diff --git a/.gitattributes b/.gitattributes index ab6f27f9a33dc5c7e704ccae7d71803ec8899258..6efc43f0f017a5c8051f3b4f6227ea0dc822bc7c 100644 --- a/.gitattributes +++ b/.gitattributes @@ -5929,3 +5929,67 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text 2023/CLIP[[:space:]]Is[[:space:]]Also[[:space:]]an[[:space:]]Efficient[[:space:]]Segmenter_[[:space:]]A[[:space:]]Text-Driven[[:space:]]Approach[[:space:]]for[[:space:]]Weakly[[:space:]]Supervised[[:space:]]Semantic[[:space:]]Segmentation/cffa99de-dd8c-4a70-9b1f-5f8553581a2e_origin.pdf filter=lfs diff=lfs merge=lfs -text 2023/CLIP[[:space:]]for[[:space:]]All[[:space:]]Things[[:space:]]Zero-Shot[[:space:]]Sketch-Based[[:space:]]Image[[:space:]]Retrieval,[[:space:]]Fine-Grained[[:space:]]or[[:space:]]Not/e8e0cef4-6aa4-43b7-996e-85dcc9fbe77b_origin.pdf filter=lfs diff=lfs merge=lfs -text 2023/CLIP[[:space:]]the[[:space:]]Gap_[[:space:]]A[[:space:]]Single[[:space:]]Domain[[:space:]]Generalization[[:space:]]Approach[[:space:]]for[[:space:]]Object[[:space:]]Detection/8bfd13ef-ca5c-4f47-a9f9-d8d7eecb88de_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/CLIP-S4_[[:space:]]Language-Guided[[:space:]]Self-Supervised[[:space:]]Semantic[[:space:]]Segmentation/c24eb4b4-f000-47c7-8244-3296ba8e2cac_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/CLIP-Sculptor_[[:space:]]Zero-Shot[[:space:]]Generation[[:space:]]of[[:space:]]High-Fidelity[[:space:]]and[[:space:]]Diverse[[:space:]]Shapes[[:space:]]From[[:space:]]Natural[[:space:]]Language/e64fac6d-0780-4018-a6c0-4f909bfb4337_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/CLIP2Protect_[[:space:]]Protecting[[:space:]]Facial[[:space:]]Privacy[[:space:]]Using[[:space:]]Text-Guided[[:space:]]Makeup[[:space:]]via[[:space:]]Adversarial[[:space:]]Latent[[:space:]]Search/d379e97e-b7a0-42bb-a735-a3de89cdb054_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/CLIP2Scene_[[:space:]]Towards[[:space:]]Label-Efficient[[:space:]]3D[[:space:]]Scene[[:space:]]Understanding[[:space:]]by[[:space:]]CLIP/58dade4c-9eec-4c3a-968d-a582ab672d68_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/CLIP2_[[:space:]]Contrastive[[:space:]]Language-Image-Point[[:space:]]Pretraining[[:space:]]From[[:space:]]Real-World[[:space:]]Point[[:space:]]Cloud[[:space:]]Data/b3efe108-020c-4b4b-bc30-fbdcfe854136_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/CLIPPING_[[:space:]]Distilling[[:space:]]CLIP-Based[[:space:]]Models[[:space:]]With[[:space:]]a[[:space:]]Student[[:space:]]Base[[:space:]]for[[:space:]]Video-Language[[:space:]]Retrieval/1eb2010b-78e6-492d-9404-5d49a6b93ef0_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/CLIPPO_[[:space:]]Image-and-Language[[:space:]]Understanding[[:space:]]From[[:space:]]Pixels[[:space:]]Only/6a9a5c21-30d9-4b55-b8d0-e491b4946877_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/CLOTH4D_[[:space:]]A[[:space:]]Dataset[[:space:]]for[[:space:]]Clothed[[:space:]]Human[[:space:]]Reconstruction/b8c1cf88-b3e6-48d2-932c-20d8bacc61eb_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/CNVid-3.5M_[[:space:]]Build,[[:space:]]Filter,[[:space:]]and[[:space:]]Pre-Train[[:space:]]the[[:space:]]Large-Scale[[:space:]]Public[[:space:]]Chinese[[:space:]]Video-Text[[:space:]]Dataset/3819ec34-d037-4584-a0be-1f25bc0cb37d_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/CODA-Prompt_[[:space:]]COntinual[[:space:]]Decomposed[[:space:]]Attention-Based[[:space:]]Prompting[[:space:]]for[[:space:]]Rehearsal-Free[[:space:]]Continual[[:space:]]Learning/44360d49-0340-4846-917d-a266d277946e_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/CORA_[[:space:]]Adapting[[:space:]]CLIP[[:space:]]for[[:space:]]Open-Vocabulary[[:space:]]Detection[[:space:]]With[[:space:]]Region[[:space:]]Prompting[[:space:]]and[[:space:]]Anchor[[:space:]]Pre-Matching/e257818b-1aea-4bd1-9288-7c6a81195cd8_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/COT_[[:space:]]Unsupervised[[:space:]]Domain[[:space:]]Adaptation[[:space:]]With[[:space:]]Clustering[[:space:]]and[[:space:]]Optimal[[:space:]]Transport/53ca0802-c684-4d48-b297-7731efdaab53_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/CP3_[[:space:]]Channel[[:space:]]Pruning[[:space:]]Plug-In[[:space:]]for[[:space:]]Point-Based[[:space:]]Networks/f7e54c81-931a-42ba-b8dd-1b55f7342f80_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/CR-FIQA_[[:space:]]Face[[:space:]]Image[[:space:]]Quality[[:space:]]Assessment[[:space:]]by[[:space:]]Learning[[:space:]]Sample[[:space:]]Relative[[:space:]]Classifiability/2ae421c8-bef3-40d8-a9e9-022734d8a26e_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/CRAFT_[[:space:]]Concept[[:space:]]Recursive[[:space:]]Activation[[:space:]]FacTorization[[:space:]]for[[:space:]]Explainability/6e462b05-7931-4758-8402-5b3e432aab49_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/CREPE_[[:space:]]Can[[:space:]]Vision-Language[[:space:]]Foundation[[:space:]]Models[[:space:]]Reason[[:space:]]Compositionally_/6b09bf56-aeff-4b97-8111-a915e425b709_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/CUDA_[[:space:]]Convolution-Based[[:space:]]Unlearnable[[:space:]]Datasets/e0179aca-27a9-460d-b895-5e7378133f26_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/CUF_[[:space:]]Continuous[[:space:]]Upsampling[[:space:]]Filters/5b0cd1d3-2b2b-407b-9bec-e263533bd51d_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/CVT-SLR_[[:space:]]Contrastive[[:space:]]Visual-Textual[[:space:]]Transformation[[:space:]]for[[:space:]]Sign[[:space:]]Language[[:space:]]Recognition[[:space:]]With[[:space:]]Variational[[:space:]]Alignment/012722bf-9201-40ff-b0cf-52287586b76f_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/CXTrack_[[:space:]]Improving[[:space:]]3D[[:space:]]Point[[:space:]]Cloud[[:space:]]Tracking[[:space:]]With[[:space:]]Contextual[[:space:]]Information/c7b22be7-4a5c-4780-8a5a-1db50c1d60af_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/CaPriDe[[:space:]]Learning_[[:space:]]Confidential[[:space:]]and[[:space:]]Private[[:space:]]Decentralized[[:space:]]Learning[[:space:]]Based[[:space:]]on[[:space:]]Encryption-Friendly[[:space:]]Distillation[[:space:]]Loss/ca732fe2-4161-4d55-be24-51fc46dde867_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/CafeBoost_[[:space:]]Causal[[:space:]]Feature[[:space:]]Boost[[:space:]]To[[:space:]]Eliminate[[:space:]]Task-Induced[[:space:]]Bias[[:space:]]for[[:space:]]Class[[:space:]]Incremental[[:space:]]Learning/6f2564b9-d9c3-4c4f-b871-b0003952027b_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Camouflaged[[:space:]]Instance[[:space:]]Segmentation[[:space:]]via[[:space:]]Explicit[[:space:]]De-Camouflaging/9acac0a8-5352-4eef-91f5-74adb1a1e4b8_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Camouflaged[[:space:]]Object[[:space:]]Detection[[:space:]]With[[:space:]]Feature[[:space:]]Decomposition[[:space:]]and[[:space:]]Edge[[:space:]]Reconstruction/69e6313d-2f4d-49ec-aeb1-e472deb81070_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Can't[[:space:]]Steal_[[:space:]]Cont-Steal![[:space:]]Contrastive[[:space:]]Stealing[[:space:]]Attacks[[:space:]]Against[[:space:]]Image[[:space:]]Encoders/d755a24e-203a-45cb-a84e-020f7fb471b0_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Canonical[[:space:]]Fields_[[:space:]]Self-Supervised[[:space:]]Learning[[:space:]]of[[:space:]]Pose-Canonicalized[[:space:]]Neural[[:space:]]Fields/54da6270-9672-4d8f-9893-3339952d04d6_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Cap4Video_[[:space:]]What[[:space:]]Can[[:space:]]Auxiliary[[:space:]]Captions[[:space:]]Do[[:space:]]for[[:space:]]Text-Video[[:space:]]Retrieval_/926c868d-a104-48e5-bcb0-c1285d26ad5d_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/CapDet_[[:space:]]Unifying[[:space:]]Dense[[:space:]]Captioning[[:space:]]and[[:space:]]Open-World[[:space:]]Detection[[:space:]]Pretraining/c3eddfd4-ae1f-40f7-a7cf-96c54f1fea7b_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Cascade[[:space:]]Evidential[[:space:]]Learning[[:space:]]for[[:space:]]Open-World[[:space:]]Weakly-Supervised[[:space:]]Temporal[[:space:]]Action[[:space:]]Localization/3c2843a4-d81f-4df7-8b2c-5c29eaeafad1_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Cascaded[[:space:]]Local[[:space:]]Implicit[[:space:]]Transformer[[:space:]]for[[:space:]]Arbitrary-Scale[[:space:]]Super-Resolution/ae047e51-87a0-4eec-807f-144baf2a48b8_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Castling-ViT_[[:space:]]Compressing[[:space:]]Self-Attention[[:space:]]via[[:space:]]Switching[[:space:]]Towards[[:space:]]Linear-Angular[[:space:]]Attention[[:space:]]at[[:space:]]Vision[[:space:]]Transformer[[:space:]]Inference/d74c49bb-effe-4caf-8858-0066e120e08d_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Catch[[:space:]]Missing[[:space:]]Details_[[:space:]]Image[[:space:]]Reconstruction[[:space:]]With[[:space:]]Frequency[[:space:]]Augmented[[:space:]]Variational[[:space:]]Autoencoder/0a253bb0-c0db-46ba-9d61-8be0d603b311_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Category[[:space:]]Query[[:space:]]Learning[[:space:]]for[[:space:]]Human-Object[[:space:]]Interaction[[:space:]]Classification/c48c8f3f-b487-4c24-815b-6385a57d02e8_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Causally-Aware[[:space:]]Intraoperative[[:space:]]Imputation[[:space:]]for[[:space:]]Overall[[:space:]]Survival[[:space:]]Time[[:space:]]Prediction/58789d3b-4b75-4c2b-ad82-075b8268b238_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/CelebV-Text_[[:space:]]A[[:space:]]Large-Scale[[:space:]]Facial[[:space:]]Text-Video[[:space:]]Dataset/b3bc8119-1bc3-419d-89b1-046a6b2d92d9_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Center[[:space:]]Focusing[[:space:]]Network[[:space:]]for[[:space:]]Real-Time[[:space:]]LiDAR[[:space:]]Panoptic[[:space:]]Segmentation/e749a0dd-41b1-4b7d-9325-ea72d56e45cb_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Change-Aware[[:space:]]Sampling[[:space:]]and[[:space:]]Contrastive[[:space:]]Learning[[:space:]]for[[:space:]]Satellite[[:space:]]Images/ef28aefe-c0fe-45ce-8981-6fc9107e6b77_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Chat2Map_[[:space:]]Efficient[[:space:]]Scene[[:space:]]Mapping[[:space:]]From[[:space:]]Multi-Ego[[:space:]]Conversations/9562f06a-6203-4663-996c-df2336adbe79_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/CiCo_[[:space:]]Domain-Aware[[:space:]]Sign[[:space:]]Language[[:space:]]Retrieval[[:space:]]via[[:space:]]Cross-Lingual[[:space:]]Contrastive[[:space:]]Learning/047059b8-c0b2-4bcc-9546-ab9b6e4bc0e5_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/CiaoSR_[[:space:]]Continuous[[:space:]]Implicit[[:space:]]Attention-in-Attention[[:space:]]Network[[:space:]]for[[:space:]]Arbitrary-Scale[[:space:]]Image[[:space:]]Super-Resolution/dcf45018-bd80-4a92-ad40-a481bde62c68_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Class[[:space:]]Adaptive[[:space:]]Network[[:space:]]Calibration/d3a59c4a-62e4-447c-b052-c6b11b8ed279_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Class[[:space:]]Attention[[:space:]]Transfer[[:space:]]Based[[:space:]]Knowledge[[:space:]]Distillation/a68b12fe-6c62-4a74-9758-0370df04b2ad_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Class[[:space:]]Balanced[[:space:]]Adaptive[[:space:]]Pseudo[[:space:]]Labeling[[:space:]]for[[:space:]]Federated[[:space:]]Semi-Supervised[[:space:]]Learning/27b4466b-7a94-415b-9a3c-c8f4f39a02f2_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Class[[:space:]]Prototypes[[:space:]]Based[[:space:]]Contrastive[[:space:]]Learning[[:space:]]for[[:space:]]Classifying[[:space:]]Multi-Label[[:space:]]and[[:space:]]Fine-Grained[[:space:]]Educational[[:space:]]Videos/bbbb89aa-912d-4dd5-8424-10e3f0db2ed4_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Class[[:space:]]Relationship[[:space:]]Embedded[[:space:]]Learning[[:space:]]for[[:space:]]Source-Free[[:space:]]Unsupervised[[:space:]]Domain[[:space:]]Adaptation/a0503f36-a779-42c8-98b9-ccfd34b3ed5c_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Class-Balancing[[:space:]]Diffusion[[:space:]]Models/eae79914-520a-4687-a1d4-1e04f53e44f8_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Class-Conditional[[:space:]]Sharpness-Aware[[:space:]]Minimization[[:space:]]for[[:space:]]Deep[[:space:]]Long-Tailed[[:space:]]Recognition/5537ac60-2d54-48e8-9091-a4f2d7a57e6f_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Class-Incremental[[:space:]]Exemplar[[:space:]]Compression[[:space:]]for[[:space:]]Class-Incremental[[:space:]]Learning/47061da4-a9f4-49a6-b036-8d247f22c620_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/CloSET_[[:space:]]Modeling[[:space:]]Clothed[[:space:]]Humans[[:space:]]on[[:space:]]Continuous[[:space:]]Surface[[:space:]]With[[:space:]]Explicit[[:space:]]Template[[:space:]]Decomposition/4725afdd-f941-4b99-a289-05a751ed5c4e_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Clothed[[:space:]]Human[[:space:]]Performance[[:space:]]Capture[[:space:]]With[[:space:]]a[[:space:]]Double-Layer[[:space:]]Neural[[:space:]]Radiance[[:space:]]Fields/edeeca28-ac72-40e3-9181-9737d0c908f2_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Cloud-Device[[:space:]]Collaborative[[:space:]]Adaptation[[:space:]]to[[:space:]]Continual[[:space:]]Changing[[:space:]]Environments[[:space:]]in[[:space:]]the[[:space:]]Real-World/071f2817-bbaa-4424-acc0-27135344a577_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Clover_[[:space:]]Towards[[:space:]]a[[:space:]]Unified[[:space:]]Video-Language[[:space:]]Alignment[[:space:]]and[[:space:]]Fusion[[:space:]]Model/eddf9a27-0fae-4139-acd3-5b79001d8460_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Co-SLAM_[[:space:]]Joint[[:space:]]Coordinate[[:space:]]and[[:space:]]Sparse[[:space:]]Parametric[[:space:]]Encodings[[:space:]]for[[:space:]]Neural[[:space:]]Real-Time[[:space:]]SLAM/98d936fc-c445-4925-9f16-8d4e0f8fc517_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Co-Salient[[:space:]]Object[[:space:]]Detection[[:space:]]With[[:space:]]Uncertainty-Aware[[:space:]]Group[[:space:]]Exchange-Masking/9cd15a5b-7ef1-4ba9-a24c-48575755f531_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Co-Speech[[:space:]]Gesture[[:space:]]Synthesis[[:space:]]by[[:space:]]Reinforcement[[:space:]]Learning[[:space:]]With[[:space:]]Contrastive[[:space:]]Pre-Trained[[:space:]]Rewards/6b1eace7-6844-4cc2-b39b-76192850eed9_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Co-Training[[:space:]]2L[[:space:]]Submodels[[:space:]]for[[:space:]]Visual[[:space:]]Recognition/51b2445a-3487-4e2d-824b-ee9b871d1274_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/CoMFormer_[[:space:]]Continual[[:space:]]Learning[[:space:]]in[[:space:]]Semantic[[:space:]]and[[:space:]]Panoptic[[:space:]]Segmentation/9493268b-7f35-48c3-8aa1-39be491edc53_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/CoWs[[:space:]]on[[:space:]]Pasture_[[:space:]]Baselines[[:space:]]and[[:space:]]Benchmarks[[:space:]]for[[:space:]]Language-Driven[[:space:]]Zero-Shot[[:space:]]Object[[:space:]]Navigation/a31a35f2-4957-4777-92dc-31eae2e5deb5_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Coaching[[:space:]]a[[:space:]]Teachable[[:space:]]Student/67002b28-2cb3-47c6-8f0a-06c8bb56b862_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/CodeTalker_[[:space:]]Speech-Driven[[:space:]]3D[[:space:]]Facial[[:space:]]Animation[[:space:]]With[[:space:]]Discrete[[:space:]]Motion[[:space:]]Prior/103a044f-c0a4-40fe-af3f-96d0d65c126f_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Collaboration[[:space:]]Helps[[:space:]]Camera[[:space:]]Overtake[[:space:]]LiDAR[[:space:]]in[[:space:]]3D[[:space:]]Detection/0d6cf930-2724-4762-9022-d5c25917e98a_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Collaborative[[:space:]]Diffusion[[:space:]]for[[:space:]]Multi-Modal[[:space:]]Face[[:space:]]Generation[[:space:]]and[[:space:]]Editing/3b7ffa39-2244-45b9-9833-f768a1b5b8c6_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Collaborative[[:space:]]Noisy[[:space:]]Label[[:space:]]Cleaner_[[:space:]]Learning[[:space:]]Scene-Aware[[:space:]]Trailers[[:space:]]for[[:space:]]Multi-Modal[[:space:]]Highlight[[:space:]]Detection[[:space:]]in[[:space:]]Movies/b5b9b0f2-713f-4176-89e7-acec845e8cbb_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Collaborative[[:space:]]Static[[:space:]]and[[:space:]]Dynamic[[:space:]]Vision-Language[[:space:]]Streams[[:space:]]for[[:space:]]Spatio-Temporal[[:space:]]Video[[:space:]]Grounding/40d01bed-f6c0-4114-9925-d43e564f9389_origin.pdf filter=lfs diff=lfs merge=lfs -text diff --git a/2023/CLIP-S4_ Language-Guided Self-Supervised Semantic Segmentation/c24eb4b4-f000-47c7-8244-3296ba8e2cac_content_list.json b/2023/CLIP-S4_ Language-Guided Self-Supervised Semantic Segmentation/c24eb4b4-f000-47c7-8244-3296ba8e2cac_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..f957c04aee2a3a4cc01adb3ff3a8d76bca49e75d --- /dev/null +++ b/2023/CLIP-S4_ Language-Guided Self-Supervised Semantic Segmentation/c24eb4b4-f000-47c7-8244-3296ba8e2cac_content_list.json @@ -0,0 +1,1571 @@ +[ + { + "type": "text", + "text": "CLIP-S $^4$ : Language-Guided Self-Supervised Semantic Segmentation", + "text_level": 1, + "bbox": [ + 138, + 130, + 828, + 151 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Wenbin He", + "bbox": [ + 174, + 180, + 266, + 196 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Suphanut Jamonnak", + "bbox": [ + 336, + 181, + 500, + 196 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Liang Gou", + "bbox": [ + 568, + 181, + 656, + 196 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Liu Ren", + "bbox": [ + 727, + 181, + 794, + 196 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Bosch Research North America & Bosch Center for Artificial Intelligence (BCAI)", + "bbox": [ + 158, + 199, + 807, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{wenbin.he2, suphanut.jamonnak, liang.gou, liu. ren}@us.bosch.com", + "bbox": [ + 199, + 219, + 764, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 268, + 312, + 282 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Existing semantic segmentation approaches are often limited by costly pixel-wise annotations and predefined classes. In this work, we present CLIP- $S^4$ that leverages self-supervised pixel representation learning and vision-language models to enable various semantic segmentation tasks (e.g., unsupervised, transfer learning, language-driven segmentation) without any human annotations and unknown class information. We first learn pixel embeddings with pixel-segment contrastive learning from different augmented views of images. To further improve the pixel embeddings and enable language-driven semantic segmentation, we design two types of consistency guided by vision-language models: 1) embedding consistency, aligning our pixel embeddings to the joint feature space of a pre-trained vision-language model, CLIP [34]; and 2) semantic consistency, forcing our model to make the same predictions as CLIP over a set of carefully designed target classes with both known and unknown prototypes. Thus, CLIP- $S^4$ enables a new task of class-free semantic segmentation where no unknown class information is needed during training. As a result, our approach shows consistent and substantial performance improvement over four popular benchmarks compared with the state-of-the-art unsupervised and language-driven semantic segmentation methods. More importantly, our method outperforms these methods on unknown class recognition by a large margin.", + "bbox": [ + 75, + 301, + 473, + 694 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 723, + 207, + 738 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Semantic segmentation aims to partition an input image into semantically meaningful regions and assign each region a semantic class label. Recent advances in semantic segmentation [6, 27, 48] heavily rely on pixel-wise human annotations, which have two limitations. First, acquiring pixel-wise annotations is extremely labor intensive and costly, which can take up to 1.5 hours to label one image [31]. Second, human annotations are often limited to a set of predefined semantic classes, with which the learned models lack the ability to recognize unknown classes [25].", + "bbox": [ + 75, + 750, + 468, + 900 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/cadd67d538f7dadd7cd4817c2093ccbe18d4c78f72380ba46acf92e465f8d87a.jpg", + "image_caption": [ + "(b) Semantic segmentation generated by different CLIP-based unsupervised approaches" + ], + "image_footnote": [], + "bbox": [ + 501, + 268, + 893, + 412 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/aec1cbba951d4512a5976fd04ee59e85632cabacb2b6ac368e140ff2bc1a30ad.jpg", + "image_caption": [ + "(c) CLIP- $S^4$ aligns the pixel embeddings and their semantics with CLIP feature space", + "Figure 1. (a) Pixel embeddings from different CLIP-based unsupervised methods: Our method, $\\mathrm{CLIP - S^4}$ , generates sharper and more coherent pixel embeddings than MaskCLIP [49] and MaskCLIP+’s [49]; (b) Language-driven semantic segmentation by different methods: $\\mathrm{CLIP - S^4}$ can recognize challenging unknown classes (e.g., moon); (c) The key idea behind $\\mathrm{CLIP - S^4}$ : aligning the pixel embeddings and their semantics with CLIP feature space." + ], + "image_footnote": [], + "bbox": [ + 503, + 422, + 888, + 501 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Various approaches have been proposed to tackle these limitations, among which we are inspired by two lines of recent research in particular. First, for unsupervised semantic segmentation (i.e., without human annotations), self-supervised pixel representation learning approaches [14,18, 19, 23, 40] have shown promising results on popular unsupervised benchmarks. The main idea is to extend self-supervised contrastive learning [7, 16] from images to pixels by attracting each pixel's embedding to its positive pairs and repelling it from negative pairs. The prior of pairs can be contours [18, 19], hierarchical groups [23], salience maps [40], and pre-trained models [14]. Although these approaches can group pixels into semantically meaningful clusters, human annotations are still needed to assign class labels to the clusters for semantic segmentation [37].", + "bbox": [ + 496, + 638, + 892, + 866 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Second, for unknown classes in semantic segmentation, large-scale vision-language models such as CLIP [34]", + "bbox": [ + 498, + 869, + 890, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 807, + 44 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "11207", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "have shown great potential. This line of research, called language-driven semantic segmentation, aims to segment images with arbitrary classes defined by texts during testing time [25, 37, 46, 49]. Among these methods, most still need training time annotations, such as pixel annotations [25] and captions [46]. Only a few recent work, MaskCLIP & MaskCLIP+ [49] attempts to address this without using additional supervision: MaskCLIP directly extracts pixel embeddings correlated with texts from CLIP, but these pixel embeddings are coarse and noisy (Fig. 1a). To address this issue, MaskCLIP+ [49] trains a segmentation model on the pseudo-labels generated by MaskCLIP for a set of predefined classes. However, the pixel embeddings of MaskCLIP+ are distorted by the predefined classes (Fig. 1a), which limits its ability to recognize unknowns (Fig. 1b). Also, it needs unknown class information during training, which hinders its real-world applications.", + "bbox": [ + 75, + 90, + 472, + 348 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We propose a language-guided self-supervised semantic segmentation approach, CLIP-S $^4$ , which takes advantage of the strengths from both lines of research and addresses their limitations accordingly. The key idea is to learn consistent pixel embeddings with respect to visual and conceptual semantics using self-supervised learning and the guidance of a vision-language model, CLIP.", + "bbox": [ + 75, + 349, + 472, + 455 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Specifically, we first train pixel embeddings with pixel-segment contrastive learning from different augmented image views [18, 19, 23] such that images can be partitioned into visually meaningful regions. To further improve pixel embedding quality and enable language-driven semantic segmentation, we introduce vision-language model guided consistency to regularize our model (Fig. 1c). The consistency is enforced from two aspects: embedding consistency and semantic consistency. First, embedding consistency aims to align the pixel embeddings generated by our model with the joint feature space of texts and images of CLIP by minimizing the distance between the pixel embeddings generated by our model and CLIP. Second, semantic consistency forces our model to make the same prediction as CLIP over a set of carefully designed target classes with both known and unknown prototypes. Note that unlike the previous methods [25,49] that use a predefined set of known classes, CLIP- $\\mathbf{S}^4$ also learns the representation of unknown classes from images during training.", + "bbox": [ + 75, + 455, + 472, + 743 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In the end, CLIP- $S^4$ also enables a new task, namely class-free semantic segmentation, as shown in Tab. 1. This new task does not need any human annotations and even assumes NO class names are given during training. This is a more challenging task than the recent work [49] that requires class names of both known and unknown.", + "bbox": [ + 75, + 744, + 470, + 835 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In summary, the contributions of this paper are threefold:", + "bbox": [ + 96, + 835, + 468, + 851 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "- We propose a self-supervised semantic segmentation approach that combines pixel-segment contrastive learning with the guidance of pre-trained vision lan", + "bbox": [ + 94, + 854, + 468, + 900 + ], + "page_idx": 1 + }, + { + "type": "table", + "img_path": "images/af8866ba4f6554164fce2a7acceea2e7e2bc3902f6c29ca08449c06e1967cbe0.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
KnownUnknownAdd. Info.
Annot.CIs NameAnnot.CIs Name
Un/Self-supervised ([19] etc.)××××Fine-Tuning
Supervised ([27] etc.)✓✓N/AN/AN/A
Zero-shot ([3] etc.)✓✓×✓Word2Vec, etc.
Language-MaskCLIP+[49]×✓CLIP
DrivenCLIP-S4×✓×CLIP
", + "bbox": [ + 500, + 88, + 893, + 242 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Table 1. Comparison of information required for training over different tasks. CLIP- $\\mathbf{S}^4$ enables a new task called class-free semantic segmentation. Compared with MaskCLIP+ [49], the new task assumes unknown class names are NOT given during training.", + "bbox": [ + 498, + 247, + 893, + 304 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "guage models. Our method can generate high-quality pixel embeddings without any human annotations and be applied to a variety of semantic segmentation tasks.", + "bbox": [ + 529, + 310, + 890, + 356 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We open up new research potentials for language-driven semantic segmentation without any human annotations by introducing and addressing a new task of class-free semantic segmentation (Tab. 1). Unlike previous work that assumes all the class names are known during training, our method can discover unknown classes from unlabelled image data without even knowing unknown class names.", + "- Consistent and substantial gains are observed with our approach over the state-of-the-art unsupervised and language-driven semantic segmentation methods on four popular datasets. More importantly, our method significantly outperforms the state-of-the-art on the segmentation of unknown classes." + ], + "bbox": [ + 517, + 357, + 890, + 570 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 587, + 640, + 603 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Unsupervised Semantic Segmentation. There are two groups of recent unsupervised semantic segmentation methods. One group of methods learns to generate consistent pixel representations or predictions between different augmentation of images with the guidance of mutual information [22, 30], clusters [8], contours [18, 19], hierarchical groups [23, 47], and saliency masks [40]. The other group of methods extracts dense features from pre-trained models based on saliency maps [36], augmentations [39], spectral decomposition [28], and feature correspondences [14]. While these methods can generate pixel embeddings with semantically meaningful clusters, annotations are needed to assign class labels to the clusters (e.g., $k$ -nearest neighbor search [19] and Hungarian algorithm [40]). Our work combines pixel-segment self-supervision with pre-trained vision-language models to enable semantic segmentation without any human annotations.", + "bbox": [ + 496, + 613, + 890, + 869 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Language-Driven Semantic Segmentation. Recently, vision-language models (e.g., CLIP [34]) trained on large-", + "bbox": [ + 500, + 869, + 890, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "11208", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/c70f39b4ee98291a5d3b8dfef69c8d525810cae5961988bd0bf15d23fa91c1f0.jpg", + "image_caption": [ + "Figure 2. Framework of CLIP-S4. (a) It starts with an encoder that maps images into pixel embeddings for semantic segmentation. Then it follows with two components: 1) self-supervised contrastive learning and 2) vision-language model guided consistency. Specifically, shown in (b), the self-supervised contrastive learning forces pixel embeddings to be consistent within visually coherent regions and among different augmented views of the same image. For vision-language model guided consistency, this framework introduces (c) embedding consistency that aligns the pixel embeddings generated by our model with CLIP embeddings, and (d) semantic consistency that forces our model to make the same predictions as CLIP for a set of target classes with both known and unknown prototypes. For known classes, the prototypes are pre-computed and fixed during training. For unknown classes, the prototypes are learned during training via clustering (e)." + ], + "image_footnote": [], + "bbox": [ + 80, + 87, + 893, + 292 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "scale image-text datasets have shown great potential on various downstream tasks such as image synthesis [21, 41], out-of-distribution detection [11], and object detection [13]. To extend vision-language models for semantic segmentation, one active research, language-driven semantic segmentation, aims to segment images with arbitrary unknown classes defined by texts during testing time [25, 37, 46, 49]. Some methods [25, 44] use pixel-wise annotations to train language-guided semantic segmentation models. Other methods [10, 46] perform large-scale pre-training on image-text pairs specifically for semantic segmentation.", + "bbox": [ + 75, + 401, + 470, + 566 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "By contrast, we directly use vision-language models that are pre-trained for classification tasks. Along this line of research, a few approaches have been proposed [35, 37, 49]. The most relevant approach to our method is MaskCLIP [49], which extends the embeddings generated by pre-trained vision-language models from image to pixel level, but these embeddings are often coarse and noisy. To address this issue, MaskCLIP + [49] fine-tunes the pixel embeddings by the pseudo-labels of a specific set of classes on top of MaskCLIP. However, it needs unknown class names during training, which may not be possible in real-world cases. Compared with [49], our method can recognize unknown classes without knowing any unknown class information during training time, and also learns fine-grained and sharper pixel embeddings with self-supervision.", + "bbox": [ + 75, + 569, + 468, + 795 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 76, + 813, + 168, + 828 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our method (Fig. 2) segments images by learning a pixel embedding function with self-supervised contrastive learning and the guidance of a pre-trained vision-language model, CLIP. We use self-supervised contrastive learning", + "bbox": [ + 75, + 839, + 470, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "to force the consistency of pixel embeddings within visually coherent regions (e.g., superpixels) and among different augmented views of the same image (Sec. 3.1). We also introduce two vision-language model guided consistency (i.e., embedding consistency and semantic consistency) to further regularize the model (Sec. 3.2). The two components are complementary to each other. On the one hand, contrastive learning mitigates the noise introduced by CLIP. On the other hand, with the knowledge extracted from CLIP, the quality of the pixel embeddings can be improved. More importantly, this approach enables us to perform language-driven semantic segmentation with our carefully designed target class prototypes of both knowns and unknowns. In the following, we discuss the two components in detail.", + "bbox": [ + 496, + 401, + 893, + 613 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Pixel-Segment Contrastive Learning", + "text_level": 1, + "bbox": [ + 500, + 619, + 815, + 636 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We train a pixel embedding function to generate consistent pixel embeddings within visually coherent regions through pixel-segment contrastive learning [18,23]. Specifically, the embedding function transforms each pixel $p$ of an image to a unit-length embedding vector $\\mathbf{z}_p$ of dimension $d$ via a deep neural network. The image is then partitioned into $|S|$ segments by clustering the pixel embeddings. The embedding $\\mathbf{v}_s$ of each segment $s$ is calculated as the average of the pixel embeddings $\\mathbf{v}_s = \\sum_{p\\in s}\\mathbf{z}_p / |s|$ , which is also normalized into a unit-length vector $\\mathbf{v}_s = \\mathbf{v}_s / \\| \\mathbf{v}_s\\|$ . For each pixel $p$ , the segments are grouped into two sets including a positive set $S^+$ and a negative set $S^-$ . The positive set $S^+$ of a pixel contains segments within the same visually coherent region of the pixel. Following the prior work [18, 23], the visually coherent region can be derived from super-pixels [1] or contours [2]. We also use data augmentation (e.g., random cropping and color jitter", + "bbox": [ + 496, + 643, + 890, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "11209", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "ing) to generate consistent pixel embeddings between different augmented views of the same image. Hence, segments within the same region of the pixel in any augmented views are considered as the positive set $S^{+}$ . Other segments in the image and segments from other images in the same batch are included in the negative set $S^{-}$ . The pixel embedding $\\mathbf{z}_p$ is then attracted to the segments in positive set $S^{+}$ and repelled from the segments in negative set $S^{-}$ with contrastive loss:", + "bbox": [ + 75, + 90, + 472, + 224 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {t} (p) = - \\log \\frac {\\sum_ {s \\in \\mathcal {S} ^ {+}} e x p \\left(\\operatorname {s i m} \\left(\\mathbf {z} _ {p} , \\mathbf {v} _ {s}\\right) \\kappa\\right)}{\\sum_ {s \\in \\mathcal {S} ^ {+} \\cup \\mathcal {S} ^ {-}} e x p \\left(\\operatorname {s i m} \\left(\\mathbf {z} _ {p} , \\mathbf {v} _ {s}\\right) \\kappa\\right)}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 114, + 237, + 468, + 273 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\kappa$ is the concentration constant and $sim(\\mathbf{z}_p, \\mathbf{v}_s)$ is the cosine similarity between the pixel embedding $\\mathbf{z}_p$ and the segment embedding $\\mathbf{v}_s$ .", + "bbox": [ + 75, + 285, + 468, + 330 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Vision-Language Model Guided Consistency", + "text_level": 1, + "bbox": [ + 76, + 344, + 454, + 359 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To enable language-driven semantic segmentation and improve the quality of pixel embeddings, we use a pretrained vision-language model such as CLIP [34] to guide the training of the pixel embedding function. The key idea is to align the output space of our pixel embedding function consistent with the feature space of CLIP (Fig. 1c). Specifically, two types of consistency are considered during training including embedding consistency and semantic consistency, which are detailed as follows.", + "bbox": [ + 75, + 368, + 468, + 503 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "**Embedding Consistency.** Our goal is to align the pixel embeddings generated from our self-supervised method (the green contour in Fig. 1c) with CLIP's pixel embeddings (the orange contour in Fig. 1c). This is done by minimizing the distance between the two pixel embedding spaces.", + "bbox": [ + 75, + 506, + 468, + 580 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We first obtain the pixel embeddings of an input image from CLIP by modifying the attention-based pooling layer of the CLIP image encoder following [49]. Specifically, we 1) remove the query and key projection layers and 2) reformulate the value projection layer and the last linear layer as two consecutive fully connected layers. In the following, we use $clip - i(\\cdot)$ as the modified CLIP image encoder and $clip - t(\\cdot)$ as CLIP text encoder.", + "bbox": [ + 75, + 582, + 468, + 703 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Then we obtain the pixel embeddings of CLIP for different augmented views of the image. Note that we use the original image to generate the CLIP pixel embeddings and perform augmentation afterwards to make sure that the CLIP pixel embeddings are consistent among different augmented views. In the end, we minimize the distance of embeddings between segments instead of pixels from our self-supervised and CLIP embedding spaces. This is because the pixel embeddings of CLIP are noisy (Fig. 2), which can be mitigated by aggregating over segments. Hence, we use the pixel embeddings generated by our model to derive segments (clusters) and then apply them to the CLIP's pixel embeddings. In the end, for each segment $s$ , the embedding", + "bbox": [ + 75, + 704, + 470, + 901 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/e1c79c194e1189f50624c1c920fb3868b965be53cee49adf0b56be8721db3348.jpg", + "image_caption": [ + "Figure 3. Computation of target class prototypes with both knowns and unknowns, $C = \\{C_k, C_u\\}$ . For a set of known classes (e.g., bird, cat), we first obtain their CLIP text embeddings, $T$ , via a set of prompt templates [13,49] (shown in $\\widehat{\\mathbf{a}}$ ); then, we calculate the normalized (via softmax) similarity between text embeddings, $T$ , and all segments' CLIP embeddings, $I$ , from training images, and average the top- $m$ similar segments' CLIP embedding as the embedding prototype for this class, as shown in $\\widehat{\\mathbf{b}}$ ; For each unknown class, we randomly select the CLIP embedding of a segment as the initial prototype (shown in $\\widehat{\\mathbf{c}}$ )." + ], + "image_footnote": [], + "bbox": [ + 498, + 87, + 893, + 301 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "consistent loss is defined as:", + "bbox": [ + 500, + 455, + 689, + 469 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {e} (s) = 1 - \\operatorname {s i m} \\left(\\mathbf {v} _ {s}, \\mathbf {i} _ {s}\\right), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 607, + 486, + 890, + 502 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathbf{v}_s$ and $\\mathbf{i}_s$ are the segment embeddings derived from our embedding function and CLIP, respectively. Here, $\\mathbf{i}_s$ is the average of the CLIP pixel embedding from segment $s$ , namely, $\\mathbf{i}_s = \\sum_{p \\in s} \\text{clip-}i(s) / |s|$ .", + "bbox": [ + 496, + 517, + 890, + 579 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Semantic Consistency In addition to embedding consistency, we introduce semantic consistency by forcing our model to make the same predictions of semantic classes as CLIP. The rationale is that we can generate better pixel embeddings if they can form distinctive clusters corresponding to different semantic classes, as the goal of semantic segmentation is to perform pixel-wise classification.", + "bbox": [ + 496, + 580, + 890, + 685 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Semantic consistency is achieved via a similar idea of pseudo-labeling [38]. Again, we force the semantic consistency at the segment level (not the pixel level) to reduce the noise in pseudo-labels. Specifically, for each segment $s$ , we first use CLIP to generate its pseudo-label $y_{s}$ over a set of target classes, which include both knowns and unknowns (we will introduce how to design these target classes later). The pseudo-label is generated based on the highest similarity between the segment embedding $\\mathbf{i}_s$ with a set of prototypes, $C = \\{\\mathbf{c}_l\\}_{0}^{L - 1}$ , of the target classes in the pixel embedding space of CLIP, namely, $y_{s} = \\operatorname{argmax}_{l\\in L}(sim(\\mathbf{i}_{s},\\mathbf{c}_{l}))$ .", + "bbox": [ + 496, + 686, + 892, + 869 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Then we define the semantic consistent loss as the cross entropy between our model's prediction $\\varphi (\\mathbf{v}_s)$ over the tar", + "bbox": [ + 498, + 869, + 890, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "11210", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "get classes and the pseudo-label $y_{s}$ :", + "bbox": [ + 76, + 90, + 313, + 106 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {s} (s) = \\mathbf {H} \\left(y _ {s}, \\varphi (\\mathbf {v} _ {s})\\right), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 192, + 117, + 468, + 135 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\varphi (\\mathbf{v}_s) = \\mathrm{softmax}(sim(\\mathbf{v}_s,C))$", + "bbox": [ + 76, + 145, + 346, + 161 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Target Class Prototypes The design of target classes and associated class prototypes, $C = \\{\\mathbf{c}_l\\}_{0}^{L - 1}$ , is crucial to achieve the semantic consistency. Here, a class prototype, $\\mathbf{c}_l$ , is an embedding vector that can represent a class in an embedding space. For example, it can be the mean vector of embeddings of all segments of a class \"car\". Currently, most existing methods [25, 49] assume that the target classes are already predefined, which is not feasible in real-world use cases without any human annotations. Thus, those methods cannot handle unknown classes hidden in the data. To address this issue, we introduce two sets of class prototypes of known, $C_k = \\{\\mathbf{c}_0,\\dots ,\\mathbf{c}_{k - 1}\\}$ , and unknown classes, $C_u = \\{\\mathbf{c}_k,\\dots ,\\mathbf{c}_{k + u}\\}$ , where the known classes are predefined by leveraging CLIP and the unknown classes are learned from image data during training. Thus, we have $C = \\{\\mathbf{c}_l\\}_{0}^{L - 1} = \\{\\mathbf{c}_0,\\dots ,\\mathbf{c}_{k - 1},\\mathbf{c}_k,\\dots ,\\mathbf{c}_{k + u}\\}$ $L = k + u$ .", + "bbox": [ + 75, + 161, + 468, + 402 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For known classes, a natural choice is to use the text embeddings generated by CLIP as their class prototype embeddings [25, 49]. However, even though the text embeddings are trained to align with image/pixel embeddings [34], there is still a huge gap between the text and image/pixel embeddings in the joint space of CLIP (Fig. 1c). Therefore, it is challenging to learn meaningful unknown classes from image features when using text embeddings as class prototypes. Hence, in this work, we use the prototype of CLIP pixel embeddings to represent each known class.", + "bbox": [ + 75, + 402, + 468, + 551 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To this end, for a set of known classes (e.g., bird, cat), $K = \\{0,\\dots ,k - 1\\}$ we first obtain their CLIP text embeddings, $T = \\{\\mathbf{t}_k\\} = \\{\\text{clip - } t(k)\\}$ , via a set of prompt templates following [13, 49], as shown in Fig. 3a. We also get a set of CLIP segment embeddings, $I = \\{\\mathbf{i}_{\\hat{s}}\\}$ , for all training images by a) feeding training images into the modified image encoder of CLIP to get pixel embeddings; b) clustering the pixel embeddings as segments, $\\hat{S}$ ; c) averaging the pixel embeddings in each segment, $\\hat{s}$ . Hence, we have embeddings for each segment: $\\mathbf{i}_{\\hat{s}} = \\sum_{p\\in \\hat{s}}\\text{clip - } i(p) / |\\hat{s}|$ . Then, we calculate the similarity between text embeddings of known classes, $T$ , and all CLIP segment embeddings $I$ , and normalize the similarities over all classes by softmax. Finally, we average the top- $m$ similar segments' embedding as the embedding prototype for each class, $C_k = \\{\\mathbf{c}_k\\} = avg_m(top-m_{\\hat{s}}(\\mathbf{softmax}_k(sim(I,T))))$ .", + "bbox": [ + 75, + 553, + 468, + 794 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The prototype embeddings of the unknown classes, $C_u$ , are randomly initialized by sampling the CLIP embeddings of all segments, namely $C_u = \\text{random}(clip - i(\\hat{\\mathcal{S}}), u)$ , with a size of the unknown class of $u$ (Fig. 3c). During training, the embedding $\\mathbf{c}_u$ of each unknown class prototype is updated by minimizing its distance to all segments that are classified as this unknown class (similar to updating the cen", + "bbox": [ + 75, + 795, + 468, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "troids in $k$ -means clustering):", + "bbox": [ + 498, + 90, + 697, + 106 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {u} = \\sum_ {s \\in S _ {u}} (1 - \\operatorname {s i m} (\\mathbf {c} _ {u}, \\operatorname {c l i p} - i (s))) / | S _ {u} |, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 553, + 117, + 890, + 150 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $S_{u}$ are the segments classified as the unknown classes. In this way, our model can also learn the pixel representation of unknown classes.", + "bbox": [ + 496, + 162, + 890, + 205 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Training and Inference", + "text_level": 1, + "bbox": [ + 498, + 217, + 714, + 233 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In summary, we train the pixel embedding function by combining the pixel-segment contrastive loss, embedding consistent loss, and semantic consistent loss:", + "bbox": [ + 496, + 241, + 890, + 285 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\mathcal {L} _ {t} + \\mathcal {L} _ {e} + \\mathcal {L} _ {s}. \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 627, + 299, + 890, + 315 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "During training, we also update the embeddings for the unknown classes with $\\mathcal{L}_u$ .", + "bbox": [ + 496, + 327, + 890, + 356 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For inference, we use the trained model to generate pixel embeddings for each input image and use the pixel embeddings for different downstream tasks, including language-driven and unsupervised semantic segmentation. For language-driven semantic segmentation, we first obtain the text embeddings of arbitrary inference classes by feeding the prompt-engineered texts into the text encoder of CLIP. Then we assign each pixel with the class label whose text embedding is the closest to $\\mathrm{CLIP - S^4}$ pixel embedding. For unsupervised semantic segmentation, we follow previous work [19, 40] that uses $k$ nearest neighbor search or linear classifier to perform semantic segmentation.", + "bbox": [ + 496, + 358, + 890, + 539 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 498, + 554, + 633, + 570 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We evaluate our model on three tasks: 1) language-driven semantic segmentation for both known and unknown classes; 2) unsupervised semantic segmentation with $k$ -means clustering/linear classification; 3) transfer learning of generated pixel embeddings for instance mask tracking. We also conduct ablation studies to understand the components of our model.", + "bbox": [ + 496, + 579, + 890, + 684 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Datasets", + "text_level": 1, + "bbox": [ + 500, + 695, + 601, + 709 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Pascal VOC 2012 [12] contains 20 object classes and a background class. It has 1,464 and 1,449 images for training and validation, respectively. Following common practice [27, 48], we augment the training data with additional annotations [15], resulting in 10,582 training images.", + "bbox": [ + 496, + 718, + 890, + 794 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Pascal Context [29] extends Pascal VOC 2010 [12] with additional annotations on 4,998 training and 5,105 validation images. Following the prior work [49], we use the most common 59 classes for evaluation.", + "bbox": [ + 496, + 794, + 890, + 853 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "COCO-Stuff [5] labels MS COCO [26] with 171 object/stuff classes. It contains 118,287 and 5,000 images for training and validation, respectively.", + "bbox": [ + 496, + 854, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "11211", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "DAVIS 2017 [33] contains video sequences for instance mask tracking. Following the prior work [18, 47], we train pixel embeddings on Pascal VOC 2012 and evaluate the validation sequences without fine-tuning.", + "bbox": [ + 76, + 90, + 467, + 151 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "It is worth mentioning that no ground truth labels of any datasets are used during training. Instead, we perform self-supervised learning on pseudo segments generated by contour detectors and owt-ucm [2]. Two contour detectors are used including HED [45] for Pascal VOC 2012 and Pascal Context and PMI [20] for COCO-Stuff.", + "bbox": [ + 75, + 152, + 467, + 242 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Implementation Details", + "text_level": 1, + "bbox": [ + 76, + 253, + 294, + 268 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For self-supervised contrastive learning, images are augmented with the same set of data augmentations as SimCLR [7], including random resizing, cropping, flipping, color jittering, and Gaussian blurring. The concentration constant $\\kappa$ is set to 10, and the number of segments is set to 36 for each augmented view.", + "bbox": [ + 75, + 277, + 467, + 366 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For vision-language guidance, we use pre-trained CLIP models [34] with modified image encoders following [49]. We use prompt-engineered texts with 85 prompt templates to generate text embeddings following [13, 49]. We use the average embedding of the top 32 segments of high probabilities as the prototype of each known class. We set the number of unknown classes to $u = 64$ .", + "bbox": [ + 75, + 367, + 467, + 470 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Following the prior work [18, 19, 23], we use PSP-Net [48] with a dilated ResNet-50 [17] backbone as the network architecture. The backbone is pre-trained on the ImageNet [9] dataset. We train our model on Pascal VOC 2012 and Pascal Context for $20\\mathrm{k}$ iterations and on COCO-Stuff for $80\\mathrm{k}$ iterations. We set the batch size to 8 with additional memory banks that cache the segment embeddings of the previous 2 batches. We set the initial learning rate to 0.001 and decay it with a polynomial learning rate policy. We use the CLIP model trained with ViT-B/16 backbone unless otherwise stated.", + "bbox": [ + 75, + 473, + 467, + 638 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Language-Driven Semantic Segmentation", + "text_level": 1, + "bbox": [ + 76, + 650, + 431, + 666 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For language-driven semantic segmentation, no human annotations are used for either training or inference. At the inference time, each pixel is assigned an arbitrarily given class label whose CLIP text embedding is the closest to this pixel's CLIP- $\\mathbf{S}^4$ embedding.", + "bbox": [ + 75, + 672, + 467, + 748 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We first compare the performance of our method with the state-of-the-art language-driven semantic segmentation approaches [37, 46, 49] on the Pascal Context and COCO-Stuff datasets. The performance is evaluated with the mean Intersection over Union (mIoU). For MaskCLIP and MaskCLIP+ [49], we obtain the results using the same hyper-parameter setting as our approach with CLIP models of two different backbones of ResNet50 and ViT-B/16. Meanwhile, GroupViT [46], ReCo [37], and $\\mathrm{ReCo + }$ [37] use completely different training mechanisms compared", + "bbox": [ + 75, + 750, + 467, + 900 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/76cce9ee09455eb2b39f22c62c07754e55fc7524d556ea13e8ebfb0e259159f2.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodCLIP ModelPascal ContextCOCO-Stuff
mIoUmIoU
GroupViT [46]-22.4-
ReCo [37]ResNet50x16 + ViT-L/14@336px26.6-
ReCo+ [37]†-18.4
MaskCLIP [49]ResNet5018.610.6
ViT-B/1625.215.2
MaskCLIP+ [49]†ResNet5023.413.9
ViT-B/1632.220.7
CLIP-S4†ResNet5028.5 (+5.1)16.7 (+2.8)
ViT-B/1633.6 (+1.4)22.1 (+1.4)
", + "bbox": [ + 500, + 88, + 897, + 292 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 2. Language-guided semantic segmentation benchmarks (mIoU). CLIP- $S^4$ consistently outperforms the state-of-the-art methods on both Pascal Context and COCO-Stuff datasets with CLIP models of different backbones. $\\dagger$ indicates the models are fine-tuned on target datasets.", + "bbox": [ + 498, + 299, + 890, + 369 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "with our method. GroupViT is trained on image-caption pairs, and $\\mathrm{ReCo} / \\mathrm{ReCo}+$ combines image retrieval and cosegmentation. For comparison, we take the best results from [37, 46] for GroupViT, ReCo, and $\\mathrm{ReCo}+$ . Tabel 2 shows the benchmarking results of the aforementioned methods. Our method consistently outperforms the state-of-the-art on both datasets with CLIP models of different backbones.", + "bbox": [ + 496, + 382, + 890, + 502 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To evaluate models' performance for class-free semantic segmentation with both known and unknown classes, we split the 59 classes of Pascal Context into 4 folds, where each fold includes around 15 classes. For each experiment, classes from one fold are considered as unknown and excluded during training. The mIoUs of known and unknown classes, as well as their harmonic mean (hIoU) are reported in Tab. 3. The performance of our method is averaged across 5 runs with randomly initialized prototypes of unknown classes. Our method achieves significant gains over MaskCLIP+ on unknown classes, which are comparable to MaskCLIP. Also, our method outperforms both MaskCLIP and MaskCLIP+ on known classes, which leads to better overall performance.", + "bbox": [ + 496, + 506, + 890, + 717 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Qualitatively, the visualization in Fig. 4a offers us some insights into why our approach can achieve better results: our model yields consistent embeddings aligned with the pre-trained CLIP model. Fig. 4a visualizes the projection of segment embeddings generated by different methods on Pascal Context and COCO-Stuff. We observe that segment embeddings generated by $\\mathrm{MaskCLIP + }$ are distorted by the given text embeddings. Meanwhile, $\\mathrm{CLIP - S^4}$ generates segment embeddings that are well aligned with the segment embeddings derived from the pre-trained CLIP model. Hence, segment embeddings generated by $\\mathrm{CLIP - S^4}$ can better capture both known and unknown classes. Fig. 4b shows", + "bbox": [ + 496, + 719, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "11212", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/d7701a5f3e1417ebfd98891b1d8054f8c67d051aaabee2afdcf12a019feed15c.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Methodfold0fold1fold2fold3
mIoUumIoUkhIoUmIoUumIoUkhIoUmIoUumIoUkhIoUmIoUumIoUkhIoU
MaskCLIP29.723.726.323.725.724.623.925.724.723.425.824.5
MaskCLIP+3.628.56.33.029.25.44.829.28.24.529.97.8
CLIP-S432.0±0.829.4±0.330.6±0.522.3±0.932.8±0.426.5±0.622.4±0.532.1±0.526.4±0.428.6±0.831.5±0.230.0±0.5
vs. MaskCLIP++28.4+0.9+24.3+19.3+3.6+21.1+17.6+2.9+18.2+24.1+1.6+22.2
", + "bbox": [ + 78, + 88, + 890, + 183 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b38239396bfa302ea59816bd846b1562552de090f3edcca588b1096df5ef28e8.jpg", + "image_caption": [ + "(a) Embeddings projection for PASCAL Context (left) and COCO-Stuff (right)" + ], + "image_footnote": [], + "bbox": [ + 80, + 263, + 468, + 353 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/9a82d32527abf1564a23ee49c7a1193fcd48a0af523e48705c30de0549f21374.jpg", + "image_caption": [ + "(b) Retrieved image segments for classes not used in training" + ], + "image_footnote": [], + "bbox": [ + 78, + 368, + 468, + 469 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Figure 4. (a) Projection of pixel embeddings generated by CLIP, MaskCLIP+, and CLIP-S $^4$ trained on Pascal Context with CLIP's ViT-B/16 model (left) and COCO-Stuff with CLIP's ResNet50 model (right). MaskCLIP+ distorts pixel embeddings with respect to the given text embeddings, while CLIP-S $^4$ aligns pixel embeddings with the pre-trained CLIP model. (b) Image segments retrieved for classes that are not used during training. Correct and incorrect retrievals are outlined in green and orange, respectively. Compared with MaskCLIP+, CLIP-S $^4$ retrieves images that are more closely related to the classes.", + "bbox": [ + 75, + 484, + 468, + 623 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "image segments retrieved from the COCO-Stuff validation set using MaskCLIP+ and CLIP-S $^4$ for a set of classes that are not used in training. For each class, we obtain its text embedding and compare it with segment embeddings to obtain top retrievals from both methods. Due to better alignment with the pre-trained CLIP model, CLIP-S $^4$ retrieves images that are more closely related to the unknown classes compared with MaskCLIP+.", + "bbox": [ + 75, + 626, + 468, + 748 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Unsupervised Semantic Segmentation", + "text_level": 1, + "bbox": [ + 76, + 756, + 401, + 772 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To study whether $\\mathrm{CLIP - S^4}$ can generate pixel embeddings that form distinctive clusters, we evaluate $\\mathrm{CLIP - S^4}$ on the unsupervised semantic segmentation task for Pascal VOC 2012. To derive semantic segmentation from pixel embeddings, we use and test two approaches, including $k$ nearest neighbor ( $k$ -NN) search [19] and linear classification [40]. For $k$ -NN search, we assign each segment a class label by the majority vote of its nearest neighbors from the", + "bbox": [ + 75, + 779, + 470, + 901 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/4b9bf12ae497ed3b31f7890395e4b14a9732db31ddf0577c9817e810e8a6c6ae.jpg", + "table_caption": [ + "Table 3. Language-guided semantic segmentation benchmarks (mIoU) for unknown classes. The classes of Pascal Context are split into 4 folds with around 15 classes each fold. For each experiment, classes of one fold are considered as unknown. The performance of CLIP- $S^4$ is averaged over 5 runs with randomly initialized unknown class embeddings. CLIP- $S^4$ significantly outperforms $\\text{MaskCLIP}+$ on unknown classes. Meanwhile, CLIP- $S^4$ archives consistent gains on known classes over $\\text{MaskCLIP}$ and $\\text{MaskCLIP}+$ , and hence leads to better overall performance." + ], + "table_footnote": [], + "table_body": "
MethodmIoU k-NNmIoU Linear Classifier
IIC [22]-28.0
SegSort [19]47.355.4
Hierarch. Group. [47]-48.8
MaskContrast (Sup.) [40]53.963.9
ConceptContrast [18]58.860.4
HSG [23]61.7-
MaskCLIP [49]67.369.5
MaskCLIP+ [49]65.170.0
CLIP-S472.0 (+4.7)73.0 (+3.0)
", + "bbox": [ + 506, + 263, + 885, + 436 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 4. Unsupervised semantic segmentation benchmarks (mIoU) on Pascal VOC 2012. CLIP- $S^4$ consistently outperforms the state-of-the-art methods on both $k$ -NN search and linear classification.", + "bbox": [ + 498, + 443, + 890, + 497 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "training set following [19]. For linear classification, we train a linear classifier on the learned pixel embeddings following [40].", + "bbox": [ + 498, + 503, + 890, + 547 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We compare our method with the state-of-the-art unsupervised and language-guided semantic segmentation approaches. We train the state-of-the-art models using the same hyper-parameter setting as our approach except for IIC [22] and Hierarchical Grouping [47] as they use different training mechanisms. For comparison, we take the best results for IIC and Hierarchical Grouping. The benchmark results are shown in Tab. 4. With the vision-language guidance, our method achieves significant gains compared with the previous non-CLIP-based approaches (i.e., $+9\\%$ for both $k$ -NN search and linear classification). Meanwhile, our method also outperforms the language-guided semantic segmentation approaches by a large margin.", + "bbox": [ + 496, + 549, + 892, + 746 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.5. Instance Mask Tracking", + "text_level": 1, + "bbox": [ + 500, + 756, + 723, + 772 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We evaluate the transferability of pixel embeddings learned from the Pascal VOC 2012 dataset. We use the pixel embeddings to track instance masks in the DAVIS 2017 validation set, where the instance masks at the first frame are given for each video. Following the prior work [48], we use the similarity between pixel embeddings cross frames to propagate the instance masks to the rest of the video frames. We evaluate the performance using the region similarity $\\mathcal{I}$", + "bbox": [ + 496, + 779, + 892, + 901 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "11213", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/e1d1565331bc282ab95547a9e31975745fe237b77298d7e8764364e7ff9ad168.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodJ(Mean)↑F(Mean)↑
MaskTrack-B [32]35.336.4
OSVOS-B [4]18.530.0
Video Colorization [42]34.632.7
CycleTime [43]41.939.4
mgPFF [24]42.246.9
Hierarchy. Group. [47]47.148.9
MaskContrast (Sup.) [40]34.336.7
ConceptContrast [18]50.453.9
MaskCLIP [49]48.149.2
MaskCLIP+ [49]42.644.2
CLIP-S452.3 (+1.9)56.8 (+2.9)
", + "bbox": [ + 89, + 88, + 455, + 281 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/02057dafdee078c008f55994e339ed50f99eb0bbf5bad751a806b01264aa2d6c.jpg", + "table_caption": [ + "Table 5. Quantitative evaluation of instance mask tracking on the DAVIS-2017 validation set. The performance is measured by the region similarity $\\mathcal{J}$ (IoU) and the contour-based accuracy $\\mathcal{F}$ defined by [33]. Our method outperforms existing supervised, unsupervised, and language-guided approaches on both metrics." + ], + "table_footnote": [], + "table_body": "
LtLeLspAccmIoUavgsim
✓--1.60.5-0.01
✓✓-48.124.30.79
✓-✓52.332.90.33
-✓✓48.631.3-
✓✓✓53.733.60.66
", + "bbox": [ + 101, + 369, + 444, + 465 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "(IoU) and the contour-based accuracy $\\mathcal{F}$ defined by [33].", + "bbox": [ + 76, + 578, + 452, + 593 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We compare our method with existing supervised [4, 32], unsupervised [18, 24, 40, 42, 43, 47], and language-guided [49] approaches (Tab. 5). Though not trained on any video sequences, our method outperforms the existing approaches by more than $1.9\\%$ and $2.9\\%$ in terms of the region similarity $\\mathcal{I}$ and contour accuracy $\\mathcal{F}$ , respectively. Note that the pixel embeddings generated by $\\mathrm{MaskCLIP + }$ [49] are distorted by the classes from Pascal VOC 2012, which hinder their transferability.", + "bbox": [ + 75, + 594, + 467, + 729 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.6. Ablation Study", + "text_level": 1, + "bbox": [ + 76, + 739, + 230, + 757 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We study the contribution of different losses of our method using the Pascal Context dataset and the language-guided semantic segmentation task. The performance is evaluated with pixel accuracy (pAcc) and mIoU. We also calculate the average cosine similarity (avgsim) between our segment embeddings and CLIP's segment embeddings to quantify the alignment. Tab. 6 shows the study results. We observe that by introducing embedding consistent loss $\\mathcal{L}_e$ the learned segment embeddings are well aligned with", + "bbox": [ + 75, + 763, + 468, + 902 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/843984697e9c406da2bd2251d274c734f4c475df4cbceb0d4a88360c90cb1f81.jpg", + "table_caption": [ + "Table 6. Ablation study on the contribution of each loss of CLIP-S4. Experimented on language-guided semantic segmentation of Pascal Context. avgsim represents the average cosine similarity between segment embeddings generated by CLIP-S4 and CLIP. By combining the embedding and semantic consistent losses $\\mathcal{L}_e$ and $\\mathcal{L}_s$ , CLIP-S4 archives better semantic segmentation performance while maintaining the alignment with CLIP's embeddings." + ], + "table_footnote": [], + "table_body": "
#unknowns(u)mIoU
1671.6
3271.9
6472.0
12872.2
25671.9
", + "bbox": [ + 509, + 88, + 674, + 190 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/79c7a2b889ed6be0a126b6dc3cb5d9b1efdde0aaf17d673b347e6498f3ddc233.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
top-m segmentsavgsim
10.921
40.977
160.996
640.997
2560.993
", + "bbox": [ + 691, + 88, + 880, + 190 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 7. Ablation study on the influence of the number of unknown class prototypes.", + "bbox": [ + 498, + 196, + 683, + 239 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 8. Ablation study on different numbers of top- $m$ segments for class prototypes.", + "bbox": [ + 687, + 196, + 885, + 239 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "CLIP's embeddings with an average cosine similarity of 0.79. However, the learned segment embeddings do not perform well on the language-guided semantic segmentation task (24.3 vs. 33.6), because the segment embeddings are not optimized to classify target classes. Meanwhile, by using semantic consistent loss $\\mathcal{L}_s$ without embedding consistent loss, the learned segment embeddings have the discriminative power to classify different classes but are not aligned with CLIP's embeddings as the average cosine similarity is 0.33. As a result, the segment embeddings are limited to the target classes used during training. Hence, we combine $\\mathcal{L}_e$ and $\\mathcal{L}_s$ to balance the discriminative power over target classes and the alignment with CLIP. Meanwhile, we observe that with pixel-segment contrastive learning, the model can archive better performance.", + "bbox": [ + 496, + 242, + 890, + 468 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Also, we study the influence of the number of unknown class prototypes on the Pascal VOC dataset for the unsupervised semantic segmentation task. The results Tab. 7 show that the semantic segmentation performance is robust to the tested number of unknown class prototypes as the mIoU varies only $0.6\\%$ . Furthermore, we investigate how the size of top- $m$ segments impacts the embeddings of class prototypes. We compare the embeddings of class prototypes generated with different numbers of top- $m$ segments on the Pascal VOC dataset. We use the embeddings of class prototypes generated with $m = 32$ segments as the reference, and compute the cosine similarity between the reference prototypes embeddings and ones with different top- $m$ segments. For each case, the cosine similarity is averaged over all class prototypes. We observe that the embeddings of class prototypes are relatively stable if moderate top- $m$ segments (e.g., $m = 32$ in this work) are used (Tab. 8).", + "bbox": [ + 496, + 469, + 890, + 724 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 739, + 617, + 753 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We propose CLIP-S $^4$ , a novel pixel representation learning approach for semantic segmentation. Our method combines self-supervised contrastive learning and guidance of CLIP to learn consistent pixel embeddings with respect to visual and conceptual semantics. Our experiments on popular semantic segmentation benchmarks demonstrate consistent gains over the state-of-the-art unsupervised semantic segmentation and language-driven semantic segmentation methods, especially for unknown classes.", + "bbox": [ + 496, + 763, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "11214", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Radhakrishna Achanta, Appu Shaji, Kevin Smith, Aurelien Lucchi, Pascal Fua, and Sabine Süsstrunk. SLIC superpixels compared to state-of-the-art superpixel methods. IEEE TPAMI, 34(11):2274-2282, 2012. 3", + "[2] Pablo Arbeláez, Michael Maire, Charless Fowlkes, and Jitendra Malik. Contour detection and hierarchical image segmentation. IEEE TPAMI, 33(5):898-916, 2011. 3, 6", + "[3] Maxime Bucher, Tuan-Hung VU, Matthieu Cord, and Patrick Pérez. Zero-shot semantic segmentation. In NeurIPS, pages 468-479, 2019. 2", + "[4] S. Caelles, K.-K. Maninis, J. Pont-Tuset, L. Leal-Taixe, D. Cremers, and L. Van Gool. One-shot video object segmentation. In CVPR, pages 5320-5329, 2017. 8", + "[5] Holger Caesar, Jasper Uijlings, and Vittorio Ferrari. COCO-Stuff: Thing and stuff classes in context. In CVPR, pages 1209-1218, 2018. 5", + "[6] Liang-Chieh Chen, George Papandreou, Iasonas Kokkinos, Kevin Murphy, and Alan L. Yuille. DeepLab: Semantic image segmentation with deep convolutional nets, atrous convolution, and fully connected CRFs. IEEE TPAMI, 40(4):834-848, 2018. 1", + "[7] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In ICML, pages 1597-1607, 2020. 1, 6", + "[8] Jang Hyun Cho, Utkarsh Mall, Kavita Bala, and Bharath Hariharan. PiCIE: Unsupervised semantic segmentation using invariance and equivariance in clustering. In CVPR, pages 16794-16804, 2021. 2", + "[9] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. ImageNet: A large-scale hierarchical image database. In CVPR, pages 248–255, 2009. 6", + "[10] Xiaoyi Dong, Yinglin Zheng, Jianmin Bao, Ting Zhang, Dongdong Chen, Hao Yang, Ming Zeng, Weiming Zhang, Lu Yuan, Dong Chen, Fang Wen, and Nenghai Yu. MaskCLIP: Masked self-distillation advances contrastive language-image pretraining. arXiv preprint arXiv:2208.12262, 2022.3", + "[11] Sepideh Esmaeilpour, Bing Liu, Eric Robertson, and Lei Shu. Zero-shot out-of-distribution detection based on the pre-trained model CLIP. In AAAI, 2022. 3", + "[12] Mark Everingham, Luc Van Gool, Christopher K. I. Williams, John Winn, and Andrew Zisserman. The pascal visual object classes (voc) challenge. *IJCV*, 88(2):303–338, 2010. 5", + "[13] Xiuye Gu, Tsung-Yi Lin, Weicheng Kuo, and Yin Cui. Open-vocabulary object detection via vision and language knowledge distillation. In ICLR, 2022. 3, 4, 5, 6", + "[14] Mark Hamilton, Zhoutong Zhang, Bharath Hariharan, Noah Snavely, and William T. Freeman. Unsupervised semantic segmentation by distilling feature correspondences. In ICLR, 2022. 1, 2", + "[15] Bharath Hariharan, Pablo Arbeláez, Lubomir Bourdev, Subhransu Maji, and Jitendra Malik. Semantic contours from inverse detectors. In ICCV, pages 991-998, 2011. 5" + ], + "bbox": [ + 78, + 114, + 468, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[16] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In CVPR, pages 9726-9735, 2020. 1", + "[17] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, pages 770-778, 2016. 6", + "[18] Wenbin He, William Surmeier, Arvind Kumar Shekar, Liang Gou, and Liu Ren. Self-supervised semantic segmentation grounded in visual concepts. In IJCAI, pages 949-955, 2022. 1, 2, 3, 6, 7, 8", + "[19] Jyh-Jing Hwang, Stella X. Yu, Jianbo Shi, Maxwell D. Collins, Tien-Ju Yang, Xiao Zhang, and Liang-Chieh Chen. SegSort: Segmentation by discriminative sorting of segments. In ICCV, pages 7333-7343, 2019. 1, 2, 5, 6, 7", + "[20] Phillip Isola, Daniel Zoran, Dilip Krishnan, and Edward H. Adelson. Crisp boundary detection using pointwise mutual information. In ECCV, pages 799-814, 2014. 6", + "[21] Ajay Jain, Matthew Tancik, and Pieter Abbeel. Putting NeRF on a diet: Semantically consistent few-shot view synthesis. In ICCV, pages 5865-5874, 2021. 3", + "[22] Xu Ji, Andrea Vedaldi, and Joao Henriques. Invariant information clustering for unsupervised image classification and segmentation. In ICCV, pages 9864-9873, 2019. 2, 7", + "[23] Tsung-Wei Ke, Jyh-Jing Hwang, Yunhui Guo, Xudong Wang, and Stella X. Yu. Unsupervised hierarchical semantic segmentation with multiview cosegmentation and clustering transformers. In CVPR, pages 2561–2571, 2022. 1, 2, 3, 6, 7", + "[24] Shu Kong and Charless Fowlkes. Multigrid predictive filter flow for unsupervised learning on videos. arXiv preprint arXiv:1904.01693, 2019. 8", + "[25] Boyi Li, Kilian Q Weinberger, Serge Belongie, Vladlen Koltun, and René Ranftl. Language-driven semantic segmentation. In ICLR, 2022. 1, 2, 3, 5", + "[26] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dolkar, and C. Lawrence Zitnick. Microsoft COCO: Common objects in context. In ECCV, pages 740-755, 2014. 5", + "[27] Jonathan Long, Evan Shelhamer, and Trevor Darrell. Fully convolutional networks for semantic segmentation. In CVPR, pages 3431-3440, 2015. 1, 2, 5", + "[28] Luke Melas-Kyriazi, Christian Rupprecht, Iro Laina, and Andrea Vedaldi. Deep spectral methods: A surprisingly strong baseline for unsupervised semantic segmentation and localization. In CVPR, pages 8354-8365, 2022. 2", + "[29] Roozbeh Mottaghi, Xianjie Chen, Xiaobai Liu, Nam-Gyu Cho, Seong-Whan Lee, Sanja Fidler, Raquel Urtasun, and Alan Yuille. The role of context for object detection and semantic segmentation in the wild. In CVPR, pages 891–898, 2014. 5", + "[30] Yassine Ouali, Céline Hudelot, and Myriam Tami. Autoregressive unsupervised image segmentation. In ECCV, pages 142-158, 2020. 2", + "[31] Dim P. Papadopoulos, Ethan Weber, and Antonio Torralba. Scaling up instance annotation via label propagation. In ICCV, pages 15344-15353, 2021. 1" + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "11215", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[32] Federico Perazzi, Anna Khoreva, Rodrigo Benenson, Bernt Schiele, and Alexander Sorkine-Hornung. Learning video object segmentation from static images. In CVPR, pages 3491-3500, 2017. 8", + "[33] F. Peruzzi, J. Pont-Tuset, B. McWilliams, L. Van Gool, M. Gross, and A. Sorkine-Hornung. A benchmark dataset and evaluation methodology for video object segmentation. In CVPR, pages 724-732, 2016. 6, 8", + "[34] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In ICML, pages 8748-8763, 2021. 1, 2, 4, 5, 6", + "[35] Yongming Rao, Wenliang Zhao, Guangyi Chen, Yansong Tang, Zheng Zhu, Guan Huang, Jie Zhou, and Jiwen Lu. DenseCLIP: Language-guided dense prediction with context-aware prompting. In CVPR, pages 18061-18070, 2022. 3", + "[36] Ramprasaath R. Selvaraju, Karan Desai, Justin Johnson, and Nikhil Naik. CASTing your model: Learning to localize improves self-supervised representations. In CVPR, pages 11053-11062, 2021. 2", + "[37] Gyungin Shin, Weidi Wie, and Samuel Albanie. ReCo: Retrieve and co-segment for zero-shot transfer. In NeurIPS, 2022. 1, 2, 3, 6", + "[38] Kihyuk Sohn, David Berthelot, Chun-Liang Li, Zizhao Zhang, Nicholas Carlini, Ekin D. Cubuk, Alex Kurakin, Han Zhang, and Colin Raffel. FixMatch: Simplifying semi-supervised learning with consistency and confidence. In NeurIPS, pages 596-608, 2020. 4", + "[39] Wouter Van Gansbeke, Simon Vandenhende, Stamatios Georgoulis, and Luc V Gool. Revisiting contrastive methods for unsupervised learning of visual representations. In NeurIPS, pages 16238-16250, 2021. 2", + "[40] Wouter Van Gansbeke, Simon Vandenhende, Stamatios Georgoulis, and Luc Van Gool. Unsupervised semantic segmentation by contrasting object mask proposals. In ICCV, pages 10032-10042, 2021. 1, 2, 5, 7, 8", + "[41] Yael Vinker, Ehsan Pajouheshgar, Jessica Y. Bo, Roman Christian Bachmann, Amit Haim Bermano, Daniel Cohen-Or, Amir Zamir, and Ariel Shamir. CLIPasso: Semantically-aware object sketching. ACM TOG, 41(4):86:1-86:11, 2022. 3", + "[42] Carl Vondrick, Abhinav Shrivastava, Alireza Fathi, Sergio Guadarrama, and Kevin Murphy. Tracking emerges by colorizing videos. In ECCV, pages 402-419, 2018. 8", + "[43] Xiaolong Wang, Allan Jabri, and Alexei A. Efros. Learning correspondence from the cycle-consistency of time. In CVPR, pages 2561-2571, 2019. 8", + "[44] Zhaoqing Wang, Yu Lu, Qiang Li, Xunqiang Tao, Yandong Guo, Mingming Gong, and Tongliang Liu. CRIS: Clip-driven referring image segmentation. In CVPR, pages 11676-11685, 2022. 3", + "[45] Saining Xie and Zhuowen Tu. Holistically-nested edge detection. In ICCV, pages 1395-1403, 2015. 6", + "[46] Jiarui Xu, Shalini De Mello, Sifei Liu, Wonmin Byeon, Thomas Breuel, Jan Kautz, and Xiaolong Wang. GroupViT:" + ], + "bbox": [ + 78, + 92, + 468, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Semantic segmentation emerges from text supervision. In CVPR, pages 18134-18144, 2022. 2, 3, 6", + "[47] Xiao Zhang and Michael Maire. Self-supervised visual representation learning from hierarchical grouping. In NeurIPS, pages 16579-16590, 2020. 2, 6, 7, 8", + "[48] Hengshuang Zhao, Jianping Shi, Xiaojuan Qi, Xiaogang Wang, and Jiaya Jia. Pyramid scene parsing network. In CVPR, pages 6230-6239, 2017. 1, 5, 6, 7", + "[49] Chong Zhou, Chen Change Loy, and Bo Dai. Extract free dense labels from CLIP. In ECCV, 2022. 1, 2, 3, 4, 5, 6, 7, 8" + ], + "bbox": [ + 501, + 92, + 890, + 233 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "11216", + "bbox": [ + 480, + 945, + 519, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2023/CLIP-S4_ Language-Guided Self-Supervised Semantic Segmentation/c24eb4b4-f000-47c7-8244-3296ba8e2cac_model.json b/2023/CLIP-S4_ Language-Guided Self-Supervised Semantic Segmentation/c24eb4b4-f000-47c7-8244-3296ba8e2cac_model.json new file mode 100644 index 0000000000000000000000000000000000000000..96c8932576d38161b78f921197d8b1b4fd8e4f18 --- /dev/null +++ b/2023/CLIP-S4_ Language-Guided Self-Supervised Semantic Segmentation/c24eb4b4-f000-47c7-8244-3296ba8e2cac_model.json @@ -0,0 +1,2134 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.808, + 0.045 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.14, + 0.131, + 0.83, + 0.152 + ], + "angle": 0, + "content": "CLIP-S\\(^4\\): Language-Guided Self-Supervised Semantic Segmentation" + }, + { + "type": "text", + "bbox": [ + 0.175, + 0.181, + 0.267, + 0.197 + ], + "angle": 0, + "content": "Wenbin He" + }, + { + "type": "text", + "bbox": [ + 0.337, + 0.182, + 0.5, + 0.198 + ], + "angle": 0, + "content": "Suphanut Jamonnak" + }, + { + "type": "text", + "bbox": [ + 0.569, + 0.182, + 0.657, + 0.198 + ], + "angle": 0, + "content": "Liang Gou" + }, + { + "type": "text", + "bbox": [ + 0.728, + 0.182, + 0.795, + 0.197 + ], + "angle": 0, + "content": "Liu Ren" + }, + { + "type": "text", + "bbox": [ + 0.16, + 0.2, + 0.808, + 0.216 + ], + "angle": 0, + "content": "Bosch Research North America & Bosch Center for Artificial Intelligence (BCAI)" + }, + { + "type": "text", + "bbox": [ + 0.2, + 0.22, + 0.765, + 0.234 + ], + "angle": 0, + "content": "{wenbin.he2, suphanut.jamonnak, liang.gou, liu. ren}@us.bosch.com" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.269, + 0.313, + 0.284 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.302, + 0.474, + 0.695 + ], + "angle": 0, + "content": "Existing semantic segmentation approaches are often limited by costly pixel-wise annotations and predefined classes. In this work, we present CLIP-\\(S^4\\) that leverages self-supervised pixel representation learning and vision-language models to enable various semantic segmentation tasks (e.g., unsupervised, transfer learning, language-driven segmentation) without any human annotations and unknown class information. We first learn pixel embeddings with pixel-segment contrastive learning from different augmented views of images. To further improve the pixel embeddings and enable language-driven semantic segmentation, we design two types of consistency guided by vision-language models: 1) embedding consistency, aligning our pixel embeddings to the joint feature space of a pre-trained vision-language model, CLIP [34]; and 2) semantic consistency, forcing our model to make the same predictions as CLIP over a set of carefully designed target classes with both known and unknown prototypes. Thus, CLIP-\\(S^4\\) enables a new task of class-free semantic segmentation where no unknown class information is needed during training. As a result, our approach shows consistent and substantial performance improvement over four popular benchmarks compared with the state-of-the-art unsupervised and language-driven semantic segmentation methods. More importantly, our method outperforms these methods on unknown class recognition by a large margin." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.724, + 0.208, + 0.739 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.47, + 0.901 + ], + "angle": 0, + "content": "Semantic segmentation aims to partition an input image into semantically meaningful regions and assign each region a semantic class label. Recent advances in semantic segmentation [6, 27, 48] heavily rely on pixel-wise human annotations, which have two limitations. First, acquiring pixel-wise annotations is extremely labor intensive and costly, which can take up to 1.5 hours to label one image [31]. Second, human annotations are often limited to a set of predefined semantic classes, with which the learned models lack the ability to recognize unknown classes [25]." + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.269, + 0.895, + 0.413 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.509, + 0.413, + 0.884, + 0.422 + ], + "angle": 0, + "content": "(b) Semantic segmentation generated by different CLIP-based unsupervised approaches" + }, + { + "type": "image", + "bbox": [ + 0.504, + 0.423, + 0.89, + 0.502 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.517, + 0.502, + 0.875, + 0.511 + ], + "angle": 0, + "content": "(c) CLIP-\\(S^4\\) aligns the pixel embeddings and their semantics with CLIP feature space" + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.518, + 0.892, + 0.629 + ], + "angle": 0, + "content": "Figure 1. (a) Pixel embeddings from different CLIP-based unsupervised methods: Our method, \\(\\mathrm{CLIP - S^4}\\), generates sharper and more coherent pixel embeddings than MaskCLIP [49] and MaskCLIP+’s [49]; (b) Language-driven semantic segmentation by different methods: \\(\\mathrm{CLIP - S^4}\\) can recognize challenging unknown classes (e.g., moon); (c) The key idea behind \\(\\mathrm{CLIP - S^4}\\): aligning the pixel embeddings and their semantics with CLIP feature space." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.64, + 0.893, + 0.867 + ], + "angle": 0, + "content": "Various approaches have been proposed to tackle these limitations, among which we are inspired by two lines of recent research in particular. First, for unsupervised semantic segmentation (i.e., without human annotations), self-supervised pixel representation learning approaches [14,18, 19, 23, 40] have shown promising results on popular unsupervised benchmarks. The main idea is to extend self-supervised contrastive learning [7, 16] from images to pixels by attracting each pixel's embedding to its positive pairs and repelling it from negative pairs. The prior of pairs can be contours [18, 19], hierarchical groups [23], salience maps [40], and pre-trained models [14]. Although these approaches can group pixels into semantically meaningful clusters, human annotations are still needed to assign class labels to the clusters for semantic segmentation [37]." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.871, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Second, for unknown classes in semantic segmentation, large-scale vision-language models such as CLIP [34]" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "11207" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.349 + ], + "angle": 0, + "content": "have shown great potential. This line of research, called language-driven semantic segmentation, aims to segment images with arbitrary classes defined by texts during testing time [25, 37, 46, 49]. Among these methods, most still need training time annotations, such as pixel annotations [25] and captions [46]. Only a few recent work, MaskCLIP & MaskCLIP+ [49] attempts to address this without using additional supervision: MaskCLIP directly extracts pixel embeddings correlated with texts from CLIP, but these pixel embeddings are coarse and noisy (Fig. 1a). To address this issue, MaskCLIP+ [49] trains a segmentation model on the pseudo-labels generated by MaskCLIP for a set of predefined classes. However, the pixel embeddings of MaskCLIP+ are distorted by the predefined classes (Fig. 1a), which limits its ability to recognize unknowns (Fig. 1b). Also, it needs unknown class information during training, which hinders its real-world applications." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.35, + 0.473, + 0.456 + ], + "angle": 0, + "content": "We propose a language-guided self-supervised semantic segmentation approach, CLIP-S\\(^4\\), which takes advantage of the strengths from both lines of research and addresses their limitations accordingly. The key idea is to learn consistent pixel embeddings with respect to visual and conceptual semantics using self-supervised learning and the guidance of a vision-language model, CLIP." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.457, + 0.473, + 0.744 + ], + "angle": 0, + "content": "Specifically, we first train pixel embeddings with pixel-segment contrastive learning from different augmented image views [18, 19, 23] such that images can be partitioned into visually meaningful regions. To further improve pixel embedding quality and enable language-driven semantic segmentation, we introduce vision-language model guided consistency to regularize our model (Fig. 1c). The consistency is enforced from two aspects: embedding consistency and semantic consistency. First, embedding consistency aims to align the pixel embeddings generated by our model with the joint feature space of texts and images of CLIP by minimizing the distance between the pixel embeddings generated by our model and CLIP. Second, semantic consistency forces our model to make the same prediction as CLIP over a set of carefully designed target classes with both known and unknown prototypes. Note that unlike the previous methods [25,49] that use a predefined set of known classes, CLIP-\\(\\mathbf{S}^4\\) also learns the representation of unknown classes from images during training." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.745, + 0.472, + 0.837 + ], + "angle": 0, + "content": "In the end, CLIP-\\(S^4\\) also enables a new task, namely class-free semantic segmentation, as shown in Tab. 1. This new task does not need any human annotations and even assumes NO class names are given during training. This is a more challenging task than the recent work [49] that requires class names of both known and unknown." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.837, + 0.47, + 0.852 + ], + "angle": 0, + "content": "In summary, the contributions of this paper are threefold:" + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.856, + 0.47, + 0.901 + ], + "angle": 0, + "content": "- We propose a self-supervised semantic segmentation approach that combines pixel-segment contrastive learning with the guidance of pre-trained vision lan" + }, + { + "type": "table", + "bbox": [ + 0.501, + 0.089, + 0.895, + 0.243 + ], + "angle": 0, + "content": "
KnownUnknownAdd. Info.
Annot.CIs NameAnnot.CIs Name
Un/Self-supervised ([19] etc.)××××Fine-Tuning
Supervised ([27] etc.)✓✓N/AN/AN/A
Zero-shot ([3] etc.)✓✓×✓Word2Vec, etc.
Language-MaskCLIP+[49]×✓CLIP
DrivenCLIP-S4×✓×CLIP
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.248, + 0.895, + 0.305 + ], + "angle": 0, + "content": "Table 1. Comparison of information required for training over different tasks. CLIP-\\(\\mathbf{S}^4\\) enables a new task called class-free semantic segmentation. Compared with MaskCLIP+ [49], the new task assumes unknown class names are NOT given during training." + }, + { + "type": "text", + "bbox": [ + 0.531, + 0.311, + 0.892, + 0.357 + ], + "angle": 0, + "content": "guage models. Our method can generate high-quality pixel embeddings without any human annotations and be applied to a variety of semantic segmentation tasks." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.358, + 0.892, + 0.478 + ], + "angle": 0, + "content": "- We open up new research potentials for language-driven semantic segmentation without any human annotations by introducing and addressing a new task of class-free semantic segmentation (Tab. 1). Unlike previous work that assumes all the class names are known during training, our method can discover unknown classes from unlabelled image data without even knowing unknown class names." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.481, + 0.892, + 0.571 + ], + "angle": 0, + "content": "- Consistent and substantial gains are observed with our approach over the state-of-the-art unsupervised and language-driven semantic segmentation methods on four popular datasets. More importantly, our method significantly outperforms the state-of-the-art on the segmentation of unknown classes." + }, + { + "type": "list", + "bbox": [ + 0.518, + 0.358, + 0.892, + 0.571 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.588, + 0.642, + 0.604 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.614, + 0.892, + 0.87 + ], + "angle": 0, + "content": "Unsupervised Semantic Segmentation. There are two groups of recent unsupervised semantic segmentation methods. One group of methods learns to generate consistent pixel representations or predictions between different augmentation of images with the guidance of mutual information [22, 30], clusters [8], contours [18, 19], hierarchical groups [23, 47], and saliency masks [40]. The other group of methods extracts dense features from pre-trained models based on saliency maps [36], augmentations [39], spectral decomposition [28], and feature correspondences [14]. While these methods can generate pixel embeddings with semantically meaningful clusters, annotations are needed to assign class labels to the clusters (e.g., \\( k \\)-nearest neighbor search [19] and Hungarian algorithm [40]). Our work combines pixel-segment self-supervision with pre-trained vision-language models to enable semantic segmentation without any human annotations." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.892, + 0.903 + ], + "angle": 0, + "content": "Language-Driven Semantic Segmentation. Recently, vision-language models (e.g., CLIP [34]) trained on large-" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "11208" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.081, + 0.088, + 0.895, + 0.294 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.298, + 0.895, + 0.397 + ], + "angle": 0, + "content": "Figure 2. Framework of CLIP-S4. (a) It starts with an encoder that maps images into pixel embeddings for semantic segmentation. Then it follows with two components: 1) self-supervised contrastive learning and 2) vision-language model guided consistency. Specifically, shown in (b), the self-supervised contrastive learning forces pixel embeddings to be consistent within visually coherent regions and among different augmented views of the same image. For vision-language model guided consistency, this framework introduces (c) embedding consistency that aligns the pixel embeddings generated by our model with CLIP embeddings, and (d) semantic consistency that forces our model to make the same predictions as CLIP for a set of target classes with both known and unknown prototypes. For known classes, the prototypes are pre-computed and fixed during training. For unknown classes, the prototypes are learned during training via clustering (e)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.402, + 0.471, + 0.568 + ], + "angle": 0, + "content": "scale image-text datasets have shown great potential on various downstream tasks such as image synthesis [21, 41], out-of-distribution detection [11], and object detection [13]. To extend vision-language models for semantic segmentation, one active research, language-driven semantic segmentation, aims to segment images with arbitrary unknown classes defined by texts during testing time [25, 37, 46, 49]. Some methods [25, 44] use pixel-wise annotations to train language-guided semantic segmentation models. Other methods [10, 46] perform large-scale pre-training on image-text pairs specifically for semantic segmentation." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.57, + 0.47, + 0.796 + ], + "angle": 0, + "content": "By contrast, we directly use vision-language models that are pre-trained for classification tasks. Along this line of research, a few approaches have been proposed [35, 37, 49]. The most relevant approach to our method is MaskCLIP [49], which extends the embeddings generated by pre-trained vision-language models from image to pixel level, but these embeddings are often coarse and noisy. To address this issue, MaskCLIP + [49] fine-tunes the pixel embeddings by the pseudo-labels of a specific set of classes on top of MaskCLIP. However, it needs unknown class names during training, which may not be possible in real-world cases. Compared with [49], our method can recognize unknown classes without knowing any unknown class information during training time, and also learns fine-grained and sharper pixel embeddings with self-supervision." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.814, + 0.169, + 0.829 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Our method (Fig. 2) segments images by learning a pixel embedding function with self-supervised contrastive learning and the guidance of a pre-trained vision-language model, CLIP. We use self-supervised contrastive learning" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.402, + 0.895, + 0.614 + ], + "angle": 0, + "content": "to force the consistency of pixel embeddings within visually coherent regions (e.g., superpixels) and among different augmented views of the same image (Sec. 3.1). We also introduce two vision-language model guided consistency (i.e., embedding consistency and semantic consistency) to further regularize the model (Sec. 3.2). The two components are complementary to each other. On the one hand, contrastive learning mitigates the noise introduced by CLIP. On the other hand, with the knowledge extracted from CLIP, the quality of the pixel embeddings can be improved. More importantly, this approach enables us to perform language-driven semantic segmentation with our carefully designed target class prototypes of both knowns and unknowns. In the following, we discuss the two components in detail." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.621, + 0.816, + 0.637 + ], + "angle": 0, + "content": "3.1. Pixel-Segment Contrastive Learning" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.644, + 0.892, + 0.902 + ], + "angle": 0, + "content": "We train a pixel embedding function to generate consistent pixel embeddings within visually coherent regions through pixel-segment contrastive learning [18,23]. Specifically, the embedding function transforms each pixel \\( p \\) of an image to a unit-length embedding vector \\( \\mathbf{z}_p \\) of dimension \\( d \\) via a deep neural network. The image is then partitioned into \\( |S| \\) segments by clustering the pixel embeddings. The embedding \\( \\mathbf{v}_s \\) of each segment \\( s \\) is calculated as the average of the pixel embeddings \\( \\mathbf{v}_s = \\sum_{p\\in s}\\mathbf{z}_p / |s| \\), which is also normalized into a unit-length vector \\( \\mathbf{v}_s = \\mathbf{v}_s / \\| \\mathbf{v}_s\\| \\). For each pixel \\( p \\), the segments are grouped into two sets including a positive set \\( S^+ \\) and a negative set \\( S^- \\). The positive set \\( S^+ \\) of a pixel contains segments within the same visually coherent region of the pixel. Following the prior work [18, 23], the visually coherent region can be derived from super-pixels [1] or contours [2]. We also use data augmentation (e.g., random cropping and color jitter" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "11209" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.226 + ], + "angle": 0, + "content": "ing) to generate consistent pixel embeddings between different augmented views of the same image. Hence, segments within the same region of the pixel in any augmented views are considered as the positive set \\( S^{+} \\). Other segments in the image and segments from other images in the same batch are included in the negative set \\( S^{-} \\). The pixel embedding \\( \\mathbf{z}_p \\) is then attracted to the segments in positive set \\( S^{+} \\) and repelled from the segments in negative set \\( S^{-} \\) with contrastive loss:" + }, + { + "type": "equation", + "bbox": [ + 0.115, + 0.238, + 0.47, + 0.274 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {t} (p) = - \\log \\frac {\\sum_ {s \\in \\mathcal {S} ^ {+}} e x p \\left(\\operatorname {s i m} \\left(\\mathbf {z} _ {p} , \\mathbf {v} _ {s}\\right) \\kappa\\right)}{\\sum_ {s \\in \\mathcal {S} ^ {+} \\cup \\mathcal {S} ^ {-}} e x p \\left(\\operatorname {s i m} \\left(\\mathbf {z} _ {p} , \\mathbf {v} _ {s}\\right) \\kappa\\right)}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.286, + 0.47, + 0.332 + ], + "angle": 0, + "content": "where \\(\\kappa\\) is the concentration constant and \\(sim(\\mathbf{z}_p, \\mathbf{v}_s)\\) is the cosine similarity between the pixel embedding \\(\\mathbf{z}_p\\) and the segment embedding \\(\\mathbf{v}_s\\)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.345, + 0.455, + 0.361 + ], + "angle": 0, + "content": "3.2. Vision-Language Model Guided Consistency" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.369, + 0.47, + 0.505 + ], + "angle": 0, + "content": "To enable language-driven semantic segmentation and improve the quality of pixel embeddings, we use a pretrained vision-language model such as CLIP [34] to guide the training of the pixel embedding function. The key idea is to align the output space of our pixel embedding function consistent with the feature space of CLIP (Fig. 1c). Specifically, two types of consistency are considered during training including embedding consistency and semantic consistency, which are detailed as follows." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.507, + 0.47, + 0.581 + ], + "angle": 0, + "content": "**Embedding Consistency.** Our goal is to align the pixel embeddings generated from our self-supervised method (the green contour in Fig. 1c) with CLIP's pixel embeddings (the orange contour in Fig. 1c). This is done by minimizing the distance between the two pixel embedding spaces." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.583, + 0.47, + 0.704 + ], + "angle": 0, + "content": "We first obtain the pixel embeddings of an input image from CLIP by modifying the attention-based pooling layer of the CLIP image encoder following [49]. Specifically, we 1) remove the query and key projection layers and 2) reformulate the value projection layer and the last linear layer as two consecutive fully connected layers. In the following, we use \\(clip - i(\\cdot)\\) as the modified CLIP image encoder and \\(clip - t(\\cdot)\\) as CLIP text encoder." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.705, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Then we obtain the pixel embeddings of CLIP for different augmented views of the image. Note that we use the original image to generate the CLIP pixel embeddings and perform augmentation afterwards to make sure that the CLIP pixel embeddings are consistent among different augmented views. In the end, we minimize the distance of embeddings between segments instead of pixels from our self-supervised and CLIP embedding spaces. This is because the pixel embeddings of CLIP are noisy (Fig. 2), which can be mitigated by aggregating over segments. Hence, we use the pixel embeddings generated by our model to derive segments (clusters) and then apply them to the CLIP's pixel embeddings. In the end, for each segment \\( s \\), the embedding" + }, + { + "type": "image", + "bbox": [ + 0.499, + 0.088, + 0.895, + 0.302 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.307, + 0.895, + 0.447 + ], + "angle": 0, + "content": "Figure 3. Computation of target class prototypes with both knowns and unknowns, \\( C = \\{C_k, C_u\\} \\). For a set of known classes (e.g., bird, cat), we first obtain their CLIP text embeddings, \\( T \\), via a set of prompt templates [13,49] (shown in \\( \\widehat{\\mathbf{a}} \\)); then, we calculate the normalized (via softmax) similarity between text embeddings, \\( T \\), and all segments' CLIP embeddings, \\( I \\), from training images, and average the top-\\( m \\) similar segments' CLIP embedding as the embedding prototype for this class, as shown in \\( \\widehat{\\mathbf{b}} \\); For each unknown class, we randomly select the CLIP embedding of a segment as the initial prototype (shown in \\( \\widehat{\\mathbf{c}} \\))." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.456, + 0.691, + 0.47 + ], + "angle": 0, + "content": "consistent loss is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.608, + 0.487, + 0.892, + 0.503 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {e} (s) = 1 - \\operatorname {s i m} \\left(\\mathbf {v} _ {s}, \\mathbf {i} _ {s}\\right), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.518, + 0.892, + 0.58 + ], + "angle": 0, + "content": "where \\(\\mathbf{v}_s\\) and \\(\\mathbf{i}_s\\) are the segment embeddings derived from our embedding function and CLIP, respectively. Here, \\(\\mathbf{i}_s\\) is the average of the CLIP pixel embedding from segment \\(s\\), namely, \\(\\mathbf{i}_s = \\sum_{p \\in s} \\text{clip-}i(s) / |s|\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.581, + 0.892, + 0.686 + ], + "angle": 0, + "content": "Semantic Consistency In addition to embedding consistency, we introduce semantic consistency by forcing our model to make the same predictions of semantic classes as CLIP. The rationale is that we can generate better pixel embeddings if they can form distinctive clusters corresponding to different semantic classes, as the goal of semantic segmentation is to perform pixel-wise classification." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.688, + 0.893, + 0.87 + ], + "angle": 0, + "content": "Semantic consistency is achieved via a similar idea of pseudo-labeling [38]. Again, we force the semantic consistency at the segment level (not the pixel level) to reduce the noise in pseudo-labels. Specifically, for each segment \\( s \\), we first use CLIP to generate its pseudo-label \\( y_{s} \\) over a set of target classes, which include both knowns and unknowns (we will introduce how to design these target classes later). The pseudo-label is generated based on the highest similarity between the segment embedding \\( \\mathbf{i}_s \\) with a set of prototypes, \\( C = \\{\\mathbf{c}_l\\}_{0}^{L - 1} \\), of the target classes in the pixel embedding space of CLIP, namely, \\( y_{s} = \\operatorname{argmax}_{l\\in L}(sim(\\mathbf{i}_{s},\\mathbf{c}_{l})) \\)." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.871, + 0.892, + 0.902 + ], + "angle": 0, + "content": "Then we define the semantic consistent loss as the cross entropy between our model's prediction \\(\\varphi (\\mathbf{v}_s)\\) over the tar" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "11210" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.314, + 0.107 + ], + "angle": 0, + "content": "get classes and the pseudo-label \\( y_{s} \\):" + }, + { + "type": "equation", + "bbox": [ + 0.193, + 0.118, + 0.47, + 0.136 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {s} (s) = \\mathbf {H} \\left(y _ {s}, \\varphi (\\mathbf {v} _ {s})\\right), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.146, + 0.348, + 0.162 + ], + "angle": 0, + "content": "where \\(\\varphi (\\mathbf{v}_s) = \\mathrm{softmax}(sim(\\mathbf{v}_s,C))\\)" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.162, + 0.47, + 0.403 + ], + "angle": 0, + "content": "Target Class Prototypes The design of target classes and associated class prototypes, \\( C = \\{\\mathbf{c}_l\\}_{0}^{L - 1} \\), is crucial to achieve the semantic consistency. Here, a class prototype, \\( \\mathbf{c}_l \\), is an embedding vector that can represent a class in an embedding space. For example, it can be the mean vector of embeddings of all segments of a class \"car\". Currently, most existing methods [25, 49] assume that the target classes are already predefined, which is not feasible in real-world use cases without any human annotations. Thus, those methods cannot handle unknown classes hidden in the data. To address this issue, we introduce two sets of class prototypes of known, \\( C_k = \\{\\mathbf{c}_0,\\dots ,\\mathbf{c}_{k - 1}\\} \\), and unknown classes, \\( C_u = \\{\\mathbf{c}_k,\\dots ,\\mathbf{c}_{k + u}\\} \\), where the known classes are predefined by leveraging CLIP and the unknown classes are learned from image data during training. Thus, we have \\( C = \\{\\mathbf{c}_l\\}_{0}^{L - 1} = \\{\\mathbf{c}_0,\\dots ,\\mathbf{c}_{k - 1},\\mathbf{c}_k,\\dots ,\\mathbf{c}_{k + u}\\} \\) \\( L = k + u \\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.403, + 0.47, + 0.553 + ], + "angle": 0, + "content": "For known classes, a natural choice is to use the text embeddings generated by CLIP as their class prototype embeddings [25, 49]. However, even though the text embeddings are trained to align with image/pixel embeddings [34], there is still a huge gap between the text and image/pixel embeddings in the joint space of CLIP (Fig. 1c). Therefore, it is challenging to learn meaningful unknown classes from image features when using text embeddings as class prototypes. Hence, in this work, we use the prototype of CLIP pixel embeddings to represent each known class." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.554, + 0.47, + 0.795 + ], + "angle": 0, + "content": "To this end, for a set of known classes (e.g., bird, cat), \\( K = \\{0,\\dots ,k - 1\\} \\) we first obtain their CLIP text embeddings, \\( T = \\{\\mathbf{t}_k\\} = \\{\\text{clip - } t(k)\\} \\), via a set of prompt templates following [13, 49], as shown in Fig. 3a. We also get a set of CLIP segment embeddings, \\( I = \\{\\mathbf{i}_{\\hat{s}}\\} \\), for all training images by a) feeding training images into the modified image encoder of CLIP to get pixel embeddings; b) clustering the pixel embeddings as segments, \\( \\hat{S} \\); c) averaging the pixel embeddings in each segment, \\( \\hat{s} \\). Hence, we have embeddings for each segment: \\( \\mathbf{i}_{\\hat{s}} = \\sum_{p\\in \\hat{s}}\\text{clip - } i(p) / |\\hat{s}| \\). Then, we calculate the similarity between text embeddings of known classes, \\( T \\), and all CLIP segment embeddings \\( I \\), and normalize the similarities over all classes by softmax. Finally, we average the top- \\( m \\) similar segments' embedding as the embedding prototype for each class, \\( C_k = \\{\\mathbf{c}_k\\} = avg_m(top-m_{\\hat{s}}(\\mathbf{softmax}_k(sim(I,T)))) \\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.796, + 0.47, + 0.902 + ], + "angle": 0, + "content": "The prototype embeddings of the unknown classes, \\( C_u \\), are randomly initialized by sampling the CLIP embeddings of all segments, namely \\( C_u = \\text{random}(clip - i(\\hat{\\mathcal{S}}), u) \\), with a size of the unknown class of \\( u \\) (Fig. 3c). During training, the embedding \\( \\mathbf{c}_u \\) of each unknown class prototype is updated by minimizing its distance to all segments that are classified as this unknown class (similar to updating the cen" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.698, + 0.107 + ], + "angle": 0, + "content": "troids in \\(k\\) -means clustering):" + }, + { + "type": "equation", + "bbox": [ + 0.555, + 0.118, + 0.892, + 0.151 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {u} = \\sum_ {s \\in S _ {u}} (1 - \\operatorname {s i m} (\\mathbf {c} _ {u}, \\operatorname {c l i p} - i (s))) / | S _ {u} |, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.163, + 0.892, + 0.207 + ], + "angle": 0, + "content": "where \\( S_{u} \\) are the segments classified as the unknown classes. In this way, our model can also learn the pixel representation of unknown classes." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.218, + 0.715, + 0.234 + ], + "angle": 0, + "content": "3.3. Training and Inference" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.242, + 0.892, + 0.286 + ], + "angle": 0, + "content": "In summary, we train the pixel embedding function by combining the pixel-segment contrastive loss, embedding consistent loss, and semantic consistent loss:" + }, + { + "type": "equation", + "bbox": [ + 0.628, + 0.3, + 0.891, + 0.316 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\mathcal {L} _ {t} + \\mathcal {L} _ {e} + \\mathcal {L} _ {s}. \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.328, + 0.891, + 0.357 + ], + "angle": 0, + "content": "During training, we also update the embeddings for the unknown classes with \\(\\mathcal{L}_u\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.359, + 0.892, + 0.54 + ], + "angle": 0, + "content": "For inference, we use the trained model to generate pixel embeddings for each input image and use the pixel embeddings for different downstream tasks, including language-driven and unsupervised semantic segmentation. For language-driven semantic segmentation, we first obtain the text embeddings of arbitrary inference classes by feeding the prompt-engineered texts into the text encoder of CLIP. Then we assign each pixel with the class label whose text embedding is the closest to \\(\\mathrm{CLIP - S^4}\\) pixel embedding. For unsupervised semantic segmentation, we follow previous work [19, 40] that uses \\(k\\) nearest neighbor search or linear classifier to perform semantic segmentation." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.555, + 0.634, + 0.571 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.58, + 0.892, + 0.685 + ], + "angle": 0, + "content": "We evaluate our model on three tasks: 1) language-driven semantic segmentation for both known and unknown classes; 2) unsupervised semantic segmentation with \\( k \\)-means clustering/linear classification; 3) transfer learning of generated pixel embeddings for instance mask tracking. We also conduct ablation studies to understand the components of our model." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.696, + 0.602, + 0.71 + ], + "angle": 0, + "content": "4.1. Datasets" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.719, + 0.892, + 0.795 + ], + "angle": 0, + "content": "Pascal VOC 2012 [12] contains 20 object classes and a background class. It has 1,464 and 1,449 images for training and validation, respectively. Following common practice [27, 48], we augment the training data with additional annotations [15], resulting in 10,582 training images." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.795, + 0.892, + 0.854 + ], + "angle": 0, + "content": "Pascal Context [29] extends Pascal VOC 2010 [12] with additional annotations on 4,998 training and 5,105 validation images. Following the prior work [49], we use the most common 59 classes for evaluation." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.892, + 0.901 + ], + "angle": 0, + "content": "COCO-Stuff [5] labels MS COCO [26] with 171 object/stuff classes. It contains 118,287 and 5,000 images for training and validation, respectively." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "11211" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.468, + 0.152 + ], + "angle": 0, + "content": "DAVIS 2017 [33] contains video sequences for instance mask tracking. Following the prior work [18, 47], we train pixel embeddings on Pascal VOC 2012 and evaluate the validation sequences without fine-tuning." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.153, + 0.468, + 0.243 + ], + "angle": 0, + "content": "It is worth mentioning that no ground truth labels of any datasets are used during training. Instead, we perform self-supervised learning on pseudo segments generated by contour detectors and owt-ucm [2]. Two contour detectors are used including HED [45] for Pascal VOC 2012 and Pascal Context and PMI [20] for COCO-Stuff." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.254, + 0.295, + 0.269 + ], + "angle": 0, + "content": "4.2. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.278, + 0.468, + 0.367 + ], + "angle": 0, + "content": "For self-supervised contrastive learning, images are augmented with the same set of data augmentations as SimCLR [7], including random resizing, cropping, flipping, color jittering, and Gaussian blurring. The concentration constant \\(\\kappa\\) is set to 10, and the number of segments is set to 36 for each augmented view." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.368, + 0.468, + 0.472 + ], + "angle": 0, + "content": "For vision-language guidance, we use pre-trained CLIP models [34] with modified image encoders following [49]. We use prompt-engineered texts with 85 prompt templates to generate text embeddings following [13, 49]. We use the average embedding of the top 32 segments of high probabilities as the prototype of each known class. We set the number of unknown classes to \\( u = 64 \\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.474, + 0.468, + 0.639 + ], + "angle": 0, + "content": "Following the prior work [18, 19, 23], we use PSP-Net [48] with a dilated ResNet-50 [17] backbone as the network architecture. The backbone is pre-trained on the ImageNet [9] dataset. We train our model on Pascal VOC 2012 and Pascal Context for \\(20\\mathrm{k}\\) iterations and on COCO-Stuff for \\(80\\mathrm{k}\\) iterations. We set the batch size to 8 with additional memory banks that cache the segment embeddings of the previous 2 batches. We set the initial learning rate to 0.001 and decay it with a polynomial learning rate policy. We use the CLIP model trained with ViT-B/16 backbone unless otherwise stated." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.651, + 0.433, + 0.667 + ], + "angle": 0, + "content": "4.3. Language-Driven Semantic Segmentation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.674, + 0.468, + 0.749 + ], + "angle": 0, + "content": "For language-driven semantic segmentation, no human annotations are used for either training or inference. At the inference time, each pixel is assigned an arbitrarily given class label whose CLIP text embedding is the closest to this pixel's CLIP-\\(\\mathbf{S}^4\\) embedding." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.468, + 0.901 + ], + "angle": 0, + "content": "We first compare the performance of our method with the state-of-the-art language-driven semantic segmentation approaches [37, 46, 49] on the Pascal Context and COCO-Stuff datasets. The performance is evaluated with the mean Intersection over Union (mIoU). For MaskCLIP and MaskCLIP+ [49], we obtain the results using the same hyper-parameter setting as our approach with CLIP models of two different backbones of ResNet50 and ViT-B/16. Meanwhile, GroupViT [46], ReCo [37], and \\(\\mathrm{ReCo + }\\) [37] use completely different training mechanisms compared" + }, + { + "type": "table", + "bbox": [ + 0.501, + 0.089, + 0.898, + 0.294 + ], + "angle": 0, + "content": "
MethodCLIP ModelPascal ContextCOCO-Stuff
mIoUmIoU
GroupViT [46]-22.4-
ReCo [37]ResNet50x16 + ViT-L/14@336px26.6-
ReCo+ [37]†-18.4
MaskCLIP [49]ResNet5018.610.6
ViT-B/1625.215.2
MaskCLIP+ [49]†ResNet5023.413.9
ViT-B/1632.220.7
CLIP-S4†ResNet5028.5 (+5.1)16.7 (+2.8)
ViT-B/1633.6 (+1.4)22.1 (+1.4)
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.3, + 0.892, + 0.37 + ], + "angle": 0, + "content": "Table 2. Language-guided semantic segmentation benchmarks (mIoU). CLIP-\\(S^4\\) consistently outperforms the state-of-the-art methods on both Pascal Context and COCO-Stuff datasets with CLIP models of different backbones. \\(\\dagger\\) indicates the models are fine-tuned on target datasets." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.383, + 0.892, + 0.503 + ], + "angle": 0, + "content": "with our method. GroupViT is trained on image-caption pairs, and \\(\\mathrm{ReCo} / \\mathrm{ReCo}+\\) combines image retrieval and cosegmentation. For comparison, we take the best results from [37, 46] for GroupViT, ReCo, and \\(\\mathrm{ReCo}+\\). Tabel 2 shows the benchmarking results of the aforementioned methods. Our method consistently outperforms the state-of-the-art on both datasets with CLIP models of different backbones." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.507, + 0.892, + 0.718 + ], + "angle": 0, + "content": "To evaluate models' performance for class-free semantic segmentation with both known and unknown classes, we split the 59 classes of Pascal Context into 4 folds, where each fold includes around 15 classes. For each experiment, classes from one fold are considered as unknown and excluded during training. The mIoUs of known and unknown classes, as well as their harmonic mean (hIoU) are reported in Tab. 3. The performance of our method is averaged across 5 runs with randomly initialized prototypes of unknown classes. Our method achieves significant gains over MaskCLIP+ on unknown classes, which are comparable to MaskCLIP. Also, our method outperforms both MaskCLIP and MaskCLIP+ on known classes, which leads to better overall performance." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Qualitatively, the visualization in Fig. 4a offers us some insights into why our approach can achieve better results: our model yields consistent embeddings aligned with the pre-trained CLIP model. Fig. 4a visualizes the projection of segment embeddings generated by different methods on Pascal Context and COCO-Stuff. We observe that segment embeddings generated by \\(\\mathrm{MaskCLIP + }\\) are distorted by the given text embeddings. Meanwhile, \\(\\mathrm{CLIP - S^4}\\) generates segment embeddings that are well aligned with the segment embeddings derived from the pre-trained CLIP model. Hence, segment embeddings generated by \\(\\mathrm{CLIP - S^4}\\) can better capture both known and unknown classes. Fig. 4b shows" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "11212" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.08, + 0.089, + 0.891, + 0.184 + ], + "angle": 0, + "content": "
Methodfold0fold1fold2fold3
mIoUumIoUkhIoUmIoUumIoUkhIoUmIoUumIoUkhIoUmIoUumIoUkhIoU
MaskCLIP29.723.726.323.725.724.623.925.724.723.425.824.5
MaskCLIP+3.628.56.33.029.25.44.829.28.24.529.97.8
CLIP-S432.0±0.829.4±0.330.6±0.522.3±0.932.8±0.426.5±0.622.4±0.532.1±0.526.4±0.428.6±0.831.5±0.230.0±0.5
vs. MaskCLIP++28.4+0.9+24.3+19.3+3.6+21.1+17.6+2.9+18.2+24.1+1.6+22.2
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.188, + 0.893, + 0.258 + ], + "angle": 0, + "content": "Table 3. Language-guided semantic segmentation benchmarks (mIoU) for unknown classes. The classes of Pascal Context are split into 4 folds with around 15 classes each fold. For each experiment, classes of one fold are considered as unknown. The performance of CLIP-\\(S^4\\) is averaged over 5 runs with randomly initialized unknown class embeddings. CLIP-\\(S^4\\) significantly outperforms \\(\\text{MaskCLIP}+\\) on unknown classes. Meanwhile, CLIP-\\(S^4\\) archives consistent gains on known classes over \\(\\text{MaskCLIP}\\) and \\(\\text{MaskCLIP}+\\), and hence leads to better overall performance." + }, + { + "type": "image", + "bbox": [ + 0.081, + 0.264, + 0.47, + 0.354 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.093, + 0.355, + 0.458, + 0.367 + ], + "angle": 0, + "content": "(a) Embeddings projection for PASCAL Context (left) and COCO-Stuff (right)" + }, + { + "type": "image", + "bbox": [ + 0.08, + 0.369, + 0.469, + 0.47 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.128, + 0.47, + 0.424, + 0.48 + ], + "angle": 0, + "content": "(b) Retrieved image segments for classes not used in training" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.485, + 0.47, + 0.624 + ], + "angle": 0, + "content": "Figure 4. (a) Projection of pixel embeddings generated by CLIP, MaskCLIP+, and CLIP-S\\(^4\\) trained on Pascal Context with CLIP's ViT-B/16 model (left) and COCO-Stuff with CLIP's ResNet50 model (right). MaskCLIP+ distorts pixel embeddings with respect to the given text embeddings, while CLIP-S\\(^4\\) aligns pixel embeddings with the pre-trained CLIP model. (b) Image segments retrieved for classes that are not used during training. Correct and incorrect retrievals are outlined in green and orange, respectively. Compared with MaskCLIP+, CLIP-S\\(^4\\) retrieves images that are more closely related to the classes." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.627, + 0.47, + 0.749 + ], + "angle": 0, + "content": "image segments retrieved from the COCO-Stuff validation set using MaskCLIP+ and CLIP-S\\(^4\\) for a set of classes that are not used in training. For each class, we obtain its text embedding and compare it with segment embeddings to obtain top retrievals from both methods. Due to better alignment with the pre-trained CLIP model, CLIP-S\\(^4\\) retrieves images that are more closely related to the unknown classes compared with MaskCLIP+." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.757, + 0.403, + 0.773 + ], + "angle": 0, + "content": "4.4. Unsupervised Semantic Segmentation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.78, + 0.471, + 0.902 + ], + "angle": 0, + "content": "To study whether \\(\\mathrm{CLIP - S^4}\\) can generate pixel embeddings that form distinctive clusters, we evaluate \\(\\mathrm{CLIP - S^4}\\) on the unsupervised semantic segmentation task for Pascal VOC 2012. To derive semantic segmentation from pixel embeddings, we use and test two approaches, including \\(k\\) nearest neighbor (\\(k\\)-NN) search [19] and linear classification [40]. For \\(k\\)-NN search, we assign each segment a class label by the majority vote of its nearest neighbors from the" + }, + { + "type": "table", + "bbox": [ + 0.508, + 0.264, + 0.886, + 0.438 + ], + "angle": 0, + "content": "
MethodmIoU k-NNmIoU Linear Classifier
IIC [22]-28.0
SegSort [19]47.355.4
Hierarch. Group. [47]-48.8
MaskContrast (Sup.) [40]53.963.9
ConceptContrast [18]58.860.4
HSG [23]61.7-
MaskCLIP [49]67.369.5
MaskCLIP+ [49]65.170.0
CLIP-S472.0 (+4.7)73.0 (+3.0)
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.444, + 0.892, + 0.498 + ], + "angle": 0, + "content": "Table 4. Unsupervised semantic segmentation benchmarks (mIoU) on Pascal VOC 2012. CLIP-\\(S^4\\) consistently outperforms the state-of-the-art methods on both \\(k\\)-NN search and linear classification." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.504, + 0.892, + 0.549 + ], + "angle": 0, + "content": "training set following [19]. For linear classification, we train a linear classifier on the learned pixel embeddings following [40]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.55, + 0.893, + 0.747 + ], + "angle": 0, + "content": "We compare our method with the state-of-the-art unsupervised and language-guided semantic segmentation approaches. We train the state-of-the-art models using the same hyper-parameter setting as our approach except for IIC [22] and Hierarchical Grouping [47] as they use different training mechanisms. For comparison, we take the best results for IIC and Hierarchical Grouping. The benchmark results are shown in Tab. 4. With the vision-language guidance, our method achieves significant gains compared with the previous non-CLIP-based approaches (i.e., \\(+9\\%\\) for both \\(k\\)-NN search and linear classification). Meanwhile, our method also outperforms the language-guided semantic segmentation approaches by a large margin." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.757, + 0.724, + 0.773 + ], + "angle": 0, + "content": "4.5. Instance Mask Tracking" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.78, + 0.893, + 0.902 + ], + "angle": 0, + "content": "We evaluate the transferability of pixel embeddings learned from the Pascal VOC 2012 dataset. We use the pixel embeddings to track instance masks in the DAVIS 2017 validation set, where the instance masks at the first frame are given for each video. Following the prior work [48], we use the similarity between pixel embeddings cross frames to propagate the instance masks to the rest of the video frames. We evaluate the performance using the region similarity \\(\\mathcal{I}\\)" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "11213" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.091, + 0.089, + 0.457, + 0.282 + ], + "angle": 0, + "content": "
MethodJ(Mean)↑F(Mean)↑
MaskTrack-B [32]35.336.4
OSVOS-B [4]18.530.0
Video Colorization [42]34.632.7
CycleTime [43]41.939.4
mgPFF [24]42.246.9
Hierarchy. Group. [47]47.148.9
MaskContrast (Sup.) [40]34.336.7
ConceptContrast [18]50.453.9
MaskCLIP [49]48.149.2
MaskCLIP+ [49]42.644.2
CLIP-S452.3 (+1.9)56.8 (+2.9)
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.288, + 0.47, + 0.358 + ], + "angle": 0, + "content": "Table 5. Quantitative evaluation of instance mask tracking on the DAVIS-2017 validation set. The performance is measured by the region similarity \\(\\mathcal{J}\\) (IoU) and the contour-based accuracy \\(\\mathcal{F}\\) defined by [33]. Our method outperforms existing supervised, unsupervised, and language-guided approaches on both metrics." + }, + { + "type": "table", + "bbox": [ + 0.102, + 0.37, + 0.446, + 0.466 + ], + "angle": 0, + "content": "
LtLeLspAccmIoUavgsim
✓--1.60.5-0.01
✓✓-48.124.30.79
✓-✓52.332.90.33
-✓✓48.631.3-
✓✓✓53.733.60.66
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.472, + 0.47, + 0.569 + ], + "angle": 0, + "content": "Table 6. Ablation study on the contribution of each loss of CLIP-S4. Experimented on language-guided semantic segmentation of Pascal Context. avgsim represents the average cosine similarity between segment embeddings generated by CLIP-S4 and CLIP. By combining the embedding and semantic consistent losses \\(\\mathcal{L}_e\\) and \\(\\mathcal{L}_s\\), CLIP-S4 archives better semantic segmentation performance while maintaining the alignment with CLIP's embeddings." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.579, + 0.453, + 0.594 + ], + "angle": 0, + "content": "(IoU) and the contour-based accuracy \\(\\mathcal{F}\\) defined by [33]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.595, + 0.468, + 0.731 + ], + "angle": 0, + "content": "We compare our method with existing supervised [4, 32], unsupervised [18, 24, 40, 42, 43, 47], and language-guided [49] approaches (Tab. 5). Though not trained on any video sequences, our method outperforms the existing approaches by more than \\(1.9\\%\\) and \\(2.9\\%\\) in terms of the region similarity \\(\\mathcal{I}\\) and contour accuracy \\(\\mathcal{F}\\), respectively. Note that the pixel embeddings generated by \\(\\mathrm{MaskCLIP + }\\) [49] are distorted by the classes from Pascal VOC 2012, which hinder their transferability." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.741, + 0.231, + 0.758 + ], + "angle": 0, + "content": "4.6. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.47, + 0.903 + ], + "angle": 0, + "content": "We study the contribution of different losses of our method using the Pascal Context dataset and the language-guided semantic segmentation task. The performance is evaluated with pixel accuracy (pAcc) and mIoU. We also calculate the average cosine similarity (avgsim) between our segment embeddings and CLIP's segment embeddings to quantify the alignment. Tab. 6 shows the study results. We observe that by introducing embedding consistent loss \\(\\mathcal{L}_e\\) the learned segment embeddings are well aligned with" + }, + { + "type": "table", + "bbox": [ + 0.511, + 0.089, + 0.675, + 0.191 + ], + "angle": 0, + "content": "
#unknowns(u)mIoU
1671.6
3271.9
6472.0
12872.2
25671.9
" + }, + { + "type": "table", + "bbox": [ + 0.692, + 0.089, + 0.882, + 0.191 + ], + "angle": 0, + "content": "
top-m segmentsavgsim
10.921
40.977
160.996
640.997
2560.993
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.197, + 0.684, + 0.24 + ], + "angle": 0, + "content": "Table 7. Ablation study on the influence of the number of unknown class prototypes." + }, + { + "type": "table_caption", + "bbox": [ + 0.688, + 0.197, + 0.887, + 0.24 + ], + "angle": 0, + "content": "Table 8. Ablation study on different numbers of top- \\( m \\) segments for class prototypes." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.243, + 0.892, + 0.469 + ], + "angle": 0, + "content": "CLIP's embeddings with an average cosine similarity of 0.79. However, the learned segment embeddings do not perform well on the language-guided semantic segmentation task (24.3 vs. 33.6), because the segment embeddings are not optimized to classify target classes. Meanwhile, by using semantic consistent loss \\(\\mathcal{L}_s\\) without embedding consistent loss, the learned segment embeddings have the discriminative power to classify different classes but are not aligned with CLIP's embeddings as the average cosine similarity is 0.33. As a result, the segment embeddings are limited to the target classes used during training. Hence, we combine \\(\\mathcal{L}_e\\) and \\(\\mathcal{L}_s\\) to balance the discriminative power over target classes and the alignment with CLIP. Meanwhile, we observe that with pixel-segment contrastive learning, the model can archive better performance." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.47, + 0.892, + 0.726 + ], + "angle": 0, + "content": "Also, we study the influence of the number of unknown class prototypes on the Pascal VOC dataset for the unsupervised semantic segmentation task. The results Tab. 7 show that the semantic segmentation performance is robust to the tested number of unknown class prototypes as the mIoU varies only \\(0.6\\%\\). Furthermore, we investigate how the size of top-\\(m\\) segments impacts the embeddings of class prototypes. We compare the embeddings of class prototypes generated with different numbers of top-\\(m\\) segments on the Pascal VOC dataset. We use the embeddings of class prototypes generated with \\(m = 32\\) segments as the reference, and compute the cosine similarity between the reference prototypes embeddings and ones with different top-\\(m\\) segments. For each case, the cosine similarity is averaged over all class prototypes. We observe that the embeddings of class prototypes are relatively stable if moderate top-\\(m\\) segments (e.g., \\(m = 32\\) in this work) are used (Tab. 8)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.74, + 0.619, + 0.755 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.892, + 0.901 + ], + "angle": 0, + "content": "We propose CLIP-S\\(^4\\), a novel pixel representation learning approach for semantic segmentation. Our method combines self-supervised contrastive learning and guidance of CLIP to learn consistent pixel embeddings with respect to visual and conceptual semantics. Our experiments on popular semantic segmentation benchmarks demonstrate consistent gains over the state-of-the-art unsupervised semantic segmentation and language-driven semantic segmentation methods, especially for unknown classes." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "11214" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.116, + 0.47, + 0.17 + ], + "angle": 0, + "content": "[1] Radhakrishna Achanta, Appu Shaji, Kevin Smith, Aurelien Lucchi, Pascal Fua, and Sabine Süsstrunk. SLIC superpixels compared to state-of-the-art superpixel methods. IEEE TPAMI, 34(11):2274-2282, 2012. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.173, + 0.47, + 0.214 + ], + "angle": 0, + "content": "[2] Pablo Arbeláez, Michael Maire, Charless Fowlkes, and Jitendra Malik. Contour detection and hierarchical image segmentation. IEEE TPAMI, 33(5):898-916, 2011. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.216, + 0.47, + 0.258 + ], + "angle": 0, + "content": "[3] Maxime Bucher, Tuan-Hung VU, Matthieu Cord, and Patrick Pérez. Zero-shot semantic segmentation. In NeurIPS, pages 468-479, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.26, + 0.47, + 0.3 + ], + "angle": 0, + "content": "[4] S. Caelles, K.-K. Maninis, J. Pont-Tuset, L. Leal-Taixe, D. Cremers, and L. Van Gool. One-shot video object segmentation. In CVPR, pages 5320-5329, 2017. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.302, + 0.47, + 0.343 + ], + "angle": 0, + "content": "[5] Holger Caesar, Jasper Uijlings, and Vittorio Ferrari. COCO-Stuff: Thing and stuff classes in context. In CVPR, pages 1209-1218, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.346, + 0.47, + 0.414 + ], + "angle": 0, + "content": "[6] Liang-Chieh Chen, George Papandreou, Iasonas Kokkinos, Kevin Murphy, and Alan L. Yuille. DeepLab: Semantic image segmentation with deep convolutional nets, atrous convolution, and fully connected CRFs. IEEE TPAMI, 40(4):834-848, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.417, + 0.47, + 0.471 + ], + "angle": 0, + "content": "[7] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In ICML, pages 1597-1607, 2020. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.474, + 0.47, + 0.529 + ], + "angle": 0, + "content": "[8] Jang Hyun Cho, Utkarsh Mall, Kavita Bala, and Bharath Hariharan. PiCIE: Unsupervised semantic segmentation using invariance and equivariance in clustering. In CVPR, pages 16794-16804, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.531, + 0.47, + 0.572 + ], + "angle": 0, + "content": "[9] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. ImageNet: A large-scale hierarchical image database. In CVPR, pages 248–255, 2009. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.574, + 0.47, + 0.657 + ], + "angle": 0, + "content": "[10] Xiaoyi Dong, Yinglin Zheng, Jianmin Bao, Ting Zhang, Dongdong Chen, Hao Yang, Ming Zeng, Weiming Zhang, Lu Yuan, Dong Chen, Fang Wen, and Nenghai Yu. MaskCLIP: Masked self-distillation advances contrastive language-image pretraining. arXiv preprint arXiv:2208.12262, 2022.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.659, + 0.47, + 0.7 + ], + "angle": 0, + "content": "[11] Sepideh Esmaeilpour, Bing Liu, Eric Robertson, and Lei Shu. Zero-shot out-of-distribution detection based on the pre-trained model CLIP. In AAAI, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.702, + 0.47, + 0.756 + ], + "angle": 0, + "content": "[12] Mark Everingham, Luc Van Gool, Christopher K. I. Williams, John Winn, and Andrew Zisserman. The pascal visual object classes (voc) challenge. *IJCV*, 88(2):303–338, 2010. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.759, + 0.468, + 0.8 + ], + "angle": 0, + "content": "[13] Xiuye Gu, Tsung-Yi Lin, Weicheng Kuo, and Yin Cui. Open-vocabulary object detection via vision and language knowledge distillation. In ICLR, 2022. 3, 4, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.803, + 0.47, + 0.857 + ], + "angle": 0, + "content": "[14] Mark Hamilton, Zhoutong Zhang, Bharath Hariharan, Noah Snavely, and William T. Freeman. Unsupervised semantic segmentation by distilling feature correspondences. In ICLR, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.859, + 0.47, + 0.901 + ], + "angle": 0, + "content": "[15] Bharath Hariharan, Pablo Arbeláez, Lubomir Bourdev, Subhransu Maji, and Jitendra Malik. Semantic contours from inverse detectors. In ICCV, pages 991-998, 2011. 5" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.116, + 0.47, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.135 + ], + "angle": 0, + "content": "[16] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In CVPR, pages 9726-9735, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.137, + 0.892, + 0.178 + ], + "angle": 0, + "content": "[17] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, pages 770-778, 2016. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.18, + 0.892, + 0.235 + ], + "angle": 0, + "content": "[18] Wenbin He, William Surmeier, Arvind Kumar Shekar, Liang Gou, and Liu Ren. Self-supervised semantic segmentation grounded in visual concepts. In IJCAI, pages 949-955, 2022. 1, 2, 3, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.238, + 0.892, + 0.293 + ], + "angle": 0, + "content": "[19] Jyh-Jing Hwang, Stella X. Yu, Jianbo Shi, Maxwell D. Collins, Tien-Ju Yang, Xiao Zhang, and Liang-Chieh Chen. SegSort: Segmentation by discriminative sorting of segments. In ICCV, pages 7333-7343, 2019. 1, 2, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.296, + 0.892, + 0.337 + ], + "angle": 0, + "content": "[20] Phillip Isola, Daniel Zoran, Dilip Krishnan, and Edward H. Adelson. Crisp boundary detection using pointwise mutual information. In ECCV, pages 799-814, 2014. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.34, + 0.892, + 0.38 + ], + "angle": 0, + "content": "[21] Ajay Jain, Matthew Tancik, and Pieter Abbeel. Putting NeRF on a diet: Semantically consistent few-shot view synthesis. In ICCV, pages 5865-5874, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.383, + 0.892, + 0.424 + ], + "angle": 0, + "content": "[22] Xu Ji, Andrea Vedaldi, and Joao Henriques. Invariant information clustering for unsupervised image classification and segmentation. In ICCV, pages 9864-9873, 2019. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.427, + 0.892, + 0.493 + ], + "angle": 0, + "content": "[23] Tsung-Wei Ke, Jyh-Jing Hwang, Yunhui Guo, Xudong Wang, and Stella X. Yu. Unsupervised hierarchical semantic segmentation with multiview cosegmentation and clustering transformers. In CVPR, pages 2561–2571, 2022. 1, 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.498, + 0.892, + 0.539 + ], + "angle": 0, + "content": "[24] Shu Kong and Charless Fowlkes. Multigrid predictive filter flow for unsupervised learning on videos. arXiv preprint arXiv:1904.01693, 2019. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.542, + 0.892, + 0.582 + ], + "angle": 0, + "content": "[25] Boyi Li, Kilian Q Weinberger, Serge Belongie, Vladlen Koltun, and René Ranftl. Language-driven semantic segmentation. In ICLR, 2022. 1, 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.585, + 0.892, + 0.641 + ], + "angle": 0, + "content": "[26] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dolkar, and C. Lawrence Zitnick. Microsoft COCO: Common objects in context. In ECCV, pages 740-755, 2014. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.643, + 0.892, + 0.685 + ], + "angle": 0, + "content": "[27] Jonathan Long, Evan Shelhamer, and Trevor Darrell. Fully convolutional networks for semantic segmentation. In CVPR, pages 3431-3440, 2015. 1, 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.687, + 0.892, + 0.742 + ], + "angle": 0, + "content": "[28] Luke Melas-Kyriazi, Christian Rupprecht, Iro Laina, and Andrea Vedaldi. Deep spectral methods: A surprisingly strong baseline for unsupervised semantic segmentation and localization. In CVPR, pages 8354-8365, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.744, + 0.892, + 0.812 + ], + "angle": 0, + "content": "[29] Roozbeh Mottaghi, Xianjie Chen, Xiaobai Liu, Nam-Gyu Cho, Seong-Whan Lee, Sanja Fidler, Raquel Urtasun, and Alan Yuille. The role of context for object detection and semantic segmentation in the wild. In CVPR, pages 891–898, 2014. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.816, + 0.892, + 0.856 + ], + "angle": 0, + "content": "[30] Yassine Ouali, Céline Hudelot, and Myriam Tami. Autoregressive unsupervised image segmentation. In ECCV, pages 142-158, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.859, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[31] Dim P. Papadopoulos, Ethan Weber, and Antonio Torralba. Scaling up instance annotation via label propagation. In ICCV, pages 15344-15353, 2021. 1" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "11215" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.093, + 0.469, + 0.146 + ], + "angle": 0, + "content": "[32] Federico Perazzi, Anna Khoreva, Rodrigo Benenson, Bernt Schiele, and Alexander Sorkine-Hornung. Learning video object segmentation from static images. In CVPR, pages 3491-3500, 2017. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.148, + 0.469, + 0.204 + ], + "angle": 0, + "content": "[33] F. Peruzzi, J. Pont-Tuset, B. McWilliams, L. Van Gool, M. Gross, and A. Sorkine-Hornung. A benchmark dataset and evaluation methodology for video object segmentation. In CVPR, pages 724-732, 2016. 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.205, + 0.469, + 0.285 + ], + "angle": 0, + "content": "[34] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In ICML, pages 8748-8763, 2021. 1, 2, 4, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.288, + 0.469, + 0.354 + ], + "angle": 0, + "content": "[35] Yongming Rao, Wenliang Zhao, Guangyi Chen, Yansong Tang, Zheng Zhu, Guan Huang, Jie Zhou, and Jiwen Lu. DenseCLIP: Language-guided dense prediction with context-aware prompting. In CVPR, pages 18061-18070, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.357, + 0.469, + 0.411 + ], + "angle": 0, + "content": "[36] Ramprasaath R. Selvaraju, Karan Desai, Justin Johnson, and Nikhil Naik. CASTing your model: Learning to localize improves self-supervised representations. In CVPR, pages 11053-11062, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.413, + 0.469, + 0.453 + ], + "angle": 0, + "content": "[37] Gyungin Shin, Weidi Wie, and Samuel Albanie. ReCo: Retrieve and co-segment for zero-shot transfer. In NeurIPS, 2022. 1, 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.455, + 0.469, + 0.523 + ], + "angle": 0, + "content": "[38] Kihyuk Sohn, David Berthelot, Chun-Liang Li, Zizhao Zhang, Nicholas Carlini, Ekin D. Cubuk, Alex Kurakin, Han Zhang, and Colin Raffel. FixMatch: Simplifying semi-supervised learning with consistency and confidence. In NeurIPS, pages 596-608, 2020. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.525, + 0.469, + 0.579 + ], + "angle": 0, + "content": "[39] Wouter Van Gansbeke, Simon Vandenhende, Stamatios Georgoulis, and Luc V Gool. Revisiting contrastive methods for unsupervised learning of visual representations. In NeurIPS, pages 16238-16250, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.581, + 0.469, + 0.635 + ], + "angle": 0, + "content": "[40] Wouter Van Gansbeke, Simon Vandenhende, Stamatios Georgoulis, and Luc Van Gool. Unsupervised semantic segmentation by contrasting object mask proposals. In ICCV, pages 10032-10042, 2021. 1, 2, 5, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.637, + 0.469, + 0.703 + ], + "angle": 0, + "content": "[41] Yael Vinker, Ehsan Pajouheshgar, Jessica Y. Bo, Roman Christian Bachmann, Amit Haim Bermano, Daniel Cohen-Or, Amir Zamir, and Ariel Shamir. CLIPasso: Semantically-aware object sketching. ACM TOG, 41(4):86:1-86:11, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.706, + 0.469, + 0.747 + ], + "angle": 0, + "content": "[42] Carl Vondrick, Abhinav Shrivastava, Alireza Fathi, Sergio Guadarrama, and Kevin Murphy. Tracking emerges by colorizing videos. In ECCV, pages 402-419, 2018. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.748, + 0.469, + 0.788 + ], + "angle": 0, + "content": "[43] Xiaolong Wang, Allan Jabri, and Alexei A. Efros. Learning correspondence from the cycle-consistency of time. In CVPR, pages 2561-2571, 2019. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.79, + 0.469, + 0.843 + ], + "angle": 0, + "content": "[44] Zhaoqing Wang, Yu Lu, Qiang Li, Xunqiang Tao, Yandong Guo, Mingming Gong, and Tongliang Liu. CRIS: Clip-driven referring image segmentation. In CVPR, pages 11676-11685, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.845, + 0.469, + 0.872 + ], + "angle": 0, + "content": "[45] Saining Xie and Zhuowen Tu. Holistically-nested edge detection. In ICCV, pages 1395-1403, 2015. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.874, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[46] Jiarui Xu, Shalini De Mello, Sifei Liu, Wonmin Byeon, Thomas Breuel, Jan Kautz, and Xiaolong Wang. GroupViT:" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.093, + 0.469, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "Semantic segmentation emerges from text supervision. In CVPR, pages 18134-18144, 2022. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.122, + 0.892, + 0.163 + ], + "angle": 0, + "content": "[47] Xiao Zhang and Michael Maire. Self-supervised visual representation learning from hierarchical grouping. In NeurIPS, pages 16579-16590, 2020. 2, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.165, + 0.892, + 0.206 + ], + "angle": 0, + "content": "[48] Hengshuang Zhao, Jianping Shi, Xiaojuan Qi, Xiaogang Wang, and Jiaya Jia. Pyramid scene parsing network. In CVPR, pages 6230-6239, 2017. 1, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.208, + 0.892, + 0.234 + ], + "angle": 0, + "content": "[49] Chong Zhou, Chen Change Loy, and Bo Dai. Extract free dense labels from CLIP. In ECCV, 2022. 1, 2, 3, 4, 5, 6, 7, 8" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.234 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "11216" + } + ] +] \ No newline at end of file diff --git a/2023/CLIP-S4_ Language-Guided Self-Supervised Semantic Segmentation/c24eb4b4-f000-47c7-8244-3296ba8e2cac_origin.pdf b/2023/CLIP-S4_ Language-Guided Self-Supervised Semantic Segmentation/c24eb4b4-f000-47c7-8244-3296ba8e2cac_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..142214c642c6931f7a5e31e83558ce28b0934a59 --- /dev/null +++ b/2023/CLIP-S4_ Language-Guided Self-Supervised Semantic Segmentation/c24eb4b4-f000-47c7-8244-3296ba8e2cac_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5cf0babdd79548e844ea50e6fcc980259a6f0f9f2744e0594f1184e125a80934 +size 5577064 diff --git a/2023/CLIP-S4_ Language-Guided Self-Supervised Semantic Segmentation/full.md b/2023/CLIP-S4_ Language-Guided Self-Supervised Semantic Segmentation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..81911db6e6222066343ffe2ba2bfc34524e27788 --- /dev/null +++ b/2023/CLIP-S4_ Language-Guided Self-Supervised Semantic Segmentation/full.md @@ -0,0 +1,305 @@ +# CLIP-S $^4$ : Language-Guided Self-Supervised Semantic Segmentation + +Wenbin He + +Suphanut Jamonnak + +Liang Gou + +Liu Ren + +Bosch Research North America & Bosch Center for Artificial Intelligence (BCAI) + +{wenbin.he2, suphanut.jamonnak, liang.gou, liu. ren}@us.bosch.com + +# Abstract + +Existing semantic segmentation approaches are often limited by costly pixel-wise annotations and predefined classes. In this work, we present CLIP- $S^4$ that leverages self-supervised pixel representation learning and vision-language models to enable various semantic segmentation tasks (e.g., unsupervised, transfer learning, language-driven segmentation) without any human annotations and unknown class information. We first learn pixel embeddings with pixel-segment contrastive learning from different augmented views of images. To further improve the pixel embeddings and enable language-driven semantic segmentation, we design two types of consistency guided by vision-language models: 1) embedding consistency, aligning our pixel embeddings to the joint feature space of a pre-trained vision-language model, CLIP [34]; and 2) semantic consistency, forcing our model to make the same predictions as CLIP over a set of carefully designed target classes with both known and unknown prototypes. Thus, CLIP- $S^4$ enables a new task of class-free semantic segmentation where no unknown class information is needed during training. As a result, our approach shows consistent and substantial performance improvement over four popular benchmarks compared with the state-of-the-art unsupervised and language-driven semantic segmentation methods. More importantly, our method outperforms these methods on unknown class recognition by a large margin. + +# 1. Introduction + +Semantic segmentation aims to partition an input image into semantically meaningful regions and assign each region a semantic class label. Recent advances in semantic segmentation [6, 27, 48] heavily rely on pixel-wise human annotations, which have two limitations. First, acquiring pixel-wise annotations is extremely labor intensive and costly, which can take up to 1.5 hours to label one image [31]. Second, human annotations are often limited to a set of predefined semantic classes, with which the learned models lack the ability to recognize unknown classes [25]. + +![](images/cadd67d538f7dadd7cd4817c2093ccbe18d4c78f72380ba46acf92e465f8d87a.jpg) +(b) Semantic segmentation generated by different CLIP-based unsupervised approaches + +![](images/aec1cbba951d4512a5976fd04ee59e85632cabacb2b6ac368e140ff2bc1a30ad.jpg) +(c) CLIP- $S^4$ aligns the pixel embeddings and their semantics with CLIP feature space +Figure 1. (a) Pixel embeddings from different CLIP-based unsupervised methods: Our method, $\mathrm{CLIP - S^4}$ , generates sharper and more coherent pixel embeddings than MaskCLIP [49] and MaskCLIP+’s [49]; (b) Language-driven semantic segmentation by different methods: $\mathrm{CLIP - S^4}$ can recognize challenging unknown classes (e.g., moon); (c) The key idea behind $\mathrm{CLIP - S^4}$ : aligning the pixel embeddings and their semantics with CLIP feature space. + +Various approaches have been proposed to tackle these limitations, among which we are inspired by two lines of recent research in particular. First, for unsupervised semantic segmentation (i.e., without human annotations), self-supervised pixel representation learning approaches [14,18, 19, 23, 40] have shown promising results on popular unsupervised benchmarks. The main idea is to extend self-supervised contrastive learning [7, 16] from images to pixels by attracting each pixel's embedding to its positive pairs and repelling it from negative pairs. The prior of pairs can be contours [18, 19], hierarchical groups [23], salience maps [40], and pre-trained models [14]. Although these approaches can group pixels into semantically meaningful clusters, human annotations are still needed to assign class labels to the clusters for semantic segmentation [37]. + +Second, for unknown classes in semantic segmentation, large-scale vision-language models such as CLIP [34] + +have shown great potential. This line of research, called language-driven semantic segmentation, aims to segment images with arbitrary classes defined by texts during testing time [25, 37, 46, 49]. Among these methods, most still need training time annotations, such as pixel annotations [25] and captions [46]. Only a few recent work, MaskCLIP & MaskCLIP+ [49] attempts to address this without using additional supervision: MaskCLIP directly extracts pixel embeddings correlated with texts from CLIP, but these pixel embeddings are coarse and noisy (Fig. 1a). To address this issue, MaskCLIP+ [49] trains a segmentation model on the pseudo-labels generated by MaskCLIP for a set of predefined classes. However, the pixel embeddings of MaskCLIP+ are distorted by the predefined classes (Fig. 1a), which limits its ability to recognize unknowns (Fig. 1b). Also, it needs unknown class information during training, which hinders its real-world applications. + +We propose a language-guided self-supervised semantic segmentation approach, CLIP-S $^4$ , which takes advantage of the strengths from both lines of research and addresses their limitations accordingly. The key idea is to learn consistent pixel embeddings with respect to visual and conceptual semantics using self-supervised learning and the guidance of a vision-language model, CLIP. + +Specifically, we first train pixel embeddings with pixel-segment contrastive learning from different augmented image views [18, 19, 23] such that images can be partitioned into visually meaningful regions. To further improve pixel embedding quality and enable language-driven semantic segmentation, we introduce vision-language model guided consistency to regularize our model (Fig. 1c). The consistency is enforced from two aspects: embedding consistency and semantic consistency. First, embedding consistency aims to align the pixel embeddings generated by our model with the joint feature space of texts and images of CLIP by minimizing the distance between the pixel embeddings generated by our model and CLIP. Second, semantic consistency forces our model to make the same prediction as CLIP over a set of carefully designed target classes with both known and unknown prototypes. Note that unlike the previous methods [25,49] that use a predefined set of known classes, CLIP- $\mathbf{S}^4$ also learns the representation of unknown classes from images during training. + +In the end, CLIP- $S^4$ also enables a new task, namely class-free semantic segmentation, as shown in Tab. 1. This new task does not need any human annotations and even assumes NO class names are given during training. This is a more challenging task than the recent work [49] that requires class names of both known and unknown. + +In summary, the contributions of this paper are threefold: + +- We propose a self-supervised semantic segmentation approach that combines pixel-segment contrastive learning with the guidance of pre-trained vision lan + +
KnownUnknownAdd. Info.
Annot.CIs NameAnnot.CIs Name
Un/Self-supervised ([19] etc.)××××Fine-Tuning
Supervised ([27] etc.)✓✓N/AN/AN/A
Zero-shot ([3] etc.)✓✓×✓Word2Vec, etc.
Language-MaskCLIP+[49]×✓CLIP
DrivenCLIP-S4×✓×CLIP
+ +Table 1. Comparison of information required for training over different tasks. CLIP- $\mathbf{S}^4$ enables a new task called class-free semantic segmentation. Compared with MaskCLIP+ [49], the new task assumes unknown class names are NOT given during training. + +guage models. Our method can generate high-quality pixel embeddings without any human annotations and be applied to a variety of semantic segmentation tasks. + +- We open up new research potentials for language-driven semantic segmentation without any human annotations by introducing and addressing a new task of class-free semantic segmentation (Tab. 1). Unlike previous work that assumes all the class names are known during training, our method can discover unknown classes from unlabelled image data without even knowing unknown class names. +- Consistent and substantial gains are observed with our approach over the state-of-the-art unsupervised and language-driven semantic segmentation methods on four popular datasets. More importantly, our method significantly outperforms the state-of-the-art on the segmentation of unknown classes. + +# 2. Related Work + +Unsupervised Semantic Segmentation. There are two groups of recent unsupervised semantic segmentation methods. One group of methods learns to generate consistent pixel representations or predictions between different augmentation of images with the guidance of mutual information [22, 30], clusters [8], contours [18, 19], hierarchical groups [23, 47], and saliency masks [40]. The other group of methods extracts dense features from pre-trained models based on saliency maps [36], augmentations [39], spectral decomposition [28], and feature correspondences [14]. While these methods can generate pixel embeddings with semantically meaningful clusters, annotations are needed to assign class labels to the clusters (e.g., $k$ -nearest neighbor search [19] and Hungarian algorithm [40]). Our work combines pixel-segment self-supervision with pre-trained vision-language models to enable semantic segmentation without any human annotations. + +Language-Driven Semantic Segmentation. Recently, vision-language models (e.g., CLIP [34]) trained on large- + +![](images/c70f39b4ee98291a5d3b8dfef69c8d525810cae5961988bd0bf15d23fa91c1f0.jpg) +Figure 2. Framework of CLIP-S4. (a) It starts with an encoder that maps images into pixel embeddings for semantic segmentation. Then it follows with two components: 1) self-supervised contrastive learning and 2) vision-language model guided consistency. Specifically, shown in (b), the self-supervised contrastive learning forces pixel embeddings to be consistent within visually coherent regions and among different augmented views of the same image. For vision-language model guided consistency, this framework introduces (c) embedding consistency that aligns the pixel embeddings generated by our model with CLIP embeddings, and (d) semantic consistency that forces our model to make the same predictions as CLIP for a set of target classes with both known and unknown prototypes. For known classes, the prototypes are pre-computed and fixed during training. For unknown classes, the prototypes are learned during training via clustering (e). + +scale image-text datasets have shown great potential on various downstream tasks such as image synthesis [21, 41], out-of-distribution detection [11], and object detection [13]. To extend vision-language models for semantic segmentation, one active research, language-driven semantic segmentation, aims to segment images with arbitrary unknown classes defined by texts during testing time [25, 37, 46, 49]. Some methods [25, 44] use pixel-wise annotations to train language-guided semantic segmentation models. Other methods [10, 46] perform large-scale pre-training on image-text pairs specifically for semantic segmentation. + +By contrast, we directly use vision-language models that are pre-trained for classification tasks. Along this line of research, a few approaches have been proposed [35, 37, 49]. The most relevant approach to our method is MaskCLIP [49], which extends the embeddings generated by pre-trained vision-language models from image to pixel level, but these embeddings are often coarse and noisy. To address this issue, MaskCLIP + [49] fine-tunes the pixel embeddings by the pseudo-labels of a specific set of classes on top of MaskCLIP. However, it needs unknown class names during training, which may not be possible in real-world cases. Compared with [49], our method can recognize unknown classes without knowing any unknown class information during training time, and also learns fine-grained and sharper pixel embeddings with self-supervision. + +# 3. Method + +Our method (Fig. 2) segments images by learning a pixel embedding function with self-supervised contrastive learning and the guidance of a pre-trained vision-language model, CLIP. We use self-supervised contrastive learning + +to force the consistency of pixel embeddings within visually coherent regions (e.g., superpixels) and among different augmented views of the same image (Sec. 3.1). We also introduce two vision-language model guided consistency (i.e., embedding consistency and semantic consistency) to further regularize the model (Sec. 3.2). The two components are complementary to each other. On the one hand, contrastive learning mitigates the noise introduced by CLIP. On the other hand, with the knowledge extracted from CLIP, the quality of the pixel embeddings can be improved. More importantly, this approach enables us to perform language-driven semantic segmentation with our carefully designed target class prototypes of both knowns and unknowns. In the following, we discuss the two components in detail. + +# 3.1. Pixel-Segment Contrastive Learning + +We train a pixel embedding function to generate consistent pixel embeddings within visually coherent regions through pixel-segment contrastive learning [18,23]. Specifically, the embedding function transforms each pixel $p$ of an image to a unit-length embedding vector $\mathbf{z}_p$ of dimension $d$ via a deep neural network. The image is then partitioned into $|S|$ segments by clustering the pixel embeddings. The embedding $\mathbf{v}_s$ of each segment $s$ is calculated as the average of the pixel embeddings $\mathbf{v}_s = \sum_{p\in s}\mathbf{z}_p / |s|$ , which is also normalized into a unit-length vector $\mathbf{v}_s = \mathbf{v}_s / \| \mathbf{v}_s\|$ . For each pixel $p$ , the segments are grouped into two sets including a positive set $S^+$ and a negative set $S^-$ . The positive set $S^+$ of a pixel contains segments within the same visually coherent region of the pixel. Following the prior work [18, 23], the visually coherent region can be derived from super-pixels [1] or contours [2]. We also use data augmentation (e.g., random cropping and color jitter + +ing) to generate consistent pixel embeddings between different augmented views of the same image. Hence, segments within the same region of the pixel in any augmented views are considered as the positive set $S^{+}$ . Other segments in the image and segments from other images in the same batch are included in the negative set $S^{-}$ . The pixel embedding $\mathbf{z}_p$ is then attracted to the segments in positive set $S^{+}$ and repelled from the segments in negative set $S^{-}$ with contrastive loss: + +$$ +\mathcal {L} _ {t} (p) = - \log \frac {\sum_ {s \in \mathcal {S} ^ {+}} e x p \left(\operatorname {s i m} \left(\mathbf {z} _ {p} , \mathbf {v} _ {s}\right) \kappa\right)}{\sum_ {s \in \mathcal {S} ^ {+} \cup \mathcal {S} ^ {-}} e x p \left(\operatorname {s i m} \left(\mathbf {z} _ {p} , \mathbf {v} _ {s}\right) \kappa\right)}, \tag {1} +$$ + +where $\kappa$ is the concentration constant and $sim(\mathbf{z}_p, \mathbf{v}_s)$ is the cosine similarity between the pixel embedding $\mathbf{z}_p$ and the segment embedding $\mathbf{v}_s$ . + +# 3.2. Vision-Language Model Guided Consistency + +To enable language-driven semantic segmentation and improve the quality of pixel embeddings, we use a pretrained vision-language model such as CLIP [34] to guide the training of the pixel embedding function. The key idea is to align the output space of our pixel embedding function consistent with the feature space of CLIP (Fig. 1c). Specifically, two types of consistency are considered during training including embedding consistency and semantic consistency, which are detailed as follows. + +**Embedding Consistency.** Our goal is to align the pixel embeddings generated from our self-supervised method (the green contour in Fig. 1c) with CLIP's pixel embeddings (the orange contour in Fig. 1c). This is done by minimizing the distance between the two pixel embedding spaces. + +We first obtain the pixel embeddings of an input image from CLIP by modifying the attention-based pooling layer of the CLIP image encoder following [49]. Specifically, we 1) remove the query and key projection layers and 2) reformulate the value projection layer and the last linear layer as two consecutive fully connected layers. In the following, we use $clip - i(\cdot)$ as the modified CLIP image encoder and $clip - t(\cdot)$ as CLIP text encoder. + +Then we obtain the pixel embeddings of CLIP for different augmented views of the image. Note that we use the original image to generate the CLIP pixel embeddings and perform augmentation afterwards to make sure that the CLIP pixel embeddings are consistent among different augmented views. In the end, we minimize the distance of embeddings between segments instead of pixels from our self-supervised and CLIP embedding spaces. This is because the pixel embeddings of CLIP are noisy (Fig. 2), which can be mitigated by aggregating over segments. Hence, we use the pixel embeddings generated by our model to derive segments (clusters) and then apply them to the CLIP's pixel embeddings. In the end, for each segment $s$ , the embedding + +![](images/e1c79c194e1189f50624c1c920fb3868b965be53cee49adf0b56be8721db3348.jpg) +Figure 3. Computation of target class prototypes with both knowns and unknowns, $C = \{C_k, C_u\}$ . For a set of known classes (e.g., bird, cat), we first obtain their CLIP text embeddings, $T$ , via a set of prompt templates [13,49] (shown in $\widehat{\mathbf{a}}$ ); then, we calculate the normalized (via softmax) similarity between text embeddings, $T$ , and all segments' CLIP embeddings, $I$ , from training images, and average the top- $m$ similar segments' CLIP embedding as the embedding prototype for this class, as shown in $\widehat{\mathbf{b}}$ ; For each unknown class, we randomly select the CLIP embedding of a segment as the initial prototype (shown in $\widehat{\mathbf{c}}$ ). + +consistent loss is defined as: + +$$ +\mathcal {L} _ {e} (s) = 1 - \operatorname {s i m} \left(\mathbf {v} _ {s}, \mathbf {i} _ {s}\right), \tag {2} +$$ + +where $\mathbf{v}_s$ and $\mathbf{i}_s$ are the segment embeddings derived from our embedding function and CLIP, respectively. Here, $\mathbf{i}_s$ is the average of the CLIP pixel embedding from segment $s$ , namely, $\mathbf{i}_s = \sum_{p \in s} \text{clip-}i(s) / |s|$ . + +Semantic Consistency In addition to embedding consistency, we introduce semantic consistency by forcing our model to make the same predictions of semantic classes as CLIP. The rationale is that we can generate better pixel embeddings if they can form distinctive clusters corresponding to different semantic classes, as the goal of semantic segmentation is to perform pixel-wise classification. + +Semantic consistency is achieved via a similar idea of pseudo-labeling [38]. Again, we force the semantic consistency at the segment level (not the pixel level) to reduce the noise in pseudo-labels. Specifically, for each segment $s$ , we first use CLIP to generate its pseudo-label $y_{s}$ over a set of target classes, which include both knowns and unknowns (we will introduce how to design these target classes later). The pseudo-label is generated based on the highest similarity between the segment embedding $\mathbf{i}_s$ with a set of prototypes, $C = \{\mathbf{c}_l\}_{0}^{L - 1}$ , of the target classes in the pixel embedding space of CLIP, namely, $y_{s} = \operatorname{argmax}_{l\in L}(sim(\mathbf{i}_{s},\mathbf{c}_{l}))$ . + +Then we define the semantic consistent loss as the cross entropy between our model's prediction $\varphi (\mathbf{v}_s)$ over the tar + +get classes and the pseudo-label $y_{s}$ : + +$$ +\mathcal {L} _ {s} (s) = \mathbf {H} \left(y _ {s}, \varphi (\mathbf {v} _ {s})\right), \tag {3} +$$ + +where $\varphi (\mathbf{v}_s) = \mathrm{softmax}(sim(\mathbf{v}_s,C))$ + +Target Class Prototypes The design of target classes and associated class prototypes, $C = \{\mathbf{c}_l\}_{0}^{L - 1}$ , is crucial to achieve the semantic consistency. Here, a class prototype, $\mathbf{c}_l$ , is an embedding vector that can represent a class in an embedding space. For example, it can be the mean vector of embeddings of all segments of a class "car". Currently, most existing methods [25, 49] assume that the target classes are already predefined, which is not feasible in real-world use cases without any human annotations. Thus, those methods cannot handle unknown classes hidden in the data. To address this issue, we introduce two sets of class prototypes of known, $C_k = \{\mathbf{c}_0,\dots ,\mathbf{c}_{k - 1}\}$ , and unknown classes, $C_u = \{\mathbf{c}_k,\dots ,\mathbf{c}_{k + u}\}$ , where the known classes are predefined by leveraging CLIP and the unknown classes are learned from image data during training. Thus, we have $C = \{\mathbf{c}_l\}_{0}^{L - 1} = \{\mathbf{c}_0,\dots ,\mathbf{c}_{k - 1},\mathbf{c}_k,\dots ,\mathbf{c}_{k + u}\}$ $L = k + u$ . + +For known classes, a natural choice is to use the text embeddings generated by CLIP as their class prototype embeddings [25, 49]. However, even though the text embeddings are trained to align with image/pixel embeddings [34], there is still a huge gap between the text and image/pixel embeddings in the joint space of CLIP (Fig. 1c). Therefore, it is challenging to learn meaningful unknown classes from image features when using text embeddings as class prototypes. Hence, in this work, we use the prototype of CLIP pixel embeddings to represent each known class. + +To this end, for a set of known classes (e.g., bird, cat), $K = \{0,\dots ,k - 1\}$ we first obtain their CLIP text embeddings, $T = \{\mathbf{t}_k\} = \{\text{clip - } t(k)\}$ , via a set of prompt templates following [13, 49], as shown in Fig. 3a. We also get a set of CLIP segment embeddings, $I = \{\mathbf{i}_{\hat{s}}\}$ , for all training images by a) feeding training images into the modified image encoder of CLIP to get pixel embeddings; b) clustering the pixel embeddings as segments, $\hat{S}$ ; c) averaging the pixel embeddings in each segment, $\hat{s}$ . Hence, we have embeddings for each segment: $\mathbf{i}_{\hat{s}} = \sum_{p\in \hat{s}}\text{clip - } i(p) / |\hat{s}|$ . Then, we calculate the similarity between text embeddings of known classes, $T$ , and all CLIP segment embeddings $I$ , and normalize the similarities over all classes by softmax. Finally, we average the top- $m$ similar segments' embedding as the embedding prototype for each class, $C_k = \{\mathbf{c}_k\} = avg_m(top-m_{\hat{s}}(\mathbf{softmax}_k(sim(I,T))))$ . + +The prototype embeddings of the unknown classes, $C_u$ , are randomly initialized by sampling the CLIP embeddings of all segments, namely $C_u = \text{random}(clip - i(\hat{\mathcal{S}}), u)$ , with a size of the unknown class of $u$ (Fig. 3c). During training, the embedding $\mathbf{c}_u$ of each unknown class prototype is updated by minimizing its distance to all segments that are classified as this unknown class (similar to updating the cen + +troids in $k$ -means clustering): + +$$ +\mathcal {L} _ {u} = \sum_ {s \in S _ {u}} (1 - \operatorname {s i m} (\mathbf {c} _ {u}, \operatorname {c l i p} - i (s))) / | S _ {u} |, \tag {4} +$$ + +where $S_{u}$ are the segments classified as the unknown classes. In this way, our model can also learn the pixel representation of unknown classes. + +# 3.3. Training and Inference + +In summary, we train the pixel embedding function by combining the pixel-segment contrastive loss, embedding consistent loss, and semantic consistent loss: + +$$ +\mathcal {L} = \mathcal {L} _ {t} + \mathcal {L} _ {e} + \mathcal {L} _ {s}. \tag {5} +$$ + +During training, we also update the embeddings for the unknown classes with $\mathcal{L}_u$ . + +For inference, we use the trained model to generate pixel embeddings for each input image and use the pixel embeddings for different downstream tasks, including language-driven and unsupervised semantic segmentation. For language-driven semantic segmentation, we first obtain the text embeddings of arbitrary inference classes by feeding the prompt-engineered texts into the text encoder of CLIP. Then we assign each pixel with the class label whose text embedding is the closest to $\mathrm{CLIP - S^4}$ pixel embedding. For unsupervised semantic segmentation, we follow previous work [19, 40] that uses $k$ nearest neighbor search or linear classifier to perform semantic segmentation. + +# 4. Experiments + +We evaluate our model on three tasks: 1) language-driven semantic segmentation for both known and unknown classes; 2) unsupervised semantic segmentation with $k$ -means clustering/linear classification; 3) transfer learning of generated pixel embeddings for instance mask tracking. We also conduct ablation studies to understand the components of our model. + +# 4.1. Datasets + +Pascal VOC 2012 [12] contains 20 object classes and a background class. It has 1,464 and 1,449 images for training and validation, respectively. Following common practice [27, 48], we augment the training data with additional annotations [15], resulting in 10,582 training images. + +Pascal Context [29] extends Pascal VOC 2010 [12] with additional annotations on 4,998 training and 5,105 validation images. Following the prior work [49], we use the most common 59 classes for evaluation. + +COCO-Stuff [5] labels MS COCO [26] with 171 object/stuff classes. It contains 118,287 and 5,000 images for training and validation, respectively. + +DAVIS 2017 [33] contains video sequences for instance mask tracking. Following the prior work [18, 47], we train pixel embeddings on Pascal VOC 2012 and evaluate the validation sequences without fine-tuning. + +It is worth mentioning that no ground truth labels of any datasets are used during training. Instead, we perform self-supervised learning on pseudo segments generated by contour detectors and owt-ucm [2]. Two contour detectors are used including HED [45] for Pascal VOC 2012 and Pascal Context and PMI [20] for COCO-Stuff. + +# 4.2. Implementation Details + +For self-supervised contrastive learning, images are augmented with the same set of data augmentations as SimCLR [7], including random resizing, cropping, flipping, color jittering, and Gaussian blurring. The concentration constant $\kappa$ is set to 10, and the number of segments is set to 36 for each augmented view. + +For vision-language guidance, we use pre-trained CLIP models [34] with modified image encoders following [49]. We use prompt-engineered texts with 85 prompt templates to generate text embeddings following [13, 49]. We use the average embedding of the top 32 segments of high probabilities as the prototype of each known class. We set the number of unknown classes to $u = 64$ . + +Following the prior work [18, 19, 23], we use PSP-Net [48] with a dilated ResNet-50 [17] backbone as the network architecture. The backbone is pre-trained on the ImageNet [9] dataset. We train our model on Pascal VOC 2012 and Pascal Context for $20\mathrm{k}$ iterations and on COCO-Stuff for $80\mathrm{k}$ iterations. We set the batch size to 8 with additional memory banks that cache the segment embeddings of the previous 2 batches. We set the initial learning rate to 0.001 and decay it with a polynomial learning rate policy. We use the CLIP model trained with ViT-B/16 backbone unless otherwise stated. + +# 4.3. Language-Driven Semantic Segmentation + +For language-driven semantic segmentation, no human annotations are used for either training or inference. At the inference time, each pixel is assigned an arbitrarily given class label whose CLIP text embedding is the closest to this pixel's CLIP- $\mathbf{S}^4$ embedding. + +We first compare the performance of our method with the state-of-the-art language-driven semantic segmentation approaches [37, 46, 49] on the Pascal Context and COCO-Stuff datasets. The performance is evaluated with the mean Intersection over Union (mIoU). For MaskCLIP and MaskCLIP+ [49], we obtain the results using the same hyper-parameter setting as our approach with CLIP models of two different backbones of ResNet50 and ViT-B/16. Meanwhile, GroupViT [46], ReCo [37], and $\mathrm{ReCo + }$ [37] use completely different training mechanisms compared + +
MethodCLIP ModelPascal ContextCOCO-Stuff
mIoUmIoU
GroupViT [46]-22.4-
ReCo [37]ResNet50x16 + ViT-L/14@336px26.6-
ReCo+ [37]†-18.4
MaskCLIP [49]ResNet5018.610.6
ViT-B/1625.215.2
MaskCLIP+ [49]†ResNet5023.413.9
ViT-B/1632.220.7
CLIP-S4†ResNet5028.5 (+5.1)16.7 (+2.8)
ViT-B/1633.6 (+1.4)22.1 (+1.4)
+ +Table 2. Language-guided semantic segmentation benchmarks (mIoU). CLIP- $S^4$ consistently outperforms the state-of-the-art methods on both Pascal Context and COCO-Stuff datasets with CLIP models of different backbones. $\dagger$ indicates the models are fine-tuned on target datasets. + +with our method. GroupViT is trained on image-caption pairs, and $\mathrm{ReCo} / \mathrm{ReCo}+$ combines image retrieval and cosegmentation. For comparison, we take the best results from [37, 46] for GroupViT, ReCo, and $\mathrm{ReCo}+$ . Tabel 2 shows the benchmarking results of the aforementioned methods. Our method consistently outperforms the state-of-the-art on both datasets with CLIP models of different backbones. + +To evaluate models' performance for class-free semantic segmentation with both known and unknown classes, we split the 59 classes of Pascal Context into 4 folds, where each fold includes around 15 classes. For each experiment, classes from one fold are considered as unknown and excluded during training. The mIoUs of known and unknown classes, as well as their harmonic mean (hIoU) are reported in Tab. 3. The performance of our method is averaged across 5 runs with randomly initialized prototypes of unknown classes. Our method achieves significant gains over MaskCLIP+ on unknown classes, which are comparable to MaskCLIP. Also, our method outperforms both MaskCLIP and MaskCLIP+ on known classes, which leads to better overall performance. + +Qualitatively, the visualization in Fig. 4a offers us some insights into why our approach can achieve better results: our model yields consistent embeddings aligned with the pre-trained CLIP model. Fig. 4a visualizes the projection of segment embeddings generated by different methods on Pascal Context and COCO-Stuff. We observe that segment embeddings generated by $\mathrm{MaskCLIP + }$ are distorted by the given text embeddings. Meanwhile, $\mathrm{CLIP - S^4}$ generates segment embeddings that are well aligned with the segment embeddings derived from the pre-trained CLIP model. Hence, segment embeddings generated by $\mathrm{CLIP - S^4}$ can better capture both known and unknown classes. Fig. 4b shows + +
Methodfold0fold1fold2fold3
mIoUumIoUkhIoUmIoUumIoUkhIoUmIoUumIoUkhIoUmIoUumIoUkhIoU
MaskCLIP29.723.726.323.725.724.623.925.724.723.425.824.5
MaskCLIP+3.628.56.33.029.25.44.829.28.24.529.97.8
CLIP-S432.0±0.829.4±0.330.6±0.522.3±0.932.8±0.426.5±0.622.4±0.532.1±0.526.4±0.428.6±0.831.5±0.230.0±0.5
vs. MaskCLIP++28.4+0.9+24.3+19.3+3.6+21.1+17.6+2.9+18.2+24.1+1.6+22.2
+ +![](images/b38239396bfa302ea59816bd846b1562552de090f3edcca588b1096df5ef28e8.jpg) +(a) Embeddings projection for PASCAL Context (left) and COCO-Stuff (right) + +![](images/9a82d32527abf1564a23ee49c7a1193fcd48a0af523e48705c30de0549f21374.jpg) +(b) Retrieved image segments for classes not used in training + +Figure 4. (a) Projection of pixel embeddings generated by CLIP, MaskCLIP+, and CLIP-S $^4$ trained on Pascal Context with CLIP's ViT-B/16 model (left) and COCO-Stuff with CLIP's ResNet50 model (right). MaskCLIP+ distorts pixel embeddings with respect to the given text embeddings, while CLIP-S $^4$ aligns pixel embeddings with the pre-trained CLIP model. (b) Image segments retrieved for classes that are not used during training. Correct and incorrect retrievals are outlined in green and orange, respectively. Compared with MaskCLIP+, CLIP-S $^4$ retrieves images that are more closely related to the classes. + +image segments retrieved from the COCO-Stuff validation set using MaskCLIP+ and CLIP-S $^4$ for a set of classes that are not used in training. For each class, we obtain its text embedding and compare it with segment embeddings to obtain top retrievals from both methods. Due to better alignment with the pre-trained CLIP model, CLIP-S $^4$ retrieves images that are more closely related to the unknown classes compared with MaskCLIP+. + +# 4.4. Unsupervised Semantic Segmentation + +To study whether $\mathrm{CLIP - S^4}$ can generate pixel embeddings that form distinctive clusters, we evaluate $\mathrm{CLIP - S^4}$ on the unsupervised semantic segmentation task for Pascal VOC 2012. To derive semantic segmentation from pixel embeddings, we use and test two approaches, including $k$ nearest neighbor ( $k$ -NN) search [19] and linear classification [40]. For $k$ -NN search, we assign each segment a class label by the majority vote of its nearest neighbors from the + +Table 3. Language-guided semantic segmentation benchmarks (mIoU) for unknown classes. The classes of Pascal Context are split into 4 folds with around 15 classes each fold. For each experiment, classes of one fold are considered as unknown. The performance of CLIP- $S^4$ is averaged over 5 runs with randomly initialized unknown class embeddings. CLIP- $S^4$ significantly outperforms $\text{MaskCLIP}+$ on unknown classes. Meanwhile, CLIP- $S^4$ archives consistent gains on known classes over $\text{MaskCLIP}$ and $\text{MaskCLIP}+$ , and hence leads to better overall performance. + +
MethodmIoU k-NNmIoU Linear Classifier
IIC [22]-28.0
SegSort [19]47.355.4
Hierarch. Group. [47]-48.8
MaskContrast (Sup.) [40]53.963.9
ConceptContrast [18]58.860.4
HSG [23]61.7-
MaskCLIP [49]67.369.5
MaskCLIP+ [49]65.170.0
CLIP-S472.0 (+4.7)73.0 (+3.0)
+ +Table 4. Unsupervised semantic segmentation benchmarks (mIoU) on Pascal VOC 2012. CLIP- $S^4$ consistently outperforms the state-of-the-art methods on both $k$ -NN search and linear classification. + +training set following [19]. For linear classification, we train a linear classifier on the learned pixel embeddings following [40]. + +We compare our method with the state-of-the-art unsupervised and language-guided semantic segmentation approaches. We train the state-of-the-art models using the same hyper-parameter setting as our approach except for IIC [22] and Hierarchical Grouping [47] as they use different training mechanisms. For comparison, we take the best results for IIC and Hierarchical Grouping. The benchmark results are shown in Tab. 4. With the vision-language guidance, our method achieves significant gains compared with the previous non-CLIP-based approaches (i.e., $+9\%$ for both $k$ -NN search and linear classification). Meanwhile, our method also outperforms the language-guided semantic segmentation approaches by a large margin. + +# 4.5. Instance Mask Tracking + +We evaluate the transferability of pixel embeddings learned from the Pascal VOC 2012 dataset. We use the pixel embeddings to track instance masks in the DAVIS 2017 validation set, where the instance masks at the first frame are given for each video. Following the prior work [48], we use the similarity between pixel embeddings cross frames to propagate the instance masks to the rest of the video frames. We evaluate the performance using the region similarity $\mathcal{I}$ + +
MethodJ(Mean)↑F(Mean)↑
MaskTrack-B [32]35.336.4
OSVOS-B [4]18.530.0
Video Colorization [42]34.632.7
CycleTime [43]41.939.4
mgPFF [24]42.246.9
Hierarchy. Group. [47]47.148.9
MaskContrast (Sup.) [40]34.336.7
ConceptContrast [18]50.453.9
MaskCLIP [49]48.149.2
MaskCLIP+ [49]42.644.2
CLIP-S452.3 (+1.9)56.8 (+2.9)
+ +Table 5. Quantitative evaluation of instance mask tracking on the DAVIS-2017 validation set. The performance is measured by the region similarity $\mathcal{J}$ (IoU) and the contour-based accuracy $\mathcal{F}$ defined by [33]. Our method outperforms existing supervised, unsupervised, and language-guided approaches on both metrics. + +
LtLeLspAccmIoUavgsim
✓--1.60.5-0.01
✓✓-48.124.30.79
✓-✓52.332.90.33
-✓✓48.631.3-
✓✓✓53.733.60.66
+ +(IoU) and the contour-based accuracy $\mathcal{F}$ defined by [33]. + +We compare our method with existing supervised [4, 32], unsupervised [18, 24, 40, 42, 43, 47], and language-guided [49] approaches (Tab. 5). Though not trained on any video sequences, our method outperforms the existing approaches by more than $1.9\%$ and $2.9\%$ in terms of the region similarity $\mathcal{I}$ and contour accuracy $\mathcal{F}$ , respectively. Note that the pixel embeddings generated by $\mathrm{MaskCLIP + }$ [49] are distorted by the classes from Pascal VOC 2012, which hinder their transferability. + +# 4.6. Ablation Study + +We study the contribution of different losses of our method using the Pascal Context dataset and the language-guided semantic segmentation task. The performance is evaluated with pixel accuracy (pAcc) and mIoU. We also calculate the average cosine similarity (avgsim) between our segment embeddings and CLIP's segment embeddings to quantify the alignment. Tab. 6 shows the study results. We observe that by introducing embedding consistent loss $\mathcal{L}_e$ the learned segment embeddings are well aligned with + +Table 6. Ablation study on the contribution of each loss of CLIP-S4. Experimented on language-guided semantic segmentation of Pascal Context. avgsim represents the average cosine similarity between segment embeddings generated by CLIP-S4 and CLIP. By combining the embedding and semantic consistent losses $\mathcal{L}_e$ and $\mathcal{L}_s$ , CLIP-S4 archives better semantic segmentation performance while maintaining the alignment with CLIP's embeddings. + +
#unknowns(u)mIoU
1671.6
3271.9
6472.0
12872.2
25671.9
+ +
top-m segmentsavgsim
10.921
40.977
160.996
640.997
2560.993
+ +Table 7. Ablation study on the influence of the number of unknown class prototypes. + +Table 8. Ablation study on different numbers of top- $m$ segments for class prototypes. + +CLIP's embeddings with an average cosine similarity of 0.79. However, the learned segment embeddings do not perform well on the language-guided semantic segmentation task (24.3 vs. 33.6), because the segment embeddings are not optimized to classify target classes. Meanwhile, by using semantic consistent loss $\mathcal{L}_s$ without embedding consistent loss, the learned segment embeddings have the discriminative power to classify different classes but are not aligned with CLIP's embeddings as the average cosine similarity is 0.33. As a result, the segment embeddings are limited to the target classes used during training. Hence, we combine $\mathcal{L}_e$ and $\mathcal{L}_s$ to balance the discriminative power over target classes and the alignment with CLIP. Meanwhile, we observe that with pixel-segment contrastive learning, the model can archive better performance. + +Also, we study the influence of the number of unknown class prototypes on the Pascal VOC dataset for the unsupervised semantic segmentation task. The results Tab. 7 show that the semantic segmentation performance is robust to the tested number of unknown class prototypes as the mIoU varies only $0.6\%$ . Furthermore, we investigate how the size of top- $m$ segments impacts the embeddings of class prototypes. We compare the embeddings of class prototypes generated with different numbers of top- $m$ segments on the Pascal VOC dataset. We use the embeddings of class prototypes generated with $m = 32$ segments as the reference, and compute the cosine similarity between the reference prototypes embeddings and ones with different top- $m$ segments. For each case, the cosine similarity is averaged over all class prototypes. We observe that the embeddings of class prototypes are relatively stable if moderate top- $m$ segments (e.g., $m = 32$ in this work) are used (Tab. 8). + +# 5. Conclusion + +We propose CLIP-S $^4$ , a novel pixel representation learning approach for semantic segmentation. Our method combines self-supervised contrastive learning and guidance of CLIP to learn consistent pixel embeddings with respect to visual and conceptual semantics. Our experiments on popular semantic segmentation benchmarks demonstrate consistent gains over the state-of-the-art unsupervised semantic segmentation and language-driven semantic segmentation methods, especially for unknown classes. + +# References + +[1] Radhakrishna Achanta, Appu Shaji, Kevin Smith, Aurelien Lucchi, Pascal Fua, and Sabine Süsstrunk. SLIC superpixels compared to state-of-the-art superpixel methods. IEEE TPAMI, 34(11):2274-2282, 2012. 3 +[2] Pablo Arbeláez, Michael Maire, Charless Fowlkes, and Jitendra Malik. Contour detection and hierarchical image segmentation. IEEE TPAMI, 33(5):898-916, 2011. 3, 6 +[3] Maxime Bucher, Tuan-Hung VU, Matthieu Cord, and Patrick Pérez. Zero-shot semantic segmentation. In NeurIPS, pages 468-479, 2019. 2 +[4] S. Caelles, K.-K. Maninis, J. Pont-Tuset, L. Leal-Taixe, D. Cremers, and L. Van Gool. One-shot video object segmentation. In CVPR, pages 5320-5329, 2017. 8 +[5] Holger Caesar, Jasper Uijlings, and Vittorio Ferrari. COCO-Stuff: Thing and stuff classes in context. In CVPR, pages 1209-1218, 2018. 5 +[6] Liang-Chieh Chen, George Papandreou, Iasonas Kokkinos, Kevin Murphy, and Alan L. Yuille. DeepLab: Semantic image segmentation with deep convolutional nets, atrous convolution, and fully connected CRFs. IEEE TPAMI, 40(4):834-848, 2018. 1 +[7] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In ICML, pages 1597-1607, 2020. 1, 6 +[8] Jang Hyun Cho, Utkarsh Mall, Kavita Bala, and Bharath Hariharan. PiCIE: Unsupervised semantic segmentation using invariance and equivariance in clustering. In CVPR, pages 16794-16804, 2021. 2 +[9] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. ImageNet: A large-scale hierarchical image database. In CVPR, pages 248–255, 2009. 6 +[10] Xiaoyi Dong, Yinglin Zheng, Jianmin Bao, Ting Zhang, Dongdong Chen, Hao Yang, Ming Zeng, Weiming Zhang, Lu Yuan, Dong Chen, Fang Wen, and Nenghai Yu. MaskCLIP: Masked self-distillation advances contrastive language-image pretraining. arXiv preprint arXiv:2208.12262, 2022.3 +[11] Sepideh Esmaeilpour, Bing Liu, Eric Robertson, and Lei Shu. Zero-shot out-of-distribution detection based on the pre-trained model CLIP. In AAAI, 2022. 3 +[12] Mark Everingham, Luc Van Gool, Christopher K. I. Williams, John Winn, and Andrew Zisserman. The pascal visual object classes (voc) challenge. *IJCV*, 88(2):303–338, 2010. 5 +[13] Xiuye Gu, Tsung-Yi Lin, Weicheng Kuo, and Yin Cui. Open-vocabulary object detection via vision and language knowledge distillation. In ICLR, 2022. 3, 4, 5, 6 +[14] Mark Hamilton, Zhoutong Zhang, Bharath Hariharan, Noah Snavely, and William T. Freeman. Unsupervised semantic segmentation by distilling feature correspondences. In ICLR, 2022. 1, 2 +[15] Bharath Hariharan, Pablo Arbeláez, Lubomir Bourdev, Subhransu Maji, and Jitendra Malik. Semantic contours from inverse detectors. In ICCV, pages 991-998, 2011. 5 + +[16] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In CVPR, pages 9726-9735, 2020. 1 +[17] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, pages 770-778, 2016. 6 +[18] Wenbin He, William Surmeier, Arvind Kumar Shekar, Liang Gou, and Liu Ren. Self-supervised semantic segmentation grounded in visual concepts. In IJCAI, pages 949-955, 2022. 1, 2, 3, 6, 7, 8 +[19] Jyh-Jing Hwang, Stella X. Yu, Jianbo Shi, Maxwell D. Collins, Tien-Ju Yang, Xiao Zhang, and Liang-Chieh Chen. SegSort: Segmentation by discriminative sorting of segments. In ICCV, pages 7333-7343, 2019. 1, 2, 5, 6, 7 +[20] Phillip Isola, Daniel Zoran, Dilip Krishnan, and Edward H. Adelson. Crisp boundary detection using pointwise mutual information. In ECCV, pages 799-814, 2014. 6 +[21] Ajay Jain, Matthew Tancik, and Pieter Abbeel. Putting NeRF on a diet: Semantically consistent few-shot view synthesis. In ICCV, pages 5865-5874, 2021. 3 +[22] Xu Ji, Andrea Vedaldi, and Joao Henriques. Invariant information clustering for unsupervised image classification and segmentation. In ICCV, pages 9864-9873, 2019. 2, 7 +[23] Tsung-Wei Ke, Jyh-Jing Hwang, Yunhui Guo, Xudong Wang, and Stella X. Yu. Unsupervised hierarchical semantic segmentation with multiview cosegmentation and clustering transformers. In CVPR, pages 2561–2571, 2022. 1, 2, 3, 6, 7 +[24] Shu Kong and Charless Fowlkes. Multigrid predictive filter flow for unsupervised learning on videos. arXiv preprint arXiv:1904.01693, 2019. 8 +[25] Boyi Li, Kilian Q Weinberger, Serge Belongie, Vladlen Koltun, and René Ranftl. Language-driven semantic segmentation. In ICLR, 2022. 1, 2, 3, 5 +[26] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dolkar, and C. Lawrence Zitnick. Microsoft COCO: Common objects in context. In ECCV, pages 740-755, 2014. 5 +[27] Jonathan Long, Evan Shelhamer, and Trevor Darrell. Fully convolutional networks for semantic segmentation. In CVPR, pages 3431-3440, 2015. 1, 2, 5 +[28] Luke Melas-Kyriazi, Christian Rupprecht, Iro Laina, and Andrea Vedaldi. Deep spectral methods: A surprisingly strong baseline for unsupervised semantic segmentation and localization. In CVPR, pages 8354-8365, 2022. 2 +[29] Roozbeh Mottaghi, Xianjie Chen, Xiaobai Liu, Nam-Gyu Cho, Seong-Whan Lee, Sanja Fidler, Raquel Urtasun, and Alan Yuille. The role of context for object detection and semantic segmentation in the wild. In CVPR, pages 891–898, 2014. 5 +[30] Yassine Ouali, Céline Hudelot, and Myriam Tami. Autoregressive unsupervised image segmentation. In ECCV, pages 142-158, 2020. 2 +[31] Dim P. Papadopoulos, Ethan Weber, and Antonio Torralba. Scaling up instance annotation via label propagation. In ICCV, pages 15344-15353, 2021. 1 + +[32] Federico Perazzi, Anna Khoreva, Rodrigo Benenson, Bernt Schiele, and Alexander Sorkine-Hornung. Learning video object segmentation from static images. In CVPR, pages 3491-3500, 2017. 8 +[33] F. Peruzzi, J. Pont-Tuset, B. McWilliams, L. Van Gool, M. Gross, and A. Sorkine-Hornung. A benchmark dataset and evaluation methodology for video object segmentation. In CVPR, pages 724-732, 2016. 6, 8 +[34] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In ICML, pages 8748-8763, 2021. 1, 2, 4, 5, 6 +[35] Yongming Rao, Wenliang Zhao, Guangyi Chen, Yansong Tang, Zheng Zhu, Guan Huang, Jie Zhou, and Jiwen Lu. DenseCLIP: Language-guided dense prediction with context-aware prompting. In CVPR, pages 18061-18070, 2022. 3 +[36] Ramprasaath R. Selvaraju, Karan Desai, Justin Johnson, and Nikhil Naik. CASTing your model: Learning to localize improves self-supervised representations. In CVPR, pages 11053-11062, 2021. 2 +[37] Gyungin Shin, Weidi Wie, and Samuel Albanie. ReCo: Retrieve and co-segment for zero-shot transfer. In NeurIPS, 2022. 1, 2, 3, 6 +[38] Kihyuk Sohn, David Berthelot, Chun-Liang Li, Zizhao Zhang, Nicholas Carlini, Ekin D. Cubuk, Alex Kurakin, Han Zhang, and Colin Raffel. FixMatch: Simplifying semi-supervised learning with consistency and confidence. In NeurIPS, pages 596-608, 2020. 4 +[39] Wouter Van Gansbeke, Simon Vandenhende, Stamatios Georgoulis, and Luc V Gool. Revisiting contrastive methods for unsupervised learning of visual representations. In NeurIPS, pages 16238-16250, 2021. 2 +[40] Wouter Van Gansbeke, Simon Vandenhende, Stamatios Georgoulis, and Luc Van Gool. Unsupervised semantic segmentation by contrasting object mask proposals. In ICCV, pages 10032-10042, 2021. 1, 2, 5, 7, 8 +[41] Yael Vinker, Ehsan Pajouheshgar, Jessica Y. Bo, Roman Christian Bachmann, Amit Haim Bermano, Daniel Cohen-Or, Amir Zamir, and Ariel Shamir. CLIPasso: Semantically-aware object sketching. ACM TOG, 41(4):86:1-86:11, 2022. 3 +[42] Carl Vondrick, Abhinav Shrivastava, Alireza Fathi, Sergio Guadarrama, and Kevin Murphy. Tracking emerges by colorizing videos. In ECCV, pages 402-419, 2018. 8 +[43] Xiaolong Wang, Allan Jabri, and Alexei A. Efros. Learning correspondence from the cycle-consistency of time. In CVPR, pages 2561-2571, 2019. 8 +[44] Zhaoqing Wang, Yu Lu, Qiang Li, Xunqiang Tao, Yandong Guo, Mingming Gong, and Tongliang Liu. CRIS: Clip-driven referring image segmentation. In CVPR, pages 11676-11685, 2022. 3 +[45] Saining Xie and Zhuowen Tu. Holistically-nested edge detection. In ICCV, pages 1395-1403, 2015. 6 +[46] Jiarui Xu, Shalini De Mello, Sifei Liu, Wonmin Byeon, Thomas Breuel, Jan Kautz, and Xiaolong Wang. GroupViT: + +Semantic segmentation emerges from text supervision. In CVPR, pages 18134-18144, 2022. 2, 3, 6 +[47] Xiao Zhang and Michael Maire. Self-supervised visual representation learning from hierarchical grouping. In NeurIPS, pages 16579-16590, 2020. 2, 6, 7, 8 +[48] Hengshuang Zhao, Jianping Shi, Xiaojuan Qi, Xiaogang Wang, and Jiaya Jia. Pyramid scene parsing network. In CVPR, pages 6230-6239, 2017. 1, 5, 6, 7 +[49] Chong Zhou, Chen Change Loy, and Bo Dai. Extract free dense labels from CLIP. In ECCV, 2022. 1, 2, 3, 4, 5, 6, 7, 8 \ No newline at end of file diff --git a/2023/CLIP-S4_ Language-Guided Self-Supervised Semantic Segmentation/images.zip b/2023/CLIP-S4_ Language-Guided Self-Supervised Semantic Segmentation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..79a9142490a17b3050031f93edc4d9636889861c --- /dev/null +++ b/2023/CLIP-S4_ Language-Guided Self-Supervised Semantic Segmentation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0f92175224f87484abe4320da53960247e64e822185eec3b183de20a455ac48e +size 437432 diff --git a/2023/CLIP-S4_ Language-Guided Self-Supervised Semantic Segmentation/layout.json b/2023/CLIP-S4_ Language-Guided Self-Supervised Semantic Segmentation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..0f8397422a40f1f96beae8c8b7d028abd0cb46dd --- /dev/null +++ b/2023/CLIP-S4_ Language-Guided Self-Supervised Semantic Segmentation/layout.json @@ -0,0 +1,9308 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 85, + 103, + 507, + 120 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 85, + 103, + 507, + 120 + ], + "spans": [ + { + "bbox": [ + 85, + 103, + 507, + 120 + ], + "type": "text", + "content": "CLIP-S" + }, + { + "bbox": [ + 85, + 103, + 507, + 120 + ], + "type": "inline_equation", + "content": "^4" + }, + { + "bbox": [ + 85, + 103, + 507, + 120 + ], + "type": "text", + "content": ": Language-Guided Self-Supervised Semantic Segmentation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 107, + 143, + 163, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 107, + 143, + 163, + 156 + ], + "spans": [ + { + "bbox": [ + 107, + 143, + 163, + 156 + ], + "type": "text", + "content": "Wenbin He" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 206, + 144, + 306, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 206, + 144, + 306, + 156 + ], + "spans": [ + { + "bbox": [ + 206, + 144, + 306, + 156 + ], + "type": "text", + "content": "Suphanut Jamonnak" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 348, + 144, + 402, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 348, + 144, + 402, + 156 + ], + "spans": [ + { + "bbox": [ + 348, + 144, + 402, + 156 + ], + "type": "text", + "content": "Liang Gou" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 445, + 144, + 486, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 445, + 144, + 486, + 156 + ], + "spans": [ + { + "bbox": [ + 445, + 144, + 486, + 156 + ], + "type": "text", + "content": "Liu Ren" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 97, + 158, + 494, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 97, + 158, + 494, + 171 + ], + "spans": [ + { + "bbox": [ + 97, + 158, + 494, + 171 + ], + "type": "text", + "content": "Bosch Research North America & Bosch Center for Artificial Intelligence (BCAI)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 122, + 174, + 468, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 122, + 174, + 468, + 185 + ], + "spans": [ + { + "bbox": [ + 122, + 174, + 468, + 185 + ], + "type": "text", + "content": "{wenbin.he2, suphanut.jamonnak, liang.gou, liu. ren}@us.bosch.com" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 143, + 213, + 191, + 224 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 213, + 191, + 224 + ], + "spans": [ + { + "bbox": [ + 143, + 213, + 191, + 224 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 239, + 290, + 550 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 239, + 290, + 550 + ], + "spans": [ + { + "bbox": [ + 46, + 239, + 290, + 550 + ], + "type": "text", + "content": "Existing semantic segmentation approaches are often limited by costly pixel-wise annotations and predefined classes. In this work, we present CLIP-" + }, + { + "bbox": [ + 46, + 239, + 290, + 550 + ], + "type": "inline_equation", + "content": "S^4" + }, + { + "bbox": [ + 46, + 239, + 290, + 550 + ], + "type": "text", + "content": " that leverages self-supervised pixel representation learning and vision-language models to enable various semantic segmentation tasks (e.g., unsupervised, transfer learning, language-driven segmentation) without any human annotations and unknown class information. We first learn pixel embeddings with pixel-segment contrastive learning from different augmented views of images. To further improve the pixel embeddings and enable language-driven semantic segmentation, we design two types of consistency guided by vision-language models: 1) embedding consistency, aligning our pixel embeddings to the joint feature space of a pre-trained vision-language model, CLIP [34]; and 2) semantic consistency, forcing our model to make the same predictions as CLIP over a set of carefully designed target classes with both known and unknown prototypes. Thus, CLIP-" + }, + { + "bbox": [ + 46, + 239, + 290, + 550 + ], + "type": "inline_equation", + "content": "S^4" + }, + { + "bbox": [ + 46, + 239, + 290, + 550 + ], + "type": "text", + "content": " enables a new task of class-free semantic segmentation where no unknown class information is needed during training. As a result, our approach shows consistent and substantial performance improvement over four popular benchmarks compared with the state-of-the-art unsupervised and language-driven semantic segmentation methods. More importantly, our method outperforms these methods on unknown class recognition by a large margin." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 573, + 127, + 585 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 573, + 127, + 585 + ], + "spans": [ + { + "bbox": [ + 47, + 573, + 127, + 585 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 594, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 287, + 713 + ], + "type": "text", + "content": "Semantic segmentation aims to partition an input image into semantically meaningful regions and assign each region a semantic class label. Recent advances in semantic segmentation [6, 27, 48] heavily rely on pixel-wise human annotations, which have two limitations. First, acquiring pixel-wise annotations is extremely labor intensive and costly, which can take up to 1.5 hours to label one image [31]. Second, human annotations are often limited to a set of predefined semantic classes, with which the learned models lack the ability to recognize unknown classes [25]." + } + ] + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 307, + 213, + 547, + 327 + ], + "blocks": [ + { + "bbox": [ + 307, + 213, + 547, + 327 + ], + "lines": [ + { + "bbox": [ + 307, + 213, + 547, + 327 + ], + "spans": [ + { + "bbox": [ + 307, + 213, + 547, + 327 + ], + "type": "image", + "image_path": "cadd67d538f7dadd7cd4817c2093ccbe18d4c78f72380ba46acf92e465f8d87a.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 311, + 327, + 541, + 334 + ], + "lines": [ + { + "bbox": [ + 311, + 327, + 541, + 334 + ], + "spans": [ + { + "bbox": [ + 311, + 327, + 541, + 334 + ], + "type": "text", + "content": "(b) Semantic segmentation generated by different CLIP-based unsupervised approaches" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 308, + 335, + 544, + 397 + ], + "blocks": [ + { + "bbox": [ + 308, + 335, + 544, + 397 + ], + "lines": [ + { + "bbox": [ + 308, + 335, + 544, + 397 + ], + "spans": [ + { + "bbox": [ + 308, + 335, + 544, + 397 + ], + "type": "image", + "image_path": "aec1cbba951d4512a5976fd04ee59e85632cabacb2b6ac368e140ff2bc1a30ad.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 316, + 397, + 535, + 404 + ], + "lines": [ + { + "bbox": [ + 316, + 397, + 535, + 404 + ], + "spans": [ + { + "bbox": [ + 316, + 397, + 535, + 404 + ], + "type": "text", + "content": "(c) CLIP-" + }, + { + "bbox": [ + 316, + 397, + 535, + 404 + ], + "type": "inline_equation", + "content": "S^4" + }, + { + "bbox": [ + 316, + 397, + 535, + 404 + ], + "type": "text", + "content": " aligns the pixel embeddings and their semantics with CLIP feature space" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 304, + 410, + 545, + 498 + ], + "lines": [ + { + "bbox": [ + 304, + 410, + 545, + 498 + ], + "spans": [ + { + "bbox": [ + 304, + 410, + 545, + 498 + ], + "type": "text", + "content": "Figure 1. (a) Pixel embeddings from different CLIP-based unsupervised methods: Our method, " + }, + { + "bbox": [ + 304, + 410, + 545, + 498 + ], + "type": "inline_equation", + "content": "\\mathrm{CLIP - S^4}" + }, + { + "bbox": [ + 304, + 410, + 545, + 498 + ], + "type": "text", + "content": ", generates sharper and more coherent pixel embeddings than MaskCLIP [49] and MaskCLIP+’s [49]; (b) Language-driven semantic segmentation by different methods: " + }, + { + "bbox": [ + 304, + 410, + 545, + 498 + ], + "type": "inline_equation", + "content": "\\mathrm{CLIP - S^4}" + }, + { + "bbox": [ + 304, + 410, + 545, + 498 + ], + "type": "text", + "content": " can recognize challenging unknown classes (e.g., moon); (c) The key idea behind " + }, + { + "bbox": [ + 304, + 410, + 545, + 498 + ], + "type": "inline_equation", + "content": "\\mathrm{CLIP - S^4}" + }, + { + "bbox": [ + 304, + 410, + 545, + 498 + ], + "type": "text", + "content": ": aligning the pixel embeddings and their semantics with CLIP feature space." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 506, + 546, + 686 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 506, + 546, + 686 + ], + "spans": [ + { + "bbox": [ + 304, + 506, + 546, + 686 + ], + "type": "text", + "content": "Various approaches have been proposed to tackle these limitations, among which we are inspired by two lines of recent research in particular. First, for unsupervised semantic segmentation (i.e., without human annotations), self-supervised pixel representation learning approaches [14,18, 19, 23, 40] have shown promising results on popular unsupervised benchmarks. The main idea is to extend self-supervised contrastive learning [7, 16] from images to pixels by attracting each pixel's embedding to its positive pairs and repelling it from negative pairs. The prior of pairs can be contours [18, 19], hierarchical groups [23], salience maps [40], and pre-trained models [14]. Although these approaches can group pixels into semantically meaningful clusters, human annotations are still needed to assign class labels to the clusters for semantic segmentation [37]." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "text", + "content": "Second, for unknown classes in semantic segmentation, large-scale vision-language models such as CLIP [34]" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 494, + 35 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 494, + 35 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 494, + 35 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "11207" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 276 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 276 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 276 + ], + "type": "text", + "content": "have shown great potential. This line of research, called language-driven semantic segmentation, aims to segment images with arbitrary classes defined by texts during testing time [25, 37, 46, 49]. Among these methods, most still need training time annotations, such as pixel annotations [25] and captions [46]. Only a few recent work, MaskCLIP & MaskCLIP+ [49] attempts to address this without using additional supervision: MaskCLIP directly extracts pixel embeddings correlated with texts from CLIP, but these pixel embeddings are coarse and noisy (Fig. 1a). To address this issue, MaskCLIP+ [49] trains a segmentation model on the pseudo-labels generated by MaskCLIP for a set of predefined classes. However, the pixel embeddings of MaskCLIP+ are distorted by the predefined classes (Fig. 1a), which limits its ability to recognize unknowns (Fig. 1b). Also, it needs unknown class information during training, which hinders its real-world applications." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 277, + 289, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 277, + 289, + 361 + ], + "spans": [ + { + "bbox": [ + 46, + 277, + 289, + 361 + ], + "type": "text", + "content": "We propose a language-guided self-supervised semantic segmentation approach, CLIP-S" + }, + { + "bbox": [ + 46, + 277, + 289, + 361 + ], + "type": "inline_equation", + "content": "^4" + }, + { + "bbox": [ + 46, + 277, + 289, + 361 + ], + "type": "text", + "content": ", which takes advantage of the strengths from both lines of research and addresses their limitations accordingly. The key idea is to learn consistent pixel embeddings with respect to visual and conceptual semantics using self-supervised learning and the guidance of a vision-language model, CLIP." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 361, + 289, + 589 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 361, + 289, + 589 + ], + "spans": [ + { + "bbox": [ + 46, + 361, + 289, + 589 + ], + "type": "text", + "content": "Specifically, we first train pixel embeddings with pixel-segment contrastive learning from different augmented image views [18, 19, 23] such that images can be partitioned into visually meaningful regions. To further improve pixel embedding quality and enable language-driven semantic segmentation, we introduce vision-language model guided consistency to regularize our model (Fig. 1c). The consistency is enforced from two aspects: embedding consistency and semantic consistency. First, embedding consistency aims to align the pixel embeddings generated by our model with the joint feature space of texts and images of CLIP by minimizing the distance between the pixel embeddings generated by our model and CLIP. Second, semantic consistency forces our model to make the same prediction as CLIP over a set of carefully designed target classes with both known and unknown prototypes. Note that unlike the previous methods [25,49] that use a predefined set of known classes, CLIP-" + }, + { + "bbox": [ + 46, + 361, + 289, + 589 + ], + "type": "inline_equation", + "content": "\\mathbf{S}^4" + }, + { + "bbox": [ + 46, + 361, + 289, + 589 + ], + "type": "text", + "content": " also learns the representation of unknown classes from images during training." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 590, + 288, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 590, + 288, + 662 + ], + "spans": [ + { + "bbox": [ + 46, + 590, + 288, + 662 + ], + "type": "text", + "content": "In the end, CLIP-" + }, + { + "bbox": [ + 46, + 590, + 288, + 662 + ], + "type": "inline_equation", + "content": "S^4" + }, + { + "bbox": [ + 46, + 590, + 288, + 662 + ], + "type": "text", + "content": " also enables a new task, namely class-free semantic segmentation, as shown in Tab. 1. This new task does not need any human annotations and even assumes NO class names are given during training. This is a more challenging task than the recent work [49] that requires class names of both known and unknown." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 59, + 662, + 287, + 674 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 662, + 287, + 674 + ], + "spans": [ + { + "bbox": [ + 59, + 662, + 287, + 674 + ], + "type": "text", + "content": "In summary, the contributions of this paper are threefold:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 58, + 677, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 677, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 58, + 677, + 287, + 713 + ], + "type": "text", + "content": "- We propose a self-supervised semantic segmentation approach that combines pixel-segment contrastive learning with the guidance of pre-trained vision lan" + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 306, + 70, + 547, + 192 + ], + "blocks": [ + { + "bbox": [ + 306, + 70, + 547, + 192 + ], + "lines": [ + { + "bbox": [ + 306, + 70, + 547, + 192 + ], + "spans": [ + { + "bbox": [ + 306, + 70, + 547, + 192 + ], + "type": "table", + "html": "
KnownUnknownAdd. Info.
Annot.CIs NameAnnot.CIs Name
Un/Self-supervised ([19] etc.)××××Fine-Tuning
Supervised ([27] etc.)✓✓N/AN/AN/A
Zero-shot ([3] etc.)✓✓×✓Word2Vec, etc.
Language-MaskCLIP+[49]×✓CLIP
DrivenCLIP-S4×✓×CLIP
", + "image_path": "af8866ba4f6554164fce2a7acceea2e7e2bc3902f6c29ca08449c06e1967cbe0.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 196, + 547, + 241 + ], + "lines": [ + { + "bbox": [ + 305, + 196, + 547, + 241 + ], + "spans": [ + { + "bbox": [ + 305, + 196, + 547, + 241 + ], + "type": "text", + "content": "Table 1. Comparison of information required for training over different tasks. CLIP-" + }, + { + "bbox": [ + 305, + 196, + 547, + 241 + ], + "type": "inline_equation", + "content": "\\mathbf{S}^4" + }, + { + "bbox": [ + 305, + 196, + 547, + 241 + ], + "type": "text", + "content": " enables a new task called class-free semantic segmentation. Compared with MaskCLIP+ [49], the new task assumes unknown class names are NOT given during training." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 324, + 246, + 545, + 282 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 324, + 246, + 545, + 282 + ], + "spans": [ + { + "bbox": [ + 324, + 246, + 545, + 282 + ], + "type": "text", + "content": "guage models. Our method can generate high-quality pixel embeddings without any human annotations and be applied to a variety of semantic segmentation tasks." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 317, + 283, + 545, + 452 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 317, + 283, + 545, + 378 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 283, + 545, + 378 + ], + "spans": [ + { + "bbox": [ + 317, + 283, + 545, + 378 + ], + "type": "text", + "content": "- We open up new research potentials for language-driven semantic segmentation without any human annotations by introducing and addressing a new task of class-free semantic segmentation (Tab. 1). Unlike previous work that assumes all the class names are known during training, our method can discover unknown classes from unlabelled image data without even knowing unknown class names." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 317, + 380, + 545, + 452 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 380, + 545, + 452 + ], + "spans": [ + { + "bbox": [ + 317, + 380, + 545, + 452 + ], + "type": "text", + "content": "- Consistent and substantial gains are observed with our approach over the state-of-the-art unsupervised and language-driven semantic segmentation methods on four popular datasets. More importantly, our method significantly outperforms the state-of-the-art on the segmentation of unknown classes." + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 465, + 392, + 478 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 465, + 392, + 478 + ], + "spans": [ + { + "bbox": [ + 306, + 465, + 392, + 478 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 486, + 545, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 486, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 304, + 486, + 545, + 689 + ], + "type": "text", + "content": "Unsupervised Semantic Segmentation. There are two groups of recent unsupervised semantic segmentation methods. One group of methods learns to generate consistent pixel representations or predictions between different augmentation of images with the guidance of mutual information [22, 30], clusters [8], contours [18, 19], hierarchical groups [23, 47], and saliency masks [40]. The other group of methods extracts dense features from pre-trained models based on saliency maps [36], augmentations [39], spectral decomposition [28], and feature correspondences [14]. While these methods can generate pixel embeddings with semantically meaningful clusters, annotations are needed to assign class labels to the clusters (e.g., " + }, + { + "bbox": [ + 304, + 486, + 545, + 689 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 486, + 545, + 689 + ], + "type": "text", + "content": "-nearest neighbor search [19] and Hungarian algorithm [40]). Our work combines pixel-segment self-supervision with pre-trained vision-language models to enable semantic segmentation without any human annotations." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 689, + 545, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 545, + 715 + ], + "type": "text", + "content": "Language-Driven Semantic Segmentation. Recently, vision-language models (e.g., CLIP [34]) trained on large-" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "11208" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 49, + 69, + 547, + 232 + ], + "blocks": [ + { + "bbox": [ + 49, + 69, + 547, + 232 + ], + "lines": [ + { + "bbox": [ + 49, + 69, + 547, + 232 + ], + "spans": [ + { + "bbox": [ + 49, + 69, + 547, + 232 + ], + "type": "image", + "image_path": "c70f39b4ee98291a5d3b8dfef69c8d525810cae5961988bd0bf15d23fa91c1f0.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 236, + 547, + 314 + ], + "lines": [ + { + "bbox": [ + 46, + 236, + 547, + 314 + ], + "spans": [ + { + "bbox": [ + 46, + 236, + 547, + 314 + ], + "type": "text", + "content": "Figure 2. Framework of CLIP-S4. (a) It starts with an encoder that maps images into pixel embeddings for semantic segmentation. Then it follows with two components: 1) self-supervised contrastive learning and 2) vision-language model guided consistency. Specifically, shown in (b), the self-supervised contrastive learning forces pixel embeddings to be consistent within visually coherent regions and among different augmented views of the same image. For vision-language model guided consistency, this framework introduces (c) embedding consistency that aligns the pixel embeddings generated by our model with CLIP embeddings, and (d) semantic consistency that forces our model to make the same predictions as CLIP for a set of target classes with both known and unknown prototypes. For known classes, the prototypes are pre-computed and fixed during training. For unknown classes, the prototypes are learned during training via clustering (e)." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 318, + 288, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 318, + 288, + 449 + ], + "spans": [ + { + "bbox": [ + 46, + 318, + 288, + 449 + ], + "type": "text", + "content": "scale image-text datasets have shown great potential on various downstream tasks such as image synthesis [21, 41], out-of-distribution detection [11], and object detection [13]. To extend vision-language models for semantic segmentation, one active research, language-driven semantic segmentation, aims to segment images with arbitrary unknown classes defined by texts during testing time [25, 37, 46, 49]. Some methods [25, 44] use pixel-wise annotations to train language-guided semantic segmentation models. Other methods [10, 46] perform large-scale pre-training on image-text pairs specifically for semantic segmentation." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 451, + 287, + 630 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 451, + 287, + 630 + ], + "spans": [ + { + "bbox": [ + 46, + 451, + 287, + 630 + ], + "type": "text", + "content": "By contrast, we directly use vision-language models that are pre-trained for classification tasks. Along this line of research, a few approaches have been proposed [35, 37, 49]. The most relevant approach to our method is MaskCLIP [49], which extends the embeddings generated by pre-trained vision-language models from image to pixel level, but these embeddings are often coarse and noisy. To address this issue, MaskCLIP + [49] fine-tunes the pixel embeddings by the pseudo-labels of a specific set of classes on top of MaskCLIP. However, it needs unknown class names during training, which may not be possible in real-world cases. Compared with [49], our method can recognize unknown classes without knowing any unknown class information during training time, and also learns fine-grained and sharper pixel embeddings with self-supervision." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 644, + 103, + 656 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 644, + 103, + 656 + ], + "spans": [ + { + "bbox": [ + 47, + 644, + 103, + 656 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 665, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 288, + 714 + ], + "type": "text", + "content": "Our method (Fig. 2) segments images by learning a pixel embedding function with self-supervised contrastive learning and the guidance of a pre-trained vision-language model, CLIP. We use self-supervised contrastive learning" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 318, + 547, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 318, + 547, + 486 + ], + "spans": [ + { + "bbox": [ + 304, + 318, + 547, + 486 + ], + "type": "text", + "content": "to force the consistency of pixel embeddings within visually coherent regions (e.g., superpixels) and among different augmented views of the same image (Sec. 3.1). We also introduce two vision-language model guided consistency (i.e., embedding consistency and semantic consistency) to further regularize the model (Sec. 3.2). The two components are complementary to each other. On the one hand, contrastive learning mitigates the noise introduced by CLIP. On the other hand, with the knowledge extracted from CLIP, the quality of the pixel embeddings can be improved. More importantly, this approach enables us to perform language-driven semantic segmentation with our carefully designed target class prototypes of both knowns and unknowns. In the following, we discuss the two components in detail." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 491, + 499, + 504 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 491, + 499, + 504 + ], + "spans": [ + { + "bbox": [ + 306, + 491, + 499, + 504 + ], + "type": "text", + "content": "3.1. Pixel-Segment Contrastive Learning" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 510, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 510, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 510, + 545, + 714 + ], + "type": "text", + "content": "We train a pixel embedding function to generate consistent pixel embeddings within visually coherent regions through pixel-segment contrastive learning [18,23]. Specifically, the embedding function transforms each pixel " + }, + { + "bbox": [ + 304, + 510, + 545, + 714 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 304, + 510, + 545, + 714 + ], + "type": "text", + "content": " of an image to a unit-length embedding vector " + }, + { + "bbox": [ + 304, + 510, + 545, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_p" + }, + { + "bbox": [ + 304, + 510, + 545, + 714 + ], + "type": "text", + "content": " of dimension " + }, + { + "bbox": [ + 304, + 510, + 545, + 714 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 304, + 510, + 545, + 714 + ], + "type": "text", + "content": " via a deep neural network. The image is then partitioned into " + }, + { + "bbox": [ + 304, + 510, + 545, + 714 + ], + "type": "inline_equation", + "content": "|S|" + }, + { + "bbox": [ + 304, + 510, + 545, + 714 + ], + "type": "text", + "content": " segments by clustering the pixel embeddings. The embedding " + }, + { + "bbox": [ + 304, + 510, + 545, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_s" + }, + { + "bbox": [ + 304, + 510, + 545, + 714 + ], + "type": "text", + "content": " of each segment " + }, + { + "bbox": [ + 304, + 510, + 545, + 714 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 304, + 510, + 545, + 714 + ], + "type": "text", + "content": " is calculated as the average of the pixel embeddings " + }, + { + "bbox": [ + 304, + 510, + 545, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_s = \\sum_{p\\in s}\\mathbf{z}_p / |s|" + }, + { + "bbox": [ + 304, + 510, + 545, + 714 + ], + "type": "text", + "content": ", which is also normalized into a unit-length vector " + }, + { + "bbox": [ + 304, + 510, + 545, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_s = \\mathbf{v}_s / \\| \\mathbf{v}_s\\|" + }, + { + "bbox": [ + 304, + 510, + 545, + 714 + ], + "type": "text", + "content": ". For each pixel " + }, + { + "bbox": [ + 304, + 510, + 545, + 714 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 304, + 510, + 545, + 714 + ], + "type": "text", + "content": ", the segments are grouped into two sets including a positive set " + }, + { + "bbox": [ + 304, + 510, + 545, + 714 + ], + "type": "inline_equation", + "content": "S^+" + }, + { + "bbox": [ + 304, + 510, + 545, + 714 + ], + "type": "text", + "content": " and a negative set " + }, + { + "bbox": [ + 304, + 510, + 545, + 714 + ], + "type": "inline_equation", + "content": "S^-" + }, + { + "bbox": [ + 304, + 510, + 545, + 714 + ], + "type": "text", + "content": ". The positive set " + }, + { + "bbox": [ + 304, + 510, + 545, + 714 + ], + "type": "inline_equation", + "content": "S^+" + }, + { + "bbox": [ + 304, + 510, + 545, + 714 + ], + "type": "text", + "content": " of a pixel contains segments within the same visually coherent region of the pixel. Following the prior work [18, 23], the visually coherent region can be derived from super-pixels [1] or contours [2]. We also use data augmentation (e.g., random cropping and color jitter" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "11209" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 178 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 178 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 178 + ], + "type": "text", + "content": "ing) to generate consistent pixel embeddings between different augmented views of the same image. Hence, segments within the same region of the pixel in any augmented views are considered as the positive set " + }, + { + "bbox": [ + 46, + 72, + 289, + 178 + ], + "type": "inline_equation", + "content": "S^{+}" + }, + { + "bbox": [ + 46, + 72, + 289, + 178 + ], + "type": "text", + "content": ". Other segments in the image and segments from other images in the same batch are included in the negative set " + }, + { + "bbox": [ + 46, + 72, + 289, + 178 + ], + "type": "inline_equation", + "content": "S^{-}" + }, + { + "bbox": [ + 46, + 72, + 289, + 178 + ], + "type": "text", + "content": ". The pixel embedding " + }, + { + "bbox": [ + 46, + 72, + 289, + 178 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_p" + }, + { + "bbox": [ + 46, + 72, + 289, + 178 + ], + "type": "text", + "content": " is then attracted to the segments in positive set " + }, + { + "bbox": [ + 46, + 72, + 289, + 178 + ], + "type": "inline_equation", + "content": "S^{+}" + }, + { + "bbox": [ + 46, + 72, + 289, + 178 + ], + "type": "text", + "content": " and repelled from the segments in negative set " + }, + { + "bbox": [ + 46, + 72, + 289, + 178 + ], + "type": "inline_equation", + "content": "S^{-}" + }, + { + "bbox": [ + 46, + 72, + 289, + 178 + ], + "type": "text", + "content": " with contrastive loss:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 70, + 188, + 287, + 217 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 188, + 287, + 217 + ], + "spans": [ + { + "bbox": [ + 70, + 188, + 287, + 217 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {t} (p) = - \\log \\frac {\\sum_ {s \\in \\mathcal {S} ^ {+}} e x p \\left(\\operatorname {s i m} \\left(\\mathbf {z} _ {p} , \\mathbf {v} _ {s}\\right) \\kappa\\right)}{\\sum_ {s \\in \\mathcal {S} ^ {+} \\cup \\mathcal {S} ^ {-}} e x p \\left(\\operatorname {s i m} \\left(\\mathbf {z} _ {p} , \\mathbf {v} _ {s}\\right) \\kappa\\right)}, \\tag {1}", + "image_path": "54258d97b82d8334095e917ca4ff3fb85c20c329f491c4de5a88a13f3b764f29.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 226, + 287, + 262 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 226, + 287, + 262 + ], + "spans": [ + { + "bbox": [ + 46, + 226, + 287, + 262 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 226, + 287, + 262 + ], + "type": "inline_equation", + "content": "\\kappa" + }, + { + "bbox": [ + 46, + 226, + 287, + 262 + ], + "type": "text", + "content": " is the concentration constant and " + }, + { + "bbox": [ + 46, + 226, + 287, + 262 + ], + "type": "inline_equation", + "content": "sim(\\mathbf{z}_p, \\mathbf{v}_s)" + }, + { + "bbox": [ + 46, + 226, + 287, + 262 + ], + "type": "text", + "content": " is the cosine similarity between the pixel embedding " + }, + { + "bbox": [ + 46, + 226, + 287, + 262 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_p" + }, + { + "bbox": [ + 46, + 226, + 287, + 262 + ], + "type": "text", + "content": " and the segment embedding " + }, + { + "bbox": [ + 46, + 226, + 287, + 262 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_s" + }, + { + "bbox": [ + 46, + 226, + 287, + 262 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 273, + 278, + 285 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 273, + 278, + 285 + ], + "spans": [ + { + "bbox": [ + 47, + 273, + 278, + 285 + ], + "type": "text", + "content": "3.2. Vision-Language Model Guided Consistency" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 292, + 287, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 292, + 287, + 399 + ], + "spans": [ + { + "bbox": [ + 46, + 292, + 287, + 399 + ], + "type": "text", + "content": "To enable language-driven semantic segmentation and improve the quality of pixel embeddings, we use a pretrained vision-language model such as CLIP [34] to guide the training of the pixel embedding function. The key idea is to align the output space of our pixel embedding function consistent with the feature space of CLIP (Fig. 1c). Specifically, two types of consistency are considered during training including embedding consistency and semantic consistency, which are detailed as follows." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 401, + 287, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 401, + 287, + 460 + ], + "spans": [ + { + "bbox": [ + 46, + 401, + 287, + 460 + ], + "type": "text", + "content": "**Embedding Consistency.** Our goal is to align the pixel embeddings generated from our self-supervised method (the green contour in Fig. 1c) with CLIP's pixel embeddings (the orange contour in Fig. 1c). This is done by minimizing the distance between the two pixel embedding spaces." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 461, + 287, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 461, + 287, + 557 + ], + "spans": [ + { + "bbox": [ + 46, + 461, + 287, + 557 + ], + "type": "text", + "content": "We first obtain the pixel embeddings of an input image from CLIP by modifying the attention-based pooling layer of the CLIP image encoder following [49]. Specifically, we 1) remove the query and key projection layers and 2) reformulate the value projection layer and the last linear layer as two consecutive fully connected layers. In the following, we use " + }, + { + "bbox": [ + 46, + 461, + 287, + 557 + ], + "type": "inline_equation", + "content": "clip - i(\\cdot)" + }, + { + "bbox": [ + 46, + 461, + 287, + 557 + ], + "type": "text", + "content": " as the modified CLIP image encoder and " + }, + { + "bbox": [ + 46, + 461, + 287, + 557 + ], + "type": "inline_equation", + "content": "clip - t(\\cdot)" + }, + { + "bbox": [ + 46, + 461, + 287, + 557 + ], + "type": "text", + "content": " as CLIP text encoder." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 558, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 558, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 558, + 288, + 714 + ], + "type": "text", + "content": "Then we obtain the pixel embeddings of CLIP for different augmented views of the image. Note that we use the original image to generate the CLIP pixel embeddings and perform augmentation afterwards to make sure that the CLIP pixel embeddings are consistent among different augmented views. In the end, we minimize the distance of embeddings between segments instead of pixels from our self-supervised and CLIP embedding spaces. This is because the pixel embeddings of CLIP are noisy (Fig. 2), which can be mitigated by aggregating over segments. Hence, we use the pixel embeddings generated by our model to derive segments (clusters) and then apply them to the CLIP's pixel embeddings. In the end, for each segment " + }, + { + "bbox": [ + 46, + 558, + 288, + 714 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 46, + 558, + 288, + 714 + ], + "type": "text", + "content": ", the embedding" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 305, + 69, + 547, + 239 + ], + "blocks": [ + { + "bbox": [ + 305, + 69, + 547, + 239 + ], + "lines": [ + { + "bbox": [ + 305, + 69, + 547, + 239 + ], + "spans": [ + { + "bbox": [ + 305, + 69, + 547, + 239 + ], + "type": "image", + "image_path": "e1c79c194e1189f50624c1c920fb3868b965be53cee49adf0b56be8721db3348.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 243, + 547, + 354 + ], + "lines": [ + { + "bbox": [ + 304, + 243, + 547, + 354 + ], + "spans": [ + { + "bbox": [ + 304, + 243, + 547, + 354 + ], + "type": "text", + "content": "Figure 3. Computation of target class prototypes with both knowns and unknowns, " + }, + { + "bbox": [ + 304, + 243, + 547, + 354 + ], + "type": "inline_equation", + "content": "C = \\{C_k, C_u\\}" + }, + { + "bbox": [ + 304, + 243, + 547, + 354 + ], + "type": "text", + "content": ". For a set of known classes (e.g., bird, cat), we first obtain their CLIP text embeddings, " + }, + { + "bbox": [ + 304, + 243, + 547, + 354 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 304, + 243, + 547, + 354 + ], + "type": "text", + "content": ", via a set of prompt templates [13,49] (shown in " + }, + { + "bbox": [ + 304, + 243, + 547, + 354 + ], + "type": "inline_equation", + "content": "\\widehat{\\mathbf{a}}" + }, + { + "bbox": [ + 304, + 243, + 547, + 354 + ], + "type": "text", + "content": "); then, we calculate the normalized (via softmax) similarity between text embeddings, " + }, + { + "bbox": [ + 304, + 243, + 547, + 354 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 304, + 243, + 547, + 354 + ], + "type": "text", + "content": ", and all segments' CLIP embeddings, " + }, + { + "bbox": [ + 304, + 243, + 547, + 354 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 304, + 243, + 547, + 354 + ], + "type": "text", + "content": ", from training images, and average the top-" + }, + { + "bbox": [ + 304, + 243, + 547, + 354 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 304, + 243, + 547, + 354 + ], + "type": "text", + "content": " similar segments' CLIP embedding as the embedding prototype for this class, as shown in " + }, + { + "bbox": [ + 304, + 243, + 547, + 354 + ], + "type": "inline_equation", + "content": "\\widehat{\\mathbf{b}}" + }, + { + "bbox": [ + 304, + 243, + 547, + 354 + ], + "type": "text", + "content": "; For each unknown class, we randomly select the CLIP embedding of a segment as the initial prototype (shown in " + }, + { + "bbox": [ + 304, + 243, + 547, + 354 + ], + "type": "inline_equation", + "content": "\\widehat{\\mathbf{c}}" + }, + { + "bbox": [ + 304, + 243, + 547, + 354 + ], + "type": "text", + "content": ")." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 361, + 422, + 372 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 361, + 422, + 372 + ], + "spans": [ + { + "bbox": [ + 306, + 361, + 422, + 372 + ], + "type": "text", + "content": "consistent loss is defined as:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 372, + 385, + 545, + 398 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 372, + 385, + 545, + 398 + ], + "spans": [ + { + "bbox": [ + 372, + 385, + 545, + 398 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {e} (s) = 1 - \\operatorname {s i m} \\left(\\mathbf {v} _ {s}, \\mathbf {i} _ {s}\\right), \\tag {2}", + "image_path": "f4f3fa7101a0cb7de3b26806205a2955af94429780dcbeb9ba8e750721a22576.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 410, + 545, + 459 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 410, + 545, + 459 + ], + "spans": [ + { + "bbox": [ + 304, + 410, + 545, + 459 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 410, + 545, + 459 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_s" + }, + { + "bbox": [ + 304, + 410, + 545, + 459 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 410, + 545, + 459 + ], + "type": "inline_equation", + "content": "\\mathbf{i}_s" + }, + { + "bbox": [ + 304, + 410, + 545, + 459 + ], + "type": "text", + "content": " are the segment embeddings derived from our embedding function and CLIP, respectively. Here, " + }, + { + "bbox": [ + 304, + 410, + 545, + 459 + ], + "type": "inline_equation", + "content": "\\mathbf{i}_s" + }, + { + "bbox": [ + 304, + 410, + 545, + 459 + ], + "type": "text", + "content": " is the average of the CLIP pixel embedding from segment " + }, + { + "bbox": [ + 304, + 410, + 545, + 459 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 304, + 410, + 545, + 459 + ], + "type": "text", + "content": ", namely, " + }, + { + "bbox": [ + 304, + 410, + 545, + 459 + ], + "type": "inline_equation", + "content": "\\mathbf{i}_s = \\sum_{p \\in s} \\text{clip-}i(s) / |s|" + }, + { + "bbox": [ + 304, + 410, + 545, + 459 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 460, + 545, + 543 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 460, + 545, + 543 + ], + "spans": [ + { + "bbox": [ + 304, + 460, + 545, + 543 + ], + "type": "text", + "content": "Semantic Consistency In addition to embedding consistency, we introduce semantic consistency by forcing our model to make the same predictions of semantic classes as CLIP. The rationale is that we can generate better pixel embeddings if they can form distinctive clusters corresponding to different semantic classes, as the goal of semantic segmentation is to perform pixel-wise classification." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 544, + 546, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 544, + 546, + 689 + ], + "spans": [ + { + "bbox": [ + 304, + 544, + 546, + 689 + ], + "type": "text", + "content": "Semantic consistency is achieved via a similar idea of pseudo-labeling [38]. Again, we force the semantic consistency at the segment level (not the pixel level) to reduce the noise in pseudo-labels. Specifically, for each segment " + }, + { + "bbox": [ + 304, + 544, + 546, + 689 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 304, + 544, + 546, + 689 + ], + "type": "text", + "content": ", we first use CLIP to generate its pseudo-label " + }, + { + "bbox": [ + 304, + 544, + 546, + 689 + ], + "type": "inline_equation", + "content": "y_{s}" + }, + { + "bbox": [ + 304, + 544, + 546, + 689 + ], + "type": "text", + "content": " over a set of target classes, which include both knowns and unknowns (we will introduce how to design these target classes later). The pseudo-label is generated based on the highest similarity between the segment embedding " + }, + { + "bbox": [ + 304, + 544, + 546, + 689 + ], + "type": "inline_equation", + "content": "\\mathbf{i}_s" + }, + { + "bbox": [ + 304, + 544, + 546, + 689 + ], + "type": "text", + "content": " with a set of prototypes, " + }, + { + "bbox": [ + 304, + 544, + 546, + 689 + ], + "type": "inline_equation", + "content": "C = \\{\\mathbf{c}_l\\}_{0}^{L - 1}" + }, + { + "bbox": [ + 304, + 544, + 546, + 689 + ], + "type": "text", + "content": ", of the target classes in the pixel embedding space of CLIP, namely, " + }, + { + "bbox": [ + 304, + 544, + 546, + 689 + ], + "type": "inline_equation", + "content": "y_{s} = \\operatorname{argmax}_{l\\in L}(sim(\\mathbf{i}_{s},\\mathbf{c}_{l}))" + }, + { + "bbox": [ + 304, + 544, + 546, + 689 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 689, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 689, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 305, + 689, + 545, + 714 + ], + "type": "text", + "content": "Then we define the semantic consistent loss as the cross entropy between our model's prediction " + }, + { + "bbox": [ + 305, + 689, + 545, + 714 + ], + "type": "inline_equation", + "content": "\\varphi (\\mathbf{v}_s)" + }, + { + "bbox": [ + 305, + 689, + 545, + 714 + ], + "type": "text", + "content": " over the tar" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "11210" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 192, + 84 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 192, + 84 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 192, + 84 + ], + "type": "text", + "content": "get classes and the pseudo-label " + }, + { + "bbox": [ + 47, + 72, + 192, + 84 + ], + "type": "inline_equation", + "content": "y_{s}" + }, + { + "bbox": [ + 47, + 72, + 192, + 84 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 118, + 93, + 287, + 107 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 118, + 93, + 287, + 107 + ], + "spans": [ + { + "bbox": [ + 118, + 93, + 287, + 107 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {s} (s) = \\mathbf {H} \\left(y _ {s}, \\varphi (\\mathbf {v} _ {s})\\right), \\tag {3}", + "image_path": "51730bc36e8c924bc939956f39f09351d905b3e8c38398859638e9941639b8f8.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 115, + 212, + 128 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 115, + 212, + 128 + ], + "spans": [ + { + "bbox": [ + 47, + 115, + 212, + 128 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 115, + 212, + 128 + ], + "type": "inline_equation", + "content": "\\varphi (\\mathbf{v}_s) = \\mathrm{softmax}(sim(\\mathbf{v}_s,C))" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 128, + 287, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 128, + 287, + 319 + ], + "spans": [ + { + "bbox": [ + 46, + 128, + 287, + 319 + ], + "type": "text", + "content": "Target Class Prototypes The design of target classes and associated class prototypes, " + }, + { + "bbox": [ + 46, + 128, + 287, + 319 + ], + "type": "inline_equation", + "content": "C = \\{\\mathbf{c}_l\\}_{0}^{L - 1}" + }, + { + "bbox": [ + 46, + 128, + 287, + 319 + ], + "type": "text", + "content": ", is crucial to achieve the semantic consistency. Here, a class prototype, " + }, + { + "bbox": [ + 46, + 128, + 287, + 319 + ], + "type": "inline_equation", + "content": "\\mathbf{c}_l" + }, + { + "bbox": [ + 46, + 128, + 287, + 319 + ], + "type": "text", + "content": ", is an embedding vector that can represent a class in an embedding space. For example, it can be the mean vector of embeddings of all segments of a class \"car\". Currently, most existing methods [25, 49] assume that the target classes are already predefined, which is not feasible in real-world use cases without any human annotations. Thus, those methods cannot handle unknown classes hidden in the data. To address this issue, we introduce two sets of class prototypes of known, " + }, + { + "bbox": [ + 46, + 128, + 287, + 319 + ], + "type": "inline_equation", + "content": "C_k = \\{\\mathbf{c}_0,\\dots ,\\mathbf{c}_{k - 1}\\}" + }, + { + "bbox": [ + 46, + 128, + 287, + 319 + ], + "type": "text", + "content": ", and unknown classes, " + }, + { + "bbox": [ + 46, + 128, + 287, + 319 + ], + "type": "inline_equation", + "content": "C_u = \\{\\mathbf{c}_k,\\dots ,\\mathbf{c}_{k + u}\\}" + }, + { + "bbox": [ + 46, + 128, + 287, + 319 + ], + "type": "text", + "content": ", where the known classes are predefined by leveraging CLIP and the unknown classes are learned from image data during training. Thus, we have " + }, + { + "bbox": [ + 46, + 128, + 287, + 319 + ], + "type": "inline_equation", + "content": "C = \\{\\mathbf{c}_l\\}_{0}^{L - 1} = \\{\\mathbf{c}_0,\\dots ,\\mathbf{c}_{k - 1},\\mathbf{c}_k,\\dots ,\\mathbf{c}_{k + u}\\}" + }, + { + "bbox": [ + 46, + 128, + 287, + 319 + ], + "type": "inline_equation", + "content": "L = k + u" + }, + { + "bbox": [ + 46, + 128, + 287, + 319 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 319, + 287, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 319, + 287, + 437 + ], + "spans": [ + { + "bbox": [ + 46, + 319, + 287, + 437 + ], + "type": "text", + "content": "For known classes, a natural choice is to use the text embeddings generated by CLIP as their class prototype embeddings [25, 49]. However, even though the text embeddings are trained to align with image/pixel embeddings [34], there is still a huge gap between the text and image/pixel embeddings in the joint space of CLIP (Fig. 1c). Therefore, it is challenging to learn meaningful unknown classes from image features when using text embeddings as class prototypes. Hence, in this work, we use the prototype of CLIP pixel embeddings to represent each known class." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 438, + 287, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 438, + 287, + 629 + ], + "spans": [ + { + "bbox": [ + 46, + 438, + 287, + 629 + ], + "type": "text", + "content": "To this end, for a set of known classes (e.g., bird, cat), " + }, + { + "bbox": [ + 46, + 438, + 287, + 629 + ], + "type": "inline_equation", + "content": "K = \\{0,\\dots ,k - 1\\}" + }, + { + "bbox": [ + 46, + 438, + 287, + 629 + ], + "type": "text", + "content": " we first obtain their CLIP text embeddings, " + }, + { + "bbox": [ + 46, + 438, + 287, + 629 + ], + "type": "inline_equation", + "content": "T = \\{\\mathbf{t}_k\\} = \\{\\text{clip - } t(k)\\}" + }, + { + "bbox": [ + 46, + 438, + 287, + 629 + ], + "type": "text", + "content": ", via a set of prompt templates following [13, 49], as shown in Fig. 3a. We also get a set of CLIP segment embeddings, " + }, + { + "bbox": [ + 46, + 438, + 287, + 629 + ], + "type": "inline_equation", + "content": "I = \\{\\mathbf{i}_{\\hat{s}}\\}" + }, + { + "bbox": [ + 46, + 438, + 287, + 629 + ], + "type": "text", + "content": ", for all training images by a) feeding training images into the modified image encoder of CLIP to get pixel embeddings; b) clustering the pixel embeddings as segments, " + }, + { + "bbox": [ + 46, + 438, + 287, + 629 + ], + "type": "inline_equation", + "content": "\\hat{S}" + }, + { + "bbox": [ + 46, + 438, + 287, + 629 + ], + "type": "text", + "content": "; c) averaging the pixel embeddings in each segment, " + }, + { + "bbox": [ + 46, + 438, + 287, + 629 + ], + "type": "inline_equation", + "content": "\\hat{s}" + }, + { + "bbox": [ + 46, + 438, + 287, + 629 + ], + "type": "text", + "content": ". Hence, we have embeddings for each segment: " + }, + { + "bbox": [ + 46, + 438, + 287, + 629 + ], + "type": "inline_equation", + "content": "\\mathbf{i}_{\\hat{s}} = \\sum_{p\\in \\hat{s}}\\text{clip - } i(p) / |\\hat{s}|" + }, + { + "bbox": [ + 46, + 438, + 287, + 629 + ], + "type": "text", + "content": ". Then, we calculate the similarity between text embeddings of known classes, " + }, + { + "bbox": [ + 46, + 438, + 287, + 629 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 46, + 438, + 287, + 629 + ], + "type": "text", + "content": ", and all CLIP segment embeddings " + }, + { + "bbox": [ + 46, + 438, + 287, + 629 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 46, + 438, + 287, + 629 + ], + "type": "text", + "content": ", and normalize the similarities over all classes by softmax. Finally, we average the top- " + }, + { + "bbox": [ + 46, + 438, + 287, + 629 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 46, + 438, + 287, + 629 + ], + "type": "text", + "content": " similar segments' embedding as the embedding prototype for each class, " + }, + { + "bbox": [ + 46, + 438, + 287, + 629 + ], + "type": "inline_equation", + "content": "C_k = \\{\\mathbf{c}_k\\} = avg_m(top-m_{\\hat{s}}(\\mathbf{softmax}_k(sim(I,T))))" + }, + { + "bbox": [ + 46, + 438, + 287, + 629 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 630, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 630, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 630, + 287, + 714 + ], + "type": "text", + "content": "The prototype embeddings of the unknown classes, " + }, + { + "bbox": [ + 46, + 630, + 287, + 714 + ], + "type": "inline_equation", + "content": "C_u" + }, + { + "bbox": [ + 46, + 630, + 287, + 714 + ], + "type": "text", + "content": ", are randomly initialized by sampling the CLIP embeddings of all segments, namely " + }, + { + "bbox": [ + 46, + 630, + 287, + 714 + ], + "type": "inline_equation", + "content": "C_u = \\text{random}(clip - i(\\hat{\\mathcal{S}}), u)" + }, + { + "bbox": [ + 46, + 630, + 287, + 714 + ], + "type": "text", + "content": ", with a size of the unknown class of " + }, + { + "bbox": [ + 46, + 630, + 287, + 714 + ], + "type": "inline_equation", + "content": "u" + }, + { + "bbox": [ + 46, + 630, + 287, + 714 + ], + "type": "text", + "content": " (Fig. 3c). During training, the embedding " + }, + { + "bbox": [ + 46, + 630, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{c}_u" + }, + { + "bbox": [ + 46, + 630, + 287, + 714 + ], + "type": "text", + "content": " of each unknown class prototype is updated by minimizing its distance to all segments that are classified as this unknown class (similar to updating the cen" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 72, + 427, + 84 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 427, + 84 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 427, + 84 + ], + "type": "text", + "content": "troids in " + }, + { + "bbox": [ + 305, + 72, + 427, + 84 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 305, + 72, + 427, + 84 + ], + "type": "text", + "content": " -means clustering):" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 339, + 93, + 545, + 119 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 339, + 93, + 545, + 119 + ], + "spans": [ + { + "bbox": [ + 339, + 93, + 545, + 119 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {u} = \\sum_ {s \\in S _ {u}} (1 - \\operatorname {s i m} (\\mathbf {c} _ {u}, \\operatorname {c l i p} - i (s))) / | S _ {u} |, \\tag {4}", + "image_path": "5976be124c22b02e66fb2e637eedef29d85f8813935aecded323bc96726fb3fe.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 129, + 545, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 129, + 545, + 163 + ], + "spans": [ + { + "bbox": [ + 304, + 129, + 545, + 163 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 129, + 545, + 163 + ], + "type": "inline_equation", + "content": "S_{u}" + }, + { + "bbox": [ + 304, + 129, + 545, + 163 + ], + "type": "text", + "content": " are the segments classified as the unknown classes. In this way, our model can also learn the pixel representation of unknown classes." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 172, + 437, + 185 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 172, + 437, + 185 + ], + "spans": [ + { + "bbox": [ + 305, + 172, + 437, + 185 + ], + "type": "text", + "content": "3.3. Training and Inference" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 191, + 545, + 226 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 191, + 545, + 226 + ], + "spans": [ + { + "bbox": [ + 304, + 191, + 545, + 226 + ], + "type": "text", + "content": "In summary, we train the pixel embedding function by combining the pixel-segment contrastive loss, embedding consistent loss, and semantic consistent loss:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 384, + 237, + 545, + 250 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 384, + 237, + 545, + 250 + ], + "spans": [ + { + "bbox": [ + 384, + 237, + 545, + 250 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\mathcal {L} _ {t} + \\mathcal {L} _ {e} + \\mathcal {L} _ {s}. \\tag {5}", + "image_path": "ea946ce76ed0cafdc9e178b2927e0b620a1dd5243cd39093c72d8e1b28ac700e.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 259, + 545, + 282 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 259, + 545, + 282 + ], + "spans": [ + { + "bbox": [ + 304, + 259, + 545, + 282 + ], + "type": "text", + "content": "During training, we also update the embeddings for the unknown classes with " + }, + { + "bbox": [ + 304, + 259, + 545, + 282 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_u" + }, + { + "bbox": [ + 304, + 259, + 545, + 282 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 284, + 545, + 427 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 284, + 545, + 427 + ], + "spans": [ + { + "bbox": [ + 304, + 284, + 545, + 427 + ], + "type": "text", + "content": "For inference, we use the trained model to generate pixel embeddings for each input image and use the pixel embeddings for different downstream tasks, including language-driven and unsupervised semantic segmentation. For language-driven semantic segmentation, we first obtain the text embeddings of arbitrary inference classes by feeding the prompt-engineered texts into the text encoder of CLIP. Then we assign each pixel with the class label whose text embedding is the closest to " + }, + { + "bbox": [ + 304, + 284, + 545, + 427 + ], + "type": "inline_equation", + "content": "\\mathrm{CLIP - S^4}" + }, + { + "bbox": [ + 304, + 284, + 545, + 427 + ], + "type": "text", + "content": " pixel embedding. For unsupervised semantic segmentation, we follow previous work [19, 40] that uses " + }, + { + "bbox": [ + 304, + 284, + 545, + 427 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 284, + 545, + 427 + ], + "type": "text", + "content": " nearest neighbor search or linear classifier to perform semantic segmentation." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 439, + 388, + 452 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 439, + 388, + 452 + ], + "spans": [ + { + "bbox": [ + 305, + 439, + 388, + 452 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 459, + 545, + 542 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 459, + 545, + 542 + ], + "spans": [ + { + "bbox": [ + 304, + 459, + 545, + 542 + ], + "type": "text", + "content": "We evaluate our model on three tasks: 1) language-driven semantic segmentation for both known and unknown classes; 2) unsupervised semantic segmentation with " + }, + { + "bbox": [ + 304, + 459, + 545, + 542 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 459, + 545, + 542 + ], + "type": "text", + "content": "-means clustering/linear classification; 3) transfer learning of generated pixel embeddings for instance mask tracking. We also conduct ablation studies to understand the components of our model." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 551, + 368, + 562 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 551, + 368, + 562 + ], + "spans": [ + { + "bbox": [ + 306, + 551, + 368, + 562 + ], + "type": "text", + "content": "4.1. Datasets" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 569, + 545, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 569, + 545, + 629 + ], + "spans": [ + { + "bbox": [ + 304, + 569, + 545, + 629 + ], + "type": "text", + "content": "Pascal VOC 2012 [12] contains 20 object classes and a background class. It has 1,464 and 1,449 images for training and validation, respectively. Following common practice [27, 48], we augment the training data with additional annotations [15], resulting in 10,582 training images." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 629, + 545, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 629, + 545, + 676 + ], + "spans": [ + { + "bbox": [ + 304, + 629, + 545, + 676 + ], + "type": "text", + "content": "Pascal Context [29] extends Pascal VOC 2010 [12] with additional annotations on 4,998 training and 5,105 validation images. Following the prior work [49], we use the most common 59 classes for evaluation." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "content": "COCO-Stuff [5] labels MS COCO [26] with 171 object/stuff classes. It contains 118,287 and 5,000 images for training and validation, respectively." + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "11211" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 286, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 286, + 120 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 286, + 120 + ], + "type": "text", + "content": "DAVIS 2017 [33] contains video sequences for instance mask tracking. Following the prior work [18, 47], we train pixel embeddings on Pascal VOC 2012 and evaluate the validation sequences without fine-tuning." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 121, + 286, + 192 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 121, + 286, + 192 + ], + "spans": [ + { + "bbox": [ + 46, + 121, + 286, + 192 + ], + "type": "text", + "content": "It is worth mentioning that no ground truth labels of any datasets are used during training. Instead, we perform self-supervised learning on pseudo segments generated by contour detectors and owt-ucm [2]. Two contour detectors are used including HED [45] for Pascal VOC 2012 and Pascal Context and PMI [20] for COCO-Stuff." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 201, + 180, + 213 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 201, + 180, + 213 + ], + "spans": [ + { + "bbox": [ + 47, + 201, + 180, + 213 + ], + "type": "text", + "content": "4.2. Implementation Details" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 220, + 286, + 290 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 220, + 286, + 290 + ], + "spans": [ + { + "bbox": [ + 46, + 220, + 286, + 290 + ], + "type": "text", + "content": "For self-supervised contrastive learning, images are augmented with the same set of data augmentations as SimCLR [7], including random resizing, cropping, flipping, color jittering, and Gaussian blurring. The concentration constant " + }, + { + "bbox": [ + 46, + 220, + 286, + 290 + ], + "type": "inline_equation", + "content": "\\kappa" + }, + { + "bbox": [ + 46, + 220, + 286, + 290 + ], + "type": "text", + "content": " is set to 10, and the number of segments is set to 36 for each augmented view." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 291, + 286, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 291, + 286, + 373 + ], + "spans": [ + { + "bbox": [ + 46, + 291, + 286, + 373 + ], + "type": "text", + "content": "For vision-language guidance, we use pre-trained CLIP models [34] with modified image encoders following [49]. We use prompt-engineered texts with 85 prompt templates to generate text embeddings following [13, 49]. We use the average embedding of the top 32 segments of high probabilities as the prototype of each known class. We set the number of unknown classes to " + }, + { + "bbox": [ + 46, + 291, + 286, + 373 + ], + "type": "inline_equation", + "content": "u = 64" + }, + { + "bbox": [ + 46, + 291, + 286, + 373 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 375, + 286, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 375, + 286, + 506 + ], + "spans": [ + { + "bbox": [ + 46, + 375, + 286, + 506 + ], + "type": "text", + "content": "Following the prior work [18, 19, 23], we use PSP-Net [48] with a dilated ResNet-50 [17] backbone as the network architecture. The backbone is pre-trained on the ImageNet [9] dataset. We train our model on Pascal VOC 2012 and Pascal Context for " + }, + { + "bbox": [ + 46, + 375, + 286, + 506 + ], + "type": "inline_equation", + "content": "20\\mathrm{k}" + }, + { + "bbox": [ + 46, + 375, + 286, + 506 + ], + "type": "text", + "content": " iterations and on COCO-Stuff for " + }, + { + "bbox": [ + 46, + 375, + 286, + 506 + ], + "type": "inline_equation", + "content": "80\\mathrm{k}" + }, + { + "bbox": [ + 46, + 375, + 286, + 506 + ], + "type": "text", + "content": " iterations. We set the batch size to 8 with additional memory banks that cache the segment embeddings of the previous 2 batches. We set the initial learning rate to 0.001 and decay it with a polynomial learning rate policy. We use the CLIP model trained with ViT-B/16 backbone unless otherwise stated." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 515, + 264, + 528 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 515, + 264, + 528 + ], + "spans": [ + { + "bbox": [ + 47, + 515, + 264, + 528 + ], + "type": "text", + "content": "4.3. Language-Driven Semantic Segmentation" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 533, + 286, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 533, + 286, + 593 + ], + "spans": [ + { + "bbox": [ + 46, + 533, + 286, + 593 + ], + "type": "text", + "content": "For language-driven semantic segmentation, no human annotations are used for either training or inference. At the inference time, each pixel is assigned an arbitrarily given class label whose CLIP text embedding is the closest to this pixel's CLIP-" + }, + { + "bbox": [ + 46, + 533, + 286, + 593 + ], + "type": "inline_equation", + "content": "\\mathbf{S}^4" + }, + { + "bbox": [ + 46, + 533, + 286, + 593 + ], + "type": "text", + "content": " embedding." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 594, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 286, + 713 + ], + "type": "text", + "content": "We first compare the performance of our method with the state-of-the-art language-driven semantic segmentation approaches [37, 46, 49] on the Pascal Context and COCO-Stuff datasets. The performance is evaluated with the mean Intersection over Union (mIoU). For MaskCLIP and MaskCLIP+ [49], we obtain the results using the same hyper-parameter setting as our approach with CLIP models of two different backbones of ResNet50 and ViT-B/16. Meanwhile, GroupViT [46], ReCo [37], and " + }, + { + "bbox": [ + 46, + 594, + 286, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{ReCo + }" + }, + { + "bbox": [ + 46, + 594, + 286, + 713 + ], + "type": "text", + "content": " [37] use completely different training mechanisms compared" + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 306, + 70, + 549, + 232 + ], + "blocks": [ + { + "bbox": [ + 306, + 70, + 549, + 232 + ], + "lines": [ + { + "bbox": [ + 306, + 70, + 549, + 232 + ], + "spans": [ + { + "bbox": [ + 306, + 70, + 549, + 232 + ], + "type": "table", + "html": "
MethodCLIP ModelPascal ContextCOCO-Stuff
mIoUmIoU
GroupViT [46]-22.4-
ReCo [37]ResNet50x16 + ViT-L/14@336px26.6-
ReCo+ [37]†-18.4
MaskCLIP [49]ResNet5018.610.6
ViT-B/1625.215.2
MaskCLIP+ [49]†ResNet5023.413.9
ViT-B/1632.220.7
CLIP-S4†ResNet5028.5 (+5.1)16.7 (+2.8)
ViT-B/1633.6 (+1.4)22.1 (+1.4)
", + "image_path": "76cce9ee09455eb2b39f22c62c07754e55fc7524d556ea13e8ebfb0e259159f2.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 237, + 545, + 293 + ], + "lines": [ + { + "bbox": [ + 305, + 237, + 545, + 293 + ], + "spans": [ + { + "bbox": [ + 305, + 237, + 545, + 293 + ], + "type": "text", + "content": "Table 2. Language-guided semantic segmentation benchmarks (mIoU). CLIP-" + }, + { + "bbox": [ + 305, + 237, + 545, + 293 + ], + "type": "inline_equation", + "content": "S^4" + }, + { + "bbox": [ + 305, + 237, + 545, + 293 + ], + "type": "text", + "content": " consistently outperforms the state-of-the-art methods on both Pascal Context and COCO-Stuff datasets with CLIP models of different backbones. " + }, + { + "bbox": [ + 305, + 237, + 545, + 293 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 305, + 237, + 545, + 293 + ], + "type": "text", + "content": " indicates the models are fine-tuned on target datasets." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 303, + 545, + 398 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 303, + 545, + 398 + ], + "spans": [ + { + "bbox": [ + 304, + 303, + 545, + 398 + ], + "type": "text", + "content": "with our method. GroupViT is trained on image-caption pairs, and " + }, + { + "bbox": [ + 304, + 303, + 545, + 398 + ], + "type": "inline_equation", + "content": "\\mathrm{ReCo} / \\mathrm{ReCo}+" + }, + { + "bbox": [ + 304, + 303, + 545, + 398 + ], + "type": "text", + "content": " combines image retrieval and cosegmentation. For comparison, we take the best results from [37, 46] for GroupViT, ReCo, and " + }, + { + "bbox": [ + 304, + 303, + 545, + 398 + ], + "type": "inline_equation", + "content": "\\mathrm{ReCo}+" + }, + { + "bbox": [ + 304, + 303, + 545, + 398 + ], + "type": "text", + "content": ". Tabel 2 shows the benchmarking results of the aforementioned methods. Our method consistently outperforms the state-of-the-art on both datasets with CLIP models of different backbones." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 401, + 545, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 401, + 545, + 568 + ], + "spans": [ + { + "bbox": [ + 304, + 401, + 545, + 568 + ], + "type": "text", + "content": "To evaluate models' performance for class-free semantic segmentation with both known and unknown classes, we split the 59 classes of Pascal Context into 4 folds, where each fold includes around 15 classes. For each experiment, classes from one fold are considered as unknown and excluded during training. The mIoUs of known and unknown classes, as well as their harmonic mean (hIoU) are reported in Tab. 3. The performance of our method is averaged across 5 runs with randomly initialized prototypes of unknown classes. Our method achieves significant gains over MaskCLIP+ on unknown classes, which are comparable to MaskCLIP. Also, our method outperforms both MaskCLIP and MaskCLIP+ on known classes, which leads to better overall performance." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "text", + "content": "Qualitatively, the visualization in Fig. 4a offers us some insights into why our approach can achieve better results: our model yields consistent embeddings aligned with the pre-trained CLIP model. Fig. 4a visualizes the projection of segment embeddings generated by different methods on Pascal Context and COCO-Stuff. We observe that segment embeddings generated by " + }, + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{MaskCLIP + }" + }, + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "text", + "content": " are distorted by the given text embeddings. Meanwhile, " + }, + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{CLIP - S^4}" + }, + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "text", + "content": " generates segment embeddings that are well aligned with the segment embeddings derived from the pre-trained CLIP model. Hence, segment embeddings generated by " + }, + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{CLIP - S^4}" + }, + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "text", + "content": " can better capture both known and unknown classes. Fig. 4b shows" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "11212" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 48, + 70, + 545, + 145 + ], + "blocks": [ + { + "bbox": [ + 48, + 70, + 545, + 145 + ], + "lines": [ + { + "bbox": [ + 48, + 70, + 545, + 145 + ], + "spans": [ + { + "bbox": [ + 48, + 70, + 545, + 145 + ], + "type": "table", + "html": "
Methodfold0fold1fold2fold3
mIoUumIoUkhIoUmIoUumIoUkhIoUmIoUumIoUkhIoUmIoUumIoUkhIoU
MaskCLIP29.723.726.323.725.724.623.925.724.723.425.824.5
MaskCLIP+3.628.56.33.029.25.44.829.28.24.529.97.8
CLIP-S432.0±0.829.4±0.330.6±0.522.3±0.932.8±0.426.5±0.622.4±0.532.1±0.526.4±0.428.6±0.831.5±0.230.0±0.5
vs. MaskCLIP++28.4+0.9+24.3+19.3+3.6+21.1+17.6+2.9+18.2+24.1+1.6+22.2
", + "image_path": "d7701a5f3e1417ebfd98891b1d8054f8c67d051aaabee2afdcf12a019feed15c.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 49, + 209, + 287, + 280 + ], + "blocks": [ + { + "bbox": [ + 49, + 209, + 287, + 280 + ], + "lines": [ + { + "bbox": [ + 49, + 209, + 287, + 280 + ], + "spans": [ + { + "bbox": [ + 49, + 209, + 287, + 280 + ], + "type": "image", + "image_path": "b38239396bfa302ea59816bd846b1562552de090f3edcca588b1096df5ef28e8.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 56, + 281, + 280, + 290 + ], + "lines": [ + { + "bbox": [ + 56, + 281, + 280, + 290 + ], + "spans": [ + { + "bbox": [ + 56, + 281, + 280, + 290 + ], + "type": "text", + "content": "(a) Embeddings projection for PASCAL Context (left) and COCO-Stuff (right)" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 48, + 292, + 287, + 372 + ], + "blocks": [ + { + "bbox": [ + 48, + 292, + 287, + 372 + ], + "lines": [ + { + "bbox": [ + 48, + 292, + 287, + 372 + ], + "spans": [ + { + "bbox": [ + 48, + 292, + 287, + 372 + ], + "type": "image", + "image_path": "9a82d32527abf1564a23ee49c7a1193fcd48a0af523e48705c30de0549f21374.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 78, + 372, + 259, + 380 + ], + "lines": [ + { + "bbox": [ + 78, + 372, + 259, + 380 + ], + "spans": [ + { + "bbox": [ + 78, + 372, + 259, + 380 + ], + "type": "text", + "content": "(b) Retrieved image segments for classes not used in training" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 384, + 287, + 494 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 384, + 287, + 494 + ], + "spans": [ + { + "bbox": [ + 46, + 384, + 287, + 494 + ], + "type": "text", + "content": "Figure 4. (a) Projection of pixel embeddings generated by CLIP, MaskCLIP+, and CLIP-S" + }, + { + "bbox": [ + 46, + 384, + 287, + 494 + ], + "type": "inline_equation", + "content": "^4" + }, + { + "bbox": [ + 46, + 384, + 287, + 494 + ], + "type": "text", + "content": " trained on Pascal Context with CLIP's ViT-B/16 model (left) and COCO-Stuff with CLIP's ResNet50 model (right). MaskCLIP+ distorts pixel embeddings with respect to the given text embeddings, while CLIP-S" + }, + { + "bbox": [ + 46, + 384, + 287, + 494 + ], + "type": "inline_equation", + "content": "^4" + }, + { + "bbox": [ + 46, + 384, + 287, + 494 + ], + "type": "text", + "content": " aligns pixel embeddings with the pre-trained CLIP model. (b) Image segments retrieved for classes that are not used during training. Correct and incorrect retrievals are outlined in green and orange, respectively. Compared with MaskCLIP+, CLIP-S" + }, + { + "bbox": [ + 46, + 384, + 287, + 494 + ], + "type": "inline_equation", + "content": "^4" + }, + { + "bbox": [ + 46, + 384, + 287, + 494 + ], + "type": "text", + "content": " retrieves images that are more closely related to the classes." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 496, + 287, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 496, + 287, + 593 + ], + "spans": [ + { + "bbox": [ + 46, + 496, + 287, + 593 + ], + "type": "text", + "content": "image segments retrieved from the COCO-Stuff validation set using MaskCLIP+ and CLIP-S" + }, + { + "bbox": [ + 46, + 496, + 287, + 593 + ], + "type": "inline_equation", + "content": "^4" + }, + { + "bbox": [ + 46, + 496, + 287, + 593 + ], + "type": "text", + "content": " for a set of classes that are not used in training. For each class, we obtain its text embedding and compare it with segment embeddings to obtain top retrievals from both methods. Due to better alignment with the pre-trained CLIP model, CLIP-S" + }, + { + "bbox": [ + 46, + 496, + 287, + 593 + ], + "type": "inline_equation", + "content": "^4" + }, + { + "bbox": [ + 46, + 496, + 287, + 593 + ], + "type": "text", + "content": " retrieves images that are more closely related to the unknown classes compared with MaskCLIP+." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 599, + 246, + 612 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 599, + 246, + 612 + ], + "spans": [ + { + "bbox": [ + 47, + 599, + 246, + 612 + ], + "type": "text", + "content": "4.4. Unsupervised Semantic Segmentation" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 617, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 617, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 617, + 288, + 714 + ], + "type": "text", + "content": "To study whether " + }, + { + "bbox": [ + 46, + 617, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\mathrm{CLIP - S^4}" + }, + { + "bbox": [ + 46, + 617, + 288, + 714 + ], + "type": "text", + "content": " can generate pixel embeddings that form distinctive clusters, we evaluate " + }, + { + "bbox": [ + 46, + 617, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\mathrm{CLIP - S^4}" + }, + { + "bbox": [ + 46, + 617, + 288, + 714 + ], + "type": "text", + "content": " on the unsupervised semantic segmentation task for Pascal VOC 2012. To derive semantic segmentation from pixel embeddings, we use and test two approaches, including " + }, + { + "bbox": [ + 46, + 617, + 288, + 714 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 617, + 288, + 714 + ], + "type": "text", + "content": " nearest neighbor (" + }, + { + "bbox": [ + 46, + 617, + 288, + 714 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 617, + 288, + 714 + ], + "type": "text", + "content": "-NN) search [19] and linear classification [40]. For " + }, + { + "bbox": [ + 46, + 617, + 288, + 714 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 617, + 288, + 714 + ], + "type": "text", + "content": "-NN search, we assign each segment a class label by the majority vote of its nearest neighbors from the" + } + ] + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 310, + 209, + 542, + 346 + ], + "blocks": [ + { + "bbox": [ + 46, + 148, + 546, + 204 + ], + "lines": [ + { + "bbox": [ + 46, + 148, + 546, + 204 + ], + "spans": [ + { + "bbox": [ + 46, + 148, + 546, + 204 + ], + "type": "text", + "content": "Table 3. Language-guided semantic segmentation benchmarks (mIoU) for unknown classes. The classes of Pascal Context are split into 4 folds with around 15 classes each fold. For each experiment, classes of one fold are considered as unknown. The performance of CLIP-" + }, + { + "bbox": [ + 46, + 148, + 546, + 204 + ], + "type": "inline_equation", + "content": "S^4" + }, + { + "bbox": [ + 46, + 148, + 546, + 204 + ], + "type": "text", + "content": " is averaged over 5 runs with randomly initialized unknown class embeddings. CLIP-" + }, + { + "bbox": [ + 46, + 148, + 546, + 204 + ], + "type": "inline_equation", + "content": "S^4" + }, + { + "bbox": [ + 46, + 148, + 546, + 204 + ], + "type": "text", + "content": " significantly outperforms " + }, + { + "bbox": [ + 46, + 148, + 546, + 204 + ], + "type": "inline_equation", + "content": "\\text{MaskCLIP}+" + }, + { + "bbox": [ + 46, + 148, + 546, + 204 + ], + "type": "text", + "content": " on unknown classes. Meanwhile, CLIP-" + }, + { + "bbox": [ + 46, + 148, + 546, + 204 + ], + "type": "inline_equation", + "content": "S^4" + }, + { + "bbox": [ + 46, + 148, + 546, + 204 + ], + "type": "text", + "content": " archives consistent gains on known classes over " + }, + { + "bbox": [ + 46, + 148, + 546, + 204 + ], + "type": "inline_equation", + "content": "\\text{MaskCLIP}" + }, + { + "bbox": [ + 46, + 148, + 546, + 204 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 148, + 546, + 204 + ], + "type": "inline_equation", + "content": "\\text{MaskCLIP}+" + }, + { + "bbox": [ + 46, + 148, + 546, + 204 + ], + "type": "text", + "content": ", and hence leads to better overall performance." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 310, + 209, + 542, + 346 + ], + "lines": [ + { + "bbox": [ + 310, + 209, + 542, + 346 + ], + "spans": [ + { + "bbox": [ + 310, + 209, + 542, + 346 + ], + "type": "table", + "html": "
MethodmIoU k-NNmIoU Linear Classifier
IIC [22]-28.0
SegSort [19]47.355.4
Hierarch. Group. [47]-48.8
MaskContrast (Sup.) [40]53.963.9
ConceptContrast [18]58.860.4
HSG [23]61.7-
MaskCLIP [49]67.369.5
MaskCLIP+ [49]65.170.0
CLIP-S472.0 (+4.7)73.0 (+3.0)
", + "image_path": "4b9bf12ae497ed3b31f7890395e4b14a9732db31ddf0577c9817e810e8a6c6ae.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 351, + 545, + 394 + ], + "lines": [ + { + "bbox": [ + 305, + 351, + 545, + 394 + ], + "spans": [ + { + "bbox": [ + 305, + 351, + 545, + 394 + ], + "type": "text", + "content": "Table 4. Unsupervised semantic segmentation benchmarks (mIoU) on Pascal VOC 2012. CLIP-" + }, + { + "bbox": [ + 305, + 351, + 545, + 394 + ], + "type": "inline_equation", + "content": "S^4" + }, + { + "bbox": [ + 305, + 351, + 545, + 394 + ], + "type": "text", + "content": " consistently outperforms the state-of-the-art methods on both " + }, + { + "bbox": [ + 305, + 351, + 545, + 394 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 305, + 351, + 545, + 394 + ], + "type": "text", + "content": "-NN search and linear classification." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 305, + 399, + 545, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 399, + 545, + 434 + ], + "spans": [ + { + "bbox": [ + 305, + 399, + 545, + 434 + ], + "type": "text", + "content": "training set following [19]. For linear classification, we train a linear classifier on the learned pixel embeddings following [40]." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 435, + 546, + 591 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 435, + 546, + 591 + ], + "spans": [ + { + "bbox": [ + 304, + 435, + 546, + 591 + ], + "type": "text", + "content": "We compare our method with the state-of-the-art unsupervised and language-guided semantic segmentation approaches. We train the state-of-the-art models using the same hyper-parameter setting as our approach except for IIC [22] and Hierarchical Grouping [47] as they use different training mechanisms. For comparison, we take the best results for IIC and Hierarchical Grouping. The benchmark results are shown in Tab. 4. With the vision-language guidance, our method achieves significant gains compared with the previous non-CLIP-based approaches (i.e., " + }, + { + "bbox": [ + 304, + 435, + 546, + 591 + ], + "type": "inline_equation", + "content": "+9\\%" + }, + { + "bbox": [ + 304, + 435, + 546, + 591 + ], + "type": "text", + "content": " for both " + }, + { + "bbox": [ + 304, + 435, + 546, + 591 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 435, + 546, + 591 + ], + "type": "text", + "content": "-NN search and linear classification). Meanwhile, our method also outperforms the language-guided semantic segmentation approaches by a large margin." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 599, + 443, + 612 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 599, + 443, + 612 + ], + "spans": [ + { + "bbox": [ + 306, + 599, + 443, + 612 + ], + "type": "text", + "content": "4.5. Instance Mask Tracking" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 617, + 546, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 617, + 546, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 617, + 546, + 714 + ], + "type": "text", + "content": "We evaluate the transferability of pixel embeddings learned from the Pascal VOC 2012 dataset. We use the pixel embeddings to track instance masks in the DAVIS 2017 validation set, where the instance masks at the first frame are given for each video. Following the prior work [48], we use the similarity between pixel embeddings cross frames to propagate the instance masks to the rest of the video frames. We evaluate the performance using the region similarity " + }, + { + "bbox": [ + 304, + 617, + 546, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{I}" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "11213" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 55, + 70, + 279, + 223 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 279, + 223 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 279, + 223 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 279, + 223 + ], + "type": "table", + "html": "
MethodJ(Mean)↑F(Mean)↑
MaskTrack-B [32]35.336.4
OSVOS-B [4]18.530.0
Video Colorization [42]34.632.7
CycleTime [43]41.939.4
mgPFF [24]42.246.9
Hierarchy. Group. [47]47.148.9
MaskContrast (Sup.) [40]34.336.7
ConceptContrast [18]50.453.9
MaskCLIP [49]48.149.2
MaskCLIP+ [49]42.644.2
CLIP-S452.3 (+1.9)56.8 (+2.9)
", + "image_path": "e1d1565331bc282ab95547a9e31975745fe237b77298d7e8764364e7ff9ad168.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 62, + 293, + 272, + 369 + ], + "blocks": [ + { + "bbox": [ + 46, + 228, + 287, + 283 + ], + "lines": [ + { + "bbox": [ + 46, + 228, + 287, + 283 + ], + "spans": [ + { + "bbox": [ + 46, + 228, + 287, + 283 + ], + "type": "text", + "content": "Table 5. Quantitative evaluation of instance mask tracking on the DAVIS-2017 validation set. The performance is measured by the region similarity " + }, + { + "bbox": [ + 46, + 228, + 287, + 283 + ], + "type": "inline_equation", + "content": "\\mathcal{J}" + }, + { + "bbox": [ + 46, + 228, + 287, + 283 + ], + "type": "text", + "content": " (IoU) and the contour-based accuracy " + }, + { + "bbox": [ + 46, + 228, + 287, + 283 + ], + "type": "inline_equation", + "content": "\\mathcal{F}" + }, + { + "bbox": [ + 46, + 228, + 287, + 283 + ], + "type": "text", + "content": " defined by [33]. Our method outperforms existing supervised, unsupervised, and language-guided approaches on both metrics." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 62, + 293, + 272, + 369 + ], + "lines": [ + { + "bbox": [ + 62, + 293, + 272, + 369 + ], + "spans": [ + { + "bbox": [ + 62, + 293, + 272, + 369 + ], + "type": "table", + "html": "
LtLeLspAccmIoUavgsim
✓--1.60.5-0.01
✓✓-48.124.30.79
✓-✓52.332.90.33
-✓✓48.631.3-
✓✓✓53.733.60.66
", + "image_path": "02057dafdee078c008f55994e339ed50f99eb0bbf5bad751a806b01264aa2d6c.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 458, + 277, + 470 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 458, + 277, + 470 + ], + "spans": [ + { + "bbox": [ + 47, + 458, + 277, + 470 + ], + "type": "text", + "content": "(IoU) and the contour-based accuracy " + }, + { + "bbox": [ + 47, + 458, + 277, + 470 + ], + "type": "inline_equation", + "content": "\\mathcal{F}" + }, + { + "bbox": [ + 47, + 458, + 277, + 470 + ], + "type": "text", + "content": " defined by [33]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 471, + 286, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 471, + 286, + 578 + ], + "spans": [ + { + "bbox": [ + 46, + 471, + 286, + 578 + ], + "type": "text", + "content": "We compare our method with existing supervised [4, 32], unsupervised [18, 24, 40, 42, 43, 47], and language-guided [49] approaches (Tab. 5). Though not trained on any video sequences, our method outperforms the existing approaches by more than " + }, + { + "bbox": [ + 46, + 471, + 286, + 578 + ], + "type": "inline_equation", + "content": "1.9\\%" + }, + { + "bbox": [ + 46, + 471, + 286, + 578 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 471, + 286, + 578 + ], + "type": "inline_equation", + "content": "2.9\\%" + }, + { + "bbox": [ + 46, + 471, + 286, + 578 + ], + "type": "text", + "content": " in terms of the region similarity " + }, + { + "bbox": [ + 46, + 471, + 286, + 578 + ], + "type": "inline_equation", + "content": "\\mathcal{I}" + }, + { + "bbox": [ + 46, + 471, + 286, + 578 + ], + "type": "text", + "content": " and contour accuracy " + }, + { + "bbox": [ + 46, + 471, + 286, + 578 + ], + "type": "inline_equation", + "content": "\\mathcal{F}" + }, + { + "bbox": [ + 46, + 471, + 286, + 578 + ], + "type": "text", + "content": ", respectively. Note that the pixel embeddings generated by " + }, + { + "bbox": [ + 46, + 471, + 286, + 578 + ], + "type": "inline_equation", + "content": "\\mathrm{MaskCLIP + }" + }, + { + "bbox": [ + 46, + 471, + 286, + 578 + ], + "type": "text", + "content": " [49] are distorted by the classes from Pascal VOC 2012, which hinder their transferability." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 586, + 141, + 600 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 586, + 141, + 600 + ], + "spans": [ + { + "bbox": [ + 47, + 586, + 141, + 600 + ], + "type": "text", + "content": "4.6. Ablation Study" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 605, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 287, + 715 + ], + "type": "text", + "content": "We study the contribution of different losses of our method using the Pascal Context dataset and the language-guided semantic segmentation task. The performance is evaluated with pixel accuracy (pAcc) and mIoU. We also calculate the average cosine similarity (avgsim) between our segment embeddings and CLIP's segment embeddings to quantify the alignment. Tab. 6 shows the study results. We observe that by introducing embedding consistent loss " + }, + { + "bbox": [ + 46, + 605, + 287, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_e" + }, + { + "bbox": [ + 46, + 605, + 287, + 715 + ], + "type": "text", + "content": " the learned segment embeddings are well aligned with" + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 312, + 70, + 413, + 151 + ], + "blocks": [ + { + "bbox": [ + 46, + 373, + 287, + 450 + ], + "lines": [ + { + "bbox": [ + 46, + 373, + 287, + 450 + ], + "spans": [ + { + "bbox": [ + 46, + 373, + 287, + 450 + ], + "type": "text", + "content": "Table 6. Ablation study on the contribution of each loss of CLIP-S4. Experimented on language-guided semantic segmentation of Pascal Context. avgsim represents the average cosine similarity between segment embeddings generated by CLIP-S4 and CLIP. By combining the embedding and semantic consistent losses " + }, + { + "bbox": [ + 46, + 373, + 287, + 450 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_e" + }, + { + "bbox": [ + 46, + 373, + 287, + 450 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 373, + 287, + 450 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_s" + }, + { + "bbox": [ + 46, + 373, + 287, + 450 + ], + "type": "text", + "content": ", CLIP-S4 archives better semantic segmentation performance while maintaining the alignment with CLIP's embeddings." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 312, + 70, + 413, + 151 + ], + "lines": [ + { + "bbox": [ + 312, + 70, + 413, + 151 + ], + "spans": [ + { + "bbox": [ + 312, + 70, + 413, + 151 + ], + "type": "table", + "html": "
#unknowns(u)mIoU
1671.6
3271.9
6472.0
12872.2
25671.9
", + "image_path": "843984697e9c406da2bd2251d274c734f4c475df4cbceb0d4a88360c90cb1f81.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 423, + 70, + 539, + 151 + ], + "blocks": [ + { + "bbox": [ + 423, + 70, + 539, + 151 + ], + "lines": [ + { + "bbox": [ + 423, + 70, + 539, + 151 + ], + "spans": [ + { + "bbox": [ + 423, + 70, + 539, + 151 + ], + "type": "table", + "html": "
top-m segmentsavgsim
10.921
40.977
160.996
640.997
2560.993
", + "image_path": "79c7a2b889ed6be0a126b6dc3cb5d9b1efdde0aaf17d673b347e6498f3ddc233.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 156, + 418, + 190 + ], + "lines": [ + { + "bbox": [ + 305, + 156, + 418, + 190 + ], + "spans": [ + { + "bbox": [ + 305, + 156, + 418, + 190 + ], + "type": "text", + "content": "Table 7. Ablation study on the influence of the number of unknown class prototypes." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 421, + 156, + 542, + 190 + ], + "lines": [ + { + "bbox": [ + 421, + 156, + 542, + 190 + ], + "spans": [ + { + "bbox": [ + 421, + 156, + 542, + 190 + ], + "type": "text", + "content": "Table 8. Ablation study on different numbers of top- " + }, + { + "bbox": [ + 421, + 156, + 542, + 190 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 421, + 156, + 542, + 190 + ], + "type": "text", + "content": " segments for class prototypes." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 192, + 545, + 371 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 192, + 545, + 371 + ], + "spans": [ + { + "bbox": [ + 304, + 192, + 545, + 371 + ], + "type": "text", + "content": "CLIP's embeddings with an average cosine similarity of 0.79. However, the learned segment embeddings do not perform well on the language-guided semantic segmentation task (24.3 vs. 33.6), because the segment embeddings are not optimized to classify target classes. Meanwhile, by using semantic consistent loss " + }, + { + "bbox": [ + 304, + 192, + 545, + 371 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_s" + }, + { + "bbox": [ + 304, + 192, + 545, + 371 + ], + "type": "text", + "content": " without embedding consistent loss, the learned segment embeddings have the discriminative power to classify different classes but are not aligned with CLIP's embeddings as the average cosine similarity is 0.33. As a result, the segment embeddings are limited to the target classes used during training. Hence, we combine " + }, + { + "bbox": [ + 304, + 192, + 545, + 371 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_e" + }, + { + "bbox": [ + 304, + 192, + 545, + 371 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 192, + 545, + 371 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_s" + }, + { + "bbox": [ + 304, + 192, + 545, + 371 + ], + "type": "text", + "content": " to balance the discriminative power over target classes and the alignment with CLIP. Meanwhile, we observe that with pixel-segment contrastive learning, the model can archive better performance." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 372, + 545, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 372, + 545, + 574 + ], + "spans": [ + { + "bbox": [ + 304, + 372, + 545, + 574 + ], + "type": "text", + "content": "Also, we study the influence of the number of unknown class prototypes on the Pascal VOC dataset for the unsupervised semantic segmentation task. The results Tab. 7 show that the semantic segmentation performance is robust to the tested number of unknown class prototypes as the mIoU varies only " + }, + { + "bbox": [ + 304, + 372, + 545, + 574 + ], + "type": "inline_equation", + "content": "0.6\\%" + }, + { + "bbox": [ + 304, + 372, + 545, + 574 + ], + "type": "text", + "content": ". Furthermore, we investigate how the size of top-" + }, + { + "bbox": [ + 304, + 372, + 545, + 574 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 304, + 372, + 545, + 574 + ], + "type": "text", + "content": " segments impacts the embeddings of class prototypes. We compare the embeddings of class prototypes generated with different numbers of top-" + }, + { + "bbox": [ + 304, + 372, + 545, + 574 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 304, + 372, + 545, + 574 + ], + "type": "text", + "content": " segments on the Pascal VOC dataset. We use the embeddings of class prototypes generated with " + }, + { + "bbox": [ + 304, + 372, + 545, + 574 + ], + "type": "inline_equation", + "content": "m = 32" + }, + { + "bbox": [ + 304, + 372, + 545, + 574 + ], + "type": "text", + "content": " segments as the reference, and compute the cosine similarity between the reference prototypes embeddings and ones with different top-" + }, + { + "bbox": [ + 304, + 372, + 545, + 574 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 304, + 372, + 545, + 574 + ], + "type": "text", + "content": " segments. For each case, the cosine similarity is averaged over all class prototypes. We observe that the embeddings of class prototypes are relatively stable if moderate top-" + }, + { + "bbox": [ + 304, + 372, + 545, + 574 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 304, + 372, + 545, + 574 + ], + "type": "text", + "content": " segments (e.g., " + }, + { + "bbox": [ + 304, + 372, + 545, + 574 + ], + "type": "inline_equation", + "content": "m = 32" + }, + { + "bbox": [ + 304, + 372, + 545, + 574 + ], + "type": "text", + "content": " in this work) are used (Tab. 8)." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 586, + 378, + 597 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 586, + 378, + 597 + ], + "spans": [ + { + "bbox": [ + 306, + 586, + 378, + 597 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": "We propose CLIP-S" + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "inline_equation", + "content": "^4" + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": ", a novel pixel representation learning approach for semantic segmentation. Our method combines self-supervised contrastive learning and guidance of CLIP to learn consistent pixel embeddings with respect to visual and conceptual semantics. Our experiments on popular semantic segmentation benchmarks demonstrate consistent gains over the state-of-the-art unsupervised semantic segmentation and language-driven semantic segmentation methods, especially for unknown classes." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "11214" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 134 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 134 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 134 + ], + "type": "text", + "content": "[1] Radhakrishna Achanta, Appu Shaji, Kevin Smith, Aurelien Lucchi, Pascal Fua, and Sabine Süsstrunk. SLIC superpixels compared to state-of-the-art superpixel methods. IEEE TPAMI, 34(11):2274-2282, 2012. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 137, + 287, + 169 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 137, + 287, + 169 + ], + "spans": [ + { + "bbox": [ + 53, + 137, + 287, + 169 + ], + "type": "text", + "content": "[2] Pablo Arbeláez, Michael Maire, Charless Fowlkes, and Jitendra Malik. Contour detection and hierarchical image segmentation. IEEE TPAMI, 33(5):898-916, 2011. 3, 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 171, + 287, + 204 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 171, + 287, + 204 + ], + "spans": [ + { + "bbox": [ + 53, + 171, + 287, + 204 + ], + "type": "text", + "content": "[3] Maxime Bucher, Tuan-Hung VU, Matthieu Cord, and Patrick Pérez. Zero-shot semantic segmentation. In NeurIPS, pages 468-479, 2019. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 205, + 287, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 205, + 287, + 237 + ], + "spans": [ + { + "bbox": [ + 53, + 205, + 287, + 237 + ], + "type": "text", + "content": "[4] S. Caelles, K.-K. Maninis, J. Pont-Tuset, L. Leal-Taixe, D. Cremers, and L. Van Gool. One-shot video object segmentation. In CVPR, pages 5320-5329, 2017. 8" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 239, + 287, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 239, + 287, + 271 + ], + "spans": [ + { + "bbox": [ + 53, + 239, + 287, + 271 + ], + "type": "text", + "content": "[5] Holger Caesar, Jasper Uijlings, and Vittorio Ferrari. COCO-Stuff: Thing and stuff classes in context. In CVPR, pages 1209-1218, 2018. 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 274, + 287, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 274, + 287, + 327 + ], + "spans": [ + { + "bbox": [ + 53, + 274, + 287, + 327 + ], + "type": "text", + "content": "[6] Liang-Chieh Chen, George Papandreou, Iasonas Kokkinos, Kevin Murphy, and Alan L. Yuille. DeepLab: Semantic image segmentation with deep convolutional nets, atrous convolution, and fully connected CRFs. IEEE TPAMI, 40(4):834-848, 2018. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 330, + 287, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 330, + 287, + 373 + ], + "spans": [ + { + "bbox": [ + 53, + 330, + 287, + 373 + ], + "type": "text", + "content": "[7] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In ICML, pages 1597-1607, 2020. 1, 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 375, + 287, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 375, + 287, + 418 + ], + "spans": [ + { + "bbox": [ + 53, + 375, + 287, + 418 + ], + "type": "text", + "content": "[8] Jang Hyun Cho, Utkarsh Mall, Kavita Bala, and Bharath Hariharan. PiCIE: Unsupervised semantic segmentation using invariance and equivariance in clustering. In CVPR, pages 16794-16804, 2021. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 420, + 287, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 420, + 287, + 453 + ], + "spans": [ + { + "bbox": [ + 53, + 420, + 287, + 453 + ], + "type": "text", + "content": "[9] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. ImageNet: A large-scale hierarchical image database. In CVPR, pages 248–255, 2009. 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 454, + 287, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 454, + 287, + 520 + ], + "spans": [ + { + "bbox": [ + 48, + 454, + 287, + 520 + ], + "type": "text", + "content": "[10] Xiaoyi Dong, Yinglin Zheng, Jianmin Bao, Ting Zhang, Dongdong Chen, Hao Yang, Ming Zeng, Weiming Zhang, Lu Yuan, Dong Chen, Fang Wen, and Nenghai Yu. MaskCLIP: Masked self-distillation advances contrastive language-image pretraining. arXiv preprint arXiv:2208.12262, 2022.3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 521, + 287, + 554 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 521, + 287, + 554 + ], + "spans": [ + { + "bbox": [ + 48, + 521, + 287, + 554 + ], + "type": "text", + "content": "[11] Sepideh Esmaeilpour, Bing Liu, Eric Robertson, and Lei Shu. Zero-shot out-of-distribution detection based on the pre-trained model CLIP. In AAAI, 2022. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 555, + 287, + 598 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 555, + 287, + 598 + ], + "spans": [ + { + "bbox": [ + 48, + 555, + 287, + 598 + ], + "type": "text", + "content": "[12] Mark Everingham, Luc Van Gool, Christopher K. I. Williams, John Winn, and Andrew Zisserman. The pascal visual object classes (voc) challenge. *IJCV*, 88(2):303–338, 2010. 5" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 601, + 286, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 601, + 286, + 633 + ], + "spans": [ + { + "bbox": [ + 48, + 601, + 286, + 633 + ], + "type": "text", + "content": "[13] Xiuye Gu, Tsung-Yi Lin, Weicheng Kuo, and Yin Cui. Open-vocabulary object detection via vision and language knowledge distillation. In ICLR, 2022. 3, 4, 5, 6" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 635, + 287, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 635, + 287, + 678 + ], + "spans": [ + { + "bbox": [ + 48, + 635, + 287, + 678 + ], + "type": "text", + "content": "[14] Mark Hamilton, Zhoutong Zhang, Bharath Hariharan, Noah Snavely, and William T. Freeman. Unsupervised semantic segmentation by distilling feature correspondences. In ICLR, 2022. 1, 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "type": "text", + "content": "[15] Bharath Hariharan, Pablo Arbeláez, Lubomir Bourdev, Subhransu Maji, and Jitendra Malik. Semantic contours from inverse detectors. In ICCV, pages 991-998, 2011. 5" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 33, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "text", + "content": "[16] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In CVPR, pages 9726-9735, 2020. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 108, + 545, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 108, + 545, + 140 + ], + "spans": [ + { + "bbox": [ + 307, + 108, + 545, + 140 + ], + "type": "text", + "content": "[17] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, pages 770-778, 2016. 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 142, + 545, + 186 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 142, + 545, + 186 + ], + "spans": [ + { + "bbox": [ + 307, + 142, + 545, + 186 + ], + "type": "text", + "content": "[18] Wenbin He, William Surmeier, Arvind Kumar Shekar, Liang Gou, and Liu Ren. Self-supervised semantic segmentation grounded in visual concepts. In IJCAI, pages 949-955, 2022. 1, 2, 3, 6, 7, 8" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 188, + 545, + 232 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 188, + 545, + 232 + ], + "spans": [ + { + "bbox": [ + 307, + 188, + 545, + 232 + ], + "type": "text", + "content": "[19] Jyh-Jing Hwang, Stella X. Yu, Jianbo Shi, Maxwell D. Collins, Tien-Ju Yang, Xiao Zhang, and Liang-Chieh Chen. SegSort: Segmentation by discriminative sorting of segments. In ICCV, pages 7333-7343, 2019. 1, 2, 5, 6, 7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 234, + 545, + 266 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 234, + 545, + 266 + ], + "spans": [ + { + "bbox": [ + 307, + 234, + 545, + 266 + ], + "type": "text", + "content": "[20] Phillip Isola, Daniel Zoran, Dilip Krishnan, and Edward H. Adelson. Crisp boundary detection using pointwise mutual information. In ECCV, pages 799-814, 2014. 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 269, + 545, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 269, + 545, + 300 + ], + "spans": [ + { + "bbox": [ + 307, + 269, + 545, + 300 + ], + "type": "text", + "content": "[21] Ajay Jain, Matthew Tancik, and Pieter Abbeel. Putting NeRF on a diet: Semantically consistent few-shot view synthesis. In ICCV, pages 5865-5874, 2021. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 303, + 545, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 303, + 545, + 335 + ], + "spans": [ + { + "bbox": [ + 307, + 303, + 545, + 335 + ], + "type": "text", + "content": "[22] Xu Ji, Andrea Vedaldi, and Joao Henriques. Invariant information clustering for unsupervised image classification and segmentation. In ICCV, pages 9864-9873, 2019. 2, 7" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 338, + 545, + 390 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 338, + 545, + 390 + ], + "spans": [ + { + "bbox": [ + 307, + 338, + 545, + 390 + ], + "type": "text", + "content": "[23] Tsung-Wei Ke, Jyh-Jing Hwang, Yunhui Guo, Xudong Wang, and Stella X. Yu. Unsupervised hierarchical semantic segmentation with multiview cosegmentation and clustering transformers. In CVPR, pages 2561–2571, 2022. 1, 2, 3, 6, 7" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 394, + 545, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 394, + 545, + 426 + ], + "spans": [ + { + "bbox": [ + 307, + 394, + 545, + 426 + ], + "type": "text", + "content": "[24] Shu Kong and Charless Fowlkes. Multigrid predictive filter flow for unsupervised learning on videos. arXiv preprint arXiv:1904.01693, 2019. 8" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 429, + 545, + 460 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 429, + 545, + 460 + ], + "spans": [ + { + "bbox": [ + 307, + 429, + 545, + 460 + ], + "type": "text", + "content": "[25] Boyi Li, Kilian Q Weinberger, Serge Belongie, Vladlen Koltun, and René Ranftl. Language-driven semantic segmentation. In ICLR, 2022. 1, 2, 3, 5" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 463, + 545, + 507 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 463, + 545, + 507 + ], + "spans": [ + { + "bbox": [ + 307, + 463, + 545, + 507 + ], + "type": "text", + "content": "[26] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dolkar, and C. Lawrence Zitnick. Microsoft COCO: Common objects in context. In ECCV, pages 740-755, 2014. 5" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 509, + 545, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 509, + 545, + 542 + ], + "spans": [ + { + "bbox": [ + 307, + 509, + 545, + 542 + ], + "type": "text", + "content": "[27] Jonathan Long, Evan Shelhamer, and Trevor Darrell. Fully convolutional networks for semantic segmentation. In CVPR, pages 3431-3440, 2015. 1, 2, 5" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 544, + 545, + 587 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 544, + 545, + 587 + ], + "spans": [ + { + "bbox": [ + 307, + 544, + 545, + 587 + ], + "type": "text", + "content": "[28] Luke Melas-Kyriazi, Christian Rupprecht, Iro Laina, and Andrea Vedaldi. Deep spectral methods: A surprisingly strong baseline for unsupervised semantic segmentation and localization. In CVPR, pages 8354-8365, 2022. 2" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 307, + 589, + 545, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 589, + 545, + 643 + ], + "spans": [ + { + "bbox": [ + 307, + 589, + 545, + 643 + ], + "type": "text", + "content": "[29] Roozbeh Mottaghi, Xianjie Chen, Xiaobai Liu, Nam-Gyu Cho, Seong-Whan Lee, Sanja Fidler, Raquel Urtasun, and Alan Yuille. The role of context for object detection and semantic segmentation in the wild. In CVPR, pages 891–898, 2014. 5" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 307, + 646, + 545, + 677 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 646, + 545, + 677 + ], + "spans": [ + { + "bbox": [ + 307, + 646, + 545, + 677 + ], + "type": "text", + "content": "[30] Yassine Ouali, Céline Hudelot, and Myriam Tami. Autoregressive unsupervised image segmentation. In ECCV, pages 142-158, 2020. 2" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 307, + 680, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 680, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 680, + 545, + 713 + ], + "type": "text", + "content": "[31] Dim P. Papadopoulos, Ethan Weber, and Antonio Torralba. Scaling up instance annotation via label propagation. In ICCV, pages 15344-15353, 2021. 1" + } + ] + } + ], + "index": 32 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "11215" + } + ] + } + ], + "index": 34 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 73, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 48, + 73, + 287, + 115 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 73, + 287, + 115 + ], + "spans": [ + { + "bbox": [ + 48, + 73, + 287, + 115 + ], + "type": "text", + "content": "[32] Federico Perazzi, Anna Khoreva, Rodrigo Benenson, Bernt Schiele, and Alexander Sorkine-Hornung. Learning video object segmentation from static images. In CVPR, pages 3491-3500, 2017. 8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 117, + 287, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 117, + 287, + 161 + ], + "spans": [ + { + "bbox": [ + 49, + 117, + 287, + 161 + ], + "type": "text", + "content": "[33] F. Peruzzi, J. Pont-Tuset, B. McWilliams, L. Van Gool, M. Gross, and A. Sorkine-Hornung. A benchmark dataset and evaluation methodology for video object segmentation. In CVPR, pages 724-732, 2016. 6, 8" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 162, + 287, + 225 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 162, + 287, + 225 + ], + "spans": [ + { + "bbox": [ + 49, + 162, + 287, + 225 + ], + "type": "text", + "content": "[34] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In ICML, pages 8748-8763, 2021. 1, 2, 4, 5, 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 228, + 287, + 280 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 228, + 287, + 280 + ], + "spans": [ + { + "bbox": [ + 49, + 228, + 287, + 280 + ], + "type": "text", + "content": "[35] Yongming Rao, Wenliang Zhao, Guangyi Chen, Yansong Tang, Zheng Zhu, Guan Huang, Jie Zhou, and Jiwen Lu. DenseCLIP: Language-guided dense prediction with context-aware prompting. In CVPR, pages 18061-18070, 2022. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 49, + 282, + 287, + 325 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 282, + 287, + 325 + ], + "spans": [ + { + "bbox": [ + 49, + 282, + 287, + 325 + ], + "type": "text", + "content": "[36] Ramprasaath R. Selvaraju, Karan Desai, Justin Johnson, and Nikhil Naik. CASTing your model: Learning to localize improves self-supervised representations. In CVPR, pages 11053-11062, 2021. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 49, + 327, + 287, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 327, + 287, + 358 + ], + "spans": [ + { + "bbox": [ + 49, + 327, + 287, + 358 + ], + "type": "text", + "content": "[37] Gyungin Shin, Weidi Wie, and Samuel Albanie. ReCo: Retrieve and co-segment for zero-shot transfer. In NeurIPS, 2022. 1, 2, 3, 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 49, + 360, + 287, + 414 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 360, + 287, + 414 + ], + "spans": [ + { + "bbox": [ + 49, + 360, + 287, + 414 + ], + "type": "text", + "content": "[38] Kihyuk Sohn, David Berthelot, Chun-Liang Li, Zizhao Zhang, Nicholas Carlini, Ekin D. Cubuk, Alex Kurakin, Han Zhang, and Colin Raffel. FixMatch: Simplifying semi-supervised learning with consistency and confidence. In NeurIPS, pages 596-608, 2020. 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 49, + 415, + 287, + 458 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 415, + 287, + 458 + ], + "spans": [ + { + "bbox": [ + 49, + 415, + 287, + 458 + ], + "type": "text", + "content": "[39] Wouter Van Gansbeke, Simon Vandenhende, Stamatios Georgoulis, and Luc V Gool. Revisiting contrastive methods for unsupervised learning of visual representations. In NeurIPS, pages 16238-16250, 2021. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 49, + 460, + 287, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 460, + 287, + 502 + ], + "spans": [ + { + "bbox": [ + 49, + 460, + 287, + 502 + ], + "type": "text", + "content": "[40] Wouter Van Gansbeke, Simon Vandenhende, Stamatios Georgoulis, and Luc Van Gool. Unsupervised semantic segmentation by contrasting object mask proposals. In ICCV, pages 10032-10042, 2021. 1, 2, 5, 7, 8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 49, + 504, + 287, + 556 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 504, + 287, + 556 + ], + "spans": [ + { + "bbox": [ + 49, + 504, + 287, + 556 + ], + "type": "text", + "content": "[41] Yael Vinker, Ehsan Pajouheshgar, Jessica Y. Bo, Roman Christian Bachmann, Amit Haim Bermano, Daniel Cohen-Or, Amir Zamir, and Ariel Shamir. CLIPasso: Semantically-aware object sketching. ACM TOG, 41(4):86:1-86:11, 2022. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 49, + 559, + 287, + 591 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 559, + 287, + 591 + ], + "spans": [ + { + "bbox": [ + 49, + 559, + 287, + 591 + ], + "type": "text", + "content": "[42] Carl Vondrick, Abhinav Shrivastava, Alireza Fathi, Sergio Guadarrama, and Kevin Murphy. Tracking emerges by colorizing videos. In ECCV, pages 402-419, 2018. 8" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 49, + 592, + 287, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 592, + 287, + 624 + ], + "spans": [ + { + "bbox": [ + 49, + 592, + 287, + 624 + ], + "type": "text", + "content": "[43] Xiaolong Wang, Allan Jabri, and Alexei A. Efros. Learning correspondence from the cycle-consistency of time. In CVPR, pages 2561-2571, 2019. 8" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 49, + 625, + 287, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 625, + 287, + 667 + ], + "spans": [ + { + "bbox": [ + 49, + 625, + 287, + 667 + ], + "type": "text", + "content": "[44] Zhaoqing Wang, Yu Lu, Qiang Li, Xunqiang Tao, Yandong Guo, Mingming Gong, and Tongliang Liu. CRIS: Clip-driven referring image segmentation. In CVPR, pages 11676-11685, 2022. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 49, + 669, + 287, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 669, + 287, + 690 + ], + "spans": [ + { + "bbox": [ + 49, + 669, + 287, + 690 + ], + "type": "text", + "content": "[45] Saining Xie and Zhuowen Tu. Holistically-nested edge detection. In ICCV, pages 1395-1403, 2015. 6" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 49, + 692, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 692, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 49, + 692, + 287, + 713 + ], + "type": "text", + "content": "[46] Jiarui Xu, Shalini De Mello, Sifei Liu, Wonmin Byeon, Thomas Breuel, Jan Kautz, and Xiaolong Wang. GroupViT:" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 185 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "type": "text", + "content": "Semantic segmentation emerges from text supervision. In CVPR, pages 18134-18144, 2022. 2, 3, 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 96, + 545, + 129 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 96, + 545, + 129 + ], + "spans": [ + { + "bbox": [ + 307, + 96, + 545, + 129 + ], + "type": "text", + "content": "[47] Xiao Zhang and Michael Maire. Self-supervised visual representation learning from hierarchical grouping. In NeurIPS, pages 16579-16590, 2020. 2, 6, 7, 8" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 130, + 545, + 163 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 130, + 545, + 163 + ], + "spans": [ + { + "bbox": [ + 307, + 130, + 545, + 163 + ], + "type": "text", + "content": "[48] Hengshuang Zhao, Jianping Shi, Xiaojuan Qi, Xiaogang Wang, and Jiaya Jia. Pyramid scene parsing network. In CVPR, pages 6230-6239, 2017. 1, 5, 6, 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 164, + 545, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 164, + 545, + 185 + ], + "spans": [ + { + "bbox": [ + 308, + 164, + 545, + 185 + ], + "type": "text", + "content": "[49] Chong Zhou, Chen Change Loy, and Bo Dai. Extract free dense labels from CLIP. In ECCV, 2022. 1, 2, 3, 4, 5, 6, 7, 8" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "text", + "content": "11216" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/CLIP-Sculptor_ Zero-Shot Generation of High-Fidelity and Diverse Shapes From Natural Language/e64fac6d-0780-4018-a6c0-4f909bfb4337_content_list.json b/2023/CLIP-Sculptor_ Zero-Shot Generation of High-Fidelity and Diverse Shapes From Natural Language/e64fac6d-0780-4018-a6c0-4f909bfb4337_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..b3d1c5b8b6033a84294c815e1279c5c925dd4820 --- /dev/null +++ b/2023/CLIP-Sculptor_ Zero-Shot Generation of High-Fidelity and Diverse Shapes From Natural Language/e64fac6d-0780-4018-a6c0-4f909bfb4337_content_list.json @@ -0,0 +1,1228 @@ +[ + { + "type": "text", + "text": "CLIP-Sculptor: Zero-Shot Generation of High-Fidelity and Diverse Shapes from Natural Language", + "text_level": 1, + "bbox": [ + 78, + 128, + 888, + 175 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Aditya Sanghi† Rao Fu‡ Vivian Liu§ Karl D.D. Willis† Hooman Shayani† Amir H. Khasahmadi† Srinath Sridhar‡ Daniel Ritchie‡ Autodesk Research† Brown University‡ Columbia University§", + "bbox": [ + 135, + 200, + 841, + 257 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/af57706f733e87fff2175388558c77f34e475f0ab3b75f9eb54c922df9ff6187.jpg", + "image_caption": [ + "Figure 1. CLIP-Sculptor is a zero-shot text-to-shape generation method. Left Top/Bottom: It generates diverse shapes that reflect the semantic meaning of the text input without requiring any text data during training. Right: The method also generates high-fidelity shapes." + ], + "image_footnote": [], + "bbox": [ + 76, + 290, + 630, + 452 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/053674f60e974e7a9db3708cd195027dec70fd24adfb6edc9f662af9bafbbe9c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 635, + 292, + 890, + 369 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/8867c4b6a962f70f75737786dbb4d89448714c10757f7851ef4c0ac23c93e158.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 635, + 377, + 890, + 450 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 513, + 313, + 530 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recent works have demonstrated that natural language can be used to generate and edit 3D shapes. However, these methods generate shapes with limited fidelity and diversity. We introduce CLIP-Sculptor, a method to address these constraints by producing high-fidelity and diverse 3D shapes without the need for (text, shape) pairs during training. CLIP-Sculptor achieves this in a multi-resolution approach that first generates in a low-dimensional latent space and then upscales to a higher resolution for improved shape fidelity. For improved shape diversity, we use a discrete latent space which is modeled using a transformer conditioned on CLIP's image-text embedding space. We also present a novel variant of classifier-free guidance, which improves the accuracy-diversity trade-off. Finally, we perform extensive experiments demonstrating that CLIP-Sculptor outperforms state-of-the-art baselines.", + "bbox": [ + 73, + 546, + 473, + 789 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 814, + 209, + 830 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In recent years, there has been rapid progress in text-conditioned image generation [31, 32, 35], which has been driven by advances in multimodal understanding learned from web-scaled paired (text, image) data. These advances", + "bbox": [ + 75, + 839, + 470, + 902 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "have led to applications in domains ranging from content creation [21, 22, 31] to human-robot interaction [40]. Unfortunately, developing the analogue of a text-conditioned 3D shape generator is challenging because it is difficult to obtain (text, 3D shape) pairs at large scale. Prior work has attempted to address this problem by collecting text-shape paired data [2, 5, 9, 23, 26], but these approaches have been limited to a small number of object categories.", + "bbox": [ + 496, + 516, + 892, + 638 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "A promising way around this data bottleneck is to use weak supervision from large-scale vision/language models such as CLIP [30]. One approach is to directly optimize a 3D representation such that (text, image render) pairs are aligned when projected into the CLIP embedding space. Prior work has applied this approach to stylize 3D meshes [25, 41] and to create abstract \"dreamlike\" objects using neural radiance fields [18] or meshes [19]. However, neither of the aforementioned methods produce realistic object geometry, and they can require expensive optimization. Another approach, more in line with text-to-image generators [31, 32], is to train a conditional generative model. The CLIP-Forge system [37] builds such a model without paired (text, shape) data by using rendered images of shapes at training time and leveraging the CLIP embedding space to bridge the modalities of image and text at inference time. CLIP-Forge demonstrates compelling zero-shot generation", + "bbox": [ + 496, + 643, + 893, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 44 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "18339", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 0 + }, + { + "type": "table", + "img_path": "images/29d39bd1a16ea94f48a8966f9eebfcae48acbf9b30970907a62e81e76f8abe76.jpg", + "table_caption": [ + "Table 1. High-level comparison between zero-shot text-to-shape generation methods. Inference time is calculated using a single NVIDIA Tesla V100 GPU." + ], + "table_footnote": [], + "table_body": "
MethodInferenceFidelityDiversity
DreamFields [18]>24 hrsLowSingle
Clip-Mesh [19]30 minLowSingle
Clip-Forge [37]6.41 msMediumMedium
CLIP-Sculptor0.91 secHighHigh
", + "bbox": [ + 80, + 142, + 467, + 234 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "abilities but produces low-fidelity shapes which do not capture the full diversity of shapes found in the training distribution.", + "bbox": [ + 75, + 258, + 467, + 303 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we propose CLIP-Sculptor, a text-conditioned 3D shape generative model that outperforms the state of the art by improving shape diversity and fidelity with only (image, shape) pairs as supervision. CLIP-Sculptor's novelty lies in its multi-resolution, voxel-based conditional generation scheme. Without (text, shape) pairs, CLIP-Sculptor learns to generate 3D shapes of common object categories from the text-image joint embedding of CLIP. To achieve high fidelity outputs, CLIP-Sculptor adopts a multi-resolution approach: it first generates a low-resolution latent grid representation that captures the semantics from text/image, then upscales it to a higher resolution latent grid representation with a super-resolution model, and finally decodes the output geometry. To generate diverse shapes, CLIP-Sculptor adopts discrete latent representations which are obtained using a vector quantization scheme that avoids posterior collapse. To further improve shape fidelity and diversity, CLIP-Sculptor uses a masked transformer architecture. We additionally propose a novel annealed strategy for the classifier-free guidance [16]. To sum up, we make the following contributions:", + "bbox": [ + 75, + 304, + 468, + 621 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- CLIP-Sculptor, a multi-resolution text-conditional shape generative model that achieves both high fidelity and diversity without the need for (text, shape) pairs.", + "- A novel variant of classifier-free guidance for generative models, with an annealed guidance schedule to achieve better quality for a given diversity level." + ], + "bbox": [ + 96, + 631, + 467, + 731 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 76, + 753, + 218, + 768 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Neural Discrete Representation. Discrete representations [28] were first proposed for image generation in the context of variational autoencoders as a method to improve image quality and avoid posterior collapse. Further work introduced multi-scale and hierarchical variants [8, 33] which improved generative capabilities. Recently, 3D discrete representations have emerged as well, such as discretized voxel and implicit grids [26, 42]. In this work, we take inspiration", + "bbox": [ + 75, + 779, + 467, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "from hierarchical VQ-VAEs [8, 33] and propose an architecture of hierarchical discrete representations capable of generating high fidelity 3D shapes.", + "bbox": [ + 498, + 90, + 890, + 137 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Latent Generative Models. In recent years, latent generative models have been widely adopted, because these models can effectively generate low-dimensional latent representations which can be used to generate high fidelity images and shapes. In the 3D domain, GAN-based latent models [1, 6, 17] have shown impressive results but tend to suffer from training instability and mode collapse. Flow-based models [37, 43] have been proposed to alleviate these problems, but they yield sample quality that is inferior to GAN-based models. Recent works in the image domain use diffusion [12, 36] or masking models [4] on the latent space which can increase the inference efficiency and still give quality outputs. Building on these works, we propose a hierarchical latent generative model that can further improve the fidelity and diversity of shape generations.", + "bbox": [ + 496, + 154, + 890, + 380 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Classifier-free Guidance. Classifier-free guidance, proposed by [16], jointly trains a conditional and unconditional generative model, whose score estimates are mixed to establish a trade-off between diversity measured by Frechet inception distance (FID) and sample accuracy/fidelity measured by Inception score (IS). Classifier-free guidance guides unconditional samples in the direction of conditional ones and has been implemented in practice to achieve state of the art results in recent work [10-12, 34]. To the best of our knowledge, guidance scale has always been kept constant outside the concurrent exploration in [15]. Their oscillating guidance technique, however does not show improvements in sample fidelity and produces more artifacts in their generations. In this work, we find that rather using annealed scheduling can give better quality versus diversity trade-offs.", + "bbox": [ + 496, + 398, + 890, + 640 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Text-to-Shape Generation. Text-to-shape generation has gained momentum in recent years. Recent works [5, 9, 23, 26] use supervised text-shape pairs to generate shapes effectively using natural language. However, a major drawback is the lack of available text-shape datasets, which limits supervised methods to generate shapes only in few categories. To solve this, several recent works have successfully leveraged the image-text embeddings of CLIP as a prior [30] by converting shapes into images using renders. One line of work uses differentiable renderers [18, 19, 25, 29, 41]. Another line of work learns a mapping from the image space to the shape space [37]. Although these methods are effective, they suffer from either long optimization time [18, 19, 25, 29] or limited quality of generated shapes [18, 19, 37]. Our method is able to generate more diverse shapes of higher fidelity within a comparably short inference time.", + "bbox": [ + 496, + 657, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "18340", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/b6b086398f90e1e54f1ccc07d1fea640e8f0c3159632c3ce5ccbaa9d0cb54d98.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 76, + 88, + 511, + 321 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/e5f560aab1b10a668d6587feb3f86cf3929f5075a0f1b53a86e7a5f887fcc25e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 514, + 88, + 890, + 321 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/3a5c639818d65f0efcc63f11a4cc592b7513cbab1bd81100d3ba95dd12ac5634.jpg", + "image_caption": [ + "Figure 2. The CLIP-Sculptor architecture during training (top) and inference (bottom). CLIP-Sculptor is trained in three stages. In Stage 1, we train two separate VQ-VAE models for $32^{3}$ and $64^{3}$ resolution voxel grids. In Stage 2 we train a coarse transformer conditioned on a CLIP embedding to generate low resolution VQ-VAE latent grids $\\mathbf{E}_{32}$ . In Stage 3, we train a fine transformer to perform super resolution on these latent grids. During inference, a text prompt is passed through the CLIP text encoder and used to condition the coarse transformer to generate a coarse latent grid $\\mathbf{E}_{32}$ . This coarse grid is then used to condition the fine transformer to generate a fine latent grid $\\mathbf{E}_{64}$ . Finally, this fine latent grid is then passed through the Training Stage 1 $64^{3}$ VQ-VAE decoder to generate the output shape." + ], + "image_footnote": [], + "bbox": [ + 78, + 321, + 890, + 452 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 76, + 585, + 168, + 599 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "CLIP-Sculptor aims to generate diverse and high-fidelity 3D shapes of common object categories from text prompts. Figure 2 illustrates the components of our CLIP-Sculptor model. We train the network using 3 stages. In the first stage, we use a vector-quantized representation [28], to efficiently represent 3D shapes in discrete voxel grids. This latent representation captures the diverse semantic meanings of the 3D shapes in common object categories and is compatible with powerful generative models. To balance computation cost and shape quality, we utilize latent representations at two resolutions. A low resolution latent representation $\\mathbf{E}_{32}$ encodes geometric information of voxels $\\mathbf{V}_{32}$ at resolutions of $32^{3}$ , which is useful for text-conditional generation. A high resolution latent representation $\\mathbf{E}_{64}$ encodes geometric information of voxel grids $\\mathbf{V}_{64}$ at resolutions of $64^{3}$ , which is useful for super-resolution and detail generation. In stage 2, we train a generative model conditioned on CLIP features obtained from images $\\{\\mathbf{I}_r|r\\in R\\}$ rendered at random views $R$ to solve the lack of paired (text, shape)", + "bbox": [ + 75, + 613, + 473, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "data. The conditional generative model consists of a mapping net and a coarse transformer $T_{c}(.)$ that is trained to unmask a masked low-resolution grid $\\mathbf{E}_{32}$ . Finally in stage 3, a fine transformer $T_{f}(.)$ is used as a super-resolution model in latent space to generate higher resolution shapes. It is trained in a similar manner to the coarse transformer but conditioned on a low-resolution grid $\\mathbf{E}_{32}$ instead of CLIP features.", + "bbox": [ + 496, + 587, + 890, + 707 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "During inference, a text prompt is sequentially converted from CLIP text embedding to a low-resolution latent representation $\\mathbf{E}_{32}$ to a high-resolution latent representation $\\mathbf{E}_{64}$ to finally a high-resolution voxel grid $\\mathbf{V}_{64}$ . A confidence-based iterative decoding scheme [4] is adopted for unmasking the fully masked initial grids for coarse and fine transformers at inference. To further control shape fidelity and diversity, we propose an annealing strategy for classifier-free guidance at inference time.", + "bbox": [ + 496, + 713, + 890, + 849 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In the following subsections, we will introduce the major components within CLIP-Sculptor. Section 3.1 describes the vector quantized representations at multiple resolutions;", + "bbox": [ + 498, + 854, + 892, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "18341", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Section 3.2 explains the conditional generation model as well as its design choices and training strategies; Section 3.3 introduces the super-resolution model; Section 3.4 explains how the classifier-free guidance is achieved at inference using an annealing strategy.", + "bbox": [ + 75, + 90, + 468, + 167 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1. Multi-Resolution Voxel VQ-VAEs", + "text_level": 1, + "bbox": [ + 76, + 174, + 372, + 191 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To learn a low-dimensional latent space that effectively represents the distribution of the training shape set $D$ , we use a vector-quantized variational autoencoder (VQVAE) [28]. The vector quantized latent representation avoids \"posterior collapse\", and it is also amenable to modeling with transformer-based generative models. To balance computational cost and shape quality, we train two separate VQ-VAEs for $32^{3}$ and $64^{3}$ resolution voxel grids $\\mathbf{V}_{32}$ and $\\mathbf{V}_{64}$ , which are useful for conditional generation and superresolution respectively.", + "bbox": [ + 75, + 198, + 468, + 348 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For both VQ-VAEs, we use ResNet-based [13] volumetric CNN encoders and decoders. A vector quantization layer maps the down-sampled voxel grids into a discrete latent space by indexing the continuous latent space with an embedding codebook. In this way, the voxel grids $\\mathbf{V}_{32}$ and $\\mathbf{V}_{64}$ are effectively represented by low-dimensional latent representations $\\mathbf{E}_{32}$ and $\\mathbf{E}_{64}$ , at the resolution of $4^{3}$ and $8^{3}$ . The VQ-VAEs are trained with mean squared error loss (MSE), commitment loss [28], and an exponential moving average strategy [33, 45].", + "bbox": [ + 75, + 349, + 468, + 501 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. CLIP-Conditioned Coarse Transformer", + "text_level": 1, + "bbox": [ + 76, + 508, + 418, + 523 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The conditional generation process aims to generate diverse shapes that semantically correspond to the text prompts. Inspired by the recent success of masking approaches [4] and diffusion models [12, 27, 36], we formulate this generation process as a conditional unmasking task. Given an input masked latent representation $\\mathbf{E}_{32}^{msk}$ and a conditional vector $\\mathbf{c}$ , the coarse transformer $T_{c}(.)$ should produce an fully unmasked latent representation $\\mathbf{E}_{32}$ that not only aligns with the semantic meaning captured by the conditional vector, but also allows for the sampling of diverse shapes at inference.", + "bbox": [ + 75, + 532, + 468, + 698 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The conditional vector $\\mathbf{c}$ provides the semantic guidance for the shape generation process. Since we assume no text data at the training time, we leverage the text-image joint-embedding of CLIP. The generation is guided by the CLIP image embedding from the rendered images $\\{\\mathbf{I}_r\\}$ during training and guided by the CLIP text embedding from the text prompt during inference. We add Gaussian noise to the CLIP image embeddings [44] to alleviate the modality gap between the CLIP text and image embeddings [20]:", + "bbox": [ + 75, + 699, + 468, + 834 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathbf {c}} = f _ {I} \\left(\\mathbf {I} _ {r}\\right) + \\frac {\\gamma \\cdot \\epsilon \\cdot \\| f _ {I} \\left(\\mathbf {I} _ {r}\\right) \\| _ {2}}{\\| \\epsilon \\| _ {2}}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 166, + 843, + 468, + 876 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $f_{I}(\\cdot)$ is the CLIP image encoder, $\\epsilon \\sim \\mathcal{N}(0,1)$ , and $\\gamma$", + "bbox": [ + 76, + 885, + 468, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "controls the level of perturbation. The conditional vector $\\hat{\\mathbf{c}}$ is then normalized and passed through a mapping network that consists of multi-layer perceptrons (MLP). The mapped vector $\\tilde{\\mathbf{c}}$ conditions on the coarse transformer $T_{c}(.)$ by predicting the affine transform parameters of each transformer block's layernorm. In order to control the diversity-quality balance of the generated shapes, we replace the conditional vector with the null embedding by a $\\rho\\%$ drop out rate for classifier-free guidance [16]. These are important design choices that affects the diversity and quality of the generated shapes at inference time.", + "bbox": [ + 496, + 90, + 890, + 256 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To train the coarse transformer $T_{c}(.)$ , we randomly replace the input indices with a special \"mask\" tokens by uniformly selecting a masking ratio between 0-100% [4]. Given the masked latent grid $\\mathbf{E}_{32}^{msk}$ and the conditional vector $\\mathbf{c}$ , the training objective is to maximize the following log-likelihood: $\\sum_{n}^{N}\\log p(\\mathbf{E}_{32,n}|\\mathbf{E}_{32,n}^{msk},\\mathbf{c}) / N$ , where $N$ is the training data size and $\\mathbf{E}_{32}^{msk}$ is the randomly masked latent grids. As the network sees only the ground truth tokens but not its own predicted samples during training, there exists an autoregressive drift resulted by the accumulative error at inference-time sampling. To alleviate this issue, we take inspiration from the NLP literature [38] and use a similar step-unrolled training (SUT) that predicts the entire un-masked grids at each time step. The final objective is to maximize the following formulation:", + "bbox": [ + 496, + 257, + 890, + 484 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\sum_ {n} ^ {N} \\left(\\log p \\left(\\mathbf {E} _ {3 2, n} \\mid \\mathbf {E} _ {3 2, n} ^ {m s k}, \\mathbf {c}\\right) + \\log p \\left(\\mathbf {E} _ {3 2, n} \\mid \\tilde {\\mathbf {E}} _ {3 2, n}, \\mathbf {c}\\right)\\right) / N, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 501, + 494, + 890, + 547 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\tilde{\\mathbf{E}}_{32}$ is the coarse transformer $T_{c}(.)$ 's own prediction.", + "bbox": [ + 500, + 547, + 888, + 564 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Super-Resolution with Fine Transformer", + "text_level": 1, + "bbox": [ + 500, + 573, + 848, + 589 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Super-resolution in the discrete latent space improves quality and upsamples the resolution of generated shapes. This process is done with a fine transformer $T_{f}(.)$ , which takes a low-resolution latent representation $\\mathbf{E}_{32}$ and a masked high-resolution latent grid as input. The fine transformer is conditioned on the coarse grid $\\mathbf{E}_{32}$ via cross attention. It unmasks the high-resolution representation $\\mathbf{E}_{64}$ . To ensure that training data resembles the inference data, we directly use the predictions $\\tilde{\\mathbf{E}}_{32}$ of the coarse transformer $T_{c}(.)$ to train the fine transformer $T_{f}(.)$ , instead of using the ground truth low-resolution latent representation $\\mathbf{E}_{32}$ .", + "bbox": [ + 496, + 597, + 890, + 763 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.4. Annealed Classifier-Free Guidance", + "text_level": 1, + "bbox": [ + 500, + 771, + 803, + 786 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "At inference time, we first convert an input text prompt into a CLIP text embedding using the CLIP text encoder $f_{T}(\\cdot)$ . The text embedding is then fed into the mapping network and used as the conditional vector $\\tilde{\\mathbf{c}}$ of the coarse transformer $T_{c}(\\cdot)$ . We use the iterative decoding scheme [4] to slowly unmask the grid over a sequence of $T$ steps. The initial input to the coarse transformer $T_{c}(\\cdot)$ is a completely", + "bbox": [ + 496, + 795, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "18342", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "masked latent grid. At each time step $t$ , we condition the coarse transformer $T_{c}(.)$ with $\\tilde{\\mathbf{c}}$ and take the output latent grid $\\mathbf{E}_{32,t-1}^{msk}$ from the prior time step as the input. The coarse transformer $T_{c}(.)$ will predict an unmasked latent grid $\\tilde{\\mathbf{E}}_{32,t}$ . We mask all the tokens except for the most confident token predictions and the tokens unmasked from previous steps. This newly masked latent grid $\\mathbf{E}_{32,t}^{msk}$ is the input to the next time step $t+1$ . We repeat this process until the process unmasks all the tokens, which is ensured by a cosine masking schedule [4].", + "bbox": [ + 75, + 90, + 472, + 241 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "During this iterative decoding scheme, we apply a new variant of classifier-free guidance [16]. Classifier-free guidance extrapolates an unconditional sample in the direction of a conditional sample. The extrapolation is controlled by an adaptive guidance scale parameter. Varying this parameter results in a trade-off between accuracy of text-to-shape generation and sample diversity. Instead of using a constant guidance scale for all time steps [16], we propose to vary the guidance scale over time according to an annealing schedule. The intuition is that during the initial steps of sampling when the input to the coarse transformer is mostly masked indices, a larger guidance scale is important to keep the network \"on task\"; later on in the sampling process, a lower guidance scale can help produce more sample diversity. Overall, the guidance is formulated as:", + "bbox": [ + 75, + 243, + 472, + 470 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {P} _ {t} (c) = P _ {t} (0) + a (t) \\left(P _ {t} (c) - P _ {t} (0)\\right), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 138, + 484, + 468, + 510 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $a(t)$ is the guidance scale annealing schedule, $P_{t}(c)$ denotes the conditional distribution $p(\\mathbf{E}_{32,t}|\\mathbf{E}_{32,t - 1}^{msk},\\mathbf{c})$ and $P_{t}(0)$ denotes $p(\\mathbf{E}_{32,t}|\\mathbf{E}_{32,t - 1}^{msk},\\mathbf{0})$ . Note that $a(t)$ is continuous and usually monotonically decreasing. By setting $a(t) = k$ for some constant $k$ , this equation is equal to the classifier-free guidance scheme with a constant guidance scale. In this paper, we experimentally evaluate other annealing schedules [4], including linear, cosine, and square root functions.", + "bbox": [ + 75, + 520, + 468, + 657 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Finally, the unmasked coarse latent grid $\\tilde{\\mathbf{E}}_{32}$ is fed to the fine transformer $T_{f}$ , which uses the iterative decoding scheme to unmask an initially masked fine latent grid $\\mathbf{E}_{64}^{msk}$ . The unmasked fine grid is then passed to the $64^3$ VQ-VAE decoder to obtain the final high-resolution voxel shape.", + "bbox": [ + 75, + 659, + 468, + 736 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 76, + 753, + 209, + 768 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this section, we report experimental results to evaluate the generation fidelity, diversity, and class accuracy of CLIP-Sculptor. We provide full details of the experiment setup in the supplementary. We run the experiments 3 times for each of the aforementioned measures and report the mean in each case, except for the final comparisons with baselines (Table 2), where we use the best seed. Additional results can also be found in the supplementary.", + "bbox": [ + 75, + 780, + 468, + 902 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Dataset. We conduct our experiments on two subsets of the ShapeNet(v2) dataset [3]. The first subset, ShapeNet13, contains 13 categories from ShapeNet as used in [7,24]. We use the same train/test split as specified in [24]. Our second subset is ShapeNet55 which contains all 55 ShapeNet categories. For ShapeNet55, we render images as described in [7]; the training dataset contains 51,784 datapoints, whereas the test set contains 6,101 datapoints.", + "bbox": [ + 496, + 90, + 890, + 210 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Evaluation Metrics. To assess the generative capabilities of CLIP-Sculptor, we use Fréchet Inception Distance (FID) ([14]) to evaluate shape diversity and quality, and use Classifier Accuracy (Acc) to evaluate the text-shape correspondence. These metrics follow [37] where they first generated single mean shapes for 234 predetermined text queries and then passed all shapes through a classifier to measure Acc. The latent space of this classifier is also used for FID.", + "bbox": [ + 496, + 212, + 890, + 332 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Baseline. We compare the performance of our method against CLIP-Forge [37], CLIP-Mesh [19] and DreamFields [18], which are currently state of the art for zero-shot text-to-shape generation. We also set up a comparison method called Zero-Shot AutoSDF(ZS-ASDF), which trains the supervised method AutoSDF [26] with the CLIP features as CLIP-Sculptor, to highlight the superiority of our conditional generation model. CLIP-Forge reports results on single shape generation for 234 text queries. These single shapes were generated using the mean of the prior (the Gaussian distribution), so we refer to these results as CF-MS. Note that this does not capture the diversity of shapes that can be generated given a text query. To capture results on more shape generations, we sample 32 shapes instead of one using a Gaussian (CF-G), truncated Gaussian (CF-TG) or clipped Gaussian (CF-CG) distribution. We also generate 32 shapes for ZS-ASDF and our method for fair comparison. We only compare qualitatively with DreamFields and CLIP-Mesh as the optimization time per query is significant (Table 1), and it does not use prior knowledge from the shape dataset.", + "bbox": [ + 496, + 334, + 892, + 652 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Evaluating Shape Diversity and Accuracy", + "text_level": 1, + "bbox": [ + 498, + 664, + 854, + 681 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this section, we first quantitatively evaluate the diversity and accuracy of a given shape matching a given text query on the ShapeNet13 dataset. We compare with CLIP-Forge and ZS-ASDF using the Acc and FID metrics. The results are shown in Table 2. The first four columns represent CLIP-Forge(CF) and its different sampling techniques. The fifth column represents the result from AutoSDF trained with CLIP features(ZS-ASDF). The other columns represent our method with different annealing scale strategies. Three major things can be observed. First, all variants of our method outperform CLIP-Forge and ZS-ASDF significantly. This indicates that the method produces more diverse and higher fidelity shapes of increasing accuracy. Second, it can be seen that annealing strategies", + "bbox": [ + 496, + 688, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "18343", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/4d2b11d33bf405305c5d336a25d1e06a4cdb9aa618c6931d990ba514a23a45c7.jpg", + "image_caption": [ + "Figure 3. Qualitative comparison of CLIP-Sculptor (rendered in purple), CLIP-Forge [37](rendered in green), DreamField [18], and CLIP-Mesh [19] on text-conditioned generation. For each text prompt, the first generated shape by CLIP-Sculptor and CLIP-Forge [37] is zoomed in for detail comparison. Comparing with other methods, CLIP-Sculptor generates a broader category of shapes that correspond with the text prompt. The generated shapes are more diverse and has higher fidelity." + ], + "image_footnote": [], + "bbox": [ + 81, + 85, + 916, + 714 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "for guidance scale give a better accuracy-diversity trade-off than constant scale guidance. Finally, we note that simply augmenting AutoSDF (another VQ-VAE) with CLIP features does not lead to improved text to shape generation, indicating the importance of our architecture design decisions.", + "bbox": [ + 75, + 806, + 472, + 898 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We also qualitatively compare our method to CLIP-Forge, CLIP-Mesh, and DreamFields. The results are shown in Figure 3. It can be observed that our method generates higher quality shapes (view \"a machine gun\"), with more detail (view \"an office chair\") and higher diversity (view \"a jet\"). CLIP-Forge usually produces the same shape", + "bbox": [ + 496, + 806, + 893, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "18344", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/8791f35eb4d219269e9aa37cbb20db618588cb99f7ca517b68cd3966ae76930d.jpg", + "table_caption": [ + "Table 2. Comparisons of CLIP-Forge(CF) baseline (across different sampling strategies), AutoSDF augmented with CLIP (ZS-ASDF) with CLIP-Sculptor (CS) on Accuracy(ACC) and FID. CLIP-Sculptor outperforms other methods in both metrics, which proves it generates shapes with higher fidelity and better diversity." + ], + "table_footnote": [], + "table_body": "
MethodCF-MSCF-GCF-TGCF-CGZS-ASDFCS-const.CS-sqrtCS-linearCS-cosine
FID ↓2425.252233.482141.612100.677332.931821.781480.111629.511725.63
ACC ↑83.3362.8168.7171.1139.1486.5987.0887.5087.27
", + "bbox": [ + 78, + 154, + 893, + 219 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/01582681405a9c0c8b21c4494d407f37e9a4e7b6e501cc6564226346267ef213.jpg", + "table_caption": [ + "Table 3. Ablations on the major components of CLIP-Sculptor. CLIP-Sculptor achieves the best performance comparing with other design choice, which proves the effectiveness of the major components.", + "(a) Effect of varying the Noise parameter." + ], + "table_footnote": [], + "table_body": "
γ×0.50.81.01.21.5
FID↓1720.021764.981484.611703.381447.911478.17
ACC↑64.8775.7377.4179.0979.6378.47
(b) Effect of varying number of layers (L) in the mapping network.
L012345
FID↓2874.871716.731518.971447.911532.501424.46
ACC↑62.7278.7079.1779.6377.1676.09
(c) Comparison with baselines on super resolution.
Method3D-UNet643-DSCLIP-Sculptor
FID↓2056.922196.961910.28
ACC↑86.6577.9286.85
", + "bbox": [ + 86, + 335, + 478, + 551 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "with small variations (view \"a rectangular table\"). We also note that DreamFields and CLIP-Mesh produce very abstract results which may not be useful in many applications. Finally, we also show results on ShapeNet55 in the last row of Figure 3. We could not get CLIP-Forge to produce sensible shapes for most text queries on ShapeNet55 (which we attribute to the data imbalance issue of ShapeNet55), whereas our method produced high fidelity shapes.", + "bbox": [ + 75, + 577, + 468, + 699 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. Major Components for Conditional Generation", + "text_level": 1, + "bbox": [ + 76, + 709, + 468, + 727 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Effect of Noise Parameter. During Stage 2 Training, to better align the image and text embeddings, we added Gaussian noise at varying levels of $\\gamma$ . We investigate the effect of $(\\gamma)$ added to the image condition vector and show the results in Table 3 (a). We keep the number of mapping layers fixed at three in this experiment. We observe that adding Gaussian noise drastically improves both the diversity and accuracy of generation, with the optimal noise parameter being around 1.2. These results indicate that adding noise during training helps align the text features seen at inference and the image features seen at training.", + "bbox": [ + 75, + 734, + 470, + 902 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Size of Mapping Network. We next probe the importance of the mapping network. We show the results in Table 3 (b), where $L$ represents the number of layers of the mapping function. In the case of 0 layers, we directly project conditional embeddings to the layernorm parameters using a linear layer. We make two observations from the results: 1) A common mapping network improves both the accuracy and FID. 2) Increasing the number of mapping network layers beyond a certain number decreases accuracy.", + "bbox": [ + 496, + 253, + 890, + 390 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Classifier-Free Guidance. We next explore the relationship between dropping out image conditioning at $\\rho \\%$ during training and using constant classifier-free guidance during test time. In Table 4, we vary the scale parameter, $a(t) = k$ , across the columns. It can be seen that as we increase the scale, accuracy typically increases while FID decreases. This indicates that the method is giving more accurate results on a given text query at the cost of shape diversity. A low dropout rate (5-15%) gives a good trade-off between accuracy and diversity.", + "bbox": [ + 496, + 391, + 892, + 542 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Step-Unrolled Training (SUT). Finally, to further improve the accuracy we investigate the use of step-unrolled training [39]. The results are shown in the last row of Table 4. From the table it can be observed that step-unrolled training enables higher accuracy across all scales. This indicates that the model learns to unmask samples it would encounter during inference time.", + "bbox": [ + 496, + 544, + 893, + 650 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Annealing Strategy for Scale Parameter", + "text_level": 1, + "bbox": [ + 498, + 664, + 841, + 681 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "An important idea we propose in this paper is a scale annealing technique for classifier-free guidance. We employ classifier-free guidance to identify a better accuracy-diversity trade-off with different annealing schedules: constant, linear, cosine and square root. To determine which annealing technique works the best, we fix the FID values and plotted the accuracy at each fixed FID value (Figure 4). As different scale parameters give different FID values, we conducted an extensive grid search over the starting scale parameter. Note that finding the exact FID is not always feasible, so we pick the closest FID. Figure 4 shows results of FID versus accuracy on three different runs of the Coarse Transformer. We find that across all three runs, the accuracy is typically lower for a given FID for constant schedules as", + "bbox": [ + 496, + 689, + 893, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "18345", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/9a6179151f3b0838fce1ae66d3985d6efa9254938e7e27275854c9776766b124.jpg", + "table_caption": [ + "Table 4. Classifier-Free Guidance and Step-Unrolled Training (SUT) experiment results. $\\rho \\%$ is the drop out rate for classifier-free guidance. SUT indicates the use of Step-Unrolled Training. The remaining columns indicate the variation in scale parameter." + ], + "table_footnote": [], + "table_body": "
ρ%SUTa(t) = 3a(t) = 2.5a(t) = 2a(t) = 1.5a(t) = 1
FID↓Acc↑FID↓Acc↑FID↓Acc↑FID↓Acc↑FID↓Acc↑
5×2059.085.731970.185.471790.985.431536.583.981227.579.17
10×1893.284.461821.284.501684.383.411522.482.341348.477.76
15×2086.985.031964.784.991851.984.361660.983.141485.578.19
20×2062.583.391972.382.951892.982.741733.381.131566.675.94
5✓2039.887.692011.187.391811.887.401678.686.241517.982.18
", + "bbox": [ + 76, + 128, + 898, + 239 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/0cd7ff696dfc73d7523e291fc7cf6e556d36b6e5463c94b1fde9453ba86509ce.jpg", + "image_caption": [ + "Figure 4. FID and Acc results with different classifier-free guidance scale annealing strategies (constant, cosine, square root, linear) across three different runs (different seeds) of the CLIP-Sculptor Stage 2 Coarse Transformer." + ], + "image_footnote": [], + "bbox": [ + 86, + 253, + 364, + 396 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/b761d6b05adefe8d1aa43a87eea9f07b6b2b79c107a38e92ae487cdbdeabaad9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 370, + 253, + 633, + 396 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/b567cbab04f84aedc64cb29439a8d1f655ece0eb4525882548c2d908d9c54e6b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 638, + 255, + 890, + 396 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "opposed to other schedules. This is especially the case at the lower range of FID values. The results indicate that having a large scale at the beginning of sampling is more important than later stages, especially in use cases where diversity is paramount.", + "bbox": [ + 75, + 465, + 468, + 542 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4. Super-Resolution", + "text_level": 1, + "bbox": [ + 76, + 550, + 246, + 566 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Finally, we investigate the importance of hierarchical super-resolution. We compare our method with two baselines. In the first baseline, we directly use the $32^{3}$ resolution results from the coarse transformer and use a 3D-UNET based super-resolution network to translate from $32^{3}$ to $64^{3}$ . For the second baseline, we train a transformer directly on $64^{3}$ VQ-VAE that is conditioned on text features as opposed to a coarse resolution grid. We refer to this as $64^{3}$ -DS (64 $^{3}$ direct synthesis). The results are shown in Table 3(c). It can be seen from the table that latent-based super-resolution outperforms the baselines in both accuracy and diversity.", + "bbox": [ + 75, + 573, + 468, + 741 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 76, + 753, + 194, + 768 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We present CLIP-Sculptor, a text-to-3D-shape generation method capable of producing shapes of high fidelity and diversity without the need for (text, shape) pairs. To achieve this, CLIP-Sculptor leverages CLIP and implements super-resolution in a discrete latent space with a hierarchical architecture and a novel annealed variant of classifier-free guidance on a mask-based model. We validate CLIP-Sculptor by comparing it with a number of base", + "bbox": [ + 75, + 779, + 470, + 902 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "lines in terms of FID and accuracy, finding that CLIP-Sculptor is the new state of the art for this problem. In experimenting with different guidance scale scheduling, we find that constant scale scheduling did not always work the best, an important finding for the diffusion modeling community that may improve generation quality. Our paper is an important step towards improved quality and diversity of 3D text-to-shape generation outcomes.", + "bbox": [ + 496, + 465, + 893, + 585 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Limitations and Societal Impact. Despite providing higher quality over baseline methods, CLIP-Sculptor still suffers from limitations in its ability to capture smaller details (e.g., chair with a hole on its back) and generate shapes from text prompts involving counting (e.g., chair with four slats). Furthermore, CLIP-Sculptor also fails on many text queries which are not present in the prior knowledge of CLIP. The method also lacks the ability to produce texture. In future work, we will address these limitations by exploring implicit representations that can capture even finer details and by investigating neural networks that can count.", + "bbox": [ + 496, + 612, + 893, + 779 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Text-to-shape approaches can make the generation of 3D content more accessible and efficient. Within the text-to-image domain there has been concern that automatic approaches may replace artists. Because our approach produces voxels, a representation that lends well to direct manipulation and handoff to 3D designers, we believe that our text-to-3D approach can augment 3D design workflows rather than replace them.", + "bbox": [ + 496, + 780, + 893, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "18346", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Panos Achlioptas, Olga Diamanti, Ioannis Mitliagkas, and Leonidas Guibas. Learning representations and generative models for 3d point clouds. In International conference on machine learning, pages 40-49. PMLR, 2018. 2", + "[2] Panos Achlioptas, Judy Fan, X.D. Robert Hawkins, D. Noah Goodman, and J. Leonidas Guibas. ShapeGlot: Learning language for shape differentiation. CoRR, abs/1905.02925, 2019. 1", + "[3] Angel X. Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, Jianxiong Xiao, Li Yi, and Fisher Yu. Shapenet: An information-rich 3d model repository, 2015. cite arxiv:1512.03012. 5", + "[4] Huiwen Chang, Han Zhang, Lu Jiang, Ce Liu, and William T Freeman. Maskgit: Masked generative image transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11315-11325, 2022. 2, 3, 4, 5", + "[5] Kevin Chen, Christopher B Choy, Manolis Savva, Angel X Chang, Thomas Funkhouser, and Silvio Savarese. Text2shape: Generating shapes from natural language by learning joint embeddings. In Asian Conference on Computer Vision, pages 100-116. Springer, 2018. 1, 2", + "[6] Zhiqin Chen and Hao Zhang. Learning implicit fields for generative shape modeling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5939-5948, 2019. 2", + "[7] Christopher B Choy, Danfei Xu, JunYoung Gwak, Kevin Chen, and Silvio Savarese. 3d-r2n2: A unified approach for single and multi-view 3d object reconstruction. In European conference on computer vision, pages 628-644. Springer, 2016. 5", + "[8] Prafulla Dhariwal, Heewoo Jun, Christine Payne, Jong Wook Kim, Alec Radford, and Ilya Sutskever. Jukebox: A generative model for music. arXiv preprint arXiv:2005.00341, 2020. 2", + "[9] Rao Fu, Xiao Zhan, Yiwen Chen, Daniel Ritchie, and Srinath Sridhar. Shapecrafter: A recursive text-conditioned 3d shape generation model, 2022. 1, 2", + "[10] Oran Gafni, Adam Polyak, Oron Ashual, Shelly Sheynin, Devi Parikh, and Yaniv Taigman. Make-a-scene: Scene-based text-to-image generation with human priors, 2022. 2", + "[11] Rinon Gal, Yuval Alaluf, Yuval Atzmon, Or Patashnik, Amit H. Bermano, Gal Chechik, and Daniel Cohen-Or. An image is worth one word: Personalizing text-to-image generation using textual inversion, 2022. 2", + "[12] Shuyang Gu, Dong Chen, Jianmin Bao, Fang Wen, Bo Zhang, Dongdong Chen, Lu Yuan, and Baining Guo. Vector quantized diffusion model for text-to-image synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10696-10706, 2022. 2, 4", + "[13] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 4" + ], + "bbox": [ + 78, + 114, + 470, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[14] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 5", + "[15] Jonathan Ho, William Chan, Chitwan Sahara, Jay Whang, Ruiqi Gao, Alexey Gritsanko, Diederik P. Kingma, Ben Poole, Mohammad Norouzi, David J. Fleet, and Tim Salimans. Imagen video: High definition video generation with diffusion models, 2022. 2", + "[16] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598, 2022. 2, 4, 5", + "[17] Moritz Ibing, Isaak Lim, and Leif Kobbelt. 3d shape generation with grid-based implicit functions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13559-13568, 2021. 2", + "[18] Ajay Jain, Ben Mildenhall, Jonathan T Barron, Pieter Abbeel, and Ben Poole. Zero-shot text-guided object generation with dream fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 867-876, 2022. 1, 2, 5, 6", + "[19] Nasir Khalid, Tianhao Xie, Eugene Belilovsky, and Tiberiu Popa. Clip-mesh: Generating textured meshes from text using pretrained image-text models. ACM Transactions on Graphics (TOG), Proc. SIGGRAPH Asia, 2022. 1, 2, 5, 6", + "[20] Weixin Liang, Yuhui Zhang, Yongchan Kwon, Serena Yeung, and James Zou. Mind the gap: Understanding the modality gap in multi-modal contrastive representation learning. arXiv preprint arXiv:2203.02053, 2022. 4", + "[21] Vivian Liu, Han Qiao, and Lydia Chilton. Opal: Multimodal image generation for news illustration, 2022. 1", + "[22] Vivian Liu, Jo Vermeulen, George Fitzmaurice, and Justin Matejka. 3dall-e: Integrating text-to-image ai in 3d design workflows, 2022. 1", + "[23] Zhengzhe Liu, Yi Wang, Xiaojuan Qi, and Chi-Wing Fu. Towards implicit text-guided 3d shape generation. arXiv preprint arXiv:2203.14622, 2022. 1, 2", + "[24] Lars Mescheder, Michael Oechsle, Michael Niemeyer, Sebastian Nowozin, and Andreas Geiger. Occupancy networks: Learning 3d reconstruction in function space. In Proceedings IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2019. 5", + "[25] Oscar Michel, Roi Bar-On, Richard Liu, Sagie Benaim, and Rana Hanocka. Text2mesh: Text-driven neural stylization for meshes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13492-13502, 2022. 1, 2", + "[26] Paritosh Mittal, Yen-Chi Cheng, Maneesh Singh, and Shubham Tulsiani. Autosdf: Shape priors for 3d completion, reconstruction and generation. arXiv preprint arXiv:2203.09516, 2022. 1, 2, 5", + "[27] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models, 2021. 4", + "[28] Aaron van den Oord, Oriol Vinyals, and Koray Kavukcuoglu. Neural discrete representation learning. arXiv preprint arXiv:1711.00937, 2017. 2, 3, 4" + ], + "bbox": [ + 501, + 92, + 893, + 901 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "18347", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[29] Ben Poole, Ajay Jain, Jonathan T. Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. arXiv, 2022. 2", + "[30] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision, 2021. 1, 2", + "[31] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 2022. 1", + "[32] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In International Conference on Machine Learning, pages 8821-8831. PMLR, 2021. 1", + "[33] Ali Razavi, Aaron Van den Oord, and Oriol Vinyals. Generating diverse high-fidelity images with vq-vae-2. Advances in neural information processing systems, 32, 2019. 2, 4", + "[34] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models, 2021. 2", + "[35] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10684-10695, June 2022. 1", + "[36] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10684-10695, 2022. 2, 4", + "[37] Aditya Sanghi, Hang Chu, Joseph G Lambourne, Ye Wang, Chin-Yi Cheng, Marco Fumero, and Kamal Rahimi Malekshan. Clip-forge: Towards zero-shot text-to-shape generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18603-18613, 2022. 1, 2, 5, 6", + "[38] Nikolay Savinov, Junyoung Chung, Mikolaj Binkowski, Erich Elsen, and Aaron van den Oord. Step-unrolled denoising autoencoders for text generation. arXiv preprint arXiv:2112.06749, 2021. 4", + "[39] Nikolay Savinov, Junyoung Chung, Mikolaj Binkowski, Erich Elsen, and Aaron van den Oord. Step-unrolled denoising autoencoders for text generation, 2021. 7", + "[40] Mohit Shridhar, Lucas Manuelli, and Dieter Fox. CIport: What and where pathways for robotic manipulation. In Proceedings of the 5th Conference on Robot Learning (CoRL), 2021. 1", + "[41] Can Wang, Menglei Chai, Mingming He, Dongdong Chen, and Jing Liao. Clip-nerf: Text-and-image driven manipulation of neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3835–3844, 2022. 1, 2", + "[42] Xingguang Yan, Liqiang Lin, Niloy J Mitra, Dani Lischinski, Daniel Cohen-Or, and Hui Huang. Shapeformer:" + ], + "bbox": [ + 78, + 90, + 467, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Transformer-based shape completion via sparse representation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6239-6249, 2022. 2", + "[43] Guandao Yang, Xun Huang, Zekun Hao, Ming-Yu Liu, Serge Belongie, and Bharath Hariharan. Pointflow: 3d point cloud generation with continuous normalizing flows. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4541-4550, 2019. 2", + "[44] Yufan Zhou, Ruiyi Zhang, Changyou Chen, Chunyuan Li, Chris Tensmeyer, Tong Yu, Jiuming Gu, Jinhui Xu, and Tong Sun. Lafite: Towards language-free training for text-to-image generation. arXiv preprint arXiv:2111.13792, 2021. 4", + "[45] Adrian Łancucki, Jan Chorowski, Guillaume Sanchez, Ricard Marxer, Nanxin Chen, Hans J. G. A. Dolfing, Sameer Khurana, Tanel Alumäe, and Antoine Laurent. Robust training of vector quantized bottleneck models, 2020. 4" + ], + "bbox": [ + 503, + 92, + 890, + 330 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "18348", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2023/CLIP-Sculptor_ Zero-Shot Generation of High-Fidelity and Diverse Shapes From Natural Language/e64fac6d-0780-4018-a6c0-4f909bfb4337_model.json b/2023/CLIP-Sculptor_ Zero-Shot Generation of High-Fidelity and Diverse Shapes From Natural Language/e64fac6d-0780-4018-a6c0-4f909bfb4337_model.json new file mode 100644 index 0000000000000000000000000000000000000000..6b5fb2a56c2a13be7c6b29a5354bad8a70786a1a --- /dev/null +++ b/2023/CLIP-Sculptor_ Zero-Shot Generation of High-Fidelity and Diverse Shapes From Natural Language/e64fac6d-0780-4018-a6c0-4f909bfb4337_model.json @@ -0,0 +1,1749 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.0, + 0.812, + 0.045 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.08, + 0.13, + 0.89, + 0.176 + ], + "angle": 0, + "content": "CLIP-Sculptor: Zero-Shot Generation of High-Fidelity and Diverse Shapes from Natural Language" + }, + { + "type": "text", + "bbox": [ + 0.136, + 0.202, + 0.842, + 0.258 + ], + "angle": 0, + "content": "Aditya Sanghi† Rao Fu‡ Vivian Liu§ Karl D.D. Willis† Hooman Shayani† Amir H. Khasahmadi† Srinath Sridhar‡ Daniel Ritchie‡ Autodesk Research† Brown University‡ Columbia University§" + }, + { + "type": "image", + "bbox": [ + 0.078, + 0.291, + 0.631, + 0.453 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.637, + 0.293, + 0.892, + 0.37 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.637, + 0.378, + 0.892, + 0.452 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.075, + 0.465, + 0.893, + 0.508 + ], + "angle": 0, + "content": "Figure 1. CLIP-Sculptor is a zero-shot text-to-shape generation method. Left Top/Bottom: It generates diverse shapes that reflect the semantic meaning of the text input without requiring any text data during training. Right: The method also generates high-fidelity shapes." + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.515, + 0.314, + 0.531 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.547, + 0.474, + 0.79 + ], + "angle": 0, + "content": "Recent works have demonstrated that natural language can be used to generate and edit 3D shapes. However, these methods generate shapes with limited fidelity and diversity. We introduce CLIP-Sculptor, a method to address these constraints by producing high-fidelity and diverse 3D shapes without the need for (text, shape) pairs during training. CLIP-Sculptor achieves this in a multi-resolution approach that first generates in a low-dimensional latent space and then upscales to a higher resolution for improved shape fidelity. For improved shape diversity, we use a discrete latent space which is modeled using a transformer conditioned on CLIP's image-text embedding space. We also present a novel variant of classifier-free guidance, which improves the accuracy-diversity trade-off. Finally, we perform extensive experiments demonstrating that CLIP-Sculptor outperforms state-of-the-art baselines." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.815, + 0.21, + 0.831 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.471, + 0.903 + ], + "angle": 0, + "content": "In recent years, there has been rapid progress in text-conditioned image generation [31, 32, 35], which has been driven by advances in multimodal understanding learned from web-scaled paired (text, image) data. These advances" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.517, + 0.893, + 0.639 + ], + "angle": 0, + "content": "have led to applications in domains ranging from content creation [21, 22, 31] to human-robot interaction [40]. Unfortunately, developing the analogue of a text-conditioned 3D shape generator is challenging because it is difficult to obtain (text, 3D shape) pairs at large scale. Prior work has attempted to address this problem by collecting text-shape paired data [2, 5, 9, 23, 26], but these approaches have been limited to a small number of object categories." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.644, + 0.895, + 0.902 + ], + "angle": 0, + "content": "A promising way around this data bottleneck is to use weak supervision from large-scale vision/language models such as CLIP [30]. One approach is to directly optimize a 3D representation such that (text, image render) pairs are aligned when projected into the CLIP embedding space. Prior work has applied this approach to stylize 3D meshes [25, 41] and to create abstract \"dreamlike\" objects using neural radiance fields [18] or meshes [19]. However, neither of the aforementioned methods produce realistic object geometry, and they can require expensive optimization. Another approach, more in line with text-to-image generators [31, 32], is to train a conditional generative model. The CLIP-Forge system [37] builds such a model without paired (text, shape) data by using rendered images of shapes at training time and leveraging the CLIP embedding space to bridge the modalities of image and text at inference time. CLIP-Forge demonstrates compelling zero-shot generation" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "18339" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.089, + 0.472, + 0.131 + ], + "angle": 0, + "content": "Table 1. High-level comparison between zero-shot text-to-shape generation methods. Inference time is calculated using a single NVIDIA Tesla V100 GPU." + }, + { + "type": "table", + "bbox": [ + 0.081, + 0.143, + 0.468, + 0.235 + ], + "angle": 0, + "content": "
MethodInferenceFidelityDiversity
DreamFields [18]>24 hrsLowSingle
Clip-Mesh [19]30 minLowSingle
Clip-Forge [37]6.41 msMediumMedium
CLIP-Sculptor0.91 secHighHigh
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.26, + 0.468, + 0.304 + ], + "angle": 0, + "content": "abilities but produces low-fidelity shapes which do not capture the full diversity of shapes found in the training distribution." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.305, + 0.47, + 0.622 + ], + "angle": 0, + "content": "In this paper, we propose CLIP-Sculptor, a text-conditioned 3D shape generative model that outperforms the state of the art by improving shape diversity and fidelity with only (image, shape) pairs as supervision. CLIP-Sculptor's novelty lies in its multi-resolution, voxel-based conditional generation scheme. Without (text, shape) pairs, CLIP-Sculptor learns to generate 3D shapes of common object categories from the text-image joint embedding of CLIP. To achieve high fidelity outputs, CLIP-Sculptor adopts a multi-resolution approach: it first generates a low-resolution latent grid representation that captures the semantics from text/image, then upscales it to a higher resolution latent grid representation with a super-resolution model, and finally decodes the output geometry. To generate diverse shapes, CLIP-Sculptor adopts discrete latent representations which are obtained using a vector quantization scheme that avoids posterior collapse. To further improve shape fidelity and diversity, CLIP-Sculptor uses a masked transformer architecture. We additionally propose a novel annealed strategy for the classifier-free guidance [16]. To sum up, we make the following contributions:" + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.632, + 0.468, + 0.677 + ], + "angle": 0, + "content": "- CLIP-Sculptor, a multi-resolution text-conditional shape generative model that achieves both high fidelity and diversity without the need for (text, shape) pairs." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.687, + 0.468, + 0.732 + ], + "angle": 0, + "content": "- A novel variant of classifier-free guidance for generative models, with an annealed guidance schedule to achieve better quality for a given diversity level." + }, + { + "type": "list", + "bbox": [ + 0.097, + 0.632, + 0.468, + 0.732 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.755, + 0.22, + 0.77 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.78, + 0.468, + 0.901 + ], + "angle": 0, + "content": "Neural Discrete Representation. Discrete representations [28] were first proposed for image generation in the context of variational autoencoders as a method to improve image quality and avoid posterior collapse. Further work introduced multi-scale and hierarchical variants [8, 33] which improved generative capabilities. Recently, 3D discrete representations have emerged as well, such as discretized voxel and implicit grids [26, 42]. In this work, we take inspiration" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.892, + 0.138 + ], + "angle": 0, + "content": "from hierarchical VQ-VAEs [8, 33] and propose an architecture of hierarchical discrete representations capable of generating high fidelity 3D shapes." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.155, + 0.892, + 0.381 + ], + "angle": 0, + "content": "Latent Generative Models. In recent years, latent generative models have been widely adopted, because these models can effectively generate low-dimensional latent representations which can be used to generate high fidelity images and shapes. In the 3D domain, GAN-based latent models [1, 6, 17] have shown impressive results but tend to suffer from training instability and mode collapse. Flow-based models [37, 43] have been proposed to alleviate these problems, but they yield sample quality that is inferior to GAN-based models. Recent works in the image domain use diffusion [12, 36] or masking models [4] on the latent space which can increase the inference efficiency and still give quality outputs. Building on these works, we propose a hierarchical latent generative model that can further improve the fidelity and diversity of shape generations." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.399, + 0.892, + 0.641 + ], + "angle": 0, + "content": "Classifier-free Guidance. Classifier-free guidance, proposed by [16], jointly trains a conditional and unconditional generative model, whose score estimates are mixed to establish a trade-off between diversity measured by Frechet inception distance (FID) and sample accuracy/fidelity measured by Inception score (IS). Classifier-free guidance guides unconditional samples in the direction of conditional ones and has been implemented in practice to achieve state of the art results in recent work [10-12, 34]. To the best of our knowledge, guidance scale has always been kept constant outside the concurrent exploration in [15]. Their oscillating guidance technique, however does not show improvements in sample fidelity and produces more artifacts in their generations. In this work, we find that rather using annealed scheduling can give better quality versus diversity trade-offs." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.659, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Text-to-Shape Generation. Text-to-shape generation has gained momentum in recent years. Recent works [5, 9, 23, 26] use supervised text-shape pairs to generate shapes effectively using natural language. However, a major drawback is the lack of available text-shape datasets, which limits supervised methods to generate shapes only in few categories. To solve this, several recent works have successfully leveraged the image-text embeddings of CLIP as a prior [30] by converting shapes into images using renders. One line of work uses differentiable renderers [18, 19, 25, 29, 41]. Another line of work learns a mapping from the image space to the shape space [37]. Although these methods are effective, they suffer from either long optimization time [18, 19, 25, 29] or limited quality of generated shapes [18, 19, 37]. Our method is able to generate more diverse shapes of higher fidelity within a comparably short inference time." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.956 + ], + "angle": 0, + "content": "18340" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.078, + 0.089, + 0.512, + 0.322 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.516, + 0.089, + 0.892, + 0.322 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.079, + 0.323, + 0.891, + 0.453 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.465, + 0.893, + 0.549 + ], + "angle": 0, + "content": "Figure 2. The CLIP-Sculptor architecture during training (top) and inference (bottom). CLIP-Sculptor is trained in three stages. In Stage 1, we train two separate VQ-VAE models for \\(32^{3}\\) and \\(64^{3}\\) resolution voxel grids. In Stage 2 we train a coarse transformer conditioned on a CLIP embedding to generate low resolution VQ-VAE latent grids \\(\\mathbf{E}_{32}\\). In Stage 3, we train a fine transformer to perform super resolution on these latent grids. During inference, a text prompt is passed through the CLIP text encoder and used to condition the coarse transformer to generate a coarse latent grid \\(\\mathbf{E}_{32}\\). This coarse grid is then used to condition the fine transformer to generate a fine latent grid \\(\\mathbf{E}_{64}\\). Finally, this fine latent grid is then passed through the Training Stage 1 \\(64^{3}\\) VQ-VAE decoder to generate the output shape." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.586, + 0.169, + 0.601 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.614, + 0.474, + 0.903 + ], + "angle": 0, + "content": "CLIP-Sculptor aims to generate diverse and high-fidelity 3D shapes of common object categories from text prompts. Figure 2 illustrates the components of our CLIP-Sculptor model. We train the network using 3 stages. In the first stage, we use a vector-quantized representation [28], to efficiently represent 3D shapes in discrete voxel grids. This latent representation captures the diverse semantic meanings of the 3D shapes in common object categories and is compatible with powerful generative models. To balance computation cost and shape quality, we utilize latent representations at two resolutions. A low resolution latent representation \\(\\mathbf{E}_{32}\\) encodes geometric information of voxels \\(\\mathbf{V}_{32}\\) at resolutions of \\(32^{3}\\), which is useful for text-conditional generation. A high resolution latent representation \\(\\mathbf{E}_{64}\\) encodes geometric information of voxel grids \\(\\mathbf{V}_{64}\\) at resolutions of \\(64^{3}\\), which is useful for super-resolution and detail generation. In stage 2, we train a generative model conditioned on CLIP features obtained from images \\(\\{\\mathbf{I}_r|r\\in R\\}\\) rendered at random views \\(R\\) to solve the lack of paired (text, shape)" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.588, + 0.892, + 0.708 + ], + "angle": 0, + "content": "data. The conditional generative model consists of a mapping net and a coarse transformer \\( T_{c}(.) \\) that is trained to unmask a masked low-resolution grid \\( \\mathbf{E}_{32} \\). Finally in stage 3, a fine transformer \\( T_{f}(.) \\) is used as a super-resolution model in latent space to generate higher resolution shapes. It is trained in a similar manner to the coarse transformer but conditioned on a low-resolution grid \\( \\mathbf{E}_{32} \\) instead of CLIP features." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.714, + 0.892, + 0.85 + ], + "angle": 0, + "content": "During inference, a text prompt is sequentially converted from CLIP text embedding to a low-resolution latent representation \\(\\mathbf{E}_{32}\\) to a high-resolution latent representation \\(\\mathbf{E}_{64}\\) to finally a high-resolution voxel grid \\(\\mathbf{V}_{64}\\). A confidence-based iterative decoding scheme [4] is adopted for unmasking the fully masked initial grids for coarse and fine transformers at inference. To further control shape fidelity and diversity, we propose an annealing strategy for classifier-free guidance at inference time." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.856, + 0.893, + 0.901 + ], + "angle": 0, + "content": "In the following subsections, we will introduce the major components within CLIP-Sculptor. Section 3.1 describes the vector quantized representations at multiple resolutions;" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "18341" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.168 + ], + "angle": 0, + "content": "Section 3.2 explains the conditional generation model as well as its design choices and training strategies; Section 3.3 introduces the super-resolution model; Section 3.4 explains how the classifier-free guidance is achieved at inference using an annealing strategy." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.175, + 0.373, + 0.192 + ], + "angle": 0, + "content": "3.1. Multi-Resolution Voxel VQ-VAEs" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.199, + 0.469, + 0.349 + ], + "angle": 0, + "content": "To learn a low-dimensional latent space that effectively represents the distribution of the training shape set \\(D\\), we use a vector-quantized variational autoencoder (VQVAE) [28]. The vector quantized latent representation avoids \"posterior collapse\", and it is also amenable to modeling with transformer-based generative models. To balance computational cost and shape quality, we train two separate VQ-VAEs for \\(32^{3}\\) and \\(64^{3}\\) resolution voxel grids \\(\\mathbf{V}_{32}\\) and \\(\\mathbf{V}_{64}\\), which are useful for conditional generation and superresolution respectively." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.35, + 0.47, + 0.502 + ], + "angle": 0, + "content": "For both VQ-VAEs, we use ResNet-based [13] volumetric CNN encoders and decoders. A vector quantization layer maps the down-sampled voxel grids into a discrete latent space by indexing the continuous latent space with an embedding codebook. In this way, the voxel grids \\(\\mathbf{V}_{32}\\) and \\(\\mathbf{V}_{64}\\) are effectively represented by low-dimensional latent representations \\(\\mathbf{E}_{32}\\) and \\(\\mathbf{E}_{64}\\), at the resolution of \\(4^{3}\\) and \\(8^{3}\\). The VQ-VAEs are trained with mean squared error loss (MSE), commitment loss [28], and an exponential moving average strategy [33, 45]." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.51, + 0.419, + 0.525 + ], + "angle": 0, + "content": "3.2. CLIP-Conditioned Coarse Transformer" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.533, + 0.469, + 0.699 + ], + "angle": 0, + "content": "The conditional generation process aims to generate diverse shapes that semantically correspond to the text prompts. Inspired by the recent success of masking approaches [4] and diffusion models [12, 27, 36], we formulate this generation process as a conditional unmasking task. Given an input masked latent representation \\(\\mathbf{E}_{32}^{msk}\\) and a conditional vector \\(\\mathbf{c}\\), the coarse transformer \\(T_{c}(.)\\) should produce an fully unmasked latent representation \\(\\mathbf{E}_{32}\\) that not only aligns with the semantic meaning captured by the conditional vector, but also allows for the sampling of diverse shapes at inference." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.7, + 0.469, + 0.835 + ], + "angle": 0, + "content": "The conditional vector \\(\\mathbf{c}\\) provides the semantic guidance for the shape generation process. Since we assume no text data at the training time, we leverage the text-image joint-embedding of CLIP. The generation is guided by the CLIP image embedding from the rendered images \\(\\{\\mathbf{I}_r\\}\\) during training and guided by the CLIP text embedding from the text prompt during inference. We add Gaussian noise to the CLIP image embeddings [44] to alleviate the modality gap between the CLIP text and image embeddings [20]:" + }, + { + "type": "equation", + "bbox": [ + 0.167, + 0.844, + 0.469, + 0.877 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathbf {c}} = f _ {I} \\left(\\mathbf {I} _ {r}\\right) + \\frac {\\gamma \\cdot \\epsilon \\cdot \\| f _ {I} \\left(\\mathbf {I} _ {r}\\right) \\| _ {2}}{\\| \\epsilon \\| _ {2}}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.886, + 0.469, + 0.902 + ], + "angle": 0, + "content": "where \\(f_{I}(\\cdot)\\) is the CLIP image encoder, \\(\\epsilon \\sim \\mathcal{N}(0,1)\\), and \\(\\gamma\\)" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.257 + ], + "angle": 0, + "content": "controls the level of perturbation. The conditional vector \\(\\hat{\\mathbf{c}}\\) is then normalized and passed through a mapping network that consists of multi-layer perceptrons (MLP). The mapped vector \\(\\tilde{\\mathbf{c}}\\) conditions on the coarse transformer \\(T_{c}(.)\\) by predicting the affine transform parameters of each transformer block's layernorm. In order to control the diversity-quality balance of the generated shapes, we replace the conditional vector with the null embedding by a \\(\\rho\\%\\) drop out rate for classifier-free guidance [16]. These are important design choices that affects the diversity and quality of the generated shapes at inference time." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.258, + 0.892, + 0.486 + ], + "angle": 0, + "content": "To train the coarse transformer \\( T_{c}(.) \\), we randomly replace the input indices with a special \"mask\" tokens by uniformly selecting a masking ratio between 0-100% [4]. Given the masked latent grid \\( \\mathbf{E}_{32}^{msk} \\) and the conditional vector \\( \\mathbf{c} \\), the training objective is to maximize the following log-likelihood: \\( \\sum_{n}^{N}\\log p(\\mathbf{E}_{32,n}|\\mathbf{E}_{32,n}^{msk},\\mathbf{c}) / N \\), where \\( N \\) is the training data size and \\( \\mathbf{E}_{32}^{msk} \\) is the randomly masked latent grids. As the network sees only the ground truth tokens but not its own predicted samples during training, there exists an autoregressive drift resulted by the accumulative error at inference-time sampling. To alleviate this issue, we take inspiration from the NLP literature [38] and use a similar step-unrolled training (SUT) that predicts the entire un-masked grids at each time step. The final objective is to maximize the following formulation:" + }, + { + "type": "equation", + "bbox": [ + 0.502, + 0.496, + 0.891, + 0.549 + ], + "angle": 0, + "content": "\\[\n\\sum_ {n} ^ {N} \\left(\\log p \\left(\\mathbf {E} _ {3 2, n} \\mid \\mathbf {E} _ {3 2, n} ^ {m s k}, \\mathbf {c}\\right) + \\log p \\left(\\mathbf {E} _ {3 2, n} \\mid \\tilde {\\mathbf {E}} _ {3 2, n}, \\mathbf {c}\\right)\\right) / N, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.549, + 0.89, + 0.565 + ], + "angle": 0, + "content": "where \\(\\tilde{\\mathbf{E}}_{32}\\) is the coarse transformer \\(T_{c}(.)\\) 's own prediction." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.574, + 0.849, + 0.59 + ], + "angle": 0, + "content": "3.3. Super-Resolution with Fine Transformer" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.598, + 0.892, + 0.764 + ], + "angle": 0, + "content": "Super-resolution in the discrete latent space improves quality and upsamples the resolution of generated shapes. This process is done with a fine transformer \\( T_{f}(.) \\), which takes a low-resolution latent representation \\( \\mathbf{E}_{32} \\) and a masked high-resolution latent grid as input. The fine transformer is conditioned on the coarse grid \\( \\mathbf{E}_{32} \\) via cross attention. It unmasks the high-resolution representation \\( \\mathbf{E}_{64} \\). To ensure that training data resembles the inference data, we directly use the predictions \\( \\tilde{\\mathbf{E}}_{32} \\) of the coarse transformer \\( T_{c}(.) \\) to train the fine transformer \\( T_{f}(.) \\), instead of using the ground truth low-resolution latent representation \\( \\mathbf{E}_{32} \\)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.772, + 0.805, + 0.787 + ], + "angle": 0, + "content": "3.4. Annealed Classifier-Free Guidance" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.796, + 0.892, + 0.901 + ], + "angle": 0, + "content": "At inference time, we first convert an input text prompt into a CLIP text embedding using the CLIP text encoder \\( f_{T}(\\cdot) \\). The text embedding is then fed into the mapping network and used as the conditional vector \\( \\tilde{\\mathbf{c}} \\) of the coarse transformer \\( T_{c}(\\cdot) \\). We use the iterative decoding scheme [4] to slowly unmask the grid over a sequence of \\( T \\) steps. The initial input to the coarse transformer \\( T_{c}(\\cdot) \\) is a completely" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18342" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.242 + ], + "angle": 0, + "content": "masked latent grid. At each time step \\( t \\), we condition the coarse transformer \\( T_{c}(.) \\) with \\( \\tilde{\\mathbf{c}} \\) and take the output latent grid \\( \\mathbf{E}_{32,t-1}^{msk} \\) from the prior time step as the input. The coarse transformer \\( T_{c}(.) \\) will predict an unmasked latent grid \\( \\tilde{\\mathbf{E}}_{32,t} \\). We mask all the tokens except for the most confident token predictions and the tokens unmasked from previous steps. This newly masked latent grid \\( \\mathbf{E}_{32,t}^{msk} \\) is the input to the next time step \\( t+1 \\). We repeat this process until the process unmasks all the tokens, which is ensured by a cosine masking schedule [4]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.244, + 0.473, + 0.471 + ], + "angle": 0, + "content": "During this iterative decoding scheme, we apply a new variant of classifier-free guidance [16]. Classifier-free guidance extrapolates an unconditional sample in the direction of a conditional sample. The extrapolation is controlled by an adaptive guidance scale parameter. Varying this parameter results in a trade-off between accuracy of text-to-shape generation and sample diversity. Instead of using a constant guidance scale for all time steps [16], we propose to vary the guidance scale over time according to an annealing schedule. The intuition is that during the initial steps of sampling when the input to the coarse transformer is mostly masked indices, a larger guidance scale is important to keep the network \"on task\"; later on in the sampling process, a lower guidance scale can help produce more sample diversity. Overall, the guidance is formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.14, + 0.486, + 0.47, + 0.511 + ], + "angle": 0, + "content": "\\[\n\\hat {P} _ {t} (c) = P _ {t} (0) + a (t) \\left(P _ {t} (c) - P _ {t} (0)\\right), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.521, + 0.47, + 0.658 + ], + "angle": 0, + "content": "where \\(a(t)\\) is the guidance scale annealing schedule, \\(P_{t}(c)\\) denotes the conditional distribution \\(p(\\mathbf{E}_{32,t}|\\mathbf{E}_{32,t - 1}^{msk},\\mathbf{c})\\) and \\(P_{t}(0)\\) denotes \\(p(\\mathbf{E}_{32,t}|\\mathbf{E}_{32,t - 1}^{msk},\\mathbf{0})\\). Note that \\(a(t)\\) is continuous and usually monotonically decreasing. By setting \\(a(t) = k\\) for some constant \\(k\\), this equation is equal to the classifier-free guidance scheme with a constant guidance scale. In this paper, we experimentally evaluate other annealing schedules [4], including linear, cosine, and square root functions." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.66, + 0.47, + 0.737 + ], + "angle": 0, + "content": "Finally, the unmasked coarse latent grid \\(\\tilde{\\mathbf{E}}_{32}\\) is fed to the fine transformer \\(T_{f}\\), which uses the iterative decoding scheme to unmask an initially masked fine latent grid \\(\\mathbf{E}_{64}^{msk}\\). The unmasked fine grid is then passed to the \\(64^3\\) VQ-VAE decoder to obtain the final high-resolution voxel shape." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.754, + 0.21, + 0.77 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.781, + 0.47, + 0.903 + ], + "angle": 0, + "content": "In this section, we report experimental results to evaluate the generation fidelity, diversity, and class accuracy of CLIP-Sculptor. We provide full details of the experiment setup in the supplementary. We run the experiments 3 times for each of the aforementioned measures and report the mean in each case, except for the final comparisons with baselines (Table 2), where we use the best seed. Additional results can also be found in the supplementary." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.212 + ], + "angle": 0, + "content": "Dataset. We conduct our experiments on two subsets of the ShapeNet(v2) dataset [3]. The first subset, ShapeNet13, contains 13 categories from ShapeNet as used in [7,24]. We use the same train/test split as specified in [24]. Our second subset is ShapeNet55 which contains all 55 ShapeNet categories. For ShapeNet55, we render images as described in [7]; the training dataset contains 51,784 datapoints, whereas the test set contains 6,101 datapoints." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.213, + 0.892, + 0.333 + ], + "angle": 0, + "content": "Evaluation Metrics. To assess the generative capabilities of CLIP-Sculptor, we use Fréchet Inception Distance (FID) ([14]) to evaluate shape diversity and quality, and use Classifier Accuracy (Acc) to evaluate the text-shape correspondence. These metrics follow [37] where they first generated single mean shapes for 234 predetermined text queries and then passed all shapes through a classifier to measure Acc. The latent space of this classifier is also used for FID." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.335, + 0.893, + 0.653 + ], + "angle": 0, + "content": "Baseline. We compare the performance of our method against CLIP-Forge [37], CLIP-Mesh [19] and DreamFields [18], which are currently state of the art for zero-shot text-to-shape generation. We also set up a comparison method called Zero-Shot AutoSDF(ZS-ASDF), which trains the supervised method AutoSDF [26] with the CLIP features as CLIP-Sculptor, to highlight the superiority of our conditional generation model. CLIP-Forge reports results on single shape generation for 234 text queries. These single shapes were generated using the mean of the prior (the Gaussian distribution), so we refer to these results as CF-MS. Note that this does not capture the diversity of shapes that can be generated given a text query. To capture results on more shape generations, we sample 32 shapes instead of one using a Gaussian (CF-G), truncated Gaussian (CF-TG) or clipped Gaussian (CF-CG) distribution. We also generate 32 shapes for ZS-ASDF and our method for fair comparison. We only compare qualitatively with DreamFields and CLIP-Mesh as the optimization time per query is significant (Table 1), and it does not use prior knowledge from the shape dataset." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.665, + 0.855, + 0.682 + ], + "angle": 0, + "content": "4.1. Evaluating Shape Diversity and Accuracy" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.689, + 0.892, + 0.901 + ], + "angle": 0, + "content": "In this section, we first quantitatively evaluate the diversity and accuracy of a given shape matching a given text query on the ShapeNet13 dataset. We compare with CLIP-Forge and ZS-ASDF using the Acc and FID metrics. The results are shown in Table 2. The first four columns represent CLIP-Forge(CF) and its different sampling techniques. The fifth column represents the result from AutoSDF trained with CLIP features(ZS-ASDF). The other columns represent our method with different annealing scale strategies. Three major things can be observed. First, all variants of our method outperform CLIP-Forge and ZS-ASDF significantly. This indicates that the method produces more diverse and higher fidelity shapes of increasing accuracy. Second, it can be seen that annealing strategies" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18343" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.083, + 0.086, + 0.917, + 0.715 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.725, + 0.895, + 0.784 + ], + "angle": 0, + "content": "Figure 3. Qualitative comparison of CLIP-Sculptor (rendered in purple), CLIP-Forge [37](rendered in green), DreamField [18], and CLIP-Mesh [19] on text-conditioned generation. For each text prompt, the first generated shape by CLIP-Sculptor and CLIP-Forge [37] is zoomed in for detail comparison. Comparing with other methods, CLIP-Sculptor generates a broader category of shapes that correspond with the text prompt. The generated shapes are more diverse and has higher fidelity." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.808, + 0.473, + 0.899 + ], + "angle": 0, + "content": "for guidance scale give a better accuracy-diversity trade-off than constant scale guidance. Finally, we note that simply augmenting AutoSDF (another VQ-VAE) with CLIP features does not lead to improved text to shape generation, indicating the importance of our architecture design decisions." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.808, + 0.895, + 0.901 + ], + "angle": 0, + "content": "We also qualitatively compare our method to CLIP-Forge, CLIP-Mesh, and DreamFields. The results are shown in Figure 3. It can be observed that our method generates higher quality shapes (view \"a machine gun\"), with more detail (view \"an office chair\") and higher diversity (view \"a jet\"). CLIP-Forge usually produces the same shape" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "18344" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.089, + 0.895, + 0.133 + ], + "angle": 0, + "content": "Table 2. Comparisons of CLIP-Forge(CF) baseline (across different sampling strategies), AutoSDF augmented with CLIP (ZS-ASDF) with CLIP-Sculptor (CS) on Accuracy(ACC) and FID. CLIP-Sculptor outperforms other methods in both metrics, which proves it generates shapes with higher fidelity and better diversity." + }, + { + "type": "table", + "bbox": [ + 0.08, + 0.155, + 0.894, + 0.22 + ], + "angle": 0, + "content": "
MethodCF-MSCF-GCF-TGCF-CGZS-ASDFCS-const.CS-sqrtCS-linearCS-cosine
FID ↓2425.252233.482141.612100.677332.931821.781480.111629.511725.63
ACC ↑83.3362.8168.7171.1139.1486.5987.0887.5087.27
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.253, + 0.472, + 0.31 + ], + "angle": 0, + "content": "Table 3. Ablations on the major components of CLIP-Sculptor. CLIP-Sculptor achieves the best performance comparing with other design choice, which proves the effectiveness of the major components." + }, + { + "type": "table_caption", + "bbox": [ + 0.173, + 0.321, + 0.393, + 0.334 + ], + "angle": 0, + "content": "(a) Effect of varying the Noise parameter." + }, + { + "type": "table", + "bbox": [ + 0.088, + 0.336, + 0.479, + 0.552 + ], + "angle": 0, + "content": "
γ×0.50.81.01.21.5
FID↓1720.021764.981484.611703.381447.911478.17
ACC↑64.8775.7377.4179.0979.6378.47
(b) Effect of varying number of layers (L) in the mapping network.
L012345
FID↓2874.871716.731518.971447.911532.501424.46
ACC↑62.7278.7079.1779.6377.1676.09
(c) Comparison with baselines on super resolution.
Method3D-UNet643-DSCLIP-Sculptor
FID↓2056.922196.961910.28
ACC↑86.6577.9286.85
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.578, + 0.47, + 0.7 + ], + "angle": 0, + "content": "with small variations (view \"a rectangular table\"). We also note that DreamFields and CLIP-Mesh produce very abstract results which may not be useful in many applications. Finally, we also show results on ShapeNet55 in the last row of Figure 3. We could not get CLIP-Forge to produce sensible shapes for most text queries on ShapeNet55 (which we attribute to the data imbalance issue of ShapeNet55), whereas our method produced high fidelity shapes." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.71, + 0.47, + 0.728 + ], + "angle": 0, + "content": "4.2. Major Components for Conditional Generation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.735, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Effect of Noise Parameter. During Stage 2 Training, to better align the image and text embeddings, we added Gaussian noise at varying levels of \\(\\gamma\\). We investigate the effect of \\((\\gamma)\\) added to the image condition vector and show the results in Table 3 (a). We keep the number of mapping layers fixed at three in this experiment. We observe that adding Gaussian noise drastically improves both the diversity and accuracy of generation, with the optimal noise parameter being around 1.2. These results indicate that adding noise during training helps align the text features seen at inference and the image features seen at training." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.255, + 0.892, + 0.391 + ], + "angle": 0, + "content": "Size of Mapping Network. We next probe the importance of the mapping network. We show the results in Table 3 (b), where \\( L \\) represents the number of layers of the mapping function. In the case of 0 layers, we directly project conditional embeddings to the layernorm parameters using a linear layer. We make two observations from the results: 1) A common mapping network improves both the accuracy and FID. 2) Increasing the number of mapping network layers beyond a certain number decreases accuracy." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.392, + 0.893, + 0.544 + ], + "angle": 0, + "content": "Classifier-Free Guidance. We next explore the relationship between dropping out image conditioning at \\(\\rho \\%\\) during training and using constant classifier-free guidance during test time. In Table 4, we vary the scale parameter, \\(a(t) = k\\), across the columns. It can be seen that as we increase the scale, accuracy typically increases while FID decreases. This indicates that the method is giving more accurate results on a given text query at the cost of shape diversity. A low dropout rate (5-15%) gives a good trade-off between accuracy and diversity." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.545, + 0.894, + 0.651 + ], + "angle": 0, + "content": "Step-Unrolled Training (SUT). Finally, to further improve the accuracy we investigate the use of step-unrolled training [39]. The results are shown in the last row of Table 4. From the table it can be observed that step-unrolled training enables higher accuracy across all scales. This indicates that the model learns to unmask samples it would encounter during inference time." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.665, + 0.842, + 0.682 + ], + "angle": 0, + "content": "4.3. Annealing Strategy for Scale Parameter" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.69, + 0.895, + 0.901 + ], + "angle": 0, + "content": "An important idea we propose in this paper is a scale annealing technique for classifier-free guidance. We employ classifier-free guidance to identify a better accuracy-diversity trade-off with different annealing schedules: constant, linear, cosine and square root. To determine which annealing technique works the best, we fix the FID values and plotted the accuracy at each fixed FID value (Figure 4). As different scale parameters give different FID values, we conducted an extensive grid search over the starting scale parameter. Note that finding the exact FID is not always feasible, so we pick the closest FID. Figure 4 shows results of FID versus accuracy on three different runs of the Coarse Transformer. We find that across all three runs, the accuracy is typically lower for a given FID for constant schedules as" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18345" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.089, + 0.892, + 0.119 + ], + "angle": 0, + "content": "Table 4. Classifier-Free Guidance and Step-Unrolled Training (SUT) experiment results. \\( \\rho \\% \\) is the drop out rate for classifier-free guidance. SUT indicates the use of Step-Unrolled Training. The remaining columns indicate the variation in scale parameter." + }, + { + "type": "table", + "bbox": [ + 0.078, + 0.129, + 0.9, + 0.241 + ], + "angle": 0, + "content": "
ρ%SUTa(t) = 3a(t) = 2.5a(t) = 2a(t) = 1.5a(t) = 1
FID↓Acc↑FID↓Acc↑FID↓Acc↑FID↓Acc↑FID↓Acc↑
5×2059.085.731970.185.471790.985.431536.583.981227.579.17
10×1893.284.461821.284.501684.383.411522.482.341348.477.76
15×2086.985.031964.784.991851.984.361660.983.141485.578.19
20×2062.583.391972.382.951892.982.741733.381.131566.675.94
5✓2039.887.692011.187.391811.887.401678.686.241517.982.18
" + }, + { + "type": "image", + "bbox": [ + 0.088, + 0.255, + 0.365, + 0.397 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.371, + 0.255, + 0.634, + 0.397 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.639, + 0.256, + 0.892, + 0.397 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.411, + 0.893, + 0.44 + ], + "angle": 0, + "content": "Figure 4. FID and Acc results with different classifier-free guidance scale annealing strategies (constant, cosine, square root, linear) across three different runs (different seeds) of the CLIP-Sculptor Stage 2 Coarse Transformer." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.466, + 0.47, + 0.543 + ], + "angle": 0, + "content": "opposed to other schedules. This is especially the case at the lower range of FID values. The results indicate that having a large scale at the beginning of sampling is more important than later stages, especially in use cases where diversity is paramount." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.551, + 0.248, + 0.567 + ], + "angle": 0, + "content": "4.4. Super-Resolution" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.574, + 0.47, + 0.742 + ], + "angle": 0, + "content": "Finally, we investigate the importance of hierarchical super-resolution. We compare our method with two baselines. In the first baseline, we directly use the \\(32^{3}\\) resolution results from the coarse transformer and use a 3D-UNET based super-resolution network to translate from \\(32^{3}\\) to \\(64^{3}\\). For the second baseline, we train a transformer directly on \\(64^{3}\\) VQ-VAE that is conditioned on text features as opposed to a coarse resolution grid. We refer to this as \\(64^{3}\\)-DS (64\\(^{3}\\) direct synthesis). The results are shown in Table 3(c). It can be seen from the table that latent-based super-resolution outperforms the baselines in both accuracy and diversity." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.755, + 0.196, + 0.77 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.78, + 0.471, + 0.903 + ], + "angle": 0, + "content": "We present CLIP-Sculptor, a text-to-3D-shape generation method capable of producing shapes of high fidelity and diversity without the need for (text, shape) pairs. To achieve this, CLIP-Sculptor leverages CLIP and implements super-resolution in a discrete latent space with a hierarchical architecture and a novel annealed variant of classifier-free guidance on a mask-based model. We validate CLIP-Sculptor by comparing it with a number of base" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.466, + 0.895, + 0.587 + ], + "angle": 0, + "content": "lines in terms of FID and accuracy, finding that CLIP-Sculptor is the new state of the art for this problem. In experimenting with different guidance scale scheduling, we find that constant scale scheduling did not always work the best, an important finding for the diffusion modeling community that may improve generation quality. Our paper is an important step towards improved quality and diversity of 3D text-to-shape generation outcomes." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.613, + 0.895, + 0.78 + ], + "angle": 0, + "content": "Limitations and Societal Impact. Despite providing higher quality over baseline methods, CLIP-Sculptor still suffers from limitations in its ability to capture smaller details (e.g., chair with a hole on its back) and generate shapes from text prompts involving counting (e.g., chair with four slats). Furthermore, CLIP-Sculptor also fails on many text queries which are not present in the prior knowledge of CLIP. The method also lacks the ability to produce texture. In future work, we will address these limitations by exploring implicit representations that can capture even finer details and by investigating neural networks that can count." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.781, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Text-to-shape approaches can make the generation of 3D content more accessible and efficient. Within the text-to-image domain there has been concern that automatic approaches may replace artists. Because our approach produces voxels, a representation that lends well to direct manipulation and handoff to 3D designers, we believe that our text-to-3D approach can augment 3D design workflows rather than replace them." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "18346" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.17 + ], + "angle": 0, + "content": "[1] Panos Achlioptas, Olga Diamanti, Ioannis Mitliagkas, and Leonidas Guibas. Learning representations and generative models for 3d point clouds. In International conference on machine learning, pages 40-49. PMLR, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.171, + 0.472, + 0.226 + ], + "angle": 0, + "content": "[2] Panos Achlioptas, Judy Fan, X.D. Robert Hawkins, D. Noah Goodman, and J. Leonidas Guibas. ShapeGlot: Learning language for shape differentiation. CoRR, abs/1905.02925, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.228, + 0.471, + 0.297 + ], + "angle": 0, + "content": "[3] Angel X. Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, Jianxiong Xiao, Li Yi, and Fisher Yu. Shapenet: An information-rich 3d model repository, 2015. cite arxiv:1512.03012. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.298, + 0.471, + 0.366 + ], + "angle": 0, + "content": "[4] Huiwen Chang, Han Zhang, Lu Jiang, Ce Liu, and William T Freeman. Maskgit: Masked generative image transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11315-11325, 2022. 2, 3, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.368, + 0.471, + 0.438 + ], + "angle": 0, + "content": "[5] Kevin Chen, Christopher B Choy, Manolis Savva, Angel X Chang, Thomas Funkhouser, and Silvio Savarese. Text2shape: Generating shapes from natural language by learning joint embeddings. In Asian Conference on Computer Vision, pages 100-116. Springer, 2018. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.438, + 0.471, + 0.493 + ], + "angle": 0, + "content": "[6] Zhiqin Chen and Hao Zhang. Learning implicit fields for generative shape modeling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5939-5948, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.494, + 0.471, + 0.563 + ], + "angle": 0, + "content": "[7] Christopher B Choy, Danfei Xu, JunYoung Gwak, Kevin Chen, and Silvio Savarese. 3d-r2n2: A unified approach for single and multi-view 3d object reconstruction. In European conference on computer vision, pages 628-644. Springer, 2016. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.565, + 0.471, + 0.619 + ], + "angle": 0, + "content": "[8] Prafulla Dhariwal, Heewoo Jun, Christine Payne, Jong Wook Kim, Alec Radford, and Ilya Sutskever. Jukebox: A generative model for music. arXiv preprint arXiv:2005.00341, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.621, + 0.471, + 0.662 + ], + "angle": 0, + "content": "[9] Rao Fu, Xiao Zhan, Yiwen Chen, Daniel Ritchie, and Srinath Sridhar. Shapecrafter: A recursive text-conditioned 3d shape generation model, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.663, + 0.47, + 0.704 + ], + "angle": 0, + "content": "[10] Oran Gafni, Adam Polyak, Oron Ashual, Shelly Sheynin, Devi Parikh, and Yaniv Taigman. Make-a-scene: Scene-based text-to-image generation with human priors, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.705, + 0.47, + 0.761 + ], + "angle": 0, + "content": "[11] Rinon Gal, Yuval Alaluf, Yuval Atzmon, Or Patashnik, Amit H. Bermano, Gal Chechik, and Daniel Cohen-Or. An image is worth one word: Personalizing text-to-image generation using textual inversion, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.761, + 0.47, + 0.843 + ], + "angle": 0, + "content": "[12] Shuyang Gu, Dong Chen, Jianmin Bao, Fang Wen, Bo Zhang, Dongdong Chen, Lu Yuan, and Baining Guo. Vector quantized diffusion model for text-to-image synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10696-10706, 2022. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.845, + 0.47, + 0.902 + ], + "angle": 0, + "content": "[13] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 4" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.472, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.161 + ], + "angle": 0, + "content": "[14] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.162, + 0.894, + 0.232 + ], + "angle": 0, + "content": "[15] Jonathan Ho, William Chan, Chitwan Sahara, Jay Whang, Ruiqi Gao, Alexey Gritsanko, Diederik P. Kingma, Ben Poole, Mohammad Norouzi, David J. Fleet, and Tim Salimans. Imagen video: High definition video generation with diffusion models, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.232, + 0.894, + 0.26 + ], + "angle": 0, + "content": "[16] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598, 2022. 2, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.26, + 0.895, + 0.316 + ], + "angle": 0, + "content": "[17] Moritz Ibing, Isaak Lim, and Leif Kobbelt. 3d shape generation with grid-based implicit functions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13559-13568, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.316, + 0.894, + 0.384 + ], + "angle": 0, + "content": "[18] Ajay Jain, Ben Mildenhall, Jonathan T Barron, Pieter Abbeel, and Ben Poole. Zero-shot text-guided object generation with dream fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 867-876, 2022. 1, 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.385, + 0.894, + 0.442 + ], + "angle": 0, + "content": "[19] Nasir Khalid, Tianhao Xie, Eugene Belilovsky, and Tiberiu Popa. Clip-mesh: Generating textured meshes from text using pretrained image-text models. ACM Transactions on Graphics (TOG), Proc. SIGGRAPH Asia, 2022. 1, 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.442, + 0.894, + 0.496 + ], + "angle": 0, + "content": "[20] Weixin Liang, Yuhui Zhang, Yongchan Kwon, Serena Yeung, and James Zou. Mind the gap: Understanding the modality gap in multi-modal contrastive representation learning. arXiv preprint arXiv:2203.02053, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.496, + 0.894, + 0.525 + ], + "angle": 0, + "content": "[21] Vivian Liu, Han Qiao, and Lydia Chilton. Opal: Multimodal image generation for news illustration, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.525, + 0.894, + 0.566 + ], + "angle": 0, + "content": "[22] Vivian Liu, Jo Vermeulen, George Fitzmaurice, and Justin Matejka. 3dall-e: Integrating text-to-image ai in 3d design workflows, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.566, + 0.894, + 0.609 + ], + "angle": 0, + "content": "[23] Zhengzhe Liu, Yi Wang, Xiaojuan Qi, and Chi-Wing Fu. Towards implicit text-guided 3d shape generation. arXiv preprint arXiv:2203.14622, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.609, + 0.894, + 0.678 + ], + "angle": 0, + "content": "[24] Lars Mescheder, Michael Oechsle, Michael Niemeyer, Sebastian Nowozin, and Andreas Geiger. Occupancy networks: Learning 3d reconstruction in function space. In Proceedings IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.679, + 0.894, + 0.748 + ], + "angle": 0, + "content": "[25] Oscar Michel, Roi Bar-On, Richard Liu, Sagie Benaim, and Rana Hanocka. Text2mesh: Text-driven neural stylization for meshes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13492-13502, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.748, + 0.894, + 0.803 + ], + "angle": 0, + "content": "[26] Paritosh Mittal, Yen-Chi Cheng, Maneesh Singh, and Shubham Tulsiani. Autosdf: Shape priors for 3d completion, reconstruction and generation. arXiv preprint arXiv:2203.09516, 2022. 1, 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.803, + 0.894, + 0.859 + ], + "angle": 0, + "content": "[27] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.859, + 0.894, + 0.902 + ], + "angle": 0, + "content": "[28] Aaron van den Oord, Oriol Vinyals, and Koray Kavukcuoglu. Neural discrete representation learning. arXiv preprint arXiv:1711.00937, 2017. 2, 3, 4" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.895, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "18347" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.468, + 0.133 + ], + "angle": 0, + "content": "[29] Ben Poole, Ajay Jain, Jonathan T. Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. arXiv, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.136, + 0.468, + 0.205 + ], + "angle": 0, + "content": "[30] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.207, + 0.468, + 0.26 + ], + "angle": 0, + "content": "[31] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.264, + 0.468, + 0.33 + ], + "angle": 0, + "content": "[32] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In International Conference on Machine Learning, pages 8821-8831. PMLR, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.334, + 0.468, + 0.375 + ], + "angle": 0, + "content": "[33] Ali Razavi, Aaron Van den Oord, and Oriol Vinyals. Generating diverse high-fidelity images with vq-vae-2. Advances in neural information processing systems, 32, 2019. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.378, + 0.468, + 0.417 + ], + "angle": 0, + "content": "[34] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.421, + 0.468, + 0.489 + ], + "angle": 0, + "content": "[35] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10684-10695, June 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.491, + 0.468, + 0.559 + ], + "angle": 0, + "content": "[36] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10684-10695, 2022. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.562, + 0.468, + 0.643 + ], + "angle": 0, + "content": "[37] Aditya Sanghi, Hang Chu, Joseph G Lambourne, Ye Wang, Chin-Yi Cheng, Marco Fumero, and Kamal Rahimi Malekshan. Clip-forge: Towards zero-shot text-to-shape generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18603-18613, 2022. 1, 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.646, + 0.468, + 0.7 + ], + "angle": 0, + "content": "[38] Nikolay Savinov, Junyoung Chung, Mikolaj Binkowski, Erich Elsen, and Aaron van den Oord. Step-unrolled denoising autoencoders for text generation. arXiv preprint arXiv:2112.06749, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.703, + 0.468, + 0.743 + ], + "angle": 0, + "content": "[39] Nikolay Savinov, Junyoung Chung, Mikolaj Binkowski, Erich Elsen, and Aaron van den Oord. Step-unrolled denoising autoencoders for text generation, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.746, + 0.468, + 0.8 + ], + "angle": 0, + "content": "[40] Mohit Shridhar, Lucas Manuelli, and Dieter Fox. CIport: What and where pathways for robotic manipulation. In Proceedings of the 5th Conference on Robot Learning (CoRL), 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.803, + 0.468, + 0.871 + ], + "angle": 0, + "content": "[41] Can Wang, Menglei Chai, Mingming He, Dongdong Chen, and Jing Liao. Clip-nerf: Text-and-image driven manipulation of neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3835–3844, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.874, + 0.468, + 0.901 + ], + "angle": 0, + "content": "[42] Xingguang Yan, Liqiang Lin, Niloy J Mitra, Dani Lischinski, Daniel Cohen-Or, and Hui Huang. Shapeformer:" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.468, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.892, + 0.147 + ], + "angle": 0, + "content": "Transformer-based shape completion via sparse representation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6239-6249, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.15, + 0.892, + 0.219 + ], + "angle": 0, + "content": "[43] Guandao Yang, Xun Huang, Zekun Hao, Ming-Yu Liu, Serge Belongie, and Bharath Hariharan. Pointflow: 3d point cloud generation with continuous normalizing flows. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4541-4550, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.221, + 0.892, + 0.275 + ], + "angle": 0, + "content": "[44] Yufan Zhou, Ruiyi Zhang, Changyou Chen, Chunyuan Li, Chris Tensmeyer, Tong Yu, Jiuming Gu, Jinhui Xu, and Tong Sun. Lafite: Towards language-free training for text-to-image generation. arXiv preprint arXiv:2111.13792, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.278, + 0.892, + 0.331 + ], + "angle": 0, + "content": "[45] Adrian Łancucki, Jan Chorowski, Guillaume Sanchez, Ricard Marxer, Nanxin Chen, Hans J. G. A. Dolfing, Sameer Khurana, Tanel Alumäe, and Antoine Laurent. Robust training of vector quantized bottleneck models, 2020. 4" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.892, + 0.331 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.956 + ], + "angle": 0, + "content": "18348" + } + ] +] \ No newline at end of file diff --git a/2023/CLIP-Sculptor_ Zero-Shot Generation of High-Fidelity and Diverse Shapes From Natural Language/e64fac6d-0780-4018-a6c0-4f909bfb4337_origin.pdf b/2023/CLIP-Sculptor_ Zero-Shot Generation of High-Fidelity and Diverse Shapes From Natural Language/e64fac6d-0780-4018-a6c0-4f909bfb4337_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..55eda46ee175a6ccb1d135477bfa2d915549c8f9 --- /dev/null +++ b/2023/CLIP-Sculptor_ Zero-Shot Generation of High-Fidelity and Diverse Shapes From Natural Language/e64fac6d-0780-4018-a6c0-4f909bfb4337_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2341910505aaf8ac902473f3e07a3758482ebce26283d2456b2ce39b36a81501 +size 9290963 diff --git a/2023/CLIP-Sculptor_ Zero-Shot Generation of High-Fidelity and Diverse Shapes From Natural Language/full.md b/2023/CLIP-Sculptor_ Zero-Shot Generation of High-Fidelity and Diverse Shapes From Natural Language/full.md new file mode 100644 index 0000000000000000000000000000000000000000..75adc9bfd20edfc9600a87c42bf30332eb908b46 --- /dev/null +++ b/2023/CLIP-Sculptor_ Zero-Shot Generation of High-Fidelity and Diverse Shapes From Natural Language/full.md @@ -0,0 +1,237 @@ +# CLIP-Sculptor: Zero-Shot Generation of High-Fidelity and Diverse Shapes from Natural Language + +Aditya Sanghi† Rao Fu‡ Vivian Liu§ Karl D.D. Willis† Hooman Shayani† Amir H. Khasahmadi† Srinath Sridhar‡ Daniel Ritchie‡ Autodesk Research† Brown University‡ Columbia University§ + +![](images/af57706f733e87fff2175388558c77f34e475f0ab3b75f9eb54c922df9ff6187.jpg) +Figure 1. CLIP-Sculptor is a zero-shot text-to-shape generation method. Left Top/Bottom: It generates diverse shapes that reflect the semantic meaning of the text input without requiring any text data during training. Right: The method also generates high-fidelity shapes. + +![](images/053674f60e974e7a9db3708cd195027dec70fd24adfb6edc9f662af9bafbbe9c.jpg) + +![](images/8867c4b6a962f70f75737786dbb4d89448714c10757f7851ef4c0ac23c93e158.jpg) + +# Abstract + +Recent works have demonstrated that natural language can be used to generate and edit 3D shapes. However, these methods generate shapes with limited fidelity and diversity. We introduce CLIP-Sculptor, a method to address these constraints by producing high-fidelity and diverse 3D shapes without the need for (text, shape) pairs during training. CLIP-Sculptor achieves this in a multi-resolution approach that first generates in a low-dimensional latent space and then upscales to a higher resolution for improved shape fidelity. For improved shape diversity, we use a discrete latent space which is modeled using a transformer conditioned on CLIP's image-text embedding space. We also present a novel variant of classifier-free guidance, which improves the accuracy-diversity trade-off. Finally, we perform extensive experiments demonstrating that CLIP-Sculptor outperforms state-of-the-art baselines. + +# 1. Introduction + +In recent years, there has been rapid progress in text-conditioned image generation [31, 32, 35], which has been driven by advances in multimodal understanding learned from web-scaled paired (text, image) data. These advances + +have led to applications in domains ranging from content creation [21, 22, 31] to human-robot interaction [40]. Unfortunately, developing the analogue of a text-conditioned 3D shape generator is challenging because it is difficult to obtain (text, 3D shape) pairs at large scale. Prior work has attempted to address this problem by collecting text-shape paired data [2, 5, 9, 23, 26], but these approaches have been limited to a small number of object categories. + +A promising way around this data bottleneck is to use weak supervision from large-scale vision/language models such as CLIP [30]. One approach is to directly optimize a 3D representation such that (text, image render) pairs are aligned when projected into the CLIP embedding space. Prior work has applied this approach to stylize 3D meshes [25, 41] and to create abstract "dreamlike" objects using neural radiance fields [18] or meshes [19]. However, neither of the aforementioned methods produce realistic object geometry, and they can require expensive optimization. Another approach, more in line with text-to-image generators [31, 32], is to train a conditional generative model. The CLIP-Forge system [37] builds such a model without paired (text, shape) data by using rendered images of shapes at training time and leveraging the CLIP embedding space to bridge the modalities of image and text at inference time. CLIP-Forge demonstrates compelling zero-shot generation + +Table 1. High-level comparison between zero-shot text-to-shape generation methods. Inference time is calculated using a single NVIDIA Tesla V100 GPU. + +
MethodInferenceFidelityDiversity
DreamFields [18]>24 hrsLowSingle
Clip-Mesh [19]30 minLowSingle
Clip-Forge [37]6.41 msMediumMedium
CLIP-Sculptor0.91 secHighHigh
+ +abilities but produces low-fidelity shapes which do not capture the full diversity of shapes found in the training distribution. + +In this paper, we propose CLIP-Sculptor, a text-conditioned 3D shape generative model that outperforms the state of the art by improving shape diversity and fidelity with only (image, shape) pairs as supervision. CLIP-Sculptor's novelty lies in its multi-resolution, voxel-based conditional generation scheme. Without (text, shape) pairs, CLIP-Sculptor learns to generate 3D shapes of common object categories from the text-image joint embedding of CLIP. To achieve high fidelity outputs, CLIP-Sculptor adopts a multi-resolution approach: it first generates a low-resolution latent grid representation that captures the semantics from text/image, then upscales it to a higher resolution latent grid representation with a super-resolution model, and finally decodes the output geometry. To generate diverse shapes, CLIP-Sculptor adopts discrete latent representations which are obtained using a vector quantization scheme that avoids posterior collapse. To further improve shape fidelity and diversity, CLIP-Sculptor uses a masked transformer architecture. We additionally propose a novel annealed strategy for the classifier-free guidance [16]. To sum up, we make the following contributions: + +- CLIP-Sculptor, a multi-resolution text-conditional shape generative model that achieves both high fidelity and diversity without the need for (text, shape) pairs. +- A novel variant of classifier-free guidance for generative models, with an annealed guidance schedule to achieve better quality for a given diversity level. + +# 2. Related Work + +Neural Discrete Representation. Discrete representations [28] were first proposed for image generation in the context of variational autoencoders as a method to improve image quality and avoid posterior collapse. Further work introduced multi-scale and hierarchical variants [8, 33] which improved generative capabilities. Recently, 3D discrete representations have emerged as well, such as discretized voxel and implicit grids [26, 42]. In this work, we take inspiration + +from hierarchical VQ-VAEs [8, 33] and propose an architecture of hierarchical discrete representations capable of generating high fidelity 3D shapes. + +Latent Generative Models. In recent years, latent generative models have been widely adopted, because these models can effectively generate low-dimensional latent representations which can be used to generate high fidelity images and shapes. In the 3D domain, GAN-based latent models [1, 6, 17] have shown impressive results but tend to suffer from training instability and mode collapse. Flow-based models [37, 43] have been proposed to alleviate these problems, but they yield sample quality that is inferior to GAN-based models. Recent works in the image domain use diffusion [12, 36] or masking models [4] on the latent space which can increase the inference efficiency and still give quality outputs. Building on these works, we propose a hierarchical latent generative model that can further improve the fidelity and diversity of shape generations. + +Classifier-free Guidance. Classifier-free guidance, proposed by [16], jointly trains a conditional and unconditional generative model, whose score estimates are mixed to establish a trade-off between diversity measured by Frechet inception distance (FID) and sample accuracy/fidelity measured by Inception score (IS). Classifier-free guidance guides unconditional samples in the direction of conditional ones and has been implemented in practice to achieve state of the art results in recent work [10-12, 34]. To the best of our knowledge, guidance scale has always been kept constant outside the concurrent exploration in [15]. Their oscillating guidance technique, however does not show improvements in sample fidelity and produces more artifacts in their generations. In this work, we find that rather using annealed scheduling can give better quality versus diversity trade-offs. + +Text-to-Shape Generation. Text-to-shape generation has gained momentum in recent years. Recent works [5, 9, 23, 26] use supervised text-shape pairs to generate shapes effectively using natural language. However, a major drawback is the lack of available text-shape datasets, which limits supervised methods to generate shapes only in few categories. To solve this, several recent works have successfully leveraged the image-text embeddings of CLIP as a prior [30] by converting shapes into images using renders. One line of work uses differentiable renderers [18, 19, 25, 29, 41]. Another line of work learns a mapping from the image space to the shape space [37]. Although these methods are effective, they suffer from either long optimization time [18, 19, 25, 29] or limited quality of generated shapes [18, 19, 37]. Our method is able to generate more diverse shapes of higher fidelity within a comparably short inference time. + +![](images/b6b086398f90e1e54f1ccc07d1fea640e8f0c3159632c3ce5ccbaa9d0cb54d98.jpg) + +![](images/e5f560aab1b10a668d6587feb3f86cf3929f5075a0f1b53a86e7a5f887fcc25e.jpg) + +![](images/3a5c639818d65f0efcc63f11a4cc592b7513cbab1bd81100d3ba95dd12ac5634.jpg) +Figure 2. The CLIP-Sculptor architecture during training (top) and inference (bottom). CLIP-Sculptor is trained in three stages. In Stage 1, we train two separate VQ-VAE models for $32^{3}$ and $64^{3}$ resolution voxel grids. In Stage 2 we train a coarse transformer conditioned on a CLIP embedding to generate low resolution VQ-VAE latent grids $\mathbf{E}_{32}$ . In Stage 3, we train a fine transformer to perform super resolution on these latent grids. During inference, a text prompt is passed through the CLIP text encoder and used to condition the coarse transformer to generate a coarse latent grid $\mathbf{E}_{32}$ . This coarse grid is then used to condition the fine transformer to generate a fine latent grid $\mathbf{E}_{64}$ . Finally, this fine latent grid is then passed through the Training Stage 1 $64^{3}$ VQ-VAE decoder to generate the output shape. + +# 3. Method + +CLIP-Sculptor aims to generate diverse and high-fidelity 3D shapes of common object categories from text prompts. Figure 2 illustrates the components of our CLIP-Sculptor model. We train the network using 3 stages. In the first stage, we use a vector-quantized representation [28], to efficiently represent 3D shapes in discrete voxel grids. This latent representation captures the diverse semantic meanings of the 3D shapes in common object categories and is compatible with powerful generative models. To balance computation cost and shape quality, we utilize latent representations at two resolutions. A low resolution latent representation $\mathbf{E}_{32}$ encodes geometric information of voxels $\mathbf{V}_{32}$ at resolutions of $32^{3}$ , which is useful for text-conditional generation. A high resolution latent representation $\mathbf{E}_{64}$ encodes geometric information of voxel grids $\mathbf{V}_{64}$ at resolutions of $64^{3}$ , which is useful for super-resolution and detail generation. In stage 2, we train a generative model conditioned on CLIP features obtained from images $\{\mathbf{I}_r|r\in R\}$ rendered at random views $R$ to solve the lack of paired (text, shape) + +data. The conditional generative model consists of a mapping net and a coarse transformer $T_{c}(.)$ that is trained to unmask a masked low-resolution grid $\mathbf{E}_{32}$ . Finally in stage 3, a fine transformer $T_{f}(.)$ is used as a super-resolution model in latent space to generate higher resolution shapes. It is trained in a similar manner to the coarse transformer but conditioned on a low-resolution grid $\mathbf{E}_{32}$ instead of CLIP features. + +During inference, a text prompt is sequentially converted from CLIP text embedding to a low-resolution latent representation $\mathbf{E}_{32}$ to a high-resolution latent representation $\mathbf{E}_{64}$ to finally a high-resolution voxel grid $\mathbf{V}_{64}$ . A confidence-based iterative decoding scheme [4] is adopted for unmasking the fully masked initial grids for coarse and fine transformers at inference. To further control shape fidelity and diversity, we propose an annealing strategy for classifier-free guidance at inference time. + +In the following subsections, we will introduce the major components within CLIP-Sculptor. Section 3.1 describes the vector quantized representations at multiple resolutions; + +Section 3.2 explains the conditional generation model as well as its design choices and training strategies; Section 3.3 introduces the super-resolution model; Section 3.4 explains how the classifier-free guidance is achieved at inference using an annealing strategy. + +# 3.1. Multi-Resolution Voxel VQ-VAEs + +To learn a low-dimensional latent space that effectively represents the distribution of the training shape set $D$ , we use a vector-quantized variational autoencoder (VQVAE) [28]. The vector quantized latent representation avoids "posterior collapse", and it is also amenable to modeling with transformer-based generative models. To balance computational cost and shape quality, we train two separate VQ-VAEs for $32^{3}$ and $64^{3}$ resolution voxel grids $\mathbf{V}_{32}$ and $\mathbf{V}_{64}$ , which are useful for conditional generation and superresolution respectively. + +For both VQ-VAEs, we use ResNet-based [13] volumetric CNN encoders and decoders. A vector quantization layer maps the down-sampled voxel grids into a discrete latent space by indexing the continuous latent space with an embedding codebook. In this way, the voxel grids $\mathbf{V}_{32}$ and $\mathbf{V}_{64}$ are effectively represented by low-dimensional latent representations $\mathbf{E}_{32}$ and $\mathbf{E}_{64}$ , at the resolution of $4^{3}$ and $8^{3}$ . The VQ-VAEs are trained with mean squared error loss (MSE), commitment loss [28], and an exponential moving average strategy [33, 45]. + +# 3.2. CLIP-Conditioned Coarse Transformer + +The conditional generation process aims to generate diverse shapes that semantically correspond to the text prompts. Inspired by the recent success of masking approaches [4] and diffusion models [12, 27, 36], we formulate this generation process as a conditional unmasking task. Given an input masked latent representation $\mathbf{E}_{32}^{msk}$ and a conditional vector $\mathbf{c}$ , the coarse transformer $T_{c}(.)$ should produce an fully unmasked latent representation $\mathbf{E}_{32}$ that not only aligns with the semantic meaning captured by the conditional vector, but also allows for the sampling of diverse shapes at inference. + +The conditional vector $\mathbf{c}$ provides the semantic guidance for the shape generation process. Since we assume no text data at the training time, we leverage the text-image joint-embedding of CLIP. The generation is guided by the CLIP image embedding from the rendered images $\{\mathbf{I}_r\}$ during training and guided by the CLIP text embedding from the text prompt during inference. We add Gaussian noise to the CLIP image embeddings [44] to alleviate the modality gap between the CLIP text and image embeddings [20]: + +$$ +\hat {\mathbf {c}} = f _ {I} \left(\mathbf {I} _ {r}\right) + \frac {\gamma \cdot \epsilon \cdot \| f _ {I} \left(\mathbf {I} _ {r}\right) \| _ {2}}{\| \epsilon \| _ {2}}, \tag {1} +$$ + +where $f_{I}(\cdot)$ is the CLIP image encoder, $\epsilon \sim \mathcal{N}(0,1)$ , and $\gamma$ + +controls the level of perturbation. The conditional vector $\hat{\mathbf{c}}$ is then normalized and passed through a mapping network that consists of multi-layer perceptrons (MLP). The mapped vector $\tilde{\mathbf{c}}$ conditions on the coarse transformer $T_{c}(.)$ by predicting the affine transform parameters of each transformer block's layernorm. In order to control the diversity-quality balance of the generated shapes, we replace the conditional vector with the null embedding by a $\rho\%$ drop out rate for classifier-free guidance [16]. These are important design choices that affects the diversity and quality of the generated shapes at inference time. + +To train the coarse transformer $T_{c}(.)$ , we randomly replace the input indices with a special "mask" tokens by uniformly selecting a masking ratio between 0-100% [4]. Given the masked latent grid $\mathbf{E}_{32}^{msk}$ and the conditional vector $\mathbf{c}$ , the training objective is to maximize the following log-likelihood: $\sum_{n}^{N}\log p(\mathbf{E}_{32,n}|\mathbf{E}_{32,n}^{msk},\mathbf{c}) / N$ , where $N$ is the training data size and $\mathbf{E}_{32}^{msk}$ is the randomly masked latent grids. As the network sees only the ground truth tokens but not its own predicted samples during training, there exists an autoregressive drift resulted by the accumulative error at inference-time sampling. To alleviate this issue, we take inspiration from the NLP literature [38] and use a similar step-unrolled training (SUT) that predicts the entire un-masked grids at each time step. The final objective is to maximize the following formulation: + +$$ +\sum_ {n} ^ {N} \left(\log p \left(\mathbf {E} _ {3 2, n} \mid \mathbf {E} _ {3 2, n} ^ {m s k}, \mathbf {c}\right) + \log p \left(\mathbf {E} _ {3 2, n} \mid \tilde {\mathbf {E}} _ {3 2, n}, \mathbf {c}\right)\right) / N, \tag {2} +$$ + +where $\tilde{\mathbf{E}}_{32}$ is the coarse transformer $T_{c}(.)$ 's own prediction. + +# 3.3. Super-Resolution with Fine Transformer + +Super-resolution in the discrete latent space improves quality and upsamples the resolution of generated shapes. This process is done with a fine transformer $T_{f}(.)$ , which takes a low-resolution latent representation $\mathbf{E}_{32}$ and a masked high-resolution latent grid as input. The fine transformer is conditioned on the coarse grid $\mathbf{E}_{32}$ via cross attention. It unmasks the high-resolution representation $\mathbf{E}_{64}$ . To ensure that training data resembles the inference data, we directly use the predictions $\tilde{\mathbf{E}}_{32}$ of the coarse transformer $T_{c}(.)$ to train the fine transformer $T_{f}(.)$ , instead of using the ground truth low-resolution latent representation $\mathbf{E}_{32}$ . + +# 3.4. Annealed Classifier-Free Guidance + +At inference time, we first convert an input text prompt into a CLIP text embedding using the CLIP text encoder $f_{T}(\cdot)$ . The text embedding is then fed into the mapping network and used as the conditional vector $\tilde{\mathbf{c}}$ of the coarse transformer $T_{c}(\cdot)$ . We use the iterative decoding scheme [4] to slowly unmask the grid over a sequence of $T$ steps. The initial input to the coarse transformer $T_{c}(\cdot)$ is a completely + +masked latent grid. At each time step $t$ , we condition the coarse transformer $T_{c}(.)$ with $\tilde{\mathbf{c}}$ and take the output latent grid $\mathbf{E}_{32,t-1}^{msk}$ from the prior time step as the input. The coarse transformer $T_{c}(.)$ will predict an unmasked latent grid $\tilde{\mathbf{E}}_{32,t}$ . We mask all the tokens except for the most confident token predictions and the tokens unmasked from previous steps. This newly masked latent grid $\mathbf{E}_{32,t}^{msk}$ is the input to the next time step $t+1$ . We repeat this process until the process unmasks all the tokens, which is ensured by a cosine masking schedule [4]. + +During this iterative decoding scheme, we apply a new variant of classifier-free guidance [16]. Classifier-free guidance extrapolates an unconditional sample in the direction of a conditional sample. The extrapolation is controlled by an adaptive guidance scale parameter. Varying this parameter results in a trade-off between accuracy of text-to-shape generation and sample diversity. Instead of using a constant guidance scale for all time steps [16], we propose to vary the guidance scale over time according to an annealing schedule. The intuition is that during the initial steps of sampling when the input to the coarse transformer is mostly masked indices, a larger guidance scale is important to keep the network "on task"; later on in the sampling process, a lower guidance scale can help produce more sample diversity. Overall, the guidance is formulated as: + +$$ +\hat {P} _ {t} (c) = P _ {t} (0) + a (t) \left(P _ {t} (c) - P _ {t} (0)\right), \tag {3} +$$ + +where $a(t)$ is the guidance scale annealing schedule, $P_{t}(c)$ denotes the conditional distribution $p(\mathbf{E}_{32,t}|\mathbf{E}_{32,t - 1}^{msk},\mathbf{c})$ and $P_{t}(0)$ denotes $p(\mathbf{E}_{32,t}|\mathbf{E}_{32,t - 1}^{msk},\mathbf{0})$ . Note that $a(t)$ is continuous and usually monotonically decreasing. By setting $a(t) = k$ for some constant $k$ , this equation is equal to the classifier-free guidance scheme with a constant guidance scale. In this paper, we experimentally evaluate other annealing schedules [4], including linear, cosine, and square root functions. + +Finally, the unmasked coarse latent grid $\tilde{\mathbf{E}}_{32}$ is fed to the fine transformer $T_{f}$ , which uses the iterative decoding scheme to unmask an initially masked fine latent grid $\mathbf{E}_{64}^{msk}$ . The unmasked fine grid is then passed to the $64^3$ VQ-VAE decoder to obtain the final high-resolution voxel shape. + +# 4. Experiments + +In this section, we report experimental results to evaluate the generation fidelity, diversity, and class accuracy of CLIP-Sculptor. We provide full details of the experiment setup in the supplementary. We run the experiments 3 times for each of the aforementioned measures and report the mean in each case, except for the final comparisons with baselines (Table 2), where we use the best seed. Additional results can also be found in the supplementary. + +Dataset. We conduct our experiments on two subsets of the ShapeNet(v2) dataset [3]. The first subset, ShapeNet13, contains 13 categories from ShapeNet as used in [7,24]. We use the same train/test split as specified in [24]. Our second subset is ShapeNet55 which contains all 55 ShapeNet categories. For ShapeNet55, we render images as described in [7]; the training dataset contains 51,784 datapoints, whereas the test set contains 6,101 datapoints. + +Evaluation Metrics. To assess the generative capabilities of CLIP-Sculptor, we use Fréchet Inception Distance (FID) ([14]) to evaluate shape diversity and quality, and use Classifier Accuracy (Acc) to evaluate the text-shape correspondence. These metrics follow [37] where they first generated single mean shapes for 234 predetermined text queries and then passed all shapes through a classifier to measure Acc. The latent space of this classifier is also used for FID. + +Baseline. We compare the performance of our method against CLIP-Forge [37], CLIP-Mesh [19] and DreamFields [18], which are currently state of the art for zero-shot text-to-shape generation. We also set up a comparison method called Zero-Shot AutoSDF(ZS-ASDF), which trains the supervised method AutoSDF [26] with the CLIP features as CLIP-Sculptor, to highlight the superiority of our conditional generation model. CLIP-Forge reports results on single shape generation for 234 text queries. These single shapes were generated using the mean of the prior (the Gaussian distribution), so we refer to these results as CF-MS. Note that this does not capture the diversity of shapes that can be generated given a text query. To capture results on more shape generations, we sample 32 shapes instead of one using a Gaussian (CF-G), truncated Gaussian (CF-TG) or clipped Gaussian (CF-CG) distribution. We also generate 32 shapes for ZS-ASDF and our method for fair comparison. We only compare qualitatively with DreamFields and CLIP-Mesh as the optimization time per query is significant (Table 1), and it does not use prior knowledge from the shape dataset. + +# 4.1. Evaluating Shape Diversity and Accuracy + +In this section, we first quantitatively evaluate the diversity and accuracy of a given shape matching a given text query on the ShapeNet13 dataset. We compare with CLIP-Forge and ZS-ASDF using the Acc and FID metrics. The results are shown in Table 2. The first four columns represent CLIP-Forge(CF) and its different sampling techniques. The fifth column represents the result from AutoSDF trained with CLIP features(ZS-ASDF). The other columns represent our method with different annealing scale strategies. Three major things can be observed. First, all variants of our method outperform CLIP-Forge and ZS-ASDF significantly. This indicates that the method produces more diverse and higher fidelity shapes of increasing accuracy. Second, it can be seen that annealing strategies + +![](images/4d2b11d33bf405305c5d336a25d1e06a4cdb9aa618c6931d990ba514a23a45c7.jpg) +Figure 3. Qualitative comparison of CLIP-Sculptor (rendered in purple), CLIP-Forge [37](rendered in green), DreamField [18], and CLIP-Mesh [19] on text-conditioned generation. For each text prompt, the first generated shape by CLIP-Sculptor and CLIP-Forge [37] is zoomed in for detail comparison. Comparing with other methods, CLIP-Sculptor generates a broader category of shapes that correspond with the text prompt. The generated shapes are more diverse and has higher fidelity. + +for guidance scale give a better accuracy-diversity trade-off than constant scale guidance. Finally, we note that simply augmenting AutoSDF (another VQ-VAE) with CLIP features does not lead to improved text to shape generation, indicating the importance of our architecture design decisions. + +We also qualitatively compare our method to CLIP-Forge, CLIP-Mesh, and DreamFields. The results are shown in Figure 3. It can be observed that our method generates higher quality shapes (view "a machine gun"), with more detail (view "an office chair") and higher diversity (view "a jet"). CLIP-Forge usually produces the same shape + +Table 2. Comparisons of CLIP-Forge(CF) baseline (across different sampling strategies), AutoSDF augmented with CLIP (ZS-ASDF) with CLIP-Sculptor (CS) on Accuracy(ACC) and FID. CLIP-Sculptor outperforms other methods in both metrics, which proves it generates shapes with higher fidelity and better diversity. + +
MethodCF-MSCF-GCF-TGCF-CGZS-ASDFCS-const.CS-sqrtCS-linearCS-cosine
FID ↓2425.252233.482141.612100.677332.931821.781480.111629.511725.63
ACC ↑83.3362.8168.7171.1139.1486.5987.0887.5087.27
+ +Table 3. Ablations on the major components of CLIP-Sculptor. CLIP-Sculptor achieves the best performance comparing with other design choice, which proves the effectiveness of the major components. +(a) Effect of varying the Noise parameter. + +
γ×0.50.81.01.21.5
FID↓1720.021764.981484.611703.381447.911478.17
ACC↑64.8775.7377.4179.0979.6378.47
(b) Effect of varying number of layers (L) in the mapping network.
L012345
FID↓2874.871716.731518.971447.911532.501424.46
ACC↑62.7278.7079.1779.6377.1676.09
(c) Comparison with baselines on super resolution.
Method3D-UNet643-DSCLIP-Sculptor
FID↓2056.922196.961910.28
ACC↑86.6577.9286.85
+ +with small variations (view "a rectangular table"). We also note that DreamFields and CLIP-Mesh produce very abstract results which may not be useful in many applications. Finally, we also show results on ShapeNet55 in the last row of Figure 3. We could not get CLIP-Forge to produce sensible shapes for most text queries on ShapeNet55 (which we attribute to the data imbalance issue of ShapeNet55), whereas our method produced high fidelity shapes. + +# 4.2. Major Components for Conditional Generation + +Effect of Noise Parameter. During Stage 2 Training, to better align the image and text embeddings, we added Gaussian noise at varying levels of $\gamma$ . We investigate the effect of $(\gamma)$ added to the image condition vector and show the results in Table 3 (a). We keep the number of mapping layers fixed at three in this experiment. We observe that adding Gaussian noise drastically improves both the diversity and accuracy of generation, with the optimal noise parameter being around 1.2. These results indicate that adding noise during training helps align the text features seen at inference and the image features seen at training. + +Size of Mapping Network. We next probe the importance of the mapping network. We show the results in Table 3 (b), where $L$ represents the number of layers of the mapping function. In the case of 0 layers, we directly project conditional embeddings to the layernorm parameters using a linear layer. We make two observations from the results: 1) A common mapping network improves both the accuracy and FID. 2) Increasing the number of mapping network layers beyond a certain number decreases accuracy. + +Classifier-Free Guidance. We next explore the relationship between dropping out image conditioning at $\rho \%$ during training and using constant classifier-free guidance during test time. In Table 4, we vary the scale parameter, $a(t) = k$ , across the columns. It can be seen that as we increase the scale, accuracy typically increases while FID decreases. This indicates that the method is giving more accurate results on a given text query at the cost of shape diversity. A low dropout rate (5-15%) gives a good trade-off between accuracy and diversity. + +Step-Unrolled Training (SUT). Finally, to further improve the accuracy we investigate the use of step-unrolled training [39]. The results are shown in the last row of Table 4. From the table it can be observed that step-unrolled training enables higher accuracy across all scales. This indicates that the model learns to unmask samples it would encounter during inference time. + +# 4.3. Annealing Strategy for Scale Parameter + +An important idea we propose in this paper is a scale annealing technique for classifier-free guidance. We employ classifier-free guidance to identify a better accuracy-diversity trade-off with different annealing schedules: constant, linear, cosine and square root. To determine which annealing technique works the best, we fix the FID values and plotted the accuracy at each fixed FID value (Figure 4). As different scale parameters give different FID values, we conducted an extensive grid search over the starting scale parameter. Note that finding the exact FID is not always feasible, so we pick the closest FID. Figure 4 shows results of FID versus accuracy on three different runs of the Coarse Transformer. We find that across all three runs, the accuracy is typically lower for a given FID for constant schedules as + +Table 4. Classifier-Free Guidance and Step-Unrolled Training (SUT) experiment results. $\rho \%$ is the drop out rate for classifier-free guidance. SUT indicates the use of Step-Unrolled Training. The remaining columns indicate the variation in scale parameter. + +
ρ%SUTa(t) = 3a(t) = 2.5a(t) = 2a(t) = 1.5a(t) = 1
FID↓Acc↑FID↓Acc↑FID↓Acc↑FID↓Acc↑FID↓Acc↑
5×2059.085.731970.185.471790.985.431536.583.981227.579.17
10×1893.284.461821.284.501684.383.411522.482.341348.477.76
15×2086.985.031964.784.991851.984.361660.983.141485.578.19
20×2062.583.391972.382.951892.982.741733.381.131566.675.94
5✓2039.887.692011.187.391811.887.401678.686.241517.982.18
+ +![](images/0cd7ff696dfc73d7523e291fc7cf6e556d36b6e5463c94b1fde9453ba86509ce.jpg) +Figure 4. FID and Acc results with different classifier-free guidance scale annealing strategies (constant, cosine, square root, linear) across three different runs (different seeds) of the CLIP-Sculptor Stage 2 Coarse Transformer. + +![](images/b761d6b05adefe8d1aa43a87eea9f07b6b2b79c107a38e92ae487cdbdeabaad9.jpg) + +![](images/b567cbab04f84aedc64cb29439a8d1f655ece0eb4525882548c2d908d9c54e6b.jpg) + +opposed to other schedules. This is especially the case at the lower range of FID values. The results indicate that having a large scale at the beginning of sampling is more important than later stages, especially in use cases where diversity is paramount. + +# 4.4. Super-Resolution + +Finally, we investigate the importance of hierarchical super-resolution. We compare our method with two baselines. In the first baseline, we directly use the $32^{3}$ resolution results from the coarse transformer and use a 3D-UNET based super-resolution network to translate from $32^{3}$ to $64^{3}$ . For the second baseline, we train a transformer directly on $64^{3}$ VQ-VAE that is conditioned on text features as opposed to a coarse resolution grid. We refer to this as $64^{3}$ -DS (64 $^{3}$ direct synthesis). The results are shown in Table 3(c). It can be seen from the table that latent-based super-resolution outperforms the baselines in both accuracy and diversity. + +# 5. Conclusion + +We present CLIP-Sculptor, a text-to-3D-shape generation method capable of producing shapes of high fidelity and diversity without the need for (text, shape) pairs. To achieve this, CLIP-Sculptor leverages CLIP and implements super-resolution in a discrete latent space with a hierarchical architecture and a novel annealed variant of classifier-free guidance on a mask-based model. We validate CLIP-Sculptor by comparing it with a number of base + +lines in terms of FID and accuracy, finding that CLIP-Sculptor is the new state of the art for this problem. In experimenting with different guidance scale scheduling, we find that constant scale scheduling did not always work the best, an important finding for the diffusion modeling community that may improve generation quality. Our paper is an important step towards improved quality and diversity of 3D text-to-shape generation outcomes. + +Limitations and Societal Impact. Despite providing higher quality over baseline methods, CLIP-Sculptor still suffers from limitations in its ability to capture smaller details (e.g., chair with a hole on its back) and generate shapes from text prompts involving counting (e.g., chair with four slats). Furthermore, CLIP-Sculptor also fails on many text queries which are not present in the prior knowledge of CLIP. The method also lacks the ability to produce texture. In future work, we will address these limitations by exploring implicit representations that can capture even finer details and by investigating neural networks that can count. + +Text-to-shape approaches can make the generation of 3D content more accessible and efficient. Within the text-to-image domain there has been concern that automatic approaches may replace artists. Because our approach produces voxels, a representation that lends well to direct manipulation and handoff to 3D designers, we believe that our text-to-3D approach can augment 3D design workflows rather than replace them. + +# References + +[1] Panos Achlioptas, Olga Diamanti, Ioannis Mitliagkas, and Leonidas Guibas. Learning representations and generative models for 3d point clouds. In International conference on machine learning, pages 40-49. PMLR, 2018. 2 +[2] Panos Achlioptas, Judy Fan, X.D. Robert Hawkins, D. Noah Goodman, and J. Leonidas Guibas. ShapeGlot: Learning language for shape differentiation. CoRR, abs/1905.02925, 2019. 1 +[3] Angel X. Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, Jianxiong Xiao, Li Yi, and Fisher Yu. Shapenet: An information-rich 3d model repository, 2015. cite arxiv:1512.03012. 5 +[4] Huiwen Chang, Han Zhang, Lu Jiang, Ce Liu, and William T Freeman. Maskgit: Masked generative image transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11315-11325, 2022. 2, 3, 4, 5 +[5] Kevin Chen, Christopher B Choy, Manolis Savva, Angel X Chang, Thomas Funkhouser, and Silvio Savarese. Text2shape: Generating shapes from natural language by learning joint embeddings. In Asian Conference on Computer Vision, pages 100-116. Springer, 2018. 1, 2 +[6] Zhiqin Chen and Hao Zhang. Learning implicit fields for generative shape modeling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5939-5948, 2019. 2 +[7] Christopher B Choy, Danfei Xu, JunYoung Gwak, Kevin Chen, and Silvio Savarese. 3d-r2n2: A unified approach for single and multi-view 3d object reconstruction. In European conference on computer vision, pages 628-644. Springer, 2016. 5 +[8] Prafulla Dhariwal, Heewoo Jun, Christine Payne, Jong Wook Kim, Alec Radford, and Ilya Sutskever. Jukebox: A generative model for music. arXiv preprint arXiv:2005.00341, 2020. 2 +[9] Rao Fu, Xiao Zhan, Yiwen Chen, Daniel Ritchie, and Srinath Sridhar. Shapecrafter: A recursive text-conditioned 3d shape generation model, 2022. 1, 2 +[10] Oran Gafni, Adam Polyak, Oron Ashual, Shelly Sheynin, Devi Parikh, and Yaniv Taigman. Make-a-scene: Scene-based text-to-image generation with human priors, 2022. 2 +[11] Rinon Gal, Yuval Alaluf, Yuval Atzmon, Or Patashnik, Amit H. Bermano, Gal Chechik, and Daniel Cohen-Or. An image is worth one word: Personalizing text-to-image generation using textual inversion, 2022. 2 +[12] Shuyang Gu, Dong Chen, Jianmin Bao, Fang Wen, Bo Zhang, Dongdong Chen, Lu Yuan, and Baining Guo. Vector quantized diffusion model for text-to-image synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10696-10706, 2022. 2, 4 +[13] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 4 + +[14] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 5 +[15] Jonathan Ho, William Chan, Chitwan Sahara, Jay Whang, Ruiqi Gao, Alexey Gritsanko, Diederik P. Kingma, Ben Poole, Mohammad Norouzi, David J. Fleet, and Tim Salimans. Imagen video: High definition video generation with diffusion models, 2022. 2 +[16] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598, 2022. 2, 4, 5 +[17] Moritz Ibing, Isaak Lim, and Leif Kobbelt. 3d shape generation with grid-based implicit functions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13559-13568, 2021. 2 +[18] Ajay Jain, Ben Mildenhall, Jonathan T Barron, Pieter Abbeel, and Ben Poole. Zero-shot text-guided object generation with dream fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 867-876, 2022. 1, 2, 5, 6 +[19] Nasir Khalid, Tianhao Xie, Eugene Belilovsky, and Tiberiu Popa. Clip-mesh: Generating textured meshes from text using pretrained image-text models. ACM Transactions on Graphics (TOG), Proc. SIGGRAPH Asia, 2022. 1, 2, 5, 6 +[20] Weixin Liang, Yuhui Zhang, Yongchan Kwon, Serena Yeung, and James Zou. Mind the gap: Understanding the modality gap in multi-modal contrastive representation learning. arXiv preprint arXiv:2203.02053, 2022. 4 +[21] Vivian Liu, Han Qiao, and Lydia Chilton. Opal: Multimodal image generation for news illustration, 2022. 1 +[22] Vivian Liu, Jo Vermeulen, George Fitzmaurice, and Justin Matejka. 3dall-e: Integrating text-to-image ai in 3d design workflows, 2022. 1 +[23] Zhengzhe Liu, Yi Wang, Xiaojuan Qi, and Chi-Wing Fu. Towards implicit text-guided 3d shape generation. arXiv preprint arXiv:2203.14622, 2022. 1, 2 +[24] Lars Mescheder, Michael Oechsle, Michael Niemeyer, Sebastian Nowozin, and Andreas Geiger. Occupancy networks: Learning 3d reconstruction in function space. In Proceedings IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2019. 5 +[25] Oscar Michel, Roi Bar-On, Richard Liu, Sagie Benaim, and Rana Hanocka. Text2mesh: Text-driven neural stylization for meshes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13492-13502, 2022. 1, 2 +[26] Paritosh Mittal, Yen-Chi Cheng, Maneesh Singh, and Shubham Tulsiani. Autosdf: Shape priors for 3d completion, reconstruction and generation. arXiv preprint arXiv:2203.09516, 2022. 1, 2, 5 +[27] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models, 2021. 4 +[28] Aaron van den Oord, Oriol Vinyals, and Koray Kavukcuoglu. Neural discrete representation learning. arXiv preprint arXiv:1711.00937, 2017. 2, 3, 4 + +[29] Ben Poole, Ajay Jain, Jonathan T. Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. arXiv, 2022. 2 +[30] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision, 2021. 1, 2 +[31] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 2022. 1 +[32] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In International Conference on Machine Learning, pages 8821-8831. PMLR, 2021. 1 +[33] Ali Razavi, Aaron Van den Oord, and Oriol Vinyals. Generating diverse high-fidelity images with vq-vae-2. Advances in neural information processing systems, 32, 2019. 2, 4 +[34] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models, 2021. 2 +[35] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10684-10695, June 2022. 1 +[36] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10684-10695, 2022. 2, 4 +[37] Aditya Sanghi, Hang Chu, Joseph G Lambourne, Ye Wang, Chin-Yi Cheng, Marco Fumero, and Kamal Rahimi Malekshan. Clip-forge: Towards zero-shot text-to-shape generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18603-18613, 2022. 1, 2, 5, 6 +[38] Nikolay Savinov, Junyoung Chung, Mikolaj Binkowski, Erich Elsen, and Aaron van den Oord. Step-unrolled denoising autoencoders for text generation. arXiv preprint arXiv:2112.06749, 2021. 4 +[39] Nikolay Savinov, Junyoung Chung, Mikolaj Binkowski, Erich Elsen, and Aaron van den Oord. Step-unrolled denoising autoencoders for text generation, 2021. 7 +[40] Mohit Shridhar, Lucas Manuelli, and Dieter Fox. CIport: What and where pathways for robotic manipulation. In Proceedings of the 5th Conference on Robot Learning (CoRL), 2021. 1 +[41] Can Wang, Menglei Chai, Mingming He, Dongdong Chen, and Jing Liao. Clip-nerf: Text-and-image driven manipulation of neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3835–3844, 2022. 1, 2 +[42] Xingguang Yan, Liqiang Lin, Niloy J Mitra, Dani Lischinski, Daniel Cohen-Or, and Hui Huang. Shapeformer: + +Transformer-based shape completion via sparse representation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6239-6249, 2022. 2 +[43] Guandao Yang, Xun Huang, Zekun Hao, Ming-Yu Liu, Serge Belongie, and Bharath Hariharan. Pointflow: 3d point cloud generation with continuous normalizing flows. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4541-4550, 2019. 2 +[44] Yufan Zhou, Ruiyi Zhang, Changyou Chen, Chunyuan Li, Chris Tensmeyer, Tong Yu, Jiuming Gu, Jinhui Xu, and Tong Sun. Lafite: Towards language-free training for text-to-image generation. arXiv preprint arXiv:2111.13792, 2021. 4 +[45] Adrian Łancucki, Jan Chorowski, Guillaume Sanchez, Ricard Marxer, Nanxin Chen, Hans J. G. A. Dolfing, Sameer Khurana, Tanel Alumäe, and Antoine Laurent. Robust training of vector quantized bottleneck models, 2020. 4 \ No newline at end of file diff --git a/2023/CLIP-Sculptor_ Zero-Shot Generation of High-Fidelity and Diverse Shapes From Natural Language/images.zip b/2023/CLIP-Sculptor_ Zero-Shot Generation of High-Fidelity and Diverse Shapes From Natural Language/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..a30578fcbbdb95db861fb57f80adcb395dea51c6 --- /dev/null +++ b/2023/CLIP-Sculptor_ Zero-Shot Generation of High-Fidelity and Diverse Shapes From Natural Language/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b1989fa2bbd74d1df2f889dd9834eac802507ab655dd47e8698403c5b917245c +size 614368 diff --git a/2023/CLIP-Sculptor_ Zero-Shot Generation of High-Fidelity and Diverse Shapes From Natural Language/layout.json b/2023/CLIP-Sculptor_ Zero-Shot Generation of High-Fidelity and Diverse Shapes From Natural Language/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..b455ad44d3a10f1eea03580150462b8626a6ce54 --- /dev/null +++ b/2023/CLIP-Sculptor_ Zero-Shot Generation of High-Fidelity and Diverse Shapes From Natural Language/layout.json @@ -0,0 +1,7327 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 48, + 102, + 544, + 139 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 102, + 544, + 139 + ], + "spans": [ + { + "bbox": [ + 48, + 102, + 544, + 139 + ], + "type": "text", + "content": "CLIP-Sculptor: Zero-Shot Generation of High-Fidelity and Diverse Shapes from Natural Language" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 83, + 159, + 515, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 83, + 159, + 515, + 204 + ], + "spans": [ + { + "bbox": [ + 83, + 159, + 515, + 204 + ], + "type": "text", + "content": "Aditya Sanghi† Rao Fu‡ Vivian Liu§ Karl D.D. Willis† Hooman Shayani† Amir H. Khasahmadi† Srinath Sridhar‡ Daniel Ritchie‡ Autodesk Research† Brown University‡ Columbia University§" + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 47, + 230, + 386, + 358 + ], + "blocks": [ + { + "bbox": [ + 47, + 230, + 386, + 358 + ], + "lines": [ + { + "bbox": [ + 47, + 230, + 386, + 358 + ], + "spans": [ + { + "bbox": [ + 47, + 230, + 386, + 358 + ], + "type": "image", + "image_path": "af57706f733e87fff2175388558c77f34e475f0ab3b75f9eb54c922df9ff6187.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 45, + 368, + 546, + 402 + ], + "lines": [ + { + "bbox": [ + 45, + 368, + 546, + 402 + ], + "spans": [ + { + "bbox": [ + 45, + 368, + 546, + 402 + ], + "type": "text", + "content": "Figure 1. CLIP-Sculptor is a zero-shot text-to-shape generation method. Left Top/Bottom: It generates diverse shapes that reflect the semantic meaning of the text input without requiring any text data during training. Right: The method also generates high-fidelity shapes." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 389, + 232, + 545, + 293 + ], + "blocks": [ + { + "bbox": [ + 389, + 232, + 545, + 293 + ], + "lines": [ + { + "bbox": [ + 389, + 232, + 545, + 293 + ], + "spans": [ + { + "bbox": [ + 389, + 232, + 545, + 293 + ], + "type": "image", + "image_path": "053674f60e974e7a9db3708cd195027dec70fd24adfb6edc9f662af9bafbbe9c.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 389, + 299, + 545, + 357 + ], + "blocks": [ + { + "bbox": [ + 389, + 299, + 545, + 357 + ], + "lines": [ + { + "bbox": [ + 389, + 299, + 545, + 357 + ], + "spans": [ + { + "bbox": [ + 389, + 299, + 545, + 357 + ], + "type": "image", + "image_path": "8867c4b6a962f70f75737786dbb4d89448714c10757f7851ef4c0ac23c93e158.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 143, + 407, + 192, + 420 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 407, + 192, + 420 + ], + "spans": [ + { + "bbox": [ + 143, + 407, + 192, + 420 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 45, + 433, + 290, + 625 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 433, + 290, + 625 + ], + "spans": [ + { + "bbox": [ + 45, + 433, + 290, + 625 + ], + "type": "text", + "content": "Recent works have demonstrated that natural language can be used to generate and edit 3D shapes. However, these methods generate shapes with limited fidelity and diversity. We introduce CLIP-Sculptor, a method to address these constraints by producing high-fidelity and diverse 3D shapes without the need for (text, shape) pairs during training. CLIP-Sculptor achieves this in a multi-resolution approach that first generates in a low-dimensional latent space and then upscales to a higher resolution for improved shape fidelity. For improved shape diversity, we use a discrete latent space which is modeled using a transformer conditioned on CLIP's image-text embedding space. We also present a novel variant of classifier-free guidance, which improves the accuracy-diversity trade-off. Finally, we perform extensive experiments demonstrating that CLIP-Sculptor outperforms state-of-the-art baselines." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 645, + 128, + 658 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 645, + 128, + 658 + ], + "spans": [ + { + "bbox": [ + 47, + 645, + 128, + 658 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "type": "text", + "content": "In recent years, there has been rapid progress in text-conditioned image generation [31, 32, 35], which has been driven by advances in multimodal understanding learned from web-scaled paired (text, image) data. These advances" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 409, + 546, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 409, + 546, + 506 + ], + "spans": [ + { + "bbox": [ + 304, + 409, + 546, + 506 + ], + "type": "text", + "content": "have led to applications in domains ranging from content creation [21, 22, 31] to human-robot interaction [40]. Unfortunately, developing the analogue of a text-conditioned 3D shape generator is challenging because it is difficult to obtain (text, 3D shape) pairs at large scale. Prior work has attempted to address this problem by collecting text-shape paired data [2, 5, 9, 23, 26], but these approaches have been limited to a small number of object categories." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 510, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 510, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 510, + 547, + 714 + ], + "type": "text", + "content": "A promising way around this data bottleneck is to use weak supervision from large-scale vision/language models such as CLIP [30]. One approach is to directly optimize a 3D representation such that (text, image render) pairs are aligned when projected into the CLIP embedding space. Prior work has applied this approach to stylize 3D meshes [25, 41] and to create abstract \"dreamlike\" objects using neural radiance fields [18] or meshes [19]. However, neither of the aforementioned methods produce realistic object geometry, and they can require expensive optimization. Another approach, more in line with text-to-image generators [31, 32], is to train a conditional generative model. The CLIP-Forge system [37] builds such a model without paired (text, shape) data by using rendered images of shapes at training time and leveraging the CLIP embedding space to bridge the modalities of image and text at inference time. CLIP-Forge demonstrates compelling zero-shot generation" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "18339" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 49, + 113, + 286, + 186 + ], + "blocks": [ + { + "bbox": [ + 47, + 70, + 288, + 103 + ], + "lines": [ + { + "bbox": [ + 47, + 70, + 288, + 103 + ], + "spans": [ + { + "bbox": [ + 47, + 70, + 288, + 103 + ], + "type": "text", + "content": "Table 1. High-level comparison between zero-shot text-to-shape generation methods. Inference time is calculated using a single NVIDIA Tesla V100 GPU." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 49, + 113, + 286, + 186 + ], + "lines": [ + { + "bbox": [ + 49, + 113, + 286, + 186 + ], + "spans": [ + { + "bbox": [ + 49, + 113, + 286, + 186 + ], + "type": "table", + "html": "
MethodInferenceFidelityDiversity
DreamFields [18]>24 hrsLowSingle
Clip-Mesh [19]30 minLowSingle
Clip-Forge [37]6.41 msMediumMedium
CLIP-Sculptor0.91 secHighHigh
", + "image_path": "29d39bd1a16ea94f48a8966f9eebfcae48acbf9b30970907a62e81e76f8abe76.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 205, + 286, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 205, + 286, + 240 + ], + "spans": [ + { + "bbox": [ + 46, + 205, + 286, + 240 + ], + "type": "text", + "content": "abilities but produces low-fidelity shapes which do not capture the full diversity of shapes found in the training distribution." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 241, + 287, + 492 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 241, + 287, + 492 + ], + "spans": [ + { + "bbox": [ + 46, + 241, + 287, + 492 + ], + "type": "text", + "content": "In this paper, we propose CLIP-Sculptor, a text-conditioned 3D shape generative model that outperforms the state of the art by improving shape diversity and fidelity with only (image, shape) pairs as supervision. CLIP-Sculptor's novelty lies in its multi-resolution, voxel-based conditional generation scheme. Without (text, shape) pairs, CLIP-Sculptor learns to generate 3D shapes of common object categories from the text-image joint embedding of CLIP. To achieve high fidelity outputs, CLIP-Sculptor adopts a multi-resolution approach: it first generates a low-resolution latent grid representation that captures the semantics from text/image, then upscales it to a higher resolution latent grid representation with a super-resolution model, and finally decodes the output geometry. To generate diverse shapes, CLIP-Sculptor adopts discrete latent representations which are obtained using a vector quantization scheme that avoids posterior collapse. To further improve shape fidelity and diversity, CLIP-Sculptor uses a masked transformer architecture. We additionally propose a novel annealed strategy for the classifier-free guidance [16]. To sum up, we make the following contributions:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 59, + 500, + 286, + 579 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 59, + 500, + 286, + 536 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 500, + 286, + 536 + ], + "spans": [ + { + "bbox": [ + 59, + 500, + 286, + 536 + ], + "type": "text", + "content": "- CLIP-Sculptor, a multi-resolution text-conditional shape generative model that achieves both high fidelity and diversity without the need for (text, shape) pairs." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 59, + 544, + 286, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 544, + 286, + 579 + ], + "spans": [ + { + "bbox": [ + 59, + 544, + 286, + 579 + ], + "type": "text", + "content": "- A novel variant of classifier-free guidance for generative models, with an annealed guidance schedule to achieve better quality for a given diversity level." + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 47, + 597, + 134, + 609 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 597, + 134, + 609 + ], + "spans": [ + { + "bbox": [ + 47, + 597, + 134, + 609 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 617, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 617, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 617, + 286, + 713 + ], + "type": "text", + "content": "Neural Discrete Representation. Discrete representations [28] were first proposed for image generation in the context of variational autoencoders as a method to improve image quality and avoid posterior collapse. Further work introduced multi-scale and hierarchical variants [8, 33] which improved generative capabilities. Recently, 3D discrete representations have emerged as well, such as discretized voxel and implicit grids [26, 42]. In this work, we take inspiration" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 72, + 545, + 109 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 545, + 109 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 545, + 109 + ], + "type": "text", + "content": "from hierarchical VQ-VAEs [8, 33] and propose an architecture of hierarchical discrete representations capable of generating high fidelity 3D shapes." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 122, + 545, + 301 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 122, + 545, + 301 + ], + "spans": [ + { + "bbox": [ + 304, + 122, + 545, + 301 + ], + "type": "text", + "content": "Latent Generative Models. In recent years, latent generative models have been widely adopted, because these models can effectively generate low-dimensional latent representations which can be used to generate high fidelity images and shapes. In the 3D domain, GAN-based latent models [1, 6, 17] have shown impressive results but tend to suffer from training instability and mode collapse. Flow-based models [37, 43] have been proposed to alleviate these problems, but they yield sample quality that is inferior to GAN-based models. Recent works in the image domain use diffusion [12, 36] or masking models [4] on the latent space which can increase the inference efficiency and still give quality outputs. Building on these works, we propose a hierarchical latent generative model that can further improve the fidelity and diversity of shape generations." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 316, + 545, + 507 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 316, + 545, + 507 + ], + "spans": [ + { + "bbox": [ + 304, + 316, + 545, + 507 + ], + "type": "text", + "content": "Classifier-free Guidance. Classifier-free guidance, proposed by [16], jointly trains a conditional and unconditional generative model, whose score estimates are mixed to establish a trade-off between diversity measured by Frechet inception distance (FID) and sample accuracy/fidelity measured by Inception score (IS). Classifier-free guidance guides unconditional samples in the direction of conditional ones and has been implemented in practice to achieve state of the art results in recent work [10-12, 34]. To the best of our knowledge, guidance scale has always been kept constant outside the concurrent exploration in [15]. Their oscillating guidance technique, however does not show improvements in sample fidelity and produces more artifacts in their generations. In this work, we find that rather using annealed scheduling can give better quality versus diversity trade-offs." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 521, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 521, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 521, + 545, + 713 + ], + "type": "text", + "content": "Text-to-Shape Generation. Text-to-shape generation has gained momentum in recent years. Recent works [5, 9, 23, 26] use supervised text-shape pairs to generate shapes effectively using natural language. However, a major drawback is the lack of available text-shape datasets, which limits supervised methods to generate shapes only in few categories. To solve this, several recent works have successfully leveraged the image-text embeddings of CLIP as a prior [30] by converting shapes into images using renders. One line of work uses differentiable renderers [18, 19, 25, 29, 41]. Another line of work learns a mapping from the image space to the shape space [37]. Although these methods are effective, they suffer from either long optimization time [18, 19, 25, 29] or limited quality of generated shapes [18, 19, 37]. Our method is able to generate more diverse shapes of higher fidelity within a comparably short inference time." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18340" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 47, + 70, + 313, + 255 + ], + "blocks": [ + { + "bbox": [ + 47, + 70, + 313, + 255 + ], + "lines": [ + { + "bbox": [ + 47, + 70, + 313, + 255 + ], + "spans": [ + { + "bbox": [ + 47, + 70, + 313, + 255 + ], + "type": "image", + "image_path": "b6b086398f90e1e54f1ccc07d1fea640e8f0c3159632c3ce5ccbaa9d0cb54d98.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 315, + 70, + 545, + 255 + ], + "blocks": [ + { + "bbox": [ + 315, + 70, + 545, + 255 + ], + "lines": [ + { + "bbox": [ + 315, + 70, + 545, + 255 + ], + "spans": [ + { + "bbox": [ + 315, + 70, + 545, + 255 + ], + "type": "image", + "image_path": "e5f560aab1b10a668d6587feb3f86cf3929f5075a0f1b53a86e7a5f887fcc25e.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 48, + 255, + 545, + 358 + ], + "blocks": [ + { + "bbox": [ + 48, + 255, + 545, + 358 + ], + "lines": [ + { + "bbox": [ + 48, + 255, + 545, + 358 + ], + "spans": [ + { + "bbox": [ + 48, + 255, + 545, + 358 + ], + "type": "image", + "image_path": "3a5c639818d65f0efcc63f11a4cc592b7513cbab1bd81100d3ba95dd12ac5634.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 368, + 546, + 434 + ], + "lines": [ + { + "bbox": [ + 46, + 368, + 546, + 434 + ], + "spans": [ + { + "bbox": [ + 46, + 368, + 546, + 434 + ], + "type": "text", + "content": "Figure 2. The CLIP-Sculptor architecture during training (top) and inference (bottom). CLIP-Sculptor is trained in three stages. In Stage 1, we train two separate VQ-VAE models for " + }, + { + "bbox": [ + 46, + 368, + 546, + 434 + ], + "type": "inline_equation", + "content": "32^{3}" + }, + { + "bbox": [ + 46, + 368, + 546, + 434 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 368, + 546, + 434 + ], + "type": "inline_equation", + "content": "64^{3}" + }, + { + "bbox": [ + 46, + 368, + 546, + 434 + ], + "type": "text", + "content": " resolution voxel grids. In Stage 2 we train a coarse transformer conditioned on a CLIP embedding to generate low resolution VQ-VAE latent grids " + }, + { + "bbox": [ + 46, + 368, + 546, + 434 + ], + "type": "inline_equation", + "content": "\\mathbf{E}_{32}" + }, + { + "bbox": [ + 46, + 368, + 546, + 434 + ], + "type": "text", + "content": ". In Stage 3, we train a fine transformer to perform super resolution on these latent grids. During inference, a text prompt is passed through the CLIP text encoder and used to condition the coarse transformer to generate a coarse latent grid " + }, + { + "bbox": [ + 46, + 368, + 546, + 434 + ], + "type": "inline_equation", + "content": "\\mathbf{E}_{32}" + }, + { + "bbox": [ + 46, + 368, + 546, + 434 + ], + "type": "text", + "content": ". This coarse grid is then used to condition the fine transformer to generate a fine latent grid " + }, + { + "bbox": [ + 46, + 368, + 546, + 434 + ], + "type": "inline_equation", + "content": "\\mathbf{E}_{64}" + }, + { + "bbox": [ + 46, + 368, + 546, + 434 + ], + "type": "text", + "content": ". Finally, this fine latent grid is then passed through the Training Stage 1 " + }, + { + "bbox": [ + 46, + 368, + 546, + 434 + ], + "type": "inline_equation", + "content": "64^{3}" + }, + { + "bbox": [ + 46, + 368, + 546, + 434 + ], + "type": "text", + "content": " VQ-VAE decoder to generate the output shape." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 464, + 103, + 475 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 464, + 103, + 475 + ], + "spans": [ + { + "bbox": [ + 47, + 464, + 103, + 475 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 486, + 290, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 486, + 290, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 486, + 290, + 715 + ], + "type": "text", + "content": "CLIP-Sculptor aims to generate diverse and high-fidelity 3D shapes of common object categories from text prompts. Figure 2 illustrates the components of our CLIP-Sculptor model. We train the network using 3 stages. In the first stage, we use a vector-quantized representation [28], to efficiently represent 3D shapes in discrete voxel grids. This latent representation captures the diverse semantic meanings of the 3D shapes in common object categories and is compatible with powerful generative models. To balance computation cost and shape quality, we utilize latent representations at two resolutions. A low resolution latent representation " + }, + { + "bbox": [ + 46, + 486, + 290, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{E}_{32}" + }, + { + "bbox": [ + 46, + 486, + 290, + 715 + ], + "type": "text", + "content": " encodes geometric information of voxels " + }, + { + "bbox": [ + 46, + 486, + 290, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{V}_{32}" + }, + { + "bbox": [ + 46, + 486, + 290, + 715 + ], + "type": "text", + "content": " at resolutions of " + }, + { + "bbox": [ + 46, + 486, + 290, + 715 + ], + "type": "inline_equation", + "content": "32^{3}" + }, + { + "bbox": [ + 46, + 486, + 290, + 715 + ], + "type": "text", + "content": ", which is useful for text-conditional generation. A high resolution latent representation " + }, + { + "bbox": [ + 46, + 486, + 290, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{E}_{64}" + }, + { + "bbox": [ + 46, + 486, + 290, + 715 + ], + "type": "text", + "content": " encodes geometric information of voxel grids " + }, + { + "bbox": [ + 46, + 486, + 290, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{V}_{64}" + }, + { + "bbox": [ + 46, + 486, + 290, + 715 + ], + "type": "text", + "content": " at resolutions of " + }, + { + "bbox": [ + 46, + 486, + 290, + 715 + ], + "type": "inline_equation", + "content": "64^{3}" + }, + { + "bbox": [ + 46, + 486, + 290, + 715 + ], + "type": "text", + "content": ", which is useful for super-resolution and detail generation. In stage 2, we train a generative model conditioned on CLIP features obtained from images " + }, + { + "bbox": [ + 46, + 486, + 290, + 715 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{I}_r|r\\in R\\}" + }, + { + "bbox": [ + 46, + 486, + 290, + 715 + ], + "type": "text", + "content": " rendered at random views " + }, + { + "bbox": [ + 46, + 486, + 290, + 715 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 46, + 486, + 290, + 715 + ], + "type": "text", + "content": " to solve the lack of paired (text, shape)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 465, + 545, + 560 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 465, + 545, + 560 + ], + "spans": [ + { + "bbox": [ + 304, + 465, + 545, + 560 + ], + "type": "text", + "content": "data. The conditional generative model consists of a mapping net and a coarse transformer " + }, + { + "bbox": [ + 304, + 465, + 545, + 560 + ], + "type": "inline_equation", + "content": "T_{c}(.)" + }, + { + "bbox": [ + 304, + 465, + 545, + 560 + ], + "type": "text", + "content": " that is trained to unmask a masked low-resolution grid " + }, + { + "bbox": [ + 304, + 465, + 545, + 560 + ], + "type": "inline_equation", + "content": "\\mathbf{E}_{32}" + }, + { + "bbox": [ + 304, + 465, + 545, + 560 + ], + "type": "text", + "content": ". Finally in stage 3, a fine transformer " + }, + { + "bbox": [ + 304, + 465, + 545, + 560 + ], + "type": "inline_equation", + "content": "T_{f}(.)" + }, + { + "bbox": [ + 304, + 465, + 545, + 560 + ], + "type": "text", + "content": " is used as a super-resolution model in latent space to generate higher resolution shapes. It is trained in a similar manner to the coarse transformer but conditioned on a low-resolution grid " + }, + { + "bbox": [ + 304, + 465, + 545, + 560 + ], + "type": "inline_equation", + "content": "\\mathbf{E}_{32}" + }, + { + "bbox": [ + 304, + 465, + 545, + 560 + ], + "type": "text", + "content": " instead of CLIP features." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 565, + 545, + 673 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 565, + 545, + 673 + ], + "spans": [ + { + "bbox": [ + 304, + 565, + 545, + 673 + ], + "type": "text", + "content": "During inference, a text prompt is sequentially converted from CLIP text embedding to a low-resolution latent representation " + }, + { + "bbox": [ + 304, + 565, + 545, + 673 + ], + "type": "inline_equation", + "content": "\\mathbf{E}_{32}" + }, + { + "bbox": [ + 304, + 565, + 545, + 673 + ], + "type": "text", + "content": " to a high-resolution latent representation " + }, + { + "bbox": [ + 304, + 565, + 545, + 673 + ], + "type": "inline_equation", + "content": "\\mathbf{E}_{64}" + }, + { + "bbox": [ + 304, + 565, + 545, + 673 + ], + "type": "text", + "content": " to finally a high-resolution voxel grid " + }, + { + "bbox": [ + 304, + 565, + 545, + 673 + ], + "type": "inline_equation", + "content": "\\mathbf{V}_{64}" + }, + { + "bbox": [ + 304, + 565, + 545, + 673 + ], + "type": "text", + "content": ". A confidence-based iterative decoding scheme [4] is adopted for unmasking the fully masked initial grids for coarse and fine transformers at inference. To further control shape fidelity and diversity, we propose an annealing strategy for classifier-free guidance at inference time." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 677, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 677, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 677, + 546, + 713 + ], + "type": "text", + "content": "In the following subsections, we will introduce the major components within CLIP-Sculptor. Section 3.1 describes the vector quantized representations at multiple resolutions;" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18341" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 133 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 133 + ], + "type": "text", + "content": "Section 3.2 explains the conditional generation model as well as its design choices and training strategies; Section 3.3 introduces the super-resolution model; Section 3.4 explains how the classifier-free guidance is achieved at inference using an annealing strategy." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 138, + 228, + 152 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 138, + 228, + 152 + ], + "spans": [ + { + "bbox": [ + 47, + 138, + 228, + 152 + ], + "type": "text", + "content": "3.1. Multi-Resolution Voxel VQ-VAEs" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 157, + 287, + 276 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 157, + 287, + 276 + ], + "spans": [ + { + "bbox": [ + 46, + 157, + 287, + 276 + ], + "type": "text", + "content": "To learn a low-dimensional latent space that effectively represents the distribution of the training shape set " + }, + { + "bbox": [ + 46, + 157, + 287, + 276 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 46, + 157, + 287, + 276 + ], + "type": "text", + "content": ", we use a vector-quantized variational autoencoder (VQVAE) [28]. The vector quantized latent representation avoids \"posterior collapse\", and it is also amenable to modeling with transformer-based generative models. To balance computational cost and shape quality, we train two separate VQ-VAEs for " + }, + { + "bbox": [ + 46, + 157, + 287, + 276 + ], + "type": "inline_equation", + "content": "32^{3}" + }, + { + "bbox": [ + 46, + 157, + 287, + 276 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 157, + 287, + 276 + ], + "type": "inline_equation", + "content": "64^{3}" + }, + { + "bbox": [ + 46, + 157, + 287, + 276 + ], + "type": "text", + "content": " resolution voxel grids " + }, + { + "bbox": [ + 46, + 157, + 287, + 276 + ], + "type": "inline_equation", + "content": "\\mathbf{V}_{32}" + }, + { + "bbox": [ + 46, + 157, + 287, + 276 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 157, + 287, + 276 + ], + "type": "inline_equation", + "content": "\\mathbf{V}_{64}" + }, + { + "bbox": [ + 46, + 157, + 287, + 276 + ], + "type": "text", + "content": ", which are useful for conditional generation and superresolution respectively." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 277, + 287, + 397 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 277, + 287, + 397 + ], + "spans": [ + { + "bbox": [ + 46, + 277, + 287, + 397 + ], + "type": "text", + "content": "For both VQ-VAEs, we use ResNet-based [13] volumetric CNN encoders and decoders. A vector quantization layer maps the down-sampled voxel grids into a discrete latent space by indexing the continuous latent space with an embedding codebook. In this way, the voxel grids " + }, + { + "bbox": [ + 46, + 277, + 287, + 397 + ], + "type": "inline_equation", + "content": "\\mathbf{V}_{32}" + }, + { + "bbox": [ + 46, + 277, + 287, + 397 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 277, + 287, + 397 + ], + "type": "inline_equation", + "content": "\\mathbf{V}_{64}" + }, + { + "bbox": [ + 46, + 277, + 287, + 397 + ], + "type": "text", + "content": " are effectively represented by low-dimensional latent representations " + }, + { + "bbox": [ + 46, + 277, + 287, + 397 + ], + "type": "inline_equation", + "content": "\\mathbf{E}_{32}" + }, + { + "bbox": [ + 46, + 277, + 287, + 397 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 277, + 287, + 397 + ], + "type": "inline_equation", + "content": "\\mathbf{E}_{64}" + }, + { + "bbox": [ + 46, + 277, + 287, + 397 + ], + "type": "text", + "content": ", at the resolution of " + }, + { + "bbox": [ + 46, + 277, + 287, + 397 + ], + "type": "inline_equation", + "content": "4^{3}" + }, + { + "bbox": [ + 46, + 277, + 287, + 397 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 277, + 287, + 397 + ], + "type": "inline_equation", + "content": "8^{3}" + }, + { + "bbox": [ + 46, + 277, + 287, + 397 + ], + "type": "text", + "content": ". The VQ-VAEs are trained with mean squared error loss (MSE), commitment loss [28], and an exponential moving average strategy [33, 45]." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 403, + 256, + 415 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 403, + 256, + 415 + ], + "spans": [ + { + "bbox": [ + 47, + 403, + 256, + 415 + ], + "type": "text", + "content": "3.2. CLIP-Conditioned Coarse Transformer" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 422, + 287, + 553 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 422, + 287, + 553 + ], + "spans": [ + { + "bbox": [ + 46, + 422, + 287, + 553 + ], + "type": "text", + "content": "The conditional generation process aims to generate diverse shapes that semantically correspond to the text prompts. Inspired by the recent success of masking approaches [4] and diffusion models [12, 27, 36], we formulate this generation process as a conditional unmasking task. Given an input masked latent representation " + }, + { + "bbox": [ + 46, + 422, + 287, + 553 + ], + "type": "inline_equation", + "content": "\\mathbf{E}_{32}^{msk}" + }, + { + "bbox": [ + 46, + 422, + 287, + 553 + ], + "type": "text", + "content": " and a conditional vector " + }, + { + "bbox": [ + 46, + 422, + 287, + 553 + ], + "type": "inline_equation", + "content": "\\mathbf{c}" + }, + { + "bbox": [ + 46, + 422, + 287, + 553 + ], + "type": "text", + "content": ", the coarse transformer " + }, + { + "bbox": [ + 46, + 422, + 287, + 553 + ], + "type": "inline_equation", + "content": "T_{c}(.)" + }, + { + "bbox": [ + 46, + 422, + 287, + 553 + ], + "type": "text", + "content": " should produce an fully unmasked latent representation " + }, + { + "bbox": [ + 46, + 422, + 287, + 553 + ], + "type": "inline_equation", + "content": "\\mathbf{E}_{32}" + }, + { + "bbox": [ + 46, + 422, + 287, + 553 + ], + "type": "text", + "content": " that not only aligns with the semantic meaning captured by the conditional vector, but also allows for the sampling of diverse shapes at inference." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 554, + 287, + 661 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 554, + 287, + 661 + ], + "spans": [ + { + "bbox": [ + 46, + 554, + 287, + 661 + ], + "type": "text", + "content": "The conditional vector " + }, + { + "bbox": [ + 46, + 554, + 287, + 661 + ], + "type": "inline_equation", + "content": "\\mathbf{c}" + }, + { + "bbox": [ + 46, + 554, + 287, + 661 + ], + "type": "text", + "content": " provides the semantic guidance for the shape generation process. Since we assume no text data at the training time, we leverage the text-image joint-embedding of CLIP. The generation is guided by the CLIP image embedding from the rendered images " + }, + { + "bbox": [ + 46, + 554, + 287, + 661 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{I}_r\\}" + }, + { + "bbox": [ + 46, + 554, + 287, + 661 + ], + "type": "text", + "content": " during training and guided by the CLIP text embedding from the text prompt during inference. We add Gaussian noise to the CLIP image embeddings [44] to alleviate the modality gap between the CLIP text and image embeddings [20]:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 102, + 668, + 287, + 694 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 102, + 668, + 287, + 694 + ], + "spans": [ + { + "bbox": [ + 102, + 668, + 287, + 694 + ], + "type": "interline_equation", + "content": "\\hat {\\mathbf {c}} = f _ {I} \\left(\\mathbf {I} _ {r}\\right) + \\frac {\\gamma \\cdot \\epsilon \\cdot \\| f _ {I} \\left(\\mathbf {I} _ {r}\\right) \\| _ {2}}{\\| \\epsilon \\| _ {2}}, \\tag {1}", + "image_path": "5414e79a658754d5e3f0f407c794e2f35912099967008b7ef32f690556a3016e.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 701, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 701, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 701, + 287, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 701, + 287, + 714 + ], + "type": "inline_equation", + "content": "f_{I}(\\cdot)" + }, + { + "bbox": [ + 47, + 701, + 287, + 714 + ], + "type": "text", + "content": " is the CLIP image encoder, " + }, + { + "bbox": [ + 47, + 701, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\epsilon \\sim \\mathcal{N}(0,1)" + }, + { + "bbox": [ + 47, + 701, + 287, + 714 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 47, + 701, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\gamma" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 72, + 545, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 203 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 203 + ], + "type": "text", + "content": "controls the level of perturbation. The conditional vector " + }, + { + "bbox": [ + 304, + 72, + 545, + 203 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{c}}" + }, + { + "bbox": [ + 304, + 72, + 545, + 203 + ], + "type": "text", + "content": " is then normalized and passed through a mapping network that consists of multi-layer perceptrons (MLP). The mapped vector " + }, + { + "bbox": [ + 304, + 72, + 545, + 203 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{c}}" + }, + { + "bbox": [ + 304, + 72, + 545, + 203 + ], + "type": "text", + "content": " conditions on the coarse transformer " + }, + { + "bbox": [ + 304, + 72, + 545, + 203 + ], + "type": "inline_equation", + "content": "T_{c}(.)" + }, + { + "bbox": [ + 304, + 72, + 545, + 203 + ], + "type": "text", + "content": " by predicting the affine transform parameters of each transformer block's layernorm. In order to control the diversity-quality balance of the generated shapes, we replace the conditional vector with the null embedding by a " + }, + { + "bbox": [ + 304, + 72, + 545, + 203 + ], + "type": "inline_equation", + "content": "\\rho\\%" + }, + { + "bbox": [ + 304, + 72, + 545, + 203 + ], + "type": "text", + "content": " drop out rate for classifier-free guidance [16]. These are important design choices that affects the diversity and quality of the generated shapes at inference time." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 204, + 545, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 204, + 545, + 384 + ], + "spans": [ + { + "bbox": [ + 304, + 204, + 545, + 384 + ], + "type": "text", + "content": "To train the coarse transformer " + }, + { + "bbox": [ + 304, + 204, + 545, + 384 + ], + "type": "inline_equation", + "content": "T_{c}(.)" + }, + { + "bbox": [ + 304, + 204, + 545, + 384 + ], + "type": "text", + "content": ", we randomly replace the input indices with a special \"mask\" tokens by uniformly selecting a masking ratio between 0-100% [4]. Given the masked latent grid " + }, + { + "bbox": [ + 304, + 204, + 545, + 384 + ], + "type": "inline_equation", + "content": "\\mathbf{E}_{32}^{msk}" + }, + { + "bbox": [ + 304, + 204, + 545, + 384 + ], + "type": "text", + "content": " and the conditional vector " + }, + { + "bbox": [ + 304, + 204, + 545, + 384 + ], + "type": "inline_equation", + "content": "\\mathbf{c}" + }, + { + "bbox": [ + 304, + 204, + 545, + 384 + ], + "type": "text", + "content": ", the training objective is to maximize the following log-likelihood: " + }, + { + "bbox": [ + 304, + 204, + 545, + 384 + ], + "type": "inline_equation", + "content": "\\sum_{n}^{N}\\log p(\\mathbf{E}_{32,n}|\\mathbf{E}_{32,n}^{msk},\\mathbf{c}) / N" + }, + { + "bbox": [ + 304, + 204, + 545, + 384 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 204, + 545, + 384 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 204, + 545, + 384 + ], + "type": "text", + "content": " is the training data size and " + }, + { + "bbox": [ + 304, + 204, + 545, + 384 + ], + "type": "inline_equation", + "content": "\\mathbf{E}_{32}^{msk}" + }, + { + "bbox": [ + 304, + 204, + 545, + 384 + ], + "type": "text", + "content": " is the randomly masked latent grids. As the network sees only the ground truth tokens but not its own predicted samples during training, there exists an autoregressive drift resulted by the accumulative error at inference-time sampling. To alleviate this issue, we take inspiration from the NLP literature [38] and use a similar step-unrolled training (SUT) that predicts the entire un-masked grids at each time step. The final objective is to maximize the following formulation:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 307, + 392, + 545, + 434 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 392, + 545, + 434 + ], + "spans": [ + { + "bbox": [ + 307, + 392, + 545, + 434 + ], + "type": "interline_equation", + "content": "\\sum_ {n} ^ {N} \\left(\\log p \\left(\\mathbf {E} _ {3 2, n} \\mid \\mathbf {E} _ {3 2, n} ^ {m s k}, \\mathbf {c}\\right) + \\log p \\left(\\mathbf {E} _ {3 2, n} \\mid \\tilde {\\mathbf {E}} _ {3 2, n}, \\mathbf {c}\\right)\\right) / N, \\tag {2}", + "image_path": "c5302208cffa994711554949381496101377c1d15b716e03cf436215fe56fb1f.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 434, + 544, + 447 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 434, + 544, + 447 + ], + "spans": [ + { + "bbox": [ + 306, + 434, + 544, + 447 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 306, + 434, + 544, + 447 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{E}}_{32}" + }, + { + "bbox": [ + 306, + 434, + 544, + 447 + ], + "type": "text", + "content": " is the coarse transformer " + }, + { + "bbox": [ + 306, + 434, + 544, + 447 + ], + "type": "inline_equation", + "content": "T_{c}(.)" + }, + { + "bbox": [ + 306, + 434, + 544, + 447 + ], + "type": "text", + "content": " 's own prediction." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 454, + 519, + 467 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 454, + 519, + 467 + ], + "spans": [ + { + "bbox": [ + 306, + 454, + 519, + 467 + ], + "type": "text", + "content": "3.3. Super-Resolution with Fine Transformer" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 473, + 545, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 473, + 545, + 605 + ], + "spans": [ + { + "bbox": [ + 304, + 473, + 545, + 605 + ], + "type": "text", + "content": "Super-resolution in the discrete latent space improves quality and upsamples the resolution of generated shapes. This process is done with a fine transformer " + }, + { + "bbox": [ + 304, + 473, + 545, + 605 + ], + "type": "inline_equation", + "content": "T_{f}(.)" + }, + { + "bbox": [ + 304, + 473, + 545, + 605 + ], + "type": "text", + "content": ", which takes a low-resolution latent representation " + }, + { + "bbox": [ + 304, + 473, + 545, + 605 + ], + "type": "inline_equation", + "content": "\\mathbf{E}_{32}" + }, + { + "bbox": [ + 304, + 473, + 545, + 605 + ], + "type": "text", + "content": " and a masked high-resolution latent grid as input. The fine transformer is conditioned on the coarse grid " + }, + { + "bbox": [ + 304, + 473, + 545, + 605 + ], + "type": "inline_equation", + "content": "\\mathbf{E}_{32}" + }, + { + "bbox": [ + 304, + 473, + 545, + 605 + ], + "type": "text", + "content": " via cross attention. It unmasks the high-resolution representation " + }, + { + "bbox": [ + 304, + 473, + 545, + 605 + ], + "type": "inline_equation", + "content": "\\mathbf{E}_{64}" + }, + { + "bbox": [ + 304, + 473, + 545, + 605 + ], + "type": "text", + "content": ". To ensure that training data resembles the inference data, we directly use the predictions " + }, + { + "bbox": [ + 304, + 473, + 545, + 605 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{E}}_{32}" + }, + { + "bbox": [ + 304, + 473, + 545, + 605 + ], + "type": "text", + "content": " of the coarse transformer " + }, + { + "bbox": [ + 304, + 473, + 545, + 605 + ], + "type": "inline_equation", + "content": "T_{c}(.)" + }, + { + "bbox": [ + 304, + 473, + 545, + 605 + ], + "type": "text", + "content": " to train the fine transformer " + }, + { + "bbox": [ + 304, + 473, + 545, + 605 + ], + "type": "inline_equation", + "content": "T_{f}(.)" + }, + { + "bbox": [ + 304, + 473, + 545, + 605 + ], + "type": "text", + "content": ", instead of using the ground truth low-resolution latent representation " + }, + { + "bbox": [ + 304, + 473, + 545, + 605 + ], + "type": "inline_equation", + "content": "\\mathbf{E}_{32}" + }, + { + "bbox": [ + 304, + 473, + 545, + 605 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 611, + 492, + 623 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 611, + 492, + 623 + ], + "spans": [ + { + "bbox": [ + 306, + 611, + 492, + 623 + ], + "type": "text", + "content": "3.4. Annealed Classifier-Free Guidance" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "content": "At inference time, we first convert an input text prompt into a CLIP text embedding using the CLIP text encoder " + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "inline_equation", + "content": "f_{T}(\\cdot)" + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "content": ". The text embedding is then fed into the mapping network and used as the conditional vector " + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{c}}" + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "content": " of the coarse transformer " + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "inline_equation", + "content": "T_{c}(\\cdot)" + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "content": ". We use the iterative decoding scheme [4] to slowly unmask the grid over a sequence of " + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "content": " steps. The initial input to the coarse transformer " + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "inline_equation", + "content": "T_{c}(\\cdot)" + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "content": " is a completely" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18342" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 191 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 191 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 191 + ], + "type": "text", + "content": "masked latent grid. At each time step " + }, + { + "bbox": [ + 46, + 72, + 289, + 191 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 72, + 289, + 191 + ], + "type": "text", + "content": ", we condition the coarse transformer " + }, + { + "bbox": [ + 46, + 72, + 289, + 191 + ], + "type": "inline_equation", + "content": "T_{c}(.)" + }, + { + "bbox": [ + 46, + 72, + 289, + 191 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 46, + 72, + 289, + 191 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{c}}" + }, + { + "bbox": [ + 46, + 72, + 289, + 191 + ], + "type": "text", + "content": " and take the output latent grid " + }, + { + "bbox": [ + 46, + 72, + 289, + 191 + ], + "type": "inline_equation", + "content": "\\mathbf{E}_{32,t-1}^{msk}" + }, + { + "bbox": [ + 46, + 72, + 289, + 191 + ], + "type": "text", + "content": " from the prior time step as the input. The coarse transformer " + }, + { + "bbox": [ + 46, + 72, + 289, + 191 + ], + "type": "inline_equation", + "content": "T_{c}(.)" + }, + { + "bbox": [ + 46, + 72, + 289, + 191 + ], + "type": "text", + "content": " will predict an unmasked latent grid " + }, + { + "bbox": [ + 46, + 72, + 289, + 191 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{E}}_{32,t}" + }, + { + "bbox": [ + 46, + 72, + 289, + 191 + ], + "type": "text", + "content": ". We mask all the tokens except for the most confident token predictions and the tokens unmasked from previous steps. This newly masked latent grid " + }, + { + "bbox": [ + 46, + 72, + 289, + 191 + ], + "type": "inline_equation", + "content": "\\mathbf{E}_{32,t}^{msk}" + }, + { + "bbox": [ + 46, + 72, + 289, + 191 + ], + "type": "text", + "content": " is the input to the next time step " + }, + { + "bbox": [ + 46, + 72, + 289, + 191 + ], + "type": "inline_equation", + "content": "t+1" + }, + { + "bbox": [ + 46, + 72, + 289, + 191 + ], + "type": "text", + "content": ". We repeat this process until the process unmasks all the tokens, which is ensured by a cosine masking schedule [4]." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 193, + 289, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 193, + 289, + 373 + ], + "spans": [ + { + "bbox": [ + 46, + 193, + 289, + 373 + ], + "type": "text", + "content": "During this iterative decoding scheme, we apply a new variant of classifier-free guidance [16]. Classifier-free guidance extrapolates an unconditional sample in the direction of a conditional sample. The extrapolation is controlled by an adaptive guidance scale parameter. Varying this parameter results in a trade-off between accuracy of text-to-shape generation and sample diversity. Instead of using a constant guidance scale for all time steps [16], we propose to vary the guidance scale over time according to an annealing schedule. The intuition is that during the initial steps of sampling when the input to the coarse transformer is mostly masked indices, a larger guidance scale is important to keep the network \"on task\"; later on in the sampling process, a lower guidance scale can help produce more sample diversity. Overall, the guidance is formulated as:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 85, + 384, + 287, + 404 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 85, + 384, + 287, + 404 + ], + "spans": [ + { + "bbox": [ + 85, + 384, + 287, + 404 + ], + "type": "interline_equation", + "content": "\\hat {P} _ {t} (c) = P _ {t} (0) + a (t) \\left(P _ {t} (c) - P _ {t} (0)\\right), \\tag {3}", + "image_path": "2bab086b8ed73ab139f989b9ecb02247535aea304c5d682221f64c7520dfa921.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 412, + 287, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 412, + 287, + 521 + ], + "spans": [ + { + "bbox": [ + 46, + 412, + 287, + 521 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 412, + 287, + 521 + ], + "type": "inline_equation", + "content": "a(t)" + }, + { + "bbox": [ + 46, + 412, + 287, + 521 + ], + "type": "text", + "content": " is the guidance scale annealing schedule, " + }, + { + "bbox": [ + 46, + 412, + 287, + 521 + ], + "type": "inline_equation", + "content": "P_{t}(c)" + }, + { + "bbox": [ + 46, + 412, + 287, + 521 + ], + "type": "text", + "content": " denotes the conditional distribution " + }, + { + "bbox": [ + 46, + 412, + 287, + 521 + ], + "type": "inline_equation", + "content": "p(\\mathbf{E}_{32,t}|\\mathbf{E}_{32,t - 1}^{msk},\\mathbf{c})" + }, + { + "bbox": [ + 46, + 412, + 287, + 521 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 412, + 287, + 521 + ], + "type": "inline_equation", + "content": "P_{t}(0)" + }, + { + "bbox": [ + 46, + 412, + 287, + 521 + ], + "type": "text", + "content": " denotes " + }, + { + "bbox": [ + 46, + 412, + 287, + 521 + ], + "type": "inline_equation", + "content": "p(\\mathbf{E}_{32,t}|\\mathbf{E}_{32,t - 1}^{msk},\\mathbf{0})" + }, + { + "bbox": [ + 46, + 412, + 287, + 521 + ], + "type": "text", + "content": ". Note that " + }, + { + "bbox": [ + 46, + 412, + 287, + 521 + ], + "type": "inline_equation", + "content": "a(t)" + }, + { + "bbox": [ + 46, + 412, + 287, + 521 + ], + "type": "text", + "content": " is continuous and usually monotonically decreasing. By setting " + }, + { + "bbox": [ + 46, + 412, + 287, + 521 + ], + "type": "inline_equation", + "content": "a(t) = k" + }, + { + "bbox": [ + 46, + 412, + 287, + 521 + ], + "type": "text", + "content": " for some constant " + }, + { + "bbox": [ + 46, + 412, + 287, + 521 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 412, + 287, + 521 + ], + "type": "text", + "content": ", this equation is equal to the classifier-free guidance scheme with a constant guidance scale. In this paper, we experimentally evaluate other annealing schedules [4], including linear, cosine, and square root functions." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 522, + 287, + 583 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 522, + 287, + 583 + ], + "spans": [ + { + "bbox": [ + 46, + 522, + 287, + 583 + ], + "type": "text", + "content": "Finally, the unmasked coarse latent grid " + }, + { + "bbox": [ + 46, + 522, + 287, + 583 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{E}}_{32}" + }, + { + "bbox": [ + 46, + 522, + 287, + 583 + ], + "type": "text", + "content": " is fed to the fine transformer " + }, + { + "bbox": [ + 46, + 522, + 287, + 583 + ], + "type": "inline_equation", + "content": "T_{f}" + }, + { + "bbox": [ + 46, + 522, + 287, + 583 + ], + "type": "text", + "content": ", which uses the iterative decoding scheme to unmask an initially masked fine latent grid " + }, + { + "bbox": [ + 46, + 522, + 287, + 583 + ], + "type": "inline_equation", + "content": "\\mathbf{E}_{64}^{msk}" + }, + { + "bbox": [ + 46, + 522, + 287, + 583 + ], + "type": "text", + "content": ". The unmasked fine grid is then passed to the " + }, + { + "bbox": [ + 46, + 522, + 287, + 583 + ], + "type": "inline_equation", + "content": "64^3" + }, + { + "bbox": [ + 46, + 522, + 287, + 583 + ], + "type": "text", + "content": " VQ-VAE decoder to obtain the final high-resolution voxel shape." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 597, + 128, + 609 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 597, + 128, + 609 + ], + "spans": [ + { + "bbox": [ + 47, + 597, + 128, + 609 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 618, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 618, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 618, + 287, + 715 + ], + "type": "text", + "content": "In this section, we report experimental results to evaluate the generation fidelity, diversity, and class accuracy of CLIP-Sculptor. We provide full details of the experiment setup in the supplementary. We run the experiments 3 times for each of the aforementioned measures and report the mean in each case, except for the final comparisons with baselines (Table 2), where we use the best seed. Additional results can also be found in the supplementary." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 72, + 545, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 167 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 167 + ], + "type": "text", + "content": "Dataset. We conduct our experiments on two subsets of the ShapeNet(v2) dataset [3]. The first subset, ShapeNet13, contains 13 categories from ShapeNet as used in [7,24]. We use the same train/test split as specified in [24]. Our second subset is ShapeNet55 which contains all 55 ShapeNet categories. For ShapeNet55, we render images as described in [7]; the training dataset contains 51,784 datapoints, whereas the test set contains 6,101 datapoints." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 168, + 545, + 263 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 168, + 545, + 263 + ], + "spans": [ + { + "bbox": [ + 304, + 168, + 545, + 263 + ], + "type": "text", + "content": "Evaluation Metrics. To assess the generative capabilities of CLIP-Sculptor, we use Fréchet Inception Distance (FID) ([14]) to evaluate shape diversity and quality, and use Classifier Accuracy (Acc) to evaluate the text-shape correspondence. These metrics follow [37] where they first generated single mean shapes for 234 predetermined text queries and then passed all shapes through a classifier to measure Acc. The latent space of this classifier is also used for FID." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 265, + 546, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 265, + 546, + 517 + ], + "spans": [ + { + "bbox": [ + 304, + 265, + 546, + 517 + ], + "type": "text", + "content": "Baseline. We compare the performance of our method against CLIP-Forge [37], CLIP-Mesh [19] and DreamFields [18], which are currently state of the art for zero-shot text-to-shape generation. We also set up a comparison method called Zero-Shot AutoSDF(ZS-ASDF), which trains the supervised method AutoSDF [26] with the CLIP features as CLIP-Sculptor, to highlight the superiority of our conditional generation model. CLIP-Forge reports results on single shape generation for 234 text queries. These single shapes were generated using the mean of the prior (the Gaussian distribution), so we refer to these results as CF-MS. Note that this does not capture the diversity of shapes that can be generated given a text query. To capture results on more shape generations, we sample 32 shapes instead of one using a Gaussian (CF-G), truncated Gaussian (CF-TG) or clipped Gaussian (CF-CG) distribution. We also generate 32 shapes for ZS-ASDF and our method for fair comparison. We only compare qualitatively with DreamFields and CLIP-Mesh as the optimization time per query is significant (Table 1), and it does not use prior knowledge from the shape dataset." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 526, + 523, + 540 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 526, + 523, + 540 + ], + "spans": [ + { + "bbox": [ + 305, + 526, + 523, + 540 + ], + "type": "text", + "content": "4.1. Evaluating Shape Diversity and Accuracy" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 545, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 545, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 545, + 545, + 713 + ], + "type": "text", + "content": "In this section, we first quantitatively evaluate the diversity and accuracy of a given shape matching a given text query on the ShapeNet13 dataset. We compare with CLIP-Forge and ZS-ASDF using the Acc and FID metrics. The results are shown in Table 2. The first four columns represent CLIP-Forge(CF) and its different sampling techniques. The fifth column represents the result from AutoSDF trained with CLIP features(ZS-ASDF). The other columns represent our method with different annealing scale strategies. Three major things can be observed. First, all variants of our method outperform CLIP-Forge and ZS-ASDF significantly. This indicates that the method produces more diverse and higher fidelity shapes of increasing accuracy. Second, it can be seen that annealing strategies" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18343" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 68, + 561, + 566 + ], + "blocks": [ + { + "bbox": [ + 50, + 68, + 561, + 566 + ], + "lines": [ + { + "bbox": [ + 50, + 68, + 561, + 566 + ], + "spans": [ + { + "bbox": [ + 50, + 68, + 561, + 566 + ], + "type": "image", + "image_path": "4d2b11d33bf405305c5d336a25d1e06a4cdb9aa618c6931d990ba514a23a45c7.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 574, + 547, + 620 + ], + "lines": [ + { + "bbox": [ + 46, + 574, + 547, + 620 + ], + "spans": [ + { + "bbox": [ + 46, + 574, + 547, + 620 + ], + "type": "text", + "content": "Figure 3. Qualitative comparison of CLIP-Sculptor (rendered in purple), CLIP-Forge [37](rendered in green), DreamField [18], and CLIP-Mesh [19] on text-conditioned generation. For each text prompt, the first generated shape by CLIP-Sculptor and CLIP-Forge [37] is zoomed in for detail comparison. Comparing with other methods, CLIP-Sculptor generates a broader category of shapes that correspond with the text prompt. The generated shapes are more diverse and has higher fidelity." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 639, + 289, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 639, + 289, + 712 + ], + "spans": [ + { + "bbox": [ + 46, + 639, + 289, + 712 + ], + "type": "text", + "content": "for guidance scale give a better accuracy-diversity trade-off than constant scale guidance. Finally, we note that simply augmenting AutoSDF (another VQ-VAE) with CLIP features does not lead to improved text to shape generation, indicating the importance of our architecture design decisions." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 304, + 639, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 639, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 639, + 547, + 713 + ], + "type": "text", + "content": "We also qualitatively compare our method to CLIP-Forge, CLIP-Mesh, and DreamFields. The results are shown in Figure 3. It can be observed that our method generates higher quality shapes (view \"a machine gun\"), with more detail (view \"an office chair\") and higher diversity (view \"a jet\"). CLIP-Forge usually produces the same shape" + } + ] + } + ], + "index": 3 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "18344" + } + ] + } + ], + "index": 4 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 48, + 122, + 547, + 174 + ], + "blocks": [ + { + "bbox": [ + 46, + 70, + 547, + 105 + ], + "lines": [ + { + "bbox": [ + 46, + 70, + 547, + 105 + ], + "spans": [ + { + "bbox": [ + 46, + 70, + 547, + 105 + ], + "type": "text", + "content": "Table 2. Comparisons of CLIP-Forge(CF) baseline (across different sampling strategies), AutoSDF augmented with CLIP (ZS-ASDF) with CLIP-Sculptor (CS) on Accuracy(ACC) and FID. CLIP-Sculptor outperforms other methods in both metrics, which proves it generates shapes with higher fidelity and better diversity." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 48, + 122, + 547, + 174 + ], + "lines": [ + { + "bbox": [ + 48, + 122, + 547, + 174 + ], + "spans": [ + { + "bbox": [ + 48, + 122, + 547, + 174 + ], + "type": "table", + "html": "
MethodCF-MSCF-GCF-TGCF-CGZS-ASDFCS-const.CS-sqrtCS-linearCS-cosine
FID ↓2425.252233.482141.612100.677332.931821.781480.111629.511725.63
ACC ↑83.3362.8168.7171.1139.1486.5987.0887.5087.27
", + "image_path": "8791f35eb4d219269e9aa37cbb20db618588cb99f7ca517b68cd3966ae76930d.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 53, + 266, + 293, + 437 + ], + "blocks": [ + { + "bbox": [ + 46, + 200, + 288, + 245 + ], + "lines": [ + { + "bbox": [ + 46, + 200, + 288, + 245 + ], + "spans": [ + { + "bbox": [ + 46, + 200, + 288, + 245 + ], + "type": "text", + "content": "Table 3. Ablations on the major components of CLIP-Sculptor. CLIP-Sculptor achieves the best performance comparing with other design choice, which proves the effectiveness of the major components." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 105, + 254, + 240, + 264 + ], + "lines": [ + { + "bbox": [ + 105, + 254, + 240, + 264 + ], + "spans": [ + { + "bbox": [ + 105, + 254, + 240, + 264 + ], + "type": "text", + "content": "(a) Effect of varying the Noise parameter." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 53, + 266, + 293, + 437 + ], + "lines": [ + { + "bbox": [ + 53, + 266, + 293, + 437 + ], + "spans": [ + { + "bbox": [ + 53, + 266, + 293, + 437 + ], + "type": "table", + "html": "
γ×0.50.81.01.21.5
FID↓1720.021764.981484.611703.381447.911478.17
ACC↑64.8775.7377.4179.0979.6378.47
(b) Effect of varying number of layers (L) in the mapping network.
L012345
FID↓2874.871716.731518.971447.911532.501424.46
ACC↑62.7278.7079.1779.6377.1676.09
(c) Comparison with baselines on super resolution.
Method3D-UNet643-DSCLIP-Sculptor
FID↓2056.922196.961910.28
ACC↑86.6577.9286.85
", + "image_path": "01582681405a9c0c8b21c4494d407f37e9a4e7b6e501cc6564226346267ef213.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 457, + 287, + 554 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 457, + 287, + 554 + ], + "spans": [ + { + "bbox": [ + 46, + 457, + 287, + 554 + ], + "type": "text", + "content": "with small variations (view \"a rectangular table\"). We also note that DreamFields and CLIP-Mesh produce very abstract results which may not be useful in many applications. Finally, we also show results on ShapeNet55 in the last row of Figure 3. We could not get CLIP-Forge to produce sensible shapes for most text queries on ShapeNet55 (which we attribute to the data imbalance issue of ShapeNet55), whereas our method produced high fidelity shapes." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 562, + 287, + 576 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 562, + 287, + 576 + ], + "spans": [ + { + "bbox": [ + 47, + 562, + 287, + 576 + ], + "type": "text", + "content": "4.2. Major Components for Conditional Generation" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 582, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 582, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 582, + 288, + 715 + ], + "type": "text", + "content": "Effect of Noise Parameter. During Stage 2 Training, to better align the image and text embeddings, we added Gaussian noise at varying levels of " + }, + { + "bbox": [ + 46, + 582, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 46, + 582, + 288, + 715 + ], + "type": "text", + "content": ". We investigate the effect of " + }, + { + "bbox": [ + 46, + 582, + 288, + 715 + ], + "type": "inline_equation", + "content": "(\\gamma)" + }, + { + "bbox": [ + 46, + 582, + 288, + 715 + ], + "type": "text", + "content": " added to the image condition vector and show the results in Table 3 (a). We keep the number of mapping layers fixed at three in this experiment. We observe that adding Gaussian noise drastically improves both the diversity and accuracy of generation, with the optimal noise parameter being around 1.2. These results indicate that adding noise during training helps align the text features seen at inference and the image features seen at training." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 201, + 545, + 309 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 201, + 545, + 309 + ], + "spans": [ + { + "bbox": [ + 304, + 201, + 545, + 309 + ], + "type": "text", + "content": "Size of Mapping Network. We next probe the importance of the mapping network. We show the results in Table 3 (b), where " + }, + { + "bbox": [ + 304, + 201, + 545, + 309 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 304, + 201, + 545, + 309 + ], + "type": "text", + "content": " represents the number of layers of the mapping function. In the case of 0 layers, we directly project conditional embeddings to the layernorm parameters using a linear layer. We make two observations from the results: 1) A common mapping network improves both the accuracy and FID. 2) Increasing the number of mapping network layers beyond a certain number decreases accuracy." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 310, + 546, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 310, + 546, + 430 + ], + "spans": [ + { + "bbox": [ + 304, + 310, + 546, + 430 + ], + "type": "text", + "content": "Classifier-Free Guidance. We next explore the relationship between dropping out image conditioning at " + }, + { + "bbox": [ + 304, + 310, + 546, + 430 + ], + "type": "inline_equation", + "content": "\\rho \\%" + }, + { + "bbox": [ + 304, + 310, + 546, + 430 + ], + "type": "text", + "content": " during training and using constant classifier-free guidance during test time. In Table 4, we vary the scale parameter, " + }, + { + "bbox": [ + 304, + 310, + 546, + 430 + ], + "type": "inline_equation", + "content": "a(t) = k" + }, + { + "bbox": [ + 304, + 310, + 546, + 430 + ], + "type": "text", + "content": ", across the columns. It can be seen that as we increase the scale, accuracy typically increases while FID decreases. This indicates that the method is giving more accurate results on a given text query at the cost of shape diversity. A low dropout rate (5-15%) gives a good trade-off between accuracy and diversity." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 431, + 547, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 431, + 547, + 515 + ], + "spans": [ + { + "bbox": [ + 304, + 431, + 547, + 515 + ], + "type": "text", + "content": "Step-Unrolled Training (SUT). Finally, to further improve the accuracy we investigate the use of step-unrolled training [39]. The results are shown in the last row of Table 4. From the table it can be observed that step-unrolled training enables higher accuracy across all scales. This indicates that the model learns to unmask samples it would encounter during inference time." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 526, + 515, + 540 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 526, + 515, + 540 + ], + "spans": [ + { + "bbox": [ + 305, + 526, + 515, + 540 + ], + "type": "text", + "content": "4.3. Annealing Strategy for Scale Parameter" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 546, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 546, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 546, + 547, + 713 + ], + "type": "text", + "content": "An important idea we propose in this paper is a scale annealing technique for classifier-free guidance. We employ classifier-free guidance to identify a better accuracy-diversity trade-off with different annealing schedules: constant, linear, cosine and square root. To determine which annealing technique works the best, we fix the FID values and plotted the accuracy at each fixed FID value (Figure 4). As different scale parameters give different FID values, we conducted an extensive grid search over the starting scale parameter. Note that finding the exact FID is not always feasible, so we pick the closest FID. Figure 4 shows results of FID versus accuracy on three different runs of the Coarse Transformer. We find that across all three runs, the accuracy is typically lower for a given FID for constant schedules as" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18345" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 47, + 102, + 550, + 190 + ], + "blocks": [ + { + "bbox": [ + 47, + 70, + 545, + 94 + ], + "lines": [ + { + "bbox": [ + 47, + 70, + 545, + 94 + ], + "spans": [ + { + "bbox": [ + 47, + 70, + 545, + 94 + ], + "type": "text", + "content": "Table 4. Classifier-Free Guidance and Step-Unrolled Training (SUT) experiment results. " + }, + { + "bbox": [ + 47, + 70, + 545, + 94 + ], + "type": "inline_equation", + "content": "\\rho \\%" + }, + { + "bbox": [ + 47, + 70, + 545, + 94 + ], + "type": "text", + "content": " is the drop out rate for classifier-free guidance. SUT indicates the use of Step-Unrolled Training. The remaining columns indicate the variation in scale parameter." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 47, + 102, + 550, + 190 + ], + "lines": [ + { + "bbox": [ + 47, + 102, + 550, + 190 + ], + "spans": [ + { + "bbox": [ + 47, + 102, + 550, + 190 + ], + "type": "table", + "html": "
ρ%SUTa(t) = 3a(t) = 2.5a(t) = 2a(t) = 1.5a(t) = 1
FID↓Acc↑FID↓Acc↑FID↓Acc↑FID↓Acc↑FID↓Acc↑
5×2059.085.731970.185.471790.985.431536.583.981227.579.17
10×1893.284.461821.284.501684.383.411522.482.341348.477.76
15×2086.985.031964.784.991851.984.361660.983.141485.578.19
20×2062.583.391972.382.951892.982.741733.381.131566.675.94
5✓2039.887.692011.187.391811.887.401678.686.241517.982.18
", + "image_path": "9a6179151f3b0838fce1ae66d3985d6efa9254938e7e27275854c9776766b124.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 53, + 201, + 223, + 314 + ], + "blocks": [ + { + "bbox": [ + 53, + 201, + 223, + 314 + ], + "lines": [ + { + "bbox": [ + 53, + 201, + 223, + 314 + ], + "spans": [ + { + "bbox": [ + 53, + 201, + 223, + 314 + ], + "type": "image", + "image_path": "0cd7ff696dfc73d7523e291fc7cf6e556d36b6e5463c94b1fde9453ba86509ce.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 325, + 546, + 348 + ], + "lines": [ + { + "bbox": [ + 46, + 325, + 546, + 348 + ], + "spans": [ + { + "bbox": [ + 46, + 325, + 546, + 348 + ], + "type": "text", + "content": "Figure 4. FID and Acc results with different classifier-free guidance scale annealing strategies (constant, cosine, square root, linear) across three different runs (different seeds) of the CLIP-Sculptor Stage 2 Coarse Transformer." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 227, + 201, + 388, + 314 + ], + "blocks": [ + { + "bbox": [ + 227, + 201, + 388, + 314 + ], + "lines": [ + { + "bbox": [ + 227, + 201, + 388, + 314 + ], + "spans": [ + { + "bbox": [ + 227, + 201, + 388, + 314 + ], + "type": "image", + "image_path": "b761d6b05adefe8d1aa43a87eea9f07b6b2b79c107a38e92ae487cdbdeabaad9.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 391, + 202, + 545, + 314 + ], + "blocks": [ + { + "bbox": [ + 391, + 202, + 545, + 314 + ], + "lines": [ + { + "bbox": [ + 391, + 202, + 545, + 314 + ], + "spans": [ + { + "bbox": [ + 391, + 202, + 545, + 314 + ], + "type": "image", + "image_path": "b567cbab04f84aedc64cb29439a8d1f655ece0eb4525882548c2d908d9c54e6b.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 369, + 287, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 369, + 287, + 430 + ], + "spans": [ + { + "bbox": [ + 46, + 369, + 287, + 430 + ], + "type": "text", + "content": "opposed to other schedules. This is especially the case at the lower range of FID values. The results indicate that having a large scale at the beginning of sampling is more important than later stages, especially in use cases where diversity is paramount." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 436, + 151, + 449 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 436, + 151, + 449 + ], + "spans": [ + { + "bbox": [ + 47, + 436, + 151, + 449 + ], + "type": "text", + "content": "4.4. Super-Resolution" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 454, + 287, + 587 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 454, + 287, + 587 + ], + "spans": [ + { + "bbox": [ + 46, + 454, + 287, + 587 + ], + "type": "text", + "content": "Finally, we investigate the importance of hierarchical super-resolution. We compare our method with two baselines. In the first baseline, we directly use the " + }, + { + "bbox": [ + 46, + 454, + 287, + 587 + ], + "type": "inline_equation", + "content": "32^{3}" + }, + { + "bbox": [ + 46, + 454, + 287, + 587 + ], + "type": "text", + "content": " resolution results from the coarse transformer and use a 3D-UNET based super-resolution network to translate from " + }, + { + "bbox": [ + 46, + 454, + 287, + 587 + ], + "type": "inline_equation", + "content": "32^{3}" + }, + { + "bbox": [ + 46, + 454, + 287, + 587 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 46, + 454, + 287, + 587 + ], + "type": "inline_equation", + "content": "64^{3}" + }, + { + "bbox": [ + 46, + 454, + 287, + 587 + ], + "type": "text", + "content": ". For the second baseline, we train a transformer directly on " + }, + { + "bbox": [ + 46, + 454, + 287, + 587 + ], + "type": "inline_equation", + "content": "64^{3}" + }, + { + "bbox": [ + 46, + 454, + 287, + 587 + ], + "type": "text", + "content": " VQ-VAE that is conditioned on text features as opposed to a coarse resolution grid. We refer to this as " + }, + { + "bbox": [ + 46, + 454, + 287, + 587 + ], + "type": "inline_equation", + "content": "64^{3}" + }, + { + "bbox": [ + 46, + 454, + 287, + 587 + ], + "type": "text", + "content": "-DS (64" + }, + { + "bbox": [ + 46, + 454, + 287, + 587 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 46, + 454, + 287, + 587 + ], + "type": "text", + "content": " direct synthesis). The results are shown in Table 3(c). It can be seen from the table that latent-based super-resolution outperforms the baselines in both accuracy and diversity." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 597, + 119, + 609 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 597, + 119, + 609 + ], + "spans": [ + { + "bbox": [ + 47, + 597, + 119, + 609 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 617, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 617, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 617, + 288, + 715 + ], + "type": "text", + "content": "We present CLIP-Sculptor, a text-to-3D-shape generation method capable of producing shapes of high fidelity and diversity without the need for (text, shape) pairs. To achieve this, CLIP-Sculptor leverages CLIP and implements super-resolution in a discrete latent space with a hierarchical architecture and a novel annealed variant of classifier-free guidance on a mask-based model. We validate CLIP-Sculptor by comparing it with a number of base" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 369, + 547, + 464 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 369, + 547, + 464 + ], + "spans": [ + { + "bbox": [ + 304, + 369, + 547, + 464 + ], + "type": "text", + "content": "lines in terms of FID and accuracy, finding that CLIP-Sculptor is the new state of the art for this problem. In experimenting with different guidance scale scheduling, we find that constant scale scheduling did not always work the best, an important finding for the diffusion modeling community that may improve generation quality. Our paper is an important step towards improved quality and diversity of 3D text-to-shape generation outcomes." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 485, + 547, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 485, + 547, + 617 + ], + "spans": [ + { + "bbox": [ + 304, + 485, + 547, + 617 + ], + "type": "text", + "content": "Limitations and Societal Impact. Despite providing higher quality over baseline methods, CLIP-Sculptor still suffers from limitations in its ability to capture smaller details (e.g., chair with a hole on its back) and generate shapes from text prompts involving counting (e.g., chair with four slats). Furthermore, CLIP-Sculptor also fails on many text queries which are not present in the prior knowledge of CLIP. The method also lacks the ability to produce texture. In future work, we will address these limitations by exploring implicit representations that can capture even finer details and by investigating neural networks that can count." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 618, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 618, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 618, + 547, + 715 + ], + "type": "text", + "content": "Text-to-shape approaches can make the generation of 3D content more accessible and efficient. Within the text-to-image domain there has been concern that automatic approaches may replace artists. Because our approach produces voxels, a representation that lends well to direct manipulation and handoff to 3D designers, we believe that our text-to-3D approach can augment 3D design workflows rather than replace them." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "18346" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 714 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 134 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 134 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 134 + ], + "type": "text", + "content": "[1] Panos Achlioptas, Olga Diamanti, Ioannis Mitliagkas, and Leonidas Guibas. Learning representations and generative models for 3d point clouds. In International conference on machine learning, pages 40-49. PMLR, 2018. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 135, + 288, + 178 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 135, + 288, + 178 + ], + "spans": [ + { + "bbox": [ + 53, + 135, + 288, + 178 + ], + "type": "text", + "content": "[2] Panos Achlioptas, Judy Fan, X.D. Robert Hawkins, D. Noah Goodman, and J. Leonidas Guibas. ShapeGlot: Learning language for shape differentiation. CoRR, abs/1905.02925, 2019. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 180, + 288, + 235 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 180, + 288, + 235 + ], + "spans": [ + { + "bbox": [ + 53, + 180, + 288, + 235 + ], + "type": "text", + "content": "[3] Angel X. Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, Jianxiong Xiao, Li Yi, and Fisher Yu. Shapenet: An information-rich 3d model repository, 2015. cite arxiv:1512.03012. 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 236, + 288, + 289 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 236, + 288, + 289 + ], + "spans": [ + { + "bbox": [ + 53, + 236, + 288, + 289 + ], + "type": "text", + "content": "[4] Huiwen Chang, Han Zhang, Lu Jiang, Ce Liu, and William T Freeman. Maskgit: Masked generative image transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11315-11325, 2022. 2, 3, 4, 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 291, + 288, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 291, + 288, + 346 + ], + "spans": [ + { + "bbox": [ + 53, + 291, + 288, + 346 + ], + "type": "text", + "content": "[5] Kevin Chen, Christopher B Choy, Manolis Savva, Angel X Chang, Thomas Funkhouser, and Silvio Savarese. Text2shape: Generating shapes from natural language by learning joint embeddings. In Asian Conference on Computer Vision, pages 100-116. Springer, 2018. 1, 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 346, + 288, + 390 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 346, + 288, + 390 + ], + "spans": [ + { + "bbox": [ + 53, + 346, + 288, + 390 + ], + "type": "text", + "content": "[6] Zhiqin Chen and Hao Zhang. Learning implicit fields for generative shape modeling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5939-5948, 2019. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 391, + 288, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 391, + 288, + 445 + ], + "spans": [ + { + "bbox": [ + 53, + 391, + 288, + 445 + ], + "type": "text", + "content": "[7] Christopher B Choy, Danfei Xu, JunYoung Gwak, Kevin Chen, and Silvio Savarese. 3d-r2n2: A unified approach for single and multi-view 3d object reconstruction. In European conference on computer vision, pages 628-644. Springer, 2016. 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 447, + 288, + 490 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 447, + 288, + 490 + ], + "spans": [ + { + "bbox": [ + 53, + 447, + 288, + 490 + ], + "type": "text", + "content": "[8] Prafulla Dhariwal, Heewoo Jun, Christine Payne, Jong Wook Kim, Alec Radford, and Ilya Sutskever. Jukebox: A generative model for music. arXiv preprint arXiv:2005.00341, 2020. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 491, + 288, + 524 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 491, + 288, + 524 + ], + "spans": [ + { + "bbox": [ + 53, + 491, + 288, + 524 + ], + "type": "text", + "content": "[9] Rao Fu, Xiao Zhan, Yiwen Chen, Daniel Ritchie, and Srinath Sridhar. Shapecrafter: A recursive text-conditioned 3d shape generation model, 2022. 1, 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 525, + 287, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 525, + 287, + 557 + ], + "spans": [ + { + "bbox": [ + 48, + 525, + 287, + 557 + ], + "type": "text", + "content": "[10] Oran Gafni, Adam Polyak, Oron Ashual, Shelly Sheynin, Devi Parikh, and Yaniv Taigman. Make-a-scene: Scene-based text-to-image generation with human priors, 2022. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 558, + 287, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 558, + 287, + 602 + ], + "spans": [ + { + "bbox": [ + 48, + 558, + 287, + 602 + ], + "type": "text", + "content": "[11] Rinon Gal, Yuval Alaluf, Yuval Atzmon, Or Patashnik, Amit H. Bermano, Gal Chechik, and Daniel Cohen-Or. An image is worth one word: Personalizing text-to-image generation using textual inversion, 2022. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 602, + 287, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 602, + 287, + 667 + ], + "spans": [ + { + "bbox": [ + 48, + 602, + 287, + 667 + ], + "type": "text", + "content": "[12] Shuyang Gu, Dong Chen, Jianmin Bao, Fang Wen, Bo Zhang, Dongdong Chen, Lu Yuan, and Baining Guo. Vector quantized diffusion model for text-to-image synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10696-10706, 2022. 2, 4" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 669, + 287, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 669, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 48, + 669, + 287, + 714 + ], + "type": "text", + "content": "[13] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 4" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 714 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 307, + 73, + 547, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 547, + 127 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 547, + 127 + ], + "type": "text", + "content": "[14] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 5" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 128, + 547, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 128, + 547, + 183 + ], + "spans": [ + { + "bbox": [ + 307, + 128, + 547, + 183 + ], + "type": "text", + "content": "[15] Jonathan Ho, William Chan, Chitwan Sahara, Jay Whang, Ruiqi Gao, Alexey Gritsanko, Diederik P. Kingma, Ben Poole, Mohammad Norouzi, David J. Fleet, and Tim Salimans. Imagen video: High definition video generation with diffusion models, 2022. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 183, + 547, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 183, + 547, + 205 + ], + "spans": [ + { + "bbox": [ + 307, + 183, + 547, + 205 + ], + "type": "text", + "content": "[16] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598, 2022. 2, 4, 5" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 205, + 547, + 250 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 205, + 547, + 250 + ], + "spans": [ + { + "bbox": [ + 307, + 205, + 547, + 250 + ], + "type": "text", + "content": "[17] Moritz Ibing, Isaak Lim, and Leif Kobbelt. 3d shape generation with grid-based implicit functions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13559-13568, 2021. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 250, + 547, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 250, + 547, + 304 + ], + "spans": [ + { + "bbox": [ + 307, + 250, + 547, + 304 + ], + "type": "text", + "content": "[18] Ajay Jain, Ben Mildenhall, Jonathan T Barron, Pieter Abbeel, and Ben Poole. Zero-shot text-guided object generation with dream fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 867-876, 2022. 1, 2, 5, 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 304, + 547, + 350 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 304, + 547, + 350 + ], + "spans": [ + { + "bbox": [ + 307, + 304, + 547, + 350 + ], + "type": "text", + "content": "[19] Nasir Khalid, Tianhao Xie, Eugene Belilovsky, and Tiberiu Popa. Clip-mesh: Generating textured meshes from text using pretrained image-text models. ACM Transactions on Graphics (TOG), Proc. SIGGRAPH Asia, 2022. 1, 2, 5, 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 350, + 547, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 350, + 547, + 392 + ], + "spans": [ + { + "bbox": [ + 307, + 350, + 547, + 392 + ], + "type": "text", + "content": "[20] Weixin Liang, Yuhui Zhang, Yongchan Kwon, Serena Yeung, and James Zou. Mind the gap: Understanding the modality gap in multi-modal contrastive representation learning. arXiv preprint arXiv:2203.02053, 2022. 4" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 392, + 547, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 392, + 547, + 415 + ], + "spans": [ + { + "bbox": [ + 307, + 392, + 547, + 415 + ], + "type": "text", + "content": "[21] Vivian Liu, Han Qiao, and Lydia Chilton. Opal: Multimodal image generation for news illustration, 2022. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 415, + 547, + 448 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 415, + 547, + 448 + ], + "spans": [ + { + "bbox": [ + 307, + 415, + 547, + 448 + ], + "type": "text", + "content": "[22] Vivian Liu, Jo Vermeulen, George Fitzmaurice, and Justin Matejka. 3dall-e: Integrating text-to-image ai in 3d design workflows, 2022. 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 448, + 547, + 482 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 448, + 547, + 482 + ], + "spans": [ + { + "bbox": [ + 307, + 448, + 547, + 482 + ], + "type": "text", + "content": "[23] Zhengzhe Liu, Yi Wang, Xiaojuan Qi, and Chi-Wing Fu. Towards implicit text-guided 3d shape generation. arXiv preprint arXiv:2203.14622, 2022. 1, 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 482, + 547, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 482, + 547, + 536 + ], + "spans": [ + { + "bbox": [ + 307, + 482, + 547, + 536 + ], + "type": "text", + "content": "[24] Lars Mescheder, Michael Oechsle, Michael Niemeyer, Sebastian Nowozin, and Andreas Geiger. Occupancy networks: Learning 3d reconstruction in function space. In Proceedings IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2019. 5" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 537, + 547, + 592 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 537, + 547, + 592 + ], + "spans": [ + { + "bbox": [ + 307, + 537, + 547, + 592 + ], + "type": "text", + "content": "[25] Oscar Michel, Roi Bar-On, Richard Liu, Sagie Benaim, and Rana Hanocka. Text2mesh: Text-driven neural stylization for meshes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13492-13502, 2022. 1, 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 592, + 547, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 592, + 547, + 635 + ], + "spans": [ + { + "bbox": [ + 307, + 592, + 547, + 635 + ], + "type": "text", + "content": "[26] Paritosh Mittal, Yen-Chi Cheng, Maneesh Singh, and Shubham Tulsiani. Autosdf: Shape priors for 3d completion, reconstruction and generation. arXiv preprint arXiv:2203.09516, 2022. 1, 2, 5" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 635, + 547, + 680 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 635, + 547, + 680 + ], + "spans": [ + { + "bbox": [ + 307, + 635, + 547, + 680 + ], + "type": "text", + "content": "[27] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models, 2021. 4" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 680, + 547, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 680, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 307, + 680, + 547, + 714 + ], + "type": "text", + "content": "[28] Aaron van den Oord, Oriol Vinyals, and Koray Kavukcuoglu. Neural discrete representation learning. arXiv preprint arXiv:1711.00937, 2017. 2, 3, 4" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "18347" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 286, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 48, + 72, + 286, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 286, + 105 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 286, + 105 + ], + "type": "text", + "content": "[29] Ben Poole, Ajay Jain, Jonathan T. Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. arXiv, 2022. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 107, + 286, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 107, + 286, + 162 + ], + "spans": [ + { + "bbox": [ + 49, + 107, + 286, + 162 + ], + "type": "text", + "content": "[30] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision, 2021. 1, 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 163, + 286, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 163, + 286, + 205 + ], + "spans": [ + { + "bbox": [ + 49, + 163, + 286, + 205 + ], + "type": "text", + "content": "[31] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 2022. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 209, + 286, + 261 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 209, + 286, + 261 + ], + "spans": [ + { + "bbox": [ + 49, + 209, + 286, + 261 + ], + "type": "text", + "content": "[32] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In International Conference on Machine Learning, pages 8821-8831. PMLR, 2021. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 49, + 264, + 286, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 264, + 286, + 297 + ], + "spans": [ + { + "bbox": [ + 49, + 264, + 286, + 297 + ], + "type": "text", + "content": "[33] Ali Razavi, Aaron Van den Oord, and Oriol Vinyals. Generating diverse high-fidelity images with vq-vae-2. Advances in neural information processing systems, 32, 2019. 2, 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 49, + 299, + 286, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 299, + 286, + 330 + ], + "spans": [ + { + "bbox": [ + 49, + 299, + 286, + 330 + ], + "type": "text", + "content": "[34] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models, 2021. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 49, + 333, + 286, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 333, + 286, + 387 + ], + "spans": [ + { + "bbox": [ + 49, + 333, + 286, + 387 + ], + "type": "text", + "content": "[35] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10684-10695, June 2022. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 49, + 388, + 286, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 388, + 286, + 442 + ], + "spans": [ + { + "bbox": [ + 49, + 388, + 286, + 442 + ], + "type": "text", + "content": "[36] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10684-10695, 2022. 2, 4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 49, + 445, + 286, + 509 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 445, + 286, + 509 + ], + "spans": [ + { + "bbox": [ + 49, + 445, + 286, + 509 + ], + "type": "text", + "content": "[37] Aditya Sanghi, Hang Chu, Joseph G Lambourne, Ye Wang, Chin-Yi Cheng, Marco Fumero, and Kamal Rahimi Malekshan. Clip-forge: Towards zero-shot text-to-shape generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18603-18613, 2022. 1, 2, 5, 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 49, + 511, + 286, + 554 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 511, + 286, + 554 + ], + "spans": [ + { + "bbox": [ + 49, + 511, + 286, + 554 + ], + "type": "text", + "content": "[38] Nikolay Savinov, Junyoung Chung, Mikolaj Binkowski, Erich Elsen, and Aaron van den Oord. Step-unrolled denoising autoencoders for text generation. arXiv preprint arXiv:2112.06749, 2021. 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 49, + 556, + 286, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 556, + 286, + 588 + ], + "spans": [ + { + "bbox": [ + 49, + 556, + 286, + 588 + ], + "type": "text", + "content": "[39] Nikolay Savinov, Junyoung Chung, Mikolaj Binkowski, Erich Elsen, and Aaron van den Oord. Step-unrolled denoising autoencoders for text generation, 2021. 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 49, + 590, + 286, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 590, + 286, + 633 + ], + "spans": [ + { + "bbox": [ + 49, + 590, + 286, + 633 + ], + "type": "text", + "content": "[40] Mohit Shridhar, Lucas Manuelli, and Dieter Fox. CIport: What and where pathways for robotic manipulation. In Proceedings of the 5th Conference on Robot Learning (CoRL), 2021. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 49, + 635, + 286, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 635, + 286, + 689 + ], + "spans": [ + { + "bbox": [ + 49, + 635, + 286, + 689 + ], + "type": "text", + "content": "[41] Can Wang, Menglei Chai, Mingming He, Dongdong Chen, and Jing Liao. Clip-nerf: Text-and-image driven manipulation of neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3835–3844, 2022. 1, 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 49, + 692, + 286, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 692, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 49, + 692, + 286, + 713 + ], + "type": "text", + "content": "[42] Xingguang Yan, Liqiang Lin, Niloy J Mitra, Dani Lischinski, Daniel Cohen-Or, and Hui Huang. Shapeformer:" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 545, + 262 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 327, + 73, + 545, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 545, + 116 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 545, + 116 + ], + "type": "text", + "content": "Transformer-based shape completion via sparse representation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6239-6249, 2022. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 118, + 545, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 118, + 545, + 173 + ], + "spans": [ + { + "bbox": [ + 308, + 118, + 545, + 173 + ], + "type": "text", + "content": "[43] Guandao Yang, Xun Huang, Zekun Hao, Ming-Yu Liu, Serge Belongie, and Bharath Hariharan. Pointflow: 3d point cloud generation with continuous normalizing flows. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4541-4550, 2019. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 175, + 545, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 175, + 545, + 217 + ], + "spans": [ + { + "bbox": [ + 308, + 175, + 545, + 217 + ], + "type": "text", + "content": "[44] Yufan Zhou, Ruiyi Zhang, Changyou Chen, Chunyuan Li, Chris Tensmeyer, Tong Yu, Jiuming Gu, Jinhui Xu, and Tong Sun. Lafite: Towards language-free training for text-to-image generation. arXiv preprint arXiv:2111.13792, 2021. 4" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 220, + 545, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 220, + 545, + 262 + ], + "spans": [ + { + "bbox": [ + 308, + 220, + 545, + 262 + ], + "type": "text", + "content": "[45] Adrian Łancucki, Jan Chorowski, Guillaume Sanchez, Ricard Marxer, Nanxin Chen, Hans J. G. A. Dolfing, Sameer Khurana, Tanel Alumäe, and Antoine Laurent. Robust training of vector quantized bottleneck models, 2020. 4" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "18348" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/CLIP2Protect_ Protecting Facial Privacy Using Text-Guided Makeup via Adversarial Latent Search/d379e97e-b7a0-42bb-a735-a3de89cdb054_content_list.json b/2023/CLIP2Protect_ Protecting Facial Privacy Using Text-Guided Makeup via Adversarial Latent Search/d379e97e-b7a0-42bb-a735-a3de89cdb054_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..374a50c95a33a059f4466a570c8d8b045748b043 --- /dev/null +++ b/2023/CLIP2Protect_ Protecting Facial Privacy Using Text-Guided Makeup via Adversarial Latent Search/d379e97e-b7a0-42bb-a735-a3de89cdb054_content_list.json @@ -0,0 +1,1461 @@ +[ + { + "type": "text", + "text": "CLIP2Protect: Protecting Facial Privacy using Text-Guided Makeup via Adversarial Latent Search", + "text_level": 1, + "bbox": [ + 120, + 119, + 849, + 162 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Fahad Shamshad Muzammal Naseer Karthik Nandakumar Mohamed Bin Zayed University of AI, UAE", + "bbox": [ + 233, + 185, + 733, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{fahad.shamshad, muzammal.naseer, karthik.nandakumar}@mbzuai.ac.ae", + "bbox": [ + 189, + 223, + 772, + 238 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/84f4e77f6d06dc8ebde7ec482aebe28aac8a0e06def5699e03738271ad09c778.jpg", + "image_caption": [ + "Protected Original" + ], + "image_footnote": [], + "bbox": [ + 102, + 270, + 883, + 468 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/bf65165568d7efeaaee26804444fe4923be66bba4fb54e90fbbc1c2f0234b563.jpg", + "image_caption": [ + "Figure 1. The proposed approach crafts \"naturalistic\" and transferable text-guided adversarial faces to deceive black-box face recognition systems. First row shows original images that need to be protected and second row shows corresponding protected images along with the user-defined makeup text prompts that guide the adversarial search. Comparison against existing methods is shown in the third row. The yellow text represents the confidence score (higher is better) output by a commercial API (Face++, when matching the protected image against the target identity shown in the bottom right. The reference image used by [22] for makeup transfer is shown at the bottom corner of the corresponding adversarial image." + ], + "image_footnote": [], + "bbox": [ + 102, + 474, + 883, + 564 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/35c8881d6894b6f19f0c6d90b1d9859eb4492ae04adf388b9c4280dcfc386773.jpg", + "image_caption": [ + "Target" + ], + "image_footnote": [], + "bbox": [ + 812, + 570, + 883, + 625 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 664, + 313, + 679 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The success of deep learning based face recognition systems has given rise to serious privacy concerns due to their ability to enable unauthorized tracking of users in the digital world. Existing methods for enhancing privacy fail to generate \"naturalistic\" images that can protect facial privacy without compromising user experience. We propose a novel two-step approach for facial privacy protection that relies on finding adversarial latent codes in the low-dimensional manifold of a pretrained generative model. The first step inverts the given face image into the latent space and finetunes the generative model to achieve an accurate reconstruction of the given image from its latent code. This step produces a good initialization, aiding the generation", + "bbox": [ + 73, + 704, + 470, + 900 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "of high-quality faces that resemble the given identity. Subsequently, user-defined makeup text prompts and identity-preserving regularization are used to guide the search for adversarial codes in the latent space. Extensive experiments demonstrate that faces generated by our approach have stronger black-box transferability with an absolute gain of $12.06\\%$ over the state-of-the-art facial privacy protection approach under the face verification task. Finally, we demonstrate the effectiveness of the proposed approach for commercial face recognition systems. Our code is available at https://github.com/fahadshamshad/Clip2Protect.", + "bbox": [ + 496, + 665, + 890, + 830 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 501, + 843, + 630, + 857 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Deep learning based face recognition (FR) systems [43, 61] have found widespread usage in multiple applications,", + "bbox": [ + 500, + 869, + 890, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 44 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "20595", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "table", + "img_path": "images/9183f8082408685d93c8263745f8beff1d5039593c7bc6625f0e43e1cbe901cf.jpg", + "table_caption": [ + "Table 1. Comparison among different facial privacy protection methods w.r.t. the natural outputs, black box setting, experiments under face verification and identification tasks, unrestricted (semantically meaningful), and more flexible text guided adversaries." + ], + "table_footnote": [], + "table_body": "
Adv-Makeup [71]TIP-IM [70]AMT-GAN [22]Ours
Natural outputsYesPartiallyPartiallyYes
Black boxYesYesYesYes
VerificationYesNoYesYes
IdentificationNoYesNoYes
UnrestrictedYesNoYesYes
Text guidedNoNoNoYes
", + "bbox": [ + 81, + 148, + 468, + 244 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "including security [63], biometrics [38], and criminal investigation [45], outperforming humans in many scenarios [12, 48, 61]. Despite positive aspects of this technology, FR systems seriously threaten personal security and privacy in the digital world because of their potential to enable mass surveillance capabilities [1, 67]. For example, government and private entities can use FR systems to track user relationships and activities by scraping face images from social media profiles such as Twitter, LinkedIn, and Facebook [18, 20]. These entities generally use proprietary FR systems, whose specifications are unknown to the public (black box model). Therefore, there is an urgent need for an effective approach that protects facial privacy against such unknown FR systems.", + "bbox": [ + 75, + 263, + 468, + 474 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "An ideal facial privacy protection algorithm must strike the right balance between naturalness and privacy protection [70, 77]. In this context, \"naturalness\" is defined as the absence of any noise artifacts that can be easily perceived by human observers and the preservation of human-perceived identity. \"Privacy protection\" refers to the fact that the protected image must be capable of deceiving a black-box malicious FR system. In other words, the protected image must closely resemble the given face image and be artifact-free for a human observer, while at the same time fool an unknown automated FR system. Since failure to generate naturalistic faces can significantly affect user experience on social media platforms, it is a necessary precondition for adoption of a privacy-enhancement algorithm.", + "bbox": [ + 75, + 477, + 468, + 686 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recent works exploit adversarial attacks [57] to conceal user identity by overlaying noise-constrained (bounded) adversarial perturbations on the original face image [6,53,74]. Since the adversarial examples are generally optimized in the image space, it is often difficult to simultaneously achieve naturalness and privacy [70]. Unlike noise-based methods, unrestricted adversarial examples are not constrained by the magnitude of perturbation in the image space and have demonstrated better perceptual realism for human observers while being adversially effective [3,55,68,76].", + "bbox": [ + 75, + 688, + 468, + 839 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Several efforts have been made to generate unrestricted adversarial examples that mislead FR systems (see Tab. 1) [22, 25, 39, 72]. Among these, adversarial makeup based methods [22, 72] are gaining increasing attention as they", + "bbox": [ + 76, + 839, + 468, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "can embed adversarial modifications in a more natural way. These approaches use generative adversarial networks [15] (GANs) to adversarily transfer makeup from a given reference image to the user's face image while impersonating a target identity. However, existing techniques based on adversarial makeup transfer have the following limitations: (i) adversarial toxicity in these methods hamper the performance of the makeup transfer module, thereby resulting in unnatural faces with makeup artifacts (see Fig. 1); (ii) the use of a reference image to define the desired makeup style affects the practicality of this approach; (iii) for every new target identity, these approaches require end-to-end retraining from scratch using large makeup datasets; and (iv) most of these methods primarily aim at impersonation of the target identity, whereas the desired privacy objective is dodging, i.e., multiple images of the user's face scraped from different social media sites must not match with each other.", + "bbox": [ + 496, + 90, + 890, + 347 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To mitigate the above problems, we propose a new approach to protect user facial privacy on online platforms (Sec. 3). The proposed approach aims to search for adversarial latent codes in a low-dimensional manifold learned by a generative model trained to generate face images [2, 27]. Our main contributions are:", + "bbox": [ + 496, + 349, + 890, + 441 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Facial Privacy-protection Framework Using Adversarial Latent Codes: Given a face image, we propose a novel two-step method to search for adversarial latent codes, which can be used by a generative model (e.g., StyleGAN) to produce face images with high visual quality that matches human-perceived identity, while deceiving black-box FR systems.", + "- Adversarial Makeup Transfer using Textual Prompts: A critical component of the above framework is a technique for leveraging user-defined textual (makeup) prompts to traverse over the latent manifold of the generative model and find transferable adversarial latent codes. Our approach effectively hides attack information in the desired makeup style, without the need for any large makeup dataset or retraining of models for different target identities.", + "- Identity Preserving Regularization: We propose a regularizer that preserves identity-related attributes within the latent space of the generative model and ensures that the protected face image visually resembles the original face." + ], + "bbox": [ + 517, + 455, + 890, + 792 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Extensive experiments (Sec. 4.1) for both face verification and identification scenarios demonstrate the effectiveness of our approach against black-box FR models and online commercial facial recognition APIs (Sec. 4.2). Furthermore, we provide detailed ablative analysis to dissect the performance of different components of our approach (Sec. 4.3).", + "bbox": [ + 496, + 809, + 890, + 901 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "20596", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 76, + 89, + 218, + 104 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Obfuscation Methods: Obfuscation is the most widely used technique [38] to protect user's facial privacy. Earlier obfuscation approaches typically degrade the quality of the original face image by applying simple operations such as masking [52, 64], filtering [33, 78], and image transformations [8, 36, 62]. While these relatively simple obfuscation techniques are reasonable for surveillance applications, they are ill-suited for online/social media platforms where user experience is critical [41]. Though deep learning based obfuscation approaches generate more realistic images [4, 7, 56, 58], they often result in a change of identity compared to the original image and occasionally produce undesirable artifacts [30, 31, 34].", + "bbox": [ + 75, + 114, + 468, + 311 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Noise-based Adversarial Examples: Adversarial attacks have been used to protect users from unauthorized FR models. Some methods [6, 53] rely on data poisoning to deceive targeted FR models, but are less practical because access to the training data or the gallery set of the unknown FR system is often not available. Other approaches have used game-theory perspective [42] in white-box settings or person-specific privacy masks (one mask per person) to generate protected images at the cost of acquiring multiple images of the same user [77]. In contrast, we aim to fool the black box FR model using only single image. In TIP-IM [70], targeted optimization was used to generate privacy masks against unknown FR models by introducing a naturalness constraint. While this approach provides effective privacy, it generates output images with perceptible noises that can affect the user experience [70].", + "bbox": [ + 75, + 311, + 470, + 551 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Unrestricted Adversarial Examples: Unrestricted adversarial attacks (UAAs) are not constrained by the perturbation norm and can induce large but semantically meaningful perturbations. These attacks have been extensively studied in image classification literature [3, 35, 55, 68, 73, 76] and it has been shown that outputs generated via UAAs are less perceptible to human observers as compared to noise-based adversarial attacks. Motivated by this observation, patch-based unrestricted attacks have been proposed to generate wearable adversarial accessories like colorful glasses [54], hat [29] or random patch [69] to fool the FR model, but such synthesized patches generally have weak transferability due to the limited editing region and the large visible pattern compromises naturalness and affects user experience. Recently, generative models [24, 50] have been leveraged to craft UAAs against FR models. However, these generative approaches are either designed for the white-box settings [46, 79] or show limited performance in query-free black-box settings [25]. Makeup-based UAAs [17, 72] have also been proposed against FR systems by embedding the perturbations into a natural makeup effect. These makeup based attacks have also been exploited to protect the user privacy by applying adversarial makeup on the user", + "bbox": [ + 75, + 551, + 470, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "face image [22]. However, interference between adversarial perturbations and makeup transfer can produce undesirable makeup artifacts in the output images. Moreover, these attacks generally assume access to large makeup datasets for training models and require a reference makeup image. In contrast, our approach finds adversarial faces on the natural image manifold in black-box setting via guidance from makeup text prompt, which makes it less susceptible to artifacts (see Fig. 1) and more practical.", + "bbox": [ + 496, + 90, + 890, + 226 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Vision-Language Modelling: Cross-modal vision-language modelling has attracted significant attention in recent years [13]. OpenAI introduced CLIP [47] that is trained on 400 million image-text pairs using contrastive objective and maps both image and text in a joint multimodal embedding space. With powerful representation embedding of CLIP, several methods have been proposed to manipulate images with text-guidance. StyleCLIP [44] and DiffusionCLIP [28, 40] leverage the powerful generative capabilities of StyleGAN and diffusion models to manipulate images with text prompts. Other similar works include HairCLIP [66], CLIP-NeRF [60], CLIPstyler [32], and CLIPDraw [14]. While these methods focus on the text-guidance ability of CLIP, our approach aims to find the adversarial latent codes in a generative model's latent space for privacy protection against black-box FR models.", + "bbox": [ + 496, + 227, + 892, + 469 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Proposed Approach for Facial Privacy", + "text_level": 1, + "bbox": [ + 500, + 481, + 843, + 500 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our goal is to protect user facial privacy on online platforms against unknown (black-box) FR models without compromising on the user's online experience. The proposed approach finds protected faces by adversarially exploring the low-dimensional latent space of a pretrained generative model that is trained on natural face images. To avoid artifacts in the protected image, we restrict the search for adversarial faces close to the clean image manifold learned by the generative model. Moreover, we propose to optimize only over identity-preserving latent codes in the latent space. This effectively preserves human-perceived identity during attack while offering high privacy against automated systems. Further, we employ natural makeup-like perturbations via guidance from a text prompt, which provides more flexibility to the user compared to reference image-based adversarial makeup transfer [22].", + "bbox": [ + 496, + 506, + 890, + 750 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Preliminaries", + "text_level": 1, + "bbox": [ + 500, + 756, + 640, + 771 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Let $\\pmb{x} \\in \\mathcal{X} \\subset \\mathbb{R}^n$ denote the given original/real face image. Let $f(\\pmb{x}) : \\mathcal{X} \\to \\mathbb{R}^d$ be a FR model that extracts a fixed-length normalized feature representation. Let $\\mathcal{D}(\\pmb{x}_1, \\pmb{x}_2) = D(f(\\pmb{x}_1), f(\\pmb{x}_2))$ be a distance metric that measures the dissimilarity between two face images $\\pmb{x}_1$ and $\\pmb{x}_2$ based on their respective representations $f(\\pmb{x}_1)$ and $f(\\pmb{x}_2)$ . Generally a FR system can operate in two modes: verification and identification. A face verification", + "bbox": [ + 496, + 779, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "20597", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/d22a9ba3df87594c2ff4ea2e1367252711931d277ddbfbe0fbd87c1f409c524f.jpg", + "image_caption": [ + "Figure 2. Overall pipeline of the proposed approach to protect users facial privacy. Our proposed approach searches for the adversarial latent codes on the generative manifold to reconstruct an adversarial face that is capable of fooling unknown FR systems for privacy protection. Our approach allows \"makeup\" editing in an adversarial manner through user defined textual prompts and thereby enhance the user's online experience. Our text-guided objective searches for such latent codes while keeping the original identity preserved." + ], + "image_footnote": [], + "bbox": [ + 99, + 85, + 869, + 304 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "system predicts that two faces belong to the same identity if $\\mathcal{D}(\\pmb{x}_1, \\pmb{x}_2) \\leq \\tau$ , where $\\tau$ is the system threshold. On the other hand, a (closed set) face identification system compares the input image (probe) against a set of face images (gallery) and outputs the identity whose representation is most similar to that of the probe. Since the attacker can employ verification or identification to determine the user identity using black-box FR models, a protection approach should conceal the user's identity in both scenarios.", + "bbox": [ + 75, + 396, + 468, + 532 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "User privacy can be protected by misleading the malicious FR model through impersonation or dodging attacks. In the context of verification, impersonation (false match) implies that the protected face matches with the face of a specific target identity and dodging (false non-match) means that the protected face does not match with some other image of the same person. Similarly, for face identification, impersonation ensures that the protected image gets matched to a specified target identity in the gallery set, while dodging prevents the protected face from matching with images of the same person in the gallery.", + "bbox": [ + 75, + 532, + 470, + 699 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Problem Statement: Given the original face image $\\pmb{x}$ , our goal is to generate a protected face image $\\pmb{x}^p$ such that $\\mathcal{D}(\\pmb{x}^p, \\pmb{x})$ is large (for successful dodging attack) and $\\mathcal{D}(\\pmb{x}^p, \\pmb{x}^t)$ is small (for successfully impersonating a target face $\\pmb{x}^t$ ), where $\\mathcal{O}(\\pmb{x}) \\neq \\mathcal{O}(\\pmb{x}^t)$ and $\\mathcal{O}$ is the oracle that gives the true identity labels. At the same time, we want to minimize $\\mathcal{H}(\\pmb{x}^p, \\pmb{x})$ , where $\\mathcal{H}$ quantifies the degree of unnaturalness introduced in the protected image $\\pmb{x}^p$ in relation to the original image $\\pmb{x}$ . Formally, the optimization problem that we aim to solve is:", + "bbox": [ + 75, + 699, + 468, + 849 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\min _ {\\boldsymbol {x} ^ {p}} \\mathcal {L} (\\boldsymbol {x} ^ {p}) = \\mathcal {D} (\\boldsymbol {x} ^ {p}, \\boldsymbol {x} ^ {t}) - \\mathcal {D} (\\boldsymbol {x} ^ {p}, \\boldsymbol {x}) \\tag {1} \\\\ \\mathrm {s . t .} \\mathcal {H} (\\boldsymbol {x} ^ {p}, \\boldsymbol {x}) \\leq \\epsilon \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 147, + 854, + 468, + 896 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/2f7422eee5a7f7e84514603fc2fbbb4edb25e2397277343c442a1b4f18c6235b.jpg", + "image_caption": [ + "(a) Original" + ], + "image_footnote": [], + "bbox": [ + 501, + 393, + 629, + 493 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/d881a1341c142b5295fe4413d8f2b62a3f46bf4fa85b858312701c8b0020eaee.jpg", + "image_caption": [ + "(b) Encoder Inversion" + ], + "image_footnote": [], + "bbox": [ + 632, + 393, + 759, + 493 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/6a8bc6961077e8abed5b5570bff0e24b561455f7b2f07958f9ec3ac193baf6f9.jpg", + "image_caption": [ + "(c) Generator finetuning", + "Figure 3. Generator finetuning allows near-perfect reconstructions of LFW dataset sample. This is crucial for the online experience of users. Matching scores returned by Face++ API are 62.38 and 98.96 for encoder and generator-finetuned inversions, respectively." + ], + "image_footnote": [], + "bbox": [ + 761, + 393, + 890, + 493 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\epsilon$ is a bound on the adversarial perturbation. For noise-based approach, $\\mathcal{H}(\\pmb{x}^p,\\pmb{x}) = \\| \\pmb{x} - \\pmb{x}^p\\|_p$ , where $\\| \\cdot \\|_p$ denotes the $L_p$ norm. However, direct enforcement of the perturbation constraint leads to visible artifacts, which affects visual quality and user experience. Constraining the solution search space to a natural image manifold using an effective image prior can produce more realistic images. Note that the distance metric $\\mathcal{D}$ is unknown since our goal is to deceive a black-box FR system.", + "bbox": [ + 496, + 593, + 890, + 729 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Makeup Text-Guided Adversarial Faces", + "text_level": 1, + "bbox": [ + 500, + 739, + 841, + 757 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our approach restricts the solution space of the protected face $\\pmb{x}^p$ to lie close to the clean face manifold $\\mathcal{X}$ . This manifold can be learned using a generative model trained on real human faces. Specifically, let $G_{\\theta}(\\pmb{w}) : \\mathcal{W} \\to \\mathbb{R}^n$ denote the pretrained generative model with weights $\\theta$ , where $\\mathcal{W}$ is the latent space. Our proposed approach consists of two stages: (i) latent code initialization (Sec. 3.2.1) and (ii) text-guided adversarial optimization (Sec. 3.2.2). The overall pipeline of the proposed approach is shown in Fig. 2.", + "bbox": [ + 496, + 763, + 890, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "20598", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2.1 Latent Code Initialization", + "text_level": 1, + "bbox": [ + 76, + 90, + 312, + 104 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The latent code initialization stage is based on GAN inversion, which aims to invert the original image $\\pmb{x}$ into the latent space $\\mathcal{W}$ , i.e., find a latent code $\\pmb{w}_{\\mathrm{inv}} \\in \\mathcal{W}$ such that $\\pmb{x}_{\\mathrm{inv}} = G_{\\theta}(\\pmb{w}_{\\mathrm{inv}}) \\approx \\pmb{x}$ . To achieve this, we first use an encoder-based inversion called e4e [59] to infer $\\pmb{w}_{\\mathrm{inv}}$ in $\\mathcal{W}$ from $\\pmb{x}$ i.e., $\\pmb{w}_{\\mathrm{inv}} = I_{\\phi}(\\pmb{x})$ , where $I_{\\phi}: \\mathcal{X} \\to \\mathcal{W}$ is the pretrained encoder with weights $\\phi$ (see Fig. 2).", + "bbox": [ + 75, + 114, + 468, + 220 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We use StyleGAN trained on a high-resolution dataset of face images as the pretrained generative model $G_{\\theta}$ due to its powerful synthesis ability and the disentangled structure of its latent space. A significant challenge during inversion is preserving the identity of the original image i.e., $\\mathcal{O}(\\pmb{x}) = \\mathcal{O}(\\pmb{x}_{\\mathrm{inv}})$ . Generally, optimization and encoder-based inversion approaches struggle to preserve identity after reconstruction [49] (see Fig. 3b). Moreover, when using these approaches, the inversion error can be large for out-of-domain face images with extreme poses and viewpoints, which are quite common in social media applications. Therefore, these approaches cannot be applied directly to invert $\\pmb{x}$ . Instead, motivated by the recent observation [49] that slight changes to the pretrained generator weights do not harm its editing abilities while achieving near-perfect reconstructions, we finetune the pretrained generator weights $\\theta$ instead of the encoder weights $\\phi$ . Specifically, we fix $w_{\\mathrm{inv}} = I_{\\phi}(\\pmb{x})$ and fine-tune $G_{\\theta}$ using the following loss:", + "bbox": [ + 75, + 220, + 470, + 494 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\theta^ {*} = \\underset {\\theta} {\\arg \\min } \\mathcal {L} _ {\\mathrm {L P I P S}} (\\boldsymbol {x}, G _ {\\theta} (\\boldsymbol {w} _ {\\mathrm {i n v}})) + \\lambda_ {2} \\mathcal {L} _ {2} (\\boldsymbol {x}, G _ {\\theta} (\\boldsymbol {w} _ {\\mathrm {i n v}})),\n$$\n", + "text_format": "latex", + "bbox": [ + 91, + 503, + 452, + 527 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathcal{L}_{\\mathrm{LPIPS}}$ is the perceptual loss and $\\mathcal{L}_2$ denotes the pixelwise similarity. The final inverted image $x_{\\mathrm{inv}}^*$ (see Fig. 3c) can be obtained by performing a forward pass of $\\pmb{w}_{\\mathrm{inv}}$ through fine-tuned generator i.e., $\\pmb{x}_{\\mathrm{inv}}^{*} = G_{\\theta^{*}}(\\pmb{w}_{\\mathrm{inv}})$ .", + "bbox": [ + 75, + 536, + 468, + 599 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2.2 Text-guided adversarial optimization", + "text_level": 1, + "bbox": [ + 76, + 618, + 388, + 633 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Given the inverted latent code $\\boldsymbol{w}_{\\mathrm{inv}}$ and fine-tuned generator $G_{\\theta^{*}}(\\cdot)$ , our goal is to adversarially perturb this latent code $\\boldsymbol{w}_{\\mathrm{inv}}$ in the low-dimensional generative manifold $\\mathcal{W}$ to generate a protected face that fools the black-box FR model, while imitating the makeup style of the text prompt $t_{\\mathrm{makeup}}$ .", + "bbox": [ + 75, + 643, + 468, + 719 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To achieve these objectives, we investigate the following questions: (i) how to effectively extract makeup style information from $t_{\\mathrm{makeup}}$ and apply it to the face image $x$ in an adversarial manner?, (ii) how to regularize the optimization process so that the output face image is not qualitatively impaired?, (iii) how to craft effective adversarial perturbations that mislead black-box FR models?, and (iv) how to preserve the human-perceived identity $\\mathcal{O}(x)$ of the original face image while ensuring high privacy?", + "bbox": [ + 75, + 719, + 468, + 854 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The first issue can be addressed by aligning the output adversarial image with the text prompt $t_{\\mathrm{makeup}}$ in the embedding space of a pretrained vision-language model. The", + "bbox": [ + 75, + 854, + 468, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "second issue is addressed by enforcing the adversarial latent code to remain close to initialization $\\boldsymbol{w}_{\\mathrm{inv}}$ . The third issue is solved by crafting transferable text-guided adversarial faces on a white-box surrogate model (or an ensemble of models) with the goal of boosting the fooling rate on the black-box FR model. Finally, we leverage the disentangled nature of latent space in the generative model and incorporate an identity-preserving regularization to effectively maintain the original visual identity. We now present the details of the loss functions used to incorporate the above ideas.", + "bbox": [ + 496, + 90, + 890, + 241 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Textual Loss: A key ingredient of the proposed approach is text-based guidance to inconspicuously hide the adversarial perturbations into the makeup effect. This can be naively achieved by aligning the representation of $t_{\\mathrm{makeup}}$ and the adversarial face $G_{\\theta^*}(\\boldsymbol{w})$ in the common embedding space of a pre-trained vision-language model (e.g. CLIP [47]). However, this approach will transform the whole output image to follow the makeup style of $t_{\\mathrm{makeup}}$ , which results in low diversity. Therefore, we use a directional CLIP loss that aligns the CLIP-space direction between the text-image pairs of the original and adversarial images. Specifically,", + "bbox": [ + 496, + 242, + 892, + 409 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {c l i p}} = 1 - \\frac {\\Delta I \\cdot \\Delta T}{| \\Delta I | | \\Delta T |}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 614, + 421, + 890, + 454 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\Delta T = E_{T}(t_{\\mathrm{makeup}}) - E_{T}(t_{\\mathrm{src}})$ and $\\Delta I = E_I(G_{\\theta^*}(\\pmb {w})) - E_I(\\pmb {x})$ . Here, $E_{T}$ and $E_{I}$ are the text and image encoders of the CLIP model and $t_\\mathrm{src}$ is the semantic text of the input image $\\pmb{x}$ . Since we are dealing with faces, $t_\\mathrm{src}$ can be simply set as \"face\". This loss localizes makeup transfer (e.g. red lipstick) without affecting privacy.", + "bbox": [ + 498, + 464, + 890, + 556 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Adversarial Loss: Our goal is to traverse over the latent space $\\mathcal{W}$ to find adversarial latent codes on the generative manifold whose face feature representation lies close to that of target image and far away from the original image itself i.e., $\\mathcal{D}(\\pmb{x}^p, \\pmb{x}) > \\mathcal{D}(\\pmb{x}^p, \\pmb{x}^t)$ . Hence, the adversarial loss is:", + "bbox": [ + 498, + 556, + 890, + 632 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {a d v}} = \\mathcal {D} \\left(G _ {\\theta^ {*}} (\\boldsymbol {w}), \\boldsymbol {x} ^ {t}\\right) - \\mathcal {D} \\left(G _ {\\theta^ {*}} (\\boldsymbol {w}), \\boldsymbol {x}\\right), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 555, + 643, + 890, + 662 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathcal{D}(\\pmb{x}_1,\\pmb{x}_2) = 1 - \\cos [f(\\pmb{x}_1),f(\\pmb{x}_2)]$ is the cosine distance. Since the malicious FR model is unknown in the black-box setting, Eq. 3 cannot be solved directly. Instead, following AMT-GAN [22], we perform adversarial optimization on an ensemble of white-box surrogate models to imitate the decision boundary of the unknown FR model.", + "bbox": [ + 496, + 672, + 890, + 763 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Identity Preservation Loss: The optimization over the generative manifold ensures that the protected image $\\pmb{x}^p$ is natural i.e., artifact-free, however, it does not explicitly enforce the protected image to preserve the identity of the original image with respect to the human observer. To mitigate the issue, we take advantage of the semantic control exhibited by StyleGAN in its latent space. The latent code $\\pmb{w} \\in \\mathcal{W}$ impacts image generation by controlling different level of semantics in the output image. Specifically, latent", + "bbox": [ + 496, + 763, + 892, + 900 + ], + "page_idx": 4 + }, + { + "type": "footer", + "text": "20599", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "codes corresponding to the initial layers of StyleGAN control high-level aspects such as pose, general hairstyle, and face shape [27]. Adversarially perturbing these latent layers can change these attributes, resulting in a change of identity (see Sec. 4.3). Latent codes corresponding to deeper layers of StyleGAN are associated with fine-level control such as makeup style [2]. Therefore, we perturb only those latent codes associated with deeper layers of StyleGAN, thereby restricting the adversarial faces to the identity preserving manifold. We further constrain the latent code to stay close to its initial value $w_{\\mathrm{inv}}$ using the following regularization:", + "bbox": [ + 75, + 90, + 472, + 257 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {l a t e n t}} = \\left\\| (\\boldsymbol {w} \\odot \\boldsymbol {m} _ {i d}) - \\left(\\boldsymbol {w} _ {\\text {i n v}} \\odot \\boldsymbol {m} _ {i d}\\right) \\right\\| _ {2}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 133, + 268, + 468, + 287 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\odot$ denotes element-wise product and $m_{id}$ is an identity preservation mask that is 0 for the initial layers and 1 only for the deeper layers of the latent code. StyleGAN has 18 layers, each having a dimension of 512. The identity preservation mask is set to 1 only from layer 8 to 18. Finally, combining the three loss functions, we have", + "bbox": [ + 75, + 295, + 468, + 387 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {t o t a l}} = \\lambda_ {\\text {a d v}} \\mathcal {L} _ {\\text {a d v}} + \\lambda_ {\\text {c l i p}} \\mathcal {L} _ {\\text {c l i p}} + \\lambda_ {\\text {l a t e n t}} \\mathcal {L} _ {\\text {l a t e n t}}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 127, + 397, + 468, + 415 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\lambda_{\\mathrm{adv.}}$ , $\\lambda_{\\mathrm{clip}}$ , and $\\lambda_{\\mathrm{latent}}$ are hyperparameters. Note that $\\mathcal{L}_{\\mathrm{adv}}$ accounts for the adversarial objective in Eq. 1, while the text-guided makeup transfer $(\\mathcal{L}_{\\mathrm{clip}})$ and identity-preserving regularization $(\\mathcal{L}_{\\mathrm{latent}})$ implicitly enforce the naturalness constraint in Eq. 1.", + "bbox": [ + 75, + 424, + 468, + 501 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 76, + 512, + 209, + 530 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation details: In all experiments, we use StyleGAN2 pretrained on the FFHQ face dataset as our generative model. For adversarial text guidance, we use a vision transformer-based CLIP model. For generator fine-tuning in the latent code initialization step, we use 450 iterations with value of $\\lambda_{2}$ in Eq. 2 set to 0.5. For the makeup text input, we collect 40 text prompts based on the makeup style of diverse nature (details in supplementary material). For adversarial optimization, we use an Adam optimizer with $\\beta_{1}$ and $\\beta_{2}$ set to 0.9 and 0.999, respectively, and a learning rate of 0.01. We run the optimizer for 50 iterations to craft protected faces. We set the value of $\\lambda_{\\mathrm{adv}}$ , $\\lambda_{\\mathrm{clip}}$ , and $\\lambda_{\\mathrm{latent}}$ to 1, 0.5, and 0.01, respectively. All our experiments are conducted on a A100 GPU with 40 GB memory.", + "bbox": [ + 75, + 537, + 468, + 750 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Datasets: We perform experiments for both face verification and identification settings. Face verification: We use CelebA-HQ [26] and LADN [16] for the impersonation attack. We select subset of 1,000 images from CelebA-HQ and report average results over 4 target identities provided by [22]. Similarly, for LADN, we divide the 332 images available into 4 groups, where images in each group aim to impersonate the target identities provided by [22]. For dodging attack, we use CelebA-HQ [26] and LFW [23] datasets. Specifically, we select 500 subjects at random and", + "bbox": [ + 75, + 750, + 470, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "each subject has a pair of faces. Face identification: For impersonation and dodging, we use CelebA-HQ [26] and LFW [23] as our evaluation set. For both datasets, we randomly select 500 subjects, each with a pair of faces. We assign one image in the pair to the gallery set and the other to the probe set. Both impersonation and dodging attacks are performed on the probe set. For impersonation, we insert 4 target identities provided by [22] into the gallery set. A more detailed description of all datasets and pre-processing steps is provided in the supplementary material.", + "bbox": [ + 496, + 90, + 890, + 242 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Target Models: We aim to protect user facial privacy by attacking four FR model with diverse back bones in the black-box settings. The target models include IRSE50 [21], IR152 [9], FaceNet [51], and MobileFace [5]. Following standard protocol, we align and crop the face images using MTCNN [75] before giving them as input to FR models. Further, we also report privacy protection performance based on commercial FR API including Face++ and Tencent Yunshentu FR platforms.", + "bbox": [ + 496, + 243, + 890, + 378 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation metrics: Following [70], we use protection success rate (PSR) to evaluate the proposed approach. PSR is defined as the fraction of protected faces missclassified by the malicious FR system. To evaluate PSR, we use the thresholding and closed set strategies for face verification and identification, respectively. For face identification, we also use Rank-N targeted identity success rate (Rank-N-T) and untargeted identity success rate (Rank-N-U), where Rank-N-T means that target image $\\boldsymbol{x}^t$ will appear at least once in the top N candidates shortlisted from the gallery and Rank-N-U implies that the top N candidate list does not have the same identity as that of original image $\\boldsymbol{x}$ . We also report results of PSNR (dB), SSIM, and FID [19] scores to evaluate the imperceptibility of method. Large PSNR and SSIM [65] indicates better match with the original images, while low FID score indicates more realistic images. For commercial APIs, we directly report the confidence score returned by the respective servers.", + "bbox": [ + 496, + 380, + 892, + 651 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Baseline methods: We compare our approach with recent noise-based and makeup based facial privacy protection approaches. Noise based methods include PGD [37], MI-FGSM [10], TI-DIM [11], and TIP-IM [70], whereas makeup-based approaches are Adv-Makeup [71] and AMTGAN [22]. We want to highlight that TIP-IM and AMTGAN are considered the state-of-the-art (SOTA) for face privacy protection against black-box FR systems in noise-based and unrestricted settings, respectively. TIP-IM also incorporate multi-target objective in its optimization to find the optimal target image among multiple targets. For fair comparison, we use its single target variant.", + "bbox": [ + 496, + 652, + 890, + 834 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Experimental Results", + "text_level": 1, + "bbox": [ + 500, + 845, + 702, + 861 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this section, we present experimental results of our approach in black-box settings on four different pretrained", + "bbox": [ + 500, + 869, + 892, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "20600", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/e1d3a30df08150fa17b6576d379ee72bdb701195b608b6c5c66bee751053456c.jpg", + "table_caption": [ + "Table 2. Protection success rate (PSR %) of black-box impersonation attack under the face verification task. For each column, the other three FR systems are used as surrogates to generate the protected faces." + ], + "table_footnote": [], + "table_body": "
MethodCelebA-HQLADN-DatasetAverage
IRSE50IR152FaceNetMobileFaceIRSE50IR152FaceNetMobileFace
Clean7.293.801.0812.682.713.610.605.114.61
Inverted5.572.770.6013.326.804.510.2511.665.68
PGD [37]36.8720.681.8543.9940.0919.593.8241.0925.60
MI-FGSM [10]45.7925.032.5845.8548.9025.576.3145.0130.63
TI-DIM [11]63.6336.1715.3057.1256.3634.1822.1148.3041.64
Adv-Makeup(ICAI'21) [71]21.959.481.3722.0029.6410.030.9722.3814.72
TIP-IM(ICCV'21) [70]54.4037.2340.7448.7265.8943.5763.5046.4850.06
AMT-GAN(CVPR'22) [22]76.9635.1316.6250.7189.6449.1232.1372.4352.84
Ours81.1048.4241.7275.2691.5753.3147.9179.9464.90
", + "bbox": [ + 125, + 121, + 846, + 272 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/2dd54fead1c19fbf6364aad49158d50daad247d31dec991723a5d518ec458cab.jpg", + "table_caption": [ + "Table 3. Protection success rate (PSR %) of black-box dodging (top) and impersonation (bottom) attacks under the face identification task for LFW dataset [23]. For each column, the other three FR systems are used as surrogates to generate the protected faces. R1-U: Rank-1-Untargeted, R5-U: Rank-5-Untargeted, R1-T: Rank-1-Targeted, R5-T: Rank-5-Targeted." + ], + "table_footnote": [], + "table_body": "
MethodIRSE50IR152FaceNetMobileFaceAverage
R1-UR5-UR1-UR5-UR1-UR5-UR1-UR5-UR1-UR5-U
MI-FGSM [10]70.242.658.441.859.234.068.047.263.941.4
TI-DIM [11]79.051.267.454.074.452.079.261.675.054.7
TIP-IM(ICCv'21) [70]81.452.271.854.676.049.882.263.077.854.9
Ours86.659.473.456.683.851.285.066.882.258.5
R1-TR5-TR1-TR5-TR1-TR5-TR1-TR5-TR1-TR5-T
MI-FGSM [10]4.010.23.214.29.018.88.422.46.1516.4
TI-DIM [11]4.013.67.819.618.032.821.639.012.8526.25
TIP-IM(ICCv'21) [70]8.028.211.631.225.256.834.051.419.741.9
Ours11.237.816.051.227.454.039.061.223.451.05
", + "bbox": [ + 114, + 324, + 859, + 486 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/104a093182231efba5c599f083417c533c5344c2a1e8d4155c769bf41e514dae.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodFID ↓PSR Gain ↑
Adv-Makeup [71]4.230
TIP-IM [70]38.7335.34
AMT-GAN [22]34.4438.12
Ours26.6250.18
", + "bbox": [ + 93, + 501, + 326, + 579 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 4. FID comparison. PSR Gain is absolute gain in PSR relative to AdvMakeup.", + "bbox": [ + 334, + 501, + 470, + 571 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "FR models under face verification and identification tasks. To generate protected images, we use three FR models as a surrogate to imitate the decision boundary of the fourth FR model. All results are averaged over 5 text based makeup styles that are provided in the supplementary material.", + "bbox": [ + 75, + 598, + 468, + 674 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "For face verification experiments, we set the system threshold value at 0.01 false match rate for each FR model i.e., IRSE50 (0.241), IR152 (0.167), FaceNet (0.409), and MobileFace (0.302). Quantitative results in terms of PSR for impersonation attack under the face verification task are shown in Tab. 2. Our approach is able to achieve an average absolute gain of about $12\\%$ and $14\\%$ over SOTA unrestricted [22] and noise-based [70] facial privacy protection methods, respectively. Qualitative results are shown in Fig. 1 which shows that protected faces generated by our approach are more realistic. Results for dodging attacks under face verification are provided in the supplementary material. In Tab. 3, we also provide PSR vales under the face identification task for dodging (untargeted) and impersonation attacks. Our approach consistently outperforms recent", + "bbox": [ + 75, + 674, + 470, + 901 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "methods at both Rank-1 and Rank-5 settings. We emphasize that we are the first to show effectiveness of generative models in offering untargeted privacy protection (dodging) in a more practical identification setting. Since AMT-GAN and Adv-Makeup are originally trained to impersonate target identity under the verification task, we have not included them in Tab. 3. Qualitative results for LFW and CelebA are provided in the supplementary material.", + "bbox": [ + 496, + 503, + 890, + 625 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We report FID scores (lower is better) of our approach in Tab. 4 for CelebA and LADN datasets to measure naturalness. Adv-Makeup has the lowest FID score as it only applies makeup to the eye region without changing the rest of the face. However, this kind of restriction results in poor PSR. Our method has lower FID compared to TIP-IM and AMT-GAN and achieves the highest PSR. We provide PSNR and SSIM results in the supplementary material.", + "bbox": [ + 496, + 625, + 890, + 748 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. Effectiveness in Real-World Applications", + "text_level": 1, + "bbox": [ + 498, + 756, + 849, + 772 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We now show the effectiveness of our approach to protect facial images (through targeted impersonation) against commercial API such as Face++ and Tencent Yunshentu FR platform operating in the verification mode. These APIs return confidence scores between 0 to 100 to measure whether two images are similar or not, where a high confidence score indicates high similarity. As the training data and model parameters of these propriety FR models are unknown, it", + "bbox": [ + 496, + 780, + 893, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "20601", + "bbox": [ + 478, + 944, + 517, + 957 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/015044b84574475d588ab89a5e52e746f610eca9ca86988dade41c43919170f7.jpg", + "image_caption": [ + "Figure 4. Average confidence score (higher is better) returned by a real-world face verification API, Face++, for impersonation attack. Our approach has a higher confidence score than state-of-the-art makeup and noise-based facial privacy protection methods." + ], + "image_footnote": [], + "bbox": [ + 96, + 88, + 457, + 224 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/a99d2a6abe74e06b9a91cadc7fa9a543dd4983e5519b616b4cc587b067313827.jpg", + "table_caption": [ + "Table 5. Impact of ${\\lambda }_{\\text{latent }}$ on FID score and PSR." + ], + "table_footnote": [], + "table_body": "
λlatent0.50.10.050.010.0050.00010
FID11.621.425.227.830.138.443.2
PSR (%)31.239.057.476.283.890.093.6
", + "bbox": [ + 89, + 325, + 460, + 377 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "effectively mimics a real-world scenario. We protect 100 faces randomly selected from CelebA-HQ using the baselines and the proposed method. In Fig. 4, we show the average confidence score returned by Face++ against these images. These results indicate that our method has a high PSR compared to baselines. We defer more details and results for Tencent Yunshentu API to supplementary material.", + "bbox": [ + 75, + 396, + 468, + 503 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.3. Ablation Studies", + "text_level": 1, + "bbox": [ + 76, + 512, + 243, + 527 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Next, we report some ablations to evaluate the contributions of our loss components.", + "bbox": [ + 75, + 536, + 468, + 566 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Makeup based text guidance: As shown in Fig. 5 (top), in the absence of text guidance, resulting images may contain artifacts due to increased perturbations induced by the adversarial objective. Text-guidance effectively hides the perturbations in the makeup, leading to more natural looking images. It also provides the user more flexibility to select a desired makeup style compared to a reference image.", + "bbox": [ + 75, + 566, + 468, + 672 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Identity preserving regularization: Optimizing over the whole latent space provides more degrees of freedom and increases the PSR. However, it does not explicitly enforce adversarial optimization to preserve the user identity as shown in Fig. 5 (bottom). The proposed identity preserving regularization effectively preserves identity, while imitating the desire makeup style.", + "bbox": [ + 75, + 672, + 470, + 779 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Impact of latent loss weight: Decreasing the weight assigned to the latent loss $\\lambda_{\\mathrm{latent}}$ results in an increase in both the FID score and PSR (and vice versa). Allowing the latent to deviate more from the initial inverted latent code of the given face image often results in artifacts caused by the adversarial loss, degrading naturalness but aiding privacy.", + "bbox": [ + 75, + 779, + 470, + 869 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Robustness against textual variations. Finally, we evaluate the impact of different textual styles on the PSR. We", + "bbox": [ + 76, + 869, + 470, + 900 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/c7f83ace38d139841302d33c12412c2033b4e3fc897af1ce0d93cac392098f4e.jpg", + "image_caption": [ + "Figure 5. Top: Effect of makeup-based text guidance on the visual quality of the output images. Output images are able to impersonate the target identity for face verification. Text-prompt is \"tanned makeup with red lipstick\". Bottom: Optimizing over all latent codes changes the identity of the protected image. Our identity-preserving regularization enforces the adversarial optimization to search for latent codes that hide the perturbations in the makeup effect while simultaneously preserving visual identity." + ], + "image_footnote": [], + "bbox": [ + 553, + 88, + 841, + 343 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/f94ed57a383db61a642c2ef9e3d49e6275dcb92c2aae330a2fa5e1247f74151c.jpg", + "table_caption": [ + "Table 6. Impact of different textual makeup styles on PSR. Makeup styles are \"tanned\", \"pale\", \"pink eyeshadows\", \"red lipstick\", and \"Matte\". Std. denotes standard deviation." + ], + "table_footnote": [], + "table_body": "
t1makeupt2makeupt3makeupt4makeupt5makeupStd.
PSR74.177.378.478.779.21.24
", + "bbox": [ + 517, + 527, + 879, + 570 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "select five text-based makeup styles to protect 1000 images of CelebA-HQ using our method. Results in Tab. 6 show that PSR does not change significantly (low standard deviation) for different makeup styles, indicating robustness of our approach wrt different text-based makeup styles.", + "bbox": [ + 496, + 589, + 893, + 666 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 678, + 619, + 694 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We have proposed a framework to protect privacy of face images on online platforms by carefully searching for adversarial codes in the low-dimensional latent manifold of a pre-trained generative model. We have shown that incorporating a makeup text-guided loss and an identity preserving regularization effectively hides the adversarial perturbations in the makeup style, provides images with high quality, and preserves human-perceived identity. While this approach is robust to the user-defined text-prompt and target identity, it would be beneficial if the text-prompt and target identity can be automatically selected based on the given face image. Limitations of our method include high computational cost at the time of protected face generation.", + "bbox": [ + 496, + 704, + 893, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "20602", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Shane Ahern, Dean Eckles, Nathaniel S Good, Simon King, Mor Naaman, and Rahul Nair. Over-exposed? privacy patterns and considerations in online and mobile photo sharing. In Proceedings of the SIGCHI conference on Human factors in computing systems, pages 357-366, 2007. 2", + "[2] Amit H Bermano, Rinon Gal, Yuval Alaluf, Ron Mokady, Yotam Nitzan, Omer Tov, Oren Patashnik, and Daniel Cohen-Or. State-of-the-art in the architecture, methods and applications of stylegan. In Computer Graphics Forum, volume 41, pages 591–611. Wiley Online Library, 2022. 2, 6", + "[3] Anand Bhattachad, Min Jin Chong, Kaizhao Liang, Bo Li, and David A Forsyth. Unrestricted adversarial examples via semantic manipulation. arXiv preprint arXiv:1904.06347, 2019. 2, 3", + "[4] Jia-Wei Chen, Li-Ju Chen, Chia-Mu Yu, and Chun-Shien Lu. Perceptual indistinguishability-net (pi-net): Facial image obfuscation with manipulable semantics. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6478–6487, 2021. 3", + "[5] Sheng Chen, Yang Liu, Xiang Gao, and Zhen Han. Mobile-facenets: Efficient cnns for accurate real-time face verification on mobile devices. In Chinese Conference on Biometric Recognition, pages 428-438. Springer, 2018. 6", + "[6] Valeriia Cherepanova, Micah Goldblum, Harrison Foley, Shiyuan Duan, John P Dickerson, Gavin Taylor, and Tom Goldstein. Lowkey: Leveraging adversarial attacks to protect social media users from facial recognition. In International Conference on Learning Representations, 2020. 2, 3", + "[7] William L Croft, Jörg-Rüdiger Sack, and Wei Shi. Differentially private facial obfuscation via generative adversarial networks. Future Generation Computer Systems, 129:358-379, 2022. 3", + "[8] Ali Dabouei, Sobhan Soleymani, Jeremy Dawson, and Nasser Nasrabadi. Fast geometrically-perturbed adversarial faces. In 2019 IEEE Winter Conference on Applications of Computer Vision (WACV), pages 1979-1988. IEEE, 2019. 3", + "[9] Jiankang Deng, Jia Guo, Niannan Xue, and Stefanos Zafeiriou. Arcface: Additive angular margin loss for deep face recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4690-4699, 2019. 6", + "[10] Yinpeng Dong, Fangzhou Liao, Tianyu Pang, Hang Su, Jun Zhu, Xiaolin Hu, and Jianguo Li. Boosting adversarial attacks with momentum. In Proceedings of the 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR'18), pages 9185-9193, 2018. 6, 7", + "[11] Yinpeng Dong, Tianyu Pang, Hang Su, and Jun Zhu. Evading defenses to transferable adversarial examples by translation-invariant attacks. In Proceedings of the 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR'19), pages 4312-4321, 2019. 6, 7", + "[12] Hang Du, Hailin Shi, Dan Zeng, Xiao-Ping Zhang, and Tao Mei. The elements of end-to-end deep face recognition: A survey of recent advances. ACM Computing Surveys (CSUR), 54(10s):1-42, 2022. 2" + ], + "bbox": [ + 78, + 114, + 470, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[13] Yifan Du, Zikang Liu, Junyi Li, and Wayne Xin Zhao. A survey of vision-language pre-trained models. arXiv preprint arXiv:2202.10936, 2022. 3", + "[14] Kevin Frans, Lisa B Soros, and Olaf Witkowski. Clipdraw: Exploring text-to-drawing synthesis through language-image encoders. arXiv preprint arXiv:2106.14843, 2021. 3", + "[15] Ian J. Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron C. Courville, and Yoshua Bengio. Generative adversarial networks. CoRR, abs/1406.2661, 2014. 2", + "[16] Qiao Gu, Guanzhi Wang, Mang Tik Chiu, Yu-Wing Tai, and Chi-Keung Tang. Ladin: Local adversarial disentangling network for facial makeup and de-makeup. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10481-10490, 2019. 6", + "[17] Nitzan Guetta, Asaf Shabtai, Inderjeet Singh, Satoru Momiyama, and Yuval Elovici. Dodging attack using carefully crafted natural makeup. arXiv preprint arXiv:2109.06467, 2021. 3", + "[18] Rebecca Heilweil. The world's scariest facial recognition company explained. Vox, May, 8, 2020. 2", + "[19] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 6", + "[20] Kashmir Hill. The secretive company that might end privacy as we know it. *The New York Times*, 18:2020, 2020. 2", + "[21] Jie Hu, Li Shen, and Gang Sun. Squeeze-and-excitation networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7132-7141, 2018. 6", + "[22] Shengshan Hu, Xiaogeng Liu, Yechao Zhang, Minghui Li, Leo Yu Zhang, Hai Jin, and Libing Wu. Protecting facial privacy: Generating adversarial identity masks via style-robust makeup transfer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15014–15023, 2022. 1, 2, 3, 5, 6, 7", + "[23] Gary B Huang, Marwan Mattar, Tamara Berg, and Eric Learned-Miller. Labeled faces in the wild: A database for studying face recognition in unconstrained environments. In Workshop on faces in 'Real-Life' Images: detection, alignment, and recognition, 2008. 6, 7", + "[24] Phillip Isola, Jun-Yan Zhu, Tinghui Zhou, and Alexei A Efros. Image-to-image translation with conditional adversarial networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1125-1134, 2017. 3", + "[25] Kazuya Kakizaki and Kosuke Yoshida. Adversarial image translation: Unrestricted adversarial examples in face recognition systems. arXiv preprint arXiv:1905.03421, 2019. 2, 3", + "[26] Tero Karras, Timo Aila, Samuli Laine, and Jaakko Lehtinen. Progressive growing of gans for improved quality, stability, and variation. In International Conference on Learning Representations, 2018. 6", + "[27] Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of stylegan. In Proceedings of" + ], + "bbox": [ + 501, + 92, + 893, + 901 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "20603", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "the IEEE/CVF conference on computer vision and pattern recognition, pages 8110-8119, 2020. 2, 6", + "[28] Gwanghyun Kim, Taesung Kwon, and Jong Chul Ye. Diffusionclip: Text-guided diffusion models for robust image manipulation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2426-2435, 2022. 3", + "[29] Stepan Komkov and Aleksandr Petiushko. Advhat: Realworld adversarial attack on arcface face id system. In 2020 25th International Conference on Pattern Recognition (ICPR), pages 819-826. IEEE, 2021. 3", + "[30] Zhenzhong Kuang, Zhiqiang Guo, Jinglong Fang, Jun Yu, Noboru Babaguchi, and Jianping Fan. Unnoticeable synthetic face replacement for image privacy protection. Neurocomputing, 457:322-333, 2021. 3", + "[31] Zhenzhong Kuang, Huigui Liu, Jun Yu, Aikui Tian, Lei Wang, Jianping Fan, and Noboru Babaguchi. Effective de-identification generative adversarial network for face anonymization. In Proceedings of the 29th ACM International Conference on Multimedia, pages 3182–3191, 2021. 3", + "[32] Gihyun Kwon and Jong Chul Ye. Clipstyler: Image style transfer with a single text condition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18062-18071, 2022. 3", + "[33] Tao Li and Min Soo Choi. Deepblur: A simple and effective method for natural image obfuscation. arXiv preprint arXiv:2104.02655, 1, 2021. 3", + "[34] Tao Li and Lei Lin. Anonymousnet: Natural face de-identification with measurable privacy. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops, pages 0–0, 2019. 3", + "[35] Fangcheng Liu, Chao Zhang, and Hongyang Zhang. Towards transferable unrestricted adversarial examples with minimum changes. arXiv preprint arXiv:2201.01102, 2022. 3", + "[36] Suolan Liu, Lizhi Kong, and Hongyuan Wang. Face detection and encryption for privacy preserving in surveillance video. In Chinese Conference on Pattern Recognition and Computer Vision (PRCV), pages 162-172. Springer, 2018. 3", + "[37] Aleksander Madry, Aleksandar Makelov, Ludwig Schmidt, Dimitris Tsipras, and Adrian Vladu. Towards deep learning models resistant to adversarial attacks. In Proceedings of the 6th International Conference on Learning Representations (ICLR'18), 2018. 6, 7", + "[38] Blaz Meden, Peter Rot, Philipp Terhörst, Naser Damer, Arjan Kuijper, Walter J Scheirer, Arun Ross, Peter Peer, and Vitomir Štruc. Privacy-enhancing face biometrics: A comprehensive survey. IEEE Transactions on Information Forensics and Security, 2021. 2, 3", + "[39] Dongbin Na, Sangwoo Ji, and Jong Kim. Unrestricted blackbox adversarial attack using gan with limited queries. arXiv preprint arXiv:2208.11613, 2022. 2", + "[40] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741, 2021. 3" + ], + "bbox": [ + 78, + 92, + 468, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[41] Seong Joon Oh, Rodrigo Benenson, Mario Fritz, and Bernt Schiele. Faceless person recognition: Privacy implications in social media. In European Conference on Computer Vision, pages 19-35. Springer, 2016. 3", + "[42] Seong Joon Oh, Mario Fritz, and Bernt Schiele. Adversarial image perturbation for privacy protection a game theory perspective. In 2017 IEEE International Conference on Computer Vision (ICCV), pages 1491-1500. IEEE, 2017. 3", + "[43] Omkar M Parkhi, Andrea Vedaldi, and Andrew Zisserman. Deep face recognition. 2015. 1", + "[44] Or Patashnik, Zongze Wu, Eli Shechtman, Daniel Cohen-Or, and Dani Lischinski. Styleclip: Text-driven manipulation of stylegan imagery. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2085–2094, 2021. 3", + "[45] P Jonathon Phillips, Amy N Yates, Ying Hu, Carina A Hahn, Eilidh Noyes, Kelsey Jackson, Jacqueline G Cava-zos, Geraldine Jeckeln, Rajeev Ranjan, Swami Sankaranarayanan, et al. Face recognition accuracy of forensic examiners, superrecognizers, and face recognition algorithms. Proceedings of the National Academy of Sciences, 115(24):6171-6176, 2018. 2", + "[46] Omid Poursaeed, Tianxing Jiang, Harry Yang, Serge Belongie, and Ser-Nam Lim. Robustness and generalization via generative adversarial training. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15711-15720, 2021. 3", + "[47] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 3, 5", + "[48] Rajeev Ranjan, Swami Sankaranarayanan, Ankan Bansal, Navaneeth Bodla, Jun-Cheng Chen, Vishal M Patel, Carlos D Castillo, and Rama Chellappa. Deep learning for understanding faces: Machines may be just as good, or better, than humans. IEEE Signal Processing Magazine, 35(1):66-83, 2018. 2", + "[49] Daniel Roich, Ron Mokady, Amit H Bermano, and Daniel Cohen-Or. Pivotal tuning for latent-based editing of real images. ACM Transactions on Graphics (TOG), 42(1):1-13, 2022. 5", + "[50] Tim Salimans, Ian Goodfellow, Wojciech Zaremba, Vicki Cheung, Alec Radford, and Xi Chen. Improved techniques for training gans. Advances in neural information processing systems, 29, 2016. 3", + "[51] Florian Schroff, Dmitry Kalenichenko, and James Philbin. Facenet: A unified embedding for face recognition and clustering. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 815-823, 2015. 6", + "[52] Sachith Seneviratne, Nuran Kasthuriarachchi, Sanka Rasnayaka, Danula Hettiachchi, and Ridwan Shariffdeen. Does a face mask protect my privacy?: Deep learning to predict protected attributes from masked face images. In Australasian Joint Conference on Artificial Intelligence, pages 91-102. Springer, 2022. 3" + ], + "bbox": [ + 501, + 92, + 890, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "20604", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[53] Shawn Shan, Emily Wenger, Jiayun Zhang, Huiying Li, Haitao Zheng, and Ben Y Zhao. Fawkes: Protecting privacy against unauthorized deep learning models. In 29th USENIX security symposium (USENIX Security 20), pages 1589-1604, 2020. 2, 3", + "[54] Mahmood Sharif, Sruti Bhagavatula, Lujo Bauer, and Michael K Reiter. A general framework for adversarial examples with objectives. ACM Transactions on Privacy and Security (TOPS), 22(3):1-30, 2019. 3", + "[55] Yang Song, Rui Shu, Nate Kushman, and Stefano Ermon. Constructing unrestricted adversarial examples with generative models. Advances in Neural Information Processing Systems, 31, 2018. 2, 3", + "[56] Qianru Sun, Liqian Ma, Seong Joon Oh, Luc Van Gool, Bernt Schiele, and Mario Fritz. Natural and effective obfuscation by head inpainting. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 5050-5059, 2018. 3", + "[57] Christian Szegedy, Wojciech Zaremba, Ilya Sutskever, Joan Bruna, Dumitru Erhan, Ian Goodfellow, and Rob Fergus. Intriguing properties of neural networks. In International Conference on Learning Representations, 2014. 2", + "[58] Huan Tian, Tianqing Zhu, and Wanlei Zhou. Fairness and privacy preservation for facial images: Gan-based methods. Computers & Security, 122:102902, 2022. 3", + "[59] Omer Tov, Yuval Alaluf, Yotam Nitzan, Or Patashnik, and Daniel Cohen-Or. Designing an encoder for stylegan image manipulation. ACM Transactions on Graphics (TOG), 40(4):1-14, 2021. 5", + "[60] Can Wang, Menglei Chai, Mingming He, Dongdong Chen, and Jing Liao. Clip-nerf: Text-and-image driven manipulation of neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3835-3844, 2022. 3", + "[61] Mei Wang and Weihong Deng. Deep face recognition: A survey. Neurocomputing, 429:215-244, 2021. 1, 2", + "[62] Shunxin Wang, Una M Kelly, and Raymond NJ Veldhuis. Gender obfuscation through face morphing. In 2021 IEEE International Workshop on Biometrics and Forensics (IWBF), pages 1-6. IEEE, 2021. 3", + "[63] Ya Wang, Tianlong Bao, Chunhui Ding, and Ming Zhu. Face recognition in real-world surveillance videos with deep learning method. In 2017 2nd international conference on image, vision and computing (icivc), pages 239-243. IEEE, 2017. 2", + "[64] Yinggui Wang, Jian Liu, Man Luo, Le Yang, and Li Wang. Privacy-preserving face recognition in the frequency domain. 2022. 3", + "[65] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600-612, 2004. 6", + "[66] Tianyi Wei, Dongdong Chen, Wenbo Zhou, Jing Liao, Zhentao Tan, Lu Yuan, Weiming Zhang, and Nenghai Yu. Hairclip: Design your hair by text and reference image. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18072-18081, 2022. 3" + ], + "bbox": [ + 78, + 90, + 470, + 900 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[67] Emily Wenger, Shawn Shan, Haitao Zheng, and Ben Y Zhao. Sok: Anti-facial recognition technology. arXiv preprint arXiv:2112.04558, 2021. 2", + "[68] Chaowei Xiao, Jun-Yan Zhu, Bo Li, Warren He, Mingyan Liu, and Dawn Song. Spatially transformed adversarial examples. arXiv preprint arXiv:1801.02612, 2018. 2, 3", + "[69] Zihao Xiao, Xianfeng Gao, Chilin Fu, Yinpeng Dong, Wei Gao, Xiaolu Zhang, Jun Zhou, and Jun Zhu. Improving transferability of adversarial patches on face recognition with generative models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11845-11854, 2021. 3", + "[70] Xiao Yang, Yinpeng Dong, Tianyu Pang, Hang Su, Jun Zhu, Yuefeng Chen, and Hui Xue. Towards face encryption by generating adversarial identity masks. In Proceedings of the 2021 IEEE/CVF International Conference on Computer Vision (ICCV'21), pages 3897-3907, 2021. 1, 2, 3, 6, 7", + "[71] Bangjie Yin, Wenxuan Wang, Taiping Yao, Junfeng Guo, Zelun Kong, Shouhong Ding, Jilin Li, and Cong Liu. Advmakeup: A new imperceptible and transferable attack on face recognition. In Proceedings of the 30th International Joint Conference on Artificial Intelligence (IJCAI'21), pages 1252-1258, 2021. 2, 6, 7", + "[72] Bangjie Yin, Wenxuan Wang, Taiping Yao, Junfeng Guo, Zelun Kong, Shouhong Ding, Jilin Li, and Cong Liu. Advmakeup: A new imperceptible and transferable attack on face recognition. arXiv preprint arXiv:2105.03162, 2021. 2, 3", + "[73] Shengming Yuan, Qilong Zhang, Lianli Gao, Yaya Cheng, and Jingkuan Song. Natural color fool: Towards boosting black-box unrestricted attacks. arXiv preprint arXiv:2210.02041, 2022. 3", + "[74] Jiaming Zhang, Jitao Sang, Xian Zhao, Xiaowen Huang, Yanfeng Sun, and Yongli Hu. Adversarial privacy-preserving filter. In Proceedings of the 28th ACM International Conference on Multimedia, pages 1423-1431, 2020. 2", + "[75] Kaipeng Zhang, Zhanpeng Zhang, Zhifeng Li, and Yu Qiao. Joint face detection and alignment using multitask cascaded convolutional networks. IEEE signal processing letters, 23(10):1499-1503, 2016. 6", + "[76] Zhengyu Zhao, Zhuoran Liu, and Martha Larson. Towards large yet imperceptible adversarial image perturbations with perceptual color distance. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1039-1048, 2020. 2, 3", + "[77] Yaoyao Zhong and Weihong Deng. Opom: Customized invisible cloak towards face privacy protection. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2022, 2, 3", + "[78] Jizhe Zhou and Chi-Man Pun. Personal privacy protection via irrelevant faces tracking and pixelation in video live streaming. IEEE Transactions on Information Forensics and Security, 16:1088-1103, 2020. 3", + "[79] Zheng-An Zhu, Yun-Zhong Lu, and Chen-Kuo Chiang. Generating adversarial examples by makeup attacks on face recognition. In 2019 IEEE International Conference on Image Processing (ICIP), pages 2516-2520. IEEE, 2019. 3" + ], + "bbox": [ + 501, + 92, + 890, + 882 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "20605", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2023/CLIP2Protect_ Protecting Facial Privacy Using Text-Guided Makeup via Adversarial Latent Search/d379e97e-b7a0-42bb-a735-a3de89cdb054_model.json b/2023/CLIP2Protect_ Protecting Facial Privacy Using Text-Guided Makeup via Adversarial Latent Search/d379e97e-b7a0-42bb-a735-a3de89cdb054_model.json new file mode 100644 index 0000000000000000000000000000000000000000..0daafb2e1f19c89068ea6497574a714575dd2887 --- /dev/null +++ b/2023/CLIP2Protect_ Protecting Facial Privacy Using Text-Guided Makeup via Adversarial Latent Search/d379e97e-b7a0-42bb-a735-a3de89cdb054_model.json @@ -0,0 +1,2378 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.0, + 0.812, + 0.045 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.122, + 0.121, + 0.85, + 0.164 + ], + "angle": 0, + "content": "CLIP2Protect: Protecting Facial Privacy using Text-Guided Makeup via Adversarial Latent Search" + }, + { + "type": "text", + "bbox": [ + 0.235, + 0.186, + 0.735, + 0.221 + ], + "angle": 0, + "content": "Fahad Shamshad Muzammal Naseer Karthik Nandakumar Mohamed Bin Zayed University of AI, UAE" + }, + { + "type": "text", + "bbox": [ + 0.191, + 0.224, + 0.773, + 0.239 + ], + "angle": 0, + "content": "{fahad.shamshad, muzammal.naseer, karthik.nandakumar}@mbzuai.ac.ae" + }, + { + "type": "image_caption", + "bbox": [ + 0.087, + 0.295, + 0.102, + 0.424 + ], + "angle": 270, + "content": "Protected Original" + }, + { + "type": "image", + "bbox": [ + 0.104, + 0.271, + 0.885, + 0.469 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.104, + 0.475, + 0.885, + 0.565 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.568, + 0.812, + 0.653 + ], + "angle": 0, + "content": "Figure 1. The proposed approach crafts \"naturalistic\" and transferable text-guided adversarial faces to deceive black-box face recognition systems. First row shows original images that need to be protected and second row shows corresponding protected images along with the user-defined makeup text prompts that guide the adversarial search. Comparison against existing methods is shown in the third row. The yellow text represents the confidence score (higher is better) output by a commercial API (Face++, when matching the protected image against the target identity shown in the bottom right. The reference image used by [22] for makeup transfer is shown at the bottom corner of the corresponding adversarial image." + }, + { + "type": "image", + "bbox": [ + 0.813, + 0.571, + 0.884, + 0.626 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.825, + 0.627, + 0.865, + 0.638 + ], + "angle": 0, + "content": "Target" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.665, + 0.314, + 0.68 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.705, + 0.471, + 0.901 + ], + "angle": 0, + "content": "The success of deep learning based face recognition systems has given rise to serious privacy concerns due to their ability to enable unauthorized tracking of users in the digital world. Existing methods for enhancing privacy fail to generate \"naturalistic\" images that can protect facial privacy without compromising user experience. We propose a novel two-step approach for facial privacy protection that relies on finding adversarial latent codes in the low-dimensional manifold of a pretrained generative model. The first step inverts the given face image into the latent space and finetunes the generative model to achieve an accurate reconstruction of the given image from its latent code. This step produces a good initialization, aiding the generation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.666, + 0.892, + 0.832 + ], + "angle": 0, + "content": "of high-quality faces that resemble the given identity. Subsequently, user-defined makeup text prompts and identity-preserving regularization are used to guide the search for adversarial codes in the latent space. Extensive experiments demonstrate that faces generated by our approach have stronger black-box transferability with an absolute gain of \\(12.06\\%\\) over the state-of-the-art facial privacy protection approach under the face verification task. Finally, we demonstrate the effectiveness of the proposed approach for commercial face recognition systems. Our code is available at https://github.com/fahadshamshad/Clip2Protect." + }, + { + "type": "title", + "bbox": [ + 0.502, + 0.844, + 0.631, + 0.858 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Deep learning based face recognition (FR) systems [43, 61] have found widespread usage in multiple applications," + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "20595" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.09, + 0.472, + 0.146 + ], + "angle": 0, + "content": "Table 1. Comparison among different facial privacy protection methods w.r.t. the natural outputs, black box setting, experiments under face verification and identification tasks, unrestricted (semantically meaningful), and more flexible text guided adversaries." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.15, + 0.47, + 0.246 + ], + "angle": 0, + "content": "
Adv-Makeup [71]TIP-IM [70]AMT-GAN [22]Ours
Natural outputsYesPartiallyPartiallyYes
Black boxYesYesYesYes
VerificationYesNoYesYes
IdentificationNoYesNoYes
UnrestrictedYesNoYesYes
Text guidedNoNoNoYes
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.265, + 0.47, + 0.476 + ], + "angle": 0, + "content": "including security [63], biometrics [38], and criminal investigation [45], outperforming humans in many scenarios [12, 48, 61]. Despite positive aspects of this technology, FR systems seriously threaten personal security and privacy in the digital world because of their potential to enable mass surveillance capabilities [1, 67]. For example, government and private entities can use FR systems to track user relationships and activities by scraping face images from social media profiles such as Twitter, LinkedIn, and Facebook [18, 20]. These entities generally use proprietary FR systems, whose specifications are unknown to the public (black box model). Therefore, there is an urgent need for an effective approach that protects facial privacy against such unknown FR systems." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.478, + 0.47, + 0.688 + ], + "angle": 0, + "content": "An ideal facial privacy protection algorithm must strike the right balance between naturalness and privacy protection [70, 77]. In this context, \"naturalness\" is defined as the absence of any noise artifacts that can be easily perceived by human observers and the preservation of human-perceived identity. \"Privacy protection\" refers to the fact that the protected image must be capable of deceiving a black-box malicious FR system. In other words, the protected image must closely resemble the given face image and be artifact-free for a human observer, while at the same time fool an unknown automated FR system. Since failure to generate naturalistic faces can significantly affect user experience on social media platforms, it is a necessary precondition for adoption of a privacy-enhancement algorithm." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.689, + 0.47, + 0.84 + ], + "angle": 0, + "content": "Recent works exploit adversarial attacks [57] to conceal user identity by overlaying noise-constrained (bounded) adversarial perturbations on the original face image [6,53,74]. Since the adversarial examples are generally optimized in the image space, it is often difficult to simultaneously achieve naturalness and privacy [70]. Unlike noise-based methods, unrestricted adversarial examples are not constrained by the magnitude of perturbation in the image space and have demonstrated better perceptual realism for human observers while being adversially effective [3,55,68,76]." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.84, + 0.47, + 0.902 + ], + "angle": 0, + "content": "Several efforts have been made to generate unrestricted adversarial examples that mislead FR systems (see Tab. 1) [22, 25, 39, 72]. Among these, adversarial makeup based methods [22, 72] are gaining increasing attention as they" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.348 + ], + "angle": 0, + "content": "can embed adversarial modifications in a more natural way. These approaches use generative adversarial networks [15] (GANs) to adversarily transfer makeup from a given reference image to the user's face image while impersonating a target identity. However, existing techniques based on adversarial makeup transfer have the following limitations: (i) adversarial toxicity in these methods hamper the performance of the makeup transfer module, thereby resulting in unnatural faces with makeup artifacts (see Fig. 1); (ii) the use of a reference image to define the desired makeup style affects the practicality of this approach; (iii) for every new target identity, these approaches require end-to-end retraining from scratch using large makeup datasets; and (iv) most of these methods primarily aim at impersonation of the target identity, whereas the desired privacy objective is dodging, i.e., multiple images of the user's face scraped from different social media sites must not match with each other." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.35, + 0.892, + 0.442 + ], + "angle": 0, + "content": "To mitigate the above problems, we propose a new approach to protect user facial privacy on online platforms (Sec. 3). The proposed approach aims to search for adversarial latent codes in a low-dimensional manifold learned by a generative model trained to generate face images [2, 27]. Our main contributions are:" + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.457, + 0.892, + 0.563 + ], + "angle": 0, + "content": "- Facial Privacy-protection Framework Using Adversarial Latent Codes: Given a face image, we propose a novel two-step method to search for adversarial latent codes, which can be used by a generative model (e.g., StyleGAN) to produce face images with high visual quality that matches human-perceived identity, while deceiving black-box FR systems." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.573, + 0.892, + 0.709 + ], + "angle": 0, + "content": "- Adversarial Makeup Transfer using Textual Prompts: A critical component of the above framework is a technique for leveraging user-defined textual (makeup) prompts to traverse over the latent manifold of the generative model and find transferable adversarial latent codes. Our approach effectively hides attack information in the desired makeup style, without the need for any large makeup dataset or retraining of models for different target identities." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.718, + 0.892, + 0.794 + ], + "angle": 0, + "content": "- Identity Preserving Regularization: We propose a regularizer that preserves identity-related attributes within the latent space of the generative model and ensures that the protected face image visually resembles the original face." + }, + { + "type": "list", + "bbox": [ + 0.518, + 0.457, + 0.892, + 0.794 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.81, + 0.892, + 0.902 + ], + "angle": 0, + "content": "Extensive experiments (Sec. 4.1) for both face verification and identification scenarios demonstrate the effectiveness of our approach against black-box FR models and online commercial facial recognition APIs (Sec. 4.2). Furthermore, we provide detailed ablative analysis to dissect the performance of different components of our approach (Sec. 4.3)." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "20596" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.078, + 0.09, + 0.22, + 0.106 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.115, + 0.47, + 0.312 + ], + "angle": 0, + "content": "Obfuscation Methods: Obfuscation is the most widely used technique [38] to protect user's facial privacy. Earlier obfuscation approaches typically degrade the quality of the original face image by applying simple operations such as masking [52, 64], filtering [33, 78], and image transformations [8, 36, 62]. While these relatively simple obfuscation techniques are reasonable for surveillance applications, they are ill-suited for online/social media platforms where user experience is critical [41]. Though deep learning based obfuscation approaches generate more realistic images [4, 7, 56, 58], they often result in a change of identity compared to the original image and occasionally produce undesirable artifacts [30, 31, 34]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.312, + 0.471, + 0.553 + ], + "angle": 0, + "content": "Noise-based Adversarial Examples: Adversarial attacks have been used to protect users from unauthorized FR models. Some methods [6, 53] rely on data poisoning to deceive targeted FR models, but are less practical because access to the training data or the gallery set of the unknown FR system is often not available. Other approaches have used game-theory perspective [42] in white-box settings or person-specific privacy masks (one mask per person) to generate protected images at the cost of acquiring multiple images of the same user [77]. In contrast, we aim to fool the black box FR model using only single image. In TIP-IM [70], targeted optimization was used to generate privacy masks against unknown FR models by introducing a naturalness constraint. While this approach provides effective privacy, it generates output images with perceptible noises that can affect the user experience [70]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.553, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Unrestricted Adversarial Examples: Unrestricted adversarial attacks (UAAs) are not constrained by the perturbation norm and can induce large but semantically meaningful perturbations. These attacks have been extensively studied in image classification literature [3, 35, 55, 68, 73, 76] and it has been shown that outputs generated via UAAs are less perceptible to human observers as compared to noise-based adversarial attacks. Motivated by this observation, patch-based unrestricted attacks have been proposed to generate wearable adversarial accessories like colorful glasses [54], hat [29] or random patch [69] to fool the FR model, but such synthesized patches generally have weak transferability due to the limited editing region and the large visible pattern compromises naturalness and affects user experience. Recently, generative models [24, 50] have been leveraged to craft UAAs against FR models. However, these generative approaches are either designed for the white-box settings [46, 79] or show limited performance in query-free black-box settings [25]. Makeup-based UAAs [17, 72] have also been proposed against FR systems by embedding the perturbations into a natural makeup effect. These makeup based attacks have also been exploited to protect the user privacy by applying adversarial makeup on the user" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.227 + ], + "angle": 0, + "content": "face image [22]. However, interference between adversarial perturbations and makeup transfer can produce undesirable makeup artifacts in the output images. Moreover, these attacks generally assume access to large makeup datasets for training models and require a reference makeup image. In contrast, our approach finds adversarial faces on the natural image manifold in black-box setting via guidance from makeup text prompt, which makes it less susceptible to artifacts (see Fig. 1) and more practical." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.228, + 0.893, + 0.47 + ], + "angle": 0, + "content": "Vision-Language Modelling: Cross-modal vision-language modelling has attracted significant attention in recent years [13]. OpenAI introduced CLIP [47] that is trained on 400 million image-text pairs using contrastive objective and maps both image and text in a joint multimodal embedding space. With powerful representation embedding of CLIP, several methods have been proposed to manipulate images with text-guidance. StyleCLIP [44] and DiffusionCLIP [28, 40] leverage the powerful generative capabilities of StyleGAN and diffusion models to manipulate images with text prompts. Other similar works include HairCLIP [66], CLIP-NeRF [60], CLIPstyler [32], and CLIPDraw [14]. While these methods focus on the text-guidance ability of CLIP, our approach aims to find the adversarial latent codes in a generative model's latent space for privacy protection against black-box FR models." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.482, + 0.844, + 0.5 + ], + "angle": 0, + "content": "3. Proposed Approach for Facial Privacy" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.507, + 0.892, + 0.75 + ], + "angle": 0, + "content": "Our goal is to protect user facial privacy on online platforms against unknown (black-box) FR models without compromising on the user's online experience. The proposed approach finds protected faces by adversarially exploring the low-dimensional latent space of a pretrained generative model that is trained on natural face images. To avoid artifacts in the protected image, we restrict the search for adversarial faces close to the clean image manifold learned by the generative model. Moreover, we propose to optimize only over identity-preserving latent codes in the latent space. This effectively preserves human-perceived identity during attack while offering high privacy against automated systems. Further, we employ natural makeup-like perturbations via guidance from a text prompt, which provides more flexibility to the user compared to reference image-based adversarial makeup transfer [22]." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.757, + 0.642, + 0.772 + ], + "angle": 0, + "content": "3.1. Preliminaries" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.78, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Let \\( \\pmb{x} \\in \\mathcal{X} \\subset \\mathbb{R}^n \\) denote the given original/real face image. Let \\( f(\\pmb{x}) : \\mathcal{X} \\to \\mathbb{R}^d \\) be a FR model that extracts a fixed-length normalized feature representation. Let \\( \\mathcal{D}(\\pmb{x}_1, \\pmb{x}_2) = D(f(\\pmb{x}_1), f(\\pmb{x}_2)) \\) be a distance metric that measures the dissimilarity between two face images \\( \\pmb{x}_1 \\) and \\( \\pmb{x}_2 \\) based on their respective representations \\( f(\\pmb{x}_1) \\) and \\( f(\\pmb{x}_2) \\). Generally a FR system can operate in two modes: verification and identification. A face verification" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "20597" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.101, + 0.087, + 0.87, + 0.305 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.315, + 0.896, + 0.375 + ], + "angle": 0, + "content": "Figure 2. Overall pipeline of the proposed approach to protect users facial privacy. Our proposed approach searches for the adversarial latent codes on the generative manifold to reconstruct an adversarial face that is capable of fooling unknown FR systems for privacy protection. Our approach allows \"makeup\" editing in an adversarial manner through user defined textual prompts and thereby enhance the user's online experience. Our text-guided objective searches for such latent codes while keeping the original identity preserved." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.397, + 0.47, + 0.533 + ], + "angle": 0, + "content": "system predicts that two faces belong to the same identity if \\(\\mathcal{D}(\\pmb{x}_1, \\pmb{x}_2) \\leq \\tau\\), where \\(\\tau\\) is the system threshold. On the other hand, a (closed set) face identification system compares the input image (probe) against a set of face images (gallery) and outputs the identity whose representation is most similar to that of the probe. Since the attacker can employ verification or identification to determine the user identity using black-box FR models, a protection approach should conceal the user's identity in both scenarios." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.534, + 0.471, + 0.7 + ], + "angle": 0, + "content": "User privacy can be protected by misleading the malicious FR model through impersonation or dodging attacks. In the context of verification, impersonation (false match) implies that the protected face matches with the face of a specific target identity and dodging (false non-match) means that the protected face does not match with some other image of the same person. Similarly, for face identification, impersonation ensures that the protected image gets matched to a specified target identity in the gallery set, while dodging prevents the protected face from matching with images of the same person in the gallery." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.7, + 0.47, + 0.85 + ], + "angle": 0, + "content": "Problem Statement: Given the original face image \\( \\pmb{x} \\), our goal is to generate a protected face image \\( \\pmb{x}^p \\) such that \\( \\mathcal{D}(\\pmb{x}^p, \\pmb{x}) \\) is large (for successful dodging attack) and \\( \\mathcal{D}(\\pmb{x}^p, \\pmb{x}^t) \\) is small (for successfully impersonating a target face \\( \\pmb{x}^t \\)), where \\( \\mathcal{O}(\\pmb{x}) \\neq \\mathcal{O}(\\pmb{x}^t) \\) and \\( \\mathcal{O} \\) is the oracle that gives the true identity labels. At the same time, we want to minimize \\( \\mathcal{H}(\\pmb{x}^p, \\pmb{x}) \\), where \\( \\mathcal{H} \\) quantifies the degree of unnaturalness introduced in the protected image \\( \\pmb{x}^p \\) in relation to the original image \\( \\pmb{x} \\). Formally, the optimization problem that we aim to solve is:" + }, + { + "type": "equation", + "bbox": [ + 0.148, + 0.855, + 0.47, + 0.897 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\min _ {\\boldsymbol {x} ^ {p}} \\mathcal {L} (\\boldsymbol {x} ^ {p}) = \\mathcal {D} (\\boldsymbol {x} ^ {p}, \\boldsymbol {x} ^ {t}) - \\mathcal {D} (\\boldsymbol {x} ^ {p}, \\boldsymbol {x}) \\tag {1} \\\\ \\mathrm {s . t .} \\mathcal {H} (\\boldsymbol {x} ^ {p}, \\boldsymbol {x}) \\leq \\epsilon \\\\ \\end{array}\n\\]" + }, + { + "type": "image", + "bbox": [ + 0.502, + 0.395, + 0.63, + 0.494 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.534, + 0.496, + 0.598, + 0.508 + ], + "angle": 0, + "content": "(a) Original" + }, + { + "type": "image", + "bbox": [ + 0.633, + 0.395, + 0.761, + 0.494 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.638, + 0.496, + 0.754, + 0.507 + ], + "angle": 0, + "content": "(b) Encoder Inversion" + }, + { + "type": "image", + "bbox": [ + 0.763, + 0.395, + 0.892, + 0.494 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.764, + 0.496, + 0.891, + 0.508 + ], + "angle": 0, + "content": "(c) Generator finetuning" + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.511, + 0.895, + 0.568 + ], + "angle": 0, + "content": "Figure 3. Generator finetuning allows near-perfect reconstructions of LFW dataset sample. This is crucial for the online experience of users. Matching scores returned by Face++ API are 62.38 and 98.96 for encoder and generator-finetuned inversions, respectively." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.594, + 0.892, + 0.73 + ], + "angle": 0, + "content": "where \\(\\epsilon\\) is a bound on the adversarial perturbation. For noise-based approach, \\(\\mathcal{H}(\\pmb{x}^p,\\pmb{x}) = \\| \\pmb{x} - \\pmb{x}^p\\|_p\\), where \\(\\| \\cdot \\|_p\\) denotes the \\(L_p\\) norm. However, direct enforcement of the perturbation constraint leads to visible artifacts, which affects visual quality and user experience. Constraining the solution search space to a natural image manifold using an effective image prior can produce more realistic images. Note that the distance metric \\(\\mathcal{D}\\) is unknown since our goal is to deceive a black-box FR system." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.741, + 0.843, + 0.758 + ], + "angle": 0, + "content": "3.2. Makeup Text-Guided Adversarial Faces" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.892, + 0.903 + ], + "angle": 0, + "content": "Our approach restricts the solution space of the protected face \\( \\pmb{x}^p \\) to lie close to the clean face manifold \\( \\mathcal{X} \\). This manifold can be learned using a generative model trained on real human faces. Specifically, let \\( G_{\\theta}(\\pmb{w}) : \\mathcal{W} \\to \\mathbb{R}^n \\) denote the pretrained generative model with weights \\( \\theta \\), where \\( \\mathcal{W} \\) is the latent space. Our proposed approach consists of two stages: (i) latent code initialization (Sec. 3.2.1) and (ii) text-guided adversarial optimization (Sec. 3.2.2). The overall pipeline of the proposed approach is shown in Fig. 2." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20598" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.077, + 0.092, + 0.313, + 0.106 + ], + "angle": 0, + "content": "3.2.1 Latent Code Initialization" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.116, + 0.469, + 0.222 + ], + "angle": 0, + "content": "The latent code initialization stage is based on GAN inversion, which aims to invert the original image \\(\\pmb{x}\\) into the latent space \\(\\mathcal{W}\\), i.e., find a latent code \\(\\pmb{w}_{\\mathrm{inv}} \\in \\mathcal{W}\\) such that \\(\\pmb{x}_{\\mathrm{inv}} = G_{\\theta}(\\pmb{w}_{\\mathrm{inv}}) \\approx \\pmb{x}\\). To achieve this, we first use an encoder-based inversion called e4e [59] to infer \\(\\pmb{w}_{\\mathrm{inv}}\\) in \\(\\mathcal{W}\\) from \\(\\pmb{x}\\) i.e., \\(\\pmb{w}_{\\mathrm{inv}} = I_{\\phi}(\\pmb{x})\\), where \\(I_{\\phi}: \\mathcal{X} \\to \\mathcal{W}\\) is the pretrained encoder with weights \\(\\phi\\) (see Fig. 2)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.222, + 0.471, + 0.495 + ], + "angle": 0, + "content": "We use StyleGAN trained on a high-resolution dataset of face images as the pretrained generative model \\( G_{\\theta} \\) due to its powerful synthesis ability and the disentangled structure of its latent space. A significant challenge during inversion is preserving the identity of the original image i.e., \\( \\mathcal{O}(\\pmb{x}) = \\mathcal{O}(\\pmb{x}_{\\mathrm{inv}}) \\). Generally, optimization and encoder-based inversion approaches struggle to preserve identity after reconstruction [49] (see Fig. 3b). Moreover, when using these approaches, the inversion error can be large for out-of-domain face images with extreme poses and viewpoints, which are quite common in social media applications. Therefore, these approaches cannot be applied directly to invert \\( \\pmb{x} \\). Instead, motivated by the recent observation [49] that slight changes to the pretrained generator weights do not harm its editing abilities while achieving near-perfect reconstructions, we finetune the pretrained generator weights \\( \\theta \\) instead of the encoder weights \\( \\phi \\). Specifically, we fix \\( w_{\\mathrm{inv}} = I_{\\phi}(\\pmb{x}) \\) and fine-tune \\( G_{\\theta} \\) using the following loss:" + }, + { + "type": "equation", + "bbox": [ + 0.093, + 0.505, + 0.454, + 0.528 + ], + "angle": 0, + "content": "\\[\n\\theta^ {*} = \\underset {\\theta} {\\arg \\min } \\mathcal {L} _ {\\mathrm {L P I P S}} (\\boldsymbol {x}, G _ {\\theta} (\\boldsymbol {w} _ {\\mathrm {i n v}})) + \\lambda_ {2} \\mathcal {L} _ {2} (\\boldsymbol {x}, G _ {\\theta} (\\boldsymbol {w} _ {\\mathrm {i n v}})),\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.537, + 0.47, + 0.6 + ], + "angle": 0, + "content": "where \\(\\mathcal{L}_{\\mathrm{LPIPS}}\\) is the perceptual loss and \\(\\mathcal{L}_2\\) denotes the pixelwise similarity. The final inverted image \\(x_{\\mathrm{inv}}^*\\) (see Fig. 3c) can be obtained by performing a forward pass of \\(\\pmb{w}_{\\mathrm{inv}}\\) through fine-tuned generator i.e., \\(\\pmb{x}_{\\mathrm{inv}}^{*} = G_{\\theta^{*}}(\\pmb{w}_{\\mathrm{inv}})\\)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.619, + 0.39, + 0.635 + ], + "angle": 0, + "content": "3.2.2 Text-guided adversarial optimization" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.644, + 0.469, + 0.72 + ], + "angle": 0, + "content": "Given the inverted latent code \\( \\boldsymbol{w}_{\\mathrm{inv}} \\) and fine-tuned generator \\( G_{\\theta^{*}}(\\cdot) \\), our goal is to adversarially perturb this latent code \\( \\boldsymbol{w}_{\\mathrm{inv}} \\) in the low-dimensional generative manifold \\( \\mathcal{W} \\) to generate a protected face that fools the black-box FR model, while imitating the makeup style of the text prompt \\( t_{\\mathrm{makeup}} \\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.72, + 0.469, + 0.856 + ], + "angle": 0, + "content": "To achieve these objectives, we investigate the following questions: (i) how to effectively extract makeup style information from \\( t_{\\mathrm{makeup}} \\) and apply it to the face image \\( x \\) in an adversarial manner?, (ii) how to regularize the optimization process so that the output face image is not qualitatively impaired?, (iii) how to craft effective adversarial perturbations that mislead black-box FR models?, and (iv) how to preserve the human-perceived identity \\( \\mathcal{O}(x) \\) of the original face image while ensuring high privacy?" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.47, + 0.901 + ], + "angle": 0, + "content": "The first issue can be addressed by aligning the output adversarial image with the text prompt \\( t_{\\mathrm{makeup}} \\) in the embedding space of a pretrained vision-language model. The" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.242 + ], + "angle": 0, + "content": "second issue is addressed by enforcing the adversarial latent code to remain close to initialization \\( \\boldsymbol{w}_{\\mathrm{inv}} \\). The third issue is solved by crafting transferable text-guided adversarial faces on a white-box surrogate model (or an ensemble of models) with the goal of boosting the fooling rate on the black-box FR model. Finally, we leverage the disentangled nature of latent space in the generative model and incorporate an identity-preserving regularization to effectively maintain the original visual identity. We now present the details of the loss functions used to incorporate the above ideas." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.243, + 0.893, + 0.41 + ], + "angle": 0, + "content": "Textual Loss: A key ingredient of the proposed approach is text-based guidance to inconspicuously hide the adversarial perturbations into the makeup effect. This can be naively achieved by aligning the representation of \\( t_{\\mathrm{makeup}} \\) and the adversarial face \\( G_{\\theta^*}(\\boldsymbol{w}) \\) in the common embedding space of a pre-trained vision-language model (e.g. CLIP [47]). However, this approach will transform the whole output image to follow the makeup style of \\( t_{\\mathrm{makeup}} \\), which results in low diversity. Therefore, we use a directional CLIP loss that aligns the CLIP-space direction between the text-image pairs of the original and adversarial images. Specifically," + }, + { + "type": "equation", + "bbox": [ + 0.616, + 0.422, + 0.892, + 0.455 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {c l i p}} = 1 - \\frac {\\Delta I \\cdot \\Delta T}{| \\Delta I | | \\Delta T |}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.465, + 0.892, + 0.557 + ], + "angle": 0, + "content": "where \\(\\Delta T = E_{T}(t_{\\mathrm{makeup}}) - E_{T}(t_{\\mathrm{src}})\\) and \\(\\Delta I = E_I(G_{\\theta^*}(\\pmb {w})) - E_I(\\pmb {x})\\). Here, \\(E_{T}\\) and \\(E_{I}\\) are the text and image encoders of the CLIP model and \\(t_\\mathrm{src}\\) is the semantic text of the input image \\(\\pmb{x}\\). Since we are dealing with faces, \\(t_\\mathrm{src}\\) can be simply set as \"face\". This loss localizes makeup transfer (e.g. red lipstick) without affecting privacy." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.557, + 0.892, + 0.633 + ], + "angle": 0, + "content": "Adversarial Loss: Our goal is to traverse over the latent space \\(\\mathcal{W}\\) to find adversarial latent codes on the generative manifold whose face feature representation lies close to that of target image and far away from the original image itself i.e., \\(\\mathcal{D}(\\pmb{x}^p, \\pmb{x}) > \\mathcal{D}(\\pmb{x}^p, \\pmb{x}^t)\\). Hence, the adversarial loss is:" + }, + { + "type": "equation", + "bbox": [ + 0.556, + 0.645, + 0.892, + 0.663 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {a d v}} = \\mathcal {D} \\left(G _ {\\theta^ {*}} (\\boldsymbol {w}), \\boldsymbol {x} ^ {t}\\right) - \\mathcal {D} \\left(G _ {\\theta^ {*}} (\\boldsymbol {w}), \\boldsymbol {x}\\right), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.673, + 0.892, + 0.764 + ], + "angle": 0, + "content": "where \\(\\mathcal{D}(\\pmb{x}_1,\\pmb{x}_2) = 1 - \\cos [f(\\pmb{x}_1),f(\\pmb{x}_2)]\\) is the cosine distance. Since the malicious FR model is unknown in the black-box setting, Eq. 3 cannot be solved directly. Instead, following AMT-GAN [22], we perform adversarial optimization on an ensemble of white-box surrogate models to imitate the decision boundary of the unknown FR model." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.893, + 0.901 + ], + "angle": 0, + "content": "Identity Preservation Loss: The optimization over the generative manifold ensures that the protected image \\( \\pmb{x}^p \\) is natural i.e., artifact-free, however, it does not explicitly enforce the protected image to preserve the identity of the original image with respect to the human observer. To mitigate the issue, we take advantage of the semantic control exhibited by StyleGAN in its latent space. The latent code \\( \\pmb{w} \\in \\mathcal{W} \\) impacts image generation by controlling different level of semantics in the output image. Specifically, latent" + }, + { + "type": "footer", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "20599" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.258 + ], + "angle": 0, + "content": "codes corresponding to the initial layers of StyleGAN control high-level aspects such as pose, general hairstyle, and face shape [27]. Adversarially perturbing these latent layers can change these attributes, resulting in a change of identity (see Sec. 4.3). Latent codes corresponding to deeper layers of StyleGAN are associated with fine-level control such as makeup style [2]. Therefore, we perturb only those latent codes associated with deeper layers of StyleGAN, thereby restricting the adversarial faces to the identity preserving manifold. We further constrain the latent code to stay close to its initial value \\( w_{\\mathrm{inv}} \\) using the following regularization:" + }, + { + "type": "equation", + "bbox": [ + 0.134, + 0.269, + 0.47, + 0.288 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {l a t e n t}} = \\left\\| (\\boldsymbol {w} \\odot \\boldsymbol {m} _ {i d}) - \\left(\\boldsymbol {w} _ {\\text {i n v}} \\odot \\boldsymbol {m} _ {i d}\\right) \\right\\| _ {2}, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.296, + 0.47, + 0.388 + ], + "angle": 0, + "content": "where \\(\\odot\\) denotes element-wise product and \\(m_{id}\\) is an identity preservation mask that is 0 for the initial layers and 1 only for the deeper layers of the latent code. StyleGAN has 18 layers, each having a dimension of 512. The identity preservation mask is set to 1 only from layer 8 to 18. Finally, combining the three loss functions, we have" + }, + { + "type": "equation", + "bbox": [ + 0.129, + 0.398, + 0.47, + 0.416 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {t o t a l}} = \\lambda_ {\\text {a d v}} \\mathcal {L} _ {\\text {a d v}} + \\lambda_ {\\text {c l i p}} \\mathcal {L} _ {\\text {c l i p}} + \\lambda_ {\\text {l a t e n t}} \\mathcal {L} _ {\\text {l a t e n t}}, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.425, + 0.47, + 0.502 + ], + "angle": 0, + "content": "where \\(\\lambda_{\\mathrm{adv.}}\\), \\(\\lambda_{\\mathrm{clip}}\\), and \\(\\lambda_{\\mathrm{latent}}\\) are hyperparameters. Note that \\(\\mathcal{L}_{\\mathrm{adv}}\\) accounts for the adversarial objective in Eq. 1, while the text-guided makeup transfer \\((\\mathcal{L}_{\\mathrm{clip}})\\) and identity-preserving regularization \\((\\mathcal{L}_{\\mathrm{latent}})\\) implicitly enforce the naturalness constraint in Eq. 1." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.513, + 0.21, + 0.531 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.539, + 0.47, + 0.75 + ], + "angle": 0, + "content": "Implementation details: In all experiments, we use StyleGAN2 pretrained on the FFHQ face dataset as our generative model. For adversarial text guidance, we use a vision transformer-based CLIP model. For generator fine-tuning in the latent code initialization step, we use 450 iterations with value of \\(\\lambda_{2}\\) in Eq. 2 set to 0.5. For the makeup text input, we collect 40 text prompts based on the makeup style of diverse nature (details in supplementary material). For adversarial optimization, we use an Adam optimizer with \\(\\beta_{1}\\) and \\(\\beta_{2}\\) set to 0.9 and 0.999, respectively, and a learning rate of 0.01. We run the optimizer for 50 iterations to craft protected faces. We set the value of \\(\\lambda_{\\mathrm{adv}}\\), \\(\\lambda_{\\mathrm{clip}}\\), and \\(\\lambda_{\\mathrm{latent}}\\) to 1, 0.5, and 0.01, respectively. All our experiments are conducted on a A100 GPU with 40 GB memory." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Datasets: We perform experiments for both face verification and identification settings. Face verification: We use CelebA-HQ [26] and LADN [16] for the impersonation attack. We select subset of 1,000 images from CelebA-HQ and report average results over 4 target identities provided by [22]. Similarly, for LADN, we divide the 332 images available into 4 groups, where images in each group aim to impersonate the target identities provided by [22]. For dodging attack, we use CelebA-HQ [26] and LFW [23] datasets. Specifically, we select 500 subjects at random and" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.243 + ], + "angle": 0, + "content": "each subject has a pair of faces. Face identification: For impersonation and dodging, we use CelebA-HQ [26] and LFW [23] as our evaluation set. For both datasets, we randomly select 500 subjects, each with a pair of faces. We assign one image in the pair to the gallery set and the other to the probe set. Both impersonation and dodging attacks are performed on the probe set. For impersonation, we insert 4 target identities provided by [22] into the gallery set. A more detailed description of all datasets and pre-processing steps is provided in the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.244, + 0.892, + 0.379 + ], + "angle": 0, + "content": "Target Models: We aim to protect user facial privacy by attacking four FR model with diverse back bones in the black-box settings. The target models include IRSE50 [21], IR152 [9], FaceNet [51], and MobileFace [5]. Following standard protocol, we align and crop the face images using MTCNN [75] before giving them as input to FR models. Further, we also report privacy protection performance based on commercial FR API including Face++ and Tencent Yunshentu FR platforms." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.381, + 0.893, + 0.652 + ], + "angle": 0, + "content": "Evaluation metrics: Following [70], we use protection success rate (PSR) to evaluate the proposed approach. PSR is defined as the fraction of protected faces missclassified by the malicious FR system. To evaluate PSR, we use the thresholding and closed set strategies for face verification and identification, respectively. For face identification, we also use Rank-N targeted identity success rate (Rank-N-T) and untargeted identity success rate (Rank-N-U), where Rank-N-T means that target image \\( \\boldsymbol{x}^t \\) will appear at least once in the top N candidates shortlisted from the gallery and Rank-N-U implies that the top N candidate list does not have the same identity as that of original image \\( \\boldsymbol{x} \\). We also report results of PSNR (dB), SSIM, and FID [19] scores to evaluate the imperceptibility of method. Large PSNR and SSIM [65] indicates better match with the original images, while low FID score indicates more realistic images. For commercial APIs, we directly report the confidence score returned by the respective servers." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.654, + 0.892, + 0.835 + ], + "angle": 0, + "content": "Baseline methods: We compare our approach with recent noise-based and makeup based facial privacy protection approaches. Noise based methods include PGD [37], MI-FGSM [10], TI-DIM [11], and TIP-IM [70], whereas makeup-based approaches are Adv-Makeup [71] and AMTGAN [22]. We want to highlight that TIP-IM and AMTGAN are considered the state-of-the-art (SOTA) for face privacy protection against black-box FR systems in noise-based and unrestricted settings, respectively. TIP-IM also incorporate multi-target objective in its optimization to find the optimal target image among multiple targets. For fair comparison, we use its single target variant." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.847, + 0.704, + 0.862 + ], + "angle": 0, + "content": "4.1. Experimental Results" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.893, + 0.902 + ], + "angle": 0, + "content": "In this section, we present experimental results of our approach in black-box settings on four different pretrained" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "20600" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.089, + 0.895, + 0.119 + ], + "angle": 0, + "content": "Table 2. Protection success rate (PSR %) of black-box impersonation attack under the face verification task. For each column, the other three FR systems are used as surrogates to generate the protected faces." + }, + { + "type": "table", + "bbox": [ + 0.127, + 0.122, + 0.847, + 0.273 + ], + "angle": 0, + "content": "
MethodCelebA-HQLADN-DatasetAverage
IRSE50IR152FaceNetMobileFaceIRSE50IR152FaceNetMobileFace
Clean7.293.801.0812.682.713.610.605.114.61
Inverted5.572.770.6013.326.804.510.2511.665.68
PGD [37]36.8720.681.8543.9940.0919.593.8241.0925.60
MI-FGSM [10]45.7925.032.5845.8548.9025.576.3145.0130.63
TI-DIM [11]63.6336.1715.3057.1256.3634.1822.1148.3041.64
Adv-Makeup(ICAI'21) [71]21.959.481.3722.0029.6410.030.9722.3814.72
TIP-IM(ICCV'21) [70]54.4037.2340.7448.7265.8943.5763.5046.4850.06
AMT-GAN(CVPR'22) [22]76.9635.1316.6250.7189.6449.1232.1372.4352.84
Ours81.1048.4241.7275.2691.5753.3147.9179.9464.90
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.279, + 0.895, + 0.322 + ], + "angle": 0, + "content": "Table 3. Protection success rate (PSR %) of black-box dodging (top) and impersonation (bottom) attacks under the face identification task for LFW dataset [23]. For each column, the other three FR systems are used as surrogates to generate the protected faces. R1-U: Rank-1-Untargeted, R5-U: Rank-5-Untargeted, R1-T: Rank-1-Targeted, R5-T: Rank-5-Targeted." + }, + { + "type": "table", + "bbox": [ + 0.116, + 0.325, + 0.86, + 0.487 + ], + "angle": 0, + "content": "
MethodIRSE50IR152FaceNetMobileFaceAverage
R1-UR5-UR1-UR5-UR1-UR5-UR1-UR5-UR1-UR5-U
MI-FGSM [10]70.242.658.441.859.234.068.047.263.941.4
TI-DIM [11]79.051.267.454.074.452.079.261.675.054.7
TIP-IM(ICCv'21) [70]81.452.271.854.676.049.882.263.077.854.9
Ours86.659.473.456.683.851.285.066.882.258.5
R1-TR5-TR1-TR5-TR1-TR5-TR1-TR5-TR1-TR5-T
MI-FGSM [10]4.010.23.214.29.018.88.422.46.1516.4
TI-DIM [11]4.013.67.819.618.032.821.639.012.8526.25
TIP-IM(ICCv'21) [70]8.028.211.631.225.256.834.051.419.741.9
Ours11.237.816.051.227.454.039.061.223.451.05
" + }, + { + "type": "table", + "bbox": [ + 0.094, + 0.502, + 0.328, + 0.58 + ], + "angle": 0, + "content": "
MethodFID ↓PSR Gain ↑
Adv-Makeup [71]4.230
TIP-IM [70]38.7335.34
AMT-GAN [22]34.4438.12
Ours26.6250.18
" + }, + { + "type": "table_caption", + "bbox": [ + 0.336, + 0.502, + 0.471, + 0.573 + ], + "angle": 0, + "content": "Table 4. FID comparison. PSR Gain is absolute gain in PSR relative to AdvMakeup." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.599, + 0.47, + 0.675 + ], + "angle": 0, + "content": "FR models under face verification and identification tasks. To generate protected images, we use three FR models as a surrogate to imitate the decision boundary of the fourth FR model. All results are averaged over 5 text based makeup styles that are provided in the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.675, + 0.471, + 0.902 + ], + "angle": 0, + "content": "For face verification experiments, we set the system threshold value at 0.01 false match rate for each FR model i.e., IRSE50 (0.241), IR152 (0.167), FaceNet (0.409), and MobileFace (0.302). Quantitative results in terms of PSR for impersonation attack under the face verification task are shown in Tab. 2. Our approach is able to achieve an average absolute gain of about \\(12\\%\\) and \\(14\\%\\) over SOTA unrestricted [22] and noise-based [70] facial privacy protection methods, respectively. Qualitative results are shown in Fig. 1 which shows that protected faces generated by our approach are more realistic. Results for dodging attacks under face verification are provided in the supplementary material. In Tab. 3, we also provide PSR vales under the face identification task for dodging (untargeted) and impersonation attacks. Our approach consistently outperforms recent" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.504, + 0.892, + 0.626 + ], + "angle": 0, + "content": "methods at both Rank-1 and Rank-5 settings. We emphasize that we are the first to show effectiveness of generative models in offering untargeted privacy protection (dodging) in a more practical identification setting. Since AMT-GAN and Adv-Makeup are originally trained to impersonate target identity under the verification task, we have not included them in Tab. 3. Qualitative results for LFW and CelebA are provided in the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.626, + 0.892, + 0.749 + ], + "angle": 0, + "content": "We report FID scores (lower is better) of our approach in Tab. 4 for CelebA and LADN datasets to measure naturalness. Adv-Makeup has the lowest FID score as it only applies makeup to the eye region without changing the rest of the face. However, this kind of restriction results in poor PSR. Our method has lower FID compared to TIP-IM and AMT-GAN and achieves the highest PSR. We provide PSNR and SSIM results in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.757, + 0.851, + 0.773 + ], + "angle": 0, + "content": "4.2. Effectiveness in Real-World Applications" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.781, + 0.895, + 0.903 + ], + "angle": 0, + "content": "We now show the effectiveness of our approach to protect facial images (through targeted impersonation) against commercial API such as Face++ and Tencent Yunshentu FR platform operating in the verification mode. These APIs return confidence scores between 0 to 100 to measure whether two images are similar or not, where a high confidence score indicates high similarity. As the training data and model parameters of these propriety FR models are unknown, it" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.958 + ], + "angle": 0, + "content": "20601" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.097, + 0.089, + 0.458, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.236, + 0.473, + 0.295 + ], + "angle": 0, + "content": "Figure 4. Average confidence score (higher is better) returned by a real-world face verification API, Face++, for impersonation attack. Our approach has a higher confidence score than state-of-the-art makeup and noise-based facial privacy protection methods." + }, + { + "type": "table_caption", + "bbox": [ + 0.13, + 0.308, + 0.416, + 0.322 + ], + "angle": 0, + "content": "Table 5. Impact of \\( {\\lambda }_{\\text{latent }} \\) on FID score and PSR." + }, + { + "type": "table", + "bbox": [ + 0.09, + 0.327, + 0.462, + 0.378 + ], + "angle": 0, + "content": "
λlatent0.50.10.050.010.0050.00010
FID11.621.425.227.830.138.443.2
PSR (%)31.239.057.476.283.890.093.6
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.397, + 0.47, + 0.504 + ], + "angle": 0, + "content": "effectively mimics a real-world scenario. We protect 100 faces randomly selected from CelebA-HQ using the baselines and the proposed method. In Fig. 4, we show the average confidence score returned by Face++ against these images. These results indicate that our method has a high PSR compared to baselines. We defer more details and results for Tencent Yunshentu API to supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.513, + 0.245, + 0.529 + ], + "angle": 0, + "content": "4.3. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.537, + 0.47, + 0.567 + ], + "angle": 0, + "content": "Next, we report some ablations to evaluate the contributions of our loss components." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.568, + 0.47, + 0.674 + ], + "angle": 0, + "content": "Makeup based text guidance: As shown in Fig. 5 (top), in the absence of text guidance, resulting images may contain artifacts due to increased perturbations induced by the adversarial objective. Text-guidance effectively hides the perturbations in the makeup, leading to more natural looking images. It also provides the user more flexibility to select a desired makeup style compared to a reference image." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.674, + 0.471, + 0.78 + ], + "angle": 0, + "content": "Identity preserving regularization: Optimizing over the whole latent space provides more degrees of freedom and increases the PSR. However, it does not explicitly enforce adversarial optimization to preserve the user identity as shown in Fig. 5 (bottom). The proposed identity preserving regularization effectively preserves identity, while imitating the desire makeup style." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.78, + 0.471, + 0.871 + ], + "angle": 0, + "content": "Impact of latent loss weight: Decreasing the weight assigned to the latent loss \\(\\lambda_{\\mathrm{latent}}\\) results in an increase in both the FID score and PSR (and vice versa). Allowing the latent to deviate more from the initial inverted latent code of the given face image often results in artifacts caused by the adversarial loss, degrading naturalness but aiding privacy." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.471, + 0.901 + ], + "angle": 0, + "content": "Robustness against textual variations. Finally, we evaluate the impact of different textual styles on the PSR. We" + }, + { + "type": "image", + "bbox": [ + 0.554, + 0.089, + 0.843, + 0.344 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.356, + 0.895, + 0.468 + ], + "angle": 0, + "content": "Figure 5. Top: Effect of makeup-based text guidance on the visual quality of the output images. Output images are able to impersonate the target identity for face verification. Text-prompt is \"tanned makeup with red lipstick\". Bottom: Optimizing over all latent codes changes the identity of the protected image. Our identity-preserving regularization enforces the adversarial optimization to search for latent codes that hide the perturbations in the makeup effect while simultaneously preserving visual identity." + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.482, + 0.895, + 0.525 + ], + "angle": 0, + "content": "Table 6. Impact of different textual makeup styles on PSR. Makeup styles are \"tanned\", \"pale\", \"pink eyeshadows\", \"red lipstick\", and \"Matte\". Std. denotes standard deviation." + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.529, + 0.88, + 0.571 + ], + "angle": 0, + "content": "
t1makeupt2makeupt3makeupt4makeupt5makeupStd.
PSR74.177.378.478.779.21.24
" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.59, + 0.895, + 0.667 + ], + "angle": 0, + "content": "select five text-based makeup styles to protect 1000 images of CelebA-HQ using our method. Results in Tab. 6 show that PSR does not change significantly (low standard deviation) for different makeup styles, indicating robustness of our approach wrt different text-based makeup styles." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.679, + 0.62, + 0.695 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.705, + 0.895, + 0.903 + ], + "angle": 0, + "content": "We have proposed a framework to protect privacy of face images on online platforms by carefully searching for adversarial codes in the low-dimensional latent manifold of a pre-trained generative model. We have shown that incorporating a makeup text-guided loss and an identity preserving regularization effectively hides the adversarial perturbations in the makeup style, provides images with high quality, and preserves human-perceived identity. While this approach is robust to the user-defined text-prompt and target identity, it would be beneficial if the text-prompt and target identity can be automatically selected based on the given face image. Limitations of our method include high computational cost at the time of protected face generation." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "20602" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.116, + 0.47, + 0.186 + ], + "angle": 0, + "content": "[1] Shane Ahern, Dean Eckles, Nathaniel S Good, Simon King, Mor Naaman, and Rahul Nair. Over-exposed? privacy patterns and considerations in online and mobile photo sharing. In Proceedings of the SIGCHI conference on Human factors in computing systems, pages 357-366, 2007. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.187, + 0.472, + 0.257 + ], + "angle": 0, + "content": "[2] Amit H Bermano, Rinon Gal, Yuval Alaluf, Ron Mokady, Yotam Nitzan, Omer Tov, Oren Patashnik, and Daniel Cohen-Or. State-of-the-art in the architecture, methods and applications of stylegan. In Computer Graphics Forum, volume 41, pages 591–611. Wiley Online Library, 2022. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.259, + 0.471, + 0.313 + ], + "angle": 0, + "content": "[3] Anand Bhattachad, Min Jin Chong, Kaizhao Liang, Bo Li, and David A Forsyth. Unrestricted adversarial examples via semantic manipulation. arXiv preprint arXiv:1904.06347, 2019. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.316, + 0.472, + 0.386 + ], + "angle": 0, + "content": "[4] Jia-Wei Chen, Li-Ju Chen, Chia-Mu Yu, and Chun-Shien Lu. Perceptual indistinguishability-net (pi-net): Facial image obfuscation with manipulable semantics. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6478–6487, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.388, + 0.47, + 0.444 + ], + "angle": 0, + "content": "[5] Sheng Chen, Yang Liu, Xiang Gao, and Zhen Han. Mobile-facenets: Efficient cnns for accurate real-time face verification on mobile devices. In Chinese Conference on Biometric Recognition, pages 428-438. Springer, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.445, + 0.47, + 0.514 + ], + "angle": 0, + "content": "[6] Valeriia Cherepanova, Micah Goldblum, Harrison Foley, Shiyuan Duan, John P Dickerson, Gavin Taylor, and Tom Goldstein. Lowkey: Leveraging adversarial attacks to protect social media users from facial recognition. In International Conference on Learning Representations, 2020. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.517, + 0.47, + 0.571 + ], + "angle": 0, + "content": "[7] William L Croft, Jörg-Rüdiger Sack, and Wei Shi. Differentially private facial obfuscation via generative adversarial networks. Future Generation Computer Systems, 129:358-379, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.574, + 0.471, + 0.63 + ], + "angle": 0, + "content": "[8] Ali Dabouei, Sobhan Soleymani, Jeremy Dawson, and Nasser Nasrabadi. Fast geometrically-perturbed adversarial faces. In 2019 IEEE Winter Conference on Applications of Computer Vision (WACV), pages 1979-1988. IEEE, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.632, + 0.47, + 0.7 + ], + "angle": 0, + "content": "[9] Jiankang Deng, Jia Guo, Niannan Xue, and Stefanos Zafeiriou. Arcface: Additive angular margin loss for deep face recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4690-4699, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.702, + 0.47, + 0.773 + ], + "angle": 0, + "content": "[10] Yinpeng Dong, Fangzhou Liao, Tianyu Pang, Hang Su, Jun Zhu, Xiaolin Hu, and Jianguo Li. Boosting adversarial attacks with momentum. In Proceedings of the 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR'18), pages 9185-9193, 2018. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.774, + 0.47, + 0.844 + ], + "angle": 0, + "content": "[11] Yinpeng Dong, Tianyu Pang, Hang Su, and Jun Zhu. Evading defenses to transferable adversarial examples by translation-invariant attacks. In Proceedings of the 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR'19), pages 4312-4321, 2019. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.845, + 0.47, + 0.9 + ], + "angle": 0, + "content": "[12] Hang Du, Hailin Shi, Dan Zeng, Xiao-Ping Zhang, and Tao Mei. The elements of end-to-end deep face recognition: A survey of recent advances. ACM Computing Surveys (CSUR), 54(10s):1-42, 2022. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.116, + 0.472, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.134 + ], + "angle": 0, + "content": "[13] Yifan Du, Zikang Liu, Junyi Li, and Wayne Xin Zhao. A survey of vision-language pre-trained models. arXiv preprint arXiv:2202.10936, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.136, + 0.892, + 0.176 + ], + "angle": 0, + "content": "[14] Kevin Frans, Lisa B Soros, and Olaf Witkowski. Clipdraw: Exploring text-to-drawing synthesis through language-image encoders. arXiv preprint arXiv:2106.14843, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.178, + 0.892, + 0.231 + ], + "angle": 0, + "content": "[15] Ian J. Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron C. Courville, and Yoshua Bengio. Generative adversarial networks. CoRR, abs/1406.2661, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.233, + 0.894, + 0.302 + ], + "angle": 0, + "content": "[16] Qiao Gu, Guanzhi Wang, Mang Tik Chiu, Yu-Wing Tai, and Chi-Keung Tang. Ladin: Local adversarial disentangling network for facial makeup and de-makeup. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10481-10490, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.303, + 0.892, + 0.357 + ], + "angle": 0, + "content": "[17] Nitzan Guetta, Asaf Shabtai, Inderjeet Singh, Satoru Momiyama, and Yuval Elovici. Dodging attack using carefully crafted natural makeup. arXiv preprint arXiv:2109.06467, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.358, + 0.892, + 0.385 + ], + "angle": 0, + "content": "[18] Rebecca Heilweil. The world's scariest facial recognition company explained. Vox, May, 8, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.386, + 0.892, + 0.453 + ], + "angle": 0, + "content": "[19] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.455, + 0.892, + 0.482 + ], + "angle": 0, + "content": "[20] Kashmir Hill. The secretive company that might end privacy as we know it. *The New York Times*, 18:2020, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.484, + 0.892, + 0.525 + ], + "angle": 0, + "content": "[21] Jie Hu, Li Shen, and Gang Sun. Squeeze-and-excitation networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7132-7141, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.526, + 0.892, + 0.607 + ], + "angle": 0, + "content": "[22] Shengshan Hu, Xiaogeng Liu, Yechao Zhang, Minghui Li, Leo Yu Zhang, Hai Jin, and Libing Wu. Protecting facial privacy: Generating adversarial identity masks via style-robust makeup transfer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15014–15023, 2022. 1, 2, 3, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.609, + 0.892, + 0.678 + ], + "angle": 0, + "content": "[23] Gary B Huang, Marwan Mattar, Tamara Berg, and Eric Learned-Miller. Labeled faces in the wild: A database for studying face recognition in unconstrained environments. In Workshop on faces in 'Real-Life' Images: detection, alignment, and recognition, 2008. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.679, + 0.892, + 0.747 + ], + "angle": 0, + "content": "[24] Phillip Isola, Jun-Yan Zhu, Tinghui Zhou, and Alexei A Efros. Image-to-image translation with conditional adversarial networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1125-1134, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.748, + 0.892, + 0.801 + ], + "angle": 0, + "content": "[25] Kazuya Kakizaki and Kosuke Yoshida. Adversarial image translation: Unrestricted adversarial examples in face recognition systems. arXiv preprint arXiv:1905.03421, 2019. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.803, + 0.892, + 0.858 + ], + "angle": 0, + "content": "[26] Tero Karras, Timo Aila, Samuli Laine, and Jaakko Lehtinen. Progressive growing of gans for improved quality, stability, and variation. In International Conference on Learning Representations, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.86, + 0.895, + 0.902 + ], + "angle": 0, + "content": "[27] Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of stylegan. In Proceedings of" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.895, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20603" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.11, + 0.093, + 0.469, + 0.12 + ], + "angle": 0, + "content": "the IEEE/CVF conference on computer vision and pattern recognition, pages 8110-8119, 2020. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.121, + 0.47, + 0.188 + ], + "angle": 0, + "content": "[28] Gwanghyun Kim, Taesung Kwon, and Jong Chul Ye. Diffusionclip: Text-guided diffusion models for robust image manipulation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2426-2435, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.19, + 0.469, + 0.245 + ], + "angle": 0, + "content": "[29] Stepan Komkov and Aleksandr Petiushko. Advhat: Realworld adversarial attack on arcface face id system. In 2020 25th International Conference on Pattern Recognition (ICPR), pages 819-826. IEEE, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.247, + 0.469, + 0.302 + ], + "angle": 0, + "content": "[30] Zhenzhong Kuang, Zhiqiang Guo, Jinglong Fang, Jun Yu, Noboru Babaguchi, and Jianping Fan. Unnoticeable synthetic face replacement for image privacy protection. Neurocomputing, 457:322-333, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.303, + 0.469, + 0.383 + ], + "angle": 0, + "content": "[31] Zhenzhong Kuang, Huigui Liu, Jun Yu, Aikui Tian, Lei Wang, Jianping Fan, and Noboru Babaguchi. Effective de-identification generative adversarial network for face anonymization. In Proceedings of the 29th ACM International Conference on Multimedia, pages 3182–3191, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.386, + 0.47, + 0.441 + ], + "angle": 0, + "content": "[32] Gihyun Kwon and Jong Chul Ye. Clipstyler: Image style transfer with a single text condition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18062-18071, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.442, + 0.469, + 0.482 + ], + "angle": 0, + "content": "[33] Tao Li and Min Soo Choi. Deepblur: A simple and effective method for natural image obfuscation. arXiv preprint arXiv:2104.02655, 1, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.484, + 0.47, + 0.539 + ], + "angle": 0, + "content": "[34] Tao Li and Lei Lin. Anonymousnet: Natural face de-identification with measurable privacy. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops, pages 0–0, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.54, + 0.469, + 0.593 + ], + "angle": 0, + "content": "[35] Fangcheng Liu, Chao Zhang, and Hongyang Zhang. Towards transferable unrestricted adversarial examples with minimum changes. arXiv preprint arXiv:2201.01102, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.595, + 0.469, + 0.65 + ], + "angle": 0, + "content": "[36] Suolan Liu, Lizhi Kong, and Hongyuan Wang. Face detection and encryption for privacy preserving in surveillance video. In Chinese Conference on Pattern Recognition and Computer Vision (PRCV), pages 162-172. Springer, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.651, + 0.469, + 0.718 + ], + "angle": 0, + "content": "[37] Aleksander Madry, Aleksandar Makelov, Ludwig Schmidt, Dimitris Tsipras, and Adrian Vladu. Towards deep learning models resistant to adversarial attacks. In Proceedings of the 6th International Conference on Learning Representations (ICLR'18), 2018. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.72, + 0.469, + 0.789 + ], + "angle": 0, + "content": "[38] Blaz Meden, Peter Rot, Philipp Terhörst, Naser Damer, Arjan Kuijper, Walter J Scheirer, Arun Ross, Peter Peer, and Vitomir Štruc. Privacy-enhancing face biometrics: A comprehensive survey. IEEE Transactions on Information Forensics and Security, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.79, + 0.469, + 0.831 + ], + "angle": 0, + "content": "[39] Dongbin Na, Sangwoo Ji, and Jong Kim. Unrestricted blackbox adversarial attack using gan with limited queries. arXiv preprint arXiv:2208.11613, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.832, + 0.469, + 0.9 + ], + "angle": 0, + "content": "[40] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741, 2021. 3" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.093, + 0.47, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.148 + ], + "angle": 0, + "content": "[41] Seong Joon Oh, Rodrigo Benenson, Mario Fritz, and Bernt Schiele. Faceless person recognition: Privacy implications in social media. In European Conference on Computer Vision, pages 19-35. Springer, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.15, + 0.892, + 0.206 + ], + "angle": 0, + "content": "[42] Seong Joon Oh, Mario Fritz, and Bernt Schiele. Adversarial image perturbation for privacy protection a game theory perspective. In 2017 IEEE International Conference on Computer Vision (ICCV), pages 1491-1500. IEEE, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.208, + 0.891, + 0.235 + ], + "angle": 0, + "content": "[43] Omkar M Parkhi, Andrea Vedaldi, and Andrew Zisserman. Deep face recognition. 2015. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.236, + 0.892, + 0.304 + ], + "angle": 0, + "content": "[44] Or Patashnik, Zongze Wu, Eli Shechtman, Daniel Cohen-Or, and Dani Lischinski. Styleclip: Text-driven manipulation of stylegan imagery. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2085–2094, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.307, + 0.892, + 0.403 + ], + "angle": 0, + "content": "[45] P Jonathon Phillips, Amy N Yates, Ying Hu, Carina A Hahn, Eilidh Noyes, Kelsey Jackson, Jacqueline G Cava-zos, Geraldine Jeckeln, Rajeev Ranjan, Swami Sankaranarayanan, et al. Face recognition accuracy of forensic examiners, superrecognizers, and face recognition algorithms. Proceedings of the National Academy of Sciences, 115(24):6171-6176, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.406, + 0.892, + 0.475 + ], + "angle": 0, + "content": "[46] Omid Poursaeed, Tianxing Jiang, Harry Yang, Serge Belongie, and Ser-Nam Lim. Robustness and generalization via generative adversarial training. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15711-15720, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.477, + 0.892, + 0.56 + ], + "angle": 0, + "content": "[47] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.562, + 0.892, + 0.644 + ], + "angle": 0, + "content": "[48] Rajeev Ranjan, Swami Sankaranarayanan, Ankan Bansal, Navaneeth Bodla, Jun-Cheng Chen, Vishal M Patel, Carlos D Castillo, and Rama Chellappa. Deep learning for understanding faces: Machines may be just as good, or better, than humans. IEEE Signal Processing Magazine, 35(1):66-83, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.647, + 0.892, + 0.701 + ], + "angle": 0, + "content": "[49] Daniel Roich, Ron Mokady, Amit H Bermano, and Daniel Cohen-Or. Pivotal tuning for latent-based editing of real images. ACM Transactions on Graphics (TOG), 42(1):1-13, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.704, + 0.892, + 0.759 + ], + "angle": 0, + "content": "[50] Tim Salimans, Ian Goodfellow, Wojciech Zaremba, Vicki Cheung, Alec Radford, and Xi Chen. Improved techniques for training gans. Advances in neural information processing systems, 29, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.761, + 0.892, + 0.816 + ], + "angle": 0, + "content": "[51] Florian Schroff, Dmitry Kalenichenko, and James Philbin. Facenet: A unified embedding for face recognition and clustering. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 815-823, 2015. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.818, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[52] Sachith Seneviratne, Nuran Kasthuriarachchi, Sanka Rasnayaka, Danula Hettiachchi, and Ridwan Shariffdeen. Does a face mask protect my privacy?: Deep learning to predict protected attributes from masked face images. In Australasian Joint Conference on Artificial Intelligence, pages 91-102. Springer, 2022. 3" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20604" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.161 + ], + "angle": 0, + "content": "[53] Shawn Shan, Emily Wenger, Jiayun Zhang, Huiying Li, Haitao Zheng, and Ben Y Zhao. Fawkes: Protecting privacy against unauthorized deep learning models. In 29th USENIX security symposium (USENIX Security 20), pages 1589-1604, 2020. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.163, + 0.471, + 0.219 + ], + "angle": 0, + "content": "[54] Mahmood Sharif, Sruti Bhagavatula, Lujo Bauer, and Michael K Reiter. A general framework for adversarial examples with objectives. ACM Transactions on Privacy and Security (TOPS), 22(3):1-30, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.221, + 0.471, + 0.276 + ], + "angle": 0, + "content": "[55] Yang Song, Rui Shu, Nate Kushman, and Stefano Ermon. Constructing unrestricted adversarial examples with generative models. Advances in Neural Information Processing Systems, 31, 2018. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.278, + 0.471, + 0.346 + ], + "angle": 0, + "content": "[56] Qianru Sun, Liqian Ma, Seong Joon Oh, Luc Van Gool, Bernt Schiele, and Mario Fritz. Natural and effective obfuscation by head inpainting. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 5050-5059, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.348, + 0.471, + 0.404 + ], + "angle": 0, + "content": "[57] Christian Szegedy, Wojciech Zaremba, Ilya Sutskever, Joan Bruna, Dumitru Erhan, Ian Goodfellow, and Rob Fergus. Intriguing properties of neural networks. In International Conference on Learning Representations, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.405, + 0.471, + 0.447 + ], + "angle": 0, + "content": "[58] Huan Tian, Tianqing Zhu, and Wanlei Zhou. Fairness and privacy preservation for facial images: Gan-based methods. Computers & Security, 122:102902, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.448, + 0.471, + 0.503 + ], + "angle": 0, + "content": "[59] Omer Tov, Yuval Alaluf, Yotam Nitzan, Or Patashnik, and Daniel Cohen-Or. Designing an encoder for stylegan image manipulation. ACM Transactions on Graphics (TOG), 40(4):1-14, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.504, + 0.471, + 0.574 + ], + "angle": 0, + "content": "[60] Can Wang, Menglei Chai, Mingming He, Dongdong Chen, and Jing Liao. Clip-nerf: Text-and-image driven manipulation of neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3835-3844, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.576, + 0.471, + 0.603 + ], + "angle": 0, + "content": "[61] Mei Wang and Weihong Deng. Deep face recognition: A survey. Neurocomputing, 429:215-244, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.604, + 0.471, + 0.66 + ], + "angle": 0, + "content": "[62] Shunxin Wang, Una M Kelly, and Raymond NJ Veldhuis. Gender obfuscation through face morphing. In 2021 IEEE International Workshop on Biometrics and Forensics (IWBF), pages 1-6. IEEE, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.661, + 0.471, + 0.73 + ], + "angle": 0, + "content": "[63] Ya Wang, Tianlong Bao, Chunhui Ding, and Ming Zhu. Face recognition in real-world surveillance videos with deep learning method. In 2017 2nd international conference on image, vision and computing (icivc), pages 239-243. IEEE, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.732, + 0.471, + 0.772 + ], + "angle": 0, + "content": "[64] Yinggui Wang, Jian Liu, Man Luo, Le Yang, and Li Wang. Privacy-preserving face recognition in the frequency domain. 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.774, + 0.471, + 0.83 + ], + "angle": 0, + "content": "[65] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600-612, 2004. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.832, + 0.471, + 0.901 + ], + "angle": 0, + "content": "[66] Tianyi Wei, Dongdong Chen, Wenbo Zhou, Jing Liao, Zhentao Tan, Lu Yuan, Weiming Zhang, and Nenghai Yu. Hairclip: Design your hair by text and reference image. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18072-18081, 2022. 3" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.471, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.134 + ], + "angle": 0, + "content": "[67] Emily Wenger, Shawn Shan, Haitao Zheng, and Ben Y Zhao. Sok: Anti-facial recognition technology. arXiv preprint arXiv:2112.04558, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.136, + 0.892, + 0.178 + ], + "angle": 0, + "content": "[68] Chaowei Xiao, Jun-Yan Zhu, Bo Li, Warren He, Mingyan Liu, and Dawn Song. Spatially transformed adversarial examples. arXiv preprint arXiv:1801.02612, 2018. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.179, + 0.892, + 0.261 + ], + "angle": 0, + "content": "[69] Zihao Xiao, Xianfeng Gao, Chilin Fu, Yinpeng Dong, Wei Gao, Xiaolu Zhang, Jun Zhou, and Jun Zhu. Improving transferability of adversarial patches on face recognition with generative models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11845-11854, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.262, + 0.892, + 0.332 + ], + "angle": 0, + "content": "[70] Xiao Yang, Yinpeng Dong, Tianyu Pang, Hang Su, Jun Zhu, Yuefeng Chen, and Hui Xue. Towards face encryption by generating adversarial identity masks. In Proceedings of the 2021 IEEE/CVF International Conference on Computer Vision (ICCV'21), pages 3897-3907, 2021. 1, 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.333, + 0.892, + 0.415 + ], + "angle": 0, + "content": "[71] Bangjie Yin, Wenxuan Wang, Taiping Yao, Junfeng Guo, Zelun Kong, Shouhong Ding, Jilin Li, and Cong Liu. Advmakeup: A new imperceptible and transferable attack on face recognition. In Proceedings of the 30th International Joint Conference on Artificial Intelligence (IJCAI'21), pages 1252-1258, 2021. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.417, + 0.892, + 0.473 + ], + "angle": 0, + "content": "[72] Bangjie Yin, Wenxuan Wang, Taiping Yao, Junfeng Guo, Zelun Kong, Shouhong Ding, Jilin Li, and Cong Liu. Advmakeup: A new imperceptible and transferable attack on face recognition. arXiv preprint arXiv:2105.03162, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.474, + 0.892, + 0.529 + ], + "angle": 0, + "content": "[73] Shengming Yuan, Qilong Zhang, Lianli Gao, Yaya Cheng, and Jingkuan Song. Natural color fool: Towards boosting black-box unrestricted attacks. arXiv preprint arXiv:2210.02041, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.531, + 0.892, + 0.586 + ], + "angle": 0, + "content": "[74] Jiaming Zhang, Jitao Sang, Xian Zhao, Xiaowen Huang, Yanfeng Sun, and Yongli Hu. Adversarial privacy-preserving filter. In Proceedings of the 28th ACM International Conference on Multimedia, pages 1423-1431, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.587, + 0.892, + 0.642 + ], + "angle": 0, + "content": "[75] Kaipeng Zhang, Zhanpeng Zhang, Zhifeng Li, and Yu Qiao. Joint face detection and alignment using multitask cascaded convolutional networks. IEEE signal processing letters, 23(10):1499-1503, 2016. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.643, + 0.892, + 0.713 + ], + "angle": 0, + "content": "[76] Zhengyu Zhao, Zhuoran Liu, and Martha Larson. Towards large yet imperceptible adversarial image perturbations with perceptual color distance. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1039-1048, 2020. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.714, + 0.892, + 0.769 + ], + "angle": 0, + "content": "[77] Yaoyao Zhong and Weihong Deng. Opom: Customized invisible cloak towards face privacy protection. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2022, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.77, + 0.892, + 0.827 + ], + "angle": 0, + "content": "[78] Jizhe Zhou and Chi-Man Pun. Personal privacy protection via irrelevant faces tracking and pixelation in video live streaming. IEEE Transactions on Information Forensics and Security, 16:1088-1103, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.828, + 0.892, + 0.883 + ], + "angle": 0, + "content": "[79] Zheng-An Zhu, Yun-Zhong Lu, and Chen-Kuo Chiang. Generating adversarial examples by makeup attacks on face recognition. In 2019 IEEE International Conference on Image Processing (ICIP), pages 2516-2520. IEEE, 2019. 3" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.883 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20605" + } + ] +] \ No newline at end of file diff --git a/2023/CLIP2Protect_ Protecting Facial Privacy Using Text-Guided Makeup via Adversarial Latent Search/d379e97e-b7a0-42bb-a735-a3de89cdb054_origin.pdf b/2023/CLIP2Protect_ Protecting Facial Privacy Using Text-Guided Makeup via Adversarial Latent Search/d379e97e-b7a0-42bb-a735-a3de89cdb054_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..ec31d54f97c428b8f2493929e6c3add9926c5a84 --- /dev/null +++ b/2023/CLIP2Protect_ Protecting Facial Privacy Using Text-Guided Makeup via Adversarial Latent Search/d379e97e-b7a0-42bb-a735-a3de89cdb054_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1413c39a84c0131921109588118f85c86ccfb6fae7e53c90027a6d65f17fd3e7 +size 1844432 diff --git a/2023/CLIP2Protect_ Protecting Facial Privacy Using Text-Guided Makeup via Adversarial Latent Search/full.md b/2023/CLIP2Protect_ Protecting Facial Privacy Using Text-Guided Makeup via Adversarial Latent Search/full.md new file mode 100644 index 0000000000000000000000000000000000000000..83360e9fddfebd15248485ed7390c422f0acbfef --- /dev/null +++ b/2023/CLIP2Protect_ Protecting Facial Privacy Using Text-Guided Makeup via Adversarial Latent Search/full.md @@ -0,0 +1,313 @@ +# CLIP2Protect: Protecting Facial Privacy using Text-Guided Makeup via Adversarial Latent Search + +Fahad Shamshad Muzammal Naseer Karthik Nandakumar Mohamed Bin Zayed University of AI, UAE + +{fahad.shamshad, muzammal.naseer, karthik.nandakumar}@mbzuai.ac.ae + +![](images/84f4e77f6d06dc8ebde7ec482aebe28aac8a0e06def5699e03738271ad09c778.jpg) +Protected Original + +![](images/bf65165568d7efeaaee26804444fe4923be66bba4fb54e90fbbc1c2f0234b563.jpg) +Figure 1. The proposed approach crafts "naturalistic" and transferable text-guided adversarial faces to deceive black-box face recognition systems. First row shows original images that need to be protected and second row shows corresponding protected images along with the user-defined makeup text prompts that guide the adversarial search. Comparison against existing methods is shown in the third row. The yellow text represents the confidence score (higher is better) output by a commercial API (Face++, when matching the protected image against the target identity shown in the bottom right. The reference image used by [22] for makeup transfer is shown at the bottom corner of the corresponding adversarial image. + +![](images/35c8881d6894b6f19f0c6d90b1d9859eb4492ae04adf388b9c4280dcfc386773.jpg) +Target + +# Abstract + +The success of deep learning based face recognition systems has given rise to serious privacy concerns due to their ability to enable unauthorized tracking of users in the digital world. Existing methods for enhancing privacy fail to generate "naturalistic" images that can protect facial privacy without compromising user experience. We propose a novel two-step approach for facial privacy protection that relies on finding adversarial latent codes in the low-dimensional manifold of a pretrained generative model. The first step inverts the given face image into the latent space and finetunes the generative model to achieve an accurate reconstruction of the given image from its latent code. This step produces a good initialization, aiding the generation + +of high-quality faces that resemble the given identity. Subsequently, user-defined makeup text prompts and identity-preserving regularization are used to guide the search for adversarial codes in the latent space. Extensive experiments demonstrate that faces generated by our approach have stronger black-box transferability with an absolute gain of $12.06\%$ over the state-of-the-art facial privacy protection approach under the face verification task. Finally, we demonstrate the effectiveness of the proposed approach for commercial face recognition systems. Our code is available at https://github.com/fahadshamshad/Clip2Protect. + +# 1. Introduction + +Deep learning based face recognition (FR) systems [43, 61] have found widespread usage in multiple applications, + +Table 1. Comparison among different facial privacy protection methods w.r.t. the natural outputs, black box setting, experiments under face verification and identification tasks, unrestricted (semantically meaningful), and more flexible text guided adversaries. + +
Adv-Makeup [71]TIP-IM [70]AMT-GAN [22]Ours
Natural outputsYesPartiallyPartiallyYes
Black boxYesYesYesYes
VerificationYesNoYesYes
IdentificationNoYesNoYes
UnrestrictedYesNoYesYes
Text guidedNoNoNoYes
+ +including security [63], biometrics [38], and criminal investigation [45], outperforming humans in many scenarios [12, 48, 61]. Despite positive aspects of this technology, FR systems seriously threaten personal security and privacy in the digital world because of their potential to enable mass surveillance capabilities [1, 67]. For example, government and private entities can use FR systems to track user relationships and activities by scraping face images from social media profiles such as Twitter, LinkedIn, and Facebook [18, 20]. These entities generally use proprietary FR systems, whose specifications are unknown to the public (black box model). Therefore, there is an urgent need for an effective approach that protects facial privacy against such unknown FR systems. + +An ideal facial privacy protection algorithm must strike the right balance between naturalness and privacy protection [70, 77]. In this context, "naturalness" is defined as the absence of any noise artifacts that can be easily perceived by human observers and the preservation of human-perceived identity. "Privacy protection" refers to the fact that the protected image must be capable of deceiving a black-box malicious FR system. In other words, the protected image must closely resemble the given face image and be artifact-free for a human observer, while at the same time fool an unknown automated FR system. Since failure to generate naturalistic faces can significantly affect user experience on social media platforms, it is a necessary precondition for adoption of a privacy-enhancement algorithm. + +Recent works exploit adversarial attacks [57] to conceal user identity by overlaying noise-constrained (bounded) adversarial perturbations on the original face image [6,53,74]. Since the adversarial examples are generally optimized in the image space, it is often difficult to simultaneously achieve naturalness and privacy [70]. Unlike noise-based methods, unrestricted adversarial examples are not constrained by the magnitude of perturbation in the image space and have demonstrated better perceptual realism for human observers while being adversially effective [3,55,68,76]. + +Several efforts have been made to generate unrestricted adversarial examples that mislead FR systems (see Tab. 1) [22, 25, 39, 72]. Among these, adversarial makeup based methods [22, 72] are gaining increasing attention as they + +can embed adversarial modifications in a more natural way. These approaches use generative adversarial networks [15] (GANs) to adversarily transfer makeup from a given reference image to the user's face image while impersonating a target identity. However, existing techniques based on adversarial makeup transfer have the following limitations: (i) adversarial toxicity in these methods hamper the performance of the makeup transfer module, thereby resulting in unnatural faces with makeup artifacts (see Fig. 1); (ii) the use of a reference image to define the desired makeup style affects the practicality of this approach; (iii) for every new target identity, these approaches require end-to-end retraining from scratch using large makeup datasets; and (iv) most of these methods primarily aim at impersonation of the target identity, whereas the desired privacy objective is dodging, i.e., multiple images of the user's face scraped from different social media sites must not match with each other. + +To mitigate the above problems, we propose a new approach to protect user facial privacy on online platforms (Sec. 3). The proposed approach aims to search for adversarial latent codes in a low-dimensional manifold learned by a generative model trained to generate face images [2, 27]. Our main contributions are: + +- Facial Privacy-protection Framework Using Adversarial Latent Codes: Given a face image, we propose a novel two-step method to search for adversarial latent codes, which can be used by a generative model (e.g., StyleGAN) to produce face images with high visual quality that matches human-perceived identity, while deceiving black-box FR systems. +- Adversarial Makeup Transfer using Textual Prompts: A critical component of the above framework is a technique for leveraging user-defined textual (makeup) prompts to traverse over the latent manifold of the generative model and find transferable adversarial latent codes. Our approach effectively hides attack information in the desired makeup style, without the need for any large makeup dataset or retraining of models for different target identities. +- Identity Preserving Regularization: We propose a regularizer that preserves identity-related attributes within the latent space of the generative model and ensures that the protected face image visually resembles the original face. + +Extensive experiments (Sec. 4.1) for both face verification and identification scenarios demonstrate the effectiveness of our approach against black-box FR models and online commercial facial recognition APIs (Sec. 4.2). Furthermore, we provide detailed ablative analysis to dissect the performance of different components of our approach (Sec. 4.3). + +# 2. Related Work + +Obfuscation Methods: Obfuscation is the most widely used technique [38] to protect user's facial privacy. Earlier obfuscation approaches typically degrade the quality of the original face image by applying simple operations such as masking [52, 64], filtering [33, 78], and image transformations [8, 36, 62]. While these relatively simple obfuscation techniques are reasonable for surveillance applications, they are ill-suited for online/social media platforms where user experience is critical [41]. Though deep learning based obfuscation approaches generate more realistic images [4, 7, 56, 58], they often result in a change of identity compared to the original image and occasionally produce undesirable artifacts [30, 31, 34]. + +Noise-based Adversarial Examples: Adversarial attacks have been used to protect users from unauthorized FR models. Some methods [6, 53] rely on data poisoning to deceive targeted FR models, but are less practical because access to the training data or the gallery set of the unknown FR system is often not available. Other approaches have used game-theory perspective [42] in white-box settings or person-specific privacy masks (one mask per person) to generate protected images at the cost of acquiring multiple images of the same user [77]. In contrast, we aim to fool the black box FR model using only single image. In TIP-IM [70], targeted optimization was used to generate privacy masks against unknown FR models by introducing a naturalness constraint. While this approach provides effective privacy, it generates output images with perceptible noises that can affect the user experience [70]. + +Unrestricted Adversarial Examples: Unrestricted adversarial attacks (UAAs) are not constrained by the perturbation norm and can induce large but semantically meaningful perturbations. These attacks have been extensively studied in image classification literature [3, 35, 55, 68, 73, 76] and it has been shown that outputs generated via UAAs are less perceptible to human observers as compared to noise-based adversarial attacks. Motivated by this observation, patch-based unrestricted attacks have been proposed to generate wearable adversarial accessories like colorful glasses [54], hat [29] or random patch [69] to fool the FR model, but such synthesized patches generally have weak transferability due to the limited editing region and the large visible pattern compromises naturalness and affects user experience. Recently, generative models [24, 50] have been leveraged to craft UAAs against FR models. However, these generative approaches are either designed for the white-box settings [46, 79] or show limited performance in query-free black-box settings [25]. Makeup-based UAAs [17, 72] have also been proposed against FR systems by embedding the perturbations into a natural makeup effect. These makeup based attacks have also been exploited to protect the user privacy by applying adversarial makeup on the user + +face image [22]. However, interference between adversarial perturbations and makeup transfer can produce undesirable makeup artifacts in the output images. Moreover, these attacks generally assume access to large makeup datasets for training models and require a reference makeup image. In contrast, our approach finds adversarial faces on the natural image manifold in black-box setting via guidance from makeup text prompt, which makes it less susceptible to artifacts (see Fig. 1) and more practical. + +Vision-Language Modelling: Cross-modal vision-language modelling has attracted significant attention in recent years [13]. OpenAI introduced CLIP [47] that is trained on 400 million image-text pairs using contrastive objective and maps both image and text in a joint multimodal embedding space. With powerful representation embedding of CLIP, several methods have been proposed to manipulate images with text-guidance. StyleCLIP [44] and DiffusionCLIP [28, 40] leverage the powerful generative capabilities of StyleGAN and diffusion models to manipulate images with text prompts. Other similar works include HairCLIP [66], CLIP-NeRF [60], CLIPstyler [32], and CLIPDraw [14]. While these methods focus on the text-guidance ability of CLIP, our approach aims to find the adversarial latent codes in a generative model's latent space for privacy protection against black-box FR models. + +# 3. Proposed Approach for Facial Privacy + +Our goal is to protect user facial privacy on online platforms against unknown (black-box) FR models without compromising on the user's online experience. The proposed approach finds protected faces by adversarially exploring the low-dimensional latent space of a pretrained generative model that is trained on natural face images. To avoid artifacts in the protected image, we restrict the search for adversarial faces close to the clean image manifold learned by the generative model. Moreover, we propose to optimize only over identity-preserving latent codes in the latent space. This effectively preserves human-perceived identity during attack while offering high privacy against automated systems. Further, we employ natural makeup-like perturbations via guidance from a text prompt, which provides more flexibility to the user compared to reference image-based adversarial makeup transfer [22]. + +# 3.1. Preliminaries + +Let $\pmb{x} \in \mathcal{X} \subset \mathbb{R}^n$ denote the given original/real face image. Let $f(\pmb{x}) : \mathcal{X} \to \mathbb{R}^d$ be a FR model that extracts a fixed-length normalized feature representation. Let $\mathcal{D}(\pmb{x}_1, \pmb{x}_2) = D(f(\pmb{x}_1), f(\pmb{x}_2))$ be a distance metric that measures the dissimilarity between two face images $\pmb{x}_1$ and $\pmb{x}_2$ based on their respective representations $f(\pmb{x}_1)$ and $f(\pmb{x}_2)$ . Generally a FR system can operate in two modes: verification and identification. A face verification + +![](images/d22a9ba3df87594c2ff4ea2e1367252711931d277ddbfbe0fbd87c1f409c524f.jpg) +Figure 2. Overall pipeline of the proposed approach to protect users facial privacy. Our proposed approach searches for the adversarial latent codes on the generative manifold to reconstruct an adversarial face that is capable of fooling unknown FR systems for privacy protection. Our approach allows "makeup" editing in an adversarial manner through user defined textual prompts and thereby enhance the user's online experience. Our text-guided objective searches for such latent codes while keeping the original identity preserved. + +system predicts that two faces belong to the same identity if $\mathcal{D}(\pmb{x}_1, \pmb{x}_2) \leq \tau$ , where $\tau$ is the system threshold. On the other hand, a (closed set) face identification system compares the input image (probe) against a set of face images (gallery) and outputs the identity whose representation is most similar to that of the probe. Since the attacker can employ verification or identification to determine the user identity using black-box FR models, a protection approach should conceal the user's identity in both scenarios. + +User privacy can be protected by misleading the malicious FR model through impersonation or dodging attacks. In the context of verification, impersonation (false match) implies that the protected face matches with the face of a specific target identity and dodging (false non-match) means that the protected face does not match with some other image of the same person. Similarly, for face identification, impersonation ensures that the protected image gets matched to a specified target identity in the gallery set, while dodging prevents the protected face from matching with images of the same person in the gallery. + +Problem Statement: Given the original face image $\pmb{x}$ , our goal is to generate a protected face image $\pmb{x}^p$ such that $\mathcal{D}(\pmb{x}^p, \pmb{x})$ is large (for successful dodging attack) and $\mathcal{D}(\pmb{x}^p, \pmb{x}^t)$ is small (for successfully impersonating a target face $\pmb{x}^t$ ), where $\mathcal{O}(\pmb{x}) \neq \mathcal{O}(\pmb{x}^t)$ and $\mathcal{O}$ is the oracle that gives the true identity labels. At the same time, we want to minimize $\mathcal{H}(\pmb{x}^p, \pmb{x})$ , where $\mathcal{H}$ quantifies the degree of unnaturalness introduced in the protected image $\pmb{x}^p$ in relation to the original image $\pmb{x}$ . Formally, the optimization problem that we aim to solve is: + +$$ +\begin{array}{l} \min _ {\boldsymbol {x} ^ {p}} \mathcal {L} (\boldsymbol {x} ^ {p}) = \mathcal {D} (\boldsymbol {x} ^ {p}, \boldsymbol {x} ^ {t}) - \mathcal {D} (\boldsymbol {x} ^ {p}, \boldsymbol {x}) \tag {1} \\ \mathrm {s . t .} \mathcal {H} (\boldsymbol {x} ^ {p}, \boldsymbol {x}) \leq \epsilon \\ \end{array} +$$ + +![](images/2f7422eee5a7f7e84514603fc2fbbb4edb25e2397277343c442a1b4f18c6235b.jpg) +(a) Original + +![](images/d881a1341c142b5295fe4413d8f2b62a3f46bf4fa85b858312701c8b0020eaee.jpg) +(b) Encoder Inversion + +![](images/6a8bc6961077e8abed5b5570bff0e24b561455f7b2f07958f9ec3ac193baf6f9.jpg) +(c) Generator finetuning +Figure 3. Generator finetuning allows near-perfect reconstructions of LFW dataset sample. This is crucial for the online experience of users. Matching scores returned by Face++ API are 62.38 and 98.96 for encoder and generator-finetuned inversions, respectively. + +where $\epsilon$ is a bound on the adversarial perturbation. For noise-based approach, $\mathcal{H}(\pmb{x}^p,\pmb{x}) = \| \pmb{x} - \pmb{x}^p\|_p$ , where $\| \cdot \|_p$ denotes the $L_p$ norm. However, direct enforcement of the perturbation constraint leads to visible artifacts, which affects visual quality and user experience. Constraining the solution search space to a natural image manifold using an effective image prior can produce more realistic images. Note that the distance metric $\mathcal{D}$ is unknown since our goal is to deceive a black-box FR system. + +# 3.2. Makeup Text-Guided Adversarial Faces + +Our approach restricts the solution space of the protected face $\pmb{x}^p$ to lie close to the clean face manifold $\mathcal{X}$ . This manifold can be learned using a generative model trained on real human faces. Specifically, let $G_{\theta}(\pmb{w}) : \mathcal{W} \to \mathbb{R}^n$ denote the pretrained generative model with weights $\theta$ , where $\mathcal{W}$ is the latent space. Our proposed approach consists of two stages: (i) latent code initialization (Sec. 3.2.1) and (ii) text-guided adversarial optimization (Sec. 3.2.2). The overall pipeline of the proposed approach is shown in Fig. 2. + +# 3.2.1 Latent Code Initialization + +The latent code initialization stage is based on GAN inversion, which aims to invert the original image $\pmb{x}$ into the latent space $\mathcal{W}$ , i.e., find a latent code $\pmb{w}_{\mathrm{inv}} \in \mathcal{W}$ such that $\pmb{x}_{\mathrm{inv}} = G_{\theta}(\pmb{w}_{\mathrm{inv}}) \approx \pmb{x}$ . To achieve this, we first use an encoder-based inversion called e4e [59] to infer $\pmb{w}_{\mathrm{inv}}$ in $\mathcal{W}$ from $\pmb{x}$ i.e., $\pmb{w}_{\mathrm{inv}} = I_{\phi}(\pmb{x})$ , where $I_{\phi}: \mathcal{X} \to \mathcal{W}$ is the pretrained encoder with weights $\phi$ (see Fig. 2). + +We use StyleGAN trained on a high-resolution dataset of face images as the pretrained generative model $G_{\theta}$ due to its powerful synthesis ability and the disentangled structure of its latent space. A significant challenge during inversion is preserving the identity of the original image i.e., $\mathcal{O}(\pmb{x}) = \mathcal{O}(\pmb{x}_{\mathrm{inv}})$ . Generally, optimization and encoder-based inversion approaches struggle to preserve identity after reconstruction [49] (see Fig. 3b). Moreover, when using these approaches, the inversion error can be large for out-of-domain face images with extreme poses and viewpoints, which are quite common in social media applications. Therefore, these approaches cannot be applied directly to invert $\pmb{x}$ . Instead, motivated by the recent observation [49] that slight changes to the pretrained generator weights do not harm its editing abilities while achieving near-perfect reconstructions, we finetune the pretrained generator weights $\theta$ instead of the encoder weights $\phi$ . Specifically, we fix $w_{\mathrm{inv}} = I_{\phi}(\pmb{x})$ and fine-tune $G_{\theta}$ using the following loss: + +$$ +\theta^ {*} = \underset {\theta} {\arg \min } \mathcal {L} _ {\mathrm {L P I P S}} (\boldsymbol {x}, G _ {\theta} (\boldsymbol {w} _ {\mathrm {i n v}})) + \lambda_ {2} \mathcal {L} _ {2} (\boldsymbol {x}, G _ {\theta} (\boldsymbol {w} _ {\mathrm {i n v}})), +$$ + +where $\mathcal{L}_{\mathrm{LPIPS}}$ is the perceptual loss and $\mathcal{L}_2$ denotes the pixelwise similarity. The final inverted image $x_{\mathrm{inv}}^*$ (see Fig. 3c) can be obtained by performing a forward pass of $\pmb{w}_{\mathrm{inv}}$ through fine-tuned generator i.e., $\pmb{x}_{\mathrm{inv}}^{*} = G_{\theta^{*}}(\pmb{w}_{\mathrm{inv}})$ . + +# 3.2.2 Text-guided adversarial optimization + +Given the inverted latent code $\boldsymbol{w}_{\mathrm{inv}}$ and fine-tuned generator $G_{\theta^{*}}(\cdot)$ , our goal is to adversarially perturb this latent code $\boldsymbol{w}_{\mathrm{inv}}$ in the low-dimensional generative manifold $\mathcal{W}$ to generate a protected face that fools the black-box FR model, while imitating the makeup style of the text prompt $t_{\mathrm{makeup}}$ . + +To achieve these objectives, we investigate the following questions: (i) how to effectively extract makeup style information from $t_{\mathrm{makeup}}$ and apply it to the face image $x$ in an adversarial manner?, (ii) how to regularize the optimization process so that the output face image is not qualitatively impaired?, (iii) how to craft effective adversarial perturbations that mislead black-box FR models?, and (iv) how to preserve the human-perceived identity $\mathcal{O}(x)$ of the original face image while ensuring high privacy? + +The first issue can be addressed by aligning the output adversarial image with the text prompt $t_{\mathrm{makeup}}$ in the embedding space of a pretrained vision-language model. The + +second issue is addressed by enforcing the adversarial latent code to remain close to initialization $\boldsymbol{w}_{\mathrm{inv}}$ . The third issue is solved by crafting transferable text-guided adversarial faces on a white-box surrogate model (or an ensemble of models) with the goal of boosting the fooling rate on the black-box FR model. Finally, we leverage the disentangled nature of latent space in the generative model and incorporate an identity-preserving regularization to effectively maintain the original visual identity. We now present the details of the loss functions used to incorporate the above ideas. + +Textual Loss: A key ingredient of the proposed approach is text-based guidance to inconspicuously hide the adversarial perturbations into the makeup effect. This can be naively achieved by aligning the representation of $t_{\mathrm{makeup}}$ and the adversarial face $G_{\theta^*}(\boldsymbol{w})$ in the common embedding space of a pre-trained vision-language model (e.g. CLIP [47]). However, this approach will transform the whole output image to follow the makeup style of $t_{\mathrm{makeup}}$ , which results in low diversity. Therefore, we use a directional CLIP loss that aligns the CLIP-space direction between the text-image pairs of the original and adversarial images. Specifically, + +$$ +\mathcal {L} _ {\mathrm {c l i p}} = 1 - \frac {\Delta I \cdot \Delta T}{| \Delta I | | \Delta T |}, \tag {2} +$$ + +where $\Delta T = E_{T}(t_{\mathrm{makeup}}) - E_{T}(t_{\mathrm{src}})$ and $\Delta I = E_I(G_{\theta^*}(\pmb {w})) - E_I(\pmb {x})$ . Here, $E_{T}$ and $E_{I}$ are the text and image encoders of the CLIP model and $t_\mathrm{src}$ is the semantic text of the input image $\pmb{x}$ . Since we are dealing with faces, $t_\mathrm{src}$ can be simply set as "face". This loss localizes makeup transfer (e.g. red lipstick) without affecting privacy. + +Adversarial Loss: Our goal is to traverse over the latent space $\mathcal{W}$ to find adversarial latent codes on the generative manifold whose face feature representation lies close to that of target image and far away from the original image itself i.e., $\mathcal{D}(\pmb{x}^p, \pmb{x}) > \mathcal{D}(\pmb{x}^p, \pmb{x}^t)$ . Hence, the adversarial loss is: + +$$ +\mathcal {L} _ {\mathrm {a d v}} = \mathcal {D} \left(G _ {\theta^ {*}} (\boldsymbol {w}), \boldsymbol {x} ^ {t}\right) - \mathcal {D} \left(G _ {\theta^ {*}} (\boldsymbol {w}), \boldsymbol {x}\right), \tag {3} +$$ + +where $\mathcal{D}(\pmb{x}_1,\pmb{x}_2) = 1 - \cos [f(\pmb{x}_1),f(\pmb{x}_2)]$ is the cosine distance. Since the malicious FR model is unknown in the black-box setting, Eq. 3 cannot be solved directly. Instead, following AMT-GAN [22], we perform adversarial optimization on an ensemble of white-box surrogate models to imitate the decision boundary of the unknown FR model. + +Identity Preservation Loss: The optimization over the generative manifold ensures that the protected image $\pmb{x}^p$ is natural i.e., artifact-free, however, it does not explicitly enforce the protected image to preserve the identity of the original image with respect to the human observer. To mitigate the issue, we take advantage of the semantic control exhibited by StyleGAN in its latent space. The latent code $\pmb{w} \in \mathcal{W}$ impacts image generation by controlling different level of semantics in the output image. Specifically, latent + +codes corresponding to the initial layers of StyleGAN control high-level aspects such as pose, general hairstyle, and face shape [27]. Adversarially perturbing these latent layers can change these attributes, resulting in a change of identity (see Sec. 4.3). Latent codes corresponding to deeper layers of StyleGAN are associated with fine-level control such as makeup style [2]. Therefore, we perturb only those latent codes associated with deeper layers of StyleGAN, thereby restricting the adversarial faces to the identity preserving manifold. We further constrain the latent code to stay close to its initial value $w_{\mathrm{inv}}$ using the following regularization: + +$$ +\mathcal {L} _ {\text {l a t e n t}} = \left\| (\boldsymbol {w} \odot \boldsymbol {m} _ {i d}) - \left(\boldsymbol {w} _ {\text {i n v}} \odot \boldsymbol {m} _ {i d}\right) \right\| _ {2}, \tag {4} +$$ + +where $\odot$ denotes element-wise product and $m_{id}$ is an identity preservation mask that is 0 for the initial layers and 1 only for the deeper layers of the latent code. StyleGAN has 18 layers, each having a dimension of 512. The identity preservation mask is set to 1 only from layer 8 to 18. Finally, combining the three loss functions, we have + +$$ +\mathcal {L} _ {\text {t o t a l}} = \lambda_ {\text {a d v}} \mathcal {L} _ {\text {a d v}} + \lambda_ {\text {c l i p}} \mathcal {L} _ {\text {c l i p}} + \lambda_ {\text {l a t e n t}} \mathcal {L} _ {\text {l a t e n t}}, \tag {5} +$$ + +where $\lambda_{\mathrm{adv.}}$ , $\lambda_{\mathrm{clip}}$ , and $\lambda_{\mathrm{latent}}$ are hyperparameters. Note that $\mathcal{L}_{\mathrm{adv}}$ accounts for the adversarial objective in Eq. 1, while the text-guided makeup transfer $(\mathcal{L}_{\mathrm{clip}})$ and identity-preserving regularization $(\mathcal{L}_{\mathrm{latent}})$ implicitly enforce the naturalness constraint in Eq. 1. + +# 4. Experiments + +Implementation details: In all experiments, we use StyleGAN2 pretrained on the FFHQ face dataset as our generative model. For adversarial text guidance, we use a vision transformer-based CLIP model. For generator fine-tuning in the latent code initialization step, we use 450 iterations with value of $\lambda_{2}$ in Eq. 2 set to 0.5. For the makeup text input, we collect 40 text prompts based on the makeup style of diverse nature (details in supplementary material). For adversarial optimization, we use an Adam optimizer with $\beta_{1}$ and $\beta_{2}$ set to 0.9 and 0.999, respectively, and a learning rate of 0.01. We run the optimizer for 50 iterations to craft protected faces. We set the value of $\lambda_{\mathrm{adv}}$ , $\lambda_{\mathrm{clip}}$ , and $\lambda_{\mathrm{latent}}$ to 1, 0.5, and 0.01, respectively. All our experiments are conducted on a A100 GPU with 40 GB memory. + +Datasets: We perform experiments for both face verification and identification settings. Face verification: We use CelebA-HQ [26] and LADN [16] for the impersonation attack. We select subset of 1,000 images from CelebA-HQ and report average results over 4 target identities provided by [22]. Similarly, for LADN, we divide the 332 images available into 4 groups, where images in each group aim to impersonate the target identities provided by [22]. For dodging attack, we use CelebA-HQ [26] and LFW [23] datasets. Specifically, we select 500 subjects at random and + +each subject has a pair of faces. Face identification: For impersonation and dodging, we use CelebA-HQ [26] and LFW [23] as our evaluation set. For both datasets, we randomly select 500 subjects, each with a pair of faces. We assign one image in the pair to the gallery set and the other to the probe set. Both impersonation and dodging attacks are performed on the probe set. For impersonation, we insert 4 target identities provided by [22] into the gallery set. A more detailed description of all datasets and pre-processing steps is provided in the supplementary material. + +Target Models: We aim to protect user facial privacy by attacking four FR model with diverse back bones in the black-box settings. The target models include IRSE50 [21], IR152 [9], FaceNet [51], and MobileFace [5]. Following standard protocol, we align and crop the face images using MTCNN [75] before giving them as input to FR models. Further, we also report privacy protection performance based on commercial FR API including Face++ and Tencent Yunshentu FR platforms. + +Evaluation metrics: Following [70], we use protection success rate (PSR) to evaluate the proposed approach. PSR is defined as the fraction of protected faces missclassified by the malicious FR system. To evaluate PSR, we use the thresholding and closed set strategies for face verification and identification, respectively. For face identification, we also use Rank-N targeted identity success rate (Rank-N-T) and untargeted identity success rate (Rank-N-U), where Rank-N-T means that target image $\boldsymbol{x}^t$ will appear at least once in the top N candidates shortlisted from the gallery and Rank-N-U implies that the top N candidate list does not have the same identity as that of original image $\boldsymbol{x}$ . We also report results of PSNR (dB), SSIM, and FID [19] scores to evaluate the imperceptibility of method. Large PSNR and SSIM [65] indicates better match with the original images, while low FID score indicates more realistic images. For commercial APIs, we directly report the confidence score returned by the respective servers. + +Baseline methods: We compare our approach with recent noise-based and makeup based facial privacy protection approaches. Noise based methods include PGD [37], MI-FGSM [10], TI-DIM [11], and TIP-IM [70], whereas makeup-based approaches are Adv-Makeup [71] and AMTGAN [22]. We want to highlight that TIP-IM and AMTGAN are considered the state-of-the-art (SOTA) for face privacy protection against black-box FR systems in noise-based and unrestricted settings, respectively. TIP-IM also incorporate multi-target objective in its optimization to find the optimal target image among multiple targets. For fair comparison, we use its single target variant. + +# 4.1. Experimental Results + +In this section, we present experimental results of our approach in black-box settings on four different pretrained + +Table 2. Protection success rate (PSR %) of black-box impersonation attack under the face verification task. For each column, the other three FR systems are used as surrogates to generate the protected faces. + +
MethodCelebA-HQLADN-DatasetAverage
IRSE50IR152FaceNetMobileFaceIRSE50IR152FaceNetMobileFace
Clean7.293.801.0812.682.713.610.605.114.61
Inverted5.572.770.6013.326.804.510.2511.665.68
PGD [37]36.8720.681.8543.9940.0919.593.8241.0925.60
MI-FGSM [10]45.7925.032.5845.8548.9025.576.3145.0130.63
TI-DIM [11]63.6336.1715.3057.1256.3634.1822.1148.3041.64
Adv-Makeup(ICAI'21) [71]21.959.481.3722.0029.6410.030.9722.3814.72
TIP-IM(ICCV'21) [70]54.4037.2340.7448.7265.8943.5763.5046.4850.06
AMT-GAN(CVPR'22) [22]76.9635.1316.6250.7189.6449.1232.1372.4352.84
Ours81.1048.4241.7275.2691.5753.3147.9179.9464.90
+ +Table 3. Protection success rate (PSR %) of black-box dodging (top) and impersonation (bottom) attacks under the face identification task for LFW dataset [23]. For each column, the other three FR systems are used as surrogates to generate the protected faces. R1-U: Rank-1-Untargeted, R5-U: Rank-5-Untargeted, R1-T: Rank-1-Targeted, R5-T: Rank-5-Targeted. + +
MethodIRSE50IR152FaceNetMobileFaceAverage
R1-UR5-UR1-UR5-UR1-UR5-UR1-UR5-UR1-UR5-U
MI-FGSM [10]70.242.658.441.859.234.068.047.263.941.4
TI-DIM [11]79.051.267.454.074.452.079.261.675.054.7
TIP-IM(ICCv'21) [70]81.452.271.854.676.049.882.263.077.854.9
Ours86.659.473.456.683.851.285.066.882.258.5
R1-TR5-TR1-TR5-TR1-TR5-TR1-TR5-TR1-TR5-T
MI-FGSM [10]4.010.23.214.29.018.88.422.46.1516.4
TI-DIM [11]4.013.67.819.618.032.821.639.012.8526.25
TIP-IM(ICCv'21) [70]8.028.211.631.225.256.834.051.419.741.9
Ours11.237.816.051.227.454.039.061.223.451.05
+ +
MethodFID ↓PSR Gain ↑
Adv-Makeup [71]4.230
TIP-IM [70]38.7335.34
AMT-GAN [22]34.4438.12
Ours26.6250.18
+ +Table 4. FID comparison. PSR Gain is absolute gain in PSR relative to AdvMakeup. + +FR models under face verification and identification tasks. To generate protected images, we use three FR models as a surrogate to imitate the decision boundary of the fourth FR model. All results are averaged over 5 text based makeup styles that are provided in the supplementary material. + +For face verification experiments, we set the system threshold value at 0.01 false match rate for each FR model i.e., IRSE50 (0.241), IR152 (0.167), FaceNet (0.409), and MobileFace (0.302). Quantitative results in terms of PSR for impersonation attack under the face verification task are shown in Tab. 2. Our approach is able to achieve an average absolute gain of about $12\%$ and $14\%$ over SOTA unrestricted [22] and noise-based [70] facial privacy protection methods, respectively. Qualitative results are shown in Fig. 1 which shows that protected faces generated by our approach are more realistic. Results for dodging attacks under face verification are provided in the supplementary material. In Tab. 3, we also provide PSR vales under the face identification task for dodging (untargeted) and impersonation attacks. Our approach consistently outperforms recent + +methods at both Rank-1 and Rank-5 settings. We emphasize that we are the first to show effectiveness of generative models in offering untargeted privacy protection (dodging) in a more practical identification setting. Since AMT-GAN and Adv-Makeup are originally trained to impersonate target identity under the verification task, we have not included them in Tab. 3. Qualitative results for LFW and CelebA are provided in the supplementary material. + +We report FID scores (lower is better) of our approach in Tab. 4 for CelebA and LADN datasets to measure naturalness. Adv-Makeup has the lowest FID score as it only applies makeup to the eye region without changing the rest of the face. However, this kind of restriction results in poor PSR. Our method has lower FID compared to TIP-IM and AMT-GAN and achieves the highest PSR. We provide PSNR and SSIM results in the supplementary material. + +# 4.2. Effectiveness in Real-World Applications + +We now show the effectiveness of our approach to protect facial images (through targeted impersonation) against commercial API such as Face++ and Tencent Yunshentu FR platform operating in the verification mode. These APIs return confidence scores between 0 to 100 to measure whether two images are similar or not, where a high confidence score indicates high similarity. As the training data and model parameters of these propriety FR models are unknown, it + +![](images/015044b84574475d588ab89a5e52e746f610eca9ca86988dade41c43919170f7.jpg) +Figure 4. Average confidence score (higher is better) returned by a real-world face verification API, Face++, for impersonation attack. Our approach has a higher confidence score than state-of-the-art makeup and noise-based facial privacy protection methods. + +Table 5. Impact of ${\lambda }_{\text{latent }}$ on FID score and PSR. + +
λlatent0.50.10.050.010.0050.00010
FID11.621.425.227.830.138.443.2
PSR (%)31.239.057.476.283.890.093.6
+ +effectively mimics a real-world scenario. We protect 100 faces randomly selected from CelebA-HQ using the baselines and the proposed method. In Fig. 4, we show the average confidence score returned by Face++ against these images. These results indicate that our method has a high PSR compared to baselines. We defer more details and results for Tencent Yunshentu API to supplementary material. + +# 4.3. Ablation Studies + +Next, we report some ablations to evaluate the contributions of our loss components. + +Makeup based text guidance: As shown in Fig. 5 (top), in the absence of text guidance, resulting images may contain artifacts due to increased perturbations induced by the adversarial objective. Text-guidance effectively hides the perturbations in the makeup, leading to more natural looking images. It also provides the user more flexibility to select a desired makeup style compared to a reference image. + +Identity preserving regularization: Optimizing over the whole latent space provides more degrees of freedom and increases the PSR. However, it does not explicitly enforce adversarial optimization to preserve the user identity as shown in Fig. 5 (bottom). The proposed identity preserving regularization effectively preserves identity, while imitating the desire makeup style. + +Impact of latent loss weight: Decreasing the weight assigned to the latent loss $\lambda_{\mathrm{latent}}$ results in an increase in both the FID score and PSR (and vice versa). Allowing the latent to deviate more from the initial inverted latent code of the given face image often results in artifacts caused by the adversarial loss, degrading naturalness but aiding privacy. + +Robustness against textual variations. Finally, we evaluate the impact of different textual styles on the PSR. We + +![](images/c7f83ace38d139841302d33c12412c2033b4e3fc897af1ce0d93cac392098f4e.jpg) +Figure 5. Top: Effect of makeup-based text guidance on the visual quality of the output images. Output images are able to impersonate the target identity for face verification. Text-prompt is "tanned makeup with red lipstick". Bottom: Optimizing over all latent codes changes the identity of the protected image. Our identity-preserving regularization enforces the adversarial optimization to search for latent codes that hide the perturbations in the makeup effect while simultaneously preserving visual identity. + +Table 6. Impact of different textual makeup styles on PSR. Makeup styles are "tanned", "pale", "pink eyeshadows", "red lipstick", and "Matte". Std. denotes standard deviation. + +
t1makeupt2makeupt3makeupt4makeupt5makeupStd.
PSR74.177.378.478.779.21.24
+ +select five text-based makeup styles to protect 1000 images of CelebA-HQ using our method. Results in Tab. 6 show that PSR does not change significantly (low standard deviation) for different makeup styles, indicating robustness of our approach wrt different text-based makeup styles. + +# 5. Conclusion + +We have proposed a framework to protect privacy of face images on online platforms by carefully searching for adversarial codes in the low-dimensional latent manifold of a pre-trained generative model. We have shown that incorporating a makeup text-guided loss and an identity preserving regularization effectively hides the adversarial perturbations in the makeup style, provides images with high quality, and preserves human-perceived identity. While this approach is robust to the user-defined text-prompt and target identity, it would be beneficial if the text-prompt and target identity can be automatically selected based on the given face image. Limitations of our method include high computational cost at the time of protected face generation. + +# References + +[1] Shane Ahern, Dean Eckles, Nathaniel S Good, Simon King, Mor Naaman, and Rahul Nair. Over-exposed? privacy patterns and considerations in online and mobile photo sharing. In Proceedings of the SIGCHI conference on Human factors in computing systems, pages 357-366, 2007. 2 +[2] Amit H Bermano, Rinon Gal, Yuval Alaluf, Ron Mokady, Yotam Nitzan, Omer Tov, Oren Patashnik, and Daniel Cohen-Or. State-of-the-art in the architecture, methods and applications of stylegan. In Computer Graphics Forum, volume 41, pages 591–611. Wiley Online Library, 2022. 2, 6 +[3] Anand Bhattachad, Min Jin Chong, Kaizhao Liang, Bo Li, and David A Forsyth. Unrestricted adversarial examples via semantic manipulation. arXiv preprint arXiv:1904.06347, 2019. 2, 3 +[4] Jia-Wei Chen, Li-Ju Chen, Chia-Mu Yu, and Chun-Shien Lu. Perceptual indistinguishability-net (pi-net): Facial image obfuscation with manipulable semantics. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6478–6487, 2021. 3 +[5] Sheng Chen, Yang Liu, Xiang Gao, and Zhen Han. Mobile-facenets: Efficient cnns for accurate real-time face verification on mobile devices. In Chinese Conference on Biometric Recognition, pages 428-438. Springer, 2018. 6 +[6] Valeriia Cherepanova, Micah Goldblum, Harrison Foley, Shiyuan Duan, John P Dickerson, Gavin Taylor, and Tom Goldstein. Lowkey: Leveraging adversarial attacks to protect social media users from facial recognition. In International Conference on Learning Representations, 2020. 2, 3 +[7] William L Croft, Jörg-Rüdiger Sack, and Wei Shi. Differentially private facial obfuscation via generative adversarial networks. Future Generation Computer Systems, 129:358-379, 2022. 3 +[8] Ali Dabouei, Sobhan Soleymani, Jeremy Dawson, and Nasser Nasrabadi. Fast geometrically-perturbed adversarial faces. In 2019 IEEE Winter Conference on Applications of Computer Vision (WACV), pages 1979-1988. IEEE, 2019. 3 +[9] Jiankang Deng, Jia Guo, Niannan Xue, and Stefanos Zafeiriou. Arcface: Additive angular margin loss for deep face recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4690-4699, 2019. 6 +[10] Yinpeng Dong, Fangzhou Liao, Tianyu Pang, Hang Su, Jun Zhu, Xiaolin Hu, and Jianguo Li. Boosting adversarial attacks with momentum. In Proceedings of the 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR'18), pages 9185-9193, 2018. 6, 7 +[11] Yinpeng Dong, Tianyu Pang, Hang Su, and Jun Zhu. Evading defenses to transferable adversarial examples by translation-invariant attacks. In Proceedings of the 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR'19), pages 4312-4321, 2019. 6, 7 +[12] Hang Du, Hailin Shi, Dan Zeng, Xiao-Ping Zhang, and Tao Mei. The elements of end-to-end deep face recognition: A survey of recent advances. ACM Computing Surveys (CSUR), 54(10s):1-42, 2022. 2 + +[13] Yifan Du, Zikang Liu, Junyi Li, and Wayne Xin Zhao. A survey of vision-language pre-trained models. arXiv preprint arXiv:2202.10936, 2022. 3 +[14] Kevin Frans, Lisa B Soros, and Olaf Witkowski. Clipdraw: Exploring text-to-drawing synthesis through language-image encoders. arXiv preprint arXiv:2106.14843, 2021. 3 +[15] Ian J. Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron C. Courville, and Yoshua Bengio. Generative adversarial networks. CoRR, abs/1406.2661, 2014. 2 +[16] Qiao Gu, Guanzhi Wang, Mang Tik Chiu, Yu-Wing Tai, and Chi-Keung Tang. Ladin: Local adversarial disentangling network for facial makeup and de-makeup. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10481-10490, 2019. 6 +[17] Nitzan Guetta, Asaf Shabtai, Inderjeet Singh, Satoru Momiyama, and Yuval Elovici. Dodging attack using carefully crafted natural makeup. arXiv preprint arXiv:2109.06467, 2021. 3 +[18] Rebecca Heilweil. The world's scariest facial recognition company explained. Vox, May, 8, 2020. 2 +[19] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 6 +[20] Kashmir Hill. The secretive company that might end privacy as we know it. *The New York Times*, 18:2020, 2020. 2 +[21] Jie Hu, Li Shen, and Gang Sun. Squeeze-and-excitation networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7132-7141, 2018. 6 +[22] Shengshan Hu, Xiaogeng Liu, Yechao Zhang, Minghui Li, Leo Yu Zhang, Hai Jin, and Libing Wu. Protecting facial privacy: Generating adversarial identity masks via style-robust makeup transfer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15014–15023, 2022. 1, 2, 3, 5, 6, 7 +[23] Gary B Huang, Marwan Mattar, Tamara Berg, and Eric Learned-Miller. Labeled faces in the wild: A database for studying face recognition in unconstrained environments. In Workshop on faces in 'Real-Life' Images: detection, alignment, and recognition, 2008. 6, 7 +[24] Phillip Isola, Jun-Yan Zhu, Tinghui Zhou, and Alexei A Efros. Image-to-image translation with conditional adversarial networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1125-1134, 2017. 3 +[25] Kazuya Kakizaki and Kosuke Yoshida. Adversarial image translation: Unrestricted adversarial examples in face recognition systems. arXiv preprint arXiv:1905.03421, 2019. 2, 3 +[26] Tero Karras, Timo Aila, Samuli Laine, and Jaakko Lehtinen. Progressive growing of gans for improved quality, stability, and variation. In International Conference on Learning Representations, 2018. 6 +[27] Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of stylegan. In Proceedings of + +the IEEE/CVF conference on computer vision and pattern recognition, pages 8110-8119, 2020. 2, 6 +[28] Gwanghyun Kim, Taesung Kwon, and Jong Chul Ye. Diffusionclip: Text-guided diffusion models for robust image manipulation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2426-2435, 2022. 3 +[29] Stepan Komkov and Aleksandr Petiushko. Advhat: Realworld adversarial attack on arcface face id system. In 2020 25th International Conference on Pattern Recognition (ICPR), pages 819-826. IEEE, 2021. 3 +[30] Zhenzhong Kuang, Zhiqiang Guo, Jinglong Fang, Jun Yu, Noboru Babaguchi, and Jianping Fan. Unnoticeable synthetic face replacement for image privacy protection. Neurocomputing, 457:322-333, 2021. 3 +[31] Zhenzhong Kuang, Huigui Liu, Jun Yu, Aikui Tian, Lei Wang, Jianping Fan, and Noboru Babaguchi. Effective de-identification generative adversarial network for face anonymization. In Proceedings of the 29th ACM International Conference on Multimedia, pages 3182–3191, 2021. 3 +[32] Gihyun Kwon and Jong Chul Ye. Clipstyler: Image style transfer with a single text condition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18062-18071, 2022. 3 +[33] Tao Li and Min Soo Choi. Deepblur: A simple and effective method for natural image obfuscation. arXiv preprint arXiv:2104.02655, 1, 2021. 3 +[34] Tao Li and Lei Lin. Anonymousnet: Natural face de-identification with measurable privacy. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops, pages 0–0, 2019. 3 +[35] Fangcheng Liu, Chao Zhang, and Hongyang Zhang. Towards transferable unrestricted adversarial examples with minimum changes. arXiv preprint arXiv:2201.01102, 2022. 3 +[36] Suolan Liu, Lizhi Kong, and Hongyuan Wang. Face detection and encryption for privacy preserving in surveillance video. In Chinese Conference on Pattern Recognition and Computer Vision (PRCV), pages 162-172. Springer, 2018. 3 +[37] Aleksander Madry, Aleksandar Makelov, Ludwig Schmidt, Dimitris Tsipras, and Adrian Vladu. Towards deep learning models resistant to adversarial attacks. In Proceedings of the 6th International Conference on Learning Representations (ICLR'18), 2018. 6, 7 +[38] Blaz Meden, Peter Rot, Philipp Terhörst, Naser Damer, Arjan Kuijper, Walter J Scheirer, Arun Ross, Peter Peer, and Vitomir Štruc. Privacy-enhancing face biometrics: A comprehensive survey. IEEE Transactions on Information Forensics and Security, 2021. 2, 3 +[39] Dongbin Na, Sangwoo Ji, and Jong Kim. Unrestricted blackbox adversarial attack using gan with limited queries. arXiv preprint arXiv:2208.11613, 2022. 2 +[40] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741, 2021. 3 + +[41] Seong Joon Oh, Rodrigo Benenson, Mario Fritz, and Bernt Schiele. Faceless person recognition: Privacy implications in social media. In European Conference on Computer Vision, pages 19-35. Springer, 2016. 3 +[42] Seong Joon Oh, Mario Fritz, and Bernt Schiele. Adversarial image perturbation for privacy protection a game theory perspective. In 2017 IEEE International Conference on Computer Vision (ICCV), pages 1491-1500. IEEE, 2017. 3 +[43] Omkar M Parkhi, Andrea Vedaldi, and Andrew Zisserman. Deep face recognition. 2015. 1 +[44] Or Patashnik, Zongze Wu, Eli Shechtman, Daniel Cohen-Or, and Dani Lischinski. Styleclip: Text-driven manipulation of stylegan imagery. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2085–2094, 2021. 3 +[45] P Jonathon Phillips, Amy N Yates, Ying Hu, Carina A Hahn, Eilidh Noyes, Kelsey Jackson, Jacqueline G Cava-zos, Geraldine Jeckeln, Rajeev Ranjan, Swami Sankaranarayanan, et al. Face recognition accuracy of forensic examiners, superrecognizers, and face recognition algorithms. Proceedings of the National Academy of Sciences, 115(24):6171-6176, 2018. 2 +[46] Omid Poursaeed, Tianxing Jiang, Harry Yang, Serge Belongie, and Ser-Nam Lim. Robustness and generalization via generative adversarial training. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15711-15720, 2021. 3 +[47] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 3, 5 +[48] Rajeev Ranjan, Swami Sankaranarayanan, Ankan Bansal, Navaneeth Bodla, Jun-Cheng Chen, Vishal M Patel, Carlos D Castillo, and Rama Chellappa. Deep learning for understanding faces: Machines may be just as good, or better, than humans. IEEE Signal Processing Magazine, 35(1):66-83, 2018. 2 +[49] Daniel Roich, Ron Mokady, Amit H Bermano, and Daniel Cohen-Or. Pivotal tuning for latent-based editing of real images. ACM Transactions on Graphics (TOG), 42(1):1-13, 2022. 5 +[50] Tim Salimans, Ian Goodfellow, Wojciech Zaremba, Vicki Cheung, Alec Radford, and Xi Chen. Improved techniques for training gans. Advances in neural information processing systems, 29, 2016. 3 +[51] Florian Schroff, Dmitry Kalenichenko, and James Philbin. Facenet: A unified embedding for face recognition and clustering. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 815-823, 2015. 6 +[52] Sachith Seneviratne, Nuran Kasthuriarachchi, Sanka Rasnayaka, Danula Hettiachchi, and Ridwan Shariffdeen. Does a face mask protect my privacy?: Deep learning to predict protected attributes from masked face images. In Australasian Joint Conference on Artificial Intelligence, pages 91-102. Springer, 2022. 3 + +[53] Shawn Shan, Emily Wenger, Jiayun Zhang, Huiying Li, Haitao Zheng, and Ben Y Zhao. Fawkes: Protecting privacy against unauthorized deep learning models. In 29th USENIX security symposium (USENIX Security 20), pages 1589-1604, 2020. 2, 3 +[54] Mahmood Sharif, Sruti Bhagavatula, Lujo Bauer, and Michael K Reiter. A general framework for adversarial examples with objectives. ACM Transactions on Privacy and Security (TOPS), 22(3):1-30, 2019. 3 +[55] Yang Song, Rui Shu, Nate Kushman, and Stefano Ermon. Constructing unrestricted adversarial examples with generative models. Advances in Neural Information Processing Systems, 31, 2018. 2, 3 +[56] Qianru Sun, Liqian Ma, Seong Joon Oh, Luc Van Gool, Bernt Schiele, and Mario Fritz. Natural and effective obfuscation by head inpainting. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 5050-5059, 2018. 3 +[57] Christian Szegedy, Wojciech Zaremba, Ilya Sutskever, Joan Bruna, Dumitru Erhan, Ian Goodfellow, and Rob Fergus. Intriguing properties of neural networks. In International Conference on Learning Representations, 2014. 2 +[58] Huan Tian, Tianqing Zhu, and Wanlei Zhou. Fairness and privacy preservation for facial images: Gan-based methods. Computers & Security, 122:102902, 2022. 3 +[59] Omer Tov, Yuval Alaluf, Yotam Nitzan, Or Patashnik, and Daniel Cohen-Or. Designing an encoder for stylegan image manipulation. ACM Transactions on Graphics (TOG), 40(4):1-14, 2021. 5 +[60] Can Wang, Menglei Chai, Mingming He, Dongdong Chen, and Jing Liao. Clip-nerf: Text-and-image driven manipulation of neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3835-3844, 2022. 3 +[61] Mei Wang and Weihong Deng. Deep face recognition: A survey. Neurocomputing, 429:215-244, 2021. 1, 2 +[62] Shunxin Wang, Una M Kelly, and Raymond NJ Veldhuis. Gender obfuscation through face morphing. In 2021 IEEE International Workshop on Biometrics and Forensics (IWBF), pages 1-6. IEEE, 2021. 3 +[63] Ya Wang, Tianlong Bao, Chunhui Ding, and Ming Zhu. Face recognition in real-world surveillance videos with deep learning method. In 2017 2nd international conference on image, vision and computing (icivc), pages 239-243. IEEE, 2017. 2 +[64] Yinggui Wang, Jian Liu, Man Luo, Le Yang, and Li Wang. Privacy-preserving face recognition in the frequency domain. 2022. 3 +[65] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600-612, 2004. 6 +[66] Tianyi Wei, Dongdong Chen, Wenbo Zhou, Jing Liao, Zhentao Tan, Lu Yuan, Weiming Zhang, and Nenghai Yu. Hairclip: Design your hair by text and reference image. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18072-18081, 2022. 3 + +[67] Emily Wenger, Shawn Shan, Haitao Zheng, and Ben Y Zhao. Sok: Anti-facial recognition technology. arXiv preprint arXiv:2112.04558, 2021. 2 +[68] Chaowei Xiao, Jun-Yan Zhu, Bo Li, Warren He, Mingyan Liu, and Dawn Song. Spatially transformed adversarial examples. arXiv preprint arXiv:1801.02612, 2018. 2, 3 +[69] Zihao Xiao, Xianfeng Gao, Chilin Fu, Yinpeng Dong, Wei Gao, Xiaolu Zhang, Jun Zhou, and Jun Zhu. Improving transferability of adversarial patches on face recognition with generative models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11845-11854, 2021. 3 +[70] Xiao Yang, Yinpeng Dong, Tianyu Pang, Hang Su, Jun Zhu, Yuefeng Chen, and Hui Xue. Towards face encryption by generating adversarial identity masks. In Proceedings of the 2021 IEEE/CVF International Conference on Computer Vision (ICCV'21), pages 3897-3907, 2021. 1, 2, 3, 6, 7 +[71] Bangjie Yin, Wenxuan Wang, Taiping Yao, Junfeng Guo, Zelun Kong, Shouhong Ding, Jilin Li, and Cong Liu. Advmakeup: A new imperceptible and transferable attack on face recognition. In Proceedings of the 30th International Joint Conference on Artificial Intelligence (IJCAI'21), pages 1252-1258, 2021. 2, 6, 7 +[72] Bangjie Yin, Wenxuan Wang, Taiping Yao, Junfeng Guo, Zelun Kong, Shouhong Ding, Jilin Li, and Cong Liu. Advmakeup: A new imperceptible and transferable attack on face recognition. arXiv preprint arXiv:2105.03162, 2021. 2, 3 +[73] Shengming Yuan, Qilong Zhang, Lianli Gao, Yaya Cheng, and Jingkuan Song. Natural color fool: Towards boosting black-box unrestricted attacks. arXiv preprint arXiv:2210.02041, 2022. 3 +[74] Jiaming Zhang, Jitao Sang, Xian Zhao, Xiaowen Huang, Yanfeng Sun, and Yongli Hu. Adversarial privacy-preserving filter. In Proceedings of the 28th ACM International Conference on Multimedia, pages 1423-1431, 2020. 2 +[75] Kaipeng Zhang, Zhanpeng Zhang, Zhifeng Li, and Yu Qiao. Joint face detection and alignment using multitask cascaded convolutional networks. IEEE signal processing letters, 23(10):1499-1503, 2016. 6 +[76] Zhengyu Zhao, Zhuoran Liu, and Martha Larson. Towards large yet imperceptible adversarial image perturbations with perceptual color distance. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1039-1048, 2020. 2, 3 +[77] Yaoyao Zhong and Weihong Deng. Opom: Customized invisible cloak towards face privacy protection. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2022, 2, 3 +[78] Jizhe Zhou and Chi-Man Pun. Personal privacy protection via irrelevant faces tracking and pixelation in video live streaming. IEEE Transactions on Information Forensics and Security, 16:1088-1103, 2020. 3 +[79] Zheng-An Zhu, Yun-Zhong Lu, and Chen-Kuo Chiang. Generating adversarial examples by makeup attacks on face recognition. In 2019 IEEE International Conference on Image Processing (ICIP), pages 2516-2520. IEEE, 2019. 3 \ No newline at end of file diff --git a/2023/CLIP2Protect_ Protecting Facial Privacy Using Text-Guided Makeup via Adversarial Latent Search/images.zip b/2023/CLIP2Protect_ Protecting Facial Privacy Using Text-Guided Makeup via Adversarial Latent Search/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..41e8eed7e17b5cb6ed2a394b4715a75d211320eb --- /dev/null +++ b/2023/CLIP2Protect_ Protecting Facial Privacy Using Text-Guided Makeup via Adversarial Latent Search/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:feaa506a6908779074db215869e6c626501c5df474956aca344cd6fee047c90c +size 565583 diff --git a/2023/CLIP2Protect_ Protecting Facial Privacy Using Text-Guided Makeup via Adversarial Latent Search/layout.json b/2023/CLIP2Protect_ Protecting Facial Privacy Using Text-Guided Makeup via Adversarial Latent Search/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..2fb2cb37769cd27ad061b8aded0ad005a6abb0c3 --- /dev/null +++ b/2023/CLIP2Protect_ Protecting Facial Privacy Using Text-Guided Makeup via Adversarial Latent Search/layout.json @@ -0,0 +1,9228 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 74, + 95, + 520, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 74, + 95, + 520, + 129 + ], + "spans": [ + { + "bbox": [ + 74, + 95, + 520, + 129 + ], + "type": "text", + "content": "CLIP2Protect: Protecting Facial Privacy using Text-Guided Makeup via Adversarial Latent Search" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 143, + 147, + 449, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 147, + 449, + 175 + ], + "spans": [ + { + "bbox": [ + 143, + 147, + 449, + 175 + ], + "type": "text", + "content": "Fahad Shamshad Muzammal Naseer Karthik Nandakumar Mohamed Bin Zayed University of AI, UAE" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 116, + 177, + 473, + 189 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 116, + 177, + 473, + 189 + ], + "spans": [ + { + "bbox": [ + 116, + 177, + 473, + 189 + ], + "type": "text", + "content": "{fahad.shamshad, muzammal.naseer, karthik.nandakumar}@mbzuai.ac.ae" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 63, + 214, + 541, + 371 + ], + "blocks": [ + { + "bbox": [ + 53, + 233, + 62, + 335 + ], + "lines": [ + { + "bbox": [ + 53, + 233, + 62, + 335 + ], + "spans": [ + { + "bbox": [ + 53, + 233, + 62, + 335 + ], + "type": "text", + "content": "Protected Original" + } + ] + } + ], + "index": 5, + "angle": 270, + "type": "image_caption" + }, + { + "bbox": [ + 63, + 214, + 541, + 371 + ], + "lines": [ + { + "bbox": [ + 63, + 214, + 541, + 371 + ], + "spans": [ + { + "bbox": [ + 63, + 214, + 541, + 371 + ], + "type": "image", + "image_path": "84f4e77f6d06dc8ebde7ec482aebe28aac8a0e06def5699e03738271ad09c778.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 63, + 376, + 541, + 447 + ], + "blocks": [ + { + "bbox": [ + 63, + 376, + 541, + 447 + ], + "lines": [ + { + "bbox": [ + 63, + 376, + 541, + 447 + ], + "spans": [ + { + "bbox": [ + 63, + 376, + 541, + 447 + ], + "type": "image", + "image_path": "bf65165568d7efeaaee26804444fe4923be66bba4fb54e90fbbc1c2f0234b563.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 449, + 496, + 517 + ], + "lines": [ + { + "bbox": [ + 46, + 449, + 496, + 517 + ], + "spans": [ + { + "bbox": [ + 46, + 449, + 496, + 517 + ], + "type": "text", + "content": "Figure 1. The proposed approach crafts \"naturalistic\" and transferable text-guided adversarial faces to deceive black-box face recognition systems. First row shows original images that need to be protected and second row shows corresponding protected images along with the user-defined makeup text prompts that guide the adversarial search. Comparison against existing methods is shown in the third row. The yellow text represents the confidence score (higher is better) output by a commercial API (Face++, when matching the protected image against the target identity shown in the bottom right. The reference image used by [22] for makeup transfer is shown at the bottom corner of the corresponding adversarial image." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 497, + 452, + 541, + 495 + ], + "blocks": [ + { + "bbox": [ + 497, + 452, + 541, + 495 + ], + "lines": [ + { + "bbox": [ + 497, + 452, + 541, + 495 + ], + "spans": [ + { + "bbox": [ + 497, + 452, + 541, + 495 + ], + "type": "image", + "image_path": "35c8881d6894b6f19f0c6d90b1d9859eb4492ae04adf388b9c4280dcfc386773.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 504, + 496, + 529, + 505 + ], + "lines": [ + { + "bbox": [ + 504, + 496, + 529, + 505 + ], + "spans": [ + { + "bbox": [ + 504, + 496, + 529, + 505 + ], + "type": "text", + "content": "Target" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 143, + 526, + 192, + 538 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 526, + 192, + 538 + ], + "spans": [ + { + "bbox": [ + 143, + 526, + 192, + 538 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 45, + 558, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 558, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 45, + 558, + 288, + 713 + ], + "type": "text", + "content": "The success of deep learning based face recognition systems has given rise to serious privacy concerns due to their ability to enable unauthorized tracking of users in the digital world. Existing methods for enhancing privacy fail to generate \"naturalistic\" images that can protect facial privacy without compromising user experience. We propose a novel two-step approach for facial privacy protection that relies on finding adversarial latent codes in the low-dimensional manifold of a pretrained generative model. The first step inverts the given face image into the latent space and finetunes the generative model to achieve an accurate reconstruction of the given image from its latent code. This step produces a good initialization, aiding the generation" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 527, + 545, + 658 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 527, + 545, + 658 + ], + "spans": [ + { + "bbox": [ + 304, + 527, + 545, + 658 + ], + "type": "text", + "content": "of high-quality faces that resemble the given identity. Subsequently, user-defined makeup text prompts and identity-preserving regularization are used to guide the search for adversarial codes in the latent space. Extensive experiments demonstrate that faces generated by our approach have stronger black-box transferability with an absolute gain of " + }, + { + "bbox": [ + 304, + 527, + 545, + 658 + ], + "type": "inline_equation", + "content": "12.06\\%" + }, + { + "bbox": [ + 304, + 527, + 545, + 658 + ], + "type": "text", + "content": " over the state-of-the-art facial privacy protection approach under the face verification task. Finally, we demonstrate the effectiveness of the proposed approach for commercial face recognition systems. Our code is available at https://github.com/fahadshamshad/Clip2Protect." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 307, + 668, + 386, + 679 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 668, + 386, + 679 + ], + "spans": [ + { + "bbox": [ + 307, + 668, + 386, + 679 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "content": "Deep learning based face recognition (FR) systems [43, 61] have found widespread usage in multiple applications," + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20595" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 50, + 118, + 287, + 194 + ], + "blocks": [ + { + "bbox": [ + 47, + 71, + 288, + 115 + ], + "lines": [ + { + "bbox": [ + 47, + 71, + 288, + 115 + ], + "spans": [ + { + "bbox": [ + 47, + 71, + 288, + 115 + ], + "type": "text", + "content": "Table 1. Comparison among different facial privacy protection methods w.r.t. the natural outputs, black box setting, experiments under face verification and identification tasks, unrestricted (semantically meaningful), and more flexible text guided adversaries." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 50, + 118, + 287, + 194 + ], + "lines": [ + { + "bbox": [ + 50, + 118, + 287, + 194 + ], + "spans": [ + { + "bbox": [ + 50, + 118, + 287, + 194 + ], + "type": "table", + "html": "
Adv-Makeup [71]TIP-IM [70]AMT-GAN [22]Ours
Natural outputsYesPartiallyPartiallyYes
Black boxYesYesYesYes
VerificationYesNoYesYes
IdentificationNoYesNoYes
UnrestrictedYesNoYesYes
Text guidedNoNoNoYes
", + "image_path": "9183f8082408685d93c8263745f8beff1d5039593c7bc6625f0e43e1cbe901cf.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 209, + 287, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 209, + 287, + 376 + ], + "spans": [ + { + "bbox": [ + 46, + 209, + 287, + 376 + ], + "type": "text", + "content": "including security [63], biometrics [38], and criminal investigation [45], outperforming humans in many scenarios [12, 48, 61]. Despite positive aspects of this technology, FR systems seriously threaten personal security and privacy in the digital world because of their potential to enable mass surveillance capabilities [1, 67]. For example, government and private entities can use FR systems to track user relationships and activities by scraping face images from social media profiles such as Twitter, LinkedIn, and Facebook [18, 20]. These entities generally use proprietary FR systems, whose specifications are unknown to the public (black box model). Therefore, there is an urgent need for an effective approach that protects facial privacy against such unknown FR systems." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 378, + 287, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 378, + 287, + 544 + ], + "spans": [ + { + "bbox": [ + 46, + 378, + 287, + 544 + ], + "type": "text", + "content": "An ideal facial privacy protection algorithm must strike the right balance between naturalness and privacy protection [70, 77]. In this context, \"naturalness\" is defined as the absence of any noise artifacts that can be easily perceived by human observers and the preservation of human-perceived identity. \"Privacy protection\" refers to the fact that the protected image must be capable of deceiving a black-box malicious FR system. In other words, the protected image must closely resemble the given face image and be artifact-free for a human observer, while at the same time fool an unknown automated FR system. Since failure to generate naturalistic faces can significantly affect user experience on social media platforms, it is a necessary precondition for adoption of a privacy-enhancement algorithm." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 545, + 287, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 545, + 287, + 665 + ], + "spans": [ + { + "bbox": [ + 46, + 545, + 287, + 665 + ], + "type": "text", + "content": "Recent works exploit adversarial attacks [57] to conceal user identity by overlaying noise-constrained (bounded) adversarial perturbations on the original face image [6,53,74]. Since the adversarial examples are generally optimized in the image space, it is often difficult to simultaneously achieve naturalness and privacy [70]. Unlike noise-based methods, unrestricted adversarial examples are not constrained by the magnitude of perturbation in the image space and have demonstrated better perceptual realism for human observers while being adversially effective [3,55,68,76]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 665, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 665, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 665, + 287, + 714 + ], + "type": "text", + "content": "Several efforts have been made to generate unrestricted adversarial examples that mislead FR systems (see Tab. 1) [22, 25, 39, 72]. Among these, adversarial makeup based methods [22, 72] are gaining increasing attention as they" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 72, + 545, + 275 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 275 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 275 + ], + "type": "text", + "content": "can embed adversarial modifications in a more natural way. These approaches use generative adversarial networks [15] (GANs) to adversarily transfer makeup from a given reference image to the user's face image while impersonating a target identity. However, existing techniques based on adversarial makeup transfer have the following limitations: (i) adversarial toxicity in these methods hamper the performance of the makeup transfer module, thereby resulting in unnatural faces with makeup artifacts (see Fig. 1); (ii) the use of a reference image to define the desired makeup style affects the practicality of this approach; (iii) for every new target identity, these approaches require end-to-end retraining from scratch using large makeup datasets; and (iv) most of these methods primarily aim at impersonation of the target identity, whereas the desired privacy objective is dodging, i.e., multiple images of the user's face scraped from different social media sites must not match with each other." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 277, + 545, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 277, + 545, + 350 + ], + "spans": [ + { + "bbox": [ + 304, + 277, + 545, + 350 + ], + "type": "text", + "content": "To mitigate the above problems, we propose a new approach to protect user facial privacy on online platforms (Sec. 3). The proposed approach aims to search for adversarial latent codes in a low-dimensional manifold learned by a generative model trained to generate face images [2, 27]. Our main contributions are:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 317, + 361, + 545, + 628 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 317, + 361, + 545, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 361, + 545, + 445 + ], + "spans": [ + { + "bbox": [ + 317, + 361, + 545, + 445 + ], + "type": "text", + "content": "- Facial Privacy-protection Framework Using Adversarial Latent Codes: Given a face image, we propose a novel two-step method to search for adversarial latent codes, which can be used by a generative model (e.g., StyleGAN) to produce face images with high visual quality that matches human-perceived identity, while deceiving black-box FR systems." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 317, + 453, + 545, + 561 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 453, + 545, + 561 + ], + "spans": [ + { + "bbox": [ + 317, + 453, + 545, + 561 + ], + "type": "text", + "content": "- Adversarial Makeup Transfer using Textual Prompts: A critical component of the above framework is a technique for leveraging user-defined textual (makeup) prompts to traverse over the latent manifold of the generative model and find transferable adversarial latent codes. Our approach effectively hides attack information in the desired makeup style, without the need for any large makeup dataset or retraining of models for different target identities." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 317, + 568, + 545, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 568, + 545, + 628 + ], + "spans": [ + { + "bbox": [ + 317, + 568, + 545, + 628 + ], + "type": "text", + "content": "- Identity Preserving Regularization: We propose a regularizer that preserves identity-related attributes within the latent space of the generative model and ensures that the protected face image visually resembles the original face." + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 304, + 641, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 641, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 641, + 545, + 714 + ], + "type": "text", + "content": "Extensive experiments (Sec. 4.1) for both face verification and identification scenarios demonstrate the effectiveness of our approach against black-box FR models and online commercial facial recognition APIs (Sec. 4.2). Furthermore, we provide detailed ablative analysis to dissect the performance of different components of our approach (Sec. 4.3)." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "20596" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 71, + 134, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 71, + 134, + 83 + ], + "spans": [ + { + "bbox": [ + 47, + 71, + 134, + 83 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 91, + 287, + 247 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 91, + 287, + 247 + ], + "spans": [ + { + "bbox": [ + 46, + 91, + 287, + 247 + ], + "type": "text", + "content": "Obfuscation Methods: Obfuscation is the most widely used technique [38] to protect user's facial privacy. Earlier obfuscation approaches typically degrade the quality of the original face image by applying simple operations such as masking [52, 64], filtering [33, 78], and image transformations [8, 36, 62]. While these relatively simple obfuscation techniques are reasonable for surveillance applications, they are ill-suited for online/social media platforms where user experience is critical [41]. Though deep learning based obfuscation approaches generate more realistic images [4, 7, 56, 58], they often result in a change of identity compared to the original image and occasionally produce undesirable artifacts [30, 31, 34]." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 247, + 288, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 247, + 288, + 437 + ], + "spans": [ + { + "bbox": [ + 46, + 247, + 288, + 437 + ], + "type": "text", + "content": "Noise-based Adversarial Examples: Adversarial attacks have been used to protect users from unauthorized FR models. Some methods [6, 53] rely on data poisoning to deceive targeted FR models, but are less practical because access to the training data or the gallery set of the unknown FR system is often not available. Other approaches have used game-theory perspective [42] in white-box settings or person-specific privacy masks (one mask per person) to generate protected images at the cost of acquiring multiple images of the same user [77]. In contrast, we aim to fool the black box FR model using only single image. In TIP-IM [70], targeted optimization was used to generate privacy masks against unknown FR models by introducing a naturalness constraint. While this approach provides effective privacy, it generates output images with perceptible noises that can affect the user experience [70]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 437, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 437, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 437, + 288, + 715 + ], + "type": "text", + "content": "Unrestricted Adversarial Examples: Unrestricted adversarial attacks (UAAs) are not constrained by the perturbation norm and can induce large but semantically meaningful perturbations. These attacks have been extensively studied in image classification literature [3, 35, 55, 68, 73, 76] and it has been shown that outputs generated via UAAs are less perceptible to human observers as compared to noise-based adversarial attacks. Motivated by this observation, patch-based unrestricted attacks have been proposed to generate wearable adversarial accessories like colorful glasses [54], hat [29] or random patch [69] to fool the FR model, but such synthesized patches generally have weak transferability due to the limited editing region and the large visible pattern compromises naturalness and affects user experience. Recently, generative models [24, 50] have been leveraged to craft UAAs against FR models. However, these generative approaches are either designed for the white-box settings [46, 79] or show limited performance in query-free black-box settings [25]. Makeup-based UAAs [17, 72] have also been proposed against FR systems by embedding the perturbations into a natural makeup effect. These makeup based attacks have also been exploited to protect the user privacy by applying adversarial makeup on the user" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 304, + 72, + 545, + 179 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 179 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 179 + ], + "type": "text", + "content": "face image [22]. However, interference between adversarial perturbations and makeup transfer can produce undesirable makeup artifacts in the output images. Moreover, these attacks generally assume access to large makeup datasets for training models and require a reference makeup image. In contrast, our approach finds adversarial faces on the natural image manifold in black-box setting via guidance from makeup text prompt, which makes it less susceptible to artifacts (see Fig. 1) and more practical." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 180, + 546, + 372 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 180, + 546, + 372 + ], + "spans": [ + { + "bbox": [ + 304, + 180, + 546, + 372 + ], + "type": "text", + "content": "Vision-Language Modelling: Cross-modal vision-language modelling has attracted significant attention in recent years [13]. OpenAI introduced CLIP [47] that is trained on 400 million image-text pairs using contrastive objective and maps both image and text in a joint multimodal embedding space. With powerful representation embedding of CLIP, several methods have been proposed to manipulate images with text-guidance. StyleCLIP [44] and DiffusionCLIP [28, 40] leverage the powerful generative capabilities of StyleGAN and diffusion models to manipulate images with text prompts. Other similar works include HairCLIP [66], CLIP-NeRF [60], CLIPstyler [32], and CLIPDraw [14]. While these methods focus on the text-guidance ability of CLIP, our approach aims to find the adversarial latent codes in a generative model's latent space for privacy protection against black-box FR models." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 381, + 516, + 396 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 381, + 516, + 396 + ], + "spans": [ + { + "bbox": [ + 306, + 381, + 516, + 396 + ], + "type": "text", + "content": "3. Proposed Approach for Facial Privacy" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 401, + 545, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 401, + 545, + 594 + ], + "spans": [ + { + "bbox": [ + 304, + 401, + 545, + 594 + ], + "type": "text", + "content": "Our goal is to protect user facial privacy on online platforms against unknown (black-box) FR models without compromising on the user's online experience. The proposed approach finds protected faces by adversarially exploring the low-dimensional latent space of a pretrained generative model that is trained on natural face images. To avoid artifacts in the protected image, we restrict the search for adversarial faces close to the clean image manifold learned by the generative model. Moreover, we propose to optimize only over identity-preserving latent codes in the latent space. This effectively preserves human-perceived identity during attack while offering high privacy against automated systems. Further, we employ natural makeup-like perturbations via guidance from a text prompt, which provides more flexibility to the user compared to reference image-based adversarial makeup transfer [22]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 599, + 392, + 611 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 599, + 392, + 611 + ], + "spans": [ + { + "bbox": [ + 306, + 599, + 392, + 611 + ], + "type": "text", + "content": "3.1. Preliminaries" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "content": "Let " + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\pmb{x} \\in \\mathcal{X} \\subset \\mathbb{R}^n" + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "content": " denote the given original/real face image. Let " + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "inline_equation", + "content": "f(\\pmb{x}) : \\mathcal{X} \\to \\mathbb{R}^d" + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "content": " be a FR model that extracts a fixed-length normalized feature representation. Let " + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{D}(\\pmb{x}_1, \\pmb{x}_2) = D(f(\\pmb{x}_1), f(\\pmb{x}_2))" + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "content": " be a distance metric that measures the dissimilarity between two face images " + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\pmb{x}_1" + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\pmb{x}_2" + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "content": " based on their respective representations " + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "inline_equation", + "content": "f(\\pmb{x}_1)" + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "inline_equation", + "content": "f(\\pmb{x}_2)" + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "content": ". Generally a FR system can operate in two modes: verification and identification. A face verification" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "20597" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 61, + 68, + 532, + 241 + ], + "blocks": [ + { + "bbox": [ + 61, + 68, + 532, + 241 + ], + "lines": [ + { + "bbox": [ + 61, + 68, + 532, + 241 + ], + "spans": [ + { + "bbox": [ + 61, + 68, + 532, + 241 + ], + "type": "image", + "image_path": "d22a9ba3df87594c2ff4ea2e1367252711931d277ddbfbe0fbd87c1f409c524f.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 249, + 548, + 297 + ], + "lines": [ + { + "bbox": [ + 46, + 249, + 548, + 297 + ], + "spans": [ + { + "bbox": [ + 46, + 249, + 548, + 297 + ], + "type": "text", + "content": "Figure 2. Overall pipeline of the proposed approach to protect users facial privacy. Our proposed approach searches for the adversarial latent codes on the generative manifold to reconstruct an adversarial face that is capable of fooling unknown FR systems for privacy protection. Our approach allows \"makeup\" editing in an adversarial manner through user defined textual prompts and thereby enhance the user's online experience. Our text-guided objective searches for such latent codes while keeping the original identity preserved." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 314, + 287, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 314, + 287, + 422 + ], + "spans": [ + { + "bbox": [ + 46, + 314, + 287, + 422 + ], + "type": "text", + "content": "system predicts that two faces belong to the same identity if " + }, + { + "bbox": [ + 46, + 314, + 287, + 422 + ], + "type": "inline_equation", + "content": "\\mathcal{D}(\\pmb{x}_1, \\pmb{x}_2) \\leq \\tau" + }, + { + "bbox": [ + 46, + 314, + 287, + 422 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 314, + 287, + 422 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 46, + 314, + 287, + 422 + ], + "type": "text", + "content": " is the system threshold. On the other hand, a (closed set) face identification system compares the input image (probe) against a set of face images (gallery) and outputs the identity whose representation is most similar to that of the probe. Since the attacker can employ verification or identification to determine the user identity using black-box FR models, a protection approach should conceal the user's identity in both scenarios." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 422, + 288, + 554 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 422, + 288, + 554 + ], + "spans": [ + { + "bbox": [ + 46, + 422, + 288, + 554 + ], + "type": "text", + "content": "User privacy can be protected by misleading the malicious FR model through impersonation or dodging attacks. In the context of verification, impersonation (false match) implies that the protected face matches with the face of a specific target identity and dodging (false non-match) means that the protected face does not match with some other image of the same person. Similarly, for face identification, impersonation ensures that the protected image gets matched to a specified target identity in the gallery set, while dodging prevents the protected face from matching with images of the same person in the gallery." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 554, + 287, + 673 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 554, + 287, + 673 + ], + "spans": [ + { + "bbox": [ + 46, + 554, + 287, + 673 + ], + "type": "text", + "content": "Problem Statement: Given the original face image " + }, + { + "bbox": [ + 46, + 554, + 287, + 673 + ], + "type": "inline_equation", + "content": "\\pmb{x}" + }, + { + "bbox": [ + 46, + 554, + 287, + 673 + ], + "type": "text", + "content": ", our goal is to generate a protected face image " + }, + { + "bbox": [ + 46, + 554, + 287, + 673 + ], + "type": "inline_equation", + "content": "\\pmb{x}^p" + }, + { + "bbox": [ + 46, + 554, + 287, + 673 + ], + "type": "text", + "content": " such that " + }, + { + "bbox": [ + 46, + 554, + 287, + 673 + ], + "type": "inline_equation", + "content": "\\mathcal{D}(\\pmb{x}^p, \\pmb{x})" + }, + { + "bbox": [ + 46, + 554, + 287, + 673 + ], + "type": "text", + "content": " is large (for successful dodging attack) and " + }, + { + "bbox": [ + 46, + 554, + 287, + 673 + ], + "type": "inline_equation", + "content": "\\mathcal{D}(\\pmb{x}^p, \\pmb{x}^t)" + }, + { + "bbox": [ + 46, + 554, + 287, + 673 + ], + "type": "text", + "content": " is small (for successfully impersonating a target face " + }, + { + "bbox": [ + 46, + 554, + 287, + 673 + ], + "type": "inline_equation", + "content": "\\pmb{x}^t" + }, + { + "bbox": [ + 46, + 554, + 287, + 673 + ], + "type": "text", + "content": "), where " + }, + { + "bbox": [ + 46, + 554, + 287, + 673 + ], + "type": "inline_equation", + "content": "\\mathcal{O}(\\pmb{x}) \\neq \\mathcal{O}(\\pmb{x}^t)" + }, + { + "bbox": [ + 46, + 554, + 287, + 673 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 554, + 287, + 673 + ], + "type": "inline_equation", + "content": "\\mathcal{O}" + }, + { + "bbox": [ + 46, + 554, + 287, + 673 + ], + "type": "text", + "content": " is the oracle that gives the true identity labels. At the same time, we want to minimize " + }, + { + "bbox": [ + 46, + 554, + 287, + 673 + ], + "type": "inline_equation", + "content": "\\mathcal{H}(\\pmb{x}^p, \\pmb{x})" + }, + { + "bbox": [ + 46, + 554, + 287, + 673 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 554, + 287, + 673 + ], + "type": "inline_equation", + "content": "\\mathcal{H}" + }, + { + "bbox": [ + 46, + 554, + 287, + 673 + ], + "type": "text", + "content": " quantifies the degree of unnaturalness introduced in the protected image " + }, + { + "bbox": [ + 46, + 554, + 287, + 673 + ], + "type": "inline_equation", + "content": "\\pmb{x}^p" + }, + { + "bbox": [ + 46, + 554, + 287, + 673 + ], + "type": "text", + "content": " in relation to the original image " + }, + { + "bbox": [ + 46, + 554, + 287, + 673 + ], + "type": "inline_equation", + "content": "\\pmb{x}" + }, + { + "bbox": [ + 46, + 554, + 287, + 673 + ], + "type": "text", + "content": ". Formally, the optimization problem that we aim to solve is:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 90, + 677, + 287, + 710 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 90, + 677, + 287, + 710 + ], + "spans": [ + { + "bbox": [ + 90, + 677, + 287, + 710 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\min _ {\\boldsymbol {x} ^ {p}} \\mathcal {L} (\\boldsymbol {x} ^ {p}) = \\mathcal {D} (\\boldsymbol {x} ^ {p}, \\boldsymbol {x} ^ {t}) - \\mathcal {D} (\\boldsymbol {x} ^ {p}, \\boldsymbol {x}) \\tag {1} \\\\ \\mathrm {s . t .} \\mathcal {H} (\\boldsymbol {x} ^ {p}, \\boldsymbol {x}) \\leq \\epsilon \\\\ \\end{array}", + "image_path": "9078484076fb797ea51beb6f1abc6805f79f795956a0413dc6aa1844fe6ff271.jpg" + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 307, + 312, + 385, + 391 + ], + "blocks": [ + { + "bbox": [ + 307, + 312, + 385, + 391 + ], + "lines": [ + { + "bbox": [ + 307, + 312, + 385, + 391 + ], + "spans": [ + { + "bbox": [ + 307, + 312, + 385, + 391 + ], + "type": "image", + "image_path": "2f7422eee5a7f7e84514603fc2fbbb4edb25e2397277343c442a1b4f18c6235b.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 326, + 392, + 365, + 402 + ], + "lines": [ + { + "bbox": [ + 326, + 392, + 365, + 402 + ], + "spans": [ + { + "bbox": [ + 326, + 392, + 365, + 402 + ], + "type": "text", + "content": "(a) Original" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 387, + 312, + 465, + 391 + ], + "blocks": [ + { + "bbox": [ + 387, + 312, + 465, + 391 + ], + "lines": [ + { + "bbox": [ + 387, + 312, + 465, + 391 + ], + "spans": [ + { + "bbox": [ + 387, + 312, + 465, + 391 + ], + "type": "image", + "image_path": "d881a1341c142b5295fe4413d8f2b62a3f46bf4fa85b858312701c8b0020eaee.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 390, + 392, + 461, + 401 + ], + "lines": [ + { + "bbox": [ + 390, + 392, + 461, + 401 + ], + "spans": [ + { + "bbox": [ + 390, + 392, + 461, + 401 + ], + "type": "text", + "content": "(b) Encoder Inversion" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 466, + 312, + 545, + 391 + ], + "blocks": [ + { + "bbox": [ + 466, + 312, + 545, + 391 + ], + "lines": [ + { + "bbox": [ + 466, + 312, + 545, + 391 + ], + "spans": [ + { + "bbox": [ + 466, + 312, + 545, + 391 + ], + "type": "image", + "image_path": "6a8bc6961077e8abed5b5570bff0e24b561455f7b2f07958f9ec3ac193baf6f9.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 467, + 392, + 545, + 402 + ], + "lines": [ + { + "bbox": [ + 467, + 392, + 545, + 402 + ], + "spans": [ + { + "bbox": [ + 467, + 392, + 545, + 402 + ], + "type": "text", + "content": "(c) Generator finetuning" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 304, + 404, + 547, + 449 + ], + "lines": [ + { + "bbox": [ + 304, + 404, + 547, + 449 + ], + "spans": [ + { + "bbox": [ + 304, + 404, + 547, + 449 + ], + "type": "text", + "content": "Figure 3. Generator finetuning allows near-perfect reconstructions of LFW dataset sample. This is crucial for the online experience of users. Matching scores returned by Face++ API are 62.38 and 98.96 for encoder and generator-finetuned inversions, respectively." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 470, + 545, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 470, + 545, + 578 + ], + "spans": [ + { + "bbox": [ + 304, + 470, + 545, + 578 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 470, + 545, + 578 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 304, + 470, + 545, + 578 + ], + "type": "text", + "content": " is a bound on the adversarial perturbation. For noise-based approach, " + }, + { + "bbox": [ + 304, + 470, + 545, + 578 + ], + "type": "inline_equation", + "content": "\\mathcal{H}(\\pmb{x}^p,\\pmb{x}) = \\| \\pmb{x} - \\pmb{x}^p\\|_p" + }, + { + "bbox": [ + 304, + 470, + 545, + 578 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 470, + 545, + 578 + ], + "type": "inline_equation", + "content": "\\| \\cdot \\|_p" + }, + { + "bbox": [ + 304, + 470, + 545, + 578 + ], + "type": "text", + "content": " denotes the " + }, + { + "bbox": [ + 304, + 470, + 545, + 578 + ], + "type": "inline_equation", + "content": "L_p" + }, + { + "bbox": [ + 304, + 470, + 545, + 578 + ], + "type": "text", + "content": " norm. However, direct enforcement of the perturbation constraint leads to visible artifacts, which affects visual quality and user experience. Constraining the solution search space to a natural image manifold using an effective image prior can produce more realistic images. Note that the distance metric " + }, + { + "bbox": [ + 304, + 470, + 545, + 578 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 304, + 470, + 545, + 578 + ], + "type": "text", + "content": " is unknown since our goal is to deceive a black-box FR system." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 586, + 515, + 600 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 586, + 515, + 600 + ], + "spans": [ + { + "bbox": [ + 306, + 586, + 515, + 600 + ], + "type": "text", + "content": "3.2. Makeup Text-Guided Adversarial Faces" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 605, + 545, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 545, + 715 + ], + "type": "text", + "content": "Our approach restricts the solution space of the protected face " + }, + { + "bbox": [ + 304, + 605, + 545, + 715 + ], + "type": "inline_equation", + "content": "\\pmb{x}^p" + }, + { + "bbox": [ + 304, + 605, + 545, + 715 + ], + "type": "text", + "content": " to lie close to the clean face manifold " + }, + { + "bbox": [ + 304, + 605, + 545, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{X}" + }, + { + "bbox": [ + 304, + 605, + 545, + 715 + ], + "type": "text", + "content": ". This manifold can be learned using a generative model trained on real human faces. Specifically, let " + }, + { + "bbox": [ + 304, + 605, + 545, + 715 + ], + "type": "inline_equation", + "content": "G_{\\theta}(\\pmb{w}) : \\mathcal{W} \\to \\mathbb{R}^n" + }, + { + "bbox": [ + 304, + 605, + 545, + 715 + ], + "type": "text", + "content": " denote the pretrained generative model with weights " + }, + { + "bbox": [ + 304, + 605, + 545, + 715 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 304, + 605, + 545, + 715 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 605, + 545, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{W}" + }, + { + "bbox": [ + 304, + 605, + 545, + 715 + ], + "type": "text", + "content": " is the latent space. Our proposed approach consists of two stages: (i) latent code initialization (Sec. 3.2.1) and (ii) text-guided adversarial optimization (Sec. 3.2.2). The overall pipeline of the proposed approach is shown in Fig. 2." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20598" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 191, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 191, + 83 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 191, + 83 + ], + "type": "text", + "content": "3.2.1 Latent Code Initialization" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 91, + 287, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 91, + 287, + 175 + ], + "spans": [ + { + "bbox": [ + 46, + 91, + 287, + 175 + ], + "type": "text", + "content": "The latent code initialization stage is based on GAN inversion, which aims to invert the original image " + }, + { + "bbox": [ + 46, + 91, + 287, + 175 + ], + "type": "inline_equation", + "content": "\\pmb{x}" + }, + { + "bbox": [ + 46, + 91, + 287, + 175 + ], + "type": "text", + "content": " into the latent space " + }, + { + "bbox": [ + 46, + 91, + 287, + 175 + ], + "type": "inline_equation", + "content": "\\mathcal{W}" + }, + { + "bbox": [ + 46, + 91, + 287, + 175 + ], + "type": "text", + "content": ", i.e., find a latent code " + }, + { + "bbox": [ + 46, + 91, + 287, + 175 + ], + "type": "inline_equation", + "content": "\\pmb{w}_{\\mathrm{inv}} \\in \\mathcal{W}" + }, + { + "bbox": [ + 46, + 91, + 287, + 175 + ], + "type": "text", + "content": " such that " + }, + { + "bbox": [ + 46, + 91, + 287, + 175 + ], + "type": "inline_equation", + "content": "\\pmb{x}_{\\mathrm{inv}} = G_{\\theta}(\\pmb{w}_{\\mathrm{inv}}) \\approx \\pmb{x}" + }, + { + "bbox": [ + 46, + 91, + 287, + 175 + ], + "type": "text", + "content": ". To achieve this, we first use an encoder-based inversion called e4e [59] to infer " + }, + { + "bbox": [ + 46, + 91, + 287, + 175 + ], + "type": "inline_equation", + "content": "\\pmb{w}_{\\mathrm{inv}}" + }, + { + "bbox": [ + 46, + 91, + 287, + 175 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 46, + 91, + 287, + 175 + ], + "type": "inline_equation", + "content": "\\mathcal{W}" + }, + { + "bbox": [ + 46, + 91, + 287, + 175 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 46, + 91, + 287, + 175 + ], + "type": "inline_equation", + "content": "\\pmb{x}" + }, + { + "bbox": [ + 46, + 91, + 287, + 175 + ], + "type": "text", + "content": " i.e., " + }, + { + "bbox": [ + 46, + 91, + 287, + 175 + ], + "type": "inline_equation", + "content": "\\pmb{w}_{\\mathrm{inv}} = I_{\\phi}(\\pmb{x})" + }, + { + "bbox": [ + 46, + 91, + 287, + 175 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 91, + 287, + 175 + ], + "type": "inline_equation", + "content": "I_{\\phi}: \\mathcal{X} \\to \\mathcal{W}" + }, + { + "bbox": [ + 46, + 91, + 287, + 175 + ], + "type": "text", + "content": " is the pretrained encoder with weights " + }, + { + "bbox": [ + 46, + 91, + 287, + 175 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 46, + 91, + 287, + 175 + ], + "type": "text", + "content": " (see Fig. 2)." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 175, + 288, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 175, + 288, + 392 + ], + "spans": [ + { + "bbox": [ + 46, + 175, + 288, + 392 + ], + "type": "text", + "content": "We use StyleGAN trained on a high-resolution dataset of face images as the pretrained generative model " + }, + { + "bbox": [ + 46, + 175, + 288, + 392 + ], + "type": "inline_equation", + "content": "G_{\\theta}" + }, + { + "bbox": [ + 46, + 175, + 288, + 392 + ], + "type": "text", + "content": " due to its powerful synthesis ability and the disentangled structure of its latent space. A significant challenge during inversion is preserving the identity of the original image i.e., " + }, + { + "bbox": [ + 46, + 175, + 288, + 392 + ], + "type": "inline_equation", + "content": "\\mathcal{O}(\\pmb{x}) = \\mathcal{O}(\\pmb{x}_{\\mathrm{inv}})" + }, + { + "bbox": [ + 46, + 175, + 288, + 392 + ], + "type": "text", + "content": ". Generally, optimization and encoder-based inversion approaches struggle to preserve identity after reconstruction [49] (see Fig. 3b). Moreover, when using these approaches, the inversion error can be large for out-of-domain face images with extreme poses and viewpoints, which are quite common in social media applications. Therefore, these approaches cannot be applied directly to invert " + }, + { + "bbox": [ + 46, + 175, + 288, + 392 + ], + "type": "inline_equation", + "content": "\\pmb{x}" + }, + { + "bbox": [ + 46, + 175, + 288, + 392 + ], + "type": "text", + "content": ". Instead, motivated by the recent observation [49] that slight changes to the pretrained generator weights do not harm its editing abilities while achieving near-perfect reconstructions, we finetune the pretrained generator weights " + }, + { + "bbox": [ + 46, + 175, + 288, + 392 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 46, + 175, + 288, + 392 + ], + "type": "text", + "content": " instead of the encoder weights " + }, + { + "bbox": [ + 46, + 175, + 288, + 392 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 46, + 175, + 288, + 392 + ], + "type": "text", + "content": ". Specifically, we fix " + }, + { + "bbox": [ + 46, + 175, + 288, + 392 + ], + "type": "inline_equation", + "content": "w_{\\mathrm{inv}} = I_{\\phi}(\\pmb{x})" + }, + { + "bbox": [ + 46, + 175, + 288, + 392 + ], + "type": "text", + "content": " and fine-tune " + }, + { + "bbox": [ + 46, + 175, + 288, + 392 + ], + "type": "inline_equation", + "content": "G_{\\theta}" + }, + { + "bbox": [ + 46, + 175, + 288, + 392 + ], + "type": "text", + "content": " using the following loss:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 399, + 277, + 418 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 399, + 277, + 418 + ], + "spans": [ + { + "bbox": [ + 56, + 399, + 277, + 418 + ], + "type": "interline_equation", + "content": "\\theta^ {*} = \\underset {\\theta} {\\arg \\min } \\mathcal {L} _ {\\mathrm {L P I P S}} (\\boldsymbol {x}, G _ {\\theta} (\\boldsymbol {w} _ {\\mathrm {i n v}})) + \\lambda_ {2} \\mathcal {L} _ {2} (\\boldsymbol {x}, G _ {\\theta} (\\boldsymbol {w} _ {\\mathrm {i n v}})),", + "image_path": "ec0a519f82e387df9a133695b3dabf6931ce0bc92b1ac5f45f04d3caa5a20958.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 425, + 287, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 425, + 287, + 475 + ], + "spans": [ + { + "bbox": [ + 46, + 425, + 287, + 475 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 425, + 287, + 475 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{LPIPS}}" + }, + { + "bbox": [ + 46, + 425, + 287, + 475 + ], + "type": "text", + "content": " is the perceptual loss and " + }, + { + "bbox": [ + 46, + 425, + 287, + 475 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_2" + }, + { + "bbox": [ + 46, + 425, + 287, + 475 + ], + "type": "text", + "content": " denotes the pixelwise similarity. The final inverted image " + }, + { + "bbox": [ + 46, + 425, + 287, + 475 + ], + "type": "inline_equation", + "content": "x_{\\mathrm{inv}}^*" + }, + { + "bbox": [ + 46, + 425, + 287, + 475 + ], + "type": "text", + "content": " (see Fig. 3c) can be obtained by performing a forward pass of " + }, + { + "bbox": [ + 46, + 425, + 287, + 475 + ], + "type": "inline_equation", + "content": "\\pmb{w}_{\\mathrm{inv}}" + }, + { + "bbox": [ + 46, + 425, + 287, + 475 + ], + "type": "text", + "content": " through fine-tuned generator i.e., " + }, + { + "bbox": [ + 46, + 425, + 287, + 475 + ], + "type": "inline_equation", + "content": "\\pmb{x}_{\\mathrm{inv}}^{*} = G_{\\theta^{*}}(\\pmb{w}_{\\mathrm{inv}})" + }, + { + "bbox": [ + 46, + 425, + 287, + 475 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 490, + 238, + 502 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 490, + 238, + 502 + ], + "spans": [ + { + "bbox": [ + 47, + 490, + 238, + 502 + ], + "type": "text", + "content": "3.2.2 Text-guided adversarial optimization" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 510, + 287, + 570 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 510, + 287, + 570 + ], + "spans": [ + { + "bbox": [ + 46, + 510, + 287, + 570 + ], + "type": "text", + "content": "Given the inverted latent code " + }, + { + "bbox": [ + 46, + 510, + 287, + 570 + ], + "type": "inline_equation", + "content": "\\boldsymbol{w}_{\\mathrm{inv}}" + }, + { + "bbox": [ + 46, + 510, + 287, + 570 + ], + "type": "text", + "content": " and fine-tuned generator " + }, + { + "bbox": [ + 46, + 510, + 287, + 570 + ], + "type": "inline_equation", + "content": "G_{\\theta^{*}}(\\cdot)" + }, + { + "bbox": [ + 46, + 510, + 287, + 570 + ], + "type": "text", + "content": ", our goal is to adversarially perturb this latent code " + }, + { + "bbox": [ + 46, + 510, + 287, + 570 + ], + "type": "inline_equation", + "content": "\\boldsymbol{w}_{\\mathrm{inv}}" + }, + { + "bbox": [ + 46, + 510, + 287, + 570 + ], + "type": "text", + "content": " in the low-dimensional generative manifold " + }, + { + "bbox": [ + 46, + 510, + 287, + 570 + ], + "type": "inline_equation", + "content": "\\mathcal{W}" + }, + { + "bbox": [ + 46, + 510, + 287, + 570 + ], + "type": "text", + "content": " to generate a protected face that fools the black-box FR model, while imitating the makeup style of the text prompt " + }, + { + "bbox": [ + 46, + 510, + 287, + 570 + ], + "type": "inline_equation", + "content": "t_{\\mathrm{makeup}}" + }, + { + "bbox": [ + 46, + 510, + 287, + 570 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 570, + 287, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 570, + 287, + 677 + ], + "spans": [ + { + "bbox": [ + 46, + 570, + 287, + 677 + ], + "type": "text", + "content": "To achieve these objectives, we investigate the following questions: (i) how to effectively extract makeup style information from " + }, + { + "bbox": [ + 46, + 570, + 287, + 677 + ], + "type": "inline_equation", + "content": "t_{\\mathrm{makeup}}" + }, + { + "bbox": [ + 46, + 570, + 287, + 677 + ], + "type": "text", + "content": " and apply it to the face image " + }, + { + "bbox": [ + 46, + 570, + 287, + 677 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 46, + 570, + 287, + 677 + ], + "type": "text", + "content": " in an adversarial manner?, (ii) how to regularize the optimization process so that the output face image is not qualitatively impaired?, (iii) how to craft effective adversarial perturbations that mislead black-box FR models?, and (iv) how to preserve the human-perceived identity " + }, + { + "bbox": [ + 46, + 570, + 287, + 677 + ], + "type": "inline_equation", + "content": "\\mathcal{O}(x)" + }, + { + "bbox": [ + 46, + 570, + 287, + 677 + ], + "type": "text", + "content": " of the original face image while ensuring high privacy?" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "text", + "content": "The first issue can be addressed by aligning the output adversarial image with the text prompt " + }, + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "inline_equation", + "content": "t_{\\mathrm{makeup}}" + }, + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "text", + "content": " in the embedding space of a pretrained vision-language model. The" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 72, + 545, + 191 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 191 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 191 + ], + "type": "text", + "content": "second issue is addressed by enforcing the adversarial latent code to remain close to initialization " + }, + { + "bbox": [ + 304, + 72, + 545, + 191 + ], + "type": "inline_equation", + "content": "\\boldsymbol{w}_{\\mathrm{inv}}" + }, + { + "bbox": [ + 304, + 72, + 545, + 191 + ], + "type": "text", + "content": ". The third issue is solved by crafting transferable text-guided adversarial faces on a white-box surrogate model (or an ensemble of models) with the goal of boosting the fooling rate on the black-box FR model. Finally, we leverage the disentangled nature of latent space in the generative model and incorporate an identity-preserving regularization to effectively maintain the original visual identity. We now present the details of the loss functions used to incorporate the above ideas." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 192, + 546, + 324 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 192, + 546, + 324 + ], + "spans": [ + { + "bbox": [ + 304, + 192, + 546, + 324 + ], + "type": "text", + "content": "Textual Loss: A key ingredient of the proposed approach is text-based guidance to inconspicuously hide the adversarial perturbations into the makeup effect. This can be naively achieved by aligning the representation of " + }, + { + "bbox": [ + 304, + 192, + 546, + 324 + ], + "type": "inline_equation", + "content": "t_{\\mathrm{makeup}}" + }, + { + "bbox": [ + 304, + 192, + 546, + 324 + ], + "type": "text", + "content": " and the adversarial face " + }, + { + "bbox": [ + 304, + 192, + 546, + 324 + ], + "type": "inline_equation", + "content": "G_{\\theta^*}(\\boldsymbol{w})" + }, + { + "bbox": [ + 304, + 192, + 546, + 324 + ], + "type": "text", + "content": " in the common embedding space of a pre-trained vision-language model (e.g. CLIP [47]). However, this approach will transform the whole output image to follow the makeup style of " + }, + { + "bbox": [ + 304, + 192, + 546, + 324 + ], + "type": "inline_equation", + "content": "t_{\\mathrm{makeup}}" + }, + { + "bbox": [ + 304, + 192, + 546, + 324 + ], + "type": "text", + "content": ", which results in low diversity. Therefore, we use a directional CLIP loss that aligns the CLIP-space direction between the text-image pairs of the original and adversarial images. Specifically," + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 376, + 334, + 545, + 360 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 376, + 334, + 545, + 360 + ], + "spans": [ + { + "bbox": [ + 376, + 334, + 545, + 360 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {c l i p}} = 1 - \\frac {\\Delta I \\cdot \\Delta T}{| \\Delta I | | \\Delta T |}, \\tag {2}", + "image_path": "f8dd1ba35c17284c02cab4a32be6b579022e55a673c3354e4a4fcbee45e11f70.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 368, + 545, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 368, + 545, + 441 + ], + "spans": [ + { + "bbox": [ + 305, + 368, + 545, + 441 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 368, + 545, + 441 + ], + "type": "inline_equation", + "content": "\\Delta T = E_{T}(t_{\\mathrm{makeup}}) - E_{T}(t_{\\mathrm{src}})" + }, + { + "bbox": [ + 305, + 368, + 545, + 441 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 368, + 545, + 441 + ], + "type": "inline_equation", + "content": "\\Delta I = E_I(G_{\\theta^*}(\\pmb {w})) - E_I(\\pmb {x})" + }, + { + "bbox": [ + 305, + 368, + 545, + 441 + ], + "type": "text", + "content": ". Here, " + }, + { + "bbox": [ + 305, + 368, + 545, + 441 + ], + "type": "inline_equation", + "content": "E_{T}" + }, + { + "bbox": [ + 305, + 368, + 545, + 441 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 368, + 545, + 441 + ], + "type": "inline_equation", + "content": "E_{I}" + }, + { + "bbox": [ + 305, + 368, + 545, + 441 + ], + "type": "text", + "content": " are the text and image encoders of the CLIP model and " + }, + { + "bbox": [ + 305, + 368, + 545, + 441 + ], + "type": "inline_equation", + "content": "t_\\mathrm{src}" + }, + { + "bbox": [ + 305, + 368, + 545, + 441 + ], + "type": "text", + "content": " is the semantic text of the input image " + }, + { + "bbox": [ + 305, + 368, + 545, + 441 + ], + "type": "inline_equation", + "content": "\\pmb{x}" + }, + { + "bbox": [ + 305, + 368, + 545, + 441 + ], + "type": "text", + "content": ". Since we are dealing with faces, " + }, + { + "bbox": [ + 305, + 368, + 545, + 441 + ], + "type": "inline_equation", + "content": "t_\\mathrm{src}" + }, + { + "bbox": [ + 305, + 368, + 545, + 441 + ], + "type": "text", + "content": " can be simply set as \"face\". This loss localizes makeup transfer (e.g. red lipstick) without affecting privacy." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 441, + 545, + 501 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 441, + 545, + 501 + ], + "spans": [ + { + "bbox": [ + 305, + 441, + 545, + 501 + ], + "type": "text", + "content": "Adversarial Loss: Our goal is to traverse over the latent space " + }, + { + "bbox": [ + 305, + 441, + 545, + 501 + ], + "type": "inline_equation", + "content": "\\mathcal{W}" + }, + { + "bbox": [ + 305, + 441, + 545, + 501 + ], + "type": "text", + "content": " to find adversarial latent codes on the generative manifold whose face feature representation lies close to that of target image and far away from the original image itself i.e., " + }, + { + "bbox": [ + 305, + 441, + 545, + 501 + ], + "type": "inline_equation", + "content": "\\mathcal{D}(\\pmb{x}^p, \\pmb{x}) > \\mathcal{D}(\\pmb{x}^p, \\pmb{x}^t)" + }, + { + "bbox": [ + 305, + 441, + 545, + 501 + ], + "type": "text", + "content": ". Hence, the adversarial loss is:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 340, + 510, + 545, + 525 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 340, + 510, + 545, + 525 + ], + "spans": [ + { + "bbox": [ + 340, + 510, + 545, + 525 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {a d v}} = \\mathcal {D} \\left(G _ {\\theta^ {*}} (\\boldsymbol {w}), \\boldsymbol {x} ^ {t}\\right) - \\mathcal {D} \\left(G _ {\\theta^ {*}} (\\boldsymbol {w}), \\boldsymbol {x}\\right), \\tag {3}", + "image_path": "3fcfdd55e02b48bf313f610bebd05c1fbf4179619eb0cc65eb92046297a2cfe3.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 533, + 545, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 533, + 545, + 605 + ], + "spans": [ + { + "bbox": [ + 304, + 533, + 545, + 605 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 533, + 545, + 605 + ], + "type": "inline_equation", + "content": "\\mathcal{D}(\\pmb{x}_1,\\pmb{x}_2) = 1 - \\cos [f(\\pmb{x}_1),f(\\pmb{x}_2)]" + }, + { + "bbox": [ + 304, + 533, + 545, + 605 + ], + "type": "text", + "content": " is the cosine distance. Since the malicious FR model is unknown in the black-box setting, Eq. 3 cannot be solved directly. Instead, following AMT-GAN [22], we perform adversarial optimization on an ensemble of white-box surrogate models to imitate the decision boundary of the unknown FR model." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 605, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 546, + 713 + ], + "type": "text", + "content": "Identity Preservation Loss: The optimization over the generative manifold ensures that the protected image " + }, + { + "bbox": [ + 304, + 605, + 546, + 713 + ], + "type": "inline_equation", + "content": "\\pmb{x}^p" + }, + { + "bbox": [ + 304, + 605, + 546, + 713 + ], + "type": "text", + "content": " is natural i.e., artifact-free, however, it does not explicitly enforce the protected image to preserve the identity of the original image with respect to the human observer. To mitigate the issue, we take advantage of the semantic control exhibited by StyleGAN in its latent space. The latent code " + }, + { + "bbox": [ + 304, + 605, + 546, + 713 + ], + "type": "inline_equation", + "content": "\\pmb{w} \\in \\mathcal{W}" + }, + { + "bbox": [ + 304, + 605, + 546, + 713 + ], + "type": "text", + "content": " impacts image generation by controlling different level of semantics in the output image. Specifically, latent" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "footer", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "20599" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 204 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 204 + ], + "type": "text", + "content": "codes corresponding to the initial layers of StyleGAN control high-level aspects such as pose, general hairstyle, and face shape [27]. Adversarially perturbing these latent layers can change these attributes, resulting in a change of identity (see Sec. 4.3). Latent codes corresponding to deeper layers of StyleGAN are associated with fine-level control such as makeup style [2]. Therefore, we perturb only those latent codes associated with deeper layers of StyleGAN, thereby restricting the adversarial faces to the identity preserving manifold. We further constrain the latent code to stay close to its initial value " + }, + { + "bbox": [ + 46, + 72, + 289, + 204 + ], + "type": "inline_equation", + "content": "w_{\\mathrm{inv}}" + }, + { + "bbox": [ + 46, + 72, + 289, + 204 + ], + "type": "text", + "content": " using the following regularization:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 82, + 213, + 287, + 228 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 82, + 213, + 287, + 228 + ], + "spans": [ + { + "bbox": [ + 82, + 213, + 287, + 228 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {l a t e n t}} = \\left\\| (\\boldsymbol {w} \\odot \\boldsymbol {m} _ {i d}) - \\left(\\boldsymbol {w} _ {\\text {i n v}} \\odot \\boldsymbol {m} _ {i d}\\right) \\right\\| _ {2}, \\tag {4}", + "image_path": "45bd211eea2a915347a075493fa4a1c87a2f3f8fb07b469e84c7be848c9df27b.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 234, + 287, + 307 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 234, + 287, + 307 + ], + "spans": [ + { + "bbox": [ + 46, + 234, + 287, + 307 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 234, + 287, + 307 + ], + "type": "inline_equation", + "content": "\\odot" + }, + { + "bbox": [ + 46, + 234, + 287, + 307 + ], + "type": "text", + "content": " denotes element-wise product and " + }, + { + "bbox": [ + 46, + 234, + 287, + 307 + ], + "type": "inline_equation", + "content": "m_{id}" + }, + { + "bbox": [ + 46, + 234, + 287, + 307 + ], + "type": "text", + "content": " is an identity preservation mask that is 0 for the initial layers and 1 only for the deeper layers of the latent code. StyleGAN has 18 layers, each having a dimension of 512. The identity preservation mask is set to 1 only from layer 8 to 18. Finally, combining the three loss functions, we have" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 78, + 315, + 287, + 329 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 78, + 315, + 287, + 329 + ], + "spans": [ + { + "bbox": [ + 78, + 315, + 287, + 329 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {t o t a l}} = \\lambda_ {\\text {a d v}} \\mathcal {L} _ {\\text {a d v}} + \\lambda_ {\\text {c l i p}} \\mathcal {L} _ {\\text {c l i p}} + \\lambda_ {\\text {l a t e n t}} \\mathcal {L} _ {\\text {l a t e n t}}, \\tag {5}", + "image_path": "794effb7df3ed84e5e33af8a946a7112ccb00b58b138788403153252262c57e3.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 336, + 287, + 397 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 336, + 287, + 397 + ], + "spans": [ + { + "bbox": [ + 46, + 336, + 287, + 397 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 336, + 287, + 397 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{adv.}}" + }, + { + "bbox": [ + 46, + 336, + 287, + 397 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 336, + 287, + 397 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{clip}}" + }, + { + "bbox": [ + 46, + 336, + 287, + 397 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 336, + 287, + 397 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{latent}}" + }, + { + "bbox": [ + 46, + 336, + 287, + 397 + ], + "type": "text", + "content": " are hyperparameters. Note that " + }, + { + "bbox": [ + 46, + 336, + 287, + 397 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{adv}}" + }, + { + "bbox": [ + 46, + 336, + 287, + 397 + ], + "type": "text", + "content": " accounts for the adversarial objective in Eq. 1, while the text-guided makeup transfer " + }, + { + "bbox": [ + 46, + 336, + 287, + 397 + ], + "type": "inline_equation", + "content": "(\\mathcal{L}_{\\mathrm{clip}})" + }, + { + "bbox": [ + 46, + 336, + 287, + 397 + ], + "type": "text", + "content": " and identity-preserving regularization " + }, + { + "bbox": [ + 46, + 336, + 287, + 397 + ], + "type": "inline_equation", + "content": "(\\mathcal{L}_{\\mathrm{latent}})" + }, + { + "bbox": [ + 46, + 336, + 287, + 397 + ], + "type": "text", + "content": " implicitly enforce the naturalness constraint in Eq. 1." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 406, + 128, + 420 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 406, + 128, + 420 + ], + "spans": [ + { + "bbox": [ + 47, + 406, + 128, + 420 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 426, + 287, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 426, + 287, + 594 + ], + "spans": [ + { + "bbox": [ + 46, + 426, + 287, + 594 + ], + "type": "text", + "content": "Implementation details: In all experiments, we use StyleGAN2 pretrained on the FFHQ face dataset as our generative model. For adversarial text guidance, we use a vision transformer-based CLIP model. For generator fine-tuning in the latent code initialization step, we use 450 iterations with value of " + }, + { + "bbox": [ + 46, + 426, + 287, + 594 + ], + "type": "inline_equation", + "content": "\\lambda_{2}" + }, + { + "bbox": [ + 46, + 426, + 287, + 594 + ], + "type": "text", + "content": " in Eq. 2 set to 0.5. For the makeup text input, we collect 40 text prompts based on the makeup style of diverse nature (details in supplementary material). For adversarial optimization, we use an Adam optimizer with " + }, + { + "bbox": [ + 46, + 426, + 287, + 594 + ], + "type": "inline_equation", + "content": "\\beta_{1}" + }, + { + "bbox": [ + 46, + 426, + 287, + 594 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 426, + 287, + 594 + ], + "type": "inline_equation", + "content": "\\beta_{2}" + }, + { + "bbox": [ + 46, + 426, + 287, + 594 + ], + "type": "text", + "content": " set to 0.9 and 0.999, respectively, and a learning rate of 0.01. We run the optimizer for 50 iterations to craft protected faces. We set the value of " + }, + { + "bbox": [ + 46, + 426, + 287, + 594 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{adv}}" + }, + { + "bbox": [ + 46, + 426, + 287, + 594 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 426, + 287, + 594 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{clip}}" + }, + { + "bbox": [ + 46, + 426, + 287, + 594 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 426, + 287, + 594 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{latent}}" + }, + { + "bbox": [ + 46, + 426, + 287, + 594 + ], + "type": "text", + "content": " to 1, 0.5, and 0.01, respectively. All our experiments are conducted on a A100 GPU with 40 GB memory." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "text", + "content": "Datasets: We perform experiments for both face verification and identification settings. Face verification: We use CelebA-HQ [26] and LADN [16] for the impersonation attack. We select subset of 1,000 images from CelebA-HQ and report average results over 4 target identities provided by [22]. Similarly, for LADN, we divide the 332 images available into 4 groups, where images in each group aim to impersonate the target identities provided by [22]. For dodging attack, we use CelebA-HQ [26] and LFW [23] datasets. Specifically, we select 500 subjects at random and" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 72, + 545, + 192 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 192 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 192 + ], + "type": "text", + "content": "each subject has a pair of faces. Face identification: For impersonation and dodging, we use CelebA-HQ [26] and LFW [23] as our evaluation set. For both datasets, we randomly select 500 subjects, each with a pair of faces. We assign one image in the pair to the gallery set and the other to the probe set. Both impersonation and dodging attacks are performed on the probe set. For impersonation, we insert 4 target identities provided by [22] into the gallery set. A more detailed description of all datasets and pre-processing steps is provided in the supplementary material." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 193, + 545, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 193, + 545, + 300 + ], + "spans": [ + { + "bbox": [ + 304, + 193, + 545, + 300 + ], + "type": "text", + "content": "Target Models: We aim to protect user facial privacy by attacking four FR model with diverse back bones in the black-box settings. The target models include IRSE50 [21], IR152 [9], FaceNet [51], and MobileFace [5]. Following standard protocol, we align and crop the face images using MTCNN [75] before giving them as input to FR models. Further, we also report privacy protection performance based on commercial FR API including Face++ and Tencent Yunshentu FR platforms." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 301, + 546, + 516 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 301, + 546, + 516 + ], + "spans": [ + { + "bbox": [ + 304, + 301, + 546, + 516 + ], + "type": "text", + "content": "Evaluation metrics: Following [70], we use protection success rate (PSR) to evaluate the proposed approach. PSR is defined as the fraction of protected faces missclassified by the malicious FR system. To evaluate PSR, we use the thresholding and closed set strategies for face verification and identification, respectively. For face identification, we also use Rank-N targeted identity success rate (Rank-N-T) and untargeted identity success rate (Rank-N-U), where Rank-N-T means that target image " + }, + { + "bbox": [ + 304, + 301, + 546, + 516 + ], + "type": "inline_equation", + "content": "\\boldsymbol{x}^t" + }, + { + "bbox": [ + 304, + 301, + 546, + 516 + ], + "type": "text", + "content": " will appear at least once in the top N candidates shortlisted from the gallery and Rank-N-U implies that the top N candidate list does not have the same identity as that of original image " + }, + { + "bbox": [ + 304, + 301, + 546, + 516 + ], + "type": "inline_equation", + "content": "\\boldsymbol{x}" + }, + { + "bbox": [ + 304, + 301, + 546, + 516 + ], + "type": "text", + "content": ". We also report results of PSNR (dB), SSIM, and FID [19] scores to evaluate the imperceptibility of method. Large PSNR and SSIM [65] indicates better match with the original images, while low FID score indicates more realistic images. For commercial APIs, we directly report the confidence score returned by the respective servers." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 517, + 545, + 661 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 517, + 545, + 661 + ], + "spans": [ + { + "bbox": [ + 304, + 517, + 545, + 661 + ], + "type": "text", + "content": "Baseline methods: We compare our approach with recent noise-based and makeup based facial privacy protection approaches. Noise based methods include PGD [37], MI-FGSM [10], TI-DIM [11], and TIP-IM [70], whereas makeup-based approaches are Adv-Makeup [71] and AMTGAN [22]. We want to highlight that TIP-IM and AMTGAN are considered the state-of-the-art (SOTA) for face privacy protection against black-box FR systems in noise-based and unrestricted settings, respectively. TIP-IM also incorporate multi-target objective in its optimization to find the optimal target image among multiple targets. For fair comparison, we use its single target variant." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 670, + 430, + 682 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 670, + 430, + 682 + ], + "spans": [ + { + "bbox": [ + 306, + 670, + 430, + 682 + ], + "type": "text", + "content": "4.1. Experimental Results" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 689, + 546, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 546, + 714 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 546, + 714 + ], + "type": "text", + "content": "In this section, we present experimental results of our approach in black-box settings on four different pretrained" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "20600" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 77, + 96, + 518, + 216 + ], + "blocks": [ + { + "bbox": [ + 46, + 70, + 547, + 94 + ], + "lines": [ + { + "bbox": [ + 46, + 70, + 547, + 94 + ], + "spans": [ + { + "bbox": [ + 46, + 70, + 547, + 94 + ], + "type": "text", + "content": "Table 2. Protection success rate (PSR %) of black-box impersonation attack under the face verification task. For each column, the other three FR systems are used as surrogates to generate the protected faces." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 77, + 96, + 518, + 216 + ], + "lines": [ + { + "bbox": [ + 77, + 96, + 518, + 216 + ], + "spans": [ + { + "bbox": [ + 77, + 96, + 518, + 216 + ], + "type": "table", + "html": "
MethodCelebA-HQLADN-DatasetAverage
IRSE50IR152FaceNetMobileFaceIRSE50IR152FaceNetMobileFace
Clean7.293.801.0812.682.713.610.605.114.61
Inverted5.572.770.6013.326.804.510.2511.665.68
PGD [37]36.8720.681.8543.9940.0919.593.8241.0925.60
MI-FGSM [10]45.7925.032.5845.8548.9025.576.3145.0130.63
TI-DIM [11]63.6336.1715.3057.1256.3634.1822.1148.3041.64
Adv-Makeup(ICAI'21) [71]21.959.481.3722.0029.6410.030.9722.3814.72
TIP-IM(ICCV'21) [70]54.4037.2340.7448.7265.8943.5763.5046.4850.06
AMT-GAN(CVPR'22) [22]76.9635.1316.6250.7189.6449.1232.1372.4352.84
Ours81.1048.4241.7275.2691.5753.3147.9179.9464.90
", + "image_path": "e1d3a30df08150fa17b6576d379ee72bdb701195b608b6c5c66bee751053456c.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 70, + 257, + 526, + 385 + ], + "blocks": [ + { + "bbox": [ + 46, + 220, + 547, + 255 + ], + "lines": [ + { + "bbox": [ + 46, + 220, + 547, + 255 + ], + "spans": [ + { + "bbox": [ + 46, + 220, + 547, + 255 + ], + "type": "text", + "content": "Table 3. Protection success rate (PSR %) of black-box dodging (top) and impersonation (bottom) attacks under the face identification task for LFW dataset [23]. For each column, the other three FR systems are used as surrogates to generate the protected faces. R1-U: Rank-1-Untargeted, R5-U: Rank-5-Untargeted, R1-T: Rank-1-Targeted, R5-T: Rank-5-Targeted." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 70, + 257, + 526, + 385 + ], + "lines": [ + { + "bbox": [ + 70, + 257, + 526, + 385 + ], + "spans": [ + { + "bbox": [ + 70, + 257, + 526, + 385 + ], + "type": "table", + "html": "
MethodIRSE50IR152FaceNetMobileFaceAverage
R1-UR5-UR1-UR5-UR1-UR5-UR1-UR5-UR1-UR5-U
MI-FGSM [10]70.242.658.441.859.234.068.047.263.941.4
TI-DIM [11]79.051.267.454.074.452.079.261.675.054.7
TIP-IM(ICCv'21) [70]81.452.271.854.676.049.882.263.077.854.9
Ours86.659.473.456.683.851.285.066.882.258.5
R1-TR5-TR1-TR5-TR1-TR5-TR1-TR5-TR1-TR5-T
MI-FGSM [10]4.010.23.214.29.018.88.422.46.1516.4
TI-DIM [11]4.013.67.819.618.032.821.639.012.8526.25
TIP-IM(ICCv'21) [70]8.028.211.631.225.256.834.051.419.741.9
Ours11.237.816.051.227.454.039.061.223.451.05
", + "image_path": "2dd54fead1c19fbf6364aad49158d50daad247d31dec991723a5d518ec458cab.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 57, + 397, + 200, + 459 + ], + "blocks": [ + { + "bbox": [ + 57, + 397, + 200, + 459 + ], + "lines": [ + { + "bbox": [ + 57, + 397, + 200, + 459 + ], + "spans": [ + { + "bbox": [ + 57, + 397, + 200, + 459 + ], + "type": "table", + "html": "
MethodFID ↓PSR Gain ↑
Adv-Makeup [71]4.230
TIP-IM [70]38.7335.34
AMT-GAN [22]34.4438.12
Ours26.6250.18
", + "image_path": "104a093182231efba5c599f083417c533c5344c2a1e8d4155c769bf41e514dae.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 205, + 397, + 288, + 453 + ], + "lines": [ + { + "bbox": [ + 205, + 397, + 288, + 453 + ], + "spans": [ + { + "bbox": [ + 205, + 397, + 288, + 453 + ], + "type": "text", + "content": "Table 4. FID comparison. PSR Gain is absolute gain in PSR relative to AdvMakeup." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 474, + 287, + 534 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 474, + 287, + 534 + ], + "spans": [ + { + "bbox": [ + 46, + 474, + 287, + 534 + ], + "type": "text", + "content": "FR models under face verification and identification tasks. To generate protected images, we use three FR models as a surrogate to imitate the decision boundary of the fourth FR model. All results are averaged over 5 text based makeup styles that are provided in the supplementary material." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 534, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 534, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 534, + 288, + 714 + ], + "type": "text", + "content": "For face verification experiments, we set the system threshold value at 0.01 false match rate for each FR model i.e., IRSE50 (0.241), IR152 (0.167), FaceNet (0.409), and MobileFace (0.302). Quantitative results in terms of PSR for impersonation attack under the face verification task are shown in Tab. 2. Our approach is able to achieve an average absolute gain of about " + }, + { + "bbox": [ + 46, + 534, + 288, + 714 + ], + "type": "inline_equation", + "content": "12\\%" + }, + { + "bbox": [ + 46, + 534, + 288, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 534, + 288, + 714 + ], + "type": "inline_equation", + "content": "14\\%" + }, + { + "bbox": [ + 46, + 534, + 288, + 714 + ], + "type": "text", + "content": " over SOTA unrestricted [22] and noise-based [70] facial privacy protection methods, respectively. Qualitative results are shown in Fig. 1 which shows that protected faces generated by our approach are more realistic. Results for dodging attacks under face verification are provided in the supplementary material. In Tab. 3, we also provide PSR vales under the face identification task for dodging (untargeted) and impersonation attacks. Our approach consistently outperforms recent" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 399, + 545, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 399, + 545, + 495 + ], + "spans": [ + { + "bbox": [ + 304, + 399, + 545, + 495 + ], + "type": "text", + "content": "methods at both Rank-1 and Rank-5 settings. We emphasize that we are the first to show effectiveness of generative models in offering untargeted privacy protection (dodging) in a more practical identification setting. Since AMT-GAN and Adv-Makeup are originally trained to impersonate target identity under the verification task, we have not included them in Tab. 3. Qualitative results for LFW and CelebA are provided in the supplementary material." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 495, + 545, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 495, + 545, + 593 + ], + "spans": [ + { + "bbox": [ + 304, + 495, + 545, + 593 + ], + "type": "text", + "content": "We report FID scores (lower is better) of our approach in Tab. 4 for CelebA and LADN datasets to measure naturalness. Adv-Makeup has the lowest FID score as it only applies makeup to the eye region without changing the rest of the face. However, this kind of restriction results in poor PSR. Our method has lower FID compared to TIP-IM and AMT-GAN and achieves the highest PSR. We provide PSNR and SSIM results in the supplementary material." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 599, + 520, + 612 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 599, + 520, + 612 + ], + "spans": [ + { + "bbox": [ + 305, + 599, + 520, + 612 + ], + "type": "text", + "content": "4.2. Effectiveness in Real-World Applications" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 618, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 618, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 618, + 547, + 715 + ], + "type": "text", + "content": "We now show the effectiveness of our approach to protect facial images (through targeted impersonation) against commercial API such as Face++ and Tencent Yunshentu FR platform operating in the verification mode. These APIs return confidence scores between 0 to 100 to measure whether two images are similar or not, where a high confidence score indicates high similarity. As the training data and model parameters of these propriety FR models are unknown, it" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "type": "text", + "content": "20601" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 70, + 280, + 178 + ], + "blocks": [ + { + "bbox": [ + 59, + 70, + 280, + 178 + ], + "lines": [ + { + "bbox": [ + 59, + 70, + 280, + 178 + ], + "spans": [ + { + "bbox": [ + 59, + 70, + 280, + 178 + ], + "type": "image", + "image_path": "015044b84574475d588ab89a5e52e746f610eca9ca86988dade41c43919170f7.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 186, + 289, + 233 + ], + "lines": [ + { + "bbox": [ + 46, + 186, + 289, + 233 + ], + "spans": [ + { + "bbox": [ + 46, + 186, + 289, + 233 + ], + "type": "text", + "content": "Figure 4. Average confidence score (higher is better) returned by a real-world face verification API, Face++, for impersonation attack. Our approach has a higher confidence score than state-of-the-art makeup and noise-based facial privacy protection methods." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 55, + 258, + 282, + 299 + ], + "blocks": [ + { + "bbox": [ + 79, + 243, + 254, + 255 + ], + "lines": [ + { + "bbox": [ + 79, + 243, + 254, + 255 + ], + "spans": [ + { + "bbox": [ + 79, + 243, + 254, + 255 + ], + "type": "text", + "content": "Table 5. Impact of " + }, + { + "bbox": [ + 79, + 243, + 254, + 255 + ], + "type": "inline_equation", + "content": "{\\lambda }_{\\text{latent }}" + }, + { + "bbox": [ + 79, + 243, + 254, + 255 + ], + "type": "text", + "content": " on FID score and PSR." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 55, + 258, + 282, + 299 + ], + "lines": [ + { + "bbox": [ + 55, + 258, + 282, + 299 + ], + "spans": [ + { + "bbox": [ + 55, + 258, + 282, + 299 + ], + "type": "table", + "html": "
λlatent0.50.10.050.010.0050.00010
FID11.621.425.227.830.138.443.2
PSR (%)31.239.057.476.283.890.093.6
", + "image_path": "a99d2a6abe74e06b9a91cadc7fa9a543dd4983e5519b616b4cc587b067313827.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 314, + 287, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 314, + 287, + 399 + ], + "spans": [ + { + "bbox": [ + 46, + 314, + 287, + 399 + ], + "type": "text", + "content": "effectively mimics a real-world scenario. We protect 100 faces randomly selected from CelebA-HQ using the baselines and the proposed method. In Fig. 4, we show the average confidence score returned by Face++ against these images. These results indicate that our method has a high PSR compared to baselines. We defer more details and results for Tencent Yunshentu API to supplementary material." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 406, + 149, + 418 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 406, + 149, + 418 + ], + "spans": [ + { + "bbox": [ + 47, + 406, + 149, + 418 + ], + "type": "text", + "content": "4.3. Ablation Studies" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 425, + 287, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 425, + 287, + 449 + ], + "spans": [ + { + "bbox": [ + 46, + 425, + 287, + 449 + ], + "type": "text", + "content": "Next, we report some ablations to evaluate the contributions of our loss components." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 449, + 287, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 449, + 287, + 533 + ], + "spans": [ + { + "bbox": [ + 46, + 449, + 287, + 533 + ], + "type": "text", + "content": "Makeup based text guidance: As shown in Fig. 5 (top), in the absence of text guidance, resulting images may contain artifacts due to increased perturbations induced by the adversarial objective. Text-guidance effectively hides the perturbations in the makeup, leading to more natural looking images. It also provides the user more flexibility to select a desired makeup style compared to a reference image." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 533, + 288, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 533, + 288, + 617 + ], + "spans": [ + { + "bbox": [ + 46, + 533, + 288, + 617 + ], + "type": "text", + "content": "Identity preserving regularization: Optimizing over the whole latent space provides more degrees of freedom and increases the PSR. However, it does not explicitly enforce adversarial optimization to preserve the user identity as shown in Fig. 5 (bottom). The proposed identity preserving regularization effectively preserves identity, while imitating the desire makeup style." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 617, + 288, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 617, + 288, + 689 + ], + "spans": [ + { + "bbox": [ + 46, + 617, + 288, + 689 + ], + "type": "text", + "content": "Impact of latent loss weight: Decreasing the weight assigned to the latent loss " + }, + { + "bbox": [ + 46, + 617, + 288, + 689 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{latent}}" + }, + { + "bbox": [ + 46, + 617, + 288, + 689 + ], + "type": "text", + "content": " results in an increase in both the FID score and PSR (and vice versa). Allowing the latent to deviate more from the initial inverted latent code of the given face image often results in artifacts caused by the adversarial loss, degrading naturalness but aiding privacy." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 689, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 288, + 713 + ], + "type": "text", + "content": "Robustness against textual variations. Finally, we evaluate the impact of different textual styles on the PSR. We" + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 339, + 70, + 515, + 272 + ], + "blocks": [ + { + "bbox": [ + 339, + 70, + 515, + 272 + ], + "lines": [ + { + "bbox": [ + 339, + 70, + 515, + 272 + ], + "spans": [ + { + "bbox": [ + 339, + 70, + 515, + 272 + ], + "type": "image", + "image_path": "c7f83ace38d139841302d33c12412c2033b4e3fc897af1ce0d93cac392098f4e.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 281, + 547, + 370 + ], + "lines": [ + { + "bbox": [ + 304, + 281, + 547, + 370 + ], + "spans": [ + { + "bbox": [ + 304, + 281, + 547, + 370 + ], + "type": "text", + "content": "Figure 5. Top: Effect of makeup-based text guidance on the visual quality of the output images. Output images are able to impersonate the target identity for face verification. Text-prompt is \"tanned makeup with red lipstick\". Bottom: Optimizing over all latent codes changes the identity of the protected image. Our identity-preserving regularization enforces the adversarial optimization to search for latent codes that hide the perturbations in the makeup effect while simultaneously preserving visual identity." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "table", + "bbox": [ + 317, + 418, + 538, + 452 + ], + "blocks": [ + { + "bbox": [ + 305, + 381, + 547, + 415 + ], + "lines": [ + { + "bbox": [ + 305, + 381, + 547, + 415 + ], + "spans": [ + { + "bbox": [ + 305, + 381, + 547, + 415 + ], + "type": "text", + "content": "Table 6. Impact of different textual makeup styles on PSR. Makeup styles are \"tanned\", \"pale\", \"pink eyeshadows\", \"red lipstick\", and \"Matte\". Std. denotes standard deviation." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 317, + 418, + 538, + 452 + ], + "lines": [ + { + "bbox": [ + 317, + 418, + 538, + 452 + ], + "spans": [ + { + "bbox": [ + 317, + 418, + 538, + 452 + ], + "type": "table", + "html": "
t1makeupt2makeupt3makeupt4makeupt5makeupStd.
PSR74.177.378.478.779.21.24
", + "image_path": "f94ed57a383db61a642c2ef9e3d49e6275dcb92c2aae330a2fa5e1247f74151c.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "table_body" + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 467, + 547, + 528 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 467, + 547, + 528 + ], + "spans": [ + { + "bbox": [ + 304, + 467, + 547, + 528 + ], + "type": "text", + "content": "select five text-based makeup styles to protect 1000 images of CelebA-HQ using our method. Results in Tab. 6 show that PSR does not change significantly (low standard deviation) for different makeup styles, indicating robustness of our approach wrt different text-based makeup styles." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 537, + 379, + 550 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 537, + 379, + 550 + ], + "spans": [ + { + "bbox": [ + 306, + 537, + 379, + 550 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 558, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 558, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 558, + 547, + 715 + ], + "type": "text", + "content": "We have proposed a framework to protect privacy of face images on online platforms by carefully searching for adversarial codes in the low-dimensional latent manifold of a pre-trained generative model. We have shown that incorporating a makeup text-guided loss and an identity preserving regularization effectively hides the adversarial perturbations in the makeup style, provides images with high quality, and preserves human-perceived identity. While this approach is robust to the user-defined text-prompt and target identity, it would be beneficial if the text-prompt and target identity can be automatically selected based on the given face image. Limitations of our method include high computational cost at the time of protected face generation." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "20602" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 147 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 147 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 147 + ], + "type": "text", + "content": "[1] Shane Ahern, Dean Eckles, Nathaniel S Good, Simon King, Mor Naaman, and Rahul Nair. Over-exposed? privacy patterns and considerations in online and mobile photo sharing. In Proceedings of the SIGCHI conference on Human factors in computing systems, pages 357-366, 2007. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 148, + 288, + 203 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 148, + 288, + 203 + ], + "spans": [ + { + "bbox": [ + 53, + 148, + 288, + 203 + ], + "type": "text", + "content": "[2] Amit H Bermano, Rinon Gal, Yuval Alaluf, Ron Mokady, Yotam Nitzan, Omer Tov, Oren Patashnik, and Daniel Cohen-Or. State-of-the-art in the architecture, methods and applications of stylegan. In Computer Graphics Forum, volume 41, pages 591–611. Wiley Online Library, 2022. 2, 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 205, + 288, + 247 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 205, + 288, + 247 + ], + "spans": [ + { + "bbox": [ + 53, + 205, + 288, + 247 + ], + "type": "text", + "content": "[3] Anand Bhattachad, Min Jin Chong, Kaizhao Liang, Bo Li, and David A Forsyth. Unrestricted adversarial examples via semantic manipulation. arXiv preprint arXiv:1904.06347, 2019. 2, 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 250, + 288, + 305 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 250, + 288, + 305 + ], + "spans": [ + { + "bbox": [ + 53, + 250, + 288, + 305 + ], + "type": "text", + "content": "[4] Jia-Wei Chen, Li-Ju Chen, Chia-Mu Yu, and Chun-Shien Lu. Perceptual indistinguishability-net (pi-net): Facial image obfuscation with manipulable semantics. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6478–6487, 2021. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 307, + 287, + 351 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 307, + 287, + 351 + ], + "spans": [ + { + "bbox": [ + 53, + 307, + 287, + 351 + ], + "type": "text", + "content": "[5] Sheng Chen, Yang Liu, Xiang Gao, and Zhen Han. Mobile-facenets: Efficient cnns for accurate real-time face verification on mobile devices. In Chinese Conference on Biometric Recognition, pages 428-438. Springer, 2018. 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 352, + 287, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 352, + 287, + 407 + ], + "spans": [ + { + "bbox": [ + 53, + 352, + 287, + 407 + ], + "type": "text", + "content": "[6] Valeriia Cherepanova, Micah Goldblum, Harrison Foley, Shiyuan Duan, John P Dickerson, Gavin Taylor, and Tom Goldstein. Lowkey: Leveraging adversarial attacks to protect social media users from facial recognition. In International Conference on Learning Representations, 2020. 2, 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 409, + 287, + 452 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 409, + 287, + 452 + ], + "spans": [ + { + "bbox": [ + 53, + 409, + 287, + 452 + ], + "type": "text", + "content": "[7] William L Croft, Jörg-Rüdiger Sack, and Wei Shi. Differentially private facial obfuscation via generative adversarial networks. Future Generation Computer Systems, 129:358-379, 2022. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 454, + 288, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 454, + 288, + 498 + ], + "spans": [ + { + "bbox": [ + 53, + 454, + 288, + 498 + ], + "type": "text", + "content": "[8] Ali Dabouei, Sobhan Soleymani, Jeremy Dawson, and Nasser Nasrabadi. Fast geometrically-perturbed adversarial faces. In 2019 IEEE Winter Conference on Applications of Computer Vision (WACV), pages 1979-1988. IEEE, 2019. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 500, + 287, + 554 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 500, + 287, + 554 + ], + "spans": [ + { + "bbox": [ + 53, + 500, + 287, + 554 + ], + "type": "text", + "content": "[9] Jiankang Deng, Jia Guo, Niannan Xue, and Stefanos Zafeiriou. Arcface: Additive angular margin loss for deep face recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4690-4699, 2019. 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 555, + 287, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 555, + 287, + 612 + ], + "spans": [ + { + "bbox": [ + 48, + 555, + 287, + 612 + ], + "type": "text", + "content": "[10] Yinpeng Dong, Fangzhou Liao, Tianyu Pang, Hang Su, Jun Zhu, Xiaolin Hu, and Jianguo Li. Boosting adversarial attacks with momentum. In Proceedings of the 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR'18), pages 9185-9193, 2018. 6, 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 613, + 287, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 613, + 287, + 668 + ], + "spans": [ + { + "bbox": [ + 48, + 613, + 287, + 668 + ], + "type": "text", + "content": "[11] Yinpeng Dong, Tianyu Pang, Hang Su, and Jun Zhu. Evading defenses to transferable adversarial examples by translation-invariant attacks. In Proceedings of the 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR'19), pages 4312-4321, 2019. 6, 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 669, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 669, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 669, + 287, + 712 + ], + "type": "text", + "content": "[12] Hang Du, Hailin Shi, Dan Zeng, Xiao-Ping Zhang, and Tao Mei. The elements of end-to-end deep face recognition: A survey of recent advances. ACM Computing Surveys (CSUR), 54(10s):1-42, 2022. 2" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 714 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "text", + "content": "[13] Yifan Du, Zikang Liu, Junyi Li, and Wayne Xin Zhao. A survey of vision-language pre-trained models. arXiv preprint arXiv:2202.10936, 2022. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 107, + 545, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 107, + 545, + 139 + ], + "spans": [ + { + "bbox": [ + 307, + 107, + 545, + 139 + ], + "type": "text", + "content": "[14] Kevin Frans, Lisa B Soros, and Olaf Witkowski. Clipdraw: Exploring text-to-drawing synthesis through language-image encoders. arXiv preprint arXiv:2106.14843, 2021. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 140, + 545, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 140, + 545, + 182 + ], + "spans": [ + { + "bbox": [ + 308, + 140, + 545, + 182 + ], + "type": "text", + "content": "[15] Ian J. Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron C. Courville, and Yoshua Bengio. Generative adversarial networks. CoRR, abs/1406.2661, 2014. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 184, + 547, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 184, + 547, + 239 + ], + "spans": [ + { + "bbox": [ + 308, + 184, + 547, + 239 + ], + "type": "text", + "content": "[16] Qiao Gu, Guanzhi Wang, Mang Tik Chiu, Yu-Wing Tai, and Chi-Keung Tang. Ladin: Local adversarial disentangling network for facial makeup and de-makeup. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10481-10490, 2019. 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 239, + 545, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 239, + 545, + 282 + ], + "spans": [ + { + "bbox": [ + 308, + 239, + 545, + 282 + ], + "type": "text", + "content": "[17] Nitzan Guetta, Asaf Shabtai, Inderjeet Singh, Satoru Momiyama, and Yuval Elovici. Dodging attack using carefully crafted natural makeup. arXiv preprint arXiv:2109.06467, 2021. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 283, + 545, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 283, + 545, + 304 + ], + "spans": [ + { + "bbox": [ + 308, + 283, + 545, + 304 + ], + "type": "text", + "content": "[18] Rebecca Heilweil. The world's scariest facial recognition company explained. Vox, May, 8, 2020. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 305, + 545, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 305, + 545, + 358 + ], + "spans": [ + { + "bbox": [ + 308, + 305, + 545, + 358 + ], + "type": "text", + "content": "[19] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 360, + 545, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 360, + 545, + 381 + ], + "spans": [ + { + "bbox": [ + 308, + 360, + 545, + 381 + ], + "type": "text", + "content": "[20] Kashmir Hill. The secretive company that might end privacy as we know it. *The New York Times*, 18:2020, 2020. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 383, + 545, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 383, + 545, + 415 + ], + "spans": [ + { + "bbox": [ + 308, + 383, + 545, + 415 + ], + "type": "text", + "content": "[21] Jie Hu, Li Shen, and Gang Sun. Squeeze-and-excitation networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7132-7141, 2018. 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 416, + 545, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 416, + 545, + 480 + ], + "spans": [ + { + "bbox": [ + 308, + 416, + 545, + 480 + ], + "type": "text", + "content": "[22] Shengshan Hu, Xiaogeng Liu, Yechao Zhang, Minghui Li, Leo Yu Zhang, Hai Jin, and Libing Wu. Protecting facial privacy: Generating adversarial identity masks via style-robust makeup transfer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15014–15023, 2022. 1, 2, 3, 5, 6, 7" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 482, + 545, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 482, + 545, + 536 + ], + "spans": [ + { + "bbox": [ + 308, + 482, + 545, + 536 + ], + "type": "text", + "content": "[23] Gary B Huang, Marwan Mattar, Tamara Berg, and Eric Learned-Miller. Labeled faces in the wild: A database for studying face recognition in unconstrained environments. In Workshop on faces in 'Real-Life' Images: detection, alignment, and recognition, 2008. 6, 7" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 537, + 545, + 591 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 537, + 545, + 591 + ], + "spans": [ + { + "bbox": [ + 308, + 537, + 545, + 591 + ], + "type": "text", + "content": "[24] Phillip Isola, Jun-Yan Zhu, Tinghui Zhou, and Alexei A Efros. Image-to-image translation with conditional adversarial networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1125-1134, 2017. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 592, + 545, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 592, + 545, + 634 + ], + "spans": [ + { + "bbox": [ + 308, + 592, + 545, + 634 + ], + "type": "text", + "content": "[25] Kazuya Kakizaki and Kosuke Yoshida. Adversarial image translation: Unrestricted adversarial examples in face recognition systems. arXiv preprint arXiv:1905.03421, 2019. 2, 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 635, + 545, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 635, + 545, + 679 + ], + "spans": [ + { + "bbox": [ + 308, + 635, + 545, + 679 + ], + "type": "text", + "content": "[26] Tero Karras, Timo Aila, Samuli Laine, and Jaakko Lehtinen. Progressive growing of gans for improved quality, stability, and variation. In International Conference on Learning Representations, 2018. 6" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 681, + 547, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 681, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 308, + 681, + 547, + 714 + ], + "type": "text", + "content": "[27] Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of stylegan. In Proceedings of" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20603" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 73, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 67, + 73, + 287, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 73, + 287, + 95 + ], + "spans": [ + { + "bbox": [ + 67, + 73, + 287, + 95 + ], + "type": "text", + "content": "the IEEE/CVF conference on computer vision and pattern recognition, pages 8110-8119, 2020. 2, 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 95, + 287, + 148 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 95, + 287, + 148 + ], + "spans": [ + { + "bbox": [ + 48, + 95, + 287, + 148 + ], + "type": "text", + "content": "[28] Gwanghyun Kim, Taesung Kwon, and Jong Chul Ye. Diffusionclip: Text-guided diffusion models for robust image manipulation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2426-2435, 2022. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 150, + 287, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 150, + 287, + 194 + ], + "spans": [ + { + "bbox": [ + 48, + 150, + 287, + 194 + ], + "type": "text", + "content": "[29] Stepan Komkov and Aleksandr Petiushko. Advhat: Realworld adversarial attack on arcface face id system. In 2020 25th International Conference on Pattern Recognition (ICPR), pages 819-826. IEEE, 2021. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 195, + 287, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 195, + 287, + 239 + ], + "spans": [ + { + "bbox": [ + 48, + 195, + 287, + 239 + ], + "type": "text", + "content": "[30] Zhenzhong Kuang, Zhiqiang Guo, Jinglong Fang, Jun Yu, Noboru Babaguchi, and Jianping Fan. Unnoticeable synthetic face replacement for image privacy protection. Neurocomputing, 457:322-333, 2021. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 239, + 287, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 239, + 287, + 303 + ], + "spans": [ + { + "bbox": [ + 48, + 239, + 287, + 303 + ], + "type": "text", + "content": "[31] Zhenzhong Kuang, Huigui Liu, Jun Yu, Aikui Tian, Lei Wang, Jianping Fan, and Noboru Babaguchi. Effective de-identification generative adversarial network for face anonymization. In Proceedings of the 29th ACM International Conference on Multimedia, pages 3182–3191, 2021. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 305, + 287, + 349 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 305, + 287, + 349 + ], + "spans": [ + { + "bbox": [ + 48, + 305, + 287, + 349 + ], + "type": "text", + "content": "[32] Gihyun Kwon and Jong Chul Ye. Clipstyler: Image style transfer with a single text condition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18062-18071, 2022. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 350, + 287, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 350, + 287, + 381 + ], + "spans": [ + { + "bbox": [ + 48, + 350, + 287, + 381 + ], + "type": "text", + "content": "[33] Tao Li and Min Soo Choi. Deepblur: A simple and effective method for natural image obfuscation. arXiv preprint arXiv:2104.02655, 1, 2021. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 383, + 287, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 383, + 287, + 426 + ], + "spans": [ + { + "bbox": [ + 48, + 383, + 287, + 426 + ], + "type": "text", + "content": "[34] Tao Li and Lei Lin. Anonymousnet: Natural face de-identification with measurable privacy. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops, pages 0–0, 2019. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 427, + 287, + 469 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 427, + 287, + 469 + ], + "spans": [ + { + "bbox": [ + 48, + 427, + 287, + 469 + ], + "type": "text", + "content": "[35] Fangcheng Liu, Chao Zhang, and Hongyang Zhang. Towards transferable unrestricted adversarial examples with minimum changes. arXiv preprint arXiv:2201.01102, 2022. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 471, + 287, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 471, + 287, + 514 + ], + "spans": [ + { + "bbox": [ + 48, + 471, + 287, + 514 + ], + "type": "text", + "content": "[36] Suolan Liu, Lizhi Kong, and Hongyuan Wang. Face detection and encryption for privacy preserving in surveillance video. In Chinese Conference on Pattern Recognition and Computer Vision (PRCV), pages 162-172. Springer, 2018. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 515, + 287, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 515, + 287, + 568 + ], + "spans": [ + { + "bbox": [ + 48, + 515, + 287, + 568 + ], + "type": "text", + "content": "[37] Aleksander Madry, Aleksandar Makelov, Ludwig Schmidt, Dimitris Tsipras, and Adrian Vladu. Towards deep learning models resistant to adversarial attacks. In Proceedings of the 6th International Conference on Learning Representations (ICLR'18), 2018. 6, 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 570, + 287, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 570, + 287, + 624 + ], + "spans": [ + { + "bbox": [ + 48, + 570, + 287, + 624 + ], + "type": "text", + "content": "[38] Blaz Meden, Peter Rot, Philipp Terhörst, Naser Damer, Arjan Kuijper, Walter J Scheirer, Arun Ross, Peter Peer, and Vitomir Štruc. Privacy-enhancing face biometrics: A comprehensive survey. IEEE Transactions on Information Forensics and Security, 2021. 2, 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 625, + 287, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 625, + 287, + 658 + ], + "spans": [ + { + "bbox": [ + 48, + 625, + 287, + 658 + ], + "type": "text", + "content": "[39] Dongbin Na, Sangwoo Ji, and Jong Kim. Unrestricted blackbox adversarial attack using gan with limited queries. arXiv preprint arXiv:2208.11613, 2022. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 658, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 287, + 712 + ], + "type": "text", + "content": "[40] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741, 2021. 3" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 117 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 117 + ], + "type": "text", + "content": "[41] Seong Joon Oh, Rodrigo Benenson, Mario Fritz, and Bernt Schiele. Faceless person recognition: Privacy implications in social media. In European Conference on Computer Vision, pages 19-35. Springer, 2016. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 118, + 545, + 163 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 545, + 163 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 545, + 163 + ], + "type": "text", + "content": "[42] Seong Joon Oh, Mario Fritz, and Bernt Schiele. Adversarial image perturbation for privacy protection a game theory perspective. In 2017 IEEE International Conference on Computer Vision (ICCV), pages 1491-1500. IEEE, 2017. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 164, + 545, + 186 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 164, + 545, + 186 + ], + "spans": [ + { + "bbox": [ + 307, + 164, + 545, + 186 + ], + "type": "text", + "content": "[43] Omkar M Parkhi, Andrea Vedaldi, and Andrew Zisserman. Deep face recognition. 2015. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 186, + 545, + 240 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 186, + 545, + 240 + ], + "spans": [ + { + "bbox": [ + 308, + 186, + 545, + 240 + ], + "type": "text", + "content": "[44] Or Patashnik, Zongze Wu, Eli Shechtman, Daniel Cohen-Or, and Dani Lischinski. Styleclip: Text-driven manipulation of stylegan imagery. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2085–2094, 2021. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 243, + 545, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 243, + 545, + 319 + ], + "spans": [ + { + "bbox": [ + 308, + 243, + 545, + 319 + ], + "type": "text", + "content": "[45] P Jonathon Phillips, Amy N Yates, Ying Hu, Carina A Hahn, Eilidh Noyes, Kelsey Jackson, Jacqueline G Cava-zos, Geraldine Jeckeln, Rajeev Ranjan, Swami Sankaranarayanan, et al. Face recognition accuracy of forensic examiners, superrecognizers, and face recognition algorithms. Proceedings of the National Academy of Sciences, 115(24):6171-6176, 2018. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 321, + 545, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 321, + 545, + 376 + ], + "spans": [ + { + "bbox": [ + 308, + 321, + 545, + 376 + ], + "type": "text", + "content": "[46] Omid Poursaeed, Tianxing Jiang, Harry Yang, Serge Belongie, and Ser-Nam Lim. Robustness and generalization via generative adversarial training. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15711-15720, 2021. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 377, + 545, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 377, + 545, + 443 + ], + "spans": [ + { + "bbox": [ + 308, + 377, + 545, + 443 + ], + "type": "text", + "content": "[47] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 3, 5" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 445, + 545, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 445, + 545, + 510 + ], + "spans": [ + { + "bbox": [ + 308, + 445, + 545, + 510 + ], + "type": "text", + "content": "[48] Rajeev Ranjan, Swami Sankaranarayanan, Ankan Bansal, Navaneeth Bodla, Jun-Cheng Chen, Vishal M Patel, Carlos D Castillo, and Rama Chellappa. Deep learning for understanding faces: Machines may be just as good, or better, than humans. IEEE Signal Processing Magazine, 35(1):66-83, 2018. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 512, + 545, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 512, + 545, + 555 + ], + "spans": [ + { + "bbox": [ + 308, + 512, + 545, + 555 + ], + "type": "text", + "content": "[49] Daniel Roich, Ron Mokady, Amit H Bermano, and Daniel Cohen-Or. Pivotal tuning for latent-based editing of real images. ACM Transactions on Graphics (TOG), 42(1):1-13, 2022. 5" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 557, + 545, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 557, + 545, + 601 + ], + "spans": [ + { + "bbox": [ + 308, + 557, + 545, + 601 + ], + "type": "text", + "content": "[50] Tim Salimans, Ian Goodfellow, Wojciech Zaremba, Vicki Cheung, Alec Radford, and Xi Chen. Improved techniques for training gans. Advances in neural information processing systems, 29, 2016. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 602, + 545, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 602, + 545, + 646 + ], + "spans": [ + { + "bbox": [ + 308, + 602, + 545, + 646 + ], + "type": "text", + "content": "[51] Florian Schroff, Dmitry Kalenichenko, and James Philbin. Facenet: A unified embedding for face recognition and clustering. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 815-823, 2015. 6" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 647, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 647, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 308, + 647, + 545, + 712 + ], + "type": "text", + "content": "[52] Sachith Seneviratne, Nuran Kasthuriarachchi, Sanka Rasnayaka, Danula Hettiachchi, and Ridwan Shariffdeen. Does a face mask protect my privacy?: Deep learning to predict protected attributes from masked face images. In Australasian Joint Conference on Artificial Intelligence, pages 91-102. Springer, 2022. 3" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "20604" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "text", + "content": "[53] Shawn Shan, Emily Wenger, Jiayun Zhang, Huiying Li, Haitao Zheng, and Ben Y Zhao. Fawkes: Protecting privacy against unauthorized deep learning models. In 29th USENIX security symposium (USENIX Security 20), pages 1589-1604, 2020. 2, 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 129, + 288, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 129, + 288, + 173 + ], + "spans": [ + { + "bbox": [ + 49, + 129, + 288, + 173 + ], + "type": "text", + "content": "[54] Mahmood Sharif, Sruti Bhagavatula, Lujo Bauer, and Michael K Reiter. A general framework for adversarial examples with objectives. ACM Transactions on Privacy and Security (TOPS), 22(3):1-30, 2019. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 175, + 288, + 218 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 175, + 288, + 218 + ], + "spans": [ + { + "bbox": [ + 48, + 175, + 288, + 218 + ], + "type": "text", + "content": "[55] Yang Song, Rui Shu, Nate Kushman, and Stefano Ermon. Constructing unrestricted adversarial examples with generative models. Advances in Neural Information Processing Systems, 31, 2018. 2, 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 220, + 288, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 220, + 288, + 274 + ], + "spans": [ + { + "bbox": [ + 48, + 220, + 288, + 274 + ], + "type": "text", + "content": "[56] Qianru Sun, Liqian Ma, Seong Joon Oh, Luc Van Gool, Bernt Schiele, and Mario Fritz. Natural and effective obfuscation by head inpainting. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 5050-5059, 2018. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 275, + 288, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 275, + 288, + 319 + ], + "spans": [ + { + "bbox": [ + 48, + 275, + 288, + 319 + ], + "type": "text", + "content": "[57] Christian Szegedy, Wojciech Zaremba, Ilya Sutskever, Joan Bruna, Dumitru Erhan, Ian Goodfellow, and Rob Fergus. Intriguing properties of neural networks. In International Conference on Learning Representations, 2014. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 320, + 288, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 320, + 288, + 354 + ], + "spans": [ + { + "bbox": [ + 48, + 320, + 288, + 354 + ], + "type": "text", + "content": "[58] Huan Tian, Tianqing Zhu, and Wanlei Zhou. Fairness and privacy preservation for facial images: Gan-based methods. Computers & Security, 122:102902, 2022. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 354, + 288, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 354, + 288, + 398 + ], + "spans": [ + { + "bbox": [ + 48, + 354, + 288, + 398 + ], + "type": "text", + "content": "[59] Omer Tov, Yuval Alaluf, Yotam Nitzan, Or Patashnik, and Daniel Cohen-Or. Designing an encoder for stylegan image manipulation. ACM Transactions on Graphics (TOG), 40(4):1-14, 2021. 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 399, + 288, + 454 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 399, + 288, + 454 + ], + "spans": [ + { + "bbox": [ + 48, + 399, + 288, + 454 + ], + "type": "text", + "content": "[60] Can Wang, Menglei Chai, Mingming He, Dongdong Chen, and Jing Liao. Clip-nerf: Text-and-image driven manipulation of neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3835-3844, 2022. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 456, + 288, + 477 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 456, + 288, + 477 + ], + "spans": [ + { + "bbox": [ + 48, + 456, + 288, + 477 + ], + "type": "text", + "content": "[61] Mei Wang and Weihong Deng. Deep face recognition: A survey. Neurocomputing, 429:215-244, 2021. 1, 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 478, + 288, + 522 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 478, + 288, + 522 + ], + "spans": [ + { + "bbox": [ + 48, + 478, + 288, + 522 + ], + "type": "text", + "content": "[62] Shunxin Wang, Una M Kelly, and Raymond NJ Veldhuis. Gender obfuscation through face morphing. In 2021 IEEE International Workshop on Biometrics and Forensics (IWBF), pages 1-6. IEEE, 2021. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 523, + 288, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 523, + 288, + 578 + ], + "spans": [ + { + "bbox": [ + 48, + 523, + 288, + 578 + ], + "type": "text", + "content": "[63] Ya Wang, Tianlong Bao, Chunhui Ding, and Ming Zhu. Face recognition in real-world surveillance videos with deep learning method. In 2017 2nd international conference on image, vision and computing (icivc), pages 239-243. IEEE, 2017. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 579, + 288, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 579, + 288, + 611 + ], + "spans": [ + { + "bbox": [ + 48, + 579, + 288, + 611 + ], + "type": "text", + "content": "[64] Yinggui Wang, Jian Liu, Man Luo, Le Yang, and Li Wang. Privacy-preserving face recognition in the frequency domain. 2022. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 613, + 288, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 613, + 288, + 657 + ], + "spans": [ + { + "bbox": [ + 48, + 613, + 288, + 657 + ], + "type": "text", + "content": "[65] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600-612, 2004. 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 658, + 288, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 288, + 713 + ], + "type": "text", + "content": "[66] Tianyi Wei, Dongdong Chen, Wenbo Zhou, Jing Liao, Zhentao Tan, Lu Yuan, Weiming Zhang, and Nenghai Yu. Hairclip: Design your hair by text and reference image. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18072-18081, 2022. 3" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 699 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "text", + "content": "[67] Emily Wenger, Shawn Shan, Haitao Zheng, and Ben Y Zhao. Sok: Anti-facial recognition technology. arXiv preprint arXiv:2112.04558, 2021. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 107, + 545, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 107, + 545, + 140 + ], + "spans": [ + { + "bbox": [ + 307, + 107, + 545, + 140 + ], + "type": "text", + "content": "[68] Chaowei Xiao, Jun-Yan Zhu, Bo Li, Warren He, Mingyan Liu, and Dawn Song. Spatially transformed adversarial examples. arXiv preprint arXiv:1801.02612, 2018. 2, 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 141, + 545, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 141, + 545, + 206 + ], + "spans": [ + { + "bbox": [ + 308, + 141, + 545, + 206 + ], + "type": "text", + "content": "[69] Zihao Xiao, Xianfeng Gao, Chilin Fu, Yinpeng Dong, Wei Gao, Xiaolu Zhang, Jun Zhou, and Jun Zhu. Improving transferability of adversarial patches on face recognition with generative models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11845-11854, 2021. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 207, + 545, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 207, + 545, + 262 + ], + "spans": [ + { + "bbox": [ + 308, + 207, + 545, + 262 + ], + "type": "text", + "content": "[70] Xiao Yang, Yinpeng Dong, Tianyu Pang, Hang Su, Jun Zhu, Yuefeng Chen, and Hui Xue. Towards face encryption by generating adversarial identity masks. In Proceedings of the 2021 IEEE/CVF International Conference on Computer Vision (ICCV'21), pages 3897-3907, 2021. 1, 2, 3, 6, 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 263, + 545, + 328 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 263, + 545, + 328 + ], + "spans": [ + { + "bbox": [ + 308, + 263, + 545, + 328 + ], + "type": "text", + "content": "[71] Bangjie Yin, Wenxuan Wang, Taiping Yao, Junfeng Guo, Zelun Kong, Shouhong Ding, Jilin Li, and Cong Liu. Advmakeup: A new imperceptible and transferable attack on face recognition. In Proceedings of the 30th International Joint Conference on Artificial Intelligence (IJCAI'21), pages 1252-1258, 2021. 2, 6, 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 330, + 545, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 330, + 545, + 374 + ], + "spans": [ + { + "bbox": [ + 308, + 330, + 545, + 374 + ], + "type": "text", + "content": "[72] Bangjie Yin, Wenxuan Wang, Taiping Yao, Junfeng Guo, Zelun Kong, Shouhong Ding, Jilin Li, and Cong Liu. Advmakeup: A new imperceptible and transferable attack on face recognition. arXiv preprint arXiv:2105.03162, 2021. 2, 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 375, + 545, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 375, + 545, + 418 + ], + "spans": [ + { + "bbox": [ + 308, + 375, + 545, + 418 + ], + "type": "text", + "content": "[73] Shengming Yuan, Qilong Zhang, Lianli Gao, Yaya Cheng, and Jingkuan Song. Natural color fool: Towards boosting black-box unrestricted attacks. arXiv preprint arXiv:2210.02041, 2022. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 420, + 545, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 420, + 545, + 464 + ], + "spans": [ + { + "bbox": [ + 308, + 420, + 545, + 464 + ], + "type": "text", + "content": "[74] Jiaming Zhang, Jitao Sang, Xian Zhao, Xiaowen Huang, Yanfeng Sun, and Yongli Hu. Adversarial privacy-preserving filter. In Proceedings of the 28th ACM International Conference on Multimedia, pages 1423-1431, 2020. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 464, + 545, + 508 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 464, + 545, + 508 + ], + "spans": [ + { + "bbox": [ + 308, + 464, + 545, + 508 + ], + "type": "text", + "content": "[75] Kaipeng Zhang, Zhanpeng Zhang, Zhifeng Li, and Yu Qiao. Joint face detection and alignment using multitask cascaded convolutional networks. IEEE signal processing letters, 23(10):1499-1503, 2016. 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 509, + 545, + 564 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 509, + 545, + 564 + ], + "spans": [ + { + "bbox": [ + 308, + 509, + 545, + 564 + ], + "type": "text", + "content": "[76] Zhengyu Zhao, Zhuoran Liu, and Martha Larson. Towards large yet imperceptible adversarial image perturbations with perceptual color distance. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1039-1048, 2020. 2, 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 565, + 545, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 565, + 545, + 609 + ], + "spans": [ + { + "bbox": [ + 308, + 565, + 545, + 609 + ], + "type": "text", + "content": "[77] Yaoyao Zhong and Weihong Deng. Opom: Customized invisible cloak towards face privacy protection. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2022, 2, 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 609, + 545, + 654 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 609, + 545, + 654 + ], + "spans": [ + { + "bbox": [ + 308, + 609, + 545, + 654 + ], + "type": "text", + "content": "[78] Jizhe Zhou and Chi-Man Pun. Personal privacy protection via irrelevant faces tracking and pixelation in video live streaming. IEEE Transactions on Information Forensics and Security, 16:1088-1103, 2020. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 655, + 545, + 699 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 655, + 545, + 699 + ], + "spans": [ + { + "bbox": [ + 308, + 655, + 545, + 699 + ], + "type": "text", + "content": "[79] Zheng-An Zhu, Yun-Zhong Lu, and Chen-Kuo Chiang. Generating adversarial examples by makeup attacks on face recognition. In 2019 IEEE International Conference on Image Processing (ICIP), pages 2516-2520. IEEE, 2019. 3" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "20605" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/CLIP2Scene_ Towards Label-Efficient 3D Scene Understanding by CLIP/58dade4c-9eec-4c3a-968d-a582ab672d68_content_list.json b/2023/CLIP2Scene_ Towards Label-Efficient 3D Scene Understanding by CLIP/58dade4c-9eec-4c3a-968d-a582ab672d68_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..c140a07612cef5b8835b67a068ef16253e72fa3c --- /dev/null +++ b/2023/CLIP2Scene_ Towards Label-Efficient 3D Scene Understanding by CLIP/58dade4c-9eec-4c3a-968d-a582ab672d68_content_list.json @@ -0,0 +1,1321 @@ +[ + { + "type": "text", + "text": "CLIP2Scene: Towards Label-efficient 3D Scene Understanding by CLIP", + "text_level": 1, + "bbox": [ + 120, + 130, + 849, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Runnan Chen $^{1,2}$ Youquan Liu $^{2,3}$ Lingdong Kong $^{2,4}$ Xinge Zhu $^{5}$ Yuexin Ma $^{6}$ Yikang Li $^{2}$ Yuenan Hou $^{2,\\dagger}$ Yu Qiao $^{2}$ Wenping Wang $^{7,\\dagger}$", + "bbox": [ + 91, + 178, + 879, + 217 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1The University of Hong Kong 2Shanghai AI Laboratory 3Hochschule Bremerhaven 4National University of Singapore 5The Chinese University of Hong Kong 6ShanghaiTech University 7Texas A&M University", + "bbox": [ + 109, + 219, + 861, + 253 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 303, + 313, + 318 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Contrastive Language-Image Pre-training (CLIP) achieves promising results in 2D zero-shot and few-shot learning. Despite the impressive performance in 2D, applying CLIP to help the learning in 3D scene understanding has yet to be explored. In this paper, we make the first attempt to investigate how CLIP knowledge benefits 3D scene understanding. We propose CLIP2Scene, a simple yet effective framework that transfers CLIP knowledge from 2D image-text pre-trained models to a 3D point cloud network. We show that the pre-trained 3D network yields impressive performance on various downstream tasks, i.e., annotation-free and fine-tuning with labelled data for semantic segmentation. Specifically, built upon CLIP, we design a Semantic-driven Cross-modal Contrastive Learning framework that pre-trains a 3D network via semantic and spatial-temporal consistency regularization. For the former, we first leverage CLIP's text semantics to select the positive and negative point samples and then employ the contrastive loss to train the 3D network. In terms of the latter, we force the consistency between the temporally coherent point cloud features and their corresponding image features. We conduct experiments on SemanticKITTI, nuScenes, and ScanNet. For the first time, our pre-trained network achieves annotation-free 3D semantic segmentation with $20.8\\%$ and $25.08\\%$ mIoU on nuScenes and ScanNet, respectively. When fine-tuned with $1\\%$ or $100\\%$ labelled data, our method significantly outperforms other self-supervised methods, with improvements of $8\\%$ and $1\\%$ mIoU, respectively. Furthermore, we demonstrate the generalizability for handling cross-domain datasets. Code is publicly available1.", + "bbox": [ + 73, + 330, + 473, + 785 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 78, + 809, + 209, + 825 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3D scene understanding is fundamental in autonomous driving, robot navigation, etc [26,28]. Current deep learning-", + "bbox": [ + 76, + 834, + 470, + 866 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/e635707803576d3ce4da203fd52c9028e20daf537b797f5ed1eded7a308ab370.jpg", + "image_caption": [ + "Semantic-driven Cross-modal Contrastive Learning", + "Figure 1. We explore how CLIP knowledge benefits 3D scene understanding. To this end, we propose CLIP2Scene, a Semantic-driven Cross-modal Contrastive Learning framework that leverages CLIP knowledge to pre-train a 3D point cloud segmentation network via semantic and spatial-temporal consistency regularization. CLIP2Scene yields impressive performance on annotation-free 3D semantic segmentation and significantly outperforms other self-supervised methods when fine-tuning on annotated data." + ], + "image_footnote": [], + "bbox": [ + 506, + 311, + 877, + 535 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "based methods have shown inspirational performance on 3D point cloud data [15, 32, 33, 38, 47, 56, 62]. However, some drawbacks hinder their real-world applications. The first one comes from their heavy reliance on the large collection of annotated point clouds, especially when high-quality 3D annotations are expensive to acquire [39, 40, 44, 51]. Besides, they typically fail to recognize novel objects that are never seen in the training data [11, 45]. As a result, it may need extra annotation efforts to train the model on recognizing these novel objects, which is both tedious and time-consuming.", + "bbox": [ + 496, + 684, + 893, + 835 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Contrastive Vision-Language Pre-training (CLIP) [48] provides a new perspective that mitigates the above issues in 2D vision. It was trained on large-scale free-available image-text pairs from websites and built vision-language", + "bbox": [ + 496, + 839, + 893, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "Symbol $\\dagger$ denotes the corresponding authors. \n1https://github.com/runnanchen/CLIP2Scene.", + "bbox": [ + 94, + 875, + 423, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "7020", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "correlation to achieve promising open-vocabulary recognition. MaskCLIP [61] further explores semantic segmentation based on CLIP. With minimal modifications to the CLIP pre-trained network, MaskCLIP can be directly used for the semantic segmentation of novel objects without additional training efforts. PointCLIP [59] reveals that the zero-shot classification ability of CLIP can be generalized from the 2D image to the 3D point cloud. It perspectively projects a point cloud frame into different views of 2D depth maps that bridge the modal gap between the image and the point cloud. The above studies indicate the potential of CLIP on enhancing the 2D segmentation and 3D classification performance. However, whether and how CLIP knowledge benefits 3D scene understanding is still under-explored.", + "bbox": [ + 75, + 90, + 472, + 303 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we explore how to leverage CLIP's 2D image-text pre-learned knowledge for 3D scene understanding. Previous cross-modal knowledge distillation methods [44,51] suffer from the optimization-conflict issue, i.e., some of the positive pairs are regarded as negative samples for contrastive learning, leading to unsatisfactory representation learning and hammering the performance of downstream tasks. Besides, they also ignore the temporal coherence of the multi-sweep point cloud, failing to utilize the rich inter-sweep correspondence. To handle the mentioned problems, we propose a novel Semantic-driven Cross-modal Contrastive Learning framework that fully leverages CLIP's semantic and visual information to regularize a 3D network. Specifically, we propose Semantic Consistency Regularization and Spatial-Temporal Consistency Regularization. In semantic consistency regularization, we utilize CLIP's text semantics to select the positive and negative point samples for less-conflict contrastive learning. For spatial-temporal consistency regularization, we take CLIP's image pixel feature to impose a soft consistency constraint on the temporally coherent point features. Such an operation also alleviates the effects of imperfect image-to-point calibration.", + "bbox": [ + 75, + 306, + 472, + 638 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We conduct several downstream tasks on the indoor and outdoor datasets to verify how the pre-trained network benefits the 3D scene understanding. The first one is annotation-free semantic segmentation. Following MaskCLIP, we place class names into multiple hand-crafted templates as prompts and average the text embeddings generated by CLIP to conduct the annotation-free segmentation. For the first time, our method achieves $20.8\\%$ and $25.08\\%$ mIoU annotation-free 3D semantic segmentation on the nuScenes [24] and ScanNet [20] datasets without training on any labelled data. Secondly, we compare with other self-supervised methods in label-efficient learning. When fine-tuning the 3D network with $1\\%$ or $100\\%$ labelled data on the nuScenes dataset, our method significantly outperforms state-of-the-art self-supervised methods, with improvements of $8\\%$ and $1\\%$ mIoU, respectively. Besides, to verify the generalization capability, we pre-train the network on the nuScenes dataset", + "bbox": [ + 75, + 643, + 472, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "and evaluate it on SemanticKITTI [3]. Our method still significantly outperforms state-of-the-art methods. The key contributions of our work are summarized as follows.", + "bbox": [ + 496, + 90, + 890, + 137 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- The first work that distils CLIP knowledge to a 3D network for 3D scene understanding.", + "- We propose a novel Semantic-driven Cross-modal Contrastive Learning framework that pre-trains a 3D network via spatial-temporal and semantic consistency regularization.", + "- We propose a novel Semantic-guided Spatial-Temporal Consistency Regularization that forces the consistency between the temporally coherent point cloud features and their corresponding image features.", + "- For the first time, our method achieves promising results on annotation-free 3D scene segmentation. When fine-tuning with labelled data, our method significantly outperforms state-of-the-art self-supervised methods." + ], + "bbox": [ + 517, + 146, + 893, + 392 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 406, + 640, + 422 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Zero-shot Learning in 3D. The objective of zero-shot learning (ZSL) is to recognize objects that are unseen in the training set. Many efforts have been devoted to the 2D recognition tasks [1, 2, 4, 8, 21, 25, 34, 37, 41–43, 46, 54, 55, 58, 60], and few works concentrate on performing ZSL in the 3D domain [11, 16–18, 45]. [18] applies ZSL to 3D tasks, where they train PointNet [47] on \"seen\" samples and test on \"unseen\" samples. Subsequent work [16] addresses the hubness problem caused by the low-quality point cloud features. [17] proposes the triplet loss to boost the performance under the transductive setting, where the \"unseen\" class is observed and unlabeled in the training phase. [11] makes the first attempt to explore the transductive zero-shot segmentation for 3D scene understanding. Recently, some studies introduced CLIP into zero-shot learning. MaskCLIP [61] investigates the problem of utilizing CLIP to help the 2D dense prediction tasks and exhibits encouraging zero-shot semantic segmentation performance. PointCLIP [59] is the pioneering work that applies CLIP to 3D recognition and shows impressive performance on zero-shot and few-shot classification tasks. Our work takes a step further to investigate how the rich semantic and visual knowledge in CLIP can benefit the 3D semantic segmentation tasks.", + "bbox": [ + 496, + 431, + 893, + 779 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Self-supervised Representation Learning. The purpose of self-supervised learning is to obtain a good representation that benefits the downstream tasks. The dominant approaches resort to contrastive learning to pre-train the network [7, 9, 9, 10, 12-14, 22, 23, 27, 30]. Recently, inspired by the success of CLIP, leveraging the pre-trained model of CLIP to the downstream tasks has raised the community's attention [35, 36, 49, 50, 57]. DenseCLIP [49] utilizes", + "bbox": [ + 496, + 780, + 895, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "7021", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/0c613f286b55db46731f70fd01a88b176c9d596b00fc0a3d0db65877657647eb.jpg", + "image_caption": [ + "Figure 2. Illustration of the Semantic-driven Cross-modal Contrastive Learning. Firstly, we obtain the text embedding $t_i$ , image pixel feature $x_i$ , and point feature $p_i$ by text encoder, image encoder, and point encoder, respectively. Secondly, we leverage CLIP knowledge to construct positive and negative samples for contrastive learning. Thus we obtain point-text pairs $\\{x_i, t_i\\}_{i=1}^M$ and all pixel-point-text pairs in a short temporal $\\{\\hat{x}_i^k, \\hat{p}_i^k, t_i^k\\}_{i=1,k=1}^M$ . Here, $\\{x_i, t_i\\}_{i=1}^M$ and $\\{\\hat{x}_i^k, \\hat{p}_i^k, t_i^k\\}_{i=1,k=1}^M$ are used for Semantic Consistency Regularization and Spatial-Temporal Consistency Regularization, respectively. Lastly, we perform Semantic Consistency Regularization by pulling the point features to their corresponding text embedding and Spatial-Temporal Consistency Regularization by mimicking the temporally coherent point features to their corresponding pixel features." + ], + "image_footnote": [], + "bbox": [ + 89, + 92, + 877, + 377 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "the CLIP's pre-trained knowledge for dense image pixel prediction. DetCLIP [57] proposes a pre-training method equipped with CLIP for open-world detection. We leverage the image-text pre-trained CLIP knowledge to help 3D scene understanding.", + "bbox": [ + 75, + 508, + 470, + 584 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Cross-modal Knowledge Distillation. Recently, increasing studies have focused on transferring knowledge from 2D images to 3D point clouds for self-supervised representation learning [44, 51]. PPKT [44] resorts to the InfoNCE loss to help the 3D network distil rich knowledge from the 2D image backbone. SLidR [51] further introduce the superpixel to boost the cross-modal knowledge distillation. In this paper, we first attempt to pre-train a 3D network with CLIP's knowledge.", + "bbox": [ + 75, + 585, + 470, + 720 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Methodology", + "text_level": 1, + "bbox": [ + 76, + 738, + 210, + 756 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Considering the impressive open-vocabulary performance achieved by CLIP in image classification and segmentation, natural curiosities have been raised. Can CLIP endow the ability to a 3D network for annotation-free scene understanding? And further, will it promote the network performance when fine-tuned on labelled data? To answer the above questions, we study the cross-modal knowledge transfer of CLIP for 3D scene understanding, termed CLIP2Scene. Our work is a pioneer in exploiting CLIP knowledge for 3D scene", + "bbox": [ + 75, + 763, + 470, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "understanding. In what follows, we revisit the CLIP applied in 2D open-vocabulary classification and semantic segmentation, then present our CLIP2Scene in detail. Our approach consists of three major components: Semantic Consistency Regularization, Semantic-Guided Spatial-Temporal Consistency Regularization, and Switchable Self-Training Strategy.", + "bbox": [ + 496, + 508, + 893, + 602 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Revisiting CLIP", + "text_level": 1, + "bbox": [ + 500, + 614, + 661, + 633 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Contrastive Vision-Language Pre-training (CLIP) mitigates the following drawbacks that dominate the computer vision field: 1. Deep models need a large amount of formatted and labelled training data, which is expensive to acquire; 2. The model's generalization ability is weak, making it difficult to migrate to a new scenario with unseen objects. CLIP consists of an image encoder (ResNet [31] or ViT [6]) and a text encoder (Transformer [53]), both respectively project the image and text representation to a joint embedding space. During training, CLIP constructs positive and negative samples from 400 million image-text pairs to train both encoders with a contrastive loss, where the large-scale image-text pairs are free-available from the Internet and assumed to contain every class of images and most concepts of text. Therefore, CLIP can achieve promising open-vocabulary recognition.", + "bbox": [ + 496, + 641, + 893, + 868 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "For 2D zero-shot classification, CLIP first places the class name into a pre-defined template to generate the text embed-", + "bbox": [ + 500, + 869, + 893, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "7022", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/abb87bad06a099b3a04ab1504989c71b187a5965afe07839adca112ec8c71c3c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 81, + 94, + 473, + 281 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/df1d6e894bf9f8cd67b1d613b7c758a6cf027723e49a24df834de1fab5e2dd5d.jpg", + "image_caption": [ + "Figure 3. Illustration of the image pixel-to-text mapping. The dense pixel-text correspondence $\\{x_{i},t_{i}\\}_{i = 1}^{M}$ is extracted by the off-the-shelf method MaskCLIP [61]." + ], + "image_footnote": [], + "bbox": [ + 80, + 282, + 464, + 431 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "dings and then encodes images to obtain image embeddings. Next, it calculates the similarities between images and text embeddings to determine the class. MaskCLIP further extends CLIP into 2D semantic segmentation. Specifically, MaskCLIP modifies the attention pooling layer of the CLIP's image encoder, thus performing pixel-level mask prediction instead of the global image-level prediction.", + "bbox": [ + 75, + 500, + 472, + 606 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. CLIP2Scene", + "text_level": 1, + "bbox": [ + 76, + 617, + 209, + 632 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As shown in Fig. 2, we first leverage CLIP and 3D network to respectively extract the text embeddings, image pixel feature, and point feature. Secondly, we construct positive and negative samples based on CLIP's knowledge. Lastly, we impose Semantic Consistency Regularization by pulling the point features to their corresponding text embedding. At the same time, we apply Spatial-Temporal Consistency Regularization by forcing the consistency between temporally coherent point features and their corresponding pixel features. In what follows, we present the details and insights.", + "bbox": [ + 75, + 642, + 472, + 792 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2.1 Semantic Consistency Regularization", + "text_level": 1, + "bbox": [ + 76, + 815, + 390, + 830 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As CLIP is pre-trained on 2D images and text, our first concern is the domain gap between 2D images and the 3D point cloud. To this end, we build dense pixel-point correspondence and transfer image knowledge to the 3D point cloud", + "bbox": [ + 75, + 839, + 470, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "via the pixel-point pairs. Specifically, we calibrate the Li-DAR point cloud with corresponding images captured by six cameras. Therefore, the dense pixel-point correspondence $\\{x_i,p_i\\}_{i = 1}^M$ can be obtained accordingly, where $x_{i}$ and $p_i$ indicates $i$ -th paired image feature and point feature, which are respectively extracted by the CLIP's image encoder and the 3D network. $M$ is the number of pairs. Note that it is an online operation and is irrelevant to the image and point data augmentation.", + "bbox": [ + 496, + 90, + 893, + 226 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Previous methods [44, 51] provide a promising solution to cross-modal knowledge transfer. They first construct positive pixel-point pairs $\\{x_i, p_i\\}_{i=1}^M$ and negative pairs $\\{x_i, p_j\\} (i \\neq j)$ , and then pull in the positive pairs while pushing away the negative pairs in the embedding space via the InfoNCE loss. Despite the encourageable performance of previous methods in transferring cross-modal knowledge, they are both confronted with the same optimization-conflict issue. For example, suppose $i$ -th pixel $x_i$ and $j$ -th point $p_j$ are in the different positions of the same instance with the same semantics. However, the InfoNCE loss will try to push them away, which is unreasonable and hammer the performance of the downstream tasks [51]. In light of this, we propose a Semantic Consistency Regularization that leverages the CLIP's semantic information to alleviate this issue. Specifically, we generate the dense pixel-text pairs $\\{x_i, t_i\\}_{i=1}^M$ by following the off-the-shelf method MaskCLIP [61] (Fig. 3), where $t_i$ is the text embedding generated from the CLIP's text encoder. Note that the pixel-text mappings are free-available from CLIP without any additional training. We then transfer pixel-text pairs to point-text pairs $\\{p_i, t_i\\}_{i=1}^M$ and utilize the text semantics to select the positive and negative point samples for contrastive learning. The objective function is as follows:", + "bbox": [ + 496, + 227, + 893, + 589 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {S - i n f o} = - \\sum_ {c = 1} ^ {C} \\log \\frac {\\sum_ {t _ {i} \\in c , p _ {i}} \\exp \\left(D \\left(t _ {i} , p _ {i}\\right) / \\tau\\right)}{\\sum_ {t _ {i} \\in c , t _ {j} \\notin c , p _ {j}} \\exp \\left(D \\left(t _ {i} , p _ {j}\\right) / \\tau\\right)}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 514, + 609, + 890, + 662 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $t_i \\in c$ indicates that $t_i$ is generated by $c$ -th classes name, and $C$ is the number of classes. $D$ denotes the scalar product operation and $\\tau$ is a temperature term $(\\tau > 0)$ .", + "bbox": [ + 496, + 662, + 890, + 708 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Since the text is composed of class names placed into pre-defined templates, the text embedding represents the semantic information of the corresponding class. Therefore, those points with the same semantics will be restricted near the same text embedding, and those with different semantics will be pushed away. To this end, our Semantic Consistency Regularization causes less conflict in contrastive learning.", + "bbox": [ + 496, + 708, + 890, + 814 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2.2 Semantic-guided Spatial-temporal Consistency Regularization", + "text_level": 1, + "bbox": [ + 500, + 830, + 890, + 862 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Besides semantic consistency regularization, we consider how image pixel features help to regularize a 3D network.", + "bbox": [ + 498, + 869, + 893, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "7023", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/c03b35ee78740b90f8de34ea9faf125257f71553320e469e344e2ff96c32183f.jpg", + "image_caption": [ + "Figure 4. Illustration of the image pixel-to-point mapping (left) and semantic-guided fusion feature generation (right). We build the grid-wise correspondence between an image $I$ and the temporally coherent LiDAR point cloud $\\{P_k\\}_{k=1}^K$ within $S$ seconds and generate semantic-guided fusion features for individual grids. Both $\\{\\hat{x}_i^k, \\hat{p}_i^k\\}_{i=1,k=1}^M, K$ and $\\{f_n\\}_{n=1}^N$ are used to perform Spatial-Temporal Consistency Regularization." + ], + "image_footnote": [], + "bbox": [ + 75, + 90, + 477, + 224 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The natural alternative directly pulls in the point feature with its corresponding pixel in the embedding space. However, the noise-assigned semantics of the image pixel and the imperfect pixel-point mapping hinder the downstream task's performance. To this end, we propose a novel semantic-guided Spatial-Temporal Consistency Regularization to alleviate the problem by imposing a soft constraint on points within local space and time.", + "bbox": [ + 75, + 354, + 472, + 477 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Specifically, given an image $I$ and temporally coherent LiDAR point cloud $\\{P_k\\}_{k=1}^K$ , where $K$ is the number of sweeps within $S$ seconds. Note that the image is matched to the first frame of the point cloud $P_1$ with pixel-point pairs $\\{\\hat{x}_i^1, \\hat{p}_i^1\\}_{i=1}^M$ . We register the rest of the point cloud to the first frame via the calibration matrices and map them to the image (Fig. 4). Thus we obtain all pixel-point-text pairs in a short temporal $\\{\\hat{x}_i^k, \\hat{p}_i^k, t_i^k\\}_{i=1,k=1}^M$ . Next, we divide the entire stitched point cloud into regular grids $\\{g_n\\}_{n=1}^N$ , where the temporally coherent points are located in the same grid. We impose the spatial-temporal consistency constraint within individual grids by the following objective function:", + "bbox": [ + 75, + 479, + 472, + 672 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {S S R} = \\sum_ {g _ {n}} \\sum_ {(\\hat {i}, \\hat {k}) \\in g _ {n}} (1 - \\operatorname {s i g m o i d} \\left(D \\left(\\hat {p} _ {\\hat {i}} ^ {\\hat {k}}, f _ {n}\\right)\\right)) / N, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 88, + 703, + 470, + 741 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $(\\hat{i},\\hat{k})\\in g_{n}$ indicates the pixel-point pair $\\{\\hat{x}_i^k,\\hat{p}_i^k\\}$ is located in the $n$ -th grid. $\\{f_n\\}_{n = 1}^N$ is a semantic-guided cross-modal fusion feature formulated by:", + "bbox": [ + 75, + 760, + 468, + 809 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nf _ {n} = \\sum_ {(\\hat {i}, \\hat {k}) \\in g _ {n}} a _ {\\hat {i}} ^ {\\hat {k}} * \\hat {x} _ {\\hat {i}} ^ {\\hat {k}} + b _ {\\hat {i}} ^ {\\hat {k}} * \\hat {p} _ {\\hat {i}} ^ {\\hat {k}}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 163, + 825, + 470, + 863 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $a_{\\hat{i}}^{\\hat{k}}$ and $b_{\\hat{i}}^{\\hat{k}}$ are attention weight calculated by:", + "bbox": [ + 76, + 882, + 418, + 905 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} a _ {i} ^ {\\hat {k}} = \\frac {\\exp (D (\\hat {x} _ {\\hat {i}} ^ {\\hat {k}} , t _ {\\hat {i}} ^ {1}) / \\lambda)}{\\sum_ {(\\hat {i} , \\hat {k}) \\in g _ {n}} \\exp (D (\\hat {x} _ {\\hat {i}} ^ {\\hat {k}} , t _ {\\hat {i}} ^ {1}) / \\lambda) + \\exp (D (\\hat {p} _ {\\hat {i}} ^ {\\hat {k}} , t _ {\\hat {i}} ^ {1}) / \\lambda)}, \\\\ b _ {\\hat {i}} ^ {\\hat {k}} = \\frac {\\exp \\left(D \\left(\\hat {p} _ {\\hat {i}} ^ {\\hat {k}} , t _ {\\hat {i}} ^ {1}\\right) / \\lambda\\right)}{\\sum_ {(\\hat {i} , \\hat {k}) \\in g _ {n}} \\exp \\left(D \\left(\\hat {x} _ {\\hat {i}} ^ {\\hat {k}}, t _ {\\hat {i}} ^ {1}\\right) / \\lambda\\right) + \\exp \\left(D \\left(\\hat {p} _ {\\hat {i}} ^ {\\hat {k}}, t _ {\\hat {i}} ^ {1}\\right) / \\lambda\\right)}, \\tag {4} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 506, + 114, + 892, + 220 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\lambda$ is the temperature term.", + "bbox": [ + 498, + 232, + 718, + 244 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Actually, those pixel and point features within the local grid $g_{n}$ are restricted near a dynamic centre $f_{n}$ . Thus, such a soft constraint alleviates the noisy prediction and calibration error issues. At the same time, it imposes Spatio-Temporal Regularization on the temporally coherent point features.", + "bbox": [ + 496, + 246, + 893, + 323 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2.3 Switchable Self-training Strategy", + "text_level": 1, + "bbox": [ + 498, + 340, + 784, + 357 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We combine the loss function $\\mathcal{L}_{S,\\text{info}}$ and $\\mathcal{L}_{SSR}$ to end-to-end train the whole network, where the CLIP's image and text encoder backbone are frozen during training. We find that method worked only when the pixel-point feature $\\{x_i,p_i\\}_{i = 1}^M$ and $\\{\\hat{x}_i^k,\\hat{p}_i^k\\}_{i = 1,k = 1}^{\\hat{M},K}$ , which are used in $\\mathcal{L}_{S,\\text{info}}$ and $\\mathcal{L}_{SSR}$ , are generated from different learnable linear layer. On top of that, we further put forward an effective strategy to promote performance. Specifically, after contrastive learning of the 3D network for a few epochs, we randomly switch the point pseudo label between the paired image pixel's pseudo label and the point's predicted label. Since different modality networks learn different feature representations, they can filter different types of error introduced by noisy pseudo labels. By this switchable operation, the error flows can be reduced by mutually [29].", + "bbox": [ + 496, + 364, + 893, + 595 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 498, + 607, + 633, + 625 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Datasets. We conduct extensive experiments on two large-scale outdoor LiDAR semantic segmentation datasets, i.e., SemanticKITTI [3] and nuScenes [5, 24], and one indoor dataset ScanNet [20]. The nuScenes dataset contains 700 scenes for training, 150 scenes for validation, and 150 scenes for testing, where 16 classes are utilized for LiDAR semantic segmentation. As for SemanticKITTI, it contains 19 classes for training and evaluation. It has 22 sequences, where sequences 00 to 10, 08, and 11 to 21 are used for training, validation, and testing, respectively. ScanNet [20] contains 1603 scans with 20 classes, where 1201 scans are for training, 312 scans are for validation, and 100 scans are for testing.", + "bbox": [ + 496, + 633, + 893, + 815 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Implementation Details. We follow SLidR [51] to pretrain the network on the nuScenes [5, 24] dataset. The network is pre-trained on all keyframes from 600 scenes. Besides, the pre-trained network is fine-tuned on SemanticKITTI [3] to verify the generalization ability. We", + "bbox": [ + 496, + 824, + 895, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "7024", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/1ac25c1eae9cb23d523696019d59d4b34347819172376f9a0849a974e1d37ca8.jpg", + "table_caption": [ + "Table 1. Comparisons (mIoU) among self-supervised methods on the nuScenes [24], SemanticKITTI [3], and ScanNet [20] val sets." + ], + "table_footnote": [], + "table_body": "
InitializationnuScenesSemanticKITTIScanNet
1%100%1%100%5%100%
Random42.269.132.552.146.163.3
PPKT [44]48.070.139.153.147.564.2
SLidR [51]48.270.439.654.347.964.9
PointContrast [55]47.269.237.152.347.664.5
CLIP2Scene56.371.542.655.048.465.1
", + "bbox": [ + 81, + 123, + 467, + 229 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/73987a8a261f6284f7819b8f3ccf3f946f4dc364e65da34b53f149617fabe18d.jpg", + "table_caption": [ + "Table 2. Annotation-free 3D semantic segmentation performance (mIoU) on the nuScenes [24] and ScanNet [20] val sets." + ], + "table_footnote": [], + "table_body": "
MethodnuScenesScanNet
CLIP2Scene20.8025.08
", + "bbox": [ + 169, + 276, + 380, + 314 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "leverage the CLIP model to generate image features and text embedding. Following MaskCLIP, we modify the attention pooling layer of the CLIP's image encoder, thus extracting the dense pixel-text correspondences. We take SPVCNN [52] as the 3D network to produce the point-wise feature. The framework is developed on PyTorch, where the CLIP model is frozen during training. The training time is about 40 hours for 20 epochs on two NVIDIA Tesla A100 GPUs. The optimizer is SGD with a cosine scheduler. We set the temperature $\\lambda$ and $\\tau$ to be 1 and 0.5, respectively. The sweep number is set to be 3 empirically. Besides, We adopt MinkowskiNet14 [19] as the backbone for evaluation on the ScanNet dataset, where the number of sweeps is set to be 1 and the training epochs is 30. As for the Switchable Self-Training Strategy, we randomly switch the point supervision signal after 10 epochs. We apply several data augmentations in contrastive learning, including random rotation along the z-axis and random flip on the point cloud, random horizontal flip, and random crop-resize on the image.", + "bbox": [ + 75, + 339, + 468, + 627 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Annotation-free Semantic Segmentation", + "text_level": 1, + "bbox": [ + 76, + 635, + 419, + 651 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "After pre-training the network, we show the performance of the 3D network when it is not fine-tuned on any annotations (Table 2). As no previous method reports the 3D annotation-free segmentation performance, we compare our method with different setups (Table 3). In what follows, we describe the experimental settings and give insights into our method and the different settings.", + "bbox": [ + 75, + 659, + 468, + 763 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Settings. We conduct experiments on the nuScenes and ScanNet datasets to evaluate the annotation-free semantic segmentation performance. Following MaskCLIP [61], we place the class name into 85 hand-craft prompts and feed it into the CLIP's text encoder to produce multiple text features. We then average the text features and feed the averaged features to the classifier for point-wise prediction. Besides, to explore how to effectively transfer CLIP's knowledge to the 3D network for annotation-free segmentation, We conduct", + "bbox": [ + 75, + 763, + 470, + 901 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/35fd1e5ae2e98391e0f0be995552b7c75a0aedf161f5d566cef15e8116b787d1.jpg", + "table_caption": [ + "Table 3. Ablation study on the nuScenes [24] val set for annotation-free 3D semantic segmentation." + ], + "table_footnote": [], + "table_body": "
Ablation TargetSettingmIoU (%)
-Baseline15.1
PromptsnuScenes15.1 (+0.0)
SemanticKITTI13.9 (-1.2)
Cityscapes11.3 (-3.8)
All15.3 (+0.2)
Regularizationw/o StCR19.8 (+4.7)
w/o SCR16.8 (+1.7)
KL0.0 (-15.1)
Training Strategyw/o S318.8 (+3.7)
ST10.1 (-4.0)
Sweeps1 sweep18.7 (+3.6)
3 sweeps20.8 (+5.7)
5 sweeps20.6 (+5.5)
merged18.6 (+3.5)
Full ConfigurationCLIP2Scene20.8 (+5.7)
", + "bbox": [ + 553, + 123, + 843, + 367 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "the following experiments to highlight the effectiveness of different modules in our framework.", + "bbox": [ + 498, + 393, + 892, + 422 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Baseline. The input of the 3D semantic segmentation network is only one sweep, and we pre-train the framework via semantic consistency regularization.", + "bbox": [ + 498, + 426, + 893, + 470 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Prompts (nuScenes, SemanticKITTI, Cityscapes, All). Based on the baseline, we respectively replace the nuScenes, SemanticKITTI, Cityscapes, and all class names into the prompts to produce the text embedding.", + "bbox": [ + 498, + 473, + 893, + 532 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Regularization (w/o StCR, w/o SCR, KL). Based on the full method, we remove the Spatial-temporal Consistency Regularization (w/o StCR) and remove the Semantic Consistency Regularization (w/o SCR). Besides, we abuse both StCR and SCR and distill the image feature to the point cloud by Kullback-Leibler (KL) divergence loss.", + "bbox": [ + 498, + 534, + 893, + 625 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Training Strategies (w/o S3, ST). We abuse the Switchable Self-Training Strategy (w/o S3) in the full method. Besides, we show the performance of only training the 3D network by their own predictions after ten epochs (ST).", + "bbox": [ + 498, + 626, + 893, + 686 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Sweeps Number (1 sweep, 3 sweeps, 5 sweeps, and merged). We set the sweep number $K$ to be 1, 3, and 5, respectively. Besides, we also take three sweeps of the point cloud as the input to pre-train the network (merged).", + "bbox": [ + 498, + 688, + 892, + 748 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Effect of Different Prompts. To verify how text embedding affects the performance, we generate various text embedding by the class name from different datasets (nuScenes, SemanticKITTI, and Cityscapes) and all classes for pre-training the framework. As shown in Table 3, we find that even learning with other datasets' text embedding (SemanticKITTI and Cityscapes), the 3D network could still recognize the nuScenes's objects with decent performance (13.9% and 11.3% mIoU, respectively). The result shows that the 3D network is capable of open-vocabulary recognition ability.", + "bbox": [ + 496, + 750, + 893, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "7025", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/89ffe74d74d82a6c97723e01d1374336ed7a758c47b237279664fa3a765805fd.jpg", + "image_caption": [ + "Figure 5. Qualitative results of annotation-free semantic segmentation on nuScenes dataset. Note that we show the results of individual classes. From the left to the right column are bus, motorcycle, car, and truck, respectively. The first row [ground truth] is the annotated semantic label. The second row [ours*] is our prediction of the highlighted target. The third row [ours] is our prediction of full classes." + ], + "image_footnote": [], + "bbox": [ + 89, + 97, + 883, + 375 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Effect of Semantic and Spatial-temporal Consistency Regularization. We remove Spatial-temporal Consistency Regularization (w/o SCR) from our method. Experiments show that the performance is dramatically decreased, indicating the effectiveness of our design. Besides, we also distill the image feature to the point cloud by KL divergence loss, where the text embeddings calculate the logits. However, such a method fails to transfer the semantic information from the image. The main reason is the noise-assigned semantics of the image pixel and the imperfect pixel-point correspondence due to the calibration error.", + "bbox": [ + 75, + 443, + 472, + 609 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Effect of Switchable Self-training Strategy. To examine the effect of the Switchable Self-Training Strategy, we either train the network with image supervision (w/o S3) or train the 3D network by their own predictions. Both trials witness a performance drop, indicating Switchable Self-Training Strategy is efficient in cross-modal self-supervised learning.", + "bbox": [ + 75, + 611, + 470, + 700 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Effect of Sweep Numbers. Intuitively, the performance of our method benefits from more sweeps information. Therefore, we also show the performance when restricting sweep size to 1, 3, and 5, respectively. However, we observe that the performance of 5 sweeps is similar to 3 sweeps but is more computationally expensive. Thus, we empirically set the sweep number to be 3.", + "bbox": [ + 75, + 702, + 470, + 808 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Qualitative Evaluation. The qualitative evaluations of individual classes (bus, motorcycle, car, and truck) are in Fig. 5, indicating that our method is able to perceive the objects even without training on any annotated data. However, we also observe the false positive predictions around the ground truth objects. We will resolve this issue in future work.", + "bbox": [ + 75, + 809, + 470, + 902 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. Annotation-efficient Semantic Segmentation", + "text_level": 1, + "bbox": [ + 498, + 443, + 874, + 459 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The pre-trained 3D network also boosts the performance when few labeled data are available for training. We directly compare SLidR [51], the only published method for image-to-Lidar self-supervised representation distillation. Besides, we also compared PPKT [44] and PointContrast [55]. In the following, we introduce SLidR and PPKT and compare them in detail.", + "bbox": [ + 496, + 469, + 893, + 574 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "PPKT. PPKT is a cross-modal self-supervised method for the RGB-D dataset. It performs 2D-to-3D knowledge distillation via pixel-to-point contrastive loss. For a fair comparison, we use the same 3D network and training protocol but replace our semantic and Spatio-Temporal Regularization with InfoNCE loss. The framework is trained 50 epochs on 4096 randomly selected image-to-point pairs.", + "bbox": [ + 496, + 577, + 893, + 684 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "SLidR. SLidR is an image-to-Lidar self-supervised method for autonomous driving data. Compared with PPKT, it introduces image super-pixel into cross-modal self-supervised learning. For a fair comparison, we replace our loss function with their superpixel-driven contrastive loss.", + "bbox": [ + 496, + 686, + 893, + 762 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Performance. As shown in Table 1, our method significantly outperforms the state-of-the-art methods when fine-tuned on $1\\%$ and $100\\%$ nuScenes dataset, with the improvement of $8.1\\%$ and $1.1\\%$ , respectively. Compared with the random initialization, the improvement is $14.1\\%$ and $2.4\\%$ , respectively, indicating the efficiency of our Semantic-driven Cross-modal Contrastive Learning framework. The qualitative results are shown in Fig. 6. Besides, we also verify the cross-domain generalization ability of our method. When pre-training the", + "bbox": [ + 496, + 763, + 895, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7026", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/97ac7c1a4bf31eb9d085e0bde550b8b14bd6ecef4676192d80d18bcd4042f676.jpg", + "image_caption": [ + "Figure 6. Qualitative results of fine-tuning on $1\\%$ nuScenes dataset. From the first row to the last row are the input LiDAR scan, ground truth, prediction of SLidR, and our prediction, respectively. Note that we show the results by error map, where the red point indicates the wrong prediction. Apparently, our method achieves decent performance." + ], + "image_footnote": [], + "bbox": [ + 86, + 97, + 890, + 398 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "3D network on the nuScenes dataset and fine-tuning on $1\\%$ and $100\\%$ SemanticKITTI dataset, our method significantly outperforms other state-of-the-art self-supervised methods.", + "bbox": [ + 75, + 474, + 470, + 518 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Discussions. PPKT and SLidR reveal that contrastive loss is promising for transferring knowledge from image to point cloud. Like self-supervised learning, constructing the positive and negative samples is vital to unsupervised cross-modal knowledge distillation. However, previous methods suffer from optimization-conflict issues, i.e., some negative paired samples are actually positive pairs. For example, the road occupies a large proportion of the point cloud in a scene and is supposed to have the same semantics in the semantic segmentation task. When randomly selecting training samples, most negatively defined road-road points are actually positive. When feedforwarding such samples into contrastive learning, the contrastive loss will push them away in the embedding space, leading to unsatisfactory representation learning and hammering the downstream tasks' performance. SLidR introduces superpixel-driven contrastive learning to alleviate such issues. The motivation is that the visual representation of the image pixel and the projected points are consistent intra-superpixel. Although avoiding selecting negative image-point pairs from the same superpixel, the conflict still exists inter-superpixel. In our CLIP2Scene, we introduce the free-available dense pixel-text correspondence to alleviate the optimization conflicts. The text embedding represents the semantic information and can be used to select more reasonable training samples for contrastive learning.", + "bbox": [ + 75, + 523, + 473, + 902 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Besides training sample selection, the previous method also ignores the temporal coherence of the multi-sweep point cloud. That is, for LiDAR points mapping to the same image pixel, their feature is restricted to be consistent. Besides, considering the calibration error between the LiDAR scan and the camera image. We relax the pixel-to-point mapping to image grid-to-point grid mapping for consistency regularization. To this end, our Spatial-temporal consistency regularization leads to a more rational point representation.", + "bbox": [ + 496, + 474, + 893, + 609 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Last but not least, we find that randomly switching the supervision signal benefits self-supervised learning. Essentially, different modality networks learn different feature representations. They can filter different types of errors introduced by noisy pseudo labels. By this switchable operation, the error flows can be reduced mutually.", + "bbox": [ + 496, + 612, + 895, + 703 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 722, + 619, + 738 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We explored how CLIP knowledge benefits 3D scene understanding in this work, termed CLIP2Scene. To efficiently transfer CLIP's image and text features to a 3D network, we propose a novel Semantic-driven Cross-modal Contrastive Learning framework including Semantic Regularization and Spatial-Temporal Regularization. For the first time, our pre-trained 3D network achieves annotation-free 3D semantic segmentation with decent performance. Besides, our method significantly outperforms state-of-the-art self-supervised methods when fine-tuning with labelled data.", + "bbox": [ + 496, + 750, + 895, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "7027", + "bbox": [ + 482, + 944, + 514, + 957 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Zeynep Akata, Florent Perronnin, Zaid Harchaoui, and Cordelia Schmid. Label-embedding for attribute-based classification. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 819–826, 2013. 2", + "[2] Zeynep Akata, Scott E. Reed, Daniel Walter, Honglak Lee, and Bernt Schiele. Evaluation of output embeddings for fine-grained image classification. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2927-2936, 2015. 2", + "[3] Jens Behley, Martin Garbade, Andres Milioto, Jan Quenzel, Sven Behnke, Cyril Stachniss, and Jurgen Gall. Semantickitti: A dataset for semantic scene understanding of lidar sequences. In IEEE/CVF International Conference on Computer Vision, pages 9297-9307, 2019. 2, 5, 6", + "[4] Max Bucher, Stéphane Herbin, and Frédéric Jurie. Generating visual representations for zero-shot classification. In IEEE/CVF International Conference on Computer Vision Workshop, pages 2666-2673, 2017. 2", + "[5] Holger Caesar, Varun Bankiti, Alex H Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giacarlo Baldan, and Oscar Beijbom. nuscenes: A multimodal dataset for autonomous driving. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11621-11631, 2020. 5", + "[6] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In European Conference on Computer Vision, pages 213-229, 2020. 3", + "[7] Mathilde Caron, Ishan Misra, Julien Mairal, Priya Goyal, Piotr Bojanowski, and Armand Joulin. Unsupervised learning of visual features by contrasting cluster assignments. In Advances in Neural Information Processing Systems, volume 33, pages 9912-9924, 2020. 2", + "[8] Soravit Changpinyo, Wei-Lun Chao, Boqing Gong, and Fei Sha. Synthesized classifiers for zero-shot learning. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5327-5336, 2016. 2", + "[9] Nenglun Chen, Lingjie Liu, Zhiming Cui, Runnan Chen, Duygu Ceylan, Changhe Tu, and Wenping Wang. Unsupervised learning of intrinsic structural representation points. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9121-9130, 2020. 2", + "[10] Runnan Chen, Zhu Xinge, Nenglun Chen, Dawei Wang, Wei Li, Yuexin Ma, Ruigang Yang, and Wenping Wang. Referring self-supervised learning on 3d point cloud. 2021. 2", + "[11] Runnan Chen, Xinge Zhu, Nenglun Chen, Wei Li, Yuexin Ma, Ruigang Yang, and Wenping Wang. Zero-shot point cloud segmentation by transferring geometric primitives. arXiv preprint arXiv:2210.09923, 2022. 1, 2", + "[12] Runnan Chen, Xinge Zhu, Nenglun Chen, Dawei Wang, Wei Li, Yuexin Ma, Ruigang Yang, and Wenping Wang. Towards 3d scene understanding by referring synthetic models. arXiv preprint arXiv:2203.10546, 2022. 2", + "[13] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning" + ], + "bbox": [ + 78, + 114, + 470, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "of visual representations. In International Conference on Machine Learning, pages 1597-1607, 2020. 2", + "[14] Xinlei Chen and Kaiming He. Exploring simple siamese representation learning. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15750-15758, 2021. 2", + "[15] Ran Cheng, Ryan Razani, Ehsan Taghavi, Enxu Li, and Bingbing Liu. (af)2-s3net: Attentive feature fusion with adaptive feature selection for sparse semantic segmentation network. In IEEE Conference on Computer Vision and Pattern Recognition, pages 12547-12556, 2021. 1", + "[16] Ali Cheraghian, Shafin Rahman, Dylan Campbell, and Lars Petersson. Mitigating the Hubness Problem for Zero-Shot Learning of 3D Objects. arXiv preprint arXiv:1907.06371, 2019.2", + "[17] Ali Cheraghian, Shafin Rahman, Townim F Chowdhury, Dylan Campbell, and Lars Petersson. Zero-shot learning on 3d point cloud objects and beyond. International Journal of Computer Vision, 130(10):2364-2384, 2022. 2", + "[18] Ali Cheraghian, Shafin Rahman, and Lars Petersson. Zero-Shot Learning of 3D Point Cloud Objects. In IEEE International Conference on Machine Vision Applications, pages 1-6, 2019. 2", + "[19] Christopher Choy, JunYoung Gwak, and Silvio Savarese. 4d spatio-temporal convnets: Minkowski convolutional neural networks. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3075-3084, 2019. 6", + "[20] Angela Dai, Angel Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5828-5839, 2017. 2, 5, 6", + "[21] Berkan Demirel, Ramazan Gokberk Cinbis, and Nazli Ikizler-Cinbis. Attributes2classname: A discriminative model for attribute-based unsupervised zero-shot learning. In IEEE/CVF International Conference on Computer Vision, pages 1241-1250, 2017. 2", + "[22] Carl Doersch, Abhinav Gupta, and Alexei A Efros. Unsupervised visual representation learning by context prediction. In IEEE/CVF International Conference on Computer Vision, pages 1422-1430, 2015. 2", + "[23] Alexey Dosovitskiy, Jost Tobias Springenberg, Martin Ried-miller, and Thomas Brox. Discriminative unsupervised feature learning with convolutional neural networks. In Advances in Neural Information Processing Systems, volume 27, 2014. 2", + "[24] Whye Kit Fong, Rohit Mohan, Juana Valeria Hurtado, Lubing Zhou, Holger Caesar, Oscar Beijbom, and Abhinav Valada. Panoptic nuscenes: A large-scale benchmark for lidar panoptic segmentation and tracking. IEEE Robotics and Automation Letters, 7:3795-3802, 2022. 2, 5, 6", + "[25] Chuang Gan, Ming Lin, Yi Yang, Yueting Zhuang, and Alexander Hauptmann. Exploring semantic inter-class relationships (sir) for zero-shot action recognition. In AAAI Conference on Artificial Intelligence, 2015. 2", + "[26] Biao Gao, Yancheng Pan, Chengkun Li, Sibo Geng, and Huijing Zhao. Are we hungry for 3d lidar data for semantic segmentation? a survey of datasets and methods. IEEE Transactions on Intelligent Transportation Systems, 2021. 1" + ], + "bbox": [ + 503, + 92, + 893, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "7028", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[27] Jean-Bastien Grill, Florian Strub, Florent Altché, Corentin Tallec, Pierre Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Avila Pires, Zhaohan Guo, Mohammad Gheshlaghi Azar, et al. Bootstrap your own latent-a new approach to self-supervised learning. In Advances in Neural Information Processing Systems, volume 33, pages 21271-21284, 2020. 2", + "[28] Yulan Guo, Hanyun Wang, Qingyong Hu, Hao Liu, Li Liu, and Mohammed Bennamoun. Deep learning for 3d point clouds: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence, 43(12):4338-4364, 2020. 1", + "[29] Bo Han, Quanming Yao, Xingrui Yu, Gang Niu, Miao Xu, Weihua Hu, Ivor Tsang, and Masashi Sugiyama. Co-teaching: Robust training of deep neural networks with extremely noisy labels. Advances in neural information processing systems, 31, 2018. 5", + "[30] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9729-9738, 2020. 2", + "[31] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 770-778, 2016. 3", + "[32] Fangzhou Hong, Lingdong Kong, Hui Zhou, Xinge Zhu, Hongsheng Li, and Ziwei Liu. Unified 3d and 4d panoptic segmentation via dynamic shifting network. arXiv preprint arXiv:2203.07186, 2022. 1", + "[33] Yuenan Hou, Xinge Zhu, Yuexin Ma, Chen Change Loy, and Yikang Li. Point-to-Voxel Knowledge Distillation for LiDAR Semantic Segmentation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8479-8488, 2022. 1", + "[34] Siyuan Huang, Yichen Xie, Song-Chun Zhu, and Yixin Zhu. Spatio-temporal self-supervised representation learning for 3d point clouds. In IEEE/CVF International Conference on Computer Vision, pages 6535-6545, 2021. 2", + "[35] Ajay Jain, Matthew Tancik, and Pieter Abbeel. Putting nerf on a diet: Semantically consistent few-shot view synthesis. In IEEE/CVF International Conference on Computer Vision, pages 5885-5894, 2021. 2", + "[36] Sosuke Kobayashi, Eiichi Matsumoto, and Vincent Sitzmann. Decomposing nef for editing via feature field distillation. arXiv preprint arXiv:2205.15585, 2022. 2", + "[37] Elyor Kodirov, Tao Xiang, and Shaogang Gong. Semantic autoencoder for zero-shot learning. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4447-4456, 2017. 2", + "[38] Lingdong Kong, Youquan Liu, Runnan Chen, Yuexin Ma, Xinge Zhu, Yikang Li, Yuenan Hou, Yu Qiao, and Ziwei Liu. Rethinking range view representation for lidar segmentation. arXiv preprint arXiv:2303.05367, 2023. 1", + "[39] Lingdong Kong, Niamul Quader, and Venice Erin Liong. Conda: Unsupervised domain adaptation for lidar segmentation via regularized domain concatenation. In IEEE International Conference on Robotics and Automation, 2023. 1", + "[40] Lingdong Kong, Jiawei Ren, Liang Pan, and Ziwei Liu. Lasermix for semi-supervised lidar semantic segmentation." + ], + "bbox": [ + 78, + 90, + 470, + 901 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023. 1", + "[41] Christoph H. Lampert, Hannes Nickisch, and Stefan Harmeling. Learning to detect unseen object classes by between-class attribute transfer. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 951-958, 2009. 2", + "[42] Christoph H. Lampert, Hannes Nickisch, and Stefan Harmeling. Attribute-based classification for zero-shot visual object categorization. IEEE Transactions on Pattern Analysis and Machine Intelligence, 36:453-465, 2014. 2", + "[43] Yan Li, Zhen Jia, Junge Zhang, Kaiqi Huang, and Tieniu Tan. Deep semantic structural constraints for zero-shot learning. In AAAI Conference on Artificial Intelligence, 2018. 2", + "[44] Yueh-Cheng Liu, Yu-Kai Huang, Hung-Yueh Chiang, Hung-Ting Su, Zhe-Yu Liu, Chin-Tang Chen, Ching-Yu Tseng, and Winston H Hsu. Learning from 2d: Contrastive pixel-topoint knowledge transfer for 3d pretraining. arXiv preprint arXiv:2104.04687, 2021. 1, 2, 3, 4, 6, 7", + "[45] Björn Michele, Alexandre Boulch, Gilles Puy, Maxime Bucher, and Renaud Marlet. Generative zero-shot learning for semantic segmentation of 3d point clouds. In International Conference on 3D Vision, pages 992-1002, 2021. 1, 2", + "[46] Ashish Mishra, M. Shiva Krishna Reddy, Anurag Mittal, and Hema A. Murthy. A generative model for zero shot learning using conditional variational autoencoders. In IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshop, pages 2269-22698, 2018. 2", + "[47] Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. Pointnet: Deep learning on point sets for 3d classification and segmentation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 652-660, 2017. 1, 2", + "[48] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 1", + "[49] Yongming Rao, Wenliang Zhao, Guangyi Chen, Yansong Tang, Zheng Zhu, Guan Huang, Jie Zhou, and Jiwen Lu. Denseclip: Language-guided dense prediction with context-aware prompting. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18082-18091, 2022. 2", + "[50] David Rozenberszki, Or Litany, and Angela Dai. Language-grounded indoor 3d semantic segmentation in the wild. In European Conference on Computer Vision, pages 125-141, 2022. 2", + "[51] Corentin Sautier, Gilles Puy, Spyros Gidaris, Alexandre Boulch, Andrei Bursuc, and Renaud Marlet. Image-to-lidar self-supervised distillation for autonomous driving data. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9891-9901, 2022. 1, 2, 3, 4, 5, 6, 7", + "[52] Haotian Tang, Zhijian Liu, Shengyu Zhao, Yujun Lin, Ji Lin, Hanrui Wang, and Song Han. Searching efficient 3d architectures with sparse point-voxel convolution. In European Conference on Computer Vision, pages 685–702, 2020. 6", + "[53] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia" + ], + "bbox": [ + 501, + 92, + 893, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "7029", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Polosukhin. Attention is all you need. In Advances in Neural Information Processing Systems, volume 30, 2017. 3", + "[54] Yongqin Xian, Zeynep Akata, Gaurav Sharma, Quynh N. Nguyen, Matthias Hein, and Bernt Schiele. Latent embeddings for zero-shot classification. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 69-77, 2016. 2", + "[55] Saining Xie, Jiatao Gu, Demi Guo, Charles R Qi, Leonidas Guibas, and Or Litany. Pointcontrast: Unsupervised pretraining for 3d point cloud understanding. In European Conference on Computer Vision, pages 574-591, 2020. 2, 6, 7", + "[56] Jianyun Xu, Ruixiang Zhang, Jian Dou, Yushi Zhu, Jie Sun, and Shiliang Pu. Rpvnet: A deep and efficient range-point-voxel fusion network for lidar point cloud segmentation. In IEEE/CVF International Conference on Computer Vision, pages 16024-16033, 2021. 1", + "[57] Lewei Yao, Jianhua Han, Youpeng Wen, Xiaodan Liang, Dan Xu, Wei Zhang, Zhenguo Li, Chunjing Xu, and Hang Xu. Detclip: Dictionary-enriched visual-concept paralleled pre-training for open-world detection. arXiv preprint arXiv:2209.09407, 2022. 2, 3", + "[58] Éloi Zablocki, Patrick Bordes, Benjamin Piwowarski, Laure Soulier, and Patrick Gallinari. Context-aware zero-shot learning for object recognition. In International Conference on Machine Learning, pages 7292-7303, 2019. 2", + "[59] Renrui Zhang, Ziyu Guo, Wei Zhang, Kunchang Li, Xupeng Miao, Bin Cui, Yu Qiao, Peng Gao, and Hongsheng Li. Pointclip: Point cloud understanding by clip. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8552-8562, 2022. 2", + "[60] Zaiwei Zhang, Rohit Girdhar, Armand Joulin, and Ishan Misra. Self-supervised pretraining of 3d features on any point-cloud. In IEEE/CVF International Conference on Computer Vision, pages 10252-10263, 2021. 2", + "[61] Chong Zhou, Chen Change Loy, and Bo Dai. Extract free dense labels from clip. In European Conference on Computer Vision, pages 696-712, 2022. 2, 4, 6", + "[62] Xinge Zhu, Hui Zhou, Tai Wang, Fangzhou Hong, Yuexin Ma, Wei Li, Hongsheng Li, and Dahua Lin. Cylindrical and asymmetrical 3d convolution networks for lidar segmentation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9939-9948, 2021. 1" + ], + "bbox": [ + 78, + 90, + 470, + 698 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "7030", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2023/CLIP2Scene_ Towards Label-Efficient 3D Scene Understanding by CLIP/58dade4c-9eec-4c3a-968d-a582ab672d68_model.json b/2023/CLIP2Scene_ Towards Label-Efficient 3D Scene Understanding by CLIP/58dade4c-9eec-4c3a-968d-a582ab672d68_model.json new file mode 100644 index 0000000000000000000000000000000000000000..66212b9c965701afee68348c7b535a866c85e7d1 --- /dev/null +++ b/2023/CLIP2Scene_ Towards Label-Efficient 3D Scene Understanding by CLIP/58dade4c-9eec-4c3a-968d-a582ab672d68_model.json @@ -0,0 +1,2070 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.121, + 0.131, + 0.85, + 0.154 + ], + "angle": 0, + "content": "CLIP2Scene: Towards Label-efficient 3D Scene Understanding by CLIP" + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.179, + 0.88, + 0.218 + ], + "angle": 0, + "content": "Runnan Chen\\(^{1,2}\\) Youquan Liu\\(^{2,3}\\) Lingdong Kong\\(^{2,4}\\) Xinge Zhu\\(^{5}\\) Yuexin Ma\\(^{6}\\) Yikang Li\\(^{2}\\) Yuenan Hou\\(^{2,\\dagger}\\) Yu Qiao\\(^{2}\\) Wenping Wang\\(^{7,\\dagger}\\)" + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.22, + 0.862, + 0.254 + ], + "angle": 0, + "content": "1The University of Hong Kong 2Shanghai AI Laboratory 3Hochschule Bremerhaven 4National University of Singapore 5The Chinese University of Hong Kong 6ShanghaiTech University 7Texas A&M University" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.304, + 0.314, + 0.319 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.331, + 0.474, + 0.786 + ], + "angle": 0, + "content": "Contrastive Language-Image Pre-training (CLIP) achieves promising results in 2D zero-shot and few-shot learning. Despite the impressive performance in 2D, applying CLIP to help the learning in 3D scene understanding has yet to be explored. In this paper, we make the first attempt to investigate how CLIP knowledge benefits 3D scene understanding. We propose CLIP2Scene, a simple yet effective framework that transfers CLIP knowledge from 2D image-text pre-trained models to a 3D point cloud network. We show that the pre-trained 3D network yields impressive performance on various downstream tasks, i.e., annotation-free and fine-tuning with labelled data for semantic segmentation. Specifically, built upon CLIP, we design a Semantic-driven Cross-modal Contrastive Learning framework that pre-trains a 3D network via semantic and spatial-temporal consistency regularization. For the former, we first leverage CLIP's text semantics to select the positive and negative point samples and then employ the contrastive loss to train the 3D network. In terms of the latter, we force the consistency between the temporally coherent point cloud features and their corresponding image features. We conduct experiments on SemanticKITTI, nuScenes, and ScanNet. For the first time, our pre-trained network achieves annotation-free 3D semantic segmentation with \\(20.8\\%\\) and \\(25.08\\%\\) mIoU on nuScenes and ScanNet, respectively. When fine-tuned with \\(1\\%\\) or \\(100\\%\\) labelled data, our method significantly outperforms other self-supervised methods, with improvements of \\(8\\%\\) and \\(1\\%\\) mIoU, respectively. Furthermore, we demonstrate the generalizability for handling cross-domain datasets. Code is publicly available1." + }, + { + "type": "title", + "bbox": [ + 0.079, + 0.81, + 0.21, + 0.826 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.835, + 0.471, + 0.867 + ], + "angle": 0, + "content": "3D scene understanding is fundamental in autonomous driving, robot navigation, etc [26,28]. Current deep learning-" + }, + { + "type": "image_caption", + "bbox": [ + 0.528, + 0.295, + 0.876, + 0.31 + ], + "angle": 0, + "content": "Semantic-driven Cross-modal Contrastive Learning" + }, + { + "type": "image", + "bbox": [ + 0.508, + 0.313, + 0.878, + 0.536 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.544, + 0.896, + 0.656 + ], + "angle": 0, + "content": "Figure 1. We explore how CLIP knowledge benefits 3D scene understanding. To this end, we propose CLIP2Scene, a Semantic-driven Cross-modal Contrastive Learning framework that leverages CLIP knowledge to pre-train a 3D point cloud segmentation network via semantic and spatial-temporal consistency regularization. CLIP2Scene yields impressive performance on annotation-free 3D semantic segmentation and significantly outperforms other self-supervised methods when fine-tuning on annotated data." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.685, + 0.895, + 0.837 + ], + "angle": 0, + "content": "based methods have shown inspirational performance on 3D point cloud data [15, 32, 33, 38, 47, 56, 62]. However, some drawbacks hinder their real-world applications. The first one comes from their heavy reliance on the large collection of annotated point clouds, especially when high-quality 3D annotations are expensive to acquire [39, 40, 44, 51]. Besides, they typically fail to recognize novel objects that are never seen in the training data [11, 45]. As a result, it may need extra annotation efforts to train the model on recognizing these novel objects, which is both tedious and time-consuming." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Contrastive Vision-Language Pre-training (CLIP) [48] provides a new perspective that mitigates the above issues in 2D vision. It was trained on large-scale free-available image-text pairs from websites and built vision-language" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.875, + 0.424, + 0.901 + ], + "angle": 0, + "content": "Symbol \\(\\dagger\\) denotes the corresponding authors. \n1https://github.com/runnanchen/CLIP2Scene." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7020" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.304 + ], + "angle": 0, + "content": "correlation to achieve promising open-vocabulary recognition. MaskCLIP [61] further explores semantic segmentation based on CLIP. With minimal modifications to the CLIP pre-trained network, MaskCLIP can be directly used for the semantic segmentation of novel objects without additional training efforts. PointCLIP [59] reveals that the zero-shot classification ability of CLIP can be generalized from the 2D image to the 3D point cloud. It perspectively projects a point cloud frame into different views of 2D depth maps that bridge the modal gap between the image and the point cloud. The above studies indicate the potential of CLIP on enhancing the 2D segmentation and 3D classification performance. However, whether and how CLIP knowledge benefits 3D scene understanding is still under-explored." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.307, + 0.473, + 0.64 + ], + "angle": 0, + "content": "In this paper, we explore how to leverage CLIP's 2D image-text pre-learned knowledge for 3D scene understanding. Previous cross-modal knowledge distillation methods [44,51] suffer from the optimization-conflict issue, i.e., some of the positive pairs are regarded as negative samples for contrastive learning, leading to unsatisfactory representation learning and hammering the performance of downstream tasks. Besides, they also ignore the temporal coherence of the multi-sweep point cloud, failing to utilize the rich inter-sweep correspondence. To handle the mentioned problems, we propose a novel Semantic-driven Cross-modal Contrastive Learning framework that fully leverages CLIP's semantic and visual information to regularize a 3D network. Specifically, we propose Semantic Consistency Regularization and Spatial-Temporal Consistency Regularization. In semantic consistency regularization, we utilize CLIP's text semantics to select the positive and negative point samples for less-conflict contrastive learning. For spatial-temporal consistency regularization, we take CLIP's image pixel feature to impose a soft consistency constraint on the temporally coherent point features. Such an operation also alleviates the effects of imperfect image-to-point calibration." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.644, + 0.473, + 0.902 + ], + "angle": 0, + "content": "We conduct several downstream tasks on the indoor and outdoor datasets to verify how the pre-trained network benefits the 3D scene understanding. The first one is annotation-free semantic segmentation. Following MaskCLIP, we place class names into multiple hand-crafted templates as prompts and average the text embeddings generated by CLIP to conduct the annotation-free segmentation. For the first time, our method achieves \\(20.8\\%\\) and \\(25.08\\%\\) mIoU annotation-free 3D semantic segmentation on the nuScenes [24] and ScanNet [20] datasets without training on any labelled data. Secondly, we compare with other self-supervised methods in label-efficient learning. When fine-tuning the 3D network with \\(1\\%\\) or \\(100\\%\\) labelled data on the nuScenes dataset, our method significantly outperforms state-of-the-art self-supervised methods, with improvements of \\(8\\%\\) and \\(1\\%\\) mIoU, respectively. Besides, to verify the generalization capability, we pre-train the network on the nuScenes dataset" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.138 + ], + "angle": 0, + "content": "and evaluate it on SemanticKITTI [3]. Our method still significantly outperforms state-of-the-art methods. The key contributions of our work are summarized as follows." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.147, + 0.892, + 0.178 + ], + "angle": 0, + "content": "- The first work that distils CLIP knowledge to a 3D network for 3D scene understanding." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.189, + 0.895, + 0.248 + ], + "angle": 0, + "content": "- We propose a novel Semantic-driven Cross-modal Contrastive Learning framework that pre-trains a 3D network via spatial-temporal and semantic consistency regularization." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.261, + 0.892, + 0.321 + ], + "angle": 0, + "content": "- We propose a novel Semantic-guided Spatial-Temporal Consistency Regularization that forces the consistency between the temporally coherent point cloud features and their corresponding image features." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.332, + 0.894, + 0.393 + ], + "angle": 0, + "content": "- For the first time, our method achieves promising results on annotation-free 3D scene segmentation. When fine-tuning with labelled data, our method significantly outperforms state-of-the-art self-supervised methods." + }, + { + "type": "list", + "bbox": [ + 0.518, + 0.147, + 0.895, + 0.393 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.407, + 0.642, + 0.423 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.433, + 0.895, + 0.78 + ], + "angle": 0, + "content": "Zero-shot Learning in 3D. The objective of zero-shot learning (ZSL) is to recognize objects that are unseen in the training set. Many efforts have been devoted to the 2D recognition tasks [1, 2, 4, 8, 21, 25, 34, 37, 41–43, 46, 54, 55, 58, 60], and few works concentrate on performing ZSL in the 3D domain [11, 16–18, 45]. [18] applies ZSL to 3D tasks, where they train PointNet [47] on \"seen\" samples and test on \"unseen\" samples. Subsequent work [16] addresses the hubness problem caused by the low-quality point cloud features. [17] proposes the triplet loss to boost the performance under the transductive setting, where the \"unseen\" class is observed and unlabeled in the training phase. [11] makes the first attempt to explore the transductive zero-shot segmentation for 3D scene understanding. Recently, some studies introduced CLIP into zero-shot learning. MaskCLIP [61] investigates the problem of utilizing CLIP to help the 2D dense prediction tasks and exhibits encouraging zero-shot semantic segmentation performance. PointCLIP [59] is the pioneering work that applies CLIP to 3D recognition and shows impressive performance on zero-shot and few-shot classification tasks. Our work takes a step further to investigate how the rich semantic and visual knowledge in CLIP can benefit the 3D semantic segmentation tasks." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.781, + 0.897, + 0.903 + ], + "angle": 0, + "content": "Self-supervised Representation Learning. The purpose of self-supervised learning is to obtain a good representation that benefits the downstream tasks. The dominant approaches resort to contrastive learning to pre-train the network [7, 9, 9, 10, 12-14, 22, 23, 27, 30]. Recently, inspired by the success of CLIP, leveraging the pre-trained model of CLIP to the downstream tasks has raised the community's attention [35, 36, 49, 50, 57]. DenseCLIP [49] utilizes" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "7021" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.091, + 0.093, + 0.878, + 0.378 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.383, + 0.896, + 0.485 + ], + "angle": 0, + "content": "Figure 2. Illustration of the Semantic-driven Cross-modal Contrastive Learning. Firstly, we obtain the text embedding \\( t_i \\), image pixel feature \\( x_i \\), and point feature \\( p_i \\) by text encoder, image encoder, and point encoder, respectively. Secondly, we leverage CLIP knowledge to construct positive and negative samples for contrastive learning. Thus we obtain point-text pairs \\( \\{x_i, t_i\\}_{i=1}^M \\) and all pixel-point-text pairs in a short temporal \\( \\{\\hat{x}_i^k, \\hat{p}_i^k, t_i^k\\}_{i=1,k=1}^M \\). Here, \\( \\{x_i, t_i\\}_{i=1}^M \\) and \\( \\{\\hat{x}_i^k, \\hat{p}_i^k, t_i^k\\}_{i=1,k=1}^M \\) are used for Semantic Consistency Regularization and Spatial-Temporal Consistency Regularization, respectively. Lastly, we perform Semantic Consistency Regularization by pulling the point features to their corresponding text embedding and Spatial-Temporal Consistency Regularization by mimicking the temporally coherent point features to their corresponding pixel features." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.51, + 0.471, + 0.585 + ], + "angle": 0, + "content": "the CLIP's pre-trained knowledge for dense image pixel prediction. DetCLIP [57] proposes a pre-training method equipped with CLIP for open-world detection. We leverage the image-text pre-trained CLIP knowledge to help 3D scene understanding." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.587, + 0.471, + 0.722 + ], + "angle": 0, + "content": "Cross-modal Knowledge Distillation. Recently, increasing studies have focused on transferring knowledge from 2D images to 3D point clouds for self-supervised representation learning [44, 51]. PPKT [44] resorts to the InfoNCE loss to help the 3D network distil rich knowledge from the 2D image backbone. SLidR [51] further introduce the superpixel to boost the cross-modal knowledge distillation. In this paper, we first attempt to pre-train a 3D network with CLIP's knowledge." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.739, + 0.212, + 0.757 + ], + "angle": 0, + "content": "3. Methodology" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.472, + 0.903 + ], + "angle": 0, + "content": "Considering the impressive open-vocabulary performance achieved by CLIP in image classification and segmentation, natural curiosities have been raised. Can CLIP endow the ability to a 3D network for annotation-free scene understanding? And further, will it promote the network performance when fine-tuned on labelled data? To answer the above questions, we study the cross-modal knowledge transfer of CLIP for 3D scene understanding, termed CLIP2Scene. Our work is a pioneer in exploiting CLIP knowledge for 3D scene" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.51, + 0.895, + 0.603 + ], + "angle": 0, + "content": "understanding. In what follows, we revisit the CLIP applied in 2D open-vocabulary classification and semantic segmentation, then present our CLIP2Scene in detail. Our approach consists of three major components: Semantic Consistency Regularization, Semantic-Guided Spatial-Temporal Consistency Regularization, and Switchable Self-Training Strategy." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.616, + 0.663, + 0.634 + ], + "angle": 0, + "content": "3.1. Revisiting CLIP" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.642, + 0.895, + 0.869 + ], + "angle": 0, + "content": "Contrastive Vision-Language Pre-training (CLIP) mitigates the following drawbacks that dominate the computer vision field: 1. Deep models need a large amount of formatted and labelled training data, which is expensive to acquire; 2. The model's generalization ability is weak, making it difficult to migrate to a new scenario with unseen objects. CLIP consists of an image encoder (ResNet [31] or ViT [6]) and a text encoder (Transformer [53]), both respectively project the image and text representation to a joint embedding space. During training, CLIP constructs positive and negative samples from 400 million image-text pairs to train both encoders with a contrastive loss, where the large-scale image-text pairs are free-available from the Internet and assumed to contain every class of images and most concepts of text. Therefore, CLIP can achieve promising open-vocabulary recognition." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.895, + 0.901 + ], + "angle": 0, + "content": "For 2D zero-shot classification, CLIP first places the class name into a pre-defined template to generate the text embed-" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7022" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.095, + 0.475, + 0.282 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.081, + 0.283, + 0.465, + 0.432 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.44, + 0.47, + 0.482 + ], + "angle": 0, + "content": "Figure 3. Illustration of the image pixel-to-text mapping. The dense pixel-text correspondence \\(\\{x_{i},t_{i}\\}_{i = 1}^{M}\\) is extracted by the off-the-shelf method MaskCLIP [61]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.501, + 0.473, + 0.607 + ], + "angle": 0, + "content": "dings and then encodes images to obtain image embeddings. Next, it calculates the similarities between images and text embeddings to determine the class. MaskCLIP further extends CLIP into 2D semantic segmentation. Specifically, MaskCLIP modifies the attention pooling layer of the CLIP's image encoder, thus performing pixel-level mask prediction instead of the global image-level prediction." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.618, + 0.21, + 0.633 + ], + "angle": 0, + "content": "3.2. CLIP2Scene" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.643, + 0.473, + 0.793 + ], + "angle": 0, + "content": "As shown in Fig. 2, we first leverage CLIP and 3D network to respectively extract the text embeddings, image pixel feature, and point feature. Secondly, we construct positive and negative samples based on CLIP's knowledge. Lastly, we impose Semantic Consistency Regularization by pulling the point features to their corresponding text embedding. At the same time, we apply Spatial-Temporal Consistency Regularization by forcing the consistency between temporally coherent point features and their corresponding pixel features. In what follows, we present the details and insights." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.816, + 0.391, + 0.831 + ], + "angle": 0, + "content": "3.2.1 Semantic Consistency Regularization" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.472, + 0.901 + ], + "angle": 0, + "content": "As CLIP is pre-trained on 2D images and text, our first concern is the domain gap between 2D images and the 3D point cloud. To this end, we build dense pixel-point correspondence and transfer image knowledge to the 3D point cloud" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.894, + 0.227 + ], + "angle": 0, + "content": "via the pixel-point pairs. Specifically, we calibrate the Li-DAR point cloud with corresponding images captured by six cameras. Therefore, the dense pixel-point correspondence \\(\\{x_i,p_i\\}_{i = 1}^M\\) can be obtained accordingly, where \\(x_{i}\\) and \\(p_i\\) indicates \\(i\\)-th paired image feature and point feature, which are respectively extracted by the CLIP's image encoder and the 3D network. \\(M\\) is the number of pairs. Note that it is an online operation and is irrelevant to the image and point data augmentation." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.228, + 0.895, + 0.59 + ], + "angle": 0, + "content": "Previous methods [44, 51] provide a promising solution to cross-modal knowledge transfer. They first construct positive pixel-point pairs \\(\\{x_i, p_i\\}_{i=1}^M\\) and negative pairs \\(\\{x_i, p_j\\} (i \\neq j)\\), and then pull in the positive pairs while pushing away the negative pairs in the embedding space via the InfoNCE loss. Despite the encourageable performance of previous methods in transferring cross-modal knowledge, they are both confronted with the same optimization-conflict issue. For example, suppose \\(i\\)-th pixel \\(x_i\\) and \\(j\\)-th point \\(p_j\\) are in the different positions of the same instance with the same semantics. However, the InfoNCE loss will try to push them away, which is unreasonable and hammer the performance of the downstream tasks [51]. In light of this, we propose a Semantic Consistency Regularization that leverages the CLIP's semantic information to alleviate this issue. Specifically, we generate the dense pixel-text pairs \\(\\{x_i, t_i\\}_{i=1}^M\\) by following the off-the-shelf method MaskCLIP [61] (Fig. 3), where \\(t_i\\) is the text embedding generated from the CLIP's text encoder. Note that the pixel-text mappings are free-available from CLIP without any additional training. We then transfer pixel-text pairs to point-text pairs \\(\\{p_i, t_i\\}_{i=1}^M\\) and utilize the text semantics to select the positive and negative point samples for contrastive learning. The objective function is as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.516, + 0.61, + 0.892, + 0.663 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {S - i n f o} = - \\sum_ {c = 1} ^ {C} \\log \\frac {\\sum_ {t _ {i} \\in c , p _ {i}} \\exp \\left(D \\left(t _ {i} , p _ {i}\\right) / \\tau\\right)}{\\sum_ {t _ {i} \\in c , t _ {j} \\notin c , p _ {j}} \\exp \\left(D \\left(t _ {i} , p _ {j}\\right) / \\tau\\right)}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.664, + 0.892, + 0.709 + ], + "angle": 0, + "content": "where \\( t_i \\in c \\) indicates that \\( t_i \\) is generated by \\( c \\)-th classes name, and \\( C \\) is the number of classes. \\( D \\) denotes the scalar product operation and \\( \\tau \\) is a temperature term \\( (\\tau > 0) \\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.709, + 0.892, + 0.815 + ], + "angle": 0, + "content": "Since the text is composed of class names placed into pre-defined templates, the text embedding represents the semantic information of the corresponding class. Therefore, those points with the same semantics will be restricted near the same text embedding, and those with different semantics will be pushed away. To this end, our Semantic Consistency Regularization causes less conflict in contrastive learning." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.832, + 0.892, + 0.863 + ], + "angle": 0, + "content": "3.2.2 Semantic-guided Spatial-temporal Consistency Regularization" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.871, + 0.894, + 0.901 + ], + "angle": 0, + "content": "Besides semantic consistency regularization, we consider how image pixel features help to regularize a 3D network." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7023" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.076, + 0.092, + 0.478, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.231, + 0.473, + 0.332 + ], + "angle": 0, + "content": "Figure 4. Illustration of the image pixel-to-point mapping (left) and semantic-guided fusion feature generation (right). We build the grid-wise correspondence between an image \\(I\\) and the temporally coherent LiDAR point cloud \\(\\{P_k\\}_{k=1}^K\\) within \\(S\\) seconds and generate semantic-guided fusion features for individual grids. Both \\(\\{\\hat{x}_i^k, \\hat{p}_i^k\\}_{i=1,k=1}^M, K\\) and \\(\\{f_n\\}_{n=1}^N\\) are used to perform Spatial-Temporal Consistency Regularization." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.356, + 0.473, + 0.478 + ], + "angle": 0, + "content": "The natural alternative directly pulls in the point feature with its corresponding pixel in the embedding space. However, the noise-assigned semantics of the image pixel and the imperfect pixel-point mapping hinder the downstream task's performance. To this end, we propose a novel semantic-guided Spatial-Temporal Consistency Regularization to alleviate the problem by imposing a soft constraint on points within local space and time." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.481, + 0.473, + 0.673 + ], + "angle": 0, + "content": "Specifically, given an image \\(I\\) and temporally coherent LiDAR point cloud \\(\\{P_k\\}_{k=1}^K\\), where \\(K\\) is the number of sweeps within \\(S\\) seconds. Note that the image is matched to the first frame of the point cloud \\(P_1\\) with pixel-point pairs \\(\\{\\hat{x}_i^1, \\hat{p}_i^1\\}_{i=1}^M\\). We register the rest of the point cloud to the first frame via the calibration matrices and map them to the image (Fig. 4). Thus we obtain all pixel-point-text pairs in a short temporal \\(\\{\\hat{x}_i^k, \\hat{p}_i^k, t_i^k\\}_{i=1,k=1}^M\\). Next, we divide the entire stitched point cloud into regular grids \\(\\{g_n\\}_{n=1}^N\\), where the temporally coherent points are located in the same grid. We impose the spatial-temporal consistency constraint within individual grids by the following objective function:" + }, + { + "type": "equation", + "bbox": [ + 0.089, + 0.704, + 0.472, + 0.742 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {S S R} = \\sum_ {g _ {n}} \\sum_ {(\\hat {i}, \\hat {k}) \\in g _ {n}} (1 - \\operatorname {s i g m o i d} \\left(D \\left(\\hat {p} _ {\\hat {i}} ^ {\\hat {k}}, f _ {n}\\right)\\right)) / N, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.761, + 0.47, + 0.81 + ], + "angle": 0, + "content": "where \\((\\hat{i},\\hat{k})\\in g_{n}\\) indicates the pixel-point pair \\(\\{\\hat{x}_i^k,\\hat{p}_i^k\\}\\) is located in the \\(n\\) -th grid. \\(\\{f_n\\}_{n = 1}^N\\) is a semantic-guided cross-modal fusion feature formulated by:" + }, + { + "type": "equation", + "bbox": [ + 0.164, + 0.826, + 0.472, + 0.864 + ], + "angle": 0, + "content": "\\[\nf _ {n} = \\sum_ {(\\hat {i}, \\hat {k}) \\in g _ {n}} a _ {\\hat {i}} ^ {\\hat {k}} * \\hat {x} _ {\\hat {i}} ^ {\\hat {k}} + b _ {\\hat {i}} ^ {\\hat {k}} * \\hat {p} _ {\\hat {i}} ^ {\\hat {k}}, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.883, + 0.419, + 0.906 + ], + "angle": 0, + "content": "where \\(a_{\\hat{i}}^{\\hat{k}}\\) and \\(b_{\\hat{i}}^{\\hat{k}}\\) are attention weight calculated by:" + }, + { + "type": "equation", + "bbox": [ + 0.507, + 0.115, + 0.893, + 0.222 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} a _ {i} ^ {\\hat {k}} = \\frac {\\exp (D (\\hat {x} _ {\\hat {i}} ^ {\\hat {k}} , t _ {\\hat {i}} ^ {1}) / \\lambda)}{\\sum_ {(\\hat {i} , \\hat {k}) \\in g _ {n}} \\exp (D (\\hat {x} _ {\\hat {i}} ^ {\\hat {k}} , t _ {\\hat {i}} ^ {1}) / \\lambda) + \\exp (D (\\hat {p} _ {\\hat {i}} ^ {\\hat {k}} , t _ {\\hat {i}} ^ {1}) / \\lambda)}, \\\\ b _ {\\hat {i}} ^ {\\hat {k}} = \\frac {\\exp \\left(D \\left(\\hat {p} _ {\\hat {i}} ^ {\\hat {k}} , t _ {\\hat {i}} ^ {1}\\right) / \\lambda\\right)}{\\sum_ {(\\hat {i} , \\hat {k}) \\in g _ {n}} \\exp \\left(D \\left(\\hat {x} _ {\\hat {i}} ^ {\\hat {k}}, t _ {\\hat {i}} ^ {1}\\right) / \\lambda\\right) + \\exp \\left(D \\left(\\hat {p} _ {\\hat {i}} ^ {\\hat {k}}, t _ {\\hat {i}} ^ {1}\\right) / \\lambda\\right)}, \\tag {4} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.233, + 0.719, + 0.246 + ], + "angle": 0, + "content": "where \\(\\lambda\\) is the temperature term." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.247, + 0.894, + 0.324 + ], + "angle": 0, + "content": "Actually, those pixel and point features within the local grid \\( g_{n} \\) are restricted near a dynamic centre \\( f_{n} \\). Thus, such a soft constraint alleviates the noisy prediction and calibration error issues. At the same time, it imposes Spatio-Temporal Regularization on the temporally coherent point features." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.341, + 0.785, + 0.358 + ], + "angle": 0, + "content": "3.2.3 Switchable Self-training Strategy" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.365, + 0.895, + 0.596 + ], + "angle": 0, + "content": "We combine the loss function \\(\\mathcal{L}_{S,\\text{info}}\\) and \\(\\mathcal{L}_{SSR}\\) to end-to-end train the whole network, where the CLIP's image and text encoder backbone are frozen during training. We find that method worked only when the pixel-point feature \\(\\{x_i,p_i\\}_{i = 1}^M\\) and \\(\\{\\hat{x}_i^k,\\hat{p}_i^k\\}_{i = 1,k = 1}^{\\hat{M},K}\\), which are used in \\(\\mathcal{L}_{S,\\text{info}}\\) and \\(\\mathcal{L}_{SSR}\\), are generated from different learnable linear layer. On top of that, we further put forward an effective strategy to promote performance. Specifically, after contrastive learning of the 3D network for a few epochs, we randomly switch the point pseudo label between the paired image pixel's pseudo label and the point's predicted label. Since different modality networks learn different feature representations, they can filter different types of error introduced by noisy pseudo labels. By this switchable operation, the error flows can be reduced by mutually [29]." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.608, + 0.634, + 0.626 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.634, + 0.895, + 0.816 + ], + "angle": 0, + "content": "Datasets. We conduct extensive experiments on two large-scale outdoor LiDAR semantic segmentation datasets, i.e., SemanticKITTI [3] and nuScenes [5, 24], and one indoor dataset ScanNet [20]. The nuScenes dataset contains 700 scenes for training, 150 scenes for validation, and 150 scenes for testing, where 16 classes are utilized for LiDAR semantic segmentation. As for SemanticKITTI, it contains 19 classes for training and evaluation. It has 22 sequences, where sequences 00 to 10, 08, and 11 to 21 are used for training, validation, and testing, respectively. ScanNet [20] contains 1603 scans with 20 classes, where 1201 scans are for training, 312 scans are for validation, and 100 scans are for testing." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.825, + 0.897, + 0.903 + ], + "angle": 0, + "content": "Implementation Details. We follow SLidR [51] to pretrain the network on the nuScenes [5, 24] dataset. The network is pre-trained on all keyframes from 600 scenes. Besides, the pre-trained network is fine-tuned on SemanticKITTI [3] to verify the generalization ability. We" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7024" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.09, + 0.47, + 0.118 + ], + "angle": 0, + "content": "Table 1. Comparisons (mIoU) among self-supervised methods on the nuScenes [24], SemanticKITTI [3], and ScanNet [20] val sets." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.124, + 0.468, + 0.23 + ], + "angle": 0, + "content": "
InitializationnuScenesSemanticKITTIScanNet
1%100%1%100%5%100%
Random42.269.132.552.146.163.3
PPKT [44]48.070.139.153.147.564.2
SLidR [51]48.270.439.654.347.964.9
PointContrast [55]47.269.237.152.347.664.5
CLIP2Scene56.371.542.655.048.465.1
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.244, + 0.468, + 0.271 + ], + "angle": 0, + "content": "Table 2. Annotation-free 3D semantic segmentation performance (mIoU) on the nuScenes [24] and ScanNet [20] val sets." + }, + { + "type": "table", + "bbox": [ + 0.17, + 0.277, + 0.382, + 0.315 + ], + "angle": 0, + "content": "
MethodnuScenesScanNet
CLIP2Scene20.8025.08
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.34, + 0.47, + 0.628 + ], + "angle": 0, + "content": "leverage the CLIP model to generate image features and text embedding. Following MaskCLIP, we modify the attention pooling layer of the CLIP's image encoder, thus extracting the dense pixel-text correspondences. We take SPVCNN [52] as the 3D network to produce the point-wise feature. The framework is developed on PyTorch, where the CLIP model is frozen during training. The training time is about 40 hours for 20 epochs on two NVIDIA Tesla A100 GPUs. The optimizer is SGD with a cosine scheduler. We set the temperature \\(\\lambda\\) and \\(\\tau\\) to be 1 and 0.5, respectively. The sweep number is set to be 3 empirically. Besides, We adopt MinkowskiNet14 [19] as the backbone for evaluation on the ScanNet dataset, where the number of sweeps is set to be 1 and the training epochs is 30. As for the Switchable Self-Training Strategy, we randomly switch the point supervision signal after 10 epochs. We apply several data augmentations in contrastive learning, including random rotation along the z-axis and random flip on the point cloud, random horizontal flip, and random crop-resize on the image." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.636, + 0.421, + 0.652 + ], + "angle": 0, + "content": "4.1. Annotation-free Semantic Segmentation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.66, + 0.47, + 0.765 + ], + "angle": 0, + "content": "After pre-training the network, we show the performance of the 3D network when it is not fine-tuned on any annotations (Table 2). As no previous method reports the 3D annotation-free segmentation performance, we compare our method with different setups (Table 3). In what follows, we describe the experimental settings and give insights into our method and the different settings." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Settings. We conduct experiments on the nuScenes and ScanNet datasets to evaluate the annotation-free semantic segmentation performance. Following MaskCLIP [61], we place the class name into 85 hand-craft prompts and feed it into the CLIP's text encoder to produce multiple text features. We then average the text features and feed the averaged features to the classifier for point-wise prediction. Besides, to explore how to effectively transfer CLIP's knowledge to the 3D network for annotation-free segmentation, We conduct" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.09, + 0.894, + 0.118 + ], + "angle": 0, + "content": "Table 3. Ablation study on the nuScenes [24] val set for annotation-free 3D semantic segmentation." + }, + { + "type": "table", + "bbox": [ + 0.554, + 0.124, + 0.844, + 0.368 + ], + "angle": 0, + "content": "
Ablation TargetSettingmIoU (%)
-Baseline15.1
PromptsnuScenes15.1 (+0.0)
SemanticKITTI13.9 (-1.2)
Cityscapes11.3 (-3.8)
All15.3 (+0.2)
Regularizationw/o StCR19.8 (+4.7)
w/o SCR16.8 (+1.7)
KL0.0 (-15.1)
Training Strategyw/o S318.8 (+3.7)
ST10.1 (-4.0)
Sweeps1 sweep18.7 (+3.6)
3 sweeps20.8 (+5.7)
5 sweeps20.6 (+5.5)
merged18.6 (+3.5)
Full ConfigurationCLIP2Scene20.8 (+5.7)
" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.395, + 0.893, + 0.424 + ], + "angle": 0, + "content": "the following experiments to highlight the effectiveness of different modules in our framework." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.427, + 0.894, + 0.472 + ], + "angle": 0, + "content": "Baseline. The input of the 3D semantic segmentation network is only one sweep, and we pre-train the framework via semantic consistency regularization." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.474, + 0.894, + 0.534 + ], + "angle": 0, + "content": "Prompts (nuScenes, SemanticKITTI, Cityscapes, All). Based on the baseline, we respectively replace the nuScenes, SemanticKITTI, Cityscapes, and all class names into the prompts to produce the text embedding." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.535, + 0.894, + 0.626 + ], + "angle": 0, + "content": "Regularization (w/o StCR, w/o SCR, KL). Based on the full method, we remove the Spatial-temporal Consistency Regularization (w/o StCR) and remove the Semantic Consistency Regularization (w/o SCR). Besides, we abuse both StCR and SCR and distill the image feature to the point cloud by Kullback-Leibler (KL) divergence loss." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.627, + 0.894, + 0.687 + ], + "angle": 0, + "content": "Training Strategies (w/o S3, ST). We abuse the Switchable Self-Training Strategy (w/o S3) in the full method. Besides, we show the performance of only training the 3D network by their own predictions after ten epochs (ST)." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.689, + 0.893, + 0.749 + ], + "angle": 0, + "content": "Sweeps Number (1 sweep, 3 sweeps, 5 sweeps, and merged). We set the sweep number \\( K \\) to be 1, 3, and 5, respectively. Besides, we also take three sweeps of the point cloud as the input to pre-train the network (merged)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.895, + 0.902 + ], + "angle": 0, + "content": "Effect of Different Prompts. To verify how text embedding affects the performance, we generate various text embedding by the class name from different datasets (nuScenes, SemanticKITTI, and Cityscapes) and all classes for pre-training the framework. As shown in Table 3, we find that even learning with other datasets' text embedding (SemanticKITTI and Cityscapes), the 3D network could still recognize the nuScenes's objects with decent performance (13.9% and 11.3% mIoU, respectively). The result shows that the 3D network is capable of open-vocabulary recognition ability." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7025" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.091, + 0.098, + 0.885, + 0.375 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.382, + 0.893, + 0.425 + ], + "angle": 0, + "content": "Figure 5. Qualitative results of annotation-free semantic segmentation on nuScenes dataset. Note that we show the results of individual classes. From the left to the right column are bus, motorcycle, car, and truck, respectively. The first row [ground truth] is the annotated semantic label. The second row [ours*] is our prediction of the highlighted target. The third row [ours] is our prediction of full classes." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.444, + 0.473, + 0.61 + ], + "angle": 0, + "content": "Effect of Semantic and Spatial-temporal Consistency Regularization. We remove Spatial-temporal Consistency Regularization (w/o SCR) from our method. Experiments show that the performance is dramatically decreased, indicating the effectiveness of our design. Besides, we also distill the image feature to the point cloud by KL divergence loss, where the text embeddings calculate the logits. However, such a method fails to transfer the semantic information from the image. The main reason is the noise-assigned semantics of the image pixel and the imperfect pixel-point correspondence due to the calibration error." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.612, + 0.472, + 0.702 + ], + "angle": 0, + "content": "Effect of Switchable Self-training Strategy. To examine the effect of the Switchable Self-Training Strategy, we either train the network with image supervision (w/o S3) or train the 3D network by their own predictions. Both trials witness a performance drop, indicating Switchable Self-Training Strategy is efficient in cross-modal self-supervised learning." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.703, + 0.472, + 0.809 + ], + "angle": 0, + "content": "Effect of Sweep Numbers. Intuitively, the performance of our method benefits from more sweeps information. Therefore, we also show the performance when restricting sweep size to 1, 3, and 5, respectively. However, we observe that the performance of 5 sweeps is similar to 3 sweeps but is more computationally expensive. Thus, we empirically set the sweep number to be 3." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.81, + 0.472, + 0.903 + ], + "angle": 0, + "content": "Qualitative Evaluation. The qualitative evaluations of individual classes (bus, motorcycle, car, and truck) are in Fig. 5, indicating that our method is able to perceive the objects even without training on any annotated data. However, we also observe the false positive predictions around the ground truth objects. We will resolve this issue in future work." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.444, + 0.875, + 0.46 + ], + "angle": 0, + "content": "4.2. Annotation-efficient Semantic Segmentation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.47, + 0.895, + 0.575 + ], + "angle": 0, + "content": "The pre-trained 3D network also boosts the performance when few labeled data are available for training. We directly compare SLidR [51], the only published method for image-to-Lidar self-supervised representation distillation. Besides, we also compared PPKT [44] and PointContrast [55]. In the following, we introduce SLidR and PPKT and compare them in detail." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.578, + 0.895, + 0.685 + ], + "angle": 0, + "content": "PPKT. PPKT is a cross-modal self-supervised method for the RGB-D dataset. It performs 2D-to-3D knowledge distillation via pixel-to-point contrastive loss. For a fair comparison, we use the same 3D network and training protocol but replace our semantic and Spatio-Temporal Regularization with InfoNCE loss. The framework is trained 50 epochs on 4096 randomly selected image-to-point pairs." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.687, + 0.895, + 0.763 + ], + "angle": 0, + "content": "SLidR. SLidR is an image-to-Lidar self-supervised method for autonomous driving data. Compared with PPKT, it introduces image super-pixel into cross-modal self-supervised learning. For a fair comparison, we replace our loss function with their superpixel-driven contrastive loss." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.897, + 0.903 + ], + "angle": 0, + "content": "Performance. As shown in Table 1, our method significantly outperforms the state-of-the-art methods when fine-tuned on \\(1\\%\\) and \\(100\\%\\) nuScenes dataset, with the improvement of \\(8.1\\%\\) and \\(1.1\\%\\), respectively. Compared with the random initialization, the improvement is \\(14.1\\%\\) and \\(2.4\\%\\), respectively, indicating the efficiency of our Semantic-driven Cross-modal Contrastive Learning framework. The qualitative results are shown in Fig. 6. Besides, we also verify the cross-domain generalization ability of our method. When pre-training the" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7026" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.088, + 0.098, + 0.891, + 0.4 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.406, + 0.895, + 0.449 + ], + "angle": 0, + "content": "Figure 6. Qualitative results of fine-tuning on \\(1\\%\\) nuScenes dataset. From the first row to the last row are the input LiDAR scan, ground truth, prediction of SLidR, and our prediction, respectively. Note that we show the results by error map, where the red point indicates the wrong prediction. Apparently, our method achieves decent performance." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.475, + 0.471, + 0.52 + ], + "angle": 0, + "content": "3D network on the nuScenes dataset and fine-tuning on \\(1\\%\\) and \\(100\\%\\) SemanticKITTI dataset, our method significantly outperforms other state-of-the-art self-supervised methods." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.524, + 0.474, + 0.903 + ], + "angle": 0, + "content": "Discussions. PPKT and SLidR reveal that contrastive loss is promising for transferring knowledge from image to point cloud. Like self-supervised learning, constructing the positive and negative samples is vital to unsupervised cross-modal knowledge distillation. However, previous methods suffer from optimization-conflict issues, i.e., some negative paired samples are actually positive pairs. For example, the road occupies a large proportion of the point cloud in a scene and is supposed to have the same semantics in the semantic segmentation task. When randomly selecting training samples, most negatively defined road-road points are actually positive. When feedforwarding such samples into contrastive learning, the contrastive loss will push them away in the embedding space, leading to unsatisfactory representation learning and hammering the downstream tasks' performance. SLidR introduces superpixel-driven contrastive learning to alleviate such issues. The motivation is that the visual representation of the image pixel and the projected points are consistent intra-superpixel. Although avoiding selecting negative image-point pairs from the same superpixel, the conflict still exists inter-superpixel. In our CLIP2Scene, we introduce the free-available dense pixel-text correspondence to alleviate the optimization conflicts. The text embedding represents the semantic information and can be used to select more reasonable training samples for contrastive learning." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.475, + 0.895, + 0.611 + ], + "angle": 0, + "content": "Besides training sample selection, the previous method also ignores the temporal coherence of the multi-sweep point cloud. That is, for LiDAR points mapping to the same image pixel, their feature is restricted to be consistent. Besides, considering the calibration error between the LiDAR scan and the camera image. We relax the pixel-to-point mapping to image grid-to-point grid mapping for consistency regularization. To this end, our Spatial-temporal consistency regularization leads to a more rational point representation." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.613, + 0.897, + 0.704 + ], + "angle": 0, + "content": "Last but not least, we find that randomly switching the supervision signal benefits self-supervised learning. Essentially, different modality networks learn different feature representations. They can filter different types of errors introduced by noisy pseudo labels. By this switchable operation, the error flows can be reduced mutually." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.723, + 0.62, + 0.739 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.897, + 0.903 + ], + "angle": 0, + "content": "We explored how CLIP knowledge benefits 3D scene understanding in this work, termed CLIP2Scene. To efficiently transfer CLIP's image and text features to a 3D network, we propose a novel Semantic-driven Cross-modal Contrastive Learning framework including Semantic Regularization and Spatial-Temporal Regularization. For the first time, our pre-trained 3D network achieves annotation-free 3D semantic segmentation with decent performance. Besides, our method significantly outperforms state-of-the-art self-supervised methods when fine-tuning with labelled data." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.958 + ], + "angle": 0, + "content": "7027" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.116, + 0.47, + 0.171 + ], + "angle": 0, + "content": "[1] Zeynep Akata, Florent Perronnin, Zaid Harchaoui, and Cordelia Schmid. Label-embedding for attribute-based classification. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 819–826, 2013. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.173, + 0.472, + 0.241 + ], + "angle": 0, + "content": "[2] Zeynep Akata, Scott E. Reed, Daniel Walter, Honglak Lee, and Bernt Schiele. Evaluation of output embeddings for fine-grained image classification. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2927-2936, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.245, + 0.472, + 0.313 + ], + "angle": 0, + "content": "[3] Jens Behley, Martin Garbade, Andres Milioto, Jan Quenzel, Sven Behnke, Cyril Stachniss, and Jurgen Gall. Semantickitti: A dataset for semantic scene understanding of lidar sequences. In IEEE/CVF International Conference on Computer Vision, pages 9297-9307, 2019. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.316, + 0.472, + 0.371 + ], + "angle": 0, + "content": "[4] Max Bucher, Stéphane Herbin, and Frédéric Jurie. Generating visual representations for zero-shot classification. In IEEE/CVF International Conference on Computer Vision Workshop, pages 2666-2673, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.373, + 0.472, + 0.455 + ], + "angle": 0, + "content": "[5] Holger Caesar, Varun Bankiti, Alex H Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giacarlo Baldan, and Oscar Beijbom. nuscenes: A multimodal dataset for autonomous driving. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11621-11631, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.458, + 0.472, + 0.514 + ], + "angle": 0, + "content": "[6] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In European Conference on Computer Vision, pages 213-229, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.515, + 0.472, + 0.584 + ], + "angle": 0, + "content": "[7] Mathilde Caron, Ishan Misra, Julien Mairal, Priya Goyal, Piotr Bojanowski, and Armand Joulin. Unsupervised learning of visual features by contrasting cluster assignments. In Advances in Neural Information Processing Systems, volume 33, pages 9912-9924, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.587, + 0.472, + 0.642 + ], + "angle": 0, + "content": "[8] Soravit Changpinyo, Wei-Lun Chao, Boqing Gong, and Fei Sha. Synthesized classifiers for zero-shot learning. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5327-5336, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.644, + 0.472, + 0.713 + ], + "angle": 0, + "content": "[9] Nenglun Chen, Lingjie Liu, Zhiming Cui, Runnan Chen, Duygu Ceylan, Changhe Tu, and Wenping Wang. Unsupervised learning of intrinsic structural representation points. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9121-9130, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.715, + 0.47, + 0.757 + ], + "angle": 0, + "content": "[10] Runnan Chen, Zhu Xinge, Nenglun Chen, Dawei Wang, Wei Li, Yuexin Ma, Ruigang Yang, and Wenping Wang. Referring self-supervised learning on 3d point cloud. 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.758, + 0.47, + 0.814 + ], + "angle": 0, + "content": "[11] Runnan Chen, Xinge Zhu, Nenglun Chen, Wei Li, Yuexin Ma, Ruigang Yang, and Wenping Wang. Zero-shot point cloud segmentation by transferring geometric primitives. arXiv preprint arXiv:2210.09923, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.816, + 0.47, + 0.871 + ], + "angle": 0, + "content": "[12] Runnan Chen, Xinge Zhu, Nenglun Chen, Dawei Wang, Wei Li, Yuexin Ma, Ruigang Yang, and Wenping Wang. Towards 3d scene understanding by referring synthetic models. arXiv preprint arXiv:2203.10546, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.47, + 0.902 + ], + "angle": 0, + "content": "[13] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.116, + 0.472, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "of visual representations. In International Conference on Machine Learning, pages 1597-1607, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.121, + 0.894, + 0.162 + ], + "angle": 0, + "content": "[14] Xinlei Chen and Kaiming He. Exploring simple siamese representation learning. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15750-15758, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.163, + 0.894, + 0.232 + ], + "angle": 0, + "content": "[15] Ran Cheng, Ryan Razani, Ehsan Taghavi, Enxu Li, and Bingbing Liu. (af)2-s3net: Attentive feature fusion with adaptive feature selection for sparse semantic segmentation network. In IEEE Conference on Computer Vision and Pattern Recognition, pages 12547-12556, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.233, + 0.894, + 0.287 + ], + "angle": 0, + "content": "[16] Ali Cheraghian, Shafin Rahman, Dylan Campbell, and Lars Petersson. Mitigating the Hubness Problem for Zero-Shot Learning of 3D Objects. arXiv preprint arXiv:1907.06371, 2019.2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.288, + 0.894, + 0.343 + ], + "angle": 0, + "content": "[17] Ali Cheraghian, Shafin Rahman, Townim F Chowdhury, Dylan Campbell, and Lars Petersson. Zero-shot learning on 3d point cloud objects and beyond. International Journal of Computer Vision, 130(10):2364-2384, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.344, + 0.894, + 0.398 + ], + "angle": 0, + "content": "[18] Ali Cheraghian, Shafin Rahman, and Lars Petersson. Zero-Shot Learning of 3D Point Cloud Objects. In IEEE International Conference on Machine Vision Applications, pages 1-6, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.4, + 0.892, + 0.455 + ], + "angle": 0, + "content": "[19] Christopher Choy, JunYoung Gwak, and Silvio Savarese. 4d spatio-temporal convnets: Minkowski convolutional neural networks. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3075-3084, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.457, + 0.894, + 0.524 + ], + "angle": 0, + "content": "[20] Angela Dai, Angel Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5828-5839, 2017. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.525, + 0.894, + 0.594 + ], + "angle": 0, + "content": "[21] Berkan Demirel, Ramazan Gokberk Cinbis, and Nazli Ikizler-Cinbis. Attributes2classname: A discriminative model for attribute-based unsupervised zero-shot learning. In IEEE/CVF International Conference on Computer Vision, pages 1241-1250, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.595, + 0.894, + 0.65 + ], + "angle": 0, + "content": "[22] Carl Doersch, Abhinav Gupta, and Alexei A Efros. Unsupervised visual representation learning by context prediction. In IEEE/CVF International Conference on Computer Vision, pages 1422-1430, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.651, + 0.894, + 0.718 + ], + "angle": 0, + "content": "[23] Alexey Dosovitskiy, Jost Tobias Springenberg, Martin Ried-miller, and Thomas Brox. Discriminative unsupervised feature learning with convolutional neural networks. In Advances in Neural Information Processing Systems, volume 27, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.72, + 0.894, + 0.788 + ], + "angle": 0, + "content": "[24] Whye Kit Fong, Rohit Mohan, Juana Valeria Hurtado, Lubing Zhou, Holger Caesar, Oscar Beijbom, and Abhinav Valada. Panoptic nuscenes: A large-scale benchmark for lidar panoptic segmentation and tracking. IEEE Robotics and Automation Letters, 7:3795-3802, 2022. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.79, + 0.894, + 0.845 + ], + "angle": 0, + "content": "[25] Chuang Gan, Ming Lin, Yi Yang, Yueting Zhuang, and Alexander Hauptmann. Exploring semantic inter-class relationships (sir) for zero-shot action recognition. In AAAI Conference on Artificial Intelligence, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.846, + 0.894, + 0.901 + ], + "angle": 0, + "content": "[26] Biao Gao, Yancheng Pan, Chengkun Li, Sibo Geng, and Huijing Zhao. Are we hungry for 3d lidar data for semantic segmentation? a survey of datasets and methods. IEEE Transactions on Intelligent Transportation Systems, 2021. 1" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.894, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "7028" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.176 + ], + "angle": 0, + "content": "[27] Jean-Bastien Grill, Florian Strub, Florent Altché, Corentin Tallec, Pierre Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Avila Pires, Zhaohan Guo, Mohammad Gheshlaghi Azar, et al. Bootstrap your own latent-a new approach to self-supervised learning. In Advances in Neural Information Processing Systems, volume 33, pages 21271-21284, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.177, + 0.472, + 0.232 + ], + "angle": 0, + "content": "[28] Yulan Guo, Hanyun Wang, Qingyong Hu, Hao Liu, Li Liu, and Mohammed Bennamoun. Deep learning for 3d point clouds: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence, 43(12):4338-4364, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.233, + 0.472, + 0.3 + ], + "angle": 0, + "content": "[29] Bo Han, Quanming Yao, Xingrui Yu, Gang Niu, Miao Xu, Weihua Hu, Ivor Tsang, and Masashi Sugiyama. Co-teaching: Robust training of deep neural networks with extremely noisy labels. Advances in neural information processing systems, 31, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.302, + 0.472, + 0.357 + ], + "angle": 0, + "content": "[30] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9729-9738, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.358, + 0.472, + 0.413 + ], + "angle": 0, + "content": "[31] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 770-778, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.414, + 0.472, + 0.468 + ], + "angle": 0, + "content": "[32] Fangzhou Hong, Lingdong Kong, Hui Zhou, Xinge Zhu, Hongsheng Li, and Ziwei Liu. Unified 3d and 4d panoptic segmentation via dynamic shifting network. arXiv preprint arXiv:2203.07186, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.469, + 0.472, + 0.537 + ], + "angle": 0, + "content": "[33] Yuenan Hou, Xinge Zhu, Yuexin Ma, Chen Change Loy, and Yikang Li. Point-to-Voxel Knowledge Distillation for LiDAR Semantic Segmentation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8479-8488, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.538, + 0.472, + 0.594 + ], + "angle": 0, + "content": "[34] Siyuan Huang, Yichen Xie, Song-Chun Zhu, and Yixin Zhu. Spatio-temporal self-supervised representation learning for 3d point clouds. In IEEE/CVF International Conference on Computer Vision, pages 6535-6545, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.595, + 0.472, + 0.65 + ], + "angle": 0, + "content": "[35] Ajay Jain, Matthew Tancik, and Pieter Abbeel. Putting nerf on a diet: Semantically consistent few-shot view synthesis. In IEEE/CVF International Conference on Computer Vision, pages 5885-5894, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.651, + 0.472, + 0.692 + ], + "angle": 0, + "content": "[36] Sosuke Kobayashi, Eiichi Matsumoto, and Vincent Sitzmann. Decomposing nef for editing via feature field distillation. arXiv preprint arXiv:2205.15585, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.692, + 0.472, + 0.747 + ], + "angle": 0, + "content": "[37] Elyor Kodirov, Tao Xiang, and Shaogang Gong. Semantic autoencoder for zero-shot learning. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4447-4456, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.748, + 0.472, + 0.803 + ], + "angle": 0, + "content": "[38] Lingdong Kong, Youquan Liu, Runnan Chen, Yuexin Ma, Xinge Zhu, Yikang Li, Yuenan Hou, Yu Qiao, and Ziwei Liu. Rethinking range view representation for lidar segmentation. arXiv preprint arXiv:2303.05367, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.804, + 0.472, + 0.871 + ], + "angle": 0, + "content": "[39] Lingdong Kong, Niamul Quader, and Venice Erin Liong. Conda: Unsupervised domain adaptation for lidar segmentation via regularized domain concatenation. In IEEE International Conference on Robotics and Automation, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.472, + 0.902 + ], + "angle": 0, + "content": "[40] Lingdong Kong, Jiawei Ren, Liang Pan, and Ziwei Liu. Lasermix for semi-supervised lidar semantic segmentation." + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.122, + 0.894, + 0.178 + ], + "angle": 0, + "content": "[41] Christoph H. Lampert, Hannes Nickisch, and Stefan Harmeling. Learning to detect unseen object classes by between-class attribute transfer. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 951-958, 2009. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.179, + 0.894, + 0.234 + ], + "angle": 0, + "content": "[42] Christoph H. Lampert, Hannes Nickisch, and Stefan Harmeling. Attribute-based classification for zero-shot visual object categorization. IEEE Transactions on Pattern Analysis and Machine Intelligence, 36:453-465, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.235, + 0.894, + 0.277 + ], + "angle": 0, + "content": "[43] Yan Li, Zhen Jia, Junge Zhang, Kaiqi Huang, and Tieniu Tan. Deep semantic structural constraints for zero-shot learning. In AAAI Conference on Artificial Intelligence, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.279, + 0.894, + 0.348 + ], + "angle": 0, + "content": "[44] Yueh-Cheng Liu, Yu-Kai Huang, Hung-Yueh Chiang, Hung-Ting Su, Zhe-Yu Liu, Chin-Tang Chen, Ching-Yu Tseng, and Winston H Hsu. Learning from 2d: Contrastive pixel-topoint knowledge transfer for 3d pretraining. arXiv preprint arXiv:2104.04687, 2021. 1, 2, 3, 4, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.349, + 0.894, + 0.405 + ], + "angle": 0, + "content": "[45] Björn Michele, Alexandre Boulch, Gilles Puy, Maxime Bucher, and Renaud Marlet. Generative zero-shot learning for semantic segmentation of 3d point clouds. In International Conference on 3D Vision, pages 992-1002, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.406, + 0.894, + 0.476 + ], + "angle": 0, + "content": "[46] Ashish Mishra, M. Shiva Krishna Reddy, Anurag Mittal, and Hema A. Murthy. A generative model for zero shot learning using conditional variational autoencoders. In IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshop, pages 2269-22698, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.477, + 0.894, + 0.533 + ], + "angle": 0, + "content": "[47] Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. Pointnet: Deep learning on point sets for 3d classification and segmentation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 652-660, 2017. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.534, + 0.894, + 0.616 + ], + "angle": 0, + "content": "[48] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.618, + 0.894, + 0.687 + ], + "angle": 0, + "content": "[49] Yongming Rao, Wenliang Zhao, Guangyi Chen, Yansong Tang, Zheng Zhu, Guan Huang, Jie Zhou, and Jiwen Lu. Denseclip: Language-guided dense prediction with context-aware prompting. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18082-18091, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.689, + 0.894, + 0.743 + ], + "angle": 0, + "content": "[50] David Rozenberszki, Or Litany, and Angela Dai. Language-grounded indoor 3d semantic segmentation in the wild. In European Conference on Computer Vision, pages 125-141, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.745, + 0.894, + 0.815 + ], + "angle": 0, + "content": "[51] Corentin Sautier, Gilles Puy, Spyros Gidaris, Alexandre Boulch, Andrei Bursuc, and Renaud Marlet. Image-to-lidar self-supervised distillation for autonomous driving data. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9891-9901, 2022. 1, 2, 3, 4, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.817, + 0.894, + 0.872 + ], + "angle": 0, + "content": "[52] Haotian Tang, Zhijian Liu, Shengyu Zhao, Yujun Lin, Ji Lin, Hanrui Wang, and Song Han. Searching efficient 3d architectures with sparse point-voxel convolution. In European Conference on Computer Vision, pages 685–702, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.873, + 0.894, + 0.901 + ], + "angle": 0, + "content": "[53] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "7029" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.092, + 0.47, + 0.12 + ], + "angle": 0, + "content": "Polosukhin. Attention is all you need. In Advances in Neural Information Processing Systems, volume 30, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.122, + 0.472, + 0.189 + ], + "angle": 0, + "content": "[54] Yongqin Xian, Zeynep Akata, Gaurav Sharma, Quynh N. Nguyen, Matthias Hein, and Bernt Schiele. Latent embeddings for zero-shot classification. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 69-77, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.192, + 0.472, + 0.26 + ], + "angle": 0, + "content": "[55] Saining Xie, Jiatao Gu, Demi Guo, Charles R Qi, Leonidas Guibas, and Or Litany. Pointcontrast: Unsupervised pretraining for 3d point cloud understanding. In European Conference on Computer Vision, pages 574-591, 2020. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.263, + 0.472, + 0.331 + ], + "angle": 0, + "content": "[56] Jianyun Xu, Ruixiang Zhang, Jian Dou, Yushi Zhu, Jie Sun, and Shiliang Pu. Rpvnet: A deep and efficient range-point-voxel fusion network for lidar point cloud segmentation. In IEEE/CVF International Conference on Computer Vision, pages 16024-16033, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.333, + 0.472, + 0.401 + ], + "angle": 0, + "content": "[57] Lewei Yao, Jianhua Han, Youpeng Wen, Xiaodan Liang, Dan Xu, Wei Zhang, Zhenguo Li, Chunjing Xu, and Hang Xu. Detclip: Dictionary-enriched visual-concept paralleled pre-training for open-world detection. arXiv preprint arXiv:2209.09407, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.403, + 0.472, + 0.459 + ], + "angle": 0, + "content": "[58] Éloi Zablocki, Patrick Bordes, Benjamin Piwowarski, Laure Soulier, and Patrick Gallinari. Context-aware zero-shot learning for object recognition. In International Conference on Machine Learning, pages 7292-7303, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.46, + 0.472, + 0.528 + ], + "angle": 0, + "content": "[59] Renrui Zhang, Ziyu Guo, Wei Zhang, Kunchang Li, Xupeng Miao, Bin Cui, Yu Qiao, Peng Gao, and Hongsheng Li. Pointclip: Point cloud understanding by clip. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8552-8562, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.531, + 0.472, + 0.586 + ], + "angle": 0, + "content": "[60] Zaiwei Zhang, Rohit Girdhar, Armand Joulin, and Ishan Misra. Self-supervised pretraining of 3d features on any point-cloud. In IEEE/CVF International Conference on Computer Vision, pages 10252-10263, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.588, + 0.472, + 0.628 + ], + "angle": 0, + "content": "[61] Chong Zhou, Chen Change Loy, and Bo Dai. Extract free dense labels from clip. In European Conference on Computer Vision, pages 696-712, 2022. 2, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.63, + 0.472, + 0.699 + ], + "angle": 0, + "content": "[62] Xinge Zhu, Hui Zhou, Tai Wang, Fangzhou Hong, Yuexin Ma, Wei Li, Hongsheng Li, and Dahua Lin. Cylindrical and asymmetrical 3d convolution networks for lidar segmentation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9939-9948, 2021. 1" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.699 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.956 + ], + "angle": 0, + "content": "7030" + } + ] +] \ No newline at end of file diff --git a/2023/CLIP2Scene_ Towards Label-Efficient 3D Scene Understanding by CLIP/58dade4c-9eec-4c3a-968d-a582ab672d68_origin.pdf b/2023/CLIP2Scene_ Towards Label-Efficient 3D Scene Understanding by CLIP/58dade4c-9eec-4c3a-968d-a582ab672d68_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..df0b0e352dda4d6da0eb343e76a810437c37940f --- /dev/null +++ b/2023/CLIP2Scene_ Towards Label-Efficient 3D Scene Understanding by CLIP/58dade4c-9eec-4c3a-968d-a582ab672d68_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:864de3983713ddb7ab02f912a20fcb371ae9698f7c0ca632263a756ad6ff817b +size 2756188 diff --git a/2023/CLIP2Scene_ Towards Label-Efficient 3D Scene Understanding by CLIP/full.md b/2023/CLIP2Scene_ Towards Label-Efficient 3D Scene Understanding by CLIP/full.md new file mode 100644 index 0000000000000000000000000000000000000000..0375b848c2b3a5b02a482bfb053ec4800d9df305 --- /dev/null +++ b/2023/CLIP2Scene_ Towards Label-Efficient 3D Scene Understanding by CLIP/full.md @@ -0,0 +1,269 @@ +# CLIP2Scene: Towards Label-efficient 3D Scene Understanding by CLIP + +Runnan Chen $^{1,2}$ Youquan Liu $^{2,3}$ Lingdong Kong $^{2,4}$ Xinge Zhu $^{5}$ Yuexin Ma $^{6}$ Yikang Li $^{2}$ Yuenan Hou $^{2,\dagger}$ Yu Qiao $^{2}$ Wenping Wang $^{7,\dagger}$ + +1The University of Hong Kong 2Shanghai AI Laboratory 3Hochschule Bremerhaven 4National University of Singapore 5The Chinese University of Hong Kong 6ShanghaiTech University 7Texas A&M University + +# Abstract + +Contrastive Language-Image Pre-training (CLIP) achieves promising results in 2D zero-shot and few-shot learning. Despite the impressive performance in 2D, applying CLIP to help the learning in 3D scene understanding has yet to be explored. In this paper, we make the first attempt to investigate how CLIP knowledge benefits 3D scene understanding. We propose CLIP2Scene, a simple yet effective framework that transfers CLIP knowledge from 2D image-text pre-trained models to a 3D point cloud network. We show that the pre-trained 3D network yields impressive performance on various downstream tasks, i.e., annotation-free and fine-tuning with labelled data for semantic segmentation. Specifically, built upon CLIP, we design a Semantic-driven Cross-modal Contrastive Learning framework that pre-trains a 3D network via semantic and spatial-temporal consistency regularization. For the former, we first leverage CLIP's text semantics to select the positive and negative point samples and then employ the contrastive loss to train the 3D network. In terms of the latter, we force the consistency between the temporally coherent point cloud features and their corresponding image features. We conduct experiments on SemanticKITTI, nuScenes, and ScanNet. For the first time, our pre-trained network achieves annotation-free 3D semantic segmentation with $20.8\%$ and $25.08\%$ mIoU on nuScenes and ScanNet, respectively. When fine-tuned with $1\%$ or $100\%$ labelled data, our method significantly outperforms other self-supervised methods, with improvements of $8\%$ and $1\%$ mIoU, respectively. Furthermore, we demonstrate the generalizability for handling cross-domain datasets. Code is publicly available1. + +# 1. Introduction + +3D scene understanding is fundamental in autonomous driving, robot navigation, etc [26,28]. Current deep learning- + +![](images/e635707803576d3ce4da203fd52c9028e20daf537b797f5ed1eded7a308ab370.jpg) +Semantic-driven Cross-modal Contrastive Learning +Figure 1. We explore how CLIP knowledge benefits 3D scene understanding. To this end, we propose CLIP2Scene, a Semantic-driven Cross-modal Contrastive Learning framework that leverages CLIP knowledge to pre-train a 3D point cloud segmentation network via semantic and spatial-temporal consistency regularization. CLIP2Scene yields impressive performance on annotation-free 3D semantic segmentation and significantly outperforms other self-supervised methods when fine-tuning on annotated data. + +based methods have shown inspirational performance on 3D point cloud data [15, 32, 33, 38, 47, 56, 62]. However, some drawbacks hinder their real-world applications. The first one comes from their heavy reliance on the large collection of annotated point clouds, especially when high-quality 3D annotations are expensive to acquire [39, 40, 44, 51]. Besides, they typically fail to recognize novel objects that are never seen in the training data [11, 45]. As a result, it may need extra annotation efforts to train the model on recognizing these novel objects, which is both tedious and time-consuming. + +Contrastive Vision-Language Pre-training (CLIP) [48] provides a new perspective that mitigates the above issues in 2D vision. It was trained on large-scale free-available image-text pairs from websites and built vision-language + +correlation to achieve promising open-vocabulary recognition. MaskCLIP [61] further explores semantic segmentation based on CLIP. With minimal modifications to the CLIP pre-trained network, MaskCLIP can be directly used for the semantic segmentation of novel objects without additional training efforts. PointCLIP [59] reveals that the zero-shot classification ability of CLIP can be generalized from the 2D image to the 3D point cloud. It perspectively projects a point cloud frame into different views of 2D depth maps that bridge the modal gap between the image and the point cloud. The above studies indicate the potential of CLIP on enhancing the 2D segmentation and 3D classification performance. However, whether and how CLIP knowledge benefits 3D scene understanding is still under-explored. + +In this paper, we explore how to leverage CLIP's 2D image-text pre-learned knowledge for 3D scene understanding. Previous cross-modal knowledge distillation methods [44,51] suffer from the optimization-conflict issue, i.e., some of the positive pairs are regarded as negative samples for contrastive learning, leading to unsatisfactory representation learning and hammering the performance of downstream tasks. Besides, they also ignore the temporal coherence of the multi-sweep point cloud, failing to utilize the rich inter-sweep correspondence. To handle the mentioned problems, we propose a novel Semantic-driven Cross-modal Contrastive Learning framework that fully leverages CLIP's semantic and visual information to regularize a 3D network. Specifically, we propose Semantic Consistency Regularization and Spatial-Temporal Consistency Regularization. In semantic consistency regularization, we utilize CLIP's text semantics to select the positive and negative point samples for less-conflict contrastive learning. For spatial-temporal consistency regularization, we take CLIP's image pixel feature to impose a soft consistency constraint on the temporally coherent point features. Such an operation also alleviates the effects of imperfect image-to-point calibration. + +We conduct several downstream tasks on the indoor and outdoor datasets to verify how the pre-trained network benefits the 3D scene understanding. The first one is annotation-free semantic segmentation. Following MaskCLIP, we place class names into multiple hand-crafted templates as prompts and average the text embeddings generated by CLIP to conduct the annotation-free segmentation. For the first time, our method achieves $20.8\%$ and $25.08\%$ mIoU annotation-free 3D semantic segmentation on the nuScenes [24] and ScanNet [20] datasets without training on any labelled data. Secondly, we compare with other self-supervised methods in label-efficient learning. When fine-tuning the 3D network with $1\%$ or $100\%$ labelled data on the nuScenes dataset, our method significantly outperforms state-of-the-art self-supervised methods, with improvements of $8\%$ and $1\%$ mIoU, respectively. Besides, to verify the generalization capability, we pre-train the network on the nuScenes dataset + +and evaluate it on SemanticKITTI [3]. Our method still significantly outperforms state-of-the-art methods. The key contributions of our work are summarized as follows. + +- The first work that distils CLIP knowledge to a 3D network for 3D scene understanding. +- We propose a novel Semantic-driven Cross-modal Contrastive Learning framework that pre-trains a 3D network via spatial-temporal and semantic consistency regularization. +- We propose a novel Semantic-guided Spatial-Temporal Consistency Regularization that forces the consistency between the temporally coherent point cloud features and their corresponding image features. +- For the first time, our method achieves promising results on annotation-free 3D scene segmentation. When fine-tuning with labelled data, our method significantly outperforms state-of-the-art self-supervised methods. + +# 2. Related Work + +Zero-shot Learning in 3D. The objective of zero-shot learning (ZSL) is to recognize objects that are unseen in the training set. Many efforts have been devoted to the 2D recognition tasks [1, 2, 4, 8, 21, 25, 34, 37, 41–43, 46, 54, 55, 58, 60], and few works concentrate on performing ZSL in the 3D domain [11, 16–18, 45]. [18] applies ZSL to 3D tasks, where they train PointNet [47] on "seen" samples and test on "unseen" samples. Subsequent work [16] addresses the hubness problem caused by the low-quality point cloud features. [17] proposes the triplet loss to boost the performance under the transductive setting, where the "unseen" class is observed and unlabeled in the training phase. [11] makes the first attempt to explore the transductive zero-shot segmentation for 3D scene understanding. Recently, some studies introduced CLIP into zero-shot learning. MaskCLIP [61] investigates the problem of utilizing CLIP to help the 2D dense prediction tasks and exhibits encouraging zero-shot semantic segmentation performance. PointCLIP [59] is the pioneering work that applies CLIP to 3D recognition and shows impressive performance on zero-shot and few-shot classification tasks. Our work takes a step further to investigate how the rich semantic and visual knowledge in CLIP can benefit the 3D semantic segmentation tasks. + +Self-supervised Representation Learning. The purpose of self-supervised learning is to obtain a good representation that benefits the downstream tasks. The dominant approaches resort to contrastive learning to pre-train the network [7, 9, 9, 10, 12-14, 22, 23, 27, 30]. Recently, inspired by the success of CLIP, leveraging the pre-trained model of CLIP to the downstream tasks has raised the community's attention [35, 36, 49, 50, 57]. DenseCLIP [49] utilizes + +![](images/0c613f286b55db46731f70fd01a88b176c9d596b00fc0a3d0db65877657647eb.jpg) +Figure 2. Illustration of the Semantic-driven Cross-modal Contrastive Learning. Firstly, we obtain the text embedding $t_i$ , image pixel feature $x_i$ , and point feature $p_i$ by text encoder, image encoder, and point encoder, respectively. Secondly, we leverage CLIP knowledge to construct positive and negative samples for contrastive learning. Thus we obtain point-text pairs $\{x_i, t_i\}_{i=1}^M$ and all pixel-point-text pairs in a short temporal $\{\hat{x}_i^k, \hat{p}_i^k, t_i^k\}_{i=1,k=1}^M$ . Here, $\{x_i, t_i\}_{i=1}^M$ and $\{\hat{x}_i^k, \hat{p}_i^k, t_i^k\}_{i=1,k=1}^M$ are used for Semantic Consistency Regularization and Spatial-Temporal Consistency Regularization, respectively. Lastly, we perform Semantic Consistency Regularization by pulling the point features to their corresponding text embedding and Spatial-Temporal Consistency Regularization by mimicking the temporally coherent point features to their corresponding pixel features. + +the CLIP's pre-trained knowledge for dense image pixel prediction. DetCLIP [57] proposes a pre-training method equipped with CLIP for open-world detection. We leverage the image-text pre-trained CLIP knowledge to help 3D scene understanding. + +Cross-modal Knowledge Distillation. Recently, increasing studies have focused on transferring knowledge from 2D images to 3D point clouds for self-supervised representation learning [44, 51]. PPKT [44] resorts to the InfoNCE loss to help the 3D network distil rich knowledge from the 2D image backbone. SLidR [51] further introduce the superpixel to boost the cross-modal knowledge distillation. In this paper, we first attempt to pre-train a 3D network with CLIP's knowledge. + +# 3. Methodology + +Considering the impressive open-vocabulary performance achieved by CLIP in image classification and segmentation, natural curiosities have been raised. Can CLIP endow the ability to a 3D network for annotation-free scene understanding? And further, will it promote the network performance when fine-tuned on labelled data? To answer the above questions, we study the cross-modal knowledge transfer of CLIP for 3D scene understanding, termed CLIP2Scene. Our work is a pioneer in exploiting CLIP knowledge for 3D scene + +understanding. In what follows, we revisit the CLIP applied in 2D open-vocabulary classification and semantic segmentation, then present our CLIP2Scene in detail. Our approach consists of three major components: Semantic Consistency Regularization, Semantic-Guided Spatial-Temporal Consistency Regularization, and Switchable Self-Training Strategy. + +# 3.1. Revisiting CLIP + +Contrastive Vision-Language Pre-training (CLIP) mitigates the following drawbacks that dominate the computer vision field: 1. Deep models need a large amount of formatted and labelled training data, which is expensive to acquire; 2. The model's generalization ability is weak, making it difficult to migrate to a new scenario with unseen objects. CLIP consists of an image encoder (ResNet [31] or ViT [6]) and a text encoder (Transformer [53]), both respectively project the image and text representation to a joint embedding space. During training, CLIP constructs positive and negative samples from 400 million image-text pairs to train both encoders with a contrastive loss, where the large-scale image-text pairs are free-available from the Internet and assumed to contain every class of images and most concepts of text. Therefore, CLIP can achieve promising open-vocabulary recognition. + +For 2D zero-shot classification, CLIP first places the class name into a pre-defined template to generate the text embed- + +![](images/abb87bad06a099b3a04ab1504989c71b187a5965afe07839adca112ec8c71c3c.jpg) + +![](images/df1d6e894bf9f8cd67b1d613b7c758a6cf027723e49a24df834de1fab5e2dd5d.jpg) +Figure 3. Illustration of the image pixel-to-text mapping. The dense pixel-text correspondence $\{x_{i},t_{i}\}_{i = 1}^{M}$ is extracted by the off-the-shelf method MaskCLIP [61]. + +dings and then encodes images to obtain image embeddings. Next, it calculates the similarities between images and text embeddings to determine the class. MaskCLIP further extends CLIP into 2D semantic segmentation. Specifically, MaskCLIP modifies the attention pooling layer of the CLIP's image encoder, thus performing pixel-level mask prediction instead of the global image-level prediction. + +# 3.2. CLIP2Scene + +As shown in Fig. 2, we first leverage CLIP and 3D network to respectively extract the text embeddings, image pixel feature, and point feature. Secondly, we construct positive and negative samples based on CLIP's knowledge. Lastly, we impose Semantic Consistency Regularization by pulling the point features to their corresponding text embedding. At the same time, we apply Spatial-Temporal Consistency Regularization by forcing the consistency between temporally coherent point features and their corresponding pixel features. In what follows, we present the details and insights. + +# 3.2.1 Semantic Consistency Regularization + +As CLIP is pre-trained on 2D images and text, our first concern is the domain gap between 2D images and the 3D point cloud. To this end, we build dense pixel-point correspondence and transfer image knowledge to the 3D point cloud + +via the pixel-point pairs. Specifically, we calibrate the Li-DAR point cloud with corresponding images captured by six cameras. Therefore, the dense pixel-point correspondence $\{x_i,p_i\}_{i = 1}^M$ can be obtained accordingly, where $x_{i}$ and $p_i$ indicates $i$ -th paired image feature and point feature, which are respectively extracted by the CLIP's image encoder and the 3D network. $M$ is the number of pairs. Note that it is an online operation and is irrelevant to the image and point data augmentation. + +Previous methods [44, 51] provide a promising solution to cross-modal knowledge transfer. They first construct positive pixel-point pairs $\{x_i, p_i\}_{i=1}^M$ and negative pairs $\{x_i, p_j\} (i \neq j)$ , and then pull in the positive pairs while pushing away the negative pairs in the embedding space via the InfoNCE loss. Despite the encourageable performance of previous methods in transferring cross-modal knowledge, they are both confronted with the same optimization-conflict issue. For example, suppose $i$ -th pixel $x_i$ and $j$ -th point $p_j$ are in the different positions of the same instance with the same semantics. However, the InfoNCE loss will try to push them away, which is unreasonable and hammer the performance of the downstream tasks [51]. In light of this, we propose a Semantic Consistency Regularization that leverages the CLIP's semantic information to alleviate this issue. Specifically, we generate the dense pixel-text pairs $\{x_i, t_i\}_{i=1}^M$ by following the off-the-shelf method MaskCLIP [61] (Fig. 3), where $t_i$ is the text embedding generated from the CLIP's text encoder. Note that the pixel-text mappings are free-available from CLIP without any additional training. We then transfer pixel-text pairs to point-text pairs $\{p_i, t_i\}_{i=1}^M$ and utilize the text semantics to select the positive and negative point samples for contrastive learning. The objective function is as follows: + +$$ +\mathcal {L} _ {S - i n f o} = - \sum_ {c = 1} ^ {C} \log \frac {\sum_ {t _ {i} \in c , p _ {i}} \exp \left(D \left(t _ {i} , p _ {i}\right) / \tau\right)}{\sum_ {t _ {i} \in c , t _ {j} \notin c , p _ {j}} \exp \left(D \left(t _ {i} , p _ {j}\right) / \tau\right)}, \tag {1} +$$ + +where $t_i \in c$ indicates that $t_i$ is generated by $c$ -th classes name, and $C$ is the number of classes. $D$ denotes the scalar product operation and $\tau$ is a temperature term $(\tau > 0)$ . + +Since the text is composed of class names placed into pre-defined templates, the text embedding represents the semantic information of the corresponding class. Therefore, those points with the same semantics will be restricted near the same text embedding, and those with different semantics will be pushed away. To this end, our Semantic Consistency Regularization causes less conflict in contrastive learning. + +# 3.2.2 Semantic-guided Spatial-temporal Consistency Regularization + +Besides semantic consistency regularization, we consider how image pixel features help to regularize a 3D network. + +![](images/c03b35ee78740b90f8de34ea9faf125257f71553320e469e344e2ff96c32183f.jpg) +Figure 4. Illustration of the image pixel-to-point mapping (left) and semantic-guided fusion feature generation (right). We build the grid-wise correspondence between an image $I$ and the temporally coherent LiDAR point cloud $\{P_k\}_{k=1}^K$ within $S$ seconds and generate semantic-guided fusion features for individual grids. Both $\{\hat{x}_i^k, \hat{p}_i^k\}_{i=1,k=1}^M, K$ and $\{f_n\}_{n=1}^N$ are used to perform Spatial-Temporal Consistency Regularization. + +The natural alternative directly pulls in the point feature with its corresponding pixel in the embedding space. However, the noise-assigned semantics of the image pixel and the imperfect pixel-point mapping hinder the downstream task's performance. To this end, we propose a novel semantic-guided Spatial-Temporal Consistency Regularization to alleviate the problem by imposing a soft constraint on points within local space and time. + +Specifically, given an image $I$ and temporally coherent LiDAR point cloud $\{P_k\}_{k=1}^K$ , where $K$ is the number of sweeps within $S$ seconds. Note that the image is matched to the first frame of the point cloud $P_1$ with pixel-point pairs $\{\hat{x}_i^1, \hat{p}_i^1\}_{i=1}^M$ . We register the rest of the point cloud to the first frame via the calibration matrices and map them to the image (Fig. 4). Thus we obtain all pixel-point-text pairs in a short temporal $\{\hat{x}_i^k, \hat{p}_i^k, t_i^k\}_{i=1,k=1}^M$ . Next, we divide the entire stitched point cloud into regular grids $\{g_n\}_{n=1}^N$ , where the temporally coherent points are located in the same grid. We impose the spatial-temporal consistency constraint within individual grids by the following objective function: + +$$ +\mathcal {L} _ {S S R} = \sum_ {g _ {n}} \sum_ {(\hat {i}, \hat {k}) \in g _ {n}} (1 - \operatorname {s i g m o i d} \left(D \left(\hat {p} _ {\hat {i}} ^ {\hat {k}}, f _ {n}\right)\right)) / N, \tag {2} +$$ + +where $(\hat{i},\hat{k})\in g_{n}$ indicates the pixel-point pair $\{\hat{x}_i^k,\hat{p}_i^k\}$ is located in the $n$ -th grid. $\{f_n\}_{n = 1}^N$ is a semantic-guided cross-modal fusion feature formulated by: + +$$ +f _ {n} = \sum_ {(\hat {i}, \hat {k}) \in g _ {n}} a _ {\hat {i}} ^ {\hat {k}} * \hat {x} _ {\hat {i}} ^ {\hat {k}} + b _ {\hat {i}} ^ {\hat {k}} * \hat {p} _ {\hat {i}} ^ {\hat {k}}, \tag {3} +$$ + +where $a_{\hat{i}}^{\hat{k}}$ and $b_{\hat{i}}^{\hat{k}}$ are attention weight calculated by: + +$$ +\begin{array}{l} a _ {i} ^ {\hat {k}} = \frac {\exp (D (\hat {x} _ {\hat {i}} ^ {\hat {k}} , t _ {\hat {i}} ^ {1}) / \lambda)}{\sum_ {(\hat {i} , \hat {k}) \in g _ {n}} \exp (D (\hat {x} _ {\hat {i}} ^ {\hat {k}} , t _ {\hat {i}} ^ {1}) / \lambda) + \exp (D (\hat {p} _ {\hat {i}} ^ {\hat {k}} , t _ {\hat {i}} ^ {1}) / \lambda)}, \\ b _ {\hat {i}} ^ {\hat {k}} = \frac {\exp \left(D \left(\hat {p} _ {\hat {i}} ^ {\hat {k}} , t _ {\hat {i}} ^ {1}\right) / \lambda\right)}{\sum_ {(\hat {i} , \hat {k}) \in g _ {n}} \exp \left(D \left(\hat {x} _ {\hat {i}} ^ {\hat {k}}, t _ {\hat {i}} ^ {1}\right) / \lambda\right) + \exp \left(D \left(\hat {p} _ {\hat {i}} ^ {\hat {k}}, t _ {\hat {i}} ^ {1}\right) / \lambda\right)}, \tag {4} \\ \end{array} +$$ + +where $\lambda$ is the temperature term. + +Actually, those pixel and point features within the local grid $g_{n}$ are restricted near a dynamic centre $f_{n}$ . Thus, such a soft constraint alleviates the noisy prediction and calibration error issues. At the same time, it imposes Spatio-Temporal Regularization on the temporally coherent point features. + +# 3.2.3 Switchable Self-training Strategy + +We combine the loss function $\mathcal{L}_{S,\text{info}}$ and $\mathcal{L}_{SSR}$ to end-to-end train the whole network, where the CLIP's image and text encoder backbone are frozen during training. We find that method worked only when the pixel-point feature $\{x_i,p_i\}_{i = 1}^M$ and $\{\hat{x}_i^k,\hat{p}_i^k\}_{i = 1,k = 1}^{\hat{M},K}$ , which are used in $\mathcal{L}_{S,\text{info}}$ and $\mathcal{L}_{SSR}$ , are generated from different learnable linear layer. On top of that, we further put forward an effective strategy to promote performance. Specifically, after contrastive learning of the 3D network for a few epochs, we randomly switch the point pseudo label between the paired image pixel's pseudo label and the point's predicted label. Since different modality networks learn different feature representations, they can filter different types of error introduced by noisy pseudo labels. By this switchable operation, the error flows can be reduced by mutually [29]. + +# 4. Experiments + +Datasets. We conduct extensive experiments on two large-scale outdoor LiDAR semantic segmentation datasets, i.e., SemanticKITTI [3] and nuScenes [5, 24], and one indoor dataset ScanNet [20]. The nuScenes dataset contains 700 scenes for training, 150 scenes for validation, and 150 scenes for testing, where 16 classes are utilized for LiDAR semantic segmentation. As for SemanticKITTI, it contains 19 classes for training and evaluation. It has 22 sequences, where sequences 00 to 10, 08, and 11 to 21 are used for training, validation, and testing, respectively. ScanNet [20] contains 1603 scans with 20 classes, where 1201 scans are for training, 312 scans are for validation, and 100 scans are for testing. + +Implementation Details. We follow SLidR [51] to pretrain the network on the nuScenes [5, 24] dataset. The network is pre-trained on all keyframes from 600 scenes. Besides, the pre-trained network is fine-tuned on SemanticKITTI [3] to verify the generalization ability. We + +Table 1. Comparisons (mIoU) among self-supervised methods on the nuScenes [24], SemanticKITTI [3], and ScanNet [20] val sets. + +
InitializationnuScenesSemanticKITTIScanNet
1%100%1%100%5%100%
Random42.269.132.552.146.163.3
PPKT [44]48.070.139.153.147.564.2
SLidR [51]48.270.439.654.347.964.9
PointContrast [55]47.269.237.152.347.664.5
CLIP2Scene56.371.542.655.048.465.1
+ +Table 2. Annotation-free 3D semantic segmentation performance (mIoU) on the nuScenes [24] and ScanNet [20] val sets. + +
MethodnuScenesScanNet
CLIP2Scene20.8025.08
+ +leverage the CLIP model to generate image features and text embedding. Following MaskCLIP, we modify the attention pooling layer of the CLIP's image encoder, thus extracting the dense pixel-text correspondences. We take SPVCNN [52] as the 3D network to produce the point-wise feature. The framework is developed on PyTorch, where the CLIP model is frozen during training. The training time is about 40 hours for 20 epochs on two NVIDIA Tesla A100 GPUs. The optimizer is SGD with a cosine scheduler. We set the temperature $\lambda$ and $\tau$ to be 1 and 0.5, respectively. The sweep number is set to be 3 empirically. Besides, We adopt MinkowskiNet14 [19] as the backbone for evaluation on the ScanNet dataset, where the number of sweeps is set to be 1 and the training epochs is 30. As for the Switchable Self-Training Strategy, we randomly switch the point supervision signal after 10 epochs. We apply several data augmentations in contrastive learning, including random rotation along the z-axis and random flip on the point cloud, random horizontal flip, and random crop-resize on the image. + +# 4.1. Annotation-free Semantic Segmentation + +After pre-training the network, we show the performance of the 3D network when it is not fine-tuned on any annotations (Table 2). As no previous method reports the 3D annotation-free segmentation performance, we compare our method with different setups (Table 3). In what follows, we describe the experimental settings and give insights into our method and the different settings. + +Settings. We conduct experiments on the nuScenes and ScanNet datasets to evaluate the annotation-free semantic segmentation performance. Following MaskCLIP [61], we place the class name into 85 hand-craft prompts and feed it into the CLIP's text encoder to produce multiple text features. We then average the text features and feed the averaged features to the classifier for point-wise prediction. Besides, to explore how to effectively transfer CLIP's knowledge to the 3D network for annotation-free segmentation, We conduct + +Table 3. Ablation study on the nuScenes [24] val set for annotation-free 3D semantic segmentation. + +
Ablation TargetSettingmIoU (%)
-Baseline15.1
PromptsnuScenes15.1 (+0.0)
SemanticKITTI13.9 (-1.2)
Cityscapes11.3 (-3.8)
All15.3 (+0.2)
Regularizationw/o StCR19.8 (+4.7)
w/o SCR16.8 (+1.7)
KL0.0 (-15.1)
Training Strategyw/o S318.8 (+3.7)
ST10.1 (-4.0)
Sweeps1 sweep18.7 (+3.6)
3 sweeps20.8 (+5.7)
5 sweeps20.6 (+5.5)
merged18.6 (+3.5)
Full ConfigurationCLIP2Scene20.8 (+5.7)
+ +the following experiments to highlight the effectiveness of different modules in our framework. + +Baseline. The input of the 3D semantic segmentation network is only one sweep, and we pre-train the framework via semantic consistency regularization. + +Prompts (nuScenes, SemanticKITTI, Cityscapes, All). Based on the baseline, we respectively replace the nuScenes, SemanticKITTI, Cityscapes, and all class names into the prompts to produce the text embedding. + +Regularization (w/o StCR, w/o SCR, KL). Based on the full method, we remove the Spatial-temporal Consistency Regularization (w/o StCR) and remove the Semantic Consistency Regularization (w/o SCR). Besides, we abuse both StCR and SCR and distill the image feature to the point cloud by Kullback-Leibler (KL) divergence loss. + +Training Strategies (w/o S3, ST). We abuse the Switchable Self-Training Strategy (w/o S3) in the full method. Besides, we show the performance of only training the 3D network by their own predictions after ten epochs (ST). + +Sweeps Number (1 sweep, 3 sweeps, 5 sweeps, and merged). We set the sweep number $K$ to be 1, 3, and 5, respectively. Besides, we also take three sweeps of the point cloud as the input to pre-train the network (merged). + +Effect of Different Prompts. To verify how text embedding affects the performance, we generate various text embedding by the class name from different datasets (nuScenes, SemanticKITTI, and Cityscapes) and all classes for pre-training the framework. As shown in Table 3, we find that even learning with other datasets' text embedding (SemanticKITTI and Cityscapes), the 3D network could still recognize the nuScenes's objects with decent performance (13.9% and 11.3% mIoU, respectively). The result shows that the 3D network is capable of open-vocabulary recognition ability. + +![](images/89ffe74d74d82a6c97723e01d1374336ed7a758c47b237279664fa3a765805fd.jpg) +Figure 5. Qualitative results of annotation-free semantic segmentation on nuScenes dataset. Note that we show the results of individual classes. From the left to the right column are bus, motorcycle, car, and truck, respectively. The first row [ground truth] is the annotated semantic label. The second row [ours*] is our prediction of the highlighted target. The third row [ours] is our prediction of full classes. + +Effect of Semantic and Spatial-temporal Consistency Regularization. We remove Spatial-temporal Consistency Regularization (w/o SCR) from our method. Experiments show that the performance is dramatically decreased, indicating the effectiveness of our design. Besides, we also distill the image feature to the point cloud by KL divergence loss, where the text embeddings calculate the logits. However, such a method fails to transfer the semantic information from the image. The main reason is the noise-assigned semantics of the image pixel and the imperfect pixel-point correspondence due to the calibration error. + +Effect of Switchable Self-training Strategy. To examine the effect of the Switchable Self-Training Strategy, we either train the network with image supervision (w/o S3) or train the 3D network by their own predictions. Both trials witness a performance drop, indicating Switchable Self-Training Strategy is efficient in cross-modal self-supervised learning. + +Effect of Sweep Numbers. Intuitively, the performance of our method benefits from more sweeps information. Therefore, we also show the performance when restricting sweep size to 1, 3, and 5, respectively. However, we observe that the performance of 5 sweeps is similar to 3 sweeps but is more computationally expensive. Thus, we empirically set the sweep number to be 3. + +Qualitative Evaluation. The qualitative evaluations of individual classes (bus, motorcycle, car, and truck) are in Fig. 5, indicating that our method is able to perceive the objects even without training on any annotated data. However, we also observe the false positive predictions around the ground truth objects. We will resolve this issue in future work. + +# 4.2. Annotation-efficient Semantic Segmentation + +The pre-trained 3D network also boosts the performance when few labeled data are available for training. We directly compare SLidR [51], the only published method for image-to-Lidar self-supervised representation distillation. Besides, we also compared PPKT [44] and PointContrast [55]. In the following, we introduce SLidR and PPKT and compare them in detail. + +PPKT. PPKT is a cross-modal self-supervised method for the RGB-D dataset. It performs 2D-to-3D knowledge distillation via pixel-to-point contrastive loss. For a fair comparison, we use the same 3D network and training protocol but replace our semantic and Spatio-Temporal Regularization with InfoNCE loss. The framework is trained 50 epochs on 4096 randomly selected image-to-point pairs. + +SLidR. SLidR is an image-to-Lidar self-supervised method for autonomous driving data. Compared with PPKT, it introduces image super-pixel into cross-modal self-supervised learning. For a fair comparison, we replace our loss function with their superpixel-driven contrastive loss. + +Performance. As shown in Table 1, our method significantly outperforms the state-of-the-art methods when fine-tuned on $1\%$ and $100\%$ nuScenes dataset, with the improvement of $8.1\%$ and $1.1\%$ , respectively. Compared with the random initialization, the improvement is $14.1\%$ and $2.4\%$ , respectively, indicating the efficiency of our Semantic-driven Cross-modal Contrastive Learning framework. The qualitative results are shown in Fig. 6. Besides, we also verify the cross-domain generalization ability of our method. When pre-training the + +![](images/97ac7c1a4bf31eb9d085e0bde550b8b14bd6ecef4676192d80d18bcd4042f676.jpg) +Figure 6. Qualitative results of fine-tuning on $1\%$ nuScenes dataset. From the first row to the last row are the input LiDAR scan, ground truth, prediction of SLidR, and our prediction, respectively. Note that we show the results by error map, where the red point indicates the wrong prediction. Apparently, our method achieves decent performance. + +3D network on the nuScenes dataset and fine-tuning on $1\%$ and $100\%$ SemanticKITTI dataset, our method significantly outperforms other state-of-the-art self-supervised methods. + +Discussions. PPKT and SLidR reveal that contrastive loss is promising for transferring knowledge from image to point cloud. Like self-supervised learning, constructing the positive and negative samples is vital to unsupervised cross-modal knowledge distillation. However, previous methods suffer from optimization-conflict issues, i.e., some negative paired samples are actually positive pairs. For example, the road occupies a large proportion of the point cloud in a scene and is supposed to have the same semantics in the semantic segmentation task. When randomly selecting training samples, most negatively defined road-road points are actually positive. When feedforwarding such samples into contrastive learning, the contrastive loss will push them away in the embedding space, leading to unsatisfactory representation learning and hammering the downstream tasks' performance. SLidR introduces superpixel-driven contrastive learning to alleviate such issues. The motivation is that the visual representation of the image pixel and the projected points are consistent intra-superpixel. Although avoiding selecting negative image-point pairs from the same superpixel, the conflict still exists inter-superpixel. In our CLIP2Scene, we introduce the free-available dense pixel-text correspondence to alleviate the optimization conflicts. The text embedding represents the semantic information and can be used to select more reasonable training samples for contrastive learning. + +Besides training sample selection, the previous method also ignores the temporal coherence of the multi-sweep point cloud. That is, for LiDAR points mapping to the same image pixel, their feature is restricted to be consistent. Besides, considering the calibration error between the LiDAR scan and the camera image. We relax the pixel-to-point mapping to image grid-to-point grid mapping for consistency regularization. To this end, our Spatial-temporal consistency regularization leads to a more rational point representation. + +Last but not least, we find that randomly switching the supervision signal benefits self-supervised learning. Essentially, different modality networks learn different feature representations. They can filter different types of errors introduced by noisy pseudo labels. By this switchable operation, the error flows can be reduced mutually. + +# 5. Conclusion + +We explored how CLIP knowledge benefits 3D scene understanding in this work, termed CLIP2Scene. To efficiently transfer CLIP's image and text features to a 3D network, we propose a novel Semantic-driven Cross-modal Contrastive Learning framework including Semantic Regularization and Spatial-Temporal Regularization. For the first time, our pre-trained 3D network achieves annotation-free 3D semantic segmentation with decent performance. Besides, our method significantly outperforms state-of-the-art self-supervised methods when fine-tuning with labelled data. + +# References + +[1] Zeynep Akata, Florent Perronnin, Zaid Harchaoui, and Cordelia Schmid. Label-embedding for attribute-based classification. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 819–826, 2013. 2 +[2] Zeynep Akata, Scott E. Reed, Daniel Walter, Honglak Lee, and Bernt Schiele. Evaluation of output embeddings for fine-grained image classification. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2927-2936, 2015. 2 +[3] Jens Behley, Martin Garbade, Andres Milioto, Jan Quenzel, Sven Behnke, Cyril Stachniss, and Jurgen Gall. Semantickitti: A dataset for semantic scene understanding of lidar sequences. In IEEE/CVF International Conference on Computer Vision, pages 9297-9307, 2019. 2, 5, 6 +[4] Max Bucher, Stéphane Herbin, and Frédéric Jurie. Generating visual representations for zero-shot classification. In IEEE/CVF International Conference on Computer Vision Workshop, pages 2666-2673, 2017. 2 +[5] Holger Caesar, Varun Bankiti, Alex H Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giacarlo Baldan, and Oscar Beijbom. nuscenes: A multimodal dataset for autonomous driving. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11621-11631, 2020. 5 +[6] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In European Conference on Computer Vision, pages 213-229, 2020. 3 +[7] Mathilde Caron, Ishan Misra, Julien Mairal, Priya Goyal, Piotr Bojanowski, and Armand Joulin. Unsupervised learning of visual features by contrasting cluster assignments. In Advances in Neural Information Processing Systems, volume 33, pages 9912-9924, 2020. 2 +[8] Soravit Changpinyo, Wei-Lun Chao, Boqing Gong, and Fei Sha. Synthesized classifiers for zero-shot learning. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5327-5336, 2016. 2 +[9] Nenglun Chen, Lingjie Liu, Zhiming Cui, Runnan Chen, Duygu Ceylan, Changhe Tu, and Wenping Wang. Unsupervised learning of intrinsic structural representation points. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9121-9130, 2020. 2 +[10] Runnan Chen, Zhu Xinge, Nenglun Chen, Dawei Wang, Wei Li, Yuexin Ma, Ruigang Yang, and Wenping Wang. Referring self-supervised learning on 3d point cloud. 2021. 2 +[11] Runnan Chen, Xinge Zhu, Nenglun Chen, Wei Li, Yuexin Ma, Ruigang Yang, and Wenping Wang. Zero-shot point cloud segmentation by transferring geometric primitives. arXiv preprint arXiv:2210.09923, 2022. 1, 2 +[12] Runnan Chen, Xinge Zhu, Nenglun Chen, Dawei Wang, Wei Li, Yuexin Ma, Ruigang Yang, and Wenping Wang. Towards 3d scene understanding by referring synthetic models. arXiv preprint arXiv:2203.10546, 2022. 2 +[13] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning + +of visual representations. In International Conference on Machine Learning, pages 1597-1607, 2020. 2 +[14] Xinlei Chen and Kaiming He. Exploring simple siamese representation learning. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15750-15758, 2021. 2 +[15] Ran Cheng, Ryan Razani, Ehsan Taghavi, Enxu Li, and Bingbing Liu. (af)2-s3net: Attentive feature fusion with adaptive feature selection for sparse semantic segmentation network. In IEEE Conference on Computer Vision and Pattern Recognition, pages 12547-12556, 2021. 1 +[16] Ali Cheraghian, Shafin Rahman, Dylan Campbell, and Lars Petersson. Mitigating the Hubness Problem for Zero-Shot Learning of 3D Objects. arXiv preprint arXiv:1907.06371, 2019.2 +[17] Ali Cheraghian, Shafin Rahman, Townim F Chowdhury, Dylan Campbell, and Lars Petersson. Zero-shot learning on 3d point cloud objects and beyond. International Journal of Computer Vision, 130(10):2364-2384, 2022. 2 +[18] Ali Cheraghian, Shafin Rahman, and Lars Petersson. Zero-Shot Learning of 3D Point Cloud Objects. In IEEE International Conference on Machine Vision Applications, pages 1-6, 2019. 2 +[19] Christopher Choy, JunYoung Gwak, and Silvio Savarese. 4d spatio-temporal convnets: Minkowski convolutional neural networks. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3075-3084, 2019. 6 +[20] Angela Dai, Angel Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5828-5839, 2017. 2, 5, 6 +[21] Berkan Demirel, Ramazan Gokberk Cinbis, and Nazli Ikizler-Cinbis. Attributes2classname: A discriminative model for attribute-based unsupervised zero-shot learning. In IEEE/CVF International Conference on Computer Vision, pages 1241-1250, 2017. 2 +[22] Carl Doersch, Abhinav Gupta, and Alexei A Efros. Unsupervised visual representation learning by context prediction. In IEEE/CVF International Conference on Computer Vision, pages 1422-1430, 2015. 2 +[23] Alexey Dosovitskiy, Jost Tobias Springenberg, Martin Ried-miller, and Thomas Brox. Discriminative unsupervised feature learning with convolutional neural networks. In Advances in Neural Information Processing Systems, volume 27, 2014. 2 +[24] Whye Kit Fong, Rohit Mohan, Juana Valeria Hurtado, Lubing Zhou, Holger Caesar, Oscar Beijbom, and Abhinav Valada. Panoptic nuscenes: A large-scale benchmark for lidar panoptic segmentation and tracking. IEEE Robotics and Automation Letters, 7:3795-3802, 2022. 2, 5, 6 +[25] Chuang Gan, Ming Lin, Yi Yang, Yueting Zhuang, and Alexander Hauptmann. Exploring semantic inter-class relationships (sir) for zero-shot action recognition. In AAAI Conference on Artificial Intelligence, 2015. 2 +[26] Biao Gao, Yancheng Pan, Chengkun Li, Sibo Geng, and Huijing Zhao. Are we hungry for 3d lidar data for semantic segmentation? a survey of datasets and methods. IEEE Transactions on Intelligent Transportation Systems, 2021. 1 + +[27] Jean-Bastien Grill, Florian Strub, Florent Altché, Corentin Tallec, Pierre Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Avila Pires, Zhaohan Guo, Mohammad Gheshlaghi Azar, et al. Bootstrap your own latent-a new approach to self-supervised learning. In Advances in Neural Information Processing Systems, volume 33, pages 21271-21284, 2020. 2 +[28] Yulan Guo, Hanyun Wang, Qingyong Hu, Hao Liu, Li Liu, and Mohammed Bennamoun. Deep learning for 3d point clouds: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence, 43(12):4338-4364, 2020. 1 +[29] Bo Han, Quanming Yao, Xingrui Yu, Gang Niu, Miao Xu, Weihua Hu, Ivor Tsang, and Masashi Sugiyama. Co-teaching: Robust training of deep neural networks with extremely noisy labels. Advances in neural information processing systems, 31, 2018. 5 +[30] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9729-9738, 2020. 2 +[31] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 770-778, 2016. 3 +[32] Fangzhou Hong, Lingdong Kong, Hui Zhou, Xinge Zhu, Hongsheng Li, and Ziwei Liu. Unified 3d and 4d panoptic segmentation via dynamic shifting network. arXiv preprint arXiv:2203.07186, 2022. 1 +[33] Yuenan Hou, Xinge Zhu, Yuexin Ma, Chen Change Loy, and Yikang Li. Point-to-Voxel Knowledge Distillation for LiDAR Semantic Segmentation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8479-8488, 2022. 1 +[34] Siyuan Huang, Yichen Xie, Song-Chun Zhu, and Yixin Zhu. Spatio-temporal self-supervised representation learning for 3d point clouds. In IEEE/CVF International Conference on Computer Vision, pages 6535-6545, 2021. 2 +[35] Ajay Jain, Matthew Tancik, and Pieter Abbeel. Putting nerf on a diet: Semantically consistent few-shot view synthesis. In IEEE/CVF International Conference on Computer Vision, pages 5885-5894, 2021. 2 +[36] Sosuke Kobayashi, Eiichi Matsumoto, and Vincent Sitzmann. Decomposing nef for editing via feature field distillation. arXiv preprint arXiv:2205.15585, 2022. 2 +[37] Elyor Kodirov, Tao Xiang, and Shaogang Gong. Semantic autoencoder for zero-shot learning. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4447-4456, 2017. 2 +[38] Lingdong Kong, Youquan Liu, Runnan Chen, Yuexin Ma, Xinge Zhu, Yikang Li, Yuenan Hou, Yu Qiao, and Ziwei Liu. Rethinking range view representation for lidar segmentation. arXiv preprint arXiv:2303.05367, 2023. 1 +[39] Lingdong Kong, Niamul Quader, and Venice Erin Liong. Conda: Unsupervised domain adaptation for lidar segmentation via regularized domain concatenation. In IEEE International Conference on Robotics and Automation, 2023. 1 +[40] Lingdong Kong, Jiawei Ren, Liang Pan, and Ziwei Liu. Lasermix for semi-supervised lidar semantic segmentation. + +In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023. 1 +[41] Christoph H. Lampert, Hannes Nickisch, and Stefan Harmeling. Learning to detect unseen object classes by between-class attribute transfer. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 951-958, 2009. 2 +[42] Christoph H. Lampert, Hannes Nickisch, and Stefan Harmeling. Attribute-based classification for zero-shot visual object categorization. IEEE Transactions on Pattern Analysis and Machine Intelligence, 36:453-465, 2014. 2 +[43] Yan Li, Zhen Jia, Junge Zhang, Kaiqi Huang, and Tieniu Tan. Deep semantic structural constraints for zero-shot learning. In AAAI Conference on Artificial Intelligence, 2018. 2 +[44] Yueh-Cheng Liu, Yu-Kai Huang, Hung-Yueh Chiang, Hung-Ting Su, Zhe-Yu Liu, Chin-Tang Chen, Ching-Yu Tseng, and Winston H Hsu. Learning from 2d: Contrastive pixel-topoint knowledge transfer for 3d pretraining. arXiv preprint arXiv:2104.04687, 2021. 1, 2, 3, 4, 6, 7 +[45] Björn Michele, Alexandre Boulch, Gilles Puy, Maxime Bucher, and Renaud Marlet. Generative zero-shot learning for semantic segmentation of 3d point clouds. In International Conference on 3D Vision, pages 992-1002, 2021. 1, 2 +[46] Ashish Mishra, M. Shiva Krishna Reddy, Anurag Mittal, and Hema A. Murthy. A generative model for zero shot learning using conditional variational autoencoders. In IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshop, pages 2269-22698, 2018. 2 +[47] Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. Pointnet: Deep learning on point sets for 3d classification and segmentation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 652-660, 2017. 1, 2 +[48] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 1 +[49] Yongming Rao, Wenliang Zhao, Guangyi Chen, Yansong Tang, Zheng Zhu, Guan Huang, Jie Zhou, and Jiwen Lu. Denseclip: Language-guided dense prediction with context-aware prompting. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18082-18091, 2022. 2 +[50] David Rozenberszki, Or Litany, and Angela Dai. Language-grounded indoor 3d semantic segmentation in the wild. In European Conference on Computer Vision, pages 125-141, 2022. 2 +[51] Corentin Sautier, Gilles Puy, Spyros Gidaris, Alexandre Boulch, Andrei Bursuc, and Renaud Marlet. Image-to-lidar self-supervised distillation for autonomous driving data. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9891-9901, 2022. 1, 2, 3, 4, 5, 6, 7 +[52] Haotian Tang, Zhijian Liu, Shengyu Zhao, Yujun Lin, Ji Lin, Hanrui Wang, and Song Han. Searching efficient 3d architectures with sparse point-voxel convolution. In European Conference on Computer Vision, pages 685–702, 2020. 6 +[53] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia + +Polosukhin. Attention is all you need. In Advances in Neural Information Processing Systems, volume 30, 2017. 3 +[54] Yongqin Xian, Zeynep Akata, Gaurav Sharma, Quynh N. Nguyen, Matthias Hein, and Bernt Schiele. Latent embeddings for zero-shot classification. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 69-77, 2016. 2 +[55] Saining Xie, Jiatao Gu, Demi Guo, Charles R Qi, Leonidas Guibas, and Or Litany. Pointcontrast: Unsupervised pretraining for 3d point cloud understanding. In European Conference on Computer Vision, pages 574-591, 2020. 2, 6, 7 +[56] Jianyun Xu, Ruixiang Zhang, Jian Dou, Yushi Zhu, Jie Sun, and Shiliang Pu. Rpvnet: A deep and efficient range-point-voxel fusion network for lidar point cloud segmentation. In IEEE/CVF International Conference on Computer Vision, pages 16024-16033, 2021. 1 +[57] Lewei Yao, Jianhua Han, Youpeng Wen, Xiaodan Liang, Dan Xu, Wei Zhang, Zhenguo Li, Chunjing Xu, and Hang Xu. Detclip: Dictionary-enriched visual-concept paralleled pre-training for open-world detection. arXiv preprint arXiv:2209.09407, 2022. 2, 3 +[58] Éloi Zablocki, Patrick Bordes, Benjamin Piwowarski, Laure Soulier, and Patrick Gallinari. Context-aware zero-shot learning for object recognition. In International Conference on Machine Learning, pages 7292-7303, 2019. 2 +[59] Renrui Zhang, Ziyu Guo, Wei Zhang, Kunchang Li, Xupeng Miao, Bin Cui, Yu Qiao, Peng Gao, and Hongsheng Li. Pointclip: Point cloud understanding by clip. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8552-8562, 2022. 2 +[60] Zaiwei Zhang, Rohit Girdhar, Armand Joulin, and Ishan Misra. Self-supervised pretraining of 3d features on any point-cloud. In IEEE/CVF International Conference on Computer Vision, pages 10252-10263, 2021. 2 +[61] Chong Zhou, Chen Change Loy, and Bo Dai. Extract free dense labels from clip. In European Conference on Computer Vision, pages 696-712, 2022. 2, 4, 6 +[62] Xinge Zhu, Hui Zhou, Tai Wang, Fangzhou Hong, Yuexin Ma, Wei Li, Hongsheng Li, and Dahua Lin. Cylindrical and asymmetrical 3d convolution networks for lidar segmentation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9939-9948, 2021. 1 \ No newline at end of file diff --git a/2023/CLIP2Scene_ Towards Label-Efficient 3D Scene Understanding by CLIP/images.zip b/2023/CLIP2Scene_ Towards Label-Efficient 3D Scene Understanding by CLIP/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..66ef5902fc135d7353c4c8f0b35c0cf46d0b9281 --- /dev/null +++ b/2023/CLIP2Scene_ Towards Label-Efficient 3D Scene Understanding by CLIP/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b6b145aee458de77fab1725cd0d05515cea27c7c9a960a6c3ec196637ad4ad67 +size 634590 diff --git a/2023/CLIP2Scene_ Towards Label-Efficient 3D Scene Understanding by CLIP/layout.json b/2023/CLIP2Scene_ Towards Label-Efficient 3D Scene Understanding by CLIP/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..1e5d75072ada0f318d7eea2d04430dd7ceb2683a --- /dev/null +++ b/2023/CLIP2Scene_ Towards Label-Efficient 3D Scene Understanding by CLIP/layout.json @@ -0,0 +1,8088 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 74, + 103, + 520, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 74, + 103, + 520, + 121 + ], + "spans": [ + { + "bbox": [ + 74, + 103, + 520, + 121 + ], + "type": "text", + "content": "CLIP2Scene: Towards Label-efficient 3D Scene Understanding by CLIP" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 141, + 538, + 172 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 141, + 538, + 172 + ], + "spans": [ + { + "bbox": [ + 56, + 141, + 538, + 172 + ], + "type": "text", + "content": "Runnan Chen" + }, + { + "bbox": [ + 56, + 141, + 538, + 172 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 56, + 141, + 538, + 172 + ], + "type": "text", + "content": " Youquan Liu" + }, + { + "bbox": [ + 56, + 141, + 538, + 172 + ], + "type": "inline_equation", + "content": "^{2,3}" + }, + { + "bbox": [ + 56, + 141, + 538, + 172 + ], + "type": "text", + "content": " Lingdong Kong" + }, + { + "bbox": [ + 56, + 141, + 538, + 172 + ], + "type": "inline_equation", + "content": "^{2,4}" + }, + { + "bbox": [ + 56, + 141, + 538, + 172 + ], + "type": "text", + "content": " Xinge Zhu" + }, + { + "bbox": [ + 56, + 141, + 538, + 172 + ], + "type": "inline_equation", + "content": "^{5}" + }, + { + "bbox": [ + 56, + 141, + 538, + 172 + ], + "type": "text", + "content": " Yuexin Ma" + }, + { + "bbox": [ + 56, + 141, + 538, + 172 + ], + "type": "inline_equation", + "content": "^{6}" + }, + { + "bbox": [ + 56, + 141, + 538, + 172 + ], + "type": "text", + "content": " Yikang Li" + }, + { + "bbox": [ + 56, + 141, + 538, + 172 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 56, + 141, + 538, + 172 + ], + "type": "text", + "content": " Yuenan Hou" + }, + { + "bbox": [ + 56, + 141, + 538, + 172 + ], + "type": "inline_equation", + "content": "^{2,\\dagger}" + }, + { + "bbox": [ + 56, + 141, + 538, + 172 + ], + "type": "text", + "content": " Yu Qiao" + }, + { + "bbox": [ + 56, + 141, + 538, + 172 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 56, + 141, + 538, + 172 + ], + "type": "text", + "content": " Wenping Wang" + }, + { + "bbox": [ + 56, + 141, + 538, + 172 + ], + "type": "inline_equation", + "content": "^{7,\\dagger}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 67, + 174, + 527, + 201 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 174, + 527, + 201 + ], + "spans": [ + { + "bbox": [ + 67, + 174, + 527, + 201 + ], + "type": "text", + "content": "1The University of Hong Kong 2Shanghai AI Laboratory 3Hochschule Bremerhaven 4National University of Singapore 5The Chinese University of Hong Kong 6ShanghaiTech University 7Texas A&M University" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 143, + 240, + 192, + 252 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 240, + 192, + 252 + ], + "spans": [ + { + "bbox": [ + 143, + 240, + 192, + 252 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 45, + 262, + 290, + 622 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 262, + 290, + 622 + ], + "spans": [ + { + "bbox": [ + 45, + 262, + 290, + 622 + ], + "type": "text", + "content": "Contrastive Language-Image Pre-training (CLIP) achieves promising results in 2D zero-shot and few-shot learning. Despite the impressive performance in 2D, applying CLIP to help the learning in 3D scene understanding has yet to be explored. In this paper, we make the first attempt to investigate how CLIP knowledge benefits 3D scene understanding. We propose CLIP2Scene, a simple yet effective framework that transfers CLIP knowledge from 2D image-text pre-trained models to a 3D point cloud network. We show that the pre-trained 3D network yields impressive performance on various downstream tasks, i.e., annotation-free and fine-tuning with labelled data for semantic segmentation. Specifically, built upon CLIP, we design a Semantic-driven Cross-modal Contrastive Learning framework that pre-trains a 3D network via semantic and spatial-temporal consistency regularization. For the former, we first leverage CLIP's text semantics to select the positive and negative point samples and then employ the contrastive loss to train the 3D network. In terms of the latter, we force the consistency between the temporally coherent point cloud features and their corresponding image features. We conduct experiments on SemanticKITTI, nuScenes, and ScanNet. For the first time, our pre-trained network achieves annotation-free 3D semantic segmentation with " + }, + { + "bbox": [ + 45, + 262, + 290, + 622 + ], + "type": "inline_equation", + "content": "20.8\\%" + }, + { + "bbox": [ + 45, + 262, + 290, + 622 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 45, + 262, + 290, + 622 + ], + "type": "inline_equation", + "content": "25.08\\%" + }, + { + "bbox": [ + 45, + 262, + 290, + 622 + ], + "type": "text", + "content": " mIoU on nuScenes and ScanNet, respectively. When fine-tuned with " + }, + { + "bbox": [ + 45, + 262, + 290, + 622 + ], + "type": "inline_equation", + "content": "1\\%" + }, + { + "bbox": [ + 45, + 262, + 290, + 622 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 45, + 262, + 290, + 622 + ], + "type": "inline_equation", + "content": "100\\%" + }, + { + "bbox": [ + 45, + 262, + 290, + 622 + ], + "type": "text", + "content": " labelled data, our method significantly outperforms other self-supervised methods, with improvements of " + }, + { + "bbox": [ + 45, + 262, + 290, + 622 + ], + "type": "inline_equation", + "content": "8\\%" + }, + { + "bbox": [ + 45, + 262, + 290, + 622 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 45, + 262, + 290, + 622 + ], + "type": "inline_equation", + "content": "1\\%" + }, + { + "bbox": [ + 45, + 262, + 290, + 622 + ], + "type": "text", + "content": " mIoU, respectively. Furthermore, we demonstrate the generalizability for handling cross-domain datasets. Code is publicly available1." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 641, + 128, + 654 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 641, + 128, + 654 + ], + "spans": [ + { + "bbox": [ + 48, + 641, + 128, + 654 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 661, + 288, + 686 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 661, + 288, + 686 + ], + "spans": [ + { + "bbox": [ + 47, + 661, + 288, + 686 + ], + "type": "text", + "content": "3D scene understanding is fundamental in autonomous driving, robot navigation, etc [26,28]. Current deep learning-" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 310, + 247, + 537, + 424 + ], + "blocks": [ + { + "bbox": [ + 323, + 233, + 536, + 245 + ], + "lines": [ + { + "bbox": [ + 323, + 233, + 536, + 245 + ], + "spans": [ + { + "bbox": [ + 323, + 233, + 536, + 245 + ], + "type": "text", + "content": "Semantic-driven Cross-modal Contrastive Learning" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 310, + 247, + 537, + 424 + ], + "lines": [ + { + "bbox": [ + 310, + 247, + 537, + 424 + ], + "spans": [ + { + "bbox": [ + 310, + 247, + 537, + 424 + ], + "type": "image", + "image_path": "e635707803576d3ce4da203fd52c9028e20daf537b797f5ed1eded7a308ab370.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 430, + 548, + 519 + ], + "lines": [ + { + "bbox": [ + 305, + 430, + 548, + 519 + ], + "spans": [ + { + "bbox": [ + 305, + 430, + 548, + 519 + ], + "type": "text", + "content": "Figure 1. We explore how CLIP knowledge benefits 3D scene understanding. To this end, we propose CLIP2Scene, a Semantic-driven Cross-modal Contrastive Learning framework that leverages CLIP knowledge to pre-train a 3D point cloud segmentation network via semantic and spatial-temporal consistency regularization. CLIP2Scene yields impressive performance on annotation-free 3D semantic segmentation and significantly outperforms other self-supervised methods when fine-tuning on annotated data." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 542, + 547, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 542, + 547, + 662 + ], + "spans": [ + { + "bbox": [ + 304, + 542, + 547, + 662 + ], + "type": "text", + "content": "based methods have shown inspirational performance on 3D point cloud data [15, 32, 33, 38, 47, 56, 62]. However, some drawbacks hinder their real-world applications. The first one comes from their heavy reliance on the large collection of annotated point clouds, especially when high-quality 3D annotations are expensive to acquire [39, 40, 44, 51]. Besides, they typically fail to recognize novel objects that are never seen in the training data [11, 45]. As a result, it may need extra annotation efforts to train the model on recognizing these novel objects, which is both tedious and time-consuming." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 665, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 547, + 715 + ], + "type": "text", + "content": "Contrastive Vision-Language Pre-training (CLIP) [48] provides a new perspective that mitigates the above issues in 2D vision. It was trained on large-scale free-available image-text pairs from websites and built vision-language" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 693, + 259, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 693, + 259, + 713 + ], + "spans": [ + { + "bbox": [ + 58, + 693, + 259, + 713 + ], + "type": "text", + "content": "Symbol " + }, + { + "bbox": [ + 58, + 693, + 259, + 713 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 58, + 693, + 259, + 713 + ], + "type": "text", + "content": " denotes the corresponding authors. \n1https://github.com/runnanchen/CLIP2Scene." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7020" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 240 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 240 + ], + "type": "text", + "content": "correlation to achieve promising open-vocabulary recognition. MaskCLIP [61] further explores semantic segmentation based on CLIP. With minimal modifications to the CLIP pre-trained network, MaskCLIP can be directly used for the semantic segmentation of novel objects without additional training efforts. PointCLIP [59] reveals that the zero-shot classification ability of CLIP can be generalized from the 2D image to the 3D point cloud. It perspectively projects a point cloud frame into different views of 2D depth maps that bridge the modal gap between the image and the point cloud. The above studies indicate the potential of CLIP on enhancing the 2D segmentation and 3D classification performance. However, whether and how CLIP knowledge benefits 3D scene understanding is still under-explored." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 243, + 289, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 243, + 289, + 506 + ], + "spans": [ + { + "bbox": [ + 46, + 243, + 289, + 506 + ], + "type": "text", + "content": "In this paper, we explore how to leverage CLIP's 2D image-text pre-learned knowledge for 3D scene understanding. Previous cross-modal knowledge distillation methods [44,51] suffer from the optimization-conflict issue, i.e., some of the positive pairs are regarded as negative samples for contrastive learning, leading to unsatisfactory representation learning and hammering the performance of downstream tasks. Besides, they also ignore the temporal coherence of the multi-sweep point cloud, failing to utilize the rich inter-sweep correspondence. To handle the mentioned problems, we propose a novel Semantic-driven Cross-modal Contrastive Learning framework that fully leverages CLIP's semantic and visual information to regularize a 3D network. Specifically, we propose Semantic Consistency Regularization and Spatial-Temporal Consistency Regularization. In semantic consistency regularization, we utilize CLIP's text semantics to select the positive and negative point samples for less-conflict contrastive learning. For spatial-temporal consistency regularization, we take CLIP's image pixel feature to impose a soft consistency constraint on the temporally coherent point features. Such an operation also alleviates the effects of imperfect image-to-point calibration." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 510, + 289, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 510, + 289, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 510, + 289, + 714 + ], + "type": "text", + "content": "We conduct several downstream tasks on the indoor and outdoor datasets to verify how the pre-trained network benefits the 3D scene understanding. The first one is annotation-free semantic segmentation. Following MaskCLIP, we place class names into multiple hand-crafted templates as prompts and average the text embeddings generated by CLIP to conduct the annotation-free segmentation. For the first time, our method achieves " + }, + { + "bbox": [ + 46, + 510, + 289, + 714 + ], + "type": "inline_equation", + "content": "20.8\\%" + }, + { + "bbox": [ + 46, + 510, + 289, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 510, + 289, + 714 + ], + "type": "inline_equation", + "content": "25.08\\%" + }, + { + "bbox": [ + 46, + 510, + 289, + 714 + ], + "type": "text", + "content": " mIoU annotation-free 3D semantic segmentation on the nuScenes [24] and ScanNet [20] datasets without training on any labelled data. Secondly, we compare with other self-supervised methods in label-efficient learning. When fine-tuning the 3D network with " + }, + { + "bbox": [ + 46, + 510, + 289, + 714 + ], + "type": "inline_equation", + "content": "1\\%" + }, + { + "bbox": [ + 46, + 510, + 289, + 714 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 46, + 510, + 289, + 714 + ], + "type": "inline_equation", + "content": "100\\%" + }, + { + "bbox": [ + 46, + 510, + 289, + 714 + ], + "type": "text", + "content": " labelled data on the nuScenes dataset, our method significantly outperforms state-of-the-art self-supervised methods, with improvements of " + }, + { + "bbox": [ + 46, + 510, + 289, + 714 + ], + "type": "inline_equation", + "content": "8\\%" + }, + { + "bbox": [ + 46, + 510, + 289, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 510, + 289, + 714 + ], + "type": "inline_equation", + "content": "1\\%" + }, + { + "bbox": [ + 46, + 510, + 289, + 714 + ], + "type": "text", + "content": " mIoU, respectively. Besides, to verify the generalization capability, we pre-train the network on the nuScenes dataset" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 304, + 72, + 545, + 109 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 109 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 109 + ], + "type": "text", + "content": "and evaluate it on SemanticKITTI [3]. Our method still significantly outperforms state-of-the-art methods. The key contributions of our work are summarized as follows." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 317, + 116, + 547, + 311 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 317, + 116, + 545, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 116, + 545, + 140 + ], + "spans": [ + { + "bbox": [ + 317, + 116, + 545, + 140 + ], + "type": "text", + "content": "- The first work that distils CLIP knowledge to a 3D network for 3D scene understanding." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 317, + 149, + 547, + 196 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 149, + 547, + 196 + ], + "spans": [ + { + "bbox": [ + 317, + 149, + 547, + 196 + ], + "type": "text", + "content": "- We propose a novel Semantic-driven Cross-modal Contrastive Learning framework that pre-trains a 3D network via spatial-temporal and semantic consistency regularization." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 317, + 206, + 545, + 254 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 206, + 545, + 254 + ], + "spans": [ + { + "bbox": [ + 317, + 206, + 545, + 254 + ], + "type": "text", + "content": "- We propose a novel Semantic-guided Spatial-Temporal Consistency Regularization that forces the consistency between the temporally coherent point cloud features and their corresponding image features." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 317, + 262, + 547, + 311 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 262, + 547, + 311 + ], + "spans": [ + { + "bbox": [ + 317, + 262, + 547, + 311 + ], + "type": "text", + "content": "- For the first time, our method achieves promising results on annotation-free 3D scene segmentation. When fine-tuning with labelled data, our method significantly outperforms state-of-the-art self-supervised methods." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 322, + 392, + 335 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 322, + 392, + 335 + ], + "spans": [ + { + "bbox": [ + 306, + 322, + 392, + 335 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 342, + 547, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 342, + 547, + 617 + ], + "spans": [ + { + "bbox": [ + 304, + 342, + 547, + 617 + ], + "type": "text", + "content": "Zero-shot Learning in 3D. The objective of zero-shot learning (ZSL) is to recognize objects that are unseen in the training set. Many efforts have been devoted to the 2D recognition tasks [1, 2, 4, 8, 21, 25, 34, 37, 41–43, 46, 54, 55, 58, 60], and few works concentrate on performing ZSL in the 3D domain [11, 16–18, 45]. [18] applies ZSL to 3D tasks, where they train PointNet [47] on \"seen\" samples and test on \"unseen\" samples. Subsequent work [16] addresses the hubness problem caused by the low-quality point cloud features. [17] proposes the triplet loss to boost the performance under the transductive setting, where the \"unseen\" class is observed and unlabeled in the training phase. [11] makes the first attempt to explore the transductive zero-shot segmentation for 3D scene understanding. Recently, some studies introduced CLIP into zero-shot learning. MaskCLIP [61] investigates the problem of utilizing CLIP to help the 2D dense prediction tasks and exhibits encouraging zero-shot semantic segmentation performance. PointCLIP [59] is the pioneering work that applies CLIP to 3D recognition and shows impressive performance on zero-shot and few-shot classification tasks. Our work takes a step further to investigate how the rich semantic and visual knowledge in CLIP can benefit the 3D semantic segmentation tasks." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 618, + 548, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 618, + 548, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 618, + 548, + 715 + ], + "type": "text", + "content": "Self-supervised Representation Learning. The purpose of self-supervised learning is to obtain a good representation that benefits the downstream tasks. The dominant approaches resort to contrastive learning to pre-train the network [7, 9, 9, 10, 12-14, 22, 23, 27, 30]. Recently, inspired by the success of CLIP, leveraging the pre-trained model of CLIP to the downstream tasks has raised the community's attention [35, 36, 49, 50, 57]. DenseCLIP [49] utilizes" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "7021" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 55, + 73, + 537, + 299 + ], + "blocks": [ + { + "bbox": [ + 55, + 73, + 537, + 299 + ], + "lines": [ + { + "bbox": [ + 55, + 73, + 537, + 299 + ], + "spans": [ + { + "bbox": [ + 55, + 73, + 537, + 299 + ], + "type": "image", + "image_path": "0c613f286b55db46731f70fd01a88b176c9d596b00fc0a3d0db65877657647eb.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 303, + 548, + 384 + ], + "lines": [ + { + "bbox": [ + 46, + 303, + 548, + 384 + ], + "spans": [ + { + "bbox": [ + 46, + 303, + 548, + 384 + ], + "type": "text", + "content": "Figure 2. Illustration of the Semantic-driven Cross-modal Contrastive Learning. Firstly, we obtain the text embedding " + }, + { + "bbox": [ + 46, + 303, + 548, + 384 + ], + "type": "inline_equation", + "content": "t_i" + }, + { + "bbox": [ + 46, + 303, + 548, + 384 + ], + "type": "text", + "content": ", image pixel feature " + }, + { + "bbox": [ + 46, + 303, + 548, + 384 + ], + "type": "inline_equation", + "content": "x_i" + }, + { + "bbox": [ + 46, + 303, + 548, + 384 + ], + "type": "text", + "content": ", and point feature " + }, + { + "bbox": [ + 46, + 303, + 548, + 384 + ], + "type": "inline_equation", + "content": "p_i" + }, + { + "bbox": [ + 46, + 303, + 548, + 384 + ], + "type": "text", + "content": " by text encoder, image encoder, and point encoder, respectively. Secondly, we leverage CLIP knowledge to construct positive and negative samples for contrastive learning. Thus we obtain point-text pairs " + }, + { + "bbox": [ + 46, + 303, + 548, + 384 + ], + "type": "inline_equation", + "content": "\\{x_i, t_i\\}_{i=1}^M" + }, + { + "bbox": [ + 46, + 303, + 548, + 384 + ], + "type": "text", + "content": " and all pixel-point-text pairs in a short temporal " + }, + { + "bbox": [ + 46, + 303, + 548, + 384 + ], + "type": "inline_equation", + "content": "\\{\\hat{x}_i^k, \\hat{p}_i^k, t_i^k\\}_{i=1,k=1}^M" + }, + { + "bbox": [ + 46, + 303, + 548, + 384 + ], + "type": "text", + "content": ". Here, " + }, + { + "bbox": [ + 46, + 303, + 548, + 384 + ], + "type": "inline_equation", + "content": "\\{x_i, t_i\\}_{i=1}^M" + }, + { + "bbox": [ + 46, + 303, + 548, + 384 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 303, + 548, + 384 + ], + "type": "inline_equation", + "content": "\\{\\hat{x}_i^k, \\hat{p}_i^k, t_i^k\\}_{i=1,k=1}^M" + }, + { + "bbox": [ + 46, + 303, + 548, + 384 + ], + "type": "text", + "content": " are used for Semantic Consistency Regularization and Spatial-Temporal Consistency Regularization, respectively. Lastly, we perform Semantic Consistency Regularization by pulling the point features to their corresponding text embedding and Spatial-Temporal Consistency Regularization by mimicking the temporally coherent point features to their corresponding pixel features." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 403, + 288, + 463 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 403, + 288, + 463 + ], + "spans": [ + { + "bbox": [ + 46, + 403, + 288, + 463 + ], + "type": "text", + "content": "the CLIP's pre-trained knowledge for dense image pixel prediction. DetCLIP [57] proposes a pre-training method equipped with CLIP for open-world detection. We leverage the image-text pre-trained CLIP knowledge to help 3D scene understanding." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 464, + 288, + 571 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 464, + 288, + 571 + ], + "spans": [ + { + "bbox": [ + 46, + 464, + 288, + 571 + ], + "type": "text", + "content": "Cross-modal Knowledge Distillation. Recently, increasing studies have focused on transferring knowledge from 2D images to 3D point clouds for self-supervised representation learning [44, 51]. PPKT [44] resorts to the InfoNCE loss to help the 3D network distil rich knowledge from the 2D image backbone. SLidR [51] further introduce the superpixel to boost the cross-modal knowledge distillation. In this paper, we first attempt to pre-train a 3D network with CLIP's knowledge." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 585, + 129, + 599 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 585, + 129, + 599 + ], + "spans": [ + { + "bbox": [ + 47, + 585, + 129, + 599 + ], + "type": "text", + "content": "3. Methodology" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "text", + "content": "Considering the impressive open-vocabulary performance achieved by CLIP in image classification and segmentation, natural curiosities have been raised. Can CLIP endow the ability to a 3D network for annotation-free scene understanding? And further, will it promote the network performance when fine-tuned on labelled data? To answer the above questions, we study the cross-modal knowledge transfer of CLIP for 3D scene understanding, termed CLIP2Scene. Our work is a pioneer in exploiting CLIP knowledge for 3D scene" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 403, + 547, + 477 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 403, + 547, + 477 + ], + "spans": [ + { + "bbox": [ + 304, + 403, + 547, + 477 + ], + "type": "text", + "content": "understanding. In what follows, we revisit the CLIP applied in 2D open-vocabulary classification and semantic segmentation, then present our CLIP2Scene in detail. Our approach consists of three major components: Semantic Consistency Regularization, Semantic-Guided Spatial-Temporal Consistency Regularization, and Switchable Self-Training Strategy." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 487, + 405, + 502 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 487, + 405, + 502 + ], + "spans": [ + { + "bbox": [ + 306, + 487, + 405, + 502 + ], + "type": "text", + "content": "3.1. Revisiting CLIP" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 508, + 547, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 508, + 547, + 688 + ], + "spans": [ + { + "bbox": [ + 304, + 508, + 547, + 688 + ], + "type": "text", + "content": "Contrastive Vision-Language Pre-training (CLIP) mitigates the following drawbacks that dominate the computer vision field: 1. Deep models need a large amount of formatted and labelled training data, which is expensive to acquire; 2. The model's generalization ability is weak, making it difficult to migrate to a new scenario with unseen objects. CLIP consists of an image encoder (ResNet [31] or ViT [6]) and a text encoder (Transformer [53]), both respectively project the image and text representation to a joint embedding space. During training, CLIP constructs positive and negative samples from 400 million image-text pairs to train both encoders with a contrastive loss, where the large-scale image-text pairs are free-available from the Internet and assumed to contain every class of images and most concepts of text. Therefore, CLIP can achieve promising open-vocabulary recognition." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 689, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 547, + 713 + ], + "type": "text", + "content": "For 2D zero-shot classification, CLIP first places the class name into a pre-defined template to generate the text embed-" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7022" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 75, + 290, + 223 + ], + "blocks": [ + { + "bbox": [ + 50, + 75, + 290, + 223 + ], + "lines": [ + { + "bbox": [ + 50, + 75, + 290, + 223 + ], + "spans": [ + { + "bbox": [ + 50, + 75, + 290, + 223 + ], + "type": "image", + "image_path": "abb87bad06a099b3a04ab1504989c71b187a5965afe07839adca112ec8c71c3c.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 49, + 224, + 284, + 342 + ], + "blocks": [ + { + "bbox": [ + 49, + 224, + 284, + 342 + ], + "lines": [ + { + "bbox": [ + 49, + 224, + 284, + 342 + ], + "spans": [ + { + "bbox": [ + 49, + 224, + 284, + 342 + ], + "type": "image", + "image_path": "df1d6e894bf9f8cd67b1d613b7c758a6cf027723e49a24df834de1fab5e2dd5d.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 348, + 287, + 381 + ], + "lines": [ + { + "bbox": [ + 46, + 348, + 287, + 381 + ], + "spans": [ + { + "bbox": [ + 46, + 348, + 287, + 381 + ], + "type": "text", + "content": "Figure 3. Illustration of the image pixel-to-text mapping. The dense pixel-text correspondence " + }, + { + "bbox": [ + 46, + 348, + 287, + 381 + ], + "type": "inline_equation", + "content": "\\{x_{i},t_{i}\\}_{i = 1}^{M}" + }, + { + "bbox": [ + 46, + 348, + 287, + 381 + ], + "type": "text", + "content": " is extracted by the off-the-shelf method MaskCLIP [61]." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 396, + 289, + 480 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 396, + 289, + 480 + ], + "spans": [ + { + "bbox": [ + 46, + 396, + 289, + 480 + ], + "type": "text", + "content": "dings and then encodes images to obtain image embeddings. Next, it calculates the similarities between images and text embeddings to determine the class. MaskCLIP further extends CLIP into 2D semantic segmentation. Specifically, MaskCLIP modifies the attention pooling layer of the CLIP's image encoder, thus performing pixel-level mask prediction instead of the global image-level prediction." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 489, + 128, + 501 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 489, + 128, + 501 + ], + "spans": [ + { + "bbox": [ + 47, + 489, + 128, + 501 + ], + "type": "text", + "content": "3.2. CLIP2Scene" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 509, + 289, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 509, + 289, + 628 + ], + "spans": [ + { + "bbox": [ + 46, + 509, + 289, + 628 + ], + "type": "text", + "content": "As shown in Fig. 2, we first leverage CLIP and 3D network to respectively extract the text embeddings, image pixel feature, and point feature. Secondly, we construct positive and negative samples based on CLIP's knowledge. Lastly, we impose Semantic Consistency Regularization by pulling the point features to their corresponding text embedding. At the same time, we apply Spatial-Temporal Consistency Regularization by forcing the consistency between temporally coherent point features and their corresponding pixel features. In what follows, we present the details and insights." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 646, + 239, + 658 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 646, + 239, + 658 + ], + "spans": [ + { + "bbox": [ + 47, + 646, + 239, + 658 + ], + "type": "text", + "content": "3.2.1 Semantic Consistency Regularization" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 665, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 288, + 713 + ], + "type": "text", + "content": "As CLIP is pre-trained on 2D images and text, our first concern is the domain gap between 2D images and the 3D point cloud. To this end, we build dense pixel-point correspondence and transfer image knowledge to the 3D point cloud" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 72, + 547, + 179 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 179 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 179 + ], + "type": "text", + "content": "via the pixel-point pairs. Specifically, we calibrate the Li-DAR point cloud with corresponding images captured by six cameras. Therefore, the dense pixel-point correspondence " + }, + { + "bbox": [ + 304, + 72, + 547, + 179 + ], + "type": "inline_equation", + "content": "\\{x_i,p_i\\}_{i = 1}^M" + }, + { + "bbox": [ + 304, + 72, + 547, + 179 + ], + "type": "text", + "content": " can be obtained accordingly, where " + }, + { + "bbox": [ + 304, + 72, + 547, + 179 + ], + "type": "inline_equation", + "content": "x_{i}" + }, + { + "bbox": [ + 304, + 72, + 547, + 179 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 72, + 547, + 179 + ], + "type": "inline_equation", + "content": "p_i" + }, + { + "bbox": [ + 304, + 72, + 547, + 179 + ], + "type": "text", + "content": " indicates " + }, + { + "bbox": [ + 304, + 72, + 547, + 179 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 72, + 547, + 179 + ], + "type": "text", + "content": "-th paired image feature and point feature, which are respectively extracted by the CLIP's image encoder and the 3D network. " + }, + { + "bbox": [ + 304, + 72, + 547, + 179 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 304, + 72, + 547, + 179 + ], + "type": "text", + "content": " is the number of pairs. Note that it is an online operation and is irrelevant to the image and point data augmentation." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 180, + 547, + 467 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 180, + 547, + 467 + ], + "spans": [ + { + "bbox": [ + 304, + 180, + 547, + 467 + ], + "type": "text", + "content": "Previous methods [44, 51] provide a promising solution to cross-modal knowledge transfer. They first construct positive pixel-point pairs " + }, + { + "bbox": [ + 304, + 180, + 547, + 467 + ], + "type": "inline_equation", + "content": "\\{x_i, p_i\\}_{i=1}^M" + }, + { + "bbox": [ + 304, + 180, + 547, + 467 + ], + "type": "text", + "content": " and negative pairs " + }, + { + "bbox": [ + 304, + 180, + 547, + 467 + ], + "type": "inline_equation", + "content": "\\{x_i, p_j\\} (i \\neq j)" + }, + { + "bbox": [ + 304, + 180, + 547, + 467 + ], + "type": "text", + "content": ", and then pull in the positive pairs while pushing away the negative pairs in the embedding space via the InfoNCE loss. Despite the encourageable performance of previous methods in transferring cross-modal knowledge, they are both confronted with the same optimization-conflict issue. For example, suppose " + }, + { + "bbox": [ + 304, + 180, + 547, + 467 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 180, + 547, + 467 + ], + "type": "text", + "content": "-th pixel " + }, + { + "bbox": [ + 304, + 180, + 547, + 467 + ], + "type": "inline_equation", + "content": "x_i" + }, + { + "bbox": [ + 304, + 180, + 547, + 467 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 180, + 547, + 467 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 304, + 180, + 547, + 467 + ], + "type": "text", + "content": "-th point " + }, + { + "bbox": [ + 304, + 180, + 547, + 467 + ], + "type": "inline_equation", + "content": "p_j" + }, + { + "bbox": [ + 304, + 180, + 547, + 467 + ], + "type": "text", + "content": " are in the different positions of the same instance with the same semantics. However, the InfoNCE loss will try to push them away, which is unreasonable and hammer the performance of the downstream tasks [51]. In light of this, we propose a Semantic Consistency Regularization that leverages the CLIP's semantic information to alleviate this issue. Specifically, we generate the dense pixel-text pairs " + }, + { + "bbox": [ + 304, + 180, + 547, + 467 + ], + "type": "inline_equation", + "content": "\\{x_i, t_i\\}_{i=1}^M" + }, + { + "bbox": [ + 304, + 180, + 547, + 467 + ], + "type": "text", + "content": " by following the off-the-shelf method MaskCLIP [61] (Fig. 3), where " + }, + { + "bbox": [ + 304, + 180, + 547, + 467 + ], + "type": "inline_equation", + "content": "t_i" + }, + { + "bbox": [ + 304, + 180, + 547, + 467 + ], + "type": "text", + "content": " is the text embedding generated from the CLIP's text encoder. Note that the pixel-text mappings are free-available from CLIP without any additional training. We then transfer pixel-text pairs to point-text pairs " + }, + { + "bbox": [ + 304, + 180, + 547, + 467 + ], + "type": "inline_equation", + "content": "\\{p_i, t_i\\}_{i=1}^M" + }, + { + "bbox": [ + 304, + 180, + 547, + 467 + ], + "type": "text", + "content": " and utilize the text semantics to select the positive and negative point samples for contrastive learning. The objective function is as follows:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 315, + 483, + 545, + 525 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 483, + 545, + 525 + ], + "spans": [ + { + "bbox": [ + 315, + 483, + 545, + 525 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {S - i n f o} = - \\sum_ {c = 1} ^ {C} \\log \\frac {\\sum_ {t _ {i} \\in c , p _ {i}} \\exp \\left(D \\left(t _ {i} , p _ {i}\\right) / \\tau\\right)}{\\sum_ {t _ {i} \\in c , t _ {j} \\notin c , p _ {j}} \\exp \\left(D \\left(t _ {i} , p _ {j}\\right) / \\tau\\right)}, \\tag {1}", + "image_path": "4f8458f03058b7e274a7c43403d0b4bafa3e7d5e7e7d3f3e06ae94dee436d2e8.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 525, + 545, + 561 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 525, + 545, + 561 + ], + "spans": [ + { + "bbox": [ + 304, + 525, + 545, + 561 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 525, + 545, + 561 + ], + "type": "inline_equation", + "content": "t_i \\in c" + }, + { + "bbox": [ + 304, + 525, + 545, + 561 + ], + "type": "text", + "content": " indicates that " + }, + { + "bbox": [ + 304, + 525, + 545, + 561 + ], + "type": "inline_equation", + "content": "t_i" + }, + { + "bbox": [ + 304, + 525, + 545, + 561 + ], + "type": "text", + "content": " is generated by " + }, + { + "bbox": [ + 304, + 525, + 545, + 561 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 304, + 525, + 545, + 561 + ], + "type": "text", + "content": "-th classes name, and " + }, + { + "bbox": [ + 304, + 525, + 545, + 561 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 304, + 525, + 545, + 561 + ], + "type": "text", + "content": " is the number of classes. " + }, + { + "bbox": [ + 304, + 525, + 545, + 561 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 304, + 525, + 545, + 561 + ], + "type": "text", + "content": " denotes the scalar product operation and " + }, + { + "bbox": [ + 304, + 525, + 545, + 561 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 304, + 525, + 545, + 561 + ], + "type": "text", + "content": " is a temperature term " + }, + { + "bbox": [ + 304, + 525, + 545, + 561 + ], + "type": "inline_equation", + "content": "(\\tau > 0)" + }, + { + "bbox": [ + 304, + 525, + 545, + 561 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 561, + 545, + 645 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 561, + 545, + 645 + ], + "spans": [ + { + "bbox": [ + 304, + 561, + 545, + 645 + ], + "type": "text", + "content": "Since the text is composed of class names placed into pre-defined templates, the text embedding represents the semantic information of the corresponding class. Therefore, those points with the same semantics will be restricted near the same text embedding, and those with different semantics will be pushed away. To this end, our Semantic Consistency Regularization causes less conflict in contrastive learning." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 658, + 545, + 683 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 658, + 545, + 683 + ], + "spans": [ + { + "bbox": [ + 306, + 658, + 545, + 683 + ], + "type": "text", + "content": "3.2.2 Semantic-guided Spatial-temporal Consistency Regularization" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 689, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 689, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 689, + 547, + 713 + ], + "type": "text", + "content": "Besides semantic consistency regularization, we consider how image pixel features help to regularize a 3D network." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7023" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 46, + 72, + 292, + 178 + ], + "blocks": [ + { + "bbox": [ + 46, + 72, + 292, + 178 + ], + "lines": [ + { + "bbox": [ + 46, + 72, + 292, + 178 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 292, + 178 + ], + "type": "image", + "image_path": "c03b35ee78740b90f8de34ea9faf125257f71553320e469e344e2ff96c32183f.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 182, + 289, + 262 + ], + "lines": [ + { + "bbox": [ + 46, + 182, + 289, + 262 + ], + "spans": [ + { + "bbox": [ + 46, + 182, + 289, + 262 + ], + "type": "text", + "content": "Figure 4. Illustration of the image pixel-to-point mapping (left) and semantic-guided fusion feature generation (right). We build the grid-wise correspondence between an image " + }, + { + "bbox": [ + 46, + 182, + 289, + 262 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 46, + 182, + 289, + 262 + ], + "type": "text", + "content": " and the temporally coherent LiDAR point cloud " + }, + { + "bbox": [ + 46, + 182, + 289, + 262 + ], + "type": "inline_equation", + "content": "\\{P_k\\}_{k=1}^K" + }, + { + "bbox": [ + 46, + 182, + 289, + 262 + ], + "type": "text", + "content": " within " + }, + { + "bbox": [ + 46, + 182, + 289, + 262 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 46, + 182, + 289, + 262 + ], + "type": "text", + "content": " seconds and generate semantic-guided fusion features for individual grids. Both " + }, + { + "bbox": [ + 46, + 182, + 289, + 262 + ], + "type": "inline_equation", + "content": "\\{\\hat{x}_i^k, \\hat{p}_i^k\\}_{i=1,k=1}^M, K" + }, + { + "bbox": [ + 46, + 182, + 289, + 262 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 182, + 289, + 262 + ], + "type": "inline_equation", + "content": "\\{f_n\\}_{n=1}^N" + }, + { + "bbox": [ + 46, + 182, + 289, + 262 + ], + "type": "text", + "content": " are used to perform Spatial-Temporal Consistency Regularization." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 281, + 289, + 378 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 281, + 289, + 378 + ], + "spans": [ + { + "bbox": [ + 46, + 281, + 289, + 378 + ], + "type": "text", + "content": "The natural alternative directly pulls in the point feature with its corresponding pixel in the embedding space. However, the noise-assigned semantics of the image pixel and the imperfect pixel-point mapping hinder the downstream task's performance. To this end, we propose a novel semantic-guided Spatial-Temporal Consistency Regularization to alleviate the problem by imposing a soft constraint on points within local space and time." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 380, + 289, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 380, + 289, + 533 + ], + "spans": [ + { + "bbox": [ + 46, + 380, + 289, + 533 + ], + "type": "text", + "content": "Specifically, given an image " + }, + { + "bbox": [ + 46, + 380, + 289, + 533 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 46, + 380, + 289, + 533 + ], + "type": "text", + "content": " and temporally coherent LiDAR point cloud " + }, + { + "bbox": [ + 46, + 380, + 289, + 533 + ], + "type": "inline_equation", + "content": "\\{P_k\\}_{k=1}^K" + }, + { + "bbox": [ + 46, + 380, + 289, + 533 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 380, + 289, + 533 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 46, + 380, + 289, + 533 + ], + "type": "text", + "content": " is the number of sweeps within " + }, + { + "bbox": [ + 46, + 380, + 289, + 533 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 46, + 380, + 289, + 533 + ], + "type": "text", + "content": " seconds. Note that the image is matched to the first frame of the point cloud " + }, + { + "bbox": [ + 46, + 380, + 289, + 533 + ], + "type": "inline_equation", + "content": "P_1" + }, + { + "bbox": [ + 46, + 380, + 289, + 533 + ], + "type": "text", + "content": " with pixel-point pairs " + }, + { + "bbox": [ + 46, + 380, + 289, + 533 + ], + "type": "inline_equation", + "content": "\\{\\hat{x}_i^1, \\hat{p}_i^1\\}_{i=1}^M" + }, + { + "bbox": [ + 46, + 380, + 289, + 533 + ], + "type": "text", + "content": ". We register the rest of the point cloud to the first frame via the calibration matrices and map them to the image (Fig. 4). Thus we obtain all pixel-point-text pairs in a short temporal " + }, + { + "bbox": [ + 46, + 380, + 289, + 533 + ], + "type": "inline_equation", + "content": "\\{\\hat{x}_i^k, \\hat{p}_i^k, t_i^k\\}_{i=1,k=1}^M" + }, + { + "bbox": [ + 46, + 380, + 289, + 533 + ], + "type": "text", + "content": ". Next, we divide the entire stitched point cloud into regular grids " + }, + { + "bbox": [ + 46, + 380, + 289, + 533 + ], + "type": "inline_equation", + "content": "\\{g_n\\}_{n=1}^N" + }, + { + "bbox": [ + 46, + 380, + 289, + 533 + ], + "type": "text", + "content": ", where the temporally coherent points are located in the same grid. We impose the spatial-temporal consistency constraint within individual grids by the following objective function:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 557, + 288, + 587 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 557, + 288, + 587 + ], + "spans": [ + { + "bbox": [ + 54, + 557, + 288, + 587 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {S S R} = \\sum_ {g _ {n}} \\sum_ {(\\hat {i}, \\hat {k}) \\in g _ {n}} (1 - \\operatorname {s i g m o i d} \\left(D \\left(\\hat {p} _ {\\hat {i}} ^ {\\hat {k}}, f _ {n}\\right)\\right)) / N, \\tag {2}", + "image_path": "a648bc9ff00aeccf4fa1a005e0d8b856247ecb2a4d5a36e9b16087a0b03d9594.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 602, + 287, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 602, + 287, + 641 + ], + "spans": [ + { + "bbox": [ + 46, + 602, + 287, + 641 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 602, + 287, + 641 + ], + "type": "inline_equation", + "content": "(\\hat{i},\\hat{k})\\in g_{n}" + }, + { + "bbox": [ + 46, + 602, + 287, + 641 + ], + "type": "text", + "content": " indicates the pixel-point pair " + }, + { + "bbox": [ + 46, + 602, + 287, + 641 + ], + "type": "inline_equation", + "content": "\\{\\hat{x}_i^k,\\hat{p}_i^k\\}" + }, + { + "bbox": [ + 46, + 602, + 287, + 641 + ], + "type": "text", + "content": " is located in the " + }, + { + "bbox": [ + 46, + 602, + 287, + 641 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 46, + 602, + 287, + 641 + ], + "type": "text", + "content": " -th grid. " + }, + { + "bbox": [ + 46, + 602, + 287, + 641 + ], + "type": "inline_equation", + "content": "\\{f_n\\}_{n = 1}^N" + }, + { + "bbox": [ + 46, + 602, + 287, + 641 + ], + "type": "text", + "content": " is a semantic-guided cross-modal fusion feature formulated by:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 100, + 654, + 288, + 684 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 100, + 654, + 288, + 684 + ], + "spans": [ + { + "bbox": [ + 100, + 654, + 288, + 684 + ], + "type": "interline_equation", + "content": "f _ {n} = \\sum_ {(\\hat {i}, \\hat {k}) \\in g _ {n}} a _ {\\hat {i}} ^ {\\hat {k}} * \\hat {x} _ {\\hat {i}} ^ {\\hat {k}} + b _ {\\hat {i}} ^ {\\hat {k}} * \\hat {p} _ {\\hat {i}} ^ {\\hat {k}}, \\tag {3}", + "image_path": "9cc047d532f5f08bceabbf9f05e697e0233a697ed667d968a2f4b67f7291acef.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 699, + 256, + 717 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 699, + 256, + 717 + ], + "spans": [ + { + "bbox": [ + 47, + 699, + 256, + 717 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 699, + 256, + 717 + ], + "type": "inline_equation", + "content": "a_{\\hat{i}}^{\\hat{k}}" + }, + { + "bbox": [ + 47, + 699, + 256, + 717 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 699, + 256, + 717 + ], + "type": "inline_equation", + "content": "b_{\\hat{i}}^{\\hat{k}}" + }, + { + "bbox": [ + 47, + 699, + 256, + 717 + ], + "type": "text", + "content": " are attention weight calculated by:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 310, + 91, + 546, + 175 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 310, + 91, + 546, + 175 + ], + "spans": [ + { + "bbox": [ + 310, + 91, + 546, + 175 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} a _ {i} ^ {\\hat {k}} = \\frac {\\exp (D (\\hat {x} _ {\\hat {i}} ^ {\\hat {k}} , t _ {\\hat {i}} ^ {1}) / \\lambda)}{\\sum_ {(\\hat {i} , \\hat {k}) \\in g _ {n}} \\exp (D (\\hat {x} _ {\\hat {i}} ^ {\\hat {k}} , t _ {\\hat {i}} ^ {1}) / \\lambda) + \\exp (D (\\hat {p} _ {\\hat {i}} ^ {\\hat {k}} , t _ {\\hat {i}} ^ {1}) / \\lambda)}, \\\\ b _ {\\hat {i}} ^ {\\hat {k}} = \\frac {\\exp \\left(D \\left(\\hat {p} _ {\\hat {i}} ^ {\\hat {k}} , t _ {\\hat {i}} ^ {1}\\right) / \\lambda\\right)}{\\sum_ {(\\hat {i} , \\hat {k}) \\in g _ {n}} \\exp \\left(D \\left(\\hat {x} _ {\\hat {i}} ^ {\\hat {k}}, t _ {\\hat {i}} ^ {1}\\right) / \\lambda\\right) + \\exp \\left(D \\left(\\hat {p} _ {\\hat {i}} ^ {\\hat {k}}, t _ {\\hat {i}} ^ {1}\\right) / \\lambda\\right)}, \\tag {4} \\\\ \\end{array}", + "image_path": "7b07f39b4271f9f1dc0ce0572fa2f84c47f691cd2fc147f644ca9772101f9206.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 184, + 440, + 194 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 184, + 440, + 194 + ], + "spans": [ + { + "bbox": [ + 305, + 184, + 440, + 194 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 184, + 440, + 194 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 305, + 184, + 440, + 194 + ], + "type": "text", + "content": " is the temperature term." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 195, + 547, + 256 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 195, + 547, + 256 + ], + "spans": [ + { + "bbox": [ + 304, + 195, + 547, + 256 + ], + "type": "text", + "content": "Actually, those pixel and point features within the local grid " + }, + { + "bbox": [ + 304, + 195, + 547, + 256 + ], + "type": "inline_equation", + "content": "g_{n}" + }, + { + "bbox": [ + 304, + 195, + 547, + 256 + ], + "type": "text", + "content": " are restricted near a dynamic centre " + }, + { + "bbox": [ + 304, + 195, + 547, + 256 + ], + "type": "inline_equation", + "content": "f_{n}" + }, + { + "bbox": [ + 304, + 195, + 547, + 256 + ], + "type": "text", + "content": ". Thus, such a soft constraint alleviates the noisy prediction and calibration error issues. At the same time, it imposes Spatio-Temporal Regularization on the temporally coherent point features." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 270, + 480, + 283 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 270, + 480, + 283 + ], + "spans": [ + { + "bbox": [ + 305, + 270, + 480, + 283 + ], + "type": "text", + "content": "3.2.3 Switchable Self-training Strategy" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 289, + 547, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 289, + 547, + 472 + ], + "spans": [ + { + "bbox": [ + 304, + 289, + 547, + 472 + ], + "type": "text", + "content": "We combine the loss function " + }, + { + "bbox": [ + 304, + 289, + 547, + 472 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{S,\\text{info}}" + }, + { + "bbox": [ + 304, + 289, + 547, + 472 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 289, + 547, + 472 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{SSR}" + }, + { + "bbox": [ + 304, + 289, + 547, + 472 + ], + "type": "text", + "content": " to end-to-end train the whole network, where the CLIP's image and text encoder backbone are frozen during training. We find that method worked only when the pixel-point feature " + }, + { + "bbox": [ + 304, + 289, + 547, + 472 + ], + "type": "inline_equation", + "content": "\\{x_i,p_i\\}_{i = 1}^M" + }, + { + "bbox": [ + 304, + 289, + 547, + 472 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 289, + 547, + 472 + ], + "type": "inline_equation", + "content": "\\{\\hat{x}_i^k,\\hat{p}_i^k\\}_{i = 1,k = 1}^{\\hat{M},K}" + }, + { + "bbox": [ + 304, + 289, + 547, + 472 + ], + "type": "text", + "content": ", which are used in " + }, + { + "bbox": [ + 304, + 289, + 547, + 472 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{S,\\text{info}}" + }, + { + "bbox": [ + 304, + 289, + 547, + 472 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 289, + 547, + 472 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{SSR}" + }, + { + "bbox": [ + 304, + 289, + 547, + 472 + ], + "type": "text", + "content": ", are generated from different learnable linear layer. On top of that, we further put forward an effective strategy to promote performance. Specifically, after contrastive learning of the 3D network for a few epochs, we randomly switch the point pseudo label between the paired image pixel's pseudo label and the point's predicted label. Since different modality networks learn different feature representations, they can filter different types of error introduced by noisy pseudo labels. By this switchable operation, the error flows can be reduced by mutually [29]." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 481, + 388, + 495 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 481, + 388, + 495 + ], + "spans": [ + { + "bbox": [ + 305, + 481, + 388, + 495 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 502, + 547, + 646 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 502, + 547, + 646 + ], + "spans": [ + { + "bbox": [ + 304, + 502, + 547, + 646 + ], + "type": "text", + "content": "Datasets. We conduct extensive experiments on two large-scale outdoor LiDAR semantic segmentation datasets, i.e., SemanticKITTI [3] and nuScenes [5, 24], and one indoor dataset ScanNet [20]. The nuScenes dataset contains 700 scenes for training, 150 scenes for validation, and 150 scenes for testing, where 16 classes are utilized for LiDAR semantic segmentation. As for SemanticKITTI, it contains 19 classes for training and evaluation. It has 22 sequences, where sequences 00 to 10, 08, and 11 to 21 are used for training, validation, and testing, respectively. ScanNet [20] contains 1603 scans with 20 classes, where 1201 scans are for training, 312 scans are for validation, and 100 scans are for testing." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 653, + 548, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 653, + 548, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 653, + 548, + 715 + ], + "type": "text", + "content": "Implementation Details. We follow SLidR [51] to pretrain the network on the nuScenes [5, 24] dataset. The network is pre-trained on all keyframes from 600 scenes. Besides, the pre-trained network is fine-tuned on SemanticKITTI [3] to verify the generalization ability. We" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7024" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 50, + 98, + 286, + 182 + ], + "blocks": [ + { + "bbox": [ + 47, + 71, + 287, + 93 + ], + "lines": [ + { + "bbox": [ + 47, + 71, + 287, + 93 + ], + "spans": [ + { + "bbox": [ + 47, + 71, + 287, + 93 + ], + "type": "text", + "content": "Table 1. Comparisons (mIoU) among self-supervised methods on the nuScenes [24], SemanticKITTI [3], and ScanNet [20] val sets." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 50, + 98, + 286, + 182 + ], + "lines": [ + { + "bbox": [ + 50, + 98, + 286, + 182 + ], + "spans": [ + { + "bbox": [ + 50, + 98, + 286, + 182 + ], + "type": "table", + "html": "
InitializationnuScenesSemanticKITTIScanNet
1%100%1%100%5%100%
Random42.269.132.552.146.163.3
PPKT [44]48.070.139.153.147.564.2
SLidR [51]48.270.439.654.347.964.9
PointContrast [55]47.269.237.152.347.664.5
CLIP2Scene56.371.542.655.048.465.1
", + "image_path": "1ac25c1eae9cb23d523696019d59d4b34347819172376f9a0849a974e1d37ca8.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 104, + 219, + 233, + 249 + ], + "blocks": [ + { + "bbox": [ + 47, + 193, + 286, + 214 + ], + "lines": [ + { + "bbox": [ + 47, + 193, + 286, + 214 + ], + "spans": [ + { + "bbox": [ + 47, + 193, + 286, + 214 + ], + "type": "text", + "content": "Table 2. Annotation-free 3D semantic segmentation performance (mIoU) on the nuScenes [24] and ScanNet [20] val sets." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 104, + 219, + 233, + 249 + ], + "lines": [ + { + "bbox": [ + 104, + 219, + 233, + 249 + ], + "spans": [ + { + "bbox": [ + 104, + 219, + 233, + 249 + ], + "type": "table", + "html": "
MethodnuScenesScanNet
CLIP2Scene20.8025.08
", + "image_path": "73987a8a261f6284f7819b8f3ccf3f946f4dc364e65da34b53f149617fabe18d.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 269, + 287, + 497 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 269, + 287, + 497 + ], + "spans": [ + { + "bbox": [ + 46, + 269, + 287, + 497 + ], + "type": "text", + "content": "leverage the CLIP model to generate image features and text embedding. Following MaskCLIP, we modify the attention pooling layer of the CLIP's image encoder, thus extracting the dense pixel-text correspondences. We take SPVCNN [52] as the 3D network to produce the point-wise feature. The framework is developed on PyTorch, where the CLIP model is frozen during training. The training time is about 40 hours for 20 epochs on two NVIDIA Tesla A100 GPUs. The optimizer is SGD with a cosine scheduler. We set the temperature " + }, + { + "bbox": [ + 46, + 269, + 287, + 497 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 46, + 269, + 287, + 497 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 269, + 287, + 497 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 46, + 269, + 287, + 497 + ], + "type": "text", + "content": " to be 1 and 0.5, respectively. The sweep number is set to be 3 empirically. Besides, We adopt MinkowskiNet14 [19] as the backbone for evaluation on the ScanNet dataset, where the number of sweeps is set to be 1 and the training epochs is 30. As for the Switchable Self-Training Strategy, we randomly switch the point supervision signal after 10 epochs. We apply several data augmentations in contrastive learning, including random rotation along the z-axis and random flip on the point cloud, random horizontal flip, and random crop-resize on the image." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 503, + 257, + 516 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 503, + 257, + 516 + ], + "spans": [ + { + "bbox": [ + 47, + 503, + 257, + 516 + ], + "type": "text", + "content": "4.1. Annotation-free Semantic Segmentation" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 522, + 287, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 522, + 287, + 605 + ], + "spans": [ + { + "bbox": [ + 46, + 522, + 287, + 605 + ], + "type": "text", + "content": "After pre-training the network, we show the performance of the 3D network when it is not fine-tuned on any annotations (Table 2). As no previous method reports the 3D annotation-free segmentation performance, we compare our method with different setups (Table 3). In what follows, we describe the experimental settings and give insights into our method and the different settings." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 605, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 288, + 714 + ], + "type": "text", + "content": "Settings. We conduct experiments on the nuScenes and ScanNet datasets to evaluate the annotation-free semantic segmentation performance. Following MaskCLIP [61], we place the class name into 85 hand-craft prompts and feed it into the CLIP's text encoder to produce multiple text features. We then average the text features and feed the averaged features to the classifier for point-wise prediction. Besides, to explore how to effectively transfer CLIP's knowledge to the 3D network for annotation-free segmentation, We conduct" + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 339, + 98, + 516, + 291 + ], + "blocks": [ + { + "bbox": [ + 305, + 71, + 547, + 93 + ], + "lines": [ + { + "bbox": [ + 305, + 71, + 547, + 93 + ], + "spans": [ + { + "bbox": [ + 305, + 71, + 547, + 93 + ], + "type": "text", + "content": "Table 3. Ablation study on the nuScenes [24] val set for annotation-free 3D semantic segmentation." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 339, + 98, + 516, + 291 + ], + "lines": [ + { + "bbox": [ + 339, + 98, + 516, + 291 + ], + "spans": [ + { + "bbox": [ + 339, + 98, + 516, + 291 + ], + "type": "table", + "html": "
Ablation TargetSettingmIoU (%)
-Baseline15.1
PromptsnuScenes15.1 (+0.0)
SemanticKITTI13.9 (-1.2)
Cityscapes11.3 (-3.8)
All15.3 (+0.2)
Regularizationw/o StCR19.8 (+4.7)
w/o SCR16.8 (+1.7)
KL0.0 (-15.1)
Training Strategyw/o S318.8 (+3.7)
ST10.1 (-4.0)
Sweeps1 sweep18.7 (+3.6)
3 sweeps20.8 (+5.7)
5 sweeps20.6 (+5.5)
merged18.6 (+3.5)
Full ConfigurationCLIP2Scene20.8 (+5.7)
", + "image_path": "35fd1e5ae2e98391e0f0be995552b7c75a0aedf161f5d566cef15e8116b787d1.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 312, + 546, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 312, + 546, + 335 + ], + "spans": [ + { + "bbox": [ + 305, + 312, + 546, + 335 + ], + "type": "text", + "content": "the following experiments to highlight the effectiveness of different modules in our framework." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 338, + 547, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 338, + 547, + 373 + ], + "spans": [ + { + "bbox": [ + 305, + 338, + 547, + 373 + ], + "type": "text", + "content": "Baseline. The input of the 3D semantic segmentation network is only one sweep, and we pre-train the framework via semantic consistency regularization." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 375, + 547, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 375, + 547, + 422 + ], + "spans": [ + { + "bbox": [ + 305, + 375, + 547, + 422 + ], + "type": "text", + "content": "Prompts (nuScenes, SemanticKITTI, Cityscapes, All). Based on the baseline, we respectively replace the nuScenes, SemanticKITTI, Cityscapes, and all class names into the prompts to produce the text embedding." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 423, + 547, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 423, + 547, + 495 + ], + "spans": [ + { + "bbox": [ + 305, + 423, + 547, + 495 + ], + "type": "text", + "content": "Regularization (w/o StCR, w/o SCR, KL). Based on the full method, we remove the Spatial-temporal Consistency Regularization (w/o StCR) and remove the Semantic Consistency Regularization (w/o SCR). Besides, we abuse both StCR and SCR and distill the image feature to the point cloud by Kullback-Leibler (KL) divergence loss." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 496, + 547, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 496, + 547, + 544 + ], + "spans": [ + { + "bbox": [ + 305, + 496, + 547, + 544 + ], + "type": "text", + "content": "Training Strategies (w/o S3, ST). We abuse the Switchable Self-Training Strategy (w/o S3) in the full method. Besides, we show the performance of only training the 3D network by their own predictions after ten epochs (ST)." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 545, + 546, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 545, + 546, + 593 + ], + "spans": [ + { + "bbox": [ + 305, + 545, + 546, + 593 + ], + "type": "text", + "content": "Sweeps Number (1 sweep, 3 sweeps, 5 sweeps, and merged). We set the sweep number " + }, + { + "bbox": [ + 305, + 545, + 546, + 593 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 305, + 545, + 546, + 593 + ], + "type": "text", + "content": " to be 1, 3, and 5, respectively. Besides, we also take three sweeps of the point cloud as the input to pre-train the network (merged)." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 594, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 547, + 714 + ], + "type": "text", + "content": "Effect of Different Prompts. To verify how text embedding affects the performance, we generate various text embedding by the class name from different datasets (nuScenes, SemanticKITTI, and Cityscapes) and all classes for pre-training the framework. As shown in Table 3, we find that even learning with other datasets' text embedding (SemanticKITTI and Cityscapes), the 3D network could still recognize the nuScenes's objects with decent performance (13.9% and 11.3% mIoU, respectively). The result shows that the 3D network is capable of open-vocabulary recognition ability." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7025" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 55, + 77, + 541, + 297 + ], + "blocks": [ + { + "bbox": [ + 55, + 77, + 541, + 297 + ], + "lines": [ + { + "bbox": [ + 55, + 77, + 541, + 297 + ], + "spans": [ + { + "bbox": [ + 55, + 77, + 541, + 297 + ], + "type": "image", + "image_path": "89ffe74d74d82a6c97723e01d1374336ed7a758c47b237279664fa3a765805fd.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 302, + 546, + 336 + ], + "lines": [ + { + "bbox": [ + 46, + 302, + 546, + 336 + ], + "spans": [ + { + "bbox": [ + 46, + 302, + 546, + 336 + ], + "type": "text", + "content": "Figure 5. Qualitative results of annotation-free semantic segmentation on nuScenes dataset. Note that we show the results of individual classes. From the left to the right column are bus, motorcycle, car, and truck, respectively. The first row [ground truth] is the annotated semantic label. The second row [ours*] is our prediction of the highlighted target. The third row [ours] is our prediction of full classes." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 351, + 289, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 351, + 289, + 483 + ], + "spans": [ + { + "bbox": [ + 46, + 351, + 289, + 483 + ], + "type": "text", + "content": "Effect of Semantic and Spatial-temporal Consistency Regularization. We remove Spatial-temporal Consistency Regularization (w/o SCR) from our method. Experiments show that the performance is dramatically decreased, indicating the effectiveness of our design. Besides, we also distill the image feature to the point cloud by KL divergence loss, where the text embeddings calculate the logits. However, such a method fails to transfer the semantic information from the image. The main reason is the noise-assigned semantics of the image pixel and the imperfect pixel-point correspondence due to the calibration error." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 484, + 288, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 484, + 288, + 555 + ], + "spans": [ + { + "bbox": [ + 46, + 484, + 288, + 555 + ], + "type": "text", + "content": "Effect of Switchable Self-training Strategy. To examine the effect of the Switchable Self-Training Strategy, we either train the network with image supervision (w/o S3) or train the 3D network by their own predictions. Both trials witness a performance drop, indicating Switchable Self-Training Strategy is efficient in cross-modal self-supervised learning." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 556, + 288, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 556, + 288, + 640 + ], + "spans": [ + { + "bbox": [ + 46, + 556, + 288, + 640 + ], + "type": "text", + "content": "Effect of Sweep Numbers. Intuitively, the performance of our method benefits from more sweeps information. Therefore, we also show the performance when restricting sweep size to 1, 3, and 5, respectively. However, we observe that the performance of 5 sweeps is similar to 3 sweeps but is more computationally expensive. Thus, we empirically set the sweep number to be 3." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 641, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 641, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 641, + 288, + 715 + ], + "type": "text", + "content": "Qualitative Evaluation. The qualitative evaluations of individual classes (bus, motorcycle, car, and truck) are in Fig. 5, indicating that our method is able to perceive the objects even without training on any annotated data. However, we also observe the false positive predictions around the ground truth objects. We will resolve this issue in future work." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 305, + 351, + 535, + 364 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 351, + 535, + 364 + ], + "spans": [ + { + "bbox": [ + 305, + 351, + 535, + 364 + ], + "type": "text", + "content": "4.2. Annotation-efficient Semantic Segmentation" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 372, + 547, + 455 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 372, + 547, + 455 + ], + "spans": [ + { + "bbox": [ + 304, + 372, + 547, + 455 + ], + "type": "text", + "content": "The pre-trained 3D network also boosts the performance when few labeled data are available for training. We directly compare SLidR [51], the only published method for image-to-Lidar self-supervised representation distillation. Besides, we also compared PPKT [44] and PointContrast [55]. In the following, we introduce SLidR and PPKT and compare them in detail." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 457, + 547, + 542 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 457, + 547, + 542 + ], + "spans": [ + { + "bbox": [ + 304, + 457, + 547, + 542 + ], + "type": "text", + "content": "PPKT. PPKT is a cross-modal self-supervised method for the RGB-D dataset. It performs 2D-to-3D knowledge distillation via pixel-to-point contrastive loss. For a fair comparison, we use the same 3D network and training protocol but replace our semantic and Spatio-Temporal Regularization with InfoNCE loss. The framework is trained 50 epochs on 4096 randomly selected image-to-point pairs." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 544, + 547, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 544, + 547, + 604 + ], + "spans": [ + { + "bbox": [ + 304, + 544, + 547, + 604 + ], + "type": "text", + "content": "SLidR. SLidR is an image-to-Lidar self-supervised method for autonomous driving data. Compared with PPKT, it introduces image super-pixel into cross-modal self-supervised learning. For a fair comparison, we replace our loss function with their superpixel-driven contrastive loss." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 605, + 548, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 548, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 548, + 715 + ], + "type": "text", + "content": "Performance. As shown in Table 1, our method significantly outperforms the state-of-the-art methods when fine-tuned on " + }, + { + "bbox": [ + 304, + 605, + 548, + 715 + ], + "type": "inline_equation", + "content": "1\\%" + }, + { + "bbox": [ + 304, + 605, + 548, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 605, + 548, + 715 + ], + "type": "inline_equation", + "content": "100\\%" + }, + { + "bbox": [ + 304, + 605, + 548, + 715 + ], + "type": "text", + "content": " nuScenes dataset, with the improvement of " + }, + { + "bbox": [ + 304, + 605, + 548, + 715 + ], + "type": "inline_equation", + "content": "8.1\\%" + }, + { + "bbox": [ + 304, + 605, + 548, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 605, + 548, + 715 + ], + "type": "inline_equation", + "content": "1.1\\%" + }, + { + "bbox": [ + 304, + 605, + 548, + 715 + ], + "type": "text", + "content": ", respectively. Compared with the random initialization, the improvement is " + }, + { + "bbox": [ + 304, + 605, + 548, + 715 + ], + "type": "inline_equation", + "content": "14.1\\%" + }, + { + "bbox": [ + 304, + 605, + 548, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 605, + 548, + 715 + ], + "type": "inline_equation", + "content": "2.4\\%" + }, + { + "bbox": [ + 304, + 605, + 548, + 715 + ], + "type": "text", + "content": ", respectively, indicating the efficiency of our Semantic-driven Cross-modal Contrastive Learning framework. The qualitative results are shown in Fig. 6. Besides, we also verify the cross-domain generalization ability of our method. When pre-training the" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7026" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 53, + 77, + 545, + 316 + ], + "blocks": [ + { + "bbox": [ + 53, + 77, + 545, + 316 + ], + "lines": [ + { + "bbox": [ + 53, + 77, + 545, + 316 + ], + "spans": [ + { + "bbox": [ + 53, + 77, + 545, + 316 + ], + "type": "image", + "image_path": "97ac7c1a4bf31eb9d085e0bde550b8b14bd6ecef4676192d80d18bcd4042f676.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 321, + 547, + 355 + ], + "lines": [ + { + "bbox": [ + 46, + 321, + 547, + 355 + ], + "spans": [ + { + "bbox": [ + 46, + 321, + 547, + 355 + ], + "type": "text", + "content": "Figure 6. Qualitative results of fine-tuning on " + }, + { + "bbox": [ + 46, + 321, + 547, + 355 + ], + "type": "inline_equation", + "content": "1\\%" + }, + { + "bbox": [ + 46, + 321, + 547, + 355 + ], + "type": "text", + "content": " nuScenes dataset. From the first row to the last row are the input LiDAR scan, ground truth, prediction of SLidR, and our prediction, respectively. Note that we show the results by error map, where the red point indicates the wrong prediction. Apparently, our method achieves decent performance." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 376, + 288, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 376, + 288, + 411 + ], + "spans": [ + { + "bbox": [ + 46, + 376, + 288, + 411 + ], + "type": "text", + "content": "3D network on the nuScenes dataset and fine-tuning on " + }, + { + "bbox": [ + 46, + 376, + 288, + 411 + ], + "type": "inline_equation", + "content": "1\\%" + }, + { + "bbox": [ + 46, + 376, + 288, + 411 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 376, + 288, + 411 + ], + "type": "inline_equation", + "content": "100\\%" + }, + { + "bbox": [ + 46, + 376, + 288, + 411 + ], + "type": "text", + "content": " SemanticKITTI dataset, our method significantly outperforms other state-of-the-art self-supervised methods." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 415, + 290, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 415, + 290, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 415, + 290, + 715 + ], + "type": "text", + "content": "Discussions. PPKT and SLidR reveal that contrastive loss is promising for transferring knowledge from image to point cloud. Like self-supervised learning, constructing the positive and negative samples is vital to unsupervised cross-modal knowledge distillation. However, previous methods suffer from optimization-conflict issues, i.e., some negative paired samples are actually positive pairs. For example, the road occupies a large proportion of the point cloud in a scene and is supposed to have the same semantics in the semantic segmentation task. When randomly selecting training samples, most negatively defined road-road points are actually positive. When feedforwarding such samples into contrastive learning, the contrastive loss will push them away in the embedding space, leading to unsatisfactory representation learning and hammering the downstream tasks' performance. SLidR introduces superpixel-driven contrastive learning to alleviate such issues. The motivation is that the visual representation of the image pixel and the projected points are consistent intra-superpixel. Although avoiding selecting negative image-point pairs from the same superpixel, the conflict still exists inter-superpixel. In our CLIP2Scene, we introduce the free-available dense pixel-text correspondence to alleviate the optimization conflicts. The text embedding represents the semantic information and can be used to select more reasonable training samples for contrastive learning." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 304, + 376, + 547, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 376, + 547, + 483 + ], + "spans": [ + { + "bbox": [ + 304, + 376, + 547, + 483 + ], + "type": "text", + "content": "Besides training sample selection, the previous method also ignores the temporal coherence of the multi-sweep point cloud. That is, for LiDAR points mapping to the same image pixel, their feature is restricted to be consistent. Besides, considering the calibration error between the LiDAR scan and the camera image. We relax the pixel-to-point mapping to image grid-to-point grid mapping for consistency regularization. To this end, our Spatial-temporal consistency regularization leads to a more rational point representation." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 485, + 548, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 485, + 548, + 557 + ], + "spans": [ + { + "bbox": [ + 304, + 485, + 548, + 557 + ], + "type": "text", + "content": "Last but not least, we find that randomly switching the supervision signal benefits self-supervised learning. Essentially, different modality networks learn different feature representations. They can filter different types of errors introduced by noisy pseudo labels. By this switchable operation, the error flows can be reduced mutually." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 572, + 379, + 585 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 572, + 379, + 585 + ], + "spans": [ + { + "bbox": [ + 306, + 572, + 379, + 585 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 594, + 548, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 548, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 548, + 715 + ], + "type": "text", + "content": "We explored how CLIP knowledge benefits 3D scene understanding in this work, termed CLIP2Scene. To efficiently transfer CLIP's image and text features to a 3D network, we propose a novel Semantic-driven Cross-modal Contrastive Learning framework including Semantic Regularization and Spatial-Temporal Regularization. For the first time, our pre-trained 3D network achieves annotation-free 3D semantic segmentation with decent performance. Besides, our method significantly outperforms state-of-the-art self-supervised methods when fine-tuning with labelled data." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 758 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 758 + ], + "type": "text", + "content": "7027" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 714 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 135 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 135 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 135 + ], + "type": "text", + "content": "[1] Zeynep Akata, Florent Perronnin, Zaid Harchaoui, and Cordelia Schmid. Label-embedding for attribute-based classification. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 819–826, 2013. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 137, + 288, + 190 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 137, + 288, + 190 + ], + "spans": [ + { + "bbox": [ + 53, + 137, + 288, + 190 + ], + "type": "text", + "content": "[2] Zeynep Akata, Scott E. Reed, Daniel Walter, Honglak Lee, and Bernt Schiele. Evaluation of output embeddings for fine-grained image classification. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2927-2936, 2015. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 194, + 288, + 247 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 194, + 288, + 247 + ], + "spans": [ + { + "bbox": [ + 53, + 194, + 288, + 247 + ], + "type": "text", + "content": "[3] Jens Behley, Martin Garbade, Andres Milioto, Jan Quenzel, Sven Behnke, Cyril Stachniss, and Jurgen Gall. Semantickitti: A dataset for semantic scene understanding of lidar sequences. In IEEE/CVF International Conference on Computer Vision, pages 9297-9307, 2019. 2, 5, 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 250, + 288, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 250, + 288, + 293 + ], + "spans": [ + { + "bbox": [ + 53, + 250, + 288, + 293 + ], + "type": "text", + "content": "[4] Max Bucher, Stéphane Herbin, and Frédéric Jurie. Generating visual representations for zero-shot classification. In IEEE/CVF International Conference on Computer Vision Workshop, pages 2666-2673, 2017. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 295, + 288, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 295, + 288, + 360 + ], + "spans": [ + { + "bbox": [ + 53, + 295, + 288, + 360 + ], + "type": "text", + "content": "[5] Holger Caesar, Varun Bankiti, Alex H Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giacarlo Baldan, and Oscar Beijbom. nuscenes: A multimodal dataset for autonomous driving. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11621-11631, 2020. 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 362, + 288, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 362, + 288, + 407 + ], + "spans": [ + { + "bbox": [ + 53, + 362, + 288, + 407 + ], + "type": "text", + "content": "[6] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In European Conference on Computer Vision, pages 213-229, 2020. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 407, + 288, + 462 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 407, + 288, + 462 + ], + "spans": [ + { + "bbox": [ + 53, + 407, + 288, + 462 + ], + "type": "text", + "content": "[7] Mathilde Caron, Ishan Misra, Julien Mairal, Priya Goyal, Piotr Bojanowski, and Armand Joulin. Unsupervised learning of visual features by contrasting cluster assignments. In Advances in Neural Information Processing Systems, volume 33, pages 9912-9924, 2020. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 464, + 288, + 508 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 464, + 288, + 508 + ], + "spans": [ + { + "bbox": [ + 53, + 464, + 288, + 508 + ], + "type": "text", + "content": "[8] Soravit Changpinyo, Wei-Lun Chao, Boqing Gong, and Fei Sha. Synthesized classifiers for zero-shot learning. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5327-5336, 2016. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 510, + 288, + 564 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 510, + 288, + 564 + ], + "spans": [ + { + "bbox": [ + 53, + 510, + 288, + 564 + ], + "type": "text", + "content": "[9] Nenglun Chen, Lingjie Liu, Zhiming Cui, Runnan Chen, Duygu Ceylan, Changhe Tu, and Wenping Wang. Unsupervised learning of intrinsic structural representation points. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9121-9130, 2020. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 566, + 287, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 566, + 287, + 599 + ], + "spans": [ + { + "bbox": [ + 48, + 566, + 287, + 599 + ], + "type": "text", + "content": "[10] Runnan Chen, Zhu Xinge, Nenglun Chen, Dawei Wang, Wei Li, Yuexin Ma, Ruigang Yang, and Wenping Wang. Referring self-supervised learning on 3d point cloud. 2021. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 600, + 287, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 600, + 287, + 644 + ], + "spans": [ + { + "bbox": [ + 48, + 600, + 287, + 644 + ], + "type": "text", + "content": "[11] Runnan Chen, Xinge Zhu, Nenglun Chen, Wei Li, Yuexin Ma, Ruigang Yang, and Wenping Wang. Zero-shot point cloud segmentation by transferring geometric primitives. arXiv preprint arXiv:2210.09923, 2022. 1, 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 646, + 287, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 646, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 646, + 287, + 689 + ], + "type": "text", + "content": "[12] Runnan Chen, Xinge Zhu, Nenglun Chen, Dawei Wang, Wei Li, Yuexin Ma, Ruigang Yang, and Wenping Wang. Towards 3d scene understanding by referring synthetic models. arXiv preprint arXiv:2203.10546, 2022. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 691, + 287, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 287, + 714 + ], + "type": "text", + "content": "[13] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 547, + 713 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "text", + "content": "of visual representations. In International Conference on Machine Learning, pages 1597-1607, 2020. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 95, + 547, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 95, + 547, + 128 + ], + "spans": [ + { + "bbox": [ + 308, + 95, + 547, + 128 + ], + "type": "text", + "content": "[14] Xinlei Chen and Kaiming He. Exploring simple siamese representation learning. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15750-15758, 2021. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 129, + 547, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 129, + 547, + 183 + ], + "spans": [ + { + "bbox": [ + 308, + 129, + 547, + 183 + ], + "type": "text", + "content": "[15] Ran Cheng, Ryan Razani, Ehsan Taghavi, Enxu Li, and Bingbing Liu. (af)2-s3net: Attentive feature fusion with adaptive feature selection for sparse semantic segmentation network. In IEEE Conference on Computer Vision and Pattern Recognition, pages 12547-12556, 2021. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 184, + 547, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 184, + 547, + 227 + ], + "spans": [ + { + "bbox": [ + 308, + 184, + 547, + 227 + ], + "type": "text", + "content": "[16] Ali Cheraghian, Shafin Rahman, Dylan Campbell, and Lars Petersson. Mitigating the Hubness Problem for Zero-Shot Learning of 3D Objects. arXiv preprint arXiv:1907.06371, 2019.2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 228, + 547, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 228, + 547, + 271 + ], + "spans": [ + { + "bbox": [ + 308, + 228, + 547, + 271 + ], + "type": "text", + "content": "[17] Ali Cheraghian, Shafin Rahman, Townim F Chowdhury, Dylan Campbell, and Lars Petersson. Zero-shot learning on 3d point cloud objects and beyond. International Journal of Computer Vision, 130(10):2364-2384, 2022. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 272, + 547, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 272, + 547, + 315 + ], + "spans": [ + { + "bbox": [ + 308, + 272, + 547, + 315 + ], + "type": "text", + "content": "[18] Ali Cheraghian, Shafin Rahman, and Lars Petersson. Zero-Shot Learning of 3D Point Cloud Objects. In IEEE International Conference on Machine Vision Applications, pages 1-6, 2019. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 316, + 545, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 316, + 545, + 360 + ], + "spans": [ + { + "bbox": [ + 308, + 316, + 545, + 360 + ], + "type": "text", + "content": "[19] Christopher Choy, JunYoung Gwak, and Silvio Savarese. 4d spatio-temporal convnets: Minkowski convolutional neural networks. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3075-3084, 2019. 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 361, + 547, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 361, + 547, + 415 + ], + "spans": [ + { + "bbox": [ + 308, + 361, + 547, + 415 + ], + "type": "text", + "content": "[20] Angela Dai, Angel Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5828-5839, 2017. 2, 5, 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 415, + 547, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 415, + 547, + 470 + ], + "spans": [ + { + "bbox": [ + 308, + 415, + 547, + 470 + ], + "type": "text", + "content": "[21] Berkan Demirel, Ramazan Gokberk Cinbis, and Nazli Ikizler-Cinbis. Attributes2classname: A discriminative model for attribute-based unsupervised zero-shot learning. In IEEE/CVF International Conference on Computer Vision, pages 1241-1250, 2017. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 471, + 547, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 471, + 547, + 514 + ], + "spans": [ + { + "bbox": [ + 308, + 471, + 547, + 514 + ], + "type": "text", + "content": "[22] Carl Doersch, Abhinav Gupta, and Alexei A Efros. Unsupervised visual representation learning by context prediction. In IEEE/CVF International Conference on Computer Vision, pages 1422-1430, 2015. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 515, + 547, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 515, + 547, + 568 + ], + "spans": [ + { + "bbox": [ + 308, + 515, + 547, + 568 + ], + "type": "text", + "content": "[23] Alexey Dosovitskiy, Jost Tobias Springenberg, Martin Ried-miller, and Thomas Brox. Discriminative unsupervised feature learning with convolutional neural networks. In Advances in Neural Information Processing Systems, volume 27, 2014. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 570, + 547, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 570, + 547, + 624 + ], + "spans": [ + { + "bbox": [ + 308, + 570, + 547, + 624 + ], + "type": "text", + "content": "[24] Whye Kit Fong, Rohit Mohan, Juana Valeria Hurtado, Lubing Zhou, Holger Caesar, Oscar Beijbom, and Abhinav Valada. Panoptic nuscenes: A large-scale benchmark for lidar panoptic segmentation and tracking. IEEE Robotics and Automation Letters, 7:3795-3802, 2022. 2, 5, 6" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 625, + 547, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 625, + 547, + 669 + ], + "spans": [ + { + "bbox": [ + 308, + 625, + 547, + 669 + ], + "type": "text", + "content": "[25] Chuang Gan, Ming Lin, Yi Yang, Yueting Zhuang, and Alexander Hauptmann. Exploring semantic inter-class relationships (sir) for zero-shot action recognition. In AAAI Conference on Artificial Intelligence, 2015. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 670, + 547, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 670, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 670, + 547, + 713 + ], + "type": "text", + "content": "[26] Biao Gao, Yancheng Pan, Chengkun Li, Sibo Geng, and Huijing Zhao. Are we hungry for 3d lidar data for semantic segmentation? a survey of datasets and methods. IEEE Transactions on Intelligent Transportation Systems, 2021. 1" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7028" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 714 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 288, + 139 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 288, + 139 + ], + "type": "text", + "content": "[27] Jean-Bastien Grill, Florian Strub, Florent Altché, Corentin Tallec, Pierre Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Avila Pires, Zhaohan Guo, Mohammad Gheshlaghi Azar, et al. Bootstrap your own latent-a new approach to self-supervised learning. In Advances in Neural Information Processing Systems, volume 33, pages 21271-21284, 2020. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 140, + 288, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 140, + 288, + 183 + ], + "spans": [ + { + "bbox": [ + 48, + 140, + 288, + 183 + ], + "type": "text", + "content": "[28] Yulan Guo, Hanyun Wang, Qingyong Hu, Hao Liu, Li Liu, and Mohammed Bennamoun. Deep learning for 3d point clouds: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence, 43(12):4338-4364, 2020. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 184, + 288, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 184, + 288, + 237 + ], + "spans": [ + { + "bbox": [ + 48, + 184, + 288, + 237 + ], + "type": "text", + "content": "[29] Bo Han, Quanming Yao, Xingrui Yu, Gang Niu, Miao Xu, Weihua Hu, Ivor Tsang, and Masashi Sugiyama. Co-teaching: Robust training of deep neural networks with extremely noisy labels. Advances in neural information processing systems, 31, 2018. 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 239, + 288, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 239, + 288, + 282 + ], + "spans": [ + { + "bbox": [ + 48, + 239, + 288, + 282 + ], + "type": "text", + "content": "[30] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9729-9738, 2020. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 283, + 288, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 283, + 288, + 327 + ], + "spans": [ + { + "bbox": [ + 48, + 283, + 288, + 327 + ], + "type": "text", + "content": "[31] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 770-778, 2016. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 327, + 288, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 327, + 288, + 370 + ], + "spans": [ + { + "bbox": [ + 48, + 327, + 288, + 370 + ], + "type": "text", + "content": "[32] Fangzhou Hong, Lingdong Kong, Hui Zhou, Xinge Zhu, Hongsheng Li, and Ziwei Liu. Unified 3d and 4d panoptic segmentation via dynamic shifting network. arXiv preprint arXiv:2203.07186, 2022. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 371, + 288, + 425 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 371, + 288, + 425 + ], + "spans": [ + { + "bbox": [ + 48, + 371, + 288, + 425 + ], + "type": "text", + "content": "[33] Yuenan Hou, Xinge Zhu, Yuexin Ma, Chen Change Loy, and Yikang Li. Point-to-Voxel Knowledge Distillation for LiDAR Semantic Segmentation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8479-8488, 2022. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 426, + 288, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 426, + 288, + 470 + ], + "spans": [ + { + "bbox": [ + 48, + 426, + 288, + 470 + ], + "type": "text", + "content": "[34] Siyuan Huang, Yichen Xie, Song-Chun Zhu, and Yixin Zhu. Spatio-temporal self-supervised representation learning for 3d point clouds. In IEEE/CVF International Conference on Computer Vision, pages 6535-6545, 2021. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 471, + 288, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 471, + 288, + 514 + ], + "spans": [ + { + "bbox": [ + 48, + 471, + 288, + 514 + ], + "type": "text", + "content": "[35] Ajay Jain, Matthew Tancik, and Pieter Abbeel. Putting nerf on a diet: Semantically consistent few-shot view synthesis. In IEEE/CVF International Conference on Computer Vision, pages 5885-5894, 2021. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 515, + 288, + 548 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 515, + 288, + 548 + ], + "spans": [ + { + "bbox": [ + 48, + 515, + 288, + 548 + ], + "type": "text", + "content": "[36] Sosuke Kobayashi, Eiichi Matsumoto, and Vincent Sitzmann. Decomposing nef for editing via feature field distillation. arXiv preprint arXiv:2205.15585, 2022. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 548, + 288, + 591 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 548, + 288, + 591 + ], + "spans": [ + { + "bbox": [ + 48, + 548, + 288, + 591 + ], + "type": "text", + "content": "[37] Elyor Kodirov, Tao Xiang, and Shaogang Gong. Semantic autoencoder for zero-shot learning. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4447-4456, 2017. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 592, + 288, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 592, + 288, + 635 + ], + "spans": [ + { + "bbox": [ + 48, + 592, + 288, + 635 + ], + "type": "text", + "content": "[38] Lingdong Kong, Youquan Liu, Runnan Chen, Yuexin Ma, Xinge Zhu, Yikang Li, Yuenan Hou, Yu Qiao, and Ziwei Liu. Rethinking range view representation for lidar segmentation. arXiv preprint arXiv:2303.05367, 2023. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 636, + 288, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 636, + 288, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 636, + 288, + 689 + ], + "type": "text", + "content": "[39] Lingdong Kong, Niamul Quader, and Venice Erin Liong. Conda: Unsupervised domain adaptation for lidar segmentation via regularized domain concatenation. In IEEE International Conference on Robotics and Automation, 2023. 1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 691, + 288, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 288, + 714 + ], + "type": "text", + "content": "[40] Lingdong Kong, Jiawei Ren, Liang Pan, and Ziwei Liu. Lasermix for semi-supervised lidar semantic segmentation." + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 713 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "text", + "content": "In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023. 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 96, + 547, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 96, + 547, + 140 + ], + "spans": [ + { + "bbox": [ + 307, + 96, + 547, + 140 + ], + "type": "text", + "content": "[41] Christoph H. Lampert, Hannes Nickisch, and Stefan Harmeling. Learning to detect unseen object classes by between-class attribute transfer. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 951-958, 2009. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 141, + 547, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 141, + 547, + 185 + ], + "spans": [ + { + "bbox": [ + 307, + 141, + 547, + 185 + ], + "type": "text", + "content": "[42] Christoph H. Lampert, Hannes Nickisch, and Stefan Harmeling. Attribute-based classification for zero-shot visual object categorization. IEEE Transactions on Pattern Analysis and Machine Intelligence, 36:453-465, 2014. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 186, + 547, + 219 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 186, + 547, + 219 + ], + "spans": [ + { + "bbox": [ + 307, + 186, + 547, + 219 + ], + "type": "text", + "content": "[43] Yan Li, Zhen Jia, Junge Zhang, Kaiqi Huang, and Tieniu Tan. Deep semantic structural constraints for zero-shot learning. In AAAI Conference on Artificial Intelligence, 2018. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 220, + 547, + 275 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 220, + 547, + 275 + ], + "spans": [ + { + "bbox": [ + 308, + 220, + 547, + 275 + ], + "type": "text", + "content": "[44] Yueh-Cheng Liu, Yu-Kai Huang, Hung-Yueh Chiang, Hung-Ting Su, Zhe-Yu Liu, Chin-Tang Chen, Ching-Yu Tseng, and Winston H Hsu. Learning from 2d: Contrastive pixel-topoint knowledge transfer for 3d pretraining. arXiv preprint arXiv:2104.04687, 2021. 1, 2, 3, 4, 6, 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 276, + 547, + 320 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 276, + 547, + 320 + ], + "spans": [ + { + "bbox": [ + 308, + 276, + 547, + 320 + ], + "type": "text", + "content": "[45] Björn Michele, Alexandre Boulch, Gilles Puy, Maxime Bucher, and Renaud Marlet. Generative zero-shot learning for semantic segmentation of 3d point clouds. In International Conference on 3D Vision, pages 992-1002, 2021. 1, 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 321, + 547, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 321, + 547, + 376 + ], + "spans": [ + { + "bbox": [ + 308, + 321, + 547, + 376 + ], + "type": "text", + "content": "[46] Ashish Mishra, M. Shiva Krishna Reddy, Anurag Mittal, and Hema A. Murthy. A generative model for zero shot learning using conditional variational autoencoders. In IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshop, pages 2269-22698, 2018. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 377, + 547, + 422 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 377, + 547, + 422 + ], + "spans": [ + { + "bbox": [ + 308, + 377, + 547, + 422 + ], + "type": "text", + "content": "[47] Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. Pointnet: Deep learning on point sets for 3d classification and segmentation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 652-660, 2017. 1, 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 422, + 547, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 422, + 547, + 487 + ], + "spans": [ + { + "bbox": [ + 308, + 422, + 547, + 487 + ], + "type": "text", + "content": "[48] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 489, + 547, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 489, + 547, + 544 + ], + "spans": [ + { + "bbox": [ + 308, + 489, + 547, + 544 + ], + "type": "text", + "content": "[49] Yongming Rao, Wenliang Zhao, Guangyi Chen, Yansong Tang, Zheng Zhu, Guan Huang, Jie Zhou, and Jiwen Lu. Denseclip: Language-guided dense prediction with context-aware prompting. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18082-18091, 2022. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 545, + 547, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 545, + 547, + 588 + ], + "spans": [ + { + "bbox": [ + 308, + 545, + 547, + 588 + ], + "type": "text", + "content": "[50] David Rozenberszki, Or Litany, and Angela Dai. Language-grounded indoor 3d semantic segmentation in the wild. In European Conference on Computer Vision, pages 125-141, 2022. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 590, + 547, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 590, + 547, + 645 + ], + "spans": [ + { + "bbox": [ + 308, + 590, + 547, + 645 + ], + "type": "text", + "content": "[51] Corentin Sautier, Gilles Puy, Spyros Gidaris, Alexandre Boulch, Andrei Bursuc, and Renaud Marlet. Image-to-lidar self-supervised distillation for autonomous driving data. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9891-9901, 2022. 1, 2, 3, 4, 5, 6, 7" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 647, + 547, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 647, + 547, + 690 + ], + "spans": [ + { + "bbox": [ + 308, + 647, + 547, + 690 + ], + "type": "text", + "content": "[52] Haotian Tang, Zhijian Liu, Shengyu Zhao, Yujun Lin, Ji Lin, Hanrui Wang, and Song Han. Searching efficient 3d architectures with sparse point-voxel convolution. In European Conference on Computer Vision, pages 685–702, 2020. 6" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 691, + 547, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 691, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 691, + 547, + 713 + ], + "type": "text", + "content": "[53] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7029" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 553 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 66, + 72, + 287, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 72, + 287, + 95 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 287, + 95 + ], + "type": "text", + "content": "Polosukhin. Attention is all you need. In Advances in Neural Information Processing Systems, volume 30, 2017. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 96, + 288, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 96, + 288, + 149 + ], + "spans": [ + { + "bbox": [ + 48, + 96, + 288, + 149 + ], + "type": "text", + "content": "[54] Yongqin Xian, Zeynep Akata, Gaurav Sharma, Quynh N. Nguyen, Matthias Hein, and Bernt Schiele. Latent embeddings for zero-shot classification. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 69-77, 2016. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 152, + 288, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 152, + 288, + 205 + ], + "spans": [ + { + "bbox": [ + 48, + 152, + 288, + 205 + ], + "type": "text", + "content": "[55] Saining Xie, Jiatao Gu, Demi Guo, Charles R Qi, Leonidas Guibas, and Or Litany. Pointcontrast: Unsupervised pretraining for 3d point cloud understanding. In European Conference on Computer Vision, pages 574-591, 2020. 2, 6, 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 208, + 288, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 208, + 288, + 262 + ], + "spans": [ + { + "bbox": [ + 48, + 208, + 288, + 262 + ], + "type": "text", + "content": "[56] Jianyun Xu, Ruixiang Zhang, Jian Dou, Yushi Zhu, Jie Sun, and Shiliang Pu. Rpvnet: A deep and efficient range-point-voxel fusion network for lidar point cloud segmentation. In IEEE/CVF International Conference on Computer Vision, pages 16024-16033, 2021. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 263, + 288, + 317 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 263, + 288, + 317 + ], + "spans": [ + { + "bbox": [ + 48, + 263, + 288, + 317 + ], + "type": "text", + "content": "[57] Lewei Yao, Jianhua Han, Youpeng Wen, Xiaodan Liang, Dan Xu, Wei Zhang, Zhenguo Li, Chunjing Xu, and Hang Xu. Detclip: Dictionary-enriched visual-concept paralleled pre-training for open-world detection. arXiv preprint arXiv:2209.09407, 2022. 2, 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 319, + 288, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 319, + 288, + 363 + ], + "spans": [ + { + "bbox": [ + 48, + 319, + 288, + 363 + ], + "type": "text", + "content": "[58] Éloi Zablocki, Patrick Bordes, Benjamin Piwowarski, Laure Soulier, and Patrick Gallinari. Context-aware zero-shot learning for object recognition. In International Conference on Machine Learning, pages 7292-7303, 2019. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 364, + 288, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 364, + 288, + 418 + ], + "spans": [ + { + "bbox": [ + 48, + 364, + 288, + 418 + ], + "type": "text", + "content": "[59] Renrui Zhang, Ziyu Guo, Wei Zhang, Kunchang Li, Xupeng Miao, Bin Cui, Yu Qiao, Peng Gao, and Hongsheng Li. Pointclip: Point cloud understanding by clip. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8552-8562, 2022. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 420, + 288, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 420, + 288, + 464 + ], + "spans": [ + { + "bbox": [ + 48, + 420, + 288, + 464 + ], + "type": "text", + "content": "[60] Zaiwei Zhang, Rohit Girdhar, Armand Joulin, and Ishan Misra. Self-supervised pretraining of 3d features on any point-cloud. In IEEE/CVF International Conference on Computer Vision, pages 10252-10263, 2021. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 465, + 288, + 497 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 465, + 288, + 497 + ], + "spans": [ + { + "bbox": [ + 48, + 465, + 288, + 497 + ], + "type": "text", + "content": "[61] Chong Zhou, Chen Change Loy, and Bo Dai. Extract free dense labels from clip. In European Conference on Computer Vision, pages 696-712, 2022. 2, 4, 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 498, + 288, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 498, + 288, + 553 + ], + "spans": [ + { + "bbox": [ + 48, + 498, + 288, + 553 + ], + "type": "text", + "content": "[62] Xinge Zhu, Hui Zhou, Tai Wang, Fangzhou Hong, Yuexin Ma, Wei Li, Hongsheng Li, and Dahua Lin. Cylindrical and asymmetrical 3d convolution networks for lidar segmentation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9939-9948, 2021. 1" + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "7030" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/CLIP2_ Contrastive Language-Image-Point Pretraining From Real-World Point Cloud Data/b3efe108-020c-4b4b-bc30-fbdcfe854136_content_list.json b/2023/CLIP2_ Contrastive Language-Image-Point Pretraining From Real-World Point Cloud Data/b3efe108-020c-4b4b-bc30-fbdcfe854136_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..d22975d39884092fa23893547ada30c34f6c90e6 --- /dev/null +++ b/2023/CLIP2_ Contrastive Language-Image-Point Pretraining From Real-World Point Cloud Data/b3efe108-020c-4b4b-bc30-fbdcfe854136_content_list.json @@ -0,0 +1,1393 @@ +[ + { + "type": "text", + "text": "CLIP $^2$ : Contrastive Language-Image-Point Pretraining from Real-World Point Cloud Data", + "text_level": 1, + "bbox": [ + 178, + 128, + 790, + 171 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yihan Zeng $^{1*}$ , Chenhan Jiang $^{2*}$ , Jiageng Mao $^{3}$ , Jianhua Han $^{1}$ , Chaoqiang Ye $^{1}$ , Qingqiu Huang $^{1}$ , Dit-Yan Yeung $^{2}$ , Zhen Yang $^{1}$ , Xiaodan Liang $^{4}$ , Hang $\\mathrm{Xu}^{1\\dagger}$", + "bbox": [ + 176, + 200, + 790, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Car, Pedestrian, Bicycle, Bus, Debris, Trailer, Table, Chair, Light, Cleaner, Closet, Backpack, Tire, Couch ...", + "bbox": [ + 316, + 270, + 867, + 282 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/cfe8c4c6d10f3a5dd955c4f4c255cbfd42736320d93dda788fc81c0f5a6ecab8.jpg", + "image_caption": [ + "Open-world Vocabularies", + "Figure 1. Illustration of our open-world recognition results. Benefiting from our $\\mathrm{CLIP}^2$ , the 3D representation is aligned to the openworld language representation, which enables flexible zero-shot transfer. Best viewed in colors." + ], + "image_footnote": [], + "bbox": [ + 81, + 290, + 488, + 465 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/e34e30f102ed2def75dd13d5cca9f46c554c1ec5db0ff05b03c664eb1a47f446.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 491, + 287, + 887, + 465 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 525, + 312, + 541 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Contrastive Language-Image Pre-training, benefiting from large-scale unlabeled text-image pairs, has demonstrated great performance in open-world vision understanding tasks. However, due to the limited Text-3D data pairs, adapting the success of 2D Vision-Language Models (VLM) to the 3D space remains an open problem. Existing works that leverage VLM for 3D understanding generally resort to constructing intermediate 2D representations for the 3D data, but at the cost of losing 3D geometry information. To take a step toward open-world 3D vision understanding, we propose Contrastive Language-Image-Point Cloud Pretraining $(CLIP^2)$ to directly learn the transferable 3D point cloud representation in realistic scenarios with a novel proxy alignment mechanism. Specifically, we exploit naturally-existed correspondences in 2D and 3D scenarios, and build well-aligned and instance-based text-image-point proxies from those complex scenarios. On top of that, we propose a cross-modal contrastive objective to learn semantic and instance-level aligned point cloud representa", + "bbox": [ + 73, + 559, + 468, + 838 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "tion. Experimental results on both indoor and outdoor scenarios show that our learned 3D representation has great transfer ability in downstream tasks, including zero-shot and few-shot 3D recognition, which boosts the state-of-the-art methods by large margins. Furthermore, we provide analyses of the capability of different representations in real scenarios and present the optional ensemble scheme.", + "bbox": [ + 500, + 513, + 890, + 614 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 500, + 652, + 630, + 667 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Powerful 3D point cloud representation plays a crucial role in various real-world applications, e.g., 3D object recognition and detection [10, 20, 31, 40, 44]. Compared to 2D images, 3D point cloud provides specific information like accurate geometry that is robust to illumination changes. However, current methods [25, 40] that learn 3D representations generally rely on the predefined number of object categories and require plenty of labor-intensive annotations. Those learned 3D representations are insufficient for safety-critical scenarios like self-driving which includes a long-tail class distribution far beyond the predefined taxonomy. Therefore, it is highly demanded to learn a transferable 3D representation equipped with zero-shot recognition ability in vocabulary scalable real-world scenes. Figure 1 shows an open-world recognition example by our $\\mathrm{CLIP}^2$ in", + "bbox": [ + 496, + 680, + 890, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Equal contribution. 1Huawei Noah's Ark Lab 2Hong Kong University of Science and Technology 3The Chinese University of Hong Kong 4Sun Yat-san University †Corresponding Author: xu.hang@huawei.com", + "bbox": [ + 75, + 851, + 467, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "15244", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "outdoor and indoor scenes, where the 3D objects can be classified with the correlation alignment between 3D representations and open-world vocabularies.", + "bbox": [ + 75, + 90, + 468, + 133 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The critical ingredient of open-world understanding is that the models learn sufficient knowledge to obtain general representations. To achieve this, recent Vision-Language Models (VLM) [14, 27, 38] leverage Internet-scale text-image pairs to conduct vision-language pretraining, which facilitates transferable 2D representation and demonstrates promising performance in 2D open-vocabulary tasks. However, 3D vision-language pretraining remains unexplored due to the limitation of existing 3D datasets in diversity and scale compared to the massive data sources in 2D counterparts [14, 15, 27, 38]. Though some recent works [12, 13, 43] try to avoid this problem by transferring the pretrained 2D VLM into the intermediate representation including projected image patches [12, 18] or depth maps [1, 43], those representations suffer from the loss of 3D geometric information and limited viewpoints under realistic scenarios. Especially the camera images are only sometimes available due to the sensor failure in 3D scenes. We believe the 3D representation based on original point cloud data retains most information and is the optimal solution for 3D real world understanding, which requires a rethink of learning the transferable 3D representation under realistic scenarios.", + "bbox": [ + 75, + 136, + 470, + 458 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To this end, we propose a Contrastive Language-Image-Point cloud Pretraining framework, short for $\\mathrm{CLIP}^2$ , which directly aligns 3D space with broader raw text and advances the 3D representation learning into an open-world era. Our learning process can be decomposed into two stages: Firstly, we introduce a Triplet Proxy Collection to alleviate the limitation of accessible pretraining data by constructing language-image-point triplets from real-world scenes. Since the large-scale realistic 3D datasets for outdoor driving [2,19] and indoor scenarios [9,32] are collected in open-world, it contains huge amounts of realistic objects that vary in semantics and diversity. Thus we consider them as potential pretraining data sources without extra human supervision. Specifically, we propose \"Proxy\" instances as the bridges between language descriptions, 2D images and 3D point clouds. Enabled by a well-aligned VLM, a scalable caption list and the geometry transformation between 2D and 3D, we automatically create more than 1 million triplets to facilitate pretraining. Secondly, we further propose a Cross-Modal Pretraining scheme to jointly optimize the feature space alignments of three modalities, i.e., point cloud, language and image. It contains both the contrastive learning objective of semantic-level text-3D correlation and instance-level image-3D correlation, which contributes to better transferability of learned 3D representation.", + "bbox": [ + 75, + 459, + 468, + 825 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We study the transferable capability of $\\mathrm{CLIP}^2$ by benchmarking the zero-shot recognition performance on four popular indoor and outdoor real-world datasets, and find a significant improvement over current methods, achieving Top1 accuracy $61.3\\%$ on SunRGBD [32], $43.8\\%$ on ScanNet [9]),", + "bbox": [ + 75, + 825, + 468, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "28.8% on nuScenes [2] and 56.0% on ONCE [19]. For a fair comparison with existing methods [1, 13, 36, 43], we conduct zero-shot and few-shot classification on single object dataset ScanObjectNN [34] and find consistent dominance, 16.1% relative improvement on zero-shot classification over previous state-of-the-art method [13]. To validate the vocabulary-increasing ability of $\\mathrm{CLIP}^2$ , we report the quantity results and visualizations to show the improved discovery of the long-tail categories. Moreover, we make ablations and analysis on different representations, and investigate ensembling alternatives to merge complementary knowledge of all available representations in realistic applications. Our contributions can be summarized as follows:", + "bbox": [ + 496, + 90, + 890, + 282 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose a novel $\\mathrm{CLIP}^2$ framework that aligns 3D space with open-world language representation, facilitating zero-shot transfer in realistic scenarios.", + "- We present a Triplet Proxies Collection scheme in real-world scenes, which alleviates the shortage of text-3D data sources and facilitates the pretraining methods.", + "- $\\mathrm{CLIP}^2$ jointly optimizes the correlation alignment between point cloud, language and image by proposed cross-modal pretraining mechanism, which enhances the transferability of learned 3D representation.", + "- Our $\\mathsf{CLIP}^2$ achieves the state-of-the-art zero-shot transfer performance on 5 datasets (indoor/outdoor scenes and single-object) and shows quality results on vocabulary-increasing discovery in real world." + ], + "bbox": [ + 517, + 292, + 890, + 527 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 540, + 640, + 555 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Vision-Language Model. Large vision language models (VLM) [14, 15, 27, 38] have demonstrated successful performance in downstream zero-shot tasks with the learned transferable 2D representations. CLIP [27] and ALIGN [14] push the limit by collecting Internet-scale image-text pairs and then learning the correlation alignment between image and language feature space with contrastive pretraining objectives. Those models can be directly transferred to zero-shot 2D recognition and achieve impressive results. Recent DetClip [38] learns to align image patches to test phrases after pretraining under hybrid supervision from detection, grounding and image-text pair data, which extends the ability to localize open-vocabulary 2D proposals in images. In this paper, we attempt to transfer the open-vocabulary ability of pre-trained VLM to the 3D domain, making language applicable to zero-shot point cloud recognition.", + "bbox": [ + 496, + 559, + 890, + 794 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Zero-shot/Open-world Learning in 3D. Recognizing 3D objects with a large vocabulary is necessary for safety-critical autonomous driving and robotic tasks, yet remains under-explored. Cheraghian et al. [5-8] first attempt to associate PointNet [25] feature with category semantic information via a projection function, and separately proposed an unsupervised skewness loss [5] to mitigate the hubness", + "bbox": [ + 496, + 797, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "15245", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/c9d4e4112331aca8c40b6ed1534e6c7f1bd3717b4519fc73bd58938f66cf5179.jpg", + "image_caption": [ + "Figure 2. Overview of $\\mathbf{CLIP}^2$ framework. The main components contain two parts, the Triplet Proxy Collection and the Cross-Modal Pretraining. The defined Triplet Proxy set $\\mathcal{D}_{\\mathrm{proxy}}$ consists of language captions $\\mathbf{X}^T$ , corresponding image instances $\\mathbf{X}^I$ and raw 3D point cloud instances $\\mathbf{X}^P$ , which come from the free data source under realistic scenarios without any labeling labor. On top of that, we pretrain a point cloud encoder $E^P$ with the cross-modal contrastive learning objective. Equipped with $\\mathbf{CLIP}^2$ , the learned 3D point cloud representation $F^P$ is well aligned to the language representation, which facilitates downstream zero-shot 3D transfer tasks in the real world." + ], + "image_footnote": [], + "bbox": [ + 86, + 88, + 890, + 361 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "problem. The transductive case [6] is discussed in which extends [5] using a triplet loss. Notably, the above works conduct experiments on synthetic datasets and need to divide datasets into \"seen\" categories as training data and \"unseen\" categories as testing data. Thus they are not suitable for realistic scenarios due to the domain gap between synthetic and real-world data, as well as the limited vocabulary-increasing ability. Recently, inspired by the success of VLMs [14, 27] in 2D tasks, some works [13, 43] propose to transfer the zero-shot recognition ability of pretrained CLIP [27] into 3D area. PointCLIP [43] directly projects point cloud into multi-view depth maps as image-like data input for pretrained CLIP to make classification predictions. While CLIP2Point [13] trains an image-depth embedding on ShapeNet [39] to better align the depth representation to the pretrained image space of CLIP. However, depth maps lost plenty of geometry information of the original point cloud data structure, resulting in poor performance especially in realistic scenarios. By contrast, we aim to learn transferable 3D representation based on the original point cloud data structure in realistic scenarios.", + "bbox": [ + 75, + 465, + 472, + 773 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3D Representation Learning. Much progress has been made in learning a comprehensive 3D representation in an unsupervised manner. Most works [1, 16, 17, 23, 30, 37, 41, 42] follow the paradigm that conducts pretraining on unlabeled datasets and then finetunes on the limited downstream annotations. Though the improved transferability of 3D representation, they can not be directly transferred to zero-shot tasks with open-world vocabularies. In this work, we con", + "bbox": [ + 75, + 782, + 472, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "duct language-image-point cloud pretraining, which learns transferable 3D representation aligned to open-vocabulary language space to facilitate the zero-shot transfer.", + "bbox": [ + 496, + 465, + 893, + 512 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 498, + 521, + 591, + 537 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we introduce $\\mathrm{CLIP}^2$ to learn a transferable 3D point cloud representation with arbitrary category recognition ability under realistic scenarios, illustrated in Figure 2. We will first present the Triplet Proxy Collection in Section 3.1, which utilizes a pretrained VLM and geometric transformation to obtain language-image-point triplets from real-world scenes. Then we will elaborate Cross-Modal Contrastive Pretraining mechanism in Section 3.2, which jointly optimizes the alignment correlations between language, image and point cloud feature space.", + "bbox": [ + 496, + 546, + 893, + 695 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Triplet Proxy Collection", + "text_level": 1, + "bbox": [ + 498, + 700, + 723, + 718 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Inspired by the significant performance of 2D VLMs on open-vocabulary tasks, we aim to develop 3D vision-language pretraining to facilitate category-increasing capacity for real-world scenarios. However, the core challenge is the shortage of pretraining data. Compared to the 2D vision-language pretraining framework CLIP [27], which takes more than 400M image-language pairs from the Internet, the largest 3D single-object dataset ShapeNet [39] only contains 50K CAD models with 55 categories. In addition to the insufficiency of data scale, pretraining on such synthetic data fails to transfer well in the real world due to the huge domain gap. Enlightened by the recent emer", + "bbox": [ + 496, + 724, + 893, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "15246", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/a8da405e6ae9f064a476ddb3c2101091ec79bc7b6889540c99430eeeaf5a9e32.jpg", + "image_caption": [ + "Figure 3. Illustration of three representation modals of two 3D objects examples under indoor and outdoor scenarios." + ], + "image_footnote": [], + "bbox": [ + 78, + 88, + 467, + 195 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "gence of large-scale point cloud datasets collected in indoor [9, 32] and outdoor scenarios [2, 19], we observe that those naturally-collected datasets potentially contain vast amounts of open-world objects that vary in semantics and diversity. Considering the data collection itself is cheap except for laborious annotation, we novelly take leverage of those available datasets without human annotations as a practical yet effective pretraining data source.", + "bbox": [ + 75, + 258, + 468, + 376 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Specifically, given the realistic scene data $S = \\{(P_s, I_s)_{s=1}^{|S|}\\}$ , where $P_s \\in \\mathbb{R}^{N_P \\times 3}$ and $I_s \\in \\mathbb{R}^{N_I \\times H \\times W \\times 3}$ are corresponding 3D point clouds and images of scene $s$ , we propose a novel concept, Proxy, as the bridge between language, image and 3D point cloud. As illustrated in Figure 2, equipped by those proxy instances, we can automatically collect a massive number of language-image-point cloud pairs $D_{\\mathrm{proxy}}$ in the format of proxies under open-world scenes. We detail the process as follows.", + "bbox": [ + 76, + 376, + 468, + 508 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Language Proxy. We set the language proxies $\\mathbf{X}^T\\in \\mathbb{R}^V$ as a raw text list from the 2D open-world dataset [11], where $V = 1206$ denotes the vocabulary size of language proxies. Image Proxy. Next, we obtain the image proxies $\\mathbf{X}^I$ by an open vocabulary detector DetCLIP [38], denoted as $M$ which is trained with open-world data and performs openset detection. Concretely, given language proxies $\\mathbf{X}^T$ and input scene image $I_{s}$ , we extract corresponding image proposals as image proxies $X_{s}^{I}$ with $M$ by the similarity between input language embeddings and proposal features as", + "bbox": [ + 76, + 508, + 468, + 657 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\left\\{X _ {s} ^ {I} \\right\\} _ {s \\in | \\mathcal {S} |} = \\mathrm {M} \\left(\\left\\{I _ {s} \\right\\} _ {s \\in | \\mathcal {S} |}, \\mathbf {X} ^ {T}\\right). \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 161, + 670, + 468, + 690 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3D Proxy. We exploit the naturally-existed geometry relations between 2D and 3D scenes to obtain 3D proxies $\\mathbf{X}^P$ , which consists of point cloud instances corresponding to image proposals in $\\mathbf{X}^I$ . We simplify the geometry transformation as $\\mathrm{G}(\\cdot)$ and formulate the relations as:", + "bbox": [ + 76, + 696, + 468, + 771 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nX _ {i} ^ {P} = \\mathrm {G} \\left(X _ {i} ^ {I}\\right). \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 220, + 780, + 468, + 799 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Detailedly, for indoor scenes equipped with RGB-D sensors, we first remove the background pixels by unsupervised segmentation algorithm [28] for each image proxy $X_{s,i}^{I}$ , $i \\in |X_s^I|$ . Since depth information is known, we then transform the segmented pixels from uvd coordinate $X_{s,i}^{I,uvd} \\in \\mathbb{R}^{n,3}$ to xyz coordinate $X_{s,i}^{P,xyz} \\in \\mathbb{R}^{n,3}$ as a 3D", + "bbox": [ + 76, + 810, + 468, + 902 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/c658027c3b01323f7c826a78da3e707d1381ceadaf174e4edb9bb24c7406eded.jpg", + "image_caption": [ + "Figure 4. Comparison of different pretraining strategies. (a) CLIP aligns image and language embedding space [27] as $L_{TI}$ based on large-scale text-image pairs. (b) PointClip [27] aligns projected depth map to CLIP language space as $L_{TD}$ . (c) Clip2Point aligns depth map to CLIP image space as $L_{ID}$ . (d) our $\\mathrm{CLIP}^2$ aligns original 3D point cloud to both CLIP language space and image space via cross-modal objective $L_{CM}$ ." + ], + "image_footnote": [], + "bbox": [ + 506, + 87, + 883, + 194 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "point cloud proxy with the given camera parameters. For outdoor scenes captured by LiDAR sensors, we first create a 3D frustum for each image proxy by extruding the 2D image proposal into 3D space following [22, 24]. Then we conduct DBSCAN algorithm [29] within frustum and select the point cloud cluster as the point proxy $X_{s,i}^{P,xyz}$ .", + "bbox": [ + 496, + 325, + 890, + 416 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Eventually, we construct Triplet Proxy $\\mathcal{D}_{\\mathrm{proxy}} = \\{\\mathbf{X}^T,X_s^I,X_s^P\\}_{s = 1}^{|\\mathcal{S}|}$ by combining corresponding language proxies $\\mathbf{X}^T$ , image proxies $\\mathbf{X}^I$ and 3D proxies $\\mathbf{X}^P$ , where $\\mathbf{X}^I = \\{X_s^I\\}_{s = 1}^{|\\mathcal{S}|}$ and $\\mathbf{X}^P = \\{X_s^P\\}_{s = 1}^{|\\mathcal{S}|}$ . $220\\mathrm{K}$ and $1.4\\mathrm{M}$ proxy triplets are formed for indoor and outdoor scenes, respectively. More details can be found in the appendix.", + "bbox": [ + 498, + 417, + 890, + 512 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Cross-Modal Contrastive Pretraining", + "text_level": 1, + "bbox": [ + 500, + 523, + 821, + 539 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "With the triplet proxies $\\mathcal{D}_{\\mathrm{proxy}}$ , a straightforward pretraining objective is forcing the alignment between the embedding spaces of point cloud $X_{i}^{P}$ and language $X_{i}^{T}$ from scratch. However, it might not promise good transferability of learned representation, since the number of language-image-point pretraining data triplets remains two orders of magnitude smaller than the language-image pairs adopted by CLIP [27] and the vocabulary size is much more limited. Therefore, we design to learn the correlation alignment based on the pretrained embedding space of CLIP.", + "bbox": [ + 496, + 547, + 890, + 694 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The comparison of current pretraining strategies [13,43] is illustrated in Figure 4, which is a series of 3D variants of CLIP. Notably, both existing methods exploit projected depth map as the intermediate representation of point cloud, which are respectively learned to align to language space [43] and image space [13]. Intuitively, as illustrated in Figure 3, depth representation lost plenty of geometry information compared to the original point cloud, especially in outdoor scenarios. Moreover, images are sometimes unavailable for 3D objects. Thus we conduct pretraining on original 3D point cloud data as an optimal representation.", + "bbox": [ + 496, + 694, + 890, + 856 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Toward learning more transferable representation, we introduce a cross-modal contrastive learning objective to jointly optimize the correlation alignment across lan", + "bbox": [ + 496, + 856, + 890, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "15247", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "guage, image and point cloud, including Semantic-Level Language-3D Alignment and Instance-Level Image-3D Alignment. Specifically, the overall architecture of $\\mathrm{CLIP}^2$ , shown in Figure 2, contains language encoder $E_{\\theta}^{T}$ , point cloud encoder $E_{\\theta}^{P}$ and visual encoder $E_{\\theta}^{I}$ , which respectively embed the triplet proxies into text feature $f^{T} \\in \\mathcal{R}^{1 \\times C_{T}}$ , point cloud feature $f^{P} \\in \\mathcal{R}^{1 \\times C_{P}}$ and image feature $f^{I} \\in \\mathcal{R}^{1 \\times C_{I}}$ , where $C$ is the embedding dimension.", + "bbox": [ + 75, + 90, + 468, + 207 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Semantic-Level Language-3D Alignment. In order to inherit the open-world recognition ability from pretrained CLIP [27], we align the point cloud feature $f^{P}$ with text embedding $f^{T}$ from well-trained CLIP with Language-Point Proxy $\\{X_{i}^{T}, X_{i}^{P}\\}$ input. We replace classname in the prompts, like \"point cloud of a {classname} .\" with raw text in proxy $X_{i}^{T}$ as language sentences. The core idea is to drive the feature centroids of 3D instances and the corresponding text prompt closer. We compute the loss function of between of language proxy and point cloud proxy as:", + "bbox": [ + 75, + 208, + 470, + 354 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nl (i, T, P) = - \\log \\frac {\\exp \\left(f _ {i} ^ {T} \\cdot f _ {i} ^ {P} / \\tau\\right)}{\\exp \\left(f _ {i} ^ {T} \\cdot f _ {i} ^ {P} / \\tau\\right) + \\sum_ {j \\in N , X _ {j} ^ {T} \\neq X _ {i} ^ {T}} \\exp \\left(f _ {i} ^ {T} \\cdot f _ {j} ^ {P} / \\tau\\right)}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 76, + 359, + 472, + 412 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $N$ is the mini-batch size, $\\tau$ is the temperature coefficient. Within a training mini-batch, the language-3D alignment objective $L(T,P)$ can be described as:", + "bbox": [ + 75, + 414, + 468, + 459 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nL (T, P) = \\frac {1}{N} \\sum_ {i \\in N} l (i, T, P). \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 181, + 467, + 468, + 500 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Instance-Level Image-3D Alignment. In addition to the alignment between semantic language and 3D proxy instances, we further introduce the contrastive alignment between instance-wise image proxy and 3D proxy instances. Note that the instance-aware visual concept has been well-studied in the embedding space of pretrained CLIP. We believe instance-sensitive learning can contribute to further correlation learning and benefits to the transferability of learned 3D representation. The contrastive aligned objective $L(I,P)$ across point cloud and image is formulated as:", + "bbox": [ + 75, + 507, + 468, + 654 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nl (i, I, P) = - \\log \\frac {\\exp \\left(f _ {i} ^ {I} \\cdot f _ {i} ^ {P} / \\tau\\right)}{\\exp \\left(f _ {i} ^ {I} \\cdot f _ {i} ^ {P} / \\tau\\right) + \\sum_ {j \\in N , j \\neq i} \\exp \\left(f _ {i} ^ {J} \\cdot f _ {j} ^ {P} / \\tau\\right)}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 78, + 674, + 467, + 728 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nL (I, P) = \\frac {1}{N} \\sum_ {i \\in N} l (i, I, P). \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 176, + 737, + 467, + 773 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Finally, we obtain the resultant cross-modal contrastive learning objective $L_{CM}(T,I,P)$ as the combination of $L(T,P)$ and $L(I,P)$ , where both alignments of semantic-level text-3D correlation and instance-level image-3D correlation are injected:", + "bbox": [ + 76, + 779, + 468, + 853 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nL _ {C M} (T, I, P) = \\lambda_ {1} L (T, P) + \\lambda_ {1} L (I, P), \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 140, + 866, + 467, + 881 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where the hyper-parameters $\\lambda_{1}$ and $\\lambda_{2}$ are both set to 0.5.", + "bbox": [ + 76, + 885, + 457, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiment", + "text_level": 1, + "bbox": [ + 500, + 90, + 624, + 107 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this section, we evaluate $\\mathrm{CLIP}^2$ on realistic indoor and outdoor scenarios. We report the zero-shot transfer results on various datasets [2, 9, 32-34] and make further analysis on the designs of pretraining strategy.", + "bbox": [ + 496, + 114, + 890, + 174 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Zero-shot Transfer", + "text_level": 1, + "bbox": [ + 498, + 181, + 681, + 195 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Setting. After pretraining, natural language is applied to reference learned 3D representation to enable following zero-shot transfer tasks. (i) Zero-Shot Recognition: we evaluate zero-shot recognition performance for realistic objects, where $K$ category names are transferred to text prompt \"point cloud of {CLASS}\" to encode the text features $F_{K} \\in \\mathbb{R}^{K \\times C}$ . Then the classification logits are calculated with the 3D feature $f^{P}$ and text features as:", + "bbox": [ + 496, + 199, + 890, + 316 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\log_ {i} = \\operatorname {s o f t m a x} \\left(f _ {i} ^ {P} \\left(F _ {K}\\right) ^ {T}\\right). \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 591, + 323, + 890, + 342 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We present the results under both indoor and outdoor scenarios in Table 1, Table 2 and Table 5, as well as the object-level benchmark in Table 6. (ii) Open-vocabulary recognition: we enlarge the category vocabularies of ScanNet to 249 and 384 to study the open-vocabulary recognition ability in Table 3. (iii) Open-vocabulary localization: we study the open-vocabulary localization ability by localizing open-world 3D objects with our proxy generation process and then classifying them with our learned 3D representation, of which the visualization is illustrated in Figure 5 and evaluation results are reported in Table 4. Notably, we investigate representation ensembling alternatives to enable knowledge merging of all available representations for realistic applications, illustrated in Table 8.", + "bbox": [ + 496, + 349, + 890, + 555 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1.1 Indoor Scenarios", + "text_level": 1, + "bbox": [ + 498, + 561, + 671, + 575 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Datasets and details. We adopt the widely used indoor 3D dataset SUN RGB-D [32] as the realistic indoor scenario that provides pretraining data source, a single-view RGB-D dataset consisting of $\\sim 10\\mathrm{K}$ scenes. To validate the transferability of learned 3D representation, we also evaluate another popular indoor 3D dataset ScanNet [9], which contains $\\sim 1.5\\mathrm{K}$ scenes of 3D reconstructed meshes. We remove objects in ScanNet with less than 5 points, leaving 384 noisy categories. For open-vocabulary recognition, we evaluate performance on the ScanNet 384-class set and a 249-class merged set. In addition to the scene-wise indoor dataset, we conduct evaluations on ScanObjectNN [34], which collects $\\sim 3\\mathrm{K}$ individual realistic objects with 15 categories and is applied in the previous zero-shot evaluation [13, 43]. During the proxy collection process, we empirically set $\\epsilon = 0.3$ in [38] as a tradeoff between filtering FPs and preserving TPs to generate image proxies. Considering the occurrence frequencies of different indoor categories vary a lot, we adopt the class balance strategy [4] to mitigate the class imbalance. During pretraining process, we adopt [26] as point cloud encoder and set the overall training epoch number to 100.", + "bbox": [ + 496, + 580, + 890, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "15248", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/d3341ade92a01b6750c52db0feb52b14e628493e390d9e0a8d93025414504f71.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodAvg.BedBookshelfChairDeskSofaTabletoiletBathhubDresserNight Stand
PointClip [43]11.50.094.00.00.00.014.70.00.06.10.0
Clip2Point [13]18.610.920.664.334.413.814.126.20.01.40.0
PointClip [43] w/ TP.38.045.3100.062.548.544.44.855.216.33.30
Clip2Point [13] w/ TP.56.978.087.636.236.664.737.482.177.567.61.2
CLIP²61.384.075.570.747.375.533.886.265.371.82.4
CLIP² w/ En.69.687.394.370.754.179.347.091.785.782.23.6
", + "bbox": [ + 84, + 89, + 885, + 205 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/cf3afe102667b94e603f4ba677287620485560fd03c2b37b1ae5f2a6f55bde1f.jpg", + "table_caption": [ + "Table 1. Zero-shot recognition results in SUN RGB-D. Avg.: the mean average Top1 accuracy across all categories. TP.: our Triplet Proxy set. En.: our optional ensembling scheme for inference." + ], + "table_footnote": [], + "table_body": "
MethodAvg.CabBedChairSofaTablDoorWindBksfPicCntrDeskCurtFridgBathShowrToilSink
PointClip [43]6.30.00.00.00.00.70.00.091.80.00.00.015.00.00.00.00.00.0
Clip2Point [13]24.90.020.885.143.326.569.90.020.91.731.727.00.01.646.50.022.425.6
PointClip [43] w/ TP.26.155.70.072.85.05.11.70.077.20.00.051.70.30.00.040.385.349.2
Clip2Point [13] w/ TP.35.23.011.845.127.610.561.52.671.90.333.629.94.711.572.292.486.134.0
CLIP²38.567.232.669.342.318.319.14.062.61.412.752.840.19.159.741.071.045.5
", + "bbox": [ + 84, + 253, + 885, + 335 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/7d54b6c0708993235172dc22d2d5d5f8017ed63f7a3a8feee6d5f58147fbdaf0.jpg", + "table_caption": [ + "Table 2. Zero-shot recognition in ScanNet. Avg.: the mean average Top1 accuracy across all categories. TP.: our Triplet Proxy set." + ], + "table_footnote": [], + "table_body": "
MethodTop5 Acc.
384cls.249cls.
[43]0.30.4
[13]6.47.0
CLIP222.031.7
", + "bbox": [ + 83, + 371, + 223, + 435 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/7c845af3cb741e054f9702aff52e6a54d8991f99880fa7b0d7bf3789526a54e7.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodmAP25IN AR25mIoUOUT P.R.
[21]1.3----
[18]13.037.7---
CLIP212.743.052.86.787.4
", + "bbox": [ + 230, + 371, + 454, + 435 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 3. Recognition results. Table 4. Zero-shot localization results of vocabulary ex-sults. IN: indoor scenario SUN RGBpansion in ScanNet. D. OUT: outdoor scenario nuScenes.", + "bbox": [ + 76, + 439, + 460, + 481 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Quantity results. For zero-shot recognition task, we take two recent works as our baselines, i.e. PointClip [43] and Clip2Point [13], which study the zero-shot classification task on 3D object-level benchmarks [34, 35] by leveraging pretrained CLIP with projected depth maps. Focusing on the real-world scenarios, we conduct comparison not only on the realistic object-level [34] as illustrated in Table 6 but also on the scene-level datasets shown in Table 1 and Table 2, where the evaluation follows the common classes split in [10, 44] and reports the instance Top1 accuracy of each class. As shown in tables, our $\\mathrm{CLIP}^2$ can outperform baselines on all benchmarks by large margins. Besides, we apply our triplet proxy generation mechanism (TP.) to baseline methods, and achieve considerable improvements on SUN RGB-D and ScanNet by $26.5\\%$ and $19.8\\%$ for Point-Clip, $38.3\\%$ and $10.3\\%$ for Clip2Point. On the one hand, the contrasts demonstrate the effectiveness of our triplet proxies for open-world understanding. On the other hand, our learned 3D representation is superior in 3D object recognition by retaining more 3D-specific information than depth representation. Besides, we present the optional ensembling scheme (En.) when camera images are available, which can take advantage of multi-modal knowledge and further boost the performance by $8.3\\%$ . To further validate the open-vocabulary recognition ability, we conduct evaluation on a larger category set of ScanNet in Table 3 and report the instance Top5 accuracy, which illustrates the superior", + "bbox": [ + 75, + 503, + 472, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "ity of our $\\mathrm{CLIP}^2$ when vocabulary increases. Beyond that, $\\mathrm{CLIP}^2$ is also equipped with zero-shot 3D localization ability by proxy generation. On indoor scenario SUN RGB-D, we compare with a SOTA indoor 3D detector 3DETR [21] and a recent work OV3D [18] that studies open-vocabulary detection, where evaluation is conducted on the same \"unseen\" split in [18]. Since $\\mathrm{CLIP}^2$ do not fit the tight bounding boxes of point cloud instances, we estimate the maximum bounding box of proxies and GT instances to conduct evaluation following the same metrics $\\mathrm{mAP}_{25}$ and $\\mathrm{AR}_{25}$ in [18], as shown in Table 4. Notably, compared to baseline works that train on \"seen\" 3D annotations and test on \"unseen\" categories, we have no access to any 3D annotations yet achieve comparable localization ability, which yields $5.3\\%$ $\\mathrm{AR}_{25}$ improvement over OV3D [18]. We further evaluate segmentation results in Table 4.", + "bbox": [ + 496, + 375, + 890, + 609 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Quality results. The visualization results of $\\mathrm{CLIP}^2$ under a indoor scene of SUN RGB-D [32] is shown in Figure 5(a). Our triplet proxy generation process can localize open-world 3D objects in a point cloud scene. Moreover, the 3D representation learned from our cross-modal pretraining can provide more accurate classification results for 3D instances by exploiting original point cloud, which corrects the mistaken \"People\" prediction in image to \"Picture\" by considering the geometry information.", + "bbox": [ + 496, + 609, + 892, + 743 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1.2 Outdoor Scenarios", + "text_level": 1, + "bbox": [ + 498, + 750, + 687, + 762 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Datasets and details. We exploit a prevalent large-scale 3D dataset nuScenes [2] as the outdoor data source and extra validate the performance on the ONCE dataset [19]. The nuScenes dataset consists of $\\sim 28\\mathrm{K}$ frames with 10 categories, while ONCE contains 6 annotated sequences with 5 categories. Similarly, we set the $\\epsilon = 0.3$ for image proxies collection and adopt the class balance strategy [4].", + "bbox": [ + 496, + 766, + 890, + 869 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Quantity results. Since the outdoor point cloud is collected by LiDAR sensors, it has a wider perception range", + "bbox": [ + 498, + 869, + 890, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "15249", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/9f1d2bd50c49e4436b3d4ca69428bf693f8fc4be2e5eb924bf82e31cf81978ac.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodAvg.nuScenesONCE
CarTruckBusPed.BicycleTrailerC.V.Motor.BarrierT.C.CarCyc.Ped.TruckBus
PointClip [43]11.70.00.00.029.141.83.40.00.142.50.00.013.879.20.27.61
Clip2Point [13]12.40.40.30.113.531.01.89.31.566.20.317.311.895.435.74.0
PointClip [43] w/ TP.28.318.805.574.017.957.01.94.52.129.751.89.299.85.046.8
Clip2Point [13] w/ TP.33.026.716.851.245.215.813.920.05.710.534.239.427.895.540.651.7
CLIP²37.841.941.322.540.321.120.624.822.417.335.352.727.377.778.544.0
", + "bbox": [ + 81, + 89, + 890, + 186 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d4d7bd0c222aa751cda04a21ff7675f53f7b9569fb4b711efc61ca0fe9b02bfd.jpg", + "image_caption": [ + "Figure 5. Visualizations of the zero-shot localization and recognition results by $\\mathrm{CLIP}^2$ under open-world (a) indoor realistic scene [32] and (b) outdoor scenes [2]. Notably, the whole pipeline of $\\mathrm{CLIP}^2$ not only has no access to human annotations, but also enables the open-world vocabularies beyond groundtruth annotations, such as 'Picture' in (a) and 'Plastic bag', 'Tire' in (b). Best viewed in colors." + ], + "image_footnote": [], + "bbox": [ + 81, + 234, + 436, + 411 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/7acfa580419ef00bd6b679fccb2d957142c4c7f782fc29dcc91736b580390ac4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 439, + 234, + 890, + 411 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/61b581bd2423c1782fe57b6ce0bee93eef5a4279d446c209605f98d694dfd3f7.jpg", + "table_caption": [ + "Table 5. Zero-shot recognition results in outdoor scenario: nuScenes (Left) and ONCE (Right). TP.: our Triplet Proxy set. Avg.: the mean average Top1 accuracy across all categories of two benchmarks." + ], + "table_footnote": [], + "table_body": "
MethodZS15-way10-way
4-S8-S16-S10-S20-S
PointNet++ [25]-41.047.655.0--
PointCLIP [43]15.446.050.055.6--
Clip2Point [13]23.3-----
CrossPoint [1]----58.7±1.864.6±1.2
CLIP239.151.359.662.560.6±2.566.3±3.2
", + "bbox": [ + 84, + 479, + 465, + 577 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 6. Zero-shot and Few-shot classification results on ScanObjectNN. ZS: zero-shot. K-way N-shot: few-shot settings.", + "bbox": [ + 75, + 580, + 468, + 609 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "than RGB-D but leads to sparse distribution. Thus the projected depth representation of baselines results in severer information lost, as illustrated in the second row in Figure 3. As shown in Table 5, our $\\mathrm{CLIP}^2$ considerably outperforms the baseline recognition results by more than $20\\%$ , and our triplet proxies respectively boost two baselines by $9.5\\%$ and $4.8\\%$ . Additionally, we evaluate the localization ability on the outdoor scenario nuScenes in Table 4. Due to the lack of works that tackle outdoor open-vocabulary localization problems, we follow classic detection accuracy metrics Precision(P.) and Recall(R.) as evaluation metrics. Specifically, we calculate the center distance between groundtruth bounding boxes and our 3D proxies that are predicted to belong to the same category as the groundtruth, and set the distance threshold as $\\lambda = 2\\mathrm{m}$ . For those matched pairs that are closer than $\\lambda$ , we count the proxies as TPs. Otherwise, for those unmatched proxies and groundtruth, we count them as FPs and FNs respectively, thus $\\mathrm{P.} = \\frac{\\mathrm{TPs}}{\\mathrm{TPs} + \\mathrm{FPs}}$ , $\\mathrm{R.} = \\frac{\\mathrm{TPs}}{\\mathrm{TPs} + \\mathrm{FNs}}$ . As shown in Table 4, our $\\mathrm{CLIP}^2$ pipeline can provide high", + "bbox": [ + 75, + 618, + 472, + 902 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "recall for outdoor objects. Since $\\mathrm{CLIP}^2$ is highly sensitive to open-world objects and can perceive categories beyond groundtruth list, it tends to create overmuch predictions thus the precision is comparably low. The perception ability of open-world objects can be viewed in Figure 5(b).", + "bbox": [ + 496, + 481, + 892, + 556 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Quality results. We show off two outdoor scenes of nuScenes [2] in Figure 5(b-i) and Figure 5(b-ii). In addition to perceiving those common categories, our $\\mathrm{CLIP}^2$ surprisingly localizes and recognizes those uncommon 3D objects in 3D scenes such as the tires of vehicles, the plastic bag in the hand of pedestrian as well as the plastic bag on the road. We believe it contributes to auto-driving safety by providing the localization and recognition of universal obstacles to facilitate follow-up driving decisions.", + "bbox": [ + 496, + 560, + 893, + 694 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. Few-shot Classification", + "text_level": 1, + "bbox": [ + 498, + 700, + 714, + 715 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Setting. Lightweight few-shot learning is practical for application by finetuning the pretraining model with given limited data annotations, which can also validate the generalization capability of our learned representation. To make a fair comparison, we follow the existing methods [1, 43] to conduct experiments under \"K-way N-shot\" setting on the challenging realistic object-level dataset ScanObjectNN [34], where we randomly sample N point cloud objects from each of the randomly selected K classes.", + "bbox": [ + 496, + 718, + 893, + 851 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Quantity results. As illustrated in Table 6, we compare with representative 3D networks PointNet++ [26], the recent zero-shot approach PointClip [43] as well as a state-", + "bbox": [ + 496, + 856, + 893, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "15250", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/7c879c8d5eb9a4f0bac0c96b6bb09d00402f7e0572a079e0a0a11c9a863c5121.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Rep.Obj.Avg..INAvg..OUTAvg..OBJ
(a)DepthLang.-Depth38.09.331.8
(b)DepthImage-Depth56.421.739.0
(c)PC.Lang.-Point52.626.837.7
(d)PC.Image-Point56.524.437.8
(e)PC.Lang.-Image-Point61.328.839.4
", + "bbox": [ + 83, + 89, + 470, + 181 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "of-the-art representation learning method CrossPoint [1], which conducts contrastive pretraining between point cloud and rendered images on CAD dataset ShapeNet [3]. As we can see, with a slight number of samples, our $\\mathrm{CLIP}^2$ can boost the classification results by a large margin, exceeding PointClip by $5.3\\%$ , $9.6\\%$ and $6.9\\%$ with 4, 8 and 16 shots. Besides, we outperform CrossPoint with considerable gain, illustrating our pretraining strategy on collected proxies can learn sufficient knowledge from realistic open world to generate transferable 3D representation, which is superior to the pretraining on a small-scale synthetic dataset.", + "bbox": [ + 75, + 244, + 470, + 407 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.3. Ablations and Analysis", + "text_level": 1, + "bbox": [ + 76, + 416, + 290, + 431 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Ablations on representation learning. To observe the transferability of different representations and the effect of different learning objectives, we conduct ablations and report the mean average Top1 accuracy across all classes of zero-shot recognition in indoor scenarios [32] (Avg..IN), outdoor scenarios [2] (Avg..OUT) and object-level benchmark [34] (Avg..OBJ), which is shown in Table 7. Firstly, for fair comparisons, we follow [13, 43] to project input point cloud into depth maps in $N_V$ different views as alternative representation. Secondly, we adopt various objectives to learn different correlation alignments across language, image and point cloud feature space or depth space. Specifically, comparing (a) and (b), aligning depth space to image space yields better transfer performance due to the similar data structure of image and depth map. In (c) and (d), point cloud representation is better when aligning to image space in indoor scenes, while better to align with language space in outdoor scenes due to the data discrepancy between image-like RGB-D points and sparse LiDAR points. Generally, 3D point cloud representation outperforms depth representation in all benchmarks due to preserving the complete 3D structure and sufficient 3D-specific information. Comparing (e) and (d), the joint alignment between three feature spaces contributes to the best 3D point cloud representation transferability on all benchmarks.", + "bbox": [ + 75, + 436, + 468, + 804 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Analysis of representation ensembling. Intuitively, different representations contain different perspectives of knowledge, which can be potentially merged to achieve the optimum results during inference. To validate the ensembling application, we adopt three optional representation modals, i.e. point clouds, projected depth maps and cor", + "bbox": [ + 75, + 811, + 468, + 901 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/95c9e5f9865366c014e808607f3680b44da597dee886ae4a757b35313df6f44e.jpg", + "table_caption": [ + "Table 7. Ablations on representation learning. Rep.: the recognition representations. Obj.: the learning objectives. PC.: point cloud. Lang.: language space." + ], + "table_footnote": [], + "table_body": "
PC.ImageDepthAvg..INAvg..OUTAvg..OBJ
(i)✓61.328.839.4
(ii)✓64.241.1-
(iii)✓56.923.939.0
(f)✓✓68.743.9-
(g)✓✓64.830.443.2
(h)✓✓✓69.642.3-
", + "bbox": [ + 506, + 89, + 887, + 193 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 8. Analysis on the representation ensembling schemes.", + "bbox": [ + 511, + 198, + 877, + 212 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "responding image patches, where depth representation is trained on our proxies and image representation is generated from pretrained image branch of CLIP. We ensemble their predicted logits by simple summation as the final output, and illustrate the separate recognition results and ensembling performance in Table 8. Benefiting from the sufficient knowledge learned from massive CLIP training data, image representation presents best performance in separate applications. By merging the complementary knowledge, our 3D representation leads to gains of $4.6\\%$ indoors [32] and $2.8\\%$ outdoors. Though further improving the indoor recognition performance with $0.9\\%$ when merged, depth representation yeilds $1.6\\%$ drop for outdoor objects, illustrating the information lost especially for outdoor scenarios. Since image representation is sometimes missing, such as in [34], our 3D representation is more robust for 3D applications.", + "bbox": [ + 496, + 226, + 890, + 462 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Limitation", + "text_level": 1, + "bbox": [ + 500, + 473, + 612, + 489 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "As a pilot work for the language-3D pretraining problem, though $\\mathrm{CLIP}^2$ enables zero-shot localization and recognition with proposed triplet proxy generation and learned transferable 3D representation, it can not provide the accurate tight bounding box for open-world 3D objects as a common detector does. We believe $\\mathrm{CLIP}^2$ can facilitate the development of open-world 3D detectors by introducing the recognition ability to general 3D detectors or providing presented 3D proxies to enable further training of 3D detectors.", + "bbox": [ + 496, + 498, + 890, + 631 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 643, + 617, + 657 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we present a novel contrastive language-image-point cloud pretraining framework, $\\mathrm{CLIP}^2$ , which consists of a triplet proxy collection scheme and a cross-modal contrastive learning mechanism. Based on the observation that realistic scenarios contain a massive amount of open-world objects, we innovatively propose to collect triplet proxies from realistic scenes as pretraining data. We then conduct cross-modal contrastive alignment across language, image and point cloud feature space to learn transferable 3D representation. The zero-shot transfer results on various indoor and outdoor benchmarks validate the ability of $\\mathrm{CLIP}^2$ for 3D open-world understanding.", + "bbox": [ + 496, + 667, + 890, + 844 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements We acknowledge the support of MindSpore1, CANN, and Ascend AI Processor in this work.", + "bbox": [ + 498, + 848, + 890, + 878 + ], + "page_idx": 7 + }, + { + "type": "page_footnote", + "text": "1https://www.mindspore.cn/", + "bbox": [ + 517, + 886, + 669, + 898 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "15251", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 174, + 106 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Mohamed Afham, Isuru Dissanayake, Dinithi Dissanayake, Amaya Dharmasiri, Kanchana Thilakarathna, and Ranga Rodrigo. Crosspoint: Self-supervised cross-modal contrastive learning for 3d point cloud understanding. In CVPR, 2022. 2, 3, 7, 8", + "[2] Holger Caesar, Varun Bankiti, Alex H Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuscenes: A multimodal dataset for autonomous driving. In CVPR, 2020. 2, 4, 5, 6, 7, 8", + "[3] Angel X Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, et al. Shapenet: An information-rich 3d model repository. arXiv preprint arXiv:1512.03012, 2015. 8", + "[4] Kai Chen, Jiaqi Wang, Jiangmiao Pang, Yuhang Cao, Yu Xiong, Xiaoxiao Li, Shuyang Sun, Wansen Feng, Ziwei Liu, Jiarui Xu, Zheng Zhang, Dazhi Cheng, Chenchen Zhu, Tianheng Cheng, Qijie Zhao, Buyu Li, Xin Lu, Rui Zhu, Yue Wu, Jifeng Dai, Jingdong Wang, Jianping Shi, Wanli Ouyang, Chen Change Loy, and Dahua Lin. MMDetection: Open mmlab detection toolbox and benchmark. arXiv preprint arXiv:1906.07155, 2019. 5, 6", + "[5] Ali Cheraghian, Shafin Rahman, Dylan Campbell, and Lars Petersson. Mitigating the hubness problem for zero-shot learning of 3d objects. In BMVC, 2019. 2, 3", + "[6] Ali Cheraghian, Shafin Rahman, Dylan Campbell, and Lars Petersson. Transductive zero-shot learning for 3d point cloud classification. In WACV, 2020. 2, 3", + "[7] Ali Cheraghian, Shafin Rahman, Townim F Chowdhury, Dylan Campbell, and Lars Petersson. Zero-shot learning on 3d point cloud objects and beyond. IJCV, 2022. 2", + "[8] Ali Cheraghian, Shafin Rahman, and Lars Petersson. Zeroshot learning of 3d point cloud objects. In MVA. IEEE, 2019. 2", + "[9] Angela Dai, Angel X Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In CVPR, 2017. 2, 4, 5", + "[10] Zhipeng Ding, Xu Han, and Marc Niethammer. Votenet: A deep learning label fusion method for multi-atlas segmentation. In MICCAI, 2019. 1, 6", + "[11] Agrim Gupta, Piotr Dollar, and Ross Girshick. Lvis: A dataset for large vocabulary instance segmentation. In CVPR, pages 5356-5364, 2019. 4", + "[12] Huy Ha and Shuran Song. Semantic abstraction: Openworld 3d scene understanding from 2d vision-language models. In CoRL, 2022. 2", + "[13] Tianyu Huang, Bowen Dong, Yunhan Yang, Xiaoshui Huang, Rynson WH Lau, Wanli Ouyang, and Wangmeng Zuo. Clip2point: Transfer clip to point cloud classification with image-depth pre-training. arXiv preprint arXiv:2210.01055, 2022. 2, 3, 4, 5, 6, 7, 8", + "[14] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In ICML. PMLR, 2021. 2, 3" + ], + "bbox": [ + 78, + 114, + 470, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[15] Junnan Li, Ramprasaath Selvaraju, Akhilesh Gotmare, Shafiq Joty, Caiming Xiong, and Steven Chu Hong Hoi. Align before fuse: Vision and language representation learning with momentum distillation. NeurIPS, 2021. 2", + "[16] Hanxue Liang, Chenhan Jiang, Dapeng Feng, Xin Chen, Hang Xu, Xiaodan Liang, Wei Zhang, Zhenguo Li, and Luc Van Gool. Exploring geometry-aware contrast and clustering harmonization for self-supervised 3d object detection. In ICCV, 2021. 3", + "[17] Haotian Liu, Mu Cai, and Yong Jae Lee. Masked discrimination for self-supervised learning on point clouds. In ECCV, 2022. 3", + "[18] Yuheng Lu, Chenfeng Xu, Xiaobao Wei, Xiaodong Xie, Masayoshi Tomizuka, Kurt Keutzer, and Shanghang Zhang. Open-vocabulary 3d detection via image-level class and de-biased cross-modal contrastive learning. arXiv preprint arXiv:2207.01987, 2022. 2, 6", + "[19] Jiageng Mao, Minzhe Niu, Chenhan Jiang, Hanxue Liang, Jingheng Chen, Xiaodan Liang, Yamin Li, Chaoqiang Ye, Wei Zhang, Zhenguo Li, et al. One million scenes for autonomous driving: Once dataset. arXiv preprint arXiv:2106.11037, 2021. 2, 4, 6", + "[20] Jiageng Mao, Shaoshuai Shi, Xiaogang Wang, and Hongsheng Li. 3d object detection for autonomous driving: A review and new outlooks. arXiv preprint arXiv:2206.09474, 2022.1", + "[21] Ishan Misra, Rohit Girdhar, and Armand Joulin. An end-to-end transformer model for 3d object detection. In ICCV, 2021. 6", + "[22] Anshul Paigwar, David Sierra-Gonzalez, Özgür Erkent, and Christian Laugier. Frustum-pointpillars: A multi-stage approach for 3d object detection using rgb camera and lidar. In ICCV, 2021. 4", + "[23] Yatian Pang, Wenxiao Wang, Francis EH Tay, Wei Liu, Yonghong Tian, and Li Yuan. Masked autoencoders for point cloud self-supervised learning. In ECCV, 2022. 3", + "[24] Charles R Qi, Wei Liu, Chenxia Wu, Hao Su, and Leonidas J Guibas. Frustum pointnets for 3d object detection from rgb-d data. In CVPR, 2018. 4", + "[25] Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. Pointnet: Deep learning on point sets for 3d classification and segmentation. In CVPR, 2017. 1, 2, 7", + "[26] Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J Guibas. Pointnet++: Deep hierarchical feature learning on point sets in a metric space. NeurIPS, 2017. 5, 7", + "[27] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, 2021. 2, 3, 4, 5", + "[28] Carsten Rother, Vladimir Kolmogorov, and Andrew Blake. \"grabcut\" interactive foreground extraction using iterated graph cuts. TOG, 2004. 4", + "[29] Erich Schubert, Jörg Sander, Martin Ester, Hans Peter Kriegel, and Xiaowei Xu. Dbscan revisited, revisited: why and how you should (still) use dbscan. *TODS*, 42(3):1-21, 2017. 4", + "[30] Charu Sharma and Manohar Kaul. Self-supervised few-shot learning on point clouds. NeurIPS, 2020. 3" + ], + "bbox": [ + 501, + 92, + 890, + 901 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "15252", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[31] Shaoshuai Shi, Xiaogang Wang, and Hongsheng Li. Pointcnn: 3D object proposal generation and detection from point cloud. In CVPR, 2019. 1", + "[32] Shuran Song, Samuel P Lichtenberg, and Jianxiong Xiao. Sun rgb-d: A rgb-d scene understanding benchmark suite. In CVPR, 2015. 2, 4, 5, 6, 7, 8", + "[33] Pei Sun, Henrik Kretzschmar, Xerxes Dotiwalla, Aurelien Chouard, Vijaysai Patnaik, Paul Tsui, James Guo, Yin Zhou, Yuning Chai, Benjamin Caine, et al. Scalability in perception for autonomous driving: Waymo open dataset. In CVPR, 2020. 5", + "[34] Mikaela Angelina Uy, Quang-Hieu Pham, Binh-Son Hua, Thanh Nguyen, and Sai-Kit Yeung. Revisiting point cloud classification: A new benchmark dataset and classification model on real-world data. In ICCV, 2019. 2, 5, 6, 7, 8", + "[35] Kashi Venkatesh Vishwanath, Diwaker Gupta, Amin Vahdat, and Ken Yocum. Modelnet: Towards a datacenter emulation environment. In 2009 IEEE Ninth International Conference on Peer-to-Peer Computing. IEEE, 2009. 6", + "[36] Hanchen Wang, Qi Liu, Xiangyu Yue, Joan Lasenby, and Matt J Kusner. Unsupervised point cloud pre-training via occlusion completion. In ICCV, 2021. 2", + "[37] Saining Xie, Jiatao Gu, Demi Guo, Charles R Qi, Leonidas Guibas, and Or Litany. Pointcontrast: Unsupervised pretraining for 3d point cloud understanding. In ECCV, 2020. 3", + "[38] Lewei Yao, Jianhua Han, Youpeng Wen, Xiaodan Liang, Dan Xu, Wei Zhang, Zhenguo Li, Chunjing Xu, and Hang Xu. Detclip: Dictionary-enriched visual-concept paralleled pretraining for open-world detection. In NeurIPS, 2022. 2, 4, 5", + "[39] Yangyang Ye, Houjin Chen, Chi Zhang, Xiaoli Hao, and Zhaoxiang Zhang. Sarpnet: Shape attention regional proposal network for lidar-based 3d object detection. Neurocomputing, 2020. 3", + "[40] Tianwei Yin, Xingyi Zhou, and Philipp Kähenbuhl. Center-based 3d object detection and tracking. arXiv preprint arXiv:2006.11275, 2020. 1", + "[41] Xumin Yu, Lulu Tang, Yongming Rao, Tiejun Huang, Jie Zhou, and Jiwen Lu. Point-bert: Pre-training 3d point cloud transformers with masked point modeling. In CVPR, 2022. 3", + "[42] Renrui Zhang, Ziyu Guo, Peng Gao, Rongyao Fang, Bin Zhao, Dong Wang, Yu Qiao, and Hongsheng Li. Pointm2ae: Multi-scale masked autoencoders for hierarchical point cloud pre-training. arXiv preprint arXiv:2205.14401, 2022.3", + "[43] Renrui Zhang, Ziyu Guo, Wei Zhang, Kunchang Li, Xu-peng Miao, Bin Cui, Yu Qiao, Peng Gao, and Hongsheng Li. Pointclip: Point cloud understanding by clip. In CVPR, 2022. 2, 3, 4, 5, 6, 7, 8", + "[44] Zaiwei Zhang, Bo Sun, Haitao Yang, and Qixing Huang. H3dnet: 3d object detection using hybrid geometric primitives. In ECCV, 2020. 1, 6" + ], + "bbox": [ + 78, + 90, + 468, + 832 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "15253", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2023/CLIP2_ Contrastive Language-Image-Point Pretraining From Real-World Point Cloud Data/b3efe108-020c-4b4b-bc30-fbdcfe854136_model.json b/2023/CLIP2_ Contrastive Language-Image-Point Pretraining From Real-World Point Cloud Data/b3efe108-020c-4b4b-bc30-fbdcfe854136_model.json new file mode 100644 index 0000000000000000000000000000000000000000..05135509eb92a00398b6ece49eac62e2a98bf319 --- /dev/null +++ b/2023/CLIP2_ Contrastive Language-Image-Point Pretraining From Real-World Point Cloud Data/b3efe108-020c-4b4b-bc30-fbdcfe854136_model.json @@ -0,0 +1,1914 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.179, + 0.13, + 0.791, + 0.172 + ], + "angle": 0, + "content": "CLIP\\(^2\\): Contrastive Language-Image-Point Pretraining from Real-World Point Cloud Data" + }, + { + "type": "text", + "bbox": [ + 0.178, + 0.201, + 0.791, + 0.24 + ], + "angle": 0, + "content": "Yihan Zeng\\(^{1*}\\), Chenhan Jiang\\(^{2*}\\), Jiageng Mao\\(^{3}\\), Jianhua Han\\(^{1}\\), Chaoqiang Ye\\(^{1}\\), Qingqiu Huang\\(^{1}\\), Dit-Yan Yeung\\(^{2}\\), Zhen Yang\\(^{1}\\), Xiaodan Liang\\(^{4}\\), Hang \\(\\mathrm{Xu}^{1\\dagger}\\)" + }, + { + "type": "image_caption", + "bbox": [ + 0.108, + 0.271, + 0.272, + 0.286 + ], + "angle": 0, + "content": "Open-world Vocabularies" + }, + { + "type": "text", + "bbox": [ + 0.318, + 0.271, + 0.868, + 0.284 + ], + "angle": 0, + "content": "Car, Pedestrian, Bicycle, Bus, Debris, Trailer, Table, Chair, Light, Cleaner, Closet, Backpack, Tire, Couch ..." + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.291, + 0.489, + 0.467 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.493, + 0.288, + 0.888, + 0.467 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.469, + 0.89, + 0.497 + ], + "angle": 0, + "content": "Figure 1. Illustration of our open-world recognition results. Benefiting from our \\(\\mathrm{CLIP}^2\\), the 3D representation is aligned to the openworld language representation, which enables flexible zero-shot transfer. Best viewed in colors." + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.526, + 0.313, + 0.542 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.56, + 0.47, + 0.839 + ], + "angle": 0, + "content": "Contrastive Language-Image Pre-training, benefiting from large-scale unlabeled text-image pairs, has demonstrated great performance in open-world vision understanding tasks. However, due to the limited Text-3D data pairs, adapting the success of 2D Vision-Language Models (VLM) to the 3D space remains an open problem. Existing works that leverage VLM for 3D understanding generally resort to constructing intermediate 2D representations for the 3D data, but at the cost of losing 3D geometry information. To take a step toward open-world 3D vision understanding, we propose Contrastive Language-Image-Point Cloud Pretraining \\((CLIP^2)\\) to directly learn the transferable 3D point cloud representation in realistic scenarios with a novel proxy alignment mechanism. Specifically, we exploit naturally-existed correspondences in 2D and 3D scenarios, and build well-aligned and instance-based text-image-point proxies from those complex scenarios. On top of that, we propose a cross-modal contrastive objective to learn semantic and instance-level aligned point cloud representa" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.514, + 0.892, + 0.616 + ], + "angle": 0, + "content": "tion. Experimental results on both indoor and outdoor scenarios show that our learned 3D representation has great transfer ability in downstream tasks, including zero-shot and few-shot 3D recognition, which boosts the state-of-the-art methods by large margins. Furthermore, we provide analyses of the capability of different representations in real scenarios and present the optional ensemble scheme." + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.653, + 0.631, + 0.669 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.681, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Powerful 3D point cloud representation plays a crucial role in various real-world applications, e.g., 3D object recognition and detection [10, 20, 31, 40, 44]. Compared to 2D images, 3D point cloud provides specific information like accurate geometry that is robust to illumination changes. However, current methods [25, 40] that learn 3D representations generally rely on the predefined number of object categories and require plenty of labor-intensive annotations. Those learned 3D representations are insufficient for safety-critical scenarios like self-driving which includes a long-tail class distribution far beyond the predefined taxonomy. Therefore, it is highly demanded to learn a transferable 3D representation equipped with zero-shot recognition ability in vocabulary scalable real-world scenes. Figure 1 shows an open-world recognition example by our \\(\\mathrm{CLIP}^2\\) in" + }, + { + "type": "page_footnote", + "bbox": [ + 0.076, + 0.852, + 0.468, + 0.9 + ], + "angle": 0, + "content": "*Equal contribution. 1Huawei Noah's Ark Lab 2Hong Kong University of Science and Technology 3The Chinese University of Hong Kong 4Sun Yat-san University †Corresponding Author: xu.hang@huawei.com" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15244" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.135 + ], + "angle": 0, + "content": "outdoor and indoor scenes, where the 3D objects can be classified with the correlation alignment between 3D representations and open-world vocabularies." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.137, + 0.471, + 0.459 + ], + "angle": 0, + "content": "The critical ingredient of open-world understanding is that the models learn sufficient knowledge to obtain general representations. To achieve this, recent Vision-Language Models (VLM) [14, 27, 38] leverage Internet-scale text-image pairs to conduct vision-language pretraining, which facilitates transferable 2D representation and demonstrates promising performance in 2D open-vocabulary tasks. However, 3D vision-language pretraining remains unexplored due to the limitation of existing 3D datasets in diversity and scale compared to the massive data sources in 2D counterparts [14, 15, 27, 38]. Though some recent works [12, 13, 43] try to avoid this problem by transferring the pretrained 2D VLM into the intermediate representation including projected image patches [12, 18] or depth maps [1, 43], those representations suffer from the loss of 3D geometric information and limited viewpoints under realistic scenarios. Especially the camera images are only sometimes available due to the sensor failure in 3D scenes. We believe the 3D representation based on original point cloud data retains most information and is the optimal solution for 3D real world understanding, which requires a rethink of learning the transferable 3D representation under realistic scenarios." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.46, + 0.47, + 0.827 + ], + "angle": 0, + "content": "To this end, we propose a Contrastive Language-Image-Point cloud Pretraining framework, short for \\(\\mathrm{CLIP}^2\\), which directly aligns 3D space with broader raw text and advances the 3D representation learning into an open-world era. Our learning process can be decomposed into two stages: Firstly, we introduce a Triplet Proxy Collection to alleviate the limitation of accessible pretraining data by constructing language-image-point triplets from real-world scenes. Since the large-scale realistic 3D datasets for outdoor driving [2,19] and indoor scenarios [9,32] are collected in open-world, it contains huge amounts of realistic objects that vary in semantics and diversity. Thus we consider them as potential pretraining data sources without extra human supervision. Specifically, we propose \"Proxy\" instances as the bridges between language descriptions, 2D images and 3D point clouds. Enabled by a well-aligned VLM, a scalable caption list and the geometry transformation between 2D and 3D, we automatically create more than 1 million triplets to facilitate pretraining. Secondly, we further propose a Cross-Modal Pretraining scheme to jointly optimize the feature space alignments of three modalities, i.e., point cloud, language and image. It contains both the contrastive learning objective of semantic-level text-3D correlation and instance-level image-3D correlation, which contributes to better transferability of learned 3D representation." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.827, + 0.469, + 0.902 + ], + "angle": 0, + "content": "We study the transferable capability of \\(\\mathrm{CLIP}^2\\) by benchmarking the zero-shot recognition performance on four popular indoor and outdoor real-world datasets, and find a significant improvement over current methods, achieving Top1 accuracy \\(61.3\\%\\) on SunRGBD [32], \\(43.8\\%\\) on ScanNet [9])," + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.283 + ], + "angle": 0, + "content": "28.8% on nuScenes [2] and 56.0% on ONCE [19]. For a fair comparison with existing methods [1, 13, 36, 43], we conduct zero-shot and few-shot classification on single object dataset ScanObjectNN [34] and find consistent dominance, 16.1% relative improvement on zero-shot classification over previous state-of-the-art method [13]. To validate the vocabulary-increasing ability of \\(\\mathrm{CLIP}^2\\), we report the quantity results and visualizations to show the improved discovery of the long-tail categories. Moreover, we make ablations and analysis on different representations, and investigate ensembling alternatives to merge complementary knowledge of all available representations in realistic applications. Our contributions can be summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.293, + 0.891, + 0.338 + ], + "angle": 0, + "content": "- We propose a novel \\(\\mathrm{CLIP}^2\\) framework that aligns 3D space with open-world language representation, facilitating zero-shot transfer in realistic scenarios." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.347, + 0.892, + 0.392 + ], + "angle": 0, + "content": "- We present a Triplet Proxies Collection scheme in real-world scenes, which alleviates the shortage of text-3D data sources and facilitates the pretraining methods." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.4, + 0.892, + 0.46 + ], + "angle": 0, + "content": "- \\(\\mathrm{CLIP}^2\\) jointly optimizes the correlation alignment between point cloud, language and image by proposed cross-modal pretraining mechanism, which enhances the transferability of learned 3D representation." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.469, + 0.892, + 0.528 + ], + "angle": 0, + "content": "- Our \\(\\mathsf{CLIP}^2\\) achieves the state-of-the-art zero-shot transfer performance on 5 datasets (indoor/outdoor scenes and single-object) and shows quality results on vocabulary-increasing discovery in real world." + }, + { + "type": "list", + "bbox": [ + 0.518, + 0.293, + 0.892, + 0.528 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.541, + 0.642, + 0.556 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.56, + 0.892, + 0.795 + ], + "angle": 0, + "content": "Vision-Language Model. Large vision language models (VLM) [14, 15, 27, 38] have demonstrated successful performance in downstream zero-shot tasks with the learned transferable 2D representations. CLIP [27] and ALIGN [14] push the limit by collecting Internet-scale image-text pairs and then learning the correlation alignment between image and language feature space with contrastive pretraining objectives. Those models can be directly transferred to zero-shot 2D recognition and achieve impressive results. Recent DetClip [38] learns to align image patches to test phrases after pretraining under hybrid supervision from detection, grounding and image-text pair data, which extends the ability to localize open-vocabulary 2D proposals in images. In this paper, we attempt to transfer the open-vocabulary ability of pre-trained VLM to the 3D domain, making language applicable to zero-shot point cloud recognition." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.798, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Zero-shot/Open-world Learning in 3D. Recognizing 3D objects with a large vocabulary is necessary for safety-critical autonomous driving and robotic tasks, yet remains under-explored. Cheraghian et al. [5-8] first attempt to associate PointNet [25] feature with category semantic information via a projection function, and separately proposed an unsupervised skewness loss [5] to mitigate the hubness" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15245" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.088, + 0.089, + 0.891, + 0.362 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.371, + 0.897, + 0.442 + ], + "angle": 0, + "content": "Figure 2. Overview of \\(\\mathbf{CLIP}^2\\) framework. The main components contain two parts, the Triplet Proxy Collection and the Cross-Modal Pretraining. The defined Triplet Proxy set \\(\\mathcal{D}_{\\mathrm{proxy}}\\) consists of language captions \\(\\mathbf{X}^T\\), corresponding image instances \\(\\mathbf{X}^I\\) and raw 3D point cloud instances \\(\\mathbf{X}^P\\), which come from the free data source under realistic scenarios without any labeling labor. On top of that, we pretrain a point cloud encoder \\(E^P\\) with the cross-modal contrastive learning objective. Equipped with \\(\\mathbf{CLIP}^2\\), the learned 3D point cloud representation \\(F^P\\) is well aligned to the language representation, which facilitates downstream zero-shot 3D transfer tasks in the real world." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.466, + 0.473, + 0.775 + ], + "angle": 0, + "content": "problem. The transductive case [6] is discussed in which extends [5] using a triplet loss. Notably, the above works conduct experiments on synthetic datasets and need to divide datasets into \"seen\" categories as training data and \"unseen\" categories as testing data. Thus they are not suitable for realistic scenarios due to the domain gap between synthetic and real-world data, as well as the limited vocabulary-increasing ability. Recently, inspired by the success of VLMs [14, 27] in 2D tasks, some works [13, 43] propose to transfer the zero-shot recognition ability of pretrained CLIP [27] into 3D area. PointCLIP [43] directly projects point cloud into multi-view depth maps as image-like data input for pretrained CLIP to make classification predictions. While CLIP2Point [13] trains an image-depth embedding on ShapeNet [39] to better align the depth representation to the pretrained image space of CLIP. However, depth maps lost plenty of geometry information of the original point cloud data structure, resulting in poor performance especially in realistic scenarios. By contrast, we aim to learn transferable 3D representation based on the original point cloud data structure in realistic scenarios." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.784, + 0.473, + 0.903 + ], + "angle": 0, + "content": "3D Representation Learning. Much progress has been made in learning a comprehensive 3D representation in an unsupervised manner. Most works [1, 16, 17, 23, 30, 37, 41, 42] follow the paradigm that conducts pretraining on unlabeled datasets and then finetunes on the limited downstream annotations. Though the improved transferability of 3D representation, they can not be directly transferred to zero-shot tasks with open-world vocabularies. In this work, we con" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.466, + 0.895, + 0.513 + ], + "angle": 0, + "content": "duct language-image-point cloud pretraining, which learns transferable 3D representation aligned to open-vocabulary language space to facilitate the zero-shot transfer." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.522, + 0.593, + 0.538 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.547, + 0.895, + 0.696 + ], + "angle": 0, + "content": "In this section, we introduce \\(\\mathrm{CLIP}^2\\) to learn a transferable 3D point cloud representation with arbitrary category recognition ability under realistic scenarios, illustrated in Figure 2. We will first present the Triplet Proxy Collection in Section 3.1, which utilizes a pretrained VLM and geometric transformation to obtain language-image-point triplets from real-world scenes. Then we will elaborate Cross-Modal Contrastive Pretraining mechanism in Section 3.2, which jointly optimizes the alignment correlations between language, image and point cloud feature space." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.702, + 0.724, + 0.719 + ], + "angle": 0, + "content": "3.1. Triplet Proxy Collection" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.725, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Inspired by the significant performance of 2D VLMs on open-vocabulary tasks, we aim to develop 3D vision-language pretraining to facilitate category-increasing capacity for real-world scenarios. However, the core challenge is the shortage of pretraining data. Compared to the 2D vision-language pretraining framework CLIP [27], which takes more than 400M image-language pairs from the Internet, the largest 3D single-object dataset ShapeNet [39] only contains 50K CAD models with 55 categories. In addition to the insufficiency of data scale, pretraining on such synthetic data fails to transfer well in the real world due to the huge domain gap. Enlightened by the recent emer" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "15246" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.079, + 0.089, + 0.468, + 0.196 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.204, + 0.47, + 0.233 + ], + "angle": 0, + "content": "Figure 3. Illustration of three representation modals of two 3D objects examples under indoor and outdoor scenarios." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.259, + 0.47, + 0.377 + ], + "angle": 0, + "content": "gence of large-scale point cloud datasets collected in indoor [9, 32] and outdoor scenarios [2, 19], we observe that those naturally-collected datasets potentially contain vast amounts of open-world objects that vary in semantics and diversity. Considering the data collection itself is cheap except for laborious annotation, we novelly take leverage of those available datasets without human annotations as a practical yet effective pretraining data source." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.377, + 0.47, + 0.51 + ], + "angle": 0, + "content": "Specifically, given the realistic scene data \\( S = \\{(P_s, I_s)_{s=1}^{|S|}\\} \\), where \\( P_s \\in \\mathbb{R}^{N_P \\times 3} \\) and \\( I_s \\in \\mathbb{R}^{N_I \\times H \\times W \\times 3} \\) are corresponding 3D point clouds and images of scene \\( s \\), we propose a novel concept, Proxy, as the bridge between language, image and 3D point cloud. As illustrated in Figure 2, equipped by those proxy instances, we can automatically collect a massive number of language-image-point cloud pairs \\( D_{\\mathrm{proxy}} \\) in the format of proxies under open-world scenes. We detail the process as follows." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.51, + 0.47, + 0.659 + ], + "angle": 0, + "content": "Language Proxy. We set the language proxies \\(\\mathbf{X}^T\\in \\mathbb{R}^V\\) as a raw text list from the 2D open-world dataset [11], where \\(V = 1206\\) denotes the vocabulary size of language proxies. Image Proxy. Next, we obtain the image proxies \\(\\mathbf{X}^I\\) by an open vocabulary detector DetCLIP [38], denoted as \\(M\\) which is trained with open-world data and performs openset detection. Concretely, given language proxies \\(\\mathbf{X}^T\\) and input scene image \\(I_{s}\\) , we extract corresponding image proposals as image proxies \\(X_{s}^{I}\\) with \\(M\\) by the similarity between input language embeddings and proposal features as" + }, + { + "type": "equation", + "bbox": [ + 0.163, + 0.671, + 0.469, + 0.691 + ], + "angle": 0, + "content": "\\[\n\\left\\{X _ {s} ^ {I} \\right\\} _ {s \\in | \\mathcal {S} |} = \\mathrm {M} \\left(\\left\\{I _ {s} \\right\\} _ {s \\in | \\mathcal {S} |}, \\mathbf {X} ^ {T}\\right). \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.697, + 0.47, + 0.772 + ], + "angle": 0, + "content": "3D Proxy. We exploit the naturally-existed geometry relations between 2D and 3D scenes to obtain 3D proxies \\(\\mathbf{X}^P\\), which consists of point cloud instances corresponding to image proposals in \\(\\mathbf{X}^I\\). We simplify the geometry transformation as \\(\\mathrm{G}(\\cdot)\\) and formulate the relations as:" + }, + { + "type": "equation", + "bbox": [ + 0.221, + 0.781, + 0.469, + 0.8 + ], + "angle": 0, + "content": "\\[\nX _ {i} ^ {P} = \\mathrm {G} \\left(X _ {i} ^ {I}\\right). \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.811, + 0.47, + 0.904 + ], + "angle": 0, + "content": "Detailedly, for indoor scenes equipped with RGB-D sensors, we first remove the background pixels by unsupervised segmentation algorithm [28] for each image proxy \\(X_{s,i}^{I}\\), \\(i \\in |X_s^I|\\). Since depth information is known, we then transform the segmented pixels from uvd coordinate \\(X_{s,i}^{I,uvd} \\in \\mathbb{R}^{n,3}\\) to xyz coordinate \\(X_{s,i}^{P,xyz} \\in \\mathbb{R}^{n,3}\\) as a 3D" + }, + { + "type": "image", + "bbox": [ + 0.508, + 0.088, + 0.885, + 0.195 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.204, + 0.892, + 0.3 + ], + "angle": 0, + "content": "Figure 4. Comparison of different pretraining strategies. (a) CLIP aligns image and language embedding space [27] as \\( L_{TI} \\) based on large-scale text-image pairs. (b) PointClip [27] aligns projected depth map to CLIP language space as \\( L_{TD} \\). (c) Clip2Point aligns depth map to CLIP image space as \\( L_{ID} \\). (d) our \\( \\mathrm{CLIP}^2 \\) aligns original 3D point cloud to both CLIP language space and image space via cross-modal objective \\( L_{CM} \\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.327, + 0.892, + 0.417 + ], + "angle": 0, + "content": "point cloud proxy with the given camera parameters. For outdoor scenes captured by LiDAR sensors, we first create a 3D frustum for each image proxy by extruding the 2D image proposal into 3D space following [22, 24]. Then we conduct DBSCAN algorithm [29] within frustum and select the point cloud cluster as the point proxy \\( X_{s,i}^{P,xyz} \\)." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.418, + 0.892, + 0.513 + ], + "angle": 0, + "content": "Eventually, we construct Triplet Proxy \\(\\mathcal{D}_{\\mathrm{proxy}} = \\{\\mathbf{X}^T,X_s^I,X_s^P\\}_{s = 1}^{|\\mathcal{S}|}\\) by combining corresponding language proxies \\(\\mathbf{X}^T\\) , image proxies \\(\\mathbf{X}^I\\) and 3D proxies \\(\\mathbf{X}^P\\) , where \\(\\mathbf{X}^I = \\{X_s^I\\}_{s = 1}^{|\\mathcal{S}|}\\) and \\(\\mathbf{X}^P = \\{X_s^P\\}_{s = 1}^{|\\mathcal{S}|}\\) . \\(220\\mathrm{K}\\) and \\(1.4\\mathrm{M}\\) proxy triplets are formed for indoor and outdoor scenes, respectively. More details can be found in the appendix." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.524, + 0.822, + 0.54 + ], + "angle": 0, + "content": "3.2. Cross-Modal Contrastive Pretraining" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.548, + 0.892, + 0.695 + ], + "angle": 0, + "content": "With the triplet proxies \\(\\mathcal{D}_{\\mathrm{proxy}}\\), a straightforward pretraining objective is forcing the alignment between the embedding spaces of point cloud \\(X_{i}^{P}\\) and language \\(X_{i}^{T}\\) from scratch. However, it might not promise good transferability of learned representation, since the number of language-image-point pretraining data triplets remains two orders of magnitude smaller than the language-image pairs adopted by CLIP [27] and the vocabulary size is much more limited. Therefore, we design to learn the correlation alignment based on the pretrained embedding space of CLIP." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.695, + 0.892, + 0.857 + ], + "angle": 0, + "content": "The comparison of current pretraining strategies [13,43] is illustrated in Figure 4, which is a series of 3D variants of CLIP. Notably, both existing methods exploit projected depth map as the intermediate representation of point cloud, which are respectively learned to align to language space [43] and image space [13]. Intuitively, as illustrated in Figure 3, depth representation lost plenty of geometry information compared to the original point cloud, especially in outdoor scenarios. Moreover, images are sometimes unavailable for 3D objects. Thus we conduct pretraining on original 3D point cloud data as an optimal representation." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.857, + 0.892, + 0.902 + ], + "angle": 0, + "content": "Toward learning more transferable representation, we introduce a cross-modal contrastive learning objective to jointly optimize the correlation alignment across lan" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15247" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.208 + ], + "angle": 0, + "content": "guage, image and point cloud, including Semantic-Level Language-3D Alignment and Instance-Level Image-3D Alignment. Specifically, the overall architecture of \\(\\mathrm{CLIP}^2\\), shown in Figure 2, contains language encoder \\(E_{\\theta}^{T}\\), point cloud encoder \\(E_{\\theta}^{P}\\) and visual encoder \\(E_{\\theta}^{I}\\), which respectively embed the triplet proxies into text feature \\(f^{T} \\in \\mathcal{R}^{1 \\times C_{T}}\\), point cloud feature \\(f^{P} \\in \\mathcal{R}^{1 \\times C_{P}}\\) and image feature \\(f^{I} \\in \\mathcal{R}^{1 \\times C_{I}}\\), where \\(C\\) is the embedding dimension." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.209, + 0.471, + 0.356 + ], + "angle": 0, + "content": "Semantic-Level Language-3D Alignment. In order to inherit the open-world recognition ability from pretrained CLIP [27], we align the point cloud feature \\( f^{P} \\) with text embedding \\( f^{T} \\) from well-trained CLIP with Language-Point Proxy \\( \\{X_{i}^{T}, X_{i}^{P}\\} \\) input. We replace classname in the prompts, like \"point cloud of a {classname} .\" with raw text in proxy \\( X_{i}^{T} \\) as language sentences. The core idea is to drive the feature centroids of 3D instances and the corresponding text prompt closer. We compute the loss function of between of language proxy and point cloud proxy as:" + }, + { + "type": "equation", + "bbox": [ + 0.077, + 0.36, + 0.473, + 0.414 + ], + "angle": 0, + "content": "\\[\nl (i, T, P) = - \\log \\frac {\\exp \\left(f _ {i} ^ {T} \\cdot f _ {i} ^ {P} / \\tau\\right)}{\\exp \\left(f _ {i} ^ {T} \\cdot f _ {i} ^ {P} / \\tau\\right) + \\sum_ {j \\in N , X _ {j} ^ {T} \\neq X _ {i} ^ {T}} \\exp \\left(f _ {i} ^ {T} \\cdot f _ {j} ^ {P} / \\tau\\right)}, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.415, + 0.47, + 0.46 + ], + "angle": 0, + "content": "where \\(N\\) is the mini-batch size, \\(\\tau\\) is the temperature coefficient. Within a training mini-batch, the language-3D alignment objective \\(L(T,P)\\) can be described as:" + }, + { + "type": "equation", + "bbox": [ + 0.182, + 0.468, + 0.47, + 0.5 + ], + "angle": 0, + "content": "\\[\nL (T, P) = \\frac {1}{N} \\sum_ {i \\in N} l (i, T, P). \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.508, + 0.47, + 0.655 + ], + "angle": 0, + "content": "Instance-Level Image-3D Alignment. In addition to the alignment between semantic language and 3D proxy instances, we further introduce the contrastive alignment between instance-wise image proxy and 3D proxy instances. Note that the instance-aware visual concept has been well-studied in the embedding space of pretrained CLIP. We believe instance-sensitive learning can contribute to further correlation learning and benefits to the transferability of learned 3D representation. The contrastive aligned objective \\( L(I,P) \\) across point cloud and image is formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.08, + 0.675, + 0.468, + 0.729 + ], + "angle": 0, + "content": "\\[\nl (i, I, P) = - \\log \\frac {\\exp \\left(f _ {i} ^ {I} \\cdot f _ {i} ^ {P} / \\tau\\right)}{\\exp \\left(f _ {i} ^ {I} \\cdot f _ {i} ^ {P} / \\tau\\right) + \\sum_ {j \\in N , j \\neq i} \\exp \\left(f _ {i} ^ {J} \\cdot f _ {j} ^ {P} / \\tau\\right)}, \\tag {5}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.178, + 0.738, + 0.468, + 0.774 + ], + "angle": 0, + "content": "\\[\nL (I, P) = \\frac {1}{N} \\sum_ {i \\in N} l (i, I, P). \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.78, + 0.47, + 0.854 + ], + "angle": 0, + "content": "Finally, we obtain the resultant cross-modal contrastive learning objective \\( L_{CM}(T,I,P) \\) as the combination of \\( L(T,P) \\) and \\( L(I,P) \\), where both alignments of semantic-level text-3D correlation and instance-level image-3D correlation are injected:" + }, + { + "type": "equation", + "bbox": [ + 0.141, + 0.867, + 0.468, + 0.882 + ], + "angle": 0, + "content": "\\[\nL _ {C M} (T, I, P) = \\lambda_ {1} L (T, P) + \\lambda_ {1} L (I, P), \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.886, + 0.458, + 0.901 + ], + "angle": 0, + "content": "where the hyper-parameters \\(\\lambda_{1}\\) and \\(\\lambda_{2}\\) are both set to 0.5." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.091, + 0.625, + 0.108 + ], + "angle": 0, + "content": "4. Experiment" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.115, + 0.892, + 0.175 + ], + "angle": 0, + "content": "In this section, we evaluate \\(\\mathrm{CLIP}^2\\) on realistic indoor and outdoor scenarios. We report the zero-shot transfer results on various datasets [2, 9, 32-34] and make further analysis on the designs of pretraining strategy." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.182, + 0.682, + 0.196 + ], + "angle": 0, + "content": "4.1. Zero-shot Transfer" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.2, + 0.892, + 0.317 + ], + "angle": 0, + "content": "Setting. After pretraining, natural language is applied to reference learned 3D representation to enable following zero-shot transfer tasks. (i) Zero-Shot Recognition: we evaluate zero-shot recognition performance for realistic objects, where \\( K \\) category names are transferred to text prompt \"point cloud of {CLASS}\" to encode the text features \\( F_{K} \\in \\mathbb{R}^{K \\times C} \\). Then the classification logits are calculated with the 3D feature \\( f^{P} \\) and text features as:" + }, + { + "type": "equation", + "bbox": [ + 0.592, + 0.324, + 0.892, + 0.343 + ], + "angle": 0, + "content": "\\[\n\\log_ {i} = \\operatorname {s o f t m a x} \\left(f _ {i} ^ {P} \\left(F _ {K}\\right) ^ {T}\\right). \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.351, + 0.892, + 0.556 + ], + "angle": 0, + "content": "We present the results under both indoor and outdoor scenarios in Table 1, Table 2 and Table 5, as well as the object-level benchmark in Table 6. (ii) Open-vocabulary recognition: we enlarge the category vocabularies of ScanNet to 249 and 384 to study the open-vocabulary recognition ability in Table 3. (iii) Open-vocabulary localization: we study the open-vocabulary localization ability by localizing open-world 3D objects with our proxy generation process and then classifying them with our learned 3D representation, of which the visualization is illustrated in Figure 5 and evaluation results are reported in Table 4. Notably, we investigate representation ensembling alternatives to enable knowledge merging of all available representations for realistic applications, illustrated in Table 8." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.563, + 0.673, + 0.577 + ], + "angle": 0, + "content": "4.1.1 Indoor Scenarios" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.581, + 0.892, + 0.904 + ], + "angle": 0, + "content": "Datasets and details. We adopt the widely used indoor 3D dataset SUN RGB-D [32] as the realistic indoor scenario that provides pretraining data source, a single-view RGB-D dataset consisting of \\(\\sim 10\\mathrm{K}\\) scenes. To validate the transferability of learned 3D representation, we also evaluate another popular indoor 3D dataset ScanNet [9], which contains \\(\\sim 1.5\\mathrm{K}\\) scenes of 3D reconstructed meshes. We remove objects in ScanNet with less than 5 points, leaving 384 noisy categories. For open-vocabulary recognition, we evaluate performance on the ScanNet 384-class set and a 249-class merged set. In addition to the scene-wise indoor dataset, we conduct evaluations on ScanObjectNN [34], which collects \\(\\sim 3\\mathrm{K}\\) individual realistic objects with 15 categories and is applied in the previous zero-shot evaluation [13, 43]. During the proxy collection process, we empirically set \\(\\epsilon = 0.3\\) in [38] as a tradeoff between filtering FPs and preserving TPs to generate image proxies. Considering the occurrence frequencies of different indoor categories vary a lot, we adopt the class balance strategy [4] to mitigate the class imbalance. During pretraining process, we adopt [26] as point cloud encoder and set the overall training epoch number to 100." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15248" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.086, + 0.09, + 0.887, + 0.207 + ], + "angle": 0, + "content": "
MethodAvg.BedBookshelfChairDeskSofaTabletoiletBathhubDresserNight Stand
PointClip [43]11.50.094.00.00.00.014.70.00.06.10.0
Clip2Point [13]18.610.920.664.334.413.814.126.20.01.40.0
PointClip [43] w/ TP.38.045.3100.062.548.544.44.855.216.33.30
Clip2Point [13] w/ TP.56.978.087.636.236.664.737.482.177.567.61.2
CLIP²61.384.075.570.747.375.533.886.265.371.82.4
CLIP² w/ En.69.687.394.370.754.179.347.091.785.782.23.6
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.211, + 0.893, + 0.24 + ], + "angle": 0, + "content": "Table 1. Zero-shot recognition results in SUN RGB-D. Avg.: the mean average Top1 accuracy across all categories. TP.: our Triplet Proxy set. En.: our optional ensembling scheme for inference." + }, + { + "type": "table", + "bbox": [ + 0.085, + 0.254, + 0.887, + 0.337 + ], + "angle": 0, + "content": "
MethodAvg.CabBedChairSofaTablDoorWindBksfPicCntrDeskCurtFridgBathShowrToilSink
PointClip [43]6.30.00.00.00.00.70.00.091.80.00.00.015.00.00.00.00.00.0
Clip2Point [13]24.90.020.885.143.326.569.90.020.91.731.727.00.01.646.50.022.425.6
PointClip [43] w/ TP.26.155.70.072.85.05.11.70.077.20.00.051.70.30.00.040.385.349.2
Clip2Point [13] w/ TP.35.23.011.845.127.610.561.52.671.90.333.629.94.711.572.292.486.134.0
CLIP²38.567.232.669.342.318.319.14.062.61.412.752.840.19.159.741.071.045.5
" + }, + { + "type": "table_caption", + "bbox": [ + 0.091, + 0.342, + 0.877, + 0.358 + ], + "angle": 0, + "content": "Table 2. Zero-shot recognition in ScanNet. Avg.: the mean average Top1 accuracy across all categories. TP.: our Triplet Proxy set." + }, + { + "type": "table", + "bbox": [ + 0.084, + 0.372, + 0.225, + 0.436 + ], + "angle": 0, + "content": "
MethodTop5 Acc.
384cls.249cls.
[43]0.30.4
[13]6.47.0
CLIP222.031.7
" + }, + { + "type": "table", + "bbox": [ + 0.231, + 0.372, + 0.455, + 0.436 + ], + "angle": 0, + "content": "
MethodmAP25IN AR25mIoUOUT P.R.
[21]1.3----
[18]13.037.7---
CLIP212.743.052.86.787.4
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.44, + 0.461, + 0.482 + ], + "angle": 0, + "content": "Table 3. Recognition results. Table 4. Zero-shot localization results of vocabulary ex-sults. IN: indoor scenario SUN RGBpansion in ScanNet. D. OUT: outdoor scenario nuScenes." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.505, + 0.473, + 0.903 + ], + "angle": 0, + "content": "Quantity results. For zero-shot recognition task, we take two recent works as our baselines, i.e. PointClip [43] and Clip2Point [13], which study the zero-shot classification task on 3D object-level benchmarks [34, 35] by leveraging pretrained CLIP with projected depth maps. Focusing on the real-world scenarios, we conduct comparison not only on the realistic object-level [34] as illustrated in Table 6 but also on the scene-level datasets shown in Table 1 and Table 2, where the evaluation follows the common classes split in [10, 44] and reports the instance Top1 accuracy of each class. As shown in tables, our \\(\\mathrm{CLIP}^2\\) can outperform baselines on all benchmarks by large margins. Besides, we apply our triplet proxy generation mechanism (TP.) to baseline methods, and achieve considerable improvements on SUN RGB-D and ScanNet by \\(26.5\\%\\) and \\(19.8\\%\\) for Point-Clip, \\(38.3\\%\\) and \\(10.3\\%\\) for Clip2Point. On the one hand, the contrasts demonstrate the effectiveness of our triplet proxies for open-world understanding. On the other hand, our learned 3D representation is superior in 3D object recognition by retaining more 3D-specific information than depth representation. Besides, we present the optional ensembling scheme (En.) when camera images are available, which can take advantage of multi-modal knowledge and further boost the performance by \\(8.3\\%\\). To further validate the open-vocabulary recognition ability, we conduct evaluation on a larger category set of ScanNet in Table 3 and report the instance Top5 accuracy, which illustrates the superior" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.375, + 0.892, + 0.61 + ], + "angle": 0, + "content": "ity of our \\(\\mathrm{CLIP}^2\\) when vocabulary increases. Beyond that, \\(\\mathrm{CLIP}^2\\) is also equipped with zero-shot 3D localization ability by proxy generation. On indoor scenario SUN RGB-D, we compare with a SOTA indoor 3D detector 3DETR [21] and a recent work OV3D [18] that studies open-vocabulary detection, where evaluation is conducted on the same \"unseen\" split in [18]. Since \\(\\mathrm{CLIP}^2\\) do not fit the tight bounding boxes of point cloud instances, we estimate the maximum bounding box of proxies and GT instances to conduct evaluation following the same metrics \\(\\mathrm{mAP}_{25}\\) and \\(\\mathrm{AR}_{25}\\) in [18], as shown in Table 4. Notably, compared to baseline works that train on \"seen\" 3D annotations and test on \"unseen\" categories, we have no access to any 3D annotations yet achieve comparable localization ability, which yields \\(5.3\\%\\) \\(\\mathrm{AR}_{25}\\) improvement over OV3D [18]. We further evaluate segmentation results in Table 4." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.611, + 0.893, + 0.744 + ], + "angle": 0, + "content": "Quality results. The visualization results of \\(\\mathrm{CLIP}^2\\) under a indoor scene of SUN RGB-D [32] is shown in Figure 5(a). Our triplet proxy generation process can localize open-world 3D objects in a point cloud scene. Moreover, the 3D representation learned from our cross-modal pretraining can provide more accurate classification results for 3D instances by exploiting original point cloud, which corrects the mistaken \"People\" prediction in image to \"Picture\" by considering the geometry information." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.75, + 0.688, + 0.763 + ], + "angle": 0, + "content": "4.1.2 Outdoor Scenarios" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.767, + 0.892, + 0.87 + ], + "angle": 0, + "content": "Datasets and details. We exploit a prevalent large-scale 3D dataset nuScenes [2] as the outdoor data source and extra validate the performance on the ONCE dataset [19]. The nuScenes dataset consists of \\(\\sim 28\\mathrm{K}\\) frames with 10 categories, while ONCE contains 6 annotated sequences with 5 categories. Similarly, we set the \\(\\epsilon = 0.3\\) for image proxies collection and adopt the class balance strategy [4]." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.871, + 0.892, + 0.903 + ], + "angle": 0, + "content": "Quantity results. Since the outdoor point cloud is collected by LiDAR sensors, it has a wider perception range" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "15249" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.082, + 0.09, + 0.891, + 0.188 + ], + "angle": 0, + "content": "
MethodAvg.nuScenesONCE
CarTruckBusPed.BicycleTrailerC.V.Motor.BarrierT.C.CarCyc.Ped.TruckBus
PointClip [43]11.70.00.00.029.141.83.40.00.142.50.00.013.879.20.27.61
Clip2Point [13]12.40.40.30.113.531.01.89.31.566.20.317.311.895.435.74.0
PointClip [43] w/ TP.28.318.805.574.017.957.01.94.52.129.751.89.299.85.046.8
Clip2Point [13] w/ TP.33.026.716.851.245.215.813.920.05.710.534.239.427.895.540.651.7
CLIP²37.841.941.322.540.321.120.624.822.417.335.352.727.377.778.544.0
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.194, + 0.893, + 0.222 + ], + "angle": 0, + "content": "Table 5. Zero-shot recognition results in outdoor scenario: nuScenes (Left) and ONCE (Right). TP.: our Triplet Proxy set. Avg.: the mean average Top1 accuracy across all categories of two benchmarks." + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.236, + 0.437, + 0.412 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.441, + 0.236, + 0.892, + 0.412 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.415, + 0.893, + 0.457 + ], + "angle": 0, + "content": "Figure 5. Visualizations of the zero-shot localization and recognition results by \\(\\mathrm{CLIP}^2\\) under open-world (a) indoor realistic scene [32] and (b) outdoor scenes [2]. Notably, the whole pipeline of \\(\\mathrm{CLIP}^2\\) not only has no access to human annotations, but also enables the open-world vocabularies beyond groundtruth annotations, such as 'Picture' in (a) and 'Plastic bag', 'Tire' in (b). Best viewed in colors." + }, + { + "type": "table", + "bbox": [ + 0.085, + 0.48, + 0.466, + 0.578 + ], + "angle": 0, + "content": "
MethodZS15-way10-way
4-S8-S16-S10-S20-S
PointNet++ [25]-41.047.655.0--
PointCLIP [43]15.446.050.055.6--
Clip2Point [13]23.3-----
CrossPoint [1]----58.7±1.864.6±1.2
CLIP239.151.359.662.560.6±2.566.3±3.2
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.582, + 0.47, + 0.611 + ], + "angle": 0, + "content": "Table 6. Zero-shot and Few-shot classification results on ScanObjectNN. ZS: zero-shot. K-way N-shot: few-shot settings." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.619, + 0.473, + 0.903 + ], + "angle": 0, + "content": "than RGB-D but leads to sparse distribution. Thus the projected depth representation of baselines results in severer information lost, as illustrated in the second row in Figure 3. As shown in Table 5, our \\(\\mathrm{CLIP}^2\\) considerably outperforms the baseline recognition results by more than \\(20\\%\\), and our triplet proxies respectively boost two baselines by \\(9.5\\%\\) and \\(4.8\\%\\). Additionally, we evaluate the localization ability on the outdoor scenario nuScenes in Table 4. Due to the lack of works that tackle outdoor open-vocabulary localization problems, we follow classic detection accuracy metrics Precision(P.) and Recall(R.) as evaluation metrics. Specifically, we calculate the center distance between groundtruth bounding boxes and our 3D proxies that are predicted to belong to the same category as the groundtruth, and set the distance threshold as \\(\\lambda = 2\\mathrm{m}\\). For those matched pairs that are closer than \\(\\lambda\\), we count the proxies as TPs. Otherwise, for those unmatched proxies and groundtruth, we count them as FPs and FNs respectively, thus \\(\\mathrm{P.} = \\frac{\\mathrm{TPs}}{\\mathrm{TPs} + \\mathrm{FPs}}\\), \\(\\mathrm{R.} = \\frac{\\mathrm{TPs}}{\\mathrm{TPs} + \\mathrm{FNs}}\\). As shown in Table 4, our \\(\\mathrm{CLIP}^2\\) pipeline can provide high" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.482, + 0.893, + 0.557 + ], + "angle": 0, + "content": "recall for outdoor objects. Since \\(\\mathrm{CLIP}^2\\) is highly sensitive to open-world objects and can perceive categories beyond groundtruth list, it tends to create overmuch predictions thus the precision is comparably low. The perception ability of open-world objects can be viewed in Figure 5(b)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.561, + 0.895, + 0.695 + ], + "angle": 0, + "content": "Quality results. We show off two outdoor scenes of nuScenes [2] in Figure 5(b-i) and Figure 5(b-ii). In addition to perceiving those common categories, our \\(\\mathrm{CLIP}^2\\) surprisingly localizes and recognizes those uncommon 3D objects in 3D scenes such as the tires of vehicles, the plastic bag in the hand of pedestrian as well as the plastic bag on the road. We believe it contributes to auto-driving safety by providing the localization and recognition of universal obstacles to facilitate follow-up driving decisions." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.702, + 0.715, + 0.716 + ], + "angle": 0, + "content": "4.2. Few-shot Classification" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.719, + 0.894, + 0.852 + ], + "angle": 0, + "content": "Setting. Lightweight few-shot learning is practical for application by finetuning the pretraining model with given limited data annotations, which can also validate the generalization capability of our learned representation. To make a fair comparison, we follow the existing methods [1, 43] to conduct experiments under \"K-way N-shot\" setting on the challenging realistic object-level dataset ScanObjectNN [34], where we randomly sample N point cloud objects from each of the randomly selected K classes." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.857, + 0.894, + 0.903 + ], + "angle": 0, + "content": "Quantity results. As illustrated in Table 6, we compare with representative 3D networks PointNet++ [26], the recent zero-shot approach PointClip [43] as well as a state-" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "15250" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.084, + 0.09, + 0.472, + 0.183 + ], + "angle": 0, + "content": "
Rep.Obj.Avg..INAvg..OUTAvg..OBJ
(a)DepthLang.-Depth38.09.331.8
(b)DepthImage-Depth56.421.739.0
(c)PC.Lang.-Point52.626.837.7
(d)PC.Image-Point56.524.437.8
(e)PC.Lang.-Image-Point61.328.839.4
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.187, + 0.47, + 0.229 + ], + "angle": 0, + "content": "Table 7. Ablations on representation learning. Rep.: the recognition representations. Obj.: the learning objectives. PC.: point cloud. Lang.: language space." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.245, + 0.471, + 0.408 + ], + "angle": 0, + "content": "of-the-art representation learning method CrossPoint [1], which conducts contrastive pretraining between point cloud and rendered images on CAD dataset ShapeNet [3]. As we can see, with a slight number of samples, our \\(\\mathrm{CLIP}^2\\) can boost the classification results by a large margin, exceeding PointClip by \\(5.3\\%\\), \\(9.6\\%\\) and \\(6.9\\%\\) with 4, 8 and 16 shots. Besides, we outperform CrossPoint with considerable gain, illustrating our pretraining strategy on collected proxies can learn sufficient knowledge from realistic open world to generate transferable 3D representation, which is superior to the pretraining on a small-scale synthetic dataset." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.417, + 0.292, + 0.433 + ], + "angle": 0, + "content": "4.3. Ablations and Analysis" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.438, + 0.47, + 0.805 + ], + "angle": 0, + "content": "Ablations on representation learning. To observe the transferability of different representations and the effect of different learning objectives, we conduct ablations and report the mean average Top1 accuracy across all classes of zero-shot recognition in indoor scenarios [32] (Avg..IN), outdoor scenarios [2] (Avg..OUT) and object-level benchmark [34] (Avg..OBJ), which is shown in Table 7. Firstly, for fair comparisons, we follow [13, 43] to project input point cloud into depth maps in \\( N_V \\) different views as alternative representation. Secondly, we adopt various objectives to learn different correlation alignments across language, image and point cloud feature space or depth space. Specifically, comparing (a) and (b), aligning depth space to image space yields better transfer performance due to the similar data structure of image and depth map. In (c) and (d), point cloud representation is better when aligning to image space in indoor scenes, while better to align with language space in outdoor scenes due to the data discrepancy between image-like RGB-D points and sparse LiDAR points. Generally, 3D point cloud representation outperforms depth representation in all benchmarks due to preserving the complete 3D structure and sufficient 3D-specific information. Comparing (e) and (d), the joint alignment between three feature spaces contributes to the best 3D point cloud representation transferability on all benchmarks." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.813, + 0.47, + 0.902 + ], + "angle": 0, + "content": "Analysis of representation ensembling. Intuitively, different representations contain different perspectives of knowledge, which can be potentially merged to achieve the optimum results during inference. To validate the ensembling application, we adopt three optional representation modals, i.e. point clouds, projected depth maps and cor" + }, + { + "type": "table", + "bbox": [ + 0.508, + 0.09, + 0.888, + 0.194 + ], + "angle": 0, + "content": "
PC.ImageDepthAvg..INAvg..OUTAvg..OBJ
(i)✓61.328.839.4
(ii)✓64.241.1-
(iii)✓56.923.939.0
(f)✓✓68.743.9-
(g)✓✓64.830.443.2
(h)✓✓✓69.642.3-
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.199, + 0.878, + 0.213 + ], + "angle": 0, + "content": "Table 8. Analysis on the representation ensembling schemes." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.227, + 0.892, + 0.463 + ], + "angle": 0, + "content": "responding image patches, where depth representation is trained on our proxies and image representation is generated from pretrained image branch of CLIP. We ensemble their predicted logits by simple summation as the final output, and illustrate the separate recognition results and ensembling performance in Table 8. Benefiting from the sufficient knowledge learned from massive CLIP training data, image representation presents best performance in separate applications. By merging the complementary knowledge, our 3D representation leads to gains of \\(4.6\\%\\) indoors [32] and \\(2.8\\%\\) outdoors. Though further improving the indoor recognition performance with \\(0.9\\%\\) when merged, depth representation yeilds \\(1.6\\%\\) drop for outdoor objects, illustrating the information lost especially for outdoor scenarios. Since image representation is sometimes missing, such as in [34], our 3D representation is more robust for 3D applications." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.474, + 0.614, + 0.49 + ], + "angle": 0, + "content": "5. Limitation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.499, + 0.892, + 0.632 + ], + "angle": 0, + "content": "As a pilot work for the language-3D pretraining problem, though \\(\\mathrm{CLIP}^2\\) enables zero-shot localization and recognition with proposed triplet proxy generation and learned transferable 3D representation, it can not provide the accurate tight bounding box for open-world 3D objects as a common detector does. We believe \\(\\mathrm{CLIP}^2\\) can facilitate the development of open-world 3D detectors by introducing the recognition ability to general 3D detectors or providing presented 3D proxies to enable further training of 3D detectors." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.644, + 0.619, + 0.659 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.669, + 0.892, + 0.845 + ], + "angle": 0, + "content": "In this paper, we present a novel contrastive language-image-point cloud pretraining framework, \\(\\mathrm{CLIP}^2\\), which consists of a triplet proxy collection scheme and a cross-modal contrastive learning mechanism. Based on the observation that realistic scenarios contain a massive amount of open-world objects, we innovatively propose to collect triplet proxies from realistic scenes as pretraining data. We then conduct cross-modal contrastive alignment across language, image and point cloud feature space to learn transferable 3D representation. The zero-shot transfer results on various indoor and outdoor benchmarks validate the ability of \\(\\mathrm{CLIP}^2\\) for 3D open-world understanding." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.849, + 0.891, + 0.879 + ], + "angle": 0, + "content": "Acknowledgements We acknowledge the support of MindSpore1, CANN, and Ascend AI Processor in this work." + }, + { + "type": "page_footnote", + "bbox": [ + 0.518, + 0.887, + 0.671, + 0.9 + ], + "angle": 0, + "content": "1https://www.mindspore.cn/" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "15251" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.079, + 0.09, + 0.176, + 0.107 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.182 + ], + "angle": 0, + "content": "[1] Mohamed Afham, Isuru Dissanayake, Dinithi Dissanayake, Amaya Dharmasiri, Kanchana Thilakarathna, and Ranga Rodrigo. Crosspoint: Self-supervised cross-modal contrastive learning for 3d point cloud understanding. In CVPR, 2022. 2, 3, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.183, + 0.472, + 0.249 + ], + "angle": 0, + "content": "[2] Holger Caesar, Varun Bankiti, Alex H Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuscenes: A multimodal dataset for autonomous driving. In CVPR, 2020. 2, 4, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.25, + 0.472, + 0.317 + ], + "angle": 0, + "content": "[3] Angel X Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, et al. Shapenet: An information-rich 3d model repository. arXiv preprint arXiv:1512.03012, 2015. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.318, + 0.472, + 0.425 + ], + "angle": 0, + "content": "[4] Kai Chen, Jiaqi Wang, Jiangmiao Pang, Yuhang Cao, Yu Xiong, Xiaoxiao Li, Shuyang Sun, Wansen Feng, Ziwei Liu, Jiarui Xu, Zheng Zhang, Dazhi Cheng, Chenchen Zhu, Tianheng Cheng, Qijie Zhao, Buyu Li, Xin Lu, Rui Zhu, Yue Wu, Jifeng Dai, Jingdong Wang, Jianping Shi, Wanli Ouyang, Chen Change Loy, and Dahua Lin. MMDetection: Open mmlab detection toolbox and benchmark. arXiv preprint arXiv:1906.07155, 2019. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.426, + 0.472, + 0.467 + ], + "angle": 0, + "content": "[5] Ali Cheraghian, Shafin Rahman, Dylan Campbell, and Lars Petersson. Mitigating the hubness problem for zero-shot learning of 3d objects. In BMVC, 2019. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.467, + 0.472, + 0.507 + ], + "angle": 0, + "content": "[6] Ali Cheraghian, Shafin Rahman, Dylan Campbell, and Lars Petersson. Transductive zero-shot learning for 3d point cloud classification. In WACV, 2020. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.508, + 0.472, + 0.548 + ], + "angle": 0, + "content": "[7] Ali Cheraghian, Shafin Rahman, Townim F Chowdhury, Dylan Campbell, and Lars Petersson. Zero-shot learning on 3d point cloud objects and beyond. IJCV, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.549, + 0.472, + 0.587 + ], + "angle": 0, + "content": "[8] Ali Cheraghian, Shafin Rahman, and Lars Petersson. Zeroshot learning of 3d point cloud objects. In MVA. IEEE, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.589, + 0.472, + 0.642 + ], + "angle": 0, + "content": "[9] Angela Dai, Angel X Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In CVPR, 2017. 2, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.643, + 0.472, + 0.683 + ], + "angle": 0, + "content": "[10] Zhipeng Ding, Xu Han, and Marc Niethammer. Votenet: A deep learning label fusion method for multi-atlas segmentation. In MICCAI, 2019. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.684, + 0.472, + 0.724 + ], + "angle": 0, + "content": "[11] Agrim Gupta, Piotr Dollar, and Ross Girshick. Lvis: A dataset for large vocabulary instance segmentation. In CVPR, pages 5356-5364, 2019. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.725, + 0.472, + 0.764 + ], + "angle": 0, + "content": "[12] Huy Ha and Shuran Song. Semantic abstraction: Openworld 3d scene understanding from 2d vision-language models. In CoRL, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.766, + 0.472, + 0.833 + ], + "angle": 0, + "content": "[13] Tianyu Huang, Bowen Dong, Yunhan Yang, Xiaoshui Huang, Rynson WH Lau, Wanli Ouyang, and Wangmeng Zuo. Clip2point: Transfer clip to point cloud classification with image-depth pre-training. arXiv preprint arXiv:2210.01055, 2022. 2, 3, 4, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.834, + 0.472, + 0.9 + ], + "angle": 0, + "content": "[14] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In ICML. PMLR, 2021. 2, 3" + }, + { + "type": "list", + "bbox": [ + 0.079, + 0.115, + 0.472, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.147 + ], + "angle": 0, + "content": "[15] Junnan Li, Ramprasaath Selvaraju, Akhilesh Gotmare, Shafiq Joty, Caiming Xiong, and Steven Chu Hong Hoi. Align before fuse: Vision and language representation learning with momentum distillation. NeurIPS, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.148, + 0.892, + 0.214 + ], + "angle": 0, + "content": "[16] Hanxue Liang, Chenhan Jiang, Dapeng Feng, Xin Chen, Hang Xu, Xiaodan Liang, Wei Zhang, Zhenguo Li, and Luc Van Gool. Exploring geometry-aware contrast and clustering harmonization for self-supervised 3d object detection. In ICCV, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.216, + 0.892, + 0.255 + ], + "angle": 0, + "content": "[17] Haotian Liu, Mu Cai, and Yong Jae Lee. Masked discrimination for self-supervised learning on point clouds. In ECCV, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.257, + 0.892, + 0.323 + ], + "angle": 0, + "content": "[18] Yuheng Lu, Chenfeng Xu, Xiaobao Wei, Xiaodong Xie, Masayoshi Tomizuka, Kurt Keutzer, and Shanghang Zhang. Open-vocabulary 3d detection via image-level class and de-biased cross-modal contrastive learning. arXiv preprint arXiv:2207.01987, 2022. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.325, + 0.892, + 0.392 + ], + "angle": 0, + "content": "[19] Jiageng Mao, Minzhe Niu, Chenhan Jiang, Hanxue Liang, Jingheng Chen, Xiaodan Liang, Yamin Li, Chaoqiang Ye, Wei Zhang, Zhenguo Li, et al. One million scenes for autonomous driving: Once dataset. arXiv preprint arXiv:2106.11037, 2021. 2, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.394, + 0.892, + 0.446 + ], + "angle": 0, + "content": "[20] Jiageng Mao, Shaoshuai Shi, Xiaogang Wang, and Hongsheng Li. 3d object detection for autonomous driving: A review and new outlooks. arXiv preprint arXiv:2206.09474, 2022.1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.448, + 0.892, + 0.488 + ], + "angle": 0, + "content": "[21] Ishan Misra, Rohit Girdhar, and Armand Joulin. An end-to-end transformer model for 3d object detection. In ICCV, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.489, + 0.892, + 0.543 + ], + "angle": 0, + "content": "[22] Anshul Paigwar, David Sierra-Gonzalez, Özgür Erkent, and Christian Laugier. Frustum-pointpillars: A multi-stage approach for 3d object detection using rgb camera and lidar. In ICCV, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.545, + 0.892, + 0.585 + ], + "angle": 0, + "content": "[23] Yatian Pang, Wenxiao Wang, Francis EH Tay, Wei Liu, Yonghong Tian, and Li Yuan. Masked autoencoders for point cloud self-supervised learning. In ECCV, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.586, + 0.892, + 0.625 + ], + "angle": 0, + "content": "[24] Charles R Qi, Wei Liu, Chenxia Wu, Hao Su, and Leonidas J Guibas. Frustum pointnets for 3d object detection from rgb-d data. In CVPR, 2018. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.627, + 0.892, + 0.667 + ], + "angle": 0, + "content": "[25] Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. Pointnet: Deep learning on point sets for 3d classification and segmentation. In CVPR, 2017. 1, 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.668, + 0.892, + 0.708 + ], + "angle": 0, + "content": "[26] Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J Guibas. Pointnet++: Deep hierarchical feature learning on point sets in a metric space. NeurIPS, 2017. 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.709, + 0.892, + 0.776 + ], + "angle": 0, + "content": "[27] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, 2021. 2, 3, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.778, + 0.892, + 0.817 + ], + "angle": 0, + "content": "[28] Carsten Rother, Vladimir Kolmogorov, and Andrew Blake. \"grabcut\" interactive foreground extraction using iterated graph cuts. TOG, 2004. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.818, + 0.892, + 0.871 + ], + "angle": 0, + "content": "[29] Erich Schubert, Jörg Sander, Martin Ester, Hans Peter Kriegel, and Xiaowei Xu. Dbscan revisited, revisited: why and how you should (still) use dbscan. *TODS*, 42(3):1-21, 2017. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.873, + 0.892, + 0.902 + ], + "angle": 0, + "content": "[30] Charu Sharma and Manohar Kaul. Self-supervised few-shot learning on point clouds. NeurIPS, 2020. 3" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "15252" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.469, + 0.133 + ], + "angle": 0, + "content": "[31] Shaoshuai Shi, Xiaogang Wang, and Hongsheng Li. Pointcnn: 3D object proposal generation and detection from point cloud. In CVPR, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.134, + 0.468, + 0.174 + ], + "angle": 0, + "content": "[32] Shuran Song, Samuel P Lichtenberg, and Jianxiong Xiao. Sun rgb-d: A rgb-d scene understanding benchmark suite. In CVPR, 2015. 2, 4, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.176, + 0.468, + 0.242 + ], + "angle": 0, + "content": "[33] Pei Sun, Henrik Kretzschmar, Xerxes Dotiwalla, Aurelien Chouard, Vijaysai Patnaik, Paul Tsui, James Guo, Yin Zhou, Yuning Chai, Benjamin Caine, et al. Scalability in perception for autonomous driving: Waymo open dataset. In CVPR, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.245, + 0.468, + 0.297 + ], + "angle": 0, + "content": "[34] Mikaela Angelina Uy, Quang-Hieu Pham, Binh-Son Hua, Thanh Nguyen, and Sai-Kit Yeung. Revisiting point cloud classification: A new benchmark dataset and classification model on real-world data. In ICCV, 2019. 2, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.3, + 0.468, + 0.353 + ], + "angle": 0, + "content": "[35] Kashi Venkatesh Vishwanath, Diwaker Gupta, Amin Vahdat, and Ken Yocum. Modelnet: Towards a datacenter emulation environment. In 2009 IEEE Ninth International Conference on Peer-to-Peer Computing. IEEE, 2009. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.355, + 0.468, + 0.394 + ], + "angle": 0, + "content": "[36] Hanchen Wang, Qi Liu, Xiangyu Yue, Joan Lasenby, and Matt J Kusner. Unsupervised point cloud pre-training via occlusion completion. In ICCV, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.396, + 0.468, + 0.448 + ], + "angle": 0, + "content": "[37] Saining Xie, Jiatao Gu, Demi Guo, Charles R Qi, Leonidas Guibas, and Or Litany. Pointcontrast: Unsupervised pretraining for 3d point cloud understanding. In ECCV, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.451, + 0.468, + 0.516 + ], + "angle": 0, + "content": "[38] Lewei Yao, Jianhua Han, Youpeng Wen, Xiaodan Liang, Dan Xu, Wei Zhang, Zhenguo Li, Chunjing Xu, and Hang Xu. Detclip: Dictionary-enriched visual-concept paralleled pretraining for open-world detection. In NeurIPS, 2022. 2, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.519, + 0.468, + 0.572 + ], + "angle": 0, + "content": "[39] Yangyang Ye, Houjin Chen, Chi Zhang, Xiaoli Hao, and Zhaoxiang Zhang. Sarpnet: Shape attention regional proposal network for lidar-based 3d object detection. Neurocomputing, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.574, + 0.468, + 0.613 + ], + "angle": 0, + "content": "[40] Tianwei Yin, Xingyi Zhou, and Philipp Kähenbuhl. Center-based 3d object detection and tracking. arXiv preprint arXiv:2006.11275, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.616, + 0.468, + 0.667 + ], + "angle": 0, + "content": "[41] Xumin Yu, Lulu Tang, Yongming Rao, Tiejun Huang, Jie Zhou, and Jiwen Lu. Point-bert: Pre-training 3d point cloud transformers with masked point modeling. In CVPR, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.67, + 0.468, + 0.735 + ], + "angle": 0, + "content": "[42] Renrui Zhang, Ziyu Guo, Peng Gao, Rongyao Fang, Bin Zhao, Dong Wang, Yu Qiao, and Hongsheng Li. Pointm2ae: Multi-scale masked autoencoders for hierarchical point cloud pre-training. arXiv preprint arXiv:2205.14401, 2022.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.739, + 0.468, + 0.791 + ], + "angle": 0, + "content": "[43] Renrui Zhang, Ziyu Guo, Wei Zhang, Kunchang Li, Xu-peng Miao, Bin Cui, Yu Qiao, Peng Gao, and Hongsheng Li. Pointclip: Point cloud understanding by clip. In CVPR, 2022. 2, 3, 4, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.794, + 0.468, + 0.833 + ], + "angle": 0, + "content": "[44] Zaiwei Zhang, Bo Sun, Haitao Yang, and Qixing Huang. H3dnet: 3d object detection using hybrid geometric primitives. In ECCV, 2020. 1, 6" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.469, + 0.833 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.956 + ], + "angle": 0, + "content": "15253" + } + ] +] \ No newline at end of file diff --git a/2023/CLIP2_ Contrastive Language-Image-Point Pretraining From Real-World Point Cloud Data/b3efe108-020c-4b4b-bc30-fbdcfe854136_origin.pdf b/2023/CLIP2_ Contrastive Language-Image-Point Pretraining From Real-World Point Cloud Data/b3efe108-020c-4b4b-bc30-fbdcfe854136_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..d9cbe9c0f3e1f7af7d64fdd8c7ae623f8210212d --- /dev/null +++ b/2023/CLIP2_ Contrastive Language-Image-Point Pretraining From Real-World Point Cloud Data/b3efe108-020c-4b4b-bc30-fbdcfe854136_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ee4546c86d65d963c4e715058db80d1cb48920f255cb695e03c05ef077da2f24 +size 4810052 diff --git a/2023/CLIP2_ Contrastive Language-Image-Point Pretraining From Real-World Point Cloud Data/full.md b/2023/CLIP2_ Contrastive Language-Image-Point Pretraining From Real-World Point Cloud Data/full.md new file mode 100644 index 0000000000000000000000000000000000000000..791a9a116f6d732785cfb0acdf60ee83a8c04a59 --- /dev/null +++ b/2023/CLIP2_ Contrastive Language-Image-Point Pretraining From Real-World Point Cloud Data/full.md @@ -0,0 +1,273 @@ +# CLIP $^2$ : Contrastive Language-Image-Point Pretraining from Real-World Point Cloud Data + +Yihan Zeng $^{1*}$ , Chenhan Jiang $^{2*}$ , Jiageng Mao $^{3}$ , Jianhua Han $^{1}$ , Chaoqiang Ye $^{1}$ , Qingqiu Huang $^{1}$ , Dit-Yan Yeung $^{2}$ , Zhen Yang $^{1}$ , Xiaodan Liang $^{4}$ , Hang $\mathrm{Xu}^{1\dagger}$ + +Car, Pedestrian, Bicycle, Bus, Debris, Trailer, Table, Chair, Light, Cleaner, Closet, Backpack, Tire, Couch ... + +![](images/cfe8c4c6d10f3a5dd955c4f4c255cbfd42736320d93dda788fc81c0f5a6ecab8.jpg) +Open-world Vocabularies +Figure 1. Illustration of our open-world recognition results. Benefiting from our $\mathrm{CLIP}^2$ , the 3D representation is aligned to the openworld language representation, which enables flexible zero-shot transfer. Best viewed in colors. + +![](images/e34e30f102ed2def75dd13d5cca9f46c554c1ec5db0ff05b03c664eb1a47f446.jpg) + +# Abstract + +Contrastive Language-Image Pre-training, benefiting from large-scale unlabeled text-image pairs, has demonstrated great performance in open-world vision understanding tasks. However, due to the limited Text-3D data pairs, adapting the success of 2D Vision-Language Models (VLM) to the 3D space remains an open problem. Existing works that leverage VLM for 3D understanding generally resort to constructing intermediate 2D representations for the 3D data, but at the cost of losing 3D geometry information. To take a step toward open-world 3D vision understanding, we propose Contrastive Language-Image-Point Cloud Pretraining $(CLIP^2)$ to directly learn the transferable 3D point cloud representation in realistic scenarios with a novel proxy alignment mechanism. Specifically, we exploit naturally-existed correspondences in 2D and 3D scenarios, and build well-aligned and instance-based text-image-point proxies from those complex scenarios. On top of that, we propose a cross-modal contrastive objective to learn semantic and instance-level aligned point cloud representa + +tion. Experimental results on both indoor and outdoor scenarios show that our learned 3D representation has great transfer ability in downstream tasks, including zero-shot and few-shot 3D recognition, which boosts the state-of-the-art methods by large margins. Furthermore, we provide analyses of the capability of different representations in real scenarios and present the optional ensemble scheme. + +# 1. Introduction + +Powerful 3D point cloud representation plays a crucial role in various real-world applications, e.g., 3D object recognition and detection [10, 20, 31, 40, 44]. Compared to 2D images, 3D point cloud provides specific information like accurate geometry that is robust to illumination changes. However, current methods [25, 40] that learn 3D representations generally rely on the predefined number of object categories and require plenty of labor-intensive annotations. Those learned 3D representations are insufficient for safety-critical scenarios like self-driving which includes a long-tail class distribution far beyond the predefined taxonomy. Therefore, it is highly demanded to learn a transferable 3D representation equipped with zero-shot recognition ability in vocabulary scalable real-world scenes. Figure 1 shows an open-world recognition example by our $\mathrm{CLIP}^2$ in + +outdoor and indoor scenes, where the 3D objects can be classified with the correlation alignment between 3D representations and open-world vocabularies. + +The critical ingredient of open-world understanding is that the models learn sufficient knowledge to obtain general representations. To achieve this, recent Vision-Language Models (VLM) [14, 27, 38] leverage Internet-scale text-image pairs to conduct vision-language pretraining, which facilitates transferable 2D representation and demonstrates promising performance in 2D open-vocabulary tasks. However, 3D vision-language pretraining remains unexplored due to the limitation of existing 3D datasets in diversity and scale compared to the massive data sources in 2D counterparts [14, 15, 27, 38]. Though some recent works [12, 13, 43] try to avoid this problem by transferring the pretrained 2D VLM into the intermediate representation including projected image patches [12, 18] or depth maps [1, 43], those representations suffer from the loss of 3D geometric information and limited viewpoints under realistic scenarios. Especially the camera images are only sometimes available due to the sensor failure in 3D scenes. We believe the 3D representation based on original point cloud data retains most information and is the optimal solution for 3D real world understanding, which requires a rethink of learning the transferable 3D representation under realistic scenarios. + +To this end, we propose a Contrastive Language-Image-Point cloud Pretraining framework, short for $\mathrm{CLIP}^2$ , which directly aligns 3D space with broader raw text and advances the 3D representation learning into an open-world era. Our learning process can be decomposed into two stages: Firstly, we introduce a Triplet Proxy Collection to alleviate the limitation of accessible pretraining data by constructing language-image-point triplets from real-world scenes. Since the large-scale realistic 3D datasets for outdoor driving [2,19] and indoor scenarios [9,32] are collected in open-world, it contains huge amounts of realistic objects that vary in semantics and diversity. Thus we consider them as potential pretraining data sources without extra human supervision. Specifically, we propose "Proxy" instances as the bridges between language descriptions, 2D images and 3D point clouds. Enabled by a well-aligned VLM, a scalable caption list and the geometry transformation between 2D and 3D, we automatically create more than 1 million triplets to facilitate pretraining. Secondly, we further propose a Cross-Modal Pretraining scheme to jointly optimize the feature space alignments of three modalities, i.e., point cloud, language and image. It contains both the contrastive learning objective of semantic-level text-3D correlation and instance-level image-3D correlation, which contributes to better transferability of learned 3D representation. + +We study the transferable capability of $\mathrm{CLIP}^2$ by benchmarking the zero-shot recognition performance on four popular indoor and outdoor real-world datasets, and find a significant improvement over current methods, achieving Top1 accuracy $61.3\%$ on SunRGBD [32], $43.8\%$ on ScanNet [9]), + +28.8% on nuScenes [2] and 56.0% on ONCE [19]. For a fair comparison with existing methods [1, 13, 36, 43], we conduct zero-shot and few-shot classification on single object dataset ScanObjectNN [34] and find consistent dominance, 16.1% relative improvement on zero-shot classification over previous state-of-the-art method [13]. To validate the vocabulary-increasing ability of $\mathrm{CLIP}^2$ , we report the quantity results and visualizations to show the improved discovery of the long-tail categories. Moreover, we make ablations and analysis on different representations, and investigate ensembling alternatives to merge complementary knowledge of all available representations in realistic applications. Our contributions can be summarized as follows: + +- We propose a novel $\mathrm{CLIP}^2$ framework that aligns 3D space with open-world language representation, facilitating zero-shot transfer in realistic scenarios. +- We present a Triplet Proxies Collection scheme in real-world scenes, which alleviates the shortage of text-3D data sources and facilitates the pretraining methods. +- $\mathrm{CLIP}^2$ jointly optimizes the correlation alignment between point cloud, language and image by proposed cross-modal pretraining mechanism, which enhances the transferability of learned 3D representation. +- Our $\mathsf{CLIP}^2$ achieves the state-of-the-art zero-shot transfer performance on 5 datasets (indoor/outdoor scenes and single-object) and shows quality results on vocabulary-increasing discovery in real world. + +# 2. Related Work + +Vision-Language Model. Large vision language models (VLM) [14, 15, 27, 38] have demonstrated successful performance in downstream zero-shot tasks with the learned transferable 2D representations. CLIP [27] and ALIGN [14] push the limit by collecting Internet-scale image-text pairs and then learning the correlation alignment between image and language feature space with contrastive pretraining objectives. Those models can be directly transferred to zero-shot 2D recognition and achieve impressive results. Recent DetClip [38] learns to align image patches to test phrases after pretraining under hybrid supervision from detection, grounding and image-text pair data, which extends the ability to localize open-vocabulary 2D proposals in images. In this paper, we attempt to transfer the open-vocabulary ability of pre-trained VLM to the 3D domain, making language applicable to zero-shot point cloud recognition. + +Zero-shot/Open-world Learning in 3D. Recognizing 3D objects with a large vocabulary is necessary for safety-critical autonomous driving and robotic tasks, yet remains under-explored. Cheraghian et al. [5-8] first attempt to associate PointNet [25] feature with category semantic information via a projection function, and separately proposed an unsupervised skewness loss [5] to mitigate the hubness + +![](images/c9d4e4112331aca8c40b6ed1534e6c7f1bd3717b4519fc73bd58938f66cf5179.jpg) +Figure 2. Overview of $\mathbf{CLIP}^2$ framework. The main components contain two parts, the Triplet Proxy Collection and the Cross-Modal Pretraining. The defined Triplet Proxy set $\mathcal{D}_{\mathrm{proxy}}$ consists of language captions $\mathbf{X}^T$ , corresponding image instances $\mathbf{X}^I$ and raw 3D point cloud instances $\mathbf{X}^P$ , which come from the free data source under realistic scenarios without any labeling labor. On top of that, we pretrain a point cloud encoder $E^P$ with the cross-modal contrastive learning objective. Equipped with $\mathbf{CLIP}^2$ , the learned 3D point cloud representation $F^P$ is well aligned to the language representation, which facilitates downstream zero-shot 3D transfer tasks in the real world. + +problem. The transductive case [6] is discussed in which extends [5] using a triplet loss. Notably, the above works conduct experiments on synthetic datasets and need to divide datasets into "seen" categories as training data and "unseen" categories as testing data. Thus they are not suitable for realistic scenarios due to the domain gap between synthetic and real-world data, as well as the limited vocabulary-increasing ability. Recently, inspired by the success of VLMs [14, 27] in 2D tasks, some works [13, 43] propose to transfer the zero-shot recognition ability of pretrained CLIP [27] into 3D area. PointCLIP [43] directly projects point cloud into multi-view depth maps as image-like data input for pretrained CLIP to make classification predictions. While CLIP2Point [13] trains an image-depth embedding on ShapeNet [39] to better align the depth representation to the pretrained image space of CLIP. However, depth maps lost plenty of geometry information of the original point cloud data structure, resulting in poor performance especially in realistic scenarios. By contrast, we aim to learn transferable 3D representation based on the original point cloud data structure in realistic scenarios. + +3D Representation Learning. Much progress has been made in learning a comprehensive 3D representation in an unsupervised manner. Most works [1, 16, 17, 23, 30, 37, 41, 42] follow the paradigm that conducts pretraining on unlabeled datasets and then finetunes on the limited downstream annotations. Though the improved transferability of 3D representation, they can not be directly transferred to zero-shot tasks with open-world vocabularies. In this work, we con + +duct language-image-point cloud pretraining, which learns transferable 3D representation aligned to open-vocabulary language space to facilitate the zero-shot transfer. + +# 3. Method + +In this section, we introduce $\mathrm{CLIP}^2$ to learn a transferable 3D point cloud representation with arbitrary category recognition ability under realistic scenarios, illustrated in Figure 2. We will first present the Triplet Proxy Collection in Section 3.1, which utilizes a pretrained VLM and geometric transformation to obtain language-image-point triplets from real-world scenes. Then we will elaborate Cross-Modal Contrastive Pretraining mechanism in Section 3.2, which jointly optimizes the alignment correlations between language, image and point cloud feature space. + +# 3.1. Triplet Proxy Collection + +Inspired by the significant performance of 2D VLMs on open-vocabulary tasks, we aim to develop 3D vision-language pretraining to facilitate category-increasing capacity for real-world scenarios. However, the core challenge is the shortage of pretraining data. Compared to the 2D vision-language pretraining framework CLIP [27], which takes more than 400M image-language pairs from the Internet, the largest 3D single-object dataset ShapeNet [39] only contains 50K CAD models with 55 categories. In addition to the insufficiency of data scale, pretraining on such synthetic data fails to transfer well in the real world due to the huge domain gap. Enlightened by the recent emer + +![](images/a8da405e6ae9f064a476ddb3c2101091ec79bc7b6889540c99430eeeaf5a9e32.jpg) +Figure 3. Illustration of three representation modals of two 3D objects examples under indoor and outdoor scenarios. + +gence of large-scale point cloud datasets collected in indoor [9, 32] and outdoor scenarios [2, 19], we observe that those naturally-collected datasets potentially contain vast amounts of open-world objects that vary in semantics and diversity. Considering the data collection itself is cheap except for laborious annotation, we novelly take leverage of those available datasets without human annotations as a practical yet effective pretraining data source. + +Specifically, given the realistic scene data $S = \{(P_s, I_s)_{s=1}^{|S|}\}$ , where $P_s \in \mathbb{R}^{N_P \times 3}$ and $I_s \in \mathbb{R}^{N_I \times H \times W \times 3}$ are corresponding 3D point clouds and images of scene $s$ , we propose a novel concept, Proxy, as the bridge between language, image and 3D point cloud. As illustrated in Figure 2, equipped by those proxy instances, we can automatically collect a massive number of language-image-point cloud pairs $D_{\mathrm{proxy}}$ in the format of proxies under open-world scenes. We detail the process as follows. + +Language Proxy. We set the language proxies $\mathbf{X}^T\in \mathbb{R}^V$ as a raw text list from the 2D open-world dataset [11], where $V = 1206$ denotes the vocabulary size of language proxies. Image Proxy. Next, we obtain the image proxies $\mathbf{X}^I$ by an open vocabulary detector DetCLIP [38], denoted as $M$ which is trained with open-world data and performs openset detection. Concretely, given language proxies $\mathbf{X}^T$ and input scene image $I_{s}$ , we extract corresponding image proposals as image proxies $X_{s}^{I}$ with $M$ by the similarity between input language embeddings and proposal features as + +$$ +\left\{X _ {s} ^ {I} \right\} _ {s \in | \mathcal {S} |} = \mathrm {M} \left(\left\{I _ {s} \right\} _ {s \in | \mathcal {S} |}, \mathbf {X} ^ {T}\right). \tag {1} +$$ + +3D Proxy. We exploit the naturally-existed geometry relations between 2D and 3D scenes to obtain 3D proxies $\mathbf{X}^P$ , which consists of point cloud instances corresponding to image proposals in $\mathbf{X}^I$ . We simplify the geometry transformation as $\mathrm{G}(\cdot)$ and formulate the relations as: + +$$ +X _ {i} ^ {P} = \mathrm {G} \left(X _ {i} ^ {I}\right). \tag {2} +$$ + +Detailedly, for indoor scenes equipped with RGB-D sensors, we first remove the background pixels by unsupervised segmentation algorithm [28] for each image proxy $X_{s,i}^{I}$ , $i \in |X_s^I|$ . Since depth information is known, we then transform the segmented pixels from uvd coordinate $X_{s,i}^{I,uvd} \in \mathbb{R}^{n,3}$ to xyz coordinate $X_{s,i}^{P,xyz} \in \mathbb{R}^{n,3}$ as a 3D + +![](images/c658027c3b01323f7c826a78da3e707d1381ceadaf174e4edb9bb24c7406eded.jpg) +Figure 4. Comparison of different pretraining strategies. (a) CLIP aligns image and language embedding space [27] as $L_{TI}$ based on large-scale text-image pairs. (b) PointClip [27] aligns projected depth map to CLIP language space as $L_{TD}$ . (c) Clip2Point aligns depth map to CLIP image space as $L_{ID}$ . (d) our $\mathrm{CLIP}^2$ aligns original 3D point cloud to both CLIP language space and image space via cross-modal objective $L_{CM}$ . + +point cloud proxy with the given camera parameters. For outdoor scenes captured by LiDAR sensors, we first create a 3D frustum for each image proxy by extruding the 2D image proposal into 3D space following [22, 24]. Then we conduct DBSCAN algorithm [29] within frustum and select the point cloud cluster as the point proxy $X_{s,i}^{P,xyz}$ . + +Eventually, we construct Triplet Proxy $\mathcal{D}_{\mathrm{proxy}} = \{\mathbf{X}^T,X_s^I,X_s^P\}_{s = 1}^{|\mathcal{S}|}$ by combining corresponding language proxies $\mathbf{X}^T$ , image proxies $\mathbf{X}^I$ and 3D proxies $\mathbf{X}^P$ , where $\mathbf{X}^I = \{X_s^I\}_{s = 1}^{|\mathcal{S}|}$ and $\mathbf{X}^P = \{X_s^P\}_{s = 1}^{|\mathcal{S}|}$ . $220\mathrm{K}$ and $1.4\mathrm{M}$ proxy triplets are formed for indoor and outdoor scenes, respectively. More details can be found in the appendix. + +# 3.2. Cross-Modal Contrastive Pretraining + +With the triplet proxies $\mathcal{D}_{\mathrm{proxy}}$ , a straightforward pretraining objective is forcing the alignment between the embedding spaces of point cloud $X_{i}^{P}$ and language $X_{i}^{T}$ from scratch. However, it might not promise good transferability of learned representation, since the number of language-image-point pretraining data triplets remains two orders of magnitude smaller than the language-image pairs adopted by CLIP [27] and the vocabulary size is much more limited. Therefore, we design to learn the correlation alignment based on the pretrained embedding space of CLIP. + +The comparison of current pretraining strategies [13,43] is illustrated in Figure 4, which is a series of 3D variants of CLIP. Notably, both existing methods exploit projected depth map as the intermediate representation of point cloud, which are respectively learned to align to language space [43] and image space [13]. Intuitively, as illustrated in Figure 3, depth representation lost plenty of geometry information compared to the original point cloud, especially in outdoor scenarios. Moreover, images are sometimes unavailable for 3D objects. Thus we conduct pretraining on original 3D point cloud data as an optimal representation. + +Toward learning more transferable representation, we introduce a cross-modal contrastive learning objective to jointly optimize the correlation alignment across lan + +guage, image and point cloud, including Semantic-Level Language-3D Alignment and Instance-Level Image-3D Alignment. Specifically, the overall architecture of $\mathrm{CLIP}^2$ , shown in Figure 2, contains language encoder $E_{\theta}^{T}$ , point cloud encoder $E_{\theta}^{P}$ and visual encoder $E_{\theta}^{I}$ , which respectively embed the triplet proxies into text feature $f^{T} \in \mathcal{R}^{1 \times C_{T}}$ , point cloud feature $f^{P} \in \mathcal{R}^{1 \times C_{P}}$ and image feature $f^{I} \in \mathcal{R}^{1 \times C_{I}}$ , where $C$ is the embedding dimension. + +Semantic-Level Language-3D Alignment. In order to inherit the open-world recognition ability from pretrained CLIP [27], we align the point cloud feature $f^{P}$ with text embedding $f^{T}$ from well-trained CLIP with Language-Point Proxy $\{X_{i}^{T}, X_{i}^{P}\}$ input. We replace classname in the prompts, like "point cloud of a {classname} ." with raw text in proxy $X_{i}^{T}$ as language sentences. The core idea is to drive the feature centroids of 3D instances and the corresponding text prompt closer. We compute the loss function of between of language proxy and point cloud proxy as: + +$$ +l (i, T, P) = - \log \frac {\exp \left(f _ {i} ^ {T} \cdot f _ {i} ^ {P} / \tau\right)}{\exp \left(f _ {i} ^ {T} \cdot f _ {i} ^ {P} / \tau\right) + \sum_ {j \in N , X _ {j} ^ {T} \neq X _ {i} ^ {T}} \exp \left(f _ {i} ^ {T} \cdot f _ {j} ^ {P} / \tau\right)}, \tag {3} +$$ + +where $N$ is the mini-batch size, $\tau$ is the temperature coefficient. Within a training mini-batch, the language-3D alignment objective $L(T,P)$ can be described as: + +$$ +L (T, P) = \frac {1}{N} \sum_ {i \in N} l (i, T, P). \tag {4} +$$ + +Instance-Level Image-3D Alignment. In addition to the alignment between semantic language and 3D proxy instances, we further introduce the contrastive alignment between instance-wise image proxy and 3D proxy instances. Note that the instance-aware visual concept has been well-studied in the embedding space of pretrained CLIP. We believe instance-sensitive learning can contribute to further correlation learning and benefits to the transferability of learned 3D representation. The contrastive aligned objective $L(I,P)$ across point cloud and image is formulated as: + +$$ +l (i, I, P) = - \log \frac {\exp \left(f _ {i} ^ {I} \cdot f _ {i} ^ {P} / \tau\right)}{\exp \left(f _ {i} ^ {I} \cdot f _ {i} ^ {P} / \tau\right) + \sum_ {j \in N , j \neq i} \exp \left(f _ {i} ^ {J} \cdot f _ {j} ^ {P} / \tau\right)}, \tag {5} +$$ + +$$ +L (I, P) = \frac {1}{N} \sum_ {i \in N} l (i, I, P). \tag {6} +$$ + +Finally, we obtain the resultant cross-modal contrastive learning objective $L_{CM}(T,I,P)$ as the combination of $L(T,P)$ and $L(I,P)$ , where both alignments of semantic-level text-3D correlation and instance-level image-3D correlation are injected: + +$$ +L _ {C M} (T, I, P) = \lambda_ {1} L (T, P) + \lambda_ {1} L (I, P), \tag {7} +$$ + +where the hyper-parameters $\lambda_{1}$ and $\lambda_{2}$ are both set to 0.5. + +# 4. Experiment + +In this section, we evaluate $\mathrm{CLIP}^2$ on realistic indoor and outdoor scenarios. We report the zero-shot transfer results on various datasets [2, 9, 32-34] and make further analysis on the designs of pretraining strategy. + +# 4.1. Zero-shot Transfer + +Setting. After pretraining, natural language is applied to reference learned 3D representation to enable following zero-shot transfer tasks. (i) Zero-Shot Recognition: we evaluate zero-shot recognition performance for realistic objects, where $K$ category names are transferred to text prompt "point cloud of {CLASS}" to encode the text features $F_{K} \in \mathbb{R}^{K \times C}$ . Then the classification logits are calculated with the 3D feature $f^{P}$ and text features as: + +$$ +\log_ {i} = \operatorname {s o f t m a x} \left(f _ {i} ^ {P} \left(F _ {K}\right) ^ {T}\right). \tag {8} +$$ + +We present the results under both indoor and outdoor scenarios in Table 1, Table 2 and Table 5, as well as the object-level benchmark in Table 6. (ii) Open-vocabulary recognition: we enlarge the category vocabularies of ScanNet to 249 and 384 to study the open-vocabulary recognition ability in Table 3. (iii) Open-vocabulary localization: we study the open-vocabulary localization ability by localizing open-world 3D objects with our proxy generation process and then classifying them with our learned 3D representation, of which the visualization is illustrated in Figure 5 and evaluation results are reported in Table 4. Notably, we investigate representation ensembling alternatives to enable knowledge merging of all available representations for realistic applications, illustrated in Table 8. + +# 4.1.1 Indoor Scenarios + +Datasets and details. We adopt the widely used indoor 3D dataset SUN RGB-D [32] as the realistic indoor scenario that provides pretraining data source, a single-view RGB-D dataset consisting of $\sim 10\mathrm{K}$ scenes. To validate the transferability of learned 3D representation, we also evaluate another popular indoor 3D dataset ScanNet [9], which contains $\sim 1.5\mathrm{K}$ scenes of 3D reconstructed meshes. We remove objects in ScanNet with less than 5 points, leaving 384 noisy categories. For open-vocabulary recognition, we evaluate performance on the ScanNet 384-class set and a 249-class merged set. In addition to the scene-wise indoor dataset, we conduct evaluations on ScanObjectNN [34], which collects $\sim 3\mathrm{K}$ individual realistic objects with 15 categories and is applied in the previous zero-shot evaluation [13, 43]. During the proxy collection process, we empirically set $\epsilon = 0.3$ in [38] as a tradeoff between filtering FPs and preserving TPs to generate image proxies. Considering the occurrence frequencies of different indoor categories vary a lot, we adopt the class balance strategy [4] to mitigate the class imbalance. During pretraining process, we adopt [26] as point cloud encoder and set the overall training epoch number to 100. + +
MethodAvg.BedBookshelfChairDeskSofaTabletoiletBathhubDresserNight Stand
PointClip [43]11.50.094.00.00.00.014.70.00.06.10.0
Clip2Point [13]18.610.920.664.334.413.814.126.20.01.40.0
PointClip [43] w/ TP.38.045.3100.062.548.544.44.855.216.33.30
Clip2Point [13] w/ TP.56.978.087.636.236.664.737.482.177.567.61.2
CLIP²61.384.075.570.747.375.533.886.265.371.82.4
CLIP² w/ En.69.687.394.370.754.179.347.091.785.782.23.6
+ +Table 1. Zero-shot recognition results in SUN RGB-D. Avg.: the mean average Top1 accuracy across all categories. TP.: our Triplet Proxy set. En.: our optional ensembling scheme for inference. + +
MethodAvg.CabBedChairSofaTablDoorWindBksfPicCntrDeskCurtFridgBathShowrToilSink
PointClip [43]6.30.00.00.00.00.70.00.091.80.00.00.015.00.00.00.00.00.0
Clip2Point [13]24.90.020.885.143.326.569.90.020.91.731.727.00.01.646.50.022.425.6
PointClip [43] w/ TP.26.155.70.072.85.05.11.70.077.20.00.051.70.30.00.040.385.349.2
Clip2Point [13] w/ TP.35.23.011.845.127.610.561.52.671.90.333.629.94.711.572.292.486.134.0
CLIP²38.567.232.669.342.318.319.14.062.61.412.752.840.19.159.741.071.045.5
+ +Table 2. Zero-shot recognition in ScanNet. Avg.: the mean average Top1 accuracy across all categories. TP.: our Triplet Proxy set. + +
MethodTop5 Acc.
384cls.249cls.
[43]0.30.4
[13]6.47.0
CLIP222.031.7
+ +
MethodmAP25IN AR25mIoUOUT P.R.
[21]1.3----
[18]13.037.7---
CLIP212.743.052.86.787.4
+ +Table 3. Recognition results. Table 4. Zero-shot localization results of vocabulary ex-sults. IN: indoor scenario SUN RGBpansion in ScanNet. D. OUT: outdoor scenario nuScenes. + +Quantity results. For zero-shot recognition task, we take two recent works as our baselines, i.e. PointClip [43] and Clip2Point [13], which study the zero-shot classification task on 3D object-level benchmarks [34, 35] by leveraging pretrained CLIP with projected depth maps. Focusing on the real-world scenarios, we conduct comparison not only on the realistic object-level [34] as illustrated in Table 6 but also on the scene-level datasets shown in Table 1 and Table 2, where the evaluation follows the common classes split in [10, 44] and reports the instance Top1 accuracy of each class. As shown in tables, our $\mathrm{CLIP}^2$ can outperform baselines on all benchmarks by large margins. Besides, we apply our triplet proxy generation mechanism (TP.) to baseline methods, and achieve considerable improvements on SUN RGB-D and ScanNet by $26.5\%$ and $19.8\%$ for Point-Clip, $38.3\%$ and $10.3\%$ for Clip2Point. On the one hand, the contrasts demonstrate the effectiveness of our triplet proxies for open-world understanding. On the other hand, our learned 3D representation is superior in 3D object recognition by retaining more 3D-specific information than depth representation. Besides, we present the optional ensembling scheme (En.) when camera images are available, which can take advantage of multi-modal knowledge and further boost the performance by $8.3\%$ . To further validate the open-vocabulary recognition ability, we conduct evaluation on a larger category set of ScanNet in Table 3 and report the instance Top5 accuracy, which illustrates the superior + +ity of our $\mathrm{CLIP}^2$ when vocabulary increases. Beyond that, $\mathrm{CLIP}^2$ is also equipped with zero-shot 3D localization ability by proxy generation. On indoor scenario SUN RGB-D, we compare with a SOTA indoor 3D detector 3DETR [21] and a recent work OV3D [18] that studies open-vocabulary detection, where evaluation is conducted on the same "unseen" split in [18]. Since $\mathrm{CLIP}^2$ do not fit the tight bounding boxes of point cloud instances, we estimate the maximum bounding box of proxies and GT instances to conduct evaluation following the same metrics $\mathrm{mAP}_{25}$ and $\mathrm{AR}_{25}$ in [18], as shown in Table 4. Notably, compared to baseline works that train on "seen" 3D annotations and test on "unseen" categories, we have no access to any 3D annotations yet achieve comparable localization ability, which yields $5.3\%$ $\mathrm{AR}_{25}$ improvement over OV3D [18]. We further evaluate segmentation results in Table 4. + +Quality results. The visualization results of $\mathrm{CLIP}^2$ under a indoor scene of SUN RGB-D [32] is shown in Figure 5(a). Our triplet proxy generation process can localize open-world 3D objects in a point cloud scene. Moreover, the 3D representation learned from our cross-modal pretraining can provide more accurate classification results for 3D instances by exploiting original point cloud, which corrects the mistaken "People" prediction in image to "Picture" by considering the geometry information. + +# 4.1.2 Outdoor Scenarios + +Datasets and details. We exploit a prevalent large-scale 3D dataset nuScenes [2] as the outdoor data source and extra validate the performance on the ONCE dataset [19]. The nuScenes dataset consists of $\sim 28\mathrm{K}$ frames with 10 categories, while ONCE contains 6 annotated sequences with 5 categories. Similarly, we set the $\epsilon = 0.3$ for image proxies collection and adopt the class balance strategy [4]. + +Quantity results. Since the outdoor point cloud is collected by LiDAR sensors, it has a wider perception range + +
MethodAvg.nuScenesONCE
CarTruckBusPed.BicycleTrailerC.V.Motor.BarrierT.C.CarCyc.Ped.TruckBus
PointClip [43]11.70.00.00.029.141.83.40.00.142.50.00.013.879.20.27.61
Clip2Point [13]12.40.40.30.113.531.01.89.31.566.20.317.311.895.435.74.0
PointClip [43] w/ TP.28.318.805.574.017.957.01.94.52.129.751.89.299.85.046.8
Clip2Point [13] w/ TP.33.026.716.851.245.215.813.920.05.710.534.239.427.895.540.651.7
CLIP²37.841.941.322.540.321.120.624.822.417.335.352.727.377.778.544.0
+ +![](images/d4d7bd0c222aa751cda04a21ff7675f53f7b9569fb4b711efc61ca0fe9b02bfd.jpg) +Figure 5. Visualizations of the zero-shot localization and recognition results by $\mathrm{CLIP}^2$ under open-world (a) indoor realistic scene [32] and (b) outdoor scenes [2]. Notably, the whole pipeline of $\mathrm{CLIP}^2$ not only has no access to human annotations, but also enables the open-world vocabularies beyond groundtruth annotations, such as 'Picture' in (a) and 'Plastic bag', 'Tire' in (b). Best viewed in colors. + +![](images/7acfa580419ef00bd6b679fccb2d957142c4c7f782fc29dcc91736b580390ac4.jpg) + +Table 5. Zero-shot recognition results in outdoor scenario: nuScenes (Left) and ONCE (Right). TP.: our Triplet Proxy set. Avg.: the mean average Top1 accuracy across all categories of two benchmarks. + +
MethodZS15-way10-way
4-S8-S16-S10-S20-S
PointNet++ [25]-41.047.655.0--
PointCLIP [43]15.446.050.055.6--
Clip2Point [13]23.3-----
CrossPoint [1]----58.7±1.864.6±1.2
CLIP239.151.359.662.560.6±2.566.3±3.2
+ +Table 6. Zero-shot and Few-shot classification results on ScanObjectNN. ZS: zero-shot. K-way N-shot: few-shot settings. + +than RGB-D but leads to sparse distribution. Thus the projected depth representation of baselines results in severer information lost, as illustrated in the second row in Figure 3. As shown in Table 5, our $\mathrm{CLIP}^2$ considerably outperforms the baseline recognition results by more than $20\%$ , and our triplet proxies respectively boost two baselines by $9.5\%$ and $4.8\%$ . Additionally, we evaluate the localization ability on the outdoor scenario nuScenes in Table 4. Due to the lack of works that tackle outdoor open-vocabulary localization problems, we follow classic detection accuracy metrics Precision(P.) and Recall(R.) as evaluation metrics. Specifically, we calculate the center distance between groundtruth bounding boxes and our 3D proxies that are predicted to belong to the same category as the groundtruth, and set the distance threshold as $\lambda = 2\mathrm{m}$ . For those matched pairs that are closer than $\lambda$ , we count the proxies as TPs. Otherwise, for those unmatched proxies and groundtruth, we count them as FPs and FNs respectively, thus $\mathrm{P.} = \frac{\mathrm{TPs}}{\mathrm{TPs} + \mathrm{FPs}}$ , $\mathrm{R.} = \frac{\mathrm{TPs}}{\mathrm{TPs} + \mathrm{FNs}}$ . As shown in Table 4, our $\mathrm{CLIP}^2$ pipeline can provide high + +recall for outdoor objects. Since $\mathrm{CLIP}^2$ is highly sensitive to open-world objects and can perceive categories beyond groundtruth list, it tends to create overmuch predictions thus the precision is comparably low. The perception ability of open-world objects can be viewed in Figure 5(b). + +Quality results. We show off two outdoor scenes of nuScenes [2] in Figure 5(b-i) and Figure 5(b-ii). In addition to perceiving those common categories, our $\mathrm{CLIP}^2$ surprisingly localizes and recognizes those uncommon 3D objects in 3D scenes such as the tires of vehicles, the plastic bag in the hand of pedestrian as well as the plastic bag on the road. We believe it contributes to auto-driving safety by providing the localization and recognition of universal obstacles to facilitate follow-up driving decisions. + +# 4.2. Few-shot Classification + +Setting. Lightweight few-shot learning is practical for application by finetuning the pretraining model with given limited data annotations, which can also validate the generalization capability of our learned representation. To make a fair comparison, we follow the existing methods [1, 43] to conduct experiments under "K-way N-shot" setting on the challenging realistic object-level dataset ScanObjectNN [34], where we randomly sample N point cloud objects from each of the randomly selected K classes. + +Quantity results. As illustrated in Table 6, we compare with representative 3D networks PointNet++ [26], the recent zero-shot approach PointClip [43] as well as a state- + +
Rep.Obj.Avg..INAvg..OUTAvg..OBJ
(a)DepthLang.-Depth38.09.331.8
(b)DepthImage-Depth56.421.739.0
(c)PC.Lang.-Point52.626.837.7
(d)PC.Image-Point56.524.437.8
(e)PC.Lang.-Image-Point61.328.839.4
+ +of-the-art representation learning method CrossPoint [1], which conducts contrastive pretraining between point cloud and rendered images on CAD dataset ShapeNet [3]. As we can see, with a slight number of samples, our $\mathrm{CLIP}^2$ can boost the classification results by a large margin, exceeding PointClip by $5.3\%$ , $9.6\%$ and $6.9\%$ with 4, 8 and 16 shots. Besides, we outperform CrossPoint with considerable gain, illustrating our pretraining strategy on collected proxies can learn sufficient knowledge from realistic open world to generate transferable 3D representation, which is superior to the pretraining on a small-scale synthetic dataset. + +# 4.3. Ablations and Analysis + +Ablations on representation learning. To observe the transferability of different representations and the effect of different learning objectives, we conduct ablations and report the mean average Top1 accuracy across all classes of zero-shot recognition in indoor scenarios [32] (Avg..IN), outdoor scenarios [2] (Avg..OUT) and object-level benchmark [34] (Avg..OBJ), which is shown in Table 7. Firstly, for fair comparisons, we follow [13, 43] to project input point cloud into depth maps in $N_V$ different views as alternative representation. Secondly, we adopt various objectives to learn different correlation alignments across language, image and point cloud feature space or depth space. Specifically, comparing (a) and (b), aligning depth space to image space yields better transfer performance due to the similar data structure of image and depth map. In (c) and (d), point cloud representation is better when aligning to image space in indoor scenes, while better to align with language space in outdoor scenes due to the data discrepancy between image-like RGB-D points and sparse LiDAR points. Generally, 3D point cloud representation outperforms depth representation in all benchmarks due to preserving the complete 3D structure and sufficient 3D-specific information. Comparing (e) and (d), the joint alignment between three feature spaces contributes to the best 3D point cloud representation transferability on all benchmarks. + +Analysis of representation ensembling. Intuitively, different representations contain different perspectives of knowledge, which can be potentially merged to achieve the optimum results during inference. To validate the ensembling application, we adopt three optional representation modals, i.e. point clouds, projected depth maps and cor + +Table 7. Ablations on representation learning. Rep.: the recognition representations. Obj.: the learning objectives. PC.: point cloud. Lang.: language space. + +
PC.ImageDepthAvg..INAvg..OUTAvg..OBJ
(i)✓61.328.839.4
(ii)✓64.241.1-
(iii)✓56.923.939.0
(f)✓✓68.743.9-
(g)✓✓64.830.443.2
(h)✓✓✓69.642.3-
+ +Table 8. Analysis on the representation ensembling schemes. + +responding image patches, where depth representation is trained on our proxies and image representation is generated from pretrained image branch of CLIP. We ensemble their predicted logits by simple summation as the final output, and illustrate the separate recognition results and ensembling performance in Table 8. Benefiting from the sufficient knowledge learned from massive CLIP training data, image representation presents best performance in separate applications. By merging the complementary knowledge, our 3D representation leads to gains of $4.6\%$ indoors [32] and $2.8\%$ outdoors. Though further improving the indoor recognition performance with $0.9\%$ when merged, depth representation yeilds $1.6\%$ drop for outdoor objects, illustrating the information lost especially for outdoor scenarios. Since image representation is sometimes missing, such as in [34], our 3D representation is more robust for 3D applications. + +# 5. Limitation + +As a pilot work for the language-3D pretraining problem, though $\mathrm{CLIP}^2$ enables zero-shot localization and recognition with proposed triplet proxy generation and learned transferable 3D representation, it can not provide the accurate tight bounding box for open-world 3D objects as a common detector does. We believe $\mathrm{CLIP}^2$ can facilitate the development of open-world 3D detectors by introducing the recognition ability to general 3D detectors or providing presented 3D proxies to enable further training of 3D detectors. + +# 6. Conclusion + +In this paper, we present a novel contrastive language-image-point cloud pretraining framework, $\mathrm{CLIP}^2$ , which consists of a triplet proxy collection scheme and a cross-modal contrastive learning mechanism. Based on the observation that realistic scenarios contain a massive amount of open-world objects, we innovatively propose to collect triplet proxies from realistic scenes as pretraining data. We then conduct cross-modal contrastive alignment across language, image and point cloud feature space to learn transferable 3D representation. The zero-shot transfer results on various indoor and outdoor benchmarks validate the ability of $\mathrm{CLIP}^2$ for 3D open-world understanding. + +Acknowledgements We acknowledge the support of MindSpore1, CANN, and Ascend AI Processor in this work. + +# References + +[1] Mohamed Afham, Isuru Dissanayake, Dinithi Dissanayake, Amaya Dharmasiri, Kanchana Thilakarathna, and Ranga Rodrigo. Crosspoint: Self-supervised cross-modal contrastive learning for 3d point cloud understanding. In CVPR, 2022. 2, 3, 7, 8 +[2] Holger Caesar, Varun Bankiti, Alex H Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuscenes: A multimodal dataset for autonomous driving. In CVPR, 2020. 2, 4, 5, 6, 7, 8 +[3] Angel X Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, et al. Shapenet: An information-rich 3d model repository. arXiv preprint arXiv:1512.03012, 2015. 8 +[4] Kai Chen, Jiaqi Wang, Jiangmiao Pang, Yuhang Cao, Yu Xiong, Xiaoxiao Li, Shuyang Sun, Wansen Feng, Ziwei Liu, Jiarui Xu, Zheng Zhang, Dazhi Cheng, Chenchen Zhu, Tianheng Cheng, Qijie Zhao, Buyu Li, Xin Lu, Rui Zhu, Yue Wu, Jifeng Dai, Jingdong Wang, Jianping Shi, Wanli Ouyang, Chen Change Loy, and Dahua Lin. MMDetection: Open mmlab detection toolbox and benchmark. arXiv preprint arXiv:1906.07155, 2019. 5, 6 +[5] Ali Cheraghian, Shafin Rahman, Dylan Campbell, and Lars Petersson. Mitigating the hubness problem for zero-shot learning of 3d objects. In BMVC, 2019. 2, 3 +[6] Ali Cheraghian, Shafin Rahman, Dylan Campbell, and Lars Petersson. Transductive zero-shot learning for 3d point cloud classification. In WACV, 2020. 2, 3 +[7] Ali Cheraghian, Shafin Rahman, Townim F Chowdhury, Dylan Campbell, and Lars Petersson. Zero-shot learning on 3d point cloud objects and beyond. IJCV, 2022. 2 +[8] Ali Cheraghian, Shafin Rahman, and Lars Petersson. Zeroshot learning of 3d point cloud objects. In MVA. IEEE, 2019. 2 +[9] Angela Dai, Angel X Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In CVPR, 2017. 2, 4, 5 +[10] Zhipeng Ding, Xu Han, and Marc Niethammer. Votenet: A deep learning label fusion method for multi-atlas segmentation. In MICCAI, 2019. 1, 6 +[11] Agrim Gupta, Piotr Dollar, and Ross Girshick. Lvis: A dataset for large vocabulary instance segmentation. In CVPR, pages 5356-5364, 2019. 4 +[12] Huy Ha and Shuran Song. Semantic abstraction: Openworld 3d scene understanding from 2d vision-language models. In CoRL, 2022. 2 +[13] Tianyu Huang, Bowen Dong, Yunhan Yang, Xiaoshui Huang, Rynson WH Lau, Wanli Ouyang, and Wangmeng Zuo. Clip2point: Transfer clip to point cloud classification with image-depth pre-training. arXiv preprint arXiv:2210.01055, 2022. 2, 3, 4, 5, 6, 7, 8 +[14] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In ICML. PMLR, 2021. 2, 3 + +[15] Junnan Li, Ramprasaath Selvaraju, Akhilesh Gotmare, Shafiq Joty, Caiming Xiong, and Steven Chu Hong Hoi. Align before fuse: Vision and language representation learning with momentum distillation. NeurIPS, 2021. 2 +[16] Hanxue Liang, Chenhan Jiang, Dapeng Feng, Xin Chen, Hang Xu, Xiaodan Liang, Wei Zhang, Zhenguo Li, and Luc Van Gool. Exploring geometry-aware contrast and clustering harmonization for self-supervised 3d object detection. In ICCV, 2021. 3 +[17] Haotian Liu, Mu Cai, and Yong Jae Lee. Masked discrimination for self-supervised learning on point clouds. In ECCV, 2022. 3 +[18] Yuheng Lu, Chenfeng Xu, Xiaobao Wei, Xiaodong Xie, Masayoshi Tomizuka, Kurt Keutzer, and Shanghang Zhang. Open-vocabulary 3d detection via image-level class and de-biased cross-modal contrastive learning. arXiv preprint arXiv:2207.01987, 2022. 2, 6 +[19] Jiageng Mao, Minzhe Niu, Chenhan Jiang, Hanxue Liang, Jingheng Chen, Xiaodan Liang, Yamin Li, Chaoqiang Ye, Wei Zhang, Zhenguo Li, et al. One million scenes for autonomous driving: Once dataset. arXiv preprint arXiv:2106.11037, 2021. 2, 4, 6 +[20] Jiageng Mao, Shaoshuai Shi, Xiaogang Wang, and Hongsheng Li. 3d object detection for autonomous driving: A review and new outlooks. arXiv preprint arXiv:2206.09474, 2022.1 +[21] Ishan Misra, Rohit Girdhar, and Armand Joulin. An end-to-end transformer model for 3d object detection. In ICCV, 2021. 6 +[22] Anshul Paigwar, David Sierra-Gonzalez, Özgür Erkent, and Christian Laugier. Frustum-pointpillars: A multi-stage approach for 3d object detection using rgb camera and lidar. In ICCV, 2021. 4 +[23] Yatian Pang, Wenxiao Wang, Francis EH Tay, Wei Liu, Yonghong Tian, and Li Yuan. Masked autoencoders for point cloud self-supervised learning. In ECCV, 2022. 3 +[24] Charles R Qi, Wei Liu, Chenxia Wu, Hao Su, and Leonidas J Guibas. Frustum pointnets for 3d object detection from rgb-d data. In CVPR, 2018. 4 +[25] Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. Pointnet: Deep learning on point sets for 3d classification and segmentation. In CVPR, 2017. 1, 2, 7 +[26] Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J Guibas. Pointnet++: Deep hierarchical feature learning on point sets in a metric space. NeurIPS, 2017. 5, 7 +[27] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, 2021. 2, 3, 4, 5 +[28] Carsten Rother, Vladimir Kolmogorov, and Andrew Blake. "grabcut" interactive foreground extraction using iterated graph cuts. TOG, 2004. 4 +[29] Erich Schubert, Jörg Sander, Martin Ester, Hans Peter Kriegel, and Xiaowei Xu. Dbscan revisited, revisited: why and how you should (still) use dbscan. *TODS*, 42(3):1-21, 2017. 4 +[30] Charu Sharma and Manohar Kaul. Self-supervised few-shot learning on point clouds. NeurIPS, 2020. 3 + +[31] Shaoshuai Shi, Xiaogang Wang, and Hongsheng Li. Pointcnn: 3D object proposal generation and detection from point cloud. In CVPR, 2019. 1 +[32] Shuran Song, Samuel P Lichtenberg, and Jianxiong Xiao. Sun rgb-d: A rgb-d scene understanding benchmark suite. In CVPR, 2015. 2, 4, 5, 6, 7, 8 +[33] Pei Sun, Henrik Kretzschmar, Xerxes Dotiwalla, Aurelien Chouard, Vijaysai Patnaik, Paul Tsui, James Guo, Yin Zhou, Yuning Chai, Benjamin Caine, et al. Scalability in perception for autonomous driving: Waymo open dataset. In CVPR, 2020. 5 +[34] Mikaela Angelina Uy, Quang-Hieu Pham, Binh-Son Hua, Thanh Nguyen, and Sai-Kit Yeung. Revisiting point cloud classification: A new benchmark dataset and classification model on real-world data. In ICCV, 2019. 2, 5, 6, 7, 8 +[35] Kashi Venkatesh Vishwanath, Diwaker Gupta, Amin Vahdat, and Ken Yocum. Modelnet: Towards a datacenter emulation environment. In 2009 IEEE Ninth International Conference on Peer-to-Peer Computing. IEEE, 2009. 6 +[36] Hanchen Wang, Qi Liu, Xiangyu Yue, Joan Lasenby, and Matt J Kusner. Unsupervised point cloud pre-training via occlusion completion. In ICCV, 2021. 2 +[37] Saining Xie, Jiatao Gu, Demi Guo, Charles R Qi, Leonidas Guibas, and Or Litany. Pointcontrast: Unsupervised pretraining for 3d point cloud understanding. In ECCV, 2020. 3 +[38] Lewei Yao, Jianhua Han, Youpeng Wen, Xiaodan Liang, Dan Xu, Wei Zhang, Zhenguo Li, Chunjing Xu, and Hang Xu. Detclip: Dictionary-enriched visual-concept paralleled pretraining for open-world detection. In NeurIPS, 2022. 2, 4, 5 +[39] Yangyang Ye, Houjin Chen, Chi Zhang, Xiaoli Hao, and Zhaoxiang Zhang. Sarpnet: Shape attention regional proposal network for lidar-based 3d object detection. Neurocomputing, 2020. 3 +[40] Tianwei Yin, Xingyi Zhou, and Philipp Kähenbuhl. Center-based 3d object detection and tracking. arXiv preprint arXiv:2006.11275, 2020. 1 +[41] Xumin Yu, Lulu Tang, Yongming Rao, Tiejun Huang, Jie Zhou, and Jiwen Lu. Point-bert: Pre-training 3d point cloud transformers with masked point modeling. In CVPR, 2022. 3 +[42] Renrui Zhang, Ziyu Guo, Peng Gao, Rongyao Fang, Bin Zhao, Dong Wang, Yu Qiao, and Hongsheng Li. Pointm2ae: Multi-scale masked autoencoders for hierarchical point cloud pre-training. arXiv preprint arXiv:2205.14401, 2022.3 +[43] Renrui Zhang, Ziyu Guo, Wei Zhang, Kunchang Li, Xu-peng Miao, Bin Cui, Yu Qiao, Peng Gao, and Hongsheng Li. Pointclip: Point cloud understanding by clip. In CVPR, 2022. 2, 3, 4, 5, 6, 7, 8 +[44] Zaiwei Zhang, Bo Sun, Haitao Yang, and Qixing Huang. H3dnet: 3d object detection using hybrid geometric primitives. In ECCV, 2020. 1, 6 \ No newline at end of file diff --git a/2023/CLIP2_ Contrastive Language-Image-Point Pretraining From Real-World Point Cloud Data/images.zip b/2023/CLIP2_ Contrastive Language-Image-Point Pretraining From Real-World Point Cloud Data/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..fb932c49d5fe41f7d6dec48781db054d51e35578 --- /dev/null +++ b/2023/CLIP2_ Contrastive Language-Image-Point Pretraining From Real-World Point Cloud Data/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:752d608b8b4f5593ad77e25c6e4894edce3d5a497fe18ce5e696ad9a226ab5cf +size 677933 diff --git a/2023/CLIP2_ Contrastive Language-Image-Point Pretraining From Real-World Point Cloud Data/layout.json b/2023/CLIP2_ Contrastive Language-Image-Point Pretraining From Real-World Point Cloud Data/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..56d5ba64457e674294301f013f0d154413aa0c52 --- /dev/null +++ b/2023/CLIP2_ Contrastive Language-Image-Point Pretraining From Real-World Point Cloud Data/layout.json @@ -0,0 +1,8672 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 109, + 102, + 484, + 136 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 109, + 102, + 484, + 136 + ], + "spans": [ + { + "bbox": [ + 109, + 102, + 484, + 136 + ], + "type": "text", + "content": "CLIP" + }, + { + "bbox": [ + 109, + 102, + 484, + 136 + ], + "type": "inline_equation", + "content": "^2" + }, + { + "bbox": [ + 109, + 102, + 484, + 136 + ], + "type": "text", + "content": ": Contrastive Language-Image-Point Pretraining from Real-World Point Cloud Data" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 108, + 159, + 484, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 108, + 159, + 484, + 190 + ], + "spans": [ + { + "bbox": [ + 108, + 159, + 484, + 190 + ], + "type": "text", + "content": "Yihan Zeng" + }, + { + "bbox": [ + 108, + 159, + 484, + 190 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 108, + 159, + 484, + 190 + ], + "type": "text", + "content": ", Chenhan Jiang" + }, + { + "bbox": [ + 108, + 159, + 484, + 190 + ], + "type": "inline_equation", + "content": "^{2*}" + }, + { + "bbox": [ + 108, + 159, + 484, + 190 + ], + "type": "text", + "content": ", Jiageng Mao" + }, + { + "bbox": [ + 108, + 159, + 484, + 190 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 108, + 159, + 484, + 190 + ], + "type": "text", + "content": ", Jianhua Han" + }, + { + "bbox": [ + 108, + 159, + 484, + 190 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 108, + 159, + 484, + 190 + ], + "type": "text", + "content": ", Chaoqiang Ye" + }, + { + "bbox": [ + 108, + 159, + 484, + 190 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 108, + 159, + 484, + 190 + ], + "type": "text", + "content": ", Qingqiu Huang" + }, + { + "bbox": [ + 108, + 159, + 484, + 190 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 108, + 159, + 484, + 190 + ], + "type": "text", + "content": ", Dit-Yan Yeung" + }, + { + "bbox": [ + 108, + 159, + 484, + 190 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 108, + 159, + 484, + 190 + ], + "type": "text", + "content": ", Zhen Yang" + }, + { + "bbox": [ + 108, + 159, + 484, + 190 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 108, + 159, + 484, + 190 + ], + "type": "text", + "content": ", Xiaodan Liang" + }, + { + "bbox": [ + 108, + 159, + 484, + 190 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 108, + 159, + 484, + 190 + ], + "type": "text", + "content": ", Hang " + }, + { + "bbox": [ + 108, + 159, + 484, + 190 + ], + "type": "inline_equation", + "content": "\\mathrm{Xu}^{1\\dagger}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 194, + 214, + 531, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 194, + 214, + 531, + 224 + ], + "spans": [ + { + "bbox": [ + 194, + 214, + 531, + 224 + ], + "type": "text", + "content": "Car, Pedestrian, Bicycle, Bus, Debris, Trailer, Table, Chair, Light, Cleaner, Closet, Backpack, Tire, Couch ..." + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 50, + 230, + 299, + 369 + ], + "blocks": [ + { + "bbox": [ + 66, + 214, + 166, + 226 + ], + "lines": [ + { + "bbox": [ + 66, + 214, + 166, + 226 + ], + "spans": [ + { + "bbox": [ + 66, + 214, + 166, + 226 + ], + "type": "text", + "content": "Open-world Vocabularies" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 50, + 230, + 299, + 369 + ], + "lines": [ + { + "bbox": [ + 50, + 230, + 299, + 369 + ], + "spans": [ + { + "bbox": [ + 50, + 230, + 299, + 369 + ], + "type": "image", + "image_path": "cfe8c4c6d10f3a5dd955c4f4c255cbfd42736320d93dda788fc81c0f5a6ecab8.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 371, + 544, + 393 + ], + "lines": [ + { + "bbox": [ + 46, + 371, + 544, + 393 + ], + "spans": [ + { + "bbox": [ + 46, + 371, + 544, + 393 + ], + "type": "text", + "content": "Figure 1. Illustration of our open-world recognition results. Benefiting from our " + }, + { + "bbox": [ + 46, + 371, + 544, + 393 + ], + "type": "inline_equation", + "content": "\\mathrm{CLIP}^2" + }, + { + "bbox": [ + 46, + 371, + 544, + 393 + ], + "type": "text", + "content": ", the 3D representation is aligned to the openworld language representation, which enables flexible zero-shot transfer. Best viewed in colors." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 301, + 228, + 543, + 369 + ], + "blocks": [ + { + "bbox": [ + 301, + 228, + 543, + 369 + ], + "lines": [ + { + "bbox": [ + 301, + 228, + 543, + 369 + ], + "spans": [ + { + "bbox": [ + 301, + 228, + 543, + 369 + ], + "type": "image", + "image_path": "e34e30f102ed2def75dd13d5cca9f46c554c1ec5db0ff05b03c664eb1a47f446.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 143, + 416, + 191, + 429 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 416, + 191, + 429 + ], + "spans": [ + { + "bbox": [ + 143, + 416, + 191, + 429 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 45, + 443, + 287, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 443, + 287, + 664 + ], + "spans": [ + { + "bbox": [ + 45, + 443, + 287, + 664 + ], + "type": "text", + "content": "Contrastive Language-Image Pre-training, benefiting from large-scale unlabeled text-image pairs, has demonstrated great performance in open-world vision understanding tasks. However, due to the limited Text-3D data pairs, adapting the success of 2D Vision-Language Models (VLM) to the 3D space remains an open problem. Existing works that leverage VLM for 3D understanding generally resort to constructing intermediate 2D representations for the 3D data, but at the cost of losing 3D geometry information. To take a step toward open-world 3D vision understanding, we propose Contrastive Language-Image-Point Cloud Pretraining " + }, + { + "bbox": [ + 45, + 443, + 287, + 664 + ], + "type": "inline_equation", + "content": "(CLIP^2)" + }, + { + "bbox": [ + 45, + 443, + 287, + 664 + ], + "type": "text", + "content": " to directly learn the transferable 3D point cloud representation in realistic scenarios with a novel proxy alignment mechanism. Specifically, we exploit naturally-existed correspondences in 2D and 3D scenarios, and build well-aligned and instance-based text-image-point proxies from those complex scenarios. On top of that, we propose a cross-modal contrastive objective to learn semantic and instance-level aligned point cloud representa" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 407, + 545, + 487 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 407, + 545, + 487 + ], + "spans": [ + { + "bbox": [ + 306, + 407, + 545, + 487 + ], + "type": "text", + "content": "tion. Experimental results on both indoor and outdoor scenarios show that our learned 3D representation has great transfer ability in downstream tasks, including zero-shot and few-shot 3D recognition, which boosts the state-of-the-art methods by large margins. Furthermore, we provide analyses of the capability of different representations in real scenarios and present the optional ensemble scheme." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 517, + 386, + 529 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 517, + 386, + 529 + ], + "spans": [ + { + "bbox": [ + 306, + 517, + 386, + 529 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 539, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 539, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 539, + 545, + 713 + ], + "type": "text", + "content": "Powerful 3D point cloud representation plays a crucial role in various real-world applications, e.g., 3D object recognition and detection [10, 20, 31, 40, 44]. Compared to 2D images, 3D point cloud provides specific information like accurate geometry that is robust to illumination changes. However, current methods [25, 40] that learn 3D representations generally rely on the predefined number of object categories and require plenty of labor-intensive annotations. Those learned 3D representations are insufficient for safety-critical scenarios like self-driving which includes a long-tail class distribution far beyond the predefined taxonomy. Therefore, it is highly demanded to learn a transferable 3D representation equipped with zero-shot recognition ability in vocabulary scalable real-world scenes. Figure 1 shows an open-world recognition example by our " + }, + { + "bbox": [ + 304, + 539, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{CLIP}^2" + }, + { + "bbox": [ + 304, + 539, + 545, + 713 + ], + "type": "text", + "content": " in" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 674, + 286, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 674, + 286, + 712 + ], + "spans": [ + { + "bbox": [ + 46, + 674, + 286, + 712 + ], + "type": "text", + "content": "*Equal contribution. 1Huawei Noah's Ark Lab 2Hong Kong University of Science and Technology 3The Chinese University of Hong Kong 4Sun Yat-san University †Corresponding Author: xu.hang@huawei.com" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15244" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 106 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 106 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 106 + ], + "type": "text", + "content": "outdoor and indoor scenes, where the 3D objects can be classified with the correlation alignment between 3D representations and open-world vocabularies." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 108, + 288, + 363 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 108, + 288, + 363 + ], + "spans": [ + { + "bbox": [ + 46, + 108, + 288, + 363 + ], + "type": "text", + "content": "The critical ingredient of open-world understanding is that the models learn sufficient knowledge to obtain general representations. To achieve this, recent Vision-Language Models (VLM) [14, 27, 38] leverage Internet-scale text-image pairs to conduct vision-language pretraining, which facilitates transferable 2D representation and demonstrates promising performance in 2D open-vocabulary tasks. However, 3D vision-language pretraining remains unexplored due to the limitation of existing 3D datasets in diversity and scale compared to the massive data sources in 2D counterparts [14, 15, 27, 38]. Though some recent works [12, 13, 43] try to avoid this problem by transferring the pretrained 2D VLM into the intermediate representation including projected image patches [12, 18] or depth maps [1, 43], those representations suffer from the loss of 3D geometric information and limited viewpoints under realistic scenarios. Especially the camera images are only sometimes available due to the sensor failure in 3D scenes. We believe the 3D representation based on original point cloud data retains most information and is the optimal solution for 3D real world understanding, which requires a rethink of learning the transferable 3D representation under realistic scenarios." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 364, + 287, + 654 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 364, + 287, + 654 + ], + "spans": [ + { + "bbox": [ + 46, + 364, + 287, + 654 + ], + "type": "text", + "content": "To this end, we propose a Contrastive Language-Image-Point cloud Pretraining framework, short for " + }, + { + "bbox": [ + 46, + 364, + 287, + 654 + ], + "type": "inline_equation", + "content": "\\mathrm{CLIP}^2" + }, + { + "bbox": [ + 46, + 364, + 287, + 654 + ], + "type": "text", + "content": ", which directly aligns 3D space with broader raw text and advances the 3D representation learning into an open-world era. Our learning process can be decomposed into two stages: Firstly, we introduce a Triplet Proxy Collection to alleviate the limitation of accessible pretraining data by constructing language-image-point triplets from real-world scenes. Since the large-scale realistic 3D datasets for outdoor driving [2,19] and indoor scenarios [9,32] are collected in open-world, it contains huge amounts of realistic objects that vary in semantics and diversity. Thus we consider them as potential pretraining data sources without extra human supervision. Specifically, we propose \"Proxy\" instances as the bridges between language descriptions, 2D images and 3D point clouds. Enabled by a well-aligned VLM, a scalable caption list and the geometry transformation between 2D and 3D, we automatically create more than 1 million triplets to facilitate pretraining. Secondly, we further propose a Cross-Modal Pretraining scheme to jointly optimize the feature space alignments of three modalities, i.e., point cloud, language and image. It contains both the contrastive learning objective of semantic-level text-3D correlation and instance-level image-3D correlation, which contributes to better transferability of learned 3D representation." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 654, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 654, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 654, + 287, + 714 + ], + "type": "text", + "content": "We study the transferable capability of " + }, + { + "bbox": [ + 46, + 654, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\mathrm{CLIP}^2" + }, + { + "bbox": [ + 46, + 654, + 287, + 714 + ], + "type": "text", + "content": " by benchmarking the zero-shot recognition performance on four popular indoor and outdoor real-world datasets, and find a significant improvement over current methods, achieving Top1 accuracy " + }, + { + "bbox": [ + 46, + 654, + 287, + 714 + ], + "type": "inline_equation", + "content": "61.3\\%" + }, + { + "bbox": [ + 46, + 654, + 287, + 714 + ], + "type": "text", + "content": " on SunRGBD [32], " + }, + { + "bbox": [ + 46, + 654, + 287, + 714 + ], + "type": "inline_equation", + "content": "43.8\\%" + }, + { + "bbox": [ + 46, + 654, + 287, + 714 + ], + "type": "text", + "content": " on ScanNet [9])," + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 304, + 72, + 545, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 224 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 224 + ], + "type": "text", + "content": "28.8% on nuScenes [2] and 56.0% on ONCE [19]. For a fair comparison with existing methods [1, 13, 36, 43], we conduct zero-shot and few-shot classification on single object dataset ScanObjectNN [34] and find consistent dominance, 16.1% relative improvement on zero-shot classification over previous state-of-the-art method [13]. To validate the vocabulary-increasing ability of " + }, + { + "bbox": [ + 304, + 72, + 545, + 224 + ], + "type": "inline_equation", + "content": "\\mathrm{CLIP}^2" + }, + { + "bbox": [ + 304, + 72, + 545, + 224 + ], + "type": "text", + "content": ", we report the quantity results and visualizations to show the improved discovery of the long-tail categories. Moreover, we make ablations and analysis on different representations, and investigate ensembling alternatives to merge complementary knowledge of all available representations in realistic applications. Our contributions can be summarized as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 317, + 232, + 545, + 418 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 317, + 232, + 545, + 267 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 232, + 545, + 267 + ], + "spans": [ + { + "bbox": [ + 317, + 232, + 545, + 267 + ], + "type": "text", + "content": "- We propose a novel " + }, + { + "bbox": [ + 317, + 232, + 545, + 267 + ], + "type": "inline_equation", + "content": "\\mathrm{CLIP}^2" + }, + { + "bbox": [ + 317, + 232, + 545, + 267 + ], + "type": "text", + "content": " framework that aligns 3D space with open-world language representation, facilitating zero-shot transfer in realistic scenarios." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 317, + 274, + 545, + 310 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 274, + 545, + 310 + ], + "spans": [ + { + "bbox": [ + 317, + 274, + 545, + 310 + ], + "type": "text", + "content": "- We present a Triplet Proxies Collection scheme in real-world scenes, which alleviates the shortage of text-3D data sources and facilitates the pretraining methods." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 317, + 316, + 545, + 364 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 316, + 545, + 364 + ], + "spans": [ + { + "bbox": [ + 317, + 316, + 545, + 364 + ], + "type": "text", + "content": "- " + }, + { + "bbox": [ + 317, + 316, + 545, + 364 + ], + "type": "inline_equation", + "content": "\\mathrm{CLIP}^2" + }, + { + "bbox": [ + 317, + 316, + 545, + 364 + ], + "type": "text", + "content": " jointly optimizes the correlation alignment between point cloud, language and image by proposed cross-modal pretraining mechanism, which enhances the transferability of learned 3D representation." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 317, + 371, + 545, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 371, + 545, + 418 + ], + "spans": [ + { + "bbox": [ + 317, + 371, + 545, + 418 + ], + "type": "text", + "content": "- Our " + }, + { + "bbox": [ + 317, + 371, + 545, + 418 + ], + "type": "inline_equation", + "content": "\\mathsf{CLIP}^2" + }, + { + "bbox": [ + 317, + 371, + 545, + 418 + ], + "type": "text", + "content": " achieves the state-of-the-art zero-shot transfer performance on 5 datasets (indoor/outdoor scenes and single-object) and shows quality results on vocabulary-increasing discovery in real world." + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 428, + 392, + 440 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 428, + 392, + 440 + ], + "spans": [ + { + "bbox": [ + 306, + 428, + 392, + 440 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 443, + 545, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 443, + 545, + 629 + ], + "spans": [ + { + "bbox": [ + 304, + 443, + 545, + 629 + ], + "type": "text", + "content": "Vision-Language Model. Large vision language models (VLM) [14, 15, 27, 38] have demonstrated successful performance in downstream zero-shot tasks with the learned transferable 2D representations. CLIP [27] and ALIGN [14] push the limit by collecting Internet-scale image-text pairs and then learning the correlation alignment between image and language feature space with contrastive pretraining objectives. Those models can be directly transferred to zero-shot 2D recognition and achieve impressive results. Recent DetClip [38] learns to align image patches to test phrases after pretraining under hybrid supervision from detection, grounding and image-text pair data, which extends the ability to localize open-vocabulary 2D proposals in images. In this paper, we attempt to transfer the open-vocabulary ability of pre-trained VLM to the 3D domain, making language applicable to zero-shot point cloud recognition." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 632, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 632, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 632, + 545, + 713 + ], + "type": "text", + "content": "Zero-shot/Open-world Learning in 3D. Recognizing 3D objects with a large vocabulary is necessary for safety-critical autonomous driving and robotic tasks, yet remains under-explored. Cheraghian et al. [5-8] first attempt to associate PointNet [25] feature with category semantic information via a projection function, and separately proposed an unsupervised skewness loss [5] to mitigate the hubness" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15245" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 53, + 70, + 545, + 286 + ], + "blocks": [ + { + "bbox": [ + 53, + 70, + 545, + 286 + ], + "lines": [ + { + "bbox": [ + 53, + 70, + 545, + 286 + ], + "spans": [ + { + "bbox": [ + 53, + 70, + 545, + 286 + ], + "type": "image", + "image_path": "c9d4e4112331aca8c40b6ed1534e6c7f1bd3717b4519fc73bd58938f66cf5179.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 293, + 548, + 350 + ], + "lines": [ + { + "bbox": [ + 46, + 293, + 548, + 350 + ], + "spans": [ + { + "bbox": [ + 46, + 293, + 548, + 350 + ], + "type": "text", + "content": "Figure 2. Overview of " + }, + { + "bbox": [ + 46, + 293, + 548, + 350 + ], + "type": "inline_equation", + "content": "\\mathbf{CLIP}^2" + }, + { + "bbox": [ + 46, + 293, + 548, + 350 + ], + "type": "text", + "content": " framework. The main components contain two parts, the Triplet Proxy Collection and the Cross-Modal Pretraining. The defined Triplet Proxy set " + }, + { + "bbox": [ + 46, + 293, + 548, + 350 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{\\mathrm{proxy}}" + }, + { + "bbox": [ + 46, + 293, + 548, + 350 + ], + "type": "text", + "content": " consists of language captions " + }, + { + "bbox": [ + 46, + 293, + 548, + 350 + ], + "type": "inline_equation", + "content": "\\mathbf{X}^T" + }, + { + "bbox": [ + 46, + 293, + 548, + 350 + ], + "type": "text", + "content": ", corresponding image instances " + }, + { + "bbox": [ + 46, + 293, + 548, + 350 + ], + "type": "inline_equation", + "content": "\\mathbf{X}^I" + }, + { + "bbox": [ + 46, + 293, + 548, + 350 + ], + "type": "text", + "content": " and raw 3D point cloud instances " + }, + { + "bbox": [ + 46, + 293, + 548, + 350 + ], + "type": "inline_equation", + "content": "\\mathbf{X}^P" + }, + { + "bbox": [ + 46, + 293, + 548, + 350 + ], + "type": "text", + "content": ", which come from the free data source under realistic scenarios without any labeling labor. On top of that, we pretrain a point cloud encoder " + }, + { + "bbox": [ + 46, + 293, + 548, + 350 + ], + "type": "inline_equation", + "content": "E^P" + }, + { + "bbox": [ + 46, + 293, + 548, + 350 + ], + "type": "text", + "content": " with the cross-modal contrastive learning objective. Equipped with " + }, + { + "bbox": [ + 46, + 293, + 548, + 350 + ], + "type": "inline_equation", + "content": "\\mathbf{CLIP}^2" + }, + { + "bbox": [ + 46, + 293, + 548, + 350 + ], + "type": "text", + "content": ", the learned 3D point cloud representation " + }, + { + "bbox": [ + 46, + 293, + 548, + 350 + ], + "type": "inline_equation", + "content": "F^P" + }, + { + "bbox": [ + 46, + 293, + 548, + 350 + ], + "type": "text", + "content": " is well aligned to the language representation, which facilitates downstream zero-shot 3D transfer tasks in the real world." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 369, + 289, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 369, + 289, + 613 + ], + "spans": [ + { + "bbox": [ + 46, + 369, + 289, + 613 + ], + "type": "text", + "content": "problem. The transductive case [6] is discussed in which extends [5] using a triplet loss. Notably, the above works conduct experiments on synthetic datasets and need to divide datasets into \"seen\" categories as training data and \"unseen\" categories as testing data. Thus they are not suitable for realistic scenarios due to the domain gap between synthetic and real-world data, as well as the limited vocabulary-increasing ability. Recently, inspired by the success of VLMs [14, 27] in 2D tasks, some works [13, 43] propose to transfer the zero-shot recognition ability of pretrained CLIP [27] into 3D area. PointCLIP [43] directly projects point cloud into multi-view depth maps as image-like data input for pretrained CLIP to make classification predictions. While CLIP2Point [13] trains an image-depth embedding on ShapeNet [39] to better align the depth representation to the pretrained image space of CLIP. However, depth maps lost plenty of geometry information of the original point cloud data structure, resulting in poor performance especially in realistic scenarios. By contrast, we aim to learn transferable 3D representation based on the original point cloud data structure in realistic scenarios." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 620, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 620, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 620, + 289, + 715 + ], + "type": "text", + "content": "3D Representation Learning. Much progress has been made in learning a comprehensive 3D representation in an unsupervised manner. Most works [1, 16, 17, 23, 30, 37, 41, 42] follow the paradigm that conducts pretraining on unlabeled datasets and then finetunes on the limited downstream annotations. Though the improved transferability of 3D representation, they can not be directly transferred to zero-shot tasks with open-world vocabularies. In this work, we con" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 304, + 369, + 547, + 406 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 369, + 547, + 406 + ], + "spans": [ + { + "bbox": [ + 304, + 369, + 547, + 406 + ], + "type": "text", + "content": "duct language-image-point cloud pretraining, which learns transferable 3D representation aligned to open-vocabulary language space to facilitate the zero-shot transfer." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 305, + 413, + 362, + 426 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 413, + 362, + 426 + ], + "spans": [ + { + "bbox": [ + 305, + 413, + 362, + 426 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 433, + 547, + 551 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 433, + 547, + 551 + ], + "spans": [ + { + "bbox": [ + 304, + 433, + 547, + 551 + ], + "type": "text", + "content": "In this section, we introduce " + }, + { + "bbox": [ + 304, + 433, + 547, + 551 + ], + "type": "inline_equation", + "content": "\\mathrm{CLIP}^2" + }, + { + "bbox": [ + 304, + 433, + 547, + 551 + ], + "type": "text", + "content": " to learn a transferable 3D point cloud representation with arbitrary category recognition ability under realistic scenarios, illustrated in Figure 2. We will first present the Triplet Proxy Collection in Section 3.1, which utilizes a pretrained VLM and geometric transformation to obtain language-image-point triplets from real-world scenes. Then we will elaborate Cross-Modal Contrastive Pretraining mechanism in Section 3.2, which jointly optimizes the alignment correlations between language, image and point cloud feature space." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 555, + 443, + 569 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 555, + 443, + 569 + ], + "spans": [ + { + "bbox": [ + 305, + 555, + 443, + 569 + ], + "type": "text", + "content": "3.1. Triplet Proxy Collection" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 574, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 574, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 574, + 547, + 715 + ], + "type": "text", + "content": "Inspired by the significant performance of 2D VLMs on open-vocabulary tasks, we aim to develop 3D vision-language pretraining to facilitate category-increasing capacity for real-world scenarios. However, the core challenge is the shortage of pretraining data. Compared to the 2D vision-language pretraining framework CLIP [27], which takes more than 400M image-language pairs from the Internet, the largest 3D single-object dataset ShapeNet [39] only contains 50K CAD models with 55 categories. In addition to the insufficiency of data scale, pretraining on such synthetic data fails to transfer well in the real world due to the huge domain gap. Enlightened by the recent emer" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "15246" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 48, + 70, + 286, + 155 + ], + "blocks": [ + { + "bbox": [ + 48, + 70, + 286, + 155 + ], + "lines": [ + { + "bbox": [ + 48, + 70, + 286, + 155 + ], + "spans": [ + { + "bbox": [ + 48, + 70, + 286, + 155 + ], + "type": "image", + "image_path": "a8da405e6ae9f064a476ddb3c2101091ec79bc7b6889540c99430eeeaf5a9e32.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 47, + 161, + 287, + 184 + ], + "lines": [ + { + "bbox": [ + 47, + 161, + 287, + 184 + ], + "spans": [ + { + "bbox": [ + 47, + 161, + 287, + 184 + ], + "type": "text", + "content": "Figure 3. Illustration of three representation modals of two 3D objects examples under indoor and outdoor scenarios." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 205, + 287, + 298 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 205, + 287, + 298 + ], + "spans": [ + { + "bbox": [ + 46, + 205, + 287, + 298 + ], + "type": "text", + "content": "gence of large-scale point cloud datasets collected in indoor [9, 32] and outdoor scenarios [2, 19], we observe that those naturally-collected datasets potentially contain vast amounts of open-world objects that vary in semantics and diversity. Considering the data collection itself is cheap except for laborious annotation, we novelly take leverage of those available datasets without human annotations as a practical yet effective pretraining data source." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 298, + 287, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 298, + 287, + 403 + ], + "spans": [ + { + "bbox": [ + 47, + 298, + 287, + 403 + ], + "type": "text", + "content": "Specifically, given the realistic scene data " + }, + { + "bbox": [ + 47, + 298, + 287, + 403 + ], + "type": "inline_equation", + "content": "S = \\{(P_s, I_s)_{s=1}^{|S|}\\}" + }, + { + "bbox": [ + 47, + 298, + 287, + 403 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 47, + 298, + 287, + 403 + ], + "type": "inline_equation", + "content": "P_s \\in \\mathbb{R}^{N_P \\times 3}" + }, + { + "bbox": [ + 47, + 298, + 287, + 403 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 298, + 287, + 403 + ], + "type": "inline_equation", + "content": "I_s \\in \\mathbb{R}^{N_I \\times H \\times W \\times 3}" + }, + { + "bbox": [ + 47, + 298, + 287, + 403 + ], + "type": "text", + "content": " are corresponding 3D point clouds and images of scene " + }, + { + "bbox": [ + 47, + 298, + 287, + 403 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 47, + 298, + 287, + 403 + ], + "type": "text", + "content": ", we propose a novel concept, Proxy, as the bridge between language, image and 3D point cloud. As illustrated in Figure 2, equipped by those proxy instances, we can automatically collect a massive number of language-image-point cloud pairs " + }, + { + "bbox": [ + 47, + 298, + 287, + 403 + ], + "type": "inline_equation", + "content": "D_{\\mathrm{proxy}}" + }, + { + "bbox": [ + 47, + 298, + 287, + 403 + ], + "type": "text", + "content": " in the format of proxies under open-world scenes. We detail the process as follows." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 403, + 287, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 403, + 287, + 521 + ], + "spans": [ + { + "bbox": [ + 47, + 403, + 287, + 521 + ], + "type": "text", + "content": "Language Proxy. We set the language proxies " + }, + { + "bbox": [ + 47, + 403, + 287, + 521 + ], + "type": "inline_equation", + "content": "\\mathbf{X}^T\\in \\mathbb{R}^V" + }, + { + "bbox": [ + 47, + 403, + 287, + 521 + ], + "type": "text", + "content": " as a raw text list from the 2D open-world dataset [11], where " + }, + { + "bbox": [ + 47, + 403, + 287, + 521 + ], + "type": "inline_equation", + "content": "V = 1206" + }, + { + "bbox": [ + 47, + 403, + 287, + 521 + ], + "type": "text", + "content": " denotes the vocabulary size of language proxies. Image Proxy. Next, we obtain the image proxies " + }, + { + "bbox": [ + 47, + 403, + 287, + 521 + ], + "type": "inline_equation", + "content": "\\mathbf{X}^I" + }, + { + "bbox": [ + 47, + 403, + 287, + 521 + ], + "type": "text", + "content": " by an open vocabulary detector DetCLIP [38], denoted as " + }, + { + "bbox": [ + 47, + 403, + 287, + 521 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 47, + 403, + 287, + 521 + ], + "type": "text", + "content": " which is trained with open-world data and performs openset detection. Concretely, given language proxies " + }, + { + "bbox": [ + 47, + 403, + 287, + 521 + ], + "type": "inline_equation", + "content": "\\mathbf{X}^T" + }, + { + "bbox": [ + 47, + 403, + 287, + 521 + ], + "type": "text", + "content": " and input scene image " + }, + { + "bbox": [ + 47, + 403, + 287, + 521 + ], + "type": "inline_equation", + "content": "I_{s}" + }, + { + "bbox": [ + 47, + 403, + 287, + 521 + ], + "type": "text", + "content": " , we extract corresponding image proposals as image proxies " + }, + { + "bbox": [ + 47, + 403, + 287, + 521 + ], + "type": "inline_equation", + "content": "X_{s}^{I}" + }, + { + "bbox": [ + 47, + 403, + 287, + 521 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 47, + 403, + 287, + 521 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 47, + 403, + 287, + 521 + ], + "type": "text", + "content": " by the similarity between input language embeddings and proposal features as" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 99, + 531, + 287, + 547 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 99, + 531, + 287, + 547 + ], + "spans": [ + { + "bbox": [ + 99, + 531, + 287, + 547 + ], + "type": "interline_equation", + "content": "\\left\\{X _ {s} ^ {I} \\right\\} _ {s \\in | \\mathcal {S} |} = \\mathrm {M} \\left(\\left\\{I _ {s} \\right\\} _ {s \\in | \\mathcal {S} |}, \\mathbf {X} ^ {T}\\right). \\tag {1}", + "image_path": "67befc026b69df5695743f1e0a357c7f12877419f71701865159431a046bcc17.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 552, + 287, + 611 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 552, + 287, + 611 + ], + "spans": [ + { + "bbox": [ + 47, + 552, + 287, + 611 + ], + "type": "text", + "content": "3D Proxy. We exploit the naturally-existed geometry relations between 2D and 3D scenes to obtain 3D proxies " + }, + { + "bbox": [ + 47, + 552, + 287, + 611 + ], + "type": "inline_equation", + "content": "\\mathbf{X}^P" + }, + { + "bbox": [ + 47, + 552, + 287, + 611 + ], + "type": "text", + "content": ", which consists of point cloud instances corresponding to image proposals in " + }, + { + "bbox": [ + 47, + 552, + 287, + 611 + ], + "type": "inline_equation", + "content": "\\mathbf{X}^I" + }, + { + "bbox": [ + 47, + 552, + 287, + 611 + ], + "type": "text", + "content": ". We simplify the geometry transformation as " + }, + { + "bbox": [ + 47, + 552, + 287, + 611 + ], + "type": "inline_equation", + "content": "\\mathrm{G}(\\cdot)" + }, + { + "bbox": [ + 47, + 552, + 287, + 611 + ], + "type": "text", + "content": " and formulate the relations as:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 135, + 618, + 287, + 633 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 618, + 287, + 633 + ], + "spans": [ + { + "bbox": [ + 135, + 618, + 287, + 633 + ], + "type": "interline_equation", + "content": "X _ {i} ^ {P} = \\mathrm {G} \\left(X _ {i} ^ {I}\\right). \\tag {2}", + "image_path": "c605a1b2cdb33b788349baebfab6d7c3dae978478e02a0ff84bcc032ce17a862.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 642, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 642, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 47, + 642, + 287, + 715 + ], + "type": "text", + "content": "Detailedly, for indoor scenes equipped with RGB-D sensors, we first remove the background pixels by unsupervised segmentation algorithm [28] for each image proxy " + }, + { + "bbox": [ + 47, + 642, + 287, + 715 + ], + "type": "inline_equation", + "content": "X_{s,i}^{I}" + }, + { + "bbox": [ + 47, + 642, + 287, + 715 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 47, + 642, + 287, + 715 + ], + "type": "inline_equation", + "content": "i \\in |X_s^I|" + }, + { + "bbox": [ + 47, + 642, + 287, + 715 + ], + "type": "text", + "content": ". Since depth information is known, we then transform the segmented pixels from uvd coordinate " + }, + { + "bbox": [ + 47, + 642, + 287, + 715 + ], + "type": "inline_equation", + "content": "X_{s,i}^{I,uvd} \\in \\mathbb{R}^{n,3}" + }, + { + "bbox": [ + 47, + 642, + 287, + 715 + ], + "type": "text", + "content": " to xyz coordinate " + }, + { + "bbox": [ + 47, + 642, + 287, + 715 + ], + "type": "inline_equation", + "content": "X_{s,i}^{P,xyz} \\in \\mathbb{R}^{n,3}" + }, + { + "bbox": [ + 47, + 642, + 287, + 715 + ], + "type": "text", + "content": " as a 3D" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 310, + 69, + 541, + 154 + ], + "blocks": [ + { + "bbox": [ + 310, + 69, + 541, + 154 + ], + "lines": [ + { + "bbox": [ + 310, + 69, + 541, + 154 + ], + "spans": [ + { + "bbox": [ + 310, + 69, + 541, + 154 + ], + "type": "image", + "image_path": "c658027c3b01323f7c826a78da3e707d1381ceadaf174e4edb9bb24c7406eded.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 161, + 545, + 237 + ], + "lines": [ + { + "bbox": [ + 304, + 161, + 545, + 237 + ], + "spans": [ + { + "bbox": [ + 304, + 161, + 545, + 237 + ], + "type": "text", + "content": "Figure 4. Comparison of different pretraining strategies. (a) CLIP aligns image and language embedding space [27] as " + }, + { + "bbox": [ + 304, + 161, + 545, + 237 + ], + "type": "inline_equation", + "content": "L_{TI}" + }, + { + "bbox": [ + 304, + 161, + 545, + 237 + ], + "type": "text", + "content": " based on large-scale text-image pairs. (b) PointClip [27] aligns projected depth map to CLIP language space as " + }, + { + "bbox": [ + 304, + 161, + 545, + 237 + ], + "type": "inline_equation", + "content": "L_{TD}" + }, + { + "bbox": [ + 304, + 161, + 545, + 237 + ], + "type": "text", + "content": ". (c) Clip2Point aligns depth map to CLIP image space as " + }, + { + "bbox": [ + 304, + 161, + 545, + 237 + ], + "type": "inline_equation", + "content": "L_{ID}" + }, + { + "bbox": [ + 304, + 161, + 545, + 237 + ], + "type": "text", + "content": ". (d) our " + }, + { + "bbox": [ + 304, + 161, + 545, + 237 + ], + "type": "inline_equation", + "content": "\\mathrm{CLIP}^2" + }, + { + "bbox": [ + 304, + 161, + 545, + 237 + ], + "type": "text", + "content": " aligns original 3D point cloud to both CLIP language space and image space via cross-modal objective " + }, + { + "bbox": [ + 304, + 161, + 545, + 237 + ], + "type": "inline_equation", + "content": "L_{CM}" + }, + { + "bbox": [ + 304, + 161, + 545, + 237 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 258, + 545, + 330 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 258, + 545, + 330 + ], + "spans": [ + { + "bbox": [ + 304, + 258, + 545, + 330 + ], + "type": "text", + "content": "point cloud proxy with the given camera parameters. For outdoor scenes captured by LiDAR sensors, we first create a 3D frustum for each image proxy by extruding the 2D image proposal into 3D space following [22, 24]. Then we conduct DBSCAN algorithm [29] within frustum and select the point cloud cluster as the point proxy " + }, + { + "bbox": [ + 304, + 258, + 545, + 330 + ], + "type": "inline_equation", + "content": "X_{s,i}^{P,xyz}" + }, + { + "bbox": [ + 304, + 258, + 545, + 330 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 331, + 545, + 406 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 331, + 545, + 406 + ], + "spans": [ + { + "bbox": [ + 305, + 331, + 545, + 406 + ], + "type": "text", + "content": "Eventually, we construct Triplet Proxy " + }, + { + "bbox": [ + 305, + 331, + 545, + 406 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{\\mathrm{proxy}} = \\{\\mathbf{X}^T,X_s^I,X_s^P\\}_{s = 1}^{|\\mathcal{S}|}" + }, + { + "bbox": [ + 305, + 331, + 545, + 406 + ], + "type": "text", + "content": " by combining corresponding language proxies " + }, + { + "bbox": [ + 305, + 331, + 545, + 406 + ], + "type": "inline_equation", + "content": "\\mathbf{X}^T" + }, + { + "bbox": [ + 305, + 331, + 545, + 406 + ], + "type": "text", + "content": " , image proxies " + }, + { + "bbox": [ + 305, + 331, + 545, + 406 + ], + "type": "inline_equation", + "content": "\\mathbf{X}^I" + }, + { + "bbox": [ + 305, + 331, + 545, + 406 + ], + "type": "text", + "content": " and 3D proxies " + }, + { + "bbox": [ + 305, + 331, + 545, + 406 + ], + "type": "inline_equation", + "content": "\\mathbf{X}^P" + }, + { + "bbox": [ + 305, + 331, + 545, + 406 + ], + "type": "text", + "content": " , where " + }, + { + "bbox": [ + 305, + 331, + 545, + 406 + ], + "type": "inline_equation", + "content": "\\mathbf{X}^I = \\{X_s^I\\}_{s = 1}^{|\\mathcal{S}|}" + }, + { + "bbox": [ + 305, + 331, + 545, + 406 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 331, + 545, + 406 + ], + "type": "inline_equation", + "content": "\\mathbf{X}^P = \\{X_s^P\\}_{s = 1}^{|\\mathcal{S}|}" + }, + { + "bbox": [ + 305, + 331, + 545, + 406 + ], + "type": "text", + "content": " . " + }, + { + "bbox": [ + 305, + 331, + 545, + 406 + ], + "type": "inline_equation", + "content": "220\\mathrm{K}" + }, + { + "bbox": [ + 305, + 331, + 545, + 406 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 331, + 545, + 406 + ], + "type": "inline_equation", + "content": "1.4\\mathrm{M}" + }, + { + "bbox": [ + 305, + 331, + 545, + 406 + ], + "type": "text", + "content": " proxy triplets are formed for indoor and outdoor scenes, respectively. More details can be found in the appendix." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 415, + 503, + 427 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 415, + 503, + 427 + ], + "spans": [ + { + "bbox": [ + 306, + 415, + 503, + 427 + ], + "type": "text", + "content": "3.2. Cross-Modal Contrastive Pretraining" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 434, + 545, + 550 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 434, + 545, + 550 + ], + "spans": [ + { + "bbox": [ + 304, + 434, + 545, + 550 + ], + "type": "text", + "content": "With the triplet proxies " + }, + { + "bbox": [ + 304, + 434, + 545, + 550 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{\\mathrm{proxy}}" + }, + { + "bbox": [ + 304, + 434, + 545, + 550 + ], + "type": "text", + "content": ", a straightforward pretraining objective is forcing the alignment between the embedding spaces of point cloud " + }, + { + "bbox": [ + 304, + 434, + 545, + 550 + ], + "type": "inline_equation", + "content": "X_{i}^{P}" + }, + { + "bbox": [ + 304, + 434, + 545, + 550 + ], + "type": "text", + "content": " and language " + }, + { + "bbox": [ + 304, + 434, + 545, + 550 + ], + "type": "inline_equation", + "content": "X_{i}^{T}" + }, + { + "bbox": [ + 304, + 434, + 545, + 550 + ], + "type": "text", + "content": " from scratch. However, it might not promise good transferability of learned representation, since the number of language-image-point pretraining data triplets remains two orders of magnitude smaller than the language-image pairs adopted by CLIP [27] and the vocabulary size is much more limited. Therefore, we design to learn the correlation alignment based on the pretrained embedding space of CLIP." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 550, + 545, + 678 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 550, + 545, + 678 + ], + "spans": [ + { + "bbox": [ + 304, + 550, + 545, + 678 + ], + "type": "text", + "content": "The comparison of current pretraining strategies [13,43] is illustrated in Figure 4, which is a series of 3D variants of CLIP. Notably, both existing methods exploit projected depth map as the intermediate representation of point cloud, which are respectively learned to align to language space [43] and image space [13]. Intuitively, as illustrated in Figure 3, depth representation lost plenty of geometry information compared to the original point cloud, especially in outdoor scenarios. Moreover, images are sometimes unavailable for 3D objects. Thus we conduct pretraining on original 3D point cloud data as an optimal representation." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 678, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 678, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 678, + 545, + 714 + ], + "type": "text", + "content": "Toward learning more transferable representation, we introduce a cross-modal contrastive learning objective to jointly optimize the correlation alignment across lan" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15247" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 164 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 164 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 164 + ], + "type": "text", + "content": "guage, image and point cloud, including Semantic-Level Language-3D Alignment and Instance-Level Image-3D Alignment. Specifically, the overall architecture of " + }, + { + "bbox": [ + 46, + 72, + 287, + 164 + ], + "type": "inline_equation", + "content": "\\mathrm{CLIP}^2" + }, + { + "bbox": [ + 46, + 72, + 287, + 164 + ], + "type": "text", + "content": ", shown in Figure 2, contains language encoder " + }, + { + "bbox": [ + 46, + 72, + 287, + 164 + ], + "type": "inline_equation", + "content": "E_{\\theta}^{T}" + }, + { + "bbox": [ + 46, + 72, + 287, + 164 + ], + "type": "text", + "content": ", point cloud encoder " + }, + { + "bbox": [ + 46, + 72, + 287, + 164 + ], + "type": "inline_equation", + "content": "E_{\\theta}^{P}" + }, + { + "bbox": [ + 46, + 72, + 287, + 164 + ], + "type": "text", + "content": " and visual encoder " + }, + { + "bbox": [ + 46, + 72, + 287, + 164 + ], + "type": "inline_equation", + "content": "E_{\\theta}^{I}" + }, + { + "bbox": [ + 46, + 72, + 287, + 164 + ], + "type": "text", + "content": ", which respectively embed the triplet proxies into text feature " + }, + { + "bbox": [ + 46, + 72, + 287, + 164 + ], + "type": "inline_equation", + "content": "f^{T} \\in \\mathcal{R}^{1 \\times C_{T}}" + }, + { + "bbox": [ + 46, + 72, + 287, + 164 + ], + "type": "text", + "content": ", point cloud feature " + }, + { + "bbox": [ + 46, + 72, + 287, + 164 + ], + "type": "inline_equation", + "content": "f^{P} \\in \\mathcal{R}^{1 \\times C_{P}}" + }, + { + "bbox": [ + 46, + 72, + 287, + 164 + ], + "type": "text", + "content": " and image feature " + }, + { + "bbox": [ + 46, + 72, + 287, + 164 + ], + "type": "inline_equation", + "content": "f^{I} \\in \\mathcal{R}^{1 \\times C_{I}}" + }, + { + "bbox": [ + 46, + 72, + 287, + 164 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 72, + 287, + 164 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 46, + 72, + 287, + 164 + ], + "type": "text", + "content": " is the embedding dimension." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 165, + 288, + 281 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 165, + 288, + 281 + ], + "spans": [ + { + "bbox": [ + 46, + 165, + 288, + 281 + ], + "type": "text", + "content": "Semantic-Level Language-3D Alignment. In order to inherit the open-world recognition ability from pretrained CLIP [27], we align the point cloud feature " + }, + { + "bbox": [ + 46, + 165, + 288, + 281 + ], + "type": "inline_equation", + "content": "f^{P}" + }, + { + "bbox": [ + 46, + 165, + 288, + 281 + ], + "type": "text", + "content": " with text embedding " + }, + { + "bbox": [ + 46, + 165, + 288, + 281 + ], + "type": "inline_equation", + "content": "f^{T}" + }, + { + "bbox": [ + 46, + 165, + 288, + 281 + ], + "type": "text", + "content": " from well-trained CLIP with Language-Point Proxy " + }, + { + "bbox": [ + 46, + 165, + 288, + 281 + ], + "type": "inline_equation", + "content": "\\{X_{i}^{T}, X_{i}^{P}\\}" + }, + { + "bbox": [ + 46, + 165, + 288, + 281 + ], + "type": "text", + "content": " input. We replace classname in the prompts, like \"point cloud of a {classname} .\" with raw text in proxy " + }, + { + "bbox": [ + 46, + 165, + 288, + 281 + ], + "type": "inline_equation", + "content": "X_{i}^{T}" + }, + { + "bbox": [ + 46, + 165, + 288, + 281 + ], + "type": "text", + "content": " as language sentences. The core idea is to drive the feature centroids of 3D instances and the corresponding text prompt closer. We compute the loss function of between of language proxy and point cloud proxy as:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 285, + 289, + 327 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 285, + 289, + 327 + ], + "spans": [ + { + "bbox": [ + 47, + 285, + 289, + 327 + ], + "type": "interline_equation", + "content": "l (i, T, P) = - \\log \\frac {\\exp \\left(f _ {i} ^ {T} \\cdot f _ {i} ^ {P} / \\tau\\right)}{\\exp \\left(f _ {i} ^ {T} \\cdot f _ {i} ^ {P} / \\tau\\right) + \\sum_ {j \\in N , X _ {j} ^ {T} \\neq X _ {i} ^ {T}} \\exp \\left(f _ {i} ^ {T} \\cdot f _ {j} ^ {P} / \\tau\\right)}, \\tag {3}", + "image_path": "287ed77131d0ac364662f1a4ee2c570763759c6c9ad149aad14cebfb95d1f049.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 328, + 287, + 364 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 328, + 287, + 364 + ], + "spans": [ + { + "bbox": [ + 46, + 328, + 287, + 364 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 328, + 287, + 364 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 328, + 287, + 364 + ], + "type": "text", + "content": " is the mini-batch size, " + }, + { + "bbox": [ + 46, + 328, + 287, + 364 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 46, + 328, + 287, + 364 + ], + "type": "text", + "content": " is the temperature coefficient. Within a training mini-batch, the language-3D alignment objective " + }, + { + "bbox": [ + 46, + 328, + 287, + 364 + ], + "type": "inline_equation", + "content": "L(T,P)" + }, + { + "bbox": [ + 46, + 328, + 287, + 364 + ], + "type": "text", + "content": " can be described as:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 111, + 370, + 287, + 396 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 111, + 370, + 287, + 396 + ], + "spans": [ + { + "bbox": [ + 111, + 370, + 287, + 396 + ], + "type": "interline_equation", + "content": "L (T, P) = \\frac {1}{N} \\sum_ {i \\in N} l (i, T, P). \\tag {4}", + "image_path": "8cf3762928b92600a0da20364e3ada81efdefbfdf0b820451bfe1a98adde545a.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 402, + 287, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 402, + 287, + 518 + ], + "spans": [ + { + "bbox": [ + 46, + 402, + 287, + 518 + ], + "type": "text", + "content": "Instance-Level Image-3D Alignment. In addition to the alignment between semantic language and 3D proxy instances, we further introduce the contrastive alignment between instance-wise image proxy and 3D proxy instances. Note that the instance-aware visual concept has been well-studied in the embedding space of pretrained CLIP. We believe instance-sensitive learning can contribute to further correlation learning and benefits to the transferability of learned 3D representation. The contrastive aligned objective " + }, + { + "bbox": [ + 46, + 402, + 287, + 518 + ], + "type": "inline_equation", + "content": "L(I,P)" + }, + { + "bbox": [ + 46, + 402, + 287, + 518 + ], + "type": "text", + "content": " across point cloud and image is formulated as:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 534, + 286, + 577 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 534, + 286, + 577 + ], + "spans": [ + { + "bbox": [ + 48, + 534, + 286, + 577 + ], + "type": "interline_equation", + "content": "l (i, I, P) = - \\log \\frac {\\exp \\left(f _ {i} ^ {I} \\cdot f _ {i} ^ {P} / \\tau\\right)}{\\exp \\left(f _ {i} ^ {I} \\cdot f _ {i} ^ {P} / \\tau\\right) + \\sum_ {j \\in N , j \\neq i} \\exp \\left(f _ {i} ^ {J} \\cdot f _ {j} ^ {P} / \\tau\\right)}, \\tag {5}", + "image_path": "7a73f1e1a4bbeb7d1cecfd9e156cfda291c6fe2bfb8244fbbdb860039bdc1032.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 108, + 584, + 286, + 613 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 108, + 584, + 286, + 613 + ], + "spans": [ + { + "bbox": [ + 108, + 584, + 286, + 613 + ], + "type": "interline_equation", + "content": "L (I, P) = \\frac {1}{N} \\sum_ {i \\in N} l (i, I, P). \\tag {6}", + "image_path": "e365566fd14f38840e71ec6b86804707e2d167600e5787570ff094e1c5d21f58.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 617, + 287, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 617, + 287, + 676 + ], + "spans": [ + { + "bbox": [ + 47, + 617, + 287, + 676 + ], + "type": "text", + "content": "Finally, we obtain the resultant cross-modal contrastive learning objective " + }, + { + "bbox": [ + 47, + 617, + 287, + 676 + ], + "type": "inline_equation", + "content": "L_{CM}(T,I,P)" + }, + { + "bbox": [ + 47, + 617, + 287, + 676 + ], + "type": "text", + "content": " as the combination of " + }, + { + "bbox": [ + 47, + 617, + 287, + 676 + ], + "type": "inline_equation", + "content": "L(T,P)" + }, + { + "bbox": [ + 47, + 617, + 287, + 676 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 617, + 287, + 676 + ], + "type": "inline_equation", + "content": "L(I,P)" + }, + { + "bbox": [ + 47, + 617, + 287, + 676 + ], + "type": "text", + "content": ", where both alignments of semantic-level text-3D correlation and instance-level image-3D correlation are injected:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 86, + 686, + 286, + 698 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 86, + 686, + 286, + 698 + ], + "spans": [ + { + "bbox": [ + 86, + 686, + 286, + 698 + ], + "type": "interline_equation", + "content": "L _ {C M} (T, I, P) = \\lambda_ {1} L (T, P) + \\lambda_ {1} L (I, P), \\tag {7}", + "image_path": "4eb06d2e37c21e598b74f23d24f613cd3105944ff252d4b93cd9e52d0992c356.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 701, + 280, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 701, + 280, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 701, + 280, + 713 + ], + "type": "text", + "content": "where the hyper-parameters " + }, + { + "bbox": [ + 47, + 701, + 280, + 713 + ], + "type": "inline_equation", + "content": "\\lambda_{1}" + }, + { + "bbox": [ + 47, + 701, + 280, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 701, + 280, + 713 + ], + "type": "inline_equation", + "content": "\\lambda_{2}" + }, + { + "bbox": [ + 47, + 701, + 280, + 713 + ], + "type": "text", + "content": " are both set to 0.5." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 72, + 382, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 382, + 85 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 382, + 85 + ], + "type": "text", + "content": "4. Experiment" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 91, + 545, + 138 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 91, + 545, + 138 + ], + "spans": [ + { + "bbox": [ + 304, + 91, + 545, + 138 + ], + "type": "text", + "content": "In this section, we evaluate " + }, + { + "bbox": [ + 304, + 91, + 545, + 138 + ], + "type": "inline_equation", + "content": "\\mathrm{CLIP}^2" + }, + { + "bbox": [ + 304, + 91, + 545, + 138 + ], + "type": "text", + "content": " on realistic indoor and outdoor scenarios. We report the zero-shot transfer results on various datasets [2, 9, 32-34] and make further analysis on the designs of pretraining strategy." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 144, + 417, + 155 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 144, + 417, + 155 + ], + "spans": [ + { + "bbox": [ + 305, + 144, + 417, + 155 + ], + "type": "text", + "content": "4.1. Zero-shot Transfer" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 158, + 545, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 158, + 545, + 251 + ], + "spans": [ + { + "bbox": [ + 304, + 158, + 545, + 251 + ], + "type": "text", + "content": "Setting. After pretraining, natural language is applied to reference learned 3D representation to enable following zero-shot transfer tasks. (i) Zero-Shot Recognition: we evaluate zero-shot recognition performance for realistic objects, where " + }, + { + "bbox": [ + 304, + 158, + 545, + 251 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 304, + 158, + 545, + 251 + ], + "type": "text", + "content": " category names are transferred to text prompt \"point cloud of {CLASS}\" to encode the text features " + }, + { + "bbox": [ + 304, + 158, + 545, + 251 + ], + "type": "inline_equation", + "content": "F_{K} \\in \\mathbb{R}^{K \\times C}" + }, + { + "bbox": [ + 304, + 158, + 545, + 251 + ], + "type": "text", + "content": ". Then the classification logits are calculated with the 3D feature " + }, + { + "bbox": [ + 304, + 158, + 545, + 251 + ], + "type": "inline_equation", + "content": "f^{P}" + }, + { + "bbox": [ + 304, + 158, + 545, + 251 + ], + "type": "text", + "content": " and text features as:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 362, + 256, + 545, + 271 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 362, + 256, + 545, + 271 + ], + "spans": [ + { + "bbox": [ + 362, + 256, + 545, + 271 + ], + "type": "interline_equation", + "content": "\\log_ {i} = \\operatorname {s o f t m a x} \\left(f _ {i} ^ {P} \\left(F _ {K}\\right) ^ {T}\\right). \\tag {8}", + "image_path": "a6f3e2c94e713f51113ad1e2471a720cbd248494934fa358eff25226632291cd.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 277, + 545, + 440 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 277, + 545, + 440 + ], + "spans": [ + { + "bbox": [ + 304, + 277, + 545, + 440 + ], + "type": "text", + "content": "We present the results under both indoor and outdoor scenarios in Table 1, Table 2 and Table 5, as well as the object-level benchmark in Table 6. (ii) Open-vocabulary recognition: we enlarge the category vocabularies of ScanNet to 249 and 384 to study the open-vocabulary recognition ability in Table 3. (iii) Open-vocabulary localization: we study the open-vocabulary localization ability by localizing open-world 3D objects with our proxy generation process and then classifying them with our learned 3D representation, of which the visualization is illustrated in Figure 5 and evaluation results are reported in Table 4. Notably, we investigate representation ensembling alternatives to enable knowledge merging of all available representations for realistic applications, illustrated in Table 8." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 305, + 445, + 411, + 456 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 445, + 411, + 456 + ], + "spans": [ + { + "bbox": [ + 305, + 445, + 411, + 456 + ], + "type": "text", + "content": "4.1.1 Indoor Scenarios" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 460, + 545, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 460, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 460, + 545, + 715 + ], + "type": "text", + "content": "Datasets and details. We adopt the widely used indoor 3D dataset SUN RGB-D [32] as the realistic indoor scenario that provides pretraining data source, a single-view RGB-D dataset consisting of " + }, + { + "bbox": [ + 304, + 460, + 545, + 715 + ], + "type": "inline_equation", + "content": "\\sim 10\\mathrm{K}" + }, + { + "bbox": [ + 304, + 460, + 545, + 715 + ], + "type": "text", + "content": " scenes. To validate the transferability of learned 3D representation, we also evaluate another popular indoor 3D dataset ScanNet [9], which contains " + }, + { + "bbox": [ + 304, + 460, + 545, + 715 + ], + "type": "inline_equation", + "content": "\\sim 1.5\\mathrm{K}" + }, + { + "bbox": [ + 304, + 460, + 545, + 715 + ], + "type": "text", + "content": " scenes of 3D reconstructed meshes. We remove objects in ScanNet with less than 5 points, leaving 384 noisy categories. For open-vocabulary recognition, we evaluate performance on the ScanNet 384-class set and a 249-class merged set. In addition to the scene-wise indoor dataset, we conduct evaluations on ScanObjectNN [34], which collects " + }, + { + "bbox": [ + 304, + 460, + 545, + 715 + ], + "type": "inline_equation", + "content": "\\sim 3\\mathrm{K}" + }, + { + "bbox": [ + 304, + 460, + 545, + 715 + ], + "type": "text", + "content": " individual realistic objects with 15 categories and is applied in the previous zero-shot evaluation [13, 43]. During the proxy collection process, we empirically set " + }, + { + "bbox": [ + 304, + 460, + 545, + 715 + ], + "type": "inline_equation", + "content": "\\epsilon = 0.3" + }, + { + "bbox": [ + 304, + 460, + 545, + 715 + ], + "type": "text", + "content": " in [38] as a tradeoff between filtering FPs and preserving TPs to generate image proxies. Considering the occurrence frequencies of different indoor categories vary a lot, we adopt the class balance strategy [4] to mitigate the class imbalance. During pretraining process, we adopt [26] as point cloud encoder and set the overall training epoch number to 100." + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15248" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 52, + 71, + 542, + 163 + ], + "blocks": [ + { + "bbox": [ + 52, + 71, + 542, + 163 + ], + "lines": [ + { + "bbox": [ + 52, + 71, + 542, + 163 + ], + "spans": [ + { + "bbox": [ + 52, + 71, + 542, + 163 + ], + "type": "table", + "html": "
MethodAvg.BedBookshelfChairDeskSofaTabletoiletBathhubDresserNight Stand
PointClip [43]11.50.094.00.00.00.014.70.00.06.10.0
Clip2Point [13]18.610.920.664.334.413.814.126.20.01.40.0
PointClip [43] w/ TP.38.045.3100.062.548.544.44.855.216.33.30
Clip2Point [13] w/ TP.56.978.087.636.236.664.737.482.177.567.61.2
CLIP²61.384.075.570.747.375.533.886.265.371.82.4
CLIP² w/ En.69.687.394.370.754.179.347.091.785.782.23.6
", + "image_path": "d3341ade92a01b6750c52db0feb52b14e628493e390d9e0a8d93025414504f71.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 52, + 201, + 542, + 266 + ], + "blocks": [ + { + "bbox": [ + 47, + 167, + 546, + 190 + ], + "lines": [ + { + "bbox": [ + 47, + 167, + 546, + 190 + ], + "spans": [ + { + "bbox": [ + 47, + 167, + 546, + 190 + ], + "type": "text", + "content": "Table 1. Zero-shot recognition results in SUN RGB-D. Avg.: the mean average Top1 accuracy across all categories. TP.: our Triplet Proxy set. En.: our optional ensembling scheme for inference." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 52, + 201, + 542, + 266 + ], + "lines": [ + { + "bbox": [ + 52, + 201, + 542, + 266 + ], + "spans": [ + { + "bbox": [ + 52, + 201, + 542, + 266 + ], + "type": "table", + "html": "
MethodAvg.CabBedChairSofaTablDoorWindBksfPicCntrDeskCurtFridgBathShowrToilSink
PointClip [43]6.30.00.00.00.00.70.00.091.80.00.00.015.00.00.00.00.00.0
Clip2Point [13]24.90.020.885.143.326.569.90.020.91.731.727.00.01.646.50.022.425.6
PointClip [43] w/ TP.26.155.70.072.85.05.11.70.077.20.00.051.70.30.00.040.385.349.2
Clip2Point [13] w/ TP.35.23.011.845.127.610.561.52.671.90.333.629.94.711.572.292.486.134.0
CLIP²38.567.232.669.342.318.319.14.062.61.412.752.840.19.159.741.071.045.5
", + "image_path": "cf3afe102667b94e603f4ba677287620485560fd03c2b37b1ae5f2a6f55bde1f.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 51, + 294, + 137, + 345 + ], + "blocks": [ + { + "bbox": [ + 55, + 270, + 536, + 283 + ], + "lines": [ + { + "bbox": [ + 55, + 270, + 536, + 283 + ], + "spans": [ + { + "bbox": [ + 55, + 270, + 536, + 283 + ], + "type": "text", + "content": "Table 2. Zero-shot recognition in ScanNet. Avg.: the mean average Top1 accuracy across all categories. TP.: our Triplet Proxy set." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 51, + 294, + 137, + 345 + ], + "lines": [ + { + "bbox": [ + 51, + 294, + 137, + 345 + ], + "spans": [ + { + "bbox": [ + 51, + 294, + 137, + 345 + ], + "type": "table", + "html": "
MethodTop5 Acc.
384cls.249cls.
[43]0.30.4
[13]6.47.0
CLIP222.031.7
", + "image_path": "7d54b6c0708993235172dc22d2d5d5f8017ed63f7a3a8feee6d5f58147fbdaf0.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 141, + 294, + 278, + 345 + ], + "blocks": [ + { + "bbox": [ + 141, + 294, + 278, + 345 + ], + "lines": [ + { + "bbox": [ + 141, + 294, + 278, + 345 + ], + "spans": [ + { + "bbox": [ + 141, + 294, + 278, + 345 + ], + "type": "table", + "html": "
MethodmAP25IN AR25mIoUOUT P.R.
[21]1.3----
[18]13.037.7---
CLIP212.743.052.86.787.4
", + "image_path": "7c845af3cb741e054f9702aff52e6a54d8991f99880fa7b0d7bf3789526a54e7.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 348, + 282, + 381 + ], + "lines": [ + { + "bbox": [ + 47, + 348, + 282, + 381 + ], + "spans": [ + { + "bbox": [ + 47, + 348, + 282, + 381 + ], + "type": "text", + "content": "Table 3. Recognition results. Table 4. Zero-shot localization results of vocabulary ex-sults. IN: indoor scenario SUN RGBpansion in ScanNet. D. OUT: outdoor scenario nuScenes." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 399, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 399, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 399, + 289, + 715 + ], + "type": "text", + "content": "Quantity results. For zero-shot recognition task, we take two recent works as our baselines, i.e. PointClip [43] and Clip2Point [13], which study the zero-shot classification task on 3D object-level benchmarks [34, 35] by leveraging pretrained CLIP with projected depth maps. Focusing on the real-world scenarios, we conduct comparison not only on the realistic object-level [34] as illustrated in Table 6 but also on the scene-level datasets shown in Table 1 and Table 2, where the evaluation follows the common classes split in [10, 44] and reports the instance Top1 accuracy of each class. As shown in tables, our " + }, + { + "bbox": [ + 46, + 399, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\mathrm{CLIP}^2" + }, + { + "bbox": [ + 46, + 399, + 289, + 715 + ], + "type": "text", + "content": " can outperform baselines on all benchmarks by large margins. Besides, we apply our triplet proxy generation mechanism (TP.) to baseline methods, and achieve considerable improvements on SUN RGB-D and ScanNet by " + }, + { + "bbox": [ + 46, + 399, + 289, + 715 + ], + "type": "inline_equation", + "content": "26.5\\%" + }, + { + "bbox": [ + 46, + 399, + 289, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 399, + 289, + 715 + ], + "type": "inline_equation", + "content": "19.8\\%" + }, + { + "bbox": [ + 46, + 399, + 289, + 715 + ], + "type": "text", + "content": " for Point-Clip, " + }, + { + "bbox": [ + 46, + 399, + 289, + 715 + ], + "type": "inline_equation", + "content": "38.3\\%" + }, + { + "bbox": [ + 46, + 399, + 289, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 399, + 289, + 715 + ], + "type": "inline_equation", + "content": "10.3\\%" + }, + { + "bbox": [ + 46, + 399, + 289, + 715 + ], + "type": "text", + "content": " for Clip2Point. On the one hand, the contrasts demonstrate the effectiveness of our triplet proxies for open-world understanding. On the other hand, our learned 3D representation is superior in 3D object recognition by retaining more 3D-specific information than depth representation. Besides, we present the optional ensembling scheme (En.) when camera images are available, which can take advantage of multi-modal knowledge and further boost the performance by " + }, + { + "bbox": [ + 46, + 399, + 289, + 715 + ], + "type": "inline_equation", + "content": "8.3\\%" + }, + { + "bbox": [ + 46, + 399, + 289, + 715 + ], + "type": "text", + "content": ". To further validate the open-vocabulary recognition ability, we conduct evaluation on a larger category set of ScanNet in Table 3 and report the instance Top5 accuracy, which illustrates the superior" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 297, + 545, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 297, + 545, + 483 + ], + "spans": [ + { + "bbox": [ + 304, + 297, + 545, + 483 + ], + "type": "text", + "content": "ity of our " + }, + { + "bbox": [ + 304, + 297, + 545, + 483 + ], + "type": "inline_equation", + "content": "\\mathrm{CLIP}^2" + }, + { + "bbox": [ + 304, + 297, + 545, + 483 + ], + "type": "text", + "content": " when vocabulary increases. Beyond that, " + }, + { + "bbox": [ + 304, + 297, + 545, + 483 + ], + "type": "inline_equation", + "content": "\\mathrm{CLIP}^2" + }, + { + "bbox": [ + 304, + 297, + 545, + 483 + ], + "type": "text", + "content": " is also equipped with zero-shot 3D localization ability by proxy generation. On indoor scenario SUN RGB-D, we compare with a SOTA indoor 3D detector 3DETR [21] and a recent work OV3D [18] that studies open-vocabulary detection, where evaluation is conducted on the same \"unseen\" split in [18]. Since " + }, + { + "bbox": [ + 304, + 297, + 545, + 483 + ], + "type": "inline_equation", + "content": "\\mathrm{CLIP}^2" + }, + { + "bbox": [ + 304, + 297, + 545, + 483 + ], + "type": "text", + "content": " do not fit the tight bounding boxes of point cloud instances, we estimate the maximum bounding box of proxies and GT instances to conduct evaluation following the same metrics " + }, + { + "bbox": [ + 304, + 297, + 545, + 483 + ], + "type": "inline_equation", + "content": "\\mathrm{mAP}_{25}" + }, + { + "bbox": [ + 304, + 297, + 545, + 483 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 297, + 545, + 483 + ], + "type": "inline_equation", + "content": "\\mathrm{AR}_{25}" + }, + { + "bbox": [ + 304, + 297, + 545, + 483 + ], + "type": "text", + "content": " in [18], as shown in Table 4. Notably, compared to baseline works that train on \"seen\" 3D annotations and test on \"unseen\" categories, we have no access to any 3D annotations yet achieve comparable localization ability, which yields " + }, + { + "bbox": [ + 304, + 297, + 545, + 483 + ], + "type": "inline_equation", + "content": "5.3\\%" + }, + { + "bbox": [ + 304, + 297, + 545, + 483 + ], + "type": "inline_equation", + "content": "\\mathrm{AR}_{25}" + }, + { + "bbox": [ + 304, + 297, + 545, + 483 + ], + "type": "text", + "content": " improvement over OV3D [18]. We further evaluate segmentation results in Table 4." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 483, + 546, + 589 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 483, + 546, + 589 + ], + "spans": [ + { + "bbox": [ + 304, + 483, + 546, + 589 + ], + "type": "text", + "content": "Quality results. The visualization results of " + }, + { + "bbox": [ + 304, + 483, + 546, + 589 + ], + "type": "inline_equation", + "content": "\\mathrm{CLIP}^2" + }, + { + "bbox": [ + 304, + 483, + 546, + 589 + ], + "type": "text", + "content": " under a indoor scene of SUN RGB-D [32] is shown in Figure 5(a). Our triplet proxy generation process can localize open-world 3D objects in a point cloud scene. Moreover, the 3D representation learned from our cross-modal pretraining can provide more accurate classification results for 3D instances by exploiting original point cloud, which corrects the mistaken \"People\" prediction in image to \"Picture\" by considering the geometry information." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 594, + 421, + 604 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 594, + 421, + 604 + ], + "spans": [ + { + "bbox": [ + 305, + 594, + 421, + 604 + ], + "type": "text", + "content": "4.1.2 Outdoor Scenarios" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 607, + 545, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 607, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 304, + 607, + 545, + 689 + ], + "type": "text", + "content": "Datasets and details. We exploit a prevalent large-scale 3D dataset nuScenes [2] as the outdoor data source and extra validate the performance on the ONCE dataset [19]. The nuScenes dataset consists of " + }, + { + "bbox": [ + 304, + 607, + 545, + 689 + ], + "type": "inline_equation", + "content": "\\sim 28\\mathrm{K}" + }, + { + "bbox": [ + 304, + 607, + 545, + 689 + ], + "type": "text", + "content": " frames with 10 categories, while ONCE contains 6 annotated sequences with 5 categories. Similarly, we set the " + }, + { + "bbox": [ + 304, + 607, + 545, + 689 + ], + "type": "inline_equation", + "content": "\\epsilon = 0.3" + }, + { + "bbox": [ + 304, + 607, + 545, + 689 + ], + "type": "text", + "content": " for image proxies collection and adopt the class balance strategy [4]." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 689, + 545, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 689, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 305, + 689, + 545, + 715 + ], + "type": "text", + "content": "Quantity results. Since the outdoor point cloud is collected by LiDAR sensors, it has a wider perception range" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "15249" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 50, + 71, + 545, + 148 + ], + "blocks": [ + { + "bbox": [ + 50, + 71, + 545, + 148 + ], + "lines": [ + { + "bbox": [ + 50, + 71, + 545, + 148 + ], + "spans": [ + { + "bbox": [ + 50, + 71, + 545, + 148 + ], + "type": "table", + "html": "
MethodAvg.nuScenesONCE
CarTruckBusPed.BicycleTrailerC.V.Motor.BarrierT.C.CarCyc.Ped.TruckBus
PointClip [43]11.70.00.00.029.141.83.40.00.142.50.00.013.879.20.27.61
Clip2Point [13]12.40.40.30.113.531.01.89.31.566.20.317.311.895.435.74.0
PointClip [43] w/ TP.28.318.805.574.017.957.01.94.52.129.751.89.299.85.046.8
Clip2Point [13] w/ TP.33.026.716.851.245.215.813.920.05.710.534.239.427.895.540.651.7
CLIP²37.841.941.322.540.321.120.624.822.417.335.352.727.377.778.544.0
", + "image_path": "9f1d2bd50c49e4436b3d4ca69428bf693f8fc4be2e5eb924bf82e31cf81978ac.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 50, + 186, + 267, + 326 + ], + "blocks": [ + { + "bbox": [ + 50, + 186, + 267, + 326 + ], + "lines": [ + { + "bbox": [ + 50, + 186, + 267, + 326 + ], + "spans": [ + { + "bbox": [ + 50, + 186, + 267, + 326 + ], + "type": "image", + "image_path": "d4d7bd0c222aa751cda04a21ff7675f53f7b9569fb4b711efc61ca0fe9b02bfd.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 328, + 546, + 361 + ], + "lines": [ + { + "bbox": [ + 46, + 328, + 546, + 361 + ], + "spans": [ + { + "bbox": [ + 46, + 328, + 546, + 361 + ], + "type": "text", + "content": "Figure 5. Visualizations of the zero-shot localization and recognition results by " + }, + { + "bbox": [ + 46, + 328, + 546, + 361 + ], + "type": "inline_equation", + "content": "\\mathrm{CLIP}^2" + }, + { + "bbox": [ + 46, + 328, + 546, + 361 + ], + "type": "text", + "content": " under open-world (a) indoor realistic scene [32] and (b) outdoor scenes [2]. Notably, the whole pipeline of " + }, + { + "bbox": [ + 46, + 328, + 546, + 361 + ], + "type": "inline_equation", + "content": "\\mathrm{CLIP}^2" + }, + { + "bbox": [ + 46, + 328, + 546, + 361 + ], + "type": "text", + "content": " not only has no access to human annotations, but also enables the open-world vocabularies beyond groundtruth annotations, such as 'Picture' in (a) and 'Plastic bag', 'Tire' in (b). Best viewed in colors." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 269, + 186, + 545, + 326 + ], + "blocks": [ + { + "bbox": [ + 269, + 186, + 545, + 326 + ], + "lines": [ + { + "bbox": [ + 269, + 186, + 545, + 326 + ], + "spans": [ + { + "bbox": [ + 269, + 186, + 545, + 326 + ], + "type": "image", + "image_path": "7acfa580419ef00bd6b679fccb2d957142c4c7f782fc29dcc91736b580390ac4.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 52, + 380, + 285, + 457 + ], + "blocks": [ + { + "bbox": [ + 46, + 153, + 546, + 175 + ], + "lines": [ + { + "bbox": [ + 46, + 153, + 546, + 175 + ], + "spans": [ + { + "bbox": [ + 46, + 153, + 546, + 175 + ], + "type": "text", + "content": "Table 5. Zero-shot recognition results in outdoor scenario: nuScenes (Left) and ONCE (Right). TP.: our Triplet Proxy set. Avg.: the mean average Top1 accuracy across all categories of two benchmarks." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 52, + 380, + 285, + 457 + ], + "lines": [ + { + "bbox": [ + 52, + 380, + 285, + 457 + ], + "spans": [ + { + "bbox": [ + 52, + 380, + 285, + 457 + ], + "type": "table", + "html": "
MethodZS15-way10-way
4-S8-S16-S10-S20-S
PointNet++ [25]-41.047.655.0--
PointCLIP [43]15.446.050.055.6--
Clip2Point [13]23.3-----
CrossPoint [1]----58.7±1.864.6±1.2
CLIP239.151.359.662.560.6±2.566.3±3.2
", + "image_path": "61b581bd2423c1782fe57b6ce0bee93eef5a4279d446c209605f98d694dfd3f7.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 460, + 287, + 483 + ], + "lines": [ + { + "bbox": [ + 46, + 460, + 287, + 483 + ], + "spans": [ + { + "bbox": [ + 46, + 460, + 287, + 483 + ], + "type": "text", + "content": "Table 6. Zero-shot and Few-shot classification results on ScanObjectNN. ZS: zero-shot. K-way N-shot: few-shot settings." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 490, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 490, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 490, + 289, + 715 + ], + "type": "text", + "content": "than RGB-D but leads to sparse distribution. Thus the projected depth representation of baselines results in severer information lost, as illustrated in the second row in Figure 3. As shown in Table 5, our " + }, + { + "bbox": [ + 46, + 490, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\mathrm{CLIP}^2" + }, + { + "bbox": [ + 46, + 490, + 289, + 715 + ], + "type": "text", + "content": " considerably outperforms the baseline recognition results by more than " + }, + { + "bbox": [ + 46, + 490, + 289, + 715 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 46, + 490, + 289, + 715 + ], + "type": "text", + "content": ", and our triplet proxies respectively boost two baselines by " + }, + { + "bbox": [ + 46, + 490, + 289, + 715 + ], + "type": "inline_equation", + "content": "9.5\\%" + }, + { + "bbox": [ + 46, + 490, + 289, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 490, + 289, + 715 + ], + "type": "inline_equation", + "content": "4.8\\%" + }, + { + "bbox": [ + 46, + 490, + 289, + 715 + ], + "type": "text", + "content": ". Additionally, we evaluate the localization ability on the outdoor scenario nuScenes in Table 4. Due to the lack of works that tackle outdoor open-vocabulary localization problems, we follow classic detection accuracy metrics Precision(P.) and Recall(R.) as evaluation metrics. Specifically, we calculate the center distance between groundtruth bounding boxes and our 3D proxies that are predicted to belong to the same category as the groundtruth, and set the distance threshold as " + }, + { + "bbox": [ + 46, + 490, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\lambda = 2\\mathrm{m}" + }, + { + "bbox": [ + 46, + 490, + 289, + 715 + ], + "type": "text", + "content": ". For those matched pairs that are closer than " + }, + { + "bbox": [ + 46, + 490, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 46, + 490, + 289, + 715 + ], + "type": "text", + "content": ", we count the proxies as TPs. Otherwise, for those unmatched proxies and groundtruth, we count them as FPs and FNs respectively, thus " + }, + { + "bbox": [ + 46, + 490, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\mathrm{P.} = \\frac{\\mathrm{TPs}}{\\mathrm{TPs} + \\mathrm{FPs}}" + }, + { + "bbox": [ + 46, + 490, + 289, + 715 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 490, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\mathrm{R.} = \\frac{\\mathrm{TPs}}{\\mathrm{TPs} + \\mathrm{FNs}}" + }, + { + "bbox": [ + 46, + 490, + 289, + 715 + ], + "type": "text", + "content": ". As shown in Table 4, our " + }, + { + "bbox": [ + 46, + 490, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\mathrm{CLIP}^2" + }, + { + "bbox": [ + 46, + 490, + 289, + 715 + ], + "type": "text", + "content": " pipeline can provide high" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 381, + 546, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 381, + 546, + 441 + ], + "spans": [ + { + "bbox": [ + 304, + 381, + 546, + 441 + ], + "type": "text", + "content": "recall for outdoor objects. Since " + }, + { + "bbox": [ + 304, + 381, + 546, + 441 + ], + "type": "inline_equation", + "content": "\\mathrm{CLIP}^2" + }, + { + "bbox": [ + 304, + 381, + 546, + 441 + ], + "type": "text", + "content": " is highly sensitive to open-world objects and can perceive categories beyond groundtruth list, it tends to create overmuch predictions thus the precision is comparably low. The perception ability of open-world objects can be viewed in Figure 5(b)." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 444, + 547, + 550 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 444, + 547, + 550 + ], + "spans": [ + { + "bbox": [ + 304, + 444, + 547, + 550 + ], + "type": "text", + "content": "Quality results. We show off two outdoor scenes of nuScenes [2] in Figure 5(b-i) and Figure 5(b-ii). In addition to perceiving those common categories, our " + }, + { + "bbox": [ + 304, + 444, + 547, + 550 + ], + "type": "inline_equation", + "content": "\\mathrm{CLIP}^2" + }, + { + "bbox": [ + 304, + 444, + 547, + 550 + ], + "type": "text", + "content": " surprisingly localizes and recognizes those uncommon 3D objects in 3D scenes such as the tires of vehicles, the plastic bag in the hand of pedestrian as well as the plastic bag on the road. We believe it contributes to auto-driving safety by providing the localization and recognition of universal obstacles to facilitate follow-up driving decisions." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 555, + 437, + 567 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 555, + 437, + 567 + ], + "spans": [ + { + "bbox": [ + 305, + 555, + 437, + 567 + ], + "type": "text", + "content": "4.2. Few-shot Classification" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 569, + 547, + 674 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 569, + 547, + 674 + ], + "spans": [ + { + "bbox": [ + 304, + 569, + 547, + 674 + ], + "type": "text", + "content": "Setting. Lightweight few-shot learning is practical for application by finetuning the pretraining model with given limited data annotations, which can also validate the generalization capability of our learned representation. To make a fair comparison, we follow the existing methods [1, 43] to conduct experiments under \"K-way N-shot\" setting on the challenging realistic object-level dataset ScanObjectNN [34], where we randomly sample N point cloud objects from each of the randomly selected K classes." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 678, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 678, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 678, + 547, + 715 + ], + "type": "text", + "content": "Quantity results. As illustrated in Table 6, we compare with representative 3D networks PointNet++ [26], the recent zero-shot approach PointClip [43] as well as a state-" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "15250" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 51, + 71, + 288, + 144 + ], + "blocks": [ + { + "bbox": [ + 51, + 71, + 288, + 144 + ], + "lines": [ + { + "bbox": [ + 51, + 71, + 288, + 144 + ], + "spans": [ + { + "bbox": [ + 51, + 71, + 288, + 144 + ], + "type": "table", + "html": "
Rep.Obj.Avg..INAvg..OUTAvg..OBJ
(a)DepthLang.-Depth38.09.331.8
(b)DepthImage-Depth56.421.739.0
(c)PC.Lang.-Point52.626.837.7
(d)PC.Image-Point56.524.437.8
(e)PC.Lang.-Image-Point61.328.839.4
", + "image_path": "7c879c8d5eb9a4f0bac0c96b6bb09d00402f7e0572a079e0a0a11c9a863c5121.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 194, + 288, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 194, + 288, + 323 + ], + "spans": [ + { + "bbox": [ + 46, + 194, + 288, + 323 + ], + "type": "text", + "content": "of-the-art representation learning method CrossPoint [1], which conducts contrastive pretraining between point cloud and rendered images on CAD dataset ShapeNet [3]. As we can see, with a slight number of samples, our " + }, + { + "bbox": [ + 46, + 194, + 288, + 323 + ], + "type": "inline_equation", + "content": "\\mathrm{CLIP}^2" + }, + { + "bbox": [ + 46, + 194, + 288, + 323 + ], + "type": "text", + "content": " can boost the classification results by a large margin, exceeding PointClip by " + }, + { + "bbox": [ + 46, + 194, + 288, + 323 + ], + "type": "inline_equation", + "content": "5.3\\%" + }, + { + "bbox": [ + 46, + 194, + 288, + 323 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 194, + 288, + 323 + ], + "type": "inline_equation", + "content": "9.6\\%" + }, + { + "bbox": [ + 46, + 194, + 288, + 323 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 194, + 288, + 323 + ], + "type": "inline_equation", + "content": "6.9\\%" + }, + { + "bbox": [ + 46, + 194, + 288, + 323 + ], + "type": "text", + "content": " with 4, 8 and 16 shots. Besides, we outperform CrossPoint with considerable gain, illustrating our pretraining strategy on collected proxies can learn sufficient knowledge from realistic open world to generate transferable 3D representation, which is superior to the pretraining on a small-scale synthetic dataset." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 330, + 178, + 342 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 330, + 178, + 342 + ], + "spans": [ + { + "bbox": [ + 47, + 330, + 178, + 342 + ], + "type": "text", + "content": "4.3. Ablations and Analysis" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 346, + 287, + 637 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 346, + 287, + 637 + ], + "spans": [ + { + "bbox": [ + 46, + 346, + 287, + 637 + ], + "type": "text", + "content": "Ablations on representation learning. To observe the transferability of different representations and the effect of different learning objectives, we conduct ablations and report the mean average Top1 accuracy across all classes of zero-shot recognition in indoor scenarios [32] (Avg..IN), outdoor scenarios [2] (Avg..OUT) and object-level benchmark [34] (Avg..OBJ), which is shown in Table 7. Firstly, for fair comparisons, we follow [13, 43] to project input point cloud into depth maps in " + }, + { + "bbox": [ + 46, + 346, + 287, + 637 + ], + "type": "inline_equation", + "content": "N_V" + }, + { + "bbox": [ + 46, + 346, + 287, + 637 + ], + "type": "text", + "content": " different views as alternative representation. Secondly, we adopt various objectives to learn different correlation alignments across language, image and point cloud feature space or depth space. Specifically, comparing (a) and (b), aligning depth space to image space yields better transfer performance due to the similar data structure of image and depth map. In (c) and (d), point cloud representation is better when aligning to image space in indoor scenes, while better to align with language space in outdoor scenes due to the data discrepancy between image-like RGB-D points and sparse LiDAR points. Generally, 3D point cloud representation outperforms depth representation in all benchmarks due to preserving the complete 3D structure and sufficient 3D-specific information. Comparing (e) and (d), the joint alignment between three feature spaces contributes to the best 3D point cloud representation transferability on all benchmarks." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 643, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 643, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 643, + 287, + 714 + ], + "type": "text", + "content": "Analysis of representation ensembling. Intuitively, different representations contain different perspectives of knowledge, which can be potentially merged to achieve the optimum results during inference. To validate the ensembling application, we adopt three optional representation modals, i.e. point clouds, projected depth maps and cor" + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 310, + 71, + 543, + 153 + ], + "blocks": [ + { + "bbox": [ + 46, + 148, + 287, + 181 + ], + "lines": [ + { + "bbox": [ + 46, + 148, + 287, + 181 + ], + "spans": [ + { + "bbox": [ + 46, + 148, + 287, + 181 + ], + "type": "text", + "content": "Table 7. Ablations on representation learning. Rep.: the recognition representations. Obj.: the learning objectives. PC.: point cloud. Lang.: language space." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 310, + 71, + 543, + 153 + ], + "lines": [ + { + "bbox": [ + 310, + 71, + 543, + 153 + ], + "spans": [ + { + "bbox": [ + 310, + 71, + 543, + 153 + ], + "type": "table", + "html": "
PC.ImageDepthAvg..INAvg..OUTAvg..OBJ
(i)✓61.328.839.4
(ii)✓64.241.1-
(iii)✓56.923.939.0
(f)✓✓68.743.9-
(g)✓✓64.830.443.2
(h)✓✓✓69.642.3-
", + "image_path": "95c9e5f9865366c014e808607f3680b44da597dee886ae4a757b35313df6f44e.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 157, + 537, + 168 + ], + "lines": [ + { + "bbox": [ + 313, + 157, + 537, + 168 + ], + "spans": [ + { + "bbox": [ + 313, + 157, + 537, + 168 + ], + "type": "text", + "content": "Table 8. Analysis on the representation ensembling schemes." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 179, + 545, + 366 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 179, + 545, + 366 + ], + "spans": [ + { + "bbox": [ + 304, + 179, + 545, + 366 + ], + "type": "text", + "content": "responding image patches, where depth representation is trained on our proxies and image representation is generated from pretrained image branch of CLIP. We ensemble their predicted logits by simple summation as the final output, and illustrate the separate recognition results and ensembling performance in Table 8. Benefiting from the sufficient knowledge learned from massive CLIP training data, image representation presents best performance in separate applications. By merging the complementary knowledge, our 3D representation leads to gains of " + }, + { + "bbox": [ + 304, + 179, + 545, + 366 + ], + "type": "inline_equation", + "content": "4.6\\%" + }, + { + "bbox": [ + 304, + 179, + 545, + 366 + ], + "type": "text", + "content": " indoors [32] and " + }, + { + "bbox": [ + 304, + 179, + 545, + 366 + ], + "type": "inline_equation", + "content": "2.8\\%" + }, + { + "bbox": [ + 304, + 179, + 545, + 366 + ], + "type": "text", + "content": " outdoors. Though further improving the indoor recognition performance with " + }, + { + "bbox": [ + 304, + 179, + 545, + 366 + ], + "type": "inline_equation", + "content": "0.9\\%" + }, + { + "bbox": [ + 304, + 179, + 545, + 366 + ], + "type": "text", + "content": " when merged, depth representation yeilds " + }, + { + "bbox": [ + 304, + 179, + 545, + 366 + ], + "type": "inline_equation", + "content": "1.6\\%" + }, + { + "bbox": [ + 304, + 179, + 545, + 366 + ], + "type": "text", + "content": " drop for outdoor objects, illustrating the information lost especially for outdoor scenarios. Since image representation is sometimes missing, such as in [34], our 3D representation is more robust for 3D applications." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 375, + 375, + 388 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 375, + 375, + 388 + ], + "spans": [ + { + "bbox": [ + 306, + 375, + 375, + 388 + ], + "type": "text", + "content": "5. Limitation" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 395, + 545, + 500 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 395, + 545, + 500 + ], + "spans": [ + { + "bbox": [ + 304, + 395, + 545, + 500 + ], + "type": "text", + "content": "As a pilot work for the language-3D pretraining problem, though " + }, + { + "bbox": [ + 304, + 395, + 545, + 500 + ], + "type": "inline_equation", + "content": "\\mathrm{CLIP}^2" + }, + { + "bbox": [ + 304, + 395, + 545, + 500 + ], + "type": "text", + "content": " enables zero-shot localization and recognition with proposed triplet proxy generation and learned transferable 3D representation, it can not provide the accurate tight bounding box for open-world 3D objects as a common detector does. We believe " + }, + { + "bbox": [ + 304, + 395, + 545, + 500 + ], + "type": "inline_equation", + "content": "\\mathrm{CLIP}^2" + }, + { + "bbox": [ + 304, + 395, + 545, + 500 + ], + "type": "text", + "content": " can facilitate the development of open-world 3D detectors by introducing the recognition ability to general 3D detectors or providing presented 3D proxies to enable further training of 3D detectors." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 510, + 378, + 521 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 510, + 378, + 521 + ], + "spans": [ + { + "bbox": [ + 306, + 510, + 378, + 521 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 529, + 545, + 669 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 529, + 545, + 669 + ], + "spans": [ + { + "bbox": [ + 304, + 529, + 545, + 669 + ], + "type": "text", + "content": "In this paper, we present a novel contrastive language-image-point cloud pretraining framework, " + }, + { + "bbox": [ + 304, + 529, + 545, + 669 + ], + "type": "inline_equation", + "content": "\\mathrm{CLIP}^2" + }, + { + "bbox": [ + 304, + 529, + 545, + 669 + ], + "type": "text", + "content": ", which consists of a triplet proxy collection scheme and a cross-modal contrastive learning mechanism. Based on the observation that realistic scenarios contain a massive amount of open-world objects, we innovatively propose to collect triplet proxies from realistic scenes as pretraining data. We then conduct cross-modal contrastive alignment across language, image and point cloud feature space to learn transferable 3D representation. The zero-shot transfer results on various indoor and outdoor benchmarks validate the ability of " + }, + { + "bbox": [ + 304, + 529, + 545, + 669 + ], + "type": "inline_equation", + "content": "\\mathrm{CLIP}^2" + }, + { + "bbox": [ + 304, + 529, + 545, + 669 + ], + "type": "text", + "content": " for 3D open-world understanding." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 672, + 545, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 672, + 545, + 696 + ], + "spans": [ + { + "bbox": [ + 305, + 672, + 545, + 696 + ], + "type": "text", + "content": "Acknowledgements We acknowledge the support of MindSpore1, CANN, and Ascend AI Processor in this work." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 317, + 702, + 410, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 702, + 410, + 712 + ], + "spans": [ + { + "bbox": [ + 317, + 702, + 410, + 712 + ], + "type": "text", + "content": "1https://www.mindspore.cn/" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15251" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 144 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 144 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 144 + ], + "type": "text", + "content": "[1] Mohamed Afham, Isuru Dissanayake, Dinithi Dissanayake, Amaya Dharmasiri, Kanchana Thilakarathna, and Ranga Rodrigo. Crosspoint: Self-supervised cross-modal contrastive learning for 3d point cloud understanding. In CVPR, 2022. 2, 3, 7, 8" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 144, + 288, + 197 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 144, + 288, + 197 + ], + "spans": [ + { + "bbox": [ + 53, + 144, + 288, + 197 + ], + "type": "text", + "content": "[2] Holger Caesar, Varun Bankiti, Alex H Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuscenes: A multimodal dataset for autonomous driving. In CVPR, 2020. 2, 4, 5, 6, 7, 8" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 198, + 288, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 198, + 288, + 251 + ], + "spans": [ + { + "bbox": [ + 53, + 198, + 288, + 251 + ], + "type": "text", + "content": "[3] Angel X Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, et al. Shapenet: An information-rich 3d model repository. arXiv preprint arXiv:1512.03012, 2015. 8" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 251, + 288, + 336 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 251, + 288, + 336 + ], + "spans": [ + { + "bbox": [ + 53, + 251, + 288, + 336 + ], + "type": "text", + "content": "[4] Kai Chen, Jiaqi Wang, Jiangmiao Pang, Yuhang Cao, Yu Xiong, Xiaoxiao Li, Shuyang Sun, Wansen Feng, Ziwei Liu, Jiarui Xu, Zheng Zhang, Dazhi Cheng, Chenchen Zhu, Tianheng Cheng, Qijie Zhao, Buyu Li, Xin Lu, Rui Zhu, Yue Wu, Jifeng Dai, Jingdong Wang, Jianping Shi, Wanli Ouyang, Chen Change Loy, and Dahua Lin. MMDetection: Open mmlab detection toolbox and benchmark. arXiv preprint arXiv:1906.07155, 2019. 5, 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 337, + 288, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 337, + 288, + 369 + ], + "spans": [ + { + "bbox": [ + 53, + 337, + 288, + 369 + ], + "type": "text", + "content": "[5] Ali Cheraghian, Shafin Rahman, Dylan Campbell, and Lars Petersson. Mitigating the hubness problem for zero-shot learning of 3d objects. In BMVC, 2019. 2, 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 369, + 288, + 401 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 369, + 288, + 401 + ], + "spans": [ + { + "bbox": [ + 53, + 369, + 288, + 401 + ], + "type": "text", + "content": "[6] Ali Cheraghian, Shafin Rahman, Dylan Campbell, and Lars Petersson. Transductive zero-shot learning for 3d point cloud classification. In WACV, 2020. 2, 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 402, + 288, + 434 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 402, + 288, + 434 + ], + "spans": [ + { + "bbox": [ + 53, + 402, + 288, + 434 + ], + "type": "text", + "content": "[7] Ali Cheraghian, Shafin Rahman, Townim F Chowdhury, Dylan Campbell, and Lars Petersson. Zero-shot learning on 3d point cloud objects and beyond. IJCV, 2022. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 434, + 288, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 434, + 288, + 464 + ], + "spans": [ + { + "bbox": [ + 53, + 434, + 288, + 464 + ], + "type": "text", + "content": "[8] Ali Cheraghian, Shafin Rahman, and Lars Petersson. Zeroshot learning of 3d point cloud objects. In MVA. IEEE, 2019. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 466, + 288, + 508 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 466, + 288, + 508 + ], + "spans": [ + { + "bbox": [ + 53, + 466, + 288, + 508 + ], + "type": "text", + "content": "[9] Angela Dai, Angel X Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In CVPR, 2017. 2, 4, 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 509, + 288, + 540 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 509, + 288, + 540 + ], + "spans": [ + { + "bbox": [ + 48, + 509, + 288, + 540 + ], + "type": "text", + "content": "[10] Zhipeng Ding, Xu Han, and Marc Niethammer. Votenet: A deep learning label fusion method for multi-atlas segmentation. In MICCAI, 2019. 1, 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 541, + 288, + 573 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 541, + 288, + 573 + ], + "spans": [ + { + "bbox": [ + 48, + 541, + 288, + 573 + ], + "type": "text", + "content": "[11] Agrim Gupta, Piotr Dollar, and Ross Girshick. Lvis: A dataset for large vocabulary instance segmentation. In CVPR, pages 5356-5364, 2019. 4" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 574, + 288, + 605 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 574, + 288, + 605 + ], + "spans": [ + { + "bbox": [ + 48, + 574, + 288, + 605 + ], + "type": "text", + "content": "[12] Huy Ha and Shuran Song. Semantic abstraction: Openworld 3d scene understanding from 2d vision-language models. In CoRL, 2022. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 606, + 288, + 659 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 606, + 288, + 659 + ], + "spans": [ + { + "bbox": [ + 48, + 606, + 288, + 659 + ], + "type": "text", + "content": "[13] Tianyu Huang, Bowen Dong, Yunhan Yang, Xiaoshui Huang, Rynson WH Lau, Wanli Ouyang, and Wangmeng Zuo. Clip2point: Transfer clip to point cloud classification with image-depth pre-training. arXiv preprint arXiv:2210.01055, 2022. 2, 3, 4, 5, 6, 7, 8" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 660, + 288, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 660, + 288, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 660, + 288, + 712 + ], + "type": "text", + "content": "[14] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In ICML. PMLR, 2021. 2, 3" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 714 + ], + "type": "list", + "angle": 0, + "index": 32, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "text", + "content": "[15] Junnan Li, Ramprasaath Selvaraju, Akhilesh Gotmare, Shafiq Joty, Caiming Xiong, and Steven Chu Hong Hoi. Align before fuse: Vision and language representation learning with momentum distillation. NeurIPS, 2021. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 117, + 545, + 169 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 117, + 545, + 169 + ], + "spans": [ + { + "bbox": [ + 307, + 117, + 545, + 169 + ], + "type": "text", + "content": "[16] Hanxue Liang, Chenhan Jiang, Dapeng Feng, Xin Chen, Hang Xu, Xiaodan Liang, Wei Zhang, Zhenguo Li, and Luc Van Gool. Exploring geometry-aware contrast and clustering harmonization for self-supervised 3d object detection. In ICCV, 2021. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 171, + 545, + 201 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 171, + 545, + 201 + ], + "spans": [ + { + "bbox": [ + 307, + 171, + 545, + 201 + ], + "type": "text", + "content": "[17] Haotian Liu, Mu Cai, and Yong Jae Lee. Masked discrimination for self-supervised learning on point clouds. In ECCV, 2022. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 203, + 545, + 255 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 203, + 545, + 255 + ], + "spans": [ + { + "bbox": [ + 307, + 203, + 545, + 255 + ], + "type": "text", + "content": "[18] Yuheng Lu, Chenfeng Xu, Xiaobao Wei, Xiaodong Xie, Masayoshi Tomizuka, Kurt Keutzer, and Shanghang Zhang. Open-vocabulary 3d detection via image-level class and de-biased cross-modal contrastive learning. arXiv preprint arXiv:2207.01987, 2022. 2, 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 257, + 545, + 310 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 257, + 545, + 310 + ], + "spans": [ + { + "bbox": [ + 307, + 257, + 545, + 310 + ], + "type": "text", + "content": "[19] Jiageng Mao, Minzhe Niu, Chenhan Jiang, Hanxue Liang, Jingheng Chen, Xiaodan Liang, Yamin Li, Chaoqiang Ye, Wei Zhang, Zhenguo Li, et al. One million scenes for autonomous driving: Once dataset. arXiv preprint arXiv:2106.11037, 2021. 2, 4, 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 312, + 545, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 312, + 545, + 353 + ], + "spans": [ + { + "bbox": [ + 307, + 312, + 545, + 353 + ], + "type": "text", + "content": "[20] Jiageng Mao, Shaoshuai Shi, Xiaogang Wang, and Hongsheng Li. 3d object detection for autonomous driving: A review and new outlooks. arXiv preprint arXiv:2206.09474, 2022.1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 354, + 545, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 354, + 545, + 386 + ], + "spans": [ + { + "bbox": [ + 307, + 354, + 545, + 386 + ], + "type": "text", + "content": "[21] Ishan Misra, Rohit Girdhar, and Armand Joulin. An end-to-end transformer model for 3d object detection. In ICCV, 2021. 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 387, + 545, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 387, + 545, + 430 + ], + "spans": [ + { + "bbox": [ + 307, + 387, + 545, + 430 + ], + "type": "text", + "content": "[22] Anshul Paigwar, David Sierra-Gonzalez, Özgür Erkent, and Christian Laugier. Frustum-pointpillars: A multi-stage approach for 3d object detection using rgb camera and lidar. In ICCV, 2021. 4" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 431, + 545, + 463 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 431, + 545, + 463 + ], + "spans": [ + { + "bbox": [ + 307, + 431, + 545, + 463 + ], + "type": "text", + "content": "[23] Yatian Pang, Wenxiao Wang, Francis EH Tay, Wei Liu, Yonghong Tian, and Li Yuan. Masked autoencoders for point cloud self-supervised learning. In ECCV, 2022. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 464, + 545, + 495 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 464, + 545, + 495 + ], + "spans": [ + { + "bbox": [ + 307, + 464, + 545, + 495 + ], + "type": "text", + "content": "[24] Charles R Qi, Wei Liu, Chenxia Wu, Hao Su, and Leonidas J Guibas. Frustum pointnets for 3d object detection from rgb-d data. In CVPR, 2018. 4" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 496, + 545, + 528 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 496, + 545, + 528 + ], + "spans": [ + { + "bbox": [ + 307, + 496, + 545, + 528 + ], + "type": "text", + "content": "[25] Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. Pointnet: Deep learning on point sets for 3d classification and segmentation. In CVPR, 2017. 1, 2, 7" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 529, + 545, + 560 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 529, + 545, + 560 + ], + "spans": [ + { + "bbox": [ + 307, + 529, + 545, + 560 + ], + "type": "text", + "content": "[26] Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J Guibas. Pointnet++: Deep hierarchical feature learning on point sets in a metric space. NeurIPS, 2017. 5, 7" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 561, + 545, + 614 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 561, + 545, + 614 + ], + "spans": [ + { + "bbox": [ + 307, + 561, + 545, + 614 + ], + "type": "text", + "content": "[27] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, 2021. 2, 3, 4, 5" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 616, + 545, + 647 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 616, + 545, + 647 + ], + "spans": [ + { + "bbox": [ + 307, + 616, + 545, + 647 + ], + "type": "text", + "content": "[28] Carsten Rother, Vladimir Kolmogorov, and Andrew Blake. \"grabcut\" interactive foreground extraction using iterated graph cuts. TOG, 2004. 4" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 307, + 647, + 545, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 647, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 307, + 647, + 545, + 689 + ], + "type": "text", + "content": "[29] Erich Schubert, Jörg Sander, Martin Ester, Hans Peter Kriegel, and Xiaowei Xu. Dbscan revisited, revisited: why and how you should (still) use dbscan. *TODS*, 42(3):1-21, 2017. 4" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 307, + 691, + 545, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 691, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 307, + 691, + 545, + 714 + ], + "type": "text", + "content": "[30] Charu Sharma and Manohar Kaul. Self-supervised few-shot learning on point clouds. NeurIPS, 2020. 3" + } + ] + } + ], + "index": 31 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "15252" + } + ] + } + ], + "index": 33 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 659 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 105 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 105 + ], + "type": "text", + "content": "[31] Shaoshuai Shi, Xiaogang Wang, and Hongsheng Li. Pointcnn: 3D object proposal generation and detection from point cloud. In CVPR, 2019. 1" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 106, + 286, + 137 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 106, + 286, + 137 + ], + "spans": [ + { + "bbox": [ + 49, + 106, + 286, + 137 + ], + "type": "text", + "content": "[32] Shuran Song, Samuel P Lichtenberg, and Jianxiong Xiao. Sun rgb-d: A rgb-d scene understanding benchmark suite. In CVPR, 2015. 2, 4, 5, 6, 7, 8" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 139, + 286, + 191 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 139, + 286, + 191 + ], + "spans": [ + { + "bbox": [ + 49, + 139, + 286, + 191 + ], + "type": "text", + "content": "[33] Pei Sun, Henrik Kretzschmar, Xerxes Dotiwalla, Aurelien Chouard, Vijaysai Patnaik, Paul Tsui, James Guo, Yin Zhou, Yuning Chai, Benjamin Caine, et al. Scalability in perception for autonomous driving: Waymo open dataset. In CVPR, 2020. 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 194, + 286, + 235 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 194, + 286, + 235 + ], + "spans": [ + { + "bbox": [ + 49, + 194, + 286, + 235 + ], + "type": "text", + "content": "[34] Mikaela Angelina Uy, Quang-Hieu Pham, Binh-Son Hua, Thanh Nguyen, and Sai-Kit Yeung. Revisiting point cloud classification: A new benchmark dataset and classification model on real-world data. In ICCV, 2019. 2, 5, 6, 7, 8" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 49, + 237, + 286, + 279 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 237, + 286, + 279 + ], + "spans": [ + { + "bbox": [ + 49, + 237, + 286, + 279 + ], + "type": "text", + "content": "[35] Kashi Venkatesh Vishwanath, Diwaker Gupta, Amin Vahdat, and Ken Yocum. Modelnet: Towards a datacenter emulation environment. In 2009 IEEE Ninth International Conference on Peer-to-Peer Computing. IEEE, 2009. 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 49, + 281, + 286, + 312 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 281, + 286, + 312 + ], + "spans": [ + { + "bbox": [ + 49, + 281, + 286, + 312 + ], + "type": "text", + "content": "[36] Hanchen Wang, Qi Liu, Xiangyu Yue, Joan Lasenby, and Matt J Kusner. Unsupervised point cloud pre-training via occlusion completion. In ICCV, 2021. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 49, + 313, + 286, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 313, + 286, + 354 + ], + "spans": [ + { + "bbox": [ + 49, + 313, + 286, + 354 + ], + "type": "text", + "content": "[37] Saining Xie, Jiatao Gu, Demi Guo, Charles R Qi, Leonidas Guibas, and Or Litany. Pointcontrast: Unsupervised pretraining for 3d point cloud understanding. In ECCV, 2020. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 49, + 357, + 286, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 357, + 286, + 408 + ], + "spans": [ + { + "bbox": [ + 49, + 357, + 286, + 408 + ], + "type": "text", + "content": "[38] Lewei Yao, Jianhua Han, Youpeng Wen, Xiaodan Liang, Dan Xu, Wei Zhang, Zhenguo Li, Chunjing Xu, and Hang Xu. Detclip: Dictionary-enriched visual-concept paralleled pretraining for open-world detection. In NeurIPS, 2022. 2, 4, 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 49, + 411, + 286, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 411, + 286, + 453 + ], + "spans": [ + { + "bbox": [ + 49, + 411, + 286, + 453 + ], + "type": "text", + "content": "[39] Yangyang Ye, Houjin Chen, Chi Zhang, Xiaoli Hao, and Zhaoxiang Zhang. Sarpnet: Shape attention regional proposal network for lidar-based 3d object detection. Neurocomputing, 2020. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 49, + 454, + 286, + 485 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 454, + 286, + 485 + ], + "spans": [ + { + "bbox": [ + 49, + 454, + 286, + 485 + ], + "type": "text", + "content": "[40] Tianwei Yin, Xingyi Zhou, and Philipp Kähenbuhl. Center-based 3d object detection and tracking. arXiv preprint arXiv:2006.11275, 2020. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 49, + 487, + 286, + 528 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 487, + 286, + 528 + ], + "spans": [ + { + "bbox": [ + 49, + 487, + 286, + 528 + ], + "type": "text", + "content": "[41] Xumin Yu, Lulu Tang, Yongming Rao, Tiejun Huang, Jie Zhou, and Jiwen Lu. Point-bert: Pre-training 3d point cloud transformers with masked point modeling. In CVPR, 2022. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 49, + 530, + 286, + 582 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 530, + 286, + 582 + ], + "spans": [ + { + "bbox": [ + 49, + 530, + 286, + 582 + ], + "type": "text", + "content": "[42] Renrui Zhang, Ziyu Guo, Peng Gao, Rongyao Fang, Bin Zhao, Dong Wang, Yu Qiao, and Hongsheng Li. Pointm2ae: Multi-scale masked autoencoders for hierarchical point cloud pre-training. arXiv preprint arXiv:2205.14401, 2022.3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 49, + 585, + 286, + 626 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 585, + 286, + 626 + ], + "spans": [ + { + "bbox": [ + 49, + 585, + 286, + 626 + ], + "type": "text", + "content": "[43] Renrui Zhang, Ziyu Guo, Wei Zhang, Kunchang Li, Xu-peng Miao, Bin Cui, Yu Qiao, Peng Gao, and Hongsheng Li. Pointclip: Point cloud understanding by clip. In CVPR, 2022. 2, 3, 4, 5, 6, 7, 8" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 49, + 628, + 286, + 659 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 628, + 286, + 659 + ], + "spans": [ + { + "bbox": [ + 49, + 628, + 286, + 659 + ], + "type": "text", + "content": "[44] Zaiwei Zhang, Bo Sun, Haitao Yang, and Qixing Huang. H3dnet: 3d object detection using hybrid geometric primitives. In ECCV, 2020. 1, 6" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "15253" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/CLIPPING_ Distilling CLIP-Based Models With a Student Base for Video-Language Retrieval/1eb2010b-78e6-492d-9404-5d49a6b93ef0_content_list.json b/2023/CLIPPING_ Distilling CLIP-Based Models With a Student Base for Video-Language Retrieval/1eb2010b-78e6-492d-9404-5d49a6b93ef0_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..6b1efe7ee9b92725f52cf3ce2f68e2e1a8837c1c --- /dev/null +++ b/2023/CLIPPING_ Distilling CLIP-Based Models With a Student Base for Video-Language Retrieval/1eb2010b-78e6-492d-9404-5d49a6b93ef0_content_list.json @@ -0,0 +1,1720 @@ +[ + { + "type": "text", + "text": "CLIPPING: Distilling CLIP-Based Models with a Student Base for Video-Language Retrieval", + "text_level": 1, + "bbox": [ + 148, + 128, + 821, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Renjing Pei, Jianzhuang Liu, Weimian Li, Bin Shao, Songcen Xu, Peng Dai, Juwei Lu, Youliang Yan Huawei Noah's Ark Lab", + "bbox": [ + 253, + 203, + 715, + 256 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{peirenjing,liu.jianzhuang,liweimian,shaobin3, xusongcen,peng.dai,juwei.lu,yanyouliang}@huawei.com", + "bbox": [ + 253, + 258, + 705, + 292 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 325, + 313, + 343 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Pre-training a vision-language model and then fine-tuning it on downstream tasks have become a popular paradigm. However, pre-trained vision-language models with the Transformer architecture usually take long inference time. Knowledge distillation has been an efficient technique to transfer the capability of a large model to a small one while maintaining the accuracy, which has achieved remarkable success in natural language processing. However, it faces many problems when applying KD to the multi-modality applications. In this paper, we propose a novel knowledge distillation method, named CLIPPING1, where the plentiful knowledge of a large teacher model that has been fine-tuned for video-language tasks with the powerful pre-trained CLIP can be effectively transferred to a small student only at the fine-tuning stage. Especially, a new layer-wise alignment with the student as the base is proposed for knowledge distillation of the intermediate layers in CLIPPING, which enables the student's layers to be the bases of the teacher, and thus allows the student to fully absorb the knowledge of the teacher. CLIPPING with MobileViT-v2 as the vision encoder without any vision-language pre-training achieves $88.1\\% - 95.3\\%$ of the performance of its teacher on three video-language retrieval benchmarks, with its vision encoder being $19.5x$ smaller. CLIPPING also significantly outperforms a state-of-the-art small baseline (ALL-in-one-B) on the MSR-VTT dataset, obtaining relatively $7.4\\%$ performance gain, with $29\\%$ fewer parameters and $86.9\\%$ fewer flops. Moreover, CLIPPING is comparable or even superior to many large pre-training models.", + "bbox": [ + 75, + 366, + 472, + 819 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 501, + 325, + 630, + 343 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recently, pre-training a vision-language model and then fine-tuning it on downstream tasks are a popular paradigm [12, 16, 20, 21, 30]. Pre-trained vision-language models (PVLMs) have achieved great success in many multi-modality tasks (e.g., video-text retrieval). However, PVLMs with the Transformer architecture (especially the vision stream) usually consume a huge amount of computation, making them difficult to be deployed on edge devices such as mobile phones. Recent small models [8, 14, 19, 22, 27, 40] show that combining Convolutional Neural Networks (CNNs) and Transformers as a hybrid architecture gets the best of both architectures, but the overall performance of these works is still far from satisfactory when compared to large pre-training models. Apparently, knowledge distillation (KD) [15] is an efficient technique to transfer the capability of a large model to a small one while maintaining the accuracy, which has achieved remarkable success in natural language processing (NLP) [17, 18]. In NLP, the knowledge distillation methods are usually performed at both the pre-training and the fine-tuning stages. However, the collection of the pre-training data for the pre-training knowledge distillation cost huge manpower in multi-modality applications (e.g., the pre-training dataset of CLIP is hard to obtain). Therefore, an efficient knowledge distillation method need to be explored for multi-modality applications. To this end, we propose a novel knowledge distillation method, named CLIPPING, where the plentiful knowledge of a large teacher model Clip4clip [26] that has been fine-tuned for video-language tasks with the powerful pre-trained CLIP [30] can be effectively transferred to a small student only at the fine-tuning stage. The contributions of CLIPPING are summarized below:", + "bbox": [ + 496, + 354, + 893, + 835 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1) We introduce an efficient approach to distill both the vision knowledge and the cross-modality knowledge from teacher to a small model at the fine-tuning stage. The resulting model shows strong performance on multi-modality", + "bbox": [ + 498, + 839, + 893, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "1In this paper, CLIPPING means cutting something to make it smaller through distilling.", + "bbox": [ + 75, + 875, + 470, + 901 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "18983", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/8097da16ed75c627f4843d5d3aeff5d1ba8aa878473aad9c97b44721c8dd0c8b.jpg", + "image_caption": [ + "(a)" + ], + "image_footnote": [], + "bbox": [ + 78, + 90, + 207, + 199 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/da544192d760903d94e60854f38c3d572730c43c338e865beb70fbb1bd558907.jpg", + "image_caption": [ + "(b)" + ], + "image_footnote": [], + "bbox": [ + 210, + 90, + 344, + 199 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/bfeb05a90cd2bf07c30252fef956f0e9ec8536a03e4fe8300eca4762e94cad2a.jpg", + "image_caption": [ + "(c)", + "Figure 1. (a) and (b) are previous methods of intermediate features' knowledge distillation. $L_{j}^{T}$ , $j = 1,2,\\dots,M$ , and $L_{i}^{S}$ , $i = 1,2,\\dots,N$ , are the outputs of the $j^{th}$ and $i^{th}$ layers of the teacher and the student, respectively. OL2OL is the most common One-Layer-to-One-Layer KD, where it selects some layers of the teacher and then distills the layers into the student one by one (e.g., from $L_{j}^{T}$ to $L_{i}^{S}$ ). Teacher-As-Base (TAB) KD uses all the teacher's layers to supervise each layer in the student, where each layer of the student (e.g., $L_{i}^{S}$ ) learns the knowledge from all the selected layers of the teacher $\\left(\\sum_{k=1}^{M} w_{ik}' L_{k}^{T}\\right)$ , where $w_{ik}'$ , $k = 1,2,\\dots,M$ , are the knowledge selection weights with $\\sum_{k=1}^{M} w_{ik}' = 1$ . (c) Our Student-As-Base (SAB) KD enables each of the teacher's layers to pass its knowledge to all the student's layers (e.g., $L_{j}^{T} \\rightarrow \\sum_{k=1}^{N} w_{jk} L_{k}^{S}$ , $\\sum_{k=1}^{N} w_{jk} = 1$ ). More detailed analysis is provided in Section 3.1." + ], + "image_footnote": [], + "bbox": [ + 346, + 90, + 470, + 199 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "text-video retrieval task.", + "bbox": [ + 76, + 446, + 238, + 460 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "2) We propose a new layer-wise alignment scheme, called Student-As-Base (SAB), for knowledge distillation of the intermediate layers from the Transformer to the CNN in CLIPPING, where the student's layers can be regarded as the bases of the teacher's feature space, forcing the student model to full absorb the knowledge of the teacher. In our experience, the SAB layer-wise alignment significantly surpasses the previous knowledge distillation methods. We believe it will become a popular paradigm for knowledge distillation of intermediate features.", + "3) We present an effective cross-modal knowledge distillation, which includes knowledge from both the global and local video-caption distributions. We use the video-caption distributions of the teacher to guide the training of the student. Besides, the student can also benefit from the powerful pre-trained CLIP that obtains certain local frame-word attention ability via learning from massive data. Experiments show that this pre-training knowledge can be effectively transferred to the student when jointly trained with our SAB KD only at the fine-tuning stage.", + "4) CLIPPING with MobileViT-v2 [27] as the vision encoder without any vision-language pre-training achieves $91.5\\% - 95.3\\%$ of the performance of its teacher on MSR-VTT video-language retrieval benchmark, with its vision encoder being $19.5 \\times$ smaller. CLIPPING also significantly outperforms a state-of-the-art small baseline (ALL-in-one-B) on the MSR-VTT dataset, obtaining relatively $7.4\\%$ performance gain, with $29\\%$ fewer parameters and $86.9\\%$ fewer flops. Moreover, CLIPPING is comparable or even" + ], + "bbox": [ + 75, + 462, + 468, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "superior to many large pre-training models.", + "bbox": [ + 500, + 90, + 787, + 106 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Relative Work", + "text_level": 1, + "bbox": [ + 500, + 123, + 645, + 138 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Vision-Language Retrieval. Learning from web-collected image-text data, large-scale Vision-Language Pretraining (VLP) models such as CLIP [30] have recently demonstrated great success across various downstream tasks. Nowadays, models such as Clip4clip [26] and MDMMT [11] extended from the pre-trained model CLIP keep appearing for vision-language retrieval. There are also some end-to-end trainable models [3, 37], which are designed to take advantage of both large-scale image and video captioning datasets. All-in-one [34] is the first work to consider both efficiency and performance for video-language retrieval tasks. It introduces a unified backbone that enables the representation learning of both video-text multimodal and unimodal inputs.", + "bbox": [ + 496, + 142, + 890, + 354 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Knowledge Distillation. Knowledge distillation (KD) [15] is a simple yet effective technique to improve the performance of a learning model. Earlier works transfer knowledge embedded in the \"logits\" learned in a large teacher model to a small student model without sacrificing much performance. Recent works [5, 6] use multiple layers of the teacher to supervise each layer in the student, where each layer of the student learns the knowledge from multiple layers of the teacher. [23] proposes a target-aware Transformer and enables the student to mimic each spatial component of the teacher in each distilled layer to boost the student's performance. For multi-modality KD, [36] designs a fusion-encoder model as the teacher and introduces cross-modal attention knowledge to train the dual-encoder student model. The distillation objective is applied at both the pre-training and the fine-tuning stages and helps the dual-encoder model learn interactions of different modalities. TinyBert [17] also introduces a two-stage learning framework that performs Transformer distillation at both the pre-training and the task-specific fine-tuning stages. In this paper, we also focus on KD for transferring the knowledge of a pre-training vision-language model to a small one but only at the fine-tuning stage.", + "bbox": [ + 496, + 354, + 892, + 702 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Transformers and CNNs. Usually, the pre-training model is a Transformer-based model [10, 25] and the small model is a CNN-based or hybrid architecture [19, 27]. For KD, most existing methods distill knowledge either from a Transformer to another Transformer (T2T) or from a CNN to another CNN (C2C) that computes the loss in the OL2OL or TAB style (see Fig. 1). To the best of our knowledge, KD from a Transformer to a CNN (T2C) has not been explored yet. The most related work [7] distills knowledge from a CNN to a Transformer (C2T) in the OL2OL way. However, previous works (e.g., [31]) have investigated the internal representation structures of ViTs and CNNs, and found striking differences between the two models, such as", + "bbox": [ + 496, + 703, + 892, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "18984", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/9d827e96f3b031b4aa6ccfb205aa5f203433525ff93a0a1ad43f822c41195bae.jpg", + "image_caption": [ + "Figure 2. Overview of CLIPPING. There are mainly four knowledge distillation (KD) parts: (1) Temporal KD. (2) Spatial KD. (3) SAB (Student-As-Base) KD. (4) Cross-modality KD. The CLIP's vision encoder is a Transformer. The temporal Transformer is an optional module for CLIPPING." + ], + "image_footnote": [], + "bbox": [ + 240, + 85, + 718, + 366 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "ViTs having highly similar representations throughout the model's layers, while CNNs showing obvious distinction of representations between lower and higher layers [31]. Considering such striking differences between ViTs and CNNs, the previous KD (OL2OL and TAB) may not be good distillation ways for our T2C task, which is verified by our experiments.", + "bbox": [ + 75, + 438, + 472, + 544 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Methodology", + "text_level": 1, + "bbox": [ + 76, + 556, + 212, + 574 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We propose an efficient approach to distill both the vision knowledge and the cross-modality knowledge from Clip4clip to a small model with MobileViT-v2 as its vision encoder for text-video retrieval task, which can get the best from both sides: powerful multimodal representations of CLIP and efficient mobile vision Transformer of MobileViT-v2. The overall architecture is illustrated in Fig. 2.", + "bbox": [ + 75, + 583, + 468, + 704 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Analysis of TAB and SAB", + "text_level": 1, + "bbox": [ + 76, + 712, + 310, + 728 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We first show the advantage of our SAB KD over traditional TAB KD. For simplicity, suppose $M = N = 2$ . Then, for SAB,", + "bbox": [ + 75, + 734, + 468, + 781 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\left\\{ \\begin{array}{l} L _ {1} ^ {T} = w _ {1 1} L _ {1} ^ {S} + w _ {1 2} L _ {2} ^ {S}, w _ {1 1} + w _ {1 2} = 1 \\\\ L _ {2} ^ {T} = w _ {2 1} L _ {1} ^ {S} + w _ {2 2} L _ {2} ^ {S}, w _ {2 1} + w _ {2 2} = 1 \\end{array} , \\right. \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 117, + 790, + 468, + 830 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "and for TAB,", + "bbox": [ + 76, + 839, + 166, + 853 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\left\\{ \\begin{array}{l} L _ {1} ^ {S} = w _ {1 1} ^ {\\prime} L _ {1} ^ {T} + w _ {1 2} ^ {\\prime} L _ {2} ^ {T}, w _ {1 1} ^ {\\prime} + w _ {1 2} ^ {\\prime} = 1 \\\\ L _ {2} ^ {S} = w _ {2 1} ^ {\\prime} L _ {1} ^ {T} + w _ {2 2} ^ {\\prime} L _ {2} ^ {T}, w _ {2 1} ^ {\\prime} + w _ {2 2} ^ {\\prime} = 1 \\end{array} \\right. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 117, + 863, + 468, + 904 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Obviously, in Eq. 1, the knowledge of the teacher $L_1^T$ or $L_2^T$ is the linear combination of $L_1^S$ and $L_2^S$ . In other words, $L_1^S$ and $L_2^S$ are the bases of $L_1^T$ and $L_2^T$ of the teacher. In Eq. 2, it can be seen that the teacher's $L_1^T$ and $L_2^T$ are as the bases for $L_1^S$ and $L_2^S$ . Eq. 2 can be converted to:", + "bbox": [ + 500, + 438, + 890, + 513 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nL _ {1} ^ {T} = A \\cdot L _ {1} ^ {S} + B \\cdot L _ {2} ^ {S}, L _ {2} ^ {T} = C \\cdot L _ {1} ^ {S} + D \\cdot L _ {2} ^ {S}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 519, + 522, + 890, + 542 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $A = \\frac{w_{22}^{\\prime}}{w_{11}^{\\prime}w_{22}^{\\prime} - w_{12}^{\\prime}w_{21}^{\\prime}}$ , $B = \\frac{-w_{12}^{\\prime}}{w_{11}^{\\prime}w_{22}^{\\prime} - w_{12}^{\\prime}w_{21}^{\\prime}}$ , $C =$", + "bbox": [ + 498, + 551, + 890, + 575 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "$\\frac{-w_{21}^{\\prime}}{w_{11}^{\\prime}w_{22}^{\\prime} - w_{12}^{\\prime}w_{21}^{\\prime}}$ and $D = \\frac{w_{11}^{\\prime}}{w_{11}^{\\prime}w_{22}^{\\prime} - w_{12}^{\\prime}w_{21}^{\\prime}}$ , which looks like Eq. 1. It seems that in this case, the student's $L_1^S$ and $L_2^S$ could also serve as the bases for $L_1^T$ and $L_2^T$ . However, (i) $A + B \\neq 1$ and $C + D \\neq 1$ , but it is usually required in the previous KD that all the coefficients in each equation should be summed to be 1 [5,6,23]; (ii) when learning with TAB KD in Eq. 2, the coefficients A, B, C and D are uncontrollable, which may be arbitrary large during learning.", + "bbox": [ + 498, + 574, + 890, + 700 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "For the reasons above, although it seems that SAB and TAB have the same matrix representation, TAB is unable to get a similar result as SAB and more likely to get to a local minimum during training. Therefore, TAB KD is not equivalent to our SAB KD. When well trained, the student with this SAB KD fully learns the knowledge of the teacher because the teacher's $L_{i}^{T}$ can be obtained by the linear combination of the student's $L_{i}^{S}$ , $i = 1,2,\\dots,N$ . Our experiment also shows this property (see Section 4.2).", + "bbox": [ + 496, + 702, + 890, + 837 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Preliminaries", + "text_level": 1, + "bbox": [ + 500, + 847, + 638, + 861 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Given a batch of videos $V$ and captions $T$ , Clip4clip learns a similarity function $s(v_{i},t_{j})$ to calculate the similar-", + "bbox": [ + 500, + 869, + 890, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "18985", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/d0412b3b6617f2558684dd17b058443e37462ad29a8cf2125df57a90e1172e52.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 178, + 85, + 795, + 202 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/b3485e9995b2dc2b7f0637287b4d2b50eb2976604aabbc3a916d59d2ef28b4f3.jpg", + "image_caption": [ + "Figure 3. Reshaping the features of (a) the student $L_{k}^{S}$ and (b) the teacher $L_{k}^{T}$ to the same shape. (c) Similarity Computation." + ], + "image_footnote": [], + "bbox": [ + 178, + 203, + 792, + 327 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "ity between a video $v_{i} \\in V$ and a caption $t_j \\in T$ . It obtains the frames' representation and the caption representation in a multi-modal embedding space via CLIP. The frames' representation is denoted as $Z_{i} = \\{z_{i1}, z_{i2}, \\dots, z_{in}\\}$ , where $n$ is the number of frames in $v_{i}$ and $z_{ik}$ ( $k = 1, 2, \\dots, n$ ) is the class token from the output of the CLIP's vision encoder corresponding to the $k^{th}$ input frame. The caption representation is denoted as $W_{j} = \\{w_{j1}, w_{j2}, \\dots, w_{jm}\\}$ , where $m$ is the number of words in $t_j$ (including [CLS] and [SEP]) and $w_{jm}$ is used as the representation of $t_j$ in Clip4clip. Clip4clip uses a temporal Transformer encoder to obtain the sequential feature, denoted as $Z_{i}' = \\{z_{i1}', z_{i2}', \\dots, z_{in}'\\} = TemporalTransformer(Z_{i})$ . Then, the mean pooling is used to aggregate the features of all frames to obtain the video representation, $z_{i} = \\frac{1}{n} \\sum_{k=1}^{n} z_{ik}'$ . Finally, the similarity functions $s(v_i, t_j)$ and $s(t_j, v_i)$ are defined as:", + "bbox": [ + 75, + 373, + 472, + 617 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\ns (v _ {i}, t _ {j}) = w _ {j m} ^ {\\top} z _ {i}, \\quad s (t _ {i}, v _ {j}) = z _ {j} ^ {\\top} w _ {i m}. \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 138, + 619, + 468, + 640 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Nowadays, models such as Clip4clip based on the pretrained model CLIP have been applied to many tasks. They pursue better performance with CLIP but have a heavy vision encoder, which makes them difficult to be deployed on edge devices such as mobile phones. Since MobileViT-v2 [27] is a light-weight and mobile-friendly hybrid network, we employ it as the vision encoder of the student model and maintain the original text encoder and temporal Transformer of Clip4clip. Note that the temporal Transformer is an optional module for the CLIPPING model structure, and the text encoder $(CLIP_{bert})$ can be further compressed through the existing methods (details are shown in the Section 4).", + "bbox": [ + 75, + 643, + 472, + 840 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. CLIPPING", + "text_level": 1, + "bbox": [ + 76, + 847, + 202, + 861 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To distill multimodal knowledge from Clip4clip to the MobileViT-v2-based model, we use four kinds of knowl-", + "bbox": [ + 76, + 869, + 468, + 898 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "edge transfer: 1) temporal KD, 2) spatial KD, 3) SAB KD, and 4) cross-modality KD.", + "bbox": [ + 498, + 373, + 890, + 402 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3.1 Temporal and Spatial Knowledge Distillation", + "text_level": 1, + "bbox": [ + 498, + 421, + 867, + 436 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The temporal KD is motivated by the previous finding that aligning multi-frame dependency from the teacher to the student can enhance the performance of the student [24], which encodes the multi-frame dependency into a latent embedding by using a recurrent unit ConvLSTM. In our architecture, multi-frame dependency can be modeled naturally through the temporal Transformer (Temporal Transformer) in Fig. 2, so we define the temporal KD loss as:", + "bbox": [ + 496, + 444, + 890, + 566 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nL _ {T K D} = \\frac {1}{B} \\sum_ {i = 1} ^ {B} D _ {K L} \\left(Z _ {i} ^ {\\prime S}, Z _ {i} ^ {\\prime T}\\right), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 580, + 574, + 890, + 614 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $B$ is the batch size, the superscripts $S$ and $T$ denote the student and the teacher, respectively, and $D_{KL}()$ is the KL divergence loss function. In addition to imitating the behavior of CLIP for each frame, we use spatial KD to align the spatial embeddings of the student and the teacher by:", + "bbox": [ + 496, + 623, + 890, + 699 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nL _ {S K D} = \\frac {1}{B} \\sum_ {i = 1} ^ {B} D _ {K L} \\left(Z _ {i} ^ {S}, Z _ {i} ^ {T}\\right). \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 584, + 707, + 890, + 746 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The Temporal Transformer is an optional module for the CLIPPING model structure in our method. In this case, $Z_{i}^{\\prime S}$ is exactly equal to $Z_{i}^{S}$ in TKD.", + "bbox": [ + 496, + 746, + 890, + 792 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3.2 SAB Knowledge Distillation", + "text_level": 1, + "bbox": [ + 500, + 806, + 748, + 821 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Recently, it is discovered that distilling intermediate features is more effective. So in addition to the spatial and temporal KD, we also force the vision encoder of the student to mimic the intermediate layers of the teacher. In the traditional OL2OL KD, some layers of the teacher are selected", + "bbox": [ + 496, + 824, + 890, + 898 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "18986", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/5aaf3acfbac4445d947fd6ee4f54a65e2b4658a636631e79e200eeb3dbdc4c65.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 240, + 85, + 415, + 209 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/8e174f102370c416bcd2396d5ae76916e10d293ca87bd242400739efdd6d36ea.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 418, + 85, + 730, + 210 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/b671099dc4d7918037e0e99b464aab4635924d5e4a39a59575ec8ba9b240699c.jpg", + "image_caption": [ + "Figure 4. (a) Masks obtained with Eq. 7, where all the $L - 1$ masks along the $k$ dimension are the same. (b) Masking the similarity tensor $W$ . (c) SAB alignment. $L_{SAB}$ is the SAB loss." + ], + "image_footnote": [], + "bbox": [ + 240, + 212, + 730, + 333 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "and distilled into the student one by one [17]. However, it is difficult to find the optimal correspondence between the student and the teacher. The recent TAB works (e.g., [5]) still cannot help the student learn enough knowledge from the teacher when there is a large architecture gap between them, which is verified in our experiments. For our task, the student (MobileViT-v2) is a hybrid architecture, which has convolution layers in the shallow stages and separable self-attention layers in the later blocks. And a separable self-attention layer differs significantly from the traditional self-attention in two ways: 1) it does not learn an explicit attention map; 2) its input and output are still 3D, which are more similar to CNN features. Except for the separable self-attention in the later blocks, the shallow stages with convolution layers undoubtedly have a huge difference from the Transformer structure in CLIP, which has been investigated in previous works [31,41]. Therefore, we design SAB layer-wise alignment for KD of the intermediate layers from the Transformer to the CNN.", + "bbox": [ + 75, + 392, + 472, + 680 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "SAB Property. As shown in Fig. 2, the student's layers in our SAB KD can be explained as the bases of the feature space, and each layer of the teacher is a linear combination of the bases. After training, if the student's layers do ideally form the bases, the knowledge of the teacher's layers is learned completely by the student. Our experiment in Section 4.2 shows that the teacher's features in different layers can be well recovered from the features of the student's layers. Next, we describe how to implement SAB KD.", + "bbox": [ + 75, + 685, + 468, + 821 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Feature Reshaping and Similarity Computation. Let the layers of the teacher and the student be $\\text{Layer}^T = [L_1^T, L_2^T, \\ldots, L_M^T]$ and $\\text{Layer}^S = [L_1^S, L_2^S, \\ldots, L_N^S]$ (usually $M > N$ ), respectively. $L_k^T \\in R^{D \\times L}$ is a 2D tensor, where $D$ is the dimensionality of the token feature", + "bbox": [ + 76, + 824, + 470, + 902 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "and $L$ is the number of tokens, while $L_{k}^{S}\\in R^{C_{k}\\times H_{k}\\times W_{k}}$ is a 3D tensor, where $C_k$ , $H_{k}$ and $W_{k}$ are the channel number, height and width of the $k^{th}$ CNN layer's feature. We calculate a similarity tensor $W$ between $L_{k}^{S}$ and $L_{k}^{T}$ which need to be reshaped first as shown in Figs. 3(a) and (b), respectively. For $L_{k}^{S}$ , we first apply a learnable linear operator and then do pixel reshuffle on the result, obtaining $\\bar{L}_k^S$ , $k = 1,2,\\dots,N$ . Next, $\\bar{L}_k^S$ is reshaped to $\\tilde{L}_k^S$ . As for $L_{k}^{T}$ , we remove the class token and obtain a $(L - 1)\\times D$ tensor, denoted as $\\tilde{L}_k^T$ , $k = 1,2,\\dots,M$ . Finally, all $\\tilde{L}_k^S$ , $k = 1,2,\\dots,N$ , and $\\tilde{L}_k^T$ , $k = 1,2,\\dots,M$ , are represented as $\\tilde{L}_{\\mathrm{er}}^{S} = [\\tilde{L}_{1}^{S},\\tilde{L}_{2}^{S},\\dots,\\tilde{L}_{N}^{S}]$ and $\\tilde{L}_{\\mathrm{er}}^{T} = [\\tilde{L}_{1}^{T},\\tilde{L}_{2}^{T},\\dots,\\tilde{L}_{M}^{T}]$ , respectively. After this reshaping, as shown in Fig. 3(c), we are able to measure the similarities between the intermediate layers of the student and the teacher through $W = \\tilde{L}_{\\mathrm{er}}^{T}\\times (\\tilde{L}_{\\mathrm{er}}^{S})^{\\top}$ , $W\\in$ $R^{(L - 1)\\times M\\times N}$ .", + "bbox": [ + 496, + 392, + 893, + 652 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Masking and SAB Alignment. To speed up the training procedure, we design SAB KD with sequential masks, which follow this formula:", + "bbox": [ + 496, + 654, + 893, + 696 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {m a s k} _ {(i, j)} = \\left\\{ \\begin{array}{l} 0, \\text {i f} j > n _ {0} \\text {a n d} i \\leq m _ {0} \\\\ 0, \\text {i f} j > n _ {1} \\text {a n d} m _ {0} < i \\leq m _ {1} \\\\ 1, \\text {e l s e} \\end{array} , \\right. \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 537, + 704, + 890, + 755 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $i = 1,2,\\dots,M$ , $j = 1,2,\\dots,N$ , $1\\leq m_0 < m_1 < M$ , $1\\leq n_0 < n_1 < N$ . And the final attention mask $\\mathsf{Mask}\\in R^{(L - 1)\\times M\\times N}$ is composed of $L - 1$ same masks, Fig. 4(a) shows one example with $m_0 = 1,n_0 = 2,m_1 = 3,n_1 = 3$ . $\\mathsf{mask}_{(i,j)} = 0$ means to mask all the elements along the $k$ dimension (Fig. 4(a)) at $(i,j)$ in $W$ , while $\\mathsf{mask}_{(i,j)} = 1$ means to maintain their similarities, as shown in Fig. 4(b). This masking is based on our experimental finding: The student's lower-level layers", + "bbox": [ + 496, + 763, + 893, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "18987", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "should learn from the teacher's lower-layers, while the student's higher layers should learn from all the teacher's layers. In Fig. 4(b), we also normalize the masks by performing softmax along the $j$ dimension on each row, obtaining $\\tilde{W} = \\sigma (W\\circ \\mathsf{Mask})$ , where $\\sigma$ is the softmax function and $\\circ$ denotes the element-wise product. Guided by $\\tilde{W}$ , as shown in Fig. 4(c), the weighted student layers are calculated as: $\\tilde{L a r e r}_{w}^{S} = \\tilde{W}\\times \\tilde{L a r e r}^{S}$ . Each component of $\\tilde{L a r e r}_{w}^{S}$ is the linear combination of $N$ student layers' features. Finally, the SAB KD loss is defined as:", + "bbox": [ + 76, + 90, + 472, + 242 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nL _ {S A B} = D _ {K L} \\left(\\tilde {L a y e r} _ {w} ^ {S}, \\tilde {L a y e r} ^ {T}\\right). \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 155, + 251, + 468, + 271 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.3.3 Cross-Modality Knowledge distillation", + "text_level": 1, + "bbox": [ + 76, + 294, + 400, + 309 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To further adapt the student's features to the teacher's multimodal feature space, we employ the teacher's knowledge of cross-modality distributions to guide the training of the student. Specifically, we characterize the cross-modal distributions from two perspectives, global video-caption distributions of the teacher and local video-caption distributions of CLIP.", + "bbox": [ + 75, + 316, + 468, + 421 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Global Video-Caption Distribution Alignment. We consider both the video-to-caption distribution $A_{GVC}$ and the caption-to-video distribution $A_{GCV}$ , the elements $(s(v_i, t_j)$ and $s(t_j, v_i))$ of which are defined via the similarities obtained by Eq. 4:", + "bbox": [ + 76, + 422, + 468, + 494 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nA _ {G V C} = \\sigma \\left( \\begin{array}{c c c} s (v _ {1}, t _ {1}) & ... & s (v _ {1}, t _ {B}) \\\\ ... & ... & ... \\\\ s (v _ {B}, t _ {1}) & ... & s (v _ {B}, t _ {j}) \\end{array} \\right),\n$$\n", + "text_format": "latex", + "bbox": [ + 135, + 501, + 401, + 546 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nA _ {G C V} = \\sigma \\left( \\begin{array}{c c c} s (t _ {1}, v _ {1}) & \\dots & s (t _ {1}, v _ {B}) \\\\ \\dots & \\dots & \\dots \\\\ s (t _ {B}, v _ {1}) & \\dots & s (t _ {B}, v _ {B}) \\end{array} \\right). \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 138, + 542, + 467, + 592 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": ", where $\\sigma$ is the softmax function along the first dimension of the similarity matrixs. Finally, the global video-caption distribution alignment loss is defined as:", + "bbox": [ + 76, + 599, + 468, + 643 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nL _ {G} = D _ {K L} \\left(A _ {G V C} ^ {S}, A _ {G V C} ^ {T}\\right) + D _ {K L} \\left(A _ {G C V} ^ {S}, A _ {G C V} ^ {T}\\right). \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 94, + 650, + 468, + 667 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Local Video-Caption Distribution Alignment. CLIP uses text prompts (such as \"A picture of a (\") for zero-shot image classification. CLIP fills them with different words (e.g., \"cat\" and \"dog\") and results in different captions (e.g., \"A picture of a cat\" and \"A picture of a dog\"). It can match the captions to the corresponding images, showing some image-word alignment ability. We transfer this pre-training knowledge to the student through a local frame-word alignment as follows. Recall that $Z_{i}^{\\prime} = \\{z_{i1}^{\\prime}, z_{i2}^{\\prime}, \\dots, z_{in}^{\\prime}\\}$ and $W_{j} = \\{w_{j1}, w_{j2}, \\dots, w_{jm}\\}$ are the features of the video $v_{i}$ and the caption $t_{j}$ , respectively, with $n$ being the number of frames in $v_{i}$ and $m$ the number of words in $t_{j}$ (Section 3.1). The similarity between the $k^{th}$ frame and the $r^{th}$ word is defined as $s_{fw}(w_{jr}, z_{ik}') = (w_{jr})^\\top z_{ik}'$ . Then we respectively define the local video-to-caption and caption-to-video similarities as:", + "bbox": [ + 76, + 676, + 470, + 900 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\ns ^ {\\prime} \\left(v _ {i}, t _ {j}\\right) = \\frac {1}{n} \\sum_ {k = 1} ^ {n} \\max _ {1 \\leq r \\leq m} \\left\\{s _ {f w} \\left(w _ {j r}, z _ {i k} ^ {\\prime}\\right) \\right\\}, \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 562, + 89, + 890, + 130 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\ns ^ {\\prime} (t _ {j}, v _ {i}) = \\frac {1}{m} \\sum_ {r = 1} ^ {m} \\max _ {1 \\leq k \\leq n} \\left\\{s _ {f w} \\left(z _ {i k} ^ {\\prime}, w _ {j r}\\right) \\right\\}.\n$$\n", + "text_format": "latex", + "bbox": [ + 562, + 127, + 828, + 161 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Finally, we have the local video-to-caption distribution $A_{LVC}$ and the local caption-to-video distribution $A_{LCV}$ :", + "bbox": [ + 500, + 169, + 890, + 199 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nA _ {L V C} = \\sigma \\left( \\begin{array}{c c c} s ^ {\\prime} \\left(v _ {1}, t _ {1}\\right) & \\dots & s ^ {\\prime} \\left(v _ {1}, t _ {B}\\right) \\\\ \\dots & \\dots & \\dots \\\\ s ^ {\\prime} \\left(v _ {B}, t _ {1}\\right) & \\dots & s ^ {\\prime} \\left(v _ {B}, t _ {B}\\right) \\end{array} \\right), \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 553, + 207, + 890, + 258 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nA _ {L C V} = \\sigma \\left( \\begin{array}{c c c} s ^ {\\prime} (t _ {1}, v _ {1}) & ... & s ^ {\\prime} (t _ {1}, v _ {B}) \\\\ ... & ... & ... \\\\ s ^ {\\prime} (t _ {B}, v _ {1}) & ... & s ^ {\\prime} (t _ {B}, v _ {B}) \\end{array} \\right),\n$$\n", + "text_format": "latex", + "bbox": [ + 562, + 253, + 836, + 297 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "and the local caption-video distribution alignment loss is defined as:", + "bbox": [ + 500, + 297, + 890, + 324 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nL _ {L} = D _ {K L} \\left(A _ {L C V} ^ {S}, A _ {L C V} ^ {T}\\right) + D _ {K L} \\left(A _ {L V C} ^ {S}, A _ {L V C} ^ {T}\\right). \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 519, + 325, + 890, + 343 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Combining $L_{G}$ and $L_{V}$ , the cross-modality KD loss is:", + "bbox": [ + 501, + 348, + 867, + 363 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nL _ {C M} = L _ {G} + L _ {L}. \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 630, + 367, + 890, + 383 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Our experiment in Section 4.2 shows that only when joint trained with SAB, our local video-caption distribution alignment shows its advantage, which also verifies the effectiveness of SAB. Finally, the total loss for training our model is:", + "bbox": [ + 496, + 388, + 890, + 463 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} L = L _ {\\text {t a s k}} + \\alpha \\cdot L _ {T K D} + \\beta \\cdot L _ {S K D} \\tag {15} \\\\ + \\gamma \\cdot L _ {C M} + \\delta \\cdot L _ {S A B}, \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 571, + 465, + 890, + 501 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $L_{task}$ is the task-specifical loss, and $\\alpha ,\\beta ,\\gamma$ and $\\delta$ are loss balance weights.", + "bbox": [ + 498, + 512, + 890, + 542 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 500, + 556, + 633, + 571 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We conduct comprehensive experiments on three benchmarks for video-text retrieval (video-to-text $(v2t)$ and text-to-video $(t2v)$ ): MSR-VTT [38], MSVD [4] and LSMDC [32]. The metrics Recall at rank K $(R@K)$ are used for evaluation.", + "bbox": [ + 498, + 580, + 890, + 655 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Comparison with State-of-the-Arts", + "text_level": 1, + "bbox": [ + 498, + 666, + 803, + 681 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "MSR-VTT. In Table 1, we compare the proposed model CLIPPING with eight state-of-the-art methods on MSR-VTT: TACo [39], VideoClip [37], Frozen [3], VIOLET [12], OA-Trans [35], BridgeFormer [13], ALL-in-one [34] and MDMMT [11]. It can be seen that CLIPPING significantly surpasses those large-scale video-text/image-text pre-training models for video-text retrieval. For example, our model exceeds VideoClip and Frozen by absolute $9.8\\%$ t2vR@1 and $8.2\\%$ t2vR@1, respectively. Note that the model MDMMT uses CLIP as its backbone, while our CLIPPING uses CLIP as the teacher. For a more efficient and edge deployable video-language retrieval model, we further compress CLIPPING into two models: $\\mathrm{CLIPPING}_{\\mathrm{W / oT}}$ and $\\mathrm{CLIPPING^{*}}$ . $\\mathrm{CLIPPING}_{\\mathrm{W / oT}}$", + "bbox": [ + 496, + 688, + 890, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "18988", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/739f78667240eaaf71dd20d5aa0773dc4969f4ca6ab6104395d9417add1981e3.jpg", + "table_caption": [], + "table_footnote": [ + "Table 1. Comparison with state-of-the-art models on MSR-VTT (1k split) for text-to-video retrieval. \"PT Datasets\": datasets used for pre-training the vision encoder. \"HT100M\": HowTo100M dataset [28]. \"C400M\": CLIP-400M dataset [30]. \"IN21K\": ImageNet21K dataset [9]. \"W2M\": WebVid-2M dataset [3]. \"C3M\": CC3M dataset [33]. All the methods in Table are the comparable methods with parameters less than 240M and flops less than 360G." + ], + "table_body": "
ModelPT DatasetsParamsR@1
TACoHT100M212M28.4
VideoClipHT100M130M30.9
FrozenC3M,W2M,COCO232M32.5
ALL-in-one-SW2M,HT100M33M33.5
VIOLETC3M,W2M198M34.5
OA-TransC3M,W2M232M35.8
BridgeFormerC3M,W2M160M37.6
ALL-in-one-BW2M,HT100M110M37.9
MDMMTC400M,AudioSet226M38.9
CLIPPING*w/o T-8.7M37.5
CLIPPINGw/o T-21.3M38.6
CLIPPING*-46.5M39.8
CLIPPING-55.0M40.6
CLIPPINGIN21K55.0M40.7
", + "bbox": [ + 93, + 88, + 450, + 280 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "does not use the temporal Transformer and CLIPPING* compresses the text encoder through TinyBERT [17]. It can be seen from Table 1 that the CLIPPING* based models also outperform the small model ALL-in-one-S even though they are smaller. In Fig. 5, we show the performances and Flops of these models. Among previous models, MDMMT has the best performance and ALL-in-one-S is the fastest. CLIPPING not only obtains relative $4.6\\%$ performance gain over MDMMT but also is faster than All-in-one-S. Table 2 provides the details of Flops and parameters for each module of our method.", + "bbox": [ + 75, + 401, + 470, + 566 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f021df41a4d4a96d5066586f05fea5c73538a85e41bd2b20fad19379d7217281.jpg", + "image_caption": [ + "Figure 5. Flops and Performances." + ], + "image_footnote": [], + "bbox": [ + 133, + 574, + 418, + 694 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Other Benchmarks. We also compare our CLIPPING with the state-of-the-art Frozen [3], MDMMT [11], NoiseEst [2] and SupportSet [29] on the LSMDC and MSVD datasets. In Table 3, CLIPPING again obtains the best performance with fewest parameters and Flops, which achieves $88.1\\% - 92.9\\%$ of the performance of its teacher on the LSMDC and MSVD datasets, showing consistent improvements and generalization across different datasets.", + "bbox": [ + 75, + 717, + 468, + 839 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. Ablation Study", + "text_level": 1, + "bbox": [ + 76, + 847, + 230, + 863 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Key Components. We provide detailed ablation study to validate each key component of our proposed method,", + "bbox": [ + 76, + 869, + 468, + 900 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/9e29ff45209e679dcc75f85e246ff9570118db5371ea61fb06a3ba4c0d366e08.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModulesParametersFlops
Vision encoder (MobileViT-v2)4.5M16.8G
Temporal Transformer37.8M1.2G
Language encoder (CLIPbert)12.6M0.15G
Language encoder (TinyBERT)4.2M0.05G
", + "bbox": [ + 532, + 88, + 859, + 157 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/52abb9087708028c103f08d42eacb6b4692654ec42e4a99dc2048b0dd56cc8bf.jpg", + "table_caption": [ + "Table 2. The Flops and parameters of each module in our method. We adopt 12 frames for each video." + ], + "table_footnote": [], + "table_body": "
ModelPT DatasetsVision EncodersR@1
LSMDC dataset
NoiseEstHT100MResnet152,ResNext-1016.4
SupportSetHT100MResNet-152,R(2+1)D-34IN21K
FrozenC3M,W2MSpace-Time15.0
MDMMTC400MCLIP_vision18.8
CLIPPING-MobileViT-v219.9
MSVD dataset
NoiseEstHT100MResnet152,ResNext-10120.3
SupportSetHT100MResNet-152,R(2+1)D-3428.4
FrozenC3M,W2MSpace-Time33.7
CLIPPING-MobileViT-v242.9
", + "bbox": [ + 501, + 205, + 903, + 364 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/bc1c66a24d1e60106d081e524581a5e57a5f034d132194fac1ccd06dc7e3e45c.jpg", + "table_caption": [ + "Table 3. Comparison with the state-of-the-art on the LSMDC and MSVD datasets." + ], + "table_footnote": [], + "table_body": "
Vision EncoderKD Typest2vR@1v2tR@1
CLIP_vision-44.542.2
MobileViTv2-25.724.5
MobileViTv2T28.827.3
MobileViTv2T,S33.032.8
MobileViTv2T,S,SAB37.636.2
MobileViTv2T,S,SAB,CMG39.639.1
MobileViTv2T,S,SAB,CMG,CML40.740.2
", + "bbox": [ + 501, + 406, + 890, + 513 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/c65775a9850dab166bad0cbd85906b5582a8282bef4d3d70e65442542f3c8f9f.jpg", + "table_caption": [ + "Table 4. Ablation study of different KD components of CLIPPING on the $1k$ validation set of MSR-VTT. The first row is the results of the teacher model (Clip4clip). $T$ , $S$ , $SAB$ , $CM_{G}$ and $CM_{L}$ denote temporal KD, spatial KD, SAB KD, global cross-modality KD and local cross-modality KD, respectively." + ], + "table_footnote": [], + "table_body": "
KD Typest2vR@1v2tR@1
T, S33.032.8
T, S, OL2OL34.633.4
T, S, TAB35.134.4
T, S, SABw/o masking37.135.7
T, S, SABw/masking37.636.2
", + "bbox": [ + 549, + 604, + 843, + 690 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 5. Ablation study of different KD types on MSR-VTT (1k split). All the models are trained for 36 epochs with the same setting (see the supplementary materials for details).", + "bbox": [ + 498, + 700, + 890, + 742 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "on MSR-VTT. From Table 4, we can see that without any KD, it gets extreme low accuracies. When simply adding the temporal KD and spatial KD, $t2vR@1$ and $v2tR@1$ increase significantly. When we further add SAB KD, and then global and local cross-modality KD, CLIPPING's performance rises gradually. This study shows that all these key components of CLIPPING are effective. Compared with the teacher, the full CLIPPING achieves about $91.5\\%$ and $95.3\\%$ of the performance of its teacher on $t2vR@1$ and $v2tR@1$ , respectively, with the vision encoder of $19.5x$", + "bbox": [ + 496, + 750, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "18989", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/73f8c893c779df8ba44d8439f9e1534e0ecf89a956bd2a0389867668625eb2de.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 119, + 89, + 240, + 167 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/742b81165d4b07bc807f031179c52135084b0e830e3e585e0a32424e21459974.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 240, + 90, + 336, + 167 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/33f57a089fd3cfcb94f2bb72e7648f0fc9d7a95a37217a34da9afc01ccefa7c5.jpg", + "image_caption": [ + "(a)" + ], + "image_footnote": [], + "bbox": [ + 122, + 175, + 240, + 250 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/ae87ddfc9026dc088bee4452a793c3d753df1d9cc269a3f17683edf6bb56cfbe.jpg", + "image_caption": [ + "Figure 6. (a) Feature tensors of the student and the teacher. (b) An example to demonstrate that the teacher's features are the linear combinations of the student features. More examples are provided in the supplementary materials." + ], + "image_footnote": [], + "bbox": [ + 240, + 175, + 364, + 251 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/9478f3c22d50742c4e8d4d5a73076b6d041a02a72c7f2e36d93cefb92076c799.jpg", + "image_caption": [ + "image" + ], + "image_footnote": [], + "bbox": [ + 379, + 126, + 491, + 200 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/1e533e4a9c58a15c471cb646cc25ff952aa2bd185a3aa88008b9ae8f586f779c.jpg", + "image_caption": [ + "(b)" + ], + "image_footnote": [], + "bbox": [ + 496, + 108, + 576, + 200 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "$\\widetilde{W}$ $\\tilde{L}_{er^S}$", + "bbox": [ + 524, + 220, + 578, + 229 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/dc1f322bc5138a618a9ab46be04c29d67cef79100d57a076656c02275a3985bf.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 584, + 107, + 669, + 199 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/d059336c5d5b7be3021a7959f093eb78e2e6fa1421e84c8a4f16aa8fb02d3c8c.jpg", + "image_caption": [ + "$\\widetilde{W}\\times \\widetilde{L}\\text{ayer}^{S} = \\widetilde{L}\\text{layer}_{w}^{S}\\longrightarrow \\widetilde{L}\\text{ayer}^{T}$" + ], + "image_footnote": [], + "bbox": [ + 599, + 199, + 658, + 215 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "$\\widetilde{W}$ $\\tilde{L}_{\\text{ayer}}^s$", + "bbox": [ + 524, + 220, + 578, + 229 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/c1f0481e564156a055f51cd1b7d0ae7526aa9c76b3d13ddc340c79915fcf8395.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 674, + 107, + 754, + 199 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/a0a6ca7b3952a317d5be7a57280e6aa18cd90dbd1d7daa7b2ba3dc42307283af.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 759, + 107, + 843, + 199 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/e6a8dc3f1aa2c988b6130a5cc0fd982a7777218e021808ec0604c0e020615efa.jpg", + "image_caption": [ + "(a)" + ], + "image_footnote": [], + "bbox": [ + 163, + 309, + 370, + 407 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/0396bafb1c9c14e77e24411b8cb9818137cc28c575ff43b4583a7613a5968841.jpg", + "image_caption": [ + "(b)" + ], + "image_footnote": [], + "bbox": [ + 380, + 309, + 589, + 407 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/844436a69d6dc0fef38a47a216ad45f5b999d87491fd9f31476371e60bd75c3b.jpg", + "image_caption": [ + "(c)", + "Figure 7. Visualization of the frame-word alignments. (a) CLIP; (b) $\\mathrm{CLIPPING}^{-}_{\\mathrm{TAB}}$ ; (c) CLIPPING." + ], + "image_footnote": [], + "bbox": [ + 598, + 309, + 805, + 407 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "smaller.", + "bbox": [ + 75, + 464, + 133, + 477 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "KD Types. In Table 5, we compare the proposed SAB KD with the previous OL2OL KD and TAB KD. For SAB and TAB, the same layers of the student and the teacher are used for KD, the details of which are given in the supplementary materials. From Table 5, we can see that TAB performers better than OL2OL, and our SAB outperforms TAB. In Section 3.3.2, we use masking to speed up the training. In this study, we also compare \"with masking\" and \"without masking\". The last two rows of Table 5 verify that the masking is beneficial.", + "bbox": [ + 75, + 483, + 468, + 633 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "SAB Property. The student's layers in our SAB KD can be explained as the bases of the feature space, and each layer of the teacher is a linear combination of the bases (Section 3.2). In Fig. 6, we show that this property holds. In SAB KD, the student has 4 layers and the teacher has 12 layers. We randomly pick 3 image examples, and for each example, we select the features of one random token (in the $k$ dimension in Fig. 6(a)). After training, the linear combinations $\\tilde{W} \\times \\tilde{L}_{\\text{layer}}^{S} = \\tilde{L}_{\\text{layer}}^{S}$ show feature patterns very similar to the teacher's features ( $\\tilde{L}_{\\text{layer}}^{T}$ ). This property verifies that the teacher's knowledge is fully absorbed by the student.", + "bbox": [ + 75, + 638, + 470, + 820 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In Fig. 7, we visualize the frame-word alignments with and without SAB KD. Fig. 7(a) is the result of the CLIP in CLIP4clip. Fig. 7(b) is the result of $\\mathrm{CLIPPING}^{-}_{\\mathrm{TAB}}$ , which is our model without SAB KD but with the previous TAB KD. It can be seen that the frame-word alignment with", + "bbox": [ + 75, + 825, + 468, + 902 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "SAB KD (Fig. 7(c)) shows clear and correct word-level attentions (e.g., \"naked\", \"child\", \"runs\" and \"through\") like CLIP. But with TAB KD, the most important words (e.g., \"naked\" and \"child\") cannot be noticed; instead, some insignificant words ([CLS] and \"a\") are activated. This study shows that the knowledge of pre-trained CLIP can be more effectively transferred to the student at the fine-tuning stage with our SAB.", + "bbox": [ + 496, + 463, + 890, + 584 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 606, + 617, + 619 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we propose a novel and efficient knowledge distillation method that is specially designed for small vision-language models. It includes temporal KD, spatial KD, SAB KD and cross-modality KD. Especially, the SAB KD has the property of the student's layers being the bases of the feature space. After training, the teacher's features are the linear combinations of the bases, indicating that the student has fully absorbed the knowledge of the teacher. CLIPPING significantly outperforms the state-of-the-art and is comparable or even superior to many large pre-training models. In the feature, we will apply CLIPPING to other vision-language models for compression.", + "bbox": [ + 496, + 626, + 890, + 808 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 500, + 828, + 666, + 844 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We gratefully acknowledge the support of MindSpore [1], CANN (Compute Architecture for Neural Networks) and Ascend AI Processor used for this research.", + "bbox": [ + 498, + 854, + 890, + 898 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "18990", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] https://www.mindspore.cn/.8", + "[2] Elad Amrani, Rami Ben-Ari, Daniel Rotman, and Alex Bronstein. Noise estimation using density estimation for self-supervised multimodal learning. In AAAI, 2021. 7", + "[3] Max Bain, Arsha Nagrani, Gül Varol, and Andrew Zisserman. Frozen in Time: A Joint Video and Image Encoder for End-to-End Retrieval. In ICCV, 2021. 2, 6, 7", + "[4] David Chen and William Dolan. Collecting highly parallel data for paraphrase evaluation. In ACL, 2011. 6", + "[5] Defang Chen, JianPing Mei, Yuan Zhang, Can Wang, Zhe Wang, Yan Feng, and Chun Chen. Distillation with semantic calibration. In AAAI, 2021. 2, 3, 5", + "[6] Pengguang Chen, Shu Liu, Hengshuang Zhao, and Jiaya Jia. Distilling Knowledge via Knowledge Review. In CVPR, 2021. 2, 3", + "[7] Xianing Chen, Qiong Cao, Yujie Zhong, Jing Zhang, Shenghua Gao, and Dacheng Tao. DearKD: Data-efficient early knowledge distillation for vision transformers. In ICLR, 2022. 2", + "[8] Yinpeng Chen, Xiyang Dai, Dongdong Chen, Mengchen Liu, Xiaoyi Dong, Lu Yuan, and Zicheng Liu. Mobile-Former: Bridging MobileNet and Transformer. In CVPR, 2022. 1", + "[9] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. ImageNet: A large-scale hierarchical image database. In CVPR, 2009. 7", + "[10] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2021. 2", + "[11] Maksim Dzabraev, Maksim Kalashnikov, Stepan Komkov, and Aleksandr Petiushko. MDMMT: Multidomain Multimodal Transformer for Video Retrieval. In CVPR, 2021. 2, 6, 7", + "[12] Tsu-Jui Fu, Linjie Li, Zhe Gan, Kevin Lin, William Yang Wang, Lijuan Wang, and Zicheng Liu. VIOLET: End-to-End Video-Language Transformers with Masked Token Modeling. In arXiv:2111.1268, 2021. 1, 6", + "[13] Yuying Ge, Yixiao Ge, Xihui Liu, Dian Li, Ying Shan, Xiaohu Qie, and Ping Luo. Bridgeformer: Bridging video-text retrieval with multiple choice questions. In CVPR, 2022. 6", + "[14] Jianyuan Guo, Kai Han, Han Wu, Yehui Tang, Xinghao Chen, Yunhe Wang, and Chang Xu. CMT: Convolutional Neural Networks Meet Vision Transformers. In CVPR, 2022. 1", + "[15] Geoffrey Hinton, Oriol Vinyals, and Jeff Dean. Distilling the Knowledge in a Neural Network. In arXiv preprint arXiv:1503.02531, 2015. 1, 2", + "[16] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In ICML, 2021. 1", + "[17] Xiaoqi Jiao, Yichun Yin, Lifeng Shang, Xin Jiang, Xiao Chen, Linlin Li, Fang Wang, and Qun Liu. TinyBERT:" + ], + "bbox": [ + 78, + 114, + 468, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Distilling BERT for Natural Language Understanding. In EMNLP, 2020. 1, 2, 5, 7", + "[18] Yoon Kim and Alexander M. Rush. Sequence-level knowledge distillation. In EMNLP, 2016. 1", + "[19] Pavan Kumar, Anasosalu Vasu, James Gabriel, Jeff Zhu, Oncel Tuzel, and Anurag Ranjan. An Improved One millisecond Mobile Backbone. In arXiv preprint arXiv:2206.04040, 2022. 1, 2", + "[20] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. BLIP: Bootstrapping Language-Image Pre-training for Unified Vision-Language Understanding and Generation. In arXiv preprint arXiv:2201.12086, 2022. 1", + "[21] Junnan Li, Ramprasaath Selvaraju, Akhilesh Gotmare, Shafiq Joty, Caiming Xiong, and Steven Chu Hong Hoi. Align before fuse: Vision and language representation learning with momentum distillation. In NIPS, 2021. 1", + "[22] Yanyu Li, Geng Yuan, Yang Wen, Eric Hu, Georgios Evangelidis, Sergey Tulyakov, Yanzhi Wang, and Jian Ren. EfficientFormer: Vision Transformers at MobileNet Speed. In arXiv:2206.01191, 2022. 1", + "[23] Sihao Lin, Hongwei Xie, Bing Wang, Kaicheng Yu, Xiaojun Chang, Xiaodan Liang, and Gang Wang. Knowledge Distillation via the Target-aware Transformer. In CVPR, 2022. 2, 3", + "[24] Yifan Liu, Chunhua Shen, Changqian Yu, and Jingdong Wang. Efficient Semantic Video Segmentation with Per-Frame Inference. In ECCV, 2020. 4", + "[25] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin Transformer: Hierarchical Vision Transformer using Shifted Windows. In ICCV, 2021. 2", + "[26] Huaishao Luo, Lei Ji, Ming Zhong, Yang Chen, Wen Lei, Nan Duan, and Tianrui Li. Clip4clip: An empirical study of clip for end to end video clip retrieval. arXiv preprint arXiv:2104.08860, 2021. 1, 2", + "[27] Sachin Mehta and Mohammad Rastegari. Separable Self-attention for Mobile Vision Transformers. In ICLR, 2022. 1, 2, 4", + "[28] Antoine Miech, Dimitri Zhukov, Jean-Baptiste Alayrac, Makarand Tapaswi, Ivan Laptev, and Josef Sivic. HowTo100M: Learning a Text-Video Embedding by Watching Hundred Million Narrated Video Clips. In ICCV, 2019. 7", + "[29] Mandela Patrick, Po-Yao Huang, Yuki Asano, Florian Metze, Alexander Hauptmann, João Henriques, and Andrea Vedaldi. Support-set bottlenecks for video-text representation learning. In ICLR, 2021. 7", + "[30] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, and Jack Clark. Learning transferable visual models from natural language supervision. ICML, 2021. 1, 2, 7", + "[31] Maithra Raghu, Thomas Unterthiner, Simon Kornblith, Chiyuan Zhang, and Alexey Dosovitskiy. Do vision transformers see like convolutional neural networks? In NIPS, 2021. 2, 3, 5", + "[32] Anna Rohrbach, Marcus Rohrbach, Niket Tandon, and Bernt Schiele. A dataset for Movie Description. In CVPR, 2015. 6" + ], + "bbox": [ + 503, + 92, + 890, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "18991", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[33] Piyush Sharma, Nan Ding, Sebastian Goodman, and Radu Soricut. Conceptual captions: A cleaned, hypernymed, image alt-text dataset for automatic image captioning. In ACL, 2018. 7", + "[34] Alex Jinpeng Wang, Yixiao Ge, Rui Yan, Yuying Ge, Xudong Lin, Guanyu Cai, Jianping Wu, Ying Shan, Xiaohu Qie, and Mike Zheng Shou. All in one: Exploring unified video-language pre-training. In arXiv preprint arXiv:2203.07303, 2022. 2, 6", + "[35] Jinpeng Wang, Yixiao Ge, Guanyu Cai, Rui Yan, Xudong Lin, Ying Shan, Xiaohu Qie, and Mike Zheng Shou. Object-aware video-language pre-training for retrieval. In CVPR, 2022. 6", + "[36] Zekun Wang, Wenhui Wang, Haichao Zhu, Ming Liu, Bing Qin, and Furu Wei. Distilled dual-encoder model for vision-language understanding. In arXiv preprint arXiv:2112.08723, 2021. 2", + "[37] Hu Xu, Gargi Ghosh, Po-Yao Huang, Dmytro Okhonko, Armen Aghajanyan, Florian Metze, Luke Zettlemoyer, and Christoph Feichtenhofer. Videoclip: Contrastive pre-training for zero-shot video-text understanding. arXiv preprint arXiv:2104.08860, 2021. 2, 6", + "[38] Jun Xu, Tao Mei, Ting Yao, and Yong Rui. MSR-VTT: A Large Video Description Dataset for Bridging Video and Language. 2016. 6", + "[39] Jianwei Yang, Yonatan Bisk, and Jianfeng Gao. Taco: Token-aware cascade contrastive learning for video-text alignment. In ICCV, 2021. 6", + "[40] Weihao Yu, Mi Luo, Pan Zhou, Chenyang Si, Yichen Zhou, Xinchao Wang, Jiashi Feng, and Shuicheng Yan. Metaformer is actually what you need for vision. In CVPR, 2022. 1", + "[41] Li Yuan, Yunpeng Chen, Tao Wang, Weihao Yu, Yujun Shi, Zi-Hang Jiang, Francis E.H. Tay, Jiashi Feng, and Shuicheng Yan. Tokens-to-Token ViT: Training Vision Transformers From Scratch on ImageNet. In ICCV, 2021. 5" + ], + "bbox": [ + 78, + 90, + 468, + 585 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "18992", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2023/CLIPPING_ Distilling CLIP-Based Models With a Student Base for Video-Language Retrieval/1eb2010b-78e6-492d-9404-5d49a6b93ef0_model.json b/2023/CLIPPING_ Distilling CLIP-Based Models With a Student Base for Video-Language Retrieval/1eb2010b-78e6-492d-9404-5d49a6b93ef0_model.json new file mode 100644 index 0000000000000000000000000000000000000000..e8f118b13ed503c629be91ff2e1cdf79d6f3bd7b --- /dev/null +++ b/2023/CLIPPING_ Distilling CLIP-Based Models With a Student Base for Video-Language Retrieval/1eb2010b-78e6-492d-9404-5d49a6b93ef0_model.json @@ -0,0 +1,2277 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.149, + 0.13, + 0.823, + 0.177 + ], + "angle": 0, + "content": "CLIPPING: Distilling CLIP-Based Models with a Student Base for Video-Language Retrieval" + }, + { + "type": "text", + "bbox": [ + 0.254, + 0.204, + 0.717, + 0.257 + ], + "angle": 0, + "content": "Renjing Pei, Jianzhuang Liu, Weimian Li, Bin Shao, Songcen Xu, Peng Dai, Juwei Lu, Youliang Yan Huawei Noah's Ark Lab" + }, + { + "type": "text", + "bbox": [ + 0.254, + 0.259, + 0.707, + 0.293 + ], + "angle": 0, + "content": "{peirenjing,liu.jianzhuang,liweimian,shaobin3, xusongcen,peng.dai,juwei.lu,yanyouliang}@huawei.com" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.327, + 0.314, + 0.344 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.367, + 0.473, + 0.82 + ], + "angle": 0, + "content": "Pre-training a vision-language model and then fine-tuning it on downstream tasks have become a popular paradigm. However, pre-trained vision-language models with the Transformer architecture usually take long inference time. Knowledge distillation has been an efficient technique to transfer the capability of a large model to a small one while maintaining the accuracy, which has achieved remarkable success in natural language processing. However, it faces many problems when applying KD to the multi-modality applications. In this paper, we propose a novel knowledge distillation method, named CLIPPING1, where the plentiful knowledge of a large teacher model that has been fine-tuned for video-language tasks with the powerful pre-trained CLIP can be effectively transferred to a small student only at the fine-tuning stage. Especially, a new layer-wise alignment with the student as the base is proposed for knowledge distillation of the intermediate layers in CLIPPING, which enables the student's layers to be the bases of the teacher, and thus allows the student to fully absorb the knowledge of the teacher. CLIPPING with MobileViT-v2 as the vision encoder without any vision-language pre-training achieves \\(88.1\\% - 95.3\\%\\) of the performance of its teacher on three video-language retrieval benchmarks, with its vision encoder being \\(19.5x\\) smaller. CLIPPING also significantly outperforms a state-of-the-art small baseline (ALL-in-one-B) on the MSR-VTT dataset, obtaining relatively \\(7.4\\%\\) performance gain, with \\(29\\%\\) fewer parameters and \\(86.9\\%\\) fewer flops. Moreover, CLIPPING is comparable or even superior to many large pre-training models." + }, + { + "type": "title", + "bbox": [ + 0.502, + 0.327, + 0.632, + 0.344 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.355, + 0.895, + 0.837 + ], + "angle": 0, + "content": "Recently, pre-training a vision-language model and then fine-tuning it on downstream tasks are a popular paradigm [12, 16, 20, 21, 30]. Pre-trained vision-language models (PVLMs) have achieved great success in many multi-modality tasks (e.g., video-text retrieval). However, PVLMs with the Transformer architecture (especially the vision stream) usually consume a huge amount of computation, making them difficult to be deployed on edge devices such as mobile phones. Recent small models [8, 14, 19, 22, 27, 40] show that combining Convolutional Neural Networks (CNNs) and Transformers as a hybrid architecture gets the best of both architectures, but the overall performance of these works is still far from satisfactory when compared to large pre-training models. Apparently, knowledge distillation (KD) [15] is an efficient technique to transfer the capability of a large model to a small one while maintaining the accuracy, which has achieved remarkable success in natural language processing (NLP) [17, 18]. In NLP, the knowledge distillation methods are usually performed at both the pre-training and the fine-tuning stages. However, the collection of the pre-training data for the pre-training knowledge distillation cost huge manpower in multi-modality applications (e.g., the pre-training dataset of CLIP is hard to obtain). Therefore, an efficient knowledge distillation method need to be explored for multi-modality applications. To this end, we propose a novel knowledge distillation method, named CLIPPING, where the plentiful knowledge of a large teacher model Clip4clip [26] that has been fine-tuned for video-language tasks with the powerful pre-trained CLIP [30] can be effectively transferred to a small student only at the fine-tuning stage. The contributions of CLIPPING are summarized below:" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.84, + 0.895, + 0.903 + ], + "angle": 0, + "content": "1) We introduce an efficient approach to distill both the vision knowledge and the cross-modality knowledge from teacher to a small model at the fine-tuning stage. The resulting model shows strong performance on multi-modality" + }, + { + "type": "page_footnote", + "bbox": [ + 0.076, + 0.875, + 0.471, + 0.902 + ], + "angle": 0, + "content": "1In this paper, CLIPPING means cutting something to make it smaller through distilling." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "18983" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.079, + 0.092, + 0.209, + 0.2 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.132, + 0.202, + 0.149, + 0.214 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.212, + 0.092, + 0.346, + 0.2 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.269, + 0.202, + 0.288, + 0.214 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image", + "bbox": [ + 0.347, + 0.092, + 0.472, + 0.2 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.404, + 0.202, + 0.421, + 0.214 + ], + "angle": 0, + "content": "(c)" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.224, + 0.47, + 0.432 + ], + "angle": 0, + "content": "Figure 1. (a) and (b) are previous methods of intermediate features' knowledge distillation. \\(L_{j}^{T}\\), \\(j = 1,2,\\dots,M\\), and \\(L_{i}^{S}\\), \\(i = 1,2,\\dots,N\\), are the outputs of the \\(j^{th}\\) and \\(i^{th}\\) layers of the teacher and the student, respectively. OL2OL is the most common One-Layer-to-One-Layer KD, where it selects some layers of the teacher and then distills the layers into the student one by one (e.g., from \\(L_{j}^{T}\\) to \\(L_{i}^{S}\\)). Teacher-As-Base (TAB) KD uses all the teacher's layers to supervise each layer in the student, where each layer of the student (e.g., \\(L_{i}^{S}\\)) learns the knowledge from all the selected layers of the teacher \\(\\left(\\sum_{k=1}^{M} w_{ik}' L_{k}^{T}\\right)\\), where \\(w_{ik}'\\), \\(k = 1,2,\\dots,M\\), are the knowledge selection weights with \\(\\sum_{k=1}^{M} w_{ik}' = 1\\). (c) Our Student-As-Base (SAB) KD enables each of the teacher's layers to pass its knowledge to all the student's layers (e.g., \\(L_{j}^{T} \\rightarrow \\sum_{k=1}^{N} w_{jk} L_{k}^{S}\\), \\(\\sum_{k=1}^{N} w_{jk} = 1\\)). More detailed analysis is provided in Section 3.1." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.448, + 0.24, + 0.461 + ], + "angle": 0, + "content": "text-video retrieval task." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.463, + 0.469, + 0.612 + ], + "angle": 0, + "content": "2) We propose a new layer-wise alignment scheme, called Student-As-Base (SAB), for knowledge distillation of the intermediate layers from the Transformer to the CNN in CLIPPING, where the student's layers can be regarded as the bases of the teacher's feature space, forcing the student model to full absorb the knowledge of the teacher. In our experience, the SAB layer-wise alignment significantly surpasses the previous knowledge distillation methods. We believe it will become a popular paradigm for knowledge distillation of intermediate features." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.614, + 0.469, + 0.765 + ], + "angle": 0, + "content": "3) We present an effective cross-modal knowledge distillation, which includes knowledge from both the global and local video-caption distributions. We use the video-caption distributions of the teacher to guide the training of the student. Besides, the student can also benefit from the powerful pre-trained CLIP that obtains certain local frame-word attention ability via learning from massive data. Experiments show that this pre-training knowledge can be effectively transferred to the student when jointly trained with our SAB KD only at the fine-tuning stage." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.469, + 0.901 + ], + "angle": 0, + "content": "4) CLIPPING with MobileViT-v2 [27] as the vision encoder without any vision-language pre-training achieves \\(91.5\\% - 95.3\\%\\) of the performance of its teacher on MSR-VTT video-language retrieval benchmark, with its vision encoder being \\(19.5 \\times\\) smaller. CLIPPING also significantly outperforms a state-of-the-art small baseline (ALL-in-one-B) on the MSR-VTT dataset, obtaining relatively \\(7.4\\%\\) performance gain, with \\(29\\%\\) fewer parameters and \\(86.9\\%\\) fewer flops. Moreover, CLIPPING is comparable or even" + }, + { + "type": "list", + "bbox": [ + 0.076, + 0.463, + 0.469, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.092, + 0.788, + 0.107 + ], + "angle": 0, + "content": "superior to many large pre-training models." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.124, + 0.646, + 0.14 + ], + "angle": 0, + "content": "2. Relative Work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.143, + 0.892, + 0.355 + ], + "angle": 0, + "content": "Vision-Language Retrieval. Learning from web-collected image-text data, large-scale Vision-Language Pretraining (VLP) models such as CLIP [30] have recently demonstrated great success across various downstream tasks. Nowadays, models such as Clip4clip [26] and MDMMT [11] extended from the pre-trained model CLIP keep appearing for vision-language retrieval. There are also some end-to-end trainable models [3, 37], which are designed to take advantage of both large-scale image and video captioning datasets. All-in-one [34] is the first work to consider both efficiency and performance for video-language retrieval tasks. It introduces a unified backbone that enables the representation learning of both video-text multimodal and unimodal inputs." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.356, + 0.893, + 0.703 + ], + "angle": 0, + "content": "Knowledge Distillation. Knowledge distillation (KD) [15] is a simple yet effective technique to improve the performance of a learning model. Earlier works transfer knowledge embedded in the \"logits\" learned in a large teacher model to a small student model without sacrificing much performance. Recent works [5, 6] use multiple layers of the teacher to supervise each layer in the student, where each layer of the student learns the knowledge from multiple layers of the teacher. [23] proposes a target-aware Transformer and enables the student to mimic each spatial component of the teacher in each distilled layer to boost the student's performance. For multi-modality KD, [36] designs a fusion-encoder model as the teacher and introduces cross-modal attention knowledge to train the dual-encoder student model. The distillation objective is applied at both the pre-training and the fine-tuning stages and helps the dual-encoder model learn interactions of different modalities. TinyBert [17] also introduces a two-stage learning framework that performs Transformer distillation at both the pre-training and the task-specific fine-tuning stages. In this paper, we also focus on KD for transferring the knowledge of a pre-training vision-language model to a small one but only at the fine-tuning stage." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.704, + 0.893, + 0.901 + ], + "angle": 0, + "content": "Transformers and CNNs. Usually, the pre-training model is a Transformer-based model [10, 25] and the small model is a CNN-based or hybrid architecture [19, 27]. For KD, most existing methods distill knowledge either from a Transformer to another Transformer (T2T) or from a CNN to another CNN (C2C) that computes the loss in the OL2OL or TAB style (see Fig. 1). To the best of our knowledge, KD from a Transformer to a CNN (T2C) has not been explored yet. The most related work [7] distills knowledge from a CNN to a Transformer (C2T) in the OL2OL way. However, previous works (e.g., [31]) have investigated the internal representation structures of ViTs and CNNs, and found striking differences between the two models, such as" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18984" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.241, + 0.087, + 0.72, + 0.367 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.383, + 0.895, + 0.425 + ], + "angle": 0, + "content": "Figure 2. Overview of CLIPPING. There are mainly four knowledge distillation (KD) parts: (1) Temporal KD. (2) Spatial KD. (3) SAB (Student-As-Base) KD. (4) Cross-modality KD. The CLIP's vision encoder is a Transformer. The temporal Transformer is an optional module for CLIPPING." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.439, + 0.473, + 0.545 + ], + "angle": 0, + "content": "ViTs having highly similar representations throughout the model's layers, while CNNs showing obvious distinction of representations between lower and higher layers [31]. Considering such striking differences between ViTs and CNNs, the previous KD (OL2OL and TAB) may not be good distillation ways for our T2C task, which is verified by our experiments." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.558, + 0.214, + 0.575 + ], + "angle": 0, + "content": "3. Methodology" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.584, + 0.47, + 0.705 + ], + "angle": 0, + "content": "We propose an efficient approach to distill both the vision knowledge and the cross-modality knowledge from Clip4clip to a small model with MobileViT-v2 as its vision encoder for text-video retrieval task, which can get the best from both sides: powerful multimodal representations of CLIP and efficient mobile vision Transformer of MobileViT-v2. The overall architecture is illustrated in Fig. 2." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.713, + 0.312, + 0.729 + ], + "angle": 0, + "content": "3.1. Analysis of TAB and SAB" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.736, + 0.47, + 0.782 + ], + "angle": 0, + "content": "We first show the advantage of our SAB KD over traditional TAB KD. For simplicity, suppose \\( M = N = 2 \\). Then, for SAB," + }, + { + "type": "equation", + "bbox": [ + 0.119, + 0.791, + 0.47, + 0.832 + ], + "angle": 0, + "content": "\\[\n\\left\\{ \\begin{array}{l} L _ {1} ^ {T} = w _ {1 1} L _ {1} ^ {S} + w _ {1 2} L _ {2} ^ {S}, w _ {1 1} + w _ {1 2} = 1 \\\\ L _ {2} ^ {T} = w _ {2 1} L _ {1} ^ {S} + w _ {2 2} L _ {2} ^ {S}, w _ {2 1} + w _ {2 2} = 1 \\end{array} , \\right. \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.84, + 0.168, + 0.854 + ], + "angle": 0, + "content": "and for TAB," + }, + { + "type": "equation", + "bbox": [ + 0.119, + 0.864, + 0.469, + 0.905 + ], + "angle": 0, + "content": "\\[\n\\left\\{ \\begin{array}{l} L _ {1} ^ {S} = w _ {1 1} ^ {\\prime} L _ {1} ^ {T} + w _ {1 2} ^ {\\prime} L _ {2} ^ {T}, w _ {1 1} ^ {\\prime} + w _ {1 2} ^ {\\prime} = 1 \\\\ L _ {2} ^ {S} = w _ {2 1} ^ {\\prime} L _ {1} ^ {T} + w _ {2 2} ^ {\\prime} L _ {2} ^ {T}, w _ {2 1} ^ {\\prime} + w _ {2 2} ^ {\\prime} = 1 \\end{array} \\right. \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.439, + 0.892, + 0.515 + ], + "angle": 0, + "content": "Obviously, in Eq. 1, the knowledge of the teacher \\( L_1^T \\) or \\( L_2^T \\) is the linear combination of \\( L_1^S \\) and \\( L_2^S \\). In other words, \\( L_1^S \\) and \\( L_2^S \\) are the bases of \\( L_1^T \\) and \\( L_2^T \\) of the teacher. In Eq. 2, it can be seen that the teacher's \\( L_1^T \\) and \\( L_2^T \\) are as the bases for \\( L_1^S \\) and \\( L_2^S \\). Eq. 2 can be converted to:" + }, + { + "type": "equation", + "bbox": [ + 0.52, + 0.523, + 0.892, + 0.543 + ], + "angle": 0, + "content": "\\[\nL _ {1} ^ {T} = A \\cdot L _ {1} ^ {S} + B \\cdot L _ {2} ^ {S}, L _ {2} ^ {T} = C \\cdot L _ {1} ^ {S} + D \\cdot L _ {2} ^ {S}, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.553, + 0.892, + 0.576 + ], + "angle": 0, + "content": "where \\(A = \\frac{w_{22}^{\\prime}}{w_{11}^{\\prime}w_{22}^{\\prime} - w_{12}^{\\prime}w_{21}^{\\prime}}\\), \\(B = \\frac{-w_{12}^{\\prime}}{w_{11}^{\\prime}w_{22}^{\\prime} - w_{12}^{\\prime}w_{21}^{\\prime}}\\), \\(C =\\)" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.575, + 0.892, + 0.702 + ], + "angle": 0, + "content": "\\(\\frac{-w_{21}^{\\prime}}{w_{11}^{\\prime}w_{22}^{\\prime} - w_{12}^{\\prime}w_{21}^{\\prime}}\\) and \\(D = \\frac{w_{11}^{\\prime}}{w_{11}^{\\prime}w_{22}^{\\prime} - w_{12}^{\\prime}w_{21}^{\\prime}}\\), which looks like Eq. 1. It seems that in this case, the student's \\(L_1^S\\) and \\(L_2^S\\) could also serve as the bases for \\(L_1^T\\) and \\(L_2^T\\). However, (i) \\(A + B \\neq 1\\) and \\(C + D \\neq 1\\), but it is usually required in the previous KD that all the coefficients in each equation should be summed to be 1 [5,6,23]; (ii) when learning with TAB KD in Eq. 2, the coefficients A, B, C and D are uncontrollable, which may be arbitrary large during learning." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.703, + 0.892, + 0.838 + ], + "angle": 0, + "content": "For the reasons above, although it seems that SAB and TAB have the same matrix representation, TAB is unable to get a similar result as SAB and more likely to get to a local minimum during training. Therefore, TAB KD is not equivalent to our SAB KD. When well trained, the student with this SAB KD fully learns the knowledge of the teacher because the teacher's \\( L_{i}^{T} \\) can be obtained by the linear combination of the student's \\( L_{i}^{S} \\), \\( i = 1,2,\\dots,N \\). Our experiment also shows this property (see Section 4.2)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.848, + 0.64, + 0.862 + ], + "angle": 0, + "content": "3.2. Preliminaries" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.892, + 0.902 + ], + "angle": 0, + "content": "Given a batch of videos \\(V\\) and captions \\(T\\), Clip4clip learns a similarity function \\(s(v_{i},t_{j})\\) to calculate the similar-" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "18985" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.179, + 0.087, + 0.796, + 0.203 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.179, + 0.204, + 0.794, + 0.328 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.115, + 0.337, + 0.854, + 0.353 + ], + "angle": 0, + "content": "Figure 3. Reshaping the features of (a) the student \\( L_{k}^{S} \\) and (b) the teacher \\( L_{k}^{T} \\) to the same shape. (c) Similarity Computation." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.374, + 0.473, + 0.618 + ], + "angle": 0, + "content": "ity between a video \\( v_{i} \\in V \\) and a caption \\( t_j \\in T \\). It obtains the frames' representation and the caption representation in a multi-modal embedding space via CLIP. The frames' representation is denoted as \\( Z_{i} = \\{z_{i1}, z_{i2}, \\dots, z_{in}\\} \\), where \\( n \\) is the number of frames in \\( v_{i} \\) and \\( z_{ik} \\) (\\( k = 1, 2, \\dots, n \\)) is the class token from the output of the CLIP's vision encoder corresponding to the \\( k^{th} \\) input frame. The caption representation is denoted as \\( W_{j} = \\{w_{j1}, w_{j2}, \\dots, w_{jm}\\} \\), where \\( m \\) is the number of words in \\( t_j \\) (including [CLS] and [SEP]) and \\( w_{jm} \\) is used as the representation of \\( t_j \\) in Clip4clip. Clip4clip uses a temporal Transformer encoder to obtain the sequential feature, denoted as \\( Z_{i}' = \\{z_{i1}', z_{i2}', \\dots, z_{in}'\\} = TemporalTransformer(Z_{i}) \\). Then, the mean pooling is used to aggregate the features of all frames to obtain the video representation, \\( z_{i} = \\frac{1}{n} \\sum_{k=1}^{n} z_{ik}' \\). Finally, the similarity functions \\( s(v_i, t_j) \\) and \\( s(t_j, v_i) \\) are defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.14, + 0.62, + 0.47, + 0.641 + ], + "angle": 0, + "content": "\\[\ns (v _ {i}, t _ {j}) = w _ {j m} ^ {\\top} z _ {i}, \\quad s (t _ {i}, v _ {j}) = z _ {j} ^ {\\top} w _ {i m}. \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.644, + 0.473, + 0.841 + ], + "angle": 0, + "content": "Nowadays, models such as Clip4clip based on the pretrained model CLIP have been applied to many tasks. They pursue better performance with CLIP but have a heavy vision encoder, which makes them difficult to be deployed on edge devices such as mobile phones. Since MobileViT-v2 [27] is a light-weight and mobile-friendly hybrid network, we employ it as the vision encoder of the student model and maintain the original text encoder and temporal Transformer of Clip4clip. Note that the temporal Transformer is an optional module for the CLIPPING model structure, and the text encoder \\((CLIP_{bert})\\) can be further compressed through the existing methods (details are shown in the Section 4)." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.848, + 0.203, + 0.862 + ], + "angle": 0, + "content": "3.3. CLIPPING" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.871, + 0.47, + 0.9 + ], + "angle": 0, + "content": "To distill multimodal knowledge from Clip4clip to the MobileViT-v2-based model, we use four kinds of knowl-" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.374, + 0.892, + 0.404 + ], + "angle": 0, + "content": "edge transfer: 1) temporal KD, 2) spatial KD, 3) SAB KD, and 4) cross-modality KD." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.422, + 0.868, + 0.437 + ], + "angle": 0, + "content": "3.3.1 Temporal and Spatial Knowledge Distillation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.445, + 0.892, + 0.567 + ], + "angle": 0, + "content": "The temporal KD is motivated by the previous finding that aligning multi-frame dependency from the teacher to the student can enhance the performance of the student [24], which encodes the multi-frame dependency into a latent embedding by using a recurrent unit ConvLSTM. In our architecture, multi-frame dependency can be modeled naturally through the temporal Transformer (Temporal Transformer) in Fig. 2, so we define the temporal KD loss as:" + }, + { + "type": "equation", + "bbox": [ + 0.581, + 0.575, + 0.892, + 0.616 + ], + "angle": 0, + "content": "\\[\nL _ {T K D} = \\frac {1}{B} \\sum_ {i = 1} ^ {B} D _ {K L} \\left(Z _ {i} ^ {\\prime S}, Z _ {i} ^ {\\prime T}\\right), \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.624, + 0.892, + 0.7 + ], + "angle": 0, + "content": "where \\(B\\) is the batch size, the superscripts \\(S\\) and \\(T\\) denote the student and the teacher, respectively, and \\(D_{KL}()\\) is the KL divergence loss function. In addition to imitating the behavior of CLIP for each frame, we use spatial KD to align the spatial embeddings of the student and the teacher by:" + }, + { + "type": "equation", + "bbox": [ + 0.586, + 0.708, + 0.891, + 0.747 + ], + "angle": 0, + "content": "\\[\nL _ {S K D} = \\frac {1}{B} \\sum_ {i = 1} ^ {B} D _ {K L} \\left(Z _ {i} ^ {S}, Z _ {i} ^ {T}\\right). \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.747, + 0.892, + 0.793 + ], + "angle": 0, + "content": "The Temporal Transformer is an optional module for the CLIPPING model structure in our method. In this case, \\( Z_{i}^{\\prime S} \\) is exactly equal to \\( Z_{i}^{S} \\) in TKD." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.807, + 0.749, + 0.822 + ], + "angle": 0, + "content": "3.3.2 SAB Knowledge Distillation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.825, + 0.892, + 0.9 + ], + "angle": 0, + "content": "Recently, it is discovered that distilling intermediate features is more effective. So in addition to the spatial and temporal KD, we also force the vision encoder of the student to mimic the intermediate layers of the teacher. In the traditional OL2OL KD, some layers of the teacher are selected" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18986" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.241, + 0.087, + 0.416, + 0.21 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.419, + 0.087, + 0.732, + 0.212 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.241, + 0.213, + 0.732, + 0.334 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.344, + 0.895, + 0.375 + ], + "angle": 0, + "content": "Figure 4. (a) Masks obtained with Eq. 7, where all the \\( L - 1 \\) masks along the \\( k \\) dimension are the same. (b) Masking the similarity tensor \\( W \\). (c) SAB alignment. \\( L_{SAB} \\) is the SAB loss." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.393, + 0.473, + 0.681 + ], + "angle": 0, + "content": "and distilled into the student one by one [17]. However, it is difficult to find the optimal correspondence between the student and the teacher. The recent TAB works (e.g., [5]) still cannot help the student learn enough knowledge from the teacher when there is a large architecture gap between them, which is verified in our experiments. For our task, the student (MobileViT-v2) is a hybrid architecture, which has convolution layers in the shallow stages and separable self-attention layers in the later blocks. And a separable self-attention layer differs significantly from the traditional self-attention in two ways: 1) it does not learn an explicit attention map; 2) its input and output are still 3D, which are more similar to CNN features. Except for the separable self-attention in the later blocks, the shallow stages with convolution layers undoubtedly have a huge difference from the Transformer structure in CLIP, which has been investigated in previous works [31,41]. Therefore, we design SAB layer-wise alignment for KD of the intermediate layers from the Transformer to the CNN." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.686, + 0.47, + 0.822 + ], + "angle": 0, + "content": "SAB Property. As shown in Fig. 2, the student's layers in our SAB KD can be explained as the bases of the feature space, and each layer of the teacher is a linear combination of the bases. After training, if the student's layers do ideally form the bases, the knowledge of the teacher's layers is learned completely by the student. Our experiment in Section 4.2 shows that the teacher's features in different layers can be well recovered from the features of the student's layers. Next, we describe how to implement SAB KD." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.825, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Feature Reshaping and Similarity Computation. Let the layers of the teacher and the student be \\( \\text{Layer}^T = [L_1^T, L_2^T, \\ldots, L_M^T] \\) and \\( \\text{Layer}^S = [L_1^S, L_2^S, \\ldots, L_N^S] \\) (usually \\( M > N \\)), respectively. \\( L_k^T \\in R^{D \\times L} \\) is a 2D tensor, where \\( D \\) is the dimensionality of the token feature" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.393, + 0.895, + 0.653 + ], + "angle": 0, + "content": "and \\(L\\) is the number of tokens, while \\(L_{k}^{S}\\in R^{C_{k}\\times H_{k}\\times W_{k}}\\) is a 3D tensor, where \\(C_k\\), \\(H_{k}\\) and \\(W_{k}\\) are the channel number, height and width of the \\(k^{th}\\) CNN layer's feature. We calculate a similarity tensor \\(W\\) between \\(L_{k}^{S}\\) and \\(L_{k}^{T}\\) which need to be reshaped first as shown in Figs. 3(a) and (b), respectively. For \\(L_{k}^{S}\\), we first apply a learnable linear operator and then do pixel reshuffle on the result, obtaining \\(\\bar{L}_k^S\\), \\(k = 1,2,\\dots,N\\). Next, \\(\\bar{L}_k^S\\) is reshaped to \\(\\tilde{L}_k^S\\). As for \\(L_{k}^{T}\\), we remove the class token and obtain a \\((L - 1)\\times D\\) tensor, denoted as \\(\\tilde{L}_k^T\\), \\(k = 1,2,\\dots,M\\). Finally, all \\(\\tilde{L}_k^S\\), \\(k = 1,2,\\dots,N\\), and \\(\\tilde{L}_k^T\\), \\(k = 1,2,\\dots,M\\), are represented as \\(\\tilde{L}_{\\mathrm{er}}^{S} = [\\tilde{L}_{1}^{S},\\tilde{L}_{2}^{S},\\dots,\\tilde{L}_{N}^{S}]\\) and \\(\\tilde{L}_{\\mathrm{er}}^{T} = [\\tilde{L}_{1}^{T},\\tilde{L}_{2}^{T},\\dots,\\tilde{L}_{M}^{T}]\\), respectively. After this reshaping, as shown in Fig. 3(c), we are able to measure the similarities between the intermediate layers of the student and the teacher through \\(W = \\tilde{L}_{\\mathrm{er}}^{T}\\times (\\tilde{L}_{\\mathrm{er}}^{S})^{\\top}\\), \\(W\\in\\) \\(R^{(L - 1)\\times M\\times N}\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.655, + 0.895, + 0.698 + ], + "angle": 0, + "content": "Masking and SAB Alignment. To speed up the training procedure, we design SAB KD with sequential masks, which follow this formula:" + }, + { + "type": "equation", + "bbox": [ + 0.538, + 0.705, + 0.892, + 0.756 + ], + "angle": 0, + "content": "\\[\n\\operatorname {m a s k} _ {(i, j)} = \\left\\{ \\begin{array}{l} 0, \\text {i f} j > n _ {0} \\text {a n d} i \\leq m _ {0} \\\\ 0, \\text {i f} j > n _ {1} \\text {a n d} m _ {0} < i \\leq m _ {1} \\\\ 1, \\text {e l s e} \\end{array} , \\right. \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.895, + 0.903 + ], + "angle": 0, + "content": "where \\( i = 1,2,\\dots,M \\), \\( j = 1,2,\\dots,N \\), \\( 1\\leq m_0 < m_1 < M \\), \\( 1\\leq n_0 < n_1 < N \\). And the final attention mask \\( \\mathsf{Mask}\\in R^{(L - 1)\\times M\\times N} \\) is composed of \\( L - 1 \\) same masks, Fig. 4(a) shows one example with \\( m_0 = 1,n_0 = 2,m_1 = 3,n_1 = 3 \\). \\( \\mathsf{mask}_{(i,j)} = 0 \\) means to mask all the elements along the \\( k \\) dimension (Fig. 4(a)) at \\( (i,j) \\) in \\( W \\), while \\( \\mathsf{mask}_{(i,j)} = 1 \\) means to maintain their similarities, as shown in Fig. 4(b). This masking is based on our experimental finding: The student's lower-level layers" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "18987" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.473, + 0.243 + ], + "angle": 0, + "content": "should learn from the teacher's lower-layers, while the student's higher layers should learn from all the teacher's layers. In Fig. 4(b), we also normalize the masks by performing softmax along the \\(j\\) dimension on each row, obtaining \\(\\tilde{W} = \\sigma (W\\circ \\mathsf{Mask})\\), where \\(\\sigma\\) is the softmax function and \\(\\circ\\) denotes the element-wise product. Guided by \\(\\tilde{W}\\), as shown in Fig. 4(c), the weighted student layers are calculated as: \\(\\tilde{L a r e r}_{w}^{S} = \\tilde{W}\\times \\tilde{L a r e r}^{S}\\). Each component of \\(\\tilde{L a r e r}_{w}^{S}\\) is the linear combination of \\(N\\) student layers' features. Finally, the SAB KD loss is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.156, + 0.252, + 0.47, + 0.272 + ], + "angle": 0, + "content": "\\[\nL _ {S A B} = D _ {K L} \\left(\\tilde {L a y e r} _ {w} ^ {S}, \\tilde {L a y e r} ^ {T}\\right). \\tag {8}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.295, + 0.401, + 0.31 + ], + "angle": 0, + "content": "3.3.3 Cross-Modality Knowledge distillation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.318, + 0.47, + 0.422 + ], + "angle": 0, + "content": "To further adapt the student's features to the teacher's multimodal feature space, we employ the teacher's knowledge of cross-modality distributions to guide the training of the student. Specifically, we characterize the cross-modal distributions from two perspectives, global video-caption distributions of the teacher and local video-caption distributions of CLIP." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.423, + 0.47, + 0.495 + ], + "angle": 0, + "content": "Global Video-Caption Distribution Alignment. We consider both the video-to-caption distribution \\( A_{GVC} \\) and the caption-to-video distribution \\( A_{GCV} \\), the elements \\( (s(v_i, t_j) \\) and \\( s(t_j, v_i)) \\) of which are defined via the similarities obtained by Eq. 4:" + }, + { + "type": "equation", + "bbox": [ + 0.136, + 0.502, + 0.402, + 0.547 + ], + "angle": 0, + "content": "\\[\nA _ {G V C} = \\sigma \\left( \\begin{array}{c c c} s (v _ {1}, t _ {1}) & ... & s (v _ {1}, t _ {B}) \\\\ ... & ... & ... \\\\ s (v _ {B}, t _ {1}) & ... & s (v _ {B}, t _ {j}) \\end{array} \\right),\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.14, + 0.543, + 0.468, + 0.593 + ], + "angle": 0, + "content": "\\[\nA _ {G C V} = \\sigma \\left( \\begin{array}{c c c} s (t _ {1}, v _ {1}) & \\dots & s (t _ {1}, v _ {B}) \\\\ \\dots & \\dots & \\dots \\\\ s (t _ {B}, v _ {1}) & \\dots & s (t _ {B}, v _ {B}) \\end{array} \\right). \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.601, + 0.47, + 0.644 + ], + "angle": 0, + "content": ", where \\(\\sigma\\) is the softmax function along the first dimension of the similarity matrixs. Finally, the global video-caption distribution alignment loss is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.096, + 0.651, + 0.47, + 0.669 + ], + "angle": 0, + "content": "\\[\nL _ {G} = D _ {K L} \\left(A _ {G V C} ^ {S}, A _ {G V C} ^ {T}\\right) + D _ {K L} \\left(A _ {G C V} ^ {S}, A _ {G C V} ^ {T}\\right). \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.677, + 0.471, + 0.901 + ], + "angle": 0, + "content": "Local Video-Caption Distribution Alignment. CLIP uses text prompts (such as \"A picture of a (\") for zero-shot image classification. CLIP fills them with different words (e.g., \"cat\" and \"dog\") and results in different captions (e.g., \"A picture of a cat\" and \"A picture of a dog\"). It can match the captions to the corresponding images, showing some image-word alignment ability. We transfer this pre-training knowledge to the student through a local frame-word alignment as follows. Recall that \\( Z_{i}^{\\prime} = \\{z_{i1}^{\\prime}, z_{i2}^{\\prime}, \\dots, z_{in}^{\\prime}\\} \\) and \\( W_{j} = \\{w_{j1}, w_{j2}, \\dots, w_{jm}\\} \\) are the features of the video \\( v_{i} \\) and the caption \\( t_{j} \\), respectively, with \\( n \\) being the number of frames in \\( v_{i} \\) and \\( m \\) the number of words in \\( t_{j} \\) (Section 3.1). The similarity between the \\( k^{th} \\) frame and the \\( r^{th} \\) word is defined as \\( s_{fw}(w_{jr}, z_{ik}') = (w_{jr})^\\top z_{ik}' \\). Then we respectively define the local video-to-caption and caption-to-video similarities as:" + }, + { + "type": "equation", + "bbox": [ + 0.563, + 0.09, + 0.892, + 0.131 + ], + "angle": 0, + "content": "\\[\ns ^ {\\prime} \\left(v _ {i}, t _ {j}\\right) = \\frac {1}{n} \\sum_ {k = 1} ^ {n} \\max _ {1 \\leq r \\leq m} \\left\\{s _ {f w} \\left(w _ {j r}, z _ {i k} ^ {\\prime}\\right) \\right\\}, \\tag {11}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.563, + 0.128, + 0.83, + 0.162 + ], + "angle": 0, + "content": "\\[\ns ^ {\\prime} (t _ {j}, v _ {i}) = \\frac {1}{m} \\sum_ {r = 1} ^ {m} \\max _ {1 \\leq k \\leq n} \\left\\{s _ {f w} \\left(z _ {i k} ^ {\\prime}, w _ {j r}\\right) \\right\\}.\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.501, + 0.17, + 0.892, + 0.2 + ], + "angle": 0, + "content": "Finally, we have the local video-to-caption distribution \\( A_{LVC} \\) and the local caption-to-video distribution \\( A_{LCV} \\):" + }, + { + "type": "equation", + "bbox": [ + 0.554, + 0.208, + 0.892, + 0.259 + ], + "angle": 0, + "content": "\\[\nA _ {L V C} = \\sigma \\left( \\begin{array}{c c c} s ^ {\\prime} \\left(v _ {1}, t _ {1}\\right) & \\dots & s ^ {\\prime} \\left(v _ {1}, t _ {B}\\right) \\\\ \\dots & \\dots & \\dots \\\\ s ^ {\\prime} \\left(v _ {B}, t _ {1}\\right) & \\dots & s ^ {\\prime} \\left(v _ {B}, t _ {B}\\right) \\end{array} \\right), \\tag {12}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.563, + 0.255, + 0.837, + 0.298 + ], + "angle": 0, + "content": "\\[\nA _ {L C V} = \\sigma \\left( \\begin{array}{c c c} s ^ {\\prime} (t _ {1}, v _ {1}) & ... & s ^ {\\prime} (t _ {1}, v _ {B}) \\\\ ... & ... & ... \\\\ s ^ {\\prime} (t _ {B}, v _ {1}) & ... & s ^ {\\prime} (t _ {B}, v _ {B}) \\end{array} \\right),\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.298, + 0.892, + 0.325 + ], + "angle": 0, + "content": "and the local caption-video distribution alignment loss is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.521, + 0.327, + 0.892, + 0.344 + ], + "angle": 0, + "content": "\\[\nL _ {L} = D _ {K L} \\left(A _ {L C V} ^ {S}, A _ {L C V} ^ {T}\\right) + D _ {K L} \\left(A _ {L V C} ^ {S}, A _ {L V C} ^ {T}\\right). \\tag {13}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.503, + 0.349, + 0.868, + 0.364 + ], + "angle": 0, + "content": "Combining \\(L_{G}\\) and \\(L_{V}\\), the cross-modality KD loss is:" + }, + { + "type": "equation", + "bbox": [ + 0.632, + 0.368, + 0.891, + 0.384 + ], + "angle": 0, + "content": "\\[\nL _ {C M} = L _ {G} + L _ {L}. \\tag {14}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.39, + 0.892, + 0.464 + ], + "angle": 0, + "content": "Our experiment in Section 4.2 shows that only when joint trained with SAB, our local video-caption distribution alignment shows its advantage, which also verifies the effectiveness of SAB. Finally, the total loss for training our model is:" + }, + { + "type": "equation", + "bbox": [ + 0.572, + 0.467, + 0.891, + 0.502 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} L = L _ {\\text {t a s k}} + \\alpha \\cdot L _ {T K D} + \\beta \\cdot L _ {S K D} \\tag {15} \\\\ + \\gamma \\cdot L _ {C M} + \\delta \\cdot L _ {S A B}, \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.513, + 0.892, + 0.543 + ], + "angle": 0, + "content": "where \\(L_{task}\\) is the task-specifical loss, and \\(\\alpha ,\\beta ,\\gamma\\) and \\(\\delta\\) are loss balance weights." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.557, + 0.634, + 0.573 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.582, + 0.892, + 0.656 + ], + "angle": 0, + "content": "We conduct comprehensive experiments on three benchmarks for video-text retrieval (video-to-text \\((v2t)\\) and text-to-video \\((t2v)\\)): MSR-VTT [38], MSVD [4] and LSMDC [32]. The metrics Recall at rank K \\((R@K)\\) are used for evaluation." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.667, + 0.805, + 0.682 + ], + "angle": 0, + "content": "4.1. Comparison with State-of-the-Arts" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.689, + 0.892, + 0.903 + ], + "angle": 0, + "content": "MSR-VTT. In Table 1, we compare the proposed model CLIPPING with eight state-of-the-art methods on MSR-VTT: TACo [39], VideoClip [37], Frozen [3], VIOLET [12], OA-Trans [35], BridgeFormer [13], ALL-in-one [34] and MDMMT [11]. It can be seen that CLIPPING significantly surpasses those large-scale video-text/image-text pre-training models for video-text retrieval. For example, our model exceeds VideoClip and Frozen by absolute \\(9.8\\%\\) t2vR@1 and \\(8.2\\%\\) t2vR@1, respectively. Note that the model MDMMT uses CLIP as its backbone, while our CLIPPING uses CLIP as the teacher. For a more efficient and edge deployable video-language retrieval model, we further compress CLIPPING into two models: \\(\\mathrm{CLIPPING}_{\\mathrm{W / oT}}\\) and \\(\\mathrm{CLIPPING^{*}}\\). \\(\\mathrm{CLIPPING}_{\\mathrm{W / oT}}\\)" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18988" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.094, + 0.089, + 0.452, + 0.281 + ], + "angle": 0, + "content": "
ModelPT DatasetsParamsR@1
TACoHT100M212M28.4
VideoClipHT100M130M30.9
FrozenC3M,W2M,COCO232M32.5
ALL-in-one-SW2M,HT100M33M33.5
VIOLETC3M,W2M198M34.5
OA-TransC3M,W2M232M35.8
BridgeFormerC3M,W2M160M37.6
ALL-in-one-BW2M,HT100M110M37.9
MDMMTC400M,AudioSet226M38.9
CLIPPING*w/o T-8.7M37.5
CLIPPINGw/o T-21.3M38.6
CLIPPING*-46.5M39.8
CLIPPING-55.0M40.6
CLIPPINGIN21K55.0M40.7
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.076, + 0.294, + 0.47, + 0.392 + ], + "angle": 0, + "content": "Table 1. Comparison with state-of-the-art models on MSR-VTT (1k split) for text-to-video retrieval. \"PT Datasets\": datasets used for pre-training the vision encoder. \"HT100M\": HowTo100M dataset [28]. \"C400M\": CLIP-400M dataset [30]. \"IN21K\": ImageNet21K dataset [9]. \"W2M\": WebVid-2M dataset [3]. \"C3M\": CC3M dataset [33]. All the methods in Table are the comparable methods with parameters less than 240M and flops less than 360G." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.402, + 0.471, + 0.567 + ], + "angle": 0, + "content": "does not use the temporal Transformer and CLIPPING* compresses the text encoder through TinyBERT [17]. It can be seen from Table 1 that the CLIPPING* based models also outperform the small model ALL-in-one-S even though they are smaller. In Fig. 5, we show the performances and Flops of these models. Among previous models, MDMMT has the best performance and ALL-in-one-S is the fastest. CLIPPING not only obtains relative \\(4.6\\%\\) performance gain over MDMMT but also is faster than All-in-one-S. Table 2 provides the details of Flops and parameters for each module of our method." + }, + { + "type": "image", + "bbox": [ + 0.134, + 0.575, + 0.419, + 0.695 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.169, + 0.702, + 0.377, + 0.716 + ], + "angle": 0, + "content": "Figure 5. Flops and Performances." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.718, + 0.47, + 0.84 + ], + "angle": 0, + "content": "Other Benchmarks. We also compare our CLIPPING with the state-of-the-art Frozen [3], MDMMT [11], NoiseEst [2] and SupportSet [29] on the LSMDC and MSVD datasets. In Table 3, CLIPPING again obtains the best performance with fewest parameters and Flops, which achieves \\(88.1\\% - 92.9\\%\\) of the performance of its teacher on the LSMDC and MSVD datasets, showing consistent improvements and generalization across different datasets." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.848, + 0.231, + 0.864 + ], + "angle": 0, + "content": "4.2. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Key Components. We provide detailed ablation study to validate each key component of our proposed method," + }, + { + "type": "table", + "bbox": [ + 0.533, + 0.089, + 0.86, + 0.159 + ], + "angle": 0, + "content": "
ModulesParametersFlops
Vision encoder (MobileViT-v2)4.5M16.8G
Temporal Transformer37.8M1.2G
Language encoder (CLIPbert)12.6M0.15G
Language encoder (TinyBERT)4.2M0.05G
" + }, + { + "type": "table_caption", + "bbox": [ + 0.5, + 0.17, + 0.89, + 0.198 + ], + "angle": 0, + "content": "Table 2. The Flops and parameters of each module in our method. We adopt 12 frames for each video." + }, + { + "type": "table", + "bbox": [ + 0.502, + 0.207, + 0.905, + 0.365 + ], + "angle": 0, + "content": "
ModelPT DatasetsVision EncodersR@1
LSMDC dataset
NoiseEstHT100MResnet152,ResNext-1016.4
SupportSetHT100MResNet-152,R(2+1)D-34IN21K
FrozenC3M,W2MSpace-Time15.0
MDMMTC400MCLIP_vision18.8
CLIPPING-MobileViT-v219.9
MSVD dataset
NoiseEstHT100MResnet152,ResNext-10120.3
SupportSetHT100MResNet-152,R(2+1)D-3428.4
FrozenC3M,W2MSpace-Time33.7
CLIPPING-MobileViT-v242.9
" + }, + { + "type": "table_caption", + "bbox": [ + 0.5, + 0.376, + 0.892, + 0.403 + ], + "angle": 0, + "content": "Table 3. Comparison with the state-of-the-art on the LSMDC and MSVD datasets." + }, + { + "type": "table", + "bbox": [ + 0.503, + 0.407, + 0.891, + 0.514 + ], + "angle": 0, + "content": "
Vision EncoderKD Typest2vR@1v2tR@1
CLIP_vision-44.542.2
MobileViTv2-25.724.5
MobileViTv2T28.827.3
MobileViTv2T,S33.032.8
MobileViTv2T,S,SAB37.636.2
MobileViTv2T,S,SAB,CMG39.639.1
MobileViTv2T,S,SAB,CMG,CML40.740.2
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.525, + 0.892, + 0.595 + ], + "angle": 0, + "content": "Table 4. Ablation study of different KD components of CLIPPING on the \\(1k\\) validation set of MSR-VTT. The first row is the results of the teacher model (Clip4clip). \\(T\\), \\(S\\), \\(SAB\\), \\(CM_{G}\\) and \\(CM_{L}\\) denote temporal KD, spatial KD, SAB KD, global cross-modality KD and local cross-modality KD, respectively." + }, + { + "type": "table", + "bbox": [ + 0.55, + 0.606, + 0.844, + 0.691 + ], + "angle": 0, + "content": "
KD Typest2vR@1v2tR@1
T, S33.032.8
T, S, OL2OL34.633.4
T, S, TAB35.134.4
T, S, SABw/o masking37.135.7
T, S, SABw/masking37.636.2
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.701, + 0.892, + 0.743 + ], + "angle": 0, + "content": "Table 5. Ablation study of different KD types on MSR-VTT (1k split). All the models are trained for 36 epochs with the same setting (see the supplementary materials for details)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.892, + 0.901 + ], + "angle": 0, + "content": "on MSR-VTT. From Table 4, we can see that without any KD, it gets extreme low accuracies. When simply adding the temporal KD and spatial KD, \\( t2vR@1 \\) and \\( v2tR@1 \\) increase significantly. When we further add SAB KD, and then global and local cross-modality KD, CLIPPING's performance rises gradually. This study shows that all these key components of CLIPPING are effective. Compared with the teacher, the full CLIPPING achieves about \\( 91.5\\% \\) and \\( 95.3\\% \\) of the performance of its teacher on \\( t2vR@1 \\) and \\( v2tR@1 \\), respectively, with the vision encoder of \\( 19.5x \\)" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18989" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.12, + 0.09, + 0.241, + 0.168 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.241, + 0.091, + 0.337, + 0.168 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.123, + 0.176, + 0.241, + 0.251 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.241, + 0.176, + 0.365, + 0.252 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.21, + 0.252, + 0.227, + 0.262 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.38, + 0.127, + 0.493, + 0.201 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.422, + 0.222, + 0.453, + 0.232 + ], + "angle": 0, + "content": "image" + }, + { + "type": "image", + "bbox": [ + 0.498, + 0.109, + 0.577, + 0.201 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.526, + 0.221, + 0.579, + 0.23 + ], + "angle": 0, + "content": "\\(\\widetilde{W}\\) \\(\\tilde{L}_{er^S}\\)" + }, + { + "type": "image", + "bbox": [ + 0.586, + 0.108, + 0.671, + 0.2 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.6, + 0.2, + 0.659, + 0.216 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.526, + 0.221, + 0.579, + 0.23 + ], + "angle": 0, + "content": "\\(\\widetilde{W}\\) \\(\\tilde{L}_{\\text{ayer}}^s\\)" + }, + { + "type": "image", + "bbox": [ + 0.675, + 0.108, + 0.755, + 0.2 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.761, + 0.108, + 0.844, + 0.2 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.65, + 0.221, + 0.788, + 0.231 + ], + "angle": 0, + "content": "\\(\\widetilde{W}\\times \\widetilde{L}\\text{ayer}^{S} = \\widetilde{L}\\text{layer}_{w}^{S}\\longrightarrow \\widetilde{L}\\text{ayer}^{T}\\)" + }, + { + "type": "image_caption", + "bbox": [ + 0.587, + 0.251, + 0.605, + 0.262 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image_caption", + "bbox": [ + 0.075, + 0.272, + 0.894, + 0.3 + ], + "angle": 0, + "content": "Figure 6. (a) Feature tensors of the student and the teacher. (b) An example to demonstrate that the teacher's features are the linear combinations of the student features. More examples are provided in the supplementary materials." + }, + { + "type": "image", + "bbox": [ + 0.165, + 0.31, + 0.371, + 0.409 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.261, + 0.414, + 0.277, + 0.425 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.382, + 0.31, + 0.59, + 0.409 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.486, + 0.414, + 0.502, + 0.425 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image", + "bbox": [ + 0.599, + 0.31, + 0.807, + 0.409 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.696, + 0.414, + 0.712, + 0.425 + ], + "angle": 0, + "content": "(c)" + }, + { + "type": "image_caption", + "bbox": [ + 0.177, + 0.433, + 0.792, + 0.448 + ], + "angle": 0, + "content": "Figure 7. Visualization of the frame-word alignments. (a) CLIP; (b) \\(\\mathrm{CLIPPING}^{-}_{\\mathrm{TAB}}\\); (c) CLIPPING." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.465, + 0.134, + 0.478 + ], + "angle": 0, + "content": "smaller." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.484, + 0.47, + 0.635 + ], + "angle": 0, + "content": "KD Types. In Table 5, we compare the proposed SAB KD with the previous OL2OL KD and TAB KD. For SAB and TAB, the same layers of the student and the teacher are used for KD, the details of which are given in the supplementary materials. From Table 5, we can see that TAB performers better than OL2OL, and our SAB outperforms TAB. In Section 3.3.2, we use masking to speed up the training. In this study, we also compare \"with masking\" and \"without masking\". The last two rows of Table 5 verify that the masking is beneficial." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.64, + 0.471, + 0.821 + ], + "angle": 0, + "content": "SAB Property. The student's layers in our SAB KD can be explained as the bases of the feature space, and each layer of the teacher is a linear combination of the bases (Section 3.2). In Fig. 6, we show that this property holds. In SAB KD, the student has 4 layers and the teacher has 12 layers. We randomly pick 3 image examples, and for each example, we select the features of one random token (in the \\(k\\) dimension in Fig. 6(a)). After training, the linear combinations \\(\\tilde{W} \\times \\tilde{L}_{\\text{layer}}^{S} = \\tilde{L}_{\\text{layer}}^{S}\\) show feature patterns very similar to the teacher's features (\\(\\tilde{L}_{\\text{layer}}^{T}\\)). This property verifies that the teacher's knowledge is fully absorbed by the student." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.826, + 0.47, + 0.903 + ], + "angle": 0, + "content": "In Fig. 7, we visualize the frame-word alignments with and without SAB KD. Fig. 7(a) is the result of the CLIP in CLIP4clip. Fig. 7(b) is the result of \\(\\mathrm{CLIPPING}^{-}_{\\mathrm{TAB}}\\), which is our model without SAB KD but with the previous TAB KD. It can be seen that the frame-word alignment with" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.464, + 0.892, + 0.585 + ], + "angle": 0, + "content": "SAB KD (Fig. 7(c)) shows clear and correct word-level attentions (e.g., \"naked\", \"child\", \"runs\" and \"through\") like CLIP. But with TAB KD, the most important words (e.g., \"naked\" and \"child\") cannot be noticed; instead, some insignificant words ([CLS] and \"a\") are activated. This study shows that the knowledge of pre-trained CLIP can be more effectively transferred to the student at the fine-tuning stage with our SAB." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.607, + 0.619, + 0.621 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.627, + 0.892, + 0.809 + ], + "angle": 0, + "content": "In this paper, we propose a novel and efficient knowledge distillation method that is specially designed for small vision-language models. It includes temporal KD, spatial KD, SAB KD and cross-modality KD. Especially, the SAB KD has the property of the student's layers being the bases of the feature space. After training, the teacher's features are the linear combinations of the bases, indicating that the student has fully absorbed the knowledge of the teacher. CLIPPING significantly outperforms the state-of-the-art and is comparable or even superior to many large pre-training models. In the feature, we will apply CLIPPING to other vision-language models for compression." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.829, + 0.667, + 0.845 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.856, + 0.892, + 0.9 + ], + "angle": 0, + "content": "We gratefully acknowledge the support of MindSpore [1], CANN (Compute Architecture for Neural Networks) and Ascend AI Processor used for this research." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18990" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.105 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.116, + 0.29, + 0.128 + ], + "angle": 0, + "content": "[1] https://www.mindspore.cn/.8" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.13, + 0.469, + 0.171 + ], + "angle": 0, + "content": "[2] Elad Amrani, Rami Ben-Ari, Daniel Rotman, and Alex Bronstein. Noise estimation using density estimation for self-supervised multimodal learning. In AAAI, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.173, + 0.469, + 0.213 + ], + "angle": 0, + "content": "[3] Max Bain, Arsha Nagrani, Gül Varol, and Andrew Zisserman. Frozen in Time: A Joint Video and Image Encoder for End-to-End Retrieval. In ICCV, 2021. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.215, + 0.469, + 0.242 + ], + "angle": 0, + "content": "[4] David Chen and William Dolan. Collecting highly parallel data for paraphrase evaluation. In ACL, 2011. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.243, + 0.469, + 0.283 + ], + "angle": 0, + "content": "[5] Defang Chen, JianPing Mei, Yuan Zhang, Can Wang, Zhe Wang, Yan Feng, and Chun Chen. Distillation with semantic calibration. In AAAI, 2021. 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.285, + 0.468, + 0.325 + ], + "angle": 0, + "content": "[6] Pengguang Chen, Shu Liu, Hengshuang Zhao, and Jiaya Jia. Distilling Knowledge via Knowledge Review. In CVPR, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.328, + 0.469, + 0.38 + ], + "angle": 0, + "content": "[7] Xianing Chen, Qiong Cao, Yujie Zhong, Jing Zhang, Shenghua Gao, and Dacheng Tao. DearKD: Data-efficient early knowledge distillation for vision transformers. In ICLR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.383, + 0.469, + 0.436 + ], + "angle": 0, + "content": "[8] Yinpeng Chen, Xiyang Dai, Dongdong Chen, Mengchen Liu, Xiaoyi Dong, Lu Yuan, and Zicheng Liu. Mobile-Former: Bridging MobileNet and Transformer. In CVPR, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.439, + 0.469, + 0.479 + ], + "angle": 0, + "content": "[9] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. ImageNet: A large-scale hierarchical image database. In CVPR, 2009. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.481, + 0.469, + 0.563 + ], + "angle": 0, + "content": "[10] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.565, + 0.469, + 0.618 + ], + "angle": 0, + "content": "[11] Maksim Dzabraev, Maksim Kalashnikov, Stepan Komkov, and Aleksandr Petiushko. MDMMT: Multidomain Multimodal Transformer for Video Retrieval. In CVPR, 2021. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.621, + 0.469, + 0.676 + ], + "angle": 0, + "content": "[12] Tsu-Jui Fu, Linjie Li, Zhe Gan, Kevin Lin, William Yang Wang, Lijuan Wang, and Zicheng Liu. VIOLET: End-to-End Video-Language Transformers with Masked Token Modeling. In arXiv:2111.1268, 2021. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.677, + 0.469, + 0.717 + ], + "angle": 0, + "content": "[13] Yuying Ge, Yixiao Ge, Xihui Liu, Dian Li, Ying Shan, Xiaohu Qie, and Ping Luo. Bridgeformer: Bridging video-text retrieval with multiple choice questions. In CVPR, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.719, + 0.469, + 0.773 + ], + "angle": 0, + "content": "[14] Jianyuan Guo, Kai Han, Han Wu, Yehui Tang, Xinghao Chen, Yunhe Wang, and Chang Xu. CMT: Convolutional Neural Networks Meet Vision Transformers. In CVPR, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.776, + 0.469, + 0.815 + ], + "angle": 0, + "content": "[15] Geoffrey Hinton, Oriol Vinyals, and Jeff Dean. Distilling the Knowledge in a Neural Network. In arXiv preprint arXiv:1503.02531, 2015. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.817, + 0.469, + 0.872 + ], + "angle": 0, + "content": "[16] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In ICML, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[17] Xiaoqi Jiao, Yichun Yin, Lifeng Shang, Xin Jiang, Xiao Chen, Linlin Li, Fang Wang, and Qun Liu. TinyBERT:" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.116, + 0.469, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.119 + ], + "angle": 0, + "content": "Distilling BERT for Natural Language Understanding. In EMNLP, 2020. 1, 2, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.121, + 0.892, + 0.148 + ], + "angle": 0, + "content": "[18] Yoon Kim and Alexander M. Rush. Sequence-level knowledge distillation. In EMNLP, 2016. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.149, + 0.892, + 0.203 + ], + "angle": 0, + "content": "[19] Pavan Kumar, Anasosalu Vasu, James Gabriel, Jeff Zhu, Oncel Tuzel, and Anurag Ranjan. An Improved One millisecond Mobile Backbone. In arXiv preprint arXiv:2206.04040, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.205, + 0.892, + 0.259 + ], + "angle": 0, + "content": "[20] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. BLIP: Bootstrapping Language-Image Pre-training for Unified Vision-Language Understanding and Generation. In arXiv preprint arXiv:2201.12086, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.261, + 0.892, + 0.315 + ], + "angle": 0, + "content": "[21] Junnan Li, Ramprasaath Selvaraju, Akhilesh Gotmare, Shafiq Joty, Caiming Xiong, and Steven Chu Hong Hoi. Align before fuse: Vision and language representation learning with momentum distillation. In NIPS, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.317, + 0.892, + 0.37 + ], + "angle": 0, + "content": "[22] Yanyu Li, Geng Yuan, Yang Wen, Eric Hu, Georgios Evangelidis, Sergey Tulyakov, Yanzhi Wang, and Jian Ren. EfficientFormer: Vision Transformers at MobileNet Speed. In arXiv:2206.01191, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.372, + 0.892, + 0.425 + ], + "angle": 0, + "content": "[23] Sihao Lin, Hongwei Xie, Bing Wang, Kaicheng Yu, Xiaojun Chang, Xiaodan Liang, and Gang Wang. Knowledge Distillation via the Target-aware Transformer. In CVPR, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.428, + 0.892, + 0.468 + ], + "angle": 0, + "content": "[24] Yifan Liu, Chunhua Shen, Changqian Yu, and Jingdong Wang. Efficient Semantic Video Segmentation with Per-Frame Inference. In ECCV, 2020. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.47, + 0.892, + 0.523 + ], + "angle": 0, + "content": "[25] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin Transformer: Hierarchical Vision Transformer using Shifted Windows. In ICCV, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.526, + 0.892, + 0.579 + ], + "angle": 0, + "content": "[26] Huaishao Luo, Lei Ji, Ming Zhong, Yang Chen, Wen Lei, Nan Duan, and Tianrui Li. Clip4clip: An empirical study of clip for end to end video clip retrieval. arXiv preprint arXiv:2104.08860, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.581, + 0.892, + 0.621 + ], + "angle": 0, + "content": "[27] Sachin Mehta and Mohammad Rastegari. Separable Self-attention for Mobile Vision Transformers. In ICLR, 2022. 1, 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.623, + 0.892, + 0.69 + ], + "angle": 0, + "content": "[28] Antoine Miech, Dimitri Zhukov, Jean-Baptiste Alayrac, Makarand Tapaswi, Ivan Laptev, and Josef Sivic. HowTo100M: Learning a Text-Video Embedding by Watching Hundred Million Narrated Video Clips. In ICCV, 2019. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.692, + 0.892, + 0.747 + ], + "angle": 0, + "content": "[29] Mandela Patrick, Po-Yao Huang, Yuki Asano, Florian Metze, Alexander Hauptmann, João Henriques, and Andrea Vedaldi. Support-set bottlenecks for video-text representation learning. In ICLR, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.749, + 0.892, + 0.815 + ], + "angle": 0, + "content": "[30] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, and Jack Clark. Learning transferable visual models from natural language supervision. ICML, 2021. 1, 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.818, + 0.892, + 0.871 + ], + "angle": 0, + "content": "[31] Maithra Raghu, Thomas Unterthiner, Simon Kornblith, Chiyuan Zhang, and Alexey Dosovitskiy. Do vision transformers see like convolutional neural networks? In NIPS, 2021. 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.873, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[32] Anna Rohrbach, Marcus Rohrbach, Niket Tandon, and Bernt Schiele. A dataset for Movie Description. In CVPR, 2015. 6" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.518, + 0.957 + ], + "angle": 0, + "content": "18991" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.146 + ], + "angle": 0, + "content": "[33] Piyush Sharma, Nan Ding, Sebastian Goodman, and Radu Soricut. Conceptual captions: A cleaned, hypernymed, image alt-text dataset for automatic image captioning. In ACL, 2018. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.149, + 0.47, + 0.218 + ], + "angle": 0, + "content": "[34] Alex Jinpeng Wang, Yixiao Ge, Rui Yan, Yuying Ge, Xudong Lin, Guanyu Cai, Jianping Wu, Ying Shan, Xiaohu Qie, and Mike Zheng Shou. All in one: Exploring unified video-language pre-training. In arXiv preprint arXiv:2203.07303, 2022. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.22, + 0.469, + 0.274 + ], + "angle": 0, + "content": "[35] Jinpeng Wang, Yixiao Ge, Guanyu Cai, Rui Yan, Xudong Lin, Ying Shan, Xiaohu Qie, and Mike Zheng Shou. Object-aware video-language pre-training for retrieval. In CVPR, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.277, + 0.469, + 0.33 + ], + "angle": 0, + "content": "[36] Zekun Wang, Wenhui Wang, Haichao Zhu, Ming Liu, Bing Qin, and Furu Wei. Distilled dual-encoder model for vision-language understanding. In arXiv preprint arXiv:2112.08723, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.333, + 0.469, + 0.401 + ], + "angle": 0, + "content": "[37] Hu Xu, Gargi Ghosh, Po-Yao Huang, Dmytro Okhonko, Armen Aghajanyan, Florian Metze, Luke Zettlemoyer, and Christoph Feichtenhofer. Videoclip: Contrastive pre-training for zero-shot video-text understanding. arXiv preprint arXiv:2104.08860, 2021. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.403, + 0.469, + 0.445 + ], + "angle": 0, + "content": "[38] Jun Xu, Tao Mei, Ting Yao, and Yong Rui. MSR-VTT: A Large Video Description Dataset for Bridging Video and Language. 2016. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.447, + 0.469, + 0.487 + ], + "angle": 0, + "content": "[39] Jianwei Yang, Yonatan Bisk, and Jianfeng Gao. Taco: Token-aware cascade contrastive learning for video-text alignment. In ICCV, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.489, + 0.469, + 0.53 + ], + "angle": 0, + "content": "[40] Weihao Yu, Mi Luo, Pan Zhou, Chenyang Si, Yichen Zhou, Xinchao Wang, Jiashi Feng, and Shuicheng Yan. Metaformer is actually what you need for vision. In CVPR, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.532, + 0.469, + 0.587 + ], + "angle": 0, + "content": "[41] Li Yuan, Yunpeng Chen, Tao Wang, Weihao Yu, Yujun Shi, Zi-Hang Jiang, Francis E.H. Tay, Jiashi Feng, and Shuicheng Yan. Tokens-to-Token ViT: Training Vision Transformers From Scratch on ImageNet. In ICCV, 2021. 5" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.587 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.956 + ], + "angle": 0, + "content": "18992" + } + ] +] \ No newline at end of file diff --git a/2023/CLIPPING_ Distilling CLIP-Based Models With a Student Base for Video-Language Retrieval/1eb2010b-78e6-492d-9404-5d49a6b93ef0_origin.pdf b/2023/CLIPPING_ Distilling CLIP-Based Models With a Student Base for Video-Language Retrieval/1eb2010b-78e6-492d-9404-5d49a6b93ef0_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..1ad4b5f05e5247c4a9620fb86cbf25f580564b2d --- /dev/null +++ b/2023/CLIPPING_ Distilling CLIP-Based Models With a Student Base for Video-Language Retrieval/1eb2010b-78e6-492d-9404-5d49a6b93ef0_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:fa14d0eaf62138abd840e64a35c0508af415d0748667f4e8ceb59cc8b3e2466a +size 3658879 diff --git a/2023/CLIPPING_ Distilling CLIP-Based Models With a Student Base for Video-Language Retrieval/full.md b/2023/CLIPPING_ Distilling CLIP-Based Models With a Student Base for Video-Language Retrieval/full.md new file mode 100644 index 0000000000000000000000000000000000000000..78f9c3ad5a6bec7f8cb82f4f168b63d81bece2f8 --- /dev/null +++ b/2023/CLIPPING_ Distilling CLIP-Based Models With a Student Base for Video-Language Retrieval/full.md @@ -0,0 +1,353 @@ +# CLIPPING: Distilling CLIP-Based Models with a Student Base for Video-Language Retrieval + +Renjing Pei, Jianzhuang Liu, Weimian Li, Bin Shao, Songcen Xu, Peng Dai, Juwei Lu, Youliang Yan Huawei Noah's Ark Lab + +{peirenjing,liu.jianzhuang,liweimian,shaobin3, xusongcen,peng.dai,juwei.lu,yanyouliang}@huawei.com + +# Abstract + +Pre-training a vision-language model and then fine-tuning it on downstream tasks have become a popular paradigm. However, pre-trained vision-language models with the Transformer architecture usually take long inference time. Knowledge distillation has been an efficient technique to transfer the capability of a large model to a small one while maintaining the accuracy, which has achieved remarkable success in natural language processing. However, it faces many problems when applying KD to the multi-modality applications. In this paper, we propose a novel knowledge distillation method, named CLIPPING1, where the plentiful knowledge of a large teacher model that has been fine-tuned for video-language tasks with the powerful pre-trained CLIP can be effectively transferred to a small student only at the fine-tuning stage. Especially, a new layer-wise alignment with the student as the base is proposed for knowledge distillation of the intermediate layers in CLIPPING, which enables the student's layers to be the bases of the teacher, and thus allows the student to fully absorb the knowledge of the teacher. CLIPPING with MobileViT-v2 as the vision encoder without any vision-language pre-training achieves $88.1\% - 95.3\%$ of the performance of its teacher on three video-language retrieval benchmarks, with its vision encoder being $19.5x$ smaller. CLIPPING also significantly outperforms a state-of-the-art small baseline (ALL-in-one-B) on the MSR-VTT dataset, obtaining relatively $7.4\%$ performance gain, with $29\%$ fewer parameters and $86.9\%$ fewer flops. Moreover, CLIPPING is comparable or even superior to many large pre-training models. + +# 1. Introduction + +Recently, pre-training a vision-language model and then fine-tuning it on downstream tasks are a popular paradigm [12, 16, 20, 21, 30]. Pre-trained vision-language models (PVLMs) have achieved great success in many multi-modality tasks (e.g., video-text retrieval). However, PVLMs with the Transformer architecture (especially the vision stream) usually consume a huge amount of computation, making them difficult to be deployed on edge devices such as mobile phones. Recent small models [8, 14, 19, 22, 27, 40] show that combining Convolutional Neural Networks (CNNs) and Transformers as a hybrid architecture gets the best of both architectures, but the overall performance of these works is still far from satisfactory when compared to large pre-training models. Apparently, knowledge distillation (KD) [15] is an efficient technique to transfer the capability of a large model to a small one while maintaining the accuracy, which has achieved remarkable success in natural language processing (NLP) [17, 18]. In NLP, the knowledge distillation methods are usually performed at both the pre-training and the fine-tuning stages. However, the collection of the pre-training data for the pre-training knowledge distillation cost huge manpower in multi-modality applications (e.g., the pre-training dataset of CLIP is hard to obtain). Therefore, an efficient knowledge distillation method need to be explored for multi-modality applications. To this end, we propose a novel knowledge distillation method, named CLIPPING, where the plentiful knowledge of a large teacher model Clip4clip [26] that has been fine-tuned for video-language tasks with the powerful pre-trained CLIP [30] can be effectively transferred to a small student only at the fine-tuning stage. The contributions of CLIPPING are summarized below: + +1) We introduce an efficient approach to distill both the vision knowledge and the cross-modality knowledge from teacher to a small model at the fine-tuning stage. The resulting model shows strong performance on multi-modality + +![](images/8097da16ed75c627f4843d5d3aeff5d1ba8aa878473aad9c97b44721c8dd0c8b.jpg) +(a) + +![](images/da544192d760903d94e60854f38c3d572730c43c338e865beb70fbb1bd558907.jpg) +(b) + +![](images/bfeb05a90cd2bf07c30252fef956f0e9ec8536a03e4fe8300eca4762e94cad2a.jpg) +(c) +Figure 1. (a) and (b) are previous methods of intermediate features' knowledge distillation. $L_{j}^{T}$ , $j = 1,2,\dots,M$ , and $L_{i}^{S}$ , $i = 1,2,\dots,N$ , are the outputs of the $j^{th}$ and $i^{th}$ layers of the teacher and the student, respectively. OL2OL is the most common One-Layer-to-One-Layer KD, where it selects some layers of the teacher and then distills the layers into the student one by one (e.g., from $L_{j}^{T}$ to $L_{i}^{S}$ ). Teacher-As-Base (TAB) KD uses all the teacher's layers to supervise each layer in the student, where each layer of the student (e.g., $L_{i}^{S}$ ) learns the knowledge from all the selected layers of the teacher $\left(\sum_{k=1}^{M} w_{ik}' L_{k}^{T}\right)$ , where $w_{ik}'$ , $k = 1,2,\dots,M$ , are the knowledge selection weights with $\sum_{k=1}^{M} w_{ik}' = 1$ . (c) Our Student-As-Base (SAB) KD enables each of the teacher's layers to pass its knowledge to all the student's layers (e.g., $L_{j}^{T} \rightarrow \sum_{k=1}^{N} w_{jk} L_{k}^{S}$ , $\sum_{k=1}^{N} w_{jk} = 1$ ). More detailed analysis is provided in Section 3.1. + +text-video retrieval task. + +2) We propose a new layer-wise alignment scheme, called Student-As-Base (SAB), for knowledge distillation of the intermediate layers from the Transformer to the CNN in CLIPPING, where the student's layers can be regarded as the bases of the teacher's feature space, forcing the student model to full absorb the knowledge of the teacher. In our experience, the SAB layer-wise alignment significantly surpasses the previous knowledge distillation methods. We believe it will become a popular paradigm for knowledge distillation of intermediate features. +3) We present an effective cross-modal knowledge distillation, which includes knowledge from both the global and local video-caption distributions. We use the video-caption distributions of the teacher to guide the training of the student. Besides, the student can also benefit from the powerful pre-trained CLIP that obtains certain local frame-word attention ability via learning from massive data. Experiments show that this pre-training knowledge can be effectively transferred to the student when jointly trained with our SAB KD only at the fine-tuning stage. +4) CLIPPING with MobileViT-v2 [27] as the vision encoder without any vision-language pre-training achieves $91.5\% - 95.3\%$ of the performance of its teacher on MSR-VTT video-language retrieval benchmark, with its vision encoder being $19.5 \times$ smaller. CLIPPING also significantly outperforms a state-of-the-art small baseline (ALL-in-one-B) on the MSR-VTT dataset, obtaining relatively $7.4\%$ performance gain, with $29\%$ fewer parameters and $86.9\%$ fewer flops. Moreover, CLIPPING is comparable or even + +superior to many large pre-training models. + +# 2. Relative Work + +Vision-Language Retrieval. Learning from web-collected image-text data, large-scale Vision-Language Pretraining (VLP) models such as CLIP [30] have recently demonstrated great success across various downstream tasks. Nowadays, models such as Clip4clip [26] and MDMMT [11] extended from the pre-trained model CLIP keep appearing for vision-language retrieval. There are also some end-to-end trainable models [3, 37], which are designed to take advantage of both large-scale image and video captioning datasets. All-in-one [34] is the first work to consider both efficiency and performance for video-language retrieval tasks. It introduces a unified backbone that enables the representation learning of both video-text multimodal and unimodal inputs. + +Knowledge Distillation. Knowledge distillation (KD) [15] is a simple yet effective technique to improve the performance of a learning model. Earlier works transfer knowledge embedded in the "logits" learned in a large teacher model to a small student model without sacrificing much performance. Recent works [5, 6] use multiple layers of the teacher to supervise each layer in the student, where each layer of the student learns the knowledge from multiple layers of the teacher. [23] proposes a target-aware Transformer and enables the student to mimic each spatial component of the teacher in each distilled layer to boost the student's performance. For multi-modality KD, [36] designs a fusion-encoder model as the teacher and introduces cross-modal attention knowledge to train the dual-encoder student model. The distillation objective is applied at both the pre-training and the fine-tuning stages and helps the dual-encoder model learn interactions of different modalities. TinyBert [17] also introduces a two-stage learning framework that performs Transformer distillation at both the pre-training and the task-specific fine-tuning stages. In this paper, we also focus on KD for transferring the knowledge of a pre-training vision-language model to a small one but only at the fine-tuning stage. + +Transformers and CNNs. Usually, the pre-training model is a Transformer-based model [10, 25] and the small model is a CNN-based or hybrid architecture [19, 27]. For KD, most existing methods distill knowledge either from a Transformer to another Transformer (T2T) or from a CNN to another CNN (C2C) that computes the loss in the OL2OL or TAB style (see Fig. 1). To the best of our knowledge, KD from a Transformer to a CNN (T2C) has not been explored yet. The most related work [7] distills knowledge from a CNN to a Transformer (C2T) in the OL2OL way. However, previous works (e.g., [31]) have investigated the internal representation structures of ViTs and CNNs, and found striking differences between the two models, such as + +![](images/9d827e96f3b031b4aa6ccfb205aa5f203433525ff93a0a1ad43f822c41195bae.jpg) +Figure 2. Overview of CLIPPING. There are mainly four knowledge distillation (KD) parts: (1) Temporal KD. (2) Spatial KD. (3) SAB (Student-As-Base) KD. (4) Cross-modality KD. The CLIP's vision encoder is a Transformer. The temporal Transformer is an optional module for CLIPPING. + +ViTs having highly similar representations throughout the model's layers, while CNNs showing obvious distinction of representations between lower and higher layers [31]. Considering such striking differences between ViTs and CNNs, the previous KD (OL2OL and TAB) may not be good distillation ways for our T2C task, which is verified by our experiments. + +# 3. Methodology + +We propose an efficient approach to distill both the vision knowledge and the cross-modality knowledge from Clip4clip to a small model with MobileViT-v2 as its vision encoder for text-video retrieval task, which can get the best from both sides: powerful multimodal representations of CLIP and efficient mobile vision Transformer of MobileViT-v2. The overall architecture is illustrated in Fig. 2. + +# 3.1. Analysis of TAB and SAB + +We first show the advantage of our SAB KD over traditional TAB KD. For simplicity, suppose $M = N = 2$ . Then, for SAB, + +$$ +\left\{ \begin{array}{l} L _ {1} ^ {T} = w _ {1 1} L _ {1} ^ {S} + w _ {1 2} L _ {2} ^ {S}, w _ {1 1} + w _ {1 2} = 1 \\ L _ {2} ^ {T} = w _ {2 1} L _ {1} ^ {S} + w _ {2 2} L _ {2} ^ {S}, w _ {2 1} + w _ {2 2} = 1 \end{array} , \right. \tag {1} +$$ + +and for TAB, + +$$ +\left\{ \begin{array}{l} L _ {1} ^ {S} = w _ {1 1} ^ {\prime} L _ {1} ^ {T} + w _ {1 2} ^ {\prime} L _ {2} ^ {T}, w _ {1 1} ^ {\prime} + w _ {1 2} ^ {\prime} = 1 \\ L _ {2} ^ {S} = w _ {2 1} ^ {\prime} L _ {1} ^ {T} + w _ {2 2} ^ {\prime} L _ {2} ^ {T}, w _ {2 1} ^ {\prime} + w _ {2 2} ^ {\prime} = 1 \end{array} \right. \tag {2} +$$ + +Obviously, in Eq. 1, the knowledge of the teacher $L_1^T$ or $L_2^T$ is the linear combination of $L_1^S$ and $L_2^S$ . In other words, $L_1^S$ and $L_2^S$ are the bases of $L_1^T$ and $L_2^T$ of the teacher. In Eq. 2, it can be seen that the teacher's $L_1^T$ and $L_2^T$ are as the bases for $L_1^S$ and $L_2^S$ . Eq. 2 can be converted to: + +$$ +L _ {1} ^ {T} = A \cdot L _ {1} ^ {S} + B \cdot L _ {2} ^ {S}, L _ {2} ^ {T} = C \cdot L _ {1} ^ {S} + D \cdot L _ {2} ^ {S}, \tag {3} +$$ + +where $A = \frac{w_{22}^{\prime}}{w_{11}^{\prime}w_{22}^{\prime} - w_{12}^{\prime}w_{21}^{\prime}}$ , $B = \frac{-w_{12}^{\prime}}{w_{11}^{\prime}w_{22}^{\prime} - w_{12}^{\prime}w_{21}^{\prime}}$ , $C =$ + +$\frac{-w_{21}^{\prime}}{w_{11}^{\prime}w_{22}^{\prime} - w_{12}^{\prime}w_{21}^{\prime}}$ and $D = \frac{w_{11}^{\prime}}{w_{11}^{\prime}w_{22}^{\prime} - w_{12}^{\prime}w_{21}^{\prime}}$ , which looks like Eq. 1. It seems that in this case, the student's $L_1^S$ and $L_2^S$ could also serve as the bases for $L_1^T$ and $L_2^T$ . However, (i) $A + B \neq 1$ and $C + D \neq 1$ , but it is usually required in the previous KD that all the coefficients in each equation should be summed to be 1 [5,6,23]; (ii) when learning with TAB KD in Eq. 2, the coefficients A, B, C and D are uncontrollable, which may be arbitrary large during learning. + +For the reasons above, although it seems that SAB and TAB have the same matrix representation, TAB is unable to get a similar result as SAB and more likely to get to a local minimum during training. Therefore, TAB KD is not equivalent to our SAB KD. When well trained, the student with this SAB KD fully learns the knowledge of the teacher because the teacher's $L_{i}^{T}$ can be obtained by the linear combination of the student's $L_{i}^{S}$ , $i = 1,2,\dots,N$ . Our experiment also shows this property (see Section 4.2). + +# 3.2. Preliminaries + +Given a batch of videos $V$ and captions $T$ , Clip4clip learns a similarity function $s(v_{i},t_{j})$ to calculate the similar- + +![](images/d0412b3b6617f2558684dd17b058443e37462ad29a8cf2125df57a90e1172e52.jpg) + +![](images/b3485e9995b2dc2b7f0637287b4d2b50eb2976604aabbc3a916d59d2ef28b4f3.jpg) +Figure 3. Reshaping the features of (a) the student $L_{k}^{S}$ and (b) the teacher $L_{k}^{T}$ to the same shape. (c) Similarity Computation. + +ity between a video $v_{i} \in V$ and a caption $t_j \in T$ . It obtains the frames' representation and the caption representation in a multi-modal embedding space via CLIP. The frames' representation is denoted as $Z_{i} = \{z_{i1}, z_{i2}, \dots, z_{in}\}$ , where $n$ is the number of frames in $v_{i}$ and $z_{ik}$ ( $k = 1, 2, \dots, n$ ) is the class token from the output of the CLIP's vision encoder corresponding to the $k^{th}$ input frame. The caption representation is denoted as $W_{j} = \{w_{j1}, w_{j2}, \dots, w_{jm}\}$ , where $m$ is the number of words in $t_j$ (including [CLS] and [SEP]) and $w_{jm}$ is used as the representation of $t_j$ in Clip4clip. Clip4clip uses a temporal Transformer encoder to obtain the sequential feature, denoted as $Z_{i}' = \{z_{i1}', z_{i2}', \dots, z_{in}'\} = TemporalTransformer(Z_{i})$ . Then, the mean pooling is used to aggregate the features of all frames to obtain the video representation, $z_{i} = \frac{1}{n} \sum_{k=1}^{n} z_{ik}'$ . Finally, the similarity functions $s(v_i, t_j)$ and $s(t_j, v_i)$ are defined as: + +$$ +s (v _ {i}, t _ {j}) = w _ {j m} ^ {\top} z _ {i}, \quad s (t _ {i}, v _ {j}) = z _ {j} ^ {\top} w _ {i m}. \tag {4} +$$ + +Nowadays, models such as Clip4clip based on the pretrained model CLIP have been applied to many tasks. They pursue better performance with CLIP but have a heavy vision encoder, which makes them difficult to be deployed on edge devices such as mobile phones. Since MobileViT-v2 [27] is a light-weight and mobile-friendly hybrid network, we employ it as the vision encoder of the student model and maintain the original text encoder and temporal Transformer of Clip4clip. Note that the temporal Transformer is an optional module for the CLIPPING model structure, and the text encoder $(CLIP_{bert})$ can be further compressed through the existing methods (details are shown in the Section 4). + +# 3.3. CLIPPING + +To distill multimodal knowledge from Clip4clip to the MobileViT-v2-based model, we use four kinds of knowl- + +edge transfer: 1) temporal KD, 2) spatial KD, 3) SAB KD, and 4) cross-modality KD. + +# 3.3.1 Temporal and Spatial Knowledge Distillation + +The temporal KD is motivated by the previous finding that aligning multi-frame dependency from the teacher to the student can enhance the performance of the student [24], which encodes the multi-frame dependency into a latent embedding by using a recurrent unit ConvLSTM. In our architecture, multi-frame dependency can be modeled naturally through the temporal Transformer (Temporal Transformer) in Fig. 2, so we define the temporal KD loss as: + +$$ +L _ {T K D} = \frac {1}{B} \sum_ {i = 1} ^ {B} D _ {K L} \left(Z _ {i} ^ {\prime S}, Z _ {i} ^ {\prime T}\right), \tag {5} +$$ + +where $B$ is the batch size, the superscripts $S$ and $T$ denote the student and the teacher, respectively, and $D_{KL}()$ is the KL divergence loss function. In addition to imitating the behavior of CLIP for each frame, we use spatial KD to align the spatial embeddings of the student and the teacher by: + +$$ +L _ {S K D} = \frac {1}{B} \sum_ {i = 1} ^ {B} D _ {K L} \left(Z _ {i} ^ {S}, Z _ {i} ^ {T}\right). \tag {6} +$$ + +The Temporal Transformer is an optional module for the CLIPPING model structure in our method. In this case, $Z_{i}^{\prime S}$ is exactly equal to $Z_{i}^{S}$ in TKD. + +# 3.3.2 SAB Knowledge Distillation + +Recently, it is discovered that distilling intermediate features is more effective. So in addition to the spatial and temporal KD, we also force the vision encoder of the student to mimic the intermediate layers of the teacher. In the traditional OL2OL KD, some layers of the teacher are selected + +![](images/5aaf3acfbac4445d947fd6ee4f54a65e2b4658a636631e79e200eeb3dbdc4c65.jpg) + +![](images/8e174f102370c416bcd2396d5ae76916e10d293ca87bd242400739efdd6d36ea.jpg) + +![](images/b671099dc4d7918037e0e99b464aab4635924d5e4a39a59575ec8ba9b240699c.jpg) +Figure 4. (a) Masks obtained with Eq. 7, where all the $L - 1$ masks along the $k$ dimension are the same. (b) Masking the similarity tensor $W$ . (c) SAB alignment. $L_{SAB}$ is the SAB loss. + +and distilled into the student one by one [17]. However, it is difficult to find the optimal correspondence between the student and the teacher. The recent TAB works (e.g., [5]) still cannot help the student learn enough knowledge from the teacher when there is a large architecture gap between them, which is verified in our experiments. For our task, the student (MobileViT-v2) is a hybrid architecture, which has convolution layers in the shallow stages and separable self-attention layers in the later blocks. And a separable self-attention layer differs significantly from the traditional self-attention in two ways: 1) it does not learn an explicit attention map; 2) its input and output are still 3D, which are more similar to CNN features. Except for the separable self-attention in the later blocks, the shallow stages with convolution layers undoubtedly have a huge difference from the Transformer structure in CLIP, which has been investigated in previous works [31,41]. Therefore, we design SAB layer-wise alignment for KD of the intermediate layers from the Transformer to the CNN. + +SAB Property. As shown in Fig. 2, the student's layers in our SAB KD can be explained as the bases of the feature space, and each layer of the teacher is a linear combination of the bases. After training, if the student's layers do ideally form the bases, the knowledge of the teacher's layers is learned completely by the student. Our experiment in Section 4.2 shows that the teacher's features in different layers can be well recovered from the features of the student's layers. Next, we describe how to implement SAB KD. + +Feature Reshaping and Similarity Computation. Let the layers of the teacher and the student be $\text{Layer}^T = [L_1^T, L_2^T, \ldots, L_M^T]$ and $\text{Layer}^S = [L_1^S, L_2^S, \ldots, L_N^S]$ (usually $M > N$ ), respectively. $L_k^T \in R^{D \times L}$ is a 2D tensor, where $D$ is the dimensionality of the token feature + +and $L$ is the number of tokens, while $L_{k}^{S}\in R^{C_{k}\times H_{k}\times W_{k}}$ is a 3D tensor, where $C_k$ , $H_{k}$ and $W_{k}$ are the channel number, height and width of the $k^{th}$ CNN layer's feature. We calculate a similarity tensor $W$ between $L_{k}^{S}$ and $L_{k}^{T}$ which need to be reshaped first as shown in Figs. 3(a) and (b), respectively. For $L_{k}^{S}$ , we first apply a learnable linear operator and then do pixel reshuffle on the result, obtaining $\bar{L}_k^S$ , $k = 1,2,\dots,N$ . Next, $\bar{L}_k^S$ is reshaped to $\tilde{L}_k^S$ . As for $L_{k}^{T}$ , we remove the class token and obtain a $(L - 1)\times D$ tensor, denoted as $\tilde{L}_k^T$ , $k = 1,2,\dots,M$ . Finally, all $\tilde{L}_k^S$ , $k = 1,2,\dots,N$ , and $\tilde{L}_k^T$ , $k = 1,2,\dots,M$ , are represented as $\tilde{L}_{\mathrm{er}}^{S} = [\tilde{L}_{1}^{S},\tilde{L}_{2}^{S},\dots,\tilde{L}_{N}^{S}]$ and $\tilde{L}_{\mathrm{er}}^{T} = [\tilde{L}_{1}^{T},\tilde{L}_{2}^{T},\dots,\tilde{L}_{M}^{T}]$ , respectively. After this reshaping, as shown in Fig. 3(c), we are able to measure the similarities between the intermediate layers of the student and the teacher through $W = \tilde{L}_{\mathrm{er}}^{T}\times (\tilde{L}_{\mathrm{er}}^{S})^{\top}$ , $W\in$ $R^{(L - 1)\times M\times N}$ . + +Masking and SAB Alignment. To speed up the training procedure, we design SAB KD with sequential masks, which follow this formula: + +$$ +\operatorname {m a s k} _ {(i, j)} = \left\{ \begin{array}{l} 0, \text {i f} j > n _ {0} \text {a n d} i \leq m _ {0} \\ 0, \text {i f} j > n _ {1} \text {a n d} m _ {0} < i \leq m _ {1} \\ 1, \text {e l s e} \end{array} , \right. \tag {7} +$$ + +where $i = 1,2,\dots,M$ , $j = 1,2,\dots,N$ , $1\leq m_0 < m_1 < M$ , $1\leq n_0 < n_1 < N$ . And the final attention mask $\mathsf{Mask}\in R^{(L - 1)\times M\times N}$ is composed of $L - 1$ same masks, Fig. 4(a) shows one example with $m_0 = 1,n_0 = 2,m_1 = 3,n_1 = 3$ . $\mathsf{mask}_{(i,j)} = 0$ means to mask all the elements along the $k$ dimension (Fig. 4(a)) at $(i,j)$ in $W$ , while $\mathsf{mask}_{(i,j)} = 1$ means to maintain their similarities, as shown in Fig. 4(b). This masking is based on our experimental finding: The student's lower-level layers + +should learn from the teacher's lower-layers, while the student's higher layers should learn from all the teacher's layers. In Fig. 4(b), we also normalize the masks by performing softmax along the $j$ dimension on each row, obtaining $\tilde{W} = \sigma (W\circ \mathsf{Mask})$ , where $\sigma$ is the softmax function and $\circ$ denotes the element-wise product. Guided by $\tilde{W}$ , as shown in Fig. 4(c), the weighted student layers are calculated as: $\tilde{L a r e r}_{w}^{S} = \tilde{W}\times \tilde{L a r e r}^{S}$ . Each component of $\tilde{L a r e r}_{w}^{S}$ is the linear combination of $N$ student layers' features. Finally, the SAB KD loss is defined as: + +$$ +L _ {S A B} = D _ {K L} \left(\tilde {L a y e r} _ {w} ^ {S}, \tilde {L a y e r} ^ {T}\right). \tag {8} +$$ + +# 3.3.3 Cross-Modality Knowledge distillation + +To further adapt the student's features to the teacher's multimodal feature space, we employ the teacher's knowledge of cross-modality distributions to guide the training of the student. Specifically, we characterize the cross-modal distributions from two perspectives, global video-caption distributions of the teacher and local video-caption distributions of CLIP. + +Global Video-Caption Distribution Alignment. We consider both the video-to-caption distribution $A_{GVC}$ and the caption-to-video distribution $A_{GCV}$ , the elements $(s(v_i, t_j)$ and $s(t_j, v_i))$ of which are defined via the similarities obtained by Eq. 4: + +$$ +A _ {G V C} = \sigma \left( \begin{array}{c c c} s (v _ {1}, t _ {1}) & ... & s (v _ {1}, t _ {B}) \\ ... & ... & ... \\ s (v _ {B}, t _ {1}) & ... & s (v _ {B}, t _ {j}) \end{array} \right), +$$ + +$$ +A _ {G C V} = \sigma \left( \begin{array}{c c c} s (t _ {1}, v _ {1}) & \dots & s (t _ {1}, v _ {B}) \\ \dots & \dots & \dots \\ s (t _ {B}, v _ {1}) & \dots & s (t _ {B}, v _ {B}) \end{array} \right). \tag {9} +$$ + +, where $\sigma$ is the softmax function along the first dimension of the similarity matrixs. Finally, the global video-caption distribution alignment loss is defined as: + +$$ +L _ {G} = D _ {K L} \left(A _ {G V C} ^ {S}, A _ {G V C} ^ {T}\right) + D _ {K L} \left(A _ {G C V} ^ {S}, A _ {G C V} ^ {T}\right). \tag {10} +$$ + +Local Video-Caption Distribution Alignment. CLIP uses text prompts (such as "A picture of a (") for zero-shot image classification. CLIP fills them with different words (e.g., "cat" and "dog") and results in different captions (e.g., "A picture of a cat" and "A picture of a dog"). It can match the captions to the corresponding images, showing some image-word alignment ability. We transfer this pre-training knowledge to the student through a local frame-word alignment as follows. Recall that $Z_{i}^{\prime} = \{z_{i1}^{\prime}, z_{i2}^{\prime}, \dots, z_{in}^{\prime}\}$ and $W_{j} = \{w_{j1}, w_{j2}, \dots, w_{jm}\}$ are the features of the video $v_{i}$ and the caption $t_{j}$ , respectively, with $n$ being the number of frames in $v_{i}$ and $m$ the number of words in $t_{j}$ (Section 3.1). The similarity between the $k^{th}$ frame and the $r^{th}$ word is defined as $s_{fw}(w_{jr}, z_{ik}') = (w_{jr})^\top z_{ik}'$ . Then we respectively define the local video-to-caption and caption-to-video similarities as: + +$$ +s ^ {\prime} \left(v _ {i}, t _ {j}\right) = \frac {1}{n} \sum_ {k = 1} ^ {n} \max _ {1 \leq r \leq m} \left\{s _ {f w} \left(w _ {j r}, z _ {i k} ^ {\prime}\right) \right\}, \tag {11} +$$ + +$$ +s ^ {\prime} (t _ {j}, v _ {i}) = \frac {1}{m} \sum_ {r = 1} ^ {m} \max _ {1 \leq k \leq n} \left\{s _ {f w} \left(z _ {i k} ^ {\prime}, w _ {j r}\right) \right\}. +$$ + +Finally, we have the local video-to-caption distribution $A_{LVC}$ and the local caption-to-video distribution $A_{LCV}$ : + +$$ +A _ {L V C} = \sigma \left( \begin{array}{c c c} s ^ {\prime} \left(v _ {1}, t _ {1}\right) & \dots & s ^ {\prime} \left(v _ {1}, t _ {B}\right) \\ \dots & \dots & \dots \\ s ^ {\prime} \left(v _ {B}, t _ {1}\right) & \dots & s ^ {\prime} \left(v _ {B}, t _ {B}\right) \end{array} \right), \tag {12} +$$ + +$$ +A _ {L C V} = \sigma \left( \begin{array}{c c c} s ^ {\prime} (t _ {1}, v _ {1}) & ... & s ^ {\prime} (t _ {1}, v _ {B}) \\ ... & ... & ... \\ s ^ {\prime} (t _ {B}, v _ {1}) & ... & s ^ {\prime} (t _ {B}, v _ {B}) \end{array} \right), +$$ + +and the local caption-video distribution alignment loss is defined as: + +$$ +L _ {L} = D _ {K L} \left(A _ {L C V} ^ {S}, A _ {L C V} ^ {T}\right) + D _ {K L} \left(A _ {L V C} ^ {S}, A _ {L V C} ^ {T}\right). \tag {13} +$$ + +Combining $L_{G}$ and $L_{V}$ , the cross-modality KD loss is: + +$$ +L _ {C M} = L _ {G} + L _ {L}. \tag {14} +$$ + +Our experiment in Section 4.2 shows that only when joint trained with SAB, our local video-caption distribution alignment shows its advantage, which also verifies the effectiveness of SAB. Finally, the total loss for training our model is: + +$$ +\begin{array}{l} L = L _ {\text {t a s k}} + \alpha \cdot L _ {T K D} + \beta \cdot L _ {S K D} \tag {15} \\ + \gamma \cdot L _ {C M} + \delta \cdot L _ {S A B}, \\ \end{array} +$$ + +where $L_{task}$ is the task-specifical loss, and $\alpha ,\beta ,\gamma$ and $\delta$ are loss balance weights. + +# 4. Experiments + +We conduct comprehensive experiments on three benchmarks for video-text retrieval (video-to-text $(v2t)$ and text-to-video $(t2v)$ ): MSR-VTT [38], MSVD [4] and LSMDC [32]. The metrics Recall at rank K $(R@K)$ are used for evaluation. + +# 4.1. Comparison with State-of-the-Arts + +MSR-VTT. In Table 1, we compare the proposed model CLIPPING with eight state-of-the-art methods on MSR-VTT: TACo [39], VideoClip [37], Frozen [3], VIOLET [12], OA-Trans [35], BridgeFormer [13], ALL-in-one [34] and MDMMT [11]. It can be seen that CLIPPING significantly surpasses those large-scale video-text/image-text pre-training models for video-text retrieval. For example, our model exceeds VideoClip and Frozen by absolute $9.8\%$ t2vR@1 and $8.2\%$ t2vR@1, respectively. Note that the model MDMMT uses CLIP as its backbone, while our CLIPPING uses CLIP as the teacher. For a more efficient and edge deployable video-language retrieval model, we further compress CLIPPING into two models: $\mathrm{CLIPPING}_{\mathrm{W / oT}}$ and $\mathrm{CLIPPING^{*}}$ . $\mathrm{CLIPPING}_{\mathrm{W / oT}}$ + +
ModelPT DatasetsParamsR@1
TACoHT100M212M28.4
VideoClipHT100M130M30.9
FrozenC3M,W2M,COCO232M32.5
ALL-in-one-SW2M,HT100M33M33.5
VIOLETC3M,W2M198M34.5
OA-TransC3M,W2M232M35.8
BridgeFormerC3M,W2M160M37.6
ALL-in-one-BW2M,HT100M110M37.9
MDMMTC400M,AudioSet226M38.9
CLIPPING*w/o T-8.7M37.5
CLIPPINGw/o T-21.3M38.6
CLIPPING*-46.5M39.8
CLIPPING-55.0M40.6
CLIPPINGIN21K55.0M40.7
+ +Table 1. Comparison with state-of-the-art models on MSR-VTT (1k split) for text-to-video retrieval. "PT Datasets": datasets used for pre-training the vision encoder. "HT100M": HowTo100M dataset [28]. "C400M": CLIP-400M dataset [30]. "IN21K": ImageNet21K dataset [9]. "W2M": WebVid-2M dataset [3]. "C3M": CC3M dataset [33]. All the methods in Table are the comparable methods with parameters less than 240M and flops less than 360G. + +does not use the temporal Transformer and CLIPPING* compresses the text encoder through TinyBERT [17]. It can be seen from Table 1 that the CLIPPING* based models also outperform the small model ALL-in-one-S even though they are smaller. In Fig. 5, we show the performances and Flops of these models. Among previous models, MDMMT has the best performance and ALL-in-one-S is the fastest. CLIPPING not only obtains relative $4.6\%$ performance gain over MDMMT but also is faster than All-in-one-S. Table 2 provides the details of Flops and parameters for each module of our method. + +![](images/f021df41a4d4a96d5066586f05fea5c73538a85e41bd2b20fad19379d7217281.jpg) +Figure 5. Flops and Performances. + +Other Benchmarks. We also compare our CLIPPING with the state-of-the-art Frozen [3], MDMMT [11], NoiseEst [2] and SupportSet [29] on the LSMDC and MSVD datasets. In Table 3, CLIPPING again obtains the best performance with fewest parameters and Flops, which achieves $88.1\% - 92.9\%$ of the performance of its teacher on the LSMDC and MSVD datasets, showing consistent improvements and generalization across different datasets. + +# 4.2. Ablation Study + +Key Components. We provide detailed ablation study to validate each key component of our proposed method, + +
ModulesParametersFlops
Vision encoder (MobileViT-v2)4.5M16.8G
Temporal Transformer37.8M1.2G
Language encoder (CLIPbert)12.6M0.15G
Language encoder (TinyBERT)4.2M0.05G
+ +Table 2. The Flops and parameters of each module in our method. We adopt 12 frames for each video. + +
ModelPT DatasetsVision EncodersR@1
LSMDC dataset
NoiseEstHT100MResnet152,ResNext-1016.4
SupportSetHT100MResNet-152,R(2+1)D-34IN21K
FrozenC3M,W2MSpace-Time15.0
MDMMTC400MCLIP_vision18.8
CLIPPING-MobileViT-v219.9
MSVD dataset
NoiseEstHT100MResnet152,ResNext-10120.3
SupportSetHT100MResNet-152,R(2+1)D-3428.4
FrozenC3M,W2MSpace-Time33.7
CLIPPING-MobileViT-v242.9
+ +Table 3. Comparison with the state-of-the-art on the LSMDC and MSVD datasets. + +
Vision EncoderKD Typest2vR@1v2tR@1
CLIP_vision-44.542.2
MobileViTv2-25.724.5
MobileViTv2T28.827.3
MobileViTv2T,S33.032.8
MobileViTv2T,S,SAB37.636.2
MobileViTv2T,S,SAB,CMG39.639.1
MobileViTv2T,S,SAB,CMG,CML40.740.2
+ +Table 4. Ablation study of different KD components of CLIPPING on the $1k$ validation set of MSR-VTT. The first row is the results of the teacher model (Clip4clip). $T$ , $S$ , $SAB$ , $CM_{G}$ and $CM_{L}$ denote temporal KD, spatial KD, SAB KD, global cross-modality KD and local cross-modality KD, respectively. + +
KD Typest2vR@1v2tR@1
T, S33.032.8
T, S, OL2OL34.633.4
T, S, TAB35.134.4
T, S, SABw/o masking37.135.7
T, S, SABw/masking37.636.2
+ +Table 5. Ablation study of different KD types on MSR-VTT (1k split). All the models are trained for 36 epochs with the same setting (see the supplementary materials for details). + +on MSR-VTT. From Table 4, we can see that without any KD, it gets extreme low accuracies. When simply adding the temporal KD and spatial KD, $t2vR@1$ and $v2tR@1$ increase significantly. When we further add SAB KD, and then global and local cross-modality KD, CLIPPING's performance rises gradually. This study shows that all these key components of CLIPPING are effective. Compared with the teacher, the full CLIPPING achieves about $91.5\%$ and $95.3\%$ of the performance of its teacher on $t2vR@1$ and $v2tR@1$ , respectively, with the vision encoder of $19.5x$ + +![](images/73f8c893c779df8ba44d8439f9e1534e0ecf89a956bd2a0389867668625eb2de.jpg) + +![](images/742b81165d4b07bc807f031179c52135084b0e830e3e585e0a32424e21459974.jpg) + +![](images/33f57a089fd3cfcb94f2bb72e7648f0fc9d7a95a37217a34da9afc01ccefa7c5.jpg) +(a) + +![](images/ae87ddfc9026dc088bee4452a793c3d753df1d9cc269a3f17683edf6bb56cfbe.jpg) +Figure 6. (a) Feature tensors of the student and the teacher. (b) An example to demonstrate that the teacher's features are the linear combinations of the student features. More examples are provided in the supplementary materials. + +![](images/9478f3c22d50742c4e8d4d5a73076b6d041a02a72c7f2e36d93cefb92076c799.jpg) +image + +![](images/1e533e4a9c58a15c471cb646cc25ff952aa2bd185a3aa88008b9ae8f586f779c.jpg) +(b) + +$\widetilde{W}$ $\tilde{L}_{er^S}$ + +![](images/dc1f322bc5138a618a9ab46be04c29d67cef79100d57a076656c02275a3985bf.jpg) + +![](images/d059336c5d5b7be3021a7959f093eb78e2e6fa1421e84c8a4f16aa8fb02d3c8c.jpg) +$\widetilde{W}\times \widetilde{L}\text{ayer}^{S} = \widetilde{L}\text{layer}_{w}^{S}\longrightarrow \widetilde{L}\text{ayer}^{T}$ + +$\widetilde{W}$ $\tilde{L}_{\text{ayer}}^s$ + +![](images/c1f0481e564156a055f51cd1b7d0ae7526aa9c76b3d13ddc340c79915fcf8395.jpg) + +![](images/a0a6ca7b3952a317d5be7a57280e6aa18cd90dbd1d7daa7b2ba3dc42307283af.jpg) + +![](images/e6a8dc3f1aa2c988b6130a5cc0fd982a7777218e021808ec0604c0e020615efa.jpg) +(a) + +![](images/0396bafb1c9c14e77e24411b8cb9818137cc28c575ff43b4583a7613a5968841.jpg) +(b) + +![](images/844436a69d6dc0fef38a47a216ad45f5b999d87491fd9f31476371e60bd75c3b.jpg) +(c) +Figure 7. Visualization of the frame-word alignments. (a) CLIP; (b) $\mathrm{CLIPPING}^{-}_{\mathrm{TAB}}$ ; (c) CLIPPING. + +smaller. + +KD Types. In Table 5, we compare the proposed SAB KD with the previous OL2OL KD and TAB KD. For SAB and TAB, the same layers of the student and the teacher are used for KD, the details of which are given in the supplementary materials. From Table 5, we can see that TAB performers better than OL2OL, and our SAB outperforms TAB. In Section 3.3.2, we use masking to speed up the training. In this study, we also compare "with masking" and "without masking". The last two rows of Table 5 verify that the masking is beneficial. + +SAB Property. The student's layers in our SAB KD can be explained as the bases of the feature space, and each layer of the teacher is a linear combination of the bases (Section 3.2). In Fig. 6, we show that this property holds. In SAB KD, the student has 4 layers and the teacher has 12 layers. We randomly pick 3 image examples, and for each example, we select the features of one random token (in the $k$ dimension in Fig. 6(a)). After training, the linear combinations $\tilde{W} \times \tilde{L}_{\text{layer}}^{S} = \tilde{L}_{\text{layer}}^{S}$ show feature patterns very similar to the teacher's features ( $\tilde{L}_{\text{layer}}^{T}$ ). This property verifies that the teacher's knowledge is fully absorbed by the student. + +In Fig. 7, we visualize the frame-word alignments with and without SAB KD. Fig. 7(a) is the result of the CLIP in CLIP4clip. Fig. 7(b) is the result of $\mathrm{CLIPPING}^{-}_{\mathrm{TAB}}$ , which is our model without SAB KD but with the previous TAB KD. It can be seen that the frame-word alignment with + +SAB KD (Fig. 7(c)) shows clear and correct word-level attentions (e.g., "naked", "child", "runs" and "through") like CLIP. But with TAB KD, the most important words (e.g., "naked" and "child") cannot be noticed; instead, some insignificant words ([CLS] and "a") are activated. This study shows that the knowledge of pre-trained CLIP can be more effectively transferred to the student at the fine-tuning stage with our SAB. + +# 5. Conclusion + +In this paper, we propose a novel and efficient knowledge distillation method that is specially designed for small vision-language models. It includes temporal KD, spatial KD, SAB KD and cross-modality KD. Especially, the SAB KD has the property of the student's layers being the bases of the feature space. After training, the teacher's features are the linear combinations of the bases, indicating that the student has fully absorbed the knowledge of the teacher. CLIPPING significantly outperforms the state-of-the-art and is comparable or even superior to many large pre-training models. In the feature, we will apply CLIPPING to other vision-language models for compression. + +# Acknowledgements + +We gratefully acknowledge the support of MindSpore [1], CANN (Compute Architecture for Neural Networks) and Ascend AI Processor used for this research. + +# References + +[1] https://www.mindspore.cn/.8 +[2] Elad Amrani, Rami Ben-Ari, Daniel Rotman, and Alex Bronstein. Noise estimation using density estimation for self-supervised multimodal learning. In AAAI, 2021. 7 +[3] Max Bain, Arsha Nagrani, Gül Varol, and Andrew Zisserman. Frozen in Time: A Joint Video and Image Encoder for End-to-End Retrieval. In ICCV, 2021. 2, 6, 7 +[4] David Chen and William Dolan. Collecting highly parallel data for paraphrase evaluation. In ACL, 2011. 6 +[5] Defang Chen, JianPing Mei, Yuan Zhang, Can Wang, Zhe Wang, Yan Feng, and Chun Chen. Distillation with semantic calibration. In AAAI, 2021. 2, 3, 5 +[6] Pengguang Chen, Shu Liu, Hengshuang Zhao, and Jiaya Jia. Distilling Knowledge via Knowledge Review. In CVPR, 2021. 2, 3 +[7] Xianing Chen, Qiong Cao, Yujie Zhong, Jing Zhang, Shenghua Gao, and Dacheng Tao. DearKD: Data-efficient early knowledge distillation for vision transformers. In ICLR, 2022. 2 +[8] Yinpeng Chen, Xiyang Dai, Dongdong Chen, Mengchen Liu, Xiaoyi Dong, Lu Yuan, and Zicheng Liu. Mobile-Former: Bridging MobileNet and Transformer. In CVPR, 2022. 1 +[9] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. ImageNet: A large-scale hierarchical image database. In CVPR, 2009. 7 +[10] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2021. 2 +[11] Maksim Dzabraev, Maksim Kalashnikov, Stepan Komkov, and Aleksandr Petiushko. MDMMT: Multidomain Multimodal Transformer for Video Retrieval. In CVPR, 2021. 2, 6, 7 +[12] Tsu-Jui Fu, Linjie Li, Zhe Gan, Kevin Lin, William Yang Wang, Lijuan Wang, and Zicheng Liu. VIOLET: End-to-End Video-Language Transformers with Masked Token Modeling. In arXiv:2111.1268, 2021. 1, 6 +[13] Yuying Ge, Yixiao Ge, Xihui Liu, Dian Li, Ying Shan, Xiaohu Qie, and Ping Luo. Bridgeformer: Bridging video-text retrieval with multiple choice questions. In CVPR, 2022. 6 +[14] Jianyuan Guo, Kai Han, Han Wu, Yehui Tang, Xinghao Chen, Yunhe Wang, and Chang Xu. CMT: Convolutional Neural Networks Meet Vision Transformers. In CVPR, 2022. 1 +[15] Geoffrey Hinton, Oriol Vinyals, and Jeff Dean. Distilling the Knowledge in a Neural Network. In arXiv preprint arXiv:1503.02531, 2015. 1, 2 +[16] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In ICML, 2021. 1 +[17] Xiaoqi Jiao, Yichun Yin, Lifeng Shang, Xin Jiang, Xiao Chen, Linlin Li, Fang Wang, and Qun Liu. TinyBERT: + +Distilling BERT for Natural Language Understanding. In EMNLP, 2020. 1, 2, 5, 7 +[18] Yoon Kim and Alexander M. Rush. Sequence-level knowledge distillation. In EMNLP, 2016. 1 +[19] Pavan Kumar, Anasosalu Vasu, James Gabriel, Jeff Zhu, Oncel Tuzel, and Anurag Ranjan. An Improved One millisecond Mobile Backbone. In arXiv preprint arXiv:2206.04040, 2022. 1, 2 +[20] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. BLIP: Bootstrapping Language-Image Pre-training for Unified Vision-Language Understanding and Generation. In arXiv preprint arXiv:2201.12086, 2022. 1 +[21] Junnan Li, Ramprasaath Selvaraju, Akhilesh Gotmare, Shafiq Joty, Caiming Xiong, and Steven Chu Hong Hoi. Align before fuse: Vision and language representation learning with momentum distillation. In NIPS, 2021. 1 +[22] Yanyu Li, Geng Yuan, Yang Wen, Eric Hu, Georgios Evangelidis, Sergey Tulyakov, Yanzhi Wang, and Jian Ren. EfficientFormer: Vision Transformers at MobileNet Speed. In arXiv:2206.01191, 2022. 1 +[23] Sihao Lin, Hongwei Xie, Bing Wang, Kaicheng Yu, Xiaojun Chang, Xiaodan Liang, and Gang Wang. Knowledge Distillation via the Target-aware Transformer. In CVPR, 2022. 2, 3 +[24] Yifan Liu, Chunhua Shen, Changqian Yu, and Jingdong Wang. Efficient Semantic Video Segmentation with Per-Frame Inference. In ECCV, 2020. 4 +[25] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin Transformer: Hierarchical Vision Transformer using Shifted Windows. In ICCV, 2021. 2 +[26] Huaishao Luo, Lei Ji, Ming Zhong, Yang Chen, Wen Lei, Nan Duan, and Tianrui Li. Clip4clip: An empirical study of clip for end to end video clip retrieval. arXiv preprint arXiv:2104.08860, 2021. 1, 2 +[27] Sachin Mehta and Mohammad Rastegari. Separable Self-attention for Mobile Vision Transformers. In ICLR, 2022. 1, 2, 4 +[28] Antoine Miech, Dimitri Zhukov, Jean-Baptiste Alayrac, Makarand Tapaswi, Ivan Laptev, and Josef Sivic. HowTo100M: Learning a Text-Video Embedding by Watching Hundred Million Narrated Video Clips. In ICCV, 2019. 7 +[29] Mandela Patrick, Po-Yao Huang, Yuki Asano, Florian Metze, Alexander Hauptmann, João Henriques, and Andrea Vedaldi. Support-set bottlenecks for video-text representation learning. In ICLR, 2021. 7 +[30] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, and Jack Clark. Learning transferable visual models from natural language supervision. ICML, 2021. 1, 2, 7 +[31] Maithra Raghu, Thomas Unterthiner, Simon Kornblith, Chiyuan Zhang, and Alexey Dosovitskiy. Do vision transformers see like convolutional neural networks? In NIPS, 2021. 2, 3, 5 +[32] Anna Rohrbach, Marcus Rohrbach, Niket Tandon, and Bernt Schiele. A dataset for Movie Description. In CVPR, 2015. 6 + +[33] Piyush Sharma, Nan Ding, Sebastian Goodman, and Radu Soricut. Conceptual captions: A cleaned, hypernymed, image alt-text dataset for automatic image captioning. In ACL, 2018. 7 +[34] Alex Jinpeng Wang, Yixiao Ge, Rui Yan, Yuying Ge, Xudong Lin, Guanyu Cai, Jianping Wu, Ying Shan, Xiaohu Qie, and Mike Zheng Shou. All in one: Exploring unified video-language pre-training. In arXiv preprint arXiv:2203.07303, 2022. 2, 6 +[35] Jinpeng Wang, Yixiao Ge, Guanyu Cai, Rui Yan, Xudong Lin, Ying Shan, Xiaohu Qie, and Mike Zheng Shou. Object-aware video-language pre-training for retrieval. In CVPR, 2022. 6 +[36] Zekun Wang, Wenhui Wang, Haichao Zhu, Ming Liu, Bing Qin, and Furu Wei. Distilled dual-encoder model for vision-language understanding. In arXiv preprint arXiv:2112.08723, 2021. 2 +[37] Hu Xu, Gargi Ghosh, Po-Yao Huang, Dmytro Okhonko, Armen Aghajanyan, Florian Metze, Luke Zettlemoyer, and Christoph Feichtenhofer. Videoclip: Contrastive pre-training for zero-shot video-text understanding. arXiv preprint arXiv:2104.08860, 2021. 2, 6 +[38] Jun Xu, Tao Mei, Ting Yao, and Yong Rui. MSR-VTT: A Large Video Description Dataset for Bridging Video and Language. 2016. 6 +[39] Jianwei Yang, Yonatan Bisk, and Jianfeng Gao. Taco: Token-aware cascade contrastive learning for video-text alignment. In ICCV, 2021. 6 +[40] Weihao Yu, Mi Luo, Pan Zhou, Chenyang Si, Yichen Zhou, Xinchao Wang, Jiashi Feng, and Shuicheng Yan. Metaformer is actually what you need for vision. In CVPR, 2022. 1 +[41] Li Yuan, Yunpeng Chen, Tao Wang, Weihao Yu, Yujun Shi, Zi-Hang Jiang, Francis E.H. Tay, Jiashi Feng, and Shuicheng Yan. Tokens-to-Token ViT: Training Vision Transformers From Scratch on ImageNet. In ICCV, 2021. 5 \ No newline at end of file diff --git a/2023/CLIPPING_ Distilling CLIP-Based Models With a Student Base for Video-Language Retrieval/images.zip b/2023/CLIPPING_ Distilling CLIP-Based Models With a Student Base for Video-Language Retrieval/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..7696a30d3195e21d90976c69e2fc128758d4165d --- /dev/null +++ b/2023/CLIPPING_ Distilling CLIP-Based Models With a Student Base for Video-Language Retrieval/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:346ffb5698fa7b87496904b1531ea7c20acdb2bf89416817a640688a8f7d7153 +size 620383 diff --git a/2023/CLIPPING_ Distilling CLIP-Based Models With a Student Base for Video-Language Retrieval/layout.json b/2023/CLIPPING_ Distilling CLIP-Based Models With a Student Base for Video-Language Retrieval/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..a0977a29435c19456ddbef4bbb01715aacb71f20 --- /dev/null +++ b/2023/CLIPPING_ Distilling CLIP-Based Models With a Student Base for Video-Language Retrieval/layout.json @@ -0,0 +1,10894 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 91, + 102, + 503, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 102, + 503, + 140 + ], + "spans": [ + { + "bbox": [ + 91, + 102, + 503, + 140 + ], + "type": "text", + "content": "CLIPPING: Distilling CLIP-Based Models with a Student Base for Video-Language Retrieval" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 155, + 161, + 438, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 155, + 161, + 438, + 203 + ], + "spans": [ + { + "bbox": [ + 155, + 161, + 438, + 203 + ], + "type": "text", + "content": "Renjing Pei, Jianzhuang Liu, Weimian Li, Bin Shao, Songcen Xu, Peng Dai, Juwei Lu, Youliang Yan Huawei Noah's Ark Lab" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 155, + 205, + 432, + 232 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 155, + 205, + 432, + 232 + ], + "spans": [ + { + "bbox": [ + 155, + 205, + 432, + 232 + ], + "type": "text", + "content": "{peirenjing,liu.jianzhuang,liweimian,shaobin3, xusongcen,peng.dai,juwei.lu,yanyouliang}@huawei.com" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 143, + 258, + 192, + 272 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 258, + 192, + 272 + ], + "spans": [ + { + "bbox": [ + 143, + 258, + 192, + 272 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 290, + 289, + 649 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 290, + 289, + 649 + ], + "spans": [ + { + "bbox": [ + 46, + 290, + 289, + 649 + ], + "type": "text", + "content": "Pre-training a vision-language model and then fine-tuning it on downstream tasks have become a popular paradigm. However, pre-trained vision-language models with the Transformer architecture usually take long inference time. Knowledge distillation has been an efficient technique to transfer the capability of a large model to a small one while maintaining the accuracy, which has achieved remarkable success in natural language processing. However, it faces many problems when applying KD to the multi-modality applications. In this paper, we propose a novel knowledge distillation method, named CLIPPING1, where the plentiful knowledge of a large teacher model that has been fine-tuned for video-language tasks with the powerful pre-trained CLIP can be effectively transferred to a small student only at the fine-tuning stage. Especially, a new layer-wise alignment with the student as the base is proposed for knowledge distillation of the intermediate layers in CLIPPING, which enables the student's layers to be the bases of the teacher, and thus allows the student to fully absorb the knowledge of the teacher. CLIPPING with MobileViT-v2 as the vision encoder without any vision-language pre-training achieves " + }, + { + "bbox": [ + 46, + 290, + 289, + 649 + ], + "type": "inline_equation", + "content": "88.1\\% - 95.3\\%" + }, + { + "bbox": [ + 46, + 290, + 289, + 649 + ], + "type": "text", + "content": " of the performance of its teacher on three video-language retrieval benchmarks, with its vision encoder being " + }, + { + "bbox": [ + 46, + 290, + 289, + 649 + ], + "type": "inline_equation", + "content": "19.5x" + }, + { + "bbox": [ + 46, + 290, + 289, + 649 + ], + "type": "text", + "content": " smaller. CLIPPING also significantly outperforms a state-of-the-art small baseline (ALL-in-one-B) on the MSR-VTT dataset, obtaining relatively " + }, + { + "bbox": [ + 46, + 290, + 289, + 649 + ], + "type": "inline_equation", + "content": "7.4\\%" + }, + { + "bbox": [ + 46, + 290, + 289, + 649 + ], + "type": "text", + "content": " performance gain, with " + }, + { + "bbox": [ + 46, + 290, + 289, + 649 + ], + "type": "inline_equation", + "content": "29\\%" + }, + { + "bbox": [ + 46, + 290, + 289, + 649 + ], + "type": "text", + "content": " fewer parameters and " + }, + { + "bbox": [ + 46, + 290, + 289, + 649 + ], + "type": "inline_equation", + "content": "86.9\\%" + }, + { + "bbox": [ + 46, + 290, + 289, + 649 + ], + "type": "text", + "content": " fewer flops. Moreover, CLIPPING is comparable or even superior to many large pre-training models." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 307, + 258, + 386, + 272 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 258, + 386, + 272 + ], + "spans": [ + { + "bbox": [ + 307, + 258, + 386, + 272 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 281, + 547, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 281, + 547, + 662 + ], + "spans": [ + { + "bbox": [ + 304, + 281, + 547, + 662 + ], + "type": "text", + "content": "Recently, pre-training a vision-language model and then fine-tuning it on downstream tasks are a popular paradigm [12, 16, 20, 21, 30]. Pre-trained vision-language models (PVLMs) have achieved great success in many multi-modality tasks (e.g., video-text retrieval). However, PVLMs with the Transformer architecture (especially the vision stream) usually consume a huge amount of computation, making them difficult to be deployed on edge devices such as mobile phones. Recent small models [8, 14, 19, 22, 27, 40] show that combining Convolutional Neural Networks (CNNs) and Transformers as a hybrid architecture gets the best of both architectures, but the overall performance of these works is still far from satisfactory when compared to large pre-training models. Apparently, knowledge distillation (KD) [15] is an efficient technique to transfer the capability of a large model to a small one while maintaining the accuracy, which has achieved remarkable success in natural language processing (NLP) [17, 18]. In NLP, the knowledge distillation methods are usually performed at both the pre-training and the fine-tuning stages. However, the collection of the pre-training data for the pre-training knowledge distillation cost huge manpower in multi-modality applications (e.g., the pre-training dataset of CLIP is hard to obtain). Therefore, an efficient knowledge distillation method need to be explored for multi-modality applications. To this end, we propose a novel knowledge distillation method, named CLIPPING, where the plentiful knowledge of a large teacher model Clip4clip [26] that has been fine-tuned for video-language tasks with the powerful pre-trained CLIP [30] can be effectively transferred to a small student only at the fine-tuning stage. The contributions of CLIPPING are summarized below:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 665, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 665, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 305, + 665, + 547, + 715 + ], + "type": "text", + "content": "1) We introduce an efficient approach to distill both the vision knowledge and the cross-modality knowledge from teacher to a small model at the fine-tuning stage. The resulting model shows strong performance on multi-modality" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 693, + 288, + 714 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 693, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 693, + 288, + 714 + ], + "type": "text", + "content": "1In this paper, CLIPPING means cutting something to make it smaller through distilling." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "18983" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 48, + 72, + 127, + 158 + ], + "blocks": [ + { + "bbox": [ + 48, + 72, + 127, + 158 + ], + "lines": [ + { + "bbox": [ + 48, + 72, + 127, + 158 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 127, + 158 + ], + "type": "image", + "image_path": "8097da16ed75c627f4843d5d3aeff5d1ba8aa878473aad9c97b44721c8dd0c8b.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 80, + 159, + 91, + 169 + ], + "lines": [ + { + "bbox": [ + 80, + 159, + 91, + 169 + ], + "spans": [ + { + "bbox": [ + 80, + 159, + 91, + 169 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 129, + 72, + 211, + 158 + ], + "blocks": [ + { + "bbox": [ + 129, + 72, + 211, + 158 + ], + "lines": [ + { + "bbox": [ + 129, + 72, + 211, + 158 + ], + "spans": [ + { + "bbox": [ + 129, + 72, + 211, + 158 + ], + "type": "image", + "image_path": "da544192d760903d94e60854f38c3d572730c43c338e865beb70fbb1bd558907.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 164, + 159, + 176, + 169 + ], + "lines": [ + { + "bbox": [ + 164, + 159, + 176, + 169 + ], + "spans": [ + { + "bbox": [ + 164, + 159, + 176, + 169 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 212, + 72, + 288, + 158 + ], + "blocks": [ + { + "bbox": [ + 212, + 72, + 288, + 158 + ], + "lines": [ + { + "bbox": [ + 212, + 72, + 288, + 158 + ], + "spans": [ + { + "bbox": [ + 212, + 72, + 288, + 158 + ], + "type": "image", + "image_path": "bfeb05a90cd2bf07c30252fef956f0e9ec8536a03e4fe8300eca4762e94cad2a.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 247, + 159, + 257, + 169 + ], + "lines": [ + { + "bbox": [ + 247, + 159, + 257, + 169 + ], + "spans": [ + { + "bbox": [ + 247, + 159, + 257, + 169 + ], + "type": "text", + "content": "(c)" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 177, + 287, + 342 + ], + "lines": [ + { + "bbox": [ + 46, + 177, + 287, + 342 + ], + "spans": [ + { + "bbox": [ + 46, + 177, + 287, + 342 + ], + "type": "text", + "content": "Figure 1. (a) and (b) are previous methods of intermediate features' knowledge distillation. " + }, + { + "bbox": [ + 46, + 177, + 287, + 342 + ], + "type": "inline_equation", + "content": "L_{j}^{T}" + }, + { + "bbox": [ + 46, + 177, + 287, + 342 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 177, + 287, + 342 + ], + "type": "inline_equation", + "content": "j = 1,2,\\dots,M" + }, + { + "bbox": [ + 46, + 177, + 287, + 342 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 177, + 287, + 342 + ], + "type": "inline_equation", + "content": "L_{i}^{S}" + }, + { + "bbox": [ + 46, + 177, + 287, + 342 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 177, + 287, + 342 + ], + "type": "inline_equation", + "content": "i = 1,2,\\dots,N" + }, + { + "bbox": [ + 46, + 177, + 287, + 342 + ], + "type": "text", + "content": ", are the outputs of the " + }, + { + "bbox": [ + 46, + 177, + 287, + 342 + ], + "type": "inline_equation", + "content": "j^{th}" + }, + { + "bbox": [ + 46, + 177, + 287, + 342 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 177, + 287, + 342 + ], + "type": "inline_equation", + "content": "i^{th}" + }, + { + "bbox": [ + 46, + 177, + 287, + 342 + ], + "type": "text", + "content": " layers of the teacher and the student, respectively. OL2OL is the most common One-Layer-to-One-Layer KD, where it selects some layers of the teacher and then distills the layers into the student one by one (e.g., from " + }, + { + "bbox": [ + 46, + 177, + 287, + 342 + ], + "type": "inline_equation", + "content": "L_{j}^{T}" + }, + { + "bbox": [ + 46, + 177, + 287, + 342 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 46, + 177, + 287, + 342 + ], + "type": "inline_equation", + "content": "L_{i}^{S}" + }, + { + "bbox": [ + 46, + 177, + 287, + 342 + ], + "type": "text", + "content": "). Teacher-As-Base (TAB) KD uses all the teacher's layers to supervise each layer in the student, where each layer of the student (e.g., " + }, + { + "bbox": [ + 46, + 177, + 287, + 342 + ], + "type": "inline_equation", + "content": "L_{i}^{S}" + }, + { + "bbox": [ + 46, + 177, + 287, + 342 + ], + "type": "text", + "content": ") learns the knowledge from all the selected layers of the teacher " + }, + { + "bbox": [ + 46, + 177, + 287, + 342 + ], + "type": "inline_equation", + "content": "\\left(\\sum_{k=1}^{M} w_{ik}' L_{k}^{T}\\right)" + }, + { + "bbox": [ + 46, + 177, + 287, + 342 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 177, + 287, + 342 + ], + "type": "inline_equation", + "content": "w_{ik}'" + }, + { + "bbox": [ + 46, + 177, + 287, + 342 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 177, + 287, + 342 + ], + "type": "inline_equation", + "content": "k = 1,2,\\dots,M" + }, + { + "bbox": [ + 46, + 177, + 287, + 342 + ], + "type": "text", + "content": ", are the knowledge selection weights with " + }, + { + "bbox": [ + 46, + 177, + 287, + 342 + ], + "type": "inline_equation", + "content": "\\sum_{k=1}^{M} w_{ik}' = 1" + }, + { + "bbox": [ + 46, + 177, + 287, + 342 + ], + "type": "text", + "content": ". (c) Our Student-As-Base (SAB) KD enables each of the teacher's layers to pass its knowledge to all the student's layers (e.g., " + }, + { + "bbox": [ + 46, + 177, + 287, + 342 + ], + "type": "inline_equation", + "content": "L_{j}^{T} \\rightarrow \\sum_{k=1}^{N} w_{jk} L_{k}^{S}" + }, + { + "bbox": [ + 46, + 177, + 287, + 342 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 177, + 287, + 342 + ], + "type": "inline_equation", + "content": "\\sum_{k=1}^{N} w_{jk} = 1" + }, + { + "bbox": [ + 46, + 177, + 287, + 342 + ], + "type": "text", + "content": "). More detailed analysis is provided in Section 3.1." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 354, + 146, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 354, + 146, + 365 + ], + "spans": [ + { + "bbox": [ + 47, + 354, + 146, + 365 + ], + "type": "text", + "content": "text-video retrieval task." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 366, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 46, + 366, + 287, + 484 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 366, + 287, + 484 + ], + "spans": [ + { + "bbox": [ + 46, + 366, + 287, + 484 + ], + "type": "text", + "content": "2) We propose a new layer-wise alignment scheme, called Student-As-Base (SAB), for knowledge distillation of the intermediate layers from the Transformer to the CNN in CLIPPING, where the student's layers can be regarded as the bases of the teacher's feature space, forcing the student model to full absorb the knowledge of the teacher. In our experience, the SAB layer-wise alignment significantly surpasses the previous knowledge distillation methods. We believe it will become a popular paradigm for knowledge distillation of intermediate features." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 486, + 287, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 486, + 287, + 605 + ], + "spans": [ + { + "bbox": [ + 46, + 486, + 287, + 605 + ], + "type": "text", + "content": "3) We present an effective cross-modal knowledge distillation, which includes knowledge from both the global and local video-caption distributions. We use the video-caption distributions of the teacher to guide the training of the student. Besides, the student can also benefit from the powerful pre-trained CLIP that obtains certain local frame-word attention ability via learning from massive data. Experiments show that this pre-training knowledge can be effectively transferred to the student when jointly trained with our SAB KD only at the fine-tuning stage." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "type": "text", + "content": "4) CLIPPING with MobileViT-v2 [27] as the vision encoder without any vision-language pre-training achieves " + }, + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "type": "inline_equation", + "content": "91.5\\% - 95.3\\%" + }, + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "type": "text", + "content": " of the performance of its teacher on MSR-VTT video-language retrieval benchmark, with its vision encoder being " + }, + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "type": "inline_equation", + "content": "19.5 \\times" + }, + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "type": "text", + "content": " smaller. CLIPPING also significantly outperforms a state-of-the-art small baseline (ALL-in-one-B) on the MSR-VTT dataset, obtaining relatively " + }, + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "type": "inline_equation", + "content": "7.4\\%" + }, + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "type": "text", + "content": " performance gain, with " + }, + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "type": "inline_equation", + "content": "29\\%" + }, + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "type": "text", + "content": " fewer parameters and " + }, + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "type": "inline_equation", + "content": "86.9\\%" + }, + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "type": "text", + "content": " fewer flops. Moreover, CLIPPING is comparable or even" + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 72, + 482, + 84 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 482, + 84 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 482, + 84 + ], + "type": "text", + "content": "superior to many large pre-training models." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 98, + 395, + 110 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 98, + 395, + 110 + ], + "spans": [ + { + "bbox": [ + 306, + 98, + 395, + 110 + ], + "type": "text", + "content": "2. Relative Work" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 113, + 545, + 281 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 113, + 545, + 281 + ], + "spans": [ + { + "bbox": [ + 304, + 113, + 545, + 281 + ], + "type": "text", + "content": "Vision-Language Retrieval. Learning from web-collected image-text data, large-scale Vision-Language Pretraining (VLP) models such as CLIP [30] have recently demonstrated great success across various downstream tasks. Nowadays, models such as Clip4clip [26] and MDMMT [11] extended from the pre-trained model CLIP keep appearing for vision-language retrieval. There are also some end-to-end trainable models [3, 37], which are designed to take advantage of both large-scale image and video captioning datasets. All-in-one [34] is the first work to consider both efficiency and performance for video-language retrieval tasks. It introduces a unified backbone that enables the representation learning of both video-text multimodal and unimodal inputs." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 281, + 546, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 281, + 546, + 556 + ], + "spans": [ + { + "bbox": [ + 304, + 281, + 546, + 556 + ], + "type": "text", + "content": "Knowledge Distillation. Knowledge distillation (KD) [15] is a simple yet effective technique to improve the performance of a learning model. Earlier works transfer knowledge embedded in the \"logits\" learned in a large teacher model to a small student model without sacrificing much performance. Recent works [5, 6] use multiple layers of the teacher to supervise each layer in the student, where each layer of the student learns the knowledge from multiple layers of the teacher. [23] proposes a target-aware Transformer and enables the student to mimic each spatial component of the teacher in each distilled layer to boost the student's performance. For multi-modality KD, [36] designs a fusion-encoder model as the teacher and introduces cross-modal attention knowledge to train the dual-encoder student model. The distillation objective is applied at both the pre-training and the fine-tuning stages and helps the dual-encoder model learn interactions of different modalities. TinyBert [17] also introduces a two-stage learning framework that performs Transformer distillation at both the pre-training and the task-specific fine-tuning stages. In this paper, we also focus on KD for transferring the knowledge of a pre-training vision-language model to a small one but only at the fine-tuning stage." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 557, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 557, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 557, + 546, + 713 + ], + "type": "text", + "content": "Transformers and CNNs. Usually, the pre-training model is a Transformer-based model [10, 25] and the small model is a CNN-based or hybrid architecture [19, 27]. For KD, most existing methods distill knowledge either from a Transformer to another Transformer (T2T) or from a CNN to another CNN (C2C) that computes the loss in the OL2OL or TAB style (see Fig. 1). To the best of our knowledge, KD from a Transformer to a CNN (T2C) has not been explored yet. The most related work [7] distills knowledge from a CNN to a Transformer (C2T) in the OL2OL way. However, previous works (e.g., [31]) have investigated the internal representation structures of ViTs and CNNs, and found striking differences between the two models, such as" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18984" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 147, + 68, + 440, + 290 + ], + "blocks": [ + { + "bbox": [ + 147, + 68, + 440, + 290 + ], + "lines": [ + { + "bbox": [ + 147, + 68, + 440, + 290 + ], + "spans": [ + { + "bbox": [ + 147, + 68, + 440, + 290 + ], + "type": "image", + "image_path": "9d827e96f3b031b4aa6ccfb205aa5f203433525ff93a0a1ad43f822c41195bae.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 303, + 547, + 336 + ], + "lines": [ + { + "bbox": [ + 46, + 303, + 547, + 336 + ], + "spans": [ + { + "bbox": [ + 46, + 303, + 547, + 336 + ], + "type": "text", + "content": "Figure 2. Overview of CLIPPING. There are mainly four knowledge distillation (KD) parts: (1) Temporal KD. (2) Spatial KD. (3) SAB (Student-As-Base) KD. (4) Cross-modality KD. The CLIP's vision encoder is a Transformer. The temporal Transformer is an optional module for CLIPPING." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 347, + 289, + 431 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 347, + 289, + 431 + ], + "spans": [ + { + "bbox": [ + 46, + 347, + 289, + 431 + ], + "type": "text", + "content": "ViTs having highly similar representations throughout the model's layers, while CNNs showing obvious distinction of representations between lower and higher layers [31]. Considering such striking differences between ViTs and CNNs, the previous KD (OL2OL and TAB) may not be good distillation ways for our T2C task, which is verified by our experiments." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 441, + 130, + 455 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 441, + 130, + 455 + ], + "spans": [ + { + "bbox": [ + 47, + 441, + 130, + 455 + ], + "type": "text", + "content": "3. Methodology" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 462, + 287, + 558 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 462, + 287, + 558 + ], + "spans": [ + { + "bbox": [ + 46, + 462, + 287, + 558 + ], + "type": "text", + "content": "We propose an efficient approach to distill both the vision knowledge and the cross-modality knowledge from Clip4clip to a small model with MobileViT-v2 as its vision encoder for text-video retrieval task, which can get the best from both sides: powerful multimodal representations of CLIP and efficient mobile vision Transformer of MobileViT-v2. The overall architecture is illustrated in Fig. 2." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 564, + 190, + 577 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 564, + 190, + 577 + ], + "spans": [ + { + "bbox": [ + 47, + 564, + 190, + 577 + ], + "type": "text", + "content": "3.1. Analysis of TAB and SAB" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 582, + 287, + 619 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 582, + 287, + 619 + ], + "spans": [ + { + "bbox": [ + 46, + 582, + 287, + 619 + ], + "type": "text", + "content": "We first show the advantage of our SAB KD over traditional TAB KD. For simplicity, suppose " + }, + { + "bbox": [ + 46, + 582, + 287, + 619 + ], + "type": "inline_equation", + "content": "M = N = 2" + }, + { + "bbox": [ + 46, + 582, + 287, + 619 + ], + "type": "text", + "content": ". Then, for SAB," + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 72, + 626, + 287, + 658 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 72, + 626, + 287, + 658 + ], + "spans": [ + { + "bbox": [ + 72, + 626, + 287, + 658 + ], + "type": "interline_equation", + "content": "\\left\\{ \\begin{array}{l} L _ {1} ^ {T} = w _ {1 1} L _ {1} ^ {S} + w _ {1 2} L _ {2} ^ {S}, w _ {1 1} + w _ {1 2} = 1 \\\\ L _ {2} ^ {T} = w _ {2 1} L _ {1} ^ {S} + w _ {2 2} L _ {2} ^ {S}, w _ {2 1} + w _ {2 2} = 1 \\end{array} , \\right. \\tag {1}", + "image_path": "dbbcbe7d6194e022211c722e9ea2e0bebcec0e8351748fa85611d77c191173b1.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 665, + 102, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 665, + 102, + 676 + ], + "spans": [ + { + "bbox": [ + 47, + 665, + 102, + 676 + ], + "type": "text", + "content": "and for TAB," + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 72, + 684, + 287, + 716 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 72, + 684, + 287, + 716 + ], + "spans": [ + { + "bbox": [ + 72, + 684, + 287, + 716 + ], + "type": "interline_equation", + "content": "\\left\\{ \\begin{array}{l} L _ {1} ^ {S} = w _ {1 1} ^ {\\prime} L _ {1} ^ {T} + w _ {1 2} ^ {\\prime} L _ {2} ^ {T}, w _ {1 1} ^ {\\prime} + w _ {1 2} ^ {\\prime} = 1 \\\\ L _ {2} ^ {S} = w _ {2 1} ^ {\\prime} L _ {1} ^ {T} + w _ {2 2} ^ {\\prime} L _ {2} ^ {T}, w _ {2 1} ^ {\\prime} + w _ {2 2} ^ {\\prime} = 1 \\end{array} \\right. \\tag {2}", + "image_path": "27f9bae02657523377efd8034b40fb3867830dc1395e1e860204919d264b3a4f.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 347, + 545, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 347, + 545, + 407 + ], + "spans": [ + { + "bbox": [ + 306, + 347, + 545, + 407 + ], + "type": "text", + "content": "Obviously, in Eq. 1, the knowledge of the teacher " + }, + { + "bbox": [ + 306, + 347, + 545, + 407 + ], + "type": "inline_equation", + "content": "L_1^T" + }, + { + "bbox": [ + 306, + 347, + 545, + 407 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 306, + 347, + 545, + 407 + ], + "type": "inline_equation", + "content": "L_2^T" + }, + { + "bbox": [ + 306, + 347, + 545, + 407 + ], + "type": "text", + "content": " is the linear combination of " + }, + { + "bbox": [ + 306, + 347, + 545, + 407 + ], + "type": "inline_equation", + "content": "L_1^S" + }, + { + "bbox": [ + 306, + 347, + 545, + 407 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 306, + 347, + 545, + 407 + ], + "type": "inline_equation", + "content": "L_2^S" + }, + { + "bbox": [ + 306, + 347, + 545, + 407 + ], + "type": "text", + "content": ". In other words, " + }, + { + "bbox": [ + 306, + 347, + 545, + 407 + ], + "type": "inline_equation", + "content": "L_1^S" + }, + { + "bbox": [ + 306, + 347, + 545, + 407 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 306, + 347, + 545, + 407 + ], + "type": "inline_equation", + "content": "L_2^S" + }, + { + "bbox": [ + 306, + 347, + 545, + 407 + ], + "type": "text", + "content": " are the bases of " + }, + { + "bbox": [ + 306, + 347, + 545, + 407 + ], + "type": "inline_equation", + "content": "L_1^T" + }, + { + "bbox": [ + 306, + 347, + 545, + 407 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 306, + 347, + 545, + 407 + ], + "type": "inline_equation", + "content": "L_2^T" + }, + { + "bbox": [ + 306, + 347, + 545, + 407 + ], + "type": "text", + "content": " of the teacher. In Eq. 2, it can be seen that the teacher's " + }, + { + "bbox": [ + 306, + 347, + 545, + 407 + ], + "type": "inline_equation", + "content": "L_1^T" + }, + { + "bbox": [ + 306, + 347, + 545, + 407 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 306, + 347, + 545, + 407 + ], + "type": "inline_equation", + "content": "L_2^T" + }, + { + "bbox": [ + 306, + 347, + 545, + 407 + ], + "type": "text", + "content": " are as the bases for " + }, + { + "bbox": [ + 306, + 347, + 545, + 407 + ], + "type": "inline_equation", + "content": "L_1^S" + }, + { + "bbox": [ + 306, + 347, + 545, + 407 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 306, + 347, + 545, + 407 + ], + "type": "inline_equation", + "content": "L_2^S" + }, + { + "bbox": [ + 306, + 347, + 545, + 407 + ], + "type": "text", + "content": ". Eq. 2 can be converted to:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 318, + 414, + 545, + 430 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 414, + 545, + 430 + ], + "spans": [ + { + "bbox": [ + 318, + 414, + 545, + 430 + ], + "type": "interline_equation", + "content": "L _ {1} ^ {T} = A \\cdot L _ {1} ^ {S} + B \\cdot L _ {2} ^ {S}, L _ {2} ^ {T} = C \\cdot L _ {1} ^ {S} + D \\cdot L _ {2} ^ {S}, \\tag {3}", + "image_path": "f9956efd78d5b0a85db75da93fac38b8b82eb45ea92d98fc8007704bddd84bba.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 437, + 545, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 437, + 545, + 456 + ], + "spans": [ + { + "bbox": [ + 305, + 437, + 545, + 456 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 437, + 545, + 456 + ], + "type": "inline_equation", + "content": "A = \\frac{w_{22}^{\\prime}}{w_{11}^{\\prime}w_{22}^{\\prime} - w_{12}^{\\prime}w_{21}^{\\prime}}" + }, + { + "bbox": [ + 305, + 437, + 545, + 456 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 305, + 437, + 545, + 456 + ], + "type": "inline_equation", + "content": "B = \\frac{-w_{12}^{\\prime}}{w_{11}^{\\prime}w_{22}^{\\prime} - w_{12}^{\\prime}w_{21}^{\\prime}}" + }, + { + "bbox": [ + 305, + 437, + 545, + 456 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 305, + 437, + 545, + 456 + ], + "type": "inline_equation", + "content": "C =" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 455, + 545, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 455, + 545, + 555 + ], + "spans": [ + { + "bbox": [ + 305, + 455, + 545, + 555 + ], + "type": "inline_equation", + "content": "\\frac{-w_{21}^{\\prime}}{w_{11}^{\\prime}w_{22}^{\\prime} - w_{12}^{\\prime}w_{21}^{\\prime}}" + }, + { + "bbox": [ + 305, + 455, + 545, + 555 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 455, + 545, + 555 + ], + "type": "inline_equation", + "content": "D = \\frac{w_{11}^{\\prime}}{w_{11}^{\\prime}w_{22}^{\\prime} - w_{12}^{\\prime}w_{21}^{\\prime}}" + }, + { + "bbox": [ + 305, + 455, + 545, + 555 + ], + "type": "text", + "content": ", which looks like Eq. 1. It seems that in this case, the student's " + }, + { + "bbox": [ + 305, + 455, + 545, + 555 + ], + "type": "inline_equation", + "content": "L_1^S" + }, + { + "bbox": [ + 305, + 455, + 545, + 555 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 455, + 545, + 555 + ], + "type": "inline_equation", + "content": "L_2^S" + }, + { + "bbox": [ + 305, + 455, + 545, + 555 + ], + "type": "text", + "content": " could also serve as the bases for " + }, + { + "bbox": [ + 305, + 455, + 545, + 555 + ], + "type": "inline_equation", + "content": "L_1^T" + }, + { + "bbox": [ + 305, + 455, + 545, + 555 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 455, + 545, + 555 + ], + "type": "inline_equation", + "content": "L_2^T" + }, + { + "bbox": [ + 305, + 455, + 545, + 555 + ], + "type": "text", + "content": ". However, (i) " + }, + { + "bbox": [ + 305, + 455, + 545, + 555 + ], + "type": "inline_equation", + "content": "A + B \\neq 1" + }, + { + "bbox": [ + 305, + 455, + 545, + 555 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 455, + 545, + 555 + ], + "type": "inline_equation", + "content": "C + D \\neq 1" + }, + { + "bbox": [ + 305, + 455, + 545, + 555 + ], + "type": "text", + "content": ", but it is usually required in the previous KD that all the coefficients in each equation should be summed to be 1 [5,6,23]; (ii) when learning with TAB KD in Eq. 2, the coefficients A, B, C and D are uncontrollable, which may be arbitrary large during learning." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 556, + 545, + 663 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 556, + 545, + 663 + ], + "spans": [ + { + "bbox": [ + 304, + 556, + 545, + 663 + ], + "type": "text", + "content": "For the reasons above, although it seems that SAB and TAB have the same matrix representation, TAB is unable to get a similar result as SAB and more likely to get to a local minimum during training. Therefore, TAB KD is not equivalent to our SAB KD. When well trained, the student with this SAB KD fully learns the knowledge of the teacher because the teacher's " + }, + { + "bbox": [ + 304, + 556, + 545, + 663 + ], + "type": "inline_equation", + "content": "L_{i}^{T}" + }, + { + "bbox": [ + 304, + 556, + 545, + 663 + ], + "type": "text", + "content": " can be obtained by the linear combination of the student's " + }, + { + "bbox": [ + 304, + 556, + 545, + 663 + ], + "type": "inline_equation", + "content": "L_{i}^{S}" + }, + { + "bbox": [ + 304, + 556, + 545, + 663 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 556, + 545, + 663 + ], + "type": "inline_equation", + "content": "i = 1,2,\\dots,N" + }, + { + "bbox": [ + 304, + 556, + 545, + 663 + ], + "type": "text", + "content": ". Our experiment also shows this property (see Section 4.2)." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 671, + 391, + 682 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 671, + 391, + 682 + ], + "spans": [ + { + "bbox": [ + 306, + 671, + 391, + 682 + ], + "type": "text", + "content": "3.2. Preliminaries" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 689, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 545, + 714 + ], + "type": "text", + "content": "Given a batch of videos " + }, + { + "bbox": [ + 306, + 689, + 545, + 714 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 306, + 689, + 545, + 714 + ], + "type": "text", + "content": " and captions " + }, + { + "bbox": [ + 306, + 689, + 545, + 714 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 306, + 689, + 545, + 714 + ], + "type": "text", + "content": ", Clip4clip learns a similarity function " + }, + { + "bbox": [ + 306, + 689, + 545, + 714 + ], + "type": "inline_equation", + "content": "s(v_{i},t_{j})" + }, + { + "bbox": [ + 306, + 689, + 545, + 714 + ], + "type": "text", + "content": " to calculate the similar-" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18985" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 109, + 68, + 487, + 160 + ], + "blocks": [ + { + "bbox": [ + 109, + 68, + 487, + 160 + ], + "lines": [ + { + "bbox": [ + 109, + 68, + 487, + 160 + ], + "spans": [ + { + "bbox": [ + 109, + 68, + 487, + 160 + ], + "type": "image", + "image_path": "d0412b3b6617f2558684dd17b058443e37462ad29a8cf2125df57a90e1172e52.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 109, + 161, + 485, + 259 + ], + "blocks": [ + { + "bbox": [ + 109, + 161, + 485, + 259 + ], + "lines": [ + { + "bbox": [ + 109, + 161, + 485, + 259 + ], + "spans": [ + { + "bbox": [ + 109, + 161, + 485, + 259 + ], + "type": "image", + "image_path": "b3485e9995b2dc2b7f0637287b4d2b50eb2976604aabbc3a916d59d2ef28b4f3.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 70, + 266, + 522, + 279 + ], + "lines": [ + { + "bbox": [ + 70, + 266, + 522, + 279 + ], + "spans": [ + { + "bbox": [ + 70, + 266, + 522, + 279 + ], + "type": "text", + "content": "Figure 3. Reshaping the features of (a) the student " + }, + { + "bbox": [ + 70, + 266, + 522, + 279 + ], + "type": "inline_equation", + "content": "L_{k}^{S}" + }, + { + "bbox": [ + 70, + 266, + 522, + 279 + ], + "type": "text", + "content": " and (b) the teacher " + }, + { + "bbox": [ + 70, + 266, + 522, + 279 + ], + "type": "inline_equation", + "content": "L_{k}^{T}" + }, + { + "bbox": [ + 70, + 266, + 522, + 279 + ], + "type": "text", + "content": " to the same shape. (c) Similarity Computation." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 296, + 289, + 489 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 296, + 289, + 489 + ], + "spans": [ + { + "bbox": [ + 46, + 296, + 289, + 489 + ], + "type": "text", + "content": "ity between a video " + }, + { + "bbox": [ + 46, + 296, + 289, + 489 + ], + "type": "inline_equation", + "content": "v_{i} \\in V" + }, + { + "bbox": [ + 46, + 296, + 289, + 489 + ], + "type": "text", + "content": " and a caption " + }, + { + "bbox": [ + 46, + 296, + 289, + 489 + ], + "type": "inline_equation", + "content": "t_j \\in T" + }, + { + "bbox": [ + 46, + 296, + 289, + 489 + ], + "type": "text", + "content": ". It obtains the frames' representation and the caption representation in a multi-modal embedding space via CLIP. The frames' representation is denoted as " + }, + { + "bbox": [ + 46, + 296, + 289, + 489 + ], + "type": "inline_equation", + "content": "Z_{i} = \\{z_{i1}, z_{i2}, \\dots, z_{in}\\}" + }, + { + "bbox": [ + 46, + 296, + 289, + 489 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 296, + 289, + 489 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 46, + 296, + 289, + 489 + ], + "type": "text", + "content": " is the number of frames in " + }, + { + "bbox": [ + 46, + 296, + 289, + 489 + ], + "type": "inline_equation", + "content": "v_{i}" + }, + { + "bbox": [ + 46, + 296, + 289, + 489 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 296, + 289, + 489 + ], + "type": "inline_equation", + "content": "z_{ik}" + }, + { + "bbox": [ + 46, + 296, + 289, + 489 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 46, + 296, + 289, + 489 + ], + "type": "inline_equation", + "content": "k = 1, 2, \\dots, n" + }, + { + "bbox": [ + 46, + 296, + 289, + 489 + ], + "type": "text", + "content": ") is the class token from the output of the CLIP's vision encoder corresponding to the " + }, + { + "bbox": [ + 46, + 296, + 289, + 489 + ], + "type": "inline_equation", + "content": "k^{th}" + }, + { + "bbox": [ + 46, + 296, + 289, + 489 + ], + "type": "text", + "content": " input frame. The caption representation is denoted as " + }, + { + "bbox": [ + 46, + 296, + 289, + 489 + ], + "type": "inline_equation", + "content": "W_{j} = \\{w_{j1}, w_{j2}, \\dots, w_{jm}\\}" + }, + { + "bbox": [ + 46, + 296, + 289, + 489 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 296, + 289, + 489 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 46, + 296, + 289, + 489 + ], + "type": "text", + "content": " is the number of words in " + }, + { + "bbox": [ + 46, + 296, + 289, + 489 + ], + "type": "inline_equation", + "content": "t_j" + }, + { + "bbox": [ + 46, + 296, + 289, + 489 + ], + "type": "text", + "content": " (including [CLS] and [SEP]) and " + }, + { + "bbox": [ + 46, + 296, + 289, + 489 + ], + "type": "inline_equation", + "content": "w_{jm}" + }, + { + "bbox": [ + 46, + 296, + 289, + 489 + ], + "type": "text", + "content": " is used as the representation of " + }, + { + "bbox": [ + 46, + 296, + 289, + 489 + ], + "type": "inline_equation", + "content": "t_j" + }, + { + "bbox": [ + 46, + 296, + 289, + 489 + ], + "type": "text", + "content": " in Clip4clip. Clip4clip uses a temporal Transformer encoder to obtain the sequential feature, denoted as " + }, + { + "bbox": [ + 46, + 296, + 289, + 489 + ], + "type": "inline_equation", + "content": "Z_{i}' = \\{z_{i1}', z_{i2}', \\dots, z_{in}'\\} = TemporalTransformer(Z_{i})" + }, + { + "bbox": [ + 46, + 296, + 289, + 489 + ], + "type": "text", + "content": ". Then, the mean pooling is used to aggregate the features of all frames to obtain the video representation, " + }, + { + "bbox": [ + 46, + 296, + 289, + 489 + ], + "type": "inline_equation", + "content": "z_{i} = \\frac{1}{n} \\sum_{k=1}^{n} z_{ik}'" + }, + { + "bbox": [ + 46, + 296, + 289, + 489 + ], + "type": "text", + "content": ". Finally, the similarity functions " + }, + { + "bbox": [ + 46, + 296, + 289, + 489 + ], + "type": "inline_equation", + "content": "s(v_i, t_j)" + }, + { + "bbox": [ + 46, + 296, + 289, + 489 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 296, + 289, + 489 + ], + "type": "inline_equation", + "content": "s(t_j, v_i)" + }, + { + "bbox": [ + 46, + 296, + 289, + 489 + ], + "type": "text", + "content": " are defined as:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 85, + 491, + 287, + 507 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 85, + 491, + 287, + 507 + ], + "spans": [ + { + "bbox": [ + 85, + 491, + 287, + 507 + ], + "type": "interline_equation", + "content": "s (v _ {i}, t _ {j}) = w _ {j m} ^ {\\top} z _ {i}, \\quad s (t _ {i}, v _ {j}) = z _ {j} ^ {\\top} w _ {i m}. \\tag {4}", + "image_path": "492859c6216ba22c080f80aeae2b57c35eccab274985d3cf003b590dd9e11b34.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 510, + 289, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 510, + 289, + 666 + ], + "spans": [ + { + "bbox": [ + 46, + 510, + 289, + 666 + ], + "type": "text", + "content": "Nowadays, models such as Clip4clip based on the pretrained model CLIP have been applied to many tasks. They pursue better performance with CLIP but have a heavy vision encoder, which makes them difficult to be deployed on edge devices such as mobile phones. Since MobileViT-v2 [27] is a light-weight and mobile-friendly hybrid network, we employ it as the vision encoder of the student model and maintain the original text encoder and temporal Transformer of Clip4clip. Note that the temporal Transformer is an optional module for the CLIPPING model structure, and the text encoder " + }, + { + "bbox": [ + 46, + 510, + 289, + 666 + ], + "type": "inline_equation", + "content": "(CLIP_{bert})" + }, + { + "bbox": [ + 46, + 510, + 289, + 666 + ], + "type": "text", + "content": " can be further compressed through the existing methods (details are shown in the Section 4)." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 671, + 124, + 682 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 671, + 124, + 682 + ], + "spans": [ + { + "bbox": [ + 47, + 671, + 124, + 682 + ], + "type": "text", + "content": "3.3. CLIPPING" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 689, + 287, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 287, + 712 + ], + "type": "text", + "content": "To distill multimodal knowledge from Clip4clip to the MobileViT-v2-based model, we use four kinds of knowl-" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 296, + 545, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 296, + 545, + 319 + ], + "spans": [ + { + "bbox": [ + 305, + 296, + 545, + 319 + ], + "type": "text", + "content": "edge transfer: 1) temporal KD, 2) spatial KD, 3) SAB KD, and 4) cross-modality KD." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 334, + 531, + 346 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 334, + 531, + 346 + ], + "spans": [ + { + "bbox": [ + 305, + 334, + 531, + 346 + ], + "type": "text", + "content": "3.3.1 Temporal and Spatial Knowledge Distillation" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 352, + 545, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 352, + 545, + 449 + ], + "spans": [ + { + "bbox": [ + 304, + 352, + 545, + 449 + ], + "type": "text", + "content": "The temporal KD is motivated by the previous finding that aligning multi-frame dependency from the teacher to the student can enhance the performance of the student [24], which encodes the multi-frame dependency into a latent embedding by using a recurrent unit ConvLSTM. In our architecture, multi-frame dependency can be modeled naturally through the temporal Transformer (Temporal Transformer) in Fig. 2, so we define the temporal KD loss as:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 355, + 455, + 545, + 487 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 355, + 455, + 545, + 487 + ], + "spans": [ + { + "bbox": [ + 355, + 455, + 545, + 487 + ], + "type": "interline_equation", + "content": "L _ {T K D} = \\frac {1}{B} \\sum_ {i = 1} ^ {B} D _ {K L} \\left(Z _ {i} ^ {\\prime S}, Z _ {i} ^ {\\prime T}\\right), \\tag {5}", + "image_path": "0b97f4990d03df5465375e9072534f8bf8cd043e1cd962238769a2f2a568747e.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 494, + 545, + 554 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 494, + 545, + 554 + ], + "spans": [ + { + "bbox": [ + 304, + 494, + 545, + 554 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 494, + 545, + 554 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 304, + 494, + 545, + 554 + ], + "type": "text", + "content": " is the batch size, the superscripts " + }, + { + "bbox": [ + 304, + 494, + 545, + 554 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 304, + 494, + 545, + 554 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 494, + 545, + 554 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 304, + 494, + 545, + 554 + ], + "type": "text", + "content": " denote the student and the teacher, respectively, and " + }, + { + "bbox": [ + 304, + 494, + 545, + 554 + ], + "type": "inline_equation", + "content": "D_{KL}()" + }, + { + "bbox": [ + 304, + 494, + 545, + 554 + ], + "type": "text", + "content": " is the KL divergence loss function. In addition to imitating the behavior of CLIP for each frame, we use spatial KD to align the spatial embeddings of the student and the teacher by:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 358, + 560, + 545, + 591 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 358, + 560, + 545, + 591 + ], + "spans": [ + { + "bbox": [ + 358, + 560, + 545, + 591 + ], + "type": "interline_equation", + "content": "L _ {S K D} = \\frac {1}{B} \\sum_ {i = 1} ^ {B} D _ {K L} \\left(Z _ {i} ^ {S}, Z _ {i} ^ {T}\\right). \\tag {6}", + "image_path": "676cbc84bb80d8c4a0c331d18e12ba5fe3defe8367b5fc69775f24132f192f37.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 591, + 545, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 591, + 545, + 628 + ], + "spans": [ + { + "bbox": [ + 304, + 591, + 545, + 628 + ], + "type": "text", + "content": "The Temporal Transformer is an optional module for the CLIPPING model structure in our method. In this case, " + }, + { + "bbox": [ + 304, + 591, + 545, + 628 + ], + "type": "inline_equation", + "content": "Z_{i}^{\\prime S}" + }, + { + "bbox": [ + 304, + 591, + 545, + 628 + ], + "type": "text", + "content": " is exactly equal to " + }, + { + "bbox": [ + 304, + 591, + 545, + 628 + ], + "type": "inline_equation", + "content": "Z_{i}^{S}" + }, + { + "bbox": [ + 304, + 591, + 545, + 628 + ], + "type": "text", + "content": " in TKD." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 639, + 458, + 651 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 639, + 458, + 651 + ], + "spans": [ + { + "bbox": [ + 306, + 639, + 458, + 651 + ], + "type": "text", + "content": "3.3.2 SAB Knowledge Distillation" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 653, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 653, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 653, + 545, + 712 + ], + "type": "text", + "content": "Recently, it is discovered that distilling intermediate features is more effective. So in addition to the spatial and temporal KD, we also force the vision encoder of the student to mimic the intermediate layers of the teacher. In the traditional OL2OL KD, some layers of the teacher are selected" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18986" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 147, + 68, + 254, + 166 + ], + "blocks": [ + { + "bbox": [ + 147, + 68, + 254, + 166 + ], + "lines": [ + { + "bbox": [ + 147, + 68, + 254, + 166 + ], + "spans": [ + { + "bbox": [ + 147, + 68, + 254, + 166 + ], + "type": "image", + "image_path": "5aaf3acfbac4445d947fd6ee4f54a65e2b4658a636631e79e200eeb3dbdc4c65.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 256, + 68, + 447, + 167 + ], + "blocks": [ + { + "bbox": [ + 256, + 68, + 447, + 167 + ], + "lines": [ + { + "bbox": [ + 256, + 68, + 447, + 167 + ], + "spans": [ + { + "bbox": [ + 256, + 68, + 447, + 167 + ], + "type": "image", + "image_path": "8e174f102370c416bcd2396d5ae76916e10d293ca87bd242400739efdd6d36ea.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 147, + 168, + 447, + 264 + ], + "blocks": [ + { + "bbox": [ + 147, + 168, + 447, + 264 + ], + "lines": [ + { + "bbox": [ + 147, + 168, + 447, + 264 + ], + "spans": [ + { + "bbox": [ + 147, + 168, + 447, + 264 + ], + "type": "image", + "image_path": "b671099dc4d7918037e0e99b464aab4635924d5e4a39a59575ec8ba9b240699c.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 272, + 547, + 297 + ], + "lines": [ + { + "bbox": [ + 46, + 272, + 547, + 297 + ], + "spans": [ + { + "bbox": [ + 46, + 272, + 547, + 297 + ], + "type": "text", + "content": "Figure 4. (a) Masks obtained with Eq. 7, where all the " + }, + { + "bbox": [ + 46, + 272, + 547, + 297 + ], + "type": "inline_equation", + "content": "L - 1" + }, + { + "bbox": [ + 46, + 272, + 547, + 297 + ], + "type": "text", + "content": " masks along the " + }, + { + "bbox": [ + 46, + 272, + 547, + 297 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 272, + 547, + 297 + ], + "type": "text", + "content": " dimension are the same. (b) Masking the similarity tensor " + }, + { + "bbox": [ + 46, + 272, + 547, + 297 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 46, + 272, + 547, + 297 + ], + "type": "text", + "content": ". (c) SAB alignment. " + }, + { + "bbox": [ + 46, + 272, + 547, + 297 + ], + "type": "inline_equation", + "content": "L_{SAB}" + }, + { + "bbox": [ + 46, + 272, + 547, + 297 + ], + "type": "text", + "content": " is the SAB loss." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 311, + 289, + 539 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 311, + 289, + 539 + ], + "spans": [ + { + "bbox": [ + 46, + 311, + 289, + 539 + ], + "type": "text", + "content": "and distilled into the student one by one [17]. However, it is difficult to find the optimal correspondence between the student and the teacher. The recent TAB works (e.g., [5]) still cannot help the student learn enough knowledge from the teacher when there is a large architecture gap between them, which is verified in our experiments. For our task, the student (MobileViT-v2) is a hybrid architecture, which has convolution layers in the shallow stages and separable self-attention layers in the later blocks. And a separable self-attention layer differs significantly from the traditional self-attention in two ways: 1) it does not learn an explicit attention map; 2) its input and output are still 3D, which are more similar to CNN features. Except for the separable self-attention in the later blocks, the shallow stages with convolution layers undoubtedly have a huge difference from the Transformer structure in CLIP, which has been investigated in previous works [31,41]. Therefore, we design SAB layer-wise alignment for KD of the intermediate layers from the Transformer to the CNN." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 543, + 287, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 543, + 287, + 651 + ], + "spans": [ + { + "bbox": [ + 46, + 543, + 287, + 651 + ], + "type": "text", + "content": "SAB Property. As shown in Fig. 2, the student's layers in our SAB KD can be explained as the bases of the feature space, and each layer of the teacher is a linear combination of the bases. After training, if the student's layers do ideally form the bases, the knowledge of the teacher's layers is learned completely by the student. Our experiment in Section 4.2 shows that the teacher's features in different layers can be well recovered from the features of the student's layers. Next, we describe how to implement SAB KD." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 653, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 653, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 47, + 653, + 288, + 715 + ], + "type": "text", + "content": "Feature Reshaping and Similarity Computation. Let the layers of the teacher and the student be " + }, + { + "bbox": [ + 47, + 653, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\text{Layer}^T = [L_1^T, L_2^T, \\ldots, L_M^T]" + }, + { + "bbox": [ + 47, + 653, + 288, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 653, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\text{Layer}^S = [L_1^S, L_2^S, \\ldots, L_N^S]" + }, + { + "bbox": [ + 47, + 653, + 288, + 715 + ], + "type": "text", + "content": " (usually " + }, + { + "bbox": [ + 47, + 653, + 288, + 715 + ], + "type": "inline_equation", + "content": "M > N" + }, + { + "bbox": [ + 47, + 653, + 288, + 715 + ], + "type": "text", + "content": "), respectively. " + }, + { + "bbox": [ + 47, + 653, + 288, + 715 + ], + "type": "inline_equation", + "content": "L_k^T \\in R^{D \\times L}" + }, + { + "bbox": [ + 47, + 653, + 288, + 715 + ], + "type": "text", + "content": " is a 2D tensor, where " + }, + { + "bbox": [ + 47, + 653, + 288, + 715 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 47, + 653, + 288, + 715 + ], + "type": "text", + "content": " is the dimensionality of the token feature" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 311, + 547, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 311, + 547, + 517 + ], + "spans": [ + { + "bbox": [ + 304, + 311, + 547, + 517 + ], + "type": "text", + "content": "and " + }, + { + "bbox": [ + 304, + 311, + 547, + 517 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 304, + 311, + 547, + 517 + ], + "type": "text", + "content": " is the number of tokens, while " + }, + { + "bbox": [ + 304, + 311, + 547, + 517 + ], + "type": "inline_equation", + "content": "L_{k}^{S}\\in R^{C_{k}\\times H_{k}\\times W_{k}}" + }, + { + "bbox": [ + 304, + 311, + 547, + 517 + ], + "type": "text", + "content": " is a 3D tensor, where " + }, + { + "bbox": [ + 304, + 311, + 547, + 517 + ], + "type": "inline_equation", + "content": "C_k" + }, + { + "bbox": [ + 304, + 311, + 547, + 517 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 311, + 547, + 517 + ], + "type": "inline_equation", + "content": "H_{k}" + }, + { + "bbox": [ + 304, + 311, + 547, + 517 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 311, + 547, + 517 + ], + "type": "inline_equation", + "content": "W_{k}" + }, + { + "bbox": [ + 304, + 311, + 547, + 517 + ], + "type": "text", + "content": " are the channel number, height and width of the " + }, + { + "bbox": [ + 304, + 311, + 547, + 517 + ], + "type": "inline_equation", + "content": "k^{th}" + }, + { + "bbox": [ + 304, + 311, + 547, + 517 + ], + "type": "text", + "content": " CNN layer's feature. We calculate a similarity tensor " + }, + { + "bbox": [ + 304, + 311, + 547, + 517 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 304, + 311, + 547, + 517 + ], + "type": "text", + "content": " between " + }, + { + "bbox": [ + 304, + 311, + 547, + 517 + ], + "type": "inline_equation", + "content": "L_{k}^{S}" + }, + { + "bbox": [ + 304, + 311, + 547, + 517 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 311, + 547, + 517 + ], + "type": "inline_equation", + "content": "L_{k}^{T}" + }, + { + "bbox": [ + 304, + 311, + 547, + 517 + ], + "type": "text", + "content": " which need to be reshaped first as shown in Figs. 3(a) and (b), respectively. For " + }, + { + "bbox": [ + 304, + 311, + 547, + 517 + ], + "type": "inline_equation", + "content": "L_{k}^{S}" + }, + { + "bbox": [ + 304, + 311, + 547, + 517 + ], + "type": "text", + "content": ", we first apply a learnable linear operator and then do pixel reshuffle on the result, obtaining " + }, + { + "bbox": [ + 304, + 311, + 547, + 517 + ], + "type": "inline_equation", + "content": "\\bar{L}_k^S" + }, + { + "bbox": [ + 304, + 311, + 547, + 517 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 311, + 547, + 517 + ], + "type": "inline_equation", + "content": "k = 1,2,\\dots,N" + }, + { + "bbox": [ + 304, + 311, + 547, + 517 + ], + "type": "text", + "content": ". Next, " + }, + { + "bbox": [ + 304, + 311, + 547, + 517 + ], + "type": "inline_equation", + "content": "\\bar{L}_k^S" + }, + { + "bbox": [ + 304, + 311, + 547, + 517 + ], + "type": "text", + "content": " is reshaped to " + }, + { + "bbox": [ + 304, + 311, + 547, + 517 + ], + "type": "inline_equation", + "content": "\\tilde{L}_k^S" + }, + { + "bbox": [ + 304, + 311, + 547, + 517 + ], + "type": "text", + "content": ". As for " + }, + { + "bbox": [ + 304, + 311, + 547, + 517 + ], + "type": "inline_equation", + "content": "L_{k}^{T}" + }, + { + "bbox": [ + 304, + 311, + 547, + 517 + ], + "type": "text", + "content": ", we remove the class token and obtain a " + }, + { + "bbox": [ + 304, + 311, + 547, + 517 + ], + "type": "inline_equation", + "content": "(L - 1)\\times D" + }, + { + "bbox": [ + 304, + 311, + 547, + 517 + ], + "type": "text", + "content": " tensor, denoted as " + }, + { + "bbox": [ + 304, + 311, + 547, + 517 + ], + "type": "inline_equation", + "content": "\\tilde{L}_k^T" + }, + { + "bbox": [ + 304, + 311, + 547, + 517 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 311, + 547, + 517 + ], + "type": "inline_equation", + "content": "k = 1,2,\\dots,M" + }, + { + "bbox": [ + 304, + 311, + 547, + 517 + ], + "type": "text", + "content": ". Finally, all " + }, + { + "bbox": [ + 304, + 311, + 547, + 517 + ], + "type": "inline_equation", + "content": "\\tilde{L}_k^S" + }, + { + "bbox": [ + 304, + 311, + 547, + 517 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 311, + 547, + 517 + ], + "type": "inline_equation", + "content": "k = 1,2,\\dots,N" + }, + { + "bbox": [ + 304, + 311, + 547, + 517 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 311, + 547, + 517 + ], + "type": "inline_equation", + "content": "\\tilde{L}_k^T" + }, + { + "bbox": [ + 304, + 311, + 547, + 517 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 311, + 547, + 517 + ], + "type": "inline_equation", + "content": "k = 1,2,\\dots,M" + }, + { + "bbox": [ + 304, + 311, + 547, + 517 + ], + "type": "text", + "content": ", are represented as " + }, + { + "bbox": [ + 304, + 311, + 547, + 517 + ], + "type": "inline_equation", + "content": "\\tilde{L}_{\\mathrm{er}}^{S} = [\\tilde{L}_{1}^{S},\\tilde{L}_{2}^{S},\\dots,\\tilde{L}_{N}^{S}]" + }, + { + "bbox": [ + 304, + 311, + 547, + 517 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 311, + 547, + 517 + ], + "type": "inline_equation", + "content": "\\tilde{L}_{\\mathrm{er}}^{T} = [\\tilde{L}_{1}^{T},\\tilde{L}_{2}^{T},\\dots,\\tilde{L}_{M}^{T}]" + }, + { + "bbox": [ + 304, + 311, + 547, + 517 + ], + "type": "text", + "content": ", respectively. After this reshaping, as shown in Fig. 3(c), we are able to measure the similarities between the intermediate layers of the student and the teacher through " + }, + { + "bbox": [ + 304, + 311, + 547, + 517 + ], + "type": "inline_equation", + "content": "W = \\tilde{L}_{\\mathrm{er}}^{T}\\times (\\tilde{L}_{\\mathrm{er}}^{S})^{\\top}" + }, + { + "bbox": [ + 304, + 311, + 547, + 517 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 311, + 547, + 517 + ], + "type": "inline_equation", + "content": "W\\in" + }, + { + "bbox": [ + 304, + 311, + 547, + 517 + ], + "type": "inline_equation", + "content": "R^{(L - 1)\\times M\\times N}" + }, + { + "bbox": [ + 304, + 311, + 547, + 517 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 518, + 547, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 518, + 547, + 552 + ], + "spans": [ + { + "bbox": [ + 304, + 518, + 547, + 552 + ], + "type": "text", + "content": "Masking and SAB Alignment. To speed up the training procedure, we design SAB KD with sequential masks, which follow this formula:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 329, + 558, + 545, + 598 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 329, + 558, + 545, + 598 + ], + "spans": [ + { + "bbox": [ + 329, + 558, + 545, + 598 + ], + "type": "interline_equation", + "content": "\\operatorname {m a s k} _ {(i, j)} = \\left\\{ \\begin{array}{l} 0, \\text {i f} j > n _ {0} \\text {a n d} i \\leq m _ {0} \\\\ 0, \\text {i f} j > n _ {1} \\text {a n d} m _ {0} < i \\leq m _ {1} \\\\ 1, \\text {e l s e} \\end{array} , \\right. \\tag {7}", + "image_path": "1e63e0c3314d7040f97ce52f6a9958280d6814496f276aa1d2bce10dbc26af80.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "type": "inline_equation", + "content": "i = 1,2,\\dots,M" + }, + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "type": "inline_equation", + "content": "j = 1,2,\\dots,N" + }, + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "type": "inline_equation", + "content": "1\\leq m_0 < m_1 < M" + }, + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "type": "inline_equation", + "content": "1\\leq n_0 < n_1 < N" + }, + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "type": "text", + "content": ". And the final attention mask " + }, + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\mathsf{Mask}\\in R^{(L - 1)\\times M\\times N}" + }, + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "type": "text", + "content": " is composed of " + }, + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "type": "inline_equation", + "content": "L - 1" + }, + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "type": "text", + "content": " same masks, Fig. 4(a) shows one example with " + }, + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "type": "inline_equation", + "content": "m_0 = 1,n_0 = 2,m_1 = 3,n_1 = 3" + }, + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\mathsf{mask}_{(i,j)} = 0" + }, + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "type": "text", + "content": " means to mask all the elements along the " + }, + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "type": "text", + "content": " dimension (Fig. 4(a)) at " + }, + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "type": "inline_equation", + "content": "(i,j)" + }, + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "type": "text", + "content": ", while " + }, + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\mathsf{mask}_{(i,j)} = 1" + }, + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "type": "text", + "content": " means to maintain their similarities, as shown in Fig. 4(b). This masking is based on our experimental finding: The student's lower-level layers" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "18987" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 289, + 192 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 289, + 192 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 289, + 192 + ], + "type": "text", + "content": "should learn from the teacher's lower-layers, while the student's higher layers should learn from all the teacher's layers. In Fig. 4(b), we also normalize the masks by performing softmax along the " + }, + { + "bbox": [ + 47, + 72, + 289, + 192 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 47, + 72, + 289, + 192 + ], + "type": "text", + "content": " dimension on each row, obtaining " + }, + { + "bbox": [ + 47, + 72, + 289, + 192 + ], + "type": "inline_equation", + "content": "\\tilde{W} = \\sigma (W\\circ \\mathsf{Mask})" + }, + { + "bbox": [ + 47, + 72, + 289, + 192 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 47, + 72, + 289, + 192 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 47, + 72, + 289, + 192 + ], + "type": "text", + "content": " is the softmax function and " + }, + { + "bbox": [ + 47, + 72, + 289, + 192 + ], + "type": "inline_equation", + "content": "\\circ" + }, + { + "bbox": [ + 47, + 72, + 289, + 192 + ], + "type": "text", + "content": " denotes the element-wise product. Guided by " + }, + { + "bbox": [ + 47, + 72, + 289, + 192 + ], + "type": "inline_equation", + "content": "\\tilde{W}" + }, + { + "bbox": [ + 47, + 72, + 289, + 192 + ], + "type": "text", + "content": ", as shown in Fig. 4(c), the weighted student layers are calculated as: " + }, + { + "bbox": [ + 47, + 72, + 289, + 192 + ], + "type": "inline_equation", + "content": "\\tilde{L a r e r}_{w}^{S} = \\tilde{W}\\times \\tilde{L a r e r}^{S}" + }, + { + "bbox": [ + 47, + 72, + 289, + 192 + ], + "type": "text", + "content": ". Each component of " + }, + { + "bbox": [ + 47, + 72, + 289, + 192 + ], + "type": "inline_equation", + "content": "\\tilde{L a r e r}_{w}^{S}" + }, + { + "bbox": [ + 47, + 72, + 289, + 192 + ], + "type": "text", + "content": " is the linear combination of " + }, + { + "bbox": [ + 47, + 72, + 289, + 192 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 47, + 72, + 289, + 192 + ], + "type": "text", + "content": " student layers' features. Finally, the SAB KD loss is defined as:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 95, + 199, + 287, + 215 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 95, + 199, + 287, + 215 + ], + "spans": [ + { + "bbox": [ + 95, + 199, + 287, + 215 + ], + "type": "interline_equation", + "content": "L _ {S A B} = D _ {K L} \\left(\\tilde {L a y e r} _ {w} ^ {S}, \\tilde {L a y e r} ^ {T}\\right). \\tag {8}", + "image_path": "610dbf011c2b175a7f4f704a13fe938ca69a157b50541b6473e10cc7a5ff7d9c.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 233, + 245, + 245 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 233, + 245, + 245 + ], + "spans": [ + { + "bbox": [ + 47, + 233, + 245, + 245 + ], + "type": "text", + "content": "3.3.3 Cross-Modality Knowledge distillation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 251, + 287, + 334 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 251, + 287, + 334 + ], + "spans": [ + { + "bbox": [ + 46, + 251, + 287, + 334 + ], + "type": "text", + "content": "To further adapt the student's features to the teacher's multimodal feature space, we employ the teacher's knowledge of cross-modality distributions to guide the training of the student. Specifically, we characterize the cross-modal distributions from two perspectives, global video-caption distributions of the teacher and local video-caption distributions of CLIP." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 335, + 287, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 335, + 287, + 392 + ], + "spans": [ + { + "bbox": [ + 47, + 335, + 287, + 392 + ], + "type": "text", + "content": "Global Video-Caption Distribution Alignment. We consider both the video-to-caption distribution " + }, + { + "bbox": [ + 47, + 335, + 287, + 392 + ], + "type": "inline_equation", + "content": "A_{GVC}" + }, + { + "bbox": [ + 47, + 335, + 287, + 392 + ], + "type": "text", + "content": " and the caption-to-video distribution " + }, + { + "bbox": [ + 47, + 335, + 287, + 392 + ], + "type": "inline_equation", + "content": "A_{GCV}" + }, + { + "bbox": [ + 47, + 335, + 287, + 392 + ], + "type": "text", + "content": ", the elements " + }, + { + "bbox": [ + 47, + 335, + 287, + 392 + ], + "type": "inline_equation", + "content": "(s(v_i, t_j)" + }, + { + "bbox": [ + 47, + 335, + 287, + 392 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 335, + 287, + 392 + ], + "type": "inline_equation", + "content": "s(t_j, v_i))" + }, + { + "bbox": [ + 47, + 335, + 287, + 392 + ], + "type": "text", + "content": " of which are defined via the similarities obtained by Eq. 4:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 83, + 397, + 246, + 433 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 83, + 397, + 246, + 433 + ], + "spans": [ + { + "bbox": [ + 83, + 397, + 246, + 433 + ], + "type": "interline_equation", + "content": "A _ {G V C} = \\sigma \\left( \\begin{array}{c c c} s (v _ {1}, t _ {1}) & ... & s (v _ {1}, t _ {B}) \\\\ ... & ... & ... \\\\ s (v _ {B}, t _ {1}) & ... & s (v _ {B}, t _ {j}) \\end{array} \\right),", + "image_path": "034d218ba36ae9f187d6fff4180542a48c712b8cc68321b5e2fce3d0f7ea980e.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 85, + 430, + 286, + 469 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 85, + 430, + 286, + 469 + ], + "spans": [ + { + "bbox": [ + 85, + 430, + 286, + 469 + ], + "type": "interline_equation", + "content": "A _ {G C V} = \\sigma \\left( \\begin{array}{c c c} s (t _ {1}, v _ {1}) & \\dots & s (t _ {1}, v _ {B}) \\\\ \\dots & \\dots & \\dots \\\\ s (t _ {B}, v _ {1}) & \\dots & s (t _ {B}, v _ {B}) \\end{array} \\right). \\tag {9}", + "image_path": "68784af36c09a797783ac85f5e03fbd7812151e93ab8c80c5bc4d2c27eb8ed64.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 475, + 287, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 475, + 287, + 510 + ], + "spans": [ + { + "bbox": [ + 47, + 475, + 287, + 510 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 47, + 475, + 287, + 510 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 47, + 475, + 287, + 510 + ], + "type": "text", + "content": " is the softmax function along the first dimension of the similarity matrixs. Finally, the global video-caption distribution alignment loss is defined as:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 58, + 515, + 287, + 529 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 515, + 287, + 529 + ], + "spans": [ + { + "bbox": [ + 58, + 515, + 287, + 529 + ], + "type": "interline_equation", + "content": "L _ {G} = D _ {K L} \\left(A _ {G V C} ^ {S}, A _ {G V C} ^ {T}\\right) + D _ {K L} \\left(A _ {G C V} ^ {S}, A _ {G C V} ^ {T}\\right). \\tag {10}", + "image_path": "6a7c3ad4204d23ccea40f15225652592454992c8895c12776afc7b6cacc02a89.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 536, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 536, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 536, + 288, + 713 + ], + "type": "text", + "content": "Local Video-Caption Distribution Alignment. CLIP uses text prompts (such as \"A picture of a (\") for zero-shot image classification. CLIP fills them with different words (e.g., \"cat\" and \"dog\") and results in different captions (e.g., \"A picture of a cat\" and \"A picture of a dog\"). It can match the captions to the corresponding images, showing some image-word alignment ability. We transfer this pre-training knowledge to the student through a local frame-word alignment as follows. Recall that " + }, + { + "bbox": [ + 47, + 536, + 288, + 713 + ], + "type": "inline_equation", + "content": "Z_{i}^{\\prime} = \\{z_{i1}^{\\prime}, z_{i2}^{\\prime}, \\dots, z_{in}^{\\prime}\\}" + }, + { + "bbox": [ + 47, + 536, + 288, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 536, + 288, + 713 + ], + "type": "inline_equation", + "content": "W_{j} = \\{w_{j1}, w_{j2}, \\dots, w_{jm}\\}" + }, + { + "bbox": [ + 47, + 536, + 288, + 713 + ], + "type": "text", + "content": " are the features of the video " + }, + { + "bbox": [ + 47, + 536, + 288, + 713 + ], + "type": "inline_equation", + "content": "v_{i}" + }, + { + "bbox": [ + 47, + 536, + 288, + 713 + ], + "type": "text", + "content": " and the caption " + }, + { + "bbox": [ + 47, + 536, + 288, + 713 + ], + "type": "inline_equation", + "content": "t_{j}" + }, + { + "bbox": [ + 47, + 536, + 288, + 713 + ], + "type": "text", + "content": ", respectively, with " + }, + { + "bbox": [ + 47, + 536, + 288, + 713 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 47, + 536, + 288, + 713 + ], + "type": "text", + "content": " being the number of frames in " + }, + { + "bbox": [ + 47, + 536, + 288, + 713 + ], + "type": "inline_equation", + "content": "v_{i}" + }, + { + "bbox": [ + 47, + 536, + 288, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 536, + 288, + 713 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 47, + 536, + 288, + 713 + ], + "type": "text", + "content": " the number of words in " + }, + { + "bbox": [ + 47, + 536, + 288, + 713 + ], + "type": "inline_equation", + "content": "t_{j}" + }, + { + "bbox": [ + 47, + 536, + 288, + 713 + ], + "type": "text", + "content": " (Section 3.1). The similarity between the " + }, + { + "bbox": [ + 47, + 536, + 288, + 713 + ], + "type": "inline_equation", + "content": "k^{th}" + }, + { + "bbox": [ + 47, + 536, + 288, + 713 + ], + "type": "text", + "content": " frame and the " + }, + { + "bbox": [ + 47, + 536, + 288, + 713 + ], + "type": "inline_equation", + "content": "r^{th}" + }, + { + "bbox": [ + 47, + 536, + 288, + 713 + ], + "type": "text", + "content": " word is defined as " + }, + { + "bbox": [ + 47, + 536, + 288, + 713 + ], + "type": "inline_equation", + "content": "s_{fw}(w_{jr}, z_{ik}') = (w_{jr})^\\top z_{ik}'" + }, + { + "bbox": [ + 47, + 536, + 288, + 713 + ], + "type": "text", + "content": ". Then we respectively define the local video-to-caption and caption-to-video similarities as:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 344, + 71, + 545, + 103 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 344, + 71, + 545, + 103 + ], + "spans": [ + { + "bbox": [ + 344, + 71, + 545, + 103 + ], + "type": "interline_equation", + "content": "s ^ {\\prime} \\left(v _ {i}, t _ {j}\\right) = \\frac {1}{n} \\sum_ {k = 1} ^ {n} \\max _ {1 \\leq r \\leq m} \\left\\{s _ {f w} \\left(w _ {j r}, z _ {i k} ^ {\\prime}\\right) \\right\\}, \\tag {11}", + "image_path": "5ff7f120a4e9b124a5b7977b904bb243282de87c99be0d12460ed2925756d10d.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 344, + 101, + 507, + 128 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 344, + 101, + 507, + 128 + ], + "spans": [ + { + "bbox": [ + 344, + 101, + 507, + 128 + ], + "type": "interline_equation", + "content": "s ^ {\\prime} (t _ {j}, v _ {i}) = \\frac {1}{m} \\sum_ {r = 1} ^ {m} \\max _ {1 \\leq k \\leq n} \\left\\{s _ {f w} \\left(z _ {i k} ^ {\\prime}, w _ {j r}\\right) \\right\\}.", + "image_path": "f17d5abfbbc13632d7b07e80a73eca07f68514c19672702c8d79a83786aa7c3d.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 134, + 545, + 158 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 134, + 545, + 158 + ], + "spans": [ + { + "bbox": [ + 306, + 134, + 545, + 158 + ], + "type": "text", + "content": "Finally, we have the local video-to-caption distribution " + }, + { + "bbox": [ + 306, + 134, + 545, + 158 + ], + "type": "inline_equation", + "content": "A_{LVC}" + }, + { + "bbox": [ + 306, + 134, + 545, + 158 + ], + "type": "text", + "content": " and the local caption-to-video distribution " + }, + { + "bbox": [ + 306, + 134, + 545, + 158 + ], + "type": "inline_equation", + "content": "A_{LCV}" + }, + { + "bbox": [ + 306, + 134, + 545, + 158 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 339, + 164, + 545, + 205 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 339, + 164, + 545, + 205 + ], + "spans": [ + { + "bbox": [ + 339, + 164, + 545, + 205 + ], + "type": "interline_equation", + "content": "A _ {L V C} = \\sigma \\left( \\begin{array}{c c c} s ^ {\\prime} \\left(v _ {1}, t _ {1}\\right) & \\dots & s ^ {\\prime} \\left(v _ {1}, t _ {B}\\right) \\\\ \\dots & \\dots & \\dots \\\\ s ^ {\\prime} \\left(v _ {B}, t _ {1}\\right) & \\dots & s ^ {\\prime} \\left(v _ {B}, t _ {B}\\right) \\end{array} \\right), \\tag {12}", + "image_path": "ce77dafc7440da18bb6124cfd0eab0fd00cae54b03d9ae1626e58fa5ccf01df3.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 344, + 201, + 512, + 236 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 344, + 201, + 512, + 236 + ], + "spans": [ + { + "bbox": [ + 344, + 201, + 512, + 236 + ], + "type": "interline_equation", + "content": "A _ {L C V} = \\sigma \\left( \\begin{array}{c c c} s ^ {\\prime} (t _ {1}, v _ {1}) & ... & s ^ {\\prime} (t _ {1}, v _ {B}) \\\\ ... & ... & ... \\\\ s ^ {\\prime} (t _ {B}, v _ {1}) & ... & s ^ {\\prime} (t _ {B}, v _ {B}) \\end{array} \\right),", + "image_path": "fc96d151163f93fc7dbb89159050aeb0260e875a6e03805f700b2c62a198e04d.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 236, + 545, + 257 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 236, + 545, + 257 + ], + "spans": [ + { + "bbox": [ + 306, + 236, + 545, + 257 + ], + "type": "text", + "content": "and the local caption-video distribution alignment loss is defined as:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 318, + 258, + 545, + 272 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 258, + 545, + 272 + ], + "spans": [ + { + "bbox": [ + 318, + 258, + 545, + 272 + ], + "type": "interline_equation", + "content": "L _ {L} = D _ {K L} \\left(A _ {L C V} ^ {S}, A _ {L C V} ^ {T}\\right) + D _ {K L} \\left(A _ {L V C} ^ {S}, A _ {L V C} ^ {T}\\right). \\tag {13}", + "image_path": "9f7688d8f208a99cf6357e81e0aaeff431d2102ed9d47b4fcdd2495173544d2d.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 276, + 531, + 288 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 276, + 531, + 288 + ], + "spans": [ + { + "bbox": [ + 307, + 276, + 531, + 288 + ], + "type": "text", + "content": "Combining " + }, + { + "bbox": [ + 307, + 276, + 531, + 288 + ], + "type": "inline_equation", + "content": "L_{G}" + }, + { + "bbox": [ + 307, + 276, + 531, + 288 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 307, + 276, + 531, + 288 + ], + "type": "inline_equation", + "content": "L_{V}" + }, + { + "bbox": [ + 307, + 276, + 531, + 288 + ], + "type": "text", + "content": ", the cross-modality KD loss is:" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 386, + 291, + 545, + 304 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 386, + 291, + 545, + 304 + ], + "spans": [ + { + "bbox": [ + 386, + 291, + 545, + 304 + ], + "type": "interline_equation", + "content": "L _ {C M} = L _ {G} + L _ {L}. \\tag {14}", + "image_path": "b0d53802e095e94b56e1f12bd54f72d13e32b9ad9872264adff09ed7fcc06b50.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 308, + 545, + 367 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 308, + 545, + 367 + ], + "spans": [ + { + "bbox": [ + 304, + 308, + 545, + 367 + ], + "type": "text", + "content": "Our experiment in Section 4.2 shows that only when joint trained with SAB, our local video-caption distribution alignment shows its advantage, which also verifies the effectiveness of SAB. Finally, the total loss for training our model is:" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 350, + 369, + 545, + 397 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 350, + 369, + 545, + 397 + ], + "spans": [ + { + "bbox": [ + 350, + 369, + 545, + 397 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} L = L _ {\\text {t a s k}} + \\alpha \\cdot L _ {T K D} + \\beta \\cdot L _ {S K D} \\tag {15} \\\\ + \\gamma \\cdot L _ {C M} + \\delta \\cdot L _ {S A B}, \\\\ \\end{array}", + "image_path": "ec4d4318d0e0d9abfa020609ce1cfc666ad8ab5fd69a8b0edd18a43f4c8b6a5d.jpg" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 305, + 406, + 545, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 406, + 545, + 430 + ], + "spans": [ + { + "bbox": [ + 305, + 406, + 545, + 430 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 406, + 545, + 430 + ], + "type": "inline_equation", + "content": "L_{task}" + }, + { + "bbox": [ + 305, + 406, + 545, + 430 + ], + "type": "text", + "content": " is the task-specifical loss, and " + }, + { + "bbox": [ + 305, + 406, + 545, + 430 + ], + "type": "inline_equation", + "content": "\\alpha ,\\beta ,\\gamma" + }, + { + "bbox": [ + 305, + 406, + 545, + 430 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 406, + 545, + 430 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 305, + 406, + 545, + 430 + ], + "type": "text", + "content": " are loss balance weights." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 306, + 441, + 388, + 453 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 441, + 388, + 453 + ], + "spans": [ + { + "bbox": [ + 306, + 441, + 388, + 453 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 305, + 460, + 545, + 519 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 460, + 545, + 519 + ], + "spans": [ + { + "bbox": [ + 305, + 460, + 545, + 519 + ], + "type": "text", + "content": "We conduct comprehensive experiments on three benchmarks for video-text retrieval (video-to-text " + }, + { + "bbox": [ + 305, + 460, + 545, + 519 + ], + "type": "inline_equation", + "content": "(v2t)" + }, + { + "bbox": [ + 305, + 460, + 545, + 519 + ], + "type": "text", + "content": " and text-to-video " + }, + { + "bbox": [ + 305, + 460, + 545, + 519 + ], + "type": "inline_equation", + "content": "(t2v)" + }, + { + "bbox": [ + 305, + 460, + 545, + 519 + ], + "type": "text", + "content": "): MSR-VTT [38], MSVD [4] and LSMDC [32]. The metrics Recall at rank K " + }, + { + "bbox": [ + 305, + 460, + 545, + 519 + ], + "type": "inline_equation", + "content": "(R@K)" + }, + { + "bbox": [ + 305, + 460, + 545, + 519 + ], + "type": "text", + "content": " are used for evaluation." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 305, + 528, + 492, + 540 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 528, + 492, + 540 + ], + "spans": [ + { + "bbox": [ + 305, + 528, + 492, + 540 + ], + "type": "text", + "content": "4.1. Comparison with State-of-the-Arts" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 304, + 545, + 545, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 545, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 545, + 545, + 715 + ], + "type": "text", + "content": "MSR-VTT. In Table 1, we compare the proposed model CLIPPING with eight state-of-the-art methods on MSR-VTT: TACo [39], VideoClip [37], Frozen [3], VIOLET [12], OA-Trans [35], BridgeFormer [13], ALL-in-one [34] and MDMMT [11]. It can be seen that CLIPPING significantly surpasses those large-scale video-text/image-text pre-training models for video-text retrieval. For example, our model exceeds VideoClip and Frozen by absolute " + }, + { + "bbox": [ + 304, + 545, + 545, + 715 + ], + "type": "inline_equation", + "content": "9.8\\%" + }, + { + "bbox": [ + 304, + 545, + 545, + 715 + ], + "type": "text", + "content": " t2vR@1 and " + }, + { + "bbox": [ + 304, + 545, + 545, + 715 + ], + "type": "inline_equation", + "content": "8.2\\%" + }, + { + "bbox": [ + 304, + 545, + 545, + 715 + ], + "type": "text", + "content": " t2vR@1, respectively. Note that the model MDMMT uses CLIP as its backbone, while our CLIPPING uses CLIP as the teacher. For a more efficient and edge deployable video-language retrieval model, we further compress CLIPPING into two models: " + }, + { + "bbox": [ + 304, + 545, + 545, + 715 + ], + "type": "inline_equation", + "content": "\\mathrm{CLIPPING}_{\\mathrm{W / oT}}" + }, + { + "bbox": [ + 304, + 545, + 545, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 545, + 545, + 715 + ], + "type": "inline_equation", + "content": "\\mathrm{CLIPPING^{*}}" + }, + { + "bbox": [ + 304, + 545, + 545, + 715 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 304, + 545, + 545, + 715 + ], + "type": "inline_equation", + "content": "\\mathrm{CLIPPING}_{\\mathrm{W / oT}}" + } + ] + } + ], + "index": 25 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18988" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 57, + 70, + 276, + 222 + ], + "blocks": [ + { + "bbox": [ + 57, + 70, + 276, + 222 + ], + "lines": [ + { + "bbox": [ + 57, + 70, + 276, + 222 + ], + "spans": [ + { + "bbox": [ + 57, + 70, + 276, + 222 + ], + "type": "table", + "html": "
ModelPT DatasetsParamsR@1
TACoHT100M212M28.4
VideoClipHT100M130M30.9
FrozenC3M,W2M,COCO232M32.5
ALL-in-one-SW2M,HT100M33M33.5
VIOLETC3M,W2M198M34.5
OA-TransC3M,W2M232M35.8
BridgeFormerC3M,W2M160M37.6
ALL-in-one-BW2M,HT100M110M37.9
MDMMTC400M,AudioSet226M38.9
CLIPPING*w/o T-8.7M37.5
CLIPPINGw/o T-21.3M38.6
CLIPPING*-46.5M39.8
CLIPPING-55.0M40.6
CLIPPINGIN21K55.0M40.7
", + "image_path": "739f78667240eaaf71dd20d5aa0773dc4969f4ca6ab6104395d9417add1981e3.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 46, + 232, + 287, + 310 + ], + "lines": [ + { + "bbox": [ + 46, + 232, + 287, + 310 + ], + "spans": [ + { + "bbox": [ + 46, + 232, + 287, + 310 + ], + "type": "text", + "content": "Table 1. Comparison with state-of-the-art models on MSR-VTT (1k split) for text-to-video retrieval. \"PT Datasets\": datasets used for pre-training the vision encoder. \"HT100M\": HowTo100M dataset [28]. \"C400M\": CLIP-400M dataset [30]. \"IN21K\": ImageNet21K dataset [9]. \"W2M\": WebVid-2M dataset [3]. \"C3M\": CC3M dataset [33]. All the methods in Table are the comparable methods with parameters less than 240M and flops less than 360G." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 318, + 288, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 318, + 288, + 449 + ], + "spans": [ + { + "bbox": [ + 46, + 318, + 288, + 449 + ], + "type": "text", + "content": "does not use the temporal Transformer and CLIPPING* compresses the text encoder through TinyBERT [17]. It can be seen from Table 1 that the CLIPPING* based models also outperform the small model ALL-in-one-S even though they are smaller. In Fig. 5, we show the performances and Flops of these models. Among previous models, MDMMT has the best performance and ALL-in-one-S is the fastest. CLIPPING not only obtains relative " + }, + { + "bbox": [ + 46, + 318, + 288, + 449 + ], + "type": "inline_equation", + "content": "4.6\\%" + }, + { + "bbox": [ + 46, + 318, + 288, + 449 + ], + "type": "text", + "content": " performance gain over MDMMT but also is faster than All-in-one-S. Table 2 provides the details of Flops and parameters for each module of our method." + } + ] + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 82, + 455, + 256, + 550 + ], + "blocks": [ + { + "bbox": [ + 82, + 455, + 256, + 550 + ], + "lines": [ + { + "bbox": [ + 82, + 455, + 256, + 550 + ], + "spans": [ + { + "bbox": [ + 82, + 455, + 256, + 550 + ], + "type": "image", + "image_path": "f021df41a4d4a96d5066586f05fea5c73538a85e41bd2b20fad19379d7217281.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 103, + 555, + 230, + 567 + ], + "lines": [ + { + "bbox": [ + 103, + 555, + 230, + 567 + ], + "spans": [ + { + "bbox": [ + 103, + 555, + 230, + 567 + ], + "type": "text", + "content": "Figure 5. Flops and Performances." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 568, + 287, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 568, + 287, + 665 + ], + "spans": [ + { + "bbox": [ + 46, + 568, + 287, + 665 + ], + "type": "text", + "content": "Other Benchmarks. We also compare our CLIPPING with the state-of-the-art Frozen [3], MDMMT [11], NoiseEst [2] and SupportSet [29] on the LSMDC and MSVD datasets. In Table 3, CLIPPING again obtains the best performance with fewest parameters and Flops, which achieves " + }, + { + "bbox": [ + 46, + 568, + 287, + 665 + ], + "type": "inline_equation", + "content": "88.1\\% - 92.9\\%" + }, + { + "bbox": [ + 46, + 568, + 287, + 665 + ], + "type": "text", + "content": " of the performance of its teacher on the LSMDC and MSVD datasets, showing consistent improvements and generalization across different datasets." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 671, + 141, + 684 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 671, + 141, + 684 + ], + "spans": [ + { + "bbox": [ + 47, + 671, + 141, + 684 + ], + "type": "text", + "content": "4.2. Ablation Study" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "content": "Key Components. We provide detailed ablation study to validate each key component of our proposed method," + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 326, + 70, + 526, + 125 + ], + "blocks": [ + { + "bbox": [ + 326, + 70, + 526, + 125 + ], + "lines": [ + { + "bbox": [ + 326, + 70, + 526, + 125 + ], + "spans": [ + { + "bbox": [ + 326, + 70, + 526, + 125 + ], + "type": "table", + "html": "
ModulesParametersFlops
Vision encoder (MobileViT-v2)4.5M16.8G
Temporal Transformer37.8M1.2G
Language encoder (CLIPbert)12.6M0.15G
Language encoder (TinyBERT)4.2M0.05G
", + "image_path": "9e29ff45209e679dcc75f85e246ff9570118db5371ea61fb06a3ba4c0d366e08.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 307, + 163, + 553, + 289 + ], + "blocks": [ + { + "bbox": [ + 306, + 134, + 544, + 156 + ], + "lines": [ + { + "bbox": [ + 306, + 134, + 544, + 156 + ], + "spans": [ + { + "bbox": [ + 306, + 134, + 544, + 156 + ], + "type": "text", + "content": "Table 2. The Flops and parameters of each module in our method. We adopt 12 frames for each video." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 307, + 163, + 553, + 289 + ], + "lines": [ + { + "bbox": [ + 307, + 163, + 553, + 289 + ], + "spans": [ + { + "bbox": [ + 307, + 163, + 553, + 289 + ], + "type": "table", + "html": "
ModelPT DatasetsVision EncodersR@1
LSMDC dataset
NoiseEstHT100MResnet152,ResNext-1016.4
SupportSetHT100MResNet-152,R(2+1)D-34IN21K
FrozenC3M,W2MSpace-Time15.0
MDMMTC400MCLIP_vision18.8
CLIPPING-MobileViT-v219.9
MSVD dataset
NoiseEstHT100MResnet152,ResNext-10120.3
SupportSetHT100MResNet-152,R(2+1)D-3428.4
FrozenC3M,W2MSpace-Time33.7
CLIPPING-MobileViT-v242.9
", + "image_path": "52abb9087708028c103f08d42eacb6b4692654ec42e4a99dc2048b0dd56cc8bf.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 307, + 322, + 545, + 407 + ], + "blocks": [ + { + "bbox": [ + 306, + 297, + 545, + 319 + ], + "lines": [ + { + "bbox": [ + 306, + 297, + 545, + 319 + ], + "spans": [ + { + "bbox": [ + 306, + 297, + 545, + 319 + ], + "type": "text", + "content": "Table 3. Comparison with the state-of-the-art on the LSMDC and MSVD datasets." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 307, + 322, + 545, + 407 + ], + "lines": [ + { + "bbox": [ + 307, + 322, + 545, + 407 + ], + "spans": [ + { + "bbox": [ + 307, + 322, + 545, + 407 + ], + "type": "table", + "html": "
Vision EncoderKD Typest2vR@1v2tR@1
CLIP_vision-44.542.2
MobileViTv2-25.724.5
MobileViTv2T28.827.3
MobileViTv2T,S33.032.8
MobileViTv2T,S,SAB37.636.2
MobileViTv2T,S,SAB,CMG39.639.1
MobileViTv2T,S,SAB,CMG,CML40.740.2
", + "image_path": "bc1c66a24d1e60106d081e524581a5e57a5f034d132194fac1ccd06dc7e3e45c.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_body" + } + ], + "index": 12 + }, + { + "type": "table", + "bbox": [ + 336, + 479, + 516, + 547 + ], + "blocks": [ + { + "bbox": [ + 305, + 415, + 545, + 471 + ], + "lines": [ + { + "bbox": [ + 305, + 415, + 545, + 471 + ], + "spans": [ + { + "bbox": [ + 305, + 415, + 545, + 471 + ], + "type": "text", + "content": "Table 4. Ablation study of different KD components of CLIPPING on the " + }, + { + "bbox": [ + 305, + 415, + 545, + 471 + ], + "type": "inline_equation", + "content": "1k" + }, + { + "bbox": [ + 305, + 415, + 545, + 471 + ], + "type": "text", + "content": " validation set of MSR-VTT. The first row is the results of the teacher model (Clip4clip). " + }, + { + "bbox": [ + 305, + 415, + 545, + 471 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 305, + 415, + 545, + 471 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 305, + 415, + 545, + 471 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 305, + 415, + 545, + 471 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 305, + 415, + 545, + 471 + ], + "type": "inline_equation", + "content": "SAB" + }, + { + "bbox": [ + 305, + 415, + 545, + 471 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 305, + 415, + 545, + 471 + ], + "type": "inline_equation", + "content": "CM_{G}" + }, + { + "bbox": [ + 305, + 415, + 545, + 471 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 415, + 545, + 471 + ], + "type": "inline_equation", + "content": "CM_{L}" + }, + { + "bbox": [ + 305, + 415, + 545, + 471 + ], + "type": "text", + "content": " denote temporal KD, spatial KD, SAB KD, global cross-modality KD and local cross-modality KD, respectively." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 336, + 479, + 516, + 547 + ], + "lines": [ + { + "bbox": [ + 336, + 479, + 516, + 547 + ], + "spans": [ + { + "bbox": [ + 336, + 479, + 516, + 547 + ], + "type": "table", + "html": "
KD Typest2vR@1v2tR@1
T, S33.032.8
T, S, OL2OL34.633.4
T, S, TAB35.134.4
T, S, SABw/o masking37.135.7
T, S, SABw/masking37.636.2
", + "image_path": "c65775a9850dab166bad0cbd85906b5582a8282bef4d3d70e65442542f3c8f9f.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "table_body" + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 555, + 545, + 588 + ], + "lines": [ + { + "bbox": [ + 305, + 555, + 545, + 588 + ], + "spans": [ + { + "bbox": [ + 305, + 555, + 545, + 588 + ], + "type": "text", + "content": "Table 5. Ablation study of different KD types on MSR-VTT (1k split). All the models are trained for 36 epochs with the same setting (see the supplementary materials for details)." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "text", + "content": "on MSR-VTT. From Table 4, we can see that without any KD, it gets extreme low accuracies. When simply adding the temporal KD and spatial KD, " + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "inline_equation", + "content": "t2vR@1" + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "inline_equation", + "content": "v2tR@1" + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "text", + "content": " increase significantly. When we further add SAB KD, and then global and local cross-modality KD, CLIPPING's performance rises gradually. This study shows that all these key components of CLIPPING are effective. Compared with the teacher, the full CLIPPING achieves about " + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "inline_equation", + "content": "91.5\\%" + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "inline_equation", + "content": "95.3\\%" + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "text", + "content": " of the performance of its teacher on " + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "inline_equation", + "content": "t2vR@1" + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "inline_equation", + "content": "v2tR@1" + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "text", + "content": ", respectively, with the vision encoder of " + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "inline_equation", + "content": "19.5x" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18989" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 73, + 71, + 147, + 133 + ], + "blocks": [ + { + "bbox": [ + 73, + 71, + 147, + 133 + ], + "lines": [ + { + "bbox": [ + 73, + 71, + 147, + 133 + ], + "spans": [ + { + "bbox": [ + 73, + 71, + 147, + 133 + ], + "type": "image", + "image_path": "73f8c893c779df8ba44d8439f9e1534e0ecf89a956bd2a0389867668625eb2de.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 147, + 72, + 206, + 133 + ], + "blocks": [ + { + "bbox": [ + 147, + 72, + 206, + 133 + ], + "lines": [ + { + "bbox": [ + 147, + 72, + 206, + 133 + ], + "spans": [ + { + "bbox": [ + 147, + 72, + 206, + 133 + ], + "type": "image", + "image_path": "742b81165d4b07bc807f031179c52135084b0e830e3e585e0a32424e21459974.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 75, + 139, + 147, + 198 + ], + "blocks": [ + { + "bbox": [ + 75, + 139, + 147, + 198 + ], + "lines": [ + { + "bbox": [ + 75, + 139, + 147, + 198 + ], + "spans": [ + { + "bbox": [ + 75, + 139, + 147, + 198 + ], + "type": "image", + "image_path": "33f57a089fd3cfcb94f2bb72e7648f0fc9d7a95a37217a34da9afc01ccefa7c5.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 128, + 199, + 138, + 207 + ], + "lines": [ + { + "bbox": [ + 128, + 199, + 138, + 207 + ], + "spans": [ + { + "bbox": [ + 128, + 199, + 138, + 207 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 147, + 139, + 223, + 199 + ], + "blocks": [ + { + "bbox": [ + 147, + 139, + 223, + 199 + ], + "lines": [ + { + "bbox": [ + 147, + 139, + 223, + 199 + ], + "spans": [ + { + "bbox": [ + 147, + 139, + 223, + 199 + ], + "type": "image", + "image_path": "ae87ddfc9026dc088bee4452a793c3d753df1d9cc269a3f17683edf6bb56cfbe.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 45, + 215, + 547, + 237 + ], + "lines": [ + { + "bbox": [ + 45, + 215, + 547, + 237 + ], + "spans": [ + { + "bbox": [ + 45, + 215, + 547, + 237 + ], + "type": "text", + "content": "Figure 6. (a) Feature tensors of the student and the teacher. (b) An example to demonstrate that the teacher's features are the linear combinations of the student features. More examples are provided in the supplementary materials." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 232, + 100, + 301, + 159 + ], + "blocks": [ + { + "bbox": [ + 232, + 100, + 301, + 159 + ], + "lines": [ + { + "bbox": [ + 232, + 100, + 301, + 159 + ], + "spans": [ + { + "bbox": [ + 232, + 100, + 301, + 159 + ], + "type": "image", + "image_path": "9478f3c22d50742c4e8d4d5a73076b6d041a02a72c7f2e36d93cefb92076c799.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 258, + 175, + 277, + 183 + ], + "lines": [ + { + "bbox": [ + 258, + 175, + 277, + 183 + ], + "spans": [ + { + "bbox": [ + 258, + 175, + 277, + 183 + ], + "type": "text", + "content": "image" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 304, + 86, + 353, + 159 + ], + "blocks": [ + { + "bbox": [ + 304, + 86, + 353, + 159 + ], + "lines": [ + { + "bbox": [ + 304, + 86, + 353, + 159 + ], + "spans": [ + { + "bbox": [ + 304, + 86, + 353, + 159 + ], + "type": "image", + "image_path": "1e533e4a9c58a15c471cb646cc25ff952aa2bd185a3aa88008b9ae8f586f779c.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 359, + 198, + 370, + 207 + ], + "lines": [ + { + "bbox": [ + 359, + 198, + 370, + 207 + ], + "spans": [ + { + "bbox": [ + 359, + 198, + 370, + 207 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 321, + 175, + 354, + 182 + ], + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 175, + 354, + 182 + ], + "spans": [ + { + "bbox": [ + 321, + 175, + 354, + 182 + ], + "type": "inline_equation", + "content": "\\widetilde{W}" + }, + { + "bbox": [ + 321, + 175, + 354, + 182 + ], + "type": "inline_equation", + "content": "\\tilde{L}_{er^S}" + } + ] + } + ], + "index": 8, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 358, + 85, + 410, + 158 + ], + "blocks": [ + { + "bbox": [ + 358, + 85, + 410, + 158 + ], + "lines": [ + { + "bbox": [ + 358, + 85, + 410, + 158 + ], + "spans": [ + { + "bbox": [ + 358, + 85, + 410, + 158 + ], + "type": "image", + "image_path": "dc1f322bc5138a618a9ab46be04c29d67cef79100d57a076656c02275a3985bf.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 367, + 158, + 403, + 171 + ], + "blocks": [ + { + "bbox": [ + 367, + 158, + 403, + 171 + ], + "lines": [ + { + "bbox": [ + 367, + 158, + 403, + 171 + ], + "spans": [ + { + "bbox": [ + 367, + 158, + 403, + 171 + ], + "type": "image", + "image_path": "d059336c5d5b7be3021a7959f093eb78e2e6fa1421e84c8a4f16aa8fb02d3c8c.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 397, + 175, + 482, + 182 + ], + "lines": [ + { + "bbox": [ + 397, + 175, + 482, + 182 + ], + "spans": [ + { + "bbox": [ + 397, + 175, + 482, + 182 + ], + "type": "inline_equation", + "content": "\\widetilde{W}\\times \\widetilde{L}\\text{ayer}^{S} = \\widetilde{L}\\text{layer}_{w}^{S}\\longrightarrow \\widetilde{L}\\text{ayer}^{T}" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 321, + 175, + 354, + 182 + ], + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 175, + 354, + 182 + ], + "spans": [ + { + "bbox": [ + 321, + 175, + 354, + 182 + ], + "type": "inline_equation", + "content": "\\widetilde{W}" + }, + { + "bbox": [ + 321, + 175, + 354, + 182 + ], + "type": "inline_equation", + "content": "\\tilde{L}_{\\text{ayer}}^s" + } + ] + } + ], + "index": 11, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 413, + 85, + 462, + 158 + ], + "blocks": [ + { + "bbox": [ + 413, + 85, + 462, + 158 + ], + "lines": [ + { + "bbox": [ + 413, + 85, + 462, + 158 + ], + "spans": [ + { + "bbox": [ + 413, + 85, + 462, + 158 + ], + "type": "image", + "image_path": "c1f0481e564156a055f51cd1b7d0ae7526aa9c76b3d13ddc340c79915fcf8395.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 465, + 85, + 516, + 158 + ], + "blocks": [ + { + "bbox": [ + 465, + 85, + 516, + 158 + ], + "lines": [ + { + "bbox": [ + 465, + 85, + 516, + 158 + ], + "spans": [ + { + "bbox": [ + 465, + 85, + 516, + 158 + ], + "type": "image", + "image_path": "a0a6ca7b3952a317d5be7a57280e6aa18cd90dbd1d7daa7b2ba3dc42307283af.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 100, + 245, + 227, + 323 + ], + "blocks": [ + { + "bbox": [ + 100, + 245, + 227, + 323 + ], + "lines": [ + { + "bbox": [ + 100, + 245, + 227, + 323 + ], + "spans": [ + { + "bbox": [ + 100, + 245, + 227, + 323 + ], + "type": "image", + "image_path": "e6a8dc3f1aa2c988b6130a5cc0fd982a7777218e021808ec0604c0e020615efa.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 159, + 327, + 169, + 336 + ], + "lines": [ + { + "bbox": [ + 159, + 327, + 169, + 336 + ], + "spans": [ + { + "bbox": [ + 159, + 327, + 169, + 336 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 233, + 245, + 361, + 323 + ], + "blocks": [ + { + "bbox": [ + 233, + 245, + 361, + 323 + ], + "lines": [ + { + "bbox": [ + 233, + 245, + 361, + 323 + ], + "spans": [ + { + "bbox": [ + 233, + 245, + 361, + 323 + ], + "type": "image", + "image_path": "0396bafb1c9c14e77e24411b8cb9818137cc28c575ff43b4583a7613a5968841.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 297, + 327, + 307, + 336 + ], + "lines": [ + { + "bbox": [ + 297, + 327, + 307, + 336 + ], + "spans": [ + { + "bbox": [ + 297, + 327, + 307, + 336 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 366, + 245, + 493, + 323 + ], + "blocks": [ + { + "bbox": [ + 366, + 245, + 493, + 323 + ], + "lines": [ + { + "bbox": [ + 366, + 245, + 493, + 323 + ], + "spans": [ + { + "bbox": [ + 366, + 245, + 493, + 323 + ], + "type": "image", + "image_path": "844436a69d6dc0fef38a47a216ad45f5b999d87491fd9f31476371e60bd75c3b.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 425, + 327, + 435, + 336 + ], + "lines": [ + { + "bbox": [ + 425, + 327, + 435, + 336 + ], + "spans": [ + { + "bbox": [ + 425, + 327, + 435, + 336 + ], + "type": "text", + "content": "(c)" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 108, + 342, + 484, + 354 + ], + "lines": [ + { + "bbox": [ + 108, + 342, + 484, + 354 + ], + "spans": [ + { + "bbox": [ + 108, + 342, + 484, + 354 + ], + "type": "text", + "content": "Figure 7. Visualization of the frame-word alignments. (a) CLIP; (b) " + }, + { + "bbox": [ + 108, + 342, + 484, + 354 + ], + "type": "inline_equation", + "content": "\\mathrm{CLIPPING}^{-}_{\\mathrm{TAB}}" + }, + { + "bbox": [ + 108, + 342, + 484, + 354 + ], + "type": "text", + "content": "; (c) CLIPPING." + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 21 + }, + { + "bbox": [ + 46, + 368, + 82, + 378 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 368, + 82, + 378 + ], + "spans": [ + { + "bbox": [ + 46, + 368, + 82, + 378 + ], + "type": "text", + "content": "smaller." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 46, + 383, + 287, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 383, + 287, + 502 + ], + "spans": [ + { + "bbox": [ + 46, + 383, + 287, + 502 + ], + "type": "text", + "content": "KD Types. In Table 5, we compare the proposed SAB KD with the previous OL2OL KD and TAB KD. For SAB and TAB, the same layers of the student and the teacher are used for KD, the details of which are given in the supplementary materials. From Table 5, we can see that TAB performers better than OL2OL, and our SAB outperforms TAB. In Section 3.3.2, we use masking to speed up the training. In this study, we also compare \"with masking\" and \"without masking\". The last two rows of Table 5 verify that the masking is beneficial." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 46, + 506, + 288, + 650 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 506, + 288, + 650 + ], + "spans": [ + { + "bbox": [ + 46, + 506, + 288, + 650 + ], + "type": "text", + "content": "SAB Property. The student's layers in our SAB KD can be explained as the bases of the feature space, and each layer of the teacher is a linear combination of the bases (Section 3.2). In Fig. 6, we show that this property holds. In SAB KD, the student has 4 layers and the teacher has 12 layers. We randomly pick 3 image examples, and for each example, we select the features of one random token (in the " + }, + { + "bbox": [ + 46, + 506, + 288, + 650 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 506, + 288, + 650 + ], + "type": "text", + "content": " dimension in Fig. 6(a)). After training, the linear combinations " + }, + { + "bbox": [ + 46, + 506, + 288, + 650 + ], + "type": "inline_equation", + "content": "\\tilde{W} \\times \\tilde{L}_{\\text{layer}}^{S} = \\tilde{L}_{\\text{layer}}^{S}" + }, + { + "bbox": [ + 46, + 506, + 288, + 650 + ], + "type": "text", + "content": " show feature patterns very similar to the teacher's features (" + }, + { + "bbox": [ + 46, + 506, + 288, + 650 + ], + "type": "inline_equation", + "content": "\\tilde{L}_{\\text{layer}}^{T}" + }, + { + "bbox": [ + 46, + 506, + 288, + 650 + ], + "type": "text", + "content": "). This property verifies that the teacher's knowledge is fully absorbed by the student." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 46, + 654, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 654, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 654, + 287, + 715 + ], + "type": "text", + "content": "In Fig. 7, we visualize the frame-word alignments with and without SAB KD. Fig. 7(a) is the result of the CLIP in CLIP4clip. Fig. 7(b) is the result of " + }, + { + "bbox": [ + 46, + 654, + 287, + 715 + ], + "type": "inline_equation", + "content": "\\mathrm{CLIPPING}^{-}_{\\mathrm{TAB}}" + }, + { + "bbox": [ + 46, + 654, + 287, + 715 + ], + "type": "text", + "content": ", which is our model without SAB KD but with the previous TAB KD. It can be seen that the frame-word alignment with" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 304, + 367, + 545, + 463 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 367, + 545, + 463 + ], + "spans": [ + { + "bbox": [ + 304, + 367, + 545, + 463 + ], + "type": "text", + "content": "SAB KD (Fig. 7(c)) shows clear and correct word-level attentions (e.g., \"naked\", \"child\", \"runs\" and \"through\") like CLIP. But with TAB KD, the most important words (e.g., \"naked\" and \"child\") cannot be noticed; instead, some insignificant words ([CLS] and \"a\") are activated. This study shows that the knowledge of pre-trained CLIP can be more effectively transferred to the student at the fine-tuning stage with our SAB." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 306, + 480, + 378, + 491 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 480, + 378, + 491 + ], + "spans": [ + { + "bbox": [ + 306, + 480, + 378, + 491 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 304, + 496, + 545, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 496, + 545, + 640 + ], + "spans": [ + { + "bbox": [ + 304, + 496, + 545, + 640 + ], + "type": "text", + "content": "In this paper, we propose a novel and efficient knowledge distillation method that is specially designed for small vision-language models. It includes temporal KD, spatial KD, SAB KD and cross-modality KD. Especially, the SAB KD has the property of the student's layers being the bases of the feature space. After training, the teacher's features are the linear combinations of the bases, indicating that the student has fully absorbed the knowledge of the teacher. CLIPPING significantly outperforms the state-of-the-art and is comparable or even superior to many large pre-training models. In the feature, we will apply CLIPPING to other vision-language models for compression." + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 306, + 656, + 408, + 669 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 656, + 408, + 669 + ], + "spans": [ + { + "bbox": [ + 306, + 656, + 408, + 669 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 305, + 677, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 677, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 305, + 677, + 545, + 712 + ], + "type": "text", + "content": "We gratefully acknowledge the support of MindSpore [1], CANN (Compute Architecture for Neural Networks) and Ascend AI Processor used for this research." + } + ] + } + ], + "index": 32 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18990" + } + ] + } + ], + "index": 33 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 53, + 91, + 177, + 101 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 177, + 101 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 177, + 101 + ], + "type": "text", + "content": "[1] https://www.mindspore.cn/.8" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 102, + 287, + 135 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 102, + 287, + 135 + ], + "spans": [ + { + "bbox": [ + 53, + 102, + 287, + 135 + ], + "type": "text", + "content": "[2] Elad Amrani, Rami Ben-Ari, Daniel Rotman, and Alex Bronstein. Noise estimation using density estimation for self-supervised multimodal learning. In AAAI, 2021. 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 137, + 287, + 168 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 137, + 287, + 168 + ], + "spans": [ + { + "bbox": [ + 53, + 137, + 287, + 168 + ], + "type": "text", + "content": "[3] Max Bain, Arsha Nagrani, Gül Varol, and Andrew Zisserman. Frozen in Time: A Joint Video and Image Encoder for End-to-End Retrieval. In ICCV, 2021. 2, 6, 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 170, + 287, + 191 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 170, + 287, + 191 + ], + "spans": [ + { + "bbox": [ + 53, + 170, + 287, + 191 + ], + "type": "text", + "content": "[4] David Chen and William Dolan. Collecting highly parallel data for paraphrase evaluation. In ACL, 2011. 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 192, + 287, + 224 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 192, + 287, + 224 + ], + "spans": [ + { + "bbox": [ + 53, + 192, + 287, + 224 + ], + "type": "text", + "content": "[5] Defang Chen, JianPing Mei, Yuan Zhang, Can Wang, Zhe Wang, Yan Feng, and Chun Chen. Distillation with semantic calibration. In AAAI, 2021. 2, 3, 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 225, + 286, + 257 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 225, + 286, + 257 + ], + "spans": [ + { + "bbox": [ + 53, + 225, + 286, + 257 + ], + "type": "text", + "content": "[6] Pengguang Chen, Shu Liu, Hengshuang Zhao, and Jiaya Jia. Distilling Knowledge via Knowledge Review. In CVPR, 2021. 2, 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 259, + 287, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 259, + 287, + 300 + ], + "spans": [ + { + "bbox": [ + 53, + 259, + 287, + 300 + ], + "type": "text", + "content": "[7] Xianing Chen, Qiong Cao, Yujie Zhong, Jing Zhang, Shenghua Gao, and Dacheng Tao. DearKD: Data-efficient early knowledge distillation for vision transformers. In ICLR, 2022. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 303, + 287, + 345 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 303, + 287, + 345 + ], + "spans": [ + { + "bbox": [ + 53, + 303, + 287, + 345 + ], + "type": "text", + "content": "[8] Yinpeng Chen, Xiyang Dai, Dongdong Chen, Mengchen Liu, Xiaoyi Dong, Lu Yuan, and Zicheng Liu. Mobile-Former: Bridging MobileNet and Transformer. In CVPR, 2022. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 347, + 287, + 379 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 347, + 287, + 379 + ], + "spans": [ + { + "bbox": [ + 53, + 347, + 287, + 379 + ], + "type": "text", + "content": "[9] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. ImageNet: A large-scale hierarchical image database. In CVPR, 2009. 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 380, + 287, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 380, + 287, + 445 + ], + "spans": [ + { + "bbox": [ + 48, + 380, + 287, + 445 + ], + "type": "text", + "content": "[10] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2021. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 447, + 287, + 489 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 447, + 287, + 489 + ], + "spans": [ + { + "bbox": [ + 48, + 447, + 287, + 489 + ], + "type": "text", + "content": "[11] Maksim Dzabraev, Maksim Kalashnikov, Stepan Komkov, and Aleksandr Petiushko. MDMMT: Multidomain Multimodal Transformer for Video Retrieval. In CVPR, 2021. 2, 6, 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 491, + 287, + 535 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 491, + 287, + 535 + ], + "spans": [ + { + "bbox": [ + 48, + 491, + 287, + 535 + ], + "type": "text", + "content": "[12] Tsu-Jui Fu, Linjie Li, Zhe Gan, Kevin Lin, William Yang Wang, Lijuan Wang, and Zicheng Liu. VIOLET: End-to-End Video-Language Transformers with Masked Token Modeling. In arXiv:2111.1268, 2021. 1, 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 536, + 287, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 536, + 287, + 567 + ], + "spans": [ + { + "bbox": [ + 48, + 536, + 287, + 567 + ], + "type": "text", + "content": "[13] Yuying Ge, Yixiao Ge, Xihui Liu, Dian Li, Ying Shan, Xiaohu Qie, and Ping Luo. Bridgeformer: Bridging video-text retrieval with multiple choice questions. In CVPR, 2022. 6" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 569, + 287, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 569, + 287, + 612 + ], + "spans": [ + { + "bbox": [ + 48, + 569, + 287, + 612 + ], + "type": "text", + "content": "[14] Jianyuan Guo, Kai Han, Han Wu, Yehui Tang, Xinghao Chen, Yunhe Wang, and Chang Xu. CMT: Convolutional Neural Networks Meet Vision Transformers. In CVPR, 2022. 1" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 614, + 287, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 614, + 287, + 645 + ], + "spans": [ + { + "bbox": [ + 48, + 614, + 287, + 645 + ], + "type": "text", + "content": "[15] Geoffrey Hinton, Oriol Vinyals, and Jeff Dean. Distilling the Knowledge in a Neural Network. In arXiv preprint arXiv:1503.02531, 2015. 1, 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 48, + 647, + 287, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 647, + 287, + 690 + ], + "spans": [ + { + "bbox": [ + 48, + 647, + 287, + 690 + ], + "type": "text", + "content": "[16] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In ICML, 2021. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "type": "text", + "content": "[17] Xiaoqi Jiao, Yichun Yin, Lifeng Shang, Xin Jiang, Xiao Chen, Linlin Li, Fang Wang, and Qun Liu. TinyBERT:" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 35, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 94 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 94 + ], + "type": "text", + "content": "Distilling BERT for Natural Language Understanding. In EMNLP, 2020. 1, 2, 5, 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 95, + 545, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 95, + 545, + 117 + ], + "spans": [ + { + "bbox": [ + 308, + 95, + 545, + 117 + ], + "type": "text", + "content": "[18] Yoon Kim and Alexander M. Rush. Sequence-level knowledge distillation. In EMNLP, 2016. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 118, + 545, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 118, + 545, + 160 + ], + "spans": [ + { + "bbox": [ + 308, + 118, + 545, + 160 + ], + "type": "text", + "content": "[19] Pavan Kumar, Anasosalu Vasu, James Gabriel, Jeff Zhu, Oncel Tuzel, and Anurag Ranjan. An Improved One millisecond Mobile Backbone. In arXiv preprint arXiv:2206.04040, 2022. 1, 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 162, + 545, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 162, + 545, + 205 + ], + "spans": [ + { + "bbox": [ + 308, + 162, + 545, + 205 + ], + "type": "text", + "content": "[20] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. BLIP: Bootstrapping Language-Image Pre-training for Unified Vision-Language Understanding and Generation. In arXiv preprint arXiv:2201.12086, 2022. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 206, + 545, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 206, + 545, + 249 + ], + "spans": [ + { + "bbox": [ + 308, + 206, + 545, + 249 + ], + "type": "text", + "content": "[21] Junnan Li, Ramprasaath Selvaraju, Akhilesh Gotmare, Shafiq Joty, Caiming Xiong, and Steven Chu Hong Hoi. Align before fuse: Vision and language representation learning with momentum distillation. In NIPS, 2021. 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 251, + 545, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 251, + 545, + 293 + ], + "spans": [ + { + "bbox": [ + 308, + 251, + 545, + 293 + ], + "type": "text", + "content": "[22] Yanyu Li, Geng Yuan, Yang Wen, Eric Hu, Georgios Evangelidis, Sergey Tulyakov, Yanzhi Wang, and Jian Ren. EfficientFormer: Vision Transformers at MobileNet Speed. In arXiv:2206.01191, 2022. 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 294, + 545, + 336 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 294, + 545, + 336 + ], + "spans": [ + { + "bbox": [ + 308, + 294, + 545, + 336 + ], + "type": "text", + "content": "[23] Sihao Lin, Hongwei Xie, Bing Wang, Kaicheng Yu, Xiaojun Chang, Xiaodan Liang, and Gang Wang. Knowledge Distillation via the Target-aware Transformer. In CVPR, 2022. 2, 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 338, + 545, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 338, + 545, + 370 + ], + "spans": [ + { + "bbox": [ + 308, + 338, + 545, + 370 + ], + "type": "text", + "content": "[24] Yifan Liu, Chunhua Shen, Changqian Yu, and Jingdong Wang. Efficient Semantic Video Segmentation with Per-Frame Inference. In ECCV, 2020. 4" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 372, + 545, + 414 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 372, + 545, + 414 + ], + "spans": [ + { + "bbox": [ + 308, + 372, + 545, + 414 + ], + "type": "text", + "content": "[25] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin Transformer: Hierarchical Vision Transformer using Shifted Windows. In ICCV, 2021. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 416, + 545, + 458 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 416, + 545, + 458 + ], + "spans": [ + { + "bbox": [ + 308, + 416, + 545, + 458 + ], + "type": "text", + "content": "[26] Huaishao Luo, Lei Ji, Ming Zhong, Yang Chen, Wen Lei, Nan Duan, and Tianrui Li. Clip4clip: An empirical study of clip for end to end video clip retrieval. arXiv preprint arXiv:2104.08860, 2021. 1, 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 308, + 460, + 545, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 460, + 545, + 491 + ], + "spans": [ + { + "bbox": [ + 308, + 460, + 545, + 491 + ], + "type": "text", + "content": "[27] Sachin Mehta and Mohammad Rastegari. Separable Self-attention for Mobile Vision Transformers. In ICLR, 2022. 1, 2, 4" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 308, + 493, + 545, + 546 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 493, + 545, + 546 + ], + "spans": [ + { + "bbox": [ + 308, + 493, + 545, + 546 + ], + "type": "text", + "content": "[28] Antoine Miech, Dimitri Zhukov, Jean-Baptiste Alayrac, Makarand Tapaswi, Ivan Laptev, and Josef Sivic. HowTo100M: Learning a Text-Video Embedding by Watching Hundred Million Narrated Video Clips. In ICCV, 2019. 7" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 308, + 548, + 545, + 591 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 548, + 545, + 591 + ], + "spans": [ + { + "bbox": [ + 308, + 548, + 545, + 591 + ], + "type": "text", + "content": "[29] Mandela Patrick, Po-Yao Huang, Yuki Asano, Florian Metze, Alexander Hauptmann, João Henriques, and Andrea Vedaldi. Support-set bottlenecks for video-text representation learning. In ICLR, 2021. 7" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 308, + 593, + 545, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 593, + 545, + 645 + ], + "spans": [ + { + "bbox": [ + 308, + 593, + 545, + 645 + ], + "type": "text", + "content": "[30] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, and Jack Clark. Learning transferable visual models from natural language supervision. ICML, 2021. 1, 2, 7" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 308, + 647, + 545, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 647, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 308, + 647, + 545, + 689 + ], + "type": "text", + "content": "[31] Maithra Raghu, Thomas Unterthiner, Simon Kornblith, Chiyuan Zhang, and Alexey Dosovitskiy. Do vision transformers see like convolutional neural networks? In NIPS, 2021. 2, 3, 5" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 308, + 691, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 691, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 691, + 545, + 713 + ], + "type": "text", + "content": "[32] Anna Rohrbach, Marcus Rohrbach, Niket Tandon, and Bernt Schiele. A dataset for Movie Description. In CVPR, 2015. 6" + } + ] + } + ], + "index": 34 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "18991" + } + ] + } + ], + "index": 36 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 464 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 115 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 115 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 115 + ], + "type": "text", + "content": "[33] Piyush Sharma, Nan Ding, Sebastian Goodman, and Radu Soricut. Conceptual captions: A cleaned, hypernymed, image alt-text dataset for automatic image captioning. In ACL, 2018. 7" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 118, + 287, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 118, + 287, + 172 + ], + "spans": [ + { + "bbox": [ + 48, + 118, + 287, + 172 + ], + "type": "text", + "content": "[34] Alex Jinpeng Wang, Yixiao Ge, Rui Yan, Yuying Ge, Xudong Lin, Guanyu Cai, Jianping Wu, Ying Shan, Xiaohu Qie, and Mike Zheng Shou. All in one: Exploring unified video-language pre-training. In arXiv preprint arXiv:2203.07303, 2022. 2, 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 174, + 287, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 174, + 287, + 217 + ], + "spans": [ + { + "bbox": [ + 48, + 174, + 287, + 217 + ], + "type": "text", + "content": "[35] Jinpeng Wang, Yixiao Ge, Guanyu Cai, Rui Yan, Xudong Lin, Ying Shan, Xiaohu Qie, and Mike Zheng Shou. Object-aware video-language pre-training for retrieval. In CVPR, 2022. 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 219, + 287, + 261 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 219, + 287, + 261 + ], + "spans": [ + { + "bbox": [ + 48, + 219, + 287, + 261 + ], + "type": "text", + "content": "[36] Zekun Wang, Wenhui Wang, Haichao Zhu, Ming Liu, Bing Qin, and Furu Wei. Distilled dual-encoder model for vision-language understanding. In arXiv preprint arXiv:2112.08723, 2021. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 263, + 287, + 317 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 263, + 287, + 317 + ], + "spans": [ + { + "bbox": [ + 48, + 263, + 287, + 317 + ], + "type": "text", + "content": "[37] Hu Xu, Gargi Ghosh, Po-Yao Huang, Dmytro Okhonko, Armen Aghajanyan, Florian Metze, Luke Zettlemoyer, and Christoph Feichtenhofer. Videoclip: Contrastive pre-training for zero-shot video-text understanding. arXiv preprint arXiv:2104.08860, 2021. 2, 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 319, + 287, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 319, + 287, + 352 + ], + "spans": [ + { + "bbox": [ + 48, + 319, + 287, + 352 + ], + "type": "text", + "content": "[38] Jun Xu, Tao Mei, Ting Yao, and Yong Rui. MSR-VTT: A Large Video Description Dataset for Bridging Video and Language. 2016. 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 354, + 287, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 354, + 287, + 385 + ], + "spans": [ + { + "bbox": [ + 48, + 354, + 287, + 385 + ], + "type": "text", + "content": "[39] Jianwei Yang, Yonatan Bisk, and Jianfeng Gao. Taco: Token-aware cascade contrastive learning for video-text alignment. In ICCV, 2021. 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 387, + 287, + 419 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 387, + 287, + 419 + ], + "spans": [ + { + "bbox": [ + 48, + 387, + 287, + 419 + ], + "type": "text", + "content": "[40] Weihao Yu, Mi Luo, Pan Zhou, Chenyang Si, Yichen Zhou, Xinchao Wang, Jiashi Feng, and Shuicheng Yan. Metaformer is actually what you need for vision. In CVPR, 2022. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 421, + 287, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 421, + 287, + 464 + ], + "spans": [ + { + "bbox": [ + 48, + 421, + 287, + 464 + ], + "type": "text", + "content": "[41] Li Yuan, Yunpeng Chen, Tao Wang, Weihao Yu, Yujun Shi, Zi-Hang Jiang, Francis E.H. Tay, Jiashi Feng, and Shuicheng Yan. Tokens-to-Token ViT: Training Vision Transformers From Scratch on ImageNet. In ICCV, 2021. 5" + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "18992" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/CLIPPO_ Image-and-Language Understanding From Pixels Only/6a9a5c21-30d9-4b55-b8d0-e491b4946877_content_list.json b/2023/CLIPPO_ Image-and-Language Understanding From Pixels Only/6a9a5c21-30d9-4b55-b8d0-e491b4946877_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..a9cf0671cc5e6c8a2be0494b4a20d008cc8031cc --- /dev/null +++ b/2023/CLIPPO_ Image-and-Language Understanding From Pixels Only/6a9a5c21-30d9-4b55-b8d0-e491b4946877_content_list.json @@ -0,0 +1,1474 @@ +[ + { + "type": "text", + "text": "CLIPPO: Image-and-Language Understanding from Pixels Only", + "text_level": 1, + "bbox": [ + 158, + 130, + 810, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Michael Tschannen, Basil Mustafa, Neil Houlsby \nGoogle Research, Brain Team, Zürich", + "bbox": [ + 287, + 180, + 679, + 217 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 251, + 313, + 267 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Multimodal models are becoming increasingly effective, in part due to unified components, such as the Transformer architecture. However, multimodal models still often consist of many task- and modality-specific pieces and training procedures. For example, CLIP (Radford et al., 2021) trains independent text and image towers via a contrastive loss. We explore an additional unification: the use of a pure pixel-based model to perform image, text, and multimodal tasks. Our model is trained with contrastive loss alone, so we call it CLIP-Pixels Only (CLIPPO). CLIPPO uses a single encoder that processes both regular images and text rendered as images. CLIPPO performs image-based tasks such as retrieval and zero-shot image classification almost as well as CLIP-style models, with half the number of parameters and no text-specific tower or embedding. When trained jointly via image-text contrastive learning and next-sentence contrastive learning, CLIPPO can perform well on natural language understanding tasks, without any word-level loss (language modelling or masked language modelling), outperforming pixel-based prior work. Surprisingly, CLIPPO can obtain good accuracy in visual question answering, simply by rendering the question and image together. Finally, we exploit the fact that CLIPPO does not require a tokenizer to show that it can achieve strong performance on multilingual multimodal retrieval without modifications.", + "bbox": [ + 76, + 282, + 473, + 660 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 689, + 209, + 704 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In recent years, large-scale multimodal training of Transformer-based models has led to improvements in the state-of-the-art in different domains including vision [2, 10, 74-76], language [6, 11], and audio [5]. In particular, in computer vision and image-language understanding, a single large pretrained model can outperform task-specific expert models [10, 74, 75]. However, large multimodal models often use modality or dataset-specific encoders and decoders, and accordingly lead to involved protocols. For example, such models frequently involve training different", + "bbox": [ + 75, + 714, + 468, + 864 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/ffdab49a0a316c2db9fc0e440b859ee63037c0ec97ed700f60899518134c1a3e.jpg", + "image_caption": [ + "Figure 1. CLIP [56] trains separate image and text encoders, each with a modality-specific preprocessing and embedding, on image/alt-text pairs with a contrastive objective. CLIPPO trains a pure pixel-based model with equivalent capabilities by rendering the alt-text as an image, encoding the resulting image pair using a shared vision encoder (in two separate forward passes), and applying same training objective as CLIP." + ], + "image_footnote": [], + "bbox": [ + 506, + 250, + 694, + 411 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/cbb943b06baad3ca5e62905b787b868708c0afc9f1dc74e14944d5e5690f8131.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 700, + 250, + 885, + 411 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "parts of the model in separate phases on their respective datasets, with dataset-specific preprocessing, or transferring different parts in a task-specific manner [75]. Such modality and task-specific components can lead to additional engineering complexity, and poses challenges when introducing new pretraining losses or downstream tasks. Developing a single end-to-end model that can process any modality, or combination of modalities, would be a valuable step for multimodal learning. Here, we focus on images and text.", + "bbox": [ + 496, + 537, + 890, + 672 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "A number of key unifications have accelerated the progress of multimodal learning. First, the Transformer architecture has been shown to work as a universal backbone, performing well on text [6, 15], vision [16], audio [5, 24, 54], and other domains [7, 34]. Second, many papers have explored mapping different modalities into a single shared embedding space to simplify the input/output interface [21, 22, 46, 69], or develop a single interface to many tasks [31, 37]. Third, alternative representations of modalities allow harnessing in one domain neural architectures or training procedures designed for another domain [28, 49, 54, 60]. For example, [60] and [28, 54] represent text and audio, respectively, by rendering these modalities as images (via a spectogram in the case of audio).", + "bbox": [ + 496, + 672, + 892, + 883 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this paper, we explore the use of a pure pixel-based", + "bbox": [ + 517, + 885, + 890, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "Code and pretrained models are available as part of big_vision [4] https://github.com/google-research/big_vision.", + "bbox": [ + 76, + 875, + 468, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "11006", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "model for multimodal learning of text and images. Our model is a single Vision Transformer [16] that processes visual input, or text, or both together, all rendered as RGB images. The same model parameters are used for all modalities, including low-level feature processing; that is, there are no modality-specific initial convolutions, tokenization algorithms, or input embedding tables. We train our model using only a single task: contrastive learning, as popularized by CLIP [56] and ALIGN [32]. We therefore call our model CLIP-Pixels Only (CLIPPO).", + "bbox": [ + 75, + 90, + 472, + 241 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We find that CLIPPO performs similarly to CLIP-style models (within $1 - 2\\%$ ) on the main tasks CLIP was designed for—image classification and text/image retrieval—despite not having modality-specific towers. Surprisingly, CLIPPO can perform complex language understanding tasks to a decent level without any left-to-right language modelling, masked language modelling, or explicit word-level losses. In particular, on the GLUE benchmark [73] CLIPPO outperforms classic NLP baselines, such as ELMO+BiLSTM+attention, outperforms prior pixel-based masked language models [60], and approaches the score of BERT [15]. Interestingly, CLIPPO obtains good performance on VQA when simply rendering the image and text together, despite never having been pretrained on such data.", + "bbox": [ + 75, + 243, + 472, + 455 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Pixel-based models have an immediate advantage over regular language models because they do not require predetermining the vocabulary/tokenizer and navigating the corresponding intricate trade-offs; consequently, we observe improved performance on multilingual retrieval compared to an equivalent model that uses a classical tokenizer.", + "bbox": [ + 75, + 455, + 472, + 547 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related work", + "text_level": 1, + "bbox": [ + 76, + 566, + 215, + 583 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Multimodal and contrastive pretraining Most closely related to CLIPPO are CLIP [67] and ALIGN [32] which developed the paradigm of large-scale contrastive training on noisy data from the web. Follow-ups [55,85] have scaled further and employed state-of-the-art image representation learning to boost performance.", + "bbox": [ + 75, + 595, + 468, + 686 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "A number of works have explored model unification via weight-sharing. In the contrastive context, LIMoE [53] and MS-CLIP [80] explore a one-tower model similar to ours, studying the use of mixture of experts and selective sharing of modules, respectively. Outside contrastive training, co-training distinct tasks [1, 46] is a popular strategy, with some approaches [44] involving knowledge distillation and gradient masking. Other works use self-supervised learning algorithms to unify task training [21]. These broadly use discriminative tasks to learn representations for various downstream modalities; generative approaches to multimodal modelling have been scaled to billions of parameters, generating text [2, 10, 74, 82], images [58, 62, 83], videos [27, 72] or audio [5] from various modalities.", + "bbox": [ + 75, + 688, + 472, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Another related domain is document and user interface (UI) understanding. Corresponding models are trained on diverse multimodal data sets and can usually solve a range of document/Ui understanding tasks. Many models rely on text extracted using an off-the-shelf OCR pipeline in combination with document images [3, 29], but image-only models are getting more popular [35, 41]. While these models can understand visual cues and text from the input image, they still rely on a tokenized text for training and inference.", + "bbox": [ + 496, + 90, + 890, + 227 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Contrastive training in NLP There is a sizable body of work on contrastive pretraining on sentence pairs (see [59] for a recent survey), which we explore as an auxiliary objective for CLIPPO. Popular augmentations to generate text pairs involve word deletion, span deletion, reordering, synonym substitution, and next-sentence-prediction [20, 47, 77]. Other methods use different realizations of dropout masks in the model to emulate sentence pairs, or supervised labels to obtain positive and negative pairs [19].", + "bbox": [ + 496, + 229, + 890, + 368 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Visual text and tokenization in NLP The most closely related method to CLIPPO from the NLP domain is PILIX [60], which is a masked autoencoder (MAE) [26] trained on rendered text. It obtains strong performance on multilingual syntactic (part-of-speech tagging, dependency parsing) and semantic language understanding (named entity recognition, sentence understanding) tasks, while being more robust to noise in the text than BERT. Other applications for which visual text has been explored include sentiment analysis [68] and machine translation [49, 63].", + "bbox": [ + 496, + 369, + 890, + 522 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Visual text side-steps the design and construction of an appropriate tokenizer, which is a large area of research of its own, and can hence simplify text processing in certain—in particular multilingual—scenarios. We refer to [52] for a survey on tokenizers. Popular models include WordPiece [15], Byte-Pair Encoding [65], and SentencePiece [39].", + "bbox": [ + 496, + 523, + 890, + 613 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Subword-based vocabularies are popular in monolingual setups and usually lead to a good performance trade-off compared to word and character based vocabularies for certain languages including English. In multilingual contexts, appropriately representing the vocabulary of all languages becomes challenging as the number of languages increases [13,61], which in turn can lead to poor performance in tasks involving underrepresented languages. A variety of mitigation strategies has been developed; we refer to [60, Sec. 5.1] for a more detailed discussion of these strategies.", + "bbox": [ + 496, + 614, + 890, + 765 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3. Contrastive language-image pretraining with pixels", + "text_level": 1, + "bbox": [ + 498, + 781, + 890, + 816 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Contrastive language-image pretraining has emerged as a powerful, scalable paradigm to train versatile vision models on web-scale data sets [56]. Concretely, this approach relies on image/alt-text pairs which can be automatically collected at large scale from the web. Thereby, the textual", + "bbox": [ + 496, + 825, + 890, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "11007", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "descriptions are usually noisy, and can e.g. consist of single keywords, sets of keywords, or potentially lengthy descriptions with many attributes describing the image content. Using this data, two encoders are jointly trained, namely a text encoder embedding the alt-texts and an image encoder embedding the corresponding images into a shared latent space. These two encoders are trained with a contrastive loss, encouraging the embeddings of matching images and alt-text to be similar, and at the same time to be dissimilar from all other image and alt-text embeddings.", + "bbox": [ + 75, + 90, + 472, + 241 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Once trained, such an encoder pair can be used in many ways: It can be specialized to classifying a fixed set of visual concepts via their textual descriptions (zero-shot classification); the embeddings can be used to retrieve images given a textual description and vice-versa; or the vision encoder can be transferred in supervised fashion to a downstream task by fine-tuning on a labeled data set or by training a head on top of the frozen image encoder representation. In principle, the text encoder can be used as a standalone text embedding, but this application—to our knowledge—has not been explored in-depth, with some authors citing the low quality of the alt-texts leading to weak language modeling performance of the text encoder [67].", + "bbox": [ + 75, + 244, + 472, + 441 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Previous works [46, 53] have shown that the image and text encoder can be realized with a single shared transformer model (henceforth referred to as single tower model, or 1T-CLIP), where the images are embedded using a patch embedding, and the tokenized text is embedded using a separate word embedding. Apart from the modality-specific embeddings, all model parameters are shared for the two modalities. While this type of sharing usually leads to a minor performance drop on image/image-language tasks it also halves the number of model parameters.", + "bbox": [ + 75, + 441, + 472, + 594 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "CLIPPO takes this idea one step further: text inputs are rendered on blank images, and are subsequently dealt with entirely as images, including the initial patch embedding (see Fig. 1 for an illustration). By training this single vision transformer contrastively as prior works, we obtain a single vision transformer model that can understand both images and text through the single interface of vision and provides a single representation which can be used to solve image, image-language, and pure language understanding tasks.", + "bbox": [ + 75, + 594, + 472, + 731 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Alongside multimodal versatility, CLIPPO alleviates common hurdles with text processing, namely the development of an appropriate tokenizer and vocabulary. This is particularly interesting in a massively multilingual setup, where the text encoder has to handle dozens of languages.", + "bbox": [ + 75, + 732, + 472, + 808 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We find that CLIPPO trained on image/alt-text pairs performs comparably with its 1T-CLIP counterpart on common image and image-language benchmarks, and is competitive with strong baseline language models on the GLUE benchmark [73]. However, due to the low quality of the alt-texts which are often not grammatical sentences, learn", + "bbox": [ + 75, + 809, + 472, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "ing language understanding exclusively from alt-texts is fundamentally limited. Therefore, we augment image/alt-text contrastive pretraining with language-based contrastive training. Specifically, we use positive pairs of consecutive sentences sampled from a text corpus which is seamlessly integrated into the contrastive training by supplementing batches of image/alt-texts with (rendered) text/text pairs.", + "bbox": [ + 496, + 90, + 893, + 198 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 498, + 210, + 633, + 227 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4.1. Training details and models", + "text_level": 1, + "bbox": [ + 498, + 234, + 750, + 251 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We rely on a single training setup for all our baselines and visual text models. This setup was tuned to produce good results for standard image/alt-text contrastive training as in [56] (using exactly the same loss function as [56], following the pseudocode in [56, Fig. 3]) and we found that it readily transfers to 1T-CLIP and CLIPPO (including variants with text/text co-training).", + "bbox": [ + 496, + 257, + 893, + 364 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our default architecture is a ViT-B/16 [16] and we perform a subset of experiments with a ViT-L/16 architecture to study the effect of scale (we equip both models a MAP head [40] to pool embeddings). In all cases, the representation dimension used for the contrastive loss is 768. We set the batch size to 10,240 and train the main models for 250k steps, using a minimum 100k training steps for ablations. For models co-trained with a certain percentage of text/text data, we scale the number of iterations such that the number of image/alt-text pairs seen matches the number of iterations of the corresponding model without text/text data (e.g. when $50\\%$ of the data is text/text pairs we increase the number of iterations from 250k to 500k). The contrastive loss is computed across the full batch. We use the Adafactor optimizer [66] with a learning rate of $10^{-3}$ and decoupled weight decay with weight $10^{-4}$ .", + "bbox": [ + 496, + 364, + 895, + 604 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Baseline CLIP-style models are trained using the T5-en SentencePiece tokenizer [57]; we use the abbreviation $\\mathrm{CLIP^{*}}$ for the two tower model from [56] trained from scratch using the setup described above, to avoid confusion with the model released by [56]. A sequence length of 196 is used, as this matches the number of visual text \"tokens\" CLIPPO can process with patch size 16 has at $224\\mathrm{px}$ resolution (which we use throughout unless noted otherwise).", + "bbox": [ + 496, + 604, + 893, + 727 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Visual text For visual text rendering [60, 63] relied on the Google Noto font family1 which supports the majority of Unicode code points. Here, we use the GNU Unifont bitmap font2, which has a similar coverage but allows for efficient, lookup-based on-the-fly rendering in our preprocessing pipeline. We emphasize that this rendering strategy does not slow down training compared to tokenizer-based models. In preliminary explorations, we found this to be performance-neutral when compared to the Noto font.", + "bbox": [ + 496, + 729, + 893, + 867 + ], + "page_idx": 2 + }, + { + "type": "page_footnote", + "text": "1https://fonts.google.com/nota", + "bbox": [ + 514, + 875, + 756, + 886 + ], + "page_idx": 2 + }, + { + "type": "page_footnote", + "text": "2http://unifoundry.com/unifont", + "bbox": [ + 517, + 887, + 754, + 898 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "11008", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 2 + }, + { + "type": "table", + "img_path": "images/23b2d87e1dacca821161ce70e2cb1c7a8632b4ec9ba6e242046e914e76fa7ca7.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
#param.training datasetI1k 10s.I1k 0s.C I→TC T→IF I→TF T→I
CLIP*203MWebLI55.865.148.531.379.259.4
1T-CLIP118MWebLI53.962.348.030.377.558.2
CLIPPO93MWebLI53.061.447.330.176.457.3
CLIPPO93MWebLI + 25%C452.157.440.726.768.951.8
CLIPPO93MWebLI + 50%C448.053.135.223.464.847.2
1T-CLIP L/16349MWebLI60.867.850.732.581.061.0
CLIPPO L/16316MWebLI60.367.450.633.479.262.6
CLIPPO L/16316MWebLI + 25%C460.566.044.529.872.957.3
CLIPPO L/16316MWebLI + 50%C456.861.739.727.370.154.7
", + "bbox": [ + 174, + 88, + 787, + 237 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Table 1. Vision and vision-language cross-modal results. We report ImageNet-1k 10-shot linear transfer validation accuracy (I1k 10s.), ImageNet-1k zero-shot transfer validation accuracy (I1k 0s.), image-to-text and text-to-image retrieval recall@1 on MS-COCO (C I→T and C T→I) and on Flickr30k (F T→I and F I→T). CLIPPO and 1T-CLIP incur a minor drop in these evaluations compared to CLIP*, while only using about half of the model parameters. Co-training with text pairs from C4 (models with + xx%C4) degrades performance on some cross-modal tasks (but leads to improved language understanding capabilities, see Table 2).", + "bbox": [ + 75, + 244, + 890, + 313 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Image/alt-text data We use the WebLI data set introduced in [10] which comprises 10 billion images with 12 billion corresponding alt-texts. Importantly, WebLI comprises alt-texts in 109 languages (unlike previous data sets such as LAION-400M [64] which only contain English alt-texts) and it is therefore a great foundation to study multilingual language-image pretraining and its applications. Please refer to [10, Fig. 3] for details on the alt-text language distribution. For English-only models we obtain English versions of non-English alt-texts via GCP Translation $\\mathrm{API}^3$ . In addition to alt-text, WebLI also provides OCR annotations, which we do not use in this paper. Finally, WebLI was processed with a de-duplication step removing all images from various splits of the image evaluation sets used in this paper. Please refer to [10, Sec. 3.2] for more details on the WebLI data set and to [10, Appendix B] for a datasheet.", + "bbox": [ + 75, + 325, + 470, + 566 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We also present a subset of results based on LAION-400M [64] and YFCC-100M [71] as an additional comparison points, see Appendix C.1 and C.2, respectively.", + "bbox": [ + 76, + 568, + 468, + 613 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Text/text data For co-training with text/text pairs we primarily rely on the publicly available Colossal Clean Crawled Corpus (C4; default/English split) [57]. We randomly sample pairs of consecutive sentences and contrastively train on these pairs, i.e., the model is trained for embedding-based next sentence prediction (NSP) [47]. We also experiment with pairs of parallel sentences in different languages from the WMT19 data set [18] as well as back-translated English sentences derived from C4 following the strategy described in [12].", + "bbox": [ + 75, + 617, + 468, + 768 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.2. Evaluations and metrics", + "text_level": 1, + "bbox": [ + 76, + 777, + 299, + 792 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To evaluate the vision and vision/language understanding capabilities of our models we use standard metrics from the literature [53, 56, 85]: \"zero-shot\" transfer, which uses (embedded) textual description of the classes to be classified/retrieved and compares these with image embeddings.", + "bbox": [ + 76, + 801, + 468, + 878 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We report the classification accuracy on ImageNet-1k [14] as well as the recall@1 for cross-modal retrieval on MSCOCO [9] and Flickr30k [81]. Furthermore, we test the low-data transfer performance of the models by means of the linear adaptation protocol from [16], reporting the 10-shot accuracy on ImageNet-1k.", + "bbox": [ + 504, + 325, + 890, + 416 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We also evaluate CLIPPO and baselines on the popular VQA benchmark VQAv2 [25]. To construct a VQA model using a single pretrained ViT we render the question at the top end of the corresponding image (using the same Unifont renderer as used for CLIPPO training) and follow the standard prediction setup where the answer is predicted as the most likely answer from the training set, i.e. by classification. Specifically, we replace the last layer of our pretrained CLIPPO and baselines with a randomly initialized one with the appropriate number of outputs and fine-tune on VQAv2. This setup tests the ability of the pretrained ViT to combine image and text in intermediate layers as it has produce a single output from a fused image/text input image, unlike in the other cross-modal tasks (and pretraining), where image and text representations are computed with two separate forward passes. Please refer to Appendix A in the supplementary material for examples images with rendered questions and Appendix B.1 for details on the fine-tuning protocol.", + "bbox": [ + 496, + 417, + 892, + 688 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Multilingual capabilities are assessed via zero-shot retrieval on CrossModal3600 [70], which is a geographically diverse set comprising 3600 images each human-annotated with captions in 36 languages. The corresponding recall metric is averaged across all languages and images.", + "bbox": [ + 496, + 689, + 890, + 763 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Finally, we evaluate the language understanding capabilities on the General Language Understanding Evaluation (GLUE) benchmark [73] which comprises natural language inference tasks (MNLI, QNLI, RTE), a sentiment analysis task (SST-2), sentence similarity tasks (QQP, STS-B, MRPC), and a linguistic acceptability task (CoLA). Following common practice, we exclude the WNLI task from the benchmark [15, 77]. We transfer our baselines and CLIPPO models by attaching a 2-hidden layer MLP with 768 units to", + "bbox": [ + 496, + 763, + 892, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_footnote", + "text": "3https://cloud.google.com/translate", + "bbox": [ + 93, + 886, + 370, + 898 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "11009", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/9b8eb1248493590b3de5a1a00f10c2f59eb466390defefe448d40e116c1814e1.jpg", + "image_caption": [ + "Figure 2. Results on the VQAv2 benchmark (test-dev set). In addition to CLIPPO and baselines produced in this work, we also compare to Pythia and MCAN models with ViT encoders from [67], and with comparably sized METER [17] and ViLT [36] models. CLIPPO outperforms $\\mathrm{CLIP^{*}}$ and 1T-CLIP clearly on \"yes/no\" questions and gets similar performance as task-specific models." + ], + "image_footnote": [], + "bbox": [ + 86, + 90, + 467, + 396 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "their representation and following precisely the fine-tuning protocol from BERT [15]. For sentence pair classification tasks we simply render both sentences on the same image, printing [SEP] to mark the start of the second sentence.", + "bbox": [ + 75, + 523, + 468, + 585 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.3. Vision and vision-language understanding", + "text_level": 1, + "bbox": [ + 76, + 598, + 434, + 614 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Image classification and retrieval Table 1 shows the performance of CLIPPO along with the baseline models on the benchmarks described in Sec. 4.2. It can be seen that the CLIPPO and 1T-CLIP incur a drop of a 2-3 percentage points absolute compared to $\\mathrm{CLIP^{*}}$ . This is not surprising and can be attributed to the fact that single tower models only have about half the parameters count of a corresponding two tower model. The difference in performance between the English-only CLIPPO and 1T-CLIP is very small for a B/16 backbone at $100\\mathrm{k}$ training steps (see Table 6 in the supplementary material), and vanishes with longer training and/or by increasing the model size, despite the fact that CLIPPO has $25\\%$ and $10\\%$ fewer parameters than 1T-CLIP for a B/16 and L/16 architecture, respectively (which is due to the absence of the text embedding in CLIPPO).", + "bbox": [ + 75, + 627, + 468, + 852 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The multilingual CLIPPO model performs somewhat worse than the corresponding 1T-CLIP, and the gap does not close completely when training longer (see Table 6).", + "bbox": [ + 76, + 854, + 468, + 902 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "However, when evaluated across a broad set of languages on the CrossModal3600 CLIPPO performs on par with or slightly better than 1T-CLIP (see Sec. 4.4 below).", + "bbox": [ + 498, + 90, + 890, + 136 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As we add sentence pairs to the training mix the performance on the cross-modal retrieval metrics decreases. This is not surprising as we keep the total batch size constant so that the effective batch size of image/alt-text contrastive training decreases, which is known to impact performance [85]. Interestingly, the 10-shot transfer performance does not move in tandem, but only decreases significantly when half of the training data is sentence pairs. In exchange, co-training with text data leads to significantly improved language understanding performance (see Sec. 4.5).", + "bbox": [ + 496, + 138, + 890, + 289 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "VQA In Fig. 2 we report the VQAv2 score of our models and baselines. It can be seen that CLIPPO outperforms $\\mathrm{CLIP^{*}}$ , 1T-CLIP, as well as a pretrained ViT-B/16 from [16] by a significant margin, achieving a score of 66.3, and constraining with $25\\%$ C4 data leads to a slight improvement of the score. The improved score of CLIPPO is mainly due to better performance in \"yes/no\" questions. Increasing the model size to L/16 adds another 2 points which originate from improvements in the \"number\" and \"other\" VQAv2 categories. However, note that for an L/16 architecture 1T-CLIP performs competitively with CLIPPO (see Table 7). One possible explanation for this could be that 1T-CLIP develops better OCR capabilities thanks to the higher model capacity (alt-texts can correlate with text in images/scene text, see [10, Fig. 3]). Increasing the resolution to 384px adds 2 to 3 points across models.", + "bbox": [ + 496, + 294, + 892, + 536 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We also compare CLIPPO with baselines from the literature. Specifically, [17] proposes a framework (called ME-TER) for multimodal tasks, where pretrained transformer-based image and text encoders are combined with a transformer-based fusion module. CLIPPO L/16 achieves performance competitive with their model combining a CLIP B/32 vision backbone with a BERT-Base language backbone, which is roughly comparable in size and computational cost with our L/16 models. Another related work is [67], which combines different CLIP vision backbones with two existing VQA systems, Pythia [33] and MCAN [84]. CLIPPO outperforms different CLIP ViT-based Pythia and MCAN models from [67]. Note, however, that ResNet-based CLIP backbones lead to better results when combined with these systems. We further note that both [17] and [67] also investigate training their models on a mix of different image-text data sets with multiple objectives such as grounded masked language modeling and text-image matching, before transferring to the VQA task, which leads to significant improvements. ViLT [36] relies on such a strategy to train a single transformer backbone jointly encoding image and text tokens. At 384px resolution, CLIPPO (with $25\\%$ C4 data) obtains a VQA score comparable with that of ViLT (and other models from the", + "bbox": [ + 496, + 537, + 892, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "11010", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "literature such as ViLBERT [48], VisualBERT [43], and PixelBERT [30]), despite only using a contrastive objective for pretraining.", + "bbox": [ + 75, + 90, + 468, + 137 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.4. Multilingual vision-language understanding", + "text_level": 1, + "bbox": [ + 76, + 147, + 449, + 165 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For typical language models, tokenizer choice can be a challenging process [78]. Commonly used English-language tokenizers generalize poorly to non-latin scripts [85]. This can be alleviated by the use of larger, multilingual vocabularies, at the expense of very large parameter counts. CLIPPO bypasses this issue, removing any language-related bias stemming from unbalanced or restrictive tokenizers. We consider multilingual image/text retrieval on Crossmodal3600 and compare CLIPPO, trained on WebLI with multilingual alt-texts, against 1T-CLIP with a number of SentencePiece tokenizers; one trained from 300M WebLI multilingual alt-texts, English (T5-en) and multilingual (T5-all) tokenizers from T5 [57], and a multilingual tokenizer (mT5) from mT5 [79], all with a vocabulary size of 32,000. The results are shown in Fig. 4. On average, CLIPPO achieves comparable retrieval performance to these baselines. In the case of mT5, the use of extra data to create the specialized vocabulary can boosts performance above that of CLIPPO; the leveraging of such extra parameters and data in the multilingual context will be an interesting future direction for CLIPPO.", + "bbox": [ + 75, + 171, + 472, + 488 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/c530045a6fa2378a022f1da0ae18ebf659dbe4382f0e14b00f1ed0bba3350633.jpg", + "image_caption": [ + "Figure 3. Tokenization efficiency analyzed in terms of the sequence length produced by a given method. CLIPPO produces smaller sequences for the majority of languages compared to 1T-CLIP with alternative tokenizers." + ], + "image_footnote": [], + "bbox": [ + 81, + 498, + 267, + 619 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Tokenization efficiency If a tokenizer is well suited to a particular dataset, it will tokenize to shorter sequences—this is especially the case when byte fallback [39] is enabled. SentencePiece tokenizers have the advantageous ability to tokenize entire—possibly quite long—words to single tokens. CLIPPO cannot learn any such compression, but benefits from equal treatment of all languages and", + "bbox": [ + 297, + 492, + 468, + 763 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "words: it will by definition generalize equally well to all data, as its tokenization schema has not been trained on a specific dataset. We analyze 20,000 samples for each of the 104 C4 languages. Each CLIPPO token is assumed to be a $16 \\times 16$ patch; though in typical computations all approaches considered here would pad to a fixed length, we compute CLIPPO's sequence length according to the last patch which contains rendered text. Fig. 3 shows the fraction of C4 languages where CLIPPO processes tokens", + "bbox": [ + 75, + 763, + 472, + 902 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/aff7e6aca6b658507c2c2b72c506fc8c84f425347c964bc30c587143bb11692f.jpg", + "image_caption": [ + "Figure 4. Zero-shot image/text retrieval performance on CrossModal3600 [70]. Although specialized (mc4) tokenizers can be leveraged to improve multilingual performance CLIPPO (dashed black line) broadly matches or exceeds comparable 1T-CLIP models trained with vocabulary size 32,000 (the word embeddings result in a $27\\%$ increase in parameter count compared to CLIPPO)." + ], + "image_footnote": [], + "bbox": [ + 504, + 89, + 632, + 181 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/7517525aa45ff1bdaec77d872d46a3f728890ca88b6860747cd07f33d286697b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 632, + 90, + 756, + 181 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/ca91c6e84a119e7ebe86b0050fec9439ddfee621ba860c69c656b9b675a62c4e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 758, + 90, + 888, + 180 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "more efficiently than the vocabularies discussed above. We conservatively define \"more efficient\" as producing a shorter token sequence for over $75\\%$ of examples. Even so, CLIPPO is indeed more efficient across the majority of languages. Per-language breakdowns of multilingual retrieval performance and tokenization efficiency are further discussed in Appendix C.3.", + "bbox": [ + 496, + 289, + 890, + 396 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.5. Language understanding", + "text_level": 1, + "bbox": [ + 500, + 407, + 728, + 422 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 2 shows the GLUE benchmark results of CLIPPO and baselines. One can observe that CLIPPO trained on WebLI performs competitively with the BiLSTM+Attn+ELMo baseline which relies on deep word embeddings trained on a large language corpus. Also, it can be seen that CLIPPO along with 1T-CLIP outperform the language encoder trained using standard contrastive language vision pretraining $(\\mathrm{CLIP}^{*})$ . This indicates that multimodal training in a single encoder benefits language understanding. Furthermore, CLIPPO achieves a much higher GLUE score than the $\\mathrm{CLIP}^*$ image encoder, which in turn leads to significantly better results than fine-tuning a ViT-B/16 from scratch on GLUE (see Appendix C.2 for additional results). Unsurprisingly, the models pretrained on WebLI cannot do better than random guessing on the CoLA evaluation which requires to assess the grammatical correctness of sentences (recall that alt-texts are rarely grammatical sentences). Also the accuracy of $\\mathrm{CLIP}^*$ and 1T-CLIP vision encoders we observe for SST-2 is in agreement with what was reported in [56, Table 10] for CLIP with a ViT-B/16 image encoder.", + "bbox": [ + 496, + 431, + 890, + 733 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Adding sentence pairs form the C4 corpus gradually improves the GLUE score, and when half of the examples are sentence pairs our model becomes competitive with PILXEL, while still retaining decent image and vision-language understanding capabilities (cf. Table 1). Note, however, that there is a trade-off between language-only tasks and tasks that involve image understanding. Finally, training CLIPPO only on sentence pairs leads to a model which outperforms PILXEL by a significant margin. However, our model has seen more sentence pairs than PILXEL, so PILXEL might improve as well when training longer.", + "bbox": [ + 496, + 734, + 890, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "11011", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/9853585471a5f44fb437f10d85e68ef027f5bbf01b43e6e573e0d9739b05fc13.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
training datasetMNLI-M/MMQQPQNLISST-2COLASTS-BMRPCRTEavg
BERT-BaseWiki + BC84.0 / 84.287.691.092.660.388.890.269.583.1
PIXELWiki + BC78.1 / 78.984.587.889.638.481.188.260.576.3
BiLSTM66.7 / 66.782.077.087.517.672.085.158.568.1
BiLSTM+Attn, ELMo72.4 / 72.483.675.291.544.156.182.152.770.0
CLIP* img enc.WebLI66.4 / 67.578.669.478.60.05.281.252.755.5
CLIP* text enc.WebLI71.8 / 72.582.773.086.26.665.081.453.865.9
1T-CLIP text enc.WebLI72.6 / 73.083.880.784.90.079.683.357.068.3
CLIPPOWebLI73.0 / 72.684.381.286.81.880.584.153.468.6
CLIPPOWebLI + 25%C477.7 / 77.285.383.190.928.283.484.559.274.4
CLIPPOWebLI + 50%C479.2 / 79.286.484.292.938.983.484.859.976.6
CLIPPOC479.9 / 80.286.785.293.350.984.786.358.578.4
CLIPPO L/16WebLI + 25%C476.6 / 75.587.179.993.248.284.184.656.076.1
CLIPPO L/16WebLI + 50%C482.3 / 82.487.986.794.255.385.885.959.280.0
", + "bbox": [ + 107, + 88, + 854, + 291 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 2. Results for the GLUE benchmark (dev set). The metric is accuracy except for the performance on QQP and MRPC, which is measured using the $F_{1}$ score, CoLA which uses Matthew's correlation, and STS-B which evaluated based on Spearman's correlation coefficient. \"avg\" corresponds to the average across all metrics. The results for BERT-Base and PILXEL are from [60, Table 3], and BiLSTM and BiLSTM+Attn, ELMo from [73, Table 6]. All encoders considered here have a transformer architecture comparable to BERT-Base (up to the text embedding layer), except for CLIPPO L/16 which uses a ViT L/16, and the two BiLSTM model variants. Wiki and BC stand for (English) Wikipedia and Bookcorpus [86] data, respectively.", + "bbox": [ + 75, + 294, + 893, + 380 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.6. Ablations and analysis", + "text_level": 1, + "bbox": [ + 76, + 390, + 285, + 406 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Impact of weight sharing across modalities The fact that CLIPPO 1) uses a shared patch embedding for regular images and text images and 2) this embedding has considerably fewer parameters than the text embedding of 1T-CLIP and $\\mathrm{CLIP^{*}}$ provokes the question of whether CLIPPO could benefit from separate patch embeddings for text images and regular images. Further, CLIPPO relies on a single head to compute the output representation for images and text, and relaxing this constraint by using separate heads for the two modalities could lead to more expressive representations. The results (deferred to Appendix D.1) show that neither of these variants lead to improved image classification or retrieval metrics compared to CLIPPO.", + "bbox": [ + 75, + 417, + 470, + 614 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Impact of the text location We test whether rendering the question at the top, middle, or bottom of the image impacts the VQA performance of CLIPPO and find that it does not, provided that we increase the learning rate of the positional embedding during fine-tuning (see Appendix D.2).", + "bbox": [ + 75, + 618, + 468, + 696 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Typographic attacks Since CLIPPO is trained on large amounts of rendered (alt-)text it is important to check whether it becomes more susceptible to typographic attacks—the tendency of CLIP-style models to zero-shot classify an image according to adversarially injected scene text unrelated to the scene [23,42,50]. In Appendix D.3 we present results indicating that CLIPPO is no more vulnerable to typographic attacks than 1T-CLIP and $\\mathrm{CLIP^{*}}$ .", + "bbox": [ + 75, + 699, + 468, + 820 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Modality gap Liang et al. [45] discovered that text and image embeddings of CLIP-style models form two distinct clusters rather than both filling the embedding space densely and occupying the same spatial region. They attribute this phenomenon to a combination of initialization", + "bbox": [ + 75, + 824, + 470, + 902 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "conditions and properties of the loss function/training dynamics. Since we consider single tower models here, and also co-train some of these models with text-only pairs it is interesting to see how this affects the modality gap. We compute the gap and visualize it following the recipe from [45] in Fig. 5 (see Appendices D.4 and D.5 for additional visualizations). CLIPPO attains a slightly lower modality gap than $\\mathrm{CLIP}^*$ , but clearly features a clustering structure for image and text embeddings. However, when training contrastively with sentence pairs in addition to image/alt-text pairs, the clustering structure disappears, the image and text embeddings overlap, and the modality gap decreases significantly. A possible explanation for this behavior could be that the additional learning pressure induced by the contrastive loss on sentence pairs encourages text embeddings to spread out more and hence the structure of all embeddings changes.", + "bbox": [ + 496, + 390, + 890, + 648 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Text/text co-training objectives To corroborate that contrastive NSP is a sensible objective to improve language understanding in the context of CLIPPO, we train CLIPPO without any image/alt-text data on pairs of parallel translated sentences (this is straight-forward in our framework since visual text is language-agnostic), as well as English back-translated data, and evaluate the resulting text representations on GLUE. Table 3 shows that NSP on C4 clearly achieves the highest GLUE score.", + "bbox": [ + 496, + 650, + 890, + 787 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5. Discussion and limitations", + "text_level": 1, + "bbox": [ + 498, + 799, + 743, + 816 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We proposed and evaluated CLIPPO which produces a single ViT that can understand images and language jointly using images as a sole input modality. Perhaps surprisingly, CLIPPO matches the performance of the 1T-CLIP baseline across many of the considered tasks, and only in", + "bbox": [ + 496, + 824, + 890, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "11012", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/4eb700574c3530918cc0b2dd483b0b1148395dec2156bcc6c3c84f45af2984be.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
WMT19WMT19 BTC4 NSP
GLUE score61.266.677.6
", + "bbox": [ + 122, + 88, + 421, + 128 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 3. Ablation of text pair-based contrastive co-training tasks: Training on parallel translated sentences (WMT19), training on parallel back-translated sentences (WMT19 BT), and NSP for sentences sampled from C4 (C4 NSP). C4 NSP leads to the highest GLUE score by a large margin.", + "bbox": [ + 75, + 138, + 468, + 209 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "curs a minor drop compared to the $\\mathrm{CLIP^{*}}$ baseline, despite having less than half the parameters of $\\mathrm{CLIP^{*}}$ . As we showed, the image-only interface enables a simple, unified data pipeline for training on and transferring to mixed modalities. CLIPPO opens the door for additional modalities (e.g. spectrograms) and, as we hope, might inspire applications of pixel-only models beyond contrastive training. Nevertheless, several limitations remain, as discussed next.", + "bbox": [ + 75, + 237, + 467, + 357 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Co-training First, to achieve language understanding performance competitive with PILXEL and BERT on GLUE, contrastive co-training with text pairs is necessary. While adding $25\\%$ C4 data to the batch seems to strike a good balance across all tasks considered, it does induce a nonnegligible drop in zero-shot image classification and image/text retrieval. This drop becomes more severe as the fraction of C4 examples increases. We observed an associated change in modality gap, and further investigation of the representation in the context of co-training might help to develop models that achieve better overall performance in the co-training setup.", + "bbox": [ + 75, + 363, + 467, + 542 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Diverse rendered text CLIPPO currently relies on cleanly rendered text as an input and its capabilities to handle text from documents or web pages without further adaption is limited (besides the basic OCR capabilities that CLIP-style models learn from image/alt-text pairs). We emphasize that sophisticated OCR and document understanding is not a goal of this paper, and training CLIPPO with augmented noisy rendered text that mimics the distribution of documents and websites is likely to lead to worse performance across the considered tasks, since image/alt-text pairs are less correlated and provide a weaker learning signal. However, developing CLIPPO further to handle less clean visual text will open many additional applications.", + "bbox": [ + 75, + 547, + 467, + 744 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Generative modeling CLIPPO, like CLIP, BERT, PILXEL and many other models, uses an encoder-only design and hence lacks the ability to generate text outputs. A common approach to equip encoder-only models with generation capabilities (e.g., for image captioning or VQA) is to simply combine them with a (potentially pretrained) language model [8, 76]. This approach naturally also applies to CLIPPO and PILXEL, but defeats the advantages of visual text in certain (e.g. multilingual) scenarios. While visual text outputs have previously been explored in the context of", + "bbox": [ + 75, + 750, + 467, + 900 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/58db6c461caeb2536f7513e0b50b180c5916ddaba4a1c4a68466f01496090d94.jpg", + "image_caption": [ + "Figure 5. Visualization of the modality gap for $\\mathrm{CLIP^{*}}$ and CLIPPO optionally trained with $25\\%$ C4 data. The visualization follows the analysis from [45] and shows embedded images (blue dots) and corresponding alt-text (orange dots) from the WebLI validation set, projected to the first two principal components of the validation data matrix. CLIPPO has a slightly smaller modality gap than $\\mathrm{CLIP^{*}}$ ; co-training with C4 data strongly reduces the gap." + ], + "image_footnote": [], + "bbox": [ + 506, + 83, + 887, + 196 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "machine translation [49], it remains unclear what a scalable tokenizer-free way to generate text is.", + "bbox": [ + 498, + 324, + 890, + 353 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Multilingual learning Finally, we showed that CLIPPO obtains strong multilingual image/text retrieval performance without requiring the development of an appropriate tokenizer. For fine-grained adjustment and balancing of the retrieval performance further steps will be necessary, including data balancing and potentially co-training with multi-lingual text data. Furthermore, similar to PILXEL, CLIPPO relies on certain ad-hoc design choices w.r.t. the visual representation, for example the left-to-right rendering of Arabic scripts. This approach leads to decent performance on average, but it is not clear what kind of unwanted effects it introduces and how these could be mitigated.", + "bbox": [ + 498, + 358, + 890, + 539 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 553, + 617, + 569 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We introduced CLIPPO, a model for processing images and text jointly through the lens of vision. This reduces design choices (in particular w.r.t. tokenization) and parameter count, simplifies data processing pipelines and transfer recipes, and increases generality across multiple languages. We also explored methods of enhancing language understanding, where traditional image/alt-text contrastive models trained on web data fall short. We demonstrated this is possible by co-training with text pairs, with CLIPPO models outperforming strong NLP baselines while maintaining solid image understanding capabilities.", + "bbox": [ + 496, + 579, + 890, + 744 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Although we presented a unified contrastive training algorithm, CLIPPO suffers somewhat when co-training on multiple tasks, and future work to harmonize the co-training could enhance the models significantly. Deeper understanding of the design choices in rendering text as images, and their impact on performance, is another interesting avenue.", + "bbox": [ + 496, + 746, + 890, + 835 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments We would like to thank Lucas Beyer, Josip Djolonga, Alexander Kolesnikov, Mario Lucic, Andreas Steiner, Xiao Wang, and Xiaohua Zhai for inspiring discussions and helpful feedback. We also thank Jeffrey Sorensen for help with the text rendering preprocessing.", + "bbox": [ + 500, + 840, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "11013", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 174, + 106 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Hassan Akbari, Liangzhe Yuan, Rui Qian, Wei-Hong Chuang, Shih-Fu Chang, Yin Cui, and Boqing Gong. VATT: transformers for multimodal self-supervised learning from raw video, audio and text. In NeurIPS, 2021. 2", + "[2] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katie Millican, Malcolm Reynolds, Roman Ring, Eliza Rutherford, Serkan Cabi, Tengda Han, Zhitao Gong, Sina Samangooei, Marianne Monteiro, Jacob Menick, Sebastian Borgeaud, Andrew Brock, Aida Nematzadeh, Sahand Sharifzadeh, Mikolaj Binkowski, Ricardo Barreira, Oriol Vinyals, Andrew Zisserman, and Karen Simonyan. Flamingo: A visual language model for few-shot learning. In NeurIPS, 2022. 1, 2", + "[3] Srikar Appalaraju, Bhavan Jasani, Bhargava Urala Kota, Yusheng Xie, and R. Manmatha. DocFormer: End-to-end transformer for document understanding. In ICCV, pages 973-983, 2021. 2", + "[4] Lucas Beyer, Xiaohua Zhai, and Alexander Kolesnikov. Big Vision. https://github.com/google-research/big visions, 2022.1", + "[5] Zalan Borsos, Raphael Marinier, Damien Vincent, Eugene Kharitonov, Olivier Pietquin, Matthew Sharifi, Olivier Teboul, David Grangier, Marco Tagliasacchi, and Neil Zeghidour. AudioLM: a language modeling approach to audio generation. CoRR, abs/2209.03143, 2022. 1, 2", + "[6] Tom B. Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert-Voss, Gretchen Krueger, Tom Henighan, et al. Language models are few-shot learners. In NeurIPS, 2020. 1", + "[7] Lili Chen, Kevin Lu, Aravind Rajeswaran, Kimin Lee, Aditya Grover, Michael Laskin, Pieter Abbeel, Aravind Srinivas, and Igor Mordatch. Decision transformer: Reinforcement learning via sequence modeling. In NeurIPS, pages 15084-15097, 2021. 1", + "[8] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey E. Hinton. A simple framework for contrastive learning of visual representations. In ICML, 2020. 8", + "[9] Xinlei Chen, Hao Fang, Tsung-Yi Lin, Ramakrishna Vedantam, Saurabh Gupta, Piotr Dólar, and C. Lawrence Zitnick. Microsoft COCO captions: Data collection and evaluation server. CoRR, abs/1504.00325, 2015. 4", + "[10] Xi Chen, Xiao Wang, Soravit Changpinyo, A. J. Piergiovanni, Piotr Padlewski, Daniel Salz, Sebastian Goodman, Adam Grycner, Basil Mustafa, Lucas Beyer, Alexander Kolesnikov, Joan Puigcerver, Nan Ding, Keran Rong, Hassan Akbari, Gaurav Mishra, Linting Xue, Ashish Thapliyal, James Bradbury, Weicheng Kuo, Mojtaba Seyedhosseini, Chao Jia, Burcu Karagol Ayan, Carlos Riquelme, Andreas Steiner, Anelia Angelova, Xiaohua Zhai, Neil Houlsby, and Radu Soricut. PaLI: A jointly-scaled multilingual language-image model. In ICLR, 2023. 1, 2, 4, 5", + "[11] Aakanksha Chowdhery, Sharan Narang, Jacob Devlin, Maarten Bosma, Gaurav Mishra, Adam Roberts, Paul" + ], + "bbox": [ + 78, + 114, + 470, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Barham, Hyung Won Chung, Charles Sutton, Sebastian Gehrmann, Parker Schuh, Kensen Shi, Sasha Tsvyashchenko, Joshua Maynez, Abhishek Rao, Parker Barnes, Yi Tay, Noam Shazeer, Vinodkumar Prabhakaran, Emily Reif, Nan Du, Ben Hutchinson, Reiner Pope, James Bradbury, Jacob Austin, Michael Isard, Guy Gur-Ari, Pengcheng Yin, Toju Duke, Anselm Levskaya, Sanjay Ghemawat, Sunipa Dev, Henryk Michalewski, Xavier Garcia, Vedant Misra, Kevin Robinson, Liam Fedus, Denny Zhou, Daphne Ippolito, David Luan, Hyeontaek Lim, Barret Zoph, Alexander Spiridonov, Ryan Sepassi, David Dohan, Shivani Agrawal, Mark Omernick, Andrew M. Dai, Thanumalayan Sankaranarayana Pillai, Marie Pellat, Aitor Lewkowycz, Erica Moreira, Rewon Child, Oleksandr Polozov, Katherine Lee, Zongwei Zhou, Xuezhi Wang, Brennan Saeta, Mark Diaz, Orhan Firat, Michele Catasta, Jason Wei, Kathy Meier-Hellstern, Douglas Eck, Jeff Dean, Slav Petrov, and Noah Fiedel. PaLM: Scaling language modeling with pathways. CoRR, abs/2204.02311, 2022. 1", + "[12] Geoffrey Cideron, Sertan Girgin, Anton Raichuk, Olivier Pietquin, Olivier Bachem, and Léonard Hussenot. vec2text with round-trip translations. CoRR, abs/2209.06792, 2022. 4", + "[13] Alexis Conneau, Kartikay Khandelwal, Naman Goyal, Vishrav Chaudhary, Guillaume Wenzek, Francisco Guzmán, Edouard Grave, Myle Ott, Luke Zettlemoyer, and Veselin Stoyanov. Unsupervised cross-lingual representation learning at scale. In ACL, pages 8440–8451, 2020. 2", + "[14] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In CVPR, 2009. 4", + "[15] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. BERT: pre-training of deep bidirectional transformers for language understanding. In NAACL-HLT, 2019, 1, 2, 4, 5", + "[16] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth $16 \\times 16$ words: Transformers for image recognition at scale. In ICLR, 2021. 1, 2, 3, 4, 5, 14, 16, 17, 24", + "[17] Zi-Yi Dou, Yichong Xu, Zhe Gan, Jianfeng Wang, Shuohang Wang, Lijuan Wang, Chenguang Zhu, Pengchuan Zhang, Lu Yuan, Nanyun Peng, Zicheng Liu, and Michael Zeng. An empirical study of training end-to-end vision-and-language transformers. In CVPR, pages 18145-18155, 2022. 5, 14, 17", + "[18] Wikipedia Foundation. ACL 2019 Fourth Conference on Machine Translation (WMT19), Shared Task: Machine Translation of News. http://www.statmt.org/wmt19/translation-task.html.4", + "[19] Tianyu Gao, Xingcheng Yao, and Danqi Chen. Simcse: Simple contrastive learning of sentence embeddings. In EMNLP, pages 6894-6910, 2021. 2", + "[20] John M. Giorgi, Osvald Nitski, Bo Wang, and Gary D. Bader. DeCLUTR: Deep contrastive learning for unsupervised textual representations. In ACL/IJCNLP, pages 879-895, 2021. 2" + ], + "bbox": [ + 501, + 92, + 890, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "11014", + "bbox": [ + 480, + 945, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[21] Rohit Girdhar, Alaaeldin El-Nouby, Mannat Singh, Kalyan Vasudev Alwala, Armand Joulin, and Ishan Misra. OmniMAE: Single model masked pretraining on images and videos. CoRR, abs/2206.08356, 2022. 1, 2", + "[22] Rohit Girdhar, Mannat Singh, Nikhila Ravi, Laurens van der Maaten, Armand Joulin, and Ishan Misra. Omnivore: A single model for many visual modalities. In CVPR, pages 16081-16091, 2022. 1", + "[23] Gabriel Goh, Nick Cammarata, Chelsea Voss, Shan Carter, Michael Petrov, Ludwig Schubert, Alec Radford, and Chris Olah. Multimodal neurons in artificial neural networks. *Distill*, 2021, 7, 21", + "[24] Yuan Gong, Yu-An Chung, and James R. Glass. AST: audio spectrogram transformer. In Interspeech, pages 571-575, 2021. 1", + "[25] Yash Goyal, Tejas Khot, Douglas Summers-Stay, Dhruv Bartra, and Devi Parikh. Making the V in VQA matter: Elevating the role of image understanding in visual question answering. In CVPR, 2017. 4, 13", + "[26] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollár, and Ross B. Girshick. Masked autoencoders are scalable vision learners. In CVPR, pages 15979-15988, 2022. 2", + "[27] Jonathan Ho, William Chan, Chitwan Saharia, Jay Whang, Ruiqi Gao, Alexey A. Gritsenko, Diederik P. Kingma, Ben Poole, Mohammad Norouzi, David J. Fleet, and Tim Salimans. Imagen video: High definition video generation with diffusion models. CoRR, abs/2210.02303, 2022. 2", + "[28] Po-Yao Huang, Hu Xu, Juncheng Li, Alexei Baevski, Michael Auli, Wojciech Galuba, Florian Metze, and Christoph Feichtenhofer. Masked autoencoders that listen. CoRR, abs/2207.06405, 2022. 1", + "[29] Yupan Huang, Tengchao Lv, Lei Cui, Yutong Lu, and Furu Wei. Layoutlmv3: Pre-training for document AI with unified text and image masking. In ACMMM, pages 4083-4091, 2022. 2", + "[30] Zhicheng Huang, Zhaoyang Zeng, Bei Liu, Dongmei Fu, and Jianlong Fu. Pixel-BERT: Aligning image pixels with text by deep multi-modal transformers. CoRR, abs/2004.00849, 2020.6", + "[31] Andrew Jaegle, Sebastian Borgeaud, Jean-Baptiste Alayrac, Carl Doersch, Catalin Ionescu, David Ding, Skanda Koppula, Daniel Zoran, Andrew Brock, Evan Shelhamer, Olivier J. Henaff, Matthew M. Botvinick, Andrew Zisserman, Oriol Vinyals, and João Carreira. Perceiver IO: A general architecture for structured inputs & outputs. In ICLR, 2022. 1", + "[32] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc V. Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In ICML, 2021. 2", + "[33] Yu Jiang, Vivek Natarajan, Xinlei Chen, Marcus Rohrbach, Dhruv Batra, and Devi Parikh. Pythia v0.1: The winning entry to the VQA challenge 2018. CoRR, abs/1807.09956, 2018. 5", + "[34] John Jumper, Richard Evans, Alexander Pritzel, Tim Green, Michael Figurnov, Olaf Ronneberger, Kathryn Tunyasuvunakool, Russ Bates, Augustin Žídek, Anna Potapenko, Alex" + ], + "bbox": [ + 78, + 90, + 468, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Bridgland, Clemens Meyer, Simon A A Kohl, Andrew J Ballard, Andrew Cowie, Bernardino Romera-Paredes, Stanislav Nikolov, Rishub Jain, Jonas Adler, Trevor Back, Stig Petersen, David Reiman, Ellen Clancy, Michal Zielinski, Martin Steinegger, Michalina Pacholska, Tamas Berghammer, Sebastian Bodenstein, David Silver, Oriol Vinyals, Andrew W Senior, Koray Kavukcuoglu, Pushmeet Kohli, and Demis Hassabis. Highly accurate protein structure prediction with AlphaFold. Nature, 596(7873):583-589, 2021. 1", + "[35] Geewook Kim, Teakgyu Hong, Moonbin Yim, JeongYeon Nam, Jinyoung Park, Jinyeong Yim, Wonseok Hwang, Sang-doo Yun, Dongyoon Han, and Seunghyun Park. OCR-Free document understanding transformer. In ECCV, pages 498- 517,2022.2", + "[36] Wonjae Kim, Bokyung Son, and Ildoo Kim. ViLT: Vision-and-language transformer without convolution or region supervision. In ICML, pages 5583-5594. PMLR, 2021. 5, 17", + "[37] Alexander Kolesnikov, André Susano Pinto, Lucas Beyer, Xiaohua Zhai, Jeremiah Harmsen, and Neil Houlsby. UViM: A unified modeling approach for vision with learned guiding codes. In NeurIPS, 2022. 1", + "[38] Simon Kornblith, Mohammad Norouzi, Honglak Lee, and Geoffrey Hinton. Similarity of neural network representations revisited. In ICML, pages 3519-3529, 2019. 22", + "[39] Taku Kudo and John Richardson. SentencePiece: A simple and language independent subword tokenizer and tokenizer for neural text processing. In EMNLP, 2018. 2, 6", + "[40] Juho Lee, Yoonho Lee, Jungtaek Kim, Adam R. Kosiorek, Seungjin Choi, and Yee Whye Teh. Set transformer: A framework for attention-based permutation-invariant neural networks. In ICML, pages 3744-3753, 2019. 3, 14", + "[41] Kenton Lee, Mandar Joshi, Iulia Turc, Hexiang Hu, Fangyu Liu, Julian Eisenschlos, Urvashi Khandelwal, Peter Shaw, Ming-Wei Chang, and Kristina Toutanova. Pix2struct: Screenshot parsing as pretraining for visual language understanding. CoRR, abs/2210.03347, 2022. 2", + "[42] Yoann Lemesle, Masataka Sawayama, Guillermo Valle-Perez, Maxime Adolphe, Hélène Sauzeon, and Pierre-Yves Oudeyer. Language-biased image classification: evaluation based on semantic representations. In ICLR, 2022. 7, 21", + "[43] Liunian Harold Li, Mark Yatskar, Da Yin, Cho-Jui Hsieh, and Kai-Wei Chang. VisualBERT: A simple and performant baseline for vision and language. CoRR, abs/1908.03557, 2019. 6", + "[44] Qing Li, Boqing Gong, Yin Cui, Dan Kondratyuk, Xianzhi Du, Ming-Hsuan Yang, and Matthew Brown. Towards a unified foundation model: Jointly pre-training transformers on unpaired images and text. CoRR, abs/2112.07074, 2021. 2", + "[45] Weixin Liang, Yuhui Zhang, Yongchan Kwon, Serena Yeung, and James Zou. Mind the gap: Understanding the modality gap in multi-modal contrastive representation learning. In NeurIPS, 2022. 7, 8, 22, 23", + "[46] Valerii Likhosherstov, Anurag Arnab, Krzysztof Choromanski, Mario Lucic, Yi Tay, Adrian Weller, and Mostafa Dehghani. Polyvit: Co-training vision transformers on images, videos and audio. Trans. Machine Learning Research, 2023. 1, 2, 3" + ], + "bbox": [ + 503, + 92, + 890, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "11015", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[47] Lajanugen Logeswaran and Honglak Lee. An efficient framework for learning sentence representations. In ICLR, 2018. 2, 4", + "[48] Jiasen Lu, Dhruv Batra, Devi Parikh, and Stefan Lee. ViLBERT: Pretraining task-agnostic visiolinguistic representations for vision-and-language tasks. In NeurIPS, pages 13–23, 2019. 6", + "[49] Elman Mansimov, Mitchell Stern, Mia Xu Chen, Orhan Firat, Jakob Uszkoreit, and Puneet Jain. Towards end-to-end in-image neural machine translation. CoRR, abs/2010.10648, 2020. 1, 2, 8", + "[50] Joanna Materzyńska, Antonio Torralba, and David Bau. Disentangling visual and written concepts in CLIP. In CVPR, pages 16410-16419, 2022. 7, 21", + "[51] Joanna Materzynska, Antonio Torralba, and David Bau. Disentangling visual and written concepts in clip. In CVPR, 2022. 21, 22", + "[52] Sabrina J. Mielke, Zaid Alyafeai, Elizabeth Salesky, Colin Raffel, Manan Dey, Matthias Galle, Arun Raja, Chenglei Si, Wilson Y. Lee, Benoit Sagot, and Samson Tan. Between words and characters: A brief history of open-vocabulary modeling and tokenization in NLP. CoRR, abs/2112.10508, 2021. 2", + "[53] Basil Mustafa, Carlos Riquelme, Joan Puigcerver, Rodolphe Jenatton, and Neil Houlsby. Multimodal contrastive learning with LIMoE: the language-image mixture of experts. In NeurIPS, 2022. 2, 3, 4", + "[54] Daisuke Niizumi, Daiki Takeuchi, Yasunori Ohishi, Noboru Harada, and Kunio Kashino. Masked spectrogram modeling using masked autoencoders for learning general-purpose audio representation. CoRR, abs/2204.12260, 2022. 1", + "[55] Hieu Pham, Zihang Dai, Golnaz Ghiasi, Kenji Kawaguchi, Hanxiao Liu, Adams Wei Yu, Jiahui Yu, Yi-Ting Chen, Minh-Thang Luong, Yonghui Wu, Mingxing Tan, and Quoc V. Le. Combined scaling for open-vocabulary image classification. CoRR, abs/2111.10050, 2021. 2", + "[56] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In ICML, 2021. 1, 2, 3, 4, 6, 14, 17", + "[57] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J. Liu. Exploring the limits of transfer learning with a unified text-to-text transformer. J. Mach. Learn. Res., 21:140:1-140:67, 2020. 3, 4, 6, 20", + "[58] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In ICML, pages 8821-8831, 2021. 2", + "[59] Nils Rethmeier and Isabelle Augenstein. A primer on contrastive pretraining in language processing: Methods, lessons learned and perspectives. ACM Computing Surveys, 2021. 2", + "[60] Phillip Rust, Jonas F. Lotz, Emanuele Bugliarello, Elizabeth Salesky, Miryam de Lhoneux, and Desmond Elliott. Language modelling with pixels. In ICLR, 2023. 1, 2, 3, 7, 18" + ], + "bbox": [ + 78, + 90, + 468, + 900 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[61] Phillip Rust, Jonas Pfeiffer, Ivan Vulic, Sebastian Ruder, and Iryna Gurevych. How good is your tokenizer? on the monolingual performance of multilingual language models. In ACL/IJCNLP, pages 3118-3135, 2021. 2", + "[62] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily Denton, Seyed Kamyar Seyed Ghasemipour, Burcu Karagol Ayan, S. Sara Mahdavi, Rapha Gontijo Lopes, Tim Salimans, Jonathan Ho, David J. Fleet, and Mohammad Norouzi. Photorealistic text-to-image diffusion models with deep language understanding. In NeurIPS, 2022. 2", + "[63] Elizabeth Salesky, David Etter, and Matt Post. Robust open-vocabulary translation from visual text representations. In EMNLP, pages 7235-7252, 2021. 2, 3", + "[64] Christoph Schuhmann, Richard Vencu, Romain Beaumont, Robert Kaczmarczyk, Clayton Mullis, Aarush Katta, Theo Coombes, Jenia Jitsev, and Aran Komatsuzaki. LAION-400M: Open dataset of CLIP-filtered 400 million image-text pairs. CoRR, abs/2111.02114, 2021. 4, 15", + "[65] Rico Sennrich, Barry Haddow, and Alexandra Birch. Neural machine translation of rare words with subword units. In ACL, 2016. 2", + "[66] Noam Shazeer and Mitchell Stern. Adafactor: Adaptive learning rates with sublinear memory cost. In ICML, 2018. 3, 14", + "[67] Sheng Shen, Liunian Harold Li, Hao Tan, Mohit Bansal, Anna Rohrbach, Kai-Wei Chang, Zhewei Yao, and Kurt Keutzer. How much can CLIP benefit vision-and-language tasks? In ICLR, 2022. 2, 3, 5, 17", + "[68] Baohua Sun, Lin Yang, Catherine Chi, Wenhan Zhang, and Michael Lin. Squared english word: A method of generating glyph to use super characters for sentiment analysis. In Workshop on Affective Content Analysis, volume 2328, pages 140-151, 2019. 2", + "[69] Zineng Tang, Jaemin Cho, Yixin Nie, and Mohit Bansal. TVLT: Textless vision-language transformer. In NeurIPS, 2022. 1", + "[70] Ashish Thapliyal, Jordi Pont-Tuset, Xi Chen, and Radu Soricut. Crossmodal-3600: A Massively Multilingual Multimodal Evaluation Dataset. In EMNLP, 2022. 4, 6, 19", + "[71] Bart Thomee, David A. Shamma, Gerald Friedland, Benjamin Elizalde, Karl Ni, Douglas Poland, Damian Borth, and Li-Jia Li. YFCC100M: the new data in multimedia research. Commun. ACM, 59(2):64-73, 2016. 4, 14", + "[72] Ruben Villegas, Mohammad Babaeizadeh, Pieter-Jan Kindermans, Hernan Moraldo, Han Zhang, Mohammad Taghi Saffar, Santiago Castro, Julius Kunze, and Dumitru Erhan. Phenaki: Variable length video generation from open domain textual description. In ICLR, 2023. 2", + "[73] Alex Wang, Amanpreet Singh, Julian Michael, Felix Hill, Omer Levy, and Samuel R. Bowman. GLUE: A multi-task benchmark and analysis platform for natural language understanding. In ICLR, 2019. 2, 3, 4, 7, 18", + "[74] Jianfeng Wang, Zhengyuan Yang, Xiaowei Hu, Linjie Li, Kevin Lin, Zhe Gan, Zicheng Liu, Ce Liu, and Lijuan Wang. GIT: A generative image-to-text transformer for vision and language. Trans. Machine Learning Research, 2022. 1, 2" + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11016", + "bbox": [ + 480, + 945, + 519, + 955 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[75] Wenhui Wang, Hangbo Bao, Li Dong, Johan Bjorck, Zhiliang Peng, Qiang Liu, Kriti Aggarwal, Owais Khan Mohammed, Saksham Singhal, Subhojit Som, and Furu Wei. Image as a foreign language: BEiT pretraining for all vision and vision-language tasks. CoRR, abs/2208.10442, 2022. 1", + "[76] Zirui Wang, Jiahui Yu, Adams Wei Yu, Zihang Dai, Yulia Tsvetkov, and Yuan Cao. SimVLM: Simple visual language model pretraining with weak supervision. In ICLR, 2022. 1, 8", + "[77] Zhuofeng Wu, Sinong Wang, Jiatao Gu, Madian Khabsa, Fei Sun, and Hao Ma. CLEAR: contrastive learning for sentence representation. CoRR, abs/2012.15466, 2020. 2, 4", + "[78] Linting Xue, Aditya Barua, Noah Constant, Rami Al-Rfou, Sharan Narang, Mihir Kale, Adam Roberts, and Colin Raffel. Byt5: Towards a token-free future with pre-trained byte-to-byte models. Trans. Assoc. Comput. Linguistics, 10:291-306, 2022. 6", + "[79] Linting Xue, Noah Constant, Adam Roberts, Mihir Kale, Rami Al-Rfou, Aditya Siddhant, Aditya Barua, and Colin Raffel. mT5: A massively multilingual pre-trained text-to-text transformer. In NAACL-HLT, 2021. 6, 20", + "[80] Haoxuan You, Luowei Zhou, Bin Xiao, Noel Codella, Yu Cheng, Ruochen Xu, Shih-Fu Chang, and Lu Yuan. Learning visual representation from modality-shared contrastive language-image pre-training. In ECCV, pages 69-87, 2022. 2", + "[81] Peter Young, Alice Lai, Micah Hodosh, and Julia Hockenmaier. From image descriptions to visual denotations: New similarity metrics for semantic inference over event descriptions. Trans. Assoc. Comput. Linguistics, 2:67-78, 2014. 4", + "[82] Jiahui Yu, Zirui Wang, Vijay Vasudevan, Legg Yeung, Mojtaba Seyedhosseini, and Yonghui Wu. Coca: Contrastive captioners are image-text foundation models. Trans. Machine Learning Research, 2022. 2", + "[83] Jiahui Yu, Yuanzhong Xu, Jing Yu Koh, Thang Luong, Gunjan Baid, Zirui Wang, Vijay Vasudevan, Alexander Ku, Yinfei Yang, Burcu Karagol Ayan, Ben Hutchinson, Wei Han, Zarana Parekh, Xin Li, Han Zhang, Jason Baldridge, and Yonghui Wu. Scaling autoregressive models for content-rich text-to-image generation. Trans. Machine Learning Research, 2022. 2", + "[84] Zhou Yu, Jun Yu, Yuhao Cui, Dacheng Tao, and Qi Tian. Deep modular co-attention networks for visual question answering. In CVPR, pages 6281–6290, 2019. 5", + "[85] Xiaohua Zhai, Xiao Wang, Basil Mustafa, Andreas Steiner, Daniel Keysers, Alexander Kolesnikov, and Lucas Beyer. LiT: Zero-shot transfer with locked-image text tuning. In CVPR, pages 18102-18112, 2022. 2, 4, 5, 6, 14", + "[86] Yukun Zhu, Ryan Kiros, Richard S. Zemel, Ruslan Salakhutdinov, Raquel Urtasun, Antonio Torralba, and Sanja Fidler. Aligning books and movies: Towards story-like visual explanations by watching movies and reading books. In ICCV, pages 19-27, 2015. 7, 18" + ], + "bbox": [ + 78, + 90, + 467, + 839 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "11017", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 11 + } +] \ No newline at end of file diff --git a/2023/CLIPPO_ Image-and-Language Understanding From Pixels Only/6a9a5c21-30d9-4b55-b8d0-e491b4946877_model.json b/2023/CLIPPO_ Image-and-Language Understanding From Pixels Only/6a9a5c21-30d9-4b55-b8d0-e491b4946877_model.json new file mode 100644 index 0000000000000000000000000000000000000000..9aa4eb4c45eac1f96726954a36a0f1ec462d5227 --- /dev/null +++ b/2023/CLIPPO_ Image-and-Language Understanding From Pixels Only/6a9a5c21-30d9-4b55-b8d0-e491b4946877_model.json @@ -0,0 +1,2380 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.159, + 0.131, + 0.812, + 0.154 + ], + "angle": 0, + "content": "CLIPPO: Image-and-Language Understanding from Pixels Only" + }, + { + "type": "text", + "bbox": [ + 0.289, + 0.181, + 0.681, + 0.218 + ], + "angle": 0, + "content": "Michael Tschannen, Basil Mustafa, Neil Houlsby \nGoogle Research, Brain Team, Zürich" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.252, + 0.314, + 0.268 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.283, + 0.474, + 0.661 + ], + "angle": 0, + "content": "Multimodal models are becoming increasingly effective, in part due to unified components, such as the Transformer architecture. However, multimodal models still often consist of many task- and modality-specific pieces and training procedures. For example, CLIP (Radford et al., 2021) trains independent text and image towers via a contrastive loss. We explore an additional unification: the use of a pure pixel-based model to perform image, text, and multimodal tasks. Our model is trained with contrastive loss alone, so we call it CLIP-Pixels Only (CLIPPO). CLIPPO uses a single encoder that processes both regular images and text rendered as images. CLIPPO performs image-based tasks such as retrieval and zero-shot image classification almost as well as CLIP-style models, with half the number of parameters and no text-specific tower or embedding. When trained jointly via image-text contrastive learning and next-sentence contrastive learning, CLIPPO can perform well on natural language understanding tasks, without any word-level loss (language modelling or masked language modelling), outperforming pixel-based prior work. Surprisingly, CLIPPO can obtain good accuracy in visual question answering, simply by rendering the question and image together. Finally, we exploit the fact that CLIPPO does not require a tokenizer to show that it can achieve strong performance on multilingual multimodal retrieval without modifications." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.69, + 0.21, + 0.705 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.715, + 0.47, + 0.866 + ], + "angle": 0, + "content": "In recent years, large-scale multimodal training of Transformer-based models has led to improvements in the state-of-the-art in different domains including vision [2, 10, 74-76], language [6, 11], and audio [5]. In particular, in computer vision and image-language understanding, a single large pretrained model can outperform task-specific expert models [10, 74, 75]. However, large multimodal models often use modality or dataset-specific encoders and decoders, and accordingly lead to involved protocols. For example, such models frequently involve training different" + }, + { + "type": "image", + "bbox": [ + 0.508, + 0.251, + 0.695, + 0.412 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.702, + 0.251, + 0.887, + 0.412 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.422, + 0.895, + 0.52 + ], + "angle": 0, + "content": "Figure 1. CLIP [56] trains separate image and text encoders, each with a modality-specific preprocessing and embedding, on image/alt-text pairs with a contrastive objective. CLIPPO trains a pure pixel-based model with equivalent capabilities by rendering the alt-text as an image, encoding the resulting image pair using a shared vision encoder (in two separate forward passes), and applying same training objective as CLIP." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.538, + 0.892, + 0.673 + ], + "angle": 0, + "content": "parts of the model in separate phases on their respective datasets, with dataset-specific preprocessing, or transferring different parts in a task-specific manner [75]. Such modality and task-specific components can lead to additional engineering complexity, and poses challenges when introducing new pretraining losses or downstream tasks. Developing a single end-to-end model that can process any modality, or combination of modalities, would be a valuable step for multimodal learning. Here, we focus on images and text." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.674, + 0.893, + 0.885 + ], + "angle": 0, + "content": "A number of key unifications have accelerated the progress of multimodal learning. First, the Transformer architecture has been shown to work as a universal backbone, performing well on text [6, 15], vision [16], audio [5, 24, 54], and other domains [7, 34]. Second, many papers have explored mapping different modalities into a single shared embedding space to simplify the input/output interface [21, 22, 46, 69], or develop a single interface to many tasks [31, 37]. Third, alternative representations of modalities allow harnessing in one domain neural architectures or training procedures designed for another domain [28, 49, 54, 60]. For example, [60] and [28, 54] represent text and audio, respectively, by rendering these modalities as images (via a spectogram in the case of audio)." + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.886, + 0.892, + 0.902 + ], + "angle": 0, + "content": "In this paper, we explore the use of a pure pixel-based" + }, + { + "type": "page_footnote", + "bbox": [ + 0.077, + 0.875, + 0.47, + 0.901 + ], + "angle": 0, + "content": "Code and pretrained models are available as part of big_vision [4] https://github.com/google-research/big_vision." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "11006" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.242 + ], + "angle": 0, + "content": "model for multimodal learning of text and images. Our model is a single Vision Transformer [16] that processes visual input, or text, or both together, all rendered as RGB images. The same model parameters are used for all modalities, including low-level feature processing; that is, there are no modality-specific initial convolutions, tokenization algorithms, or input embedding tables. We train our model using only a single task: contrastive learning, as popularized by CLIP [56] and ALIGN [32]. We therefore call our model CLIP-Pixels Only (CLIPPO)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.244, + 0.473, + 0.456 + ], + "angle": 0, + "content": "We find that CLIPPO performs similarly to CLIP-style models (within \\(1 - 2\\%\\)) on the main tasks CLIP was designed for—image classification and text/image retrieval—despite not having modality-specific towers. Surprisingly, CLIPPO can perform complex language understanding tasks to a decent level without any left-to-right language modelling, masked language modelling, or explicit word-level losses. In particular, on the GLUE benchmark [73] CLIPPO outperforms classic NLP baselines, such as ELMO+BiLSTM+attention, outperforms prior pixel-based masked language models [60], and approaches the score of BERT [15]. Interestingly, CLIPPO obtains good performance on VQA when simply rendering the image and text together, despite never having been pretrained on such data." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.457, + 0.473, + 0.549 + ], + "angle": 0, + "content": "Pixel-based models have an immediate advantage over regular language models because they do not require predetermining the vocabulary/tokenizer and navigating the corresponding intricate trade-offs; consequently, we observe improved performance on multilingual retrieval compared to an equivalent model that uses a classical tokenizer." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.567, + 0.216, + 0.584 + ], + "angle": 0, + "content": "2. Related work" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.597, + 0.47, + 0.688 + ], + "angle": 0, + "content": "Multimodal and contrastive pretraining Most closely related to CLIPPO are CLIP [67] and ALIGN [32] which developed the paradigm of large-scale contrastive training on noisy data from the web. Follow-ups [55,85] have scaled further and employed state-of-the-art image representation learning to boost performance." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.689, + 0.473, + 0.903 + ], + "angle": 0, + "content": "A number of works have explored model unification via weight-sharing. In the contrastive context, LIMoE [53] and MS-CLIP [80] explore a one-tower model similar to ours, studying the use of mixture of experts and selective sharing of modules, respectively. Outside contrastive training, co-training distinct tasks [1, 46] is a popular strategy, with some approaches [44] involving knowledge distillation and gradient masking. Other works use self-supervised learning algorithms to unify task training [21]. These broadly use discriminative tasks to learn representations for various downstream modalities; generative approaches to multimodal modelling have been scaled to billions of parameters, generating text [2, 10, 74, 82], images [58, 62, 83], videos [27, 72] or audio [5] from various modalities." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.228 + ], + "angle": 0, + "content": "Another related domain is document and user interface (UI) understanding. Corresponding models are trained on diverse multimodal data sets and can usually solve a range of document/Ui understanding tasks. Many models rely on text extracted using an off-the-shelf OCR pipeline in combination with document images [3, 29], but image-only models are getting more popular [35, 41]. While these models can understand visual cues and text from the input image, they still rely on a tokenized text for training and inference." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.231, + 0.892, + 0.369 + ], + "angle": 0, + "content": "Contrastive training in NLP There is a sizable body of work on contrastive pretraining on sentence pairs (see [59] for a recent survey), which we explore as an auxiliary objective for CLIPPO. Popular augmentations to generate text pairs involve word deletion, span deletion, reordering, synonym substitution, and next-sentence-prediction [20, 47, 77]. Other methods use different realizations of dropout masks in the model to emulate sentence pairs, or supervised labels to obtain positive and negative pairs [19]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.371, + 0.892, + 0.523 + ], + "angle": 0, + "content": "Visual text and tokenization in NLP The most closely related method to CLIPPO from the NLP domain is PILIX [60], which is a masked autoencoder (MAE) [26] trained on rendered text. It obtains strong performance on multilingual syntactic (part-of-speech tagging, dependency parsing) and semantic language understanding (named entity recognition, sentence understanding) tasks, while being more robust to noise in the text than BERT. Other applications for which visual text has been explored include sentiment analysis [68] and machine translation [49, 63]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.524, + 0.892, + 0.614 + ], + "angle": 0, + "content": "Visual text side-steps the design and construction of an appropriate tokenizer, which is a large area of research of its own, and can hence simplify text processing in certain—in particular multilingual—scenarios. We refer to [52] for a survey on tokenizers. Popular models include WordPiece [15], Byte-Pair Encoding [65], and SentencePiece [39]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.615, + 0.892, + 0.766 + ], + "angle": 0, + "content": "Subword-based vocabularies are popular in monolingual setups and usually lead to a good performance trade-off compared to word and character based vocabularies for certain languages including English. In multilingual contexts, appropriately representing the vocabulary of all languages becomes challenging as the number of languages increases [13,61], which in turn can lead to poor performance in tasks involving underrepresented languages. A variety of mitigation strategies has been developed; we refer to [60, Sec. 5.1] for a more detailed discussion of these strategies." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.782, + 0.892, + 0.817 + ], + "angle": 0, + "content": "3. Contrastive language-image pretraining with pixels" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.826, + 0.892, + 0.903 + ], + "angle": 0, + "content": "Contrastive language-image pretraining has emerged as a powerful, scalable paradigm to train versatile vision models on web-scale data sets [56]. Concretely, this approach relies on image/alt-text pairs which can be automatically collected at large scale from the web. Thereby, the textual" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "11007" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.242 + ], + "angle": 0, + "content": "descriptions are usually noisy, and can e.g. consist of single keywords, sets of keywords, or potentially lengthy descriptions with many attributes describing the image content. Using this data, two encoders are jointly trained, namely a text encoder embedding the alt-texts and an image encoder embedding the corresponding images into a shared latent space. These two encoders are trained with a contrastive loss, encouraging the embeddings of matching images and alt-text to be similar, and at the same time to be dissimilar from all other image and alt-text embeddings." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.245, + 0.473, + 0.442 + ], + "angle": 0, + "content": "Once trained, such an encoder pair can be used in many ways: It can be specialized to classifying a fixed set of visual concepts via their textual descriptions (zero-shot classification); the embeddings can be used to retrieve images given a textual description and vice-versa; or the vision encoder can be transferred in supervised fashion to a downstream task by fine-tuning on a labeled data set or by training a head on top of the frozen image encoder representation. In principle, the text encoder can be used as a standalone text embedding, but this application—to our knowledge—has not been explored in-depth, with some authors citing the low quality of the alt-texts leading to weak language modeling performance of the text encoder [67]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.442, + 0.473, + 0.595 + ], + "angle": 0, + "content": "Previous works [46, 53] have shown that the image and text encoder can be realized with a single shared transformer model (henceforth referred to as single tower model, or 1T-CLIP), where the images are embedded using a patch embedding, and the tokenized text is embedded using a separate word embedding. Apart from the modality-specific embeddings, all model parameters are shared for the two modalities. While this type of sharing usually leads to a minor performance drop on image/image-language tasks it also halves the number of model parameters." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.595, + 0.473, + 0.732 + ], + "angle": 0, + "content": "CLIPPO takes this idea one step further: text inputs are rendered on blank images, and are subsequently dealt with entirely as images, including the initial patch embedding (see Fig. 1 for an illustration). By training this single vision transformer contrastively as prior works, we obtain a single vision transformer model that can understand both images and text through the single interface of vision and provides a single representation which can be used to solve image, image-language, and pure language understanding tasks." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.733, + 0.473, + 0.809 + ], + "angle": 0, + "content": "Alongside multimodal versatility, CLIPPO alleviates common hurdles with text processing, namely the development of an appropriate tokenizer and vocabulary. This is particularly interesting in a massively multilingual setup, where the text encoder has to handle dozens of languages." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.81, + 0.473, + 0.903 + ], + "angle": 0, + "content": "We find that CLIPPO trained on image/alt-text pairs performs comparably with its 1T-CLIP counterpart on common image and image-language benchmarks, and is competitive with strong baseline language models on the GLUE benchmark [73]. However, due to the low quality of the alt-texts which are often not grammatical sentences, learn" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.895, + 0.199 + ], + "angle": 0, + "content": "ing language understanding exclusively from alt-texts is fundamentally limited. Therefore, we augment image/alt-text contrastive pretraining with language-based contrastive training. Specifically, we use positive pairs of consecutive sentences sampled from a text corpus which is seamlessly integrated into the contrastive training by supplementing batches of image/alt-texts with (rendered) text/text pairs." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.211, + 0.634, + 0.228 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.236, + 0.75, + 0.252 + ], + "angle": 0, + "content": "4.1. Training details and models" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.258, + 0.895, + 0.365 + ], + "angle": 0, + "content": "We rely on a single training setup for all our baselines and visual text models. This setup was tuned to produce good results for standard image/alt-text contrastive training as in [56] (using exactly the same loss function as [56], following the pseudocode in [56, Fig. 3]) and we found that it readily transfers to 1T-CLIP and CLIPPO (including variants with text/text co-training)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.365, + 0.896, + 0.606 + ], + "angle": 0, + "content": "Our default architecture is a ViT-B/16 [16] and we perform a subset of experiments with a ViT-L/16 architecture to study the effect of scale (we equip both models a MAP head [40] to pool embeddings). In all cases, the representation dimension used for the contrastive loss is 768. We set the batch size to 10,240 and train the main models for 250k steps, using a minimum 100k training steps for ablations. For models co-trained with a certain percentage of text/text data, we scale the number of iterations such that the number of image/alt-text pairs seen matches the number of iterations of the corresponding model without text/text data (e.g. when \\(50\\%\\) of the data is text/text pairs we increase the number of iterations from 250k to 500k). The contrastive loss is computed across the full batch. We use the Adafactor optimizer [66] with a learning rate of \\(10^{-3}\\) and decoupled weight decay with weight \\(10^{-4}\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.606, + 0.895, + 0.728 + ], + "angle": 0, + "content": "Baseline CLIP-style models are trained using the T5-en SentencePiece tokenizer [57]; we use the abbreviation \\(\\mathrm{CLIP^{*}}\\) for the two tower model from [56] trained from scratch using the setup described above, to avoid confusion with the model released by [56]. A sequence length of 196 is used, as this matches the number of visual text \"tokens\" CLIPPO can process with patch size 16 has at \\(224\\mathrm{px}\\) resolution (which we use throughout unless noted otherwise)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.73, + 0.895, + 0.868 + ], + "angle": 0, + "content": "Visual text For visual text rendering [60, 63] relied on the Google Noto font family1 which supports the majority of Unicode code points. Here, we use the GNU Unifont bitmap font2, which has a similar coverage but allows for efficient, lookup-based on-the-fly rendering in our preprocessing pipeline. We emphasize that this rendering strategy does not slow down training compared to tokenizer-based models. In preliminary explorations, we found this to be performance-neutral when compared to the Noto font." + }, + { + "type": "page_footnote", + "bbox": [ + 0.516, + 0.875, + 0.757, + 0.887 + ], + "angle": 0, + "content": "1https://fonts.google.com/nota" + }, + { + "type": "page_footnote", + "bbox": [ + 0.519, + 0.888, + 0.755, + 0.9 + ], + "angle": 0, + "content": "2http://unifoundry.com/unifont" + }, + { + "type": "list", + "bbox": [ + 0.516, + 0.875, + 0.757, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "11008" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.176, + 0.089, + 0.789, + 0.238 + ], + "angle": 0, + "content": "
#param.training datasetI1k 10s.I1k 0s.C I→TC T→IF I→TF T→I
CLIP*203MWebLI55.865.148.531.379.259.4
1T-CLIP118MWebLI53.962.348.030.377.558.2
CLIPPO93MWebLI53.061.447.330.176.457.3
CLIPPO93MWebLI + 25%C452.157.440.726.768.951.8
CLIPPO93MWebLI + 50%C448.053.135.223.464.847.2
1T-CLIP L/16349MWebLI60.867.850.732.581.061.0
CLIPPO L/16316MWebLI60.367.450.633.479.262.6
CLIPPO L/16316MWebLI + 25%C460.566.044.529.872.957.3
CLIPPO L/16316MWebLI + 50%C456.861.739.727.370.154.7
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.245, + 0.892, + 0.314 + ], + "angle": 0, + "content": "Table 1. Vision and vision-language cross-modal results. We report ImageNet-1k 10-shot linear transfer validation accuracy (I1k 10s.), ImageNet-1k zero-shot transfer validation accuracy (I1k 0s.), image-to-text and text-to-image retrieval recall@1 on MS-COCO (C I→T and C T→I) and on Flickr30k (F T→I and F I→T). CLIPPO and 1T-CLIP incur a minor drop in these evaluations compared to CLIP*, while only using about half of the model parameters. Co-training with text pairs from C4 (models with + xx%C4) degrades performance on some cross-modal tasks (but leads to improved language understanding capabilities, see Table 2)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.327, + 0.471, + 0.568 + ], + "angle": 0, + "content": "Image/alt-text data We use the WebLI data set introduced in [10] which comprises 10 billion images with 12 billion corresponding alt-texts. Importantly, WebLI comprises alt-texts in 109 languages (unlike previous data sets such as LAION-400M [64] which only contain English alt-texts) and it is therefore a great foundation to study multilingual language-image pretraining and its applications. Please refer to [10, Fig. 3] for details on the alt-text language distribution. For English-only models we obtain English versions of non-English alt-texts via GCP Translation \\(\\mathrm{API}^3\\). In addition to alt-text, WebLI also provides OCR annotations, which we do not use in this paper. Finally, WebLI was processed with a de-duplication step removing all images from various splits of the image evaluation sets used in this paper. Please refer to [10, Sec. 3.2] for more details on the WebLI data set and to [10, Appendix B] for a datasheet." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.569, + 0.47, + 0.614 + ], + "angle": 0, + "content": "We also present a subset of results based on LAION-400M [64] and YFCC-100M [71] as an additional comparison points, see Appendix C.1 and C.2, respectively." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.618, + 0.47, + 0.769 + ], + "angle": 0, + "content": "Text/text data For co-training with text/text pairs we primarily rely on the publicly available Colossal Clean Crawled Corpus (C4; default/English split) [57]. We randomly sample pairs of consecutive sentences and contrastively train on these pairs, i.e., the model is trained for embedding-based next sentence prediction (NSP) [47]. We also experiment with pairs of parallel sentences in different languages from the WMT19 data set [18] as well as back-translated English sentences derived from C4 following the strategy described in [12]." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.779, + 0.3, + 0.793 + ], + "angle": 0, + "content": "4.2. Evaluations and metrics" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.802, + 0.469, + 0.879 + ], + "angle": 0, + "content": "To evaluate the vision and vision/language understanding capabilities of our models we use standard metrics from the literature [53, 56, 85]: \"zero-shot\" transfer, which uses (embedded) textual description of the classes to be classified/retrieved and compares these with image embeddings." + }, + { + "type": "text", + "bbox": [ + 0.506, + 0.327, + 0.892, + 0.417 + ], + "angle": 0, + "content": "We report the classification accuracy on ImageNet-1k [14] as well as the recall@1 for cross-modal retrieval on MSCOCO [9] and Flickr30k [81]. Furthermore, we test the low-data transfer performance of the models by means of the linear adaptation protocol from [16], reporting the 10-shot accuracy on ImageNet-1k." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.418, + 0.893, + 0.689 + ], + "angle": 0, + "content": "We also evaluate CLIPPO and baselines on the popular VQA benchmark VQAv2 [25]. To construct a VQA model using a single pretrained ViT we render the question at the top end of the corresponding image (using the same Unifont renderer as used for CLIPPO training) and follow the standard prediction setup where the answer is predicted as the most likely answer from the training set, i.e. by classification. Specifically, we replace the last layer of our pretrained CLIPPO and baselines with a randomly initialized one with the appropriate number of outputs and fine-tune on VQAv2. This setup tests the ability of the pretrained ViT to combine image and text in intermediate layers as it has produce a single output from a fused image/text input image, unlike in the other cross-modal tasks (and pretraining), where image and text representations are computed with two separate forward passes. Please refer to Appendix A in the supplementary material for examples images with rendered questions and Appendix B.1 for details on the fine-tuning protocol." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.69, + 0.892, + 0.765 + ], + "angle": 0, + "content": "Multilingual capabilities are assessed via zero-shot retrieval on CrossModal3600 [70], which is a geographically diverse set comprising 3600 images each human-annotated with captions in 36 languages. The corresponding recall metric is averaged across all languages and images." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.893, + 0.901 + ], + "angle": 0, + "content": "Finally, we evaluate the language understanding capabilities on the General Language Understanding Evaluation (GLUE) benchmark [73] which comprises natural language inference tasks (MNLI, QNLI, RTE), a sentiment analysis task (SST-2), sentence similarity tasks (QQP, STS-B, MRPC), and a linguistic acceptability task (CoLA). Following common practice, we exclude the WNLI task from the benchmark [15, 77]. We transfer our baselines and CLIPPO models by attaching a 2-hidden layer MLP with 768 units to" + }, + { + "type": "page_footnote", + "bbox": [ + 0.094, + 0.887, + 0.372, + 0.9 + ], + "angle": 0, + "content": "3https://cloud.google.com/translate" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "11009" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.087, + 0.092, + 0.468, + 0.397 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.41, + 0.47, + 0.496 + ], + "angle": 0, + "content": "Figure 2. Results on the VQAv2 benchmark (test-dev set). In addition to CLIPPO and baselines produced in this work, we also compare to Pythia and MCAN models with ViT encoders from [67], and with comparably sized METER [17] and ViLT [36] models. CLIPPO outperforms \\(\\mathrm{CLIP^{*}}\\) and 1T-CLIP clearly on \"yes/no\" questions and gets similar performance as task-specific models." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.524, + 0.47, + 0.586 + ], + "angle": 0, + "content": "their representation and following precisely the fine-tuning protocol from BERT [15]. For sentence pair classification tasks we simply render both sentences on the same image, printing [SEP] to mark the start of the second sentence." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.599, + 0.436, + 0.615 + ], + "angle": 0, + "content": "4.3. Vision and vision-language understanding" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.628, + 0.47, + 0.853 + ], + "angle": 0, + "content": "Image classification and retrieval Table 1 shows the performance of CLIPPO along with the baseline models on the benchmarks described in Sec. 4.2. It can be seen that the CLIPPO and 1T-CLIP incur a drop of a 2-3 percentage points absolute compared to \\(\\mathrm{CLIP^{*}}\\). This is not surprising and can be attributed to the fact that single tower models only have about half the parameters count of a corresponding two tower model. The difference in performance between the English-only CLIPPO and 1T-CLIP is very small for a B/16 backbone at \\(100\\mathrm{k}\\) training steps (see Table 6 in the supplementary material), and vanishes with longer training and/or by increasing the model size, despite the fact that CLIPPO has \\(25\\%\\) and \\(10\\%\\) fewer parameters than 1T-CLIP for a B/16 and L/16 architecture, respectively (which is due to the absence of the text embedding in CLIPPO)." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.856, + 0.47, + 0.903 + ], + "angle": 0, + "content": "The multilingual CLIPPO model performs somewhat worse than the corresponding 1T-CLIP, and the gap does not close completely when training longer (see Table 6)." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.892, + 0.137 + ], + "angle": 0, + "content": "However, when evaluated across a broad set of languages on the CrossModal3600 CLIPPO performs on par with or slightly better than 1T-CLIP (see Sec. 4.4 below)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.139, + 0.892, + 0.29 + ], + "angle": 0, + "content": "As we add sentence pairs to the training mix the performance on the cross-modal retrieval metrics decreases. This is not surprising as we keep the total batch size constant so that the effective batch size of image/alt-text contrastive training decreases, which is known to impact performance [85]. Interestingly, the 10-shot transfer performance does not move in tandem, but only decreases significantly when half of the training data is sentence pairs. In exchange, co-training with text data leads to significantly improved language understanding performance (see Sec. 4.5)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.295, + 0.893, + 0.537 + ], + "angle": 0, + "content": "VQA In Fig. 2 we report the VQAv2 score of our models and baselines. It can be seen that CLIPPO outperforms \\(\\mathrm{CLIP^{*}}\\), 1T-CLIP, as well as a pretrained ViT-B/16 from [16] by a significant margin, achieving a score of 66.3, and constraining with \\(25\\%\\) C4 data leads to a slight improvement of the score. The improved score of CLIPPO is mainly due to better performance in \"yes/no\" questions. Increasing the model size to L/16 adds another 2 points which originate from improvements in the \"number\" and \"other\" VQAv2 categories. However, note that for an L/16 architecture 1T-CLIP performs competitively with CLIPPO (see Table 7). One possible explanation for this could be that 1T-CLIP develops better OCR capabilities thanks to the higher model capacity (alt-texts can correlate with text in images/scene text, see [10, Fig. 3]). Increasing the resolution to 384px adds 2 to 3 points across models." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.539, + 0.893, + 0.901 + ], + "angle": 0, + "content": "We also compare CLIPPO with baselines from the literature. Specifically, [17] proposes a framework (called ME-TER) for multimodal tasks, where pretrained transformer-based image and text encoders are combined with a transformer-based fusion module. CLIPPO L/16 achieves performance competitive with their model combining a CLIP B/32 vision backbone with a BERT-Base language backbone, which is roughly comparable in size and computational cost with our L/16 models. Another related work is [67], which combines different CLIP vision backbones with two existing VQA systems, Pythia [33] and MCAN [84]. CLIPPO outperforms different CLIP ViT-based Pythia and MCAN models from [67]. Note, however, that ResNet-based CLIP backbones lead to better results when combined with these systems. We further note that both [17] and [67] also investigate training their models on a mix of different image-text data sets with multiple objectives such as grounded masked language modeling and text-image matching, before transferring to the VQA task, which leads to significant improvements. ViLT [36] relies on such a strategy to train a single transformer backbone jointly encoding image and text tokens. At 384px resolution, CLIPPO (with \\(25\\%\\) C4 data) obtains a VQA score comparable with that of ViLT (and other models from the" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "11010" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.091, + 0.47, + 0.138 + ], + "angle": 0, + "content": "literature such as ViLBERT [48], VisualBERT [43], and PixelBERT [30]), despite only using a contrastive objective for pretraining." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.148, + 0.45, + 0.166 + ], + "angle": 0, + "content": "4.4. Multilingual vision-language understanding" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.172, + 0.473, + 0.489 + ], + "angle": 0, + "content": "For typical language models, tokenizer choice can be a challenging process [78]. Commonly used English-language tokenizers generalize poorly to non-latin scripts [85]. This can be alleviated by the use of larger, multilingual vocabularies, at the expense of very large parameter counts. CLIPPO bypasses this issue, removing any language-related bias stemming from unbalanced or restrictive tokenizers. We consider multilingual image/text retrieval on Crossmodal3600 and compare CLIPPO, trained on WebLI with multilingual alt-texts, against 1T-CLIP with a number of SentencePiece tokenizers; one trained from 300M WebLI multilingual alt-texts, English (T5-en) and multilingual (T5-all) tokenizers from T5 [57], and a multilingual tokenizer (mT5) from mT5 [79], all with a vocabulary size of 32,000. The results are shown in Fig. 4. On average, CLIPPO achieves comparable retrieval performance to these baselines. In the case of mT5, the use of extra data to create the specialized vocabulary can boosts performance above that of CLIPPO; the leveraging of such extra parameters and data in the multilingual context will be an interesting future direction for CLIPPO." + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.499, + 0.268, + 0.62 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.633, + 0.27, + 0.744 + ], + "angle": 0, + "content": "Figure 3. Tokenization efficiency analyzed in terms of the sequence length produced by a given method. CLIPPO produces smaller sequences for the majority of languages compared to 1T-CLIP with alternative tokenizers." + }, + { + "type": "text", + "bbox": [ + 0.298, + 0.493, + 0.47, + 0.765 + ], + "angle": 0, + "content": "Tokenization efficiency If a tokenizer is well suited to a particular dataset, it will tokenize to shorter sequences—this is especially the case when byte fallback [39] is enabled. SentencePiece tokenizers have the advantageous ability to tokenize entire—possibly quite long—words to single tokens. CLIPPO cannot learn any such compression, but benefits from equal treatment of all languages and" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.473, + 0.903 + ], + "angle": 0, + "content": "words: it will by definition generalize equally well to all data, as its tokenization schema has not been trained on a specific dataset. We analyze 20,000 samples for each of the 104 C4 languages. Each CLIPPO token is assumed to be a \\(16 \\times 16\\) patch; though in typical computations all approaches considered here would pad to a fixed length, we compute CLIPPO's sequence length according to the last patch which contains rendered text. Fig. 3 shows the fraction of C4 languages where CLIPPO processes tokens" + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.09, + 0.633, + 0.182 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.633, + 0.091, + 0.758, + 0.182 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.759, + 0.091, + 0.889, + 0.181 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.189, + 0.892, + 0.273 + ], + "angle": 0, + "content": "Figure 4. Zero-shot image/text retrieval performance on CrossModal3600 [70]. Although specialized (mc4) tokenizers can be leveraged to improve multilingual performance CLIPPO (dashed black line) broadly matches or exceeds comparable 1T-CLIP models trained with vocabulary size 32,000 (the word embeddings result in a \\(27\\%\\) increase in parameter count compared to CLIPPO)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.29, + 0.892, + 0.397 + ], + "angle": 0, + "content": "more efficiently than the vocabularies discussed above. We conservatively define \"more efficient\" as producing a shorter token sequence for over \\(75\\%\\) of examples. Even so, CLIPPO is indeed more efficient across the majority of languages. Per-language breakdowns of multilingual retrieval performance and tokenization efficiency are further discussed in Appendix C.3." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.408, + 0.729, + 0.424 + ], + "angle": 0, + "content": "4.5. Language understanding" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.432, + 0.892, + 0.734 + ], + "angle": 0, + "content": "Table 2 shows the GLUE benchmark results of CLIPPO and baselines. One can observe that CLIPPO trained on WebLI performs competitively with the BiLSTM+Attn+ELMo baseline which relies on deep word embeddings trained on a large language corpus. Also, it can be seen that CLIPPO along with 1T-CLIP outperform the language encoder trained using standard contrastive language vision pretraining \\((\\mathrm{CLIP}^{*})\\). This indicates that multimodal training in a single encoder benefits language understanding. Furthermore, CLIPPO achieves a much higher GLUE score than the \\(\\mathrm{CLIP}^*\\) image encoder, which in turn leads to significantly better results than fine-tuning a ViT-B/16 from scratch on GLUE (see Appendix C.2 for additional results). Unsurprisingly, the models pretrained on WebLI cannot do better than random guessing on the CoLA evaluation which requires to assess the grammatical correctness of sentences (recall that alt-texts are rarely grammatical sentences). Also the accuracy of \\(\\mathrm{CLIP}^*\\) and 1T-CLIP vision encoders we observe for SST-2 is in agreement with what was reported in [56, Table 10] for CLIP with a ViT-B/16 image encoder." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.735, + 0.892, + 0.902 + ], + "angle": 0, + "content": "Adding sentence pairs form the C4 corpus gradually improves the GLUE score, and when half of the examples are sentence pairs our model becomes competitive with PILXEL, while still retaining decent image and vision-language understanding capabilities (cf. Table 1). Note, however, that there is a trade-off between language-only tasks and tasks that involve image understanding. Finally, training CLIPPO only on sentence pairs leads to a model which outperforms PILXEL by a significant margin. However, our model has seen more sentence pairs than PILXEL, so PILXEL might improve as well when training longer." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "11011" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.109, + 0.089, + 0.855, + 0.292 + ], + "angle": 0, + "content": "
training datasetMNLI-M/MMQQPQNLISST-2COLASTS-BMRPCRTEavg
BERT-BaseWiki + BC84.0 / 84.287.691.092.660.388.890.269.583.1
PIXELWiki + BC78.1 / 78.984.587.889.638.481.188.260.576.3
BiLSTM66.7 / 66.782.077.087.517.672.085.158.568.1
BiLSTM+Attn, ELMo72.4 / 72.483.675.291.544.156.182.152.770.0
CLIP* img enc.WebLI66.4 / 67.578.669.478.60.05.281.252.755.5
CLIP* text enc.WebLI71.8 / 72.582.773.086.26.665.081.453.865.9
1T-CLIP text enc.WebLI72.6 / 73.083.880.784.90.079.683.357.068.3
CLIPPOWebLI73.0 / 72.684.381.286.81.880.584.153.468.6
CLIPPOWebLI + 25%C477.7 / 77.285.383.190.928.283.484.559.274.4
CLIPPOWebLI + 50%C479.2 / 79.286.484.292.938.983.484.859.976.6
CLIPPOC479.9 / 80.286.785.293.350.984.786.358.578.4
CLIPPO L/16WebLI + 25%C476.6 / 75.587.179.993.248.284.184.656.076.1
CLIPPO L/16WebLI + 50%C482.3 / 82.487.986.794.255.385.885.959.280.0
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.295, + 0.895, + 0.381 + ], + "angle": 0, + "content": "Table 2. Results for the GLUE benchmark (dev set). The metric is accuracy except for the performance on QQP and MRPC, which is measured using the \\(F_{1}\\) score, CoLA which uses Matthew's correlation, and STS-B which evaluated based on Spearman's correlation coefficient. \"avg\" corresponds to the average across all metrics. The results for BERT-Base and PILXEL are from [60, Table 3], and BiLSTM and BiLSTM+Attn, ELMo from [73, Table 6]. All encoders considered here have a transformer architecture comparable to BERT-Base (up to the text embedding layer), except for CLIPPO L/16 which uses a ViT L/16, and the two BiLSTM model variants. Wiki and BC stand for (English) Wikipedia and Bookcorpus [86] data, respectively." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.391, + 0.287, + 0.407 + ], + "angle": 0, + "content": "4.6. Ablations and analysis" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.419, + 0.471, + 0.616 + ], + "angle": 0, + "content": "Impact of weight sharing across modalities The fact that CLIPPO 1) uses a shared patch embedding for regular images and text images and 2) this embedding has considerably fewer parameters than the text embedding of 1T-CLIP and \\(\\mathrm{CLIP^{*}}\\) provokes the question of whether CLIPPO could benefit from separate patch embeddings for text images and regular images. Further, CLIPPO relies on a single head to compute the output representation for images and text, and relaxing this constraint by using separate heads for the two modalities could lead to more expressive representations. The results (deferred to Appendix D.1) show that neither of these variants lead to improved image classification or retrieval metrics compared to CLIPPO." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.619, + 0.47, + 0.697 + ], + "angle": 0, + "content": "Impact of the text location We test whether rendering the question at the top, middle, or bottom of the image impacts the VQA performance of CLIPPO and find that it does not, provided that we increase the learning rate of the positional embedding during fine-tuning (see Appendix D.2)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.7, + 0.47, + 0.821 + ], + "angle": 0, + "content": "Typographic attacks Since CLIPPO is trained on large amounts of rendered (alt-)text it is important to check whether it becomes more susceptible to typographic attacks—the tendency of CLIP-style models to zero-shot classify an image according to adversarially injected scene text unrelated to the scene [23,42,50]. In Appendix D.3 we present results indicating that CLIPPO is no more vulnerable to typographic attacks than 1T-CLIP and \\(\\mathrm{CLIP^{*}}\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Modality gap Liang et al. [45] discovered that text and image embeddings of CLIP-style models form two distinct clusters rather than both filling the embedding space densely and occupying the same spatial region. They attribute this phenomenon to a combination of initialization" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.391, + 0.892, + 0.649 + ], + "angle": 0, + "content": "conditions and properties of the loss function/training dynamics. Since we consider single tower models here, and also co-train some of these models with text-only pairs it is interesting to see how this affects the modality gap. We compute the gap and visualize it following the recipe from [45] in Fig. 5 (see Appendices D.4 and D.5 for additional visualizations). CLIPPO attains a slightly lower modality gap than \\(\\mathrm{CLIP}^*\\), but clearly features a clustering structure for image and text embeddings. However, when training contrastively with sentence pairs in addition to image/alt-text pairs, the clustering structure disappears, the image and text embeddings overlap, and the modality gap decreases significantly. A possible explanation for this behavior could be that the additional learning pressure induced by the contrastive loss on sentence pairs encourages text embeddings to spread out more and hence the structure of all embeddings changes." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.651, + 0.892, + 0.789 + ], + "angle": 0, + "content": "Text/text co-training objectives To corroborate that contrastive NSP is a sensible objective to improve language understanding in the context of CLIPPO, we train CLIPPO without any image/alt-text data on pairs of parallel translated sentences (this is straight-forward in our framework since visual text is language-agnostic), as well as English back-translated data, and evaluate the resulting text representations on GLUE. Table 3 shows that NSP on C4 clearly achieves the highest GLUE score." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.8, + 0.744, + 0.817 + ], + "angle": 0, + "content": "5. Discussion and limitations" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.825, + 0.892, + 0.903 + ], + "angle": 0, + "content": "We proposed and evaluated CLIPPO which produces a single ViT that can understand images and language jointly using images as a sole input modality. Perhaps surprisingly, CLIPPO matches the performance of the 1T-CLIP baseline across many of the considered tasks, and only in" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "11012" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.123, + 0.089, + 0.422, + 0.129 + ], + "angle": 0, + "content": "
WMT19WMT19 BTC4 NSP
GLUE score61.266.677.6
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.14, + 0.47, + 0.21 + ], + "angle": 0, + "content": "Table 3. Ablation of text pair-based contrastive co-training tasks: Training on parallel translated sentences (WMT19), training on parallel back-translated sentences (WMT19 BT), and NSP for sentences sampled from C4 (C4 NSP). C4 NSP leads to the highest GLUE score by a large margin." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.238, + 0.468, + 0.358 + ], + "angle": 0, + "content": "curs a minor drop compared to the \\(\\mathrm{CLIP^{*}}\\) baseline, despite having less than half the parameters of \\(\\mathrm{CLIP^{*}}\\). As we showed, the image-only interface enables a simple, unified data pipeline for training on and transferring to mixed modalities. CLIPPO opens the door for additional modalities (e.g. spectrograms) and, as we hope, might inspire applications of pixel-only models beyond contrastive training. Nevertheless, several limitations remain, as discussed next." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.364, + 0.468, + 0.544 + ], + "angle": 0, + "content": "Co-training First, to achieve language understanding performance competitive with PILXEL and BERT on GLUE, contrastive co-training with text pairs is necessary. While adding \\(25\\%\\) C4 data to the batch seems to strike a good balance across all tasks considered, it does induce a nonnegligible drop in zero-shot image classification and image/text retrieval. This drop becomes more severe as the fraction of C4 examples increases. We observed an associated change in modality gap, and further investigation of the representation in the context of co-training might help to develop models that achieve better overall performance in the co-training setup." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.549, + 0.468, + 0.745 + ], + "angle": 0, + "content": "Diverse rendered text CLIPPO currently relies on cleanly rendered text as an input and its capabilities to handle text from documents or web pages without further adaption is limited (besides the basic OCR capabilities that CLIP-style models learn from image/alt-text pairs). We emphasize that sophisticated OCR and document understanding is not a goal of this paper, and training CLIPPO with augmented noisy rendered text that mimics the distribution of documents and websites is likely to lead to worse performance across the considered tasks, since image/alt-text pairs are less correlated and provide a weaker learning signal. However, developing CLIPPO further to handle less clean visual text will open many additional applications." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.468, + 0.901 + ], + "angle": 0, + "content": "Generative modeling CLIPPO, like CLIP, BERT, PILXEL and many other models, uses an encoder-only design and hence lacks the ability to generate text outputs. A common approach to equip encoder-only models with generation capabilities (e.g., for image captioning or VQA) is to simply combine them with a (potentially pretrained) language model [8, 76]. This approach naturally also applies to CLIPPO and PILXEL, but defeats the advantages of visual text in certain (e.g. multilingual) scenarios. While visual text outputs have previously been explored in the context of" + }, + { + "type": "image", + "bbox": [ + 0.507, + 0.084, + 0.888, + 0.197 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.207, + 0.892, + 0.305 + ], + "angle": 0, + "content": "Figure 5. Visualization of the modality gap for \\(\\mathrm{CLIP^{*}}\\) and CLIPPO optionally trained with \\(25\\%\\) C4 data. The visualization follows the analysis from [45] and shows embedded images (blue dots) and corresponding alt-text (orange dots) from the WebLI validation set, projected to the first two principal components of the validation data matrix. CLIPPO has a slightly smaller modality gap than \\(\\mathrm{CLIP^{*}}\\); co-training with C4 data strongly reduces the gap." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.325, + 0.891, + 0.354 + ], + "angle": 0, + "content": "machine translation [49], it remains unclear what a scalable tokenizer-free way to generate text is." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.359, + 0.892, + 0.54 + ], + "angle": 0, + "content": "Multilingual learning Finally, we showed that CLIPPO obtains strong multilingual image/text retrieval performance without requiring the development of an appropriate tokenizer. For fine-grained adjustment and balancing of the retrieval performance further steps will be necessary, including data balancing and potentially co-training with multi-lingual text data. Furthermore, similar to PILXEL, CLIPPO relies on certain ad-hoc design choices w.r.t. the visual representation, for example the left-to-right rendering of Arabic scripts. This approach leads to decent performance on average, but it is not clear what kind of unwanted effects it introduces and how these could be mitigated." + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.554, + 0.618, + 0.57 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.58, + 0.892, + 0.746 + ], + "angle": 0, + "content": "We introduced CLIPPO, a model for processing images and text jointly through the lens of vision. This reduces design choices (in particular w.r.t. tokenization) and parameter count, simplifies data processing pipelines and transfer recipes, and increases generality across multiple languages. We also explored methods of enhancing language understanding, where traditional image/alt-text contrastive models trained on web data fall short. We demonstrated this is possible by co-training with text pairs, with CLIPPO models outperforming strong NLP baselines while maintaining solid image understanding capabilities." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.747, + 0.892, + 0.837 + ], + "angle": 0, + "content": "Although we presented a unified contrastive training algorithm, CLIPPO suffers somewhat when co-training on multiple tasks, and future work to harmonize the co-training could enhance the models significantly. Deeper understanding of the design choices in rendering text as images, and their impact on performance, is another interesting avenue." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.842, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Acknowledgments We would like to thank Lucas Beyer, Josip Djolonga, Alexander Kolesnikov, Mario Lucic, Andreas Steiner, Xiao Wang, and Xiaohua Zhai for inspiring discussions and helpful feedback. We also thank Jeffrey Sorensen for help with the text rendering preprocessing." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "11013" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.079, + 0.09, + 0.176, + 0.107 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.17 + ], + "angle": 0, + "content": "[1] Hassan Akbari, Liangzhe Yuan, Rui Qian, Wei-Hong Chuang, Shih-Fu Chang, Yin Cui, and Boqing Gong. VATT: transformers for multimodal self-supervised learning from raw video, audio and text. In NeurIPS, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.172, + 0.472, + 0.31 + ], + "angle": 0, + "content": "[2] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katie Millican, Malcolm Reynolds, Roman Ring, Eliza Rutherford, Serkan Cabi, Tengda Han, Zhitao Gong, Sina Samangooei, Marianne Monteiro, Jacob Menick, Sebastian Borgeaud, Andrew Brock, Aida Nematzadeh, Sahand Sharifzadeh, Mikolaj Binkowski, Ricardo Barreira, Oriol Vinyals, Andrew Zisserman, and Karen Simonyan. Flamingo: A visual language model for few-shot learning. In NeurIPS, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.311, + 0.47, + 0.366 + ], + "angle": 0, + "content": "[3] Srikar Appalaraju, Bhavan Jasani, Bhargava Urala Kota, Yusheng Xie, and R. Manmatha. DocFormer: End-to-end transformer for document understanding. In ICCV, pages 973-983, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.368, + 0.469, + 0.408 + ], + "angle": 0, + "content": "[4] Lucas Beyer, Xiaohua Zhai, and Alexander Kolesnikov. Big Vision. https://github.com/google-research/big visions, 2022.1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.41, + 0.47, + 0.479 + ], + "angle": 0, + "content": "[5] Zalan Borsos, Raphael Marinier, Damien Vincent, Eugene Kharitonov, Olivier Pietquin, Matthew Sharifi, Olivier Teboul, David Grangier, Marco Tagliasacchi, and Neil Zeghidour. AudioLM: a language modeling approach to audio generation. CoRR, abs/2209.03143, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.481, + 0.47, + 0.563 + ], + "angle": 0, + "content": "[6] Tom B. Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert-Voss, Gretchen Krueger, Tom Henighan, et al. Language models are few-shot learners. In NeurIPS, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.565, + 0.47, + 0.633 + ], + "angle": 0, + "content": "[7] Lili Chen, Kevin Lu, Aravind Rajeswaran, Kimin Lee, Aditya Grover, Michael Laskin, Pieter Abbeel, Aravind Srinivas, and Igor Mordatch. Decision transformer: Reinforcement learning via sequence modeling. In NeurIPS, pages 15084-15097, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.635, + 0.469, + 0.676 + ], + "angle": 0, + "content": "[8] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey E. Hinton. A simple framework for contrastive learning of visual representations. In ICML, 2020. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.678, + 0.47, + 0.732 + ], + "angle": 0, + "content": "[9] Xinlei Chen, Hao Fang, Tsung-Yi Lin, Ramakrishna Vedantam, Saurabh Gupta, Piotr Dólar, and C. Lawrence Zitnick. Microsoft COCO captions: Data collection and evaluation server. CoRR, abs/1504.00325, 2015. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.734, + 0.47, + 0.872 + ], + "angle": 0, + "content": "[10] Xi Chen, Xiao Wang, Soravit Changpinyo, A. J. Piergiovanni, Piotr Padlewski, Daniel Salz, Sebastian Goodman, Adam Grycner, Basil Mustafa, Lucas Beyer, Alexander Kolesnikov, Joan Puigcerver, Nan Ding, Keran Rong, Hassan Akbari, Gaurav Mishra, Linting Xue, Ashish Thapliyal, James Bradbury, Weicheng Kuo, Mojtaba Seyedhosseini, Chao Jia, Burcu Karagol Ayan, Carlos Riquelme, Andreas Steiner, Anelia Angelova, Xiaohua Zhai, Neil Houlsby, and Radu Soricut. PaLI: A jointly-scaled multilingual language-image model. In ICLR, 2023. 1, 2, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.873, + 0.47, + 0.9 + ], + "angle": 0, + "content": "[11] Aakanksha Chowdhery, Sharan Narang, Jacob Devlin, Maarten Bosma, Gaurav Mishra, Adam Roberts, Paul" + }, + { + "type": "list", + "bbox": [ + 0.079, + 0.115, + 0.472, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.356 + ], + "angle": 0, + "content": "Barham, Hyung Won Chung, Charles Sutton, Sebastian Gehrmann, Parker Schuh, Kensen Shi, Sasha Tsvyashchenko, Joshua Maynez, Abhishek Rao, Parker Barnes, Yi Tay, Noam Shazeer, Vinodkumar Prabhakaran, Emily Reif, Nan Du, Ben Hutchinson, Reiner Pope, James Bradbury, Jacob Austin, Michael Isard, Guy Gur-Ari, Pengcheng Yin, Toju Duke, Anselm Levskaya, Sanjay Ghemawat, Sunipa Dev, Henryk Michalewski, Xavier Garcia, Vedant Misra, Kevin Robinson, Liam Fedus, Denny Zhou, Daphne Ippolito, David Luan, Hyeontaek Lim, Barret Zoph, Alexander Spiridonov, Ryan Sepassi, David Dohan, Shivani Agrawal, Mark Omernick, Andrew M. Dai, Thanumalayan Sankaranarayana Pillai, Marie Pellat, Aitor Lewkowycz, Erica Moreira, Rewon Child, Oleksandr Polozov, Katherine Lee, Zongwei Zhou, Xuezhi Wang, Brennan Saeta, Mark Diaz, Orhan Firat, Michele Catasta, Jason Wei, Kathy Meier-Hellstern, Douglas Eck, Jeff Dean, Slav Petrov, and Noah Fiedel. PaLM: Scaling language modeling with pathways. CoRR, abs/2204.02311, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.358, + 0.892, + 0.412 + ], + "angle": 0, + "content": "[12] Geoffrey Cideron, Sertan Girgin, Anton Raichuk, Olivier Pietquin, Olivier Bachem, and Léonard Hussenot. vec2text with round-trip translations. CoRR, abs/2209.06792, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.415, + 0.892, + 0.485 + ], + "angle": 0, + "content": "[13] Alexis Conneau, Kartikay Khandelwal, Naman Goyal, Vishrav Chaudhary, Guillaume Wenzek, Francisco Guzmán, Edouard Grave, Myle Ott, Luke Zettlemoyer, and Veselin Stoyanov. Unsupervised cross-lingual representation learning at scale. In ACL, pages 8440–8451, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.487, + 0.892, + 0.528 + ], + "angle": 0, + "content": "[14] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In CVPR, 2009. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.53, + 0.892, + 0.585 + ], + "angle": 0, + "content": "[15] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. BERT: pre-training of deep bidirectional transformers for language understanding. In NAACL-HLT, 2019, 1, 2, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.588, + 0.892, + 0.67 + ], + "angle": 0, + "content": "[16] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth \\(16 \\times 16\\) words: Transformers for image recognition at scale. In ICLR, 2021. 1, 2, 3, 4, 5, 14, 16, 17, 24" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.672, + 0.892, + 0.742 + ], + "angle": 0, + "content": "[17] Zi-Yi Dou, Yichong Xu, Zhe Gan, Jianfeng Wang, Shuohang Wang, Lijuan Wang, Chenguang Zhu, Pengchuan Zhang, Lu Yuan, Nanyun Peng, Zicheng Liu, and Michael Zeng. An empirical study of training end-to-end vision-and-language transformers. In CVPR, pages 18145-18155, 2022. 5, 14, 17" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.744, + 0.892, + 0.799 + ], + "angle": 0, + "content": "[18] Wikipedia Foundation. ACL 2019 Fourth Conference on Machine Translation (WMT19), Shared Task: Machine Translation of News. http://www.statmt.org/wmt19/translation-task.html.4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.802, + 0.892, + 0.843 + ], + "angle": 0, + "content": "[19] Tianyu Gao, Xingcheng Yao, and Danqi Chen. Simcse: Simple contrastive learning of sentence embeddings. In EMNLP, pages 6894-6910, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.845, + 0.892, + 0.899 + ], + "angle": 0, + "content": "[20] John M. Giorgi, Osvald Nitski, Bo Wang, and Gary D. Bader. DeCLUTR: Deep contrastive learning for unsupervised textual representations. In ACL/IJCNLP, pages 879-895, 2021. 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "11014" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.147 + ], + "angle": 0, + "content": "[21] Rohit Girdhar, Alaaeldin El-Nouby, Mannat Singh, Kalyan Vasudev Alwala, Armand Joulin, and Ishan Misra. OmniMAE: Single model masked pretraining on images and videos. CoRR, abs/2206.08356, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.149, + 0.47, + 0.204 + ], + "angle": 0, + "content": "[22] Rohit Girdhar, Mannat Singh, Nikhila Ravi, Laurens van der Maaten, Armand Joulin, and Ishan Misra. Omnivore: A single model for many visual modalities. In CVPR, pages 16081-16091, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.207, + 0.469, + 0.261 + ], + "angle": 0, + "content": "[23] Gabriel Goh, Nick Cammarata, Chelsea Voss, Shan Carter, Michael Petrov, Ludwig Schubert, Alec Radford, and Chris Olah. Multimodal neurons in artificial neural networks. *Distill*, 2021, 7, 21" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.264, + 0.469, + 0.303 + ], + "angle": 0, + "content": "[24] Yuan Gong, Yu-An Chung, and James R. Glass. AST: audio spectrogram transformer. In Interspeech, pages 571-575, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.306, + 0.469, + 0.361 + ], + "angle": 0, + "content": "[25] Yash Goyal, Tejas Khot, Douglas Summers-Stay, Dhruv Bartra, and Devi Parikh. Making the V in VQA matter: Elevating the role of image understanding in visual question answering. In CVPR, 2017. 4, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.363, + 0.469, + 0.404 + ], + "angle": 0, + "content": "[26] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollár, and Ross B. Girshick. Masked autoencoders are scalable vision learners. In CVPR, pages 15979-15988, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.406, + 0.469, + 0.474 + ], + "angle": 0, + "content": "[27] Jonathan Ho, William Chan, Chitwan Saharia, Jay Whang, Ruiqi Gao, Alexey A. Gritsenko, Diederik P. Kingma, Ben Poole, Mohammad Norouzi, David J. Fleet, and Tim Salimans. Imagen video: High definition video generation with diffusion models. CoRR, abs/2210.02303, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.477, + 0.469, + 0.531 + ], + "angle": 0, + "content": "[28] Po-Yao Huang, Hu Xu, Juncheng Li, Alexei Baevski, Michael Auli, Wojciech Galuba, Florian Metze, and Christoph Feichtenhofer. Masked autoencoders that listen. CoRR, abs/2207.06405, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.534, + 0.469, + 0.587 + ], + "angle": 0, + "content": "[29] Yupan Huang, Tengchao Lv, Lei Cui, Yutong Lu, and Furu Wei. Layoutlmv3: Pre-training for document AI with unified text and image masking. In ACMMM, pages 4083-4091, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.591, + 0.469, + 0.644 + ], + "angle": 0, + "content": "[30] Zhicheng Huang, Zhaoyang Zeng, Bei Liu, Dongmei Fu, and Jianlong Fu. Pixel-BERT: Aligning image pixels with text by deep multi-modal transformers. CoRR, abs/2004.00849, 2020.6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.647, + 0.469, + 0.743 + ], + "angle": 0, + "content": "[31] Andrew Jaegle, Sebastian Borgeaud, Jean-Baptiste Alayrac, Carl Doersch, Catalin Ionescu, David Ding, Skanda Koppula, Daniel Zoran, Andrew Brock, Evan Shelhamer, Olivier J. Henaff, Matthew M. Botvinick, Andrew Zisserman, Oriol Vinyals, and João Carreira. Perceiver IO: A general architecture for structured inputs & outputs. In ICLR, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.746, + 0.469, + 0.801 + ], + "angle": 0, + "content": "[32] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc V. Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In ICML, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.803, + 0.469, + 0.856 + ], + "angle": 0, + "content": "[33] Yu Jiang, Vivek Natarajan, Xinlei Chen, Marcus Rohrbach, Dhruv Batra, and Devi Parikh. Pythia v0.1: The winning entry to the VQA challenge 2018. CoRR, abs/1807.09956, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.859, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[34] John Jumper, Richard Evans, Alexander Pritzel, Tim Green, Michael Figurnov, Olaf Ronneberger, Kathryn Tunyasuvunakool, Russ Bates, Augustin Žídek, Anna Potapenko, Alex" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.892, + 0.217 + ], + "angle": 0, + "content": "Bridgland, Clemens Meyer, Simon A A Kohl, Andrew J Ballard, Andrew Cowie, Bernardino Romera-Paredes, Stanislav Nikolov, Rishub Jain, Jonas Adler, Trevor Back, Stig Petersen, David Reiman, Ellen Clancy, Michal Zielinski, Martin Steinegger, Michalina Pacholska, Tamas Berghammer, Sebastian Bodenstein, David Silver, Oriol Vinyals, Andrew W Senior, Koray Kavukcuoglu, Pushmeet Kohli, and Demis Hassabis. Highly accurate protein structure prediction with AlphaFold. Nature, 596(7873):583-589, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.219, + 0.892, + 0.287 + ], + "angle": 0, + "content": "[35] Geewook Kim, Teakgyu Hong, Moonbin Yim, JeongYeon Nam, Jinyoung Park, Jinyeong Yim, Wonseok Hwang, Sang-doo Yun, Dongyoon Han, and Seunghyun Park. OCR-Free document understanding transformer. In ECCV, pages 498- 517,2022.2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.29, + 0.892, + 0.331 + ], + "angle": 0, + "content": "[36] Wonjae Kim, Bokyung Son, and Ildoo Kim. ViLT: Vision-and-language transformer without convolution or region supervision. In ICML, pages 5583-5594. PMLR, 2021. 5, 17" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.333, + 0.892, + 0.387 + ], + "angle": 0, + "content": "[37] Alexander Kolesnikov, André Susano Pinto, Lucas Beyer, Xiaohua Zhai, Jeremiah Harmsen, and Neil Houlsby. UViM: A unified modeling approach for vision with learned guiding codes. In NeurIPS, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.39, + 0.892, + 0.431 + ], + "angle": 0, + "content": "[38] Simon Kornblith, Mohammad Norouzi, Honglak Lee, and Geoffrey Hinton. Similarity of neural network representations revisited. In ICML, pages 3519-3529, 2019. 22" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.433, + 0.892, + 0.474 + ], + "angle": 0, + "content": "[39] Taku Kudo and John Richardson. SentencePiece: A simple and language independent subword tokenizer and tokenizer for neural text processing. In EMNLP, 2018. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.476, + 0.892, + 0.531 + ], + "angle": 0, + "content": "[40] Juho Lee, Yoonho Lee, Jungtaek Kim, Adam R. Kosiorek, Seungjin Choi, and Yee Whye Teh. Set transformer: A framework for attention-based permutation-invariant neural networks. In ICML, pages 3744-3753, 2019. 3, 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.533, + 0.892, + 0.601 + ], + "angle": 0, + "content": "[41] Kenton Lee, Mandar Joshi, Iulia Turc, Hexiang Hu, Fangyu Liu, Julian Eisenschlos, Urvashi Khandelwal, Peter Shaw, Ming-Wei Chang, and Kristina Toutanova. Pix2struct: Screenshot parsing as pretraining for visual language understanding. CoRR, abs/2210.03347, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.604, + 0.892, + 0.658 + ], + "angle": 0, + "content": "[42] Yoann Lemesle, Masataka Sawayama, Guillermo Valle-Perez, Maxime Adolphe, Hélène Sauzeon, and Pierre-Yves Oudeyer. Language-biased image classification: evaluation based on semantic representations. In ICLR, 2022. 7, 21" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.661, + 0.892, + 0.715 + ], + "angle": 0, + "content": "[43] Liunian Harold Li, Mark Yatskar, Da Yin, Cho-Jui Hsieh, and Kai-Wei Chang. VisualBERT: A simple and performant baseline for vision and language. CoRR, abs/1908.03557, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.718, + 0.892, + 0.773 + ], + "angle": 0, + "content": "[44] Qing Li, Boqing Gong, Yin Cui, Dan Kondratyuk, Xianzhi Du, Ming-Hsuan Yang, and Matthew Brown. Towards a unified foundation model: Jointly pre-training transformers on unpaired images and text. CoRR, abs/2112.07074, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.775, + 0.892, + 0.829 + ], + "angle": 0, + "content": "[45] Weixin Liang, Yuhui Zhang, Yongchan Kwon, Serena Yeung, and James Zou. Mind the gap: Understanding the modality gap in multi-modal contrastive representation learning. In NeurIPS, 2022. 7, 8, 22, 23" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.832, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[46] Valerii Likhosherstov, Anurag Arnab, Krzysztof Choromanski, Mario Lucic, Yi Tay, Adrian Weller, and Mostafa Dehghani. Polyvit: Co-training vision transformers on images, videos and audio. Trans. Machine Learning Research, 2023. 1, 2, 3" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "11015" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.133 + ], + "angle": 0, + "content": "[47] Lajanugen Logeswaran and Honglak Lee. An efficient framework for learning sentence representations. In ICLR, 2018. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.136, + 0.47, + 0.19 + ], + "angle": 0, + "content": "[48] Jiasen Lu, Dhruv Batra, Devi Parikh, and Stefan Lee. ViLBERT: Pretraining task-agnostic visiolinguistic representations for vision-and-language tasks. In NeurIPS, pages 13–23, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.193, + 0.469, + 0.247 + ], + "angle": 0, + "content": "[49] Elman Mansimov, Mitchell Stern, Mia Xu Chen, Orhan Firat, Jakob Uszkoreit, and Puneet Jain. Towards end-to-end in-image neural machine translation. CoRR, abs/2010.10648, 2020. 1, 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.25, + 0.469, + 0.291 + ], + "angle": 0, + "content": "[50] Joanna Materzyńska, Antonio Torralba, and David Bau. Disentangling visual and written concepts in CLIP. In CVPR, pages 16410-16419, 2022. 7, 21" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.293, + 0.469, + 0.333 + ], + "angle": 0, + "content": "[51] Joanna Materzynska, Antonio Torralba, and David Bau. Disentangling visual and written concepts in clip. In CVPR, 2022. 21, 22" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.336, + 0.469, + 0.417 + ], + "angle": 0, + "content": "[52] Sabrina J. Mielke, Zaid Alyafeai, Elizabeth Salesky, Colin Raffel, Manan Dey, Matthias Galle, Arun Raja, Chenglei Si, Wilson Y. Lee, Benoit Sagot, and Samson Tan. Between words and characters: A brief history of open-vocabulary modeling and tokenization in NLP. CoRR, abs/2112.10508, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.42, + 0.469, + 0.475 + ], + "angle": 0, + "content": "[53] Basil Mustafa, Carlos Riquelme, Joan Puigcerver, Rodolphe Jenatton, and Neil Houlsby. Multimodal contrastive learning with LIMoE: the language-image mixture of experts. In NeurIPS, 2022. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.478, + 0.469, + 0.532 + ], + "angle": 0, + "content": "[54] Daisuke Niizumi, Daiki Takeuchi, Yasunori Ohishi, Noboru Harada, and Kunio Kashino. Masked spectrogram modeling using masked autoencoders for learning general-purpose audio representation. CoRR, abs/2204.12260, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.534, + 0.469, + 0.602 + ], + "angle": 0, + "content": "[55] Hieu Pham, Zihang Dai, Golnaz Ghiasi, Kenji Kawaguchi, Hanxiao Liu, Adams Wei Yu, Jiahui Yu, Yi-Ting Chen, Minh-Thang Luong, Yonghui Wu, Mingxing Tan, and Quoc V. Le. Combined scaling for open-vocabulary image classification. CoRR, abs/2111.10050, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.605, + 0.469, + 0.687 + ], + "angle": 0, + "content": "[56] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In ICML, 2021. 1, 2, 3, 4, 6, 14, 17" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.689, + 0.469, + 0.757 + ], + "angle": 0, + "content": "[57] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J. Liu. Exploring the limits of transfer learning with a unified text-to-text transformer. J. Mach. Learn. Res., 21:140:1-140:67, 2020. 3, 4, 6, 20" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.76, + 0.469, + 0.814 + ], + "angle": 0, + "content": "[58] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In ICML, pages 8821-8831, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.817, + 0.469, + 0.858 + ], + "angle": 0, + "content": "[59] Nils Rethmeier and Isabelle Augenstein. A primer on contrastive pretraining in language processing: Methods, lessons learned and perspectives. ACM Computing Surveys, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.86, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[60] Phillip Rust, Jonas F. Lotz, Emanuele Bugliarello, Elizabeth Salesky, Miryam de Lhoneux, and Desmond Elliott. Language modelling with pixels. In ICLR, 2023. 1, 2, 3, 7, 18" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.147 + ], + "angle": 0, + "content": "[61] Phillip Rust, Jonas Pfeiffer, Ivan Vulic, Sebastian Ruder, and Iryna Gurevych. How good is your tokenizer? on the monolingual performance of multilingual language models. In ACL/IJCNLP, pages 3118-3135, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.15, + 0.892, + 0.245 + ], + "angle": 0, + "content": "[62] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily Denton, Seyed Kamyar Seyed Ghasemipour, Burcu Karagol Ayan, S. Sara Mahdavi, Rapha Gontijo Lopes, Tim Salimans, Jonathan Ho, David J. Fleet, and Mohammad Norouzi. Photorealistic text-to-image diffusion models with deep language understanding. In NeurIPS, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.249, + 0.892, + 0.289 + ], + "angle": 0, + "content": "[63] Elizabeth Salesky, David Etter, and Matt Post. Robust open-vocabulary translation from visual text representations. In EMNLP, pages 7235-7252, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.292, + 0.892, + 0.359 + ], + "angle": 0, + "content": "[64] Christoph Schuhmann, Richard Vencu, Romain Beaumont, Robert Kaczmarczyk, Clayton Mullis, Aarush Katta, Theo Coombes, Jenia Jitsev, and Aran Komatsuzaki. LAION-400M: Open dataset of CLIP-filtered 400 million image-text pairs. CoRR, abs/2111.02114, 2021. 4, 15" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.362, + 0.892, + 0.402 + ], + "angle": 0, + "content": "[65] Rico Sennrich, Barry Haddow, and Alexandra Birch. Neural machine translation of rare words with subword units. In ACL, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.405, + 0.892, + 0.445 + ], + "angle": 0, + "content": "[66] Noam Shazeer and Mitchell Stern. Adafactor: Adaptive learning rates with sublinear memory cost. In ICML, 2018. 3, 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.448, + 0.892, + 0.502 + ], + "angle": 0, + "content": "[67] Sheng Shen, Liunian Harold Li, Hao Tan, Mohit Bansal, Anna Rohrbach, Kai-Wei Chang, Zhewei Yao, and Kurt Keutzer. How much can CLIP benefit vision-and-language tasks? In ICLR, 2022. 2, 3, 5, 17" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.505, + 0.892, + 0.572 + ], + "angle": 0, + "content": "[68] Baohua Sun, Lin Yang, Catherine Chi, Wenhan Zhang, and Michael Lin. Squared english word: A method of generating glyph to use super characters for sentiment analysis. In Workshop on Affective Content Analysis, volume 2328, pages 140-151, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.576, + 0.892, + 0.615 + ], + "angle": 0, + "content": "[69] Zineng Tang, Jaemin Cho, Yixin Nie, and Mohit Bansal. TVLT: Textless vision-language transformer. In NeurIPS, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.619, + 0.892, + 0.658 + ], + "angle": 0, + "content": "[70] Ashish Thapliyal, Jordi Pont-Tuset, Xi Chen, and Radu Soricut. Crossmodal-3600: A Massively Multilingual Multimodal Evaluation Dataset. In EMNLP, 2022. 4, 6, 19" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.662, + 0.892, + 0.716 + ], + "angle": 0, + "content": "[71] Bart Thomee, David A. Shamma, Gerald Friedland, Benjamin Elizalde, Karl Ni, Douglas Poland, Damian Borth, and Li-Jia Li. YFCC100M: the new data in multimedia research. Commun. ACM, 59(2):64-73, 2016. 4, 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.719, + 0.892, + 0.786 + ], + "angle": 0, + "content": "[72] Ruben Villegas, Mohammad Babaeizadeh, Pieter-Jan Kindermans, Hernan Moraldo, Han Zhang, Mohammad Taghi Saffar, Santiago Castro, Julius Kunze, and Dumitru Erhan. Phenaki: Variable length video generation from open domain textual description. In ICLR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.789, + 0.892, + 0.843 + ], + "angle": 0, + "content": "[73] Alex Wang, Amanpreet Singh, Julian Michael, Felix Hill, Omer Levy, and Samuel R. Bowman. GLUE: A multi-task benchmark and analysis platform for natural language understanding. In ICLR, 2019. 2, 3, 4, 7, 18" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.846, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[74] Jianfeng Wang, Zhengyuan Yang, Xiaowei Hu, Linjie Li, Kevin Lin, Zhe Gan, Zicheng Liu, Ce Liu, and Lijuan Wang. GIT: A generative image-to-text transformer for vision and language. Trans. Machine Learning Research, 2022. 1, 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "11016" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.468, + 0.161 + ], + "angle": 0, + "content": "[75] Wenhui Wang, Hangbo Bao, Li Dong, Johan Bjorck, Zhiliang Peng, Qiang Liu, Kriti Aggarwal, Owais Khan Mohammed, Saksham Singhal, Subhojit Som, and Furu Wei. Image as a foreign language: BEiT pretraining for all vision and vision-language tasks. CoRR, abs/2208.10442, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.163, + 0.468, + 0.217 + ], + "angle": 0, + "content": "[76] Zirui Wang, Jiahui Yu, Adams Wei Yu, Zihang Dai, Yulia Tsvetkov, and Yuan Cao. SimVLM: Simple visual language model pretraining with weak supervision. In ICLR, 2022. 1, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.219, + 0.468, + 0.261 + ], + "angle": 0, + "content": "[77] Zhuofeng Wu, Sinong Wang, Jiatao Gu, Madian Khabsa, Fei Sun, and Hao Ma. CLEAR: contrastive learning for sentence representation. CoRR, abs/2012.15466, 2020. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.263, + 0.468, + 0.33 + ], + "angle": 0, + "content": "[78] Linting Xue, Aditya Barua, Noah Constant, Rami Al-Rfou, Sharan Narang, Mihir Kale, Adam Roberts, and Colin Raffel. Byt5: Towards a token-free future with pre-trained byte-to-byte models. Trans. Assoc. Comput. Linguistics, 10:291-306, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.333, + 0.468, + 0.387 + ], + "angle": 0, + "content": "[79] Linting Xue, Noah Constant, Adam Roberts, Mihir Kale, Rami Al-Rfou, Aditya Siddhant, Aditya Barua, and Colin Raffel. mT5: A massively multilingual pre-trained text-to-text transformer. In NAACL-HLT, 2021. 6, 20" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.39, + 0.468, + 0.457 + ], + "angle": 0, + "content": "[80] Haoxuan You, Luowei Zhou, Bin Xiao, Noel Codella, Yu Cheng, Ruochen Xu, Shih-Fu Chang, and Lu Yuan. Learning visual representation from modality-shared contrastive language-image pre-training. In ECCV, pages 69-87, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.46, + 0.468, + 0.514 + ], + "angle": 0, + "content": "[81] Peter Young, Alice Lai, Micah Hodosh, and Julia Hockenmaier. From image descriptions to visual denotations: New similarity metrics for semantic inference over event descriptions. Trans. Assoc. Comput. Linguistics, 2:67-78, 2014. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.517, + 0.468, + 0.572 + ], + "angle": 0, + "content": "[82] Jiahui Yu, Zirui Wang, Vijay Vasudevan, Legg Yeung, Mojtaba Seyedhosseini, and Yonghui Wu. Coca: Contrastive captioners are image-text foundation models. Trans. Machine Learning Research, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.574, + 0.468, + 0.668 + ], + "angle": 0, + "content": "[83] Jiahui Yu, Yuanzhong Xu, Jing Yu Koh, Thang Luong, Gunjan Baid, Zirui Wang, Vijay Vasudevan, Alexander Ku, Yinfei Yang, Burcu Karagol Ayan, Ben Hutchinson, Wei Han, Zarana Parekh, Xin Li, Han Zhang, Jason Baldridge, and Yonghui Wu. Scaling autoregressive models for content-rich text-to-image generation. Trans. Machine Learning Research, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.672, + 0.468, + 0.713 + ], + "angle": 0, + "content": "[84] Zhou Yu, Jun Yu, Yuhao Cui, Dacheng Tao, and Qi Tian. Deep modular co-attention networks for visual question answering. In CVPR, pages 6281–6290, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.715, + 0.468, + 0.769 + ], + "angle": 0, + "content": "[85] Xiaohua Zhai, Xiao Wang, Basil Mustafa, Andreas Steiner, Daniel Keysers, Alexander Kolesnikov, and Lucas Beyer. LiT: Zero-shot transfer with locked-image text tuning. In CVPR, pages 18102-18112, 2022. 2, 4, 5, 6, 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.771, + 0.468, + 0.84 + ], + "angle": 0, + "content": "[86] Yukun Zhu, Ryan Kiros, Richard S. Zemel, Ruslan Salakhutdinov, Raquel Urtasun, Antonio Torralba, and Sanja Fidler. Aligning books and movies: Towards story-like visual explanations by watching movies and reading books. In ICCV, pages 19-27, 2015. 7, 18" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.468, + 0.84 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.956 + ], + "angle": 0, + "content": "11017" + } + ] +] \ No newline at end of file diff --git a/2023/CLIPPO_ Image-and-Language Understanding From Pixels Only/6a9a5c21-30d9-4b55-b8d0-e491b4946877_origin.pdf b/2023/CLIPPO_ Image-and-Language Understanding From Pixels Only/6a9a5c21-30d9-4b55-b8d0-e491b4946877_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..b2ad1df84eaebc398ea7be58929fe3524842195e --- /dev/null +++ b/2023/CLIPPO_ Image-and-Language Understanding From Pixels Only/6a9a5c21-30d9-4b55-b8d0-e491b4946877_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e76e8c2c08be4b7d40102351a10b9f4ab1b7cf7d88a4aacd009ce2cc8b5f52f1 +size 3073653 diff --git a/2023/CLIPPO_ Image-and-Language Understanding From Pixels Only/full.md b/2023/CLIPPO_ Image-and-Language Understanding From Pixels Only/full.md new file mode 100644 index 0000000000000000000000000000000000000000..c80c8de36bc3b9b38ffdfab0cdddc30d4c29c7c3 --- /dev/null +++ b/2023/CLIPPO_ Image-and-Language Understanding From Pixels Only/full.md @@ -0,0 +1,290 @@ +# CLIPPO: Image-and-Language Understanding from Pixels Only + +Michael Tschannen, Basil Mustafa, Neil Houlsby +Google Research, Brain Team, Zürich + +# Abstract + +Multimodal models are becoming increasingly effective, in part due to unified components, such as the Transformer architecture. However, multimodal models still often consist of many task- and modality-specific pieces and training procedures. For example, CLIP (Radford et al., 2021) trains independent text and image towers via a contrastive loss. We explore an additional unification: the use of a pure pixel-based model to perform image, text, and multimodal tasks. Our model is trained with contrastive loss alone, so we call it CLIP-Pixels Only (CLIPPO). CLIPPO uses a single encoder that processes both regular images and text rendered as images. CLIPPO performs image-based tasks such as retrieval and zero-shot image classification almost as well as CLIP-style models, with half the number of parameters and no text-specific tower or embedding. When trained jointly via image-text contrastive learning and next-sentence contrastive learning, CLIPPO can perform well on natural language understanding tasks, without any word-level loss (language modelling or masked language modelling), outperforming pixel-based prior work. Surprisingly, CLIPPO can obtain good accuracy in visual question answering, simply by rendering the question and image together. Finally, we exploit the fact that CLIPPO does not require a tokenizer to show that it can achieve strong performance on multilingual multimodal retrieval without modifications. + +# 1. Introduction + +In recent years, large-scale multimodal training of Transformer-based models has led to improvements in the state-of-the-art in different domains including vision [2, 10, 74-76], language [6, 11], and audio [5]. In particular, in computer vision and image-language understanding, a single large pretrained model can outperform task-specific expert models [10, 74, 75]. However, large multimodal models often use modality or dataset-specific encoders and decoders, and accordingly lead to involved protocols. For example, such models frequently involve training different + +![](images/ffdab49a0a316c2db9fc0e440b859ee63037c0ec97ed700f60899518134c1a3e.jpg) +Figure 1. CLIP [56] trains separate image and text encoders, each with a modality-specific preprocessing and embedding, on image/alt-text pairs with a contrastive objective. CLIPPO trains a pure pixel-based model with equivalent capabilities by rendering the alt-text as an image, encoding the resulting image pair using a shared vision encoder (in two separate forward passes), and applying same training objective as CLIP. + +![](images/cbb943b06baad3ca5e62905b787b868708c0afc9f1dc74e14944d5e5690f8131.jpg) + +parts of the model in separate phases on their respective datasets, with dataset-specific preprocessing, or transferring different parts in a task-specific manner [75]. Such modality and task-specific components can lead to additional engineering complexity, and poses challenges when introducing new pretraining losses or downstream tasks. Developing a single end-to-end model that can process any modality, or combination of modalities, would be a valuable step for multimodal learning. Here, we focus on images and text. + +A number of key unifications have accelerated the progress of multimodal learning. First, the Transformer architecture has been shown to work as a universal backbone, performing well on text [6, 15], vision [16], audio [5, 24, 54], and other domains [7, 34]. Second, many papers have explored mapping different modalities into a single shared embedding space to simplify the input/output interface [21, 22, 46, 69], or develop a single interface to many tasks [31, 37]. Third, alternative representations of modalities allow harnessing in one domain neural architectures or training procedures designed for another domain [28, 49, 54, 60]. For example, [60] and [28, 54] represent text and audio, respectively, by rendering these modalities as images (via a spectogram in the case of audio). + +In this paper, we explore the use of a pure pixel-based + +model for multimodal learning of text and images. Our model is a single Vision Transformer [16] that processes visual input, or text, or both together, all rendered as RGB images. The same model parameters are used for all modalities, including low-level feature processing; that is, there are no modality-specific initial convolutions, tokenization algorithms, or input embedding tables. We train our model using only a single task: contrastive learning, as popularized by CLIP [56] and ALIGN [32]. We therefore call our model CLIP-Pixels Only (CLIPPO). + +We find that CLIPPO performs similarly to CLIP-style models (within $1 - 2\%$ ) on the main tasks CLIP was designed for—image classification and text/image retrieval—despite not having modality-specific towers. Surprisingly, CLIPPO can perform complex language understanding tasks to a decent level without any left-to-right language modelling, masked language modelling, or explicit word-level losses. In particular, on the GLUE benchmark [73] CLIPPO outperforms classic NLP baselines, such as ELMO+BiLSTM+attention, outperforms prior pixel-based masked language models [60], and approaches the score of BERT [15]. Interestingly, CLIPPO obtains good performance on VQA when simply rendering the image and text together, despite never having been pretrained on such data. + +Pixel-based models have an immediate advantage over regular language models because they do not require predetermining the vocabulary/tokenizer and navigating the corresponding intricate trade-offs; consequently, we observe improved performance on multilingual retrieval compared to an equivalent model that uses a classical tokenizer. + +# 2. Related work + +Multimodal and contrastive pretraining Most closely related to CLIPPO are CLIP [67] and ALIGN [32] which developed the paradigm of large-scale contrastive training on noisy data from the web. Follow-ups [55,85] have scaled further and employed state-of-the-art image representation learning to boost performance. + +A number of works have explored model unification via weight-sharing. In the contrastive context, LIMoE [53] and MS-CLIP [80] explore a one-tower model similar to ours, studying the use of mixture of experts and selective sharing of modules, respectively. Outside contrastive training, co-training distinct tasks [1, 46] is a popular strategy, with some approaches [44] involving knowledge distillation and gradient masking. Other works use self-supervised learning algorithms to unify task training [21]. These broadly use discriminative tasks to learn representations for various downstream modalities; generative approaches to multimodal modelling have been scaled to billions of parameters, generating text [2, 10, 74, 82], images [58, 62, 83], videos [27, 72] or audio [5] from various modalities. + +Another related domain is document and user interface (UI) understanding. Corresponding models are trained on diverse multimodal data sets and can usually solve a range of document/Ui understanding tasks. Many models rely on text extracted using an off-the-shelf OCR pipeline in combination with document images [3, 29], but image-only models are getting more popular [35, 41]. While these models can understand visual cues and text from the input image, they still rely on a tokenized text for training and inference. + +Contrastive training in NLP There is a sizable body of work on contrastive pretraining on sentence pairs (see [59] for a recent survey), which we explore as an auxiliary objective for CLIPPO. Popular augmentations to generate text pairs involve word deletion, span deletion, reordering, synonym substitution, and next-sentence-prediction [20, 47, 77]. Other methods use different realizations of dropout masks in the model to emulate sentence pairs, or supervised labels to obtain positive and negative pairs [19]. + +Visual text and tokenization in NLP The most closely related method to CLIPPO from the NLP domain is PILIX [60], which is a masked autoencoder (MAE) [26] trained on rendered text. It obtains strong performance on multilingual syntactic (part-of-speech tagging, dependency parsing) and semantic language understanding (named entity recognition, sentence understanding) tasks, while being more robust to noise in the text than BERT. Other applications for which visual text has been explored include sentiment analysis [68] and machine translation [49, 63]. + +Visual text side-steps the design and construction of an appropriate tokenizer, which is a large area of research of its own, and can hence simplify text processing in certain—in particular multilingual—scenarios. We refer to [52] for a survey on tokenizers. Popular models include WordPiece [15], Byte-Pair Encoding [65], and SentencePiece [39]. + +Subword-based vocabularies are popular in monolingual setups and usually lead to a good performance trade-off compared to word and character based vocabularies for certain languages including English. In multilingual contexts, appropriately representing the vocabulary of all languages becomes challenging as the number of languages increases [13,61], which in turn can lead to poor performance in tasks involving underrepresented languages. A variety of mitigation strategies has been developed; we refer to [60, Sec. 5.1] for a more detailed discussion of these strategies. + +# 3. Contrastive language-image pretraining with pixels + +Contrastive language-image pretraining has emerged as a powerful, scalable paradigm to train versatile vision models on web-scale data sets [56]. Concretely, this approach relies on image/alt-text pairs which can be automatically collected at large scale from the web. Thereby, the textual + +descriptions are usually noisy, and can e.g. consist of single keywords, sets of keywords, or potentially lengthy descriptions with many attributes describing the image content. Using this data, two encoders are jointly trained, namely a text encoder embedding the alt-texts and an image encoder embedding the corresponding images into a shared latent space. These two encoders are trained with a contrastive loss, encouraging the embeddings of matching images and alt-text to be similar, and at the same time to be dissimilar from all other image and alt-text embeddings. + +Once trained, such an encoder pair can be used in many ways: It can be specialized to classifying a fixed set of visual concepts via their textual descriptions (zero-shot classification); the embeddings can be used to retrieve images given a textual description and vice-versa; or the vision encoder can be transferred in supervised fashion to a downstream task by fine-tuning on a labeled data set or by training a head on top of the frozen image encoder representation. In principle, the text encoder can be used as a standalone text embedding, but this application—to our knowledge—has not been explored in-depth, with some authors citing the low quality of the alt-texts leading to weak language modeling performance of the text encoder [67]. + +Previous works [46, 53] have shown that the image and text encoder can be realized with a single shared transformer model (henceforth referred to as single tower model, or 1T-CLIP), where the images are embedded using a patch embedding, and the tokenized text is embedded using a separate word embedding. Apart from the modality-specific embeddings, all model parameters are shared for the two modalities. While this type of sharing usually leads to a minor performance drop on image/image-language tasks it also halves the number of model parameters. + +CLIPPO takes this idea one step further: text inputs are rendered on blank images, and are subsequently dealt with entirely as images, including the initial patch embedding (see Fig. 1 for an illustration). By training this single vision transformer contrastively as prior works, we obtain a single vision transformer model that can understand both images and text through the single interface of vision and provides a single representation which can be used to solve image, image-language, and pure language understanding tasks. + +Alongside multimodal versatility, CLIPPO alleviates common hurdles with text processing, namely the development of an appropriate tokenizer and vocabulary. This is particularly interesting in a massively multilingual setup, where the text encoder has to handle dozens of languages. + +We find that CLIPPO trained on image/alt-text pairs performs comparably with its 1T-CLIP counterpart on common image and image-language benchmarks, and is competitive with strong baseline language models on the GLUE benchmark [73]. However, due to the low quality of the alt-texts which are often not grammatical sentences, learn + +ing language understanding exclusively from alt-texts is fundamentally limited. Therefore, we augment image/alt-text contrastive pretraining with language-based contrastive training. Specifically, we use positive pairs of consecutive sentences sampled from a text corpus which is seamlessly integrated into the contrastive training by supplementing batches of image/alt-texts with (rendered) text/text pairs. + +# 4. Experiments + +# 4.1. Training details and models + +We rely on a single training setup for all our baselines and visual text models. This setup was tuned to produce good results for standard image/alt-text contrastive training as in [56] (using exactly the same loss function as [56], following the pseudocode in [56, Fig. 3]) and we found that it readily transfers to 1T-CLIP and CLIPPO (including variants with text/text co-training). + +Our default architecture is a ViT-B/16 [16] and we perform a subset of experiments with a ViT-L/16 architecture to study the effect of scale (we equip both models a MAP head [40] to pool embeddings). In all cases, the representation dimension used for the contrastive loss is 768. We set the batch size to 10,240 and train the main models for 250k steps, using a minimum 100k training steps for ablations. For models co-trained with a certain percentage of text/text data, we scale the number of iterations such that the number of image/alt-text pairs seen matches the number of iterations of the corresponding model without text/text data (e.g. when $50\%$ of the data is text/text pairs we increase the number of iterations from 250k to 500k). The contrastive loss is computed across the full batch. We use the Adafactor optimizer [66] with a learning rate of $10^{-3}$ and decoupled weight decay with weight $10^{-4}$ . + +Baseline CLIP-style models are trained using the T5-en SentencePiece tokenizer [57]; we use the abbreviation $\mathrm{CLIP^{*}}$ for the two tower model from [56] trained from scratch using the setup described above, to avoid confusion with the model released by [56]. A sequence length of 196 is used, as this matches the number of visual text "tokens" CLIPPO can process with patch size 16 has at $224\mathrm{px}$ resolution (which we use throughout unless noted otherwise). + +Visual text For visual text rendering [60, 63] relied on the Google Noto font family1 which supports the majority of Unicode code points. Here, we use the GNU Unifont bitmap font2, which has a similar coverage but allows for efficient, lookup-based on-the-fly rendering in our preprocessing pipeline. We emphasize that this rendering strategy does not slow down training compared to tokenizer-based models. In preliminary explorations, we found this to be performance-neutral when compared to the Noto font. + +
#param.training datasetI1k 10s.I1k 0s.C I→TC T→IF I→TF T→I
CLIP*203MWebLI55.865.148.531.379.259.4
1T-CLIP118MWebLI53.962.348.030.377.558.2
CLIPPO93MWebLI53.061.447.330.176.457.3
CLIPPO93MWebLI + 25%C452.157.440.726.768.951.8
CLIPPO93MWebLI + 50%C448.053.135.223.464.847.2
1T-CLIP L/16349MWebLI60.867.850.732.581.061.0
CLIPPO L/16316MWebLI60.367.450.633.479.262.6
CLIPPO L/16316MWebLI + 25%C460.566.044.529.872.957.3
CLIPPO L/16316MWebLI + 50%C456.861.739.727.370.154.7
+ +Table 1. Vision and vision-language cross-modal results. We report ImageNet-1k 10-shot linear transfer validation accuracy (I1k 10s.), ImageNet-1k zero-shot transfer validation accuracy (I1k 0s.), image-to-text and text-to-image retrieval recall@1 on MS-COCO (C I→T and C T→I) and on Flickr30k (F T→I and F I→T). CLIPPO and 1T-CLIP incur a minor drop in these evaluations compared to CLIP*, while only using about half of the model parameters. Co-training with text pairs from C4 (models with + xx%C4) degrades performance on some cross-modal tasks (but leads to improved language understanding capabilities, see Table 2). + +Image/alt-text data We use the WebLI data set introduced in [10] which comprises 10 billion images with 12 billion corresponding alt-texts. Importantly, WebLI comprises alt-texts in 109 languages (unlike previous data sets such as LAION-400M [64] which only contain English alt-texts) and it is therefore a great foundation to study multilingual language-image pretraining and its applications. Please refer to [10, Fig. 3] for details on the alt-text language distribution. For English-only models we obtain English versions of non-English alt-texts via GCP Translation $\mathrm{API}^3$ . In addition to alt-text, WebLI also provides OCR annotations, which we do not use in this paper. Finally, WebLI was processed with a de-duplication step removing all images from various splits of the image evaluation sets used in this paper. Please refer to [10, Sec. 3.2] for more details on the WebLI data set and to [10, Appendix B] for a datasheet. + +We also present a subset of results based on LAION-400M [64] and YFCC-100M [71] as an additional comparison points, see Appendix C.1 and C.2, respectively. + +Text/text data For co-training with text/text pairs we primarily rely on the publicly available Colossal Clean Crawled Corpus (C4; default/English split) [57]. We randomly sample pairs of consecutive sentences and contrastively train on these pairs, i.e., the model is trained for embedding-based next sentence prediction (NSP) [47]. We also experiment with pairs of parallel sentences in different languages from the WMT19 data set [18] as well as back-translated English sentences derived from C4 following the strategy described in [12]. + +# 4.2. Evaluations and metrics + +To evaluate the vision and vision/language understanding capabilities of our models we use standard metrics from the literature [53, 56, 85]: "zero-shot" transfer, which uses (embedded) textual description of the classes to be classified/retrieved and compares these with image embeddings. + +We report the classification accuracy on ImageNet-1k [14] as well as the recall@1 for cross-modal retrieval on MSCOCO [9] and Flickr30k [81]. Furthermore, we test the low-data transfer performance of the models by means of the linear adaptation protocol from [16], reporting the 10-shot accuracy on ImageNet-1k. + +We also evaluate CLIPPO and baselines on the popular VQA benchmark VQAv2 [25]. To construct a VQA model using a single pretrained ViT we render the question at the top end of the corresponding image (using the same Unifont renderer as used for CLIPPO training) and follow the standard prediction setup where the answer is predicted as the most likely answer from the training set, i.e. by classification. Specifically, we replace the last layer of our pretrained CLIPPO and baselines with a randomly initialized one with the appropriate number of outputs and fine-tune on VQAv2. This setup tests the ability of the pretrained ViT to combine image and text in intermediate layers as it has produce a single output from a fused image/text input image, unlike in the other cross-modal tasks (and pretraining), where image and text representations are computed with two separate forward passes. Please refer to Appendix A in the supplementary material for examples images with rendered questions and Appendix B.1 for details on the fine-tuning protocol. + +Multilingual capabilities are assessed via zero-shot retrieval on CrossModal3600 [70], which is a geographically diverse set comprising 3600 images each human-annotated with captions in 36 languages. The corresponding recall metric is averaged across all languages and images. + +Finally, we evaluate the language understanding capabilities on the General Language Understanding Evaluation (GLUE) benchmark [73] which comprises natural language inference tasks (MNLI, QNLI, RTE), a sentiment analysis task (SST-2), sentence similarity tasks (QQP, STS-B, MRPC), and a linguistic acceptability task (CoLA). Following common practice, we exclude the WNLI task from the benchmark [15, 77]. We transfer our baselines and CLIPPO models by attaching a 2-hidden layer MLP with 768 units to + +![](images/9b8eb1248493590b3de5a1a00f10c2f59eb466390defefe448d40e116c1814e1.jpg) +Figure 2. Results on the VQAv2 benchmark (test-dev set). In addition to CLIPPO and baselines produced in this work, we also compare to Pythia and MCAN models with ViT encoders from [67], and with comparably sized METER [17] and ViLT [36] models. CLIPPO outperforms $\mathrm{CLIP^{*}}$ and 1T-CLIP clearly on "yes/no" questions and gets similar performance as task-specific models. + +their representation and following precisely the fine-tuning protocol from BERT [15]. For sentence pair classification tasks we simply render both sentences on the same image, printing [SEP] to mark the start of the second sentence. + +# 4.3. Vision and vision-language understanding + +Image classification and retrieval Table 1 shows the performance of CLIPPO along with the baseline models on the benchmarks described in Sec. 4.2. It can be seen that the CLIPPO and 1T-CLIP incur a drop of a 2-3 percentage points absolute compared to $\mathrm{CLIP^{*}}$ . This is not surprising and can be attributed to the fact that single tower models only have about half the parameters count of a corresponding two tower model. The difference in performance between the English-only CLIPPO and 1T-CLIP is very small for a B/16 backbone at $100\mathrm{k}$ training steps (see Table 6 in the supplementary material), and vanishes with longer training and/or by increasing the model size, despite the fact that CLIPPO has $25\%$ and $10\%$ fewer parameters than 1T-CLIP for a B/16 and L/16 architecture, respectively (which is due to the absence of the text embedding in CLIPPO). + +The multilingual CLIPPO model performs somewhat worse than the corresponding 1T-CLIP, and the gap does not close completely when training longer (see Table 6). + +However, when evaluated across a broad set of languages on the CrossModal3600 CLIPPO performs on par with or slightly better than 1T-CLIP (see Sec. 4.4 below). + +As we add sentence pairs to the training mix the performance on the cross-modal retrieval metrics decreases. This is not surprising as we keep the total batch size constant so that the effective batch size of image/alt-text contrastive training decreases, which is known to impact performance [85]. Interestingly, the 10-shot transfer performance does not move in tandem, but only decreases significantly when half of the training data is sentence pairs. In exchange, co-training with text data leads to significantly improved language understanding performance (see Sec. 4.5). + +VQA In Fig. 2 we report the VQAv2 score of our models and baselines. It can be seen that CLIPPO outperforms $\mathrm{CLIP^{*}}$ , 1T-CLIP, as well as a pretrained ViT-B/16 from [16] by a significant margin, achieving a score of 66.3, and constraining with $25\%$ C4 data leads to a slight improvement of the score. The improved score of CLIPPO is mainly due to better performance in "yes/no" questions. Increasing the model size to L/16 adds another 2 points which originate from improvements in the "number" and "other" VQAv2 categories. However, note that for an L/16 architecture 1T-CLIP performs competitively with CLIPPO (see Table 7). One possible explanation for this could be that 1T-CLIP develops better OCR capabilities thanks to the higher model capacity (alt-texts can correlate with text in images/scene text, see [10, Fig. 3]). Increasing the resolution to 384px adds 2 to 3 points across models. + +We also compare CLIPPO with baselines from the literature. Specifically, [17] proposes a framework (called ME-TER) for multimodal tasks, where pretrained transformer-based image and text encoders are combined with a transformer-based fusion module. CLIPPO L/16 achieves performance competitive with their model combining a CLIP B/32 vision backbone with a BERT-Base language backbone, which is roughly comparable in size and computational cost with our L/16 models. Another related work is [67], which combines different CLIP vision backbones with two existing VQA systems, Pythia [33] and MCAN [84]. CLIPPO outperforms different CLIP ViT-based Pythia and MCAN models from [67]. Note, however, that ResNet-based CLIP backbones lead to better results when combined with these systems. We further note that both [17] and [67] also investigate training their models on a mix of different image-text data sets with multiple objectives such as grounded masked language modeling and text-image matching, before transferring to the VQA task, which leads to significant improvements. ViLT [36] relies on such a strategy to train a single transformer backbone jointly encoding image and text tokens. At 384px resolution, CLIPPO (with $25\%$ C4 data) obtains a VQA score comparable with that of ViLT (and other models from the + +literature such as ViLBERT [48], VisualBERT [43], and PixelBERT [30]), despite only using a contrastive objective for pretraining. + +# 4.4. Multilingual vision-language understanding + +For typical language models, tokenizer choice can be a challenging process [78]. Commonly used English-language tokenizers generalize poorly to non-latin scripts [85]. This can be alleviated by the use of larger, multilingual vocabularies, at the expense of very large parameter counts. CLIPPO bypasses this issue, removing any language-related bias stemming from unbalanced or restrictive tokenizers. We consider multilingual image/text retrieval on Crossmodal3600 and compare CLIPPO, trained on WebLI with multilingual alt-texts, against 1T-CLIP with a number of SentencePiece tokenizers; one trained from 300M WebLI multilingual alt-texts, English (T5-en) and multilingual (T5-all) tokenizers from T5 [57], and a multilingual tokenizer (mT5) from mT5 [79], all with a vocabulary size of 32,000. The results are shown in Fig. 4. On average, CLIPPO achieves comparable retrieval performance to these baselines. In the case of mT5, the use of extra data to create the specialized vocabulary can boosts performance above that of CLIPPO; the leveraging of such extra parameters and data in the multilingual context will be an interesting future direction for CLIPPO. + +![](images/c530045a6fa2378a022f1da0ae18ebf659dbe4382f0e14b00f1ed0bba3350633.jpg) +Figure 3. Tokenization efficiency analyzed in terms of the sequence length produced by a given method. CLIPPO produces smaller sequences for the majority of languages compared to 1T-CLIP with alternative tokenizers. + +Tokenization efficiency If a tokenizer is well suited to a particular dataset, it will tokenize to shorter sequences—this is especially the case when byte fallback [39] is enabled. SentencePiece tokenizers have the advantageous ability to tokenize entire—possibly quite long—words to single tokens. CLIPPO cannot learn any such compression, but benefits from equal treatment of all languages and + +words: it will by definition generalize equally well to all data, as its tokenization schema has not been trained on a specific dataset. We analyze 20,000 samples for each of the 104 C4 languages. Each CLIPPO token is assumed to be a $16 \times 16$ patch; though in typical computations all approaches considered here would pad to a fixed length, we compute CLIPPO's sequence length according to the last patch which contains rendered text. Fig. 3 shows the fraction of C4 languages where CLIPPO processes tokens + +![](images/aff7e6aca6b658507c2c2b72c506fc8c84f425347c964bc30c587143bb11692f.jpg) +Figure 4. Zero-shot image/text retrieval performance on CrossModal3600 [70]. Although specialized (mc4) tokenizers can be leveraged to improve multilingual performance CLIPPO (dashed black line) broadly matches or exceeds comparable 1T-CLIP models trained with vocabulary size 32,000 (the word embeddings result in a $27\%$ increase in parameter count compared to CLIPPO). + +![](images/7517525aa45ff1bdaec77d872d46a3f728890ca88b6860747cd07f33d286697b.jpg) + +![](images/ca91c6e84a119e7ebe86b0050fec9439ddfee621ba860c69c656b9b675a62c4e.jpg) + +more efficiently than the vocabularies discussed above. We conservatively define "more efficient" as producing a shorter token sequence for over $75\%$ of examples. Even so, CLIPPO is indeed more efficient across the majority of languages. Per-language breakdowns of multilingual retrieval performance and tokenization efficiency are further discussed in Appendix C.3. + +# 4.5. Language understanding + +Table 2 shows the GLUE benchmark results of CLIPPO and baselines. One can observe that CLIPPO trained on WebLI performs competitively with the BiLSTM+Attn+ELMo baseline which relies on deep word embeddings trained on a large language corpus. Also, it can be seen that CLIPPO along with 1T-CLIP outperform the language encoder trained using standard contrastive language vision pretraining $(\mathrm{CLIP}^{*})$ . This indicates that multimodal training in a single encoder benefits language understanding. Furthermore, CLIPPO achieves a much higher GLUE score than the $\mathrm{CLIP}^*$ image encoder, which in turn leads to significantly better results than fine-tuning a ViT-B/16 from scratch on GLUE (see Appendix C.2 for additional results). Unsurprisingly, the models pretrained on WebLI cannot do better than random guessing on the CoLA evaluation which requires to assess the grammatical correctness of sentences (recall that alt-texts are rarely grammatical sentences). Also the accuracy of $\mathrm{CLIP}^*$ and 1T-CLIP vision encoders we observe for SST-2 is in agreement with what was reported in [56, Table 10] for CLIP with a ViT-B/16 image encoder. + +Adding sentence pairs form the C4 corpus gradually improves the GLUE score, and when half of the examples are sentence pairs our model becomes competitive with PILXEL, while still retaining decent image and vision-language understanding capabilities (cf. Table 1). Note, however, that there is a trade-off between language-only tasks and tasks that involve image understanding. Finally, training CLIPPO only on sentence pairs leads to a model which outperforms PILXEL by a significant margin. However, our model has seen more sentence pairs than PILXEL, so PILXEL might improve as well when training longer. + +
training datasetMNLI-M/MMQQPQNLISST-2COLASTS-BMRPCRTEavg
BERT-BaseWiki + BC84.0 / 84.287.691.092.660.388.890.269.583.1
PIXELWiki + BC78.1 / 78.984.587.889.638.481.188.260.576.3
BiLSTM66.7 / 66.782.077.087.517.672.085.158.568.1
BiLSTM+Attn, ELMo72.4 / 72.483.675.291.544.156.182.152.770.0
CLIP* img enc.WebLI66.4 / 67.578.669.478.60.05.281.252.755.5
CLIP* text enc.WebLI71.8 / 72.582.773.086.26.665.081.453.865.9
1T-CLIP text enc.WebLI72.6 / 73.083.880.784.90.079.683.357.068.3
CLIPPOWebLI73.0 / 72.684.381.286.81.880.584.153.468.6
CLIPPOWebLI + 25%C477.7 / 77.285.383.190.928.283.484.559.274.4
CLIPPOWebLI + 50%C479.2 / 79.286.484.292.938.983.484.859.976.6
CLIPPOC479.9 / 80.286.785.293.350.984.786.358.578.4
CLIPPO L/16WebLI + 25%C476.6 / 75.587.179.993.248.284.184.656.076.1
CLIPPO L/16WebLI + 50%C482.3 / 82.487.986.794.255.385.885.959.280.0
+ +Table 2. Results for the GLUE benchmark (dev set). The metric is accuracy except for the performance on QQP and MRPC, which is measured using the $F_{1}$ score, CoLA which uses Matthew's correlation, and STS-B which evaluated based on Spearman's correlation coefficient. "avg" corresponds to the average across all metrics. The results for BERT-Base and PILXEL are from [60, Table 3], and BiLSTM and BiLSTM+Attn, ELMo from [73, Table 6]. All encoders considered here have a transformer architecture comparable to BERT-Base (up to the text embedding layer), except for CLIPPO L/16 which uses a ViT L/16, and the two BiLSTM model variants. Wiki and BC stand for (English) Wikipedia and Bookcorpus [86] data, respectively. + +# 4.6. Ablations and analysis + +Impact of weight sharing across modalities The fact that CLIPPO 1) uses a shared patch embedding for regular images and text images and 2) this embedding has considerably fewer parameters than the text embedding of 1T-CLIP and $\mathrm{CLIP^{*}}$ provokes the question of whether CLIPPO could benefit from separate patch embeddings for text images and regular images. Further, CLIPPO relies on a single head to compute the output representation for images and text, and relaxing this constraint by using separate heads for the two modalities could lead to more expressive representations. The results (deferred to Appendix D.1) show that neither of these variants lead to improved image classification or retrieval metrics compared to CLIPPO. + +Impact of the text location We test whether rendering the question at the top, middle, or bottom of the image impacts the VQA performance of CLIPPO and find that it does not, provided that we increase the learning rate of the positional embedding during fine-tuning (see Appendix D.2). + +Typographic attacks Since CLIPPO is trained on large amounts of rendered (alt-)text it is important to check whether it becomes more susceptible to typographic attacks—the tendency of CLIP-style models to zero-shot classify an image according to adversarially injected scene text unrelated to the scene [23,42,50]. In Appendix D.3 we present results indicating that CLIPPO is no more vulnerable to typographic attacks than 1T-CLIP and $\mathrm{CLIP^{*}}$ . + +Modality gap Liang et al. [45] discovered that text and image embeddings of CLIP-style models form two distinct clusters rather than both filling the embedding space densely and occupying the same spatial region. They attribute this phenomenon to a combination of initialization + +conditions and properties of the loss function/training dynamics. Since we consider single tower models here, and also co-train some of these models with text-only pairs it is interesting to see how this affects the modality gap. We compute the gap and visualize it following the recipe from [45] in Fig. 5 (see Appendices D.4 and D.5 for additional visualizations). CLIPPO attains a slightly lower modality gap than $\mathrm{CLIP}^*$ , but clearly features a clustering structure for image and text embeddings. However, when training contrastively with sentence pairs in addition to image/alt-text pairs, the clustering structure disappears, the image and text embeddings overlap, and the modality gap decreases significantly. A possible explanation for this behavior could be that the additional learning pressure induced by the contrastive loss on sentence pairs encourages text embeddings to spread out more and hence the structure of all embeddings changes. + +Text/text co-training objectives To corroborate that contrastive NSP is a sensible objective to improve language understanding in the context of CLIPPO, we train CLIPPO without any image/alt-text data on pairs of parallel translated sentences (this is straight-forward in our framework since visual text is language-agnostic), as well as English back-translated data, and evaluate the resulting text representations on GLUE. Table 3 shows that NSP on C4 clearly achieves the highest GLUE score. + +# 5. Discussion and limitations + +We proposed and evaluated CLIPPO which produces a single ViT that can understand images and language jointly using images as a sole input modality. Perhaps surprisingly, CLIPPO matches the performance of the 1T-CLIP baseline across many of the considered tasks, and only in + +
WMT19WMT19 BTC4 NSP
GLUE score61.266.677.6
+ +Table 3. Ablation of text pair-based contrastive co-training tasks: Training on parallel translated sentences (WMT19), training on parallel back-translated sentences (WMT19 BT), and NSP for sentences sampled from C4 (C4 NSP). C4 NSP leads to the highest GLUE score by a large margin. + +curs a minor drop compared to the $\mathrm{CLIP^{*}}$ baseline, despite having less than half the parameters of $\mathrm{CLIP^{*}}$ . As we showed, the image-only interface enables a simple, unified data pipeline for training on and transferring to mixed modalities. CLIPPO opens the door for additional modalities (e.g. spectrograms) and, as we hope, might inspire applications of pixel-only models beyond contrastive training. Nevertheless, several limitations remain, as discussed next. + +Co-training First, to achieve language understanding performance competitive with PILXEL and BERT on GLUE, contrastive co-training with text pairs is necessary. While adding $25\%$ C4 data to the batch seems to strike a good balance across all tasks considered, it does induce a nonnegligible drop in zero-shot image classification and image/text retrieval. This drop becomes more severe as the fraction of C4 examples increases. We observed an associated change in modality gap, and further investigation of the representation in the context of co-training might help to develop models that achieve better overall performance in the co-training setup. + +Diverse rendered text CLIPPO currently relies on cleanly rendered text as an input and its capabilities to handle text from documents or web pages without further adaption is limited (besides the basic OCR capabilities that CLIP-style models learn from image/alt-text pairs). We emphasize that sophisticated OCR and document understanding is not a goal of this paper, and training CLIPPO with augmented noisy rendered text that mimics the distribution of documents and websites is likely to lead to worse performance across the considered tasks, since image/alt-text pairs are less correlated and provide a weaker learning signal. However, developing CLIPPO further to handle less clean visual text will open many additional applications. + +Generative modeling CLIPPO, like CLIP, BERT, PILXEL and many other models, uses an encoder-only design and hence lacks the ability to generate text outputs. A common approach to equip encoder-only models with generation capabilities (e.g., for image captioning or VQA) is to simply combine them with a (potentially pretrained) language model [8, 76]. This approach naturally also applies to CLIPPO and PILXEL, but defeats the advantages of visual text in certain (e.g. multilingual) scenarios. While visual text outputs have previously been explored in the context of + +![](images/58db6c461caeb2536f7513e0b50b180c5916ddaba4a1c4a68466f01496090d94.jpg) +Figure 5. Visualization of the modality gap for $\mathrm{CLIP^{*}}$ and CLIPPO optionally trained with $25\%$ C4 data. The visualization follows the analysis from [45] and shows embedded images (blue dots) and corresponding alt-text (orange dots) from the WebLI validation set, projected to the first two principal components of the validation data matrix. CLIPPO has a slightly smaller modality gap than $\mathrm{CLIP^{*}}$ ; co-training with C4 data strongly reduces the gap. + +machine translation [49], it remains unclear what a scalable tokenizer-free way to generate text is. + +Multilingual learning Finally, we showed that CLIPPO obtains strong multilingual image/text retrieval performance without requiring the development of an appropriate tokenizer. For fine-grained adjustment and balancing of the retrieval performance further steps will be necessary, including data balancing and potentially co-training with multi-lingual text data. Furthermore, similar to PILXEL, CLIPPO relies on certain ad-hoc design choices w.r.t. the visual representation, for example the left-to-right rendering of Arabic scripts. This approach leads to decent performance on average, but it is not clear what kind of unwanted effects it introduces and how these could be mitigated. + +# 6. Conclusion + +We introduced CLIPPO, a model for processing images and text jointly through the lens of vision. This reduces design choices (in particular w.r.t. tokenization) and parameter count, simplifies data processing pipelines and transfer recipes, and increases generality across multiple languages. We also explored methods of enhancing language understanding, where traditional image/alt-text contrastive models trained on web data fall short. We demonstrated this is possible by co-training with text pairs, with CLIPPO models outperforming strong NLP baselines while maintaining solid image understanding capabilities. + +Although we presented a unified contrastive training algorithm, CLIPPO suffers somewhat when co-training on multiple tasks, and future work to harmonize the co-training could enhance the models significantly. Deeper understanding of the design choices in rendering text as images, and their impact on performance, is another interesting avenue. + +Acknowledgments We would like to thank Lucas Beyer, Josip Djolonga, Alexander Kolesnikov, Mario Lucic, Andreas Steiner, Xiao Wang, and Xiaohua Zhai for inspiring discussions and helpful feedback. We also thank Jeffrey Sorensen for help with the text rendering preprocessing. + +# References + +[1] Hassan Akbari, Liangzhe Yuan, Rui Qian, Wei-Hong Chuang, Shih-Fu Chang, Yin Cui, and Boqing Gong. VATT: transformers for multimodal self-supervised learning from raw video, audio and text. In NeurIPS, 2021. 2 +[2] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katie Millican, Malcolm Reynolds, Roman Ring, Eliza Rutherford, Serkan Cabi, Tengda Han, Zhitao Gong, Sina Samangooei, Marianne Monteiro, Jacob Menick, Sebastian Borgeaud, Andrew Brock, Aida Nematzadeh, Sahand Sharifzadeh, Mikolaj Binkowski, Ricardo Barreira, Oriol Vinyals, Andrew Zisserman, and Karen Simonyan. Flamingo: A visual language model for few-shot learning. In NeurIPS, 2022. 1, 2 +[3] Srikar Appalaraju, Bhavan Jasani, Bhargava Urala Kota, Yusheng Xie, and R. Manmatha. DocFormer: End-to-end transformer for document understanding. In ICCV, pages 973-983, 2021. 2 +[4] Lucas Beyer, Xiaohua Zhai, and Alexander Kolesnikov. Big Vision. https://github.com/google-research/big visions, 2022.1 +[5] Zalan Borsos, Raphael Marinier, Damien Vincent, Eugene Kharitonov, Olivier Pietquin, Matthew Sharifi, Olivier Teboul, David Grangier, Marco Tagliasacchi, and Neil Zeghidour. AudioLM: a language modeling approach to audio generation. CoRR, abs/2209.03143, 2022. 1, 2 +[6] Tom B. Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert-Voss, Gretchen Krueger, Tom Henighan, et al. Language models are few-shot learners. In NeurIPS, 2020. 1 +[7] Lili Chen, Kevin Lu, Aravind Rajeswaran, Kimin Lee, Aditya Grover, Michael Laskin, Pieter Abbeel, Aravind Srinivas, and Igor Mordatch. Decision transformer: Reinforcement learning via sequence modeling. In NeurIPS, pages 15084-15097, 2021. 1 +[8] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey E. Hinton. A simple framework for contrastive learning of visual representations. In ICML, 2020. 8 +[9] Xinlei Chen, Hao Fang, Tsung-Yi Lin, Ramakrishna Vedantam, Saurabh Gupta, Piotr Dólar, and C. Lawrence Zitnick. Microsoft COCO captions: Data collection and evaluation server. CoRR, abs/1504.00325, 2015. 4 +[10] Xi Chen, Xiao Wang, Soravit Changpinyo, A. J. Piergiovanni, Piotr Padlewski, Daniel Salz, Sebastian Goodman, Adam Grycner, Basil Mustafa, Lucas Beyer, Alexander Kolesnikov, Joan Puigcerver, Nan Ding, Keran Rong, Hassan Akbari, Gaurav Mishra, Linting Xue, Ashish Thapliyal, James Bradbury, Weicheng Kuo, Mojtaba Seyedhosseini, Chao Jia, Burcu Karagol Ayan, Carlos Riquelme, Andreas Steiner, Anelia Angelova, Xiaohua Zhai, Neil Houlsby, and Radu Soricut. PaLI: A jointly-scaled multilingual language-image model. In ICLR, 2023. 1, 2, 4, 5 +[11] Aakanksha Chowdhery, Sharan Narang, Jacob Devlin, Maarten Bosma, Gaurav Mishra, Adam Roberts, Paul + +Barham, Hyung Won Chung, Charles Sutton, Sebastian Gehrmann, Parker Schuh, Kensen Shi, Sasha Tsvyashchenko, Joshua Maynez, Abhishek Rao, Parker Barnes, Yi Tay, Noam Shazeer, Vinodkumar Prabhakaran, Emily Reif, Nan Du, Ben Hutchinson, Reiner Pope, James Bradbury, Jacob Austin, Michael Isard, Guy Gur-Ari, Pengcheng Yin, Toju Duke, Anselm Levskaya, Sanjay Ghemawat, Sunipa Dev, Henryk Michalewski, Xavier Garcia, Vedant Misra, Kevin Robinson, Liam Fedus, Denny Zhou, Daphne Ippolito, David Luan, Hyeontaek Lim, Barret Zoph, Alexander Spiridonov, Ryan Sepassi, David Dohan, Shivani Agrawal, Mark Omernick, Andrew M. Dai, Thanumalayan Sankaranarayana Pillai, Marie Pellat, Aitor Lewkowycz, Erica Moreira, Rewon Child, Oleksandr Polozov, Katherine Lee, Zongwei Zhou, Xuezhi Wang, Brennan Saeta, Mark Diaz, Orhan Firat, Michele Catasta, Jason Wei, Kathy Meier-Hellstern, Douglas Eck, Jeff Dean, Slav Petrov, and Noah Fiedel. PaLM: Scaling language modeling with pathways. CoRR, abs/2204.02311, 2022. 1 +[12] Geoffrey Cideron, Sertan Girgin, Anton Raichuk, Olivier Pietquin, Olivier Bachem, and Léonard Hussenot. vec2text with round-trip translations. CoRR, abs/2209.06792, 2022. 4 +[13] Alexis Conneau, Kartikay Khandelwal, Naman Goyal, Vishrav Chaudhary, Guillaume Wenzek, Francisco Guzmán, Edouard Grave, Myle Ott, Luke Zettlemoyer, and Veselin Stoyanov. Unsupervised cross-lingual representation learning at scale. In ACL, pages 8440–8451, 2020. 2 +[14] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In CVPR, 2009. 4 +[15] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. BERT: pre-training of deep bidirectional transformers for language understanding. In NAACL-HLT, 2019, 1, 2, 4, 5 +[16] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth $16 \times 16$ words: Transformers for image recognition at scale. In ICLR, 2021. 1, 2, 3, 4, 5, 14, 16, 17, 24 +[17] Zi-Yi Dou, Yichong Xu, Zhe Gan, Jianfeng Wang, Shuohang Wang, Lijuan Wang, Chenguang Zhu, Pengchuan Zhang, Lu Yuan, Nanyun Peng, Zicheng Liu, and Michael Zeng. An empirical study of training end-to-end vision-and-language transformers. In CVPR, pages 18145-18155, 2022. 5, 14, 17 +[18] Wikipedia Foundation. ACL 2019 Fourth Conference on Machine Translation (WMT19), Shared Task: Machine Translation of News. http://www.statmt.org/wmt19/translation-task.html.4 +[19] Tianyu Gao, Xingcheng Yao, and Danqi Chen. Simcse: Simple contrastive learning of sentence embeddings. In EMNLP, pages 6894-6910, 2021. 2 +[20] John M. Giorgi, Osvald Nitski, Bo Wang, and Gary D. Bader. DeCLUTR: Deep contrastive learning for unsupervised textual representations. In ACL/IJCNLP, pages 879-895, 2021. 2 + +[21] Rohit Girdhar, Alaaeldin El-Nouby, Mannat Singh, Kalyan Vasudev Alwala, Armand Joulin, and Ishan Misra. OmniMAE: Single model masked pretraining on images and videos. CoRR, abs/2206.08356, 2022. 1, 2 +[22] Rohit Girdhar, Mannat Singh, Nikhila Ravi, Laurens van der Maaten, Armand Joulin, and Ishan Misra. Omnivore: A single model for many visual modalities. In CVPR, pages 16081-16091, 2022. 1 +[23] Gabriel Goh, Nick Cammarata, Chelsea Voss, Shan Carter, Michael Petrov, Ludwig Schubert, Alec Radford, and Chris Olah. Multimodal neurons in artificial neural networks. *Distill*, 2021, 7, 21 +[24] Yuan Gong, Yu-An Chung, and James R. Glass. AST: audio spectrogram transformer. In Interspeech, pages 571-575, 2021. 1 +[25] Yash Goyal, Tejas Khot, Douglas Summers-Stay, Dhruv Bartra, and Devi Parikh. Making the V in VQA matter: Elevating the role of image understanding in visual question answering. In CVPR, 2017. 4, 13 +[26] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollár, and Ross B. Girshick. Masked autoencoders are scalable vision learners. In CVPR, pages 15979-15988, 2022. 2 +[27] Jonathan Ho, William Chan, Chitwan Saharia, Jay Whang, Ruiqi Gao, Alexey A. Gritsenko, Diederik P. Kingma, Ben Poole, Mohammad Norouzi, David J. Fleet, and Tim Salimans. Imagen video: High definition video generation with diffusion models. CoRR, abs/2210.02303, 2022. 2 +[28] Po-Yao Huang, Hu Xu, Juncheng Li, Alexei Baevski, Michael Auli, Wojciech Galuba, Florian Metze, and Christoph Feichtenhofer. Masked autoencoders that listen. CoRR, abs/2207.06405, 2022. 1 +[29] Yupan Huang, Tengchao Lv, Lei Cui, Yutong Lu, and Furu Wei. Layoutlmv3: Pre-training for document AI with unified text and image masking. In ACMMM, pages 4083-4091, 2022. 2 +[30] Zhicheng Huang, Zhaoyang Zeng, Bei Liu, Dongmei Fu, and Jianlong Fu. Pixel-BERT: Aligning image pixels with text by deep multi-modal transformers. CoRR, abs/2004.00849, 2020.6 +[31] Andrew Jaegle, Sebastian Borgeaud, Jean-Baptiste Alayrac, Carl Doersch, Catalin Ionescu, David Ding, Skanda Koppula, Daniel Zoran, Andrew Brock, Evan Shelhamer, Olivier J. Henaff, Matthew M. Botvinick, Andrew Zisserman, Oriol Vinyals, and João Carreira. Perceiver IO: A general architecture for structured inputs & outputs. In ICLR, 2022. 1 +[32] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc V. Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In ICML, 2021. 2 +[33] Yu Jiang, Vivek Natarajan, Xinlei Chen, Marcus Rohrbach, Dhruv Batra, and Devi Parikh. Pythia v0.1: The winning entry to the VQA challenge 2018. CoRR, abs/1807.09956, 2018. 5 +[34] John Jumper, Richard Evans, Alexander Pritzel, Tim Green, Michael Figurnov, Olaf Ronneberger, Kathryn Tunyasuvunakool, Russ Bates, Augustin Žídek, Anna Potapenko, Alex + +Bridgland, Clemens Meyer, Simon A A Kohl, Andrew J Ballard, Andrew Cowie, Bernardino Romera-Paredes, Stanislav Nikolov, Rishub Jain, Jonas Adler, Trevor Back, Stig Petersen, David Reiman, Ellen Clancy, Michal Zielinski, Martin Steinegger, Michalina Pacholska, Tamas Berghammer, Sebastian Bodenstein, David Silver, Oriol Vinyals, Andrew W Senior, Koray Kavukcuoglu, Pushmeet Kohli, and Demis Hassabis. Highly accurate protein structure prediction with AlphaFold. Nature, 596(7873):583-589, 2021. 1 +[35] Geewook Kim, Teakgyu Hong, Moonbin Yim, JeongYeon Nam, Jinyoung Park, Jinyeong Yim, Wonseok Hwang, Sang-doo Yun, Dongyoon Han, and Seunghyun Park. OCR-Free document understanding transformer. In ECCV, pages 498- 517,2022.2 +[36] Wonjae Kim, Bokyung Son, and Ildoo Kim. ViLT: Vision-and-language transformer without convolution or region supervision. In ICML, pages 5583-5594. PMLR, 2021. 5, 17 +[37] Alexander Kolesnikov, André Susano Pinto, Lucas Beyer, Xiaohua Zhai, Jeremiah Harmsen, and Neil Houlsby. UViM: A unified modeling approach for vision with learned guiding codes. In NeurIPS, 2022. 1 +[38] Simon Kornblith, Mohammad Norouzi, Honglak Lee, and Geoffrey Hinton. Similarity of neural network representations revisited. In ICML, pages 3519-3529, 2019. 22 +[39] Taku Kudo and John Richardson. SentencePiece: A simple and language independent subword tokenizer and tokenizer for neural text processing. In EMNLP, 2018. 2, 6 +[40] Juho Lee, Yoonho Lee, Jungtaek Kim, Adam R. Kosiorek, Seungjin Choi, and Yee Whye Teh. Set transformer: A framework for attention-based permutation-invariant neural networks. In ICML, pages 3744-3753, 2019. 3, 14 +[41] Kenton Lee, Mandar Joshi, Iulia Turc, Hexiang Hu, Fangyu Liu, Julian Eisenschlos, Urvashi Khandelwal, Peter Shaw, Ming-Wei Chang, and Kristina Toutanova. Pix2struct: Screenshot parsing as pretraining for visual language understanding. CoRR, abs/2210.03347, 2022. 2 +[42] Yoann Lemesle, Masataka Sawayama, Guillermo Valle-Perez, Maxime Adolphe, Hélène Sauzeon, and Pierre-Yves Oudeyer. Language-biased image classification: evaluation based on semantic representations. In ICLR, 2022. 7, 21 +[43] Liunian Harold Li, Mark Yatskar, Da Yin, Cho-Jui Hsieh, and Kai-Wei Chang. VisualBERT: A simple and performant baseline for vision and language. CoRR, abs/1908.03557, 2019. 6 +[44] Qing Li, Boqing Gong, Yin Cui, Dan Kondratyuk, Xianzhi Du, Ming-Hsuan Yang, and Matthew Brown. Towards a unified foundation model: Jointly pre-training transformers on unpaired images and text. CoRR, abs/2112.07074, 2021. 2 +[45] Weixin Liang, Yuhui Zhang, Yongchan Kwon, Serena Yeung, and James Zou. Mind the gap: Understanding the modality gap in multi-modal contrastive representation learning. In NeurIPS, 2022. 7, 8, 22, 23 +[46] Valerii Likhosherstov, Anurag Arnab, Krzysztof Choromanski, Mario Lucic, Yi Tay, Adrian Weller, and Mostafa Dehghani. Polyvit: Co-training vision transformers on images, videos and audio. Trans. Machine Learning Research, 2023. 1, 2, 3 + +[47] Lajanugen Logeswaran and Honglak Lee. An efficient framework for learning sentence representations. In ICLR, 2018. 2, 4 +[48] Jiasen Lu, Dhruv Batra, Devi Parikh, and Stefan Lee. ViLBERT: Pretraining task-agnostic visiolinguistic representations for vision-and-language tasks. In NeurIPS, pages 13–23, 2019. 6 +[49] Elman Mansimov, Mitchell Stern, Mia Xu Chen, Orhan Firat, Jakob Uszkoreit, and Puneet Jain. Towards end-to-end in-image neural machine translation. CoRR, abs/2010.10648, 2020. 1, 2, 8 +[50] Joanna Materzyńska, Antonio Torralba, and David Bau. Disentangling visual and written concepts in CLIP. In CVPR, pages 16410-16419, 2022. 7, 21 +[51] Joanna Materzynska, Antonio Torralba, and David Bau. Disentangling visual and written concepts in clip. In CVPR, 2022. 21, 22 +[52] Sabrina J. Mielke, Zaid Alyafeai, Elizabeth Salesky, Colin Raffel, Manan Dey, Matthias Galle, Arun Raja, Chenglei Si, Wilson Y. Lee, Benoit Sagot, and Samson Tan. Between words and characters: A brief history of open-vocabulary modeling and tokenization in NLP. CoRR, abs/2112.10508, 2021. 2 +[53] Basil Mustafa, Carlos Riquelme, Joan Puigcerver, Rodolphe Jenatton, and Neil Houlsby. Multimodal contrastive learning with LIMoE: the language-image mixture of experts. In NeurIPS, 2022. 2, 3, 4 +[54] Daisuke Niizumi, Daiki Takeuchi, Yasunori Ohishi, Noboru Harada, and Kunio Kashino. Masked spectrogram modeling using masked autoencoders for learning general-purpose audio representation. CoRR, abs/2204.12260, 2022. 1 +[55] Hieu Pham, Zihang Dai, Golnaz Ghiasi, Kenji Kawaguchi, Hanxiao Liu, Adams Wei Yu, Jiahui Yu, Yi-Ting Chen, Minh-Thang Luong, Yonghui Wu, Mingxing Tan, and Quoc V. Le. Combined scaling for open-vocabulary image classification. CoRR, abs/2111.10050, 2021. 2 +[56] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In ICML, 2021. 1, 2, 3, 4, 6, 14, 17 +[57] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J. Liu. Exploring the limits of transfer learning with a unified text-to-text transformer. J. Mach. Learn. Res., 21:140:1-140:67, 2020. 3, 4, 6, 20 +[58] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In ICML, pages 8821-8831, 2021. 2 +[59] Nils Rethmeier and Isabelle Augenstein. A primer on contrastive pretraining in language processing: Methods, lessons learned and perspectives. ACM Computing Surveys, 2021. 2 +[60] Phillip Rust, Jonas F. Lotz, Emanuele Bugliarello, Elizabeth Salesky, Miryam de Lhoneux, and Desmond Elliott. Language modelling with pixels. In ICLR, 2023. 1, 2, 3, 7, 18 + +[61] Phillip Rust, Jonas Pfeiffer, Ivan Vulic, Sebastian Ruder, and Iryna Gurevych. How good is your tokenizer? on the monolingual performance of multilingual language models. In ACL/IJCNLP, pages 3118-3135, 2021. 2 +[62] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily Denton, Seyed Kamyar Seyed Ghasemipour, Burcu Karagol Ayan, S. Sara Mahdavi, Rapha Gontijo Lopes, Tim Salimans, Jonathan Ho, David J. Fleet, and Mohammad Norouzi. Photorealistic text-to-image diffusion models with deep language understanding. In NeurIPS, 2022. 2 +[63] Elizabeth Salesky, David Etter, and Matt Post. Robust open-vocabulary translation from visual text representations. In EMNLP, pages 7235-7252, 2021. 2, 3 +[64] Christoph Schuhmann, Richard Vencu, Romain Beaumont, Robert Kaczmarczyk, Clayton Mullis, Aarush Katta, Theo Coombes, Jenia Jitsev, and Aran Komatsuzaki. LAION-400M: Open dataset of CLIP-filtered 400 million image-text pairs. CoRR, abs/2111.02114, 2021. 4, 15 +[65] Rico Sennrich, Barry Haddow, and Alexandra Birch. Neural machine translation of rare words with subword units. In ACL, 2016. 2 +[66] Noam Shazeer and Mitchell Stern. Adafactor: Adaptive learning rates with sublinear memory cost. In ICML, 2018. 3, 14 +[67] Sheng Shen, Liunian Harold Li, Hao Tan, Mohit Bansal, Anna Rohrbach, Kai-Wei Chang, Zhewei Yao, and Kurt Keutzer. How much can CLIP benefit vision-and-language tasks? In ICLR, 2022. 2, 3, 5, 17 +[68] Baohua Sun, Lin Yang, Catherine Chi, Wenhan Zhang, and Michael Lin. Squared english word: A method of generating glyph to use super characters for sentiment analysis. In Workshop on Affective Content Analysis, volume 2328, pages 140-151, 2019. 2 +[69] Zineng Tang, Jaemin Cho, Yixin Nie, and Mohit Bansal. TVLT: Textless vision-language transformer. In NeurIPS, 2022. 1 +[70] Ashish Thapliyal, Jordi Pont-Tuset, Xi Chen, and Radu Soricut. Crossmodal-3600: A Massively Multilingual Multimodal Evaluation Dataset. In EMNLP, 2022. 4, 6, 19 +[71] Bart Thomee, David A. Shamma, Gerald Friedland, Benjamin Elizalde, Karl Ni, Douglas Poland, Damian Borth, and Li-Jia Li. YFCC100M: the new data in multimedia research. Commun. ACM, 59(2):64-73, 2016. 4, 14 +[72] Ruben Villegas, Mohammad Babaeizadeh, Pieter-Jan Kindermans, Hernan Moraldo, Han Zhang, Mohammad Taghi Saffar, Santiago Castro, Julius Kunze, and Dumitru Erhan. Phenaki: Variable length video generation from open domain textual description. In ICLR, 2023. 2 +[73] Alex Wang, Amanpreet Singh, Julian Michael, Felix Hill, Omer Levy, and Samuel R. Bowman. GLUE: A multi-task benchmark and analysis platform for natural language understanding. In ICLR, 2019. 2, 3, 4, 7, 18 +[74] Jianfeng Wang, Zhengyuan Yang, Xiaowei Hu, Linjie Li, Kevin Lin, Zhe Gan, Zicheng Liu, Ce Liu, and Lijuan Wang. GIT: A generative image-to-text transformer for vision and language. Trans. Machine Learning Research, 2022. 1, 2 + +[75] Wenhui Wang, Hangbo Bao, Li Dong, Johan Bjorck, Zhiliang Peng, Qiang Liu, Kriti Aggarwal, Owais Khan Mohammed, Saksham Singhal, Subhojit Som, and Furu Wei. Image as a foreign language: BEiT pretraining for all vision and vision-language tasks. CoRR, abs/2208.10442, 2022. 1 +[76] Zirui Wang, Jiahui Yu, Adams Wei Yu, Zihang Dai, Yulia Tsvetkov, and Yuan Cao. SimVLM: Simple visual language model pretraining with weak supervision. In ICLR, 2022. 1, 8 +[77] Zhuofeng Wu, Sinong Wang, Jiatao Gu, Madian Khabsa, Fei Sun, and Hao Ma. CLEAR: contrastive learning for sentence representation. CoRR, abs/2012.15466, 2020. 2, 4 +[78] Linting Xue, Aditya Barua, Noah Constant, Rami Al-Rfou, Sharan Narang, Mihir Kale, Adam Roberts, and Colin Raffel. Byt5: Towards a token-free future with pre-trained byte-to-byte models. Trans. Assoc. Comput. Linguistics, 10:291-306, 2022. 6 +[79] Linting Xue, Noah Constant, Adam Roberts, Mihir Kale, Rami Al-Rfou, Aditya Siddhant, Aditya Barua, and Colin Raffel. mT5: A massively multilingual pre-trained text-to-text transformer. In NAACL-HLT, 2021. 6, 20 +[80] Haoxuan You, Luowei Zhou, Bin Xiao, Noel Codella, Yu Cheng, Ruochen Xu, Shih-Fu Chang, and Lu Yuan. Learning visual representation from modality-shared contrastive language-image pre-training. In ECCV, pages 69-87, 2022. 2 +[81] Peter Young, Alice Lai, Micah Hodosh, and Julia Hockenmaier. From image descriptions to visual denotations: New similarity metrics for semantic inference over event descriptions. Trans. Assoc. Comput. Linguistics, 2:67-78, 2014. 4 +[82] Jiahui Yu, Zirui Wang, Vijay Vasudevan, Legg Yeung, Mojtaba Seyedhosseini, and Yonghui Wu. Coca: Contrastive captioners are image-text foundation models. Trans. Machine Learning Research, 2022. 2 +[83] Jiahui Yu, Yuanzhong Xu, Jing Yu Koh, Thang Luong, Gunjan Baid, Zirui Wang, Vijay Vasudevan, Alexander Ku, Yinfei Yang, Burcu Karagol Ayan, Ben Hutchinson, Wei Han, Zarana Parekh, Xin Li, Han Zhang, Jason Baldridge, and Yonghui Wu. Scaling autoregressive models for content-rich text-to-image generation. Trans. Machine Learning Research, 2022. 2 +[84] Zhou Yu, Jun Yu, Yuhao Cui, Dacheng Tao, and Qi Tian. Deep modular co-attention networks for visual question answering. In CVPR, pages 6281–6290, 2019. 5 +[85] Xiaohua Zhai, Xiao Wang, Basil Mustafa, Andreas Steiner, Daniel Keysers, Alexander Kolesnikov, and Lucas Beyer. LiT: Zero-shot transfer with locked-image text tuning. In CVPR, pages 18102-18112, 2022. 2, 4, 5, 6, 14 +[86] Yukun Zhu, Ryan Kiros, Richard S. Zemel, Ruslan Salakhutdinov, Raquel Urtasun, Antonio Torralba, and Sanja Fidler. Aligning books and movies: Towards story-like visual explanations by watching movies and reading books. In ICCV, pages 19-27, 2015. 7, 18 \ No newline at end of file diff --git a/2023/CLIPPO_ Image-and-Language Understanding From Pixels Only/images.zip b/2023/CLIPPO_ Image-and-Language Understanding From Pixels Only/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..51b8fe4f71ccaed22fabcdd83ccff9dbe51a8b96 --- /dev/null +++ b/2023/CLIPPO_ Image-and-Language Understanding From Pixels Only/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:32b8638d4015f3ffe463636441543b85a17d9eb3c9fb24af6a5e609a6fa9f243 +size 283635 diff --git a/2023/CLIPPO_ Image-and-Language Understanding From Pixels Only/layout.json b/2023/CLIPPO_ Image-and-Language Understanding From Pixels Only/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..8009cc034c62b48501793aeaa98b9a5dfd5d6cb7 --- /dev/null +++ b/2023/CLIPPO_ Image-and-Language Understanding From Pixels Only/layout.json @@ -0,0 +1,7809 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 97, + 103, + 496, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 97, + 103, + 496, + 121 + ], + "spans": [ + { + "bbox": [ + 97, + 103, + 496, + 121 + ], + "type": "text", + "content": "CLIPPO: Image-and-Language Understanding from Pixels Only" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 176, + 143, + 416, + 172 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 176, + 143, + 416, + 172 + ], + "spans": [ + { + "bbox": [ + 176, + 143, + 416, + 172 + ], + "type": "text", + "content": "Michael Tschannen, Basil Mustafa, Neil Houlsby \nGoogle Research, Brain Team, Zürich" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 143, + 199, + 192, + 212 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 199, + 192, + 212 + ], + "spans": [ + { + "bbox": [ + 143, + 199, + 192, + 212 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 224, + 290, + 523 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 224, + 290, + 523 + ], + "spans": [ + { + "bbox": [ + 47, + 224, + 290, + 523 + ], + "type": "text", + "content": "Multimodal models are becoming increasingly effective, in part due to unified components, such as the Transformer architecture. However, multimodal models still often consist of many task- and modality-specific pieces and training procedures. For example, CLIP (Radford et al., 2021) trains independent text and image towers via a contrastive loss. We explore an additional unification: the use of a pure pixel-based model to perform image, text, and multimodal tasks. Our model is trained with contrastive loss alone, so we call it CLIP-Pixels Only (CLIPPO). CLIPPO uses a single encoder that processes both regular images and text rendered as images. CLIPPO performs image-based tasks such as retrieval and zero-shot image classification almost as well as CLIP-style models, with half the number of parameters and no text-specific tower or embedding. When trained jointly via image-text contrastive learning and next-sentence contrastive learning, CLIPPO can perform well on natural language understanding tasks, without any word-level loss (language modelling or masked language modelling), outperforming pixel-based prior work. Surprisingly, CLIPPO can obtain good accuracy in visual question answering, simply by rendering the question and image together. Finally, we exploit the fact that CLIPPO does not require a tokenizer to show that it can achieve strong performance on multilingual multimodal retrieval without modifications." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 546, + 128, + 558 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 546, + 128, + 558 + ], + "spans": [ + { + "bbox": [ + 47, + 546, + 128, + 558 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 566, + 287, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 566, + 287, + 685 + ], + "spans": [ + { + "bbox": [ + 46, + 566, + 287, + 685 + ], + "type": "text", + "content": "In recent years, large-scale multimodal training of Transformer-based models has led to improvements in the state-of-the-art in different domains including vision [2, 10, 74-76], language [6, 11], and audio [5]. In particular, in computer vision and image-language understanding, a single large pretrained model can outperform task-specific expert models [10, 74, 75]. However, large multimodal models often use modality or dataset-specific encoders and decoders, and accordingly lead to involved protocols. For example, such models frequently involve training different" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 310, + 198, + 425, + 326 + ], + "blocks": [ + { + "bbox": [ + 310, + 198, + 425, + 326 + ], + "lines": [ + { + "bbox": [ + 310, + 198, + 425, + 326 + ], + "spans": [ + { + "bbox": [ + 310, + 198, + 425, + 326 + ], + "type": "image", + "image_path": "ffdab49a0a316c2db9fc0e440b859ee63037c0ec97ed700f60899518134c1a3e.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 334, + 547, + 411 + ], + "lines": [ + { + "bbox": [ + 304, + 334, + 547, + 411 + ], + "spans": [ + { + "bbox": [ + 304, + 334, + 547, + 411 + ], + "type": "text", + "content": "Figure 1. CLIP [56] trains separate image and text encoders, each with a modality-specific preprocessing and embedding, on image/alt-text pairs with a contrastive objective. CLIPPO trains a pure pixel-based model with equivalent capabilities by rendering the alt-text as an image, encoding the resulting image pair using a shared vision encoder (in two separate forward passes), and applying same training objective as CLIP." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 429, + 198, + 542, + 326 + ], + "blocks": [ + { + "bbox": [ + 429, + 198, + 542, + 326 + ], + "lines": [ + { + "bbox": [ + 429, + 198, + 542, + 326 + ], + "spans": [ + { + "bbox": [ + 429, + 198, + 542, + 326 + ], + "type": "image", + "image_path": "cbb943b06baad3ca5e62905b787b868708c0afc9f1dc74e14944d5e5690f8131.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 426, + 545, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 426, + 545, + 533 + ], + "spans": [ + { + "bbox": [ + 304, + 426, + 545, + 533 + ], + "type": "text", + "content": "parts of the model in separate phases on their respective datasets, with dataset-specific preprocessing, or transferring different parts in a task-specific manner [75]. Such modality and task-specific components can lead to additional engineering complexity, and poses challenges when introducing new pretraining losses or downstream tasks. Developing a single end-to-end model that can process any modality, or combination of modalities, would be a valuable step for multimodal learning. Here, we focus on images and text." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 533, + 546, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 533, + 546, + 700 + ], + "spans": [ + { + "bbox": [ + 304, + 533, + 546, + 700 + ], + "type": "text", + "content": "A number of key unifications have accelerated the progress of multimodal learning. First, the Transformer architecture has been shown to work as a universal backbone, performing well on text [6, 15], vision [16], audio [5, 24, 54], and other domains [7, 34]. Second, many papers have explored mapping different modalities into a single shared embedding space to simplify the input/output interface [21, 22, 46, 69], or develop a single interface to many tasks [31, 37]. Third, alternative representations of modalities allow harnessing in one domain neural architectures or training procedures designed for another domain [28, 49, 54, 60]. For example, [60] and [28, 54] represent text and audio, respectively, by rendering these modalities as images (via a spectogram in the case of audio)." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 317, + 701, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 701, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 317, + 701, + 545, + 714 + ], + "type": "text", + "content": "In this paper, we explore the use of a pure pixel-based" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 693, + 287, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 693, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 693, + 287, + 713 + ], + "type": "text", + "content": "Code and pretrained models are available as part of big_vision [4] https://github.com/google-research/big_vision." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "11006" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 191 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 191 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 191 + ], + "type": "text", + "content": "model for multimodal learning of text and images. Our model is a single Vision Transformer [16] that processes visual input, or text, or both together, all rendered as RGB images. The same model parameters are used for all modalities, including low-level feature processing; that is, there are no modality-specific initial convolutions, tokenization algorithms, or input embedding tables. We train our model using only a single task: contrastive learning, as popularized by CLIP [56] and ALIGN [32]. We therefore call our model CLIP-Pixels Only (CLIPPO)." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 193, + 289, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 193, + 289, + 361 + ], + "spans": [ + { + "bbox": [ + 46, + 193, + 289, + 361 + ], + "type": "text", + "content": "We find that CLIPPO performs similarly to CLIP-style models (within " + }, + { + "bbox": [ + 46, + 193, + 289, + 361 + ], + "type": "inline_equation", + "content": "1 - 2\\%" + }, + { + "bbox": [ + 46, + 193, + 289, + 361 + ], + "type": "text", + "content": ") on the main tasks CLIP was designed for—image classification and text/image retrieval—despite not having modality-specific towers. Surprisingly, CLIPPO can perform complex language understanding tasks to a decent level without any left-to-right language modelling, masked language modelling, or explicit word-level losses. In particular, on the GLUE benchmark [73] CLIPPO outperforms classic NLP baselines, such as ELMO+BiLSTM+attention, outperforms prior pixel-based masked language models [60], and approaches the score of BERT [15]. Interestingly, CLIPPO obtains good performance on VQA when simply rendering the image and text together, despite never having been pretrained on such data." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 361, + 289, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 361, + 289, + 434 + ], + "spans": [ + { + "bbox": [ + 46, + 361, + 289, + 434 + ], + "type": "text", + "content": "Pixel-based models have an immediate advantage over regular language models because they do not require predetermining the vocabulary/tokenizer and navigating the corresponding intricate trade-offs; consequently, we observe improved performance on multilingual retrieval compared to an equivalent model that uses a classical tokenizer." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 449, + 132, + 462 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 449, + 132, + 462 + ], + "spans": [ + { + "bbox": [ + 47, + 449, + 132, + 462 + ], + "type": "text", + "content": "2. Related work" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 472, + 287, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 472, + 287, + 544 + ], + "spans": [ + { + "bbox": [ + 46, + 472, + 287, + 544 + ], + "type": "text", + "content": "Multimodal and contrastive pretraining Most closely related to CLIPPO are CLIP [67] and ALIGN [32] which developed the paradigm of large-scale contrastive training on noisy data from the web. Follow-ups [55,85] have scaled further and employed state-of-the-art image representation learning to boost performance." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 545, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 545, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 545, + 289, + 715 + ], + "type": "text", + "content": "A number of works have explored model unification via weight-sharing. In the contrastive context, LIMoE [53] and MS-CLIP [80] explore a one-tower model similar to ours, studying the use of mixture of experts and selective sharing of modules, respectively. Outside contrastive training, co-training distinct tasks [1, 46] is a popular strategy, with some approaches [44] involving knowledge distillation and gradient masking. Other works use self-supervised learning algorithms to unify task training [21]. These broadly use discriminative tasks to learn representations for various downstream modalities; generative approaches to multimodal modelling have been scaled to billions of parameters, generating text [2, 10, 74, 82], images [58, 62, 83], videos [27, 72] or audio [5] from various modalities." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 72, + 545, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 180 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 180 + ], + "type": "text", + "content": "Another related domain is document and user interface (UI) understanding. Corresponding models are trained on diverse multimodal data sets and can usually solve a range of document/Ui understanding tasks. Many models rely on text extracted using an off-the-shelf OCR pipeline in combination with document images [3, 29], but image-only models are getting more popular [35, 41]. While these models can understand visual cues and text from the input image, they still rely on a tokenized text for training and inference." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 182, + 545, + 292 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 182, + 545, + 292 + ], + "spans": [ + { + "bbox": [ + 304, + 182, + 545, + 292 + ], + "type": "text", + "content": "Contrastive training in NLP There is a sizable body of work on contrastive pretraining on sentence pairs (see [59] for a recent survey), which we explore as an auxiliary objective for CLIPPO. Popular augmentations to generate text pairs involve word deletion, span deletion, reordering, synonym substitution, and next-sentence-prediction [20, 47, 77]. Other methods use different realizations of dropout masks in the model to emulate sentence pairs, or supervised labels to obtain positive and negative pairs [19]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 293, + 545, + 414 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 293, + 545, + 414 + ], + "spans": [ + { + "bbox": [ + 304, + 293, + 545, + 414 + ], + "type": "text", + "content": "Visual text and tokenization in NLP The most closely related method to CLIPPO from the NLP domain is PILIX [60], which is a masked autoencoder (MAE) [26] trained on rendered text. It obtains strong performance on multilingual syntactic (part-of-speech tagging, dependency parsing) and semantic language understanding (named entity recognition, sentence understanding) tasks, while being more robust to noise in the text than BERT. Other applications for which visual text has been explored include sentiment analysis [68] and machine translation [49, 63]." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 415, + 545, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 415, + 545, + 486 + ], + "spans": [ + { + "bbox": [ + 304, + 415, + 545, + 486 + ], + "type": "text", + "content": "Visual text side-steps the design and construction of an appropriate tokenizer, which is a large area of research of its own, and can hence simplify text processing in certain—in particular multilingual—scenarios. We refer to [52] for a survey on tokenizers. Popular models include WordPiece [15], Byte-Pair Encoding [65], and SentencePiece [39]." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 487, + 545, + 606 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 487, + 545, + 606 + ], + "spans": [ + { + "bbox": [ + 304, + 487, + 545, + 606 + ], + "type": "text", + "content": "Subword-based vocabularies are popular in monolingual setups and usually lead to a good performance trade-off compared to word and character based vocabularies for certain languages including English. In multilingual contexts, appropriately representing the vocabulary of all languages becomes challenging as the number of languages increases [13,61], which in turn can lead to poor performance in tasks involving underrepresented languages. A variety of mitigation strategies has been developed; we refer to [60, Sec. 5.1] for a more detailed discussion of these strategies." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 619, + 545, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 619, + 545, + 647 + ], + "spans": [ + { + "bbox": [ + 305, + 619, + 545, + 647 + ], + "type": "text", + "content": "3. Contrastive language-image pretraining with pixels" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 654, + 545, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 545, + 715 + ], + "type": "text", + "content": "Contrastive language-image pretraining has emerged as a powerful, scalable paradigm to train versatile vision models on web-scale data sets [56]. Concretely, this approach relies on image/alt-text pairs which can be automatically collected at large scale from the web. Thereby, the textual" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "11007" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 191 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 191 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 191 + ], + "type": "text", + "content": "descriptions are usually noisy, and can e.g. consist of single keywords, sets of keywords, or potentially lengthy descriptions with many attributes describing the image content. Using this data, two encoders are jointly trained, namely a text encoder embedding the alt-texts and an image encoder embedding the corresponding images into a shared latent space. These two encoders are trained with a contrastive loss, encouraging the embeddings of matching images and alt-text to be similar, and at the same time to be dissimilar from all other image and alt-text embeddings." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 194, + 289, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 194, + 289, + 350 + ], + "spans": [ + { + "bbox": [ + 46, + 194, + 289, + 350 + ], + "type": "text", + "content": "Once trained, such an encoder pair can be used in many ways: It can be specialized to classifying a fixed set of visual concepts via their textual descriptions (zero-shot classification); the embeddings can be used to retrieve images given a textual description and vice-versa; or the vision encoder can be transferred in supervised fashion to a downstream task by fine-tuning on a labeled data set or by training a head on top of the frozen image encoder representation. In principle, the text encoder can be used as a standalone text embedding, but this application—to our knowledge—has not been explored in-depth, with some authors citing the low quality of the alt-texts leading to weak language modeling performance of the text encoder [67]." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 350, + 289, + 471 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 350, + 289, + 471 + ], + "spans": [ + { + "bbox": [ + 46, + 350, + 289, + 471 + ], + "type": "text", + "content": "Previous works [46, 53] have shown that the image and text encoder can be realized with a single shared transformer model (henceforth referred to as single tower model, or 1T-CLIP), where the images are embedded using a patch embedding, and the tokenized text is embedded using a separate word embedding. Apart from the modality-specific embeddings, all model parameters are shared for the two modalities. While this type of sharing usually leads to a minor performance drop on image/image-language tasks it also halves the number of model parameters." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 471, + 289, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 471, + 289, + 579 + ], + "spans": [ + { + "bbox": [ + 46, + 471, + 289, + 579 + ], + "type": "text", + "content": "CLIPPO takes this idea one step further: text inputs are rendered on blank images, and are subsequently dealt with entirely as images, including the initial patch embedding (see Fig. 1 for an illustration). By training this single vision transformer contrastively as prior works, we obtain a single vision transformer model that can understand both images and text through the single interface of vision and provides a single representation which can be used to solve image, image-language, and pure language understanding tasks." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 580, + 289, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 580, + 289, + 640 + ], + "spans": [ + { + "bbox": [ + 46, + 580, + 289, + 640 + ], + "type": "text", + "content": "Alongside multimodal versatility, CLIPPO alleviates common hurdles with text processing, namely the development of an appropriate tokenizer and vocabulary. This is particularly interesting in a massively multilingual setup, where the text encoder has to handle dozens of languages." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 641, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 641, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 641, + 289, + 715 + ], + "type": "text", + "content": "We find that CLIPPO trained on image/alt-text pairs performs comparably with its 1T-CLIP counterpart on common image and image-language benchmarks, and is competitive with strong baseline language models on the GLUE benchmark [73]. However, due to the low quality of the alt-texts which are often not grammatical sentences, learn" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 72, + 547, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 157 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 157 + ], + "type": "text", + "content": "ing language understanding exclusively from alt-texts is fundamentally limited. Therefore, we augment image/alt-text contrastive pretraining with language-based contrastive training. Specifically, we use positive pairs of consecutive sentences sampled from a text corpus which is seamlessly integrated into the contrastive training by supplementing batches of image/alt-texts with (rendered) text/text pairs." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 167, + 388, + 180 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 167, + 388, + 180 + ], + "spans": [ + { + "bbox": [ + 305, + 167, + 388, + 180 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 186, + 459, + 199 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 186, + 459, + 199 + ], + "spans": [ + { + "bbox": [ + 305, + 186, + 459, + 199 + ], + "type": "text", + "content": "4.1. Training details and models" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 204, + 547, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 204, + 547, + 289 + ], + "spans": [ + { + "bbox": [ + 304, + 204, + 547, + 289 + ], + "type": "text", + "content": "We rely on a single training setup for all our baselines and visual text models. This setup was tuned to produce good results for standard image/alt-text contrastive training as in [56] (using exactly the same loss function as [56], following the pseudocode in [56, Fig. 3]) and we found that it readily transfers to 1T-CLIP and CLIPPO (including variants with text/text co-training)." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 289, + 548, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 289, + 548, + 479 + ], + "spans": [ + { + "bbox": [ + 304, + 289, + 548, + 479 + ], + "type": "text", + "content": "Our default architecture is a ViT-B/16 [16] and we perform a subset of experiments with a ViT-L/16 architecture to study the effect of scale (we equip both models a MAP head [40] to pool embeddings). In all cases, the representation dimension used for the contrastive loss is 768. We set the batch size to 10,240 and train the main models for 250k steps, using a minimum 100k training steps for ablations. For models co-trained with a certain percentage of text/text data, we scale the number of iterations such that the number of image/alt-text pairs seen matches the number of iterations of the corresponding model without text/text data (e.g. when " + }, + { + "bbox": [ + 304, + 289, + 548, + 479 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 304, + 289, + 548, + 479 + ], + "type": "text", + "content": " of the data is text/text pairs we increase the number of iterations from 250k to 500k). The contrastive loss is computed across the full batch. We use the Adafactor optimizer [66] with a learning rate of " + }, + { + "bbox": [ + 304, + 289, + 548, + 479 + ], + "type": "inline_equation", + "content": "10^{-3}" + }, + { + "bbox": [ + 304, + 289, + 548, + 479 + ], + "type": "text", + "content": " and decoupled weight decay with weight " + }, + { + "bbox": [ + 304, + 289, + 548, + 479 + ], + "type": "inline_equation", + "content": "10^{-4}" + }, + { + "bbox": [ + 304, + 289, + 548, + 479 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 479, + 547, + 576 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 479, + 547, + 576 + ], + "spans": [ + { + "bbox": [ + 304, + 479, + 547, + 576 + ], + "type": "text", + "content": "Baseline CLIP-style models are trained using the T5-en SentencePiece tokenizer [57]; we use the abbreviation " + }, + { + "bbox": [ + 304, + 479, + 547, + 576 + ], + "type": "inline_equation", + "content": "\\mathrm{CLIP^{*}}" + }, + { + "bbox": [ + 304, + 479, + 547, + 576 + ], + "type": "text", + "content": " for the two tower model from [56] trained from scratch using the setup described above, to avoid confusion with the model released by [56]. A sequence length of 196 is used, as this matches the number of visual text \"tokens\" CLIPPO can process with patch size 16 has at " + }, + { + "bbox": [ + 304, + 479, + 547, + 576 + ], + "type": "inline_equation", + "content": "224\\mathrm{px}" + }, + { + "bbox": [ + 304, + 479, + 547, + 576 + ], + "type": "text", + "content": " resolution (which we use throughout unless noted otherwise)." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 578, + 547, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 578, + 547, + 687 + ], + "spans": [ + { + "bbox": [ + 304, + 578, + 547, + 687 + ], + "type": "text", + "content": "Visual text For visual text rendering [60, 63] relied on the Google Noto font family1 which supports the majority of Unicode code points. Here, we use the GNU Unifont bitmap font2, which has a similar coverage but allows for efficient, lookup-based on-the-fly rendering in our preprocessing pipeline. We emphasize that this rendering strategy does not slow down training compared to tokenizer-based models. In preliminary explorations, we found this to be performance-neutral when compared to the Noto font." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 315, + 693, + 463, + 702 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 693, + 463, + 702 + ], + "spans": [ + { + "bbox": [ + 315, + 693, + 463, + 702 + ], + "type": "text", + "content": "1https://fonts.google.com/nota" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 317, + 703, + 462, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 703, + 462, + 712 + ], + "spans": [ + { + "bbox": [ + 317, + 703, + 462, + 712 + ], + "type": "text", + "content": "2http://unifoundry.com/unifont" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "11008" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 107, + 70, + 482, + 188 + ], + "blocks": [ + { + "bbox": [ + 107, + 70, + 482, + 188 + ], + "lines": [ + { + "bbox": [ + 107, + 70, + 482, + 188 + ], + "spans": [ + { + "bbox": [ + 107, + 70, + 482, + 188 + ], + "type": "table", + "html": "
#param.training datasetI1k 10s.I1k 0s.C I→TC T→IF I→TF T→I
CLIP*203MWebLI55.865.148.531.379.259.4
1T-CLIP118MWebLI53.962.348.030.377.558.2
CLIPPO93MWebLI53.061.447.330.176.457.3
CLIPPO93MWebLI + 25%C452.157.440.726.768.951.8
CLIPPO93MWebLI + 50%C448.053.135.223.464.847.2
1T-CLIP L/16349MWebLI60.867.850.732.581.061.0
CLIPPO L/16316MWebLI60.367.450.633.479.262.6
CLIPPO L/16316MWebLI + 25%C460.566.044.529.872.957.3
CLIPPO L/16316MWebLI + 50%C456.861.739.727.370.154.7
", + "image_path": "23b2d87e1dacca821161ce70e2cb1c7a8632b4ec9ba6e242046e914e76fa7ca7.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 194, + 545, + 248 + ], + "lines": [ + { + "bbox": [ + 46, + 194, + 545, + 248 + ], + "spans": [ + { + "bbox": [ + 46, + 194, + 545, + 248 + ], + "type": "text", + "content": "Table 1. Vision and vision-language cross-modal results. We report ImageNet-1k 10-shot linear transfer validation accuracy (I1k 10s.), ImageNet-1k zero-shot transfer validation accuracy (I1k 0s.), image-to-text and text-to-image retrieval recall@1 on MS-COCO (C I→T and C T→I) and on Flickr30k (F T→I and F I→T). CLIPPO and 1T-CLIP incur a minor drop in these evaluations compared to CLIP*, while only using about half of the model parameters. Co-training with text pairs from C4 (models with + xx%C4) degrades performance on some cross-modal tasks (but leads to improved language understanding capabilities, see Table 2)." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 258, + 288, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 258, + 288, + 449 + ], + "spans": [ + { + "bbox": [ + 46, + 258, + 288, + 449 + ], + "type": "text", + "content": "Image/alt-text data We use the WebLI data set introduced in [10] which comprises 10 billion images with 12 billion corresponding alt-texts. Importantly, WebLI comprises alt-texts in 109 languages (unlike previous data sets such as LAION-400M [64] which only contain English alt-texts) and it is therefore a great foundation to study multilingual language-image pretraining and its applications. Please refer to [10, Fig. 3] for details on the alt-text language distribution. For English-only models we obtain English versions of non-English alt-texts via GCP Translation " + }, + { + "bbox": [ + 46, + 258, + 288, + 449 + ], + "type": "inline_equation", + "content": "\\mathrm{API}^3" + }, + { + "bbox": [ + 46, + 258, + 288, + 449 + ], + "type": "text", + "content": ". In addition to alt-text, WebLI also provides OCR annotations, which we do not use in this paper. Finally, WebLI was processed with a de-duplication step removing all images from various splits of the image evaluation sets used in this paper. Please refer to [10, Sec. 3.2] for more details on the WebLI data set and to [10, Appendix B] for a datasheet." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 450, + 287, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 450, + 287, + 486 + ], + "spans": [ + { + "bbox": [ + 47, + 450, + 287, + 486 + ], + "type": "text", + "content": "We also present a subset of results based on LAION-400M [64] and YFCC-100M [71] as an additional comparison points, see Appendix C.1 and C.2, respectively." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 489, + 287, + 609 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 489, + 287, + 609 + ], + "spans": [ + { + "bbox": [ + 46, + 489, + 287, + 609 + ], + "type": "text", + "content": "Text/text data For co-training with text/text pairs we primarily rely on the publicly available Colossal Clean Crawled Corpus (C4; default/English split) [57]. We randomly sample pairs of consecutive sentences and contrastively train on these pairs, i.e., the model is trained for embedding-based next sentence prediction (NSP) [47]. We also experiment with pairs of parallel sentences in different languages from the WMT19 data set [18] as well as back-translated English sentences derived from C4 following the strategy described in [12]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 616, + 183, + 628 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 616, + 183, + 628 + ], + "spans": [ + { + "bbox": [ + 47, + 616, + 183, + 628 + ], + "type": "text", + "content": "4.2. Evaluations and metrics" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 635, + 287, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 635, + 287, + 696 + ], + "spans": [ + { + "bbox": [ + 47, + 635, + 287, + 696 + ], + "type": "text", + "content": "To evaluate the vision and vision/language understanding capabilities of our models we use standard metrics from the literature [53, 56, 85]: \"zero-shot\" transfer, which uses (embedded) textual description of the classes to be classified/retrieved and compares these with image embeddings." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 309, + 258, + 545, + 330 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 309, + 258, + 545, + 330 + ], + "spans": [ + { + "bbox": [ + 309, + 258, + 545, + 330 + ], + "type": "text", + "content": "We report the classification accuracy on ImageNet-1k [14] as well as the recall@1 for cross-modal retrieval on MSCOCO [9] and Flickr30k [81]. Furthermore, we test the low-data transfer performance of the models by means of the linear adaptation protocol from [16], reporting the 10-shot accuracy on ImageNet-1k." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 331, + 546, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 331, + 546, + 545 + ], + "spans": [ + { + "bbox": [ + 304, + 331, + 546, + 545 + ], + "type": "text", + "content": "We also evaluate CLIPPO and baselines on the popular VQA benchmark VQAv2 [25]. To construct a VQA model using a single pretrained ViT we render the question at the top end of the corresponding image (using the same Unifont renderer as used for CLIPPO training) and follow the standard prediction setup where the answer is predicted as the most likely answer from the training set, i.e. by classification. Specifically, we replace the last layer of our pretrained CLIPPO and baselines with a randomly initialized one with the appropriate number of outputs and fine-tune on VQAv2. This setup tests the ability of the pretrained ViT to combine image and text in intermediate layers as it has produce a single output from a fused image/text input image, unlike in the other cross-modal tasks (and pretraining), where image and text representations are computed with two separate forward passes. Please refer to Appendix A in the supplementary material for examples images with rendered questions and Appendix B.1 for details on the fine-tuning protocol." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 546, + 545, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 546, + 545, + 605 + ], + "spans": [ + { + "bbox": [ + 304, + 546, + 545, + 605 + ], + "type": "text", + "content": "Multilingual capabilities are assessed via zero-shot retrieval on CrossModal3600 [70], which is a geographically diverse set comprising 3600 images each human-annotated with captions in 36 languages. The corresponding recall metric is averaged across all languages and images." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 605, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 546, + 713 + ], + "type": "text", + "content": "Finally, we evaluate the language understanding capabilities on the General Language Understanding Evaluation (GLUE) benchmark [73] which comprises natural language inference tasks (MNLI, QNLI, RTE), a sentiment analysis task (SST-2), sentence similarity tasks (QQP, STS-B, MRPC), and a linguistic acceptability task (CoLA). Following common practice, we exclude the WNLI task from the benchmark [15, 77]. We transfer our baselines and CLIPPO models by attaching a 2-hidden layer MLP with 768 units to" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 57, + 702, + 227, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 702, + 227, + 712 + ], + "spans": [ + { + "bbox": [ + 57, + 702, + 227, + 712 + ], + "type": "text", + "content": "3https://cloud.google.com/translate" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "11009" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 53, + 72, + 286, + 314 + ], + "blocks": [ + { + "bbox": [ + 53, + 72, + 286, + 314 + ], + "lines": [ + { + "bbox": [ + 53, + 72, + 286, + 314 + ], + "spans": [ + { + "bbox": [ + 53, + 72, + 286, + 314 + ], + "type": "image", + "image_path": "9b8eb1248493590b3de5a1a00f10c2f59eb466390defefe448d40e116c1814e1.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 324, + 287, + 392 + ], + "lines": [ + { + "bbox": [ + 46, + 324, + 287, + 392 + ], + "spans": [ + { + "bbox": [ + 46, + 324, + 287, + 392 + ], + "type": "text", + "content": "Figure 2. Results on the VQAv2 benchmark (test-dev set). In addition to CLIPPO and baselines produced in this work, we also compare to Pythia and MCAN models with ViT encoders from [67], and with comparably sized METER [17] and ViLT [36] models. CLIPPO outperforms " + }, + { + "bbox": [ + 46, + 324, + 287, + 392 + ], + "type": "inline_equation", + "content": "\\mathrm{CLIP^{*}}" + }, + { + "bbox": [ + 46, + 324, + 287, + 392 + ], + "type": "text", + "content": " and 1T-CLIP clearly on \"yes/no\" questions and gets similar performance as task-specific models." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 415, + 287, + 464 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 415, + 287, + 464 + ], + "spans": [ + { + "bbox": [ + 46, + 415, + 287, + 464 + ], + "type": "text", + "content": "their representation and following precisely the fine-tuning protocol from BERT [15]. For sentence pair classification tasks we simply render both sentences on the same image, printing [SEP] to mark the start of the second sentence." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 474, + 266, + 487 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 474, + 266, + 487 + ], + "spans": [ + { + "bbox": [ + 47, + 474, + 266, + 487 + ], + "type": "text", + "content": "4.3. Vision and vision-language understanding" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 497, + 287, + 675 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 497, + 287, + 675 + ], + "spans": [ + { + "bbox": [ + 46, + 497, + 287, + 675 + ], + "type": "text", + "content": "Image classification and retrieval Table 1 shows the performance of CLIPPO along with the baseline models on the benchmarks described in Sec. 4.2. It can be seen that the CLIPPO and 1T-CLIP incur a drop of a 2-3 percentage points absolute compared to " + }, + { + "bbox": [ + 46, + 497, + 287, + 675 + ], + "type": "inline_equation", + "content": "\\mathrm{CLIP^{*}}" + }, + { + "bbox": [ + 46, + 497, + 287, + 675 + ], + "type": "text", + "content": ". This is not surprising and can be attributed to the fact that single tower models only have about half the parameters count of a corresponding two tower model. The difference in performance between the English-only CLIPPO and 1T-CLIP is very small for a B/16 backbone at " + }, + { + "bbox": [ + 46, + 497, + 287, + 675 + ], + "type": "inline_equation", + "content": "100\\mathrm{k}" + }, + { + "bbox": [ + 46, + 497, + 287, + 675 + ], + "type": "text", + "content": " training steps (see Table 6 in the supplementary material), and vanishes with longer training and/or by increasing the model size, despite the fact that CLIPPO has " + }, + { + "bbox": [ + 46, + 497, + 287, + 675 + ], + "type": "inline_equation", + "content": "25\\%" + }, + { + "bbox": [ + 46, + 497, + 287, + 675 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 497, + 287, + 675 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 46, + 497, + 287, + 675 + ], + "type": "text", + "content": " fewer parameters than 1T-CLIP for a B/16 and L/16 architecture, respectively (which is due to the absence of the text embedding in CLIPPO)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 677, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 677, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 47, + 677, + 287, + 715 + ], + "type": "text", + "content": "The multilingual CLIPPO model performs somewhat worse than the corresponding 1T-CLIP, and the gap does not close completely when training longer (see Table 6)." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 305, + 72, + 545, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 545, + 108 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 545, + 108 + ], + "type": "text", + "content": "However, when evaluated across a broad set of languages on the CrossModal3600 CLIPPO performs on par with or slightly better than 1T-CLIP (see Sec. 4.4 below)." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 110, + 545, + 229 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 110, + 545, + 229 + ], + "spans": [ + { + "bbox": [ + 304, + 110, + 545, + 229 + ], + "type": "text", + "content": "As we add sentence pairs to the training mix the performance on the cross-modal retrieval metrics decreases. This is not surprising as we keep the total batch size constant so that the effective batch size of image/alt-text contrastive training decreases, which is known to impact performance [85]. Interestingly, the 10-shot transfer performance does not move in tandem, but only decreases significantly when half of the training data is sentence pairs. In exchange, co-training with text data leads to significantly improved language understanding performance (see Sec. 4.5)." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 233, + 546, + 425 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 233, + 546, + 425 + ], + "spans": [ + { + "bbox": [ + 304, + 233, + 546, + 425 + ], + "type": "text", + "content": "VQA In Fig. 2 we report the VQAv2 score of our models and baselines. It can be seen that CLIPPO outperforms " + }, + { + "bbox": [ + 304, + 233, + 546, + 425 + ], + "type": "inline_equation", + "content": "\\mathrm{CLIP^{*}}" + }, + { + "bbox": [ + 304, + 233, + 546, + 425 + ], + "type": "text", + "content": ", 1T-CLIP, as well as a pretrained ViT-B/16 from [16] by a significant margin, achieving a score of 66.3, and constraining with " + }, + { + "bbox": [ + 304, + 233, + 546, + 425 + ], + "type": "inline_equation", + "content": "25\\%" + }, + { + "bbox": [ + 304, + 233, + 546, + 425 + ], + "type": "text", + "content": " C4 data leads to a slight improvement of the score. The improved score of CLIPPO is mainly due to better performance in \"yes/no\" questions. Increasing the model size to L/16 adds another 2 points which originate from improvements in the \"number\" and \"other\" VQAv2 categories. However, note that for an L/16 architecture 1T-CLIP performs competitively with CLIPPO (see Table 7). One possible explanation for this could be that 1T-CLIP develops better OCR capabilities thanks to the higher model capacity (alt-texts can correlate with text in images/scene text, see [10, Fig. 3]). Increasing the resolution to 384px adds 2 to 3 points across models." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 426, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 426, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 426, + 546, + 713 + ], + "type": "text", + "content": "We also compare CLIPPO with baselines from the literature. Specifically, [17] proposes a framework (called ME-TER) for multimodal tasks, where pretrained transformer-based image and text encoders are combined with a transformer-based fusion module. CLIPPO L/16 achieves performance competitive with their model combining a CLIP B/32 vision backbone with a BERT-Base language backbone, which is roughly comparable in size and computational cost with our L/16 models. Another related work is [67], which combines different CLIP vision backbones with two existing VQA systems, Pythia [33] and MCAN [84]. CLIPPO outperforms different CLIP ViT-based Pythia and MCAN models from [67]. Note, however, that ResNet-based CLIP backbones lead to better results when combined with these systems. We further note that both [17] and [67] also investigate training their models on a mix of different image-text data sets with multiple objectives such as grounded masked language modeling and text-image matching, before transferring to the VQA task, which leads to significant improvements. ViLT [36] relies on such a strategy to train a single transformer backbone jointly encoding image and text tokens. At 384px resolution, CLIPPO (with " + }, + { + "bbox": [ + 304, + 426, + 546, + 713 + ], + "type": "inline_equation", + "content": "25\\%" + }, + { + "bbox": [ + 304, + 426, + 546, + 713 + ], + "type": "text", + "content": " C4 data) obtains a VQA score comparable with that of ViLT (and other models from the" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "11010" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 109 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 109 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 109 + ], + "type": "text", + "content": "literature such as ViLBERT [48], VisualBERT [43], and PixelBERT [30]), despite only using a contrastive objective for pretraining." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 117, + 275, + 131 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 117, + 275, + 131 + ], + "spans": [ + { + "bbox": [ + 47, + 117, + 275, + 131 + ], + "type": "text", + "content": "4.4. Multilingual vision-language understanding" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 136, + 289, + 387 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 136, + 289, + 387 + ], + "spans": [ + { + "bbox": [ + 46, + 136, + 289, + 387 + ], + "type": "text", + "content": "For typical language models, tokenizer choice can be a challenging process [78]. Commonly used English-language tokenizers generalize poorly to non-latin scripts [85]. This can be alleviated by the use of larger, multilingual vocabularies, at the expense of very large parameter counts. CLIPPO bypasses this issue, removing any language-related bias stemming from unbalanced or restrictive tokenizers. We consider multilingual image/text retrieval on Crossmodal3600 and compare CLIPPO, trained on WebLI with multilingual alt-texts, against 1T-CLIP with a number of SentencePiece tokenizers; one trained from 300M WebLI multilingual alt-texts, English (T5-en) and multilingual (T5-all) tokenizers from T5 [57], and a multilingual tokenizer (mT5) from mT5 [79], all with a vocabulary size of 32,000. The results are shown in Fig. 4. On average, CLIPPO achieves comparable retrieval performance to these baselines. In the case of mT5, the use of extra data to create the specialized vocabulary can boosts performance above that of CLIPPO; the leveraging of such extra parameters and data in the multilingual context will be an interesting future direction for CLIPPO." + } + ] + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 50, + 395, + 164, + 491 + ], + "blocks": [ + { + "bbox": [ + 50, + 395, + 164, + 491 + ], + "lines": [ + { + "bbox": [ + 50, + 395, + 164, + 491 + ], + "spans": [ + { + "bbox": [ + 50, + 395, + 164, + 491 + ], + "type": "image", + "image_path": "c530045a6fa2378a022f1da0ae18ebf659dbe4382f0e14b00f1ed0bba3350633.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 501, + 165, + 589 + ], + "lines": [ + { + "bbox": [ + 46, + 501, + 165, + 589 + ], + "spans": [ + { + "bbox": [ + 46, + 501, + 165, + 589 + ], + "type": "text", + "content": "Figure 3. Tokenization efficiency analyzed in terms of the sequence length produced by a given method. CLIPPO produces smaller sequences for the majority of languages compared to 1T-CLIP with alternative tokenizers." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "bbox": [ + 182, + 390, + 287, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 182, + 390, + 287, + 605 + ], + "spans": [ + { + "bbox": [ + 182, + 390, + 287, + 605 + ], + "type": "text", + "content": "Tokenization efficiency If a tokenizer is well suited to a particular dataset, it will tokenize to shorter sequences—this is especially the case when byte fallback [39] is enabled. SentencePiece tokenizers have the advantageous ability to tokenize entire—possibly quite long—words to single tokens. CLIPPO cannot learn any such compression, but benefits from equal treatment of all languages and" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 605, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 289, + 715 + ], + "type": "text", + "content": "words: it will by definition generalize equally well to all data, as its tokenization schema has not been trained on a specific dataset. We analyze 20,000 samples for each of the 104 C4 languages. Each CLIPPO token is assumed to be a " + }, + { + "bbox": [ + 46, + 605, + 289, + 715 + ], + "type": "inline_equation", + "content": "16 \\times 16" + }, + { + "bbox": [ + 46, + 605, + 289, + 715 + ], + "type": "text", + "content": " patch; though in typical computations all approaches considered here would pad to a fixed length, we compute CLIPPO's sequence length according to the last patch which contains rendered text. Fig. 3 shows the fraction of C4 languages where CLIPPO processes tokens" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 309, + 71, + 387, + 144 + ], + "blocks": [ + { + "bbox": [ + 309, + 71, + 387, + 144 + ], + "lines": [ + { + "bbox": [ + 309, + 71, + 387, + 144 + ], + "spans": [ + { + "bbox": [ + 309, + 71, + 387, + 144 + ], + "type": "image", + "image_path": "aff7e6aca6b658507c2c2b72c506fc8c84f425347c964bc30c587143bb11692f.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 149, + 545, + 216 + ], + "lines": [ + { + "bbox": [ + 304, + 149, + 545, + 216 + ], + "spans": [ + { + "bbox": [ + 304, + 149, + 545, + 216 + ], + "type": "text", + "content": "Figure 4. Zero-shot image/text retrieval performance on CrossModal3600 [70]. Although specialized (mc4) tokenizers can be leveraged to improve multilingual performance CLIPPO (dashed black line) broadly matches or exceeds comparable 1T-CLIP models trained with vocabulary size 32,000 (the word embeddings result in a " + }, + { + "bbox": [ + 304, + 149, + 545, + 216 + ], + "type": "inline_equation", + "content": "27\\%" + }, + { + "bbox": [ + 304, + 149, + 545, + 216 + ], + "type": "text", + "content": " increase in parameter count compared to CLIPPO)." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 387, + 72, + 463, + 144 + ], + "blocks": [ + { + "bbox": [ + 387, + 72, + 463, + 144 + ], + "lines": [ + { + "bbox": [ + 387, + 72, + 463, + 144 + ], + "spans": [ + { + "bbox": [ + 387, + 72, + 463, + 144 + ], + "type": "image", + "image_path": "7517525aa45ff1bdaec77d872d46a3f728890ca88b6860747cd07f33d286697b.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 464, + 72, + 544, + 143 + ], + "blocks": [ + { + "bbox": [ + 464, + 72, + 544, + 143 + ], + "lines": [ + { + "bbox": [ + 464, + 72, + 544, + 143 + ], + "spans": [ + { + "bbox": [ + 464, + 72, + 544, + 143 + ], + "type": "image", + "image_path": "ca91c6e84a119e7ebe86b0050fec9439ddfee621ba860c69c656b9b675a62c4e.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 229, + 545, + 314 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 229, + 545, + 314 + ], + "spans": [ + { + "bbox": [ + 304, + 229, + 545, + 314 + ], + "type": "text", + "content": "more efficiently than the vocabularies discussed above. We conservatively define \"more efficient\" as producing a shorter token sequence for over " + }, + { + "bbox": [ + 304, + 229, + 545, + 314 + ], + "type": "inline_equation", + "content": "75\\%" + }, + { + "bbox": [ + 304, + 229, + 545, + 314 + ], + "type": "text", + "content": " of examples. Even so, CLIPPO is indeed more efficient across the majority of languages. Per-language breakdowns of multilingual retrieval performance and tokenization efficiency are further discussed in Appendix C.3." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 323, + 446, + 335 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 323, + 446, + 335 + ], + "spans": [ + { + "bbox": [ + 306, + 323, + 446, + 335 + ], + "type": "text", + "content": "4.5. Language understanding" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 342, + 545, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 342, + 545, + 581 + ], + "spans": [ + { + "bbox": [ + 304, + 342, + 545, + 581 + ], + "type": "text", + "content": "Table 2 shows the GLUE benchmark results of CLIPPO and baselines. One can observe that CLIPPO trained on WebLI performs competitively with the BiLSTM+Attn+ELMo baseline which relies on deep word embeddings trained on a large language corpus. Also, it can be seen that CLIPPO along with 1T-CLIP outperform the language encoder trained using standard contrastive language vision pretraining " + }, + { + "bbox": [ + 304, + 342, + 545, + 581 + ], + "type": "inline_equation", + "content": "(\\mathrm{CLIP}^{*})" + }, + { + "bbox": [ + 304, + 342, + 545, + 581 + ], + "type": "text", + "content": ". This indicates that multimodal training in a single encoder benefits language understanding. Furthermore, CLIPPO achieves a much higher GLUE score than the " + }, + { + "bbox": [ + 304, + 342, + 545, + 581 + ], + "type": "inline_equation", + "content": "\\mathrm{CLIP}^*" + }, + { + "bbox": [ + 304, + 342, + 545, + 581 + ], + "type": "text", + "content": " image encoder, which in turn leads to significantly better results than fine-tuning a ViT-B/16 from scratch on GLUE (see Appendix C.2 for additional results). Unsurprisingly, the models pretrained on WebLI cannot do better than random guessing on the CoLA evaluation which requires to assess the grammatical correctness of sentences (recall that alt-texts are rarely grammatical sentences). Also the accuracy of " + }, + { + "bbox": [ + 304, + 342, + 545, + 581 + ], + "type": "inline_equation", + "content": "\\mathrm{CLIP}^*" + }, + { + "bbox": [ + 304, + 342, + 545, + 581 + ], + "type": "text", + "content": " and 1T-CLIP vision encoders we observe for SST-2 is in agreement with what was reported in [56, Table 10] for CLIP with a ViT-B/16 image encoder." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 582, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 582, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 582, + 545, + 714 + ], + "type": "text", + "content": "Adding sentence pairs form the C4 corpus gradually improves the GLUE score, and when half of the examples are sentence pairs our model becomes competitive with PILXEL, while still retaining decent image and vision-language understanding capabilities (cf. Table 1). Note, however, that there is a trade-off between language-only tasks and tasks that involve image understanding. Finally, training CLIPPO only on sentence pairs leads to a model which outperforms PILXEL by a significant margin. However, our model has seen more sentence pairs than PILXEL, so PILXEL might improve as well when training longer." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "11011" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 66, + 70, + 523, + 231 + ], + "blocks": [ + { + "bbox": [ + 66, + 70, + 523, + 231 + ], + "lines": [ + { + "bbox": [ + 66, + 70, + 523, + 231 + ], + "spans": [ + { + "bbox": [ + 66, + 70, + 523, + 231 + ], + "type": "table", + "html": "
training datasetMNLI-M/MMQQPQNLISST-2COLASTS-BMRPCRTEavg
BERT-BaseWiki + BC84.0 / 84.287.691.092.660.388.890.269.583.1
PIXELWiki + BC78.1 / 78.984.587.889.638.481.188.260.576.3
BiLSTM66.7 / 66.782.077.087.517.672.085.158.568.1
BiLSTM+Attn, ELMo72.4 / 72.483.675.291.544.156.182.152.770.0
CLIP* img enc.WebLI66.4 / 67.578.669.478.60.05.281.252.755.5
CLIP* text enc.WebLI71.8 / 72.582.773.086.26.665.081.453.865.9
1T-CLIP text enc.WebLI72.6 / 73.083.880.784.90.079.683.357.068.3
CLIPPOWebLI73.0 / 72.684.381.286.81.880.584.153.468.6
CLIPPOWebLI + 25%C477.7 / 77.285.383.190.928.283.484.559.274.4
CLIPPOWebLI + 50%C479.2 / 79.286.484.292.938.983.484.859.976.6
CLIPPOC479.9 / 80.286.785.293.350.984.786.358.578.4
CLIPPO L/16WebLI + 25%C476.6 / 75.587.179.993.248.284.184.656.076.1
CLIPPO L/16WebLI + 50%C482.3 / 82.487.986.794.255.385.885.959.280.0
", + "image_path": "9853585471a5f44fb437f10d85e68ef027f5bbf01b43e6e573e0d9739b05fc13.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 233, + 547, + 301 + ], + "lines": [ + { + "bbox": [ + 46, + 233, + 547, + 301 + ], + "spans": [ + { + "bbox": [ + 46, + 233, + 547, + 301 + ], + "type": "text", + "content": "Table 2. Results for the GLUE benchmark (dev set). The metric is accuracy except for the performance on QQP and MRPC, which is measured using the " + }, + { + "bbox": [ + 46, + 233, + 547, + 301 + ], + "type": "inline_equation", + "content": "F_{1}" + }, + { + "bbox": [ + 46, + 233, + 547, + 301 + ], + "type": "text", + "content": " score, CoLA which uses Matthew's correlation, and STS-B which evaluated based on Spearman's correlation coefficient. \"avg\" corresponds to the average across all metrics. The results for BERT-Base and PILXEL are from [60, Table 3], and BiLSTM and BiLSTM+Attn, ELMo from [73, Table 6]. All encoders considered here have a transformer architecture comparable to BERT-Base (up to the text embedding layer), except for CLIPPO L/16 which uses a ViT L/16, and the two BiLSTM model variants. Wiki and BC stand for (English) Wikipedia and Bookcorpus [86] data, respectively." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 47, + 309, + 175, + 322 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 309, + 175, + 322 + ], + "spans": [ + { + "bbox": [ + 47, + 309, + 175, + 322 + ], + "type": "text", + "content": "4.6. Ablations and analysis" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 331, + 288, + 487 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 331, + 288, + 487 + ], + "spans": [ + { + "bbox": [ + 46, + 331, + 288, + 487 + ], + "type": "text", + "content": "Impact of weight sharing across modalities The fact that CLIPPO 1) uses a shared patch embedding for regular images and text images and 2) this embedding has considerably fewer parameters than the text embedding of 1T-CLIP and " + }, + { + "bbox": [ + 46, + 331, + 288, + 487 + ], + "type": "inline_equation", + "content": "\\mathrm{CLIP^{*}}" + }, + { + "bbox": [ + 46, + 331, + 288, + 487 + ], + "type": "text", + "content": " provokes the question of whether CLIPPO could benefit from separate patch embeddings for text images and regular images. Further, CLIPPO relies on a single head to compute the output representation for images and text, and relaxing this constraint by using separate heads for the two modalities could lead to more expressive representations. The results (deferred to Appendix D.1) show that neither of these variants lead to improved image classification or retrieval metrics compared to CLIPPO." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 490, + 287, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 490, + 287, + 552 + ], + "spans": [ + { + "bbox": [ + 46, + 490, + 287, + 552 + ], + "type": "text", + "content": "Impact of the text location We test whether rendering the question at the top, middle, or bottom of the image impacts the VQA performance of CLIPPO and find that it does not, provided that we increase the learning rate of the positional embedding during fine-tuning (see Appendix D.2)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 554, + 287, + 650 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 554, + 287, + 650 + ], + "spans": [ + { + "bbox": [ + 46, + 554, + 287, + 650 + ], + "type": "text", + "content": "Typographic attacks Since CLIPPO is trained on large amounts of rendered (alt-)text it is important to check whether it becomes more susceptible to typographic attacks—the tendency of CLIP-style models to zero-shot classify an image according to adversarially injected scene text unrelated to the scene [23,42,50]. In Appendix D.3 we present results indicating that CLIPPO is no more vulnerable to typographic attacks than 1T-CLIP and " + }, + { + "bbox": [ + 46, + 554, + 287, + 650 + ], + "type": "inline_equation", + "content": "\\mathrm{CLIP^{*}}" + }, + { + "bbox": [ + 46, + 554, + 287, + 650 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "type": "text", + "content": "Modality gap Liang et al. [45] discovered that text and image embeddings of CLIP-style models form two distinct clusters rather than both filling the embedding space densely and occupying the same spatial region. They attribute this phenomenon to a combination of initialization" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 309, + 545, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 309, + 545, + 514 + ], + "spans": [ + { + "bbox": [ + 304, + 309, + 545, + 514 + ], + "type": "text", + "content": "conditions and properties of the loss function/training dynamics. Since we consider single tower models here, and also co-train some of these models with text-only pairs it is interesting to see how this affects the modality gap. We compute the gap and visualize it following the recipe from [45] in Fig. 5 (see Appendices D.4 and D.5 for additional visualizations). CLIPPO attains a slightly lower modality gap than " + }, + { + "bbox": [ + 304, + 309, + 545, + 514 + ], + "type": "inline_equation", + "content": "\\mathrm{CLIP}^*" + }, + { + "bbox": [ + 304, + 309, + 545, + 514 + ], + "type": "text", + "content": ", but clearly features a clustering structure for image and text embeddings. However, when training contrastively with sentence pairs in addition to image/alt-text pairs, the clustering structure disappears, the image and text embeddings overlap, and the modality gap decreases significantly. A possible explanation for this behavior could be that the additional learning pressure induced by the contrastive loss on sentence pairs encourages text embeddings to spread out more and hence the structure of all embeddings changes." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 515, + 545, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 515, + 545, + 624 + ], + "spans": [ + { + "bbox": [ + 304, + 515, + 545, + 624 + ], + "type": "text", + "content": "Text/text co-training objectives To corroborate that contrastive NSP is a sensible objective to improve language understanding in the context of CLIPPO, we train CLIPPO without any image/alt-text data on pairs of parallel translated sentences (this is straight-forward in our framework since visual text is language-agnostic), as well as English back-translated data, and evaluate the resulting text representations on GLUE. Table 3 shows that NSP on C4 clearly achieves the highest GLUE score." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 633, + 455, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 633, + 455, + 647 + ], + "spans": [ + { + "bbox": [ + 305, + 633, + 455, + 647 + ], + "type": "text", + "content": "5. Discussion and limitations" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 653, + 545, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 653, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 653, + 545, + 715 + ], + "type": "text", + "content": "We proposed and evaluated CLIPPO which produces a single ViT that can understand images and language jointly using images as a sole input modality. Perhaps surprisingly, CLIPPO matches the performance of the 1T-CLIP baseline across many of the considered tasks, and only in" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "11012" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 75, + 70, + 258, + 102 + ], + "blocks": [ + { + "bbox": [ + 75, + 70, + 258, + 102 + ], + "lines": [ + { + "bbox": [ + 75, + 70, + 258, + 102 + ], + "spans": [ + { + "bbox": [ + 75, + 70, + 258, + 102 + ], + "type": "table", + "html": "
WMT19WMT19 BTC4 NSP
GLUE score61.266.677.6
", + "image_path": "4eb700574c3530918cc0b2dd483b0b1148395dec2156bcc6c3c84f45af2984be.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 110, + 287, + 166 + ], + "lines": [ + { + "bbox": [ + 46, + 110, + 287, + 166 + ], + "spans": [ + { + "bbox": [ + 46, + 110, + 287, + 166 + ], + "type": "text", + "content": "Table 3. Ablation of text pair-based contrastive co-training tasks: Training on parallel translated sentences (WMT19), training on parallel back-translated sentences (WMT19 BT), and NSP for sentences sampled from C4 (C4 NSP). C4 NSP leads to the highest GLUE score by a large margin." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 188, + 286, + 283 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 188, + 286, + 283 + ], + "spans": [ + { + "bbox": [ + 46, + 188, + 286, + 283 + ], + "type": "text", + "content": "curs a minor drop compared to the " + }, + { + "bbox": [ + 46, + 188, + 286, + 283 + ], + "type": "inline_equation", + "content": "\\mathrm{CLIP^{*}}" + }, + { + "bbox": [ + 46, + 188, + 286, + 283 + ], + "type": "text", + "content": " baseline, despite having less than half the parameters of " + }, + { + "bbox": [ + 46, + 188, + 286, + 283 + ], + "type": "inline_equation", + "content": "\\mathrm{CLIP^{*}}" + }, + { + "bbox": [ + 46, + 188, + 286, + 283 + ], + "type": "text", + "content": ". As we showed, the image-only interface enables a simple, unified data pipeline for training on and transferring to mixed modalities. CLIPPO opens the door for additional modalities (e.g. spectrograms) and, as we hope, might inspire applications of pixel-only models beyond contrastive training. Nevertheless, several limitations remain, as discussed next." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 288, + 286, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 288, + 286, + 430 + ], + "spans": [ + { + "bbox": [ + 46, + 288, + 286, + 430 + ], + "type": "text", + "content": "Co-training First, to achieve language understanding performance competitive with PILXEL and BERT on GLUE, contrastive co-training with text pairs is necessary. While adding " + }, + { + "bbox": [ + 46, + 288, + 286, + 430 + ], + "type": "inline_equation", + "content": "25\\%" + }, + { + "bbox": [ + 46, + 288, + 286, + 430 + ], + "type": "text", + "content": " C4 data to the batch seems to strike a good balance across all tasks considered, it does induce a nonnegligible drop in zero-shot image classification and image/text retrieval. This drop becomes more severe as the fraction of C4 examples increases. We observed an associated change in modality gap, and further investigation of the representation in the context of co-training might help to develop models that achieve better overall performance in the co-training setup." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 434, + 286, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 434, + 286, + 590 + ], + "spans": [ + { + "bbox": [ + 46, + 434, + 286, + 590 + ], + "type": "text", + "content": "Diverse rendered text CLIPPO currently relies on cleanly rendered text as an input and its capabilities to handle text from documents or web pages without further adaption is limited (besides the basic OCR capabilities that CLIP-style models learn from image/alt-text pairs). We emphasize that sophisticated OCR and document understanding is not a goal of this paper, and training CLIPPO with augmented noisy rendered text that mimics the distribution of documents and websites is likely to lead to worse performance across the considered tasks, since image/alt-text pairs are less correlated and provide a weaker learning signal. However, developing CLIPPO further to handle less clean visual text will open many additional applications." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 594, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 286, + 713 + ], + "type": "text", + "content": "Generative modeling CLIPPO, like CLIP, BERT, PILXEL and many other models, uses an encoder-only design and hence lacks the ability to generate text outputs. A common approach to equip encoder-only models with generation capabilities (e.g., for image captioning or VQA) is to simply combine them with a (potentially pretrained) language model [8, 76]. This approach naturally also applies to CLIPPO and PILXEL, but defeats the advantages of visual text in certain (e.g. multilingual) scenarios. While visual text outputs have previously been explored in the context of" + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 310, + 66, + 543, + 156 + ], + "blocks": [ + { + "bbox": [ + 310, + 66, + 543, + 156 + ], + "lines": [ + { + "bbox": [ + 310, + 66, + 543, + 156 + ], + "spans": [ + { + "bbox": [ + 310, + 66, + 543, + 156 + ], + "type": "image", + "image_path": "58db6c461caeb2536f7513e0b50b180c5916ddaba4a1c4a68466f01496090d94.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 163, + 545, + 241 + ], + "lines": [ + { + "bbox": [ + 305, + 163, + 545, + 241 + ], + "spans": [ + { + "bbox": [ + 305, + 163, + 545, + 241 + ], + "type": "text", + "content": "Figure 5. Visualization of the modality gap for " + }, + { + "bbox": [ + 305, + 163, + 545, + 241 + ], + "type": "inline_equation", + "content": "\\mathrm{CLIP^{*}}" + }, + { + "bbox": [ + 305, + 163, + 545, + 241 + ], + "type": "text", + "content": " and CLIPPO optionally trained with " + }, + { + "bbox": [ + 305, + 163, + 545, + 241 + ], + "type": "inline_equation", + "content": "25\\%" + }, + { + "bbox": [ + 305, + 163, + 545, + 241 + ], + "type": "text", + "content": " C4 data. The visualization follows the analysis from [45] and shows embedded images (blue dots) and corresponding alt-text (orange dots) from the WebLI validation set, projected to the first two principal components of the validation data matrix. CLIPPO has a slightly smaller modality gap than " + }, + { + "bbox": [ + 305, + 163, + 545, + 241 + ], + "type": "inline_equation", + "content": "\\mathrm{CLIP^{*}}" + }, + { + "bbox": [ + 305, + 163, + 545, + 241 + ], + "type": "text", + "content": "; co-training with C4 data strongly reduces the gap." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 257, + 545, + 280 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 257, + 545, + 280 + ], + "spans": [ + { + "bbox": [ + 305, + 257, + 545, + 280 + ], + "type": "text", + "content": "machine translation [49], it remains unclear what a scalable tokenizer-free way to generate text is." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 284, + 545, + 427 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 284, + 545, + 427 + ], + "spans": [ + { + "bbox": [ + 305, + 284, + 545, + 427 + ], + "type": "text", + "content": "Multilingual learning Finally, we showed that CLIPPO obtains strong multilingual image/text retrieval performance without requiring the development of an appropriate tokenizer. For fine-grained adjustment and balancing of the retrieval performance further steps will be necessary, including data balancing and potentially co-training with multi-lingual text data. Furthermore, similar to PILXEL, CLIPPO relies on certain ad-hoc design choices w.r.t. the visual representation, for example the left-to-right rendering of Arabic scripts. This approach leads to decent performance on average, but it is not clear what kind of unwanted effects it introduces and how these could be mitigated." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 438, + 378, + 451 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 438, + 378, + 451 + ], + "spans": [ + { + "bbox": [ + 306, + 438, + 378, + 451 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 459, + 545, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 459, + 545, + 590 + ], + "spans": [ + { + "bbox": [ + 304, + 459, + 545, + 590 + ], + "type": "text", + "content": "We introduced CLIPPO, a model for processing images and text jointly through the lens of vision. This reduces design choices (in particular w.r.t. tokenization) and parameter count, simplifies data processing pipelines and transfer recipes, and increases generality across multiple languages. We also explored methods of enhancing language understanding, where traditional image/alt-text contrastive models trained on web data fall short. We demonstrated this is possible by co-training with text pairs, with CLIPPO models outperforming strong NLP baselines while maintaining solid image understanding capabilities." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 591, + 545, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 591, + 545, + 662 + ], + "spans": [ + { + "bbox": [ + 304, + 591, + 545, + 662 + ], + "type": "text", + "content": "Although we presented a unified contrastive training algorithm, CLIPPO suffers somewhat when co-training on multiple tasks, and future work to harmonize the co-training could enhance the models significantly. Deeper understanding of the design choices in rendering text as images, and their impact on performance, is another interesting avenue." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 666, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 666, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 666, + 545, + 713 + ], + "type": "text", + "content": "Acknowledgments We would like to thank Lucas Beyer, Josip Djolonga, Alexander Kolesnikov, Mario Lucic, Andreas Steiner, Xiao Wang, and Xiaohua Zhai for inspiring discussions and helpful feedback. We also thank Jeffrey Sorensen for help with the text rendering preprocessing." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "11013" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 134 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 134 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 134 + ], + "type": "text", + "content": "[1] Hassan Akbari, Liangzhe Yuan, Rui Qian, Wei-Hong Chuang, Shih-Fu Chang, Yin Cui, and Boqing Gong. VATT: transformers for multimodal self-supervised learning from raw video, audio and text. In NeurIPS, 2021. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 136, + 288, + 245 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 136, + 288, + 245 + ], + "spans": [ + { + "bbox": [ + 53, + 136, + 288, + 245 + ], + "type": "text", + "content": "[2] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katie Millican, Malcolm Reynolds, Roman Ring, Eliza Rutherford, Serkan Cabi, Tengda Han, Zhitao Gong, Sina Samangooei, Marianne Monteiro, Jacob Menick, Sebastian Borgeaud, Andrew Brock, Aida Nematzadeh, Sahand Sharifzadeh, Mikolaj Binkowski, Ricardo Barreira, Oriol Vinyals, Andrew Zisserman, and Karen Simonyan. Flamingo: A visual language model for few-shot learning. In NeurIPS, 2022. 1, 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 246, + 287, + 289 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 246, + 287, + 289 + ], + "spans": [ + { + "bbox": [ + 53, + 246, + 287, + 289 + ], + "type": "text", + "content": "[3] Srikar Appalaraju, Bhavan Jasani, Bhargava Urala Kota, Yusheng Xie, and R. Manmatha. DocFormer: End-to-end transformer for document understanding. In ICCV, pages 973-983, 2021. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 291, + 287, + 323 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 291, + 287, + 323 + ], + "spans": [ + { + "bbox": [ + 53, + 291, + 287, + 323 + ], + "type": "text", + "content": "[4] Lucas Beyer, Xiaohua Zhai, and Alexander Kolesnikov. Big Vision. https://github.com/google-research/big visions, 2022.1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 324, + 287, + 379 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 324, + 287, + 379 + ], + "spans": [ + { + "bbox": [ + 53, + 324, + 287, + 379 + ], + "type": "text", + "content": "[5] Zalan Borsos, Raphael Marinier, Damien Vincent, Eugene Kharitonov, Olivier Pietquin, Matthew Sharifi, Olivier Teboul, David Grangier, Marco Tagliasacchi, and Neil Zeghidour. AudioLM: a language modeling approach to audio generation. CoRR, abs/2209.03143, 2022. 1, 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 380, + 287, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 380, + 287, + 445 + ], + "spans": [ + { + "bbox": [ + 53, + 380, + 287, + 445 + ], + "type": "text", + "content": "[6] Tom B. Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert-Voss, Gretchen Krueger, Tom Henighan, et al. Language models are few-shot learners. In NeurIPS, 2020. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 447, + 287, + 501 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 447, + 287, + 501 + ], + "spans": [ + { + "bbox": [ + 53, + 447, + 287, + 501 + ], + "type": "text", + "content": "[7] Lili Chen, Kevin Lu, Aravind Rajeswaran, Kimin Lee, Aditya Grover, Michael Laskin, Pieter Abbeel, Aravind Srinivas, and Igor Mordatch. Decision transformer: Reinforcement learning via sequence modeling. In NeurIPS, pages 15084-15097, 2021. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 502, + 287, + 535 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 502, + 287, + 535 + ], + "spans": [ + { + "bbox": [ + 53, + 502, + 287, + 535 + ], + "type": "text", + "content": "[8] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey E. Hinton. A simple framework for contrastive learning of visual representations. In ICML, 2020. 8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 536, + 287, + 579 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 536, + 287, + 579 + ], + "spans": [ + { + "bbox": [ + 53, + 536, + 287, + 579 + ], + "type": "text", + "content": "[9] Xinlei Chen, Hao Fang, Tsung-Yi Lin, Ramakrishna Vedantam, Saurabh Gupta, Piotr Dólar, and C. Lawrence Zitnick. Microsoft COCO captions: Data collection and evaluation server. CoRR, abs/1504.00325, 2015. 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 581, + 287, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 581, + 287, + 690 + ], + "spans": [ + { + "bbox": [ + 48, + 581, + 287, + 690 + ], + "type": "text", + "content": "[10] Xi Chen, Xiao Wang, Soravit Changpinyo, A. J. Piergiovanni, Piotr Padlewski, Daniel Salz, Sebastian Goodman, Adam Grycner, Basil Mustafa, Lucas Beyer, Alexander Kolesnikov, Joan Puigcerver, Nan Ding, Keran Rong, Hassan Akbari, Gaurav Mishra, Linting Xue, Ashish Thapliyal, James Bradbury, Weicheng Kuo, Mojtaba Seyedhosseini, Chao Jia, Burcu Karagol Ayan, Carlos Riquelme, Andreas Steiner, Anelia Angelova, Xiaohua Zhai, Neil Houlsby, and Radu Soricut. PaLI: A jointly-scaled multilingual language-image model. In ICLR, 2023. 1, 2, 4, 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 691, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 287, + 712 + ], + "type": "text", + "content": "[11] Aakanksha Chowdhery, Sharan Narang, Jacob Devlin, Maarten Bosma, Gaurav Mishra, Adam Roberts, Paul" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 23, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 281 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 281 + ], + "type": "text", + "content": "Barham, Hyung Won Chung, Charles Sutton, Sebastian Gehrmann, Parker Schuh, Kensen Shi, Sasha Tsvyashchenko, Joshua Maynez, Abhishek Rao, Parker Barnes, Yi Tay, Noam Shazeer, Vinodkumar Prabhakaran, Emily Reif, Nan Du, Ben Hutchinson, Reiner Pope, James Bradbury, Jacob Austin, Michael Isard, Guy Gur-Ari, Pengcheng Yin, Toju Duke, Anselm Levskaya, Sanjay Ghemawat, Sunipa Dev, Henryk Michalewski, Xavier Garcia, Vedant Misra, Kevin Robinson, Liam Fedus, Denny Zhou, Daphne Ippolito, David Luan, Hyeontaek Lim, Barret Zoph, Alexander Spiridonov, Ryan Sepassi, David Dohan, Shivani Agrawal, Mark Omernick, Andrew M. Dai, Thanumalayan Sankaranarayana Pillai, Marie Pellat, Aitor Lewkowycz, Erica Moreira, Rewon Child, Oleksandr Polozov, Katherine Lee, Zongwei Zhou, Xuezhi Wang, Brennan Saeta, Mark Diaz, Orhan Firat, Michele Catasta, Jason Wei, Kathy Meier-Hellstern, Douglas Eck, Jeff Dean, Slav Petrov, and Noah Fiedel. PaLM: Scaling language modeling with pathways. CoRR, abs/2204.02311, 2022. 1" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 307, + 283, + 545, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 283, + 545, + 326 + ], + "spans": [ + { + "bbox": [ + 307, + 283, + 545, + 326 + ], + "type": "text", + "content": "[12] Geoffrey Cideron, Sertan Girgin, Anton Raichuk, Olivier Pietquin, Olivier Bachem, and Léonard Hussenot. vec2text with round-trip translations. CoRR, abs/2209.06792, 2022. 4" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 328, + 545, + 384 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 328, + 545, + 384 + ], + "spans": [ + { + "bbox": [ + 307, + 328, + 545, + 384 + ], + "type": "text", + "content": "[13] Alexis Conneau, Kartikay Khandelwal, Naman Goyal, Vishrav Chaudhary, Guillaume Wenzek, Francisco Guzmán, Edouard Grave, Myle Ott, Luke Zettlemoyer, and Veselin Stoyanov. Unsupervised cross-lingual representation learning at scale. In ACL, pages 8440–8451, 2020. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 385, + 545, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 385, + 545, + 418 + ], + "spans": [ + { + "bbox": [ + 307, + 385, + 545, + 418 + ], + "type": "text", + "content": "[14] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In CVPR, 2009. 4" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 419, + 545, + 463 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 419, + 545, + 463 + ], + "spans": [ + { + "bbox": [ + 307, + 419, + 545, + 463 + ], + "type": "text", + "content": "[15] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. BERT: pre-training of deep bidirectional transformers for language understanding. In NAACL-HLT, 2019, 1, 2, 4, 5" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 465, + 545, + 530 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 465, + 545, + 530 + ], + "spans": [ + { + "bbox": [ + 307, + 465, + 545, + 530 + ], + "type": "text", + "content": "[16] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth " + }, + { + "bbox": [ + 307, + 465, + 545, + 530 + ], + "type": "inline_equation", + "content": "16 \\times 16" + }, + { + "bbox": [ + 307, + 465, + 545, + 530 + ], + "type": "text", + "content": " words: Transformers for image recognition at scale. In ICLR, 2021. 1, 2, 3, 4, 5, 14, 16, 17, 24" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 532, + 545, + 587 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 532, + 545, + 587 + ], + "spans": [ + { + "bbox": [ + 307, + 532, + 545, + 587 + ], + "type": "text", + "content": "[17] Zi-Yi Dou, Yichong Xu, Zhe Gan, Jianfeng Wang, Shuohang Wang, Lijuan Wang, Chenguang Zhu, Pengchuan Zhang, Lu Yuan, Nanyun Peng, Zicheng Liu, and Michael Zeng. An empirical study of training end-to-end vision-and-language transformers. In CVPR, pages 18145-18155, 2022. 5, 14, 17" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 589, + 545, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 589, + 545, + 632 + ], + "spans": [ + { + "bbox": [ + 307, + 589, + 545, + 632 + ], + "type": "text", + "content": "[18] Wikipedia Foundation. ACL 2019 Fourth Conference on Machine Translation (WMT19), Shared Task: Machine Translation of News. http://www.statmt.org/wmt19/translation-task.html.4" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 635, + 545, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 635, + 545, + 667 + ], + "spans": [ + { + "bbox": [ + 307, + 635, + 545, + 667 + ], + "type": "text", + "content": "[19] Tianyu Gao, Xingcheng Yao, and Danqi Chen. Simcse: Simple contrastive learning of sentence embeddings. In EMNLP, pages 6894-6910, 2021. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 669, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 669, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 307, + 669, + 545, + 712 + ], + "type": "text", + "content": "[20] John M. Giorgi, Osvald Nitski, Bo Wang, and Gary D. Bader. DeCLUTR: Deep contrastive learning for unsupervised textual representations. In ACL/IJCNLP, pages 879-895, 2021. 2" + } + ] + } + ], + "index": 22 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "text", + "content": "11014" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "text", + "content": "[21] Rohit Girdhar, Alaaeldin El-Nouby, Mannat Singh, Kalyan Vasudev Alwala, Armand Joulin, and Ishan Misra. OmniMAE: Single model masked pretraining on images and videos. CoRR, abs/2206.08356, 2022. 1, 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 118, + 287, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 118, + 287, + 161 + ], + "spans": [ + { + "bbox": [ + 48, + 118, + 287, + 161 + ], + "type": "text", + "content": "[22] Rohit Girdhar, Mannat Singh, Nikhila Ravi, Laurens van der Maaten, Armand Joulin, and Ishan Misra. Omnivore: A single model for many visual modalities. In CVPR, pages 16081-16091, 2022. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 163, + 287, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 163, + 287, + 206 + ], + "spans": [ + { + "bbox": [ + 48, + 163, + 287, + 206 + ], + "type": "text", + "content": "[23] Gabriel Goh, Nick Cammarata, Chelsea Voss, Shan Carter, Michael Petrov, Ludwig Schubert, Alec Radford, and Chris Olah. Multimodal neurons in artificial neural networks. *Distill*, 2021, 7, 21" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 209, + 287, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 209, + 287, + 239 + ], + "spans": [ + { + "bbox": [ + 48, + 209, + 287, + 239 + ], + "type": "text", + "content": "[24] Yuan Gong, Yu-An Chung, and James R. Glass. AST: audio spectrogram transformer. In Interspeech, pages 571-575, 2021. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 242, + 287, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 242, + 287, + 285 + ], + "spans": [ + { + "bbox": [ + 48, + 242, + 287, + 285 + ], + "type": "text", + "content": "[25] Yash Goyal, Tejas Khot, Douglas Summers-Stay, Dhruv Bartra, and Devi Parikh. Making the V in VQA matter: Elevating the role of image understanding in visual question answering. In CVPR, 2017. 4, 13" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 287, + 287, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 287, + 287, + 319 + ], + "spans": [ + { + "bbox": [ + 48, + 287, + 287, + 319 + ], + "type": "text", + "content": "[26] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollár, and Ross B. Girshick. Masked autoencoders are scalable vision learners. In CVPR, pages 15979-15988, 2022. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 321, + 287, + 375 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 321, + 287, + 375 + ], + "spans": [ + { + "bbox": [ + 48, + 321, + 287, + 375 + ], + "type": "text", + "content": "[27] Jonathan Ho, William Chan, Chitwan Saharia, Jay Whang, Ruiqi Gao, Alexey A. Gritsenko, Diederik P. Kingma, Ben Poole, Mohammad Norouzi, David J. Fleet, and Tim Salimans. Imagen video: High definition video generation with diffusion models. CoRR, abs/2210.02303, 2022. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 377, + 287, + 420 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 377, + 287, + 420 + ], + "spans": [ + { + "bbox": [ + 48, + 377, + 287, + 420 + ], + "type": "text", + "content": "[28] Po-Yao Huang, Hu Xu, Juncheng Li, Alexei Baevski, Michael Auli, Wojciech Galuba, Florian Metze, and Christoph Feichtenhofer. Masked autoencoders that listen. CoRR, abs/2207.06405, 2022. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 422, + 287, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 422, + 287, + 464 + ], + "spans": [ + { + "bbox": [ + 48, + 422, + 287, + 464 + ], + "type": "text", + "content": "[29] Yupan Huang, Tengchao Lv, Lei Cui, Yutong Lu, and Furu Wei. Layoutlmv3: Pre-training for document AI with unified text and image masking. In ACMMM, pages 4083-4091, 2022. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 468, + 287, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 468, + 287, + 510 + ], + "spans": [ + { + "bbox": [ + 48, + 468, + 287, + 510 + ], + "type": "text", + "content": "[30] Zhicheng Huang, Zhaoyang Zeng, Bei Liu, Dongmei Fu, and Jianlong Fu. Pixel-BERT: Aligning image pixels with text by deep multi-modal transformers. CoRR, abs/2004.00849, 2020.6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 512, + 287, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 512, + 287, + 588 + ], + "spans": [ + { + "bbox": [ + 48, + 512, + 287, + 588 + ], + "type": "text", + "content": "[31] Andrew Jaegle, Sebastian Borgeaud, Jean-Baptiste Alayrac, Carl Doersch, Catalin Ionescu, David Ding, Skanda Koppula, Daniel Zoran, Andrew Brock, Evan Shelhamer, Olivier J. Henaff, Matthew M. Botvinick, Andrew Zisserman, Oriol Vinyals, and João Carreira. Perceiver IO: A general architecture for structured inputs & outputs. In ICLR, 2022. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 590, + 287, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 590, + 287, + 634 + ], + "spans": [ + { + "bbox": [ + 48, + 590, + 287, + 634 + ], + "type": "text", + "content": "[32] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc V. Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In ICML, 2021. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 635, + 287, + 677 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 635, + 287, + 677 + ], + "spans": [ + { + "bbox": [ + 48, + 635, + 287, + 677 + ], + "type": "text", + "content": "[33] Yu Jiang, Vivek Natarajan, Xinlei Chen, Marcus Rohrbach, Dhruv Batra, and Devi Parikh. Pythia v0.1: The winning entry to the VQA challenge 2018. CoRR, abs/1807.09956, 2018. 5" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "type": "text", + "content": "[34] John Jumper, Richard Evans, Alexander Pritzel, Tim Green, Michael Figurnov, Olaf Ronneberger, Kathryn Tunyasuvunakool, Russ Bates, Augustin Žídek, Anna Potapenko, Alex" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 327, + 73, + 545, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 545, + 171 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 545, + 171 + ], + "type": "text", + "content": "Bridgland, Clemens Meyer, Simon A A Kohl, Andrew J Ballard, Andrew Cowie, Bernardino Romera-Paredes, Stanislav Nikolov, Rishub Jain, Jonas Adler, Trevor Back, Stig Petersen, David Reiman, Ellen Clancy, Michal Zielinski, Martin Steinegger, Michalina Pacholska, Tamas Berghammer, Sebastian Bodenstein, David Silver, Oriol Vinyals, Andrew W Senior, Koray Kavukcuoglu, Pushmeet Kohli, and Demis Hassabis. Highly accurate protein structure prediction with AlphaFold. Nature, 596(7873):583-589, 2021. 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 173, + 545, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 173, + 545, + 227 + ], + "spans": [ + { + "bbox": [ + 308, + 173, + 545, + 227 + ], + "type": "text", + "content": "[35] Geewook Kim, Teakgyu Hong, Moonbin Yim, JeongYeon Nam, Jinyoung Park, Jinyeong Yim, Wonseok Hwang, Sang-doo Yun, Dongyoon Han, and Seunghyun Park. OCR-Free document understanding transformer. In ECCV, pages 498- 517,2022.2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 229, + 545, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 229, + 545, + 262 + ], + "spans": [ + { + "bbox": [ + 308, + 229, + 545, + 262 + ], + "type": "text", + "content": "[36] Wonjae Kim, Bokyung Son, and Ildoo Kim. ViLT: Vision-and-language transformer without convolution or region supervision. In ICML, pages 5583-5594. PMLR, 2021. 5, 17" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 263, + 545, + 306 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 263, + 545, + 306 + ], + "spans": [ + { + "bbox": [ + 308, + 263, + 545, + 306 + ], + "type": "text", + "content": "[37] Alexander Kolesnikov, André Susano Pinto, Lucas Beyer, Xiaohua Zhai, Jeremiah Harmsen, and Neil Houlsby. UViM: A unified modeling approach for vision with learned guiding codes. In NeurIPS, 2022. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 308, + 545, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 308, + 545, + 341 + ], + "spans": [ + { + "bbox": [ + 308, + 308, + 545, + 341 + ], + "type": "text", + "content": "[38] Simon Kornblith, Mohammad Norouzi, Honglak Lee, and Geoffrey Hinton. Similarity of neural network representations revisited. In ICML, pages 3519-3529, 2019. 22" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 342, + 545, + 375 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 342, + 545, + 375 + ], + "spans": [ + { + "bbox": [ + 308, + 342, + 545, + 375 + ], + "type": "text", + "content": "[39] Taku Kudo and John Richardson. SentencePiece: A simple and language independent subword tokenizer and tokenizer for neural text processing. In EMNLP, 2018. 2, 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 376, + 545, + 420 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 376, + 545, + 420 + ], + "spans": [ + { + "bbox": [ + 308, + 376, + 545, + 420 + ], + "type": "text", + "content": "[40] Juho Lee, Yoonho Lee, Jungtaek Kim, Adam R. Kosiorek, Seungjin Choi, and Yee Whye Teh. Set transformer: A framework for attention-based permutation-invariant neural networks. In ICML, pages 3744-3753, 2019. 3, 14" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 422, + 545, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 422, + 545, + 475 + ], + "spans": [ + { + "bbox": [ + 308, + 422, + 545, + 475 + ], + "type": "text", + "content": "[41] Kenton Lee, Mandar Joshi, Iulia Turc, Hexiang Hu, Fangyu Liu, Julian Eisenschlos, Urvashi Khandelwal, Peter Shaw, Ming-Wei Chang, and Kristina Toutanova. Pix2struct: Screenshot parsing as pretraining for visual language understanding. CoRR, abs/2210.03347, 2022. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 478, + 545, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 478, + 545, + 521 + ], + "spans": [ + { + "bbox": [ + 308, + 478, + 545, + 521 + ], + "type": "text", + "content": "[42] Yoann Lemesle, Masataka Sawayama, Guillermo Valle-Perez, Maxime Adolphe, Hélène Sauzeon, and Pierre-Yves Oudeyer. Language-biased image classification: evaluation based on semantic representations. In ICLR, 2022. 7, 21" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 523, + 545, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 523, + 545, + 566 + ], + "spans": [ + { + "bbox": [ + 308, + 523, + 545, + 566 + ], + "type": "text", + "content": "[43] Liunian Harold Li, Mark Yatskar, Da Yin, Cho-Jui Hsieh, and Kai-Wei Chang. VisualBERT: A simple and performant baseline for vision and language. CoRR, abs/1908.03557, 2019. 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 568, + 545, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 568, + 545, + 612 + ], + "spans": [ + { + "bbox": [ + 308, + 568, + 545, + 612 + ], + "type": "text", + "content": "[44] Qing Li, Boqing Gong, Yin Cui, Dan Kondratyuk, Xianzhi Du, Ming-Hsuan Yang, and Matthew Brown. Towards a unified foundation model: Jointly pre-training transformers on unpaired images and text. CoRR, abs/2112.07074, 2021. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 613, + 545, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 613, + 545, + 656 + ], + "spans": [ + { + "bbox": [ + 308, + 613, + 545, + 656 + ], + "type": "text", + "content": "[45] Weixin Liang, Yuhui Zhang, Yongchan Kwon, Serena Yeung, and James Zou. Mind the gap: Understanding the modality gap in multi-modal contrastive representation learning. In NeurIPS, 2022. 7, 8, 22, 23" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 658, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 658, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 308, + 658, + 545, + 712 + ], + "type": "text", + "content": "[46] Valerii Likhosherstov, Anurag Arnab, Krzysztof Choromanski, Mario Lucic, Yi Tay, Adrian Weller, and Mostafa Dehghani. Polyvit: Co-training vision transformers on images, videos and audio. Trans. Machine Learning Research, 2023. 1, 2, 3" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "11015" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 105 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 105 + ], + "type": "text", + "content": "[47] Lajanugen Logeswaran and Honglak Lee. An efficient framework for learning sentence representations. In ICLR, 2018. 2, 4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 107, + 287, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 107, + 287, + 150 + ], + "spans": [ + { + "bbox": [ + 48, + 107, + 287, + 150 + ], + "type": "text", + "content": "[48] Jiasen Lu, Dhruv Batra, Devi Parikh, and Stefan Lee. ViLBERT: Pretraining task-agnostic visiolinguistic representations for vision-and-language tasks. In NeurIPS, pages 13–23, 2019. 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 152, + 287, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 152, + 287, + 195 + ], + "spans": [ + { + "bbox": [ + 48, + 152, + 287, + 195 + ], + "type": "text", + "content": "[49] Elman Mansimov, Mitchell Stern, Mia Xu Chen, Orhan Firat, Jakob Uszkoreit, and Puneet Jain. Towards end-to-end in-image neural machine translation. CoRR, abs/2010.10648, 2020. 1, 2, 8" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 198, + 287, + 230 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 198, + 287, + 230 + ], + "spans": [ + { + "bbox": [ + 48, + 198, + 287, + 230 + ], + "type": "text", + "content": "[50] Joanna Materzyńska, Antonio Torralba, and David Bau. Disentangling visual and written concepts in CLIP. In CVPR, pages 16410-16419, 2022. 7, 21" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 232, + 287, + 263 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 232, + 287, + 263 + ], + "spans": [ + { + "bbox": [ + 48, + 232, + 287, + 263 + ], + "type": "text", + "content": "[51] Joanna Materzynska, Antonio Torralba, and David Bau. Disentangling visual and written concepts in clip. In CVPR, 2022. 21, 22" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 266, + 287, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 266, + 287, + 330 + ], + "spans": [ + { + "bbox": [ + 48, + 266, + 287, + 330 + ], + "type": "text", + "content": "[52] Sabrina J. Mielke, Zaid Alyafeai, Elizabeth Salesky, Colin Raffel, Manan Dey, Matthias Galle, Arun Raja, Chenglei Si, Wilson Y. Lee, Benoit Sagot, and Samson Tan. Between words and characters: A brief history of open-vocabulary modeling and tokenization in NLP. CoRR, abs/2112.10508, 2021. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 332, + 287, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 332, + 287, + 376 + ], + "spans": [ + { + "bbox": [ + 48, + 332, + 287, + 376 + ], + "type": "text", + "content": "[53] Basil Mustafa, Carlos Riquelme, Joan Puigcerver, Rodolphe Jenatton, and Neil Houlsby. Multimodal contrastive learning with LIMoE: the language-image mixture of experts. In NeurIPS, 2022. 2, 3, 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 378, + 287, + 421 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 378, + 287, + 421 + ], + "spans": [ + { + "bbox": [ + 48, + 378, + 287, + 421 + ], + "type": "text", + "content": "[54] Daisuke Niizumi, Daiki Takeuchi, Yasunori Ohishi, Noboru Harada, and Kunio Kashino. Masked spectrogram modeling using masked autoencoders for learning general-purpose audio representation. CoRR, abs/2204.12260, 2022. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 422, + 287, + 476 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 422, + 287, + 476 + ], + "spans": [ + { + "bbox": [ + 48, + 422, + 287, + 476 + ], + "type": "text", + "content": "[55] Hieu Pham, Zihang Dai, Golnaz Ghiasi, Kenji Kawaguchi, Hanxiao Liu, Adams Wei Yu, Jiahui Yu, Yi-Ting Chen, Minh-Thang Luong, Yonghui Wu, Mingxing Tan, and Quoc V. Le. Combined scaling for open-vocabulary image classification. CoRR, abs/2111.10050, 2021. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 479, + 287, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 479, + 287, + 544 + ], + "spans": [ + { + "bbox": [ + 48, + 479, + 287, + 544 + ], + "type": "text", + "content": "[56] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In ICML, 2021. 1, 2, 3, 4, 6, 14, 17" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 545, + 287, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 545, + 287, + 599 + ], + "spans": [ + { + "bbox": [ + 48, + 545, + 287, + 599 + ], + "type": "text", + "content": "[57] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J. Liu. Exploring the limits of transfer learning with a unified text-to-text transformer. J. Mach. Learn. Res., 21:140:1-140:67, 2020. 3, 4, 6, 20" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 601, + 287, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 601, + 287, + 644 + ], + "spans": [ + { + "bbox": [ + 48, + 601, + 287, + 644 + ], + "type": "text", + "content": "[58] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In ICML, pages 8821-8831, 2021. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 647, + 287, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 647, + 287, + 679 + ], + "spans": [ + { + "bbox": [ + 48, + 647, + 287, + 679 + ], + "type": "text", + "content": "[59] Nils Rethmeier and Isabelle Augenstein. A primer on contrastive pretraining in language processing: Methods, lessons learned and perspectives. ACM Computing Surveys, 2021. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 681, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 681, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 681, + 287, + 713 + ], + "type": "text", + "content": "[60] Phillip Rust, Jonas F. Lotz, Emanuele Bugliarello, Elizabeth Salesky, Miryam de Lhoneux, and Desmond Elliott. Language modelling with pixels. In ICLR, 2023. 1, 2, 3, 7, 18" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "text", + "content": "[61] Phillip Rust, Jonas Pfeiffer, Ivan Vulic, Sebastian Ruder, and Iryna Gurevych. How good is your tokenizer? on the monolingual performance of multilingual language models. In ACL/IJCNLP, pages 3118-3135, 2021. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 118, + 545, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 545, + 194 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 545, + 194 + ], + "type": "text", + "content": "[62] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily Denton, Seyed Kamyar Seyed Ghasemipour, Burcu Karagol Ayan, S. Sara Mahdavi, Rapha Gontijo Lopes, Tim Salimans, Jonathan Ho, David J. Fleet, and Mohammad Norouzi. Photorealistic text-to-image diffusion models with deep language understanding. In NeurIPS, 2022. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 197, + 545, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 197, + 545, + 228 + ], + "spans": [ + { + "bbox": [ + 307, + 197, + 545, + 228 + ], + "type": "text", + "content": "[63] Elizabeth Salesky, David Etter, and Matt Post. Robust open-vocabulary translation from visual text representations. In EMNLP, pages 7235-7252, 2021. 2, 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 231, + 545, + 284 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 231, + 545, + 284 + ], + "spans": [ + { + "bbox": [ + 308, + 231, + 545, + 284 + ], + "type": "text", + "content": "[64] Christoph Schuhmann, Richard Vencu, Romain Beaumont, Robert Kaczmarczyk, Clayton Mullis, Aarush Katta, Theo Coombes, Jenia Jitsev, and Aran Komatsuzaki. LAION-400M: Open dataset of CLIP-filtered 400 million image-text pairs. CoRR, abs/2111.02114, 2021. 4, 15" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 286, + 545, + 318 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 286, + 545, + 318 + ], + "spans": [ + { + "bbox": [ + 308, + 286, + 545, + 318 + ], + "type": "text", + "content": "[65] Rico Sennrich, Barry Haddow, and Alexandra Birch. Neural machine translation of rare words with subword units. In ACL, 2016. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 320, + 545, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 320, + 545, + 352 + ], + "spans": [ + { + "bbox": [ + 308, + 320, + 545, + 352 + ], + "type": "text", + "content": "[66] Noam Shazeer and Mitchell Stern. Adafactor: Adaptive learning rates with sublinear memory cost. In ICML, 2018. 3, 14" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 354, + 545, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 354, + 545, + 397 + ], + "spans": [ + { + "bbox": [ + 308, + 354, + 545, + 397 + ], + "type": "text", + "content": "[67] Sheng Shen, Liunian Harold Li, Hao Tan, Mohit Bansal, Anna Rohrbach, Kai-Wei Chang, Zhewei Yao, and Kurt Keutzer. How much can CLIP benefit vision-and-language tasks? In ICLR, 2022. 2, 3, 5, 17" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 399, + 545, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 399, + 545, + 453 + ], + "spans": [ + { + "bbox": [ + 308, + 399, + 545, + 453 + ], + "type": "text", + "content": "[68] Baohua Sun, Lin Yang, Catherine Chi, Wenhan Zhang, and Michael Lin. Squared english word: A method of generating glyph to use super characters for sentiment analysis. In Workshop on Affective Content Analysis, volume 2328, pages 140-151, 2019. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 456, + 545, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 456, + 545, + 487 + ], + "spans": [ + { + "bbox": [ + 308, + 456, + 545, + 487 + ], + "type": "text", + "content": "[69] Zineng Tang, Jaemin Cho, Yixin Nie, and Mohit Bansal. TVLT: Textless vision-language transformer. In NeurIPS, 2022. 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 490, + 545, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 490, + 545, + 521 + ], + "spans": [ + { + "bbox": [ + 308, + 490, + 545, + 521 + ], + "type": "text", + "content": "[70] Ashish Thapliyal, Jordi Pont-Tuset, Xi Chen, and Radu Soricut. Crossmodal-3600: A Massively Multilingual Multimodal Evaluation Dataset. In EMNLP, 2022. 4, 6, 19" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 524, + 545, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 524, + 545, + 567 + ], + "spans": [ + { + "bbox": [ + 308, + 524, + 545, + 567 + ], + "type": "text", + "content": "[71] Bart Thomee, David A. Shamma, Gerald Friedland, Benjamin Elizalde, Karl Ni, Douglas Poland, Damian Borth, and Li-Jia Li. YFCC100M: the new data in multimedia research. Commun. ACM, 59(2):64-73, 2016. 4, 14" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 569, + 545, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 569, + 545, + 622 + ], + "spans": [ + { + "bbox": [ + 308, + 569, + 545, + 622 + ], + "type": "text", + "content": "[72] Ruben Villegas, Mohammad Babaeizadeh, Pieter-Jan Kindermans, Hernan Moraldo, Han Zhang, Mohammad Taghi Saffar, Santiago Castro, Julius Kunze, and Dumitru Erhan. Phenaki: Variable length video generation from open domain textual description. In ICLR, 2023. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 624, + 545, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 624, + 545, + 667 + ], + "spans": [ + { + "bbox": [ + 308, + 624, + 545, + 667 + ], + "type": "text", + "content": "[73] Alex Wang, Amanpreet Singh, Julian Michael, Felix Hill, Omer Levy, and Samuel R. Bowman. GLUE: A multi-task benchmark and analysis platform for natural language understanding. In ICLR, 2019. 2, 3, 4, 7, 18" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 670, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 670, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 670, + 545, + 713 + ], + "type": "text", + "content": "[74] Jianfeng Wang, Zhengyuan Yang, Xiaowei Hu, Linjie Li, Kevin Lin, Zhe Gan, Zicheng Liu, Ce Liu, and Lijuan Wang. GIT: A generative image-to-text transformer for vision and language. Trans. Machine Learning Research, 2022. 1, 2" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "text", + "content": "11016" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 286, + 665 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 48, + 72, + 286, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 286, + 127 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 286, + 127 + ], + "type": "text", + "content": "[75] Wenhui Wang, Hangbo Bao, Li Dong, Johan Bjorck, Zhiliang Peng, Qiang Liu, Kriti Aggarwal, Owais Khan Mohammed, Saksham Singhal, Subhojit Som, and Furu Wei. Image as a foreign language: BEiT pretraining for all vision and vision-language tasks. CoRR, abs/2208.10442, 2022. 1" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 129, + 286, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 129, + 286, + 171 + ], + "spans": [ + { + "bbox": [ + 49, + 129, + 286, + 171 + ], + "type": "text", + "content": "[76] Zirui Wang, Jiahui Yu, Adams Wei Yu, Zihang Dai, Yulia Tsvetkov, and Yuan Cao. SimVLM: Simple visual language model pretraining with weak supervision. In ICLR, 2022. 1, 8" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 173, + 286, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 173, + 286, + 206 + ], + "spans": [ + { + "bbox": [ + 49, + 173, + 286, + 206 + ], + "type": "text", + "content": "[77] Zhuofeng Wu, Sinong Wang, Jiatao Gu, Madian Khabsa, Fei Sun, and Hao Ma. CLEAR: contrastive learning for sentence representation. CoRR, abs/2012.15466, 2020. 2, 4" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 208, + 286, + 261 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 208, + 286, + 261 + ], + "spans": [ + { + "bbox": [ + 49, + 208, + 286, + 261 + ], + "type": "text", + "content": "[78] Linting Xue, Aditya Barua, Noah Constant, Rami Al-Rfou, Sharan Narang, Mihir Kale, Adam Roberts, and Colin Raffel. Byt5: Towards a token-free future with pre-trained byte-to-byte models. Trans. Assoc. Comput. Linguistics, 10:291-306, 2022. 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 49, + 263, + 286, + 306 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 263, + 286, + 306 + ], + "spans": [ + { + "bbox": [ + 49, + 263, + 286, + 306 + ], + "type": "text", + "content": "[79] Linting Xue, Noah Constant, Adam Roberts, Mihir Kale, Rami Al-Rfou, Aditya Siddhant, Aditya Barua, and Colin Raffel. mT5: A massively multilingual pre-trained text-to-text transformer. In NAACL-HLT, 2021. 6, 20" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 49, + 308, + 286, + 361 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 308, + 286, + 361 + ], + "spans": [ + { + "bbox": [ + 49, + 308, + 286, + 361 + ], + "type": "text", + "content": "[80] Haoxuan You, Luowei Zhou, Bin Xiao, Noel Codella, Yu Cheng, Ruochen Xu, Shih-Fu Chang, and Lu Yuan. Learning visual representation from modality-shared contrastive language-image pre-training. In ECCV, pages 69-87, 2022. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 49, + 364, + 286, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 364, + 286, + 407 + ], + "spans": [ + { + "bbox": [ + 49, + 364, + 286, + 407 + ], + "type": "text", + "content": "[81] Peter Young, Alice Lai, Micah Hodosh, and Julia Hockenmaier. From image descriptions to visual denotations: New similarity metrics for semantic inference over event descriptions. Trans. Assoc. Comput. Linguistics, 2:67-78, 2014. 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 49, + 409, + 286, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 409, + 286, + 453 + ], + "spans": [ + { + "bbox": [ + 49, + 409, + 286, + 453 + ], + "type": "text", + "content": "[82] Jiahui Yu, Zirui Wang, Vijay Vasudevan, Legg Yeung, Mojtaba Seyedhosseini, and Yonghui Wu. Coca: Contrastive captioners are image-text foundation models. Trans. Machine Learning Research, 2022. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 49, + 454, + 286, + 529 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 454, + 286, + 529 + ], + "spans": [ + { + "bbox": [ + 49, + 454, + 286, + 529 + ], + "type": "text", + "content": "[83] Jiahui Yu, Yuanzhong Xu, Jing Yu Koh, Thang Luong, Gunjan Baid, Zirui Wang, Vijay Vasudevan, Alexander Ku, Yinfei Yang, Burcu Karagol Ayan, Ben Hutchinson, Wei Han, Zarana Parekh, Xin Li, Han Zhang, Jason Baldridge, and Yonghui Wu. Scaling autoregressive models for content-rich text-to-image generation. Trans. Machine Learning Research, 2022. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 49, + 532, + 286, + 564 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 532, + 286, + 564 + ], + "spans": [ + { + "bbox": [ + 49, + 532, + 286, + 564 + ], + "type": "text", + "content": "[84] Zhou Yu, Jun Yu, Yuhao Cui, Dacheng Tao, and Qi Tian. Deep modular co-attention networks for visual question answering. In CVPR, pages 6281–6290, 2019. 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 49, + 566, + 286, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 566, + 286, + 609 + ], + "spans": [ + { + "bbox": [ + 49, + 566, + 286, + 609 + ], + "type": "text", + "content": "[85] Xiaohua Zhai, Xiao Wang, Basil Mustafa, Andreas Steiner, Daniel Keysers, Alexander Kolesnikov, and Lucas Beyer. LiT: Zero-shot transfer with locked-image text tuning. In CVPR, pages 18102-18112, 2022. 2, 4, 5, 6, 14" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 49, + 610, + 286, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 610, + 286, + 665 + ], + "spans": [ + { + "bbox": [ + 49, + 610, + 286, + 665 + ], + "type": "text", + "content": "[86] Yukun Zhu, Ryan Kiros, Richard S. Zemel, Ruslan Salakhutdinov, Raquel Urtasun, Antonio Torralba, and Sanja Fidler. Aligning books and movies: Towards story-like visual explanations by watching movies and reading books. In ICCV, pages 19-27, 2015. 7, 18" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "11017" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/CLOTH4D_ A Dataset for Clothed Human Reconstruction/b8c1cf88-b3e6-48d2-932c-20d8bacc61eb_content_list.json b/2023/CLOTH4D_ A Dataset for Clothed Human Reconstruction/b8c1cf88-b3e6-48d2-932c-20d8bacc61eb_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..aa7281a2eaa1881a9bdd8af33b1d2a0104952cba --- /dev/null +++ b/2023/CLOTH4D_ A Dataset for Clothed Human Reconstruction/b8c1cf88-b3e6-48d2-932c-20d8bacc61eb_content_list.json @@ -0,0 +1,1419 @@ +[ + { + "type": "text", + "text": "CLOTH4D: A Dataset for Clothed Human Reconstruction", + "text_level": 1, + "bbox": [ + 189, + 130, + 779, + 150 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Xingxing Zou $^{1,3*}$ , Xintong Han $^{2*}$ , Waikeung Wong $^{3,1\\dagger}$", + "bbox": [ + 267, + 179, + 699, + 199 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Laboratory for Artificial Intelligence in Design, $^{2}$ Huya Inc.", + "bbox": [ + 248, + 199, + 720, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{3}$ School of Fashion and Textiles, The Hong Kong Polytechnic University", + "bbox": [ + 197, + 215, + 769, + 234 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "aemika.zou@connect.polyu.hk,hanxintong@huya.com,calvin.wong@polyu.edu.hk", + "bbox": [ + 163, + 237, + 803, + 250 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/7957263b34aa05b1ffe126ee43aefb27d252d1994001e46367c23d43ae8d3cd7.jpg", + "image_caption": [ + "(a) 2D patterns" + ], + "image_footnote": [], + "bbox": [ + 96, + 292, + 277, + 363 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/60c3a88e7c0cd270ce27af08f8d5e0b983a4420960d3720027e2a41795918ae7.jpg", + "image_caption": [ + "(b) varied poses" + ], + "image_footnote": [], + "bbox": [ + 91, + 395, + 129, + 458 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/e196cf514ca353ba391e3c384d7b33444d7ace70293da689ddc40022b5de99ea.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 145, + 393, + 176, + 458 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/9c2de006c3de326f318b48c04f10f190d6d781d9037556a51f393895c3ff717a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 197, + 395, + 225, + 458 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/8bae1e14ed893ffdd96d688f0481315579be8255ca31cb9c9b0730d0c7bab51d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 241, + 395, + 279, + 458 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/87003073c460fc8613142f8a555c2d3e56a91a172f2d495f967dcbec51eefd5b.jpg", + "image_caption": [ + "(c) multi views" + ], + "image_footnote": [], + "bbox": [ + 83, + 488, + 132, + 550 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/909b4dadf1846c304d4f45c96200b0f3f4cd046aca1dc61b6a7536a08c784cf7.jpg", + "image_caption": [ + "Figure 1. An instance in CLOTH4D. Given (a) hand-crafted 2D sewing patterns, clothes are simulated with virtual fashion software to have (b) varied poses, (c) multiple views, and (d) high-quality and physically plausible series of paired dynamic 3D meshes with textures." + ], + "image_footnote": [], + "bbox": [ + 140, + 489, + 178, + 550 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/b359570e2358d4a1935537b35a2ff34baa8318f3f03d4f6d962c1fd98a1af424.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 197, + 489, + 225, + 550 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/251d0b8892bc10cd17cb25bc423f69373897f71b62804ea10ec7e256dde8ae1b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 233, + 489, + 282, + 550 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/01b3afc8b98f15453b3e90fc54650f95ee8e2455a796af0535a6d5c2367fac95.jpg", + "image_caption": [ + "(d) series of paired dynamic textured meshes" + ], + "image_footnote": [], + "bbox": [ + 303, + 287, + 883, + 553 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 609, + 313, + 625 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Clothed human reconstruction is the cornerstone for creating the virtual world. To a great extent, the quality of recovered avatars decides whether the Metaverse is a passing fad. In this work, we introduce CLOTH4D, a clothed human dataset containing 1,000 subjects with varied appearances, 1,000 3D outfits, and over 100,000 clothed meshes with paired unclothed humans, to fill the gap in large-scale and high-quality 4D clothing data. It enjoys appealing characteristics: 1) Accurate and detailed clothing textured meshes—all clothing items are manually created and then simulated in professional software, strictly following the general standard in fashion design. 2) Separated textured clothing and under-clothing body meshes, closer to the physical world than single-layer raw scans. 3) Clothed human motion sequences simulated given a set of 289 actions, covering fundamental and complicated dynamics. Upon CLOTH4D, we novelly designed a series of", + "bbox": [ + 73, + 643, + 473, + 902 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "temporally-aware metrics to evaluate the temporal stability of the generated 3D human meshes, which has been overlooked previously. Moreover, by assessing and retraining current state-of-the-art clothed human reconstruction methods, we reveal insights, present improved performance, and propose potential future research directions, confirming our dataset's advancement. The dataset is available at1.", + "bbox": [ + 496, + 611, + 893, + 717 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 500, + 734, + 632, + 750 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "As we enter the volumetric and XR content era, researchers have been trailblazing their way into the Metaverse. With the converging of technologies and practical applications, e.g., fashion NFTs (non-fungible tokens), immersive AR and VR, and games, clothed human reconstruction demands are rapidly growing. While current research has made astonishing results in creating digital hu", + "bbox": [ + 496, + 758, + 893, + 866 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "1www.github.com/AemikaChow/AiDLab-fAshIon-Data", + "bbox": [ + 514, + 875, + 880, + 887 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*X. Zou and X. Han contribute equally. † Corresponding author.", + "bbox": [ + 519, + 887, + 857, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "12847", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "table", + "img_path": "images/ab23e9cfda56ffa8f1f12c44bc7a869b618ee39838f597bc30021929e168233e.jpg", + "table_caption": [ + "Table 1. Comparisons of CLOTH4D with existing representative datasets. Gray color indicates synthetic datasets generated with graphics engines. #Subjects: number of peoples in different appearances; #Action: number of actions adopted; #Scans: numbers of 3D meshes; 2D Pattern: 2D clothing pattern; TexCloth: with textured clothed model; TexHuman: with textured naked human model. w/ SMPL: with registered SMPL [33] parameters. Public: publicly available and free of charge. Photorealistic: whether the images in the dataset are realistic. -: not applicable or reported. CLOTH4D presents more desirable characteristics compared with others." + ], + "table_footnote": [], + "table_body": "
Dataset#Subjects#Action#Scan2D PatternTexClothTexHumanw/ SMPLPublicPhotorealistic
BUFF [52]6-13.6k-✓-✓✓✓
RenderPeople [1]--825-✓-✓-✓
DeepWrinkles [30]229.2k-✓---✓
CAPE [35]15600140k---✓✓✓
THuman2.0 [51]200-525-✓-✓✓✓
DRAPE [16]72324.5k------
Wang et al. [47]--24k✓✓-✓✓-
3DPeople [40]8072--✓---✓
DCA [43]-567.1k---✓--
GarNet [17]600-18.8k---✓✓-
TailorNet [38]9-5.5k-✓-✓✓-
Cloth3D [8]8.5k7.9k2.1M-✓-✓✓-
Cloth3D++ [36]9.7k8k2.2M✓✓✓✓✓-
CLOTH4D1k289100k✓✓✓✓✓✓
", + "bbox": [ + 80, + 166, + 888, + 383 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "mans, these reconstructed meshes have issues, e.g., flexible body motions and diverse appearances, owing to the lack of datasets with richness in clothing and realistic dynamics of garments. To this end, we introduce CLOTH4D, an open-sourced dataset facilitating physically plausible dynamic clothed human reconstruction.", + "bbox": [ + 75, + 390, + 470, + 481 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Prior to us, many datasets have been collected, and we sort out them in Table 1. Currently, scanned datasets are widely adopted as they are photorealistic and can be easily processed to watertight meshes, which does an excellent favor for current deep models to learn an implicit function (e.g., signed distance function) followed by marching cubes [34] for surface reconstruction. However, it is born with some weaknesses: 1) Scanned meshes are single-layer and inherently fail to capture the space between clothing and skin surface. Thus, body shape under clothing cannot be accurately inferred, let alone the multi-layer and thin clothing structures as in the real physical world. 2) It is time-consuming and expensive to obtain high-quality and large-scale temporal scanned sequences (i.e., 4D scanned sequences) due to the limited efficiency and precision of 4D scanners, especially for complicated clothing and large motions. Although synthetic datasets can to some extent overcome these limitations, existing synthetic datasets are either of small scale in terms of appearances and motions or are highly unrealistic. Moreover, many datasets are not made publicly available and free.", + "bbox": [ + 75, + 486, + 470, + 805 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In contrast, CLOTH4D possesses several attractive attributes: 1) We made great efforts to the diversity and quality of clothing. All clothes are manually designed in CLO [3] and cater to the requirement of the fashion industry. 2) Meshes in CLOTH4D are clothing/humans separated. Such flexibility makes studying and modeling the", + "bbox": [ + 75, + 810, + 470, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "relations and interactions between clothing simulation and body movement possible. 3) CLOTH4D provides plenty of temporal motion sequences with realistic clothing dynamics. As the human body moves, the dressed clothing, e.g., the skirt in Figure 1, naturally deforms. 4) The dataset is large-scale and openly accessible.", + "bbox": [ + 496, + 390, + 893, + 481 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To demonstrate the advantages of CLOTH4D, we use it to evaluate the state-of-the-art (SOTA) clothed human reconstruction methods. In addition to the generally adopted static evaluation metrics, we propose a set of temporally-aware metrics to assess the temporal coherence in a video inference scenario thanks to the rich and true-to-life 4D synthetic sequences in the dataset. Quantitative and qualitative results of SOTA methods on CLOTH4D suggest that our dataset is challenging and the temporal stability of the reconstructed mesh is vital for evaluating the perceptual quality. Meanwhile, we retrain SOTA methods on CLOTH4D, revealing interesting observations of how they perform on multi-layer meshes with thin clothing structures. With in-depth analysis and a summary of challenges for the existing approaches, CLOTH4D makes an essential step toward more realistic reconstructions of clothed humans and stimulates several exciting future work directions. All in all:", + "bbox": [ + 496, + 483, + 893, + 741 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We contribute CLOTH4D, a large-scale, high-quality, and open-accessible 4D synthetic dataset for clothed human reconstruction.", + "- We introduce a series of temporally-aware metrics to evaluate the reconstructed performance in the aspect of temporal consistency.", + "- With the proposed dataset and metrics, we thoroughly analyze the pros and cons of SOTAs, summarize the existing challenges toward more realistic 3D modeling, and propose potential new directions." + ], + "bbox": [ + 496, + 743, + 893, + 901 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "12848", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/c77154e2a7d5a2da5c3da834fcbbf1d35ab89da51d61a44605cdb11f7896b8bd.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 78, + 79, + 356, + 202 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/8f54d30d586003c4a09b136119930a84e4d9af79c112ce884e7c546ad7abffbc.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 372, + 80, + 617, + 199 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/adb9343b9720e9b1131e722f6e956721e10478ebf2d18970f7b93c5f7ffad7a9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 622, + 80, + 892, + 205 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/6ce2fdad0e1c361c30f99d83c9cbe305b7f6a863ed87ae912b47c167ad7ae29f.jpg", + "image_caption": [ + "(i) pattern making" + ], + "image_footnote": [], + "bbox": [ + 78, + 200, + 356, + 268 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/a5f69fbd618a86af8c964ba33f91c83eb123be469fa771684d47c23b99d7e244.jpg", + "image_caption": [ + "(ii) clothed 3D avatar" + ], + "image_footnote": [], + "bbox": [ + 277, + 220, + 354, + 276 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/4b01f44e2357bd9ac5ad117d021bf54e1d4a0cd6c1f0f2cfeec5cc14fafca584.jpg", + "image_caption": [ + "(iii) clothing simulation" + ], + "image_footnote": [], + "bbox": [ + 401, + 223, + 522, + 279 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/5d88420ffa946a9c1f1cbf55aabf38823672789539611c05e6ed8b30d5814714.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 555, + 252, + 578, + 262 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/37d67aa9ddb6bb7824306184d553b8b21117402a5214bf91ed6d74ec43f4bb8d.jpg", + "image_caption": [ + "CLOTH4D with varied 3D sequences", + "Figure 2. Pipeline for creating instances in CLOTH4D, which primarily adopts CLO for clothing design and simulation, Mixamo for animation, and Blender for processing and exporting meshes." + ], + "image_footnote": [], + "bbox": [ + 591, + 214, + 892, + 292 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 76, + 342, + 218, + 357 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Clothed Human Reconstruction. Clothed human reconstruction aims to recover a 3D mesh from a monocular person image. Estimating dressed people by modeling clothing geometry as 3D displacement on top of a parametric body model (e.g., SMPL [33]) is the leading solution to deal with this task [5,6,28,31,39,48]. By transferring the skinning weights from the body model to the offset clothing mesh, the reconstructed clothed mesh can be readily deformed and animated in the same way as the under-clothing 3D parametric body model. However, it assumes the clothed human to have the same topology as the naked body, leading to unsatisfactory reconstructions of long hair, skirts, dresses, etc. Although methods such as [9,22,26,38] try to isolate the reconstruction of clothing by constructing category-specific clothing templates or statistical models, they fail to generalize to unseen or complex types of clothing.", + "bbox": [ + 75, + 369, + 468, + 609 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "As another line of research, deep implicit function networks have drawn broader attention recently [7, 15, 20, 21, 24, 41, 42, 54]. PIFu [41] conditions on pixel-aligned features to build deep implicit functions for reconstructing human meshes, and PIFuHD [42] goes a further step towards enhancing 3D geometric details by predicting front and back normals. ARCH [24] and $\\mathrm{ARCH}++$ [21] enable animating the reconstructed meshes by deforming the semantic information into the SMPL canonical space. More recently, leveraging SMPL body models as prior, PaMIR [54] and ICON [49] further improve the reconstruction quality, especially on challenging poses. PHORHUM [7] achieves more accurate results via jointly estimating geometry, albedo, and shading information. It is worth noting that most works rely on training data that are not available free of charge, making an open-sourced dataset of clothed avatars of great significance to advance research in 3D body/clothing modeling.", + "bbox": [ + 75, + 612, + 468, + 868 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Clothed Human Datasets. As summarized in Table 1, existing clothed human datasets can be divided into two", + "bbox": [ + 75, + 869, + 468, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "types, i.e., scanned datasets and synthetic datasets. The former [10, 25, 44, 46] utilizes multiple synchronized cameras to capture motions that have difficulties enriching the scalability and diversity and obtaining highly accurate 4D ground truth. Moreover, it inherently cannot mimic the layering of clothing. The synthetic datasets mitigate these limitations. However, existing synthetic datasets, no matter whether static [17, 38, 47] or dynamic [8, 16, 39, 43], either only contain a few clothing types or are highly unrealistic. Cloth3D++ [36] is developed from Cloth3D [8], which contains a total of 2.2 million scans, covering 9.7k subjects dressed in 12.9k clothing, which is the most advanced in scale. However, the clothing created based on garment templates are only base patterns with an immense gap compared to clothing in real life. To intuitively demonstrate the advantages of CLOTH4D, we put visual comparisons of these datasets in Figure A in the supplementary material.", + "bbox": [ + 496, + 343, + 892, + 599 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. CLOTH4D Dataset", + "text_level": 1, + "bbox": [ + 500, + 616, + 686, + 632 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We depict the pipeline of creating CLOTH4D in Figure 2, including (1) preparing the referenced clothing images and the unclothed 3D human avatars; (2) uploading 3D human avatars to Mixamo to obtain the FBX files with various animations; (3) designing 3D clothing in CLO, integrating the FBX files obtained in (2), and conducting clothing simulation; and (4) exporting the sequenced mesh files.", + "bbox": [ + 496, + 642, + 890, + 748 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Clothes and Models. Clothes are manually created by professional fashion designers using CLO by producing the 2D garment patterns and then auto-simulating them to 3D. We put a video illustrating the production process of a 3D garment in the supplementary material (Video A) instead of demonstrating details in the paper. CLOTH4D covers 1,000 different 3D outfits spansning over 500 prints and 50 fabrics. Over $40\\%$ of the clothes are designed referring to the newest collections of varying design houses (e.g., Prada, Moschino, and Alexander McQueen) to ensure visual re", + "bbox": [ + 496, + 750, + 892, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "12849", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "alism, variance, and fashionability. In this paper, we focus on women's wear owing to its diversity covering most characteristics of garments. For avatars, CLO provides human avatars with varied physical appearances in terms of hairstyles, faces, skin colors, body figures, etc., and we directly adopt these available avatars.", + "bbox": [ + 75, + 90, + 468, + 181 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Animations and Simulations. For each human avatar, we use Mixamo [2] for rigging and generating motion sequences. Given a motion sequence and a 3D garment, CLO runs a clothing simulation of the 3D garment according to the motion sequence at 30fps. A total of 289 animations are utilized, such as Billy Dance, Offensive Idle, Jumping Rope, etc. Unlike previous work [8] directly uses Blender for cloth simulation, CLO simulates clothing with richer details and dynamics. The clothes exhibit wrinkles and fold with the body's movement in a natural and physically plausible way, especially for skirts or dresses that differ a lot from the body topology as shown in Figure 1.", + "bbox": [ + 75, + 181, + 470, + 362 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Paired Multi-Modal Data. Then, we can obtain paired data in the following representations: a 2D clothing pattern, 3D mesh sequences (clothed and naked human meshes with corresponding fitted SMPL parameters/meshes, and separate clothing meshes), and a UV texture map. The textured mesh can be rendered into multi-view normal images, depth images, and RGB images given varying light conditions. We also translate all these dynamic meshes to watertight with simplification using [23], thus they can be readily used to train an implicit function for mesh reconstruction. The number of triangles ranges from $170K$ to $4M$ for simulation and becomes $200K$ after simplification. Besides human reconstruction, many other tasks could also benefit from this paired data (e.g., clothing capture [50], human pose transfer [18], and fashion-related tasks [19, 56, 57]).", + "bbox": [ + 75, + 363, + 470, + 590 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Evaluations", + "text_level": 1, + "bbox": [ + 76, + 602, + 202, + 618 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this section, we evaluate the state-of-the-art clothed human reconstruction methods on CLOTH4D to demonstrate new insights that the dataset can provide. Further, we retrain SOTA methods on CLOTH4D and make several interesting observations of how they perform on multi-layer meshes with thin clothing structures. We also present the challenges for existing approaches and propose potential research directions with a comprehensive analysis.", + "bbox": [ + 75, + 627, + 468, + 750 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1. Baselines", + "text_level": 1, + "bbox": [ + 76, + 756, + 184, + 771 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We mainly report results of four SOTA approaches, PIFu [41], PIFuHD [42], PaMIR [54], and ICON [49] owing to other works do not release their codes or models, such as PHORHUM [7], ARCH++ [21], or which have already been extensively compared with the listed methods above. We use PIFu, PIFuHD, PaMIR, and ICON to denote their released pretrained testing models. $\\mathrm{PaMIR}^{gt}$ and $\\mathrm{ICON}^{gt}$ indicate the testing results using the fitted ground truth", + "bbox": [ + 75, + 779, + 470, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "SMPL mesh as conditions rather than the one predicted using off-the-shelf human mesh recovery (HMR) methods (GraphCMR [29] for PaMIR, and PyMAf [53] for ICON as in their released code). Based on the characteristics of these methods, they can be divided into three types: 1) pixel-aligned methods (PIFu, PIFuHD); 2) GT-SMPL-guided + pixel-aligned methods $(\\mathrm{PaMIR}^{gt}$ and $\\mathrm{ICON}^{gt})$ ; 3) HMR-SMPL-guided + pixel-aligned methods (PaMIR and ICON).", + "bbox": [ + 496, + 90, + 890, + 210 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Furthermore, we retrain PIFu, PaMIR, and ICON on CLOTH4D, which are denoted as $\\mathrm{PIF_u_{clo}}$ , $\\mathrm{PaMIR}_{clo}$ , and $\\mathrm{ICON}_{clo}$ , respectively. For these retrained models, we follow the re-implementation setting introduced in ICON [4], which allows us to train all these baselines with the same training protocol and hyper-parameters for a fair comparison. Similarly, $\\mathrm{PaMIR}_{clo}^{gt}$ , and $\\mathrm{ICON}_{clo}^{gt}$ are tested with the ground truth SMPL fits. We also use the cloth-refinement module in the ICON's released code for post-processing.", + "bbox": [ + 496, + 212, + 890, + 348 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.2. Datasets and Metrics", + "text_level": 1, + "bbox": [ + 498, + 354, + 699, + 371 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Datasets and implementation details. We organize the sequences in CLOTH4D into a $80\\% / 10\\% / 10\\%$ train/val/test split. We render each mesh into 8 views using a weak perspective camera and pre-computed radiance transfer [45] with dynamic light conditions following [41, 49]. All rendered images are $512 \\times 512$ . The 2D keypoints used in all methods are generated by OpenPose [11]. In addition, we also evaluate all models on CAPE [35] test set adopted in [49] to investigate the generalization ability.", + "bbox": [ + 496, + 378, + 890, + 515 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Static Metrics. We report the quantitative results on normal reprojection error, Chamfer distance, and P2S distance for evaluation as [15, 41, 42, 49]. As all compared methods use a weak perspective or orthographic camera, the estimated meshes may not be well aligned with the ground truth meshes in the $z$ -direction (i.e., view direction). Thus, we shift the estimated meshes to have the same $z$ -axis mean as the ground truths following [15] for a fair comparison.", + "bbox": [ + 496, + 513, + 890, + 635 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Temporal Metrics. The aforementioned static metrics ignore the temporal consistency of the reconstructed meshes across time, which is essential for real-time applications since meshes presenting jitters and flickers highly affect the perceptual quality. Thanks to the rich temporal dynamics provided in CLOTH4D, we are the first to introduce temporally-aware metrics to evaluate the temporal coherence of the generated mesh sequences. Referring to temporal metrics SSDdt and dtSSD used in video matting tasks [14, 32], we compute two metrics measuring the temporal coherence of the predicted mesh normal:", + "bbox": [ + 496, + 635, + 890, + 800 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\text {N o r m a l s} _ {d d t} = \\frac {1}{T} \\sum_ {t} \\left| \\left(\\mathcal {N} _ {t} ^ {p r} - \\mathcal {N} _ {t} ^ {g t}\\right) ^ {2} - \\left(\\mathcal {N} _ {t + 1} ^ {p r} - \\mathcal {N} _ {t + 1} ^ {g t}\\right) ^ {2} \\right|, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 509, + 804, + 890, + 832 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {N o r m a l s} _ {d t d} = \\frac {1}{T} \\sum_ {t} \\left| \\left(\\mathcal {N} _ {t} ^ {p r} - \\mathcal {N} _ {t + 1} ^ {p r}\\right) ^ {2} - \\left(\\mathcal {N} _ {t} ^ {g t} - \\mathcal {N} _ {t + 1} ^ {g t}\\right) ^ {2} \\right|, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 509, + 837, + 890, + 866 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $T$ is the length of the sequence. $\\mathcal{N}_t^{pr}$ and $\\mathcal{N}_t^{gt}$ denote the rendered normal images from the predicted mesh", + "bbox": [ + 500, + 869, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "12850", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/a6f99568675b623722d33408c121630d3fc11daa39813c1fae221bf12bd82afe.jpg", + "table_caption": [ + "Table 2. Quantitative evaluation on CLOTH4D. PaMIRgt and ICONgt denote that the fitted ground truth SMPL is used during the inference. Gray color indicates the results trained on CLOTH4D." + ], + "table_footnote": [], + "table_body": "
MethodPIFuPIFuHDPaMIR\\( PaMIR^{gt} \\)ICON\\( ICON^{gt} \\)\\( PIFu_{clo} \\)\\( PaMIR_{clo} \\)\\( PaMIR^{gt}_{clo} \\)\\( ICON_{clo} \\)\\( ICON^{gt}_{clo} \\)
Normals \\( \\downarrow \\)0.1820.1810.2240.1450.2110.1180.1500.2300.1140.1980.103
P2S \\( \\downarrow \\)3.9113.5183.8272.6414.6602.4732.7935.0372.6193.7112.068
Chamfer \\( \\downarrow \\)3.5782.4874.1572.4874.1961.6312.4124.1861.6183.4991.367
\\( Normals_{ddt} \\) \\( \\downarrow \\)0.0080.0070.0320.0250.0130.0300.0090.0210.0330.0130.035
\\( Normals_{dtd} \\) \\( \\downarrow \\)0.0250.0230.0430.0350.0340.0280.0170.0400.0270.0380.033
\\( P2S_{ddt} \\) \\( \\downarrow \\)0.2220.1670.2530.2470.3830.3690.1850.4170.4020.3210.367
\\( P2S_{dtd} \\) \\( \\downarrow \\)0.7330.5781.0250.7020.9790.5540.5480.8900.5360.8780.526
Chamfer \\( _{ddt} \\) \\( \\downarrow \\)0.2070.1810.3310.3160.3590.3580.1570.3500.3820.3310.369
Chamfer \\( _{dtd} \\) \\( \\downarrow \\)0.7290.5761.0170.7010.9740.5530.5460.8860.5360.8730.525
", + "bbox": [ + 81, + 119, + 888, + 262 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/930820348bdb765f11610271205d471c5a4f5822f27945ab874c5514ca54e2c4.jpg", + "table_caption": [ + "Table 3. Quantitative evaluation on CAPE. Table notations are the same as Table 2." + ], + "table_footnote": [], + "table_body": "
MethodPIFuPIFuHDPaMIRPaMIRgtICONICONgtPIFucloPaMIRcloPaMIRgtcloICONcloICONgtclo
Normals ↓0.1610.1600.1830.0860.1600.0560.1640.1760.0930.1560.077
P2S ↓4.2593.7953.8401.1934.0141.0674.6524.2351.4913.3041.193
Chamfer ↓4.2043.9274.2581.6543.9621.0384.3814.0801.4273.5441.397
", + "bbox": [ + 81, + 284, + 888, + 351 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "and the ground truth mesh at time step $t$ , respectively. The subscript $ddt$ is short for distance delta time, which captures the stability of errors between two consecutive meshes. And $dtd$ (delta time distance) penalizes large temporal change of the prediction with respect to the change of the ground truth. These two metrics indicate unstable mesh variations and ignore temporally coherent errors [14]. The $ddt$ and $dtd$ of Chamfer and P2S distances are similarly defined. More details can be found in the supplementary material.", + "bbox": [ + 75, + 359, + 472, + 494 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.3. Baseline Evaluation", + "text_level": 1, + "bbox": [ + 76, + 508, + 266, + 523 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Quantitative results. Table 2 gives quantitative results on the CLOTH4D test set using the evaluation metrics described in Section 4.2. As indicated in the non-gray part, we made the following observations:", + "bbox": [ + 75, + 534, + 468, + 594 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1) In terms of static metrics, $\\mathrm{ICON}^{gt} > \\mathrm{PaMIR}^{gt} > \\mathrm{PI}-\\mathrm{FuHD} > \\mathrm{PIF_u} > \\mathrm{PaMIR} > \\mathrm{ICON}$ . I.e., GT-SMPL-guided + pixel-aligned methods > pure pixel-aligned methods > HMR-SMPL-guided + pixel-aligned methods.", + "2) With the strong guidance of ground truth SMPL mesh, the performance of $\\mathrm{ICON}^{gt}$ and $\\mathrm{PaMIR}^{gt}$ significantly improves compared to their counterparts with estimated SMPL (i.e., ICON and PaMIR). However, ground truth SMPL meshes are unavailable at test time, which suggests that previous comparisons [15, 49] between GT-SMPL-based methods and others may be unfair. And pure pixel-aligned methods may be even more favorable than SOTA SMPL-based methods for the in-the-wild scenario.", + "3) From the perspective of temporally-aware metrics, pure pixel-aligned methods have higher reconstruction stability (see Figure 4). We attribute this to the fact that ICON and PaMIR strongly rely on the SMPL body prior, thus failing to generate far-from-the-body clothes (e.g., skirts and dresses) that present rich temporal dynamics as shown in Figure 1. Plus, the jittery and unstable pose estimation of" + ], + "bbox": [ + 75, + 595, + 470, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "the off-the-shelf HMR methods further prevents ICON and PaMIR from generating temporally coherent results.", + "bbox": [ + 496, + 359, + 890, + 391 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "4) ICON outperforms PaMIR as ICON better models mesh-based local features while PaMIR depends more on global information. Moreover, PaMIR loses high-frequency details due to the limited resolution of its volumetric representation. Our observations on CAPE are in line with previous works. For simplicity, we do not expand the narrative.", + "Qualitative results. We present the qualitative results on CLOTH4D in Figure 3 and draw the following insights that have not been fully explored since there are no such large-scale and diverse datasets like CLOTH4D.", + "1) Global shape vs. local details. All baselines can reconstruct the overall shape conditioned on the input RGB image. PIFuHD presents the finest details, followed by ICON, PaMIR, and PIFu, as PIFuHD enlarges the spatial resolution of pixel-aligned features and ICON takes advantage of mesh-based local features (signed distance, surface normal, etc.). However, focusing on local features suffers from overfitting and poor generalization to complicated clothing (e.g., incomplete dress in the 3rd and 5th examples in Figure 3) and large motions (e.g., artifacts in the arm regions in the 2nd, 3rd, 6th, and 8th examples in Figure 3). Comparatively, thanks to the global feature encoder, PaMIR, and PIFu can generate more holistic clothing but sacrifice details. Thus, it is an important future research direction to explore better strategies for balancing the local and global reconstruction quality.", + "2) Human body priors. The last three rows in Figure 3 and Figure 4 (also shows side views) present results on relatively challenging poses. $\\mathrm{ICON}^{gt}$ and ICON robustly recover the poses, while PIFu, PIFuHD, and PaMIR are prone to producing broken limbs or anatomically improbable shapes to different extents. For the side view, we can find that $\\mathrm{PaMIR}^{gt}$ and $\\mathrm{ICON}^{gt}$ are more similar to the" + ], + "bbox": [ + 496, + 393, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "12851", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/e19d45237e7f35dc621f920218b042fef81b42e57a3a7fe951d6051da6e588d2.jpg", + "image_caption": [ + "Figure 3. Qualitative results on CLOTH4D." + ], + "image_footnote": [], + "bbox": [ + 81, + 87, + 472, + 492 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "ground truth mesh, which is unsurprising as they are given the ground truth SMPL as prior. Comparatively, PaMIR and ICON, which adopt the estimated SMPL, face the common problems of HMR-SMPL-based reconstruction methods—the predicted SMPL body bends legs or hunches over due to the depth ambiguity—leading to large reconstruction error. Similarly, PIFu and PIFuHD tend to have forward heads with slightly bending legs as they are not aware of any human body priors. One potential research direction is to jointly train or optimize body prior (e.g., SMPL, keypoints, human parsing) with mesh reconstruction.", + "bbox": [ + 75, + 518, + 468, + 685 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3) Ambiguity of geometry and appearance. As shown in the 6th-8th rows of Figure 3, the clothing prints can affect the surface reconstruction due to the ambiguity of geometry and appearance in single-view rendering. Among all these baselines, ICON and PIFuHD show high robustness to the input clothing prints as they predict normal images as intermediate representations, which reduces this ambiguity compared to directly inputting RGB images. Note that ICON only takes the normal images as input to the reconstruction module without using the RGB images, further mitigating the ambiguity and bringing even higher robustness than PIFuHD (the 8th example). Motivated by this observation, future research could shed more light on disentangling the geometry and appearance either implicitly [37] or explicit", + "bbox": [ + 75, + 688, + 470, + 902 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/6e62e8b77b9da05f07ed72c7370e6a471563bba9b6837cbffd73fd5f8cebed2c.jpg", + "image_caption": [ + "Figure 4. Temporal qualitative results. The 1st, 3rd, and 5th rows are three consecutive frames, and the 2nd, 4th, and 6th rows are the side views predicted from the corresponding front-view RGB. Refer to the supplementary material (Video B) for video results." + ], + "image_footnote": [], + "bbox": [ + 506, + 85, + 890, + 391 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "ity [7]. Also, predicting more comprehensive intermediate 2D/3D representations (e.g., depth, illumination, keypoints, segmentation) may also improve the performance.", + "bbox": [ + 498, + 460, + 890, + 507 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4) Temporal consistency. For real-time applications, e.g., streaming from a monocular camera and importing the reconstructed motion sequence into a virtual scene, the temporal coherence of the generated meshes over time is vital for a high-quality user experience. We show the reconstructed meshes for three consecutive frames in Figure 4. As can be found from the side views of the reconstructed meshes, although only subtle motions are present in these frames, the HMR-SMPL-guided methods (PaMIR and ICON) suffer from unstable SMPL predictions and generate temporally inconsistent meshes. On the other hand, pure pixel-aligned methods (PIFu and PIFuHD) fail to predict accurate human pose but produce temporally consistent errors, thus having small $ddt$ values. Given the ground truth SMPL meshes, $\\mathrm{PaMIR}^{gt}$ is still more sensitive to global pose than $\\mathrm{ICON}^{gt}$ as also noted by [49]. These observations are consistent with the quantitative temporal metrics reported in Table 2. It would be interesting to investigate temporal modeling of implicit functions (e.g., incorporating a recurrent neural network [27, 32, 55] to train the implicit function on the 4D dataset, or applying test-time fitting to refine the reconstructed meshes with temporal loss terms). We make such an attempt by adding a temporal term (penalize Chamfer distance between two successive frames) to the refinement process of $\\mathrm{PIFu}_{clo}$ and achieve better temporal consistency (Chamfer_ddt: $0.157\\rightarrow 0.123$ and $\\mathrm{Chamfer}_{dtd}$ :", + "bbox": [ + 496, + 508, + 892, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "12852", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/cc8d9eb1464f0be0cb588a3b7b5f238d88672be921d44dac08256fe4abed7336.jpg", + "image_caption": [ + "Figure 5. Qualitative results of baseline methods tested on CLOTH4D with front view and side view." + ], + "image_footnote": [], + "bbox": [ + 78, + 85, + 890, + 684 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "$0.546\\rightarrow 0.454)$", + "bbox": [ + 76, + 702, + 181, + 715 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Baseline Enhancement", + "text_level": 1, + "bbox": [ + 76, + 726, + 287, + 739 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In addition to the findings and insights discovered in Section 4.3 by evaluating on CLOTH4D. CLOTH4D, containing various clothing types and motion sequences, as well as true-to-life multi-layer thin 3D clothing structures, poses new challenges to clothed human reconstruction research. To investigate the SOTA performance when trained on CLOTH4D, we re-train PIFu, PaMIR, and ICON as described in Section 4.1, denoting as $\\mathrm{PIFu}_{clo}$ , $\\mathrm{PaMIR}_{clo}$ , and $\\mathrm{ICON}_{clo}$ , respectively. We report the results of models trained on CLOTH4D in the gray columns in Table 2 and", + "bbox": [ + 75, + 750, + 468, + 900 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Figure 5. The following findings can be made:", + "bbox": [ + 500, + 703, + 808, + 717 + ], + "page_idx": 6 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1) In terms of quantitative metrics, the models trained on CLOTH4D generally outperform the original models trained on scan datasets as the data distributions of the training and testing datasets are closer. $\\mathrm{ICON}_{clo}^{gt}$ achieves the highest accuracy on the CLOTH4D test set. Furthermore, the qualitative results show that more high-frequency details are generated after training on CLOTH4D.", + "2) As the re-implementation setting in ICON's code allows the normal image to be input into all methods, the influence of garment print to PIFu and PaMIR are slightly relieved, validating the hypothesis we made in Section 4.3 that predicting intermediate representations like normals can re" + ], + "bbox": [ + 496, + 718, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "12853", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "duce the ambiguity of geometry and appearance.", + "bbox": [ + 76, + 90, + 398, + 106 + ], + "page_idx": 7 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "3) SOTAs fail to model layered and thin clothing structures as shown in the dress and skirt regions in Figure 5, where holes and tattered pieces are generated. Notably, the original PIFu and PaMIR can roughly generate the overall shapes of loose clothing, but they fail when trained on CLOTH4D. Since SOTA methods learn an occupancy field by sampling query points in the 3D space, for PIFu, whose spatial feature resolution is low, and for PaMIR, whose volumetric feature space is also of low resolution, it is hard to sample informative points near the thin surface. It is even harder for the network to learn if a querying point is inside or outside the mesh near the thin structure, as the inside and outside samples have very similar local features.", + "4) The difficulties of learning the occupancy field for thin faces motivates developing methods that focus more on the query points near the thin surface for better reconstruction. Future research may also seek better implicit representations to boost the performance of reconstructing multi-layer thin structures e.g., [12, 13]. However, as shown in the supplementary material (Figure B), the state-of-the-art implicit representation cannot achieve satisfactory multi-layer thin structure reconstruction even if we feed the ground truth mesh as the input to the implicit function." + ], + "bbox": [ + 75, + 107, + 470, + 455 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "As shown in Table 3 and Figure 6, the performance on the CAPE dataset drops after training on CLOTH4D due to two reasons. 1) The CAPE dataset is collected in a controlled lab environment with dim lighting, which further enlarges the domain gap between CAPE and CLOTH4D. Consequently, the predicted normal images and reconstructed meshes are less accurate. 2) CAPE dataset contains single-layer scan meshes and tight clothing. However, models trained on the multi-layer CLOTH4D dataset tend to generate gaps between the clothing and skin, resulting in holes in the clothing and unsmooth surfaces. This also verifies that local features are sensitive to overfitting. It remains a challenging but interesting problem to have a unified representation of single-layer and multi-layer data, thus different types of datasets could be trained together to yield better generalizability. Finally, we show the reconstructed samples of the in-the-wild images in Figure 7.", + "bbox": [ + 75, + 455, + 468, + 714 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.5. Limitations", + "text_level": 1, + "bbox": [ + 76, + 724, + 204, + 739 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Firstly, the simulation results are generated via graphics software, which may crash in some cases. We show some detailed examples in the supplementary material (Figure C). Meanwhile, the results on CAPE show that the clothing features of current CLOTH4D can well represent basic men's wear. However, considering the completeness, the dataset scale and diversity of subjects will be further improved to cover males and kids. In addition, the original simulated mesh sequences are non-watertight, which must be converted to watertight (and such conversion is usually lossy)", + "bbox": [ + 75, + 750, + 470, + 902 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/935dbe458a4ad764c608e737541a3c291f4300f0bca0ca5d20cc9762cedf07be.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 501, + 85, + 893, + 335 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/74a8c03af2716bb742b5369fb13107d2b3e7ae6e416f976470e6eb90d8657c01.jpg", + "image_caption": [ + "Figure 7. In-the-wild front and side view reconstructions. SOTAs with implicit functions tend to generate broken results." + ], + "image_footnote": [], + "bbox": [ + 501, + 376, + 890, + 500 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "before feeding into current clothed human reconstruction methods. It would be essential to conduct future research that reconstructs meshes with arbitrary topologies. Finally, as indicated by the results on CAPE, the generalization ability to scan data is yet to be further explored.", + "bbox": [ + 498, + 534, + 893, + 611 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 628, + 619, + 645 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We introduce CLOTH4D containing realistic and rich categories of clothing, avatars, and animations, and will release it for free, hoping to push the research on clothed human reconstruction. We evaluate current SOTAs with the newly introduced temporally-aware metrics and in-depth analyze their pros and cons by leveraging the advantages of CLOTH4D. We retrain those SOTAs on CLOTH4D, discuss the challenges the new dataset brings, and propose potential research directions. Although layering clothing in CLOTH4D brings immense difficulties to current research, we believe it is an important step toward more realistic and temporally coherent clothed human reconstruction.", + "bbox": [ + 496, + 656, + 890, + 837 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgement: This work is supported by Laboratory for Artificial Intelligence in Design (Project Code: RP 3-1) under InnoHK Research Clusters, Hong Kong SAR Government.", + "bbox": [ + 496, + 839, + 893, + 898 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "12854", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 174, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] 3dpeople. https://www.3dpeople.com. 2", + "[2] Adobe system incorporated. https://www MIXamo.com. 4", + "[3] Clo virtual fashion llc. https://www.clo3d.com. 2", + "[4] Official code of icon. https://github.com/YuliangXiu/ICON. 4", + "[5] Thiemo Alldieck, Marcus Magnor, Weipeng Xu, Christian Theobalt, and Gerard Pons-Moll. Detailed human avatars from monocular video. In 2018 International Conference on 3D Vision (3DV), pages 98-109. IEEE, 2018. 3", + "[6] Thiemo Alldieck, Gerard Pons-Moll, Christian Theobalt, and Marcus Magnor. Tex2shape: Detailed full human body geometry from a single image. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2293-2303, 2019. 3", + "[7] Thiemo Alldieck, Mihai Zanfir, and Cristian Sminchisescu. Photorealistic monocular 3d reconstruction of humans wearing clothing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1506-1515, 2022. 3, 4, 6", + "[8] Hugo Bertiche, Meysam Madadi, and Sergio Escalera. Cloth3d: clothed 3d humans. In European Conference on Computer Vision, pages 344-359. Springer, 2020. 2, 3, 4", + "[9] Bharat Lal Bhatnagar, Garvita Tiwari, Christian Theobalt, and Gerard Pons-Moll. Multi-garment net: Learning to dress 3d people from images. In proceedings of the IEEE/CVF international conference on computer vision, pages 5420-5430, 2019. 3", + "[10] Zhongang Cai, Daxuan Ren, Ailing Zeng, Zhengyu Lin, Tao Yu, Wenjia Wang, Xiangyu Fan, Yang Gao, Yifan Yu, Liang Pan, et al. Humman: Multi-modal 4d human dataset for versatile sensing and modeling. arXiv preprint arXiv:2204.13686, 2022. 3", + "[11] Z. Cao, G. Hidalgo Martinez, T. Simon, S. Wei, and Y. A. Sheikh. Openpose: Realtime multi-person 2d pose estimation using part affinity fields. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2019. 4", + "[12] Weikai Chen, Cheng Lin, Weiyang Li, and Bo Yang. 3psdf: Three-pole signed distance function for learning surfaces with arbitrary topologies. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18522-18531, 2022. 8", + "[13] Julian Chibane, Gerard Pons-Moll, et al. Neural unsigned distance fields for implicit function learning. Advances in Neural Information Processing Systems, 33:21638-21652, 2020. 8", + "[14] Mikhail Erofeev, Yury Gitman, Dmitriy S Vatolin, Alexey Fedorov, and Jue Wang. Perceptually motivated benchmark for video matting. In BMVC, pages 99-1, 2015. 4, 5", + "[15] Qiao Feng, Yebin Liu, Yu-Kun Lai, Jingyu Yang, and Kun Li. Fof: Learning fourier occupancy field for monocular real-time human reconstruction. arXiv preprint arXiv:2206.02194, 2022. 3, 4, 5", + "[16] Peng Guan, Loretta Reiss, David A Hirshberg, Alexander Weiss, and Michael J Black. Drape: Dressing any person. ACM Transactions on Graphics (ToG), 31(4):1-10, 2012. 2, 3" + ], + "bbox": [ + 78, + 114, + 467, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[17] Erhan Gundogdu, Victor Constantin, Amrollah Seifoddini, Minh Dang, Mathieu Salzmann, and Pascal Fua. Garnet: A two-stream network for fast and accurate 3d cloth draping. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8739-8748, 2019. 2, 3", + "[18] Xintong Han, Xiaojun Hu, Weilin Huang, and Matthew R Scott. Clothflow: A flow-based model for clothed person generation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10471-10480, 2019. 4", + "[19] Xintong Han, Zuxuan Wu, Zhe Wu, Ruichi Yu, and Larry S Davis. Viton: An image-based virtual try-on network. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7543-7552, 2018. 4", + "[20] Tong He, John Collomosse, Hailin Jin, and Stefano Soatto. Geo-pifu: Geometry and pixel aligned implicit functions for single-view human reconstruction. Advances in Neural Information Processing Systems, 33:9276-9287, 2020. 3", + "[21] Tong He, Yuanlu Xu, Shunsuke Saito, Stefano Soatto, and Tony Tung. Arch++: Animation-ready clothed human reconstruction revisited. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11046-11056, 2021. 3, 4", + "[22] Zhu Heming, Cao Yu, Jin Hang, Chen Weikai, Du Dong, Wang Zhangye, Cui Shuguang, and Han Xiaoguang. Deep fashion3d: A dataset and benchmark for 3d garment reconstruction from single images. In Computer Vision - ECCV 2020, pages 512-530. Springer International Publishing, 2020. 3", + "[23] Jingwei Huang, Yichao Zhou, and Leonidas Guibas. Manifoldplus: A robust and scalable watertight manifold surface generation method for triangle soups. arXiv preprint arXiv:2005.11621, 2020.4", + "[24] Zeng Huang, Yuanlu Xu, Christoph Lassner, Hao Li, and Tony Tung. Arch: Animatable reconstruction of clothed humans. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3093-3102, 2020. 3", + "[25] Catalin Ionescu, Dragos Papava, Vlad Olaru, and Cristian Sminchisescu. Human3.6m: Large scale datasets and predictive methods for 3d human sensing in natural environments. IEEE transactions on pattern analysis and machine intelligence, 36(7):1325-1339, 2013. 3", + "[26] Boyi Jiang, Juyong Zhang, Yang Hong, Jinhao Luo, Ligang Liu, and Hujun Bao. BCnet: Learning body and cloth shape from a single image. In European Conference on Computer Vision, pages 18-35. Springer, 2020. 3", + "[27] Muhammed Kocabas, Nikos Athanasiou, and Michael J Black. Vibe: Video inference for human body pose and shape estimation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5253-5263, 2020. 6", + "[28] Nikos Kolotouros, Georgios Pavlakos, Michael J Black, and Kostas Daniilidis. Learning to reconstruct 3d human pose and shape via model-fitting in the loop. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2252-2261, 2019. 3" + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "12855", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[29] Nikos Kolotouros, Georgios Pavlakos, and Kostas Dani-ilidis. Convolutional mesh regression for single-image human shape reconstruction. In CVPR, 2019. 4", + "[30] Zorah Lahner, Daniel Cremers, and Tony Tung. Deepwrinkles: Accurate and realistic clothing modeling. In Proceedings of the European conference on computer vision (ECCV), pages 667-684, 2018. 2", + "[31] Verica Lazova, Eldar Insafutdinov, and Gerard Pons-Moll. 360-degree textures of people in clothing from a single image. In 2019 International Conference on 3D Vision (3DV), pages 643-653. IEEE, 2019. 3", + "[32] Shanchuan Lin, Linjie Yang, Imran Saleemi, and Soumyadip Sengupta. Robust high-resolution video matting with temporal guidance. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 238-247, 2022. 4, 6", + "[33] Matthew Loper, Naureen Mahmood, Javier Romero, Gerard Pons-Moll, and Michael J Black. SMPL: A skinned multiperson linear model. ACM transactions on graphics (TOG), 34(6):1-16, 2015. 2, 3", + "[34] William E Lorensen and Harvey E Cline. Marching cubes: A high resolution 3d surface construction algorithm. ACM siggraph computer graphics, 21(4):163-169, 1987. 2", + "[35] Qianli Ma, Jinlong Yang, Anurag Ranjan, Sergi Pujades, Gerard Pons-Moll, Siyu Tang, and Michael J Black. Learning to dress 3d people in generative clothing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6469-6478, 2020. 2, 4", + "[36] Meysam Madadi, Hugo Bertiche, Wafa Bouzouita, Isabelle Guyon, and Sergio Escalera. Learning cloth dynamics: 3d + texture garment reconstruction benchmark. In Proceedings of the NeurIPS 2020 Competition and Demonstration Track, PMLR, volume 133, pages 57-76, 2021. 2, 3", + "[37] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In ECCV, 2020. 6", + "[38] Chaitanya Patel, Zhouyingcheng Liao, and Gerard Pons-Moll. Tailornet: Predicting clothing in 3d as a function of human pose, shape and garment style. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR). IEEE, jun 2020. 2, 3", + "[39] Gerard Pons-Moll, Sergi Pujades, Sonny Hu, and Michael J Black. Clothcap: Seamless 4d clothing capture and retargeting. ACM Transactions on Graphics (ToG), 36(4):1-15, 2017. 3", + "[40] Albert Pumarola, Jordi Sanchez, G. Choi, A. Sanfeliu, and F. Moreno-Noguer. 3dpeople: Modeling the geometry of dressed humans. 2019 IEEE/CVF International Conference on Computer Vision (ICCV), pages 2242-2251, 2019. 2", + "[41] Shunsuke Saito, Zeng Huang, Ryota Natsume, Shigeo Morishima, Angjoo Kanazawa, and Hao Li. Pifu: Pixel-aligned implicit function for high-resolution clothed human digitization. In Proceedings of the IEEE International Conference on Computer Vision, pages 2304-2314, 2019. 3, 4", + "[42] Shunsuke Saito, Tomas Simon, Jason Saragih, and Hanbyul Joo. Pifuhd: Multi-level pixel-aligned implicit function for" + ], + "bbox": [ + 78, + 90, + 468, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "high-resolution 3d human digitization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 84-93, 2020. 3, 4", + "[43] Igor Santesteban, Miguel A Otaduy, and Dan Casas. Learning-based animation of clothing for virtual try-on. In Computer Graphics Forum, volume 38, pages 355-366. Wiley Online Library, 2019. 2, 3", + "[44] Leonid Sigal, Alexandru O Balan, and Michael J Black. Humaneva: Synchronized video and motion capture dataset and baseline algorithm for evaluation of articulated human motion. International journal of computer vision, 87(1):4-27, 2010. 3", + "[45] Peter-Pike Sloan, Jan Kautz, and John Snyder. Precomputed radiance transfer for real-time rendering in dynamic, low-frequency lighting environments. In Proceedings of the 29th annual conference on Computer graphics and interactive techniques, pages 527-536, 2002. 4", + "[46] Matthew Trumble, Andrew Gilbert, Charles Malleson, A. Hilton, and J. Collomosse. Total capture: 3d human pose estimation fusing video and inertial sensors. In BMVC, 2017. 3", + "[47] Tuanfeng Y Wang, Duygu Ceylan, Jovan Popovic, and Niloy J Mitra. Learning a shared shape space for multimodal garment design. arXiv preprint arXiv:1806.11335, 2018. 2, 3", + "[48] Donglai Xiang, Fabian Prada, Chenglei Wu, and Jessica Hodgins. Monoclothcap: Towards temporally coherent clothing capture from monocular rgb video. In 2020 International Conference on 3D Vision (3DV), pages 322-332. IEEE, 2020. 3", + "[49] Yuliang Xiu, Jinlong Yang, Dimitrios Tzionas, and Michael J Black. Icon: Implicit clothed humans obtained from normals. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 13286-13296. IEEE, 2022. 3, 4, 5, 6", + "[50] Shan Yang, Tanya Ambert, Zherong Pan, Ke Wang, Licheng Yu, Tamara Berg, and Ming C Lin. Detailed garment recovery from a single-view image. arXiv preprint arXiv:1608.01250, 2016. 4", + "[51] Tao Yu, Zerong Zheng, Kaiwen Guo, Pengpeng Liu, Qionghai Dai, and Yebin Liu. Function4d: Real-time human volumetric capture from very sparse consumer rgbd sensors. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR2021), June 2021. 2", + "[52] Chao Zhang, Sergi Pujades, Michael J Black, and Gerard Pons-Moll. Detailed, accurate, human shape estimation from clothed 3d scan sequences. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 4191-4200, 2017. 2", + "[53] Hongwen Zhang, Yating Tian, Xinchi Zhou, Wanli Ouyang, Yebin Liu, Limin Wang, and Zhenan Sun. Pymaf: 3d human pose and shape regression with pyramidal mesh alignment feedback loop. In Proceedings of the IEEE International Conference on Computer Vision, 2021. 4", + "[54] Zerong Zheng, Tao Yu, Yebin Liu, and Qionghai Dai. Pamir: Parametric model-conditioned implicit representation for image-based human reconstruction. IEEE transactions on" + ], + "bbox": [ + 503, + 92, + 892, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "12856", + "bbox": [ + 480, + 945, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "pattern analysis and machine intelligence, 44(6):3170-3184, 2021. 3, 4", + "[55] Boyao Zhou, Jean-Sébastien Franco, Federica Bogo, and Edmond Boyer. Spatio-temporal human shape completion with implicit function networks. In 2021 International Conference on 3D Vision (3DV), pages 669-678. IEEE, 2021. 6", + "[56] Xingxing Zou, Xiangheng Kong, Waikeung Wong, Congde Wang, Yuguang Liu, and Yang Cao. Fashionai: A hierarchical dataset for fashion understanding. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops, pages 0-0, 2019. 4", + "[57] Xingxing Zou and Waikeung Wong. fashion after fashion: A report of ai in fashion. arXiv preprint arXiv:2105.03050, 2021.4" + ], + "bbox": [ + 78, + 90, + 468, + 287 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "12857", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2023/CLOTH4D_ A Dataset for Clothed Human Reconstruction/b8c1cf88-b3e6-48d2-932c-20d8bacc61eb_model.json b/2023/CLOTH4D_ A Dataset for Clothed Human Reconstruction/b8c1cf88-b3e6-48d2-932c-20d8bacc61eb_model.json new file mode 100644 index 0000000000000000000000000000000000000000..f8c6cee2d09d3b7ca431ad2fab7d26ca183504b5 --- /dev/null +++ b/2023/CLOTH4D_ A Dataset for Clothed Human Reconstruction/b8c1cf88-b3e6-48d2-932c-20d8bacc61eb_model.json @@ -0,0 +1,2246 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.19, + 0.131, + 0.78, + 0.151 + ], + "angle": 0, + "content": "CLOTH4D: A Dataset for Clothed Human Reconstruction" + }, + { + "type": "text", + "bbox": [ + 0.269, + 0.18, + 0.7, + 0.2 + ], + "angle": 0, + "content": "Xingxing Zou\\(^{1,3*}\\), Xintong Han\\(^{2*}\\), Waikeung Wong\\(^{3,1\\dagger}\\)" + }, + { + "type": "text", + "bbox": [ + 0.249, + 0.2, + 0.721, + 0.217 + ], + "angle": 0, + "content": "\\(^{1}\\)Laboratory for Artificial Intelligence in Design, \\(^{2}\\)Huya Inc." + }, + { + "type": "text", + "bbox": [ + 0.199, + 0.217, + 0.771, + 0.235 + ], + "angle": 0, + "content": "\\(^{3}\\)School of Fashion and Textiles, The Hong Kong Polytechnic University" + }, + { + "type": "text", + "bbox": [ + 0.165, + 0.238, + 0.804, + 0.251 + ], + "angle": 0, + "content": "aemika.zou@connect.polyu.hk,hanxintong@huya.com,calvin.wong@polyu.edu.hk" + }, + { + "type": "image", + "bbox": [ + 0.098, + 0.294, + 0.279, + 0.364 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.135, + 0.369, + 0.227, + 0.383 + ], + "angle": 0, + "content": "(a) 2D patterns" + }, + { + "type": "image", + "bbox": [ + 0.093, + 0.396, + 0.13, + 0.459 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.147, + 0.395, + 0.177, + 0.459 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.198, + 0.396, + 0.227, + 0.459 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.242, + 0.396, + 0.28, + 0.459 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.132, + 0.465, + 0.228, + 0.479 + ], + "angle": 0, + "content": "(b) varied poses" + }, + { + "type": "image", + "bbox": [ + 0.084, + 0.489, + 0.133, + 0.551 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.141, + 0.49, + 0.179, + 0.551 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.198, + 0.49, + 0.226, + 0.551 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.235, + 0.49, + 0.284, + 0.551 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.134, + 0.556, + 0.227, + 0.568 + ], + "angle": 0, + "content": "(c) multi views" + }, + { + "type": "image", + "bbox": [ + 0.304, + 0.288, + 0.884, + 0.554 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.493, + 0.556, + 0.758, + 0.569 + ], + "angle": 0, + "content": "(d) series of paired dynamic textured meshes" + }, + { + "type": "image_caption", + "bbox": [ + 0.075, + 0.57, + 0.893, + 0.598 + ], + "angle": 0, + "content": "Figure 1. An instance in CLOTH4D. Given (a) hand-crafted 2D sewing patterns, clothes are simulated with virtual fashion software to have (b) varied poses, (c) multiple views, and (d) high-quality and physically plausible series of paired dynamic 3D meshes with textures." + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.61, + 0.314, + 0.626 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.645, + 0.474, + 0.903 + ], + "angle": 0, + "content": "Clothed human reconstruction is the cornerstone for creating the virtual world. To a great extent, the quality of recovered avatars decides whether the Metaverse is a passing fad. In this work, we introduce CLOTH4D, a clothed human dataset containing 1,000 subjects with varied appearances, 1,000 3D outfits, and over 100,000 clothed meshes with paired unclothed humans, to fill the gap in large-scale and high-quality 4D clothing data. It enjoys appealing characteristics: 1) Accurate and detailed clothing textured meshes—all clothing items are manually created and then simulated in professional software, strictly following the general standard in fashion design. 2) Separated textured clothing and under-clothing body meshes, closer to the physical world than single-layer raw scans. 3) Clothed human motion sequences simulated given a set of 289 actions, covering fundamental and complicated dynamics. Upon CLOTH4D, we novelly designed a series of" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.612, + 0.895, + 0.718 + ], + "angle": 0, + "content": "temporally-aware metrics to evaluate the temporal stability of the generated 3D human meshes, which has been overlooked previously. Moreover, by assessing and retraining current state-of-the-art clothed human reconstruction methods, we reveal insights, present improved performance, and propose potential future research directions, confirming our dataset's advancement. The dataset is available at1." + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.735, + 0.633, + 0.751 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.76, + 0.894, + 0.867 + ], + "angle": 0, + "content": "As we enter the volumetric and XR content era, researchers have been trailblazing their way into the Metaverse. With the converging of technologies and practical applications, e.g., fashion NFTs (non-fungible tokens), immersive AR and VR, and games, clothed human reconstruction demands are rapidly growing. While current research has made astonishing results in creating digital hu" + }, + { + "type": "page_footnote", + "bbox": [ + 0.516, + 0.875, + 0.882, + 0.888 + ], + "angle": 0, + "content": "1www.github.com/AemikaChow/AiDLab-fAshIon-Data" + }, + { + "type": "page_footnote", + "bbox": [ + 0.521, + 0.888, + 0.859, + 0.901 + ], + "angle": 0, + "content": "*X. Zou and X. Han contribute equally. † Corresponding author." + }, + { + "type": "list", + "bbox": [ + 0.516, + 0.875, + 0.882, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12847" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.09, + 0.895, + 0.165 + ], + "angle": 0, + "content": "Table 1. Comparisons of CLOTH4D with existing representative datasets. Gray color indicates synthetic datasets generated with graphics engines. #Subjects: number of peoples in different appearances; #Action: number of actions adopted; #Scans: numbers of 3D meshes; 2D Pattern: 2D clothing pattern; TexCloth: with textured clothed model; TexHuman: with textured naked human model. w/ SMPL: with registered SMPL [33] parameters. Public: publicly available and free of charge. Photorealistic: whether the images in the dataset are realistic. -: not applicable or reported. CLOTH4D presents more desirable characteristics compared with others." + }, + { + "type": "table", + "bbox": [ + 0.081, + 0.167, + 0.89, + 0.384 + ], + "angle": 0, + "content": "
Dataset#Subjects#Action#Scan2D PatternTexClothTexHumanw/ SMPLPublicPhotorealistic
BUFF [52]6-13.6k-✓-✓✓✓
RenderPeople [1]--825-✓-✓-✓
DeepWrinkles [30]229.2k-✓---✓
CAPE [35]15600140k---✓✓✓
THuman2.0 [51]200-525-✓-✓✓✓
DRAPE [16]72324.5k------
Wang et al. [47]--24k✓✓-✓✓-
3DPeople [40]8072--✓---✓
DCA [43]-567.1k---✓--
GarNet [17]600-18.8k---✓✓-
TailorNet [38]9-5.5k-✓-✓✓-
Cloth3D [8]8.5k7.9k2.1M-✓-✓✓-
Cloth3D++ [36]9.7k8k2.2M✓✓✓✓✓-
CLOTH4D1k289100k✓✓✓✓✓✓
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.391, + 0.471, + 0.482 + ], + "angle": 0, + "content": "mans, these reconstructed meshes have issues, e.g., flexible body motions and diverse appearances, owing to the lack of datasets with richness in clothing and realistic dynamics of garments. To this end, we introduce CLOTH4D, an open-sourced dataset facilitating physically plausible dynamic clothed human reconstruction." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.487, + 0.471, + 0.806 + ], + "angle": 0, + "content": "Prior to us, many datasets have been collected, and we sort out them in Table 1. Currently, scanned datasets are widely adopted as they are photorealistic and can be easily processed to watertight meshes, which does an excellent favor for current deep models to learn an implicit function (e.g., signed distance function) followed by marching cubes [34] for surface reconstruction. However, it is born with some weaknesses: 1) Scanned meshes are single-layer and inherently fail to capture the space between clothing and skin surface. Thus, body shape under clothing cannot be accurately inferred, let alone the multi-layer and thin clothing structures as in the real physical world. 2) It is time-consuming and expensive to obtain high-quality and large-scale temporal scanned sequences (i.e., 4D scanned sequences) due to the limited efficiency and precision of 4D scanners, especially for complicated clothing and large motions. Although synthetic datasets can to some extent overcome these limitations, existing synthetic datasets are either of small scale in terms of appearances and motions or are highly unrealistic. Moreover, many datasets are not made publicly available and free." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.811, + 0.472, + 0.904 + ], + "angle": 0, + "content": "In contrast, CLOTH4D possesses several attractive attributes: 1) We made great efforts to the diversity and quality of clothing. All clothes are manually designed in CLO [3] and cater to the requirement of the fashion industry. 2) Meshes in CLOTH4D are clothing/humans separated. Such flexibility makes studying and modeling the" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.391, + 0.895, + 0.482 + ], + "angle": 0, + "content": "relations and interactions between clothing simulation and body movement possible. 3) CLOTH4D provides plenty of temporal motion sequences with realistic clothing dynamics. As the human body moves, the dressed clothing, e.g., the skirt in Figure 1, naturally deforms. 4) The dataset is large-scale and openly accessible." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.484, + 0.895, + 0.742 + ], + "angle": 0, + "content": "To demonstrate the advantages of CLOTH4D, we use it to evaluate the state-of-the-art (SOTA) clothed human reconstruction methods. In addition to the generally adopted static evaluation metrics, we propose a set of temporally-aware metrics to assess the temporal coherence in a video inference scenario thanks to the rich and true-to-life 4D synthetic sequences in the dataset. Quantitative and qualitative results of SOTA methods on CLOTH4D suggest that our dataset is challenging and the temporal stability of the reconstructed mesh is vital for evaluating the perceptual quality. Meanwhile, we retrain SOTA methods on CLOTH4D, revealing interesting observations of how they perform on multi-layer meshes with thin clothing structures. With in-depth analysis and a summary of challenges for the existing approaches, CLOTH4D makes an essential step toward more realistic reconstructions of clothed humans and stimulates several exciting future work directions. All in all:" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.744, + 0.895, + 0.789 + ], + "angle": 0, + "content": "- We contribute CLOTH4D, a large-scale, high-quality, and open-accessible 4D synthetic dataset for clothed human reconstruction." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.792, + 0.895, + 0.838 + ], + "angle": 0, + "content": "- We introduce a series of temporally-aware metrics to evaluate the reconstructed performance in the aspect of temporal consistency." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.895, + 0.902 + ], + "angle": 0, + "content": "- With the proposed dataset and metrics, we thoroughly analyze the pros and cons of SOTAs, summarize the existing challenges toward more realistic 3D modeling, and propose potential new directions." + }, + { + "type": "list", + "bbox": [ + 0.498, + 0.744, + 0.895, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "12848" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.08, + 0.08, + 0.357, + 0.203 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.373, + 0.081, + 0.619, + 0.2 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.624, + 0.081, + 0.893, + 0.207 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.08, + 0.202, + 0.357, + 0.269 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.129, + 0.279, + 0.201, + 0.289 + ], + "angle": 0, + "content": "(i) pattern making" + }, + { + "type": "image", + "bbox": [ + 0.278, + 0.221, + 0.356, + 0.277 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.272, + 0.28, + 0.358, + 0.289 + ], + "angle": 0, + "content": "(ii) clothed 3D avatar" + }, + { + "type": "image", + "bbox": [ + 0.402, + 0.224, + 0.523, + 0.28 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.41, + 0.281, + 0.503, + 0.289 + ], + "angle": 0, + "content": "(iii) clothing simulation" + }, + { + "type": "image", + "bbox": [ + 0.556, + 0.253, + 0.58, + 0.263 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.593, + 0.215, + 0.893, + 0.293 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.669, + 0.295, + 0.817, + 0.304 + ], + "angle": 0, + "content": "CLOTH4D with varied 3D sequences" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.309, + 0.893, + 0.337 + ], + "angle": 0, + "content": "Figure 2. Pipeline for creating instances in CLOTH4D, which primarily adopts CLO for clothing design and simulation, Mixamo for animation, and Blender for processing and exporting meshes." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.343, + 0.219, + 0.358 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.37, + 0.47, + 0.611 + ], + "angle": 0, + "content": "Clothed Human Reconstruction. Clothed human reconstruction aims to recover a 3D mesh from a monocular person image. Estimating dressed people by modeling clothing geometry as 3D displacement on top of a parametric body model (e.g., SMPL [33]) is the leading solution to deal with this task [5,6,28,31,39,48]. By transferring the skinning weights from the body model to the offset clothing mesh, the reconstructed clothed mesh can be readily deformed and animated in the same way as the under-clothing 3D parametric body model. However, it assumes the clothed human to have the same topology as the naked body, leading to unsatisfactory reconstructions of long hair, skirts, dresses, etc. Although methods such as [9,22,26,38] try to isolate the reconstruction of clothing by constructing category-specific clothing templates or statistical models, they fail to generalize to unseen or complex types of clothing." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.613, + 0.47, + 0.869 + ], + "angle": 0, + "content": "As another line of research, deep implicit function networks have drawn broader attention recently [7, 15, 20, 21, 24, 41, 42, 54]. PIFu [41] conditions on pixel-aligned features to build deep implicit functions for reconstructing human meshes, and PIFuHD [42] goes a further step towards enhancing 3D geometric details by predicting front and back normals. ARCH [24] and \\(\\mathrm{ARCH}++\\) [21] enable animating the reconstructed meshes by deforming the semantic information into the SMPL canonical space. More recently, leveraging SMPL body models as prior, PaMIR [54] and ICON [49] further improve the reconstruction quality, especially on challenging poses. PHORHUM [7] achieves more accurate results via jointly estimating geometry, albedo, and shading information. It is worth noting that most works rely on training data that are not available free of charge, making an open-sourced dataset of clothed avatars of great significance to advance research in 3D body/clothing modeling." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.871, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Clothed Human Datasets. As summarized in Table 1, existing clothed human datasets can be divided into two" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.344, + 0.893, + 0.601 + ], + "angle": 0, + "content": "types, i.e., scanned datasets and synthetic datasets. The former [10, 25, 44, 46] utilizes multiple synchronized cameras to capture motions that have difficulties enriching the scalability and diversity and obtaining highly accurate 4D ground truth. Moreover, it inherently cannot mimic the layering of clothing. The synthetic datasets mitigate these limitations. However, existing synthetic datasets, no matter whether static [17, 38, 47] or dynamic [8, 16, 39, 43], either only contain a few clothing types or are highly unrealistic. Cloth3D++ [36] is developed from Cloth3D [8], which contains a total of 2.2 million scans, covering 9.7k subjects dressed in 12.9k clothing, which is the most advanced in scale. However, the clothing created based on garment templates are only base patterns with an immense gap compared to clothing in real life. To intuitively demonstrate the advantages of CLOTH4D, we put visual comparisons of these datasets in Figure A in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.617, + 0.687, + 0.633 + ], + "angle": 0, + "content": "3. CLOTH4D Dataset" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.643, + 0.892, + 0.749 + ], + "angle": 0, + "content": "We depict the pipeline of creating CLOTH4D in Figure 2, including (1) preparing the referenced clothing images and the unclothed 3D human avatars; (2) uploading 3D human avatars to Mixamo to obtain the FBX files with various animations; (3) designing 3D clothing in CLO, integrating the FBX files obtained in (2), and conducting clothing simulation; and (4) exporting the sequenced mesh files." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.893, + 0.901 + ], + "angle": 0, + "content": "Clothes and Models. Clothes are manually created by professional fashion designers using CLO by producing the 2D garment patterns and then auto-simulating them to 3D. We put a video illustrating the production process of a 3D garment in the supplementary material (Video A) instead of demonstrating details in the paper. CLOTH4D covers 1,000 different 3D outfits spansning over 500 prints and 50 fabrics. Over \\(40\\%\\) of the clothes are designed referring to the newest collections of varying design houses (e.g., Prada, Moschino, and Alexander McQueen) to ensure visual re" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12849" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.182 + ], + "angle": 0, + "content": "alism, variance, and fashionability. In this paper, we focus on women's wear owing to its diversity covering most characteristics of garments. For avatars, CLO provides human avatars with varied physical appearances in terms of hairstyles, faces, skin colors, body figures, etc., and we directly adopt these available avatars." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.183, + 0.471, + 0.363 + ], + "angle": 0, + "content": "Animations and Simulations. For each human avatar, we use Mixamo [2] for rigging and generating motion sequences. Given a motion sequence and a 3D garment, CLO runs a clothing simulation of the 3D garment according to the motion sequence at 30fps. A total of 289 animations are utilized, such as Billy Dance, Offensive Idle, Jumping Rope, etc. Unlike previous work [8] directly uses Blender for cloth simulation, CLO simulates clothing with richer details and dynamics. The clothes exhibit wrinkles and fold with the body's movement in a natural and physically plausible way, especially for skirts or dresses that differ a lot from the body topology as shown in Figure 1." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.364, + 0.471, + 0.591 + ], + "angle": 0, + "content": "Paired Multi-Modal Data. Then, we can obtain paired data in the following representations: a 2D clothing pattern, 3D mesh sequences (clothed and naked human meshes with corresponding fitted SMPL parameters/meshes, and separate clothing meshes), and a UV texture map. The textured mesh can be rendered into multi-view normal images, depth images, and RGB images given varying light conditions. We also translate all these dynamic meshes to watertight with simplification using [23], thus they can be readily used to train an implicit function for mesh reconstruction. The number of triangles ranges from \\(170K\\) to \\(4M\\) for simulation and becomes \\(200K\\) after simplification. Besides human reconstruction, many other tasks could also benefit from this paired data (e.g., clothing capture [50], human pose transfer [18], and fashion-related tasks [19, 56, 57])." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.603, + 0.203, + 0.619 + ], + "angle": 0, + "content": "4. Evaluations" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.628, + 0.47, + 0.75 + ], + "angle": 0, + "content": "In this section, we evaluate the state-of-the-art clothed human reconstruction methods on CLOTH4D to demonstrate new insights that the dataset can provide. Further, we retrain SOTA methods on CLOTH4D and make several interesting observations of how they perform on multi-layer meshes with thin clothing structures. We also present the challenges for existing approaches and propose potential research directions with a comprehensive analysis." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.757, + 0.185, + 0.772 + ], + "angle": 0, + "content": "4.1. Baselines" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.78, + 0.471, + 0.903 + ], + "angle": 0, + "content": "We mainly report results of four SOTA approaches, PIFu [41], PIFuHD [42], PaMIR [54], and ICON [49] owing to other works do not release their codes or models, such as PHORHUM [7], ARCH++ [21], or which have already been extensively compared with the listed methods above. We use PIFu, PIFuHD, PaMIR, and ICON to denote their released pretrained testing models. \\(\\mathrm{PaMIR}^{gt}\\) and \\(\\mathrm{ICON}^{gt}\\) indicate the testing results using the fitted ground truth" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.212 + ], + "angle": 0, + "content": "SMPL mesh as conditions rather than the one predicted using off-the-shelf human mesh recovery (HMR) methods (GraphCMR [29] for PaMIR, and PyMAf [53] for ICON as in their released code). Based on the characteristics of these methods, they can be divided into three types: 1) pixel-aligned methods (PIFu, PIFuHD); 2) GT-SMPL-guided + pixel-aligned methods \\((\\mathrm{PaMIR}^{gt}\\) and \\(\\mathrm{ICON}^{gt})\\); 3) HMR-SMPL-guided + pixel-aligned methods (PaMIR and ICON)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.213, + 0.892, + 0.349 + ], + "angle": 0, + "content": "Furthermore, we retrain PIFu, PaMIR, and ICON on CLOTH4D, which are denoted as \\(\\mathrm{PIF_u_{clo}}\\), \\(\\mathrm{PaMIR}_{clo}\\), and \\(\\mathrm{ICON}_{clo}\\), respectively. For these retrained models, we follow the re-implementation setting introduced in ICON [4], which allows us to train all these baselines with the same training protocol and hyper-parameters for a fair comparison. Similarly, \\(\\mathrm{PaMIR}_{clo}^{gt}\\), and \\(\\mathrm{ICON}_{clo}^{gt}\\) are tested with the ground truth SMPL fits. We also use the cloth-refinement module in the ICON's released code for post-processing." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.356, + 0.7, + 0.372 + ], + "angle": 0, + "content": "4.2. Datasets and Metrics" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.38, + 0.892, + 0.516 + ], + "angle": 0, + "content": "Datasets and implementation details. We organize the sequences in CLOTH4D into a \\(80\\% / 10\\% / 10\\%\\) train/val/test split. We render each mesh into 8 views using a weak perspective camera and pre-computed radiance transfer [45] with dynamic light conditions following [41, 49]. All rendered images are \\(512 \\times 512\\). The 2D keypoints used in all methods are generated by OpenPose [11]. In addition, we also evaluate all models on CAPE [35] test set adopted in [49] to investigate the generalization ability." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.515, + 0.892, + 0.636 + ], + "angle": 0, + "content": "Static Metrics. We report the quantitative results on normal reprojection error, Chamfer distance, and P2S distance for evaluation as [15, 41, 42, 49]. As all compared methods use a weak perspective or orthographic camera, the estimated meshes may not be well aligned with the ground truth meshes in the \\(z\\)-direction (i.e., view direction). Thus, we shift the estimated meshes to have the same \\(z\\)-axis mean as the ground truths following [15] for a fair comparison." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.636, + 0.892, + 0.801 + ], + "angle": 0, + "content": "Temporal Metrics. The aforementioned static metrics ignore the temporal consistency of the reconstructed meshes across time, which is essential for real-time applications since meshes presenting jitters and flickers highly affect the perceptual quality. Thanks to the rich temporal dynamics provided in CLOTH4D, we are the first to introduce temporally-aware metrics to evaluate the temporal coherence of the generated mesh sequences. Referring to temporal metrics SSDdt and dtSSD used in video matting tasks [14, 32], we compute two metrics measuring the temporal coherence of the predicted mesh normal:" + }, + { + "type": "equation", + "bbox": [ + 0.511, + 0.805, + 0.892, + 0.833 + ], + "angle": 0, + "content": "\\[\n\\text {N o r m a l s} _ {d d t} = \\frac {1}{T} \\sum_ {t} \\left| \\left(\\mathcal {N} _ {t} ^ {p r} - \\mathcal {N} _ {t} ^ {g t}\\right) ^ {2} - \\left(\\mathcal {N} _ {t + 1} ^ {p r} - \\mathcal {N} _ {t + 1} ^ {g t}\\right) ^ {2} \\right|, \\tag {1}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.511, + 0.838, + 0.892, + 0.867 + ], + "angle": 0, + "content": "\\[\n\\operatorname {N o r m a l s} _ {d t d} = \\frac {1}{T} \\sum_ {t} \\left| \\left(\\mathcal {N} _ {t} ^ {p r} - \\mathcal {N} _ {t + 1} ^ {p r}\\right) ^ {2} - \\left(\\mathcal {N} _ {t} ^ {g t} - \\mathcal {N} _ {t + 1} ^ {g t}\\right) ^ {2} \\right|, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.87, + 0.892, + 0.901 + ], + "angle": 0, + "content": "where \\(T\\) is the length of the sequence. \\(\\mathcal{N}_t^{pr}\\) and \\(\\mathcal{N}_t^{gt}\\) denote the rendered normal images from the predicted mesh" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12850" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.089, + 0.892, + 0.119 + ], + "angle": 0, + "content": "Table 2. Quantitative evaluation on CLOTH4D. PaMIRgt and ICONgt denote that the fitted ground truth SMPL is used during the inference. Gray color indicates the results trained on CLOTH4D." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.121, + 0.89, + 0.263 + ], + "angle": 0, + "content": "
MethodPIFuPIFuHDPaMIR\\( PaMIR^{gt} \\)ICON\\( ICON^{gt} \\)\\( PIFu_{clo} \\)\\( PaMIR_{clo} \\)\\( PaMIR^{gt}_{clo} \\)\\( ICON_{clo} \\)\\( ICON^{gt}_{clo} \\)
Normals \\( \\downarrow \\)0.1820.1810.2240.1450.2110.1180.1500.2300.1140.1980.103
P2S \\( \\downarrow \\)3.9113.5183.8272.6414.6602.4732.7935.0372.6193.7112.068
Chamfer \\( \\downarrow \\)3.5782.4874.1572.4874.1961.6312.4124.1861.6183.4991.367
\\( Normals_{ddt} \\) \\( \\downarrow \\)0.0080.0070.0320.0250.0130.0300.0090.0210.0330.0130.035
\\( Normals_{dtd} \\) \\( \\downarrow \\)0.0250.0230.0430.0350.0340.0280.0170.0400.0270.0380.033
\\( P2S_{ddt} \\) \\( \\downarrow \\)0.2220.1670.2530.2470.3830.3690.1850.4170.4020.3210.367
\\( P2S_{dtd} \\) \\( \\downarrow \\)0.7330.5781.0250.7020.9790.5540.5480.8900.5360.8780.526
Chamfer \\( _{ddt} \\) \\( \\downarrow \\)0.2070.1810.3310.3160.3590.3580.1570.3500.3820.3310.369
Chamfer \\( _{dtd} \\) \\( \\downarrow \\)0.7290.5761.0170.7010.9740.5530.5460.8860.5360.8730.525
" + }, + { + "type": "table_caption", + "bbox": [ + 0.238, + 0.269, + 0.731, + 0.283 + ], + "angle": 0, + "content": "Table 3. Quantitative evaluation on CAPE. Table notations are the same as Table 2." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.285, + 0.89, + 0.352 + ], + "angle": 0, + "content": "
MethodPIFuPIFuHDPaMIRPaMIRgtICONICONgtPIFucloPaMIRcloPaMIRgtcloICONcloICONgtclo
Normals ↓0.1610.1600.1830.0860.1600.0560.1640.1760.0930.1560.077
P2S ↓4.2593.7953.8401.1934.0141.0674.6524.2351.4913.3041.193
Chamfer ↓4.2043.9274.2581.6543.9621.0384.3814.0801.4273.5441.397
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.36, + 0.473, + 0.496 + ], + "angle": 0, + "content": "and the ground truth mesh at time step \\( t \\), respectively. The subscript \\( ddt \\) is short for distance delta time, which captures the stability of errors between two consecutive meshes. And \\( dtd \\) (delta time distance) penalizes large temporal change of the prediction with respect to the change of the ground truth. These two metrics indicate unstable mesh variations and ignore temporally coherent errors [14]. The \\( ddt \\) and \\( dtd \\) of Chamfer and P2S distances are similarly defined. More details can be found in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.51, + 0.267, + 0.525 + ], + "angle": 0, + "content": "4.3. Baseline Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.535, + 0.47, + 0.595 + ], + "angle": 0, + "content": "Quantitative results. Table 2 gives quantitative results on the CLOTH4D test set using the evaluation metrics described in Section 4.2. As indicated in the non-gray part, we made the following observations:" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.596, + 0.47, + 0.657 + ], + "angle": 0, + "content": "1) In terms of static metrics, \\(\\mathrm{ICON}^{gt} > \\mathrm{PaMIR}^{gt} > \\mathrm{PI}-\\mathrm{FuHD} > \\mathrm{PIF_u} > \\mathrm{PaMIR} > \\mathrm{ICON}\\). I.e., GT-SMPL-guided + pixel-aligned methods > pure pixel-aligned methods > HMR-SMPL-guided + pixel-aligned methods." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.658, + 0.47, + 0.793 + ], + "angle": 0, + "content": "2) With the strong guidance of ground truth SMPL mesh, the performance of \\(\\mathrm{ICON}^{gt}\\) and \\(\\mathrm{PaMIR}^{gt}\\) significantly improves compared to their counterparts with estimated SMPL (i.e., ICON and PaMIR). However, ground truth SMPL meshes are unavailable at test time, which suggests that previous comparisons [15, 49] between GT-SMPL-based methods and others may be unfair. And pure pixel-aligned methods may be even more favorable than SOTA SMPL-based methods for the in-the-wild scenario." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.796, + 0.471, + 0.902 + ], + "angle": 0, + "content": "3) From the perspective of temporally-aware metrics, pure pixel-aligned methods have higher reconstruction stability (see Figure 4). We attribute this to the fact that ICON and PaMIR strongly rely on the SMPL body prior, thus failing to generate far-from-the-body clothes (e.g., skirts and dresses) that present rich temporal dynamics as shown in Figure 1. Plus, the jittery and unstable pose estimation of" + }, + { + "type": "list", + "bbox": [ + 0.076, + 0.596, + 0.471, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.36, + 0.892, + 0.392 + ], + "angle": 0, + "content": "the off-the-shelf HMR methods further prevents ICON and PaMIR from generating temporally coherent results." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.394, + 0.892, + 0.485 + ], + "angle": 0, + "content": "4) ICON outperforms PaMIR as ICON better models mesh-based local features while PaMIR depends more on global information. Moreover, PaMIR loses high-frequency details due to the limited resolution of its volumetric representation. Our observations on CAPE are in line with previous works. For simplicity, we do not expand the narrative." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.487, + 0.892, + 0.547 + ], + "angle": 0, + "content": "Qualitative results. We present the qualitative results on CLOTH4D in Figure 3 and draw the following insights that have not been fully explored since there are no such large-scale and diverse datasets like CLOTH4D." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.551, + 0.892, + 0.793 + ], + "angle": 0, + "content": "1) Global shape vs. local details. All baselines can reconstruct the overall shape conditioned on the input RGB image. PIFuHD presents the finest details, followed by ICON, PaMIR, and PIFu, as PIFuHD enlarges the spatial resolution of pixel-aligned features and ICON takes advantage of mesh-based local features (signed distance, surface normal, etc.). However, focusing on local features suffers from overfitting and poor generalization to complicated clothing (e.g., incomplete dress in the 3rd and 5th examples in Figure 3) and large motions (e.g., artifacts in the arm regions in the 2nd, 3rd, 6th, and 8th examples in Figure 3). Comparatively, thanks to the global feature encoder, PaMIR, and PIFu can generate more holistic clothing but sacrifice details. Thus, it is an important future research direction to explore better strategies for balancing the local and global reconstruction quality." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.796, + 0.892, + 0.901 + ], + "angle": 0, + "content": "2) Human body priors. The last three rows in Figure 3 and Figure 4 (also shows side views) present results on relatively challenging poses. \\(\\mathrm{ICON}^{gt}\\) and ICON robustly recover the poses, while PIFu, PIFuHD, and PaMIR are prone to producing broken limbs or anatomically improbable shapes to different extents. For the side view, we can find that \\(\\mathrm{PaMIR}^{gt}\\) and \\(\\mathrm{ICON}^{gt}\\) are more similar to the" + }, + { + "type": "list", + "bbox": [ + 0.498, + 0.394, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "12851" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.088, + 0.473, + 0.493 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.143, + 0.493, + 0.404, + 0.506 + ], + "angle": 0, + "content": "Figure 3. Qualitative results on CLOTH4D." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.519, + 0.47, + 0.686 + ], + "angle": 0, + "content": "ground truth mesh, which is unsurprising as they are given the ground truth SMPL as prior. Comparatively, PaMIR and ICON, which adopt the estimated SMPL, face the common problems of HMR-SMPL-based reconstruction methods—the predicted SMPL body bends legs or hunches over due to the depth ambiguity—leading to large reconstruction error. Similarly, PIFu and PIFuHD tend to have forward heads with slightly bending legs as they are not aware of any human body priors. One potential research direction is to jointly train or optimize body prior (e.g., SMPL, keypoints, human parsing) with mesh reconstruction." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.689, + 0.471, + 0.903 + ], + "angle": 0, + "content": "3) Ambiguity of geometry and appearance. As shown in the 6th-8th rows of Figure 3, the clothing prints can affect the surface reconstruction due to the ambiguity of geometry and appearance in single-view rendering. Among all these baselines, ICON and PIFuHD show high robustness to the input clothing prints as they predict normal images as intermediate representations, which reduces this ambiguity compared to directly inputting RGB images. Note that ICON only takes the normal images as input to the reconstruction module without using the RGB images, further mitigating the ambiguity and bringing even higher robustness than PIFuHD (the 8th example). Motivated by this observation, future research could shed more light on disentangling the geometry and appearance either implicitly [37] or explicit" + }, + { + "type": "image", + "bbox": [ + 0.507, + 0.087, + 0.891, + 0.392 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.394, + 0.892, + 0.451 + ], + "angle": 0, + "content": "Figure 4. Temporal qualitative results. The 1st, 3rd, and 5th rows are three consecutive frames, and the 2nd, 4th, and 6th rows are the side views predicted from the corresponding front-view RGB. Refer to the supplementary material (Video B) for video results." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.462, + 0.892, + 0.508 + ], + "angle": 0, + "content": "ity [7]. Also, predicting more comprehensive intermediate 2D/3D representations (e.g., depth, illumination, keypoints, segmentation) may also improve the performance." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.509, + 0.893, + 0.902 + ], + "angle": 0, + "content": "4) Temporal consistency. For real-time applications, e.g., streaming from a monocular camera and importing the reconstructed motion sequence into a virtual scene, the temporal coherence of the generated meshes over time is vital for a high-quality user experience. We show the reconstructed meshes for three consecutive frames in Figure 4. As can be found from the side views of the reconstructed meshes, although only subtle motions are present in these frames, the HMR-SMPL-guided methods (PaMIR and ICON) suffer from unstable SMPL predictions and generate temporally inconsistent meshes. On the other hand, pure pixel-aligned methods (PIFu and PIFuHD) fail to predict accurate human pose but produce temporally consistent errors, thus having small \\(ddt\\) values. Given the ground truth SMPL meshes, \\(\\mathrm{PaMIR}^{gt}\\) is still more sensitive to global pose than \\(\\mathrm{ICON}^{gt}\\) as also noted by [49]. These observations are consistent with the quantitative temporal metrics reported in Table 2. It would be interesting to investigate temporal modeling of implicit functions (e.g., incorporating a recurrent neural network [27, 32, 55] to train the implicit function on the 4D dataset, or applying test-time fitting to refine the reconstructed meshes with temporal loss terms). We make such an attempt by adding a temporal term (penalize Chamfer distance between two successive frames) to the refinement process of \\(\\mathrm{PIFu}_{clo}\\) and achieve better temporal consistency (Chamfer_ddt: \\(0.157\\rightarrow 0.123\\) and \\(\\mathrm{Chamfer}_{dtd}\\):" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12852" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.08, + 0.086, + 0.891, + 0.685 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.185, + 0.686, + 0.784, + 0.7 + ], + "angle": 0, + "content": "Figure 5. Qualitative results of baseline methods tested on CLOTH4D with front view and side view." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.703, + 0.182, + 0.717 + ], + "angle": 0, + "content": "\\(0.546\\rightarrow 0.454)\\)" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.727, + 0.288, + 0.74 + ], + "angle": 0, + "content": "4.4. Baseline Enhancement" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.47, + 0.901 + ], + "angle": 0, + "content": "In addition to the findings and insights discovered in Section 4.3 by evaluating on CLOTH4D. CLOTH4D, containing various clothing types and motion sequences, as well as true-to-life multi-layer thin 3D clothing structures, poses new challenges to clothed human reconstruction research. To investigate the SOTA performance when trained on CLOTH4D, we re-train PIFu, PaMIR, and ICON as described in Section 4.1, denoting as \\(\\mathrm{PIFu}_{clo}\\), \\(\\mathrm{PaMIR}_{clo}\\), and \\(\\mathrm{ICON}_{clo}\\), respectively. We report the results of models trained on CLOTH4D in the gray columns in Table 2 and" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.704, + 0.809, + 0.718 + ], + "angle": 0, + "content": "Figure 5. The following findings can be made:" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.719, + 0.892, + 0.824 + ], + "angle": 0, + "content": "1) In terms of quantitative metrics, the models trained on CLOTH4D generally outperform the original models trained on scan datasets as the data distributions of the training and testing datasets are closer. \\(\\mathrm{ICON}_{clo}^{gt}\\) achieves the highest accuracy on the CLOTH4D test set. Furthermore, the qualitative results show that more high-frequency details are generated after training on CLOTH4D." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.825, + 0.892, + 0.901 + ], + "angle": 0, + "content": "2) As the re-implementation setting in ICON's code allows the normal image to be input into all methods, the influence of garment print to PIFu and PaMIR are slightly relieved, validating the hypothesis we made in Section 4.3 that predicting intermediate representations like normals can re" + }, + { + "type": "list", + "bbox": [ + 0.498, + 0.719, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.518, + 0.957 + ], + "angle": 0, + "content": "12853" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.4, + 0.107 + ], + "angle": 0, + "content": "duce the ambiguity of geometry and appearance." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.108, + 0.47, + 0.304 + ], + "angle": 0, + "content": "3) SOTAs fail to model layered and thin clothing structures as shown in the dress and skirt regions in Figure 5, where holes and tattered pieces are generated. Notably, the original PIFu and PaMIR can roughly generate the overall shapes of loose clothing, but they fail when trained on CLOTH4D. Since SOTA methods learn an occupancy field by sampling query points in the 3D space, for PIFu, whose spatial feature resolution is low, and for PaMIR, whose volumetric feature space is also of low resolution, it is hard to sample informative points near the thin surface. It is even harder for the network to learn if a querying point is inside or outside the mesh near the thin structure, as the inside and outside samples have very similar local features." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.305, + 0.471, + 0.456 + ], + "angle": 0, + "content": "4) The difficulties of learning the occupancy field for thin faces motivates developing methods that focus more on the query points near the thin surface for better reconstruction. Future research may also seek better implicit representations to boost the performance of reconstructing multi-layer thin structures e.g., [12, 13]. However, as shown in the supplementary material (Figure B), the state-of-the-art implicit representation cannot achieve satisfactory multi-layer thin structure reconstruction even if we feed the ground truth mesh as the input to the implicit function." + }, + { + "type": "list", + "bbox": [ + 0.076, + 0.108, + 0.471, + 0.456 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.457, + 0.47, + 0.715 + ], + "angle": 0, + "content": "As shown in Table 3 and Figure 6, the performance on the CAPE dataset drops after training on CLOTH4D due to two reasons. 1) The CAPE dataset is collected in a controlled lab environment with dim lighting, which further enlarges the domain gap between CAPE and CLOTH4D. Consequently, the predicted normal images and reconstructed meshes are less accurate. 2) CAPE dataset contains single-layer scan meshes and tight clothing. However, models trained on the multi-layer CLOTH4D dataset tend to generate gaps between the clothing and skin, resulting in holes in the clothing and unsmooth surfaces. This also verifies that local features are sensitive to overfitting. It remains a challenging but interesting problem to have a unified representation of single-layer and multi-layer data, thus different types of datasets could be trained together to yield better generalizability. Finally, we show the reconstructed samples of the in-the-wild images in Figure 7." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.726, + 0.205, + 0.74 + ], + "angle": 0, + "content": "4.5. Limitations" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Firstly, the simulation results are generated via graphics software, which may crash in some cases. We show some detailed examples in the supplementary material (Figure C). Meanwhile, the results on CAPE show that the clothing features of current CLOTH4D can well represent basic men's wear. However, considering the completeness, the dataset scale and diversity of subjects will be further improved to cover males and kids. In addition, the original simulated mesh sequences are non-watertight, which must be converted to watertight (and such conversion is usually lossy)" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.087, + 0.895, + 0.337 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.502, + 0.377, + 0.892, + 0.5 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.5, + 0.5, + 0.892, + 0.528 + ], + "angle": 0, + "content": "Figure 7. In-the-wild front and side view reconstructions. SOTAs with implicit functions tend to generate broken results." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.535, + 0.894, + 0.612 + ], + "angle": 0, + "content": "before feeding into current clothed human reconstruction methods. It would be essential to conduct future research that reconstructs meshes with arbitrary topologies. Finally, as indicated by the results on CAPE, the generalization ability to scan data is yet to be further explored." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.63, + 0.62, + 0.646 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.657, + 0.892, + 0.838 + ], + "angle": 0, + "content": "We introduce CLOTH4D containing realistic and rich categories of clothing, avatars, and animations, and will release it for free, hoping to push the research on clothed human reconstruction. We evaluate current SOTAs with the newly introduced temporally-aware metrics and in-depth analyze their pros and cons by leveraging the advantages of CLOTH4D. We retrain those SOTAs on CLOTH4D, discuss the challenges the new dataset brings, and propose potential research directions. Although layering clothing in CLOTH4D brings immense difficulties to current research, we believe it is an important step toward more realistic and temporally coherent clothed human reconstruction." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.894, + 0.9 + ], + "angle": 0, + "content": "Acknowledgement: This work is supported by Laboratory for Artificial Intelligence in Design (Project Code: RP 3-1) under InnoHK Research Clusters, Hong Kong SAR Government." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "12854" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.175, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.116, + 0.352, + 0.129 + ], + "angle": 0, + "content": "[1] 3dpeople. https://www.3dpeople.com. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.13, + 0.454, + 0.143 + ], + "angle": 0, + "content": "[2] Adobe system incorporated. https://www MIXamo.com. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.145, + 0.406, + 0.158 + ], + "angle": 0, + "content": "[3] Clo virtual fashion llc. https://www.clo3d.com. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.16, + 0.468, + 0.184 + ], + "angle": 0, + "content": "[4] Official code of icon. https://github.com/YuliangXiu/ICON. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.188, + 0.468, + 0.242 + ], + "angle": 0, + "content": "[5] Thiemo Alldieck, Marcus Magnor, Weipeng Xu, Christian Theobalt, and Gerard Pons-Moll. Detailed human avatars from monocular video. In 2018 International Conference on 3D Vision (3DV), pages 98-109. IEEE, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.244, + 0.468, + 0.313 + ], + "angle": 0, + "content": "[6] Thiemo Alldieck, Gerard Pons-Moll, Christian Theobalt, and Marcus Magnor. Tex2shape: Detailed full human body geometry from a single image. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2293-2303, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.315, + 0.468, + 0.382 + ], + "angle": 0, + "content": "[7] Thiemo Alldieck, Mihai Zanfir, and Cristian Sminchisescu. Photorealistic monocular 3d reconstruction of humans wearing clothing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1506-1515, 2022. 3, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.384, + 0.468, + 0.424 + ], + "angle": 0, + "content": "[8] Hugo Bertiche, Meysam Madadi, and Sergio Escalera. Cloth3d: clothed 3d humans. In European Conference on Computer Vision, pages 344-359. Springer, 2020. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.426, + 0.468, + 0.493 + ], + "angle": 0, + "content": "[9] Bharat Lal Bhatnagar, Garvita Tiwari, Christian Theobalt, and Gerard Pons-Moll. Multi-garment net: Learning to dress 3d people from images. In proceedings of the IEEE/CVF international conference on computer vision, pages 5420-5430, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.496, + 0.468, + 0.563 + ], + "angle": 0, + "content": "[10] Zhongang Cai, Daxuan Ren, Ailing Zeng, Zhengyu Lin, Tao Yu, Wenjia Wang, Xiangyu Fan, Yang Gao, Yifan Yu, Liang Pan, et al. Humman: Multi-modal 4d human dataset for versatile sensing and modeling. arXiv preprint arXiv:2204.13686, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.566, + 0.468, + 0.62 + ], + "angle": 0, + "content": "[11] Z. Cao, G. Hidalgo Martinez, T. Simon, S. Wei, and Y. A. Sheikh. Openpose: Realtime multi-person 2d pose estimation using part affinity fields. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2019. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.622, + 0.468, + 0.69 + ], + "angle": 0, + "content": "[12] Weikai Chen, Cheng Lin, Weiyang Li, and Bo Yang. 3psdf: Three-pole signed distance function for learning surfaces with arbitrary topologies. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18522-18531, 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.692, + 0.468, + 0.745 + ], + "angle": 0, + "content": "[13] Julian Chibane, Gerard Pons-Moll, et al. Neural unsigned distance fields for implicit function learning. Advances in Neural Information Processing Systems, 33:21638-21652, 2020. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.748, + 0.468, + 0.789 + ], + "angle": 0, + "content": "[14] Mikhail Erofeev, Yury Gitman, Dmitriy S Vatolin, Alexey Fedorov, and Jue Wang. Perceptually motivated benchmark for video matting. In BMVC, pages 99-1, 2015. 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.79, + 0.468, + 0.844 + ], + "angle": 0, + "content": "[15] Qiao Feng, Yebin Liu, Yu-Kun Lai, Jingyu Yang, and Kun Li. Fof: Learning fourier occupancy field for monocular real-time human reconstruction. arXiv preprint arXiv:2206.02194, 2022. 3, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.846, + 0.468, + 0.899 + ], + "angle": 0, + "content": "[16] Peng Guan, Loretta Reiss, David A Hirshberg, Alexander Weiss, and Michael J Black. Drape: Dressing any person. ACM Transactions on Graphics (ToG), 31(4):1-10, 2012. 2, 3" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.116, + 0.468, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.162 + ], + "angle": 0, + "content": "[17] Erhan Gundogdu, Victor Constantin, Amrollah Seifoddini, Minh Dang, Mathieu Salzmann, and Pascal Fua. Garnet: A two-stream network for fast and accurate 3d cloth draping. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8739-8748, 2019. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.165, + 0.892, + 0.232 + ], + "angle": 0, + "content": "[18] Xintong Han, Xiaojun Hu, Weilin Huang, and Matthew R Scott. Clothflow: A flow-based model for clothed person generation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10471-10480, 2019. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.235, + 0.892, + 0.29 + ], + "angle": 0, + "content": "[19] Xintong Han, Zuxuan Wu, Zhe Wu, Ruichi Yu, and Larry S Davis. Viton: An image-based virtual try-on network. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7543-7552, 2018. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.292, + 0.892, + 0.347 + ], + "angle": 0, + "content": "[20] Tong He, John Collomosse, Hailin Jin, and Stefano Soatto. Geo-pifu: Geometry and pixel aligned implicit functions for single-view human reconstruction. Advances in Neural Information Processing Systems, 33:9276-9287, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.349, + 0.892, + 0.417 + ], + "angle": 0, + "content": "[21] Tong He, Yuanlu Xu, Shunsuke Saito, Stefano Soatto, and Tony Tung. Arch++: Animation-ready clothed human reconstruction revisited. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11046-11056, 2021. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.42, + 0.892, + 0.503 + ], + "angle": 0, + "content": "[22] Zhu Heming, Cao Yu, Jin Hang, Chen Weikai, Du Dong, Wang Zhangye, Cui Shuguang, and Han Xiaoguang. Deep fashion3d: A dataset and benchmark for 3d garment reconstruction from single images. In Computer Vision - ECCV 2020, pages 512-530. Springer International Publishing, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.505, + 0.892, + 0.559 + ], + "angle": 0, + "content": "[23] Jingwei Huang, Yichao Zhou, and Leonidas Guibas. Manifoldplus: A robust and scalable watertight manifold surface generation method for triangle soups. arXiv preprint arXiv:2005.11621, 2020.4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.562, + 0.892, + 0.63 + ], + "angle": 0, + "content": "[24] Zeng Huang, Yuanlu Xu, Christoph Lassner, Hao Li, and Tony Tung. Arch: Animatable reconstruction of clothed humans. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3093-3102, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.633, + 0.892, + 0.702 + ], + "angle": 0, + "content": "[25] Catalin Ionescu, Dragos Papava, Vlad Olaru, and Cristian Sminchisescu. Human3.6m: Large scale datasets and predictive methods for 3d human sensing in natural environments. IEEE transactions on pattern analysis and machine intelligence, 36(7):1325-1339, 2013. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.704, + 0.892, + 0.759 + ], + "angle": 0, + "content": "[26] Boyi Jiang, Juyong Zhang, Yang Hong, Jinhao Luo, Ligang Liu, and Hujun Bao. BCnet: Learning body and cloth shape from a single image. In European Conference on Computer Vision, pages 18-35. Springer, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.761, + 0.892, + 0.829 + ], + "angle": 0, + "content": "[27] Muhammed Kocabas, Nikos Athanasiou, and Michael J Black. Vibe: Video inference for human body pose and shape estimation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5253-5263, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.832, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[28] Nikos Kolotouros, Georgios Pavlakos, Michael J Black, and Kostas Daniilidis. Learning to reconstruct 3d human pose and shape via model-fitting in the loop. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2252-2261, 2019. 3" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12855" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.134 + ], + "angle": 0, + "content": "[29] Nikos Kolotouros, Georgios Pavlakos, and Kostas Dani-ilidis. Convolutional mesh regression for single-image human shape reconstruction. In CVPR, 2019. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.136, + 0.47, + 0.192 + ], + "angle": 0, + "content": "[30] Zorah Lahner, Daniel Cremers, and Tony Tung. Deepwrinkles: Accurate and realistic clothing modeling. In Proceedings of the European conference on computer vision (ECCV), pages 667-684, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.192, + 0.469, + 0.248 + ], + "angle": 0, + "content": "[31] Verica Lazova, Eldar Insafutdinov, and Gerard Pons-Moll. 360-degree textures of people in clothing from a single image. In 2019 International Conference on 3D Vision (3DV), pages 643-653. IEEE, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.25, + 0.469, + 0.318 + ], + "angle": 0, + "content": "[32] Shanchuan Lin, Linjie Yang, Imran Saleemi, and Soumyadip Sengupta. Robust high-resolution video matting with temporal guidance. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 238-247, 2022. 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.321, + 0.469, + 0.375 + ], + "angle": 0, + "content": "[33] Matthew Loper, Naureen Mahmood, Javier Romero, Gerard Pons-Moll, and Michael J Black. SMPL: A skinned multiperson linear model. ACM transactions on graphics (TOG), 34(6):1-16, 2015. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.378, + 0.469, + 0.419 + ], + "angle": 0, + "content": "[34] William E Lorensen and Harvey E Cline. Marching cubes: A high resolution 3d surface construction algorithm. ACM siggraph computer graphics, 21(4):163-169, 1987. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.421, + 0.469, + 0.489 + ], + "angle": 0, + "content": "[35] Qianli Ma, Jinlong Yang, Anurag Ranjan, Sergi Pujades, Gerard Pons-Moll, Siyu Tang, and Michael J Black. Learning to dress 3d people in generative clothing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6469-6478, 2020. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.491, + 0.469, + 0.56 + ], + "angle": 0, + "content": "[36] Meysam Madadi, Hugo Bertiche, Wafa Bouzouita, Isabelle Guyon, and Sergio Escalera. Learning cloth dynamics: 3d + texture garment reconstruction benchmark. In Proceedings of the NeurIPS 2020 Competition and Demonstration Track, PMLR, volume 133, pages 57-76, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.562, + 0.469, + 0.616 + ], + "angle": 0, + "content": "[37] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In ECCV, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.619, + 0.469, + 0.687 + ], + "angle": 0, + "content": "[38] Chaitanya Patel, Zhouyingcheng Liao, and Gerard Pons-Moll. Tailornet: Predicting clothing in 3d as a function of human pose, shape and garment style. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR). IEEE, jun 2020. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.689, + 0.469, + 0.743 + ], + "angle": 0, + "content": "[39] Gerard Pons-Moll, Sergi Pujades, Sonny Hu, and Michael J Black. Clothcap: Seamless 4d clothing capture and retargeting. ACM Transactions on Graphics (ToG), 36(4):1-15, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.746, + 0.469, + 0.802 + ], + "angle": 0, + "content": "[40] Albert Pumarola, Jordi Sanchez, G. Choi, A. Sanfeliu, and F. Moreno-Noguer. 3dpeople: Modeling the geometry of dressed humans. 2019 IEEE/CVF International Conference on Computer Vision (ICCV), pages 2242-2251, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.804, + 0.469, + 0.871 + ], + "angle": 0, + "content": "[41] Shunsuke Saito, Zeng Huang, Ryota Natsume, Shigeo Morishima, Angjoo Kanazawa, and Hao Li. Pifu: Pixel-aligned implicit function for high-resolution clothed human digitization. In Proceedings of the IEEE International Conference on Computer Vision, pages 2304-2314, 2019. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.874, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[42] Shunsuke Saito, Tomas Simon, Jason Saragih, and Hanbyul Joo. Pifuhd: Multi-level pixel-aligned implicit function for" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.893, + 0.135 + ], + "angle": 0, + "content": "high-resolution 3d human digitization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 84-93, 2020. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.136, + 0.893, + 0.192 + ], + "angle": 0, + "content": "[43] Igor Santesteban, Miguel A Otaduy, and Dan Casas. Learning-based animation of clothing for virtual try-on. In Computer Graphics Forum, volume 38, pages 355-366. Wiley Online Library, 2019. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.194, + 0.893, + 0.261 + ], + "angle": 0, + "content": "[44] Leonid Sigal, Alexandru O Balan, and Michael J Black. Humaneva: Synchronized video and motion capture dataset and baseline algorithm for evaluation of articulated human motion. International journal of computer vision, 87(1):4-27, 2010. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.264, + 0.893, + 0.332 + ], + "angle": 0, + "content": "[45] Peter-Pike Sloan, Jan Kautz, and John Snyder. Precomputed radiance transfer for real-time rendering in dynamic, low-frequency lighting environments. In Proceedings of the 29th annual conference on Computer graphics and interactive techniques, pages 527-536, 2002. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.334, + 0.893, + 0.388 + ], + "angle": 0, + "content": "[46] Matthew Trumble, Andrew Gilbert, Charles Malleson, A. Hilton, and J. Collomosse. Total capture: 3d human pose estimation fusing video and inertial sensors. In BMVC, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.392, + 0.893, + 0.445 + ], + "angle": 0, + "content": "[47] Tuanfeng Y Wang, Duygu Ceylan, Jovan Popovic, and Niloy J Mitra. Learning a shared shape space for multimodal garment design. arXiv preprint arXiv:1806.11335, 2018. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.448, + 0.893, + 0.516 + ], + "angle": 0, + "content": "[48] Donglai Xiang, Fabian Prada, Chenglei Wu, and Jessica Hodgins. Monoclothcap: Towards temporally coherent clothing capture from monocular rgb video. In 2020 International Conference on 3D Vision (3DV), pages 322-332. IEEE, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.519, + 0.893, + 0.588 + ], + "angle": 0, + "content": "[49] Yuliang Xiu, Jinlong Yang, Dimitrios Tzionas, and Michael J Black. Icon: Implicit clothed humans obtained from normals. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 13286-13296. IEEE, 2022. 3, 4, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.591, + 0.893, + 0.645 + ], + "angle": 0, + "content": "[50] Shan Yang, Tanya Ambert, Zherong Pan, Ke Wang, Licheng Yu, Tamara Berg, and Ming C Lin. Detailed garment recovery from a single-view image. arXiv preprint arXiv:1608.01250, 2016. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.648, + 0.893, + 0.716 + ], + "angle": 0, + "content": "[51] Tao Yu, Zerong Zheng, Kaiwen Guo, Pengpeng Liu, Qionghai Dai, and Yebin Liu. Function4d: Real-time human volumetric capture from very sparse consumer rgbd sensors. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR2021), June 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.718, + 0.893, + 0.786 + ], + "angle": 0, + "content": "[52] Chao Zhang, Sergi Pujades, Michael J Black, and Gerard Pons-Moll. Detailed, accurate, human shape estimation from clothed 3d scan sequences. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 4191-4200, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.789, + 0.893, + 0.857 + ], + "angle": 0, + "content": "[53] Hongwen Zhang, Yating Tian, Xinchi Zhou, Wanli Ouyang, Yebin Liu, Limin Wang, and Zhenan Sun. Pymaf: 3d human pose and shape regression with pyramidal mesh alignment feedback loop. In Proceedings of the IEEE International Conference on Computer Vision, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.86, + 0.893, + 0.901 + ], + "angle": 0, + "content": "[54] Zerong Zheng, Tao Yu, Yebin Liu, and Qionghai Dai. Pamir: Parametric model-conditioned implicit representation for image-based human reconstruction. IEEE transactions on" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.893, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "12856" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.11, + 0.092, + 0.468, + 0.119 + ], + "angle": 0, + "content": "pattern analysis and machine intelligence, 44(6):3170-3184, 2021. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.122, + 0.469, + 0.177 + ], + "angle": 0, + "content": "[55] Boyao Zhou, Jean-Sébastien Franco, Federica Bogo, and Edmond Boyer. Spatio-temporal human shape completion with implicit function networks. In 2021 International Conference on 3D Vision (3DV), pages 669-678. IEEE, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.179, + 0.469, + 0.248 + ], + "angle": 0, + "content": "[56] Xingxing Zou, Xiangheng Kong, Waikeung Wong, Congde Wang, Yuguang Liu, and Yang Cao. Fashionai: A hierarchical dataset for fashion understanding. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops, pages 0-0, 2019. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.25, + 0.468, + 0.289 + ], + "angle": 0, + "content": "[57] Xingxing Zou and Waikeung Wong. fashion after fashion: A report of ai in fashion. arXiv preprint arXiv:2105.03050, 2021.4" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.469, + 0.289 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.518, + 0.956 + ], + "angle": 0, + "content": "12857" + } + ] +] \ No newline at end of file diff --git a/2023/CLOTH4D_ A Dataset for Clothed Human Reconstruction/b8c1cf88-b3e6-48d2-932c-20d8bacc61eb_origin.pdf b/2023/CLOTH4D_ A Dataset for Clothed Human Reconstruction/b8c1cf88-b3e6-48d2-932c-20d8bacc61eb_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..e6eb6d127affb42b547dbf7b5ffc7384441bed32 --- /dev/null +++ b/2023/CLOTH4D_ A Dataset for Clothed Human Reconstruction/b8c1cf88-b3e6-48d2-932c-20d8bacc61eb_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2de34ec404da44739fae6efcfceaad56e062f9a737accf2010be28e74864fad3 +size 7779898 diff --git a/2023/CLOTH4D_ A Dataset for Clothed Human Reconstruction/full.md b/2023/CLOTH4D_ A Dataset for Clothed Human Reconstruction/full.md new file mode 100644 index 0000000000000000000000000000000000000000..b400ff934275d003cb950c7f19b0d0c253ec7c06 --- /dev/null +++ b/2023/CLOTH4D_ A Dataset for Clothed Human Reconstruction/full.md @@ -0,0 +1,278 @@ +# CLOTH4D: A Dataset for Clothed Human Reconstruction + +Xingxing Zou $^{1,3*}$ , Xintong Han $^{2*}$ , Waikeung Wong $^{3,1\dagger}$ + +$^{1}$ Laboratory for Artificial Intelligence in Design, $^{2}$ Huya Inc. + +$^{3}$ School of Fashion and Textiles, The Hong Kong Polytechnic University + +aemika.zou@connect.polyu.hk,hanxintong@huya.com,calvin.wong@polyu.edu.hk + +![](images/7957263b34aa05b1ffe126ee43aefb27d252d1994001e46367c23d43ae8d3cd7.jpg) +(a) 2D patterns + +![](images/60c3a88e7c0cd270ce27af08f8d5e0b983a4420960d3720027e2a41795918ae7.jpg) +(b) varied poses + +![](images/e196cf514ca353ba391e3c384d7b33444d7ace70293da689ddc40022b5de99ea.jpg) + +![](images/9c2de006c3de326f318b48c04f10f190d6d781d9037556a51f393895c3ff717a.jpg) + +![](images/8bae1e14ed893ffdd96d688f0481315579be8255ca31cb9c9b0730d0c7bab51d.jpg) + +![](images/87003073c460fc8613142f8a555c2d3e56a91a172f2d495f967dcbec51eefd5b.jpg) +(c) multi views + +![](images/909b4dadf1846c304d4f45c96200b0f3f4cd046aca1dc61b6a7536a08c784cf7.jpg) +Figure 1. An instance in CLOTH4D. Given (a) hand-crafted 2D sewing patterns, clothes are simulated with virtual fashion software to have (b) varied poses, (c) multiple views, and (d) high-quality and physically plausible series of paired dynamic 3D meshes with textures. + +![](images/b359570e2358d4a1935537b35a2ff34baa8318f3f03d4f6d962c1fd98a1af424.jpg) + +![](images/251d0b8892bc10cd17cb25bc423f69373897f71b62804ea10ec7e256dde8ae1b.jpg) + +![](images/01b3afc8b98f15453b3e90fc54650f95ee8e2455a796af0535a6d5c2367fac95.jpg) +(d) series of paired dynamic textured meshes + +# Abstract + +Clothed human reconstruction is the cornerstone for creating the virtual world. To a great extent, the quality of recovered avatars decides whether the Metaverse is a passing fad. In this work, we introduce CLOTH4D, a clothed human dataset containing 1,000 subjects with varied appearances, 1,000 3D outfits, and over 100,000 clothed meshes with paired unclothed humans, to fill the gap in large-scale and high-quality 4D clothing data. It enjoys appealing characteristics: 1) Accurate and detailed clothing textured meshes—all clothing items are manually created and then simulated in professional software, strictly following the general standard in fashion design. 2) Separated textured clothing and under-clothing body meshes, closer to the physical world than single-layer raw scans. 3) Clothed human motion sequences simulated given a set of 289 actions, covering fundamental and complicated dynamics. Upon CLOTH4D, we novelly designed a series of + +temporally-aware metrics to evaluate the temporal stability of the generated 3D human meshes, which has been overlooked previously. Moreover, by assessing and retraining current state-of-the-art clothed human reconstruction methods, we reveal insights, present improved performance, and propose potential future research directions, confirming our dataset's advancement. The dataset is available at1. + +# 1. Introduction + +As we enter the volumetric and XR content era, researchers have been trailblazing their way into the Metaverse. With the converging of technologies and practical applications, e.g., fashion NFTs (non-fungible tokens), immersive AR and VR, and games, clothed human reconstruction demands are rapidly growing. While current research has made astonishing results in creating digital hu + +Table 1. Comparisons of CLOTH4D with existing representative datasets. Gray color indicates synthetic datasets generated with graphics engines. #Subjects: number of peoples in different appearances; #Action: number of actions adopted; #Scans: numbers of 3D meshes; 2D Pattern: 2D clothing pattern; TexCloth: with textured clothed model; TexHuman: with textured naked human model. w/ SMPL: with registered SMPL [33] parameters. Public: publicly available and free of charge. Photorealistic: whether the images in the dataset are realistic. -: not applicable or reported. CLOTH4D presents more desirable characteristics compared with others. + +
Dataset#Subjects#Action#Scan2D PatternTexClothTexHumanw/ SMPLPublicPhotorealistic
BUFF [52]6-13.6k-✓-✓✓✓
RenderPeople [1]--825-✓-✓-✓
DeepWrinkles [30]229.2k-✓---✓
CAPE [35]15600140k---✓✓✓
THuman2.0 [51]200-525-✓-✓✓✓
DRAPE [16]72324.5k------
Wang et al. [47]--24k✓✓-✓✓-
3DPeople [40]8072--✓---✓
DCA [43]-567.1k---✓--
GarNet [17]600-18.8k---✓✓-
TailorNet [38]9-5.5k-✓-✓✓-
Cloth3D [8]8.5k7.9k2.1M-✓-✓✓-
Cloth3D++ [36]9.7k8k2.2M✓✓✓✓✓-
CLOTH4D1k289100k✓✓✓✓✓✓
+ +mans, these reconstructed meshes have issues, e.g., flexible body motions and diverse appearances, owing to the lack of datasets with richness in clothing and realistic dynamics of garments. To this end, we introduce CLOTH4D, an open-sourced dataset facilitating physically plausible dynamic clothed human reconstruction. + +Prior to us, many datasets have been collected, and we sort out them in Table 1. Currently, scanned datasets are widely adopted as they are photorealistic and can be easily processed to watertight meshes, which does an excellent favor for current deep models to learn an implicit function (e.g., signed distance function) followed by marching cubes [34] for surface reconstruction. However, it is born with some weaknesses: 1) Scanned meshes are single-layer and inherently fail to capture the space between clothing and skin surface. Thus, body shape under clothing cannot be accurately inferred, let alone the multi-layer and thin clothing structures as in the real physical world. 2) It is time-consuming and expensive to obtain high-quality and large-scale temporal scanned sequences (i.e., 4D scanned sequences) due to the limited efficiency and precision of 4D scanners, especially for complicated clothing and large motions. Although synthetic datasets can to some extent overcome these limitations, existing synthetic datasets are either of small scale in terms of appearances and motions or are highly unrealistic. Moreover, many datasets are not made publicly available and free. + +In contrast, CLOTH4D possesses several attractive attributes: 1) We made great efforts to the diversity and quality of clothing. All clothes are manually designed in CLO [3] and cater to the requirement of the fashion industry. 2) Meshes in CLOTH4D are clothing/humans separated. Such flexibility makes studying and modeling the + +relations and interactions between clothing simulation and body movement possible. 3) CLOTH4D provides plenty of temporal motion sequences with realistic clothing dynamics. As the human body moves, the dressed clothing, e.g., the skirt in Figure 1, naturally deforms. 4) The dataset is large-scale and openly accessible. + +To demonstrate the advantages of CLOTH4D, we use it to evaluate the state-of-the-art (SOTA) clothed human reconstruction methods. In addition to the generally adopted static evaluation metrics, we propose a set of temporally-aware metrics to assess the temporal coherence in a video inference scenario thanks to the rich and true-to-life 4D synthetic sequences in the dataset. Quantitative and qualitative results of SOTA methods on CLOTH4D suggest that our dataset is challenging and the temporal stability of the reconstructed mesh is vital for evaluating the perceptual quality. Meanwhile, we retrain SOTA methods on CLOTH4D, revealing interesting observations of how they perform on multi-layer meshes with thin clothing structures. With in-depth analysis and a summary of challenges for the existing approaches, CLOTH4D makes an essential step toward more realistic reconstructions of clothed humans and stimulates several exciting future work directions. All in all: + +- We contribute CLOTH4D, a large-scale, high-quality, and open-accessible 4D synthetic dataset for clothed human reconstruction. +- We introduce a series of temporally-aware metrics to evaluate the reconstructed performance in the aspect of temporal consistency. +- With the proposed dataset and metrics, we thoroughly analyze the pros and cons of SOTAs, summarize the existing challenges toward more realistic 3D modeling, and propose potential new directions. + +![](images/c77154e2a7d5a2da5c3da834fcbbf1d35ab89da51d61a44605cdb11f7896b8bd.jpg) + +![](images/8f54d30d586003c4a09b136119930a84e4d9af79c112ce884e7c546ad7abffbc.jpg) + +![](images/adb9343b9720e9b1131e722f6e956721e10478ebf2d18970f7b93c5f7ffad7a9.jpg) + +![](images/6ce2fdad0e1c361c30f99d83c9cbe305b7f6a863ed87ae912b47c167ad7ae29f.jpg) +(i) pattern making + +![](images/a5f69fbd618a86af8c964ba33f91c83eb123be469fa771684d47c23b99d7e244.jpg) +(ii) clothed 3D avatar + +![](images/4b01f44e2357bd9ac5ad117d021bf54e1d4a0cd6c1f0f2cfeec5cc14fafca584.jpg) +(iii) clothing simulation + +![](images/5d88420ffa946a9c1f1cbf55aabf38823672789539611c05e6ed8b30d5814714.jpg) + +![](images/37d67aa9ddb6bb7824306184d553b8b21117402a5214bf91ed6d74ec43f4bb8d.jpg) +CLOTH4D with varied 3D sequences +Figure 2. Pipeline for creating instances in CLOTH4D, which primarily adopts CLO for clothing design and simulation, Mixamo for animation, and Blender for processing and exporting meshes. + +# 2. Related Work + +Clothed Human Reconstruction. Clothed human reconstruction aims to recover a 3D mesh from a monocular person image. Estimating dressed people by modeling clothing geometry as 3D displacement on top of a parametric body model (e.g., SMPL [33]) is the leading solution to deal with this task [5,6,28,31,39,48]. By transferring the skinning weights from the body model to the offset clothing mesh, the reconstructed clothed mesh can be readily deformed and animated in the same way as the under-clothing 3D parametric body model. However, it assumes the clothed human to have the same topology as the naked body, leading to unsatisfactory reconstructions of long hair, skirts, dresses, etc. Although methods such as [9,22,26,38] try to isolate the reconstruction of clothing by constructing category-specific clothing templates or statistical models, they fail to generalize to unseen or complex types of clothing. + +As another line of research, deep implicit function networks have drawn broader attention recently [7, 15, 20, 21, 24, 41, 42, 54]. PIFu [41] conditions on pixel-aligned features to build deep implicit functions for reconstructing human meshes, and PIFuHD [42] goes a further step towards enhancing 3D geometric details by predicting front and back normals. ARCH [24] and $\mathrm{ARCH}++$ [21] enable animating the reconstructed meshes by deforming the semantic information into the SMPL canonical space. More recently, leveraging SMPL body models as prior, PaMIR [54] and ICON [49] further improve the reconstruction quality, especially on challenging poses. PHORHUM [7] achieves more accurate results via jointly estimating geometry, albedo, and shading information. It is worth noting that most works rely on training data that are not available free of charge, making an open-sourced dataset of clothed avatars of great significance to advance research in 3D body/clothing modeling. + +Clothed Human Datasets. As summarized in Table 1, existing clothed human datasets can be divided into two + +types, i.e., scanned datasets and synthetic datasets. The former [10, 25, 44, 46] utilizes multiple synchronized cameras to capture motions that have difficulties enriching the scalability and diversity and obtaining highly accurate 4D ground truth. Moreover, it inherently cannot mimic the layering of clothing. The synthetic datasets mitigate these limitations. However, existing synthetic datasets, no matter whether static [17, 38, 47] or dynamic [8, 16, 39, 43], either only contain a few clothing types or are highly unrealistic. Cloth3D++ [36] is developed from Cloth3D [8], which contains a total of 2.2 million scans, covering 9.7k subjects dressed in 12.9k clothing, which is the most advanced in scale. However, the clothing created based on garment templates are only base patterns with an immense gap compared to clothing in real life. To intuitively demonstrate the advantages of CLOTH4D, we put visual comparisons of these datasets in Figure A in the supplementary material. + +# 3. CLOTH4D Dataset + +We depict the pipeline of creating CLOTH4D in Figure 2, including (1) preparing the referenced clothing images and the unclothed 3D human avatars; (2) uploading 3D human avatars to Mixamo to obtain the FBX files with various animations; (3) designing 3D clothing in CLO, integrating the FBX files obtained in (2), and conducting clothing simulation; and (4) exporting the sequenced mesh files. + +Clothes and Models. Clothes are manually created by professional fashion designers using CLO by producing the 2D garment patterns and then auto-simulating them to 3D. We put a video illustrating the production process of a 3D garment in the supplementary material (Video A) instead of demonstrating details in the paper. CLOTH4D covers 1,000 different 3D outfits spansning over 500 prints and 50 fabrics. Over $40\%$ of the clothes are designed referring to the newest collections of varying design houses (e.g., Prada, Moschino, and Alexander McQueen) to ensure visual re + +alism, variance, and fashionability. In this paper, we focus on women's wear owing to its diversity covering most characteristics of garments. For avatars, CLO provides human avatars with varied physical appearances in terms of hairstyles, faces, skin colors, body figures, etc., and we directly adopt these available avatars. + +Animations and Simulations. For each human avatar, we use Mixamo [2] for rigging and generating motion sequences. Given a motion sequence and a 3D garment, CLO runs a clothing simulation of the 3D garment according to the motion sequence at 30fps. A total of 289 animations are utilized, such as Billy Dance, Offensive Idle, Jumping Rope, etc. Unlike previous work [8] directly uses Blender for cloth simulation, CLO simulates clothing with richer details and dynamics. The clothes exhibit wrinkles and fold with the body's movement in a natural and physically plausible way, especially for skirts or dresses that differ a lot from the body topology as shown in Figure 1. + +Paired Multi-Modal Data. Then, we can obtain paired data in the following representations: a 2D clothing pattern, 3D mesh sequences (clothed and naked human meshes with corresponding fitted SMPL parameters/meshes, and separate clothing meshes), and a UV texture map. The textured mesh can be rendered into multi-view normal images, depth images, and RGB images given varying light conditions. We also translate all these dynamic meshes to watertight with simplification using [23], thus they can be readily used to train an implicit function for mesh reconstruction. The number of triangles ranges from $170K$ to $4M$ for simulation and becomes $200K$ after simplification. Besides human reconstruction, many other tasks could also benefit from this paired data (e.g., clothing capture [50], human pose transfer [18], and fashion-related tasks [19, 56, 57]). + +# 4. Evaluations + +In this section, we evaluate the state-of-the-art clothed human reconstruction methods on CLOTH4D to demonstrate new insights that the dataset can provide. Further, we retrain SOTA methods on CLOTH4D and make several interesting observations of how they perform on multi-layer meshes with thin clothing structures. We also present the challenges for existing approaches and propose potential research directions with a comprehensive analysis. + +# 4.1. Baselines + +We mainly report results of four SOTA approaches, PIFu [41], PIFuHD [42], PaMIR [54], and ICON [49] owing to other works do not release their codes or models, such as PHORHUM [7], ARCH++ [21], or which have already been extensively compared with the listed methods above. We use PIFu, PIFuHD, PaMIR, and ICON to denote their released pretrained testing models. $\mathrm{PaMIR}^{gt}$ and $\mathrm{ICON}^{gt}$ indicate the testing results using the fitted ground truth + +SMPL mesh as conditions rather than the one predicted using off-the-shelf human mesh recovery (HMR) methods (GraphCMR [29] for PaMIR, and PyMAf [53] for ICON as in their released code). Based on the characteristics of these methods, they can be divided into three types: 1) pixel-aligned methods (PIFu, PIFuHD); 2) GT-SMPL-guided + pixel-aligned methods $(\mathrm{PaMIR}^{gt}$ and $\mathrm{ICON}^{gt})$ ; 3) HMR-SMPL-guided + pixel-aligned methods (PaMIR and ICON). + +Furthermore, we retrain PIFu, PaMIR, and ICON on CLOTH4D, which are denoted as $\mathrm{PIF_u_{clo}}$ , $\mathrm{PaMIR}_{clo}$ , and $\mathrm{ICON}_{clo}$ , respectively. For these retrained models, we follow the re-implementation setting introduced in ICON [4], which allows us to train all these baselines with the same training protocol and hyper-parameters for a fair comparison. Similarly, $\mathrm{PaMIR}_{clo}^{gt}$ , and $\mathrm{ICON}_{clo}^{gt}$ are tested with the ground truth SMPL fits. We also use the cloth-refinement module in the ICON's released code for post-processing. + +# 4.2. Datasets and Metrics + +Datasets and implementation details. We organize the sequences in CLOTH4D into a $80\% / 10\% / 10\%$ train/val/test split. We render each mesh into 8 views using a weak perspective camera and pre-computed radiance transfer [45] with dynamic light conditions following [41, 49]. All rendered images are $512 \times 512$ . The 2D keypoints used in all methods are generated by OpenPose [11]. In addition, we also evaluate all models on CAPE [35] test set adopted in [49] to investigate the generalization ability. + +Static Metrics. We report the quantitative results on normal reprojection error, Chamfer distance, and P2S distance for evaluation as [15, 41, 42, 49]. As all compared methods use a weak perspective or orthographic camera, the estimated meshes may not be well aligned with the ground truth meshes in the $z$ -direction (i.e., view direction). Thus, we shift the estimated meshes to have the same $z$ -axis mean as the ground truths following [15] for a fair comparison. + +Temporal Metrics. The aforementioned static metrics ignore the temporal consistency of the reconstructed meshes across time, which is essential for real-time applications since meshes presenting jitters and flickers highly affect the perceptual quality. Thanks to the rich temporal dynamics provided in CLOTH4D, we are the first to introduce temporally-aware metrics to evaluate the temporal coherence of the generated mesh sequences. Referring to temporal metrics SSDdt and dtSSD used in video matting tasks [14, 32], we compute two metrics measuring the temporal coherence of the predicted mesh normal: + +$$ +\text {N o r m a l s} _ {d d t} = \frac {1}{T} \sum_ {t} \left| \left(\mathcal {N} _ {t} ^ {p r} - \mathcal {N} _ {t} ^ {g t}\right) ^ {2} - \left(\mathcal {N} _ {t + 1} ^ {p r} - \mathcal {N} _ {t + 1} ^ {g t}\right) ^ {2} \right|, \tag {1} +$$ + +$$ +\operatorname {N o r m a l s} _ {d t d} = \frac {1}{T} \sum_ {t} \left| \left(\mathcal {N} _ {t} ^ {p r} - \mathcal {N} _ {t + 1} ^ {p r}\right) ^ {2} - \left(\mathcal {N} _ {t} ^ {g t} - \mathcal {N} _ {t + 1} ^ {g t}\right) ^ {2} \right|, \tag {2} +$$ + +where $T$ is the length of the sequence. $\mathcal{N}_t^{pr}$ and $\mathcal{N}_t^{gt}$ denote the rendered normal images from the predicted mesh + +Table 2. Quantitative evaluation on CLOTH4D. PaMIRgt and ICONgt denote that the fitted ground truth SMPL is used during the inference. Gray color indicates the results trained on CLOTH4D. + +
MethodPIFuPIFuHDPaMIR\( PaMIR^{gt} \)ICON\( ICON^{gt} \)\( PIFu_{clo} \)\( PaMIR_{clo} \)\( PaMIR^{gt}_{clo} \)\( ICON_{clo} \)\( ICON^{gt}_{clo} \)
Normals \( \downarrow \)0.1820.1810.2240.1450.2110.1180.1500.2300.1140.1980.103
P2S \( \downarrow \)3.9113.5183.8272.6414.6602.4732.7935.0372.6193.7112.068
Chamfer \( \downarrow \)3.5782.4874.1572.4874.1961.6312.4124.1861.6183.4991.367
\( Normals_{ddt} \) \( \downarrow \)0.0080.0070.0320.0250.0130.0300.0090.0210.0330.0130.035
\( Normals_{dtd} \) \( \downarrow \)0.0250.0230.0430.0350.0340.0280.0170.0400.0270.0380.033
\( P2S_{ddt} \) \( \downarrow \)0.2220.1670.2530.2470.3830.3690.1850.4170.4020.3210.367
\( P2S_{dtd} \) \( \downarrow \)0.7330.5781.0250.7020.9790.5540.5480.8900.5360.8780.526
Chamfer \( _{ddt} \) \( \downarrow \)0.2070.1810.3310.3160.3590.3580.1570.3500.3820.3310.369
Chamfer \( _{dtd} \) \( \downarrow \)0.7290.5761.0170.7010.9740.5530.5460.8860.5360.8730.525
+ +Table 3. Quantitative evaluation on CAPE. Table notations are the same as Table 2. + +
MethodPIFuPIFuHDPaMIRPaMIRgtICONICONgtPIFucloPaMIRcloPaMIRgtcloICONcloICONgtclo
Normals ↓0.1610.1600.1830.0860.1600.0560.1640.1760.0930.1560.077
P2S ↓4.2593.7953.8401.1934.0141.0674.6524.2351.4913.3041.193
Chamfer ↓4.2043.9274.2581.6543.9621.0384.3814.0801.4273.5441.397
+ +and the ground truth mesh at time step $t$ , respectively. The subscript $ddt$ is short for distance delta time, which captures the stability of errors between two consecutive meshes. And $dtd$ (delta time distance) penalizes large temporal change of the prediction with respect to the change of the ground truth. These two metrics indicate unstable mesh variations and ignore temporally coherent errors [14]. The $ddt$ and $dtd$ of Chamfer and P2S distances are similarly defined. More details can be found in the supplementary material. + +# 4.3. Baseline Evaluation + +Quantitative results. Table 2 gives quantitative results on the CLOTH4D test set using the evaluation metrics described in Section 4.2. As indicated in the non-gray part, we made the following observations: + +1) In terms of static metrics, $\mathrm{ICON}^{gt} > \mathrm{PaMIR}^{gt} > \mathrm{PI}-\mathrm{FuHD} > \mathrm{PIF_u} > \mathrm{PaMIR} > \mathrm{ICON}$ . I.e., GT-SMPL-guided + pixel-aligned methods > pure pixel-aligned methods > HMR-SMPL-guided + pixel-aligned methods. +2) With the strong guidance of ground truth SMPL mesh, the performance of $\mathrm{ICON}^{gt}$ and $\mathrm{PaMIR}^{gt}$ significantly improves compared to their counterparts with estimated SMPL (i.e., ICON and PaMIR). However, ground truth SMPL meshes are unavailable at test time, which suggests that previous comparisons [15, 49] between GT-SMPL-based methods and others may be unfair. And pure pixel-aligned methods may be even more favorable than SOTA SMPL-based methods for the in-the-wild scenario. +3) From the perspective of temporally-aware metrics, pure pixel-aligned methods have higher reconstruction stability (see Figure 4). We attribute this to the fact that ICON and PaMIR strongly rely on the SMPL body prior, thus failing to generate far-from-the-body clothes (e.g., skirts and dresses) that present rich temporal dynamics as shown in Figure 1. Plus, the jittery and unstable pose estimation of + +the off-the-shelf HMR methods further prevents ICON and PaMIR from generating temporally coherent results. + +4) ICON outperforms PaMIR as ICON better models mesh-based local features while PaMIR depends more on global information. Moreover, PaMIR loses high-frequency details due to the limited resolution of its volumetric representation. Our observations on CAPE are in line with previous works. For simplicity, we do not expand the narrative. +Qualitative results. We present the qualitative results on CLOTH4D in Figure 3 and draw the following insights that have not been fully explored since there are no such large-scale and diverse datasets like CLOTH4D. +1) Global shape vs. local details. All baselines can reconstruct the overall shape conditioned on the input RGB image. PIFuHD presents the finest details, followed by ICON, PaMIR, and PIFu, as PIFuHD enlarges the spatial resolution of pixel-aligned features and ICON takes advantage of mesh-based local features (signed distance, surface normal, etc.). However, focusing on local features suffers from overfitting and poor generalization to complicated clothing (e.g., incomplete dress in the 3rd and 5th examples in Figure 3) and large motions (e.g., artifacts in the arm regions in the 2nd, 3rd, 6th, and 8th examples in Figure 3). Comparatively, thanks to the global feature encoder, PaMIR, and PIFu can generate more holistic clothing but sacrifice details. Thus, it is an important future research direction to explore better strategies for balancing the local and global reconstruction quality. +2) Human body priors. The last three rows in Figure 3 and Figure 4 (also shows side views) present results on relatively challenging poses. $\mathrm{ICON}^{gt}$ and ICON robustly recover the poses, while PIFu, PIFuHD, and PaMIR are prone to producing broken limbs or anatomically improbable shapes to different extents. For the side view, we can find that $\mathrm{PaMIR}^{gt}$ and $\mathrm{ICON}^{gt}$ are more similar to the + +![](images/e19d45237e7f35dc621f920218b042fef81b42e57a3a7fe951d6051da6e588d2.jpg) +Figure 3. Qualitative results on CLOTH4D. + +ground truth mesh, which is unsurprising as they are given the ground truth SMPL as prior. Comparatively, PaMIR and ICON, which adopt the estimated SMPL, face the common problems of HMR-SMPL-based reconstruction methods—the predicted SMPL body bends legs or hunches over due to the depth ambiguity—leading to large reconstruction error. Similarly, PIFu and PIFuHD tend to have forward heads with slightly bending legs as they are not aware of any human body priors. One potential research direction is to jointly train or optimize body prior (e.g., SMPL, keypoints, human parsing) with mesh reconstruction. + +3) Ambiguity of geometry and appearance. As shown in the 6th-8th rows of Figure 3, the clothing prints can affect the surface reconstruction due to the ambiguity of geometry and appearance in single-view rendering. Among all these baselines, ICON and PIFuHD show high robustness to the input clothing prints as they predict normal images as intermediate representations, which reduces this ambiguity compared to directly inputting RGB images. Note that ICON only takes the normal images as input to the reconstruction module without using the RGB images, further mitigating the ambiguity and bringing even higher robustness than PIFuHD (the 8th example). Motivated by this observation, future research could shed more light on disentangling the geometry and appearance either implicitly [37] or explicit + +![](images/6e62e8b77b9da05f07ed72c7370e6a471563bba9b6837cbffd73fd5f8cebed2c.jpg) +Figure 4. Temporal qualitative results. The 1st, 3rd, and 5th rows are three consecutive frames, and the 2nd, 4th, and 6th rows are the side views predicted from the corresponding front-view RGB. Refer to the supplementary material (Video B) for video results. + +ity [7]. Also, predicting more comprehensive intermediate 2D/3D representations (e.g., depth, illumination, keypoints, segmentation) may also improve the performance. + +4) Temporal consistency. For real-time applications, e.g., streaming from a monocular camera and importing the reconstructed motion sequence into a virtual scene, the temporal coherence of the generated meshes over time is vital for a high-quality user experience. We show the reconstructed meshes for three consecutive frames in Figure 4. As can be found from the side views of the reconstructed meshes, although only subtle motions are present in these frames, the HMR-SMPL-guided methods (PaMIR and ICON) suffer from unstable SMPL predictions and generate temporally inconsistent meshes. On the other hand, pure pixel-aligned methods (PIFu and PIFuHD) fail to predict accurate human pose but produce temporally consistent errors, thus having small $ddt$ values. Given the ground truth SMPL meshes, $\mathrm{PaMIR}^{gt}$ is still more sensitive to global pose than $\mathrm{ICON}^{gt}$ as also noted by [49]. These observations are consistent with the quantitative temporal metrics reported in Table 2. It would be interesting to investigate temporal modeling of implicit functions (e.g., incorporating a recurrent neural network [27, 32, 55] to train the implicit function on the 4D dataset, or applying test-time fitting to refine the reconstructed meshes with temporal loss terms). We make such an attempt by adding a temporal term (penalize Chamfer distance between two successive frames) to the refinement process of $\mathrm{PIFu}_{clo}$ and achieve better temporal consistency (Chamfer_ddt: $0.157\rightarrow 0.123$ and $\mathrm{Chamfer}_{dtd}$ : + +![](images/cc8d9eb1464f0be0cb588a3b7b5f238d88672be921d44dac08256fe4abed7336.jpg) +Figure 5. Qualitative results of baseline methods tested on CLOTH4D with front view and side view. + +$0.546\rightarrow 0.454)$ + +# 4.4. Baseline Enhancement + +In addition to the findings and insights discovered in Section 4.3 by evaluating on CLOTH4D. CLOTH4D, containing various clothing types and motion sequences, as well as true-to-life multi-layer thin 3D clothing structures, poses new challenges to clothed human reconstruction research. To investigate the SOTA performance when trained on CLOTH4D, we re-train PIFu, PaMIR, and ICON as described in Section 4.1, denoting as $\mathrm{PIFu}_{clo}$ , $\mathrm{PaMIR}_{clo}$ , and $\mathrm{ICON}_{clo}$ , respectively. We report the results of models trained on CLOTH4D in the gray columns in Table 2 and + +Figure 5. The following findings can be made: + +1) In terms of quantitative metrics, the models trained on CLOTH4D generally outperform the original models trained on scan datasets as the data distributions of the training and testing datasets are closer. $\mathrm{ICON}_{clo}^{gt}$ achieves the highest accuracy on the CLOTH4D test set. Furthermore, the qualitative results show that more high-frequency details are generated after training on CLOTH4D. +2) As the re-implementation setting in ICON's code allows the normal image to be input into all methods, the influence of garment print to PIFu and PaMIR are slightly relieved, validating the hypothesis we made in Section 4.3 that predicting intermediate representations like normals can re + +duce the ambiguity of geometry and appearance. + +3) SOTAs fail to model layered and thin clothing structures as shown in the dress and skirt regions in Figure 5, where holes and tattered pieces are generated. Notably, the original PIFu and PaMIR can roughly generate the overall shapes of loose clothing, but they fail when trained on CLOTH4D. Since SOTA methods learn an occupancy field by sampling query points in the 3D space, for PIFu, whose spatial feature resolution is low, and for PaMIR, whose volumetric feature space is also of low resolution, it is hard to sample informative points near the thin surface. It is even harder for the network to learn if a querying point is inside or outside the mesh near the thin structure, as the inside and outside samples have very similar local features. +4) The difficulties of learning the occupancy field for thin faces motivates developing methods that focus more on the query points near the thin surface for better reconstruction. Future research may also seek better implicit representations to boost the performance of reconstructing multi-layer thin structures e.g., [12, 13]. However, as shown in the supplementary material (Figure B), the state-of-the-art implicit representation cannot achieve satisfactory multi-layer thin structure reconstruction even if we feed the ground truth mesh as the input to the implicit function. + +As shown in Table 3 and Figure 6, the performance on the CAPE dataset drops after training on CLOTH4D due to two reasons. 1) The CAPE dataset is collected in a controlled lab environment with dim lighting, which further enlarges the domain gap between CAPE and CLOTH4D. Consequently, the predicted normal images and reconstructed meshes are less accurate. 2) CAPE dataset contains single-layer scan meshes and tight clothing. However, models trained on the multi-layer CLOTH4D dataset tend to generate gaps between the clothing and skin, resulting in holes in the clothing and unsmooth surfaces. This also verifies that local features are sensitive to overfitting. It remains a challenging but interesting problem to have a unified representation of single-layer and multi-layer data, thus different types of datasets could be trained together to yield better generalizability. Finally, we show the reconstructed samples of the in-the-wild images in Figure 7. + +# 4.5. Limitations + +Firstly, the simulation results are generated via graphics software, which may crash in some cases. We show some detailed examples in the supplementary material (Figure C). Meanwhile, the results on CAPE show that the clothing features of current CLOTH4D can well represent basic men's wear. However, considering the completeness, the dataset scale and diversity of subjects will be further improved to cover males and kids. In addition, the original simulated mesh sequences are non-watertight, which must be converted to watertight (and such conversion is usually lossy) + +![](images/935dbe458a4ad764c608e737541a3c291f4300f0bca0ca5d20cc9762cedf07be.jpg) + +![](images/74a8c03af2716bb742b5369fb13107d2b3e7ae6e416f976470e6eb90d8657c01.jpg) +Figure 7. In-the-wild front and side view reconstructions. SOTAs with implicit functions tend to generate broken results. + +before feeding into current clothed human reconstruction methods. It would be essential to conduct future research that reconstructs meshes with arbitrary topologies. Finally, as indicated by the results on CAPE, the generalization ability to scan data is yet to be further explored. + +# 5. Conclusion + +We introduce CLOTH4D containing realistic and rich categories of clothing, avatars, and animations, and will release it for free, hoping to push the research on clothed human reconstruction. We evaluate current SOTAs with the newly introduced temporally-aware metrics and in-depth analyze their pros and cons by leveraging the advantages of CLOTH4D. We retrain those SOTAs on CLOTH4D, discuss the challenges the new dataset brings, and propose potential research directions. Although layering clothing in CLOTH4D brings immense difficulties to current research, we believe it is an important step toward more realistic and temporally coherent clothed human reconstruction. + +Acknowledgement: This work is supported by Laboratory for Artificial Intelligence in Design (Project Code: RP 3-1) under InnoHK Research Clusters, Hong Kong SAR Government. + +# References + +[1] 3dpeople. https://www.3dpeople.com. 2 +[2] Adobe system incorporated. https://www MIXamo.com. 4 +[3] Clo virtual fashion llc. https://www.clo3d.com. 2 +[4] Official code of icon. https://github.com/YuliangXiu/ICON. 4 +[5] Thiemo Alldieck, Marcus Magnor, Weipeng Xu, Christian Theobalt, and Gerard Pons-Moll. Detailed human avatars from monocular video. In 2018 International Conference on 3D Vision (3DV), pages 98-109. IEEE, 2018. 3 +[6] Thiemo Alldieck, Gerard Pons-Moll, Christian Theobalt, and Marcus Magnor. Tex2shape: Detailed full human body geometry from a single image. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2293-2303, 2019. 3 +[7] Thiemo Alldieck, Mihai Zanfir, and Cristian Sminchisescu. Photorealistic monocular 3d reconstruction of humans wearing clothing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1506-1515, 2022. 3, 4, 6 +[8] Hugo Bertiche, Meysam Madadi, and Sergio Escalera. Cloth3d: clothed 3d humans. In European Conference on Computer Vision, pages 344-359. Springer, 2020. 2, 3, 4 +[9] Bharat Lal Bhatnagar, Garvita Tiwari, Christian Theobalt, and Gerard Pons-Moll. Multi-garment net: Learning to dress 3d people from images. In proceedings of the IEEE/CVF international conference on computer vision, pages 5420-5430, 2019. 3 +[10] Zhongang Cai, Daxuan Ren, Ailing Zeng, Zhengyu Lin, Tao Yu, Wenjia Wang, Xiangyu Fan, Yang Gao, Yifan Yu, Liang Pan, et al. Humman: Multi-modal 4d human dataset for versatile sensing and modeling. arXiv preprint arXiv:2204.13686, 2022. 3 +[11] Z. Cao, G. Hidalgo Martinez, T. Simon, S. Wei, and Y. A. Sheikh. Openpose: Realtime multi-person 2d pose estimation using part affinity fields. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2019. 4 +[12] Weikai Chen, Cheng Lin, Weiyang Li, and Bo Yang. 3psdf: Three-pole signed distance function for learning surfaces with arbitrary topologies. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18522-18531, 2022. 8 +[13] Julian Chibane, Gerard Pons-Moll, et al. Neural unsigned distance fields for implicit function learning. Advances in Neural Information Processing Systems, 33:21638-21652, 2020. 8 +[14] Mikhail Erofeev, Yury Gitman, Dmitriy S Vatolin, Alexey Fedorov, and Jue Wang. Perceptually motivated benchmark for video matting. In BMVC, pages 99-1, 2015. 4, 5 +[15] Qiao Feng, Yebin Liu, Yu-Kun Lai, Jingyu Yang, and Kun Li. Fof: Learning fourier occupancy field for monocular real-time human reconstruction. arXiv preprint arXiv:2206.02194, 2022. 3, 4, 5 +[16] Peng Guan, Loretta Reiss, David A Hirshberg, Alexander Weiss, and Michael J Black. Drape: Dressing any person. ACM Transactions on Graphics (ToG), 31(4):1-10, 2012. 2, 3 + +[17] Erhan Gundogdu, Victor Constantin, Amrollah Seifoddini, Minh Dang, Mathieu Salzmann, and Pascal Fua. Garnet: A two-stream network for fast and accurate 3d cloth draping. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8739-8748, 2019. 2, 3 +[18] Xintong Han, Xiaojun Hu, Weilin Huang, and Matthew R Scott. Clothflow: A flow-based model for clothed person generation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10471-10480, 2019. 4 +[19] Xintong Han, Zuxuan Wu, Zhe Wu, Ruichi Yu, and Larry S Davis. Viton: An image-based virtual try-on network. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7543-7552, 2018. 4 +[20] Tong He, John Collomosse, Hailin Jin, and Stefano Soatto. Geo-pifu: Geometry and pixel aligned implicit functions for single-view human reconstruction. Advances in Neural Information Processing Systems, 33:9276-9287, 2020. 3 +[21] Tong He, Yuanlu Xu, Shunsuke Saito, Stefano Soatto, and Tony Tung. Arch++: Animation-ready clothed human reconstruction revisited. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11046-11056, 2021. 3, 4 +[22] Zhu Heming, Cao Yu, Jin Hang, Chen Weikai, Du Dong, Wang Zhangye, Cui Shuguang, and Han Xiaoguang. Deep fashion3d: A dataset and benchmark for 3d garment reconstruction from single images. In Computer Vision - ECCV 2020, pages 512-530. Springer International Publishing, 2020. 3 +[23] Jingwei Huang, Yichao Zhou, and Leonidas Guibas. Manifoldplus: A robust and scalable watertight manifold surface generation method for triangle soups. arXiv preprint arXiv:2005.11621, 2020.4 +[24] Zeng Huang, Yuanlu Xu, Christoph Lassner, Hao Li, and Tony Tung. Arch: Animatable reconstruction of clothed humans. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3093-3102, 2020. 3 +[25] Catalin Ionescu, Dragos Papava, Vlad Olaru, and Cristian Sminchisescu. Human3.6m: Large scale datasets and predictive methods for 3d human sensing in natural environments. IEEE transactions on pattern analysis and machine intelligence, 36(7):1325-1339, 2013. 3 +[26] Boyi Jiang, Juyong Zhang, Yang Hong, Jinhao Luo, Ligang Liu, and Hujun Bao. BCnet: Learning body and cloth shape from a single image. In European Conference on Computer Vision, pages 18-35. Springer, 2020. 3 +[27] Muhammed Kocabas, Nikos Athanasiou, and Michael J Black. Vibe: Video inference for human body pose and shape estimation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5253-5263, 2020. 6 +[28] Nikos Kolotouros, Georgios Pavlakos, Michael J Black, and Kostas Daniilidis. Learning to reconstruct 3d human pose and shape via model-fitting in the loop. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2252-2261, 2019. 3 + +[29] Nikos Kolotouros, Georgios Pavlakos, and Kostas Dani-ilidis. Convolutional mesh regression for single-image human shape reconstruction. In CVPR, 2019. 4 +[30] Zorah Lahner, Daniel Cremers, and Tony Tung. Deepwrinkles: Accurate and realistic clothing modeling. In Proceedings of the European conference on computer vision (ECCV), pages 667-684, 2018. 2 +[31] Verica Lazova, Eldar Insafutdinov, and Gerard Pons-Moll. 360-degree textures of people in clothing from a single image. In 2019 International Conference on 3D Vision (3DV), pages 643-653. IEEE, 2019. 3 +[32] Shanchuan Lin, Linjie Yang, Imran Saleemi, and Soumyadip Sengupta. Robust high-resolution video matting with temporal guidance. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 238-247, 2022. 4, 6 +[33] Matthew Loper, Naureen Mahmood, Javier Romero, Gerard Pons-Moll, and Michael J Black. SMPL: A skinned multiperson linear model. ACM transactions on graphics (TOG), 34(6):1-16, 2015. 2, 3 +[34] William E Lorensen and Harvey E Cline. Marching cubes: A high resolution 3d surface construction algorithm. ACM siggraph computer graphics, 21(4):163-169, 1987. 2 +[35] Qianli Ma, Jinlong Yang, Anurag Ranjan, Sergi Pujades, Gerard Pons-Moll, Siyu Tang, and Michael J Black. Learning to dress 3d people in generative clothing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6469-6478, 2020. 2, 4 +[36] Meysam Madadi, Hugo Bertiche, Wafa Bouzouita, Isabelle Guyon, and Sergio Escalera. Learning cloth dynamics: 3d + texture garment reconstruction benchmark. In Proceedings of the NeurIPS 2020 Competition and Demonstration Track, PMLR, volume 133, pages 57-76, 2021. 2, 3 +[37] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In ECCV, 2020. 6 +[38] Chaitanya Patel, Zhouyingcheng Liao, and Gerard Pons-Moll. Tailornet: Predicting clothing in 3d as a function of human pose, shape and garment style. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR). IEEE, jun 2020. 2, 3 +[39] Gerard Pons-Moll, Sergi Pujades, Sonny Hu, and Michael J Black. Clothcap: Seamless 4d clothing capture and retargeting. ACM Transactions on Graphics (ToG), 36(4):1-15, 2017. 3 +[40] Albert Pumarola, Jordi Sanchez, G. Choi, A. Sanfeliu, and F. Moreno-Noguer. 3dpeople: Modeling the geometry of dressed humans. 2019 IEEE/CVF International Conference on Computer Vision (ICCV), pages 2242-2251, 2019. 2 +[41] Shunsuke Saito, Zeng Huang, Ryota Natsume, Shigeo Morishima, Angjoo Kanazawa, and Hao Li. Pifu: Pixel-aligned implicit function for high-resolution clothed human digitization. In Proceedings of the IEEE International Conference on Computer Vision, pages 2304-2314, 2019. 3, 4 +[42] Shunsuke Saito, Tomas Simon, Jason Saragih, and Hanbyul Joo. Pifuhd: Multi-level pixel-aligned implicit function for + +high-resolution 3d human digitization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 84-93, 2020. 3, 4 +[43] Igor Santesteban, Miguel A Otaduy, and Dan Casas. Learning-based animation of clothing for virtual try-on. In Computer Graphics Forum, volume 38, pages 355-366. Wiley Online Library, 2019. 2, 3 +[44] Leonid Sigal, Alexandru O Balan, and Michael J Black. Humaneva: Synchronized video and motion capture dataset and baseline algorithm for evaluation of articulated human motion. International journal of computer vision, 87(1):4-27, 2010. 3 +[45] Peter-Pike Sloan, Jan Kautz, and John Snyder. Precomputed radiance transfer for real-time rendering in dynamic, low-frequency lighting environments. In Proceedings of the 29th annual conference on Computer graphics and interactive techniques, pages 527-536, 2002. 4 +[46] Matthew Trumble, Andrew Gilbert, Charles Malleson, A. Hilton, and J. Collomosse. Total capture: 3d human pose estimation fusing video and inertial sensors. In BMVC, 2017. 3 +[47] Tuanfeng Y Wang, Duygu Ceylan, Jovan Popovic, and Niloy J Mitra. Learning a shared shape space for multimodal garment design. arXiv preprint arXiv:1806.11335, 2018. 2, 3 +[48] Donglai Xiang, Fabian Prada, Chenglei Wu, and Jessica Hodgins. Monoclothcap: Towards temporally coherent clothing capture from monocular rgb video. In 2020 International Conference on 3D Vision (3DV), pages 322-332. IEEE, 2020. 3 +[49] Yuliang Xiu, Jinlong Yang, Dimitrios Tzionas, and Michael J Black. Icon: Implicit clothed humans obtained from normals. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 13286-13296. IEEE, 2022. 3, 4, 5, 6 +[50] Shan Yang, Tanya Ambert, Zherong Pan, Ke Wang, Licheng Yu, Tamara Berg, and Ming C Lin. Detailed garment recovery from a single-view image. arXiv preprint arXiv:1608.01250, 2016. 4 +[51] Tao Yu, Zerong Zheng, Kaiwen Guo, Pengpeng Liu, Qionghai Dai, and Yebin Liu. Function4d: Real-time human volumetric capture from very sparse consumer rgbd sensors. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR2021), June 2021. 2 +[52] Chao Zhang, Sergi Pujades, Michael J Black, and Gerard Pons-Moll. Detailed, accurate, human shape estimation from clothed 3d scan sequences. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 4191-4200, 2017. 2 +[53] Hongwen Zhang, Yating Tian, Xinchi Zhou, Wanli Ouyang, Yebin Liu, Limin Wang, and Zhenan Sun. Pymaf: 3d human pose and shape regression with pyramidal mesh alignment feedback loop. In Proceedings of the IEEE International Conference on Computer Vision, 2021. 4 +[54] Zerong Zheng, Tao Yu, Yebin Liu, and Qionghai Dai. Pamir: Parametric model-conditioned implicit representation for image-based human reconstruction. IEEE transactions on + +pattern analysis and machine intelligence, 44(6):3170-3184, 2021. 3, 4 +[55] Boyao Zhou, Jean-Sébastien Franco, Federica Bogo, and Edmond Boyer. Spatio-temporal human shape completion with implicit function networks. In 2021 International Conference on 3D Vision (3DV), pages 669-678. IEEE, 2021. 6 +[56] Xingxing Zou, Xiangheng Kong, Waikeung Wong, Congde Wang, Yuguang Liu, and Yang Cao. Fashionai: A hierarchical dataset for fashion understanding. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops, pages 0-0, 2019. 4 +[57] Xingxing Zou and Waikeung Wong. fashion after fashion: A report of ai in fashion. arXiv preprint arXiv:2105.03050, 2021.4 \ No newline at end of file diff --git a/2023/CLOTH4D_ A Dataset for Clothed Human Reconstruction/images.zip b/2023/CLOTH4D_ A Dataset for Clothed Human Reconstruction/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..7c126095db4f6184a23c608a9af7ef9d1e50c82a --- /dev/null +++ b/2023/CLOTH4D_ A Dataset for Clothed Human Reconstruction/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f417623bb313e787374d69eab59018be219024a056012ab53d4f966cd86e19af +size 876789 diff --git a/2023/CLOTH4D_ A Dataset for Clothed Human Reconstruction/layout.json b/2023/CLOTH4D_ A Dataset for Clothed Human Reconstruction/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..ad5a24e4987b9f434f0cc741a7debdad7d88a17f --- /dev/null +++ b/2023/CLOTH4D_ A Dataset for Clothed Human Reconstruction/layout.json @@ -0,0 +1,7817 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 116, + 103, + 477, + 119 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 116, + 103, + 477, + 119 + ], + "spans": [ + { + "bbox": [ + 116, + 103, + 477, + 119 + ], + "type": "text", + "content": "CLOTH4D: A Dataset for Clothed Human Reconstruction" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 164, + 142, + 428, + 158 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 164, + 142, + 428, + 158 + ], + "spans": [ + { + "bbox": [ + 164, + 142, + 428, + 158 + ], + "type": "text", + "content": "Xingxing Zou" + }, + { + "bbox": [ + 164, + 142, + 428, + 158 + ], + "type": "inline_equation", + "content": "^{1,3*}" + }, + { + "bbox": [ + 164, + 142, + 428, + 158 + ], + "type": "text", + "content": ", Xintong Han" + }, + { + "bbox": [ + 164, + 142, + 428, + 158 + ], + "type": "inline_equation", + "content": "^{2*}" + }, + { + "bbox": [ + 164, + 142, + 428, + 158 + ], + "type": "text", + "content": ", Waikeung Wong" + }, + { + "bbox": [ + 164, + 142, + 428, + 158 + ], + "type": "inline_equation", + "content": "^{3,1\\dagger}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 152, + 158, + 441, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 152, + 158, + 441, + 171 + ], + "spans": [ + { + "bbox": [ + 152, + 158, + 441, + 171 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 152, + 158, + 441, + 171 + ], + "type": "text", + "content": "Laboratory for Artificial Intelligence in Design, " + }, + { + "bbox": [ + 152, + 158, + 441, + 171 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 152, + 158, + 441, + 171 + ], + "type": "text", + "content": "Huya Inc." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 121, + 171, + 471, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 121, + 171, + 471, + 186 + ], + "spans": [ + { + "bbox": [ + 121, + 171, + 471, + 186 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 121, + 171, + 471, + 186 + ], + "type": "text", + "content": "School of Fashion and Textiles, The Hong Kong Polytechnic University" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 100, + 188, + 492, + 198 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 100, + 188, + 492, + 198 + ], + "spans": [ + { + "bbox": [ + 100, + 188, + 492, + 198 + ], + "type": "text", + "content": "aemika.zou@connect.polyu.hk,hanxintong@huya.com,calvin.wong@polyu.edu.hk" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 59, + 232, + 170, + 288 + ], + "blocks": [ + { + "bbox": [ + 59, + 232, + 170, + 288 + ], + "lines": [ + { + "bbox": [ + 59, + 232, + 170, + 288 + ], + "spans": [ + { + "bbox": [ + 59, + 232, + 170, + 288 + ], + "type": "image", + "image_path": "7957263b34aa05b1ffe126ee43aefb27d252d1994001e46367c23d43ae8d3cd7.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 82, + 292, + 138, + 303 + ], + "lines": [ + { + "bbox": [ + 82, + 292, + 138, + 303 + ], + "spans": [ + { + "bbox": [ + 82, + 292, + 138, + 303 + ], + "type": "text", + "content": "(a) 2D patterns" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 56, + 313, + 79, + 363 + ], + "blocks": [ + { + "bbox": [ + 56, + 313, + 79, + 363 + ], + "lines": [ + { + "bbox": [ + 56, + 313, + 79, + 363 + ], + "spans": [ + { + "bbox": [ + 56, + 313, + 79, + 363 + ], + "type": "image", + "image_path": "60c3a88e7c0cd270ce27af08f8d5e0b983a4420960d3720027e2a41795918ae7.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 80, + 368, + 139, + 379 + ], + "lines": [ + { + "bbox": [ + 80, + 368, + 139, + 379 + ], + "spans": [ + { + "bbox": [ + 80, + 368, + 139, + 379 + ], + "type": "text", + "content": "(b) varied poses" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 89, + 312, + 108, + 363 + ], + "blocks": [ + { + "bbox": [ + 89, + 312, + 108, + 363 + ], + "lines": [ + { + "bbox": [ + 89, + 312, + 108, + 363 + ], + "spans": [ + { + "bbox": [ + 89, + 312, + 108, + 363 + ], + "type": "image", + "image_path": "e196cf514ca353ba391e3c384d7b33444d7ace70293da689ddc40022b5de99ea.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 121, + 313, + 138, + 363 + ], + "blocks": [ + { + "bbox": [ + 121, + 313, + 138, + 363 + ], + "lines": [ + { + "bbox": [ + 121, + 313, + 138, + 363 + ], + "spans": [ + { + "bbox": [ + 121, + 313, + 138, + 363 + ], + "type": "image", + "image_path": "9c2de006c3de326f318b48c04f10f190d6d781d9037556a51f393895c3ff717a.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 148, + 313, + 171, + 363 + ], + "blocks": [ + { + "bbox": [ + 148, + 313, + 171, + 363 + ], + "lines": [ + { + "bbox": [ + 148, + 313, + 171, + 363 + ], + "spans": [ + { + "bbox": [ + 148, + 313, + 171, + 363 + ], + "type": "image", + "image_path": "8bae1e14ed893ffdd96d688f0481315579be8255ca31cb9c9b0730d0c7bab51d.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 51, + 387, + 81, + 436 + ], + "blocks": [ + { + "bbox": [ + 51, + 387, + 81, + 436 + ], + "lines": [ + { + "bbox": [ + 51, + 387, + 81, + 436 + ], + "spans": [ + { + "bbox": [ + 51, + 387, + 81, + 436 + ], + "type": "image", + "image_path": "87003073c460fc8613142f8a555c2d3e56a91a172f2d495f967dcbec51eefd5b.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 82, + 440, + 138, + 449 + ], + "lines": [ + { + "bbox": [ + 82, + 440, + 138, + 449 + ], + "spans": [ + { + "bbox": [ + 82, + 440, + 138, + 449 + ], + "type": "text", + "content": "(c) multi views" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 86, + 388, + 109, + 436 + ], + "blocks": [ + { + "bbox": [ + 86, + 388, + 109, + 436 + ], + "lines": [ + { + "bbox": [ + 86, + 388, + 109, + 436 + ], + "spans": [ + { + "bbox": [ + 86, + 388, + 109, + 436 + ], + "type": "image", + "image_path": "909b4dadf1846c304d4f45c96200b0f3f4cd046aca1dc61b6a7536a08c784cf7.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 45, + 451, + 546, + 473 + ], + "lines": [ + { + "bbox": [ + 45, + 451, + 546, + 473 + ], + "spans": [ + { + "bbox": [ + 45, + 451, + 546, + 473 + ], + "type": "text", + "content": "Figure 1. An instance in CLOTH4D. Given (a) hand-crafted 2D sewing patterns, clothes are simulated with virtual fashion software to have (b) varied poses, (c) multiple views, and (d) high-quality and physically plausible series of paired dynamic 3D meshes with textures." + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 121, + 388, + 138, + 436 + ], + "blocks": [ + { + "bbox": [ + 121, + 388, + 138, + 436 + ], + "lines": [ + { + "bbox": [ + 121, + 388, + 138, + 436 + ], + "spans": [ + { + "bbox": [ + 121, + 388, + 138, + 436 + ], + "type": "image", + "image_path": "b359570e2358d4a1935537b35a2ff34baa8318f3f03d4f6d962c1fd98a1af424.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 143, + 388, + 173, + 436 + ], + "blocks": [ + { + "bbox": [ + 143, + 388, + 173, + 436 + ], + "lines": [ + { + "bbox": [ + 143, + 388, + 173, + 436 + ], + "spans": [ + { + "bbox": [ + 143, + 388, + 173, + 436 + ], + "type": "image", + "image_path": "251d0b8892bc10cd17cb25bc423f69373897f71b62804ea10ec7e256dde8ae1b.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 186, + 228, + 541, + 438 + ], + "blocks": [ + { + "bbox": [ + 186, + 228, + 541, + 438 + ], + "lines": [ + { + "bbox": [ + 186, + 228, + 541, + 438 + ], + "spans": [ + { + "bbox": [ + 186, + 228, + 541, + 438 + ], + "type": "image", + "image_path": "01b3afc8b98f15453b3e90fc54650f95ee8e2455a796af0535a6d5c2367fac95.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 301, + 440, + 463, + 450 + ], + "lines": [ + { + "bbox": [ + 301, + 440, + 463, + 450 + ], + "spans": [ + { + "bbox": [ + 301, + 440, + 463, + 450 + ], + "type": "text", + "content": "(d) series of paired dynamic textured meshes" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "bbox": [ + 143, + 483, + 192, + 495 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 483, + 192, + 495 + ], + "spans": [ + { + "bbox": [ + 143, + 483, + 192, + 495 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 45, + 510, + 290, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 510, + 290, + 715 + ], + "spans": [ + { + "bbox": [ + 45, + 510, + 290, + 715 + ], + "type": "text", + "content": "Clothed human reconstruction is the cornerstone for creating the virtual world. To a great extent, the quality of recovered avatars decides whether the Metaverse is a passing fad. In this work, we introduce CLOTH4D, a clothed human dataset containing 1,000 subjects with varied appearances, 1,000 3D outfits, and over 100,000 clothed meshes with paired unclothed humans, to fill the gap in large-scale and high-quality 4D clothing data. It enjoys appealing characteristics: 1) Accurate and detailed clothing textured meshes—all clothing items are manually created and then simulated in professional software, strictly following the general standard in fashion design. 2) Separated textured clothing and under-clothing body meshes, closer to the physical world than single-layer raw scans. 3) Clothed human motion sequences simulated given a set of 289 actions, covering fundamental and complicated dynamics. Upon CLOTH4D, we novelly designed a series of" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 304, + 484, + 547, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 484, + 547, + 568 + ], + "spans": [ + { + "bbox": [ + 304, + 484, + 547, + 568 + ], + "type": "text", + "content": "temporally-aware metrics to evaluate the temporal stability of the generated 3D human meshes, which has been overlooked previously. Moreover, by assessing and retraining current state-of-the-art clothed human reconstruction methods, we reveal insights, present improved performance, and propose potential future research directions, confirming our dataset's advancement. The dataset is available at1." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 306, + 582, + 387, + 594 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 582, + 387, + 594 + ], + "spans": [ + { + "bbox": [ + 306, + 582, + 387, + 594 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 304, + 601, + 547, + 686 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 601, + 547, + 686 + ], + "spans": [ + { + "bbox": [ + 304, + 601, + 547, + 686 + ], + "type": "text", + "content": "As we enter the volumetric and XR content era, researchers have been trailblazing their way into the Metaverse. With the converging of technologies and practical applications, e.g., fashion NFTs (non-fungible tokens), immersive AR and VR, and games, clothed human reconstruction demands are rapidly growing. While current research has made astonishing results in creating digital hu" + } + ] + } + ], + "index": 26 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 315, + 693, + 539, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 693, + 539, + 703 + ], + "spans": [ + { + "bbox": [ + 315, + 693, + 539, + 703 + ], + "type": "text", + "content": "1www.github.com/AemikaChow/AiDLab-fAshIon-Data" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 318, + 703, + 525, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 703, + 525, + 713 + ], + "spans": [ + { + "bbox": [ + 318, + 703, + 525, + 713 + ], + "type": "text", + "content": "*X. Zou and X. Han contribute equally. † Corresponding author." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12847" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 49, + 132, + 544, + 304 + ], + "blocks": [ + { + "bbox": [ + 46, + 71, + 547, + 130 + ], + "lines": [ + { + "bbox": [ + 46, + 71, + 547, + 130 + ], + "spans": [ + { + "bbox": [ + 46, + 71, + 547, + 130 + ], + "type": "text", + "content": "Table 1. Comparisons of CLOTH4D with existing representative datasets. Gray color indicates synthetic datasets generated with graphics engines. #Subjects: number of peoples in different appearances; #Action: number of actions adopted; #Scans: numbers of 3D meshes; 2D Pattern: 2D clothing pattern; TexCloth: with textured clothed model; TexHuman: with textured naked human model. w/ SMPL: with registered SMPL [33] parameters. Public: publicly available and free of charge. Photorealistic: whether the images in the dataset are realistic. -: not applicable or reported. CLOTH4D presents more desirable characteristics compared with others." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 49, + 132, + 544, + 304 + ], + "lines": [ + { + "bbox": [ + 49, + 132, + 544, + 304 + ], + "spans": [ + { + "bbox": [ + 49, + 132, + 544, + 304 + ], + "type": "table", + "html": "
Dataset#Subjects#Action#Scan2D PatternTexClothTexHumanw/ SMPLPublicPhotorealistic
BUFF [52]6-13.6k-✓-✓✓✓
RenderPeople [1]--825-✓-✓-✓
DeepWrinkles [30]229.2k-✓---✓
CAPE [35]15600140k---✓✓✓
THuman2.0 [51]200-525-✓-✓✓✓
DRAPE [16]72324.5k------
Wang et al. [47]--24k✓✓-✓✓-
3DPeople [40]8072--✓---✓
DCA [43]-567.1k---✓--
GarNet [17]600-18.8k---✓✓-
TailorNet [38]9-5.5k-✓-✓✓-
Cloth3D [8]8.5k7.9k2.1M-✓-✓✓-
Cloth3D++ [36]9.7k8k2.2M✓✓✓✓✓-
CLOTH4D1k289100k✓✓✓✓✓✓
", + "image_path": "ab23e9cfda56ffa8f1f12c44bc7a869b618ee39838f597bc30021929e168233e.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 309, + 288, + 381 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 309, + 288, + 381 + ], + "spans": [ + { + "bbox": [ + 46, + 309, + 288, + 381 + ], + "type": "text", + "content": "mans, these reconstructed meshes have issues, e.g., flexible body motions and diverse appearances, owing to the lack of datasets with richness in clothing and realistic dynamics of garments. To this end, we introduce CLOTH4D, an open-sourced dataset facilitating physically plausible dynamic clothed human reconstruction." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 385, + 288, + 638 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 385, + 288, + 638 + ], + "spans": [ + { + "bbox": [ + 46, + 385, + 288, + 638 + ], + "type": "text", + "content": "Prior to us, many datasets have been collected, and we sort out them in Table 1. Currently, scanned datasets are widely adopted as they are photorealistic and can be easily processed to watertight meshes, which does an excellent favor for current deep models to learn an implicit function (e.g., signed distance function) followed by marching cubes [34] for surface reconstruction. However, it is born with some weaknesses: 1) Scanned meshes are single-layer and inherently fail to capture the space between clothing and skin surface. Thus, body shape under clothing cannot be accurately inferred, let alone the multi-layer and thin clothing structures as in the real physical world. 2) It is time-consuming and expensive to obtain high-quality and large-scale temporal scanned sequences (i.e., 4D scanned sequences) due to the limited efficiency and precision of 4D scanners, especially for complicated clothing and large motions. Although synthetic datasets can to some extent overcome these limitations, existing synthetic datasets are either of small scale in terms of appearances and motions or are highly unrealistic. Moreover, many datasets are not made publicly available and free." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 642, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 642, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 642, + 288, + 715 + ], + "type": "text", + "content": "In contrast, CLOTH4D possesses several attractive attributes: 1) We made great efforts to the diversity and quality of clothing. All clothes are manually designed in CLO [3] and cater to the requirement of the fashion industry. 2) Meshes in CLOTH4D are clothing/humans separated. Such flexibility makes studying and modeling the" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 309, + 547, + 381 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 309, + 547, + 381 + ], + "spans": [ + { + "bbox": [ + 304, + 309, + 547, + 381 + ], + "type": "text", + "content": "relations and interactions between clothing simulation and body movement possible. 3) CLOTH4D provides plenty of temporal motion sequences with realistic clothing dynamics. As the human body moves, the dressed clothing, e.g., the skirt in Figure 1, naturally deforms. 4) The dataset is large-scale and openly accessible." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 383, + 547, + 587 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 383, + 547, + 587 + ], + "spans": [ + { + "bbox": [ + 304, + 383, + 547, + 587 + ], + "type": "text", + "content": "To demonstrate the advantages of CLOTH4D, we use it to evaluate the state-of-the-art (SOTA) clothed human reconstruction methods. In addition to the generally adopted static evaluation metrics, we propose a set of temporally-aware metrics to assess the temporal coherence in a video inference scenario thanks to the rich and true-to-life 4D synthetic sequences in the dataset. Quantitative and qualitative results of SOTA methods on CLOTH4D suggest that our dataset is challenging and the temporal stability of the reconstructed mesh is vital for evaluating the perceptual quality. Meanwhile, we retrain SOTA methods on CLOTH4D, revealing interesting observations of how they perform on multi-layer meshes with thin clothing structures. With in-depth analysis and a summary of challenges for the existing approaches, CLOTH4D makes an essential step toward more realistic reconstructions of clothed humans and stimulates several exciting future work directions. All in all:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 589, + 547, + 714 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 304, + 589, + 547, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 589, + 547, + 624 + ], + "spans": [ + { + "bbox": [ + 304, + 589, + 547, + 624 + ], + "type": "text", + "content": "- We contribute CLOTH4D, a large-scale, high-quality, and open-accessible 4D synthetic dataset for clothed human reconstruction." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 627, + 547, + 663 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 627, + 547, + 663 + ], + "spans": [ + { + "bbox": [ + 304, + 627, + 547, + 663 + ], + "type": "text", + "content": "- We introduce a series of temporally-aware metrics to evaluate the reconstructed performance in the aspect of temporal consistency." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 665, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 547, + 714 + ], + "type": "text", + "content": "- With the proposed dataset and metrics, we thoroughly analyze the pros and cons of SOTAs, summarize the existing challenges toward more realistic 3D modeling, and propose potential new directions." + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "12848" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 48, + 63, + 218, + 160 + ], + "blocks": [ + { + "bbox": [ + 48, + 63, + 218, + 160 + ], + "lines": [ + { + "bbox": [ + 48, + 63, + 218, + 160 + ], + "spans": [ + { + "bbox": [ + 48, + 63, + 218, + 160 + ], + "type": "image", + "image_path": "c77154e2a7d5a2da5c3da834fcbbf1d35ab89da51d61a44605cdb11f7896b8bd.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 228, + 64, + 378, + 158 + ], + "blocks": [ + { + "bbox": [ + 228, + 64, + 378, + 158 + ], + "lines": [ + { + "bbox": [ + 228, + 64, + 378, + 158 + ], + "spans": [ + { + "bbox": [ + 228, + 64, + 378, + 158 + ], + "type": "image", + "image_path": "8f54d30d586003c4a09b136119930a84e4d9af79c112ce884e7c546ad7abffbc.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 381, + 64, + 546, + 163 + ], + "blocks": [ + { + "bbox": [ + 381, + 64, + 546, + 163 + ], + "lines": [ + { + "bbox": [ + 381, + 64, + 546, + 163 + ], + "spans": [ + { + "bbox": [ + 381, + 64, + 546, + 163 + ], + "type": "image", + "image_path": "adb9343b9720e9b1131e722f6e956721e10478ebf2d18970f7b93c5f7ffad7a9.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 48, + 159, + 218, + 213 + ], + "blocks": [ + { + "bbox": [ + 48, + 159, + 218, + 213 + ], + "lines": [ + { + "bbox": [ + 48, + 159, + 218, + 213 + ], + "spans": [ + { + "bbox": [ + 48, + 159, + 218, + 213 + ], + "type": "image", + "image_path": "6ce2fdad0e1c361c30f99d83c9cbe305b7f6a863ed87ae912b47c167ad7ae29f.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 78, + 220, + 123, + 228 + ], + "lines": [ + { + "bbox": [ + 78, + 220, + 123, + 228 + ], + "spans": [ + { + "bbox": [ + 78, + 220, + 123, + 228 + ], + "type": "text", + "content": "(i) pattern making" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 170, + 175, + 217, + 219 + ], + "blocks": [ + { + "bbox": [ + 170, + 175, + 217, + 219 + ], + "lines": [ + { + "bbox": [ + 170, + 175, + 217, + 219 + ], + "spans": [ + { + "bbox": [ + 170, + 175, + 217, + 219 + ], + "type": "image", + "image_path": "a5f69fbd618a86af8c964ba33f91c83eb123be469fa771684d47c23b99d7e244.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 166, + 221, + 219, + 228 + ], + "lines": [ + { + "bbox": [ + 166, + 221, + 219, + 228 + ], + "spans": [ + { + "bbox": [ + 166, + 221, + 219, + 228 + ], + "type": "text", + "content": "(ii) clothed 3D avatar" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 246, + 177, + 320, + 221 + ], + "blocks": [ + { + "bbox": [ + 246, + 177, + 320, + 221 + ], + "lines": [ + { + "bbox": [ + 246, + 177, + 320, + 221 + ], + "spans": [ + { + "bbox": [ + 246, + 177, + 320, + 221 + ], + "type": "image", + "image_path": "4b01f44e2357bd9ac5ad117d021bf54e1d4a0cd6c1f0f2cfeec5cc14fafca584.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 250, + 222, + 307, + 228 + ], + "lines": [ + { + "bbox": [ + 250, + 222, + 307, + 228 + ], + "spans": [ + { + "bbox": [ + 250, + 222, + 307, + 228 + ], + "type": "text", + "content": "(iii) clothing simulation" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 340, + 200, + 354, + 208 + ], + "blocks": [ + { + "bbox": [ + 340, + 200, + 354, + 208 + ], + "lines": [ + { + "bbox": [ + 340, + 200, + 354, + 208 + ], + "spans": [ + { + "bbox": [ + 340, + 200, + 354, + 208 + ], + "type": "image", + "image_path": "5d88420ffa946a9c1f1cbf55aabf38823672789539611c05e6ed8b30d5814714.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 362, + 170, + 546, + 232 + ], + "blocks": [ + { + "bbox": [ + 362, + 170, + 546, + 232 + ], + "lines": [ + { + "bbox": [ + 362, + 170, + 546, + 232 + ], + "spans": [ + { + "bbox": [ + 362, + 170, + 546, + 232 + ], + "type": "image", + "image_path": "37d67aa9ddb6bb7824306184d553b8b21117402a5214bf91ed6d74ec43f4bb8d.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 409, + 233, + 500, + 240 + ], + "lines": [ + { + "bbox": [ + 409, + 233, + 500, + 240 + ], + "spans": [ + { + "bbox": [ + 409, + 233, + 500, + 240 + ], + "type": "text", + "content": "CLOTH4D with varied 3D sequences" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 244, + 546, + 266 + ], + "lines": [ + { + "bbox": [ + 46, + 244, + 546, + 266 + ], + "spans": [ + { + "bbox": [ + 46, + 244, + 546, + 266 + ], + "type": "text", + "content": "Figure 2. Pipeline for creating instances in CLOTH4D, which primarily adopts CLO for clothing design and simulation, Mixamo for animation, and Blender for processing and exporting meshes." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 271, + 134, + 283 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 271, + 134, + 283 + ], + "spans": [ + { + "bbox": [ + 47, + 271, + 134, + 283 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 46, + 293, + 287, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 293, + 287, + 483 + ], + "spans": [ + { + "bbox": [ + 46, + 293, + 287, + 483 + ], + "type": "text", + "content": "Clothed Human Reconstruction. Clothed human reconstruction aims to recover a 3D mesh from a monocular person image. Estimating dressed people by modeling clothing geometry as 3D displacement on top of a parametric body model (e.g., SMPL [33]) is the leading solution to deal with this task [5,6,28,31,39,48]. By transferring the skinning weights from the body model to the offset clothing mesh, the reconstructed clothed mesh can be readily deformed and animated in the same way as the under-clothing 3D parametric body model. However, it assumes the clothed human to have the same topology as the naked body, leading to unsatisfactory reconstructions of long hair, skirts, dresses, etc. Although methods such as [9,22,26,38] try to isolate the reconstruction of clothing by constructing category-specific clothing templates or statistical models, they fail to generalize to unseen or complex types of clothing." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 46, + 485, + 287, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 485, + 287, + 688 + ], + "spans": [ + { + "bbox": [ + 46, + 485, + 287, + 688 + ], + "type": "text", + "content": "As another line of research, deep implicit function networks have drawn broader attention recently [7, 15, 20, 21, 24, 41, 42, 54]. PIFu [41] conditions on pixel-aligned features to build deep implicit functions for reconstructing human meshes, and PIFuHD [42] goes a further step towards enhancing 3D geometric details by predicting front and back normals. ARCH [24] and " + }, + { + "bbox": [ + 46, + 485, + 287, + 688 + ], + "type": "inline_equation", + "content": "\\mathrm{ARCH}++" + }, + { + "bbox": [ + 46, + 485, + 287, + 688 + ], + "type": "text", + "content": " [21] enable animating the reconstructed meshes by deforming the semantic information into the SMPL canonical space. More recently, leveraging SMPL body models as prior, PaMIR [54] and ICON [49] further improve the reconstruction quality, especially on challenging poses. PHORHUM [7] achieves more accurate results via jointly estimating geometry, albedo, and shading information. It is worth noting that most works rely on training data that are not available free of charge, making an open-sourced dataset of clothed avatars of great significance to advance research in 3D body/clothing modeling." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 46, + 689, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 689, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 689, + 287, + 713 + ], + "type": "text", + "content": "Clothed Human Datasets. As summarized in Table 1, existing clothed human datasets can be divided into two" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 272, + 546, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 272, + 546, + 475 + ], + "spans": [ + { + "bbox": [ + 304, + 272, + 546, + 475 + ], + "type": "text", + "content": "types, i.e., scanned datasets and synthetic datasets. The former [10, 25, 44, 46] utilizes multiple synchronized cameras to capture motions that have difficulties enriching the scalability and diversity and obtaining highly accurate 4D ground truth. Moreover, it inherently cannot mimic the layering of clothing. The synthetic datasets mitigate these limitations. However, existing synthetic datasets, no matter whether static [17, 38, 47] or dynamic [8, 16, 39, 43], either only contain a few clothing types or are highly unrealistic. Cloth3D++ [36] is developed from Cloth3D [8], which contains a total of 2.2 million scans, covering 9.7k subjects dressed in 12.9k clothing, which is the most advanced in scale. However, the clothing created based on garment templates are only base patterns with an immense gap compared to clothing in real life. To intuitively demonstrate the advantages of CLOTH4D, we put visual comparisons of these datasets in Figure A in the supplementary material." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 306, + 488, + 420, + 501 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 488, + 420, + 501 + ], + "spans": [ + { + "bbox": [ + 306, + 488, + 420, + 501 + ], + "type": "text", + "content": "3. CLOTH4D Dataset" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 509, + 545, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 509, + 545, + 593 + ], + "spans": [ + { + "bbox": [ + 304, + 509, + 545, + 593 + ], + "type": "text", + "content": "We depict the pipeline of creating CLOTH4D in Figure 2, including (1) preparing the referenced clothing images and the unclothed 3D human avatars; (2) uploading 3D human avatars to Mixamo to obtain the FBX files with various animations; (3) designing 3D clothing in CLO, integrating the FBX files obtained in (2), and conducting clothing simulation; and (4) exporting the sequenced mesh files." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 594, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 546, + 713 + ], + "type": "text", + "content": "Clothes and Models. Clothes are manually created by professional fashion designers using CLO by producing the 2D garment patterns and then auto-simulating them to 3D. We put a video illustrating the production process of a 3D garment in the supplementary material (Video A) instead of demonstrating details in the paper. CLOTH4D covers 1,000 different 3D outfits spansning over 500 prints and 50 fabrics. Over " + }, + { + "bbox": [ + 304, + 594, + 546, + 713 + ], + "type": "inline_equation", + "content": "40\\%" + }, + { + "bbox": [ + 304, + 594, + 546, + 713 + ], + "type": "text", + "content": " of the clothes are designed referring to the newest collections of varying design houses (e.g., Prada, Moschino, and Alexander McQueen) to ensure visual re" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12849" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 144 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 144 + ], + "type": "text", + "content": "alism, variance, and fashionability. In this paper, we focus on women's wear owing to its diversity covering most characteristics of garments. For avatars, CLO provides human avatars with varied physical appearances in terms of hairstyles, faces, skin colors, body figures, etc., and we directly adopt these available avatars." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 144, + 288, + 287 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 144, + 288, + 287 + ], + "spans": [ + { + "bbox": [ + 46, + 144, + 288, + 287 + ], + "type": "text", + "content": "Animations and Simulations. For each human avatar, we use Mixamo [2] for rigging and generating motion sequences. Given a motion sequence and a 3D garment, CLO runs a clothing simulation of the 3D garment according to the motion sequence at 30fps. A total of 289 animations are utilized, such as Billy Dance, Offensive Idle, Jumping Rope, etc. Unlike previous work [8] directly uses Blender for cloth simulation, CLO simulates clothing with richer details and dynamics. The clothes exhibit wrinkles and fold with the body's movement in a natural and physically plausible way, especially for skirts or dresses that differ a lot from the body topology as shown in Figure 1." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 288, + 288, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 288, + 288, + 468 + ], + "spans": [ + { + "bbox": [ + 46, + 288, + 288, + 468 + ], + "type": "text", + "content": "Paired Multi-Modal Data. Then, we can obtain paired data in the following representations: a 2D clothing pattern, 3D mesh sequences (clothed and naked human meshes with corresponding fitted SMPL parameters/meshes, and separate clothing meshes), and a UV texture map. The textured mesh can be rendered into multi-view normal images, depth images, and RGB images given varying light conditions. We also translate all these dynamic meshes to watertight with simplification using [23], thus they can be readily used to train an implicit function for mesh reconstruction. The number of triangles ranges from " + }, + { + "bbox": [ + 46, + 288, + 288, + 468 + ], + "type": "inline_equation", + "content": "170K" + }, + { + "bbox": [ + 46, + 288, + 288, + 468 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 46, + 288, + 288, + 468 + ], + "type": "inline_equation", + "content": "4M" + }, + { + "bbox": [ + 46, + 288, + 288, + 468 + ], + "type": "text", + "content": " for simulation and becomes " + }, + { + "bbox": [ + 46, + 288, + 288, + 468 + ], + "type": "inline_equation", + "content": "200K" + }, + { + "bbox": [ + 46, + 288, + 288, + 468 + ], + "type": "text", + "content": " after simplification. Besides human reconstruction, many other tasks could also benefit from this paired data (e.g., clothing capture [50], human pose transfer [18], and fashion-related tasks [19, 56, 57])." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 477, + 124, + 490 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 477, + 124, + 490 + ], + "spans": [ + { + "bbox": [ + 47, + 477, + 124, + 490 + ], + "type": "text", + "content": "4. Evaluations" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 497, + 287, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 497, + 287, + 594 + ], + "spans": [ + { + "bbox": [ + 46, + 497, + 287, + 594 + ], + "type": "text", + "content": "In this section, we evaluate the state-of-the-art clothed human reconstruction methods on CLOTH4D to demonstrate new insights that the dataset can provide. Further, we retrain SOTA methods on CLOTH4D and make several interesting observations of how they perform on multi-layer meshes with thin clothing structures. We also present the challenges for existing approaches and propose potential research directions with a comprehensive analysis." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 599, + 113, + 611 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 599, + 113, + 611 + ], + "spans": [ + { + "bbox": [ + 47, + 599, + 113, + 611 + ], + "type": "text", + "content": "4.1. Baselines" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 617, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 617, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 617, + 288, + 715 + ], + "type": "text", + "content": "We mainly report results of four SOTA approaches, PIFu [41], PIFuHD [42], PaMIR [54], and ICON [49] owing to other works do not release their codes or models, such as PHORHUM [7], ARCH++ [21], or which have already been extensively compared with the listed methods above. We use PIFu, PIFuHD, PaMIR, and ICON to denote their released pretrained testing models. " + }, + { + "bbox": [ + 46, + 617, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\mathrm{PaMIR}^{gt}" + }, + { + "bbox": [ + 46, + 617, + 288, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 617, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\mathrm{ICON}^{gt}" + }, + { + "bbox": [ + 46, + 617, + 288, + 715 + ], + "type": "text", + "content": " indicate the testing results using the fitted ground truth" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 72, + 545, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 167 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 167 + ], + "type": "text", + "content": "SMPL mesh as conditions rather than the one predicted using off-the-shelf human mesh recovery (HMR) methods (GraphCMR [29] for PaMIR, and PyMAf [53] for ICON as in their released code). Based on the characteristics of these methods, they can be divided into three types: 1) pixel-aligned methods (PIFu, PIFuHD); 2) GT-SMPL-guided + pixel-aligned methods " + }, + { + "bbox": [ + 304, + 72, + 545, + 167 + ], + "type": "inline_equation", + "content": "(\\mathrm{PaMIR}^{gt}" + }, + { + "bbox": [ + 304, + 72, + 545, + 167 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 72, + 545, + 167 + ], + "type": "inline_equation", + "content": "\\mathrm{ICON}^{gt})" + }, + { + "bbox": [ + 304, + 72, + 545, + 167 + ], + "type": "text", + "content": "; 3) HMR-SMPL-guided + pixel-aligned methods (PaMIR and ICON)." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 168, + 545, + 276 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 168, + 545, + 276 + ], + "spans": [ + { + "bbox": [ + 304, + 168, + 545, + 276 + ], + "type": "text", + "content": "Furthermore, we retrain PIFu, PaMIR, and ICON on CLOTH4D, which are denoted as " + }, + { + "bbox": [ + 304, + 168, + 545, + 276 + ], + "type": "inline_equation", + "content": "\\mathrm{PIF_u_{clo}}" + }, + { + "bbox": [ + 304, + 168, + 545, + 276 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 168, + 545, + 276 + ], + "type": "inline_equation", + "content": "\\mathrm{PaMIR}_{clo}" + }, + { + "bbox": [ + 304, + 168, + 545, + 276 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 168, + 545, + 276 + ], + "type": "inline_equation", + "content": "\\mathrm{ICON}_{clo}" + }, + { + "bbox": [ + 304, + 168, + 545, + 276 + ], + "type": "text", + "content": ", respectively. For these retrained models, we follow the re-implementation setting introduced in ICON [4], which allows us to train all these baselines with the same training protocol and hyper-parameters for a fair comparison. Similarly, " + }, + { + "bbox": [ + 304, + 168, + 545, + 276 + ], + "type": "inline_equation", + "content": "\\mathrm{PaMIR}_{clo}^{gt}" + }, + { + "bbox": [ + 304, + 168, + 545, + 276 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 168, + 545, + 276 + ], + "type": "inline_equation", + "content": "\\mathrm{ICON}_{clo}^{gt}" + }, + { + "bbox": [ + 304, + 168, + 545, + 276 + ], + "type": "text", + "content": " are tested with the ground truth SMPL fits. We also use the cloth-refinement module in the ICON's released code for post-processing." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 281, + 428, + 294 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 281, + 428, + 294 + ], + "spans": [ + { + "bbox": [ + 305, + 281, + 428, + 294 + ], + "type": "text", + "content": "4.2. Datasets and Metrics" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 300, + 545, + 408 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 300, + 545, + 408 + ], + "spans": [ + { + "bbox": [ + 304, + 300, + 545, + 408 + ], + "type": "text", + "content": "Datasets and implementation details. We organize the sequences in CLOTH4D into a " + }, + { + "bbox": [ + 304, + 300, + 545, + 408 + ], + "type": "inline_equation", + "content": "80\\% / 10\\% / 10\\%" + }, + { + "bbox": [ + 304, + 300, + 545, + 408 + ], + "type": "text", + "content": " train/val/test split. We render each mesh into 8 views using a weak perspective camera and pre-computed radiance transfer [45] with dynamic light conditions following [41, 49]. All rendered images are " + }, + { + "bbox": [ + 304, + 300, + 545, + 408 + ], + "type": "inline_equation", + "content": "512 \\times 512" + }, + { + "bbox": [ + 304, + 300, + 545, + 408 + ], + "type": "text", + "content": ". The 2D keypoints used in all methods are generated by OpenPose [11]. In addition, we also evaluate all models on CAPE [35] test set adopted in [49] to investigate the generalization ability." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 407, + 545, + 503 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 407, + 545, + 503 + ], + "spans": [ + { + "bbox": [ + 304, + 407, + 545, + 503 + ], + "type": "text", + "content": "Static Metrics. We report the quantitative results on normal reprojection error, Chamfer distance, and P2S distance for evaluation as [15, 41, 42, 49]. As all compared methods use a weak perspective or orthographic camera, the estimated meshes may not be well aligned with the ground truth meshes in the " + }, + { + "bbox": [ + 304, + 407, + 545, + 503 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 304, + 407, + 545, + 503 + ], + "type": "text", + "content": "-direction (i.e., view direction). Thus, we shift the estimated meshes to have the same " + }, + { + "bbox": [ + 304, + 407, + 545, + 503 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 304, + 407, + 545, + 503 + ], + "type": "text", + "content": "-axis mean as the ground truths following [15] for a fair comparison." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 503, + 545, + 634 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 503, + 545, + 634 + ], + "spans": [ + { + "bbox": [ + 304, + 503, + 545, + 634 + ], + "type": "text", + "content": "Temporal Metrics. The aforementioned static metrics ignore the temporal consistency of the reconstructed meshes across time, which is essential for real-time applications since meshes presenting jitters and flickers highly affect the perceptual quality. Thanks to the rich temporal dynamics provided in CLOTH4D, we are the first to introduce temporally-aware metrics to evaluate the temporal coherence of the generated mesh sequences. Referring to temporal metrics SSDdt and dtSSD used in video matting tasks [14, 32], we compute two metrics measuring the temporal coherence of the predicted mesh normal:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 312, + 637, + 545, + 659 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 637, + 545, + 659 + ], + "spans": [ + { + "bbox": [ + 312, + 637, + 545, + 659 + ], + "type": "interline_equation", + "content": "\\text {N o r m a l s} _ {d d t} = \\frac {1}{T} \\sum_ {t} \\left| \\left(\\mathcal {N} _ {t} ^ {p r} - \\mathcal {N} _ {t} ^ {g t}\\right) ^ {2} - \\left(\\mathcal {N} _ {t + 1} ^ {p r} - \\mathcal {N} _ {t + 1} ^ {g t}\\right) ^ {2} \\right|, \\tag {1}", + "image_path": "61fd9d6efa3168fdf28c8bc9019e0e85c0ec83dba06cc4546df385a0c77f758e.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 312, + 663, + 545, + 686 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 663, + 545, + 686 + ], + "spans": [ + { + "bbox": [ + 312, + 663, + 545, + 686 + ], + "type": "interline_equation", + "content": "\\operatorname {N o r m a l s} _ {d t d} = \\frac {1}{T} \\sum_ {t} \\left| \\left(\\mathcal {N} _ {t} ^ {p r} - \\mathcal {N} _ {t + 1} ^ {p r}\\right) ^ {2} - \\left(\\mathcal {N} _ {t} ^ {g t} - \\mathcal {N} _ {t + 1} ^ {g t}\\right) ^ {2} \\right|, \\tag {2}", + "image_path": "0569a6345748380a9148f2118084acd3df99b6cce8ac72559c11cc442c34bcbc.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "content": " is the length of the sequence. " + }, + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{N}_t^{pr}" + }, + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{N}_t^{gt}" + }, + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "content": " denote the rendered normal images from the predicted mesh" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12850" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 50, + 95, + 544, + 208 + ], + "blocks": [ + { + "bbox": [ + 47, + 70, + 545, + 94 + ], + "lines": [ + { + "bbox": [ + 47, + 70, + 545, + 94 + ], + "spans": [ + { + "bbox": [ + 47, + 70, + 545, + 94 + ], + "type": "text", + "content": "Table 2. Quantitative evaluation on CLOTH4D. PaMIRgt and ICONgt denote that the fitted ground truth SMPL is used during the inference. Gray color indicates the results trained on CLOTH4D." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 50, + 95, + 544, + 208 + ], + "lines": [ + { + "bbox": [ + 50, + 95, + 544, + 208 + ], + "spans": [ + { + "bbox": [ + 50, + 95, + 544, + 208 + ], + "type": "table", + "html": "
MethodPIFuPIFuHDPaMIR\\( PaMIR^{gt} \\)ICON\\( ICON^{gt} \\)\\( PIFu_{clo} \\)\\( PaMIR_{clo} \\)\\( PaMIR^{gt}_{clo} \\)\\( ICON_{clo} \\)\\( ICON^{gt}_{clo} \\)
Normals \\( \\downarrow \\)0.1820.1810.2240.1450.2110.1180.1500.2300.1140.1980.103
P2S \\( \\downarrow \\)3.9113.5183.8272.6414.6602.4732.7935.0372.6193.7112.068
Chamfer \\( \\downarrow \\)3.5782.4874.1572.4874.1961.6312.4124.1861.6183.4991.367
\\( Normals_{ddt} \\) \\( \\downarrow \\)0.0080.0070.0320.0250.0130.0300.0090.0210.0330.0130.035
\\( Normals_{dtd} \\) \\( \\downarrow \\)0.0250.0230.0430.0350.0340.0280.0170.0400.0270.0380.033
\\( P2S_{ddt} \\) \\( \\downarrow \\)0.2220.1670.2530.2470.3830.3690.1850.4170.4020.3210.367
\\( P2S_{dtd} \\) \\( \\downarrow \\)0.7330.5781.0250.7020.9790.5540.5480.8900.5360.8780.526
Chamfer \\( _{ddt} \\) \\( \\downarrow \\)0.2070.1810.3310.3160.3590.3580.1570.3500.3820.3310.369
Chamfer \\( _{dtd} \\) \\( \\downarrow \\)0.7290.5761.0170.7010.9740.5530.5460.8860.5360.8730.525
", + "image_path": "a6f99568675b623722d33408c121630d3fc11daa39813c1fae221bf12bd82afe.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 50, + 225, + 544, + 278 + ], + "blocks": [ + { + "bbox": [ + 145, + 213, + 447, + 224 + ], + "lines": [ + { + "bbox": [ + 145, + 213, + 447, + 224 + ], + "spans": [ + { + "bbox": [ + 145, + 213, + 447, + 224 + ], + "type": "text", + "content": "Table 3. Quantitative evaluation on CAPE. Table notations are the same as Table 2." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 50, + 225, + 544, + 278 + ], + "lines": [ + { + "bbox": [ + 50, + 225, + 544, + 278 + ], + "spans": [ + { + "bbox": [ + 50, + 225, + 544, + 278 + ], + "type": "table", + "html": "
MethodPIFuPIFuHDPaMIRPaMIRgtICONICONgtPIFucloPaMIRcloPaMIRgtcloICONcloICONgtclo
Normals ↓0.1610.1600.1830.0860.1600.0560.1640.1760.0930.1560.077
P2S ↓4.2593.7953.8401.1934.0141.0674.6524.2351.4913.3041.193
Chamfer ↓4.2043.9274.2581.6543.9621.0384.3814.0801.4273.5441.397
", + "image_path": "930820348bdb765f11610271205d471c5a4f5822f27945ab874c5514ca54e2c4.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 285, + 289, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 285, + 289, + 392 + ], + "spans": [ + { + "bbox": [ + 46, + 285, + 289, + 392 + ], + "type": "text", + "content": "and the ground truth mesh at time step " + }, + { + "bbox": [ + 46, + 285, + 289, + 392 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 285, + 289, + 392 + ], + "type": "text", + "content": ", respectively. The subscript " + }, + { + "bbox": [ + 46, + 285, + 289, + 392 + ], + "type": "inline_equation", + "content": "ddt" + }, + { + "bbox": [ + 46, + 285, + 289, + 392 + ], + "type": "text", + "content": " is short for distance delta time, which captures the stability of errors between two consecutive meshes. And " + }, + { + "bbox": [ + 46, + 285, + 289, + 392 + ], + "type": "inline_equation", + "content": "dtd" + }, + { + "bbox": [ + 46, + 285, + 289, + 392 + ], + "type": "text", + "content": " (delta time distance) penalizes large temporal change of the prediction with respect to the change of the ground truth. These two metrics indicate unstable mesh variations and ignore temporally coherent errors [14]. The " + }, + { + "bbox": [ + 46, + 285, + 289, + 392 + ], + "type": "inline_equation", + "content": "ddt" + }, + { + "bbox": [ + 46, + 285, + 289, + 392 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 285, + 289, + 392 + ], + "type": "inline_equation", + "content": "dtd" + }, + { + "bbox": [ + 46, + 285, + 289, + 392 + ], + "type": "text", + "content": " of Chamfer and P2S distances are similarly defined. More details can be found in the supplementary material." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 403, + 163, + 415 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 403, + 163, + 415 + ], + "spans": [ + { + "bbox": [ + 47, + 403, + 163, + 415 + ], + "type": "text", + "content": "4.3. Baseline Evaluation" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 423, + 287, + 471 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 423, + 287, + 471 + ], + "spans": [ + { + "bbox": [ + 46, + 423, + 287, + 471 + ], + "type": "text", + "content": "Quantitative results. Table 2 gives quantitative results on the CLOTH4D test set using the evaluation metrics described in Section 4.2. As indicated in the non-gray part, we made the following observations:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 472, + 288, + 714 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 46, + 472, + 287, + 520 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 472, + 287, + 520 + ], + "spans": [ + { + "bbox": [ + 46, + 472, + 287, + 520 + ], + "type": "text", + "content": "1) In terms of static metrics, " + }, + { + "bbox": [ + 46, + 472, + 287, + 520 + ], + "type": "inline_equation", + "content": "\\mathrm{ICON}^{gt} > \\mathrm{PaMIR}^{gt} > \\mathrm{PI}-\\mathrm{FuHD} > \\mathrm{PIF_u} > \\mathrm{PaMIR} > \\mathrm{ICON}" + }, + { + "bbox": [ + 46, + 472, + 287, + 520 + ], + "type": "text", + "content": ". I.e., GT-SMPL-guided + pixel-aligned methods > pure pixel-aligned methods > HMR-SMPL-guided + pixel-aligned methods." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 521, + 287, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 521, + 287, + 628 + ], + "spans": [ + { + "bbox": [ + 46, + 521, + 287, + 628 + ], + "type": "text", + "content": "2) With the strong guidance of ground truth SMPL mesh, the performance of " + }, + { + "bbox": [ + 46, + 521, + 287, + 628 + ], + "type": "inline_equation", + "content": "\\mathrm{ICON}^{gt}" + }, + { + "bbox": [ + 46, + 521, + 287, + 628 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 521, + 287, + 628 + ], + "type": "inline_equation", + "content": "\\mathrm{PaMIR}^{gt}" + }, + { + "bbox": [ + 46, + 521, + 287, + 628 + ], + "type": "text", + "content": " significantly improves compared to their counterparts with estimated SMPL (i.e., ICON and PaMIR). However, ground truth SMPL meshes are unavailable at test time, which suggests that previous comparisons [15, 49] between GT-SMPL-based methods and others may be unfair. And pure pixel-aligned methods may be even more favorable than SOTA SMPL-based methods for the in-the-wild scenario." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 630, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 630, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 630, + 288, + 714 + ], + "type": "text", + "content": "3) From the perspective of temporally-aware metrics, pure pixel-aligned methods have higher reconstruction stability (see Figure 4). We attribute this to the fact that ICON and PaMIR strongly rely on the SMPL body prior, thus failing to generate far-from-the-body clothes (e.g., skirts and dresses) that present rich temporal dynamics as shown in Figure 1. Plus, the jittery and unstable pose estimation of" + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 304, + 285, + 545, + 310 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 285, + 545, + 310 + ], + "spans": [ + { + "bbox": [ + 304, + 285, + 545, + 310 + ], + "type": "text", + "content": "the off-the-shelf HMR methods further prevents ICON and PaMIR from generating temporally coherent results." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 312, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 304, + 312, + 545, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 312, + 545, + 384 + ], + "spans": [ + { + "bbox": [ + 304, + 312, + 545, + 384 + ], + "type": "text", + "content": "4) ICON outperforms PaMIR as ICON better models mesh-based local features while PaMIR depends more on global information. Moreover, PaMIR loses high-frequency details due to the limited resolution of its volumetric representation. Our observations on CAPE are in line with previous works. For simplicity, we do not expand the narrative." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 385, + 545, + 433 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 385, + 545, + 433 + ], + "spans": [ + { + "bbox": [ + 304, + 385, + 545, + 433 + ], + "type": "text", + "content": "Qualitative results. We present the qualitative results on CLOTH4D in Figure 3 and draw the following insights that have not been fully explored since there are no such large-scale and diverse datasets like CLOTH4D." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 436, + 545, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 436, + 545, + 628 + ], + "spans": [ + { + "bbox": [ + 304, + 436, + 545, + 628 + ], + "type": "text", + "content": "1) Global shape vs. local details. All baselines can reconstruct the overall shape conditioned on the input RGB image. PIFuHD presents the finest details, followed by ICON, PaMIR, and PIFu, as PIFuHD enlarges the spatial resolution of pixel-aligned features and ICON takes advantage of mesh-based local features (signed distance, surface normal, etc.). However, focusing on local features suffers from overfitting and poor generalization to complicated clothing (e.g., incomplete dress in the 3rd and 5th examples in Figure 3) and large motions (e.g., artifacts in the arm regions in the 2nd, 3rd, 6th, and 8th examples in Figure 3). Comparatively, thanks to the global feature encoder, PaMIR, and PIFu can generate more holistic clothing but sacrifice details. Thus, it is an important future research direction to explore better strategies for balancing the local and global reconstruction quality." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "content": "2) Human body priors. The last three rows in Figure 3 and Figure 4 (also shows side views) present results on relatively challenging poses. " + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{ICON}^{gt}" + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "content": " and ICON robustly recover the poses, while PIFu, PIFuHD, and PaMIR are prone to producing broken limbs or anatomically improbable shapes to different extents. For the side view, we can find that " + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{PaMIR}^{gt}" + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{ICON}^{gt}" + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "content": " are more similar to the" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12851" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 69, + 289, + 390 + ], + "blocks": [ + { + "bbox": [ + 50, + 69, + 289, + 390 + ], + "lines": [ + { + "bbox": [ + 50, + 69, + 289, + 390 + ], + "spans": [ + { + "bbox": [ + 50, + 69, + 289, + 390 + ], + "type": "image", + "image_path": "e19d45237e7f35dc621f920218b042fef81b42e57a3a7fe951d6051da6e588d2.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 87, + 390, + 247, + 400 + ], + "lines": [ + { + "bbox": [ + 87, + 390, + 247, + 400 + ], + "spans": [ + { + "bbox": [ + 87, + 390, + 247, + 400 + ], + "type": "text", + "content": "Figure 3. Qualitative results on CLOTH4D." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 411, + 287, + 543 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 411, + 287, + 543 + ], + "spans": [ + { + "bbox": [ + 46, + 411, + 287, + 543 + ], + "type": "text", + "content": "ground truth mesh, which is unsurprising as they are given the ground truth SMPL as prior. Comparatively, PaMIR and ICON, which adopt the estimated SMPL, face the common problems of HMR-SMPL-based reconstruction methods—the predicted SMPL body bends legs or hunches over due to the depth ambiguity—leading to large reconstruction error. Similarly, PIFu and PIFuHD tend to have forward heads with slightly bending legs as they are not aware of any human body priors. One potential research direction is to jointly train or optimize body prior (e.g., SMPL, keypoints, human parsing) with mesh reconstruction." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 545, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 545, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 545, + 288, + 715 + ], + "type": "text", + "content": "3) Ambiguity of geometry and appearance. As shown in the 6th-8th rows of Figure 3, the clothing prints can affect the surface reconstruction due to the ambiguity of geometry and appearance in single-view rendering. Among all these baselines, ICON and PIFuHD show high robustness to the input clothing prints as they predict normal images as intermediate representations, which reduces this ambiguity compared to directly inputting RGB images. Note that ICON only takes the normal images as input to the reconstruction module without using the RGB images, further mitigating the ambiguity and bringing even higher robustness than PIFuHD (the 8th example). Motivated by this observation, future research could shed more light on disentangling the geometry and appearance either implicitly [37] or explicit" + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 310, + 68, + 545, + 310 + ], + "blocks": [ + { + "bbox": [ + 310, + 68, + 545, + 310 + ], + "lines": [ + { + "bbox": [ + 310, + 68, + 545, + 310 + ], + "spans": [ + { + "bbox": [ + 310, + 68, + 545, + 310 + ], + "type": "image", + "image_path": "6e62e8b77b9da05f07ed72c7370e6a471563bba9b6837cbffd73fd5f8cebed2c.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 312, + 545, + 357 + ], + "lines": [ + { + "bbox": [ + 305, + 312, + 545, + 357 + ], + "spans": [ + { + "bbox": [ + 305, + 312, + 545, + 357 + ], + "type": "text", + "content": "Figure 4. Temporal qualitative results. The 1st, 3rd, and 5th rows are three consecutive frames, and the 2nd, 4th, and 6th rows are the side views predicted from the corresponding front-view RGB. Refer to the supplementary material (Video B) for video results." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 305, + 365, + 545, + 402 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 365, + 545, + 402 + ], + "spans": [ + { + "bbox": [ + 305, + 365, + 545, + 402 + ], + "type": "text", + "content": "ity [7]. Also, predicting more comprehensive intermediate 2D/3D representations (e.g., depth, illumination, keypoints, segmentation) may also improve the performance." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 403, + 546, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 403, + 546, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 403, + 546, + 714 + ], + "type": "text", + "content": "4) Temporal consistency. For real-time applications, e.g., streaming from a monocular camera and importing the reconstructed motion sequence into a virtual scene, the temporal coherence of the generated meshes over time is vital for a high-quality user experience. We show the reconstructed meshes for three consecutive frames in Figure 4. As can be found from the side views of the reconstructed meshes, although only subtle motions are present in these frames, the HMR-SMPL-guided methods (PaMIR and ICON) suffer from unstable SMPL predictions and generate temporally inconsistent meshes. On the other hand, pure pixel-aligned methods (PIFu and PIFuHD) fail to predict accurate human pose but produce temporally consistent errors, thus having small " + }, + { + "bbox": [ + 304, + 403, + 546, + 714 + ], + "type": "inline_equation", + "content": "ddt" + }, + { + "bbox": [ + 304, + 403, + 546, + 714 + ], + "type": "text", + "content": " values. Given the ground truth SMPL meshes, " + }, + { + "bbox": [ + 304, + 403, + 546, + 714 + ], + "type": "inline_equation", + "content": "\\mathrm{PaMIR}^{gt}" + }, + { + "bbox": [ + 304, + 403, + 546, + 714 + ], + "type": "text", + "content": " is still more sensitive to global pose than " + }, + { + "bbox": [ + 304, + 403, + 546, + 714 + ], + "type": "inline_equation", + "content": "\\mathrm{ICON}^{gt}" + }, + { + "bbox": [ + 304, + 403, + 546, + 714 + ], + "type": "text", + "content": " as also noted by [49]. These observations are consistent with the quantitative temporal metrics reported in Table 2. It would be interesting to investigate temporal modeling of implicit functions (e.g., incorporating a recurrent neural network [27, 32, 55] to train the implicit function on the 4D dataset, or applying test-time fitting to refine the reconstructed meshes with temporal loss terms). We make such an attempt by adding a temporal term (penalize Chamfer distance between two successive frames) to the refinement process of " + }, + { + "bbox": [ + 304, + 403, + 546, + 714 + ], + "type": "inline_equation", + "content": "\\mathrm{PIFu}_{clo}" + }, + { + "bbox": [ + 304, + 403, + 546, + 714 + ], + "type": "text", + "content": " and achieve better temporal consistency (Chamfer_ddt: " + }, + { + "bbox": [ + 304, + 403, + 546, + 714 + ], + "type": "inline_equation", + "content": "0.157\\rightarrow 0.123" + }, + { + "bbox": [ + 304, + 403, + 546, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 403, + 546, + 714 + ], + "type": "inline_equation", + "content": "\\mathrm{Chamfer}_{dtd}" + }, + { + "bbox": [ + 304, + 403, + 546, + 714 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12852" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 48, + 68, + 545, + 542 + ], + "blocks": [ + { + "bbox": [ + 48, + 68, + 545, + 542 + ], + "lines": [ + { + "bbox": [ + 48, + 68, + 545, + 542 + ], + "spans": [ + { + "bbox": [ + 48, + 68, + 545, + 542 + ], + "type": "image", + "image_path": "cc8d9eb1464f0be0cb588a3b7b5f238d88672be921d44dac08256fe4abed7336.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 113, + 543, + 479, + 554 + ], + "lines": [ + { + "bbox": [ + 113, + 543, + 479, + 554 + ], + "spans": [ + { + "bbox": [ + 113, + 543, + 479, + 554 + ], + "type": "text", + "content": "Figure 5. Qualitative results of baseline methods tested on CLOTH4D with front view and side view." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 556, + 111, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 556, + 111, + 567 + ], + "spans": [ + { + "bbox": [ + 47, + 556, + 111, + 567 + ], + "type": "inline_equation", + "content": "0.546\\rightarrow 0.454)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 575, + 176, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 575, + 176, + 586 + ], + "spans": [ + { + "bbox": [ + 47, + 575, + 176, + 586 + ], + "type": "text", + "content": "4.4. Baseline Enhancement" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 594, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 287, + 713 + ], + "type": "text", + "content": "In addition to the findings and insights discovered in Section 4.3 by evaluating on CLOTH4D. CLOTH4D, containing various clothing types and motion sequences, as well as true-to-life multi-layer thin 3D clothing structures, poses new challenges to clothed human reconstruction research. To investigate the SOTA performance when trained on CLOTH4D, we re-train PIFu, PaMIR, and ICON as described in Section 4.1, denoting as " + }, + { + "bbox": [ + 46, + 594, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{PIFu}_{clo}" + }, + { + "bbox": [ + 46, + 594, + 287, + 713 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 594, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{PaMIR}_{clo}" + }, + { + "bbox": [ + 46, + 594, + 287, + 713 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 594, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{ICON}_{clo}" + }, + { + "bbox": [ + 46, + 594, + 287, + 713 + ], + "type": "text", + "content": ", respectively. We report the results of models trained on CLOTH4D in the gray columns in Table 2 and" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 306, + 557, + 495, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 557, + 495, + 568 + ], + "spans": [ + { + "bbox": [ + 306, + 557, + 495, + 568 + ], + "type": "text", + "content": "Figure 5. The following findings can be made:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 569, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 304, + 569, + 545, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 569, + 545, + 652 + ], + "spans": [ + { + "bbox": [ + 304, + 569, + 545, + 652 + ], + "type": "text", + "content": "1) In terms of quantitative metrics, the models trained on CLOTH4D generally outperform the original models trained on scan datasets as the data distributions of the training and testing datasets are closer. " + }, + { + "bbox": [ + 304, + 569, + 545, + 652 + ], + "type": "inline_equation", + "content": "\\mathrm{ICON}_{clo}^{gt}" + }, + { + "bbox": [ + 304, + 569, + 545, + 652 + ], + "type": "text", + "content": " achieves the highest accuracy on the CLOTH4D test set. Furthermore, the qualitative results show that more high-frequency details are generated after training on CLOTH4D." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "type": "text", + "content": "2) As the re-implementation setting in ICON's code allows the normal image to be input into all methods, the influence of garment print to PIFu and PaMIR are slightly relieved, validating the hypothesis we made in Section 4.3 that predicting intermediate representations like normals can re" + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "12853" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 244, + 84 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 244, + 84 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 244, + 84 + ], + "type": "text", + "content": "duce the ambiguity of geometry and appearance." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 85, + 288, + 361 + ], + "type": "list", + "angle": 0, + "index": 3, + "blocks": [ + { + "bbox": [ + 46, + 85, + 287, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 85, + 287, + 240 + ], + "spans": [ + { + "bbox": [ + 46, + 85, + 287, + 240 + ], + "type": "text", + "content": "3) SOTAs fail to model layered and thin clothing structures as shown in the dress and skirt regions in Figure 5, where holes and tattered pieces are generated. Notably, the original PIFu and PaMIR can roughly generate the overall shapes of loose clothing, but they fail when trained on CLOTH4D. Since SOTA methods learn an occupancy field by sampling query points in the 3D space, for PIFu, whose spatial feature resolution is low, and for PaMIR, whose volumetric feature space is also of low resolution, it is hard to sample informative points near the thin surface. It is even harder for the network to learn if a querying point is inside or outside the mesh near the thin structure, as the inside and outside samples have very similar local features." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 241, + 288, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 241, + 288, + 361 + ], + "spans": [ + { + "bbox": [ + 46, + 241, + 288, + 361 + ], + "type": "text", + "content": "4) The difficulties of learning the occupancy field for thin faces motivates developing methods that focus more on the query points near the thin surface for better reconstruction. Future research may also seek better implicit representations to boost the performance of reconstructing multi-layer thin structures e.g., [12, 13]. However, as shown in the supplementary material (Figure B), the state-of-the-art implicit representation cannot achieve satisfactory multi-layer thin structure reconstruction even if we feed the ground truth mesh as the input to the implicit function." + } + ] + } + ], + "index": 2 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 46, + 361, + 287, + 566 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 361, + 287, + 566 + ], + "spans": [ + { + "bbox": [ + 46, + 361, + 287, + 566 + ], + "type": "text", + "content": "As shown in Table 3 and Figure 6, the performance on the CAPE dataset drops after training on CLOTH4D due to two reasons. 1) The CAPE dataset is collected in a controlled lab environment with dim lighting, which further enlarges the domain gap between CAPE and CLOTH4D. Consequently, the predicted normal images and reconstructed meshes are less accurate. 2) CAPE dataset contains single-layer scan meshes and tight clothing. However, models trained on the multi-layer CLOTH4D dataset tend to generate gaps between the clothing and skin, resulting in holes in the clothing and unsmooth surfaces. This also verifies that local features are sensitive to overfitting. It remains a challenging but interesting problem to have a unified representation of single-layer and multi-layer data, thus different types of datasets could be trained together to yield better generalizability. Finally, we show the reconstructed samples of the in-the-wild images in Figure 7." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 574, + 125, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 574, + 125, + 586 + ], + "spans": [ + { + "bbox": [ + 47, + 574, + 125, + 586 + ], + "type": "text", + "content": "4.5. Limitations" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 594, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 288, + 715 + ], + "type": "text", + "content": "Firstly, the simulation results are generated via graphics software, which may crash in some cases. We show some detailed examples in the supplementary material (Figure C). Meanwhile, the results on CAPE show that the clothing features of current CLOTH4D can well represent basic men's wear. However, considering the completeness, the dataset scale and diversity of subjects will be further improved to cover males and kids. In addition, the original simulated mesh sequences are non-watertight, which must be converted to watertight (and such conversion is usually lossy)" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 307, + 68, + 547, + 266 + ], + "blocks": [ + { + "bbox": [ + 307, + 68, + 547, + 266 + ], + "lines": [ + { + "bbox": [ + 307, + 68, + 547, + 266 + ], + "spans": [ + { + "bbox": [ + 307, + 68, + 547, + 266 + ], + "type": "image", + "image_path": "935dbe458a4ad764c608e737541a3c291f4300f0bca0ca5d20cc9762cedf07be.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 307, + 298, + 545, + 396 + ], + "blocks": [ + { + "bbox": [ + 307, + 298, + 545, + 396 + ], + "lines": [ + { + "bbox": [ + 307, + 298, + 545, + 396 + ], + "spans": [ + { + "bbox": [ + 307, + 298, + 545, + 396 + ], + "type": "image", + "image_path": "74a8c03af2716bb742b5369fb13107d2b3e7ae6e416f976470e6eb90d8657c01.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 306, + 396, + 545, + 418 + ], + "lines": [ + { + "bbox": [ + 306, + 396, + 545, + 418 + ], + "spans": [ + { + "bbox": [ + 306, + 396, + 545, + 418 + ], + "type": "text", + "content": "Figure 7. In-the-wild front and side view reconstructions. SOTAs with implicit functions tend to generate broken results." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 423, + 547, + 484 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 423, + 547, + 484 + ], + "spans": [ + { + "bbox": [ + 305, + 423, + 547, + 484 + ], + "type": "text", + "content": "before feeding into current clothed human reconstruction methods. It would be essential to conduct future research that reconstructs meshes with arbitrary topologies. Finally, as indicated by the results on CAPE, the generalization ability to scan data is yet to be further explored." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 498, + 379, + 511 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 498, + 379, + 511 + ], + "spans": [ + { + "bbox": [ + 306, + 498, + 379, + 511 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 520, + 545, + 663 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 520, + 545, + 663 + ], + "spans": [ + { + "bbox": [ + 304, + 520, + 545, + 663 + ], + "type": "text", + "content": "We introduce CLOTH4D containing realistic and rich categories of clothing, avatars, and animations, and will release it for free, hoping to push the research on clothed human reconstruction. We evaluate current SOTAs with the newly introduced temporally-aware metrics and in-depth analyze their pros and cons by leveraging the advantages of CLOTH4D. We retrain those SOTAs on CLOTH4D, discuss the challenges the new dataset brings, and propose potential research directions. Although layering clothing in CLOTH4D brings immense difficulties to current research, we believe it is an important step toward more realistic and temporally coherent clothed human reconstruction." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 665, + 547, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 547, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 547, + 712 + ], + "type": "text", + "content": "Acknowledgement: This work is supported by Laboratory for Artificial Intelligence in Design (Project Code: RP 3-1) under InnoHK Research Clusters, Hong Kong SAR Government." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "12854" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 286, + 712 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 53, + 91, + 215, + 102 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 215, + 102 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 215, + 102 + ], + "type": "text", + "content": "[1] 3dpeople. https://www.3dpeople.com. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 102, + 277, + 113 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 102, + 277, + 113 + ], + "spans": [ + { + "bbox": [ + 53, + 102, + 277, + 113 + ], + "type": "text", + "content": "[2] Adobe system incorporated. https://www MIXamo.com. 4" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 114, + 248, + 125 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 114, + 248, + 125 + ], + "spans": [ + { + "bbox": [ + 53, + 114, + 248, + 125 + ], + "type": "text", + "content": "[3] Clo virtual fashion llc. https://www.clo3d.com. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 126, + 286, + 145 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 126, + 286, + 145 + ], + "spans": [ + { + "bbox": [ + 53, + 126, + 286, + 145 + ], + "type": "text", + "content": "[4] Official code of icon. https://github.com/YuliangXiu/ICON. 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 148, + 286, + 191 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 148, + 286, + 191 + ], + "spans": [ + { + "bbox": [ + 53, + 148, + 286, + 191 + ], + "type": "text", + "content": "[5] Thiemo Alldieck, Marcus Magnor, Weipeng Xu, Christian Theobalt, and Gerard Pons-Moll. Detailed human avatars from monocular video. In 2018 International Conference on 3D Vision (3DV), pages 98-109. IEEE, 2018. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 193, + 286, + 247 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 193, + 286, + 247 + ], + "spans": [ + { + "bbox": [ + 53, + 193, + 286, + 247 + ], + "type": "text", + "content": "[6] Thiemo Alldieck, Gerard Pons-Moll, Christian Theobalt, and Marcus Magnor. Tex2shape: Detailed full human body geometry from a single image. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2293-2303, 2019. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 249, + 286, + 302 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 249, + 286, + 302 + ], + "spans": [ + { + "bbox": [ + 53, + 249, + 286, + 302 + ], + "type": "text", + "content": "[7] Thiemo Alldieck, Mihai Zanfir, and Cristian Sminchisescu. Photorealistic monocular 3d reconstruction of humans wearing clothing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1506-1515, 2022. 3, 4, 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 304, + 286, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 304, + 286, + 335 + ], + "spans": [ + { + "bbox": [ + 53, + 304, + 286, + 335 + ], + "type": "text", + "content": "[8] Hugo Bertiche, Meysam Madadi, and Sergio Escalera. Cloth3d: clothed 3d humans. In European Conference on Computer Vision, pages 344-359. Springer, 2020. 2, 3, 4" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 337, + 286, + 390 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 337, + 286, + 390 + ], + "spans": [ + { + "bbox": [ + 53, + 337, + 286, + 390 + ], + "type": "text", + "content": "[9] Bharat Lal Bhatnagar, Garvita Tiwari, Christian Theobalt, and Gerard Pons-Moll. Multi-garment net: Learning to dress 3d people from images. In proceedings of the IEEE/CVF international conference on computer vision, pages 5420-5430, 2019. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 392, + 286, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 392, + 286, + 445 + ], + "spans": [ + { + "bbox": [ + 48, + 392, + 286, + 445 + ], + "type": "text", + "content": "[10] Zhongang Cai, Daxuan Ren, Ailing Zeng, Zhengyu Lin, Tao Yu, Wenjia Wang, Xiangyu Fan, Yang Gao, Yifan Yu, Liang Pan, et al. Humman: Multi-modal 4d human dataset for versatile sensing and modeling. arXiv preprint arXiv:2204.13686, 2022. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 448, + 286, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 448, + 286, + 491 + ], + "spans": [ + { + "bbox": [ + 48, + 448, + 286, + 491 + ], + "type": "text", + "content": "[11] Z. Cao, G. Hidalgo Martinez, T. Simon, S. Wei, and Y. A. Sheikh. Openpose: Realtime multi-person 2d pose estimation using part affinity fields. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2019. 4" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 492, + 286, + 546 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 492, + 286, + 546 + ], + "spans": [ + { + "bbox": [ + 48, + 492, + 286, + 546 + ], + "type": "text", + "content": "[12] Weikai Chen, Cheng Lin, Weiyang Li, and Bo Yang. 3psdf: Three-pole signed distance function for learning surfaces with arbitrary topologies. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18522-18531, 2022. 8" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 548, + 286, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 548, + 286, + 590 + ], + "spans": [ + { + "bbox": [ + 48, + 548, + 286, + 590 + ], + "type": "text", + "content": "[13] Julian Chibane, Gerard Pons-Moll, et al. Neural unsigned distance fields for implicit function learning. Advances in Neural Information Processing Systems, 33:21638-21652, 2020. 8" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 592, + 286, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 592, + 286, + 624 + ], + "spans": [ + { + "bbox": [ + 48, + 592, + 286, + 624 + ], + "type": "text", + "content": "[14] Mikhail Erofeev, Yury Gitman, Dmitriy S Vatolin, Alexey Fedorov, and Jue Wang. Perceptually motivated benchmark for video matting. In BMVC, pages 99-1, 2015. 4, 5" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 625, + 286, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 625, + 286, + 668 + ], + "spans": [ + { + "bbox": [ + 48, + 625, + 286, + 668 + ], + "type": "text", + "content": "[15] Qiao Feng, Yebin Liu, Yu-Kun Lai, Jingyu Yang, and Kun Li. Fof: Learning fourier occupancy field for monocular real-time human reconstruction. arXiv preprint arXiv:2206.02194, 2022. 3, 4, 5" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 48, + 670, + 286, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 670, + 286, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 670, + 286, + 712 + ], + "type": "text", + "content": "[16] Peng Guan, Loretta Reiss, David A Hirshberg, Alexander Weiss, and Michael J Black. Drape: Dressing any person. ACM Transactions on Graphics (ToG), 31(4):1-10, 2012. 2, 3" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 128 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 128 + ], + "type": "text", + "content": "[17] Erhan Gundogdu, Victor Constantin, Amrollah Seifoddini, Minh Dang, Mathieu Salzmann, and Pascal Fua. Garnet: A two-stream network for fast and accurate 3d cloth draping. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8739-8748, 2019. 2, 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 130, + 545, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 130, + 545, + 183 + ], + "spans": [ + { + "bbox": [ + 307, + 130, + 545, + 183 + ], + "type": "text", + "content": "[18] Xintong Han, Xiaojun Hu, Weilin Huang, and Matthew R Scott. Clothflow: A flow-based model for clothed person generation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10471-10480, 2019. 4" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 186, + 545, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 186, + 545, + 229 + ], + "spans": [ + { + "bbox": [ + 307, + 186, + 545, + 229 + ], + "type": "text", + "content": "[19] Xintong Han, Zuxuan Wu, Zhe Wu, Ruichi Yu, and Larry S Davis. Viton: An image-based virtual try-on network. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7543-7552, 2018. 4" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 231, + 545, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 231, + 545, + 274 + ], + "spans": [ + { + "bbox": [ + 307, + 231, + 545, + 274 + ], + "type": "text", + "content": "[20] Tong He, John Collomosse, Hailin Jin, and Stefano Soatto. Geo-pifu: Geometry and pixel aligned implicit functions for single-view human reconstruction. Advances in Neural Information Processing Systems, 33:9276-9287, 2020. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 276, + 545, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 276, + 545, + 330 + ], + "spans": [ + { + "bbox": [ + 307, + 276, + 545, + 330 + ], + "type": "text", + "content": "[21] Tong He, Yuanlu Xu, Shunsuke Saito, Stefano Soatto, and Tony Tung. Arch++: Animation-ready clothed human reconstruction revisited. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11046-11056, 2021. 3, 4" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 332, + 545, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 332, + 545, + 398 + ], + "spans": [ + { + "bbox": [ + 307, + 332, + 545, + 398 + ], + "type": "text", + "content": "[22] Zhu Heming, Cao Yu, Jin Hang, Chen Weikai, Du Dong, Wang Zhangye, Cui Shuguang, and Han Xiaoguang. Deep fashion3d: A dataset and benchmark for 3d garment reconstruction from single images. In Computer Vision - ECCV 2020, pages 512-530. Springer International Publishing, 2020. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 399, + 545, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 399, + 545, + 442 + ], + "spans": [ + { + "bbox": [ + 307, + 399, + 545, + 442 + ], + "type": "text", + "content": "[23] Jingwei Huang, Yichao Zhou, and Leonidas Guibas. Manifoldplus: A robust and scalable watertight manifold surface generation method for triangle soups. arXiv preprint arXiv:2005.11621, 2020.4" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 445, + 545, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 445, + 545, + 498 + ], + "spans": [ + { + "bbox": [ + 307, + 445, + 545, + 498 + ], + "type": "text", + "content": "[24] Zeng Huang, Yuanlu Xu, Christoph Lassner, Hao Li, and Tony Tung. Arch: Animatable reconstruction of clothed humans. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3093-3102, 2020. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 501, + 545, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 501, + 545, + 555 + ], + "spans": [ + { + "bbox": [ + 307, + 501, + 545, + 555 + ], + "type": "text", + "content": "[25] Catalin Ionescu, Dragos Papava, Vlad Olaru, and Cristian Sminchisescu. Human3.6m: Large scale datasets and predictive methods for 3d human sensing in natural environments. IEEE transactions on pattern analysis and machine intelligence, 36(7):1325-1339, 2013. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 557, + 545, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 557, + 545, + 601 + ], + "spans": [ + { + "bbox": [ + 307, + 557, + 545, + 601 + ], + "type": "text", + "content": "[26] Boyi Jiang, Juyong Zhang, Yang Hong, Jinhao Luo, Ligang Liu, and Hujun Bao. BCnet: Learning body and cloth shape from a single image. In European Conference on Computer Vision, pages 18-35. Springer, 2020. 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 602, + 545, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 602, + 545, + 656 + ], + "spans": [ + { + "bbox": [ + 307, + 602, + 545, + 656 + ], + "type": "text", + "content": "[27] Muhammed Kocabas, Nikos Athanasiou, and Michael J Black. Vibe: Video inference for human body pose and shape estimation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5253-5263, 2020. 6" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 658, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 658, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 658, + 545, + 713 + ], + "type": "text", + "content": "[28] Nikos Kolotouros, Georgios Pavlakos, Michael J Black, and Kostas Daniilidis. Learning to reconstruct 3d human pose and shape via model-fitting in the loop. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2252-2261, 2019. 3" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "12855" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "type": "text", + "content": "[29] Nikos Kolotouros, Georgios Pavlakos, and Kostas Dani-ilidis. Convolutional mesh regression for single-image human shape reconstruction. In CVPR, 2019. 4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 107, + 287, + 152 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 107, + 287, + 152 + ], + "spans": [ + { + "bbox": [ + 48, + 107, + 287, + 152 + ], + "type": "text", + "content": "[30] Zorah Lahner, Daniel Cremers, and Tony Tung. Deepwrinkles: Accurate and realistic clothing modeling. In Proceedings of the European conference on computer vision (ECCV), pages 667-684, 2018. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 152, + 287, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 152, + 287, + 196 + ], + "spans": [ + { + "bbox": [ + 48, + 152, + 287, + 196 + ], + "type": "text", + "content": "[31] Verica Lazova, Eldar Insafutdinov, and Gerard Pons-Moll. 360-degree textures of people in clothing from a single image. In 2019 International Conference on 3D Vision (3DV), pages 643-653. IEEE, 2019. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 198, + 287, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 198, + 287, + 251 + ], + "spans": [ + { + "bbox": [ + 48, + 198, + 287, + 251 + ], + "type": "text", + "content": "[32] Shanchuan Lin, Linjie Yang, Imran Saleemi, and Soumyadip Sengupta. Robust high-resolution video matting with temporal guidance. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 238-247, 2022. 4, 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 254, + 287, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 254, + 287, + 297 + ], + "spans": [ + { + "bbox": [ + 48, + 254, + 287, + 297 + ], + "type": "text", + "content": "[33] Matthew Loper, Naureen Mahmood, Javier Romero, Gerard Pons-Moll, and Michael J Black. SMPL: A skinned multiperson linear model. ACM transactions on graphics (TOG), 34(6):1-16, 2015. 2, 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 299, + 287, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 299, + 287, + 331 + ], + "spans": [ + { + "bbox": [ + 48, + 299, + 287, + 331 + ], + "type": "text", + "content": "[34] William E Lorensen and Harvey E Cline. Marching cubes: A high resolution 3d surface construction algorithm. ACM siggraph computer graphics, 21(4):163-169, 1987. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 333, + 287, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 333, + 287, + 387 + ], + "spans": [ + { + "bbox": [ + 48, + 333, + 287, + 387 + ], + "type": "text", + "content": "[35] Qianli Ma, Jinlong Yang, Anurag Ranjan, Sergi Pujades, Gerard Pons-Moll, Siyu Tang, and Michael J Black. Learning to dress 3d people in generative clothing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6469-6478, 2020. 2, 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 388, + 287, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 388, + 287, + 443 + ], + "spans": [ + { + "bbox": [ + 48, + 388, + 287, + 443 + ], + "type": "text", + "content": "[36] Meysam Madadi, Hugo Bertiche, Wafa Bouzouita, Isabelle Guyon, and Sergio Escalera. Learning cloth dynamics: 3d + texture garment reconstruction benchmark. In Proceedings of the NeurIPS 2020 Competition and Demonstration Track, PMLR, volume 133, pages 57-76, 2021. 2, 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 445, + 287, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 445, + 287, + 487 + ], + "spans": [ + { + "bbox": [ + 48, + 445, + 287, + 487 + ], + "type": "text", + "content": "[37] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In ECCV, 2020. 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 490, + 287, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 490, + 287, + 544 + ], + "spans": [ + { + "bbox": [ + 48, + 490, + 287, + 544 + ], + "type": "text", + "content": "[38] Chaitanya Patel, Zhouyingcheng Liao, and Gerard Pons-Moll. Tailornet: Predicting clothing in 3d as a function of human pose, shape and garment style. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR). IEEE, jun 2020. 2, 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 545, + 287, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 545, + 287, + 588 + ], + "spans": [ + { + "bbox": [ + 48, + 545, + 287, + 588 + ], + "type": "text", + "content": "[39] Gerard Pons-Moll, Sergi Pujades, Sonny Hu, and Michael J Black. Clothcap: Seamless 4d clothing capture and retargeting. ACM Transactions on Graphics (ToG), 36(4):1-15, 2017. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 590, + 287, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 590, + 287, + 635 + ], + "spans": [ + { + "bbox": [ + 48, + 590, + 287, + 635 + ], + "type": "text", + "content": "[40] Albert Pumarola, Jordi Sanchez, G. Choi, A. Sanfeliu, and F. Moreno-Noguer. 3dpeople: Modeling the geometry of dressed humans. 2019 IEEE/CVF International Conference on Computer Vision (ICCV), pages 2242-2251, 2019. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 636, + 287, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 636, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 636, + 287, + 689 + ], + "type": "text", + "content": "[41] Shunsuke Saito, Zeng Huang, Ryota Natsume, Shigeo Morishima, Angjoo Kanazawa, and Hao Li. Pifu: Pixel-aligned implicit function for high-resolution clothed human digitization. In Proceedings of the IEEE International Conference on Computer Vision, pages 2304-2314, 2019. 3, 4" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 692, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 692, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 692, + 287, + 713 + ], + "type": "text", + "content": "[42] Shunsuke Saito, Tomas Simon, Jason Saragih, and Hanbyul Joo. Pifuhd: Multi-level pixel-aligned implicit function for" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 546, + 713 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 327, + 73, + 546, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 546, + 106 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 546, + 106 + ], + "type": "text", + "content": "high-resolution 3d human digitization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 84-93, 2020. 3, 4" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 107, + 546, + 152 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 107, + 546, + 152 + ], + "spans": [ + { + "bbox": [ + 308, + 107, + 546, + 152 + ], + "type": "text", + "content": "[43] Igor Santesteban, Miguel A Otaduy, and Dan Casas. Learning-based animation of clothing for virtual try-on. In Computer Graphics Forum, volume 38, pages 355-366. Wiley Online Library, 2019. 2, 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 153, + 546, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 153, + 546, + 206 + ], + "spans": [ + { + "bbox": [ + 308, + 153, + 546, + 206 + ], + "type": "text", + "content": "[44] Leonid Sigal, Alexandru O Balan, and Michael J Black. Humaneva: Synchronized video and motion capture dataset and baseline algorithm for evaluation of articulated human motion. International journal of computer vision, 87(1):4-27, 2010. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 209, + 546, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 209, + 546, + 262 + ], + "spans": [ + { + "bbox": [ + 308, + 209, + 546, + 262 + ], + "type": "text", + "content": "[45] Peter-Pike Sloan, Jan Kautz, and John Snyder. Precomputed radiance transfer for real-time rendering in dynamic, low-frequency lighting environments. In Proceedings of the 29th annual conference on Computer graphics and interactive techniques, pages 527-536, 2002. 4" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 264, + 546, + 307 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 264, + 546, + 307 + ], + "spans": [ + { + "bbox": [ + 308, + 264, + 546, + 307 + ], + "type": "text", + "content": "[46] Matthew Trumble, Andrew Gilbert, Charles Malleson, A. Hilton, and J. Collomosse. Total capture: 3d human pose estimation fusing video and inertial sensors. In BMVC, 2017. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 310, + 546, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 310, + 546, + 352 + ], + "spans": [ + { + "bbox": [ + 308, + 310, + 546, + 352 + ], + "type": "text", + "content": "[47] Tuanfeng Y Wang, Duygu Ceylan, Jovan Popovic, and Niloy J Mitra. Learning a shared shape space for multimodal garment design. arXiv preprint arXiv:1806.11335, 2018. 2, 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 354, + 546, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 354, + 546, + 408 + ], + "spans": [ + { + "bbox": [ + 308, + 354, + 546, + 408 + ], + "type": "text", + "content": "[48] Donglai Xiang, Fabian Prada, Chenglei Wu, and Jessica Hodgins. Monoclothcap: Towards temporally coherent clothing capture from monocular rgb video. In 2020 International Conference on 3D Vision (3DV), pages 322-332. IEEE, 2020. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 411, + 546, + 465 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 411, + 546, + 465 + ], + "spans": [ + { + "bbox": [ + 308, + 411, + 546, + 465 + ], + "type": "text", + "content": "[49] Yuliang Xiu, Jinlong Yang, Dimitrios Tzionas, and Michael J Black. Icon: Implicit clothed humans obtained from normals. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 13286-13296. IEEE, 2022. 3, 4, 5, 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 468, + 546, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 468, + 546, + 510 + ], + "spans": [ + { + "bbox": [ + 308, + 468, + 546, + 510 + ], + "type": "text", + "content": "[50] Shan Yang, Tanya Ambert, Zherong Pan, Ke Wang, Licheng Yu, Tamara Berg, and Ming C Lin. Detailed garment recovery from a single-view image. arXiv preprint arXiv:1608.01250, 2016. 4" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 513, + 546, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 513, + 546, + 567 + ], + "spans": [ + { + "bbox": [ + 308, + 513, + 546, + 567 + ], + "type": "text", + "content": "[51] Tao Yu, Zerong Zheng, Kaiwen Guo, Pengpeng Liu, Qionghai Dai, and Yebin Liu. Function4d: Real-time human volumetric capture from very sparse consumer rgbd sensors. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR2021), June 2021. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 568, + 546, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 568, + 546, + 622 + ], + "spans": [ + { + "bbox": [ + 308, + 568, + 546, + 622 + ], + "type": "text", + "content": "[52] Chao Zhang, Sergi Pujades, Michael J Black, and Gerard Pons-Moll. Detailed, accurate, human shape estimation from clothed 3d scan sequences. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 4191-4200, 2017. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 624, + 546, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 624, + 546, + 678 + ], + "spans": [ + { + "bbox": [ + 308, + 624, + 546, + 678 + ], + "type": "text", + "content": "[53] Hongwen Zhang, Yating Tian, Xinchi Zhou, Wanli Ouyang, Yebin Liu, Limin Wang, and Zhenan Sun. Pymaf: 3d human pose and shape regression with pyramidal mesh alignment feedback loop. In Proceedings of the IEEE International Conference on Computer Vision, 2021. 4" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 681, + 546, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 681, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 681, + 546, + 713 + ], + "type": "text", + "content": "[54] Zerong Zheng, Tao Yu, Yebin Liu, and Qionghai Dai. Pamir: Parametric model-conditioned implicit representation for image-based human reconstruction. IEEE transactions on" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "text", + "content": "12856" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 228 + ], + "type": "list", + "angle": 0, + "index": 4, + "blocks": [ + { + "bbox": [ + 67, + 72, + 286, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 72, + 286, + 94 + ], + "spans": [ + { + "bbox": [ + 67, + 72, + 286, + 94 + ], + "type": "text", + "content": "pattern analysis and machine intelligence, 44(6):3170-3184, 2021. 3, 4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 96, + 287, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 96, + 287, + 140 + ], + "spans": [ + { + "bbox": [ + 48, + 96, + 287, + 140 + ], + "type": "text", + "content": "[55] Boyao Zhou, Jean-Sébastien Franco, Federica Bogo, and Edmond Boyer. Spatio-temporal human shape completion with implicit function networks. In 2021 International Conference on 3D Vision (3DV), pages 669-678. IEEE, 2021. 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 141, + 287, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 141, + 287, + 196 + ], + "spans": [ + { + "bbox": [ + 48, + 141, + 287, + 196 + ], + "type": "text", + "content": "[56] Xingxing Zou, Xiangheng Kong, Waikeung Wong, Congde Wang, Yuguang Liu, and Yang Cao. Fashionai: A hierarchical dataset for fashion understanding. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops, pages 0-0, 2019. 4" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 198, + 286, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 198, + 286, + 228 + ], + "spans": [ + { + "bbox": [ + 48, + 198, + 286, + 228 + ], + "type": "text", + "content": "[57] Xingxing Zou and Waikeung Wong. fashion after fashion: A report of ai in fashion. arXiv preprint arXiv:2105.03050, 2021.4" + } + ] + } + ], + "index": 3 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "12857" + } + ] + } + ], + "index": 5 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/CNVid-3.5M_ Build, Filter, and Pre-Train the Large-Scale Public Chinese Video-Text Dataset/3819ec34-d037-4584-a0be-1f25bc0cb37d_content_list.json b/2023/CNVid-3.5M_ Build, Filter, and Pre-Train the Large-Scale Public Chinese Video-Text Dataset/3819ec34-d037-4584-a0be-1f25bc0cb37d_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..9fbe877bfc9de1877e9b959d0666b22e560bfd02 --- /dev/null +++ b/2023/CNVid-3.5M_ Build, Filter, and Pre-Train the Large-Scale Public Chinese Video-Text Dataset/3819ec34-d037-4584-a0be-1f25bc0cb37d_content_list.json @@ -0,0 +1,1892 @@ +[ + { + "type": "text", + "text": "CNVid-3.5M: Build, Filter, and Pre-train the Large-scale Public Chinese Video-text Dataset", + "text_level": 1, + "bbox": [ + 233, + 130, + 736, + 174 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Tian Gan $^{1*}$ , Qing Wang $^{2*}$ , Xingning Dong $^{1}$ , Xiangyuan Ren $^{2}$ , Liqiang Nie $^{3}$ , Qingpei Guo $^{2\\dagger}$", + "bbox": [ + 106, + 202, + 908, + 222 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1Shandong University,", + "bbox": [ + 200, + 224, + 351, + 241 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ Ant Group,", + "bbox": [ + 382, + 224, + 468, + 241 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{3}$ Harbin Institute of Technology (Shenzhen)", + "bbox": [ + 500, + 224, + 790, + 241 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "gantian@sdu.edu.cn,", + "bbox": [ + 181, + 244, + 349, + 258 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "wq176625@antgroup.com,", + "bbox": [ + 372, + 244, + 568, + 258 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "dongxingning1998@gmail.com", + "bbox": [ + 589, + 244, + 823, + 258 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "xiangyuan.rxy@antgroup.com, nieliqiang@gmail.com, qingpei.gqp@antgroup.com", + "bbox": [ + 148, + 262, + 834, + 276 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 311, + 313, + 328 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Owing to well-designed large-scale video-text datasets, recent years have witnessed tremendous progress in videotext pre-training. However, existing large-scale video-text datasets are mostly English-only. Though there are certain methods studying the Chinese video-text pre-training, they pre-train their models on private datasets whose videos and text are unavailable. This lack of large-scale public datasets and benchmarks in Chinese hampers the research and downstream applications of Chinese video-text pretraining. Towards this end, we release and benchmark CNVid-3.5M, a large-scale public cross-modal dataset containing over 3.5M Chinese video-text pairs. We summarize our contributions by three verbs, i.e., \"Build\", \"Filter\", and \"Pre-train\": 1) To build a public Chinese video-text dataset, we collect over 4.5M videos from the Chinese websites. 2) To improve the data quality, we propose a novel method to filter out 1M weakly-paired videos, resulting in the CNVid-3.5M dataset. And 3) we benchmark CNVid-3.5M with three mainstream pixel-level pre-training architectures. At last, we propose the Hard Sample Curriculum Learning strategy to promote the pre-training performance. To the best of our knowledge, CNVid-3.5M is the largest public video-text dataset in Chinese, and we provide the first pixel-level benchmarks for Chinese video-text pretraining. The dataset, codebase, and pre-trained models are available at https://github.com/CNVid/CNVid-3.5M.", + "bbox": [ + 76, + 343, + 472, + 736 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 763, + 209, + 780 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Owing to well-designed large-scale datasets, video-text pre-training [15, 17, 19] has achieved superior performance in various downstream tasks, such as video-text retrieval [4, 10, 36], video question answering [27, 34, 42], and video captioning [1, 22, 30]. However, recent large-scale video-", + "bbox": [ + 75, + 789, + 468, + 866 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/9b7a8bdb48e124ef90a9d5b532991186db25efb56d5530433476257688e18c7f.jpg", + "image_caption": [ + "Figure 1. Here presents the motivations of this paper, based on which we highly summarize our contributions with three verbs: \"Build\", \"Filter\", and \"Pre-train\"." + ], + "image_footnote": [], + "bbox": [ + 504, + 310, + 630, + 414 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/a4d4cffe965fc815b477273e86bb78783bbe166255e5abaa4f22807664ae1cd0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 638, + 314, + 771, + 412 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/b46894051c7f630db7c77289f2833c7996d7b3725c9d2e194d6e2cb1d0955b60.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 776, + 314, + 900, + 414 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "text datasets are mostly English-only (e.g., Howto100M [25] and WebVid-2.5M [4]). Though some methods [14,26, 45] turn to study the Chinese video-text pre-training, they pre-train their models on private datasets whose videos and text are unavailable. Therefore, the research towards Chinese video-text pre-training is still in its infancy due to the lack of large-scale public datasets.", + "bbox": [ + 496, + 484, + 892, + 590 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Towards this problem, directly translating English text into Chinese is a simple solution. However, it may result in unacceptable performance degradation for two reasons: 1) Translation errors are inevitable. Moreover, since most of the large-scale video-text datasets employ the Automatic Speech Recognition (ASR) system to generate text, the language translator would amplify the error from the incomplete and noisy ASR text. And 2) there remains an intrinsic linguistic gap between English and Chinese. Many widely-used English idioms and slang can hardly find their Chinese counterparts, leading some translated text incomprehensible and even contrary to the original meaning.", + "bbox": [ + 496, + 594, + 893, + 775 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this paper, we aim to release and benchmark a large-scale public Chinese video-text dataset to facilitate future researchers and the community. As illustrated in Figure 1, three verbs could highly summarize our contributions, i.e., \"Build\", \"Filter\", and \"Pre-train\".", + "bbox": [ + 496, + 777, + 893, + 852 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To build a large-scale Chinese video-text dataset, we collect over 4.5M videos from Chinese websites. All videos are associated with user-uploaded titles and ASR text.", + "bbox": [ + 496, + 854, + 892, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Equal contribution.", + "bbox": [ + 94, + 875, + 207, + 887 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "† Corresponding author.", + "bbox": [ + 96, + 887, + 220, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "14815", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We filter out the weakly-paired data by a novel method to improve the data quality. As some work [25,26] pointed out, the pre-training performance would suffer from the noisy ASR text that fails to accurately describe the video content. Unfortunately, the problem is raised with few practical solutions. Therefore, we employ a well-trained image-text model to evaluate the video-text consistency for three reasons: 1) The text information in existing image-text datasets [11] are usually manually-written titles or captions, whose consistency is guaranteed. 2) Some video-text pre-training architectures [23,35] are based upon image-text ones. And 3) it is cheap and efficient to \"hire\" a well-trained model to check millions of videos. In this way, we filter out about 1M weakly-paired videos based on the balance between the pre-training performance and efficiency, deriving the proposed CNVid-3.5M dataset.", + "bbox": [ + 75, + 90, + 472, + 332 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We pre-train various models to benchmark our CNVid-3.5M dataset. Current video-text pre-training methods could be roughly divided into two categories: 1) feature-level pre-training methods [24, 33, 40] that employ offline video and textual feature extractors, and 2) pixel-level ones [4, 15, 36] that learn cross-modal representations end-to-end from raw videos and text. Since there remain domain gaps between pre-training datasets and frozen feature extractors, pixel-level pre-training methods usually achieve better performance and have been widely employed in recent years. However, existing Chinese video-text pre-training methods [14, 26, 45] are all feature-level ones pre-trained on private datasets, limiting their contributions on the development of Chinese video-text pre-training techniques. Hence, we adopt three mainstream pixel-level pre-training frameworks, which are the first pixel-level benchmarks for Chinese video-text pre-training.", + "bbox": [ + 75, + 333, + 472, + 589 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Moreover, we propose the novel Hard Sample Curriculum Learning strategy to promote the pre-training performance. Since contrastive learning is a significant component in video-text pre-training, some methods [16, 18, 43] employ the hard sample mining [12, 29] strategy to promote the cross-modal alignment. However, hard sample mining would bring side effects to pre-training when the model is far from convergence. Suppose that a model is incapable of discriminating the ground-truth video-text pairs, recklessly introducing hard negatives would lead to the sub-optimal performance. Inspired by the curriculum learning [32, 37] strategy that \"starts small\" and gradually \"learns hard\", we combine these two strategies and propose the novel Hard Sample Curriculum Learning (HSCL). By gradually and smoothly emphasizing those hard samples, HSCL could effectively improve the pre-training performance.", + "bbox": [ + 75, + 590, + 472, + 832 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our contributions are summarized in four folds:", + "bbox": [ + 96, + 833, + 413, + 847 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "- To fill in the blank of large-scale public Chinese video-text datasets, we collect over 4.5M videos associated with titles and ASR text from the websites.", + "bbox": [ + 94, + 854, + 468, + 900 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- To improve the data quality, we propose a novel method to filter out 1M weakly-paired videos, resulting in the CNVid-3.5M dataset.", + "- To promote the pre-training performance, we propose the novel Hard Sample Curriculum Learning strategy for better cross-modal contrastive learning.", + "- To the best of our knowledge, the constructed CNVid-3.5M is the largest public Chinese video-text dataset. Moreover, we provide the first Chinese pixel-level benchmarks based on CNVid-3.5M. The dataset, codebase, and benchmarks are available at https://github.com/CNVid/CNVid-3.5M." + ], + "bbox": [ + 517, + 90, + 893, + 272 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 296, + 640, + 313 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Video-Text Pre-training and Fine-Tuning. Video-text pre-training has attracted increasing interest in recent years, which could be roughly divided into two categories, feature-level pre-training methods [24, 33, 40] and pixel-level ones [15, 17, 35]. The former approaches employ offline visual and textual features extracted from frozen models, while the latter ones learn cross-modal representations from raw videos and text in an end-to-end manner. Since there remain domain gaps between pre-training datasets and frozen feature extractors, pixel-level pre-training methods usually achieve better performance than the feature-level ones.", + "bbox": [ + 496, + 321, + 890, + 488 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "There are three mainstream architectures in pixel-level video-text pre-training methods, i.e., three-encoders-fusion (3-E-F) [15, 17, 30], two-encoders-fusion (2-E-F) [9], and twin-towers-crossed (T-T-C) [8, 23, 36]. 3-E-F methods contain three separate encoders to model visual, textual, and cross-modal features; 2-E-F methods simplify the 3-E-F architecture by employing a shared encoder to embed textual and cross-modal features; while T-T-C methods are mainly based on the widely-adopted CLIP [28] architecture, which includes a couple of text and visual encoders to learn the cross-modal alignment by contrastive learning.", + "bbox": [ + 496, + 489, + 890, + 656 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Though the pixel-level paradigm has been widely adopted recently, existing Chinese video-text models [14,26,45] are all feature-level ones pre-trained on private datasets. In this work, we provide the first pixel-level benchmarks pre-trained on our CNVid-3.5M dataset.", + "bbox": [ + 496, + 657, + 890, + 733 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Video-Text Dataset. Well-designed large-scale datasets are a prerequisite for the successful applications of the \"Pre-training & Fine-tuning\" paradigm. There exist a few large-scale public English datasets (e.g., Howto100M [25], WebVid-2.5M [4]) for video-text pre-training. However, existing large-scale Chinese video datasets (ALIVOL [14], Kwai-SVC [26], CREATE [45]) are all private, whose videos and text are unavailable to the public. Therefore, we aimed to release a public one to facilitate future researchers.", + "bbox": [ + 496, + 733, + 890, + 868 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Hard Sample Mining in Contrastive Learning. Hard Sample Mining [12, 29] is an effective strategy to promote", + "bbox": [ + 496, + 869, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "14816", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 1 + }, + { + "type": "table", + "img_path": "images/b96e5adf0d4bb57206450bfdebed46c78382f64c9a3ec9b6c64b09fb5c1f4f78.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Dataset Name#(Videos)#(Text)Duration (hours)Video SourceText TypeAvailability
Off-FeatRaw-VidState
Datasets for Downstream Chinese Video-text Tasks
VATEX [38]41,269825,380115kinetics-600Caption✓✓All-public
BFVD [44]43,16643,166140E-CommerceTitle✓-Partly-public
FFVD [44]32,76332,763252E-CommerceTitle✓-Partly-public
CREATE210K [45]216,303268,5931,800Open WebsitesCaption--NOT-public
Kwai-SVC [26]222,077143,5693,500Open WebsitesTitle & ASR✓-Partly-public
Datasets for Chinese Video-text Pre-training
ALIVOL-10M [14]10,300,00011,000,00098,801E-CommerceTitle--NOT-public
Kwai-SVC-11M [26]11,075,0843,931,879177,200Open WebsitesTitle & ASR--NOT-public
CREATE-10M [45]10,000,00010,000,00083,000Open WebsitesTitle--NOT-public
CNVid-3.5M (ours)3,508,1203,508,12035,414Open WebsitesTitle & ASR✓✓All-public
", + "bbox": [ + 76, + 87, + 916, + 277 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Table 1. Data analysis of CNVid-3.5M and other Chinese video-text datasets. \"Off-Feat\" and \"Raw-Vid\" represent whether offline features and raw videos are available. As far as we know, CNVid-3.5M is the largest public video-text dataset in Chinese.", + "bbox": [ + 75, + 289, + 890, + 316 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/0807d4357590c810673f4124c45b001682cf285b1fa50e9e71996407f54d152e.jpg", + "image_caption": [ + "Figure 2. An example of video-text pairs in our CNVid-3.5M dataset. We present three types of text information: 1) Caption (manually written), 2) Title (user uploaded), and 3) ASR text." + ], + "image_footnote": [], + "bbox": [ + 80, + 340, + 478, + 566 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "the performance of contrastive learning. In video-text pretraining, TACo [43] proposes a cascade sampling method to construct hard negative examples for better cross-modal alignment. However, we conjectured that hard sample mining may hinder the video-text pre-training at the beginning, where the model is far from convergence and could hardly discriminate the ground-truth video-text pairs. Therefore, we introduced curriculum learning [32, 37] and proposed the Hard Sample Curriculum Learning (HSCL) strategy. HSCL would guide the pre-training procedure from \"starting small\" gradually to \"learning hard\", mitigating the side effect of conventional hard sample mining methods.", + "bbox": [ + 75, + 633, + 468, + 815 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. The CNVid-3.5M Dataset", + "text_level": 1, + "bbox": [ + 76, + 829, + 316, + 845 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To fill in the blank of public Chinese video-text datasets, we release CNVid-3.5M, a large-scale cross-modal dataset containing over 3.5M Chinese video-text pairs.", + "bbox": [ + 75, + 854, + 468, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Dataset Building", + "text_level": 1, + "bbox": [ + 500, + 343, + 666, + 359 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We collect raw videos from Douyin*, a popular Chinese social website, where millions of active users upload various types of videos every day. To enrich the data diversity, we do not limit video categories like HowTo100M [25] (mainly instructional videos) and cover various topics (e.g., food, entertainment, and technology). I.e., we first form up a set of keywords and then leverage them to search for videos and corresponding captions. Moreover, we span the data created time at a range of 3 years and filter the abnormal ones whose videos could not be played. In this way, we collect over 4.5M raw videos in Chinese.", + "bbox": [ + 496, + 367, + 890, + 532 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Besides, the supplementary material will present more details about fairness and privacy for safety and ethics.", + "bbox": [ + 498, + 534, + 890, + 566 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Captions, Titles, and ASR Text", + "text_level": 1, + "bbox": [ + 500, + 577, + 772, + 593 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "As CREATE [45] pointed out, three types of text information are widely-employed in current Chinese video-text datasets. We show an example in Figure 2, based on which we summarize their characteristics and applications:", + "bbox": [ + 496, + 601, + 890, + 661 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "1) Captions are manually written by human annotators. They could accurately describe the video content in a formal manner. However, obtaining a large number of captions is extremely expensive. Therefore, captions are mainly employed in small-scale downstream video-text datasets like VATEX [38]. Moreover, captions are indeed global descriptions of the whole videos. Therefore, they could hardly describe the frame-level information of videos.", + "bbox": [ + 496, + 662, + 890, + 782 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2) Titles are written and uploaded by users. Compared with captions, it is not unfeasible to obtain millions of user-uploaded titles. However, to increase the attractiveness and click rate, users may over-polish titles with some \"appealing but redundant\" words, which may decrease the video-text consistency. Similar to captions, titles are video-level", + "bbox": [ + 496, + 784, + 890, + 875 + ], + "page_idx": 2 + }, + { + "type": "page_footnote", + "text": "*Official Website: https://www.douyin.com/", + "bbox": [ + 517, + 887, + 753, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "14817", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "descriptions rather than frame-level ones. Therefore, titles and captions are mainly employed in feature-level videotext pre-training methods rather than pixel-level ones.", + "bbox": [ + 75, + 90, + 468, + 136 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3) ASR text is obtained by the ASR toolkit. Similar to titles, it is easy to obtain a large number of ASR text. However, some ASR text is incomplete, noisy, and even faulty, which may decrease the pre-training performance. Besides, ASR text is a frame-level description, which is widely employed in the pixel-level video-text pre-training.", + "bbox": [ + 75, + 136, + 467, + 227 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Following previous methods [25, 26], we collect user-uploaded titles and obtain the ASR text of each video. Since we provide both video-level titles and frame-level ASR text, our dataset could support the video-text pre-training in both feature-level and pixel-level manner.", + "bbox": [ + 75, + 227, + 467, + 303 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Weakly-paired Data Filtering", + "text_level": 1, + "bbox": [ + 76, + 313, + 339, + 330 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As some work [25, 26] has pointed out, some ASR text may not match with associated videos (e.g., videos have background music), which would decrease the pre-training performance. However, the problem is only raised with few practical solutions. Therefore, we propose a novel weakly-paired data filtering strategy by leveraging a well-trained image-text model to automatically evaluate the video-text consistency for three reasons: 1) The text information in existing image-text datasets [11, 13] are manually-written titles or captions, whose quality and consistency are guaranteed. 2) Pixel-level pre-training methods [8, 15, 23] usually sample several frames from raw videos to serve as the visual input, whose frameworks are closely associated with image-text pre-trained models. 3) Compared with manually checking millions of videos, the proposed filtering strategy remarkably reduces the cost and improves the efficiency.", + "bbox": [ + 75, + 337, + 467, + 578 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Specifically, we first pre-train an image-text model on Wukong-100M [11]. We then leverage it to calculate the similarity score between videos and ASR text, and filter out the last 1M videos by their ranks. In this way, we obtain the CNVid-3.5M dataset, based on which we provide three mainstream pixel-level pre-training benchmarks.", + "bbox": [ + 75, + 579, + 467, + 670 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.4. Comparison of Dataset Statistics", + "text_level": 1, + "bbox": [ + 76, + 680, + 362, + 696 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Table 1 presents the statistics of CNVid-3.5M and other Chinese video-text datasets. To the best of our knowledge, CNVid-3.5M is the largest public dataset for Chinese video-text pre-training, with a total of 3.5M videos that last for 35.4K hours. Though ALIVOL-10M [14], KwaiSVC-11M [26], and CREATE-10M [45] contain more Chinese videos than ours, yet they do not release raw videos and text information to the public. Moreover, these three methods only provide conventional feature-level pretrained models, whose performance highly depends on frozen feature extractors. Towards this end, we provide the first pixel-level benchmarks based on our CNVid-3.5M dataset, which would pave the way for future research on", + "bbox": [ + 75, + 703, + 467, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Chinese video-text pre-training.", + "bbox": [ + 498, + 90, + 712, + 106 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Moreover, we present additional statistics of our CNVid-3.5M dataset in the supplementary material, including the distributions of topics, keywords, video durations, and Part-of-Speech (POS) tags of the ASR text.", + "bbox": [ + 496, + 106, + 890, + 167 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Pixel-level Pre-training Paradigm", + "text_level": 1, + "bbox": [ + 498, + 181, + 805, + 200 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Since mainstream pixel-level video-text pre-training methods [17, 23, 35] extract cross-modal representations from raw data directly, they usually outperform the conventional feature-level ones [2, 24, 33]. Therefore, we benchmark our CNVid-3.5M dataset with three mainstream pixel-level architectures. As illustrated in Figure 3, we split the pixel-level pre-training process into three steps, namely: 1) data pre-processing to process raw videos and text into patches and tokens, 2) model architectures to generate cross-modal representations, and 3) proxy tasks to determine overall pre-training objectives.", + "bbox": [ + 496, + 208, + 890, + 375 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1. Data Pre-processing", + "text_level": 1, + "bbox": [ + 498, + 386, + 691, + 402 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Given a mini-batch (denoted as $\\mathcal{B}$ ) of videos $\\{V_i\\}_{i=1}^{|B|}$ and their corresponding ASR text $\\{T_i\\}_{i=1}^{|B|}$ , pixel-level pre-training methods would first sparsely (and randomly) sample $N_v$ frames from each video. We then slice each frame into patches, obtaining visual patch embeddings $\\mathbf{P} \\in \\mathbb{R}^{N_v * N_p * d}$ , where $N_p$ is the number of sliced patches, and $d$ is the dimension of the embedding. Simultaneously, a BERT embedder is employed to process the text $T$ into fixed-length word embeddings $\\mathbf{W} = [\\mathbf{w}^{cls}, \\mathbf{w}^1, \\mathbf{w}^2, \\dots, \\mathbf{w}^{N_t - 1}]$ , where $\\mathbf{W} \\in \\mathbb{R}^{N_t * d}$ , $N_t$ is the length of text tokens.", + "bbox": [ + 496, + 409, + 890, + 580 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.2. Model Architectures", + "text_level": 1, + "bbox": [ + 498, + 590, + 692, + 606 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "There are three mainstream pixel-level video-text pretraining architectures, namely three-encoders-fusion [15, 17, 30], two-encoders-fusion [9], and twin-towers-crossed [8, 23, 36]. We benchmark CNVid-3.5M with these three pixel-level pre-training architectures.", + "bbox": [ + 496, + 614, + 890, + 691 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Three-encoders-fusion (3-E-F) pre-training methods contain a visual encoder $E_{vis}$ , a text encoder $E_{txt}$ , and a cross-modal encoder $E_{mul}$ . After obtaining visual features $\\mathbf{V} = E_{vis}(\\mathbf{P})$ and textual features $\\mathbf{T} = E_{txt}(\\mathbf{W})$ , we concatenate these two-modal features into $[\\mathbf{T},\\mathbf{V}] = [\\mathbf{t}^{cls},\\mathbf{t}^1,\\mathbf{t}^2,\\dots ,\\mathbf{t}^{N_t - 1},\\mathbf{v}^{cls},\\mathbf{v}^1,\\mathbf{v}^2,\\dots ,\\mathbf{v}^{N_v}]$ . Following CLIP4Clip [23], we set position embeddings to zero, and obtain cross-modal features $\\mathbf{M} = E_{mul}([\\mathbf{T},\\mathbf{V}])$ . This forwarding process could be formulated as follows:", + "bbox": [ + 496, + 691, + 890, + 828 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {M} = E _ {m u l} ([ E _ {t x t} (\\mathbf {W}), E _ {v i s} (\\mathbf {P}) ]), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 578, + 840, + 890, + 858 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathbf{M} = [\\mathbf{m}^{cls},\\mathbf{m}^1,\\mathbf{m}^2,\\dots ,\\mathbf{m}^{N_t + N_v}]$ , and $[\\cdot ,\\cdot ]$ denotes the concatenation operation.", + "bbox": [ + 496, + 869, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "14818", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/b49f4b4679379e1001da74ede97fe47584a7138ef4d59dddec2cdeff492ba927.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
1. Data Pre-processing2. Three Types of Model Architecture3. Proxy Tasks
Text Embedding\nC | | | | BERT Embedder\n| | | | | [CLS] a cat plays piano\n[CLS] 猫弹钢琴\nASR Caption\n| | | | | Input VideoThree-Encoders-Fusion\nCross-Modal Feature\n| | | | | Cross-Modal Encoder\n| | | | | Emut | | | | Text FeatureEbert\nShared Encoder\n| | | | | C | | | | Video Feature\n| | | | | Shared Encoder\n| | | | | C | | | | Video Encoder\n| | | | | C | | | | Cross-Modal EncodingTwo-Encoders-Fusion\nCross-Modal Feature\n| | | | | C | | | | Video Feature\n| | | | | C | | | | Video Encoder\n| | | | | C | | | | EvisTwin-Towers-Crossed\nContrastive Learning\n| | | | | C | | | | Video Feature\n| | | | | C | | | | Video Encoder\n| | | | | C | | | | EvisMasked Language Modeling (MLM)\nFC\n[LMSK]\n[C | | | | F | | | | F | | | | F | | | | F | | | | F | | | | F | | | | F | | | | F | | | | F | | | | F | | | | F | | | | F | | | | F | | | | F | | | | F | | | | F | | | | F | | | | F | | | | F | | | | F | | | |
", + "bbox": [ + 78, + 88, + 893, + 247 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Figure 3. The pipeline of mainstream pixel-level video-text pre-training methods, which contains three key steps: 1) Data pre-processing, 2) model architectures, and 3) proxy tasks. Note that we employ three mainstream architectures to enrich pixel-level pre-training benchmarks.", + "bbox": [ + 75, + 253, + 893, + 284 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Two-encoders-fusion (2-E-F) pre-training methods simplify the 3-E-F architecture by employing a shared BERT-type encoder $E_{bert}$ to embed textual and cross-modal features simultaneously. The pipeline of obtaining textual features $\\mathbf{T}$ and cross-modal features $\\mathbf{M}$ could be formulated as follows:", + "bbox": [ + 75, + 295, + 473, + 386 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\left\\{ \\begin{array}{l} \\mathbf {T} = E _ {b e r t} (\\mathbf {W}), \\\\ \\mathbf {M} = E _ {b e r t} ([ \\mathbf {W}, E _ {v i s} (\\mathbf {P}) ]). \\end{array} \\right. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 169, + 393, + 468, + 436 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Twin-towers-crossed $(T-T-C)$ pre-training methods mainly follow the conventional CLIP [28] architecture. After obtaining visual and textual features, $T-T-C$ methods employ contrastive learning to optimize the twin encoders $E_{vis}$ and $E_{txt}$ for better cross-modal alignment. Note that $T-T-C$ methods do not generate cross-modal video representations like $3-E-F$ and $2-E-F$ .", + "bbox": [ + 76, + 445, + 472, + 551 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.3. Proxy Tasks", + "text_level": 1, + "bbox": [ + 76, + 561, + 209, + 577 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Proxy tasks are crucial for video-text pre-training as they directly determine the final optimization objectives. Following the conventional protocol [4, 15, 43], we adopt two widely-employed proxy tasks, i.e., Masked Language Modeling (MLM) and Video-Text Matching (VTM).", + "bbox": [ + 75, + 584, + 468, + 660 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "MLM first masks out a certain percentage of words in a given sentence, and then forces the model to restore these clozes according to visual and textual cues. MLM is calculated twice for textual features $\\mathbf{T}(\\mathcal{L}_1)$ and cross-modal features $\\mathbf{M}(\\mathcal{L}_2)$ as follows:", + "bbox": [ + 75, + 660, + 470, + 736 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {1} = \\frac {1}{| \\mathcal {Q} |} \\sum_ {q \\in \\mathcal {Q}} \\mathcal {L} _ {C E} \\left(y ^ {q}, \\Theta_ {1} \\left(\\mathbf {t} ^ {q}\\right)\\right), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 160, + 744, + 468, + 782 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {2} = \\frac {1}{| \\mathcal {Q} |} \\sum_ {q \\in \\mathcal {Q}} \\mathcal {L} _ {C E} \\left(y ^ {q}, \\Theta_ {2} \\left(\\mathbf {m} ^ {q}\\right)\\right), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 156, + 795, + 468, + 833 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathcal{Q}$ records the location of masked tokens, $|\\cdot|$ denotes the length of a given set, $y^{q}$ denotes the ground-truth token label, $\\Theta$ is a Multi-Layer Perception (MLP), and $\\mathcal{L}_{CE}$ is the regular Cross-Entropy cost function.", + "bbox": [ + 75, + 839, + 472, + 902 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "VTM aims to promote the cross-modal alignment between videos and text, which is calculated in a parameter-free $(\\mathcal{L}_3)$ and parameter-employed $(\\mathcal{L}_4)$ way:", + "bbox": [ + 498, + 295, + 892, + 342 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {3} = - \\sum_ {i = 1} ^ {| \\mathcal {B} |} \\log \\frac {\\sum_ {k = 1} ^ {N _ {v}} \\exp \\left\\langle \\mathbf {v} _ {i} ^ {k} , \\mathbf {t} _ {i} ^ {c l s} \\right\\rangle}{\\sum_ {k = 1} ^ {N _ {v}} \\left(\\exp \\left\\langle \\mathbf {v} _ {i} ^ {k} , \\mathbf {t} _ {i} ^ {c l s} \\right\\rangle + \\sum_ {j \\neq i} \\exp \\left\\langle \\mathbf {v} _ {j} ^ {k} , \\mathbf {t} _ {i} ^ {c l s} \\right\\rangle\\right)}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 501, + 353, + 890, + 410 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {4} = - \\sum_ {i = 1} ^ {| \\mathcal {B} |} \\log \\frac {\\exp^ {\\Theta_ {4} \\left(\\mathbf {f} _ {i , i}\\right)}}{\\exp^ {\\Theta_ {4} \\left(\\mathbf {f} _ {i , i}\\right)} + \\sum_ {j \\neq i} \\exp^ {\\Theta_ {4} \\left(\\mathbf {f} _ {j , i}\\right)}}, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 526, + 434, + 890, + 478 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $|\\mathcal{B}|$ is the length of a mini-batch, $\\langle \\cdot ,\\cdot \\rangle$ denotes the matrix multiplication operation, and $\\mathbf{f}_{j,i}$ is global-pooling cross-modal features of the video-text pair $(V_{j},S_{i})$", + "bbox": [ + 498, + 489, + 890, + 535 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For $3 - E - F$ and $2 - E - F$ models, the objective functions $\\mathcal{L}_{3EF} / \\mathcal{L}_{2EF}$ are calculated as follows:", + "bbox": [ + 500, + 535, + 890, + 566 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {3 E F} \\left(\\mathcal {L} _ {2 E F}\\right) = \\mathcal {L} _ {1} + \\mathcal {L} _ {2} + \\mathcal {L} _ {3} + \\mathcal {L} _ {4}. \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 566, + 578, + 890, + 595 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "While for $T$ - $T-C$ models, since they do not contain a cross-modal encoder to generate video representations, their objective function $\\mathcal{L}_{TTC}$ is calculated as follows:", + "bbox": [ + 498, + 607, + 890, + 654 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {T T C} = \\mathcal {L} _ {1} + \\mathcal {L} _ {3}. \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 630, + 667, + 890, + 683 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.4. Fine-tuning", + "text_level": 1, + "bbox": [ + 500, + 695, + 627, + 710 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In order to thoroughly evaluate the pre-training performance, we fine-tune our models on the Text-to-Video Retrieval (TVR) task. TVR is a widely-employed downstream video-text task, aiming to retrieve the most relevant videos according to text queries. Since there has only one Chinese TVR dataset (VATEX [38]), we translate two widely-adopted English datasets (MSRVTT [41] and DiDemo [3]) into Chinese to verify the superiority and robustness of the constructed dataset and proposed methods.", + "bbox": [ + 496, + 719, + 890, + 853 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Specifically, we fine-tune our pre-trained models by reusing Video-Text Matching objectives ( $\\mathcal{L}_3$ in Eq. 5 and $\\mathcal{L}_4$ in Eq. 6, we only employ $\\mathcal{L}_3$ for $T$ - $T-C$ models).", + "bbox": [ + 498, + 854, + 892, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "14819", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/9706241771516cafa10cc722eb7557c1d8ff1513a795cdc863411f7b57d921cc.jpg", + "image_caption": [ + "Figure 4. The pipeline of Hard Sample Curriculum Learning (HSCL), which contains three components to promote the performance of video-text pre-training." + ], + "image_footnote": [], + "bbox": [ + 81, + 89, + 477, + 209 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Hard Sample Curriculum Learning", + "text_level": 1, + "bbox": [ + 76, + 273, + 400, + 292 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We propose Hard Sample Curriculum Learning (HSCL), a model-agnostic strategy to promote the video-text pretraining performance. HSCL is based on two motivations. 1) The negative samples involved in the calculation of parameter-employed VTM loss $(\\mathcal{L}_4$ in Eq. 6) are not representative. Suppose that the batch size is set to be $N_{k}$ , a model needs to process $N_{k} * N_{k}$ video-text pairs for calculating $\\mathcal{L}_4$ , whose calculation cost is hardly affordable. Therefore, previous methods [9, 15, 22] tend to cut the number of negative samples to $N_{k}^{\\prime}$ by random sampling. However, since $N_{k}^{\\prime}$ is usually much smaller than $N_{k}$ , it may result in the sub-optimal performance. 2) If a model is far from convergence, directly introducing hard negative samples would amplify the burden of learning, which would decrease the pre-training performance.", + "bbox": [ + 75, + 299, + 468, + 526 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Therefore, inspired by hard sample mining and curriculum learning, we propose the HSCL strategy. By leveraging the similarity score calculated by the parameter-free VTM loss $(\\mathcal{L}_3$ in Eq. 5), HSCL could gradually and smoothly introduce hard samples for better cross-modal alignment. As illustrated in Figure 4, HSCL contains the following three key components:", + "bbox": [ + 75, + 527, + 468, + 633 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Positive Column Weighting (PCW) aims to decrease the learning coefficient $\\phi$ of well-learned positive samples, forcing the model to emphasize those hard positive ones. Specifically, we re-weigh half of samples within the batch according to their similarity scores $z$ calculated by $\\mathcal{L}_3$ , which could be formulated as follows:", + "bbox": [ + 75, + 633, + 468, + 723 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nz _ {j, i} = \\frac {1}{N _ {v}} \\sum_ {k = 1} ^ {N _ {v}} \\exp \\left\\langle \\mathbf {v} _ {j} ^ {k}, \\mathbf {t} _ {i} ^ {c l s} \\right\\rangle , \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 179, + 732, + 468, + 773 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\phi_ {i} = \\operatorname {M i n} \\left(\\frac {\\operatorname {M e d} \\left(Z _ {d i a g}\\right) - \\operatorname {M i n} \\left(Z _ {d i a g}\\right)}{z _ {i , i} - \\operatorname {M i n} \\left(Z _ {d i a g}\\right) + e _ {0}}, 1. 0\\right), \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 109, + 785, + 468, + 820 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $Z_{diag} = [z_{1,1}, z_{2,2}, \\dots, z_{i,i}, \\dots]$ , $\\operatorname{Med}(\\cdot)$ and $\\operatorname{Min}(\\cdot)$ represent the median and minimum result of the list, $e_0$ is a pre-defined coefficient to avoid a zero denominator.", + "bbox": [ + 76, + 824, + 468, + 869 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Negative Line Sampling (NLS) aims to sample $N_{k}^{\\prime}$ hard negative examples for better contrastive learning. Specifi-", + "bbox": [ + 76, + 869, + 468, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "cally, for each text $T_{j}$ , we choose the Nearliest $N_{k}^{\\prime}$ aligned negative videos according to their similarity scores in the list $Z_{j} = [z_{j,1},z_{j,2},\\dots ,z_{j,i},\\dots ]$ . I.e., Pick the minimum $N_{k}^{\\prime}$ samples in the list $Z_{j}^{\\prime} = \\{|z_{j,i} - z_{j,j}|\\}_{j\\neq i}$", + "bbox": [ + 496, + 89, + 890, + 152 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Curriculum Learning (CL) aims to mitigate the side effect of hard sample mining when the model is under convergence. We set a gradually increasing coefficient $\\beta \\in (0,1)$ . $\\beta$ is initialized with zero at the beginning of pre-training. It will gradually increase by a fixed number $\\beta_0$ for each $N_d$ iteration until reaching 1.0. For PCW, the actual coefficient $\\phi^*$ is calculated as follows:", + "bbox": [ + 496, + 152, + 892, + 258 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\phi_ {i} ^ {*} = (1. 0 - \\beta) + \\beta * \\phi_ {i}. \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 607, + 271, + 890, + 287 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For NLS, we first generate a random number $rand \\in (0,1)$ , and only perform the NLS step if $rand < \\beta$ . In this way, the model would conduct conventional pre-training at the beginning, and smoothly change to learn hard negative samples to pursue better pre-training performance.", + "bbox": [ + 496, + 299, + 890, + 376 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "6. Experiments and Benchmarks", + "text_level": 1, + "bbox": [ + 498, + 388, + 777, + 405 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "6.1. Pre-training Datasets", + "text_level": 1, + "bbox": [ + 500, + 412, + 700, + 429 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In Section 3.3, the Chinese image-text model that helps to filter out the weakly-paired data is pre-trained on Wukong-100M [11] (100M image-text pairs). The Chinese video-text models are pre-trained on our CNVid-3.5M dataset, which contains 3.5M video-text pairs.", + "bbox": [ + 496, + 436, + 890, + 511 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To verify the generality of the proposed Hard Sample Curriculum Learning (HSCL) strategy, we also pre-trained our models on four large-scale English datasets, they are 1) COCO [20] (0.6M image-text pairs), 2) VG [13] (5.4M image-text pairs), 3) CC [31] (3.1M image-text pairs), and 4) WebVid-2.5M (2.5M video-text pairs).", + "bbox": [ + 496, + 512, + 890, + 603 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "6.2. Fine-tuning Datasets", + "text_level": 1, + "bbox": [ + 500, + 612, + 694, + 628 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We fine-tuned our pre-trained models on the Text-to-Video Retrieval (TVR) task, which includes three widely-adopted datasets as follows: 1) VATEX [38] contains 28K video clips, and each video is associated with 10 Chinese and 10 English diverse captions. 2) MSRVTT [41] contains 10K video clips associated with 200K English captions. Following [15, 22, 43], we used 7K videos for training and randomly selected 1K videos from the remaining ones for testing (7K-1K split). 3) DiDemo [3] contains 10K Flickr videos associated with 40K English captions.", + "bbox": [ + 496, + 635, + 890, + 786 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Moreover, for a comprehensive performance evaluation of Chinese video-text pre-trained models, we employ Google $^{\\dagger}$ to translate all captions in MSRVTT and DiDemo from English to Chinese.", + "bbox": [ + 496, + 787, + 890, + 845 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Metrics. We employed Recall@K (R@K, K=1/5/10) and Median Rank (MdR) to measure the TVR performance.", + "bbox": [ + 496, + 847, + 890, + 877 + ], + "page_idx": 5 + }, + { + "type": "page_footnote", + "text": "†Official Website: https://translate.google.com/", + "bbox": [ + 514, + 886, + 769, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "14820", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/d48ae3dcf253a0419016bd4f4f0af66578ac2931aa42d38096bd3e31fb981a1c.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
No.Model TypeData VolumeMasking StrategyMatching StrategyVATEX (Chinese)MSRVTT (Translated)DiDemo (Translated)
R@1/5/10↑(MdR↓)R@1/5/10↑(MdR↓)R@1/5/10↑(MdR↓)
", + "bbox": [ + 78, + 74, + 901, + 104 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/6c36804f02c728b7dce0a29e4ba094b4ef9463280689c3cb4ad2532aa2955733.jpg", + "table_caption": [ + "For \"Building\": Ablation Study of the constructed CNVid-3.5M dataset" + ], + "table_footnote": [], + "table_body": "
A12-E-FNo Pre-training36.4 / 75.4 / 85.2 (2)15.9 / 42.1 / 54.1 (8)8.0 / 24.8 / 39.2 (18)
A22-E-F3.5MMLMVTM39.9 / 77.2 / 87.0 (2)20.7 / 47.9 / 61.2 (6)13.3 / 34.6 / 45.9 (13)
", + "bbox": [ + 80, + 122, + 901, + 154 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/67d242b659bdf66d52807d5700a4c2d0cdbd1d3e72b475eb42d2a28db4b24754.jpg", + "table_caption": [ + "For \"Filtering\": Ablation Study of Weakly-Paired Data Filtering" + ], + "table_footnote": [], + "table_body": "
B12-E-F4.5MMLMVTM38.8 / 76.4 / 86.4 (2)19.4 / 45.6 / 58.6 (6)10.6 / 32.7 / 45.0 (14)
B22-E-F4.0MMLMVTM39.5 / 77.0 / 86.5 (2)19.8 / 47.2 / 60.3 (6)12.4 / 32.8 / 45.6 (14)
B32-E-F3.5MMLMVTM39.9 / 77.2 / 87.0 (2)20.7 / 47.9 / 61.2 (6)13.3 / 34.6 / 45.9 (13)
B42-E-F3.0MMLMVTM40.5 / 76.8 / 86.5 (2)20.6 / 47.3 / 61.3 (6)12.2 / 34.3 / 45.6 (14)
B52-E-F2.5MMLMVTM39.5 / 76.6 / 86.9 (2)20.1 / 47.4 / 60.8 (6)12.1 / 33.6 / 46.0 (14)
", + "bbox": [ + 80, + 171, + 901, + 244 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/a9f7c9fe68f54926257ccb8785e6da493d86fb817e7a71854359443a05809cca.jpg", + "table_caption": [ + "For \"Pre-training\": Ablation Study of Hard Sample Curriculum Learning" + ], + "table_footnote": [], + "table_body": "
C12-E-F3.5MMLMVTM39.9 / 77.2 / 87.0 (2)20.7 / 47.9 / 61.2 (6)13.3 / 34.6 / 45.9 (13)
C22-E-F3.5MMLMVTM+HSM40.7 / 76.8 / 86.9 (2)21.1 / 47.3 / 60.1 (6)13.4 / 34.3 / 46.4 (12)
C32-E-F3.5MMLMVTM+HSCL41.5 / 78.2 / 87.2 (2)23.3 / 48.0 / 61.2 (6)13.6 / 34.4 / 47.3 (12)
", + "bbox": [ + 80, + 263, + 901, + 308 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/5b5c0f38fb5079c16556621b6bc591ddb4e9200bddcf4814a34f326dab1d1a22.jpg", + "table_caption": [ + "Table 2. For Chinese video-text pre-training: Ablation study of Weakly-Paired Data Filtering and Hard Sample Curriculum Learning (HSCL) on the Text-to-Video Retrieval task of three datasets (VATEX-Chinese, MSRVTT-Translated, and DiDemo-Translated). \"HSM\" denotes the conventional Hard Sample Mining strategy without curriculum learning. Note that $A2$ , $B3$ , and $C1$ are the same model." + ], + "table_footnote": [], + "table_body": "
No.Pre-training DatasetMasking StrategyMatching StrategyVATEX (English)MSRVTTDiDemo
R@1/5/10 ↑ (MdR ↓)R@1/5/10 ↑ (MdR ↓)R@1/5/10 ↑ (MdR ↓)
D0No Pre-training39.5 / 78.3 / 88.6 (2)19.2 / 46.2 / 59.5 (6)14.2 / 34.1 / 46.1 (12)
D1COCO+VG+CCMLMVTM45.5 / 81.5 / 91.0 (2)28.6 / 55.5 / 66.1 (4)25.3 / 50.3 / 62.4 (5)
D2COCO+VG+CCMLMVTM+HSCL48.7 / 83.1 / 91.1 (2)31.6 / 56.5 / 66.7 (4)27.1 / 53.8 / 63.8 (4)
D3CC+WebVidMLMVTM50.5 / 85.6 / 92.5 (1)29.4 / 56.9 / 68.3 (4)29.0 / 57.3 / 67.8 (4)
D4CC+WebVidMLMVTM+HSCL53.4 / 86.2 / 92.9 (1)32.6 / 58.8 / 69.7 (3)29.8 / 58.0 / 68.9 (3)
", + "bbox": [ + 78, + 375, + 898, + 484 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 3. For English video-text pre-training: Ablation study of Hard Sample Curriculum Learning (HSCL) on the Text-to-Video Retrieval task of three datasets (VATEX-English, MSRVTT, and DiDemo). All models are based on the two-encoders-fusion $(2 - E - F)$ architecture.", + "bbox": [ + 75, + 494, + 892, + 523 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "6.3. Experimental Settings", + "text_level": 1, + "bbox": [ + 76, + 535, + 284, + 551 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Settings for Weakly-paired Data Filtering. We pretrained the image-text model on the Wukong-100M [11] dataset, whose pre-training details are presented in the supplementary material. We then employed the pre-trained model to obtain the parameter-free video-text consistency score $(z_{i,i}$ in Eq. 9) of 4.5M video-text pairs. To improve the reliability, we randomly sampled four frames to calculate the consistency score, and repeated this step for three times. The final score of each video is the average of three results calculated under different random seeds. Ultimately, we sorted the consistency score in descending order, and filtered out the last 1M videos to refine the dataset.", + "bbox": [ + 75, + 558, + 468, + 738 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Settings for Chinese Video-Text Pre-training. For model architectures, we employed the Video Swin Transformer [21] to serve as the visual encoder, which is initialized with parameters pre-trained on ImageNet [6]. The text, cross-modal, and shared encoders belong to the BERT-Base [7] model, which are initialized with parameters privdied by Hugging Face $\\ddagger$ . We sparsely sampled $4(N_v)$ frames from raw videos during pre-training. For the MLM task, the masking rate is $15\\%$ . For the proposed Hard Sample", + "bbox": [ + 75, + 739, + 468, + 876 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Curriculum Learning, the coefficient $\\beta$ would increase by $0.15(\\beta_0)$ for each $5000(N_d)$ iterations until reaching 1.0.", + "bbox": [ + 500, + 536, + 890, + 565 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "For hyper-parameters, we set the length of text tokens $N_{t} = 30$ , and the dimension of the hidden state $d = 768$ . All models are pre-trained by the Adam optimizer with a momentum of 0.9. The total pre-training stage lasts for 10 epochs with a batch size of 128. The initial learning rate is 5e-5 and is decayed by the factor of 10 after 5 epochs. The whole pre-training on CNVid-3.5M takes about 4 days to complete on 8 NVIDIA V100 GPUs.", + "bbox": [ + 498, + 566, + 890, + 685 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Settings for Fine-tuning on Downstream Chinese Videotext Tasks. The optimizer and hyper-parameters in fine-tuning remain the same as the pre-training configuration. The total fine-tuning stage lasts for 25,000 steps. The batch size is set to 128. The initial learning rate is set to 1e-5.", + "bbox": [ + 498, + 686, + 890, + 762 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "6.4. Ablation Study", + "text_level": 1, + "bbox": [ + 500, + 771, + 653, + 787 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "As aforementioned, we summarized our work with three verbs, i.e., \"Build\", \"Filter\", and \"Pre-train\". To evaluate the effectiveness of these three contributions, we took the 2- $E-F$ architecture as an example model in the ablation study. Ablation Study of \"Building\". We compared the performance obtained by models with or without pre-training. As illustrated in Table 2 (A1 vs. A2), models pre-trained on", + "bbox": [ + 498, + 795, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_footnote", + "text": "$^\\ddagger$ Official Website: https://huggingface.co/bert-base-chinese", + "bbox": [ + 94, + 886, + 410, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "14821", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/0654043d02b55207e91364054e3ee9d60860821669438fc65b92ec74a541b209.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
No.Model TypeDatasetPre-training StrategyVATEX (Chinese)MSRVTT (Translated)DiDemo (Translated)
R@1/5/10 ↑ (MdR ↓)R@1/5/10 ↑ (MdR ↓)R@1/5/10 ↑ (MdR ↓)
E13-E-FCNVid-3.5MMLM+VTM39.9 / 78.2 / 87.6 (2)20.6 / 47.1 / 60.0 (6)13.8 / 34.5 / 46.6 (13)
E22-E-FCNVid-3.5MMLM+VTM39.9 / 77.2 / 87.0 (2)20.7 / 47.9 / 61.2 (6)13.3 / 34.6 / 45.9 (13)
E3T-T-CCNVid-3.5MMLM+VTM40.5 / 77.3 / 87.0 (2)21.1 / 46.5 / 58.9 (6)13.0 / 32.6 / 44.6 (15)
", + "bbox": [ + 81, + 88, + 885, + 162 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 4. We benchmark CNVid-3.5M with three mainstream pixel-level pre-training architectures, which is the first pixel-level benchmarks for Chinese video-text pre-training.", + "bbox": [ + 75, + 172, + 892, + 202 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "our CNVid-3.5M dataset (A2) would achieve a remarkable performance gain compared with those without pre-training (A1), which demonstrates the effectiveness of building a large-scale Chinese video-text dataset.", + "bbox": [ + 75, + 213, + 467, + 273 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Ablation Study of \"Filtering\". We compared the performance obtained by models pre-trained under different numbers of video-text pairs. For example, \"4.5M\", \"4.0M\", and \"3.5M\" denotes that we employ the 4.5M full set, the Top-4.0M split set, and the Top-3.5M split set (CNVid-3.5M) according to the sorted video-text consistency scores calculated by weakly-paired data filtering. As illustrated in Table 2 (B1-B5), we have two conclusions as follows:", + "bbox": [ + 75, + 275, + 467, + 393 + ], + "page_idx": 7 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1) The proposed weakly-paired data filtering strategy could effectively improve the quality of datasets. The fine-tuning performance on three downstream datasets first increases rapidly when filtering out videos with low video-text consistency scores ( $B1 \\to B2 \\to B3$ ). It would remain stable ( $B3 \\to B4$ ) or drop slightly ( $B4 \\to B5$ ) when continuously removing videos after the Top-3.5M split.", + "2) In general, employing the Top-3.5M split (B3) achieves the best performance. The results among all ablation models (B1-B5) may prove that, for the video-text pre-training, it is not \"more data is better\", but \"more good data is better\". Therefore, we filtered out the last 1M videos, resulting in the CNVid-3.5M dataset." + ], + "bbox": [ + 75, + 397, + 467, + 593 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Ablation Study of \"Pre-training\" (HSCL). To prove the effectiveness of Hard Sample Curriculum Learning (HSCL), we pre-trained several ablation models on Chinese video-text datasets ( $C1-C3$ in Table 2) and English ones ( $D1-D4$ in Table 3). We have two conclusions as follows:", + "bbox": [ + 75, + 595, + 467, + 670 + ], + "page_idx": 7 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1) HSCL is a model-agnostic strategy, which could evidently improve the pre-training performance. In Table 2, compared with the baseline (C1), HSCL (C3) achieves an improvement of $1.6\\%$ , $2.6\\%$ , and $0.3\\%$ at R@1 of VATEX, MSRVTT, and DiDemo dataset, respectively. While in Table 3, HSCL outperforms the baseline by $3.2 / 3.0 / 1.8\\%$ ( $D1$ vs. $D2$ on $\\mathrm{COCO + VG + CC}$ ) and $2.9 / 3.2 / 0.8\\%$ ( $D3$ vs. $D4$ on $\\mathrm{CC + WebVid}$ ) at R@1 of three TVR datasets.", + "2) The proposed Curriculum Learning (CL) strategy in HSCL could effectively mitigate the side effect of conventional Hard Sample Mining (HSM) when the model is far from convergence. As illustrated in Table 2 (C1 vs. C2), models equipped with HSCL perform better than HSM." + ], + "bbox": [ + 75, + 672, + 467, + 868 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Moreover, we present the detailed quantitative analyses of PCW and NLS in the supplementary material.", + "bbox": [ + 75, + 869, + 467, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6.5. Chinese Video-text Pre-training Benchmarks", + "text_level": 1, + "bbox": [ + 498, + 213, + 879, + 229 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Based on the constructed CNVid-3.5M dataset, We provided the first pixel-level benchmarks for Chinese videotext pre-training. Following existing video-text pre-training methods, we adopted three mainstream pixel-level architectures, i.e., three-encoders-fusion (3-E-F) [15, 17, 30], two-encoders-fusion (2-E-F) [9], and twin-towers-crossed (T-T-C) [8, 23, 36]. Detailed performance comparisons of three pixel-level benchmarks are presented in Table 4.", + "bbox": [ + 496, + 236, + 890, + 356 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "7. Conclusion", + "text_level": 1, + "bbox": [ + 498, + 371, + 617, + 386 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this work, we release the largest public Chinese video-text dataset, i.e., CNVid-3.5M. Besides, we provide the first pixel-level benchmarks for Chinese video-text pretraining. Moreover, we propose a novel weakly-paired data filtering method to improve the quality of datasets. We also design a novel Hard Sample Curriculum Learning strategy to promote the pre-training performance. In conclusion, we \"build\", \"filter\", and \"pre-train\" the large-scale Chinese video-text dataset. We believe this work would pave the way for future research on Chinese video-text pre-training.", + "bbox": [ + 496, + 396, + 890, + 547 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "8. Limitations and Future Work", + "text_level": 1, + "bbox": [ + 498, + 560, + 769, + 575 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Address the Difference between English and Chinese. One potential limitation of our work is that we adopt pixel-level architectures proved effective in English video-text pre-training, ignoring the linguistic gap between Chinese and English. E.g., it may be better to employ the Whole Word Masking (WWM) [5, 46] strategy rather than the conventional MLM for Chinese video-text pre-training. We plan to explore more Chinese-specific methods in the future. Need More Downstream Chinese Video-Text Datasets. Though we release a large-scale Chinese video-text dataset, there still lack some downstream datasets (e.g., MSRVTT [41], MSVD-QA [39] in English) to comprehensively evaluate the performance of Chinese video-text pre-trained models. Currently, there only exists VATEX [38] for Chinese cross-modal retrieval and some captioning datasets [26, 45]. Therefore, building more downstream Chinese video datasets is a promising future direction.", + "bbox": [ + 496, + 585, + 890, + 843 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments. This work is supported by the National Natural Science Foundation of China, No.: 62176137; Ant Group.", + "bbox": [ + 496, + 852, + 890, + 898 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "14822", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 174, + 106 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Nayyer Aafaq, Ajmal Mian, Wei Liu, Syed Zulqarnain Gilani, and Mubarak Shah. Video description: A survey of methods, datasets, and evaluation metrics. ACM Computing Surveys, 52(6):1-37, 2019. 1", + "[2] Hassan Akbari, Liangzhe Yuan, Rui Qian, Wei-Hong Chuang, Shih-Fu Chang, Yin Cui, and Boqing Gong. Vatt: Transformers for multimodal self-supervised learning from raw video, audio and text. Advances in Neural Information Processing Systems, 34, 2021. 4", + "[3] Lisa Anne Hendricks, Oliver Wang, Eli Shechtman, Josef Sivic, Trevor Darrell, and Bryan Russell. Localizing moments in video with natural language. In Proceedings of the IEEE international conference on computer vision, pages 5803-5812, 2017. 5, 6", + "[4] Max Bain, Arsha Nagrani, Gül Varol, and Andrew Zisserman. Frozen in time: A joint video and image encoder for end-to-end retrieval. In Proceedings of the IEEE International Conference on Computer Vision, pages 1728-1738, 2021. 1, 2, 5", + "[5] Yiming Cui, Wanxiang Che, Ting Liu, Bing Qin, and Ziqing Yang. Pre-training with whole word masking for chinese bert. IEEE Transactions on Audio, Speech, and Language Processing, 29:3504-3514, 2021. 8", + "[6] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In IEEE Conference on Computer Vision and Pattern Recognition, pages 248-255, 2009. 7", + "[7] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018. 7", + "[8] Han Fang, Pengfei Xiong, Luhui Xu, and Yu Chen. Clip2video: Mastering video-text retrieval via image clip. arXiv preprint arXiv:2106.11097, 2021. 2, 4, 8", + "[9] Tsu-Jui Fu, Linjie Li, Zhe Gan, Kevin Lin, William Yang Wang, Lijuan Wang, and Zicheng Liu. Violet: End-to-end video-language transformers with masked visual-token modeling. arXiv preprint arXiv:2111.12681, 2021. 2, 4, 6, 8", + "[10] Yuying Ge, Yixiao Ge, Xihui Liu, Dian Li, Ying Shan, Xiaohu Qie, and Ping Luo. Bridging video-text retrieval with multiple choice questions. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 16167-16176, 2022. 1", + "[11] Jiaxi Gu, Xiaojun Meng, Guansong Lu, Lu Hou, Minzhe Niu, Hang Xu, Xiaodan Liang, Wei Zhang, Xin Jiang, and Chunjing Xu. Wukong: 100 million large-scale chinese cross-modal pre-training dataset and a foundation framework. arXiv preprint arXiv:2202.06767, 2022. 2, 4, 6, 7", + "[12] Yannis Kalantidis, Mert Bulent Sariyildiz, Noe Pion, Philippe Weinzaepfel, and Diane Larlus. Hard negative mixing for contrastive learning. Advances in Neural Information Processing Systems, 33:21798-21809, 2020. 2", + "[13] Ranjay Krishna, Yuke Zhu, Oliver Groth, Justin Johnson, Kenji Hata, Joshua Kravitz, Stephanie Chen, Yannis Kalantidis, Li-Jia Li, David A Shamma, et al. Visual genome:" + ], + "bbox": [ + 78, + 114, + 472, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Connecting language and vision using crowdsourced dense image annotations. International Journal of Computer Vision, 123(1):32-73, 2017. 4, 6", + "[14] Chenyi Lei, Shixian Luo, Yong Liu, Wanggui He, Jiamang Wang, Guoxin Wang, Haihong Tang, Chunyan Miao, and Houqiang Li. Understanding chinese video and language via contrastive multimodal pre-training. In Proceedings of the ACM International Conference on Multimedia, pages 2567-2576, 2021. 1, 2, 3, 4", + "[15] Jie Lei, Linjie Li, Luowei Zhou, Zhe Gan, Tamara L Berg, Mohit Bansal, and Jingjing Liu. Less is more: Clipbert for video-and-language learning via sparse sampling. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 7331-7341, 2021. 1, 2, 4, 5, 6, 8", + "[16] Chunyuan Li, Xiujun Li, Lei Zhang, Baolin Peng, Mingyuan Zhou, and Jianfeng Gao. Self-supervised pre-training with hard examples improves visual representations. arXiv preprint arXiv:2012.13493, 2020. 2", + "[17] Dongxu Li, Junnan Li, Hongdong Li, Juan Carlos Niebles, and Steven CH Hoi. Align and prompt: Video-and-language pre-training with entity prompts. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 4953-4963, 2022. 1, 2, 4, 8", + "[18] Junnan Li, Ramprasaath Selvaraju, Akhilesh Gotmare, Shafiq Joty, Caiming Xiong, and Steven Chu Hong Hoi. Align before fuse: Vision and language representation learning with momentum distillation. Advances in neural information processing systems, 34:9694-9705, 2021. 2", + "[19] Linjie Li, Yen-Chun Chen, Yu Cheng, Zhe Gan, Licheng Yu, and Jingjing Liu. Hero: Hierarchical encoder for video+language omni-representation pre-training. arXiv preprint arXiv:2005.00200, 2020. 1", + "[20] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Proceedings of the European Conference on Computer Vision, pages 740-755. Springer, 2014. 6", + "[21] Ze Liu, Jia Ning, Yue Cao, Yixuan Wei, Zheng Zhang, Stephen Lin, and Han Hu. Video swim transformer. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 3202-3211, 2022. 7", + "[22] Huaishao Luo, Lei Ji, Botian Shi, Haoyang Huang, Nan Duan, Tianrui Li, Jason Li, Taroon Bharti, and Ming Zhou. Univl: A unified video and language pre-training model for multimodal understanding and generation. arXiv preprint arXiv:2002.06353, 2020. 1, 6", + "[23] Huaishao Luo, Lei Ji, Ming Zhong, Yang Chen, Wen Lei, Nan Duan, and Tianrui Li. Clip4clip: An empirical study of clip for end to end video clip retrieval. arXiv preprint arXiv:2104.08860, 2021. 2, 4, 8", + "[24] Jianjie Luo, Yehao Li, Yingwei Pan, Ting Yao, Hongyang Chao, and Tao Mei. Coco-bert: Improving video-language pre-training with contrastive cross-modal matching and denoising. In Proceedings of the ACM International Conference on Multimedia, pages 5600-5608, 2021. 2, 4", + "[25] Antoine Miech, Dimitri Zhukov, Jean-Baptiste Alayrac, Makarand Tapaswi, Ivan Laptev, and Josef Sivic." + ], + "bbox": [ + 503, + 92, + 893, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "14823", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Howto100m: Learning a text-video embedding by watching hundred million narrated video clips. In Proceedings of the IEEE International Conference on Computer Vision, pages 2630-2640, 2019. 1, 2, 3, 4", + "[26] Liqiang Nie, Leigang Qu, Dai Meng, Min Zhang, Qi Tian, and Alberto Del Bimbo. Search-oriented microvideo captioning. In Proceedings of the ACM International Conference on Multimedia, pages 3234-3243, 2022. 1, 2, 3, 4, 8", + "[27] Devshree Patel, Ratnam Parikh, and Yesha Shastri. Recent advances in video question answering: A review of datasets and methods. In International Conference on Pattern Recognition, pages 339-356. Springer, 2021. 1", + "[28] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 2, 5", + "[29] Joshua David Robinson, Ching-Yao Chuang, Suvrit Sra, and Stefanie Jegelka. Contrastive learning with hard negative samples. In International Conference on Learning Representations, 2020. 2", + "[30] Paul Hongsuck Seo, Arsha Nagrani, Anurag Arnab, and Cordelia Schmid. End-to-end generative pretraining for multimodal video captioning. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 17959-17968, 2022. 1, 2, 4, 8", + "[31] Piyush Sharma, Nan Ding, Sebastian Goodman, and Radu Soricut. Conceptual captions: A cleaned, hypernymed, image alt-text dataset for automatic image captioning. In Proceedings of the Annual Meeting of the Association for Computational Linguistics, pages 2556–2565, 2018. 6", + "[32] Petru Soviany, Radu Tudor Ionescu, Paolo Rota, and Nicu Sebe. Curriculum learning: A survey. International Journal of Computer Vision, pages 1-40, 2022. 2, 3", + "[33] Chen Sun, Austin Myers, Carl Vondrick, Kevin Murphy, and Cordelia Schmid. Videobert: A joint model for video and language representation learning. In Proceedings of the IEEE International Conference on Computer Vision, pages 7464-7473, 2019. 2, 4", + "[34] Guanglu Sun, Lili Liang, Tianlin Li, Bo Yu, Meng Wu, and Bolun Zhang. Video question answering: a survey of models and datasets. Mobile Networks and Applications, 26(5):1904-1937, 2021. 1", + "[35] Alex Jinpeng Wang, Yixiao Ge, Rui Yan, Yuying Ge, Xudong Lin, Guanyu Cai, Jianping Wu, Ying Shan, Xiaohu Qie, and Mike Zheng Shou. All in one: Exploring unified video-language pre-training. arXiv preprint arXiv:2203.07303, 2022. 2, 4", + "[36] Jinpeng Wang, Yixiao Ge, Guanyu Cai, Rui Yan, Xudong Lin, Ying Shan, Xiaohu Qie, and Mike Zheng Shou. Object-aware video-language pre-training for retrieval. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 3313-3322, 2022. 1, 2, 4, 8" + ], + "bbox": [ + 78, + 90, + 468, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[37] Xin Wang, Yudong Chen, and Wenwu Zhu. A survey on curriculum learning. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2021. 2, 3", + "[38] Xin Wang, Jiawei Wu, Junkun Chen, Lei Li, Yuan-Fang Wang, and William Yang Wang. Vatex: A large-scale, high-quality multilingual dataset for video-and-language research. In Proceedings of the IEEE International Conference on Computer Vision, pages 4581-4591, 2019. 3, 5, 6, 8", + "[39] Dejing Xu, Zhou Zhao, Jun Xiao, Fei Wu, Hanwang Zhang, Xiangnan He, and Yueting Zhuang. Video question answering via gradually refined attention over appearance and motion. In Proceedings of the ACM International Conference on Multimedia, pages 1645-1653, 2017. 8", + "[40] Hu Xu, Gargi Ghosh, Po-Yao Huang, Prahal Arora, Masoumeh Aminzadeh, Christoph Feichtenhofer, Florian Metze, and Luke Zettlemoyer. Vlm: Task-agnostic video-language model pre-training for video understanding. arXiv preprint arXiv:2105.09996, 2021. 2", + "[41] Jun Xu, Tao Mei, Ting Yao, and Yong Rui. Msr-vtt: A large video description dataset for bridging video and language. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 5288-5296, 2016. 5, 6, 8", + "[42] Antoine Yang, Antoine Miech, Josef Sivic, Ivan Laptev, and Cordelia Schmid. Just ask: Learning to answer questions from millions of narrated videos. In Proceedings of the IEEE International Conference on Computer Vision, pages 1686-1697, 2021. 1", + "[43] Jianwei Yang, Yonatan Bisk, and Jianfeng Gao. Taco: Token-aware cascade contrastive learning for video-text alignment. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11562-11572, 2021. 2, 3, 5, 6", + "[44] Shengyu Zhang, Ziqi Tan, Jin Yu, Zhou Zhao, Kun Kuang, Jie Liu, Jingren Zhou, Hongxia Yang, and Fei Wu. Poet: Product-oriented video captioner for e-commerce. In Proceedings of the ACM International Conference on Multimedia, pages 1292-1301, 2020. 3", + "[45] Ziqi Zhang, Yuxin Chen, Zongyang Ma, Zhongang Qi, Chunfeng Yuan, Bing Li, Ying Shan, and Weiming Hu. Create: A benchmark for chinese short video retrieval and title generation. arXiv preprint arXiv:2203.16763, 2022. 1, 2, 3, 4, 8", + "[46] Wei Zhu. Mvp-bert: Multi-vocab pre-training for chinese bert. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing: Student Research Workshop, pages 260-269, 2021. 8" + ], + "bbox": [ + 501, + 92, + 890, + 766 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "14824", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2023/CNVid-3.5M_ Build, Filter, and Pre-Train the Large-Scale Public Chinese Video-Text Dataset/3819ec34-d037-4584-a0be-1f25bc0cb37d_model.json b/2023/CNVid-3.5M_ Build, Filter, and Pre-Train the Large-Scale Public Chinese Video-Text Dataset/3819ec34-d037-4584-a0be-1f25bc0cb37d_model.json new file mode 100644 index 0000000000000000000000000000000000000000..71c478949baca49ce02ad104a9f8d45709759a6b --- /dev/null +++ b/2023/CNVid-3.5M_ Build, Filter, and Pre-Train the Large-Scale Public Chinese Video-Text Dataset/3819ec34-d037-4584-a0be-1f25bc0cb37d_model.json @@ -0,0 +1,2453 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.234, + 0.131, + 0.738, + 0.175 + ], + "angle": 0, + "content": "CNVid-3.5M: Build, Filter, and Pre-train the Large-scale Public Chinese Video-text Dataset" + }, + { + "type": "text", + "bbox": [ + 0.107, + 0.203, + 0.91, + 0.223 + ], + "angle": 0, + "content": "Tian Gan\\(^{1*}\\), Qing Wang\\(^{2*}\\), Xingning Dong\\(^{1}\\), Xiangyuan Ren\\(^{2}\\), Liqiang Nie\\(^{3}\\), Qingpei Guo\\(^{2\\dagger}\\)" + }, + { + "type": "text", + "bbox": [ + 0.201, + 0.225, + 0.352, + 0.242 + ], + "angle": 0, + "content": "1Shandong University," + }, + { + "type": "text", + "bbox": [ + 0.383, + 0.226, + 0.47, + 0.242 + ], + "angle": 0, + "content": "\\(^{2}\\)Ant Group," + }, + { + "type": "text", + "bbox": [ + 0.501, + 0.226, + 0.792, + 0.242 + ], + "angle": 0, + "content": "\\(^{3}\\)Harbin Institute of Technology (Shenzhen)" + }, + { + "type": "text", + "bbox": [ + 0.182, + 0.246, + 0.351, + 0.259 + ], + "angle": 0, + "content": "gantian@sdu.edu.cn," + }, + { + "type": "text", + "bbox": [ + 0.374, + 0.246, + 0.569, + 0.259 + ], + "angle": 0, + "content": "wq176625@antgroup.com," + }, + { + "type": "text", + "bbox": [ + 0.591, + 0.246, + 0.824, + 0.259 + ], + "angle": 0, + "content": "dongxingning1998@gmail.com" + }, + { + "type": "text", + "bbox": [ + 0.15, + 0.263, + 0.836, + 0.277 + ], + "angle": 0, + "content": "xiangyuan.rxy@antgroup.com, nieliqiang@gmail.com, qingpei.gqp@antgroup.com" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.312, + 0.314, + 0.329 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.344, + 0.473, + 0.737 + ], + "angle": 0, + "content": "Owing to well-designed large-scale video-text datasets, recent years have witnessed tremendous progress in videotext pre-training. However, existing large-scale video-text datasets are mostly English-only. Though there are certain methods studying the Chinese video-text pre-training, they pre-train their models on private datasets whose videos and text are unavailable. This lack of large-scale public datasets and benchmarks in Chinese hampers the research and downstream applications of Chinese video-text pretraining. Towards this end, we release and benchmark CNVid-3.5M, a large-scale public cross-modal dataset containing over 3.5M Chinese video-text pairs. We summarize our contributions by three verbs, i.e., \"Build\", \"Filter\", and \"Pre-train\": 1) To build a public Chinese video-text dataset, we collect over 4.5M videos from the Chinese websites. 2) To improve the data quality, we propose a novel method to filter out 1M weakly-paired videos, resulting in the CNVid-3.5M dataset. And 3) we benchmark CNVid-3.5M with three mainstream pixel-level pre-training architectures. At last, we propose the Hard Sample Curriculum Learning strategy to promote the pre-training performance. To the best of our knowledge, CNVid-3.5M is the largest public video-text dataset in Chinese, and we provide the first pixel-level benchmarks for Chinese video-text pretraining. The dataset, codebase, and pre-trained models are available at https://github.com/CNVid/CNVid-3.5M." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.765, + 0.21, + 0.781 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.79, + 0.47, + 0.867 + ], + "angle": 0, + "content": "Owing to well-designed large-scale datasets, video-text pre-training [15, 17, 19] has achieved superior performance in various downstream tasks, such as video-text retrieval [4, 10, 36], video question answering [27, 34, 42], and video captioning [1, 22, 30]. However, recent large-scale video-" + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.311, + 0.631, + 0.415 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.639, + 0.315, + 0.772, + 0.414 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.777, + 0.315, + 0.901, + 0.415 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.427, + 0.893, + 0.469 + ], + "angle": 0, + "content": "Figure 1. Here presents the motivations of this paper, based on which we highly summarize our contributions with three verbs: \"Build\", \"Filter\", and \"Pre-train\"." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.486, + 0.893, + 0.592 + ], + "angle": 0, + "content": "text datasets are mostly English-only (e.g., Howto100M [25] and WebVid-2.5M [4]). Though some methods [14,26, 45] turn to study the Chinese video-text pre-training, they pre-train their models on private datasets whose videos and text are unavailable. Therefore, the research towards Chinese video-text pre-training is still in its infancy due to the lack of large-scale public datasets." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.595, + 0.895, + 0.776 + ], + "angle": 0, + "content": "Towards this problem, directly translating English text into Chinese is a simple solution. However, it may result in unacceptable performance degradation for two reasons: 1) Translation errors are inevitable. Moreover, since most of the large-scale video-text datasets employ the Automatic Speech Recognition (ASR) system to generate text, the language translator would amplify the error from the incomplete and noisy ASR text. And 2) there remains an intrinsic linguistic gap between English and Chinese. Many widely-used English idioms and slang can hardly find their Chinese counterparts, leading some translated text incomprehensible and even contrary to the original meaning." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.778, + 0.894, + 0.853 + ], + "angle": 0, + "content": "In this paper, we aim to release and benchmark a large-scale public Chinese video-text dataset to facilitate future researchers and the community. As illustrated in Figure 1, three verbs could highly summarize our contributions, i.e., \"Build\", \"Filter\", and \"Pre-train\"." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.893, + 0.901 + ], + "angle": 0, + "content": "To build a large-scale Chinese video-text dataset, we collect over 4.5M videos from Chinese websites. All videos are associated with user-uploaded titles and ASR text." + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.875, + 0.208, + 0.888 + ], + "angle": 0, + "content": "*Equal contribution." + }, + { + "type": "page_footnote", + "bbox": [ + 0.097, + 0.888, + 0.222, + 0.9 + ], + "angle": 0, + "content": "† Corresponding author." + }, + { + "type": "list", + "bbox": [ + 0.096, + 0.875, + 0.222, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "14815" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.333 + ], + "angle": 0, + "content": "We filter out the weakly-paired data by a novel method to improve the data quality. As some work [25,26] pointed out, the pre-training performance would suffer from the noisy ASR text that fails to accurately describe the video content. Unfortunately, the problem is raised with few practical solutions. Therefore, we employ a well-trained image-text model to evaluate the video-text consistency for three reasons: 1) The text information in existing image-text datasets [11] are usually manually-written titles or captions, whose consistency is guaranteed. 2) Some video-text pre-training architectures [23,35] are based upon image-text ones. And 3) it is cheap and efficient to \"hire\" a well-trained model to check millions of videos. In this way, we filter out about 1M weakly-paired videos based on the balance between the pre-training performance and efficiency, deriving the proposed CNVid-3.5M dataset." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.334, + 0.473, + 0.59 + ], + "angle": 0, + "content": "We pre-train various models to benchmark our CNVid-3.5M dataset. Current video-text pre-training methods could be roughly divided into two categories: 1) feature-level pre-training methods [24, 33, 40] that employ offline video and textual feature extractors, and 2) pixel-level ones [4, 15, 36] that learn cross-modal representations end-to-end from raw videos and text. Since there remain domain gaps between pre-training datasets and frozen feature extractors, pixel-level pre-training methods usually achieve better performance and have been widely employed in recent years. However, existing Chinese video-text pre-training methods [14, 26, 45] are all feature-level ones pre-trained on private datasets, limiting their contributions on the development of Chinese video-text pre-training techniques. Hence, we adopt three mainstream pixel-level pre-training frameworks, which are the first pixel-level benchmarks for Chinese video-text pre-training." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.591, + 0.473, + 0.833 + ], + "angle": 0, + "content": "Moreover, we propose the novel Hard Sample Curriculum Learning strategy to promote the pre-training performance. Since contrastive learning is a significant component in video-text pre-training, some methods [16, 18, 43] employ the hard sample mining [12, 29] strategy to promote the cross-modal alignment. However, hard sample mining would bring side effects to pre-training when the model is far from convergence. Suppose that a model is incapable of discriminating the ground-truth video-text pairs, recklessly introducing hard negatives would lead to the sub-optimal performance. Inspired by the curriculum learning [32, 37] strategy that \"starts small\" and gradually \"learns hard\", we combine these two strategies and propose the novel Hard Sample Curriculum Learning (HSCL). By gradually and smoothly emphasizing those hard samples, HSCL could effectively improve the pre-training performance." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.834, + 0.415, + 0.848 + ], + "angle": 0, + "content": "Our contributions are summarized in four folds:" + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.856, + 0.47, + 0.901 + ], + "angle": 0, + "content": "- To fill in the blank of large-scale public Chinese video-text datasets, we collect over 4.5M videos associated with titles and ASR text from the websites." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.092, + 0.892, + 0.136 + ], + "angle": 0, + "content": "- To improve the data quality, we propose a novel method to filter out 1M weakly-paired videos, resulting in the CNVid-3.5M dataset." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.137, + 0.892, + 0.182 + ], + "angle": 0, + "content": "- To promote the pre-training performance, we propose the novel Hard Sample Curriculum Learning strategy for better cross-modal contrastive learning." + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.183, + 0.894, + 0.273 + ], + "angle": 0, + "content": "- To the best of our knowledge, the constructed CNVid-3.5M is the largest public Chinese video-text dataset. Moreover, we provide the first Chinese pixel-level benchmarks based on CNVid-3.5M. The dataset, codebase, and benchmarks are available at https://github.com/CNVid/CNVid-3.5M." + }, + { + "type": "list", + "bbox": [ + 0.518, + 0.092, + 0.894, + 0.273 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.297, + 0.642, + 0.314 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.323, + 0.892, + 0.489 + ], + "angle": 0, + "content": "Video-Text Pre-training and Fine-Tuning. Video-text pre-training has attracted increasing interest in recent years, which could be roughly divided into two categories, feature-level pre-training methods [24, 33, 40] and pixel-level ones [15, 17, 35]. The former approaches employ offline visual and textual features extracted from frozen models, while the latter ones learn cross-modal representations from raw videos and text in an end-to-end manner. Since there remain domain gaps between pre-training datasets and frozen feature extractors, pixel-level pre-training methods usually achieve better performance than the feature-level ones." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.49, + 0.892, + 0.657 + ], + "angle": 0, + "content": "There are three mainstream architectures in pixel-level video-text pre-training methods, i.e., three-encoders-fusion (3-E-F) [15, 17, 30], two-encoders-fusion (2-E-F) [9], and twin-towers-crossed (T-T-C) [8, 23, 36]. 3-E-F methods contain three separate encoders to model visual, textual, and cross-modal features; 2-E-F methods simplify the 3-E-F architecture by employing a shared encoder to embed textual and cross-modal features; while T-T-C methods are mainly based on the widely-adopted CLIP [28] architecture, which includes a couple of text and visual encoders to learn the cross-modal alignment by contrastive learning." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.658, + 0.892, + 0.734 + ], + "angle": 0, + "content": "Though the pixel-level paradigm has been widely adopted recently, existing Chinese video-text models [14,26,45] are all feature-level ones pre-trained on private datasets. In this work, we provide the first pixel-level benchmarks pre-trained on our CNVid-3.5M dataset." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.734, + 0.892, + 0.869 + ], + "angle": 0, + "content": "Video-Text Dataset. Well-designed large-scale datasets are a prerequisite for the successful applications of the \"Pre-training & Fine-tuning\" paradigm. There exist a few large-scale public English datasets (e.g., Howto100M [25], WebVid-2.5M [4]) for video-text pre-training. However, existing large-scale Chinese video datasets (ALIVOL [14], Kwai-SVC [26], CREATE [45]) are all private, whose videos and text are unavailable to the public. Therefore, we aimed to release a public one to facilitate future researchers." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.87, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Hard Sample Mining in Contrastive Learning. Hard Sample Mining [12, 29] is an effective strategy to promote" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "14816" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.078, + 0.088, + 0.917, + 0.278 + ], + "angle": 0, + "content": "
Dataset Name#(Videos)#(Text)Duration (hours)Video SourceText TypeAvailability
Off-FeatRaw-VidState
Datasets for Downstream Chinese Video-text Tasks
VATEX [38]41,269825,380115kinetics-600Caption✓✓All-public
BFVD [44]43,16643,166140E-CommerceTitle✓-Partly-public
FFVD [44]32,76332,763252E-CommerceTitle✓-Partly-public
CREATE210K [45]216,303268,5931,800Open WebsitesCaption--NOT-public
Kwai-SVC [26]222,077143,5693,500Open WebsitesTitle & ASR✓-Partly-public
Datasets for Chinese Video-text Pre-training
ALIVOL-10M [14]10,300,00011,000,00098,801E-CommerceTitle--NOT-public
Kwai-SVC-11M [26]11,075,0843,931,879177,200Open WebsitesTitle & ASR--NOT-public
CREATE-10M [45]10,000,00010,000,00083,000Open WebsitesTitle--NOT-public
CNVid-3.5M (ours)3,508,1203,508,12035,414Open WebsitesTitle & ASR✓✓All-public
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.29, + 0.892, + 0.318 + ], + "angle": 0, + "content": "Table 1. Data analysis of CNVid-3.5M and other Chinese video-text datasets. \"Off-Feat\" and \"Raw-Vid\" represent whether offline features and raw videos are available. As far as we know, CNVid-3.5M is the largest public video-text dataset in Chinese." + }, + { + "type": "image", + "bbox": [ + 0.081, + 0.342, + 0.479, + 0.567 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.579, + 0.47, + 0.622 + ], + "angle": 0, + "content": "Figure 2. An example of video-text pairs in our CNVid-3.5M dataset. We present three types of text information: 1) Caption (manually written), 2) Title (user uploaded), and 3) ASR text." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.635, + 0.47, + 0.816 + ], + "angle": 0, + "content": "the performance of contrastive learning. In video-text pretraining, TACo [43] proposes a cascade sampling method to construct hard negative examples for better cross-modal alignment. However, we conjectured that hard sample mining may hinder the video-text pre-training at the beginning, where the model is far from convergence and could hardly discriminate the ground-truth video-text pairs. Therefore, we introduced curriculum learning [32, 37] and proposed the Hard Sample Curriculum Learning (HSCL) strategy. HSCL would guide the pre-training procedure from \"starting small\" gradually to \"learning hard\", mitigating the side effect of conventional hard sample mining methods." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.83, + 0.317, + 0.846 + ], + "angle": 0, + "content": "3. The CNVid-3.5M Dataset" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.47, + 0.903 + ], + "angle": 0, + "content": "To fill in the blank of public Chinese video-text datasets, we release CNVid-3.5M, a large-scale cross-modal dataset containing over 3.5M Chinese video-text pairs." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.344, + 0.667, + 0.361 + ], + "angle": 0, + "content": "3.1. Dataset Building" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.368, + 0.892, + 0.534 + ], + "angle": 0, + "content": "We collect raw videos from Douyin*, a popular Chinese social website, where millions of active users upload various types of videos every day. To enrich the data diversity, we do not limit video categories like HowTo100M [25] (mainly instructional videos) and cover various topics (e.g., food, entertainment, and technology). I.e., we first form up a set of keywords and then leverage them to search for videos and corresponding captions. Moreover, we span the data created time at a range of 3 years and filter the abnormal ones whose videos could not be played. In this way, we collect over 4.5M raw videos in Chinese." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.535, + 0.892, + 0.567 + ], + "angle": 0, + "content": "Besides, the supplementary material will present more details about fairness and privacy for safety and ethics." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.578, + 0.774, + 0.594 + ], + "angle": 0, + "content": "3.2. Captions, Titles, and ASR Text" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.602, + 0.892, + 0.662 + ], + "angle": 0, + "content": "As CREATE [45] pointed out, three types of text information are widely-employed in current Chinese video-text datasets. We show an example in Figure 2, based on which we summarize their characteristics and applications:" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.663, + 0.892, + 0.783 + ], + "angle": 0, + "content": "1) Captions are manually written by human annotators. They could accurately describe the video content in a formal manner. However, obtaining a large number of captions is extremely expensive. Therefore, captions are mainly employed in small-scale downstream video-text datasets like VATEX [38]. Moreover, captions are indeed global descriptions of the whole videos. Therefore, they could hardly describe the frame-level information of videos." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.785, + 0.892, + 0.875 + ], + "angle": 0, + "content": "2) Titles are written and uploaded by users. Compared with captions, it is not unfeasible to obtain millions of user-uploaded titles. However, to increase the attractiveness and click rate, users may over-polish titles with some \"appealing but redundant\" words, which may decrease the video-text consistency. Similar to captions, titles are video-level" + }, + { + "type": "page_footnote", + "bbox": [ + 0.518, + 0.888, + 0.754, + 0.901 + ], + "angle": 0, + "content": "*Official Website: https://www.douyin.com/" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "14817" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.137 + ], + "angle": 0, + "content": "descriptions rather than frame-level ones. Therefore, titles and captions are mainly employed in feature-level videotext pre-training methods rather than pixel-level ones." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.137, + 0.468, + 0.228 + ], + "angle": 0, + "content": "3) ASR text is obtained by the ASR toolkit. Similar to titles, it is easy to obtain a large number of ASR text. However, some ASR text is incomplete, noisy, and even faulty, which may decrease the pre-training performance. Besides, ASR text is a frame-level description, which is widely employed in the pixel-level video-text pre-training." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.228, + 0.468, + 0.304 + ], + "angle": 0, + "content": "Following previous methods [25, 26], we collect user-uploaded titles and obtain the ASR text of each video. Since we provide both video-level titles and frame-level ASR text, our dataset could support the video-text pre-training in both feature-level and pixel-level manner." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.314, + 0.34, + 0.331 + ], + "angle": 0, + "content": "3.3. Weakly-paired Data Filtering" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.338, + 0.468, + 0.579 + ], + "angle": 0, + "content": "As some work [25, 26] has pointed out, some ASR text may not match with associated videos (e.g., videos have background music), which would decrease the pre-training performance. However, the problem is only raised with few practical solutions. Therefore, we propose a novel weakly-paired data filtering strategy by leveraging a well-trained image-text model to automatically evaluate the video-text consistency for three reasons: 1) The text information in existing image-text datasets [11, 13] are manually-written titles or captions, whose quality and consistency are guaranteed. 2) Pixel-level pre-training methods [8, 15, 23] usually sample several frames from raw videos to serve as the visual input, whose frameworks are closely associated with image-text pre-trained models. 3) Compared with manually checking millions of videos, the proposed filtering strategy remarkably reduces the cost and improves the efficiency." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.58, + 0.468, + 0.671 + ], + "angle": 0, + "content": "Specifically, we first pre-train an image-text model on Wukong-100M [11]. We then leverage it to calculate the similarity score between videos and ASR text, and filter out the last 1M videos by their ranks. In this way, we obtain the CNVid-3.5M dataset, based on which we provide three mainstream pixel-level pre-training benchmarks." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.681, + 0.364, + 0.697 + ], + "angle": 0, + "content": "3.4. Comparison of Dataset Statistics" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.704, + 0.468, + 0.901 + ], + "angle": 0, + "content": "Table 1 presents the statistics of CNVid-3.5M and other Chinese video-text datasets. To the best of our knowledge, CNVid-3.5M is the largest public dataset for Chinese video-text pre-training, with a total of 3.5M videos that last for 35.4K hours. Though ALIVOL-10M [14], KwaiSVC-11M [26], and CREATE-10M [45] contain more Chinese videos than ours, yet they do not release raw videos and text information to the public. Moreover, these three methods only provide conventional feature-level pretrained models, whose performance highly depends on frozen feature extractors. Towards this end, we provide the first pixel-level benchmarks based on our CNVid-3.5M dataset, which would pave the way for future research on" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.713, + 0.107 + ], + "angle": 0, + "content": "Chinese video-text pre-training." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.107, + 0.892, + 0.168 + ], + "angle": 0, + "content": "Moreover, we present additional statistics of our CNVid-3.5M dataset in the supplementary material, including the distributions of topics, keywords, video durations, and Part-of-Speech (POS) tags of the ASR text." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.183, + 0.806, + 0.201 + ], + "angle": 0, + "content": "4. Pixel-level Pre-training Paradigm" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.209, + 0.892, + 0.376 + ], + "angle": 0, + "content": "Since mainstream pixel-level video-text pre-training methods [17, 23, 35] extract cross-modal representations from raw data directly, they usually outperform the conventional feature-level ones [2, 24, 33]. Therefore, we benchmark our CNVid-3.5M dataset with three mainstream pixel-level architectures. As illustrated in Figure 3, we split the pixel-level pre-training process into three steps, namely: 1) data pre-processing to process raw videos and text into patches and tokens, 2) model architectures to generate cross-modal representations, and 3) proxy tasks to determine overall pre-training objectives." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.387, + 0.692, + 0.403 + ], + "angle": 0, + "content": "4.1. Data Pre-processing" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.41, + 0.892, + 0.581 + ], + "angle": 0, + "content": "Given a mini-batch (denoted as \\(\\mathcal{B}\\)) of videos \\(\\{V_i\\}_{i=1}^{|B|}\\) and their corresponding ASR text \\(\\{T_i\\}_{i=1}^{|B|}\\), pixel-level pre-training methods would first sparsely (and randomly) sample \\(N_v\\) frames from each video. We then slice each frame into patches, obtaining visual patch embeddings \\(\\mathbf{P} \\in \\mathbb{R}^{N_v * N_p * d}\\), where \\(N_p\\) is the number of sliced patches, and \\(d\\) is the dimension of the embedding. Simultaneously, a BERT embedder is employed to process the text \\(T\\) into fixed-length word embeddings \\(\\mathbf{W} = [\\mathbf{w}^{cls}, \\mathbf{w}^1, \\mathbf{w}^2, \\dots, \\mathbf{w}^{N_t - 1}]\\), where \\(\\mathbf{W} \\in \\mathbb{R}^{N_t * d}\\), \\(N_t\\) is the length of text tokens." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.592, + 0.694, + 0.607 + ], + "angle": 0, + "content": "4.2. Model Architectures" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.616, + 0.892, + 0.692 + ], + "angle": 0, + "content": "There are three mainstream pixel-level video-text pretraining architectures, namely three-encoders-fusion [15, 17, 30], two-encoders-fusion [9], and twin-towers-crossed [8, 23, 36]. We benchmark CNVid-3.5M with these three pixel-level pre-training architectures." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.692, + 0.892, + 0.829 + ], + "angle": 0, + "content": "Three-encoders-fusion (3-E-F) pre-training methods contain a visual encoder \\( E_{vis} \\), a text encoder \\( E_{txt} \\), and a cross-modal encoder \\( E_{mul} \\). After obtaining visual features \\( \\mathbf{V} = E_{vis}(\\mathbf{P}) \\) and textual features \\( \\mathbf{T} = E_{txt}(\\mathbf{W}) \\), we concatenate these two-modal features into \\( [\\mathbf{T},\\mathbf{V}] = [\\mathbf{t}^{cls},\\mathbf{t}^1,\\mathbf{t}^2,\\dots ,\\mathbf{t}^{N_t - 1},\\mathbf{v}^{cls},\\mathbf{v}^1,\\mathbf{v}^2,\\dots ,\\mathbf{v}^{N_v}] \\). Following CLIP4Clip [23], we set position embeddings to zero, and obtain cross-modal features \\( \\mathbf{M} = E_{mul}([\\mathbf{T},\\mathbf{V}]) \\). This forwarding process could be formulated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.579, + 0.841, + 0.891, + 0.859 + ], + "angle": 0, + "content": "\\[\n\\mathbf {M} = E _ {m u l} ([ E _ {t x t} (\\mathbf {W}), E _ {v i s} (\\mathbf {P}) ]), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.87, + 0.891, + 0.901 + ], + "angle": 0, + "content": "where \\(\\mathbf{M} = [\\mathbf{m}^{cls},\\mathbf{m}^1,\\mathbf{m}^2,\\dots ,\\mathbf{m}^{N_t + N_v}]\\), and \\([\\cdot ,\\cdot ]\\) denotes the concatenation operation." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "14818" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.08, + 0.089, + 0.895, + 0.248 + ], + "angle": 0, + "content": "
1. Data Pre-processing2. Three Types of Model Architecture3. Proxy Tasks
Text Embedding\nC | | | | BERT Embedder\n| | | | | [CLS] a cat plays piano\n[CLS] 猫弹钢琴\nASR Caption\n| | | | | Input VideoThree-Encoders-Fusion\nCross-Modal Feature\n| | | | | Cross-Modal Encoder\n| | | | | Emut | | | | Text FeatureEbert\nShared Encoder\n| | | | | C | | | | Video Feature\n| | | | | Shared Encoder\n| | | | | C | | | | Video Encoder\n| | | | | C | | | | Cross-Modal EncodingTwo-Encoders-Fusion\nCross-Modal Feature\n| | | | | C | | | | Video Feature\n| | | | | C | | | | Video Encoder\n| | | | | C | | | | EvisTwin-Towers-Crossed\nContrastive Learning\n| | | | | C | | | | Video Feature\n| | | | | C | | | | Video Encoder\n| | | | | C | | | | EvisMasked Language Modeling (MLM)\nFC\n[LMSK]\n[C | | | | F | | | | F | | | | F | | | | F | | | | F | | | | F | | | | F | | | | F | | | | F | | | | F | | | | F | | | | F | | | | F | | | | F | | | | F | | | | F | | | | F | | | | F | | | | F | | | | F | | | |
" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.255, + 0.895, + 0.285 + ], + "angle": 0, + "content": "Figure 3. The pipeline of mainstream pixel-level video-text pre-training methods, which contains three key steps: 1) Data pre-processing, 2) model architectures, and 3) proxy tasks. Note that we employ three mainstream architectures to enrich pixel-level pre-training benchmarks." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.296, + 0.475, + 0.387 + ], + "angle": 0, + "content": "Two-encoders-fusion (2-E-F) pre-training methods simplify the 3-E-F architecture by employing a shared BERT-type encoder \\( E_{bert} \\) to embed textual and cross-modal features simultaneously. The pipeline of obtaining textual features \\( \\mathbf{T} \\) and cross-modal features \\( \\mathbf{M} \\) could be formulated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.171, + 0.395, + 0.47, + 0.437 + ], + "angle": 0, + "content": "\\[\n\\left\\{ \\begin{array}{l} \\mathbf {T} = E _ {b e r t} (\\mathbf {W}), \\\\ \\mathbf {M} = E _ {b e r t} ([ \\mathbf {W}, E _ {v i s} (\\mathbf {P}) ]). \\end{array} \\right. \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.446, + 0.473, + 0.553 + ], + "angle": 0, + "content": "Twin-towers-crossed \\((T-T-C)\\) pre-training methods mainly follow the conventional CLIP [28] architecture. After obtaining visual and textual features, \\(T-T-C\\) methods employ contrastive learning to optimize the twin encoders \\(E_{vis}\\) and \\(E_{txt}\\) for better cross-modal alignment. Note that \\(T-T-C\\) methods do not generate cross-modal video representations like \\(3-E-F\\) and \\(2-E-F\\)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.562, + 0.21, + 0.578 + ], + "angle": 0, + "content": "4.3. Proxy Tasks" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.585, + 0.47, + 0.661 + ], + "angle": 0, + "content": "Proxy tasks are crucial for video-text pre-training as they directly determine the final optimization objectives. Following the conventional protocol [4, 15, 43], we adopt two widely-employed proxy tasks, i.e., Masked Language Modeling (MLM) and Video-Text Matching (VTM)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.661, + 0.471, + 0.737 + ], + "angle": 0, + "content": "MLM first masks out a certain percentage of words in a given sentence, and then forces the model to restore these clozes according to visual and textual cues. MLM is calculated twice for textual features \\(\\mathbf{T}(\\mathcal{L}_1)\\) and cross-modal features \\(\\mathbf{M}(\\mathcal{L}_2)\\) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.161, + 0.746, + 0.47, + 0.784 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {1} = \\frac {1}{| \\mathcal {Q} |} \\sum_ {q \\in \\mathcal {Q}} \\mathcal {L} _ {C E} \\left(y ^ {q}, \\Theta_ {1} \\left(\\mathbf {t} ^ {q}\\right)\\right), \\tag {3}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.158, + 0.796, + 0.47, + 0.834 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {2} = \\frac {1}{| \\mathcal {Q} |} \\sum_ {q \\in \\mathcal {Q}} \\mathcal {L} _ {C E} \\left(y ^ {q}, \\Theta_ {2} \\left(\\mathbf {m} ^ {q}\\right)\\right), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.473, + 0.903 + ], + "angle": 0, + "content": "where \\(\\mathcal{Q}\\) records the location of masked tokens, \\(|\\cdot|\\) denotes the length of a given set, \\(y^{q}\\) denotes the ground-truth token label, \\(\\Theta\\) is a Multi-Layer Perception (MLP), and \\(\\mathcal{L}_{CE}\\) is the regular Cross-Entropy cost function." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.296, + 0.893, + 0.343 + ], + "angle": 0, + "content": "VTM aims to promote the cross-modal alignment between videos and text, which is calculated in a parameter-free \\((\\mathcal{L}_3)\\) and parameter-employed \\((\\mathcal{L}_4)\\) way:" + }, + { + "type": "equation", + "bbox": [ + 0.503, + 0.354, + 0.892, + 0.411 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {3} = - \\sum_ {i = 1} ^ {| \\mathcal {B} |} \\log \\frac {\\sum_ {k = 1} ^ {N _ {v}} \\exp \\left\\langle \\mathbf {v} _ {i} ^ {k} , \\mathbf {t} _ {i} ^ {c l s} \\right\\rangle}{\\sum_ {k = 1} ^ {N _ {v}} \\left(\\exp \\left\\langle \\mathbf {v} _ {i} ^ {k} , \\mathbf {t} _ {i} ^ {c l s} \\right\\rangle + \\sum_ {j \\neq i} \\exp \\left\\langle \\mathbf {v} _ {j} ^ {k} , \\mathbf {t} _ {i} ^ {c l s} \\right\\rangle\\right)}, \\tag {5}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.527, + 0.435, + 0.892, + 0.479 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {4} = - \\sum_ {i = 1} ^ {| \\mathcal {B} |} \\log \\frac {\\exp^ {\\Theta_ {4} \\left(\\mathbf {f} _ {i , i}\\right)}}{\\exp^ {\\Theta_ {4} \\left(\\mathbf {f} _ {i , i}\\right)} + \\sum_ {j \\neq i} \\exp^ {\\Theta_ {4} \\left(\\mathbf {f} _ {j , i}\\right)}}, \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.49, + 0.892, + 0.536 + ], + "angle": 0, + "content": "where \\(|\\mathcal{B}|\\) is the length of a mini-batch, \\(\\langle \\cdot ,\\cdot \\rangle\\) denotes the matrix multiplication operation, and \\(\\mathbf{f}_{j,i}\\) is global-pooling cross-modal features of the video-text pair \\((V_{j},S_{i})\\)" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.536, + 0.892, + 0.567 + ], + "angle": 0, + "content": "For \\(3 - E - F\\) and \\(2 - E - F\\) models, the objective functions \\(\\mathcal{L}_{3EF} / \\mathcal{L}_{2EF}\\) are calculated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.568, + 0.579, + 0.892, + 0.596 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {3 E F} \\left(\\mathcal {L} _ {2 E F}\\right) = \\mathcal {L} _ {1} + \\mathcal {L} _ {2} + \\mathcal {L} _ {3} + \\mathcal {L} _ {4}. \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.608, + 0.892, + 0.655 + ], + "angle": 0, + "content": "While for \\(T\\)-\\(T-C\\) models, since they do not contain a cross-modal encoder to generate video representations, their objective function \\(\\mathcal{L}_{TTC}\\) is calculated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.631, + 0.668, + 0.892, + 0.684 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {T T C} = \\mathcal {L} _ {1} + \\mathcal {L} _ {3}. \\tag {8}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.696, + 0.628, + 0.712 + ], + "angle": 0, + "content": "4.4. Fine-tuning" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.892, + 0.854 + ], + "angle": 0, + "content": "In order to thoroughly evaluate the pre-training performance, we fine-tune our models on the Text-to-Video Retrieval (TVR) task. TVR is a widely-employed downstream video-text task, aiming to retrieve the most relevant videos according to text queries. Since there has only one Chinese TVR dataset (VATEX [38]), we translate two widely-adopted English datasets (MSRVTT [41] and DiDemo [3]) into Chinese to verify the superiority and robustness of the constructed dataset and proposed methods." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.856, + 0.893, + 0.902 + ], + "angle": 0, + "content": "Specifically, we fine-tune our pre-trained models by reusing Video-Text Matching objectives (\\(\\mathcal{L}_3\\) in Eq. 5 and \\(\\mathcal{L}_4\\) in Eq. 6, we only employ \\(\\mathcal{L}_3\\) for \\(T\\)-\\(T-C\\) models)." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "14819" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.09, + 0.478, + 0.21 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.222, + 0.47, + 0.265 + ], + "angle": 0, + "content": "Figure 4. The pipeline of Hard Sample Curriculum Learning (HSCL), which contains three components to promote the performance of video-text pre-training." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.275, + 0.401, + 0.293 + ], + "angle": 0, + "content": "5. Hard Sample Curriculum Learning" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.3, + 0.47, + 0.527 + ], + "angle": 0, + "content": "We propose Hard Sample Curriculum Learning (HSCL), a model-agnostic strategy to promote the video-text pretraining performance. HSCL is based on two motivations. 1) The negative samples involved in the calculation of parameter-employed VTM loss \\((\\mathcal{L}_4\\) in Eq. 6) are not representative. Suppose that the batch size is set to be \\(N_{k}\\), a model needs to process \\(N_{k} * N_{k}\\) video-text pairs for calculating \\(\\mathcal{L}_4\\), whose calculation cost is hardly affordable. Therefore, previous methods [9, 15, 22] tend to cut the number of negative samples to \\(N_{k}^{\\prime}\\) by random sampling. However, since \\(N_{k}^{\\prime}\\) is usually much smaller than \\(N_{k}\\), it may result in the sub-optimal performance. 2) If a model is far from convergence, directly introducing hard negative samples would amplify the burden of learning, which would decrease the pre-training performance." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.528, + 0.47, + 0.634 + ], + "angle": 0, + "content": "Therefore, inspired by hard sample mining and curriculum learning, we propose the HSCL strategy. By leveraging the similarity score calculated by the parameter-free VTM loss \\((\\mathcal{L}_3\\) in Eq. 5), HSCL could gradually and smoothly introduce hard samples for better cross-modal alignment. As illustrated in Figure 4, HSCL contains the following three key components:" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.634, + 0.469, + 0.724 + ], + "angle": 0, + "content": "Positive Column Weighting (PCW) aims to decrease the learning coefficient \\(\\phi\\) of well-learned positive samples, forcing the model to emphasize those hard positive ones. Specifically, we re-weigh half of samples within the batch according to their similarity scores \\(z\\) calculated by \\(\\mathcal{L}_3\\), which could be formulated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.18, + 0.733, + 0.469, + 0.775 + ], + "angle": 0, + "content": "\\[\nz _ {j, i} = \\frac {1}{N _ {v}} \\sum_ {k = 1} ^ {N _ {v}} \\exp \\left\\langle \\mathbf {v} _ {j} ^ {k}, \\mathbf {t} _ {i} ^ {c l s} \\right\\rangle , \\tag {9}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.11, + 0.786, + 0.469, + 0.821 + ], + "angle": 0, + "content": "\\[\n\\phi_ {i} = \\operatorname {M i n} \\left(\\frac {\\operatorname {M e d} \\left(Z _ {d i a g}\\right) - \\operatorname {M i n} \\left(Z _ {d i a g}\\right)}{z _ {i , i} - \\operatorname {M i n} \\left(Z _ {d i a g}\\right) + e _ {0}}, 1. 0\\right), \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.825, + 0.469, + 0.87 + ], + "angle": 0, + "content": "where \\( Z_{diag} = [z_{1,1}, z_{2,2}, \\dots, z_{i,i}, \\dots] \\), \\( \\operatorname{Med}(\\cdot) \\) and \\( \\operatorname{Min}(\\cdot) \\) represent the median and minimum result of the list, \\( e_0 \\) is a pre-defined coefficient to avoid a zero denominator." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.469, + 0.902 + ], + "angle": 0, + "content": "Negative Line Sampling (NLS) aims to sample \\( N_{k}^{\\prime} \\) hard negative examples for better contrastive learning. Specifi-" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.09, + 0.892, + 0.154 + ], + "angle": 0, + "content": "cally, for each text \\(T_{j}\\) , we choose the Nearliest \\(N_{k}^{\\prime}\\) aligned negative videos according to their similarity scores in the list \\(Z_{j} = [z_{j,1},z_{j,2},\\dots ,z_{j,i},\\dots ]\\) . I.e., Pick the minimum \\(N_{k}^{\\prime}\\) samples in the list \\(Z_{j}^{\\prime} = \\{|z_{j,i} - z_{j,j}|\\}_{j\\neq i}\\)" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.154, + 0.893, + 0.259 + ], + "angle": 0, + "content": "Curriculum Learning (CL) aims to mitigate the side effect of hard sample mining when the model is under convergence. We set a gradually increasing coefficient \\(\\beta \\in (0,1)\\). \\(\\beta\\) is initialized with zero at the beginning of pre-training. It will gradually increase by a fixed number \\(\\beta_0\\) for each \\(N_d\\) iteration until reaching 1.0. For PCW, the actual coefficient \\(\\phi^*\\) is calculated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.609, + 0.272, + 0.892, + 0.288 + ], + "angle": 0, + "content": "\\[\n\\phi_ {i} ^ {*} = (1. 0 - \\beta) + \\beta * \\phi_ {i}. \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.3, + 0.892, + 0.377 + ], + "angle": 0, + "content": "For NLS, we first generate a random number \\( rand \\in (0,1) \\), and only perform the NLS step if \\( rand < \\beta \\). In this way, the model would conduct conventional pre-training at the beginning, and smoothly change to learn hard negative samples to pursue better pre-training performance." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.389, + 0.779, + 0.406 + ], + "angle": 0, + "content": "6. Experiments and Benchmarks" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.414, + 0.702, + 0.43 + ], + "angle": 0, + "content": "6.1. Pre-training Datasets" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.437, + 0.892, + 0.512 + ], + "angle": 0, + "content": "In Section 3.3, the Chinese image-text model that helps to filter out the weakly-paired data is pre-trained on Wukong-100M [11] (100M image-text pairs). The Chinese video-text models are pre-trained on our CNVid-3.5M dataset, which contains 3.5M video-text pairs." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.513, + 0.892, + 0.604 + ], + "angle": 0, + "content": "To verify the generality of the proposed Hard Sample Curriculum Learning (HSCL) strategy, we also pre-trained our models on four large-scale English datasets, they are 1) COCO [20] (0.6M image-text pairs), 2) VG [13] (5.4M image-text pairs), 3) CC [31] (3.1M image-text pairs), and 4) WebVid-2.5M (2.5M video-text pairs)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.613, + 0.696, + 0.629 + ], + "angle": 0, + "content": "6.2. Fine-tuning Datasets" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.636, + 0.892, + 0.787 + ], + "angle": 0, + "content": "We fine-tuned our pre-trained models on the Text-to-Video Retrieval (TVR) task, which includes three widely-adopted datasets as follows: 1) VATEX [38] contains 28K video clips, and each video is associated with 10 Chinese and 10 English diverse captions. 2) MSRVTT [41] contains 10K video clips associated with 200K English captions. Following [15, 22, 43], we used 7K videos for training and randomly selected 1K videos from the remaining ones for testing (7K-1K split). 3) DiDemo [3] contains 10K Flickr videos associated with 40K English captions." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.788, + 0.892, + 0.847 + ], + "angle": 0, + "content": "Moreover, for a comprehensive performance evaluation of Chinese video-text pre-trained models, we employ Google \\(^{\\dagger}\\) to translate all captions in MSRVTT and DiDemo from English to Chinese." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.848, + 0.892, + 0.878 + ], + "angle": 0, + "content": "Metrics. We employed Recall@K (R@K, K=1/5/10) and Median Rank (MdR) to measure the TVR performance." + }, + { + "type": "page_footnote", + "bbox": [ + 0.516, + 0.887, + 0.771, + 0.901 + ], + "angle": 0, + "content": "†Official Website: https://translate.google.com/" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "14820" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.08, + 0.075, + 0.903, + 0.106 + ], + "angle": 0, + "content": "
No.Model TypeData VolumeMasking StrategyMatching StrategyVATEX (Chinese)MSRVTT (Translated)DiDemo (Translated)
R@1/5/10↑(MdR↓)R@1/5/10↑(MdR↓)R@1/5/10↑(MdR↓)
" + }, + { + "type": "table_caption", + "bbox": [ + 0.278, + 0.108, + 0.704, + 0.121 + ], + "angle": 0, + "content": "For \"Building\": Ablation Study of the constructed CNVid-3.5M dataset" + }, + { + "type": "table", + "bbox": [ + 0.081, + 0.123, + 0.903, + 0.155 + ], + "angle": 0, + "content": "
A12-E-FNo Pre-training36.4 / 75.4 / 85.2 (2)15.9 / 42.1 / 54.1 (8)8.0 / 24.8 / 39.2 (18)
A22-E-F3.5MMLMVTM39.9 / 77.2 / 87.0 (2)20.7 / 47.9 / 61.2 (6)13.3 / 34.6 / 45.9 (13)
" + }, + { + "type": "table_caption", + "bbox": [ + 0.299, + 0.157, + 0.684, + 0.17 + ], + "angle": 0, + "content": "For \"Filtering\": Ablation Study of Weakly-Paired Data Filtering" + }, + { + "type": "table", + "bbox": [ + 0.081, + 0.172, + 0.902, + 0.245 + ], + "angle": 0, + "content": "
B12-E-F4.5MMLMVTM38.8 / 76.4 / 86.4 (2)19.4 / 45.6 / 58.6 (6)10.6 / 32.7 / 45.0 (14)
B22-E-F4.0MMLMVTM39.5 / 77.0 / 86.5 (2)19.8 / 47.2 / 60.3 (6)12.4 / 32.8 / 45.6 (14)
B32-E-F3.5MMLMVTM39.9 / 77.2 / 87.0 (2)20.7 / 47.9 / 61.2 (6)13.3 / 34.6 / 45.9 (13)
B42-E-F3.0MMLMVTM40.5 / 76.8 / 86.5 (2)20.6 / 47.3 / 61.3 (6)12.2 / 34.3 / 45.6 (14)
B52-E-F2.5MMLMVTM39.5 / 76.6 / 86.9 (2)20.1 / 47.4 / 60.8 (6)12.1 / 33.6 / 46.0 (14)
" + }, + { + "type": "table_caption", + "bbox": [ + 0.271, + 0.248, + 0.71, + 0.262 + ], + "angle": 0, + "content": "For \"Pre-training\": Ablation Study of Hard Sample Curriculum Learning" + }, + { + "type": "table", + "bbox": [ + 0.081, + 0.264, + 0.902, + 0.309 + ], + "angle": 0, + "content": "
C12-E-F3.5MMLMVTM39.9 / 77.2 / 87.0 (2)20.7 / 47.9 / 61.2 (6)13.3 / 34.6 / 45.9 (13)
C22-E-F3.5MMLMVTM+HSM40.7 / 76.8 / 86.9 (2)21.1 / 47.3 / 60.1 (6)13.4 / 34.3 / 46.4 (12)
C32-E-F3.5MMLMVTM+HSCL41.5 / 78.2 / 87.2 (2)23.3 / 48.0 / 61.2 (6)13.6 / 34.4 / 47.3 (12)
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.32, + 0.895, + 0.362 + ], + "angle": 0, + "content": "Table 2. For Chinese video-text pre-training: Ablation study of Weakly-Paired Data Filtering and Hard Sample Curriculum Learning (HSCL) on the Text-to-Video Retrieval task of three datasets (VATEX-Chinese, MSRVTT-Translated, and DiDemo-Translated). \"HSM\" denotes the conventional Hard Sample Mining strategy without curriculum learning. Note that \\(A2\\), \\(B3\\), and \\(C1\\) are the same model." + }, + { + "type": "table", + "bbox": [ + 0.079, + 0.376, + 0.9, + 0.485 + ], + "angle": 0, + "content": "
No.Pre-training DatasetMasking StrategyMatching StrategyVATEX (English)MSRVTTDiDemo
R@1/5/10 ↑ (MdR ↓)R@1/5/10 ↑ (MdR ↓)R@1/5/10 ↑ (MdR ↓)
D0No Pre-training39.5 / 78.3 / 88.6 (2)19.2 / 46.2 / 59.5 (6)14.2 / 34.1 / 46.1 (12)
D1COCO+VG+CCMLMVTM45.5 / 81.5 / 91.0 (2)28.6 / 55.5 / 66.1 (4)25.3 / 50.3 / 62.4 (5)
D2COCO+VG+CCMLMVTM+HSCL48.7 / 83.1 / 91.1 (2)31.6 / 56.5 / 66.7 (4)27.1 / 53.8 / 63.8 (4)
D3CC+WebVidMLMVTM50.5 / 85.6 / 92.5 (1)29.4 / 56.9 / 68.3 (4)29.0 / 57.3 / 67.8 (4)
D4CC+WebVidMLMVTM+HSCL53.4 / 86.2 / 92.9 (1)32.6 / 58.8 / 69.7 (3)29.8 / 58.0 / 68.9 (3)
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.495, + 0.893, + 0.525 + ], + "angle": 0, + "content": "Table 3. For English video-text pre-training: Ablation study of Hard Sample Curriculum Learning (HSCL) on the Text-to-Video Retrieval task of three datasets (VATEX-English, MSRVTT, and DiDemo). All models are based on the two-encoders-fusion \\((2 - E - F)\\) architecture." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.536, + 0.285, + 0.552 + ], + "angle": 0, + "content": "6.3. Experimental Settings" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.559, + 0.47, + 0.739 + ], + "angle": 0, + "content": "Settings for Weakly-paired Data Filtering. We pretrained the image-text model on the Wukong-100M [11] dataset, whose pre-training details are presented in the supplementary material. We then employed the pre-trained model to obtain the parameter-free video-text consistency score \\((z_{i,i}\\) in Eq. 9) of 4.5M video-text pairs. To improve the reliability, we randomly sampled four frames to calculate the consistency score, and repeated this step for three times. The final score of each video is the average of three results calculated under different random seeds. Ultimately, we sorted the consistency score in descending order, and filtered out the last 1M videos to refine the dataset." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.741, + 0.47, + 0.877 + ], + "angle": 0, + "content": "Settings for Chinese Video-Text Pre-training. For model architectures, we employed the Video Swin Transformer [21] to serve as the visual encoder, which is initialized with parameters pre-trained on ImageNet [6]. The text, cross-modal, and shared encoders belong to the BERT-Base [7] model, which are initialized with parameters privdied by Hugging Face \\(\\ddagger\\). We sparsely sampled \\(4(N_v)\\) frames from raw videos during pre-training. For the MLM task, the masking rate is \\(15\\%\\). For the proposed Hard Sample" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.537, + 0.891, + 0.566 + ], + "angle": 0, + "content": "Curriculum Learning, the coefficient \\(\\beta\\) would increase by \\(0.15(\\beta_0)\\) for each \\(5000(N_d)\\) iterations until reaching 1.0." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.567, + 0.892, + 0.686 + ], + "angle": 0, + "content": "For hyper-parameters, we set the length of text tokens \\( N_{t} = 30 \\), and the dimension of the hidden state \\( d = 768 \\). All models are pre-trained by the Adam optimizer with a momentum of 0.9. The total pre-training stage lasts for 10 epochs with a batch size of 128. The initial learning rate is 5e-5 and is decayed by the factor of 10 after 5 epochs. The whole pre-training on CNVid-3.5M takes about 4 days to complete on 8 NVIDIA V100 GPUs." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.687, + 0.892, + 0.763 + ], + "angle": 0, + "content": "Settings for Fine-tuning on Downstream Chinese Videotext Tasks. The optimizer and hyper-parameters in fine-tuning remain the same as the pre-training configuration. The total fine-tuning stage lasts for 25,000 steps. The batch size is set to 128. The initial learning rate is set to 1e-5." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.772, + 0.655, + 0.789 + ], + "angle": 0, + "content": "6.4. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.796, + 0.892, + 0.901 + ], + "angle": 0, + "content": "As aforementioned, we summarized our work with three verbs, i.e., \"Build\", \"Filter\", and \"Pre-train\". To evaluate the effectiveness of these three contributions, we took the 2-\\(E-F\\) architecture as an example model in the ablation study. Ablation Study of \"Building\". We compared the performance obtained by models with or without pre-training. As illustrated in Table 2 (A1 vs. A2), models pre-trained on" + }, + { + "type": "page_footnote", + "bbox": [ + 0.095, + 0.887, + 0.411, + 0.901 + ], + "angle": 0, + "content": "\\(^\\ddagger\\) Official Website: https://huggingface.co/bert-base-chinese" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "14821" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.082, + 0.089, + 0.887, + 0.164 + ], + "angle": 0, + "content": "
No.Model TypeDatasetPre-training StrategyVATEX (Chinese)MSRVTT (Translated)DiDemo (Translated)
R@1/5/10 ↑ (MdR ↓)R@1/5/10 ↑ (MdR ↓)R@1/5/10 ↑ (MdR ↓)
E13-E-FCNVid-3.5MMLM+VTM39.9 / 78.2 / 87.6 (2)20.6 / 47.1 / 60.0 (6)13.8 / 34.5 / 46.6 (13)
E22-E-FCNVid-3.5MMLM+VTM39.9 / 77.2 / 87.0 (2)20.7 / 47.9 / 61.2 (6)13.3 / 34.6 / 45.9 (13)
E3T-T-CCNVid-3.5MMLM+VTM40.5 / 77.3 / 87.0 (2)21.1 / 46.5 / 58.9 (6)13.0 / 32.6 / 44.6 (15)
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.174, + 0.893, + 0.203 + ], + "angle": 0, + "content": "Table 4. We benchmark CNVid-3.5M with three mainstream pixel-level pre-training architectures, which is the first pixel-level benchmarks for Chinese video-text pre-training." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.214, + 0.468, + 0.274 + ], + "angle": 0, + "content": "our CNVid-3.5M dataset (A2) would achieve a remarkable performance gain compared with those without pre-training (A1), which demonstrates the effectiveness of building a large-scale Chinese video-text dataset." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.276, + 0.468, + 0.395 + ], + "angle": 0, + "content": "Ablation Study of \"Filtering\". We compared the performance obtained by models pre-trained under different numbers of video-text pairs. For example, \"4.5M\", \"4.0M\", and \"3.5M\" denotes that we employ the 4.5M full set, the Top-4.0M split set, and the Top-3.5M split set (CNVid-3.5M) according to the sorted video-text consistency scores calculated by weakly-paired data filtering. As illustrated in Table 2 (B1-B5), we have two conclusions as follows:" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.398, + 0.468, + 0.502 + ], + "angle": 0, + "content": "1) The proposed weakly-paired data filtering strategy could effectively improve the quality of datasets. The fine-tuning performance on three downstream datasets first increases rapidly when filtering out videos with low video-text consistency scores (\\(B1 \\to B2 \\to B3\\)). It would remain stable (\\(B3 \\to B4\\)) or drop slightly (\\(B4 \\to B5\\)) when continuously removing videos after the Top-3.5M split." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.505, + 0.468, + 0.594 + ], + "angle": 0, + "content": "2) In general, employing the Top-3.5M split (B3) achieves the best performance. The results among all ablation models (B1-B5) may prove that, for the video-text pre-training, it is not \"more data is better\", but \"more good data is better\". Therefore, we filtered out the last 1M videos, resulting in the CNVid-3.5M dataset." + }, + { + "type": "list", + "bbox": [ + 0.076, + 0.398, + 0.468, + 0.594 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.596, + 0.468, + 0.671 + ], + "angle": 0, + "content": "Ablation Study of \"Pre-training\" (HSCL). To prove the effectiveness of Hard Sample Curriculum Learning (HSCL), we pre-trained several ablation models on Chinese video-text datasets (\\(C1-C3\\) in Table 2) and English ones (\\(D1-D4\\) in Table 3). We have two conclusions as follows:" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.673, + 0.468, + 0.792 + ], + "angle": 0, + "content": "1) HSCL is a model-agnostic strategy, which could evidently improve the pre-training performance. In Table 2, compared with the baseline (C1), HSCL (C3) achieves an improvement of \\(1.6\\%\\), \\(2.6\\%\\), and \\(0.3\\%\\) at R@1 of VATEX, MSRVTT, and DiDemo dataset, respectively. While in Table 3, HSCL outperforms the baseline by \\(3.2 / 3.0 / 1.8\\%\\) (\\(D1\\) vs. \\(D2\\) on \\(\\mathrm{COCO + VG + CC}\\)) and \\(2.9 / 3.2 / 0.8\\%\\) (\\(D3\\) vs. \\(D4\\) on \\(\\mathrm{CC + WebVid}\\)) at R@1 of three TVR datasets." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.794, + 0.468, + 0.869 + ], + "angle": 0, + "content": "2) The proposed Curriculum Learning (CL) strategy in HSCL could effectively mitigate the side effect of conventional Hard Sample Mining (HSM) when the model is far from convergence. As illustrated in Table 2 (C1 vs. C2), models equipped with HSCL perform better than HSM." + }, + { + "type": "list", + "bbox": [ + 0.076, + 0.673, + 0.468, + 0.869 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.871, + 0.468, + 0.901 + ], + "angle": 0, + "content": "Moreover, we present the detailed quantitative analyses of PCW and NLS in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.214, + 0.88, + 0.23 + ], + "angle": 0, + "content": "6.5. Chinese Video-text Pre-training Benchmarks" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.237, + 0.892, + 0.357 + ], + "angle": 0, + "content": "Based on the constructed CNVid-3.5M dataset, We provided the first pixel-level benchmarks for Chinese videotext pre-training. Following existing video-text pre-training methods, we adopted three mainstream pixel-level architectures, i.e., three-encoders-fusion (3-E-F) [15, 17, 30], two-encoders-fusion (2-E-F) [9], and twin-towers-crossed (T-T-C) [8, 23, 36]. Detailed performance comparisons of three pixel-level benchmarks are presented in Table 4." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.372, + 0.618, + 0.387 + ], + "angle": 0, + "content": "7. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.397, + 0.892, + 0.548 + ], + "angle": 0, + "content": "In this work, we release the largest public Chinese video-text dataset, i.e., CNVid-3.5M. Besides, we provide the first pixel-level benchmarks for Chinese video-text pretraining. Moreover, we propose a novel weakly-paired data filtering method to improve the quality of datasets. We also design a novel Hard Sample Curriculum Learning strategy to promote the pre-training performance. In conclusion, we \"build\", \"filter\", and \"pre-train\" the large-scale Chinese video-text dataset. We believe this work would pave the way for future research on Chinese video-text pre-training." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.561, + 0.771, + 0.577 + ], + "angle": 0, + "content": "8. Limitations and Future Work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.587, + 0.892, + 0.844 + ], + "angle": 0, + "content": "Address the Difference between English and Chinese. One potential limitation of our work is that we adopt pixel-level architectures proved effective in English video-text pre-training, ignoring the linguistic gap between Chinese and English. E.g., it may be better to employ the Whole Word Masking (WWM) [5, 46] strategy rather than the conventional MLM for Chinese video-text pre-training. We plan to explore more Chinese-specific methods in the future. Need More Downstream Chinese Video-Text Datasets. Though we release a large-scale Chinese video-text dataset, there still lack some downstream datasets (e.g., MSRVTT [41], MSVD-QA [39] in English) to comprehensively evaluate the performance of Chinese video-text pre-trained models. Currently, there only exists VATEX [38] for Chinese cross-modal retrieval and some captioning datasets [26, 45]. Therefore, building more downstream Chinese video datasets is a promising future direction." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.853, + 0.892, + 0.899 + ], + "angle": 0, + "content": "Acknowledgments. This work is supported by the National Natural Science Foundation of China, No.: 62176137; Ant Group." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "14822" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.079, + 0.09, + 0.176, + 0.107 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.17 + ], + "angle": 0, + "content": "[1] Nayyer Aafaq, Ajmal Mian, Wei Liu, Syed Zulqarnain Gilani, and Mubarak Shah. Video description: A survey of methods, datasets, and evaluation metrics. ACM Computing Surveys, 52(6):1-37, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.171, + 0.472, + 0.24 + ], + "angle": 0, + "content": "[2] Hassan Akbari, Liangzhe Yuan, Rui Qian, Wei-Hong Chuang, Shih-Fu Chang, Yin Cui, and Boqing Gong. Vatt: Transformers for multimodal self-supervised learning from raw video, audio and text. Advances in Neural Information Processing Systems, 34, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.242, + 0.473, + 0.31 + ], + "angle": 0, + "content": "[3] Lisa Anne Hendricks, Oliver Wang, Eli Shechtman, Josef Sivic, Trevor Darrell, and Bryan Russell. Localizing moments in video with natural language. In Proceedings of the IEEE international conference on computer vision, pages 5803-5812, 2017. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.311, + 0.472, + 0.38 + ], + "angle": 0, + "content": "[4] Max Bain, Arsha Nagrani, Gül Varol, and Andrew Zisserman. Frozen in time: A joint video and image encoder for end-to-end retrieval. In Proceedings of the IEEE International Conference on Computer Vision, pages 1728-1738, 2021. 1, 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.381, + 0.472, + 0.436 + ], + "angle": 0, + "content": "[5] Yiming Cui, Wanxiang Che, Ting Liu, Bing Qin, and Ziqing Yang. Pre-training with whole word masking for chinese bert. IEEE Transactions on Audio, Speech, and Language Processing, 29:3504-3514, 2021. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.438, + 0.472, + 0.493 + ], + "angle": 0, + "content": "[6] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In IEEE Conference on Computer Vision and Pattern Recognition, pages 248-255, 2009. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.494, + 0.472, + 0.549 + ], + "angle": 0, + "content": "[7] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.55, + 0.472, + 0.592 + ], + "angle": 0, + "content": "[8] Han Fang, Pengfei Xiong, Luhui Xu, and Yu Chen. Clip2video: Mastering video-text retrieval via image clip. arXiv preprint arXiv:2106.11097, 2021. 2, 4, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.593, + 0.472, + 0.648 + ], + "angle": 0, + "content": "[9] Tsu-Jui Fu, Linjie Li, Zhe Gan, Kevin Lin, William Yang Wang, Lijuan Wang, and Zicheng Liu. Violet: End-to-end video-language transformers with masked visual-token modeling. arXiv preprint arXiv:2111.12681, 2021. 2, 4, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.649, + 0.472, + 0.718 + ], + "angle": 0, + "content": "[10] Yuying Ge, Yixiao Ge, Xihui Liu, Dian Li, Ying Shan, Xiaohu Qie, and Ping Luo. Bridging video-text retrieval with multiple choice questions. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 16167-16176, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.719, + 0.472, + 0.802 + ], + "angle": 0, + "content": "[11] Jiaxi Gu, Xiaojun Meng, Guansong Lu, Lu Hou, Minzhe Niu, Hang Xu, Xiaodan Liang, Wei Zhang, Xin Jiang, and Chunjing Xu. Wukong: 100 million large-scale chinese cross-modal pre-training dataset and a foundation framework. arXiv preprint arXiv:2202.06767, 2022. 2, 4, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.803, + 0.472, + 0.858 + ], + "angle": 0, + "content": "[12] Yannis Kalantidis, Mert Bulent Sariyildiz, Noe Pion, Philippe Weinzaepfel, and Diane Larlus. Hard negative mixing for contrastive learning. Advances in Neural Information Processing Systems, 33:21798-21809, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.859, + 0.472, + 0.901 + ], + "angle": 0, + "content": "[13] Ranjay Krishna, Yuke Zhu, Oliver Groth, Justin Johnson, Kenji Hata, Joshua Kravitz, Stephanie Chen, Yannis Kalantidis, Li-Jia Li, David A Shamma, et al. Visual genome:" + }, + { + "type": "list", + "bbox": [ + 0.079, + 0.115, + 0.473, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.894, + 0.134 + ], + "angle": 0, + "content": "Connecting language and vision using crowdsourced dense image annotations. International Journal of Computer Vision, 123(1):32-73, 2017. 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.135, + 0.894, + 0.217 + ], + "angle": 0, + "content": "[14] Chenyi Lei, Shixian Luo, Yong Liu, Wanggui He, Jiamang Wang, Guoxin Wang, Haihong Tang, Chunyan Miao, and Houqiang Li. Understanding chinese video and language via contrastive multimodal pre-training. In Proceedings of the ACM International Conference on Multimedia, pages 2567-2576, 2021. 1, 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.218, + 0.894, + 0.3 + ], + "angle": 0, + "content": "[15] Jie Lei, Linjie Li, Luowei Zhou, Zhe Gan, Tamara L Berg, Mohit Bansal, and Jingjing Liu. Less is more: Clipbert for video-and-language learning via sparse sampling. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 7331-7341, 2021. 1, 2, 4, 5, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.301, + 0.894, + 0.357 + ], + "angle": 0, + "content": "[16] Chunyuan Li, Xiujun Li, Lei Zhang, Baolin Peng, Mingyuan Zhou, and Jianfeng Gao. Self-supervised pre-training with hard examples improves visual representations. arXiv preprint arXiv:2012.13493, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.358, + 0.894, + 0.426 + ], + "angle": 0, + "content": "[17] Dongxu Li, Junnan Li, Hongdong Li, Juan Carlos Niebles, and Steven CH Hoi. Align and prompt: Video-and-language pre-training with entity prompts. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 4953-4963, 2022. 1, 2, 4, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.427, + 0.894, + 0.496 + ], + "angle": 0, + "content": "[18] Junnan Li, Ramprasaath Selvaraju, Akhilesh Gotmare, Shafiq Joty, Caiming Xiong, and Steven Chu Hong Hoi. Align before fuse: Vision and language representation learning with momentum distillation. Advances in neural information processing systems, 34:9694-9705, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.497, + 0.894, + 0.552 + ], + "angle": 0, + "content": "[19] Linjie Li, Yen-Chun Chen, Yu Cheng, Zhe Gan, Licheng Yu, and Jingjing Liu. Hero: Hierarchical encoder for video+language omni-representation pre-training. arXiv preprint arXiv:2005.00200, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.553, + 0.894, + 0.621 + ], + "angle": 0, + "content": "[20] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Proceedings of the European Conference on Computer Vision, pages 740-755. Springer, 2014. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.622, + 0.894, + 0.678 + ], + "angle": 0, + "content": "[21] Ze Liu, Jia Ning, Yue Cao, Yixuan Wei, Zheng Zhang, Stephen Lin, and Han Hu. Video swim transformer. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 3202-3211, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.679, + 0.894, + 0.747 + ], + "angle": 0, + "content": "[22] Huaishao Luo, Lei Ji, Botian Shi, Haoyang Huang, Nan Duan, Tianrui Li, Jason Li, Taroon Bharti, and Ming Zhou. Univl: A unified video and language pre-training model for multimodal understanding and generation. arXiv preprint arXiv:2002.06353, 2020. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.748, + 0.894, + 0.802 + ], + "angle": 0, + "content": "[23] Huaishao Luo, Lei Ji, Ming Zhong, Yang Chen, Wen Lei, Nan Duan, and Tianrui Li. Clip4clip: An empirical study of clip for end to end video clip retrieval. arXiv preprint arXiv:2104.08860, 2021. 2, 4, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.803, + 0.894, + 0.873 + ], + "angle": 0, + "content": "[24] Jianjie Luo, Yehao Li, Yingwei Pan, Ting Yao, Hongyang Chao, and Tao Mei. Coco-bert: Improving video-language pre-training with contrastive cross-modal matching and denoising. In Proceedings of the ACM International Conference on Multimedia, pages 5600-5608, 2021. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.873, + 0.894, + 0.901 + ], + "angle": 0, + "content": "[25] Antoine Miech, Dimitri Zhukov, Jean-Baptiste Alayrac, Makarand Tapaswi, Ivan Laptev, and Josef Sivic." + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.894, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "14823" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.092, + 0.47, + 0.147 + ], + "angle": 0, + "content": "Howto100m: Learning a text-video embedding by watching hundred million narrated video clips. In Proceedings of the IEEE International Conference on Computer Vision, pages 2630-2640, 2019. 1, 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.151, + 0.47, + 0.22 + ], + "angle": 0, + "content": "[26] Liqiang Nie, Leigang Qu, Dai Meng, Min Zhang, Qi Tian, and Alberto Del Bimbo. Search-oriented microvideo captioning. In Proceedings of the ACM International Conference on Multimedia, pages 3234-3243, 2022. 1, 2, 3, 4, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.224, + 0.47, + 0.279 + ], + "angle": 0, + "content": "[27] Devshree Patel, Ratnam Parikh, and Yesha Shastri. Recent advances in video question answering: A review of datasets and methods. In International Conference on Pattern Recognition, pages 339-356. Springer, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.282, + 0.469, + 0.365 + ], + "angle": 0, + "content": "[28] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.368, + 0.469, + 0.422 + ], + "angle": 0, + "content": "[29] Joshua David Robinson, Ching-Yao Chuang, Suvrit Sra, and Stefanie Jegelka. Contrastive learning with hard negative samples. In International Conference on Learning Representations, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.426, + 0.469, + 0.495 + ], + "angle": 0, + "content": "[30] Paul Hongsuck Seo, Arsha Nagrani, Anurag Arnab, and Cordelia Schmid. End-to-end generative pretraining for multimodal video captioning. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 17959-17968, 2022. 1, 2, 4, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.498, + 0.469, + 0.568 + ], + "angle": 0, + "content": "[31] Piyush Sharma, Nan Ding, Sebastian Goodman, and Radu Soricut. Conceptual captions: A cleaned, hypernymed, image alt-text dataset for automatic image captioning. In Proceedings of the Annual Meeting of the Association for Computational Linguistics, pages 2556–2565, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.571, + 0.469, + 0.613 + ], + "angle": 0, + "content": "[32] Petru Soviany, Radu Tudor Ionescu, Paolo Rota, and Nicu Sebe. Curriculum learning: A survey. International Journal of Computer Vision, pages 1-40, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.616, + 0.469, + 0.683 + ], + "angle": 0, + "content": "[33] Chen Sun, Austin Myers, Carl Vondrick, Kevin Murphy, and Cordelia Schmid. Videobert: A joint model for video and language representation learning. In Proceedings of the IEEE International Conference on Computer Vision, pages 7464-7473, 2019. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.687, + 0.469, + 0.742 + ], + "angle": 0, + "content": "[34] Guanglu Sun, Lili Liang, Tianlin Li, Bo Yu, Meng Wu, and Bolun Zhang. Video question answering: a survey of models and datasets. Mobile Networks and Applications, 26(5):1904-1937, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.746, + 0.469, + 0.815 + ], + "angle": 0, + "content": "[35] Alex Jinpeng Wang, Yixiao Ge, Rui Yan, Yuying Ge, Xudong Lin, Guanyu Cai, Jianping Wu, Ying Shan, Xiaohu Qie, and Mike Zheng Shou. All in one: Exploring unified video-language pre-training. arXiv preprint arXiv:2203.07303, 2022. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.818, + 0.469, + 0.899 + ], + "angle": 0, + "content": "[36] Jinpeng Wang, Yixiao Ge, Guanyu Cai, Rui Yan, Xudong Lin, Ying Shan, Xiaohu Qie, and Mike Zheng Shou. Object-aware video-language pre-training for retrieval. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 3313-3322, 2022. 1, 2, 4, 8" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.134 + ], + "angle": 0, + "content": "[37] Xin Wang, Yudong Chen, and Wenwu Zhu. A survey on curriculum learning. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.136, + 0.892, + 0.205 + ], + "angle": 0, + "content": "[38] Xin Wang, Jiawei Wu, Junkun Chen, Lei Li, Yuan-Fang Wang, and William Yang Wang. Vatex: A large-scale, high-quality multilingual dataset for video-and-language research. In Proceedings of the IEEE International Conference on Computer Vision, pages 4581-4591, 2019. 3, 5, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.207, + 0.892, + 0.275 + ], + "angle": 0, + "content": "[39] Dejing Xu, Zhou Zhao, Jun Xiao, Fei Wu, Hanwang Zhang, Xiangnan He, and Yueting Zhuang. Video question answering via gradually refined attention over appearance and motion. In Proceedings of the ACM International Conference on Multimedia, pages 1645-1653, 2017. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.277, + 0.892, + 0.345 + ], + "angle": 0, + "content": "[40] Hu Xu, Gargi Ghosh, Po-Yao Huang, Prahal Arora, Masoumeh Aminzadeh, Christoph Feichtenhofer, Florian Metze, and Luke Zettlemoyer. Vlm: Task-agnostic video-language model pre-training for video understanding. arXiv preprint arXiv:2105.09996, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.347, + 0.892, + 0.403 + ], + "angle": 0, + "content": "[41] Jun Xu, Tao Mei, Ting Yao, and Yong Rui. Msr-vtt: A large video description dataset for bridging video and language. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 5288-5296, 2016. 5, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.404, + 0.892, + 0.471 + ], + "angle": 0, + "content": "[42] Antoine Yang, Antoine Miech, Josef Sivic, Ivan Laptev, and Cordelia Schmid. Just ask: Learning to answer questions from millions of narrated videos. In Proceedings of the IEEE International Conference on Computer Vision, pages 1686-1697, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.474, + 0.892, + 0.543 + ], + "angle": 0, + "content": "[43] Jianwei Yang, Yonatan Bisk, and Jianfeng Gao. Taco: Token-aware cascade contrastive learning for video-text alignment. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11562-11572, 2021. 2, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.545, + 0.892, + 0.613 + ], + "angle": 0, + "content": "[44] Shengyu Zhang, Ziqi Tan, Jin Yu, Zhou Zhao, Kun Kuang, Jie Liu, Jingren Zhou, Hongxia Yang, and Fei Wu. Poet: Product-oriented video captioner for e-commerce. In Proceedings of the ACM International Conference on Multimedia, pages 1292-1301, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.615, + 0.892, + 0.683 + ], + "angle": 0, + "content": "[45] Ziqi Zhang, Yuxin Chen, Zongyang Ma, Zhongang Qi, Chunfeng Yuan, Bing Li, Ying Shan, and Weiming Hu. Create: A benchmark for chinese short video retrieval and title generation. arXiv preprint arXiv:2203.16763, 2022. 1, 2, 3, 4, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.686, + 0.892, + 0.767 + ], + "angle": 0, + "content": "[46] Wei Zhu. Mvp-bert: Multi-vocab pre-training for chinese bert. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing: Student Research Workshop, pages 260-269, 2021. 8" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.767 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "14824" + } + ] +] \ No newline at end of file diff --git a/2023/CNVid-3.5M_ Build, Filter, and Pre-Train the Large-Scale Public Chinese Video-Text Dataset/3819ec34-d037-4584-a0be-1f25bc0cb37d_origin.pdf b/2023/CNVid-3.5M_ Build, Filter, and Pre-Train the Large-Scale Public Chinese Video-Text Dataset/3819ec34-d037-4584-a0be-1f25bc0cb37d_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..d26c4f42a3611fa7e7ccee63e1a7b7ce57e95aab --- /dev/null +++ b/2023/CNVid-3.5M_ Build, Filter, and Pre-Train the Large-Scale Public Chinese Video-Text Dataset/3819ec34-d037-4584-a0be-1f25bc0cb37d_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e7f758339a0ee91feb7f9ca98a770abe650e20d3c9caf497ba483004037cdd21 +size 1821782 diff --git a/2023/CNVid-3.5M_ Build, Filter, and Pre-Train the Large-Scale Public Chinese Video-Text Dataset/full.md b/2023/CNVid-3.5M_ Build, Filter, and Pre-Train the Large-Scale Public Chinese Video-Text Dataset/full.md new file mode 100644 index 0000000000000000000000000000000000000000..e57dec22be0c2e8965b24d3d54fee5d6dba7b94c --- /dev/null +++ b/2023/CNVid-3.5M_ Build, Filter, and Pre-Train the Large-Scale Public Chinese Video-Text Dataset/full.md @@ -0,0 +1,386 @@ +# CNVid-3.5M: Build, Filter, and Pre-train the Large-scale Public Chinese Video-text Dataset + +Tian Gan $^{1*}$ , Qing Wang $^{2*}$ , Xingning Dong $^{1}$ , Xiangyuan Ren $^{2}$ , Liqiang Nie $^{3}$ , Qingpei Guo $^{2\dagger}$ + +1Shandong University, + +$^{2}$ Ant Group, + +$^{3}$ Harbin Institute of Technology (Shenzhen) + +gantian@sdu.edu.cn, + +wq176625@antgroup.com, + +dongxingning1998@gmail.com + +xiangyuan.rxy@antgroup.com, nieliqiang@gmail.com, qingpei.gqp@antgroup.com + +# Abstract + +Owing to well-designed large-scale video-text datasets, recent years have witnessed tremendous progress in videotext pre-training. However, existing large-scale video-text datasets are mostly English-only. Though there are certain methods studying the Chinese video-text pre-training, they pre-train their models on private datasets whose videos and text are unavailable. This lack of large-scale public datasets and benchmarks in Chinese hampers the research and downstream applications of Chinese video-text pretraining. Towards this end, we release and benchmark CNVid-3.5M, a large-scale public cross-modal dataset containing over 3.5M Chinese video-text pairs. We summarize our contributions by three verbs, i.e., "Build", "Filter", and "Pre-train": 1) To build a public Chinese video-text dataset, we collect over 4.5M videos from the Chinese websites. 2) To improve the data quality, we propose a novel method to filter out 1M weakly-paired videos, resulting in the CNVid-3.5M dataset. And 3) we benchmark CNVid-3.5M with three mainstream pixel-level pre-training architectures. At last, we propose the Hard Sample Curriculum Learning strategy to promote the pre-training performance. To the best of our knowledge, CNVid-3.5M is the largest public video-text dataset in Chinese, and we provide the first pixel-level benchmarks for Chinese video-text pretraining. The dataset, codebase, and pre-trained models are available at https://github.com/CNVid/CNVid-3.5M. + +# 1. Introduction + +Owing to well-designed large-scale datasets, video-text pre-training [15, 17, 19] has achieved superior performance in various downstream tasks, such as video-text retrieval [4, 10, 36], video question answering [27, 34, 42], and video captioning [1, 22, 30]. However, recent large-scale video- + +![](images/9b7a8bdb48e124ef90a9d5b532991186db25efb56d5530433476257688e18c7f.jpg) +Figure 1. Here presents the motivations of this paper, based on which we highly summarize our contributions with three verbs: "Build", "Filter", and "Pre-train". + +![](images/a4d4cffe965fc815b477273e86bb78783bbe166255e5abaa4f22807664ae1cd0.jpg) + +![](images/b46894051c7f630db7c77289f2833c7996d7b3725c9d2e194d6e2cb1d0955b60.jpg) + +text datasets are mostly English-only (e.g., Howto100M [25] and WebVid-2.5M [4]). Though some methods [14,26, 45] turn to study the Chinese video-text pre-training, they pre-train their models on private datasets whose videos and text are unavailable. Therefore, the research towards Chinese video-text pre-training is still in its infancy due to the lack of large-scale public datasets. + +Towards this problem, directly translating English text into Chinese is a simple solution. However, it may result in unacceptable performance degradation for two reasons: 1) Translation errors are inevitable. Moreover, since most of the large-scale video-text datasets employ the Automatic Speech Recognition (ASR) system to generate text, the language translator would amplify the error from the incomplete and noisy ASR text. And 2) there remains an intrinsic linguistic gap between English and Chinese. Many widely-used English idioms and slang can hardly find their Chinese counterparts, leading some translated text incomprehensible and even contrary to the original meaning. + +In this paper, we aim to release and benchmark a large-scale public Chinese video-text dataset to facilitate future researchers and the community. As illustrated in Figure 1, three verbs could highly summarize our contributions, i.e., "Build", "Filter", and "Pre-train". + +To build a large-scale Chinese video-text dataset, we collect over 4.5M videos from Chinese websites. All videos are associated with user-uploaded titles and ASR text. + +We filter out the weakly-paired data by a novel method to improve the data quality. As some work [25,26] pointed out, the pre-training performance would suffer from the noisy ASR text that fails to accurately describe the video content. Unfortunately, the problem is raised with few practical solutions. Therefore, we employ a well-trained image-text model to evaluate the video-text consistency for three reasons: 1) The text information in existing image-text datasets [11] are usually manually-written titles or captions, whose consistency is guaranteed. 2) Some video-text pre-training architectures [23,35] are based upon image-text ones. And 3) it is cheap and efficient to "hire" a well-trained model to check millions of videos. In this way, we filter out about 1M weakly-paired videos based on the balance between the pre-training performance and efficiency, deriving the proposed CNVid-3.5M dataset. + +We pre-train various models to benchmark our CNVid-3.5M dataset. Current video-text pre-training methods could be roughly divided into two categories: 1) feature-level pre-training methods [24, 33, 40] that employ offline video and textual feature extractors, and 2) pixel-level ones [4, 15, 36] that learn cross-modal representations end-to-end from raw videos and text. Since there remain domain gaps between pre-training datasets and frozen feature extractors, pixel-level pre-training methods usually achieve better performance and have been widely employed in recent years. However, existing Chinese video-text pre-training methods [14, 26, 45] are all feature-level ones pre-trained on private datasets, limiting their contributions on the development of Chinese video-text pre-training techniques. Hence, we adopt three mainstream pixel-level pre-training frameworks, which are the first pixel-level benchmarks for Chinese video-text pre-training. + +Moreover, we propose the novel Hard Sample Curriculum Learning strategy to promote the pre-training performance. Since contrastive learning is a significant component in video-text pre-training, some methods [16, 18, 43] employ the hard sample mining [12, 29] strategy to promote the cross-modal alignment. However, hard sample mining would bring side effects to pre-training when the model is far from convergence. Suppose that a model is incapable of discriminating the ground-truth video-text pairs, recklessly introducing hard negatives would lead to the sub-optimal performance. Inspired by the curriculum learning [32, 37] strategy that "starts small" and gradually "learns hard", we combine these two strategies and propose the novel Hard Sample Curriculum Learning (HSCL). By gradually and smoothly emphasizing those hard samples, HSCL could effectively improve the pre-training performance. + +Our contributions are summarized in four folds: + +- To fill in the blank of large-scale public Chinese video-text datasets, we collect over 4.5M videos associated with titles and ASR text from the websites. + +- To improve the data quality, we propose a novel method to filter out 1M weakly-paired videos, resulting in the CNVid-3.5M dataset. +- To promote the pre-training performance, we propose the novel Hard Sample Curriculum Learning strategy for better cross-modal contrastive learning. +- To the best of our knowledge, the constructed CNVid-3.5M is the largest public Chinese video-text dataset. Moreover, we provide the first Chinese pixel-level benchmarks based on CNVid-3.5M. The dataset, codebase, and benchmarks are available at https://github.com/CNVid/CNVid-3.5M. + +# 2. Related Work + +Video-Text Pre-training and Fine-Tuning. Video-text pre-training has attracted increasing interest in recent years, which could be roughly divided into two categories, feature-level pre-training methods [24, 33, 40] and pixel-level ones [15, 17, 35]. The former approaches employ offline visual and textual features extracted from frozen models, while the latter ones learn cross-modal representations from raw videos and text in an end-to-end manner. Since there remain domain gaps between pre-training datasets and frozen feature extractors, pixel-level pre-training methods usually achieve better performance than the feature-level ones. + +There are three mainstream architectures in pixel-level video-text pre-training methods, i.e., three-encoders-fusion (3-E-F) [15, 17, 30], two-encoders-fusion (2-E-F) [9], and twin-towers-crossed (T-T-C) [8, 23, 36]. 3-E-F methods contain three separate encoders to model visual, textual, and cross-modal features; 2-E-F methods simplify the 3-E-F architecture by employing a shared encoder to embed textual and cross-modal features; while T-T-C methods are mainly based on the widely-adopted CLIP [28] architecture, which includes a couple of text and visual encoders to learn the cross-modal alignment by contrastive learning. + +Though the pixel-level paradigm has been widely adopted recently, existing Chinese video-text models [14,26,45] are all feature-level ones pre-trained on private datasets. In this work, we provide the first pixel-level benchmarks pre-trained on our CNVid-3.5M dataset. + +Video-Text Dataset. Well-designed large-scale datasets are a prerequisite for the successful applications of the "Pre-training & Fine-tuning" paradigm. There exist a few large-scale public English datasets (e.g., Howto100M [25], WebVid-2.5M [4]) for video-text pre-training. However, existing large-scale Chinese video datasets (ALIVOL [14], Kwai-SVC [26], CREATE [45]) are all private, whose videos and text are unavailable to the public. Therefore, we aimed to release a public one to facilitate future researchers. + +Hard Sample Mining in Contrastive Learning. Hard Sample Mining [12, 29] is an effective strategy to promote + +
Dataset Name#(Videos)#(Text)Duration (hours)Video SourceText TypeAvailability
Off-FeatRaw-VidState
Datasets for Downstream Chinese Video-text Tasks
VATEX [38]41,269825,380115kinetics-600Caption✓✓All-public
BFVD [44]43,16643,166140E-CommerceTitle✓-Partly-public
FFVD [44]32,76332,763252E-CommerceTitle✓-Partly-public
CREATE210K [45]216,303268,5931,800Open WebsitesCaption--NOT-public
Kwai-SVC [26]222,077143,5693,500Open WebsitesTitle & ASR✓-Partly-public
Datasets for Chinese Video-text Pre-training
ALIVOL-10M [14]10,300,00011,000,00098,801E-CommerceTitle--NOT-public
Kwai-SVC-11M [26]11,075,0843,931,879177,200Open WebsitesTitle & ASR--NOT-public
CREATE-10M [45]10,000,00010,000,00083,000Open WebsitesTitle--NOT-public
CNVid-3.5M (ours)3,508,1203,508,12035,414Open WebsitesTitle & ASR✓✓All-public
+ +Table 1. Data analysis of CNVid-3.5M and other Chinese video-text datasets. "Off-Feat" and "Raw-Vid" represent whether offline features and raw videos are available. As far as we know, CNVid-3.5M is the largest public video-text dataset in Chinese. + +![](images/0807d4357590c810673f4124c45b001682cf285b1fa50e9e71996407f54d152e.jpg) +Figure 2. An example of video-text pairs in our CNVid-3.5M dataset. We present three types of text information: 1) Caption (manually written), 2) Title (user uploaded), and 3) ASR text. + +the performance of contrastive learning. In video-text pretraining, TACo [43] proposes a cascade sampling method to construct hard negative examples for better cross-modal alignment. However, we conjectured that hard sample mining may hinder the video-text pre-training at the beginning, where the model is far from convergence and could hardly discriminate the ground-truth video-text pairs. Therefore, we introduced curriculum learning [32, 37] and proposed the Hard Sample Curriculum Learning (HSCL) strategy. HSCL would guide the pre-training procedure from "starting small" gradually to "learning hard", mitigating the side effect of conventional hard sample mining methods. + +# 3. The CNVid-3.5M Dataset + +To fill in the blank of public Chinese video-text datasets, we release CNVid-3.5M, a large-scale cross-modal dataset containing over 3.5M Chinese video-text pairs. + +# 3.1. Dataset Building + +We collect raw videos from Douyin*, a popular Chinese social website, where millions of active users upload various types of videos every day. To enrich the data diversity, we do not limit video categories like HowTo100M [25] (mainly instructional videos) and cover various topics (e.g., food, entertainment, and technology). I.e., we first form up a set of keywords and then leverage them to search for videos and corresponding captions. Moreover, we span the data created time at a range of 3 years and filter the abnormal ones whose videos could not be played. In this way, we collect over 4.5M raw videos in Chinese. + +Besides, the supplementary material will present more details about fairness and privacy for safety and ethics. + +# 3.2. Captions, Titles, and ASR Text + +As CREATE [45] pointed out, three types of text information are widely-employed in current Chinese video-text datasets. We show an example in Figure 2, based on which we summarize their characteristics and applications: + +1) Captions are manually written by human annotators. They could accurately describe the video content in a formal manner. However, obtaining a large number of captions is extremely expensive. Therefore, captions are mainly employed in small-scale downstream video-text datasets like VATEX [38]. Moreover, captions are indeed global descriptions of the whole videos. Therefore, they could hardly describe the frame-level information of videos. + +2) Titles are written and uploaded by users. Compared with captions, it is not unfeasible to obtain millions of user-uploaded titles. However, to increase the attractiveness and click rate, users may over-polish titles with some "appealing but redundant" words, which may decrease the video-text consistency. Similar to captions, titles are video-level + +descriptions rather than frame-level ones. Therefore, titles and captions are mainly employed in feature-level videotext pre-training methods rather than pixel-level ones. + +3) ASR text is obtained by the ASR toolkit. Similar to titles, it is easy to obtain a large number of ASR text. However, some ASR text is incomplete, noisy, and even faulty, which may decrease the pre-training performance. Besides, ASR text is a frame-level description, which is widely employed in the pixel-level video-text pre-training. + +Following previous methods [25, 26], we collect user-uploaded titles and obtain the ASR text of each video. Since we provide both video-level titles and frame-level ASR text, our dataset could support the video-text pre-training in both feature-level and pixel-level manner. + +# 3.3. Weakly-paired Data Filtering + +As some work [25, 26] has pointed out, some ASR text may not match with associated videos (e.g., videos have background music), which would decrease the pre-training performance. However, the problem is only raised with few practical solutions. Therefore, we propose a novel weakly-paired data filtering strategy by leveraging a well-trained image-text model to automatically evaluate the video-text consistency for three reasons: 1) The text information in existing image-text datasets [11, 13] are manually-written titles or captions, whose quality and consistency are guaranteed. 2) Pixel-level pre-training methods [8, 15, 23] usually sample several frames from raw videos to serve as the visual input, whose frameworks are closely associated with image-text pre-trained models. 3) Compared with manually checking millions of videos, the proposed filtering strategy remarkably reduces the cost and improves the efficiency. + +Specifically, we first pre-train an image-text model on Wukong-100M [11]. We then leverage it to calculate the similarity score between videos and ASR text, and filter out the last 1M videos by their ranks. In this way, we obtain the CNVid-3.5M dataset, based on which we provide three mainstream pixel-level pre-training benchmarks. + +# 3.4. Comparison of Dataset Statistics + +Table 1 presents the statistics of CNVid-3.5M and other Chinese video-text datasets. To the best of our knowledge, CNVid-3.5M is the largest public dataset for Chinese video-text pre-training, with a total of 3.5M videos that last for 35.4K hours. Though ALIVOL-10M [14], KwaiSVC-11M [26], and CREATE-10M [45] contain more Chinese videos than ours, yet they do not release raw videos and text information to the public. Moreover, these three methods only provide conventional feature-level pretrained models, whose performance highly depends on frozen feature extractors. Towards this end, we provide the first pixel-level benchmarks based on our CNVid-3.5M dataset, which would pave the way for future research on + +Chinese video-text pre-training. + +Moreover, we present additional statistics of our CNVid-3.5M dataset in the supplementary material, including the distributions of topics, keywords, video durations, and Part-of-Speech (POS) tags of the ASR text. + +# 4. Pixel-level Pre-training Paradigm + +Since mainstream pixel-level video-text pre-training methods [17, 23, 35] extract cross-modal representations from raw data directly, they usually outperform the conventional feature-level ones [2, 24, 33]. Therefore, we benchmark our CNVid-3.5M dataset with three mainstream pixel-level architectures. As illustrated in Figure 3, we split the pixel-level pre-training process into three steps, namely: 1) data pre-processing to process raw videos and text into patches and tokens, 2) model architectures to generate cross-modal representations, and 3) proxy tasks to determine overall pre-training objectives. + +# 4.1. Data Pre-processing + +Given a mini-batch (denoted as $\mathcal{B}$ ) of videos $\{V_i\}_{i=1}^{|B|}$ and their corresponding ASR text $\{T_i\}_{i=1}^{|B|}$ , pixel-level pre-training methods would first sparsely (and randomly) sample $N_v$ frames from each video. We then slice each frame into patches, obtaining visual patch embeddings $\mathbf{P} \in \mathbb{R}^{N_v * N_p * d}$ , where $N_p$ is the number of sliced patches, and $d$ is the dimension of the embedding. Simultaneously, a BERT embedder is employed to process the text $T$ into fixed-length word embeddings $\mathbf{W} = [\mathbf{w}^{cls}, \mathbf{w}^1, \mathbf{w}^2, \dots, \mathbf{w}^{N_t - 1}]$ , where $\mathbf{W} \in \mathbb{R}^{N_t * d}$ , $N_t$ is the length of text tokens. + +# 4.2. Model Architectures + +There are three mainstream pixel-level video-text pretraining architectures, namely three-encoders-fusion [15, 17, 30], two-encoders-fusion [9], and twin-towers-crossed [8, 23, 36]. We benchmark CNVid-3.5M with these three pixel-level pre-training architectures. + +Three-encoders-fusion (3-E-F) pre-training methods contain a visual encoder $E_{vis}$ , a text encoder $E_{txt}$ , and a cross-modal encoder $E_{mul}$ . After obtaining visual features $\mathbf{V} = E_{vis}(\mathbf{P})$ and textual features $\mathbf{T} = E_{txt}(\mathbf{W})$ , we concatenate these two-modal features into $[\mathbf{T},\mathbf{V}] = [\mathbf{t}^{cls},\mathbf{t}^1,\mathbf{t}^2,\dots ,\mathbf{t}^{N_t - 1},\mathbf{v}^{cls},\mathbf{v}^1,\mathbf{v}^2,\dots ,\mathbf{v}^{N_v}]$ . Following CLIP4Clip [23], we set position embeddings to zero, and obtain cross-modal features $\mathbf{M} = E_{mul}([\mathbf{T},\mathbf{V}])$ . This forwarding process could be formulated as follows: + +$$ +\mathbf {M} = E _ {m u l} ([ E _ {t x t} (\mathbf {W}), E _ {v i s} (\mathbf {P}) ]), \tag {1} +$$ + +where $\mathbf{M} = [\mathbf{m}^{cls},\mathbf{m}^1,\mathbf{m}^2,\dots ,\mathbf{m}^{N_t + N_v}]$ , and $[\cdot ,\cdot ]$ denotes the concatenation operation. + +
1. Data Pre-processing2. Three Types of Model Architecture3. Proxy Tasks
Text Embedding +C | | | | BERT Embedder +| | | | | [CLS] a cat plays piano +[CLS] 猫弹钢琴 +ASR Caption +| | | | | Input VideoThree-Encoders-Fusion +Cross-Modal Feature +| | | | | Cross-Modal Encoder +| | | | | Emut | | | | Text FeatureEbert +Shared Encoder +| | | | | C | | | | Video Feature +| | | | | Shared Encoder +| | | | | C | | | | Video Encoder +| | | | | C | | | | Cross-Modal EncodingTwo-Encoders-Fusion +Cross-Modal Feature +| | | | | C | | | | Video Feature +| | | | | C | | | | Video Encoder +| | | | | C | | | | EvisTwin-Towers-Crossed +Contrastive Learning +| | | | | C | | | | Video Feature +| | | | | C | | | | Video Encoder +| | | | | C | | | | EvisMasked Language Modeling (MLM) +FC +[LMSK] +[C | | | | F | | | | F | | | | F | | | | F | | | | F | | | | F | | | | F | | | | F | | | | F | | | | F | | | | F | | | | F | | | | F | | | | F | | | | F | | | | F | | | | F | | | | F | | | | F | | | | F | | | |
+ +Figure 3. The pipeline of mainstream pixel-level video-text pre-training methods, which contains three key steps: 1) Data pre-processing, 2) model architectures, and 3) proxy tasks. Note that we employ three mainstream architectures to enrich pixel-level pre-training benchmarks. + +Two-encoders-fusion (2-E-F) pre-training methods simplify the 3-E-F architecture by employing a shared BERT-type encoder $E_{bert}$ to embed textual and cross-modal features simultaneously. The pipeline of obtaining textual features $\mathbf{T}$ and cross-modal features $\mathbf{M}$ could be formulated as follows: + +$$ +\left\{ \begin{array}{l} \mathbf {T} = E _ {b e r t} (\mathbf {W}), \\ \mathbf {M} = E _ {b e r t} ([ \mathbf {W}, E _ {v i s} (\mathbf {P}) ]). \end{array} \right. \tag {2} +$$ + +Twin-towers-crossed $(T-T-C)$ pre-training methods mainly follow the conventional CLIP [28] architecture. After obtaining visual and textual features, $T-T-C$ methods employ contrastive learning to optimize the twin encoders $E_{vis}$ and $E_{txt}$ for better cross-modal alignment. Note that $T-T-C$ methods do not generate cross-modal video representations like $3-E-F$ and $2-E-F$ . + +# 4.3. Proxy Tasks + +Proxy tasks are crucial for video-text pre-training as they directly determine the final optimization objectives. Following the conventional protocol [4, 15, 43], we adopt two widely-employed proxy tasks, i.e., Masked Language Modeling (MLM) and Video-Text Matching (VTM). + +MLM first masks out a certain percentage of words in a given sentence, and then forces the model to restore these clozes according to visual and textual cues. MLM is calculated twice for textual features $\mathbf{T}(\mathcal{L}_1)$ and cross-modal features $\mathbf{M}(\mathcal{L}_2)$ as follows: + +$$ +\mathcal {L} _ {1} = \frac {1}{| \mathcal {Q} |} \sum_ {q \in \mathcal {Q}} \mathcal {L} _ {C E} \left(y ^ {q}, \Theta_ {1} \left(\mathbf {t} ^ {q}\right)\right), \tag {3} +$$ + +$$ +\mathcal {L} _ {2} = \frac {1}{| \mathcal {Q} |} \sum_ {q \in \mathcal {Q}} \mathcal {L} _ {C E} \left(y ^ {q}, \Theta_ {2} \left(\mathbf {m} ^ {q}\right)\right), \tag {4} +$$ + +where $\mathcal{Q}$ records the location of masked tokens, $|\cdot|$ denotes the length of a given set, $y^{q}$ denotes the ground-truth token label, $\Theta$ is a Multi-Layer Perception (MLP), and $\mathcal{L}_{CE}$ is the regular Cross-Entropy cost function. + +VTM aims to promote the cross-modal alignment between videos and text, which is calculated in a parameter-free $(\mathcal{L}_3)$ and parameter-employed $(\mathcal{L}_4)$ way: + +$$ +\mathcal {L} _ {3} = - \sum_ {i = 1} ^ {| \mathcal {B} |} \log \frac {\sum_ {k = 1} ^ {N _ {v}} \exp \left\langle \mathbf {v} _ {i} ^ {k} , \mathbf {t} _ {i} ^ {c l s} \right\rangle}{\sum_ {k = 1} ^ {N _ {v}} \left(\exp \left\langle \mathbf {v} _ {i} ^ {k} , \mathbf {t} _ {i} ^ {c l s} \right\rangle + \sum_ {j \neq i} \exp \left\langle \mathbf {v} _ {j} ^ {k} , \mathbf {t} _ {i} ^ {c l s} \right\rangle\right)}, \tag {5} +$$ + +$$ +\mathcal {L} _ {4} = - \sum_ {i = 1} ^ {| \mathcal {B} |} \log \frac {\exp^ {\Theta_ {4} \left(\mathbf {f} _ {i , i}\right)}}{\exp^ {\Theta_ {4} \left(\mathbf {f} _ {i , i}\right)} + \sum_ {j \neq i} \exp^ {\Theta_ {4} \left(\mathbf {f} _ {j , i}\right)}}, \tag {6} +$$ + +where $|\mathcal{B}|$ is the length of a mini-batch, $\langle \cdot ,\cdot \rangle$ denotes the matrix multiplication operation, and $\mathbf{f}_{j,i}$ is global-pooling cross-modal features of the video-text pair $(V_{j},S_{i})$ + +For $3 - E - F$ and $2 - E - F$ models, the objective functions $\mathcal{L}_{3EF} / \mathcal{L}_{2EF}$ are calculated as follows: + +$$ +\mathcal {L} _ {3 E F} \left(\mathcal {L} _ {2 E F}\right) = \mathcal {L} _ {1} + \mathcal {L} _ {2} + \mathcal {L} _ {3} + \mathcal {L} _ {4}. \tag {7} +$$ + +While for $T$ - $T-C$ models, since they do not contain a cross-modal encoder to generate video representations, their objective function $\mathcal{L}_{TTC}$ is calculated as follows: + +$$ +\mathcal {L} _ {T T C} = \mathcal {L} _ {1} + \mathcal {L} _ {3}. \tag {8} +$$ + +# 4.4. Fine-tuning + +In order to thoroughly evaluate the pre-training performance, we fine-tune our models on the Text-to-Video Retrieval (TVR) task. TVR is a widely-employed downstream video-text task, aiming to retrieve the most relevant videos according to text queries. Since there has only one Chinese TVR dataset (VATEX [38]), we translate two widely-adopted English datasets (MSRVTT [41] and DiDemo [3]) into Chinese to verify the superiority and robustness of the constructed dataset and proposed methods. + +Specifically, we fine-tune our pre-trained models by reusing Video-Text Matching objectives ( $\mathcal{L}_3$ in Eq. 5 and $\mathcal{L}_4$ in Eq. 6, we only employ $\mathcal{L}_3$ for $T$ - $T-C$ models). + +![](images/9706241771516cafa10cc722eb7557c1d8ff1513a795cdc863411f7b57d921cc.jpg) +Figure 4. The pipeline of Hard Sample Curriculum Learning (HSCL), which contains three components to promote the performance of video-text pre-training. + +# 5. Hard Sample Curriculum Learning + +We propose Hard Sample Curriculum Learning (HSCL), a model-agnostic strategy to promote the video-text pretraining performance. HSCL is based on two motivations. 1) The negative samples involved in the calculation of parameter-employed VTM loss $(\mathcal{L}_4$ in Eq. 6) are not representative. Suppose that the batch size is set to be $N_{k}$ , a model needs to process $N_{k} * N_{k}$ video-text pairs for calculating $\mathcal{L}_4$ , whose calculation cost is hardly affordable. Therefore, previous methods [9, 15, 22] tend to cut the number of negative samples to $N_{k}^{\prime}$ by random sampling. However, since $N_{k}^{\prime}$ is usually much smaller than $N_{k}$ , it may result in the sub-optimal performance. 2) If a model is far from convergence, directly introducing hard negative samples would amplify the burden of learning, which would decrease the pre-training performance. + +Therefore, inspired by hard sample mining and curriculum learning, we propose the HSCL strategy. By leveraging the similarity score calculated by the parameter-free VTM loss $(\mathcal{L}_3$ in Eq. 5), HSCL could gradually and smoothly introduce hard samples for better cross-modal alignment. As illustrated in Figure 4, HSCL contains the following three key components: + +Positive Column Weighting (PCW) aims to decrease the learning coefficient $\phi$ of well-learned positive samples, forcing the model to emphasize those hard positive ones. Specifically, we re-weigh half of samples within the batch according to their similarity scores $z$ calculated by $\mathcal{L}_3$ , which could be formulated as follows: + +$$ +z _ {j, i} = \frac {1}{N _ {v}} \sum_ {k = 1} ^ {N _ {v}} \exp \left\langle \mathbf {v} _ {j} ^ {k}, \mathbf {t} _ {i} ^ {c l s} \right\rangle , \tag {9} +$$ + +$$ +\phi_ {i} = \operatorname {M i n} \left(\frac {\operatorname {M e d} \left(Z _ {d i a g}\right) - \operatorname {M i n} \left(Z _ {d i a g}\right)}{z _ {i , i} - \operatorname {M i n} \left(Z _ {d i a g}\right) + e _ {0}}, 1. 0\right), \tag {10} +$$ + +where $Z_{diag} = [z_{1,1}, z_{2,2}, \dots, z_{i,i}, \dots]$ , $\operatorname{Med}(\cdot)$ and $\operatorname{Min}(\cdot)$ represent the median and minimum result of the list, $e_0$ is a pre-defined coefficient to avoid a zero denominator. + +Negative Line Sampling (NLS) aims to sample $N_{k}^{\prime}$ hard negative examples for better contrastive learning. Specifi- + +cally, for each text $T_{j}$ , we choose the Nearliest $N_{k}^{\prime}$ aligned negative videos according to their similarity scores in the list $Z_{j} = [z_{j,1},z_{j,2},\dots ,z_{j,i},\dots ]$ . I.e., Pick the minimum $N_{k}^{\prime}$ samples in the list $Z_{j}^{\prime} = \{|z_{j,i} - z_{j,j}|\}_{j\neq i}$ + +Curriculum Learning (CL) aims to mitigate the side effect of hard sample mining when the model is under convergence. We set a gradually increasing coefficient $\beta \in (0,1)$ . $\beta$ is initialized with zero at the beginning of pre-training. It will gradually increase by a fixed number $\beta_0$ for each $N_d$ iteration until reaching 1.0. For PCW, the actual coefficient $\phi^*$ is calculated as follows: + +$$ +\phi_ {i} ^ {*} = (1. 0 - \beta) + \beta * \phi_ {i}. \tag {11} +$$ + +For NLS, we first generate a random number $rand \in (0,1)$ , and only perform the NLS step if $rand < \beta$ . In this way, the model would conduct conventional pre-training at the beginning, and smoothly change to learn hard negative samples to pursue better pre-training performance. + +# 6. Experiments and Benchmarks + +# 6.1. Pre-training Datasets + +In Section 3.3, the Chinese image-text model that helps to filter out the weakly-paired data is pre-trained on Wukong-100M [11] (100M image-text pairs). The Chinese video-text models are pre-trained on our CNVid-3.5M dataset, which contains 3.5M video-text pairs. + +To verify the generality of the proposed Hard Sample Curriculum Learning (HSCL) strategy, we also pre-trained our models on four large-scale English datasets, they are 1) COCO [20] (0.6M image-text pairs), 2) VG [13] (5.4M image-text pairs), 3) CC [31] (3.1M image-text pairs), and 4) WebVid-2.5M (2.5M video-text pairs). + +# 6.2. Fine-tuning Datasets + +We fine-tuned our pre-trained models on the Text-to-Video Retrieval (TVR) task, which includes three widely-adopted datasets as follows: 1) VATEX [38] contains 28K video clips, and each video is associated with 10 Chinese and 10 English diverse captions. 2) MSRVTT [41] contains 10K video clips associated with 200K English captions. Following [15, 22, 43], we used 7K videos for training and randomly selected 1K videos from the remaining ones for testing (7K-1K split). 3) DiDemo [3] contains 10K Flickr videos associated with 40K English captions. + +Moreover, for a comprehensive performance evaluation of Chinese video-text pre-trained models, we employ Google $^{\dagger}$ to translate all captions in MSRVTT and DiDemo from English to Chinese. + +Metrics. We employed Recall@K (R@K, K=1/5/10) and Median Rank (MdR) to measure the TVR performance. + +
No.Model TypeData VolumeMasking StrategyMatching StrategyVATEX (Chinese)MSRVTT (Translated)DiDemo (Translated)
R@1/5/10↑(MdR↓)R@1/5/10↑(MdR↓)R@1/5/10↑(MdR↓)
+ +For "Building": Ablation Study of the constructed CNVid-3.5M dataset + +
A12-E-FNo Pre-training36.4 / 75.4 / 85.2 (2)15.9 / 42.1 / 54.1 (8)8.0 / 24.8 / 39.2 (18)
A22-E-F3.5MMLMVTM39.9 / 77.2 / 87.0 (2)20.7 / 47.9 / 61.2 (6)13.3 / 34.6 / 45.9 (13)
+ +For "Filtering": Ablation Study of Weakly-Paired Data Filtering + +
B12-E-F4.5MMLMVTM38.8 / 76.4 / 86.4 (2)19.4 / 45.6 / 58.6 (6)10.6 / 32.7 / 45.0 (14)
B22-E-F4.0MMLMVTM39.5 / 77.0 / 86.5 (2)19.8 / 47.2 / 60.3 (6)12.4 / 32.8 / 45.6 (14)
B32-E-F3.5MMLMVTM39.9 / 77.2 / 87.0 (2)20.7 / 47.9 / 61.2 (6)13.3 / 34.6 / 45.9 (13)
B42-E-F3.0MMLMVTM40.5 / 76.8 / 86.5 (2)20.6 / 47.3 / 61.3 (6)12.2 / 34.3 / 45.6 (14)
B52-E-F2.5MMLMVTM39.5 / 76.6 / 86.9 (2)20.1 / 47.4 / 60.8 (6)12.1 / 33.6 / 46.0 (14)
+ +For "Pre-training": Ablation Study of Hard Sample Curriculum Learning + +
C12-E-F3.5MMLMVTM39.9 / 77.2 / 87.0 (2)20.7 / 47.9 / 61.2 (6)13.3 / 34.6 / 45.9 (13)
C22-E-F3.5MMLMVTM+HSM40.7 / 76.8 / 86.9 (2)21.1 / 47.3 / 60.1 (6)13.4 / 34.3 / 46.4 (12)
C32-E-F3.5MMLMVTM+HSCL41.5 / 78.2 / 87.2 (2)23.3 / 48.0 / 61.2 (6)13.6 / 34.4 / 47.3 (12)
+ +Table 2. For Chinese video-text pre-training: Ablation study of Weakly-Paired Data Filtering and Hard Sample Curriculum Learning (HSCL) on the Text-to-Video Retrieval task of three datasets (VATEX-Chinese, MSRVTT-Translated, and DiDemo-Translated). "HSM" denotes the conventional Hard Sample Mining strategy without curriculum learning. Note that $A2$ , $B3$ , and $C1$ are the same model. + +
No.Pre-training DatasetMasking StrategyMatching StrategyVATEX (English)MSRVTTDiDemo
R@1/5/10 ↑ (MdR ↓)R@1/5/10 ↑ (MdR ↓)R@1/5/10 ↑ (MdR ↓)
D0No Pre-training39.5 / 78.3 / 88.6 (2)19.2 / 46.2 / 59.5 (6)14.2 / 34.1 / 46.1 (12)
D1COCO+VG+CCMLMVTM45.5 / 81.5 / 91.0 (2)28.6 / 55.5 / 66.1 (4)25.3 / 50.3 / 62.4 (5)
D2COCO+VG+CCMLMVTM+HSCL48.7 / 83.1 / 91.1 (2)31.6 / 56.5 / 66.7 (4)27.1 / 53.8 / 63.8 (4)
D3CC+WebVidMLMVTM50.5 / 85.6 / 92.5 (1)29.4 / 56.9 / 68.3 (4)29.0 / 57.3 / 67.8 (4)
D4CC+WebVidMLMVTM+HSCL53.4 / 86.2 / 92.9 (1)32.6 / 58.8 / 69.7 (3)29.8 / 58.0 / 68.9 (3)
+ +Table 3. For English video-text pre-training: Ablation study of Hard Sample Curriculum Learning (HSCL) on the Text-to-Video Retrieval task of three datasets (VATEX-English, MSRVTT, and DiDemo). All models are based on the two-encoders-fusion $(2 - E - F)$ architecture. + +# 6.3. Experimental Settings + +Settings for Weakly-paired Data Filtering. We pretrained the image-text model on the Wukong-100M [11] dataset, whose pre-training details are presented in the supplementary material. We then employed the pre-trained model to obtain the parameter-free video-text consistency score $(z_{i,i}$ in Eq. 9) of 4.5M video-text pairs. To improve the reliability, we randomly sampled four frames to calculate the consistency score, and repeated this step for three times. The final score of each video is the average of three results calculated under different random seeds. Ultimately, we sorted the consistency score in descending order, and filtered out the last 1M videos to refine the dataset. + +Settings for Chinese Video-Text Pre-training. For model architectures, we employed the Video Swin Transformer [21] to serve as the visual encoder, which is initialized with parameters pre-trained on ImageNet [6]. The text, cross-modal, and shared encoders belong to the BERT-Base [7] model, which are initialized with parameters privdied by Hugging Face $\ddagger$ . We sparsely sampled $4(N_v)$ frames from raw videos during pre-training. For the MLM task, the masking rate is $15\%$ . For the proposed Hard Sample + +Curriculum Learning, the coefficient $\beta$ would increase by $0.15(\beta_0)$ for each $5000(N_d)$ iterations until reaching 1.0. + +For hyper-parameters, we set the length of text tokens $N_{t} = 30$ , and the dimension of the hidden state $d = 768$ . All models are pre-trained by the Adam optimizer with a momentum of 0.9. The total pre-training stage lasts for 10 epochs with a batch size of 128. The initial learning rate is 5e-5 and is decayed by the factor of 10 after 5 epochs. The whole pre-training on CNVid-3.5M takes about 4 days to complete on 8 NVIDIA V100 GPUs. + +Settings for Fine-tuning on Downstream Chinese Videotext Tasks. The optimizer and hyper-parameters in fine-tuning remain the same as the pre-training configuration. The total fine-tuning stage lasts for 25,000 steps. The batch size is set to 128. The initial learning rate is set to 1e-5. + +# 6.4. Ablation Study + +As aforementioned, we summarized our work with three verbs, i.e., "Build", "Filter", and "Pre-train". To evaluate the effectiveness of these three contributions, we took the 2- $E-F$ architecture as an example model in the ablation study. Ablation Study of "Building". We compared the performance obtained by models with or without pre-training. As illustrated in Table 2 (A1 vs. A2), models pre-trained on + +
No.Model TypeDatasetPre-training StrategyVATEX (Chinese)MSRVTT (Translated)DiDemo (Translated)
R@1/5/10 ↑ (MdR ↓)R@1/5/10 ↑ (MdR ↓)R@1/5/10 ↑ (MdR ↓)
E13-E-FCNVid-3.5MMLM+VTM39.9 / 78.2 / 87.6 (2)20.6 / 47.1 / 60.0 (6)13.8 / 34.5 / 46.6 (13)
E22-E-FCNVid-3.5MMLM+VTM39.9 / 77.2 / 87.0 (2)20.7 / 47.9 / 61.2 (6)13.3 / 34.6 / 45.9 (13)
E3T-T-CCNVid-3.5MMLM+VTM40.5 / 77.3 / 87.0 (2)21.1 / 46.5 / 58.9 (6)13.0 / 32.6 / 44.6 (15)
+ +Table 4. We benchmark CNVid-3.5M with three mainstream pixel-level pre-training architectures, which is the first pixel-level benchmarks for Chinese video-text pre-training. + +our CNVid-3.5M dataset (A2) would achieve a remarkable performance gain compared with those without pre-training (A1), which demonstrates the effectiveness of building a large-scale Chinese video-text dataset. + +Ablation Study of "Filtering". We compared the performance obtained by models pre-trained under different numbers of video-text pairs. For example, "4.5M", "4.0M", and "3.5M" denotes that we employ the 4.5M full set, the Top-4.0M split set, and the Top-3.5M split set (CNVid-3.5M) according to the sorted video-text consistency scores calculated by weakly-paired data filtering. As illustrated in Table 2 (B1-B5), we have two conclusions as follows: + +1) The proposed weakly-paired data filtering strategy could effectively improve the quality of datasets. The fine-tuning performance on three downstream datasets first increases rapidly when filtering out videos with low video-text consistency scores ( $B1 \to B2 \to B3$ ). It would remain stable ( $B3 \to B4$ ) or drop slightly ( $B4 \to B5$ ) when continuously removing videos after the Top-3.5M split. +2) In general, employing the Top-3.5M split (B3) achieves the best performance. The results among all ablation models (B1-B5) may prove that, for the video-text pre-training, it is not "more data is better", but "more good data is better". Therefore, we filtered out the last 1M videos, resulting in the CNVid-3.5M dataset. + +Ablation Study of "Pre-training" (HSCL). To prove the effectiveness of Hard Sample Curriculum Learning (HSCL), we pre-trained several ablation models on Chinese video-text datasets ( $C1-C3$ in Table 2) and English ones ( $D1-D4$ in Table 3). We have two conclusions as follows: + +1) HSCL is a model-agnostic strategy, which could evidently improve the pre-training performance. In Table 2, compared with the baseline (C1), HSCL (C3) achieves an improvement of $1.6\%$ , $2.6\%$ , and $0.3\%$ at R@1 of VATEX, MSRVTT, and DiDemo dataset, respectively. While in Table 3, HSCL outperforms the baseline by $3.2 / 3.0 / 1.8\%$ ( $D1$ vs. $D2$ on $\mathrm{COCO + VG + CC}$ ) and $2.9 / 3.2 / 0.8\%$ ( $D3$ vs. $D4$ on $\mathrm{CC + WebVid}$ ) at R@1 of three TVR datasets. +2) The proposed Curriculum Learning (CL) strategy in HSCL could effectively mitigate the side effect of conventional Hard Sample Mining (HSM) when the model is far from convergence. As illustrated in Table 2 (C1 vs. C2), models equipped with HSCL perform better than HSM. + +Moreover, we present the detailed quantitative analyses of PCW and NLS in the supplementary material. + +# 6.5. Chinese Video-text Pre-training Benchmarks + +Based on the constructed CNVid-3.5M dataset, We provided the first pixel-level benchmarks for Chinese videotext pre-training. Following existing video-text pre-training methods, we adopted three mainstream pixel-level architectures, i.e., three-encoders-fusion (3-E-F) [15, 17, 30], two-encoders-fusion (2-E-F) [9], and twin-towers-crossed (T-T-C) [8, 23, 36]. Detailed performance comparisons of three pixel-level benchmarks are presented in Table 4. + +# 7. Conclusion + +In this work, we release the largest public Chinese video-text dataset, i.e., CNVid-3.5M. Besides, we provide the first pixel-level benchmarks for Chinese video-text pretraining. Moreover, we propose a novel weakly-paired data filtering method to improve the quality of datasets. We also design a novel Hard Sample Curriculum Learning strategy to promote the pre-training performance. In conclusion, we "build", "filter", and "pre-train" the large-scale Chinese video-text dataset. We believe this work would pave the way for future research on Chinese video-text pre-training. + +# 8. Limitations and Future Work + +Address the Difference between English and Chinese. One potential limitation of our work is that we adopt pixel-level architectures proved effective in English video-text pre-training, ignoring the linguistic gap between Chinese and English. E.g., it may be better to employ the Whole Word Masking (WWM) [5, 46] strategy rather than the conventional MLM for Chinese video-text pre-training. We plan to explore more Chinese-specific methods in the future. Need More Downstream Chinese Video-Text Datasets. Though we release a large-scale Chinese video-text dataset, there still lack some downstream datasets (e.g., MSRVTT [41], MSVD-QA [39] in English) to comprehensively evaluate the performance of Chinese video-text pre-trained models. Currently, there only exists VATEX [38] for Chinese cross-modal retrieval and some captioning datasets [26, 45]. Therefore, building more downstream Chinese video datasets is a promising future direction. + +Acknowledgments. This work is supported by the National Natural Science Foundation of China, No.: 62176137; Ant Group. + +# References + +[1] Nayyer Aafaq, Ajmal Mian, Wei Liu, Syed Zulqarnain Gilani, and Mubarak Shah. Video description: A survey of methods, datasets, and evaluation metrics. ACM Computing Surveys, 52(6):1-37, 2019. 1 +[2] Hassan Akbari, Liangzhe Yuan, Rui Qian, Wei-Hong Chuang, Shih-Fu Chang, Yin Cui, and Boqing Gong. Vatt: Transformers for multimodal self-supervised learning from raw video, audio and text. Advances in Neural Information Processing Systems, 34, 2021. 4 +[3] Lisa Anne Hendricks, Oliver Wang, Eli Shechtman, Josef Sivic, Trevor Darrell, and Bryan Russell. Localizing moments in video with natural language. In Proceedings of the IEEE international conference on computer vision, pages 5803-5812, 2017. 5, 6 +[4] Max Bain, Arsha Nagrani, Gül Varol, and Andrew Zisserman. Frozen in time: A joint video and image encoder for end-to-end retrieval. In Proceedings of the IEEE International Conference on Computer Vision, pages 1728-1738, 2021. 1, 2, 5 +[5] Yiming Cui, Wanxiang Che, Ting Liu, Bing Qin, and Ziqing Yang. Pre-training with whole word masking for chinese bert. IEEE Transactions on Audio, Speech, and Language Processing, 29:3504-3514, 2021. 8 +[6] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In IEEE Conference on Computer Vision and Pattern Recognition, pages 248-255, 2009. 7 +[7] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018. 7 +[8] Han Fang, Pengfei Xiong, Luhui Xu, and Yu Chen. Clip2video: Mastering video-text retrieval via image clip. arXiv preprint arXiv:2106.11097, 2021. 2, 4, 8 +[9] Tsu-Jui Fu, Linjie Li, Zhe Gan, Kevin Lin, William Yang Wang, Lijuan Wang, and Zicheng Liu. Violet: End-to-end video-language transformers with masked visual-token modeling. arXiv preprint arXiv:2111.12681, 2021. 2, 4, 6, 8 +[10] Yuying Ge, Yixiao Ge, Xihui Liu, Dian Li, Ying Shan, Xiaohu Qie, and Ping Luo. Bridging video-text retrieval with multiple choice questions. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 16167-16176, 2022. 1 +[11] Jiaxi Gu, Xiaojun Meng, Guansong Lu, Lu Hou, Minzhe Niu, Hang Xu, Xiaodan Liang, Wei Zhang, Xin Jiang, and Chunjing Xu. Wukong: 100 million large-scale chinese cross-modal pre-training dataset and a foundation framework. arXiv preprint arXiv:2202.06767, 2022. 2, 4, 6, 7 +[12] Yannis Kalantidis, Mert Bulent Sariyildiz, Noe Pion, Philippe Weinzaepfel, and Diane Larlus. Hard negative mixing for contrastive learning. Advances in Neural Information Processing Systems, 33:21798-21809, 2020. 2 +[13] Ranjay Krishna, Yuke Zhu, Oliver Groth, Justin Johnson, Kenji Hata, Joshua Kravitz, Stephanie Chen, Yannis Kalantidis, Li-Jia Li, David A Shamma, et al. Visual genome: + +Connecting language and vision using crowdsourced dense image annotations. International Journal of Computer Vision, 123(1):32-73, 2017. 4, 6 +[14] Chenyi Lei, Shixian Luo, Yong Liu, Wanggui He, Jiamang Wang, Guoxin Wang, Haihong Tang, Chunyan Miao, and Houqiang Li. Understanding chinese video and language via contrastive multimodal pre-training. In Proceedings of the ACM International Conference on Multimedia, pages 2567-2576, 2021. 1, 2, 3, 4 +[15] Jie Lei, Linjie Li, Luowei Zhou, Zhe Gan, Tamara L Berg, Mohit Bansal, and Jingjing Liu. Less is more: Clipbert for video-and-language learning via sparse sampling. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 7331-7341, 2021. 1, 2, 4, 5, 6, 8 +[16] Chunyuan Li, Xiujun Li, Lei Zhang, Baolin Peng, Mingyuan Zhou, and Jianfeng Gao. Self-supervised pre-training with hard examples improves visual representations. arXiv preprint arXiv:2012.13493, 2020. 2 +[17] Dongxu Li, Junnan Li, Hongdong Li, Juan Carlos Niebles, and Steven CH Hoi. Align and prompt: Video-and-language pre-training with entity prompts. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 4953-4963, 2022. 1, 2, 4, 8 +[18] Junnan Li, Ramprasaath Selvaraju, Akhilesh Gotmare, Shafiq Joty, Caiming Xiong, and Steven Chu Hong Hoi. Align before fuse: Vision and language representation learning with momentum distillation. Advances in neural information processing systems, 34:9694-9705, 2021. 2 +[19] Linjie Li, Yen-Chun Chen, Yu Cheng, Zhe Gan, Licheng Yu, and Jingjing Liu. Hero: Hierarchical encoder for video+language omni-representation pre-training. arXiv preprint arXiv:2005.00200, 2020. 1 +[20] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Proceedings of the European Conference on Computer Vision, pages 740-755. Springer, 2014. 6 +[21] Ze Liu, Jia Ning, Yue Cao, Yixuan Wei, Zheng Zhang, Stephen Lin, and Han Hu. Video swim transformer. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 3202-3211, 2022. 7 +[22] Huaishao Luo, Lei Ji, Botian Shi, Haoyang Huang, Nan Duan, Tianrui Li, Jason Li, Taroon Bharti, and Ming Zhou. Univl: A unified video and language pre-training model for multimodal understanding and generation. arXiv preprint arXiv:2002.06353, 2020. 1, 6 +[23] Huaishao Luo, Lei Ji, Ming Zhong, Yang Chen, Wen Lei, Nan Duan, and Tianrui Li. Clip4clip: An empirical study of clip for end to end video clip retrieval. arXiv preprint arXiv:2104.08860, 2021. 2, 4, 8 +[24] Jianjie Luo, Yehao Li, Yingwei Pan, Ting Yao, Hongyang Chao, and Tao Mei. Coco-bert: Improving video-language pre-training with contrastive cross-modal matching and denoising. In Proceedings of the ACM International Conference on Multimedia, pages 5600-5608, 2021. 2, 4 +[25] Antoine Miech, Dimitri Zhukov, Jean-Baptiste Alayrac, Makarand Tapaswi, Ivan Laptev, and Josef Sivic. + +Howto100m: Learning a text-video embedding by watching hundred million narrated video clips. In Proceedings of the IEEE International Conference on Computer Vision, pages 2630-2640, 2019. 1, 2, 3, 4 +[26] Liqiang Nie, Leigang Qu, Dai Meng, Min Zhang, Qi Tian, and Alberto Del Bimbo. Search-oriented microvideo captioning. In Proceedings of the ACM International Conference on Multimedia, pages 3234-3243, 2022. 1, 2, 3, 4, 8 +[27] Devshree Patel, Ratnam Parikh, and Yesha Shastri. Recent advances in video question answering: A review of datasets and methods. In International Conference on Pattern Recognition, pages 339-356. Springer, 2021. 1 +[28] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 2, 5 +[29] Joshua David Robinson, Ching-Yao Chuang, Suvrit Sra, and Stefanie Jegelka. Contrastive learning with hard negative samples. In International Conference on Learning Representations, 2020. 2 +[30] Paul Hongsuck Seo, Arsha Nagrani, Anurag Arnab, and Cordelia Schmid. End-to-end generative pretraining for multimodal video captioning. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 17959-17968, 2022. 1, 2, 4, 8 +[31] Piyush Sharma, Nan Ding, Sebastian Goodman, and Radu Soricut. Conceptual captions: A cleaned, hypernymed, image alt-text dataset for automatic image captioning. In Proceedings of the Annual Meeting of the Association for Computational Linguistics, pages 2556–2565, 2018. 6 +[32] Petru Soviany, Radu Tudor Ionescu, Paolo Rota, and Nicu Sebe. Curriculum learning: A survey. International Journal of Computer Vision, pages 1-40, 2022. 2, 3 +[33] Chen Sun, Austin Myers, Carl Vondrick, Kevin Murphy, and Cordelia Schmid. Videobert: A joint model for video and language representation learning. In Proceedings of the IEEE International Conference on Computer Vision, pages 7464-7473, 2019. 2, 4 +[34] Guanglu Sun, Lili Liang, Tianlin Li, Bo Yu, Meng Wu, and Bolun Zhang. Video question answering: a survey of models and datasets. Mobile Networks and Applications, 26(5):1904-1937, 2021. 1 +[35] Alex Jinpeng Wang, Yixiao Ge, Rui Yan, Yuying Ge, Xudong Lin, Guanyu Cai, Jianping Wu, Ying Shan, Xiaohu Qie, and Mike Zheng Shou. All in one: Exploring unified video-language pre-training. arXiv preprint arXiv:2203.07303, 2022. 2, 4 +[36] Jinpeng Wang, Yixiao Ge, Guanyu Cai, Rui Yan, Xudong Lin, Ying Shan, Xiaohu Qie, and Mike Zheng Shou. Object-aware video-language pre-training for retrieval. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 3313-3322, 2022. 1, 2, 4, 8 + +[37] Xin Wang, Yudong Chen, and Wenwu Zhu. A survey on curriculum learning. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2021. 2, 3 +[38] Xin Wang, Jiawei Wu, Junkun Chen, Lei Li, Yuan-Fang Wang, and William Yang Wang. Vatex: A large-scale, high-quality multilingual dataset for video-and-language research. In Proceedings of the IEEE International Conference on Computer Vision, pages 4581-4591, 2019. 3, 5, 6, 8 +[39] Dejing Xu, Zhou Zhao, Jun Xiao, Fei Wu, Hanwang Zhang, Xiangnan He, and Yueting Zhuang. Video question answering via gradually refined attention over appearance and motion. In Proceedings of the ACM International Conference on Multimedia, pages 1645-1653, 2017. 8 +[40] Hu Xu, Gargi Ghosh, Po-Yao Huang, Prahal Arora, Masoumeh Aminzadeh, Christoph Feichtenhofer, Florian Metze, and Luke Zettlemoyer. Vlm: Task-agnostic video-language model pre-training for video understanding. arXiv preprint arXiv:2105.09996, 2021. 2 +[41] Jun Xu, Tao Mei, Ting Yao, and Yong Rui. Msr-vtt: A large video description dataset for bridging video and language. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 5288-5296, 2016. 5, 6, 8 +[42] Antoine Yang, Antoine Miech, Josef Sivic, Ivan Laptev, and Cordelia Schmid. Just ask: Learning to answer questions from millions of narrated videos. In Proceedings of the IEEE International Conference on Computer Vision, pages 1686-1697, 2021. 1 +[43] Jianwei Yang, Yonatan Bisk, and Jianfeng Gao. Taco: Token-aware cascade contrastive learning for video-text alignment. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11562-11572, 2021. 2, 3, 5, 6 +[44] Shengyu Zhang, Ziqi Tan, Jin Yu, Zhou Zhao, Kun Kuang, Jie Liu, Jingren Zhou, Hongxia Yang, and Fei Wu. Poet: Product-oriented video captioner for e-commerce. In Proceedings of the ACM International Conference on Multimedia, pages 1292-1301, 2020. 3 +[45] Ziqi Zhang, Yuxin Chen, Zongyang Ma, Zhongang Qi, Chunfeng Yuan, Bing Li, Ying Shan, and Weiming Hu. Create: A benchmark for chinese short video retrieval and title generation. arXiv preprint arXiv:2203.16763, 2022. 1, 2, 3, 4, 8 +[46] Wei Zhu. Mvp-bert: Multi-vocab pre-training for chinese bert. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing: Student Research Workshop, pages 260-269, 2021. 8 \ No newline at end of file diff --git a/2023/CNVid-3.5M_ Build, Filter, and Pre-Train the Large-Scale Public Chinese Video-Text Dataset/images.zip b/2023/CNVid-3.5M_ Build, Filter, and Pre-Train the Large-Scale Public Chinese Video-Text Dataset/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..d366c235cd4e4553e2440bcfc6b900b86343fecb --- /dev/null +++ b/2023/CNVid-3.5M_ Build, Filter, and Pre-Train the Large-Scale Public Chinese Video-Text Dataset/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:70169a0dac79f2cc7de9a15f8aa7e371f4b4f6251ce94798eedf126fcc0098af +size 686306 diff --git a/2023/CNVid-3.5M_ Build, Filter, and Pre-Train the Large-Scale Public Chinese Video-Text Dataset/layout.json b/2023/CNVid-3.5M_ Build, Filter, and Pre-Train the Large-Scale Public Chinese Video-Text Dataset/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..fb1c2017ef8166193eaa36ef93c1dc9cba9f7aa2 --- /dev/null +++ b/2023/CNVid-3.5M_ Build, Filter, and Pre-Train the Large-Scale Public Chinese Video-Text Dataset/layout.json @@ -0,0 +1,9798 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 143, + 103, + 451, + 138 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 103, + 451, + 138 + ], + "spans": [ + { + "bbox": [ + 143, + 103, + 451, + 138 + ], + "type": "text", + "content": "CNVid-3.5M: Build, Filter, and Pre-train the Large-scale Public Chinese Video-text Dataset" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 65, + 160, + 556, + 176 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 160, + 556, + 176 + ], + "spans": [ + { + "bbox": [ + 65, + 160, + 556, + 176 + ], + "type": "text", + "content": "Tian Gan" + }, + { + "bbox": [ + 65, + 160, + 556, + 176 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 65, + 160, + 556, + 176 + ], + "type": "text", + "content": ", Qing Wang" + }, + { + "bbox": [ + 65, + 160, + 556, + 176 + ], + "type": "inline_equation", + "content": "^{2*}" + }, + { + "bbox": [ + 65, + 160, + 556, + 176 + ], + "type": "text", + "content": ", Xingning Dong" + }, + { + "bbox": [ + 65, + 160, + 556, + 176 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 65, + 160, + 556, + 176 + ], + "type": "text", + "content": ", Xiangyuan Ren" + }, + { + "bbox": [ + 65, + 160, + 556, + 176 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 65, + 160, + 556, + 176 + ], + "type": "text", + "content": ", Liqiang Nie" + }, + { + "bbox": [ + 65, + 160, + 556, + 176 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 65, + 160, + 556, + 176 + ], + "type": "text", + "content": ", Qingpei Guo" + }, + { + "bbox": [ + 65, + 160, + 556, + 176 + ], + "type": "inline_equation", + "content": "^{2\\dagger}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 123, + 178, + 215, + 191 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 123, + 178, + 215, + 191 + ], + "spans": [ + { + "bbox": [ + 123, + 178, + 215, + 191 + ], + "type": "text", + "content": "1Shandong University," + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 234, + 178, + 287, + 191 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 234, + 178, + 287, + 191 + ], + "spans": [ + { + "bbox": [ + 234, + 178, + 287, + 191 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 234, + 178, + 287, + 191 + ], + "type": "text", + "content": "Ant Group," + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 178, + 484, + 191 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 178, + 484, + 191 + ], + "spans": [ + { + "bbox": [ + 306, + 178, + 484, + 191 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 306, + 178, + 484, + 191 + ], + "type": "text", + "content": "Harbin Institute of Technology (Shenzhen)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 111, + 194, + 214, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 111, + 194, + 214, + 205 + ], + "spans": [ + { + "bbox": [ + 111, + 194, + 214, + 205 + ], + "type": "text", + "content": "gantian@sdu.edu.cn," + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 228, + 194, + 348, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 228, + 194, + 348, + 205 + ], + "spans": [ + { + "bbox": [ + 228, + 194, + 348, + 205 + ], + "type": "text", + "content": "wq176625@antgroup.com," + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 361, + 194, + 504, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 361, + 194, + 504, + 205 + ], + "spans": [ + { + "bbox": [ + 361, + 194, + 504, + 205 + ], + "type": "text", + "content": "dongxingning1998@gmail.com" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 91, + 208, + 511, + 219 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 208, + 511, + 219 + ], + "spans": [ + { + "bbox": [ + 91, + 208, + 511, + 219 + ], + "type": "text", + "content": "xiangyuan.rxy@antgroup.com, nieliqiang@gmail.com, qingpei.gqp@antgroup.com" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 143, + 247, + 192, + 260 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 247, + 192, + 260 + ], + "spans": [ + { + "bbox": [ + 143, + 247, + 192, + 260 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 47, + 272, + 289, + 583 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 272, + 289, + 583 + ], + "spans": [ + { + "bbox": [ + 47, + 272, + 289, + 583 + ], + "type": "text", + "content": "Owing to well-designed large-scale video-text datasets, recent years have witnessed tremendous progress in videotext pre-training. However, existing large-scale video-text datasets are mostly English-only. Though there are certain methods studying the Chinese video-text pre-training, they pre-train their models on private datasets whose videos and text are unavailable. This lack of large-scale public datasets and benchmarks in Chinese hampers the research and downstream applications of Chinese video-text pretraining. Towards this end, we release and benchmark CNVid-3.5M, a large-scale public cross-modal dataset containing over 3.5M Chinese video-text pairs. We summarize our contributions by three verbs, i.e., \"Build\", \"Filter\", and \"Pre-train\": 1) To build a public Chinese video-text dataset, we collect over 4.5M videos from the Chinese websites. 2) To improve the data quality, we propose a novel method to filter out 1M weakly-paired videos, resulting in the CNVid-3.5M dataset. And 3) we benchmark CNVid-3.5M with three mainstream pixel-level pre-training architectures. At last, we propose the Hard Sample Curriculum Learning strategy to promote the pre-training performance. To the best of our knowledge, CNVid-3.5M is the largest public video-text dataset in Chinese, and we provide the first pixel-level benchmarks for Chinese video-text pretraining. The dataset, codebase, and pre-trained models are available at https://github.com/CNVid/CNVid-3.5M." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 47, + 605, + 128, + 618 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 605, + 128, + 618 + ], + "spans": [ + { + "bbox": [ + 47, + 605, + 128, + 618 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 46, + 625, + 287, + 686 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 625, + 287, + 686 + ], + "spans": [ + { + "bbox": [ + 46, + 625, + 287, + 686 + ], + "type": "text", + "content": "Owing to well-designed large-scale datasets, video-text pre-training [15, 17, 19] has achieved superior performance in various downstream tasks, such as video-text retrieval [4, 10, 36], video question answering [27, 34, 42], and video captioning [1, 22, 30]. However, recent large-scale video-" + } + ] + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 309, + 246, + 386, + 328 + ], + "blocks": [ + { + "bbox": [ + 309, + 246, + 386, + 328 + ], + "lines": [ + { + "bbox": [ + 309, + 246, + 386, + 328 + ], + "spans": [ + { + "bbox": [ + 309, + 246, + 386, + 328 + ], + "type": "image", + "image_path": "9b7a8bdb48e124ef90a9d5b532991186db25efb56d5530433476257688e18c7f.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 338, + 546, + 371 + ], + "lines": [ + { + "bbox": [ + 305, + 338, + 546, + 371 + ], + "spans": [ + { + "bbox": [ + 305, + 338, + 546, + 371 + ], + "type": "text", + "content": "Figure 1. Here presents the motivations of this paper, based on which we highly summarize our contributions with three verbs: \"Build\", \"Filter\", and \"Pre-train\"." + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 391, + 249, + 472, + 327 + ], + "blocks": [ + { + "bbox": [ + 391, + 249, + 472, + 327 + ], + "lines": [ + { + "bbox": [ + 391, + 249, + 472, + 327 + ], + "spans": [ + { + "bbox": [ + 391, + 249, + 472, + 327 + ], + "type": "image", + "image_path": "a4d4cffe965fc815b477273e86bb78783bbe166255e5abaa4f22807664ae1cd0.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 475, + 249, + 551, + 328 + ], + "blocks": [ + { + "bbox": [ + 475, + 249, + 551, + 328 + ], + "lines": [ + { + "bbox": [ + 475, + 249, + 551, + 328 + ], + "spans": [ + { + "bbox": [ + 475, + 249, + 551, + 328 + ], + "type": "image", + "image_path": "b46894051c7f630db7c77289f2833c7996d7b3725c9d2e194d6e2cb1d0955b60.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 384, + 546, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 384, + 546, + 468 + ], + "spans": [ + { + "bbox": [ + 304, + 384, + 546, + 468 + ], + "type": "text", + "content": "text datasets are mostly English-only (e.g., Howto100M [25] and WebVid-2.5M [4]). Though some methods [14,26, 45] turn to study the Chinese video-text pre-training, they pre-train their models on private datasets whose videos and text are unavailable. Therefore, the research towards Chinese video-text pre-training is still in its infancy due to the lack of large-scale public datasets." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 471, + 547, + 614 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 471, + 547, + 614 + ], + "spans": [ + { + "bbox": [ + 304, + 471, + 547, + 614 + ], + "type": "text", + "content": "Towards this problem, directly translating English text into Chinese is a simple solution. However, it may result in unacceptable performance degradation for two reasons: 1) Translation errors are inevitable. Moreover, since most of the large-scale video-text datasets employ the Automatic Speech Recognition (ASR) system to generate text, the language translator would amplify the error from the incomplete and noisy ASR text. And 2) there remains an intrinsic linguistic gap between English and Chinese. Many widely-used English idioms and slang can hardly find their Chinese counterparts, leading some translated text incomprehensible and even contrary to the original meaning." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 616, + 547, + 675 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 616, + 547, + 675 + ], + "spans": [ + { + "bbox": [ + 304, + 616, + 547, + 675 + ], + "type": "text", + "content": "In this paper, we aim to release and benchmark a large-scale public Chinese video-text dataset to facilitate future researchers and the community. As illustrated in Figure 1, three verbs could highly summarize our contributions, i.e., \"Build\", \"Filter\", and \"Pre-train\"." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 304, + 677, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 546, + 713 + ], + "type": "text", + "content": "To build a large-scale Chinese video-text dataset, we collect over 4.5M videos from Chinese websites. All videos are associated with user-uploaded titles and ASR text." + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 693, + 127, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 693, + 127, + 703 + ], + "spans": [ + { + "bbox": [ + 58, + 693, + 127, + 703 + ], + "type": "text", + "content": "*Equal contribution." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 59, + 703, + 135, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 703, + 135, + 712 + ], + "spans": [ + { + "bbox": [ + 59, + 703, + 135, + 712 + ], + "type": "text", + "content": "† Corresponding author." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14815" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 263 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 263 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 263 + ], + "type": "text", + "content": "We filter out the weakly-paired data by a novel method to improve the data quality. As some work [25,26] pointed out, the pre-training performance would suffer from the noisy ASR text that fails to accurately describe the video content. Unfortunately, the problem is raised with few practical solutions. Therefore, we employ a well-trained image-text model to evaluate the video-text consistency for three reasons: 1) The text information in existing image-text datasets [11] are usually manually-written titles or captions, whose consistency is guaranteed. 2) Some video-text pre-training architectures [23,35] are based upon image-text ones. And 3) it is cheap and efficient to \"hire\" a well-trained model to check millions of videos. In this way, we filter out about 1M weakly-paired videos based on the balance between the pre-training performance and efficiency, deriving the proposed CNVid-3.5M dataset." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 264, + 289, + 467 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 264, + 289, + 467 + ], + "spans": [ + { + "bbox": [ + 46, + 264, + 289, + 467 + ], + "type": "text", + "content": "We pre-train various models to benchmark our CNVid-3.5M dataset. Current video-text pre-training methods could be roughly divided into two categories: 1) feature-level pre-training methods [24, 33, 40] that employ offline video and textual feature extractors, and 2) pixel-level ones [4, 15, 36] that learn cross-modal representations end-to-end from raw videos and text. Since there remain domain gaps between pre-training datasets and frozen feature extractors, pixel-level pre-training methods usually achieve better performance and have been widely employed in recent years. However, existing Chinese video-text pre-training methods [14, 26, 45] are all feature-level ones pre-trained on private datasets, limiting their contributions on the development of Chinese video-text pre-training techniques. Hence, we adopt three mainstream pixel-level pre-training frameworks, which are the first pixel-level benchmarks for Chinese video-text pre-training." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 468, + 289, + 659 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 468, + 289, + 659 + ], + "spans": [ + { + "bbox": [ + 46, + 468, + 289, + 659 + ], + "type": "text", + "content": "Moreover, we propose the novel Hard Sample Curriculum Learning strategy to promote the pre-training performance. Since contrastive learning is a significant component in video-text pre-training, some methods [16, 18, 43] employ the hard sample mining [12, 29] strategy to promote the cross-modal alignment. However, hard sample mining would bring side effects to pre-training when the model is far from convergence. Suppose that a model is incapable of discriminating the ground-truth video-text pairs, recklessly introducing hard negatives would lead to the sub-optimal performance. Inspired by the curriculum learning [32, 37] strategy that \"starts small\" and gradually \"learns hard\", we combine these two strategies and propose the novel Hard Sample Curriculum Learning (HSCL). By gradually and smoothly emphasizing those hard samples, HSCL could effectively improve the pre-training performance." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 59, + 660, + 253, + 671 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 660, + 253, + 671 + ], + "spans": [ + { + "bbox": [ + 59, + 660, + 253, + 671 + ], + "type": "text", + "content": "Our contributions are summarized in four folds:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 58, + 677, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 677, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 58, + 677, + 287, + 713 + ], + "type": "text", + "content": "- To fill in the blank of large-scale public Chinese video-text datasets, we collect over 4.5M videos associated with titles and ASR text from the websites." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 317, + 72, + 547, + 216 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 317, + 72, + 545, + 107 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 72, + 545, + 107 + ], + "spans": [ + { + "bbox": [ + 317, + 72, + 545, + 107 + ], + "type": "text", + "content": "- To improve the data quality, we propose a novel method to filter out 1M weakly-paired videos, resulting in the CNVid-3.5M dataset." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 317, + 108, + 545, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 108, + 545, + 144 + ], + "spans": [ + { + "bbox": [ + 317, + 108, + 545, + 144 + ], + "type": "text", + "content": "- To promote the pre-training performance, we propose the novel Hard Sample Curriculum Learning strategy for better cross-modal contrastive learning." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 317, + 144, + 547, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 144, + 547, + 216 + ], + "spans": [ + { + "bbox": [ + 317, + 144, + 547, + 216 + ], + "type": "text", + "content": "- To the best of our knowledge, the constructed CNVid-3.5M is the largest public Chinese video-text dataset. Moreover, we provide the first Chinese pixel-level benchmarks based on CNVid-3.5M. The dataset, codebase, and benchmarks are available at https://github.com/CNVid/CNVid-3.5M." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 235, + 392, + 248 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 235, + 392, + 248 + ], + "spans": [ + { + "bbox": [ + 306, + 235, + 392, + 248 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 255, + 545, + 387 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 255, + 545, + 387 + ], + "spans": [ + { + "bbox": [ + 304, + 255, + 545, + 387 + ], + "type": "text", + "content": "Video-Text Pre-training and Fine-Tuning. Video-text pre-training has attracted increasing interest in recent years, which could be roughly divided into two categories, feature-level pre-training methods [24, 33, 40] and pixel-level ones [15, 17, 35]. The former approaches employ offline visual and textual features extracted from frozen models, while the latter ones learn cross-modal representations from raw videos and text in an end-to-end manner. Since there remain domain gaps between pre-training datasets and frozen feature extractors, pixel-level pre-training methods usually achieve better performance than the feature-level ones." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 388, + 545, + 520 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 388, + 545, + 520 + ], + "spans": [ + { + "bbox": [ + 304, + 388, + 545, + 520 + ], + "type": "text", + "content": "There are three mainstream architectures in pixel-level video-text pre-training methods, i.e., three-encoders-fusion (3-E-F) [15, 17, 30], two-encoders-fusion (2-E-F) [9], and twin-towers-crossed (T-T-C) [8, 23, 36]. 3-E-F methods contain three separate encoders to model visual, textual, and cross-modal features; 2-E-F methods simplify the 3-E-F architecture by employing a shared encoder to embed textual and cross-modal features; while T-T-C methods are mainly based on the widely-adopted CLIP [28] architecture, which includes a couple of text and visual encoders to learn the cross-modal alignment by contrastive learning." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 521, + 545, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 521, + 545, + 581 + ], + "spans": [ + { + "bbox": [ + 304, + 521, + 545, + 581 + ], + "type": "text", + "content": "Though the pixel-level paradigm has been widely adopted recently, existing Chinese video-text models [14,26,45] are all feature-level ones pre-trained on private datasets. In this work, we provide the first pixel-level benchmarks pre-trained on our CNVid-3.5M dataset." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 581, + 545, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 581, + 545, + 688 + ], + "spans": [ + { + "bbox": [ + 304, + 581, + 545, + 688 + ], + "type": "text", + "content": "Video-Text Dataset. Well-designed large-scale datasets are a prerequisite for the successful applications of the \"Pre-training & Fine-tuning\" paradigm. There exist a few large-scale public English datasets (e.g., Howto100M [25], WebVid-2.5M [4]) for video-text pre-training. However, existing large-scale Chinese video datasets (ALIVOL [14], Kwai-SVC [26], CREATE [45]) are all private, whose videos and text are unavailable to the public. Therefore, we aimed to release a public one to facilitate future researchers." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 689, + 545, + 713 + ], + "type": "text", + "content": "Hard Sample Mining in Contrastive Learning. Hard Sample Mining [12, 29] is an effective strategy to promote" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "14816" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 47, + 69, + 561, + 220 + ], + "blocks": [ + { + "bbox": [ + 47, + 69, + 561, + 220 + ], + "lines": [ + { + "bbox": [ + 47, + 69, + 561, + 220 + ], + "spans": [ + { + "bbox": [ + 47, + 69, + 561, + 220 + ], + "type": "table", + "html": "
Dataset Name#(Videos)#(Text)Duration (hours)Video SourceText TypeAvailability
Off-FeatRaw-VidState
Datasets for Downstream Chinese Video-text Tasks
VATEX [38]41,269825,380115kinetics-600Caption✓✓All-public
BFVD [44]43,16643,166140E-CommerceTitle✓-Partly-public
FFVD [44]32,76332,763252E-CommerceTitle✓-Partly-public
CREATE210K [45]216,303268,5931,800Open WebsitesCaption--NOT-public
Kwai-SVC [26]222,077143,5693,500Open WebsitesTitle & ASR✓-Partly-public
Datasets for Chinese Video-text Pre-training
ALIVOL-10M [14]10,300,00011,000,00098,801E-CommerceTitle--NOT-public
Kwai-SVC-11M [26]11,075,0843,931,879177,200Open WebsitesTitle & ASR--NOT-public
CREATE-10M [45]10,000,00010,000,00083,000Open WebsitesTitle--NOT-public
CNVid-3.5M (ours)3,508,1203,508,12035,414Open WebsitesTitle & ASR✓✓All-public
", + "image_path": "b96e5adf0d4bb57206450bfdebed46c78382f64c9a3ec9b6c64b09fb5c1f4f78.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 229, + 545, + 251 + ], + "lines": [ + { + "bbox": [ + 46, + 229, + 545, + 251 + ], + "spans": [ + { + "bbox": [ + 46, + 229, + 545, + 251 + ], + "type": "text", + "content": "Table 1. Data analysis of CNVid-3.5M and other Chinese video-text datasets. \"Off-Feat\" and \"Raw-Vid\" represent whether offline features and raw videos are available. As far as we know, CNVid-3.5M is the largest public video-text dataset in Chinese." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 49, + 270, + 293, + 449 + ], + "blocks": [ + { + "bbox": [ + 49, + 270, + 293, + 449 + ], + "lines": [ + { + "bbox": [ + 49, + 270, + 293, + 449 + ], + "spans": [ + { + "bbox": [ + 49, + 270, + 293, + 449 + ], + "type": "image", + "image_path": "0807d4357590c810673f4124c45b001682cf285b1fa50e9e71996407f54d152e.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 458, + 287, + 492 + ], + "lines": [ + { + "bbox": [ + 46, + 458, + 287, + 492 + ], + "spans": [ + { + "bbox": [ + 46, + 458, + 287, + 492 + ], + "type": "text", + "content": "Figure 2. An example of video-text pairs in our CNVid-3.5M dataset. We present three types of text information: 1) Caption (manually written), 2) Title (user uploaded), and 3) ASR text." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 502, + 287, + 646 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 502, + 287, + 646 + ], + "spans": [ + { + "bbox": [ + 46, + 502, + 287, + 646 + ], + "type": "text", + "content": "the performance of contrastive learning. In video-text pretraining, TACo [43] proposes a cascade sampling method to construct hard negative examples for better cross-modal alignment. However, we conjectured that hard sample mining may hinder the video-text pre-training at the beginning, where the model is far from convergence and could hardly discriminate the ground-truth video-text pairs. Therefore, we introduced curriculum learning [32, 37] and proposed the Hard Sample Curriculum Learning (HSCL) strategy. HSCL would guide the pre-training procedure from \"starting small\" gradually to \"learning hard\", mitigating the side effect of conventional hard sample mining methods." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 657, + 194, + 670 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 657, + 194, + 670 + ], + "spans": [ + { + "bbox": [ + 47, + 657, + 194, + 670 + ], + "type": "text", + "content": "3. The CNVid-3.5M Dataset" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 677, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 287, + 715 + ], + "type": "text", + "content": "To fill in the blank of public Chinese video-text datasets, we release CNVid-3.5M, a large-scale cross-modal dataset containing over 3.5M Chinese video-text pairs." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 272, + 408, + 285 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 272, + 408, + 285 + ], + "spans": [ + { + "bbox": [ + 306, + 272, + 408, + 285 + ], + "type": "text", + "content": "3.1. Dataset Building" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 291, + 545, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 291, + 545, + 422 + ], + "spans": [ + { + "bbox": [ + 304, + 291, + 545, + 422 + ], + "type": "text", + "content": "We collect raw videos from Douyin*, a popular Chinese social website, where millions of active users upload various types of videos every day. To enrich the data diversity, we do not limit video categories like HowTo100M [25] (mainly instructional videos) and cover various topics (e.g., food, entertainment, and technology). I.e., we first form up a set of keywords and then leverage them to search for videos and corresponding captions. Moreover, we span the data created time at a range of 3 years and filter the abnormal ones whose videos could not be played. In this way, we collect over 4.5M raw videos in Chinese." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 423, + 545, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 423, + 545, + 449 + ], + "spans": [ + { + "bbox": [ + 305, + 423, + 545, + 449 + ], + "type": "text", + "content": "Besides, the supplementary material will present more details about fairness and privacy for safety and ethics." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 457, + 473, + 470 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 457, + 473, + 470 + ], + "spans": [ + { + "bbox": [ + 306, + 457, + 473, + 470 + ], + "type": "text", + "content": "3.2. Captions, Titles, and ASR Text" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 476, + 545, + 524 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 476, + 545, + 524 + ], + "spans": [ + { + "bbox": [ + 304, + 476, + 545, + 524 + ], + "type": "text", + "content": "As CREATE [45] pointed out, three types of text information are widely-employed in current Chinese video-text datasets. We show an example in Figure 2, based on which we summarize their characteristics and applications:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 525, + 545, + 620 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 525, + 545, + 620 + ], + "spans": [ + { + "bbox": [ + 304, + 525, + 545, + 620 + ], + "type": "text", + "content": "1) Captions are manually written by human annotators. They could accurately describe the video content in a formal manner. However, obtaining a large number of captions is extremely expensive. Therefore, captions are mainly employed in small-scale downstream video-text datasets like VATEX [38]. Moreover, captions are indeed global descriptions of the whole videos. Therefore, they could hardly describe the frame-level information of videos." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 621, + 545, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 621, + 545, + 693 + ], + "spans": [ + { + "bbox": [ + 304, + 621, + 545, + 693 + ], + "type": "text", + "content": "2) Titles are written and uploaded by users. Compared with captions, it is not unfeasible to obtain millions of user-uploaded titles. However, to increase the attractiveness and click rate, users may over-polish titles with some \"appealing but redundant\" words, which may decrease the video-text consistency. Similar to captions, titles are video-level" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 317, + 703, + 461, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 703, + 461, + 713 + ], + "spans": [ + { + "bbox": [ + 317, + 703, + 461, + 713 + ], + "type": "text", + "content": "*Official Website: https://www.douyin.com/" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14817" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 108 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 108 + ], + "type": "text", + "content": "descriptions rather than frame-level ones. Therefore, titles and captions are mainly employed in feature-level videotext pre-training methods rather than pixel-level ones." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 108, + 286, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 108, + 286, + 180 + ], + "spans": [ + { + "bbox": [ + 46, + 108, + 286, + 180 + ], + "type": "text", + "content": "3) ASR text is obtained by the ASR toolkit. Similar to titles, it is easy to obtain a large number of ASR text. However, some ASR text is incomplete, noisy, and even faulty, which may decrease the pre-training performance. Besides, ASR text is a frame-level description, which is widely employed in the pixel-level video-text pre-training." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 180, + 286, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 180, + 286, + 240 + ], + "spans": [ + { + "bbox": [ + 46, + 180, + 286, + 240 + ], + "type": "text", + "content": "Following previous methods [25, 26], we collect user-uploaded titles and obtain the ASR text of each video. Since we provide both video-level titles and frame-level ASR text, our dataset could support the video-text pre-training in both feature-level and pixel-level manner." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 248, + 208, + 262 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 248, + 208, + 262 + ], + "spans": [ + { + "bbox": [ + 47, + 248, + 208, + 262 + ], + "type": "text", + "content": "3.3. Weakly-paired Data Filtering" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 267, + 286, + 458 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 267, + 286, + 458 + ], + "spans": [ + { + "bbox": [ + 46, + 267, + 286, + 458 + ], + "type": "text", + "content": "As some work [25, 26] has pointed out, some ASR text may not match with associated videos (e.g., videos have background music), which would decrease the pre-training performance. However, the problem is only raised with few practical solutions. Therefore, we propose a novel weakly-paired data filtering strategy by leveraging a well-trained image-text model to automatically evaluate the video-text consistency for three reasons: 1) The text information in existing image-text datasets [11, 13] are manually-written titles or captions, whose quality and consistency are guaranteed. 2) Pixel-level pre-training methods [8, 15, 23] usually sample several frames from raw videos to serve as the visual input, whose frameworks are closely associated with image-text pre-trained models. 3) Compared with manually checking millions of videos, the proposed filtering strategy remarkably reduces the cost and improves the efficiency." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 459, + 286, + 531 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 459, + 286, + 531 + ], + "spans": [ + { + "bbox": [ + 46, + 459, + 286, + 531 + ], + "type": "text", + "content": "Specifically, we first pre-train an image-text model on Wukong-100M [11]. We then leverage it to calculate the similarity score between videos and ASR text, and filter out the last 1M videos by their ranks. In this way, we obtain the CNVid-3.5M dataset, based on which we provide three mainstream pixel-level pre-training benchmarks." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 539, + 222, + 552 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 539, + 222, + 552 + ], + "spans": [ + { + "bbox": [ + 47, + 539, + 222, + 552 + ], + "type": "text", + "content": "3.4. Comparison of Dataset Statistics" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 557, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 557, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 557, + 286, + 713 + ], + "type": "text", + "content": "Table 1 presents the statistics of CNVid-3.5M and other Chinese video-text datasets. To the best of our knowledge, CNVid-3.5M is the largest public dataset for Chinese video-text pre-training, with a total of 3.5M videos that last for 35.4K hours. Though ALIVOL-10M [14], KwaiSVC-11M [26], and CREATE-10M [45] contain more Chinese videos than ours, yet they do not release raw videos and text information to the public. Moreover, these three methods only provide conventional feature-level pretrained models, whose performance highly depends on frozen feature extractors. Towards this end, we provide the first pixel-level benchmarks based on our CNVid-3.5M dataset, which would pave the way for future research on" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 72, + 436, + 84 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 436, + 84 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 436, + 84 + ], + "type": "text", + "content": "Chinese video-text pre-training." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 84, + 545, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 84, + 545, + 133 + ], + "spans": [ + { + "bbox": [ + 304, + 84, + 545, + 133 + ], + "type": "text", + "content": "Moreover, we present additional statistics of our CNVid-3.5M dataset in the supplementary material, including the distributions of topics, keywords, video durations, and Part-of-Speech (POS) tags of the ASR text." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 144, + 493, + 159 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 144, + 493, + 159 + ], + "spans": [ + { + "bbox": [ + 305, + 144, + 493, + 159 + ], + "type": "text", + "content": "4. Pixel-level Pre-training Paradigm" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 165, + 545, + 297 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 165, + 545, + 297 + ], + "spans": [ + { + "bbox": [ + 304, + 165, + 545, + 297 + ], + "type": "text", + "content": "Since mainstream pixel-level video-text pre-training methods [17, 23, 35] extract cross-modal representations from raw data directly, they usually outperform the conventional feature-level ones [2, 24, 33]. Therefore, we benchmark our CNVid-3.5M dataset with three mainstream pixel-level architectures. As illustrated in Figure 3, we split the pixel-level pre-training process into three steps, namely: 1) data pre-processing to process raw videos and text into patches and tokens, 2) model architectures to generate cross-modal representations, and 3) proxy tasks to determine overall pre-training objectives." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 306, + 423, + 319 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 306, + 423, + 319 + ], + "spans": [ + { + "bbox": [ + 305, + 306, + 423, + 319 + ], + "type": "text", + "content": "4.1. Data Pre-processing" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 324, + 545, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 324, + 545, + 460 + ], + "spans": [ + { + "bbox": [ + 304, + 324, + 545, + 460 + ], + "type": "text", + "content": "Given a mini-batch (denoted as " + }, + { + "bbox": [ + 304, + 324, + 545, + 460 + ], + "type": "inline_equation", + "content": "\\mathcal{B}" + }, + { + "bbox": [ + 304, + 324, + 545, + 460 + ], + "type": "text", + "content": ") of videos " + }, + { + "bbox": [ + 304, + 324, + 545, + 460 + ], + "type": "inline_equation", + "content": "\\{V_i\\}_{i=1}^{|B|}" + }, + { + "bbox": [ + 304, + 324, + 545, + 460 + ], + "type": "text", + "content": " and their corresponding ASR text " + }, + { + "bbox": [ + 304, + 324, + 545, + 460 + ], + "type": "inline_equation", + "content": "\\{T_i\\}_{i=1}^{|B|}" + }, + { + "bbox": [ + 304, + 324, + 545, + 460 + ], + "type": "text", + "content": ", pixel-level pre-training methods would first sparsely (and randomly) sample " + }, + { + "bbox": [ + 304, + 324, + 545, + 460 + ], + "type": "inline_equation", + "content": "N_v" + }, + { + "bbox": [ + 304, + 324, + 545, + 460 + ], + "type": "text", + "content": " frames from each video. We then slice each frame into patches, obtaining visual patch embeddings " + }, + { + "bbox": [ + 304, + 324, + 545, + 460 + ], + "type": "inline_equation", + "content": "\\mathbf{P} \\in \\mathbb{R}^{N_v * N_p * d}" + }, + { + "bbox": [ + 304, + 324, + 545, + 460 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 324, + 545, + 460 + ], + "type": "inline_equation", + "content": "N_p" + }, + { + "bbox": [ + 304, + 324, + 545, + 460 + ], + "type": "text", + "content": " is the number of sliced patches, and " + }, + { + "bbox": [ + 304, + 324, + 545, + 460 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 304, + 324, + 545, + 460 + ], + "type": "text", + "content": " is the dimension of the embedding. Simultaneously, a BERT embedder is employed to process the text " + }, + { + "bbox": [ + 304, + 324, + 545, + 460 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 304, + 324, + 545, + 460 + ], + "type": "text", + "content": " into fixed-length word embeddings " + }, + { + "bbox": [ + 304, + 324, + 545, + 460 + ], + "type": "inline_equation", + "content": "\\mathbf{W} = [\\mathbf{w}^{cls}, \\mathbf{w}^1, \\mathbf{w}^2, \\dots, \\mathbf{w}^{N_t - 1}]" + }, + { + "bbox": [ + 304, + 324, + 545, + 460 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 324, + 545, + 460 + ], + "type": "inline_equation", + "content": "\\mathbf{W} \\in \\mathbb{R}^{N_t * d}" + }, + { + "bbox": [ + 304, + 324, + 545, + 460 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 324, + 545, + 460 + ], + "type": "inline_equation", + "content": "N_t" + }, + { + "bbox": [ + 304, + 324, + 545, + 460 + ], + "type": "text", + "content": " is the length of text tokens." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 468, + 424, + 480 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 468, + 424, + 480 + ], + "spans": [ + { + "bbox": [ + 305, + 468, + 424, + 480 + ], + "type": "text", + "content": "4.2. Model Architectures" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 487, + 545, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 487, + 545, + 548 + ], + "spans": [ + { + "bbox": [ + 304, + 487, + 545, + 548 + ], + "type": "text", + "content": "There are three mainstream pixel-level video-text pretraining architectures, namely three-encoders-fusion [15, 17, 30], two-encoders-fusion [9], and twin-towers-crossed [8, 23, 36]. We benchmark CNVid-3.5M with these three pixel-level pre-training architectures." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 548, + 545, + 656 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 548, + 545, + 656 + ], + "spans": [ + { + "bbox": [ + 304, + 548, + 545, + 656 + ], + "type": "text", + "content": "Three-encoders-fusion (3-E-F) pre-training methods contain a visual encoder " + }, + { + "bbox": [ + 304, + 548, + 545, + 656 + ], + "type": "inline_equation", + "content": "E_{vis}" + }, + { + "bbox": [ + 304, + 548, + 545, + 656 + ], + "type": "text", + "content": ", a text encoder " + }, + { + "bbox": [ + 304, + 548, + 545, + 656 + ], + "type": "inline_equation", + "content": "E_{txt}" + }, + { + "bbox": [ + 304, + 548, + 545, + 656 + ], + "type": "text", + "content": ", and a cross-modal encoder " + }, + { + "bbox": [ + 304, + 548, + 545, + 656 + ], + "type": "inline_equation", + "content": "E_{mul}" + }, + { + "bbox": [ + 304, + 548, + 545, + 656 + ], + "type": "text", + "content": ". After obtaining visual features " + }, + { + "bbox": [ + 304, + 548, + 545, + 656 + ], + "type": "inline_equation", + "content": "\\mathbf{V} = E_{vis}(\\mathbf{P})" + }, + { + "bbox": [ + 304, + 548, + 545, + 656 + ], + "type": "text", + "content": " and textual features " + }, + { + "bbox": [ + 304, + 548, + 545, + 656 + ], + "type": "inline_equation", + "content": "\\mathbf{T} = E_{txt}(\\mathbf{W})" + }, + { + "bbox": [ + 304, + 548, + 545, + 656 + ], + "type": "text", + "content": ", we concatenate these two-modal features into " + }, + { + "bbox": [ + 304, + 548, + 545, + 656 + ], + "type": "inline_equation", + "content": "[\\mathbf{T},\\mathbf{V}] = [\\mathbf{t}^{cls},\\mathbf{t}^1,\\mathbf{t}^2,\\dots ,\\mathbf{t}^{N_t - 1},\\mathbf{v}^{cls},\\mathbf{v}^1,\\mathbf{v}^2,\\dots ,\\mathbf{v}^{N_v}]" + }, + { + "bbox": [ + 304, + 548, + 545, + 656 + ], + "type": "text", + "content": ". Following CLIP4Clip [23], we set position embeddings to zero, and obtain cross-modal features " + }, + { + "bbox": [ + 304, + 548, + 545, + 656 + ], + "type": "inline_equation", + "content": "\\mathbf{M} = E_{mul}([\\mathbf{T},\\mathbf{V}])" + }, + { + "bbox": [ + 304, + 548, + 545, + 656 + ], + "type": "text", + "content": ". This forwarding process could be formulated as follows:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 354, + 666, + 545, + 680 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 354, + 666, + 545, + 680 + ], + "spans": [ + { + "bbox": [ + 354, + 666, + 545, + 680 + ], + "type": "interline_equation", + "content": "\\mathbf {M} = E _ {m u l} ([ E _ {t x t} (\\mathbf {W}), E _ {v i s} (\\mathbf {P}) ]), \\tag {1}", + "image_path": "8b797b0f41bc5142b7321080bf6a4f88d76f2a25c1f527aa31d1c921b082830e.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 689, + 545, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 689, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{M} = [\\mathbf{m}^{cls},\\mathbf{m}^1,\\mathbf{m}^2,\\dots ,\\mathbf{m}^{N_t + N_v}]" + }, + { + "bbox": [ + 304, + 689, + 545, + 713 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 689, + 545, + 713 + ], + "type": "inline_equation", + "content": "[\\cdot ,\\cdot ]" + }, + { + "bbox": [ + 304, + 689, + 545, + 713 + ], + "type": "text", + "content": " denotes the concatenation operation." + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14818" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 48, + 70, + 547, + 196 + ], + "blocks": [ + { + "bbox": [ + 48, + 70, + 547, + 196 + ], + "lines": [ + { + "bbox": [ + 48, + 70, + 547, + 196 + ], + "spans": [ + { + "bbox": [ + 48, + 70, + 547, + 196 + ], + "type": "table", + "html": "
1. Data Pre-processing2. Three Types of Model Architecture3. Proxy Tasks
Text Embedding\nC | | | | BERT Embedder\n| | | | | [CLS] a cat plays piano\n[CLS] 猫弹钢琴\nASR Caption\n| | | | | Input VideoThree-Encoders-Fusion\nCross-Modal Feature\n| | | | | Cross-Modal Encoder\n| | | | | Emut | | | | Text FeatureEbert\nShared Encoder\n| | | | | C | | | | Video Feature\n| | | | | Shared Encoder\n| | | | | C | | | | Video Encoder\n| | | | | C | | | | Cross-Modal EncodingTwo-Encoders-Fusion\nCross-Modal Feature\n| | | | | C | | | | Video Feature\n| | | | | C | | | | Video Encoder\n| | | | | C | | | | EvisTwin-Towers-Crossed\nContrastive Learning\n| | | | | C | | | | Video Feature\n| | | | | C | | | | Video Encoder\n| | | | | C | | | | EvisMasked Language Modeling (MLM)\nFC\n[LMSK]\n[C | | | | F | | | | F | | | | F | | | | F | | | | F | | | | F | | | | F | | | | F | | | | F | | | | F | | | | F | | | | F | | | | F | | | | F | | | | F | | | | F | | | | F | | | | F | | | | F | | | | F | | | |
", + "image_path": "b49f4b4679379e1001da74ede97fe47584a7138ef4d59dddec2cdeff492ba927.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 201, + 547, + 225 + ], + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 201, + 547, + 225 + ], + "spans": [ + { + "bbox": [ + 46, + 201, + 547, + 225 + ], + "type": "text", + "content": "Figure 3. The pipeline of mainstream pixel-level video-text pre-training methods, which contains three key steps: 1) Data pre-processing, 2) model architectures, and 3) proxy tasks. Note that we employ three mainstream architectures to enrich pixel-level pre-training benchmarks." + } + ] + } + ], + "index": 1, + "type": "text" + }, + { + "bbox": [ + 46, + 234, + 290, + 306 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 234, + 290, + 306 + ], + "spans": [ + { + "bbox": [ + 46, + 234, + 290, + 306 + ], + "type": "text", + "content": "Two-encoders-fusion (2-E-F) pre-training methods simplify the 3-E-F architecture by employing a shared BERT-type encoder " + }, + { + "bbox": [ + 46, + 234, + 290, + 306 + ], + "type": "inline_equation", + "content": "E_{bert}" + }, + { + "bbox": [ + 46, + 234, + 290, + 306 + ], + "type": "text", + "content": " to embed textual and cross-modal features simultaneously. The pipeline of obtaining textual features " + }, + { + "bbox": [ + 46, + 234, + 290, + 306 + ], + "type": "inline_equation", + "content": "\\mathbf{T}" + }, + { + "bbox": [ + 46, + 234, + 290, + 306 + ], + "type": "text", + "content": " and cross-modal features " + }, + { + "bbox": [ + 46, + 234, + 290, + 306 + ], + "type": "inline_equation", + "content": "\\mathbf{M}" + }, + { + "bbox": [ + 46, + 234, + 290, + 306 + ], + "type": "text", + "content": " could be formulated as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 104, + 312, + 287, + 346 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 104, + 312, + 287, + 346 + ], + "spans": [ + { + "bbox": [ + 104, + 312, + 287, + 346 + ], + "type": "interline_equation", + "content": "\\left\\{ \\begin{array}{l} \\mathbf {T} = E _ {b e r t} (\\mathbf {W}), \\\\ \\mathbf {M} = E _ {b e r t} ([ \\mathbf {W}, E _ {v i s} (\\mathbf {P}) ]). \\end{array} \\right. \\tag {2}", + "image_path": "0fa6bcf95b5d6dd67db314dd36bec71b38fb157bcc59b58031ebe4a007b9310c.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 353, + 289, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 353, + 289, + 437 + ], + "spans": [ + { + "bbox": [ + 47, + 353, + 289, + 437 + ], + "type": "text", + "content": "Twin-towers-crossed " + }, + { + "bbox": [ + 47, + 353, + 289, + 437 + ], + "type": "inline_equation", + "content": "(T-T-C)" + }, + { + "bbox": [ + 47, + 353, + 289, + 437 + ], + "type": "text", + "content": " pre-training methods mainly follow the conventional CLIP [28] architecture. After obtaining visual and textual features, " + }, + { + "bbox": [ + 47, + 353, + 289, + 437 + ], + "type": "inline_equation", + "content": "T-T-C" + }, + { + "bbox": [ + 47, + 353, + 289, + 437 + ], + "type": "text", + "content": " methods employ contrastive learning to optimize the twin encoders " + }, + { + "bbox": [ + 47, + 353, + 289, + 437 + ], + "type": "inline_equation", + "content": "E_{vis}" + }, + { + "bbox": [ + 47, + 353, + 289, + 437 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 353, + 289, + 437 + ], + "type": "inline_equation", + "content": "E_{txt}" + }, + { + "bbox": [ + 47, + 353, + 289, + 437 + ], + "type": "text", + "content": " for better cross-modal alignment. Note that " + }, + { + "bbox": [ + 47, + 353, + 289, + 437 + ], + "type": "inline_equation", + "content": "T-T-C" + }, + { + "bbox": [ + 47, + 353, + 289, + 437 + ], + "type": "text", + "content": " methods do not generate cross-modal video representations like " + }, + { + "bbox": [ + 47, + 353, + 289, + 437 + ], + "type": "inline_equation", + "content": "3-E-F" + }, + { + "bbox": [ + 47, + 353, + 289, + 437 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 353, + 289, + 437 + ], + "type": "inline_equation", + "content": "2-E-F" + }, + { + "bbox": [ + 47, + 353, + 289, + 437 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 445, + 128, + 457 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 445, + 128, + 457 + ], + "spans": [ + { + "bbox": [ + 47, + 445, + 128, + 457 + ], + "type": "text", + "content": "4.3. Proxy Tasks" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 463, + 287, + 523 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 463, + 287, + 523 + ], + "spans": [ + { + "bbox": [ + 46, + 463, + 287, + 523 + ], + "type": "text", + "content": "Proxy tasks are crucial for video-text pre-training as they directly determine the final optimization objectives. Following the conventional protocol [4, 15, 43], we adopt two widely-employed proxy tasks, i.e., Masked Language Modeling (MLM) and Video-Text Matching (VTM)." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 523, + 288, + 583 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 523, + 288, + 583 + ], + "spans": [ + { + "bbox": [ + 46, + 523, + 288, + 583 + ], + "type": "text", + "content": "MLM first masks out a certain percentage of words in a given sentence, and then forces the model to restore these clozes according to visual and textual cues. MLM is calculated twice for textual features " + }, + { + "bbox": [ + 46, + 523, + 288, + 583 + ], + "type": "inline_equation", + "content": "\\mathbf{T}(\\mathcal{L}_1)" + }, + { + "bbox": [ + 46, + 523, + 288, + 583 + ], + "type": "text", + "content": " and cross-modal features " + }, + { + "bbox": [ + 46, + 523, + 288, + 583 + ], + "type": "inline_equation", + "content": "\\mathbf{M}(\\mathcal{L}_2)" + }, + { + "bbox": [ + 46, + 523, + 288, + 583 + ], + "type": "text", + "content": " as follows:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 98, + 590, + 287, + 620 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 98, + 590, + 287, + 620 + ], + "spans": [ + { + "bbox": [ + 98, + 590, + 287, + 620 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {1} = \\frac {1}{| \\mathcal {Q} |} \\sum_ {q \\in \\mathcal {Q}} \\mathcal {L} _ {C E} \\left(y ^ {q}, \\Theta_ {1} \\left(\\mathbf {t} ^ {q}\\right)\\right), \\tag {3}", + "image_path": "50dadd559fdc49825ac39afb5c5c53e80e44cf55987fb787003edddc2689ca37.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 96, + 630, + 287, + 660 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 96, + 630, + 287, + 660 + ], + "spans": [ + { + "bbox": [ + 96, + 630, + 287, + 660 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {2} = \\frac {1}{| \\mathcal {Q} |} \\sum_ {q \\in \\mathcal {Q}} \\mathcal {L} _ {C E} \\left(y ^ {q}, \\Theta_ {2} \\left(\\mathbf {m} ^ {q}\\right)\\right), \\tag {4}", + "image_path": "ad5b6e1749636629041ed8e0039c5191705b2e0418cbc053a3c193d80ae2127a.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 665, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 289, + 715 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 665, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{Q}" + }, + { + "bbox": [ + 46, + 665, + 289, + 715 + ], + "type": "text", + "content": " records the location of masked tokens, " + }, + { + "bbox": [ + 46, + 665, + 289, + 715 + ], + "type": "inline_equation", + "content": "|\\cdot|" + }, + { + "bbox": [ + 46, + 665, + 289, + 715 + ], + "type": "text", + "content": " denotes the length of a given set, " + }, + { + "bbox": [ + 46, + 665, + 289, + 715 + ], + "type": "inline_equation", + "content": "y^{q}" + }, + { + "bbox": [ + 46, + 665, + 289, + 715 + ], + "type": "text", + "content": " denotes the ground-truth token label, " + }, + { + "bbox": [ + 46, + 665, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\Theta" + }, + { + "bbox": [ + 46, + 665, + 289, + 715 + ], + "type": "text", + "content": " is a Multi-Layer Perception (MLP), and " + }, + { + "bbox": [ + 46, + 665, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{CE}" + }, + { + "bbox": [ + 46, + 665, + 289, + 715 + ], + "type": "text", + "content": " is the regular Cross-Entropy cost function." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 234, + 546, + 271 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 234, + 546, + 271 + ], + "spans": [ + { + "bbox": [ + 305, + 234, + 546, + 271 + ], + "type": "text", + "content": "VTM aims to promote the cross-modal alignment between videos and text, which is calculated in a parameter-free " + }, + { + "bbox": [ + 305, + 234, + 546, + 271 + ], + "type": "inline_equation", + "content": "(\\mathcal{L}_3)" + }, + { + "bbox": [ + 305, + 234, + 546, + 271 + ], + "type": "text", + "content": " and parameter-employed " + }, + { + "bbox": [ + 305, + 234, + 546, + 271 + ], + "type": "inline_equation", + "content": "(\\mathcal{L}_4)" + }, + { + "bbox": [ + 305, + 234, + 546, + 271 + ], + "type": "text", + "content": " way:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 307, + 280, + 545, + 325 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 280, + 545, + 325 + ], + "spans": [ + { + "bbox": [ + 307, + 280, + 545, + 325 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {3} = - \\sum_ {i = 1} ^ {| \\mathcal {B} |} \\log \\frac {\\sum_ {k = 1} ^ {N _ {v}} \\exp \\left\\langle \\mathbf {v} _ {i} ^ {k} , \\mathbf {t} _ {i} ^ {c l s} \\right\\rangle}{\\sum_ {k = 1} ^ {N _ {v}} \\left(\\exp \\left\\langle \\mathbf {v} _ {i} ^ {k} , \\mathbf {t} _ {i} ^ {c l s} \\right\\rangle + \\sum_ {j \\neq i} \\exp \\left\\langle \\mathbf {v} _ {j} ^ {k} , \\mathbf {t} _ {i} ^ {c l s} \\right\\rangle\\right)}, \\tag {5}", + "image_path": "e43e5b1643d2cc1d8d9a0df074b67995c370682e726d01e99d47bb943c1c531b.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 322, + 344, + 545, + 379 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 322, + 344, + 545, + 379 + ], + "spans": [ + { + "bbox": [ + 322, + 344, + 545, + 379 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {4} = - \\sum_ {i = 1} ^ {| \\mathcal {B} |} \\log \\frac {\\exp^ {\\Theta_ {4} \\left(\\mathbf {f} _ {i , i}\\right)}}{\\exp^ {\\Theta_ {4} \\left(\\mathbf {f} _ {i , i}\\right)} + \\sum_ {j \\neq i} \\exp^ {\\Theta_ {4} \\left(\\mathbf {f} _ {j , i}\\right)}}, \\tag {6}", + "image_path": "6ab545c4426e2a3d0d078cbabdfe0369ef78d6cf3f2de5ae2430a8f30e6884ba.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 388, + 545, + 424 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 388, + 545, + 424 + ], + "spans": [ + { + "bbox": [ + 305, + 388, + 545, + 424 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 388, + 545, + 424 + ], + "type": "inline_equation", + "content": "|\\mathcal{B}|" + }, + { + "bbox": [ + 305, + 388, + 545, + 424 + ], + "type": "text", + "content": " is the length of a mini-batch, " + }, + { + "bbox": [ + 305, + 388, + 545, + 424 + ], + "type": "inline_equation", + "content": "\\langle \\cdot ,\\cdot \\rangle" + }, + { + "bbox": [ + 305, + 388, + 545, + 424 + ], + "type": "text", + "content": " denotes the matrix multiplication operation, and " + }, + { + "bbox": [ + 305, + 388, + 545, + 424 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_{j,i}" + }, + { + "bbox": [ + 305, + 388, + 545, + 424 + ], + "type": "text", + "content": " is global-pooling cross-modal features of the video-text pair " + }, + { + "bbox": [ + 305, + 388, + 545, + 424 + ], + "type": "inline_equation", + "content": "(V_{j},S_{i})" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 424, + 545, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 424, + 545, + 449 + ], + "spans": [ + { + "bbox": [ + 306, + 424, + 545, + 449 + ], + "type": "text", + "content": "For " + }, + { + "bbox": [ + 306, + 424, + 545, + 449 + ], + "type": "inline_equation", + "content": "3 - E - F" + }, + { + "bbox": [ + 306, + 424, + 545, + 449 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 306, + 424, + 545, + 449 + ], + "type": "inline_equation", + "content": "2 - E - F" + }, + { + "bbox": [ + 306, + 424, + 545, + 449 + ], + "type": "text", + "content": " models, the objective functions " + }, + { + "bbox": [ + 306, + 424, + 545, + 449 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{3EF} / \\mathcal{L}_{2EF}" + }, + { + "bbox": [ + 306, + 424, + 545, + 449 + ], + "type": "text", + "content": " are calculated as follows:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 347, + 458, + 545, + 472 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 347, + 458, + 545, + 472 + ], + "spans": [ + { + "bbox": [ + 347, + 458, + 545, + 472 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {3 E F} \\left(\\mathcal {L} _ {2 E F}\\right) = \\mathcal {L} _ {1} + \\mathcal {L} _ {2} + \\mathcal {L} _ {3} + \\mathcal {L} _ {4}. \\tag {7}", + "image_path": "91fd65709e8ebc253822cd679f23598fcbc790ef9c50a4ba0765cff65741aa8f.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 305, + 481, + 545, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 481, + 545, + 518 + ], + "spans": [ + { + "bbox": [ + 305, + 481, + 545, + 518 + ], + "type": "text", + "content": "While for " + }, + { + "bbox": [ + 305, + 481, + 545, + 518 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 305, + 481, + 545, + 518 + ], + "type": "text", + "content": "-" + }, + { + "bbox": [ + 305, + 481, + 545, + 518 + ], + "type": "inline_equation", + "content": "T-C" + }, + { + "bbox": [ + 305, + 481, + 545, + 518 + ], + "type": "text", + "content": " models, since they do not contain a cross-modal encoder to generate video representations, their objective function " + }, + { + "bbox": [ + 305, + 481, + 545, + 518 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{TTC}" + }, + { + "bbox": [ + 305, + 481, + 545, + 518 + ], + "type": "text", + "content": " is calculated as follows:" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 386, + 529, + 545, + 541 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 386, + 529, + 545, + 541 + ], + "spans": [ + { + "bbox": [ + 386, + 529, + 545, + 541 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {T T C} = \\mathcal {L} _ {1} + \\mathcal {L} _ {3}. \\tag {8}", + "image_path": "f1742636f772f7b84db29f53ad4927398ae0c11efb2d16b016be49ff23a51929.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 306, + 551, + 384, + 563 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 551, + 384, + 563 + ], + "spans": [ + { + "bbox": [ + 306, + 551, + 384, + 563 + ], + "type": "text", + "content": "4.4. Fine-tuning" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 570, + 545, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 545, + 676 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 545, + 676 + ], + "type": "text", + "content": "In order to thoroughly evaluate the pre-training performance, we fine-tune our models on the Text-to-Video Retrieval (TVR) task. TVR is a widely-employed downstream video-text task, aiming to retrieve the most relevant videos according to text queries. Since there has only one Chinese TVR dataset (VATEX [38]), we translate two widely-adopted English datasets (MSRVTT [41] and DiDemo [3]) into Chinese to verify the superiority and robustness of the constructed dataset and proposed methods." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 305, + 677, + 546, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 677, + 546, + 714 + ], + "spans": [ + { + "bbox": [ + 305, + 677, + 546, + 714 + ], + "type": "text", + "content": "Specifically, we fine-tune our pre-trained models by reusing Video-Text Matching objectives (" + }, + { + "bbox": [ + 305, + 677, + 546, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_3" + }, + { + "bbox": [ + 305, + 677, + 546, + 714 + ], + "type": "text", + "content": " in Eq. 5 and " + }, + { + "bbox": [ + 305, + 677, + 546, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_4" + }, + { + "bbox": [ + 305, + 677, + 546, + 714 + ], + "type": "text", + "content": " in Eq. 6, we only employ " + }, + { + "bbox": [ + 305, + 677, + 546, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_3" + }, + { + "bbox": [ + 305, + 677, + 546, + 714 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 305, + 677, + 546, + 714 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 305, + 677, + 546, + 714 + ], + "type": "text", + "content": "-" + }, + { + "bbox": [ + 305, + 677, + 546, + 714 + ], + "type": "inline_equation", + "content": "T-C" + }, + { + "bbox": [ + 305, + 677, + 546, + 714 + ], + "type": "text", + "content": " models)." + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "14819" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 71, + 292, + 166 + ], + "blocks": [ + { + "bbox": [ + 50, + 71, + 292, + 166 + ], + "lines": [ + { + "bbox": [ + 50, + 71, + 292, + 166 + ], + "spans": [ + { + "bbox": [ + 50, + 71, + 292, + 166 + ], + "type": "image", + "image_path": "9706241771516cafa10cc722eb7557c1d8ff1513a795cdc863411f7b57d921cc.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 175, + 287, + 209 + ], + "lines": [ + { + "bbox": [ + 46, + 175, + 287, + 209 + ], + "spans": [ + { + "bbox": [ + 46, + 175, + 287, + 209 + ], + "type": "text", + "content": "Figure 4. The pipeline of Hard Sample Curriculum Learning (HSCL), which contains three components to promote the performance of video-text pre-training." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 217, + 245, + 232 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 217, + 245, + 232 + ], + "spans": [ + { + "bbox": [ + 47, + 217, + 245, + 232 + ], + "type": "text", + "content": "5. Hard Sample Curriculum Learning" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 237, + 287, + 417 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 237, + 287, + 417 + ], + "spans": [ + { + "bbox": [ + 46, + 237, + 287, + 417 + ], + "type": "text", + "content": "We propose Hard Sample Curriculum Learning (HSCL), a model-agnostic strategy to promote the video-text pretraining performance. HSCL is based on two motivations. 1) The negative samples involved in the calculation of parameter-employed VTM loss " + }, + { + "bbox": [ + 46, + 237, + 287, + 417 + ], + "type": "inline_equation", + "content": "(\\mathcal{L}_4" + }, + { + "bbox": [ + 46, + 237, + 287, + 417 + ], + "type": "text", + "content": " in Eq. 6) are not representative. Suppose that the batch size is set to be " + }, + { + "bbox": [ + 46, + 237, + 287, + 417 + ], + "type": "inline_equation", + "content": "N_{k}" + }, + { + "bbox": [ + 46, + 237, + 287, + 417 + ], + "type": "text", + "content": ", a model needs to process " + }, + { + "bbox": [ + 46, + 237, + 287, + 417 + ], + "type": "inline_equation", + "content": "N_{k} * N_{k}" + }, + { + "bbox": [ + 46, + 237, + 287, + 417 + ], + "type": "text", + "content": " video-text pairs for calculating " + }, + { + "bbox": [ + 46, + 237, + 287, + 417 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_4" + }, + { + "bbox": [ + 46, + 237, + 287, + 417 + ], + "type": "text", + "content": ", whose calculation cost is hardly affordable. Therefore, previous methods [9, 15, 22] tend to cut the number of negative samples to " + }, + { + "bbox": [ + 46, + 237, + 287, + 417 + ], + "type": "inline_equation", + "content": "N_{k}^{\\prime}" + }, + { + "bbox": [ + 46, + 237, + 287, + 417 + ], + "type": "text", + "content": " by random sampling. However, since " + }, + { + "bbox": [ + 46, + 237, + 287, + 417 + ], + "type": "inline_equation", + "content": "N_{k}^{\\prime}" + }, + { + "bbox": [ + 46, + 237, + 287, + 417 + ], + "type": "text", + "content": " is usually much smaller than " + }, + { + "bbox": [ + 46, + 237, + 287, + 417 + ], + "type": "inline_equation", + "content": "N_{k}" + }, + { + "bbox": [ + 46, + 237, + 287, + 417 + ], + "type": "text", + "content": ", it may result in the sub-optimal performance. 2) If a model is far from convergence, directly introducing hard negative samples would amplify the burden of learning, which would decrease the pre-training performance." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 418, + 287, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 418, + 287, + 502 + ], + "spans": [ + { + "bbox": [ + 46, + 418, + 287, + 502 + ], + "type": "text", + "content": "Therefore, inspired by hard sample mining and curriculum learning, we propose the HSCL strategy. By leveraging the similarity score calculated by the parameter-free VTM loss " + }, + { + "bbox": [ + 46, + 418, + 287, + 502 + ], + "type": "inline_equation", + "content": "(\\mathcal{L}_3" + }, + { + "bbox": [ + 46, + 418, + 287, + 502 + ], + "type": "text", + "content": " in Eq. 5), HSCL could gradually and smoothly introduce hard samples for better cross-modal alignment. As illustrated in Figure 4, HSCL contains the following three key components:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 502, + 287, + 573 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 502, + 287, + 573 + ], + "spans": [ + { + "bbox": [ + 46, + 502, + 287, + 573 + ], + "type": "text", + "content": "Positive Column Weighting (PCW) aims to decrease the learning coefficient " + }, + { + "bbox": [ + 46, + 502, + 287, + 573 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 46, + 502, + 287, + 573 + ], + "type": "text", + "content": " of well-learned positive samples, forcing the model to emphasize those hard positive ones. Specifically, we re-weigh half of samples within the batch according to their similarity scores " + }, + { + "bbox": [ + 46, + 502, + 287, + 573 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 46, + 502, + 287, + 573 + ], + "type": "text", + "content": " calculated by " + }, + { + "bbox": [ + 46, + 502, + 287, + 573 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_3" + }, + { + "bbox": [ + 46, + 502, + 287, + 573 + ], + "type": "text", + "content": ", which could be formulated as follows:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 110, + 580, + 287, + 613 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 110, + 580, + 287, + 613 + ], + "spans": [ + { + "bbox": [ + 110, + 580, + 287, + 613 + ], + "type": "interline_equation", + "content": "z _ {j, i} = \\frac {1}{N _ {v}} \\sum_ {k = 1} ^ {N _ {v}} \\exp \\left\\langle \\mathbf {v} _ {j} ^ {k}, \\mathbf {t} _ {i} ^ {c l s} \\right\\rangle , \\tag {9}", + "image_path": "fb91ff0503fcd3ce9ff19ae9ddec8c168aa766ecc480057345d78509062f1730.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 67, + 622, + 287, + 650 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 622, + 287, + 650 + ], + "spans": [ + { + "bbox": [ + 67, + 622, + 287, + 650 + ], + "type": "interline_equation", + "content": "\\phi_ {i} = \\operatorname {M i n} \\left(\\frac {\\operatorname {M e d} \\left(Z _ {d i a g}\\right) - \\operatorname {M i n} \\left(Z _ {d i a g}\\right)}{z _ {i , i} - \\operatorname {M i n} \\left(Z _ {d i a g}\\right) + e _ {0}}, 1. 0\\right), \\tag {10}", + "image_path": "568688a473bea81f0414bf6ae3f695698f3d3d5e299e36218bac5225540d7f06.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 653, + 287, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 653, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 47, + 653, + 287, + 689 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 653, + 287, + 689 + ], + "type": "inline_equation", + "content": "Z_{diag} = [z_{1,1}, z_{2,2}, \\dots, z_{i,i}, \\dots]" + }, + { + "bbox": [ + 47, + 653, + 287, + 689 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 47, + 653, + 287, + 689 + ], + "type": "inline_equation", + "content": "\\operatorname{Med}(\\cdot)" + }, + { + "bbox": [ + 47, + 653, + 287, + 689 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 653, + 287, + 689 + ], + "type": "inline_equation", + "content": "\\operatorname{Min}(\\cdot)" + }, + { + "bbox": [ + 47, + 653, + 287, + 689 + ], + "type": "text", + "content": " represent the median and minimum result of the list, " + }, + { + "bbox": [ + 47, + 653, + 287, + 689 + ], + "type": "inline_equation", + "content": "e_0" + }, + { + "bbox": [ + 47, + 653, + 287, + 689 + ], + "type": "text", + "content": " is a pre-defined coefficient to avoid a zero denominator." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "type": "text", + "content": "Negative Line Sampling (NLS) aims to sample " + }, + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "type": "inline_equation", + "content": "N_{k}^{\\prime}" + }, + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "type": "text", + "content": " hard negative examples for better contrastive learning. Specifi-" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 71, + 545, + 121 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 71, + 545, + 121 + ], + "spans": [ + { + "bbox": [ + 304, + 71, + 545, + 121 + ], + "type": "text", + "content": "cally, for each text " + }, + { + "bbox": [ + 304, + 71, + 545, + 121 + ], + "type": "inline_equation", + "content": "T_{j}" + }, + { + "bbox": [ + 304, + 71, + 545, + 121 + ], + "type": "text", + "content": " , we choose the Nearliest " + }, + { + "bbox": [ + 304, + 71, + 545, + 121 + ], + "type": "inline_equation", + "content": "N_{k}^{\\prime}" + }, + { + "bbox": [ + 304, + 71, + 545, + 121 + ], + "type": "text", + "content": " aligned negative videos according to their similarity scores in the list " + }, + { + "bbox": [ + 304, + 71, + 545, + 121 + ], + "type": "inline_equation", + "content": "Z_{j} = [z_{j,1},z_{j,2},\\dots ,z_{j,i},\\dots ]" + }, + { + "bbox": [ + 304, + 71, + 545, + 121 + ], + "type": "text", + "content": " . I.e., Pick the minimum " + }, + { + "bbox": [ + 304, + 71, + 545, + 121 + ], + "type": "inline_equation", + "content": "N_{k}^{\\prime}" + }, + { + "bbox": [ + 304, + 71, + 545, + 121 + ], + "type": "text", + "content": " samples in the list " + }, + { + "bbox": [ + 304, + 71, + 545, + 121 + ], + "type": "inline_equation", + "content": "Z_{j}^{\\prime} = \\{|z_{j,i} - z_{j,j}|\\}_{j\\neq i}" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 121, + 546, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 121, + 546, + 205 + ], + "spans": [ + { + "bbox": [ + 304, + 121, + 546, + 205 + ], + "type": "text", + "content": "Curriculum Learning (CL) aims to mitigate the side effect of hard sample mining when the model is under convergence. We set a gradually increasing coefficient " + }, + { + "bbox": [ + 304, + 121, + 546, + 205 + ], + "type": "inline_equation", + "content": "\\beta \\in (0,1)" + }, + { + "bbox": [ + 304, + 121, + 546, + 205 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 304, + 121, + 546, + 205 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 304, + 121, + 546, + 205 + ], + "type": "text", + "content": " is initialized with zero at the beginning of pre-training. It will gradually increase by a fixed number " + }, + { + "bbox": [ + 304, + 121, + 546, + 205 + ], + "type": "inline_equation", + "content": "\\beta_0" + }, + { + "bbox": [ + 304, + 121, + 546, + 205 + ], + "type": "text", + "content": " for each " + }, + { + "bbox": [ + 304, + 121, + 546, + 205 + ], + "type": "inline_equation", + "content": "N_d" + }, + { + "bbox": [ + 304, + 121, + 546, + 205 + ], + "type": "text", + "content": " iteration until reaching 1.0. For PCW, the actual coefficient " + }, + { + "bbox": [ + 304, + 121, + 546, + 205 + ], + "type": "inline_equation", + "content": "\\phi^*" + }, + { + "bbox": [ + 304, + 121, + 546, + 205 + ], + "type": "text", + "content": " is calculated as follows:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 372, + 215, + 545, + 228 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 372, + 215, + 545, + 228 + ], + "spans": [ + { + "bbox": [ + 372, + 215, + 545, + 228 + ], + "type": "interline_equation", + "content": "\\phi_ {i} ^ {*} = (1. 0 - \\beta) + \\beta * \\phi_ {i}. \\tag {11}", + "image_path": "c26c08ae9cbfb1221772e1f34e3e077de67502a529089a1d8c8f36a619a2bc87.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 237, + 545, + 298 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 237, + 545, + 298 + ], + "spans": [ + { + "bbox": [ + 304, + 237, + 545, + 298 + ], + "type": "text", + "content": "For NLS, we first generate a random number " + }, + { + "bbox": [ + 304, + 237, + 545, + 298 + ], + "type": "inline_equation", + "content": "rand \\in (0,1)" + }, + { + "bbox": [ + 304, + 237, + 545, + 298 + ], + "type": "text", + "content": ", and only perform the NLS step if " + }, + { + "bbox": [ + 304, + 237, + 545, + 298 + ], + "type": "inline_equation", + "content": "rand < \\beta" + }, + { + "bbox": [ + 304, + 237, + 545, + 298 + ], + "type": "text", + "content": ". In this way, the model would conduct conventional pre-training at the beginning, and smoothly change to learn hard negative samples to pursue better pre-training performance." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 308, + 476, + 321 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 308, + 476, + 321 + ], + "spans": [ + { + "bbox": [ + 305, + 308, + 476, + 321 + ], + "type": "text", + "content": "6. Experiments and Benchmarks" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 327, + 429, + 340 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 327, + 429, + 340 + ], + "spans": [ + { + "bbox": [ + 306, + 327, + 429, + 340 + ], + "type": "text", + "content": "6.1. Pre-training Datasets" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 346, + 545, + 405 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 346, + 545, + 405 + ], + "spans": [ + { + "bbox": [ + 304, + 346, + 545, + 405 + ], + "type": "text", + "content": "In Section 3.3, the Chinese image-text model that helps to filter out the weakly-paired data is pre-trained on Wukong-100M [11] (100M image-text pairs). The Chinese video-text models are pre-trained on our CNVid-3.5M dataset, which contains 3.5M video-text pairs." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 406, + 545, + 478 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 406, + 545, + 478 + ], + "spans": [ + { + "bbox": [ + 304, + 406, + 545, + 478 + ], + "type": "text", + "content": "To verify the generality of the proposed Hard Sample Curriculum Learning (HSCL) strategy, we also pre-trained our models on four large-scale English datasets, they are 1) COCO [20] (0.6M image-text pairs), 2) VG [13] (5.4M image-text pairs), 3) CC [31] (3.1M image-text pairs), and 4) WebVid-2.5M (2.5M video-text pairs)." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 306, + 485, + 425, + 498 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 485, + 425, + 498 + ], + "spans": [ + { + "bbox": [ + 306, + 485, + 425, + 498 + ], + "type": "text", + "content": "6.2. Fine-tuning Datasets" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 503, + 545, + 623 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 503, + 545, + 623 + ], + "spans": [ + { + "bbox": [ + 304, + 503, + 545, + 623 + ], + "type": "text", + "content": "We fine-tuned our pre-trained models on the Text-to-Video Retrieval (TVR) task, which includes three widely-adopted datasets as follows: 1) VATEX [38] contains 28K video clips, and each video is associated with 10 Chinese and 10 English diverse captions. 2) MSRVTT [41] contains 10K video clips associated with 200K English captions. Following [15, 22, 43], we used 7K videos for training and randomly selected 1K videos from the remaining ones for testing (7K-1K split). 3) DiDemo [3] contains 10K Flickr videos associated with 40K English captions." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 624, + 545, + 670 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 624, + 545, + 670 + ], + "spans": [ + { + "bbox": [ + 304, + 624, + 545, + 670 + ], + "type": "text", + "content": "Moreover, for a comprehensive performance evaluation of Chinese video-text pre-trained models, we employ Google " + }, + { + "bbox": [ + 304, + 624, + 545, + 670 + ], + "type": "inline_equation", + "content": "^{\\dagger}" + }, + { + "bbox": [ + 304, + 624, + 545, + 670 + ], + "type": "text", + "content": " to translate all captions in MSRVTT and DiDemo from English to Chinese." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 671, + 545, + 695 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 671, + 545, + 695 + ], + "spans": [ + { + "bbox": [ + 304, + 671, + 545, + 695 + ], + "type": "text", + "content": "Metrics. We employed Recall@K (R@K, K=1/5/10) and Median Rank (MdR) to measure the TVR performance." + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 315, + 702, + 471, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 702, + 471, + 713 + ], + "spans": [ + { + "bbox": [ + 315, + 702, + 471, + 713 + ], + "type": "text", + "content": "†Official Website: https://translate.google.com/" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14820" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 48, + 59, + 552, + 83 + ], + "blocks": [ + { + "bbox": [ + 48, + 59, + 552, + 83 + ], + "lines": [ + { + "bbox": [ + 48, + 59, + 552, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 59, + 552, + 83 + ], + "type": "table", + "html": "
No.Model TypeData VolumeMasking StrategyMatching StrategyVATEX (Chinese)MSRVTT (Translated)DiDemo (Translated)
R@1/5/10↑(MdR↓)R@1/5/10↑(MdR↓)R@1/5/10↑(MdR↓)
", + "image_path": "d48ae3dcf253a0419016bd4f4f0af66578ac2931aa42d38096bd3e31fb981a1c.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 49, + 97, + 552, + 122 + ], + "blocks": [ + { + "bbox": [ + 170, + 85, + 430, + 95 + ], + "lines": [ + { + "bbox": [ + 170, + 85, + 430, + 95 + ], + "spans": [ + { + "bbox": [ + 170, + 85, + 430, + 95 + ], + "type": "text", + "content": "For \"Building\": Ablation Study of the constructed CNVid-3.5M dataset" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 49, + 97, + 552, + 122 + ], + "lines": [ + { + "bbox": [ + 49, + 97, + 552, + 122 + ], + "spans": [ + { + "bbox": [ + 49, + 97, + 552, + 122 + ], + "type": "table", + "html": "
A12-E-FNo Pre-training36.4 / 75.4 / 85.2 (2)15.9 / 42.1 / 54.1 (8)8.0 / 24.8 / 39.2 (18)
A22-E-F3.5MMLMVTM39.9 / 77.2 / 87.0 (2)20.7 / 47.9 / 61.2 (6)13.3 / 34.6 / 45.9 (13)
", + "image_path": "6c36804f02c728b7dce0a29e4ba094b4ef9463280689c3cb4ad2532aa2955733.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 49, + 136, + 552, + 194 + ], + "blocks": [ + { + "bbox": [ + 182, + 124, + 418, + 134 + ], + "lines": [ + { + "bbox": [ + 182, + 124, + 418, + 134 + ], + "spans": [ + { + "bbox": [ + 182, + 124, + 418, + 134 + ], + "type": "text", + "content": "For \"Filtering\": Ablation Study of Weakly-Paired Data Filtering" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 49, + 136, + 552, + 194 + ], + "lines": [ + { + "bbox": [ + 49, + 136, + 552, + 194 + ], + "spans": [ + { + "bbox": [ + 49, + 136, + 552, + 194 + ], + "type": "table", + "html": "
B12-E-F4.5MMLMVTM38.8 / 76.4 / 86.4 (2)19.4 / 45.6 / 58.6 (6)10.6 / 32.7 / 45.0 (14)
B22-E-F4.0MMLMVTM39.5 / 77.0 / 86.5 (2)19.8 / 47.2 / 60.3 (6)12.4 / 32.8 / 45.6 (14)
B32-E-F3.5MMLMVTM39.9 / 77.2 / 87.0 (2)20.7 / 47.9 / 61.2 (6)13.3 / 34.6 / 45.9 (13)
B42-E-F3.0MMLMVTM40.5 / 76.8 / 86.5 (2)20.6 / 47.3 / 61.3 (6)12.2 / 34.3 / 45.6 (14)
B52-E-F2.5MMLMVTM39.5 / 76.6 / 86.9 (2)20.1 / 47.4 / 60.8 (6)12.1 / 33.6 / 46.0 (14)
", + "image_path": "67d242b659bdf66d52807d5700a4c2d0cdbd1d3e72b475eb42d2a28db4b24754.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 49, + 209, + 552, + 244 + ], + "blocks": [ + { + "bbox": [ + 165, + 196, + 434, + 207 + ], + "lines": [ + { + "bbox": [ + 165, + 196, + 434, + 207 + ], + "spans": [ + { + "bbox": [ + 165, + 196, + 434, + 207 + ], + "type": "text", + "content": "For \"Pre-training\": Ablation Study of Hard Sample Curriculum Learning" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 49, + 209, + 552, + 244 + ], + "lines": [ + { + "bbox": [ + 49, + 209, + 552, + 244 + ], + "spans": [ + { + "bbox": [ + 49, + 209, + 552, + 244 + ], + "type": "table", + "html": "
C12-E-F3.5MMLMVTM39.9 / 77.2 / 87.0 (2)20.7 / 47.9 / 61.2 (6)13.3 / 34.6 / 45.9 (13)
C22-E-F3.5MMLMVTM+HSM40.7 / 76.8 / 86.9 (2)21.1 / 47.3 / 60.1 (6)13.4 / 34.3 / 46.4 (12)
C32-E-F3.5MMLMVTM+HSCL41.5 / 78.2 / 87.2 (2)23.3 / 48.0 / 61.2 (6)13.6 / 34.4 / 47.3 (12)
", + "image_path": "a9f7c9fe68f54926257ccb8785e6da493d86fb817e7a71854359443a05809cca.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 48, + 297, + 550, + 384 + ], + "blocks": [ + { + "bbox": [ + 46, + 253, + 547, + 286 + ], + "lines": [ + { + "bbox": [ + 46, + 253, + 547, + 286 + ], + "spans": [ + { + "bbox": [ + 46, + 253, + 547, + 286 + ], + "type": "text", + "content": "Table 2. For Chinese video-text pre-training: Ablation study of Weakly-Paired Data Filtering and Hard Sample Curriculum Learning (HSCL) on the Text-to-Video Retrieval task of three datasets (VATEX-Chinese, MSRVTT-Translated, and DiDemo-Translated). \"HSM\" denotes the conventional Hard Sample Mining strategy without curriculum learning. Note that " + }, + { + "bbox": [ + 46, + 253, + 547, + 286 + ], + "type": "inline_equation", + "content": "A2" + }, + { + "bbox": [ + 46, + 253, + 547, + 286 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 253, + 547, + 286 + ], + "type": "inline_equation", + "content": "B3" + }, + { + "bbox": [ + 46, + 253, + 547, + 286 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 253, + 547, + 286 + ], + "type": "inline_equation", + "content": "C1" + }, + { + "bbox": [ + 46, + 253, + 547, + 286 + ], + "type": "text", + "content": " are the same model." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 48, + 297, + 550, + 384 + ], + "lines": [ + { + "bbox": [ + 48, + 297, + 550, + 384 + ], + "spans": [ + { + "bbox": [ + 48, + 297, + 550, + 384 + ], + "type": "table", + "html": "
No.Pre-training DatasetMasking StrategyMatching StrategyVATEX (English)MSRVTTDiDemo
R@1/5/10 ↑ (MdR ↓)R@1/5/10 ↑ (MdR ↓)R@1/5/10 ↑ (MdR ↓)
D0No Pre-training39.5 / 78.3 / 88.6 (2)19.2 / 46.2 / 59.5 (6)14.2 / 34.1 / 46.1 (12)
D1COCO+VG+CCMLMVTM45.5 / 81.5 / 91.0 (2)28.6 / 55.5 / 66.1 (4)25.3 / 50.3 / 62.4 (5)
D2COCO+VG+CCMLMVTM+HSCL48.7 / 83.1 / 91.1 (2)31.6 / 56.5 / 66.7 (4)27.1 / 53.8 / 63.8 (4)
D3CC+WebVidMLMVTM50.5 / 85.6 / 92.5 (1)29.4 / 56.9 / 68.3 (4)29.0 / 57.3 / 67.8 (4)
D4CC+WebVidMLMVTM+HSCL53.4 / 86.2 / 92.9 (1)32.6 / 58.8 / 69.7 (3)29.8 / 58.0 / 68.9 (3)
", + "image_path": "5b5c0f38fb5079c16556621b6bc591ddb4e9200bddcf4814a34f326dab1d1a22.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 392, + 546, + 415 + ], + "lines": [ + { + "bbox": [ + 46, + 392, + 546, + 415 + ], + "spans": [ + { + "bbox": [ + 46, + 392, + 546, + 415 + ], + "type": "text", + "content": "Table 3. For English video-text pre-training: Ablation study of Hard Sample Curriculum Learning (HSCL) on the Text-to-Video Retrieval task of three datasets (VATEX-English, MSRVTT, and DiDemo). All models are based on the two-encoders-fusion " + }, + { + "bbox": [ + 46, + 392, + 546, + 415 + ], + "type": "inline_equation", + "content": "(2 - E - F)" + }, + { + "bbox": [ + 46, + 392, + 546, + 415 + ], + "type": "text", + "content": " architecture." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 47, + 424, + 174, + 437 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 424, + 174, + 437 + ], + "spans": [ + { + "bbox": [ + 47, + 424, + 174, + 437 + ], + "type": "text", + "content": "6.3. Experimental Settings" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 442, + 287, + 585 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 442, + 287, + 585 + ], + "spans": [ + { + "bbox": [ + 46, + 442, + 287, + 585 + ], + "type": "text", + "content": "Settings for Weakly-paired Data Filtering. We pretrained the image-text model on the Wukong-100M [11] dataset, whose pre-training details are presented in the supplementary material. We then employed the pre-trained model to obtain the parameter-free video-text consistency score " + }, + { + "bbox": [ + 46, + 442, + 287, + 585 + ], + "type": "inline_equation", + "content": "(z_{i,i}" + }, + { + "bbox": [ + 46, + 442, + 287, + 585 + ], + "type": "text", + "content": " in Eq. 9) of 4.5M video-text pairs. To improve the reliability, we randomly sampled four frames to calculate the consistency score, and repeated this step for three times. The final score of each video is the average of three results calculated under different random seeds. Ultimately, we sorted the consistency score in descending order, and filtered out the last 1M videos to refine the dataset." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 586, + 287, + 694 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 586, + 287, + 694 + ], + "spans": [ + { + "bbox": [ + 46, + 586, + 287, + 694 + ], + "type": "text", + "content": "Settings for Chinese Video-Text Pre-training. For model architectures, we employed the Video Swin Transformer [21] to serve as the visual encoder, which is initialized with parameters pre-trained on ImageNet [6]. The text, cross-modal, and shared encoders belong to the BERT-Base [7] model, which are initialized with parameters privdied by Hugging Face " + }, + { + "bbox": [ + 46, + 586, + 287, + 694 + ], + "type": "inline_equation", + "content": "\\ddagger" + }, + { + "bbox": [ + 46, + 586, + 287, + 694 + ], + "type": "text", + "content": ". We sparsely sampled " + }, + { + "bbox": [ + 46, + 586, + 287, + 694 + ], + "type": "inline_equation", + "content": "4(N_v)" + }, + { + "bbox": [ + 46, + 586, + 287, + 694 + ], + "type": "text", + "content": " frames from raw videos during pre-training. For the MLM task, the masking rate is " + }, + { + "bbox": [ + 46, + 586, + 287, + 694 + ], + "type": "inline_equation", + "content": "15\\%" + }, + { + "bbox": [ + 46, + 586, + 287, + 694 + ], + "type": "text", + "content": ". For the proposed Hard Sample" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 425, + 545, + 448 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 425, + 545, + 448 + ], + "spans": [ + { + "bbox": [ + 306, + 425, + 545, + 448 + ], + "type": "text", + "content": "Curriculum Learning, the coefficient " + }, + { + "bbox": [ + 306, + 425, + 545, + 448 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 306, + 425, + 545, + 448 + ], + "type": "text", + "content": " would increase by " + }, + { + "bbox": [ + 306, + 425, + 545, + 448 + ], + "type": "inline_equation", + "content": "0.15(\\beta_0)" + }, + { + "bbox": [ + 306, + 425, + 545, + 448 + ], + "type": "text", + "content": " for each " + }, + { + "bbox": [ + 306, + 425, + 545, + 448 + ], + "type": "inline_equation", + "content": "5000(N_d)" + }, + { + "bbox": [ + 306, + 425, + 545, + 448 + ], + "type": "text", + "content": " iterations until reaching 1.0." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 449, + 545, + 543 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 449, + 545, + 543 + ], + "spans": [ + { + "bbox": [ + 305, + 449, + 545, + 543 + ], + "type": "text", + "content": "For hyper-parameters, we set the length of text tokens " + }, + { + "bbox": [ + 305, + 449, + 545, + 543 + ], + "type": "inline_equation", + "content": "N_{t} = 30" + }, + { + "bbox": [ + 305, + 449, + 545, + 543 + ], + "type": "text", + "content": ", and the dimension of the hidden state " + }, + { + "bbox": [ + 305, + 449, + 545, + 543 + ], + "type": "inline_equation", + "content": "d = 768" + }, + { + "bbox": [ + 305, + 449, + 545, + 543 + ], + "type": "text", + "content": ". All models are pre-trained by the Adam optimizer with a momentum of 0.9. The total pre-training stage lasts for 10 epochs with a batch size of 128. The initial learning rate is 5e-5 and is decayed by the factor of 10 after 5 epochs. The whole pre-training on CNVid-3.5M takes about 4 days to complete on 8 NVIDIA V100 GPUs." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 544, + 545, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 544, + 545, + 604 + ], + "spans": [ + { + "bbox": [ + 305, + 544, + 545, + 604 + ], + "type": "text", + "content": "Settings for Fine-tuning on Downstream Chinese Videotext Tasks. The optimizer and hyper-parameters in fine-tuning remain the same as the pre-training configuration. The total fine-tuning stage lasts for 25,000 steps. The batch size is set to 128. The initial learning rate is set to 1e-5." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 611, + 400, + 624 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 611, + 400, + 624 + ], + "spans": [ + { + "bbox": [ + 306, + 611, + 400, + 624 + ], + "type": "text", + "content": "6.4. Ablation Study" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 305, + 630, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 630, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 630, + 545, + 713 + ], + "type": "text", + "content": "As aforementioned, we summarized our work with three verbs, i.e., \"Build\", \"Filter\", and \"Pre-train\". To evaluate the effectiveness of these three contributions, we took the 2-" + }, + { + "bbox": [ + 305, + 630, + 545, + 713 + ], + "type": "inline_equation", + "content": "E-F" + }, + { + "bbox": [ + 305, + 630, + 545, + 713 + ], + "type": "text", + "content": " architecture as an example model in the ablation study. Ablation Study of \"Building\". We compared the performance obtained by models with or without pre-training. As illustrated in Table 2 (A1 vs. A2), models pre-trained on" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 58, + 702, + 251, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 702, + 251, + 713 + ], + "spans": [ + { + "bbox": [ + 58, + 702, + 251, + 713 + ], + "type": "inline_equation", + "content": "^\\ddagger" + }, + { + "bbox": [ + 58, + 702, + 251, + 713 + ], + "type": "text", + "content": " Official Website: https://huggingface.co/bert-base-chinese" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14821" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 50, + 70, + 542, + 129 + ], + "blocks": [ + { + "bbox": [ + 50, + 70, + 542, + 129 + ], + "lines": [ + { + "bbox": [ + 50, + 70, + 542, + 129 + ], + "spans": [ + { + "bbox": [ + 50, + 70, + 542, + 129 + ], + "type": "table", + "html": "
No.Model TypeDatasetPre-training StrategyVATEX (Chinese)MSRVTT (Translated)DiDemo (Translated)
R@1/5/10 ↑ (MdR ↓)R@1/5/10 ↑ (MdR ↓)R@1/5/10 ↑ (MdR ↓)
E13-E-FCNVid-3.5MMLM+VTM39.9 / 78.2 / 87.6 (2)20.6 / 47.1 / 60.0 (6)13.8 / 34.5 / 46.6 (13)
E22-E-FCNVid-3.5MMLM+VTM39.9 / 77.2 / 87.0 (2)20.7 / 47.9 / 61.2 (6)13.3 / 34.6 / 45.9 (13)
E3T-T-CCNVid-3.5MMLM+VTM40.5 / 77.3 / 87.0 (2)21.1 / 46.5 / 58.9 (6)13.0 / 32.6 / 44.6 (15)
", + "image_path": "0654043d02b55207e91364054e3ee9d60860821669438fc65b92ec74a541b209.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 137, + 546, + 160 + ], + "lines": [ + { + "bbox": [ + 46, + 137, + 546, + 160 + ], + "spans": [ + { + "bbox": [ + 46, + 137, + 546, + 160 + ], + "type": "text", + "content": "Table 4. We benchmark CNVid-3.5M with three mainstream pixel-level pre-training architectures, which is the first pixel-level benchmarks for Chinese video-text pre-training." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 169, + 286, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 169, + 286, + 217 + ], + "spans": [ + { + "bbox": [ + 46, + 169, + 286, + 217 + ], + "type": "text", + "content": "our CNVid-3.5M dataset (A2) would achieve a remarkable performance gain compared with those without pre-training (A1), which demonstrates the effectiveness of building a large-scale Chinese video-text dataset." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 218, + 286, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 218, + 286, + 312 + ], + "spans": [ + { + "bbox": [ + 46, + 218, + 286, + 312 + ], + "type": "text", + "content": "Ablation Study of \"Filtering\". We compared the performance obtained by models pre-trained under different numbers of video-text pairs. For example, \"4.5M\", \"4.0M\", and \"3.5M\" denotes that we employ the 4.5M full set, the Top-4.0M split set, and the Top-3.5M split set (CNVid-3.5M) according to the sorted video-text consistency scores calculated by weakly-paired data filtering. As illustrated in Table 2 (B1-B5), we have two conclusions as follows:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 315, + 286, + 470 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 46, + 315, + 286, + 397 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 315, + 286, + 397 + ], + "spans": [ + { + "bbox": [ + 46, + 315, + 286, + 397 + ], + "type": "text", + "content": "1) The proposed weakly-paired data filtering strategy could effectively improve the quality of datasets. The fine-tuning performance on three downstream datasets first increases rapidly when filtering out videos with low video-text consistency scores (" + }, + { + "bbox": [ + 46, + 315, + 286, + 397 + ], + "type": "inline_equation", + "content": "B1 \\to B2 \\to B3" + }, + { + "bbox": [ + 46, + 315, + 286, + 397 + ], + "type": "text", + "content": "). It would remain stable (" + }, + { + "bbox": [ + 46, + 315, + 286, + 397 + ], + "type": "inline_equation", + "content": "B3 \\to B4" + }, + { + "bbox": [ + 46, + 315, + 286, + 397 + ], + "type": "text", + "content": ") or drop slightly (" + }, + { + "bbox": [ + 46, + 315, + 286, + 397 + ], + "type": "inline_equation", + "content": "B4 \\to B5" + }, + { + "bbox": [ + 46, + 315, + 286, + 397 + ], + "type": "text", + "content": ") when continuously removing videos after the Top-3.5M split." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 399, + 286, + 470 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 399, + 286, + 470 + ], + "spans": [ + { + "bbox": [ + 46, + 399, + 286, + 470 + ], + "type": "text", + "content": "2) In general, employing the Top-3.5M split (B3) achieves the best performance. The results among all ablation models (B1-B5) may prove that, for the video-text pre-training, it is not \"more data is better\", but \"more good data is better\". Therefore, we filtered out the last 1M videos, resulting in the CNVid-3.5M dataset." + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 46, + 472, + 286, + 531 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 472, + 286, + 531 + ], + "spans": [ + { + "bbox": [ + 46, + 472, + 286, + 531 + ], + "type": "text", + "content": "Ablation Study of \"Pre-training\" (HSCL). To prove the effectiveness of Hard Sample Curriculum Learning (HSCL), we pre-trained several ablation models on Chinese video-text datasets (" + }, + { + "bbox": [ + 46, + 472, + 286, + 531 + ], + "type": "inline_equation", + "content": "C1-C3" + }, + { + "bbox": [ + 46, + 472, + 286, + 531 + ], + "type": "text", + "content": " in Table 2) and English ones (" + }, + { + "bbox": [ + 46, + 472, + 286, + 531 + ], + "type": "inline_equation", + "content": "D1-D4" + }, + { + "bbox": [ + 46, + 472, + 286, + 531 + ], + "type": "text", + "content": " in Table 3). We have two conclusions as follows:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 533, + 286, + 688 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 46, + 533, + 286, + 627 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 533, + 286, + 627 + ], + "spans": [ + { + "bbox": [ + 46, + 533, + 286, + 627 + ], + "type": "text", + "content": "1) HSCL is a model-agnostic strategy, which could evidently improve the pre-training performance. In Table 2, compared with the baseline (C1), HSCL (C3) achieves an improvement of " + }, + { + "bbox": [ + 46, + 533, + 286, + 627 + ], + "type": "inline_equation", + "content": "1.6\\%" + }, + { + "bbox": [ + 46, + 533, + 286, + 627 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 533, + 286, + 627 + ], + "type": "inline_equation", + "content": "2.6\\%" + }, + { + "bbox": [ + 46, + 533, + 286, + 627 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 533, + 286, + 627 + ], + "type": "inline_equation", + "content": "0.3\\%" + }, + { + "bbox": [ + 46, + 533, + 286, + 627 + ], + "type": "text", + "content": " at R@1 of VATEX, MSRVTT, and DiDemo dataset, respectively. While in Table 3, HSCL outperforms the baseline by " + }, + { + "bbox": [ + 46, + 533, + 286, + 627 + ], + "type": "inline_equation", + "content": "3.2 / 3.0 / 1.8\\%" + }, + { + "bbox": [ + 46, + 533, + 286, + 627 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 46, + 533, + 286, + 627 + ], + "type": "inline_equation", + "content": "D1" + }, + { + "bbox": [ + 46, + 533, + 286, + 627 + ], + "type": "text", + "content": " vs. " + }, + { + "bbox": [ + 46, + 533, + 286, + 627 + ], + "type": "inline_equation", + "content": "D2" + }, + { + "bbox": [ + 46, + 533, + 286, + 627 + ], + "type": "text", + "content": " on " + }, + { + "bbox": [ + 46, + 533, + 286, + 627 + ], + "type": "inline_equation", + "content": "\\mathrm{COCO + VG + CC}" + }, + { + "bbox": [ + 46, + 533, + 286, + 627 + ], + "type": "text", + "content": ") and " + }, + { + "bbox": [ + 46, + 533, + 286, + 627 + ], + "type": "inline_equation", + "content": "2.9 / 3.2 / 0.8\\%" + }, + { + "bbox": [ + 46, + 533, + 286, + 627 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 46, + 533, + 286, + 627 + ], + "type": "inline_equation", + "content": "D3" + }, + { + "bbox": [ + 46, + 533, + 286, + 627 + ], + "type": "text", + "content": " vs. " + }, + { + "bbox": [ + 46, + 533, + 286, + 627 + ], + "type": "inline_equation", + "content": "D4" + }, + { + "bbox": [ + 46, + 533, + 286, + 627 + ], + "type": "text", + "content": " on " + }, + { + "bbox": [ + 46, + 533, + 286, + 627 + ], + "type": "inline_equation", + "content": "\\mathrm{CC + WebVid}" + }, + { + "bbox": [ + 46, + 533, + 286, + 627 + ], + "type": "text", + "content": ") at R@1 of three TVR datasets." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 628, + 286, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 628, + 286, + 688 + ], + "spans": [ + { + "bbox": [ + 46, + 628, + 286, + 688 + ], + "type": "text", + "content": "2) The proposed Curriculum Learning (CL) strategy in HSCL could effectively mitigate the side effect of conventional Hard Sample Mining (HSM) when the model is far from convergence. As illustrated in Table 2 (C1 vs. C2), models equipped with HSCL perform better than HSM." + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 46, + 689, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 689, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 689, + 286, + 713 + ], + "type": "text", + "content": "Moreover, we present the detailed quantitative analyses of PCW and NLS in the supplementary material." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 169, + 538, + 182 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 169, + 538, + 182 + ], + "spans": [ + { + "bbox": [ + 305, + 169, + 538, + 182 + ], + "type": "text", + "content": "6.5. Chinese Video-text Pre-training Benchmarks" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 187, + 545, + 282 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 187, + 545, + 282 + ], + "spans": [ + { + "bbox": [ + 304, + 187, + 545, + 282 + ], + "type": "text", + "content": "Based on the constructed CNVid-3.5M dataset, We provided the first pixel-level benchmarks for Chinese videotext pre-training. Following existing video-text pre-training methods, we adopted three mainstream pixel-level architectures, i.e., three-encoders-fusion (3-E-F) [15, 17, 30], two-encoders-fusion (2-E-F) [9], and twin-towers-crossed (T-T-C) [8, 23, 36]. Detailed performance comparisons of three pixel-level benchmarks are presented in Table 4." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 294, + 378, + 306 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 294, + 378, + 306 + ], + "spans": [ + { + "bbox": [ + 305, + 294, + 378, + 306 + ], + "type": "text", + "content": "7. Conclusion" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 314, + 545, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 314, + 545, + 434 + ], + "spans": [ + { + "bbox": [ + 304, + 314, + 545, + 434 + ], + "type": "text", + "content": "In this work, we release the largest public Chinese video-text dataset, i.e., CNVid-3.5M. Besides, we provide the first pixel-level benchmarks for Chinese video-text pretraining. Moreover, we propose a novel weakly-paired data filtering method to improve the quality of datasets. We also design a novel Hard Sample Curriculum Learning strategy to promote the pre-training performance. In conclusion, we \"build\", \"filter\", and \"pre-train\" the large-scale Chinese video-text dataset. We believe this work would pave the way for future research on Chinese video-text pre-training." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 305, + 444, + 471, + 456 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 444, + 471, + 456 + ], + "spans": [ + { + "bbox": [ + 305, + 444, + 471, + 456 + ], + "type": "text", + "content": "8. Limitations and Future Work" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 464, + 545, + 668 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 464, + 545, + 668 + ], + "spans": [ + { + "bbox": [ + 304, + 464, + 545, + 668 + ], + "type": "text", + "content": "Address the Difference between English and Chinese. One potential limitation of our work is that we adopt pixel-level architectures proved effective in English video-text pre-training, ignoring the linguistic gap between Chinese and English. E.g., it may be better to employ the Whole Word Masking (WWM) [5, 46] strategy rather than the conventional MLM for Chinese video-text pre-training. We plan to explore more Chinese-specific methods in the future. Need More Downstream Chinese Video-Text Datasets. Though we release a large-scale Chinese video-text dataset, there still lack some downstream datasets (e.g., MSRVTT [41], MSVD-QA [39] in English) to comprehensively evaluate the performance of Chinese video-text pre-trained models. Currently, there only exists VATEX [38] for Chinese cross-modal retrieval and some captioning datasets [26, 45]. Therefore, building more downstream Chinese video datasets is a promising future direction." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 675, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 675, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 675, + 545, + 712 + ], + "type": "text", + "content": "Acknowledgments. This work is supported by the National Natural Science Foundation of China, No.: 62176137; Ant Group." + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14822" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 289, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 134 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 134 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 134 + ], + "type": "text", + "content": "[1] Nayyer Aafaq, Ajmal Mian, Wei Liu, Syed Zulqarnain Gilani, and Mubarak Shah. Video description: A survey of methods, datasets, and evaluation metrics. ACM Computing Surveys, 52(6):1-37, 2019. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 135, + 288, + 190 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 135, + 288, + 190 + ], + "spans": [ + { + "bbox": [ + 53, + 135, + 288, + 190 + ], + "type": "text", + "content": "[2] Hassan Akbari, Liangzhe Yuan, Rui Qian, Wei-Hong Chuang, Shih-Fu Chang, Yin Cui, and Boqing Gong. Vatt: Transformers for multimodal self-supervised learning from raw video, audio and text. Advances in Neural Information Processing Systems, 34, 2021. 4" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 191, + 289, + 245 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 191, + 289, + 245 + ], + "spans": [ + { + "bbox": [ + 53, + 191, + 289, + 245 + ], + "type": "text", + "content": "[3] Lisa Anne Hendricks, Oliver Wang, Eli Shechtman, Josef Sivic, Trevor Darrell, and Bryan Russell. Localizing moments in video with natural language. In Proceedings of the IEEE international conference on computer vision, pages 5803-5812, 2017. 5, 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 246, + 288, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 246, + 288, + 300 + ], + "spans": [ + { + "bbox": [ + 53, + 246, + 288, + 300 + ], + "type": "text", + "content": "[4] Max Bain, Arsha Nagrani, Gül Varol, and Andrew Zisserman. Frozen in time: A joint video and image encoder for end-to-end retrieval. In Proceedings of the IEEE International Conference on Computer Vision, pages 1728-1738, 2021. 1, 2, 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 301, + 288, + 345 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 301, + 288, + 345 + ], + "spans": [ + { + "bbox": [ + 53, + 301, + 288, + 345 + ], + "type": "text", + "content": "[5] Yiming Cui, Wanxiang Che, Ting Liu, Bing Qin, and Ziqing Yang. Pre-training with whole word masking for chinese bert. IEEE Transactions on Audio, Speech, and Language Processing, 29:3504-3514, 2021. 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 346, + 288, + 390 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 346, + 288, + 390 + ], + "spans": [ + { + "bbox": [ + 53, + 346, + 288, + 390 + ], + "type": "text", + "content": "[6] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In IEEE Conference on Computer Vision and Pattern Recognition, pages 248-255, 2009. 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 391, + 288, + 434 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 391, + 288, + 434 + ], + "spans": [ + { + "bbox": [ + 53, + 391, + 288, + 434 + ], + "type": "text", + "content": "[7] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018. 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 435, + 288, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 435, + 288, + 468 + ], + "spans": [ + { + "bbox": [ + 53, + 435, + 288, + 468 + ], + "type": "text", + "content": "[8] Han Fang, Pengfei Xiong, Luhui Xu, and Yu Chen. Clip2video: Mastering video-text retrieval via image clip. arXiv preprint arXiv:2106.11097, 2021. 2, 4, 8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 469, + 288, + 513 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 469, + 288, + 513 + ], + "spans": [ + { + "bbox": [ + 53, + 469, + 288, + 513 + ], + "type": "text", + "content": "[9] Tsu-Jui Fu, Linjie Li, Zhe Gan, Kevin Lin, William Yang Wang, Lijuan Wang, and Zicheng Liu. Violet: End-to-end video-language transformers with masked visual-token modeling. arXiv preprint arXiv:2111.12681, 2021. 2, 4, 6, 8" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 514, + 288, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 514, + 288, + 568 + ], + "spans": [ + { + "bbox": [ + 48, + 514, + 288, + 568 + ], + "type": "text", + "content": "[10] Yuying Ge, Yixiao Ge, Xihui Liu, Dian Li, Ying Shan, Xiaohu Qie, and Ping Luo. Bridging video-text retrieval with multiple choice questions. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 16167-16176, 2022. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 569, + 288, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 569, + 288, + 635 + ], + "spans": [ + { + "bbox": [ + 48, + 569, + 288, + 635 + ], + "type": "text", + "content": "[11] Jiaxi Gu, Xiaojun Meng, Guansong Lu, Lu Hou, Minzhe Niu, Hang Xu, Xiaodan Liang, Wei Zhang, Xin Jiang, and Chunjing Xu. Wukong: 100 million large-scale chinese cross-modal pre-training dataset and a foundation framework. arXiv preprint arXiv:2202.06767, 2022. 2, 4, 6, 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 635, + 288, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 635, + 288, + 679 + ], + "spans": [ + { + "bbox": [ + 48, + 635, + 288, + 679 + ], + "type": "text", + "content": "[12] Yannis Kalantidis, Mert Bulent Sariyildiz, Noe Pion, Philippe Weinzaepfel, and Diane Larlus. Hard negative mixing for contrastive learning. Advances in Neural Information Processing Systems, 33:21798-21809, 2020. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 680, + 288, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 680, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 680, + 288, + 713 + ], + "type": "text", + "content": "[13] Ranjay Krishna, Yuke Zhu, Oliver Groth, Justin Johnson, Kenji Hata, Joshua Kravitz, Stephanie Chen, Yannis Kalantidis, Li-Jia Li, David A Shamma, et al. Visual genome:" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 547, + 713 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 326, + 73, + 547, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 547, + 106 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 547, + 106 + ], + "type": "text", + "content": "Connecting language and vision using crowdsourced dense image annotations. International Journal of Computer Vision, 123(1):32-73, 2017. 4, 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 106, + 547, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 106, + 547, + 171 + ], + "spans": [ + { + "bbox": [ + 308, + 106, + 547, + 171 + ], + "type": "text", + "content": "[14] Chenyi Lei, Shixian Luo, Yong Liu, Wanggui He, Jiamang Wang, Guoxin Wang, Haihong Tang, Chunyan Miao, and Houqiang Li. Understanding chinese video and language via contrastive multimodal pre-training. In Proceedings of the ACM International Conference on Multimedia, pages 2567-2576, 2021. 1, 2, 3, 4" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 172, + 547, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 172, + 547, + 237 + ], + "spans": [ + { + "bbox": [ + 308, + 172, + 547, + 237 + ], + "type": "text", + "content": "[15] Jie Lei, Linjie Li, Luowei Zhou, Zhe Gan, Tamara L Berg, Mohit Bansal, and Jingjing Liu. Less is more: Clipbert for video-and-language learning via sparse sampling. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 7331-7341, 2021. 1, 2, 4, 5, 6, 8" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 238, + 547, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 238, + 547, + 282 + ], + "spans": [ + { + "bbox": [ + 308, + 238, + 547, + 282 + ], + "type": "text", + "content": "[16] Chunyuan Li, Xiujun Li, Lei Zhang, Baolin Peng, Mingyuan Zhou, and Jianfeng Gao. Self-supervised pre-training with hard examples improves visual representations. arXiv preprint arXiv:2012.13493, 2020. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 283, + 547, + 337 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 283, + 547, + 337 + ], + "spans": [ + { + "bbox": [ + 308, + 283, + 547, + 337 + ], + "type": "text", + "content": "[17] Dongxu Li, Junnan Li, Hongdong Li, Juan Carlos Niebles, and Steven CH Hoi. Align and prompt: Video-and-language pre-training with entity prompts. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 4953-4963, 2022. 1, 2, 4, 8" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 338, + 547, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 338, + 547, + 392 + ], + "spans": [ + { + "bbox": [ + 308, + 338, + 547, + 392 + ], + "type": "text", + "content": "[18] Junnan Li, Ramprasaath Selvaraju, Akhilesh Gotmare, Shafiq Joty, Caiming Xiong, and Steven Chu Hong Hoi. Align before fuse: Vision and language representation learning with momentum distillation. Advances in neural information processing systems, 34:9694-9705, 2021. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 393, + 547, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 393, + 547, + 437 + ], + "spans": [ + { + "bbox": [ + 308, + 393, + 547, + 437 + ], + "type": "text", + "content": "[19] Linjie Li, Yen-Chun Chen, Yu Cheng, Zhe Gan, Licheng Yu, and Jingjing Liu. Hero: Hierarchical encoder for video+language omni-representation pre-training. arXiv preprint arXiv:2005.00200, 2020. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 437, + 547, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 437, + 547, + 491 + ], + "spans": [ + { + "bbox": [ + 308, + 437, + 547, + 491 + ], + "type": "text", + "content": "[20] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Proceedings of the European Conference on Computer Vision, pages 740-755. Springer, 2014. 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 492, + 547, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 492, + 547, + 536 + ], + "spans": [ + { + "bbox": [ + 308, + 492, + 547, + 536 + ], + "type": "text", + "content": "[21] Ze Liu, Jia Ning, Yue Cao, Yixuan Wei, Zheng Zhang, Stephen Lin, and Han Hu. Video swim transformer. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 3202-3211, 2022. 7" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 537, + 547, + 591 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 537, + 547, + 591 + ], + "spans": [ + { + "bbox": [ + 308, + 537, + 547, + 591 + ], + "type": "text", + "content": "[22] Huaishao Luo, Lei Ji, Botian Shi, Haoyang Huang, Nan Duan, Tianrui Li, Jason Li, Taroon Bharti, and Ming Zhou. Univl: A unified video and language pre-training model for multimodal understanding and generation. arXiv preprint arXiv:2002.06353, 2020. 1, 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 592, + 547, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 592, + 547, + 635 + ], + "spans": [ + { + "bbox": [ + 308, + 592, + 547, + 635 + ], + "type": "text", + "content": "[23] Huaishao Luo, Lei Ji, Ming Zhong, Yang Chen, Wen Lei, Nan Duan, and Tianrui Li. Clip4clip: An empirical study of clip for end to end video clip retrieval. arXiv preprint arXiv:2104.08860, 2021. 2, 4, 8" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 635, + 547, + 691 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 635, + 547, + 691 + ], + "spans": [ + { + "bbox": [ + 308, + 635, + 547, + 691 + ], + "type": "text", + "content": "[24] Jianjie Luo, Yehao Li, Yingwei Pan, Ting Yao, Hongyang Chao, and Tao Mei. Coco-bert: Improving video-language pre-training with contrastive cross-modal matching and denoising. In Proceedings of the ACM International Conference on Multimedia, pages 5600-5608, 2021. 2, 4" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 691, + 547, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 691, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 691, + 547, + 713 + ], + "type": "text", + "content": "[25] Antoine Miech, Dimitri Zhukov, Jean-Baptiste Alayrac, Makarand Tapaswi, Ivan Laptev, and Josef Sivic." + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "14823" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 66, + 72, + 287, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 72, + 287, + 116 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 287, + 116 + ], + "type": "text", + "content": "Howto100m: Learning a text-video embedding by watching hundred million narrated video clips. In Proceedings of the IEEE International Conference on Computer Vision, pages 2630-2640, 2019. 1, 2, 3, 4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 119, + 287, + 174 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 119, + 287, + 174 + ], + "spans": [ + { + "bbox": [ + 48, + 119, + 287, + 174 + ], + "type": "text", + "content": "[26] Liqiang Nie, Leigang Qu, Dai Meng, Min Zhang, Qi Tian, and Alberto Del Bimbo. Search-oriented microvideo captioning. In Proceedings of the ACM International Conference on Multimedia, pages 3234-3243, 2022. 1, 2, 3, 4, 8" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 177, + 287, + 220 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 177, + 287, + 220 + ], + "spans": [ + { + "bbox": [ + 48, + 177, + 287, + 220 + ], + "type": "text", + "content": "[27] Devshree Patel, Ratnam Parikh, and Yesha Shastri. Recent advances in video question answering: A review of datasets and methods. In International Conference on Pattern Recognition, pages 339-356. Springer, 2021. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 223, + 287, + 289 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 223, + 287, + 289 + ], + "spans": [ + { + "bbox": [ + 48, + 223, + 287, + 289 + ], + "type": "text", + "content": "[28] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 2, 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 291, + 287, + 334 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 291, + 287, + 334 + ], + "spans": [ + { + "bbox": [ + 48, + 291, + 287, + 334 + ], + "type": "text", + "content": "[29] Joshua David Robinson, Ching-Yao Chuang, Suvrit Sra, and Stefanie Jegelka. Contrastive learning with hard negative samples. In International Conference on Learning Representations, 2020. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 337, + 287, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 337, + 287, + 392 + ], + "spans": [ + { + "bbox": [ + 48, + 337, + 287, + 392 + ], + "type": "text", + "content": "[30] Paul Hongsuck Seo, Arsha Nagrani, Anurag Arnab, and Cordelia Schmid. End-to-end generative pretraining for multimodal video captioning. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 17959-17968, 2022. 1, 2, 4, 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 394, + 287, + 449 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 394, + 287, + 449 + ], + "spans": [ + { + "bbox": [ + 48, + 394, + 287, + 449 + ], + "type": "text", + "content": "[31] Piyush Sharma, Nan Ding, Sebastian Goodman, and Radu Soricut. Conceptual captions: A cleaned, hypernymed, image alt-text dataset for automatic image captioning. In Proceedings of the Annual Meeting of the Association for Computational Linguistics, pages 2556–2565, 2018. 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 452, + 287, + 485 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 452, + 287, + 485 + ], + "spans": [ + { + "bbox": [ + 48, + 452, + 287, + 485 + ], + "type": "text", + "content": "[32] Petru Soviany, Radu Tudor Ionescu, Paolo Rota, and Nicu Sebe. Curriculum learning: A survey. International Journal of Computer Vision, pages 1-40, 2022. 2, 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 487, + 287, + 540 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 487, + 287, + 540 + ], + "spans": [ + { + "bbox": [ + 48, + 487, + 287, + 540 + ], + "type": "text", + "content": "[33] Chen Sun, Austin Myers, Carl Vondrick, Kevin Murphy, and Cordelia Schmid. Videobert: A joint model for video and language representation learning. In Proceedings of the IEEE International Conference on Computer Vision, pages 7464-7473, 2019. 2, 4" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 544, + 287, + 587 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 544, + 287, + 587 + ], + "spans": [ + { + "bbox": [ + 48, + 544, + 287, + 587 + ], + "type": "text", + "content": "[34] Guanglu Sun, Lili Liang, Tianlin Li, Bo Yu, Meng Wu, and Bolun Zhang. Video question answering: a survey of models and datasets. Mobile Networks and Applications, 26(5):1904-1937, 2021. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 590, + 287, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 590, + 287, + 645 + ], + "spans": [ + { + "bbox": [ + 48, + 590, + 287, + 645 + ], + "type": "text", + "content": "[35] Alex Jinpeng Wang, Yixiao Ge, Rui Yan, Yuying Ge, Xudong Lin, Guanyu Cai, Jianping Wu, Ying Shan, Xiaohu Qie, and Mike Zheng Shou. All in one: Exploring unified video-language pre-training. arXiv preprint arXiv:2203.07303, 2022. 2, 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 647, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 647, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 647, + 287, + 712 + ], + "type": "text", + "content": "[36] Jinpeng Wang, Yixiao Ge, Guanyu Cai, Rui Yan, Xudong Lin, Ying Shan, Xiaohu Qie, and Mike Zheng Shou. Object-aware video-language pre-training for retrieval. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 3313-3322, 2022. 1, 2, 4, 8" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 607 + ], + "type": "list", + "angle": 0, + "index": 23, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "text", + "content": "[37] Xin Wang, Yudong Chen, and Wenwu Zhu. A survey on curriculum learning. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2021. 2, 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 307, + 107, + 545, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 107, + 545, + 162 + ], + "spans": [ + { + "bbox": [ + 307, + 107, + 545, + 162 + ], + "type": "text", + "content": "[38] Xin Wang, Jiawei Wu, Junkun Chen, Lei Li, Yuan-Fang Wang, and William Yang Wang. Vatex: A large-scale, high-quality multilingual dataset for video-and-language research. In Proceedings of the IEEE International Conference on Computer Vision, pages 4581-4591, 2019. 3, 5, 6, 8" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 163, + 545, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 163, + 545, + 217 + ], + "spans": [ + { + "bbox": [ + 307, + 163, + 545, + 217 + ], + "type": "text", + "content": "[39] Dejing Xu, Zhou Zhao, Jun Xiao, Fei Wu, Hanwang Zhang, Xiangnan He, and Yueting Zhuang. Video question answering via gradually refined attention over appearance and motion. In Proceedings of the ACM International Conference on Multimedia, pages 1645-1653, 2017. 8" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 219, + 545, + 273 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 219, + 545, + 273 + ], + "spans": [ + { + "bbox": [ + 307, + 219, + 545, + 273 + ], + "type": "text", + "content": "[40] Hu Xu, Gargi Ghosh, Po-Yao Huang, Prahal Arora, Masoumeh Aminzadeh, Christoph Feichtenhofer, Florian Metze, and Luke Zettlemoyer. Vlm: Task-agnostic video-language model pre-training for video understanding. arXiv preprint arXiv:2105.09996, 2021. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 274, + 545, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 274, + 545, + 319 + ], + "spans": [ + { + "bbox": [ + 307, + 274, + 545, + 319 + ], + "type": "text", + "content": "[41] Jun Xu, Tao Mei, Ting Yao, and Yong Rui. Msr-vtt: A large video description dataset for bridging video and language. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 5288-5296, 2016. 5, 6, 8" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 319, + 545, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 319, + 545, + 373 + ], + "spans": [ + { + "bbox": [ + 307, + 319, + 545, + 373 + ], + "type": "text", + "content": "[42] Antoine Yang, Antoine Miech, Josef Sivic, Ivan Laptev, and Cordelia Schmid. Just ask: Learning to answer questions from millions of narrated videos. In Proceedings of the IEEE International Conference on Computer Vision, pages 1686-1697, 2021. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 375, + 545, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 375, + 545, + 430 + ], + "spans": [ + { + "bbox": [ + 307, + 375, + 545, + 430 + ], + "type": "text", + "content": "[43] Jianwei Yang, Yonatan Bisk, and Jianfeng Gao. Taco: Token-aware cascade contrastive learning for video-text alignment. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11562-11572, 2021. 2, 3, 5, 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 431, + 545, + 485 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 431, + 545, + 485 + ], + "spans": [ + { + "bbox": [ + 307, + 431, + 545, + 485 + ], + "type": "text", + "content": "[44] Shengyu Zhang, Ziqi Tan, Jin Yu, Zhou Zhao, Kun Kuang, Jie Liu, Jingren Zhou, Hongxia Yang, and Fei Wu. Poet: Product-oriented video captioner for e-commerce. In Proceedings of the ACM International Conference on Multimedia, pages 1292-1301, 2020. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 487, + 545, + 540 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 487, + 545, + 540 + ], + "spans": [ + { + "bbox": [ + 307, + 487, + 545, + 540 + ], + "type": "text", + "content": "[45] Ziqi Zhang, Yuxin Chen, Zongyang Ma, Zhongang Qi, Chunfeng Yuan, Bing Li, Ying Shan, and Weiming Hu. Create: A benchmark for chinese short video retrieval and title generation. arXiv preprint arXiv:2203.16763, 2022. 1, 2, 3, 4, 8" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 543, + 545, + 607 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 543, + 545, + 607 + ], + "spans": [ + { + "bbox": [ + 307, + 543, + 545, + 607 + ], + "type": "text", + "content": "[46] Wei Zhu. Mvp-bert: Multi-vocab pre-training for chinese bert. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing: Student Research Workshop, pages 260-269, 2021. 8" + } + ] + } + ], + "index": 22 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "14824" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/CODA-Prompt_ COntinual Decomposed Attention-Based Prompting for Rehearsal-Free Continual Learning/44360d49-0340-4846-917d-a266d277946e_content_list.json b/2023/CODA-Prompt_ COntinual Decomposed Attention-Based Prompting for Rehearsal-Free Continual Learning/44360d49-0340-4846-917d-a266d277946e_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..fdae26f0cd524ee4604e3e816b9225220086b307 --- /dev/null +++ b/2023/CODA-Prompt_ COntinual Decomposed Attention-Based Prompting for Rehearsal-Free Continual Learning/44360d49-0340-4846-917d-a266d277946e_content_list.json @@ -0,0 +1,1519 @@ +[ + { + "type": "text", + "text": "CODA-Prompt: CContinual Decomposed Attention-based Prompting for Rehearsal-Free Continual Learning", + "text_level": 1, + "bbox": [ + 122, + 128, + 849, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "James Seale Smith $^{1,2}$ Leonid Karlinsky $^{2,4}$ Vyshnavi Gutta $^{1}$ Paola Cascante-Bonilla $^{2,3}$ Donghyun Kim $^{2,4}$ Assaf Arbelle $^{4}$ Rameswar Panda $^{2,4}$ Rogerio Feris $^{2,4}$ Zsolt Kira $^{1}$", + "bbox": [ + 102, + 200, + 867, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1Georgia Institute of Technology 2MIT-IBM Watson AI Lab 3Rice University 4IBM Research", + "bbox": [ + 150, + 239, + 813, + 255 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 291, + 313, + 306 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Computer vision models suffer from a phenomenon known as catastrophic forgetting when learning novel concepts from continuously shifting training data. Typical solutions for this continual learning problem require extensive rehearsal of previously seen data, which increases memory costs and may violate data privacy. Recently, the emergence of large-scale pre-trained vision transformer models has enabled prompting approaches as an alternative to data-rehearsal. These approaches rely on a key-query mechanism to generate prompts and have been found to be highly resistant to catastrophic forgetting in the well-established rehearsal-free continual learning setting. However, the key mechanism of these methods is not trained end-to-end with the task sequence. Our experiments show that this leads to a reduction in their plasticity, hence sacrificing new task accuracy, and inability to benefit from expanded parameter capacity. We instead propose to learn a set of prompt components which are assembled with input-conditioned weights to produce input-conditioned prompts, resulting in a novel attention-based end-to-end key-query scheme. Our experiments show that we outperform the current SOTA method DualPrompt on established benchmarks by as much as $4.5\\%$ in average final accuracy. We also outperform the state of art by as much as $4.4\\%$ accuracy on a continual learning benchmark which contains both class-incremental and domain-incremental task shifts, corresponding to many practical settings. Our code is available at https://github.com/GT-RIPL/CODA-Prompt", + "bbox": [ + 76, + 323, + 473, + 748 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 776, + 209, + 791 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "For a computer vision model to succeed in the real world, it must overcome brittle assumptions that the concepts it will encounter after deployment will match those learned $a$ -priori during training. The real world is indeed dynamic and contains continuously emerging objects and categories. Once deployed, models will encounter a range of differences from their training sets, requiring us to continuously", + "bbox": [ + 75, + 794, + 470, + 902 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/9fc104dd3bfbb342f5ac031223ecc62e911b4c3f4618e42f7592753a4f003667.jpg", + "image_caption": [ + "Prior Work" + ], + "image_footnote": [], + "bbox": [ + 509, + 291, + 885, + 395 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/377f964a93e17e9799fa6fdfaab3d0f5dde02db371cebdffb0392499006b34ce.jpg", + "image_caption": [ + "Our Work", + "Figure 1. Prior work [65] learns a pool of key-value pairs to select learnable prompts which are inserted into several layers of a pretrained ViT (the prompting parameters are unique to each layer). Our work introduces a decomposed prompt which consists of learnable prompt components that assemble to produce attention-conditioned prompts. Unlike prior work, ours is optimized in an end-to-end fashion (denoted with thick, green lines)." + ], + "image_footnote": [], + "bbox": [ + 509, + 415, + 885, + 521 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "update them to avoid stale and decaying performance.", + "bbox": [ + 500, + 672, + 854, + 686 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "One way to update a model is to collect additional training data, combine this new training data with the old training data, and then retrain the model from scratch. While this will guarantee high performance, it is not practical for large-scale applications which may require lengthy training times and aggressive replacement timelines, such as models for self-driving cars or social media platforms. This could incur high financial [26] and environmental [33] costs if the replacement is frequent. We could instead update the model by only training on the new training data, but this leads to a phenomenon known as catastrophic forgetting [46] where the model overwrites existing knowledge when learning the new data, a problem known as continual learning.", + "bbox": [ + 496, + 688, + 892, + 883 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The most effective approaches proposed by the continual", + "bbox": [ + 519, + 885, + 892, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "11909", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "learning community involve saving [50] or generating [55] a subset of past training data and mixing it with future task data, a strategy referred to as rehearsal. Yet many important applications are unable to store this data because they work with private user data that cannot be stored long term.", + "bbox": [ + 75, + 90, + 467, + 166 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we instead consider the highly-impactful and well-established setting of rehearsal-free continual learning [38, 56, 57, 65, 66], limiting our scope to strategies for continual learning which do not store training data. While rehearsal-free approaches have classically under-performed rehearsal-based approaches in this challenging setting by wide-margins [57], recent prompt-based approaches [65, 66], leveraging pre-trained Vision transformers (ViT), have had tremendous success and can even outperform state-of-the-art (SOTA) rehearsal-based methods. These prompting approaches boast a strong protection against catastrophic forgetting by learning a small pool of insert-able model embeddings (prompts) rather than modifying vision encoder parameters directly. One drawback of these approaches, however, is that they cannot be optimized in an end-to-end fashion as they use a key and query to select a prompt index from the pool of prompts, and thus rely on a second, localized optimization to learn the keys because the model gradient cannot backpropagate through the key/query index selection. Furthermore, these methods reduce forgetting by sacrificing new task accuracy (i.e., they lack sufficient plasticity). We show that expanding the prompt size does not increase the plasticity, motivating us to grow learning capacity from a new perspective.", + "bbox": [ + 75, + 167, + 468, + 527 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We replace the prompt pool with a decomposed prompt that consists of a weighted sum of learnable prompt components (Figure 1). This decomposition enables higher prompting capacity by expanding in a new dimension (the number of components). Furthermore, this inherently encourages knowledge re-use, as future task prompts will include contributions from prior task components. We also introduce a novel attention-based component-weighting scheme, which allows for our entire method to be optimized in an end-to-end fashion unlike the existing works, increasing our plasticity to better learn future tasks. We boast significant performance gains on existing rehearsal-free continual learning benchmarks w.r.t. SOTA prompting baselines in addition to a challenging dual-shift benchmark containing both incremental concept shifts and covariate domain shifts, highlighting our method's impact and generality. Our method improves results even under equivalent parameter counts, but importantly can scale performance by increasing capacity, unlike prior methods which quickly plateau. In summary, we make the following contributions:", + "bbox": [ + 75, + 529, + 467, + 830 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. We introduce a decomposed attention-based prompting for rehearsal-free continual learning, characterized by an expanded learning capacity compared to existing continual prompting approaches. Importantly, our approach can be optimized in an end-to-end fashion, unlike the prior SOTA.", + "2. We establish a new SOTA for rehearsal-free continual learning on the well-established ImageNet-R [21, 65] and CIFAR-100 [32] benchmarks, beating the previous SOTA method DualPrompt [65] by as much as $4.5\\%$ .", + "3. We evaluate on a challenging benchmark with dual distribution shifts (semantic and covariate) using the ImageNet-R dataset, and again outperform the state of art, highlighting the real-world impact and generality of our approach." + ], + "bbox": [ + 511, + 90, + 890, + 343 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Background and Related Work", + "text_level": 1, + "bbox": [ + 500, + 359, + 784, + 376 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Continual Learning: Continual learning approaches can be organized into a few broad categories which are all useful depending on the problem setting and constraints. One group of approaches expands a model's architecture as new tasks are encountered; these are highly effective for applications where a model growing with tasks is practical [16, 35, 39, 43, 54]. Another approach is to regularize the model with respect to past task knowledge while training the new task. This can either be done by regularizing the model in the weight space (i.e., penalize changes to model parameters) [2, 15, 30, 58, 71] or the prediction space (i.e., penalize changes to model predictions) [1, 8, 23, 34, 38]. Regularizing knowledge in the prediction space is done using knowledge distillation [22] and it has been found to perform better than model regularization-based methods for continual learning when task labels are not given [36, 60].", + "bbox": [ + 496, + 385, + 890, + 626 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Rehearsal with stored data [3-5, 7, 9, 10, 18, 19, 24, 29, 41, 50-52, 63, 67] or samples from a generative model [27, 28, 47, 55, 59] is highly effective when storing training data or training/saving a generative model is possible. Unfortunately, for many machine learning applications long-term storage of training data will violate data privacy, as well as incur a large memory cost. With respect to the generative model, this training process is much more computationally and memory intensive compared to a classification model and additionally may violate data legality concerns because using a generative model increases the chance of memorizing potentially sensitive data [45]. This motivates us to work on the important setting of rehearsal-free approaches to mitigate catastrophic forgetting.", + "bbox": [ + 496, + 628, + 890, + 839 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Rehearsal-Free Continual Learning: Recent works propose producing images for rehearsal using deep-model inversion [12, 17, 56, 68]. While these methods perform well compared to generative modeling approaches and simply", + "bbox": [ + 496, + 839, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_footnote", + "text": "1We focus on continual learning over a single, expanding classification head called class-incremental continual learning. This is different from the multi-task continual learning setting, known as task-incremental continual learning, where we learn separate classification heads for each task and the task label is provided during inference. [25, 61]", + "bbox": [ + 75, + 838, + 467, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "11910", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "rehearse from a small number of stored images, model-inversion is a slow process associated with high computational costs in the continual learning setting, and furthermore these methods under-perform rehearsal-based approaches by significant margins [56]. Other works have looked at rehearsal-free continual learning from an online \"streaming\" learning perspective using a frozen, pre-trained model [20, 40]. Because we allow our models to train to convergence on task data (as is common for continual learning [67]), our setting is very different.", + "bbox": [ + 75, + 90, + 472, + 241 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Continual Learning in Vision Transformers: Recent work [6, 37, 69] has proven transformers to generalize well to unseen domains. For example, one study varies the number of attention heads in Vision transformers and conclude that ViTs offer more robustness to forgetting with respect to CNN-based equivalents [44]. Another [70] shows that the vanilla counterparts of ViTs are more prone to forgetting when trained from scratch. Finally, DyTox [14] learns a unified model with a parameter-isolation approach and dynamically expands the tokens processed by the last layer to mitigate forgetting. For each task, they learn a new task-specific token per head using task-attention based decoder blocks. However, the above works either rely on exemplars to defy forgetting or they need to train a new transformer model from scratch, a costly operation, hence differing from our objective of exemplar-free CL using pre-trained ViTs.", + "bbox": [ + 75, + 243, + 472, + 484 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Prompting for Continual Learning: Prompt-based approaches for continual learning boast a strong protection against catastrophic forgetting by learning a small number of insert-able model instructions (prompts) rather than modifying encoder parameters directly. The current state-of-the-art approaches for our setting, DualPrompt [65] and L2P [66], create a pool of prompts to be selected for insertion into the model, matching input data to prompts without task id with a local clustering-like optimization. We build upon the foundation of these methods, as discussed later in this paper. We note that the recent S-Prompts [64] method is similar to these approaches but designed for domain-incremental learning (i.e., learning the same set of classes under covariate distribution task shifts), which is different than the class-incremental continual learning setting of our paper (i.e., learn emerging object classes in new tasks).", + "bbox": [ + 75, + 486, + 472, + 728 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Preliminaries", + "text_level": 1, + "bbox": [ + 76, + 744, + 215, + 758 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Continual Learning", + "text_level": 1, + "bbox": [ + 76, + 770, + 266, + 786 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In our continual learning setting, a model is sequentially shown $N$ tasks corresponding to non-overlapping subsets of semantic object classes. Each class appears in only a single task, and the goal is to incrementally learn to classify new object classes as they are introduced while retaining performance on previously learned classes. To describe our models, we denote $\\theta$ as a pre-trained vision encoder and $\\phi_{n}$", + "bbox": [ + 75, + 794, + 468, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "as our classifier head with logits corresponding to task $n$ classes. In this paper, we deal with the class-incremental continual learning setting rather than the task-incremental continual learning setting. Class-incremental continual learning is challenging because the learner must support classification across all classes seen up to task $n$ [25] (i.e., no task labels are provided to the learner during inference). Task-incremental continual learning is a simpler multi-task setting where the task labels are given during both training and inference.", + "bbox": [ + 496, + 90, + 893, + 241 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We also include experiments that contain both class-incremental and domain-incremental continual learning at the same time. The difference between this setting and the former is that we also inject covariate distribution shifts (e.g., task 1 might include real images and clip-art, whereas task 2 might include corrupted photos and artistic paintings). The motivation is to make continual learning more practical - as in the real world we might encounter these dual types of domain shifts as well.", + "bbox": [ + 496, + 243, + 893, + 378 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Prompting with Prefix-Tuning", + "text_level": 1, + "bbox": [ + 498, + 392, + 769, + 409 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In our work, we do not change the technical foundations of prompting from the prior state-of-the-art DualPrompt [65]. We instead focus on the selection and formation of the prompt (including enabling an end-to-end optimization of all prompting components), and the resulting prompt is used by the vision transformer encoder in an identical fashion to DualPrompt. This allows us to make a fair comparison with respect to our novel contributions.", + "bbox": [ + 496, + 417, + 893, + 537 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "As done in DualPrompt, we pass prompts to several multi-head self-attention (MSA) layers in a pre-trained ViT transformer [13, 62] and use prefix-tuning over prompt-tuning $^2$ , which prepends prompts to the keys and values of an MSA layer rather than prepending prompts to the input tokens. The layers in which prompting occurs is set as a hyperparameter, and the prompting parameters are unique between layers. We define a prompt parameter as $\\pmb{p} \\in \\mathbb{R}^{L_p \\times D}$ where $L_{p}$ is the prompt length (chosen as a hyperparameter) and $D$ is the embedding dimension (768). Consider an MSA layer with input $\\pmb{h} \\in \\mathbb{R}^{L \\times D}$ , and query, key, and values given as $\\pmb{h}_{Q}, \\pmb{h}_{K}, \\pmb{h}_{V}$ , respectively. In the ViT model, $\\pmb{h}_{Q} = \\pmb{h}_{K} = \\pmb{h}_{V} = \\pmb{h}$ . The output of this layer is given as:", + "bbox": [ + 496, + 539, + 893, + 736 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\operatorname {M S A} \\left(\\boldsymbol {h} _ {Q}, \\boldsymbol {h} _ {K}, \\boldsymbol {h} _ {V}\\right) = \\operatorname {C o n c a t} \\left(\\mathrm {h} _ {1}, \\dots , \\mathrm {h} _ {\\mathrm {m}}\\right) W ^ {O} \\\\ \\text {w h e r e} \\mathrm {h} _ {\\mathrm {i}} = \\text {A t t e n t i o n} \\left(\\boldsymbol {h} _ {Q} W _ {i} ^ {Q}, \\boldsymbol {h} _ {K} W _ {i} ^ {K}, \\boldsymbol {h} _ {V} W _ {i} ^ {V}\\right) \\tag {1} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 514, + 750, + 890, + 796 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $W^{O}, W_{i}^{Q}, W_{i}^{K}$ , and $W_{i}^{V}$ are projection matrices and $m$ is the number of heads. We split our prompt $\\pmb{p}$ into $\\{\\pmb{p}_K, \\pmb{p}_V\\} \\in \\mathbb{R}^{\\frac{L_p}{2} \\times D}$ and prepend them to $h_K$ and $h_V$ with", + "bbox": [ + 500, + 809, + 893, + 861 + ], + "page_idx": 2 + }, + { + "type": "page_footnote", + "text": "2We refer the reader to sections 4.2 and 5.4 of the the DualPrompt [65] paper for a discussion on this design choice.", + "bbox": [ + 500, + 875, + 893, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "11911", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "prefix-tuning (P-T) as:", + "bbox": [ + 76, + 90, + 230, + 107 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nf _ {P - T} (\\boldsymbol {p}, \\boldsymbol {h}) = \\operatorname {M S A} \\left(\\boldsymbol {h} _ {Q}, \\left[ \\boldsymbol {p} _ {K}; \\boldsymbol {h} _ {K} \\right], \\left[ \\boldsymbol {p} _ {V}; \\boldsymbol {h} _ {V} \\right]\\right) \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 104, + 113, + 468, + 132 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The result is that we now only train a small number of parameter (the prompts) while leaving the rest of the transformer encoder unmodified. The critical question that now remains is how to select and update prompts in a continuous fashion? The next subsection will discuss how prior works select and update prompts.", + "bbox": [ + 73, + 138, + 468, + 229 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. L2P and DualPrompt", + "text_level": 1, + "bbox": [ + 76, + 237, + 279, + 255 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "L2P [66] and DualPrompt [65] select prompts from a pool using an image-wise prompt query. Specifically, these methods use a key-value pair based query strategy to dynamically select instance-specific prompts from a pool of candidate prompts. Each prompt $\\pmb{p}_m$ is selected from a pool of learnable keys $\\pmb{k}_m \\in \\mathbb{R}^D$ where $M$ is the size of the prompt pool, based on cosine similarity to an input-conditioned query. Queries are produced as: $q(\\pmb{x}) \\in \\mathbb{R}^D = \\theta(\\pmb{x})$ where $\\theta$ is the pretrained ViT encoder and $x$ is the input image4. A query $q(\\pmb{x})$ is matched to a key $\\pmb{k}_m$ by selecting the key with the highest cosine similarity, denoted as: $\\gamma(q(\\pmb{x}), \\pmb{k}_m)$ . The keys are learned with a separate optimization from the task classification loss by simply maximizing the cosine similarity between each matched $q(\\pmb{x})$ and $\\pmb{k}_m$ , acting as a clustering-like approach. However, this approach does not directly update the key with gradients from the task loss, which is critical in our experiments.", + "bbox": [ + 73, + 260, + 472, + 518 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Method", + "text_level": 1, + "bbox": [ + 76, + 530, + 169, + 544 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1. Prompt Formation", + "text_level": 1, + "bbox": [ + 76, + 554, + 258, + 570 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "While prompting has been shown to perform exceptionally well in terms of mitigating catastrophic forgetting, the existing state-of-the-art approach DualPrompt lacks the ability to expand learning capacity within a given task. Specifically, DualPrompt learns a single prompt for each new task - regardless if the task is easy (such as learning a handful of new object classes) or complex (such as learning a large number of new object classes). One might try to increase the length of the prompt, but we show in our experiments (Section 5.3) that increasing the prompt length has saturated returns. Intuitively, we instead desire a method where the learning capacity is correlated to the underlying complexity of task data, rather than a single prompt. Additionally, we desire an approach that is end-to-end differentiable, which we conjecture increases the ability to learn a new task with higher accuracy.", + "bbox": [ + 73, + 577, + 468, + 819 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3We note that the difference between L2P and DualPrompt w.r.t. prompt querying is that the size of the prompt pool in L2P is chosen as a hyperparameter, but in DualPrompt it is equal to the number of tasks and, during training, DualPrompt selects the prompt using task-id (and selects the closest key-query match during inference).", + "bbox": [ + 75, + 825, + 468, + 887 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4The output is read from the class token.", + "bbox": [ + 94, + 887, + 310, + 898 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We grow our learning capacity by introducing a new axis: a set of prompt components. Rather than pick and choose prompts from a pool, we learn a set of prompt components which, via a weighted summation, form a decomposed5 prompt that is passed to the corresponding MSA layer. This allows us to grow our prompting capacity to arbitrary depth and capture the rich underlying complexity of any task while maintaining a fixed prompt length. In addition, prompting in new tasks will inherently reuse the previously acquired knowledge of past tasks rather than initializing a new task prompt from scratch.", + "bbox": [ + 496, + 90, + 890, + 256 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Specifically, we replace learnable prompt parameter $\\pmb{p}$ with a weighted summation over the prompt components:", + "bbox": [ + 498, + 257, + 893, + 287 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {p} = \\sum_ {m} \\alpha_ {m} \\boldsymbol {P} _ {m} \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 638, + 297, + 890, + 327 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $P \\in \\mathbb{R}^{L_p \\times D \\times M}$ is our set of prompt components, $M$ is the length of our set (i.e., the introduced axis of additional capacity), and $\\alpha$ is a weighting vector that determines the contribution of each component which will be discussed in the next subsection.", + "bbox": [ + 496, + 338, + 890, + 412 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "A distinct difference between our method and L2P/DualPrompt is that all of our learnable parameters are optimized using the classification loss rather than splitting into two separate loss optimizations. This allows for better end-to-end learning in our method, and we argue this is a key reason why our method performs better in the benchmarks described in Section 5.", + "bbox": [ + 496, + 415, + 890, + 518 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.2. Prompt-Component Weighting", + "text_level": 1, + "bbox": [ + 500, + 529, + 774, + 546 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Rather than a pool of prompts, we have a set of prompt components and desire to produce a weight vector $\\alpha$ given a query $\\theta(x)$ , rather than a prompt index. We compute the cosine similarity between a query and keys to produce our weighting vector as:", + "bbox": [ + 496, + 551, + 890, + 628 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\alpha = \\gamma (q (\\boldsymbol {x}), \\boldsymbol {K}) \\\\ = \\left\\{\\gamma (q (\\boldsymbol {x}), \\boldsymbol {K} _ {1}), \\gamma (q (\\boldsymbol {x}), \\boldsymbol {K} _ {2}), \\dots , \\gamma (q (\\boldsymbol {x}), \\boldsymbol {K} _ {M}) \\right\\} \\tag {4} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 506, + 638, + 890, + 674 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\pmb{K} \\in \\mathbb{R}^{D \\times M}$ contains keys corresponding to our prompt components. The intuition here is that the contributing magnitude of each prompt component $\\pmb{P}_{m}$ to the final prompt $\\pmb{p}$ is determined by the similarity between the query $q(\\pmb{x})$ and corresponding key $\\pmb{K}_{m}$ .", + "bbox": [ + 496, + 684, + 890, + 760 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The prompt-query matching can be thought of as clustering in a high dimension space $D$ which is a well-known and difficult problem [31]. To mitigate this drawback, we introduce another component to our key-query matching: attention. Each $P_{m}$ has a corresponding attention vector $A_{m}$ in addition to the key $K_{m}$ . This allows the query to focus on certain features from the high dimensional query", + "bbox": [ + 496, + 761, + 890, + 866 + ], + "page_idx": 3 + }, + { + "type": "page_footnote", + "text": "5 Decomposed in the sense that our prompt is a weighted summation of components.", + "bbox": [ + 500, + 875, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "11912", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/4d37f8d1134de127a55dcc201666115e9e5d18d45881b047a51a9426b060cab2.jpg", + "image_caption": [ + "Figure 2. Our CODA-Prompt approach to rehearsal-free continual learning. An input (e.g., image) is passed through a query function (we use the pretrained ViT encoder) and used for a novel attention-based prompt-selection scheme to produce prompts which are passed to multiple layers of a large-scale, pre-trained transformer. Our method is parameterized by an expanding set of small prompt components, with each prompt having a corresponding key and attention vector. Both the input and produced prompts are passed through the transformer and sent to the task head. Only the prompting and task parameters are learned which is parameter efficient and no training data is stored for replay which is memory-efficient and privacy preserving. Importantly, our prompt selection scheme can be optimized end-to-end, unlike prior works which have a local similarity-based optimization for key-query matching." + ], + "image_footnote": [], + "bbox": [ + 98, + 94, + 872, + 287 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "$q(\\pmb{x})$ output - for example, a prompt component used for mammalian face features to classify dogs and cats would be able to focus on relevant query features such as eye placement and skin texture, and furthermore ignore features of unrelated semantics, such as features useful for automobile classification. We use a simple feature-selection attention scheme with element-wise multiplication between the query vector and attention vector to create an attended-query which is then used for the key-similarity matching. Specifically, our updated approach to producing our weighting vector is:", + "bbox": [ + 75, + 405, + 473, + 571 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\begin{array}{l} \\alpha = \\gamma (q (\\boldsymbol {x}) \\odot \\boldsymbol {A}, \\boldsymbol {K}) \\\\ = \\left\\{\\gamma (q (\\boldsymbol {x}) \\odot \\boldsymbol {A}, \\boldsymbol {K}), \\quad \\gamma (q (\\boldsymbol {x}) \\odot \\boldsymbol {A}, \\boldsymbol {K}), \\right\\} \\end{array} \\tag {5} \\\\ = \\left\\{\\gamma (q (\\boldsymbol {x}) \\odot \\boldsymbol {A} _ {1}, \\boldsymbol {K} _ {1}), \\dots , \\gamma (q (\\boldsymbol {x}) \\odot \\boldsymbol {A} _ {M}, \\boldsymbol {K} _ {M}) \\right\\} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 84, + 580, + 468, + 616 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathbf{A} \\in \\mathbb{R}^{D \\times M}$ contains learnable parameters (attention vectors) corresponding to our prompt components and $\\odot$ is the element-wise product (also known as the Hadamard product). Notice that our attention vectors are learnable feature weightings rather than input-conditioned modules - we found that this fixed representation was less vulnerable to forgetting by its simple design, similar to the our prompt component keys.", + "bbox": [ + 75, + 626, + 468, + 748 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.3. Expansion & Orthogonality", + "text_level": 1, + "bbox": [ + 76, + 756, + 326, + 773 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Crucial to mitigating catastrophic forgetting is the avoidance of overwriting knowledge acquired in prior tasks. When we visit a new task we freeze our current components and expand the set, only updating the new components. This is visualized in the bottom of Figure 2 where the existing parameters are locked and only the expanded parameters are optimized. Specifically, in task $t$ we learn $\\frac{M}{N}$ components where $N$ is the number of tasks and $M$ is a", + "bbox": [ + 75, + 779, + 468, + 904 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "chosen hyperparameter, keeping the previous $\\frac{(t - 1)\\cdot M}{N}$ components frozen. This expansion is enabled by our attention-based component-weighting scheme, in that expanding our parameters do not alter the computation of weights $\\alpha$ corresponding previously learned components. For example, expanding the capacity of a nonlinear model like an MLP module, transformer attention head [13], or hypernetwork [63] would alter the weight outputs of previously learned components.", + "bbox": [ + 496, + 404, + 890, + 541 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "While the prior heuristic is focused on reducing catastrophic forgetting (not destroying previously acquired knowledge), we also want to reduce interference between existing and new knowledge. To do this, we add an orthogonality constraint to $P$ , $K$ , and $A$ . The intuition is that orthogonal vectors have less interference with each other. For example, we would want a key and prompt learned in task 2 to not attract task 1 data, otherwise the encoder output for task 1 data will change, resulting in misclassified data. We introduce a simple orthogonality penalty loss to our method as:", + "bbox": [ + 496, + 542, + 892, + 705 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {o r t h o}} (B) = \\left| \\left| B B ^ {\\top} - I \\right| \\right| _ {2} \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 601, + 705, + 890, + 724 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $B$ represents any arbitrary matrix.", + "bbox": [ + 500, + 729, + 769, + 746 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.4. Full Optimization", + "text_level": 1, + "bbox": [ + 500, + 753, + 671, + 771 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Given task classification loss $\\mathcal{L}$ , our full optimization is:", + "bbox": [ + 517, + 777, + 890, + 792 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\min _ {\\boldsymbol {P} ^ {n}, \\boldsymbol {K} ^ {n}, \\boldsymbol {A} ^ {n}, \\phi_ {n}} \\mathcal {L} \\left(f _ {\\phi} \\left(f _ {\\theta , \\boldsymbol {P}, \\boldsymbol {K}, \\boldsymbol {A}} (\\boldsymbol {x})\\right), y\\right) + \\tag {7} \\\\ \\lambda \\left(\\mathcal {L} _ {\\text {o r t h o}} (\\boldsymbol {P}) + \\mathcal {L} _ {\\text {o r t h o}} (\\boldsymbol {K}) + \\mathcal {L} _ {\\text {o r t h o}} (\\boldsymbol {A})\\right) \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 524, + 816, + 890, + 859 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $P^n, K^n, A^n$ refer to the prompt components and corresponding keys/attention vectors that are unfrozen and", + "bbox": [ + 498, + 869, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "11913", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "trained during task $n$ (see Section 4.3) and $\\lambda$ is a hyperparameter balancing the orthogonality loss. Note that we do not update the logits of past task classes, consistent with L2P and DualPrompt. We formally refer to our method as COntinual Decomposed Attention-based Prompting, or simply CODA-P.", + "bbox": [ + 75, + 90, + 472, + 183 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 76, + 196, + 209, + 213 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We benchmark our approach with several image datasets in the class-incremental continual learning setting. We implemented baselines which do not store training data for rehearsal: Learning without Forgetting (LwF) [38], Learning to Prompt (L2P) [66], and DualPrompt [65]. Additionally, we report the upper bound performance (i.e., trained offline) and performance for a neural network trained only on classification loss using the new task training data (we refer to this as FT), and include an improved version of FT which uses the same classifier implementation as L2P/DualPrompt (referred to as $\\mathrm{FT} + +$ ). We also compare to Experience Replay [11] to provide additional context to our results. We implement our method and all baselines in PyTorch [48] using the ViT-B/16 backbone [13] pre-trained on ImageNet-1K [53]. Our contribution includes faithful PyTorch implementations of the popular prompting baselines L2P and DualPrompt, which were released in JAX [65, 66]. Our implementations of the competing methods actually achieve slight boosts in the performance of DualPrompt in most benchmarks, as well as significant performance boosts in L2P due to the improved prompting type (which we refer to as L2P++.", + "bbox": [ + 75, + 222, + 472, + 554 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "DualPrompt uses length 5 prompts in layers 1-2 (referred to as general prompts) and length 20 prompts in layers 3-5 (referred to as task-expert) prompts. We insert prompts into the same layers as DualPrompt (layers 1-5) for CODA-P and use a prompt length of 8 and 100 prompt components, which were chosen with a hyperparameter sweep on validation data to have the best trade-off between performance and parameter efficiency. Because our approach introduces more learnable parameters compared to DualPrompt, we include a variant of out method CODA-P-S which uses a smaller number of parameters equal to DualPrompt for additional comparison. We show that our method outperforms other methods even under this condition, while still retaining the capability to scale if desired.", + "bbox": [ + 75, + 555, + 472, + 767 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We report results on the test dataset, but emphasize that all hyperparameters and design decisions (including for the baselines) were made using validation data (20% of the training data). Unlike DualPrompt, we run our benchmarks for several different shuffles of the task class order and re", + "bbox": [ + 75, + 767, + 470, + 842 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "port the mean and standard deviation of these runs. We do this with a consistent seed (different for each trial) so that results can be directly compared. This is crucial because the results with different shuffling order can be different due to differences in task difficulty and their appearance order. We include additional implementation details and results in the Supplementary Material (SM).", + "bbox": [ + 500, + 90, + 890, + 196 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation Metrics: We evaluate methods using (1) average final accuracy $A_N$ , or the final accuracy with respect to all past classes averaged over $N$ tasks, and (2) average forgetting [9, 41, 42] $F_N$ , or the drop in task performance averaged over $N$ tasks. The reader is referred to our SM for additional details about these metrics. We emphasize that $A_N$ is the more important metric and encompasses both method plasticity and forgetting, whereas $F_N$ simply provides additional context.", + "bbox": [ + 498, + 198, + 893, + 335 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1. CODA-P sets the SOTA on existing benchmarks", + "text_level": 1, + "bbox": [ + 498, + 345, + 890, + 363 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We first evaluate our method and state of art on several well-established benchmarks. Table 1 provides results for ImageNet-R [21, 65] which is composed of 200 object classes with a wide collection of image styles, including cartoon, graffiti, and hard examples from the original ImageNet dataset [53]. This benchmark is attractive because the distribution of training data has significant distance to the pre-training data (ImageNet), thus providing a fair and challenging problem setting. In addition to the original 10-task benchmark, we also provide results with a smaller number of large tasks (5-task) and a larger number of small tasks (20 task). We first notice that our reported performance for DualPrompt is a few percentage points higher compared to the original DualPrompt [65] paper. We reimplemented the method from scratch in a different framework (PyTorch [48]) and suspect the difference has to do with our averaging over different class-order shuffling. We note that our implementation of L2P $(\\mathrm{L}2\\mathrm{P}++)$ performs significantly better than originally reported because we use the same form of prompting as DualPrompt (for fair comparison). Furthermore, our \"Deep L2P $+ + \"$ , which prompts at the same 5 layers as DualPrompt (rather than only at layer 1), actually performs similarly to DualPrompt. See supplementary for discussion and figures illustrating this.", + "bbox": [ + 496, + 371, + 890, + 734 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Importantly, we see that our method has strong gains in average accuracy across all three task lengths, with as much as $4.5\\%$ improvement in average accuracy over DualPrompt. We remind the reader that CODA-P is our proposed method with tuned prompt length and component set size, whereas CODA-P-S is downsized (see SM for exact details) to exactly match the number of learnable parameters as DualPrompt. We notice that our method often has slightly higher average forgetting compared to DualPrompt. Given that average accuracy is the crucial metric that captures practical performance, we", + "bbox": [ + 496, + 734, + 893, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_footnote", + "text": "See the SM for additional details", + "bbox": [ + 94, + 849, + 277, + 862 + ], + "page_idx": 5 + }, + { + "type": "page_footnote", + "text": "7As discussed in Section 3.2, we use pre-fix tuning over prompt-tuning for all implementations as is shown to work best for continual learning [65].", + "bbox": [ + 76, + 862, + 468, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "11914", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/18504fcbac243d3e7e92a242435b566eeef3b7d264cca8d558592f1cf28c6f18.jpg", + "table_caption": [ + "Table 1. Results (\\%) on ImageNet-R. Results are included for 5 tasks (40 classes per task), 10 tasks (20 classes per task), and 20 tasks (10 classes per task). $A_N$ gives the accuracy averaged over tasks and $F_N$ gives the average forgetting. We report results over 5 trials." + ], + "table_footnote": [], + "table_body": "
Tasks51020
MethodAN(↑)FN(↓)AN(↑)FN(↓)AN(↑)FN(↓)
UB77.13-77.13-77.13-
ER (5000)71.72 ± 0.7113.70 ± 0.2664.43 ± 1.1610.30 ± 0.0552.43 ± 0.877.70 ± 0.13
FT18.74 ± 0.4441.49 ± 0.5210.12 ± 0.5125.69 ± 0.234.75 ± 0.4016.34 ± 0.19
FT++60.42 ± 0.8714.66 ± 0.2448.93 ± 1.159.81 ± 0.3135.98 ± 1.386.63 ± 0.11
LwF.MC74.56 ± 0.594.98 ± 0.3766.73 ± 1.253.52 ± 0.3954.05 ± 2.662.86 ± 0.26
L2P++70.83 ± 0.583.36 ± 0.1869.29 ± 0.732.03 ± 0.1965.89 ± 1.301.24 ± 0.14
Deep L2P++73.93 ± 0.372.69 ± 0.1071.66 ± 0.641.78 ± 0.1668.42 ± 1.201.12 ± 0.13
DualPrompt73.05 ± 0.502.64 ± 0.1771.32 ± 0.621.71 ± 0.2467.87 ± 1.391.07 ± 0.14
CODA-P-S75.19 ± 0.472.65 ± 0.1573.93 ± 0.491.60 ± 0.2070.53 ± 1.241.00 ± 0.15
CODA-P76.51 ± 0.382.99 ± 0.1975.45 ± 0.561.64 ± 0.1072.37 ± 1.190.96 ± 0.15
", + "bbox": [ + 83, + 127, + 883, + 319 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/961f1c4dc6c1f2a18876ec0c70cc1e925e6bc79285b6e160ba38edadb2a69f09.jpg", + "table_caption": [ + "Table 2. Results (%) on CIFAR-100 and DomainNet. $A_N$ gives the accuracy averaged over tasks and $F_N$ gives the average forgetting. We report results over 5 trials for CIFAR-100 and 3 trials for DomainNet (due to dataset size).", + "(a) 10-task CIFAR-100 (10 classes per task)" + ], + "table_footnote": [], + "table_body": "
MethodAN(↑)FN(↓)
Upper-Bound89.30-
ER (5000)76.20 ± 1.048.50 ± 0.37
FT9.92 ± 0.2729.21 ± 0.18
FT++49.91 ± 0.4212.30 ± 0.23
LwF.MC64.83 ± 1.035.27 ± 0.39
L2P++82.50 ± 1.101.75 ± 0.42
Deep L2P++84.30 ± 1.031.53 ± 0.40
DualPrompt83.05 ± 1.161.72 ± 0.40
CODA-P-S84.59 ± 0.871.76 ± 0.28
CODA-P86.25 ± 0.741.67 ± 0.26
", + "bbox": [ + 107, + 375, + 450, + 549 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/522e3afaf015e8444d6031a17ccb291f0b0785b085e64b94f40b1a3e02bfd7a4.jpg", + "table_caption": [ + "(b) 5-task DomainNet (69 classes per task)" + ], + "table_footnote": [], + "table_body": "
MethodAN(↑)FN(↓)
Upper-Bound79.65-
ER (5000)58.32 ± 0.4726.25 ± 0.24
FT18.00 ± 0.2643.55 ± 0.27
FT++39.28 ± 0.2144.39 ± 0.31
LwF.MC74.78 ± 0.435.01 ± 0.14
L2P++69.58 ± 0.392.25 ± 0.08
Deep L2P++70.54 ± 0.512.05 ± 0.07
DualPrompt70.73 ± 0.492.03 ± 0.22
CODA-P-S71.80 ± 0.572.54 ± 0.10
CODA-P73.24 ± 0.593.46 ± 0.09
", + "bbox": [ + 519, + 375, + 862, + 549 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "are not worried by the slight uptick in forgetting. In fact, we argue it is very reasonable and reflects the strength of our approach: our method has a larger capacity to learn new tasks, and thus we can sacrifice marginally higher forgetting. Crucially, we see that as the task sequence length grows, the forgetting metric of our method versus DualPrompt begins to converge to a similar value.", + "bbox": [ + 75, + 573, + 468, + 679 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We provide results for experience replay to provide additional context of our results. While the gap between rehearsal with a substantial coreset size (i.e., the buffer size of replay data) and our approach is small for the short task sequence, we see that our method strongly outperforms rehearsal in the longer task sequence.", + "bbox": [ + 75, + 680, + 468, + 770 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Tables 2a and 2b provide results on additional8 benchmarks ten-task CIFAR-100 [32] and five-task DomainNet [49]. While neither dataset is as impactful as ImageNet-R in terms of challenge and distance from the pre-trained distribution, these tables provide additional", + "bbox": [ + 75, + 771, + 468, + 849 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "context to our method's performance. These tables illustrate a similar story to the ImageNet-R benchmarks, with improvements of $+3.2\\%$ and $+2.5\\%$ , respectively. We do note that, surprisingly, LwF [38] slightly outperforms prompting on this benchmark.", + "bbox": [ + 500, + 573, + 890, + 650 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.2. CODA-P sets the SOTA for a new dual-shift benchmark", + "text_level": 1, + "bbox": [ + 500, + 662, + 890, + 693 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We also evaluate on a challenging dual-shift benchmark using the ImageNet-R dataset. Our motivation is to show robustness to two different types of continual distribution shifts: semantic and covariate. We accomplish this by randomly selecting image types from the ImageNet-R dataset to include in each task's training data (while keeping the evaluation data unmodified). The reader is refereed to our SM for more details. Results for this benchmark are provided in Table 3. Compared to 5-task ImageNet-R where our method improves SOTA by $4.5\\%$ average accuracy, we see a $4.4\\%$ improvement on this 5-task benchmark with similar forgetting margins. Our results indicate that CODA-P better generalizes to real-world type shifts.", + "bbox": [ + 496, + 703, + 892, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_footnote", + "text": "Note that the hyperparameters for both DualPrompt and CODA-P were tuned on ImageNet-R, which was intentionally done to evaluate robustness of all methods.", + "bbox": [ + 76, + 862, + 468, + 898 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "11915", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/1bfc8d2b79099318087cab0529e71c891696e34d600913fe7173104b54262094.jpg", + "table_caption": [ + "Table 3. Results (\\%) on ImageNet-R with covariate domain shifts. Results are included for 5 tasks (40 classes per task). We simulate domain shifts by randomly removing $50\\%$ of the dataset's domains (e.g., clipart, paintings, and cartoon) for the training data of each task (see SM for more details). $A_N$ gives the accuracy averaged over tasks and $F_N$ gives the average forgetting. We report results over 5 trials." + ], + "table_footnote": [], + "table_body": "
MethodAN(↑)FN(↓)
Upper-Bound77.132.66
ER (5000)67.39 ± 0.3711.94 ± 0.17
FT17.93 ± 0.2737.49 ± 0.28
FT++54.51 ± 0.6814.41 ± 0.33
LwF.MC64.02 ± 1.557.05 ± 0.27
L2P++65.08 ± 0.292.79 ± 0.32
Deep L2P++65.74 ± 0.122.48 ± 0.30
DualPrompt66.98 ± 0.082.21 ± 0.28
CODA-P-S69.73 ± 0.182.35 ± 0.19
CODA-P71.35 ± 0.082.56 ± 0.26
", + "bbox": [ + 99, + 196, + 444, + 371 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/52f6d029241dc8683481d33ec1b55a79a7dc5411cab03fdfb6e03f51c3fb34fd.jpg", + "table_caption": [ + "Table 4. Ablation Results (\\%) on 10-task ImageNet-R. $A_N$ gives the accuracy averaged over tasks and $F_N$ gives the average forgetting. We report results over 5 trials." + ], + "table_footnote": [], + "table_body": "
MethodAN(↑)FN(↓)
CODA-P75.45 ± 0.561.64 ± 0.10
Ablate Attention74.52 ± 0.651.67 ± 0.13
Ablate Freezing74.60 ± 0.642.29 ± 0.10
Ablate Orthogonality70.66 ± 0.601.74 ± 0.25
", + "bbox": [ + 84, + 431, + 457, + 513 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.3. Ablations and Additional Analysis", + "text_level": 1, + "bbox": [ + 76, + 527, + 375, + 544 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We take a closer look at our method with ablation studies and additional analysis. In Table 4 we show the effect of removing a few key components of our method: attention keys, freezing of past task components, and our orthogonality regularization. We show that removing attention slightly reduces forgetting degrades performance on average accuracy (the more important metric). This makes sense because removing the attention keys makes our query processing more aligned with the L2P/DualPrompt methods which boast low forgetting but lack sufficient learning capacity. We see larger drops when removing freezing and orthogonality. This indicates that these aspects are crucial to our approach. Our intuition is that, without these, our prompt formation is similar to a shallow MLP module which, unregularized, should suffer from high forgetting.", + "bbox": [ + 75, + 551, + 468, + 779 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We also look at our ability to grow in model capacity along our newly introduced prompt component dimension using the 5-task ImageNet-R benchmark (with validation data). We show in Figure 3 that, with a number of prompt components equal to the number of prompts learned in DualPrompt, we achieve higher performance. Importantly, when we increase the number of components, we are able to leverage the scaled parameters for significantly higher", + "bbox": [ + 75, + 779, + 470, + 901 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/5833e8ab7aa279e3f28487b002243e4ec312317a46bc80612c4a1fca517349d6.jpg", + "image_caption": [ + "Figure 3. Analysis of average accuracy $A_N$ vs prompt components (or pool size) for CODA-P, L2P, and DualPrompt. The setting is 5 task ImageNet-R, and we report the mean over 3 trials." + ], + "image_footnote": [], + "bbox": [ + 535, + 92, + 851, + 212 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/f9367098ec6e4e37c56405c9a15b36d0517bd413b3844573b15769021f86a476.jpg", + "image_caption": [ + "Figure 4. Analysis of average accuracy $A_N$ vs prompt length for CODA-P, L2P, and DualPrompt. The setting is 5 task ImageNet-R, and we report the mean over 3 trials." + ], + "image_footnote": [], + "bbox": [ + 534, + 268, + 854, + 388 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "performance, finding that our method is closer to upper bound performance than it is to DualPrompt in average accuracy. Because the prompt pool in L2P can be increased to arbitrary size as well, we include results in this analysis as well. However, we show that the L2P method peaks with a pool size equal to twice the task sequence length, and then drops in performance. This reflects that our prompt components benefit from scale, whereas the existing prompt pool actually suffer.", + "bbox": [ + 496, + 453, + 890, + 589 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Finally, we show average accuracy versus prompt length in Figure 4 using the same setting as above. The purpose of this experiment is to emphasize that accuracy saturates with prompt length, thus motivating the need to expand in our introduced component dimension. Additional analysis and hyperparameter sweeps is available in our SM.", + "bbox": [ + 496, + 589, + 890, + 679 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 694, + 617, + 709 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We present CContinual decomposed attention-based prompting (CODA-Prompt) for rehearsal-free continual learning. Our approach assembles learnable prompt components which are then inserted into a pre-trained ViT encoder for image classification. Importantly, CODA-Prompt is optimized in an end-to-end fashion (unlike prior SOTA methods which involve two, separate optimizations). Furthermore, CODA-Prompt can scale prompting capacity to arbitrary sizes. We set a new SOTA on both well-established benchmarks and a dual-distribution shift benchmark (containing semantic and covariate shifts), highlighting the potential real world impact and generality of our approach.", + "bbox": [ + 496, + 719, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "11916", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Hongjoon Ahn, Jihwan Kwak, Subin Lim, Hyeonsu Bang, Hyojun Kim, and Taesup Moon. Ss-il: Separated softmax for incremental learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 844-853, October 2021. 2", + "[2] Rahaf Aljundi, Francesca Babiloni, Mohamed Elhoseiny, Marcus Rohrbach, and Tinne Tuytelaars. Memory aware synapses: Learning what (not) to forget. In ECCV, 2018. 2", + "[3] Rahaf Aljundi, Eugene Belilovsky, Tinne Tuytelaars, Laurent Charlin, Massimo Caccia, Min Lin, and Lucas Page-Caccia. Online continual learning with maximal interfered retrieval. In Advances in Neural Information Processing Systems, pages 11849-11860, 2019. 2", + "[4] Rahaf Aljundi, Min Lin, Baptiste Goujaud, and Yoshua Bengio. Gradient based sample selection for online continual learning. In Advances in Neural Information Processing Systems, pages 11816-11825, 2019. 2", + "[5] Jihwan Bang, Heesu Kim, YoungJoon Yoo, Jung-Woo Ha, and Jonghyun Choi. Rainbow memory: Continual learning with a memory of diverse samples. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8218-8227, 2021. 2", + "[6] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. Advances in neural information processing systems, 33:1877-1901, 2020. 3", + "[7] Pietro Buzzega, Matteo Boschini, Angelo Porrello, Davide Abati, and Simone Calderara. Dark experience for general continual learning: a strong, simple baseline. Advances in neural information processing systems, 33:15920-15930, 2020. 2", + "[8] Francisco M Castro, Manuel J Marín-Jiménez, Nicolas Guil, Cordelia Schmid, and Karteek Alahari. End-to-end incremental learning. In Proceedings of the European Conference on Computer Vision (ECCV), pages 233-248, 2018. 2", + "[9] Arslan Chaudhry, Marc' Aurelio Ranzato, Marcus Rohrbach, and Mohamed Elhoseiny. Efficient lifelong learning with a GEM. In International Conference on Learning Representations, 2019. 2, 6", + "[10] Arslan Chaudhry, Marcus Rohrbach, Mohamed Elhoseiny, Thalaiyasingam Ajanthan, Puneet K Dokania, Philip HS Torr, and Marc'Aurelio Ranzato. Continual learning with tiny episodic memories. arXiv preprint arXiv:1902.10486, 2019. 2", + "[11] Arslan Chaudhry, Marcus Rohrbach, Mohamed Elhoseiny, Thalaiyasingam Ajanthan, Puneet K Dokania, Philip HS Torr, and Marc'Aurelio Ranzato. On tiny episodic memories in continual learning. arXiv preprint arXiv:1902.10486, 2019. 6", + "[12] Yoojin Choi, Mostafa El-Khamy, and Jungwon Lee. Dualteacher class-incremental learning with data-free generative replay. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3543-3552, 2021. 2" + ], + "bbox": [ + 78, + 114, + 470, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[13] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 3, 5, 6", + "[14] Arthur Douillard, Alexandre Ramé, Guillaume Couairon, and Matthieu Cord. Dytox: Transformers for continual learning with dynamic token expansion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9285-9295, 2022. 3", + "[15] Sayna Ebrahimi, Mohamed Elhoseiny, Trevor Darrell, and Marcus Rohrbach. Uncertainty-guided continual learning with bayesian neural networks. arXiv preprint arXiv:1906.02425, 2019. 2", + "[16] Sayna Ebrahimi, Franziska Meier, Roberto Calandra, Trevor Darrell, and Marcus Rohrbach. Adversarial continual learning. arXiv preprint arXiv:2003.09553, 2020. 2", + "[17] Qiankun Gao, Chen Zhao, Bernard Ghanem, and Jian Zhang. R-dfcil: Relation-guided representation learning for data-free class incremental learning. arXiv preprint arXiv:2203.13104, 2022. 2", + "[18] Alexander Gepperth and Cem Karaoguz. Incremental learning with self-organizing maps. 2017 12th International Workshop on Self-Organizing Maps and Learning Vector Quantization, Clustering and Data Visualization (WSOM), pages 1–8, 2017. 2", + "[19] Tyler L Hayes, Nathan D Cahill, and Christopher Kanan. Memory efficient experience replay for streaming learning. In 2019 International Conference on Robotics and Automation (ICRA), pages 9769-9776. IEEE, 2019. 2", + "[20] Tyler L Hayes and Christopher Kanan. Lifelong machine learning with deep streaming linear discriminant analysis. arXiv preprint arXiv:1909.01520, 2019. 3", + "[21] Dan Hendrycks, Steven Basart, Norman Mu, Saurav Kadavath, Frank Wang, Evan Dorundo, Rahul Desai, Tyler Zhu, Samyak Parajuli, Mike Guo, Dawn Song, Jacob Steinhardt, and Justin Gilmer. The many faces of robustness: A critical analysis of out-of-distribution generalization. ICCV, 2021. 2, 6", + "[22] Geoffrey Hinton, Oriol Vinyals, and Jeff Dean. Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531, 2015. 2", + "[23] Saihui Hou, Xinyu Pan, Chen Change Loy, Zilei Wang, and Dahua Lin. Lifelong learning via progressive distillation and retrospection. In Proceedings of the European Conference on Computer Vision (ECCV), pages 437-452, 2018. 2", + "[24] Saihui Hou, Xinyu Pan, Chen Change Loy, Zilei Wang, and Dahua Lin. Learning a unified classifier incrementally via rebalancing. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 831-839, 2019. 2", + "[25] Yen-Chang Hsu, Yen-Cheng Liu, Anita Ramasamy, and Zsolt Kira. Re-evaluating continual learning scenarios: A categorization and case for strong baselines. arXiv preprint arXiv:1810.12488, 2018. 2, 3", + "[26] Daniel Justus, John Brennan, Stephen Bonner, and Andrew Stephen McGough. Predicting the computational cost" + ], + "bbox": [ + 501, + 92, + 893, + 901 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "11917", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "of deep learning models. In 2018 IEEE international conference on big data (Big Data), pages 3873-3882. IEEE, 2018. 1", + "[27] Nitin Kamra, Umang Gupta, and Yan Liu. Deep generative dual memory network for continual learning. arXiv preprint arXiv:1710.10368, 2017. 2", + "[28] Ronald Kemker and Christopher Kanan. Fearnet: Brain-inspired model for incremental learning. International Conference on Learning Representations (ICLR), 2018. 2", + "[29] Ronald Kemker, Marc McClure, Angelina Abitino, Tyler Hayes, and Christopher Kanan. Measuring catastrophic forgetting in neural networks. AAAI Conference on Artificial Intelligence, 2018. 2", + "[30] James Kirkpatrick, Razvan Pascanu, Neil Rabinowitz, Joel Veness, Guillaume Desjardins, Andrei A Rusu, Kieran Milan, John Quan, Tiago Ramalho, Agnieszka Grabska-Barwinska, et al. Overcoming catastrophic forgetting in neural networks. Proceedings of the national academy of sciences, 2017. 2", + "[31] Mario Köppen. The curse of dimensionality. In 5th online world conference on soft computing in industrial applications (WSC5), volume 1, pages 4-8, 2000. 4", + "[32] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. Tech Report, 2009. 2, 7", + "[33] Alexandre Lacoste, Alexandra Luccioni, Victor Schmidt, and Thomas Dandres. Quantifying the carbon emissions of machine learning. arXiv preprint arXiv:1910.09700, 2019. 1", + "[34] Kibok Lee, Kimin Lee, Jinwoo Shin, and Honglak Lee. Overcoming catastrophic forgetting with unlabeled data in the wild. In Proceedings of the IEEE International Conference on Computer Vision, pages 312-321, 2019. 2", + "[35] Soochan Lee, Junsoo Ha, Dongsu Zhang, and Gunhee Kim. A neural dirichlet process mixture model for task-free continual learning. arXiv preprint arXiv:2001.00689, 2020. 2", + "[36] Timothee Lesort, Hugo Caselles-Dupré, Michael Garcia-Ortiz, Andrei Stoian, and David Filliat. Generative models from the perspective of continual learning. In 2019 International Joint Conference on Neural Networks (IJCNN), pages 1–8. IEEE, 2019. 2", + "[37] Duo Li, Guimei Cao, Yunlu Xu, Zhanzhan Cheng, and Yi Niu. Technical report for iccv 2021 challenge sslad-track3b: Transformers are better continual learners. arXiv preprint arXiv:2201.04924, 2022. 3", + "[38] Zhizhong Li and Derek Hoiem. Learning without forgetting. IEEE transactions on pattern analysis and machine intelligence, 40(12):2935-2947, 2017. 2, 6, 7", + "[39] Vincenzo Lomonaco and Davide Maltoni. Core50: a new dataset and benchmark for continuous object recognition. arXiv preprint arXiv:1705.03550, 2017. 2", + "[40] Vincenzo Lomonaco, Davide Maltoni, and Lorenzo Pellegrini. Rehearsal-free continual learning over small non-iid batches. In CVPR Workshops, pages 989-998, 2020. 3", + "[41] David Lopez-Paz and Marc'Aurelio Ranzato. Gradient episodic memory for continual learning. In Proceedings of the 31st International Conference on Neural Information Processing Systems, NIPS'17, pages 6470-6479, USA, 2017. Curran Associates Inc. 2, 6" + ], + "bbox": [ + 78, + 90, + 468, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[42] Zheda Mai, Ruiwen Li, Jihwan Jeong, David Quispe, Hyunwoo Kim, and Scott Sanner. Online continual learning in image classification: An empirical survey. Neurocomputing, 469:28-51, 2022. 6", + "[43] Davide Maltoni and Vincenzo Lomonaco. Continuous learning in single-incremental-task scenarios. Neural Networks, 116:56-73, 2019. 2", + "[44] Seyed Iman Mirzadeh, Arslan Chaudhry, Dong Yin, Timothy Nguyen, Razvan Pascanu, Dilan Gorur, and Mehrdad Farajtabar. Architecture matters in continual learning. arXiv preprint arXiv:2202.00275, 2022. 3", + "[45] Vaishnavh Nagarajan, Colin Raffel, and Ian J Goodfellow. Theoretical insights into memorization in gans. In Neural Information Processing Systems Workshop, 2018. 2", + "[46] Cuong V Nguyen, Alessandro Achille, Michael Lam, Tal Hassner, Vijay Mahadevan, and Stefano Soatto. Toward understanding catastrophic forgetting in continual learning. arXiv preprint arXiv:1908.01091, 2019. 1", + "[47] Oleksiy Ostapenko, Mihai Puscas, Tassilo Klein, Patrick Jahnichen, and Moin Nabi. Learning to remember: A synaptic plasticity driven framework for continual learning. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 11321-11329, 2019. 2", + "[48] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. Advances in neural information processing systems, 32, 2019. 6", + "[49] Xingchao Peng, Qinxun Bai, Xide Xia, Zijun Huang, Kate Saenko, and Bo Wang. Moment matching for multi-source domain adaptation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 1406-1415, 2019. 7", + "[50] Sylvestre-Alvise Rebuffi, Alexander Kolesnikov, Georg Sperl, and Christoph H. Lampert. icarl: Incremental classifier and representation learning. In 2017 IEEE Conference on Computer Vision and Pattern Recognition, CVPR'17, pages 5533-5542, 2017. 2", + "[51] Anthony Robins. Catastrophic forgetting, rehearsal and pseudorehearsal. Connection Science, 7(2):123-146, 1995. 2", + "[52] David Rolnick, Arun Ahuja, Jonathan Schwarz, Timothy Lillicrap, and Gregory Wayne. Experience replay for continual learning. In Advances in Neural Information Processing Systems, pages 348-358, 2019. 2", + "[53] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, et al. Imagenet large scale visual recognition challenge. International journal of computer vision, 115(3):211-252, 2015. 6", + "[54] Andrei A Rusu, Neil C Rabinowitz, Guillaume Desjardins, Hubert Soyer, James Kirkpatrick, Koray Kavukcuoglu, Razvan Pascanu, and Raia Hadsell. Progressive neural networks. arXiv preprint arXiv:1606.04671, 2016. 2", + "[55] Hanul Shin, Jung Kwon Lee, Jaehong Kim, and Jiwon Kim. Continual learning with deep generative replay. In I. Guyon," + ], + "bbox": [ + 501, + 92, + 890, + 901 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "11918", + "bbox": [ + 480, + 945, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "U. V. Luxburg, S. Bengio, H. Wallach, R. Fergus, S. Vishwanathan, and R. Garnett, editors, Advances in Neural Information Processing Systems 30, pages 2990-2999. Curran Associates, Inc., 2017. 2", + "[56] James Smith, Yen-Chang Hsu, Jonathan Balloch, Yilin Shen, Hongxia Jin, and Zsolt Kira. Always be dreaming: A new approach for data-free class-incremental learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 9374-9384, October 2021. 2, 3", + "[57] James Seale Smith, Junjiao Tian, Yen-Chang Hsu, and Zsolt Kira. A closer look at rehearsal-free continual learning. arXiv preprint arXiv:2203.17269, 2022. 2", + "[58] Michalis K Titsias, Jonathan Schwarz, Alexander G de G Matthews, Razvan Pascanu, and Yee Whye Teh. Functional regularisation for continual learning with gaussian processes. In International Conference on Learning Representations, 2019. 2", + "[59] Gido M van de Ven, Hava T Siegelmann, and Andreas S Tolias. Brain-inspired replay for continual learning with artificial neural networks. Nature communications, 11(1):1-14, 2020. 2", + "[60] Gido M van de Ven and Andreas S Tolias. Generative replay with feedback connections as a general strategy for continual learning. arXiv preprint arXiv:1809.10635, 2018. 2", + "[61] Gido M Van de Ven and Andreas S Tolias. Three scenarios for continual learning. arXiv preprint arXiv:1904.07734, 2019. 2", + "[62] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 3", + "[63] Johannes von Oswald, Christian Henning, João Sacramento, and Benjamin F Grewe. Continual learning with hypernetworks. arXiv preprint arXiv:1906.00695, 2019. 2, 5", + "[64] Yabin Wang, Zhiwu Huang, and Xiaopeng Hong. S-prompts learning with pre-trained transformers: An occam's razor for domain incremental learning. In Conference on Neural Information Processing Systems (NeurIPS), 2022. 3", + "[65] Zifeng Wang, Zizhao Zhang, Sayna Ebrahimi, Ruoxi Sun, Han Zhang, Chen-Yu Lee, Xiaoqi Ren, Guolong Su, Vincent Perot, Jennifer Dy, et al. Dualprompt: Complementary prompting for rehearsal-free continual learning. arXiv preprint arXiv:2204.04799, 2022. 1, 2, 3, 4, 6", + "[66] Zifeng Wang, Zizhao Zhang, Chen-Yu Lee, Han Zhang, Ruoxi Sun, Xiaqi Ren, Guolong Su, Vincent Perot, Jennifer Dy, and Tomas Pfister. Learning to prompt for continual learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 139-149, 2022. 2, 3, 4, 6", + "[67] Yue Wu, Yinpeng Chen, Lijuan Wang, Yuancheng Ye, Zicheng Liu, Yandong Guo, and Yun Fu. Large scale incremental learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 374-382, 2019. 2, 3", + "[68] Hongxu Yin, Pavlo Molchanov, Jose M Alvarez, Zhizhong Li, Arun Mallya, Derek Hoiem, Niraj K Jha, and Jan Kautz." + ], + "bbox": [ + 78, + 90, + 468, + 900 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Dreaming to distill: Data-free knowledge transfer via deep-inversion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8715-8724, 2020. 2", + "[69] Wenpeng Yin, Jamaal Hay, and Dan Roth. Benchmarking zero-shot text classification: Datasets, evaluation and entailment approach. arXiv preprint arXiv:1909.00161, 2019. 3", + "[70] Pei Yu, Yinpeng Chen, Ying Jin, and Zicheng Liu. Improving vision transformers for incremental learning. arXiv preprint arXiv:2112.06103, 2021. 3", + "[71] Friedemann Zenke, Ben Poole, and Surya Ganguli. Continual learning through synaptic intelligence. In International Conference on Machine Learning, 2017. 2" + ], + "bbox": [ + 501, + 92, + 890, + 275 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11919", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2023/CODA-Prompt_ COntinual Decomposed Attention-Based Prompting for Rehearsal-Free Continual Learning/44360d49-0340-4846-917d-a266d277946e_model.json b/2023/CODA-Prompt_ COntinual Decomposed Attention-Based Prompting for Rehearsal-Free Continual Learning/44360d49-0340-4846-917d-a266d277946e_model.json new file mode 100644 index 0000000000000000000000000000000000000000..9fd1d898ae1a10bb9cdcb308fec54fdb39f71a3f --- /dev/null +++ b/2023/CODA-Prompt_ COntinual Decomposed Attention-Based Prompting for Rehearsal-Free Continual Learning/44360d49-0340-4846-917d-a266d277946e_model.json @@ -0,0 +1,2367 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.123, + 0.13, + 0.85, + 0.177 + ], + "angle": 0, + "content": "CODA-Prompt: CContinual Decomposed Attention-based Prompting for Rehearsal-Free Continual Learning" + }, + { + "type": "text", + "bbox": [ + 0.103, + 0.202, + 0.868, + 0.24 + ], + "angle": 0, + "content": "James Seale Smith\\(^{1,2}\\) Leonid Karlinsky\\(^{2,4}\\) Vyshnavi Gutta\\(^{1}\\) Paola Cascante-Bonilla\\(^{2,3}\\) Donghyun Kim\\(^{2,4}\\) Assaf Arbelle\\(^{4}\\) Rameswar Panda\\(^{2,4}\\) Rogerio Feris\\(^{2,4}\\) Zsolt Kira\\(^{1}\\)" + }, + { + "type": "text", + "bbox": [ + 0.151, + 0.24, + 0.815, + 0.256 + ], + "angle": 0, + "content": "1Georgia Institute of Technology 2MIT-IBM Watson AI Lab 3Rice University 4IBM Research" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.292, + 0.314, + 0.308 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.324, + 0.474, + 0.749 + ], + "angle": 0, + "content": "Computer vision models suffer from a phenomenon known as catastrophic forgetting when learning novel concepts from continuously shifting training data. Typical solutions for this continual learning problem require extensive rehearsal of previously seen data, which increases memory costs and may violate data privacy. Recently, the emergence of large-scale pre-trained vision transformer models has enabled prompting approaches as an alternative to data-rehearsal. These approaches rely on a key-query mechanism to generate prompts and have been found to be highly resistant to catastrophic forgetting in the well-established rehearsal-free continual learning setting. However, the key mechanism of these methods is not trained end-to-end with the task sequence. Our experiments show that this leads to a reduction in their plasticity, hence sacrificing new task accuracy, and inability to benefit from expanded parameter capacity. We instead propose to learn a set of prompt components which are assembled with input-conditioned weights to produce input-conditioned prompts, resulting in a novel attention-based end-to-end key-query scheme. Our experiments show that we outperform the current SOTA method DualPrompt on established benchmarks by as much as \\(4.5\\%\\) in average final accuracy. We also outperform the state of art by as much as \\(4.4\\%\\) accuracy on a continual learning benchmark which contains both class-incremental and domain-incremental task shifts, corresponding to many practical settings. Our code is available at https://github.com/GT-RIPL/CODA-Prompt" + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.777, + 0.21, + 0.792 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.795, + 0.471, + 0.903 + ], + "angle": 0, + "content": "For a computer vision model to succeed in the real world, it must overcome brittle assumptions that the concepts it will encounter after deployment will match those learned \\(a\\)-priori during training. The real world is indeed dynamic and contains continuously emerging objects and categories. Once deployed, models will encounter a range of differences from their training sets, requiring us to continuously" + }, + { + "type": "image", + "bbox": [ + 0.51, + 0.292, + 0.886, + 0.396 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.653, + 0.398, + 0.741, + 0.411 + ], + "angle": 0, + "content": "Prior Work" + }, + { + "type": "image", + "bbox": [ + 0.511, + 0.416, + 0.886, + 0.522 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.661, + 0.525, + 0.738, + 0.538 + ], + "angle": 0, + "content": "Our Work" + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.55, + 0.895, + 0.648 + ], + "angle": 0, + "content": "Figure 1. Prior work [65] learns a pool of key-value pairs to select learnable prompts which are inserted into several layers of a pretrained ViT (the prompting parameters are unique to each layer). Our work introduces a decomposed prompt which consists of learnable prompt components that assemble to produce attention-conditioned prompts. Unlike prior work, ours is optimized in an end-to-end fashion (denoted with thick, green lines)." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.673, + 0.856, + 0.688 + ], + "angle": 0, + "content": "update them to avoid stale and decaying performance." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.689, + 0.893, + 0.885 + ], + "angle": 0, + "content": "One way to update a model is to collect additional training data, combine this new training data with the old training data, and then retrain the model from scratch. While this will guarantee high performance, it is not practical for large-scale applications which may require lengthy training times and aggressive replacement timelines, such as models for self-driving cars or social media platforms. This could incur high financial [26] and environmental [33] costs if the replacement is frequent. We could instead update the model by only training on the new training data, but this leads to a phenomenon known as catastrophic forgetting [46] where the model overwrites existing knowledge when learning the new data, a problem known as continual learning." + }, + { + "type": "text", + "bbox": [ + 0.52, + 0.886, + 0.893, + 0.902 + ], + "angle": 0, + "content": "The most effective approaches proposed by the continual" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "11909" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.468, + 0.167 + ], + "angle": 0, + "content": "learning community involve saving [50] or generating [55] a subset of past training data and mixing it with future task data, a strategy referred to as rehearsal. Yet many important applications are unable to store this data because they work with private user data that cannot be stored long term." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.168, + 0.47, + 0.529 + ], + "angle": 0, + "content": "In this paper, we instead consider the highly-impactful and well-established setting of rehearsal-free continual learning [38, 56, 57, 65, 66], limiting our scope to strategies for continual learning which do not store training data. While rehearsal-free approaches have classically under-performed rehearsal-based approaches in this challenging setting by wide-margins [57], recent prompt-based approaches [65, 66], leveraging pre-trained Vision transformers (ViT), have had tremendous success and can even outperform state-of-the-art (SOTA) rehearsal-based methods. These prompting approaches boast a strong protection against catastrophic forgetting by learning a small pool of insert-able model embeddings (prompts) rather than modifying vision encoder parameters directly. One drawback of these approaches, however, is that they cannot be optimized in an end-to-end fashion as they use a key and query to select a prompt index from the pool of prompts, and thus rely on a second, localized optimization to learn the keys because the model gradient cannot backpropagate through the key/query index selection. Furthermore, these methods reduce forgetting by sacrificing new task accuracy (i.e., they lack sufficient plasticity). We show that expanding the prompt size does not increase the plasticity, motivating us to grow learning capacity from a new perspective." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.53, + 0.468, + 0.831 + ], + "angle": 0, + "content": "We replace the prompt pool with a decomposed prompt that consists of a weighted sum of learnable prompt components (Figure 1). This decomposition enables higher prompting capacity by expanding in a new dimension (the number of components). Furthermore, this inherently encourages knowledge re-use, as future task prompts will include contributions from prior task components. We also introduce a novel attention-based component-weighting scheme, which allows for our entire method to be optimized in an end-to-end fashion unlike the existing works, increasing our plasticity to better learn future tasks. We boast significant performance gains on existing rehearsal-free continual learning benchmarks w.r.t. SOTA prompting baselines in addition to a challenging dual-shift benchmark containing both incremental concept shifts and covariate domain shifts, highlighting our method's impact and generality. Our method improves results even under equivalent parameter counts, but importantly can scale performance by increasing capacity, unlike prior methods which quickly plateau. In summary, we make the following contributions:" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.892, + 0.182 + ], + "angle": 0, + "content": "1. We introduce a decomposed attention-based prompting for rehearsal-free continual learning, characterized by an expanded learning capacity compared to existing continual prompting approaches. Importantly, our approach can be optimized in an end-to-end fashion, unlike the prior SOTA." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.195, + 0.892, + 0.256 + ], + "angle": 0, + "content": "2. We establish a new SOTA for rehearsal-free continual learning on the well-established ImageNet-R [21, 65] and CIFAR-100 [32] benchmarks, beating the previous SOTA method DualPrompt [65] by as much as \\(4.5\\%\\)." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.268, + 0.892, + 0.344 + ], + "angle": 0, + "content": "3. We evaluate on a challenging benchmark with dual distribution shifts (semantic and covariate) using the ImageNet-R dataset, and again outperform the state of art, highlighting the real-world impact and generality of our approach." + }, + { + "type": "list", + "bbox": [ + 0.513, + 0.092, + 0.892, + 0.344 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.36, + 0.785, + 0.377 + ], + "angle": 0, + "content": "2. Background and Related Work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.386, + 0.892, + 0.627 + ], + "angle": 0, + "content": "Continual Learning: Continual learning approaches can be organized into a few broad categories which are all useful depending on the problem setting and constraints. One group of approaches expands a model's architecture as new tasks are encountered; these are highly effective for applications where a model growing with tasks is practical [16, 35, 39, 43, 54]. Another approach is to regularize the model with respect to past task knowledge while training the new task. This can either be done by regularizing the model in the weight space (i.e., penalize changes to model parameters) [2, 15, 30, 58, 71] or the prediction space (i.e., penalize changes to model predictions) [1, 8, 23, 34, 38]. Regularizing knowledge in the prediction space is done using knowledge distillation [22] and it has been found to perform better than model regularization-based methods for continual learning when task labels are not given [36, 60]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.629, + 0.892, + 0.84 + ], + "angle": 0, + "content": "Rehearsal with stored data [3-5, 7, 9, 10, 18, 19, 24, 29, 41, 50-52, 63, 67] or samples from a generative model [27, 28, 47, 55, 59] is highly effective when storing training data or training/saving a generative model is possible. Unfortunately, for many machine learning applications long-term storage of training data will violate data privacy, as well as incur a large memory cost. With respect to the generative model, this training process is much more computationally and memory intensive compared to a classification model and additionally may violate data legality concerns because using a generative model increases the chance of memorizing potentially sensitive data [45]. This motivates us to work on the important setting of rehearsal-free approaches to mitigate catastrophic forgetting." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Rehearsal-Free Continual Learning: Recent works propose producing images for rehearsal using deep-model inversion [12, 17, 56, 68]. While these methods perform well compared to generative modeling approaches and simply" + }, + { + "type": "page_footnote", + "bbox": [ + 0.076, + 0.839, + 0.468, + 0.901 + ], + "angle": 0, + "content": "1We focus on continual learning over a single, expanding classification head called class-incremental continual learning. This is different from the multi-task continual learning setting, known as task-incremental continual learning, where we learn separate classification heads for each task and the task label is provided during inference. [25, 61]" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "11910" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.242 + ], + "angle": 0, + "content": "rehearse from a small number of stored images, model-inversion is a slow process associated with high computational costs in the continual learning setting, and furthermore these methods under-perform rehearsal-based approaches by significant margins [56]. Other works have looked at rehearsal-free continual learning from an online \"streaming\" learning perspective using a frozen, pre-trained model [20, 40]. Because we allow our models to train to convergence on task data (as is common for continual learning [67]), our setting is very different." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.244, + 0.473, + 0.486 + ], + "angle": 0, + "content": "Continual Learning in Vision Transformers: Recent work [6, 37, 69] has proven transformers to generalize well to unseen domains. For example, one study varies the number of attention heads in Vision transformers and conclude that ViTs offer more robustness to forgetting with respect to CNN-based equivalents [44]. Another [70] shows that the vanilla counterparts of ViTs are more prone to forgetting when trained from scratch. Finally, DyTox [14] learns a unified model with a parameter-isolation approach and dynamically expands the tokens processed by the last layer to mitigate forgetting. For each task, they learn a new task-specific token per head using task-attention based decoder blocks. However, the above works either rely on exemplars to defy forgetting or they need to train a new transformer model from scratch, a costly operation, hence differing from our objective of exemplar-free CL using pre-trained ViTs." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.487, + 0.473, + 0.729 + ], + "angle": 0, + "content": "Prompting for Continual Learning: Prompt-based approaches for continual learning boast a strong protection against catastrophic forgetting by learning a small number of insert-able model instructions (prompts) rather than modifying encoder parameters directly. The current state-of-the-art approaches for our setting, DualPrompt [65] and L2P [66], create a pool of prompts to be selected for insertion into the model, matching input data to prompts without task id with a local clustering-like optimization. We build upon the foundation of these methods, as discussed later in this paper. We note that the recent S-Prompts [64] method is similar to these approaches but designed for domain-incremental learning (i.e., learning the same set of classes under covariate distribution task shifts), which is different than the class-incremental continual learning setting of our paper (i.e., learn emerging object classes in new tasks)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.745, + 0.217, + 0.76 + ], + "angle": 0, + "content": "3. Preliminaries" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.771, + 0.267, + 0.787 + ], + "angle": 0, + "content": "3.1. Continual Learning" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.795, + 0.47, + 0.902 + ], + "angle": 0, + "content": "In our continual learning setting, a model is sequentially shown \\(N\\) tasks corresponding to non-overlapping subsets of semantic object classes. Each class appears in only a single task, and the goal is to incrementally learn to classify new object classes as they are introduced while retaining performance on previously learned classes. To describe our models, we denote \\(\\theta\\) as a pre-trained vision encoder and \\(\\phi_{n}\\)" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.895, + 0.242 + ], + "angle": 0, + "content": "as our classifier head with logits corresponding to task \\( n \\) classes. In this paper, we deal with the class-incremental continual learning setting rather than the task-incremental continual learning setting. Class-incremental continual learning is challenging because the learner must support classification across all classes seen up to task \\( n \\) [25] (i.e., no task labels are provided to the learner during inference). Task-incremental continual learning is a simpler multi-task setting where the task labels are given during both training and inference." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.244, + 0.895, + 0.38 + ], + "angle": 0, + "content": "We also include experiments that contain both class-incremental and domain-incremental continual learning at the same time. The difference between this setting and the former is that we also inject covariate distribution shifts (e.g., task 1 might include real images and clip-art, whereas task 2 might include corrupted photos and artistic paintings). The motivation is to make continual learning more practical - as in the real world we might encounter these dual types of domain shifts as well." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.393, + 0.77, + 0.41 + ], + "angle": 0, + "content": "3.2. Prompting with Prefix-Tuning" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.418, + 0.894, + 0.538 + ], + "angle": 0, + "content": "In our work, we do not change the technical foundations of prompting from the prior state-of-the-art DualPrompt [65]. We instead focus on the selection and formation of the prompt (including enabling an end-to-end optimization of all prompting components), and the resulting prompt is used by the vision transformer encoder in an identical fashion to DualPrompt. This allows us to make a fair comparison with respect to our novel contributions." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.54, + 0.895, + 0.737 + ], + "angle": 0, + "content": "As done in DualPrompt, we pass prompts to several multi-head self-attention (MSA) layers in a pre-trained ViT transformer [13, 62] and use prefix-tuning over prompt-tuning\\(^2\\), which prepends prompts to the keys and values of an MSA layer rather than prepending prompts to the input tokens. The layers in which prompting occurs is set as a hyperparameter, and the prompting parameters are unique between layers. We define a prompt parameter as \\(\\pmb{p} \\in \\mathbb{R}^{L_p \\times D}\\) where \\(L_{p}\\) is the prompt length (chosen as a hyperparameter) and \\(D\\) is the embedding dimension (768). Consider an MSA layer with input \\(\\pmb{h} \\in \\mathbb{R}^{L \\times D}\\), and query, key, and values given as \\(\\pmb{h}_{Q}, \\pmb{h}_{K}, \\pmb{h}_{V}\\), respectively. In the ViT model, \\(\\pmb{h}_{Q} = \\pmb{h}_{K} = \\pmb{h}_{V} = \\pmb{h}\\). The output of this layer is given as:" + }, + { + "type": "equation", + "bbox": [ + 0.515, + 0.75, + 0.892, + 0.797 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\operatorname {M S A} \\left(\\boldsymbol {h} _ {Q}, \\boldsymbol {h} _ {K}, \\boldsymbol {h} _ {V}\\right) = \\operatorname {C o n c a t} \\left(\\mathrm {h} _ {1}, \\dots , \\mathrm {h} _ {\\mathrm {m}}\\right) W ^ {O} \\\\ \\text {w h e r e} \\mathrm {h} _ {\\mathrm {i}} = \\text {A t t e n t i o n} \\left(\\boldsymbol {h} _ {Q} W _ {i} ^ {Q}, \\boldsymbol {h} _ {K} W _ {i} ^ {K}, \\boldsymbol {h} _ {V} W _ {i} ^ {V}\\right) \\tag {1} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.81, + 0.894, + 0.862 + ], + "angle": 0, + "content": "where \\( W^{O}, W_{i}^{Q}, W_{i}^{K} \\), and \\( W_{i}^{V} \\) are projection matrices and \\( m \\) is the number of heads. We split our prompt \\( \\pmb{p} \\) into \\( \\{\\pmb{p}_K, \\pmb{p}_V\\} \\in \\mathbb{R}^{\\frac{L_p}{2} \\times D} \\) and prepend them to \\( h_K \\) and \\( h_V \\) with" + }, + { + "type": "page_footnote", + "bbox": [ + 0.5, + 0.875, + 0.894, + 0.901 + ], + "angle": 0, + "content": "2We refer the reader to sections 4.2 and 5.4 of the the DualPrompt [65] paper for a discussion on this design choice." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "11911" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.232, + 0.108 + ], + "angle": 0, + "content": "prefix-tuning (P-T) as:" + }, + { + "type": "equation", + "bbox": [ + 0.106, + 0.114, + 0.47, + 0.133 + ], + "angle": 0, + "content": "\\[\nf _ {P - T} (\\boldsymbol {p}, \\boldsymbol {h}) = \\operatorname {M S A} \\left(\\boldsymbol {h} _ {Q}, \\left[ \\boldsymbol {p} _ {K}; \\boldsymbol {h} _ {K} \\right], \\left[ \\boldsymbol {p} _ {V}; \\boldsymbol {h} _ {V} \\right]\\right) \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.139, + 0.47, + 0.231 + ], + "angle": 0, + "content": "The result is that we now only train a small number of parameter (the prompts) while leaving the rest of the transformer encoder unmodified. The critical question that now remains is how to select and update prompts in a continuous fashion? The next subsection will discuss how prior works select and update prompts." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.238, + 0.281, + 0.256 + ], + "angle": 0, + "content": "3.3. L2P and DualPrompt" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.261, + 0.473, + 0.519 + ], + "angle": 0, + "content": "L2P [66] and DualPrompt [65] select prompts from a pool using an image-wise prompt query. Specifically, these methods use a key-value pair based query strategy to dynamically select instance-specific prompts from a pool of candidate prompts. Each prompt \\( \\pmb{p}_m \\) is selected from a pool of learnable keys \\( \\pmb{k}_m \\in \\mathbb{R}^D \\) where \\( M \\) is the size of the prompt pool, based on cosine similarity to an input-conditioned query. Queries are produced as: \\( q(\\pmb{x}) \\in \\mathbb{R}^D = \\theta(\\pmb{x}) \\) where \\( \\theta \\) is the pretrained ViT encoder and \\( x \\) is the input image4. A query \\( q(\\pmb{x}) \\) is matched to a key \\( \\pmb{k}_m \\) by selecting the key with the highest cosine similarity, denoted as: \\( \\gamma(q(\\pmb{x}), \\pmb{k}_m) \\). The keys are learned with a separate optimization from the task classification loss by simply maximizing the cosine similarity between each matched \\( q(\\pmb{x}) \\) and \\( \\pmb{k}_m \\), acting as a clustering-like approach. However, this approach does not directly update the key with gradients from the task loss, which is critical in our experiments." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.531, + 0.17, + 0.545 + ], + "angle": 0, + "content": "4. Method" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.555, + 0.259, + 0.571 + ], + "angle": 0, + "content": "4.1. Prompt Formation" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.578, + 0.47, + 0.82 + ], + "angle": 0, + "content": "While prompting has been shown to perform exceptionally well in terms of mitigating catastrophic forgetting, the existing state-of-the-art approach DualPrompt lacks the ability to expand learning capacity within a given task. Specifically, DualPrompt learns a single prompt for each new task - regardless if the task is easy (such as learning a handful of new object classes) or complex (such as learning a large number of new object classes). One might try to increase the length of the prompt, but we show in our experiments (Section 5.3) that increasing the prompt length has saturated returns. Intuitively, we instead desire a method where the learning capacity is correlated to the underlying complexity of task data, rather than a single prompt. Additionally, we desire an approach that is end-to-end differentiable, which we conjecture increases the ability to learn a new task with higher accuracy." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.827, + 0.47, + 0.888 + ], + "angle": 0, + "content": "3We note that the difference between L2P and DualPrompt w.r.t. prompt querying is that the size of the prompt pool in L2P is chosen as a hyperparameter, but in DualPrompt it is equal to the number of tasks and, during training, DualPrompt selects the prompt using task-id (and selects the closest key-query match during inference)." + }, + { + "type": "text", + "bbox": [ + 0.095, + 0.888, + 0.312, + 0.9 + ], + "angle": 0, + "content": "4The output is read from the class token." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.257 + ], + "angle": 0, + "content": "We grow our learning capacity by introducing a new axis: a set of prompt components. Rather than pick and choose prompts from a pool, we learn a set of prompt components which, via a weighted summation, form a decomposed5 prompt that is passed to the corresponding MSA layer. This allows us to grow our prompting capacity to arbitrary depth and capture the rich underlying complexity of any task while maintaining a fixed prompt length. In addition, prompting in new tasks will inherently reuse the previously acquired knowledge of past tasks rather than initializing a new task prompt from scratch." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.258, + 0.894, + 0.289 + ], + "angle": 0, + "content": "Specifically, we replace learnable prompt parameter \\( \\pmb{p} \\) with a weighted summation over the prompt components:" + }, + { + "type": "equation", + "bbox": [ + 0.639, + 0.298, + 0.892, + 0.328 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {p} = \\sum_ {m} \\alpha_ {m} \\boldsymbol {P} _ {m} \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.339, + 0.892, + 0.414 + ], + "angle": 0, + "content": "where \\( P \\in \\mathbb{R}^{L_p \\times D \\times M} \\) is our set of prompt components, \\( M \\) is the length of our set (i.e., the introduced axis of additional capacity), and \\( \\alpha \\) is a weighting vector that determines the contribution of each component which will be discussed in the next subsection." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.416, + 0.892, + 0.52 + ], + "angle": 0, + "content": "A distinct difference between our method and L2P/DualPrompt is that all of our learnable parameters are optimized using the classification loss rather than splitting into two separate loss optimizations. This allows for better end-to-end learning in our method, and we argue this is a key reason why our method performs better in the benchmarks described in Section 5." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.53, + 0.776, + 0.547 + ], + "angle": 0, + "content": "4.2. Prompt-Component Weighting" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.553, + 0.892, + 0.629 + ], + "angle": 0, + "content": "Rather than a pool of prompts, we have a set of prompt components and desire to produce a weight vector \\(\\alpha\\) given a query \\(\\theta(x)\\), rather than a prompt index. We compute the cosine similarity between a query and keys to produce our weighting vector as:" + }, + { + "type": "equation", + "bbox": [ + 0.508, + 0.639, + 0.891, + 0.675 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\alpha = \\gamma (q (\\boldsymbol {x}), \\boldsymbol {K}) \\\\ = \\left\\{\\gamma (q (\\boldsymbol {x}), \\boldsymbol {K} _ {1}), \\gamma (q (\\boldsymbol {x}), \\boldsymbol {K} _ {2}), \\dots , \\gamma (q (\\boldsymbol {x}), \\boldsymbol {K} _ {M}) \\right\\} \\tag {4} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.685, + 0.892, + 0.761 + ], + "angle": 0, + "content": "where \\(\\pmb{K} \\in \\mathbb{R}^{D \\times M}\\) contains keys corresponding to our prompt components. The intuition here is that the contributing magnitude of each prompt component \\(\\pmb{P}_{m}\\) to the final prompt \\(\\pmb{p}\\) is determined by the similarity between the query \\(q(\\pmb{x})\\) and corresponding key \\(\\pmb{K}_{m}\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.762, + 0.892, + 0.867 + ], + "angle": 0, + "content": "The prompt-query matching can be thought of as clustering in a high dimension space \\(D\\) which is a well-known and difficult problem [31]. To mitigate this drawback, we introduce another component to our key-query matching: attention. Each \\(P_{m}\\) has a corresponding attention vector \\(A_{m}\\) in addition to the key \\(K_{m}\\). This allows the query to focus on certain features from the high dimensional query" + }, + { + "type": "page_footnote", + "bbox": [ + 0.5, + 0.875, + 0.892, + 0.901 + ], + "angle": 0, + "content": "5 Decomposed in the sense that our prompt is a weighted summation of components." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "11912" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.099, + 0.095, + 0.873, + 0.288 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.297, + 0.892, + 0.396 + ], + "angle": 0, + "content": "Figure 2. Our CODA-Prompt approach to rehearsal-free continual learning. An input (e.g., image) is passed through a query function (we use the pretrained ViT encoder) and used for a novel attention-based prompt-selection scheme to produce prompts which are passed to multiple layers of a large-scale, pre-trained transformer. Our method is parameterized by an expanding set of small prompt components, with each prompt having a corresponding key and attention vector. Both the input and produced prompts are passed through the transformer and sent to the task head. Only the prompting and task parameters are learned which is parameter efficient and no training data is stored for replay which is memory-efficient and privacy preserving. Importantly, our prompt selection scheme can be optimized end-to-end, unlike prior works which have a local similarity-based optimization for key-query matching." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.406, + 0.474, + 0.573 + ], + "angle": 0, + "content": "\\(q(\\pmb{x})\\) output - for example, a prompt component used for mammalian face features to classify dogs and cats would be able to focus on relevant query features such as eye placement and skin texture, and furthermore ignore features of unrelated semantics, such as features useful for automobile classification. We use a simple feature-selection attention scheme with element-wise multiplication between the query vector and attention vector to create an attended-query which is then used for the key-similarity matching. Specifically, our updated approach to producing our weighting vector is:" + }, + { + "type": "equation", + "bbox": [ + 0.085, + 0.582, + 0.469, + 0.617 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\begin{array}{l} \\alpha = \\gamma (q (\\boldsymbol {x}) \\odot \\boldsymbol {A}, \\boldsymbol {K}) \\\\ = \\left\\{\\gamma (q (\\boldsymbol {x}) \\odot \\boldsymbol {A}, \\boldsymbol {K}), \\quad \\gamma (q (\\boldsymbol {x}) \\odot \\boldsymbol {A}, \\boldsymbol {K}), \\right\\} \\end{array} \\tag {5} \\\\ = \\left\\{\\gamma (q (\\boldsymbol {x}) \\odot \\boldsymbol {A} _ {1}, \\boldsymbol {K} _ {1}), \\dots , \\gamma (q (\\boldsymbol {x}) \\odot \\boldsymbol {A} _ {M}, \\boldsymbol {K} _ {M}) \\right\\} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.627, + 0.47, + 0.749 + ], + "angle": 0, + "content": "where \\( \\mathbf{A} \\in \\mathbb{R}^{D \\times M} \\) contains learnable parameters (attention vectors) corresponding to our prompt components and \\( \\odot \\) is the element-wise product (also known as the Hadamard product). Notice that our attention vectors are learnable feature weightings rather than input-conditioned modules - we found that this fixed representation was less vulnerable to forgetting by its simple design, similar to the our prompt component keys." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.757, + 0.327, + 0.774 + ], + "angle": 0, + "content": "4.3. Expansion & Orthogonality" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.78, + 0.47, + 0.905 + ], + "angle": 0, + "content": "Crucial to mitigating catastrophic forgetting is the avoidance of overwriting knowledge acquired in prior tasks. When we visit a new task we freeze our current components and expand the set, only updating the new components. This is visualized in the bottom of Figure 2 where the existing parameters are locked and only the expanded parameters are optimized. Specifically, in task \\( t \\) we learn \\( \\frac{M}{N} \\) components where \\( N \\) is the number of tasks and \\( M \\) is a" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.405, + 0.892, + 0.542 + ], + "angle": 0, + "content": "chosen hyperparameter, keeping the previous \\(\\frac{(t - 1)\\cdot M}{N}\\) components frozen. This expansion is enabled by our attention-based component-weighting scheme, in that expanding our parameters do not alter the computation of weights \\(\\alpha\\) corresponding previously learned components. For example, expanding the capacity of a nonlinear model like an MLP module, transformer attention head [13], or hypernetwork [63] would alter the weight outputs of previously learned components." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.543, + 0.893, + 0.706 + ], + "angle": 0, + "content": "While the prior heuristic is focused on reducing catastrophic forgetting (not destroying previously acquired knowledge), we also want to reduce interference between existing and new knowledge. To do this, we add an orthogonality constraint to \\( P \\), \\( K \\), and \\( A \\). The intuition is that orthogonal vectors have less interference with each other. For example, we would want a key and prompt learned in task 2 to not attract task 1 data, otherwise the encoder output for task 1 data will change, resulting in misclassified data. We introduce a simple orthogonality penalty loss to our method as:" + }, + { + "type": "equation", + "bbox": [ + 0.602, + 0.707, + 0.892, + 0.726 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {o r t h o}} (B) = \\left| \\left| B B ^ {\\top} - I \\right| \\right| _ {2} \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.731, + 0.771, + 0.747 + ], + "angle": 0, + "content": "where \\(B\\) represents any arbitrary matrix." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.755, + 0.673, + 0.772 + ], + "angle": 0, + "content": "4.4. Full Optimization" + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.778, + 0.891, + 0.794 + ], + "angle": 0, + "content": "Given task classification loss \\(\\mathcal{L}\\), our full optimization is:" + }, + { + "type": "equation", + "bbox": [ + 0.526, + 0.817, + 0.891, + 0.861 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\min _ {\\boldsymbol {P} ^ {n}, \\boldsymbol {K} ^ {n}, \\boldsymbol {A} ^ {n}, \\phi_ {n}} \\mathcal {L} \\left(f _ {\\phi} \\left(f _ {\\theta , \\boldsymbol {P}, \\boldsymbol {K}, \\boldsymbol {A}} (\\boldsymbol {x})\\right), y\\right) + \\tag {7} \\\\ \\lambda \\left(\\mathcal {L} _ {\\text {o r t h o}} (\\boldsymbol {P}) + \\mathcal {L} _ {\\text {o r t h o}} (\\boldsymbol {K}) + \\mathcal {L} _ {\\text {o r t h o}} (\\boldsymbol {A})\\right) \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.871, + 0.892, + 0.901 + ], + "angle": 0, + "content": "where \\(P^n, K^n, A^n\\) refer to the prompt components and corresponding keys/attention vectors that are unfrozen and" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "11913" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.184 + ], + "angle": 0, + "content": "trained during task \\( n \\) (see Section 4.3) and \\( \\lambda \\) is a hyperparameter balancing the orthogonality loss. Note that we do not update the logits of past task classes, consistent with L2P and DualPrompt. We formally refer to our method as COntinual Decomposed Attention-based Prompting, or simply CODA-P." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.197, + 0.21, + 0.214 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.223, + 0.473, + 0.555 + ], + "angle": 0, + "content": "We benchmark our approach with several image datasets in the class-incremental continual learning setting. We implemented baselines which do not store training data for rehearsal: Learning without Forgetting (LwF) [38], Learning to Prompt (L2P) [66], and DualPrompt [65]. Additionally, we report the upper bound performance (i.e., trained offline) and performance for a neural network trained only on classification loss using the new task training data (we refer to this as FT), and include an improved version of FT which uses the same classifier implementation as L2P/DualPrompt (referred to as \\(\\mathrm{FT} + +\\)). We also compare to Experience Replay [11] to provide additional context to our results. We implement our method and all baselines in PyTorch [48] using the ViT-B/16 backbone [13] pre-trained on ImageNet-1K [53]. Our contribution includes faithful PyTorch implementations of the popular prompting baselines L2P and DualPrompt, which were released in JAX [65, 66]. Our implementations of the competing methods actually achieve slight boosts in the performance of DualPrompt in most benchmarks, as well as significant performance boosts in L2P due to the improved prompting type (which we refer to as L2P++." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.556, + 0.473, + 0.768 + ], + "angle": 0, + "content": "DualPrompt uses length 5 prompts in layers 1-2 (referred to as general prompts) and length 20 prompts in layers 3-5 (referred to as task-expert) prompts. We insert prompts into the same layers as DualPrompt (layers 1-5) for CODA-P and use a prompt length of 8 and 100 prompt components, which were chosen with a hyperparameter sweep on validation data to have the best trade-off between performance and parameter efficiency. Because our approach introduces more learnable parameters compared to DualPrompt, we include a variant of out method CODA-P-S which uses a smaller number of parameters equal to DualPrompt for additional comparison. We show that our method outperforms other methods even under this condition, while still retaining the capability to scale if desired." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.768, + 0.472, + 0.843 + ], + "angle": 0, + "content": "We report results on the test dataset, but emphasize that all hyperparameters and design decisions (including for the baselines) were made using validation data (20% of the training data). Unlike DualPrompt, we run our benchmarks for several different shuffles of the task class order and re" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.092, + 0.892, + 0.198 + ], + "angle": 0, + "content": "port the mean and standard deviation of these runs. We do this with a consistent seed (different for each trial) so that results can be directly compared. This is crucial because the results with different shuffling order can be different due to differences in task difficulty and their appearance order. We include additional implementation details and results in the Supplementary Material (SM)." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.199, + 0.895, + 0.336 + ], + "angle": 0, + "content": "Evaluation Metrics: We evaluate methods using (1) average final accuracy \\( A_N \\), or the final accuracy with respect to all past classes averaged over \\( N \\) tasks, and (2) average forgetting [9, 41, 42] \\( F_N \\), or the drop in task performance averaged over \\( N \\) tasks. The reader is referred to our SM for additional details about these metrics. We emphasize that \\( A_N \\) is the more important metric and encompasses both method plasticity and forgetting, whereas \\( F_N \\) simply provides additional context." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.347, + 0.892, + 0.364 + ], + "angle": 0, + "content": "5.1. CODA-P sets the SOTA on existing benchmarks" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.372, + 0.892, + 0.735 + ], + "angle": 0, + "content": "We first evaluate our method and state of art on several well-established benchmarks. Table 1 provides results for ImageNet-R [21, 65] which is composed of 200 object classes with a wide collection of image styles, including cartoon, graffiti, and hard examples from the original ImageNet dataset [53]. This benchmark is attractive because the distribution of training data has significant distance to the pre-training data (ImageNet), thus providing a fair and challenging problem setting. In addition to the original 10-task benchmark, we also provide results with a smaller number of large tasks (5-task) and a larger number of small tasks (20 task). We first notice that our reported performance for DualPrompt is a few percentage points higher compared to the original DualPrompt [65] paper. We reimplemented the method from scratch in a different framework (PyTorch [48]) and suspect the difference has to do with our averaging over different class-order shuffling. We note that our implementation of L2P \\((\\mathrm{L}2\\mathrm{P}++)\\) performs significantly better than originally reported because we use the same form of prompting as DualPrompt (for fair comparison). Furthermore, our \"Deep L2P \\(+ + \"\\), which prompts at the same 5 layers as DualPrompt (rather than only at layer 1), actually performs similarly to DualPrompt. See supplementary for discussion and figures illustrating this." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.736, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Importantly, we see that our method has strong gains in average accuracy across all three task lengths, with as much as \\(4.5\\%\\) improvement in average accuracy over DualPrompt. We remind the reader that CODA-P is our proposed method with tuned prompt length and component set size, whereas CODA-P-S is downsized (see SM for exact details) to exactly match the number of learnable parameters as DualPrompt. We notice that our method often has slightly higher average forgetting compared to DualPrompt. Given that average accuracy is the crucial metric that captures practical performance, we" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.851, + 0.279, + 0.863 + ], + "angle": 0, + "content": "See the SM for additional details" + }, + { + "type": "page_footnote", + "bbox": [ + 0.078, + 0.863, + 0.47, + 0.902 + ], + "angle": 0, + "content": "7As discussed in Section 3.2, we use pre-fix tuning over prompt-tuning for all implementations as is shown to work best for continual learning [65]." + }, + { + "type": "list", + "bbox": [ + 0.078, + 0.851, + 0.47, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "11914" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.089, + 0.892, + 0.118 + ], + "angle": 0, + "content": "Table 1. Results (\\%) on ImageNet-R. Results are included for 5 tasks (40 classes per task), 10 tasks (20 classes per task), and 20 tasks (10 classes per task). \\(A_N\\) gives the accuracy averaged over tasks and \\(F_N\\) gives the average forgetting. We report results over 5 trials." + }, + { + "type": "table", + "bbox": [ + 0.084, + 0.128, + 0.885, + 0.32 + ], + "angle": 0, + "content": "
Tasks51020
MethodAN(↑)FN(↓)AN(↑)FN(↓)AN(↑)FN(↓)
UB77.13-77.13-77.13-
ER (5000)71.72 ± 0.7113.70 ± 0.2664.43 ± 1.1610.30 ± 0.0552.43 ± 0.877.70 ± 0.13
FT18.74 ± 0.4441.49 ± 0.5210.12 ± 0.5125.69 ± 0.234.75 ± 0.4016.34 ± 0.19
FT++60.42 ± 0.8714.66 ± 0.2448.93 ± 1.159.81 ± 0.3135.98 ± 1.386.63 ± 0.11
LwF.MC74.56 ± 0.594.98 ± 0.3766.73 ± 1.253.52 ± 0.3954.05 ± 2.662.86 ± 0.26
L2P++70.83 ± 0.583.36 ± 0.1869.29 ± 0.732.03 ± 0.1965.89 ± 1.301.24 ± 0.14
Deep L2P++73.93 ± 0.372.69 ± 0.1071.66 ± 0.641.78 ± 0.1668.42 ± 1.201.12 ± 0.13
DualPrompt73.05 ± 0.502.64 ± 0.1771.32 ± 0.621.71 ± 0.2467.87 ± 1.391.07 ± 0.14
CODA-P-S75.19 ± 0.472.65 ± 0.1573.93 ± 0.491.60 ± 0.2070.53 ± 1.241.00 ± 0.15
CODA-P76.51 ± 0.382.99 ± 0.1975.45 ± 0.561.64 ± 0.1072.37 ± 1.190.96 ± 0.15
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.328, + 0.89, + 0.356 + ], + "angle": 0, + "content": "Table 2. Results (%) on CIFAR-100 and DomainNet. \\(A_N\\) gives the accuracy averaged over tasks and \\(F_N\\) gives the average forgetting. We report results over 5 trials for CIFAR-100 and 3 trials for DomainNet (due to dataset size)." + }, + { + "type": "table_caption", + "bbox": [ + 0.162, + 0.361, + 0.395, + 0.373 + ], + "angle": 0, + "content": "(a) 10-task CIFAR-100 (10 classes per task)" + }, + { + "type": "table", + "bbox": [ + 0.109, + 0.375, + 0.451, + 0.55 + ], + "angle": 0, + "content": "
MethodAN(↑)FN(↓)
Upper-Bound89.30-
ER (5000)76.20 ± 1.048.50 ± 0.37
FT9.92 ± 0.2729.21 ± 0.18
FT++49.91 ± 0.4212.30 ± 0.23
LwF.MC64.83 ± 1.035.27 ± 0.39
L2P++82.50 ± 1.101.75 ± 0.42
Deep L2P++84.30 ± 1.031.53 ± 0.40
DualPrompt83.05 ± 1.161.72 ± 0.40
CODA-P-S84.59 ± 0.871.76 ± 0.28
CODA-P86.25 ± 0.741.67 ± 0.26
" + }, + { + "type": "table_caption", + "bbox": [ + 0.578, + 0.361, + 0.805, + 0.373 + ], + "angle": 0, + "content": "(b) 5-task DomainNet (69 classes per task)" + }, + { + "type": "table", + "bbox": [ + 0.521, + 0.375, + 0.864, + 0.55 + ], + "angle": 0, + "content": "
MethodAN(↑)FN(↓)
Upper-Bound79.65-
ER (5000)58.32 ± 0.4726.25 ± 0.24
FT18.00 ± 0.2643.55 ± 0.27
FT++39.28 ± 0.2144.39 ± 0.31
LwF.MC74.78 ± 0.435.01 ± 0.14
L2P++69.58 ± 0.392.25 ± 0.08
Deep L2P++70.54 ± 0.512.05 ± 0.07
DualPrompt70.73 ± 0.492.03 ± 0.22
CODA-P-S71.80 ± 0.572.54 ± 0.10
CODA-P73.24 ± 0.593.46 ± 0.09
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.574, + 0.47, + 0.68 + ], + "angle": 0, + "content": "are not worried by the slight uptick in forgetting. In fact, we argue it is very reasonable and reflects the strength of our approach: our method has a larger capacity to learn new tasks, and thus we can sacrifice marginally higher forgetting. Crucially, we see that as the task sequence length grows, the forgetting metric of our method versus DualPrompt begins to converge to a similar value." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.681, + 0.469, + 0.771 + ], + "angle": 0, + "content": "We provide results for experience replay to provide additional context of our results. While the gap between rehearsal with a substantial coreset size (i.e., the buffer size of replay data) and our approach is small for the short task sequence, we see that our method strongly outperforms rehearsal in the longer task sequence." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.772, + 0.47, + 0.85 + ], + "angle": 0, + "content": "Tables 2a and 2b provide results on additional8 benchmarks ten-task CIFAR-100 [32] and five-task DomainNet [49]. While neither dataset is as impactful as ImageNet-R in terms of challenge and distance from the pre-trained distribution, these tables provide additional" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.574, + 0.892, + 0.651 + ], + "angle": 0, + "content": "context to our method's performance. These tables illustrate a similar story to the ImageNet-R benchmarks, with improvements of \\(+3.2\\%\\) and \\(+2.5\\%\\), respectively. We do note that, surprisingly, LwF [38] slightly outperforms prompting on this benchmark." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.664, + 0.892, + 0.694 + ], + "angle": 0, + "content": "5.2. CODA-P sets the SOTA for a new dual-shift benchmark" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.704, + 0.893, + 0.901 + ], + "angle": 0, + "content": "We also evaluate on a challenging dual-shift benchmark using the ImageNet-R dataset. Our motivation is to show robustness to two different types of continual distribution shifts: semantic and covariate. We accomplish this by randomly selecting image types from the ImageNet-R dataset to include in each task's training data (while keeping the evaluation data unmodified). The reader is refereed to our SM for more details. Results for this benchmark are provided in Table 3. Compared to 5-task ImageNet-R where our method improves SOTA by \\(4.5\\%\\) average accuracy, we see a \\(4.4\\%\\) improvement on this 5-task benchmark with similar forgetting margins. Our results indicate that CODA-P better generalizes to real-world type shifts." + }, + { + "type": "page_footnote", + "bbox": [ + 0.077, + 0.863, + 0.47, + 0.9 + ], + "angle": 0, + "content": "Note that the hyperparameters for both DualPrompt and CODA-P were tuned on ImageNet-R, which was intentionally done to evaluate robustness of all methods." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "11915" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.089, + 0.472, + 0.186 + ], + "angle": 0, + "content": "Table 3. Results (\\%) on ImageNet-R with covariate domain shifts. Results are included for 5 tasks (40 classes per task). We simulate domain shifts by randomly removing \\(50\\%\\) of the dataset's domains (e.g., clipart, paintings, and cartoon) for the training data of each task (see SM for more details). \\(A_N\\) gives the accuracy averaged over tasks and \\(F_N\\) gives the average forgetting. We report results over 5 trials." + }, + { + "type": "table", + "bbox": [ + 0.1, + 0.198, + 0.445, + 0.372 + ], + "angle": 0, + "content": "
MethodAN(↑)FN(↓)
Upper-Bound77.132.66
ER (5000)67.39 ± 0.3711.94 ± 0.17
FT17.93 ± 0.2737.49 ± 0.28
FT++54.51 ± 0.6814.41 ± 0.33
LwF.MC64.02 ± 1.557.05 ± 0.27
L2P++65.08 ± 0.292.79 ± 0.32
Deep L2P++65.74 ± 0.122.48 ± 0.30
DualPrompt66.98 ± 0.082.21 ± 0.28
CODA-P-S69.73 ± 0.182.35 ± 0.19
CODA-P71.35 ± 0.082.56 ± 0.26
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.378, + 0.47, + 0.421 + ], + "angle": 0, + "content": "Table 4. Ablation Results (\\%) on 10-task ImageNet-R. \\(A_N\\) gives the accuracy averaged over tasks and \\(F_N\\) gives the average forgetting. We report results over 5 trials." + }, + { + "type": "table", + "bbox": [ + 0.086, + 0.432, + 0.458, + 0.515 + ], + "angle": 0, + "content": "
MethodAN(↑)FN(↓)
CODA-P75.45 ± 0.561.64 ± 0.10
Ablate Attention74.52 ± 0.651.67 ± 0.13
Ablate Freezing74.60 ± 0.642.29 ± 0.10
Ablate Orthogonality70.66 ± 0.601.74 ± 0.25
" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.529, + 0.377, + 0.545 + ], + "angle": 0, + "content": "5.3. Ablations and Additional Analysis" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.553, + 0.47, + 0.78 + ], + "angle": 0, + "content": "We take a closer look at our method with ablation studies and additional analysis. In Table 4 we show the effect of removing a few key components of our method: attention keys, freezing of past task components, and our orthogonality regularization. We show that removing attention slightly reduces forgetting degrades performance on average accuracy (the more important metric). This makes sense because removing the attention keys makes our query processing more aligned with the L2P/DualPrompt methods which boast low forgetting but lack sufficient learning capacity. We see larger drops when removing freezing and orthogonality. This indicates that these aspects are crucial to our approach. Our intuition is that, without these, our prompt formation is similar to a shallow MLP module which, unregularized, should suffer from high forgetting." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.78, + 0.471, + 0.902 + ], + "angle": 0, + "content": "We also look at our ability to grow in model capacity along our newly introduced prompt component dimension using the 5-task ImageNet-R benchmark (with validation data). We show in Figure 3 that, with a number of prompt components equal to the number of prompts learned in DualPrompt, we achieve higher performance. Importantly, when we increase the number of components, we are able to leverage the scaled parameters for significantly higher" + }, + { + "type": "image", + "bbox": [ + 0.536, + 0.093, + 0.852, + 0.213 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.22, + 0.892, + 0.262 + ], + "angle": 0, + "content": "Figure 3. Analysis of average accuracy \\( A_N \\) vs prompt components (or pool size) for CODA-P, L2P, and DualPrompt. The setting is 5 task ImageNet-R, and we report the mean over 3 trials." + }, + { + "type": "image", + "bbox": [ + 0.535, + 0.269, + 0.855, + 0.389 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.396, + 0.892, + 0.438 + ], + "angle": 0, + "content": "Figure 4. Analysis of average accuracy \\( A_N \\) vs prompt length for CODA-P, L2P, and DualPrompt. The setting is 5 task ImageNet-R, and we report the mean over 3 trials." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.454, + 0.892, + 0.59 + ], + "angle": 0, + "content": "performance, finding that our method is closer to upper bound performance than it is to DualPrompt in average accuracy. Because the prompt pool in L2P can be increased to arbitrary size as well, we include results in this analysis as well. However, we show that the L2P method peaks with a pool size equal to twice the task sequence length, and then drops in performance. This reflects that our prompt components benefit from scale, whereas the existing prompt pool actually suffer." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.59, + 0.892, + 0.68 + ], + "angle": 0, + "content": "Finally, we show average accuracy versus prompt length in Figure 4 using the same setting as above. The purpose of this experiment is to emphasize that accuracy saturates with prompt length, thus motivating the need to expand in our introduced component dimension. Additional analysis and hyperparameter sweeps is available in our SM." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.695, + 0.618, + 0.71 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.892, + 0.901 + ], + "angle": 0, + "content": "We present CContinual decomposed attention-based prompting (CODA-Prompt) for rehearsal-free continual learning. Our approach assembles learnable prompt components which are then inserted into a pre-trained ViT encoder for image classification. Importantly, CODA-Prompt is optimized in an end-to-end fashion (unlike prior SOTA methods which involve two, separate optimizations). Furthermore, CODA-Prompt can scale prompting capacity to arbitrary sizes. We set a new SOTA on both well-established benchmarks and a dual-distribution shift benchmark (containing semantic and covariate shifts), highlighting the potential real world impact and generality of our approach." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "11916" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.184 + ], + "angle": 0, + "content": "[1] Hongjoon Ahn, Jihwan Kwak, Subin Lim, Hyeonsu Bang, Hyojun Kim, and Taesup Moon. Ss-il: Separated softmax for incremental learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 844-853, October 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.185, + 0.472, + 0.24 + ], + "angle": 0, + "content": "[2] Rahaf Aljundi, Francesca Babiloni, Mohamed Elhoseiny, Marcus Rohrbach, and Tinne Tuytelaars. Memory aware synapses: Learning what (not) to forget. In ECCV, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.242, + 0.471, + 0.311 + ], + "angle": 0, + "content": "[3] Rahaf Aljundi, Eugene Belilovsky, Tinne Tuytelaars, Laurent Charlin, Massimo Caccia, Min Lin, and Lucas Page-Caccia. Online continual learning with maximal interfered retrieval. In Advances in Neural Information Processing Systems, pages 11849-11860, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.312, + 0.471, + 0.368 + ], + "angle": 0, + "content": "[4] Rahaf Aljundi, Min Lin, Baptiste Goujaud, and Yoshua Bengio. Gradient based sample selection for online continual learning. In Advances in Neural Information Processing Systems, pages 11816-11825, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.368, + 0.472, + 0.437 + ], + "angle": 0, + "content": "[5] Jihwan Bang, Heesu Kim, YoungJoon Yoo, Jung-Woo Ha, and Jonghyun Choi. Rainbow memory: Continual learning with a memory of diverse samples. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8218-8227, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.438, + 0.471, + 0.508 + ], + "angle": 0, + "content": "[6] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. Advances in neural information processing systems, 33:1877-1901, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.509, + 0.471, + 0.577 + ], + "angle": 0, + "content": "[7] Pietro Buzzega, Matteo Boschini, Angelo Porrello, Davide Abati, and Simone Calderara. Dark experience for general continual learning: a strong, simple baseline. Advances in neural information processing systems, 33:15920-15930, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.579, + 0.471, + 0.635 + ], + "angle": 0, + "content": "[8] Francisco M Castro, Manuel J Marín-Jiménez, Nicolas Guil, Cordelia Schmid, and Karteek Alahari. End-to-end incremental learning. In Proceedings of the European Conference on Computer Vision (ECCV), pages 233-248, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.635, + 0.471, + 0.69 + ], + "angle": 0, + "content": "[9] Arslan Chaudhry, Marc' Aurelio Ranzato, Marcus Rohrbach, and Mohamed Elhoseiny. Efficient lifelong learning with a GEM. In International Conference on Learning Representations, 2019. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.691, + 0.47, + 0.759 + ], + "angle": 0, + "content": "[10] Arslan Chaudhry, Marcus Rohrbach, Mohamed Elhoseiny, Thalaiyasingam Ajanthan, Puneet K Dokania, Philip HS Torr, and Marc'Aurelio Ranzato. Continual learning with tiny episodic memories. arXiv preprint arXiv:1902.10486, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.761, + 0.47, + 0.83 + ], + "angle": 0, + "content": "[11] Arslan Chaudhry, Marcus Rohrbach, Mohamed Elhoseiny, Thalaiyasingam Ajanthan, Puneet K Dokania, Philip HS Torr, and Marc'Aurelio Ranzato. On tiny episodic memories in continual learning. arXiv preprint arXiv:1902.10486, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.831, + 0.47, + 0.9 + ], + "angle": 0, + "content": "[12] Yoojin Choi, Mostafa El-Khamy, and Jungwon Lee. Dualteacher class-incremental learning with data-free generative replay. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3543-3552, 2021. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.472, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.175 + ], + "angle": 0, + "content": "[13] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.177, + 0.895, + 0.245 + ], + "angle": 0, + "content": "[14] Arthur Douillard, Alexandre Ramé, Guillaume Couairon, and Matthieu Cord. Dytox: Transformers for continual learning with dynamic token expansion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9285-9295, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.246, + 0.894, + 0.3 + ], + "angle": 0, + "content": "[15] Sayna Ebrahimi, Mohamed Elhoseiny, Trevor Darrell, and Marcus Rohrbach. Uncertainty-guided continual learning with bayesian neural networks. arXiv preprint arXiv:1906.02425, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.301, + 0.894, + 0.343 + ], + "angle": 0, + "content": "[16] Sayna Ebrahimi, Franziska Meier, Roberto Calandra, Trevor Darrell, and Marcus Rohrbach. Adversarial continual learning. arXiv preprint arXiv:2003.09553, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.344, + 0.894, + 0.398 + ], + "angle": 0, + "content": "[17] Qiankun Gao, Chen Zhao, Bernard Ghanem, and Jian Zhang. R-dfcil: Relation-guided representation learning for data-free class incremental learning. arXiv preprint arXiv:2203.13104, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.4, + 0.894, + 0.469 + ], + "angle": 0, + "content": "[18] Alexander Gepperth and Cem Karaoguz. Incremental learning with self-organizing maps. 2017 12th International Workshop on Self-Organizing Maps and Learning Vector Quantization, Clustering and Data Visualization (WSOM), pages 1–8, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.469, + 0.893, + 0.525 + ], + "angle": 0, + "content": "[19] Tyler L Hayes, Nathan D Cahill, and Christopher Kanan. Memory efficient experience replay for streaming learning. In 2019 International Conference on Robotics and Automation (ICRA), pages 9769-9776. IEEE, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.525, + 0.893, + 0.567 + ], + "angle": 0, + "content": "[20] Tyler L Hayes and Christopher Kanan. Lifelong machine learning with deep streaming linear discriminant analysis. arXiv preprint arXiv:1909.01520, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.567, + 0.893, + 0.649 + ], + "angle": 0, + "content": "[21] Dan Hendrycks, Steven Basart, Norman Mu, Saurav Kadavath, Frank Wang, Evan Dorundo, Rahul Desai, Tyler Zhu, Samyak Parajuli, Mike Guo, Dawn Song, Jacob Steinhardt, and Justin Gilmer. The many faces of robustness: A critical analysis of out-of-distribution generalization. ICCV, 2021. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.65, + 0.894, + 0.691 + ], + "angle": 0, + "content": "[22] Geoffrey Hinton, Oriol Vinyals, and Jeff Dean. Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.692, + 0.894, + 0.748 + ], + "angle": 0, + "content": "[23] Saihui Hou, Xinyu Pan, Chen Change Loy, Zilei Wang, and Dahua Lin. Lifelong learning via progressive distillation and retrospection. In Proceedings of the European Conference on Computer Vision (ECCV), pages 437-452, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.748, + 0.893, + 0.816 + ], + "angle": 0, + "content": "[24] Saihui Hou, Xinyu Pan, Chen Change Loy, Zilei Wang, and Dahua Lin. Learning a unified classifier incrementally via rebalancing. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 831-839, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.817, + 0.893, + 0.872 + ], + "angle": 0, + "content": "[25] Yen-Chang Hsu, Yen-Cheng Liu, Anita Ramasamy, and Zsolt Kira. Re-evaluating continual learning scenarios: A categorization and case for strong baselines. arXiv preprint arXiv:1810.12488, 2018. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.873, + 0.893, + 0.902 + ], + "angle": 0, + "content": "[26] Daniel Justus, John Brennan, Stephen Bonner, and Andrew Stephen McGough. Predicting the computational cost" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.895, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "11917" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.092, + 0.47, + 0.134 + ], + "angle": 0, + "content": "of deep learning models. In 2018 IEEE international conference on big data (Big Data), pages 3873-3882. IEEE, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.135, + 0.47, + 0.176 + ], + "angle": 0, + "content": "[27] Nitin Kamra, Umang Gupta, and Yan Liu. Deep generative dual memory network for continual learning. arXiv preprint arXiv:1710.10368, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.177, + 0.469, + 0.218 + ], + "angle": 0, + "content": "[28] Ronald Kemker and Christopher Kanan. Fearnet: Brain-inspired model for incremental learning. International Conference on Learning Representations (ICLR), 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.219, + 0.469, + 0.273 + ], + "angle": 0, + "content": "[29] Ronald Kemker, Marc McClure, Angelina Abitino, Tyler Hayes, and Christopher Kanan. Measuring catastrophic forgetting in neural networks. AAAI Conference on Artificial Intelligence, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.274, + 0.469, + 0.356 + ], + "angle": 0, + "content": "[30] James Kirkpatrick, Razvan Pascanu, Neil Rabinowitz, Joel Veness, Guillaume Desjardins, Andrei A Rusu, Kieran Milan, John Quan, Tiago Ramalho, Agnieszka Grabska-Barwinska, et al. Overcoming catastrophic forgetting in neural networks. Proceedings of the national academy of sciences, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.358, + 0.469, + 0.399 + ], + "angle": 0, + "content": "[31] Mario Köppen. The curse of dimensionality. In 5th online world conference on soft computing in industrial applications (WSC5), volume 1, pages 4-8, 2000. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.4, + 0.469, + 0.427 + ], + "angle": 0, + "content": "[32] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. Tech Report, 2009. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.428, + 0.469, + 0.481 + ], + "angle": 0, + "content": "[33] Alexandre Lacoste, Alexandra Luccioni, Victor Schmidt, and Thomas Dandres. Quantifying the carbon emissions of machine learning. arXiv preprint arXiv:1910.09700, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.483, + 0.469, + 0.538 + ], + "angle": 0, + "content": "[34] Kibok Lee, Kimin Lee, Jinwoo Shin, and Honglak Lee. Overcoming catastrophic forgetting with unlabeled data in the wild. In Proceedings of the IEEE International Conference on Computer Vision, pages 312-321, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.539, + 0.469, + 0.58 + ], + "angle": 0, + "content": "[35] Soochan Lee, Junsoo Ha, Dongsu Zhang, and Gunhee Kim. A neural dirichlet process mixture model for task-free continual learning. arXiv preprint arXiv:2001.00689, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.581, + 0.469, + 0.649 + ], + "angle": 0, + "content": "[36] Timothee Lesort, Hugo Caselles-Dupré, Michael Garcia-Ortiz, Andrei Stoian, and David Filliat. Generative models from the perspective of continual learning. In 2019 International Joint Conference on Neural Networks (IJCNN), pages 1–8. IEEE, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.65, + 0.469, + 0.704 + ], + "angle": 0, + "content": "[37] Duo Li, Guimei Cao, Yunlu Xu, Zhanzhan Cheng, and Yi Niu. Technical report for iccv 2021 challenge sslad-track3b: Transformers are better continual learners. arXiv preprint arXiv:2201.04924, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.706, + 0.469, + 0.747 + ], + "angle": 0, + "content": "[38] Zhizhong Li and Derek Hoiem. Learning without forgetting. IEEE transactions on pattern analysis and machine intelligence, 40(12):2935-2947, 2017. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.748, + 0.469, + 0.789 + ], + "angle": 0, + "content": "[39] Vincenzo Lomonaco and Davide Maltoni. Core50: a new dataset and benchmark for continuous object recognition. arXiv preprint arXiv:1705.03550, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.79, + 0.469, + 0.831 + ], + "angle": 0, + "content": "[40] Vincenzo Lomonaco, Davide Maltoni, and Lorenzo Pellegrini. Rehearsal-free continual learning over small non-iid batches. In CVPR Workshops, pages 989-998, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.832, + 0.469, + 0.9 + ], + "angle": 0, + "content": "[41] David Lopez-Paz and Marc'Aurelio Ranzato. Gradient episodic memory for continual learning. In Proceedings of the 31st International Conference on Neural Information Processing Systems, NIPS'17, pages 6470-6479, USA, 2017. Curran Associates Inc. 2, 6" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.147 + ], + "angle": 0, + "content": "[42] Zheda Mai, Ruiwen Li, Jihwan Jeong, David Quispe, Hyunwoo Kim, and Scott Sanner. Online continual learning in image classification: An empirical survey. Neurocomputing, 469:28-51, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.149, + 0.892, + 0.19 + ], + "angle": 0, + "content": "[43] Davide Maltoni and Vincenzo Lomonaco. Continuous learning in single-incremental-task scenarios. Neural Networks, 116:56-73, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.192, + 0.892, + 0.248 + ], + "angle": 0, + "content": "[44] Seyed Iman Mirzadeh, Arslan Chaudhry, Dong Yin, Timothy Nguyen, Razvan Pascanu, Dilan Gorur, and Mehrdad Farajtabar. Architecture matters in continual learning. arXiv preprint arXiv:2202.00275, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.249, + 0.892, + 0.291 + ], + "angle": 0, + "content": "[45] Vaishnavh Nagarajan, Colin Raffel, and Ian J Goodfellow. Theoretical insights into memorization in gans. In Neural Information Processing Systems Workshop, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.293, + 0.892, + 0.347 + ], + "angle": 0, + "content": "[46] Cuong V Nguyen, Alessandro Achille, Michael Lam, Tal Hassner, Vijay Mahadevan, and Stefano Soatto. Toward understanding catastrophic forgetting in continual learning. arXiv preprint arXiv:1908.01091, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.349, + 0.892, + 0.419 + ], + "angle": 0, + "content": "[47] Oleksiy Ostapenko, Mihai Puscas, Tassilo Klein, Patrick Jahnichen, and Moin Nabi. Learning to remember: A synaptic plasticity driven framework for continual learning. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 11321-11329, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.42, + 0.892, + 0.502 + ], + "angle": 0, + "content": "[48] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. Advances in neural information processing systems, 32, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.504, + 0.892, + 0.572 + ], + "angle": 0, + "content": "[49] Xingchao Peng, Qinxun Bai, Xide Xia, Zijun Huang, Kate Saenko, and Bo Wang. Moment matching for multi-source domain adaptation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 1406-1415, 2019. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.575, + 0.892, + 0.643 + ], + "angle": 0, + "content": "[50] Sylvestre-Alvise Rebuffi, Alexander Kolesnikov, Georg Sperl, and Christoph H. Lampert. icarl: Incremental classifier and representation learning. In 2017 IEEE Conference on Computer Vision and Pattern Recognition, CVPR'17, pages 5533-5542, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.645, + 0.892, + 0.686 + ], + "angle": 0, + "content": "[51] Anthony Robins. Catastrophic forgetting, rehearsal and pseudorehearsal. Connection Science, 7(2):123-146, 1995. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.688, + 0.892, + 0.744 + ], + "angle": 0, + "content": "[52] David Rolnick, Arun Ahuja, Jonathan Schwarz, Timothy Lillicrap, and Gregory Wayne. Experience replay for continual learning. In Advances in Neural Information Processing Systems, pages 348-358, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.745, + 0.892, + 0.815 + ], + "angle": 0, + "content": "[53] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, et al. Imagenet large scale visual recognition challenge. International journal of computer vision, 115(3):211-252, 2015. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.817, + 0.892, + 0.871 + ], + "angle": 0, + "content": "[54] Andrei A Rusu, Neil C Rabinowitz, Guillaume Desjardins, Hubert Soyer, James Kirkpatrick, Koray Kavukcuoglu, Razvan Pascanu, and Raia Hadsell. Progressive neural networks. arXiv preprint arXiv:1606.04671, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.873, + 0.892, + 0.902 + ], + "angle": 0, + "content": "[55] Hanul Shin, Jung Kwon Lee, Jaehong Kim, and Jiwon Kim. Continual learning with deep generative replay. In I. Guyon," + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "11918" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.092, + 0.47, + 0.147 + ], + "angle": 0, + "content": "U. V. Luxburg, S. Bengio, H. Wallach, R. Fergus, S. Vishwanathan, and R. Garnett, editors, Advances in Neural Information Processing Systems 30, pages 2990-2999. Curran Associates, Inc., 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.15, + 0.47, + 0.232 + ], + "angle": 0, + "content": "[56] James Smith, Yen-Chang Hsu, Jonathan Balloch, Yilin Shen, Hongxia Jin, and Zsolt Kira. Always be dreaming: A new approach for data-free class-incremental learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 9374-9384, October 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.234, + 0.469, + 0.275 + ], + "angle": 0, + "content": "[57] James Seale Smith, Junjiao Tian, Yen-Chang Hsu, and Zsolt Kira. A closer look at rehearsal-free continual learning. arXiv preprint arXiv:2203.17269, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.278, + 0.469, + 0.345 + ], + "angle": 0, + "content": "[58] Michalis K Titsias, Jonathan Schwarz, Alexander G de G Matthews, Razvan Pascanu, and Yee Whye Teh. Functional regularisation for continual learning with gaussian processes. In International Conference on Learning Representations, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.348, + 0.469, + 0.401 + ], + "angle": 0, + "content": "[59] Gido M van de Ven, Hava T Siegelmann, and Andreas S Tolias. Brain-inspired replay for continual learning with artificial neural networks. Nature communications, 11(1):1-14, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.405, + 0.469, + 0.446 + ], + "angle": 0, + "content": "[60] Gido M van de Ven and Andreas S Tolias. Generative replay with feedback connections as a general strategy for continual learning. arXiv preprint arXiv:1809.10635, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.448, + 0.469, + 0.488 + ], + "angle": 0, + "content": "[61] Gido M Van de Ven and Andreas S Tolias. Three scenarios for continual learning. arXiv preprint arXiv:1904.07734, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.491, + 0.469, + 0.546 + ], + "angle": 0, + "content": "[62] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.548, + 0.469, + 0.589 + ], + "angle": 0, + "content": "[63] Johannes von Oswald, Christian Henning, João Sacramento, and Benjamin F Grewe. Continual learning with hypernetworks. arXiv preprint arXiv:1906.00695, 2019. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.591, + 0.469, + 0.646 + ], + "angle": 0, + "content": "[64] Yabin Wang, Zhiwu Huang, and Xiaopeng Hong. S-prompts learning with pre-trained transformers: An occam's razor for domain incremental learning. In Conference on Neural Information Processing Systems (NeurIPS), 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.648, + 0.469, + 0.716 + ], + "angle": 0, + "content": "[65] Zifeng Wang, Zizhao Zhang, Sayna Ebrahimi, Ruoxi Sun, Han Zhang, Chen-Yu Lee, Xiaoqi Ren, Guolong Su, Vincent Perot, Jennifer Dy, et al. Dualprompt: Complementary prompting for rehearsal-free continual learning. arXiv preprint arXiv:2204.04799, 2022. 1, 2, 3, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.718, + 0.469, + 0.8 + ], + "angle": 0, + "content": "[66] Zifeng Wang, Zizhao Zhang, Chen-Yu Lee, Han Zhang, Ruoxi Sun, Xiaqi Ren, Guolong Su, Vincent Perot, Jennifer Dy, and Tomas Pfister. Learning to prompt for continual learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 139-149, 2022. 2, 3, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.803, + 0.469, + 0.87 + ], + "angle": 0, + "content": "[67] Yue Wu, Yinpeng Chen, Lijuan Wang, Yuancheng Ye, Zicheng Liu, Yandong Guo, and Yun Fu. Large scale incremental learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 374-382, 2019. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[68] Hongxu Yin, Pavlo Molchanov, Jose M Alvarez, Zhizhong Li, Arun Mallya, Derek Hoiem, Niraj K Jha, and Jan Kautz." + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.147 + ], + "angle": 0, + "content": "Dreaming to distill: Data-free knowledge transfer via deep-inversion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8715-8724, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.149, + 0.892, + 0.19 + ], + "angle": 0, + "content": "[69] Wenpeng Yin, Jamaal Hay, and Dan Roth. Benchmarking zero-shot text classification: Datasets, evaluation and entailment approach. arXiv preprint arXiv:1909.00161, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.193, + 0.892, + 0.233 + ], + "angle": 0, + "content": "[70] Pei Yu, Yinpeng Chen, Ying Jin, and Zicheng Liu. Improving vision transformers for incremental learning. arXiv preprint arXiv:2112.06103, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.235, + 0.892, + 0.276 + ], + "angle": 0, + "content": "[71] Friedemann Zenke, Ben Poole, and Surya Ganguli. Continual learning through synaptic intelligence. In International Conference on Machine Learning, 2017. 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.276 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "11919" + } + ] +] \ No newline at end of file diff --git a/2023/CODA-Prompt_ COntinual Decomposed Attention-Based Prompting for Rehearsal-Free Continual Learning/44360d49-0340-4846-917d-a266d277946e_origin.pdf b/2023/CODA-Prompt_ COntinual Decomposed Attention-Based Prompting for Rehearsal-Free Continual Learning/44360d49-0340-4846-917d-a266d277946e_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..10bb6d3d526d472ed3184a05a03178cc33f2cd18 --- /dev/null +++ b/2023/CODA-Prompt_ COntinual Decomposed Attention-Based Prompting for Rehearsal-Free Continual Learning/44360d49-0340-4846-917d-a266d277946e_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:91a111093c125607aa0dacc70006091faaa6f10f48ee31ef03e513025697f4ba +size 598433 diff --git a/2023/CODA-Prompt_ COntinual Decomposed Attention-Based Prompting for Rehearsal-Free Continual Learning/full.md b/2023/CODA-Prompt_ COntinual Decomposed Attention-Based Prompting for Rehearsal-Free Continual Learning/full.md new file mode 100644 index 0000000000000000000000000000000000000000..36064521ffc1558f772332f683cac13e80954589 --- /dev/null +++ b/2023/CODA-Prompt_ COntinual Decomposed Attention-Based Prompting for Rehearsal-Free Continual Learning/full.md @@ -0,0 +1,308 @@ +# CODA-Prompt: CContinual Decomposed Attention-based Prompting for Rehearsal-Free Continual Learning + +James Seale Smith $^{1,2}$ Leonid Karlinsky $^{2,4}$ Vyshnavi Gutta $^{1}$ Paola Cascante-Bonilla $^{2,3}$ Donghyun Kim $^{2,4}$ Assaf Arbelle $^{4}$ Rameswar Panda $^{2,4}$ Rogerio Feris $^{2,4}$ Zsolt Kira $^{1}$ + +1Georgia Institute of Technology 2MIT-IBM Watson AI Lab 3Rice University 4IBM Research + +# Abstract + +Computer vision models suffer from a phenomenon known as catastrophic forgetting when learning novel concepts from continuously shifting training data. Typical solutions for this continual learning problem require extensive rehearsal of previously seen data, which increases memory costs and may violate data privacy. Recently, the emergence of large-scale pre-trained vision transformer models has enabled prompting approaches as an alternative to data-rehearsal. These approaches rely on a key-query mechanism to generate prompts and have been found to be highly resistant to catastrophic forgetting in the well-established rehearsal-free continual learning setting. However, the key mechanism of these methods is not trained end-to-end with the task sequence. Our experiments show that this leads to a reduction in their plasticity, hence sacrificing new task accuracy, and inability to benefit from expanded parameter capacity. We instead propose to learn a set of prompt components which are assembled with input-conditioned weights to produce input-conditioned prompts, resulting in a novel attention-based end-to-end key-query scheme. Our experiments show that we outperform the current SOTA method DualPrompt on established benchmarks by as much as $4.5\%$ in average final accuracy. We also outperform the state of art by as much as $4.4\%$ accuracy on a continual learning benchmark which contains both class-incremental and domain-incremental task shifts, corresponding to many practical settings. Our code is available at https://github.com/GT-RIPL/CODA-Prompt + +# 1. Introduction + +For a computer vision model to succeed in the real world, it must overcome brittle assumptions that the concepts it will encounter after deployment will match those learned $a$ -priori during training. The real world is indeed dynamic and contains continuously emerging objects and categories. Once deployed, models will encounter a range of differences from their training sets, requiring us to continuously + +![](images/9fc104dd3bfbb342f5ac031223ecc62e911b4c3f4618e42f7592753a4f003667.jpg) +Prior Work + +![](images/377f964a93e17e9799fa6fdfaab3d0f5dde02db371cebdffb0392499006b34ce.jpg) +Our Work +Figure 1. Prior work [65] learns a pool of key-value pairs to select learnable prompts which are inserted into several layers of a pretrained ViT (the prompting parameters are unique to each layer). Our work introduces a decomposed prompt which consists of learnable prompt components that assemble to produce attention-conditioned prompts. Unlike prior work, ours is optimized in an end-to-end fashion (denoted with thick, green lines). + +update them to avoid stale and decaying performance. + +One way to update a model is to collect additional training data, combine this new training data with the old training data, and then retrain the model from scratch. While this will guarantee high performance, it is not practical for large-scale applications which may require lengthy training times and aggressive replacement timelines, such as models for self-driving cars or social media platforms. This could incur high financial [26] and environmental [33] costs if the replacement is frequent. We could instead update the model by only training on the new training data, but this leads to a phenomenon known as catastrophic forgetting [46] where the model overwrites existing knowledge when learning the new data, a problem known as continual learning. + +The most effective approaches proposed by the continual + +learning community involve saving [50] or generating [55] a subset of past training data and mixing it with future task data, a strategy referred to as rehearsal. Yet many important applications are unable to store this data because they work with private user data that cannot be stored long term. + +In this paper, we instead consider the highly-impactful and well-established setting of rehearsal-free continual learning [38, 56, 57, 65, 66], limiting our scope to strategies for continual learning which do not store training data. While rehearsal-free approaches have classically under-performed rehearsal-based approaches in this challenging setting by wide-margins [57], recent prompt-based approaches [65, 66], leveraging pre-trained Vision transformers (ViT), have had tremendous success and can even outperform state-of-the-art (SOTA) rehearsal-based methods. These prompting approaches boast a strong protection against catastrophic forgetting by learning a small pool of insert-able model embeddings (prompts) rather than modifying vision encoder parameters directly. One drawback of these approaches, however, is that they cannot be optimized in an end-to-end fashion as they use a key and query to select a prompt index from the pool of prompts, and thus rely on a second, localized optimization to learn the keys because the model gradient cannot backpropagate through the key/query index selection. Furthermore, these methods reduce forgetting by sacrificing new task accuracy (i.e., they lack sufficient plasticity). We show that expanding the prompt size does not increase the plasticity, motivating us to grow learning capacity from a new perspective. + +We replace the prompt pool with a decomposed prompt that consists of a weighted sum of learnable prompt components (Figure 1). This decomposition enables higher prompting capacity by expanding in a new dimension (the number of components). Furthermore, this inherently encourages knowledge re-use, as future task prompts will include contributions from prior task components. We also introduce a novel attention-based component-weighting scheme, which allows for our entire method to be optimized in an end-to-end fashion unlike the existing works, increasing our plasticity to better learn future tasks. We boast significant performance gains on existing rehearsal-free continual learning benchmarks w.r.t. SOTA prompting baselines in addition to a challenging dual-shift benchmark containing both incremental concept shifts and covariate domain shifts, highlighting our method's impact and generality. Our method improves results even under equivalent parameter counts, but importantly can scale performance by increasing capacity, unlike prior methods which quickly plateau. In summary, we make the following contributions: + +1. We introduce a decomposed attention-based prompting for rehearsal-free continual learning, characterized by an expanded learning capacity compared to existing continual prompting approaches. Importantly, our approach can be optimized in an end-to-end fashion, unlike the prior SOTA. +2. We establish a new SOTA for rehearsal-free continual learning on the well-established ImageNet-R [21, 65] and CIFAR-100 [32] benchmarks, beating the previous SOTA method DualPrompt [65] by as much as $4.5\%$ . +3. We evaluate on a challenging benchmark with dual distribution shifts (semantic and covariate) using the ImageNet-R dataset, and again outperform the state of art, highlighting the real-world impact and generality of our approach. + +# 2. Background and Related Work + +Continual Learning: Continual learning approaches can be organized into a few broad categories which are all useful depending on the problem setting and constraints. One group of approaches expands a model's architecture as new tasks are encountered; these are highly effective for applications where a model growing with tasks is practical [16, 35, 39, 43, 54]. Another approach is to regularize the model with respect to past task knowledge while training the new task. This can either be done by regularizing the model in the weight space (i.e., penalize changes to model parameters) [2, 15, 30, 58, 71] or the prediction space (i.e., penalize changes to model predictions) [1, 8, 23, 34, 38]. Regularizing knowledge in the prediction space is done using knowledge distillation [22] and it has been found to perform better than model regularization-based methods for continual learning when task labels are not given [36, 60]. + +Rehearsal with stored data [3-5, 7, 9, 10, 18, 19, 24, 29, 41, 50-52, 63, 67] or samples from a generative model [27, 28, 47, 55, 59] is highly effective when storing training data or training/saving a generative model is possible. Unfortunately, for many machine learning applications long-term storage of training data will violate data privacy, as well as incur a large memory cost. With respect to the generative model, this training process is much more computationally and memory intensive compared to a classification model and additionally may violate data legality concerns because using a generative model increases the chance of memorizing potentially sensitive data [45]. This motivates us to work on the important setting of rehearsal-free approaches to mitigate catastrophic forgetting. + +Rehearsal-Free Continual Learning: Recent works propose producing images for rehearsal using deep-model inversion [12, 17, 56, 68]. While these methods perform well compared to generative modeling approaches and simply + +rehearse from a small number of stored images, model-inversion is a slow process associated with high computational costs in the continual learning setting, and furthermore these methods under-perform rehearsal-based approaches by significant margins [56]. Other works have looked at rehearsal-free continual learning from an online "streaming" learning perspective using a frozen, pre-trained model [20, 40]. Because we allow our models to train to convergence on task data (as is common for continual learning [67]), our setting is very different. + +Continual Learning in Vision Transformers: Recent work [6, 37, 69] has proven transformers to generalize well to unseen domains. For example, one study varies the number of attention heads in Vision transformers and conclude that ViTs offer more robustness to forgetting with respect to CNN-based equivalents [44]. Another [70] shows that the vanilla counterparts of ViTs are more prone to forgetting when trained from scratch. Finally, DyTox [14] learns a unified model with a parameter-isolation approach and dynamically expands the tokens processed by the last layer to mitigate forgetting. For each task, they learn a new task-specific token per head using task-attention based decoder blocks. However, the above works either rely on exemplars to defy forgetting or they need to train a new transformer model from scratch, a costly operation, hence differing from our objective of exemplar-free CL using pre-trained ViTs. + +Prompting for Continual Learning: Prompt-based approaches for continual learning boast a strong protection against catastrophic forgetting by learning a small number of insert-able model instructions (prompts) rather than modifying encoder parameters directly. The current state-of-the-art approaches for our setting, DualPrompt [65] and L2P [66], create a pool of prompts to be selected for insertion into the model, matching input data to prompts without task id with a local clustering-like optimization. We build upon the foundation of these methods, as discussed later in this paper. We note that the recent S-Prompts [64] method is similar to these approaches but designed for domain-incremental learning (i.e., learning the same set of classes under covariate distribution task shifts), which is different than the class-incremental continual learning setting of our paper (i.e., learn emerging object classes in new tasks). + +# 3. Preliminaries + +# 3.1. Continual Learning + +In our continual learning setting, a model is sequentially shown $N$ tasks corresponding to non-overlapping subsets of semantic object classes. Each class appears in only a single task, and the goal is to incrementally learn to classify new object classes as they are introduced while retaining performance on previously learned classes. To describe our models, we denote $\theta$ as a pre-trained vision encoder and $\phi_{n}$ + +as our classifier head with logits corresponding to task $n$ classes. In this paper, we deal with the class-incremental continual learning setting rather than the task-incremental continual learning setting. Class-incremental continual learning is challenging because the learner must support classification across all classes seen up to task $n$ [25] (i.e., no task labels are provided to the learner during inference). Task-incremental continual learning is a simpler multi-task setting where the task labels are given during both training and inference. + +We also include experiments that contain both class-incremental and domain-incremental continual learning at the same time. The difference between this setting and the former is that we also inject covariate distribution shifts (e.g., task 1 might include real images and clip-art, whereas task 2 might include corrupted photos and artistic paintings). The motivation is to make continual learning more practical - as in the real world we might encounter these dual types of domain shifts as well. + +# 3.2. Prompting with Prefix-Tuning + +In our work, we do not change the technical foundations of prompting from the prior state-of-the-art DualPrompt [65]. We instead focus on the selection and formation of the prompt (including enabling an end-to-end optimization of all prompting components), and the resulting prompt is used by the vision transformer encoder in an identical fashion to DualPrompt. This allows us to make a fair comparison with respect to our novel contributions. + +As done in DualPrompt, we pass prompts to several multi-head self-attention (MSA) layers in a pre-trained ViT transformer [13, 62] and use prefix-tuning over prompt-tuning $^2$ , which prepends prompts to the keys and values of an MSA layer rather than prepending prompts to the input tokens. The layers in which prompting occurs is set as a hyperparameter, and the prompting parameters are unique between layers. We define a prompt parameter as $\pmb{p} \in \mathbb{R}^{L_p \times D}$ where $L_{p}$ is the prompt length (chosen as a hyperparameter) and $D$ is the embedding dimension (768). Consider an MSA layer with input $\pmb{h} \in \mathbb{R}^{L \times D}$ , and query, key, and values given as $\pmb{h}_{Q}, \pmb{h}_{K}, \pmb{h}_{V}$ , respectively. In the ViT model, $\pmb{h}_{Q} = \pmb{h}_{K} = \pmb{h}_{V} = \pmb{h}$ . The output of this layer is given as: + +$$ +\begin{array}{l} \operatorname {M S A} \left(\boldsymbol {h} _ {Q}, \boldsymbol {h} _ {K}, \boldsymbol {h} _ {V}\right) = \operatorname {C o n c a t} \left(\mathrm {h} _ {1}, \dots , \mathrm {h} _ {\mathrm {m}}\right) W ^ {O} \\ \text {w h e r e} \mathrm {h} _ {\mathrm {i}} = \text {A t t e n t i o n} \left(\boldsymbol {h} _ {Q} W _ {i} ^ {Q}, \boldsymbol {h} _ {K} W _ {i} ^ {K}, \boldsymbol {h} _ {V} W _ {i} ^ {V}\right) \tag {1} \\ \end{array} +$$ + +where $W^{O}, W_{i}^{Q}, W_{i}^{K}$ , and $W_{i}^{V}$ are projection matrices and $m$ is the number of heads. We split our prompt $\pmb{p}$ into $\{\pmb{p}_K, \pmb{p}_V\} \in \mathbb{R}^{\frac{L_p}{2} \times D}$ and prepend them to $h_K$ and $h_V$ with + +prefix-tuning (P-T) as: + +$$ +f _ {P - T} (\boldsymbol {p}, \boldsymbol {h}) = \operatorname {M S A} \left(\boldsymbol {h} _ {Q}, \left[ \boldsymbol {p} _ {K}; \boldsymbol {h} _ {K} \right], \left[ \boldsymbol {p} _ {V}; \boldsymbol {h} _ {V} \right]\right) \tag {2} +$$ + +The result is that we now only train a small number of parameter (the prompts) while leaving the rest of the transformer encoder unmodified. The critical question that now remains is how to select and update prompts in a continuous fashion? The next subsection will discuss how prior works select and update prompts. + +# 3.3. L2P and DualPrompt + +L2P [66] and DualPrompt [65] select prompts from a pool using an image-wise prompt query. Specifically, these methods use a key-value pair based query strategy to dynamically select instance-specific prompts from a pool of candidate prompts. Each prompt $\pmb{p}_m$ is selected from a pool of learnable keys $\pmb{k}_m \in \mathbb{R}^D$ where $M$ is the size of the prompt pool, based on cosine similarity to an input-conditioned query. Queries are produced as: $q(\pmb{x}) \in \mathbb{R}^D = \theta(\pmb{x})$ where $\theta$ is the pretrained ViT encoder and $x$ is the input image4. A query $q(\pmb{x})$ is matched to a key $\pmb{k}_m$ by selecting the key with the highest cosine similarity, denoted as: $\gamma(q(\pmb{x}), \pmb{k}_m)$ . The keys are learned with a separate optimization from the task classification loss by simply maximizing the cosine similarity between each matched $q(\pmb{x})$ and $\pmb{k}_m$ , acting as a clustering-like approach. However, this approach does not directly update the key with gradients from the task loss, which is critical in our experiments. + +# 4. Method + +# 4.1. Prompt Formation + +While prompting has been shown to perform exceptionally well in terms of mitigating catastrophic forgetting, the existing state-of-the-art approach DualPrompt lacks the ability to expand learning capacity within a given task. Specifically, DualPrompt learns a single prompt for each new task - regardless if the task is easy (such as learning a handful of new object classes) or complex (such as learning a large number of new object classes). One might try to increase the length of the prompt, but we show in our experiments (Section 5.3) that increasing the prompt length has saturated returns. Intuitively, we instead desire a method where the learning capacity is correlated to the underlying complexity of task data, rather than a single prompt. Additionally, we desire an approach that is end-to-end differentiable, which we conjecture increases the ability to learn a new task with higher accuracy. + +3We note that the difference between L2P and DualPrompt w.r.t. prompt querying is that the size of the prompt pool in L2P is chosen as a hyperparameter, but in DualPrompt it is equal to the number of tasks and, during training, DualPrompt selects the prompt using task-id (and selects the closest key-query match during inference). + +4The output is read from the class token. + +We grow our learning capacity by introducing a new axis: a set of prompt components. Rather than pick and choose prompts from a pool, we learn a set of prompt components which, via a weighted summation, form a decomposed5 prompt that is passed to the corresponding MSA layer. This allows us to grow our prompting capacity to arbitrary depth and capture the rich underlying complexity of any task while maintaining a fixed prompt length. In addition, prompting in new tasks will inherently reuse the previously acquired knowledge of past tasks rather than initializing a new task prompt from scratch. + +Specifically, we replace learnable prompt parameter $\pmb{p}$ with a weighted summation over the prompt components: + +$$ +\boldsymbol {p} = \sum_ {m} \alpha_ {m} \boldsymbol {P} _ {m} \tag {3} +$$ + +where $P \in \mathbb{R}^{L_p \times D \times M}$ is our set of prompt components, $M$ is the length of our set (i.e., the introduced axis of additional capacity), and $\alpha$ is a weighting vector that determines the contribution of each component which will be discussed in the next subsection. + +A distinct difference between our method and L2P/DualPrompt is that all of our learnable parameters are optimized using the classification loss rather than splitting into two separate loss optimizations. This allows for better end-to-end learning in our method, and we argue this is a key reason why our method performs better in the benchmarks described in Section 5. + +# 4.2. Prompt-Component Weighting + +Rather than a pool of prompts, we have a set of prompt components and desire to produce a weight vector $\alpha$ given a query $\theta(x)$ , rather than a prompt index. We compute the cosine similarity between a query and keys to produce our weighting vector as: + +$$ +\begin{array}{l} \alpha = \gamma (q (\boldsymbol {x}), \boldsymbol {K}) \\ = \left\{\gamma (q (\boldsymbol {x}), \boldsymbol {K} _ {1}), \gamma (q (\boldsymbol {x}), \boldsymbol {K} _ {2}), \dots , \gamma (q (\boldsymbol {x}), \boldsymbol {K} _ {M}) \right\} \tag {4} \\ \end{array} +$$ + +where $\pmb{K} \in \mathbb{R}^{D \times M}$ contains keys corresponding to our prompt components. The intuition here is that the contributing magnitude of each prompt component $\pmb{P}_{m}$ to the final prompt $\pmb{p}$ is determined by the similarity between the query $q(\pmb{x})$ and corresponding key $\pmb{K}_{m}$ . + +The prompt-query matching can be thought of as clustering in a high dimension space $D$ which is a well-known and difficult problem [31]. To mitigate this drawback, we introduce another component to our key-query matching: attention. Each $P_{m}$ has a corresponding attention vector $A_{m}$ in addition to the key $K_{m}$ . This allows the query to focus on certain features from the high dimensional query + +![](images/4d37f8d1134de127a55dcc201666115e9e5d18d45881b047a51a9426b060cab2.jpg) +Figure 2. Our CODA-Prompt approach to rehearsal-free continual learning. An input (e.g., image) is passed through a query function (we use the pretrained ViT encoder) and used for a novel attention-based prompt-selection scheme to produce prompts which are passed to multiple layers of a large-scale, pre-trained transformer. Our method is parameterized by an expanding set of small prompt components, with each prompt having a corresponding key and attention vector. Both the input and produced prompts are passed through the transformer and sent to the task head. Only the prompting and task parameters are learned which is parameter efficient and no training data is stored for replay which is memory-efficient and privacy preserving. Importantly, our prompt selection scheme can be optimized end-to-end, unlike prior works which have a local similarity-based optimization for key-query matching. + +$q(\pmb{x})$ output - for example, a prompt component used for mammalian face features to classify dogs and cats would be able to focus on relevant query features such as eye placement and skin texture, and furthermore ignore features of unrelated semantics, such as features useful for automobile classification. We use a simple feature-selection attention scheme with element-wise multiplication between the query vector and attention vector to create an attended-query which is then used for the key-similarity matching. Specifically, our updated approach to producing our weighting vector is: + +$$ +\begin{array}{l} \begin{array}{l} \alpha = \gamma (q (\boldsymbol {x}) \odot \boldsymbol {A}, \boldsymbol {K}) \\ = \left\{\gamma (q (\boldsymbol {x}) \odot \boldsymbol {A}, \boldsymbol {K}), \quad \gamma (q (\boldsymbol {x}) \odot \boldsymbol {A}, \boldsymbol {K}), \right\} \end{array} \tag {5} \\ = \left\{\gamma (q (\boldsymbol {x}) \odot \boldsymbol {A} _ {1}, \boldsymbol {K} _ {1}), \dots , \gamma (q (\boldsymbol {x}) \odot \boldsymbol {A} _ {M}, \boldsymbol {K} _ {M}) \right\} \\ \end{array} +$$ + +where $\mathbf{A} \in \mathbb{R}^{D \times M}$ contains learnable parameters (attention vectors) corresponding to our prompt components and $\odot$ is the element-wise product (also known as the Hadamard product). Notice that our attention vectors are learnable feature weightings rather than input-conditioned modules - we found that this fixed representation was less vulnerable to forgetting by its simple design, similar to the our prompt component keys. + +# 4.3. Expansion & Orthogonality + +Crucial to mitigating catastrophic forgetting is the avoidance of overwriting knowledge acquired in prior tasks. When we visit a new task we freeze our current components and expand the set, only updating the new components. This is visualized in the bottom of Figure 2 where the existing parameters are locked and only the expanded parameters are optimized. Specifically, in task $t$ we learn $\frac{M}{N}$ components where $N$ is the number of tasks and $M$ is a + +chosen hyperparameter, keeping the previous $\frac{(t - 1)\cdot M}{N}$ components frozen. This expansion is enabled by our attention-based component-weighting scheme, in that expanding our parameters do not alter the computation of weights $\alpha$ corresponding previously learned components. For example, expanding the capacity of a nonlinear model like an MLP module, transformer attention head [13], or hypernetwork [63] would alter the weight outputs of previously learned components. + +While the prior heuristic is focused on reducing catastrophic forgetting (not destroying previously acquired knowledge), we also want to reduce interference between existing and new knowledge. To do this, we add an orthogonality constraint to $P$ , $K$ , and $A$ . The intuition is that orthogonal vectors have less interference with each other. For example, we would want a key and prompt learned in task 2 to not attract task 1 data, otherwise the encoder output for task 1 data will change, resulting in misclassified data. We introduce a simple orthogonality penalty loss to our method as: + +$$ +\mathcal {L} _ {\text {o r t h o}} (B) = \left| \left| B B ^ {\top} - I \right| \right| _ {2} \tag {6} +$$ + +where $B$ represents any arbitrary matrix. + +# 4.4. Full Optimization + +Given task classification loss $\mathcal{L}$ , our full optimization is: + +$$ +\begin{array}{l} \min _ {\boldsymbol {P} ^ {n}, \boldsymbol {K} ^ {n}, \boldsymbol {A} ^ {n}, \phi_ {n}} \mathcal {L} \left(f _ {\phi} \left(f _ {\theta , \boldsymbol {P}, \boldsymbol {K}, \boldsymbol {A}} (\boldsymbol {x})\right), y\right) + \tag {7} \\ \lambda \left(\mathcal {L} _ {\text {o r t h o}} (\boldsymbol {P}) + \mathcal {L} _ {\text {o r t h o}} (\boldsymbol {K}) + \mathcal {L} _ {\text {o r t h o}} (\boldsymbol {A})\right) \\ \end{array} +$$ + +where $P^n, K^n, A^n$ refer to the prompt components and corresponding keys/attention vectors that are unfrozen and + +trained during task $n$ (see Section 4.3) and $\lambda$ is a hyperparameter balancing the orthogonality loss. Note that we do not update the logits of past task classes, consistent with L2P and DualPrompt. We formally refer to our method as COntinual Decomposed Attention-based Prompting, or simply CODA-P. + +# 5. Experiments + +We benchmark our approach with several image datasets in the class-incremental continual learning setting. We implemented baselines which do not store training data for rehearsal: Learning without Forgetting (LwF) [38], Learning to Prompt (L2P) [66], and DualPrompt [65]. Additionally, we report the upper bound performance (i.e., trained offline) and performance for a neural network trained only on classification loss using the new task training data (we refer to this as FT), and include an improved version of FT which uses the same classifier implementation as L2P/DualPrompt (referred to as $\mathrm{FT} + +$ ). We also compare to Experience Replay [11] to provide additional context to our results. We implement our method and all baselines in PyTorch [48] using the ViT-B/16 backbone [13] pre-trained on ImageNet-1K [53]. Our contribution includes faithful PyTorch implementations of the popular prompting baselines L2P and DualPrompt, which were released in JAX [65, 66]. Our implementations of the competing methods actually achieve slight boosts in the performance of DualPrompt in most benchmarks, as well as significant performance boosts in L2P due to the improved prompting type (which we refer to as L2P++. + +DualPrompt uses length 5 prompts in layers 1-2 (referred to as general prompts) and length 20 prompts in layers 3-5 (referred to as task-expert) prompts. We insert prompts into the same layers as DualPrompt (layers 1-5) for CODA-P and use a prompt length of 8 and 100 prompt components, which were chosen with a hyperparameter sweep on validation data to have the best trade-off between performance and parameter efficiency. Because our approach introduces more learnable parameters compared to DualPrompt, we include a variant of out method CODA-P-S which uses a smaller number of parameters equal to DualPrompt for additional comparison. We show that our method outperforms other methods even under this condition, while still retaining the capability to scale if desired. + +We report results on the test dataset, but emphasize that all hyperparameters and design decisions (including for the baselines) were made using validation data (20% of the training data). Unlike DualPrompt, we run our benchmarks for several different shuffles of the task class order and re + +port the mean and standard deviation of these runs. We do this with a consistent seed (different for each trial) so that results can be directly compared. This is crucial because the results with different shuffling order can be different due to differences in task difficulty and their appearance order. We include additional implementation details and results in the Supplementary Material (SM). + +Evaluation Metrics: We evaluate methods using (1) average final accuracy $A_N$ , or the final accuracy with respect to all past classes averaged over $N$ tasks, and (2) average forgetting [9, 41, 42] $F_N$ , or the drop in task performance averaged over $N$ tasks. The reader is referred to our SM for additional details about these metrics. We emphasize that $A_N$ is the more important metric and encompasses both method plasticity and forgetting, whereas $F_N$ simply provides additional context. + +# 5.1. CODA-P sets the SOTA on existing benchmarks + +We first evaluate our method and state of art on several well-established benchmarks. Table 1 provides results for ImageNet-R [21, 65] which is composed of 200 object classes with a wide collection of image styles, including cartoon, graffiti, and hard examples from the original ImageNet dataset [53]. This benchmark is attractive because the distribution of training data has significant distance to the pre-training data (ImageNet), thus providing a fair and challenging problem setting. In addition to the original 10-task benchmark, we also provide results with a smaller number of large tasks (5-task) and a larger number of small tasks (20 task). We first notice that our reported performance for DualPrompt is a few percentage points higher compared to the original DualPrompt [65] paper. We reimplemented the method from scratch in a different framework (PyTorch [48]) and suspect the difference has to do with our averaging over different class-order shuffling. We note that our implementation of L2P $(\mathrm{L}2\mathrm{P}++)$ performs significantly better than originally reported because we use the same form of prompting as DualPrompt (for fair comparison). Furthermore, our "Deep L2P $+ + "$ , which prompts at the same 5 layers as DualPrompt (rather than only at layer 1), actually performs similarly to DualPrompt. See supplementary for discussion and figures illustrating this. + +Importantly, we see that our method has strong gains in average accuracy across all three task lengths, with as much as $4.5\%$ improvement in average accuracy over DualPrompt. We remind the reader that CODA-P is our proposed method with tuned prompt length and component set size, whereas CODA-P-S is downsized (see SM for exact details) to exactly match the number of learnable parameters as DualPrompt. We notice that our method often has slightly higher average forgetting compared to DualPrompt. Given that average accuracy is the crucial metric that captures practical performance, we + +Table 1. Results (\%) on ImageNet-R. Results are included for 5 tasks (40 classes per task), 10 tasks (20 classes per task), and 20 tasks (10 classes per task). $A_N$ gives the accuracy averaged over tasks and $F_N$ gives the average forgetting. We report results over 5 trials. + +
Tasks51020
MethodAN(↑)FN(↓)AN(↑)FN(↓)AN(↑)FN(↓)
UB77.13-77.13-77.13-
ER (5000)71.72 ± 0.7113.70 ± 0.2664.43 ± 1.1610.30 ± 0.0552.43 ± 0.877.70 ± 0.13
FT18.74 ± 0.4441.49 ± 0.5210.12 ± 0.5125.69 ± 0.234.75 ± 0.4016.34 ± 0.19
FT++60.42 ± 0.8714.66 ± 0.2448.93 ± 1.159.81 ± 0.3135.98 ± 1.386.63 ± 0.11
LwF.MC74.56 ± 0.594.98 ± 0.3766.73 ± 1.253.52 ± 0.3954.05 ± 2.662.86 ± 0.26
L2P++70.83 ± 0.583.36 ± 0.1869.29 ± 0.732.03 ± 0.1965.89 ± 1.301.24 ± 0.14
Deep L2P++73.93 ± 0.372.69 ± 0.1071.66 ± 0.641.78 ± 0.1668.42 ± 1.201.12 ± 0.13
DualPrompt73.05 ± 0.502.64 ± 0.1771.32 ± 0.621.71 ± 0.2467.87 ± 1.391.07 ± 0.14
CODA-P-S75.19 ± 0.472.65 ± 0.1573.93 ± 0.491.60 ± 0.2070.53 ± 1.241.00 ± 0.15
CODA-P76.51 ± 0.382.99 ± 0.1975.45 ± 0.561.64 ± 0.1072.37 ± 1.190.96 ± 0.15
+ +Table 2. Results (%) on CIFAR-100 and DomainNet. $A_N$ gives the accuracy averaged over tasks and $F_N$ gives the average forgetting. We report results over 5 trials for CIFAR-100 and 3 trials for DomainNet (due to dataset size). +(a) 10-task CIFAR-100 (10 classes per task) + +
MethodAN(↑)FN(↓)
Upper-Bound89.30-
ER (5000)76.20 ± 1.048.50 ± 0.37
FT9.92 ± 0.2729.21 ± 0.18
FT++49.91 ± 0.4212.30 ± 0.23
LwF.MC64.83 ± 1.035.27 ± 0.39
L2P++82.50 ± 1.101.75 ± 0.42
Deep L2P++84.30 ± 1.031.53 ± 0.40
DualPrompt83.05 ± 1.161.72 ± 0.40
CODA-P-S84.59 ± 0.871.76 ± 0.28
CODA-P86.25 ± 0.741.67 ± 0.26
+ +(b) 5-task DomainNet (69 classes per task) + +
MethodAN(↑)FN(↓)
Upper-Bound79.65-
ER (5000)58.32 ± 0.4726.25 ± 0.24
FT18.00 ± 0.2643.55 ± 0.27
FT++39.28 ± 0.2144.39 ± 0.31
LwF.MC74.78 ± 0.435.01 ± 0.14
L2P++69.58 ± 0.392.25 ± 0.08
Deep L2P++70.54 ± 0.512.05 ± 0.07
DualPrompt70.73 ± 0.492.03 ± 0.22
CODA-P-S71.80 ± 0.572.54 ± 0.10
CODA-P73.24 ± 0.593.46 ± 0.09
+ +are not worried by the slight uptick in forgetting. In fact, we argue it is very reasonable and reflects the strength of our approach: our method has a larger capacity to learn new tasks, and thus we can sacrifice marginally higher forgetting. Crucially, we see that as the task sequence length grows, the forgetting metric of our method versus DualPrompt begins to converge to a similar value. + +We provide results for experience replay to provide additional context of our results. While the gap between rehearsal with a substantial coreset size (i.e., the buffer size of replay data) and our approach is small for the short task sequence, we see that our method strongly outperforms rehearsal in the longer task sequence. + +Tables 2a and 2b provide results on additional8 benchmarks ten-task CIFAR-100 [32] and five-task DomainNet [49]. While neither dataset is as impactful as ImageNet-R in terms of challenge and distance from the pre-trained distribution, these tables provide additional + +context to our method's performance. These tables illustrate a similar story to the ImageNet-R benchmarks, with improvements of $+3.2\%$ and $+2.5\%$ , respectively. We do note that, surprisingly, LwF [38] slightly outperforms prompting on this benchmark. + +# 5.2. CODA-P sets the SOTA for a new dual-shift benchmark + +We also evaluate on a challenging dual-shift benchmark using the ImageNet-R dataset. Our motivation is to show robustness to two different types of continual distribution shifts: semantic and covariate. We accomplish this by randomly selecting image types from the ImageNet-R dataset to include in each task's training data (while keeping the evaluation data unmodified). The reader is refereed to our SM for more details. Results for this benchmark are provided in Table 3. Compared to 5-task ImageNet-R where our method improves SOTA by $4.5\%$ average accuracy, we see a $4.4\%$ improvement on this 5-task benchmark with similar forgetting margins. Our results indicate that CODA-P better generalizes to real-world type shifts. + +Table 3. Results (\%) on ImageNet-R with covariate domain shifts. Results are included for 5 tasks (40 classes per task). We simulate domain shifts by randomly removing $50\%$ of the dataset's domains (e.g., clipart, paintings, and cartoon) for the training data of each task (see SM for more details). $A_N$ gives the accuracy averaged over tasks and $F_N$ gives the average forgetting. We report results over 5 trials. + +
MethodAN(↑)FN(↓)
Upper-Bound77.132.66
ER (5000)67.39 ± 0.3711.94 ± 0.17
FT17.93 ± 0.2737.49 ± 0.28
FT++54.51 ± 0.6814.41 ± 0.33
LwF.MC64.02 ± 1.557.05 ± 0.27
L2P++65.08 ± 0.292.79 ± 0.32
Deep L2P++65.74 ± 0.122.48 ± 0.30
DualPrompt66.98 ± 0.082.21 ± 0.28
CODA-P-S69.73 ± 0.182.35 ± 0.19
CODA-P71.35 ± 0.082.56 ± 0.26
+ +Table 4. Ablation Results (\%) on 10-task ImageNet-R. $A_N$ gives the accuracy averaged over tasks and $F_N$ gives the average forgetting. We report results over 5 trials. + +
MethodAN(↑)FN(↓)
CODA-P75.45 ± 0.561.64 ± 0.10
Ablate Attention74.52 ± 0.651.67 ± 0.13
Ablate Freezing74.60 ± 0.642.29 ± 0.10
Ablate Orthogonality70.66 ± 0.601.74 ± 0.25
+ +# 5.3. Ablations and Additional Analysis + +We take a closer look at our method with ablation studies and additional analysis. In Table 4 we show the effect of removing a few key components of our method: attention keys, freezing of past task components, and our orthogonality regularization. We show that removing attention slightly reduces forgetting degrades performance on average accuracy (the more important metric). This makes sense because removing the attention keys makes our query processing more aligned with the L2P/DualPrompt methods which boast low forgetting but lack sufficient learning capacity. We see larger drops when removing freezing and orthogonality. This indicates that these aspects are crucial to our approach. Our intuition is that, without these, our prompt formation is similar to a shallow MLP module which, unregularized, should suffer from high forgetting. + +We also look at our ability to grow in model capacity along our newly introduced prompt component dimension using the 5-task ImageNet-R benchmark (with validation data). We show in Figure 3 that, with a number of prompt components equal to the number of prompts learned in DualPrompt, we achieve higher performance. Importantly, when we increase the number of components, we are able to leverage the scaled parameters for significantly higher + +![](images/5833e8ab7aa279e3f28487b002243e4ec312317a46bc80612c4a1fca517349d6.jpg) +Figure 3. Analysis of average accuracy $A_N$ vs prompt components (or pool size) for CODA-P, L2P, and DualPrompt. The setting is 5 task ImageNet-R, and we report the mean over 3 trials. + +![](images/f9367098ec6e4e37c56405c9a15b36d0517bd413b3844573b15769021f86a476.jpg) +Figure 4. Analysis of average accuracy $A_N$ vs prompt length for CODA-P, L2P, and DualPrompt. The setting is 5 task ImageNet-R, and we report the mean over 3 trials. + +performance, finding that our method is closer to upper bound performance than it is to DualPrompt in average accuracy. Because the prompt pool in L2P can be increased to arbitrary size as well, we include results in this analysis as well. However, we show that the L2P method peaks with a pool size equal to twice the task sequence length, and then drops in performance. This reflects that our prompt components benefit from scale, whereas the existing prompt pool actually suffer. + +Finally, we show average accuracy versus prompt length in Figure 4 using the same setting as above. The purpose of this experiment is to emphasize that accuracy saturates with prompt length, thus motivating the need to expand in our introduced component dimension. Additional analysis and hyperparameter sweeps is available in our SM. + +# 6. Conclusion + +We present CContinual decomposed attention-based prompting (CODA-Prompt) for rehearsal-free continual learning. Our approach assembles learnable prompt components which are then inserted into a pre-trained ViT encoder for image classification. Importantly, CODA-Prompt is optimized in an end-to-end fashion (unlike prior SOTA methods which involve two, separate optimizations). Furthermore, CODA-Prompt can scale prompting capacity to arbitrary sizes. We set a new SOTA on both well-established benchmarks and a dual-distribution shift benchmark (containing semantic and covariate shifts), highlighting the potential real world impact and generality of our approach. + +# References + +[1] Hongjoon Ahn, Jihwan Kwak, Subin Lim, Hyeonsu Bang, Hyojun Kim, and Taesup Moon. Ss-il: Separated softmax for incremental learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 844-853, October 2021. 2 +[2] Rahaf Aljundi, Francesca Babiloni, Mohamed Elhoseiny, Marcus Rohrbach, and Tinne Tuytelaars. Memory aware synapses: Learning what (not) to forget. In ECCV, 2018. 2 +[3] Rahaf Aljundi, Eugene Belilovsky, Tinne Tuytelaars, Laurent Charlin, Massimo Caccia, Min Lin, and Lucas Page-Caccia. Online continual learning with maximal interfered retrieval. In Advances in Neural Information Processing Systems, pages 11849-11860, 2019. 2 +[4] Rahaf Aljundi, Min Lin, Baptiste Goujaud, and Yoshua Bengio. Gradient based sample selection for online continual learning. In Advances in Neural Information Processing Systems, pages 11816-11825, 2019. 2 +[5] Jihwan Bang, Heesu Kim, YoungJoon Yoo, Jung-Woo Ha, and Jonghyun Choi. Rainbow memory: Continual learning with a memory of diverse samples. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8218-8227, 2021. 2 +[6] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. Advances in neural information processing systems, 33:1877-1901, 2020. 3 +[7] Pietro Buzzega, Matteo Boschini, Angelo Porrello, Davide Abati, and Simone Calderara. Dark experience for general continual learning: a strong, simple baseline. Advances in neural information processing systems, 33:15920-15930, 2020. 2 +[8] Francisco M Castro, Manuel J Marín-Jiménez, Nicolas Guil, Cordelia Schmid, and Karteek Alahari. End-to-end incremental learning. In Proceedings of the European Conference on Computer Vision (ECCV), pages 233-248, 2018. 2 +[9] Arslan Chaudhry, Marc' Aurelio Ranzato, Marcus Rohrbach, and Mohamed Elhoseiny. Efficient lifelong learning with a GEM. In International Conference on Learning Representations, 2019. 2, 6 +[10] Arslan Chaudhry, Marcus Rohrbach, Mohamed Elhoseiny, Thalaiyasingam Ajanthan, Puneet K Dokania, Philip HS Torr, and Marc'Aurelio Ranzato. Continual learning with tiny episodic memories. arXiv preprint arXiv:1902.10486, 2019. 2 +[11] Arslan Chaudhry, Marcus Rohrbach, Mohamed Elhoseiny, Thalaiyasingam Ajanthan, Puneet K Dokania, Philip HS Torr, and Marc'Aurelio Ranzato. On tiny episodic memories in continual learning. arXiv preprint arXiv:1902.10486, 2019. 6 +[12] Yoojin Choi, Mostafa El-Khamy, and Jungwon Lee. Dualteacher class-incremental learning with data-free generative replay. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3543-3552, 2021. 2 + +[13] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 3, 5, 6 +[14] Arthur Douillard, Alexandre Ramé, Guillaume Couairon, and Matthieu Cord. Dytox: Transformers for continual learning with dynamic token expansion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9285-9295, 2022. 3 +[15] Sayna Ebrahimi, Mohamed Elhoseiny, Trevor Darrell, and Marcus Rohrbach. Uncertainty-guided continual learning with bayesian neural networks. arXiv preprint arXiv:1906.02425, 2019. 2 +[16] Sayna Ebrahimi, Franziska Meier, Roberto Calandra, Trevor Darrell, and Marcus Rohrbach. Adversarial continual learning. arXiv preprint arXiv:2003.09553, 2020. 2 +[17] Qiankun Gao, Chen Zhao, Bernard Ghanem, and Jian Zhang. R-dfcil: Relation-guided representation learning for data-free class incremental learning. arXiv preprint arXiv:2203.13104, 2022. 2 +[18] Alexander Gepperth and Cem Karaoguz. Incremental learning with self-organizing maps. 2017 12th International Workshop on Self-Organizing Maps and Learning Vector Quantization, Clustering and Data Visualization (WSOM), pages 1–8, 2017. 2 +[19] Tyler L Hayes, Nathan D Cahill, and Christopher Kanan. Memory efficient experience replay for streaming learning. In 2019 International Conference on Robotics and Automation (ICRA), pages 9769-9776. IEEE, 2019. 2 +[20] Tyler L Hayes and Christopher Kanan. Lifelong machine learning with deep streaming linear discriminant analysis. arXiv preprint arXiv:1909.01520, 2019. 3 +[21] Dan Hendrycks, Steven Basart, Norman Mu, Saurav Kadavath, Frank Wang, Evan Dorundo, Rahul Desai, Tyler Zhu, Samyak Parajuli, Mike Guo, Dawn Song, Jacob Steinhardt, and Justin Gilmer. The many faces of robustness: A critical analysis of out-of-distribution generalization. ICCV, 2021. 2, 6 +[22] Geoffrey Hinton, Oriol Vinyals, and Jeff Dean. Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531, 2015. 2 +[23] Saihui Hou, Xinyu Pan, Chen Change Loy, Zilei Wang, and Dahua Lin. Lifelong learning via progressive distillation and retrospection. In Proceedings of the European Conference on Computer Vision (ECCV), pages 437-452, 2018. 2 +[24] Saihui Hou, Xinyu Pan, Chen Change Loy, Zilei Wang, and Dahua Lin. Learning a unified classifier incrementally via rebalancing. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 831-839, 2019. 2 +[25] Yen-Chang Hsu, Yen-Cheng Liu, Anita Ramasamy, and Zsolt Kira. Re-evaluating continual learning scenarios: A categorization and case for strong baselines. arXiv preprint arXiv:1810.12488, 2018. 2, 3 +[26] Daniel Justus, John Brennan, Stephen Bonner, and Andrew Stephen McGough. Predicting the computational cost + +of deep learning models. In 2018 IEEE international conference on big data (Big Data), pages 3873-3882. IEEE, 2018. 1 +[27] Nitin Kamra, Umang Gupta, and Yan Liu. Deep generative dual memory network for continual learning. arXiv preprint arXiv:1710.10368, 2017. 2 +[28] Ronald Kemker and Christopher Kanan. Fearnet: Brain-inspired model for incremental learning. International Conference on Learning Representations (ICLR), 2018. 2 +[29] Ronald Kemker, Marc McClure, Angelina Abitino, Tyler Hayes, and Christopher Kanan. Measuring catastrophic forgetting in neural networks. AAAI Conference on Artificial Intelligence, 2018. 2 +[30] James Kirkpatrick, Razvan Pascanu, Neil Rabinowitz, Joel Veness, Guillaume Desjardins, Andrei A Rusu, Kieran Milan, John Quan, Tiago Ramalho, Agnieszka Grabska-Barwinska, et al. Overcoming catastrophic forgetting in neural networks. Proceedings of the national academy of sciences, 2017. 2 +[31] Mario Köppen. The curse of dimensionality. In 5th online world conference on soft computing in industrial applications (WSC5), volume 1, pages 4-8, 2000. 4 +[32] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. Tech Report, 2009. 2, 7 +[33] Alexandre Lacoste, Alexandra Luccioni, Victor Schmidt, and Thomas Dandres. Quantifying the carbon emissions of machine learning. arXiv preprint arXiv:1910.09700, 2019. 1 +[34] Kibok Lee, Kimin Lee, Jinwoo Shin, and Honglak Lee. Overcoming catastrophic forgetting with unlabeled data in the wild. In Proceedings of the IEEE International Conference on Computer Vision, pages 312-321, 2019. 2 +[35] Soochan Lee, Junsoo Ha, Dongsu Zhang, and Gunhee Kim. A neural dirichlet process mixture model for task-free continual learning. arXiv preprint arXiv:2001.00689, 2020. 2 +[36] Timothee Lesort, Hugo Caselles-Dupré, Michael Garcia-Ortiz, Andrei Stoian, and David Filliat. Generative models from the perspective of continual learning. In 2019 International Joint Conference on Neural Networks (IJCNN), pages 1–8. IEEE, 2019. 2 +[37] Duo Li, Guimei Cao, Yunlu Xu, Zhanzhan Cheng, and Yi Niu. Technical report for iccv 2021 challenge sslad-track3b: Transformers are better continual learners. arXiv preprint arXiv:2201.04924, 2022. 3 +[38] Zhizhong Li and Derek Hoiem. Learning without forgetting. IEEE transactions on pattern analysis and machine intelligence, 40(12):2935-2947, 2017. 2, 6, 7 +[39] Vincenzo Lomonaco and Davide Maltoni. Core50: a new dataset and benchmark for continuous object recognition. arXiv preprint arXiv:1705.03550, 2017. 2 +[40] Vincenzo Lomonaco, Davide Maltoni, and Lorenzo Pellegrini. Rehearsal-free continual learning over small non-iid batches. In CVPR Workshops, pages 989-998, 2020. 3 +[41] David Lopez-Paz and Marc'Aurelio Ranzato. Gradient episodic memory for continual learning. In Proceedings of the 31st International Conference on Neural Information Processing Systems, NIPS'17, pages 6470-6479, USA, 2017. Curran Associates Inc. 2, 6 + +[42] Zheda Mai, Ruiwen Li, Jihwan Jeong, David Quispe, Hyunwoo Kim, and Scott Sanner. Online continual learning in image classification: An empirical survey. Neurocomputing, 469:28-51, 2022. 6 +[43] Davide Maltoni and Vincenzo Lomonaco. Continuous learning in single-incremental-task scenarios. Neural Networks, 116:56-73, 2019. 2 +[44] Seyed Iman Mirzadeh, Arslan Chaudhry, Dong Yin, Timothy Nguyen, Razvan Pascanu, Dilan Gorur, and Mehrdad Farajtabar. Architecture matters in continual learning. arXiv preprint arXiv:2202.00275, 2022. 3 +[45] Vaishnavh Nagarajan, Colin Raffel, and Ian J Goodfellow. Theoretical insights into memorization in gans. In Neural Information Processing Systems Workshop, 2018. 2 +[46] Cuong V Nguyen, Alessandro Achille, Michael Lam, Tal Hassner, Vijay Mahadevan, and Stefano Soatto. Toward understanding catastrophic forgetting in continual learning. arXiv preprint arXiv:1908.01091, 2019. 1 +[47] Oleksiy Ostapenko, Mihai Puscas, Tassilo Klein, Patrick Jahnichen, and Moin Nabi. Learning to remember: A synaptic plasticity driven framework for continual learning. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 11321-11329, 2019. 2 +[48] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. Advances in neural information processing systems, 32, 2019. 6 +[49] Xingchao Peng, Qinxun Bai, Xide Xia, Zijun Huang, Kate Saenko, and Bo Wang. Moment matching for multi-source domain adaptation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 1406-1415, 2019. 7 +[50] Sylvestre-Alvise Rebuffi, Alexander Kolesnikov, Georg Sperl, and Christoph H. Lampert. icarl: Incremental classifier and representation learning. In 2017 IEEE Conference on Computer Vision and Pattern Recognition, CVPR'17, pages 5533-5542, 2017. 2 +[51] Anthony Robins. Catastrophic forgetting, rehearsal and pseudorehearsal. Connection Science, 7(2):123-146, 1995. 2 +[52] David Rolnick, Arun Ahuja, Jonathan Schwarz, Timothy Lillicrap, and Gregory Wayne. Experience replay for continual learning. In Advances in Neural Information Processing Systems, pages 348-358, 2019. 2 +[53] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, et al. Imagenet large scale visual recognition challenge. International journal of computer vision, 115(3):211-252, 2015. 6 +[54] Andrei A Rusu, Neil C Rabinowitz, Guillaume Desjardins, Hubert Soyer, James Kirkpatrick, Koray Kavukcuoglu, Razvan Pascanu, and Raia Hadsell. Progressive neural networks. arXiv preprint arXiv:1606.04671, 2016. 2 +[55] Hanul Shin, Jung Kwon Lee, Jaehong Kim, and Jiwon Kim. Continual learning with deep generative replay. In I. Guyon, + +U. V. Luxburg, S. Bengio, H. Wallach, R. Fergus, S. Vishwanathan, and R. Garnett, editors, Advances in Neural Information Processing Systems 30, pages 2990-2999. Curran Associates, Inc., 2017. 2 +[56] James Smith, Yen-Chang Hsu, Jonathan Balloch, Yilin Shen, Hongxia Jin, and Zsolt Kira. Always be dreaming: A new approach for data-free class-incremental learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 9374-9384, October 2021. 2, 3 +[57] James Seale Smith, Junjiao Tian, Yen-Chang Hsu, and Zsolt Kira. A closer look at rehearsal-free continual learning. arXiv preprint arXiv:2203.17269, 2022. 2 +[58] Michalis K Titsias, Jonathan Schwarz, Alexander G de G Matthews, Razvan Pascanu, and Yee Whye Teh. Functional regularisation for continual learning with gaussian processes. In International Conference on Learning Representations, 2019. 2 +[59] Gido M van de Ven, Hava T Siegelmann, and Andreas S Tolias. Brain-inspired replay for continual learning with artificial neural networks. Nature communications, 11(1):1-14, 2020. 2 +[60] Gido M van de Ven and Andreas S Tolias. Generative replay with feedback connections as a general strategy for continual learning. arXiv preprint arXiv:1809.10635, 2018. 2 +[61] Gido M Van de Ven and Andreas S Tolias. Three scenarios for continual learning. arXiv preprint arXiv:1904.07734, 2019. 2 +[62] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 3 +[63] Johannes von Oswald, Christian Henning, João Sacramento, and Benjamin F Grewe. Continual learning with hypernetworks. arXiv preprint arXiv:1906.00695, 2019. 2, 5 +[64] Yabin Wang, Zhiwu Huang, and Xiaopeng Hong. S-prompts learning with pre-trained transformers: An occam's razor for domain incremental learning. In Conference on Neural Information Processing Systems (NeurIPS), 2022. 3 +[65] Zifeng Wang, Zizhao Zhang, Sayna Ebrahimi, Ruoxi Sun, Han Zhang, Chen-Yu Lee, Xiaoqi Ren, Guolong Su, Vincent Perot, Jennifer Dy, et al. Dualprompt: Complementary prompting for rehearsal-free continual learning. arXiv preprint arXiv:2204.04799, 2022. 1, 2, 3, 4, 6 +[66] Zifeng Wang, Zizhao Zhang, Chen-Yu Lee, Han Zhang, Ruoxi Sun, Xiaqi Ren, Guolong Su, Vincent Perot, Jennifer Dy, and Tomas Pfister. Learning to prompt for continual learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 139-149, 2022. 2, 3, 4, 6 +[67] Yue Wu, Yinpeng Chen, Lijuan Wang, Yuancheng Ye, Zicheng Liu, Yandong Guo, and Yun Fu. Large scale incremental learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 374-382, 2019. 2, 3 +[68] Hongxu Yin, Pavlo Molchanov, Jose M Alvarez, Zhizhong Li, Arun Mallya, Derek Hoiem, Niraj K Jha, and Jan Kautz. + +Dreaming to distill: Data-free knowledge transfer via deep-inversion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8715-8724, 2020. 2 +[69] Wenpeng Yin, Jamaal Hay, and Dan Roth. Benchmarking zero-shot text classification: Datasets, evaluation and entailment approach. arXiv preprint arXiv:1909.00161, 2019. 3 +[70] Pei Yu, Yinpeng Chen, Ying Jin, and Zicheng Liu. Improving vision transformers for incremental learning. arXiv preprint arXiv:2112.06103, 2021. 3 +[71] Friedemann Zenke, Ben Poole, and Surya Ganguli. Continual learning through synaptic intelligence. In International Conference on Machine Learning, 2017. 2 \ No newline at end of file diff --git a/2023/CODA-Prompt_ COntinual Decomposed Attention-Based Prompting for Rehearsal-Free Continual Learning/images.zip b/2023/CODA-Prompt_ COntinual Decomposed Attention-Based Prompting for Rehearsal-Free Continual Learning/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..7dc3dd61446b2c4183fd23673909a87c98cbba5f --- /dev/null +++ b/2023/CODA-Prompt_ COntinual Decomposed Attention-Based Prompting for Rehearsal-Free Continual Learning/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:da3516172782a16fb78db127ae632046a6b7a729fcac771538b6eecc1f877a26 +size 455652 diff --git a/2023/CODA-Prompt_ COntinual Decomposed Attention-Based Prompting for Rehearsal-Free Continual Learning/layout.json b/2023/CODA-Prompt_ COntinual Decomposed Attention-Based Prompting for Rehearsal-Free Continual Learning/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..1a9e7ee2b7fab3292f7764fde90780256af0c332 --- /dev/null +++ b/2023/CODA-Prompt_ COntinual Decomposed Attention-Based Prompting for Rehearsal-Free Continual Learning/layout.json @@ -0,0 +1,9092 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 75, + 102, + 520, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 102, + 520, + 140 + ], + "spans": [ + { + "bbox": [ + 75, + 102, + 520, + 140 + ], + "type": "text", + "content": "CODA-Prompt: CContinual Decomposed Attention-based Prompting for Rehearsal-Free Continual Learning" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 63, + 159, + 531, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 159, + 531, + 190 + ], + "spans": [ + { + "bbox": [ + 63, + 159, + 531, + 190 + ], + "type": "text", + "content": "James Seale Smith" + }, + { + "bbox": [ + 63, + 159, + 531, + 190 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 63, + 159, + 531, + 190 + ], + "type": "text", + "content": " Leonid Karlinsky" + }, + { + "bbox": [ + 63, + 159, + 531, + 190 + ], + "type": "inline_equation", + "content": "^{2,4}" + }, + { + "bbox": [ + 63, + 159, + 531, + 190 + ], + "type": "text", + "content": " Vyshnavi Gutta" + }, + { + "bbox": [ + 63, + 159, + 531, + 190 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 63, + 159, + 531, + 190 + ], + "type": "text", + "content": " Paola Cascante-Bonilla" + }, + { + "bbox": [ + 63, + 159, + 531, + 190 + ], + "type": "inline_equation", + "content": "^{2,3}" + }, + { + "bbox": [ + 63, + 159, + 531, + 190 + ], + "type": "text", + "content": " Donghyun Kim" + }, + { + "bbox": [ + 63, + 159, + 531, + 190 + ], + "type": "inline_equation", + "content": "^{2,4}" + }, + { + "bbox": [ + 63, + 159, + 531, + 190 + ], + "type": "text", + "content": " Assaf Arbelle" + }, + { + "bbox": [ + 63, + 159, + 531, + 190 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 63, + 159, + 531, + 190 + ], + "type": "text", + "content": " Rameswar Panda" + }, + { + "bbox": [ + 63, + 159, + 531, + 190 + ], + "type": "inline_equation", + "content": "^{2,4}" + }, + { + "bbox": [ + 63, + 159, + 531, + 190 + ], + "type": "text", + "content": " Rogerio Feris" + }, + { + "bbox": [ + 63, + 159, + 531, + 190 + ], + "type": "inline_equation", + "content": "^{2,4}" + }, + { + "bbox": [ + 63, + 159, + 531, + 190 + ], + "type": "text", + "content": " Zsolt Kira" + }, + { + "bbox": [ + 63, + 159, + 531, + 190 + ], + "type": "inline_equation", + "content": "^{1}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 92, + 190, + 498, + 202 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 92, + 190, + 498, + 202 + ], + "spans": [ + { + "bbox": [ + 92, + 190, + 498, + 202 + ], + "type": "text", + "content": "1Georgia Institute of Technology 2MIT-IBM Watson AI Lab 3Rice University 4IBM Research" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 143, + 231, + 192, + 243 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 231, + 192, + 243 + ], + "spans": [ + { + "bbox": [ + 143, + 231, + 192, + 243 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 256, + 290, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 256, + 290, + 593 + ], + "spans": [ + { + "bbox": [ + 47, + 256, + 290, + 593 + ], + "type": "text", + "content": "Computer vision models suffer from a phenomenon known as catastrophic forgetting when learning novel concepts from continuously shifting training data. Typical solutions for this continual learning problem require extensive rehearsal of previously seen data, which increases memory costs and may violate data privacy. Recently, the emergence of large-scale pre-trained vision transformer models has enabled prompting approaches as an alternative to data-rehearsal. These approaches rely on a key-query mechanism to generate prompts and have been found to be highly resistant to catastrophic forgetting in the well-established rehearsal-free continual learning setting. However, the key mechanism of these methods is not trained end-to-end with the task sequence. Our experiments show that this leads to a reduction in their plasticity, hence sacrificing new task accuracy, and inability to benefit from expanded parameter capacity. We instead propose to learn a set of prompt components which are assembled with input-conditioned weights to produce input-conditioned prompts, resulting in a novel attention-based end-to-end key-query scheme. Our experiments show that we outperform the current SOTA method DualPrompt on established benchmarks by as much as " + }, + { + "bbox": [ + 47, + 256, + 290, + 593 + ], + "type": "inline_equation", + "content": "4.5\\%" + }, + { + "bbox": [ + 47, + 256, + 290, + 593 + ], + "type": "text", + "content": " in average final accuracy. We also outperform the state of art by as much as " + }, + { + "bbox": [ + 47, + 256, + 290, + 593 + ], + "type": "inline_equation", + "content": "4.4\\%" + }, + { + "bbox": [ + 47, + 256, + 290, + 593 + ], + "type": "text", + "content": " accuracy on a continual learning benchmark which contains both class-incremental and domain-incremental task shifts, corresponding to many practical settings. Our code is available at https://github.com/GT-RIPL/CODA-Prompt" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 615, + 128, + 627 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 615, + 128, + 627 + ], + "spans": [ + { + "bbox": [ + 47, + 615, + 128, + 627 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "type": "text", + "content": "For a computer vision model to succeed in the real world, it must overcome brittle assumptions that the concepts it will encounter after deployment will match those learned " + }, + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "type": "inline_equation", + "content": "a" + }, + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "type": "text", + "content": "-priori during training. The real world is indeed dynamic and contains continuously emerging objects and categories. Once deployed, models will encounter a range of differences from their training sets, requiring us to continuously" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 312, + 231, + 542, + 313 + ], + "blocks": [ + { + "bbox": [ + 312, + 231, + 542, + 313 + ], + "lines": [ + { + "bbox": [ + 312, + 231, + 542, + 313 + ], + "spans": [ + { + "bbox": [ + 312, + 231, + 542, + 313 + ], + "type": "image", + "image_path": "9fc104dd3bfbb342f5ac031223ecc62e911b4c3f4618e42f7592753a4f003667.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 399, + 315, + 453, + 325 + ], + "lines": [ + { + "bbox": [ + 399, + 315, + 453, + 325 + ], + "spans": [ + { + "bbox": [ + 399, + 315, + 453, + 325 + ], + "type": "text", + "content": "Prior Work" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 312, + 329, + 542, + 413 + ], + "blocks": [ + { + "bbox": [ + 312, + 329, + 542, + 413 + ], + "lines": [ + { + "bbox": [ + 312, + 329, + 542, + 413 + ], + "spans": [ + { + "bbox": [ + 312, + 329, + 542, + 413 + ], + "type": "image", + "image_path": "377f964a93e17e9799fa6fdfaab3d0f5dde02db371cebdffb0392499006b34ce.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 404, + 415, + 451, + 426 + ], + "lines": [ + { + "bbox": [ + 404, + 415, + 451, + 426 + ], + "spans": [ + { + "bbox": [ + 404, + 415, + 451, + 426 + ], + "type": "text", + "content": "Our Work" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 305, + 435, + 547, + 513 + ], + "lines": [ + { + "bbox": [ + 305, + 435, + 547, + 513 + ], + "spans": [ + { + "bbox": [ + 305, + 435, + 547, + 513 + ], + "type": "text", + "content": "Figure 1. Prior work [65] learns a pool of key-value pairs to select learnable prompts which are inserted into several layers of a pretrained ViT (the prompting parameters are unique to each layer). Our work introduces a decomposed prompt which consists of learnable prompt components that assemble to produce attention-conditioned prompts. Unlike prior work, ours is optimized in an end-to-end fashion (denoted with thick, green lines)." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 533, + 523, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 533, + 523, + 544 + ], + "spans": [ + { + "bbox": [ + 306, + 533, + 523, + 544 + ], + "type": "text", + "content": "update them to avoid stale and decaying performance." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 545, + 546, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 545, + 546, + 700 + ], + "spans": [ + { + "bbox": [ + 304, + 545, + 546, + 700 + ], + "type": "text", + "content": "One way to update a model is to collect additional training data, combine this new training data with the old training data, and then retrain the model from scratch. While this will guarantee high performance, it is not practical for large-scale applications which may require lengthy training times and aggressive replacement timelines, such as models for self-driving cars or social media platforms. This could incur high financial [26] and environmental [33] costs if the replacement is frequent. We could instead update the model by only training on the new training data, but this leads to a phenomenon known as catastrophic forgetting [46] where the model overwrites existing knowledge when learning the new data, a problem known as continual learning." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 318, + 701, + 546, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 701, + 546, + 714 + ], + "spans": [ + { + "bbox": [ + 318, + 701, + 546, + 714 + ], + "type": "text", + "content": "The most effective approaches proposed by the continual" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "11909" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 286, + 132 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 286, + 132 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 286, + 132 + ], + "type": "text", + "content": "learning community involve saving [50] or generating [55] a subset of past training data and mixing it with future task data, a strategy referred to as rehearsal. Yet many important applications are unable to store this data because they work with private user data that cannot be stored long term." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 133, + 287, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 133, + 287, + 418 + ], + "spans": [ + { + "bbox": [ + 46, + 133, + 287, + 418 + ], + "type": "text", + "content": "In this paper, we instead consider the highly-impactful and well-established setting of rehearsal-free continual learning [38, 56, 57, 65, 66], limiting our scope to strategies for continual learning which do not store training data. While rehearsal-free approaches have classically under-performed rehearsal-based approaches in this challenging setting by wide-margins [57], recent prompt-based approaches [65, 66], leveraging pre-trained Vision transformers (ViT), have had tremendous success and can even outperform state-of-the-art (SOTA) rehearsal-based methods. These prompting approaches boast a strong protection against catastrophic forgetting by learning a small pool of insert-able model embeddings (prompts) rather than modifying vision encoder parameters directly. One drawback of these approaches, however, is that they cannot be optimized in an end-to-end fashion as they use a key and query to select a prompt index from the pool of prompts, and thus rely on a second, localized optimization to learn the keys because the model gradient cannot backpropagate through the key/query index selection. Furthermore, these methods reduce forgetting by sacrificing new task accuracy (i.e., they lack sufficient plasticity). We show that expanding the prompt size does not increase the plasticity, motivating us to grow learning capacity from a new perspective." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 419, + 286, + 658 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 419, + 286, + 658 + ], + "spans": [ + { + "bbox": [ + 46, + 419, + 286, + 658 + ], + "type": "text", + "content": "We replace the prompt pool with a decomposed prompt that consists of a weighted sum of learnable prompt components (Figure 1). This decomposition enables higher prompting capacity by expanding in a new dimension (the number of components). Furthermore, this inherently encourages knowledge re-use, as future task prompts will include contributions from prior task components. We also introduce a novel attention-based component-weighting scheme, which allows for our entire method to be optimized in an end-to-end fashion unlike the existing works, increasing our plasticity to better learn future tasks. We boast significant performance gains on existing rehearsal-free continual learning benchmarks w.r.t. SOTA prompting baselines in addition to a challenging dual-shift benchmark containing both incremental concept shifts and covariate domain shifts, highlighting our method's impact and generality. Our method improves results even under equivalent parameter counts, but importantly can scale performance by increasing capacity, unlike prior methods which quickly plateau. In summary, we make the following contributions:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 313, + 72, + 545, + 272 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 313, + 72, + 545, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 545, + 144 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 545, + 144 + ], + "type": "text", + "content": "1. We introduce a decomposed attention-based prompting for rehearsal-free continual learning, characterized by an expanded learning capacity compared to existing continual prompting approaches. Importantly, our approach can be optimized in an end-to-end fashion, unlike the prior SOTA." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 154, + 545, + 202 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 154, + 545, + 202 + ], + "spans": [ + { + "bbox": [ + 313, + 154, + 545, + 202 + ], + "type": "text", + "content": "2. We establish a new SOTA for rehearsal-free continual learning on the well-established ImageNet-R [21, 65] and CIFAR-100 [32] benchmarks, beating the previous SOTA method DualPrompt [65] by as much as " + }, + { + "bbox": [ + 313, + 154, + 545, + 202 + ], + "type": "inline_equation", + "content": "4.5\\%" + }, + { + "bbox": [ + 313, + 154, + 545, + 202 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 212, + 545, + 272 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 212, + 545, + 272 + ], + "spans": [ + { + "bbox": [ + 313, + 212, + 545, + 272 + ], + "type": "text", + "content": "3. We evaluate on a challenging benchmark with dual distribution shifts (semantic and covariate) using the ImageNet-R dataset, and again outperform the state of art, highlighting the real-world impact and generality of our approach." + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 285, + 480, + 298 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 285, + 480, + 298 + ], + "spans": [ + { + "bbox": [ + 306, + 285, + 480, + 298 + ], + "type": "text", + "content": "2. Background and Related Work" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 305, + 545, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 305, + 545, + 496 + ], + "spans": [ + { + "bbox": [ + 304, + 305, + 545, + 496 + ], + "type": "text", + "content": "Continual Learning: Continual learning approaches can be organized into a few broad categories which are all useful depending on the problem setting and constraints. One group of approaches expands a model's architecture as new tasks are encountered; these are highly effective for applications where a model growing with tasks is practical [16, 35, 39, 43, 54]. Another approach is to regularize the model with respect to past task knowledge while training the new task. This can either be done by regularizing the model in the weight space (i.e., penalize changes to model parameters) [2, 15, 30, 58, 71] or the prediction space (i.e., penalize changes to model predictions) [1, 8, 23, 34, 38]. Regularizing knowledge in the prediction space is done using knowledge distillation [22] and it has been found to perform better than model regularization-based methods for continual learning when task labels are not given [36, 60]." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 498, + 545, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 498, + 545, + 665 + ], + "spans": [ + { + "bbox": [ + 304, + 498, + 545, + 665 + ], + "type": "text", + "content": "Rehearsal with stored data [3-5, 7, 9, 10, 18, 19, 24, 29, 41, 50-52, 63, 67] or samples from a generative model [27, 28, 47, 55, 59] is highly effective when storing training data or training/saving a generative model is possible. Unfortunately, for many machine learning applications long-term storage of training data will violate data privacy, as well as incur a large memory cost. With respect to the generative model, this training process is much more computationally and memory intensive compared to a classification model and additionally may violate data legality concerns because using a generative model increases the chance of memorizing potentially sensitive data [45]. This motivates us to work on the important setting of rehearsal-free approaches to mitigate catastrophic forgetting." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "content": "Rehearsal-Free Continual Learning: Recent works propose producing images for rehearsal using deep-model inversion [12, 17, 56, 68]. While these methods perform well compared to generative modeling approaches and simply" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 46, + 664, + 286, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 664, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 664, + 286, + 713 + ], + "type": "text", + "content": "1We focus on continual learning over a single, expanding classification head called class-incremental continual learning. This is different from the multi-task continual learning setting, known as task-incremental continual learning, where we learn separate classification heads for each task and the task label is provided during inference. [25, 61]" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "11910" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 191 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 191 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 191 + ], + "type": "text", + "content": "rehearse from a small number of stored images, model-inversion is a slow process associated with high computational costs in the continual learning setting, and furthermore these methods under-perform rehearsal-based approaches by significant margins [56]. Other works have looked at rehearsal-free continual learning from an online \"streaming\" learning perspective using a frozen, pre-trained model [20, 40]. Because we allow our models to train to convergence on task data (as is common for continual learning [67]), our setting is very different." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 193, + 289, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 193, + 289, + 384 + ], + "spans": [ + { + "bbox": [ + 46, + 193, + 289, + 384 + ], + "type": "text", + "content": "Continual Learning in Vision Transformers: Recent work [6, 37, 69] has proven transformers to generalize well to unseen domains. For example, one study varies the number of attention heads in Vision transformers and conclude that ViTs offer more robustness to forgetting with respect to CNN-based equivalents [44]. Another [70] shows that the vanilla counterparts of ViTs are more prone to forgetting when trained from scratch. Finally, DyTox [14] learns a unified model with a parameter-isolation approach and dynamically expands the tokens processed by the last layer to mitigate forgetting. For each task, they learn a new task-specific token per head using task-attention based decoder blocks. However, the above works either rely on exemplars to defy forgetting or they need to train a new transformer model from scratch, a costly operation, hence differing from our objective of exemplar-free CL using pre-trained ViTs." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 385, + 289, + 577 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 385, + 289, + 577 + ], + "spans": [ + { + "bbox": [ + 46, + 385, + 289, + 577 + ], + "type": "text", + "content": "Prompting for Continual Learning: Prompt-based approaches for continual learning boast a strong protection against catastrophic forgetting by learning a small number of insert-able model instructions (prompts) rather than modifying encoder parameters directly. The current state-of-the-art approaches for our setting, DualPrompt [65] and L2P [66], create a pool of prompts to be selected for insertion into the model, matching input data to prompts without task id with a local clustering-like optimization. We build upon the foundation of these methods, as discussed later in this paper. We note that the recent S-Prompts [64] method is similar to these approaches but designed for domain-incremental learning (i.e., learning the same set of classes under covariate distribution task shifts), which is different than the class-incremental continual learning setting of our paper (i.e., learn emerging object classes in new tasks)." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 590, + 132, + 601 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 590, + 132, + 601 + ], + "spans": [ + { + "bbox": [ + 47, + 590, + 132, + 601 + ], + "type": "text", + "content": "3. Preliminaries" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 610, + 163, + 623 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 610, + 163, + 623 + ], + "spans": [ + { + "bbox": [ + 47, + 610, + 163, + 623 + ], + "type": "text", + "content": "3.1. Continual Learning" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 629, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 629, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 629, + 287, + 714 + ], + "type": "text", + "content": "In our continual learning setting, a model is sequentially shown " + }, + { + "bbox": [ + 46, + 629, + 287, + 714 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 629, + 287, + 714 + ], + "type": "text", + "content": " tasks corresponding to non-overlapping subsets of semantic object classes. Each class appears in only a single task, and the goal is to incrementally learn to classify new object classes as they are introduced while retaining performance on previously learned classes. To describe our models, we denote " + }, + { + "bbox": [ + 46, + 629, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 46, + 629, + 287, + 714 + ], + "type": "text", + "content": " as a pre-trained vision encoder and " + }, + { + "bbox": [ + 46, + 629, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\phi_{n}" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 72, + 547, + 191 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 191 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 191 + ], + "type": "text", + "content": "as our classifier head with logits corresponding to task " + }, + { + "bbox": [ + 304, + 72, + 547, + 191 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 304, + 72, + 547, + 191 + ], + "type": "text", + "content": " classes. In this paper, we deal with the class-incremental continual learning setting rather than the task-incremental continual learning setting. Class-incremental continual learning is challenging because the learner must support classification across all classes seen up to task " + }, + { + "bbox": [ + 304, + 72, + 547, + 191 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 304, + 72, + 547, + 191 + ], + "type": "text", + "content": " [25] (i.e., no task labels are provided to the learner during inference). Task-incremental continual learning is a simpler multi-task setting where the task labels are given during both training and inference." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 193, + 547, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 193, + 547, + 300 + ], + "spans": [ + { + "bbox": [ + 304, + 193, + 547, + 300 + ], + "type": "text", + "content": "We also include experiments that contain both class-incremental and domain-incremental continual learning at the same time. The difference between this setting and the former is that we also inject covariate distribution shifts (e.g., task 1 might include real images and clip-art, whereas task 2 might include corrupted photos and artistic paintings). The motivation is to make continual learning more practical - as in the real world we might encounter these dual types of domain shifts as well." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 311, + 471, + 324 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 311, + 471, + 324 + ], + "spans": [ + { + "bbox": [ + 305, + 311, + 471, + 324 + ], + "type": "text", + "content": "3.2. Prompting with Prefix-Tuning" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 331, + 547, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 331, + 547, + 426 + ], + "spans": [ + { + "bbox": [ + 304, + 331, + 547, + 426 + ], + "type": "text", + "content": "In our work, we do not change the technical foundations of prompting from the prior state-of-the-art DualPrompt [65]. We instead focus on the selection and formation of the prompt (including enabling an end-to-end optimization of all prompting components), and the resulting prompt is used by the vision transformer encoder in an identical fashion to DualPrompt. This allows us to make a fair comparison with respect to our novel contributions." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 427, + 547, + 583 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 427, + 547, + 583 + ], + "spans": [ + { + "bbox": [ + 304, + 427, + 547, + 583 + ], + "type": "text", + "content": "As done in DualPrompt, we pass prompts to several multi-head self-attention (MSA) layers in a pre-trained ViT transformer [13, 62] and use prefix-tuning over prompt-tuning" + }, + { + "bbox": [ + 304, + 427, + 547, + 583 + ], + "type": "inline_equation", + "content": "^2" + }, + { + "bbox": [ + 304, + 427, + 547, + 583 + ], + "type": "text", + "content": ", which prepends prompts to the keys and values of an MSA layer rather than prepending prompts to the input tokens. The layers in which prompting occurs is set as a hyperparameter, and the prompting parameters are unique between layers. We define a prompt parameter as " + }, + { + "bbox": [ + 304, + 427, + 547, + 583 + ], + "type": "inline_equation", + "content": "\\pmb{p} \\in \\mathbb{R}^{L_p \\times D}" + }, + { + "bbox": [ + 304, + 427, + 547, + 583 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 304, + 427, + 547, + 583 + ], + "type": "inline_equation", + "content": "L_{p}" + }, + { + "bbox": [ + 304, + 427, + 547, + 583 + ], + "type": "text", + "content": " is the prompt length (chosen as a hyperparameter) and " + }, + { + "bbox": [ + 304, + 427, + 547, + 583 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 304, + 427, + 547, + 583 + ], + "type": "text", + "content": " is the embedding dimension (768). Consider an MSA layer with input " + }, + { + "bbox": [ + 304, + 427, + 547, + 583 + ], + "type": "inline_equation", + "content": "\\pmb{h} \\in \\mathbb{R}^{L \\times D}" + }, + { + "bbox": [ + 304, + 427, + 547, + 583 + ], + "type": "text", + "content": ", and query, key, and values given as " + }, + { + "bbox": [ + 304, + 427, + 547, + 583 + ], + "type": "inline_equation", + "content": "\\pmb{h}_{Q}, \\pmb{h}_{K}, \\pmb{h}_{V}" + }, + { + "bbox": [ + 304, + 427, + 547, + 583 + ], + "type": "text", + "content": ", respectively. In the ViT model, " + }, + { + "bbox": [ + 304, + 427, + 547, + 583 + ], + "type": "inline_equation", + "content": "\\pmb{h}_{Q} = \\pmb{h}_{K} = \\pmb{h}_{V} = \\pmb{h}" + }, + { + "bbox": [ + 304, + 427, + 547, + 583 + ], + "type": "text", + "content": ". The output of this layer is given as:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 315, + 594, + 545, + 631 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 594, + 545, + 631 + ], + "spans": [ + { + "bbox": [ + 315, + 594, + 545, + 631 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\operatorname {M S A} \\left(\\boldsymbol {h} _ {Q}, \\boldsymbol {h} _ {K}, \\boldsymbol {h} _ {V}\\right) = \\operatorname {C o n c a t} \\left(\\mathrm {h} _ {1}, \\dots , \\mathrm {h} _ {\\mathrm {m}}\\right) W ^ {O} \\\\ \\text {w h e r e} \\mathrm {h} _ {\\mathrm {i}} = \\text {A t t e n t i o n} \\left(\\boldsymbol {h} _ {Q} W _ {i} ^ {Q}, \\boldsymbol {h} _ {K} W _ {i} ^ {K}, \\boldsymbol {h} _ {V} W _ {i} ^ {V}\\right) \\tag {1} \\\\ \\end{array}", + "image_path": "87b13d99145219a7507a1105c538cc3ad1b13a51b1a9de7c93d3eeabccd19b0d.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 641, + 547, + 682 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 641, + 547, + 682 + ], + "spans": [ + { + "bbox": [ + 306, + 641, + 547, + 682 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 306, + 641, + 547, + 682 + ], + "type": "inline_equation", + "content": "W^{O}, W_{i}^{Q}, W_{i}^{K}" + }, + { + "bbox": [ + 306, + 641, + 547, + 682 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 306, + 641, + 547, + 682 + ], + "type": "inline_equation", + "content": "W_{i}^{V}" + }, + { + "bbox": [ + 306, + 641, + 547, + 682 + ], + "type": "text", + "content": " are projection matrices and " + }, + { + "bbox": [ + 306, + 641, + 547, + 682 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 306, + 641, + 547, + 682 + ], + "type": "text", + "content": " is the number of heads. We split our prompt " + }, + { + "bbox": [ + 306, + 641, + 547, + 682 + ], + "type": "inline_equation", + "content": "\\pmb{p}" + }, + { + "bbox": [ + 306, + 641, + 547, + 682 + ], + "type": "text", + "content": " into " + }, + { + "bbox": [ + 306, + 641, + 547, + 682 + ], + "type": "inline_equation", + "content": "\\{\\pmb{p}_K, \\pmb{p}_V\\} \\in \\mathbb{R}^{\\frac{L_p}{2} \\times D}" + }, + { + "bbox": [ + 306, + 641, + 547, + 682 + ], + "type": "text", + "content": " and prepend them to " + }, + { + "bbox": [ + 306, + 641, + 547, + 682 + ], + "type": "inline_equation", + "content": "h_K" + }, + { + "bbox": [ + 306, + 641, + 547, + 682 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 306, + 641, + 547, + 682 + ], + "type": "inline_equation", + "content": "h_V" + }, + { + "bbox": [ + 306, + 641, + 547, + 682 + ], + "type": "text", + "content": " with" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 306, + 693, + 547, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 693, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 693, + 547, + 713 + ], + "type": "text", + "content": "2We refer the reader to sections 4.2 and 5.4 of the the DualPrompt [65] paper for a discussion on this design choice." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "11911" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 141, + 85 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 141, + 85 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 141, + 85 + ], + "type": "text", + "content": "prefix-tuning (P-T) as:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 64, + 90, + 287, + 105 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 90, + 287, + 105 + ], + "spans": [ + { + "bbox": [ + 64, + 90, + 287, + 105 + ], + "type": "interline_equation", + "content": "f _ {P - T} (\\boldsymbol {p}, \\boldsymbol {h}) = \\operatorname {M S A} \\left(\\boldsymbol {h} _ {Q}, \\left[ \\boldsymbol {p} _ {K}; \\boldsymbol {h} _ {K} \\right], \\left[ \\boldsymbol {p} _ {V}; \\boldsymbol {h} _ {V} \\right]\\right) \\tag {2}", + "image_path": "dffd4a2f84deed5a22f4d8e88281e6ee54ea0478bb1f3b97c5629fd62cdf35fb.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 45, + 110, + 287, + 182 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 110, + 287, + 182 + ], + "spans": [ + { + "bbox": [ + 45, + 110, + 287, + 182 + ], + "type": "text", + "content": "The result is that we now only train a small number of parameter (the prompts) while leaving the rest of the transformer encoder unmodified. The critical question that now remains is how to select and update prompts in a continuous fashion? The next subsection will discuss how prior works select and update prompts." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 188, + 171, + 202 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 188, + 171, + 202 + ], + "spans": [ + { + "bbox": [ + 47, + 188, + 171, + 202 + ], + "type": "text", + "content": "3.3. L2P and DualPrompt" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 45, + 206, + 289, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 206, + 289, + 411 + ], + "spans": [ + { + "bbox": [ + 45, + 206, + 289, + 411 + ], + "type": "text", + "content": "L2P [66] and DualPrompt [65] select prompts from a pool using an image-wise prompt query. Specifically, these methods use a key-value pair based query strategy to dynamically select instance-specific prompts from a pool of candidate prompts. Each prompt " + }, + { + "bbox": [ + 45, + 206, + 289, + 411 + ], + "type": "inline_equation", + "content": "\\pmb{p}_m" + }, + { + "bbox": [ + 45, + 206, + 289, + 411 + ], + "type": "text", + "content": " is selected from a pool of learnable keys " + }, + { + "bbox": [ + 45, + 206, + 289, + 411 + ], + "type": "inline_equation", + "content": "\\pmb{k}_m \\in \\mathbb{R}^D" + }, + { + "bbox": [ + 45, + 206, + 289, + 411 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 45, + 206, + 289, + 411 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 45, + 206, + 289, + 411 + ], + "type": "text", + "content": " is the size of the prompt pool, based on cosine similarity to an input-conditioned query. Queries are produced as: " + }, + { + "bbox": [ + 45, + 206, + 289, + 411 + ], + "type": "inline_equation", + "content": "q(\\pmb{x}) \\in \\mathbb{R}^D = \\theta(\\pmb{x})" + }, + { + "bbox": [ + 45, + 206, + 289, + 411 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 45, + 206, + 289, + 411 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 45, + 206, + 289, + 411 + ], + "type": "text", + "content": " is the pretrained ViT encoder and " + }, + { + "bbox": [ + 45, + 206, + 289, + 411 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 45, + 206, + 289, + 411 + ], + "type": "text", + "content": " is the input image4. A query " + }, + { + "bbox": [ + 45, + 206, + 289, + 411 + ], + "type": "inline_equation", + "content": "q(\\pmb{x})" + }, + { + "bbox": [ + 45, + 206, + 289, + 411 + ], + "type": "text", + "content": " is matched to a key " + }, + { + "bbox": [ + 45, + 206, + 289, + 411 + ], + "type": "inline_equation", + "content": "\\pmb{k}_m" + }, + { + "bbox": [ + 45, + 206, + 289, + 411 + ], + "type": "text", + "content": " by selecting the key with the highest cosine similarity, denoted as: " + }, + { + "bbox": [ + 45, + 206, + 289, + 411 + ], + "type": "inline_equation", + "content": "\\gamma(q(\\pmb{x}), \\pmb{k}_m)" + }, + { + "bbox": [ + 45, + 206, + 289, + 411 + ], + "type": "text", + "content": ". The keys are learned with a separate optimization from the task classification loss by simply maximizing the cosine similarity between each matched " + }, + { + "bbox": [ + 45, + 206, + 289, + 411 + ], + "type": "inline_equation", + "content": "q(\\pmb{x})" + }, + { + "bbox": [ + 45, + 206, + 289, + 411 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 45, + 206, + 289, + 411 + ], + "type": "inline_equation", + "content": "\\pmb{k}_m" + }, + { + "bbox": [ + 45, + 206, + 289, + 411 + ], + "type": "text", + "content": ", acting as a clustering-like approach. However, this approach does not directly update the key with gradients from the task loss, which is critical in our experiments." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 420, + 104, + 431 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 420, + 104, + 431 + ], + "spans": [ + { + "bbox": [ + 47, + 420, + 104, + 431 + ], + "type": "text", + "content": "4. Method" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 439, + 158, + 452 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 439, + 158, + 452 + ], + "spans": [ + { + "bbox": [ + 47, + 439, + 158, + 452 + ], + "type": "text", + "content": "4.1. Prompt Formation" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 45, + 457, + 287, + 649 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 457, + 287, + 649 + ], + "spans": [ + { + "bbox": [ + 45, + 457, + 287, + 649 + ], + "type": "text", + "content": "While prompting has been shown to perform exceptionally well in terms of mitigating catastrophic forgetting, the existing state-of-the-art approach DualPrompt lacks the ability to expand learning capacity within a given task. Specifically, DualPrompt learns a single prompt for each new task - regardless if the task is easy (such as learning a handful of new object classes) or complex (such as learning a large number of new object classes). One might try to increase the length of the prompt, but we show in our experiments (Section 5.3) that increasing the prompt length has saturated returns. Intuitively, we instead desire a method where the learning capacity is correlated to the underlying complexity of task data, rather than a single prompt. Additionally, we desire an approach that is end-to-end differentiable, which we conjecture increases the ability to learn a new task with higher accuracy." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 654, + 287, + 703 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 654, + 287, + 703 + ], + "spans": [ + { + "bbox": [ + 46, + 654, + 287, + 703 + ], + "type": "text", + "content": "3We note that the difference between L2P and DualPrompt w.r.t. prompt querying is that the size of the prompt pool in L2P is chosen as a hyperparameter, but in DualPrompt it is equal to the number of tasks and, during training, DualPrompt selects the prompt using task-id (and selects the closest key-query match during inference)." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 58, + 703, + 190, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 703, + 190, + 712 + ], + "spans": [ + { + "bbox": [ + 58, + 703, + 190, + 712 + ], + "type": "text", + "content": "4The output is read from the class token." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 72, + 545, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 203 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 203 + ], + "type": "text", + "content": "We grow our learning capacity by introducing a new axis: a set of prompt components. Rather than pick and choose prompts from a pool, we learn a set of prompt components which, via a weighted summation, form a decomposed5 prompt that is passed to the corresponding MSA layer. This allows us to grow our prompting capacity to arbitrary depth and capture the rich underlying complexity of any task while maintaining a fixed prompt length. In addition, prompting in new tasks will inherently reuse the previously acquired knowledge of past tasks rather than initializing a new task prompt from scratch." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 204, + 547, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 204, + 547, + 228 + ], + "spans": [ + { + "bbox": [ + 305, + 204, + 547, + 228 + ], + "type": "text", + "content": "Specifically, we replace learnable prompt parameter " + }, + { + "bbox": [ + 305, + 204, + 547, + 228 + ], + "type": "inline_equation", + "content": "\\pmb{p}" + }, + { + "bbox": [ + 305, + 204, + 547, + 228 + ], + "type": "text", + "content": " with a weighted summation over the prompt components:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 391, + 236, + 545, + 259 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 391, + 236, + 545, + 259 + ], + "spans": [ + { + "bbox": [ + 391, + 236, + 545, + 259 + ], + "type": "interline_equation", + "content": "\\boldsymbol {p} = \\sum_ {m} \\alpha_ {m} \\boldsymbol {P} _ {m} \\tag {3}", + "image_path": "698c3cb1bc9bfeccf3d23c7f636503ea7fce9f993e63d33038682503fd61b102.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 268, + 545, + 327 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 268, + 545, + 327 + ], + "spans": [ + { + "bbox": [ + 304, + 268, + 545, + 327 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 268, + 545, + 327 + ], + "type": "inline_equation", + "content": "P \\in \\mathbb{R}^{L_p \\times D \\times M}" + }, + { + "bbox": [ + 304, + 268, + 545, + 327 + ], + "type": "text", + "content": " is our set of prompt components, " + }, + { + "bbox": [ + 304, + 268, + 545, + 327 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 304, + 268, + 545, + 327 + ], + "type": "text", + "content": " is the length of our set (i.e., the introduced axis of additional capacity), and " + }, + { + "bbox": [ + 304, + 268, + 545, + 327 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 304, + 268, + 545, + 327 + ], + "type": "text", + "content": " is a weighting vector that determines the contribution of each component which will be discussed in the next subsection." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 329, + 545, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 329, + 545, + 411 + ], + "spans": [ + { + "bbox": [ + 304, + 329, + 545, + 411 + ], + "type": "text", + "content": "A distinct difference between our method and L2P/DualPrompt is that all of our learnable parameters are optimized using the classification loss rather than splitting into two separate loss optimizations. This allows for better end-to-end learning in our method, and we argue this is a key reason why our method performs better in the benchmarks described in Section 5." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 419, + 474, + 433 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 419, + 474, + 433 + ], + "spans": [ + { + "bbox": [ + 306, + 419, + 474, + 433 + ], + "type": "text", + "content": "4.2. Prompt-Component Weighting" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 437, + 545, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 437, + 545, + 498 + ], + "spans": [ + { + "bbox": [ + 304, + 437, + 545, + 498 + ], + "type": "text", + "content": "Rather than a pool of prompts, we have a set of prompt components and desire to produce a weight vector " + }, + { + "bbox": [ + 304, + 437, + 545, + 498 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 304, + 437, + 545, + 498 + ], + "type": "text", + "content": " given a query " + }, + { + "bbox": [ + 304, + 437, + 545, + 498 + ], + "type": "inline_equation", + "content": "\\theta(x)" + }, + { + "bbox": [ + 304, + 437, + 545, + 498 + ], + "type": "text", + "content": ", rather than a prompt index. We compute the cosine similarity between a query and keys to produce our weighting vector as:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 310, + 506, + 545, + 534 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 310, + 506, + 545, + 534 + ], + "spans": [ + { + "bbox": [ + 310, + 506, + 545, + 534 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\alpha = \\gamma (q (\\boldsymbol {x}), \\boldsymbol {K}) \\\\ = \\left\\{\\gamma (q (\\boldsymbol {x}), \\boldsymbol {K} _ {1}), \\gamma (q (\\boldsymbol {x}), \\boldsymbol {K} _ {2}), \\dots , \\gamma (q (\\boldsymbol {x}), \\boldsymbol {K} _ {M}) \\right\\} \\tag {4} \\\\ \\end{array}", + "image_path": "c602d2d9b5fc790e63072d9daa48913400fdc564c146953af52d72c62ef76249.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 542, + 545, + 602 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 542, + 545, + 602 + ], + "spans": [ + { + "bbox": [ + 304, + 542, + 545, + 602 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 542, + 545, + 602 + ], + "type": "inline_equation", + "content": "\\pmb{K} \\in \\mathbb{R}^{D \\times M}" + }, + { + "bbox": [ + 304, + 542, + 545, + 602 + ], + "type": "text", + "content": " contains keys corresponding to our prompt components. The intuition here is that the contributing magnitude of each prompt component " + }, + { + "bbox": [ + 304, + 542, + 545, + 602 + ], + "type": "inline_equation", + "content": "\\pmb{P}_{m}" + }, + { + "bbox": [ + 304, + 542, + 545, + 602 + ], + "type": "text", + "content": " to the final prompt " + }, + { + "bbox": [ + 304, + 542, + 545, + 602 + ], + "type": "inline_equation", + "content": "\\pmb{p}" + }, + { + "bbox": [ + 304, + 542, + 545, + 602 + ], + "type": "text", + "content": " is determined by the similarity between the query " + }, + { + "bbox": [ + 304, + 542, + 545, + 602 + ], + "type": "inline_equation", + "content": "q(\\pmb{x})" + }, + { + "bbox": [ + 304, + 542, + 545, + 602 + ], + "type": "text", + "content": " and corresponding key " + }, + { + "bbox": [ + 304, + 542, + 545, + 602 + ], + "type": "inline_equation", + "content": "\\pmb{K}_{m}" + }, + { + "bbox": [ + 304, + 542, + 545, + 602 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 603, + 545, + 686 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 603, + 545, + 686 + ], + "spans": [ + { + "bbox": [ + 304, + 603, + 545, + 686 + ], + "type": "text", + "content": "The prompt-query matching can be thought of as clustering in a high dimension space " + }, + { + "bbox": [ + 304, + 603, + 545, + 686 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 304, + 603, + 545, + 686 + ], + "type": "text", + "content": " which is a well-known and difficult problem [31]. To mitigate this drawback, we introduce another component to our key-query matching: attention. Each " + }, + { + "bbox": [ + 304, + 603, + 545, + 686 + ], + "type": "inline_equation", + "content": "P_{m}" + }, + { + "bbox": [ + 304, + 603, + 545, + 686 + ], + "type": "text", + "content": " has a corresponding attention vector " + }, + { + "bbox": [ + 304, + 603, + 545, + 686 + ], + "type": "inline_equation", + "content": "A_{m}" + }, + { + "bbox": [ + 304, + 603, + 545, + 686 + ], + "type": "text", + "content": " in addition to the key " + }, + { + "bbox": [ + 304, + 603, + 545, + 686 + ], + "type": "inline_equation", + "content": "K_{m}" + }, + { + "bbox": [ + 304, + 603, + 545, + 686 + ], + "type": "text", + "content": ". This allows the query to focus on certain features from the high dimensional query" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 306, + 693, + 545, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 693, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 693, + 545, + 713 + ], + "type": "text", + "content": "5 Decomposed in the sense that our prompt is a weighted summation of components." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "11912" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 60, + 75, + 534, + 228 + ], + "blocks": [ + { + "bbox": [ + 60, + 75, + 534, + 228 + ], + "lines": [ + { + "bbox": [ + 60, + 75, + 534, + 228 + ], + "spans": [ + { + "bbox": [ + 60, + 75, + 534, + 228 + ], + "type": "image", + "image_path": "4d37f8d1134de127a55dcc201666115e9e5d18d45881b047a51a9426b060cab2.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 235, + 545, + 313 + ], + "lines": [ + { + "bbox": [ + 46, + 235, + 545, + 313 + ], + "spans": [ + { + "bbox": [ + 46, + 235, + 545, + 313 + ], + "type": "text", + "content": "Figure 2. Our CODA-Prompt approach to rehearsal-free continual learning. An input (e.g., image) is passed through a query function (we use the pretrained ViT encoder) and used for a novel attention-based prompt-selection scheme to produce prompts which are passed to multiple layers of a large-scale, pre-trained transformer. Our method is parameterized by an expanding set of small prompt components, with each prompt having a corresponding key and attention vector. Both the input and produced prompts are passed through the transformer and sent to the task head. Only the prompting and task parameters are learned which is parameter efficient and no training data is stored for replay which is memory-efficient and privacy preserving. Importantly, our prompt selection scheme can be optimized end-to-end, unlike prior works which have a local similarity-based optimization for key-query matching." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 321, + 290, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 321, + 290, + 453 + ], + "spans": [ + { + "bbox": [ + 46, + 321, + 290, + 453 + ], + "type": "inline_equation", + "content": "q(\\pmb{x})" + }, + { + "bbox": [ + 46, + 321, + 290, + 453 + ], + "type": "text", + "content": " output - for example, a prompt component used for mammalian face features to classify dogs and cats would be able to focus on relevant query features such as eye placement and skin texture, and furthermore ignore features of unrelated semantics, such as features useful for automobile classification. We use a simple feature-selection attention scheme with element-wise multiplication between the query vector and attention vector to create an attended-query which is then used for the key-similarity matching. Specifically, our updated approach to producing our weighting vector is:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 52, + 460, + 287, + 488 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 52, + 460, + 287, + 488 + ], + "spans": [ + { + "bbox": [ + 52, + 460, + 287, + 488 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\begin{array}{l} \\alpha = \\gamma (q (\\boldsymbol {x}) \\odot \\boldsymbol {A}, \\boldsymbol {K}) \\\\ = \\left\\{\\gamma (q (\\boldsymbol {x}) \\odot \\boldsymbol {A}, \\boldsymbol {K}), \\quad \\gamma (q (\\boldsymbol {x}) \\odot \\boldsymbol {A}, \\boldsymbol {K}), \\right\\} \\end{array} \\tag {5} \\\\ = \\left\\{\\gamma (q (\\boldsymbol {x}) \\odot \\boldsymbol {A} _ {1}, \\boldsymbol {K} _ {1}), \\dots , \\gamma (q (\\boldsymbol {x}) \\odot \\boldsymbol {A} _ {M}, \\boldsymbol {K} _ {M}) \\right\\} \\\\ \\end{array}", + "image_path": "b9e9f75c8c538f9e88ad4022b3b0807a33ece22d0e88fb07ea7f074e0dc13975.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 496, + 287, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 496, + 287, + 593 + ], + "spans": [ + { + "bbox": [ + 46, + 496, + 287, + 593 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 496, + 287, + 593 + ], + "type": "inline_equation", + "content": "\\mathbf{A} \\in \\mathbb{R}^{D \\times M}" + }, + { + "bbox": [ + 46, + 496, + 287, + 593 + ], + "type": "text", + "content": " contains learnable parameters (attention vectors) corresponding to our prompt components and " + }, + { + "bbox": [ + 46, + 496, + 287, + 593 + ], + "type": "inline_equation", + "content": "\\odot" + }, + { + "bbox": [ + 46, + 496, + 287, + 593 + ], + "type": "text", + "content": " is the element-wise product (also known as the Hadamard product). Notice that our attention vectors are learnable feature weightings rather than input-conditioned modules - we found that this fixed representation was less vulnerable to forgetting by its simple design, similar to the our prompt component keys." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 599, + 200, + 613 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 599, + 200, + 613 + ], + "spans": [ + { + "bbox": [ + 47, + 599, + 200, + 613 + ], + "type": "text", + "content": "4.3. Expansion & Orthogonality" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 617, + 287, + 716 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 617, + 287, + 716 + ], + "spans": [ + { + "bbox": [ + 46, + 617, + 287, + 716 + ], + "type": "text", + "content": "Crucial to mitigating catastrophic forgetting is the avoidance of overwriting knowledge acquired in prior tasks. When we visit a new task we freeze our current components and expand the set, only updating the new components. This is visualized in the bottom of Figure 2 where the existing parameters are locked and only the expanded parameters are optimized. Specifically, in task " + }, + { + "bbox": [ + 46, + 617, + 287, + 716 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 617, + 287, + 716 + ], + "type": "text", + "content": " we learn " + }, + { + "bbox": [ + 46, + 617, + 287, + 716 + ], + "type": "inline_equation", + "content": "\\frac{M}{N}" + }, + { + "bbox": [ + 46, + 617, + 287, + 716 + ], + "type": "text", + "content": " components where " + }, + { + "bbox": [ + 46, + 617, + 287, + 716 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 617, + 287, + 716 + ], + "type": "text", + "content": " is the number of tasks and " + }, + { + "bbox": [ + 46, + 617, + 287, + 716 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 46, + 617, + 287, + 716 + ], + "type": "text", + "content": " is a" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 320, + 545, + 429 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 320, + 545, + 429 + ], + "spans": [ + { + "bbox": [ + 304, + 320, + 545, + 429 + ], + "type": "text", + "content": "chosen hyperparameter, keeping the previous " + }, + { + "bbox": [ + 304, + 320, + 545, + 429 + ], + "type": "inline_equation", + "content": "\\frac{(t - 1)\\cdot M}{N}" + }, + { + "bbox": [ + 304, + 320, + 545, + 429 + ], + "type": "text", + "content": " components frozen. This expansion is enabled by our attention-based component-weighting scheme, in that expanding our parameters do not alter the computation of weights " + }, + { + "bbox": [ + 304, + 320, + 545, + 429 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 304, + 320, + 545, + 429 + ], + "type": "text", + "content": " corresponding previously learned components. For example, expanding the capacity of a nonlinear model like an MLP module, transformer attention head [13], or hypernetwork [63] would alter the weight outputs of previously learned components." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 430, + 546, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 430, + 546, + 559 + ], + "spans": [ + { + "bbox": [ + 304, + 430, + 546, + 559 + ], + "type": "text", + "content": "While the prior heuristic is focused on reducing catastrophic forgetting (not destroying previously acquired knowledge), we also want to reduce interference between existing and new knowledge. To do this, we add an orthogonality constraint to " + }, + { + "bbox": [ + 304, + 430, + 546, + 559 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 304, + 430, + 546, + 559 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 430, + 546, + 559 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 304, + 430, + 546, + 559 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 430, + 546, + 559 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 304, + 430, + 546, + 559 + ], + "type": "text", + "content": ". The intuition is that orthogonal vectors have less interference with each other. For example, we would want a key and prompt learned in task 2 to not attract task 1 data, otherwise the encoder output for task 1 data will change, resulting in misclassified data. We introduce a simple orthogonality penalty loss to our method as:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 368, + 559, + 545, + 574 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 368, + 559, + 545, + 574 + ], + "spans": [ + { + "bbox": [ + 368, + 559, + 545, + 574 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {o r t h o}} (B) = \\left| \\left| B B ^ {\\top} - I \\right| \\right| _ {2} \\tag {6}", + "image_path": "94f90131e60049001a14ac010427f30132db153d83cbf70798c9e8e482c14044.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 578, + 471, + 591 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 578, + 471, + 591 + ], + "spans": [ + { + "bbox": [ + 306, + 578, + 471, + 591 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 306, + 578, + 471, + 591 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 306, + 578, + 471, + 591 + ], + "type": "text", + "content": " represents any arbitrary matrix." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 597, + 411, + 611 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 597, + 411, + 611 + ], + "spans": [ + { + "bbox": [ + 306, + 597, + 411, + 611 + ], + "type": "text", + "content": "4.4. Full Optimization" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 317, + 616, + 545, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 616, + 545, + 628 + ], + "spans": [ + { + "bbox": [ + 317, + 616, + 545, + 628 + ], + "type": "text", + "content": "Given task classification loss " + }, + { + "bbox": [ + 317, + 616, + 545, + 628 + ], + "type": "inline_equation", + "content": "\\mathcal{L}" + }, + { + "bbox": [ + 317, + 616, + 545, + 628 + ], + "type": "text", + "content": ", our full optimization is:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 321, + 647, + 545, + 681 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 647, + 545, + 681 + ], + "spans": [ + { + "bbox": [ + 321, + 647, + 545, + 681 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\min _ {\\boldsymbol {P} ^ {n}, \\boldsymbol {K} ^ {n}, \\boldsymbol {A} ^ {n}, \\phi_ {n}} \\mathcal {L} \\left(f _ {\\phi} \\left(f _ {\\theta , \\boldsymbol {P}, \\boldsymbol {K}, \\boldsymbol {A}} (\\boldsymbol {x})\\right), y\\right) + \\tag {7} \\\\ \\lambda \\left(\\mathcal {L} _ {\\text {o r t h o}} (\\boldsymbol {P}) + \\mathcal {L} _ {\\text {o r t h o}} (\\boldsymbol {K}) + \\mathcal {L} _ {\\text {o r t h o}} (\\boldsymbol {A})\\right) \\\\ \\end{array}", + "image_path": "3ecb1d44608625e180c8309ea921a4ceed1d245a0190357344e15ab433cf2323.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "inline_equation", + "content": "P^n, K^n, A^n" + }, + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "text", + "content": " refer to the prompt components and corresponding keys/attention vectors that are unfrozen and" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "11913" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 145 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 145 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 145 + ], + "type": "text", + "content": "trained during task " + }, + { + "bbox": [ + 46, + 72, + 289, + 145 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 46, + 72, + 289, + 145 + ], + "type": "text", + "content": " (see Section 4.3) and " + }, + { + "bbox": [ + 46, + 72, + 289, + 145 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 46, + 72, + 289, + 145 + ], + "type": "text", + "content": " is a hyperparameter balancing the orthogonality loss. Note that we do not update the logits of past task classes, consistent with L2P and DualPrompt. We formally refer to our method as COntinual Decomposed Attention-based Prompting, or simply CODA-P." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 156, + 128, + 169 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 156, + 128, + 169 + ], + "spans": [ + { + "bbox": [ + 47, + 156, + 128, + 169 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 176, + 289, + 439 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 176, + 289, + 439 + ], + "spans": [ + { + "bbox": [ + 46, + 176, + 289, + 439 + ], + "type": "text", + "content": "We benchmark our approach with several image datasets in the class-incremental continual learning setting. We implemented baselines which do not store training data for rehearsal: Learning without Forgetting (LwF) [38], Learning to Prompt (L2P) [66], and DualPrompt [65]. Additionally, we report the upper bound performance (i.e., trained offline) and performance for a neural network trained only on classification loss using the new task training data (we refer to this as FT), and include an improved version of FT which uses the same classifier implementation as L2P/DualPrompt (referred to as " + }, + { + "bbox": [ + 46, + 176, + 289, + 439 + ], + "type": "inline_equation", + "content": "\\mathrm{FT} + +" + }, + { + "bbox": [ + 46, + 176, + 289, + 439 + ], + "type": "text", + "content": "). We also compare to Experience Replay [11] to provide additional context to our results. We implement our method and all baselines in PyTorch [48] using the ViT-B/16 backbone [13] pre-trained on ImageNet-1K [53]. Our contribution includes faithful PyTorch implementations of the popular prompting baselines L2P and DualPrompt, which were released in JAX [65, 66]. Our implementations of the competing methods actually achieve slight boosts in the performance of DualPrompt in most benchmarks, as well as significant performance boosts in L2P due to the improved prompting type (which we refer to as L2P++." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 440, + 289, + 608 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 440, + 289, + 608 + ], + "spans": [ + { + "bbox": [ + 46, + 440, + 289, + 608 + ], + "type": "text", + "content": "DualPrompt uses length 5 prompts in layers 1-2 (referred to as general prompts) and length 20 prompts in layers 3-5 (referred to as task-expert) prompts. We insert prompts into the same layers as DualPrompt (layers 1-5) for CODA-P and use a prompt length of 8 and 100 prompt components, which were chosen with a hyperparameter sweep on validation data to have the best trade-off between performance and parameter efficiency. Because our approach introduces more learnable parameters compared to DualPrompt, we include a variant of out method CODA-P-S which uses a smaller number of parameters equal to DualPrompt for additional comparison. We show that our method outperforms other methods even under this condition, while still retaining the capability to scale if desired." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 608, + 288, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 608, + 288, + 667 + ], + "spans": [ + { + "bbox": [ + 46, + 608, + 288, + 667 + ], + "type": "text", + "content": "We report results on the test dataset, but emphasize that all hyperparameters and design decisions (including for the baselines) were made using validation data (20% of the training data). Unlike DualPrompt, we run our benchmarks for several different shuffles of the task class order and re" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 306, + 72, + 545, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 545, + 156 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 545, + 156 + ], + "type": "text", + "content": "port the mean and standard deviation of these runs. We do this with a consistent seed (different for each trial) so that results can be directly compared. This is crucial because the results with different shuffling order can be different due to differences in task difficulty and their appearance order. We include additional implementation details and results in the Supplementary Material (SM)." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 305, + 157, + 547, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 157, + 547, + 266 + ], + "spans": [ + { + "bbox": [ + 305, + 157, + 547, + 266 + ], + "type": "text", + "content": "Evaluation Metrics: We evaluate methods using (1) average final accuracy " + }, + { + "bbox": [ + 305, + 157, + 547, + 266 + ], + "type": "inline_equation", + "content": "A_N" + }, + { + "bbox": [ + 305, + 157, + 547, + 266 + ], + "type": "text", + "content": ", or the final accuracy with respect to all past classes averaged over " + }, + { + "bbox": [ + 305, + 157, + 547, + 266 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 305, + 157, + 547, + 266 + ], + "type": "text", + "content": " tasks, and (2) average forgetting [9, 41, 42] " + }, + { + "bbox": [ + 305, + 157, + 547, + 266 + ], + "type": "inline_equation", + "content": "F_N" + }, + { + "bbox": [ + 305, + 157, + 547, + 266 + ], + "type": "text", + "content": ", or the drop in task performance averaged over " + }, + { + "bbox": [ + 305, + 157, + 547, + 266 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 305, + 157, + 547, + 266 + ], + "type": "text", + "content": " tasks. The reader is referred to our SM for additional details about these metrics. We emphasize that " + }, + { + "bbox": [ + 305, + 157, + 547, + 266 + ], + "type": "inline_equation", + "content": "A_N" + }, + { + "bbox": [ + 305, + 157, + 547, + 266 + ], + "type": "text", + "content": " is the more important metric and encompasses both method plasticity and forgetting, whereas " + }, + { + "bbox": [ + 305, + 157, + 547, + 266 + ], + "type": "inline_equation", + "content": "F_N" + }, + { + "bbox": [ + 305, + 157, + 547, + 266 + ], + "type": "text", + "content": " simply provides additional context." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 274, + 545, + 288 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 274, + 545, + 288 + ], + "spans": [ + { + "bbox": [ + 305, + 274, + 545, + 288 + ], + "type": "text", + "content": "5.1. CODA-P sets the SOTA on existing benchmarks" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 294, + 545, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 294, + 545, + 582 + ], + "spans": [ + { + "bbox": [ + 304, + 294, + 545, + 582 + ], + "type": "text", + "content": "We first evaluate our method and state of art on several well-established benchmarks. Table 1 provides results for ImageNet-R [21, 65] which is composed of 200 object classes with a wide collection of image styles, including cartoon, graffiti, and hard examples from the original ImageNet dataset [53]. This benchmark is attractive because the distribution of training data has significant distance to the pre-training data (ImageNet), thus providing a fair and challenging problem setting. In addition to the original 10-task benchmark, we also provide results with a smaller number of large tasks (5-task) and a larger number of small tasks (20 task). We first notice that our reported performance for DualPrompt is a few percentage points higher compared to the original DualPrompt [65] paper. We reimplemented the method from scratch in a different framework (PyTorch [48]) and suspect the difference has to do with our averaging over different class-order shuffling. We note that our implementation of L2P " + }, + { + "bbox": [ + 304, + 294, + 545, + 582 + ], + "type": "inline_equation", + "content": "(\\mathrm{L}2\\mathrm{P}++)" + }, + { + "bbox": [ + 304, + 294, + 545, + 582 + ], + "type": "text", + "content": " performs significantly better than originally reported because we use the same form of prompting as DualPrompt (for fair comparison). Furthermore, our \"Deep L2P " + }, + { + "bbox": [ + 304, + 294, + 545, + 582 + ], + "type": "inline_equation", + "content": "+ + \"" + }, + { + "bbox": [ + 304, + 294, + 545, + 582 + ], + "type": "text", + "content": ", which prompts at the same 5 layers as DualPrompt (rather than only at layer 1), actually performs similarly to DualPrompt. See supplementary for discussion and figures illustrating this." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 582, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 582, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 582, + 547, + 715 + ], + "type": "text", + "content": "Importantly, we see that our method has strong gains in average accuracy across all three task lengths, with as much as " + }, + { + "bbox": [ + 304, + 582, + 547, + 715 + ], + "type": "inline_equation", + "content": "4.5\\%" + }, + { + "bbox": [ + 304, + 582, + 547, + 715 + ], + "type": "text", + "content": " improvement in average accuracy over DualPrompt. We remind the reader that CODA-P is our proposed method with tuned prompt length and component set size, whereas CODA-P-S is downsized (see SM for exact details) to exactly match the number of learnable parameters as DualPrompt. We notice that our method often has slightly higher average forgetting compared to DualPrompt. Given that average accuracy is the crucial metric that captures practical performance, we" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 58, + 673, + 170, + 683 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 673, + 170, + 683 + ], + "spans": [ + { + "bbox": [ + 58, + 673, + 170, + 683 + ], + "type": "text", + "content": "See the SM for additional details" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 683, + 287, + 714 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 683, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 683, + 287, + 714 + ], + "type": "text", + "content": "7As discussed in Section 3.2, we use pre-fix tuning over prompt-tuning for all implementations as is shown to work best for continual learning [65]." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "11914" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 51, + 101, + 541, + 253 + ], + "blocks": [ + { + "bbox": [ + 47, + 70, + 545, + 93 + ], + "lines": [ + { + "bbox": [ + 47, + 70, + 545, + 93 + ], + "spans": [ + { + "bbox": [ + 47, + 70, + 545, + 93 + ], + "type": "text", + "content": "Table 1. Results (\\%) on ImageNet-R. Results are included for 5 tasks (40 classes per task), 10 tasks (20 classes per task), and 20 tasks (10 classes per task). " + }, + { + "bbox": [ + 47, + 70, + 545, + 93 + ], + "type": "inline_equation", + "content": "A_N" + }, + { + "bbox": [ + 47, + 70, + 545, + 93 + ], + "type": "text", + "content": " gives the accuracy averaged over tasks and " + }, + { + "bbox": [ + 47, + 70, + 545, + 93 + ], + "type": "inline_equation", + "content": "F_N" + }, + { + "bbox": [ + 47, + 70, + 545, + 93 + ], + "type": "text", + "content": " gives the average forgetting. We report results over 5 trials." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 51, + 101, + 541, + 253 + ], + "lines": [ + { + "bbox": [ + 51, + 101, + 541, + 253 + ], + "spans": [ + { + "bbox": [ + 51, + 101, + 541, + 253 + ], + "type": "table", + "html": "
Tasks51020
MethodAN(↑)FN(↓)AN(↑)FN(↓)AN(↑)FN(↓)
UB77.13-77.13-77.13-
ER (5000)71.72 ± 0.7113.70 ± 0.2664.43 ± 1.1610.30 ± 0.0552.43 ± 0.877.70 ± 0.13
FT18.74 ± 0.4441.49 ± 0.5210.12 ± 0.5125.69 ± 0.234.75 ± 0.4016.34 ± 0.19
FT++60.42 ± 0.8714.66 ± 0.2448.93 ± 1.159.81 ± 0.3135.98 ± 1.386.63 ± 0.11
LwF.MC74.56 ± 0.594.98 ± 0.3766.73 ± 1.253.52 ± 0.3954.05 ± 2.662.86 ± 0.26
L2P++70.83 ± 0.583.36 ± 0.1869.29 ± 0.732.03 ± 0.1965.89 ± 1.301.24 ± 0.14
Deep L2P++73.93 ± 0.372.69 ± 0.1071.66 ± 0.641.78 ± 0.1668.42 ± 1.201.12 ± 0.13
DualPrompt73.05 ± 0.502.64 ± 0.1771.32 ± 0.621.71 ± 0.2467.87 ± 1.391.07 ± 0.14
CODA-P-S75.19 ± 0.472.65 ± 0.1573.93 ± 0.491.60 ± 0.2070.53 ± 1.241.00 ± 0.15
CODA-P76.51 ± 0.382.99 ± 0.1975.45 ± 0.561.64 ± 0.1072.37 ± 1.190.96 ± 0.15
", + "image_path": "18504fcbac243d3e7e92a242435b566eeef3b7d264cca8d558592f1cf28c6f18.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 66, + 297, + 276, + 435 + ], + "blocks": [ + { + "bbox": [ + 47, + 259, + 544, + 281 + ], + "lines": [ + { + "bbox": [ + 47, + 259, + 544, + 281 + ], + "spans": [ + { + "bbox": [ + 47, + 259, + 544, + 281 + ], + "type": "text", + "content": "Table 2. Results (%) on CIFAR-100 and DomainNet. " + }, + { + "bbox": [ + 47, + 259, + 544, + 281 + ], + "type": "inline_equation", + "content": "A_N" + }, + { + "bbox": [ + 47, + 259, + 544, + 281 + ], + "type": "text", + "content": " gives the accuracy averaged over tasks and " + }, + { + "bbox": [ + 47, + 259, + 544, + 281 + ], + "type": "inline_equation", + "content": "F_N" + }, + { + "bbox": [ + 47, + 259, + 544, + 281 + ], + "type": "text", + "content": " gives the average forgetting. We report results over 5 trials for CIFAR-100 and 3 trials for DomainNet (due to dataset size)." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 99, + 285, + 241, + 295 + ], + "lines": [ + { + "bbox": [ + 99, + 285, + 241, + 295 + ], + "spans": [ + { + "bbox": [ + 99, + 285, + 241, + 295 + ], + "type": "text", + "content": "(a) 10-task CIFAR-100 (10 classes per task)" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 66, + 297, + 276, + 435 + ], + "lines": [ + { + "bbox": [ + 66, + 297, + 276, + 435 + ], + "spans": [ + { + "bbox": [ + 66, + 297, + 276, + 435 + ], + "type": "table", + "html": "
MethodAN(↑)FN(↓)
Upper-Bound89.30-
ER (5000)76.20 ± 1.048.50 ± 0.37
FT9.92 ± 0.2729.21 ± 0.18
FT++49.91 ± 0.4212.30 ± 0.23
LwF.MC64.83 ± 1.035.27 ± 0.39
L2P++82.50 ± 1.101.75 ± 0.42
Deep L2P++84.30 ± 1.031.53 ± 0.40
DualPrompt83.05 ± 1.161.72 ± 0.40
CODA-P-S84.59 ± 0.871.76 ± 0.28
CODA-P86.25 ± 0.741.67 ± 0.26
", + "image_path": "961f1c4dc6c1f2a18876ec0c70cc1e925e6bc79285b6e160ba38edadb2a69f09.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 318, + 297, + 528, + 435 + ], + "blocks": [ + { + "bbox": [ + 353, + 285, + 492, + 295 + ], + "lines": [ + { + "bbox": [ + 353, + 285, + 492, + 295 + ], + "spans": [ + { + "bbox": [ + 353, + 285, + 492, + 295 + ], + "type": "text", + "content": "(b) 5-task DomainNet (69 classes per task)" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 318, + 297, + 528, + 435 + ], + "lines": [ + { + "bbox": [ + 318, + 297, + 528, + 435 + ], + "spans": [ + { + "bbox": [ + 318, + 297, + 528, + 435 + ], + "type": "table", + "html": "
MethodAN(↑)FN(↓)
Upper-Bound79.65-
ER (5000)58.32 ± 0.4726.25 ± 0.24
FT18.00 ± 0.2643.55 ± 0.27
FT++39.28 ± 0.2144.39 ± 0.31
LwF.MC74.78 ± 0.435.01 ± 0.14
L2P++69.58 ± 0.392.25 ± 0.08
Deep L2P++70.54 ± 0.512.05 ± 0.07
DualPrompt70.73 ± 0.492.03 ± 0.22
CODA-P-S71.80 ± 0.572.54 ± 0.10
CODA-P73.24 ± 0.593.46 ± 0.09
", + "image_path": "522e3afaf015e8444d6031a17ccb291f0b0785b085e64b94f40b1a3e02bfd7a4.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 454, + 287, + 538 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 454, + 287, + 538 + ], + "spans": [ + { + "bbox": [ + 46, + 454, + 287, + 538 + ], + "type": "text", + "content": "are not worried by the slight uptick in forgetting. In fact, we argue it is very reasonable and reflects the strength of our approach: our method has a larger capacity to learn new tasks, and thus we can sacrifice marginally higher forgetting. Crucially, we see that as the task sequence length grows, the forgetting metric of our method versus DualPrompt begins to converge to a similar value." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 539, + 287, + 610 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 539, + 287, + 610 + ], + "spans": [ + { + "bbox": [ + 46, + 539, + 287, + 610 + ], + "type": "text", + "content": "We provide results for experience replay to provide additional context of our results. While the gap between rehearsal with a substantial coreset size (i.e., the buffer size of replay data) and our approach is small for the short task sequence, we see that our method strongly outperforms rehearsal in the longer task sequence." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 611, + 287, + 673 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 611, + 287, + 673 + ], + "spans": [ + { + "bbox": [ + 46, + 611, + 287, + 673 + ], + "type": "text", + "content": "Tables 2a and 2b provide results on additional8 benchmarks ten-task CIFAR-100 [32] and five-task DomainNet [49]. While neither dataset is as impactful as ImageNet-R in terms of challenge and distance from the pre-trained distribution, these tables provide additional" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 454, + 545, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 454, + 545, + 515 + ], + "spans": [ + { + "bbox": [ + 306, + 454, + 545, + 515 + ], + "type": "text", + "content": "context to our method's performance. These tables illustrate a similar story to the ImageNet-R benchmarks, with improvements of " + }, + { + "bbox": [ + 306, + 454, + 545, + 515 + ], + "type": "inline_equation", + "content": "+3.2\\%" + }, + { + "bbox": [ + 306, + 454, + 545, + 515 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 306, + 454, + 545, + 515 + ], + "type": "inline_equation", + "content": "+2.5\\%" + }, + { + "bbox": [ + 306, + 454, + 545, + 515 + ], + "type": "text", + "content": ", respectively. We do note that, surprisingly, LwF [38] slightly outperforms prompting on this benchmark." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 525, + 545, + 549 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 525, + 545, + 549 + ], + "spans": [ + { + "bbox": [ + 306, + 525, + 545, + 549 + ], + "type": "text", + "content": "5.2. CODA-P sets the SOTA for a new dual-shift benchmark" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 557, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 557, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 557, + 546, + 713 + ], + "type": "text", + "content": "We also evaluate on a challenging dual-shift benchmark using the ImageNet-R dataset. Our motivation is to show robustness to two different types of continual distribution shifts: semantic and covariate. We accomplish this by randomly selecting image types from the ImageNet-R dataset to include in each task's training data (while keeping the evaluation data unmodified). The reader is refereed to our SM for more details. Results for this benchmark are provided in Table 3. Compared to 5-task ImageNet-R where our method improves SOTA by " + }, + { + "bbox": [ + 304, + 557, + 546, + 713 + ], + "type": "inline_equation", + "content": "4.5\\%" + }, + { + "bbox": [ + 304, + 557, + 546, + 713 + ], + "type": "text", + "content": " average accuracy, we see a " + }, + { + "bbox": [ + 304, + 557, + 546, + 713 + ], + "type": "inline_equation", + "content": "4.4\\%" + }, + { + "bbox": [ + 304, + 557, + 546, + 713 + ], + "type": "text", + "content": " improvement on this 5-task benchmark with similar forgetting margins. Our results indicate that CODA-P better generalizes to real-world type shifts." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 47, + 683, + 287, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 683, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 47, + 683, + 287, + 712 + ], + "type": "text", + "content": "Note that the hyperparameters for both DualPrompt and CODA-P were tuned on ImageNet-R, which was intentionally done to evaluate robustness of all methods." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "11915" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 61, + 156, + 272, + 294 + ], + "blocks": [ + { + "bbox": [ + 46, + 70, + 288, + 147 + ], + "lines": [ + { + "bbox": [ + 46, + 70, + 288, + 147 + ], + "spans": [ + { + "bbox": [ + 46, + 70, + 288, + 147 + ], + "type": "text", + "content": "Table 3. Results (\\%) on ImageNet-R with covariate domain shifts. Results are included for 5 tasks (40 classes per task). We simulate domain shifts by randomly removing " + }, + { + "bbox": [ + 46, + 70, + 288, + 147 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 46, + 70, + 288, + 147 + ], + "type": "text", + "content": " of the dataset's domains (e.g., clipart, paintings, and cartoon) for the training data of each task (see SM for more details). " + }, + { + "bbox": [ + 46, + 70, + 288, + 147 + ], + "type": "inline_equation", + "content": "A_N" + }, + { + "bbox": [ + 46, + 70, + 288, + 147 + ], + "type": "text", + "content": " gives the accuracy averaged over tasks and " + }, + { + "bbox": [ + 46, + 70, + 288, + 147 + ], + "type": "inline_equation", + "content": "F_N" + }, + { + "bbox": [ + 46, + 70, + 288, + 147 + ], + "type": "text", + "content": " gives the average forgetting. We report results over 5 trials." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 61, + 156, + 272, + 294 + ], + "lines": [ + { + "bbox": [ + 61, + 156, + 272, + 294 + ], + "spans": [ + { + "bbox": [ + 61, + 156, + 272, + 294 + ], + "type": "table", + "html": "
MethodAN(↑)FN(↓)
Upper-Bound77.132.66
ER (5000)67.39 ± 0.3711.94 ± 0.17
FT17.93 ± 0.2737.49 ± 0.28
FT++54.51 ± 0.6814.41 ± 0.33
LwF.MC64.02 ± 1.557.05 ± 0.27
L2P++65.08 ± 0.292.79 ± 0.32
Deep L2P++65.74 ± 0.122.48 ± 0.30
DualPrompt66.98 ± 0.082.21 ± 0.28
CODA-P-S69.73 ± 0.182.35 ± 0.19
CODA-P71.35 ± 0.082.56 ± 0.26
", + "image_path": "1bfc8d2b79099318087cab0529e71c891696e34d600913fe7173104b54262094.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 52, + 342, + 280, + 407 + ], + "blocks": [ + { + "bbox": [ + 46, + 299, + 287, + 333 + ], + "lines": [ + { + "bbox": [ + 46, + 299, + 287, + 333 + ], + "spans": [ + { + "bbox": [ + 46, + 299, + 287, + 333 + ], + "type": "text", + "content": "Table 4. Ablation Results (\\%) on 10-task ImageNet-R. " + }, + { + "bbox": [ + 46, + 299, + 287, + 333 + ], + "type": "inline_equation", + "content": "A_N" + }, + { + "bbox": [ + 46, + 299, + 287, + 333 + ], + "type": "text", + "content": " gives the accuracy averaged over tasks and " + }, + { + "bbox": [ + 46, + 299, + 287, + 333 + ], + "type": "inline_equation", + "content": "F_N" + }, + { + "bbox": [ + 46, + 299, + 287, + 333 + ], + "type": "text", + "content": " gives the average forgetting. We report results over 5 trials." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 52, + 342, + 280, + 407 + ], + "lines": [ + { + "bbox": [ + 52, + 342, + 280, + 407 + ], + "spans": [ + { + "bbox": [ + 52, + 342, + 280, + 407 + ], + "type": "table", + "html": "
MethodAN(↑)FN(↓)
CODA-P75.45 ± 0.561.64 ± 0.10
Ablate Attention74.52 ± 0.651.67 ± 0.13
Ablate Freezing74.60 ± 0.642.29 ± 0.10
Ablate Orthogonality70.66 ± 0.601.74 ± 0.25
", + "image_path": "52f6d029241dc8683481d33ec1b55a79a7dc5411cab03fdfb6e03f51c3fb34fd.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 418, + 230, + 431 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 418, + 230, + 431 + ], + "spans": [ + { + "bbox": [ + 47, + 418, + 230, + 431 + ], + "type": "text", + "content": "5.3. Ablations and Additional Analysis" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 437, + 287, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 437, + 287, + 617 + ], + "spans": [ + { + "bbox": [ + 46, + 437, + 287, + 617 + ], + "type": "text", + "content": "We take a closer look at our method with ablation studies and additional analysis. In Table 4 we show the effect of removing a few key components of our method: attention keys, freezing of past task components, and our orthogonality regularization. We show that removing attention slightly reduces forgetting degrades performance on average accuracy (the more important metric). This makes sense because removing the attention keys makes our query processing more aligned with the L2P/DualPrompt methods which boast low forgetting but lack sufficient learning capacity. We see larger drops when removing freezing and orthogonality. This indicates that these aspects are crucial to our approach. Our intuition is that, without these, our prompt formation is similar to a shallow MLP module which, unregularized, should suffer from high forgetting." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 617, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 617, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 617, + 288, + 714 + ], + "type": "text", + "content": "We also look at our ability to grow in model capacity along our newly introduced prompt component dimension using the 5-task ImageNet-R benchmark (with validation data). We show in Figure 3 that, with a number of prompt components equal to the number of prompts learned in DualPrompt, we achieve higher performance. Importantly, when we increase the number of components, we are able to leverage the scaled parameters for significantly higher" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 328, + 73, + 521, + 168 + ], + "blocks": [ + { + "bbox": [ + 328, + 73, + 521, + 168 + ], + "lines": [ + { + "bbox": [ + 328, + 73, + 521, + 168 + ], + "spans": [ + { + "bbox": [ + 328, + 73, + 521, + 168 + ], + "type": "image", + "image_path": "5833e8ab7aa279e3f28487b002243e4ec312317a46bc80612c4a1fca517349d6.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 174, + 545, + 207 + ], + "lines": [ + { + "bbox": [ + 305, + 174, + 545, + 207 + ], + "spans": [ + { + "bbox": [ + 305, + 174, + 545, + 207 + ], + "type": "text", + "content": "Figure 3. Analysis of average accuracy " + }, + { + "bbox": [ + 305, + 174, + 545, + 207 + ], + "type": "inline_equation", + "content": "A_N" + }, + { + "bbox": [ + 305, + 174, + 545, + 207 + ], + "type": "text", + "content": " vs prompt components (or pool size) for CODA-P, L2P, and DualPrompt. The setting is 5 task ImageNet-R, and we report the mean over 3 trials." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 327, + 213, + 523, + 308 + ], + "blocks": [ + { + "bbox": [ + 327, + 213, + 523, + 308 + ], + "lines": [ + { + "bbox": [ + 327, + 213, + 523, + 308 + ], + "spans": [ + { + "bbox": [ + 327, + 213, + 523, + 308 + ], + "type": "image", + "image_path": "f9367098ec6e4e37c56405c9a15b36d0517bd413b3844573b15769021f86a476.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 313, + 545, + 346 + ], + "lines": [ + { + "bbox": [ + 305, + 313, + 545, + 346 + ], + "spans": [ + { + "bbox": [ + 305, + 313, + 545, + 346 + ], + "type": "text", + "content": "Figure 4. Analysis of average accuracy " + }, + { + "bbox": [ + 305, + 313, + 545, + 346 + ], + "type": "inline_equation", + "content": "A_N" + }, + { + "bbox": [ + 305, + 313, + 545, + 346 + ], + "type": "text", + "content": " vs prompt length for CODA-P, L2P, and DualPrompt. The setting is 5 task ImageNet-R, and we report the mean over 3 trials." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 359, + 545, + 467 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 359, + 545, + 467 + ], + "spans": [ + { + "bbox": [ + 304, + 359, + 545, + 467 + ], + "type": "text", + "content": "performance, finding that our method is closer to upper bound performance than it is to DualPrompt in average accuracy. Because the prompt pool in L2P can be increased to arbitrary size as well, we include results in this analysis as well. However, we show that the L2P method peaks with a pool size equal to twice the task sequence length, and then drops in performance. This reflects that our prompt components benefit from scale, whereas the existing prompt pool actually suffer." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 467, + 545, + 538 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 467, + 545, + 538 + ], + "spans": [ + { + "bbox": [ + 304, + 467, + 545, + 538 + ], + "type": "text", + "content": "Finally, we show average accuracy versus prompt length in Figure 4 using the same setting as above. The purpose of this experiment is to emphasize that accuracy saturates with prompt length, thus motivating the need to expand in our introduced component dimension. Additional analysis and hyperparameter sweeps is available in our SM." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 550, + 378, + 562 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 550, + 378, + 562 + ], + "spans": [ + { + "bbox": [ + 306, + 550, + 378, + 562 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "text", + "content": "We present CContinual decomposed attention-based prompting (CODA-Prompt) for rehearsal-free continual learning. Our approach assembles learnable prompt components which are then inserted into a pre-trained ViT encoder for image classification. Importantly, CODA-Prompt is optimized in an end-to-end fashion (unlike prior SOTA methods which involve two, separate optimizations). Furthermore, CODA-Prompt can scale prompting capacity to arbitrary sizes. We set a new SOTA on both well-established benchmarks and a dual-distribution shift benchmark (containing semantic and covariate shifts), highlighting the potential real world impact and generality of our approach." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "11916" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "type": "text", + "content": "[1] Hongjoon Ahn, Jihwan Kwak, Subin Lim, Hyeonsu Bang, Hyojun Kim, and Taesup Moon. Ss-il: Separated softmax for incremental learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 844-853, October 2021. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 146, + 288, + 190 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 146, + 288, + 190 + ], + "spans": [ + { + "bbox": [ + 53, + 146, + 288, + 190 + ], + "type": "text", + "content": "[2] Rahaf Aljundi, Francesca Babiloni, Mohamed Elhoseiny, Marcus Rohrbach, and Tinne Tuytelaars. Memory aware synapses: Learning what (not) to forget. In ECCV, 2018. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 191, + 288, + 246 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 191, + 288, + 246 + ], + "spans": [ + { + "bbox": [ + 53, + 191, + 288, + 246 + ], + "type": "text", + "content": "[3] Rahaf Aljundi, Eugene Belilovsky, Tinne Tuytelaars, Laurent Charlin, Massimo Caccia, Min Lin, and Lucas Page-Caccia. Online continual learning with maximal interfered retrieval. In Advances in Neural Information Processing Systems, pages 11849-11860, 2019. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 247, + 288, + 291 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 247, + 288, + 291 + ], + "spans": [ + { + "bbox": [ + 53, + 247, + 288, + 291 + ], + "type": "text", + "content": "[4] Rahaf Aljundi, Min Lin, Baptiste Goujaud, and Yoshua Bengio. Gradient based sample selection for online continual learning. In Advances in Neural Information Processing Systems, pages 11816-11825, 2019. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 291, + 288, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 291, + 288, + 346 + ], + "spans": [ + { + "bbox": [ + 53, + 291, + 288, + 346 + ], + "type": "text", + "content": "[5] Jihwan Bang, Heesu Kim, YoungJoon Yoo, Jung-Woo Ha, and Jonghyun Choi. Rainbow memory: Continual learning with a memory of diverse samples. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8218-8227, 2021. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 346, + 288, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 346, + 288, + 402 + ], + "spans": [ + { + "bbox": [ + 53, + 346, + 288, + 402 + ], + "type": "text", + "content": "[6] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. Advances in neural information processing systems, 33:1877-1901, 2020. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 403, + 288, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 403, + 288, + 456 + ], + "spans": [ + { + "bbox": [ + 53, + 403, + 288, + 456 + ], + "type": "text", + "content": "[7] Pietro Buzzega, Matteo Boschini, Angelo Porrello, Davide Abati, and Simone Calderara. Dark experience for general continual learning: a strong, simple baseline. Advances in neural information processing systems, 33:15920-15930, 2020. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 458, + 288, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 458, + 288, + 502 + ], + "spans": [ + { + "bbox": [ + 53, + 458, + 288, + 502 + ], + "type": "text", + "content": "[8] Francisco M Castro, Manuel J Marín-Jiménez, Nicolas Guil, Cordelia Schmid, and Karteek Alahari. End-to-end incremental learning. In Proceedings of the European Conference on Computer Vision (ECCV), pages 233-248, 2018. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 502, + 288, + 546 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 502, + 288, + 546 + ], + "spans": [ + { + "bbox": [ + 53, + 502, + 288, + 546 + ], + "type": "text", + "content": "[9] Arslan Chaudhry, Marc' Aurelio Ranzato, Marcus Rohrbach, and Mohamed Elhoseiny. Efficient lifelong learning with a GEM. In International Conference on Learning Representations, 2019. 2, 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 547, + 287, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 547, + 287, + 601 + ], + "spans": [ + { + "bbox": [ + 48, + 547, + 287, + 601 + ], + "type": "text", + "content": "[10] Arslan Chaudhry, Marcus Rohrbach, Mohamed Elhoseiny, Thalaiyasingam Ajanthan, Puneet K Dokania, Philip HS Torr, and Marc'Aurelio Ranzato. Continual learning with tiny episodic memories. arXiv preprint arXiv:1902.10486, 2019. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 602, + 287, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 602, + 287, + 657 + ], + "spans": [ + { + "bbox": [ + 48, + 602, + 287, + 657 + ], + "type": "text", + "content": "[11] Arslan Chaudhry, Marcus Rohrbach, Mohamed Elhoseiny, Thalaiyasingam Ajanthan, Puneet K Dokania, Philip HS Torr, and Marc'Aurelio Ranzato. On tiny episodic memories in continual learning. arXiv preprint arXiv:1902.10486, 2019. 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 658, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 287, + 712 + ], + "type": "text", + "content": "[12] Yoojin Choi, Mostafa El-Khamy, and Jungwon Lee. Dualteacher class-incremental learning with data-free generative replay. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3543-3552, 2021. 2" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 714 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 307, + 73, + 547, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 547, + 138 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 547, + 138 + ], + "type": "text", + "content": "[13] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 3, 5, 6" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 140, + 547, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 140, + 547, + 194 + ], + "spans": [ + { + "bbox": [ + 307, + 140, + 547, + 194 + ], + "type": "text", + "content": "[14] Arthur Douillard, Alexandre Ramé, Guillaume Couairon, and Matthieu Cord. Dytox: Transformers for continual learning with dynamic token expansion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9285-9295, 2022. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 194, + 547, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 194, + 547, + 237 + ], + "spans": [ + { + "bbox": [ + 307, + 194, + 547, + 237 + ], + "type": "text", + "content": "[15] Sayna Ebrahimi, Mohamed Elhoseiny, Trevor Darrell, and Marcus Rohrbach. Uncertainty-guided continual learning with bayesian neural networks. arXiv preprint arXiv:1906.02425, 2019. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 238, + 547, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 238, + 547, + 271 + ], + "spans": [ + { + "bbox": [ + 307, + 238, + 547, + 271 + ], + "type": "text", + "content": "[16] Sayna Ebrahimi, Franziska Meier, Roberto Calandra, Trevor Darrell, and Marcus Rohrbach. Adversarial continual learning. arXiv preprint arXiv:2003.09553, 2020. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 272, + 547, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 272, + 547, + 315 + ], + "spans": [ + { + "bbox": [ + 307, + 272, + 547, + 315 + ], + "type": "text", + "content": "[17] Qiankun Gao, Chen Zhao, Bernard Ghanem, and Jian Zhang. R-dfcil: Relation-guided representation learning for data-free class incremental learning. arXiv preprint arXiv:2203.13104, 2022. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 316, + 547, + 371 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 316, + 547, + 371 + ], + "spans": [ + { + "bbox": [ + 307, + 316, + 547, + 371 + ], + "type": "text", + "content": "[18] Alexander Gepperth and Cem Karaoguz. Incremental learning with self-organizing maps. 2017 12th International Workshop on Self-Organizing Maps and Learning Vector Quantization, Clustering and Data Visualization (WSOM), pages 1–8, 2017. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 371, + 546, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 371, + 546, + 415 + ], + "spans": [ + { + "bbox": [ + 307, + 371, + 546, + 415 + ], + "type": "text", + "content": "[19] Tyler L Hayes, Nathan D Cahill, and Christopher Kanan. Memory efficient experience replay for streaming learning. In 2019 International Conference on Robotics and Automation (ICRA), pages 9769-9776. IEEE, 2019. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 415, + 546, + 449 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 415, + 546, + 449 + ], + "spans": [ + { + "bbox": [ + 307, + 415, + 546, + 449 + ], + "type": "text", + "content": "[20] Tyler L Hayes and Christopher Kanan. Lifelong machine learning with deep streaming linear discriminant analysis. arXiv preprint arXiv:1909.01520, 2019. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 449, + 546, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 449, + 546, + 514 + ], + "spans": [ + { + "bbox": [ + 307, + 449, + 546, + 514 + ], + "type": "text", + "content": "[21] Dan Hendrycks, Steven Basart, Norman Mu, Saurav Kadavath, Frank Wang, Evan Dorundo, Rahul Desai, Tyler Zhu, Samyak Parajuli, Mike Guo, Dawn Song, Jacob Steinhardt, and Justin Gilmer. The many faces of robustness: A critical analysis of out-of-distribution generalization. ICCV, 2021. 2, 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 514, + 547, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 514, + 547, + 547 + ], + "spans": [ + { + "bbox": [ + 307, + 514, + 547, + 547 + ], + "type": "text", + "content": "[22] Geoffrey Hinton, Oriol Vinyals, and Jeff Dean. Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531, 2015. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 548, + 547, + 592 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 548, + 547, + 592 + ], + "spans": [ + { + "bbox": [ + 307, + 548, + 547, + 592 + ], + "type": "text", + "content": "[23] Saihui Hou, Xinyu Pan, Chen Change Loy, Zilei Wang, and Dahua Lin. Lifelong learning via progressive distillation and retrospection. In Proceedings of the European Conference on Computer Vision (ECCV), pages 437-452, 2018. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 592, + 546, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 592, + 546, + 646 + ], + "spans": [ + { + "bbox": [ + 307, + 592, + 546, + 646 + ], + "type": "text", + "content": "[24] Saihui Hou, Xinyu Pan, Chen Change Loy, Zilei Wang, and Dahua Lin. Learning a unified classifier incrementally via rebalancing. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 831-839, 2019. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 647, + 546, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 647, + 546, + 690 + ], + "spans": [ + { + "bbox": [ + 307, + 647, + 546, + 690 + ], + "type": "text", + "content": "[25] Yen-Chang Hsu, Yen-Cheng Liu, Anita Ramasamy, and Zsolt Kira. Re-evaluating continual learning scenarios: A categorization and case for strong baselines. arXiv preprint arXiv:1810.12488, 2018. 2, 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 691, + 546, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 691, + 546, + 714 + ], + "spans": [ + { + "bbox": [ + 307, + 691, + 546, + 714 + ], + "type": "text", + "content": "[26] Daniel Justus, John Brennan, Stephen Bonner, and Andrew Stephen McGough. Predicting the computational cost" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "11917" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 66, + 72, + 287, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 72, + 287, + 106 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 287, + 106 + ], + "type": "text", + "content": "of deep learning models. In 2018 IEEE international conference on big data (Big Data), pages 3873-3882. IEEE, 2018. 1" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 106, + 287, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 106, + 287, + 139 + ], + "spans": [ + { + "bbox": [ + 48, + 106, + 287, + 139 + ], + "type": "text", + "content": "[27] Nitin Kamra, Umang Gupta, and Yan Liu. Deep generative dual memory network for continual learning. arXiv preprint arXiv:1710.10368, 2017. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 140, + 287, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 140, + 287, + 172 + ], + "spans": [ + { + "bbox": [ + 48, + 140, + 287, + 172 + ], + "type": "text", + "content": "[28] Ronald Kemker and Christopher Kanan. Fearnet: Brain-inspired model for incremental learning. International Conference on Learning Representations (ICLR), 2018. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 173, + 287, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 173, + 287, + 216 + ], + "spans": [ + { + "bbox": [ + 48, + 173, + 287, + 216 + ], + "type": "text", + "content": "[29] Ronald Kemker, Marc McClure, Angelina Abitino, Tyler Hayes, and Christopher Kanan. Measuring catastrophic forgetting in neural networks. AAAI Conference on Artificial Intelligence, 2018. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 217, + 287, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 217, + 287, + 281 + ], + "spans": [ + { + "bbox": [ + 48, + 217, + 287, + 281 + ], + "type": "text", + "content": "[30] James Kirkpatrick, Razvan Pascanu, Neil Rabinowitz, Joel Veness, Guillaume Desjardins, Andrei A Rusu, Kieran Milan, John Quan, Tiago Ramalho, Agnieszka Grabska-Barwinska, et al. Overcoming catastrophic forgetting in neural networks. Proceedings of the national academy of sciences, 2017. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 283, + 287, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 283, + 287, + 316 + ], + "spans": [ + { + "bbox": [ + 48, + 283, + 287, + 316 + ], + "type": "text", + "content": "[31] Mario Köppen. The curse of dimensionality. In 5th online world conference on soft computing in industrial applications (WSC5), volume 1, pages 4-8, 2000. 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 316, + 287, + 338 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 316, + 287, + 338 + ], + "spans": [ + { + "bbox": [ + 48, + 316, + 287, + 338 + ], + "type": "text", + "content": "[32] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. Tech Report, 2009. 2, 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 338, + 287, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 338, + 287, + 380 + ], + "spans": [ + { + "bbox": [ + 48, + 338, + 287, + 380 + ], + "type": "text", + "content": "[33] Alexandre Lacoste, Alexandra Luccioni, Victor Schmidt, and Thomas Dandres. Quantifying the carbon emissions of machine learning. arXiv preprint arXiv:1910.09700, 2019. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 382, + 287, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 382, + 287, + 426 + ], + "spans": [ + { + "bbox": [ + 48, + 382, + 287, + 426 + ], + "type": "text", + "content": "[34] Kibok Lee, Kimin Lee, Jinwoo Shin, and Honglak Lee. Overcoming catastrophic forgetting with unlabeled data in the wild. In Proceedings of the IEEE International Conference on Computer Vision, pages 312-321, 2019. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 426, + 287, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 426, + 287, + 459 + ], + "spans": [ + { + "bbox": [ + 48, + 426, + 287, + 459 + ], + "type": "text", + "content": "[35] Soochan Lee, Junsoo Ha, Dongsu Zhang, and Gunhee Kim. A neural dirichlet process mixture model for task-free continual learning. arXiv preprint arXiv:2001.00689, 2020. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 460, + 287, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 460, + 287, + 514 + ], + "spans": [ + { + "bbox": [ + 48, + 460, + 287, + 514 + ], + "type": "text", + "content": "[36] Timothee Lesort, Hugo Caselles-Dupré, Michael Garcia-Ortiz, Andrei Stoian, and David Filliat. Generative models from the perspective of continual learning. In 2019 International Joint Conference on Neural Networks (IJCNN), pages 1–8. IEEE, 2019. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 514, + 287, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 514, + 287, + 557 + ], + "spans": [ + { + "bbox": [ + 48, + 514, + 287, + 557 + ], + "type": "text", + "content": "[37] Duo Li, Guimei Cao, Yunlu Xu, Zhanzhan Cheng, and Yi Niu. Technical report for iccv 2021 challenge sslad-track3b: Transformers are better continual learners. arXiv preprint arXiv:2201.04924, 2022. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 559, + 287, + 591 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 559, + 287, + 591 + ], + "spans": [ + { + "bbox": [ + 48, + 559, + 287, + 591 + ], + "type": "text", + "content": "[38] Zhizhong Li and Derek Hoiem. Learning without forgetting. IEEE transactions on pattern analysis and machine intelligence, 40(12):2935-2947, 2017. 2, 6, 7" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 592, + 287, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 592, + 287, + 624 + ], + "spans": [ + { + "bbox": [ + 48, + 592, + 287, + 624 + ], + "type": "text", + "content": "[39] Vincenzo Lomonaco and Davide Maltoni. Core50: a new dataset and benchmark for continuous object recognition. arXiv preprint arXiv:1705.03550, 2017. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 625, + 287, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 625, + 287, + 658 + ], + "spans": [ + { + "bbox": [ + 48, + 625, + 287, + 658 + ], + "type": "text", + "content": "[40] Vincenzo Lomonaco, Davide Maltoni, and Lorenzo Pellegrini. Rehearsal-free continual learning over small non-iid batches. In CVPR Workshops, pages 989-998, 2020. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 658, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 287, + 712 + ], + "type": "text", + "content": "[41] David Lopez-Paz and Marc'Aurelio Ranzato. Gradient episodic memory for continual learning. In Proceedings of the 31st International Conference on Neural Information Processing Systems, NIPS'17, pages 6470-6479, USA, 2017. Curran Associates Inc. 2, 6" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 714 + ], + "type": "list", + "angle": 0, + "index": 31, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "text", + "content": "[42] Zheda Mai, Ruiwen Li, Jihwan Jeong, David Quispe, Hyunwoo Kim, and Scott Sanner. Online continual learning in image classification: An empirical survey. Neurocomputing, 469:28-51, 2022. 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 118, + 545, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 545, + 150 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 545, + 150 + ], + "type": "text", + "content": "[43] Davide Maltoni and Vincenzo Lomonaco. Continuous learning in single-incremental-task scenarios. Neural Networks, 116:56-73, 2019. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 152, + 545, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 152, + 545, + 196 + ], + "spans": [ + { + "bbox": [ + 307, + 152, + 545, + 196 + ], + "type": "text", + "content": "[44] Seyed Iman Mirzadeh, Arslan Chaudhry, Dong Yin, Timothy Nguyen, Razvan Pascanu, Dilan Gorur, and Mehrdad Farajtabar. Architecture matters in continual learning. arXiv preprint arXiv:2202.00275, 2022. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 197, + 545, + 230 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 197, + 545, + 230 + ], + "spans": [ + { + "bbox": [ + 307, + 197, + 545, + 230 + ], + "type": "text", + "content": "[45] Vaishnavh Nagarajan, Colin Raffel, and Ian J Goodfellow. Theoretical insights into memorization in gans. In Neural Information Processing Systems Workshop, 2018. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 232, + 545, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 232, + 545, + 274 + ], + "spans": [ + { + "bbox": [ + 307, + 232, + 545, + 274 + ], + "type": "text", + "content": "[46] Cuong V Nguyen, Alessandro Achille, Michael Lam, Tal Hassner, Vijay Mahadevan, and Stefano Soatto. Toward understanding catastrophic forgetting in continual learning. arXiv preprint arXiv:1908.01091, 2019. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 276, + 545, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 276, + 545, + 331 + ], + "spans": [ + { + "bbox": [ + 307, + 276, + 545, + 331 + ], + "type": "text", + "content": "[47] Oleksiy Ostapenko, Mihai Puscas, Tassilo Klein, Patrick Jahnichen, and Moin Nabi. Learning to remember: A synaptic plasticity driven framework for continual learning. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 11321-11329, 2019. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 332, + 545, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 332, + 545, + 397 + ], + "spans": [ + { + "bbox": [ + 307, + 332, + 545, + 397 + ], + "type": "text", + "content": "[48] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. Advances in neural information processing systems, 32, 2019. 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 399, + 545, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 399, + 545, + 453 + ], + "spans": [ + { + "bbox": [ + 307, + 399, + 545, + 453 + ], + "type": "text", + "content": "[49] Xingchao Peng, Qinxun Bai, Xide Xia, Zijun Huang, Kate Saenko, and Bo Wang. Moment matching for multi-source domain adaptation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 1406-1415, 2019. 7" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 455, + 545, + 509 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 455, + 545, + 509 + ], + "spans": [ + { + "bbox": [ + 307, + 455, + 545, + 509 + ], + "type": "text", + "content": "[50] Sylvestre-Alvise Rebuffi, Alexander Kolesnikov, Georg Sperl, and Christoph H. Lampert. icarl: Incremental classifier and representation learning. In 2017 IEEE Conference on Computer Vision and Pattern Recognition, CVPR'17, pages 5533-5542, 2017. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 510, + 545, + 543 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 510, + 545, + 543 + ], + "spans": [ + { + "bbox": [ + 307, + 510, + 545, + 543 + ], + "type": "text", + "content": "[51] Anthony Robins. Catastrophic forgetting, rehearsal and pseudorehearsal. Connection Science, 7(2):123-146, 1995. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 544, + 545, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 544, + 545, + 589 + ], + "spans": [ + { + "bbox": [ + 307, + 544, + 545, + 589 + ], + "type": "text", + "content": "[52] David Rolnick, Arun Ahuja, Jonathan Schwarz, Timothy Lillicrap, and Gregory Wayne. Experience replay for continual learning. In Advances in Neural Information Processing Systems, pages 348-358, 2019. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 590, + 545, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 590, + 545, + 645 + ], + "spans": [ + { + "bbox": [ + 307, + 590, + 545, + 645 + ], + "type": "text", + "content": "[53] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, et al. Imagenet large scale visual recognition challenge. International journal of computer vision, 115(3):211-252, 2015. 6" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 647, + 545, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 647, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 307, + 647, + 545, + 689 + ], + "type": "text", + "content": "[54] Andrei A Rusu, Neil C Rabinowitz, Guillaume Desjardins, Hubert Soyer, James Kirkpatrick, Koray Kavukcuoglu, Razvan Pascanu, and Raia Hadsell. Progressive neural networks. arXiv preprint arXiv:1606.04671, 2016. 2" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 307, + 691, + 545, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 691, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 307, + 691, + 545, + 714 + ], + "type": "text", + "content": "[55] Hanul Shin, Jung Kwon Lee, Jaehong Kim, and Jiwon Kim. Continual learning with deep generative replay. In I. Guyon," + } + ] + } + ], + "index": 30 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "text", + "content": "11918" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 66, + 72, + 287, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 72, + 287, + 116 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 287, + 116 + ], + "type": "text", + "content": "U. V. Luxburg, S. Bengio, H. Wallach, R. Fergus, S. Vishwanathan, and R. Garnett, editors, Advances in Neural Information Processing Systems 30, pages 2990-2999. Curran Associates, Inc., 2017. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 118, + 287, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 118, + 287, + 183 + ], + "spans": [ + { + "bbox": [ + 48, + 118, + 287, + 183 + ], + "type": "text", + "content": "[56] James Smith, Yen-Chang Hsu, Jonathan Balloch, Yilin Shen, Hongxia Jin, and Zsolt Kira. Always be dreaming: A new approach for data-free class-incremental learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 9374-9384, October 2021. 2, 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 185, + 287, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 185, + 287, + 217 + ], + "spans": [ + { + "bbox": [ + 48, + 185, + 287, + 217 + ], + "type": "text", + "content": "[57] James Seale Smith, Junjiao Tian, Yen-Chang Hsu, and Zsolt Kira. A closer look at rehearsal-free continual learning. arXiv preprint arXiv:2203.17269, 2022. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 220, + 287, + 273 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 220, + 287, + 273 + ], + "spans": [ + { + "bbox": [ + 48, + 220, + 287, + 273 + ], + "type": "text", + "content": "[58] Michalis K Titsias, Jonathan Schwarz, Alexander G de G Matthews, Razvan Pascanu, and Yee Whye Teh. Functional regularisation for continual learning with gaussian processes. In International Conference on Learning Representations, 2019. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 275, + 287, + 317 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 275, + 287, + 317 + ], + "spans": [ + { + "bbox": [ + 48, + 275, + 287, + 317 + ], + "type": "text", + "content": "[59] Gido M van de Ven, Hava T Siegelmann, and Andreas S Tolias. Brain-inspired replay for continual learning with artificial neural networks. Nature communications, 11(1):1-14, 2020. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 320, + 287, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 320, + 287, + 353 + ], + "spans": [ + { + "bbox": [ + 48, + 320, + 287, + 353 + ], + "type": "text", + "content": "[60] Gido M van de Ven and Andreas S Tolias. Generative replay with feedback connections as a general strategy for continual learning. arXiv preprint arXiv:1809.10635, 2018. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 354, + 287, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 354, + 287, + 386 + ], + "spans": [ + { + "bbox": [ + 48, + 354, + 287, + 386 + ], + "type": "text", + "content": "[61] Gido M Van de Ven and Andreas S Tolias. Three scenarios for continual learning. arXiv preprint arXiv:1904.07734, 2019. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 388, + 287, + 432 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 388, + 287, + 432 + ], + "spans": [ + { + "bbox": [ + 48, + 388, + 287, + 432 + ], + "type": "text", + "content": "[62] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 434, + 287, + 466 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 434, + 287, + 466 + ], + "spans": [ + { + "bbox": [ + 48, + 434, + 287, + 466 + ], + "type": "text", + "content": "[63] Johannes von Oswald, Christian Henning, João Sacramento, and Benjamin F Grewe. Continual learning with hypernetworks. arXiv preprint arXiv:1906.00695, 2019. 2, 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 468, + 287, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 468, + 287, + 511 + ], + "spans": [ + { + "bbox": [ + 48, + 468, + 287, + 511 + ], + "type": "text", + "content": "[64] Yabin Wang, Zhiwu Huang, and Xiaopeng Hong. S-prompts learning with pre-trained transformers: An occam's razor for domain incremental learning. In Conference on Neural Information Processing Systems (NeurIPS), 2022. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 513, + 287, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 513, + 287, + 567 + ], + "spans": [ + { + "bbox": [ + 48, + 513, + 287, + 567 + ], + "type": "text", + "content": "[65] Zifeng Wang, Zizhao Zhang, Sayna Ebrahimi, Ruoxi Sun, Han Zhang, Chen-Yu Lee, Xiaoqi Ren, Guolong Su, Vincent Perot, Jennifer Dy, et al. Dualprompt: Complementary prompting for rehearsal-free continual learning. arXiv preprint arXiv:2204.04799, 2022. 1, 2, 3, 4, 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 568, + 287, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 568, + 287, + 633 + ], + "spans": [ + { + "bbox": [ + 48, + 568, + 287, + 633 + ], + "type": "text", + "content": "[66] Zifeng Wang, Zizhao Zhang, Chen-Yu Lee, Han Zhang, Ruoxi Sun, Xiaqi Ren, Guolong Su, Vincent Perot, Jennifer Dy, and Tomas Pfister. Learning to prompt for continual learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 139-149, 2022. 2, 3, 4, 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 635, + 287, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 635, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 635, + 287, + 689 + ], + "type": "text", + "content": "[67] Yue Wu, Yinpeng Chen, Lijuan Wang, Yuancheng Ye, Zicheng Liu, Yandong Guo, and Yun Fu. Large scale incremental learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 374-382, 2019. 2, 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "type": "text", + "content": "[68] Hongxu Yin, Pavlo Molchanov, Jose M Alvarez, Zhizhong Li, Arun Mallya, Derek Hoiem, Niraj K Jha, and Jan Kautz." + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 218 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 116 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 116 + ], + "type": "text", + "content": "Dreaming to distill: Data-free knowledge transfer via deep-inversion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8715-8724, 2020. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 118, + 545, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 545, + 150 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 545, + 150 + ], + "type": "text", + "content": "[69] Wenpeng Yin, Jamaal Hay, and Dan Roth. Benchmarking zero-shot text classification: Datasets, evaluation and entailment approach. arXiv preprint arXiv:1909.00161, 2019. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 152, + 545, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 152, + 545, + 184 + ], + "spans": [ + { + "bbox": [ + 307, + 152, + 545, + 184 + ], + "type": "text", + "content": "[70] Pei Yu, Yinpeng Chen, Ying Jin, and Zicheng Liu. Improving vision transformers for incremental learning. arXiv preprint arXiv:2112.06103, 2021. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 186, + 545, + 218 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 186, + 545, + 218 + ], + "spans": [ + { + "bbox": [ + 307, + 186, + 545, + 218 + ], + "type": "text", + "content": "[71] Friedemann Zenke, Ben Poole, and Surya Ganguli. Continual learning through synaptic intelligence. In International Conference on Machine Learning, 2017. 2" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "11919" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/CORA_ Adapting CLIP for Open-Vocabulary Detection With Region Prompting and Anchor Pre-Matching/e257818b-1aea-4bd1-9288-7c6a81195cd8_content_list.json b/2023/CORA_ Adapting CLIP for Open-Vocabulary Detection With Region Prompting and Anchor Pre-Matching/e257818b-1aea-4bd1-9288-7c6a81195cd8_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..4eb213d0cb04dc32ef16cf95e53e8d726cd703b2 --- /dev/null +++ b/2023/CORA_ Adapting CLIP for Open-Vocabulary Detection With Region Prompting and Anchor Pre-Matching/e257818b-1aea-4bd1-9288-7c6a81195cd8_content_list.json @@ -0,0 +1,1501 @@ +[ + { + "type": "text", + "text": "CORA: Adapting CLIP for Open-Vocabulary Detection with Region Prompting and Anchor Pre-Matching", + "text_level": 1, + "bbox": [ + 179, + 128, + 789, + 172 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Xiaoshi Wu $^{1}$ , Feng Zhu $^{2}$ , Rui Zhao $^{2,3}$ , Hongsheng Li $^{1,4}$", + "bbox": [ + 259, + 199, + 702, + 218 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1Multimedia Laboratory, The Chinese University of Hong Kong", + "bbox": [ + 233, + 219, + 738, + 237 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ SenseTime Research $^{3}$ Qing Yuan Research Institute, Shanghai Jiao Tong University", + "bbox": [ + 143, + 237, + 826, + 253 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{4}$ Centre for Perceptual and Interactive Intelligence (CPII)", + "bbox": [ + 259, + 253, + 709, + 270 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{wuxiaoshi@link, hsli@ee}.cuhk.edu.hk,{zhufeng, zhaorui}@sentetime.com", + "bbox": [ + 174, + 272, + 790, + 287 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 321, + 312, + 338 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Open-vocabulary detection (OVD) is an object detection task aiming at detecting objects from novel categories beyond the base categories on which the detector is trained. Recent OVD methods rely on large-scale visual-language pre-trained models, such as CLIP, for recognizing novel objects. We identify the two core obstacles that need to be tackled when incorporating these models into detector training: (1) the distribution mismatch that happens when applying a VL-model trained on whole images to region recognition tasks; (2) the difficulty of localizing objects of unseen classes. To overcome these obstacles, we propose CORA, a DETR-style framework that adapts CLIP for Open-vocabulary detection by Region prompting and Anchor pre-matching. Region prompting mitigates the whole-to-region distribution gap by prompting the region features of the CLIP-based region classifier. Anchor pre-matching helps learning generalizable object localization by a class-aware matching mechanism. We evaluate CORA on the COCO OVD benchmark, where we achieve 41.7 AP50 on novel classes, which outperforms the previous SOTA by 2.4 AP50 even without resorting to extra training data. When extra training data is available, we train $\\mathrm{CORA^{+}}$ on both ground-truth base-category annotations and additional pseudo bounding box labels computed by CORA. $\\mathrm{CORA^{+}}$ achieves 43.1 AP50 on the COCO OVD benchmark and 28.1 box APr on the LVIS OVD benchmark. The code is available at https://github.com/tgxs002/CORA.", + "bbox": [ + 76, + 354, + 472, + 741 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 787, + 207, + 803 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Object detection is a fundamental vision problem that involves localizing and classifying objects from images. Classical object detection requires detecting objects from a closed set of categories. Extra annotations and training are required if objects of unseen categories need to be detected. It has attracted much attention on detecting novel categories", + "bbox": [ + 75, + 813, + 468, + 901 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "without tedious annotations, or even detect object from new category, which is currently referred as open-vocabulary detection (OVD) [36].", + "bbox": [ + 498, + 323, + 890, + 366 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recent advances on large-scale vision-language pretrained models, such as CLIP [30], enable new solutions for tackling OVD. CLIP learns a joint embedding space of images and text from a large-scale image-text dataset, which shows remarkable capability on visual recognition tasks. The general idea of applying CLIP for OVD is to treat it as an open-vocabulary classifier. However, there are two obstacles hindering the effective use of CLIP on tackling OVD.", + "bbox": [ + 496, + 367, + 892, + 494 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "How to adapt CLIP for region-level tasks? One trivial solution is to crop regions and treat them as separate images, which has been adopted by multiple recent works [7, 14, 31, 35]. But the distribution gap between region crops and full images leads to inferior classification accuracy. MEDet [7] mitigates this issue by augmenting the text feature with image features. However, it requires extra image-text pairs to prevent overfitting to so-called \"base\" classes that are seen during training. RegionCLIP [40] directly acquires regional features by RoIAlign [17], which is more efficient but cannot generalize well to novel classes without finetuning. The finetuning is costly when adopting a larger CLIP model.", + "bbox": [ + 496, + 496, + 893, + 681 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "How to learn generalizable object proposals? ViLD [14], OV-DETR [35], Object-Centric-OVD [31], Region-CLIP [40] need RPN or class-agnostic object detectors [29] to mine potential novel class objects. However, these RPNs are strongly biased towards the base classes on which they are trained, while perform poorly on the novel classes. MEDet [7] and VL-PLM [39] identify this problem and adopt several handcrafted policies to rule out or merge low-quality boxes, but the performance is still bounded by the frozen RPN. OV-DETR [35] learns generalizable object localization by conditioning box regression on class name embeddings, but at the cost of efficiency issue induced by repetitive per-class inference.", + "bbox": [ + 496, + 684, + 893, + 869 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this work, we propose a new framework based on DETection TR transformers (DETR) [6] that incorporates CLIP", + "bbox": [ + 500, + 869, + 890, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "7031", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "into detector training to achieve open-vocabulary detection without additional image-text data. Specifically, we use a DETR-style object localizer for class-aware object localization, and the predicted boxes are encoded by pooling the intermediate feature map of the CLIP image encoder, which are classified by the CLIP text encoder with class names. However, there is a distribution gap between whole-image features from CLIP's original visual encoder and the newly pooled region features, leading to an inferior classification accuracy. Thus, we propose Region Prompting to adapt the CLIP image encoder, which boosts the classification performance, and also demonstrates better generalization capability than existing methods. We adopt DAB-DETR [26] as the localizer, in which object queries are associated with dynamic anchor boxes. By pre-matching the dynamic anchor boxes with the input categories before box regression (Anchor Pre-Matching), class-aware regression can be achieved without the cost of repetitive per-class inference.", + "bbox": [ + 80, + 90, + 467, + 349 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We validate our method on COCO [24] and LVIS v1.0 [16] OVD benchmarks. On the COCO OVD benchmark, our method improves AP50 of novel categories over the previous best method [40] by 2.4 AP50 without training on extra data, and achieves consistent gain on CLIP models of different scales. When compared under a fairer setting with extra training data, our method significantly outperforms the existing methods by 3.8 AP50 on novel categories and achieves comparable performance on base categories. On the LVIS OVD benchmark, our method achieves 22.2/28.1 APr with/w.o. extra data, which significantly outperforms existing methods that are also trained with/w.o. extra data. By applying region prompting on the base classes of COCO, the classification performance on the novel classes is boosted from $63.9\\%$ to $74.1\\%$ , whereas other prompting or adaptation methods easily bias towards the base classes.", + "bbox": [ + 80, + 351, + 467, + 594 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The contributions of this work are summarized as follows: (1) Our proposed region prompting effectively mitigates the gap between whole image features and region features, and generalize well in the open-vocabulary setting. (2) Anchor Pre-Matching enables DETR for generalizable object localization efficiently. (3) We achieve state-of-the-art performance on COCO and LVIS OVD benchmarks.", + "bbox": [ + 80, + 595, + 467, + 696 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 80, + 715, + 223, + 729 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Open-Vocabulary Object Detection OVR-CNN [36] firstly put forth this new formulation of detection, and proposes its baseline solution by aligning region features with nouns in captions that are paired with the image. Mingfeng [11] et al. mine pseudo labels by utilizing the localization ability of pre-trained vision-language models. PromptDet [10] addresses the gap between image and region classification by adding learnable prompts when encoding the class names, namely regional prompt learning (RPL), which is expected to generalize from base to novel categories. OV-DETR [35] is the first DETR-style open", + "bbox": [ + 80, + 741, + 467, + 898 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "vocabulary detector, which proposes conditional matching to solve the missing novel class problem in assignment, but at the cost of inefficient inference. RegionCLIP [40] propose a second-stage pre-training mechanism to adapt the CLIP model to encode region features, and demonstrates its capability on OVD and zero-shot transfer setting. GLIP [21] jointly learns object localization and VL alignment. Matthias et al. [15] proposes to finetune a VL aligned model for detection, while we fix the pre-trained VL model for better generalization towards novel categories.", + "bbox": [ + 503, + 90, + 888, + 234 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Detection Transformers DETR [6] is an object detection architecture based on transformers that formulates object detection as a set-to-set matching problem, which greatly simplifies the pipeline. Several works address the slow convergence problem of DETR by architectural improvement [1, 13,26,38] or special training strategies [8,18]. Zhu et al. [1] proposes multi-scale deformable attention module to efficiently aggregate information from multi-scale feature maps. Gao et al. [13] proposes to modulate the cross-attention in the transformer decoder by anchor box coordinates to accelerate the detector convergence. DABDETR [26] formulates the queries in DETR architecture as anchor boxes, which accelerates detector training. Chen et al. [38] proposes Group DETR, which adds auxiliary object queries during training to take advantage of one-to-many matching for faster convergence.", + "bbox": [ + 503, + 239, + 888, + 470 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Prompt Tuning Prompting is originated from NLP, and it refers to prepending task instructions before the input sequence to give the language model the hint about the task [5]. Later works [22, 27] explores tuning continuous prompt vectors when few-shot data is available. VPT [19], Visual Prompting [2, 3] explore prompting in the pixel space. [20] and [25] prompts pre-trained model for video recognition tasks. [34] proposes class-aware visual prompt tuning to generalize the learned prompts to unseen categories. Recent works demonstrate that prompt tuning is an effective and parameter-efficient way to adapt large-scale pre-trained models to downstream tasks.", + "bbox": [ + 503, + 477, + 888, + 648 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 503, + 681, + 586, + 696 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Open-vocabulary detection (OVD) is an object detection task aiming at detecting objects from novel categories beyond the base categories on which the detector is trained. Formally, the detector is trained on a detection dataset with base-category $C^B$ box annotations, and tested on a new input image $I \\in \\mathbb{R}^{H \\times W \\times 3}$ to detect objects belonging to a novel category set $C^N$ , where $C^B \\cap C^N = \\emptyset$ . In this section, we introduce CORA, a framework that adapts CLIP for the OVD task by Region prompting and Anchor prematching. For fairer comparisons with existing methods, we also experiment with a broader setting where extra data is available, which is referred to as $\\mathrm{CORA}^+$ , and will be introduced along with the experiments.", + "bbox": [ + 503, + 712, + 888, + 898 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "7032", + "bbox": [ + 483, + 946, + 513, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/3a099067d201f95e92c3386783275c33f2e4c8b2a56b18f3527f8edca9ed3c16.jpg", + "image_caption": [ + "Figure 1. Overview of our method. The image is encoded into a feature map by the CLIP image encoder for both localization and classification. The regional feature is extracted by pooling the feature map, and then prompted before classified by the CLIP class name embeddings. The anchor boxes are pre-matched and conditioned on a class before decoding. During training, per-class post-matching is conducted. During inference, the box predictions are classified by the region classifier." + ], + "image_footnote": [], + "bbox": [ + 109, + 99, + 859, + 435 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Overview", + "text_level": 1, + "bbox": [ + 76, + 537, + 187, + 551 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The overall framework of CORA is illustrated in Fig. 1. Given an image as input, we acquire the spatial feature map using the ResNet backbone from the pre-trained CLIP image encoder, which is shared by both region classification and object localization branches. Unlike conventional detectors, localization and classification are decoupled and sequentially conducted in our framework to better fit the characteristic of the OVD problem. We train a DETR-style object localizer that refines a set of object queries together with their associated anchor boxes to localize the objects, which are then classified by a region classifier adapted from CLIP.", + "bbox": [ + 75, + 561, + 468, + 733 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Region Classification. Given a region to be classified (anchor box or box prediction), we adopt RoIAlign to obtain the region feature, followed by the attention pooling module of CLIP to generate region embeddings, which can be classified by class embeddings obtained from the CLIP text encoder, as done in CLIP. We name this module as CLIP-based region classifier (Fig. 1-(a)).", + "bbox": [ + 75, + 738, + 468, + 839 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Object Localization. The visual feature map is firstly refined by the DETR-like encoder, and then fed into the DETR-like decoder. The queries of anchor boxes are firstly classified by the CLIP-based region classifier, which are", + "bbox": [ + 75, + 842, + 468, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "then conditioned on their predicted labels before iteratively refined by the DETR-like decoder for better localization. The decoder also estimates the matchability of the query with the previously predicted label. During training, the predicted boxes are one-to-one matched with ground truth boxes that has the same labels (Fig. 1-(b)), and trained as in DETR. In the inference stage, the class labels of the boxes are adjusted by the CLIP-based region classifier.", + "bbox": [ + 496, + 537, + 890, + 652 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "As mentioned in Sec. 1, there are two obstacles to be addressed: (1) object detection conducts recognition on image regions, while the CLIP model is trained on whole-image input, leading to a distribution gap that hinders classification performance. (2) the detector needs to learn object localization for novel classes, while we only have annotations on a limited number of base classes. To solve the first obstacle, we propose region prompting to modulate the region features for better generalizable region embeddings, which will be introduced in Sec. 3.2. To solve the second obstacle, we put forward anchor pre-matching to encourage class-aware object localization that can generalize to novel classes during inference, which will be introduced in Sec. 3.3.", + "bbox": [ + 496, + 652, + 892, + 839 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Region Prompting", + "text_level": 1, + "bbox": [ + 500, + 848, + 679, + 864 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "OVD requires the detector to classify image regions into a given category list. In this section, we will elaborate how", + "bbox": [ + 500, + 869, + 890, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "7033", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/d4b96808f0036c773207a589026eb6d4894e7a822810cea81d943805e2e19cff.jpg", + "image_caption": [ + "(a) Whole-image encoding" + ], + "image_footnote": [], + "bbox": [ + 81, + 95, + 460, + 167 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/093ce500705802d0b9bcbb337745ddd9a613030762f101636c7d7f5bd7cc1949.jpg", + "image_caption": [ + "(b) Regional encoding" + ], + "image_footnote": [], + "bbox": [ + 81, + 203, + 460, + 292 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "a pre-trained CLIP model is adapted to formulate the CLIP-based region classifier. Given the CLIP model, region classification can be realized by comparing the similarity between the regional embedding from the CLIP image encoder and the class name embedding from the CLIP text encoder.", + "bbox": [ + 75, + 402, + 467, + 489 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Region Prompting. As illustrated in Fig. 2, given an image and a set of region of interest (RoI), we firstly encode the whole image into a feature map by the CLIP encoder's first 3 blocks, which are then pooled by RoIAlign [32] either according to anchor boxes or predicted boxes into region features, before encoded by the last block of the CLIP image encoder backbone. There exists a distribution gap between the CLIP image encoder's whole-image feature map and the pooled regional features. We propose region prompting to fix the misalignment by augmenting the region feature with learnable prompts $p \\in \\mathbb{R}^{S \\times S \\times C}$ , where $S$ is the spatial size of the regional feature, and $C$ is the dimension of the regional feature. Specifically, given the input regional feature $f_{\\mathrm{region}}$ , the region prompting is conducted as", + "bbox": [ + 75, + 489, + 467, + 691 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nv _ {\\text {p r o m p t}} = P \\left(f _ {\\text {r e g i o n}} \\oplus p\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 181, + 702, + 468, + 719 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\oplus$ denotes element-wise addition, $P$ is the attention pooling module of the CLIP visual encoder.", + "bbox": [ + 75, + 727, + 467, + 755 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Optimizing Region Prompts. We train the region prompts on a detection dataset with base-class annotations. The class name embeddings are pre-computed by the CLIP text encoder, which are used as classifier weights later. We train the prompts by a standard cross-entropy loss to classify the ground truth boxes with their pooled regional features $f_{\\mathrm{region}}$ . When optimizing the region prompts, we keep other model weights frozen, and only make the region prompts to be learned.", + "bbox": [ + 75, + 756, + 467, + 883 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Comparison with existing methods. The previous com", + "bbox": [ + 76, + 885, + 467, + 900 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/5f22df4d1ddf32c8fd272c91cd51d2a1f68a26cae1494f8fb0add14df39bedca.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 509, + 90, + 867, + 167 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/ea644c72a4a7a810463b136d9cc3252acc7d45a5b9b07513a6397e57585bfe83.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 509, + 172, + 866, + 251 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/c79b27b5ef350e5f411263794dd0c7246d457b7f0fd6fb47c5cf91687b8f533b.jpg", + "image_caption": [ + "(a) Conditional matching" + ], + "image_footnote": [], + "bbox": [ + 524, + 260, + 866, + 335 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/8518f2ee0e91295835e023b3060fabc7fff020afa782f935077b24e536c1bee2.jpg", + "image_caption": [ + "Figure 2. Comparison between the CLIP-based region classifier and the vanilla pipeline. We pool regional feature from the feature map, rather than cropping the region patch and classify them as a separate image.", + "(b) Anchor pre-matching", + "Figure 3. Comparison between anchor pre-matching and conditional matching. Anchor pre-matching decodes a constant number of object queries, and assigns different numbers of object queries based on the image content, avoiding repetitive decoding as in conditional matching." + ], + "image_footnote": [], + "bbox": [ + 522, + 358, + 869, + 460 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "mon practice for region classification with CLIP is to crop the RoIs, and encode them as separate images, before comparing with text embeddings. This pipeline is not efficient when encoding regions with overlaps, since the overlapping regions are encoded more than once in different region crops. Its accuracy also suffers from the missing context information. In contrast, our regional prompting is more efficient and preserves richer context.", + "bbox": [ + 496, + 580, + 890, + 696 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The region prompts contain less than 1M parameters, which is in line with recent advances of the prompt tuning and adapter literatures. Region prompting generalizes well to unseen novel classes. We attribute the generalization capability to the fact that region prompting directly fix the distribution mismatch right after where it occurs (after region pooling), whereas the existing methods tweak irrelevant parameters to compensate for the distribution mismatch.", + "bbox": [ + 496, + 696, + 890, + 810 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Anchor Pre-Matching", + "text_level": 1, + "bbox": [ + 500, + 819, + 705, + 835 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Region prompting helps solve the region classification problem. Object localization is the other critical subtask of object detection. Considering the inferior performance of pre-trained RPN on novel classes, we introduce", + "bbox": [ + 496, + 842, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "7034", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "a class-aware query-based object localizer, which demonstrates better generalization capability on unseen classes. As shown in Fig. 1, Given the visual feature map from the frozen CLIP image encoder, the object queries are prematched to the class name embeddings by the CLIP text encoder.", + "bbox": [ + 75, + 90, + 468, + 176 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Anchor Pre-Matching. The object localizer is implemented by a DETR-style transformer encoder-decoder structure, where the encoder refines the feature map, and the decoder decodes a set of object queries into box predictions. We adopt DAB-DETR [26], where each object query is associated with an anchor box. Each ground truth box is pre-matched to a set of queries with the same label. The label $\\hat{c}_i$ of an object query is assigned by classifying the associated anchor box $b_i$", + "bbox": [ + 75, + 176, + 470, + 306 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {c} _ {i} = \\underset {c \\in C ^ {B}} {\\arg \\max } \\operatorname {c o s i n e} \\left(v _ {i}, l _ {c}\\right), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 176, + 319, + 468, + 344 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $v_{i}$ is the region feature of anchor box $b_{i}, l_{c}$ is the class name embedding of class $c$ , and cosine denotes the cosine similarity. After pre-matching, each object query conditions on the predicted class embedding to allow class-aware box regression. The conditioned object query is given by", + "bbox": [ + 75, + 356, + 468, + 429 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nq _ {i} = \\mathrm {M L P} \\left(l _ {\\mathrm {c} _ {\\mathrm {i}}}\\right). \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 217, + 440, + 468, + 455 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The DETR-like decoder iteratively refines each object query with its associated anchor box $(q_i, b_i)$ into $\\hat{y}_i = (\\hat{p}_i, \\hat{b}_i)$ , where $\\hat{b}_i$ is the refined box coordinates and $\\hat{p}_i$ is the matching probability to the query's pre-matched class $\\hat{c}_i$ .", + "bbox": [ + 76, + 468, + 468, + 541 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Given the model predictions, the assignment between ground truth boxes and model predictions is conducted by performing bipartite matching for each class separately. We only allow each ground truth box to be assigned to the prediction with the same pre-matched label, in order to enforce the decoder to be aware of the conditioned text embedding.", + "bbox": [ + 75, + 542, + 468, + 628 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Specifically, for class $c$ , given the $N^c$ box predictions $\\hat{y}^c = \\{\\hat{y}_i \\mid \\hat{c}_i = c\\}$ that are pre-matched to class $c$ , and the set of ground truth boxes $y^c$ in class $c$ , we optimize a permutation of $N^c$ elements $\\sigma \\in \\mathfrak{S}_{\\mathrm{N}^c}$ that minimizes the following cost", + "bbox": [ + 75, + 628, + 468, + 700 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\sigma} _ {c} = \\underset {\\sigma \\in \\mathfrak {S} _ {\\mathrm {N} ^ {\\mathrm {c}}}} {\\arg \\max } \\sum_ {i} ^ {N ^ {c}} \\mathcal {L} _ {\\operatorname {c o s t}} \\left(y _ {i} ^ {c}, \\hat {y} _ {\\sigma (\\mathrm {i})} ^ {c}\\right), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 155, + 714, + 468, + 753 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where the matching cost is defined as", + "bbox": [ + 76, + 767, + 326, + 781 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {c o s t}} (y, \\hat {y}) = \\mathcal {L} _ {\\text {m a t c h}} (p, \\hat {p}) + \\mathcal {L} _ {\\text {b o x}} (b, \\hat {b}). \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 135, + 794, + 468, + 811 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "$\\mathcal{L}_{\\mathrm{match}}(p,\\hat{p})$ is a binary classification loss, and $\\mathcal{L}_{\\mathrm{box}}(b,\\hat{b})$ characterizes the localization error of $\\hat{b}$ w.r.t $b$ . In our case, we implement $\\mathcal{L}_{\\mathrm{match}}$ by the focal loss [23]. $\\mathcal{L}_{\\mathrm{box}}$ is implemented by a weighted sum of $L_{1}$ loss and GIoU [37] loss following prior works.", + "bbox": [ + 76, + 825, + 468, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The model is optimized by the following loss", + "bbox": [ + 519, + 90, + 821, + 106 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} = \\sum_ {c \\in C ^ {B}} \\mathcal {L} _ {\\text {m a t c h}} \\left(p ^ {c}, \\hat {p} _ {\\hat {\\sigma} _ {c}} ^ {c}\\right) + \\mathcal {L} _ {\\text {b o x}} \\left(b ^ {c}, \\hat {b} _ {\\hat {\\sigma} _ {c}} ^ {c}\\right) \\tag {6} \\\\ = \\lambda_ {\\text {f o c a l}} \\mathcal {L} _ {\\text {f o c a l}} + \\lambda_ {\\mathrm {L} _ {1}} \\mathcal {L} _ {\\mathrm {L} _ {1}} + \\lambda_ {\\mathrm {G I o U}} \\mathcal {L} _ {\\mathrm {G I o U}}. \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 547, + 114, + 890, + 166 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "During inference, we adopt the region classifier introduced in Sec. 3.2 to classify the predicted boxes $\\{\\hat{b}_i\\}$ for better classification accuracy. The class score is multiplied by the pre-matching score to account for the box quality", + "bbox": [ + 496, + 174, + 890, + 234 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {P} \\left(\\hat {b} _ {i} \\in c\\right) = \\hat {p} _ {i} \\operatorname {c o s i n e} \\left(\\hat {v} _ {i}, l _ {c}\\right). \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 596, + 243, + 890, + 260 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Comparison with Conditional Matching [35]. Conditional Matching in OV-DETR [35] also proposes to condition the queries on the text embedding for class-aware regression. But it suffers from repetitive per-class inference. Specifically, as shown in Fig. 3, each class in $C^N$ needs to be separately localized with the same group of query, which means both computation and memory consumption scales linearly with the vocabulary size. During training, the number of negative classes sampled in each iteration is limited due to the memory constraint, which hinders convergence. During inference, repetitive per-class decoding is required and results in low inference efficiency, especially when the vocabulary size is large.", + "bbox": [ + 496, + 268, + 890, + 455 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Contrary to conditional matching, our anchor pre-matching mechanism assigns anchor boxes for different classes adaptively according to the image content, which ensures a constant number of query decoupled from the category size. By anchor pre-matching, all the classes can be decoded together in one pass, eliminating the need for repetitive per-class decoding.", + "bbox": [ + 496, + 455, + 890, + 556 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To improve the generalization capability and training convergence of open-vocabulary detectors equipped with Anchor Pre-Matching, we also introduce two effective training techniques, namely, \"Drop Class\" and \"CLIP-Aligned Labeling\".", + "bbox": [ + 496, + 556, + 890, + 627 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Class Dropout. The generalization capability of the model can be further boosted by randomly dropping categories during training. Since our goal is to train a detector that can detect objects from a user specified category list, training on a fixed list of categories leads to bias. We mitigate this bias by randomly dropping out the base categories during training. For efficiency reason, we implement this idea by splitting the base classes into two complementary groups and train on both of them, instead of training on one group while dropping the other group. Specifically, in each training iteration, we split $C^B$ and the ground truth boxes at a probability of $p$ , and train the detector on the same image with two complementary sets of categories. It enforces the model to condition its prediction on the query categories. Since a ground truth box in one group does not appear in the other group, the model needs to be aware of the categories to treat differently for the two sets of annotations.", + "bbox": [ + 496, + 627, + 890, + 869 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "CLIP-Aligned Labeling. Directly training the localizer on the original COCO dataset suffers from convergence issue.", + "bbox": [ + 500, + 871, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "7035", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "By anchor pre-matching mechanism, a ground truth box incorporates training only when at least one query with the same pre-matched label exists. Otherwise, it is ignored, which hinders convergence. This issue can be partially attributed to the inaccurate anchor box. However, even if a ground truth box has an accurate anchor box, it may still be ignored due to the limited recognition accuracy limitation of the region classifier, or in other words, the ground truth box label is not aligned with the CLIP region classifier used for pre-matching. Thus, we relabel the boxes in the training dataset with the region classifier, which we refer to as CLIP-Aligned labeling. With this technique, more ground truth boxes can be matched.", + "bbox": [ + 75, + 90, + 472, + 277 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 76, + 292, + 209, + 309 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this section, we comprehensively evaluate our CORA on the open-vocabulary detection task. Datasets and evaluation protocols are introduced in Sec. 4.1, and implementation details of our method are provided in Sec. 4.2. We compare with state-of-the-art methods in Sec. 4.3, demonstrating advantages of our CORA, and then validate the effectiveness of the proposed Region Prompting and Anchor Pre-Matching in Sec. 4.4 and Sec. 4.5, respectively.", + "bbox": [ + 75, + 318, + 470, + 434 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Dataset & Training & Evaluation", + "text_level": 1, + "bbox": [ + 76, + 444, + 370, + 460 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Following the convention of the COCO OVD benchmark proposed in [4], the 80 classes in the COCO dataset [24] classes are divided into 48 base classes and 17 novel classes. The model is trained on the 48 base classes, which contains 107,761 images and 665,387 instances. The model is then evaluated on the validation set of novel classes, which contains 4,836 images and 33,152 instances from both the 48 base classes and the 17 novel classes. We also conduct experiments on the LVIS v1.0 [16] dataset. On LVIS dataset, the model is trained on 461 common classes and 405 frequent classes, which contains 100,170 images and 1,264,883 instances. After training, the model is evaluated on the LVIS validation set, which contains 19,809 images and 244,707 instances.", + "bbox": [ + 75, + 468, + 468, + 667 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To make fairer comparisons with other methods, we propose $\\mathrm{CORA^{+}}$ , which utilizes extra dataset or target novel class names. $\\mathrm{CORA^{+}}$ is a detector trained with both ground-truth base-category annotations and additional pseudo bounding box labels. When target class names are provided, we generate pseudo boxes of novel classes on the base training dataset using CORA, and when extra imagedataset (text could be captions or class names) is available, pseudo boxes of all objects mentioned by the corresponding text are generated. We use standard detector architecture and training target to train $\\mathrm{CORA^{+}}$ . SAMDETR [38] is used for the experiment on COCO and CenterNet2 [43] is used for the experiment on LVIS.", + "bbox": [ + 75, + 670, + 468, + 856 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In the OVD task, we evaluate our model under the \"generalized\" setting, in which the model needs to predict objects from both base and novel classes, and then evaluated", + "bbox": [ + 75, + 857, + 468, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "on novel objects. On the COCO benchmark, we take AP50 as our evaluation metric, which counts the average precision at an intersection over union (IoU) of $50\\%$ for each class, and then averages among all the classes. For the LVIS OVD benchmark, we evaluate on the full validation dataset, and report the mean AP of boxes from the novel classes to compare with prior works [40]. For the region classification task, we let the model classify the ground truth boxes in the COCO dataset, the performance is evaluated in terms of mAP.", + "bbox": [ + 496, + 90, + 890, + 234 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Implementation Details", + "text_level": 1, + "bbox": [ + 498, + 243, + 718, + 258 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Model Specifications. We use DAB-DETR [26] as the object localizer. Specifically, localizer is configured to have 1,000 object queries, 3 encoder layers and 6 decoder layers. We use a multi-layer perceptron (MLP) with 128 hidden neurons to transform class name embeddings into an object query. Following CLIP [30], each class embedding is computed as the average text embedding of the class name over 80 context prompts. In class dropout, each class is randomly assigned to one of the two groups with equal probabilities. When adopting region prompting on LVIS, classes in the common and frequent group are sampled with equal weights, meaning that the objects in the less frequent classes are over sampled. When training our method on LVIS, we sample 100 categories (including the ground truth categories) in each iteration. Since the number of classes in the LVIS dataset is much larger than that of COCO, we relax the matching constraint in anchor pre-matching, such that ground truth boxes can be post-matched to the anchor boxes with a similar label. Specifically, classes with a cosine similarity greater than 0.7 are considered similar.", + "bbox": [ + 496, + 265, + 890, + 551 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Training & Hyperparameters. We train the region prompts for 5 epochs with a base learning rate of $10^{-4}$ , which decays after the $4^{th}$ epoch by a factor of 0.1. The localizer is trained for 35 epochs with a learning rate of $10^{-4}$ without learning rate decay. Both the region prompts and the localizer are trained with batch size 32 by the AdamW optimizer [28] with $10^{-4}$ weight decay. We apply gradient clipping with a maximal norm of 0.1. To stabilize training, we evaluate on the exponential moving average (EMA) of the model after training. The class dropout probability $p$ is set as 0.2. $\\lambda_{\\mathrm{focal}}$ , $\\lambda_{\\mathrm{L_1}}$ and $\\lambda_{\\mathrm{GIOU}}$ are set as 2.0, 5.0, 2.0, respectively. For the experiment on LVIS, we use repeat factor sampling [16] with default hyperparameters to balance the training samples. We use non-maximum suppression (NMS) with an IoU threshold 0.5 during inference.", + "bbox": [ + 496, + 553, + 890, + 768 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Comparison with State-of-the-Art Methods", + "text_level": 1, + "bbox": [ + 498, + 777, + 867, + 792 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Tab. 1 summarizes our main results. Since the pretrained model is crucial to the open-vocabulary capability of the detector, we compare our method with baseline methods that are trained with the same CLIP model. When compared with methods trained on CLIP RN50, CORA outperforms VL-PLM by 0.7 AP50 on novel classes. With a larger pre-trained model, our method improves the previous state", + "bbox": [ + 496, + 799, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "7036", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/eb02657d0b54306fd0a5370a1888ca26714cd5469306edd68bb85124c15d28ba.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodExtra DatasetDetector Training \nPre-train ModelRequire Novel ClassGeneralized (17 + 48)
NovelBaseAll
OVR-CNN [36]COCO Captions [9]-X22.846.039.9
Detic [42]COCO Captions [9]CLIP (text encoder)X27.847.145.0
RegionCLIP [40]CC3M [33]CLIP (RN50)X31.457.150.4
VL-PLM [39]-CLIP (RN50)✓34.460.253.5
CORA (Ours)-CLIP (RN50)X35.135.535.4
ViLD [14]-CLIP (ViT-B/32)✓27.659.951.3
OV-DETR [35]-CLIP (ViT-B/32)✓29.461.052.7
MEDet [7]COCO Captions [9]CLIP (ViT-B/32)X32.654.049.4
RegionCLIP [40]CC3M [33]CLIP (RN50x4)X39.361.655.7
CORA (Ours)-CLIP (RN50x4)X41.744.543.8
CORA+ (Ours)COCO Captions [9]CLIP (RN50x4)X43.160.956.2
", + "bbox": [ + 122, + 88, + 849, + 301 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/343112e54549f70889ebd375951aeb3822800bcda2e494503042d7638696ffeb.jpg", + "table_caption": [ + "Table 1. Main results on the COCO OVD benchmark. We report AP50 as the evaluation metric. The baseline methods are grouped by their pre-trained model. We also list the extra dataset requirement of each method, and whether they require the novel class to be provided during training." + ], + "table_footnote": [], + "table_body": "
MethodDetector TrainingLVIS APr
Ext. DataPre-train Model
ViLD [14]-CLIP (ViT-B/32)16.3
OV-DETR [35]-CLIP (ViT-B/32)17.4
RegionCLIP [40]CC3MCLIP (RN50x4)22.0
CORA (Ours)-CLIP (RN50x4)22.2
MEDet [7]CC3MCLIP (ViT-B/32)22.4
Detic [42]IN-21kCLIP (text encoder)26.2
CORA+ (Ours)IN-21kCLIP (text encoder)28.1
", + "bbox": [ + 84, + 376, + 460, + 512 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "of-the-art RegionCLIP [40] by 2.4 AP50. When extra data is available, the performance can be further boosted to 43.1 AP50. The results on LVIS [16] OVD benchmark is shown in Tab. 2.", + "bbox": [ + 75, + 566, + 468, + 625 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Note that among the baseline methods, VL-PLM [39], ViLD [14] and OV-DETR [35] use novel class names during training in order to recognize the potential novel objects and assign pseudo labels for them. Consequently, a new detector needs to be trained whenever there is a new set of categories to be detected. CORA can generalize to any combination of novel categories once trained without tedious re-training. Other compared methods relying on CLIP require image tag annotations extracted from language descriptions [7,31,40] or image labels [42] during training. We argue that the extra annotations do not provide additional information over CLIP. Instead, they serve as a media to transfer the knowledge from CLIP to the detector. Our method directly adapts the CLIP model to obtain region classifier, thus no extra image-text data is needed.", + "bbox": [ + 75, + 625, + 468, + 840 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In this work, both region prompting and anchor prematching aim to generalize the knowledge learned from base classes to novel classes. Consequently, the performance gap between novel and base classes is significantly", + "bbox": [ + 75, + 842, + 468, + 902 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/ae4284a3c3b9606e4f54b7d79ab2b6b62ffd01d7ef793cfd40c6d38815a35908.jpg", + "table_caption": [ + "Table 2. Results on the LVIS [16] OVD benchmark." + ], + "table_footnote": [], + "table_body": "
MethodCLIP modelNovelBase
CLIPRN5058.258.6
CLIP-Adapter [12]RN5063.080.6
CoOp [41]RN5064.475.7
CORARN5065.170.0
CLIPRN50x463.962.7
CORARN50x474.176.0
", + "bbox": [ + 516, + 376, + 877, + 500 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 3. Results on the region classification task evaluated in mAP. We compare our method with the original CLIP regional classifier and other baseline methods.", + "bbox": [ + 498, + 508, + 890, + 550 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "lowered than the compared methods. Note that in OVD, the performance is evaluated by the generalization capability towards novel classes, rather than the bases classes on which they are trained.", + "bbox": [ + 498, + 577, + 890, + 637 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Effectiveness of Region Prompting", + "text_level": 1, + "bbox": [ + 498, + 646, + 802, + 662 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Region Prompting. Since object classification is decoupled from localization in this work, the CLIP-based region classifier can be directly evaluated by the region classification task. After trained on the base-class annotations of COCO dataset, we evaluate the CLIP-based region classifier on both base and novel classes in the validation set. We use mean average precision (mAP) as our evaluation metric.", + "bbox": [ + 496, + 669, + 890, + 770 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Tab. 3 shows our main result. Directly evaluating on the CLIP model without further training already achieves considerable performance of $58.2\\mathrm{mAP}$ on the novel classes. We compare our result with two competitive methods from the adapter and prompt tuning literatures. CLIP-Adapter [12] adopts an additional bottleneck layer to learn new features and performs residual style feature blending with the original pre-trained features. CoOp [41] prepends shared learnable prompts to the text embeddings before en", + "bbox": [ + 496, + 771, + 890, + 901 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7037", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/49bde419eb84b97d7b60b73215d3a45690acca51aa890599dc13192a4173edb7.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
CLIP modelFeatureNovelBase
RN50whole-image43.840.5
RN50regional58.658.2
", + "bbox": [ + 114, + 88, + 433, + 148 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/3fde7f0a72ade9a5c93754a8870e94aaa2daa5f867fdd7e6188d62fd7aef9b43.jpg", + "table_caption": [ + "Table 4. Comparison of the whole-image classification pipeline and the regional classification pipeline. Results are reported in mAP." + ], + "table_footnote": [], + "table_body": "
Query ConditioningPer-Class Post-MatchingCLIP-Aligned LabelingNovel
XX✓26.0
✓X✓29.9
✓✓X40.9
✓✓✓41.7
", + "bbox": [ + 84, + 213, + 460, + 311 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 5. Ablation studies on anchor pre-matching and CLIP-aligned labeling. Anchor pre-matching consists of query conditioning and per-class post-matching.", + "bbox": [ + 76, + 321, + 467, + 362 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "coded by the CLIP text encoder. Experiments show that the compared methods bias strongly towards base classes. Note that the baseline methods adapts the CLIP model by tuning the text input or output feature, which are irrelevant or distant to the regional feature, where the mismatch between regional and whole-image feature occurs. On the contrary, region prompting directly prompts the mismatched features, thus generalizes better to the novel classes, achieving a performance gain of $6.9\\mathrm{mAP}$ over CLIP on the novel classes. Region prompting also scales with larger backbones. On the $\\mathrm{RN50x4}$ CLIP backbone, region prompting further boosts CLIP by $10.2\\%$ mAP over the corresponding CLIP model.", + "bbox": [ + 75, + 388, + 467, + 561 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Comparison with whole-image classification. A common practice of region classification by CLIP is to crop the regions and classify them as separate images. We compare the region classifier with the common practice on the original CLIP weights without region prompting. As shown in Tab. 4, when classifying a region as a whole-image, the performance is significantly lower than using the regional feature, despite the extra computation. We attribute the performance gap to the missing context of cropped images.", + "bbox": [ + 75, + 563, + 467, + 691 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.5. Effectiveness of Anchor Pre-Matching", + "text_level": 1, + "bbox": [ + 76, + 704, + 405, + 720 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We conduct ablation studies to validate anchor pre-matching and the proposed training techniques.", + "bbox": [ + 76, + 727, + 467, + 755 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Anchor Pre-Matching. Anchor pre-matching consists of two separate operations: query conditioning and per-class post-matching. We analyze their effects in Tab. 5. Firstly, we train a model without anchor pre-matching. We keep using DAB-DETR [26] as the localizer, and use the same number of object queries and anchor boxes for fair comparison. The object queries are not pre-matched to classes, and are initialized as 0 as done in DAB-DETR. Since the model predictions are not pre-matched, the per-class post-matching is replaced by the vanilla one-to-one matching", + "bbox": [ + 75, + 756, + 467, + 900 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/d07fdb4c783b7b97ee3fc4edcd3da4d29bc26fa0f53d2efa680078d996c5d90e.jpg", + "image_caption": [ + "Figure 4. Different choices of $p$ for class dropout." + ], + "image_footnote": [], + "bbox": [ + 527, + 112, + 857, + 294 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "mechanism in DETR. Then, we classify the anchor boxes and condition the object queries on the class name embeddings, but do not set constraint on the post-matching, which boosts the performance on novel classes by 3.9 AP50. After adopting the full anchor pre-matching, the performance is significantly boosted by 10.8 AP50.", + "bbox": [ + 496, + 345, + 890, + 430 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Training Techniques. Class dropout and CLIP-aligned labeling are two training techniques that help the model generalize better. In Fig. 4, we examine the effect of different dropout probability. We find that the models trained with class dropout consistently outperform the baseline, and $p = 0.2$ gives the best performance. In Tab. 5, we validate the effectiveness of CLIP-Aligned Labeling.", + "bbox": [ + 496, + 431, + 890, + 532 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 545, + 617, + 560 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The core challenge in open-vocabulary detection is how to effectively transfer the knowledge learned from the base classes to the unseen novel classes for evaluation. In this work, we directly adapt the CLIP into a region classifier, and mitigate the distribution gap between whole-image feature and regional feature through region prompting, which successfully generalizes to the novel categories. Different from prior works that rely on a fixed RPN for novel class localization, we achieve efficient class-aware localization by the proposed anchor pre-matching mechanism. Experiments show that our method can better transfer the knowledge from base classes to unseen novel classes with a smaller gap than prior works. We hope our work can help other researchers gain better insight on the OVD problem and develop better open-vocabulary detectors.", + "bbox": [ + 496, + 570, + 890, + 785 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgement. This project is funded in part by National Key R&D Program of China Project 2022ZD0161100, by the Centre for Perceptual and Interactive Intelligence (CPII) Ltd under the Innovation and Technology Commission (ITC)'s InnoHK, by General Research Fund of Hong Kong RGC Project 14204021. Hongsheng Li is a PI of CPII under the InnoHK. This project is also supported by SenseTime Collaborative Research Grant.", + "bbox": [ + 496, + 785, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "7038", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Deformable detr: Deformable transformers. 2020. 2", + "[2] Hyojin Bahng, Ali Jahanian, Swami Sankaranarayanan, and Phillip Isola. Exploring visual prompts for adapting large-scale models. arXiv preprint arXiv:2203.17274, 2022. 2", + "[3] Hyojin Bahng, Ali Jahanian, Swami Sankaranarayanan, and Phillip Isola. Visual prompting: Modifying pixel space to adapt pre-trained models. ArXiv, abs/2203.17274, 2022. 2", + "[4] Ankan Bansal, Karan Sikka, Gaurav Sharma, Rama Chellappa, and Ajay Divakaran. Zero-shot object detection. ArXiv, abs/1804.04340, 2018. 6", + "[5] Tom B. Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert-Voss, Gretchen Krueger, T. J. Henighan, Rewon Child, Aditya Ramesh, Daniel M. Ziegler, Jeff Wu, Clemens Winter, Christopher Hesse, Mark Chen, Eric Sigler, Mateusz Litwin, Scott Gray, Benjamin Chess, Jack Clark, Christopher Berner, Sam McCandlish, Alec Radford, Ilya Sutskever, and Dario Amodei. Language models are few-shot learners. ArXiv, abs/2005.14165, 2020. 2", + "[6] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. ArXiv, abs/2005.12872, 2020. 1, 2", + "[7] Peixian Chen, Kekai Sheng, Mengdan Zhang, Yunhang Shen, Ke Li, and Chunhua Shen. Open vocabulary object detection with proposal mining and prediction equalization. CoRR, abs/2206.11134, 2022. 1, 7", + "[8] Qiang Chen, Xiaokang Chen, Gang Zeng, and Jingdong Wang. Group detr: Fast training convergence with decoupled one-to-many label assignment. ArXiv, abs/2207.13085, 2022. 2", + "[9] Xinlei Chen, Hao Fang, Tsung-Yi Lin, Ramakrishna Vedantam, Saurabh Gupta, Piotr Dólar, and C. Lawrence Zitnick. Microsoft coco captions: Data collection and evaluation server. ArXiv, abs/1504.00325, 2015. 7", + "[10] Chengjian Feng, Yujie Zhong, Zequn Jie, Xiangxiang Chu, Haibing Ren, Xiaolin Wei, Weidi Xie, and Lin Ma. Promptdet: Towards open-vocabulary detection using uncurated images. 2022. 2", + "[11] Mingfei Gao, Chen Xing, Juan Carlos Niebles, Junnan Li, Ran Xu, Wenhao Liu, and Caiming Xiong. Open vocabulary object detection with pseudo bounding-box labels. arXiv preprint arXiv:2111.09452, 2021. 2", + "[12] Peng Gao, Shijie Geng, Renrui Zhang, Teli Ma, Rongyao Fang, Yongfeng Zhang, Hongsheng Li, and Yu Jiao Qiao. Clip-adapter: Better vision-language models with feature adapters. ArXiv, abs/2110.04544, 2021. 7", + "[13] Peng Gao, Minghang Zheng, Xiaogang Wang, Jifeng Dai, and Hongsheng Li. Fast convergence of detr with spatially modulated co-attention. 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 3601–3610, 2021. 2", + "[14] Xiuye Gu, Tsung-Yi Lin, Weicheng Kuo, and Yin Cui. Open-vocabulary object detection via vision and language knowledge distillation. In ICLR, 2022. 1, 7", + "[15] Xiuye Gu, Tsung-Yi Lin, Weicheng Kuo, and Yin Cui. Open-vocabulary object detection via vision and language knowledge distillation. In ICLR, 2022. 2" + ], + "bbox": [ + 78, + 114, + 468, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[16] Agrim Gupta, Piotr Dollar, and Ross Girshick. LVIS: A dataset for large vocabulary instance segmentation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2019. 2, 6, 7", + "[17] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In 2017 IEEE International Conference on Computer Vision (ICCV), pages 2980-2988, 2017. 1", + "[18] Ding Jia, Yuhui Yuan, Haodi He, Xiao pei Wu, Haojun Yu, Weihong Lin, Lei huan Sun, Chao Zhang, and Hanhua Hu. Detrs with hybrid matching. ArXiv, abs/2207.13080, 2022. 2", + "[19] Menglin Jia, Luming Tang, Bor-Chun Chen, Claire Cardie, Serge J. Belongie, Bharath Hariharan, and Ser Nam Lim. Visual prompt tuning. ArXiv, abs/2203.12119, 2022. 2", + "[20] Chen Ju, Tengda Han, Kunhao Zheng, Ya Zhang, and Weidi Xie. Prompting visual-language models for efficient video understanding. ArXiv, abs/2112.04478, 2021. 2", + "[21] Liunian Harold Li, Pengchuan Zhang, Haotian Zhang, Jianwei Yang, Chunyuan Li, Yiwu Zhong, Lijuan Wang, Lu Yuan, Lei Zhang, Jenq-Neng Hwang, Kai-Wei Chang, and Jianfeng Gao. Grounded language-image pre-training. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10955–10965, 2022. 2", + "[22] Xiang Lisa Li and Percy Liang. Prefix-tuning: Optimizing continuous prompts for generation. Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), abs/2101.00190, 2021. 2", + "[23] Tsung-Yi Lin, Priya Goyal, Ross B. Girshick, Kaiming He, and Piotr Dólar. Focal loss for dense object detection. 2017 IEEE International Conference on Computer Vision (ICCV), pages 2999-3007, 2017. 5", + "[24] Tsung-Yi Lin, Michael Maire, Serge J. Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C. Lawrence Zitnick. Microsoft coco: Common objects in context. In ECCV, 2014. 2, 6", + "[25] Ziyi Lin, Shijie Geng, Renrui Zhang, Peng Gao, Gerard de Melo, Xiaogang Wang, Jifeng Dai, Yu Qiao, and Hongsheng Li. Frozen clip models are efficient video learners. arXiv preprint arXiv:2208.03550, 2022. 2", + "[26] Shilong Liu, Feng Li, Hao Zhang, Xiao Bin Yang, Xianbiao Qi, Hang Su, Jun Zhu, and Lei Zhang. Dab-detr: Dynamic anchor boxes are better queries for detr. ArXiv, abs/2201.12329, 2022. 2, 5, 6, 8", + "[27] Xiao Liu, Kaixuan Ji, Yicheng Fu, Zhengxiao Du, Zhilin Yang, and Jie Tang. P-tuning v2: Prompt tuning can be comparable to fine-tuning universally across scales and tasks. CoRR, abs/2110.07602, 2021. 2", + "[28] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In ICLR, 2019. 6", + "[29] Muhammad Maaz, Hanoona Rasheed, Salman Khan, Fahad Shahbaz Khan, Rao Muhammad Anwer, and Ming-Hsuan Yang. Class-agnostic object detection with multimodal transformer. In 17th European Conference on Computer Vision (ECCV). Springer, 2022. 1", + "[30] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual" + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "7039", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "models from natural language supervision. In ICML, 2021. 1, 6", + "[31] Hanoona Rasheed, Muhammad Maaz, Muhammad Uzair Khattak, Salman Khan, and Fahad Shahbaz Khan. Bridging the gap between object and image-level representations for open-vocabulary detection. ArXiv, abs/2207.03482, 2022. 1, 7", + "[32] Shaoqing Ren, Kaiming He, Ross B. Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. IEEE Transactions on Pattern Analysis and Machine Intelligence, 39:1137-1149, 2015. 4", + "[33] Piyush Sharma, Nan Ding, Sebastian Goodman, and Radu Soricut. Conceptual captions: A cleaned, hypernymed, image alt-text dataset for automatic image captioning. In ACL, 2018. 7", + "[34] Yinghui Xing, Qirui Wu, De Cheng, Shizhou Zhang, Guoqiang Liang, and Yanning Zhang. Class-aware visual prompt tuning for vision-language pre-trained model. CoRR, abs/2208.08340, 2022. 2", + "[35] Yuhang Zang, Wei Li, Kaiyang Zhou, Chen Huang, and Chen Change Loy. Open-vocabulary detr with conditional matching. ArXiv, abs/2203.11876, 2022. 1, 2, 5, 7", + "[36] Alireza Zareian, Kevin Dela Rosa, Derek Hao Hu, and Shih-Fu Chang. Open-vocabulary object detection using captions. In 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 14388-14397, 2021. 1, 2, 7", + "[37] Hongyu Zhai, Jian Cheng, and Mengyong Wang. Rethink the iou-based loss functions for bounding box regression. 2020 IEEE 9th Joint International Information Technology and Artificial Intelligence Conference (ITAIC), 9:1522-1528, 2020. 5", + "[38] Gongjie Zhang, Zhipeng Luo, Yingchen Yu, Kaiwen Cui, and Shijian Lu. Accelerating DETR convergence via semantic-aligned matching. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 949-958, 2022. 2, 6", + "[39] Shiyu Zhao, Zhixing Zhang, Samuel Schulter, Long Zhao, Vijay Kumar BG, Anastasis Stathopoulos, Manmohan Chandraker, and Dimitris N. Metaxas. Exploiting unlabeled data with vision and language models for object detection. CoRR, abs/2207.08954, 2022. 1, 7", + "[40] Yiwu Zhong, Jianwei Yang, Pengchuan Zhang, Chengkun Li, Noel C. F. Codella, Liunian Harold Li, Luowei Zhou, Xiyang Dai, Lu Yuan, Yin Li, and Jianfeng Gao. Regionclip: Region-based language-image pretraining. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 16772-16782, 2022. 1, 2, 6, 7", + "[41] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Learning to prompt for vision-language models. Int. J. Comput. Vis., 130:2337-2348, 2022. 7", + "[42] Xingyi Zhou, Rohit Girdhar, Armand Joulin, Philipp Krahenbihl, and Ishan Misra. Detecting twenty-thousand classes using image-level supervision. In ECCV, 2022. 7", + "[43] Xingyi Zhou, Vladlen Koltun, and Philipp Krahenbuhl. Probabilistic two-stage detection. In arXiv preprint arXiv:2103.07461, 2021. 6" + ], + "bbox": [ + 78, + 90, + 468, + 858 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "7040", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2023/CORA_ Adapting CLIP for Open-Vocabulary Detection With Region Prompting and Anchor Pre-Matching/e257818b-1aea-4bd1-9288-7c6a81195cd8_model.json b/2023/CORA_ Adapting CLIP for Open-Vocabulary Detection With Region Prompting and Anchor Pre-Matching/e257818b-1aea-4bd1-9288-7c6a81195cd8_model.json new file mode 100644 index 0000000000000000000000000000000000000000..b733051c5b3377b6959135dff58b83ef3ecd4579 --- /dev/null +++ b/2023/CORA_ Adapting CLIP for Open-Vocabulary Detection With Region Prompting and Anchor Pre-Matching/e257818b-1aea-4bd1-9288-7c6a81195cd8_model.json @@ -0,0 +1,2002 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.18, + 0.13, + 0.79, + 0.174 + ], + "angle": 0, + "content": "CORA: Adapting CLIP for Open-Vocabulary Detection with Region Prompting and Anchor Pre-Matching" + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.2, + 0.704, + 0.219 + ], + "angle": 0, + "content": "Xiaoshi Wu\\(^{1}\\), Feng Zhu\\(^{2}\\), Rui Zhao\\(^{2,3}\\), Hongsheng Li\\(^{1,4}\\)" + }, + { + "type": "text", + "bbox": [ + 0.234, + 0.22, + 0.739, + 0.238 + ], + "angle": 0, + "content": "1Multimedia Laboratory, The Chinese University of Hong Kong" + }, + { + "type": "text", + "bbox": [ + 0.144, + 0.238, + 0.827, + 0.254 + ], + "angle": 0, + "content": "\\(^{2}\\)SenseTime Research \\(^{3}\\)Qing Yuan Research Institute, Shanghai Jiao Tong University" + }, + { + "type": "text", + "bbox": [ + 0.26, + 0.254, + 0.71, + 0.271 + ], + "angle": 0, + "content": "\\(^{4}\\)Centre for Perceptual and Interactive Intelligence (CPII)" + }, + { + "type": "text", + "bbox": [ + 0.176, + 0.273, + 0.791, + 0.288 + ], + "angle": 0, + "content": "{wuxiaoshi@link, hsli@ee}.cuhk.edu.hk,{zhufeng, zhaorui}@sentetime.com" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.323, + 0.313, + 0.339 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.355, + 0.473, + 0.742 + ], + "angle": 0, + "content": "Open-vocabulary detection (OVD) is an object detection task aiming at detecting objects from novel categories beyond the base categories on which the detector is trained. Recent OVD methods rely on large-scale visual-language pre-trained models, such as CLIP, for recognizing novel objects. We identify the two core obstacles that need to be tackled when incorporating these models into detector training: (1) the distribution mismatch that happens when applying a VL-model trained on whole images to region recognition tasks; (2) the difficulty of localizing objects of unseen classes. To overcome these obstacles, we propose CORA, a DETR-style framework that adapts CLIP for Open-vocabulary detection by Region prompting and Anchor pre-matching. Region prompting mitigates the whole-to-region distribution gap by prompting the region features of the CLIP-based region classifier. Anchor pre-matching helps learning generalizable object localization by a class-aware matching mechanism. We evaluate CORA on the COCO OVD benchmark, where we achieve 41.7 AP50 on novel classes, which outperforms the previous SOTA by 2.4 AP50 even without resorting to extra training data. When extra training data is available, we train \\(\\mathrm{CORA^{+}}\\) on both ground-truth base-category annotations and additional pseudo bounding box labels computed by CORA. \\(\\mathrm{CORA^{+}}\\) achieves 43.1 AP50 on the COCO OVD benchmark and 28.1 box APr on the LVIS OVD benchmark. The code is available at https://github.com/tgxs002/CORA." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.789, + 0.208, + 0.804 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.814, + 0.47, + 0.902 + ], + "angle": 0, + "content": "Object detection is a fundamental vision problem that involves localizing and classifying objects from images. Classical object detection requires detecting objects from a closed set of categories. Extra annotations and training are required if objects of unseen categories need to be detected. It has attracted much attention on detecting novel categories" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.324, + 0.892, + 0.367 + ], + "angle": 0, + "content": "without tedious annotations, or even detect object from new category, which is currently referred as open-vocabulary detection (OVD) [36]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.368, + 0.893, + 0.496 + ], + "angle": 0, + "content": "Recent advances on large-scale vision-language pretrained models, such as CLIP [30], enable new solutions for tackling OVD. CLIP learns a joint embedding space of images and text from a large-scale image-text dataset, which shows remarkable capability on visual recognition tasks. The general idea of applying CLIP for OVD is to treat it as an open-vocabulary classifier. However, there are two obstacles hindering the effective use of CLIP on tackling OVD." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.497, + 0.895, + 0.683 + ], + "angle": 0, + "content": "How to adapt CLIP for region-level tasks? One trivial solution is to crop regions and treat them as separate images, which has been adopted by multiple recent works [7, 14, 31, 35]. But the distribution gap between region crops and full images leads to inferior classification accuracy. MEDet [7] mitigates this issue by augmenting the text feature with image features. However, it requires extra image-text pairs to prevent overfitting to so-called \"base\" classes that are seen during training. RegionCLIP [40] directly acquires regional features by RoIAlign [17], which is more efficient but cannot generalize well to novel classes without finetuning. The finetuning is costly when adopting a larger CLIP model." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.685, + 0.895, + 0.87 + ], + "angle": 0, + "content": "How to learn generalizable object proposals? ViLD [14], OV-DETR [35], Object-Centric-OVD [31], Region-CLIP [40] need RPN or class-agnostic object detectors [29] to mine potential novel class objects. However, these RPNs are strongly biased towards the base classes on which they are trained, while perform poorly on the novel classes. MEDet [7] and VL-PLM [39] identify this problem and adopt several handcrafted policies to rule out or merge low-quality boxes, but the performance is still bounded by the frozen RPN. OV-DETR [35] learns generalizable object localization by conditioning box regression on class name embeddings, but at the cost of efficiency issue induced by repetitive per-class inference." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.892, + 0.901 + ], + "angle": 0, + "content": "In this work, we propose a new framework based on DETection TR transformers (DETR) [6] that incorporates CLIP" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "7031" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.081, + 0.092, + 0.468, + 0.35 + ], + "angle": 0, + "content": "into detector training to achieve open-vocabulary detection without additional image-text data. Specifically, we use a DETR-style object localizer for class-aware object localization, and the predicted boxes are encoded by pooling the intermediate feature map of the CLIP image encoder, which are classified by the CLIP text encoder with class names. However, there is a distribution gap between whole-image features from CLIP's original visual encoder and the newly pooled region features, leading to an inferior classification accuracy. Thus, we propose Region Prompting to adapt the CLIP image encoder, which boosts the classification performance, and also demonstrates better generalization capability than existing methods. We adopt DAB-DETR [26] as the localizer, in which object queries are associated with dynamic anchor boxes. By pre-matching the dynamic anchor boxes with the input categories before box regression (Anchor Pre-Matching), class-aware regression can be achieved without the cost of repetitive per-class inference." + }, + { + "type": "text", + "bbox": [ + 0.081, + 0.352, + 0.468, + 0.595 + ], + "angle": 0, + "content": "We validate our method on COCO [24] and LVIS v1.0 [16] OVD benchmarks. On the COCO OVD benchmark, our method improves AP50 of novel categories over the previous best method [40] by 2.4 AP50 without training on extra data, and achieves consistent gain on CLIP models of different scales. When compared under a fairer setting with extra training data, our method significantly outperforms the existing methods by 3.8 AP50 on novel categories and achieves comparable performance on base categories. On the LVIS OVD benchmark, our method achieves 22.2/28.1 APr with/w.o. extra data, which significantly outperforms existing methods that are also trained with/w.o. extra data. By applying region prompting on the base classes of COCO, the classification performance on the novel classes is boosted from \\(63.9\\%\\) to \\(74.1\\%\\), whereas other prompting or adaptation methods easily bias towards the base classes." + }, + { + "type": "text", + "bbox": [ + 0.081, + 0.597, + 0.468, + 0.698 + ], + "angle": 0, + "content": "The contributions of this work are summarized as follows: (1) Our proposed region prompting effectively mitigates the gap between whole image features and region features, and generalize well in the open-vocabulary setting. (2) Anchor Pre-Matching enables DETR for generalizable object localization efficiently. (3) We achieve state-of-the-art performance on COCO and LVIS OVD benchmarks." + }, + { + "type": "title", + "bbox": [ + 0.081, + 0.716, + 0.225, + 0.731 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "text", + "bbox": [ + 0.081, + 0.742, + 0.468, + 0.9 + ], + "angle": 0, + "content": "Open-Vocabulary Object Detection OVR-CNN [36] firstly put forth this new formulation of detection, and proposes its baseline solution by aligning region features with nouns in captions that are paired with the image. Mingfeng [11] et al. mine pseudo labels by utilizing the localization ability of pre-trained vision-language models. PromptDet [10] addresses the gap between image and region classification by adding learnable prompts when encoding the class names, namely regional prompt learning (RPL), which is expected to generalize from base to novel categories. OV-DETR [35] is the first DETR-style open" + }, + { + "type": "text", + "bbox": [ + 0.504, + 0.092, + 0.89, + 0.235 + ], + "angle": 0, + "content": "vocabulary detector, which proposes conditional matching to solve the missing novel class problem in assignment, but at the cost of inefficient inference. RegionCLIP [40] propose a second-stage pre-training mechanism to adapt the CLIP model to encode region features, and demonstrates its capability on OVD and zero-shot transfer setting. GLIP [21] jointly learns object localization and VL alignment. Matthias et al. [15] proposes to finetune a VL aligned model for detection, while we fix the pre-trained VL model for better generalization towards novel categories." + }, + { + "type": "text", + "bbox": [ + 0.504, + 0.241, + 0.89, + 0.472 + ], + "angle": 0, + "content": "Detection Transformers DETR [6] is an object detection architecture based on transformers that formulates object detection as a set-to-set matching problem, which greatly simplifies the pipeline. Several works address the slow convergence problem of DETR by architectural improvement [1, 13,26,38] or special training strategies [8,18]. Zhu et al. [1] proposes multi-scale deformable attention module to efficiently aggregate information from multi-scale feature maps. Gao et al. [13] proposes to modulate the cross-attention in the transformer decoder by anchor box coordinates to accelerate the detector convergence. DABDETR [26] formulates the queries in DETR architecture as anchor boxes, which accelerates detector training. Chen et al. [38] proposes Group DETR, which adds auxiliary object queries during training to take advantage of one-to-many matching for faster convergence." + }, + { + "type": "text", + "bbox": [ + 0.504, + 0.478, + 0.89, + 0.65 + ], + "angle": 0, + "content": "Prompt Tuning Prompting is originated from NLP, and it refers to prepending task instructions before the input sequence to give the language model the hint about the task [5]. Later works [22, 27] explores tuning continuous prompt vectors when few-shot data is available. VPT [19], Visual Prompting [2, 3] explore prompting in the pixel space. [20] and [25] prompts pre-trained model for video recognition tasks. [34] proposes class-aware visual prompt tuning to generalize the learned prompts to unseen categories. Recent works demonstrate that prompt tuning is an effective and parameter-efficient way to adapt large-scale pre-trained models to downstream tasks." + }, + { + "type": "title", + "bbox": [ + 0.504, + 0.682, + 0.588, + 0.698 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.504, + 0.713, + 0.89, + 0.9 + ], + "angle": 0, + "content": "Open-vocabulary detection (OVD) is an object detection task aiming at detecting objects from novel categories beyond the base categories on which the detector is trained. Formally, the detector is trained on a detection dataset with base-category \\( C^B \\) box annotations, and tested on a new input image \\( I \\in \\mathbb{R}^{H \\times W \\times 3} \\) to detect objects belonging to a novel category set \\( C^N \\), where \\( C^B \\cap C^N = \\emptyset \\). In this section, we introduce CORA, a framework that adapts CLIP for the OVD task by Region prompting and Anchor prematching. For fairer comparisons with existing methods, we also experiment with a broader setting where extra data is available, which is referred to as \\( \\mathrm{CORA}^+ \\), and will be introduced along with the experiments." + }, + { + "type": "page_number", + "bbox": [ + 0.485, + 0.947, + 0.514, + 0.957 + ], + "angle": 0, + "content": "7032" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.111, + 0.1, + 0.861, + 0.436 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.458, + 0.895, + 0.514 + ], + "angle": 0, + "content": "Figure 1. Overview of our method. The image is encoded into a feature map by the CLIP image encoder for both localization and classification. The regional feature is extracted by pooling the feature map, and then prompted before classified by the CLIP class name embeddings. The anchor boxes are pre-matched and conditioned on a class before decoding. During training, per-class post-matching is conducted. During inference, the box predictions are classified by the region classifier." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.538, + 0.188, + 0.552 + ], + "angle": 0, + "content": "3.1. Overview" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.563, + 0.47, + 0.734 + ], + "angle": 0, + "content": "The overall framework of CORA is illustrated in Fig. 1. Given an image as input, we acquire the spatial feature map using the ResNet backbone from the pre-trained CLIP image encoder, which is shared by both region classification and object localization branches. Unlike conventional detectors, localization and classification are decoupled and sequentially conducted in our framework to better fit the characteristic of the OVD problem. We train a DETR-style object localizer that refines a set of object queries together with their associated anchor boxes to localize the objects, which are then classified by a region classifier adapted from CLIP." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.739, + 0.469, + 0.84 + ], + "angle": 0, + "content": "Region Classification. Given a region to be classified (anchor box or box prediction), we adopt RoIAlign to obtain the region feature, followed by the attention pooling module of CLIP to generate region embeddings, which can be classified by class embeddings obtained from the CLIP text encoder, as done in CLIP. We name this module as CLIP-based region classifier (Fig. 1-(a))." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.843, + 0.47, + 0.903 + ], + "angle": 0, + "content": "Object Localization. The visual feature map is firstly refined by the DETR-like encoder, and then fed into the DETR-like decoder. The queries of anchor boxes are firstly classified by the CLIP-based region classifier, which are" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.538, + 0.892, + 0.653 + ], + "angle": 0, + "content": "then conditioned on their predicted labels before iteratively refined by the DETR-like decoder for better localization. The decoder also estimates the matchability of the query with the previously predicted label. During training, the predicted boxes are one-to-one matched with ground truth boxes that has the same labels (Fig. 1-(b)), and trained as in DETR. In the inference stage, the class labels of the boxes are adjusted by the CLIP-based region classifier." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.654, + 0.893, + 0.84 + ], + "angle": 0, + "content": "As mentioned in Sec. 1, there are two obstacles to be addressed: (1) object detection conducts recognition on image regions, while the CLIP model is trained on whole-image input, leading to a distribution gap that hinders classification performance. (2) the detector needs to learn object localization for novel classes, while we only have annotations on a limited number of base classes. To solve the first obstacle, we propose region prompting to modulate the region features for better generalizable region embeddings, which will be introduced in Sec. 3.2. To solve the second obstacle, we put forward anchor pre-matching to encourage class-aware object localization that can generalize to novel classes during inference, which will be introduced in Sec. 3.3." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.849, + 0.68, + 0.866 + ], + "angle": 0, + "content": "3.2. Region Prompting" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.892, + 0.902 + ], + "angle": 0, + "content": "OVD requires the detector to classify image regions into a given category list. In this section, we will elaborate how" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7033" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.083, + 0.096, + 0.462, + 0.168 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.201, + 0.178, + 0.347, + 0.192 + ], + "angle": 0, + "content": "(a) Whole-image encoding" + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.204, + 0.462, + 0.293 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.301, + 0.334, + 0.314 + ], + "angle": 0, + "content": "(b) Regional encoding" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.325, + 0.471, + 0.38 + ], + "angle": 0, + "content": "Figure 2. Comparison between the CLIP-based region classifier and the vanilla pipeline. We pool regional feature from the feature map, rather than cropping the region patch and classify them as a separate image." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.404, + 0.468, + 0.49 + ], + "angle": 0, + "content": "a pre-trained CLIP model is adapted to formulate the CLIP-based region classifier. Given the CLIP model, region classification can be realized by comparing the similarity between the regional embedding from the CLIP image encoder and the class name embedding from the CLIP text encoder." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.491, + 0.468, + 0.692 + ], + "angle": 0, + "content": "Region Prompting. As illustrated in Fig. 2, given an image and a set of region of interest (RoI), we firstly encode the whole image into a feature map by the CLIP encoder's first 3 blocks, which are then pooled by RoIAlign [32] either according to anchor boxes or predicted boxes into region features, before encoded by the last block of the CLIP image encoder backbone. There exists a distribution gap between the CLIP image encoder's whole-image feature map and the pooled regional features. We propose region prompting to fix the misalignment by augmenting the region feature with learnable prompts \\( p \\in \\mathbb{R}^{S \\times S \\times C} \\), where \\( S \\) is the spatial size of the regional feature, and \\( C \\) is the dimension of the regional feature. Specifically, given the input regional feature \\( f_{\\mathrm{region}} \\), the region prompting is conducted as" + }, + { + "type": "equation", + "bbox": [ + 0.183, + 0.703, + 0.469, + 0.72 + ], + "angle": 0, + "content": "\\[\nv _ {\\text {p r o m p t}} = P \\left(f _ {\\text {r e g i o n}} \\oplus p\\right), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.728, + 0.468, + 0.756 + ], + "angle": 0, + "content": "where \\(\\oplus\\) denotes element-wise addition, \\(P\\) is the attention pooling module of the CLIP visual encoder." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.757, + 0.468, + 0.884 + ], + "angle": 0, + "content": "Optimizing Region Prompts. We train the region prompts on a detection dataset with base-class annotations. The class name embeddings are pre-computed by the CLIP text encoder, which are used as classifier weights later. We train the prompts by a standard cross-entropy loss to classify the ground truth boxes with their pooled regional features \\( f_{\\mathrm{region}} \\). When optimizing the region prompts, we keep other model weights frozen, and only make the region prompts to be learned." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.886, + 0.468, + 0.901 + ], + "angle": 0, + "content": "Comparison with existing methods. The previous com" + }, + { + "type": "image", + "bbox": [ + 0.511, + 0.091, + 0.868, + 0.168 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.51, + 0.174, + 0.867, + 0.252 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.526, + 0.261, + 0.867, + 0.337 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.63, + 0.342, + 0.765, + 0.354 + ], + "angle": 0, + "content": "(a) Conditional matching" + }, + { + "type": "image", + "bbox": [ + 0.523, + 0.359, + 0.87, + 0.461 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.63, + 0.465, + 0.764, + 0.478 + ], + "angle": 0, + "content": "(b) Anchor pre-matching" + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.489, + 0.892, + 0.556 + ], + "angle": 0, + "content": "Figure 3. Comparison between anchor pre-matching and conditional matching. Anchor pre-matching decodes a constant number of object queries, and assigns different numbers of object queries based on the image content, avoiding repetitive decoding as in conditional matching." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.582, + 0.892, + 0.697 + ], + "angle": 0, + "content": "mon practice for region classification with CLIP is to crop the RoIs, and encode them as separate images, before comparing with text embeddings. This pipeline is not efficient when encoding regions with overlaps, since the overlapping regions are encoded more than once in different region crops. Its accuracy also suffers from the missing context information. In contrast, our regional prompting is more efficient and preserves richer context." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.697, + 0.892, + 0.811 + ], + "angle": 0, + "content": "The region prompts contain less than 1M parameters, which is in line with recent advances of the prompt tuning and adapter literatures. Region prompting generalizes well to unseen novel classes. We attribute the generalization capability to the fact that region prompting directly fix the distribution mismatch right after where it occurs (after region pooling), whereas the existing methods tweak irrelevant parameters to compensate for the distribution mismatch." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.82, + 0.706, + 0.837 + ], + "angle": 0, + "content": "3.3. Anchor Pre-Matching" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.843, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Region prompting helps solve the region classification problem. Object localization is the other critical subtask of object detection. Considering the inferior performance of pre-trained RPN on novel classes, we introduce" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7034" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.177 + ], + "angle": 0, + "content": "a class-aware query-based object localizer, which demonstrates better generalization capability on unseen classes. As shown in Fig. 1, Given the visual feature map from the frozen CLIP image encoder, the object queries are prematched to the class name embeddings by the CLIP text encoder." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.178, + 0.471, + 0.307 + ], + "angle": 0, + "content": "Anchor Pre-Matching. The object localizer is implemented by a DETR-style transformer encoder-decoder structure, where the encoder refines the feature map, and the decoder decodes a set of object queries into box predictions. We adopt DAB-DETR [26], where each object query is associated with an anchor box. Each ground truth box is pre-matched to a set of queries with the same label. The label \\(\\hat{c}_i\\) of an object query is assigned by classifying the associated anchor box \\(b_i\\)" + }, + { + "type": "equation", + "bbox": [ + 0.177, + 0.32, + 0.47, + 0.345 + ], + "angle": 0, + "content": "\\[\n\\hat {c} _ {i} = \\underset {c \\in C ^ {B}} {\\arg \\max } \\operatorname {c o s i n e} \\left(v _ {i}, l _ {c}\\right), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.357, + 0.469, + 0.43 + ], + "angle": 0, + "content": "where \\( v_{i} \\) is the region feature of anchor box \\( b_{i}, l_{c} \\) is the class name embedding of class \\( c \\), and cosine denotes the cosine similarity. After pre-matching, each object query conditions on the predicted class embedding to allow class-aware box regression. The conditioned object query is given by" + }, + { + "type": "equation", + "bbox": [ + 0.218, + 0.441, + 0.469, + 0.457 + ], + "angle": 0, + "content": "\\[\nq _ {i} = \\mathrm {M L P} \\left(l _ {\\mathrm {c} _ {\\mathrm {i}}}\\right). \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.469, + 0.469, + 0.542 + ], + "angle": 0, + "content": "The DETR-like decoder iteratively refines each object query with its associated anchor box \\((q_i, b_i)\\) into \\(\\hat{y}_i = (\\hat{p}_i, \\hat{b}_i)\\), where \\(\\hat{b}_i\\) is the refined box coordinates and \\(\\hat{p}_i\\) is the matching probability to the query's pre-matched class \\(\\hat{c}_i\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.544, + 0.469, + 0.63 + ], + "angle": 0, + "content": "Given the model predictions, the assignment between ground truth boxes and model predictions is conducted by performing bipartite matching for each class separately. We only allow each ground truth box to be assigned to the prediction with the same pre-matched label, in order to enforce the decoder to be aware of the conditioned text embedding." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.63, + 0.469, + 0.702 + ], + "angle": 0, + "content": "Specifically, for class \\(c\\), given the \\(N^c\\) box predictions \\(\\hat{y}^c = \\{\\hat{y}_i \\mid \\hat{c}_i = c\\}\\) that are pre-matched to class \\(c\\), and the set of ground truth boxes \\(y^c\\) in class \\(c\\), we optimize a permutation of \\(N^c\\) elements \\(\\sigma \\in \\mathfrak{S}_{\\mathrm{N}^c}\\) that minimizes the following cost" + }, + { + "type": "equation", + "bbox": [ + 0.156, + 0.715, + 0.469, + 0.755 + ], + "angle": 0, + "content": "\\[\n\\hat {\\sigma} _ {c} = \\underset {\\sigma \\in \\mathfrak {S} _ {\\mathrm {N} ^ {\\mathrm {c}}}} {\\arg \\max } \\sum_ {i} ^ {N ^ {c}} \\mathcal {L} _ {\\operatorname {c o s t}} \\left(y _ {i} ^ {c}, \\hat {y} _ {\\sigma (\\mathrm {i})} ^ {c}\\right), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.768, + 0.327, + 0.782 + ], + "angle": 0, + "content": "where the matching cost is defined as" + }, + { + "type": "equation", + "bbox": [ + 0.137, + 0.795, + 0.469, + 0.813 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {c o s t}} (y, \\hat {y}) = \\mathcal {L} _ {\\text {m a t c h}} (p, \\hat {p}) + \\mathcal {L} _ {\\text {b o x}} (b, \\hat {b}). \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.826, + 0.469, + 0.901 + ], + "angle": 0, + "content": "\\(\\mathcal{L}_{\\mathrm{match}}(p,\\hat{p})\\) is a binary classification loss, and \\(\\mathcal{L}_{\\mathrm{box}}(b,\\hat{b})\\) characterizes the localization error of \\(\\hat{b}\\) w.r.t \\(b\\). In our case, we implement \\(\\mathcal{L}_{\\mathrm{match}}\\) by the focal loss [23]. \\(\\mathcal{L}_{\\mathrm{box}}\\) is implemented by a weighted sum of \\(L_{1}\\) loss and GIoU [37] loss following prior works." + }, + { + "type": "text", + "bbox": [ + 0.52, + 0.092, + 0.822, + 0.107 + ], + "angle": 0, + "content": "The model is optimized by the following loss" + }, + { + "type": "equation", + "bbox": [ + 0.548, + 0.115, + 0.891, + 0.167 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} = \\sum_ {c \\in C ^ {B}} \\mathcal {L} _ {\\text {m a t c h}} \\left(p ^ {c}, \\hat {p} _ {\\hat {\\sigma} _ {c}} ^ {c}\\right) + \\mathcal {L} _ {\\text {b o x}} \\left(b ^ {c}, \\hat {b} _ {\\hat {\\sigma} _ {c}} ^ {c}\\right) \\tag {6} \\\\ = \\lambda_ {\\text {f o c a l}} \\mathcal {L} _ {\\text {f o c a l}} + \\lambda_ {\\mathrm {L} _ {1}} \\mathcal {L} _ {\\mathrm {L} _ {1}} + \\lambda_ {\\mathrm {G I o U}} \\mathcal {L} _ {\\mathrm {G I o U}}. \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.175, + 0.892, + 0.235 + ], + "angle": 0, + "content": "During inference, we adopt the region classifier introduced in Sec. 3.2 to classify the predicted boxes \\(\\{\\hat{b}_i\\}\\) for better classification accuracy. The class score is multiplied by the pre-matching score to account for the box quality" + }, + { + "type": "equation", + "bbox": [ + 0.598, + 0.244, + 0.891, + 0.261 + ], + "angle": 0, + "content": "\\[\n\\mathrm {P} \\left(\\hat {b} _ {i} \\in c\\right) = \\hat {p} _ {i} \\operatorname {c o s i n e} \\left(\\hat {v} _ {i}, l _ {c}\\right). \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.269, + 0.892, + 0.456 + ], + "angle": 0, + "content": "Comparison with Conditional Matching [35]. Conditional Matching in OV-DETR [35] also proposes to condition the queries on the text embedding for class-aware regression. But it suffers from repetitive per-class inference. Specifically, as shown in Fig. 3, each class in \\( C^N \\) needs to be separately localized with the same group of query, which means both computation and memory consumption scales linearly with the vocabulary size. During training, the number of negative classes sampled in each iteration is limited due to the memory constraint, which hinders convergence. During inference, repetitive per-class decoding is required and results in low inference efficiency, especially when the vocabulary size is large." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.457, + 0.892, + 0.557 + ], + "angle": 0, + "content": "Contrary to conditional matching, our anchor pre-matching mechanism assigns anchor boxes for different classes adaptively according to the image content, which ensures a constant number of query decoupled from the category size. By anchor pre-matching, all the classes can be decoded together in one pass, eliminating the need for repetitive per-class decoding." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.557, + 0.892, + 0.628 + ], + "angle": 0, + "content": "To improve the generalization capability and training convergence of open-vocabulary detectors equipped with Anchor Pre-Matching, we also introduce two effective training techniques, namely, \"Drop Class\" and \"CLIP-Aligned Labeling\"." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.628, + 0.892, + 0.871 + ], + "angle": 0, + "content": "Class Dropout. The generalization capability of the model can be further boosted by randomly dropping categories during training. Since our goal is to train a detector that can detect objects from a user specified category list, training on a fixed list of categories leads to bias. We mitigate this bias by randomly dropping out the base categories during training. For efficiency reason, we implement this idea by splitting the base classes into two complementary groups and train on both of them, instead of training on one group while dropping the other group. Specifically, in each training iteration, we split \\( C^B \\) and the ground truth boxes at a probability of \\( p \\), and train the detector on the same image with two complementary sets of categories. It enforces the model to condition its prediction on the query categories. Since a ground truth box in one group does not appear in the other group, the model needs to be aware of the categories to treat differently for the two sets of annotations." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.872, + 0.891, + 0.901 + ], + "angle": 0, + "content": "CLIP-Aligned Labeling. Directly training the localizer on the original COCO dataset suffers from convergence issue." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.515, + 0.957 + ], + "angle": 0, + "content": "7035" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.278 + ], + "angle": 0, + "content": "By anchor pre-matching mechanism, a ground truth box incorporates training only when at least one query with the same pre-matched label exists. Otherwise, it is ignored, which hinders convergence. This issue can be partially attributed to the inaccurate anchor box. However, even if a ground truth box has an accurate anchor box, it may still be ignored due to the limited recognition accuracy limitation of the region classifier, or in other words, the ground truth box label is not aligned with the CLIP region classifier used for pre-matching. Thus, we relabel the boxes in the training dataset with the region classifier, which we refer to as CLIP-Aligned labeling. With this technique, more ground truth boxes can be matched." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.294, + 0.21, + 0.31 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.319, + 0.471, + 0.435 + ], + "angle": 0, + "content": "In this section, we comprehensively evaluate our CORA on the open-vocabulary detection task. Datasets and evaluation protocols are introduced in Sec. 4.1, and implementation details of our method are provided in Sec. 4.2. We compare with state-of-the-art methods in Sec. 4.3, demonstrating advantages of our CORA, and then validate the effectiveness of the proposed Region Prompting and Anchor Pre-Matching in Sec. 4.4 and Sec. 4.5, respectively." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.445, + 0.371, + 0.461 + ], + "angle": 0, + "content": "4.1. Dataset & Training & Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.469, + 0.47, + 0.669 + ], + "angle": 0, + "content": "Following the convention of the COCO OVD benchmark proposed in [4], the 80 classes in the COCO dataset [24] classes are divided into 48 base classes and 17 novel classes. The model is trained on the 48 base classes, which contains 107,761 images and 665,387 instances. The model is then evaluated on the validation set of novel classes, which contains 4,836 images and 33,152 instances from both the 48 base classes and the 17 novel classes. We also conduct experiments on the LVIS v1.0 [16] dataset. On LVIS dataset, the model is trained on 461 common classes and 405 frequent classes, which contains 100,170 images and 1,264,883 instances. After training, the model is evaluated on the LVIS validation set, which contains 19,809 images and 244,707 instances." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.671, + 0.47, + 0.857 + ], + "angle": 0, + "content": "To make fairer comparisons with other methods, we propose \\(\\mathrm{CORA^{+}}\\), which utilizes extra dataset or target novel class names. \\(\\mathrm{CORA^{+}}\\) is a detector trained with both ground-truth base-category annotations and additional pseudo bounding box labels. When target class names are provided, we generate pseudo boxes of novel classes on the base training dataset using CORA, and when extra imagedataset (text could be captions or class names) is available, pseudo boxes of all objects mentioned by the corresponding text are generated. We use standard detector architecture and training target to train \\(\\mathrm{CORA^{+}}\\). SAMDETR [38] is used for the experiment on COCO and CenterNet2 [43] is used for the experiment on LVIS." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.858, + 0.47, + 0.901 + ], + "angle": 0, + "content": "In the OVD task, we evaluate our model under the \"generalized\" setting, in which the model needs to predict objects from both base and novel classes, and then evaluated" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.236 + ], + "angle": 0, + "content": "on novel objects. On the COCO benchmark, we take AP50 as our evaluation metric, which counts the average precision at an intersection over union (IoU) of \\(50\\%\\) for each class, and then averages among all the classes. For the LVIS OVD benchmark, we evaluate on the full validation dataset, and report the mean AP of boxes from the novel classes to compare with prior works [40]. For the region classification task, we let the model classify the ground truth boxes in the COCO dataset, the performance is evaluated in terms of mAP." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.244, + 0.719, + 0.26 + ], + "angle": 0, + "content": "4.2. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.266, + 0.892, + 0.553 + ], + "angle": 0, + "content": "Model Specifications. We use DAB-DETR [26] as the object localizer. Specifically, localizer is configured to have 1,000 object queries, 3 encoder layers and 6 decoder layers. We use a multi-layer perceptron (MLP) with 128 hidden neurons to transform class name embeddings into an object query. Following CLIP [30], each class embedding is computed as the average text embedding of the class name over 80 context prompts. In class dropout, each class is randomly assigned to one of the two groups with equal probabilities. When adopting region prompting on LVIS, classes in the common and frequent group are sampled with equal weights, meaning that the objects in the less frequent classes are over sampled. When training our method on LVIS, we sample 100 categories (including the ground truth categories) in each iteration. Since the number of classes in the LVIS dataset is much larger than that of COCO, we relax the matching constraint in anchor pre-matching, such that ground truth boxes can be post-matched to the anchor boxes with a similar label. Specifically, classes with a cosine similarity greater than 0.7 are considered similar." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.554, + 0.892, + 0.77 + ], + "angle": 0, + "content": "Training & Hyperparameters. We train the region prompts for 5 epochs with a base learning rate of \\(10^{-4}\\), which decays after the \\(4^{th}\\) epoch by a factor of 0.1. The localizer is trained for 35 epochs with a learning rate of \\(10^{-4}\\) without learning rate decay. Both the region prompts and the localizer are trained with batch size 32 by the AdamW optimizer [28] with \\(10^{-4}\\) weight decay. We apply gradient clipping with a maximal norm of 0.1. To stabilize training, we evaluate on the exponential moving average (EMA) of the model after training. The class dropout probability \\(p\\) is set as 0.2. \\(\\lambda_{\\mathrm{focal}}\\), \\(\\lambda_{\\mathrm{L_1}}\\) and \\(\\lambda_{\\mathrm{GIOU}}\\) are set as 2.0, 5.0, 2.0, respectively. For the experiment on LVIS, we use repeat factor sampling [16] with default hyperparameters to balance the training samples. We use non-maximum suppression (NMS) with an IoU threshold 0.5 during inference." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.778, + 0.868, + 0.793 + ], + "angle": 0, + "content": "4.3. Comparison with State-of-the-Art Methods" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.8, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Tab. 1 summarizes our main results. Since the pretrained model is crucial to the open-vocabulary capability of the detector, we compare our method with baseline methods that are trained with the same CLIP model. When compared with methods trained on CLIP RN50, CORA outperforms VL-PLM by 0.7 AP50 on novel classes. With a larger pre-trained model, our method improves the previous state" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "7036" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.123, + 0.089, + 0.85, + 0.303 + ], + "angle": 0, + "content": "
MethodExtra DatasetDetector Training \nPre-train ModelRequire Novel ClassGeneralized (17 + 48)
NovelBaseAll
OVR-CNN [36]COCO Captions [9]-X22.846.039.9
Detic [42]COCO Captions [9]CLIP (text encoder)X27.847.145.0
RegionCLIP [40]CC3M [33]CLIP (RN50)X31.457.150.4
VL-PLM [39]-CLIP (RN50)✓34.460.253.5
CORA (Ours)-CLIP (RN50)X35.135.535.4
ViLD [14]-CLIP (ViT-B/32)✓27.659.951.3
OV-DETR [35]-CLIP (ViT-B/32)✓29.461.052.7
MEDet [7]COCO Captions [9]CLIP (ViT-B/32)X32.654.049.4
RegionCLIP [40]CC3M [33]CLIP (RN50x4)X39.361.655.7
CORA (Ours)-CLIP (RN50x4)X41.744.543.8
CORA+ (Ours)COCO Captions [9]CLIP (RN50x4)X43.160.956.2
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.312, + 0.895, + 0.355 + ], + "angle": 0, + "content": "Table 1. Main results on the COCO OVD benchmark. We report AP50 as the evaluation metric. The baseline methods are grouped by their pre-trained model. We also list the extra dataset requirement of each method, and whether they require the novel class to be provided during training." + }, + { + "type": "table", + "bbox": [ + 0.086, + 0.377, + 0.462, + 0.513 + ], + "angle": 0, + "content": "
MethodDetector TrainingLVIS APr
Ext. DataPre-train Model
ViLD [14]-CLIP (ViT-B/32)16.3
OV-DETR [35]-CLIP (ViT-B/32)17.4
RegionCLIP [40]CC3MCLIP (RN50x4)22.0
CORA (Ours)-CLIP (RN50x4)22.2
MEDet [7]CC3MCLIP (ViT-B/32)22.4
Detic [42]IN-21kCLIP (text encoder)26.2
CORA+ (Ours)IN-21kCLIP (text encoder)28.1
" + }, + { + "type": "table_caption", + "bbox": [ + 0.118, + 0.522, + 0.428, + 0.536 + ], + "angle": 0, + "content": "Table 2. Results on the LVIS [16] OVD benchmark." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.567, + 0.47, + 0.625 + ], + "angle": 0, + "content": "of-the-art RegionCLIP [40] by 2.4 AP50. When extra data is available, the performance can be further boosted to 43.1 AP50. The results on LVIS [16] OVD benchmark is shown in Tab. 2." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.626, + 0.47, + 0.842 + ], + "angle": 0, + "content": "Note that among the baseline methods, VL-PLM [39], ViLD [14] and OV-DETR [35] use novel class names during training in order to recognize the potential novel objects and assign pseudo labels for them. Consequently, a new detector needs to be trained whenever there is a new set of categories to be detected. CORA can generalize to any combination of novel categories once trained without tedious re-training. Other compared methods relying on CLIP require image tag annotations extracted from language descriptions [7,31,40] or image labels [42] during training. We argue that the extra annotations do not provide additional information over CLIP. Instead, they serve as a media to transfer the knowledge from CLIP to the detector. Our method directly adapts the CLIP model to obtain region classifier, thus no extra image-text data is needed." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.843, + 0.47, + 0.903 + ], + "angle": 0, + "content": "In this work, both region prompting and anchor prematching aim to generalize the knowledge learned from base classes to novel classes. Consequently, the performance gap between novel and base classes is significantly" + }, + { + "type": "table", + "bbox": [ + 0.517, + 0.377, + 0.878, + 0.501 + ], + "angle": 0, + "content": "
MethodCLIP modelNovelBase
CLIPRN5058.258.6
CLIP-Adapter [12]RN5063.080.6
CoOp [41]RN5064.475.7
CORARN5065.170.0
CLIPRN50x463.962.7
CORARN50x474.176.0
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.51, + 0.892, + 0.551 + ], + "angle": 0, + "content": "Table 3. Results on the region classification task evaluated in mAP. We compare our method with the original CLIP regional classifier and other baseline methods." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.578, + 0.892, + 0.638 + ], + "angle": 0, + "content": "lowered than the compared methods. Note that in OVD, the performance is evaluated by the generalization capability towards novel classes, rather than the bases classes on which they are trained." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.647, + 0.803, + 0.664 + ], + "angle": 0, + "content": "4.4. Effectiveness of Region Prompting" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.67, + 0.892, + 0.771 + ], + "angle": 0, + "content": "Region Prompting. Since object classification is decoupled from localization in this work, the CLIP-based region classifier can be directly evaluated by the region classification task. After trained on the base-class annotations of COCO dataset, we evaluate the CLIP-based region classifier on both base and novel classes in the validation set. We use mean average precision (mAP) as our evaluation metric." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.772, + 0.892, + 0.902 + ], + "angle": 0, + "content": "Tab. 3 shows our main result. Directly evaluating on the CLIP model without further training already achieves considerable performance of \\(58.2\\mathrm{mAP}\\) on the novel classes. We compare our result with two competitive methods from the adapter and prompt tuning literatures. CLIP-Adapter [12] adopts an additional bottleneck layer to learn new features and performs residual style feature blending with the original pre-trained features. CoOp [41] prepends shared learnable prompts to the text embeddings before en" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7037" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.115, + 0.089, + 0.434, + 0.149 + ], + "angle": 0, + "content": "
CLIP modelFeatureNovelBase
RN50whole-image43.840.5
RN50regional58.658.2
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.158, + 0.47, + 0.198 + ], + "angle": 0, + "content": "Table 4. Comparison of the whole-image classification pipeline and the regional classification pipeline. Results are reported in mAP." + }, + { + "type": "table", + "bbox": [ + 0.086, + 0.214, + 0.462, + 0.312 + ], + "angle": 0, + "content": "
Query ConditioningPer-Class Post-MatchingCLIP-Aligned LabelingNovel
XX✓26.0
✓X✓29.9
✓✓X40.9
✓✓✓41.7
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.322, + 0.468, + 0.363 + ], + "angle": 0, + "content": "Table 5. Ablation studies on anchor pre-matching and CLIP-aligned labeling. Anchor pre-matching consists of query conditioning and per-class post-matching." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.39, + 0.468, + 0.563 + ], + "angle": 0, + "content": "coded by the CLIP text encoder. Experiments show that the compared methods bias strongly towards base classes. Note that the baseline methods adapts the CLIP model by tuning the text input or output feature, which are irrelevant or distant to the regional feature, where the mismatch between regional and whole-image feature occurs. On the contrary, region prompting directly prompts the mismatched features, thus generalizes better to the novel classes, achieving a performance gain of \\(6.9\\mathrm{mAP}\\) over CLIP on the novel classes. Region prompting also scales with larger backbones. On the \\(\\mathrm{RN50x4}\\) CLIP backbone, region prompting further boosts CLIP by \\(10.2\\%\\) mAP over the corresponding CLIP model." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.564, + 0.468, + 0.693 + ], + "angle": 0, + "content": "Comparison with whole-image classification. A common practice of region classification by CLIP is to crop the regions and classify them as separate images. We compare the region classifier with the common practice on the original CLIP weights without region prompting. As shown in Tab. 4, when classifying a region as a whole-image, the performance is significantly lower than using the regional feature, despite the extra computation. We attribute the performance gap to the missing context of cropped images." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.705, + 0.406, + 0.721 + ], + "angle": 0, + "content": "4.5. Effectiveness of Anchor Pre-Matching" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.728, + 0.468, + 0.756 + ], + "angle": 0, + "content": "We conduct ablation studies to validate anchor pre-matching and the proposed training techniques." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.757, + 0.468, + 0.901 + ], + "angle": 0, + "content": "Anchor Pre-Matching. Anchor pre-matching consists of two separate operations: query conditioning and per-class post-matching. We analyze their effects in Tab. 5. Firstly, we train a model without anchor pre-matching. We keep using DAB-DETR [26] as the localizer, and use the same number of object queries and anchor boxes for fair comparison. The object queries are not pre-matched to classes, and are initialized as 0 as done in DAB-DETR. Since the model predictions are not pre-matched, the per-class post-matching is replaced by the vanilla one-to-one matching" + }, + { + "type": "image", + "bbox": [ + 0.528, + 0.113, + 0.858, + 0.295 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.547, + 0.305, + 0.843, + 0.319 + ], + "angle": 0, + "content": "Figure 4. Different choices of \\( p \\) for class dropout." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.346, + 0.892, + 0.431 + ], + "angle": 0, + "content": "mechanism in DETR. Then, we classify the anchor boxes and condition the object queries on the class name embeddings, but do not set constraint on the post-matching, which boosts the performance on novel classes by 3.9 AP50. After adopting the full anchor pre-matching, the performance is significantly boosted by 10.8 AP50." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.432, + 0.892, + 0.533 + ], + "angle": 0, + "content": "Training Techniques. Class dropout and CLIP-aligned labeling are two training techniques that help the model generalize better. In Fig. 4, we examine the effect of different dropout probability. We find that the models trained with class dropout consistently outperform the baseline, and \\( p = 0.2 \\) gives the best performance. In Tab. 5, we validate the effectiveness of CLIP-Aligned Labeling." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.546, + 0.619, + 0.561 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.571, + 0.892, + 0.786 + ], + "angle": 0, + "content": "The core challenge in open-vocabulary detection is how to effectively transfer the knowledge learned from the base classes to the unseen novel classes for evaluation. In this work, we directly adapt the CLIP into a region classifier, and mitigate the distribution gap between whole-image feature and regional feature through region prompting, which successfully generalizes to the novel categories. Different from prior works that rely on a fixed RPN for novel class localization, we achieve efficient class-aware localization by the proposed anchor pre-matching mechanism. Experiments show that our method can better transfer the knowledge from base classes to unseen novel classes with a smaller gap than prior works. We hope our work can help other researchers gain better insight on the OVD problem and develop better open-vocabulary detectors." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.786, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Acknowledgement. This project is funded in part by National Key R&D Program of China Project 2022ZD0161100, by the Centre for Perceptual and Interactive Intelligence (CPII) Ltd under the Innovation and Technology Commission (ITC)'s InnoHK, by General Research Fund of Hong Kong RGC Project 14204021. Hongsheng Li is a PI of CPII under the InnoHK. This project is also supported by SenseTime Collaborative Research Grant." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "7038" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.105 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.115, + 0.423, + 0.127 + ], + "angle": 0, + "content": "[1] Deformable detr: Deformable transformers. 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.129, + 0.47, + 0.168 + ], + "angle": 0, + "content": "[2] Hyojin Bahng, Ali Jahanian, Swami Sankaranarayanan, and Phillip Isola. Exploring visual prompts for adapting large-scale models. arXiv preprint arXiv:2203.17274, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.17, + 0.469, + 0.209 + ], + "angle": 0, + "content": "[3] Hyojin Bahng, Ali Jahanian, Swami Sankaranarayanan, and Phillip Isola. Visual prompting: Modifying pixel space to adapt pre-trained models. ArXiv, abs/2203.17274, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.21, + 0.469, + 0.248 + ], + "angle": 0, + "content": "[4] Ankan Bansal, Karan Sikka, Gaurav Sharma, Rama Chellappa, and Ajay Divakaran. Zero-shot object detection. ArXiv, abs/1804.04340, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.25, + 0.469, + 0.38 + ], + "angle": 0, + "content": "[5] Tom B. Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert-Voss, Gretchen Krueger, T. J. Henighan, Rewon Child, Aditya Ramesh, Daniel M. Ziegler, Jeff Wu, Clemens Winter, Christopher Hesse, Mark Chen, Eric Sigler, Mateusz Litwin, Scott Gray, Benjamin Chess, Jack Clark, Christopher Berner, Sam McCandlish, Alec Radford, Ilya Sutskever, and Dario Amodei. Language models are few-shot learners. ArXiv, abs/2005.14165, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.382, + 0.469, + 0.434 + ], + "angle": 0, + "content": "[6] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. ArXiv, abs/2005.12872, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.436, + 0.469, + 0.488 + ], + "angle": 0, + "content": "[7] Peixian Chen, Kekai Sheng, Mengdan Zhang, Yunhang Shen, Ke Li, and Chunhua Shen. Open vocabulary object detection with proposal mining and prediction equalization. CoRR, abs/2206.11134, 2022. 1, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.489, + 0.469, + 0.539 + ], + "angle": 0, + "content": "[8] Qiang Chen, Xiaokang Chen, Gang Zeng, and Jingdong Wang. Group detr: Fast training convergence with decoupled one-to-many label assignment. ArXiv, abs/2207.13085, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.542, + 0.469, + 0.594 + ], + "angle": 0, + "content": "[9] Xinlei Chen, Hao Fang, Tsung-Yi Lin, Ramakrishna Vedantam, Saurabh Gupta, Piotr Dólar, and C. Lawrence Zitnick. Microsoft coco captions: Data collection and evaluation server. ArXiv, abs/1504.00325, 2015. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.595, + 0.469, + 0.647 + ], + "angle": 0, + "content": "[10] Chengjian Feng, Yujie Zhong, Zequn Jie, Xiangxiang Chu, Haibing Ren, Xiaolin Wei, Weidi Xie, and Lin Ma. Promptdet: Towards open-vocabulary detection using uncurated images. 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.649, + 0.469, + 0.7 + ], + "angle": 0, + "content": "[11] Mingfei Gao, Chen Xing, Juan Carlos Niebles, Junnan Li, Ran Xu, Wenhao Liu, and Caiming Xiong. Open vocabulary object detection with pseudo bounding-box labels. arXiv preprint arXiv:2111.09452, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.702, + 0.469, + 0.753 + ], + "angle": 0, + "content": "[12] Peng Gao, Shijie Geng, Renrui Zhang, Teli Ma, Rongyao Fang, Yongfeng Zhang, Hongsheng Li, and Yu Jiao Qiao. Clip-adapter: Better vision-language models with feature adapters. ArXiv, abs/2110.04544, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.755, + 0.469, + 0.819 + ], + "angle": 0, + "content": "[13] Peng Gao, Minghang Zheng, Xiaogang Wang, Jifeng Dai, and Hongsheng Li. Fast convergence of detr with spatially modulated co-attention. 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 3601–3610, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.821, + 0.469, + 0.86 + ], + "angle": 0, + "content": "[14] Xiuye Gu, Tsung-Yi Lin, Weicheng Kuo, and Yin Cui. Open-vocabulary object detection via vision and language knowledge distillation. In ICLR, 2022. 1, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.862, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[15] Xiuye Gu, Tsung-Yi Lin, Weicheng Kuo, and Yin Cui. Open-vocabulary object detection via vision and language knowledge distillation. In ICLR, 2022. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.47, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.146 + ], + "angle": 0, + "content": "[16] Agrim Gupta, Piotr Dollar, and Ross Girshick. LVIS: A dataset for large vocabulary instance segmentation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2019. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.148, + 0.892, + 0.187 + ], + "angle": 0, + "content": "[17] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In 2017 IEEE International Conference on Computer Vision (ICCV), pages 2980-2988, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.188, + 0.892, + 0.239 + ], + "angle": 0, + "content": "[18] Ding Jia, Yuhui Yuan, Haodi He, Xiao pei Wu, Haojun Yu, Weihong Lin, Lei huan Sun, Chao Zhang, and Hanhua Hu. Detrs with hybrid matching. ArXiv, abs/2207.13080, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.242, + 0.892, + 0.281 + ], + "angle": 0, + "content": "[19] Menglin Jia, Luming Tang, Bor-Chun Chen, Claire Cardie, Serge J. Belongie, Bharath Hariharan, and Ser Nam Lim. Visual prompt tuning. ArXiv, abs/2203.12119, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.283, + 0.892, + 0.322 + ], + "angle": 0, + "content": "[20] Chen Ju, Tengda Han, Kunhao Zheng, Ya Zhang, and Weidi Xie. Prompting visual-language models for efficient video understanding. ArXiv, abs/2112.04478, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.324, + 0.892, + 0.403 + ], + "angle": 0, + "content": "[21] Liunian Harold Li, Pengchuan Zhang, Haotian Zhang, Jianwei Yang, Chunyuan Li, Yiwu Zhong, Lijuan Wang, Lu Yuan, Lei Zhang, Jenq-Neng Hwang, Kai-Wei Chang, and Jianfeng Gao. Grounded language-image pre-training. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10955–10965, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.404, + 0.892, + 0.482 + ], + "angle": 0, + "content": "[22] Xiang Lisa Li and Percy Liang. Prefix-tuning: Optimizing continuous prompts for generation. Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), abs/2101.00190, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.484, + 0.892, + 0.536 + ], + "angle": 0, + "content": "[23] Tsung-Yi Lin, Priya Goyal, Ross B. Girshick, Kaiming He, and Piotr Dólar. Focal loss for dense object detection. 2017 IEEE International Conference on Computer Vision (ICCV), pages 2999-3007, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.538, + 0.892, + 0.59 + ], + "angle": 0, + "content": "[24] Tsung-Yi Lin, Michael Maire, Serge J. Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C. Lawrence Zitnick. Microsoft coco: Common objects in context. In ECCV, 2014. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.592, + 0.892, + 0.644 + ], + "angle": 0, + "content": "[25] Ziyi Lin, Shijie Geng, Renrui Zhang, Peng Gao, Gerard de Melo, Xiaogang Wang, Jifeng Dai, Yu Qiao, and Hongsheng Li. Frozen clip models are efficient video learners. arXiv preprint arXiv:2208.03550, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.646, + 0.892, + 0.697 + ], + "angle": 0, + "content": "[26] Shilong Liu, Feng Li, Hao Zhang, Xiao Bin Yang, Xianbiao Qi, Hang Su, Jun Zhu, and Lei Zhang. Dab-detr: Dynamic anchor boxes are better queries for detr. ArXiv, abs/2201.12329, 2022. 2, 5, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.7, + 0.892, + 0.751 + ], + "angle": 0, + "content": "[27] Xiao Liu, Kaixuan Ji, Yicheng Fu, Zhengxiao Du, Zhilin Yang, and Jie Tang. P-tuning v2: Prompt tuning can be comparable to fine-tuning universally across scales and tasks. CoRR, abs/2110.07602, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.754, + 0.892, + 0.779 + ], + "angle": 0, + "content": "[28] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In ICLR, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.781, + 0.892, + 0.847 + ], + "angle": 0, + "content": "[29] Muhammad Maaz, Hanoona Rasheed, Salman Khan, Fahad Shahbaz Khan, Rao Muhammad Anwer, and Ming-Hsuan Yang. Class-agnostic object detection with multimodal transformer. In 17th European Conference on Computer Vision (ECCV). Springer, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.849, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[30] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "7039" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.112, + 0.092, + 0.468, + 0.119 + ], + "angle": 0, + "content": "models from natural language supervision. In ICML, 2021. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.121, + 0.469, + 0.185 + ], + "angle": 0, + "content": "[31] Hanoona Rasheed, Muhammad Maaz, Muhammad Uzair Khattak, Salman Khan, and Fahad Shahbaz Khan. Bridging the gap between object and image-level representations for open-vocabulary detection. ArXiv, abs/2207.03482, 2022. 1, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.188, + 0.469, + 0.24 + ], + "angle": 0, + "content": "[32] Shaoqing Ren, Kaiming He, Ross B. Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. IEEE Transactions on Pattern Analysis and Machine Intelligence, 39:1137-1149, 2015. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.242, + 0.469, + 0.293 + ], + "angle": 0, + "content": "[33] Piyush Sharma, Nan Ding, Sebastian Goodman, and Radu Soricut. Conceptual captions: A cleaned, hypernymed, image alt-text dataset for automatic image captioning. In ACL, 2018. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.296, + 0.469, + 0.347 + ], + "angle": 0, + "content": "[34] Yinghui Xing, Qirui Wu, De Cheng, Shizhou Zhang, Guoqiang Liang, and Yanning Zhang. Class-aware visual prompt tuning for vision-language pre-trained model. CoRR, abs/2208.08340, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.349, + 0.469, + 0.389 + ], + "angle": 0, + "content": "[35] Yuhang Zang, Wei Li, Kaiyang Zhou, Chen Huang, and Chen Change Loy. Open-vocabulary detr with conditional matching. ArXiv, abs/2203.11876, 2022. 1, 2, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.391, + 0.469, + 0.455 + ], + "angle": 0, + "content": "[36] Alireza Zareian, Kevin Dela Rosa, Derek Hao Hu, and Shih-Fu Chang. Open-vocabulary object detection using captions. In 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 14388-14397, 2021. 1, 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.457, + 0.469, + 0.522 + ], + "angle": 0, + "content": "[37] Hongyu Zhai, Jian Cheng, and Mengyong Wang. Rethink the iou-based loss functions for bounding box regression. 2020 IEEE 9th Joint International Information Technology and Artificial Intelligence Conference (ITAIC), 9:1522-1528, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.524, + 0.469, + 0.59 + ], + "angle": 0, + "content": "[38] Gongjie Zhang, Zhipeng Luo, Yingchen Yu, Kaiwen Cui, and Shijian Lu. Accelerating DETR convergence via semantic-aligned matching. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 949-958, 2022. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.591, + 0.469, + 0.656 + ], + "angle": 0, + "content": "[39] Shiyu Zhao, Zhixing Zhang, Samuel Schulter, Long Zhao, Vijay Kumar BG, Anastasis Stathopoulos, Manmohan Chandraker, and Dimitris N. Metaxas. Exploiting unlabeled data with vision and language models for object detection. CoRR, abs/2207.08954, 2022. 1, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.658, + 0.469, + 0.736 + ], + "angle": 0, + "content": "[40] Yiwu Zhong, Jianwei Yang, Pengchuan Zhang, Chengkun Li, Noel C. F. Codella, Liunian Harold Li, Luowei Zhou, Xiyang Dai, Lu Yuan, Yin Li, and Jianfeng Gao. Regionclip: Region-based language-image pretraining. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 16772-16782, 2022. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.738, + 0.469, + 0.777 + ], + "angle": 0, + "content": "[41] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Learning to prompt for vision-language models. Int. J. Comput. Vis., 130:2337-2348, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.779, + 0.469, + 0.818 + ], + "angle": 0, + "content": "[42] Xingyi Zhou, Rohit Girdhar, Armand Joulin, Philipp Krahenbihl, and Ishan Misra. Detecting twenty-thousand classes using image-level supervision. In ECCV, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.819, + 0.469, + 0.859 + ], + "angle": 0, + "content": "[43] Xingyi Zhou, Vladlen Koltun, and Philipp Krahenbuhl. Probabilistic two-stage detection. In arXiv preprint arXiv:2103.07461, 2021. 6" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.469, + 0.859 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.957 + ], + "angle": 0, + "content": "7040" + } + ] +] \ No newline at end of file diff --git a/2023/CORA_ Adapting CLIP for Open-Vocabulary Detection With Region Prompting and Anchor Pre-Matching/e257818b-1aea-4bd1-9288-7c6a81195cd8_origin.pdf b/2023/CORA_ Adapting CLIP for Open-Vocabulary Detection With Region Prompting and Anchor Pre-Matching/e257818b-1aea-4bd1-9288-7c6a81195cd8_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..be135170ebac46fa930aab1d98efde814a2322f4 --- /dev/null +++ b/2023/CORA_ Adapting CLIP for Open-Vocabulary Detection With Region Prompting and Anchor Pre-Matching/e257818b-1aea-4bd1-9288-7c6a81195cd8_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ac721693dda688e69def226de99a610657138f78008b608e5256923c23410cc4 +size 712057 diff --git a/2023/CORA_ Adapting CLIP for Open-Vocabulary Detection With Region Prompting and Anchor Pre-Matching/full.md b/2023/CORA_ Adapting CLIP for Open-Vocabulary Detection With Region Prompting and Anchor Pre-Matching/full.md new file mode 100644 index 0000000000000000000000000000000000000000..82d7ed0f06167fcd9c1fa59711b0dd959181a90e --- /dev/null +++ b/2023/CORA_ Adapting CLIP for Open-Vocabulary Detection With Region Prompting and Anchor Pre-Matching/full.md @@ -0,0 +1,295 @@ +# CORA: Adapting CLIP for Open-Vocabulary Detection with Region Prompting and Anchor Pre-Matching + +Xiaoshi Wu $^{1}$ , Feng Zhu $^{2}$ , Rui Zhao $^{2,3}$ , Hongsheng Li $^{1,4}$ + +1Multimedia Laboratory, The Chinese University of Hong Kong + +$^{2}$ SenseTime Research $^{3}$ Qing Yuan Research Institute, Shanghai Jiao Tong University + +$^{4}$ Centre for Perceptual and Interactive Intelligence (CPII) + +{wuxiaoshi@link, hsli@ee}.cuhk.edu.hk,{zhufeng, zhaorui}@sentetime.com + +# Abstract + +Open-vocabulary detection (OVD) is an object detection task aiming at detecting objects from novel categories beyond the base categories on which the detector is trained. Recent OVD methods rely on large-scale visual-language pre-trained models, such as CLIP, for recognizing novel objects. We identify the two core obstacles that need to be tackled when incorporating these models into detector training: (1) the distribution mismatch that happens when applying a VL-model trained on whole images to region recognition tasks; (2) the difficulty of localizing objects of unseen classes. To overcome these obstacles, we propose CORA, a DETR-style framework that adapts CLIP for Open-vocabulary detection by Region prompting and Anchor pre-matching. Region prompting mitigates the whole-to-region distribution gap by prompting the region features of the CLIP-based region classifier. Anchor pre-matching helps learning generalizable object localization by a class-aware matching mechanism. We evaluate CORA on the COCO OVD benchmark, where we achieve 41.7 AP50 on novel classes, which outperforms the previous SOTA by 2.4 AP50 even without resorting to extra training data. When extra training data is available, we train $\mathrm{CORA^{+}}$ on both ground-truth base-category annotations and additional pseudo bounding box labels computed by CORA. $\mathrm{CORA^{+}}$ achieves 43.1 AP50 on the COCO OVD benchmark and 28.1 box APr on the LVIS OVD benchmark. The code is available at https://github.com/tgxs002/CORA. + +# 1. Introduction + +Object detection is a fundamental vision problem that involves localizing and classifying objects from images. Classical object detection requires detecting objects from a closed set of categories. Extra annotations and training are required if objects of unseen categories need to be detected. It has attracted much attention on detecting novel categories + +without tedious annotations, or even detect object from new category, which is currently referred as open-vocabulary detection (OVD) [36]. + +Recent advances on large-scale vision-language pretrained models, such as CLIP [30], enable new solutions for tackling OVD. CLIP learns a joint embedding space of images and text from a large-scale image-text dataset, which shows remarkable capability on visual recognition tasks. The general idea of applying CLIP for OVD is to treat it as an open-vocabulary classifier. However, there are two obstacles hindering the effective use of CLIP on tackling OVD. + +How to adapt CLIP for region-level tasks? One trivial solution is to crop regions and treat them as separate images, which has been adopted by multiple recent works [7, 14, 31, 35]. But the distribution gap between region crops and full images leads to inferior classification accuracy. MEDet [7] mitigates this issue by augmenting the text feature with image features. However, it requires extra image-text pairs to prevent overfitting to so-called "base" classes that are seen during training. RegionCLIP [40] directly acquires regional features by RoIAlign [17], which is more efficient but cannot generalize well to novel classes without finetuning. The finetuning is costly when adopting a larger CLIP model. + +How to learn generalizable object proposals? ViLD [14], OV-DETR [35], Object-Centric-OVD [31], Region-CLIP [40] need RPN or class-agnostic object detectors [29] to mine potential novel class objects. However, these RPNs are strongly biased towards the base classes on which they are trained, while perform poorly on the novel classes. MEDet [7] and VL-PLM [39] identify this problem and adopt several handcrafted policies to rule out or merge low-quality boxes, but the performance is still bounded by the frozen RPN. OV-DETR [35] learns generalizable object localization by conditioning box regression on class name embeddings, but at the cost of efficiency issue induced by repetitive per-class inference. + +In this work, we propose a new framework based on DETection TR transformers (DETR) [6] that incorporates CLIP + +into detector training to achieve open-vocabulary detection without additional image-text data. Specifically, we use a DETR-style object localizer for class-aware object localization, and the predicted boxes are encoded by pooling the intermediate feature map of the CLIP image encoder, which are classified by the CLIP text encoder with class names. However, there is a distribution gap between whole-image features from CLIP's original visual encoder and the newly pooled region features, leading to an inferior classification accuracy. Thus, we propose Region Prompting to adapt the CLIP image encoder, which boosts the classification performance, and also demonstrates better generalization capability than existing methods. We adopt DAB-DETR [26] as the localizer, in which object queries are associated with dynamic anchor boxes. By pre-matching the dynamic anchor boxes with the input categories before box regression (Anchor Pre-Matching), class-aware regression can be achieved without the cost of repetitive per-class inference. + +We validate our method on COCO [24] and LVIS v1.0 [16] OVD benchmarks. On the COCO OVD benchmark, our method improves AP50 of novel categories over the previous best method [40] by 2.4 AP50 without training on extra data, and achieves consistent gain on CLIP models of different scales. When compared under a fairer setting with extra training data, our method significantly outperforms the existing methods by 3.8 AP50 on novel categories and achieves comparable performance on base categories. On the LVIS OVD benchmark, our method achieves 22.2/28.1 APr with/w.o. extra data, which significantly outperforms existing methods that are also trained with/w.o. extra data. By applying region prompting on the base classes of COCO, the classification performance on the novel classes is boosted from $63.9\%$ to $74.1\%$ , whereas other prompting or adaptation methods easily bias towards the base classes. + +The contributions of this work are summarized as follows: (1) Our proposed region prompting effectively mitigates the gap between whole image features and region features, and generalize well in the open-vocabulary setting. (2) Anchor Pre-Matching enables DETR for generalizable object localization efficiently. (3) We achieve state-of-the-art performance on COCO and LVIS OVD benchmarks. + +# 2. Related Works + +Open-Vocabulary Object Detection OVR-CNN [36] firstly put forth this new formulation of detection, and proposes its baseline solution by aligning region features with nouns in captions that are paired with the image. Mingfeng [11] et al. mine pseudo labels by utilizing the localization ability of pre-trained vision-language models. PromptDet [10] addresses the gap between image and region classification by adding learnable prompts when encoding the class names, namely regional prompt learning (RPL), which is expected to generalize from base to novel categories. OV-DETR [35] is the first DETR-style open + +vocabulary detector, which proposes conditional matching to solve the missing novel class problem in assignment, but at the cost of inefficient inference. RegionCLIP [40] propose a second-stage pre-training mechanism to adapt the CLIP model to encode region features, and demonstrates its capability on OVD and zero-shot transfer setting. GLIP [21] jointly learns object localization and VL alignment. Matthias et al. [15] proposes to finetune a VL aligned model for detection, while we fix the pre-trained VL model for better generalization towards novel categories. + +Detection Transformers DETR [6] is an object detection architecture based on transformers that formulates object detection as a set-to-set matching problem, which greatly simplifies the pipeline. Several works address the slow convergence problem of DETR by architectural improvement [1, 13,26,38] or special training strategies [8,18]. Zhu et al. [1] proposes multi-scale deformable attention module to efficiently aggregate information from multi-scale feature maps. Gao et al. [13] proposes to modulate the cross-attention in the transformer decoder by anchor box coordinates to accelerate the detector convergence. DABDETR [26] formulates the queries in DETR architecture as anchor boxes, which accelerates detector training. Chen et al. [38] proposes Group DETR, which adds auxiliary object queries during training to take advantage of one-to-many matching for faster convergence. + +Prompt Tuning Prompting is originated from NLP, and it refers to prepending task instructions before the input sequence to give the language model the hint about the task [5]. Later works [22, 27] explores tuning continuous prompt vectors when few-shot data is available. VPT [19], Visual Prompting [2, 3] explore prompting in the pixel space. [20] and [25] prompts pre-trained model for video recognition tasks. [34] proposes class-aware visual prompt tuning to generalize the learned prompts to unseen categories. Recent works demonstrate that prompt tuning is an effective and parameter-efficient way to adapt large-scale pre-trained models to downstream tasks. + +# 3. Method + +Open-vocabulary detection (OVD) is an object detection task aiming at detecting objects from novel categories beyond the base categories on which the detector is trained. Formally, the detector is trained on a detection dataset with base-category $C^B$ box annotations, and tested on a new input image $I \in \mathbb{R}^{H \times W \times 3}$ to detect objects belonging to a novel category set $C^N$ , where $C^B \cap C^N = \emptyset$ . In this section, we introduce CORA, a framework that adapts CLIP for the OVD task by Region prompting and Anchor prematching. For fairer comparisons with existing methods, we also experiment with a broader setting where extra data is available, which is referred to as $\mathrm{CORA}^+$ , and will be introduced along with the experiments. + +![](images/3a099067d201f95e92c3386783275c33f2e4c8b2a56b18f3527f8edca9ed3c16.jpg) +Figure 1. Overview of our method. The image is encoded into a feature map by the CLIP image encoder for both localization and classification. The regional feature is extracted by pooling the feature map, and then prompted before classified by the CLIP class name embeddings. The anchor boxes are pre-matched and conditioned on a class before decoding. During training, per-class post-matching is conducted. During inference, the box predictions are classified by the region classifier. + +# 3.1. Overview + +The overall framework of CORA is illustrated in Fig. 1. Given an image as input, we acquire the spatial feature map using the ResNet backbone from the pre-trained CLIP image encoder, which is shared by both region classification and object localization branches. Unlike conventional detectors, localization and classification are decoupled and sequentially conducted in our framework to better fit the characteristic of the OVD problem. We train a DETR-style object localizer that refines a set of object queries together with their associated anchor boxes to localize the objects, which are then classified by a region classifier adapted from CLIP. + +Region Classification. Given a region to be classified (anchor box or box prediction), we adopt RoIAlign to obtain the region feature, followed by the attention pooling module of CLIP to generate region embeddings, which can be classified by class embeddings obtained from the CLIP text encoder, as done in CLIP. We name this module as CLIP-based region classifier (Fig. 1-(a)). + +Object Localization. The visual feature map is firstly refined by the DETR-like encoder, and then fed into the DETR-like decoder. The queries of anchor boxes are firstly classified by the CLIP-based region classifier, which are + +then conditioned on their predicted labels before iteratively refined by the DETR-like decoder for better localization. The decoder also estimates the matchability of the query with the previously predicted label. During training, the predicted boxes are one-to-one matched with ground truth boxes that has the same labels (Fig. 1-(b)), and trained as in DETR. In the inference stage, the class labels of the boxes are adjusted by the CLIP-based region classifier. + +As mentioned in Sec. 1, there are two obstacles to be addressed: (1) object detection conducts recognition on image regions, while the CLIP model is trained on whole-image input, leading to a distribution gap that hinders classification performance. (2) the detector needs to learn object localization for novel classes, while we only have annotations on a limited number of base classes. To solve the first obstacle, we propose region prompting to modulate the region features for better generalizable region embeddings, which will be introduced in Sec. 3.2. To solve the second obstacle, we put forward anchor pre-matching to encourage class-aware object localization that can generalize to novel classes during inference, which will be introduced in Sec. 3.3. + +# 3.2. Region Prompting + +OVD requires the detector to classify image regions into a given category list. In this section, we will elaborate how + +![](images/d4b96808f0036c773207a589026eb6d4894e7a822810cea81d943805e2e19cff.jpg) +(a) Whole-image encoding + +![](images/093ce500705802d0b9bcbb337745ddd9a613030762f101636c7d7f5bd7cc1949.jpg) +(b) Regional encoding + +a pre-trained CLIP model is adapted to formulate the CLIP-based region classifier. Given the CLIP model, region classification can be realized by comparing the similarity between the regional embedding from the CLIP image encoder and the class name embedding from the CLIP text encoder. + +Region Prompting. As illustrated in Fig. 2, given an image and a set of region of interest (RoI), we firstly encode the whole image into a feature map by the CLIP encoder's first 3 blocks, which are then pooled by RoIAlign [32] either according to anchor boxes or predicted boxes into region features, before encoded by the last block of the CLIP image encoder backbone. There exists a distribution gap between the CLIP image encoder's whole-image feature map and the pooled regional features. We propose region prompting to fix the misalignment by augmenting the region feature with learnable prompts $p \in \mathbb{R}^{S \times S \times C}$ , where $S$ is the spatial size of the regional feature, and $C$ is the dimension of the regional feature. Specifically, given the input regional feature $f_{\mathrm{region}}$ , the region prompting is conducted as + +$$ +v _ {\text {p r o m p t}} = P \left(f _ {\text {r e g i o n}} \oplus p\right), \tag {1} +$$ + +where $\oplus$ denotes element-wise addition, $P$ is the attention pooling module of the CLIP visual encoder. + +Optimizing Region Prompts. We train the region prompts on a detection dataset with base-class annotations. The class name embeddings are pre-computed by the CLIP text encoder, which are used as classifier weights later. We train the prompts by a standard cross-entropy loss to classify the ground truth boxes with their pooled regional features $f_{\mathrm{region}}$ . When optimizing the region prompts, we keep other model weights frozen, and only make the region prompts to be learned. + +Comparison with existing methods. The previous com + +![](images/5f22df4d1ddf32c8fd272c91cd51d2a1f68a26cae1494f8fb0add14df39bedca.jpg) + +![](images/ea644c72a4a7a810463b136d9cc3252acc7d45a5b9b07513a6397e57585bfe83.jpg) + +![](images/c79b27b5ef350e5f411263794dd0c7246d457b7f0fd6fb47c5cf91687b8f533b.jpg) +(a) Conditional matching + +![](images/8518f2ee0e91295835e023b3060fabc7fff020afa782f935077b24e536c1bee2.jpg) +Figure 2. Comparison between the CLIP-based region classifier and the vanilla pipeline. We pool regional feature from the feature map, rather than cropping the region patch and classify them as a separate image. +(b) Anchor pre-matching +Figure 3. Comparison between anchor pre-matching and conditional matching. Anchor pre-matching decodes a constant number of object queries, and assigns different numbers of object queries based on the image content, avoiding repetitive decoding as in conditional matching. + +mon practice for region classification with CLIP is to crop the RoIs, and encode them as separate images, before comparing with text embeddings. This pipeline is not efficient when encoding regions with overlaps, since the overlapping regions are encoded more than once in different region crops. Its accuracy also suffers from the missing context information. In contrast, our regional prompting is more efficient and preserves richer context. + +The region prompts contain less than 1M parameters, which is in line with recent advances of the prompt tuning and adapter literatures. Region prompting generalizes well to unseen novel classes. We attribute the generalization capability to the fact that region prompting directly fix the distribution mismatch right after where it occurs (after region pooling), whereas the existing methods tweak irrelevant parameters to compensate for the distribution mismatch. + +# 3.3. Anchor Pre-Matching + +Region prompting helps solve the region classification problem. Object localization is the other critical subtask of object detection. Considering the inferior performance of pre-trained RPN on novel classes, we introduce + +a class-aware query-based object localizer, which demonstrates better generalization capability on unseen classes. As shown in Fig. 1, Given the visual feature map from the frozen CLIP image encoder, the object queries are prematched to the class name embeddings by the CLIP text encoder. + +Anchor Pre-Matching. The object localizer is implemented by a DETR-style transformer encoder-decoder structure, where the encoder refines the feature map, and the decoder decodes a set of object queries into box predictions. We adopt DAB-DETR [26], where each object query is associated with an anchor box. Each ground truth box is pre-matched to a set of queries with the same label. The label $\hat{c}_i$ of an object query is assigned by classifying the associated anchor box $b_i$ + +$$ +\hat {c} _ {i} = \underset {c \in C ^ {B}} {\arg \max } \operatorname {c o s i n e} \left(v _ {i}, l _ {c}\right), \tag {2} +$$ + +where $v_{i}$ is the region feature of anchor box $b_{i}, l_{c}$ is the class name embedding of class $c$ , and cosine denotes the cosine similarity. After pre-matching, each object query conditions on the predicted class embedding to allow class-aware box regression. The conditioned object query is given by + +$$ +q _ {i} = \mathrm {M L P} \left(l _ {\mathrm {c} _ {\mathrm {i}}}\right). \tag {3} +$$ + +The DETR-like decoder iteratively refines each object query with its associated anchor box $(q_i, b_i)$ into $\hat{y}_i = (\hat{p}_i, \hat{b}_i)$ , where $\hat{b}_i$ is the refined box coordinates and $\hat{p}_i$ is the matching probability to the query's pre-matched class $\hat{c}_i$ . + +Given the model predictions, the assignment between ground truth boxes and model predictions is conducted by performing bipartite matching for each class separately. We only allow each ground truth box to be assigned to the prediction with the same pre-matched label, in order to enforce the decoder to be aware of the conditioned text embedding. + +Specifically, for class $c$ , given the $N^c$ box predictions $\hat{y}^c = \{\hat{y}_i \mid \hat{c}_i = c\}$ that are pre-matched to class $c$ , and the set of ground truth boxes $y^c$ in class $c$ , we optimize a permutation of $N^c$ elements $\sigma \in \mathfrak{S}_{\mathrm{N}^c}$ that minimizes the following cost + +$$ +\hat {\sigma} _ {c} = \underset {\sigma \in \mathfrak {S} _ {\mathrm {N} ^ {\mathrm {c}}}} {\arg \max } \sum_ {i} ^ {N ^ {c}} \mathcal {L} _ {\operatorname {c o s t}} \left(y _ {i} ^ {c}, \hat {y} _ {\sigma (\mathrm {i})} ^ {c}\right), \tag {4} +$$ + +where the matching cost is defined as + +$$ +\mathcal {L} _ {\text {c o s t}} (y, \hat {y}) = \mathcal {L} _ {\text {m a t c h}} (p, \hat {p}) + \mathcal {L} _ {\text {b o x}} (b, \hat {b}). \tag {5} +$$ + +$\mathcal{L}_{\mathrm{match}}(p,\hat{p})$ is a binary classification loss, and $\mathcal{L}_{\mathrm{box}}(b,\hat{b})$ characterizes the localization error of $\hat{b}$ w.r.t $b$ . In our case, we implement $\mathcal{L}_{\mathrm{match}}$ by the focal loss [23]. $\mathcal{L}_{\mathrm{box}}$ is implemented by a weighted sum of $L_{1}$ loss and GIoU [37] loss following prior works. + +The model is optimized by the following loss + +$$ +\begin{array}{l} \mathcal {L} = \sum_ {c \in C ^ {B}} \mathcal {L} _ {\text {m a t c h}} \left(p ^ {c}, \hat {p} _ {\hat {\sigma} _ {c}} ^ {c}\right) + \mathcal {L} _ {\text {b o x}} \left(b ^ {c}, \hat {b} _ {\hat {\sigma} _ {c}} ^ {c}\right) \tag {6} \\ = \lambda_ {\text {f o c a l}} \mathcal {L} _ {\text {f o c a l}} + \lambda_ {\mathrm {L} _ {1}} \mathcal {L} _ {\mathrm {L} _ {1}} + \lambda_ {\mathrm {G I o U}} \mathcal {L} _ {\mathrm {G I o U}}. \\ \end{array} +$$ + +During inference, we adopt the region classifier introduced in Sec. 3.2 to classify the predicted boxes $\{\hat{b}_i\}$ for better classification accuracy. The class score is multiplied by the pre-matching score to account for the box quality + +$$ +\mathrm {P} \left(\hat {b} _ {i} \in c\right) = \hat {p} _ {i} \operatorname {c o s i n e} \left(\hat {v} _ {i}, l _ {c}\right). \tag {7} +$$ + +Comparison with Conditional Matching [35]. Conditional Matching in OV-DETR [35] also proposes to condition the queries on the text embedding for class-aware regression. But it suffers from repetitive per-class inference. Specifically, as shown in Fig. 3, each class in $C^N$ needs to be separately localized with the same group of query, which means both computation and memory consumption scales linearly with the vocabulary size. During training, the number of negative classes sampled in each iteration is limited due to the memory constraint, which hinders convergence. During inference, repetitive per-class decoding is required and results in low inference efficiency, especially when the vocabulary size is large. + +Contrary to conditional matching, our anchor pre-matching mechanism assigns anchor boxes for different classes adaptively according to the image content, which ensures a constant number of query decoupled from the category size. By anchor pre-matching, all the classes can be decoded together in one pass, eliminating the need for repetitive per-class decoding. + +To improve the generalization capability and training convergence of open-vocabulary detectors equipped with Anchor Pre-Matching, we also introduce two effective training techniques, namely, "Drop Class" and "CLIP-Aligned Labeling". + +Class Dropout. The generalization capability of the model can be further boosted by randomly dropping categories during training. Since our goal is to train a detector that can detect objects from a user specified category list, training on a fixed list of categories leads to bias. We mitigate this bias by randomly dropping out the base categories during training. For efficiency reason, we implement this idea by splitting the base classes into two complementary groups and train on both of them, instead of training on one group while dropping the other group. Specifically, in each training iteration, we split $C^B$ and the ground truth boxes at a probability of $p$ , and train the detector on the same image with two complementary sets of categories. It enforces the model to condition its prediction on the query categories. Since a ground truth box in one group does not appear in the other group, the model needs to be aware of the categories to treat differently for the two sets of annotations. + +CLIP-Aligned Labeling. Directly training the localizer on the original COCO dataset suffers from convergence issue. + +By anchor pre-matching mechanism, a ground truth box incorporates training only when at least one query with the same pre-matched label exists. Otherwise, it is ignored, which hinders convergence. This issue can be partially attributed to the inaccurate anchor box. However, even if a ground truth box has an accurate anchor box, it may still be ignored due to the limited recognition accuracy limitation of the region classifier, or in other words, the ground truth box label is not aligned with the CLIP region classifier used for pre-matching. Thus, we relabel the boxes in the training dataset with the region classifier, which we refer to as CLIP-Aligned labeling. With this technique, more ground truth boxes can be matched. + +# 4. Experiments + +In this section, we comprehensively evaluate our CORA on the open-vocabulary detection task. Datasets and evaluation protocols are introduced in Sec. 4.1, and implementation details of our method are provided in Sec. 4.2. We compare with state-of-the-art methods in Sec. 4.3, demonstrating advantages of our CORA, and then validate the effectiveness of the proposed Region Prompting and Anchor Pre-Matching in Sec. 4.4 and Sec. 4.5, respectively. + +# 4.1. Dataset & Training & Evaluation + +Following the convention of the COCO OVD benchmark proposed in [4], the 80 classes in the COCO dataset [24] classes are divided into 48 base classes and 17 novel classes. The model is trained on the 48 base classes, which contains 107,761 images and 665,387 instances. The model is then evaluated on the validation set of novel classes, which contains 4,836 images and 33,152 instances from both the 48 base classes and the 17 novel classes. We also conduct experiments on the LVIS v1.0 [16] dataset. On LVIS dataset, the model is trained on 461 common classes and 405 frequent classes, which contains 100,170 images and 1,264,883 instances. After training, the model is evaluated on the LVIS validation set, which contains 19,809 images and 244,707 instances. + +To make fairer comparisons with other methods, we propose $\mathrm{CORA^{+}}$ , which utilizes extra dataset or target novel class names. $\mathrm{CORA^{+}}$ is a detector trained with both ground-truth base-category annotations and additional pseudo bounding box labels. When target class names are provided, we generate pseudo boxes of novel classes on the base training dataset using CORA, and when extra imagedataset (text could be captions or class names) is available, pseudo boxes of all objects mentioned by the corresponding text are generated. We use standard detector architecture and training target to train $\mathrm{CORA^{+}}$ . SAMDETR [38] is used for the experiment on COCO and CenterNet2 [43] is used for the experiment on LVIS. + +In the OVD task, we evaluate our model under the "generalized" setting, in which the model needs to predict objects from both base and novel classes, and then evaluated + +on novel objects. On the COCO benchmark, we take AP50 as our evaluation metric, which counts the average precision at an intersection over union (IoU) of $50\%$ for each class, and then averages among all the classes. For the LVIS OVD benchmark, we evaluate on the full validation dataset, and report the mean AP of boxes from the novel classes to compare with prior works [40]. For the region classification task, we let the model classify the ground truth boxes in the COCO dataset, the performance is evaluated in terms of mAP. + +# 4.2. Implementation Details + +Model Specifications. We use DAB-DETR [26] as the object localizer. Specifically, localizer is configured to have 1,000 object queries, 3 encoder layers and 6 decoder layers. We use a multi-layer perceptron (MLP) with 128 hidden neurons to transform class name embeddings into an object query. Following CLIP [30], each class embedding is computed as the average text embedding of the class name over 80 context prompts. In class dropout, each class is randomly assigned to one of the two groups with equal probabilities. When adopting region prompting on LVIS, classes in the common and frequent group are sampled with equal weights, meaning that the objects in the less frequent classes are over sampled. When training our method on LVIS, we sample 100 categories (including the ground truth categories) in each iteration. Since the number of classes in the LVIS dataset is much larger than that of COCO, we relax the matching constraint in anchor pre-matching, such that ground truth boxes can be post-matched to the anchor boxes with a similar label. Specifically, classes with a cosine similarity greater than 0.7 are considered similar. + +Training & Hyperparameters. We train the region prompts for 5 epochs with a base learning rate of $10^{-4}$ , which decays after the $4^{th}$ epoch by a factor of 0.1. The localizer is trained for 35 epochs with a learning rate of $10^{-4}$ without learning rate decay. Both the region prompts and the localizer are trained with batch size 32 by the AdamW optimizer [28] with $10^{-4}$ weight decay. We apply gradient clipping with a maximal norm of 0.1. To stabilize training, we evaluate on the exponential moving average (EMA) of the model after training. The class dropout probability $p$ is set as 0.2. $\lambda_{\mathrm{focal}}$ , $\lambda_{\mathrm{L_1}}$ and $\lambda_{\mathrm{GIOU}}$ are set as 2.0, 5.0, 2.0, respectively. For the experiment on LVIS, we use repeat factor sampling [16] with default hyperparameters to balance the training samples. We use non-maximum suppression (NMS) with an IoU threshold 0.5 during inference. + +# 4.3. Comparison with State-of-the-Art Methods + +Tab. 1 summarizes our main results. Since the pretrained model is crucial to the open-vocabulary capability of the detector, we compare our method with baseline methods that are trained with the same CLIP model. When compared with methods trained on CLIP RN50, CORA outperforms VL-PLM by 0.7 AP50 on novel classes. With a larger pre-trained model, our method improves the previous state + +
MethodExtra DatasetDetector Training +Pre-train ModelRequire Novel ClassGeneralized (17 + 48)
NovelBaseAll
OVR-CNN [36]COCO Captions [9]-X22.846.039.9
Detic [42]COCO Captions [9]CLIP (text encoder)X27.847.145.0
RegionCLIP [40]CC3M [33]CLIP (RN50)X31.457.150.4
VL-PLM [39]-CLIP (RN50)✓34.460.253.5
CORA (Ours)-CLIP (RN50)X35.135.535.4
ViLD [14]-CLIP (ViT-B/32)✓27.659.951.3
OV-DETR [35]-CLIP (ViT-B/32)✓29.461.052.7
MEDet [7]COCO Captions [9]CLIP (ViT-B/32)X32.654.049.4
RegionCLIP [40]CC3M [33]CLIP (RN50x4)X39.361.655.7
CORA (Ours)-CLIP (RN50x4)X41.744.543.8
CORA+ (Ours)COCO Captions [9]CLIP (RN50x4)X43.160.956.2
+ +Table 1. Main results on the COCO OVD benchmark. We report AP50 as the evaluation metric. The baseline methods are grouped by their pre-trained model. We also list the extra dataset requirement of each method, and whether they require the novel class to be provided during training. + +
MethodDetector TrainingLVIS APr
Ext. DataPre-train Model
ViLD [14]-CLIP (ViT-B/32)16.3
OV-DETR [35]-CLIP (ViT-B/32)17.4
RegionCLIP [40]CC3MCLIP (RN50x4)22.0
CORA (Ours)-CLIP (RN50x4)22.2
MEDet [7]CC3MCLIP (ViT-B/32)22.4
Detic [42]IN-21kCLIP (text encoder)26.2
CORA+ (Ours)IN-21kCLIP (text encoder)28.1
+ +of-the-art RegionCLIP [40] by 2.4 AP50. When extra data is available, the performance can be further boosted to 43.1 AP50. The results on LVIS [16] OVD benchmark is shown in Tab. 2. + +Note that among the baseline methods, VL-PLM [39], ViLD [14] and OV-DETR [35] use novel class names during training in order to recognize the potential novel objects and assign pseudo labels for them. Consequently, a new detector needs to be trained whenever there is a new set of categories to be detected. CORA can generalize to any combination of novel categories once trained without tedious re-training. Other compared methods relying on CLIP require image tag annotations extracted from language descriptions [7,31,40] or image labels [42] during training. We argue that the extra annotations do not provide additional information over CLIP. Instead, they serve as a media to transfer the knowledge from CLIP to the detector. Our method directly adapts the CLIP model to obtain region classifier, thus no extra image-text data is needed. + +In this work, both region prompting and anchor prematching aim to generalize the knowledge learned from base classes to novel classes. Consequently, the performance gap between novel and base classes is significantly + +Table 2. Results on the LVIS [16] OVD benchmark. + +
MethodCLIP modelNovelBase
CLIPRN5058.258.6
CLIP-Adapter [12]RN5063.080.6
CoOp [41]RN5064.475.7
CORARN5065.170.0
CLIPRN50x463.962.7
CORARN50x474.176.0
+ +Table 3. Results on the region classification task evaluated in mAP. We compare our method with the original CLIP regional classifier and other baseline methods. + +lowered than the compared methods. Note that in OVD, the performance is evaluated by the generalization capability towards novel classes, rather than the bases classes on which they are trained. + +# 4.4. Effectiveness of Region Prompting + +Region Prompting. Since object classification is decoupled from localization in this work, the CLIP-based region classifier can be directly evaluated by the region classification task. After trained on the base-class annotations of COCO dataset, we evaluate the CLIP-based region classifier on both base and novel classes in the validation set. We use mean average precision (mAP) as our evaluation metric. + +Tab. 3 shows our main result. Directly evaluating on the CLIP model without further training already achieves considerable performance of $58.2\mathrm{mAP}$ on the novel classes. We compare our result with two competitive methods from the adapter and prompt tuning literatures. CLIP-Adapter [12] adopts an additional bottleneck layer to learn new features and performs residual style feature blending with the original pre-trained features. CoOp [41] prepends shared learnable prompts to the text embeddings before en + +
CLIP modelFeatureNovelBase
RN50whole-image43.840.5
RN50regional58.658.2
+ +Table 4. Comparison of the whole-image classification pipeline and the regional classification pipeline. Results are reported in mAP. + +
Query ConditioningPer-Class Post-MatchingCLIP-Aligned LabelingNovel
XX✓26.0
✓X✓29.9
✓✓X40.9
✓✓✓41.7
+ +Table 5. Ablation studies on anchor pre-matching and CLIP-aligned labeling. Anchor pre-matching consists of query conditioning and per-class post-matching. + +coded by the CLIP text encoder. Experiments show that the compared methods bias strongly towards base classes. Note that the baseline methods adapts the CLIP model by tuning the text input or output feature, which are irrelevant or distant to the regional feature, where the mismatch between regional and whole-image feature occurs. On the contrary, region prompting directly prompts the mismatched features, thus generalizes better to the novel classes, achieving a performance gain of $6.9\mathrm{mAP}$ over CLIP on the novel classes. Region prompting also scales with larger backbones. On the $\mathrm{RN50x4}$ CLIP backbone, region prompting further boosts CLIP by $10.2\%$ mAP over the corresponding CLIP model. + +Comparison with whole-image classification. A common practice of region classification by CLIP is to crop the regions and classify them as separate images. We compare the region classifier with the common practice on the original CLIP weights without region prompting. As shown in Tab. 4, when classifying a region as a whole-image, the performance is significantly lower than using the regional feature, despite the extra computation. We attribute the performance gap to the missing context of cropped images. + +# 4.5. Effectiveness of Anchor Pre-Matching + +We conduct ablation studies to validate anchor pre-matching and the proposed training techniques. + +Anchor Pre-Matching. Anchor pre-matching consists of two separate operations: query conditioning and per-class post-matching. We analyze their effects in Tab. 5. Firstly, we train a model without anchor pre-matching. We keep using DAB-DETR [26] as the localizer, and use the same number of object queries and anchor boxes for fair comparison. The object queries are not pre-matched to classes, and are initialized as 0 as done in DAB-DETR. Since the model predictions are not pre-matched, the per-class post-matching is replaced by the vanilla one-to-one matching + +![](images/d07fdb4c783b7b97ee3fc4edcd3da4d29bc26fa0f53d2efa680078d996c5d90e.jpg) +Figure 4. Different choices of $p$ for class dropout. + +mechanism in DETR. Then, we classify the anchor boxes and condition the object queries on the class name embeddings, but do not set constraint on the post-matching, which boosts the performance on novel classes by 3.9 AP50. After adopting the full anchor pre-matching, the performance is significantly boosted by 10.8 AP50. + +Training Techniques. Class dropout and CLIP-aligned labeling are two training techniques that help the model generalize better. In Fig. 4, we examine the effect of different dropout probability. We find that the models trained with class dropout consistently outperform the baseline, and $p = 0.2$ gives the best performance. In Tab. 5, we validate the effectiveness of CLIP-Aligned Labeling. + +# 5. Conclusion + +The core challenge in open-vocabulary detection is how to effectively transfer the knowledge learned from the base classes to the unseen novel classes for evaluation. In this work, we directly adapt the CLIP into a region classifier, and mitigate the distribution gap between whole-image feature and regional feature through region prompting, which successfully generalizes to the novel categories. Different from prior works that rely on a fixed RPN for novel class localization, we achieve efficient class-aware localization by the proposed anchor pre-matching mechanism. Experiments show that our method can better transfer the knowledge from base classes to unseen novel classes with a smaller gap than prior works. We hope our work can help other researchers gain better insight on the OVD problem and develop better open-vocabulary detectors. + +Acknowledgement. This project is funded in part by National Key R&D Program of China Project 2022ZD0161100, by the Centre for Perceptual and Interactive Intelligence (CPII) Ltd under the Innovation and Technology Commission (ITC)'s InnoHK, by General Research Fund of Hong Kong RGC Project 14204021. Hongsheng Li is a PI of CPII under the InnoHK. This project is also supported by SenseTime Collaborative Research Grant. + +# References + +[1] Deformable detr: Deformable transformers. 2020. 2 +[2] Hyojin Bahng, Ali Jahanian, Swami Sankaranarayanan, and Phillip Isola. Exploring visual prompts for adapting large-scale models. arXiv preprint arXiv:2203.17274, 2022. 2 +[3] Hyojin Bahng, Ali Jahanian, Swami Sankaranarayanan, and Phillip Isola. Visual prompting: Modifying pixel space to adapt pre-trained models. ArXiv, abs/2203.17274, 2022. 2 +[4] Ankan Bansal, Karan Sikka, Gaurav Sharma, Rama Chellappa, and Ajay Divakaran. Zero-shot object detection. ArXiv, abs/1804.04340, 2018. 6 +[5] Tom B. Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert-Voss, Gretchen Krueger, T. J. Henighan, Rewon Child, Aditya Ramesh, Daniel M. Ziegler, Jeff Wu, Clemens Winter, Christopher Hesse, Mark Chen, Eric Sigler, Mateusz Litwin, Scott Gray, Benjamin Chess, Jack Clark, Christopher Berner, Sam McCandlish, Alec Radford, Ilya Sutskever, and Dario Amodei. Language models are few-shot learners. ArXiv, abs/2005.14165, 2020. 2 +[6] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. ArXiv, abs/2005.12872, 2020. 1, 2 +[7] Peixian Chen, Kekai Sheng, Mengdan Zhang, Yunhang Shen, Ke Li, and Chunhua Shen. Open vocabulary object detection with proposal mining and prediction equalization. CoRR, abs/2206.11134, 2022. 1, 7 +[8] Qiang Chen, Xiaokang Chen, Gang Zeng, and Jingdong Wang. Group detr: Fast training convergence with decoupled one-to-many label assignment. ArXiv, abs/2207.13085, 2022. 2 +[9] Xinlei Chen, Hao Fang, Tsung-Yi Lin, Ramakrishna Vedantam, Saurabh Gupta, Piotr Dólar, and C. Lawrence Zitnick. Microsoft coco captions: Data collection and evaluation server. ArXiv, abs/1504.00325, 2015. 7 +[10] Chengjian Feng, Yujie Zhong, Zequn Jie, Xiangxiang Chu, Haibing Ren, Xiaolin Wei, Weidi Xie, and Lin Ma. Promptdet: Towards open-vocabulary detection using uncurated images. 2022. 2 +[11] Mingfei Gao, Chen Xing, Juan Carlos Niebles, Junnan Li, Ran Xu, Wenhao Liu, and Caiming Xiong. Open vocabulary object detection with pseudo bounding-box labels. arXiv preprint arXiv:2111.09452, 2021. 2 +[12] Peng Gao, Shijie Geng, Renrui Zhang, Teli Ma, Rongyao Fang, Yongfeng Zhang, Hongsheng Li, and Yu Jiao Qiao. Clip-adapter: Better vision-language models with feature adapters. ArXiv, abs/2110.04544, 2021. 7 +[13] Peng Gao, Minghang Zheng, Xiaogang Wang, Jifeng Dai, and Hongsheng Li. Fast convergence of detr with spatially modulated co-attention. 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 3601–3610, 2021. 2 +[14] Xiuye Gu, Tsung-Yi Lin, Weicheng Kuo, and Yin Cui. Open-vocabulary object detection via vision and language knowledge distillation. In ICLR, 2022. 1, 7 +[15] Xiuye Gu, Tsung-Yi Lin, Weicheng Kuo, and Yin Cui. Open-vocabulary object detection via vision and language knowledge distillation. In ICLR, 2022. 2 + +[16] Agrim Gupta, Piotr Dollar, and Ross Girshick. LVIS: A dataset for large vocabulary instance segmentation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2019. 2, 6, 7 +[17] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In 2017 IEEE International Conference on Computer Vision (ICCV), pages 2980-2988, 2017. 1 +[18] Ding Jia, Yuhui Yuan, Haodi He, Xiao pei Wu, Haojun Yu, Weihong Lin, Lei huan Sun, Chao Zhang, and Hanhua Hu. Detrs with hybrid matching. ArXiv, abs/2207.13080, 2022. 2 +[19] Menglin Jia, Luming Tang, Bor-Chun Chen, Claire Cardie, Serge J. Belongie, Bharath Hariharan, and Ser Nam Lim. Visual prompt tuning. ArXiv, abs/2203.12119, 2022. 2 +[20] Chen Ju, Tengda Han, Kunhao Zheng, Ya Zhang, and Weidi Xie. Prompting visual-language models for efficient video understanding. ArXiv, abs/2112.04478, 2021. 2 +[21] Liunian Harold Li, Pengchuan Zhang, Haotian Zhang, Jianwei Yang, Chunyuan Li, Yiwu Zhong, Lijuan Wang, Lu Yuan, Lei Zhang, Jenq-Neng Hwang, Kai-Wei Chang, and Jianfeng Gao. Grounded language-image pre-training. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10955–10965, 2022. 2 +[22] Xiang Lisa Li and Percy Liang. Prefix-tuning: Optimizing continuous prompts for generation. Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), abs/2101.00190, 2021. 2 +[23] Tsung-Yi Lin, Priya Goyal, Ross B. Girshick, Kaiming He, and Piotr Dólar. Focal loss for dense object detection. 2017 IEEE International Conference on Computer Vision (ICCV), pages 2999-3007, 2017. 5 +[24] Tsung-Yi Lin, Michael Maire, Serge J. Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C. Lawrence Zitnick. Microsoft coco: Common objects in context. In ECCV, 2014. 2, 6 +[25] Ziyi Lin, Shijie Geng, Renrui Zhang, Peng Gao, Gerard de Melo, Xiaogang Wang, Jifeng Dai, Yu Qiao, and Hongsheng Li. Frozen clip models are efficient video learners. arXiv preprint arXiv:2208.03550, 2022. 2 +[26] Shilong Liu, Feng Li, Hao Zhang, Xiao Bin Yang, Xianbiao Qi, Hang Su, Jun Zhu, and Lei Zhang. Dab-detr: Dynamic anchor boxes are better queries for detr. ArXiv, abs/2201.12329, 2022. 2, 5, 6, 8 +[27] Xiao Liu, Kaixuan Ji, Yicheng Fu, Zhengxiao Du, Zhilin Yang, and Jie Tang. P-tuning v2: Prompt tuning can be comparable to fine-tuning universally across scales and tasks. CoRR, abs/2110.07602, 2021. 2 +[28] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In ICLR, 2019. 6 +[29] Muhammad Maaz, Hanoona Rasheed, Salman Khan, Fahad Shahbaz Khan, Rao Muhammad Anwer, and Ming-Hsuan Yang. Class-agnostic object detection with multimodal transformer. In 17th European Conference on Computer Vision (ECCV). Springer, 2022. 1 +[30] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual + +models from natural language supervision. In ICML, 2021. 1, 6 +[31] Hanoona Rasheed, Muhammad Maaz, Muhammad Uzair Khattak, Salman Khan, and Fahad Shahbaz Khan. Bridging the gap between object and image-level representations for open-vocabulary detection. ArXiv, abs/2207.03482, 2022. 1, 7 +[32] Shaoqing Ren, Kaiming He, Ross B. Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. IEEE Transactions on Pattern Analysis and Machine Intelligence, 39:1137-1149, 2015. 4 +[33] Piyush Sharma, Nan Ding, Sebastian Goodman, and Radu Soricut. Conceptual captions: A cleaned, hypernymed, image alt-text dataset for automatic image captioning. In ACL, 2018. 7 +[34] Yinghui Xing, Qirui Wu, De Cheng, Shizhou Zhang, Guoqiang Liang, and Yanning Zhang. Class-aware visual prompt tuning for vision-language pre-trained model. CoRR, abs/2208.08340, 2022. 2 +[35] Yuhang Zang, Wei Li, Kaiyang Zhou, Chen Huang, and Chen Change Loy. Open-vocabulary detr with conditional matching. ArXiv, abs/2203.11876, 2022. 1, 2, 5, 7 +[36] Alireza Zareian, Kevin Dela Rosa, Derek Hao Hu, and Shih-Fu Chang. Open-vocabulary object detection using captions. In 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 14388-14397, 2021. 1, 2, 7 +[37] Hongyu Zhai, Jian Cheng, and Mengyong Wang. Rethink the iou-based loss functions for bounding box regression. 2020 IEEE 9th Joint International Information Technology and Artificial Intelligence Conference (ITAIC), 9:1522-1528, 2020. 5 +[38] Gongjie Zhang, Zhipeng Luo, Yingchen Yu, Kaiwen Cui, and Shijian Lu. Accelerating DETR convergence via semantic-aligned matching. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 949-958, 2022. 2, 6 +[39] Shiyu Zhao, Zhixing Zhang, Samuel Schulter, Long Zhao, Vijay Kumar BG, Anastasis Stathopoulos, Manmohan Chandraker, and Dimitris N. Metaxas. Exploiting unlabeled data with vision and language models for object detection. CoRR, abs/2207.08954, 2022. 1, 7 +[40] Yiwu Zhong, Jianwei Yang, Pengchuan Zhang, Chengkun Li, Noel C. F. Codella, Liunian Harold Li, Luowei Zhou, Xiyang Dai, Lu Yuan, Yin Li, and Jianfeng Gao. Regionclip: Region-based language-image pretraining. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 16772-16782, 2022. 1, 2, 6, 7 +[41] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Learning to prompt for vision-language models. Int. J. Comput. Vis., 130:2337-2348, 2022. 7 +[42] Xingyi Zhou, Rohit Girdhar, Armand Joulin, Philipp Krahenbihl, and Ishan Misra. Detecting twenty-thousand classes using image-level supervision. In ECCV, 2022. 7 +[43] Xingyi Zhou, Vladlen Koltun, and Philipp Krahenbuhl. Probabilistic two-stage detection. In arXiv preprint arXiv:2103.07461, 2021. 6 \ No newline at end of file diff --git a/2023/CORA_ Adapting CLIP for Open-Vocabulary Detection With Region Prompting and Anchor Pre-Matching/images.zip b/2023/CORA_ Adapting CLIP for Open-Vocabulary Detection With Region Prompting and Anchor Pre-Matching/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..edf2715b0f849009922639562353d114e2ef8a58 --- /dev/null +++ b/2023/CORA_ Adapting CLIP for Open-Vocabulary Detection With Region Prompting and Anchor Pre-Matching/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:543badb63c7dae6d0d1b068b9b69be85d8807306d4a212bd0c2874022ee12879 +size 405290 diff --git a/2023/CORA_ Adapting CLIP for Open-Vocabulary Detection With Region Prompting and Anchor Pre-Matching/layout.json b/2023/CORA_ Adapting CLIP for Open-Vocabulary Detection With Region Prompting and Anchor Pre-Matching/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..850a187483d113da3de7f840c24bf714c747191a --- /dev/null +++ b/2023/CORA_ Adapting CLIP for Open-Vocabulary Detection With Region Prompting and Anchor Pre-Matching/layout.json @@ -0,0 +1,7517 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 110, + 102, + 483, + 137 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 110, + 102, + 483, + 137 + ], + "spans": [ + { + "bbox": [ + 110, + 102, + 483, + 137 + ], + "type": "text", + "content": "CORA: Adapting CLIP for Open-Vocabulary Detection with Region Prompting and Anchor Pre-Matching" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 159, + 158, + 430, + 173 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 158, + 430, + 173 + ], + "spans": [ + { + "bbox": [ + 159, + 158, + 430, + 173 + ], + "type": "text", + "content": "Xiaoshi Wu" + }, + { + "bbox": [ + 159, + 158, + 430, + 173 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 159, + 158, + 430, + 173 + ], + "type": "text", + "content": ", Feng Zhu" + }, + { + "bbox": [ + 159, + 158, + 430, + 173 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 159, + 158, + 430, + 173 + ], + "type": "text", + "content": ", Rui Zhao" + }, + { + "bbox": [ + 159, + 158, + 430, + 173 + ], + "type": "inline_equation", + "content": "^{2,3}" + }, + { + "bbox": [ + 159, + 158, + 430, + 173 + ], + "type": "text", + "content": ", Hongsheng Li" + }, + { + "bbox": [ + 159, + 158, + 430, + 173 + ], + "type": "inline_equation", + "content": "^{1,4}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 143, + 174, + 452, + 188 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 174, + 452, + 188 + ], + "spans": [ + { + "bbox": [ + 143, + 174, + 452, + 188 + ], + "type": "text", + "content": "1Multimedia Laboratory, The Chinese University of Hong Kong" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 88, + 188, + 506, + 201 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 88, + 188, + 506, + 201 + ], + "spans": [ + { + "bbox": [ + 88, + 188, + 506, + 201 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 88, + 188, + 506, + 201 + ], + "type": "text", + "content": "SenseTime Research " + }, + { + "bbox": [ + 88, + 188, + 506, + 201 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 88, + 188, + 506, + 201 + ], + "type": "text", + "content": "Qing Yuan Research Institute, Shanghai Jiao Tong University" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 159, + 201, + 434, + 214 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 201, + 434, + 214 + ], + "spans": [ + { + "bbox": [ + 159, + 201, + 434, + 214 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 159, + 201, + 434, + 214 + ], + "type": "text", + "content": "Centre for Perceptual and Interactive Intelligence (CPII)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 107, + 216, + 484, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 107, + 216, + 484, + 228 + ], + "spans": [ + { + "bbox": [ + 107, + 216, + 484, + 228 + ], + "type": "text", + "content": "{wuxiaoshi@link, hsli@ee}.cuhk.edu.hk,{zhufeng, zhaorui}@sentetime.com" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 143, + 255, + 191, + 268 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 255, + 191, + 268 + ], + "spans": [ + { + "bbox": [ + 143, + 255, + 191, + 268 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 281, + 289, + 587 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 281, + 289, + 587 + ], + "spans": [ + { + "bbox": [ + 47, + 281, + 289, + 587 + ], + "type": "text", + "content": "Open-vocabulary detection (OVD) is an object detection task aiming at detecting objects from novel categories beyond the base categories on which the detector is trained. Recent OVD methods rely on large-scale visual-language pre-trained models, such as CLIP, for recognizing novel objects. We identify the two core obstacles that need to be tackled when incorporating these models into detector training: (1) the distribution mismatch that happens when applying a VL-model trained on whole images to region recognition tasks; (2) the difficulty of localizing objects of unseen classes. To overcome these obstacles, we propose CORA, a DETR-style framework that adapts CLIP for Open-vocabulary detection by Region prompting and Anchor pre-matching. Region prompting mitigates the whole-to-region distribution gap by prompting the region features of the CLIP-based region classifier. Anchor pre-matching helps learning generalizable object localization by a class-aware matching mechanism. We evaluate CORA on the COCO OVD benchmark, where we achieve 41.7 AP50 on novel classes, which outperforms the previous SOTA by 2.4 AP50 even without resorting to extra training data. When extra training data is available, we train " + }, + { + "bbox": [ + 47, + 281, + 289, + 587 + ], + "type": "inline_equation", + "content": "\\mathrm{CORA^{+}}" + }, + { + "bbox": [ + 47, + 281, + 289, + 587 + ], + "type": "text", + "content": " on both ground-truth base-category annotations and additional pseudo bounding box labels computed by CORA. " + }, + { + "bbox": [ + 47, + 281, + 289, + 587 + ], + "type": "inline_equation", + "content": "\\mathrm{CORA^{+}}" + }, + { + "bbox": [ + 47, + 281, + 289, + 587 + ], + "type": "text", + "content": " achieves 43.1 AP50 on the COCO OVD benchmark and 28.1 box APr on the LVIS OVD benchmark. The code is available at https://github.com/tgxs002/CORA." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 624, + 127, + 636 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 624, + 127, + 636 + ], + "spans": [ + { + "bbox": [ + 47, + 624, + 127, + 636 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 644, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 644, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 644, + 287, + 714 + ], + "type": "text", + "content": "Object detection is a fundamental vision problem that involves localizing and classifying objects from images. Classical object detection requires detecting objects from a closed set of categories. Extra annotations and training are required if objects of unseen categories need to be detected. It has attracted much attention on detecting novel categories" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 256, + 545, + 290 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 256, + 545, + 290 + ], + "spans": [ + { + "bbox": [ + 305, + 256, + 545, + 290 + ], + "type": "text", + "content": "without tedious annotations, or even detect object from new category, which is currently referred as open-vocabulary detection (OVD) [36]." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 291, + 546, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 291, + 546, + 392 + ], + "spans": [ + { + "bbox": [ + 304, + 291, + 546, + 392 + ], + "type": "text", + "content": "Recent advances on large-scale vision-language pretrained models, such as CLIP [30], enable new solutions for tackling OVD. CLIP learns a joint embedding space of images and text from a large-scale image-text dataset, which shows remarkable capability on visual recognition tasks. The general idea of applying CLIP for OVD is to treat it as an open-vocabulary classifier. However, there are two obstacles hindering the effective use of CLIP on tackling OVD." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 393, + 547, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 393, + 547, + 540 + ], + "spans": [ + { + "bbox": [ + 304, + 393, + 547, + 540 + ], + "type": "text", + "content": "How to adapt CLIP for region-level tasks? One trivial solution is to crop regions and treat them as separate images, which has been adopted by multiple recent works [7, 14, 31, 35]. But the distribution gap between region crops and full images leads to inferior classification accuracy. MEDet [7] mitigates this issue by augmenting the text feature with image features. However, it requires extra image-text pairs to prevent overfitting to so-called \"base\" classes that are seen during training. RegionCLIP [40] directly acquires regional features by RoIAlign [17], which is more efficient but cannot generalize well to novel classes without finetuning. The finetuning is costly when adopting a larger CLIP model." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 542, + 547, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 542, + 547, + 689 + ], + "spans": [ + { + "bbox": [ + 304, + 542, + 547, + 689 + ], + "type": "text", + "content": "How to learn generalizable object proposals? ViLD [14], OV-DETR [35], Object-Centric-OVD [31], Region-CLIP [40] need RPN or class-agnostic object detectors [29] to mine potential novel class objects. However, these RPNs are strongly biased towards the base classes on which they are trained, while perform poorly on the novel classes. MEDet [7] and VL-PLM [39] identify this problem and adopt several handcrafted policies to rule out or merge low-quality boxes, but the performance is still bounded by the frozen RPN. OV-DETR [35] learns generalizable object localization by conditioning box regression on class name embeddings, but at the cost of efficiency issue induced by repetitive per-class inference." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "content": "In this work, we propose a new framework based on DETection TR transformers (DETR) [6] that incorporates CLIP" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "7031" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 49, + 72, + 286, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 72, + 286, + 277 + ], + "spans": [ + { + "bbox": [ + 49, + 72, + 286, + 277 + ], + "type": "text", + "content": "into detector training to achieve open-vocabulary detection without additional image-text data. Specifically, we use a DETR-style object localizer for class-aware object localization, and the predicted boxes are encoded by pooling the intermediate feature map of the CLIP image encoder, which are classified by the CLIP text encoder with class names. However, there is a distribution gap between whole-image features from CLIP's original visual encoder and the newly pooled region features, leading to an inferior classification accuracy. Thus, we propose Region Prompting to adapt the CLIP image encoder, which boosts the classification performance, and also demonstrates better generalization capability than existing methods. We adopt DAB-DETR [26] as the localizer, in which object queries are associated with dynamic anchor boxes. By pre-matching the dynamic anchor boxes with the input categories before box regression (Anchor Pre-Matching), class-aware regression can be achieved without the cost of repetitive per-class inference." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 278, + 286, + 471 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 278, + 286, + 471 + ], + "spans": [ + { + "bbox": [ + 49, + 278, + 286, + 471 + ], + "type": "text", + "content": "We validate our method on COCO [24] and LVIS v1.0 [16] OVD benchmarks. On the COCO OVD benchmark, our method improves AP50 of novel categories over the previous best method [40] by 2.4 AP50 without training on extra data, and achieves consistent gain on CLIP models of different scales. When compared under a fairer setting with extra training data, our method significantly outperforms the existing methods by 3.8 AP50 on novel categories and achieves comparable performance on base categories. On the LVIS OVD benchmark, our method achieves 22.2/28.1 APr with/w.o. extra data, which significantly outperforms existing methods that are also trained with/w.o. extra data. By applying region prompting on the base classes of COCO, the classification performance on the novel classes is boosted from " + }, + { + "bbox": [ + 49, + 278, + 286, + 471 + ], + "type": "inline_equation", + "content": "63.9\\%" + }, + { + "bbox": [ + 49, + 278, + 286, + 471 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 49, + 278, + 286, + 471 + ], + "type": "inline_equation", + "content": "74.1\\%" + }, + { + "bbox": [ + 49, + 278, + 286, + 471 + ], + "type": "text", + "content": ", whereas other prompting or adaptation methods easily bias towards the base classes." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 472, + 286, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 472, + 286, + 552 + ], + "spans": [ + { + "bbox": [ + 49, + 472, + 286, + 552 + ], + "type": "text", + "content": "The contributions of this work are summarized as follows: (1) Our proposed region prompting effectively mitigates the gap between whole image features and region features, and generalize well in the open-vocabulary setting. (2) Anchor Pre-Matching enables DETR for generalizable object localization efficiently. (3) We achieve state-of-the-art performance on COCO and LVIS OVD benchmarks." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 567, + 137, + 578 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 567, + 137, + 578 + ], + "spans": [ + { + "bbox": [ + 49, + 567, + 137, + 578 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 49, + 587, + 286, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 587, + 286, + 712 + ], + "spans": [ + { + "bbox": [ + 49, + 587, + 286, + 712 + ], + "type": "text", + "content": "Open-Vocabulary Object Detection OVR-CNN [36] firstly put forth this new formulation of detection, and proposes its baseline solution by aligning region features with nouns in captions that are paired with the image. Mingfeng [11] et al. mine pseudo labels by utilizing the localization ability of pre-trained vision-language models. PromptDet [10] addresses the gap between image and region classification by adding learnable prompts when encoding the class names, namely regional prompt learning (RPL), which is expected to generalize from base to novel categories. OV-DETR [35] is the first DETR-style open" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 308, + 72, + 544, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 72, + 544, + 186 + ], + "spans": [ + { + "bbox": [ + 308, + 72, + 544, + 186 + ], + "type": "text", + "content": "vocabulary detector, which proposes conditional matching to solve the missing novel class problem in assignment, but at the cost of inefficient inference. RegionCLIP [40] propose a second-stage pre-training mechanism to adapt the CLIP model to encode region features, and demonstrates its capability on OVD and zero-shot transfer setting. GLIP [21] jointly learns object localization and VL alignment. Matthias et al. [15] proposes to finetune a VL aligned model for detection, while we fix the pre-trained VL model for better generalization towards novel categories." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 308, + 190, + 544, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 190, + 544, + 373 + ], + "spans": [ + { + "bbox": [ + 308, + 190, + 544, + 373 + ], + "type": "text", + "content": "Detection Transformers DETR [6] is an object detection architecture based on transformers that formulates object detection as a set-to-set matching problem, which greatly simplifies the pipeline. Several works address the slow convergence problem of DETR by architectural improvement [1, 13,26,38] or special training strategies [8,18]. Zhu et al. [1] proposes multi-scale deformable attention module to efficiently aggregate information from multi-scale feature maps. Gao et al. [13] proposes to modulate the cross-attention in the transformer decoder by anchor box coordinates to accelerate the detector convergence. DABDETR [26] formulates the queries in DETR architecture as anchor boxes, which accelerates detector training. Chen et al. [38] proposes Group DETR, which adds auxiliary object queries during training to take advantage of one-to-many matching for faster convergence." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 308, + 378, + 544, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 378, + 544, + 514 + ], + "spans": [ + { + "bbox": [ + 308, + 378, + 544, + 514 + ], + "type": "text", + "content": "Prompt Tuning Prompting is originated from NLP, and it refers to prepending task instructions before the input sequence to give the language model the hint about the task [5]. Later works [22, 27] explores tuning continuous prompt vectors when few-shot data is available. VPT [19], Visual Prompting [2, 3] explore prompting in the pixel space. [20] and [25] prompts pre-trained model for video recognition tasks. [34] proposes class-aware visual prompt tuning to generalize the learned prompts to unseen categories. Recent works demonstrate that prompt tuning is an effective and parameter-efficient way to adapt large-scale pre-trained models to downstream tasks." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 308, + 540, + 359, + 552 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 540, + 359, + 552 + ], + "spans": [ + { + "bbox": [ + 308, + 540, + 359, + 552 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 308, + 564, + 544, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 564, + 544, + 712 + ], + "spans": [ + { + "bbox": [ + 308, + 564, + 544, + 712 + ], + "type": "text", + "content": "Open-vocabulary detection (OVD) is an object detection task aiming at detecting objects from novel categories beyond the base categories on which the detector is trained. Formally, the detector is trained on a detection dataset with base-category " + }, + { + "bbox": [ + 308, + 564, + 544, + 712 + ], + "type": "inline_equation", + "content": "C^B" + }, + { + "bbox": [ + 308, + 564, + 544, + 712 + ], + "type": "text", + "content": " box annotations, and tested on a new input image " + }, + { + "bbox": [ + 308, + 564, + 544, + 712 + ], + "type": "inline_equation", + "content": "I \\in \\mathbb{R}^{H \\times W \\times 3}" + }, + { + "bbox": [ + 308, + 564, + 544, + 712 + ], + "type": "text", + "content": " to detect objects belonging to a novel category set " + }, + { + "bbox": [ + 308, + 564, + 544, + 712 + ], + "type": "inline_equation", + "content": "C^N" + }, + { + "bbox": [ + 308, + 564, + 544, + 712 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 308, + 564, + 544, + 712 + ], + "type": "inline_equation", + "content": "C^B \\cap C^N = \\emptyset" + }, + { + "bbox": [ + 308, + 564, + 544, + 712 + ], + "type": "text", + "content": ". In this section, we introduce CORA, a framework that adapts CLIP for the OVD task by Region prompting and Anchor prematching. For fairer comparisons with existing methods, we also experiment with a broader setting where extra data is available, which is referred to as " + }, + { + "bbox": [ + 308, + 564, + 544, + 712 + ], + "type": "inline_equation", + "content": "\\mathrm{CORA}^+" + }, + { + "bbox": [ + 308, + 564, + 544, + 712 + ], + "type": "text", + "content": ", and will be introduced along with the experiments." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 750, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 750, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 750, + 314, + 757 + ], + "type": "text", + "content": "7032" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 67, + 79, + 526, + 345 + ], + "blocks": [ + { + "bbox": [ + 67, + 79, + 526, + 345 + ], + "lines": [ + { + "bbox": [ + 67, + 79, + 526, + 345 + ], + "spans": [ + { + "bbox": [ + 67, + 79, + 526, + 345 + ], + "type": "image", + "image_path": "3a099067d201f95e92c3386783275c33f2e4c8b2a56b18f3527f8edca9ed3c16.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 362, + 547, + 407 + ], + "lines": [ + { + "bbox": [ + 46, + 362, + 547, + 407 + ], + "spans": [ + { + "bbox": [ + 46, + 362, + 547, + 407 + ], + "type": "text", + "content": "Figure 1. Overview of our method. The image is encoded into a feature map by the CLIP image encoder for both localization and classification. The regional feature is extracted by pooling the feature map, and then prompted before classified by the CLIP class name embeddings. The anchor boxes are pre-matched and conditioned on a class before decoding. During training, per-class post-matching is conducted. During inference, the box predictions are classified by the region classifier." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 426, + 115, + 437 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 426, + 115, + 437 + ], + "spans": [ + { + "bbox": [ + 47, + 426, + 115, + 437 + ], + "type": "text", + "content": "3.1. Overview" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 445, + 287, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 445, + 287, + 581 + ], + "spans": [ + { + "bbox": [ + 46, + 445, + 287, + 581 + ], + "type": "text", + "content": "The overall framework of CORA is illustrated in Fig. 1. Given an image as input, we acquire the spatial feature map using the ResNet backbone from the pre-trained CLIP image encoder, which is shared by both region classification and object localization branches. Unlike conventional detectors, localization and classification are decoupled and sequentially conducted in our framework to better fit the characteristic of the OVD problem. We train a DETR-style object localizer that refines a set of object queries together with their associated anchor boxes to localize the objects, which are then classified by a region classifier adapted from CLIP." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 585, + 287, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 585, + 287, + 665 + ], + "spans": [ + { + "bbox": [ + 46, + 585, + 287, + 665 + ], + "type": "text", + "content": "Region Classification. Given a region to be classified (anchor box or box prediction), we adopt RoIAlign to obtain the region feature, followed by the attention pooling module of CLIP to generate region embeddings, which can be classified by class embeddings obtained from the CLIP text encoder, as done in CLIP. We name this module as CLIP-based region classifier (Fig. 1-(a))." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 667, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 667, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 667, + 287, + 715 + ], + "type": "text", + "content": "Object Localization. The visual feature map is firstly refined by the DETR-like encoder, and then fed into the DETR-like decoder. The queries of anchor boxes are firstly classified by the CLIP-based region classifier, which are" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 426, + 545, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 426, + 545, + 517 + ], + "spans": [ + { + "bbox": [ + 304, + 426, + 545, + 517 + ], + "type": "text", + "content": "then conditioned on their predicted labels before iteratively refined by the DETR-like decoder for better localization. The decoder also estimates the matchability of the query with the previously predicted label. During training, the predicted boxes are one-to-one matched with ground truth boxes that has the same labels (Fig. 1-(b)), and trained as in DETR. In the inference stage, the class labels of the boxes are adjusted by the CLIP-based region classifier." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 517, + 546, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 517, + 546, + 665 + ], + "spans": [ + { + "bbox": [ + 304, + 517, + 546, + 665 + ], + "type": "text", + "content": "As mentioned in Sec. 1, there are two obstacles to be addressed: (1) object detection conducts recognition on image regions, while the CLIP model is trained on whole-image input, leading to a distribution gap that hinders classification performance. (2) the detector needs to learn object localization for novel classes, while we only have annotations on a limited number of base classes. To solve the first obstacle, we propose region prompting to modulate the region features for better generalizable region embeddings, which will be introduced in Sec. 3.2. To solve the second obstacle, we put forward anchor pre-matching to encourage class-aware object localization that can generalize to novel classes during inference, which will be introduced in Sec. 3.3." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 672, + 416, + 685 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 672, + 416, + 685 + ], + "spans": [ + { + "bbox": [ + 306, + 672, + 416, + 685 + ], + "type": "text", + "content": "3.2. Region Prompting" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 689, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 545, + 714 + ], + "type": "text", + "content": "OVD requires the detector to classify image regions into a given category list. In this section, we will elaborate how" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7033" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 76, + 282, + 133 + ], + "blocks": [ + { + "bbox": [ + 50, + 76, + 282, + 133 + ], + "lines": [ + { + "bbox": [ + 50, + 76, + 282, + 133 + ], + "spans": [ + { + "bbox": [ + 50, + 76, + 282, + 133 + ], + "type": "image", + "image_path": "d4b96808f0036c773207a589026eb6d4894e7a822810cea81d943805e2e19cff.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 123, + 140, + 212, + 152 + ], + "lines": [ + { + "bbox": [ + 123, + 140, + 212, + 152 + ], + "spans": [ + { + "bbox": [ + 123, + 140, + 212, + 152 + ], + "type": "text", + "content": "(a) Whole-image encoding" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 50, + 161, + 282, + 232 + ], + "blocks": [ + { + "bbox": [ + 50, + 161, + 282, + 232 + ], + "lines": [ + { + "bbox": [ + 50, + 161, + 282, + 232 + ], + "spans": [ + { + "bbox": [ + 50, + 161, + 282, + 232 + ], + "type": "image", + "image_path": "093ce500705802d0b9bcbb337745ddd9a613030762f101636c7d7f5bd7cc1949.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 238, + 204, + 248 + ], + "lines": [ + { + "bbox": [ + 130, + 238, + 204, + 248 + ], + "spans": [ + { + "bbox": [ + 130, + 238, + 204, + 248 + ], + "type": "text", + "content": "(b) Regional encoding" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 319, + 286, + 388 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 319, + 286, + 388 + ], + "spans": [ + { + "bbox": [ + 46, + 319, + 286, + 388 + ], + "type": "text", + "content": "a pre-trained CLIP model is adapted to formulate the CLIP-based region classifier. Given the CLIP model, region classification can be realized by comparing the similarity between the regional embedding from the CLIP image encoder and the class name embedding from the CLIP text encoder." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 388, + 286, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 388, + 286, + 548 + ], + "spans": [ + { + "bbox": [ + 46, + 388, + 286, + 548 + ], + "type": "text", + "content": "Region Prompting. As illustrated in Fig. 2, given an image and a set of region of interest (RoI), we firstly encode the whole image into a feature map by the CLIP encoder's first 3 blocks, which are then pooled by RoIAlign [32] either according to anchor boxes or predicted boxes into region features, before encoded by the last block of the CLIP image encoder backbone. There exists a distribution gap between the CLIP image encoder's whole-image feature map and the pooled regional features. We propose region prompting to fix the misalignment by augmenting the region feature with learnable prompts " + }, + { + "bbox": [ + 46, + 388, + 286, + 548 + ], + "type": "inline_equation", + "content": "p \\in \\mathbb{R}^{S \\times S \\times C}" + }, + { + "bbox": [ + 46, + 388, + 286, + 548 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 388, + 286, + 548 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 46, + 388, + 286, + 548 + ], + "type": "text", + "content": " is the spatial size of the regional feature, and " + }, + { + "bbox": [ + 46, + 388, + 286, + 548 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 46, + 388, + 286, + 548 + ], + "type": "text", + "content": " is the dimension of the regional feature. Specifically, given the input regional feature " + }, + { + "bbox": [ + 46, + 388, + 286, + 548 + ], + "type": "inline_equation", + "content": "f_{\\mathrm{region}}" + }, + { + "bbox": [ + 46, + 388, + 286, + 548 + ], + "type": "text", + "content": ", the region prompting is conducted as" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 111, + 556, + 287, + 570 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 111, + 556, + 287, + 570 + ], + "spans": [ + { + "bbox": [ + 111, + 556, + 287, + 570 + ], + "type": "interline_equation", + "content": "v _ {\\text {p r o m p t}} = P \\left(f _ {\\text {r e g i o n}} \\oplus p\\right), \\tag {1}", + "image_path": "80732b803f321762721ed32919a23e866d14251e5987e76d1e25dea69f53ad76.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 576, + 286, + 598 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 576, + 286, + 598 + ], + "spans": [ + { + "bbox": [ + 46, + 576, + 286, + 598 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 576, + 286, + 598 + ], + "type": "inline_equation", + "content": "\\oplus" + }, + { + "bbox": [ + 46, + 576, + 286, + 598 + ], + "type": "text", + "content": " denotes element-wise addition, " + }, + { + "bbox": [ + 46, + 576, + 286, + 598 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 46, + 576, + 286, + 598 + ], + "type": "text", + "content": " is the attention pooling module of the CLIP visual encoder." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 599, + 286, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 599, + 286, + 700 + ], + "spans": [ + { + "bbox": [ + 46, + 599, + 286, + 700 + ], + "type": "text", + "content": "Optimizing Region Prompts. We train the region prompts on a detection dataset with base-class annotations. The class name embeddings are pre-computed by the CLIP text encoder, which are used as classifier weights later. We train the prompts by a standard cross-entropy loss to classify the ground truth boxes with their pooled regional features " + }, + { + "bbox": [ + 46, + 599, + 286, + 700 + ], + "type": "inline_equation", + "content": "f_{\\mathrm{region}}" + }, + { + "bbox": [ + 46, + 599, + 286, + 700 + ], + "type": "text", + "content": ". When optimizing the region prompts, we keep other model weights frozen, and only make the region prompts to be learned." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 701, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 701, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 701, + 286, + 713 + ], + "type": "text", + "content": "Comparison with existing methods. The previous com" + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 312, + 72, + 531, + 133 + ], + "blocks": [ + { + "bbox": [ + 312, + 72, + 531, + 133 + ], + "lines": [ + { + "bbox": [ + 312, + 72, + 531, + 133 + ], + "spans": [ + { + "bbox": [ + 312, + 72, + 531, + 133 + ], + "type": "image", + "image_path": "5f22df4d1ddf32c8fd272c91cd51d2a1f68a26cae1494f8fb0add14df39bedca.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 312, + 137, + 530, + 199 + ], + "blocks": [ + { + "bbox": [ + 312, + 137, + 530, + 199 + ], + "lines": [ + { + "bbox": [ + 312, + 137, + 530, + 199 + ], + "spans": [ + { + "bbox": [ + 312, + 137, + 530, + 199 + ], + "type": "image", + "image_path": "ea644c72a4a7a810463b136d9cc3252acc7d45a5b9b07513a6397e57585bfe83.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 321, + 206, + 530, + 266 + ], + "blocks": [ + { + "bbox": [ + 321, + 206, + 530, + 266 + ], + "lines": [ + { + "bbox": [ + 321, + 206, + 530, + 266 + ], + "spans": [ + { + "bbox": [ + 321, + 206, + 530, + 266 + ], + "type": "image", + "image_path": "c79b27b5ef350e5f411263794dd0c7246d457b7f0fd6fb47c5cf91687b8f533b.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 385, + 270, + 468, + 280 + ], + "lines": [ + { + "bbox": [ + 385, + 270, + 468, + 280 + ], + "spans": [ + { + "bbox": [ + 385, + 270, + 468, + 280 + ], + "type": "text", + "content": "(a) Conditional matching" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 320, + 284, + 532, + 365 + ], + "blocks": [ + { + "bbox": [ + 46, + 257, + 288, + 300 + ], + "lines": [ + { + "bbox": [ + 46, + 257, + 288, + 300 + ], + "spans": [ + { + "bbox": [ + 46, + 257, + 288, + 300 + ], + "type": "text", + "content": "Figure 2. Comparison between the CLIP-based region classifier and the vanilla pipeline. We pool regional feature from the feature map, rather than cropping the region patch and classify them as a separate image." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 320, + 284, + 532, + 365 + ], + "lines": [ + { + "bbox": [ + 320, + 284, + 532, + 365 + ], + "spans": [ + { + "bbox": [ + 320, + 284, + 532, + 365 + ], + "type": "image", + "image_path": "8518f2ee0e91295835e023b3060fabc7fff020afa782f935077b24e536c1bee2.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 385, + 368, + 467, + 378 + ], + "lines": [ + { + "bbox": [ + 385, + 368, + 467, + 378 + ], + "spans": [ + { + "bbox": [ + 385, + 368, + 467, + 378 + ], + "type": "text", + "content": "(b) Anchor pre-matching" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 304, + 387, + 545, + 440 + ], + "lines": [ + { + "bbox": [ + 304, + 387, + 545, + 440 + ], + "spans": [ + { + "bbox": [ + 304, + 387, + 545, + 440 + ], + "type": "text", + "content": "Figure 3. Comparison between anchor pre-matching and conditional matching. Anchor pre-matching decodes a constant number of object queries, and assigns different numbers of object queries based on the image content, avoiding repetitive decoding as in conditional matching." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 460, + 545, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 460, + 545, + 552 + ], + "spans": [ + { + "bbox": [ + 304, + 460, + 545, + 552 + ], + "type": "text", + "content": "mon practice for region classification with CLIP is to crop the RoIs, and encode them as separate images, before comparing with text embeddings. This pipeline is not efficient when encoding regions with overlaps, since the overlapping regions are encoded more than once in different region crops. Its accuracy also suffers from the missing context information. In contrast, our regional prompting is more efficient and preserves richer context." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 552, + 545, + 642 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 552, + 545, + 642 + ], + "spans": [ + { + "bbox": [ + 304, + 552, + 545, + 642 + ], + "type": "text", + "content": "The region prompts contain less than 1M parameters, which is in line with recent advances of the prompt tuning and adapter literatures. Region prompting generalizes well to unseen novel classes. We attribute the generalization capability to the fact that region prompting directly fix the distribution mismatch right after where it occurs (after region pooling), whereas the existing methods tweak irrelevant parameters to compensate for the distribution mismatch." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 306, + 649, + 432, + 662 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 649, + 432, + 662 + ], + "spans": [ + { + "bbox": [ + 306, + 649, + 432, + 662 + ], + "type": "text", + "content": "3.3. Anchor Pre-Matching" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 667, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 667, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 667, + 545, + 713 + ], + "type": "text", + "content": "Region prompting helps solve the region classification problem. Object localization is the other critical subtask of object detection. Considering the inferior performance of pre-trained RPN on novel classes, we introduce" + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7034" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 140 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 140 + ], + "type": "text", + "content": "a class-aware query-based object localizer, which demonstrates better generalization capability on unseen classes. As shown in Fig. 1, Given the visual feature map from the frozen CLIP image encoder, the object queries are prematched to the class name embeddings by the CLIP text encoder." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 140, + 288, + 243 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 140, + 288, + 243 + ], + "spans": [ + { + "bbox": [ + 46, + 140, + 288, + 243 + ], + "type": "text", + "content": "Anchor Pre-Matching. The object localizer is implemented by a DETR-style transformer encoder-decoder structure, where the encoder refines the feature map, and the decoder decodes a set of object queries into box predictions. We adopt DAB-DETR [26], where each object query is associated with an anchor box. Each ground truth box is pre-matched to a set of queries with the same label. The label " + }, + { + "bbox": [ + 46, + 140, + 288, + 243 + ], + "type": "inline_equation", + "content": "\\hat{c}_i" + }, + { + "bbox": [ + 46, + 140, + 288, + 243 + ], + "type": "text", + "content": " of an object query is assigned by classifying the associated anchor box " + }, + { + "bbox": [ + 46, + 140, + 288, + 243 + ], + "type": "inline_equation", + "content": "b_i" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 108, + 253, + 287, + 273 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 108, + 253, + 287, + 273 + ], + "spans": [ + { + "bbox": [ + 108, + 253, + 287, + 273 + ], + "type": "interline_equation", + "content": "\\hat {c} _ {i} = \\underset {c \\in C ^ {B}} {\\arg \\max } \\operatorname {c o s i n e} \\left(v _ {i}, l _ {c}\\right), \\tag {2}", + "image_path": "5160d92970e0bd8da96915f3ce406e88293bdb57c4561e476905263e99f8f71e.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 282, + 287, + 340 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 282, + 287, + 340 + ], + "spans": [ + { + "bbox": [ + 46, + 282, + 287, + 340 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 282, + 287, + 340 + ], + "type": "inline_equation", + "content": "v_{i}" + }, + { + "bbox": [ + 46, + 282, + 287, + 340 + ], + "type": "text", + "content": " is the region feature of anchor box " + }, + { + "bbox": [ + 46, + 282, + 287, + 340 + ], + "type": "inline_equation", + "content": "b_{i}, l_{c}" + }, + { + "bbox": [ + 46, + 282, + 287, + 340 + ], + "type": "text", + "content": " is the class name embedding of class " + }, + { + "bbox": [ + 46, + 282, + 287, + 340 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 46, + 282, + 287, + 340 + ], + "type": "text", + "content": ", and cosine denotes the cosine similarity. After pre-matching, each object query conditions on the predicted class embedding to allow class-aware box regression. The conditioned object query is given by" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 349, + 287, + 361 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 349, + 287, + 361 + ], + "spans": [ + { + "bbox": [ + 133, + 349, + 287, + 361 + ], + "type": "interline_equation", + "content": "q _ {i} = \\mathrm {M L P} \\left(l _ {\\mathrm {c} _ {\\mathrm {i}}}\\right). \\tag {3}", + "image_path": "e9946b6cf1c68fbcacc0841b819454fe4c119f07949e714e033f4d91a6d83f2f.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 371, + 287, + 429 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 371, + 287, + 429 + ], + "spans": [ + { + "bbox": [ + 47, + 371, + 287, + 429 + ], + "type": "text", + "content": "The DETR-like decoder iteratively refines each object query with its associated anchor box " + }, + { + "bbox": [ + 47, + 371, + 287, + 429 + ], + "type": "inline_equation", + "content": "(q_i, b_i)" + }, + { + "bbox": [ + 47, + 371, + 287, + 429 + ], + "type": "text", + "content": " into " + }, + { + "bbox": [ + 47, + 371, + 287, + 429 + ], + "type": "inline_equation", + "content": "\\hat{y}_i = (\\hat{p}_i, \\hat{b}_i)" + }, + { + "bbox": [ + 47, + 371, + 287, + 429 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 47, + 371, + 287, + 429 + ], + "type": "inline_equation", + "content": "\\hat{b}_i" + }, + { + "bbox": [ + 47, + 371, + 287, + 429 + ], + "type": "text", + "content": " is the refined box coordinates and " + }, + { + "bbox": [ + 47, + 371, + 287, + 429 + ], + "type": "inline_equation", + "content": "\\hat{p}_i" + }, + { + "bbox": [ + 47, + 371, + 287, + 429 + ], + "type": "text", + "content": " is the matching probability to the query's pre-matched class " + }, + { + "bbox": [ + 47, + 371, + 287, + 429 + ], + "type": "inline_equation", + "content": "\\hat{c}_i" + }, + { + "bbox": [ + 47, + 371, + 287, + 429 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 430, + 287, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 430, + 287, + 498 + ], + "spans": [ + { + "bbox": [ + 46, + 430, + 287, + 498 + ], + "type": "text", + "content": "Given the model predictions, the assignment between ground truth boxes and model predictions is conducted by performing bipartite matching for each class separately. We only allow each ground truth box to be assigned to the prediction with the same pre-matched label, in order to enforce the decoder to be aware of the conditioned text embedding." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 498, + 287, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 498, + 287, + 555 + ], + "spans": [ + { + "bbox": [ + 46, + 498, + 287, + 555 + ], + "type": "text", + "content": "Specifically, for class " + }, + { + "bbox": [ + 46, + 498, + 287, + 555 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 46, + 498, + 287, + 555 + ], + "type": "text", + "content": ", given the " + }, + { + "bbox": [ + 46, + 498, + 287, + 555 + ], + "type": "inline_equation", + "content": "N^c" + }, + { + "bbox": [ + 46, + 498, + 287, + 555 + ], + "type": "text", + "content": " box predictions " + }, + { + "bbox": [ + 46, + 498, + 287, + 555 + ], + "type": "inline_equation", + "content": "\\hat{y}^c = \\{\\hat{y}_i \\mid \\hat{c}_i = c\\}" + }, + { + "bbox": [ + 46, + 498, + 287, + 555 + ], + "type": "text", + "content": " that are pre-matched to class " + }, + { + "bbox": [ + 46, + 498, + 287, + 555 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 46, + 498, + 287, + 555 + ], + "type": "text", + "content": ", and the set of ground truth boxes " + }, + { + "bbox": [ + 46, + 498, + 287, + 555 + ], + "type": "inline_equation", + "content": "y^c" + }, + { + "bbox": [ + 46, + 498, + 287, + 555 + ], + "type": "text", + "content": " in class " + }, + { + "bbox": [ + 46, + 498, + 287, + 555 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 46, + 498, + 287, + 555 + ], + "type": "text", + "content": ", we optimize a permutation of " + }, + { + "bbox": [ + 46, + 498, + 287, + 555 + ], + "type": "inline_equation", + "content": "N^c" + }, + { + "bbox": [ + 46, + 498, + 287, + 555 + ], + "type": "text", + "content": " elements " + }, + { + "bbox": [ + 46, + 498, + 287, + 555 + ], + "type": "inline_equation", + "content": "\\sigma \\in \\mathfrak{S}_{\\mathrm{N}^c}" + }, + { + "bbox": [ + 46, + 498, + 287, + 555 + ], + "type": "text", + "content": " that minimizes the following cost" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 95, + 566, + 287, + 597 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 95, + 566, + 287, + 597 + ], + "spans": [ + { + "bbox": [ + 95, + 566, + 287, + 597 + ], + "type": "interline_equation", + "content": "\\hat {\\sigma} _ {c} = \\underset {\\sigma \\in \\mathfrak {S} _ {\\mathrm {N} ^ {\\mathrm {c}}}} {\\arg \\max } \\sum_ {i} ^ {N ^ {c}} \\mathcal {L} _ {\\operatorname {c o s t}} \\left(y _ {i} ^ {c}, \\hat {y} _ {\\sigma (\\mathrm {i})} ^ {c}\\right), \\tag {4}", + "image_path": "af96c5fcd5dcd9403c62ee0f303029acc9877f8450c923835ccfdeff49e50981.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 608, + 200, + 619 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 608, + 200, + 619 + ], + "spans": [ + { + "bbox": [ + 47, + 608, + 200, + 619 + ], + "type": "text", + "content": "where the matching cost is defined as" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 83, + 629, + 287, + 643 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 83, + 629, + 287, + 643 + ], + "spans": [ + { + "bbox": [ + 83, + 629, + 287, + 643 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {c o s t}} (y, \\hat {y}) = \\mathcal {L} _ {\\text {m a t c h}} (p, \\hat {p}) + \\mathcal {L} _ {\\text {b o x}} (b, \\hat {b}). \\tag {5}", + "image_path": "797e0bf4b319a8d2d2f1156922ce2ecdb9433b65809f9637dc4a4d583b838cb7.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 654, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 654, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 654, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{match}}(p,\\hat{p})" + }, + { + "bbox": [ + 47, + 654, + 287, + 713 + ], + "type": "text", + "content": " is a binary classification loss, and " + }, + { + "bbox": [ + 47, + 654, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{box}}(b,\\hat{b})" + }, + { + "bbox": [ + 47, + 654, + 287, + 713 + ], + "type": "text", + "content": " characterizes the localization error of " + }, + { + "bbox": [ + 47, + 654, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\hat{b}" + }, + { + "bbox": [ + 47, + 654, + 287, + 713 + ], + "type": "text", + "content": " w.r.t " + }, + { + "bbox": [ + 47, + 654, + 287, + 713 + ], + "type": "inline_equation", + "content": "b" + }, + { + "bbox": [ + 47, + 654, + 287, + 713 + ], + "type": "text", + "content": ". In our case, we implement " + }, + { + "bbox": [ + 47, + 654, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{match}}" + }, + { + "bbox": [ + 47, + 654, + 287, + 713 + ], + "type": "text", + "content": " by the focal loss [23]. " + }, + { + "bbox": [ + 47, + 654, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{box}}" + }, + { + "bbox": [ + 47, + 654, + 287, + 713 + ], + "type": "text", + "content": " is implemented by a weighted sum of " + }, + { + "bbox": [ + 47, + 654, + 287, + 713 + ], + "type": "inline_equation", + "content": "L_{1}" + }, + { + "bbox": [ + 47, + 654, + 287, + 713 + ], + "type": "text", + "content": " loss and GIoU [37] loss following prior works." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 318, + 72, + 503, + 84 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 72, + 503, + 84 + ], + "spans": [ + { + "bbox": [ + 318, + 72, + 503, + 84 + ], + "type": "text", + "content": "The model is optimized by the following loss" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 335, + 91, + 545, + 132 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 91, + 545, + 132 + ], + "spans": [ + { + "bbox": [ + 335, + 91, + 545, + 132 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} = \\sum_ {c \\in C ^ {B}} \\mathcal {L} _ {\\text {m a t c h}} \\left(p ^ {c}, \\hat {p} _ {\\hat {\\sigma} _ {c}} ^ {c}\\right) + \\mathcal {L} _ {\\text {b o x}} \\left(b ^ {c}, \\hat {b} _ {\\hat {\\sigma} _ {c}} ^ {c}\\right) \\tag {6} \\\\ = \\lambda_ {\\text {f o c a l}} \\mathcal {L} _ {\\text {f o c a l}} + \\lambda_ {\\mathrm {L} _ {1}} \\mathcal {L} _ {\\mathrm {L} _ {1}} + \\lambda_ {\\mathrm {G I o U}} \\mathcal {L} _ {\\mathrm {G I o U}}. \\\\ \\end{array}", + "image_path": "82c6b77f2e0e77622540ed1cee388206e51ad40c7d5cd78f6dcd9b9daea9ebda.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 138, + 545, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 138, + 545, + 186 + ], + "spans": [ + { + "bbox": [ + 304, + 138, + 545, + 186 + ], + "type": "text", + "content": "During inference, we adopt the region classifier introduced in Sec. 3.2 to classify the predicted boxes " + }, + { + "bbox": [ + 304, + 138, + 545, + 186 + ], + "type": "inline_equation", + "content": "\\{\\hat{b}_i\\}" + }, + { + "bbox": [ + 304, + 138, + 545, + 186 + ], + "type": "text", + "content": " for better classification accuracy. The class score is multiplied by the pre-matching score to account for the box quality" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 365, + 193, + 545, + 206 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 365, + 193, + 545, + 206 + ], + "spans": [ + { + "bbox": [ + 365, + 193, + 545, + 206 + ], + "type": "interline_equation", + "content": "\\mathrm {P} \\left(\\hat {b} _ {i} \\in c\\right) = \\hat {p} _ {i} \\operatorname {c o s i n e} \\left(\\hat {v} _ {i}, l _ {c}\\right). \\tag {7}", + "image_path": "eea9413c0731ed939e2aa1a82926db998f7b0367c3df2e0cdd798bfe2605ce19.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 213, + 545, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 213, + 545, + 361 + ], + "spans": [ + { + "bbox": [ + 304, + 213, + 545, + 361 + ], + "type": "text", + "content": "Comparison with Conditional Matching [35]. Conditional Matching in OV-DETR [35] also proposes to condition the queries on the text embedding for class-aware regression. But it suffers from repetitive per-class inference. Specifically, as shown in Fig. 3, each class in " + }, + { + "bbox": [ + 304, + 213, + 545, + 361 + ], + "type": "inline_equation", + "content": "C^N" + }, + { + "bbox": [ + 304, + 213, + 545, + 361 + ], + "type": "text", + "content": " needs to be separately localized with the same group of query, which means both computation and memory consumption scales linearly with the vocabulary size. During training, the number of negative classes sampled in each iteration is limited due to the memory constraint, which hinders convergence. During inference, repetitive per-class decoding is required and results in low inference efficiency, especially when the vocabulary size is large." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 361, + 545, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 361, + 545, + 441 + ], + "spans": [ + { + "bbox": [ + 304, + 361, + 545, + 441 + ], + "type": "text", + "content": "Contrary to conditional matching, our anchor pre-matching mechanism assigns anchor boxes for different classes adaptively according to the image content, which ensures a constant number of query decoupled from the category size. By anchor pre-matching, all the classes can be decoded together in one pass, eliminating the need for repetitive per-class decoding." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 441, + 545, + 497 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 441, + 545, + 497 + ], + "spans": [ + { + "bbox": [ + 304, + 441, + 545, + 497 + ], + "type": "text", + "content": "To improve the generalization capability and training convergence of open-vocabulary detectors equipped with Anchor Pre-Matching, we also introduce two effective training techniques, namely, \"Drop Class\" and \"CLIP-Aligned Labeling\"." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 497, + 545, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 497, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 304, + 497, + 545, + 689 + ], + "type": "text", + "content": "Class Dropout. The generalization capability of the model can be further boosted by randomly dropping categories during training. Since our goal is to train a detector that can detect objects from a user specified category list, training on a fixed list of categories leads to bias. We mitigate this bias by randomly dropping out the base categories during training. For efficiency reason, we implement this idea by splitting the base classes into two complementary groups and train on both of them, instead of training on one group while dropping the other group. Specifically, in each training iteration, we split " + }, + { + "bbox": [ + 304, + 497, + 545, + 689 + ], + "type": "inline_equation", + "content": "C^B" + }, + { + "bbox": [ + 304, + 497, + 545, + 689 + ], + "type": "text", + "content": " and the ground truth boxes at a probability of " + }, + { + "bbox": [ + 304, + 497, + 545, + 689 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 304, + 497, + 545, + 689 + ], + "type": "text", + "content": ", and train the detector on the same image with two complementary sets of categories. It enforces the model to condition its prediction on the query categories. Since a ground truth box in one group does not appear in the other group, the model needs to be aware of the categories to treat differently for the two sets of annotations." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 306, + 690, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 690, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 690, + 545, + 713 + ], + "type": "text", + "content": "CLIP-Aligned Labeling. Directly training the localizer on the original COCO dataset suffers from convergence issue." + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7035" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 220 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 220 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 220 + ], + "type": "text", + "content": "By anchor pre-matching mechanism, a ground truth box incorporates training only when at least one query with the same pre-matched label exists. Otherwise, it is ignored, which hinders convergence. This issue can be partially attributed to the inaccurate anchor box. However, even if a ground truth box has an accurate anchor box, it may still be ignored due to the limited recognition accuracy limitation of the region classifier, or in other words, the ground truth box label is not aligned with the CLIP region classifier used for pre-matching. Thus, we relabel the boxes in the training dataset with the region classifier, which we refer to as CLIP-Aligned labeling. With this technique, more ground truth boxes can be matched." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 232, + 128, + 245 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 232, + 128, + 245 + ], + "spans": [ + { + "bbox": [ + 47, + 232, + 128, + 245 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 252, + 288, + 344 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 252, + 288, + 344 + ], + "spans": [ + { + "bbox": [ + 46, + 252, + 288, + 344 + ], + "type": "text", + "content": "In this section, we comprehensively evaluate our CORA on the open-vocabulary detection task. Datasets and evaluation protocols are introduced in Sec. 4.1, and implementation details of our method are provided in Sec. 4.2. We compare with state-of-the-art methods in Sec. 4.3, demonstrating advantages of our CORA, and then validate the effectiveness of the proposed Region Prompting and Anchor Pre-Matching in Sec. 4.4 and Sec. 4.5, respectively." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 352, + 227, + 365 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 352, + 227, + 365 + ], + "spans": [ + { + "bbox": [ + 47, + 352, + 227, + 365 + ], + "type": "text", + "content": "4.1. Dataset & Training & Evaluation" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 371, + 287, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 371, + 287, + 529 + ], + "spans": [ + { + "bbox": [ + 46, + 371, + 287, + 529 + ], + "type": "text", + "content": "Following the convention of the COCO OVD benchmark proposed in [4], the 80 classes in the COCO dataset [24] classes are divided into 48 base classes and 17 novel classes. The model is trained on the 48 base classes, which contains 107,761 images and 665,387 instances. The model is then evaluated on the validation set of novel classes, which contains 4,836 images and 33,152 instances from both the 48 base classes and the 17 novel classes. We also conduct experiments on the LVIS v1.0 [16] dataset. On LVIS dataset, the model is trained on 461 common classes and 405 frequent classes, which contains 100,170 images and 1,264,883 instances. After training, the model is evaluated on the LVIS validation set, which contains 19,809 images and 244,707 instances." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 531, + 287, + 678 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 531, + 287, + 678 + ], + "spans": [ + { + "bbox": [ + 46, + 531, + 287, + 678 + ], + "type": "text", + "content": "To make fairer comparisons with other methods, we propose " + }, + { + "bbox": [ + 46, + 531, + 287, + 678 + ], + "type": "inline_equation", + "content": "\\mathrm{CORA^{+}}" + }, + { + "bbox": [ + 46, + 531, + 287, + 678 + ], + "type": "text", + "content": ", which utilizes extra dataset or target novel class names. " + }, + { + "bbox": [ + 46, + 531, + 287, + 678 + ], + "type": "inline_equation", + "content": "\\mathrm{CORA^{+}}" + }, + { + "bbox": [ + 46, + 531, + 287, + 678 + ], + "type": "text", + "content": " is a detector trained with both ground-truth base-category annotations and additional pseudo bounding box labels. When target class names are provided, we generate pseudo boxes of novel classes on the base training dataset using CORA, and when extra imagedataset (text could be captions or class names) is available, pseudo boxes of all objects mentioned by the corresponding text are generated. We use standard detector architecture and training target to train " + }, + { + "bbox": [ + 46, + 531, + 287, + 678 + ], + "type": "inline_equation", + "content": "\\mathrm{CORA^{+}}" + }, + { + "bbox": [ + 46, + 531, + 287, + 678 + ], + "type": "text", + "content": ". SAMDETR [38] is used for the experiment on COCO and CenterNet2 [43] is used for the experiment on LVIS." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 679, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 679, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 679, + 287, + 713 + ], + "type": "text", + "content": "In the OVD task, we evaluate our model under the \"generalized\" setting, in which the model needs to predict objects from both base and novel classes, and then evaluated" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 72, + 545, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 186 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 186 + ], + "type": "text", + "content": "on novel objects. On the COCO benchmark, we take AP50 as our evaluation metric, which counts the average precision at an intersection over union (IoU) of " + }, + { + "bbox": [ + 304, + 72, + 545, + 186 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 304, + 72, + 545, + 186 + ], + "type": "text", + "content": " for each class, and then averages among all the classes. For the LVIS OVD benchmark, we evaluate on the full validation dataset, and report the mean AP of boxes from the novel classes to compare with prior works [40]. For the region classification task, we let the model classify the ground truth boxes in the COCO dataset, the performance is evaluated in terms of mAP." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 193, + 440, + 205 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 193, + 440, + 205 + ], + "spans": [ + { + "bbox": [ + 305, + 193, + 440, + 205 + ], + "type": "text", + "content": "4.2. Implementation Details" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 210, + 545, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 210, + 545, + 437 + ], + "spans": [ + { + "bbox": [ + 304, + 210, + 545, + 437 + ], + "type": "text", + "content": "Model Specifications. We use DAB-DETR [26] as the object localizer. Specifically, localizer is configured to have 1,000 object queries, 3 encoder layers and 6 decoder layers. We use a multi-layer perceptron (MLP) with 128 hidden neurons to transform class name embeddings into an object query. Following CLIP [30], each class embedding is computed as the average text embedding of the class name over 80 context prompts. In class dropout, each class is randomly assigned to one of the two groups with equal probabilities. When adopting region prompting on LVIS, classes in the common and frequent group are sampled with equal weights, meaning that the objects in the less frequent classes are over sampled. When training our method on LVIS, we sample 100 categories (including the ground truth categories) in each iteration. Since the number of classes in the LVIS dataset is much larger than that of COCO, we relax the matching constraint in anchor pre-matching, such that ground truth boxes can be post-matched to the anchor boxes with a similar label. Specifically, classes with a cosine similarity greater than 0.7 are considered similar." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 438, + 545, + 609 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 438, + 545, + 609 + ], + "spans": [ + { + "bbox": [ + 304, + 438, + 545, + 609 + ], + "type": "text", + "content": "Training & Hyperparameters. We train the region prompts for 5 epochs with a base learning rate of " + }, + { + "bbox": [ + 304, + 438, + 545, + 609 + ], + "type": "inline_equation", + "content": "10^{-4}" + }, + { + "bbox": [ + 304, + 438, + 545, + 609 + ], + "type": "text", + "content": ", which decays after the " + }, + { + "bbox": [ + 304, + 438, + 545, + 609 + ], + "type": "inline_equation", + "content": "4^{th}" + }, + { + "bbox": [ + 304, + 438, + 545, + 609 + ], + "type": "text", + "content": " epoch by a factor of 0.1. The localizer is trained for 35 epochs with a learning rate of " + }, + { + "bbox": [ + 304, + 438, + 545, + 609 + ], + "type": "inline_equation", + "content": "10^{-4}" + }, + { + "bbox": [ + 304, + 438, + 545, + 609 + ], + "type": "text", + "content": " without learning rate decay. Both the region prompts and the localizer are trained with batch size 32 by the AdamW optimizer [28] with " + }, + { + "bbox": [ + 304, + 438, + 545, + 609 + ], + "type": "inline_equation", + "content": "10^{-4}" + }, + { + "bbox": [ + 304, + 438, + 545, + 609 + ], + "type": "text", + "content": " weight decay. We apply gradient clipping with a maximal norm of 0.1. To stabilize training, we evaluate on the exponential moving average (EMA) of the model after training. The class dropout probability " + }, + { + "bbox": [ + 304, + 438, + 545, + 609 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 304, + 438, + 545, + 609 + ], + "type": "text", + "content": " is set as 0.2. " + }, + { + "bbox": [ + 304, + 438, + 545, + 609 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{focal}}" + }, + { + "bbox": [ + 304, + 438, + 545, + 609 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 438, + 545, + 609 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{L_1}}" + }, + { + "bbox": [ + 304, + 438, + 545, + 609 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 438, + 545, + 609 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{GIOU}}" + }, + { + "bbox": [ + 304, + 438, + 545, + 609 + ], + "type": "text", + "content": " are set as 2.0, 5.0, 2.0, respectively. For the experiment on LVIS, we use repeat factor sampling [16] with default hyperparameters to balance the training samples. We use non-maximum suppression (NMS) with an IoU threshold 0.5 during inference." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 616, + 531, + 628 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 616, + 531, + 628 + ], + "spans": [ + { + "bbox": [ + 305, + 616, + 531, + 628 + ], + "type": "text", + "content": "4.3. Comparison with State-of-the-Art Methods" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 633, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 633, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 633, + 545, + 713 + ], + "type": "text", + "content": "Tab. 1 summarizes our main results. Since the pretrained model is crucial to the open-vocabulary capability of the detector, we compare our method with baseline methods that are trained with the same CLIP model. When compared with methods trained on CLIP RN50, CORA outperforms VL-PLM by 0.7 AP50 on novel classes. With a larger pre-trained model, our method improves the previous state" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7036" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 75, + 70, + 520, + 239 + ], + "blocks": [ + { + "bbox": [ + 75, + 70, + 520, + 239 + ], + "lines": [ + { + "bbox": [ + 75, + 70, + 520, + 239 + ], + "spans": [ + { + "bbox": [ + 75, + 70, + 520, + 239 + ], + "type": "table", + "html": "
MethodExtra DatasetDetector Training \nPre-train ModelRequire Novel ClassGeneralized (17 + 48)
NovelBaseAll
OVR-CNN [36]COCO Captions [9]-X22.846.039.9
Detic [42]COCO Captions [9]CLIP (text encoder)X27.847.145.0
RegionCLIP [40]CC3M [33]CLIP (RN50)X31.457.150.4
VL-PLM [39]-CLIP (RN50)✓34.460.253.5
CORA (Ours)-CLIP (RN50)X35.135.535.4
ViLD [14]-CLIP (ViT-B/32)✓27.659.951.3
OV-DETR [35]-CLIP (ViT-B/32)✓29.461.052.7
MEDet [7]COCO Captions [9]CLIP (ViT-B/32)X32.654.049.4
RegionCLIP [40]CC3M [33]CLIP (RN50x4)X39.361.655.7
CORA (Ours)-CLIP (RN50x4)X41.744.543.8
CORA+ (Ours)COCO Captions [9]CLIP (RN50x4)X43.160.956.2
", + "image_path": "eb02657d0b54306fd0a5370a1888ca26714cd5469306edd68bb85124c15d28ba.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 52, + 298, + 282, + 406 + ], + "blocks": [ + { + "bbox": [ + 46, + 247, + 547, + 281 + ], + "lines": [ + { + "bbox": [ + 46, + 247, + 547, + 281 + ], + "spans": [ + { + "bbox": [ + 46, + 247, + 547, + 281 + ], + "type": "text", + "content": "Table 1. Main results on the COCO OVD benchmark. We report AP50 as the evaluation metric. The baseline methods are grouped by their pre-trained model. We also list the extra dataset requirement of each method, and whether they require the novel class to be provided during training." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 52, + 298, + 282, + 406 + ], + "lines": [ + { + "bbox": [ + 52, + 298, + 282, + 406 + ], + "spans": [ + { + "bbox": [ + 52, + 298, + 282, + 406 + ], + "type": "table", + "html": "
MethodDetector TrainingLVIS APr
Ext. DataPre-train Model
ViLD [14]-CLIP (ViT-B/32)16.3
OV-DETR [35]-CLIP (ViT-B/32)17.4
RegionCLIP [40]CC3MCLIP (RN50x4)22.0
CORA (Ours)-CLIP (RN50x4)22.2
MEDet [7]CC3MCLIP (ViT-B/32)22.4
Detic [42]IN-21kCLIP (text encoder)26.2
CORA+ (Ours)IN-21kCLIP (text encoder)28.1
", + "image_path": "343112e54549f70889ebd375951aeb3822800bcda2e494503042d7638696ffeb.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 449, + 287, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 449, + 287, + 495 + ], + "spans": [ + { + "bbox": [ + 46, + 449, + 287, + 495 + ], + "type": "text", + "content": "of-the-art RegionCLIP [40] by 2.4 AP50. When extra data is available, the performance can be further boosted to 43.1 AP50. The results on LVIS [16] OVD benchmark is shown in Tab. 2." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 495, + 287, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 495, + 287, + 666 + ], + "spans": [ + { + "bbox": [ + 46, + 495, + 287, + 666 + ], + "type": "text", + "content": "Note that among the baseline methods, VL-PLM [39], ViLD [14] and OV-DETR [35] use novel class names during training in order to recognize the potential novel objects and assign pseudo labels for them. Consequently, a new detector needs to be trained whenever there is a new set of categories to be detected. CORA can generalize to any combination of novel categories once trained without tedious re-training. Other compared methods relying on CLIP require image tag annotations extracted from language descriptions [7,31,40] or image labels [42] during training. We argue that the extra annotations do not provide additional information over CLIP. Instead, they serve as a media to transfer the knowledge from CLIP to the detector. Our method directly adapts the CLIP model to obtain region classifier, thus no extra image-text data is needed." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 667, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 667, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 667, + 287, + 715 + ], + "type": "text", + "content": "In this work, both region prompting and anchor prematching aim to generalize the knowledge learned from base classes to novel classes. Consequently, the performance gap between novel and base classes is significantly" + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 316, + 298, + 537, + 396 + ], + "blocks": [ + { + "bbox": [ + 72, + 413, + 261, + 424 + ], + "lines": [ + { + "bbox": [ + 72, + 413, + 261, + 424 + ], + "spans": [ + { + "bbox": [ + 72, + 413, + 261, + 424 + ], + "type": "text", + "content": "Table 2. Results on the LVIS [16] OVD benchmark." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 316, + 298, + 537, + 396 + ], + "lines": [ + { + "bbox": [ + 316, + 298, + 537, + 396 + ], + "spans": [ + { + "bbox": [ + 316, + 298, + 537, + 396 + ], + "type": "table", + "html": "
MethodCLIP modelNovelBase
CLIPRN5058.258.6
CLIP-Adapter [12]RN5063.080.6
CoOp [41]RN5064.475.7
CORARN5065.170.0
CLIPRN50x463.962.7
CORARN50x474.176.0
", + "image_path": "ae4284a3c3b9606e4f54b7d79ab2b6b62ffd01d7ef793cfd40c6d38815a35908.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 403, + 545, + 436 + ], + "lines": [ + { + "bbox": [ + 305, + 403, + 545, + 436 + ], + "spans": [ + { + "bbox": [ + 305, + 403, + 545, + 436 + ], + "type": "text", + "content": "Table 3. Results on the region classification task evaluated in mAP. We compare our method with the original CLIP regional classifier and other baseline methods." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 305, + 457, + 545, + 505 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 457, + 545, + 505 + ], + "spans": [ + { + "bbox": [ + 305, + 457, + 545, + 505 + ], + "type": "text", + "content": "lowered than the compared methods. Note that in OVD, the performance is evaluated by the generalization capability towards novel classes, rather than the bases classes on which they are trained." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 512, + 491, + 525 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 512, + 491, + 525 + ], + "spans": [ + { + "bbox": [ + 305, + 512, + 491, + 525 + ], + "type": "text", + "content": "4.4. Effectiveness of Region Prompting" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 530, + 545, + 610 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 530, + 545, + 610 + ], + "spans": [ + { + "bbox": [ + 304, + 530, + 545, + 610 + ], + "type": "text", + "content": "Region Prompting. Since object classification is decoupled from localization in this work, the CLIP-based region classifier can be directly evaluated by the region classification task. After trained on the base-class annotations of COCO dataset, we evaluate the CLIP-based region classifier on both base and novel classes in the validation set. We use mean average precision (mAP) as our evaluation metric." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 611, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 611, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 611, + 545, + 714 + ], + "type": "text", + "content": "Tab. 3 shows our main result. Directly evaluating on the CLIP model without further training already achieves considerable performance of " + }, + { + "bbox": [ + 304, + 611, + 545, + 714 + ], + "type": "inline_equation", + "content": "58.2\\mathrm{mAP}" + }, + { + "bbox": [ + 304, + 611, + 545, + 714 + ], + "type": "text", + "content": " on the novel classes. We compare our result with two competitive methods from the adapter and prompt tuning literatures. CLIP-Adapter [12] adopts an additional bottleneck layer to learn new features and performs residual style feature blending with the original pre-trained features. CoOp [41] prepends shared learnable prompts to the text embeddings before en" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7037" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 70, + 70, + 265, + 118 + ], + "blocks": [ + { + "bbox": [ + 70, + 70, + 265, + 118 + ], + "lines": [ + { + "bbox": [ + 70, + 70, + 265, + 118 + ], + "spans": [ + { + "bbox": [ + 70, + 70, + 265, + 118 + ], + "type": "table", + "html": "
CLIP modelFeatureNovelBase
RN50whole-image43.840.5
RN50regional58.658.2
", + "image_path": "49bde419eb84b97d7b60b73215d3a45690acca51aa890599dc13192a4173edb7.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 52, + 169, + 282, + 247 + ], + "blocks": [ + { + "bbox": [ + 47, + 125, + 287, + 156 + ], + "lines": [ + { + "bbox": [ + 47, + 125, + 287, + 156 + ], + "spans": [ + { + "bbox": [ + 47, + 125, + 287, + 156 + ], + "type": "text", + "content": "Table 4. Comparison of the whole-image classification pipeline and the regional classification pipeline. Results are reported in mAP." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 52, + 169, + 282, + 247 + ], + "lines": [ + { + "bbox": [ + 52, + 169, + 282, + 247 + ], + "spans": [ + { + "bbox": [ + 52, + 169, + 282, + 247 + ], + "type": "table", + "html": "
Query ConditioningPer-Class Post-MatchingCLIP-Aligned LabelingNovel
XX✓26.0
✓X✓29.9
✓✓X40.9
✓✓✓41.7
", + "image_path": "3fde7f0a72ade9a5c93754a8870e94aaa2daa5f867fdd7e6188d62fd7aef9b43.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 255, + 286, + 287 + ], + "lines": [ + { + "bbox": [ + 47, + 255, + 286, + 287 + ], + "spans": [ + { + "bbox": [ + 47, + 255, + 286, + 287 + ], + "type": "text", + "content": "Table 5. Ablation studies on anchor pre-matching and CLIP-aligned labeling. Anchor pre-matching consists of query conditioning and per-class post-matching." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 308, + 286, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 308, + 286, + 445 + ], + "spans": [ + { + "bbox": [ + 46, + 308, + 286, + 445 + ], + "type": "text", + "content": "coded by the CLIP text encoder. Experiments show that the compared methods bias strongly towards base classes. Note that the baseline methods adapts the CLIP model by tuning the text input or output feature, which are irrelevant or distant to the regional feature, where the mismatch between regional and whole-image feature occurs. On the contrary, region prompting directly prompts the mismatched features, thus generalizes better to the novel classes, achieving a performance gain of " + }, + { + "bbox": [ + 46, + 308, + 286, + 445 + ], + "type": "inline_equation", + "content": "6.9\\mathrm{mAP}" + }, + { + "bbox": [ + 46, + 308, + 286, + 445 + ], + "type": "text", + "content": " over CLIP on the novel classes. Region prompting also scales with larger backbones. On the " + }, + { + "bbox": [ + 46, + 308, + 286, + 445 + ], + "type": "inline_equation", + "content": "\\mathrm{RN50x4}" + }, + { + "bbox": [ + 46, + 308, + 286, + 445 + ], + "type": "text", + "content": " CLIP backbone, region prompting further boosts CLIP by " + }, + { + "bbox": [ + 46, + 308, + 286, + 445 + ], + "type": "inline_equation", + "content": "10.2\\%" + }, + { + "bbox": [ + 46, + 308, + 286, + 445 + ], + "type": "text", + "content": " mAP over the corresponding CLIP model." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 446, + 286, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 446, + 286, + 548 + ], + "spans": [ + { + "bbox": [ + 46, + 446, + 286, + 548 + ], + "type": "text", + "content": "Comparison with whole-image classification. A common practice of region classification by CLIP is to crop the regions and classify them as separate images. We compare the region classifier with the common practice on the original CLIP weights without region prompting. As shown in Tab. 4, when classifying a region as a whole-image, the performance is significantly lower than using the regional feature, despite the extra computation. We attribute the performance gap to the missing context of cropped images." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 558, + 248, + 571 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 558, + 248, + 571 + ], + "spans": [ + { + "bbox": [ + 47, + 558, + 248, + 571 + ], + "type": "text", + "content": "4.5. Effectiveness of Anchor Pre-Matching" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 576, + 286, + 598 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 576, + 286, + 598 + ], + "spans": [ + { + "bbox": [ + 47, + 576, + 286, + 598 + ], + "type": "text", + "content": "We conduct ablation studies to validate anchor pre-matching and the proposed training techniques." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 599, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 599, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 599, + 286, + 713 + ], + "type": "text", + "content": "Anchor Pre-Matching. Anchor pre-matching consists of two separate operations: query conditioning and per-class post-matching. We analyze their effects in Tab. 5. Firstly, we train a model without anchor pre-matching. We keep using DAB-DETR [26] as the localizer, and use the same number of object queries and anchor boxes for fair comparison. The object queries are not pre-matched to classes, and are initialized as 0 as done in DAB-DETR. Since the model predictions are not pre-matched, the per-class post-matching is replaced by the vanilla one-to-one matching" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 323, + 89, + 525, + 233 + ], + "blocks": [ + { + "bbox": [ + 323, + 89, + 525, + 233 + ], + "lines": [ + { + "bbox": [ + 323, + 89, + 525, + 233 + ], + "spans": [ + { + "bbox": [ + 323, + 89, + 525, + 233 + ], + "type": "image", + "image_path": "d07fdb4c783b7b97ee3fc4edcd3da4d29bc26fa0f53d2efa680078d996c5d90e.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 334, + 241, + 515, + 252 + ], + "lines": [ + { + "bbox": [ + 334, + 241, + 515, + 252 + ], + "spans": [ + { + "bbox": [ + 334, + 241, + 515, + 252 + ], + "type": "text", + "content": "Figure 4. Different choices of " + }, + { + "bbox": [ + 334, + 241, + 515, + 252 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 334, + 241, + 515, + 252 + ], + "type": "text", + "content": " for class dropout." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 274, + 545, + 341 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 274, + 545, + 341 + ], + "spans": [ + { + "bbox": [ + 304, + 274, + 545, + 341 + ], + "type": "text", + "content": "mechanism in DETR. Then, we classify the anchor boxes and condition the object queries on the class name embeddings, but do not set constraint on the post-matching, which boosts the performance on novel classes by 3.9 AP50. After adopting the full anchor pre-matching, the performance is significantly boosted by 10.8 AP50." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 342, + 545, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 342, + 545, + 422 + ], + "spans": [ + { + "bbox": [ + 304, + 342, + 545, + 422 + ], + "type": "text", + "content": "Training Techniques. Class dropout and CLIP-aligned labeling are two training techniques that help the model generalize better. In Fig. 4, we examine the effect of different dropout probability. We find that the models trained with class dropout consistently outperform the baseline, and " + }, + { + "bbox": [ + 304, + 342, + 545, + 422 + ], + "type": "inline_equation", + "content": "p = 0.2" + }, + { + "bbox": [ + 304, + 342, + 545, + 422 + ], + "type": "text", + "content": " gives the best performance. In Tab. 5, we validate the effectiveness of CLIP-Aligned Labeling." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 432, + 378, + 444 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 432, + 378, + 444 + ], + "spans": [ + { + "bbox": [ + 306, + 432, + 378, + 444 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 452, + 545, + 622 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 452, + 545, + 622 + ], + "spans": [ + { + "bbox": [ + 304, + 452, + 545, + 622 + ], + "type": "text", + "content": "The core challenge in open-vocabulary detection is how to effectively transfer the knowledge learned from the base classes to the unseen novel classes for evaluation. In this work, we directly adapt the CLIP into a region classifier, and mitigate the distribution gap between whole-image feature and regional feature through region prompting, which successfully generalizes to the novel categories. Different from prior works that rely on a fixed RPN for novel class localization, we achieve efficient class-aware localization by the proposed anchor pre-matching mechanism. Experiments show that our method can better transfer the knowledge from base classes to unseen novel classes with a smaller gap than prior works. We hope our work can help other researchers gain better insight on the OVD problem and develop better open-vocabulary detectors." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 622, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 622, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 622, + 545, + 713 + ], + "type": "text", + "content": "Acknowledgement. This project is funded in part by National Key R&D Program of China Project 2022ZD0161100, by the Centre for Perceptual and Interactive Intelligence (CPII) Ltd under the Innovation and Technology Commission (ITC)'s InnoHK, by General Research Fund of Hong Kong RGC Project 14204021. Hongsheng Li is a PI of CPII under the InnoHK. This project is also supported by SenseTime Collaborative Research Grant." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7038" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 53, + 91, + 258, + 100 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 258, + 100 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 258, + 100 + ], + "type": "text", + "content": "[1] Deformable detr: Deformable transformers. 2020. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 102, + 287, + 133 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 102, + 287, + 133 + ], + "spans": [ + { + "bbox": [ + 53, + 102, + 287, + 133 + ], + "type": "text", + "content": "[2] Hyojin Bahng, Ali Jahanian, Swami Sankaranarayanan, and Phillip Isola. Exploring visual prompts for adapting large-scale models. arXiv preprint arXiv:2203.17274, 2022. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 134, + 287, + 165 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 134, + 287, + 165 + ], + "spans": [ + { + "bbox": [ + 53, + 134, + 287, + 165 + ], + "type": "text", + "content": "[3] Hyojin Bahng, Ali Jahanian, Swami Sankaranarayanan, and Phillip Isola. Visual prompting: Modifying pixel space to adapt pre-trained models. ArXiv, abs/2203.17274, 2022. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 166, + 287, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 166, + 287, + 196 + ], + "spans": [ + { + "bbox": [ + 53, + 166, + 287, + 196 + ], + "type": "text", + "content": "[4] Ankan Bansal, Karan Sikka, Gaurav Sharma, Rama Chellappa, and Ajay Divakaran. Zero-shot object detection. ArXiv, abs/1804.04340, 2018. 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 198, + 287, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 198, + 287, + 300 + ], + "spans": [ + { + "bbox": [ + 53, + 198, + 287, + 300 + ], + "type": "text", + "content": "[5] Tom B. Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert-Voss, Gretchen Krueger, T. J. Henighan, Rewon Child, Aditya Ramesh, Daniel M. Ziegler, Jeff Wu, Clemens Winter, Christopher Hesse, Mark Chen, Eric Sigler, Mateusz Litwin, Scott Gray, Benjamin Chess, Jack Clark, Christopher Berner, Sam McCandlish, Alec Radford, Ilya Sutskever, and Dario Amodei. Language models are few-shot learners. ArXiv, abs/2005.14165, 2020. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 302, + 287, + 343 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 302, + 287, + 343 + ], + "spans": [ + { + "bbox": [ + 53, + 302, + 287, + 343 + ], + "type": "text", + "content": "[6] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. ArXiv, abs/2005.12872, 2020. 1, 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 345, + 287, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 345, + 287, + 386 + ], + "spans": [ + { + "bbox": [ + 53, + 345, + 287, + 386 + ], + "type": "text", + "content": "[7] Peixian Chen, Kekai Sheng, Mengdan Zhang, Yunhang Shen, Ke Li, and Chunhua Shen. Open vocabulary object detection with proposal mining and prediction equalization. CoRR, abs/2206.11134, 2022. 1, 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 387, + 287, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 387, + 287, + 426 + ], + "spans": [ + { + "bbox": [ + 53, + 387, + 287, + 426 + ], + "type": "text", + "content": "[8] Qiang Chen, Xiaokang Chen, Gang Zeng, and Jingdong Wang. Group detr: Fast training convergence with decoupled one-to-many label assignment. ArXiv, abs/2207.13085, 2022. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 429, + 287, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 429, + 287, + 470 + ], + "spans": [ + { + "bbox": [ + 53, + 429, + 287, + 470 + ], + "type": "text", + "content": "[9] Xinlei Chen, Hao Fang, Tsung-Yi Lin, Ramakrishna Vedantam, Saurabh Gupta, Piotr Dólar, and C. Lawrence Zitnick. Microsoft coco captions: Data collection and evaluation server. ArXiv, abs/1504.00325, 2015. 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 471, + 287, + 512 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 471, + 287, + 512 + ], + "spans": [ + { + "bbox": [ + 48, + 471, + 287, + 512 + ], + "type": "text", + "content": "[10] Chengjian Feng, Yujie Zhong, Zequn Jie, Xiangxiang Chu, Haibing Ren, Xiaolin Wei, Weidi Xie, and Lin Ma. Promptdet: Towards open-vocabulary detection using uncurated images. 2022. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 514, + 287, + 554 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 514, + 287, + 554 + ], + "spans": [ + { + "bbox": [ + 48, + 514, + 287, + 554 + ], + "type": "text", + "content": "[11] Mingfei Gao, Chen Xing, Juan Carlos Niebles, Junnan Li, Ran Xu, Wenhao Liu, and Caiming Xiong. Open vocabulary object detection with pseudo bounding-box labels. arXiv preprint arXiv:2111.09452, 2021. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 555, + 287, + 596 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 555, + 287, + 596 + ], + "spans": [ + { + "bbox": [ + 48, + 555, + 287, + 596 + ], + "type": "text", + "content": "[12] Peng Gao, Shijie Geng, Renrui Zhang, Teli Ma, Rongyao Fang, Yongfeng Zhang, Hongsheng Li, and Yu Jiao Qiao. Clip-adapter: Better vision-language models with feature adapters. ArXiv, abs/2110.04544, 2021. 7" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 597, + 287, + 648 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 597, + 287, + 648 + ], + "spans": [ + { + "bbox": [ + 48, + 597, + 287, + 648 + ], + "type": "text", + "content": "[13] Peng Gao, Minghang Zheng, Xiaogang Wang, Jifeng Dai, and Hongsheng Li. Fast convergence of detr with spatially modulated co-attention. 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 3601–3610, 2021. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 650, + 287, + 681 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 650, + 287, + 681 + ], + "spans": [ + { + "bbox": [ + 48, + 650, + 287, + 681 + ], + "type": "text", + "content": "[14] Xiuye Gu, Tsung-Yi Lin, Weicheng Kuo, and Yin Cui. Open-vocabulary object detection via vision and language knowledge distillation. In ICLR, 2022. 1, 7" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 682, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 682, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 682, + 287, + 713 + ], + "type": "text", + "content": "[15] Xiuye Gu, Tsung-Yi Lin, Weicheng Kuo, and Yin Cui. Open-vocabulary object detection via vision and language knowledge distillation. In ICLR, 2022. 2" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 32, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 115 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 115 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 115 + ], + "type": "text", + "content": "[16] Agrim Gupta, Piotr Dollar, and Ross Girshick. LVIS: A dataset for large vocabulary instance segmentation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2019. 2, 6, 7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 117, + 545, + 148 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 117, + 545, + 148 + ], + "spans": [ + { + "bbox": [ + 307, + 117, + 545, + 148 + ], + "type": "text", + "content": "[17] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In 2017 IEEE International Conference on Computer Vision (ICCV), pages 2980-2988, 2017. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 148, + 545, + 189 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 148, + 545, + 189 + ], + "spans": [ + { + "bbox": [ + 308, + 148, + 545, + 189 + ], + "type": "text", + "content": "[18] Ding Jia, Yuhui Yuan, Haodi He, Xiao pei Wu, Haojun Yu, Weihong Lin, Lei huan Sun, Chao Zhang, and Hanhua Hu. Detrs with hybrid matching. ArXiv, abs/2207.13080, 2022. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 191, + 545, + 222 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 191, + 545, + 222 + ], + "spans": [ + { + "bbox": [ + 308, + 191, + 545, + 222 + ], + "type": "text", + "content": "[19] Menglin Jia, Luming Tang, Bor-Chun Chen, Claire Cardie, Serge J. Belongie, Bharath Hariharan, and Ser Nam Lim. Visual prompt tuning. ArXiv, abs/2203.12119, 2022. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 224, + 545, + 255 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 224, + 545, + 255 + ], + "spans": [ + { + "bbox": [ + 308, + 224, + 545, + 255 + ], + "type": "text", + "content": "[20] Chen Ju, Tengda Han, Kunhao Zheng, Ya Zhang, and Weidi Xie. Prompting visual-language models for efficient video understanding. ArXiv, abs/2112.04478, 2021. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 256, + 545, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 256, + 545, + 319 + ], + "spans": [ + { + "bbox": [ + 308, + 256, + 545, + 319 + ], + "type": "text", + "content": "[21] Liunian Harold Li, Pengchuan Zhang, Haotian Zhang, Jianwei Yang, Chunyuan Li, Yiwu Zhong, Lijuan Wang, Lu Yuan, Lei Zhang, Jenq-Neng Hwang, Kai-Wei Chang, and Jianfeng Gao. Grounded language-image pre-training. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10955–10965, 2022. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 319, + 545, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 319, + 545, + 381 + ], + "spans": [ + { + "bbox": [ + 308, + 319, + 545, + 381 + ], + "type": "text", + "content": "[22] Xiang Lisa Li and Percy Liang. Prefix-tuning: Optimizing continuous prompts for generation. Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), abs/2101.00190, 2021. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 383, + 545, + 424 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 383, + 545, + 424 + ], + "spans": [ + { + "bbox": [ + 308, + 383, + 545, + 424 + ], + "type": "text", + "content": "[23] Tsung-Yi Lin, Priya Goyal, Ross B. Girshick, Kaiming He, and Piotr Dólar. Focal loss for dense object detection. 2017 IEEE International Conference on Computer Vision (ICCV), pages 2999-3007, 2017. 5" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 426, + 545, + 467 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 426, + 545, + 467 + ], + "spans": [ + { + "bbox": [ + 308, + 426, + 545, + 467 + ], + "type": "text", + "content": "[24] Tsung-Yi Lin, Michael Maire, Serge J. Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C. Lawrence Zitnick. Microsoft coco: Common objects in context. In ECCV, 2014. 2, 6" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 468, + 545, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 468, + 545, + 510 + ], + "spans": [ + { + "bbox": [ + 308, + 468, + 545, + 510 + ], + "type": "text", + "content": "[25] Ziyi Lin, Shijie Geng, Renrui Zhang, Peng Gao, Gerard de Melo, Xiaogang Wang, Jifeng Dai, Yu Qiao, and Hongsheng Li. Frozen clip models are efficient video learners. arXiv preprint arXiv:2208.03550, 2022. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 511, + 545, + 552 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 511, + 545, + 552 + ], + "spans": [ + { + "bbox": [ + 308, + 511, + 545, + 552 + ], + "type": "text", + "content": "[26] Shilong Liu, Feng Li, Hao Zhang, Xiao Bin Yang, Xianbiao Qi, Hang Su, Jun Zhu, and Lei Zhang. Dab-detr: Dynamic anchor boxes are better queries for detr. ArXiv, abs/2201.12329, 2022. 2, 5, 6, 8" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 554, + 545, + 594 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 554, + 545, + 594 + ], + "spans": [ + { + "bbox": [ + 308, + 554, + 545, + 594 + ], + "type": "text", + "content": "[27] Xiao Liu, Kaixuan Ji, Yicheng Fu, Zhengxiao Du, Zhilin Yang, and Jie Tang. P-tuning v2: Prompt tuning can be comparable to fine-tuning universally across scales and tasks. CoRR, abs/2110.07602, 2021. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 308, + 597, + 545, + 616 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 597, + 545, + 616 + ], + "spans": [ + { + "bbox": [ + 308, + 597, + 545, + 616 + ], + "type": "text", + "content": "[28] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In ICLR, 2019. 6" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 308, + 618, + 545, + 670 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 618, + 545, + 670 + ], + "spans": [ + { + "bbox": [ + 308, + 618, + 545, + 670 + ], + "type": "text", + "content": "[29] Muhammad Maaz, Hanoona Rasheed, Salman Khan, Fahad Shahbaz Khan, Rao Muhammad Anwer, and Ming-Hsuan Yang. Class-agnostic object detection with multimodal transformer. In 17th European Conference on Computer Vision (ECCV). Springer, 2022. 1" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 308, + 672, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 672, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 672, + 545, + 713 + ], + "type": "text", + "content": "[30] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual" + } + ] + } + ], + "index": 31 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7039" + } + ] + } + ], + "index": 33 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 680 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 68, + 72, + 286, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 72, + 286, + 94 + ], + "spans": [ + { + "bbox": [ + 68, + 72, + 286, + 94 + ], + "type": "text", + "content": "models from natural language supervision. In ICML, 2021. 1, 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 95, + 287, + 146 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 95, + 287, + 146 + ], + "spans": [ + { + "bbox": [ + 48, + 95, + 287, + 146 + ], + "type": "text", + "content": "[31] Hanoona Rasheed, Muhammad Maaz, Muhammad Uzair Khattak, Salman Khan, and Fahad Shahbaz Khan. Bridging the gap between object and image-level representations for open-vocabulary detection. ArXiv, abs/2207.03482, 2022. 1, 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 148, + 287, + 190 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 148, + 287, + 190 + ], + "spans": [ + { + "bbox": [ + 48, + 148, + 287, + 190 + ], + "type": "text", + "content": "[32] Shaoqing Ren, Kaiming He, Ross B. Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. IEEE Transactions on Pattern Analysis and Machine Intelligence, 39:1137-1149, 2015. 4" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 191, + 287, + 232 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 191, + 287, + 232 + ], + "spans": [ + { + "bbox": [ + 48, + 191, + 287, + 232 + ], + "type": "text", + "content": "[33] Piyush Sharma, Nan Ding, Sebastian Goodman, and Radu Soricut. Conceptual captions: A cleaned, hypernymed, image alt-text dataset for automatic image captioning. In ACL, 2018. 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 234, + 287, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 234, + 287, + 274 + ], + "spans": [ + { + "bbox": [ + 48, + 234, + 287, + 274 + ], + "type": "text", + "content": "[34] Yinghui Xing, Qirui Wu, De Cheng, Shizhou Zhang, Guoqiang Liang, and Yanning Zhang. Class-aware visual prompt tuning for vision-language pre-trained model. CoRR, abs/2208.08340, 2022. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 276, + 287, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 276, + 287, + 308 + ], + "spans": [ + { + "bbox": [ + 48, + 276, + 287, + 308 + ], + "type": "text", + "content": "[35] Yuhang Zang, Wei Li, Kaiyang Zhou, Chen Huang, and Chen Change Loy. Open-vocabulary detr with conditional matching. ArXiv, abs/2203.11876, 2022. 1, 2, 5, 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 309, + 287, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 309, + 287, + 360 + ], + "spans": [ + { + "bbox": [ + 48, + 309, + 287, + 360 + ], + "type": "text", + "content": "[36] Alireza Zareian, Kevin Dela Rosa, Derek Hao Hu, and Shih-Fu Chang. Open-vocabulary object detection using captions. In 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 14388-14397, 2021. 1, 2, 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 361, + 287, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 361, + 287, + 413 + ], + "spans": [ + { + "bbox": [ + 48, + 361, + 287, + 413 + ], + "type": "text", + "content": "[37] Hongyu Zhai, Jian Cheng, and Mengyong Wang. Rethink the iou-based loss functions for bounding box regression. 2020 IEEE 9th Joint International Information Technology and Artificial Intelligence Conference (ITAIC), 9:1522-1528, 2020. 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 415, + 287, + 467 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 415, + 287, + 467 + ], + "spans": [ + { + "bbox": [ + 48, + 415, + 287, + 467 + ], + "type": "text", + "content": "[38] Gongjie Zhang, Zhipeng Luo, Yingchen Yu, Kaiwen Cui, and Shijian Lu. Accelerating DETR convergence via semantic-aligned matching. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 949-958, 2022. 2, 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 468, + 287, + 519 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 468, + 287, + 519 + ], + "spans": [ + { + "bbox": [ + 48, + 468, + 287, + 519 + ], + "type": "text", + "content": "[39] Shiyu Zhao, Zhixing Zhang, Samuel Schulter, Long Zhao, Vijay Kumar BG, Anastasis Stathopoulos, Manmohan Chandraker, and Dimitris N. Metaxas. Exploiting unlabeled data with vision and language models for object detection. CoRR, abs/2207.08954, 2022. 1, 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 521, + 287, + 582 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 521, + 287, + 582 + ], + "spans": [ + { + "bbox": [ + 48, + 521, + 287, + 582 + ], + "type": "text", + "content": "[40] Yiwu Zhong, Jianwei Yang, Pengchuan Zhang, Chengkun Li, Noel C. F. Codella, Liunian Harold Li, Luowei Zhou, Xiyang Dai, Lu Yuan, Yin Li, and Jianfeng Gao. Regionclip: Region-based language-image pretraining. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 16772-16782, 2022. 1, 2, 6, 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 584, + 287, + 615 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 584, + 287, + 615 + ], + "spans": [ + { + "bbox": [ + 48, + 584, + 287, + 615 + ], + "type": "text", + "content": "[41] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Learning to prompt for vision-language models. Int. J. Comput. Vis., 130:2337-2348, 2022. 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 616, + 287, + 647 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 616, + 287, + 647 + ], + "spans": [ + { + "bbox": [ + 48, + 616, + 287, + 647 + ], + "type": "text", + "content": "[42] Xingyi Zhou, Rohit Girdhar, Armand Joulin, Philipp Krahenbihl, and Ishan Misra. Detecting twenty-thousand classes using image-level supervision. In ECCV, 2022. 7" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 648, + 287, + 680 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 648, + 287, + 680 + ], + "spans": [ + { + "bbox": [ + 48, + 648, + 287, + 680 + ], + "type": "text", + "content": "[43] Xingyi Zhou, Vladlen Koltun, and Philipp Krahenbuhl. Probabilistic two-stage detection. In arXiv preprint arXiv:2103.07461, 2021. 6" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "7040" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/COT_ Unsupervised Domain Adaptation With Clustering and Optimal Transport/53ca0802-c684-4d48-b297-7731efdaab53_content_list.json b/2023/COT_ Unsupervised Domain Adaptation With Clustering and Optimal Transport/53ca0802-c684-4d48-b297-7731efdaab53_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..3b23e7bcd2443a4573bebdca68603d32d634c1bf --- /dev/null +++ b/2023/COT_ Unsupervised Domain Adaptation With Clustering and Optimal Transport/53ca0802-c684-4d48-b297-7731efdaab53_content_list.json @@ -0,0 +1,1743 @@ +[ + { + "type": "text", + "text": "COT: Unsupervised Domain Adaptation with Clustering and Optimal Transport", + "text_level": 1, + "bbox": [ + 80, + 128, + 890, + 154 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yang Liu $^{1*}$ Zhipeng Zhou $^{1*}$ Baigui Sun $^{1\\dagger}$ 1Alibaba Group", + "bbox": [ + 282, + 178, + 679, + 218 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 250, + 313, + 268 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Unsupervised domain adaptation (UDA) aims to transfer the knowledge from a labeled source domain to an unlabeled target domain. Typically, to guarantee desirable knowledge transfer, aligning the distribution between source and target domain from a global perspective is widely adopted in UDA. Recent researchers further point out the importance of local-level alignment and propose to construct instance-pair alignment by leveraging on Optimal Transport (OT) theory. However, existing OT-based UDA approaches are limited to handling class imbalance challenges and introduce a heavy computation overhead when considering a large-scale training situation. To cope with two aforementioned issues, we propose a Clustering-based Optimal Transport (COT) algorithm, which formulates the alignment procedure as an Optimal Transport problem and constructs a mapping between clustering centers in the source and target domain via an end-to-end manner. With this alignment on clustering centers, our COT eliminates the negative effect caused by class imbalance and reduces the computation cost simultaneously. Empirically, our COT achieves state-of-the-art performance on several authoritative benchmark datasets.", + "bbox": [ + 75, + 282, + 473, + 616 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 661, + 209, + 678 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Benefiting from the availability of large-scale data, deep learning has achieved tremendous success over the past few years. However, directly applying a well-trained convolution neural network on a new domain frequently suffers from the domain gap/discrepancy challenge, resulting in spurious predictions on the new domain. To remedy this, Unsupervised Domain Adaptation (UDA) has attracted many researchers' attention, which can transfer the knowledge from a labeled domain to an unlabeled domain.", + "bbox": [ + 75, + 686, + 468, + 823 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "A major line of UDA approaches [1, 1, 28, 42, 49, 53] aim", + "bbox": [ + 96, + 824, + 468, + 840 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "to learn a global domain shift by aligning the global source and target distribution while ignoring the local-level alignment between two domains. By leveraging on global domain adaptation, the global distributions of source and target domain are almost the same, thus losing the fine-grained information for each class (class-structure) on the source and target domain.", + "bbox": [ + 500, + 252, + 893, + 358 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recently, to preserve class structure in both domains, several works [6, 15, 23, 30, 38, 40, 44, 51, 54] adopt optimal transport (OT) to minimize the sample-level transportation cost between source and target domain, achieving a significant performance on UDA. However, there exist two issues on recent OT-based UDA approaches. (i) When considering a realistic situation, i.e. the class imbalance $^{1}$ phenomenon occurs between the source and target domain, samples belonging to the same class in the target domain are assigned with false pseudo labels due to the mechanism of optimal transport, which requires each sample in source domain can be mapped to target samples. As a result, existing OT-based UDA methods provide poor pair-wise matching when facing class imbalance challenges. (ii) OT-based UDA methods tend to find a sample-level optimal counterpart, which requires a large amount of computation overhead, especially training on large-scale datasets.", + "bbox": [ + 496, + 359, + 893, + 616 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To solve two aforementioned issues, we propose a Clustering-based Optimal Transport algorithm, termed COT, to construct a clustering-level instead of sample-level mapping between source and target domain. Clusters in the source domain are obtained from the classifiers supervised by the labeled source domain data. While for the target domain, COT utilizes a set of learnable clusters to represent the feature distribution of the target domain, which can describe the sub-domain information [50, 57]. For instance, in many object recognition tasks [13, 20] an object could contain many attributes. Each attribute can be viewed as a sub-domain. To this end, the clusters on the source and target domain can represent the individual sub-domain information, respectively, such that optimal transport between clusters intrinsically provides a local mapping from the subdomain in the source domain to those in the target domain. Moreover, we provide a theoretical analysis and compre", + "bbox": [ + 496, + 617, + 895, + 875 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "1 label distribution are different in two domains, $P_{s}(y)\\neq P_{t}(y)$", + "bbox": [ + 517, + 886, + 864, + 901 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "Email: ly261666@alibaba-inc.com", + "bbox": [ + 76, + 851, + 290, + 863 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Equal Contribution", + "bbox": [ + 78, + 864, + 191, + 875 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "† Corresponding Author", + "bbox": [ + 78, + 875, + 210, + 888 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "19998", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "hensive experimental results to guarantee that (i) COT can alleviate the negative effect caused by class imbalance; (ii) Compared to existing OT-based UDA approaches, our COT economizes much computation head.", + "bbox": [ + 75, + 90, + 468, + 150 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In summary, our main contributions include:", + "bbox": [ + 96, + 152, + 392, + 166 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose a novel Clustering-based Optimal Transport module as well as a specially designed loss derived from the discrete type of Kantorovich dual form, which resolves two aforementioned challenges on the existing OT-based UDA algorithms, facilitating the development of OT-based UDA community.", + "- We provide a theoretical analysis to guarantee the advantages of our COT.", + "- Our COT achieves state-of-the-art performance on several UDA benchmark datasets." + ], + "bbox": [ + 94, + 176, + 467, + 345 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related work", + "text_level": 1, + "bbox": [ + 76, + 361, + 215, + 376 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Pseudo Label based Domain Adaptation Inspired by the observation that samples in the target domain can be clustered within the feature space, for accurate pseudolabeling, [48] propose a selective pseudo-labeling strategy based on structural predictions which utilize the unsupervised clustering analysis. [41] introduce a confidence-based weighting scheme for obtaining pseudo-labels and an adaptive threshold adjustment strategy to provide sufficient and accurate pseudo-labels during the training process. The confidence-based weighting scheme generates pseudo-labels that can enable the performance less sensitive to the threshold determined by the pseudo-labels. In the task of person re-identification, [17] propose an unsupervised framework called Mutual Mean-Teaching to learn better features from the target domain by refining the hard pseudo labels offline and soft pseudo labels online alternatively to mitigate the effects of noisy pseudo labels caused by the clustering algorithms.", + "bbox": [ + 75, + 386, + 468, + 657 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Optimal Transport based Domain Adaptation As a way to find a minimal effort strategy to the transport of a given mass of dirt into a given hole, [37] put forward the optimal transport problem for the first time. [22] provide an extension of the original problem of Monge. Recently, by applying the optimal transport to domain adaptation, some new computation strategies have emerged.", + "bbox": [ + 75, + 659, + 467, + 763 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "[7] propose regularized unsupervised optimal transport model to align the representation of features between different domains. The regularization schemes encoding class-structure in source domain during estimation of transport map enforce the intuition that samples of the same class must undergo a similar transformation. [5] minimize the optimal transport loss between the joint source distribution and the estimated target joint distribution depending on a function that is introduced to predict an output value given", + "bbox": [ + 75, + 763, + 467, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "input from the source domain. For reducing the discrepancy between multiple domains, [40] propose Joint Class Proportion and Optimal Transport which performs multisource domain adaptation and target shift correction simultaneously by learning the predicted class probability of the unlabeled target data and the coupling to align the distributions between source and target domain. For better alignment between different domains, a relation between target error and the magnitude of different Wasserstein distances are proposed in [23] which optimize the metric for domain adaptation.", + "bbox": [ + 496, + 90, + 890, + 257 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3. Preliminary", + "text_level": 1, + "bbox": [ + 500, + 268, + 625, + 286 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this section, we introduce the basic knowledge for optimal transport.", + "bbox": [ + 498, + 295, + 890, + 325 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3.1. Optimal Transport", + "text_level": 1, + "bbox": [ + 500, + 334, + 681, + 349 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Let $X \\subseteq \\mathbb{R}^d$ be a measurable space and the labels are denoted as $\\mathcal{V}$ . We denote the set of all probability distributions on $X$ as $\\mathcal{P}(X)$ . The source and target domains are space $X$ equipped with two distinct probability distributions $\\mu_{\\mathcal{S}}$ and $\\mu_{\\mathcal{T}}$ . Suppose we have source dataset $\\{x_i^s\\}_{i=1}^{n_s} \\subset X_{\\mathcal{S}} = (X, \\mu_{\\mathcal{S}})$ associated with label set $\\{y_i^s\\}_{i=1}^{n_s}$ where $y_i^s \\in \\mathcal{V}$ . The target dataset is $\\{x_j^t\\}_{j=1}^{n_t} \\subset X_{\\mathcal{T}} = (X, \\mu_{\\mathcal{T}})$ without labels. The goal of optimal transport is to minimize the inter-domain transportation cost by finding a feasible map to preserve measure.", + "bbox": [ + 496, + 357, + 890, + 508 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Definition 1 (Kantorovich) For given joint distribution $\\rho(x^s, x^t)$ which satisfies for every measurable Borel set $O_S \\subset X_S$ , $O_T \\subset X_T$ , we have", + "bbox": [ + 498, + 516, + 890, + 561 + ], + "page_idx": 1 + }, + { + "type": "equation", + "text": "\n$$\n\\rho \\left(O _ {\\mathcal {S}} \\times X _ {\\mathcal {T}}\\right) = \\mu_ {\\mathcal {S}} \\left(O _ {\\mathcal {S}}\\right), \\rho \\left(X _ {\\mathcal {S}} \\times O _ {\\mathcal {T}}\\right) = \\mu_ {\\mathcal {T}} \\left(O _ {\\mathcal {T}}\\right) \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 514, + 571, + 890, + 589 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "For convenience, we denote the projection maps from $X_{\\mathcal{S}} \\times X_{\\mathcal{T}}$ to $X_{\\mathcal{S}}$ and $X_{\\mathcal{T}}$ as $\\pi_{\\mathcal{S}}, \\pi_{\\mathcal{T}}$ . The above equation can be denoted as $\\pi_{\\mathcal{S}\\#} \\rho = \\mu_{\\mathcal{S}}$ and $\\pi_{\\mathcal{T}\\#} \\rho = \\mu_{\\mathcal{T}}$ . The corresponding transportation cost is", + "bbox": [ + 498, + 598, + 890, + 657 + ], + "page_idx": 1 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {C} (\\rho) = \\int_ {X _ {\\mathcal {S}} \\times X _ {\\mathcal {T}}} c \\left(x ^ {s}, x ^ {t}\\right) d \\rho \\left(x ^ {s}, x ^ {t}\\right) \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 570, + 667, + 890, + 700 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "where $c(x^s, x^t)$ is pointwise transportation cost between $x^s \\in X_S$ and $x^t \\in X_T$ . The optimal transport problem is proposed to minimize the $\\mathcal{C}(\\rho)$ under the measure preserving as the following:", + "bbox": [ + 498, + 709, + 890, + 770 + ], + "page_idx": 1 + }, + { + "type": "equation", + "text": "\n$$\nW _ {c} := \\inf _ {\\rho} \\left\\{\\mathcal {C} (\\rho) | | \\pi_ {\\mathcal {S} \\#} \\rho = \\mu_ {\\mathcal {S}}, \\pi_ {\\mathcal {T} \\#} \\rho = \\mu_ {\\mathcal {T}} \\right\\} \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 545, + 779, + 890, + 801 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "By convex optimization theory, we can consider the Kantorovich's dual problem as:", + "bbox": [ + 498, + 811, + 890, + 842 + ], + "page_idx": 1 + }, + { + "type": "equation", + "text": "\n$$\nW _ {c} := \\left\\{ \\begin{array}{l} \\max _ {\\varphi , \\psi} \\int_ {X _ {\\mathcal {S}}} \\varphi \\left(x ^ {s}\\right) d \\mu_ {\\mathcal {S}} + \\int_ {X _ {\\mathcal {T}}} \\psi \\left(x ^ {t}\\right) d \\mu_ {\\mathcal {T}} \\\\ \\text {s . t .} \\quad \\varphi \\left(x ^ {s}\\right) + \\psi \\left(x ^ {t}\\right) \\leq c \\left(x ^ {s}, x ^ {t}\\right) \\end{array} \\right. \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 521, + 851, + 890, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "19999", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "where $\\varphi$ and $\\psi$ are real functions from $X_{\\mathcal{S}}$ and $X_{\\mathcal{T}}$ to $\\mathbb{R}$ . Moreover, the Kantorovich problem can be formulated as", + "bbox": [ + 76, + 90, + 468, + 121 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\max _ {\\varphi} \\left\\{\\int_ {X _ {\\mathcal {S}}} \\varphi \\left(x ^ {s}\\right) d \\mu_ {\\mathcal {S}} + \\int_ {X _ {\\mathcal {T}}} \\varphi^ {c} \\left(x ^ {t}\\right) d \\mu_ {\\mathcal {T}} \\right\\} \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 132, + 132, + 468, + 166 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\varphi^c (x^t) = \\inf_{x_s\\in X_S}\\{c(x^s,x^t) - \\varphi (x^s)\\}$ is called the $c$ -transform of $\\varphi$", + "bbox": [ + 76, + 176, + 468, + 213 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "By classical optimal transport theory, different choices of cost function will influence the difficulty to solving the optimal transport problem. When we choose $c(x^s,x^t) = \\| x^s -x^t\\| _2$ , the problem stated in Equation (5) is equivalent to", + "bbox": [ + 76, + 213, + 468, + 285 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\max _ {\\varphi} \\left\\{\\int_ {X _ {\\mathcal {S}}} \\varphi \\left(x ^ {s}\\right) d \\mu_ {\\mathcal {S}} - \\int_ {X _ {\\mathcal {T}}} \\varphi \\left(x ^ {t}\\right) d \\mu_ {\\mathcal {T}} \\right\\} \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 135, + 285, + 468, + 318 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\varphi$ is under the constraint that $|\\varphi(x) - \\varphi(x')| \\leq \\|x - x'\\|_2$ . WGAN [2] is inspired by above cost setting, during the implementation of optimal transport in WGAN, they utilize the gradient clip to guarantee the Lipschitz constant of $\\varphi$ is bounded from above by 1. When we set the cost function as $c(x^s, x^t) = \\|x^s - x^t\\|_2^2$ , [16] guarantees the existence and uniqueness of optimal transport map.", + "bbox": [ + 76, + 323, + 468, + 429 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4. Methodology", + "text_level": 1, + "bbox": [ + 76, + 443, + 210, + 460 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we present the detail of our COT and demonstrate the advantages of our COT theoretically.", + "bbox": [ + 76, + 469, + 468, + 500 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4.1. Clustering-based Optimal Transport", + "text_level": 1, + "bbox": [ + 76, + 508, + 395, + 525 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Instead of aligning instance-level features between source and target domain, we propose a novel clustering-based optimal transport (COT) module for unsupervised domain adaptation in this subsection. Firstly, we extract features from the source and target domain by ImageNet pretrained CNNs. Then we utilize learnable clusters to represent the sub-domains in the source and target domain, respectively. Finally, we apply a Kantorovich dual form-based loss to implement the optimal transport between clusters from both domains.", + "bbox": [ + 75, + 532, + 468, + 681 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Feature Extractor We utilize an ImageNet pretrained (without fully connected layers) CNNs (e.g. ResNet50/ResNet101) to extract features $\\{x_{i}^{s}\\}_{i = 1}^{n_{s}}$ and $\\{x_j^t\\}_{j = 1}^{n_t}$ from the source and target dataset respectively at the beginning of training process. Note that the distributions of features vary during the training phase.", + "bbox": [ + 76, + 683, + 468, + 773 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Clustering As for each sample from source domain, i.e., $x_{i}^{s}$ and corresponding ground-truth $y_{i}^{s} \\in \\mathcal{Y}$ . We denote the fully-connected layer which outputs the classification logits as $W = [w_{1}^{s},\\dots,w_{|\\mathcal{Y}|}^{s}]^{\\top} \\in \\mathbb{R}^{|\\mathcal{Y}|\\times c}$ , where $|\\mathcal{Y}|$ is the number of categories and $c$ is the number of feature channels. The predicted classification probability is $P(\\hat{y}_i^s = v|x_i^s) = \\frac{e^{x_i^s\\top w_v^s}}{\\sum_{u = 1}^{|\\mathcal{Y}|}e^{x_i^s\\top w_u^s}}$ . The corresponding cross-entropy loss is shown", + "bbox": [ + 76, + 775, + 468, + 904 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "as follows:", + "bbox": [ + 500, + 92, + 576, + 104 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {c r o s s - e n t r o p y}} = \\frac {1}{b} \\sum_ {i = 1} ^ {b} - y _ {i} ^ {s} \\cdot \\log \\left(P \\left(\\hat {y} _ {i} ^ {s} \\mid x _ {i} ^ {s}\\right)\\right) \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 544, + 112, + 890, + 152 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "For the source domain, since $\\{w_v^s\\}_{v=1}^{|\\mathcal{Y}|}$ [13, 47] have been shown effective for clustering representation, we take the classifiers $\\{w_v^s\\}_{v=1}^{|\\mathcal{Y}|}$ as clusters for feature space of source domain. For the target domain, a set of learnable clusters termed as $\\{w_u^t\\}_{u=1}^K$ are proposed to represent the sub-domains, where $K = Q \\cdot |\\mathcal{Y}|$ is a hyper-parameter which stands for the number of sub-domains in the target domain, $Q$ is a positive integer which represents the number of sub-domains for each class. The motivation of introducing $Q$ is to preserve the sub-domain information in the target domain during the COT optimization procedure. For instance, if we set the same number of sub-domains on the source and target domain, the sub-domain information in the target domain would be seriously influenced by those in the source domain during the COT procedure, since the latter is optimized with hard supervision signal while the former is supervised by soft signal. On the contrary, $Q$ helps the target domain generate more sub-domains, which can preserve the sub-domain information in the target domain because not all of them are optimized during the COT phase. Empirically, the performance is superior when $Q$ is set to 2. $\\{w_u^t \\in \\mathbb{R}^c\\}_{u=(l-1)\\cdot Q+1}^{l\\cdot Q}$ represent the clusters for class $l$ , $\\forall 1 \\leq l \\leq |\\mathcal{Y}|$ . For each feature $x_j^t$ , we assign it to the closest cluster in $\\{w_u^t\\}_{u=1}^K$ . We utilize the $L_2$ distance to measure the distance between features and clusters and pull the features back to corresponding clusters.", + "bbox": [ + 496, + 164, + 890, + 565 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\left\\{ \\begin{array}{c} \\mathcal {L} _ {\\text {c l u s t e r}} = \\sum_ {j = 1} ^ {n _ {t}} \\| w _ {u *} ^ {t} - x _ {j} ^ {t} \\| ^ {2} \\\\ s. t. u _ {j} ^ {*} = \\underset {1 \\leq u \\leq K} {\\arg \\min } \\left(\\left\\| w _ {u} ^ {t} - x _ {j} ^ {t} \\right\\| ^ {2}\\right), \\forall 1 \\leq j \\leq n _ {t} \\end{array} \\right. \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 524, + 585, + 890, + 657 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Optimal Transport With clusters $\\{w_v^s\\}_{v=1}^{|\\mathcal{Y}|}$ and $\\{w_u^t\\}_{u=1}^K$ from source and target domain respectively, we design the clustering based optimal transport as follows:", + "bbox": [ + 496, + 666, + 890, + 715 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\left\\{ \\begin{array}{c} \\min _ {T \\in \\mathbb {R} ^ {| \\mathcal {Y} | \\times K}} T _ {v u} c _ {v u} \\\\ s. t. \\sum_ {u = 1} ^ {K} T _ {v u} = \\frac {1}{K}, \\sum_ {v = 1} ^ {| \\mathcal {Y} |} T _ {v u} = \\frac {1}{| \\mathcal {Y} |}, \\\\ T _ {v u} \\geq 0, \\forall 1 \\leq v \\leq | \\mathcal {Y} |, 1 \\leq u \\leq K \\end{array} \\right. \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 557, + 723, + 890, + 811 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $c_{vu} = \\| w_v^s - w_u^t\\|_2^2$ . Similar to Equation (5), we can get the discrete Kantorovich dual problem of Equation (9).", + "bbox": [ + 496, + 821, + 890, + 853 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\max _ {\\psi} \\left\\{\\frac {1}{K} \\sum_ {u = 1} ^ {K} \\psi \\left(w _ {u} ^ {t}\\right) + \\frac {1}{| \\mathcal {Y} |} \\sum_ {v = 1} ^ {| \\mathcal {Y} |} \\psi^ {c} \\left(w _ {v} ^ {s}\\right) \\right\\} \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 560, + 861, + 890, + 904 + ], + "page_idx": 2 + }, + { + "type": "footer", + "text": "20000", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Algorithm 1 Clustering-based Optimal Transport", + "text_level": 1, + "bbox": [ + 78, + 90, + 406, + 106 + ], + "page_idx": 3 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1: Set number of epochs for training as $E$ , learnable clusters for target domain as $\\{w_u^t\\}_{u=1}^{|\\mathcal{Y}|}$ , classifiers/clusters for source domain as $\\{w_v^s\\}_{v=1}^{|\\mathcal{Y}|}$ ;", + "2: for k-th training epoch while $k \\leq E$ do", + "3: for $t$ -th iteration in $k$ -th epoch do", + "4: Take mini-batch of samples from source and target domain as input for feature extractor CNNs with parameters $\\theta$ , the output features are $\\{x_i^s\\}_{i=1}^b$ and $\\{x_j^t\\}_{j=1}^b$ ;", + "5: Compute the $L_{\\mathrm{cluster}}$ for $\\{x_j^t\\}_{j=1}^b$ , $L_{\\mathrm{cross-entropy}}$ for $\\{x_i^s\\}_{i=1}^b$ in the $l$ -th batch, and $\\mathcal{L}_{OT}$ ;", + "6: if $1 \\leq t \\leq k / (b * 10)$ then", + "7: we find the current optimal map from clusters of source domain to those of target domain by maximizing $L_{\\mathrm{OT}}$", + "8: end if", + "9: Minimize $\\mathcal{L}_{COT}$", + "10: end for", + "11: end for" + ], + "bbox": [ + 81, + 109, + 468, + 404 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\psi^c (w_v^s) = \\inf_{u = 1}^{K}(c_{vu} - \\psi (w_u^t))$ . According to Equation 10, We can seek for the optimal transportation map between clusters by optimizing the following loss (line 7 in Algorithm 1):", + "bbox": [ + 76, + 430, + 468, + 503 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {O T}} = \\frac {1}{K} \\sum_ {u = 1} ^ {K} \\lambda_ {u} ^ {t} + \\frac {1}{| \\mathcal {Y} |} \\sum_ {v = 1} ^ {| \\mathcal {Y} |} \\left(\\inf _ {u = 1} ^ {K} \\left(c _ {v u} - \\lambda_ {u} ^ {t}\\right)\\right) \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 107, + 513, + 468, + 556 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\{\\lambda_u^t\\}_{u = 1}^K$ represent the value of function $\\psi$ at points $\\{w_{u}^{t}\\}_{u = 1}^{K}$ . Furthermore, it is worth noting that cost $c_{vu}$ is frozen during the optimization of $\\mathcal{L}_{\\mathrm{OT}}$ , indicating that only $\\{\\lambda_v^t\\}_{v = 1}^{|\\mathcal{V}|}$ is updated in this step. Then based on the founded optimal map, we minimize the following loss $(\\mathcal{L}_{\\mathrm{COT}})$ to close the matched clusters from the feature representation perspective, thus the domain-invariant feature representation is learned via this step (line 9 in Algorithm 1). Correspondingly, we freeze the $\\{\\lambda_v^t\\}_{v = 1}^{|\\mathcal{V}|}$ and only update $c_{vu}$ in this step.", + "bbox": [ + 76, + 566, + 468, + 723 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {C O T}} = \\mathcal {L} _ {\\text {c r o s s - e n t r o p y}} + \\alpha_ {1} \\cdot \\mathcal {L} _ {\\text {c l u s t e r}} + \\alpha_ {2} \\cdot \\mathcal {L} _ {\\mathrm {O T}} \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 107, + 734, + 468, + 750 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\alpha_{1}$ and $\\alpha_{2}$ are loss weights. Finally, we elaborate the overall optimization strategy of our COT in Algorithm 1. Line 6 of Algorithm 1 represents we only optimize $\\mathcal{L}_{OT}$ in the initial $10\\%$ iterations of each epoch.", + "bbox": [ + 76, + 761, + 468, + 821 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.2. Theoretical Analysis on Instance/Clustering Optimal Transport", + "text_level": 1, + "bbox": [ + 76, + 830, + 468, + 862 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Given features $\\{x_i^s\\}_{i = 1}^{n_s}$ and $\\{x_j^t\\}_{j = 1}^{n_t}$ from source and target domain respectively, where $x_{i}^{s}$ and $x_{j}^{t}$ are output from", + "bbox": [ + 76, + 867, + 468, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "shared-parameters neural network for feature extractor. We consider the discrete Kantorovich problem", + "bbox": [ + 498, + 90, + 890, + 121 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\left\\{ \\begin{array}{c} \\min _ {T \\in \\mathbb {R} ^ {n _ {s} \\times n _ {t}}} T _ {i j} c _ {i j} \\\\ s. t. \\sum_ {j = 1} ^ {n _ {t}} T _ {i j} = \\frac {1}{n _ {s}}, \\sum_ {i = 1} ^ {n _ {s}} T _ {i j} = \\frac {1}{n _ {t}}, \\\\ T _ {i j} \\geq 0, \\forall 1 \\leq i \\leq n _ {s}, 1 \\leq j \\leq n _ {t}. \\end{array} \\right. \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 560, + 133, + 890, + 222 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $c_{ij} = \\| x_i^s -x_j^t\\| _2^2$", + "bbox": [ + 500, + 234, + 669, + 253 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Considering the distance and inner-product between features and classifiers:", + "bbox": [ + 498, + 253, + 890, + 281 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\left\\| x _ {i} ^ {s} - w _ {v _ {1} ^ {s}} \\right\\| _ {2} ^ {2} - \\left\\| x _ {i} ^ {s} - w _ {v _ {2}} ^ {s} \\right\\| _ {2} ^ {2} \\\\ = \\| w _ {v _ {1}} ^ {s} \\| _ {2} ^ {2} - \\| w _ {v _ {2}} ^ {s} \\| _ {2} ^ {2} \\tag {14} \\\\ + 2 \\big (\\| w _ {v _ {2}} ^ {s} \\| _ {2} x _ {i} ^ {s} ^ {\\top} \\frac {w _ {v _ {2}} ^ {s}}{\\| w _ {v _ {2}} ^ {s} \\| _ {2}} - \\| w _ {v _ {1}} ^ {s} \\| _ {2} x _ {i} ^ {s} ^ {\\top} \\frac {w _ {v _ {1}} ^ {s}}{\\| w _ {v _ {1}} ^ {s} \\| _ {2}} \\big) \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 526, + 292, + 890, + 369 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In the Bayesian view, we can consider $\\| w_v^s\\| _2$ as the prior probability of class $v$ , $x_{i}^{s}$ is feature representation of a sample and $\\frac{w_s^s}{\\|w_v^s\\|_2}$ is the cluster for class $v$ . $x_{i}^{s\\top}\\frac{w_{v}^{s}}{\\|w_{v}^{s}\\|_{2}}$ measures the similarity between feature and cluster. When classifiers in $\\{\\| w_v^s\\| _2\\}_{v = 1}^{\\lfloor \\mathcal{V}\\rfloor}$ are of the same magnitude, we conclude that the similarity between features and clusters are almost equivalent to the distance between features and classifiers. With labels as supervision, the optimization of crossentropy can promote the inter-class discrepancy which implies", + "bbox": [ + 498, + 381, + 890, + 537 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nx _ {i} ^ {s \\top} \\frac {w _ {y _ {i} ^ {s}}}{\\| w _ {y _ {i} ^ {s}} \\| _ {2}} \\gg x _ {i} ^ {s \\top} \\frac {w _ {v} ^ {s}}{\\| w _ {v} ^ {s} \\| _ {2}}, \\forall v \\neq y _ {i} ^ {s} \\tag {15}\n$$\n", + "text_format": "latex", + "bbox": [ + 571, + 540, + 890, + 578 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "which also provide the following result", + "bbox": [ + 500, + 585, + 759, + 601 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\left\\| x _ {i} ^ {s} - w _ {y _ {i} ^ {s}} ^ {s} \\right\\| _ {2} ^ {2} \\ll \\left\\| x _ {i} ^ {s} - w _ {v} ^ {s} \\right\\| _ {2} ^ {2}, \\forall v \\neq y _ {i} ^ {s} \\tag {16}\n$$\n", + "text_format": "latex", + "bbox": [ + 566, + 614, + 890, + 635 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "If clustering doesn't work sufficiently well, it happens that some samples in the source domain with label $v$ are assigned to samples in the target domain with label $u \\neq v$ . When clustering performs well, We have $c_{ij} \\sim \\bar{c}_{vj} = \\| w_v^s - x_j^t\\|_2^2$ , where $\\sim$ means these two numbers are almost the same. Then we get", + "bbox": [ + 498, + 647, + 890, + 737 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\sum_ {i, j} T _ {i j} c _ {i j} \\sim \\sum_ {v, j} \\left(\\sum_ {x _ {i} ^ {s} \\in X _ {v} ^ {s}} T _ {i j}\\right) \\bar {c} _ {v j} \\tag {17}\n$$\n", + "text_format": "latex", + "bbox": [ + 586, + 751, + 890, + 785 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $X_{v}^{s}$ is the set of samples with label $v$ in the source domain. We denote the number of samples with class $v$ as $n_v$ , then we get", + "bbox": [ + 498, + 797, + 890, + 844 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\sum_ {j = 1} ^ {n _ {t}} \\left(\\sum_ {x _ {i} ^ {s} \\in X _ {v} ^ {s}} T _ {i j}\\right) = \\frac {n _ {v}}{n _ {s}}, \\sum_ {v = 1} ^ {| \\mathcal {Y} |} \\left(\\sum_ {x _ {i} ^ {s} \\in X _ {v} ^ {s}} T _ {i j}\\right) = \\frac {1}{n _ {t}} \\tag {18}\n$$\n", + "text_format": "latex", + "bbox": [ + 529, + 857, + 890, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "20001", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/95f2782fb7eed4b97c3c520b577cf220a14a103344c188df8a4d62993446558a.jpg", + "table_caption": [ + "Table 156 Accuracy (%) on Office-31 for UDA (ResNet-50).The best result is in bold." + ], + "table_footnote": [], + "table_body": "
MethodA→WD→WW→DA→DD→AW→AAvg
Common UDAADDA [46]86.296.298.477.869.568.982.9
JAN [34]85.497.499.884.768.670.084.3
MCD [42]88.698.5100.092.269.569.786.5
GTA [43]89.597.999.887.772.871.486.5
CDAN [33]94.198.6100.092.971.069.387.7
TAT [32]92.599.3100.093.273.172.188.4
MDD [25]94.598.4100.093.574.672.288.9
GSP [19]92.998.799.894.575.974.989.5
DANN [1]82.096.999.179.768.267.482.2
SHOT [29]94.090.174.798.474.399.988.6
MCC [21]95.598.6100.094.472.974.989.4
GVB-GD [9]94.898.7100.095.073.473.789.3
TSA [28]96.098.7100.095.476.776.890.6
SRDC [45]95.799.2100.095.876.777.190.8
OT-basedJDOT [5]84.797.8100.086.464.467.783.5
DeepJDOT [11]88.998.599.688.272.170.186.2
MLOT [23]92.898.5100.090.872.871.687.8
RWOT [51]95.199.5100.094.577.577.990.8
DANN [1] + MMI [28]95.298.6100.094.474.675.289.7
DANN + MMI + COT (Ours)96.599.1100.096.176.777.491.0
", + "bbox": [ + 163, + 113, + 810, + 410 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Inspired by Equation (17) and (18), we consider the following optimal transport between clusters from the source domain and instances from the target domain instead of solving the Kantorovich problem in Equation (13).", + "bbox": [ + 75, + 428, + 468, + 491 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\left\\{ \\begin{array}{c} \\min _ {\\bar {T} \\in \\mathbb {R} ^ {| Y | \\times n _ {t}}} \\bar {T} _ {v j} \\bar {c} _ {v j} \\\\ s. t. \\sum_ {j = 1} ^ {n _ {t}} \\bar {T} _ {v j} = \\frac {n _ {v}}{n _ {s}}, \\sum_ {v = 1} ^ {| \\mathcal {Y} |} \\bar {T} _ {v j} = \\frac {1}{n _ {t}}, \\\\ \\bar {T} _ {v j} \\geq 0, \\forall 1 \\leq v \\leq | \\mathcal {Y} |, 1 \\leq j \\leq n _ {t} \\end{array} \\right. \\tag {19}\n$$\n", + "text_format": "latex", + "bbox": [ + 135, + 494, + 468, + 588 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In general, because of the class imbalance, the empirical label distribution between source and target dataset are different", + "bbox": [ + 75, + 592, + 468, + 637 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\exists \\delta > 0, \\text {s . t .} \\| \\left(\\frac {n _ {1} ^ {s}}{n _ {s}}, \\dots , \\frac {n _ {| \\mathcal {Y} |} ^ {s}}{n _ {s}}\\right) - \\left(\\frac {n _ {1} ^ {t}}{n _ {t}}, \\dots , \\frac {n _ {| \\mathcal {Y} |} ^ {t}}{n _ {t}}\\right) \\| _ {2} \\geq \\delta \\tag {20}\n$$\n", + "text_format": "latex", + "bbox": [ + 84, + 641, + 468, + 676 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\delta$ is a constant which measures the label distribution between source and target domain. There must exist some index $i$ such that $\\frac{n_v^s}{n_s} >\\frac{n_v^t}{n_t}$ , which means that some samples with label $v$ in the source domain will be assigned to samples in the target domain with label $u\\neq v$ . This will result in samples belonging to the same category in the target domain are given different pseudo labels, which increase the difficulty of training and cause the degradation of the performance of deep learning methods on the target domain.", + "bbox": [ + 75, + 679, + 468, + 818 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "When we utilize the clustering based optimal transport, for source domain, we have $\\sum_{u=1}^{K} T_{vu} = \\frac{1}{|\\mathcal{V}|}$ . For target domain, $\\sum_{v=1}^{|\\mathcal{V}|} \\sum_{u=(l-1) \\cdot Q+1}^{l \\cdot Q} T_{vu} = \\frac{Q}{K} = \\frac{1}{|\\mathcal{V}|}$ , which ease the", + "bbox": [ + 75, + 819, + 468, + 902 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "negative effect from class imbalance in domain adaptation based on optimal transport.", + "bbox": [ + 498, + 428, + 890, + 460 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.3. Computation Cost", + "text_level": 1, + "bbox": [ + 498, + 470, + 679, + 488 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In terms of instance-based optimal transport, firstly we need to obtain the features of all samples from the source and target domain, computation cost on the feature extractor is shown as follows:", + "bbox": [ + 498, + 494, + 890, + 555 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {O} \\left(n _ {s} + n _ {t}\\right) \\cdot \\mathcal {O} (\\text {f e a t u r e - e x t r a c t o r}) \\tag {21}\n$$\n", + "text_format": "latex", + "bbox": [ + 580, + 570, + 890, + 587 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathcal{O}$ (feature-extractor) means the computation cost on single sample when extracting the feature. Then considering the optimization of optimal transport, every iteration will need $\\mathcal{O}(n_s\\cdot n_t)$ . In comparison, for cluster-based optimal transport, the main computation cost on optimal transport is $\\mathcal{O}(|\\mathcal{Y}|\\cdot K)$ . For a large-scale dataset, the clustering-based optimal transport cost is much less than instance-based optimal transport.", + "bbox": [ + 498, + 599, + 890, + 722 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 500, + 737, + 633, + 756 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this section, we first elaborate on implementation details and 3 authoritative benchmark datasets in the field of UDA. Then we compare our COT with existing OT-based UDA algorithms and state-of-the-art UDA methods, illustrating that our COT achieves the dominant result in the field of OT-based UDA and competitive performance on the common UDA realm, respectively. Furthermore, we provide some qualitative analysis to illustrate the advantage of our COT when compared with some clustering-based DA", + "bbox": [ + 496, + 763, + 893, + 902 + ], + "page_idx": 4 + }, + { + "type": "footer", + "text": "20002", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/cff87c2047d2a31a9f09d6910ad6aef326eeab843b9de3fd90077c3ae3e289a7.jpg", + "table_caption": [ + "Table 256 Accuracy $(\\%)$ on Office-Home for UDA.The best result is in bold." + ], + "table_footnote": [], + "table_body": "
MethodAr→ClAr→PrAr→RwCl→ArCl→PrCl→RwPr→ArPr→ClPr→RwRw→ArRw→ClRw→PrAvg
Common UDAJAN [34]45.961.268.950.459.761.045.843.470.363.952.476.858.3
TAT [32]51.669.575.459.469.568.659.550.576.870.956.681.665.8
TPN [39]51.271.276.065.172.972.855.448.976.570.953.480.466.2
ETD [26]51.371.985.757.669.273.757.851.279.370.257.582.167.3
SymNets [56]47.772.978.564.271.374.264.248.879.574.552.682.767.6
BNM [8]52.373.980.063.372.974.961.749.579.770.553.682.267.9
MDD [25]54.973.777.860.071.471.861.253.678.172.560.282.368.1
GSP [19]56.875.578.961.369.474.961.352.679.973.354.283.268.4
MCD [42]48.968.374.661.367.668.85747.175.169.152.279.667.8
DANN [1]45.659.370.147.058.560.946.143.768.563.251.876.857.6
CDAN [33]50.770.676.057.670.070.057.450.977.370.956.781.665.8
BSP [4]52.068.676.158.070.370.258.650.277.672.259.381.966.3
TSA [28]55.873.779.061.974.674.560.753.280.172.758.484.369.1
GVB-GD [9]57.074.779.864.674.174.665.255.181.074.659.784.370.4
SRDC [45]52.376.381.069.576.278.068.753.881.776.357.185.071.3
OT-basedJDOT [5]44.763.468.655.164.460.754.348.472.867.457.269.460.5
DeepJDOT [11]39.750.462.539.554.453.236.739.263.652.345.470.550.6
MLOT [23]52.472.677.858.872.472.853.250.876.570.954.482.366.2
DANN [1] + MMI [28]55.774.880.463.674.176.664.254.780.574.358.281.369.9
DANN + MMI + COT (Ours)57.675.283.267.876.275.765.456.282.475.160.784.771.7
Ours + CDTrans [52]70.186.386.482.288.689.177.265.789.283.267.489.881.3
", + "bbox": [ + 80, + 114, + 893, + 347 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "methods in the Optimal Transport realm. Finally, we conduct ablative experiments together with visualization results to further demonstrate the effectiveness of our COT.", + "bbox": [ + 75, + 378, + 468, + 425 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1. Datasets and Implementation Details", + "text_level": 1, + "bbox": [ + 76, + 436, + 397, + 452 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Office-31 is a benchmark dataset on the real-world unsupervised domain adaptation. It has 4110 images for 31 classes drawn from three domains: Amazon (A), DSLR (D), and Webcam (W). The 31 classes in the dataset consist of objects that commonly appeared in office settings, such as keyboards, file cabinets, and laptops.", + "bbox": [ + 75, + 460, + 468, + 549 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Office-Home is a challenging benchmark dataset for domain adaptation which has 4 domains where each domain consists of 65 categories. The four domains are Art - artistic images in the form of sketches, paintings, ornamentation, etc.; Clipart - a collection of clipart images; Product - images of objects without a background and Real-World - images of objects captured with a regular camera. It contains 15,500 images in 65 classes.", + "bbox": [ + 75, + 551, + 468, + 672 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "VisDa-2017 is a large-scale simulation-to-real dataset for domain adaptation, which has over 280,000 images across 12 categories in the training, validation, and testing domains. The training images are generated from the same object under different circumstances, while the validation images are collected from MSCOCO ([31]).", + "bbox": [ + 75, + 672, + 468, + 763 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation Details Follow GVB ([10]), we adopt ResNet-50 pretrained on the ImageNet ([12]) as our backbone for Office-31 and Office-Home benchmarks and ResNet-101 for VisDa-2017 dataset. Note that our COT is a plug-and-play module, indicating it can integrate with a vast body of existing UDA approaches, e.g. DANN [1], CDAN [33]. If there is no extra statement, our COT is implemented on DANN with Maximization Mutual Information Loss (MMI) [28]. In this paper, all experiments are", + "bbox": [ + 75, + 763, + 470, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "implemented by PyTorch. For the optimizer schedule, we adopt SGD with a momentum of 0.9. The total number of training epochs is 40,40,30 on Office-31, Office-Home, and VisDA-2017.", + "bbox": [ + 496, + 378, + 893, + 440 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.2. Results", + "text_level": 1, + "bbox": [ + 500, + 467, + 591, + 482 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Results on Office-Home. Table 2 presents the results of OT-based and common UDA methods on Office-Home dataset. Our COT achieves the highest accuracy $71.7\\%$ , outperforming other OT-based UDA algorithms by $1.8\\%$ at least. Such tremendous improvements demonstrate that our COT can capture a more accurate local-level alignment than previous DA-based methods'. Besides, when compared with other state-of-the-art common UDA methods, to the best of our knowledge, our COT is the first OT-based UDA work that can perform a significant competitive result on the Office-Home dataset.", + "bbox": [ + 496, + 496, + 893, + 662 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Results on Office-31. The results are reported in the Table 1. Our COT achieves the best performance $(91.0\\%)$ both on the common and OT-based UDA.", + "bbox": [ + 496, + 667, + 893, + 713 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Results on VisDA-2017. Table 3 shows the performance of OT-based and common UDA methods on the challenging VisDa-2017 dataset. Our COT outperforms the existing OT-based / common UDA methods (CNN backbone) by $3.1\\%$ / $3.2\\%$ at least, illustrating the dominant role of our COT on UDA task. Furthermore, due to the plug-and-play property of our COT, we directly add this module into the sota UDA method (CDTrans [52] on the challenging dataset VisDa2017 and Office-Home. The results in Table 3 show that our method achieves a new state-of-the-art performance, indicating the superiority and appealing potential of our COT.", + "bbox": [ + 496, + 719, + 893, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "20003", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/f48ea8f42b93b423813898cc2d8464e520e5cd9dc7cbd58e2a279ef01ef44911.jpg", + "table_caption": [ + "Table 356 Accuracy (%) on VisDA-2017 for UDA (ResNet-101). The best result is in bold." + ], + "table_footnote": [], + "table_body": "
Methodplanebcyblbuscarhorseknifemcylepersnplantsktbtraintruckmean
Common UDADANN [1]81.977.782.844.381.229.565.128.651.954.682.87.857.4
MinEnt [18]87.455.075.363.887.443.689.372.582.978.685.627.470.7
TSA [28]93.077.882.250.889.928.077.170.085.280.086.143.071.9
BSP [4]92.272.583.847.587.054.086.872.480.666.984.537.172.1
MCC [21]90.479.872.355.190.586.886.680.094.276.990.049.679.4
MODEL [27]94.873.468.874.893.195.488.684.789.184.783.548.181.6
STAR [35]95.084.084.673.091.691.885.978.494.484.787.042.282.7
BNM [8]89.661.576.955.089.369.181.365.590.047.389.130.170.4
MSTN+DSBN [3]94.786.776.072.095.275.187.981.391.168.988.345.580.2
CGDM [14]92.885.176.364.591.093.281.379.392.483.085.644.880.8
SHOT [29]94.388.580.157.393.193.180.780.391.589.186.358.282.9
TVT [55]92.985.677.560.593.698.289.476.493.692.091.755.783.9
CDTrans [52]98.086.987.980.997.997.396.885.397.683.294.054.488.4
OT-basedJDOT [5]78.470.879.468.882.380.584.270.788.468.878.445.774.7
DeepJDOT [11]85.473.477.387.384.164.791.579.391.944.488.561.877.4
MLOT [23]88.270.477.350.284.877.280.474.483.868.282.338.773.0
RWOT [51]95.180.383.790.092.468.092.582.287.978.490.468.284.0
DANN+MMI+COT (Ours)96.989.684.274.196.496.588.682.096.094.185.162.187.1
Ours + CDTrans [52]98.289.487.682.398.097.296.486.298.392.692.258.189.7
", + "bbox": [ + 81, + 114, + 893, + 376 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/808b1059f7ca154c322db66d294d8ce68f07873d475e1a1d2cc9ca18626644e4.jpg", + "image_caption": [ + "(a) RWOT" + ], + "image_footnote": [], + "bbox": [ + 116, + 417, + 254, + 523 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/39b0392be670efa9997417b0c3a9d2a55ae9b92433ea01c16e37cb857b0cf4b4.jpg", + "image_caption": [ + "(b)COT", + "Figure 1. t-SNE of classifier responses by RWOT and COT on CI50 (red: Amazon, blue: Webcam)." + ], + "image_footnote": [], + "bbox": [ + 287, + 417, + 426, + 522 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.3. Ablation study", + "text_level": 1, + "bbox": [ + 76, + 618, + 227, + 633 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Ability on Resisting Class-imbalance Challenge. As the results illustrated above, our COT brings little gain than the existing sota OT-based UDA method (RWOT [51]) on the Office-31 dataset while achieving the dominant performance on Office-Home and VisDA-2017 datasets. To find the reason for this, we compute the label distribution of the Office-31 dataset, where the class imbalance phenomenon is not occurred obviously compared to the Office-Home and VisDA-2017 datasets (shown in Fig. 2). Thus, to verify the ability of our COT on resisting class-imbalance challenges among all UDA benchmark datasets, we construct 3 types of class-imbalance evaluation benchmarks by randomly erasing $30\\%$ , $50\\%$ , $70\\%$ samples from certain classes on Office-31 datasets target domain, termed CI30, CI50, and CI70, respectively. Table 4 presents the results of our COT and existing OT-based UDA methods on Office-31, CI30, CI50, and CI70. According to this quantitative result, we dis", + "bbox": [ + 75, + 643, + 472, + 901 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "cover that the more severe the class imbalance phenomenon occurs, the larger the performance gap between our COT and RWOT, demonstrating that our COT embraces excellent ability on handling class-imbalance challenges.", + "bbox": [ + 498, + 393, + 890, + 455 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/a592e923b6e8d7883c52125824420db63305344c22118575346e3f4da81c3429.jpg", + "table_caption": [ + "Table 4. Mean accuracy of OT-based UDA methods on class-imbalance datasets." + ], + "table_footnote": [], + "table_body": "
DatasetCOTRWOTMLOT
Office-3191.090.887.8
CI3088.284.783.6
CI5087.778.380.4
CI7085.470.575.2
", + "bbox": [ + 513, + 505, + 879, + 585 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Computation Cost. Here, we compute the computation cost of our COT and RWOT, where RWOT is a representative work of a sample-level OT matching mechanism. By looking at the results in Fig. 3, compared with RWOT, our COT significantly saves the computational cost with the increasing of training scale. Integrating the above theoretical analysis with this quantitative result demonstrates the superiority of our COT in economizing the computation head.", + "bbox": [ + 496, + 598, + 890, + 718 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Effects of Loss Weight. $\\alpha_{1}$ and $\\alpha_{2}$ are employed to weight the importance of cluster and OT loss, respectively. In our experiment, $\\alpha_{1}$ and $\\alpha_{2}$ are the same and selected from a broad range of $\\{0.1, 0.2, 0.4, 1.0, 2.0\\}$ . Results are evaluated on the Office-31 and shown in Table 5. When $\\alpha_{1}$ and $\\alpha_{2}$ are set to large value, the performance will tend to drop. In our opinion, the larger $\\alpha_{1}$ and $\\alpha_{2}$ would put more attention to local alignment while weakening the feature representation on both domains. As a result, $\\alpha_{1}$ and $\\alpha_{2}$ are set to 0.2, 0.2 for all datasets.", + "bbox": [ + 496, + 719, + 890, + 869 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Effects of the Number of Sub-domains. $Y$ , defined in Section 4, represents the number of sub-domains for each", + "bbox": [ + 500, + 869, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "20004", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/fb6d561851bbfd5ede1e570d6e0345edb266678544eb7e5e269c84179527bdce.jpg", + "image_caption": [ + "(a) Office-31" + ], + "image_footnote": [], + "bbox": [ + 109, + 114, + 330, + 238 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/dd567420a6c5c3048c62f644a724b8751e00bdb499bb01e14a007fd68a4a1be7.jpg", + "image_caption": [ + "(b) Office-Home" + ], + "image_footnote": [], + "bbox": [ + 364, + 114, + 584, + 237 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/6b349db665b116784e576912a6477a04dd180ee24790c94a95edecfcee1dc072.jpg", + "image_caption": [ + "(c) VisDA-2017" + ], + "image_footnote": [], + "bbox": [ + 616, + 114, + 839, + 237 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/65737c917147038b2e2e94e1cb82ad1fe7f20e462da8a4b312a56d52b90bc40a.jpg", + "image_caption": [ + "Figure 2. Class distribution on Office-31, Office-Home and VisDA-2017. Note that the visualized class distribution is randomly selected from corresponding datasets.", + "Figure 3. Computation cost of COT and RWOT" + ], + "image_footnote": [], + "bbox": [ + 160, + 310, + 375, + 436 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/03ada669735cbf5ac53de42b60db0959d21feec3ab691d655889ea06ada2f178.jpg", + "table_caption": [ + "Table 5. Mean accuracy of our COT with different $\\alpha_{1}$ and $\\alpha_{2}$ on Office-31 dataset." + ], + "table_footnote": [], + "table_body": "
α1α2avg_acc
0.10.190.8
0.20.291.0
0.40.490.6
1.01.089.4
2.02.087.7
", + "bbox": [ + 84, + 518, + 460, + 614 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "class. We investigate a broad range of $Y$ to find an appropriate value. Based on the results shown in Table 6, we set $Y$ to 2 for all test settings.", + "bbox": [ + 75, + 642, + 468, + 688 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/e7bb15cf2a2ec1a22a58b2aee3e5a7993a5349544abcdf059cd001ce417829d6.jpg", + "table_caption": [ + "Table 6. Mean accuracy of our COT with different $Y$ on Office-31 dataset." + ], + "table_footnote": [], + "table_body": "
Yavg_acc
190.7
291.0
390.8
590.4
", + "bbox": [ + 112, + 741, + 434, + 823 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Qualitative Analysis on COT As discussed above, we present 2 advantages of our COT when compared with existing OT-based methods [23,51]. In this part, we further illustrate another advantage compared to the clustering-based", + "bbox": [ + 75, + 839, + 468, + 902 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "OT algorithm in the conventional OT realm [24, 36]. Our COT is optimized in an end-to-end manner, thus the feature representation and optimal transportation map are both updated asynchronously. While for [24] and [36], they focus on obtaining offline cluster centers firstly and then compute the optimal transportation map. Such a synchronous optimization manner inevitably leads to a sub-optimal mapping result, because the computation procedure of an optimal transportation map is based on the fixed cluster centers. This is why we adopt an online clustering mode instead of an offline one.", + "bbox": [ + 496, + 297, + 893, + 464 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.4. Visualization", + "text_level": 1, + "bbox": [ + 500, + 484, + 635, + 500 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We present the t-SNE visualization of feature representation for our COT and RWOT [51] on CI50. Under the challenge of class imbalance, the learned features from RWOT in some classes are ambiguous to provide accurate representation compared to our COT. These visualization results further demonstrate the excellent ability of our COT on handling class-imbalance challenges from the feature representation perspective.", + "bbox": [ + 496, + 513, + 890, + 633 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 660, + 617, + 674 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we propose a novel method to advance the OT-based domain adaptation community, which integrates optimal transport theory with clustering operation, termed Clustering-based Optimal Transport (COT). Concretely, COT applies the loss derived from discrete Kantorovich dual form to cluster and align centers in the source and target domain, thus transferring knowledge from the source domain to the target domain. Moreover, our COT can eliminate the negative effect brought by the class imbalance phenomenon and reduce the computation cost simultaneously, which are two challenging problems for existing OT-based UDA algorithms. Besides, we also provide comprehensive theoretical analysis and experimental results to guarantee the advantages of our COT.", + "bbox": [ + 496, + 689, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "20005", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1135] Hana Ajakan, Pascal Germain, Hugo Larochelle, François Laviolette, and Mario Marchand. Domain-adversarial neural networks. arXiv preprint arXiv:1412.4446, 2014. 1, 5, 6, 7", + "[2] M. Arjovsky, S. Chintala, and L. Bottou. Wasserstein gan. 2017. 3", + "[3] Woong-Gi Chang, Tackgeun You, Seonguk Seo, Suha Kwak, and Bohyung Han. Domain-specific batch normalization for unsupervised domain adaptation. In CVPR, 2019. 7", + "[4] Xinyang Chen, Sinan Wang, Mingsheng Long, and Jianmin Wang. Transferability vs. discriminability: Batch spectral penalization for adversarial domain adaptation. In International conference on machine learning, pages 1081-1090. PMLR, 2019. 6, 7", + "[5] Nicolas Courty, Rémi Flamary, Amaury Habrard, and Alain Rakotomamonjy. Joint distribution optimal transportation for domain adaptation. arXiv preprint arXiv:1705.08848, 2017. 2, 5, 6, 7", + "[6] Nicolas Courty, Rémi Flamary, and Devis Tuia. Domain adaptation with regularized optimal transport. In Joint European Conference on Machine Learning and Knowledge Discovery in Databases, pages 274-289. Springer, 2014. 1", + "[7] Nicolas Courty, Rémi Flamary, Devis Tuia, and Alain Rakotomamonjy. Optimal transport for domain adaptation. IEEE transactions on pattern analysis and machine intelligence, 39(9):1853-1865, 2016. 2", + "[8] Shuhao Cui, Shuhui Wang, Junbao Zhuo, Liang Li, Qingming Huang, and Qi Tian. Towards discriminability and diversity: Batch nuclear-norm maximization under label insufficient situations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3941-3950, 2020. 6, 7", + "[9] Shuhao Cui, Shuhui Wang, Junbao Zhuo, Chi Su, Qingming Huang, and Tian Qi. Gradually vanishing bridge for adversarial domain adaptation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2020. 5, 6", + "[10] Shuhao Cui, Shuhui Wang, Junbao Zhuo, Chi Su, Qingming Huang, and Qi Tian. Gradually vanishing bridge for adversarial domain adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12455-12464, 2020. 6", + "[11] Bharath Bhushan Damodaran, Benjamin Kellenberger, Rémi Flamary, Devis Tuia, and Nicolas Courty. Deepjoint: Deep joint distribution optimal transport for unsupervised domain adaptation. In Proceedings of the European Conference on Computer Vision (ECCV), pages 447-463, 2018. 5, 6, 7", + "[12] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 6", + "[13] Jiankang Deng, Jia Guo, Niannan Xue, and Stefanos Zafeiriou. Arcface: Additive angular margin loss for deep face recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4690-4699, 2019. 1, 3" + ], + "bbox": [ + 78, + 114, + 485, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[14] Zhekai Du, Jingjing Li, Hongzu Su, Lei Zhu, and Ke Lu. Cross-domain gradient discrepancy minimization for unsupervised domain adaptation. 2021. 7", + "[15] Kilian Fatras, Thibault Séjourne, Rémi Flamary, and Nicolas Courty. Unbalanced minibatch optimal transport; applications to domain adaptation. In International Conference on Machine Learning, pages 3186-3197. PMLR, 2021. 1", + "[16] Wilfrid Gangbo and Robert J McCann. The geometry of optimal transportation. Acta Mathematica, 177(2):113-161, 1996. 3", + "[17] Yixiao Ge, Dapeng Chen, and Hongsheng Li. Mutual meanteaching: Pseudo label refinery for unsupervised domain adaptation on person re-identification. In International Conference on Learning Representations, 2020. 2", + "[18] Yves Grandvalet, Yoshua Bengio, et al. Semi-supervised learning by entropy minimization. In CAP, pages 281-296, 2005. 7", + "[19] Sahand Hajifar and Hongyue Sun. Online domain adaptation for continuous cross-subject liver viability evaluation based on irregular thermal data, 2020. 5, 6", + "[20] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. 2016. 1", + "[21] Ying Jin, Ximei Wang, Mingsheng Long, and Jianmin Wang. Minimum class confusion for versatile domain adaptation, 2020. 5, 7", + "[22] Leonid V Kantorovich. On the translocation of masses. Journal of mathematical sciences, 133(4):1381-1382, 2006. 2", + "[23] Tanguy Kerdoncuff, Rémi Emonet, and Marc Sebban. Metric learning in optimal transport for domain adaptation. 2020. 1, 2, 5, 6, 7, 8", + "[24] John Lee, Max Dabagia, Eva Dyer, and Christopher Rozell. Hierarchical optimal transport for multimodal distribution alignment. Advances in Neural Information Processing Systems, 32, 2019. 8", + "[25] Jingjing Li, Erpeng Chen, Zhengming Ding, Lei Zhu, Ke Lu, and Heng Tao Shen. Maximum density divergence for domain adaptation. IEEE Transactions on Pattern Analysis and Machine Intelligence, page 1-1, 2020. 5, 6", + "[26] Mengxue Li, Yi-Ming Zhai, You-Wei Luo, Peng-Fei Ge, and Chuan-Xian Ren. Enhanced transport distance for unsupervised domain adaptation. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 13933–13941, 2020. 6", + "[27] Rui Li, Qianfen Jiao, Wenming Cao, Hau-San Wong, and Si Wu. Model adaptation: Unsupervised domain adaptation without source data. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9641-9650, 2020. 7", + "[28] Shuang Li, Mixue Xie, Kaixiong Gong, Chi Harold Liu, Yulin Wang, and Wei Li. Transferable semantic augmentation for domain adaptation. arXiv preprint arXiv:2103.12562, 2021. 1, 5, 6, 7", + "[29] Jian Liang, Dapeng Hu, and Jiashi Feng. Do we really need to access the source data? source hypothesis transfer for unsupervised domain adaptation. 2020. 5, 7", + "[30] Chi-Heng Lin, Mehdi Azabou, and Eva L Dyer. Making transport more robust and interpretable by moving data" + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "20006", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "through a small number of anchor points. Proceedings of machine learning research, 139:6631, 2021. 1", + "[31] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dálár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In European conference on computer vision, pages 740–755. Springer, 2014. 6", + "[32] Hong Liu, Mingsheng Long, Jianmin Wang, and Michael Jordan. Transferable adversarial training: A general approach to adapting deep classifiers. In International Conference on Machine Learning, pages 4013-4022. PMLR, 2019. 5, 6", + "[33] Mingsheng Long, Zhangjie Cao, Jianmin Wang, and Michael I Jordan. Conditional adversarial domain adaptation. arXiv preprint arXiv:1705.10667, 2017. 5, 6", + "[34] Mingsheng Long, Han Zhu, Jianmin Wang, and Michael I Jordan. Deep transfer learning with joint adaptation networks. In International conference on machine learning, pages 2208-2217. PMLR, 2017. 5, 6", + "[35] Zhihe Lu, Yongxin Yang, Xiatian Zhu, Cong Liu, Yi-Zhe Song, and Tao Xiang. Stochastic classifiers for unsupervised domain adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9111-9120, 2020. 7", + "[36] Liang Mi, Wen Zhang, and Yalin Wang. Regularized wasserstein means for aligning distributional data. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 34, pages 5166-5173, 2020. 8", + "[37] Gaspard Monge. Mémoire sur la théorie des déblais et des remblais. Histoire de l'Académie Royale des Sciences de Paris, 1781. 2", + "[38] Khai Nguyen, Dang Nguyen, Quoc Nguyen, Tung Pham, Hung Bui, Dinh Phung, Trung Le, and Nhat Ho. On transportation of mini-batches: A hierarchical approach. arXiv preprint arXiv:2102.05912, 2021. 1", + "[39] Yingwei Pan, Ting Yao, Yehao Li, Yu Wang, Chong-Wah Ngo, and Tao Mei. Transferrable prototypical networks for unsupervised domain adaptation, 2019. 6", + "[40] Ievgen Redko, Nicolas Courty, Rémi Flamary, and Devis Tuia. Optimal transport for multi-source domain adaptation under target shift. In The 22nd International Conference on Artificial Intelligence and Statistics, pages 849-858. PMLR, 2019. 1, 2", + "[41] Hochang Rhee and Nam Ik Cho. Efficient and robust pseudolabeling for unsupervised domain adaptation. In 2019 Asia-Pacific Signal and Information Processing Association Annual Summit and Conference (APSIPA ASC), pages 980–985, 2019. 2", + "[42] Kuniaki Saito, Kohei Watanabe, Yoshitaka Ushiku, and Tatsuya Harada. Maximum classifier discrepancy for unsupervised domain adaptation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3723-3732, 2018. 1, 5, 6", + "[43] Swami Sankaranarayanan, Yogesh Balaji, Carlos D Castillo, and Rama Chellappa. Generate to adapt: Aligning domains using generative adversarial networks. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 8503-8512, 2018. 5" + ], + "bbox": [ + 78, + 90, + 470, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[44] Filippo Santambrogio. Optimal transport for applied mathematicians. Birkhäuser, NY, 55(58-63):94, 2015. 1", + "[45] Hui Tang, Ke Chen, and Kui Jia. Unsupervised domain adaptation via structurally regularized deep clustering. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8725-8735, 2020. 5, 6", + "[46] Eric Tzeng, Judy Hoffman, Kate Saenko, and Trevor Darrell. Adversarial discriminative domain adaptation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7167-7176, 2017. 5", + "[47] Hao Wang, Yitong Wang, Zheng Zhou, Xing Ji, Dihong Gong, Jingchao Zhou, Zhifeng Li, and Wei Liu. Cosface: Large margin cosine loss for deep face recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5265-5274, 2018. 3", + "[48] Qian Wang and Toby Breckon. Unsupervised domain adaptation via structured prediction based selective pseudolabeling. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 34, pages 6243-6250, 2020. 2", + "[49] Wei Wang, Haojie Li, Zhengming Ding, and Zhihui Wang. Rethink maximum mean discrepancy for domain adaptation. arXiv preprint arXiv:2007.00689, 2020. 1", + "[50] Pengfei Wei, Yiping Ke, Xinghua Qu, and Tze-Yun Leong. Subdomain adaptation with manifolds discrepancy alignment, 2020. 1", + "[51] Renjun Xu, Pelen Liu, Liyan Wang, Chao Chen, and Jindong Wang. Reliable weighted optimal transport for unsupervised domain adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4394-4403, 2020. 1, 5, 7, 8", + "[52] Tongkun Xu, Weihua Chen, Pichao Wang, Fan Wang, Hao Li, and Rong Jin. Cdtrans: Cross-domain transformer for unsupervised domain adaptation. arXiv preprint arXiv:2109.06165, 2021. 6, 7", + "[53] Hongliang Yan, Yukang Ding, Peihua Li, Qilong Wang, Yong Xu, and Wangmeng Zuo. Mind the class weight bias: Weighted maximum mean discrepancy for unsupervised domain adaptation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 2272-2281, 2017. 1", + "[54] Yuguang Yan, Wen Li, Hanrui Wu, Huaqing Min, Mingkui Tan, and Qingyao Wu. Semi-supervised optimal transport for heterogeneous domain adaptation. In IJCAI, volume 7, pages 2969-2975, 2018. 1", + "[55] Jinyu Yang, Jingjing Liu, Ning Xu, and Junzhou Huang. Tvt: Transferable vision transformer for unsupervised domain adaptation. arXiv preprint arXiv:2108.05988, 2021. 7", + "[56] Yabin Zhang, Hui Tang, Kui Jia, and Mingkui Tan. Domain-symmetric networks for adversarial domain adaptation, 2019. 6", + "[57] Yongchun Zhu, Fuzhen Zhuang, Jindong Wang, Guolin Ke, Jingwu Chen, Jiang Bian, Hui Xiong, and Qing He. Deep subdomain adaptation network for image classification. IEEE Transactions on Neural Networks and Learning Systems, 32(4):1713-1722, Apr 2021. 1" + ], + "bbox": [ + 501, + 92, + 890, + 869 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "20007", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2023/COT_ Unsupervised Domain Adaptation With Clustering and Optimal Transport/53ca0802-c684-4d48-b297-7731efdaab53_model.json b/2023/COT_ Unsupervised Domain Adaptation With Clustering and Optimal Transport/53ca0802-c684-4d48-b297-7731efdaab53_model.json new file mode 100644 index 0000000000000000000000000000000000000000..bf4e075d61691974cd200c5e6a89ed137b78a2b9 --- /dev/null +++ b/2023/COT_ Unsupervised Domain Adaptation With Clustering and Optimal Transport/53ca0802-c684-4d48-b297-7731efdaab53_model.json @@ -0,0 +1,2541 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.081, + 0.13, + 0.892, + 0.155 + ], + "angle": 0, + "content": "COT: Unsupervised Domain Adaptation with Clustering and Optimal Transport" + }, + { + "type": "text", + "bbox": [ + 0.283, + 0.179, + 0.681, + 0.219 + ], + "angle": 0, + "content": "Yang Liu \\(^{1*}\\) Zhipeng Zhou \\(^{1*}\\) Baigui Sun \\(^{1\\dagger}\\) 1Alibaba Group" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.251, + 0.314, + 0.269 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.284, + 0.474, + 0.617 + ], + "angle": 0, + "content": "Unsupervised domain adaptation (UDA) aims to transfer the knowledge from a labeled source domain to an unlabeled target domain. Typically, to guarantee desirable knowledge transfer, aligning the distribution between source and target domain from a global perspective is widely adopted in UDA. Recent researchers further point out the importance of local-level alignment and propose to construct instance-pair alignment by leveraging on Optimal Transport (OT) theory. However, existing OT-based UDA approaches are limited to handling class imbalance challenges and introduce a heavy computation overhead when considering a large-scale training situation. To cope with two aforementioned issues, we propose a Clustering-based Optimal Transport (COT) algorithm, which formulates the alignment procedure as an Optimal Transport problem and constructs a mapping between clustering centers in the source and target domain via an end-to-end manner. With this alignment on clustering centers, our COT eliminates the negative effect caused by class imbalance and reduces the computation cost simultaneously. Empirically, our COT achieves state-of-the-art performance on several authoritative benchmark datasets." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.662, + 0.21, + 0.679 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.688, + 0.47, + 0.824 + ], + "angle": 0, + "content": "Benefiting from the availability of large-scale data, deep learning has achieved tremendous success over the past few years. However, directly applying a well-trained convolution neural network on a new domain frequently suffers from the domain gap/discrepancy challenge, resulting in spurious predictions on the new domain. To remedy this, Unsupervised Domain Adaptation (UDA) has attracted many researchers' attention, which can transfer the knowledge from a labeled domain to an unlabeled domain." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.825, + 0.47, + 0.841 + ], + "angle": 0, + "content": "A major line of UDA approaches [1, 1, 28, 42, 49, 53] aim" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.253, + 0.895, + 0.359 + ], + "angle": 0, + "content": "to learn a global domain shift by aligning the global source and target distribution while ignoring the local-level alignment between two domains. By leveraging on global domain adaptation, the global distributions of source and target domain are almost the same, thus losing the fine-grained information for each class (class-structure) on the source and target domain." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.36, + 0.895, + 0.617 + ], + "angle": 0, + "content": "Recently, to preserve class structure in both domains, several works [6, 15, 23, 30, 38, 40, 44, 51, 54] adopt optimal transport (OT) to minimize the sample-level transportation cost between source and target domain, achieving a significant performance on UDA. However, there exist two issues on recent OT-based UDA approaches. (i) When considering a realistic situation, i.e. the class imbalance\\(^{1}\\) phenomenon occurs between the source and target domain, samples belonging to the same class in the target domain are assigned with false pseudo labels due to the mechanism of optimal transport, which requires each sample in source domain can be mapped to target samples. As a result, existing OT-based UDA methods provide poor pair-wise matching when facing class imbalance challenges. (ii) OT-based UDA methods tend to find a sample-level optimal counterpart, which requires a large amount of computation overhead, especially training on large-scale datasets." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.618, + 0.896, + 0.875 + ], + "angle": 0, + "content": "To solve two aforementioned issues, we propose a Clustering-based Optimal Transport algorithm, termed COT, to construct a clustering-level instead of sample-level mapping between source and target domain. Clusters in the source domain are obtained from the classifiers supervised by the labeled source domain data. While for the target domain, COT utilizes a set of learnable clusters to represent the feature distribution of the target domain, which can describe the sub-domain information [50, 57]. For instance, in many object recognition tasks [13, 20] an object could contain many attributes. Each attribute can be viewed as a sub-domain. To this end, the clusters on the source and target domain can represent the individual sub-domain information, respectively, such that optimal transport between clusters intrinsically provides a local mapping from the subdomain in the source domain to those in the target domain. Moreover, we provide a theoretical analysis and compre" + }, + { + "type": "page_footnote", + "bbox": [ + 0.518, + 0.887, + 0.866, + 0.902 + ], + "angle": 0, + "content": "1 label distribution are different in two domains, \\(P_{s}(y)\\neq P_{t}(y)\\)" + }, + { + "type": "page_footnote", + "bbox": [ + 0.078, + 0.852, + 0.291, + 0.864 + ], + "angle": 0, + "content": "Email: ly261666@alibaba-inc.com" + }, + { + "type": "page_footnote", + "bbox": [ + 0.08, + 0.865, + 0.192, + 0.876 + ], + "angle": 0, + "content": "* Equal Contribution" + }, + { + "type": "page_footnote", + "bbox": [ + 0.08, + 0.876, + 0.211, + 0.889 + ], + "angle": 0, + "content": "† Corresponding Author" + }, + { + "type": "list", + "bbox": [ + 0.078, + 0.852, + 0.291, + 0.889 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "19998" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.151 + ], + "angle": 0, + "content": "hensive experimental results to guarantee that (i) COT can alleviate the negative effect caused by class imbalance; (ii) Compared to existing OT-based UDA approaches, our COT economizes much computation head." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.153, + 0.393, + 0.167 + ], + "angle": 0, + "content": "In summary, our main contributions include:" + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.178, + 0.468, + 0.268 + ], + "angle": 0, + "content": "- We propose a novel Clustering-based Optimal Transport module as well as a specially designed loss derived from the discrete type of Kantorovich dual form, which resolves two aforementioned challenges on the existing OT-based UDA algorithms, facilitating the development of OT-based UDA community." + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.278, + 0.468, + 0.307 + ], + "angle": 0, + "content": "- We provide a theoretical analysis to guarantee the advantages of our COT." + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.318, + 0.468, + 0.346 + ], + "angle": 0, + "content": "- Our COT achieves state-of-the-art performance on several UDA benchmark datasets." + }, + { + "type": "list", + "bbox": [ + 0.096, + 0.178, + 0.468, + 0.346 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.362, + 0.216, + 0.377 + ], + "angle": 0, + "content": "2. Related work" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.387, + 0.47, + 0.658 + ], + "angle": 0, + "content": "Pseudo Label based Domain Adaptation Inspired by the observation that samples in the target domain can be clustered within the feature space, for accurate pseudolabeling, [48] propose a selective pseudo-labeling strategy based on structural predictions which utilize the unsupervised clustering analysis. [41] introduce a confidence-based weighting scheme for obtaining pseudo-labels and an adaptive threshold adjustment strategy to provide sufficient and accurate pseudo-labels during the training process. The confidence-based weighting scheme generates pseudo-labels that can enable the performance less sensitive to the threshold determined by the pseudo-labels. In the task of person re-identification, [17] propose an unsupervised framework called Mutual Mean-Teaching to learn better features from the target domain by refining the hard pseudo labels offline and soft pseudo labels online alternatively to mitigate the effects of noisy pseudo labels caused by the clustering algorithms." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.66, + 0.468, + 0.764 + ], + "angle": 0, + "content": "Optimal Transport based Domain Adaptation As a way to find a minimal effort strategy to the transport of a given mass of dirt into a given hole, [37] put forward the optimal transport problem for the first time. [22] provide an extension of the original problem of Monge. Recently, by applying the optimal transport to domain adaptation, some new computation strategies have emerged." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.468, + 0.901 + ], + "angle": 0, + "content": "[7] propose regularized unsupervised optimal transport model to align the representation of features between different domains. The regularization schemes encoding class-structure in source domain during estimation of transport map enforce the intuition that samples of the same class must undergo a similar transformation. [5] minimize the optimal transport loss between the joint source distribution and the estimated target joint distribution depending on a function that is introduced to predict an output value given" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.258 + ], + "angle": 0, + "content": "input from the source domain. For reducing the discrepancy between multiple domains, [40] propose Joint Class Proportion and Optimal Transport which performs multisource domain adaptation and target shift correction simultaneously by learning the predicted class probability of the unlabeled target data and the coupling to align the distributions between source and target domain. For better alignment between different domains, a relation between target error and the magnitude of different Wasserstein distances are proposed in [23] which optimize the metric for domain adaptation." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.27, + 0.627, + 0.287 + ], + "angle": 0, + "content": "3. Preliminary" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.296, + 0.891, + 0.326 + ], + "angle": 0, + "content": "In this section, we introduce the basic knowledge for optimal transport." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.335, + 0.683, + 0.351 + ], + "angle": 0, + "content": "3.1. Optimal Transport" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.358, + 0.892, + 0.509 + ], + "angle": 0, + "content": "Let \\( X \\subseteq \\mathbb{R}^d \\) be a measurable space and the labels are denoted as \\( \\mathcal{V} \\). We denote the set of all probability distributions on \\( X \\) as \\( \\mathcal{P}(X) \\). The source and target domains are space \\( X \\) equipped with two distinct probability distributions \\( \\mu_{\\mathcal{S}} \\) and \\( \\mu_{\\mathcal{T}} \\). Suppose we have source dataset \\( \\{x_i^s\\}_{i=1}^{n_s} \\subset X_{\\mathcal{S}} = (X, \\mu_{\\mathcal{S}}) \\) associated with label set \\( \\{y_i^s\\}_{i=1}^{n_s} \\) where \\( y_i^s \\in \\mathcal{V} \\). The target dataset is \\( \\{x_j^t\\}_{j=1}^{n_t} \\subset X_{\\mathcal{T}} = (X, \\mu_{\\mathcal{T}}) \\) without labels. The goal of optimal transport is to minimize the inter-domain transportation cost by finding a feasible map to preserve measure." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.517, + 0.892, + 0.563 + ], + "angle": 0, + "content": "Definition 1 (Kantorovich) For given joint distribution \\(\\rho(x^s, x^t)\\) which satisfies for every measurable Borel set \\(O_S \\subset X_S\\), \\(O_T \\subset X_T\\), we have" + }, + { + "type": "equation", + "bbox": [ + 0.515, + 0.573, + 0.891, + 0.59 + ], + "angle": 0, + "content": "\\[\n\\rho \\left(O _ {\\mathcal {S}} \\times X _ {\\mathcal {T}}\\right) = \\mu_ {\\mathcal {S}} \\left(O _ {\\mathcal {S}}\\right), \\rho \\left(X _ {\\mathcal {S}} \\times O _ {\\mathcal {T}}\\right) = \\mu_ {\\mathcal {T}} \\left(O _ {\\mathcal {T}}\\right) \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.599, + 0.891, + 0.659 + ], + "angle": 0, + "content": "For convenience, we denote the projection maps from \\(X_{\\mathcal{S}} \\times X_{\\mathcal{T}}\\) to \\(X_{\\mathcal{S}}\\) and \\(X_{\\mathcal{T}}\\) as \\(\\pi_{\\mathcal{S}}, \\pi_{\\mathcal{T}}\\). The above equation can be denoted as \\(\\pi_{\\mathcal{S}\\#} \\rho = \\mu_{\\mathcal{S}}\\) and \\(\\pi_{\\mathcal{T}\\#} \\rho = \\mu_{\\mathcal{T}}\\). The corresponding transportation cost is" + }, + { + "type": "equation", + "bbox": [ + 0.571, + 0.668, + 0.891, + 0.701 + ], + "angle": 0, + "content": "\\[\n\\mathcal {C} (\\rho) = \\int_ {X _ {\\mathcal {S}} \\times X _ {\\mathcal {T}}} c \\left(x ^ {s}, x ^ {t}\\right) d \\rho \\left(x ^ {s}, x ^ {t}\\right) \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.71, + 0.891, + 0.771 + ], + "angle": 0, + "content": "where \\( c(x^s, x^t) \\) is pointwise transportation cost between \\( x^s \\in X_S \\) and \\( x^t \\in X_T \\). The optimal transport problem is proposed to minimize the \\( \\mathcal{C}(\\rho) \\) under the measure preserving as the following:" + }, + { + "type": "equation", + "bbox": [ + 0.547, + 0.78, + 0.891, + 0.803 + ], + "angle": 0, + "content": "\\[\nW _ {c} := \\inf _ {\\rho} \\left\\{\\mathcal {C} (\\rho) | | \\pi_ {\\mathcal {S} \\#} \\rho = \\mu_ {\\mathcal {S}}, \\pi_ {\\mathcal {T} \\#} \\rho = \\mu_ {\\mathcal {T}} \\right\\} \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.812, + 0.891, + 0.843 + ], + "angle": 0, + "content": "By convex optimization theory, we can consider the Kantorovich's dual problem as:" + }, + { + "type": "equation", + "bbox": [ + 0.522, + 0.852, + 0.891, + 0.904 + ], + "angle": 0, + "content": "\\[\nW _ {c} := \\left\\{ \\begin{array}{l} \\max _ {\\varphi , \\psi} \\int_ {X _ {\\mathcal {S}}} \\varphi \\left(x ^ {s}\\right) d \\mu_ {\\mathcal {S}} + \\int_ {X _ {\\mathcal {T}}} \\psi \\left(x ^ {t}\\right) d \\mu_ {\\mathcal {T}} \\\\ \\text {s . t .} \\quad \\varphi \\left(x ^ {s}\\right) + \\psi \\left(x ^ {t}\\right) \\leq c \\left(x ^ {s}, x ^ {t}\\right) \\end{array} \\right. \\tag {4}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19999" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.47, + 0.122 + ], + "angle": 0, + "content": "where \\(\\varphi\\) and \\(\\psi\\) are real functions from \\(X_{\\mathcal{S}}\\) and \\(X_{\\mathcal{T}}\\) to \\(\\mathbb{R}\\). Moreover, the Kantorovich problem can be formulated as" + }, + { + "type": "equation", + "bbox": [ + 0.133, + 0.133, + 0.47, + 0.167 + ], + "angle": 0, + "content": "\\[\n\\max _ {\\varphi} \\left\\{\\int_ {X _ {\\mathcal {S}}} \\varphi \\left(x ^ {s}\\right) d \\mu_ {\\mathcal {S}} + \\int_ {X _ {\\mathcal {T}}} \\varphi^ {c} \\left(x ^ {t}\\right) d \\mu_ {\\mathcal {T}} \\right\\} \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.177, + 0.469, + 0.214 + ], + "angle": 0, + "content": "where \\(\\varphi^c (x^t) = \\inf_{x_s\\in X_S}\\{c(x^s,x^t) - \\varphi (x^s)\\}\\) is called the \\(c\\) -transform of \\(\\varphi\\)" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.214, + 0.469, + 0.286 + ], + "angle": 0, + "content": "By classical optimal transport theory, different choices of cost function will influence the difficulty to solving the optimal transport problem. When we choose \\( c(x^s,x^t) = \\| x^s -x^t\\| _2 \\), the problem stated in Equation (5) is equivalent to" + }, + { + "type": "equation", + "bbox": [ + 0.137, + 0.286, + 0.469, + 0.319 + ], + "angle": 0, + "content": "\\[\n\\max _ {\\varphi} \\left\\{\\int_ {X _ {\\mathcal {S}}} \\varphi \\left(x ^ {s}\\right) d \\mu_ {\\mathcal {S}} - \\int_ {X _ {\\mathcal {T}}} \\varphi \\left(x ^ {t}\\right) d \\mu_ {\\mathcal {T}} \\right\\} \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.324, + 0.469, + 0.43 + ], + "angle": 0, + "content": "where \\(\\varphi\\) is under the constraint that \\(|\\varphi(x) - \\varphi(x')| \\leq \\|x - x'\\|_2\\). WGAN [2] is inspired by above cost setting, during the implementation of optimal transport in WGAN, they utilize the gradient clip to guarantee the Lipschitz constant of \\(\\varphi\\) is bounded from above by 1. When we set the cost function as \\(c(x^s, x^t) = \\|x^s - x^t\\|_2^2\\), [16] guarantees the existence and uniqueness of optimal transport map." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.444, + 0.212, + 0.461 + ], + "angle": 0, + "content": "4. Methodology" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.47, + 0.469, + 0.501 + ], + "angle": 0, + "content": "In this section, we present the detail of our COT and demonstrate the advantages of our COT theoretically." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.51, + 0.396, + 0.526 + ], + "angle": 0, + "content": "4.1. Clustering-based Optimal Transport" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.533, + 0.469, + 0.682 + ], + "angle": 0, + "content": "Instead of aligning instance-level features between source and target domain, we propose a novel clustering-based optimal transport (COT) module for unsupervised domain adaptation in this subsection. Firstly, we extract features from the source and target domain by ImageNet pretrained CNNs. Then we utilize learnable clusters to represent the sub-domains in the source and target domain, respectively. Finally, we apply a Kantorovich dual form-based loss to implement the optimal transport between clusters from both domains." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.684, + 0.469, + 0.775 + ], + "angle": 0, + "content": "Feature Extractor We utilize an ImageNet pretrained (without fully connected layers) CNNs (e.g. ResNet50/ResNet101) to extract features \\(\\{x_{i}^{s}\\}_{i = 1}^{n_{s}}\\) and \\(\\{x_j^t\\}_{j = 1}^{n_t}\\) from the source and target dataset respectively at the beginning of training process. Note that the distributions of features vary during the training phase." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.776, + 0.469, + 0.905 + ], + "angle": 0, + "content": "Clustering As for each sample from source domain, i.e., \\( x_{i}^{s} \\) and corresponding ground-truth \\( y_{i}^{s} \\in \\mathcal{Y} \\). We denote the fully-connected layer which outputs the classification logits as \\( W = [w_{1}^{s},\\dots,w_{|\\mathcal{Y}|}^{s}]^{\\top} \\in \\mathbb{R}^{|\\mathcal{Y}|\\times c} \\), where \\( |\\mathcal{Y}| \\) is the number of categories and \\( c \\) is the number of feature channels. The predicted classification probability is \\( P(\\hat{y}_i^s = v|x_i^s) = \\frac{e^{x_i^s\\top w_v^s}}{\\sum_{u = 1}^{|\\mathcal{Y}|}e^{x_i^s\\top w_u^s}} \\). The corresponding cross-entropy loss is shown" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.093, + 0.577, + 0.106 + ], + "angle": 0, + "content": "as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.545, + 0.113, + 0.892, + 0.154 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {c r o s s - e n t r o p y}} = \\frac {1}{b} \\sum_ {i = 1} ^ {b} - y _ {i} ^ {s} \\cdot \\log \\left(P \\left(\\hat {y} _ {i} ^ {s} \\mid x _ {i} ^ {s}\\right)\\right) \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.165, + 0.892, + 0.566 + ], + "angle": 0, + "content": "For the source domain, since \\(\\{w_v^s\\}_{v=1}^{|\\mathcal{Y}|}\\) [13, 47] have been shown effective for clustering representation, we take the classifiers \\(\\{w_v^s\\}_{v=1}^{|\\mathcal{Y}|}\\) as clusters for feature space of source domain. For the target domain, a set of learnable clusters termed as \\(\\{w_u^t\\}_{u=1}^K\\) are proposed to represent the sub-domains, where \\(K = Q \\cdot |\\mathcal{Y}|\\) is a hyper-parameter which stands for the number of sub-domains in the target domain, \\(Q\\) is a positive integer which represents the number of sub-domains for each class. The motivation of introducing \\(Q\\) is to preserve the sub-domain information in the target domain during the COT optimization procedure. For instance, if we set the same number of sub-domains on the source and target domain, the sub-domain information in the target domain would be seriously influenced by those in the source domain during the COT procedure, since the latter is optimized with hard supervision signal while the former is supervised by soft signal. On the contrary, \\(Q\\) helps the target domain generate more sub-domains, which can preserve the sub-domain information in the target domain because not all of them are optimized during the COT phase. Empirically, the performance is superior when \\(Q\\) is set to 2. \\(\\{w_u^t \\in \\mathbb{R}^c\\}_{u=(l-1)\\cdot Q+1}^{l\\cdot Q}\\) represent the clusters for class \\(l\\), \\(\\forall 1 \\leq l \\leq |\\mathcal{Y}|\\). For each feature \\(x_j^t\\), we assign it to the closest cluster in \\(\\{w_u^t\\}_{u=1}^K\\). We utilize the \\(L_2\\) distance to measure the distance between features and clusters and pull the features back to corresponding clusters." + }, + { + "type": "equation", + "bbox": [ + 0.525, + 0.586, + 0.891, + 0.658 + ], + "angle": 0, + "content": "\\[\n\\left\\{ \\begin{array}{c} \\mathcal {L} _ {\\text {c l u s t e r}} = \\sum_ {j = 1} ^ {n _ {t}} \\| w _ {u *} ^ {t} - x _ {j} ^ {t} \\| ^ {2} \\\\ s. t. u _ {j} ^ {*} = \\underset {1 \\leq u \\leq K} {\\arg \\min } \\left(\\left\\| w _ {u} ^ {t} - x _ {j} ^ {t} \\right\\| ^ {2}\\right), \\forall 1 \\leq j \\leq n _ {t} \\end{array} \\right. \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.667, + 0.892, + 0.716 + ], + "angle": 0, + "content": "Optimal Transport With clusters \\(\\{w_v^s\\}_{v=1}^{|\\mathcal{Y}|}\\) and \\(\\{w_u^t\\}_{u=1}^K\\) from source and target domain respectively, we design the clustering based optimal transport as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.558, + 0.724, + 0.892, + 0.813 + ], + "angle": 0, + "content": "\\[\n\\left\\{ \\begin{array}{c} \\min _ {T \\in \\mathbb {R} ^ {| \\mathcal {Y} | \\times K}} T _ {v u} c _ {v u} \\\\ s. t. \\sum_ {u = 1} ^ {K} T _ {v u} = \\frac {1}{K}, \\sum_ {v = 1} ^ {| \\mathcal {Y} |} T _ {v u} = \\frac {1}{| \\mathcal {Y} |}, \\\\ T _ {v u} \\geq 0, \\forall 1 \\leq v \\leq | \\mathcal {Y} |, 1 \\leq u \\leq K \\end{array} \\right. \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.822, + 0.892, + 0.854 + ], + "angle": 0, + "content": "where \\( c_{vu} = \\| w_v^s - w_u^t\\|_2^2 \\). Similar to Equation (5), we can get the discrete Kantorovich dual problem of Equation (9)." + }, + { + "type": "equation", + "bbox": [ + 0.561, + 0.862, + 0.892, + 0.905 + ], + "angle": 0, + "content": "\\[\n\\max _ {\\psi} \\left\\{\\frac {1}{K} \\sum_ {u = 1} ^ {K} \\psi \\left(w _ {u} ^ {t}\\right) + \\frac {1}{| \\mathcal {Y} |} \\sum_ {v = 1} ^ {| \\mathcal {Y} |} \\psi^ {c} \\left(w _ {v} ^ {s}\\right) \\right\\} \\tag {10}\n\\]" + }, + { + "type": "footer", + "bbox": [ + 0.479, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "20000" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.092, + 0.408, + 0.107 + ], + "angle": 0, + "content": "Algorithm 1 Clustering-based Optimal Transport" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.11, + 0.47, + 0.161 + ], + "angle": 0, + "content": "1: Set number of epochs for training as \\( E \\), learnable clusters for target domain as \\( \\{w_u^t\\}_{u=1}^{|\\mathcal{Y}|} \\), classifiers/clusters for source domain as \\( \\{w_v^s\\}_{v=1}^{|\\mathcal{Y}|} \\);" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.161, + 0.37, + 0.176 + ], + "angle": 0, + "content": "2: for k-th training epoch while \\(k \\leq E\\) do" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.177, + 0.347, + 0.19 + ], + "angle": 0, + "content": "3: for \\( t \\)-th iteration in \\( k \\)-th epoch do" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.191, + 0.47, + 0.253 + ], + "angle": 0, + "content": "4: Take mini-batch of samples from source and target domain as input for feature extractor CNNs with parameters \\(\\theta\\), the output features are \\(\\{x_i^s\\}_{i=1}^b\\) and \\(\\{x_j^t\\}_{j=1}^b\\);" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.253, + 0.469, + 0.285 + ], + "angle": 0, + "content": "5: Compute the \\(L_{\\mathrm{cluster}}\\) for \\(\\{x_j^t\\}_{j=1}^b\\), \\(L_{\\mathrm{cross-entropy}}\\) for \\(\\{x_i^s\\}_{i=1}^b\\) in the \\(l\\)-th batch, and \\(\\mathcal{L}_{OT}\\);" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.285, + 0.319, + 0.3 + ], + "angle": 0, + "content": "6: if \\( 1 \\leq t \\leq k / (b * 10) \\) then" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.301, + 0.469, + 0.346 + ], + "angle": 0, + "content": "7: we find the current optimal map from clusters of source domain to those of target domain by maximizing \\(L_{\\mathrm{OT}}\\)" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.346, + 0.184, + 0.36 + ], + "angle": 0, + "content": "8: end if" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.361, + 0.261, + 0.376 + ], + "angle": 0, + "content": "9: Minimize \\(\\mathcal{L}_{COT}\\)" + }, + { + "type": "ref_text", + "bbox": [ + 0.082, + 0.377, + 0.177, + 0.39 + ], + "angle": 0, + "content": "10: end for" + }, + { + "type": "ref_text", + "bbox": [ + 0.082, + 0.392, + 0.16, + 0.405 + ], + "angle": 0, + "content": "11: end for" + }, + { + "type": "list", + "bbox": [ + 0.082, + 0.11, + 0.47, + 0.405 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.431, + 0.47, + 0.504 + ], + "angle": 0, + "content": "where \\(\\psi^c (w_v^s) = \\inf_{u = 1}^{K}(c_{vu} - \\psi (w_u^t))\\). According to Equation 10, We can seek for the optimal transportation map between clusters by optimizing the following loss (line 7 in Algorithm 1):" + }, + { + "type": "equation", + "bbox": [ + 0.108, + 0.514, + 0.469, + 0.557 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {O T}} = \\frac {1}{K} \\sum_ {u = 1} ^ {K} \\lambda_ {u} ^ {t} + \\frac {1}{| \\mathcal {Y} |} \\sum_ {v = 1} ^ {| \\mathcal {Y} |} \\left(\\inf _ {u = 1} ^ {K} \\left(c _ {v u} - \\lambda_ {u} ^ {t}\\right)\\right) \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.567, + 0.47, + 0.724 + ], + "angle": 0, + "content": "where \\(\\{\\lambda_u^t\\}_{u = 1}^K\\) represent the value of function \\(\\psi\\) at points \\(\\{w_{u}^{t}\\}_{u = 1}^{K}\\). Furthermore, it is worth noting that cost \\(c_{vu}\\) is frozen during the optimization of \\(\\mathcal{L}_{\\mathrm{OT}}\\), indicating that only \\(\\{\\lambda_v^t\\}_{v = 1}^{|\\mathcal{V}|}\\) is updated in this step. Then based on the founded optimal map, we minimize the following loss \\((\\mathcal{L}_{\\mathrm{COT}})\\) to close the matched clusters from the feature representation perspective, thus the domain-invariant feature representation is learned via this step (line 9 in Algorithm 1). Correspondingly, we freeze the \\(\\{\\lambda_v^t\\}_{v = 1}^{|\\mathcal{V}|}\\) and only update \\(c_{vu}\\) in this step." + }, + { + "type": "equation", + "bbox": [ + 0.108, + 0.735, + 0.469, + 0.751 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {C O T}} = \\mathcal {L} _ {\\text {c r o s s - e n t r o p y}} + \\alpha_ {1} \\cdot \\mathcal {L} _ {\\text {c l u s t e r}} + \\alpha_ {2} \\cdot \\mathcal {L} _ {\\mathrm {O T}} \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.762, + 0.469, + 0.822 + ], + "angle": 0, + "content": "where \\(\\alpha_{1}\\) and \\(\\alpha_{2}\\) are loss weights. Finally, we elaborate the overall optimization strategy of our COT in Algorithm 1. Line 6 of Algorithm 1 represents we only optimize \\(\\mathcal{L}_{OT}\\) in the initial \\(10\\%\\) iterations of each epoch." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.831, + 0.47, + 0.863 + ], + "angle": 0, + "content": "4.2. Theoretical Analysis on Instance/Clustering Optimal Transport" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.868, + 0.47, + 0.903 + ], + "angle": 0, + "content": "Given features \\(\\{x_i^s\\}_{i = 1}^{n_s}\\) and \\(\\{x_j^t\\}_{j = 1}^{n_t}\\) from source and target domain respectively, where \\(x_{i}^{s}\\) and \\(x_{j}^{t}\\) are output from" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.892, + 0.122 + ], + "angle": 0, + "content": "shared-parameters neural network for feature extractor. We consider the discrete Kantorovich problem" + }, + { + "type": "equation", + "bbox": [ + 0.561, + 0.135, + 0.892, + 0.223 + ], + "angle": 0, + "content": "\\[\n\\left\\{ \\begin{array}{c} \\min _ {T \\in \\mathbb {R} ^ {n _ {s} \\times n _ {t}}} T _ {i j} c _ {i j} \\\\ s. t. \\sum_ {j = 1} ^ {n _ {t}} T _ {i j} = \\frac {1}{n _ {s}}, \\sum_ {i = 1} ^ {n _ {s}} T _ {i j} = \\frac {1}{n _ {t}}, \\\\ T _ {i j} \\geq 0, \\forall 1 \\leq i \\leq n _ {s}, 1 \\leq j \\leq n _ {t}. \\end{array} \\right. \\tag {13}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.236, + 0.67, + 0.254 + ], + "angle": 0, + "content": "where \\(c_{ij} = \\| x_i^s -x_j^t\\| _2^2\\)" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.254, + 0.891, + 0.282 + ], + "angle": 0, + "content": "Considering the distance and inner-product between features and classifiers:" + }, + { + "type": "equation", + "bbox": [ + 0.527, + 0.294, + 0.892, + 0.37 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\left\\| x _ {i} ^ {s} - w _ {v _ {1} ^ {s}} \\right\\| _ {2} ^ {2} - \\left\\| x _ {i} ^ {s} - w _ {v _ {2}} ^ {s} \\right\\| _ {2} ^ {2} \\\\ = \\| w _ {v _ {1}} ^ {s} \\| _ {2} ^ {2} - \\| w _ {v _ {2}} ^ {s} \\| _ {2} ^ {2} \\tag {14} \\\\ + 2 \\big (\\| w _ {v _ {2}} ^ {s} \\| _ {2} x _ {i} ^ {s} ^ {\\top} \\frac {w _ {v _ {2}} ^ {s}}{\\| w _ {v _ {2}} ^ {s} \\| _ {2}} - \\| w _ {v _ {1}} ^ {s} \\| _ {2} x _ {i} ^ {s} ^ {\\top} \\frac {w _ {v _ {1}} ^ {s}}{\\| w _ {v _ {1}} ^ {s} \\| _ {2}} \\big) \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.382, + 0.892, + 0.538 + ], + "angle": 0, + "content": "In the Bayesian view, we can consider \\(\\| w_v^s\\| _2\\) as the prior probability of class \\(v\\), \\(x_{i}^{s}\\) is feature representation of a sample and \\(\\frac{w_s^s}{\\|w_v^s\\|_2}\\) is the cluster for class \\(v\\). \\(x_{i}^{s\\top}\\frac{w_{v}^{s}}{\\|w_{v}^{s}\\|_{2}}\\) measures the similarity between feature and cluster. When classifiers in \\(\\{\\| w_v^s\\| _2\\}_{v = 1}^{\\lfloor \\mathcal{V}\\rfloor}\\) are of the same magnitude, we conclude that the similarity between features and clusters are almost equivalent to the distance between features and classifiers. With labels as supervision, the optimization of crossentropy can promote the inter-class discrepancy which implies" + }, + { + "type": "equation", + "bbox": [ + 0.573, + 0.541, + 0.892, + 0.579 + ], + "angle": 0, + "content": "\\[\nx _ {i} ^ {s \\top} \\frac {w _ {y _ {i} ^ {s}}}{\\| w _ {y _ {i} ^ {s}} \\| _ {2}} \\gg x _ {i} ^ {s \\top} \\frac {w _ {v} ^ {s}}{\\| w _ {v} ^ {s} \\| _ {2}}, \\forall v \\neq y _ {i} ^ {s} \\tag {15}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.587, + 0.761, + 0.602 + ], + "angle": 0, + "content": "which also provide the following result" + }, + { + "type": "equation", + "bbox": [ + 0.568, + 0.615, + 0.892, + 0.636 + ], + "angle": 0, + "content": "\\[\n\\left\\| x _ {i} ^ {s} - w _ {y _ {i} ^ {s}} ^ {s} \\right\\| _ {2} ^ {2} \\ll \\left\\| x _ {i} ^ {s} - w _ {v} ^ {s} \\right\\| _ {2} ^ {2}, \\forall v \\neq y _ {i} ^ {s} \\tag {16}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.648, + 0.892, + 0.738 + ], + "angle": 0, + "content": "If clustering doesn't work sufficiently well, it happens that some samples in the source domain with label \\( v \\) are assigned to samples in the target domain with label \\( u \\neq v \\). When clustering performs well, We have \\( c_{ij} \\sim \\bar{c}_{vj} = \\| w_v^s - x_j^t\\|_2^2 \\), where \\( \\sim \\) means these two numbers are almost the same. Then we get" + }, + { + "type": "equation", + "bbox": [ + 0.588, + 0.752, + 0.892, + 0.786 + ], + "angle": 0, + "content": "\\[\n\\sum_ {i, j} T _ {i j} c _ {i j} \\sim \\sum_ {v, j} \\left(\\sum_ {x _ {i} ^ {s} \\in X _ {v} ^ {s}} T _ {i j}\\right) \\bar {c} _ {v j} \\tag {17}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.799, + 0.892, + 0.845 + ], + "angle": 0, + "content": "where \\( X_{v}^{s} \\) is the set of samples with label \\( v \\) in the source domain. We denote the number of samples with class \\( v \\) as \\( n_v \\), then we get" + }, + { + "type": "equation", + "bbox": [ + 0.531, + 0.858, + 0.892, + 0.904 + ], + "angle": 0, + "content": "\\[\n\\sum_ {j = 1} ^ {n _ {t}} \\left(\\sum_ {x _ {i} ^ {s} \\in X _ {v} ^ {s}} T _ {i j}\\right) = \\frac {n _ {v}}{n _ {s}}, \\sum_ {v = 1} ^ {| \\mathcal {Y} |} \\left(\\sum_ {x _ {i} ^ {s} \\in X _ {v} ^ {s}} T _ {i j}\\right) = \\frac {1}{n _ {t}} \\tag {18}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "20001" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.235, + 0.089, + 0.746, + 0.104 + ], + "angle": 0, + "content": "Table 156 Accuracy (%) on Office-31 for UDA (ResNet-50).The best result is in bold." + }, + { + "type": "table", + "bbox": [ + 0.164, + 0.114, + 0.812, + 0.411 + ], + "angle": 0, + "content": "
MethodA→WD→WW→DA→DD→AW→AAvg
Common UDAADDA [46]86.296.298.477.869.568.982.9
JAN [34]85.497.499.884.768.670.084.3
MCD [42]88.698.5100.092.269.569.786.5
GTA [43]89.597.999.887.772.871.486.5
CDAN [33]94.198.6100.092.971.069.387.7
TAT [32]92.599.3100.093.273.172.188.4
MDD [25]94.598.4100.093.574.672.288.9
GSP [19]92.998.799.894.575.974.989.5
DANN [1]82.096.999.179.768.267.482.2
SHOT [29]94.090.174.798.474.399.988.6
MCC [21]95.598.6100.094.472.974.989.4
GVB-GD [9]94.898.7100.095.073.473.789.3
TSA [28]96.098.7100.095.476.776.890.6
SRDC [45]95.799.2100.095.876.777.190.8
OT-basedJDOT [5]84.797.8100.086.464.467.783.5
DeepJDOT [11]88.998.599.688.272.170.186.2
MLOT [23]92.898.5100.090.872.871.687.8
RWOT [51]95.199.5100.094.577.577.990.8
DANN [1] + MMI [28]95.298.6100.094.474.675.289.7
DANN + MMI + COT (Ours)96.599.1100.096.176.777.491.0
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.429, + 0.47, + 0.492 + ], + "angle": 0, + "content": "Inspired by Equation (17) and (18), we consider the following optimal transport between clusters from the source domain and instances from the target domain instead of solving the Kantorovich problem in Equation (13)." + }, + { + "type": "equation", + "bbox": [ + 0.137, + 0.495, + 0.47, + 0.589 + ], + "angle": 0, + "content": "\\[\n\\left\\{ \\begin{array}{c} \\min _ {\\bar {T} \\in \\mathbb {R} ^ {| Y | \\times n _ {t}}} \\bar {T} _ {v j} \\bar {c} _ {v j} \\\\ s. t. \\sum_ {j = 1} ^ {n _ {t}} \\bar {T} _ {v j} = \\frac {n _ {v}}{n _ {s}}, \\sum_ {v = 1} ^ {| \\mathcal {Y} |} \\bar {T} _ {v j} = \\frac {1}{n _ {t}}, \\\\ \\bar {T} _ {v j} \\geq 0, \\forall 1 \\leq v \\leq | \\mathcal {Y} |, 1 \\leq j \\leq n _ {t} \\end{array} \\right. \\tag {19}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.593, + 0.47, + 0.638 + ], + "angle": 0, + "content": "In general, because of the class imbalance, the empirical label distribution between source and target dataset are different" + }, + { + "type": "equation", + "bbox": [ + 0.085, + 0.642, + 0.469, + 0.677 + ], + "angle": 0, + "content": "\\[\n\\exists \\delta > 0, \\text {s . t .} \\| \\left(\\frac {n _ {1} ^ {s}}{n _ {s}}, \\dots , \\frac {n _ {| \\mathcal {Y} |} ^ {s}}{n _ {s}}\\right) - \\left(\\frac {n _ {1} ^ {t}}{n _ {t}}, \\dots , \\frac {n _ {| \\mathcal {Y} |} ^ {t}}{n _ {t}}\\right) \\| _ {2} \\geq \\delta \\tag {20}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.68, + 0.469, + 0.819 + ], + "angle": 0, + "content": "where \\(\\delta\\) is a constant which measures the label distribution between source and target domain. There must exist some index \\(i\\) such that \\(\\frac{n_v^s}{n_s} >\\frac{n_v^t}{n_t}\\), which means that some samples with label \\(v\\) in the source domain will be assigned to samples in the target domain with label \\(u\\neq v\\). This will result in samples belonging to the same category in the target domain are given different pseudo labels, which increase the difficulty of training and cause the degradation of the performance of deep learning methods on the target domain." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.82, + 0.47, + 0.904 + ], + "angle": 0, + "content": "When we utilize the clustering based optimal transport, for source domain, we have \\(\\sum_{u=1}^{K} T_{vu} = \\frac{1}{|\\mathcal{V}|}\\). For target domain, \\(\\sum_{v=1}^{|\\mathcal{V}|} \\sum_{u=(l-1) \\cdot Q+1}^{l \\cdot Q} T_{vu} = \\frac{Q}{K} = \\frac{1}{|\\mathcal{V}|}\\), which ease the" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.429, + 0.892, + 0.461 + ], + "angle": 0, + "content": "negative effect from class imbalance in domain adaptation based on optimal transport." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.472, + 0.68, + 0.489 + ], + "angle": 0, + "content": "4.3. Computation Cost" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.496, + 0.892, + 0.556 + ], + "angle": 0, + "content": "In terms of instance-based optimal transport, firstly we need to obtain the features of all samples from the source and target domain, computation cost on the feature extractor is shown as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.581, + 0.571, + 0.891, + 0.588 + ], + "angle": 0, + "content": "\\[\n\\mathcal {O} \\left(n _ {s} + n _ {t}\\right) \\cdot \\mathcal {O} (\\text {f e a t u r e - e x t r a c t o r}) \\tag {21}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.601, + 0.892, + 0.723 + ], + "angle": 0, + "content": "where \\(\\mathcal{O}\\) (feature-extractor) means the computation cost on single sample when extracting the feature. Then considering the optimization of optimal transport, every iteration will need \\(\\mathcal{O}(n_s\\cdot n_t)\\). In comparison, for cluster-based optimal transport, the main computation cost on optimal transport is \\(\\mathcal{O}(|\\mathcal{Y}|\\cdot K)\\). For a large-scale dataset, the clustering-based optimal transport cost is much less than instance-based optimal transport." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.738, + 0.634, + 0.757 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.894, + 0.903 + ], + "angle": 0, + "content": "In this section, we first elaborate on implementation details and 3 authoritative benchmark datasets in the field of UDA. Then we compare our COT with existing OT-based UDA algorithms and state-of-the-art UDA methods, illustrating that our COT achieves the dominant result in the field of OT-based UDA and competitive performance on the common UDA realm, respectively. Furthermore, we provide some qualitative analysis to illustrate the advantage of our COT when compared with some clustering-based DA" + }, + { + "type": "footer", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "20002" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.264, + 0.089, + 0.719, + 0.104 + ], + "angle": 0, + "content": "Table 256 Accuracy \\((\\%)\\) on Office-Home for UDA.The best result is in bold." + }, + { + "type": "table", + "bbox": [ + 0.081, + 0.115, + 0.894, + 0.348 + ], + "angle": 0, + "content": "
MethodAr→ClAr→PrAr→RwCl→ArCl→PrCl→RwPr→ArPr→ClPr→RwRw→ArRw→ClRw→PrAvg
Common UDAJAN [34]45.961.268.950.459.761.045.843.470.363.952.476.858.3
TAT [32]51.669.575.459.469.568.659.550.576.870.956.681.665.8
TPN [39]51.271.276.065.172.972.855.448.976.570.953.480.466.2
ETD [26]51.371.985.757.669.273.757.851.279.370.257.582.167.3
SymNets [56]47.772.978.564.271.374.264.248.879.574.552.682.767.6
BNM [8]52.373.980.063.372.974.961.749.579.770.553.682.267.9
MDD [25]54.973.777.860.071.471.861.253.678.172.560.282.368.1
GSP [19]56.875.578.961.369.474.961.352.679.973.354.283.268.4
MCD [42]48.968.374.661.367.668.85747.175.169.152.279.667.8
DANN [1]45.659.370.147.058.560.946.143.768.563.251.876.857.6
CDAN [33]50.770.676.057.670.070.057.450.977.370.956.781.665.8
BSP [4]52.068.676.158.070.370.258.650.277.672.259.381.966.3
TSA [28]55.873.779.061.974.674.560.753.280.172.758.484.369.1
GVB-GD [9]57.074.779.864.674.174.665.255.181.074.659.784.370.4
SRDC [45]52.376.381.069.576.278.068.753.881.776.357.185.071.3
OT-basedJDOT [5]44.763.468.655.164.460.754.348.472.867.457.269.460.5
DeepJDOT [11]39.750.462.539.554.453.236.739.263.652.345.470.550.6
MLOT [23]52.472.677.858.872.472.853.250.876.570.954.482.366.2
DANN [1] + MMI [28]55.774.880.463.674.176.664.254.780.574.358.281.369.9
DANN + MMI + COT (Ours)57.675.283.267.876.275.765.456.282.475.160.784.771.7
Ours + CDTrans [52]70.186.386.482.288.689.177.265.789.283.267.489.881.3
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.38, + 0.47, + 0.426 + ], + "angle": 0, + "content": "methods in the Optimal Transport realm. Finally, we conduct ablative experiments together with visualization results to further demonstrate the effectiveness of our COT." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.437, + 0.398, + 0.453 + ], + "angle": 0, + "content": "5.1. Datasets and Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.461, + 0.47, + 0.55 + ], + "angle": 0, + "content": "Office-31 is a benchmark dataset on the real-world unsupervised domain adaptation. It has 4110 images for 31 classes drawn from three domains: Amazon (A), DSLR (D), and Webcam (W). The 31 classes in the dataset consist of objects that commonly appeared in office settings, such as keyboards, file cabinets, and laptops." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.552, + 0.47, + 0.673 + ], + "angle": 0, + "content": "Office-Home is a challenging benchmark dataset for domain adaptation which has 4 domains where each domain consists of 65 categories. The four domains are Art - artistic images in the form of sketches, paintings, ornamentation, etc.; Clipart - a collection of clipart images; Product - images of objects without a background and Real-World - images of objects captured with a regular camera. It contains 15,500 images in 65 classes." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.673, + 0.47, + 0.765 + ], + "angle": 0, + "content": "VisDa-2017 is a large-scale simulation-to-real dataset for domain adaptation, which has over 280,000 images across 12 categories in the training, validation, and testing domains. The training images are generated from the same object under different circumstances, while the validation images are collected from MSCOCO ([31])." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Implementation Details Follow GVB ([10]), we adopt ResNet-50 pretrained on the ImageNet ([12]) as our backbone for Office-31 and Office-Home benchmarks and ResNet-101 for VisDa-2017 dataset. Note that our COT is a plug-and-play module, indicating it can integrate with a vast body of existing UDA approaches, e.g. DANN [1], CDAN [33]. If there is no extra statement, our COT is implemented on DANN with Maximization Mutual Information Loss (MMI) [28]. In this paper, all experiments are" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.38, + 0.895, + 0.441 + ], + "angle": 0, + "content": "implemented by PyTorch. For the optimizer schedule, we adopt SGD with a momentum of 0.9. The total number of training epochs is 40,40,30 on Office-31, Office-Home, and VisDA-2017." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.468, + 0.593, + 0.483 + ], + "angle": 0, + "content": "5.2. Results" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.497, + 0.895, + 0.663 + ], + "angle": 0, + "content": "Results on Office-Home. Table 2 presents the results of OT-based and common UDA methods on Office-Home dataset. Our COT achieves the highest accuracy \\(71.7\\%\\), outperforming other OT-based UDA algorithms by \\(1.8\\%\\) at least. Such tremendous improvements demonstrate that our COT can capture a more accurate local-level alignment than previous DA-based methods'. Besides, when compared with other state-of-the-art common UDA methods, to the best of our knowledge, our COT is the first OT-based UDA work that can perform a significant competitive result on the Office-Home dataset." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.669, + 0.895, + 0.714 + ], + "angle": 0, + "content": "Results on Office-31. The results are reported in the Table 1. Our COT achieves the best performance \\((91.0\\%)\\) both on the common and OT-based UDA." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.895, + 0.901 + ], + "angle": 0, + "content": "Results on VisDA-2017. Table 3 shows the performance of OT-based and common UDA methods on the challenging VisDa-2017 dataset. Our COT outperforms the existing OT-based / common UDA methods (CNN backbone) by \\(3.1\\%\\) / \\(3.2\\%\\) at least, illustrating the dominant role of our COT on UDA task. Furthermore, due to the plug-and-play property of our COT, we directly add this module into the sota UDA method (CDTrans [52] on the challenging dataset VisDa2017 and Office-Home. The results in Table 3 show that our method achieves a new state-of-the-art performance, indicating the superiority and appealing potential of our COT." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20003" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.22, + 0.089, + 0.761, + 0.104 + ], + "angle": 0, + "content": "Table 356 Accuracy (%) on VisDA-2017 for UDA (ResNet-101). The best result is in bold." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.115, + 0.895, + 0.377 + ], + "angle": 0, + "content": "
Methodplanebcyblbuscarhorseknifemcylepersnplantsktbtraintruckmean
Common UDADANN [1]81.977.782.844.381.229.565.128.651.954.682.87.857.4
MinEnt [18]87.455.075.363.887.443.689.372.582.978.685.627.470.7
TSA [28]93.077.882.250.889.928.077.170.085.280.086.143.071.9
BSP [4]92.272.583.847.587.054.086.872.480.666.984.537.172.1
MCC [21]90.479.872.355.190.586.886.680.094.276.990.049.679.4
MODEL [27]94.873.468.874.893.195.488.684.789.184.783.548.181.6
STAR [35]95.084.084.673.091.691.885.978.494.484.787.042.282.7
BNM [8]89.661.576.955.089.369.181.365.590.047.389.130.170.4
MSTN+DSBN [3]94.786.776.072.095.275.187.981.391.168.988.345.580.2
CGDM [14]92.885.176.364.591.093.281.379.392.483.085.644.880.8
SHOT [29]94.388.580.157.393.193.180.780.391.589.186.358.282.9
TVT [55]92.985.677.560.593.698.289.476.493.692.091.755.783.9
CDTrans [52]98.086.987.980.997.997.396.885.397.683.294.054.488.4
OT-basedJDOT [5]78.470.879.468.882.380.584.270.788.468.878.445.774.7
DeepJDOT [11]85.473.477.387.384.164.791.579.391.944.488.561.877.4
MLOT [23]88.270.477.350.284.877.280.474.483.868.282.338.773.0
RWOT [51]95.180.383.790.092.468.092.582.287.978.490.468.284.0
DANN+MMI+COT (Ours)96.989.684.274.196.496.588.682.096.094.185.162.187.1
Ours + CDTrans [52]98.289.487.682.398.097.296.486.298.392.692.258.189.7
" + }, + { + "type": "image", + "bbox": [ + 0.117, + 0.419, + 0.255, + 0.525 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.158, + 0.535, + 0.216, + 0.548 + ], + "angle": 0, + "content": "(a) RWOT" + }, + { + "type": "image", + "bbox": [ + 0.289, + 0.418, + 0.428, + 0.523 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.338, + 0.535, + 0.385, + 0.547 + ], + "angle": 0, + "content": "(b)COT" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.56, + 0.47, + 0.589 + ], + "angle": 0, + "content": "Figure 1. t-SNE of classifier responses by RWOT and COT on CI50 (red: Amazon, blue: Webcam)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.619, + 0.228, + 0.635 + ], + "angle": 0, + "content": "5.3. Ablation study" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.644, + 0.473, + 0.902 + ], + "angle": 0, + "content": "Ability on Resisting Class-imbalance Challenge. As the results illustrated above, our COT brings little gain than the existing sota OT-based UDA method (RWOT [51]) on the Office-31 dataset while achieving the dominant performance on Office-Home and VisDA-2017 datasets. To find the reason for this, we compute the label distribution of the Office-31 dataset, where the class imbalance phenomenon is not occurred obviously compared to the Office-Home and VisDA-2017 datasets (shown in Fig. 2). Thus, to verify the ability of our COT on resisting class-imbalance challenges among all UDA benchmark datasets, we construct 3 types of class-imbalance evaluation benchmarks by randomly erasing \\(30\\%\\), \\(50\\%\\), \\(70\\%\\) samples from certain classes on Office-31 datasets target domain, termed CI30, CI50, and CI70, respectively. Table 4 presents the results of our COT and existing OT-based UDA methods on Office-31, CI30, CI50, and CI70. According to this quantitative result, we dis" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.395, + 0.892, + 0.456 + ], + "angle": 0, + "content": "cover that the more severe the class imbalance phenomenon occurs, the larger the performance gap between our COT and RWOT, demonstrating that our COT embraces excellent ability on handling class-imbalance challenges." + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.466, + 0.892, + 0.494 + ], + "angle": 0, + "content": "Table 4. Mean accuracy of OT-based UDA methods on class-imbalance datasets." + }, + { + "type": "table", + "bbox": [ + 0.514, + 0.506, + 0.88, + 0.587 + ], + "angle": 0, + "content": "
DatasetCOTRWOTMLOT
Office-3191.090.887.8
CI3088.284.783.6
CI5087.778.380.4
CI7085.470.575.2
" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.599, + 0.892, + 0.719 + ], + "angle": 0, + "content": "Computation Cost. Here, we compute the computation cost of our COT and RWOT, where RWOT is a representative work of a sample-level OT matching mechanism. By looking at the results in Fig. 3, compared with RWOT, our COT significantly saves the computational cost with the increasing of training scale. Integrating the above theoretical analysis with this quantitative result demonstrates the superiority of our COT in economizing the computation head." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.892, + 0.87 + ], + "angle": 0, + "content": "Effects of Loss Weight. \\(\\alpha_{1}\\) and \\(\\alpha_{2}\\) are employed to weight the importance of cluster and OT loss, respectively. In our experiment, \\(\\alpha_{1}\\) and \\(\\alpha_{2}\\) are the same and selected from a broad range of \\(\\{0.1, 0.2, 0.4, 1.0, 2.0\\}\\). Results are evaluated on the Office-31 and shown in Table 5. When \\(\\alpha_{1}\\) and \\(\\alpha_{2}\\) are set to large value, the performance will tend to drop. In our opinion, the larger \\(\\alpha_{1}\\) and \\(\\alpha_{2}\\) would put more attention to local alignment while weakening the feature representation on both domains. As a result, \\(\\alpha_{1}\\) and \\(\\alpha_{2}\\) are set to 0.2, 0.2 for all datasets." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Effects of the Number of Sub-domains. \\( Y \\), defined in Section 4, represents the number of sub-domains for each" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "20004" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.11, + 0.115, + 0.331, + 0.239 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.195, + 0.248, + 0.266, + 0.26 + ], + "angle": 0, + "content": "(a) Office-31" + }, + { + "type": "image", + "bbox": [ + 0.365, + 0.115, + 0.585, + 0.238 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.44, + 0.248, + 0.531, + 0.26 + ], + "angle": 0, + "content": "(b) Office-Home" + }, + { + "type": "image", + "bbox": [ + 0.617, + 0.115, + 0.841, + 0.238 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.697, + 0.248, + 0.784, + 0.26 + ], + "angle": 0, + "content": "(c) VisDA-2017" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.272, + 0.892, + 0.3 + ], + "angle": 0, + "content": "Figure 2. Class distribution on Office-31, Office-Home and VisDA-2017. Note that the visualized class distribution is randomly selected from corresponding datasets." + }, + { + "type": "image", + "bbox": [ + 0.161, + 0.311, + 0.376, + 0.438 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.131, + 0.45, + 0.416, + 0.464 + ], + "angle": 0, + "content": "Figure 3. Computation cost of COT and RWOT" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.481, + 0.47, + 0.509 + ], + "angle": 0, + "content": "Table 5. Mean accuracy of our COT with different \\( \\alpha_{1} \\) and \\( \\alpha_{2} \\) on Office-31 dataset." + }, + { + "type": "table", + "bbox": [ + 0.085, + 0.52, + 0.462, + 0.616 + ], + "angle": 0, + "content": "
α1α2avg_acc
0.10.190.8
0.20.291.0
0.40.490.6
1.01.089.4
2.02.087.7
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.643, + 0.47, + 0.689 + ], + "angle": 0, + "content": "class. We investigate a broad range of \\( Y \\) to find an appropriate value. Based on the results shown in Table 6, we set \\( Y \\) to 2 for all test settings." + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.703, + 0.468, + 0.731 + ], + "angle": 0, + "content": "Table 6. Mean accuracy of our COT with different \\( Y \\) on Office-31 dataset." + }, + { + "type": "table", + "bbox": [ + 0.114, + 0.742, + 0.435, + 0.824 + ], + "angle": 0, + "content": "
Yavg_acc
190.7
291.0
390.8
590.4
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.47, + 0.903 + ], + "angle": 0, + "content": "Qualitative Analysis on COT As discussed above, we present 2 advantages of our COT when compared with existing OT-based methods [23,51]. In this part, we further illustrate another advantage compared to the clustering-based" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.298, + 0.895, + 0.465 + ], + "angle": 0, + "content": "OT algorithm in the conventional OT realm [24, 36]. Our COT is optimized in an end-to-end manner, thus the feature representation and optimal transportation map are both updated asynchronously. While for [24] and [36], they focus on obtaining offline cluster centers firstly and then compute the optimal transportation map. Such a synchronous optimization manner inevitably leads to a sub-optimal mapping result, because the computation procedure of an optimal transportation map is based on the fixed cluster centers. This is why we adopt an online clustering mode instead of an offline one." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.486, + 0.637, + 0.501 + ], + "angle": 0, + "content": "5.4. Visualization" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.514, + 0.892, + 0.635 + ], + "angle": 0, + "content": "We present the t-SNE visualization of feature representation for our COT and RWOT [51] on CI50. Under the challenge of class imbalance, the learned features from RWOT in some classes are ambiguous to provide accurate representation compared to our COT. These visualization results further demonstrate the excellent ability of our COT on handling class-imbalance challenges from the feature representation perspective." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.661, + 0.619, + 0.675 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.69, + 0.892, + 0.901 + ], + "angle": 0, + "content": "In this paper, we propose a novel method to advance the OT-based domain adaptation community, which integrates optimal transport theory with clustering operation, termed Clustering-based Optimal Transport (COT). Concretely, COT applies the loss derived from discrete Kantorovich dual form to cluster and align centers in the source and target domain, thus transferring knowledge from the source domain to the target domain. Moreover, our COT can eliminate the negative effect brought by the class imbalance phenomenon and reduce the computation cost simultaneously, which are two challenging problems for existing OT-based UDA algorithms. Besides, we also provide comprehensive theoretical analysis and experimental results to guarantee the advantages of our COT." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20005" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.116, + 0.486, + 0.158 + ], + "angle": 0, + "content": "[1135] Hana Ajakan, Pascal Germain, Hugo Larochelle, François Laviolette, and Mario Marchand. Domain-adversarial neural networks. arXiv preprint arXiv:1412.4446, 2014. 1, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.16, + 0.468, + 0.186 + ], + "angle": 0, + "content": "[2] M. Arjovsky, S. Chintala, and L. Bottou. Wasserstein gan. 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.189, + 0.468, + 0.231 + ], + "angle": 0, + "content": "[3] Woong-Gi Chang, Tackgeun You, Seonguk Seo, Suha Kwak, and Bohyung Han. Domain-specific batch normalization for unsupervised domain adaptation. In CVPR, 2019. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.233, + 0.468, + 0.301 + ], + "angle": 0, + "content": "[4] Xinyang Chen, Sinan Wang, Mingsheng Long, and Jianmin Wang. Transferability vs. discriminability: Batch spectral penalization for adversarial domain adaptation. In International conference on machine learning, pages 1081-1090. PMLR, 2019. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.304, + 0.468, + 0.358 + ], + "angle": 0, + "content": "[5] Nicolas Courty, Rémi Flamary, Amaury Habrard, and Alain Rakotomamonjy. Joint distribution optimal transportation for domain adaptation. arXiv preprint arXiv:1705.08848, 2017. 2, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.362, + 0.468, + 0.417 + ], + "angle": 0, + "content": "[6] Nicolas Courty, Rémi Flamary, and Devis Tuia. Domain adaptation with regularized optimal transport. In Joint European Conference on Machine Learning and Knowledge Discovery in Databases, pages 274-289. Springer, 2014. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.419, + 0.468, + 0.473 + ], + "angle": 0, + "content": "[7] Nicolas Courty, Rémi Flamary, Devis Tuia, and Alain Rakotomamonjy. Optimal transport for domain adaptation. IEEE transactions on pattern analysis and machine intelligence, 39(9):1853-1865, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.476, + 0.468, + 0.558 + ], + "angle": 0, + "content": "[8] Shuhao Cui, Shuhui Wang, Junbao Zhuo, Liang Li, Qingming Huang, and Qi Tian. Towards discriminability and diversity: Batch nuclear-norm maximization under label insufficient situations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3941-3950, 2020. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.562, + 0.468, + 0.629 + ], + "angle": 0, + "content": "[9] Shuhao Cui, Shuhui Wang, Junbao Zhuo, Chi Su, Qingming Huang, and Tian Qi. Gradually vanishing bridge for adversarial domain adaptation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2020. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.633, + 0.468, + 0.701 + ], + "angle": 0, + "content": "[10] Shuhao Cui, Shuhui Wang, Junbao Zhuo, Chi Su, Qingming Huang, and Qi Tian. Gradually vanishing bridge for adversarial domain adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12455-12464, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.704, + 0.468, + 0.771 + ], + "angle": 0, + "content": "[11] Bharath Bhushan Damodaran, Benjamin Kellenberger, Rémi Flamary, Devis Tuia, and Nicolas Courty. Deepjoint: Deep joint distribution optimal transport for unsupervised domain adaptation. In Proceedings of the European Conference on Computer Vision (ECCV), pages 447-463, 2018. 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.775, + 0.468, + 0.829 + ], + "angle": 0, + "content": "[12] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.832, + 0.468, + 0.9 + ], + "angle": 0, + "content": "[13] Jiankang Deng, Jia Guo, Niannan Xue, and Stefanos Zafeiriou. Arcface: Additive angular margin loss for deep face recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4690-4699, 2019. 1, 3" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.116, + 0.486, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.135 + ], + "angle": 0, + "content": "[14] Zhekai Du, Jingjing Li, Hongzu Su, Lei Zhu, and Ke Lu. Cross-domain gradient discrepancy minimization for unsupervised domain adaptation. 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.137, + 0.892, + 0.19 + ], + "angle": 0, + "content": "[15] Kilian Fatras, Thibault Séjourne, Rémi Flamary, and Nicolas Courty. Unbalanced minibatch optimal transport; applications to domain adaptation. In International Conference on Machine Learning, pages 3186-3197. PMLR, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.193, + 0.892, + 0.233 + ], + "angle": 0, + "content": "[16] Wilfrid Gangbo and Robert J McCann. The geometry of optimal transportation. Acta Mathematica, 177(2):113-161, 1996. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.235, + 0.892, + 0.29 + ], + "angle": 0, + "content": "[17] Yixiao Ge, Dapeng Chen, and Hongsheng Li. Mutual meanteaching: Pseudo label refinery for unsupervised domain adaptation on person re-identification. In International Conference on Learning Representations, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.292, + 0.892, + 0.331 + ], + "angle": 0, + "content": "[18] Yves Grandvalet, Yoshua Bengio, et al. Semi-supervised learning by entropy minimization. In CAP, pages 281-296, 2005. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.334, + 0.892, + 0.375 + ], + "angle": 0, + "content": "[19] Sahand Hajifar and Hongyue Sun. Online domain adaptation for continuous cross-subject liver viability evaluation based on irregular thermal data, 2020. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.377, + 0.892, + 0.404 + ], + "angle": 0, + "content": "[20] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. 2016. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.407, + 0.892, + 0.446 + ], + "angle": 0, + "content": "[21] Ying Jin, Ximei Wang, Mingsheng Long, and Jianmin Wang. Minimum class confusion for versatile domain adaptation, 2020. 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.449, + 0.892, + 0.475 + ], + "angle": 0, + "content": "[22] Leonid V Kantorovich. On the translocation of masses. Journal of mathematical sciences, 133(4):1381-1382, 2006. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.478, + 0.892, + 0.517 + ], + "angle": 0, + "content": "[23] Tanguy Kerdoncuff, Rémi Emonet, and Marc Sebban. Metric learning in optimal transport for domain adaptation. 2020. 1, 2, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.521, + 0.892, + 0.574 + ], + "angle": 0, + "content": "[24] John Lee, Max Dabagia, Eva Dyer, and Christopher Rozell. Hierarchical optimal transport for multimodal distribution alignment. Advances in Neural Information Processing Systems, 32, 2019. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.577, + 0.892, + 0.631 + ], + "angle": 0, + "content": "[25] Jingjing Li, Erpeng Chen, Zhengming Ding, Lei Zhu, Ke Lu, and Heng Tao Shen. Maximum density divergence for domain adaptation. IEEE Transactions on Pattern Analysis and Machine Intelligence, page 1-1, 2020. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.634, + 0.892, + 0.701 + ], + "angle": 0, + "content": "[26] Mengxue Li, Yi-Ming Zhai, You-Wei Luo, Peng-Fei Ge, and Chuan-Xian Ren. Enhanced transport distance for unsupervised domain adaptation. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 13933–13941, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.704, + 0.892, + 0.771 + ], + "angle": 0, + "content": "[27] Rui Li, Qianfen Jiao, Wenming Cao, Hau-San Wong, and Si Wu. Model adaptation: Unsupervised domain adaptation without source data. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9641-9650, 2020. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.774, + 0.892, + 0.829 + ], + "angle": 0, + "content": "[28] Shuang Li, Mixue Xie, Kaixiong Gong, Chi Harold Liu, Yulin Wang, and Wei Li. Transferable semantic augmentation for domain adaptation. arXiv preprint arXiv:2103.12562, 2021. 1, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.831, + 0.892, + 0.871 + ], + "angle": 0, + "content": "[29] Jian Liang, Dapeng Hu, and Jiashi Feng. Do we really need to access the source data? source hypothesis transfer for unsupervised domain adaptation. 2020. 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.874, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[30] Chi-Heng Lin, Mehdi Azabou, and Eva L Dyer. Making transport more robust and interpretable by moving data" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20006" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.092, + 0.472, + 0.12 + ], + "angle": 0, + "content": "through a small number of anchor points. Proceedings of machine learning research, 139:6631, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.121, + 0.47, + 0.189 + ], + "angle": 0, + "content": "[31] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dálár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In European conference on computer vision, pages 740–755. Springer, 2014. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.19, + 0.469, + 0.258 + ], + "angle": 0, + "content": "[32] Hong Liu, Mingsheng Long, Jianmin Wang, and Michael Jordan. Transferable adversarial training: A general approach to adapting deep classifiers. In International Conference on Machine Learning, pages 4013-4022. PMLR, 2019. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.26, + 0.469, + 0.302 + ], + "angle": 0, + "content": "[33] Mingsheng Long, Zhangjie Cao, Jianmin Wang, and Michael I Jordan. Conditional adversarial domain adaptation. arXiv preprint arXiv:1705.10667, 2017. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.302, + 0.469, + 0.357 + ], + "angle": 0, + "content": "[34] Mingsheng Long, Han Zhu, Jianmin Wang, and Michael I Jordan. Deep transfer learning with joint adaptation networks. In International conference on machine learning, pages 2208-2217. PMLR, 2017. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.359, + 0.469, + 0.426 + ], + "angle": 0, + "content": "[35] Zhihe Lu, Yongxin Yang, Xiatian Zhu, Cong Liu, Yi-Zhe Song, and Tao Xiang. Stochastic classifiers for unsupervised domain adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9111-9120, 2020. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.428, + 0.469, + 0.482 + ], + "angle": 0, + "content": "[36] Liang Mi, Wen Zhang, and Yalin Wang. Regularized wasserstein means for aligning distributional data. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 34, pages 5166-5173, 2020. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.484, + 0.469, + 0.524 + ], + "angle": 0, + "content": "[37] Gaspard Monge. Mémoire sur la théorie des déblais et des remblais. Histoire de l'Académie Royale des Sciences de Paris, 1781. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.525, + 0.469, + 0.58 + ], + "angle": 0, + "content": "[38] Khai Nguyen, Dang Nguyen, Quoc Nguyen, Tung Pham, Hung Bui, Dinh Phung, Trung Le, and Nhat Ho. On transportation of mini-batches: A hierarchical approach. arXiv preprint arXiv:2102.05912, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.581, + 0.469, + 0.622 + ], + "angle": 0, + "content": "[39] Yingwei Pan, Ting Yao, Yehao Li, Yu Wang, Chong-Wah Ngo, and Tao Mei. Transferrable prototypical networks for unsupervised domain adaptation, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.623, + 0.469, + 0.69 + ], + "angle": 0, + "content": "[40] Ievgen Redko, Nicolas Courty, Rémi Flamary, and Devis Tuia. Optimal transport for multi-source domain adaptation under target shift. In The 22nd International Conference on Artificial Intelligence and Statistics, pages 849-858. PMLR, 2019. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.692, + 0.469, + 0.76 + ], + "angle": 0, + "content": "[41] Hochang Rhee and Nam Ik Cho. Efficient and robust pseudolabeling for unsupervised domain adaptation. In 2019 Asia-Pacific Signal and Information Processing Association Annual Summit and Conference (APSIPA ASC), pages 980–985, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.762, + 0.469, + 0.83 + ], + "angle": 0, + "content": "[42] Kuniaki Saito, Kohei Watanabe, Yoshitaka Ushiku, and Tatsuya Harada. Maximum classifier discrepancy for unsupervised domain adaptation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3723-3732, 2018. 1, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.832, + 0.469, + 0.9 + ], + "angle": 0, + "content": "[43] Swami Sankaranarayanan, Yogesh Balaji, Carlos D Castillo, and Rama Chellappa. Generate to adapt: Aligning domains using generative adversarial networks. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 8503-8512, 2018. 5" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "[44] Filippo Santambrogio. Optimal transport for applied mathematicians. Birkhäuser, NY, 55(58-63):94, 2015. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.121, + 0.892, + 0.177 + ], + "angle": 0, + "content": "[45] Hui Tang, Ke Chen, and Kui Jia. Unsupervised domain adaptation via structurally regularized deep clustering. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8725-8735, 2020. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.179, + 0.892, + 0.234 + ], + "angle": 0, + "content": "[46] Eric Tzeng, Judy Hoffman, Kate Saenko, and Trevor Darrell. Adversarial discriminative domain adaptation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7167-7176, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.235, + 0.892, + 0.304 + ], + "angle": 0, + "content": "[47] Hao Wang, Yitong Wang, Zheng Zhou, Xing Ji, Dihong Gong, Jingchao Zhou, Zhifeng Li, and Wei Liu. Cosface: Large margin cosine loss for deep face recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5265-5274, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.305, + 0.892, + 0.36 + ], + "angle": 0, + "content": "[48] Qian Wang and Toby Breckon. Unsupervised domain adaptation via structured prediction based selective pseudolabeling. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 34, pages 6243-6250, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.362, + 0.892, + 0.403 + ], + "angle": 0, + "content": "[49] Wei Wang, Haojie Li, Zhengming Ding, and Zhihui Wang. Rethink maximum mean discrepancy for domain adaptation. arXiv preprint arXiv:2007.00689, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.405, + 0.892, + 0.445 + ], + "angle": 0, + "content": "[50] Pengfei Wei, Yiping Ke, Xinghua Qu, and Tze-Yun Leong. Subdomain adaptation with manifolds discrepancy alignment, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.447, + 0.892, + 0.516 + ], + "angle": 0, + "content": "[51] Renjun Xu, Pelen Liu, Liyan Wang, Chao Chen, and Jindong Wang. Reliable weighted optimal transport for unsupervised domain adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4394-4403, 2020. 1, 5, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.518, + 0.892, + 0.572 + ], + "angle": 0, + "content": "[52] Tongkun Xu, Weihua Chen, Pichao Wang, Fan Wang, Hao Li, and Rong Jin. Cdtrans: Cross-domain transformer for unsupervised domain adaptation. arXiv preprint arXiv:2109.06165, 2021. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.575, + 0.892, + 0.656 + ], + "angle": 0, + "content": "[53] Hongliang Yan, Yukang Ding, Peihua Li, Qilong Wang, Yong Xu, and Wangmeng Zuo. Mind the class weight bias: Weighted maximum mean discrepancy for unsupervised domain adaptation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 2272-2281, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.659, + 0.892, + 0.714 + ], + "angle": 0, + "content": "[54] Yuguang Yan, Wen Li, Hanrui Wu, Huaqing Min, Mingkui Tan, and Qingyao Wu. Semi-supervised optimal transport for heterogeneous domain adaptation. In IJCAI, volume 7, pages 2969-2975, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.715, + 0.892, + 0.757 + ], + "angle": 0, + "content": "[55] Jinyu Yang, Jingjing Liu, Ning Xu, and Junzhou Huang. Tvt: Transferable vision transformer for unsupervised domain adaptation. arXiv preprint arXiv:2108.05988, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.758, + 0.892, + 0.799 + ], + "angle": 0, + "content": "[56] Yabin Zhang, Hui Tang, Kui Jia, and Mingkui Tan. Domain-symmetric networks for adversarial domain adaptation, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.801, + 0.892, + 0.87 + ], + "angle": 0, + "content": "[57] Yongchun Zhu, Fuzhen Zhuang, Jindong Wang, Guolin Ke, Jingwu Chen, Jiang Bian, Hui Xiong, and Qing He. Deep subdomain adaptation network for image classification. IEEE Transactions on Neural Networks and Learning Systems, 32(4):1713-1722, Apr 2021. 1" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.87 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20007" + } + ] +] \ No newline at end of file diff --git a/2023/COT_ Unsupervised Domain Adaptation With Clustering and Optimal Transport/53ca0802-c684-4d48-b297-7731efdaab53_origin.pdf b/2023/COT_ Unsupervised Domain Adaptation With Clustering and Optimal Transport/53ca0802-c684-4d48-b297-7731efdaab53_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..ba3eaf2dbfb34b914264224a5c9d15b9b2e2aa5f --- /dev/null +++ b/2023/COT_ Unsupervised Domain Adaptation With Clustering and Optimal Transport/53ca0802-c684-4d48-b297-7731efdaab53_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6078db4acbb4e5580507e1f266928fc1f6f2988d0a303effa3e64b0e2e0c4f45 +size 518155 diff --git a/2023/COT_ Unsupervised Domain Adaptation With Clustering and Optimal Transport/full.md b/2023/COT_ Unsupervised Domain Adaptation With Clustering and Optimal Transport/full.md new file mode 100644 index 0000000000000000000000000000000000000000..70923cd71ddd659e62abcd97c5f132311966f388 --- /dev/null +++ b/2023/COT_ Unsupervised Domain Adaptation With Clustering and Optimal Transport/full.md @@ -0,0 +1,379 @@ +# COT: Unsupervised Domain Adaptation with Clustering and Optimal Transport + +Yang Liu $^{1*}$ Zhipeng Zhou $^{1*}$ Baigui Sun $^{1\dagger}$ 1Alibaba Group + +# Abstract + +Unsupervised domain adaptation (UDA) aims to transfer the knowledge from a labeled source domain to an unlabeled target domain. Typically, to guarantee desirable knowledge transfer, aligning the distribution between source and target domain from a global perspective is widely adopted in UDA. Recent researchers further point out the importance of local-level alignment and propose to construct instance-pair alignment by leveraging on Optimal Transport (OT) theory. However, existing OT-based UDA approaches are limited to handling class imbalance challenges and introduce a heavy computation overhead when considering a large-scale training situation. To cope with two aforementioned issues, we propose a Clustering-based Optimal Transport (COT) algorithm, which formulates the alignment procedure as an Optimal Transport problem and constructs a mapping between clustering centers in the source and target domain via an end-to-end manner. With this alignment on clustering centers, our COT eliminates the negative effect caused by class imbalance and reduces the computation cost simultaneously. Empirically, our COT achieves state-of-the-art performance on several authoritative benchmark datasets. + +# 1. Introduction + +Benefiting from the availability of large-scale data, deep learning has achieved tremendous success over the past few years. However, directly applying a well-trained convolution neural network on a new domain frequently suffers from the domain gap/discrepancy challenge, resulting in spurious predictions on the new domain. To remedy this, Unsupervised Domain Adaptation (UDA) has attracted many researchers' attention, which can transfer the knowledge from a labeled domain to an unlabeled domain. + +A major line of UDA approaches [1, 1, 28, 42, 49, 53] aim + +to learn a global domain shift by aligning the global source and target distribution while ignoring the local-level alignment between two domains. By leveraging on global domain adaptation, the global distributions of source and target domain are almost the same, thus losing the fine-grained information for each class (class-structure) on the source and target domain. + +Recently, to preserve class structure in both domains, several works [6, 15, 23, 30, 38, 40, 44, 51, 54] adopt optimal transport (OT) to minimize the sample-level transportation cost between source and target domain, achieving a significant performance on UDA. However, there exist two issues on recent OT-based UDA approaches. (i) When considering a realistic situation, i.e. the class imbalance $^{1}$ phenomenon occurs between the source and target domain, samples belonging to the same class in the target domain are assigned with false pseudo labels due to the mechanism of optimal transport, which requires each sample in source domain can be mapped to target samples. As a result, existing OT-based UDA methods provide poor pair-wise matching when facing class imbalance challenges. (ii) OT-based UDA methods tend to find a sample-level optimal counterpart, which requires a large amount of computation overhead, especially training on large-scale datasets. + +To solve two aforementioned issues, we propose a Clustering-based Optimal Transport algorithm, termed COT, to construct a clustering-level instead of sample-level mapping between source and target domain. Clusters in the source domain are obtained from the classifiers supervised by the labeled source domain data. While for the target domain, COT utilizes a set of learnable clusters to represent the feature distribution of the target domain, which can describe the sub-domain information [50, 57]. For instance, in many object recognition tasks [13, 20] an object could contain many attributes. Each attribute can be viewed as a sub-domain. To this end, the clusters on the source and target domain can represent the individual sub-domain information, respectively, such that optimal transport between clusters intrinsically provides a local mapping from the subdomain in the source domain to those in the target domain. Moreover, we provide a theoretical analysis and compre + +hensive experimental results to guarantee that (i) COT can alleviate the negative effect caused by class imbalance; (ii) Compared to existing OT-based UDA approaches, our COT economizes much computation head. + +In summary, our main contributions include: + +- We propose a novel Clustering-based Optimal Transport module as well as a specially designed loss derived from the discrete type of Kantorovich dual form, which resolves two aforementioned challenges on the existing OT-based UDA algorithms, facilitating the development of OT-based UDA community. +- We provide a theoretical analysis to guarantee the advantages of our COT. +- Our COT achieves state-of-the-art performance on several UDA benchmark datasets. + +# 2. Related work + +Pseudo Label based Domain Adaptation Inspired by the observation that samples in the target domain can be clustered within the feature space, for accurate pseudolabeling, [48] propose a selective pseudo-labeling strategy based on structural predictions which utilize the unsupervised clustering analysis. [41] introduce a confidence-based weighting scheme for obtaining pseudo-labels and an adaptive threshold adjustment strategy to provide sufficient and accurate pseudo-labels during the training process. The confidence-based weighting scheme generates pseudo-labels that can enable the performance less sensitive to the threshold determined by the pseudo-labels. In the task of person re-identification, [17] propose an unsupervised framework called Mutual Mean-Teaching to learn better features from the target domain by refining the hard pseudo labels offline and soft pseudo labels online alternatively to mitigate the effects of noisy pseudo labels caused by the clustering algorithms. + +Optimal Transport based Domain Adaptation As a way to find a minimal effort strategy to the transport of a given mass of dirt into a given hole, [37] put forward the optimal transport problem for the first time. [22] provide an extension of the original problem of Monge. Recently, by applying the optimal transport to domain adaptation, some new computation strategies have emerged. + +[7] propose regularized unsupervised optimal transport model to align the representation of features between different domains. The regularization schemes encoding class-structure in source domain during estimation of transport map enforce the intuition that samples of the same class must undergo a similar transformation. [5] minimize the optimal transport loss between the joint source distribution and the estimated target joint distribution depending on a function that is introduced to predict an output value given + +input from the source domain. For reducing the discrepancy between multiple domains, [40] propose Joint Class Proportion and Optimal Transport which performs multisource domain adaptation and target shift correction simultaneously by learning the predicted class probability of the unlabeled target data and the coupling to align the distributions between source and target domain. For better alignment between different domains, a relation between target error and the magnitude of different Wasserstein distances are proposed in [23] which optimize the metric for domain adaptation. + +# 3. Preliminary + +In this section, we introduce the basic knowledge for optimal transport. + +# 3.1. Optimal Transport + +Let $X \subseteq \mathbb{R}^d$ be a measurable space and the labels are denoted as $\mathcal{V}$ . We denote the set of all probability distributions on $X$ as $\mathcal{P}(X)$ . The source and target domains are space $X$ equipped with two distinct probability distributions $\mu_{\mathcal{S}}$ and $\mu_{\mathcal{T}}$ . Suppose we have source dataset $\{x_i^s\}_{i=1}^{n_s} \subset X_{\mathcal{S}} = (X, \mu_{\mathcal{S}})$ associated with label set $\{y_i^s\}_{i=1}^{n_s}$ where $y_i^s \in \mathcal{V}$ . The target dataset is $\{x_j^t\}_{j=1}^{n_t} \subset X_{\mathcal{T}} = (X, \mu_{\mathcal{T}})$ without labels. The goal of optimal transport is to minimize the inter-domain transportation cost by finding a feasible map to preserve measure. + +Definition 1 (Kantorovich) For given joint distribution $\rho(x^s, x^t)$ which satisfies for every measurable Borel set $O_S \subset X_S$ , $O_T \subset X_T$ , we have + +$$ +\rho \left(O _ {\mathcal {S}} \times X _ {\mathcal {T}}\right) = \mu_ {\mathcal {S}} \left(O _ {\mathcal {S}}\right), \rho \left(X _ {\mathcal {S}} \times O _ {\mathcal {T}}\right) = \mu_ {\mathcal {T}} \left(O _ {\mathcal {T}}\right) \tag {1} +$$ + +For convenience, we denote the projection maps from $X_{\mathcal{S}} \times X_{\mathcal{T}}$ to $X_{\mathcal{S}}$ and $X_{\mathcal{T}}$ as $\pi_{\mathcal{S}}, \pi_{\mathcal{T}}$ . The above equation can be denoted as $\pi_{\mathcal{S}\#} \rho = \mu_{\mathcal{S}}$ and $\pi_{\mathcal{T}\#} \rho = \mu_{\mathcal{T}}$ . The corresponding transportation cost is + +$$ +\mathcal {C} (\rho) = \int_ {X _ {\mathcal {S}} \times X _ {\mathcal {T}}} c \left(x ^ {s}, x ^ {t}\right) d \rho \left(x ^ {s}, x ^ {t}\right) \tag {2} +$$ + +where $c(x^s, x^t)$ is pointwise transportation cost between $x^s \in X_S$ and $x^t \in X_T$ . The optimal transport problem is proposed to minimize the $\mathcal{C}(\rho)$ under the measure preserving as the following: + +$$ +W _ {c} := \inf _ {\rho} \left\{\mathcal {C} (\rho) | | \pi_ {\mathcal {S} \#} \rho = \mu_ {\mathcal {S}}, \pi_ {\mathcal {T} \#} \rho = \mu_ {\mathcal {T}} \right\} \tag {3} +$$ + +By convex optimization theory, we can consider the Kantorovich's dual problem as: + +$$ +W _ {c} := \left\{ \begin{array}{l} \max _ {\varphi , \psi} \int_ {X _ {\mathcal {S}}} \varphi \left(x ^ {s}\right) d \mu_ {\mathcal {S}} + \int_ {X _ {\mathcal {T}}} \psi \left(x ^ {t}\right) d \mu_ {\mathcal {T}} \\ \text {s . t .} \quad \varphi \left(x ^ {s}\right) + \psi \left(x ^ {t}\right) \leq c \left(x ^ {s}, x ^ {t}\right) \end{array} \right. \tag {4} +$$ + +where $\varphi$ and $\psi$ are real functions from $X_{\mathcal{S}}$ and $X_{\mathcal{T}}$ to $\mathbb{R}$ . Moreover, the Kantorovich problem can be formulated as + +$$ +\max _ {\varphi} \left\{\int_ {X _ {\mathcal {S}}} \varphi \left(x ^ {s}\right) d \mu_ {\mathcal {S}} + \int_ {X _ {\mathcal {T}}} \varphi^ {c} \left(x ^ {t}\right) d \mu_ {\mathcal {T}} \right\} \tag {5} +$$ + +where $\varphi^c (x^t) = \inf_{x_s\in X_S}\{c(x^s,x^t) - \varphi (x^s)\}$ is called the $c$ -transform of $\varphi$ + +By classical optimal transport theory, different choices of cost function will influence the difficulty to solving the optimal transport problem. When we choose $c(x^s,x^t) = \| x^s -x^t\| _2$ , the problem stated in Equation (5) is equivalent to + +$$ +\max _ {\varphi} \left\{\int_ {X _ {\mathcal {S}}} \varphi \left(x ^ {s}\right) d \mu_ {\mathcal {S}} - \int_ {X _ {\mathcal {T}}} \varphi \left(x ^ {t}\right) d \mu_ {\mathcal {T}} \right\} \tag {6} +$$ + +where $\varphi$ is under the constraint that $|\varphi(x) - \varphi(x')| \leq \|x - x'\|_2$ . WGAN [2] is inspired by above cost setting, during the implementation of optimal transport in WGAN, they utilize the gradient clip to guarantee the Lipschitz constant of $\varphi$ is bounded from above by 1. When we set the cost function as $c(x^s, x^t) = \|x^s - x^t\|_2^2$ , [16] guarantees the existence and uniqueness of optimal transport map. + +# 4. Methodology + +In this section, we present the detail of our COT and demonstrate the advantages of our COT theoretically. + +# 4.1. Clustering-based Optimal Transport + +Instead of aligning instance-level features between source and target domain, we propose a novel clustering-based optimal transport (COT) module for unsupervised domain adaptation in this subsection. Firstly, we extract features from the source and target domain by ImageNet pretrained CNNs. Then we utilize learnable clusters to represent the sub-domains in the source and target domain, respectively. Finally, we apply a Kantorovich dual form-based loss to implement the optimal transport between clusters from both domains. + +Feature Extractor We utilize an ImageNet pretrained (without fully connected layers) CNNs (e.g. ResNet50/ResNet101) to extract features $\{x_{i}^{s}\}_{i = 1}^{n_{s}}$ and $\{x_j^t\}_{j = 1}^{n_t}$ from the source and target dataset respectively at the beginning of training process. Note that the distributions of features vary during the training phase. + +Clustering As for each sample from source domain, i.e., $x_{i}^{s}$ and corresponding ground-truth $y_{i}^{s} \in \mathcal{Y}$ . We denote the fully-connected layer which outputs the classification logits as $W = [w_{1}^{s},\dots,w_{|\mathcal{Y}|}^{s}]^{\top} \in \mathbb{R}^{|\mathcal{Y}|\times c}$ , where $|\mathcal{Y}|$ is the number of categories and $c$ is the number of feature channels. The predicted classification probability is $P(\hat{y}_i^s = v|x_i^s) = \frac{e^{x_i^s\top w_v^s}}{\sum_{u = 1}^{|\mathcal{Y}|}e^{x_i^s\top w_u^s}}$ . The corresponding cross-entropy loss is shown + +as follows: + +$$ +\mathcal {L} _ {\text {c r o s s - e n t r o p y}} = \frac {1}{b} \sum_ {i = 1} ^ {b} - y _ {i} ^ {s} \cdot \log \left(P \left(\hat {y} _ {i} ^ {s} \mid x _ {i} ^ {s}\right)\right) \tag {7} +$$ + +For the source domain, since $\{w_v^s\}_{v=1}^{|\mathcal{Y}|}$ [13, 47] have been shown effective for clustering representation, we take the classifiers $\{w_v^s\}_{v=1}^{|\mathcal{Y}|}$ as clusters for feature space of source domain. For the target domain, a set of learnable clusters termed as $\{w_u^t\}_{u=1}^K$ are proposed to represent the sub-domains, where $K = Q \cdot |\mathcal{Y}|$ is a hyper-parameter which stands for the number of sub-domains in the target domain, $Q$ is a positive integer which represents the number of sub-domains for each class. The motivation of introducing $Q$ is to preserve the sub-domain information in the target domain during the COT optimization procedure. For instance, if we set the same number of sub-domains on the source and target domain, the sub-domain information in the target domain would be seriously influenced by those in the source domain during the COT procedure, since the latter is optimized with hard supervision signal while the former is supervised by soft signal. On the contrary, $Q$ helps the target domain generate more sub-domains, which can preserve the sub-domain information in the target domain because not all of them are optimized during the COT phase. Empirically, the performance is superior when $Q$ is set to 2. $\{w_u^t \in \mathbb{R}^c\}_{u=(l-1)\cdot Q+1}^{l\cdot Q}$ represent the clusters for class $l$ , $\forall 1 \leq l \leq |\mathcal{Y}|$ . For each feature $x_j^t$ , we assign it to the closest cluster in $\{w_u^t\}_{u=1}^K$ . We utilize the $L_2$ distance to measure the distance between features and clusters and pull the features back to corresponding clusters. + +$$ +\left\{ \begin{array}{c} \mathcal {L} _ {\text {c l u s t e r}} = \sum_ {j = 1} ^ {n _ {t}} \| w _ {u *} ^ {t} - x _ {j} ^ {t} \| ^ {2} \\ s. t. u _ {j} ^ {*} = \underset {1 \leq u \leq K} {\arg \min } \left(\left\| w _ {u} ^ {t} - x _ {j} ^ {t} \right\| ^ {2}\right), \forall 1 \leq j \leq n _ {t} \end{array} \right. \tag {8} +$$ + +Optimal Transport With clusters $\{w_v^s\}_{v=1}^{|\mathcal{Y}|}$ and $\{w_u^t\}_{u=1}^K$ from source and target domain respectively, we design the clustering based optimal transport as follows: + +$$ +\left\{ \begin{array}{c} \min _ {T \in \mathbb {R} ^ {| \mathcal {Y} | \times K}} T _ {v u} c _ {v u} \\ s. t. \sum_ {u = 1} ^ {K} T _ {v u} = \frac {1}{K}, \sum_ {v = 1} ^ {| \mathcal {Y} |} T _ {v u} = \frac {1}{| \mathcal {Y} |}, \\ T _ {v u} \geq 0, \forall 1 \leq v \leq | \mathcal {Y} |, 1 \leq u \leq K \end{array} \right. \tag {9} +$$ + +where $c_{vu} = \| w_v^s - w_u^t\|_2^2$ . Similar to Equation (5), we can get the discrete Kantorovich dual problem of Equation (9). + +$$ +\max _ {\psi} \left\{\frac {1}{K} \sum_ {u = 1} ^ {K} \psi \left(w _ {u} ^ {t}\right) + \frac {1}{| \mathcal {Y} |} \sum_ {v = 1} ^ {| \mathcal {Y} |} \psi^ {c} \left(w _ {v} ^ {s}\right) \right\} \tag {10} +$$ + +# Algorithm 1 Clustering-based Optimal Transport + +1: Set number of epochs for training as $E$ , learnable clusters for target domain as $\{w_u^t\}_{u=1}^{|\mathcal{Y}|}$ , classifiers/clusters for source domain as $\{w_v^s\}_{v=1}^{|\mathcal{Y}|}$ ; +2: for k-th training epoch while $k \leq E$ do +3: for $t$ -th iteration in $k$ -th epoch do +4: Take mini-batch of samples from source and target domain as input for feature extractor CNNs with parameters $\theta$ , the output features are $\{x_i^s\}_{i=1}^b$ and $\{x_j^t\}_{j=1}^b$ ; +5: Compute the $L_{\mathrm{cluster}}$ for $\{x_j^t\}_{j=1}^b$ , $L_{\mathrm{cross-entropy}}$ for $\{x_i^s\}_{i=1}^b$ in the $l$ -th batch, and $\mathcal{L}_{OT}$ ; +6: if $1 \leq t \leq k / (b * 10)$ then +7: we find the current optimal map from clusters of source domain to those of target domain by maximizing $L_{\mathrm{OT}}$ +8: end if +9: Minimize $\mathcal{L}_{COT}$ +10: end for +11: end for + +where $\psi^c (w_v^s) = \inf_{u = 1}^{K}(c_{vu} - \psi (w_u^t))$ . According to Equation 10, We can seek for the optimal transportation map between clusters by optimizing the following loss (line 7 in Algorithm 1): + +$$ +\mathcal {L} _ {\mathrm {O T}} = \frac {1}{K} \sum_ {u = 1} ^ {K} \lambda_ {u} ^ {t} + \frac {1}{| \mathcal {Y} |} \sum_ {v = 1} ^ {| \mathcal {Y} |} \left(\inf _ {u = 1} ^ {K} \left(c _ {v u} - \lambda_ {u} ^ {t}\right)\right) \tag {11} +$$ + +where $\{\lambda_u^t\}_{u = 1}^K$ represent the value of function $\psi$ at points $\{w_{u}^{t}\}_{u = 1}^{K}$ . Furthermore, it is worth noting that cost $c_{vu}$ is frozen during the optimization of $\mathcal{L}_{\mathrm{OT}}$ , indicating that only $\{\lambda_v^t\}_{v = 1}^{|\mathcal{V}|}$ is updated in this step. Then based on the founded optimal map, we minimize the following loss $(\mathcal{L}_{\mathrm{COT}})$ to close the matched clusters from the feature representation perspective, thus the domain-invariant feature representation is learned via this step (line 9 in Algorithm 1). Correspondingly, we freeze the $\{\lambda_v^t\}_{v = 1}^{|\mathcal{V}|}$ and only update $c_{vu}$ in this step. + +$$ +\mathcal {L} _ {\mathrm {C O T}} = \mathcal {L} _ {\text {c r o s s - e n t r o p y}} + \alpha_ {1} \cdot \mathcal {L} _ {\text {c l u s t e r}} + \alpha_ {2} \cdot \mathcal {L} _ {\mathrm {O T}} \tag {12} +$$ + +where $\alpha_{1}$ and $\alpha_{2}$ are loss weights. Finally, we elaborate the overall optimization strategy of our COT in Algorithm 1. Line 6 of Algorithm 1 represents we only optimize $\mathcal{L}_{OT}$ in the initial $10\%$ iterations of each epoch. + +# 4.2. Theoretical Analysis on Instance/Clustering Optimal Transport + +Given features $\{x_i^s\}_{i = 1}^{n_s}$ and $\{x_j^t\}_{j = 1}^{n_t}$ from source and target domain respectively, where $x_{i}^{s}$ and $x_{j}^{t}$ are output from + +shared-parameters neural network for feature extractor. We consider the discrete Kantorovich problem + +$$ +\left\{ \begin{array}{c} \min _ {T \in \mathbb {R} ^ {n _ {s} \times n _ {t}}} T _ {i j} c _ {i j} \\ s. t. \sum_ {j = 1} ^ {n _ {t}} T _ {i j} = \frac {1}{n _ {s}}, \sum_ {i = 1} ^ {n _ {s}} T _ {i j} = \frac {1}{n _ {t}}, \\ T _ {i j} \geq 0, \forall 1 \leq i \leq n _ {s}, 1 \leq j \leq n _ {t}. \end{array} \right. \tag {13} +$$ + +where $c_{ij} = \| x_i^s -x_j^t\| _2^2$ + +Considering the distance and inner-product between features and classifiers: + +$$ +\begin{array}{l} \left\| x _ {i} ^ {s} - w _ {v _ {1} ^ {s}} \right\| _ {2} ^ {2} - \left\| x _ {i} ^ {s} - w _ {v _ {2}} ^ {s} \right\| _ {2} ^ {2} \\ = \| w _ {v _ {1}} ^ {s} \| _ {2} ^ {2} - \| w _ {v _ {2}} ^ {s} \| _ {2} ^ {2} \tag {14} \\ + 2 \big (\| w _ {v _ {2}} ^ {s} \| _ {2} x _ {i} ^ {s} ^ {\top} \frac {w _ {v _ {2}} ^ {s}}{\| w _ {v _ {2}} ^ {s} \| _ {2}} - \| w _ {v _ {1}} ^ {s} \| _ {2} x _ {i} ^ {s} ^ {\top} \frac {w _ {v _ {1}} ^ {s}}{\| w _ {v _ {1}} ^ {s} \| _ {2}} \big) \\ \end{array} +$$ + +In the Bayesian view, we can consider $\| w_v^s\| _2$ as the prior probability of class $v$ , $x_{i}^{s}$ is feature representation of a sample and $\frac{w_s^s}{\|w_v^s\|_2}$ is the cluster for class $v$ . $x_{i}^{s\top}\frac{w_{v}^{s}}{\|w_{v}^{s}\|_{2}}$ measures the similarity between feature and cluster. When classifiers in $\{\| w_v^s\| _2\}_{v = 1}^{\lfloor \mathcal{V}\rfloor}$ are of the same magnitude, we conclude that the similarity between features and clusters are almost equivalent to the distance between features and classifiers. With labels as supervision, the optimization of crossentropy can promote the inter-class discrepancy which implies + +$$ +x _ {i} ^ {s \top} \frac {w _ {y _ {i} ^ {s}}}{\| w _ {y _ {i} ^ {s}} \| _ {2}} \gg x _ {i} ^ {s \top} \frac {w _ {v} ^ {s}}{\| w _ {v} ^ {s} \| _ {2}}, \forall v \neq y _ {i} ^ {s} \tag {15} +$$ + +which also provide the following result + +$$ +\left\| x _ {i} ^ {s} - w _ {y _ {i} ^ {s}} ^ {s} \right\| _ {2} ^ {2} \ll \left\| x _ {i} ^ {s} - w _ {v} ^ {s} \right\| _ {2} ^ {2}, \forall v \neq y _ {i} ^ {s} \tag {16} +$$ + +If clustering doesn't work sufficiently well, it happens that some samples in the source domain with label $v$ are assigned to samples in the target domain with label $u \neq v$ . When clustering performs well, We have $c_{ij} \sim \bar{c}_{vj} = \| w_v^s - x_j^t\|_2^2$ , where $\sim$ means these two numbers are almost the same. Then we get + +$$ +\sum_ {i, j} T _ {i j} c _ {i j} \sim \sum_ {v, j} \left(\sum_ {x _ {i} ^ {s} \in X _ {v} ^ {s}} T _ {i j}\right) \bar {c} _ {v j} \tag {17} +$$ + +where $X_{v}^{s}$ is the set of samples with label $v$ in the source domain. We denote the number of samples with class $v$ as $n_v$ , then we get + +$$ +\sum_ {j = 1} ^ {n _ {t}} \left(\sum_ {x _ {i} ^ {s} \in X _ {v} ^ {s}} T _ {i j}\right) = \frac {n _ {v}}{n _ {s}}, \sum_ {v = 1} ^ {| \mathcal {Y} |} \left(\sum_ {x _ {i} ^ {s} \in X _ {v} ^ {s}} T _ {i j}\right) = \frac {1}{n _ {t}} \tag {18} +$$ + +Table 156 Accuracy (%) on Office-31 for UDA (ResNet-50).The best result is in bold. + +
MethodA→WD→WW→DA→DD→AW→AAvg
Common UDAADDA [46]86.296.298.477.869.568.982.9
JAN [34]85.497.499.884.768.670.084.3
MCD [42]88.698.5100.092.269.569.786.5
GTA [43]89.597.999.887.772.871.486.5
CDAN [33]94.198.6100.092.971.069.387.7
TAT [32]92.599.3100.093.273.172.188.4
MDD [25]94.598.4100.093.574.672.288.9
GSP [19]92.998.799.894.575.974.989.5
DANN [1]82.096.999.179.768.267.482.2
SHOT [29]94.090.174.798.474.399.988.6
MCC [21]95.598.6100.094.472.974.989.4
GVB-GD [9]94.898.7100.095.073.473.789.3
TSA [28]96.098.7100.095.476.776.890.6
SRDC [45]95.799.2100.095.876.777.190.8
OT-basedJDOT [5]84.797.8100.086.464.467.783.5
DeepJDOT [11]88.998.599.688.272.170.186.2
MLOT [23]92.898.5100.090.872.871.687.8
RWOT [51]95.199.5100.094.577.577.990.8
DANN [1] + MMI [28]95.298.6100.094.474.675.289.7
DANN + MMI + COT (Ours)96.599.1100.096.176.777.491.0
+ +Inspired by Equation (17) and (18), we consider the following optimal transport between clusters from the source domain and instances from the target domain instead of solving the Kantorovich problem in Equation (13). + +$$ +\left\{ \begin{array}{c} \min _ {\bar {T} \in \mathbb {R} ^ {| Y | \times n _ {t}}} \bar {T} _ {v j} \bar {c} _ {v j} \\ s. t. \sum_ {j = 1} ^ {n _ {t}} \bar {T} _ {v j} = \frac {n _ {v}}{n _ {s}}, \sum_ {v = 1} ^ {| \mathcal {Y} |} \bar {T} _ {v j} = \frac {1}{n _ {t}}, \\ \bar {T} _ {v j} \geq 0, \forall 1 \leq v \leq | \mathcal {Y} |, 1 \leq j \leq n _ {t} \end{array} \right. \tag {19} +$$ + +In general, because of the class imbalance, the empirical label distribution between source and target dataset are different + +$$ +\exists \delta > 0, \text {s . t .} \| \left(\frac {n _ {1} ^ {s}}{n _ {s}}, \dots , \frac {n _ {| \mathcal {Y} |} ^ {s}}{n _ {s}}\right) - \left(\frac {n _ {1} ^ {t}}{n _ {t}}, \dots , \frac {n _ {| \mathcal {Y} |} ^ {t}}{n _ {t}}\right) \| _ {2} \geq \delta \tag {20} +$$ + +where $\delta$ is a constant which measures the label distribution between source and target domain. There must exist some index $i$ such that $\frac{n_v^s}{n_s} >\frac{n_v^t}{n_t}$ , which means that some samples with label $v$ in the source domain will be assigned to samples in the target domain with label $u\neq v$ . This will result in samples belonging to the same category in the target domain are given different pseudo labels, which increase the difficulty of training and cause the degradation of the performance of deep learning methods on the target domain. + +When we utilize the clustering based optimal transport, for source domain, we have $\sum_{u=1}^{K} T_{vu} = \frac{1}{|\mathcal{V}|}$ . For target domain, $\sum_{v=1}^{|\mathcal{V}|} \sum_{u=(l-1) \cdot Q+1}^{l \cdot Q} T_{vu} = \frac{Q}{K} = \frac{1}{|\mathcal{V}|}$ , which ease the + +negative effect from class imbalance in domain adaptation based on optimal transport. + +# 4.3. Computation Cost + +In terms of instance-based optimal transport, firstly we need to obtain the features of all samples from the source and target domain, computation cost on the feature extractor is shown as follows: + +$$ +\mathcal {O} \left(n _ {s} + n _ {t}\right) \cdot \mathcal {O} (\text {f e a t u r e - e x t r a c t o r}) \tag {21} +$$ + +where $\mathcal{O}$ (feature-extractor) means the computation cost on single sample when extracting the feature. Then considering the optimization of optimal transport, every iteration will need $\mathcal{O}(n_s\cdot n_t)$ . In comparison, for cluster-based optimal transport, the main computation cost on optimal transport is $\mathcal{O}(|\mathcal{Y}|\cdot K)$ . For a large-scale dataset, the clustering-based optimal transport cost is much less than instance-based optimal transport. + +# 5. Experiments + +In this section, we first elaborate on implementation details and 3 authoritative benchmark datasets in the field of UDA. Then we compare our COT with existing OT-based UDA algorithms and state-of-the-art UDA methods, illustrating that our COT achieves the dominant result in the field of OT-based UDA and competitive performance on the common UDA realm, respectively. Furthermore, we provide some qualitative analysis to illustrate the advantage of our COT when compared with some clustering-based DA + +Table 256 Accuracy $(\%)$ on Office-Home for UDA.The best result is in bold. + +
MethodAr→ClAr→PrAr→RwCl→ArCl→PrCl→RwPr→ArPr→ClPr→RwRw→ArRw→ClRw→PrAvg
Common UDAJAN [34]45.961.268.950.459.761.045.843.470.363.952.476.858.3
TAT [32]51.669.575.459.469.568.659.550.576.870.956.681.665.8
TPN [39]51.271.276.065.172.972.855.448.976.570.953.480.466.2
ETD [26]51.371.985.757.669.273.757.851.279.370.257.582.167.3
SymNets [56]47.772.978.564.271.374.264.248.879.574.552.682.767.6
BNM [8]52.373.980.063.372.974.961.749.579.770.553.682.267.9
MDD [25]54.973.777.860.071.471.861.253.678.172.560.282.368.1
GSP [19]56.875.578.961.369.474.961.352.679.973.354.283.268.4
MCD [42]48.968.374.661.367.668.85747.175.169.152.279.667.8
DANN [1]45.659.370.147.058.560.946.143.768.563.251.876.857.6
CDAN [33]50.770.676.057.670.070.057.450.977.370.956.781.665.8
BSP [4]52.068.676.158.070.370.258.650.277.672.259.381.966.3
TSA [28]55.873.779.061.974.674.560.753.280.172.758.484.369.1
GVB-GD [9]57.074.779.864.674.174.665.255.181.074.659.784.370.4
SRDC [45]52.376.381.069.576.278.068.753.881.776.357.185.071.3
OT-basedJDOT [5]44.763.468.655.164.460.754.348.472.867.457.269.460.5
DeepJDOT [11]39.750.462.539.554.453.236.739.263.652.345.470.550.6
MLOT [23]52.472.677.858.872.472.853.250.876.570.954.482.366.2
DANN [1] + MMI [28]55.774.880.463.674.176.664.254.780.574.358.281.369.9
DANN + MMI + COT (Ours)57.675.283.267.876.275.765.456.282.475.160.784.771.7
Ours + CDTrans [52]70.186.386.482.288.689.177.265.789.283.267.489.881.3
+ +methods in the Optimal Transport realm. Finally, we conduct ablative experiments together with visualization results to further demonstrate the effectiveness of our COT. + +# 5.1. Datasets and Implementation Details + +Office-31 is a benchmark dataset on the real-world unsupervised domain adaptation. It has 4110 images for 31 classes drawn from three domains: Amazon (A), DSLR (D), and Webcam (W). The 31 classes in the dataset consist of objects that commonly appeared in office settings, such as keyboards, file cabinets, and laptops. + +Office-Home is a challenging benchmark dataset for domain adaptation which has 4 domains where each domain consists of 65 categories. The four domains are Art - artistic images in the form of sketches, paintings, ornamentation, etc.; Clipart - a collection of clipart images; Product - images of objects without a background and Real-World - images of objects captured with a regular camera. It contains 15,500 images in 65 classes. + +VisDa-2017 is a large-scale simulation-to-real dataset for domain adaptation, which has over 280,000 images across 12 categories in the training, validation, and testing domains. The training images are generated from the same object under different circumstances, while the validation images are collected from MSCOCO ([31]). + +Implementation Details Follow GVB ([10]), we adopt ResNet-50 pretrained on the ImageNet ([12]) as our backbone for Office-31 and Office-Home benchmarks and ResNet-101 for VisDa-2017 dataset. Note that our COT is a plug-and-play module, indicating it can integrate with a vast body of existing UDA approaches, e.g. DANN [1], CDAN [33]. If there is no extra statement, our COT is implemented on DANN with Maximization Mutual Information Loss (MMI) [28]. In this paper, all experiments are + +implemented by PyTorch. For the optimizer schedule, we adopt SGD with a momentum of 0.9. The total number of training epochs is 40,40,30 on Office-31, Office-Home, and VisDA-2017. + +# 5.2. Results + +Results on Office-Home. Table 2 presents the results of OT-based and common UDA methods on Office-Home dataset. Our COT achieves the highest accuracy $71.7\%$ , outperforming other OT-based UDA algorithms by $1.8\%$ at least. Such tremendous improvements demonstrate that our COT can capture a more accurate local-level alignment than previous DA-based methods'. Besides, when compared with other state-of-the-art common UDA methods, to the best of our knowledge, our COT is the first OT-based UDA work that can perform a significant competitive result on the Office-Home dataset. + +Results on Office-31. The results are reported in the Table 1. Our COT achieves the best performance $(91.0\%)$ both on the common and OT-based UDA. + +Results on VisDA-2017. Table 3 shows the performance of OT-based and common UDA methods on the challenging VisDa-2017 dataset. Our COT outperforms the existing OT-based / common UDA methods (CNN backbone) by $3.1\%$ / $3.2\%$ at least, illustrating the dominant role of our COT on UDA task. Furthermore, due to the plug-and-play property of our COT, we directly add this module into the sota UDA method (CDTrans [52] on the challenging dataset VisDa2017 and Office-Home. The results in Table 3 show that our method achieves a new state-of-the-art performance, indicating the superiority and appealing potential of our COT. + +Table 356 Accuracy (%) on VisDA-2017 for UDA (ResNet-101). The best result is in bold. + +
Methodplanebcyblbuscarhorseknifemcylepersnplantsktbtraintruckmean
Common UDADANN [1]81.977.782.844.381.229.565.128.651.954.682.87.857.4
MinEnt [18]87.455.075.363.887.443.689.372.582.978.685.627.470.7
TSA [28]93.077.882.250.889.928.077.170.085.280.086.143.071.9
BSP [4]92.272.583.847.587.054.086.872.480.666.984.537.172.1
MCC [21]90.479.872.355.190.586.886.680.094.276.990.049.679.4
MODEL [27]94.873.468.874.893.195.488.684.789.184.783.548.181.6
STAR [35]95.084.084.673.091.691.885.978.494.484.787.042.282.7
BNM [8]89.661.576.955.089.369.181.365.590.047.389.130.170.4
MSTN+DSBN [3]94.786.776.072.095.275.187.981.391.168.988.345.580.2
CGDM [14]92.885.176.364.591.093.281.379.392.483.085.644.880.8
SHOT [29]94.388.580.157.393.193.180.780.391.589.186.358.282.9
TVT [55]92.985.677.560.593.698.289.476.493.692.091.755.783.9
CDTrans [52]98.086.987.980.997.997.396.885.397.683.294.054.488.4
OT-basedJDOT [5]78.470.879.468.882.380.584.270.788.468.878.445.774.7
DeepJDOT [11]85.473.477.387.384.164.791.579.391.944.488.561.877.4
MLOT [23]88.270.477.350.284.877.280.474.483.868.282.338.773.0
RWOT [51]95.180.383.790.092.468.092.582.287.978.490.468.284.0
DANN+MMI+COT (Ours)96.989.684.274.196.496.588.682.096.094.185.162.187.1
Ours + CDTrans [52]98.289.487.682.398.097.296.486.298.392.692.258.189.7
+ +![](images/808b1059f7ca154c322db66d294d8ce68f07873d475e1a1d2cc9ca18626644e4.jpg) +(a) RWOT + +![](images/39b0392be670efa9997417b0c3a9d2a55ae9b92433ea01c16e37cb857b0cf4b4.jpg) +(b)COT +Figure 1. t-SNE of classifier responses by RWOT and COT on CI50 (red: Amazon, blue: Webcam). + +# 5.3. Ablation study + +Ability on Resisting Class-imbalance Challenge. As the results illustrated above, our COT brings little gain than the existing sota OT-based UDA method (RWOT [51]) on the Office-31 dataset while achieving the dominant performance on Office-Home and VisDA-2017 datasets. To find the reason for this, we compute the label distribution of the Office-31 dataset, where the class imbalance phenomenon is not occurred obviously compared to the Office-Home and VisDA-2017 datasets (shown in Fig. 2). Thus, to verify the ability of our COT on resisting class-imbalance challenges among all UDA benchmark datasets, we construct 3 types of class-imbalance evaluation benchmarks by randomly erasing $30\%$ , $50\%$ , $70\%$ samples from certain classes on Office-31 datasets target domain, termed CI30, CI50, and CI70, respectively. Table 4 presents the results of our COT and existing OT-based UDA methods on Office-31, CI30, CI50, and CI70. According to this quantitative result, we dis + +cover that the more severe the class imbalance phenomenon occurs, the larger the performance gap between our COT and RWOT, demonstrating that our COT embraces excellent ability on handling class-imbalance challenges. + +Table 4. Mean accuracy of OT-based UDA methods on class-imbalance datasets. + +
DatasetCOTRWOTMLOT
Office-3191.090.887.8
CI3088.284.783.6
CI5087.778.380.4
CI7085.470.575.2
+ +Computation Cost. Here, we compute the computation cost of our COT and RWOT, where RWOT is a representative work of a sample-level OT matching mechanism. By looking at the results in Fig. 3, compared with RWOT, our COT significantly saves the computational cost with the increasing of training scale. Integrating the above theoretical analysis with this quantitative result demonstrates the superiority of our COT in economizing the computation head. + +Effects of Loss Weight. $\alpha_{1}$ and $\alpha_{2}$ are employed to weight the importance of cluster and OT loss, respectively. In our experiment, $\alpha_{1}$ and $\alpha_{2}$ are the same and selected from a broad range of $\{0.1, 0.2, 0.4, 1.0, 2.0\}$ . Results are evaluated on the Office-31 and shown in Table 5. When $\alpha_{1}$ and $\alpha_{2}$ are set to large value, the performance will tend to drop. In our opinion, the larger $\alpha_{1}$ and $\alpha_{2}$ would put more attention to local alignment while weakening the feature representation on both domains. As a result, $\alpha_{1}$ and $\alpha_{2}$ are set to 0.2, 0.2 for all datasets. + +Effects of the Number of Sub-domains. $Y$ , defined in Section 4, represents the number of sub-domains for each + +![](images/fb6d561851bbfd5ede1e570d6e0345edb266678544eb7e5e269c84179527bdce.jpg) +(a) Office-31 + +![](images/dd567420a6c5c3048c62f644a724b8751e00bdb499bb01e14a007fd68a4a1be7.jpg) +(b) Office-Home + +![](images/6b349db665b116784e576912a6477a04dd180ee24790c94a95edecfcee1dc072.jpg) +(c) VisDA-2017 + +![](images/65737c917147038b2e2e94e1cb82ad1fe7f20e462da8a4b312a56d52b90bc40a.jpg) +Figure 2. Class distribution on Office-31, Office-Home and VisDA-2017. Note that the visualized class distribution is randomly selected from corresponding datasets. +Figure 3. Computation cost of COT and RWOT + +Table 5. Mean accuracy of our COT with different $\alpha_{1}$ and $\alpha_{2}$ on Office-31 dataset. + +
α1α2avg_acc
0.10.190.8
0.20.291.0
0.40.490.6
1.01.089.4
2.02.087.7
+ +class. We investigate a broad range of $Y$ to find an appropriate value. Based on the results shown in Table 6, we set $Y$ to 2 for all test settings. + +Table 6. Mean accuracy of our COT with different $Y$ on Office-31 dataset. + +
Yavg_acc
190.7
291.0
390.8
590.4
+ +Qualitative Analysis on COT As discussed above, we present 2 advantages of our COT when compared with existing OT-based methods [23,51]. In this part, we further illustrate another advantage compared to the clustering-based + +OT algorithm in the conventional OT realm [24, 36]. Our COT is optimized in an end-to-end manner, thus the feature representation and optimal transportation map are both updated asynchronously. While for [24] and [36], they focus on obtaining offline cluster centers firstly and then compute the optimal transportation map. Such a synchronous optimization manner inevitably leads to a sub-optimal mapping result, because the computation procedure of an optimal transportation map is based on the fixed cluster centers. This is why we adopt an online clustering mode instead of an offline one. + +# 5.4. Visualization + +We present the t-SNE visualization of feature representation for our COT and RWOT [51] on CI50. Under the challenge of class imbalance, the learned features from RWOT in some classes are ambiguous to provide accurate representation compared to our COT. These visualization results further demonstrate the excellent ability of our COT on handling class-imbalance challenges from the feature representation perspective. + +# 6. Conclusion + +In this paper, we propose a novel method to advance the OT-based domain adaptation community, which integrates optimal transport theory with clustering operation, termed Clustering-based Optimal Transport (COT). Concretely, COT applies the loss derived from discrete Kantorovich dual form to cluster and align centers in the source and target domain, thus transferring knowledge from the source domain to the target domain. Moreover, our COT can eliminate the negative effect brought by the class imbalance phenomenon and reduce the computation cost simultaneously, which are two challenging problems for existing OT-based UDA algorithms. Besides, we also provide comprehensive theoretical analysis and experimental results to guarantee the advantages of our COT. + +# References + +[1135] Hana Ajakan, Pascal Germain, Hugo Larochelle, François Laviolette, and Mario Marchand. Domain-adversarial neural networks. arXiv preprint arXiv:1412.4446, 2014. 1, 5, 6, 7 +[2] M. Arjovsky, S. Chintala, and L. Bottou. Wasserstein gan. 2017. 3 +[3] Woong-Gi Chang, Tackgeun You, Seonguk Seo, Suha Kwak, and Bohyung Han. Domain-specific batch normalization for unsupervised domain adaptation. In CVPR, 2019. 7 +[4] Xinyang Chen, Sinan Wang, Mingsheng Long, and Jianmin Wang. Transferability vs. discriminability: Batch spectral penalization for adversarial domain adaptation. In International conference on machine learning, pages 1081-1090. PMLR, 2019. 6, 7 +[5] Nicolas Courty, Rémi Flamary, Amaury Habrard, and Alain Rakotomamonjy. Joint distribution optimal transportation for domain adaptation. arXiv preprint arXiv:1705.08848, 2017. 2, 5, 6, 7 +[6] Nicolas Courty, Rémi Flamary, and Devis Tuia. Domain adaptation with regularized optimal transport. In Joint European Conference on Machine Learning and Knowledge Discovery in Databases, pages 274-289. Springer, 2014. 1 +[7] Nicolas Courty, Rémi Flamary, Devis Tuia, and Alain Rakotomamonjy. Optimal transport for domain adaptation. IEEE transactions on pattern analysis and machine intelligence, 39(9):1853-1865, 2016. 2 +[8] Shuhao Cui, Shuhui Wang, Junbao Zhuo, Liang Li, Qingming Huang, and Qi Tian. Towards discriminability and diversity: Batch nuclear-norm maximization under label insufficient situations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3941-3950, 2020. 6, 7 +[9] Shuhao Cui, Shuhui Wang, Junbao Zhuo, Chi Su, Qingming Huang, and Tian Qi. Gradually vanishing bridge for adversarial domain adaptation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2020. 5, 6 +[10] Shuhao Cui, Shuhui Wang, Junbao Zhuo, Chi Su, Qingming Huang, and Qi Tian. Gradually vanishing bridge for adversarial domain adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12455-12464, 2020. 6 +[11] Bharath Bhushan Damodaran, Benjamin Kellenberger, Rémi Flamary, Devis Tuia, and Nicolas Courty. Deepjoint: Deep joint distribution optimal transport for unsupervised domain adaptation. In Proceedings of the European Conference on Computer Vision (ECCV), pages 447-463, 2018. 5, 6, 7 +[12] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 6 +[13] Jiankang Deng, Jia Guo, Niannan Xue, and Stefanos Zafeiriou. Arcface: Additive angular margin loss for deep face recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4690-4699, 2019. 1, 3 + +[14] Zhekai Du, Jingjing Li, Hongzu Su, Lei Zhu, and Ke Lu. Cross-domain gradient discrepancy minimization for unsupervised domain adaptation. 2021. 7 +[15] Kilian Fatras, Thibault Séjourne, Rémi Flamary, and Nicolas Courty. Unbalanced minibatch optimal transport; applications to domain adaptation. In International Conference on Machine Learning, pages 3186-3197. PMLR, 2021. 1 +[16] Wilfrid Gangbo and Robert J McCann. The geometry of optimal transportation. Acta Mathematica, 177(2):113-161, 1996. 3 +[17] Yixiao Ge, Dapeng Chen, and Hongsheng Li. Mutual meanteaching: Pseudo label refinery for unsupervised domain adaptation on person re-identification. In International Conference on Learning Representations, 2020. 2 +[18] Yves Grandvalet, Yoshua Bengio, et al. Semi-supervised learning by entropy minimization. In CAP, pages 281-296, 2005. 7 +[19] Sahand Hajifar and Hongyue Sun. Online domain adaptation for continuous cross-subject liver viability evaluation based on irregular thermal data, 2020. 5, 6 +[20] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. 2016. 1 +[21] Ying Jin, Ximei Wang, Mingsheng Long, and Jianmin Wang. Minimum class confusion for versatile domain adaptation, 2020. 5, 7 +[22] Leonid V Kantorovich. On the translocation of masses. Journal of mathematical sciences, 133(4):1381-1382, 2006. 2 +[23] Tanguy Kerdoncuff, Rémi Emonet, and Marc Sebban. Metric learning in optimal transport for domain adaptation. 2020. 1, 2, 5, 6, 7, 8 +[24] John Lee, Max Dabagia, Eva Dyer, and Christopher Rozell. Hierarchical optimal transport for multimodal distribution alignment. Advances in Neural Information Processing Systems, 32, 2019. 8 +[25] Jingjing Li, Erpeng Chen, Zhengming Ding, Lei Zhu, Ke Lu, and Heng Tao Shen. Maximum density divergence for domain adaptation. IEEE Transactions on Pattern Analysis and Machine Intelligence, page 1-1, 2020. 5, 6 +[26] Mengxue Li, Yi-Ming Zhai, You-Wei Luo, Peng-Fei Ge, and Chuan-Xian Ren. Enhanced transport distance for unsupervised domain adaptation. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 13933–13941, 2020. 6 +[27] Rui Li, Qianfen Jiao, Wenming Cao, Hau-San Wong, and Si Wu. Model adaptation: Unsupervised domain adaptation without source data. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9641-9650, 2020. 7 +[28] Shuang Li, Mixue Xie, Kaixiong Gong, Chi Harold Liu, Yulin Wang, and Wei Li. Transferable semantic augmentation for domain adaptation. arXiv preprint arXiv:2103.12562, 2021. 1, 5, 6, 7 +[29] Jian Liang, Dapeng Hu, and Jiashi Feng. Do we really need to access the source data? source hypothesis transfer for unsupervised domain adaptation. 2020. 5, 7 +[30] Chi-Heng Lin, Mehdi Azabou, and Eva L Dyer. Making transport more robust and interpretable by moving data + +through a small number of anchor points. Proceedings of machine learning research, 139:6631, 2021. 1 +[31] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dálár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In European conference on computer vision, pages 740–755. Springer, 2014. 6 +[32] Hong Liu, Mingsheng Long, Jianmin Wang, and Michael Jordan. Transferable adversarial training: A general approach to adapting deep classifiers. In International Conference on Machine Learning, pages 4013-4022. PMLR, 2019. 5, 6 +[33] Mingsheng Long, Zhangjie Cao, Jianmin Wang, and Michael I Jordan. Conditional adversarial domain adaptation. arXiv preprint arXiv:1705.10667, 2017. 5, 6 +[34] Mingsheng Long, Han Zhu, Jianmin Wang, and Michael I Jordan. Deep transfer learning with joint adaptation networks. In International conference on machine learning, pages 2208-2217. PMLR, 2017. 5, 6 +[35] Zhihe Lu, Yongxin Yang, Xiatian Zhu, Cong Liu, Yi-Zhe Song, and Tao Xiang. Stochastic classifiers for unsupervised domain adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9111-9120, 2020. 7 +[36] Liang Mi, Wen Zhang, and Yalin Wang. Regularized wasserstein means for aligning distributional data. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 34, pages 5166-5173, 2020. 8 +[37] Gaspard Monge. Mémoire sur la théorie des déblais et des remblais. Histoire de l'Académie Royale des Sciences de Paris, 1781. 2 +[38] Khai Nguyen, Dang Nguyen, Quoc Nguyen, Tung Pham, Hung Bui, Dinh Phung, Trung Le, and Nhat Ho. On transportation of mini-batches: A hierarchical approach. arXiv preprint arXiv:2102.05912, 2021. 1 +[39] Yingwei Pan, Ting Yao, Yehao Li, Yu Wang, Chong-Wah Ngo, and Tao Mei. Transferrable prototypical networks for unsupervised domain adaptation, 2019. 6 +[40] Ievgen Redko, Nicolas Courty, Rémi Flamary, and Devis Tuia. Optimal transport for multi-source domain adaptation under target shift. In The 22nd International Conference on Artificial Intelligence and Statistics, pages 849-858. PMLR, 2019. 1, 2 +[41] Hochang Rhee and Nam Ik Cho. Efficient and robust pseudolabeling for unsupervised domain adaptation. In 2019 Asia-Pacific Signal and Information Processing Association Annual Summit and Conference (APSIPA ASC), pages 980–985, 2019. 2 +[42] Kuniaki Saito, Kohei Watanabe, Yoshitaka Ushiku, and Tatsuya Harada. Maximum classifier discrepancy for unsupervised domain adaptation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3723-3732, 2018. 1, 5, 6 +[43] Swami Sankaranarayanan, Yogesh Balaji, Carlos D Castillo, and Rama Chellappa. Generate to adapt: Aligning domains using generative adversarial networks. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 8503-8512, 2018. 5 + +[44] Filippo Santambrogio. Optimal transport for applied mathematicians. Birkhäuser, NY, 55(58-63):94, 2015. 1 +[45] Hui Tang, Ke Chen, and Kui Jia. Unsupervised domain adaptation via structurally regularized deep clustering. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8725-8735, 2020. 5, 6 +[46] Eric Tzeng, Judy Hoffman, Kate Saenko, and Trevor Darrell. Adversarial discriminative domain adaptation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7167-7176, 2017. 5 +[47] Hao Wang, Yitong Wang, Zheng Zhou, Xing Ji, Dihong Gong, Jingchao Zhou, Zhifeng Li, and Wei Liu. Cosface: Large margin cosine loss for deep face recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5265-5274, 2018. 3 +[48] Qian Wang and Toby Breckon. Unsupervised domain adaptation via structured prediction based selective pseudolabeling. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 34, pages 6243-6250, 2020. 2 +[49] Wei Wang, Haojie Li, Zhengming Ding, and Zhihui Wang. Rethink maximum mean discrepancy for domain adaptation. arXiv preprint arXiv:2007.00689, 2020. 1 +[50] Pengfei Wei, Yiping Ke, Xinghua Qu, and Tze-Yun Leong. Subdomain adaptation with manifolds discrepancy alignment, 2020. 1 +[51] Renjun Xu, Pelen Liu, Liyan Wang, Chao Chen, and Jindong Wang. Reliable weighted optimal transport for unsupervised domain adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4394-4403, 2020. 1, 5, 7, 8 +[52] Tongkun Xu, Weihua Chen, Pichao Wang, Fan Wang, Hao Li, and Rong Jin. Cdtrans: Cross-domain transformer for unsupervised domain adaptation. arXiv preprint arXiv:2109.06165, 2021. 6, 7 +[53] Hongliang Yan, Yukang Ding, Peihua Li, Qilong Wang, Yong Xu, and Wangmeng Zuo. Mind the class weight bias: Weighted maximum mean discrepancy for unsupervised domain adaptation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 2272-2281, 2017. 1 +[54] Yuguang Yan, Wen Li, Hanrui Wu, Huaqing Min, Mingkui Tan, and Qingyao Wu. Semi-supervised optimal transport for heterogeneous domain adaptation. In IJCAI, volume 7, pages 2969-2975, 2018. 1 +[55] Jinyu Yang, Jingjing Liu, Ning Xu, and Junzhou Huang. Tvt: Transferable vision transformer for unsupervised domain adaptation. arXiv preprint arXiv:2108.05988, 2021. 7 +[56] Yabin Zhang, Hui Tang, Kui Jia, and Mingkui Tan. Domain-symmetric networks for adversarial domain adaptation, 2019. 6 +[57] Yongchun Zhu, Fuzhen Zhuang, Jindong Wang, Guolin Ke, Jingwu Chen, Jiang Bian, Hui Xiong, and Qing He. Deep subdomain adaptation network for image classification. IEEE Transactions on Neural Networks and Learning Systems, 32(4):1713-1722, Apr 2021. 1 \ No newline at end of file diff --git a/2023/COT_ Unsupervised Domain Adaptation With Clustering and Optimal Transport/images.zip b/2023/COT_ Unsupervised Domain Adaptation With Clustering and Optimal Transport/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..ff64b8a887e1b578cfaf81f74841a332ca35e047 --- /dev/null +++ b/2023/COT_ Unsupervised Domain Adaptation With Clustering and Optimal Transport/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5313daf10a4ff096093eaa6cc957f88a582281a483d0b2f73e6869ab189d578d +size 639966 diff --git a/2023/COT_ Unsupervised Domain Adaptation With Clustering and Optimal Transport/layout.json b/2023/COT_ Unsupervised Domain Adaptation With Clustering and Optimal Transport/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..868e6b3ad40573901321ace5d1a9284bc86c1713 --- /dev/null +++ b/2023/COT_ Unsupervised Domain Adaptation With Clustering and Optimal Transport/layout.json @@ -0,0 +1,10747 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 49, + 102, + 545, + 122 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 102, + 545, + 122 + ], + "spans": [ + { + "bbox": [ + 49, + 102, + 545, + 122 + ], + "type": "text", + "content": "COT: Unsupervised Domain Adaptation with Clustering and Optimal Transport" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 173, + 141, + 416, + 173 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 173, + 141, + 416, + 173 + ], + "spans": [ + { + "bbox": [ + 173, + 141, + 416, + 173 + ], + "type": "text", + "content": "Yang Liu " + }, + { + "bbox": [ + 173, + 141, + 416, + 173 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 173, + 141, + 416, + 173 + ], + "type": "text", + "content": " Zhipeng Zhou " + }, + { + "bbox": [ + 173, + 141, + 416, + 173 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 173, + 141, + 416, + 173 + ], + "type": "text", + "content": " Baigui Sun " + }, + { + "bbox": [ + 173, + 141, + 416, + 173 + ], + "type": "inline_equation", + "content": "^{1\\dagger}" + }, + { + "bbox": [ + 173, + 141, + 416, + 173 + ], + "type": "text", + "content": " 1Alibaba Group" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 143, + 198, + 192, + 213 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 198, + 192, + 213 + ], + "spans": [ + { + "bbox": [ + 143, + 198, + 192, + 213 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 224, + 290, + 488 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 224, + 290, + 488 + ], + "spans": [ + { + "bbox": [ + 46, + 224, + 290, + 488 + ], + "type": "text", + "content": "Unsupervised domain adaptation (UDA) aims to transfer the knowledge from a labeled source domain to an unlabeled target domain. Typically, to guarantee desirable knowledge transfer, aligning the distribution between source and target domain from a global perspective is widely adopted in UDA. Recent researchers further point out the importance of local-level alignment and propose to construct instance-pair alignment by leveraging on Optimal Transport (OT) theory. However, existing OT-based UDA approaches are limited to handling class imbalance challenges and introduce a heavy computation overhead when considering a large-scale training situation. To cope with two aforementioned issues, we propose a Clustering-based Optimal Transport (COT) algorithm, which formulates the alignment procedure as an Optimal Transport problem and constructs a mapping between clustering centers in the source and target domain via an end-to-end manner. With this alignment on clustering centers, our COT eliminates the negative effect caused by class imbalance and reduces the computation cost simultaneously. Empirically, our COT achieves state-of-the-art performance on several authoritative benchmark datasets." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 524, + 128, + 537 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 524, + 128, + 537 + ], + "spans": [ + { + "bbox": [ + 47, + 524, + 128, + 537 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 544, + 287, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 544, + 287, + 652 + ], + "spans": [ + { + "bbox": [ + 46, + 544, + 287, + 652 + ], + "type": "text", + "content": "Benefiting from the availability of large-scale data, deep learning has achieved tremendous success over the past few years. However, directly applying a well-trained convolution neural network on a new domain frequently suffers from the domain gap/discrepancy challenge, resulting in spurious predictions on the new domain. To remedy this, Unsupervised Domain Adaptation (UDA) has attracted many researchers' attention, which can transfer the knowledge from a labeled domain to an unlabeled domain." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 59, + 653, + 287, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 653, + 287, + 666 + ], + "spans": [ + { + "bbox": [ + 59, + 653, + 287, + 666 + ], + "type": "text", + "content": "A major line of UDA approaches [1, 1, 28, 42, 49, 53] aim" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 200, + 547, + 284 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 200, + 547, + 284 + ], + "spans": [ + { + "bbox": [ + 306, + 200, + 547, + 284 + ], + "type": "text", + "content": "to learn a global domain shift by aligning the global source and target distribution while ignoring the local-level alignment between two domains. By leveraging on global domain adaptation, the global distributions of source and target domain are almost the same, thus losing the fine-grained information for each class (class-structure) on the source and target domain." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 285, + 547, + 488 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 285, + 547, + 488 + ], + "spans": [ + { + "bbox": [ + 304, + 285, + 547, + 488 + ], + "type": "text", + "content": "Recently, to preserve class structure in both domains, several works [6, 15, 23, 30, 38, 40, 44, 51, 54] adopt optimal transport (OT) to minimize the sample-level transportation cost between source and target domain, achieving a significant performance on UDA. However, there exist two issues on recent OT-based UDA approaches. (i) When considering a realistic situation, i.e. the class imbalance" + }, + { + "bbox": [ + 304, + 285, + 547, + 488 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 304, + 285, + 547, + 488 + ], + "type": "text", + "content": " phenomenon occurs between the source and target domain, samples belonging to the same class in the target domain are assigned with false pseudo labels due to the mechanism of optimal transport, which requires each sample in source domain can be mapped to target samples. As a result, existing OT-based UDA methods provide poor pair-wise matching when facing class imbalance challenges. (ii) OT-based UDA methods tend to find a sample-level optimal counterpart, which requires a large amount of computation overhead, especially training on large-scale datasets." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 489, + 548, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 489, + 548, + 693 + ], + "spans": [ + { + "bbox": [ + 304, + 489, + 548, + 693 + ], + "type": "text", + "content": "To solve two aforementioned issues, we propose a Clustering-based Optimal Transport algorithm, termed COT, to construct a clustering-level instead of sample-level mapping between source and target domain. Clusters in the source domain are obtained from the classifiers supervised by the labeled source domain data. While for the target domain, COT utilizes a set of learnable clusters to represent the feature distribution of the target domain, which can describe the sub-domain information [50, 57]. For instance, in many object recognition tasks [13, 20] an object could contain many attributes. Each attribute can be viewed as a sub-domain. To this end, the clusters on the source and target domain can represent the individual sub-domain information, respectively, such that optimal transport between clusters intrinsically provides a local mapping from the subdomain in the source domain to those in the target domain. Moreover, we provide a theoretical analysis and compre" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 317, + 702, + 529, + 714 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 702, + 529, + 714 + ], + "spans": [ + { + "bbox": [ + 317, + 702, + 529, + 714 + ], + "type": "text", + "content": "1 label distribution are different in two domains, " + }, + { + "bbox": [ + 317, + 702, + 529, + 714 + ], + "type": "inline_equation", + "content": "P_{s}(y)\\neq P_{t}(y)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 47, + 674, + 178, + 684 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 674, + 178, + 684 + ], + "spans": [ + { + "bbox": [ + 47, + 674, + 178, + 684 + ], + "type": "text", + "content": "Email: ly261666@alibaba-inc.com" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 685, + 117, + 693 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 685, + 117, + 693 + ], + "spans": [ + { + "bbox": [ + 48, + 685, + 117, + 693 + ], + "type": "text", + "content": "* Equal Contribution" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 693, + 129, + 704 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 693, + 129, + 704 + ], + "spans": [ + { + "bbox": [ + 48, + 693, + 129, + 704 + ], + "type": "text", + "content": "† Corresponding Author" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "19998" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 119 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 119 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 119 + ], + "type": "text", + "content": "hensive experimental results to guarantee that (i) COT can alleviate the negative effect caused by class imbalance; (ii) Compared to existing OT-based UDA approaches, our COT economizes much computation head." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 59, + 121, + 240, + 132 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 121, + 240, + 132 + ], + "spans": [ + { + "bbox": [ + 59, + 121, + 240, + 132 + ], + "type": "text", + "content": "In summary, our main contributions include:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 140, + 286, + 274 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 58, + 140, + 286, + 212 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 140, + 286, + 212 + ], + "spans": [ + { + "bbox": [ + 58, + 140, + 286, + 212 + ], + "type": "text", + "content": "- We propose a novel Clustering-based Optimal Transport module as well as a specially designed loss derived from the discrete type of Kantorovich dual form, which resolves two aforementioned challenges on the existing OT-based UDA algorithms, facilitating the development of OT-based UDA community." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 58, + 220, + 286, + 243 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 220, + 286, + 243 + ], + "spans": [ + { + "bbox": [ + 58, + 220, + 286, + 243 + ], + "type": "text", + "content": "- We provide a theoretical analysis to guarantee the advantages of our COT." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 58, + 251, + 286, + 274 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 251, + 286, + 274 + ], + "spans": [ + { + "bbox": [ + 58, + 251, + 286, + 274 + ], + "type": "text", + "content": "- Our COT achieves state-of-the-art performance on several UDA benchmark datasets." + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 47, + 286, + 132, + 298 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 286, + 132, + 298 + ], + "spans": [ + { + "bbox": [ + 47, + 286, + 132, + 298 + ], + "type": "text", + "content": "2. Related work" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 306, + 287, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 306, + 287, + 521 + ], + "spans": [ + { + "bbox": [ + 46, + 306, + 287, + 521 + ], + "type": "text", + "content": "Pseudo Label based Domain Adaptation Inspired by the observation that samples in the target domain can be clustered within the feature space, for accurate pseudolabeling, [48] propose a selective pseudo-labeling strategy based on structural predictions which utilize the unsupervised clustering analysis. [41] introduce a confidence-based weighting scheme for obtaining pseudo-labels and an adaptive threshold adjustment strategy to provide sufficient and accurate pseudo-labels during the training process. The confidence-based weighting scheme generates pseudo-labels that can enable the performance less sensitive to the threshold determined by the pseudo-labels. In the task of person re-identification, [17] propose an unsupervised framework called Mutual Mean-Teaching to learn better features from the target domain by refining the hard pseudo labels offline and soft pseudo labels online alternatively to mitigate the effects of noisy pseudo labels caused by the clustering algorithms." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 522, + 286, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 522, + 286, + 605 + ], + "spans": [ + { + "bbox": [ + 46, + 522, + 286, + 605 + ], + "type": "text", + "content": "Optimal Transport based Domain Adaptation As a way to find a minimal effort strategy to the transport of a given mass of dirt into a given hole, [37] put forward the optimal transport problem for the first time. [22] provide an extension of the original problem of Monge. Recently, by applying the optimal transport to domain adaptation, some new computation strategies have emerged." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 605, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 286, + 713 + ], + "type": "text", + "content": "[7] propose regularized unsupervised optimal transport model to align the representation of features between different domains. The regularization schemes encoding class-structure in source domain during estimation of transport map enforce the intuition that samples of the same class must undergo a similar transformation. [5] minimize the optimal transport loss between the joint source distribution and the estimated target joint distribution depending on a function that is introduced to predict an output value given" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 72, + 545, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 204 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 204 + ], + "type": "text", + "content": "input from the source domain. For reducing the discrepancy between multiple domains, [40] propose Joint Class Proportion and Optimal Transport which performs multisource domain adaptation and target shift correction simultaneously by learning the predicted class probability of the unlabeled target data and the coupling to align the distributions between source and target domain. For better alignment between different domains, a relation between target error and the magnitude of different Wasserstein distances are proposed in [23] which optimize the metric for domain adaptation." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 213, + 383, + 227 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 213, + 383, + 227 + ], + "spans": [ + { + "bbox": [ + 306, + 213, + 383, + 227 + ], + "type": "text", + "content": "3. Preliminary" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 234, + 545, + 258 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 234, + 545, + 258 + ], + "spans": [ + { + "bbox": [ + 305, + 234, + 545, + 258 + ], + "type": "text", + "content": "In this section, we introduce the basic knowledge for optimal transport." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 265, + 417, + 277 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 265, + 417, + 277 + ], + "spans": [ + { + "bbox": [ + 306, + 265, + 417, + 277 + ], + "type": "text", + "content": "3.1. Optimal Transport" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 283, + 545, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 283, + 545, + 403 + ], + "spans": [ + { + "bbox": [ + 304, + 283, + 545, + 403 + ], + "type": "text", + "content": "Let " + }, + { + "bbox": [ + 304, + 283, + 545, + 403 + ], + "type": "inline_equation", + "content": "X \\subseteq \\mathbb{R}^d" + }, + { + "bbox": [ + 304, + 283, + 545, + 403 + ], + "type": "text", + "content": " be a measurable space and the labels are denoted as " + }, + { + "bbox": [ + 304, + 283, + 545, + 403 + ], + "type": "inline_equation", + "content": "\\mathcal{V}" + }, + { + "bbox": [ + 304, + 283, + 545, + 403 + ], + "type": "text", + "content": ". We denote the set of all probability distributions on " + }, + { + "bbox": [ + 304, + 283, + 545, + 403 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 304, + 283, + 545, + 403 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 304, + 283, + 545, + 403 + ], + "type": "inline_equation", + "content": "\\mathcal{P}(X)" + }, + { + "bbox": [ + 304, + 283, + 545, + 403 + ], + "type": "text", + "content": ". The source and target domains are space " + }, + { + "bbox": [ + 304, + 283, + 545, + 403 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 304, + 283, + 545, + 403 + ], + "type": "text", + "content": " equipped with two distinct probability distributions " + }, + { + "bbox": [ + 304, + 283, + 545, + 403 + ], + "type": "inline_equation", + "content": "\\mu_{\\mathcal{S}}" + }, + { + "bbox": [ + 304, + 283, + 545, + 403 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 283, + 545, + 403 + ], + "type": "inline_equation", + "content": "\\mu_{\\mathcal{T}}" + }, + { + "bbox": [ + 304, + 283, + 545, + 403 + ], + "type": "text", + "content": ". Suppose we have source dataset " + }, + { + "bbox": [ + 304, + 283, + 545, + 403 + ], + "type": "inline_equation", + "content": "\\{x_i^s\\}_{i=1}^{n_s} \\subset X_{\\mathcal{S}} = (X, \\mu_{\\mathcal{S}})" + }, + { + "bbox": [ + 304, + 283, + 545, + 403 + ], + "type": "text", + "content": " associated with label set " + }, + { + "bbox": [ + 304, + 283, + 545, + 403 + ], + "type": "inline_equation", + "content": "\\{y_i^s\\}_{i=1}^{n_s}" + }, + { + "bbox": [ + 304, + 283, + 545, + 403 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 304, + 283, + 545, + 403 + ], + "type": "inline_equation", + "content": "y_i^s \\in \\mathcal{V}" + }, + { + "bbox": [ + 304, + 283, + 545, + 403 + ], + "type": "text", + "content": ". The target dataset is " + }, + { + "bbox": [ + 304, + 283, + 545, + 403 + ], + "type": "inline_equation", + "content": "\\{x_j^t\\}_{j=1}^{n_t} \\subset X_{\\mathcal{T}} = (X, \\mu_{\\mathcal{T}})" + }, + { + "bbox": [ + 304, + 283, + 545, + 403 + ], + "type": "text", + "content": " without labels. The goal of optimal transport is to minimize the inter-domain transportation cost by finding a feasible map to preserve measure." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 409, + 545, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 409, + 545, + 445 + ], + "spans": [ + { + "bbox": [ + 305, + 409, + 545, + 445 + ], + "type": "text", + "content": "Definition 1 (Kantorovich) For given joint distribution " + }, + { + "bbox": [ + 305, + 409, + 545, + 445 + ], + "type": "inline_equation", + "content": "\\rho(x^s, x^t)" + }, + { + "bbox": [ + 305, + 409, + 545, + 445 + ], + "type": "text", + "content": " which satisfies for every measurable Borel set " + }, + { + "bbox": [ + 305, + 409, + 545, + 445 + ], + "type": "inline_equation", + "content": "O_S \\subset X_S" + }, + { + "bbox": [ + 305, + 409, + 545, + 445 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 305, + 409, + 545, + 445 + ], + "type": "inline_equation", + "content": "O_T \\subset X_T" + }, + { + "bbox": [ + 305, + 409, + 545, + 445 + ], + "type": "text", + "content": ", we have" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 315, + 453, + 545, + 467 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 453, + 545, + 467 + ], + "spans": [ + { + "bbox": [ + 315, + 453, + 545, + 467 + ], + "type": "interline_equation", + "content": "\\rho \\left(O _ {\\mathcal {S}} \\times X _ {\\mathcal {T}}\\right) = \\mu_ {\\mathcal {S}} \\left(O _ {\\mathcal {S}}\\right), \\rho \\left(X _ {\\mathcal {S}} \\times O _ {\\mathcal {T}}\\right) = \\mu_ {\\mathcal {T}} \\left(O _ {\\mathcal {T}}\\right) \\tag {1}", + "image_path": "d8a04d4e1d3bbf6f268fb3ae774730b05e59bde3d33be153fccedda72f76b12e.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 305, + 474, + 545, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 474, + 545, + 521 + ], + "spans": [ + { + "bbox": [ + 305, + 474, + 545, + 521 + ], + "type": "text", + "content": "For convenience, we denote the projection maps from " + }, + { + "bbox": [ + 305, + 474, + 545, + 521 + ], + "type": "inline_equation", + "content": "X_{\\mathcal{S}} \\times X_{\\mathcal{T}}" + }, + { + "bbox": [ + 305, + 474, + 545, + 521 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 305, + 474, + 545, + 521 + ], + "type": "inline_equation", + "content": "X_{\\mathcal{S}}" + }, + { + "bbox": [ + 305, + 474, + 545, + 521 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 474, + 545, + 521 + ], + "type": "inline_equation", + "content": "X_{\\mathcal{T}}" + }, + { + "bbox": [ + 305, + 474, + 545, + 521 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 305, + 474, + 545, + 521 + ], + "type": "inline_equation", + "content": "\\pi_{\\mathcal{S}}, \\pi_{\\mathcal{T}}" + }, + { + "bbox": [ + 305, + 474, + 545, + 521 + ], + "type": "text", + "content": ". The above equation can be denoted as " + }, + { + "bbox": [ + 305, + 474, + 545, + 521 + ], + "type": "inline_equation", + "content": "\\pi_{\\mathcal{S}\\#} \\rho = \\mu_{\\mathcal{S}}" + }, + { + "bbox": [ + 305, + 474, + 545, + 521 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 474, + 545, + 521 + ], + "type": "inline_equation", + "content": "\\pi_{\\mathcal{T}\\#} \\rho = \\mu_{\\mathcal{T}}" + }, + { + "bbox": [ + 305, + 474, + 545, + 521 + ], + "type": "text", + "content": ". The corresponding transportation cost is" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 349, + 529, + 545, + 555 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 349, + 529, + 545, + 555 + ], + "spans": [ + { + "bbox": [ + 349, + 529, + 545, + 555 + ], + "type": "interline_equation", + "content": "\\mathcal {C} (\\rho) = \\int_ {X _ {\\mathcal {S}} \\times X _ {\\mathcal {T}}} c \\left(x ^ {s}, x ^ {t}\\right) d \\rho \\left(x ^ {s}, x ^ {t}\\right) \\tag {2}", + "image_path": "0b484571b67181c514a26d32576f184753ca8b0d7551e0bdcbb160313a7dbbbc.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 305, + 562, + 545, + 610 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 562, + 545, + 610 + ], + "spans": [ + { + "bbox": [ + 305, + 562, + 545, + 610 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 562, + 545, + 610 + ], + "type": "inline_equation", + "content": "c(x^s, x^t)" + }, + { + "bbox": [ + 305, + 562, + 545, + 610 + ], + "type": "text", + "content": " is pointwise transportation cost between " + }, + { + "bbox": [ + 305, + 562, + 545, + 610 + ], + "type": "inline_equation", + "content": "x^s \\in X_S" + }, + { + "bbox": [ + 305, + 562, + 545, + 610 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 562, + 545, + 610 + ], + "type": "inline_equation", + "content": "x^t \\in X_T" + }, + { + "bbox": [ + 305, + 562, + 545, + 610 + ], + "type": "text", + "content": ". The optimal transport problem is proposed to minimize the " + }, + { + "bbox": [ + 305, + 562, + 545, + 610 + ], + "type": "inline_equation", + "content": "\\mathcal{C}(\\rho)" + }, + { + "bbox": [ + 305, + 562, + 545, + 610 + ], + "type": "text", + "content": " under the measure preserving as the following:" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 334, + 617, + 545, + 635 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 334, + 617, + 545, + 635 + ], + "spans": [ + { + "bbox": [ + 334, + 617, + 545, + 635 + ], + "type": "interline_equation", + "content": "W _ {c} := \\inf _ {\\rho} \\left\\{\\mathcal {C} (\\rho) | | \\pi_ {\\mathcal {S} \\#} \\rho = \\mu_ {\\mathcal {S}}, \\pi_ {\\mathcal {T} \\#} \\rho = \\mu_ {\\mathcal {T}} \\right\\} \\tag {3}", + "image_path": "1bd7a021abfd7d20b1d3161d872e7b330fdb991ab7685e8ca6e2e7e6c2a84f80.jpg" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 305, + 643, + 545, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 643, + 545, + 667 + ], + "spans": [ + { + "bbox": [ + 305, + 643, + 545, + 667 + ], + "type": "text", + "content": "By convex optimization theory, we can consider the Kantorovich's dual problem as:" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 319, + 674, + 545, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 674, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 319, + 674, + 545, + 715 + ], + "type": "interline_equation", + "content": "W _ {c} := \\left\\{ \\begin{array}{l} \\max _ {\\varphi , \\psi} \\int_ {X _ {\\mathcal {S}}} \\varphi \\left(x ^ {s}\\right) d \\mu_ {\\mathcal {S}} + \\int_ {X _ {\\mathcal {T}}} \\psi \\left(x ^ {t}\\right) d \\mu_ {\\mathcal {T}} \\\\ \\text {s . t .} \\quad \\varphi \\left(x ^ {s}\\right) + \\psi \\left(x ^ {t}\\right) \\leq c \\left(x ^ {s}, x ^ {t}\\right) \\end{array} \\right. \\tag {4}", + "image_path": "1836942fff43b20d0e072011dd7ca5dcab6298d05167c75240a4b02fcffebe4d.jpg" + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19999" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 287, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 287, + 96 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 287, + 96 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 72, + 287, + 96 + ], + "type": "inline_equation", + "content": "\\varphi" + }, + { + "bbox": [ + 47, + 72, + 287, + 96 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 72, + 287, + 96 + ], + "type": "inline_equation", + "content": "\\psi" + }, + { + "bbox": [ + 47, + 72, + 287, + 96 + ], + "type": "text", + "content": " are real functions from " + }, + { + "bbox": [ + 47, + 72, + 287, + 96 + ], + "type": "inline_equation", + "content": "X_{\\mathcal{S}}" + }, + { + "bbox": [ + 47, + 72, + 287, + 96 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 72, + 287, + 96 + ], + "type": "inline_equation", + "content": "X_{\\mathcal{T}}" + }, + { + "bbox": [ + 47, + 72, + 287, + 96 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 47, + 72, + 287, + 96 + ], + "type": "inline_equation", + "content": "\\mathbb{R}" + }, + { + "bbox": [ + 47, + 72, + 287, + 96 + ], + "type": "text", + "content": ". Moreover, the Kantorovich problem can be formulated as" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 81, + 105, + 287, + 132 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 81, + 105, + 287, + 132 + ], + "spans": [ + { + "bbox": [ + 81, + 105, + 287, + 132 + ], + "type": "interline_equation", + "content": "\\max _ {\\varphi} \\left\\{\\int_ {X _ {\\mathcal {S}}} \\varphi \\left(x ^ {s}\\right) d \\mu_ {\\mathcal {S}} + \\int_ {X _ {\\mathcal {T}}} \\varphi^ {c} \\left(x ^ {t}\\right) d \\mu_ {\\mathcal {T}} \\right\\} \\tag {5}", + "image_path": "95b557a2588ed0bd5a8e17858bea81fd88f7091e63b116fb508a58f85dac85a8.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 140, + 287, + 169 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 140, + 287, + 169 + ], + "spans": [ + { + "bbox": [ + 47, + 140, + 287, + 169 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 140, + 287, + 169 + ], + "type": "inline_equation", + "content": "\\varphi^c (x^t) = \\inf_{x_s\\in X_S}\\{c(x^s,x^t) - \\varphi (x^s)\\}" + }, + { + "bbox": [ + 47, + 140, + 287, + 169 + ], + "type": "text", + "content": " is called the " + }, + { + "bbox": [ + 47, + 140, + 287, + 169 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 47, + 140, + 287, + 169 + ], + "type": "text", + "content": " -transform of " + }, + { + "bbox": [ + 47, + 140, + 287, + 169 + ], + "type": "inline_equation", + "content": "\\varphi" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 169, + 287, + 226 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 169, + 287, + 226 + ], + "spans": [ + { + "bbox": [ + 47, + 169, + 287, + 226 + ], + "type": "text", + "content": "By classical optimal transport theory, different choices of cost function will influence the difficulty to solving the optimal transport problem. When we choose " + }, + { + "bbox": [ + 47, + 169, + 287, + 226 + ], + "type": "inline_equation", + "content": "c(x^s,x^t) = \\| x^s -x^t\\| _2" + }, + { + "bbox": [ + 47, + 169, + 287, + 226 + ], + "type": "text", + "content": ", the problem stated in Equation (5) is equivalent to" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 83, + 226, + 287, + 252 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 83, + 226, + 287, + 252 + ], + "spans": [ + { + "bbox": [ + 83, + 226, + 287, + 252 + ], + "type": "interline_equation", + "content": "\\max _ {\\varphi} \\left\\{\\int_ {X _ {\\mathcal {S}}} \\varphi \\left(x ^ {s}\\right) d \\mu_ {\\mathcal {S}} - \\int_ {X _ {\\mathcal {T}}} \\varphi \\left(x ^ {t}\\right) d \\mu_ {\\mathcal {T}} \\right\\} \\tag {6}", + "image_path": "a48104bbbe4ab564aee61b50f9ac63f4a6bd1a420f63715503dbd03ebd1856d9.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 256, + 287, + 340 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 256, + 287, + 340 + ], + "spans": [ + { + "bbox": [ + 47, + 256, + 287, + 340 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 256, + 287, + 340 + ], + "type": "inline_equation", + "content": "\\varphi" + }, + { + "bbox": [ + 47, + 256, + 287, + 340 + ], + "type": "text", + "content": " is under the constraint that " + }, + { + "bbox": [ + 47, + 256, + 287, + 340 + ], + "type": "inline_equation", + "content": "|\\varphi(x) - \\varphi(x')| \\leq \\|x - x'\\|_2" + }, + { + "bbox": [ + 47, + 256, + 287, + 340 + ], + "type": "text", + "content": ". WGAN [2] is inspired by above cost setting, during the implementation of optimal transport in WGAN, they utilize the gradient clip to guarantee the Lipschitz constant of " + }, + { + "bbox": [ + 47, + 256, + 287, + 340 + ], + "type": "inline_equation", + "content": "\\varphi" + }, + { + "bbox": [ + 47, + 256, + 287, + 340 + ], + "type": "text", + "content": " is bounded from above by 1. When we set the cost function as " + }, + { + "bbox": [ + 47, + 256, + 287, + 340 + ], + "type": "inline_equation", + "content": "c(x^s, x^t) = \\|x^s - x^t\\|_2^2" + }, + { + "bbox": [ + 47, + 256, + 287, + 340 + ], + "type": "text", + "content": ", [16] guarantees the existence and uniqueness of optimal transport map." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 351, + 129, + 365 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 351, + 129, + 365 + ], + "spans": [ + { + "bbox": [ + 47, + 351, + 129, + 365 + ], + "type": "text", + "content": "4. Methodology" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 372, + 287, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 372, + 287, + 396 + ], + "spans": [ + { + "bbox": [ + 47, + 372, + 287, + 396 + ], + "type": "text", + "content": "In this section, we present the detail of our COT and demonstrate the advantages of our COT theoretically." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 403, + 242, + 416 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 403, + 242, + 416 + ], + "spans": [ + { + "bbox": [ + 47, + 403, + 242, + 416 + ], + "type": "text", + "content": "4.1. Clustering-based Optimal Transport" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 422, + 287, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 422, + 287, + 540 + ], + "spans": [ + { + "bbox": [ + 46, + 422, + 287, + 540 + ], + "type": "text", + "content": "Instead of aligning instance-level features between source and target domain, we propose a novel clustering-based optimal transport (COT) module for unsupervised domain adaptation in this subsection. Firstly, we extract features from the source and target domain by ImageNet pretrained CNNs. Then we utilize learnable clusters to represent the sub-domains in the source and target domain, respectively. Finally, we apply a Kantorovich dual form-based loss to implement the optimal transport between clusters from both domains." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 541, + 287, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 541, + 287, + 613 + ], + "spans": [ + { + "bbox": [ + 47, + 541, + 287, + 613 + ], + "type": "text", + "content": "Feature Extractor We utilize an ImageNet pretrained (without fully connected layers) CNNs (e.g. ResNet50/ResNet101) to extract features " + }, + { + "bbox": [ + 47, + 541, + 287, + 613 + ], + "type": "inline_equation", + "content": "\\{x_{i}^{s}\\}_{i = 1}^{n_{s}}" + }, + { + "bbox": [ + 47, + 541, + 287, + 613 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 541, + 287, + 613 + ], + "type": "inline_equation", + "content": "\\{x_j^t\\}_{j = 1}^{n_t}" + }, + { + "bbox": [ + 47, + 541, + 287, + 613 + ], + "type": "text", + "content": " from the source and target dataset respectively at the beginning of training process. Note that the distributions of features vary during the training phase." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 614, + 287, + 716 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 614, + 287, + 716 + ], + "spans": [ + { + "bbox": [ + 47, + 614, + 287, + 716 + ], + "type": "text", + "content": "Clustering As for each sample from source domain, i.e., " + }, + { + "bbox": [ + 47, + 614, + 287, + 716 + ], + "type": "inline_equation", + "content": "x_{i}^{s}" + }, + { + "bbox": [ + 47, + 614, + 287, + 716 + ], + "type": "text", + "content": " and corresponding ground-truth " + }, + { + "bbox": [ + 47, + 614, + 287, + 716 + ], + "type": "inline_equation", + "content": "y_{i}^{s} \\in \\mathcal{Y}" + }, + { + "bbox": [ + 47, + 614, + 287, + 716 + ], + "type": "text", + "content": ". We denote the fully-connected layer which outputs the classification logits as " + }, + { + "bbox": [ + 47, + 614, + 287, + 716 + ], + "type": "inline_equation", + "content": "W = [w_{1}^{s},\\dots,w_{|\\mathcal{Y}|}^{s}]^{\\top} \\in \\mathbb{R}^{|\\mathcal{Y}|\\times c}" + }, + { + "bbox": [ + 47, + 614, + 287, + 716 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 47, + 614, + 287, + 716 + ], + "type": "inline_equation", + "content": "|\\mathcal{Y}|" + }, + { + "bbox": [ + 47, + 614, + 287, + 716 + ], + "type": "text", + "content": " is the number of categories and " + }, + { + "bbox": [ + 47, + 614, + 287, + 716 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 47, + 614, + 287, + 716 + ], + "type": "text", + "content": " is the number of feature channels. The predicted classification probability is " + }, + { + "bbox": [ + 47, + 614, + 287, + 716 + ], + "type": "inline_equation", + "content": "P(\\hat{y}_i^s = v|x_i^s) = \\frac{e^{x_i^s\\top w_v^s}}{\\sum_{u = 1}^{|\\mathcal{Y}|}e^{x_i^s\\top w_u^s}}" + }, + { + "bbox": [ + 47, + 614, + 287, + 716 + ], + "type": "text", + "content": ". The corresponding cross-entropy loss is shown" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 73, + 353, + 83 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 73, + 353, + 83 + ], + "spans": [ + { + "bbox": [ + 306, + 73, + 353, + 83 + ], + "type": "text", + "content": "as follows:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 333, + 89, + 545, + 121 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 89, + 545, + 121 + ], + "spans": [ + { + "bbox": [ + 333, + 89, + 545, + 121 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {c r o s s - e n t r o p y}} = \\frac {1}{b} \\sum_ {i = 1} ^ {b} - y _ {i} ^ {s} \\cdot \\log \\left(P \\left(\\hat {y} _ {i} ^ {s} \\mid x _ {i} ^ {s}\\right)\\right) \\tag {7}", + "image_path": "3dfaaef601995e66c76a8f985eec58d3812aeec7e2ad1bb825bb523f5d652fb1.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 130, + 545, + 448 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 130, + 545, + 448 + ], + "spans": [ + { + "bbox": [ + 304, + 130, + 545, + 448 + ], + "type": "text", + "content": "For the source domain, since " + }, + { + "bbox": [ + 304, + 130, + 545, + 448 + ], + "type": "inline_equation", + "content": "\\{w_v^s\\}_{v=1}^{|\\mathcal{Y}|}" + }, + { + "bbox": [ + 304, + 130, + 545, + 448 + ], + "type": "text", + "content": " [13, 47] have been shown effective for clustering representation, we take the classifiers " + }, + { + "bbox": [ + 304, + 130, + 545, + 448 + ], + "type": "inline_equation", + "content": "\\{w_v^s\\}_{v=1}^{|\\mathcal{Y}|}" + }, + { + "bbox": [ + 304, + 130, + 545, + 448 + ], + "type": "text", + "content": " as clusters for feature space of source domain. For the target domain, a set of learnable clusters termed as " + }, + { + "bbox": [ + 304, + 130, + 545, + 448 + ], + "type": "inline_equation", + "content": "\\{w_u^t\\}_{u=1}^K" + }, + { + "bbox": [ + 304, + 130, + 545, + 448 + ], + "type": "text", + "content": " are proposed to represent the sub-domains, where " + }, + { + "bbox": [ + 304, + 130, + 545, + 448 + ], + "type": "inline_equation", + "content": "K = Q \\cdot |\\mathcal{Y}|" + }, + { + "bbox": [ + 304, + 130, + 545, + 448 + ], + "type": "text", + "content": " is a hyper-parameter which stands for the number of sub-domains in the target domain, " + }, + { + "bbox": [ + 304, + 130, + 545, + 448 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 304, + 130, + 545, + 448 + ], + "type": "text", + "content": " is a positive integer which represents the number of sub-domains for each class. The motivation of introducing " + }, + { + "bbox": [ + 304, + 130, + 545, + 448 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 304, + 130, + 545, + 448 + ], + "type": "text", + "content": " is to preserve the sub-domain information in the target domain during the COT optimization procedure. For instance, if we set the same number of sub-domains on the source and target domain, the sub-domain information in the target domain would be seriously influenced by those in the source domain during the COT procedure, since the latter is optimized with hard supervision signal while the former is supervised by soft signal. On the contrary, " + }, + { + "bbox": [ + 304, + 130, + 545, + 448 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 304, + 130, + 545, + 448 + ], + "type": "text", + "content": " helps the target domain generate more sub-domains, which can preserve the sub-domain information in the target domain because not all of them are optimized during the COT phase. Empirically, the performance is superior when " + }, + { + "bbox": [ + 304, + 130, + 545, + 448 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 304, + 130, + 545, + 448 + ], + "type": "text", + "content": " is set to 2. " + }, + { + "bbox": [ + 304, + 130, + 545, + 448 + ], + "type": "inline_equation", + "content": "\\{w_u^t \\in \\mathbb{R}^c\\}_{u=(l-1)\\cdot Q+1}^{l\\cdot Q}" + }, + { + "bbox": [ + 304, + 130, + 545, + 448 + ], + "type": "text", + "content": " represent the clusters for class " + }, + { + "bbox": [ + 304, + 130, + 545, + 448 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 304, + 130, + 545, + 448 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 130, + 545, + 448 + ], + "type": "inline_equation", + "content": "\\forall 1 \\leq l \\leq |\\mathcal{Y}|" + }, + { + "bbox": [ + 304, + 130, + 545, + 448 + ], + "type": "text", + "content": ". For each feature " + }, + { + "bbox": [ + 304, + 130, + 545, + 448 + ], + "type": "inline_equation", + "content": "x_j^t" + }, + { + "bbox": [ + 304, + 130, + 545, + 448 + ], + "type": "text", + "content": ", we assign it to the closest cluster in " + }, + { + "bbox": [ + 304, + 130, + 545, + 448 + ], + "type": "inline_equation", + "content": "\\{w_u^t\\}_{u=1}^K" + }, + { + "bbox": [ + 304, + 130, + 545, + 448 + ], + "type": "text", + "content": ". We utilize the " + }, + { + "bbox": [ + 304, + 130, + 545, + 448 + ], + "type": "inline_equation", + "content": "L_2" + }, + { + "bbox": [ + 304, + 130, + 545, + 448 + ], + "type": "text", + "content": " distance to measure the distance between features and clusters and pull the features back to corresponding clusters." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 321, + 464, + 545, + 521 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 464, + 545, + 521 + ], + "spans": [ + { + "bbox": [ + 321, + 464, + 545, + 521 + ], + "type": "interline_equation", + "content": "\\left\\{ \\begin{array}{c} \\mathcal {L} _ {\\text {c l u s t e r}} = \\sum_ {j = 1} ^ {n _ {t}} \\| w _ {u *} ^ {t} - x _ {j} ^ {t} \\| ^ {2} \\\\ s. t. u _ {j} ^ {*} = \\underset {1 \\leq u \\leq K} {\\arg \\min } \\left(\\left\\| w _ {u} ^ {t} - x _ {j} ^ {t} \\right\\| ^ {2}\\right), \\forall 1 \\leq j \\leq n _ {t} \\end{array} \\right. \\tag {8}", + "image_path": "5adefec1909912aab5babea3e61620bfecedac1c364c7f2ab0fd80e8c4ac99af.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 528, + 545, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 528, + 545, + 567 + ], + "spans": [ + { + "bbox": [ + 304, + 528, + 545, + 567 + ], + "type": "text", + "content": "Optimal Transport With clusters " + }, + { + "bbox": [ + 304, + 528, + 545, + 567 + ], + "type": "inline_equation", + "content": "\\{w_v^s\\}_{v=1}^{|\\mathcal{Y}|}" + }, + { + "bbox": [ + 304, + 528, + 545, + 567 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 528, + 545, + 567 + ], + "type": "inline_equation", + "content": "\\{w_u^t\\}_{u=1}^K" + }, + { + "bbox": [ + 304, + 528, + 545, + 567 + ], + "type": "text", + "content": " from source and target domain respectively, we design the clustering based optimal transport as follows:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 341, + 573, + 545, + 643 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 341, + 573, + 545, + 643 + ], + "spans": [ + { + "bbox": [ + 341, + 573, + 545, + 643 + ], + "type": "interline_equation", + "content": "\\left\\{ \\begin{array}{c} \\min _ {T \\in \\mathbb {R} ^ {| \\mathcal {Y} | \\times K}} T _ {v u} c _ {v u} \\\\ s. t. \\sum_ {u = 1} ^ {K} T _ {v u} = \\frac {1}{K}, \\sum_ {v = 1} ^ {| \\mathcal {Y} |} T _ {v u} = \\frac {1}{| \\mathcal {Y} |}, \\\\ T _ {v u} \\geq 0, \\forall 1 \\leq v \\leq | \\mathcal {Y} |, 1 \\leq u \\leq K \\end{array} \\right. \\tag {9}", + "image_path": "5c819a55eb302e9fcfd5168883aa5f4d6243b2f4819afab0607fc36f56705275.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 651, + 545, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 651, + 545, + 676 + ], + "spans": [ + { + "bbox": [ + 304, + 651, + 545, + 676 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 651, + 545, + 676 + ], + "type": "inline_equation", + "content": "c_{vu} = \\| w_v^s - w_u^t\\|_2^2" + }, + { + "bbox": [ + 304, + 651, + 545, + 676 + ], + "type": "text", + "content": ". Similar to Equation (5), we can get the discrete Kantorovich dual problem of Equation (9)." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 343, + 682, + 545, + 716 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 343, + 682, + 545, + 716 + ], + "spans": [ + { + "bbox": [ + 343, + 682, + 545, + 716 + ], + "type": "interline_equation", + "content": "\\max _ {\\psi} \\left\\{\\frac {1}{K} \\sum_ {u = 1} ^ {K} \\psi \\left(w _ {u} ^ {t}\\right) + \\frac {1}{| \\mathcal {Y} |} \\sum_ {v = 1} ^ {| \\mathcal {Y} |} \\psi^ {c} \\left(w _ {v} ^ {s}\\right) \\right\\} \\tag {10}", + "image_path": "62bdd98094e18dcea6c01964b336c3ecf9824f5a4b857efe27b2a607cad2854e.jpg" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "footer", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "20000" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 249, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 249, + 84 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 249, + 84 + ], + "type": "text", + "content": "Algorithm 1 Clustering-based Optimal Transport" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 50, + 87, + 287, + 320 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 53, + 87, + 287, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 87, + 287, + 127 + ], + "spans": [ + { + "bbox": [ + 53, + 87, + 287, + 127 + ], + "type": "text", + "content": "1: Set number of epochs for training as " + }, + { + "bbox": [ + 53, + 87, + 287, + 127 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 53, + 87, + 287, + 127 + ], + "type": "text", + "content": ", learnable clusters for target domain as " + }, + { + "bbox": [ + 53, + 87, + 287, + 127 + ], + "type": "inline_equation", + "content": "\\{w_u^t\\}_{u=1}^{|\\mathcal{Y}|}" + }, + { + "bbox": [ + 53, + 87, + 287, + 127 + ], + "type": "text", + "content": ", classifiers/clusters for source domain as " + }, + { + "bbox": [ + 53, + 87, + 287, + 127 + ], + "type": "inline_equation", + "content": "\\{w_v^s\\}_{v=1}^{|\\mathcal{Y}|}" + }, + { + "bbox": [ + 53, + 87, + 287, + 127 + ], + "type": "text", + "content": ";" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 127, + 226, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 127, + 226, + 139 + ], + "spans": [ + { + "bbox": [ + 53, + 127, + 226, + 139 + ], + "type": "text", + "content": "2: for k-th training epoch while " + }, + { + "bbox": [ + 53, + 127, + 226, + 139 + ], + "type": "inline_equation", + "content": "k \\leq E" + }, + { + "bbox": [ + 53, + 127, + 226, + 139 + ], + "type": "text", + "content": " do" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 140, + 212, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 140, + 212, + 150 + ], + "spans": [ + { + "bbox": [ + 53, + 140, + 212, + 150 + ], + "type": "text", + "content": "3: for " + }, + { + "bbox": [ + 53, + 140, + 212, + 150 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 53, + 140, + 212, + 150 + ], + "type": "text", + "content": "-th iteration in " + }, + { + "bbox": [ + 53, + 140, + 212, + 150 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 53, + 140, + 212, + 150 + ], + "type": "text", + "content": "-th epoch do" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 151, + 287, + 200 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 151, + 287, + 200 + ], + "spans": [ + { + "bbox": [ + 53, + 151, + 287, + 200 + ], + "type": "text", + "content": "4: Take mini-batch of samples from source and target domain as input for feature extractor CNNs with parameters " + }, + { + "bbox": [ + 53, + 151, + 287, + 200 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 53, + 151, + 287, + 200 + ], + "type": "text", + "content": ", the output features are " + }, + { + "bbox": [ + 53, + 151, + 287, + 200 + ], + "type": "inline_equation", + "content": "\\{x_i^s\\}_{i=1}^b" + }, + { + "bbox": [ + 53, + 151, + 287, + 200 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 53, + 151, + 287, + 200 + ], + "type": "inline_equation", + "content": "\\{x_j^t\\}_{j=1}^b" + }, + { + "bbox": [ + 53, + 151, + 287, + 200 + ], + "type": "text", + "content": ";" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 200, + 287, + 225 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 200, + 287, + 225 + ], + "spans": [ + { + "bbox": [ + 53, + 200, + 287, + 225 + ], + "type": "text", + "content": "5: Compute the " + }, + { + "bbox": [ + 53, + 200, + 287, + 225 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{cluster}}" + }, + { + "bbox": [ + 53, + 200, + 287, + 225 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 53, + 200, + 287, + 225 + ], + "type": "inline_equation", + "content": "\\{x_j^t\\}_{j=1}^b" + }, + { + "bbox": [ + 53, + 200, + 287, + 225 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 53, + 200, + 287, + 225 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{cross-entropy}}" + }, + { + "bbox": [ + 53, + 200, + 287, + 225 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 53, + 200, + 287, + 225 + ], + "type": "inline_equation", + "content": "\\{x_i^s\\}_{i=1}^b" + }, + { + "bbox": [ + 53, + 200, + 287, + 225 + ], + "type": "text", + "content": " in the " + }, + { + "bbox": [ + 53, + 200, + 287, + 225 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 53, + 200, + 287, + 225 + ], + "type": "text", + "content": "-th batch, and " + }, + { + "bbox": [ + 53, + 200, + 287, + 225 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{OT}" + }, + { + "bbox": [ + 53, + 200, + 287, + 225 + ], + "type": "text", + "content": ";" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 225, + 195, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 225, + 195, + 237 + ], + "spans": [ + { + "bbox": [ + 53, + 225, + 195, + 237 + ], + "type": "text", + "content": "6: if " + }, + { + "bbox": [ + 53, + 225, + 195, + 237 + ], + "type": "inline_equation", + "content": "1 \\leq t \\leq k / (b * 10)" + }, + { + "bbox": [ + 53, + 225, + 195, + 237 + ], + "type": "text", + "content": " then" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 238, + 287, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 238, + 287, + 274 + ], + "spans": [ + { + "bbox": [ + 53, + 238, + 287, + 274 + ], + "type": "text", + "content": "7: we find the current optimal map from clusters of source domain to those of target domain by maximizing " + }, + { + "bbox": [ + 53, + 238, + 287, + 274 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{OT}}" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 274, + 112, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 274, + 112, + 285 + ], + "spans": [ + { + "bbox": [ + 53, + 274, + 112, + 285 + ], + "type": "text", + "content": "8: end if" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 285, + 159, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 285, + 159, + 297 + ], + "spans": [ + { + "bbox": [ + 53, + 285, + 159, + 297 + ], + "type": "text", + "content": "9: Minimize " + }, + { + "bbox": [ + 53, + 285, + 159, + 297 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{COT}" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 50, + 298, + 108, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 50, + 298, + 108, + 308 + ], + "spans": [ + { + "bbox": [ + 50, + 298, + 108, + 308 + ], + "type": "text", + "content": "10: end for" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 50, + 310, + 97, + 320 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 50, + 310, + 97, + 320 + ], + "spans": [ + { + "bbox": [ + 50, + 310, + 97, + 320 + ], + "type": "text", + "content": "11: end for" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 47, + 341, + 287, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 341, + 287, + 399 + ], + "spans": [ + { + "bbox": [ + 47, + 341, + 287, + 399 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 341, + 287, + 399 + ], + "type": "inline_equation", + "content": "\\psi^c (w_v^s) = \\inf_{u = 1}^{K}(c_{vu} - \\psi (w_u^t))" + }, + { + "bbox": [ + 47, + 341, + 287, + 399 + ], + "type": "text", + "content": ". According to Equation 10, We can seek for the optimal transportation map between clusters by optimizing the following loss (line 7 in Algorithm 1):" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 66, + 407, + 287, + 441 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 407, + 287, + 441 + ], + "spans": [ + { + "bbox": [ + 66, + 407, + 287, + 441 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {O T}} = \\frac {1}{K} \\sum_ {u = 1} ^ {K} \\lambda_ {u} ^ {t} + \\frac {1}{| \\mathcal {Y} |} \\sum_ {v = 1} ^ {| \\mathcal {Y} |} \\left(\\inf _ {u = 1} ^ {K} \\left(c _ {v u} - \\lambda_ {u} ^ {t}\\right)\\right) \\tag {11}", + "image_path": "a26d8355f9cba15b1efe8e7f63062ad4a906b72e1ca88a2ea8f9516121a3cad7.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 47, + 449, + 287, + 573 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 449, + 287, + 573 + ], + "spans": [ + { + "bbox": [ + 47, + 449, + 287, + 573 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 449, + 287, + 573 + ], + "type": "inline_equation", + "content": "\\{\\lambda_u^t\\}_{u = 1}^K" + }, + { + "bbox": [ + 47, + 449, + 287, + 573 + ], + "type": "text", + "content": " represent the value of function " + }, + { + "bbox": [ + 47, + 449, + 287, + 573 + ], + "type": "inline_equation", + "content": "\\psi" + }, + { + "bbox": [ + 47, + 449, + 287, + 573 + ], + "type": "text", + "content": " at points " + }, + { + "bbox": [ + 47, + 449, + 287, + 573 + ], + "type": "inline_equation", + "content": "\\{w_{u}^{t}\\}_{u = 1}^{K}" + }, + { + "bbox": [ + 47, + 449, + 287, + 573 + ], + "type": "text", + "content": ". Furthermore, it is worth noting that cost " + }, + { + "bbox": [ + 47, + 449, + 287, + 573 + ], + "type": "inline_equation", + "content": "c_{vu}" + }, + { + "bbox": [ + 47, + 449, + 287, + 573 + ], + "type": "text", + "content": " is frozen during the optimization of " + }, + { + "bbox": [ + 47, + 449, + 287, + 573 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{OT}}" + }, + { + "bbox": [ + 47, + 449, + 287, + 573 + ], + "type": "text", + "content": ", indicating that only " + }, + { + "bbox": [ + 47, + 449, + 287, + 573 + ], + "type": "inline_equation", + "content": "\\{\\lambda_v^t\\}_{v = 1}^{|\\mathcal{V}|}" + }, + { + "bbox": [ + 47, + 449, + 287, + 573 + ], + "type": "text", + "content": " is updated in this step. Then based on the founded optimal map, we minimize the following loss " + }, + { + "bbox": [ + 47, + 449, + 287, + 573 + ], + "type": "inline_equation", + "content": "(\\mathcal{L}_{\\mathrm{COT}})" + }, + { + "bbox": [ + 47, + 449, + 287, + 573 + ], + "type": "text", + "content": " to close the matched clusters from the feature representation perspective, thus the domain-invariant feature representation is learned via this step (line 9 in Algorithm 1). Correspondingly, we freeze the " + }, + { + "bbox": [ + 47, + 449, + 287, + 573 + ], + "type": "inline_equation", + "content": "\\{\\lambda_v^t\\}_{v = 1}^{|\\mathcal{V}|}" + }, + { + "bbox": [ + 47, + 449, + 287, + 573 + ], + "type": "text", + "content": " and only update " + }, + { + "bbox": [ + 47, + 449, + 287, + 573 + ], + "type": "inline_equation", + "content": "c_{vu}" + }, + { + "bbox": [ + 47, + 449, + 287, + 573 + ], + "type": "text", + "content": " in this step." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 66, + 582, + 287, + 594 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 582, + 287, + 594 + ], + "spans": [ + { + "bbox": [ + 66, + 582, + 287, + 594 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {C O T}} = \\mathcal {L} _ {\\text {c r o s s - e n t r o p y}} + \\alpha_ {1} \\cdot \\mathcal {L} _ {\\text {c l u s t e r}} + \\alpha_ {2} \\cdot \\mathcal {L} _ {\\mathrm {O T}} \\tag {12}", + "image_path": "224cc32683f139602f8cb94918138e43fb42984d1bd078dd55a671110eb12138.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 47, + 603, + 287, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 603, + 287, + 651 + ], + "spans": [ + { + "bbox": [ + 47, + 603, + 287, + 651 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 603, + 287, + 651 + ], + "type": "inline_equation", + "content": "\\alpha_{1}" + }, + { + "bbox": [ + 47, + 603, + 287, + 651 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 603, + 287, + 651 + ], + "type": "inline_equation", + "content": "\\alpha_{2}" + }, + { + "bbox": [ + 47, + 603, + 287, + 651 + ], + "type": "text", + "content": " are loss weights. Finally, we elaborate the overall optimization strategy of our COT in Algorithm 1. Line 6 of Algorithm 1 represents we only optimize " + }, + { + "bbox": [ + 47, + 603, + 287, + 651 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{OT}" + }, + { + "bbox": [ + 47, + 603, + 287, + 651 + ], + "type": "text", + "content": " in the initial " + }, + { + "bbox": [ + 47, + 603, + 287, + 651 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 47, + 603, + 287, + 651 + ], + "type": "text", + "content": " iterations of each epoch." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 47, + 658, + 287, + 683 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 658, + 287, + 683 + ], + "spans": [ + { + "bbox": [ + 47, + 658, + 287, + 683 + ], + "type": "text", + "content": "4.2. Theoretical Analysis on Instance/Clustering Optimal Transport" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 47, + 687, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 687, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 47, + 687, + 287, + 715 + ], + "type": "text", + "content": "Given features " + }, + { + "bbox": [ + 47, + 687, + 287, + 715 + ], + "type": "inline_equation", + "content": "\\{x_i^s\\}_{i = 1}^{n_s}" + }, + { + "bbox": [ + 47, + 687, + 287, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 687, + 287, + 715 + ], + "type": "inline_equation", + "content": "\\{x_j^t\\}_{j = 1}^{n_t}" + }, + { + "bbox": [ + 47, + 687, + 287, + 715 + ], + "type": "text", + "content": " from source and target domain respectively, where " + }, + { + "bbox": [ + 47, + 687, + 287, + 715 + ], + "type": "inline_equation", + "content": "x_{i}^{s}" + }, + { + "bbox": [ + 47, + 687, + 287, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 687, + 287, + 715 + ], + "type": "inline_equation", + "content": "x_{j}^{t}" + }, + { + "bbox": [ + 47, + 687, + 287, + 715 + ], + "type": "text", + "content": " are output from" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 305, + 72, + 545, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 545, + 96 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 545, + 96 + ], + "type": "text", + "content": "shared-parameters neural network for feature extractor. We consider the discrete Kantorovich problem" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 343, + 106, + 545, + 176 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 343, + 106, + 545, + 176 + ], + "spans": [ + { + "bbox": [ + 343, + 106, + 545, + 176 + ], + "type": "interline_equation", + "content": "\\left\\{ \\begin{array}{c} \\min _ {T \\in \\mathbb {R} ^ {n _ {s} \\times n _ {t}}} T _ {i j} c _ {i j} \\\\ s. t. \\sum_ {j = 1} ^ {n _ {t}} T _ {i j} = \\frac {1}{n _ {s}}, \\sum_ {i = 1} ^ {n _ {s}} T _ {i j} = \\frac {1}{n _ {t}}, \\\\ T _ {i j} \\geq 0, \\forall 1 \\leq i \\leq n _ {s}, 1 \\leq j \\leq n _ {t}. \\end{array} \\right. \\tag {13}", + "image_path": "333fa10569ac503f1380be6df9152c703443fe48152f52ed7bce54cf61753749.jpg" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 306, + 186, + 410, + 201 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 186, + 410, + 201 + ], + "spans": [ + { + "bbox": [ + 306, + 186, + 410, + 201 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 306, + 186, + 410, + 201 + ], + "type": "inline_equation", + "content": "c_{ij} = \\| x_i^s -x_j^t\\| _2^2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 305, + 201, + 545, + 223 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 201, + 545, + 223 + ], + "spans": [ + { + "bbox": [ + 305, + 201, + 545, + 223 + ], + "type": "text", + "content": "Considering the distance and inner-product between features and classifiers:" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 322, + 232, + 545, + 293 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 322, + 232, + 545, + 293 + ], + "spans": [ + { + "bbox": [ + 322, + 232, + 545, + 293 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\left\\| x _ {i} ^ {s} - w _ {v _ {1} ^ {s}} \\right\\| _ {2} ^ {2} - \\left\\| x _ {i} ^ {s} - w _ {v _ {2}} ^ {s} \\right\\| _ {2} ^ {2} \\\\ = \\| w _ {v _ {1}} ^ {s} \\| _ {2} ^ {2} - \\| w _ {v _ {2}} ^ {s} \\| _ {2} ^ {2} \\tag {14} \\\\ + 2 \\big (\\| w _ {v _ {2}} ^ {s} \\| _ {2} x _ {i} ^ {s} ^ {\\top} \\frac {w _ {v _ {2}} ^ {s}}{\\| w _ {v _ {2}} ^ {s} \\| _ {2}} - \\| w _ {v _ {1}} ^ {s} \\| _ {2} x _ {i} ^ {s} ^ {\\top} \\frac {w _ {v _ {1}} ^ {s}}{\\| w _ {v _ {1}} ^ {s} \\| _ {2}} \\big) \\\\ \\end{array}", + "image_path": "259187f0dc6325ac507a91c2f08d9b559ad59f22fe67bf57dedfae14241e7e4d.jpg" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 305, + 302, + 545, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 302, + 545, + 426 + ], + "spans": [ + { + "bbox": [ + 305, + 302, + 545, + 426 + ], + "type": "text", + "content": "In the Bayesian view, we can consider " + }, + { + "bbox": [ + 305, + 302, + 545, + 426 + ], + "type": "inline_equation", + "content": "\\| w_v^s\\| _2" + }, + { + "bbox": [ + 305, + 302, + 545, + 426 + ], + "type": "text", + "content": " as the prior probability of class " + }, + { + "bbox": [ + 305, + 302, + 545, + 426 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 305, + 302, + 545, + 426 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 305, + 302, + 545, + 426 + ], + "type": "inline_equation", + "content": "x_{i}^{s}" + }, + { + "bbox": [ + 305, + 302, + 545, + 426 + ], + "type": "text", + "content": " is feature representation of a sample and " + }, + { + "bbox": [ + 305, + 302, + 545, + 426 + ], + "type": "inline_equation", + "content": "\\frac{w_s^s}{\\|w_v^s\\|_2}" + }, + { + "bbox": [ + 305, + 302, + 545, + 426 + ], + "type": "text", + "content": " is the cluster for class " + }, + { + "bbox": [ + 305, + 302, + 545, + 426 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 305, + 302, + 545, + 426 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 305, + 302, + 545, + 426 + ], + "type": "inline_equation", + "content": "x_{i}^{s\\top}\\frac{w_{v}^{s}}{\\|w_{v}^{s}\\|_{2}}" + }, + { + "bbox": [ + 305, + 302, + 545, + 426 + ], + "type": "text", + "content": " measures the similarity between feature and cluster. When classifiers in " + }, + { + "bbox": [ + 305, + 302, + 545, + 426 + ], + "type": "inline_equation", + "content": "\\{\\| w_v^s\\| _2\\}_{v = 1}^{\\lfloor \\mathcal{V}\\rfloor}" + }, + { + "bbox": [ + 305, + 302, + 545, + 426 + ], + "type": "text", + "content": " are of the same magnitude, we conclude that the similarity between features and clusters are almost equivalent to the distance between features and classifiers. With labels as supervision, the optimization of crossentropy can promote the inter-class discrepancy which implies" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 350, + 428, + 545, + 458 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 350, + 428, + 545, + 458 + ], + "spans": [ + { + "bbox": [ + 350, + 428, + 545, + 458 + ], + "type": "interline_equation", + "content": "x _ {i} ^ {s \\top} \\frac {w _ {y _ {i} ^ {s}}}{\\| w _ {y _ {i} ^ {s}} \\| _ {2}} \\gg x _ {i} ^ {s \\top} \\frac {w _ {v} ^ {s}}{\\| w _ {v} ^ {s} \\| _ {2}}, \\forall v \\neq y _ {i} ^ {s} \\tag {15}", + "image_path": "2418c6636a232b92ff68f5436e2751725d62533545321bcf883b5f62f00466ac.jpg" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 306, + 464, + 465, + 476 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 464, + 465, + 476 + ], + "spans": [ + { + "bbox": [ + 306, + 464, + 465, + 476 + ], + "type": "text", + "content": "which also provide the following result" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 347, + 487, + 545, + 503 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 347, + 487, + 545, + 503 + ], + "spans": [ + { + "bbox": [ + 347, + 487, + 545, + 503 + ], + "type": "interline_equation", + "content": "\\left\\| x _ {i} ^ {s} - w _ {y _ {i} ^ {s}} ^ {s} \\right\\| _ {2} ^ {2} \\ll \\left\\| x _ {i} ^ {s} - w _ {v} ^ {s} \\right\\| _ {2} ^ {2}, \\forall v \\neq y _ {i} ^ {s} \\tag {16}", + "image_path": "711d12ed757f8244f228d2da8c62c8fbc0b9b07e5cfa49dc43c09d92aef684dc.jpg" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 305, + 513, + 545, + 584 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 513, + 545, + 584 + ], + "spans": [ + { + "bbox": [ + 305, + 513, + 545, + 584 + ], + "type": "text", + "content": "If clustering doesn't work sufficiently well, it happens that some samples in the source domain with label " + }, + { + "bbox": [ + 305, + 513, + 545, + 584 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 305, + 513, + 545, + 584 + ], + "type": "text", + "content": " are assigned to samples in the target domain with label " + }, + { + "bbox": [ + 305, + 513, + 545, + 584 + ], + "type": "inline_equation", + "content": "u \\neq v" + }, + { + "bbox": [ + 305, + 513, + 545, + 584 + ], + "type": "text", + "content": ". When clustering performs well, We have " + }, + { + "bbox": [ + 305, + 513, + 545, + 584 + ], + "type": "inline_equation", + "content": "c_{ij} \\sim \\bar{c}_{vj} = \\| w_v^s - x_j^t\\|_2^2" + }, + { + "bbox": [ + 305, + 513, + 545, + 584 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 305, + 513, + 545, + 584 + ], + "type": "inline_equation", + "content": "\\sim" + }, + { + "bbox": [ + 305, + 513, + 545, + 584 + ], + "type": "text", + "content": " means these two numbers are almost the same. Then we get" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 359, + 595, + 545, + 622 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 359, + 595, + 545, + 622 + ], + "spans": [ + { + "bbox": [ + 359, + 595, + 545, + 622 + ], + "type": "interline_equation", + "content": "\\sum_ {i, j} T _ {i j} c _ {i j} \\sim \\sum_ {v, j} \\left(\\sum_ {x _ {i} ^ {s} \\in X _ {v} ^ {s}} T _ {i j}\\right) \\bar {c} _ {v j} \\tag {17}", + "image_path": "af461f059b7fa6e2abbf35133a4937eb849724975208def390832ae28a0d9d4e.jpg" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 305, + 632, + 545, + 669 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 632, + 545, + 669 + ], + "spans": [ + { + "bbox": [ + 305, + 632, + 545, + 669 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 632, + 545, + 669 + ], + "type": "inline_equation", + "content": "X_{v}^{s}" + }, + { + "bbox": [ + 305, + 632, + 545, + 669 + ], + "type": "text", + "content": " is the set of samples with label " + }, + { + "bbox": [ + 305, + 632, + 545, + 669 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 305, + 632, + 545, + 669 + ], + "type": "text", + "content": " in the source domain. We denote the number of samples with class " + }, + { + "bbox": [ + 305, + 632, + 545, + 669 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 305, + 632, + 545, + 669 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 305, + 632, + 545, + 669 + ], + "type": "inline_equation", + "content": "n_v" + }, + { + "bbox": [ + 305, + 632, + 545, + 669 + ], + "type": "text", + "content": ", then we get" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 324, + 679, + 545, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 324, + 679, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 324, + 679, + 545, + 715 + ], + "type": "interline_equation", + "content": "\\sum_ {j = 1} ^ {n _ {t}} \\left(\\sum_ {x _ {i} ^ {s} \\in X _ {v} ^ {s}} T _ {i j}\\right) = \\frac {n _ {v}}{n _ {s}}, \\sum_ {v = 1} ^ {| \\mathcal {Y} |} \\left(\\sum_ {x _ {i} ^ {s} \\in X _ {v} ^ {s}} T _ {i j}\\right) = \\frac {1}{n _ {t}} \\tag {18}", + "image_path": "ff6d17d54de906f858e7864600ac5affd4bc2ff99a9f22905e6a4cf406200800.jpg" + } + ] + } + ], + "index": 32 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20001" + } + ] + } + ], + "index": 33 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 100, + 90, + 496, + 325 + ], + "blocks": [ + { + "bbox": [ + 143, + 70, + 456, + 82 + ], + "lines": [ + { + "bbox": [ + 143, + 70, + 456, + 82 + ], + "spans": [ + { + "bbox": [ + 143, + 70, + 456, + 82 + ], + "type": "text", + "content": "Table 156 Accuracy (%) on Office-31 for UDA (ResNet-50).The best result is in bold." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 100, + 90, + 496, + 325 + ], + "lines": [ + { + "bbox": [ + 100, + 90, + 496, + 325 + ], + "spans": [ + { + "bbox": [ + 100, + 90, + 496, + 325 + ], + "type": "table", + "html": "
MethodA→WD→WW→DA→DD→AW→AAvg
Common UDAADDA [46]86.296.298.477.869.568.982.9
JAN [34]85.497.499.884.768.670.084.3
MCD [42]88.698.5100.092.269.569.786.5
GTA [43]89.597.999.887.772.871.486.5
CDAN [33]94.198.6100.092.971.069.387.7
TAT [32]92.599.3100.093.273.172.188.4
MDD [25]94.598.4100.093.574.672.288.9
GSP [19]92.998.799.894.575.974.989.5
DANN [1]82.096.999.179.768.267.482.2
SHOT [29]94.090.174.798.474.399.988.6
MCC [21]95.598.6100.094.472.974.989.4
GVB-GD [9]94.898.7100.095.073.473.789.3
TSA [28]96.098.7100.095.476.776.890.6
SRDC [45]95.799.2100.095.876.777.190.8
OT-basedJDOT [5]84.797.8100.086.464.467.783.5
DeepJDOT [11]88.998.599.688.272.170.186.2
MLOT [23]92.898.5100.090.872.871.687.8
RWOT [51]95.199.5100.094.577.577.990.8
DANN [1] + MMI [28]95.298.6100.094.474.675.289.7
DANN + MMI + COT (Ours)96.599.1100.096.176.777.491.0
", + "image_path": "95f2782fb7eed4b97c3c520b577cf220a14a103344c188df8a4d62993446558a.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 339, + 287, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 339, + 287, + 389 + ], + "spans": [ + { + "bbox": [ + 46, + 339, + 287, + 389 + ], + "type": "text", + "content": "Inspired by Equation (17) and (18), we consider the following optimal transport between clusters from the source domain and instances from the target domain instead of solving the Kantorovich problem in Equation (13)." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 83, + 392, + 287, + 466 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 83, + 392, + 287, + 466 + ], + "spans": [ + { + "bbox": [ + 83, + 392, + 287, + 466 + ], + "type": "interline_equation", + "content": "\\left\\{ \\begin{array}{c} \\min _ {\\bar {T} \\in \\mathbb {R} ^ {| Y | \\times n _ {t}}} \\bar {T} _ {v j} \\bar {c} _ {v j} \\\\ s. t. \\sum_ {j = 1} ^ {n _ {t}} \\bar {T} _ {v j} = \\frac {n _ {v}}{n _ {s}}, \\sum_ {v = 1} ^ {| \\mathcal {Y} |} \\bar {T} _ {v j} = \\frac {1}{n _ {t}}, \\\\ \\bar {T} _ {v j} \\geq 0, \\forall 1 \\leq v \\leq | \\mathcal {Y} |, 1 \\leq j \\leq n _ {t} \\end{array} \\right. \\tag {19}", + "image_path": "3b31123720e84ad0dc82c8f07271ea392a32b57f406c23bbc2f996e24eff17af.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 469, + 287, + 505 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 469, + 287, + 505 + ], + "spans": [ + { + "bbox": [ + 46, + 469, + 287, + 505 + ], + "type": "text", + "content": "In general, because of the class imbalance, the empirical label distribution between source and target dataset are different" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 52, + 508, + 287, + 536 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 52, + 508, + 287, + 536 + ], + "spans": [ + { + "bbox": [ + 52, + 508, + 287, + 536 + ], + "type": "interline_equation", + "content": "\\exists \\delta > 0, \\text {s . t .} \\| \\left(\\frac {n _ {1} ^ {s}}{n _ {s}}, \\dots , \\frac {n _ {| \\mathcal {Y} |} ^ {s}}{n _ {s}}\\right) - \\left(\\frac {n _ {1} ^ {t}}{n _ {t}}, \\dots , \\frac {n _ {| \\mathcal {Y} |} ^ {t}}{n _ {t}}\\right) \\| _ {2} \\geq \\delta \\tag {20}", + "image_path": "c7ef269624352593c2b3c3e533ae885f3c93a00378abc29768cbef2c3102e555.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 538, + 287, + 648 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 538, + 287, + 648 + ], + "spans": [ + { + "bbox": [ + 46, + 538, + 287, + 648 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 538, + 287, + 648 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 46, + 538, + 287, + 648 + ], + "type": "text", + "content": " is a constant which measures the label distribution between source and target domain. There must exist some index " + }, + { + "bbox": [ + 46, + 538, + 287, + 648 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 538, + 287, + 648 + ], + "type": "text", + "content": " such that " + }, + { + "bbox": [ + 46, + 538, + 287, + 648 + ], + "type": "inline_equation", + "content": "\\frac{n_v^s}{n_s} >\\frac{n_v^t}{n_t}" + }, + { + "bbox": [ + 46, + 538, + 287, + 648 + ], + "type": "text", + "content": ", which means that some samples with label " + }, + { + "bbox": [ + 46, + 538, + 287, + 648 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 46, + 538, + 287, + 648 + ], + "type": "text", + "content": " in the source domain will be assigned to samples in the target domain with label " + }, + { + "bbox": [ + 46, + 538, + 287, + 648 + ], + "type": "inline_equation", + "content": "u\\neq v" + }, + { + "bbox": [ + 46, + 538, + 287, + 648 + ], + "type": "text", + "content": ". This will result in samples belonging to the same category in the target domain are given different pseudo labels, which increase the difficulty of training and cause the degradation of the performance of deep learning methods on the target domain." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 649, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 649, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 649, + 287, + 715 + ], + "type": "text", + "content": "When we utilize the clustering based optimal transport, for source domain, we have " + }, + { + "bbox": [ + 46, + 649, + 287, + 715 + ], + "type": "inline_equation", + "content": "\\sum_{u=1}^{K} T_{vu} = \\frac{1}{|\\mathcal{V}|}" + }, + { + "bbox": [ + 46, + 649, + 287, + 715 + ], + "type": "text", + "content": ". For target domain, " + }, + { + "bbox": [ + 46, + 649, + 287, + 715 + ], + "type": "inline_equation", + "content": "\\sum_{v=1}^{|\\mathcal{V}|} \\sum_{u=(l-1) \\cdot Q+1}^{l \\cdot Q} T_{vu} = \\frac{Q}{K} = \\frac{1}{|\\mathcal{V}|}" + }, + { + "bbox": [ + 46, + 649, + 287, + 715 + ], + "type": "text", + "content": ", which ease the" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 339, + 545, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 339, + 545, + 365 + ], + "spans": [ + { + "bbox": [ + 305, + 339, + 545, + 365 + ], + "type": "text", + "content": "negative effect from class imbalance in domain adaptation based on optimal transport." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 373, + 416, + 387 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 373, + 416, + 387 + ], + "spans": [ + { + "bbox": [ + 305, + 373, + 416, + 387 + ], + "type": "text", + "content": "4.3. Computation Cost" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 392, + 545, + 440 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 392, + 545, + 440 + ], + "spans": [ + { + "bbox": [ + 305, + 392, + 545, + 440 + ], + "type": "text", + "content": "In terms of instance-based optimal transport, firstly we need to obtain the features of all samples from the source and target domain, computation cost on the feature extractor is shown as follows:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 355, + 452, + 545, + 465 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 355, + 452, + 545, + 465 + ], + "spans": [ + { + "bbox": [ + 355, + 452, + 545, + 465 + ], + "type": "interline_equation", + "content": "\\mathcal {O} \\left(n _ {s} + n _ {t}\\right) \\cdot \\mathcal {O} (\\text {f e a t u r e - e x t r a c t o r}) \\tag {21}", + "image_path": "0af5f6fa957c67322493a5046c601d0eb46cb3b579b2e26d8e7b2e630d25c64a.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 475, + 545, + 572 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 475, + 545, + 572 + ], + "spans": [ + { + "bbox": [ + 305, + 475, + 545, + 572 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 475, + 545, + 572 + ], + "type": "inline_equation", + "content": "\\mathcal{O}" + }, + { + "bbox": [ + 305, + 475, + 545, + 572 + ], + "type": "text", + "content": " (feature-extractor) means the computation cost on single sample when extracting the feature. Then considering the optimization of optimal transport, every iteration will need " + }, + { + "bbox": [ + 305, + 475, + 545, + 572 + ], + "type": "inline_equation", + "content": "\\mathcal{O}(n_s\\cdot n_t)" + }, + { + "bbox": [ + 305, + 475, + 545, + 572 + ], + "type": "text", + "content": ". In comparison, for cluster-based optimal transport, the main computation cost on optimal transport is " + }, + { + "bbox": [ + 305, + 475, + 545, + 572 + ], + "type": "inline_equation", + "content": "\\mathcal{O}(|\\mathcal{Y}|\\cdot K)" + }, + { + "bbox": [ + 305, + 475, + 545, + 572 + ], + "type": "text", + "content": ". For a large-scale dataset, the clustering-based optimal transport cost is much less than instance-based optimal transport." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 584, + 388, + 599 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 584, + 388, + 599 + ], + "spans": [ + { + "bbox": [ + 306, + 584, + 388, + 599 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "type": "text", + "content": "In this section, we first elaborate on implementation details and 3 authoritative benchmark datasets in the field of UDA. Then we compare our COT with existing OT-based UDA algorithms and state-of-the-art UDA methods, illustrating that our COT achieves the dominant result in the field of OT-based UDA and competitive performance on the common UDA realm, respectively. Furthermore, we provide some qualitative analysis to illustrate the advantage of our COT when compared with some clustering-based DA" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "footer", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "20002" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 49, + 91, + 547, + 275 + ], + "blocks": [ + { + "bbox": [ + 161, + 70, + 440, + 82 + ], + "lines": [ + { + "bbox": [ + 161, + 70, + 440, + 82 + ], + "spans": [ + { + "bbox": [ + 161, + 70, + 440, + 82 + ], + "type": "text", + "content": "Table 256 Accuracy " + }, + { + "bbox": [ + 161, + 70, + 440, + 82 + ], + "type": "inline_equation", + "content": "(\\%)" + }, + { + "bbox": [ + 161, + 70, + 440, + 82 + ], + "type": "text", + "content": " on Office-Home for UDA.The best result is in bold." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 49, + 91, + 547, + 275 + ], + "lines": [ + { + "bbox": [ + 49, + 91, + 547, + 275 + ], + "spans": [ + { + "bbox": [ + 49, + 91, + 547, + 275 + ], + "type": "table", + "html": "
MethodAr→ClAr→PrAr→RwCl→ArCl→PrCl→RwPr→ArPr→ClPr→RwRw→ArRw→ClRw→PrAvg
Common UDAJAN [34]45.961.268.950.459.761.045.843.470.363.952.476.858.3
TAT [32]51.669.575.459.469.568.659.550.576.870.956.681.665.8
TPN [39]51.271.276.065.172.972.855.448.976.570.953.480.466.2
ETD [26]51.371.985.757.669.273.757.851.279.370.257.582.167.3
SymNets [56]47.772.978.564.271.374.264.248.879.574.552.682.767.6
BNM [8]52.373.980.063.372.974.961.749.579.770.553.682.267.9
MDD [25]54.973.777.860.071.471.861.253.678.172.560.282.368.1
GSP [19]56.875.578.961.369.474.961.352.679.973.354.283.268.4
MCD [42]48.968.374.661.367.668.85747.175.169.152.279.667.8
DANN [1]45.659.370.147.058.560.946.143.768.563.251.876.857.6
CDAN [33]50.770.676.057.670.070.057.450.977.370.956.781.665.8
BSP [4]52.068.676.158.070.370.258.650.277.672.259.381.966.3
TSA [28]55.873.779.061.974.674.560.753.280.172.758.484.369.1
GVB-GD [9]57.074.779.864.674.174.665.255.181.074.659.784.370.4
SRDC [45]52.376.381.069.576.278.068.753.881.776.357.185.071.3
OT-basedJDOT [5]44.763.468.655.164.460.754.348.472.867.457.269.460.5
DeepJDOT [11]39.750.462.539.554.453.236.739.263.652.345.470.550.6
MLOT [23]52.472.677.858.872.472.853.250.876.570.954.482.366.2
DANN [1] + MMI [28]55.774.880.463.674.176.664.254.780.574.358.281.369.9
DANN + MMI + COT (Ours)57.675.283.267.876.275.765.456.282.475.160.784.771.7
Ours + CDTrans [52]70.186.386.482.288.689.177.265.789.283.267.489.881.3
", + "image_path": "cff87c2047d2a31a9f09d6910ad6aef326eeab843b9de3fd90077c3ae3e289a7.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 300, + 287, + 337 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 300, + 287, + 337 + ], + "spans": [ + { + "bbox": [ + 46, + 300, + 287, + 337 + ], + "type": "text", + "content": "methods in the Optimal Transport realm. Finally, we conduct ablative experiments together with visualization results to further demonstrate the effectiveness of our COT." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 346, + 243, + 358 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 346, + 243, + 358 + ], + "spans": [ + { + "bbox": [ + 47, + 346, + 243, + 358 + ], + "type": "text", + "content": "5.1. Datasets and Implementation Details" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 365, + 287, + 435 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 365, + 287, + 435 + ], + "spans": [ + { + "bbox": [ + 46, + 365, + 287, + 435 + ], + "type": "text", + "content": "Office-31 is a benchmark dataset on the real-world unsupervised domain adaptation. It has 4110 images for 31 classes drawn from three domains: Amazon (A), DSLR (D), and Webcam (W). The 31 classes in the dataset consist of objects that commonly appeared in office settings, such as keyboards, file cabinets, and laptops." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 437, + 287, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 437, + 287, + 533 + ], + "spans": [ + { + "bbox": [ + 46, + 437, + 287, + 533 + ], + "type": "text", + "content": "Office-Home is a challenging benchmark dataset for domain adaptation which has 4 domains where each domain consists of 65 categories. The four domains are Art - artistic images in the form of sketches, paintings, ornamentation, etc.; Clipart - a collection of clipart images; Product - images of objects without a background and Real-World - images of objects captured with a regular camera. It contains 15,500 images in 65 classes." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 533, + 287, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 533, + 287, + 605 + ], + "spans": [ + { + "bbox": [ + 46, + 533, + 287, + 605 + ], + "type": "text", + "content": "VisDa-2017 is a large-scale simulation-to-real dataset for domain adaptation, which has over 280,000 images across 12 categories in the training, validation, and testing domains. The training images are generated from the same object under different circumstances, while the validation images are collected from MSCOCO ([31])." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "text", + "content": "Implementation Details Follow GVB ([10]), we adopt ResNet-50 pretrained on the ImageNet ([12]) as our backbone for Office-31 and Office-Home benchmarks and ResNet-101 for VisDa-2017 dataset. Note that our COT is a plug-and-play module, indicating it can integrate with a vast body of existing UDA approaches, e.g. DANN [1], CDAN [33]. If there is no extra statement, our COT is implemented on DANN with Maximization Mutual Information Loss (MMI) [28]. In this paper, all experiments are" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 300, + 547, + 349 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 300, + 547, + 349 + ], + "spans": [ + { + "bbox": [ + 304, + 300, + 547, + 349 + ], + "type": "text", + "content": "implemented by PyTorch. For the optimizer schedule, we adopt SGD with a momentum of 0.9. The total number of training epochs is 40,40,30 on Office-31, Office-Home, and VisDA-2017." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 370, + 362, + 382 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 370, + 362, + 382 + ], + "spans": [ + { + "bbox": [ + 306, + 370, + 362, + 382 + ], + "type": "text", + "content": "5.2. Results" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 393, + 547, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 393, + 547, + 525 + ], + "spans": [ + { + "bbox": [ + 304, + 393, + 547, + 525 + ], + "type": "text", + "content": "Results on Office-Home. Table 2 presents the results of OT-based and common UDA methods on Office-Home dataset. Our COT achieves the highest accuracy " + }, + { + "bbox": [ + 304, + 393, + 547, + 525 + ], + "type": "inline_equation", + "content": "71.7\\%" + }, + { + "bbox": [ + 304, + 393, + 547, + 525 + ], + "type": "text", + "content": ", outperforming other OT-based UDA algorithms by " + }, + { + "bbox": [ + 304, + 393, + 547, + 525 + ], + "type": "inline_equation", + "content": "1.8\\%" + }, + { + "bbox": [ + 304, + 393, + 547, + 525 + ], + "type": "text", + "content": " at least. Such tremendous improvements demonstrate that our COT can capture a more accurate local-level alignment than previous DA-based methods'. Besides, when compared with other state-of-the-art common UDA methods, to the best of our knowledge, our COT is the first OT-based UDA work that can perform a significant competitive result on the Office-Home dataset." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 529, + 547, + 565 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 529, + 547, + 565 + ], + "spans": [ + { + "bbox": [ + 304, + 529, + 547, + 565 + ], + "type": "text", + "content": "Results on Office-31. The results are reported in the Table 1. Our COT achieves the best performance " + }, + { + "bbox": [ + 304, + 529, + 547, + 565 + ], + "type": "inline_equation", + "content": "(91.0\\%)" + }, + { + "bbox": [ + 304, + 529, + 547, + 565 + ], + "type": "text", + "content": " both on the common and OT-based UDA." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 570, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 547, + 713 + ], + "type": "text", + "content": "Results on VisDA-2017. Table 3 shows the performance of OT-based and common UDA methods on the challenging VisDa-2017 dataset. Our COT outperforms the existing OT-based / common UDA methods (CNN backbone) by " + }, + { + "bbox": [ + 304, + 570, + 547, + 713 + ], + "type": "inline_equation", + "content": "3.1\\%" + }, + { + "bbox": [ + 304, + 570, + 547, + 713 + ], + "type": "text", + "content": " / " + }, + { + "bbox": [ + 304, + 570, + 547, + 713 + ], + "type": "inline_equation", + "content": "3.2\\%" + }, + { + "bbox": [ + 304, + 570, + 547, + 713 + ], + "type": "text", + "content": " at least, illustrating the dominant role of our COT on UDA task. Furthermore, due to the plug-and-play property of our COT, we directly add this module into the sota UDA method (CDTrans [52] on the challenging dataset VisDa2017 and Office-Home. The results in Table 3 show that our method achieves a new state-of-the-art performance, indicating the superiority and appealing potential of our COT." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20003" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 50, + 91, + 547, + 298 + ], + "blocks": [ + { + "bbox": [ + 134, + 70, + 465, + 82 + ], + "lines": [ + { + "bbox": [ + 134, + 70, + 465, + 82 + ], + "spans": [ + { + "bbox": [ + 134, + 70, + 465, + 82 + ], + "type": "text", + "content": "Table 356 Accuracy (%) on VisDA-2017 for UDA (ResNet-101). The best result is in bold." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 50, + 91, + 547, + 298 + ], + "lines": [ + { + "bbox": [ + 50, + 91, + 547, + 298 + ], + "spans": [ + { + "bbox": [ + 50, + 91, + 547, + 298 + ], + "type": "table", + "html": "
Methodplanebcyblbuscarhorseknifemcylepersnplantsktbtraintruckmean
Common UDADANN [1]81.977.782.844.381.229.565.128.651.954.682.87.857.4
MinEnt [18]87.455.075.363.887.443.689.372.582.978.685.627.470.7
TSA [28]93.077.882.250.889.928.077.170.085.280.086.143.071.9
BSP [4]92.272.583.847.587.054.086.872.480.666.984.537.172.1
MCC [21]90.479.872.355.190.586.886.680.094.276.990.049.679.4
MODEL [27]94.873.468.874.893.195.488.684.789.184.783.548.181.6
STAR [35]95.084.084.673.091.691.885.978.494.484.787.042.282.7
BNM [8]89.661.576.955.089.369.181.365.590.047.389.130.170.4
MSTN+DSBN [3]94.786.776.072.095.275.187.981.391.168.988.345.580.2
CGDM [14]92.885.176.364.591.093.281.379.392.483.085.644.880.8
SHOT [29]94.388.580.157.393.193.180.780.391.589.186.358.282.9
TVT [55]92.985.677.560.593.698.289.476.493.692.091.755.783.9
CDTrans [52]98.086.987.980.997.997.396.885.397.683.294.054.488.4
OT-basedJDOT [5]78.470.879.468.882.380.584.270.788.468.878.445.774.7
DeepJDOT [11]85.473.477.387.384.164.791.579.391.944.488.561.877.4
MLOT [23]88.270.477.350.284.877.280.474.483.868.282.338.773.0
RWOT [51]95.180.383.790.092.468.092.582.287.978.490.468.284.0
DANN+MMI+COT (Ours)96.989.684.274.196.496.588.682.096.094.185.162.187.1
Ours + CDTrans [52]98.289.487.682.398.097.296.486.298.392.692.258.189.7
", + "image_path": "f48ea8f42b93b423813898cc2d8464e520e5cd9dc7cbd58e2a279ef01ef44911.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 71, + 331, + 156, + 415 + ], + "blocks": [ + { + "bbox": [ + 71, + 331, + 156, + 415 + ], + "lines": [ + { + "bbox": [ + 71, + 331, + 156, + 415 + ], + "spans": [ + { + "bbox": [ + 71, + 331, + 156, + 415 + ], + "type": "image", + "image_path": "808b1059f7ca154c322db66d294d8ce68f07873d475e1a1d2cc9ca18626644e4.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 96, + 423, + 132, + 434 + ], + "lines": [ + { + "bbox": [ + 96, + 423, + 132, + 434 + ], + "spans": [ + { + "bbox": [ + 96, + 423, + 132, + 434 + ], + "type": "text", + "content": "(a) RWOT" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 176, + 331, + 261, + 414 + ], + "blocks": [ + { + "bbox": [ + 176, + 331, + 261, + 414 + ], + "lines": [ + { + "bbox": [ + 176, + 331, + 261, + 414 + ], + "spans": [ + { + "bbox": [ + 176, + 331, + 261, + 414 + ], + "type": "image", + "image_path": "39b0392be670efa9997417b0c3a9d2a55ae9b92433ea01c16e37cb857b0cf4b4.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 206, + 423, + 235, + 433 + ], + "lines": [ + { + "bbox": [ + 206, + 423, + 235, + 433 + ], + "spans": [ + { + "bbox": [ + 206, + 423, + 235, + 433 + ], + "type": "text", + "content": "(b)COT" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 443, + 287, + 466 + ], + "lines": [ + { + "bbox": [ + 46, + 443, + 287, + 466 + ], + "spans": [ + { + "bbox": [ + 46, + 443, + 287, + 466 + ], + "type": "text", + "content": "Figure 1. t-SNE of classifier responses by RWOT and COT on CI50 (red: Amazon, blue: Webcam)." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 490, + 139, + 502 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 490, + 139, + 502 + ], + "spans": [ + { + "bbox": [ + 47, + 490, + 139, + 502 + ], + "type": "text", + "content": "5.3. Ablation study" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 510, + 289, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 510, + 289, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 510, + 289, + 714 + ], + "type": "text", + "content": "Ability on Resisting Class-imbalance Challenge. As the results illustrated above, our COT brings little gain than the existing sota OT-based UDA method (RWOT [51]) on the Office-31 dataset while achieving the dominant performance on Office-Home and VisDA-2017 datasets. To find the reason for this, we compute the label distribution of the Office-31 dataset, where the class imbalance phenomenon is not occurred obviously compared to the Office-Home and VisDA-2017 datasets (shown in Fig. 2). Thus, to verify the ability of our COT on resisting class-imbalance challenges among all UDA benchmark datasets, we construct 3 types of class-imbalance evaluation benchmarks by randomly erasing " + }, + { + "bbox": [ + 46, + 510, + 289, + 714 + ], + "type": "inline_equation", + "content": "30\\%" + }, + { + "bbox": [ + 46, + 510, + 289, + 714 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 510, + 289, + 714 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 46, + 510, + 289, + 714 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 510, + 289, + 714 + ], + "type": "inline_equation", + "content": "70\\%" + }, + { + "bbox": [ + 46, + 510, + 289, + 714 + ], + "type": "text", + "content": " samples from certain classes on Office-31 datasets target domain, termed CI30, CI50, and CI70, respectively. Table 4 presents the results of our COT and existing OT-based UDA methods on Office-31, CI30, CI50, and CI70. According to this quantitative result, we dis" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 312, + 545, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 312, + 545, + 361 + ], + "spans": [ + { + "bbox": [ + 305, + 312, + 545, + 361 + ], + "type": "text", + "content": "cover that the more severe the class imbalance phenomenon occurs, the larger the performance gap between our COT and RWOT, demonstrating that our COT embraces excellent ability on handling class-imbalance challenges." + } + ] + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 314, + 400, + 538, + 464 + ], + "blocks": [ + { + "bbox": [ + 305, + 369, + 545, + 391 + ], + "lines": [ + { + "bbox": [ + 305, + 369, + 545, + 391 + ], + "spans": [ + { + "bbox": [ + 305, + 369, + 545, + 391 + ], + "type": "text", + "content": "Table 4. Mean accuracy of OT-based UDA methods on class-imbalance datasets." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 314, + 400, + 538, + 464 + ], + "lines": [ + { + "bbox": [ + 314, + 400, + 538, + 464 + ], + "spans": [ + { + "bbox": [ + 314, + 400, + 538, + 464 + ], + "type": "table", + "html": "
DatasetCOTRWOTMLOT
Office-3191.090.887.8
CI3088.284.783.6
CI5087.778.380.4
CI7085.470.575.2
", + "image_path": "a592e923b6e8d7883c52125824420db63305344c22118575346e3f4da81c3429.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 474, + 545, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 474, + 545, + 569 + ], + "spans": [ + { + "bbox": [ + 304, + 474, + 545, + 569 + ], + "type": "text", + "content": "Computation Cost. Here, we compute the computation cost of our COT and RWOT, where RWOT is a representative work of a sample-level OT matching mechanism. By looking at the results in Fig. 3, compared with RWOT, our COT significantly saves the computational cost with the increasing of training scale. Integrating the above theoretical analysis with this quantitative result demonstrates the superiority of our COT in economizing the computation head." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 570, + 545, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 545, + 689 + ], + "type": "text", + "content": "Effects of Loss Weight. " + }, + { + "bbox": [ + 304, + 570, + 545, + 689 + ], + "type": "inline_equation", + "content": "\\alpha_{1}" + }, + { + "bbox": [ + 304, + 570, + 545, + 689 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 570, + 545, + 689 + ], + "type": "inline_equation", + "content": "\\alpha_{2}" + }, + { + "bbox": [ + 304, + 570, + 545, + 689 + ], + "type": "text", + "content": " are employed to weight the importance of cluster and OT loss, respectively. In our experiment, " + }, + { + "bbox": [ + 304, + 570, + 545, + 689 + ], + "type": "inline_equation", + "content": "\\alpha_{1}" + }, + { + "bbox": [ + 304, + 570, + 545, + 689 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 570, + 545, + 689 + ], + "type": "inline_equation", + "content": "\\alpha_{2}" + }, + { + "bbox": [ + 304, + 570, + 545, + 689 + ], + "type": "text", + "content": " are the same and selected from a broad range of " + }, + { + "bbox": [ + 304, + 570, + 545, + 689 + ], + "type": "inline_equation", + "content": "\\{0.1, 0.2, 0.4, 1.0, 2.0\\}" + }, + { + "bbox": [ + 304, + 570, + 545, + 689 + ], + "type": "text", + "content": ". Results are evaluated on the Office-31 and shown in Table 5. When " + }, + { + "bbox": [ + 304, + 570, + 545, + 689 + ], + "type": "inline_equation", + "content": "\\alpha_{1}" + }, + { + "bbox": [ + 304, + 570, + 545, + 689 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 570, + 545, + 689 + ], + "type": "inline_equation", + "content": "\\alpha_{2}" + }, + { + "bbox": [ + 304, + 570, + 545, + 689 + ], + "type": "text", + "content": " are set to large value, the performance will tend to drop. In our opinion, the larger " + }, + { + "bbox": [ + 304, + 570, + 545, + 689 + ], + "type": "inline_equation", + "content": "\\alpha_{1}" + }, + { + "bbox": [ + 304, + 570, + 545, + 689 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 570, + 545, + 689 + ], + "type": "inline_equation", + "content": "\\alpha_{2}" + }, + { + "bbox": [ + 304, + 570, + 545, + 689 + ], + "type": "text", + "content": " would put more attention to local alignment while weakening the feature representation on both domains. As a result, " + }, + { + "bbox": [ + 304, + 570, + 545, + 689 + ], + "type": "inline_equation", + "content": "\\alpha_{1}" + }, + { + "bbox": [ + 304, + 570, + 545, + 689 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 570, + 545, + 689 + ], + "type": "inline_equation", + "content": "\\alpha_{2}" + }, + { + "bbox": [ + 304, + 570, + 545, + 689 + ], + "type": "text", + "content": " are set to 0.2, 0.2 for all datasets." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "content": "Effects of the Number of Sub-domains. " + }, + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "inline_equation", + "content": "Y" + }, + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "content": ", defined in Section 4, represents the number of sub-domains for each" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "20004" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 67, + 91, + 202, + 189 + ], + "blocks": [ + { + "bbox": [ + 67, + 91, + 202, + 189 + ], + "lines": [ + { + "bbox": [ + 67, + 91, + 202, + 189 + ], + "spans": [ + { + "bbox": [ + 67, + 91, + 202, + 189 + ], + "type": "image", + "image_path": "fb6d561851bbfd5ede1e570d6e0345edb266678544eb7e5e269c84179527bdce.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 119, + 196, + 162, + 205 + ], + "lines": [ + { + "bbox": [ + 119, + 196, + 162, + 205 + ], + "spans": [ + { + "bbox": [ + 119, + 196, + 162, + 205 + ], + "type": "text", + "content": "(a) Office-31" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 223, + 91, + 358, + 188 + ], + "blocks": [ + { + "bbox": [ + 223, + 91, + 358, + 188 + ], + "lines": [ + { + "bbox": [ + 223, + 91, + 358, + 188 + ], + "spans": [ + { + "bbox": [ + 223, + 91, + 358, + 188 + ], + "type": "image", + "image_path": "dd567420a6c5c3048c62f644a724b8751e00bdb499bb01e14a007fd68a4a1be7.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 269, + 196, + 324, + 205 + ], + "lines": [ + { + "bbox": [ + 269, + 196, + 324, + 205 + ], + "spans": [ + { + "bbox": [ + 269, + 196, + 324, + 205 + ], + "type": "text", + "content": "(b) Office-Home" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 377, + 91, + 514, + 188 + ], + "blocks": [ + { + "bbox": [ + 377, + 91, + 514, + 188 + ], + "lines": [ + { + "bbox": [ + 377, + 91, + 514, + 188 + ], + "spans": [ + { + "bbox": [ + 377, + 91, + 514, + 188 + ], + "type": "image", + "image_path": "6b349db665b116784e576912a6477a04dd180ee24790c94a95edecfcee1dc072.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 426, + 196, + 479, + 205 + ], + "lines": [ + { + "bbox": [ + 426, + 196, + 479, + 205 + ], + "spans": [ + { + "bbox": [ + 426, + 196, + 479, + 205 + ], + "type": "text", + "content": "(c) VisDA-2017" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 98, + 246, + 230, + 346 + ], + "blocks": [ + { + "bbox": [ + 46, + 215, + 545, + 237 + ], + "lines": [ + { + "bbox": [ + 46, + 215, + 545, + 237 + ], + "spans": [ + { + "bbox": [ + 46, + 215, + 545, + 237 + ], + "type": "text", + "content": "Figure 2. Class distribution on Office-31, Office-Home and VisDA-2017. Note that the visualized class distribution is randomly selected from corresponding datasets." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 98, + 246, + 230, + 346 + ], + "lines": [ + { + "bbox": [ + 98, + 246, + 230, + 346 + ], + "spans": [ + { + "bbox": [ + 98, + 246, + 230, + 346 + ], + "type": "image", + "image_path": "65737c917147038b2e2e94e1cb82ad1fe7f20e462da8a4b312a56d52b90bc40a.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 80, + 356, + 254, + 367 + ], + "lines": [ + { + "bbox": [ + 80, + 356, + 254, + 367 + ], + "spans": [ + { + "bbox": [ + 80, + 356, + 254, + 367 + ], + "type": "text", + "content": "Figure 3. Computation cost of COT and RWOT" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 52, + 411, + 282, + 487 + ], + "blocks": [ + { + "bbox": [ + 46, + 380, + 287, + 403 + ], + "lines": [ + { + "bbox": [ + 46, + 380, + 287, + 403 + ], + "spans": [ + { + "bbox": [ + 46, + 380, + 287, + 403 + ], + "type": "text", + "content": "Table 5. Mean accuracy of our COT with different " + }, + { + "bbox": [ + 46, + 380, + 287, + 403 + ], + "type": "inline_equation", + "content": "\\alpha_{1}" + }, + { + "bbox": [ + 46, + 380, + 287, + 403 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 380, + 287, + 403 + ], + "type": "inline_equation", + "content": "\\alpha_{2}" + }, + { + "bbox": [ + 46, + 380, + 287, + 403 + ], + "type": "text", + "content": " on Office-31 dataset." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 52, + 411, + 282, + 487 + ], + "lines": [ + { + "bbox": [ + 52, + 411, + 282, + 487 + ], + "spans": [ + { + "bbox": [ + 52, + 411, + 282, + 487 + ], + "type": "table", + "html": "
α1α2avg_acc
0.10.190.8
0.20.291.0
0.40.490.6
1.01.089.4
2.02.087.7
", + "image_path": "03ada669735cbf5ac53de42b60db0959d21feec3ab691d655889ea06ada2f178.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 509, + 287, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 509, + 287, + 545 + ], + "spans": [ + { + "bbox": [ + 46, + 509, + 287, + 545 + ], + "type": "text", + "content": "class. We investigate a broad range of " + }, + { + "bbox": [ + 46, + 509, + 287, + 545 + ], + "type": "inline_equation", + "content": "Y" + }, + { + "bbox": [ + 46, + 509, + 287, + 545 + ], + "type": "text", + "content": " to find an appropriate value. Based on the results shown in Table 6, we set " + }, + { + "bbox": [ + 46, + 509, + 287, + 545 + ], + "type": "inline_equation", + "content": "Y" + }, + { + "bbox": [ + 46, + 509, + 287, + 545 + ], + "type": "text", + "content": " to 2 for all test settings." + } + ] + } + ], + "index": 11 + }, + { + "type": "table", + "bbox": [ + 69, + 587, + 266, + 652 + ], + "blocks": [ + { + "bbox": [ + 46, + 556, + 286, + 578 + ], + "lines": [ + { + "bbox": [ + 46, + 556, + 286, + 578 + ], + "spans": [ + { + "bbox": [ + 46, + 556, + 286, + 578 + ], + "type": "text", + "content": "Table 6. Mean accuracy of our COT with different " + }, + { + "bbox": [ + 46, + 556, + 286, + 578 + ], + "type": "inline_equation", + "content": "Y" + }, + { + "bbox": [ + 46, + 556, + 286, + 578 + ], + "type": "text", + "content": " on Office-31 dataset." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 69, + 587, + 266, + 652 + ], + "lines": [ + { + "bbox": [ + 69, + 587, + 266, + 652 + ], + "spans": [ + { + "bbox": [ + 69, + 587, + 266, + 652 + ], + "type": "table", + "html": "
Yavg_acc
190.7
291.0
390.8
590.4
", + "image_path": "e7bb15cf2a2ec1a22a58b2aee3e5a7993a5349544abcdf059cd001ce417829d6.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_body" + } + ], + "index": 13 + }, + { + "bbox": [ + 46, + 665, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 287, + 715 + ], + "type": "text", + "content": "Qualitative Analysis on COT As discussed above, we present 2 advantages of our COT when compared with existing OT-based methods [23,51]. In this part, we further illustrate another advantage compared to the clustering-based" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 236, + 547, + 368 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 236, + 547, + 368 + ], + "spans": [ + { + "bbox": [ + 304, + 236, + 547, + 368 + ], + "type": "text", + "content": "OT algorithm in the conventional OT realm [24, 36]. Our COT is optimized in an end-to-end manner, thus the feature representation and optimal transportation map are both updated asynchronously. While for [24] and [36], they focus on obtaining offline cluster centers firstly and then compute the optimal transportation map. Such a synchronous optimization manner inevitably leads to a sub-optimal mapping result, because the computation procedure of an optimal transportation map is based on the fixed cluster centers. This is why we adopt an online clustering mode instead of an offline one." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 384, + 389, + 396 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 384, + 389, + 396 + ], + "spans": [ + { + "bbox": [ + 306, + 384, + 389, + 396 + ], + "type": "text", + "content": "5.4. Visualization" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 407, + 545, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 407, + 545, + 502 + ], + "spans": [ + { + "bbox": [ + 304, + 407, + 545, + 502 + ], + "type": "text", + "content": "We present the t-SNE visualization of feature representation for our COT and RWOT [51] on CI50. Under the challenge of class imbalance, the learned features from RWOT in some classes are ambiguous to provide accurate representation compared to our COT. These visualization results further demonstrate the excellent ability of our COT on handling class-imbalance challenges from the feature representation perspective." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 306, + 523, + 378, + 534 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 523, + 378, + 534 + ], + "spans": [ + { + "bbox": [ + 306, + 523, + 378, + 534 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 546, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 546, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 546, + 545, + 713 + ], + "type": "text", + "content": "In this paper, we propose a novel method to advance the OT-based domain adaptation community, which integrates optimal transport theory with clustering operation, termed Clustering-based Optimal Transport (COT). Concretely, COT applies the loss derived from discrete Kantorovich dual form to cluster and align centers in the source and target domain, thus transferring knowledge from the source domain to the target domain. Moreover, our COT can eliminate the negative effect brought by the class imbalance phenomenon and reduce the computation cost simultaneously, which are two challenging problems for existing OT-based UDA algorithms. Besides, we also provide comprehensive theoretical analysis and experimental results to guarantee the advantages of our COT." + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20005" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 297, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 53, + 91, + 297, + 125 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 297, + 125 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 297, + 125 + ], + "type": "text", + "content": "[1135] Hana Ajakan, Pascal Germain, Hugo Larochelle, François Laviolette, and Mario Marchand. Domain-adversarial neural networks. arXiv preprint arXiv:1412.4446, 2014. 1, 5, 6, 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 126, + 286, + 147 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 126, + 286, + 147 + ], + "spans": [ + { + "bbox": [ + 53, + 126, + 286, + 147 + ], + "type": "text", + "content": "[2] M. Arjovsky, S. Chintala, and L. Bottou. Wasserstein gan. 2017. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 149, + 286, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 149, + 286, + 182 + ], + "spans": [ + { + "bbox": [ + 53, + 149, + 286, + 182 + ], + "type": "text", + "content": "[3] Woong-Gi Chang, Tackgeun You, Seonguk Seo, Suha Kwak, and Bohyung Han. Domain-specific batch normalization for unsupervised domain adaptation. In CVPR, 2019. 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 184, + 286, + 238 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 184, + 286, + 238 + ], + "spans": [ + { + "bbox": [ + 53, + 184, + 286, + 238 + ], + "type": "text", + "content": "[4] Xinyang Chen, Sinan Wang, Mingsheng Long, and Jianmin Wang. Transferability vs. discriminability: Batch spectral penalization for adversarial domain adaptation. In International conference on machine learning, pages 1081-1090. PMLR, 2019. 6, 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 240, + 286, + 283 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 240, + 286, + 283 + ], + "spans": [ + { + "bbox": [ + 53, + 240, + 286, + 283 + ], + "type": "text", + "content": "[5] Nicolas Courty, Rémi Flamary, Amaury Habrard, and Alain Rakotomamonjy. Joint distribution optimal transportation for domain adaptation. arXiv preprint arXiv:1705.08848, 2017. 2, 5, 6, 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 286, + 286, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 286, + 286, + 330 + ], + "spans": [ + { + "bbox": [ + 53, + 286, + 286, + 330 + ], + "type": "text", + "content": "[6] Nicolas Courty, Rémi Flamary, and Devis Tuia. Domain adaptation with regularized optimal transport. In Joint European Conference on Machine Learning and Knowledge Discovery in Databases, pages 274-289. Springer, 2014. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 331, + 286, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 331, + 286, + 374 + ], + "spans": [ + { + "bbox": [ + 53, + 331, + 286, + 374 + ], + "type": "text", + "content": "[7] Nicolas Courty, Rémi Flamary, Devis Tuia, and Alain Rakotomamonjy. Optimal transport for domain adaptation. IEEE transactions on pattern analysis and machine intelligence, 39(9):1853-1865, 2016. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 376, + 286, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 376, + 286, + 441 + ], + "spans": [ + { + "bbox": [ + 53, + 376, + 286, + 441 + ], + "type": "text", + "content": "[8] Shuhao Cui, Shuhui Wang, Junbao Zhuo, Liang Li, Qingming Huang, and Qi Tian. Towards discriminability and diversity: Batch nuclear-norm maximization under label insufficient situations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3941-3950, 2020. 6, 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 445, + 286, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 445, + 286, + 498 + ], + "spans": [ + { + "bbox": [ + 53, + 445, + 286, + 498 + ], + "type": "text", + "content": "[9] Shuhao Cui, Shuhui Wang, Junbao Zhuo, Chi Su, Qingming Huang, and Tian Qi. Gradually vanishing bridge for adversarial domain adaptation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2020. 5, 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 501, + 286, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 501, + 286, + 555 + ], + "spans": [ + { + "bbox": [ + 48, + 501, + 286, + 555 + ], + "type": "text", + "content": "[10] Shuhao Cui, Shuhui Wang, Junbao Zhuo, Chi Su, Qingming Huang, and Qi Tian. Gradually vanishing bridge for adversarial domain adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12455-12464, 2020. 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 557, + 286, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 557, + 286, + 610 + ], + "spans": [ + { + "bbox": [ + 48, + 557, + 286, + 610 + ], + "type": "text", + "content": "[11] Bharath Bhushan Damodaran, Benjamin Kellenberger, Rémi Flamary, Devis Tuia, and Nicolas Courty. Deepjoint: Deep joint distribution optimal transport for unsupervised domain adaptation. In Proceedings of the European Conference on Computer Vision (ECCV), pages 447-463, 2018. 5, 6, 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 613, + 286, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 613, + 286, + 656 + ], + "spans": [ + { + "bbox": [ + 48, + 613, + 286, + 656 + ], + "type": "text", + "content": "[12] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 658, + 286, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 286, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 286, + 712 + ], + "type": "text", + "content": "[13] Jiankang Deng, Jia Guo, Niannan Xue, and Stefanos Zafeiriou. Arcface: Additive angular margin loss for deep face recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4690-4699, 2019. 1, 3" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 32, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "text", + "content": "[14] Zhekai Du, Jingjing Li, Hongzu Su, Lei Zhu, and Ke Lu. Cross-domain gradient discrepancy minimization for unsupervised domain adaptation. 2021. 7" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 108, + 545, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 108, + 545, + 150 + ], + "spans": [ + { + "bbox": [ + 307, + 108, + 545, + 150 + ], + "type": "text", + "content": "[15] Kilian Fatras, Thibault Séjourne, Rémi Flamary, and Nicolas Courty. Unbalanced minibatch optimal transport; applications to domain adaptation. In International Conference on Machine Learning, pages 3186-3197. PMLR, 2021. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 152, + 545, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 152, + 545, + 184 + ], + "spans": [ + { + "bbox": [ + 307, + 152, + 545, + 184 + ], + "type": "text", + "content": "[16] Wilfrid Gangbo and Robert J McCann. The geometry of optimal transportation. Acta Mathematica, 177(2):113-161, 1996. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 186, + 545, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 186, + 545, + 229 + ], + "spans": [ + { + "bbox": [ + 307, + 186, + 545, + 229 + ], + "type": "text", + "content": "[17] Yixiao Ge, Dapeng Chen, and Hongsheng Li. Mutual meanteaching: Pseudo label refinery for unsupervised domain adaptation on person re-identification. In International Conference on Learning Representations, 2020. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 231, + 545, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 231, + 545, + 262 + ], + "spans": [ + { + "bbox": [ + 307, + 231, + 545, + 262 + ], + "type": "text", + "content": "[18] Yves Grandvalet, Yoshua Bengio, et al. Semi-supervised learning by entropy minimization. In CAP, pages 281-296, 2005. 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 264, + 545, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 264, + 545, + 297 + ], + "spans": [ + { + "bbox": [ + 307, + 264, + 545, + 297 + ], + "type": "text", + "content": "[19] Sahand Hajifar and Hongyue Sun. Online domain adaptation for continuous cross-subject liver viability evaluation based on irregular thermal data, 2020. 5, 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 298, + 545, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 298, + 545, + 319 + ], + "spans": [ + { + "bbox": [ + 307, + 298, + 545, + 319 + ], + "type": "text", + "content": "[20] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. 2016. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 322, + 545, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 322, + 545, + 353 + ], + "spans": [ + { + "bbox": [ + 307, + 322, + 545, + 353 + ], + "type": "text", + "content": "[21] Ying Jin, Ximei Wang, Mingsheng Long, and Jianmin Wang. Minimum class confusion for versatile domain adaptation, 2020. 5, 7" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 355, + 545, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 355, + 545, + 376 + ], + "spans": [ + { + "bbox": [ + 307, + 355, + 545, + 376 + ], + "type": "text", + "content": "[22] Leonid V Kantorovich. On the translocation of masses. Journal of mathematical sciences, 133(4):1381-1382, 2006. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 378, + 545, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 378, + 545, + 409 + ], + "spans": [ + { + "bbox": [ + 307, + 378, + 545, + 409 + ], + "type": "text", + "content": "[23] Tanguy Kerdoncuff, Rémi Emonet, and Marc Sebban. Metric learning in optimal transport for domain adaptation. 2020. 1, 2, 5, 6, 7, 8" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 412, + 545, + 454 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 412, + 545, + 454 + ], + "spans": [ + { + "bbox": [ + 307, + 412, + 545, + 454 + ], + "type": "text", + "content": "[24] John Lee, Max Dabagia, Eva Dyer, and Christopher Rozell. Hierarchical optimal transport for multimodal distribution alignment. Advances in Neural Information Processing Systems, 32, 2019. 8" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 456, + 545, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 456, + 545, + 499 + ], + "spans": [ + { + "bbox": [ + 307, + 456, + 545, + 499 + ], + "type": "text", + "content": "[25] Jingjing Li, Erpeng Chen, Zhengming Ding, Lei Zhu, Ke Lu, and Heng Tao Shen. Maximum density divergence for domain adaptation. IEEE Transactions on Pattern Analysis and Machine Intelligence, page 1-1, 2020. 5, 6" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 502, + 545, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 502, + 545, + 555 + ], + "spans": [ + { + "bbox": [ + 307, + 502, + 545, + 555 + ], + "type": "text", + "content": "[26] Mengxue Li, Yi-Ming Zhai, You-Wei Luo, Peng-Fei Ge, and Chuan-Xian Ren. Enhanced transport distance for unsupervised domain adaptation. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 13933–13941, 2020. 6" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 557, + 545, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 557, + 545, + 610 + ], + "spans": [ + { + "bbox": [ + 307, + 557, + 545, + 610 + ], + "type": "text", + "content": "[27] Rui Li, Qianfen Jiao, Wenming Cao, Hau-San Wong, and Si Wu. Model adaptation: Unsupervised domain adaptation without source data. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9641-9650, 2020. 7" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 613, + 545, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 613, + 545, + 656 + ], + "spans": [ + { + "bbox": [ + 307, + 613, + 545, + 656 + ], + "type": "text", + "content": "[28] Shuang Li, Mixue Xie, Kaixiong Gong, Chi Harold Liu, Yulin Wang, and Wei Li. Transferable semantic augmentation for domain adaptation. arXiv preprint arXiv:2103.12562, 2021. 1, 5, 6, 7" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 307, + 658, + 545, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 658, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 307, + 658, + 545, + 689 + ], + "type": "text", + "content": "[29] Jian Liang, Dapeng Hu, and Jiashi Feng. Do we really need to access the source data? source hypothesis transfer for unsupervised domain adaptation. 2020. 5, 7" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 307, + 692, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 692, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 692, + 545, + 713 + ], + "type": "text", + "content": "[30] Chi-Heng Lin, Mehdi Azabou, and Eva L Dyer. Making transport more robust and interpretable by moving data" + } + ] + } + ], + "index": 31 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "20006" + } + ] + } + ], + "index": 33 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 66, + 72, + 288, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 72, + 288, + 95 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 288, + 95 + ], + "type": "text", + "content": "through a small number of anchor points. Proceedings of machine learning research, 139:6631, 2021. 1" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 95, + 287, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 95, + 287, + 149 + ], + "spans": [ + { + "bbox": [ + 48, + 95, + 287, + 149 + ], + "type": "text", + "content": "[31] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dálár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In European conference on computer vision, pages 740–755. Springer, 2014. 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 150, + 287, + 204 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 150, + 287, + 204 + ], + "spans": [ + { + "bbox": [ + 48, + 150, + 287, + 204 + ], + "type": "text", + "content": "[32] Hong Liu, Mingsheng Long, Jianmin Wang, and Michael Jordan. Transferable adversarial training: A general approach to adapting deep classifiers. In International Conference on Machine Learning, pages 4013-4022. PMLR, 2019. 5, 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 205, + 287, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 205, + 287, + 239 + ], + "spans": [ + { + "bbox": [ + 48, + 205, + 287, + 239 + ], + "type": "text", + "content": "[33] Mingsheng Long, Zhangjie Cao, Jianmin Wang, and Michael I Jordan. Conditional adversarial domain adaptation. arXiv preprint arXiv:1705.10667, 2017. 5, 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 239, + 287, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 239, + 287, + 282 + ], + "spans": [ + { + "bbox": [ + 48, + 239, + 287, + 282 + ], + "type": "text", + "content": "[34] Mingsheng Long, Han Zhu, Jianmin Wang, and Michael I Jordan. Deep transfer learning with joint adaptation networks. In International conference on machine learning, pages 2208-2217. PMLR, 2017. 5, 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 284, + 287, + 337 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 284, + 287, + 337 + ], + "spans": [ + { + "bbox": [ + 48, + 284, + 287, + 337 + ], + "type": "text", + "content": "[35] Zhihe Lu, Yongxin Yang, Xiatian Zhu, Cong Liu, Yi-Zhe Song, and Tao Xiang. Stochastic classifiers for unsupervised domain adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9111-9120, 2020. 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 338, + 287, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 338, + 287, + 381 + ], + "spans": [ + { + "bbox": [ + 48, + 338, + 287, + 381 + ], + "type": "text", + "content": "[36] Liang Mi, Wen Zhang, and Yalin Wang. Regularized wasserstein means for aligning distributional data. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 34, pages 5166-5173, 2020. 8" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 383, + 287, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 383, + 287, + 415 + ], + "spans": [ + { + "bbox": [ + 48, + 383, + 287, + 415 + ], + "type": "text", + "content": "[37] Gaspard Monge. Mémoire sur la théorie des déblais et des remblais. Histoire de l'Académie Royale des Sciences de Paris, 1781. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 415, + 287, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 415, + 287, + 459 + ], + "spans": [ + { + "bbox": [ + 48, + 415, + 287, + 459 + ], + "type": "text", + "content": "[38] Khai Nguyen, Dang Nguyen, Quoc Nguyen, Tung Pham, Hung Bui, Dinh Phung, Trung Le, and Nhat Ho. On transportation of mini-batches: A hierarchical approach. arXiv preprint arXiv:2102.05912, 2021. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 460, + 287, + 492 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 460, + 287, + 492 + ], + "spans": [ + { + "bbox": [ + 48, + 460, + 287, + 492 + ], + "type": "text", + "content": "[39] Yingwei Pan, Ting Yao, Yehao Li, Yu Wang, Chong-Wah Ngo, and Tao Mei. Transferrable prototypical networks for unsupervised domain adaptation, 2019. 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 493, + 287, + 546 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 493, + 287, + 546 + ], + "spans": [ + { + "bbox": [ + 48, + 493, + 287, + 546 + ], + "type": "text", + "content": "[40] Ievgen Redko, Nicolas Courty, Rémi Flamary, and Devis Tuia. Optimal transport for multi-source domain adaptation under target shift. In The 22nd International Conference on Artificial Intelligence and Statistics, pages 849-858. PMLR, 2019. 1, 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 548, + 287, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 548, + 287, + 601 + ], + "spans": [ + { + "bbox": [ + 48, + 548, + 287, + 601 + ], + "type": "text", + "content": "[41] Hochang Rhee and Nam Ik Cho. Efficient and robust pseudolabeling for unsupervised domain adaptation. In 2019 Asia-Pacific Signal and Information Processing Association Annual Summit and Conference (APSIPA ASC), pages 980–985, 2019. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 603, + 287, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 603, + 287, + 657 + ], + "spans": [ + { + "bbox": [ + 48, + 603, + 287, + 657 + ], + "type": "text", + "content": "[42] Kuniaki Saito, Kohei Watanabe, Yoshitaka Ushiku, and Tatsuya Harada. Maximum classifier discrepancy for unsupervised domain adaptation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3723-3732, 2018. 1, 5, 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 658, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 287, + 712 + ], + "type": "text", + "content": "[43] Swami Sankaranarayanan, Yogesh Balaji, Carlos D Castillo, and Rama Chellappa. Generate to adapt: Aligning domains using generative adversarial networks. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 8503-8512, 2018. 5" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 689 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 95 + ], + "type": "text", + "content": "[44] Filippo Santambrogio. Optimal transport for applied mathematicians. Birkhäuser, NY, 55(58-63):94, 2015. 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 95, + 545, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 95, + 545, + 140 + ], + "spans": [ + { + "bbox": [ + 307, + 95, + 545, + 140 + ], + "type": "text", + "content": "[45] Hui Tang, Ke Chen, and Kui Jia. Unsupervised domain adaptation via structurally regularized deep clustering. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8725-8735, 2020. 5, 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 141, + 545, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 141, + 545, + 185 + ], + "spans": [ + { + "bbox": [ + 307, + 141, + 545, + 185 + ], + "type": "text", + "content": "[46] Eric Tzeng, Judy Hoffman, Kate Saenko, and Trevor Darrell. Adversarial discriminative domain adaptation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7167-7176, 2017. 5" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 186, + 545, + 240 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 186, + 545, + 240 + ], + "spans": [ + { + "bbox": [ + 308, + 186, + 545, + 240 + ], + "type": "text", + "content": "[47] Hao Wang, Yitong Wang, Zheng Zhou, Xing Ji, Dihong Gong, Jingchao Zhou, Zhifeng Li, and Wei Liu. Cosface: Large margin cosine loss for deep face recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5265-5274, 2018. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 241, + 545, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 241, + 545, + 285 + ], + "spans": [ + { + "bbox": [ + 308, + 241, + 545, + 285 + ], + "type": "text", + "content": "[48] Qian Wang and Toby Breckon. Unsupervised domain adaptation via structured prediction based selective pseudolabeling. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 34, pages 6243-6250, 2020. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 286, + 545, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 286, + 545, + 319 + ], + "spans": [ + { + "bbox": [ + 308, + 286, + 545, + 319 + ], + "type": "text", + "content": "[49] Wei Wang, Haojie Li, Zhengming Ding, and Zhihui Wang. Rethink maximum mean discrepancy for domain adaptation. arXiv preprint arXiv:2007.00689, 2020. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 320, + 545, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 320, + 545, + 352 + ], + "spans": [ + { + "bbox": [ + 308, + 320, + 545, + 352 + ], + "type": "text", + "content": "[50] Pengfei Wei, Yiping Ke, Xinghua Qu, and Tze-Yun Leong. Subdomain adaptation with manifolds discrepancy alignment, 2020. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 354, + 545, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 354, + 545, + 408 + ], + "spans": [ + { + "bbox": [ + 308, + 354, + 545, + 408 + ], + "type": "text", + "content": "[51] Renjun Xu, Pelen Liu, Liyan Wang, Chao Chen, and Jindong Wang. Reliable weighted optimal transport for unsupervised domain adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4394-4403, 2020. 1, 5, 7, 8" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 410, + 545, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 410, + 545, + 453 + ], + "spans": [ + { + "bbox": [ + 308, + 410, + 545, + 453 + ], + "type": "text", + "content": "[52] Tongkun Xu, Weihua Chen, Pichao Wang, Fan Wang, Hao Li, and Rong Jin. Cdtrans: Cross-domain transformer for unsupervised domain adaptation. arXiv preprint arXiv:2109.06165, 2021. 6, 7" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 455, + 545, + 519 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 455, + 545, + 519 + ], + "spans": [ + { + "bbox": [ + 308, + 455, + 545, + 519 + ], + "type": "text", + "content": "[53] Hongliang Yan, Yukang Ding, Peihua Li, Qilong Wang, Yong Xu, and Wangmeng Zuo. Mind the class weight bias: Weighted maximum mean discrepancy for unsupervised domain adaptation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 2272-2281, 2017. 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 521, + 545, + 565 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 521, + 545, + 565 + ], + "spans": [ + { + "bbox": [ + 308, + 521, + 545, + 565 + ], + "type": "text", + "content": "[54] Yuguang Yan, Wen Li, Hanrui Wu, Huaqing Min, Mingkui Tan, and Qingyao Wu. Semi-supervised optimal transport for heterogeneous domain adaptation. In IJCAI, volume 7, pages 2969-2975, 2018. 1" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 566, + 545, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 566, + 545, + 599 + ], + "spans": [ + { + "bbox": [ + 308, + 566, + 545, + 599 + ], + "type": "text", + "content": "[55] Jinyu Yang, Jingjing Liu, Ning Xu, and Junzhou Huang. Tvt: Transferable vision transformer for unsupervised domain adaptation. arXiv preprint arXiv:2108.05988, 2021. 7" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 600, + 545, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 600, + 545, + 632 + ], + "spans": [ + { + "bbox": [ + 308, + 600, + 545, + 632 + ], + "type": "text", + "content": "[56] Yabin Zhang, Hui Tang, Kui Jia, and Mingkui Tan. Domain-symmetric networks for adversarial domain adaptation, 2019. 6" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 634, + 545, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 634, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 308, + 634, + 545, + 689 + ], + "type": "text", + "content": "[57] Yongchun Zhu, Fuzhen Zhuang, Jindong Wang, Guolin Ke, Jingwu Chen, Jiang Bian, Hui Xiong, and Qing He. Deep subdomain adaptation network for image classification. IEEE Transactions on Neural Networks and Learning Systems, 32(4):1713-1722, Apr 2021. 1" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20007" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/CP3_ Channel Pruning Plug-In for Point-Based Networks/f7e54c81-931a-42ba-b8dd-1b55f7342f80_content_list.json b/2023/CP3_ Channel Pruning Plug-In for Point-Based Networks/f7e54c81-931a-42ba-b8dd-1b55f7342f80_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..f3389209c9de68e5ff767df5c8d0f1b19fd20bf4 --- /dev/null +++ b/2023/CP3_ Channel Pruning Plug-In for Point-Based Networks/f7e54c81-931a-42ba-b8dd-1b55f7342f80_content_list.json @@ -0,0 +1,1432 @@ +[ + { + "type": "text", + "text": "$\\mathbf{CP}^{3}$ : Channel Pruning Plug-in for Point-based Networks", + "text_level": 1, + "bbox": [ + 197, + 128, + 772, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$\\mathrm{Yaxin Peng}^{3}$ Ning Liu $^{2,*}$ Zhiying Che $^{2}$ Zhiyuan Xu $^{2}$ Chaomin Shen $^{1}$ Guixu Zhang $^{1, \\boxtimes}$ Xinmei Liu $^{1}$ Feifei Feng $^{2}$ Jian Tang $^{2, \\boxtimes}$", + "bbox": [ + 102, + 176, + 867, + 226 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ School of Computer Science, East China Normal University $^{2}$ Midea Group", + "bbox": [ + 243, + 231, + 727, + 268 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{3}$ Department of Mathematics, School of Science, Shanghai University", + "bbox": [ + 207, + 268, + 761, + 287 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1{51205901049,51205901078}@stu.ecnu.edu.cn 1{cmshen,gxzhang}@cs.ecnu.edu.cn", + "bbox": [ + 197, + 289, + 772, + 305 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "2{liuning22, chezp, xuzy70, feifei.feng, tangjian22}@midea.com yaxin.peng@shu.edu.cn", + "bbox": [ + 171, + 306, + 792, + 321 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 373, + 312, + 388 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Channel pruning can effectively reduce both computational cost and memory footprint of the original network while keeping a comparable accuracy performance. Though great success has been achieved in channel pruning for 2D image-based convolutional networks (CNNs), existing works seldom extend the channel pruning methods to 3D point-based neural networks (PNNs). Directly implementing the 2D CNN channel pruning methods to PNNs undermine the performance of PNNs because of the different representations of 2D images and 3D point clouds as well as the network architecture disparity. In this paper, we proposed $CP^3$ , which is a Channel Pruning Plugin for Point-based network. $CP^3$ is elaborately designed to leverage the characteristics of point clouds and PNNs in order to enable 2D channel pruning methods for PNNs. Specifically, it presents a coordinate-enhanced channel importance metric to reflect the correlation between dimensional information and individual channel features, and it recycles the discarded points in PNN's sampling process and reconsiders their potentially-exclusive information to enhance the robustness of channel pruning. Experiments on various PNN architectures show that $CP^3$ constantly improves state-of-the-art 2D CNN pruning approaches on different point cloud tasks. For instance, our compressed PointNeXt-S on ScanObjectNN achieves an accuracy of $88.52\\%$ with a pruning rate of $57.8\\%$ , outperforming the baseline pruning methods with an accuracy gain of $1.94\\%$ .", + "bbox": [ + 75, + 407, + 473, + 816 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 500, + 357, + 630, + 372 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Convolutional Neural Networks (CNNs) often encounter the problems of overloaded computation and overweighted storage. The cumbersome instantiation of a CNN model leads to inefficient, uneconomic, or even impossible deployment in practice. Therefore, light-weight models that provide comparable results with much fewer computational costs are in great demand for nearly all applications. Channel pruning is a promising solution to delivering efficient networks. In recent years, 2D CNN channel pruning, e.g., pruning classical VGGNets [37], ResNets [14], MobileNets [16], and many other neural networks for processing 2D images [6, 7, 12, 24, 26, 29, 40], has been successfully conducted. Most channel pruning approaches focus on identifying redundant convolution filters (i.e., channels) by evaluating their importance. The cornerstone of 2D channel pruning methods is the diversified yet effective channel evaluation metrics. For instance, HRank [24] uses the rank of the feature map as the pruning metric and removes the low-rank filters that are considered to contain less information. CHIP [40] leverages channel independence to represent the importance of each feature mapping and eliminates less important channels.", + "bbox": [ + 496, + 385, + 893, + 717 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "With the widespread application of depth-sensing technology, 3D vision tasks [9, 10, 36, 44] are a rapidly growing field starving for powerful methods. Apart from straightforwardly applying 2D CNNs, models built with Point-based Neural Networks (PNNs), which directly process point clouds from the beginning without unnecessary rendering, show their merits and are widely deployed on edge devices for various applications such as robots [22, 49] and self-driving [2, 53]. Compressing PNNs is crucial due to the limited resources of edge devices and multiple models for different tasks are likely to run simultaneously [8, 30]. Given the huge success of 2D channel pruning and the great", + "bbox": [ + 496, + 719, + 895, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Equal contributions.", + "bbox": [ + 107, + 851, + 225, + 863 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "Corresponding authors.", + "bbox": [ + 107, + 864, + 245, + 875 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "This work is done during Yaomin Huang and Xinmei Liu's internship at Midea Group.", + "bbox": [ + 78, + 876, + 468, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "5302", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "demand for efficient 3D PNNs, we intuitively raise one question: shall we directly implement the existing pruning methods to PNNs following the proposed channel importance metrics in 2D CNNs pruning?", + "bbox": [ + 75, + 90, + 468, + 151 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "With this question in mind, we investigate the fundamental factors that potentially impair 2D pruning effectiveness on PNNs. Previous works [19, 48] have shown that point clouds record visual and semantic information in a significantly different way from 2D images. Specifically, a point cloud consists of a set of unordered points on objects' and environments' surfaces, and each point encodes its features, such as intensity along with the spatial coordinates $(x,y,z)$ . In contrast, 2D images organize visual features in a dense and regular pixel array. Such data representation differences between 3D point clouds and 2D images lead to a) different ways of exploiting information from data and b) contrasting network architectures of PNNs and 2D CNNs. It is credible that only the pruning methods considering the two aspects (definitely not existing 2D CNN pruners) may obtain superior performance on PNNs.", + "bbox": [ + 75, + 154, + 470, + 396 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "From the perspective of data representations, 3D point clouds provide more 3D feature representations than 2D images, but the representations are more sensitive to network channels. To be more specific, for 2D images, all three RGB channels represent basic information in an isotropic and homogeneous way so that the latent representations extracted by CNNs applied to the images. On the other hand, point clouds explicitly encode the spatial information in three coordinate channels, which are indispensable for extracting visual and semantic information from other channels. Moreover, PNNs employ the coordinate information in multiple layers as concatenated inputs for deeper feature extraction. Nevertheless, existing CNN pruning methods are designed only suitable for the plain arrangements of 2D data but fail to consider how the informative 3D information should be extracted from point clouds.", + "bbox": [ + 75, + 398, + 470, + 641 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Moreover, the network architectures of PNNs are designed substantially different from 2D CNNs. While using smaller kernels [37] is shown to benefit 2D CNNs [37], it does not apply to networks for 3D point clouds. On the contrary, PNNs leverage neighborhoods at multiple scales to obtain both robust and detailed features. The reason is that small neighborhoods (analogous to small kernels in 2D CNNs) in point clouds consist of few points for PNNs to capture robust features. Due to the necessary sampling steps, the knowledge insufficiency issue becomes more severe for deeper PNN layers. In addition, PNNs use the random input dropout procedure during training to adaptively weight patterns detected at different scales and combine multi-scale features. This procedure randomly discards a large proportion of points and loses much exclusive information of the original data. Thus, the architecture disparity between 2D CNNs and PNNs affects the performance of", + "bbox": [ + 75, + 643, + 470, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "directly applying existing pruning methods to PNNs.", + "bbox": [ + 498, + 90, + 846, + 106 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, by explicitly dealing with the two characteristics of 3D task, namely the data representation and the PNN architecture design, we propose a Channel Pruning Plug-in for Point-based network named $\\mathrm{CP}^3$ , which can be applied to most 2D channel pruning methods for compressing PNN models. The proposed $\\mathrm{CP}^3$ refines the channel importance, the key factor of pruning methods, from two aspects. Firstly, considering the point coordinates $(x, y, \\text{and } z)$ encode the spatial information and deeply affects feature extraction procedures in PNN layers, we determine the channel importance by evaluating the correlation between the feature map and its corresponding point coordinates by introducing a coordinate-enhancement module. Secondly, calculating channel importance in channel pruning is data-driven and sensitive to the input, and the intrinsic sampling steps in PNN naturally makes pruning methods unstable. To settle this problem, we make full use of the discarded points in the sampling process via a knowledge recycling module to supplement the evaluation of channel importance. This reduces the data sampling bias impact on the channel importance calculation and increases the robustness of the pruning results. Notably, both the coordinates and recycled points in $\\mathrm{CP}^3$ do not participate in network training (with back-propagation) but only assist channel importance calculation in the reasoning phase. Thus, $\\mathrm{CP}^3$ does not increase any computational cost of the pruned network. The contributions of this paper are as follows:", + "bbox": [ + 496, + 107, + 890, + 513 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We systematically consider the characteristics of PNNs and propose a channel pruning plug-in named $\\mathrm{CP}^3$ to enhance 2D CNN channel pruning approaches on 3D PNNs. To the best of our knowledge, $\\mathrm{CP}^3$ is the first method to export existing 2D pruning methods to PNNs.", + "- We propose a coordinate-enhanced channel importance score to guide point clouds network pruning, by evaluating the correlation between feature maps and corresponding point coordinates.", + "- We design a knowledge recycling pruning scheme that increases the robustness of the pruning procedure, using the discarded points to improve the channel importance evaluation.", + "- We show that using $\\mathrm{CP}^3$ is consistently superior to directly transplanting 2D pruning methods to PNNs by extensive experiments on three 3D tasks and five datasets with different PNN models and pruning baselines." + ], + "bbox": [ + 501, + 515, + 890, + 779 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 792, + 640, + 806 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. 2D Channel Pruning", + "text_level": 1, + "bbox": [ + 500, + 816, + 694, + 833 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Channel pruning (a.k.a., filter pruning) methods reduce the redundant filters while maintaining the original structure of CNNs and is friendly to prevailing inference acceleration engines such as TensorFlow-Lite (TFLite) [11] and", + "bbox": [ + 498, + 839, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "5303", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Mobile Neural Network (MNN) [18]. Mainstream channel pruning methods [6, 7, 12, 29] usually first evaluate the importance of channels by certain metrics and then prune (i.e., remove) the less important channels. Early work [21] uses the $l_{1}$ norm of filters as importance score for channel pruning. Afterwards, learning parameters, such as the scaling factor $\\gamma$ in the batch norm layer [26] and the reconstruction error in the final network layer [51], are considered as the importance scores for channel selection. The importance sampling distribution of channels [23] is also used for pruning. Recent works [15, 40] measure the correlation of multiple feature maps to determine the importance score of the filter for pruning. HRank [24] proposes a method for pruning filters based on the theory that low-rank feature maps contain less information. [50] leverages the statistical distribution of activation gradient and takes the smaller gradient as low importance score for pruning. [46] calculates the average importance of both the input feature maps and their corresponding output feature maps to determine the overall importance. [13, 45] compress CNNs from multiple dimensions While most channel pruning methods are designed for and tested on 2D CNNs, our $\\mathrm{CP}^3$ can work in tandem with existing pruners for 3D point-based networks.", + "bbox": [ + 76, + 90, + 472, + 439 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. Point-based Networks for Point Cloud Data", + "text_level": 1, + "bbox": [ + 76, + 452, + 447, + 468 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Point-based Neural Networks (PNNs) directly process point cloud data with a flexible range of receptive field, have no positioning information loss, and thus keep more accurate spatial information. As a pioneer work, PointNet [32] learns the spatial encoding directly from the input point clouds and uses the characteristics of all points to obtain the global representations. PointNet++ [33] further proposes a multi-level feature extraction structure to extract local and global features more effectively. KPCov [42] proposes a new point convolution operation to learn local movements applied to kernel points. ASSANet [34] proposes a separable set abstraction module that decomposes the normal SA module in PointNet++ into two separate learning phases for channel and space. PointMLP [28] uses residual point blocks to extract local features, transforms local points using geometric affine modules, and extracts geometric features before and after the aggregation operation. PointNeXt [35] uses inverted residual bottleneck and separable multilayer perceptrons to achieve more efficient model scaling. Besides classification, PNNs also serve as backbones for other 3D tasks. VoteNet [31] effectively improves the 3D object detection accuracy through the Hough voting mechanism [20]. PointTransformer [52] designs models improving prior work across domains and tasks. GroupFree3D [27] uses the attention mechanism to automatically learn the contribution of each point to the object. In this paper, we show that $\\mathrm{CP}^3$ can be widely applied to point-based networks on a variety of point cloud bench", + "bbox": [ + 76, + 478, + 472, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "marks and representative original networks.", + "bbox": [ + 500, + 90, + 789, + 106 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Methodology", + "text_level": 1, + "bbox": [ + 500, + 119, + 635, + 137 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Although point-based networks are similar to CNN in concrete realization, they have fundamental differences in data representation and network architecture design. To extend the success of CNN pruning to PNN, two modules are proposed in $\\mathrm{CP}^3$ taking advantage from the dimensional information and discarded points: 1) coordinate-enhancement (CE) module, which produces a coordinate-enhanced score to estimate the channel importance by combining dimensional and feature information, and 2) knowledge recycling module reusing the discarded points to improve the channel importance evaluation criteria and increase the robustness.", + "bbox": [ + 496, + 145, + 893, + 327 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Formulations and Motivation", + "text_level": 1, + "bbox": [ + 500, + 337, + 763, + 352 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Point-based networks PNN is a unified architecture that directly takes point clouds as input. It builds hierarchical groups of points and progressively abstracts larger local regions along the hierarchy. PNN is structurally composed by a number of set abstraction (SA) blocks. Each SA block consists of 1) a sampling layer iteratively samples the farthest point to choose a subset of points from input points, 2) a group layer gathers neighbors of centroid points to a local region, 3) a set of shared Multi-Layer Perceptrons (MLPs) to extract features, and 4) a reduction layer to aggregate features in the neighbors. Formally speaking, a SA block takes an $n_{i-1} \\times (d + c_{i-1})$ matrix as input that is from $n_{i-1}$ points with $d$ -dim coordinates and $c_{i-1}$ -dim point feature. It outputs an $n_i \\times (d + c_i)$ matrix of $n_i$ subsampled points with $d$ -dimensional coordinates (i.e., $d = 3$ ) and new $c_i$ -dimensional feature vectors summarizing local context. The SA block is formulated as:", + "bbox": [ + 496, + 359, + 893, + 618 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {F} _ {i} ^ {l + 1} = \\mathcal {R} \\left\\{h _ {\\Theta} \\left(\\left[ \\mathbf {F} _ {j} ^ {l}; \\mathbf {x} _ {j} ^ {l} - \\mathbf {x} _ {i} ^ {l} \\right]\\right) \\right\\}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 583, + 625, + 890, + 652 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $h_\\Theta$ is MLPs to extract grouped points feature, $\\mathcal{R}$ is the reduction layer (e.g. max-pooling) to aggregate features in the neighbors $\\{j : (i,j) \\in \\mathcal{N}\\}$ , $F_j^l$ is the features of neighbor $j$ in the $l$ -th layer, $x_i^l$ and $x_j^l$ are input points coordinates and coordinates of neighbor $j$ in the $l$ -th layer.", + "bbox": [ + 496, + 661, + 893, + 742 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Channel pruning Assume a pre-trained PNN model has a set of $K$ convolutional layers, and $\\mathcal{A}^l$ is the $l$ -th convolution layer. The parameters in $\\mathcal{A}^l$ can be represented as a set of filters $\\mathcal{W}_{\\mathcal{A}^l} = \\{\\mathbf{w}_1^l,\\mathbf{w}_2^l,\\dots ,\\mathbf{w}_{c_l}^l\\} \\in \\mathbb{R}^{(d + c_l)\\times (d + c_{l - 1})\\times k_l\\times k_l}$ where $j$ -th filter is $\\mathbf{w}_j^l\\in \\mathbb{R}^{(d + c_{l - 1})\\times k_l\\times k_j}$ . $(d + c_{l})$ represents the number of filters in $\\mathcal{A}^l$ and $k_{l}$ denotes the kernel size. The outputs of filter, i.e., feature map, are denoted as $\\mathcal{F}^l = \\{\\mathbf{f}_1^l,\\mathbf{f}_2^l,\\ldots ,\\mathbf{f}_{n_i}^l\\} \\in \\mathbb{R}^{n_i\\times (d + c_i)}$ . Channel pruning aims to identify and remove the less importance filter from the", + "bbox": [ + 496, + 758, + 893, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "5304", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/625020ddf58bb812d6da62b4428f393d1ffe2051d1ea2a09bcce9be253eb92d9.jpg", + "image_caption": [ + "Figure 1. The framework of $\\mathrm{CP}^3$ . The figure shows the specific pruning process of one of the SA blocks. Whether a channel in a PNN is pruned is determined by three parts: 1) Original channel importance: obtained from the original CNNs channel pruning method (e.g., HRank [24], CHIP [40]). 2) Discarded channel importance: obtained from the Knowledge-Recycling module by leveraging the discarded points in the network to supplement the channel importance evaluation of the corresponding points and improve the robustness of the channel selection. 3) CE (Coordinate-Enhanced) channel importance: obtained from calculating the correlation between the feature map and its corresponding points coordinates to guide point clouds network pruning." + ], + "image_footnote": [], + "bbox": [ + 81, + 94, + 883, + 280 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "original networks. In general, channel pruning can be formulated as the following optimization problem:", + "bbox": [ + 75, + 388, + 470, + 420 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {\\delta_ {i j}} \\sum_ {i = 1} ^ {K} \\sum_ {j = 1} ^ {n _ {i}} \\delta_ {i j} \\mathcal {L} \\left(\\mathbf {w} _ {j} ^ {i}\\right), \\text {s . t .} \\sum_ {j = 1} ^ {n _ {i}} \\delta_ {i j} = k _ {l}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 145, + 425, + 468, + 465 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\delta_{ij}$ is an indicator which is 1 if $\\mathbf{w}_j^i$ is to be pruned or 0 if $\\mathbf{w}_j^i$ is to be kept, $\\mathcal{L}(\\cdot)$ measures the importance of a filter and $k_l$ is the kept filter number.", + "bbox": [ + 76, + 472, + 470, + 521 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Robust importance metric for channel pruning The metrics for evaluating the importance of filters is critical. Existing CNN pruning methods design a variety of $\\mathcal{L}(\\cdot)$ on the filters. Consider the feature maps, contain rich and important information of both filter and input data, approaches using feature information have become popular and achieved state-of-the-art performance for channel pruning. The results of the feature maps may vary depending on the variability of the input data. Therefore, when the importance of one filter solely depends on the information represented by its own generated feature map, the measurement of the importance may be unstable and sensitive to the slight change of input data. So we have taken into account the characteristics of point clouds data and point-based networks architecture to improve the robustness of channel importance in point-based networks. On the one hand, we propose a coordinate-enhancement module by evaluating the correlation between the feature map and its corresponding points coordinates to guide point clouds network pruning, which will be described in Sec. 3.2. On the other hand, we design a knowledge recycling pruning schema, using discarded points to improve the channel importance evaluation criteria and increase the robustness of the pruning module, which will be described in detail in Sec. 3.3.", + "bbox": [ + 75, + 537, + 472, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Coordinate-Enhanced Channel Importance", + "text_level": 1, + "bbox": [ + 500, + 387, + 870, + 404 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Dimensional information is critical in PNNs. The dimensional information (i.e., coordinates of the points) are usually adopted as input for feature extraction. Namely, the input and output of each SA block are concatenated with the coordinates of the points. Meanwhile, the intermediate feature maps reflect not only the information of the original input data but also the corresponding channel information. Therefore, the importance of the channel can be obtained from the feature maps, i.e., the importance of the corresponding channel. The dimensional information is crucial in point-based tasks and should be considered as part of importance metric. Thus the critical problem falls in designing a function that can well reflect the dimensional information richness of feature maps. The feature map, obtained by encoding points spatial $x$ , $y$ , and $z$ coordinates, should be closely related to the original corresponding points coordinates. Therefore, we use the correlation between the current feature map and the corresponding input points coordinates to determine the importance of the filter. The designed Coordinate-Enhancement (CE) module based on Eq. (2):", + "bbox": [ + 496, + 410, + 890, + 714 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {\\delta_ {i j}} \\sum_ {i = 1} ^ {K} \\sum_ {j = 1} ^ {n _ {i}} \\delta_ {i j} \\mathcal {L} _ {c} \\left(\\mathbf {F} _ {j} ^ {i}\\right), \\text {s . t .} \\sum_ {j = 1} ^ {n _ {i}} \\delta_ {i j} = k _ {l}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 566, + 719, + 890, + 758 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\delta_{ij}$ is an indicator which is 1 if $\\mathbf{w}_j^i$ is to be pruned or 0 if $\\mathbf{w}_j^i$ is to be kept and $k_l$ is the kept channel number. $\\mathcal{L}_c(\\cdot)$ measures the importance of a channel from take account of the relationship between feature map and points coordinates which can be formulated as:", + "bbox": [ + 496, + 763, + 893, + 842 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {c} \\left(F _ {j} ^ {i}\\right) = \\mathcal {M} \\left\\{\\mathrm {C E} \\left(F _ {j} ^ {i}, x ^ {i}\\right) \\right\\}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 602, + 845, + 890, + 867 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where CE obtains the coordinate-enhanced score by calculating correlation of each channel in the feature map with", + "bbox": [ + 496, + 869, + 893, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "5305", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/d6e02e42d971152899bb8eb6c23bfaed2cbd9531ba79ab85d19eea7ba094a7c6.jpg", + "image_caption": [ + "Coordinate-Enhancing for Filter Importance", + "Figure 2. Comparisons on the pruned model accuracy with different pruning metrics with CE scores. Results are on ModelNet40 with PointNeXt-S (C=64)." + ], + "image_footnote": [], + "bbox": [ + 114, + 101, + 426, + 218 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "the original coordinates, and $\\mathcal{M}$ takes the maximum value. Hence, higher coordinate-enhanced score (i.e., $\\mathcal{L}_c$ ) serve as a reliable measurement for information richness.", + "bbox": [ + 75, + 285, + 468, + 330 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We evaluated the effectiveness of the CE module on ModelNet40 with PointNeXt-S (C=64). To demonstrate the experiment's validity, we compared the overall accuracy (OA) for pruning rates of $40\\%$ , $50\\%$ , $60\\%$ , and $70\\%$ , respectively. Three sets of experiments are carried out. First, the 'CE' group selected filters in order of value from the CE module. Secondly, the 'Random' group is used to randomly select filters for pruning, and finally, the 'Reverse' is used to select filters according to the coordinate-enhanced score from low to high. As shown in Fig. 2, the channel with a higher coordinate-enhanced score has higher accuracy, which means that the channel with a higher coordinate-enhanced score has higher importance and should be retained in the pruning process.", + "bbox": [ + 75, + 330, + 468, + 542 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Knowledge Recycling", + "text_level": 1, + "bbox": [ + 76, + 553, + 279, + 569 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Sec. 3.2 shows that feature maps can reflect the importance of the corresponding channels. Another problem is that the importance determination may be unstable and sensitive to small changes in the input data as the feature maps are highly related to the samples of input data. Therefore, we aim to reduce the impact of such data variation. Through the analysis of the PNNs in Sec. 3.1, we found that some points are discarded to obtain hierarchical points set feature. These discarded points are informative as well since the sampling mechanism are highly random and can be leveraged to reduce the impact of data variation. The Knowledge Recycling (KR) module is proposed to reuse the discarded points to improve the robustness of channel pruning.", + "bbox": [ + 75, + 577, + 468, + 772 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For those centroids that are computed in $(l - 1)$ -th layer but discarded in $l$ -th layer due to sampling, which are equivalent to the sampled points. Therefore, the discarded centroids $x_{dis}$ are fed into $l$ -th convolutional layer to generate the feature map $\\mathbf{f}_{\\mathrm{dis}}^l$ , and $\\mathcal{L}_k(F_{dis_j}^i)$ is taken in use for the evaluation of channel importance.", + "bbox": [ + 75, + 773, + 468, + 869 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We calculate the relevant feature maps from the network parameters trained from the sampled points and use them as", + "bbox": [ + 76, + 869, + 468, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "part of the importance calculation for the current SA layer channels. Specifically, for each layer in the SA module, we obtain the features of the discard points by Eq. 1:", + "bbox": [ + 496, + 90, + 890, + 137 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {F} _ {d i s} ^ {l} = \\left\\{\\mathbf {f} _ {\\mathbf {d i s} _ {1}} ^ {l}, \\mathbf {f} _ {\\mathbf {d i s} _ {2}} ^ {l}, \\dots , \\mathbf {f} _ {\\mathbf {d i s} n _ {i}} ^ {l} \\right\\} \\in \\mathbb {R} ^ {n _ {i} \\times (3 + c _ {i})},\n$$\n", + "text_format": "latex", + "bbox": [ + 552, + 148, + 836, + 174 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $n_i$ is sampled points number, $c_i$ is the points feature dimension. So the supplement importance is:", + "bbox": [ + 498, + 185, + 890, + 215 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} _ {k} \\left(F _ {j} ^ {i}\\right) = \\mathcal {L} _ {k} \\left(F _ {\\text {d i s} j} ^ {i}\\right) \\tag {5} \\\\ = \\mathcal {M} \\left\\{\\mathrm {C E} \\left(F _ {j} ^ {i}, x _ {d i s} ^ {i}\\right) \\right\\}, \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 596, + 226, + 890, + 273 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $x_{dis}^{i}$ are discard points for recycling.", + "bbox": [ + 498, + 286, + 782, + 301 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "It should be noted that the KR module only needs to calculate $\\mathcal{F}_{dis}^{l}$ from the parameters trained by the sampled points and does not incur much additional overhead.", + "bbox": [ + 496, + 301, + 890, + 347 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Using $\\mathbf{CP}^3$ in Pruning Methods", + "text_level": 1, + "bbox": [ + 498, + 357, + 772, + 375 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The overall $\\mathrm{CP}^3$ improve the existing CNNs pruning methods by considering the input data of PNNs and the PNN structure in Sec. 3.2 and Sec. 3.3, respectively. In fact, $\\mathrm{CP}^3$ can complement the existing pruning methods, i.e., as a plug-in to the existing pruning methods, to improve the pruning performance on PNNs. Specifically, combining Eq. (4) and Eq. (5), we obtain the final pruning formula according to Eq. (2):", + "bbox": [ + 496, + 380, + 890, + 503 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\min _ {\\delta_ {i j}} \\sum_ {i = 1} ^ {K} \\sum_ {j = 1} ^ {n _ {i}} \\delta_ {i j} \\left(\\mathcal {L} \\left(\\mathbf {w} _ {j} ^ {i}\\right) + \\mathcal {L} _ {c} \\left(F _ {j} ^ {i}\\right) + \\mathcal {L} _ {k} \\left(F _ {j} ^ {i}\\right)\\right), \\tag {6} \\\\ \\text {s . t .} \\sum_ {j = 1} ^ {n _ {i}} \\delta_ {i j} = k _ {l}, \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 549, + 513, + 890, + 594 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\delta_{ij}$ is an indicator which is 1 if $\\mathbf{w}_j^i$ is to be pruned or 0 if $\\mathbf{w}_j^i$ is to be kept, $\\mathcal{L}(\\cdot)$ is original CNNs pruning method measure importance of a channel, $\\mathcal{L}_c$ and $\\mathcal{L}_k$ are coordinate-enhanced score and knowledge-recycling score, $k_l$ is the kept channel number.", + "bbox": [ + 496, + 606, + 890, + 686 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 500, + 700, + 632, + 717 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Experimental Settings", + "text_level": 1, + "bbox": [ + 500, + 724, + 705, + 741 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Baseline models and datasets To demonstrate the effectiveness and generality of the proposed $\\mathrm{CP}^3$ , we tested it on three different 3D tasks and five datasets with various PNNs and three recent advanced channel pruning methods. The evaluated pruning methods include HRank (2020) [24], ResRep (2021) [7], and CHIP (2021) [40]. For the classification task, we chose the classical PointNet++ [33] and PointNeXt [35] models as the original networks and conducted experiments on ModelNet40 [47] and ScanObjectNN [43]. Specifically, for PointNeXt-S we tested", + "bbox": [ + 496, + 750, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5306", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/d903a810808ffa5be20adff5f3999911979761de4397f1cf4c62e5b4d46bcea6.jpg", + "table_caption": [ + "Table 1. Comparisons of classification on the ModelNet40 [47] test set with PointNet++ [33], PointNeXt-S (C=32) [35], and PointNeXt-S (C=64). For PointNeXt-S (C=32), we report the baseline results from the original paper. For PointNet++ and PointNeXt-S (C=64), we report the baseline results obtained by OpenPoints [35] re-implementations trained with the improved strategies." + ], + "table_footnote": [], + "table_body": "
MethodPointNet++PointNeXt-S (C=32)PointNeXt-S (C=64)
OAmAccParams. (M)GFLOPs (↓%)OAmAccParams. (M)GFLOPs (↓%)OAmAccParams. (M)GFLOPs (↓%)
Baseline92.8089.901.471.71 (−)92.9989.61.371.64 (−)93.4491.054.526.49 (−)
HRank92.5989.830.860.77 (55.0)92.8789.970.740.71 (56.7)92.2389.812.122.69 (56.7)
HRank +CP392.9589.910.840.75 (56.1)93.2390.560.710.67 (59.1)93.5290.332.012.58 (59.1)
HRank91.7988.820.590.42 (75.4)92.6389.120.500.39 (76.2)92.7190.451.331.56 (76.2)
HRank +CP392.5488.520.570.39 (77.2)93.0390.920.490.38 (76.8)93.0790.551.281.50 (76.8)
HRank91.3488.180.360.15 (91.2)92.7389.980.430.29 (82.3)92.8389.990.740.71 (82.3)
HRank +CP391.7188.680.340.13 (92.4)92.9990.110.400.27 (83.5)93.1190.520.710.67 (83.5)
ResRep92.7190.390.850.73 (57.3)92.8390.150.810.69 (57.9)91.6188.422.082.34 (57.9)
ResRep+CP393.2790.480.820.70 (59.1)93.3590.930.790.67 (59.1)92.9390.661.892.02 (59.1)
ResRep92.5089.250.570.41 (76.0)92.6490.010.510.40 (75.6)91.6789.271.131.92 (75.6)
ResRep+CP392.4689.430.540.40 (76.6)93.4190.870.490.38 (76.8)93.1190.821.021.89 (76.8)
ResRep92.1189.000.550.24 (86.0))92.3089.310.340.21 (87.2)89.5488.540.720.69 (87.2)
ResRep+CP392.4889.210.580.21 (87.7)92.9590.700.330.18 (89.0)91.0289.820.690.65 (89.0)
CHIP92.7989.230.820.73 (57.3)93.1190.270.710.67 (59.1)93.0390.601.481.44 (59.1)
CHIP+CP392.9990.660.810.70 (59.1)93.3590.800.690.65 (60.4)93.3591.111.451.40 (60.4)
CHIP92.4589.190.570.39 (77.2)92.7189.900.490.38 (76.8)92.7990.390.920.76 (76.8)
CHIP+CP392.9189.650.540.35 (79.5)93.0390.510.480.37 (77.4)93.2390.300.890.74 (77.4)
CHIP92.2689.560.360.15 (91.2)92.4289.130.320.16 (90.2)90.8388.700.650.46 (90.2)
CHIP+CP392.7190.410.340.13 (92.4)92.5090.350.300.14 (91.5)92.8790.250.630.44 (91.5)
", + "bbox": [ + 80, + 136, + 890, + 390 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "two settings with widths of 32 and 64. For the segmentation task, we conducted experiments on S3DIS [1] with PointNeXt-B and PointNeXt-L [35] as the original PNNs. For the object detection task, we pruned two point-based detectors (VoteNet [31] and GroupFree3D [25]) on SUN RGB-D [38] and ScanNetV2 [5].", + "bbox": [ + 75, + 415, + 470, + 506 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation details We conducted the classification and segmentation experiments with OpenPoints [35] and the object detection experiments with MMdetection3D [3], all on NVIDIA P100 GPUs. For a fair comparison, we used the same hyperparameter settings for each group of experiments. We either 1) measured the parameter/FLOP reductions of the pruned networks with similar performance or 2) measured the performance of the pruned networks with a similar amount of parameter/FLOP reductions. For all experiments, we reported the number of FLOPs ('GFLOPs') and parameters ('Params.') as well as task-specific metrics to be described in each experiment. More experimental results are available in the supplementary.", + "bbox": [ + 75, + 530, + 468, + 728 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Results on Classification", + "text_level": 1, + "bbox": [ + 76, + 739, + 302, + 755 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "ModelNet40 ModelNet40 [47] contains 9843 training and 2468 testing meshed CAD models belonging to 40 categories. Following the standard practice [33], we report the class-average accuracy (mAcc) and the overall accuracy (OA) on the testing set. We compared the pruned networks directly by HRank, ResRep, and CHIP and those pruned by the three pruning methods with $\\mathrm{CP}^3$ . As shown in Tab. 1, our $\\mathrm{CP}^3$ improved the performance of existing CNNs pruning methods for different PNNs with various pruning", + "bbox": [ + 75, + 763, + 470, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "rates. $\\mathrm{CP}^3$ had higher accuracy scores with a similar (and mostly higher) pruning rates. Notably, with the pruning rate of $58\\%$ , $\\mathrm{CP}^3$ usually produced compact PNNs with even better accuracy scores than the original PNNs, which is difficult for pruning methods without $\\mathrm{CP}^3$ .", + "bbox": [ + 496, + 414, + 893, + 492 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "ScanObjectNN We also conducted experiments on the ScanObjectNN benchmark [43]. ScanObjectNN contains 15000 objects categorized into 15 classes with 2902 unique object instances in the real world. As reported in Tab. 2, $\\mathrm{CP}^3$ surpassed existing CNN pruning methods directly applied to PNNs. For example, comparing with the baseline pruning method HRank, $\\mathrm{CP}^3$ boosts the OA score of PointNet++ by $1.14\\%$ (85.01→86.15), $1.14\\%$ (84.94→86.08), and $0.59\\%$ (84.03→84.62) for the three different pruning rates. Similarly, $\\mathrm{CP}^3$ obtains much higher OA and mAcc scores than the three baselines with different pruning rates on PointNet++ and PointNeXt-S. With the extensive experimental results on classification tasks, we show that $\\mathrm{CP}^3$ surely improved the pruned network's performance to a distinct extent compared to direct applications of 2D CNN pruning methods.", + "bbox": [ + 496, + 515, + 893, + 757 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Results on Semantic Segmentation", + "text_level": 1, + "bbox": [ + 498, + 770, + 802, + 787 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "S3DIS [1] is a challenging benchmark composed of 6 large-scale indoor areas, 271 rooms, and 13 semantic categories in total. Following a common protocol [41], we evaluated the presented approach in Area-5, which means to test on Area-5 and to train on the rest. For evaluation metrics, we used the mean classwise intersection over union (mIoU), the mean classwise accuracy (mAcc), and the overall point", + "bbox": [ + 496, + 794, + 893, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "5307", + "bbox": [ + 482, + 944, + 514, + 957 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/c32ffb58de38c076cb4eff328015779d5b80f96e7ad439c04ead9f3015ae68d9.jpg", + "table_caption": [ + "Table 2. Comparisons of classification on the ScanObjectNN [43] test set with PointNet++ [33], PointNeXt-S (C=32) [35], and PointNeXt-S (C=64). For PointNeXt-S (C=32), we report the baseline results from the original paper. For PointNet++ and PointNeXt-S (C=64), we report the baseline results obtained by OpenPoints [35] re-implementations trained with the improved strategies." + ], + "table_footnote": [], + "table_body": "
MethodPointNet++PointNeXt-S (C=32)PointNeXt-S (C=64)
OAmAccParams. (M)GFLOPs (↓%)OAmAccParams. (M)GFLOPs (↓%)OAmAccParams. (M)GFLOPs (↓%)
Baseline86.2084.401.471.71 (−)87.4085.391.371.64 (−)88.2086.844.526.49 (−)
HRank85.0183.330.820.73 (57.3)87.0285.850.720.69 (57.9)87.5185.482.222.91 (55.2)
HRank+CP386.1584.400.800.70 (59.1)87.4786.210.700.67 (59.1)87.9586.022.162.82 (56.5)
HRank84.9483.430.590.43 (74.9)84.7981.930.500.39 (76.2)84.6681.001.331.56 (76.0)
HRank+CP386.0884.510.590.41 (76.0)86.4083.940.480.37 (77.4)86.4384.941.281.50 (76.9)
HRank84.0382.160.370.16 (90.6)81.3378.320.320.17 (89.6)85.3983.830.730.71 (89.1)
HRank+CP384.6282.290.360.15 (91.2)84.2182.130.310.16 (90.2)86.2684.360.700.67 (89.7)
ResRep86.7784.780.820.76 (55.6)86.3284.450.690.65 (60.4)86.5884.022.122.82 (56.5)
ResRep+CP387.1185.500.820.71 (58.5)86.6684.750.670.64 (61.0)88.5286.002.032.74 (57.8)
ResRep83.9982.340.600.44 (74.3)85.2883.270.480.38 (76.8)85.2282.031.391.65 (74.6)
ResRep+CP384.3283.680.590.43 (74.9)86.2284.320.470.37 (77.4)86.7384.231.321.54 (76.3)
ResRep83.7981.910.400.29 (83.0)83.6681.660.410.25 (84.8)85.0281.270.600.68 (89.5)
ResRep+CP384.8082.830.400.26 (84.8)85.0183.030.400.23 (86.0)86.1383.400.580.52 (92.0)
CHIP86.1484.980.820.73 (57.3)87.1385.090.700.67 (59.1)88.4587.402.112.74 (57.8)
CHIP+CP386.2584.680.800.70 (59.1)87.5485.670.690.65 (60.4)88.5886.452.052.65 (59.2)
CHIP84.4582.250.590.42 (75.4)85.2883.280.500.4 (75.6)86.6885.121.371.64 (74.7)
CHIP+CP385.5984.420.570.40 (76.6)86.2983.990.490.38 (76.8)87.7986.811.321.56 (76.0)
CHIP83.2781.150.360.15 (91.2)81.3778.990.340.19 (88.4)83.9081.830.440.32 (95.1)
CHIP+CP384.0382.220.350.14 (91.8)82.1279.410.330.18 (89.0)84.2582.170.420.29 (95.5)
", + "bbox": [ + 80, + 136, + 890, + 390 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/37bc540125a381ca2b601431a899cece853123c8738147b065a32d5d163166c2.jpg", + "table_caption": [ + "Table 3. Comparisons of semantic segmentation on the S3DIS dataset (evaluated in Area-5) with PointNeXt-B and PointNeXt-L [35]." + ], + "table_footnote": [], + "table_body": "
MethodPointNeXt-BPointNeXt-L
OAmAccmIoUParams. (M)GFLOPs (↓%)OAmAccmIoUParams. (M)GFLOPs (↓%)
Baseline89.4073.9067.503.838.80 (-)90.1075.7069.307.1315.24 (-)
HRank89.0472.1465.661.724.01 (54.4)88.8873.6166.803.206.83 (55.2)
HRank+CP389.2473.7666.951.613.80 (56.8)89.4474.2767.533.006.48 (57.5)
HRank88.8172.1665.580.852.04 (76.8)88.2170.9364.301.583.47 (77.2)
HRank+CP389.0272.8266.410.781.85 (78.9)88.3571.2664.711.443.14 (79.4)
CHIP88.8973.2666.571.663.93 (55.3)89.1673.7267.093.096.68 (56.2)
CHIP+CP389.6873.1466.801.563.65 (58.5)89.4774.2067.282.916.22 (59.2)
CHIP88.6773.2466.680.811.92 (78.2)88.5871.5865.181.503.27 (78.5)
CHIP+CP389.8173.3666.950.741.79 (79.7)89.2071.6665.241.383.04 (80.1)
", + "bbox": [ + 122, + 422, + 849, + 584 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/a9110dbc115029590592b206cc65db3176840b3a33f994ea2c7465ed9ede1bb6.jpg", + "table_caption": [ + "Table 4. Comparisons of object detection on the ScanNet dataset." + ], + "table_footnote": [], + "table_body": "
MethodmAP@0.25mAP@0.50Params. (K)GFLOPs (↓%)
Baseline (VoteNet)62.3440.82641.925.78 (−)
ResRep62.4540.95251.232.45 (57.6)
ResRep+CP363.9241.47242.262.41 (58.1)
ResRep61.7840.54180.491.83 (68.3)
ResRep+CP362.9840.94160.481.78 (69.2)
", + "bbox": [ + 81, + 626, + 465, + 708 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "wise accuracy (OA). As the segmentation task is relatively difficult and the segmentation network structures are relatively complex, we pruned only the encoder part of the network and kept the original decoder part. The results are presented in Tab. 3. As expected, the performance of the pruned networks degraded more from the original networks than those in the classification experiments, but the performance was acceptable. Meanwhile, in all cases, with our $\\mathrm{CP}^3$ , PNNs have a higher accuracy at a higher pruning rate than without $\\mathrm{CP}^3$ . For example, for PointNeXt-B, comparing with directly applying CHIP, incorporating $\\mathrm{CP}^3$", + "bbox": [ + 75, + 734, + 472, + 902 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/0b34a508152e7039850b67eeac9c800a37b7e17337f048d20860f21f9eb35b3a.jpg", + "table_caption": [ + "Table 5. Comparisons of object detection on SUN RGB-D." + ], + "table_footnote": [], + "table_body": "
MethodmAP@0.25mAP@0.50Params. (K)GFLOPs (↓%)
Baseline (VoteNet)59.7835.77641.925.78 (-)
ResRep59.3736.80179.912.42 (58.13)
ResRep+CP360.1037.37172.352.20 (61.93)
ResRep59.0135.91135.131.84 (68.17)
ResRep+CP359.1836.24129.641.83 (68.34)
", + "bbox": [ + 504, + 626, + 887, + 708 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "obtained much higher OA, mAcc and mIoU scores (1.2% on OA and 0.3% on mIoU). The results about segmentation have well validated the generalization of $\\mathrm{CP}^3$ to new and difficult tasks.", + "bbox": [ + 496, + 734, + 890, + 792 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Results on 3D Object Detection", + "text_level": 1, + "bbox": [ + 500, + 806, + 777, + 821 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4.1 Evaluation and Comparison of VoteNet", + "text_level": 1, + "bbox": [ + 500, + 830, + 828, + 845 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Tabs. 4 and 5 show the results of the pruned VoteNet models on the ScanNetV2 and SUN RGB-D datasets, respectively. We evaluated the performance of our proposed method in", + "bbox": [ + 496, + 854, + 893, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "5308", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/2638ef9571e7a1e22007a34dc06e2208c14ead28e6bdf2e2d3b6c7d371a142c7.jpg", + "table_caption": [ + "Table 6. Comparisons of object detection on the ScanNet dataset. The baseline PNN model is GroupFree3D." + ], + "table_footnote": [], + "table_body": "
MethodmAP@0.25mAP@0.50Params. (K)GFLOPs (↓%)
Baseline68.2252.612438.3421.78 (-)
ResRep68.2451.481910.7310.90 (49.95)
ResRep+CP368.8652.081654.3510.88 (50.05)
ResRep67.2151.281703.348.71 (60.01)
ResRep+CP368.5751.851501.468.68 (60.14)
", + "bbox": [ + 78, + 121, + 467, + 209 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "terms of the mean average precision at IOU thresholds of 0.25 and 0.50 (mAP@25 and mAP@50).", + "bbox": [ + 75, + 233, + 468, + 263 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "ScanNetV2 ScanNetV2 [5] is a richly annotated dataset of 3D reconstructed meshes of indoor scenes. It contains about 1200 training examples collected from hundreds of different rooms and is annotated with semantic and instance segmentation for 18 object categories. Tab. 4 shows the results of directly applying ResRep and with $\\mathrm{CP}^3$ . As can be seen from the table, the accuracy of the 2D method directly applied to the 3D network decreased by a flops drop of about $60\\%$ , while our method achieves $1.58\\%$ and $0.65\\%$ improvements at mAP@0.25 and mAP@0.5 with a drop rate of $58.13\\%$ FLOPS. When FLOPs drop to about $70\\%$ , the accuracy of the direct porting CNNs pruning method works poorly, while the improvement of mAP@0.25 and mAP@0.5 of our method is $0.64\\%$ and $0.12\\%$ , respectively.", + "bbox": [ + 75, + 281, + 468, + 492 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "SUN RGB-D The SUN RGB-D dataset [38] consists of 10355 single-view indoor RGB-D images annotated with over 64000 3D bounding boxes and semantic labels for 37 categories. We conducted experiments on SUN RGB-D with the same setup as those on ScanNetV2. The findings are also similar to those in ScanNetV2. It can be observed from Tab. 5 that the accuracy of directly transplanted CNNs pruning method is reduced to some extent (mAP@0.25 reduced by 0.41) when FLOPs drop by $58.13\\%$ , while $\\mathrm{CP}^3$ improved the detection model by $0.32\\%$ mAP@0.25 and $1.60\\%$ mAP@0.5 with a FLOPs drop of $61.93\\%$ . Even when FLOPs drop to $70\\%$ , our method's mAP@0.25 drop only $0.6\\%$ , which is obviously better.", + "bbox": [ + 75, + 510, + 468, + 708 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4.2 Evaluation and Comparison on GroupFree3D", + "text_level": 1, + "bbox": [ + 76, + 724, + 449, + 742 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We also conducted experiments on another point-based 3D detection model, GroupFree3D, on ScanNetV2. Tab. 6 summarizes the pruning performance of our approach for GroupFree3D on the ScanNetV2 dataset. When targeting a moderate compression ratio, our approach can achieve $32.15\\%$ and $50.05\\%$ storage and computation reductions, respectively, with a $0.64\\%$ accuracy increase for mAP@0.25 over the baseline model. In the case of higher compression ratio, $\\mathrm{CP}^3$ still achieves superior performance to other methods. Specifically, the ResRep loses accuracy", + "bbox": [ + 75, + 750, + 468, + 901 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/dc0c9fe09136d8ad6d86a4fccd1996da0214aeee20dbd1400cd015aee157f51e.jpg", + "table_caption": [ + "Table 7. Ablation study of different components in $\\mathrm{CP}^3$ . Results are of classification on the ScanObjectNN dataset with PointNeXt-S (C=32) as the baseline. 'CE' represents the coordinate-enhanced module, and 'KR' represents the knowledge recycling module." + ], + "table_footnote": [], + "table_body": "
SettingCEKRPruning RateOAmAcc
Baseline-88.2086.40
HRank0.7584.7981.93
✓0.7585.6382.97
✓0.7585.1182.13
HRank+CP3✓✓0.7586.6383.63
HRank0.9081.3378.32
✓0.9083.6681.32
✓0.9083.1080.47
HRank+CP3✓✓0.9084.8382.74
", + "bbox": [ + 532, + 148, + 859, + 289 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "by $1.01\\%$ mAP@0.25 when the parameters and flops drop by $30.14\\%$ and $60.01\\%$ , while in our method, the accuracy increases $0.35\\%$ for mAP@0.25.", + "bbox": [ + 496, + 315, + 890, + 359 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.5. Ablation Studies", + "text_level": 1, + "bbox": [ + 500, + 371, + 663, + 386 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We conducted ablation studies to validate the Coordinate-Enhancement (CE) module and the KnowledgeRecycling (KR) module in $\\mathrm{CP}^3$ . All results provided in the section are tested on ScanObjectNN with PointNeXt-S $(C = 32)$ as the baseline and HRank as the pruning method. We evaluated the pruned networks' performance with the FLOPs drop of $75\\%$ and $90\\%$ , and with/without the CE and KR modules to the pruning method. From Tab. 7, we find that KR improves OA of $0.84\\%$ and $2.33\\%$ when pruning rate is $75\\%$ and $90\\%$ , respectively, while CE improves OA of $0.32\\%$ and $1.77\\%$ at $75\\%$ and $90\\%$ pruning rates, respectively. Bringing the two modules together, the OA improvement is $1.84\\%$ and $3.50\\%$ at $75\\%$ and $90\\%$ pruning rates, respectively. The ablation study results have validated the effectiveness of all designs in $\\mathrm{CP}^3$ .", + "bbox": [ + 496, + 395, + 890, + 622 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 637, + 617, + 652 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we focus on 3D point-based network pruning and design a 3D channel pruning plug-in $(\\mathrm{CP}^3)$ that can be used with existing 2D CNN pruning methods. To the best of our knowledge, this is the first pruning work explicitly considering the characteristics of point cloud data and point-based networks. Empirically, we show that the proposed $\\mathrm{CP}^3$ is universally effective for a wide range of point-based networks and 3D tasks.", + "bbox": [ + 496, + 662, + 890, + 784 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgement", + "text_level": 1, + "bbox": [ + 500, + 799, + 658, + 815 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This work was partially supported by National Natural Science Foundation of China (NSFC) under Grant 62271203, the Science and Technology Innovation Action Plan of Shanghai under Grant 2251110540, and Shanghai Pujiang Program under Grant 21PJ1420300.", + "bbox": [ + 496, + 824, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "5309", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Iro Armeni, Ozan Sener, Amir R. Zamir, Helen Jiang, Ioannis K. Brilakis, Martin Fischer, and Silvio Savarese. 3D semantic parsing of large-scale indoor spaces. In IEEE Conference on Computer Vision and Pattern Recognition, CVPR, pages 1534-1543, 2016. 6, 12", + "[2] Chen Chen, Zhe Chen, Jing Zhang, and Dacheng Tao. SASA: semantics-augmented set abstraction for point-based 3D object detection. In Thirty-Sixth AAAI Conference on Artificial Intelligence, AAAI, pages 221-229, 2022. 1", + "[3] MMDetection3D Contributors. MMDetection3D: Open-MMLab next-generation platform for general 3D object detection. https://github.com/open-mmlab/mmdetection3d, 2020.6", + "[4] Angela Dai, Angel X Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3D reconstructions of indoor scenes. In IEEE Conference on Computer Vision and Pattern Recognition, pages 5828-5839, 2017. 12", + "[5] Angela Dai, Angel X. Chang, Manolis Savva, Maciej Halber, Thomas A. Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3D reconstructions of indoor scenes. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR, pages 2432-2443, 2017. 6, 8", + "[6] Xiaohan Ding, Guiguang Ding, Yuchen Guo, and Jungong Han. Centripetal sgd for pruning very deep convolutional networks with complicated structure. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 4943-4953, 2019. 1, 3", + "[7] Xiaohan Ding, Tianxiang Hao, Jianchao Tan, Ji Liu, Jungong Han, Yuchen Guo, and Guiguang Ding. ResRep: Lossless CNN pruning via decoupling remembering and forgetting. In IEEE/CVF International Conference on Computer Vision, ICCV, pages 4510-4520, 2021. 1, 3, 5", + "[8] Anand Dubey, Avik Santra, Jonas Fuchs, Maximilian Lübke, Robert Weigel, and Fabian Lurz. Haradnet: Anchor-free target detection for radar point clouds using hierarchical attention and multi-task learning. Machine Learning with Applications, 8:100275, 2022. 1", + "[9] Hamidreza Fazlali, Yixuan Xu, Yuan Ren, and Bingbing Liu. A versatile multi-view framework for LiDAR-based 3D object detection with guidance from panoptic segmentation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR, pages 17171-17180, 2022. 1", + "[10] Kehong Gong, Bingbing Li, Jianfeng Zhang, Tao Wang, Jing Huang, Michael Bi Mi, Jiashi Feng, and Xinchao Wang. PoseTriplet: Co-evolving 3d human pose estimation, imitation, and hallucination under self-supervision. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR, pages 11017-11027, 2022. 1", + "[11] Google. TensorFlow lite. https://www.tensorflow.org/ lite, 2017. 2", + "[12] Yushuo Guan, Ning Liu, Pengyu Zhao, Zhengping Che, Kaigui Bian, Yanzhi Wang, and Jian Tang. DAIS: Automatic channel pruning via differentiable annealing indicator search. IEEE Transactions on Neural Networks and Learning Systems, TNNLS, pages 1-12, 2022. 1, 3" + ], + "bbox": [ + 78, + 114, + 470, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[13] Kai Han, Yunhe Wang, Qiulin Zhang, Wei Zhang, Chunjing Xu, and Tong Zhang. Model rubik's cube: Twisting resolution, depth and width for tinynets. Advances in Neural Information Processing Systems, 33:19353-19364, 2020. 3", + "[14] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In IEEE/CVF conference on Computer Vision and Pattern Recognition, CVPR, pages 770-778, 2016. 1", + "[15] Zejiang Hou, Minghai Qin, Fei Sun, Xiaolong Ma, Kun Yuan, Yi Xu, Yen-Kuang Chen, Rong Jin, Yuan Xie, and Sun-Yuan Kung. CHEX: channel exploration for CNN model compression. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR, pages 12277-12288, 2022. 3", + "[16] Andrew G. Howard, Menglong Zhu, Bo Chen, Dmitry Kalenichenko, Weijun Wang, Tobias Weyand, Marco Andretto, and Hartwig Adam. MobileNets: Efficient convolutional neural networks for mobile vision applications. CoRR, abs/1704.04861, 2017. 1", + "[17] Fengwei Jia, Xuan Wang, Jian Guan, Huale Li, Chen Qiu, and Shuhan Qi. Arank: Toward specific model pruning via advantage rank for multiple salient objects detection. Image and Vision Computing, 111:104192, 2021. 12", + "[18] Xiaotang Jiang, Huan Wang, Yiliu Chen, Ziqi Wu, Lichuan Wang, Bin Zou, Yafeng Yang, Zongyang Cui, Yu Cai, Tianhang Yu, et al. MNN: A universal and efficient inference engine. The Conference on Machine Learning and Systems, MLSys, pages 1-13, 2020. 3", + "[19] Hema Koppula, Abhishek Anand, Thorsten Joachims, and Ashutosh Saxena. Semantic labeling of 3d point clouds for indoor scenes. Advances in Neural Information Processing Systems, NeurIPS, pages 244-252, 2011. 2", + "[20] Bastian Leibe, Ales Leonardis, and Bernt Schiele. Robust object detection with interleaved categorization and segmentation. International Journal of Computer Vision, 77(1):259-289, 2008. 3", + "[21] Hao Li, Asim Kadav, Igor Durdanovic, Hanan Samet, and Hans Peter Graf. Pruning filters for efficient convnets. In International Conference on Learning Representations, ICLR, 2017. 3", + "[22] Kelin Li, Nicholas Baron, Xian Zhang, and Nicolas Rojas. Efficientgrasp: A unified data-efficient learning to grasp method for multi-fingered robot hands. IEEE Robotics and Automation Letters, 7(4):8619-8626, 2022. 1", + "[23] Lucas Liebenwein, Cenk Baykal, Harry Lang, Dan Feldman, and Daniela Rus. Provable filter pruning for efficient neural networks. In 8th International Conference on Learning Representations, ICLR, 2020. 3", + "[24] Mingbao Lin, Rongrong Ji, Yan Wang, Yichen Zhang, Baochang Zhang, Yonghong Tian, and Ling Shao. HRank: Filter pruning using high-rank feature map. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR, pages 1529-1538, 2020. 1, 3, 4, 5", + "[25] Liyang Liu, Shilong Zhang, Zhanghui Kuang, Aojun Zhou, Jing-Hao Xue, Xinjiang Wang, Yimin Chen, Wenming Yang, Qingmin Liao, and Wayne Zhang. Group Fisher pruning for practical network compression. In International Conference on Machine Learning, ICML, pages 7021-7032, 2021. 6" + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "5310", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[26] Zhuang Liu, Jianguo Li, Zhiqiang Shen, Gao Huang, Shoumeng Yan, and Changshui Zhang. Learning efficient convolutional networks through network slimming. In IEEE/CVF International Conference on Computer Vision, ICCV, pages 2736-2744, 2017. 1, 3", + "[27] Ze Liu, Zheng Zhang, Yue Cao, Han Hu, and Xin Tong. Group-free 3D object detection via transformers. In IEEE/CVF International Conference on Computer Vision, CVPR, pages 2949-2958, 2021. 3", + "[28] Xu Ma, Can Qin, Haoxuan You, Haoxi Ran, and Yun Fu. Rethinking network design and local geometry in point cloud: A simple residual MLP framework. In International Conference on Learning Representations, ICLR, 2022. 3", + "[29] Fanxu Meng, Hao Cheng, Ke Li, Huixiang Luo, Xiaowei Guo, Guangming Lu, and Xing Sun. Pruning filter in filter. In Advances in Neural Information Processing Systems, NeurIPS, volume 33, pages 17629-17640, 2020. 1, 3", + "[30] Quang-Hieu Pham, Thanh Nguyen, Binh-Son Hua, Gemma Roig, and Sai-Kit Yeung. JSIS3D: Joint semantic-instance segmentation of 3d point clouds with multi-task pointwise networks and multi-value conditional random fields. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR, 2019. 1", + "[31] Charles R Qi, Or Litany, Kaiming He, and Leonidas J Guibas. Deep hough voting for 3D object detection in point clouds. In IEEE/CVF International Conference on Computer Vision, ICCV, 2019. 3, 6, 12", + "[32] Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. Pointnet: Deep learning on point sets for 3D classification and segmentation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR, pages 652-660, 2017. 3", + "[33] Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J Guibas. Pointnet++: Deep hierarchical feature learning on point sets in a metric space. In Advances in Neural Information Processing Systems, NeurIPS, pages 5100-5109, 2017. 3, 5, 6, 7", + "[34] Guocheng Qian, Hasan Hammoud, Guohao Li, Ali Thabet, and Bernard Ghanem. ASSANet: An anisotropic separable set abstraction for efficient point cloud representation learning. In Advances in Neural Information Processing Systems, NeurIPS, pages 28119-28130, 2021. 3", + "[35] Guocheng Qian, Yuchen Li, Houwen Peng, Jinjie Mai, Hasan Hammoud, Mohamed Elhoseiny, and Bernard Ghanem. PointNeXt: Revisiting pointnet++ with improved training and scaling strategies. In Advances in Neural Information Processing Systems, NeurIPS, 2022. 3, 5, 6, 7, 12, 13", + "[36] Ruizhi Shao, Zerong Zheng, Hongwen Zhang, Jingxiang Sun, and Yebin Liu. DiffuStereo: High quality human reconstruction via diffusion-based stereo using sparse cameras. In European Conference on Computer Vision, ECCV, pages 702-720, 2022. 1", + "[37] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. In International Conference on Learning Representations, ICLR, 2015. 1, 2" + ], + "bbox": [ + 78, + 90, + 468, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[38] Shuran Song, Samuel P Lichtenberg, and Jianxiong Xiao. Sun RGB-D: A RGB-D scene understanding benchmark suite. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR, pages 567-576, 2015. 6, 8", + "[39] Shuran Song, Samuel P Lichtenberg, and Jianxiong Xiao. Sun RGB-D: A RGB-D scene understanding benchmark suite. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR, pages 567-576, 2015. 12", + "[40] Yang Sui, Miao Yin, Yi Xie, Huy Phan, Saman Aliari Zonouz, and Bo Yuan. CHIP: Channel independence-based pruning for compact neural networks. In Advances in Neural Information Processing Systems, NeurIPS, pages 24604-24616, 2021. 1, 3, 4, 5, 12", + "[41] Lyne P. Tchapmi, Christopher B. Choy, Iro Armeni, JunYoung Gwak, and Silvio Savarese. SEGCloud: Semantic segmentation of 3d point clouds. In International Conference on 3D Vision, 3DV, pages 537-547, 2017. 6", + "[42] Hugues Thomas, Charles R Qi, Jean-Emmanuel Deschaud, Beatz Marcotegui, François Goulette, and Leonidas J Guibas. KpConv: Flexible and deformable convolution for point clouds. In IEEE/CVF international conference on computer vision, CVPR, pages 6411–6420, 2019. 3", + "[43] Mikaela Angelina Uy, Quang-Hieu Pham, Binh-Son Hua, Duc Thanh Nguyen, and Sai-Kit Yeung. Revisiting point cloud classification: A new benchmark dataset and classification model on real-world data. In IEEE/CVF International Conference on Computer Vision, ICCV, pages 1588–1597, 2019. 5, 6, 7", + "[44] Haowen Wang, Mingyuan Wang, Zhengping Che, Zhiyuan Xu, Xiuquan Qiao, Mengshi Qi, Feifei Feng, and Jian Tang. RGB-Depth fusion GAN for indoor depth completion. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR, pages 6199-6208, 2022. 1", + "[45] Wenxiao Wang, Minghao Chen, Shuai Zhao, Long Chen, Jinming Hu, Haifeng Liu, Deng Cai, Xiaofei He, and Wei Liu. Accelerate cnns from three dimensions: A comprehensive pruning framework. In Marina Meila and Tong Zhang, editors, Proceedings of the 38th International Conference on Machine Learning, ICML 2021, 18-24 July 2021, Virtual Event, volume 139 of Proceedings of Machine Learning Research, pages 10717-10726. PMLR, 2021. 3", + "[46] Wenxiao Wang, Cong Fu, Jishun Guo, Deng Cai, and Xiaofei He. COP: customized deep model compression via regularized correlation-based filter-level pruning. In Sarit Kraus, editor, Proceedings of the Twenty-Eighth International Joint Conference on Artificial Intelligence, IJCAI 2019, Macao, China, August 10-16, 2019, pages 3785-3791. ijcai.org, 2019. 3", + "[47] Zhirong Wu, Shuran Song, Aditya Khosla, Fisher Yu, Linguuang Zhang, Xiaou Tang, and Jianxiong Xiao. 3D ShapeNets: A deep representation for volumetric shapes. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR, pages 1912-1920, 2015. 5, 6", + "[48] Chenfeng Xu, Shijia Yang, Tomer Galanti, Bichen Wu, Xiangyu Yue, Bohan Zhai, Wei Zhan, Peter Vajda, Kurt Keutzer, and Masayoshi Tomizuka. Image2point: 3d point-cloud understanding with 2d image pretrained models. In" + ], + "bbox": [ + 501, + 92, + 892, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "5311", + "bbox": [ + 482, + 945, + 513, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "European Conference on Computer Vision, pages 638-656. Springer, 2022. 2", + "[49] Wei Yang, Chris Paxton, Maya Cakmak, and Dieter Fox. Human grasp classification for reactive human-to-robot handovers. In IEEE/RSJ International Conference on Intelligent Robots and Systems, IROS, pages 11123–11130, 2020. 1", + "[50] Xucheng Ye, Pengcheng Dai, Junyu Luo, Xin Guo, Yingjie Qi, Jianlei Yang, and Yiran Chen. Accelerating CNN training by pruning activation gradients. In European Conference on Computer Vision, ECCV, pages 322-338, 2020. 3", + "[51] Ruichi Yu, Ang Li, Chun-Fu Chen, Jui-Hsin Lai, Vlad I Morariu, Xintong Han, Mingfei Gao, Ching-Yung Lin, and Larry S Davis. Nisp: Pruning networks using neuron importance score propagation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR, pages 9194-9203, 2018. 3", + "[52] Hengshuang Zhao, Li Jiang, Jiaya Jia, Philip H. S. Torr, and Vladlen Koltun. Point transformer. In IEEE/CVF International Conference on Computer Vision, ICCV, pages 16239-16248, 2021. 3", + "[53] Wu Zheng, Weiliang Tang, Li Jiang, and Chi-Wing Fu. SESSD: self-ensembling single-stage object detector from point cloud. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR, pages 14494–14503, 2021. 1" + ], + "bbox": [ + 78, + 90, + 468, + 430 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "5312", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2023/CP3_ Channel Pruning Plug-In for Point-Based Networks/f7e54c81-931a-42ba-b8dd-1b55f7342f80_model.json b/2023/CP3_ Channel Pruning Plug-In for Point-Based Networks/f7e54c81-931a-42ba-b8dd-1b55f7342f80_model.json new file mode 100644 index 0000000000000000000000000000000000000000..26ab48d9846b693193bff47fe63a8bb52d2855ab --- /dev/null +++ b/2023/CP3_ Channel Pruning Plug-In for Point-Based Networks/f7e54c81-931a-42ba-b8dd-1b55f7342f80_model.json @@ -0,0 +1,2070 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.198, + 0.13, + 0.773, + 0.154 + ], + "angle": 0, + "content": "\\(\\mathbf{CP}^{3}\\): Channel Pruning Plug-in for Point-based Networks" + }, + { + "type": "text", + "bbox": [ + 0.103, + 0.178, + 0.868, + 0.227 + ], + "angle": 0, + "content": "\\(\\mathrm{Yaxin Peng}^{3}\\) Ning Liu\\(^{2,*}\\) Zhiying Che\\(^{2}\\) Zhiyuan Xu\\(^{2}\\) Chaomin Shen\\(^{1}\\) Guixu Zhang\\(^{1, \\boxtimes}\\) Xinmei Liu\\(^{1}\\) Feifei Feng\\(^{2}\\) Jian Tang\\(^{2, \\boxtimes}\\)" + }, + { + "type": "text", + "bbox": [ + 0.245, + 0.232, + 0.728, + 0.27 + ], + "angle": 0, + "content": "\\(^{1}\\)School of Computer Science, East China Normal University \\(^{2}\\)Midea Group" + }, + { + "type": "text", + "bbox": [ + 0.209, + 0.27, + 0.762, + 0.289 + ], + "angle": 0, + "content": "\\(^{3}\\)Department of Mathematics, School of Science, Shanghai University" + }, + { + "type": "text", + "bbox": [ + 0.198, + 0.29, + 0.774, + 0.306 + ], + "angle": 0, + "content": "1{51205901049,51205901078}@stu.ecnu.edu.cn 1{cmshen,gxzhang}@cs.ecnu.edu.cn" + }, + { + "type": "text", + "bbox": [ + 0.173, + 0.307, + 0.793, + 0.323 + ], + "angle": 0, + "content": "2{liuning22, chezp, xuzy70, feifei.feng, tangjian22}@midea.com yaxin.peng@shu.edu.cn" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.374, + 0.313, + 0.389 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.409, + 0.474, + 0.817 + ], + "angle": 0, + "content": "Channel pruning can effectively reduce both computational cost and memory footprint of the original network while keeping a comparable accuracy performance. Though great success has been achieved in channel pruning for 2D image-based convolutional networks (CNNs), existing works seldom extend the channel pruning methods to 3D point-based neural networks (PNNs). Directly implementing the 2D CNN channel pruning methods to PNNs undermine the performance of PNNs because of the different representations of 2D images and 3D point clouds as well as the network architecture disparity. In this paper, we proposed \\( CP^3 \\), which is a Channel Pruning Plugin for Point-based network. \\( CP^3 \\) is elaborately designed to leverage the characteristics of point clouds and PNNs in order to enable 2D channel pruning methods for PNNs. Specifically, it presents a coordinate-enhanced channel importance metric to reflect the correlation between dimensional information and individual channel features, and it recycles the discarded points in PNN's sampling process and reconsiders their potentially-exclusive information to enhance the robustness of channel pruning. Experiments on various PNN architectures show that \\( CP^3 \\) constantly improves state-of-the-art 2D CNN pruning approaches on different point cloud tasks. For instance, our compressed PointNeXt-S on ScanObjectNN achieves an accuracy of \\( 88.52\\% \\) with a pruning rate of \\( 57.8\\% \\), outperforming the baseline pruning methods with an accuracy gain of \\( 1.94\\% \\)." + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.358, + 0.631, + 0.373 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.386, + 0.895, + 0.718 + ], + "angle": 0, + "content": "Convolutional Neural Networks (CNNs) often encounter the problems of overloaded computation and overweighted storage. The cumbersome instantiation of a CNN model leads to inefficient, uneconomic, or even impossible deployment in practice. Therefore, light-weight models that provide comparable results with much fewer computational costs are in great demand for nearly all applications. Channel pruning is a promising solution to delivering efficient networks. In recent years, 2D CNN channel pruning, e.g., pruning classical VGGNets [37], ResNets [14], MobileNets [16], and many other neural networks for processing 2D images [6, 7, 12, 24, 26, 29, 40], has been successfully conducted. Most channel pruning approaches focus on identifying redundant convolution filters (i.e., channels) by evaluating their importance. The cornerstone of 2D channel pruning methods is the diversified yet effective channel evaluation metrics. For instance, HRank [24] uses the rank of the feature map as the pruning metric and removes the low-rank filters that are considered to contain less information. CHIP [40] leverages channel independence to represent the importance of each feature mapping and eliminates less important channels." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.896, + 0.902 + ], + "angle": 0, + "content": "With the widespread application of depth-sensing technology, 3D vision tasks [9, 10, 36, 44] are a rapidly growing field starving for powerful methods. Apart from straightforwardly applying 2D CNNs, models built with Point-based Neural Networks (PNNs), which directly process point clouds from the beginning without unnecessary rendering, show their merits and are widely deployed on edge devices for various applications such as robots [22, 49] and self-driving [2, 53]. Compressing PNNs is crucial due to the limited resources of edge devices and multiple models for different tasks are likely to run simultaneously [8, 30]. Given the huge success of 2D channel pruning and the great" + }, + { + "type": "page_footnote", + "bbox": [ + 0.108, + 0.852, + 0.227, + 0.864 + ], + "angle": 0, + "content": "* Equal contributions." + }, + { + "type": "page_footnote", + "bbox": [ + 0.109, + 0.865, + 0.246, + 0.876 + ], + "angle": 0, + "content": "Corresponding authors." + }, + { + "type": "page_footnote", + "bbox": [ + 0.08, + 0.877, + 0.469, + 0.9 + ], + "angle": 0, + "content": "This work is done during Yaomin Huang and Xinmei Liu's internship at Midea Group." + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.852, + 0.469, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5302" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.152 + ], + "angle": 0, + "content": "demand for efficient 3D PNNs, we intuitively raise one question: shall we directly implement the existing pruning methods to PNNs following the proposed channel importance metrics in 2D CNNs pruning?" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.155, + 0.471, + 0.397 + ], + "angle": 0, + "content": "With this question in mind, we investigate the fundamental factors that potentially impair 2D pruning effectiveness on PNNs. Previous works [19, 48] have shown that point clouds record visual and semantic information in a significantly different way from 2D images. Specifically, a point cloud consists of a set of unordered points on objects' and environments' surfaces, and each point encodes its features, such as intensity along with the spatial coordinates \\((x,y,z)\\). In contrast, 2D images organize visual features in a dense and regular pixel array. Such data representation differences between 3D point clouds and 2D images lead to a) different ways of exploiting information from data and b) contrasting network architectures of PNNs and 2D CNNs. It is credible that only the pruning methods considering the two aspects (definitely not existing 2D CNN pruners) may obtain superior performance on PNNs." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.4, + 0.471, + 0.642 + ], + "angle": 0, + "content": "From the perspective of data representations, 3D point clouds provide more 3D feature representations than 2D images, but the representations are more sensitive to network channels. To be more specific, for 2D images, all three RGB channels represent basic information in an isotropic and homogeneous way so that the latent representations extracted by CNNs applied to the images. On the other hand, point clouds explicitly encode the spatial information in three coordinate channels, which are indispensable for extracting visual and semantic information from other channels. Moreover, PNNs employ the coordinate information in multiple layers as concatenated inputs for deeper feature extraction. Nevertheless, existing CNN pruning methods are designed only suitable for the plain arrangements of 2D data but fail to consider how the informative 3D information should be extracted from point clouds." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.645, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Moreover, the network architectures of PNNs are designed substantially different from 2D CNNs. While using smaller kernels [37] is shown to benefit 2D CNNs [37], it does not apply to networks for 3D point clouds. On the contrary, PNNs leverage neighborhoods at multiple scales to obtain both robust and detailed features. The reason is that small neighborhoods (analogous to small kernels in 2D CNNs) in point clouds consist of few points for PNNs to capture robust features. Due to the necessary sampling steps, the knowledge insufficiency issue becomes more severe for deeper PNN layers. In addition, PNNs use the random input dropout procedure during training to adaptively weight patterns detected at different scales and combine multi-scale features. This procedure randomly discards a large proportion of points and loses much exclusive information of the original data. Thus, the architecture disparity between 2D CNNs and PNNs affects the performance of" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.848, + 0.107 + ], + "angle": 0, + "content": "directly applying existing pruning methods to PNNs." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.108, + 0.892, + 0.514 + ], + "angle": 0, + "content": "In this paper, by explicitly dealing with the two characteristics of 3D task, namely the data representation and the PNN architecture design, we propose a Channel Pruning Plug-in for Point-based network named \\(\\mathrm{CP}^3\\), which can be applied to most 2D channel pruning methods for compressing PNN models. The proposed \\(\\mathrm{CP}^3\\) refines the channel importance, the key factor of pruning methods, from two aspects. Firstly, considering the point coordinates \\((x, y, \\text{and } z)\\) encode the spatial information and deeply affects feature extraction procedures in PNN layers, we determine the channel importance by evaluating the correlation between the feature map and its corresponding point coordinates by introducing a coordinate-enhancement module. Secondly, calculating channel importance in channel pruning is data-driven and sensitive to the input, and the intrinsic sampling steps in PNN naturally makes pruning methods unstable. To settle this problem, we make full use of the discarded points in the sampling process via a knowledge recycling module to supplement the evaluation of channel importance. This reduces the data sampling bias impact on the channel importance calculation and increases the robustness of the pruning results. Notably, both the coordinates and recycled points in \\(\\mathrm{CP}^3\\) do not participate in network training (with back-propagation) but only assist channel importance calculation in the reasoning phase. Thus, \\(\\mathrm{CP}^3\\) does not increase any computational cost of the pruned network. The contributions of this paper are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.503, + 0.516, + 0.892, + 0.591 + ], + "angle": 0, + "content": "- We systematically consider the characteristics of PNNs and propose a channel pruning plug-in named \\(\\mathrm{CP}^3\\) to enhance 2D CNN channel pruning approaches on 3D PNNs. To the best of our knowledge, \\(\\mathrm{CP}^3\\) is the first method to export existing 2D pruning methods to PNNs." + }, + { + "type": "text", + "bbox": [ + 0.503, + 0.594, + 0.892, + 0.654 + ], + "angle": 0, + "content": "- We propose a coordinate-enhanced channel importance score to guide point clouds network pruning, by evaluating the correlation between feature maps and corresponding point coordinates." + }, + { + "type": "text", + "bbox": [ + 0.503, + 0.657, + 0.892, + 0.716 + ], + "angle": 0, + "content": "- We design a knowledge recycling pruning scheme that increases the robustness of the pruning procedure, using the discarded points to improve the channel importance evaluation." + }, + { + "type": "text", + "bbox": [ + 0.503, + 0.719, + 0.892, + 0.78 + ], + "angle": 0, + "content": "- We show that using \\(\\mathrm{CP}^3\\) is consistently superior to directly transplanting 2D pruning methods to PNNs by extensive experiments on three 3D tasks and five datasets with different PNN models and pruning baselines." + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.516, + 0.892, + 0.78 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.793, + 0.642, + 0.808 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.818, + 0.696, + 0.834 + ], + "angle": 0, + "content": "2.1. 2D Channel Pruning" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.84, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Channel pruning (a.k.a., filter pruning) methods reduce the redundant filters while maintaining the original structure of CNNs and is friendly to prevailing inference acceleration engines such as TensorFlow-Lite (TFLite) [11] and" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5303" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.473, + 0.44 + ], + "angle": 0, + "content": "Mobile Neural Network (MNN) [18]. Mainstream channel pruning methods [6, 7, 12, 29] usually first evaluate the importance of channels by certain metrics and then prune (i.e., remove) the less important channels. Early work [21] uses the \\( l_{1} \\) norm of filters as importance score for channel pruning. Afterwards, learning parameters, such as the scaling factor \\( \\gamma \\) in the batch norm layer [26] and the reconstruction error in the final network layer [51], are considered as the importance scores for channel selection. The importance sampling distribution of channels [23] is also used for pruning. Recent works [15, 40] measure the correlation of multiple feature maps to determine the importance score of the filter for pruning. HRank [24] proposes a method for pruning filters based on the theory that low-rank feature maps contain less information. [50] leverages the statistical distribution of activation gradient and takes the smaller gradient as low importance score for pruning. [46] calculates the average importance of both the input feature maps and their corresponding output feature maps to determine the overall importance. [13, 45] compress CNNs from multiple dimensions While most channel pruning methods are designed for and tested on 2D CNNs, our \\( \\mathrm{CP}^3 \\) can work in tandem with existing pruners for 3D point-based networks." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.453, + 0.449, + 0.469 + ], + "angle": 0, + "content": "2.2. Point-based Networks for Point Cloud Data" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.479, + 0.473, + 0.903 + ], + "angle": 0, + "content": "Point-based Neural Networks (PNNs) directly process point cloud data with a flexible range of receptive field, have no positioning information loss, and thus keep more accurate spatial information. As a pioneer work, PointNet [32] learns the spatial encoding directly from the input point clouds and uses the characteristics of all points to obtain the global representations. PointNet++ [33] further proposes a multi-level feature extraction structure to extract local and global features more effectively. KPCov [42] proposes a new point convolution operation to learn local movements applied to kernel points. ASSANet [34] proposes a separable set abstraction module that decomposes the normal SA module in PointNet++ into two separate learning phases for channel and space. PointMLP [28] uses residual point blocks to extract local features, transforms local points using geometric affine modules, and extracts geometric features before and after the aggregation operation. PointNeXt [35] uses inverted residual bottleneck and separable multilayer perceptrons to achieve more efficient model scaling. Besides classification, PNNs also serve as backbones for other 3D tasks. VoteNet [31] effectively improves the 3D object detection accuracy through the Hough voting mechanism [20]. PointTransformer [52] designs models improving prior work across domains and tasks. GroupFree3D [27] uses the attention mechanism to automatically learn the contribution of each point to the object. In this paper, we show that \\(\\mathrm{CP}^3\\) can be widely applied to point-based networks on a variety of point cloud bench" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.092, + 0.79, + 0.107 + ], + "angle": 0, + "content": "marks and representative original networks." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.121, + 0.636, + 0.138 + ], + "angle": 0, + "content": "3. Methodology" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.146, + 0.895, + 0.328 + ], + "angle": 0, + "content": "Although point-based networks are similar to CNN in concrete realization, they have fundamental differences in data representation and network architecture design. To extend the success of CNN pruning to PNN, two modules are proposed in \\(\\mathrm{CP}^3\\) taking advantage from the dimensional information and discarded points: 1) coordinate-enhancement (CE) module, which produces a coordinate-enhanced score to estimate the channel importance by combining dimensional and feature information, and 2) knowledge recycling module reusing the discarded points to improve the channel importance evaluation criteria and increase the robustness." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.338, + 0.764, + 0.353 + ], + "angle": 0, + "content": "3.1. Formulations and Motivation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.361, + 0.895, + 0.619 + ], + "angle": 0, + "content": "Point-based networks PNN is a unified architecture that directly takes point clouds as input. It builds hierarchical groups of points and progressively abstracts larger local regions along the hierarchy. PNN is structurally composed by a number of set abstraction (SA) blocks. Each SA block consists of 1) a sampling layer iteratively samples the farthest point to choose a subset of points from input points, 2) a group layer gathers neighbors of centroid points to a local region, 3) a set of shared Multi-Layer Perceptrons (MLPs) to extract features, and 4) a reduction layer to aggregate features in the neighbors. Formally speaking, a SA block takes an \\( n_{i-1} \\times (d + c_{i-1}) \\) matrix as input that is from \\( n_{i-1} \\) points with \\( d \\)-dim coordinates and \\( c_{i-1} \\)-dim point feature. It outputs an \\( n_i \\times (d + c_i) \\) matrix of \\( n_i \\) subsampled points with \\( d \\)-dimensional coordinates (i.e., \\( d = 3 \\)) and new \\( c_i \\)-dimensional feature vectors summarizing local context. The SA block is formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.584, + 0.625, + 0.892, + 0.653 + ], + "angle": 0, + "content": "\\[\n\\mathbf {F} _ {i} ^ {l + 1} = \\mathcal {R} \\left\\{h _ {\\Theta} \\left(\\left[ \\mathbf {F} _ {j} ^ {l}; \\mathbf {x} _ {j} ^ {l} - \\mathbf {x} _ {i} ^ {l} \\right]\\right) \\right\\}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.662, + 0.895, + 0.743 + ], + "angle": 0, + "content": "where \\( h_\\Theta \\) is MLPs to extract grouped points feature, \\( \\mathcal{R} \\) is the reduction layer (e.g. max-pooling) to aggregate features in the neighbors \\( \\{j : (i,j) \\in \\mathcal{N}\\} \\), \\( F_j^l \\) is the features of neighbor \\( j \\) in the \\( l \\)-th layer, \\( x_i^l \\) and \\( x_j^l \\) are input points coordinates and coordinates of neighbor \\( j \\) in the \\( l \\)-th layer." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.759, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Channel pruning Assume a pre-trained PNN model has a set of \\(K\\) convolutional layers, and \\(\\mathcal{A}^l\\) is the \\(l\\)-th convolution layer. The parameters in \\(\\mathcal{A}^l\\) can be represented as a set of filters \\(\\mathcal{W}_{\\mathcal{A}^l} = \\{\\mathbf{w}_1^l,\\mathbf{w}_2^l,\\dots ,\\mathbf{w}_{c_l}^l\\} \\in \\mathbb{R}^{(d + c_l)\\times (d + c_{l - 1})\\times k_l\\times k_l}\\) where \\(j\\)-th filter is \\(\\mathbf{w}_j^l\\in \\mathbb{R}^{(d + c_{l - 1})\\times k_l\\times k_j}\\). \\((d + c_{l})\\) represents the number of filters in \\(\\mathcal{A}^l\\) and \\(k_{l}\\) denotes the kernel size. The outputs of filter, i.e., feature map, are denoted as \\(\\mathcal{F}^l = \\{\\mathbf{f}_1^l,\\mathbf{f}_2^l,\\ldots ,\\mathbf{f}_{n_i}^l\\} \\in \\mathbb{R}^{n_i\\times (d + c_i)}\\). Channel pruning aims to identify and remove the less importance filter from the" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.957 + ], + "angle": 0, + "content": "5304" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.095, + 0.885, + 0.281 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.285, + 0.893, + 0.373 + ], + "angle": 0, + "content": "Figure 1. The framework of \\(\\mathrm{CP}^3\\). The figure shows the specific pruning process of one of the SA blocks. Whether a channel in a PNN is pruned is determined by three parts: 1) Original channel importance: obtained from the original CNNs channel pruning method (e.g., HRank [24], CHIP [40]). 2) Discarded channel importance: obtained from the Knowledge-Recycling module by leveraging the discarded points in the network to supplement the channel importance evaluation of the corresponding points and improve the robustness of the channel selection. 3) CE (Coordinate-Enhanced) channel importance: obtained from calculating the correlation between the feature map and its corresponding points coordinates to guide point clouds network pruning." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.389, + 0.471, + 0.421 + ], + "angle": 0, + "content": "original networks. In general, channel pruning can be formulated as the following optimization problem:" + }, + { + "type": "equation", + "bbox": [ + 0.146, + 0.426, + 0.47, + 0.466 + ], + "angle": 0, + "content": "\\[\n\\min _ {\\delta_ {i j}} \\sum_ {i = 1} ^ {K} \\sum_ {j = 1} ^ {n _ {i}} \\delta_ {i j} \\mathcal {L} \\left(\\mathbf {w} _ {j} ^ {i}\\right), \\text {s . t .} \\sum_ {j = 1} ^ {n _ {i}} \\delta_ {i j} = k _ {l}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.473, + 0.471, + 0.522 + ], + "angle": 0, + "content": "where \\(\\delta_{ij}\\) is an indicator which is 1 if \\(\\mathbf{w}_j^i\\) is to be pruned or 0 if \\(\\mathbf{w}_j^i\\) is to be kept, \\(\\mathcal{L}(\\cdot)\\) measures the importance of a filter and \\(k_l\\) is the kept filter number." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.539, + 0.473, + 0.903 + ], + "angle": 0, + "content": "Robust importance metric for channel pruning The metrics for evaluating the importance of filters is critical. Existing CNN pruning methods design a variety of \\(\\mathcal{L}(\\cdot)\\) on the filters. Consider the feature maps, contain rich and important information of both filter and input data, approaches using feature information have become popular and achieved state-of-the-art performance for channel pruning. The results of the feature maps may vary depending on the variability of the input data. Therefore, when the importance of one filter solely depends on the information represented by its own generated feature map, the measurement of the importance may be unstable and sensitive to the slight change of input data. So we have taken into account the characteristics of point clouds data and point-based networks architecture to improve the robustness of channel importance in point-based networks. On the one hand, we propose a coordinate-enhancement module by evaluating the correlation between the feature map and its corresponding points coordinates to guide point clouds network pruning, which will be described in Sec. 3.2. On the other hand, we design a knowledge recycling pruning schema, using discarded points to improve the channel importance evaluation criteria and increase the robustness of the pruning module, which will be described in detail in Sec. 3.3." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.388, + 0.871, + 0.405 + ], + "angle": 0, + "content": "3.2. Coordinate-Enhanced Channel Importance" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.411, + 0.892, + 0.715 + ], + "angle": 0, + "content": "Dimensional information is critical in PNNs. The dimensional information (i.e., coordinates of the points) are usually adopted as input for feature extraction. Namely, the input and output of each SA block are concatenated with the coordinates of the points. Meanwhile, the intermediate feature maps reflect not only the information of the original input data but also the corresponding channel information. Therefore, the importance of the channel can be obtained from the feature maps, i.e., the importance of the corresponding channel. The dimensional information is crucial in point-based tasks and should be considered as part of importance metric. Thus the critical problem falls in designing a function that can well reflect the dimensional information richness of feature maps. The feature map, obtained by encoding points spatial \\( x \\), \\( y \\), and \\( z \\) coordinates, should be closely related to the original corresponding points coordinates. Therefore, we use the correlation between the current feature map and the corresponding input points coordinates to determine the importance of the filter. The designed Coordinate-Enhancement (CE) module based on Eq. (2):" + }, + { + "type": "equation", + "bbox": [ + 0.567, + 0.72, + 0.892, + 0.759 + ], + "angle": 0, + "content": "\\[\n\\min _ {\\delta_ {i j}} \\sum_ {i = 1} ^ {K} \\sum_ {j = 1} ^ {n _ {i}} \\delta_ {i j} \\mathcal {L} _ {c} \\left(\\mathbf {F} _ {j} ^ {i}\\right), \\text {s . t .} \\sum_ {j = 1} ^ {n _ {i}} \\delta_ {i j} = k _ {l}, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.764, + 0.895, + 0.843 + ], + "angle": 0, + "content": "where \\(\\delta_{ij}\\) is an indicator which is 1 if \\(\\mathbf{w}_j^i\\) is to be pruned or 0 if \\(\\mathbf{w}_j^i\\) is to be kept and \\(k_l\\) is the kept channel number. \\(\\mathcal{L}_c(\\cdot)\\) measures the importance of a channel from take account of the relationship between feature map and points coordinates which can be formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.604, + 0.847, + 0.892, + 0.868 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {c} \\left(F _ {j} ^ {i}\\right) = \\mathcal {M} \\left\\{\\mathrm {C E} \\left(F _ {j} ^ {i}, x ^ {i}\\right) \\right\\}, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.871, + 0.895, + 0.903 + ], + "angle": 0, + "content": "where CE obtains the coordinate-enhanced score by calculating correlation of each channel in the feature map with" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5305" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.156, + 0.089, + 0.386, + 0.102 + ], + "angle": 0, + "content": "Coordinate-Enhancing for Filter Importance" + }, + { + "type": "image", + "bbox": [ + 0.115, + 0.102, + 0.428, + 0.219 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.224, + 0.47, + 0.266 + ], + "angle": 0, + "content": "Figure 2. Comparisons on the pruned model accuracy with different pruning metrics with CE scores. Results are on ModelNet40 with PointNeXt-S (C=64)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.286, + 0.469, + 0.331 + ], + "angle": 0, + "content": "the original coordinates, and \\(\\mathcal{M}\\) takes the maximum value. Hence, higher coordinate-enhanced score (i.e., \\(\\mathcal{L}_c\\)) serve as a reliable measurement for information richness." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.332, + 0.47, + 0.543 + ], + "angle": 0, + "content": "We evaluated the effectiveness of the CE module on ModelNet40 with PointNeXt-S (C=64). To demonstrate the experiment's validity, we compared the overall accuracy (OA) for pruning rates of \\(40\\%\\), \\(50\\%\\), \\(60\\%\\), and \\(70\\%\\), respectively. Three sets of experiments are carried out. First, the 'CE' group selected filters in order of value from the CE module. Secondly, the 'Random' group is used to randomly select filters for pruning, and finally, the 'Reverse' is used to select filters according to the coordinate-enhanced score from low to high. As shown in Fig. 2, the channel with a higher coordinate-enhanced score has higher accuracy, which means that the channel with a higher coordinate-enhanced score has higher importance and should be retained in the pruning process." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.554, + 0.28, + 0.57 + ], + "angle": 0, + "content": "3.3. Knowledge Recycling" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.578, + 0.469, + 0.773 + ], + "angle": 0, + "content": "Sec. 3.2 shows that feature maps can reflect the importance of the corresponding channels. Another problem is that the importance determination may be unstable and sensitive to small changes in the input data as the feature maps are highly related to the samples of input data. Therefore, we aim to reduce the impact of such data variation. Through the analysis of the PNNs in Sec. 3.1, we found that some points are discarded to obtain hierarchical points set feature. These discarded points are informative as well since the sampling mechanism are highly random and can be leveraged to reduce the impact of data variation. The Knowledge Recycling (KR) module is proposed to reuse the discarded points to improve the robustness of channel pruning." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.774, + 0.469, + 0.87 + ], + "angle": 0, + "content": "For those centroids that are computed in \\((l - 1)\\)-th layer but discarded in \\(l\\)-th layer due to sampling, which are equivalent to the sampled points. Therefore, the discarded centroids \\(x_{dis}\\) are fed into \\(l\\)-th convolutional layer to generate the feature map \\(\\mathbf{f}_{\\mathrm{dis}}^l\\), and \\(\\mathcal{L}_k(F_{dis_j}^i)\\) is taken in use for the evaluation of channel importance." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.469, + 0.901 + ], + "angle": 0, + "content": "We calculate the relevant feature maps from the network parameters trained from the sampled points and use them as" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.138 + ], + "angle": 0, + "content": "part of the importance calculation for the current SA layer channels. Specifically, for each layer in the SA module, we obtain the features of the discard points by Eq. 1:" + }, + { + "type": "equation", + "bbox": [ + 0.553, + 0.149, + 0.838, + 0.175 + ], + "angle": 0, + "content": "\\[\n\\mathcal {F} _ {d i s} ^ {l} = \\left\\{\\mathbf {f} _ {\\mathbf {d i s} _ {1}} ^ {l}, \\mathbf {f} _ {\\mathbf {d i s} _ {2}} ^ {l}, \\dots , \\mathbf {f} _ {\\mathbf {d i s} n _ {i}} ^ {l} \\right\\} \\in \\mathbb {R} ^ {n _ {i} \\times (3 + c _ {i})},\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.186, + 0.892, + 0.216 + ], + "angle": 0, + "content": "where \\( n_i \\) is sampled points number, \\( c_i \\) is the points feature dimension. So the supplement importance is:" + }, + { + "type": "equation", + "bbox": [ + 0.598, + 0.227, + 0.892, + 0.274 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} _ {k} \\left(F _ {j} ^ {i}\\right) = \\mathcal {L} _ {k} \\left(F _ {\\text {d i s} j} ^ {i}\\right) \\tag {5} \\\\ = \\mathcal {M} \\left\\{\\mathrm {C E} \\left(F _ {j} ^ {i}, x _ {d i s} ^ {i}\\right) \\right\\}, \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.287, + 0.783, + 0.303 + ], + "angle": 0, + "content": "where \\(x_{dis}^{i}\\) are discard points for recycling." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.303, + 0.892, + 0.348 + ], + "angle": 0, + "content": "It should be noted that the KR module only needs to calculate \\(\\mathcal{F}_{dis}^{l}\\) from the parameters trained by the sampled points and does not incur much additional overhead." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.358, + 0.774, + 0.375 + ], + "angle": 0, + "content": "3.4. Using \\(\\mathbf{CP}^3\\) in Pruning Methods" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.381, + 0.892, + 0.504 + ], + "angle": 0, + "content": "The overall \\(\\mathrm{CP}^3\\) improve the existing CNNs pruning methods by considering the input data of PNNs and the PNN structure in Sec. 3.2 and Sec. 3.3, respectively. In fact, \\(\\mathrm{CP}^3\\) can complement the existing pruning methods, i.e., as a plug-in to the existing pruning methods, to improve the pruning performance on PNNs. Specifically, combining Eq. (4) and Eq. (5), we obtain the final pruning formula according to Eq. (2):" + }, + { + "type": "equation", + "bbox": [ + 0.55, + 0.515, + 0.892, + 0.595 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\min _ {\\delta_ {i j}} \\sum_ {i = 1} ^ {K} \\sum_ {j = 1} ^ {n _ {i}} \\delta_ {i j} \\left(\\mathcal {L} \\left(\\mathbf {w} _ {j} ^ {i}\\right) + \\mathcal {L} _ {c} \\left(F _ {j} ^ {i}\\right) + \\mathcal {L} _ {k} \\left(F _ {j} ^ {i}\\right)\\right), \\tag {6} \\\\ \\text {s . t .} \\sum_ {j = 1} ^ {n _ {i}} \\delta_ {i j} = k _ {l}, \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.607, + 0.892, + 0.687 + ], + "angle": 0, + "content": "where \\(\\delta_{ij}\\) is an indicator which is 1 if \\(\\mathbf{w}_j^i\\) is to be pruned or 0 if \\(\\mathbf{w}_j^i\\) is to be kept, \\(\\mathcal{L}(\\cdot)\\) is original CNNs pruning method measure importance of a channel, \\(\\mathcal{L}_c\\) and \\(\\mathcal{L}_k\\) are coordinate-enhanced score and knowledge-recycling score, \\(k_l\\) is the kept channel number." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.701, + 0.633, + 0.718 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.726, + 0.706, + 0.742 + ], + "angle": 0, + "content": "4.1. Experimental Settings" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Baseline models and datasets To demonstrate the effectiveness and generality of the proposed \\(\\mathrm{CP}^3\\), we tested it on three different 3D tasks and five datasets with various PNNs and three recent advanced channel pruning methods. The evaluated pruning methods include HRank (2020) [24], ResRep (2021) [7], and CHIP (2021) [40]. For the classification task, we chose the classical PointNet++ [33] and PointNeXt [35] models as the original networks and conducted experiments on ModelNet40 [47] and ScanObjectNN [43]. Specifically, for PointNeXt-S we tested" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5306" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.089, + 0.895, + 0.133 + ], + "angle": 0, + "content": "Table 1. Comparisons of classification on the ModelNet40 [47] test set with PointNet++ [33], PointNeXt-S (C=32) [35], and PointNeXt-S (C=64). For PointNeXt-S (C=32), we report the baseline results from the original paper. For PointNet++ and PointNeXt-S (C=64), we report the baseline results obtained by OpenPoints [35] re-implementations trained with the improved strategies." + }, + { + "type": "table", + "bbox": [ + 0.081, + 0.137, + 0.892, + 0.391 + ], + "angle": 0, + "content": "
MethodPointNet++PointNeXt-S (C=32)PointNeXt-S (C=64)
OAmAccParams. (M)GFLOPs (↓%)OAmAccParams. (M)GFLOPs (↓%)OAmAccParams. (M)GFLOPs (↓%)
Baseline92.8089.901.471.71 (−)92.9989.61.371.64 (−)93.4491.054.526.49 (−)
HRank92.5989.830.860.77 (55.0)92.8789.970.740.71 (56.7)92.2389.812.122.69 (56.7)
HRank +CP392.9589.910.840.75 (56.1)93.2390.560.710.67 (59.1)93.5290.332.012.58 (59.1)
HRank91.7988.820.590.42 (75.4)92.6389.120.500.39 (76.2)92.7190.451.331.56 (76.2)
HRank +CP392.5488.520.570.39 (77.2)93.0390.920.490.38 (76.8)93.0790.551.281.50 (76.8)
HRank91.3488.180.360.15 (91.2)92.7389.980.430.29 (82.3)92.8389.990.740.71 (82.3)
HRank +CP391.7188.680.340.13 (92.4)92.9990.110.400.27 (83.5)93.1190.520.710.67 (83.5)
ResRep92.7190.390.850.73 (57.3)92.8390.150.810.69 (57.9)91.6188.422.082.34 (57.9)
ResRep+CP393.2790.480.820.70 (59.1)93.3590.930.790.67 (59.1)92.9390.661.892.02 (59.1)
ResRep92.5089.250.570.41 (76.0)92.6490.010.510.40 (75.6)91.6789.271.131.92 (75.6)
ResRep+CP392.4689.430.540.40 (76.6)93.4190.870.490.38 (76.8)93.1190.821.021.89 (76.8)
ResRep92.1189.000.550.24 (86.0))92.3089.310.340.21 (87.2)89.5488.540.720.69 (87.2)
ResRep+CP392.4889.210.580.21 (87.7)92.9590.700.330.18 (89.0)91.0289.820.690.65 (89.0)
CHIP92.7989.230.820.73 (57.3)93.1190.270.710.67 (59.1)93.0390.601.481.44 (59.1)
CHIP+CP392.9990.660.810.70 (59.1)93.3590.800.690.65 (60.4)93.3591.111.451.40 (60.4)
CHIP92.4589.190.570.39 (77.2)92.7189.900.490.38 (76.8)92.7990.390.920.76 (76.8)
CHIP+CP392.9189.650.540.35 (79.5)93.0390.510.480.37 (77.4)93.2390.300.890.74 (77.4)
CHIP92.2689.560.360.15 (91.2)92.4289.130.320.16 (90.2)90.8388.700.650.46 (90.2)
CHIP+CP392.7190.410.340.13 (92.4)92.5090.350.300.14 (91.5)92.8790.250.630.44 (91.5)
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.416, + 0.471, + 0.507 + ], + "angle": 0, + "content": "two settings with widths of 32 and 64. For the segmentation task, we conducted experiments on S3DIS [1] with PointNeXt-B and PointNeXt-L [35] as the original PNNs. For the object detection task, we pruned two point-based detectors (VoteNet [31] and GroupFree3D [25]) on SUN RGB-D [38] and ScanNetV2 [5]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.531, + 0.47, + 0.729 + ], + "angle": 0, + "content": "Implementation details We conducted the classification and segmentation experiments with OpenPoints [35] and the object detection experiments with MMdetection3D [3], all on NVIDIA P100 GPUs. For a fair comparison, we used the same hyperparameter settings for each group of experiments. We either 1) measured the parameter/FLOP reductions of the pruned networks with similar performance or 2) measured the performance of the pruned networks with a similar amount of parameter/FLOP reductions. For all experiments, we reported the number of FLOPs ('GFLOPs') and parameters ('Params.') as well as task-specific metrics to be described in each experiment. More experimental results are available in the supplementary." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.741, + 0.303, + 0.756 + ], + "angle": 0, + "content": "4.2. Results on Classification" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.471, + 0.903 + ], + "angle": 0, + "content": "ModelNet40 ModelNet40 [47] contains 9843 training and 2468 testing meshed CAD models belonging to 40 categories. Following the standard practice [33], we report the class-average accuracy (mAcc) and the overall accuracy (OA) on the testing set. We compared the pruned networks directly by HRank, ResRep, and CHIP and those pruned by the three pruning methods with \\(\\mathrm{CP}^3\\). As shown in Tab. 1, our \\(\\mathrm{CP}^3\\) improved the performance of existing CNNs pruning methods for different PNNs with various pruning" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.415, + 0.895, + 0.493 + ], + "angle": 0, + "content": "rates. \\(\\mathrm{CP}^3\\) had higher accuracy scores with a similar (and mostly higher) pruning rates. Notably, with the pruning rate of \\(58\\%\\), \\(\\mathrm{CP}^3\\) usually produced compact PNNs with even better accuracy scores than the original PNNs, which is difficult for pruning methods without \\(\\mathrm{CP}^3\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.516, + 0.895, + 0.758 + ], + "angle": 0, + "content": "ScanObjectNN We also conducted experiments on the ScanObjectNN benchmark [43]. ScanObjectNN contains 15000 objects categorized into 15 classes with 2902 unique object instances in the real world. As reported in Tab. 2, \\(\\mathrm{CP}^3\\) surpassed existing CNN pruning methods directly applied to PNNs. For example, comparing with the baseline pruning method HRank, \\(\\mathrm{CP}^3\\) boosts the OA score of PointNet++ by \\(1.14\\%\\) (85.01→86.15), \\(1.14\\%\\) (84.94→86.08), and \\(0.59\\%\\) (84.03→84.62) for the three different pruning rates. Similarly, \\(\\mathrm{CP}^3\\) obtains much higher OA and mAcc scores than the three baselines with different pruning rates on PointNet++ and PointNeXt-S. With the extensive experimental results on classification tasks, we show that \\(\\mathrm{CP}^3\\) surely improved the pruned network's performance to a distinct extent compared to direct applications of 2D CNN pruning methods." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.771, + 0.803, + 0.788 + ], + "angle": 0, + "content": "4.3. Results on Semantic Segmentation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.795, + 0.895, + 0.903 + ], + "angle": 0, + "content": "S3DIS [1] is a challenging benchmark composed of 6 large-scale indoor areas, 271 rooms, and 13 semantic categories in total. Following a common protocol [41], we evaluated the presented approach in Area-5, which means to test on Area-5 and to train on the rest. For evaluation metrics, we used the mean classwise intersection over union (mIoU), the mean classwise accuracy (mAcc), and the overall point" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.958 + ], + "angle": 0, + "content": "5307" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.089, + 0.895, + 0.133 + ], + "angle": 0, + "content": "Table 2. Comparisons of classification on the ScanObjectNN [43] test set with PointNet++ [33], PointNeXt-S (C=32) [35], and PointNeXt-S (C=64). For PointNeXt-S (C=32), we report the baseline results from the original paper. For PointNet++ and PointNeXt-S (C=64), we report the baseline results obtained by OpenPoints [35] re-implementations trained with the improved strategies." + }, + { + "type": "table", + "bbox": [ + 0.081, + 0.137, + 0.892, + 0.391 + ], + "angle": 0, + "content": "
MethodPointNet++PointNeXt-S (C=32)PointNeXt-S (C=64)
OAmAccParams. (M)GFLOPs (↓%)OAmAccParams. (M)GFLOPs (↓%)OAmAccParams. (M)GFLOPs (↓%)
Baseline86.2084.401.471.71 (−)87.4085.391.371.64 (−)88.2086.844.526.49 (−)
HRank85.0183.330.820.73 (57.3)87.0285.850.720.69 (57.9)87.5185.482.222.91 (55.2)
HRank+CP386.1584.400.800.70 (59.1)87.4786.210.700.67 (59.1)87.9586.022.162.82 (56.5)
HRank84.9483.430.590.43 (74.9)84.7981.930.500.39 (76.2)84.6681.001.331.56 (76.0)
HRank+CP386.0884.510.590.41 (76.0)86.4083.940.480.37 (77.4)86.4384.941.281.50 (76.9)
HRank84.0382.160.370.16 (90.6)81.3378.320.320.17 (89.6)85.3983.830.730.71 (89.1)
HRank+CP384.6282.290.360.15 (91.2)84.2182.130.310.16 (90.2)86.2684.360.700.67 (89.7)
ResRep86.7784.780.820.76 (55.6)86.3284.450.690.65 (60.4)86.5884.022.122.82 (56.5)
ResRep+CP387.1185.500.820.71 (58.5)86.6684.750.670.64 (61.0)88.5286.002.032.74 (57.8)
ResRep83.9982.340.600.44 (74.3)85.2883.270.480.38 (76.8)85.2282.031.391.65 (74.6)
ResRep+CP384.3283.680.590.43 (74.9)86.2284.320.470.37 (77.4)86.7384.231.321.54 (76.3)
ResRep83.7981.910.400.29 (83.0)83.6681.660.410.25 (84.8)85.0281.270.600.68 (89.5)
ResRep+CP384.8082.830.400.26 (84.8)85.0183.030.400.23 (86.0)86.1383.400.580.52 (92.0)
CHIP86.1484.980.820.73 (57.3)87.1385.090.700.67 (59.1)88.4587.402.112.74 (57.8)
CHIP+CP386.2584.680.800.70 (59.1)87.5485.670.690.65 (60.4)88.5886.452.052.65 (59.2)
CHIP84.4582.250.590.42 (75.4)85.2883.280.500.4 (75.6)86.6885.121.371.64 (74.7)
CHIP+CP385.5984.420.570.40 (76.6)86.2983.990.490.38 (76.8)87.7986.811.321.56 (76.0)
CHIP83.2781.150.360.15 (91.2)81.3778.990.340.19 (88.4)83.9081.830.440.32 (95.1)
CHIP+CP384.0382.220.350.14 (91.8)82.1279.410.330.18 (89.0)84.2582.170.420.29 (95.5)
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.404, + 0.88, + 0.419 + ], + "angle": 0, + "content": "Table 3. Comparisons of semantic segmentation on the S3DIS dataset (evaluated in Area-5) with PointNeXt-B and PointNeXt-L [35]." + }, + { + "type": "table", + "bbox": [ + 0.123, + 0.424, + 0.851, + 0.585 + ], + "angle": 0, + "content": "
MethodPointNeXt-BPointNeXt-L
OAmAccmIoUParams. (M)GFLOPs (↓%)OAmAccmIoUParams. (M)GFLOPs (↓%)
Baseline89.4073.9067.503.838.80 (-)90.1075.7069.307.1315.24 (-)
HRank89.0472.1465.661.724.01 (54.4)88.8873.6166.803.206.83 (55.2)
HRank+CP389.2473.7666.951.613.80 (56.8)89.4474.2767.533.006.48 (57.5)
HRank88.8172.1665.580.852.04 (76.8)88.2170.9364.301.583.47 (77.2)
HRank+CP389.0272.8266.410.781.85 (78.9)88.3571.2664.711.443.14 (79.4)
CHIP88.8973.2666.571.663.93 (55.3)89.1673.7267.093.096.68 (56.2)
CHIP+CP389.6873.1466.801.563.65 (58.5)89.4774.2067.282.916.22 (59.2)
CHIP88.6773.2466.680.811.92 (78.2)88.5871.5865.181.503.27 (78.5)
CHIP+CP389.8173.3666.950.741.79 (79.7)89.2071.6665.241.383.04 (80.1)
" + }, + { + "type": "table_caption", + "bbox": [ + 0.079, + 0.608, + 0.467, + 0.622 + ], + "angle": 0, + "content": "Table 4. Comparisons of object detection on the ScanNet dataset." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.627, + 0.466, + 0.709 + ], + "angle": 0, + "content": "
MethodmAP@0.25mAP@0.50Params. (K)GFLOPs (↓%)
Baseline (VoteNet)62.3440.82641.925.78 (−)
ResRep62.4540.95251.232.45 (57.6)
ResRep+CP363.9241.47242.262.41 (58.1)
ResRep61.7840.54180.491.83 (68.3)
ResRep+CP362.9840.94160.481.78 (69.2)
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.735, + 0.473, + 0.903 + ], + "angle": 0, + "content": "wise accuracy (OA). As the segmentation task is relatively difficult and the segmentation network structures are relatively complex, we pruned only the encoder part of the network and kept the original decoder part. The results are presented in Tab. 3. As expected, the performance of the pruned networks degraded more from the original networks than those in the classification experiments, but the performance was acceptable. Meanwhile, in all cases, with our \\(\\mathrm{CP}^3\\), PNNs have a higher accuracy at a higher pruning rate than without \\(\\mathrm{CP}^3\\). For example, for PointNeXt-B, comparing with directly applying CHIP, incorporating \\(\\mathrm{CP}^3\\)" + }, + { + "type": "table_caption", + "bbox": [ + 0.52, + 0.608, + 0.871, + 0.622 + ], + "angle": 0, + "content": "Table 5. Comparisons of object detection on SUN RGB-D." + }, + { + "type": "table", + "bbox": [ + 0.505, + 0.627, + 0.888, + 0.709 + ], + "angle": 0, + "content": "
MethodmAP@0.25mAP@0.50Params. (K)GFLOPs (↓%)
Baseline (VoteNet)59.7835.77641.925.78 (-)
ResRep59.3736.80179.912.42 (58.13)
ResRep+CP360.1037.37172.352.20 (61.93)
ResRep59.0135.91135.131.84 (68.17)
ResRep+CP359.1836.24129.641.83 (68.34)
" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.735, + 0.892, + 0.794 + ], + "angle": 0, + "content": "obtained much higher OA, mAcc and mIoU scores (1.2% on OA and 0.3% on mIoU). The results about segmentation have well validated the generalization of \\(\\mathrm{CP}^3\\) to new and difficult tasks." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.807, + 0.779, + 0.823 + ], + "angle": 0, + "content": "4.4. Results on 3D Object Detection" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.831, + 0.829, + 0.846 + ], + "angle": 0, + "content": "4.4.1 Evaluation and Comparison of VoteNet" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.894, + 0.903 + ], + "angle": 0, + "content": "Tabs. 4 and 5 show the results of the pruned VoteNet models on the ScanNetV2 and SUN RGB-D datasets, respectively. We evaluated the performance of our proposed method in" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5308" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.078, + 0.089, + 0.468, + 0.117 + ], + "angle": 0, + "content": "Table 6. Comparisons of object detection on the ScanNet dataset. The baseline PNN model is GroupFree3D." + }, + { + "type": "table", + "bbox": [ + 0.08, + 0.122, + 0.468, + 0.21 + ], + "angle": 0, + "content": "
MethodmAP@0.25mAP@0.50Params. (K)GFLOPs (↓%)
Baseline68.2252.612438.3421.78 (-)
ResRep68.2451.481910.7310.90 (49.95)
ResRep+CP368.8652.081654.3510.88 (50.05)
ResRep67.2151.281703.348.71 (60.01)
ResRep+CP368.5751.851501.468.68 (60.14)
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.234, + 0.47, + 0.264 + ], + "angle": 0, + "content": "terms of the mean average precision at IOU thresholds of 0.25 and 0.50 (mAP@25 and mAP@50)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.282, + 0.47, + 0.493 + ], + "angle": 0, + "content": "ScanNetV2 ScanNetV2 [5] is a richly annotated dataset of 3D reconstructed meshes of indoor scenes. It contains about 1200 training examples collected from hundreds of different rooms and is annotated with semantic and instance segmentation for 18 object categories. Tab. 4 shows the results of directly applying ResRep and with \\(\\mathrm{CP}^3\\). As can be seen from the table, the accuracy of the 2D method directly applied to the 3D network decreased by a flops drop of about \\(60\\%\\), while our method achieves \\(1.58\\%\\) and \\(0.65\\%\\) improvements at mAP@0.25 and mAP@0.5 with a drop rate of \\(58.13\\%\\) FLOPS. When FLOPs drop to about \\(70\\%\\), the accuracy of the direct porting CNNs pruning method works poorly, while the improvement of mAP@0.25 and mAP@0.5 of our method is \\(0.64\\%\\) and \\(0.12\\%\\), respectively." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.511, + 0.47, + 0.709 + ], + "angle": 0, + "content": "SUN RGB-D The SUN RGB-D dataset [38] consists of 10355 single-view indoor RGB-D images annotated with over 64000 3D bounding boxes and semantic labels for 37 categories. We conducted experiments on SUN RGB-D with the same setup as those on ScanNetV2. The findings are also similar to those in ScanNetV2. It can be observed from Tab. 5 that the accuracy of directly transplanted CNNs pruning method is reduced to some extent (mAP@0.25 reduced by 0.41) when FLOPs drop by \\(58.13\\%\\), while \\(\\mathrm{CP}^3\\) improved the detection model by \\(0.32\\%\\) mAP@0.25 and \\(1.60\\%\\) mAP@0.5 with a FLOPs drop of \\(61.93\\%\\). Even when FLOPs drop to \\(70\\%\\), our method's mAP@0.25 drop only \\(0.6\\%\\), which is obviously better." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.726, + 0.45, + 0.743 + ], + "angle": 0, + "content": "4.4.2 Evaluation and Comparison on GroupFree3D" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.47, + 0.902 + ], + "angle": 0, + "content": "We also conducted experiments on another point-based 3D detection model, GroupFree3D, on ScanNetV2. Tab. 6 summarizes the pruning performance of our approach for GroupFree3D on the ScanNetV2 dataset. When targeting a moderate compression ratio, our approach can achieve \\(32.15\\%\\) and \\(50.05\\%\\) storage and computation reductions, respectively, with a \\(0.64\\%\\) accuracy increase for mAP@0.25 over the baseline model. In the case of higher compression ratio, \\(\\mathrm{CP}^3\\) still achieves superior performance to other methods. Specifically, the ResRep loses accuracy" + }, + { + "type": "table_caption", + "bbox": [ + 0.498, + 0.089, + 0.892, + 0.146 + ], + "angle": 0, + "content": "Table 7. Ablation study of different components in \\(\\mathrm{CP}^3\\). Results are of classification on the ScanObjectNN dataset with PointNeXt-S (C=32) as the baseline. 'CE' represents the coordinate-enhanced module, and 'KR' represents the knowledge recycling module." + }, + { + "type": "table", + "bbox": [ + 0.534, + 0.15, + 0.86, + 0.29 + ], + "angle": 0, + "content": "
SettingCEKRPruning RateOAmAcc
Baseline-88.2086.40
HRank0.7584.7981.93
✓0.7585.6382.97
✓0.7585.1182.13
HRank+CP3✓✓0.7586.6383.63
HRank0.9081.3378.32
✓0.9083.6681.32
✓0.9083.1080.47
HRank+CP3✓✓0.9084.8382.74
" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.316, + 0.892, + 0.361 + ], + "angle": 0, + "content": "by \\(1.01\\%\\) mAP@0.25 when the parameters and flops drop by \\(30.14\\%\\) and \\(60.01\\%\\), while in our method, the accuracy increases \\(0.35\\%\\) for mAP@0.25." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.372, + 0.665, + 0.387 + ], + "angle": 0, + "content": "4.5. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.396, + 0.892, + 0.623 + ], + "angle": 0, + "content": "We conducted ablation studies to validate the Coordinate-Enhancement (CE) module and the KnowledgeRecycling (KR) module in \\(\\mathrm{CP}^3\\). All results provided in the section are tested on ScanObjectNN with PointNeXt-S \\((C = 32)\\) as the baseline and HRank as the pruning method. We evaluated the pruned networks' performance with the FLOPs drop of \\(75\\%\\) and \\(90\\%\\), and with/without the CE and KR modules to the pruning method. From Tab. 7, we find that KR improves OA of \\(0.84\\%\\) and \\(2.33\\%\\) when pruning rate is \\(75\\%\\) and \\(90\\%\\), respectively, while CE improves OA of \\(0.32\\%\\) and \\(1.77\\%\\) at \\(75\\%\\) and \\(90\\%\\) pruning rates, respectively. Bringing the two modules together, the OA improvement is \\(1.84\\%\\) and \\(3.50\\%\\) at \\(75\\%\\) and \\(90\\%\\) pruning rates, respectively. The ablation study results have validated the effectiveness of all designs in \\(\\mathrm{CP}^3\\)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.638, + 0.618, + 0.654 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.664, + 0.892, + 0.785 + ], + "angle": 0, + "content": "In this paper, we focus on 3D point-based network pruning and design a 3D channel pruning plug-in \\((\\mathrm{CP}^3)\\) that can be used with existing 2D CNN pruning methods. To the best of our knowledge, this is the first pruning work explicitly considering the characteristics of point cloud data and point-based networks. Empirically, we show that the proposed \\(\\mathrm{CP}^3\\) is universally effective for a wide range of point-based networks and 3D tasks." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.8, + 0.66, + 0.816 + ], + "angle": 0, + "content": "Acknowledgement" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.825, + 0.892, + 0.901 + ], + "angle": 0, + "content": "This work was partially supported by National Natural Science Foundation of China (NSFC) under Grant 62271203, the Science and Technology Innovation Action Plan of Shanghai under Grant 2251110540, and Shanghai Pujiang Program under Grant 21PJ1420300." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5309" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.184 + ], + "angle": 0, + "content": "[1] Iro Armeni, Ozan Sener, Amir R. Zamir, Helen Jiang, Ioannis K. Brilakis, Martin Fischer, and Silvio Savarese. 3D semantic parsing of large-scale indoor spaces. In IEEE Conference on Computer Vision and Pattern Recognition, CVPR, pages 1534-1543, 2016. 6, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.186, + 0.472, + 0.242 + ], + "angle": 0, + "content": "[2] Chen Chen, Zhe Chen, Jing Zhang, and Dacheng Tao. SASA: semantics-augmented set abstraction for point-based 3D object detection. In Thirty-Sixth AAAI Conference on Artificial Intelligence, AAAI, pages 221-229, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.243, + 0.47, + 0.297 + ], + "angle": 0, + "content": "[3] MMDetection3D Contributors. MMDetection3D: Open-MMLab next-generation platform for general 3D object detection. https://github.com/open-mmlab/mmdetection3d, 2020.6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.298, + 0.47, + 0.367 + ], + "angle": 0, + "content": "[4] Angela Dai, Angel X Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3D reconstructions of indoor scenes. In IEEE Conference on Computer Vision and Pattern Recognition, pages 5828-5839, 2017. 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.368, + 0.47, + 0.438 + ], + "angle": 0, + "content": "[5] Angela Dai, Angel X. Chang, Manolis Savva, Maciej Halber, Thomas A. Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3D reconstructions of indoor scenes. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR, pages 2432-2443, 2017. 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.439, + 0.47, + 0.508 + ], + "angle": 0, + "content": "[6] Xiaohan Ding, Guiguang Ding, Yuchen Guo, and Jungong Han. Centripetal sgd for pruning very deep convolutional networks with complicated structure. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 4943-4953, 2019. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.509, + 0.47, + 0.577 + ], + "angle": 0, + "content": "[7] Xiaohan Ding, Tianxiang Hao, Jianchao Tan, Ji Liu, Jungong Han, Yuchen Guo, and Guiguang Ding. ResRep: Lossless CNN pruning via decoupling remembering and forgetting. In IEEE/CVF International Conference on Computer Vision, ICCV, pages 4510-4520, 2021. 1, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.579, + 0.47, + 0.648 + ], + "angle": 0, + "content": "[8] Anand Dubey, Avik Santra, Jonas Fuchs, Maximilian Lübke, Robert Weigel, and Fabian Lurz. Haradnet: Anchor-free target detection for radar point clouds using hierarchical attention and multi-task learning. Machine Learning with Applications, 8:100275, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.649, + 0.47, + 0.718 + ], + "angle": 0, + "content": "[9] Hamidreza Fazlali, Yixuan Xu, Yuan Ren, and Bingbing Liu. A versatile multi-view framework for LiDAR-based 3D object detection with guidance from panoptic segmentation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR, pages 17171-17180, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.719, + 0.47, + 0.802 + ], + "angle": 0, + "content": "[10] Kehong Gong, Bingbing Li, Jianfeng Zhang, Tao Wang, Jing Huang, Michael Bi Mi, Jiashi Feng, and Xinchao Wang. PoseTriplet: Co-evolving 3d human pose estimation, imitation, and hallucination under self-supervision. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR, pages 11017-11027, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.803, + 0.469, + 0.83 + ], + "angle": 0, + "content": "[11] Google. TensorFlow lite. https://www.tensorflow.org/ lite, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.832, + 0.47, + 0.901 + ], + "angle": 0, + "content": "[12] Yushuo Guan, Ning Liu, Pengyu Zhao, Zhengping Che, Kaigui Bian, Yanzhi Wang, and Jian Tang. DAIS: Automatic channel pruning via differentiable annealing indicator search. IEEE Transactions on Neural Networks and Learning Systems, TNNLS, pages 1-12, 2022. 1, 3" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.148 + ], + "angle": 0, + "content": "[13] Kai Han, Yunhe Wang, Qiulin Zhang, Wei Zhang, Chunjing Xu, and Tong Zhang. Model rubik's cube: Twisting resolution, depth and width for tinynets. Advances in Neural Information Processing Systems, 33:19353-19364, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.149, + 0.892, + 0.204 + ], + "angle": 0, + "content": "[14] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In IEEE/CVF conference on Computer Vision and Pattern Recognition, CVPR, pages 770-778, 2016. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.205, + 0.892, + 0.286 + ], + "angle": 0, + "content": "[15] Zejiang Hou, Minghai Qin, Fei Sun, Xiaolong Ma, Kun Yuan, Yi Xu, Yen-Kuang Chen, Rong Jin, Yuan Xie, and Sun-Yuan Kung. CHEX: channel exploration for CNN model compression. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR, pages 12277-12288, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.288, + 0.892, + 0.356 + ], + "angle": 0, + "content": "[16] Andrew G. Howard, Menglong Zhu, Bo Chen, Dmitry Kalenichenko, Weijun Wang, Tobias Weyand, Marco Andretto, and Hartwig Adam. MobileNets: Efficient convolutional neural networks for mobile vision applications. CoRR, abs/1704.04861, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.358, + 0.892, + 0.413 + ], + "angle": 0, + "content": "[17] Fengwei Jia, Xuan Wang, Jian Guan, Huale Li, Chen Qiu, and Shuhan Qi. Arank: Toward specific model pruning via advantage rank for multiple salient objects detection. Image and Vision Computing, 111:104192, 2021. 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.414, + 0.892, + 0.483 + ], + "angle": 0, + "content": "[18] Xiaotang Jiang, Huan Wang, Yiliu Chen, Ziqi Wu, Lichuan Wang, Bin Zou, Yafeng Yang, Zongyang Cui, Yu Cai, Tianhang Yu, et al. MNN: A universal and efficient inference engine. The Conference on Machine Learning and Systems, MLSys, pages 1-13, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.484, + 0.892, + 0.538 + ], + "angle": 0, + "content": "[19] Hema Koppula, Abhishek Anand, Thorsten Joachims, and Ashutosh Saxena. Semantic labeling of 3d point clouds for indoor scenes. Advances in Neural Information Processing Systems, NeurIPS, pages 244-252, 2011. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.539, + 0.892, + 0.593 + ], + "angle": 0, + "content": "[20] Bastian Leibe, Ales Leonardis, and Bernt Schiele. Robust object detection with interleaved categorization and segmentation. International Journal of Computer Vision, 77(1):259-289, 2008. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.595, + 0.892, + 0.649 + ], + "angle": 0, + "content": "[21] Hao Li, Asim Kadav, Igor Durdanovic, Hanan Samet, and Hans Peter Graf. Pruning filters for efficient convnets. In International Conference on Learning Representations, ICLR, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.65, + 0.892, + 0.705 + ], + "angle": 0, + "content": "[22] Kelin Li, Nicholas Baron, Xian Zhang, and Nicolas Rojas. Efficientgrasp: A unified data-efficient learning to grasp method for multi-fingered robot hands. IEEE Robotics and Automation Letters, 7(4):8619-8626, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.706, + 0.892, + 0.761 + ], + "angle": 0, + "content": "[23] Lucas Liebenwein, Cenk Baykal, Harry Lang, Dan Feldman, and Daniela Rus. Provable filter pruning for efficient neural networks. In 8th International Conference on Learning Representations, ICLR, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.762, + 0.892, + 0.831 + ], + "angle": 0, + "content": "[24] Mingbao Lin, Rongrong Ji, Yan Wang, Yichen Zhang, Baochang Zhang, Yonghong Tian, and Ling Shao. HRank: Filter pruning using high-rank feature map. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR, pages 1529-1538, 2020. 1, 3, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.832, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[25] Liyang Liu, Shilong Zhang, Zhanghui Kuang, Aojun Zhou, Jing-Hao Xue, Xinjiang Wang, Yimin Chen, Wenming Yang, Qingmin Liao, and Wayne Zhang. Group Fisher pruning for practical network compression. In International Conference on Machine Learning, ICML, pages 7021-7032, 2021. 6" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5310" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.161 + ], + "angle": 0, + "content": "[26] Zhuang Liu, Jianguo Li, Zhiqiang Shen, Gao Huang, Shoumeng Yan, and Changshui Zhang. Learning efficient convolutional networks through network slimming. In IEEE/CVF International Conference on Computer Vision, ICCV, pages 2736-2744, 2017. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.164, + 0.47, + 0.219 + ], + "angle": 0, + "content": "[27] Ze Liu, Zheng Zhang, Yue Cao, Han Hu, and Xin Tong. Group-free 3D object detection via transformers. In IEEE/CVF International Conference on Computer Vision, CVPR, pages 2949-2958, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.221, + 0.47, + 0.276 + ], + "angle": 0, + "content": "[28] Xu Ma, Can Qin, Haoxuan You, Haoxi Ran, and Yun Fu. Rethinking network design and local geometry in point cloud: A simple residual MLP framework. In International Conference on Learning Representations, ICLR, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.278, + 0.47, + 0.333 + ], + "angle": 0, + "content": "[29] Fanxu Meng, Hao Cheng, Ke Li, Huixiang Luo, Xiaowei Guo, Guangming Lu, and Xing Sun. Pruning filter in filter. In Advances in Neural Information Processing Systems, NeurIPS, volume 33, pages 17629-17640, 2020. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.335, + 0.47, + 0.417 + ], + "angle": 0, + "content": "[30] Quang-Hieu Pham, Thanh Nguyen, Binh-Son Hua, Gemma Roig, and Sai-Kit Yeung. JSIS3D: Joint semantic-instance segmentation of 3d point clouds with multi-task pointwise networks and multi-value conditional random fields. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.42, + 0.47, + 0.475 + ], + "angle": 0, + "content": "[31] Charles R Qi, Or Litany, Kaiming He, and Leonidas J Guibas. Deep hough voting for 3D object detection in point clouds. In IEEE/CVF International Conference on Computer Vision, ICCV, 2019. 3, 6, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.477, + 0.47, + 0.546 + ], + "angle": 0, + "content": "[32] Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. Pointnet: Deep learning on point sets for 3D classification and segmentation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR, pages 652-660, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.548, + 0.47, + 0.616 + ], + "angle": 0, + "content": "[33] Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J Guibas. Pointnet++: Deep hierarchical feature learning on point sets in a metric space. In Advances in Neural Information Processing Systems, NeurIPS, pages 5100-5109, 2017. 3, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.619, + 0.47, + 0.687 + ], + "angle": 0, + "content": "[34] Guocheng Qian, Hasan Hammoud, Guohao Li, Ali Thabet, and Bernard Ghanem. ASSANet: An anisotropic separable set abstraction for efficient point cloud representation learning. In Advances in Neural Information Processing Systems, NeurIPS, pages 28119-28130, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.69, + 0.47, + 0.771 + ], + "angle": 0, + "content": "[35] Guocheng Qian, Yuchen Li, Houwen Peng, Jinjie Mai, Hasan Hammoud, Mohamed Elhoseiny, and Bernard Ghanem. PointNeXt: Revisiting pointnet++ with improved training and scaling strategies. In Advances in Neural Information Processing Systems, NeurIPS, 2022. 3, 5, 6, 7, 12, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.775, + 0.47, + 0.843 + ], + "angle": 0, + "content": "[36] Ruizhi Shao, Zerong Zheng, Hongwen Zhang, Jingxiang Sun, and Yebin Liu. DiffuStereo: High quality human reconstruction via diffusion-based stereo using sparse cameras. In European Conference on Computer Vision, ECCV, pages 702-720, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.846, + 0.47, + 0.9 + ], + "angle": 0, + "content": "[37] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. In International Conference on Learning Representations, ICLR, 2015. 1, 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.893, + 0.148 + ], + "angle": 0, + "content": "[38] Shuran Song, Samuel P Lichtenberg, and Jianxiong Xiao. Sun RGB-D: A RGB-D scene understanding benchmark suite. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR, pages 567-576, 2015. 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.151, + 0.893, + 0.206 + ], + "angle": 0, + "content": "[39] Shuran Song, Samuel P Lichtenberg, and Jianxiong Xiao. Sun RGB-D: A RGB-D scene understanding benchmark suite. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR, pages 567-576, 2015. 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.208, + 0.893, + 0.275 + ], + "angle": 0, + "content": "[40] Yang Sui, Miao Yin, Yi Xie, Huy Phan, Saman Aliari Zonouz, and Bo Yuan. CHIP: Channel independence-based pruning for compact neural networks. In Advances in Neural Information Processing Systems, NeurIPS, pages 24604-24616, 2021. 1, 3, 4, 5, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.279, + 0.893, + 0.333 + ], + "angle": 0, + "content": "[41] Lyne P. Tchapmi, Christopher B. Choy, Iro Armeni, JunYoung Gwak, and Silvio Savarese. SEGCloud: Semantic segmentation of 3d point clouds. In International Conference on 3D Vision, 3DV, pages 537-547, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.335, + 0.893, + 0.405 + ], + "angle": 0, + "content": "[42] Hugues Thomas, Charles R Qi, Jean-Emmanuel Deschaud, Beatz Marcotegui, François Goulette, and Leonidas J Guibas. KpConv: Flexible and deformable convolution for point clouds. In IEEE/CVF international conference on computer vision, CVPR, pages 6411–6420, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.407, + 0.893, + 0.489 + ], + "angle": 0, + "content": "[43] Mikaela Angelina Uy, Quang-Hieu Pham, Binh-Son Hua, Duc Thanh Nguyen, and Sai-Kit Yeung. Revisiting point cloud classification: A new benchmark dataset and classification model on real-world data. In IEEE/CVF International Conference on Computer Vision, ICCV, pages 1588–1597, 2019. 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.492, + 0.893, + 0.56 + ], + "angle": 0, + "content": "[44] Haowen Wang, Mingyuan Wang, Zhengping Che, Zhiyuan Xu, Xiuquan Qiao, Mengshi Qi, Feifei Feng, and Jian Tang. RGB-Depth fusion GAN for indoor depth completion. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR, pages 6199-6208, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.563, + 0.893, + 0.673 + ], + "angle": 0, + "content": "[45] Wenxiao Wang, Minghao Chen, Shuai Zhao, Long Chen, Jinming Hu, Haifeng Liu, Deng Cai, Xiaofei He, and Wei Liu. Accelerate cnns from three dimensions: A comprehensive pruning framework. In Marina Meila and Tong Zhang, editors, Proceedings of the 38th International Conference on Machine Learning, ICML 2021, 18-24 July 2021, Virtual Event, volume 139 of Proceedings of Machine Learning Research, pages 10717-10726. PMLR, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.676, + 0.893, + 0.771 + ], + "angle": 0, + "content": "[46] Wenxiao Wang, Cong Fu, Jishun Guo, Deng Cai, and Xiaofei He. COP: customized deep model compression via regularized correlation-based filter-level pruning. In Sarit Kraus, editor, Proceedings of the Twenty-Eighth International Joint Conference on Artificial Intelligence, IJCAI 2019, Macao, China, August 10-16, 2019, pages 3785-3791. ijcai.org, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.775, + 0.893, + 0.843 + ], + "angle": 0, + "content": "[47] Zhirong Wu, Shuran Song, Aditya Khosla, Fisher Yu, Linguuang Zhang, Xiaou Tang, and Jianxiong Xiao. 3D ShapeNets: A deep representation for volumetric shapes. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR, pages 1912-1920, 2015. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.846, + 0.893, + 0.9 + ], + "angle": 0, + "content": "[48] Chenfeng Xu, Shijia Yang, Tomer Galanti, Bichen Wu, Xiangyu Yue, Bohan Zhai, Wei Zhan, Peter Vajda, Kurt Keutzer, and Masayoshi Tomizuka. Image2point: 3d point-cloud understanding with 2d image pretrained models. In" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.893, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.514, + 0.957 + ], + "angle": 0, + "content": "5311" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.11, + 0.092, + 0.468, + 0.12 + ], + "angle": 0, + "content": "European Conference on Computer Vision, pages 638-656. Springer, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.122, + 0.47, + 0.177 + ], + "angle": 0, + "content": "[49] Wei Yang, Chris Paxton, Maya Cakmak, and Dieter Fox. Human grasp classification for reactive human-to-robot handovers. In IEEE/RSJ International Conference on Intelligent Robots and Systems, IROS, pages 11123–11130, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.179, + 0.469, + 0.233 + ], + "angle": 0, + "content": "[50] Xucheng Ye, Pengcheng Dai, Junyu Luo, Xin Guo, Yingjie Qi, Jianlei Yang, and Yiran Chen. Accelerating CNN training by pruning activation gradients. In European Conference on Computer Vision, ECCV, pages 322-338, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.235, + 0.469, + 0.317 + ], + "angle": 0, + "content": "[51] Ruichi Yu, Ang Li, Chun-Fu Chen, Jui-Hsin Lai, Vlad I Morariu, Xintong Han, Mingfei Gao, Ching-Yung Lin, and Larry S Davis. Nisp: Pruning networks using neuron importance score propagation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR, pages 9194-9203, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.32, + 0.469, + 0.374 + ], + "angle": 0, + "content": "[52] Hengshuang Zhao, Li Jiang, Jiaya Jia, Philip H. S. Torr, and Vladlen Koltun. Point transformer. In IEEE/CVF International Conference on Computer Vision, ICCV, pages 16239-16248, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.376, + 0.469, + 0.431 + ], + "angle": 0, + "content": "[53] Wu Zheng, Weiliang Tang, Li Jiang, and Chi-Wing Fu. SESSD: self-ensembling single-stage object detector from point cloud. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR, pages 14494–14503, 2021. 1" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.431 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "5312" + } + ] +] \ No newline at end of file diff --git a/2023/CP3_ Channel Pruning Plug-In for Point-Based Networks/f7e54c81-931a-42ba-b8dd-1b55f7342f80_origin.pdf b/2023/CP3_ Channel Pruning Plug-In for Point-Based Networks/f7e54c81-931a-42ba-b8dd-1b55f7342f80_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..8279df60553fa2c7d7966e36e159630ca4811a35 --- /dev/null +++ b/2023/CP3_ Channel Pruning Plug-In for Point-Based Networks/f7e54c81-931a-42ba-b8dd-1b55f7342f80_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3506312bdd67f80e0895fe0f01c1b0970b8f59c32699673fd403feff8e8b8b9a +size 332046 diff --git a/2023/CP3_ Channel Pruning Plug-In for Point-Based Networks/full.md b/2023/CP3_ Channel Pruning Plug-In for Point-Based Networks/full.md new file mode 100644 index 0000000000000000000000000000000000000000..e8983d745af43f8e3b722614db41f22b7834e7a3 --- /dev/null +++ b/2023/CP3_ Channel Pruning Plug-In for Point-Based Networks/full.md @@ -0,0 +1,284 @@ +# $\mathbf{CP}^{3}$ : Channel Pruning Plug-in for Point-based Networks + +$\mathrm{Yaxin Peng}^{3}$ Ning Liu $^{2,*}$ Zhiying Che $^{2}$ Zhiyuan Xu $^{2}$ Chaomin Shen $^{1}$ Guixu Zhang $^{1, \boxtimes}$ Xinmei Liu $^{1}$ Feifei Feng $^{2}$ Jian Tang $^{2, \boxtimes}$ + +$^{1}$ School of Computer Science, East China Normal University $^{2}$ Midea Group + +$^{3}$ Department of Mathematics, School of Science, Shanghai University + +1{51205901049,51205901078}@stu.ecnu.edu.cn 1{cmshen,gxzhang}@cs.ecnu.edu.cn + +2{liuning22, chezp, xuzy70, feifei.feng, tangjian22}@midea.com yaxin.peng@shu.edu.cn + +# Abstract + +Channel pruning can effectively reduce both computational cost and memory footprint of the original network while keeping a comparable accuracy performance. Though great success has been achieved in channel pruning for 2D image-based convolutional networks (CNNs), existing works seldom extend the channel pruning methods to 3D point-based neural networks (PNNs). Directly implementing the 2D CNN channel pruning methods to PNNs undermine the performance of PNNs because of the different representations of 2D images and 3D point clouds as well as the network architecture disparity. In this paper, we proposed $CP^3$ , which is a Channel Pruning Plugin for Point-based network. $CP^3$ is elaborately designed to leverage the characteristics of point clouds and PNNs in order to enable 2D channel pruning methods for PNNs. Specifically, it presents a coordinate-enhanced channel importance metric to reflect the correlation between dimensional information and individual channel features, and it recycles the discarded points in PNN's sampling process and reconsiders their potentially-exclusive information to enhance the robustness of channel pruning. Experiments on various PNN architectures show that $CP^3$ constantly improves state-of-the-art 2D CNN pruning approaches on different point cloud tasks. For instance, our compressed PointNeXt-S on ScanObjectNN achieves an accuracy of $88.52\%$ with a pruning rate of $57.8\%$ , outperforming the baseline pruning methods with an accuracy gain of $1.94\%$ . + +# 1. Introduction + +Convolutional Neural Networks (CNNs) often encounter the problems of overloaded computation and overweighted storage. The cumbersome instantiation of a CNN model leads to inefficient, uneconomic, or even impossible deployment in practice. Therefore, light-weight models that provide comparable results with much fewer computational costs are in great demand for nearly all applications. Channel pruning is a promising solution to delivering efficient networks. In recent years, 2D CNN channel pruning, e.g., pruning classical VGGNets [37], ResNets [14], MobileNets [16], and many other neural networks for processing 2D images [6, 7, 12, 24, 26, 29, 40], has been successfully conducted. Most channel pruning approaches focus on identifying redundant convolution filters (i.e., channels) by evaluating their importance. The cornerstone of 2D channel pruning methods is the diversified yet effective channel evaluation metrics. For instance, HRank [24] uses the rank of the feature map as the pruning metric and removes the low-rank filters that are considered to contain less information. CHIP [40] leverages channel independence to represent the importance of each feature mapping and eliminates less important channels. + +With the widespread application of depth-sensing technology, 3D vision tasks [9, 10, 36, 44] are a rapidly growing field starving for powerful methods. Apart from straightforwardly applying 2D CNNs, models built with Point-based Neural Networks (PNNs), which directly process point clouds from the beginning without unnecessary rendering, show their merits and are widely deployed on edge devices for various applications such as robots [22, 49] and self-driving [2, 53]. Compressing PNNs is crucial due to the limited resources of edge devices and multiple models for different tasks are likely to run simultaneously [8, 30]. Given the huge success of 2D channel pruning and the great + +demand for efficient 3D PNNs, we intuitively raise one question: shall we directly implement the existing pruning methods to PNNs following the proposed channel importance metrics in 2D CNNs pruning? + +With this question in mind, we investigate the fundamental factors that potentially impair 2D pruning effectiveness on PNNs. Previous works [19, 48] have shown that point clouds record visual and semantic information in a significantly different way from 2D images. Specifically, a point cloud consists of a set of unordered points on objects' and environments' surfaces, and each point encodes its features, such as intensity along with the spatial coordinates $(x,y,z)$ . In contrast, 2D images organize visual features in a dense and regular pixel array. Such data representation differences between 3D point clouds and 2D images lead to a) different ways of exploiting information from data and b) contrasting network architectures of PNNs and 2D CNNs. It is credible that only the pruning methods considering the two aspects (definitely not existing 2D CNN pruners) may obtain superior performance on PNNs. + +From the perspective of data representations, 3D point clouds provide more 3D feature representations than 2D images, but the representations are more sensitive to network channels. To be more specific, for 2D images, all three RGB channels represent basic information in an isotropic and homogeneous way so that the latent representations extracted by CNNs applied to the images. On the other hand, point clouds explicitly encode the spatial information in three coordinate channels, which are indispensable for extracting visual and semantic information from other channels. Moreover, PNNs employ the coordinate information in multiple layers as concatenated inputs for deeper feature extraction. Nevertheless, existing CNN pruning methods are designed only suitable for the plain arrangements of 2D data but fail to consider how the informative 3D information should be extracted from point clouds. + +Moreover, the network architectures of PNNs are designed substantially different from 2D CNNs. While using smaller kernels [37] is shown to benefit 2D CNNs [37], it does not apply to networks for 3D point clouds. On the contrary, PNNs leverage neighborhoods at multiple scales to obtain both robust and detailed features. The reason is that small neighborhoods (analogous to small kernels in 2D CNNs) in point clouds consist of few points for PNNs to capture robust features. Due to the necessary sampling steps, the knowledge insufficiency issue becomes more severe for deeper PNN layers. In addition, PNNs use the random input dropout procedure during training to adaptively weight patterns detected at different scales and combine multi-scale features. This procedure randomly discards a large proportion of points and loses much exclusive information of the original data. Thus, the architecture disparity between 2D CNNs and PNNs affects the performance of + +directly applying existing pruning methods to PNNs. + +In this paper, by explicitly dealing with the two characteristics of 3D task, namely the data representation and the PNN architecture design, we propose a Channel Pruning Plug-in for Point-based network named $\mathrm{CP}^3$ , which can be applied to most 2D channel pruning methods for compressing PNN models. The proposed $\mathrm{CP}^3$ refines the channel importance, the key factor of pruning methods, from two aspects. Firstly, considering the point coordinates $(x, y, \text{and } z)$ encode the spatial information and deeply affects feature extraction procedures in PNN layers, we determine the channel importance by evaluating the correlation between the feature map and its corresponding point coordinates by introducing a coordinate-enhancement module. Secondly, calculating channel importance in channel pruning is data-driven and sensitive to the input, and the intrinsic sampling steps in PNN naturally makes pruning methods unstable. To settle this problem, we make full use of the discarded points in the sampling process via a knowledge recycling module to supplement the evaluation of channel importance. This reduces the data sampling bias impact on the channel importance calculation and increases the robustness of the pruning results. Notably, both the coordinates and recycled points in $\mathrm{CP}^3$ do not participate in network training (with back-propagation) but only assist channel importance calculation in the reasoning phase. Thus, $\mathrm{CP}^3$ does not increase any computational cost of the pruned network. The contributions of this paper are as follows: + +- We systematically consider the characteristics of PNNs and propose a channel pruning plug-in named $\mathrm{CP}^3$ to enhance 2D CNN channel pruning approaches on 3D PNNs. To the best of our knowledge, $\mathrm{CP}^3$ is the first method to export existing 2D pruning methods to PNNs. +- We propose a coordinate-enhanced channel importance score to guide point clouds network pruning, by evaluating the correlation between feature maps and corresponding point coordinates. +- We design a knowledge recycling pruning scheme that increases the robustness of the pruning procedure, using the discarded points to improve the channel importance evaluation. +- We show that using $\mathrm{CP}^3$ is consistently superior to directly transplanting 2D pruning methods to PNNs by extensive experiments on three 3D tasks and five datasets with different PNN models and pruning baselines. + +# 2. Related Work + +# 2.1. 2D Channel Pruning + +Channel pruning (a.k.a., filter pruning) methods reduce the redundant filters while maintaining the original structure of CNNs and is friendly to prevailing inference acceleration engines such as TensorFlow-Lite (TFLite) [11] and + +Mobile Neural Network (MNN) [18]. Mainstream channel pruning methods [6, 7, 12, 29] usually first evaluate the importance of channels by certain metrics and then prune (i.e., remove) the less important channels. Early work [21] uses the $l_{1}$ norm of filters as importance score for channel pruning. Afterwards, learning parameters, such as the scaling factor $\gamma$ in the batch norm layer [26] and the reconstruction error in the final network layer [51], are considered as the importance scores for channel selection. The importance sampling distribution of channels [23] is also used for pruning. Recent works [15, 40] measure the correlation of multiple feature maps to determine the importance score of the filter for pruning. HRank [24] proposes a method for pruning filters based on the theory that low-rank feature maps contain less information. [50] leverages the statistical distribution of activation gradient and takes the smaller gradient as low importance score for pruning. [46] calculates the average importance of both the input feature maps and their corresponding output feature maps to determine the overall importance. [13, 45] compress CNNs from multiple dimensions While most channel pruning methods are designed for and tested on 2D CNNs, our $\mathrm{CP}^3$ can work in tandem with existing pruners for 3D point-based networks. + +# 2.2. Point-based Networks for Point Cloud Data + +Point-based Neural Networks (PNNs) directly process point cloud data with a flexible range of receptive field, have no positioning information loss, and thus keep more accurate spatial information. As a pioneer work, PointNet [32] learns the spatial encoding directly from the input point clouds and uses the characteristics of all points to obtain the global representations. PointNet++ [33] further proposes a multi-level feature extraction structure to extract local and global features more effectively. KPCov [42] proposes a new point convolution operation to learn local movements applied to kernel points. ASSANet [34] proposes a separable set abstraction module that decomposes the normal SA module in PointNet++ into two separate learning phases for channel and space. PointMLP [28] uses residual point blocks to extract local features, transforms local points using geometric affine modules, and extracts geometric features before and after the aggregation operation. PointNeXt [35] uses inverted residual bottleneck and separable multilayer perceptrons to achieve more efficient model scaling. Besides classification, PNNs also serve as backbones for other 3D tasks. VoteNet [31] effectively improves the 3D object detection accuracy through the Hough voting mechanism [20]. PointTransformer [52] designs models improving prior work across domains and tasks. GroupFree3D [27] uses the attention mechanism to automatically learn the contribution of each point to the object. In this paper, we show that $\mathrm{CP}^3$ can be widely applied to point-based networks on a variety of point cloud bench + +marks and representative original networks. + +# 3. Methodology + +Although point-based networks are similar to CNN in concrete realization, they have fundamental differences in data representation and network architecture design. To extend the success of CNN pruning to PNN, two modules are proposed in $\mathrm{CP}^3$ taking advantage from the dimensional information and discarded points: 1) coordinate-enhancement (CE) module, which produces a coordinate-enhanced score to estimate the channel importance by combining dimensional and feature information, and 2) knowledge recycling module reusing the discarded points to improve the channel importance evaluation criteria and increase the robustness. + +# 3.1. Formulations and Motivation + +Point-based networks PNN is a unified architecture that directly takes point clouds as input. It builds hierarchical groups of points and progressively abstracts larger local regions along the hierarchy. PNN is structurally composed by a number of set abstraction (SA) blocks. Each SA block consists of 1) a sampling layer iteratively samples the farthest point to choose a subset of points from input points, 2) a group layer gathers neighbors of centroid points to a local region, 3) a set of shared Multi-Layer Perceptrons (MLPs) to extract features, and 4) a reduction layer to aggregate features in the neighbors. Formally speaking, a SA block takes an $n_{i-1} \times (d + c_{i-1})$ matrix as input that is from $n_{i-1}$ points with $d$ -dim coordinates and $c_{i-1}$ -dim point feature. It outputs an $n_i \times (d + c_i)$ matrix of $n_i$ subsampled points with $d$ -dimensional coordinates (i.e., $d = 3$ ) and new $c_i$ -dimensional feature vectors summarizing local context. The SA block is formulated as: + +$$ +\mathbf {F} _ {i} ^ {l + 1} = \mathcal {R} \left\{h _ {\Theta} \left(\left[ \mathbf {F} _ {j} ^ {l}; \mathbf {x} _ {j} ^ {l} - \mathbf {x} _ {i} ^ {l} \right]\right) \right\}, \tag {1} +$$ + +where $h_\Theta$ is MLPs to extract grouped points feature, $\mathcal{R}$ is the reduction layer (e.g. max-pooling) to aggregate features in the neighbors $\{j : (i,j) \in \mathcal{N}\}$ , $F_j^l$ is the features of neighbor $j$ in the $l$ -th layer, $x_i^l$ and $x_j^l$ are input points coordinates and coordinates of neighbor $j$ in the $l$ -th layer. + +Channel pruning Assume a pre-trained PNN model has a set of $K$ convolutional layers, and $\mathcal{A}^l$ is the $l$ -th convolution layer. The parameters in $\mathcal{A}^l$ can be represented as a set of filters $\mathcal{W}_{\mathcal{A}^l} = \{\mathbf{w}_1^l,\mathbf{w}_2^l,\dots ,\mathbf{w}_{c_l}^l\} \in \mathbb{R}^{(d + c_l)\times (d + c_{l - 1})\times k_l\times k_l}$ where $j$ -th filter is $\mathbf{w}_j^l\in \mathbb{R}^{(d + c_{l - 1})\times k_l\times k_j}$ . $(d + c_{l})$ represents the number of filters in $\mathcal{A}^l$ and $k_{l}$ denotes the kernel size. The outputs of filter, i.e., feature map, are denoted as $\mathcal{F}^l = \{\mathbf{f}_1^l,\mathbf{f}_2^l,\ldots ,\mathbf{f}_{n_i}^l\} \in \mathbb{R}^{n_i\times (d + c_i)}$ . Channel pruning aims to identify and remove the less importance filter from the + +![](images/625020ddf58bb812d6da62b4428f393d1ffe2051d1ea2a09bcce9be253eb92d9.jpg) +Figure 1. The framework of $\mathrm{CP}^3$ . The figure shows the specific pruning process of one of the SA blocks. Whether a channel in a PNN is pruned is determined by three parts: 1) Original channel importance: obtained from the original CNNs channel pruning method (e.g., HRank [24], CHIP [40]). 2) Discarded channel importance: obtained from the Knowledge-Recycling module by leveraging the discarded points in the network to supplement the channel importance evaluation of the corresponding points and improve the robustness of the channel selection. 3) CE (Coordinate-Enhanced) channel importance: obtained from calculating the correlation between the feature map and its corresponding points coordinates to guide point clouds network pruning. + +original networks. In general, channel pruning can be formulated as the following optimization problem: + +$$ +\min _ {\delta_ {i j}} \sum_ {i = 1} ^ {K} \sum_ {j = 1} ^ {n _ {i}} \delta_ {i j} \mathcal {L} \left(\mathbf {w} _ {j} ^ {i}\right), \text {s . t .} \sum_ {j = 1} ^ {n _ {i}} \delta_ {i j} = k _ {l}, \tag {2} +$$ + +where $\delta_{ij}$ is an indicator which is 1 if $\mathbf{w}_j^i$ is to be pruned or 0 if $\mathbf{w}_j^i$ is to be kept, $\mathcal{L}(\cdot)$ measures the importance of a filter and $k_l$ is the kept filter number. + +Robust importance metric for channel pruning The metrics for evaluating the importance of filters is critical. Existing CNN pruning methods design a variety of $\mathcal{L}(\cdot)$ on the filters. Consider the feature maps, contain rich and important information of both filter and input data, approaches using feature information have become popular and achieved state-of-the-art performance for channel pruning. The results of the feature maps may vary depending on the variability of the input data. Therefore, when the importance of one filter solely depends on the information represented by its own generated feature map, the measurement of the importance may be unstable and sensitive to the slight change of input data. So we have taken into account the characteristics of point clouds data and point-based networks architecture to improve the robustness of channel importance in point-based networks. On the one hand, we propose a coordinate-enhancement module by evaluating the correlation between the feature map and its corresponding points coordinates to guide point clouds network pruning, which will be described in Sec. 3.2. On the other hand, we design a knowledge recycling pruning schema, using discarded points to improve the channel importance evaluation criteria and increase the robustness of the pruning module, which will be described in detail in Sec. 3.3. + +# 3.2. Coordinate-Enhanced Channel Importance + +Dimensional information is critical in PNNs. The dimensional information (i.e., coordinates of the points) are usually adopted as input for feature extraction. Namely, the input and output of each SA block are concatenated with the coordinates of the points. Meanwhile, the intermediate feature maps reflect not only the information of the original input data but also the corresponding channel information. Therefore, the importance of the channel can be obtained from the feature maps, i.e., the importance of the corresponding channel. The dimensional information is crucial in point-based tasks and should be considered as part of importance metric. Thus the critical problem falls in designing a function that can well reflect the dimensional information richness of feature maps. The feature map, obtained by encoding points spatial $x$ , $y$ , and $z$ coordinates, should be closely related to the original corresponding points coordinates. Therefore, we use the correlation between the current feature map and the corresponding input points coordinates to determine the importance of the filter. The designed Coordinate-Enhancement (CE) module based on Eq. (2): + +$$ +\min _ {\delta_ {i j}} \sum_ {i = 1} ^ {K} \sum_ {j = 1} ^ {n _ {i}} \delta_ {i j} \mathcal {L} _ {c} \left(\mathbf {F} _ {j} ^ {i}\right), \text {s . t .} \sum_ {j = 1} ^ {n _ {i}} \delta_ {i j} = k _ {l}, \tag {3} +$$ + +where $\delta_{ij}$ is an indicator which is 1 if $\mathbf{w}_j^i$ is to be pruned or 0 if $\mathbf{w}_j^i$ is to be kept and $k_l$ is the kept channel number. $\mathcal{L}_c(\cdot)$ measures the importance of a channel from take account of the relationship between feature map and points coordinates which can be formulated as: + +$$ +\mathcal {L} _ {c} \left(F _ {j} ^ {i}\right) = \mathcal {M} \left\{\mathrm {C E} \left(F _ {j} ^ {i}, x ^ {i}\right) \right\}, \tag {4} +$$ + +where CE obtains the coordinate-enhanced score by calculating correlation of each channel in the feature map with + +![](images/d6e02e42d971152899bb8eb6c23bfaed2cbd9531ba79ab85d19eea7ba094a7c6.jpg) +Coordinate-Enhancing for Filter Importance +Figure 2. Comparisons on the pruned model accuracy with different pruning metrics with CE scores. Results are on ModelNet40 with PointNeXt-S (C=64). + +the original coordinates, and $\mathcal{M}$ takes the maximum value. Hence, higher coordinate-enhanced score (i.e., $\mathcal{L}_c$ ) serve as a reliable measurement for information richness. + +We evaluated the effectiveness of the CE module on ModelNet40 with PointNeXt-S (C=64). To demonstrate the experiment's validity, we compared the overall accuracy (OA) for pruning rates of $40\%$ , $50\%$ , $60\%$ , and $70\%$ , respectively. Three sets of experiments are carried out. First, the 'CE' group selected filters in order of value from the CE module. Secondly, the 'Random' group is used to randomly select filters for pruning, and finally, the 'Reverse' is used to select filters according to the coordinate-enhanced score from low to high. As shown in Fig. 2, the channel with a higher coordinate-enhanced score has higher accuracy, which means that the channel with a higher coordinate-enhanced score has higher importance and should be retained in the pruning process. + +# 3.3. Knowledge Recycling + +Sec. 3.2 shows that feature maps can reflect the importance of the corresponding channels. Another problem is that the importance determination may be unstable and sensitive to small changes in the input data as the feature maps are highly related to the samples of input data. Therefore, we aim to reduce the impact of such data variation. Through the analysis of the PNNs in Sec. 3.1, we found that some points are discarded to obtain hierarchical points set feature. These discarded points are informative as well since the sampling mechanism are highly random and can be leveraged to reduce the impact of data variation. The Knowledge Recycling (KR) module is proposed to reuse the discarded points to improve the robustness of channel pruning. + +For those centroids that are computed in $(l - 1)$ -th layer but discarded in $l$ -th layer due to sampling, which are equivalent to the sampled points. Therefore, the discarded centroids $x_{dis}$ are fed into $l$ -th convolutional layer to generate the feature map $\mathbf{f}_{\mathrm{dis}}^l$ , and $\mathcal{L}_k(F_{dis_j}^i)$ is taken in use for the evaluation of channel importance. + +We calculate the relevant feature maps from the network parameters trained from the sampled points and use them as + +part of the importance calculation for the current SA layer channels. Specifically, for each layer in the SA module, we obtain the features of the discard points by Eq. 1: + +$$ +\mathcal {F} _ {d i s} ^ {l} = \left\{\mathbf {f} _ {\mathbf {d i s} _ {1}} ^ {l}, \mathbf {f} _ {\mathbf {d i s} _ {2}} ^ {l}, \dots , \mathbf {f} _ {\mathbf {d i s} n _ {i}} ^ {l} \right\} \in \mathbb {R} ^ {n _ {i} \times (3 + c _ {i})}, +$$ + +where $n_i$ is sampled points number, $c_i$ is the points feature dimension. So the supplement importance is: + +$$ +\begin{array}{l} \mathcal {L} _ {k} \left(F _ {j} ^ {i}\right) = \mathcal {L} _ {k} \left(F _ {\text {d i s} j} ^ {i}\right) \tag {5} \\ = \mathcal {M} \left\{\mathrm {C E} \left(F _ {j} ^ {i}, x _ {d i s} ^ {i}\right) \right\}, \\ \end{array} +$$ + +where $x_{dis}^{i}$ are discard points for recycling. + +It should be noted that the KR module only needs to calculate $\mathcal{F}_{dis}^{l}$ from the parameters trained by the sampled points and does not incur much additional overhead. + +# 3.4. Using $\mathbf{CP}^3$ in Pruning Methods + +The overall $\mathrm{CP}^3$ improve the existing CNNs pruning methods by considering the input data of PNNs and the PNN structure in Sec. 3.2 and Sec. 3.3, respectively. In fact, $\mathrm{CP}^3$ can complement the existing pruning methods, i.e., as a plug-in to the existing pruning methods, to improve the pruning performance on PNNs. Specifically, combining Eq. (4) and Eq. (5), we obtain the final pruning formula according to Eq. (2): + +$$ +\begin{array}{l} \min _ {\delta_ {i j}} \sum_ {i = 1} ^ {K} \sum_ {j = 1} ^ {n _ {i}} \delta_ {i j} \left(\mathcal {L} \left(\mathbf {w} _ {j} ^ {i}\right) + \mathcal {L} _ {c} \left(F _ {j} ^ {i}\right) + \mathcal {L} _ {k} \left(F _ {j} ^ {i}\right)\right), \tag {6} \\ \text {s . t .} \sum_ {j = 1} ^ {n _ {i}} \delta_ {i j} = k _ {l}, \\ \end{array} +$$ + +where $\delta_{ij}$ is an indicator which is 1 if $\mathbf{w}_j^i$ is to be pruned or 0 if $\mathbf{w}_j^i$ is to be kept, $\mathcal{L}(\cdot)$ is original CNNs pruning method measure importance of a channel, $\mathcal{L}_c$ and $\mathcal{L}_k$ are coordinate-enhanced score and knowledge-recycling score, $k_l$ is the kept channel number. + +# 4. Experiments + +# 4.1. Experimental Settings + +Baseline models and datasets To demonstrate the effectiveness and generality of the proposed $\mathrm{CP}^3$ , we tested it on three different 3D tasks and five datasets with various PNNs and three recent advanced channel pruning methods. The evaluated pruning methods include HRank (2020) [24], ResRep (2021) [7], and CHIP (2021) [40]. For the classification task, we chose the classical PointNet++ [33] and PointNeXt [35] models as the original networks and conducted experiments on ModelNet40 [47] and ScanObjectNN [43]. Specifically, for PointNeXt-S we tested + +Table 1. Comparisons of classification on the ModelNet40 [47] test set with PointNet++ [33], PointNeXt-S (C=32) [35], and PointNeXt-S (C=64). For PointNeXt-S (C=32), we report the baseline results from the original paper. For PointNet++ and PointNeXt-S (C=64), we report the baseline results obtained by OpenPoints [35] re-implementations trained with the improved strategies. + +
MethodPointNet++PointNeXt-S (C=32)PointNeXt-S (C=64)
OAmAccParams. (M)GFLOPs (↓%)OAmAccParams. (M)GFLOPs (↓%)OAmAccParams. (M)GFLOPs (↓%)
Baseline92.8089.901.471.71 (−)92.9989.61.371.64 (−)93.4491.054.526.49 (−)
HRank92.5989.830.860.77 (55.0)92.8789.970.740.71 (56.7)92.2389.812.122.69 (56.7)
HRank +CP392.9589.910.840.75 (56.1)93.2390.560.710.67 (59.1)93.5290.332.012.58 (59.1)
HRank91.7988.820.590.42 (75.4)92.6389.120.500.39 (76.2)92.7190.451.331.56 (76.2)
HRank +CP392.5488.520.570.39 (77.2)93.0390.920.490.38 (76.8)93.0790.551.281.50 (76.8)
HRank91.3488.180.360.15 (91.2)92.7389.980.430.29 (82.3)92.8389.990.740.71 (82.3)
HRank +CP391.7188.680.340.13 (92.4)92.9990.110.400.27 (83.5)93.1190.520.710.67 (83.5)
ResRep92.7190.390.850.73 (57.3)92.8390.150.810.69 (57.9)91.6188.422.082.34 (57.9)
ResRep+CP393.2790.480.820.70 (59.1)93.3590.930.790.67 (59.1)92.9390.661.892.02 (59.1)
ResRep92.5089.250.570.41 (76.0)92.6490.010.510.40 (75.6)91.6789.271.131.92 (75.6)
ResRep+CP392.4689.430.540.40 (76.6)93.4190.870.490.38 (76.8)93.1190.821.021.89 (76.8)
ResRep92.1189.000.550.24 (86.0))92.3089.310.340.21 (87.2)89.5488.540.720.69 (87.2)
ResRep+CP392.4889.210.580.21 (87.7)92.9590.700.330.18 (89.0)91.0289.820.690.65 (89.0)
CHIP92.7989.230.820.73 (57.3)93.1190.270.710.67 (59.1)93.0390.601.481.44 (59.1)
CHIP+CP392.9990.660.810.70 (59.1)93.3590.800.690.65 (60.4)93.3591.111.451.40 (60.4)
CHIP92.4589.190.570.39 (77.2)92.7189.900.490.38 (76.8)92.7990.390.920.76 (76.8)
CHIP+CP392.9189.650.540.35 (79.5)93.0390.510.480.37 (77.4)93.2390.300.890.74 (77.4)
CHIP92.2689.560.360.15 (91.2)92.4289.130.320.16 (90.2)90.8388.700.650.46 (90.2)
CHIP+CP392.7190.410.340.13 (92.4)92.5090.350.300.14 (91.5)92.8790.250.630.44 (91.5)
+ +two settings with widths of 32 and 64. For the segmentation task, we conducted experiments on S3DIS [1] with PointNeXt-B and PointNeXt-L [35] as the original PNNs. For the object detection task, we pruned two point-based detectors (VoteNet [31] and GroupFree3D [25]) on SUN RGB-D [38] and ScanNetV2 [5]. + +Implementation details We conducted the classification and segmentation experiments with OpenPoints [35] and the object detection experiments with MMdetection3D [3], all on NVIDIA P100 GPUs. For a fair comparison, we used the same hyperparameter settings for each group of experiments. We either 1) measured the parameter/FLOP reductions of the pruned networks with similar performance or 2) measured the performance of the pruned networks with a similar amount of parameter/FLOP reductions. For all experiments, we reported the number of FLOPs ('GFLOPs') and parameters ('Params.') as well as task-specific metrics to be described in each experiment. More experimental results are available in the supplementary. + +# 4.2. Results on Classification + +ModelNet40 ModelNet40 [47] contains 9843 training and 2468 testing meshed CAD models belonging to 40 categories. Following the standard practice [33], we report the class-average accuracy (mAcc) and the overall accuracy (OA) on the testing set. We compared the pruned networks directly by HRank, ResRep, and CHIP and those pruned by the three pruning methods with $\mathrm{CP}^3$ . As shown in Tab. 1, our $\mathrm{CP}^3$ improved the performance of existing CNNs pruning methods for different PNNs with various pruning + +rates. $\mathrm{CP}^3$ had higher accuracy scores with a similar (and mostly higher) pruning rates. Notably, with the pruning rate of $58\%$ , $\mathrm{CP}^3$ usually produced compact PNNs with even better accuracy scores than the original PNNs, which is difficult for pruning methods without $\mathrm{CP}^3$ . + +ScanObjectNN We also conducted experiments on the ScanObjectNN benchmark [43]. ScanObjectNN contains 15000 objects categorized into 15 classes with 2902 unique object instances in the real world. As reported in Tab. 2, $\mathrm{CP}^3$ surpassed existing CNN pruning methods directly applied to PNNs. For example, comparing with the baseline pruning method HRank, $\mathrm{CP}^3$ boosts the OA score of PointNet++ by $1.14\%$ (85.01→86.15), $1.14\%$ (84.94→86.08), and $0.59\%$ (84.03→84.62) for the three different pruning rates. Similarly, $\mathrm{CP}^3$ obtains much higher OA and mAcc scores than the three baselines with different pruning rates on PointNet++ and PointNeXt-S. With the extensive experimental results on classification tasks, we show that $\mathrm{CP}^3$ surely improved the pruned network's performance to a distinct extent compared to direct applications of 2D CNN pruning methods. + +# 4.3. Results on Semantic Segmentation + +S3DIS [1] is a challenging benchmark composed of 6 large-scale indoor areas, 271 rooms, and 13 semantic categories in total. Following a common protocol [41], we evaluated the presented approach in Area-5, which means to test on Area-5 and to train on the rest. For evaluation metrics, we used the mean classwise intersection over union (mIoU), the mean classwise accuracy (mAcc), and the overall point + +Table 2. Comparisons of classification on the ScanObjectNN [43] test set with PointNet++ [33], PointNeXt-S (C=32) [35], and PointNeXt-S (C=64). For PointNeXt-S (C=32), we report the baseline results from the original paper. For PointNet++ and PointNeXt-S (C=64), we report the baseline results obtained by OpenPoints [35] re-implementations trained with the improved strategies. + +
MethodPointNet++PointNeXt-S (C=32)PointNeXt-S (C=64)
OAmAccParams. (M)GFLOPs (↓%)OAmAccParams. (M)GFLOPs (↓%)OAmAccParams. (M)GFLOPs (↓%)
Baseline86.2084.401.471.71 (−)87.4085.391.371.64 (−)88.2086.844.526.49 (−)
HRank85.0183.330.820.73 (57.3)87.0285.850.720.69 (57.9)87.5185.482.222.91 (55.2)
HRank+CP386.1584.400.800.70 (59.1)87.4786.210.700.67 (59.1)87.9586.022.162.82 (56.5)
HRank84.9483.430.590.43 (74.9)84.7981.930.500.39 (76.2)84.6681.001.331.56 (76.0)
HRank+CP386.0884.510.590.41 (76.0)86.4083.940.480.37 (77.4)86.4384.941.281.50 (76.9)
HRank84.0382.160.370.16 (90.6)81.3378.320.320.17 (89.6)85.3983.830.730.71 (89.1)
HRank+CP384.6282.290.360.15 (91.2)84.2182.130.310.16 (90.2)86.2684.360.700.67 (89.7)
ResRep86.7784.780.820.76 (55.6)86.3284.450.690.65 (60.4)86.5884.022.122.82 (56.5)
ResRep+CP387.1185.500.820.71 (58.5)86.6684.750.670.64 (61.0)88.5286.002.032.74 (57.8)
ResRep83.9982.340.600.44 (74.3)85.2883.270.480.38 (76.8)85.2282.031.391.65 (74.6)
ResRep+CP384.3283.680.590.43 (74.9)86.2284.320.470.37 (77.4)86.7384.231.321.54 (76.3)
ResRep83.7981.910.400.29 (83.0)83.6681.660.410.25 (84.8)85.0281.270.600.68 (89.5)
ResRep+CP384.8082.830.400.26 (84.8)85.0183.030.400.23 (86.0)86.1383.400.580.52 (92.0)
CHIP86.1484.980.820.73 (57.3)87.1385.090.700.67 (59.1)88.4587.402.112.74 (57.8)
CHIP+CP386.2584.680.800.70 (59.1)87.5485.670.690.65 (60.4)88.5886.452.052.65 (59.2)
CHIP84.4582.250.590.42 (75.4)85.2883.280.500.4 (75.6)86.6885.121.371.64 (74.7)
CHIP+CP385.5984.420.570.40 (76.6)86.2983.990.490.38 (76.8)87.7986.811.321.56 (76.0)
CHIP83.2781.150.360.15 (91.2)81.3778.990.340.19 (88.4)83.9081.830.440.32 (95.1)
CHIP+CP384.0382.220.350.14 (91.8)82.1279.410.330.18 (89.0)84.2582.170.420.29 (95.5)
+ +Table 3. Comparisons of semantic segmentation on the S3DIS dataset (evaluated in Area-5) with PointNeXt-B and PointNeXt-L [35]. + +
MethodPointNeXt-BPointNeXt-L
OAmAccmIoUParams. (M)GFLOPs (↓%)OAmAccmIoUParams. (M)GFLOPs (↓%)
Baseline89.4073.9067.503.838.80 (-)90.1075.7069.307.1315.24 (-)
HRank89.0472.1465.661.724.01 (54.4)88.8873.6166.803.206.83 (55.2)
HRank+CP389.2473.7666.951.613.80 (56.8)89.4474.2767.533.006.48 (57.5)
HRank88.8172.1665.580.852.04 (76.8)88.2170.9364.301.583.47 (77.2)
HRank+CP389.0272.8266.410.781.85 (78.9)88.3571.2664.711.443.14 (79.4)
CHIP88.8973.2666.571.663.93 (55.3)89.1673.7267.093.096.68 (56.2)
CHIP+CP389.6873.1466.801.563.65 (58.5)89.4774.2067.282.916.22 (59.2)
CHIP88.6773.2466.680.811.92 (78.2)88.5871.5865.181.503.27 (78.5)
CHIP+CP389.8173.3666.950.741.79 (79.7)89.2071.6665.241.383.04 (80.1)
+ +Table 4. Comparisons of object detection on the ScanNet dataset. + +
MethodmAP@0.25mAP@0.50Params. (K)GFLOPs (↓%)
Baseline (VoteNet)62.3440.82641.925.78 (−)
ResRep62.4540.95251.232.45 (57.6)
ResRep+CP363.9241.47242.262.41 (58.1)
ResRep61.7840.54180.491.83 (68.3)
ResRep+CP362.9840.94160.481.78 (69.2)
+ +wise accuracy (OA). As the segmentation task is relatively difficult and the segmentation network structures are relatively complex, we pruned only the encoder part of the network and kept the original decoder part. The results are presented in Tab. 3. As expected, the performance of the pruned networks degraded more from the original networks than those in the classification experiments, but the performance was acceptable. Meanwhile, in all cases, with our $\mathrm{CP}^3$ , PNNs have a higher accuracy at a higher pruning rate than without $\mathrm{CP}^3$ . For example, for PointNeXt-B, comparing with directly applying CHIP, incorporating $\mathrm{CP}^3$ + +Table 5. Comparisons of object detection on SUN RGB-D. + +
MethodmAP@0.25mAP@0.50Params. (K)GFLOPs (↓%)
Baseline (VoteNet)59.7835.77641.925.78 (-)
ResRep59.3736.80179.912.42 (58.13)
ResRep+CP360.1037.37172.352.20 (61.93)
ResRep59.0135.91135.131.84 (68.17)
ResRep+CP359.1836.24129.641.83 (68.34)
+ +obtained much higher OA, mAcc and mIoU scores (1.2% on OA and 0.3% on mIoU). The results about segmentation have well validated the generalization of $\mathrm{CP}^3$ to new and difficult tasks. + +# 4.4. Results on 3D Object Detection + +# 4.4.1 Evaluation and Comparison of VoteNet + +Tabs. 4 and 5 show the results of the pruned VoteNet models on the ScanNetV2 and SUN RGB-D datasets, respectively. We evaluated the performance of our proposed method in + +Table 6. Comparisons of object detection on the ScanNet dataset. The baseline PNN model is GroupFree3D. + +
MethodmAP@0.25mAP@0.50Params. (K)GFLOPs (↓%)
Baseline68.2252.612438.3421.78 (-)
ResRep68.2451.481910.7310.90 (49.95)
ResRep+CP368.8652.081654.3510.88 (50.05)
ResRep67.2151.281703.348.71 (60.01)
ResRep+CP368.5751.851501.468.68 (60.14)
+ +terms of the mean average precision at IOU thresholds of 0.25 and 0.50 (mAP@25 and mAP@50). + +ScanNetV2 ScanNetV2 [5] is a richly annotated dataset of 3D reconstructed meshes of indoor scenes. It contains about 1200 training examples collected from hundreds of different rooms and is annotated with semantic and instance segmentation for 18 object categories. Tab. 4 shows the results of directly applying ResRep and with $\mathrm{CP}^3$ . As can be seen from the table, the accuracy of the 2D method directly applied to the 3D network decreased by a flops drop of about $60\%$ , while our method achieves $1.58\%$ and $0.65\%$ improvements at mAP@0.25 and mAP@0.5 with a drop rate of $58.13\%$ FLOPS. When FLOPs drop to about $70\%$ , the accuracy of the direct porting CNNs pruning method works poorly, while the improvement of mAP@0.25 and mAP@0.5 of our method is $0.64\%$ and $0.12\%$ , respectively. + +SUN RGB-D The SUN RGB-D dataset [38] consists of 10355 single-view indoor RGB-D images annotated with over 64000 3D bounding boxes and semantic labels for 37 categories. We conducted experiments on SUN RGB-D with the same setup as those on ScanNetV2. The findings are also similar to those in ScanNetV2. It can be observed from Tab. 5 that the accuracy of directly transplanted CNNs pruning method is reduced to some extent (mAP@0.25 reduced by 0.41) when FLOPs drop by $58.13\%$ , while $\mathrm{CP}^3$ improved the detection model by $0.32\%$ mAP@0.25 and $1.60\%$ mAP@0.5 with a FLOPs drop of $61.93\%$ . Even when FLOPs drop to $70\%$ , our method's mAP@0.25 drop only $0.6\%$ , which is obviously better. + +# 4.4.2 Evaluation and Comparison on GroupFree3D + +We also conducted experiments on another point-based 3D detection model, GroupFree3D, on ScanNetV2. Tab. 6 summarizes the pruning performance of our approach for GroupFree3D on the ScanNetV2 dataset. When targeting a moderate compression ratio, our approach can achieve $32.15\%$ and $50.05\%$ storage and computation reductions, respectively, with a $0.64\%$ accuracy increase for mAP@0.25 over the baseline model. In the case of higher compression ratio, $\mathrm{CP}^3$ still achieves superior performance to other methods. Specifically, the ResRep loses accuracy + +Table 7. Ablation study of different components in $\mathrm{CP}^3$ . Results are of classification on the ScanObjectNN dataset with PointNeXt-S (C=32) as the baseline. 'CE' represents the coordinate-enhanced module, and 'KR' represents the knowledge recycling module. + +
SettingCEKRPruning RateOAmAcc
Baseline-88.2086.40
HRank0.7584.7981.93
✓0.7585.6382.97
✓0.7585.1182.13
HRank+CP3✓✓0.7586.6383.63
HRank0.9081.3378.32
✓0.9083.6681.32
✓0.9083.1080.47
HRank+CP3✓✓0.9084.8382.74
+ +by $1.01\%$ mAP@0.25 when the parameters and flops drop by $30.14\%$ and $60.01\%$ , while in our method, the accuracy increases $0.35\%$ for mAP@0.25. + +# 4.5. Ablation Studies + +We conducted ablation studies to validate the Coordinate-Enhancement (CE) module and the KnowledgeRecycling (KR) module in $\mathrm{CP}^3$ . All results provided in the section are tested on ScanObjectNN with PointNeXt-S $(C = 32)$ as the baseline and HRank as the pruning method. We evaluated the pruned networks' performance with the FLOPs drop of $75\%$ and $90\%$ , and with/without the CE and KR modules to the pruning method. From Tab. 7, we find that KR improves OA of $0.84\%$ and $2.33\%$ when pruning rate is $75\%$ and $90\%$ , respectively, while CE improves OA of $0.32\%$ and $1.77\%$ at $75\%$ and $90\%$ pruning rates, respectively. Bringing the two modules together, the OA improvement is $1.84\%$ and $3.50\%$ at $75\%$ and $90\%$ pruning rates, respectively. The ablation study results have validated the effectiveness of all designs in $\mathrm{CP}^3$ . + +# 5. Conclusion + +In this paper, we focus on 3D point-based network pruning and design a 3D channel pruning plug-in $(\mathrm{CP}^3)$ that can be used with existing 2D CNN pruning methods. To the best of our knowledge, this is the first pruning work explicitly considering the characteristics of point cloud data and point-based networks. Empirically, we show that the proposed $\mathrm{CP}^3$ is universally effective for a wide range of point-based networks and 3D tasks. + +# Acknowledgement + +This work was partially supported by National Natural Science Foundation of China (NSFC) under Grant 62271203, the Science and Technology Innovation Action Plan of Shanghai under Grant 2251110540, and Shanghai Pujiang Program under Grant 21PJ1420300. + +# References + +[1] Iro Armeni, Ozan Sener, Amir R. Zamir, Helen Jiang, Ioannis K. Brilakis, Martin Fischer, and Silvio Savarese. 3D semantic parsing of large-scale indoor spaces. In IEEE Conference on Computer Vision and Pattern Recognition, CVPR, pages 1534-1543, 2016. 6, 12 +[2] Chen Chen, Zhe Chen, Jing Zhang, and Dacheng Tao. SASA: semantics-augmented set abstraction for point-based 3D object detection. In Thirty-Sixth AAAI Conference on Artificial Intelligence, AAAI, pages 221-229, 2022. 1 +[3] MMDetection3D Contributors. MMDetection3D: Open-MMLab next-generation platform for general 3D object detection. https://github.com/open-mmlab/mmdetection3d, 2020.6 +[4] Angela Dai, Angel X Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3D reconstructions of indoor scenes. In IEEE Conference on Computer Vision and Pattern Recognition, pages 5828-5839, 2017. 12 +[5] Angela Dai, Angel X. Chang, Manolis Savva, Maciej Halber, Thomas A. Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3D reconstructions of indoor scenes. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR, pages 2432-2443, 2017. 6, 8 +[6] Xiaohan Ding, Guiguang Ding, Yuchen Guo, and Jungong Han. Centripetal sgd for pruning very deep convolutional networks with complicated structure. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 4943-4953, 2019. 1, 3 +[7] Xiaohan Ding, Tianxiang Hao, Jianchao Tan, Ji Liu, Jungong Han, Yuchen Guo, and Guiguang Ding. ResRep: Lossless CNN pruning via decoupling remembering and forgetting. In IEEE/CVF International Conference on Computer Vision, ICCV, pages 4510-4520, 2021. 1, 3, 5 +[8] Anand Dubey, Avik Santra, Jonas Fuchs, Maximilian Lübke, Robert Weigel, and Fabian Lurz. Haradnet: Anchor-free target detection for radar point clouds using hierarchical attention and multi-task learning. Machine Learning with Applications, 8:100275, 2022. 1 +[9] Hamidreza Fazlali, Yixuan Xu, Yuan Ren, and Bingbing Liu. A versatile multi-view framework for LiDAR-based 3D object detection with guidance from panoptic segmentation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR, pages 17171-17180, 2022. 1 +[10] Kehong Gong, Bingbing Li, Jianfeng Zhang, Tao Wang, Jing Huang, Michael Bi Mi, Jiashi Feng, and Xinchao Wang. PoseTriplet: Co-evolving 3d human pose estimation, imitation, and hallucination under self-supervision. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR, pages 11017-11027, 2022. 1 +[11] Google. TensorFlow lite. https://www.tensorflow.org/ lite, 2017. 2 +[12] Yushuo Guan, Ning Liu, Pengyu Zhao, Zhengping Che, Kaigui Bian, Yanzhi Wang, and Jian Tang. DAIS: Automatic channel pruning via differentiable annealing indicator search. IEEE Transactions on Neural Networks and Learning Systems, TNNLS, pages 1-12, 2022. 1, 3 + +[13] Kai Han, Yunhe Wang, Qiulin Zhang, Wei Zhang, Chunjing Xu, and Tong Zhang. Model rubik's cube: Twisting resolution, depth and width for tinynets. Advances in Neural Information Processing Systems, 33:19353-19364, 2020. 3 +[14] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In IEEE/CVF conference on Computer Vision and Pattern Recognition, CVPR, pages 770-778, 2016. 1 +[15] Zejiang Hou, Minghai Qin, Fei Sun, Xiaolong Ma, Kun Yuan, Yi Xu, Yen-Kuang Chen, Rong Jin, Yuan Xie, and Sun-Yuan Kung. CHEX: channel exploration for CNN model compression. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR, pages 12277-12288, 2022. 3 +[16] Andrew G. Howard, Menglong Zhu, Bo Chen, Dmitry Kalenichenko, Weijun Wang, Tobias Weyand, Marco Andretto, and Hartwig Adam. MobileNets: Efficient convolutional neural networks for mobile vision applications. CoRR, abs/1704.04861, 2017. 1 +[17] Fengwei Jia, Xuan Wang, Jian Guan, Huale Li, Chen Qiu, and Shuhan Qi. Arank: Toward specific model pruning via advantage rank for multiple salient objects detection. Image and Vision Computing, 111:104192, 2021. 12 +[18] Xiaotang Jiang, Huan Wang, Yiliu Chen, Ziqi Wu, Lichuan Wang, Bin Zou, Yafeng Yang, Zongyang Cui, Yu Cai, Tianhang Yu, et al. MNN: A universal and efficient inference engine. The Conference on Machine Learning and Systems, MLSys, pages 1-13, 2020. 3 +[19] Hema Koppula, Abhishek Anand, Thorsten Joachims, and Ashutosh Saxena. Semantic labeling of 3d point clouds for indoor scenes. Advances in Neural Information Processing Systems, NeurIPS, pages 244-252, 2011. 2 +[20] Bastian Leibe, Ales Leonardis, and Bernt Schiele. Robust object detection with interleaved categorization and segmentation. International Journal of Computer Vision, 77(1):259-289, 2008. 3 +[21] Hao Li, Asim Kadav, Igor Durdanovic, Hanan Samet, and Hans Peter Graf. Pruning filters for efficient convnets. In International Conference on Learning Representations, ICLR, 2017. 3 +[22] Kelin Li, Nicholas Baron, Xian Zhang, and Nicolas Rojas. Efficientgrasp: A unified data-efficient learning to grasp method for multi-fingered robot hands. IEEE Robotics and Automation Letters, 7(4):8619-8626, 2022. 1 +[23] Lucas Liebenwein, Cenk Baykal, Harry Lang, Dan Feldman, and Daniela Rus. Provable filter pruning for efficient neural networks. In 8th International Conference on Learning Representations, ICLR, 2020. 3 +[24] Mingbao Lin, Rongrong Ji, Yan Wang, Yichen Zhang, Baochang Zhang, Yonghong Tian, and Ling Shao. HRank: Filter pruning using high-rank feature map. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR, pages 1529-1538, 2020. 1, 3, 4, 5 +[25] Liyang Liu, Shilong Zhang, Zhanghui Kuang, Aojun Zhou, Jing-Hao Xue, Xinjiang Wang, Yimin Chen, Wenming Yang, Qingmin Liao, and Wayne Zhang. Group Fisher pruning for practical network compression. In International Conference on Machine Learning, ICML, pages 7021-7032, 2021. 6 + +[26] Zhuang Liu, Jianguo Li, Zhiqiang Shen, Gao Huang, Shoumeng Yan, and Changshui Zhang. Learning efficient convolutional networks through network slimming. In IEEE/CVF International Conference on Computer Vision, ICCV, pages 2736-2744, 2017. 1, 3 +[27] Ze Liu, Zheng Zhang, Yue Cao, Han Hu, and Xin Tong. Group-free 3D object detection via transformers. In IEEE/CVF International Conference on Computer Vision, CVPR, pages 2949-2958, 2021. 3 +[28] Xu Ma, Can Qin, Haoxuan You, Haoxi Ran, and Yun Fu. Rethinking network design and local geometry in point cloud: A simple residual MLP framework. In International Conference on Learning Representations, ICLR, 2022. 3 +[29] Fanxu Meng, Hao Cheng, Ke Li, Huixiang Luo, Xiaowei Guo, Guangming Lu, and Xing Sun. Pruning filter in filter. In Advances in Neural Information Processing Systems, NeurIPS, volume 33, pages 17629-17640, 2020. 1, 3 +[30] Quang-Hieu Pham, Thanh Nguyen, Binh-Son Hua, Gemma Roig, and Sai-Kit Yeung. JSIS3D: Joint semantic-instance segmentation of 3d point clouds with multi-task pointwise networks and multi-value conditional random fields. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR, 2019. 1 +[31] Charles R Qi, Or Litany, Kaiming He, and Leonidas J Guibas. Deep hough voting for 3D object detection in point clouds. In IEEE/CVF International Conference on Computer Vision, ICCV, 2019. 3, 6, 12 +[32] Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. Pointnet: Deep learning on point sets for 3D classification and segmentation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR, pages 652-660, 2017. 3 +[33] Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J Guibas. Pointnet++: Deep hierarchical feature learning on point sets in a metric space. In Advances in Neural Information Processing Systems, NeurIPS, pages 5100-5109, 2017. 3, 5, 6, 7 +[34] Guocheng Qian, Hasan Hammoud, Guohao Li, Ali Thabet, and Bernard Ghanem. ASSANet: An anisotropic separable set abstraction for efficient point cloud representation learning. In Advances in Neural Information Processing Systems, NeurIPS, pages 28119-28130, 2021. 3 +[35] Guocheng Qian, Yuchen Li, Houwen Peng, Jinjie Mai, Hasan Hammoud, Mohamed Elhoseiny, and Bernard Ghanem. PointNeXt: Revisiting pointnet++ with improved training and scaling strategies. In Advances in Neural Information Processing Systems, NeurIPS, 2022. 3, 5, 6, 7, 12, 13 +[36] Ruizhi Shao, Zerong Zheng, Hongwen Zhang, Jingxiang Sun, and Yebin Liu. DiffuStereo: High quality human reconstruction via diffusion-based stereo using sparse cameras. In European Conference on Computer Vision, ECCV, pages 702-720, 2022. 1 +[37] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. In International Conference on Learning Representations, ICLR, 2015. 1, 2 + +[38] Shuran Song, Samuel P Lichtenberg, and Jianxiong Xiao. Sun RGB-D: A RGB-D scene understanding benchmark suite. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR, pages 567-576, 2015. 6, 8 +[39] Shuran Song, Samuel P Lichtenberg, and Jianxiong Xiao. Sun RGB-D: A RGB-D scene understanding benchmark suite. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR, pages 567-576, 2015. 12 +[40] Yang Sui, Miao Yin, Yi Xie, Huy Phan, Saman Aliari Zonouz, and Bo Yuan. CHIP: Channel independence-based pruning for compact neural networks. In Advances in Neural Information Processing Systems, NeurIPS, pages 24604-24616, 2021. 1, 3, 4, 5, 12 +[41] Lyne P. Tchapmi, Christopher B. Choy, Iro Armeni, JunYoung Gwak, and Silvio Savarese. SEGCloud: Semantic segmentation of 3d point clouds. In International Conference on 3D Vision, 3DV, pages 537-547, 2017. 6 +[42] Hugues Thomas, Charles R Qi, Jean-Emmanuel Deschaud, Beatz Marcotegui, François Goulette, and Leonidas J Guibas. KpConv: Flexible and deformable convolution for point clouds. In IEEE/CVF international conference on computer vision, CVPR, pages 6411–6420, 2019. 3 +[43] Mikaela Angelina Uy, Quang-Hieu Pham, Binh-Son Hua, Duc Thanh Nguyen, and Sai-Kit Yeung. Revisiting point cloud classification: A new benchmark dataset and classification model on real-world data. In IEEE/CVF International Conference on Computer Vision, ICCV, pages 1588–1597, 2019. 5, 6, 7 +[44] Haowen Wang, Mingyuan Wang, Zhengping Che, Zhiyuan Xu, Xiuquan Qiao, Mengshi Qi, Feifei Feng, and Jian Tang. RGB-Depth fusion GAN for indoor depth completion. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR, pages 6199-6208, 2022. 1 +[45] Wenxiao Wang, Minghao Chen, Shuai Zhao, Long Chen, Jinming Hu, Haifeng Liu, Deng Cai, Xiaofei He, and Wei Liu. Accelerate cnns from three dimensions: A comprehensive pruning framework. In Marina Meila and Tong Zhang, editors, Proceedings of the 38th International Conference on Machine Learning, ICML 2021, 18-24 July 2021, Virtual Event, volume 139 of Proceedings of Machine Learning Research, pages 10717-10726. PMLR, 2021. 3 +[46] Wenxiao Wang, Cong Fu, Jishun Guo, Deng Cai, and Xiaofei He. COP: customized deep model compression via regularized correlation-based filter-level pruning. In Sarit Kraus, editor, Proceedings of the Twenty-Eighth International Joint Conference on Artificial Intelligence, IJCAI 2019, Macao, China, August 10-16, 2019, pages 3785-3791. ijcai.org, 2019. 3 +[47] Zhirong Wu, Shuran Song, Aditya Khosla, Fisher Yu, Linguuang Zhang, Xiaou Tang, and Jianxiong Xiao. 3D ShapeNets: A deep representation for volumetric shapes. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR, pages 1912-1920, 2015. 5, 6 +[48] Chenfeng Xu, Shijia Yang, Tomer Galanti, Bichen Wu, Xiangyu Yue, Bohan Zhai, Wei Zhan, Peter Vajda, Kurt Keutzer, and Masayoshi Tomizuka. Image2point: 3d point-cloud understanding with 2d image pretrained models. In + +European Conference on Computer Vision, pages 638-656. Springer, 2022. 2 +[49] Wei Yang, Chris Paxton, Maya Cakmak, and Dieter Fox. Human grasp classification for reactive human-to-robot handovers. In IEEE/RSJ International Conference on Intelligent Robots and Systems, IROS, pages 11123–11130, 2020. 1 +[50] Xucheng Ye, Pengcheng Dai, Junyu Luo, Xin Guo, Yingjie Qi, Jianlei Yang, and Yiran Chen. Accelerating CNN training by pruning activation gradients. In European Conference on Computer Vision, ECCV, pages 322-338, 2020. 3 +[51] Ruichi Yu, Ang Li, Chun-Fu Chen, Jui-Hsin Lai, Vlad I Morariu, Xintong Han, Mingfei Gao, Ching-Yung Lin, and Larry S Davis. Nisp: Pruning networks using neuron importance score propagation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR, pages 9194-9203, 2018. 3 +[52] Hengshuang Zhao, Li Jiang, Jiaya Jia, Philip H. S. Torr, and Vladlen Koltun. Point transformer. In IEEE/CVF International Conference on Computer Vision, ICCV, pages 16239-16248, 2021. 3 +[53] Wu Zheng, Weiliang Tang, Li Jiang, and Chi-Wing Fu. SESSD: self-ensembling single-stage object detector from point cloud. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR, pages 14494–14503, 2021. 1 \ No newline at end of file diff --git a/2023/CP3_ Channel Pruning Plug-In for Point-Based Networks/images.zip b/2023/CP3_ Channel Pruning Plug-In for Point-Based Networks/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..3e55a59b6d7ee3c7109ed710ce159bb64c5723fd --- /dev/null +++ b/2023/CP3_ Channel Pruning Plug-In for Point-Based Networks/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:49b729fe48b2333c1685ae6b63146c7f86798383ca59c53d3d8ef8a7ccb104c7 +size 571850 diff --git a/2023/CP3_ Channel Pruning Plug-In for Point-Based Networks/layout.json b/2023/CP3_ Channel Pruning Plug-In for Point-Based Networks/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..05e00eab7fc76be01c278ec16f44576e1a0899f7 --- /dev/null +++ b/2023/CP3_ Channel Pruning Plug-In for Point-Based Networks/layout.json @@ -0,0 +1,9480 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 121, + 102, + 473, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 121, + 102, + 473, + 121 + ], + "spans": [ + { + "bbox": [ + 121, + 102, + 473, + 121 + ], + "type": "inline_equation", + "content": "\\mathbf{CP}^{3}" + }, + { + "bbox": [ + 121, + 102, + 473, + 121 + ], + "type": "text", + "content": ": Channel Pruning Plug-in for Point-based Networks" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 63, + 140, + 531, + 179 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 140, + 531, + 179 + ], + "spans": [ + { + "bbox": [ + 63, + 140, + 531, + 179 + ], + "type": "inline_equation", + "content": "\\mathrm{Yaxin Peng}^{3}" + }, + { + "bbox": [ + 63, + 140, + 531, + 179 + ], + "type": "text", + "content": " Ning Liu" + }, + { + "bbox": [ + 63, + 140, + 531, + 179 + ], + "type": "inline_equation", + "content": "^{2,*}" + }, + { + "bbox": [ + 63, + 140, + 531, + 179 + ], + "type": "text", + "content": " Zhiying Che" + }, + { + "bbox": [ + 63, + 140, + 531, + 179 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 63, + 140, + 531, + 179 + ], + "type": "text", + "content": " Zhiyuan Xu" + }, + { + "bbox": [ + 63, + 140, + 531, + 179 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 63, + 140, + 531, + 179 + ], + "type": "text", + "content": " Chaomin Shen" + }, + { + "bbox": [ + 63, + 140, + 531, + 179 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 63, + 140, + 531, + 179 + ], + "type": "text", + "content": " Guixu Zhang" + }, + { + "bbox": [ + 63, + 140, + 531, + 179 + ], + "type": "inline_equation", + "content": "^{1, \\boxtimes}" + }, + { + "bbox": [ + 63, + 140, + 531, + 179 + ], + "type": "text", + "content": " Xinmei Liu" + }, + { + "bbox": [ + 63, + 140, + 531, + 179 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 63, + 140, + 531, + 179 + ], + "type": "text", + "content": " Feifei Feng" + }, + { + "bbox": [ + 63, + 140, + 531, + 179 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 63, + 140, + 531, + 179 + ], + "type": "text", + "content": " Jian Tang" + }, + { + "bbox": [ + 63, + 140, + 531, + 179 + ], + "type": "inline_equation", + "content": "^{2, \\boxtimes}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 149, + 183, + 445, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 149, + 183, + 445, + 213 + ], + "spans": [ + { + "bbox": [ + 149, + 183, + 445, + 213 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 149, + 183, + 445, + 213 + ], + "type": "text", + "content": "School of Computer Science, East China Normal University " + }, + { + "bbox": [ + 149, + 183, + 445, + 213 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 149, + 183, + 445, + 213 + ], + "type": "text", + "content": "Midea Group" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 127, + 213, + 466, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 127, + 213, + 466, + 228 + ], + "spans": [ + { + "bbox": [ + 127, + 213, + 466, + 228 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 127, + 213, + 466, + 228 + ], + "type": "text", + "content": "Department of Mathematics, School of Science, Shanghai University" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 121, + 229, + 473, + 242 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 121, + 229, + 473, + 242 + ], + "spans": [ + { + "bbox": [ + 121, + 229, + 473, + 242 + ], + "type": "text", + "content": "1{51205901049,51205901078}@stu.ecnu.edu.cn 1{cmshen,gxzhang}@cs.ecnu.edu.cn" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 243, + 485, + 255 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 105, + 243, + 485, + 255 + ], + "spans": [ + { + "bbox": [ + 105, + 243, + 485, + 255 + ], + "type": "text", + "content": "2{liuning22, chezp, xuzy70, feifei.feng, tangjian22}@midea.com yaxin.peng@shu.edu.cn" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 143, + 296, + 191, + 308 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 296, + 191, + 308 + ], + "spans": [ + { + "bbox": [ + 143, + 296, + 191, + 308 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 323, + 290, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 323, + 290, + 647 + ], + "spans": [ + { + "bbox": [ + 46, + 323, + 290, + 647 + ], + "type": "text", + "content": "Channel pruning can effectively reduce both computational cost and memory footprint of the original network while keeping a comparable accuracy performance. Though great success has been achieved in channel pruning for 2D image-based convolutional networks (CNNs), existing works seldom extend the channel pruning methods to 3D point-based neural networks (PNNs). Directly implementing the 2D CNN channel pruning methods to PNNs undermine the performance of PNNs because of the different representations of 2D images and 3D point clouds as well as the network architecture disparity. In this paper, we proposed " + }, + { + "bbox": [ + 46, + 323, + 290, + 647 + ], + "type": "inline_equation", + "content": "CP^3" + }, + { + "bbox": [ + 46, + 323, + 290, + 647 + ], + "type": "text", + "content": ", which is a Channel Pruning Plugin for Point-based network. " + }, + { + "bbox": [ + 46, + 323, + 290, + 647 + ], + "type": "inline_equation", + "content": "CP^3" + }, + { + "bbox": [ + 46, + 323, + 290, + 647 + ], + "type": "text", + "content": " is elaborately designed to leverage the characteristics of point clouds and PNNs in order to enable 2D channel pruning methods for PNNs. Specifically, it presents a coordinate-enhanced channel importance metric to reflect the correlation between dimensional information and individual channel features, and it recycles the discarded points in PNN's sampling process and reconsiders their potentially-exclusive information to enhance the robustness of channel pruning. Experiments on various PNN architectures show that " + }, + { + "bbox": [ + 46, + 323, + 290, + 647 + ], + "type": "inline_equation", + "content": "CP^3" + }, + { + "bbox": [ + 46, + 323, + 290, + 647 + ], + "type": "text", + "content": " constantly improves state-of-the-art 2D CNN pruning approaches on different point cloud tasks. For instance, our compressed PointNeXt-S on ScanObjectNN achieves an accuracy of " + }, + { + "bbox": [ + 46, + 323, + 290, + 647 + ], + "type": "inline_equation", + "content": "88.52\\%" + }, + { + "bbox": [ + 46, + 323, + 290, + 647 + ], + "type": "text", + "content": " with a pruning rate of " + }, + { + "bbox": [ + 46, + 323, + 290, + 647 + ], + "type": "inline_equation", + "content": "57.8\\%" + }, + { + "bbox": [ + 46, + 323, + 290, + 647 + ], + "type": "text", + "content": ", outperforming the baseline pruning methods with an accuracy gain of " + }, + { + "bbox": [ + 46, + 323, + 290, + 647 + ], + "type": "inline_equation", + "content": "1.94\\%" + }, + { + "bbox": [ + 46, + 323, + 290, + 647 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 283, + 386, + 295 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 283, + 386, + 295 + ], + "spans": [ + { + "bbox": [ + 306, + 283, + 386, + 295 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 305, + 547, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 305, + 547, + 568 + ], + "spans": [ + { + "bbox": [ + 304, + 305, + 547, + 568 + ], + "type": "text", + "content": "Convolutional Neural Networks (CNNs) often encounter the problems of overloaded computation and overweighted storage. The cumbersome instantiation of a CNN model leads to inefficient, uneconomic, or even impossible deployment in practice. Therefore, light-weight models that provide comparable results with much fewer computational costs are in great demand for nearly all applications. Channel pruning is a promising solution to delivering efficient networks. In recent years, 2D CNN channel pruning, e.g., pruning classical VGGNets [37], ResNets [14], MobileNets [16], and many other neural networks for processing 2D images [6, 7, 12, 24, 26, 29, 40], has been successfully conducted. Most channel pruning approaches focus on identifying redundant convolution filters (i.e., channels) by evaluating their importance. The cornerstone of 2D channel pruning methods is the diversified yet effective channel evaluation metrics. For instance, HRank [24] uses the rank of the feature map as the pruning metric and removes the low-rank filters that are considered to contain less information. CHIP [40] leverages channel independence to represent the importance of each feature mapping and eliminates less important channels." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 570, + 548, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 548, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 548, + 714 + ], + "type": "text", + "content": "With the widespread application of depth-sensing technology, 3D vision tasks [9, 10, 36, 44] are a rapidly growing field starving for powerful methods. Apart from straightforwardly applying 2D CNNs, models built with Point-based Neural Networks (PNNs), which directly process point clouds from the beginning without unnecessary rendering, show their merits and are widely deployed on edge devices for various applications such as robots [22, 49] and self-driving [2, 53]. Compressing PNNs is crucial due to the limited resources of edge devices and multiple models for different tasks are likely to run simultaneously [8, 30]. Given the huge success of 2D channel pruning and the great" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 66, + 674, + 138, + 684 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 674, + 138, + 684 + ], + "spans": [ + { + "bbox": [ + 66, + 674, + 138, + 684 + ], + "type": "text", + "content": "* Equal contributions." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 66, + 685, + 150, + 693 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 685, + 150, + 693 + ], + "spans": [ + { + "bbox": [ + 66, + 685, + 150, + 693 + ], + "type": "text", + "content": "Corresponding authors." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 694, + 287, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 694, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 694, + 287, + 712 + ], + "type": "text", + "content": "This work is done during Yaomin Huang and Xinmei Liu's internship at Midea Group." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5302" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "text", + "content": "demand for efficient 3D PNNs, we intuitively raise one question: shall we directly implement the existing pruning methods to PNNs following the proposed channel importance metrics in 2D CNNs pruning?" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 122, + 288, + 314 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 122, + 288, + 314 + ], + "spans": [ + { + "bbox": [ + 46, + 122, + 288, + 314 + ], + "type": "text", + "content": "With this question in mind, we investigate the fundamental factors that potentially impair 2D pruning effectiveness on PNNs. Previous works [19, 48] have shown that point clouds record visual and semantic information in a significantly different way from 2D images. Specifically, a point cloud consists of a set of unordered points on objects' and environments' surfaces, and each point encodes its features, such as intensity along with the spatial coordinates " + }, + { + "bbox": [ + 46, + 122, + 288, + 314 + ], + "type": "inline_equation", + "content": "(x,y,z)" + }, + { + "bbox": [ + 46, + 122, + 288, + 314 + ], + "type": "text", + "content": ". In contrast, 2D images organize visual features in a dense and regular pixel array. Such data representation differences between 3D point clouds and 2D images lead to a) different ways of exploiting information from data and b) contrasting network architectures of PNNs and 2D CNNs. It is credible that only the pruning methods considering the two aspects (definitely not existing 2D CNN pruners) may obtain superior performance on PNNs." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 316, + 288, + 508 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 316, + 288, + 508 + ], + "spans": [ + { + "bbox": [ + 46, + 316, + 288, + 508 + ], + "type": "text", + "content": "From the perspective of data representations, 3D point clouds provide more 3D feature representations than 2D images, but the representations are more sensitive to network channels. To be more specific, for 2D images, all three RGB channels represent basic information in an isotropic and homogeneous way so that the latent representations extracted by CNNs applied to the images. On the other hand, point clouds explicitly encode the spatial information in three coordinate channels, which are indispensable for extracting visual and semantic information from other channels. Moreover, PNNs employ the coordinate information in multiple layers as concatenated inputs for deeper feature extraction. Nevertheless, existing CNN pruning methods are designed only suitable for the plain arrangements of 2D data but fail to consider how the informative 3D information should be extracted from point clouds." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 510, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 510, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 510, + 288, + 714 + ], + "type": "text", + "content": "Moreover, the network architectures of PNNs are designed substantially different from 2D CNNs. While using smaller kernels [37] is shown to benefit 2D CNNs [37], it does not apply to networks for 3D point clouds. On the contrary, PNNs leverage neighborhoods at multiple scales to obtain both robust and detailed features. The reason is that small neighborhoods (analogous to small kernels in 2D CNNs) in point clouds consist of few points for PNNs to capture robust features. Due to the necessary sampling steps, the knowledge insufficiency issue becomes more severe for deeper PNN layers. In addition, PNNs use the random input dropout procedure during training to adaptively weight patterns detected at different scales and combine multi-scale features. This procedure randomly discards a large proportion of points and loses much exclusive information of the original data. Thus, the architecture disparity between 2D CNNs and PNNs affects the performance of" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 305, + 72, + 518, + 84 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 518, + 84 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 518, + 84 + ], + "type": "text", + "content": "directly applying existing pruning methods to PNNs." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 85, + 545, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 85, + 545, + 407 + ], + "spans": [ + { + "bbox": [ + 304, + 85, + 545, + 407 + ], + "type": "text", + "content": "In this paper, by explicitly dealing with the two characteristics of 3D task, namely the data representation and the PNN architecture design, we propose a Channel Pruning Plug-in for Point-based network named " + }, + { + "bbox": [ + 304, + 85, + 545, + 407 + ], + "type": "inline_equation", + "content": "\\mathrm{CP}^3" + }, + { + "bbox": [ + 304, + 85, + 545, + 407 + ], + "type": "text", + "content": ", which can be applied to most 2D channel pruning methods for compressing PNN models. The proposed " + }, + { + "bbox": [ + 304, + 85, + 545, + 407 + ], + "type": "inline_equation", + "content": "\\mathrm{CP}^3" + }, + { + "bbox": [ + 304, + 85, + 545, + 407 + ], + "type": "text", + "content": " refines the channel importance, the key factor of pruning methods, from two aspects. Firstly, considering the point coordinates " + }, + { + "bbox": [ + 304, + 85, + 545, + 407 + ], + "type": "inline_equation", + "content": "(x, y, \\text{and } z)" + }, + { + "bbox": [ + 304, + 85, + 545, + 407 + ], + "type": "text", + "content": " encode the spatial information and deeply affects feature extraction procedures in PNN layers, we determine the channel importance by evaluating the correlation between the feature map and its corresponding point coordinates by introducing a coordinate-enhancement module. Secondly, calculating channel importance in channel pruning is data-driven and sensitive to the input, and the intrinsic sampling steps in PNN naturally makes pruning methods unstable. To settle this problem, we make full use of the discarded points in the sampling process via a knowledge recycling module to supplement the evaluation of channel importance. This reduces the data sampling bias impact on the channel importance calculation and increases the robustness of the pruning results. Notably, both the coordinates and recycled points in " + }, + { + "bbox": [ + 304, + 85, + 545, + 407 + ], + "type": "inline_equation", + "content": "\\mathrm{CP}^3" + }, + { + "bbox": [ + 304, + 85, + 545, + 407 + ], + "type": "text", + "content": " do not participate in network training (with back-propagation) but only assist channel importance calculation in the reasoning phase. Thus, " + }, + { + "bbox": [ + 304, + 85, + 545, + 407 + ], + "type": "inline_equation", + "content": "\\mathrm{CP}^3" + }, + { + "bbox": [ + 304, + 85, + 545, + 407 + ], + "type": "text", + "content": " does not increase any computational cost of the pruned network. The contributions of this paper are as follows:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 307, + 408, + 545, + 617 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 307, + 408, + 545, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 408, + 545, + 468 + ], + "spans": [ + { + "bbox": [ + 307, + 408, + 545, + 468 + ], + "type": "text", + "content": "- We systematically consider the characteristics of PNNs and propose a channel pruning plug-in named " + }, + { + "bbox": [ + 307, + 408, + 545, + 468 + ], + "type": "inline_equation", + "content": "\\mathrm{CP}^3" + }, + { + "bbox": [ + 307, + 408, + 545, + 468 + ], + "type": "text", + "content": " to enhance 2D CNN channel pruning approaches on 3D PNNs. To the best of our knowledge, " + }, + { + "bbox": [ + 307, + 408, + 545, + 468 + ], + "type": "inline_equation", + "content": "\\mathrm{CP}^3" + }, + { + "bbox": [ + 307, + 408, + 545, + 468 + ], + "type": "text", + "content": " is the first method to export existing 2D pruning methods to PNNs." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 307, + 470, + 545, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 470, + 545, + 517 + ], + "spans": [ + { + "bbox": [ + 307, + 470, + 545, + 517 + ], + "type": "text", + "content": "- We propose a coordinate-enhanced channel importance score to guide point clouds network pruning, by evaluating the correlation between feature maps and corresponding point coordinates." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 307, + 520, + 545, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 520, + 545, + 567 + ], + "spans": [ + { + "bbox": [ + 307, + 520, + 545, + 567 + ], + "type": "text", + "content": "- We design a knowledge recycling pruning scheme that increases the robustness of the pruning procedure, using the discarded points to improve the channel importance evaluation." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 307, + 569, + 545, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 569, + 545, + 617 + ], + "spans": [ + { + "bbox": [ + 307, + 569, + 545, + 617 + ], + "type": "text", + "content": "- We show that using " + }, + { + "bbox": [ + 307, + 569, + 545, + 617 + ], + "type": "inline_equation", + "content": "\\mathrm{CP}^3" + }, + { + "bbox": [ + 307, + 569, + 545, + 617 + ], + "type": "text", + "content": " is consistently superior to directly transplanting 2D pruning methods to PNNs by extensive experiments on three 3D tasks and five datasets with different PNN models and pruning baselines." + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 628, + 392, + 639 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 628, + 392, + 639 + ], + "spans": [ + { + "bbox": [ + 306, + 628, + 392, + 639 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 647, + 425, + 660 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 647, + 425, + 660 + ], + "spans": [ + { + "bbox": [ + 306, + 647, + 425, + 660 + ], + "type": "text", + "content": "2.1. 2D Channel Pruning" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 665, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 665, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 665, + 545, + 713 + ], + "type": "text", + "content": "Channel pruning (a.k.a., filter pruning) methods reduce the redundant filters while maintaining the original structure of CNNs and is friendly to prevailing inference acceleration engines such as TensorFlow-Lite (TFLite) [11] and" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5303" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 289, + 348 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 289, + 348 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 289, + 348 + ], + "type": "text", + "content": "Mobile Neural Network (MNN) [18]. Mainstream channel pruning methods [6, 7, 12, 29] usually first evaluate the importance of channels by certain metrics and then prune (i.e., remove) the less important channels. Early work [21] uses the " + }, + { + "bbox": [ + 47, + 72, + 289, + 348 + ], + "type": "inline_equation", + "content": "l_{1}" + }, + { + "bbox": [ + 47, + 72, + 289, + 348 + ], + "type": "text", + "content": " norm of filters as importance score for channel pruning. Afterwards, learning parameters, such as the scaling factor " + }, + { + "bbox": [ + 47, + 72, + 289, + 348 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 47, + 72, + 289, + 348 + ], + "type": "text", + "content": " in the batch norm layer [26] and the reconstruction error in the final network layer [51], are considered as the importance scores for channel selection. The importance sampling distribution of channels [23] is also used for pruning. Recent works [15, 40] measure the correlation of multiple feature maps to determine the importance score of the filter for pruning. HRank [24] proposes a method for pruning filters based on the theory that low-rank feature maps contain less information. [50] leverages the statistical distribution of activation gradient and takes the smaller gradient as low importance score for pruning. [46] calculates the average importance of both the input feature maps and their corresponding output feature maps to determine the overall importance. [13, 45] compress CNNs from multiple dimensions While most channel pruning methods are designed for and tested on 2D CNNs, our " + }, + { + "bbox": [ + 47, + 72, + 289, + 348 + ], + "type": "inline_equation", + "content": "\\mathrm{CP}^3" + }, + { + "bbox": [ + 47, + 72, + 289, + 348 + ], + "type": "text", + "content": " can work in tandem with existing pruners for 3D point-based networks." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 358, + 274, + 371 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 358, + 274, + 371 + ], + "spans": [ + { + "bbox": [ + 47, + 358, + 274, + 371 + ], + "type": "text", + "content": "2.2. Point-based Networks for Point Cloud Data" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 379, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 379, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 47, + 379, + 289, + 715 + ], + "type": "text", + "content": "Point-based Neural Networks (PNNs) directly process point cloud data with a flexible range of receptive field, have no positioning information loss, and thus keep more accurate spatial information. As a pioneer work, PointNet [32] learns the spatial encoding directly from the input point clouds and uses the characteristics of all points to obtain the global representations. PointNet++ [33] further proposes a multi-level feature extraction structure to extract local and global features more effectively. KPCov [42] proposes a new point convolution operation to learn local movements applied to kernel points. ASSANet [34] proposes a separable set abstraction module that decomposes the normal SA module in PointNet++ into two separate learning phases for channel and space. PointMLP [28] uses residual point blocks to extract local features, transforms local points using geometric affine modules, and extracts geometric features before and after the aggregation operation. PointNeXt [35] uses inverted residual bottleneck and separable multilayer perceptrons to achieve more efficient model scaling. Besides classification, PNNs also serve as backbones for other 3D tasks. VoteNet [31] effectively improves the 3D object detection accuracy through the Hough voting mechanism [20]. PointTransformer [52] designs models improving prior work across domains and tasks. GroupFree3D [27] uses the attention mechanism to automatically learn the contribution of each point to the object. In this paper, we show that " + }, + { + "bbox": [ + 47, + 379, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\mathrm{CP}^3" + }, + { + "bbox": [ + 47, + 379, + 289, + 715 + ], + "type": "text", + "content": " can be widely applied to point-based networks on a variety of point cloud bench" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 306, + 72, + 483, + 84 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 483, + 84 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 483, + 84 + ], + "type": "text", + "content": "marks and representative original networks." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 306, + 95, + 389, + 109 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 95, + 389, + 109 + ], + "spans": [ + { + "bbox": [ + 306, + 95, + 389, + 109 + ], + "type": "text", + "content": "3. Methodology" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 115, + 547, + 259 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 115, + 547, + 259 + ], + "spans": [ + { + "bbox": [ + 304, + 115, + 547, + 259 + ], + "type": "text", + "content": "Although point-based networks are similar to CNN in concrete realization, they have fundamental differences in data representation and network architecture design. To extend the success of CNN pruning to PNN, two modules are proposed in " + }, + { + "bbox": [ + 304, + 115, + 547, + 259 + ], + "type": "inline_equation", + "content": "\\mathrm{CP}^3" + }, + { + "bbox": [ + 304, + 115, + 547, + 259 + ], + "type": "text", + "content": " taking advantage from the dimensional information and discarded points: 1) coordinate-enhancement (CE) module, which produces a coordinate-enhanced score to estimate the channel importance by combining dimensional and feature information, and 2) knowledge recycling module reusing the discarded points to improve the channel importance evaluation criteria and increase the robustness." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 267, + 467, + 279 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 267, + 467, + 279 + ], + "spans": [ + { + "bbox": [ + 306, + 267, + 467, + 279 + ], + "type": "text", + "content": "3.1. Formulations and Motivation" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 285, + 547, + 490 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 285, + 547, + 490 + ], + "spans": [ + { + "bbox": [ + 304, + 285, + 547, + 490 + ], + "type": "text", + "content": "Point-based networks PNN is a unified architecture that directly takes point clouds as input. It builds hierarchical groups of points and progressively abstracts larger local regions along the hierarchy. PNN is structurally composed by a number of set abstraction (SA) blocks. Each SA block consists of 1) a sampling layer iteratively samples the farthest point to choose a subset of points from input points, 2) a group layer gathers neighbors of centroid points to a local region, 3) a set of shared Multi-Layer Perceptrons (MLPs) to extract features, and 4) a reduction layer to aggregate features in the neighbors. Formally speaking, a SA block takes an " + }, + { + "bbox": [ + 304, + 285, + 547, + 490 + ], + "type": "inline_equation", + "content": "n_{i-1} \\times (d + c_{i-1})" + }, + { + "bbox": [ + 304, + 285, + 547, + 490 + ], + "type": "text", + "content": " matrix as input that is from " + }, + { + "bbox": [ + 304, + 285, + 547, + 490 + ], + "type": "inline_equation", + "content": "n_{i-1}" + }, + { + "bbox": [ + 304, + 285, + 547, + 490 + ], + "type": "text", + "content": " points with " + }, + { + "bbox": [ + 304, + 285, + 547, + 490 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 304, + 285, + 547, + 490 + ], + "type": "text", + "content": "-dim coordinates and " + }, + { + "bbox": [ + 304, + 285, + 547, + 490 + ], + "type": "inline_equation", + "content": "c_{i-1}" + }, + { + "bbox": [ + 304, + 285, + 547, + 490 + ], + "type": "text", + "content": "-dim point feature. It outputs an " + }, + { + "bbox": [ + 304, + 285, + 547, + 490 + ], + "type": "inline_equation", + "content": "n_i \\times (d + c_i)" + }, + { + "bbox": [ + 304, + 285, + 547, + 490 + ], + "type": "text", + "content": " matrix of " + }, + { + "bbox": [ + 304, + 285, + 547, + 490 + ], + "type": "inline_equation", + "content": "n_i" + }, + { + "bbox": [ + 304, + 285, + 547, + 490 + ], + "type": "text", + "content": " subsampled points with " + }, + { + "bbox": [ + 304, + 285, + 547, + 490 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 304, + 285, + 547, + 490 + ], + "type": "text", + "content": "-dimensional coordinates (i.e., " + }, + { + "bbox": [ + 304, + 285, + 547, + 490 + ], + "type": "inline_equation", + "content": "d = 3" + }, + { + "bbox": [ + 304, + 285, + 547, + 490 + ], + "type": "text", + "content": ") and new " + }, + { + "bbox": [ + 304, + 285, + 547, + 490 + ], + "type": "inline_equation", + "content": "c_i" + }, + { + "bbox": [ + 304, + 285, + 547, + 490 + ], + "type": "text", + "content": "-dimensional feature vectors summarizing local context. The SA block is formulated as:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 357, + 495, + 545, + 517 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 357, + 495, + 545, + 517 + ], + "spans": [ + { + "bbox": [ + 357, + 495, + 545, + 517 + ], + "type": "interline_equation", + "content": "\\mathbf {F} _ {i} ^ {l + 1} = \\mathcal {R} \\left\\{h _ {\\Theta} \\left(\\left[ \\mathbf {F} _ {j} ^ {l}; \\mathbf {x} _ {j} ^ {l} - \\mathbf {x} _ {i} ^ {l} \\right]\\right) \\right\\}, \\tag {1}", + "image_path": "5962a4202b9dd085c2954174e1e2c80da74dbc6af38923954645296d11e0a5f6.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 524, + 547, + 588 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 524, + 547, + 588 + ], + "spans": [ + { + "bbox": [ + 304, + 524, + 547, + 588 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 524, + 547, + 588 + ], + "type": "inline_equation", + "content": "h_\\Theta" + }, + { + "bbox": [ + 304, + 524, + 547, + 588 + ], + "type": "text", + "content": " is MLPs to extract grouped points feature, " + }, + { + "bbox": [ + 304, + 524, + 547, + 588 + ], + "type": "inline_equation", + "content": "\\mathcal{R}" + }, + { + "bbox": [ + 304, + 524, + 547, + 588 + ], + "type": "text", + "content": " is the reduction layer (e.g. max-pooling) to aggregate features in the neighbors " + }, + { + "bbox": [ + 304, + 524, + 547, + 588 + ], + "type": "inline_equation", + "content": "\\{j : (i,j) \\in \\mathcal{N}\\}" + }, + { + "bbox": [ + 304, + 524, + 547, + 588 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 524, + 547, + 588 + ], + "type": "inline_equation", + "content": "F_j^l" + }, + { + "bbox": [ + 304, + 524, + 547, + 588 + ], + "type": "text", + "content": " is the features of neighbor " + }, + { + "bbox": [ + 304, + 524, + 547, + 588 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 304, + 524, + 547, + 588 + ], + "type": "text", + "content": " in the " + }, + { + "bbox": [ + 304, + 524, + 547, + 588 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 304, + 524, + 547, + 588 + ], + "type": "text", + "content": "-th layer, " + }, + { + "bbox": [ + 304, + 524, + 547, + 588 + ], + "type": "inline_equation", + "content": "x_i^l" + }, + { + "bbox": [ + 304, + 524, + 547, + 588 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 524, + 547, + 588 + ], + "type": "inline_equation", + "content": "x_j^l" + }, + { + "bbox": [ + 304, + 524, + 547, + 588 + ], + "type": "text", + "content": " are input points coordinates and coordinates of neighbor " + }, + { + "bbox": [ + 304, + 524, + 547, + 588 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 304, + 524, + 547, + 588 + ], + "type": "text", + "content": " in the " + }, + { + "bbox": [ + 304, + 524, + 547, + 588 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 304, + 524, + 547, + 588 + ], + "type": "text", + "content": "-th layer." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 601, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 601, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 601, + 547, + 715 + ], + "type": "text", + "content": "Channel pruning Assume a pre-trained PNN model has a set of " + }, + { + "bbox": [ + 304, + 601, + 547, + 715 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 304, + 601, + 547, + 715 + ], + "type": "text", + "content": " convolutional layers, and " + }, + { + "bbox": [ + 304, + 601, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{A}^l" + }, + { + "bbox": [ + 304, + 601, + 547, + 715 + ], + "type": "text", + "content": " is the " + }, + { + "bbox": [ + 304, + 601, + 547, + 715 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 304, + 601, + 547, + 715 + ], + "type": "text", + "content": "-th convolution layer. The parameters in " + }, + { + "bbox": [ + 304, + 601, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{A}^l" + }, + { + "bbox": [ + 304, + 601, + 547, + 715 + ], + "type": "text", + "content": " can be represented as a set of filters " + }, + { + "bbox": [ + 304, + 601, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{W}_{\\mathcal{A}^l} = \\{\\mathbf{w}_1^l,\\mathbf{w}_2^l,\\dots ,\\mathbf{w}_{c_l}^l\\} \\in \\mathbb{R}^{(d + c_l)\\times (d + c_{l - 1})\\times k_l\\times k_l}" + }, + { + "bbox": [ + 304, + 601, + 547, + 715 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 304, + 601, + 547, + 715 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 304, + 601, + 547, + 715 + ], + "type": "text", + "content": "-th filter is " + }, + { + "bbox": [ + 304, + 601, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{w}_j^l\\in \\mathbb{R}^{(d + c_{l - 1})\\times k_l\\times k_j}" + }, + { + "bbox": [ + 304, + 601, + 547, + 715 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 304, + 601, + 547, + 715 + ], + "type": "inline_equation", + "content": "(d + c_{l})" + }, + { + "bbox": [ + 304, + 601, + 547, + 715 + ], + "type": "text", + "content": " represents the number of filters in " + }, + { + "bbox": [ + 304, + 601, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{A}^l" + }, + { + "bbox": [ + 304, + 601, + 547, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 601, + 547, + 715 + ], + "type": "inline_equation", + "content": "k_{l}" + }, + { + "bbox": [ + 304, + 601, + 547, + 715 + ], + "type": "text", + "content": " denotes the kernel size. The outputs of filter, i.e., feature map, are denoted as " + }, + { + "bbox": [ + 304, + 601, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{F}^l = \\{\\mathbf{f}_1^l,\\mathbf{f}_2^l,\\ldots ,\\mathbf{f}_{n_i}^l\\} \\in \\mathbb{R}^{n_i\\times (d + c_i)}" + }, + { + "bbox": [ + 304, + 601, + 547, + 715 + ], + "type": "text", + "content": ". Channel pruning aims to identify and remove the less importance filter from the" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "5304" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 75, + 541, + 222 + ], + "blocks": [ + { + "bbox": [ + 50, + 75, + 541, + 222 + ], + "lines": [ + { + "bbox": [ + 50, + 75, + 541, + 222 + ], + "spans": [ + { + "bbox": [ + 50, + 75, + 541, + 222 + ], + "type": "image", + "image_path": "625020ddf58bb812d6da62b4428f393d1ffe2051d1ea2a09bcce9be253eb92d9.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 225, + 546, + 295 + ], + "lines": [ + { + "bbox": [ + 46, + 225, + 546, + 295 + ], + "spans": [ + { + "bbox": [ + 46, + 225, + 546, + 295 + ], + "type": "text", + "content": "Figure 1. The framework of " + }, + { + "bbox": [ + 46, + 225, + 546, + 295 + ], + "type": "inline_equation", + "content": "\\mathrm{CP}^3" + }, + { + "bbox": [ + 46, + 225, + 546, + 295 + ], + "type": "text", + "content": ". The figure shows the specific pruning process of one of the SA blocks. Whether a channel in a PNN is pruned is determined by three parts: 1) Original channel importance: obtained from the original CNNs channel pruning method (e.g., HRank [24], CHIP [40]). 2) Discarded channel importance: obtained from the Knowledge-Recycling module by leveraging the discarded points in the network to supplement the channel importance evaluation of the corresponding points and improve the robustness of the channel selection. 3) CE (Coordinate-Enhanced) channel importance: obtained from calculating the correlation between the feature map and its corresponding points coordinates to guide point clouds network pruning." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 308, + 288, + 333 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 308, + 288, + 333 + ], + "spans": [ + { + "bbox": [ + 46, + 308, + 288, + 333 + ], + "type": "text", + "content": "original networks. In general, channel pruning can be formulated as the following optimization problem:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 89, + 337, + 287, + 369 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 337, + 287, + 369 + ], + "spans": [ + { + "bbox": [ + 89, + 337, + 287, + 369 + ], + "type": "interline_equation", + "content": "\\min _ {\\delta_ {i j}} \\sum_ {i = 1} ^ {K} \\sum_ {j = 1} ^ {n _ {i}} \\delta_ {i j} \\mathcal {L} \\left(\\mathbf {w} _ {j} ^ {i}\\right), \\text {s . t .} \\sum_ {j = 1} ^ {n _ {i}} \\delta_ {i j} = k _ {l}, \\tag {2}", + "image_path": "e3501b2d26091a18a12ba79c44a773d19367d08ff824b73494205fbd7ff2ad3d.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 374, + 288, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 374, + 288, + 413 + ], + "spans": [ + { + "bbox": [ + 47, + 374, + 288, + 413 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 374, + 288, + 413 + ], + "type": "inline_equation", + "content": "\\delta_{ij}" + }, + { + "bbox": [ + 47, + 374, + 288, + 413 + ], + "type": "text", + "content": " is an indicator which is 1 if " + }, + { + "bbox": [ + 47, + 374, + 288, + 413 + ], + "type": "inline_equation", + "content": "\\mathbf{w}_j^i" + }, + { + "bbox": [ + 47, + 374, + 288, + 413 + ], + "type": "text", + "content": " is to be pruned or 0 if " + }, + { + "bbox": [ + 47, + 374, + 288, + 413 + ], + "type": "inline_equation", + "content": "\\mathbf{w}_j^i" + }, + { + "bbox": [ + 47, + 374, + 288, + 413 + ], + "type": "text", + "content": " is to be kept, " + }, + { + "bbox": [ + 47, + 374, + 288, + 413 + ], + "type": "inline_equation", + "content": "\\mathcal{L}(\\cdot)" + }, + { + "bbox": [ + 47, + 374, + 288, + 413 + ], + "type": "text", + "content": " measures the importance of a filter and " + }, + { + "bbox": [ + 47, + 374, + 288, + 413 + ], + "type": "inline_equation", + "content": "k_l" + }, + { + "bbox": [ + 47, + 374, + 288, + 413 + ], + "type": "text", + "content": " is the kept filter number." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 426, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 426, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 426, + 289, + 715 + ], + "type": "text", + "content": "Robust importance metric for channel pruning The metrics for evaluating the importance of filters is critical. Existing CNN pruning methods design a variety of " + }, + { + "bbox": [ + 46, + 426, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{L}(\\cdot)" + }, + { + "bbox": [ + 46, + 426, + 289, + 715 + ], + "type": "text", + "content": " on the filters. Consider the feature maps, contain rich and important information of both filter and input data, approaches using feature information have become popular and achieved state-of-the-art performance for channel pruning. The results of the feature maps may vary depending on the variability of the input data. Therefore, when the importance of one filter solely depends on the information represented by its own generated feature map, the measurement of the importance may be unstable and sensitive to the slight change of input data. So we have taken into account the characteristics of point clouds data and point-based networks architecture to improve the robustness of channel importance in point-based networks. On the one hand, we propose a coordinate-enhancement module by evaluating the correlation between the feature map and its corresponding points coordinates to guide point clouds network pruning, which will be described in Sec. 3.2. On the other hand, we design a knowledge recycling pruning schema, using discarded points to improve the channel importance evaluation criteria and increase the robustness of the pruning module, which will be described in detail in Sec. 3.3." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 307, + 533, + 320 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 307, + 533, + 320 + ], + "spans": [ + { + "bbox": [ + 306, + 307, + 533, + 320 + ], + "type": "text", + "content": "3.2. Coordinate-Enhanced Channel Importance" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 325, + 545, + 566 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 325, + 545, + 566 + ], + "spans": [ + { + "bbox": [ + 304, + 325, + 545, + 566 + ], + "type": "text", + "content": "Dimensional information is critical in PNNs. The dimensional information (i.e., coordinates of the points) are usually adopted as input for feature extraction. Namely, the input and output of each SA block are concatenated with the coordinates of the points. Meanwhile, the intermediate feature maps reflect not only the information of the original input data but also the corresponding channel information. Therefore, the importance of the channel can be obtained from the feature maps, i.e., the importance of the corresponding channel. The dimensional information is crucial in point-based tasks and should be considered as part of importance metric. Thus the critical problem falls in designing a function that can well reflect the dimensional information richness of feature maps. The feature map, obtained by encoding points spatial " + }, + { + "bbox": [ + 304, + 325, + 545, + 566 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 304, + 325, + 545, + 566 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 325, + 545, + 566 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 304, + 325, + 545, + 566 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 325, + 545, + 566 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 304, + 325, + 545, + 566 + ], + "type": "text", + "content": " coordinates, should be closely related to the original corresponding points coordinates. Therefore, we use the correlation between the current feature map and the corresponding input points coordinates to determine the importance of the filter. The designed Coordinate-Enhancement (CE) module based on Eq. (2):" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 347, + 570, + 545, + 601 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 347, + 570, + 545, + 601 + ], + "spans": [ + { + "bbox": [ + 347, + 570, + 545, + 601 + ], + "type": "interline_equation", + "content": "\\min _ {\\delta_ {i j}} \\sum_ {i = 1} ^ {K} \\sum_ {j = 1} ^ {n _ {i}} \\delta_ {i j} \\mathcal {L} _ {c} \\left(\\mathbf {F} _ {j} ^ {i}\\right), \\text {s . t .} \\sum_ {j = 1} ^ {n _ {i}} \\delta_ {i j} = k _ {l}, \\tag {3}", + "image_path": "d2b5043218042a604e411af8d48897a16bb5efb8185e89575aef87e0f9448899.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 605, + 547, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 547, + 667 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 547, + 667 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 605, + 547, + 667 + ], + "type": "inline_equation", + "content": "\\delta_{ij}" + }, + { + "bbox": [ + 304, + 605, + 547, + 667 + ], + "type": "text", + "content": " is an indicator which is 1 if " + }, + { + "bbox": [ + 304, + 605, + 547, + 667 + ], + "type": "inline_equation", + "content": "\\mathbf{w}_j^i" + }, + { + "bbox": [ + 304, + 605, + 547, + 667 + ], + "type": "text", + "content": " is to be pruned or 0 if " + }, + { + "bbox": [ + 304, + 605, + 547, + 667 + ], + "type": "inline_equation", + "content": "\\mathbf{w}_j^i" + }, + { + "bbox": [ + 304, + 605, + 547, + 667 + ], + "type": "text", + "content": " is to be kept and " + }, + { + "bbox": [ + 304, + 605, + 547, + 667 + ], + "type": "inline_equation", + "content": "k_l" + }, + { + "bbox": [ + 304, + 605, + 547, + 667 + ], + "type": "text", + "content": " is the kept channel number. " + }, + { + "bbox": [ + 304, + 605, + 547, + 667 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_c(\\cdot)" + }, + { + "bbox": [ + 304, + 605, + 547, + 667 + ], + "type": "text", + "content": " measures the importance of a channel from take account of the relationship between feature map and points coordinates which can be formulated as:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 369, + 670, + 545, + 687 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 369, + 670, + 545, + 687 + ], + "spans": [ + { + "bbox": [ + 369, + 670, + 545, + 687 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {c} \\left(F _ {j} ^ {i}\\right) = \\mathcal {M} \\left\\{\\mathrm {C E} \\left(F _ {j} ^ {i}, x ^ {i}\\right) \\right\\}, \\tag {4}", + "image_path": "bc2bb4cea699be5f639ffcd7b54b29363355e8dcb1f5748a0f54f09922fdf772.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 689, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 689, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 689, + 547, + 715 + ], + "type": "text", + "content": "where CE obtains the coordinate-enhanced score by calculating correlation of each channel in the feature map with" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5305" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 70, + 80, + 261, + 173 + ], + "blocks": [ + { + "bbox": [ + 95, + 70, + 236, + 80 + ], + "lines": [ + { + "bbox": [ + 95, + 70, + 236, + 80 + ], + "spans": [ + { + "bbox": [ + 95, + 70, + 236, + 80 + ], + "type": "text", + "content": "Coordinate-Enhancing for Filter Importance" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 70, + 80, + 261, + 173 + ], + "lines": [ + { + "bbox": [ + 70, + 80, + 261, + 173 + ], + "spans": [ + { + "bbox": [ + 70, + 80, + 261, + 173 + ], + "type": "image", + "image_path": "d6e02e42d971152899bb8eb6c23bfaed2cbd9531ba79ab85d19eea7ba094a7c6.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 177, + 287, + 210 + ], + "lines": [ + { + "bbox": [ + 46, + 177, + 287, + 210 + ], + "spans": [ + { + "bbox": [ + 46, + 177, + 287, + 210 + ], + "type": "text", + "content": "Figure 2. Comparisons on the pruned model accuracy with different pruning metrics with CE scores. Results are on ModelNet40 with PointNeXt-S (C=64)." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 226, + 287, + 262 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 226, + 287, + 262 + ], + "spans": [ + { + "bbox": [ + 46, + 226, + 287, + 262 + ], + "type": "text", + "content": "the original coordinates, and " + }, + { + "bbox": [ + 46, + 226, + 287, + 262 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 46, + 226, + 287, + 262 + ], + "type": "text", + "content": " takes the maximum value. Hence, higher coordinate-enhanced score (i.e., " + }, + { + "bbox": [ + 46, + 226, + 287, + 262 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_c" + }, + { + "bbox": [ + 46, + 226, + 287, + 262 + ], + "type": "text", + "content": ") serve as a reliable measurement for information richness." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 262, + 287, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 262, + 287, + 430 + ], + "spans": [ + { + "bbox": [ + 46, + 262, + 287, + 430 + ], + "type": "text", + "content": "We evaluated the effectiveness of the CE module on ModelNet40 with PointNeXt-S (C=64). To demonstrate the experiment's validity, we compared the overall accuracy (OA) for pruning rates of " + }, + { + "bbox": [ + 46, + 262, + 287, + 430 + ], + "type": "inline_equation", + "content": "40\\%" + }, + { + "bbox": [ + 46, + 262, + 287, + 430 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 262, + 287, + 430 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 46, + 262, + 287, + 430 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 262, + 287, + 430 + ], + "type": "inline_equation", + "content": "60\\%" + }, + { + "bbox": [ + 46, + 262, + 287, + 430 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 262, + 287, + 430 + ], + "type": "inline_equation", + "content": "70\\%" + }, + { + "bbox": [ + 46, + 262, + 287, + 430 + ], + "type": "text", + "content": ", respectively. Three sets of experiments are carried out. First, the 'CE' group selected filters in order of value from the CE module. Secondly, the 'Random' group is used to randomly select filters for pruning, and finally, the 'Reverse' is used to select filters according to the coordinate-enhanced score from low to high. As shown in Fig. 2, the channel with a higher coordinate-enhanced score has higher accuracy, which means that the channel with a higher coordinate-enhanced score has higher importance and should be retained in the pruning process." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 438, + 171, + 451 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 438, + 171, + 451 + ], + "spans": [ + { + "bbox": [ + 47, + 438, + 171, + 451 + ], + "type": "text", + "content": "3.3. Knowledge Recycling" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 457, + 287, + 612 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 457, + 287, + 612 + ], + "spans": [ + { + "bbox": [ + 46, + 457, + 287, + 612 + ], + "type": "text", + "content": "Sec. 3.2 shows that feature maps can reflect the importance of the corresponding channels. Another problem is that the importance determination may be unstable and sensitive to small changes in the input data as the feature maps are highly related to the samples of input data. Therefore, we aim to reduce the impact of such data variation. Through the analysis of the PNNs in Sec. 3.1, we found that some points are discarded to obtain hierarchical points set feature. These discarded points are informative as well since the sampling mechanism are highly random and can be leveraged to reduce the impact of data variation. The Knowledge Recycling (KR) module is proposed to reuse the discarded points to improve the robustness of channel pruning." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 613, + 287, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 613, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 46, + 613, + 287, + 689 + ], + "type": "text", + "content": "For those centroids that are computed in " + }, + { + "bbox": [ + 46, + 613, + 287, + 689 + ], + "type": "inline_equation", + "content": "(l - 1)" + }, + { + "bbox": [ + 46, + 613, + 287, + 689 + ], + "type": "text", + "content": "-th layer but discarded in " + }, + { + "bbox": [ + 46, + 613, + 287, + 689 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 46, + 613, + 287, + 689 + ], + "type": "text", + "content": "-th layer due to sampling, which are equivalent to the sampled points. Therefore, the discarded centroids " + }, + { + "bbox": [ + 46, + 613, + 287, + 689 + ], + "type": "inline_equation", + "content": "x_{dis}" + }, + { + "bbox": [ + 46, + 613, + 287, + 689 + ], + "type": "text", + "content": " are fed into " + }, + { + "bbox": [ + 46, + 613, + 287, + 689 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 46, + 613, + 287, + 689 + ], + "type": "text", + "content": "-th convolutional layer to generate the feature map " + }, + { + "bbox": [ + 46, + 613, + 287, + 689 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_{\\mathrm{dis}}^l" + }, + { + "bbox": [ + 46, + 613, + 287, + 689 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 613, + 287, + 689 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_k(F_{dis_j}^i)" + }, + { + "bbox": [ + 46, + 613, + 287, + 689 + ], + "type": "text", + "content": " is taken in use for the evaluation of channel importance." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "content": "We calculate the relevant feature maps from the network parameters trained from the sampled points and use them as" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 72, + 545, + 109 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 109 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 109 + ], + "type": "text", + "content": "part of the importance calculation for the current SA layer channels. Specifically, for each layer in the SA module, we obtain the features of the discard points by Eq. 1:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 338, + 118, + 512, + 138 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 338, + 118, + 512, + 138 + ], + "spans": [ + { + "bbox": [ + 338, + 118, + 512, + 138 + ], + "type": "interline_equation", + "content": "\\mathcal {F} _ {d i s} ^ {l} = \\left\\{\\mathbf {f} _ {\\mathbf {d i s} _ {1}} ^ {l}, \\mathbf {f} _ {\\mathbf {d i s} _ {2}} ^ {l}, \\dots , \\mathbf {f} _ {\\mathbf {d i s} n _ {i}} ^ {l} \\right\\} \\in \\mathbb {R} ^ {n _ {i} \\times (3 + c _ {i})},", + "image_path": "f32e2416cd848cd79102a373aafcad538ac967848f6001137744ab64b0707b2a.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 147, + 545, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 147, + 545, + 171 + ], + "spans": [ + { + "bbox": [ + 305, + 147, + 545, + 171 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 147, + 545, + 171 + ], + "type": "inline_equation", + "content": "n_i" + }, + { + "bbox": [ + 305, + 147, + 545, + 171 + ], + "type": "text", + "content": " is sampled points number, " + }, + { + "bbox": [ + 305, + 147, + 545, + 171 + ], + "type": "inline_equation", + "content": "c_i" + }, + { + "bbox": [ + 305, + 147, + 545, + 171 + ], + "type": "text", + "content": " is the points feature dimension. So the supplement importance is:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 365, + 179, + 545, + 217 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 365, + 179, + 545, + 217 + ], + "spans": [ + { + "bbox": [ + 365, + 179, + 545, + 217 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} _ {k} \\left(F _ {j} ^ {i}\\right) = \\mathcal {L} _ {k} \\left(F _ {\\text {d i s} j} ^ {i}\\right) \\tag {5} \\\\ = \\mathcal {M} \\left\\{\\mathrm {C E} \\left(F _ {j} ^ {i}, x _ {d i s} ^ {i}\\right) \\right\\}, \\\\ \\end{array}", + "image_path": "da3683f25b10de0772daa590750afb972192645382535cf3b34d3d5c02f00d54.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 227, + 479, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 227, + 479, + 239 + ], + "spans": [ + { + "bbox": [ + 305, + 227, + 479, + 239 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 227, + 479, + 239 + ], + "type": "inline_equation", + "content": "x_{dis}^{i}" + }, + { + "bbox": [ + 305, + 227, + 479, + 239 + ], + "type": "text", + "content": " are discard points for recycling." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 239, + 545, + 275 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 239, + 545, + 275 + ], + "spans": [ + { + "bbox": [ + 304, + 239, + 545, + 275 + ], + "type": "text", + "content": "It should be noted that the KR module only needs to calculate " + }, + { + "bbox": [ + 304, + 239, + 545, + 275 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{dis}^{l}" + }, + { + "bbox": [ + 304, + 239, + 545, + 275 + ], + "type": "text", + "content": " from the parameters trained by the sampled points and does not incur much additional overhead." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 283, + 473, + 297 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 283, + 473, + 297 + ], + "spans": [ + { + "bbox": [ + 305, + 283, + 473, + 297 + ], + "type": "text", + "content": "3.4. Using " + }, + { + "bbox": [ + 305, + 283, + 473, + 297 + ], + "type": "inline_equation", + "content": "\\mathbf{CP}^3" + }, + { + "bbox": [ + 305, + 283, + 473, + 297 + ], + "type": "text", + "content": " in Pruning Methods" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 301, + 545, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 301, + 545, + 399 + ], + "spans": [ + { + "bbox": [ + 304, + 301, + 545, + 399 + ], + "type": "text", + "content": "The overall " + }, + { + "bbox": [ + 304, + 301, + 545, + 399 + ], + "type": "inline_equation", + "content": "\\mathrm{CP}^3" + }, + { + "bbox": [ + 304, + 301, + 545, + 399 + ], + "type": "text", + "content": " improve the existing CNNs pruning methods by considering the input data of PNNs and the PNN structure in Sec. 3.2 and Sec. 3.3, respectively. In fact, " + }, + { + "bbox": [ + 304, + 301, + 545, + 399 + ], + "type": "inline_equation", + "content": "\\mathrm{CP}^3" + }, + { + "bbox": [ + 304, + 301, + 545, + 399 + ], + "type": "text", + "content": " can complement the existing pruning methods, i.e., as a plug-in to the existing pruning methods, to improve the pruning performance on PNNs. Specifically, combining Eq. (4) and Eq. (5), we obtain the final pruning formula according to Eq. (2):" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 336, + 407, + 545, + 471 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 336, + 407, + 545, + 471 + ], + "spans": [ + { + "bbox": [ + 336, + 407, + 545, + 471 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\min _ {\\delta_ {i j}} \\sum_ {i = 1} ^ {K} \\sum_ {j = 1} ^ {n _ {i}} \\delta_ {i j} \\left(\\mathcal {L} \\left(\\mathbf {w} _ {j} ^ {i}\\right) + \\mathcal {L} _ {c} \\left(F _ {j} ^ {i}\\right) + \\mathcal {L} _ {k} \\left(F _ {j} ^ {i}\\right)\\right), \\tag {6} \\\\ \\text {s . t .} \\sum_ {j = 1} ^ {n _ {i}} \\delta_ {i j} = k _ {l}, \\\\ \\end{array}", + "image_path": "22370ff039d3ea1747f832fc9ff49af471c1609ca4133024cf0afabcf7cea376.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 480, + 545, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 480, + 545, + 544 + ], + "spans": [ + { + "bbox": [ + 304, + 480, + 545, + 544 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 480, + 545, + 544 + ], + "type": "inline_equation", + "content": "\\delta_{ij}" + }, + { + "bbox": [ + 304, + 480, + 545, + 544 + ], + "type": "text", + "content": " is an indicator which is 1 if " + }, + { + "bbox": [ + 304, + 480, + 545, + 544 + ], + "type": "inline_equation", + "content": "\\mathbf{w}_j^i" + }, + { + "bbox": [ + 304, + 480, + 545, + 544 + ], + "type": "text", + "content": " is to be pruned or 0 if " + }, + { + "bbox": [ + 304, + 480, + 545, + 544 + ], + "type": "inline_equation", + "content": "\\mathbf{w}_j^i" + }, + { + "bbox": [ + 304, + 480, + 545, + 544 + ], + "type": "text", + "content": " is to be kept, " + }, + { + "bbox": [ + 304, + 480, + 545, + 544 + ], + "type": "inline_equation", + "content": "\\mathcal{L}(\\cdot)" + }, + { + "bbox": [ + 304, + 480, + 545, + 544 + ], + "type": "text", + "content": " is original CNNs pruning method measure importance of a channel, " + }, + { + "bbox": [ + 304, + 480, + 545, + 544 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_c" + }, + { + "bbox": [ + 304, + 480, + 545, + 544 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 480, + 545, + 544 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_k" + }, + { + "bbox": [ + 304, + 480, + 545, + 544 + ], + "type": "text", + "content": " are coordinate-enhanced score and knowledge-recycling score, " + }, + { + "bbox": [ + 304, + 480, + 545, + 544 + ], + "type": "inline_equation", + "content": "k_l" + }, + { + "bbox": [ + 304, + 480, + 545, + 544 + ], + "type": "text", + "content": " is the kept channel number." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 306, + 555, + 387, + 568 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 555, + 387, + 568 + ], + "spans": [ + { + "bbox": [ + 306, + 555, + 387, + 568 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 306, + 574, + 432, + 587 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 574, + 432, + 587 + ], + "spans": [ + { + "bbox": [ + 306, + 574, + 432, + 587 + ], + "type": "text", + "content": "4.1. Experimental Settings" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "text", + "content": "Baseline models and datasets To demonstrate the effectiveness and generality of the proposed " + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{CP}^3" + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "text", + "content": ", we tested it on three different 3D tasks and five datasets with various PNNs and three recent advanced channel pruning methods. The evaluated pruning methods include HRank (2020) [24], ResRep (2021) [7], and CHIP (2021) [40]. For the classification task, we chose the classical PointNet++ [33] and PointNeXt [35] models as the original networks and conducted experiments on ModelNet40 [47] and ScanObjectNN [43]. Specifically, for PointNeXt-S we tested" + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5306" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 49, + 108, + 545, + 309 + ], + "blocks": [ + { + "bbox": [ + 46, + 70, + 547, + 105 + ], + "lines": [ + { + "bbox": [ + 46, + 70, + 547, + 105 + ], + "spans": [ + { + "bbox": [ + 46, + 70, + 547, + 105 + ], + "type": "text", + "content": "Table 1. Comparisons of classification on the ModelNet40 [47] test set with PointNet++ [33], PointNeXt-S (C=32) [35], and PointNeXt-S (C=64). For PointNeXt-S (C=32), we report the baseline results from the original paper. For PointNet++ and PointNeXt-S (C=64), we report the baseline results obtained by OpenPoints [35] re-implementations trained with the improved strategies." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 49, + 108, + 545, + 309 + ], + "lines": [ + { + "bbox": [ + 49, + 108, + 545, + 309 + ], + "spans": [ + { + "bbox": [ + 49, + 108, + 545, + 309 + ], + "type": "table", + "html": "
MethodPointNet++PointNeXt-S (C=32)PointNeXt-S (C=64)
OAmAccParams. (M)GFLOPs (↓%)OAmAccParams. (M)GFLOPs (↓%)OAmAccParams. (M)GFLOPs (↓%)
Baseline92.8089.901.471.71 (−)92.9989.61.371.64 (−)93.4491.054.526.49 (−)
HRank92.5989.830.860.77 (55.0)92.8789.970.740.71 (56.7)92.2389.812.122.69 (56.7)
HRank +CP392.9589.910.840.75 (56.1)93.2390.560.710.67 (59.1)93.5290.332.012.58 (59.1)
HRank91.7988.820.590.42 (75.4)92.6389.120.500.39 (76.2)92.7190.451.331.56 (76.2)
HRank +CP392.5488.520.570.39 (77.2)93.0390.920.490.38 (76.8)93.0790.551.281.50 (76.8)
HRank91.3488.180.360.15 (91.2)92.7389.980.430.29 (82.3)92.8389.990.740.71 (82.3)
HRank +CP391.7188.680.340.13 (92.4)92.9990.110.400.27 (83.5)93.1190.520.710.67 (83.5)
ResRep92.7190.390.850.73 (57.3)92.8390.150.810.69 (57.9)91.6188.422.082.34 (57.9)
ResRep+CP393.2790.480.820.70 (59.1)93.3590.930.790.67 (59.1)92.9390.661.892.02 (59.1)
ResRep92.5089.250.570.41 (76.0)92.6490.010.510.40 (75.6)91.6789.271.131.92 (75.6)
ResRep+CP392.4689.430.540.40 (76.6)93.4190.870.490.38 (76.8)93.1190.821.021.89 (76.8)
ResRep92.1189.000.550.24 (86.0))92.3089.310.340.21 (87.2)89.5488.540.720.69 (87.2)
ResRep+CP392.4889.210.580.21 (87.7)92.9590.700.330.18 (89.0)91.0289.820.690.65 (89.0)
CHIP92.7989.230.820.73 (57.3)93.1190.270.710.67 (59.1)93.0390.601.481.44 (59.1)
CHIP+CP392.9990.660.810.70 (59.1)93.3590.800.690.65 (60.4)93.3591.111.451.40 (60.4)
CHIP92.4589.190.570.39 (77.2)92.7189.900.490.38 (76.8)92.7990.390.920.76 (76.8)
CHIP+CP392.9189.650.540.35 (79.5)93.0390.510.480.37 (77.4)93.2390.300.890.74 (77.4)
CHIP92.2689.560.360.15 (91.2)92.4289.130.320.16 (90.2)90.8388.700.650.46 (90.2)
CHIP+CP392.7190.410.340.13 (92.4)92.5090.350.300.14 (91.5)92.8790.250.630.44 (91.5)
", + "image_path": "d903a810808ffa5be20adff5f3999911979761de4397f1cf4c62e5b4d46bcea6.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 329, + 288, + 401 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 329, + 288, + 401 + ], + "spans": [ + { + "bbox": [ + 46, + 329, + 288, + 401 + ], + "type": "text", + "content": "two settings with widths of 32 and 64. For the segmentation task, we conducted experiments on S3DIS [1] with PointNeXt-B and PointNeXt-L [35] as the original PNNs. For the object detection task, we pruned two point-based detectors (VoteNet [31] and GroupFree3D [25]) on SUN RGB-D [38] and ScanNetV2 [5]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 420, + 287, + 577 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 420, + 287, + 577 + ], + "spans": [ + { + "bbox": [ + 46, + 420, + 287, + 577 + ], + "type": "text", + "content": "Implementation details We conducted the classification and segmentation experiments with OpenPoints [35] and the object detection experiments with MMdetection3D [3], all on NVIDIA P100 GPUs. For a fair comparison, we used the same hyperparameter settings for each group of experiments. We either 1) measured the parameter/FLOP reductions of the pruned networks with similar performance or 2) measured the performance of the pruned networks with a similar amount of parameter/FLOP reductions. For all experiments, we reported the number of FLOPs ('GFLOPs') and parameters ('Params.') as well as task-specific metrics to be described in each experiment. More experimental results are available in the supplementary." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 586, + 185, + 598 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 586, + 185, + 598 + ], + "spans": [ + { + "bbox": [ + 47, + 586, + 185, + 598 + ], + "type": "text", + "content": "4.2. Results on Classification" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "text", + "content": "ModelNet40 ModelNet40 [47] contains 9843 training and 2468 testing meshed CAD models belonging to 40 categories. Following the standard practice [33], we report the class-average accuracy (mAcc) and the overall accuracy (OA) on the testing set. We compared the pruned networks directly by HRank, ResRep, and CHIP and those pruned by the three pruning methods with " + }, + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\mathrm{CP}^3" + }, + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "text", + "content": ". As shown in Tab. 1, our " + }, + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\mathrm{CP}^3" + }, + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "text", + "content": " improved the performance of existing CNNs pruning methods for different PNNs with various pruning" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 328, + 547, + 390 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 328, + 547, + 390 + ], + "spans": [ + { + "bbox": [ + 304, + 328, + 547, + 390 + ], + "type": "text", + "content": "rates. " + }, + { + "bbox": [ + 304, + 328, + 547, + 390 + ], + "type": "inline_equation", + "content": "\\mathrm{CP}^3" + }, + { + "bbox": [ + 304, + 328, + 547, + 390 + ], + "type": "text", + "content": " had higher accuracy scores with a similar (and mostly higher) pruning rates. Notably, with the pruning rate of " + }, + { + "bbox": [ + 304, + 328, + 547, + 390 + ], + "type": "inline_equation", + "content": "58\\%" + }, + { + "bbox": [ + 304, + 328, + 547, + 390 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 328, + 547, + 390 + ], + "type": "inline_equation", + "content": "\\mathrm{CP}^3" + }, + { + "bbox": [ + 304, + 328, + 547, + 390 + ], + "type": "text", + "content": " usually produced compact PNNs with even better accuracy scores than the original PNNs, which is difficult for pruning methods without " + }, + { + "bbox": [ + 304, + 328, + 547, + 390 + ], + "type": "inline_equation", + "content": "\\mathrm{CP}^3" + }, + { + "bbox": [ + 304, + 328, + 547, + 390 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 408, + 547, + 600 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 408, + 547, + 600 + ], + "spans": [ + { + "bbox": [ + 304, + 408, + 547, + 600 + ], + "type": "text", + "content": "ScanObjectNN We also conducted experiments on the ScanObjectNN benchmark [43]. ScanObjectNN contains 15000 objects categorized into 15 classes with 2902 unique object instances in the real world. As reported in Tab. 2, " + }, + { + "bbox": [ + 304, + 408, + 547, + 600 + ], + "type": "inline_equation", + "content": "\\mathrm{CP}^3" + }, + { + "bbox": [ + 304, + 408, + 547, + 600 + ], + "type": "text", + "content": " surpassed existing CNN pruning methods directly applied to PNNs. For example, comparing with the baseline pruning method HRank, " + }, + { + "bbox": [ + 304, + 408, + 547, + 600 + ], + "type": "inline_equation", + "content": "\\mathrm{CP}^3" + }, + { + "bbox": [ + 304, + 408, + 547, + 600 + ], + "type": "text", + "content": " boosts the OA score of PointNet++ by " + }, + { + "bbox": [ + 304, + 408, + 547, + 600 + ], + "type": "inline_equation", + "content": "1.14\\%" + }, + { + "bbox": [ + 304, + 408, + 547, + 600 + ], + "type": "text", + "content": " (85.01→86.15), " + }, + { + "bbox": [ + 304, + 408, + 547, + 600 + ], + "type": "inline_equation", + "content": "1.14\\%" + }, + { + "bbox": [ + 304, + 408, + 547, + 600 + ], + "type": "text", + "content": " (84.94→86.08), and " + }, + { + "bbox": [ + 304, + 408, + 547, + 600 + ], + "type": "inline_equation", + "content": "0.59\\%" + }, + { + "bbox": [ + 304, + 408, + 547, + 600 + ], + "type": "text", + "content": " (84.03→84.62) for the three different pruning rates. Similarly, " + }, + { + "bbox": [ + 304, + 408, + 547, + 600 + ], + "type": "inline_equation", + "content": "\\mathrm{CP}^3" + }, + { + "bbox": [ + 304, + 408, + 547, + 600 + ], + "type": "text", + "content": " obtains much higher OA and mAcc scores than the three baselines with different pruning rates on PointNet++ and PointNeXt-S. With the extensive experimental results on classification tasks, we show that " + }, + { + "bbox": [ + 304, + 408, + 547, + 600 + ], + "type": "inline_equation", + "content": "\\mathrm{CP}^3" + }, + { + "bbox": [ + 304, + 408, + 547, + 600 + ], + "type": "text", + "content": " surely improved the pruned network's performance to a distinct extent compared to direct applications of 2D CNN pruning methods." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 610, + 491, + 624 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 610, + 491, + 624 + ], + "spans": [ + { + "bbox": [ + 305, + 610, + 491, + 624 + ], + "type": "text", + "content": "4.3. Results on Semantic Segmentation" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 629, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 629, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 629, + 547, + 715 + ], + "type": "text", + "content": "S3DIS [1] is a challenging benchmark composed of 6 large-scale indoor areas, 271 rooms, and 13 semantic categories in total. Following a common protocol [41], we evaluated the presented approach in Area-5, which means to test on Area-5 and to train on the rest. For evaluation metrics, we used the mean classwise intersection over union (mIoU), the mean classwise accuracy (mAcc), and the overall point" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 758 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 758 + ], + "type": "text", + "content": "5307" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 49, + 108, + 545, + 309 + ], + "blocks": [ + { + "bbox": [ + 46, + 70, + 547, + 105 + ], + "lines": [ + { + "bbox": [ + 46, + 70, + 547, + 105 + ], + "spans": [ + { + "bbox": [ + 46, + 70, + 547, + 105 + ], + "type": "text", + "content": "Table 2. Comparisons of classification on the ScanObjectNN [43] test set with PointNet++ [33], PointNeXt-S (C=32) [35], and PointNeXt-S (C=64). For PointNeXt-S (C=32), we report the baseline results from the original paper. For PointNet++ and PointNeXt-S (C=64), we report the baseline results obtained by OpenPoints [35] re-implementations trained with the improved strategies." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 49, + 108, + 545, + 309 + ], + "lines": [ + { + "bbox": [ + 49, + 108, + 545, + 309 + ], + "spans": [ + { + "bbox": [ + 49, + 108, + 545, + 309 + ], + "type": "table", + "html": "
MethodPointNet++PointNeXt-S (C=32)PointNeXt-S (C=64)
OAmAccParams. (M)GFLOPs (↓%)OAmAccParams. (M)GFLOPs (↓%)OAmAccParams. (M)GFLOPs (↓%)
Baseline86.2084.401.471.71 (−)87.4085.391.371.64 (−)88.2086.844.526.49 (−)
HRank85.0183.330.820.73 (57.3)87.0285.850.720.69 (57.9)87.5185.482.222.91 (55.2)
HRank+CP386.1584.400.800.70 (59.1)87.4786.210.700.67 (59.1)87.9586.022.162.82 (56.5)
HRank84.9483.430.590.43 (74.9)84.7981.930.500.39 (76.2)84.6681.001.331.56 (76.0)
HRank+CP386.0884.510.590.41 (76.0)86.4083.940.480.37 (77.4)86.4384.941.281.50 (76.9)
HRank84.0382.160.370.16 (90.6)81.3378.320.320.17 (89.6)85.3983.830.730.71 (89.1)
HRank+CP384.6282.290.360.15 (91.2)84.2182.130.310.16 (90.2)86.2684.360.700.67 (89.7)
ResRep86.7784.780.820.76 (55.6)86.3284.450.690.65 (60.4)86.5884.022.122.82 (56.5)
ResRep+CP387.1185.500.820.71 (58.5)86.6684.750.670.64 (61.0)88.5286.002.032.74 (57.8)
ResRep83.9982.340.600.44 (74.3)85.2883.270.480.38 (76.8)85.2282.031.391.65 (74.6)
ResRep+CP384.3283.680.590.43 (74.9)86.2284.320.470.37 (77.4)86.7384.231.321.54 (76.3)
ResRep83.7981.910.400.29 (83.0)83.6681.660.410.25 (84.8)85.0281.270.600.68 (89.5)
ResRep+CP384.8082.830.400.26 (84.8)85.0183.030.400.23 (86.0)86.1383.400.580.52 (92.0)
CHIP86.1484.980.820.73 (57.3)87.1385.090.700.67 (59.1)88.4587.402.112.74 (57.8)
CHIP+CP386.2584.680.800.70 (59.1)87.5485.670.690.65 (60.4)88.5886.452.052.65 (59.2)
CHIP84.4582.250.590.42 (75.4)85.2883.280.500.4 (75.6)86.6885.121.371.64 (74.7)
CHIP+CP385.5984.420.570.40 (76.6)86.2983.990.490.38 (76.8)87.7986.811.321.56 (76.0)
CHIP83.2781.150.360.15 (91.2)81.3778.990.340.19 (88.4)83.9081.830.440.32 (95.1)
CHIP+CP384.0382.220.350.14 (91.8)82.1279.410.330.18 (89.0)84.2582.170.420.29 (95.5)
", + "image_path": "c32ffb58de38c076cb4eff328015779d5b80f96e7ad439c04ead9f3015ae68d9.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 75, + 335, + 520, + 463 + ], + "blocks": [ + { + "bbox": [ + 55, + 319, + 538, + 331 + ], + "lines": [ + { + "bbox": [ + 55, + 319, + 538, + 331 + ], + "spans": [ + { + "bbox": [ + 55, + 319, + 538, + 331 + ], + "type": "text", + "content": "Table 3. Comparisons of semantic segmentation on the S3DIS dataset (evaluated in Area-5) with PointNeXt-B and PointNeXt-L [35]." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 75, + 335, + 520, + 463 + ], + "lines": [ + { + "bbox": [ + 75, + 335, + 520, + 463 + ], + "spans": [ + { + "bbox": [ + 75, + 335, + 520, + 463 + ], + "type": "table", + "html": "
MethodPointNeXt-BPointNeXt-L
OAmAccmIoUParams. (M)GFLOPs (↓%)OAmAccmIoUParams. (M)GFLOPs (↓%)
Baseline89.4073.9067.503.838.80 (-)90.1075.7069.307.1315.24 (-)
HRank89.0472.1465.661.724.01 (54.4)88.8873.6166.803.206.83 (55.2)
HRank+CP389.2473.7666.951.613.80 (56.8)89.4474.2767.533.006.48 (57.5)
HRank88.8172.1665.580.852.04 (76.8)88.2170.9364.301.583.47 (77.2)
HRank+CP389.0272.8266.410.781.85 (78.9)88.3571.2664.711.443.14 (79.4)
CHIP88.8973.2666.571.663.93 (55.3)89.1673.7267.093.096.68 (56.2)
CHIP+CP389.6873.1466.801.563.65 (58.5)89.4774.2067.282.916.22 (59.2)
CHIP88.6773.2466.680.811.92 (78.2)88.5871.5865.181.503.27 (78.5)
CHIP+CP389.8173.3666.950.741.79 (79.7)89.2071.6665.241.383.04 (80.1)
", + "image_path": "37bc540125a381ca2b601431a899cece853123c8738147b065a32d5d163166c2.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 50, + 496, + 285, + 561 + ], + "blocks": [ + { + "bbox": [ + 48, + 481, + 285, + 492 + ], + "lines": [ + { + "bbox": [ + 48, + 481, + 285, + 492 + ], + "spans": [ + { + "bbox": [ + 48, + 481, + 285, + 492 + ], + "type": "text", + "content": "Table 4. Comparisons of object detection on the ScanNet dataset." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 50, + 496, + 285, + 561 + ], + "lines": [ + { + "bbox": [ + 50, + 496, + 285, + 561 + ], + "spans": [ + { + "bbox": [ + 50, + 496, + 285, + 561 + ], + "type": "table", + "html": "
MethodmAP@0.25mAP@0.50Params. (K)GFLOPs (↓%)
Baseline (VoteNet)62.3440.82641.925.78 (−)
ResRep62.4540.95251.232.45 (57.6)
ResRep+CP363.9241.47242.262.41 (58.1)
ResRep61.7840.54180.491.83 (68.3)
ResRep+CP362.9840.94160.481.78 (69.2)
", + "image_path": "a9110dbc115029590592b206cc65db3176840b3a33f994ea2c7465ed9ede1bb6.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "text", + "content": "wise accuracy (OA). As the segmentation task is relatively difficult and the segmentation network structures are relatively complex, we pruned only the encoder part of the network and kept the original decoder part. The results are presented in Tab. 3. As expected, the performance of the pruned networks degraded more from the original networks than those in the classification experiments, but the performance was acceptable. Meanwhile, in all cases, with our " + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\mathrm{CP}^3" + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "text", + "content": ", PNNs have a higher accuracy at a higher pruning rate than without " + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\mathrm{CP}^3" + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "text", + "content": ". For example, for PointNeXt-B, comparing with directly applying CHIP, incorporating " + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\mathrm{CP}^3" + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 309, + 496, + 543, + 561 + ], + "blocks": [ + { + "bbox": [ + 318, + 481, + 533, + 492 + ], + "lines": [ + { + "bbox": [ + 318, + 481, + 533, + 492 + ], + "spans": [ + { + "bbox": [ + 318, + 481, + 533, + 492 + ], + "type": "text", + "content": "Table 5. Comparisons of object detection on SUN RGB-D." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 309, + 496, + 543, + 561 + ], + "lines": [ + { + "bbox": [ + 309, + 496, + 543, + 561 + ], + "spans": [ + { + "bbox": [ + 309, + 496, + 543, + 561 + ], + "type": "table", + "html": "
MethodmAP@0.25mAP@0.50Params. (K)GFLOPs (↓%)
Baseline (VoteNet)59.7835.77641.925.78 (-)
ResRep59.3736.80179.912.42 (58.13)
ResRep+CP360.1037.37172.352.20 (61.93)
ResRep59.0135.91135.131.84 (68.17)
ResRep+CP359.1836.24129.641.83 (68.34)
", + "image_path": "0b34a508152e7039850b67eeac9c800a37b7e17337f048d20860f21f9eb35b3a.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 582, + 545, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 582, + 545, + 628 + ], + "spans": [ + { + "bbox": [ + 304, + 582, + 545, + 628 + ], + "type": "text", + "content": "obtained much higher OA, mAcc and mIoU scores (1.2% on OA and 0.3% on mIoU). The results about segmentation have well validated the generalization of " + }, + { + "bbox": [ + 304, + 582, + 545, + 628 + ], + "type": "inline_equation", + "content": "\\mathrm{CP}^3" + }, + { + "bbox": [ + 304, + 582, + 545, + 628 + ], + "type": "text", + "content": " to new and difficult tasks." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 639, + 476, + 651 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 639, + 476, + 651 + ], + "spans": [ + { + "bbox": [ + 306, + 639, + 476, + 651 + ], + "type": "text", + "content": "4.4. Results on 3D Object Detection" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 658, + 507, + 670 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 658, + 507, + 670 + ], + "spans": [ + { + "bbox": [ + 306, + 658, + 507, + 670 + ], + "type": "text", + "content": "4.4.1 Evaluation and Comparison of VoteNet" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 677, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 547, + 715 + ], + "type": "text", + "content": "Tabs. 4 and 5 show the results of the pruned VoteNet models on the ScanNetV2 and SUN RGB-D datasets, respectively. We evaluated the performance of our proposed method in" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5308" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 48, + 96, + 286, + 166 + ], + "blocks": [ + { + "bbox": [ + 47, + 70, + 286, + 92 + ], + "lines": [ + { + "bbox": [ + 47, + 70, + 286, + 92 + ], + "spans": [ + { + "bbox": [ + 47, + 70, + 286, + 92 + ], + "type": "text", + "content": "Table 6. Comparisons of object detection on the ScanNet dataset. The baseline PNN model is GroupFree3D." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 48, + 96, + 286, + 166 + ], + "lines": [ + { + "bbox": [ + 48, + 96, + 286, + 166 + ], + "spans": [ + { + "bbox": [ + 48, + 96, + 286, + 166 + ], + "type": "table", + "html": "
MethodmAP@0.25mAP@0.50Params. (K)GFLOPs (↓%)
Baseline68.2252.612438.3421.78 (-)
ResRep68.2451.481910.7310.90 (49.95)
ResRep+CP368.8652.081654.3510.88 (50.05)
ResRep67.2151.281703.348.71 (60.01)
ResRep+CP368.5751.851501.468.68 (60.14)
", + "image_path": "2638ef9571e7a1e22007a34dc06e2208c14ead28e6bdf2e2d3b6c7d371a142c7.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 185, + 287, + 209 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 185, + 287, + 209 + ], + "spans": [ + { + "bbox": [ + 46, + 185, + 287, + 209 + ], + "type": "text", + "content": "terms of the mean average precision at IOU thresholds of 0.25 and 0.50 (mAP@25 and mAP@50)." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 223, + 287, + 390 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 223, + 287, + 390 + ], + "spans": [ + { + "bbox": [ + 46, + 223, + 287, + 390 + ], + "type": "text", + "content": "ScanNetV2 ScanNetV2 [5] is a richly annotated dataset of 3D reconstructed meshes of indoor scenes. It contains about 1200 training examples collected from hundreds of different rooms and is annotated with semantic and instance segmentation for 18 object categories. Tab. 4 shows the results of directly applying ResRep and with " + }, + { + "bbox": [ + 46, + 223, + 287, + 390 + ], + "type": "inline_equation", + "content": "\\mathrm{CP}^3" + }, + { + "bbox": [ + 46, + 223, + 287, + 390 + ], + "type": "text", + "content": ". As can be seen from the table, the accuracy of the 2D method directly applied to the 3D network decreased by a flops drop of about " + }, + { + "bbox": [ + 46, + 223, + 287, + 390 + ], + "type": "inline_equation", + "content": "60\\%" + }, + { + "bbox": [ + 46, + 223, + 287, + 390 + ], + "type": "text", + "content": ", while our method achieves " + }, + { + "bbox": [ + 46, + 223, + 287, + 390 + ], + "type": "inline_equation", + "content": "1.58\\%" + }, + { + "bbox": [ + 46, + 223, + 287, + 390 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 223, + 287, + 390 + ], + "type": "inline_equation", + "content": "0.65\\%" + }, + { + "bbox": [ + 46, + 223, + 287, + 390 + ], + "type": "text", + "content": " improvements at mAP@0.25 and mAP@0.5 with a drop rate of " + }, + { + "bbox": [ + 46, + 223, + 287, + 390 + ], + "type": "inline_equation", + "content": "58.13\\%" + }, + { + "bbox": [ + 46, + 223, + 287, + 390 + ], + "type": "text", + "content": " FLOPS. When FLOPs drop to about " + }, + { + "bbox": [ + 46, + 223, + 287, + 390 + ], + "type": "inline_equation", + "content": "70\\%" + }, + { + "bbox": [ + 46, + 223, + 287, + 390 + ], + "type": "text", + "content": ", the accuracy of the direct porting CNNs pruning method works poorly, while the improvement of mAP@0.25 and mAP@0.5 of our method is " + }, + { + "bbox": [ + 46, + 223, + 287, + 390 + ], + "type": "inline_equation", + "content": "0.64\\%" + }, + { + "bbox": [ + 46, + 223, + 287, + 390 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 223, + 287, + 390 + ], + "type": "inline_equation", + "content": "0.12\\%" + }, + { + "bbox": [ + 46, + 223, + 287, + 390 + ], + "type": "text", + "content": ", respectively." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 404, + 287, + 561 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 404, + 287, + 561 + ], + "spans": [ + { + "bbox": [ + 46, + 404, + 287, + 561 + ], + "type": "text", + "content": "SUN RGB-D The SUN RGB-D dataset [38] consists of 10355 single-view indoor RGB-D images annotated with over 64000 3D bounding boxes and semantic labels for 37 categories. We conducted experiments on SUN RGB-D with the same setup as those on ScanNetV2. The findings are also similar to those in ScanNetV2. It can be observed from Tab. 5 that the accuracy of directly transplanted CNNs pruning method is reduced to some extent (mAP@0.25 reduced by 0.41) when FLOPs drop by " + }, + { + "bbox": [ + 46, + 404, + 287, + 561 + ], + "type": "inline_equation", + "content": "58.13\\%" + }, + { + "bbox": [ + 46, + 404, + 287, + 561 + ], + "type": "text", + "content": ", while " + }, + { + "bbox": [ + 46, + 404, + 287, + 561 + ], + "type": "inline_equation", + "content": "\\mathrm{CP}^3" + }, + { + "bbox": [ + 46, + 404, + 287, + 561 + ], + "type": "text", + "content": " improved the detection model by " + }, + { + "bbox": [ + 46, + 404, + 287, + 561 + ], + "type": "inline_equation", + "content": "0.32\\%" + }, + { + "bbox": [ + 46, + 404, + 287, + 561 + ], + "type": "text", + "content": " mAP@0.25 and " + }, + { + "bbox": [ + 46, + 404, + 287, + 561 + ], + "type": "inline_equation", + "content": "1.60\\%" + }, + { + "bbox": [ + 46, + 404, + 287, + 561 + ], + "type": "text", + "content": " mAP@0.5 with a FLOPs drop of " + }, + { + "bbox": [ + 46, + 404, + 287, + 561 + ], + "type": "inline_equation", + "content": "61.93\\%" + }, + { + "bbox": [ + 46, + 404, + 287, + 561 + ], + "type": "text", + "content": ". Even when FLOPs drop to " + }, + { + "bbox": [ + 46, + 404, + 287, + 561 + ], + "type": "inline_equation", + "content": "70\\%" + }, + { + "bbox": [ + 46, + 404, + 287, + 561 + ], + "type": "text", + "content": ", our method's mAP@0.25 drop only " + }, + { + "bbox": [ + 46, + 404, + 287, + 561 + ], + "type": "inline_equation", + "content": "0.6\\%" + }, + { + "bbox": [ + 46, + 404, + 287, + 561 + ], + "type": "text", + "content": ", which is obviously better." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 574, + 275, + 588 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 574, + 275, + 588 + ], + "spans": [ + { + "bbox": [ + 47, + 574, + 275, + 588 + ], + "type": "text", + "content": "4.4.2 Evaluation and Comparison on GroupFree3D" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 594, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 287, + 714 + ], + "type": "text", + "content": "We also conducted experiments on another point-based 3D detection model, GroupFree3D, on ScanNetV2. Tab. 6 summarizes the pruning performance of our approach for GroupFree3D on the ScanNetV2 dataset. When targeting a moderate compression ratio, our approach can achieve " + }, + { + "bbox": [ + 46, + 594, + 287, + 714 + ], + "type": "inline_equation", + "content": "32.15\\%" + }, + { + "bbox": [ + 46, + 594, + 287, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 594, + 287, + 714 + ], + "type": "inline_equation", + "content": "50.05\\%" + }, + { + "bbox": [ + 46, + 594, + 287, + 714 + ], + "type": "text", + "content": " storage and computation reductions, respectively, with a " + }, + { + "bbox": [ + 46, + 594, + 287, + 714 + ], + "type": "inline_equation", + "content": "0.64\\%" + }, + { + "bbox": [ + 46, + 594, + 287, + 714 + ], + "type": "text", + "content": " accuracy increase for mAP@0.25 over the baseline model. In the case of higher compression ratio, " + }, + { + "bbox": [ + 46, + 594, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\mathrm{CP}^3" + }, + { + "bbox": [ + 46, + 594, + 287, + 714 + ], + "type": "text", + "content": " still achieves superior performance to other methods. Specifically, the ResRep loses accuracy" + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 326, + 118, + 526, + 229 + ], + "blocks": [ + { + "bbox": [ + 304, + 70, + 545, + 115 + ], + "lines": [ + { + "bbox": [ + 304, + 70, + 545, + 115 + ], + "spans": [ + { + "bbox": [ + 304, + 70, + 545, + 115 + ], + "type": "text", + "content": "Table 7. Ablation study of different components in " + }, + { + "bbox": [ + 304, + 70, + 545, + 115 + ], + "type": "inline_equation", + "content": "\\mathrm{CP}^3" + }, + { + "bbox": [ + 304, + 70, + 545, + 115 + ], + "type": "text", + "content": ". Results are of classification on the ScanObjectNN dataset with PointNeXt-S (C=32) as the baseline. 'CE' represents the coordinate-enhanced module, and 'KR' represents the knowledge recycling module." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 326, + 118, + 526, + 229 + ], + "lines": [ + { + "bbox": [ + 326, + 118, + 526, + 229 + ], + "spans": [ + { + "bbox": [ + 326, + 118, + 526, + 229 + ], + "type": "table", + "html": "
SettingCEKRPruning RateOAmAcc
Baseline-88.2086.40
HRank0.7584.7981.93
✓0.7585.6382.97
✓0.7585.1182.13
HRank+CP3✓✓0.7586.6383.63
HRank0.9081.3378.32
✓0.9083.6681.32
✓0.9083.1080.47
HRank+CP3✓✓0.9084.8382.74
", + "image_path": "dc0c9fe09136d8ad6d86a4fccd1996da0214aeee20dbd1400cd015aee157f51e.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 250, + 545, + 285 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 250, + 545, + 285 + ], + "spans": [ + { + "bbox": [ + 304, + 250, + 545, + 285 + ], + "type": "text", + "content": "by " + }, + { + "bbox": [ + 304, + 250, + 545, + 285 + ], + "type": "inline_equation", + "content": "1.01\\%" + }, + { + "bbox": [ + 304, + 250, + 545, + 285 + ], + "type": "text", + "content": " mAP@0.25 when the parameters and flops drop by " + }, + { + "bbox": [ + 304, + 250, + 545, + 285 + ], + "type": "inline_equation", + "content": "30.14\\%" + }, + { + "bbox": [ + 304, + 250, + 545, + 285 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 250, + 545, + 285 + ], + "type": "inline_equation", + "content": "60.01\\%" + }, + { + "bbox": [ + 304, + 250, + 545, + 285 + ], + "type": "text", + "content": ", while in our method, the accuracy increases " + }, + { + "bbox": [ + 304, + 250, + 545, + 285 + ], + "type": "inline_equation", + "content": "0.35\\%" + }, + { + "bbox": [ + 304, + 250, + 545, + 285 + ], + "type": "text", + "content": " for mAP@0.25." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 294, + 406, + 306 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 294, + 406, + 306 + ], + "spans": [ + { + "bbox": [ + 306, + 294, + 406, + 306 + ], + "type": "text", + "content": "4.5. Ablation Studies" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 313, + 545, + 493 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 313, + 545, + 493 + ], + "spans": [ + { + "bbox": [ + 304, + 313, + 545, + 493 + ], + "type": "text", + "content": "We conducted ablation studies to validate the Coordinate-Enhancement (CE) module and the KnowledgeRecycling (KR) module in " + }, + { + "bbox": [ + 304, + 313, + 545, + 493 + ], + "type": "inline_equation", + "content": "\\mathrm{CP}^3" + }, + { + "bbox": [ + 304, + 313, + 545, + 493 + ], + "type": "text", + "content": ". All results provided in the section are tested on ScanObjectNN with PointNeXt-S " + }, + { + "bbox": [ + 304, + 313, + 545, + 493 + ], + "type": "inline_equation", + "content": "(C = 32)" + }, + { + "bbox": [ + 304, + 313, + 545, + 493 + ], + "type": "text", + "content": " as the baseline and HRank as the pruning method. We evaluated the pruned networks' performance with the FLOPs drop of " + }, + { + "bbox": [ + 304, + 313, + 545, + 493 + ], + "type": "inline_equation", + "content": "75\\%" + }, + { + "bbox": [ + 304, + 313, + 545, + 493 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 313, + 545, + 493 + ], + "type": "inline_equation", + "content": "90\\%" + }, + { + "bbox": [ + 304, + 313, + 545, + 493 + ], + "type": "text", + "content": ", and with/without the CE and KR modules to the pruning method. From Tab. 7, we find that KR improves OA of " + }, + { + "bbox": [ + 304, + 313, + 545, + 493 + ], + "type": "inline_equation", + "content": "0.84\\%" + }, + { + "bbox": [ + 304, + 313, + 545, + 493 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 313, + 545, + 493 + ], + "type": "inline_equation", + "content": "2.33\\%" + }, + { + "bbox": [ + 304, + 313, + 545, + 493 + ], + "type": "text", + "content": " when pruning rate is " + }, + { + "bbox": [ + 304, + 313, + 545, + 493 + ], + "type": "inline_equation", + "content": "75\\%" + }, + { + "bbox": [ + 304, + 313, + 545, + 493 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 313, + 545, + 493 + ], + "type": "inline_equation", + "content": "90\\%" + }, + { + "bbox": [ + 304, + 313, + 545, + 493 + ], + "type": "text", + "content": ", respectively, while CE improves OA of " + }, + { + "bbox": [ + 304, + 313, + 545, + 493 + ], + "type": "inline_equation", + "content": "0.32\\%" + }, + { + "bbox": [ + 304, + 313, + 545, + 493 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 313, + 545, + 493 + ], + "type": "inline_equation", + "content": "1.77\\%" + }, + { + "bbox": [ + 304, + 313, + 545, + 493 + ], + "type": "text", + "content": " at " + }, + { + "bbox": [ + 304, + 313, + 545, + 493 + ], + "type": "inline_equation", + "content": "75\\%" + }, + { + "bbox": [ + 304, + 313, + 545, + 493 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 313, + 545, + 493 + ], + "type": "inline_equation", + "content": "90\\%" + }, + { + "bbox": [ + 304, + 313, + 545, + 493 + ], + "type": "text", + "content": " pruning rates, respectively. Bringing the two modules together, the OA improvement is " + }, + { + "bbox": [ + 304, + 313, + 545, + 493 + ], + "type": "inline_equation", + "content": "1.84\\%" + }, + { + "bbox": [ + 304, + 313, + 545, + 493 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 313, + 545, + 493 + ], + "type": "inline_equation", + "content": "3.50\\%" + }, + { + "bbox": [ + 304, + 313, + 545, + 493 + ], + "type": "text", + "content": " at " + }, + { + "bbox": [ + 304, + 313, + 545, + 493 + ], + "type": "inline_equation", + "content": "75\\%" + }, + { + "bbox": [ + 304, + 313, + 545, + 493 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 313, + 545, + 493 + ], + "type": "inline_equation", + "content": "90\\%" + }, + { + "bbox": [ + 304, + 313, + 545, + 493 + ], + "type": "text", + "content": " pruning rates, respectively. The ablation study results have validated the effectiveness of all designs in " + }, + { + "bbox": [ + 304, + 313, + 545, + 493 + ], + "type": "inline_equation", + "content": "\\mathrm{CP}^3" + }, + { + "bbox": [ + 304, + 313, + 545, + 493 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 505, + 378, + 517 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 505, + 378, + 517 + ], + "spans": [ + { + "bbox": [ + 306, + 505, + 378, + 517 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 525, + 545, + 621 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 525, + 545, + 621 + ], + "spans": [ + { + "bbox": [ + 304, + 525, + 545, + 621 + ], + "type": "text", + "content": "In this paper, we focus on 3D point-based network pruning and design a 3D channel pruning plug-in " + }, + { + "bbox": [ + 304, + 525, + 545, + 621 + ], + "type": "inline_equation", + "content": "(\\mathrm{CP}^3)" + }, + { + "bbox": [ + 304, + 525, + 545, + 621 + ], + "type": "text", + "content": " that can be used with existing 2D CNN pruning methods. To the best of our knowledge, this is the first pruning work explicitly considering the characteristics of point cloud data and point-based networks. Empirically, we show that the proposed " + }, + { + "bbox": [ + 304, + 525, + 545, + 621 + ], + "type": "inline_equation", + "content": "\\mathrm{CP}^3" + }, + { + "bbox": [ + 304, + 525, + 545, + 621 + ], + "type": "text", + "content": " is universally effective for a wide range of point-based networks and 3D tasks." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 633, + 403, + 646 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 633, + 403, + 646 + ], + "spans": [ + { + "bbox": [ + 306, + 633, + 403, + 646 + ], + "type": "text", + "content": "Acknowledgement" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "type": "text", + "content": "This work was partially supported by National Natural Science Foundation of China (NSFC) under Grant 62271203, the Science and Technology Innovation Action Plan of Shanghai under Grant 2251110540, and Shanghai Pujiang Program under Grant 21PJ1420300." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5309" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "type": "text", + "content": "[1] Iro Armeni, Ozan Sener, Amir R. Zamir, Helen Jiang, Ioannis K. Brilakis, Martin Fischer, and Silvio Savarese. 3D semantic parsing of large-scale indoor spaces. In IEEE Conference on Computer Vision and Pattern Recognition, CVPR, pages 1534-1543, 2016. 6, 12" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 147, + 288, + 191 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 147, + 288, + 191 + ], + "spans": [ + { + "bbox": [ + 53, + 147, + 288, + 191 + ], + "type": "text", + "content": "[2] Chen Chen, Zhe Chen, Jing Zhang, and Dacheng Tao. SASA: semantics-augmented set abstraction for point-based 3D object detection. In Thirty-Sixth AAAI Conference on Artificial Intelligence, AAAI, pages 221-229, 2022. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 192, + 287, + 235 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 192, + 287, + 235 + ], + "spans": [ + { + "bbox": [ + 53, + 192, + 287, + 235 + ], + "type": "text", + "content": "[3] MMDetection3D Contributors. MMDetection3D: Open-MMLab next-generation platform for general 3D object detection. https://github.com/open-mmlab/mmdetection3d, 2020.6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 236, + 287, + 290 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 236, + 287, + 290 + ], + "spans": [ + { + "bbox": [ + 53, + 236, + 287, + 290 + ], + "type": "text", + "content": "[4] Angela Dai, Angel X Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3D reconstructions of indoor scenes. In IEEE Conference on Computer Vision and Pattern Recognition, pages 5828-5839, 2017. 12" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 291, + 287, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 291, + 287, + 346 + ], + "spans": [ + { + "bbox": [ + 53, + 291, + 287, + 346 + ], + "type": "text", + "content": "[5] Angela Dai, Angel X. Chang, Manolis Savva, Maciej Halber, Thomas A. Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3D reconstructions of indoor scenes. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR, pages 2432-2443, 2017. 6, 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 347, + 287, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 347, + 287, + 402 + ], + "spans": [ + { + "bbox": [ + 53, + 347, + 287, + 402 + ], + "type": "text", + "content": "[6] Xiaohan Ding, Guiguang Ding, Yuchen Guo, and Jungong Han. Centripetal sgd for pruning very deep convolutional networks with complicated structure. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 4943-4953, 2019. 1, 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 403, + 287, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 403, + 287, + 456 + ], + "spans": [ + { + "bbox": [ + 53, + 403, + 287, + 456 + ], + "type": "text", + "content": "[7] Xiaohan Ding, Tianxiang Hao, Jianchao Tan, Ji Liu, Jungong Han, Yuchen Guo, and Guiguang Ding. ResRep: Lossless CNN pruning via decoupling remembering and forgetting. In IEEE/CVF International Conference on Computer Vision, ICCV, pages 4510-4520, 2021. 1, 3, 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 458, + 287, + 513 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 458, + 287, + 513 + ], + "spans": [ + { + "bbox": [ + 53, + 458, + 287, + 513 + ], + "type": "text", + "content": "[8] Anand Dubey, Avik Santra, Jonas Fuchs, Maximilian Lübke, Robert Weigel, and Fabian Lurz. Haradnet: Anchor-free target detection for radar point clouds using hierarchical attention and multi-task learning. Machine Learning with Applications, 8:100275, 2022. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 514, + 287, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 514, + 287, + 568 + ], + "spans": [ + { + "bbox": [ + 53, + 514, + 287, + 568 + ], + "type": "text", + "content": "[9] Hamidreza Fazlali, Yixuan Xu, Yuan Ren, and Bingbing Liu. A versatile multi-view framework for LiDAR-based 3D object detection with guidance from panoptic segmentation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR, pages 17171-17180, 2022. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 569, + 287, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 569, + 287, + 635 + ], + "spans": [ + { + "bbox": [ + 48, + 569, + 287, + 635 + ], + "type": "text", + "content": "[10] Kehong Gong, Bingbing Li, Jianfeng Zhang, Tao Wang, Jing Huang, Michael Bi Mi, Jiashi Feng, and Xinchao Wang. PoseTriplet: Co-evolving 3d human pose estimation, imitation, and hallucination under self-supervision. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR, pages 11017-11027, 2022. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 635, + 287, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 635, + 287, + 657 + ], + "spans": [ + { + "bbox": [ + 48, + 635, + 287, + 657 + ], + "type": "text", + "content": "[11] Google. TensorFlow lite. https://www.tensorflow.org/ lite, 2017. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 658, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 287, + 713 + ], + "type": "text", + "content": "[12] Yushuo Guan, Ning Liu, Pengyu Zhao, Zhengping Che, Kaigui Bian, Yanzhi Wang, and Jian Tang. DAIS: Automatic channel pruning via differentiable annealing indicator search. IEEE Transactions on Neural Networks and Learning Systems, TNNLS, pages 1-12, 2022. 1, 3" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 117 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 117 + ], + "type": "text", + "content": "[13] Kai Han, Yunhe Wang, Qiulin Zhang, Wei Zhang, Chunjing Xu, and Tong Zhang. Model rubik's cube: Twisting resolution, depth and width for tinynets. Advances in Neural Information Processing Systems, 33:19353-19364, 2020. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 118, + 545, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 545, + 161 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 545, + 161 + ], + "type": "text", + "content": "[14] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In IEEE/CVF conference on Computer Vision and Pattern Recognition, CVPR, pages 770-778, 2016. 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 162, + 545, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 162, + 545, + 226 + ], + "spans": [ + { + "bbox": [ + 308, + 162, + 545, + 226 + ], + "type": "text", + "content": "[15] Zejiang Hou, Minghai Qin, Fei Sun, Xiaolong Ma, Kun Yuan, Yi Xu, Yen-Kuang Chen, Rong Jin, Yuan Xie, and Sun-Yuan Kung. CHEX: channel exploration for CNN model compression. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR, pages 12277-12288, 2022. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 228, + 545, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 228, + 545, + 281 + ], + "spans": [ + { + "bbox": [ + 308, + 228, + 545, + 281 + ], + "type": "text", + "content": "[16] Andrew G. Howard, Menglong Zhu, Bo Chen, Dmitry Kalenichenko, Weijun Wang, Tobias Weyand, Marco Andretto, and Hartwig Adam. MobileNets: Efficient convolutional neural networks for mobile vision applications. CoRR, abs/1704.04861, 2017. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 283, + 545, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 283, + 545, + 327 + ], + "spans": [ + { + "bbox": [ + 308, + 283, + 545, + 327 + ], + "type": "text", + "content": "[17] Fengwei Jia, Xuan Wang, Jian Guan, Huale Li, Chen Qiu, and Shuhan Qi. Arank: Toward specific model pruning via advantage rank for multiple salient objects detection. Image and Vision Computing, 111:104192, 2021. 12" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 327, + 545, + 382 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 327, + 545, + 382 + ], + "spans": [ + { + "bbox": [ + 308, + 327, + 545, + 382 + ], + "type": "text", + "content": "[18] Xiaotang Jiang, Huan Wang, Yiliu Chen, Ziqi Wu, Lichuan Wang, Bin Zou, Yafeng Yang, Zongyang Cui, Yu Cai, Tianhang Yu, et al. MNN: A universal and efficient inference engine. The Conference on Machine Learning and Systems, MLSys, pages 1-13, 2020. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 383, + 545, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 383, + 545, + 426 + ], + "spans": [ + { + "bbox": [ + 308, + 383, + 545, + 426 + ], + "type": "text", + "content": "[19] Hema Koppula, Abhishek Anand, Thorsten Joachims, and Ashutosh Saxena. Semantic labeling of 3d point clouds for indoor scenes. Advances in Neural Information Processing Systems, NeurIPS, pages 244-252, 2011. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 426, + 545, + 469 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 426, + 545, + 469 + ], + "spans": [ + { + "bbox": [ + 308, + 426, + 545, + 469 + ], + "type": "text", + "content": "[20] Bastian Leibe, Ales Leonardis, and Bernt Schiele. Robust object detection with interleaved categorization and segmentation. International Journal of Computer Vision, 77(1):259-289, 2008. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 471, + 545, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 471, + 545, + 514 + ], + "spans": [ + { + "bbox": [ + 308, + 471, + 545, + 514 + ], + "type": "text", + "content": "[21] Hao Li, Asim Kadav, Igor Durdanovic, Hanan Samet, and Hans Peter Graf. Pruning filters for efficient convnets. In International Conference on Learning Representations, ICLR, 2017. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 514, + 545, + 558 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 514, + 545, + 558 + ], + "spans": [ + { + "bbox": [ + 308, + 514, + 545, + 558 + ], + "type": "text", + "content": "[22] Kelin Li, Nicholas Baron, Xian Zhang, and Nicolas Rojas. Efficientgrasp: A unified data-efficient learning to grasp method for multi-fingered robot hands. IEEE Robotics and Automation Letters, 7(4):8619-8626, 2022. 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 559, + 545, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 559, + 545, + 602 + ], + "spans": [ + { + "bbox": [ + 308, + 559, + 545, + 602 + ], + "type": "text", + "content": "[23] Lucas Liebenwein, Cenk Baykal, Harry Lang, Dan Feldman, and Daniela Rus. Provable filter pruning for efficient neural networks. In 8th International Conference on Learning Representations, ICLR, 2020. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 603, + 545, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 603, + 545, + 658 + ], + "spans": [ + { + "bbox": [ + 308, + 603, + 545, + 658 + ], + "type": "text", + "content": "[24] Mingbao Lin, Rongrong Ji, Yan Wang, Yichen Zhang, Baochang Zhang, Yonghong Tian, and Ling Shao. HRank: Filter pruning using high-rank feature map. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR, pages 1529-1538, 2020. 1, 3, 4, 5" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 658, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 658, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 658, + 545, + 713 + ], + "type": "text", + "content": "[25] Liyang Liu, Shilong Zhang, Zhanghui Kuang, Aojun Zhou, Jing-Hao Xue, Xinjiang Wang, Yimin Chen, Wenming Yang, Qingmin Liao, and Wayne Zhang. Group Fisher pruning for practical network compression. In International Conference on Machine Learning, ICML, pages 7021-7032, 2021. 6" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5310" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "text", + "content": "[26] Zhuang Liu, Jianguo Li, Zhiqiang Shen, Gao Huang, Shoumeng Yan, and Changshui Zhang. Learning efficient convolutional networks through network slimming. In IEEE/CVF International Conference on Computer Vision, ICCV, pages 2736-2744, 2017. 1, 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 129, + 287, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 129, + 287, + 173 + ], + "spans": [ + { + "bbox": [ + 48, + 129, + 287, + 173 + ], + "type": "text", + "content": "[27] Ze Liu, Zheng Zhang, Yue Cao, Han Hu, and Xin Tong. Group-free 3D object detection via transformers. In IEEE/CVF International Conference on Computer Vision, CVPR, pages 2949-2958, 2021. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 175, + 287, + 218 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 175, + 287, + 218 + ], + "spans": [ + { + "bbox": [ + 48, + 175, + 287, + 218 + ], + "type": "text", + "content": "[28] Xu Ma, Can Qin, Haoxuan You, Haoxi Ran, and Yun Fu. Rethinking network design and local geometry in point cloud: A simple residual MLP framework. In International Conference on Learning Representations, ICLR, 2022. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 220, + 287, + 263 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 220, + 287, + 263 + ], + "spans": [ + { + "bbox": [ + 48, + 220, + 287, + 263 + ], + "type": "text", + "content": "[29] Fanxu Meng, Hao Cheng, Ke Li, Huixiang Luo, Xiaowei Guo, Guangming Lu, and Xing Sun. Pruning filter in filter. In Advances in Neural Information Processing Systems, NeurIPS, volume 33, pages 17629-17640, 2020. 1, 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 265, + 287, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 265, + 287, + 330 + ], + "spans": [ + { + "bbox": [ + 48, + 265, + 287, + 330 + ], + "type": "text", + "content": "[30] Quang-Hieu Pham, Thanh Nguyen, Binh-Son Hua, Gemma Roig, and Sai-Kit Yeung. JSIS3D: Joint semantic-instance segmentation of 3d point clouds with multi-task pointwise networks and multi-value conditional random fields. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR, 2019. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 332, + 287, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 332, + 287, + 376 + ], + "spans": [ + { + "bbox": [ + 48, + 332, + 287, + 376 + ], + "type": "text", + "content": "[31] Charles R Qi, Or Litany, Kaiming He, and Leonidas J Guibas. Deep hough voting for 3D object detection in point clouds. In IEEE/CVF International Conference on Computer Vision, ICCV, 2019. 3, 6, 12" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 377, + 287, + 432 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 377, + 287, + 432 + ], + "spans": [ + { + "bbox": [ + 48, + 377, + 287, + 432 + ], + "type": "text", + "content": "[32] Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. Pointnet: Deep learning on point sets for 3D classification and segmentation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR, pages 652-660, 2017. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 434, + 287, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 434, + 287, + 487 + ], + "spans": [ + { + "bbox": [ + 48, + 434, + 287, + 487 + ], + "type": "text", + "content": "[33] Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J Guibas. Pointnet++: Deep hierarchical feature learning on point sets in a metric space. In Advances in Neural Information Processing Systems, NeurIPS, pages 5100-5109, 2017. 3, 5, 6, 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 490, + 287, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 490, + 287, + 544 + ], + "spans": [ + { + "bbox": [ + 48, + 490, + 287, + 544 + ], + "type": "text", + "content": "[34] Guocheng Qian, Hasan Hammoud, Guohao Li, Ali Thabet, and Bernard Ghanem. ASSANet: An anisotropic separable set abstraction for efficient point cloud representation learning. In Advances in Neural Information Processing Systems, NeurIPS, pages 28119-28130, 2021. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 546, + 287, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 546, + 287, + 610 + ], + "spans": [ + { + "bbox": [ + 48, + 546, + 287, + 610 + ], + "type": "text", + "content": "[35] Guocheng Qian, Yuchen Li, Houwen Peng, Jinjie Mai, Hasan Hammoud, Mohamed Elhoseiny, and Bernard Ghanem. PointNeXt: Revisiting pointnet++ with improved training and scaling strategies. In Advances in Neural Information Processing Systems, NeurIPS, 2022. 3, 5, 6, 7, 12, 13" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 613, + 287, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 613, + 287, + 667 + ], + "spans": [ + { + "bbox": [ + 48, + 613, + 287, + 667 + ], + "type": "text", + "content": "[36] Ruizhi Shao, Zerong Zheng, Hongwen Zhang, Jingxiang Sun, and Yebin Liu. DiffuStereo: High quality human reconstruction via diffusion-based stereo using sparse cameras. In European Conference on Computer Vision, ECCV, pages 702-720, 2022. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "type": "text", + "content": "[37] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. In International Conference on Learning Representations, ICLR, 2015. 1, 2" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 546, + 712 + ], + "type": "list", + "angle": 0, + "index": 24, + "blocks": [ + { + "bbox": [ + 307, + 73, + 546, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 546, + 117 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 546, + 117 + ], + "type": "text", + "content": "[38] Shuran Song, Samuel P Lichtenberg, and Jianxiong Xiao. Sun RGB-D: A RGB-D scene understanding benchmark suite. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR, pages 567-576, 2015. 6, 8" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 307, + 119, + 546, + 163 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 119, + 546, + 163 + ], + "spans": [ + { + "bbox": [ + 307, + 119, + 546, + 163 + ], + "type": "text", + "content": "[39] Shuran Song, Samuel P Lichtenberg, and Jianxiong Xiao. Sun RGB-D: A RGB-D scene understanding benchmark suite. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR, pages 567-576, 2015. 12" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 164, + 546, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 164, + 546, + 217 + ], + "spans": [ + { + "bbox": [ + 307, + 164, + 546, + 217 + ], + "type": "text", + "content": "[40] Yang Sui, Miao Yin, Yi Xie, Huy Phan, Saman Aliari Zonouz, and Bo Yuan. CHIP: Channel independence-based pruning for compact neural networks. In Advances in Neural Information Processing Systems, NeurIPS, pages 24604-24616, 2021. 1, 3, 4, 5, 12" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 220, + 546, + 263 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 220, + 546, + 263 + ], + "spans": [ + { + "bbox": [ + 307, + 220, + 546, + 263 + ], + "type": "text", + "content": "[41] Lyne P. Tchapmi, Christopher B. Choy, Iro Armeni, JunYoung Gwak, and Silvio Savarese. SEGCloud: Semantic segmentation of 3d point clouds. In International Conference on 3D Vision, 3DV, pages 537-547, 2017. 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 265, + 546, + 320 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 265, + 546, + 320 + ], + "spans": [ + { + "bbox": [ + 307, + 265, + 546, + 320 + ], + "type": "text", + "content": "[42] Hugues Thomas, Charles R Qi, Jean-Emmanuel Deschaud, Beatz Marcotegui, François Goulette, and Leonidas J Guibas. KpConv: Flexible and deformable convolution for point clouds. In IEEE/CVF international conference on computer vision, CVPR, pages 6411–6420, 2019. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 322, + 546, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 322, + 546, + 387 + ], + "spans": [ + { + "bbox": [ + 307, + 322, + 546, + 387 + ], + "type": "text", + "content": "[43] Mikaela Angelina Uy, Quang-Hieu Pham, Binh-Son Hua, Duc Thanh Nguyen, and Sai-Kit Yeung. Revisiting point cloud classification: A new benchmark dataset and classification model on real-world data. In IEEE/CVF International Conference on Computer Vision, ICCV, pages 1588–1597, 2019. 5, 6, 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 389, + 546, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 389, + 546, + 443 + ], + "spans": [ + { + "bbox": [ + 307, + 389, + 546, + 443 + ], + "type": "text", + "content": "[44] Haowen Wang, Mingyuan Wang, Zhengping Che, Zhiyuan Xu, Xiuquan Qiao, Mengshi Qi, Feifei Feng, and Jian Tang. RGB-Depth fusion GAN for indoor depth completion. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR, pages 6199-6208, 2022. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 445, + 546, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 445, + 546, + 533 + ], + "spans": [ + { + "bbox": [ + 307, + 445, + 546, + 533 + ], + "type": "text", + "content": "[45] Wenxiao Wang, Minghao Chen, Shuai Zhao, Long Chen, Jinming Hu, Haifeng Liu, Deng Cai, Xiaofei He, and Wei Liu. Accelerate cnns from three dimensions: A comprehensive pruning framework. In Marina Meila and Tong Zhang, editors, Proceedings of the 38th International Conference on Machine Learning, ICML 2021, 18-24 July 2021, Virtual Event, volume 139 of Proceedings of Machine Learning Research, pages 10717-10726. PMLR, 2021. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 535, + 546, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 535, + 546, + 610 + ], + "spans": [ + { + "bbox": [ + 307, + 535, + 546, + 610 + ], + "type": "text", + "content": "[46] Wenxiao Wang, Cong Fu, Jishun Guo, Deng Cai, and Xiaofei He. COP: customized deep model compression via regularized correlation-based filter-level pruning. In Sarit Kraus, editor, Proceedings of the Twenty-Eighth International Joint Conference on Artificial Intelligence, IJCAI 2019, Macao, China, August 10-16, 2019, pages 3785-3791. ijcai.org, 2019. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 613, + 546, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 613, + 546, + 667 + ], + "spans": [ + { + "bbox": [ + 307, + 613, + 546, + 667 + ], + "type": "text", + "content": "[47] Zhirong Wu, Shuran Song, Aditya Khosla, Fisher Yu, Linguuang Zhang, Xiaou Tang, and Jianxiong Xiao. 3D ShapeNets: A deep representation for volumetric shapes. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR, pages 1912-1920, 2015. 5, 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 670, + 546, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 670, + 546, + 712 + ], + "spans": [ + { + "bbox": [ + 307, + 670, + 546, + 712 + ], + "type": "text", + "content": "[48] Chenfeng Xu, Shijia Yang, Tomer Galanti, Bichen Wu, Xiangyu Yue, Bohan Zhai, Wei Zhan, Peter Vajda, Kurt Keutzer, and Masayoshi Tomizuka. Image2point: 3d point-cloud understanding with 2d image pretrained models. In" + } + ] + } + ], + "index": 23 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 314, + 757 + ], + "type": "text", + "content": "5311" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 341 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 67, + 72, + 286, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 72, + 286, + 95 + ], + "spans": [ + { + "bbox": [ + 67, + 72, + 286, + 95 + ], + "type": "text", + "content": "European Conference on Computer Vision, pages 638-656. Springer, 2022. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 96, + 287, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 96, + 287, + 140 + ], + "spans": [ + { + "bbox": [ + 48, + 96, + 287, + 140 + ], + "type": "text", + "content": "[49] Wei Yang, Chris Paxton, Maya Cakmak, and Dieter Fox. Human grasp classification for reactive human-to-robot handovers. In IEEE/RSJ International Conference on Intelligent Robots and Systems, IROS, pages 11123–11130, 2020. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 141, + 287, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 141, + 287, + 184 + ], + "spans": [ + { + "bbox": [ + 48, + 141, + 287, + 184 + ], + "type": "text", + "content": "[50] Xucheng Ye, Pengcheng Dai, Junyu Luo, Xin Guo, Yingjie Qi, Jianlei Yang, and Yiran Chen. Accelerating CNN training by pruning activation gradients. In European Conference on Computer Vision, ECCV, pages 322-338, 2020. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 186, + 287, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 186, + 287, + 251 + ], + "spans": [ + { + "bbox": [ + 48, + 186, + 287, + 251 + ], + "type": "text", + "content": "[51] Ruichi Yu, Ang Li, Chun-Fu Chen, Jui-Hsin Lai, Vlad I Morariu, Xintong Han, Mingfei Gao, Ching-Yung Lin, and Larry S Davis. Nisp: Pruning networks using neuron importance score propagation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR, pages 9194-9203, 2018. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 253, + 287, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 253, + 287, + 296 + ], + "spans": [ + { + "bbox": [ + 48, + 253, + 287, + 296 + ], + "type": "text", + "content": "[52] Hengshuang Zhao, Li Jiang, Jiaya Jia, Philip H. S. Torr, and Vladlen Koltun. Point transformer. In IEEE/CVF International Conference on Computer Vision, ICCV, pages 16239-16248, 2021. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 297, + 287, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 297, + 287, + 341 + ], + "spans": [ + { + "bbox": [ + 48, + 297, + 287, + 341 + ], + "type": "text", + "content": "[53] Wu Zheng, Weiliang Tang, Li Jiang, and Chi-Wing Fu. SESSD: self-ensembling single-stage object detector from point cloud. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR, pages 14494–14503, 2021. 1" + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5312" + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/CR-FIQA_ Face Image Quality Assessment by Learning Sample Relative Classifiability/2ae421c8-bef3-40d8-a9e9-022734d8a26e_content_list.json b/2023/CR-FIQA_ Face Image Quality Assessment by Learning Sample Relative Classifiability/2ae421c8-bef3-40d8-a9e9-022734d8a26e_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..b0f46da1d472330275d1b8352b0a34dd91d95d04 --- /dev/null +++ b/2023/CR-FIQA_ Face Image Quality Assessment by Learning Sample Relative Classifiability/2ae421c8-bef3-40d8-a9e9-022734d8a26e_content_list.json @@ -0,0 +1,1369 @@ +[ + { + "type": "text", + "text": "CR-FIQA: Face Image Quality Assessment by Learning Sample Relative Classifiability", + "text_level": 1, + "bbox": [ + 119, + 128, + 851, + 175 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Fadi Boutros1, Meiling Fang1,2, Marcel Klemt1, Biying Fu1, Naser Damer1,2 \n1Fraunhofer Institute for Computer Graphics Research IGD, Darmstadt, Germany \n2Department of Computer Science, TU Darmstadt, Darmstadt, Germany \nEmail: fadi.boutros@igd.fraunhofer.de", + "bbox": [ + 161, + 202, + 807, + 273 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 308, + 313, + 325 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Face image quality assessment (FIQA) estimates the utility of the captured image in achieving reliable and accurate recognition performance. This work proposes a novel FIQA method, CR-FIQA, that estimates the face image quality of a sample by learning to predict its relative classifiability. This classifiability is measured based on the allocation of the training sample feature representation in angular space with respect to its class center and the nearest negative class center. We experimentally illustrate the correlation between the face image quality and the sample relative classifiability. As such property is only observable for the training dataset, we propose to learn this property by probing internal network observations during the training process and utilizing it to predict the quality of unseen samples. Through extensive evaluation experiments on eight benchmarks and four face recognition models, we demonstrate the superiority of our proposed CR-FIQA over state-of-the-art (SOTA) FIQA algorithms.", + "bbox": [ + 75, + 340, + 473, + 612 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 619, + 209, + 633 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Face image utility indicates the utility (value) of an image to face recognition (FR) algorithms [1, 19]. This utility is measured with a scalar, namely the face image quality (FIQ) score, following the definition in ISO/IEC 2382-37 [20] and the FR Vendor Test (FRVT) for FIQA [10].", + "bbox": [ + 75, + 637, + 468, + 712 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "As FIQA measures the face utility to FR algorithm, it does not necessary reflects, and does not aim at measuring, the perceived image quality, e.g. a profile face image can be of high perceived quality but of low utility to FR algorithm [35]. Assessing this perceived image quality has been addressed in the literature by general image quality assessment (IQA) methods [26, 29, 30] and is different than assessing the utility of an the image for FR. This is reflected by FIQA methods [28, 32, 36] significantly outperforming IQA methods [26, 29, 30] in measuring the utility [19] of face images in FR, as demonstrated in [8, 28, 36].", + "bbox": [ + 75, + 712, + 470, + 878 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "SOTA FIQA methods focused either on creating concepts to label the training data with FIQ scores and then learn a regression problem [14, 15, 32], or on developing a link between face embedding properties under certain scenarios and the FIQ [28, 34, 36]. Generally, the second approach led to better FIQA performances with most works mentioning the error-prone labeling of the ground truth quality in the first research direction as a possible reason [28, 36]. However, in the second category, transferring the information in network embeddings into an FIQ score is not a learnable process, but rather a form of statistical analysis, which might not be optimal.", + "bbox": [ + 500, + 310, + 892, + 491 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "This paper proposes a novel learning paradigm to assess FIQ, namely the CR-FIQA. Our concept is based on learning to predict the classifiability of FR training samples by probing internal network observations that point to the relative proximity of these samples to their class centers and negative class centers. This regression is learned simultaneously with a conventional FR training process that minimizes the distance between the training samples and their class centers. Linking the properties that cause high/low classifiability of a training sample to the properties leading to high/low FIQ, we can use our CR-FIQA to predict the FIQ of any given sample. We empirically prove the theorized link between classifiability (Section 3.3) and FIQ and conduct thorough ablation studies on key aspects of our CR-FIQA design (Section 5). The proposed CR-FIQA is evaluated on eight benchmarks along with SOTA FIQAs. The reported results on four FR models demonstrate the superiority of our proposed CR-FIQA over SOTA methods and the stability of its performance across different FR models. An overview of the proposed CR-FIQA is presented in Figure 1 and will be clarified in detail in this paper.", + "bbox": [ + 496, + 493, + 893, + 811 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "2. Related work", + "text_level": 1, + "bbox": [ + 500, + 819, + 638, + 835 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The recent SOTA FIQA approaches can be roughly grouped into two main categories. The first are approaches that learn a straight forward regressions problem to assess a FIQ score [1,14,15,32,38]. The second category uses prop", + "bbox": [ + 498, + 839, + 893, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "1https://github.com/fdbtrs/CR-FIQA", + "bbox": [ + 94, + 886, + 364, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "5836", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/113af72237b96da2e5c4dbc25fd3f98375e9d2f01344258c80c376e07ffa2f6e.jpg", + "image_caption": [ + "Figure 1. An overview of our CR-FIQA training paradigm. We propose to simultaneously learn to optimize the class center $(\\mathcal{L}_{Arc})$ , while learning to predict an internal network observation i.e. the allocation of the feature representation of sample $x$ in feature space, with respect to its class center $w_{1}$ and nearest negative class center $w_{2}$ $(\\mathcal{L}_{\\mathcal{CR}})$ . The figure in the red rectangle illustrates the angle between two samples $x_{1}$ and $x_{2}$ (belong to identity 1) and their class center $w_{1}$ . The plot on the right of the figure shows the distribution of the cosine similarity between training samples and their class centers (CCS) and nearest negative class centers (NNCCS) obtained from ResNet-50 trained on CASIA-WebFace [39]. The example images on the top-right of this plot are of high CCS values and the ones on the top-left are of low CCS values (notice the correspondence to perceived quality). These samples are selected from CASIA-WebFace [39]. During the testing mode, the classification layer is removed and the output of the regression layer is used to predict the FIQA of testing samples." + ], + "image_footnote": [], + "bbox": [ + 181, + 89, + 789, + 290 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "erties of the FR model responses to face samples to estimate the sample quality without explicitly learning a typical supervised regression that requires quality labels [28, 34, 36]. In the first category, the innovation focused on creating the FIQ labels for training. These quality labels included human-labeled quality labels [1], the FR genuine comparison score between a sample and an ICAO [18] compliant sample [14, 15], the FR comparison score involving the labeled sample (assumed to have the lower quality in the comparison pair) [38], and the Wasserstein distance between a randomly selected genuine and imposter FR comparisons with the labeled sample [32]. These solutions generally trained a regression network to predict the quality label, using both, trained-from-scratch networks in some cases [38], and pre-trained FR networks in other cases [14, 15, 32]. A slightly different approach, however also based on learning from labels, focuses on learning to predict the sample quality as a rank [22] based on FR performance-based training rank labels of a set of databases [4]. In the second category, the innovation was rather focused on linking face embedding properties under certain scenarios to the FIQ, without the explicit need for quality-labeled data. In [36], the assessed sample is passed through an FR network multiple times, each with a different random dropout pattern. The robustness of the resulting embeddings, represented by the sigmoid of the negative mean of the Euclidean distances between the embeddings, is considered the FIQ score. In [28], the FIQ score is calculated as the magnitude of the sample embedding. This is based on training the FR model using a loss that adapts the penalty margin loss based on this magnitude, and thus links the closeness of a sample to its class", + "bbox": [ + 75, + 421, + 472, + 891 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "center to the unnormalized embedding magnitude. While in [34], the solution produces both, an FR embedding and a gaussian variance (uncertainty) vector, from a face sample. The inverse of harmonic mean of the uncertainty vector is considered as the FIQ score. Our CR-FIQA learns a regression problem to estimate the FIQ score, however, unlike previous works, without relying on preset labels, but rather learn a dynamic internal network observations (during training) that point out sample classifiability.", + "bbox": [ + 496, + 421, + 893, + 556 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3. Approach", + "text_level": 1, + "bbox": [ + 500, + 564, + 609, + 580 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "This section presents our proposed Certainty Ratio Face Image Quality Assessment (CR-FIQA) approach, which inspects internal network observations to learn to predict the sample relative classifiability. This classifiability prediction is then used to estimate the FIQ. An overview of the proposed CR-FIQA approach is presented in Figure 1. During the training phase of an FR model, the model can conveniently push the high-quality samples close to their class center and relatively far from other class centers. Conversely, the FR is not able to push, to the same degree, low-quality samples to their class center, and thus they will remain relatively farther from their class center than the high-quality ones. Based on this assumption, we theorize our approach by stating that the properties that cause a face sample to lay relatively closer to its class center during training are the ones that make it a high-quality sample, and vice versa. Therefore, learning to predict such properties in any given sample would lead to learning to assess this sample quality. To learn to perform such assessment, our training paradigm targets learning internal network observations that evolve during the FR training phase, where these observa", + "bbox": [ + 496, + 583, + 893, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "5837", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "tions act as a training objective. The predictions of such training paradigm can be simply stated as answering a question: if a given sample was hypothetically part of the FR model training (which it is not), how relatively close would it be to its class center? Answering this question would give us an indication of this sample quality as will be shown in detail in this paper.", + "bbox": [ + 75, + 90, + 468, + 196 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In the rest of this section, We formalize and empirically rationalize our proposed CR-FIQA approach and its components. To do that, we start by shortly revisiting angular margin penalty-based softmax loss utilized to optimize the class centers of the FR model. Then present a detailed description of our proposed CR-FIQA concept and the associated training paradigm.", + "bbox": [ + 75, + 196, + 467, + 303 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Revisiting Margin Penalty-based Softmax Loss", + "text_level": 1, + "bbox": [ + 76, + 306, + 467, + 321 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Angular margin penalty-based softmax is a widely used loss function for training FR models [3,5,17,28]. It extends over softmax loss by deploying angular penalty margin on the angle between the deep features and their corresponding weights. Margin penalty-based softmax loss aims to push the decision boundary of softmax, and thus enhance intraclass compactness and inter-class discrepancy. From this family of loss functions, this work utilizes ArcFace loss [5] to optimize the distance between the training samples and their class center. Our choice of ArcFace loss is based on the SOTA performance achieved by ResNet-100 network trained with ArcFace on mainstream benchmarks [5]. Formally, ArcFace loss is defined as follows:", + "bbox": [ + 75, + 325, + 467, + 521 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {A r c} = \\frac {1}{N} \\sum_ {i \\in N} - \\log \\frac {e ^ {s \\left(\\cos \\left(\\theta_ {y _ {i}} + m\\right)\\right)}}{e ^ {s \\left(\\cos \\left(\\theta_ {y _ {i}} + m\\right)\\right)} + \\sum_ {j = 1 , j \\neq y _ {i}} ^ {C} e ^ {s \\left(\\cos \\left(\\theta_ {j}\\right)\\right)}}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 78, + 523, + 467, + 589 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $N$ is the batch size, $C$ is the number of classes (identities), $y_{i}$ is the class label of sample $i$ (in range $[1,C])$ , $\\theta_{y_i}$ is the angle between the features $x_{i}$ and the $y_{i}$ -th class center $w_{y_i}$ . $x_{i} \\in R^{d}$ is the deep feature embedding of the last fully connected layer of size $d$ . $w_{y_i}$ is the $y_{i}$ -th column of weights $W \\in R_C^d$ of the classification layer. $\\theta_{y_i}$ is defined as $x_{i} w_{y_{i}}^{T} = \\| x_{i} \\| \\| w_{y_{i}} \\| \\cos (\\theta_{y_{i}})$ [24]. The weights and the feature norms are fixed to $\\| w_{y_i} \\| = 1$ and $\\| x_{i} \\| = 1$ , respectively, using $l_2$ normalization as defined in [24, 37]. The decision boundary, in this case, depends on the angle cosine between $x_{i}$ and $w_{y_i}$ . $m > 0$ is an additive angular margin proposed by ArcFace [5] to enhance the intra-class compactness and inter-class discrepancy. Lastly, $s$ is the scaling parameter [37].", + "bbox": [ + 75, + 590, + 467, + 801 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Certainty Ratio", + "text_level": 1, + "bbox": [ + 76, + 805, + 233, + 821 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we formulate and empirically rationalize the main concepts that build our FIQA solution. We derive our Certainty Ratio (CR) to estimate the sample relative classifiability. Additionally, we experimentally illustrate the strong relationship between our CR measure and FIQ.", + "bbox": [ + 75, + 824, + 467, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Certainty Ratio During the FR model training phase, the model is trained to enhance the separability between the classes (identities) by pushing each sample $x_{i}$ to be close to its class center $w_{y_i}$ and far from the other (negative) class centers $w_{j}, j \\neq y_{i}$ . Based on this, we first define the Class Center Angular Similarity (CCS) as the proximity between $x_{i}$ and its class center $w_{y_i}$ , as follows:", + "bbox": [ + 496, + 90, + 890, + 196 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nC C S _ {x _ {i}} = \\cos \\left(\\theta_ {y _ {i}}\\right), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 625, + 200, + 890, + 217 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\theta_{y_i}$ is the angle between $x_i$ and its class center $w_{y_i}$ , where the weights of the last fully connected layer of the FR model trained with softmax loss are considered as the centers for each class [5,25]. Then, we define the Closest Nearest Negative Class Center Angular Similarity (NNCCS) as proximity between $x_i$ and the nearest negative class center $w_j$ , $j \\neq y_i$ . Formally, NNCCS is defined as follows:", + "bbox": [ + 496, + 229, + 890, + 335 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nN N C C S _ {x _ {i}} = \\max _ {j = 1, j \\neq y _ {i}} ^ {C} (\\cos \\left(\\theta_ {j}\\right)), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 581, + 339, + 890, + 368 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\theta_{j}$ is the angle between $x_{i}$ and $w_{j}$ . As we theorize, when the FR model converges, the high-quality samples are pushed closer to their class centers (high CCS) in relation to their distance to neighbouring negative class centers (low NNCCS). However, low-quality samples can not be pushed as close to their class centers. A sample able to achieve high CCS with respect to NNCCS is a sample easily correctly classified during training, and thus is relatively highly classifiable. We thus measure this relative classifiability by the ratio of CCS to NNCS, which we note as the Certainty Ratio (CR), as follows:", + "bbox": [ + 496, + 378, + 890, + 542 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nC R _ {x _ {i}} = \\frac {C C S _ {x _ {i}}}{N N C C S _ {x _ {i}} + (1 + \\epsilon)}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 584, + 545, + 890, + 579 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where the $1 + \\epsilon$ term is added to insure a positive above zero denominator, i.e. shift the NNCCS value range from $[-1, +1]$ to $[\\epsilon, 2 + \\epsilon]$ . This ensures that the CR of a sample with a lower NNCCS is relatively higher than a sample with a higher NNCCS, given the same CCS, i.e. NNCCS regulates the CCS value in relation to neighbouring classes. The $\\epsilon$ is set to $1e - 9$ in our experiments. The optimal CR is obtained when the CCS is approaching the maximum cosine similarity value (+1) and the NNCCS is approaching the minimum cosine similarity value (-1), i.e. the training sample is capable of being pushed to its class center, and far away from the closest negative class center, and thus it is highly classifiable.", + "bbox": [ + 496, + 588, + 890, + 785 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.3. Relation between the CR and FIQ", + "text_level": 1, + "bbox": [ + 500, + 790, + 797, + 806 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Here, we empirically prove the theorized relationship between the CR and FIQ (defined earlier as image utility). Namely, we want to answer: if the CR values achieved by training samples of an FR model were used as FIQ, would they behave as expected from an optimal FIQ? If yes, then the face image properties leading to high/low CR", + "bbox": [ + 496, + 810, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "5838", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "do also theoretically lead to high/low FIQ. To answer this question we conducted an experiment on a ResNet-50 [13] FR model trained on CASIA-WebFace [39] with ArcFace loss [5] (noted as R50(CASIA)). Specifically, we calculate the CR, CCS, and NNCCS values from the trained model for all samples in the training dataset (0.5M images of 10K identities). An insight on the resulting CCS and NNCCS values (CR being a derivative measure) is given as value distributions in Figure 1, showing that these measures vary between different samples. Furthermore, based on the calculated scores, we plot Error vs. Reject Curves (ERC) (described in Section 4) to demonstrate the relationship between the CR, as an FIQ measure, and FR performance. To calculate the FR performance in the ERC curve, we extract the feature embedding of CASIA-WebFace [39] using a ResNet-100 model [13] trained on MS1M-V2 [5, 12] with ArcFace (noted as R100(MS1M-V2)). We utilize a different model (trained on a different database) to extract the embedding (R100(MS1M-V2)) than the one used to calculate CR, CCS, and NNCCS (R50(CASIA)) to provide fair evaluation where the FR performance is evaluated on unseen data. Then, we perform $n : n$ comparisons between all samples of CASIA-WebFace using feature embedding obtained from the R100(MS1M-V2).", + "bbox": [ + 76, + 90, + 472, + 454 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/f127b852bbb10049bcd2c891ff2f7b7cc4b4eb23f68e608b52856ec8d7617c78.jpg", + "image_caption": [ + "(a)" + ], + "image_footnote": [], + "bbox": [ + 96, + 463, + 259, + 535 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/bc38a9ae33aced9bb1c910c20058afdc20a61b75cf6cbfab187e61b6df40bf26.jpg", + "image_caption": [ + "(b)", + "Figure 2. ERCs showing the verification performance as False None Match Rate (FNMR) at False Match Rate (FMR) of 1e-3 (2a) and 1e-4 (2b) with CCS, NNCCS and CR as FIQ vs. rejection ratio. This ERC plots show the effectiveness of rejecting samples with the lowest CCS and CR on the performance" + ], + "image_footnote": [], + "bbox": [ + 272, + 463, + 436, + 534 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Figure 2 presents the ERC of CR, CCS, and NNCCS experimentally used as FIQ. An FIQ measure would cause the ERC to drop as rapidly as possible when rejecting a larger fraction of low-quality samples (moving to the right).", + "bbox": [ + 75, + 627, + 468, + 688 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "It can be clearly noticed in Figure 2 that the CCS and CR do behave as we would expect from a good performing FIQ, as the verification error value drops rapidly when rejecting low quality (low CCS and CR) samples. It can be also observed that the CR does that more steadily when compared to CCS. This points out that adding the scaling term NNCCS in CR calculation can enhance the representation of the CCS as an FIQ measure, which will be clearer later when we experimentally evaluate our CR-FIQA approach in Section 5. As expected, the NNCCS measure by itself does not strongly act as an FIQ measure would, demonstrated by the relatively flat ERC in Figure 2, as it only considers the distance to the nearest negative class. This empirical evaluation does provide a confirming answer to the previ", + "bbox": [ + 75, + 688, + 470, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "ously stated question by affirming that the CR does act as expected from an FIQ measure and thus, theoretically, one can strongly link the image properties that cause high/low CR in the FR training data to these causing high/low FIQ.", + "bbox": [ + 496, + 90, + 890, + 151 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.4. Quality Estimation Training Paradigm", + "text_level": 1, + "bbox": [ + 498, + 159, + 833, + 175 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In the previous section, we proved that the CR does behave as an FIQ would, and thus, it can also relate to image properties that dictate FIQ. However, the CR measure is only observable for samples in the FR training dataset, where the class centers are known. In a real case scenario, the FIQ measure should be assessed to any single image, i.e. unseen evaluation data. Considering this, and in an effort to predict what the CR value would be for a given sample if hypothetically it was part of the FR training, we propose to simultaneously learn to predict the CR from the training dataset while optimizing the class centers (typical FR training) during the training phase, i.e. the CR-FIQA model. To enable this, we add a single regression layer to the FR model. The input of the regression layer is a feature embedding $x_{i}$ and the output is an estimation of the CR. The output of this regression layer is used later to predict the FIQ score of the unseen sample, e.g. from the evaluation dataset. Thus, we capture the properties that make the CR high/low to predict the FIQ of any given sample. Towards this goal, during the training phase, the model (in Figure 1) has two learning objectives: a) It is trained to optimize the distance between the samples and the class centers using ArcFace loss defined in Equation 1. b) It is trained to predict the internal network observation, CR, using Smooth L1-Loss [9] applied between the output of the regression layer (P) and the CR calculated as in Equation 5. Smooth L1-loss can be interpreted as a combination of L1 and L2-losses by defining a threshold $\\beta$ that changes between them [9]. Our choice for smooth-l1 loss is based on: 1) It is less sensitive to outliers than l2. The derivative of L2 loss increases when the difference between the prediction and ground-truth label is increased, making the derivative of loss values large at the early stage of the training, leading to unstable training. Additionally, L2 loss can easily generate gradient explosion [9] when there are outliers in the training data. 2) L1 loss can lead to stable training. However, the absolute values of the difference between prediction and ground truth are small, especially in the later stage of the training. Therefore, the model accuracy can hardly be improved at a later stage of the training as the loss function will fluctuate around a stable value. Combining L1 and L2 as in Smooth L1-loss avoids gradient explosion, which might be caused by L2 and facilitates better convergence than L1. The loss leading to the second objective is then given as:", + "bbox": [ + 496, + 179, + 890, + 843 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {C R} = \\frac {1}{N} \\sum_ {i \\in N} \\left\\{ \\begin{array}{l l} \\frac {0 . 5 \\times \\left(C R _ {x _ {i}} - P _ {i}\\right) ^ {2}}{\\beta} & \\text {i f} \\left| C R _ {x _ {i}} - P _ {i} \\right| < \\beta \\\\ \\left| C R _ {x _ {i}} - P _ {i} \\right| - 0. 5 \\times \\beta & \\text {o t h e r w i s e} \\end{array} \\right. \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 500, + 847, + 893, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "5839", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/c14e758581959881b3fd5bc9341c6344a0bd6d5f4843358172bff43491fe7809.jpg", + "image_caption": [ + "(a)" + ], + "image_footnote": [], + "bbox": [ + 89, + 88, + 285, + 215 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/c0a0ffe667186e01fc06bdb62efb807dc28b8ac45759bbbdd4e3eed889c96c18.jpg", + "image_caption": [ + "(b)" + ], + "image_footnote": [], + "bbox": [ + 289, + 88, + 483, + 214 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/7ca45a0c251b02023d85b195cf69060a7183d277e6ce53f7a6ceffe1c590a47e.jpg", + "image_caption": [ + "(c)", + "Figure 3. ERC comparison between CR-FIQA(S), CCS-FIQA(S), CR-FIQA(S) (On top) and CCS-FIQA(S) (On top). The plots show the effect of rejecting samples of lowest quality, on the verification error (FNMR at FMR1e-3). CR-FIQA(S) and CCS-FIQA(S) outperformed the on-top solutions, and CR-FIQA(S) performs generally better than CCS-FIQA(S) (curve decays faster with more rejected samples)" + ], + "image_footnote": [], + "bbox": [ + 486, + 88, + 681, + 215 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/f78a781dae2143838300b454c28e485eb6a63d039a537a4192f615f032204c64.jpg", + "image_caption": [ + "(d)" + ], + "image_footnote": [], + "bbox": [ + 684, + 90, + 875, + 215 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The final loss combining both objectives for training our CR-FIQA model is defined as follows:", + "bbox": [ + 75, + 282, + 470, + 313 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\mathcal {L} _ {A r c} + \\lambda \\times \\mathcal {L} _ {C R}, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 192, + 316, + 468, + 333 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\lambda$ is a hyper-parameter used to control the balance between the two losses. At the beginning of model training, the value range of $\\mathcal{L}_{CR}$ is very small ( $\\leq 2$ ) in comparison to $\\mathcal{L}_{Arc} (\\sim 45)$ . Setting $\\lambda$ to a small value, the model will only focus on $\\mathcal{L}_{Arc}$ . Besides, setting $\\lambda$ to a large value, i.e. $>10$ , we observed that the model did not converge. Therefore, we set $\\lambda$ to 10 in all the experiments in this paper.", + "bbox": [ + 75, + 344, + 468, + 450 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experimental Setup", + "text_level": 1, + "bbox": [ + 76, + 455, + 269, + 473 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Implementation Details We demonstrate our proposed CR-FIQA under two protocols (small and large) based on the training dataset and the training model architecture. We utilize widely used architectures in the SOTA FR solutions, ResNet100 and ResNet50 [13], both modified as described in Section 3.4. For the small protocol, we utilize ResNet50 and the CASIA-WebFace [39] training data (noted as CR-FIQA(S)) and for the large protocol, we utilize ResNet100 and the MS1MV2 [5, 12] training data (noted as CR-FIQA(L)). The MS1MV2 is a refined version of the MS-Celeb-1M [12] by [5] containing 5.8M images of 85K identities. The CASIA-WebFace contains 0.5m images of 10K identities [39]. We follow the ArcFace training setting [5] to set the scale parameter $s$ to 64 and the margin $m$ to 0.5. We set the mini-batch size to 512. All models are trained with Stochastic Gradient Descent (SGD) optimizer with an initial learning rate of 1e-1. During the training, we use random horizontal flipping with a probability of 0.5 for data augmentation. We set the momentum to 0.9 and the weight decay to 5e-4. For CR-FIQA(S), the learning rate is divided by 10 at 20K and at 28K training iterations, following [5]. The training is stopped after 32K iterations. For CR-FIQA(L), the learning rate is divided by 10 at 100K and 160K training iterations, following [5]. The training is stopped after 180K iterations. All the images in evaluation and training datasets are aligned and cropped to $112 \\times 112$ , as described in [5]. All the training and testing images are normalized to have pixel values between -1 and 1. Both", + "bbox": [ + 75, + 478, + 470, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "models are trained using the loss defined in Equation 6.", + "bbox": [ + 498, + 282, + 864, + 299 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Evaluation Benchmarks We reported the achieved results on eight different benchmarks: Labeled Faces in the Wild (LFW) [16], AgeDB-30 [31], Celebrities in Frontal-Profile in the Wild (CFP-FP) [33], Cross-age LFW (CALFW) [41], Adience [6], Cross-Pose LFW (CPLFW) [40], Cross-Quality LFW (XQLFW) [23], IARPA Janus Benchmark-C (IJB-C) [27]. These benchmarks are chosen to provide a wide comparison to SOTA FIQA algorithms and give an insight into the CR-FIQA generalizability.", + "bbox": [ + 496, + 301, + 890, + 438 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Evaluation Metric We evaluate the FIQA by plotting ERCs [11]. The ERC is a widely used representation of the FIQA performance [10, 11] by demonstrating the effect of rejecting a fraction face images, of the lowest quality, on face verification performance in terms of False None Match Rate [21] (FNMR) at a specific threshold calculated at fixed False Match Rate [21] (FMR). The ERC curves for all benchmarks are plotted at two fixed FMRs, 1e-3 (as recommended for border control operations by Frontex [7]) and 1e-4 (the latter is provided in the supplementary material). We also report the Area under the Curve (AUC) of the ERC, to provide a quantitative aggregate measure of verification performance across all rejection ratios.", + "bbox": [ + 496, + 440, + 892, + 637 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Additionally, motivated by evaluating the FIQ as a weighting term for face embedding [32, 34], we follow the IJB-C 1:1 mixed verification benchmark [27] by weighting the frames such that all frames belonging to the same subject within a video have a combined weight equal to a single still image as described in IJB-C benchmark [27]. We do that by using the CR-FIQA quality scores as well as all SOTA methods. We report the verification performance of IJB-C as true acceptance rates (TAR) at false acceptance rates (FAR) of 1e-4, 1e-5, and 1e-6, as defined in [27].", + "bbox": [ + 496, + 640, + 893, + 791 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Face Recognition Models We utilize four different SOTA FR models to report the verification performance at different quality rejection rate to inspect the generalizability of FIQA over FR solutions. The FR models are ArcFace [5], ElasticFace (ElasticFace-Arc) [3], MagFace [28], and CurricularFace [17]. All models process $112 \\times 112$ aligned and cropped image to produce 512-D feature em", + "bbox": [ + 496, + 795, + 893, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5840", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "bedding. We used the officially released pretrained ResNet-100 models trained on MS1MV2 released by the four FR solutions. Although, the presented solution in this paper does not aim, and is not presented as, a solution to extract face embeddings, but rather an FIQA solution, we opted to evaluate CR-FIQA(L) backbone as a FR model on mainstream FR benchmarks for sake of providing complete experiment evaluation and probe the possibility of simultaneously using it as both FIQA and FR model. The evaluation results of CR-FIQA(L) backbone as a FR model are provided in the supplementary material.", + "bbox": [ + 75, + 90, + 472, + 257 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Baseline We compare our CR-FIQA approach with nine quality assessment methods. Three are general IQA methods that have been proven in [8] to correlate well to face utility i.e. BRISQUE [29], RankIQA [26], and DeepIQA [2], and six are SOTA face-specific FIQA methods, namely RankIQ [4], PFE [34], SER-FIQ [36], FaceQnet (v1 [14]) [14, 15], MagFace [28], and SDD-FIQA [32], all as officially released in the respective works.", + "bbox": [ + 75, + 257, + 472, + 378 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Ablation Studies", + "text_level": 1, + "bbox": [ + 76, + 383, + 241, + 398 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "This section provides experimental proof of the two main design choices in CR-FIQA.", + "bbox": [ + 75, + 402, + 468, + 431 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Does CR-FIQA benefit from the NNCCS scaling term? To answer this, we conducted additional experiments using ResNet-50 model trained on CASIA-WebFace [39] using the experimental setup described in Section 4. This model is noted as CCS-FIAQ(S). The only difference from CR-FIAQ(S) is that the CCS-FIAQ(S) is trained to learn CCS (instead of CR) by replacing the $CR_{x_i}$ in Equation 5 with $CCS_{x_i}$ , thus neglecting the NNCCS scaling term in the equation. Figure 3 presents the ERCs along with AUC using CR-FIAQ(S) and CCS-FIAQ(S) on Adience, AgeDb-30, CALLFW and CRF-FP. The verification error, FNMR at FMR1e-3, is calculated using ArcFace FR model (described in Section 4). The ERCs and AUC values show that the reduction in the error is more evident for CR-FIAQ(S) than CCS-FIQA(S). Thus, adding the scaling term NNCCS in CR enhanced the performance of the FIQA.", + "bbox": [ + 75, + 433, + 472, + 674 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Dose the simultaneous learning in CR-FIQA lead to better performance in comparison to on-the-top learning? We consider the possibility of learning to estimate CR after finalizing the FR training, in comparison to the simultaneously learning in CR-FIQA. We conducted two additional experiments using pretrained ResNet-50 trained with ArcFace loss [5] on CASIA-WebFace [39]. Specifically, we add an additional single regression layer to this pretrained model. We freeze the weights of the pretrained model and train only the regression layer to learn the internal network observation of the pretrained model using only the $\\mathcal{L}_{\\mathcal{CR}}$ (Equation 5). Using this setting, we present two instances, CR-FIQA(S) (On top) and CCS-FIQA(S) (On top), that learned to predict CR and CCS, respectively. Each of CR-FIQA(S) (On top) and CCS-FIQA(S) (On top) is fine", + "bbox": [ + 75, + 674, + 472, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "tuned for 32K iteration with an initial learning rate of 0.01. The learning rate is divided by 10 at 20K and 28K training iterations, similarly to CR-FIAQ(S) and CCS-FIAQ(S). The results (ERCs and AUCs) of these models are compared to CR-FIQA(S) and CCS-FIQA(S) in Figure 3. The ERCs in Figure 3 presented the evaluation on Adience, AgeDb-30, CALFW and CFP-FP using the ArcFace FR model. The ERCs and AUCs in Figure 3 show that both CR-FIQA(S) and CCS-FIQA(S) lead to stronger reductions in the error than the CR-FIQA(S) (On top) and the CCS-FIQA(S) (On top), when rejecting low-quality samples. This supports our training paradigm that simultaneously learns the internal network observation, CR, while optimizing the class centers. This can be related to the step-wise convergence towards the final CR value in the simultaneously training. For both ablation study questions, ERCs and AUC for all the remaining benchmarks and FR models (mentioned in Section 4) lead to similar conclusions and are provided in the supplementary material.", + "bbox": [ + 496, + 90, + 893, + 378 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "6. Result and Discussion", + "text_level": 1, + "bbox": [ + 498, + 383, + 707, + 398 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "All CR-FIQA performances reported in this paper are obtained under cross-model settings. The proposed CR-FIQA is used only to predict FIQ and not to extract feature representation of face images. None of the utilized FR models (ArcFace [5], ElasticFace [3], MagFace [28], and CurricularFace [17]) is trained with our paradigm. Instead, we trained a separate model for CR-FIQA and used the official pretrained FR models (as described in Section 4) for feature extraction. The verification performances as AUC at FMR1e-3 and FMR1e-4 are presented in Table 1. The visual verification performances as ERC curves (Figure 4) using ArcFace and ElasticFace FR models are reported at FMR1e-3. The ERC curves at FMR1e-4 and for MagFace and CurricularFace FR models at FMR1e-3 are provided in the supplementary material.", + "bbox": [ + 496, + 402, + 893, + 628 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The ERC curves (Figure 4) and the AUC values (Table 1) show that our proposed CR-FIQA(S) and CR-FIQA(L) outperformed the SOTA methods by significant margins in almost all settings. Observing the results on IJB-C, Adience, CFP-FP, CALFW, and CPLFW at FMR1e-3 and FMR1e-4 (Figure 4 and Table 1), our proposed CR-FIQA outperformed all SOTA methods on all the considered FR models. On the AgeDB-30 benchmark, our proposed CR-FIQA ranked first in five out of eight settings and second in the other three settings (Table 1). On the LFW benchmark, our proposed CR-FIQA ranked behind the MagFace and the RankIQ. This is the only case that our models did not outperform all SOTA methods. However, it can be noticed from the ERC curves in Figure 4 that none of the SOTA methods were able to achieve stable behavior (smoothly decaying curve) on LFW. The main reason for such unstable ERC behavior on LFW is that the FR performance on LFW is nearly saturated (all models achieved above $99.80\\%$ accu", + "bbox": [ + 496, + 628, + 895, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "5841", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/a83dde4923042bfae4c611ef0b68d1a56fc0c639dc7c63ba3adc4eb9b92e81a7.jpg", + "table_caption": [ + "Table 1. The AUCs of ERC achieved by our CR-FIQA and the SOTA methods under different experimental settings. CR-FIQA achieved the best performance (lowest AUC) in almost all settings. On XQLFW, the SER-FIQ (marked with *) is used for the sample selection of the XQLFW benchmark. The best result for each experimental setting is in bold and the second-ranked one is in italic. The notions of $1e - 3$ and $1e - 4$ indicate the value of the fixed FMR at which the ERC curve (FNMR vs. reject) was calculated at" + ], + "table_footnote": [], + "table_body": "
FRMethodAdience [6]AgeDB-30 [31]CFP-FP [33]LFW [16]CALFW [41]CPLFW [40]XQLFW [23]IJB-C [27]
1e-31e-41e-31e-41e-31e-41e-31e-41e-31e-41e-31e-41e-31e-41e-31e-4
ArcFace [5]QABRISQUE [29]0.05650.12850.04000.05850.03430.04330.00430.00490.07550.08130.25580.30370.66800.71220.03810.0656
RankIQA [26]0.04000.09330.03720.05230.03010.03840.00390.00450.08460.09150.24370.29690.65840.70390.03850.0640
DeepIQA [2]0.05680.13720.04030.05230.02380.02920.00490.00560.07930.08500.23090.28560.59580.64580.03830.0640
FIQARankIQ [4]0.03530.08730.03220.04200.01520.02600.00180.00240.06080.06720.06330.08480.27890.33320.02270.0342
PFE [34]0.02120.04280.01720.02260.00920.01290.00230.00280.06470.06810.04500.06380.23020.27100.01760.0248
SER-FIQ [36]0.02230.04340.01670.02230.00650.01030.00230.00280.05950.06270.03890.05840.1812*0.2295*0.01610.0241
FaceQnet [14, 15]0.03460.07340.01970.02450.02400.02730.00220.00270.07740.08220.15040.17510.58290.61360.02700.0376
MagFace [28]0.02070.04250.01560.01980.00730.01050.00160.00210.05680.06020.04920.06420.40220.46360.01710.0254
SDD-FIQA [32]0.02480.05620.01860.02060.01220.01930.00210.00270.06410.06980.05170.06700.30900.35610.01860.0270
CR-FIQA(S)(Our)0.02410.05170.01440.01870.00900.01450.00200.00250.05210.05540.03910.05670.23770.27400.01710.0250
CR-FIQA(L)(Our)0.02040.03530.01590.01890.00500.00820.00230.00290.06160.06320.03600.05150.20840.24410.01380.0207
ElasticFace [3]QABRISQUE [29]0.06440.11840.03750.04030.02810.03720.00340.00470.07260.07470.26410.46880.63430.69640.03570.0621
RankIQA [26]0.04330.08620.03740.04360.02690.03180.00330.00450.08100.08350.23250.43060.61890.68560.03660.0599
DeepIQA [2]0.06450.12030.03840.04110.01910.02560.00430.00560.07560.07720.24010.45410.54000.58320.03790.0590
FIQARankIQ [4]0.04000.07770.03090.03370.01490.01800.00130.00200.05980.06140.05810.07270.24680.27760.02260.0334
PFE [34]0.02220.03810.01630.01720.00880.01130.00180.00250.06280.06430.04190.08950.21120.24360.01710.0247
SER-FIQ [36]0.02400.04170.01630.01790.00610.00850.00210.00280.05740.05900.03870.05130.1576*0.1868*0.01560.0235
FaceQnet [14, 15]0.03690.06670.01940.02070.02270.02470.00210.00260.07630.07770.14200.28800.55490.58440.02630.0370
MagFace [28]0.02250.03850.01500.01580.00690.00950.00140.00210.05530.05630.04740.05970.39730.42820.01660.0243
SDD-FIQA [32]0.02770.05120.01870.02000.00980.01180.00190.00270.06240.06380.04930.06340.30520.35620.01830.0266
CR-FIQA(S)(Our)0.02570.04650.01460.01600.00700.00960.00150.00220.05090.05220.03830.05020.20930.28350.01670.0244
CR-FIQA(L)(Our)0.02140.03570.01490.01590.00450.00650.00180.00250.05940.06080.03500.04620.17980.20600.01350.0203
MagFace [28]QABRISQUE [29]0.05940.13080.04420.07990.04220.05890.00430.00580.07580.07880.46490.68090.69110.72290.04620.0787
RankIQA [26]0.04070.08890.03700.06810.03690.05430.00410.00560.08290.08570.32510.64750.67060.70460.04610.0750
DeepIQA [2]0.05710.13020.04170.07210.03220.05450.00480.00590.07870.08090.36720.66320.61620.65190.04740.0765
FIQARankIQ [4]0.03590.08370.03610.05310.02130.03320.00790.00270.06020.06290.06590.16420.30760.34750.02690.0383
PFE [34]0.02150.04230.01920.03170.01070.01380.00230.00290.06400.06520.04490.14350.26150.29260.02000.0283
SER-FIQ [36]0.02330.04510.01850.02930.00800.01390.00250.00330.05900.06070.03970.52180.60160.62100.03050.0422
FaceQnet [14, 15]0.03650.07200.02170.03140.02710.03510.00220.00270.07630.07730.29880.52180.64750.71880.19550.1979
MagFace [28]0.02120.04170.01590.02470.00850.01290.00170.00220.05620.05780.05060.88870.44780.49000.01950.0279
SDD-FIQA [32]0.02530.05620.02160.03050.01460.02010.00210.06430.06570.05250.11880.34040.39280.39280.02150.0307
CR-FIQA(S)(Our)0.02440.05070.01650.02340.01020.01210.00200.00280.05160.05280.04090.08400.26700.33360.01980.0284
CR-FIQA(L)(Our)0.02110.03720.01740.02350.00620.00800.00230.00280.06140.06280.03740.06790.23690.28390.01620.0236
CurricularFace [17]QABRISQUE [29]0.05020.10950.04330.04910.03230.03570.00410.00470.07550.07840.27090.50570.61460.63360.03620.0589
RankIQA [26]0.03590.07520.03940.05100.02980.03560.00390.00450.88060.86560.23460.46540.56440.62120.13610.1558
DeepIQA [2]0.04920.10700.04070.04760.02270.02780.00500.00560.7640.7860.24880.49610.51650.55260.13760.1558
FIQARankIQ [4]0.03140.07150.03650.04170.01860.02490.00180.00240.59590.66451515*1515*1515*1515*1515*1515*
PFE [34]0.01980.03650.01970.02270.01030.01340.00240.00280.63630.65751515*1515*1515*1515*1515*1515*
SER-FIQ [36]0.02110.03810.01670.01930.00740.01110.00250.03561518*1519*1518*1519*1519*1519*1519*1519*
FaceQnet [14, 15]0.03260.06260.02210.02670.02260.02740.00220.02771518*1519*1518*1519*1519*1519*1519*1519*
MagFace [28]0.02000.03640.01670.01950.00780.01110.00160.00211516*1517*1518*1519*1519*1519*1519*1519*
SDD-FIQA [32]0.02300.04620.02190.02540.01380.01850.00210.06271517*1518*1518*1519*1519*1519*1519*1519*
CR-FIQA(S)(Our)0.02270.04460.01560.01980.00970.01480.00200.02551513*1514*1514*1515*1515*1515*1515*1515*
CR-FIQA(L)(Our)0.01980.03360.01620.02000.00540.08800.00230.02991516*1517*1518*1519*1519*1519*1519*1519*
", + "bbox": [ + 109, + 146, + 862, + 568 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/bccaa1745a6c4cd0aba0c843c803d8fbe3c42f672e186bd9926fea5ef137c70a.jpg", + "table_caption": [ + "Table 2. Verification performance on the IJB-C (1:1 mixed verification) [27]. CR-FIQA outperformed SOTA methods under all settings" + ], + "table_footnote": [], + "table_body": "
Quality Estimation1:1 mixed Verification: TAR (%) at
ArcFace [5]ElasticFace [3]MagFace [28]CurricularFace [17]
FAR=1e-6FAR=1e-5FAR=1e-4FAR=1e-6FAR=1e-5FAR=1e-4FAR=1e-6FAR=1e-5FAR=1e-4FAR=1e-6FAR=1e-5FAR=1e-4
-89.8594.4796.2889.1594.5496.4985.6793.0896.6590.4694.8996.58
IAQBRISQUE [29]86.6593.6295.9885.6893.5195.6581.1190.6494.8288.1693.9896.29
RankIQA [26]86.3793.6195.8386.7193.4696.0080.7890.7594.8688.1694.1196.22
DeepIQA [2]81.9791.6494.6778.9391.5994.8173.5386.3492.9082.6592.0495.00
FIQARankIQ [4]88.7894.4296.2088.8894.6496.4585.6392.6695.7090.0094.9396.53
PFE [34]89.5094.5196.3189.1094.6796.5184.9392.4495.6090.3695.0496.54
SER-FIQ [36]89.7494.6596.3290.0594.7996.5786.0293.3595.8090.6695.1196.58
FaceQNet [14, 15]87.8794.0496.1286.2694.0996.2582.9190.5695.0389.6194.6596.36
MagFace [28]89.4994.4196.2289.3794.6996.4685.7592.7195.5490.3495.0296.50
SDD-FIQA [32]89.3994.6196.3488.0794.8296.4984.6992.8395.7389.9195.1296.63
CR-FIQA(S)(Our)89.5994.7896.3590.3094.9796.6386.4593.4895.9590.8295.1396.64
CR-FIQA(L)(Our)90.1694.7596.3690.0094.9296.5887.1293.6795.9090.7995.1296.58
", + "bbox": [ + 104, + 590, + 857, + 734 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "racy [3,5,17,28]), leaving very few samples causing errors and thus lowering the significance of the measured FNMR.", + "bbox": [ + 76, + 750, + 467, + 777 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The XQLFW benchmark is derived from LFW to contain pairs with a maximum difference in quality. The XQLFW images are chosen based on BRISQUE [29] and SER-FIQ [23] quality scores to be either extremely high or low. The use of SER-FIQ [36] in this selection might give a biased edge for SER-FIQ on this benchmark. On XQLFW, our CR-FIQA achieved very close performance to the selection method (SER-FIQ) and is far ahead of all other SOTA", + "bbox": [ + 75, + 779, + 468, + 900 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "methods. Lastly, our proposed CR-FIQA(S) achieved very comparable performance to our CR-FIQA(L), pointing out the robustness of our approach, regardless of the training database and architecture complexity.", + "bbox": [ + 496, + 750, + 890, + 809 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 2 presents the verification performance on the IJB-C 1:1 mixed verification benchmark [27] using quality scores as an embedding weighting term (as defined in [27]) under different experimental settings. For each of the FR models, we report in the first row the evaluation result of the corresponding FR model as defined in the protocol [27] and", + "bbox": [ + 496, + 810, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "5842", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a82f22ad1d52348cae670ed95b1d433880540117ff65dfeac6f27bed43eab300.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 107, + 88, + 294, + 195 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/35fd13f5e6b524ca30ce69a4ca87e66bb352cb00e02ffce4323012f9e319af6b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 295, + 88, + 480, + 196 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/dd3f2ff2716f695cb1b20f5a21e8994a64b92decdde2434660c1580bf37c4872.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 488, + 88, + 674, + 196 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/eeab048702330d160cf068de1e61c9c8c719c70d4a5234bb2c1a337493f81ec3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 674, + 89, + 861, + 196 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/fadcbf964e12e0dbe9bcadd5f6854b07c15d34449b1b5e0006c456b408d746ae.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 106, + 198, + 294, + 305 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/7741785b04c7e29ac5e32d5519d423666998ec255da71cdf7abe0cd322fb48f3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 295, + 198, + 480, + 305 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/8ab76e9fddc9a173b612786ef91a3f8c3ec9eabb5edfcfa96acb622fa9884c91.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 488, + 198, + 674, + 305 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/e2e3e85d37ad943874fb5cf960272f0d698373ca90379092b9d3c4a0fa897358.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 674, + 198, + 861, + 305 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/771b35430170d9df9908ce8d99f810923a68162605e46e89191629d401bdf2a7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 106, + 306, + 295, + 414 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/80a3098b1c61c8fa07b1fe66a09f99bd622791f47301a62ec369cf59f48fbed5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 295, + 306, + 480, + 414 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/b74275bebd231178c2913a54158c46f8c1aafdcd4751a5c33b91115c833e7f16.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 488, + 306, + 674, + 412 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/754b9d7570e68dac2860bd33ad61da7fb0e41879677bb01c05e17149e0f8f39e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 674, + 306, + 861, + 412 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/fd3370c640db7d3890be16d27af380026231c669c8b5ff0e18eee497f6a5dac8.jpg", + "image_caption": [ + "Figure 4. ERC (FNMR at FMR1e-3 vs reject) curves for all evaluated benchmarks using ArcFace and ElasticFace FR models corresponding to Table 1 results. The visual evaluation, ERC curves, using MagFace and CurricularFace FR models are provided in the supplementary material. The proposed CR-FIQA(L) and CR-FIQA(S) are marked with solid blue and red lines, respectively. CR-FIQA leads to lower verification error, when rejecting a fraction of images, of the lowest quality, in comparison to SOTA methods (faster decaying curve) under most experimental settings." + ], + "image_footnote": [ + "BRIQUE RankIQ DeepIQ $\\rightarrow$ RankIQ PFE SER-FIQ FaceQnet MagFace SDD-FIQA CR-FIQA(S) Our) CR-FIQA(L) Our" + ], + "bbox": [ + 106, + 415, + 295, + 523 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/c51273b871eb43e818f7674904a61d55ce6efead9e5ec90af238d4680af0f9b2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 295, + 415, + 480, + 523 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/4574db913f7abfee3cf4c2214a57e182c99d16898a22d7961980aa2737ed2843.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 488, + 415, + 674, + 523 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/e0e060cd3b1e3f8ff022c16b61b1dd512397d2ca7e17851a36ebbc283c3e656c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 674, + 415, + 861, + 523 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "the corresponding released evaluation scripts [3, 5, 17, 28], i.e. without considering the FIQ. Our proposed CR-FIQA significantly leads to higher verification performance than all evaluated SOTA methods, when the quality score is used as an embedding weighting term (Table 2). This achievement is observable under all experimental settings (Table 2). Another outcome of this evaluation is that the integration of CR-FIQA leads to SOTA verification performance on one of the most challenging FR benchmarks, IJB-C [27].", + "bbox": [ + 75, + 613, + 470, + 750 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "7. Conclusion", + "text_level": 1, + "bbox": [ + 76, + 756, + 194, + 771 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this work, we propose the CR-FIQA approach that probes the relative classifiability of training samples of the FR model and utilize this observation to learn to predict the utility of any given sample in achieving an accurate FR performance. We experimentally prove the theorized relationship between the sample relative classifiability and FIQ and build on that towards our CR-FIQA. The CR-FIQA training paradigm simultaneously learns to optimize the class center", + "bbox": [ + 75, + 779, + 470, + 902 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "while learning to predict sample relative classifiability. The presented ablation studies and the extensive experimental results prove the effectiveness of the proposed CR-FIQA approach, and its design choices, as an FIQ method. The reported results demonstrated that our proposed CR-FIQA outperformed SOTA methods repeatedly across multiple FR models and on multiple benchmarks, including ones with a large age gap (AgeDb-30, Adience, CALFW), large quality difference (XQLFW), large pose variation (CPLFW, CFPFP), and extremely large-scale and challenging FR benchmarks (IJB-C).", + "bbox": [ + 496, + 613, + 890, + 779 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgment This research work has been funded by the German Federal Ministry of Education and Research and the Hessian Ministry of Higher Education, Research, Science and the Arts within their joint support of the National Research Center for Applied Cybersecurity ATHENE. This work has been partially funded by the German Federal Ministry of Education and Research (BMBF) through the Software Campus Project.", + "bbox": [ + 496, + 779, + 892, + 901 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "5843", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Lacey Best-Rowden and Anil K. Jain. Learning face image quality from human assessments. IEEE Trans. Inf. Forensics Secur., 13(12):3064–3077, 2018. 1, 2", + "[2] Sebastian Bosse, Dominique Maniry, Klaus-Robert Müller, Thomas Wiegand, and Wojciech Samek. Deep neural networks for no-reference and full-reference image quality assessment. IEEE Trans. Image Process., 27(1):206-219, 2018. 6, 7", + "[3] Fadi Boutros, Naser Damer, Florian Kirchbuchner, and Arjan Kuijper. Elasticface: Elastic margin loss for deep face recognition. In IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, CVPR Workshops 2022, New Orleans, LA, USA, June 19-20, 2022, pages 1577-1586. IEEE, 2022. 3, 5, 6, 7, 8", + "[4] Jiansheng Chen, Yu Deng, Gaocheng Bai, and Guangda Su. Face image quality assessment based on learning to rank. IEEE Signal Process. Lett., 22(1):90-94, 2015. 2, 6, 7", + "[5] Jiankang Deng, Jia Guo, Niannan Xue, and Stefanos Zafeiriou. Arcface: Additive angular margin loss for deep face recognition. In IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2019, Long Beach, CA, USA, June 16-20, 2019, pages 4690-4699. Computer Vision Foundation / IEEE, 2019. 3, 4, 5, 6, 7, 8", + "[6] Eran Eidinger, Roee Enbar, and Tal Hassner. Age and gender estimation of unfiltered faces. IEEE Trans. Inf. Forensics Secur., 9(12):2170-2179, 2014. 5, 7", + "[7] Frontex. Best practice technical guidelines for automated border control (abc) systems, 2015. 5", + "[8] Biying Fu, Cong Chen, Olaf Henniger, and Naser Damer. A deep insight into measuring face image utility with general and face-specific image quality metrics. pages 1121-1130, 2022. 1, 6", + "[9] Ross B. Girshick. Fast R-CNN. In 2015 IEEE International Conference on Computer Vision, ICCV 2015, Santiago, Chile, December 7-13, 2015, pages 1440-1448. IEEE Computer Society, 2015. 4", + "[10] P. Grother, M. Ngan A. Hom, and K. Hanaoka. Ongoing face recognition vendor test (frvt) part 5: Face image quality assessment (4th draft). In National Institute of Standards and Technology. Tech. Rep., Sep. 2021. 1, 5", + "[11] P. Grother and E. Tabassi. Performance of biometric quality measures. IEEE Trans. on Pattern Analysis and Machine Intelligence, 29(4):531-543, Apr. 2007. 5", + "[12] Yandong Guo, Lei Zhang, Yuxiao Hu, Xiaodong He, and Jianfeng Gao. Ms-celeb-Im: A dataset and benchmark for large-scale face recognition. In Bastian Leibe, Jiri Matas, Nicu Sebe, and Max Welling, editors, Computer Vision - ECCV 2016 - 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part III, volume 9907 of Lecture Notes in Computer Science, pages 87-102. Springer, 2016. 4, 5", + "[13] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In 2016 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2016, Las Vegas, NV, USA, June 27-30, 2016, pages 770-778. IEEE Computer Society, 2016. 4, 5" + ], + "bbox": [ + 78, + 114, + 470, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[14] Javier Hernandez-Ortega, Javier Galbally, Julian Fierrez, and Laurent Beslay. Biometric quality: Review and application to face recognition with faceqnet. CoRR, abs/2006.03298, 2020. 1, 2, 6, 7", + "[15] Javier Hernandez-Ortega, Javier Galbally, Julian Fierrez, Rudolf Haraksim, and Laurent Beslay. Faceqnet: Quality assessment for face recognition based on deep learning. In 2019 International Conference on Biometrics, ICB 2019, Crete, Greece, June 4-7, 2019, pages 1-8. IEEE, 2019. 1, 2, 6, 7", + "[16] Gary B. Huang, Manu Ramesh, Tamara Berg, and Erik Learned-Miller. Labeled faces in the wild: A database for studying face recognition in unconstrained environments. Technical Report 07-49, University of Massachusetts, Amherst, October 2007. 5, 7", + "[17] Yuge Huang, Yuhan Wang, Ying Tai, Xiaoming Liu, Pengcheng Shen, Shaoxin Li, Jilin Li, and Feiyue Huang. Curricularface: Adaptive curriculum learning loss for deep face recognition. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2020, Seattle, WA, USA, June 13-19, 2020, pages 5900-5909. Computer Vision Foundation / IEEE, 2020. 3, 5, 6, 7, 8", + "[18] ISO/IEC JTC1 SC17 WG3. Portrait Quality - Reference Facial Images for MRTD. International Civil Aviation Organization, 2018. 2", + "[19] ISO/IEC JTC1 SC37 Biometrics. ISO/IEC 29794-1:2016 Information technology - Biometric sample quality - Part 1: Framework. International Organization for Standardization, 2016. 1", + "[20] ISO/IEC JTC1 SC37 Biometrics. ISO/IEC 2382-37:2017 Information technology - Vocabulary - Part 37: Biometrics. International Organization for Standardization, 2017. 1", + "[21] ISO/IEC JTC1 SC37 Biometrics. ISO/IEC 19795-1:2021 Information technology — Biometric performance testing and reporting — Part 1: Principles and framework. International Organization for Standardization, 2021. 5", + "[22] Thorsten Joachims. Optimizing search engines using click-through data. In Proceedings of the Eighth ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, July 23-26, 2002, Edmonton, Alberta, Canada, pages 133-142. ACM, 2002. 2", + "[23] Martin Knoche, Stefan Hörmann, and Gerhard Rigoll. Cross-quality LFW: A database for analyzing cross-resolution image face recognition in unconstrained environments. In 16th IEEE International Conference on Automatic Face and Gesture Recognition, FG 2021, Jodhpur, India, December 15-18, 2021, pages 1-5. IEEE, 2021. 5, 7", + "[24] Weiyang Liu, Yandong Wen, Zhiding Yu, Ming Li, Bhiksha Raj, and Le Song. Spherical face: Deep hypersphere embedding for face recognition. In 2017 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2017, Honolulu, HI, USA, July 21-26, 2017, pages 6738-6746. IEEE Computer Society, 2017. 3", + "[25] Weiyang Liu, Yandong Wen, Zhiding Yu, Ming Li, Bhiksha Raj, and Le Song. SphereFace: Deep hypersphere embedding for face recognition. In Proc. of the IEEE Conf. on Computer Vision and Pattern Recognition, pages 212-220, 2017. 3" + ], + "bbox": [ + 501, + 92, + 890, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "5844", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[26] Xialei Liu, Joost van de Weijer, and Andrew D. Bagdanov. Rankiaq: Learning from rankings for no-reference image quality assessment. In IEEE International Conference on Computer Vision, ICCV 2017, Venice, Italy, October 22-29, 2017, pages 1040-1049. IEEE Computer Society, 2017. 1, 6, 7", + "[27] Brianna Maze, Jocelyn C. Adams, James A. Duncan, Nathan D. Kalka, Tim Miller, Charles Otto, Anil K. Jain, W. Tyler Niggel, Janet Anderson, Jordan Cheney, and Patrick Grother. IARPA janus benchmark - C: face dataset and protocol. In 2018 International Conference on Biometrics, ICB 2018, Gold Coast, Australia, February 20-23, 2018, pages 158-165. IEEE, 2018. 5, 7, 8", + "[28] Qiang Meng, Shichao Zhao, Zhida Huang, and Feng Zhou. Magface: A universal representation for face recognition and quality assessment. In IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2021, virtual, June 19-25, 2021, pages 14225-14234. Computer Vision Foundation / IEEE, 2021. 1, 2, 3, 5, 6, 7, 8", + "[29] Anish Mittal, Anush Krishna Moorthy, and Alan Conrad Bovik. No-reference image quality assessment in the spatial domain. IEEE Trans. Image Process., 21(12):4695-4708, 2012. 1, 6, 7", + "[30] Anish Mittal, Rajiv Soundararajan, and Alan C. Bovik. Making a \"completely blind\" image quality analyzer. IEEE Signal Process. Lett., 20(3):209-212, 2013. 1", + "[31] Stylianos Moschoglou, Athanasios Papaioannou, Christos Sagonas, Jiankang Deng, Irene Kotsia, and Stefanos Zafeiriou. Agedb: The first manually collected, in-the-wild age database. In 2017 IEEE CVPRW, CVPR Workshops 2017, Honolulu, HI, USA, July 21-26, 2017, pages 1997-2005. IEEE Computer Society, 2017. 5, 7", + "[32] Fu-Zhao Ou, Xingyu Chen, Ruixin Zhang, Yuge Huang, Shaoxin Li, Jilin Li, Yong Li, Liujuan Cao, and Yuan-Gen Wang. SDD-FIQA: unsupervised face image quality assessment with similarity distribution distance. In IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2021, virtual, June 19-25, 2021, pages 7670-7679. Computer Vision Foundation / IEEE, 2021. 1, 2, 5, 6, 7", + "[33] Soumyadip Sengupta, Jun-Cheng Chen, Carlos Domingo Castillo, Vishal M. Patel, Rama Chellappa, and David W. Jacobs. Frontal to profile face verification in the wild. In 2016 IEEE Winter Conference on Applications of Computer Vision, WACV 2016, Lake Placid, NY, USA, March 7-10, 2016, pages 1-9. IEEE Computer Society, 2016. 5, 7", + "[34] Yichun Shi and Anil K. Jain. Probabilistic face embeddings. In 2019 IEEE/CVF International Conference on Computer Vision, ICCV 2019, Seoul, Korea (South), October 27 - November 2, 2019, pages 6901-6910. IEEE, 2019. 1, 2, 5, 6, 7", + "[35] Philipp Terhörst, Jan Niklas Kolf, Naser Damer, Florian Kirchbuchner, and Arjan Kuijper. Face quality estimation and its correlation to demographic and non-demographic bias in face recognition. In 2020 IEEE International Joint Conference on Biometrics, IJCB 2020, Houston, TX, USA, September 28 - October 1, 2020, pages 1-11. IEEE, 2020. 1", + "[36] Philipp Terhörst, Jan Niklas Kolf, Naser Damer, Florian Kirchbuchner, and Arjan Kuijper. SER-FIQ: unsupervised" + ], + "bbox": [ + 78, + 90, + 470, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "estimation of face image quality based on stochastic embedding robustness. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2020, Seattle, WA, USA, June 13-19, 2020, pages 5650-5659. Computer Vision Foundation / IEEE, 2020. 1, 2, 6, 7", + "[37] Hao Wang, Yitong Wang, Zheng Zhou, Xing Ji, Dihong Gong, Jingchao Zhou, Zhifeng Li, and Wei Liu. Cosface: Large margin cosine loss for deep face recognition. In 2018 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2018, Salt Lake City, UT, USA, June 18-22, 2018, pages 5265-5274. IEEE Computer Society, 2018. 3", + "[38] Weidi Xie, Jeffrey Byrne, and Andrew Zisserman. Inducing predictive uncertainty estimation for face verification. In 31st British Machine Vision Conference 2020, BMVC 2020, Virtual Event, UK, September 7-10, 2020. BMVA Press, 2020. 1, 2", + "[39] Dong Yi, Zhen Lei, Shengcai Liao, and Stan Z. Li. Learning face representation from scratch. CoRR, abs/1411.7923, 2014. 2, 4, 5, 6", + "[40] T. Zheng and W. Deng. Cross-posed lfw: A database for studying cross-posed face recognition in unconstrained environments. Technical Report 18-01, Beijing University of Posts and Telecommunications, February 2018. 5, 7", + "[41] Tianyue Zheng, Weihong Deng, and Jiani Hu. Cross-age LFW: A database for studying cross-age face recognition in unconstrained environments. CoRR, abs/1708.08197, 2017. 5, 7" + ], + "bbox": [ + 501, + 92, + 890, + 470 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "5845", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2023/CR-FIQA_ Face Image Quality Assessment by Learning Sample Relative Classifiability/2ae421c8-bef3-40d8-a9e9-022734d8a26e_model.json b/2023/CR-FIQA_ Face Image Quality Assessment by Learning Sample Relative Classifiability/2ae421c8-bef3-40d8-a9e9-022734d8a26e_model.json new file mode 100644 index 0000000000000000000000000000000000000000..8f34bf93c8771af625be3c1f0c19899138ecc5d3 --- /dev/null +++ b/2023/CR-FIQA_ Face Image Quality Assessment by Learning Sample Relative Classifiability/2ae421c8-bef3-40d8-a9e9-022734d8a26e_model.json @@ -0,0 +1,1848 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.12, + 0.13, + 0.852, + 0.176 + ], + "angle": 0, + "content": "CR-FIQA: Face Image Quality Assessment by Learning Sample Relative Classifiability" + }, + { + "type": "text", + "bbox": [ + 0.163, + 0.203, + 0.808, + 0.275 + ], + "angle": 0, + "content": "Fadi Boutros1, Meiling Fang1,2, Marcel Klemt1, Biying Fu1, Naser Damer1,2 \n1Fraunhofer Institute for Computer Graphics Research IGD, Darmstadt, Germany \n2Department of Computer Science, TU Darmstadt, Darmstadt, Germany \nEmail: fadi.boutros@igd.fraunhofer.de" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.309, + 0.314, + 0.327 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.341, + 0.474, + 0.613 + ], + "angle": 0, + "content": "Face image quality assessment (FIQA) estimates the utility of the captured image in achieving reliable and accurate recognition performance. This work proposes a novel FIQA method, CR-FIQA, that estimates the face image quality of a sample by learning to predict its relative classifiability. This classifiability is measured based on the allocation of the training sample feature representation in angular space with respect to its class center and the nearest negative class center. We experimentally illustrate the correlation between the face image quality and the sample relative classifiability. As such property is only observable for the training dataset, we propose to learn this property by probing internal network observations during the training process and utilizing it to predict the quality of unseen samples. Through extensive evaluation experiments on eight benchmarks and four face recognition models, we demonstrate the superiority of our proposed CR-FIQA over state-of-the-art (SOTA) FIQA algorithms." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.62, + 0.21, + 0.635 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.638, + 0.47, + 0.713 + ], + "angle": 0, + "content": "Face image utility indicates the utility (value) of an image to face recognition (FR) algorithms [1, 19]. This utility is measured with a scalar, namely the face image quality (FIQ) score, following the definition in ISO/IEC 2382-37 [20] and the FR Vendor Test (FRVT) for FIQA [10]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.713, + 0.471, + 0.88 + ], + "angle": 0, + "content": "As FIQA measures the face utility to FR algorithm, it does not necessary reflects, and does not aim at measuring, the perceived image quality, e.g. a profile face image can be of high perceived quality but of low utility to FR algorithm [35]. Assessing this perceived image quality has been addressed in the literature by general image quality assessment (IQA) methods [26, 29, 30] and is different than assessing the utility of an the image for FR. This is reflected by FIQA methods [28, 32, 36] significantly outperforming IQA methods [26, 29, 30] in measuring the utility [19] of face images in FR, as demonstrated in [8, 28, 36]." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.311, + 0.893, + 0.492 + ], + "angle": 0, + "content": "SOTA FIQA methods focused either on creating concepts to label the training data with FIQ scores and then learn a regression problem [14, 15, 32], or on developing a link between face embedding properties under certain scenarios and the FIQ [28, 34, 36]. Generally, the second approach led to better FIQA performances with most works mentioning the error-prone labeling of the ground truth quality in the first research direction as a possible reason [28, 36]. However, in the second category, transferring the information in network embeddings into an FIQ score is not a learnable process, but rather a form of statistical analysis, which might not be optimal." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.494, + 0.895, + 0.812 + ], + "angle": 0, + "content": "This paper proposes a novel learning paradigm to assess FIQ, namely the CR-FIQA. Our concept is based on learning to predict the classifiability of FR training samples by probing internal network observations that point to the relative proximity of these samples to their class centers and negative class centers. This regression is learned simultaneously with a conventional FR training process that minimizes the distance between the training samples and their class centers. Linking the properties that cause high/low classifiability of a training sample to the properties leading to high/low FIQ, we can use our CR-FIQA to predict the FIQ of any given sample. We empirically prove the theorized link between classifiability (Section 3.3) and FIQ and conduct thorough ablation studies on key aspects of our CR-FIQA design (Section 5). The proposed CR-FIQA is evaluated on eight benchmarks along with SOTA FIQAs. The reported results on four FR models demonstrate the superiority of our proposed CR-FIQA over SOTA methods and the stability of its performance across different FR models. An overview of the proposed CR-FIQA is presented in Figure 1 and will be clarified in detail in this paper." + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.82, + 0.64, + 0.836 + ], + "angle": 0, + "content": "2. Related work" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.84, + 0.895, + 0.903 + ], + "angle": 0, + "content": "The recent SOTA FIQA approaches can be roughly grouped into two main categories. The first are approaches that learn a straight forward regressions problem to assess a FIQ score [1,14,15,32,38]. The second category uses prop" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.887, + 0.366, + 0.9 + ], + "angle": 0, + "content": "1https://github.com/fdbtrs/CR-FIQA" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.957 + ], + "angle": 0, + "content": "5836" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.182, + 0.09, + 0.79, + 0.291 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.302, + 0.895, + 0.414 + ], + "angle": 0, + "content": "Figure 1. An overview of our CR-FIQA training paradigm. We propose to simultaneously learn to optimize the class center \\((\\mathcal{L}_{Arc})\\), while learning to predict an internal network observation i.e. the allocation of the feature representation of sample \\(x\\) in feature space, with respect to its class center \\(w_{1}\\) and nearest negative class center \\(w_{2}\\) \\((\\mathcal{L}_{\\mathcal{CR}})\\). The figure in the red rectangle illustrates the angle between two samples \\(x_{1}\\) and \\(x_{2}\\) (belong to identity 1) and their class center \\(w_{1}\\). The plot on the right of the figure shows the distribution of the cosine similarity between training samples and their class centers (CCS) and nearest negative class centers (NNCCS) obtained from ResNet-50 trained on CASIA-WebFace [39]. The example images on the top-right of this plot are of high CCS values and the ones on the top-left are of low CCS values (notice the correspondence to perceived quality). These samples are selected from CASIA-WebFace [39]. During the testing mode, the classification layer is removed and the output of the regression layer is used to predict the FIQA of testing samples." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.422, + 0.473, + 0.892 + ], + "angle": 0, + "content": "erties of the FR model responses to face samples to estimate the sample quality without explicitly learning a typical supervised regression that requires quality labels [28, 34, 36]. In the first category, the innovation focused on creating the FIQ labels for training. These quality labels included human-labeled quality labels [1], the FR genuine comparison score between a sample and an ICAO [18] compliant sample [14, 15], the FR comparison score involving the labeled sample (assumed to have the lower quality in the comparison pair) [38], and the Wasserstein distance between a randomly selected genuine and imposter FR comparisons with the labeled sample [32]. These solutions generally trained a regression network to predict the quality label, using both, trained-from-scratch networks in some cases [38], and pre-trained FR networks in other cases [14, 15, 32]. A slightly different approach, however also based on learning from labels, focuses on learning to predict the sample quality as a rank [22] based on FR performance-based training rank labels of a set of databases [4]. In the second category, the innovation was rather focused on linking face embedding properties under certain scenarios to the FIQ, without the explicit need for quality-labeled data. In [36], the assessed sample is passed through an FR network multiple times, each with a different random dropout pattern. The robustness of the resulting embeddings, represented by the sigmoid of the negative mean of the Euclidean distances between the embeddings, is considered the FIQ score. In [28], the FIQ score is calculated as the magnitude of the sample embedding. This is based on training the FR model using a loss that adapts the penalty margin loss based on this magnitude, and thus links the closeness of a sample to its class" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.422, + 0.895, + 0.558 + ], + "angle": 0, + "content": "center to the unnormalized embedding magnitude. While in [34], the solution produces both, an FR embedding and a gaussian variance (uncertainty) vector, from a face sample. The inverse of harmonic mean of the uncertainty vector is considered as the FIQ score. Our CR-FIQA learns a regression problem to estimate the FIQ score, however, unlike previous works, without relying on preset labels, but rather learn a dynamic internal network observations (during training) that point out sample classifiability." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.565, + 0.61, + 0.581 + ], + "angle": 0, + "content": "3. Approach" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.584, + 0.895, + 0.903 + ], + "angle": 0, + "content": "This section presents our proposed Certainty Ratio Face Image Quality Assessment (CR-FIQA) approach, which inspects internal network observations to learn to predict the sample relative classifiability. This classifiability prediction is then used to estimate the FIQ. An overview of the proposed CR-FIQA approach is presented in Figure 1. During the training phase of an FR model, the model can conveniently push the high-quality samples close to their class center and relatively far from other class centers. Conversely, the FR is not able to push, to the same degree, low-quality samples to their class center, and thus they will remain relatively farther from their class center than the high-quality ones. Based on this assumption, we theorize our approach by stating that the properties that cause a face sample to lay relatively closer to its class center during training are the ones that make it a high-quality sample, and vice versa. Therefore, learning to predict such properties in any given sample would lead to learning to assess this sample quality. To learn to perform such assessment, our training paradigm targets learning internal network observations that evolve during the FR training phase, where these observa" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5837" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.198 + ], + "angle": 0, + "content": "tions act as a training objective. The predictions of such training paradigm can be simply stated as answering a question: if a given sample was hypothetically part of the FR model training (which it is not), how relatively close would it be to its class center? Answering this question would give us an indication of this sample quality as will be shown in detail in this paper." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.198, + 0.468, + 0.304 + ], + "angle": 0, + "content": "In the rest of this section, We formalize and empirically rationalize our proposed CR-FIQA approach and its components. To do that, we start by shortly revisiting angular margin penalty-based softmax loss utilized to optimize the class centers of the FR model. Then present a detailed description of our proposed CR-FIQA concept and the associated training paradigm." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.307, + 0.468, + 0.323 + ], + "angle": 0, + "content": "3.1. Revisiting Margin Penalty-based Softmax Loss" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.326, + 0.468, + 0.522 + ], + "angle": 0, + "content": "Angular margin penalty-based softmax is a widely used loss function for training FR models [3,5,17,28]. It extends over softmax loss by deploying angular penalty margin on the angle between the deep features and their corresponding weights. Margin penalty-based softmax loss aims to push the decision boundary of softmax, and thus enhance intraclass compactness and inter-class discrepancy. From this family of loss functions, this work utilizes ArcFace loss [5] to optimize the distance between the training samples and their class center. Our choice of ArcFace loss is based on the SOTA performance achieved by ResNet-100 network trained with ArcFace on mainstream benchmarks [5]. Formally, ArcFace loss is defined as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.079, + 0.524, + 0.468, + 0.59 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {A r c} = \\frac {1}{N} \\sum_ {i \\in N} - \\log \\frac {e ^ {s \\left(\\cos \\left(\\theta_ {y _ {i}} + m\\right)\\right)}}{e ^ {s \\left(\\cos \\left(\\theta_ {y _ {i}} + m\\right)\\right)} + \\sum_ {j = 1 , j \\neq y _ {i}} ^ {C} e ^ {s \\left(\\cos \\left(\\theta_ {j}\\right)\\right)}}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.591, + 0.468, + 0.802 + ], + "angle": 0, + "content": "where \\(N\\) is the batch size, \\(C\\) is the number of classes (identities), \\(y_{i}\\) is the class label of sample \\(i\\) (in range \\([1,C])\\), \\(\\theta_{y_i}\\) is the angle between the features \\(x_{i}\\) and the \\(y_{i}\\)-th class center \\(w_{y_i}\\). \\(x_{i} \\in R^{d}\\) is the deep feature embedding of the last fully connected layer of size \\(d\\). \\(w_{y_i}\\) is the \\(y_{i}\\)-th column of weights \\(W \\in R_C^d\\) of the classification layer. \\(\\theta_{y_i}\\) is defined as \\(x_{i} w_{y_{i}}^{T} = \\| x_{i} \\| \\| w_{y_{i}} \\| \\cos (\\theta_{y_{i}})\\) [24]. The weights and the feature norms are fixed to \\(\\| w_{y_i} \\| = 1\\) and \\(\\| x_{i} \\| = 1\\), respectively, using \\(l_2\\) normalization as defined in [24, 37]. The decision boundary, in this case, depends on the angle cosine between \\(x_{i}\\) and \\(w_{y_i}\\). \\(m > 0\\) is an additive angular margin proposed by ArcFace [5] to enhance the intra-class compactness and inter-class discrepancy. Lastly, \\(s\\) is the scaling parameter [37]." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.806, + 0.235, + 0.822 + ], + "angle": 0, + "content": "3.2. Certainty Ratio" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.468, + 0.901 + ], + "angle": 0, + "content": "In this section, we formulate and empirically rationalize the main concepts that build our FIQA solution. We derive our Certainty Ratio (CR) to estimate the sample relative classifiability. Additionally, we experimentally illustrate the strong relationship between our CR measure and FIQ." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.198 + ], + "angle": 0, + "content": "Certainty Ratio During the FR model training phase, the model is trained to enhance the separability between the classes (identities) by pushing each sample \\( x_{i} \\) to be close to its class center \\( w_{y_i} \\) and far from the other (negative) class centers \\( w_{j}, j \\neq y_{i} \\). Based on this, we first define the Class Center Angular Similarity (CCS) as the proximity between \\( x_{i} \\) and its class center \\( w_{y_i} \\), as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.627, + 0.202, + 0.892, + 0.218 + ], + "angle": 0, + "content": "\\[\nC C S _ {x _ {i}} = \\cos \\left(\\theta_ {y _ {i}}\\right), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.23, + 0.892, + 0.336 + ], + "angle": 0, + "content": "where \\(\\theta_{y_i}\\) is the angle between \\(x_i\\) and its class center \\(w_{y_i}\\), where the weights of the last fully connected layer of the FR model trained with softmax loss are considered as the centers for each class [5,25]. Then, we define the Closest Nearest Negative Class Center Angular Similarity (NNCCS) as proximity between \\(x_i\\) and the nearest negative class center \\(w_j\\), \\(j \\neq y_i\\). Formally, NNCCS is defined as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.583, + 0.34, + 0.892, + 0.369 + ], + "angle": 0, + "content": "\\[\nN N C C S _ {x _ {i}} = \\max _ {j = 1, j \\neq y _ {i}} ^ {C} (\\cos \\left(\\theta_ {j}\\right)), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.379, + 0.892, + 0.543 + ], + "angle": 0, + "content": "where \\(\\theta_{j}\\) is the angle between \\(x_{i}\\) and \\(w_{j}\\). As we theorize, when the FR model converges, the high-quality samples are pushed closer to their class centers (high CCS) in relation to their distance to neighbouring negative class centers (low NNCCS). However, low-quality samples can not be pushed as close to their class centers. A sample able to achieve high CCS with respect to NNCCS is a sample easily correctly classified during training, and thus is relatively highly classifiable. We thus measure this relative classifiability by the ratio of CCS to NNCS, which we note as the Certainty Ratio (CR), as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.585, + 0.546, + 0.892, + 0.58 + ], + "angle": 0, + "content": "\\[\nC R _ {x _ {i}} = \\frac {C C S _ {x _ {i}}}{N N C C S _ {x _ {i}} + (1 + \\epsilon)}, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.589, + 0.892, + 0.786 + ], + "angle": 0, + "content": "where the \\(1 + \\epsilon\\) term is added to insure a positive above zero denominator, i.e. shift the NNCCS value range from \\([-1, +1]\\) to \\([\\epsilon, 2 + \\epsilon]\\). This ensures that the CR of a sample with a lower NNCCS is relatively higher than a sample with a higher NNCCS, given the same CCS, i.e. NNCCS regulates the CCS value in relation to neighbouring classes. The \\(\\epsilon\\) is set to \\(1e - 9\\) in our experiments. The optimal CR is obtained when the CCS is approaching the maximum cosine similarity value (+1) and the NNCCS is approaching the minimum cosine similarity value (-1), i.e. the training sample is capable of being pushed to its class center, and far away from the closest negative class center, and thus it is highly classifiable." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.791, + 0.798, + 0.807 + ], + "angle": 0, + "content": "3.3. Relation between the CR and FIQ" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.811, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Here, we empirically prove the theorized relationship between the CR and FIQ (defined earlier as image utility). Namely, we want to answer: if the CR values achieved by training samples of an FR model were used as FIQ, would they behave as expected from an optimal FIQ? If yes, then the face image properties leading to high/low CR" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5838" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.473, + 0.455 + ], + "angle": 0, + "content": "do also theoretically lead to high/low FIQ. To answer this question we conducted an experiment on a ResNet-50 [13] FR model trained on CASIA-WebFace [39] with ArcFace loss [5] (noted as R50(CASIA)). Specifically, we calculate the CR, CCS, and NNCCS values from the trained model for all samples in the training dataset (0.5M images of 10K identities). An insight on the resulting CCS and NNCCS values (CR being a derivative measure) is given as value distributions in Figure 1, showing that these measures vary between different samples. Furthermore, based on the calculated scores, we plot Error vs. Reject Curves (ERC) (described in Section 4) to demonstrate the relationship between the CR, as an FIQ measure, and FR performance. To calculate the FR performance in the ERC curve, we extract the feature embedding of CASIA-WebFace [39] using a ResNet-100 model [13] trained on MS1M-V2 [5, 12] with ArcFace (noted as R100(MS1M-V2)). We utilize a different model (trained on a different database) to extract the embedding (R100(MS1M-V2)) than the one used to calculate CR, CCS, and NNCCS (R50(CASIA)) to provide fair evaluation where the FR performance is evaluated on unseen data. Then, we perform \\( n : n \\) comparisons between all samples of CASIA-WebFace using feature embedding obtained from the R100(MS1M-V2)." + }, + { + "type": "image", + "bbox": [ + 0.097, + 0.464, + 0.26, + 0.536 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.174, + 0.539, + 0.192, + 0.549 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.274, + 0.464, + 0.437, + 0.535 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.352, + 0.538, + 0.371, + 0.549 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.551, + 0.47, + 0.621 + ], + "angle": 0, + "content": "Figure 2. ERCs showing the verification performance as False None Match Rate (FNMR) at False Match Rate (FMR) of 1e-3 (2a) and 1e-4 (2b) with CCS, NNCCS and CR as FIQ vs. rejection ratio. This ERC plots show the effectiveness of rejecting samples with the lowest CCS and CR on the performance" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.628, + 0.47, + 0.689 + ], + "angle": 0, + "content": "Figure 2 presents the ERC of CR, CCS, and NNCCS experimentally used as FIQ. An FIQ measure would cause the ERC to drop as rapidly as possible when rejecting a larger fraction of low-quality samples (moving to the right)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.689, + 0.471, + 0.901 + ], + "angle": 0, + "content": "It can be clearly noticed in Figure 2 that the CCS and CR do behave as we would expect from a good performing FIQ, as the verification error value drops rapidly when rejecting low quality (low CCS and CR) samples. It can be also observed that the CR does that more steadily when compared to CCS. This points out that adding the scaling term NNCCS in CR calculation can enhance the representation of the CCS as an FIQ measure, which will be clearer later when we experimentally evaluate our CR-FIQA approach in Section 5. As expected, the NNCCS measure by itself does not strongly act as an FIQ measure would, demonstrated by the relatively flat ERC in Figure 2, as it only considers the distance to the nearest negative class. This empirical evaluation does provide a confirming answer to the previ" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.152 + ], + "angle": 0, + "content": "ously stated question by affirming that the CR does act as expected from an FIQ measure and thus, theoretically, one can strongly link the image properties that cause high/low CR in the FR training data to these causing high/low FIQ." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.16, + 0.834, + 0.176 + ], + "angle": 0, + "content": "3.4. Quality Estimation Training Paradigm" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.18, + 0.892, + 0.844 + ], + "angle": 0, + "content": "In the previous section, we proved that the CR does behave as an FIQ would, and thus, it can also relate to image properties that dictate FIQ. However, the CR measure is only observable for samples in the FR training dataset, where the class centers are known. In a real case scenario, the FIQ measure should be assessed to any single image, i.e. unseen evaluation data. Considering this, and in an effort to predict what the CR value would be for a given sample if hypothetically it was part of the FR training, we propose to simultaneously learn to predict the CR from the training dataset while optimizing the class centers (typical FR training) during the training phase, i.e. the CR-FIQA model. To enable this, we add a single regression layer to the FR model. The input of the regression layer is a feature embedding \\( x_{i} \\) and the output is an estimation of the CR. The output of this regression layer is used later to predict the FIQ score of the unseen sample, e.g. from the evaluation dataset. Thus, we capture the properties that make the CR high/low to predict the FIQ of any given sample. Towards this goal, during the training phase, the model (in Figure 1) has two learning objectives: a) It is trained to optimize the distance between the samples and the class centers using ArcFace loss defined in Equation 1. b) It is trained to predict the internal network observation, CR, using Smooth L1-Loss [9] applied between the output of the regression layer (P) and the CR calculated as in Equation 5. Smooth L1-loss can be interpreted as a combination of L1 and L2-losses by defining a threshold \\( \\beta \\) that changes between them [9]. Our choice for smooth-l1 loss is based on: 1) It is less sensitive to outliers than l2. The derivative of L2 loss increases when the difference between the prediction and ground-truth label is increased, making the derivative of loss values large at the early stage of the training, leading to unstable training. Additionally, L2 loss can easily generate gradient explosion [9] when there are outliers in the training data. 2) L1 loss can lead to stable training. However, the absolute values of the difference between prediction and ground truth are small, especially in the later stage of the training. Therefore, the model accuracy can hardly be improved at a later stage of the training as the loss function will fluctuate around a stable value. Combining L1 and L2 as in Smooth L1-loss avoids gradient explosion, which might be caused by L2 and facilitates better convergence than L1. The loss leading to the second objective is then given as:" + }, + { + "type": "equation", + "bbox": [ + 0.5, + 0.848, + 0.894, + 0.901 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {C R} = \\frac {1}{N} \\sum_ {i \\in N} \\left\\{ \\begin{array}{l l} \\frac {0 . 5 \\times \\left(C R _ {x _ {i}} - P _ {i}\\right) ^ {2}}{\\beta} & \\text {i f} \\left| C R _ {x _ {i}} - P _ {i} \\right| < \\beta \\\\ \\left| C R _ {x _ {i}} - P _ {i} \\right| - 0. 5 \\times \\beta & \\text {o t h e r w i s e} \\end{array} \\right. \\tag {5}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "5839" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.091, + 0.089, + 0.287, + 0.216 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.179, + 0.22, + 0.197, + 0.231 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.29, + 0.089, + 0.485, + 0.215 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.377, + 0.22, + 0.396, + 0.231 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image", + "bbox": [ + 0.487, + 0.089, + 0.682, + 0.216 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.576, + 0.22, + 0.594, + 0.231 + ], + "angle": 0, + "content": "(c)" + }, + { + "type": "image", + "bbox": [ + 0.685, + 0.091, + 0.877, + 0.216 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.772, + 0.22, + 0.79, + 0.231 + ], + "angle": 0, + "content": "(d)" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.233, + 0.894, + 0.276 + ], + "angle": 0, + "content": "Figure 3. ERC comparison between CR-FIQA(S), CCS-FIQA(S), CR-FIQA(S) (On top) and CCS-FIQA(S) (On top). The plots show the effect of rejecting samples of lowest quality, on the verification error (FNMR at FMR1e-3). CR-FIQA(S) and CCS-FIQA(S) outperformed the on-top solutions, and CR-FIQA(S) performs generally better than CCS-FIQA(S) (curve decays faster with more rejected samples)" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.284, + 0.471, + 0.314 + ], + "angle": 0, + "content": "The final loss combining both objectives for training our CR-FIQA model is defined as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.194, + 0.318, + 0.469, + 0.334 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\mathcal {L} _ {A r c} + \\lambda \\times \\mathcal {L} _ {C R}, \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.345, + 0.47, + 0.451 + ], + "angle": 0, + "content": "where \\(\\lambda\\) is a hyper-parameter used to control the balance between the two losses. At the beginning of model training, the value range of \\(\\mathcal{L}_{CR}\\) is very small (\\(\\leq 2\\)) in comparison to \\(\\mathcal{L}_{Arc} (\\sim 45)\\). Setting \\(\\lambda\\) to a small value, the model will only focus on \\(\\mathcal{L}_{Arc}\\). Besides, setting \\(\\lambda\\) to a large value, i.e. \\(>10\\), we observed that the model did not converge. Therefore, we set \\(\\lambda\\) to 10 in all the experiments in this paper." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.457, + 0.271, + 0.474 + ], + "angle": 0, + "content": "4. Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.479, + 0.471, + 0.901 + ], + "angle": 0, + "content": "Implementation Details We demonstrate our proposed CR-FIQA under two protocols (small and large) based on the training dataset and the training model architecture. We utilize widely used architectures in the SOTA FR solutions, ResNet100 and ResNet50 [13], both modified as described in Section 3.4. For the small protocol, we utilize ResNet50 and the CASIA-WebFace [39] training data (noted as CR-FIQA(S)) and for the large protocol, we utilize ResNet100 and the MS1MV2 [5, 12] training data (noted as CR-FIQA(L)). The MS1MV2 is a refined version of the MS-Celeb-1M [12] by [5] containing 5.8M images of 85K identities. The CASIA-WebFace contains 0.5m images of 10K identities [39]. We follow the ArcFace training setting [5] to set the scale parameter \\( s \\) to 64 and the margin \\( m \\) to 0.5. We set the mini-batch size to 512. All models are trained with Stochastic Gradient Descent (SGD) optimizer with an initial learning rate of 1e-1. During the training, we use random horizontal flipping with a probability of 0.5 for data augmentation. We set the momentum to 0.9 and the weight decay to 5e-4. For CR-FIQA(S), the learning rate is divided by 10 at 20K and at 28K training iterations, following [5]. The training is stopped after 32K iterations. For CR-FIQA(L), the learning rate is divided by 10 at 100K and 160K training iterations, following [5]. The training is stopped after 180K iterations. All the images in evaluation and training datasets are aligned and cropped to \\( 112 \\times 112 \\), as described in [5]. All the training and testing images are normalized to have pixel values between -1 and 1. Both" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.284, + 0.866, + 0.3 + ], + "angle": 0, + "content": "models are trained using the loss defined in Equation 6." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.302, + 0.892, + 0.439 + ], + "angle": 0, + "content": "Evaluation Benchmarks We reported the achieved results on eight different benchmarks: Labeled Faces in the Wild (LFW) [16], AgeDB-30 [31], Celebrities in Frontal-Profile in the Wild (CFP-FP) [33], Cross-age LFW (CALFW) [41], Adience [6], Cross-Pose LFW (CPLFW) [40], Cross-Quality LFW (XQLFW) [23], IARPA Janus Benchmark-C (IJB-C) [27]. These benchmarks are chosen to provide a wide comparison to SOTA FIQA algorithms and give an insight into the CR-FIQA generalizability." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.441, + 0.893, + 0.638 + ], + "angle": 0, + "content": "Evaluation Metric We evaluate the FIQA by plotting ERCs [11]. The ERC is a widely used representation of the FIQA performance [10, 11] by demonstrating the effect of rejecting a fraction face images, of the lowest quality, on face verification performance in terms of False None Match Rate [21] (FNMR) at a specific threshold calculated at fixed False Match Rate [21] (FMR). The ERC curves for all benchmarks are plotted at two fixed FMRs, 1e-3 (as recommended for border control operations by Frontex [7]) and 1e-4 (the latter is provided in the supplementary material). We also report the Area under the Curve (AUC) of the ERC, to provide a quantitative aggregate measure of verification performance across all rejection ratios." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.641, + 0.894, + 0.792 + ], + "angle": 0, + "content": "Additionally, motivated by evaluating the FIQ as a weighting term for face embedding [32, 34], we follow the IJB-C 1:1 mixed verification benchmark [27] by weighting the frames such that all frames belonging to the same subject within a video have a combined weight equal to a single still image as described in IJB-C benchmark [27]. We do that by using the CR-FIQA quality scores as well as all SOTA methods. We report the verification performance of IJB-C as true acceptance rates (TAR) at false acceptance rates (FAR) of 1e-4, 1e-5, and 1e-6, as defined in [27]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.796, + 0.894, + 0.901 + ], + "angle": 0, + "content": "Face Recognition Models We utilize four different SOTA FR models to report the verification performance at different quality rejection rate to inspect the generalizability of FIQA over FR solutions. The FR models are ArcFace [5], ElasticFace (ElasticFace-Arc) [3], MagFace [28], and CurricularFace [17]. All models process \\(112 \\times 112\\) aligned and cropped image to produce 512-D feature em" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5840" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.258 + ], + "angle": 0, + "content": "bedding. We used the officially released pretrained ResNet-100 models trained on MS1MV2 released by the four FR solutions. Although, the presented solution in this paper does not aim, and is not presented as, a solution to extract face embeddings, but rather an FIQA solution, we opted to evaluate CR-FIQA(L) backbone as a FR model on mainstream FR benchmarks for sake of providing complete experiment evaluation and probe the possibility of simultaneously using it as both FIQA and FR model. The evaluation results of CR-FIQA(L) backbone as a FR model are provided in the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.258, + 0.473, + 0.38 + ], + "angle": 0, + "content": "Baseline We compare our CR-FIQA approach with nine quality assessment methods. Three are general IQA methods that have been proven in [8] to correlate well to face utility i.e. BRISQUE [29], RankIQA [26], and DeepIQA [2], and six are SOTA face-specific FIQA methods, namely RankIQ [4], PFE [34], SER-FIQ [36], FaceQnet (v1 [14]) [14, 15], MagFace [28], and SDD-FIQA [32], all as officially released in the respective works." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.385, + 0.242, + 0.4 + ], + "angle": 0, + "content": "5. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.403, + 0.47, + 0.433 + ], + "angle": 0, + "content": "This section provides experimental proof of the two main design choices in CR-FIQA." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.434, + 0.473, + 0.675 + ], + "angle": 0, + "content": "Does CR-FIQA benefit from the NNCCS scaling term? To answer this, we conducted additional experiments using ResNet-50 model trained on CASIA-WebFace [39] using the experimental setup described in Section 4. This model is noted as CCS-FIAQ(S). The only difference from CR-FIAQ(S) is that the CCS-FIAQ(S) is trained to learn CCS (instead of CR) by replacing the \\( CR_{x_i} \\) in Equation 5 with \\( CCS_{x_i} \\), thus neglecting the NNCCS scaling term in the equation. Figure 3 presents the ERCs along with AUC using CR-FIAQ(S) and CCS-FIAQ(S) on Adience, AgeDb-30, CALLFW and CRF-FP. The verification error, FNMR at FMR1e-3, is calculated using ArcFace FR model (described in Section 4). The ERCs and AUC values show that the reduction in the error is more evident for CR-FIAQ(S) than CCS-FIQA(S). Thus, adding the scaling term NNCCS in CR enhanced the performance of the FIQA." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.675, + 0.473, + 0.903 + ], + "angle": 0, + "content": "Dose the simultaneous learning in CR-FIQA lead to better performance in comparison to on-the-top learning? We consider the possibility of learning to estimate CR after finalizing the FR training, in comparison to the simultaneously learning in CR-FIQA. We conducted two additional experiments using pretrained ResNet-50 trained with ArcFace loss [5] on CASIA-WebFace [39]. Specifically, we add an additional single regression layer to this pretrained model. We freeze the weights of the pretrained model and train only the regression layer to learn the internal network observation of the pretrained model using only the \\(\\mathcal{L}_{\\mathcal{CR}}\\) (Equation 5). Using this setting, we present two instances, CR-FIQA(S) (On top) and CCS-FIQA(S) (On top), that learned to predict CR and CCS, respectively. Each of CR-FIQA(S) (On top) and CCS-FIQA(S) (On top) is fine" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.895, + 0.38 + ], + "angle": 0, + "content": "tuned for 32K iteration with an initial learning rate of 0.01. The learning rate is divided by 10 at 20K and 28K training iterations, similarly to CR-FIAQ(S) and CCS-FIAQ(S). The results (ERCs and AUCs) of these models are compared to CR-FIQA(S) and CCS-FIQA(S) in Figure 3. The ERCs in Figure 3 presented the evaluation on Adience, AgeDb-30, CALFW and CFP-FP using the ArcFace FR model. The ERCs and AUCs in Figure 3 show that both CR-FIQA(S) and CCS-FIQA(S) lead to stronger reductions in the error than the CR-FIQA(S) (On top) and the CCS-FIQA(S) (On top), when rejecting low-quality samples. This supports our training paradigm that simultaneously learns the internal network observation, CR, while optimizing the class centers. This can be related to the step-wise convergence towards the final CR value in the simultaneously training. For both ablation study questions, ERCs and AUC for all the remaining benchmarks and FR models (mentioned in Section 4) lead to similar conclusions and are provided in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.385, + 0.708, + 0.4 + ], + "angle": 0, + "content": "6. Result and Discussion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.403, + 0.895, + 0.629 + ], + "angle": 0, + "content": "All CR-FIQA performances reported in this paper are obtained under cross-model settings. The proposed CR-FIQA is used only to predict FIQ and not to extract feature representation of face images. None of the utilized FR models (ArcFace [5], ElasticFace [3], MagFace [28], and CurricularFace [17]) is trained with our paradigm. Instead, we trained a separate model for CR-FIQA and used the official pretrained FR models (as described in Section 4) for feature extraction. The verification performances as AUC at FMR1e-3 and FMR1e-4 are presented in Table 1. The visual verification performances as ERC curves (Figure 4) using ArcFace and ElasticFace FR models are reported at FMR1e-3. The ERC curves at FMR1e-4 and for MagFace and CurricularFace FR models at FMR1e-3 are provided in the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.63, + 0.896, + 0.902 + ], + "angle": 0, + "content": "The ERC curves (Figure 4) and the AUC values (Table 1) show that our proposed CR-FIQA(S) and CR-FIQA(L) outperformed the SOTA methods by significant margins in almost all settings. Observing the results on IJB-C, Adience, CFP-FP, CALFW, and CPLFW at FMR1e-3 and FMR1e-4 (Figure 4 and Table 1), our proposed CR-FIQA outperformed all SOTA methods on all the considered FR models. On the AgeDB-30 benchmark, our proposed CR-FIQA ranked first in five out of eight settings and second in the other three settings (Table 1). On the LFW benchmark, our proposed CR-FIQA ranked behind the MagFace and the RankIQ. This is the only case that our models did not outperform all SOTA methods. However, it can be noticed from the ERC curves in Figure 4 that none of the SOTA methods were able to achieve stable behavior (smoothly decaying curve) on LFW. The main reason for such unstable ERC behavior on LFW is that the FR performance on LFW is nearly saturated (all models achieved above \\(99.80\\%\\) accu" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "5841" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.089, + 0.892, + 0.146 + ], + "angle": 0, + "content": "Table 1. The AUCs of ERC achieved by our CR-FIQA and the SOTA methods under different experimental settings. CR-FIQA achieved the best performance (lowest AUC) in almost all settings. On XQLFW, the SER-FIQ (marked with *) is used for the sample selection of the XQLFW benchmark. The best result for each experimental setting is in bold and the second-ranked one is in italic. The notions of \\(1e - 3\\) and \\(1e - 4\\) indicate the value of the fixed FMR at which the ERC curve (FNMR vs. reject) was calculated at" + }, + { + "type": "table", + "bbox": [ + 0.11, + 0.147, + 0.864, + 0.569 + ], + "angle": 0, + "content": "
FRMethodAdience [6]AgeDB-30 [31]CFP-FP [33]LFW [16]CALFW [41]CPLFW [40]XQLFW [23]IJB-C [27]
1e-31e-41e-31e-41e-31e-41e-31e-41e-31e-41e-31e-41e-31e-41e-31e-4
ArcFace [5]QABRISQUE [29]0.05650.12850.04000.05850.03430.04330.00430.00490.07550.08130.25580.30370.66800.71220.03810.0656
RankIQA [26]0.04000.09330.03720.05230.03010.03840.00390.00450.08460.09150.24370.29690.65840.70390.03850.0640
DeepIQA [2]0.05680.13720.04030.05230.02380.02920.00490.00560.07930.08500.23090.28560.59580.64580.03830.0640
FIQARankIQ [4]0.03530.08730.03220.04200.01520.02600.00180.00240.06080.06720.06330.08480.27890.33320.02270.0342
PFE [34]0.02120.04280.01720.02260.00920.01290.00230.00280.06470.06810.04500.06380.23020.27100.01760.0248
SER-FIQ [36]0.02230.04340.01670.02230.00650.01030.00230.00280.05950.06270.03890.05840.1812*0.2295*0.01610.0241
FaceQnet [14, 15]0.03460.07340.01970.02450.02400.02730.00220.00270.07740.08220.15040.17510.58290.61360.02700.0376
MagFace [28]0.02070.04250.01560.01980.00730.01050.00160.00210.05680.06020.04920.06420.40220.46360.01710.0254
SDD-FIQA [32]0.02480.05620.01860.02060.01220.01930.00210.00270.06410.06980.05170.06700.30900.35610.01860.0270
CR-FIQA(S)(Our)0.02410.05170.01440.01870.00900.01450.00200.00250.05210.05540.03910.05670.23770.27400.01710.0250
CR-FIQA(L)(Our)0.02040.03530.01590.01890.00500.00820.00230.00290.06160.06320.03600.05150.20840.24410.01380.0207
ElasticFace [3]QABRISQUE [29]0.06440.11840.03750.04030.02810.03720.00340.00470.07260.07470.26410.46880.63430.69640.03570.0621
RankIQA [26]0.04330.08620.03740.04360.02690.03180.00330.00450.08100.08350.23250.43060.61890.68560.03660.0599
DeepIQA [2]0.06450.12030.03840.04110.01910.02560.00430.00560.07560.07720.24010.45410.54000.58320.03790.0590
FIQARankIQ [4]0.04000.07770.03090.03370.01490.01800.00130.00200.05980.06140.05810.07270.24680.27760.02260.0334
PFE [34]0.02220.03810.01630.01720.00880.01130.00180.00250.06280.06430.04190.08950.21120.24360.01710.0247
SER-FIQ [36]0.02400.04170.01630.01790.00610.00850.00210.00280.05740.05900.03870.05130.1576*0.1868*0.01560.0235
FaceQnet [14, 15]0.03690.06670.01940.02070.02270.02470.00210.00260.07630.07770.14200.28800.55490.58440.02630.0370
MagFace [28]0.02250.03850.01500.01580.00690.00950.00140.00210.05530.05630.04740.05970.39730.42820.01660.0243
SDD-FIQA [32]0.02770.05120.01870.02000.00980.01180.00190.00270.06240.06380.04930.06340.30520.35620.01830.0266
CR-FIQA(S)(Our)0.02570.04650.01460.01600.00700.00960.00150.00220.05090.05220.03830.05020.20930.28350.01670.0244
CR-FIQA(L)(Our)0.02140.03570.01490.01590.00450.00650.00180.00250.05940.06080.03500.04620.17980.20600.01350.0203
MagFace [28]QABRISQUE [29]0.05940.13080.04420.07990.04220.05890.00430.00580.07580.07880.46490.68090.69110.72290.04620.0787
RankIQA [26]0.04070.08890.03700.06810.03690.05430.00410.00560.08290.08570.32510.64750.67060.70460.04610.0750
DeepIQA [2]0.05710.13020.04170.07210.03220.05450.00480.00590.07870.08090.36720.66320.61620.65190.04740.0765
FIQARankIQ [4]0.03590.08370.03610.05310.02130.03320.00790.00270.06020.06290.06590.16420.30760.34750.02690.0383
PFE [34]0.02150.04230.01920.03170.01070.01380.00230.00290.06400.06520.04490.14350.26150.29260.02000.0283
SER-FIQ [36]0.02330.04510.01850.02930.00800.01390.00250.00330.05900.06070.03970.52180.60160.62100.03050.0422
FaceQnet [14, 15]0.03650.07200.02170.03140.02710.03510.00220.00270.07630.07730.29880.52180.64750.71880.19550.1979
MagFace [28]0.02120.04170.01590.02470.00850.01290.00170.00220.05620.05780.05060.88870.44780.49000.01950.0279
SDD-FIQA [32]0.02530.05620.02160.03050.01460.02010.00210.06430.06570.05250.11880.34040.39280.39280.02150.0307
CR-FIQA(S)(Our)0.02440.05070.01650.02340.01020.01210.00200.00280.05160.05280.04090.08400.26700.33360.01980.0284
CR-FIQA(L)(Our)0.02110.03720.01740.02350.00620.00800.00230.00280.06140.06280.03740.06790.23690.28390.01620.0236
CurricularFace [17]QABRISQUE [29]0.05020.10950.04330.04910.03230.03570.00410.00470.07550.07840.27090.50570.61460.63360.03620.0589
RankIQA [26]0.03590.07520.03940.05100.02980.03560.00390.00450.88060.86560.23460.46540.56440.62120.13610.1558
DeepIQA [2]0.04920.10700.04070.04760.02270.02780.00500.00560.7640.7860.24880.49610.51650.55260.13760.1558
FIQARankIQ [4]0.03140.07150.03650.04170.01860.02490.00180.00240.59590.66451515*1515*1515*1515*1515*1515*
PFE [34]0.01980.03650.01970.02270.01030.01340.00240.00280.63630.65751515*1515*1515*1515*1515*1515*
SER-FIQ [36]0.02110.03810.01670.01930.00740.01110.00250.03561518*1519*1518*1519*1519*1519*1519*1519*
FaceQnet [14, 15]0.03260.06260.02210.02670.02260.02740.00220.02771518*1519*1518*1519*1519*1519*1519*1519*
MagFace [28]0.02000.03640.01670.01950.00780.01110.00160.00211516*1517*1518*1519*1519*1519*1519*1519*
SDD-FIQA [32]0.02300.04620.02190.02540.01380.01850.00210.06271517*1518*1518*1519*1519*1519*1519*1519*
CR-FIQA(S)(Our)0.02270.04460.01560.01980.00970.01480.00200.02551513*1514*1514*1515*1515*1515*1515*1515*
CR-FIQA(L)(Our)0.01980.03360.01620.02000.00540.08800.00230.02991516*1517*1518*1519*1519*1519*1519*1519*
" + }, + { + "type": "table_caption", + "bbox": [ + 0.086, + 0.577, + 0.884, + 0.591 + ], + "angle": 0, + "content": "Table 2. Verification performance on the IJB-C (1:1 mixed verification) [27]. CR-FIQA outperformed SOTA methods under all settings" + }, + { + "type": "table", + "bbox": [ + 0.106, + 0.592, + 0.859, + 0.735 + ], + "angle": 0, + "content": "
Quality Estimation1:1 mixed Verification: TAR (%) at
ArcFace [5]ElasticFace [3]MagFace [28]CurricularFace [17]
FAR=1e-6FAR=1e-5FAR=1e-4FAR=1e-6FAR=1e-5FAR=1e-4FAR=1e-6FAR=1e-5FAR=1e-4FAR=1e-6FAR=1e-5FAR=1e-4
-89.8594.4796.2889.1594.5496.4985.6793.0896.6590.4694.8996.58
IAQBRISQUE [29]86.6593.6295.9885.6893.5195.6581.1190.6494.8288.1693.9896.29
RankIQA [26]86.3793.6195.8386.7193.4696.0080.7890.7594.8688.1694.1196.22
DeepIQA [2]81.9791.6494.6778.9391.5994.8173.5386.3492.9082.6592.0495.00
FIQARankIQ [4]88.7894.4296.2088.8894.6496.4585.6392.6695.7090.0094.9396.53
PFE [34]89.5094.5196.3189.1094.6796.5184.9392.4495.6090.3695.0496.54
SER-FIQ [36]89.7494.6596.3290.0594.7996.5786.0293.3595.8090.6695.1196.58
FaceQNet [14, 15]87.8794.0496.1286.2694.0996.2582.9190.5695.0389.6194.6596.36
MagFace [28]89.4994.4196.2289.3794.6996.4685.7592.7195.5490.3495.0296.50
SDD-FIQA [32]89.3994.6196.3488.0794.8296.4984.6992.8395.7389.9195.1296.63
CR-FIQA(S)(Our)89.5994.7896.3590.3094.9796.6386.4593.4895.9590.8295.1396.64
CR-FIQA(L)(Our)90.1694.7596.3690.0094.9296.5887.1293.6795.9090.7995.1296.58
" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.75, + 0.468, + 0.779 + ], + "angle": 0, + "content": "racy [3,5,17,28]), leaving very few samples causing errors and thus lowering the significance of the measured FNMR." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.78, + 0.47, + 0.901 + ], + "angle": 0, + "content": "The XQLFW benchmark is derived from LFW to contain pairs with a maximum difference in quality. The XQLFW images are chosen based on BRISQUE [29] and SER-FIQ [23] quality scores to be either extremely high or low. The use of SER-FIQ [36] in this selection might give a biased edge for SER-FIQ on this benchmark. On XQLFW, our CR-FIQA achieved very close performance to the selection method (SER-FIQ) and is far ahead of all other SOTA" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.892, + 0.81 + ], + "angle": 0, + "content": "methods. Lastly, our proposed CR-FIQA(S) achieved very comparable performance to our CR-FIQA(L), pointing out the robustness of our approach, regardless of the training database and architecture complexity." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.811, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Table 2 presents the verification performance on the IJB-C 1:1 mixed verification benchmark [27] using quality scores as an embedding weighting term (as defined in [27]) under different experimental settings. For each of the FR models, we report in the first row the evaluation result of the corresponding FR model as defined in the protocol [27] and" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5842" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.108, + 0.089, + 0.295, + 0.196 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.297, + 0.089, + 0.482, + 0.197 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.489, + 0.089, + 0.675, + 0.197 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.676, + 0.09, + 0.862, + 0.197 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.107, + 0.199, + 0.295, + 0.306 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.296, + 0.199, + 0.482, + 0.306 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.489, + 0.199, + 0.675, + 0.306 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.676, + 0.199, + 0.862, + 0.306 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.107, + 0.307, + 0.296, + 0.415 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.297, + 0.307, + 0.482, + 0.415 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.489, + 0.307, + 0.675, + 0.414 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.676, + 0.307, + 0.862, + 0.414 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.107, + 0.416, + 0.296, + 0.524 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.297, + 0.416, + 0.482, + 0.524 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.489, + 0.416, + 0.675, + 0.524 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.676, + 0.416, + 0.862, + 0.524 + ], + "angle": 0, + "content": null + }, + { + "type": "image_footnote", + "bbox": [ + 0.179, + 0.527, + 0.787, + 0.536 + ], + "angle": 0, + "content": "BRIQUE RankIQ DeepIQ \\(\\rightarrow\\) RankIQ PFE SER-FIQ FaceQnet MagFace SDD-FIQA CR-FIQA(S) Our) CR-FIQA(L) Our" + }, + { + "type": "image_caption", + "bbox": [ + 0.075, + 0.54, + 0.892, + 0.61 + ], + "angle": 0, + "content": "Figure 4. ERC (FNMR at FMR1e-3 vs reject) curves for all evaluated benchmarks using ArcFace and ElasticFace FR models corresponding to Table 1 results. The visual evaluation, ERC curves, using MagFace and CurricularFace FR models are provided in the supplementary material. The proposed CR-FIQA(L) and CR-FIQA(S) are marked with solid blue and red lines, respectively. CR-FIQA leads to lower verification error, when rejecting a fraction of images, of the lowest quality, in comparison to SOTA methods (faster decaying curve) under most experimental settings." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.614, + 0.471, + 0.751 + ], + "angle": 0, + "content": "the corresponding released evaluation scripts [3, 5, 17, 28], i.e. without considering the FIQ. Our proposed CR-FIQA significantly leads to higher verification performance than all evaluated SOTA methods, when the quality score is used as an embedding weighting term (Table 2). This achievement is observable under all experimental settings (Table 2). Another outcome of this evaluation is that the integration of CR-FIQA leads to SOTA verification performance on one of the most challenging FR benchmarks, IJB-C [27]." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.757, + 0.196, + 0.772 + ], + "angle": 0, + "content": "7. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.78, + 0.471, + 0.903 + ], + "angle": 0, + "content": "In this work, we propose the CR-FIQA approach that probes the relative classifiability of training samples of the FR model and utilize this observation to learn to predict the utility of any given sample in achieving an accurate FR performance. We experimentally prove the theorized relationship between the sample relative classifiability and FIQ and build on that towards our CR-FIQA. The CR-FIQA training paradigm simultaneously learns to optimize the class center" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.614, + 0.892, + 0.78 + ], + "angle": 0, + "content": "while learning to predict sample relative classifiability. The presented ablation studies and the extensive experimental results prove the effectiveness of the proposed CR-FIQA approach, and its design choices, as an FIQ method. The reported results demonstrated that our proposed CR-FIQA outperformed SOTA methods repeatedly across multiple FR models and on multiple benchmarks, including ones with a large age gap (AgeDb-30, Adience, CALFW), large quality difference (XQLFW), large pose variation (CPLFW, CFPFP), and extremely large-scale and challenging FR benchmarks (IJB-C)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.78, + 0.893, + 0.902 + ], + "angle": 0, + "content": "Acknowledgment This research work has been funded by the German Federal Ministry of Education and Research and the Hessian Ministry of Higher Education, Research, Science and the Arts within their joint support of the National Research Center for Applied Cybersecurity ATHENE. This work has been partially funded by the German Federal Ministry of Education and Research (BMBF) through the Software Campus Project." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5843" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.156 + ], + "angle": 0, + "content": "[1] Lacey Best-Rowden and Anil K. Jain. Learning face image quality from human assessments. IEEE Trans. Inf. Forensics Secur., 13(12):3064–3077, 2018. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.158, + 0.472, + 0.226 + ], + "angle": 0, + "content": "[2] Sebastian Bosse, Dominique Maniry, Klaus-Robert Müller, Thomas Wiegand, and Wojciech Samek. Deep neural networks for no-reference and full-reference image quality assessment. IEEE Trans. Image Process., 27(1):206-219, 2018. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.228, + 0.47, + 0.31 + ], + "angle": 0, + "content": "[3] Fadi Boutros, Naser Damer, Florian Kirchbuchner, and Arjan Kuijper. Elasticface: Elastic margin loss for deep face recognition. In IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, CVPR Workshops 2022, New Orleans, LA, USA, June 19-20, 2022, pages 1577-1586. IEEE, 2022. 3, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.312, + 0.468, + 0.353 + ], + "angle": 0, + "content": "[4] Jiansheng Chen, Yu Deng, Gaocheng Bai, and Guangda Su. Face image quality assessment based on learning to rank. IEEE Signal Process. Lett., 22(1):90-94, 2015. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.354, + 0.468, + 0.436 + ], + "angle": 0, + "content": "[5] Jiankang Deng, Jia Guo, Niannan Xue, and Stefanos Zafeiriou. Arcface: Additive angular margin loss for deep face recognition. In IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2019, Long Beach, CA, USA, June 16-20, 2019, pages 4690-4699. Computer Vision Foundation / IEEE, 2019. 3, 4, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.438, + 0.468, + 0.479 + ], + "angle": 0, + "content": "[6] Eran Eidinger, Roee Enbar, and Tal Hassner. Age and gender estimation of unfiltered faces. IEEE Trans. Inf. Forensics Secur., 9(12):2170-2179, 2014. 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.481, + 0.468, + 0.508 + ], + "angle": 0, + "content": "[7] Frontex. Best practice technical guidelines for automated border control (abc) systems, 2015. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.509, + 0.468, + 0.564 + ], + "angle": 0, + "content": "[8] Biying Fu, Cong Chen, Olaf Henniger, and Naser Damer. A deep insight into measuring face image utility with general and face-specific image quality metrics. pages 1121-1130, 2022. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.565, + 0.468, + 0.62 + ], + "angle": 0, + "content": "[9] Ross B. Girshick. Fast R-CNN. In 2015 IEEE International Conference on Computer Vision, ICCV 2015, Santiago, Chile, December 7-13, 2015, pages 1440-1448. IEEE Computer Society, 2015. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.622, + 0.468, + 0.677 + ], + "angle": 0, + "content": "[10] P. Grother, M. Ngan A. Hom, and K. Hanaoka. Ongoing face recognition vendor test (frvt) part 5: Face image quality assessment (4th draft). In National Institute of Standards and Technology. Tech. Rep., Sep. 2021. 1, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.679, + 0.468, + 0.718 + ], + "angle": 0, + "content": "[11] P. Grother and E. Tabassi. Performance of biometric quality measures. IEEE Trans. on Pattern Analysis and Machine Intelligence, 29(4):531-543, Apr. 2007. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.72, + 0.468, + 0.831 + ], + "angle": 0, + "content": "[12] Yandong Guo, Lei Zhang, Yuxiao Hu, Xiaodong He, and Jianfeng Gao. Ms-celeb-Im: A dataset and benchmark for large-scale face recognition. In Bastian Leibe, Jiri Matas, Nicu Sebe, and Max Welling, editors, Computer Vision - ECCV 2016 - 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part III, volume 9907 of Lecture Notes in Computer Science, pages 87-102. Springer, 2016. 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.832, + 0.468, + 0.901 + ], + "angle": 0, + "content": "[13] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In 2016 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2016, Las Vegas, NV, USA, June 27-30, 2016, pages 770-778. IEEE Computer Society, 2016. 4, 5" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.147 + ], + "angle": 0, + "content": "[14] Javier Hernandez-Ortega, Javier Galbally, Julian Fierrez, and Laurent Beslay. Biometric quality: Review and application to face recognition with faceqnet. CoRR, abs/2006.03298, 2020. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.149, + 0.892, + 0.232 + ], + "angle": 0, + "content": "[15] Javier Hernandez-Ortega, Javier Galbally, Julian Fierrez, Rudolf Haraksim, and Laurent Beslay. Faceqnet: Quality assessment for face recognition based on deep learning. In 2019 International Conference on Biometrics, ICB 2019, Crete, Greece, June 4-7, 2019, pages 1-8. IEEE, 2019. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.233, + 0.892, + 0.301 + ], + "angle": 0, + "content": "[16] Gary B. Huang, Manu Ramesh, Tamara Berg, and Erik Learned-Miller. Labeled faces in the wild: A database for studying face recognition in unconstrained environments. Technical Report 07-49, University of Massachusetts, Amherst, October 2007. 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.302, + 0.892, + 0.398 + ], + "angle": 0, + "content": "[17] Yuge Huang, Yuhan Wang, Ying Tai, Xiaoming Liu, Pengcheng Shen, Shaoxin Li, Jilin Li, and Feiyue Huang. Curricularface: Adaptive curriculum learning loss for deep face recognition. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2020, Seattle, WA, USA, June 13-19, 2020, pages 5900-5909. Computer Vision Foundation / IEEE, 2020. 3, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.4, + 0.892, + 0.439 + ], + "angle": 0, + "content": "[18] ISO/IEC JTC1 SC17 WG3. Portrait Quality - Reference Facial Images for MRTD. International Civil Aviation Organization, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.441, + 0.892, + 0.495 + ], + "angle": 0, + "content": "[19] ISO/IEC JTC1 SC37 Biometrics. ISO/IEC 29794-1:2016 Information technology - Biometric sample quality - Part 1: Framework. International Organization for Standardization, 2016. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.497, + 0.892, + 0.538 + ], + "angle": 0, + "content": "[20] ISO/IEC JTC1 SC37 Biometrics. ISO/IEC 2382-37:2017 Information technology - Vocabulary - Part 37: Biometrics. International Organization for Standardization, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.539, + 0.892, + 0.594 + ], + "angle": 0, + "content": "[21] ISO/IEC JTC1 SC37 Biometrics. ISO/IEC 19795-1:2021 Information technology — Biometric performance testing and reporting — Part 1: Principles and framework. International Organization for Standardization, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.595, + 0.892, + 0.664 + ], + "angle": 0, + "content": "[22] Thorsten Joachims. Optimizing search engines using click-through data. In Proceedings of the Eighth ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, July 23-26, 2002, Edmonton, Alberta, Canada, pages 133-142. ACM, 2002. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.665, + 0.892, + 0.748 + ], + "angle": 0, + "content": "[23] Martin Knoche, Stefan Hörmann, and Gerhard Rigoll. Cross-quality LFW: A database for analyzing cross-resolution image face recognition in unconstrained environments. In 16th IEEE International Conference on Automatic Face and Gesture Recognition, FG 2021, Jodhpur, India, December 15-18, 2021, pages 1-5. IEEE, 2021. 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.748, + 0.892, + 0.831 + ], + "angle": 0, + "content": "[24] Weiyang Liu, Yandong Wen, Zhiding Yu, Ming Li, Bhiksha Raj, and Le Song. Spherical face: Deep hypersphere embedding for face recognition. In 2017 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2017, Honolulu, HI, USA, July 21-26, 2017, pages 6738-6746. IEEE Computer Society, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.832, + 0.892, + 0.899 + ], + "angle": 0, + "content": "[25] Weiyang Liu, Yandong Wen, Zhiding Yu, Ming Li, Bhiksha Raj, and Le Song. SphereFace: Deep hypersphere embedding for face recognition. In Proc. of the IEEE Conf. on Computer Vision and Pattern Recognition, pages 212-220, 2017. 3" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5844" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.174 + ], + "angle": 0, + "content": "[26] Xialei Liu, Joost van de Weijer, and Andrew D. Bagdanov. Rankiaq: Learning from rankings for no-reference image quality assessment. In IEEE International Conference on Computer Vision, ICCV 2017, Venice, Italy, October 22-29, 2017, pages 1040-1049. IEEE Computer Society, 2017. 1, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.177, + 0.471, + 0.272 + ], + "angle": 0, + "content": "[27] Brianna Maze, Jocelyn C. Adams, James A. Duncan, Nathan D. Kalka, Tim Miller, Charles Otto, Anil K. Jain, W. Tyler Niggel, Janet Anderson, Jordan Cheney, and Patrick Grother. IARPA janus benchmark - C: face dataset and protocol. In 2018 International Conference on Biometrics, ICB 2018, Gold Coast, Australia, February 20-23, 2018, pages 158-165. IEEE, 2018. 5, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.274, + 0.471, + 0.356 + ], + "angle": 0, + "content": "[28] Qiang Meng, Shichao Zhao, Zhida Huang, and Feng Zhou. Magface: A universal representation for face recognition and quality assessment. In IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2021, virtual, June 19-25, 2021, pages 14225-14234. Computer Vision Foundation / IEEE, 2021. 1, 2, 3, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.358, + 0.47, + 0.412 + ], + "angle": 0, + "content": "[29] Anish Mittal, Anush Krishna Moorthy, and Alan Conrad Bovik. No-reference image quality assessment in the spatial domain. IEEE Trans. Image Process., 21(12):4695-4708, 2012. 1, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.414, + 0.47, + 0.455 + ], + "angle": 0, + "content": "[30] Anish Mittal, Rajiv Soundararajan, and Alan C. Bovik. Making a \"completely blind\" image quality analyzer. IEEE Signal Process. Lett., 20(3):209-212, 2013. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.456, + 0.47, + 0.538 + ], + "angle": 0, + "content": "[31] Stylianos Moschoglou, Athanasios Papaioannou, Christos Sagonas, Jiankang Deng, Irene Kotsia, and Stefanos Zafeiriou. Agedb: The first manually collected, in-the-wild age database. In 2017 IEEE CVPRW, CVPR Workshops 2017, Honolulu, HI, USA, July 21-26, 2017, pages 1997-2005. IEEE Computer Society, 2017. 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.539, + 0.47, + 0.636 + ], + "angle": 0, + "content": "[32] Fu-Zhao Ou, Xingyu Chen, Ruixin Zhang, Yuge Huang, Shaoxin Li, Jilin Li, Yong Li, Liujuan Cao, and Yuan-Gen Wang. SDD-FIQA: unsupervised face image quality assessment with similarity distribution distance. In IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2021, virtual, June 19-25, 2021, pages 7670-7679. Computer Vision Foundation / IEEE, 2021. 1, 2, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.637, + 0.47, + 0.72 + ], + "angle": 0, + "content": "[33] Soumyadip Sengupta, Jun-Cheng Chen, Carlos Domingo Castillo, Vishal M. Patel, Rama Chellappa, and David W. Jacobs. Frontal to profile face verification in the wild. In 2016 IEEE Winter Conference on Applications of Computer Vision, WACV 2016, Lake Placid, NY, USA, March 7-10, 2016, pages 1-9. IEEE Computer Society, 2016. 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.721, + 0.47, + 0.788 + ], + "angle": 0, + "content": "[34] Yichun Shi and Anil K. Jain. Probabilistic face embeddings. In 2019 IEEE/CVF International Conference on Computer Vision, ICCV 2019, Seoul, Korea (South), October 27 - November 2, 2019, pages 6901-6910. IEEE, 2019. 1, 2, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.79, + 0.47, + 0.873 + ], + "angle": 0, + "content": "[35] Philipp Terhörst, Jan Niklas Kolf, Naser Damer, Florian Kirchbuchner, and Arjan Kuijper. Face quality estimation and its correlation to demographic and non-demographic bias in face recognition. In 2020 IEEE International Joint Conference on Biometrics, IJCB 2020, Houston, TX, USA, September 28 - October 1, 2020, pages 1-11. IEEE, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.874, + 0.47, + 0.901 + ], + "angle": 0, + "content": "[36] Philipp Terhörst, Jan Niklas Kolf, Naser Damer, Florian Kirchbuchner, and Arjan Kuijper. SER-FIQ: unsupervised" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.471, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.162 + ], + "angle": 0, + "content": "estimation of face image quality based on stochastic embedding robustness. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2020, Seattle, WA, USA, June 13-19, 2020, pages 5650-5659. Computer Vision Foundation / IEEE, 2020. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.163, + 0.892, + 0.247 + ], + "angle": 0, + "content": "[37] Hao Wang, Yitong Wang, Zheng Zhou, Xing Ji, Dihong Gong, Jingchao Zhou, Zhifeng Li, and Wei Liu. Cosface: Large margin cosine loss for deep face recognition. In 2018 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2018, Salt Lake City, UT, USA, June 18-22, 2018, pages 5265-5274. IEEE Computer Society, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.248, + 0.892, + 0.316 + ], + "angle": 0, + "content": "[38] Weidi Xie, Jeffrey Byrne, and Andrew Zisserman. Inducing predictive uncertainty estimation for face verification. In 31st British Machine Vision Conference 2020, BMVC 2020, Virtual Event, UK, September 7-10, 2020. BMVA Press, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.318, + 0.892, + 0.358 + ], + "angle": 0, + "content": "[39] Dong Yi, Zhen Lei, Shengcai Liao, and Stan Z. Li. Learning face representation from scratch. CoRR, abs/1411.7923, 2014. 2, 4, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.361, + 0.892, + 0.416 + ], + "angle": 0, + "content": "[40] T. Zheng and W. Deng. Cross-posed lfw: A database for studying cross-posed face recognition in unconstrained environments. Technical Report 18-01, Beijing University of Posts and Telecommunications, February 2018. 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.418, + 0.892, + 0.472 + ], + "angle": 0, + "content": "[41] Tianyue Zheng, Weihong Deng, and Jiani Hu. Cross-age LFW: A database for studying cross-age face recognition in unconstrained environments. CoRR, abs/1708.08197, 2017. 5, 7" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.472 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "5845" + } + ] +] \ No newline at end of file diff --git a/2023/CR-FIQA_ Face Image Quality Assessment by Learning Sample Relative Classifiability/2ae421c8-bef3-40d8-a9e9-022734d8a26e_origin.pdf b/2023/CR-FIQA_ Face Image Quality Assessment by Learning Sample Relative Classifiability/2ae421c8-bef3-40d8-a9e9-022734d8a26e_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..16e9e87b6709ba220b29043d47910bd917c47ded --- /dev/null +++ b/2023/CR-FIQA_ Face Image Quality Assessment by Learning Sample Relative Classifiability/2ae421c8-bef3-40d8-a9e9-022734d8a26e_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8f91bb3fbc3850ae7e19896f595f5ecf2aef93a54aeb79a36e95a971ae0db834 +size 1648861 diff --git a/2023/CR-FIQA_ Face Image Quality Assessment by Learning Sample Relative Classifiability/full.md b/2023/CR-FIQA_ Face Image Quality Assessment by Learning Sample Relative Classifiability/full.md new file mode 100644 index 0000000000000000000000000000000000000000..bb3cbc00e10d29047ff33c6bcacb6ae87ef80e43 --- /dev/null +++ b/2023/CR-FIQA_ Face Image Quality Assessment by Learning Sample Relative Classifiability/full.md @@ -0,0 +1,263 @@ +# CR-FIQA: Face Image Quality Assessment by Learning Sample Relative Classifiability + +Fadi Boutros1, Meiling Fang1,2, Marcel Klemt1, Biying Fu1, Naser Damer1,2 +1Fraunhofer Institute for Computer Graphics Research IGD, Darmstadt, Germany +2Department of Computer Science, TU Darmstadt, Darmstadt, Germany +Email: fadi.boutros@igd.fraunhofer.de + +# Abstract + +Face image quality assessment (FIQA) estimates the utility of the captured image in achieving reliable and accurate recognition performance. This work proposes a novel FIQA method, CR-FIQA, that estimates the face image quality of a sample by learning to predict its relative classifiability. This classifiability is measured based on the allocation of the training sample feature representation in angular space with respect to its class center and the nearest negative class center. We experimentally illustrate the correlation between the face image quality and the sample relative classifiability. As such property is only observable for the training dataset, we propose to learn this property by probing internal network observations during the training process and utilizing it to predict the quality of unseen samples. Through extensive evaluation experiments on eight benchmarks and four face recognition models, we demonstrate the superiority of our proposed CR-FIQA over state-of-the-art (SOTA) FIQA algorithms. + +# 1. Introduction + +Face image utility indicates the utility (value) of an image to face recognition (FR) algorithms [1, 19]. This utility is measured with a scalar, namely the face image quality (FIQ) score, following the definition in ISO/IEC 2382-37 [20] and the FR Vendor Test (FRVT) for FIQA [10]. + +As FIQA measures the face utility to FR algorithm, it does not necessary reflects, and does not aim at measuring, the perceived image quality, e.g. a profile face image can be of high perceived quality but of low utility to FR algorithm [35]. Assessing this perceived image quality has been addressed in the literature by general image quality assessment (IQA) methods [26, 29, 30] and is different than assessing the utility of an the image for FR. This is reflected by FIQA methods [28, 32, 36] significantly outperforming IQA methods [26, 29, 30] in measuring the utility [19] of face images in FR, as demonstrated in [8, 28, 36]. + +SOTA FIQA methods focused either on creating concepts to label the training data with FIQ scores and then learn a regression problem [14, 15, 32], or on developing a link between face embedding properties under certain scenarios and the FIQ [28, 34, 36]. Generally, the second approach led to better FIQA performances with most works mentioning the error-prone labeling of the ground truth quality in the first research direction as a possible reason [28, 36]. However, in the second category, transferring the information in network embeddings into an FIQ score is not a learnable process, but rather a form of statistical analysis, which might not be optimal. + +This paper proposes a novel learning paradigm to assess FIQ, namely the CR-FIQA. Our concept is based on learning to predict the classifiability of FR training samples by probing internal network observations that point to the relative proximity of these samples to their class centers and negative class centers. This regression is learned simultaneously with a conventional FR training process that minimizes the distance between the training samples and their class centers. Linking the properties that cause high/low classifiability of a training sample to the properties leading to high/low FIQ, we can use our CR-FIQA to predict the FIQ of any given sample. We empirically prove the theorized link between classifiability (Section 3.3) and FIQ and conduct thorough ablation studies on key aspects of our CR-FIQA design (Section 5). The proposed CR-FIQA is evaluated on eight benchmarks along with SOTA FIQAs. The reported results on four FR models demonstrate the superiority of our proposed CR-FIQA over SOTA methods and the stability of its performance across different FR models. An overview of the proposed CR-FIQA is presented in Figure 1 and will be clarified in detail in this paper. + +# 2. Related work + +The recent SOTA FIQA approaches can be roughly grouped into two main categories. The first are approaches that learn a straight forward regressions problem to assess a FIQ score [1,14,15,32,38]. The second category uses prop + +![](images/113af72237b96da2e5c4dbc25fd3f98375e9d2f01344258c80c376e07ffa2f6e.jpg) +Figure 1. An overview of our CR-FIQA training paradigm. We propose to simultaneously learn to optimize the class center $(\mathcal{L}_{Arc})$ , while learning to predict an internal network observation i.e. the allocation of the feature representation of sample $x$ in feature space, with respect to its class center $w_{1}$ and nearest negative class center $w_{2}$ $(\mathcal{L}_{\mathcal{CR}})$ . The figure in the red rectangle illustrates the angle between two samples $x_{1}$ and $x_{2}$ (belong to identity 1) and their class center $w_{1}$ . The plot on the right of the figure shows the distribution of the cosine similarity between training samples and their class centers (CCS) and nearest negative class centers (NNCCS) obtained from ResNet-50 trained on CASIA-WebFace [39]. The example images on the top-right of this plot are of high CCS values and the ones on the top-left are of low CCS values (notice the correspondence to perceived quality). These samples are selected from CASIA-WebFace [39]. During the testing mode, the classification layer is removed and the output of the regression layer is used to predict the FIQA of testing samples. + +erties of the FR model responses to face samples to estimate the sample quality without explicitly learning a typical supervised regression that requires quality labels [28, 34, 36]. In the first category, the innovation focused on creating the FIQ labels for training. These quality labels included human-labeled quality labels [1], the FR genuine comparison score between a sample and an ICAO [18] compliant sample [14, 15], the FR comparison score involving the labeled sample (assumed to have the lower quality in the comparison pair) [38], and the Wasserstein distance between a randomly selected genuine and imposter FR comparisons with the labeled sample [32]. These solutions generally trained a regression network to predict the quality label, using both, trained-from-scratch networks in some cases [38], and pre-trained FR networks in other cases [14, 15, 32]. A slightly different approach, however also based on learning from labels, focuses on learning to predict the sample quality as a rank [22] based on FR performance-based training rank labels of a set of databases [4]. In the second category, the innovation was rather focused on linking face embedding properties under certain scenarios to the FIQ, without the explicit need for quality-labeled data. In [36], the assessed sample is passed through an FR network multiple times, each with a different random dropout pattern. The robustness of the resulting embeddings, represented by the sigmoid of the negative mean of the Euclidean distances between the embeddings, is considered the FIQ score. In [28], the FIQ score is calculated as the magnitude of the sample embedding. This is based on training the FR model using a loss that adapts the penalty margin loss based on this magnitude, and thus links the closeness of a sample to its class + +center to the unnormalized embedding magnitude. While in [34], the solution produces both, an FR embedding and a gaussian variance (uncertainty) vector, from a face sample. The inverse of harmonic mean of the uncertainty vector is considered as the FIQ score. Our CR-FIQA learns a regression problem to estimate the FIQ score, however, unlike previous works, without relying on preset labels, but rather learn a dynamic internal network observations (during training) that point out sample classifiability. + +# 3. Approach + +This section presents our proposed Certainty Ratio Face Image Quality Assessment (CR-FIQA) approach, which inspects internal network observations to learn to predict the sample relative classifiability. This classifiability prediction is then used to estimate the FIQ. An overview of the proposed CR-FIQA approach is presented in Figure 1. During the training phase of an FR model, the model can conveniently push the high-quality samples close to their class center and relatively far from other class centers. Conversely, the FR is not able to push, to the same degree, low-quality samples to their class center, and thus they will remain relatively farther from their class center than the high-quality ones. Based on this assumption, we theorize our approach by stating that the properties that cause a face sample to lay relatively closer to its class center during training are the ones that make it a high-quality sample, and vice versa. Therefore, learning to predict such properties in any given sample would lead to learning to assess this sample quality. To learn to perform such assessment, our training paradigm targets learning internal network observations that evolve during the FR training phase, where these observa + +tions act as a training objective. The predictions of such training paradigm can be simply stated as answering a question: if a given sample was hypothetically part of the FR model training (which it is not), how relatively close would it be to its class center? Answering this question would give us an indication of this sample quality as will be shown in detail in this paper. + +In the rest of this section, We formalize and empirically rationalize our proposed CR-FIQA approach and its components. To do that, we start by shortly revisiting angular margin penalty-based softmax loss utilized to optimize the class centers of the FR model. Then present a detailed description of our proposed CR-FIQA concept and the associated training paradigm. + +# 3.1. Revisiting Margin Penalty-based Softmax Loss + +Angular margin penalty-based softmax is a widely used loss function for training FR models [3,5,17,28]. It extends over softmax loss by deploying angular penalty margin on the angle between the deep features and their corresponding weights. Margin penalty-based softmax loss aims to push the decision boundary of softmax, and thus enhance intraclass compactness and inter-class discrepancy. From this family of loss functions, this work utilizes ArcFace loss [5] to optimize the distance between the training samples and their class center. Our choice of ArcFace loss is based on the SOTA performance achieved by ResNet-100 network trained with ArcFace on mainstream benchmarks [5]. Formally, ArcFace loss is defined as follows: + +$$ +\mathcal {L} _ {A r c} = \frac {1}{N} \sum_ {i \in N} - \log \frac {e ^ {s \left(\cos \left(\theta_ {y _ {i}} + m\right)\right)}}{e ^ {s \left(\cos \left(\theta_ {y _ {i}} + m\right)\right)} + \sum_ {j = 1 , j \neq y _ {i}} ^ {C} e ^ {s \left(\cos \left(\theta_ {j}\right)\right)}}, \tag {1} +$$ + +where $N$ is the batch size, $C$ is the number of classes (identities), $y_{i}$ is the class label of sample $i$ (in range $[1,C])$ , $\theta_{y_i}$ is the angle between the features $x_{i}$ and the $y_{i}$ -th class center $w_{y_i}$ . $x_{i} \in R^{d}$ is the deep feature embedding of the last fully connected layer of size $d$ . $w_{y_i}$ is the $y_{i}$ -th column of weights $W \in R_C^d$ of the classification layer. $\theta_{y_i}$ is defined as $x_{i} w_{y_{i}}^{T} = \| x_{i} \| \| w_{y_{i}} \| \cos (\theta_{y_{i}})$ [24]. The weights and the feature norms are fixed to $\| w_{y_i} \| = 1$ and $\| x_{i} \| = 1$ , respectively, using $l_2$ normalization as defined in [24, 37]. The decision boundary, in this case, depends on the angle cosine between $x_{i}$ and $w_{y_i}$ . $m > 0$ is an additive angular margin proposed by ArcFace [5] to enhance the intra-class compactness and inter-class discrepancy. Lastly, $s$ is the scaling parameter [37]. + +# 3.2. Certainty Ratio + +In this section, we formulate and empirically rationalize the main concepts that build our FIQA solution. We derive our Certainty Ratio (CR) to estimate the sample relative classifiability. Additionally, we experimentally illustrate the strong relationship between our CR measure and FIQ. + +Certainty Ratio During the FR model training phase, the model is trained to enhance the separability between the classes (identities) by pushing each sample $x_{i}$ to be close to its class center $w_{y_i}$ and far from the other (negative) class centers $w_{j}, j \neq y_{i}$ . Based on this, we first define the Class Center Angular Similarity (CCS) as the proximity between $x_{i}$ and its class center $w_{y_i}$ , as follows: + +$$ +C C S _ {x _ {i}} = \cos \left(\theta_ {y _ {i}}\right), \tag {2} +$$ + +where $\theta_{y_i}$ is the angle between $x_i$ and its class center $w_{y_i}$ , where the weights of the last fully connected layer of the FR model trained with softmax loss are considered as the centers for each class [5,25]. Then, we define the Closest Nearest Negative Class Center Angular Similarity (NNCCS) as proximity between $x_i$ and the nearest negative class center $w_j$ , $j \neq y_i$ . Formally, NNCCS is defined as follows: + +$$ +N N C C S _ {x _ {i}} = \max _ {j = 1, j \neq y _ {i}} ^ {C} (\cos \left(\theta_ {j}\right)), \tag {3} +$$ + +where $\theta_{j}$ is the angle between $x_{i}$ and $w_{j}$ . As we theorize, when the FR model converges, the high-quality samples are pushed closer to their class centers (high CCS) in relation to their distance to neighbouring negative class centers (low NNCCS). However, low-quality samples can not be pushed as close to their class centers. A sample able to achieve high CCS with respect to NNCCS is a sample easily correctly classified during training, and thus is relatively highly classifiable. We thus measure this relative classifiability by the ratio of CCS to NNCS, which we note as the Certainty Ratio (CR), as follows: + +$$ +C R _ {x _ {i}} = \frac {C C S _ {x _ {i}}}{N N C C S _ {x _ {i}} + (1 + \epsilon)}, \tag {4} +$$ + +where the $1 + \epsilon$ term is added to insure a positive above zero denominator, i.e. shift the NNCCS value range from $[-1, +1]$ to $[\epsilon, 2 + \epsilon]$ . This ensures that the CR of a sample with a lower NNCCS is relatively higher than a sample with a higher NNCCS, given the same CCS, i.e. NNCCS regulates the CCS value in relation to neighbouring classes. The $\epsilon$ is set to $1e - 9$ in our experiments. The optimal CR is obtained when the CCS is approaching the maximum cosine similarity value (+1) and the NNCCS is approaching the minimum cosine similarity value (-1), i.e. the training sample is capable of being pushed to its class center, and far away from the closest negative class center, and thus it is highly classifiable. + +# 3.3. Relation between the CR and FIQ + +Here, we empirically prove the theorized relationship between the CR and FIQ (defined earlier as image utility). Namely, we want to answer: if the CR values achieved by training samples of an FR model were used as FIQ, would they behave as expected from an optimal FIQ? If yes, then the face image properties leading to high/low CR + +do also theoretically lead to high/low FIQ. To answer this question we conducted an experiment on a ResNet-50 [13] FR model trained on CASIA-WebFace [39] with ArcFace loss [5] (noted as R50(CASIA)). Specifically, we calculate the CR, CCS, and NNCCS values from the trained model for all samples in the training dataset (0.5M images of 10K identities). An insight on the resulting CCS and NNCCS values (CR being a derivative measure) is given as value distributions in Figure 1, showing that these measures vary between different samples. Furthermore, based on the calculated scores, we plot Error vs. Reject Curves (ERC) (described in Section 4) to demonstrate the relationship between the CR, as an FIQ measure, and FR performance. To calculate the FR performance in the ERC curve, we extract the feature embedding of CASIA-WebFace [39] using a ResNet-100 model [13] trained on MS1M-V2 [5, 12] with ArcFace (noted as R100(MS1M-V2)). We utilize a different model (trained on a different database) to extract the embedding (R100(MS1M-V2)) than the one used to calculate CR, CCS, and NNCCS (R50(CASIA)) to provide fair evaluation where the FR performance is evaluated on unseen data. Then, we perform $n : n$ comparisons between all samples of CASIA-WebFace using feature embedding obtained from the R100(MS1M-V2). + +![](images/f127b852bbb10049bcd2c891ff2f7b7cc4b4eb23f68e608b52856ec8d7617c78.jpg) +(a) + +![](images/bc38a9ae33aced9bb1c910c20058afdc20a61b75cf6cbfab187e61b6df40bf26.jpg) +(b) +Figure 2. ERCs showing the verification performance as False None Match Rate (FNMR) at False Match Rate (FMR) of 1e-3 (2a) and 1e-4 (2b) with CCS, NNCCS and CR as FIQ vs. rejection ratio. This ERC plots show the effectiveness of rejecting samples with the lowest CCS and CR on the performance + +Figure 2 presents the ERC of CR, CCS, and NNCCS experimentally used as FIQ. An FIQ measure would cause the ERC to drop as rapidly as possible when rejecting a larger fraction of low-quality samples (moving to the right). + +It can be clearly noticed in Figure 2 that the CCS and CR do behave as we would expect from a good performing FIQ, as the verification error value drops rapidly when rejecting low quality (low CCS and CR) samples. It can be also observed that the CR does that more steadily when compared to CCS. This points out that adding the scaling term NNCCS in CR calculation can enhance the representation of the CCS as an FIQ measure, which will be clearer later when we experimentally evaluate our CR-FIQA approach in Section 5. As expected, the NNCCS measure by itself does not strongly act as an FIQ measure would, demonstrated by the relatively flat ERC in Figure 2, as it only considers the distance to the nearest negative class. This empirical evaluation does provide a confirming answer to the previ + +ously stated question by affirming that the CR does act as expected from an FIQ measure and thus, theoretically, one can strongly link the image properties that cause high/low CR in the FR training data to these causing high/low FIQ. + +# 3.4. Quality Estimation Training Paradigm + +In the previous section, we proved that the CR does behave as an FIQ would, and thus, it can also relate to image properties that dictate FIQ. However, the CR measure is only observable for samples in the FR training dataset, where the class centers are known. In a real case scenario, the FIQ measure should be assessed to any single image, i.e. unseen evaluation data. Considering this, and in an effort to predict what the CR value would be for a given sample if hypothetically it was part of the FR training, we propose to simultaneously learn to predict the CR from the training dataset while optimizing the class centers (typical FR training) during the training phase, i.e. the CR-FIQA model. To enable this, we add a single regression layer to the FR model. The input of the regression layer is a feature embedding $x_{i}$ and the output is an estimation of the CR. The output of this regression layer is used later to predict the FIQ score of the unseen sample, e.g. from the evaluation dataset. Thus, we capture the properties that make the CR high/low to predict the FIQ of any given sample. Towards this goal, during the training phase, the model (in Figure 1) has two learning objectives: a) It is trained to optimize the distance between the samples and the class centers using ArcFace loss defined in Equation 1. b) It is trained to predict the internal network observation, CR, using Smooth L1-Loss [9] applied between the output of the regression layer (P) and the CR calculated as in Equation 5. Smooth L1-loss can be interpreted as a combination of L1 and L2-losses by defining a threshold $\beta$ that changes between them [9]. Our choice for smooth-l1 loss is based on: 1) It is less sensitive to outliers than l2. The derivative of L2 loss increases when the difference between the prediction and ground-truth label is increased, making the derivative of loss values large at the early stage of the training, leading to unstable training. Additionally, L2 loss can easily generate gradient explosion [9] when there are outliers in the training data. 2) L1 loss can lead to stable training. However, the absolute values of the difference between prediction and ground truth are small, especially in the later stage of the training. Therefore, the model accuracy can hardly be improved at a later stage of the training as the loss function will fluctuate around a stable value. Combining L1 and L2 as in Smooth L1-loss avoids gradient explosion, which might be caused by L2 and facilitates better convergence than L1. The loss leading to the second objective is then given as: + +$$ +\mathcal {L} _ {C R} = \frac {1}{N} \sum_ {i \in N} \left\{ \begin{array}{l l} \frac {0 . 5 \times \left(C R _ {x _ {i}} - P _ {i}\right) ^ {2}}{\beta} & \text {i f} \left| C R _ {x _ {i}} - P _ {i} \right| < \beta \\ \left| C R _ {x _ {i}} - P _ {i} \right| - 0. 5 \times \beta & \text {o t h e r w i s e} \end{array} \right. \tag {5} +$$ + +![](images/c14e758581959881b3fd5bc9341c6344a0bd6d5f4843358172bff43491fe7809.jpg) +(a) + +![](images/c0a0ffe667186e01fc06bdb62efb807dc28b8ac45759bbbdd4e3eed889c96c18.jpg) +(b) + +![](images/7ca45a0c251b02023d85b195cf69060a7183d277e6ce53f7a6ceffe1c590a47e.jpg) +(c) +Figure 3. ERC comparison between CR-FIQA(S), CCS-FIQA(S), CR-FIQA(S) (On top) and CCS-FIQA(S) (On top). The plots show the effect of rejecting samples of lowest quality, on the verification error (FNMR at FMR1e-3). CR-FIQA(S) and CCS-FIQA(S) outperformed the on-top solutions, and CR-FIQA(S) performs generally better than CCS-FIQA(S) (curve decays faster with more rejected samples) + +![](images/f78a781dae2143838300b454c28e485eb6a63d039a537a4192f615f032204c64.jpg) +(d) + +The final loss combining both objectives for training our CR-FIQA model is defined as follows: + +$$ +\mathcal {L} = \mathcal {L} _ {A r c} + \lambda \times \mathcal {L} _ {C R}, \tag {6} +$$ + +where $\lambda$ is a hyper-parameter used to control the balance between the two losses. At the beginning of model training, the value range of $\mathcal{L}_{CR}$ is very small ( $\leq 2$ ) in comparison to $\mathcal{L}_{Arc} (\sim 45)$ . Setting $\lambda$ to a small value, the model will only focus on $\mathcal{L}_{Arc}$ . Besides, setting $\lambda$ to a large value, i.e. $>10$ , we observed that the model did not converge. Therefore, we set $\lambda$ to 10 in all the experiments in this paper. + +# 4. Experimental Setup + +Implementation Details We demonstrate our proposed CR-FIQA under two protocols (small and large) based on the training dataset and the training model architecture. We utilize widely used architectures in the SOTA FR solutions, ResNet100 and ResNet50 [13], both modified as described in Section 3.4. For the small protocol, we utilize ResNet50 and the CASIA-WebFace [39] training data (noted as CR-FIQA(S)) and for the large protocol, we utilize ResNet100 and the MS1MV2 [5, 12] training data (noted as CR-FIQA(L)). The MS1MV2 is a refined version of the MS-Celeb-1M [12] by [5] containing 5.8M images of 85K identities. The CASIA-WebFace contains 0.5m images of 10K identities [39]. We follow the ArcFace training setting [5] to set the scale parameter $s$ to 64 and the margin $m$ to 0.5. We set the mini-batch size to 512. All models are trained with Stochastic Gradient Descent (SGD) optimizer with an initial learning rate of 1e-1. During the training, we use random horizontal flipping with a probability of 0.5 for data augmentation. We set the momentum to 0.9 and the weight decay to 5e-4. For CR-FIQA(S), the learning rate is divided by 10 at 20K and at 28K training iterations, following [5]. The training is stopped after 32K iterations. For CR-FIQA(L), the learning rate is divided by 10 at 100K and 160K training iterations, following [5]. The training is stopped after 180K iterations. All the images in evaluation and training datasets are aligned and cropped to $112 \times 112$ , as described in [5]. All the training and testing images are normalized to have pixel values between -1 and 1. Both + +models are trained using the loss defined in Equation 6. + +Evaluation Benchmarks We reported the achieved results on eight different benchmarks: Labeled Faces in the Wild (LFW) [16], AgeDB-30 [31], Celebrities in Frontal-Profile in the Wild (CFP-FP) [33], Cross-age LFW (CALFW) [41], Adience [6], Cross-Pose LFW (CPLFW) [40], Cross-Quality LFW (XQLFW) [23], IARPA Janus Benchmark-C (IJB-C) [27]. These benchmarks are chosen to provide a wide comparison to SOTA FIQA algorithms and give an insight into the CR-FIQA generalizability. + +Evaluation Metric We evaluate the FIQA by plotting ERCs [11]. The ERC is a widely used representation of the FIQA performance [10, 11] by demonstrating the effect of rejecting a fraction face images, of the lowest quality, on face verification performance in terms of False None Match Rate [21] (FNMR) at a specific threshold calculated at fixed False Match Rate [21] (FMR). The ERC curves for all benchmarks are plotted at two fixed FMRs, 1e-3 (as recommended for border control operations by Frontex [7]) and 1e-4 (the latter is provided in the supplementary material). We also report the Area under the Curve (AUC) of the ERC, to provide a quantitative aggregate measure of verification performance across all rejection ratios. + +Additionally, motivated by evaluating the FIQ as a weighting term for face embedding [32, 34], we follow the IJB-C 1:1 mixed verification benchmark [27] by weighting the frames such that all frames belonging to the same subject within a video have a combined weight equal to a single still image as described in IJB-C benchmark [27]. We do that by using the CR-FIQA quality scores as well as all SOTA methods. We report the verification performance of IJB-C as true acceptance rates (TAR) at false acceptance rates (FAR) of 1e-4, 1e-5, and 1e-6, as defined in [27]. + +Face Recognition Models We utilize four different SOTA FR models to report the verification performance at different quality rejection rate to inspect the generalizability of FIQA over FR solutions. The FR models are ArcFace [5], ElasticFace (ElasticFace-Arc) [3], MagFace [28], and CurricularFace [17]. All models process $112 \times 112$ aligned and cropped image to produce 512-D feature em + +bedding. We used the officially released pretrained ResNet-100 models trained on MS1MV2 released by the four FR solutions. Although, the presented solution in this paper does not aim, and is not presented as, a solution to extract face embeddings, but rather an FIQA solution, we opted to evaluate CR-FIQA(L) backbone as a FR model on mainstream FR benchmarks for sake of providing complete experiment evaluation and probe the possibility of simultaneously using it as both FIQA and FR model. The evaluation results of CR-FIQA(L) backbone as a FR model are provided in the supplementary material. + +Baseline We compare our CR-FIQA approach with nine quality assessment methods. Three are general IQA methods that have been proven in [8] to correlate well to face utility i.e. BRISQUE [29], RankIQA [26], and DeepIQA [2], and six are SOTA face-specific FIQA methods, namely RankIQ [4], PFE [34], SER-FIQ [36], FaceQnet (v1 [14]) [14, 15], MagFace [28], and SDD-FIQA [32], all as officially released in the respective works. + +# 5. Ablation Studies + +This section provides experimental proof of the two main design choices in CR-FIQA. + +Does CR-FIQA benefit from the NNCCS scaling term? To answer this, we conducted additional experiments using ResNet-50 model trained on CASIA-WebFace [39] using the experimental setup described in Section 4. This model is noted as CCS-FIAQ(S). The only difference from CR-FIAQ(S) is that the CCS-FIAQ(S) is trained to learn CCS (instead of CR) by replacing the $CR_{x_i}$ in Equation 5 with $CCS_{x_i}$ , thus neglecting the NNCCS scaling term in the equation. Figure 3 presents the ERCs along with AUC using CR-FIAQ(S) and CCS-FIAQ(S) on Adience, AgeDb-30, CALLFW and CRF-FP. The verification error, FNMR at FMR1e-3, is calculated using ArcFace FR model (described in Section 4). The ERCs and AUC values show that the reduction in the error is more evident for CR-FIAQ(S) than CCS-FIQA(S). Thus, adding the scaling term NNCCS in CR enhanced the performance of the FIQA. + +Dose the simultaneous learning in CR-FIQA lead to better performance in comparison to on-the-top learning? We consider the possibility of learning to estimate CR after finalizing the FR training, in comparison to the simultaneously learning in CR-FIQA. We conducted two additional experiments using pretrained ResNet-50 trained with ArcFace loss [5] on CASIA-WebFace [39]. Specifically, we add an additional single regression layer to this pretrained model. We freeze the weights of the pretrained model and train only the regression layer to learn the internal network observation of the pretrained model using only the $\mathcal{L}_{\mathcal{CR}}$ (Equation 5). Using this setting, we present two instances, CR-FIQA(S) (On top) and CCS-FIQA(S) (On top), that learned to predict CR and CCS, respectively. Each of CR-FIQA(S) (On top) and CCS-FIQA(S) (On top) is fine + +tuned for 32K iteration with an initial learning rate of 0.01. The learning rate is divided by 10 at 20K and 28K training iterations, similarly to CR-FIAQ(S) and CCS-FIAQ(S). The results (ERCs and AUCs) of these models are compared to CR-FIQA(S) and CCS-FIQA(S) in Figure 3. The ERCs in Figure 3 presented the evaluation on Adience, AgeDb-30, CALFW and CFP-FP using the ArcFace FR model. The ERCs and AUCs in Figure 3 show that both CR-FIQA(S) and CCS-FIQA(S) lead to stronger reductions in the error than the CR-FIQA(S) (On top) and the CCS-FIQA(S) (On top), when rejecting low-quality samples. This supports our training paradigm that simultaneously learns the internal network observation, CR, while optimizing the class centers. This can be related to the step-wise convergence towards the final CR value in the simultaneously training. For both ablation study questions, ERCs and AUC for all the remaining benchmarks and FR models (mentioned in Section 4) lead to similar conclusions and are provided in the supplementary material. + +# 6. Result and Discussion + +All CR-FIQA performances reported in this paper are obtained under cross-model settings. The proposed CR-FIQA is used only to predict FIQ and not to extract feature representation of face images. None of the utilized FR models (ArcFace [5], ElasticFace [3], MagFace [28], and CurricularFace [17]) is trained with our paradigm. Instead, we trained a separate model for CR-FIQA and used the official pretrained FR models (as described in Section 4) for feature extraction. The verification performances as AUC at FMR1e-3 and FMR1e-4 are presented in Table 1. The visual verification performances as ERC curves (Figure 4) using ArcFace and ElasticFace FR models are reported at FMR1e-3. The ERC curves at FMR1e-4 and for MagFace and CurricularFace FR models at FMR1e-3 are provided in the supplementary material. + +The ERC curves (Figure 4) and the AUC values (Table 1) show that our proposed CR-FIQA(S) and CR-FIQA(L) outperformed the SOTA methods by significant margins in almost all settings. Observing the results on IJB-C, Adience, CFP-FP, CALFW, and CPLFW at FMR1e-3 and FMR1e-4 (Figure 4 and Table 1), our proposed CR-FIQA outperformed all SOTA methods on all the considered FR models. On the AgeDB-30 benchmark, our proposed CR-FIQA ranked first in five out of eight settings and second in the other three settings (Table 1). On the LFW benchmark, our proposed CR-FIQA ranked behind the MagFace and the RankIQ. This is the only case that our models did not outperform all SOTA methods. However, it can be noticed from the ERC curves in Figure 4 that none of the SOTA methods were able to achieve stable behavior (smoothly decaying curve) on LFW. The main reason for such unstable ERC behavior on LFW is that the FR performance on LFW is nearly saturated (all models achieved above $99.80\%$ accu + +Table 1. The AUCs of ERC achieved by our CR-FIQA and the SOTA methods under different experimental settings. CR-FIQA achieved the best performance (lowest AUC) in almost all settings. On XQLFW, the SER-FIQ (marked with *) is used for the sample selection of the XQLFW benchmark. The best result for each experimental setting is in bold and the second-ranked one is in italic. The notions of $1e - 3$ and $1e - 4$ indicate the value of the fixed FMR at which the ERC curve (FNMR vs. reject) was calculated at + +
FRMethodAdience [6]AgeDB-30 [31]CFP-FP [33]LFW [16]CALFW [41]CPLFW [40]XQLFW [23]IJB-C [27]
1e-31e-41e-31e-41e-31e-41e-31e-41e-31e-41e-31e-41e-31e-41e-31e-4
ArcFace [5]QABRISQUE [29]0.05650.12850.04000.05850.03430.04330.00430.00490.07550.08130.25580.30370.66800.71220.03810.0656
RankIQA [26]0.04000.09330.03720.05230.03010.03840.00390.00450.08460.09150.24370.29690.65840.70390.03850.0640
DeepIQA [2]0.05680.13720.04030.05230.02380.02920.00490.00560.07930.08500.23090.28560.59580.64580.03830.0640
FIQARankIQ [4]0.03530.08730.03220.04200.01520.02600.00180.00240.06080.06720.06330.08480.27890.33320.02270.0342
PFE [34]0.02120.04280.01720.02260.00920.01290.00230.00280.06470.06810.04500.06380.23020.27100.01760.0248
SER-FIQ [36]0.02230.04340.01670.02230.00650.01030.00230.00280.05950.06270.03890.05840.1812*0.2295*0.01610.0241
FaceQnet [14, 15]0.03460.07340.01970.02450.02400.02730.00220.00270.07740.08220.15040.17510.58290.61360.02700.0376
MagFace [28]0.02070.04250.01560.01980.00730.01050.00160.00210.05680.06020.04920.06420.40220.46360.01710.0254
SDD-FIQA [32]0.02480.05620.01860.02060.01220.01930.00210.00270.06410.06980.05170.06700.30900.35610.01860.0270
CR-FIQA(S)(Our)0.02410.05170.01440.01870.00900.01450.00200.00250.05210.05540.03910.05670.23770.27400.01710.0250
CR-FIQA(L)(Our)0.02040.03530.01590.01890.00500.00820.00230.00290.06160.06320.03600.05150.20840.24410.01380.0207
ElasticFace [3]QABRISQUE [29]0.06440.11840.03750.04030.02810.03720.00340.00470.07260.07470.26410.46880.63430.69640.03570.0621
RankIQA [26]0.04330.08620.03740.04360.02690.03180.00330.00450.08100.08350.23250.43060.61890.68560.03660.0599
DeepIQA [2]0.06450.12030.03840.04110.01910.02560.00430.00560.07560.07720.24010.45410.54000.58320.03790.0590
FIQARankIQ [4]0.04000.07770.03090.03370.01490.01800.00130.00200.05980.06140.05810.07270.24680.27760.02260.0334
PFE [34]0.02220.03810.01630.01720.00880.01130.00180.00250.06280.06430.04190.08950.21120.24360.01710.0247
SER-FIQ [36]0.02400.04170.01630.01790.00610.00850.00210.00280.05740.05900.03870.05130.1576*0.1868*0.01560.0235
FaceQnet [14, 15]0.03690.06670.01940.02070.02270.02470.00210.00260.07630.07770.14200.28800.55490.58440.02630.0370
MagFace [28]0.02250.03850.01500.01580.00690.00950.00140.00210.05530.05630.04740.05970.39730.42820.01660.0243
SDD-FIQA [32]0.02770.05120.01870.02000.00980.01180.00190.00270.06240.06380.04930.06340.30520.35620.01830.0266
CR-FIQA(S)(Our)0.02570.04650.01460.01600.00700.00960.00150.00220.05090.05220.03830.05020.20930.28350.01670.0244
CR-FIQA(L)(Our)0.02140.03570.01490.01590.00450.00650.00180.00250.05940.06080.03500.04620.17980.20600.01350.0203
MagFace [28]QABRISQUE [29]0.05940.13080.04420.07990.04220.05890.00430.00580.07580.07880.46490.68090.69110.72290.04620.0787
RankIQA [26]0.04070.08890.03700.06810.03690.05430.00410.00560.08290.08570.32510.64750.67060.70460.04610.0750
DeepIQA [2]0.05710.13020.04170.07210.03220.05450.00480.00590.07870.08090.36720.66320.61620.65190.04740.0765
FIQARankIQ [4]0.03590.08370.03610.05310.02130.03320.00790.00270.06020.06290.06590.16420.30760.34750.02690.0383
PFE [34]0.02150.04230.01920.03170.01070.01380.00230.00290.06400.06520.04490.14350.26150.29260.02000.0283
SER-FIQ [36]0.02330.04510.01850.02930.00800.01390.00250.00330.05900.06070.03970.52180.60160.62100.03050.0422
FaceQnet [14, 15]0.03650.07200.02170.03140.02710.03510.00220.00270.07630.07730.29880.52180.64750.71880.19550.1979
MagFace [28]0.02120.04170.01590.02470.00850.01290.00170.00220.05620.05780.05060.88870.44780.49000.01950.0279
SDD-FIQA [32]0.02530.05620.02160.03050.01460.02010.00210.06430.06570.05250.11880.34040.39280.39280.02150.0307
CR-FIQA(S)(Our)0.02440.05070.01650.02340.01020.01210.00200.00280.05160.05280.04090.08400.26700.33360.01980.0284
CR-FIQA(L)(Our)0.02110.03720.01740.02350.00620.00800.00230.00280.06140.06280.03740.06790.23690.28390.01620.0236
CurricularFace [17]QABRISQUE [29]0.05020.10950.04330.04910.03230.03570.00410.00470.07550.07840.27090.50570.61460.63360.03620.0589
RankIQA [26]0.03590.07520.03940.05100.02980.03560.00390.00450.88060.86560.23460.46540.56440.62120.13610.1558
DeepIQA [2]0.04920.10700.04070.04760.02270.02780.00500.00560.7640.7860.24880.49610.51650.55260.13760.1558
FIQARankIQ [4]0.03140.07150.03650.04170.01860.02490.00180.00240.59590.66451515*1515*1515*1515*1515*1515*
PFE [34]0.01980.03650.01970.02270.01030.01340.00240.00280.63630.65751515*1515*1515*1515*1515*1515*
SER-FIQ [36]0.02110.03810.01670.01930.00740.01110.00250.03561518*1519*1518*1519*1519*1519*1519*1519*
FaceQnet [14, 15]0.03260.06260.02210.02670.02260.02740.00220.02771518*1519*1518*1519*1519*1519*1519*1519*
MagFace [28]0.02000.03640.01670.01950.00780.01110.00160.00211516*1517*1518*1519*1519*1519*1519*1519*
SDD-FIQA [32]0.02300.04620.02190.02540.01380.01850.00210.06271517*1518*1518*1519*1519*1519*1519*1519*
CR-FIQA(S)(Our)0.02270.04460.01560.01980.00970.01480.00200.02551513*1514*1514*1515*1515*1515*1515*1515*
CR-FIQA(L)(Our)0.01980.03360.01620.02000.00540.08800.00230.02991516*1517*1518*1519*1519*1519*1519*1519*
+ +Table 2. Verification performance on the IJB-C (1:1 mixed verification) [27]. CR-FIQA outperformed SOTA methods under all settings + +
Quality Estimation1:1 mixed Verification: TAR (%) at
ArcFace [5]ElasticFace [3]MagFace [28]CurricularFace [17]
FAR=1e-6FAR=1e-5FAR=1e-4FAR=1e-6FAR=1e-5FAR=1e-4FAR=1e-6FAR=1e-5FAR=1e-4FAR=1e-6FAR=1e-5FAR=1e-4
-89.8594.4796.2889.1594.5496.4985.6793.0896.6590.4694.8996.58
IAQBRISQUE [29]86.6593.6295.9885.6893.5195.6581.1190.6494.8288.1693.9896.29
RankIQA [26]86.3793.6195.8386.7193.4696.0080.7890.7594.8688.1694.1196.22
DeepIQA [2]81.9791.6494.6778.9391.5994.8173.5386.3492.9082.6592.0495.00
FIQARankIQ [4]88.7894.4296.2088.8894.6496.4585.6392.6695.7090.0094.9396.53
PFE [34]89.5094.5196.3189.1094.6796.5184.9392.4495.6090.3695.0496.54
SER-FIQ [36]89.7494.6596.3290.0594.7996.5786.0293.3595.8090.6695.1196.58
FaceQNet [14, 15]87.8794.0496.1286.2694.0996.2582.9190.5695.0389.6194.6596.36
MagFace [28]89.4994.4196.2289.3794.6996.4685.7592.7195.5490.3495.0296.50
SDD-FIQA [32]89.3994.6196.3488.0794.8296.4984.6992.8395.7389.9195.1296.63
CR-FIQA(S)(Our)89.5994.7896.3590.3094.9796.6386.4593.4895.9590.8295.1396.64
CR-FIQA(L)(Our)90.1694.7596.3690.0094.9296.5887.1293.6795.9090.7995.1296.58
+ +racy [3,5,17,28]), leaving very few samples causing errors and thus lowering the significance of the measured FNMR. + +The XQLFW benchmark is derived from LFW to contain pairs with a maximum difference in quality. The XQLFW images are chosen based on BRISQUE [29] and SER-FIQ [23] quality scores to be either extremely high or low. The use of SER-FIQ [36] in this selection might give a biased edge for SER-FIQ on this benchmark. On XQLFW, our CR-FIQA achieved very close performance to the selection method (SER-FIQ) and is far ahead of all other SOTA + +methods. Lastly, our proposed CR-FIQA(S) achieved very comparable performance to our CR-FIQA(L), pointing out the robustness of our approach, regardless of the training database and architecture complexity. + +Table 2 presents the verification performance on the IJB-C 1:1 mixed verification benchmark [27] using quality scores as an embedding weighting term (as defined in [27]) under different experimental settings. For each of the FR models, we report in the first row the evaluation result of the corresponding FR model as defined in the protocol [27] and + +![](images/a82f22ad1d52348cae670ed95b1d433880540117ff65dfeac6f27bed43eab300.jpg) + +![](images/35fd13f5e6b524ca30ce69a4ca87e66bb352cb00e02ffce4323012f9e319af6b.jpg) + +![](images/dd3f2ff2716f695cb1b20f5a21e8994a64b92decdde2434660c1580bf37c4872.jpg) + +![](images/eeab048702330d160cf068de1e61c9c8c719c70d4a5234bb2c1a337493f81ec3.jpg) + +![](images/fadcbf964e12e0dbe9bcadd5f6854b07c15d34449b1b5e0006c456b408d746ae.jpg) + +![](images/7741785b04c7e29ac5e32d5519d423666998ec255da71cdf7abe0cd322fb48f3.jpg) + +![](images/8ab76e9fddc9a173b612786ef91a3f8c3ec9eabb5edfcfa96acb622fa9884c91.jpg) + +![](images/e2e3e85d37ad943874fb5cf960272f0d698373ca90379092b9d3c4a0fa897358.jpg) + +![](images/771b35430170d9df9908ce8d99f810923a68162605e46e89191629d401bdf2a7.jpg) + +![](images/80a3098b1c61c8fa07b1fe66a09f99bd622791f47301a62ec369cf59f48fbed5.jpg) + +![](images/b74275bebd231178c2913a54158c46f8c1aafdcd4751a5c33b91115c833e7f16.jpg) + +![](images/754b9d7570e68dac2860bd33ad61da7fb0e41879677bb01c05e17149e0f8f39e.jpg) + +Figure 4. ERC (FNMR at FMR1e-3 vs reject) curves for all evaluated benchmarks using ArcFace and ElasticFace FR models corresponding to Table 1 results. The visual evaluation, ERC curves, using MagFace and CurricularFace FR models are provided in the supplementary material. The proposed CR-FIQA(L) and CR-FIQA(S) are marked with solid blue and red lines, respectively. CR-FIQA leads to lower verification error, when rejecting a fraction of images, of the lowest quality, in comparison to SOTA methods (faster decaying curve) under most experimental settings. +![](images/fd3370c640db7d3890be16d27af380026231c669c8b5ff0e18eee497f6a5dac8.jpg) +BRIQUE RankIQ DeepIQ $\rightarrow$ RankIQ PFE SER-FIQ FaceQnet MagFace SDD-FIQA CR-FIQA(S) Our) CR-FIQA(L) Our + +![](images/c51273b871eb43e818f7674904a61d55ce6efead9e5ec90af238d4680af0f9b2.jpg) + +![](images/4574db913f7abfee3cf4c2214a57e182c99d16898a22d7961980aa2737ed2843.jpg) + +![](images/e0e060cd3b1e3f8ff022c16b61b1dd512397d2ca7e17851a36ebbc283c3e656c.jpg) + +the corresponding released evaluation scripts [3, 5, 17, 28], i.e. without considering the FIQ. Our proposed CR-FIQA significantly leads to higher verification performance than all evaluated SOTA methods, when the quality score is used as an embedding weighting term (Table 2). This achievement is observable under all experimental settings (Table 2). Another outcome of this evaluation is that the integration of CR-FIQA leads to SOTA verification performance on one of the most challenging FR benchmarks, IJB-C [27]. + +# 7. Conclusion + +In this work, we propose the CR-FIQA approach that probes the relative classifiability of training samples of the FR model and utilize this observation to learn to predict the utility of any given sample in achieving an accurate FR performance. We experimentally prove the theorized relationship between the sample relative classifiability and FIQ and build on that towards our CR-FIQA. The CR-FIQA training paradigm simultaneously learns to optimize the class center + +while learning to predict sample relative classifiability. The presented ablation studies and the extensive experimental results prove the effectiveness of the proposed CR-FIQA approach, and its design choices, as an FIQ method. The reported results demonstrated that our proposed CR-FIQA outperformed SOTA methods repeatedly across multiple FR models and on multiple benchmarks, including ones with a large age gap (AgeDb-30, Adience, CALFW), large quality difference (XQLFW), large pose variation (CPLFW, CFPFP), and extremely large-scale and challenging FR benchmarks (IJB-C). + +Acknowledgment This research work has been funded by the German Federal Ministry of Education and Research and the Hessian Ministry of Higher Education, Research, Science and the Arts within their joint support of the National Research Center for Applied Cybersecurity ATHENE. This work has been partially funded by the German Federal Ministry of Education and Research (BMBF) through the Software Campus Project. + +# References + +[1] Lacey Best-Rowden and Anil K. Jain. Learning face image quality from human assessments. IEEE Trans. Inf. Forensics Secur., 13(12):3064–3077, 2018. 1, 2 +[2] Sebastian Bosse, Dominique Maniry, Klaus-Robert Müller, Thomas Wiegand, and Wojciech Samek. Deep neural networks for no-reference and full-reference image quality assessment. IEEE Trans. Image Process., 27(1):206-219, 2018. 6, 7 +[3] Fadi Boutros, Naser Damer, Florian Kirchbuchner, and Arjan Kuijper. Elasticface: Elastic margin loss for deep face recognition. In IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, CVPR Workshops 2022, New Orleans, LA, USA, June 19-20, 2022, pages 1577-1586. IEEE, 2022. 3, 5, 6, 7, 8 +[4] Jiansheng Chen, Yu Deng, Gaocheng Bai, and Guangda Su. Face image quality assessment based on learning to rank. IEEE Signal Process. Lett., 22(1):90-94, 2015. 2, 6, 7 +[5] Jiankang Deng, Jia Guo, Niannan Xue, and Stefanos Zafeiriou. Arcface: Additive angular margin loss for deep face recognition. In IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2019, Long Beach, CA, USA, June 16-20, 2019, pages 4690-4699. Computer Vision Foundation / IEEE, 2019. 3, 4, 5, 6, 7, 8 +[6] Eran Eidinger, Roee Enbar, and Tal Hassner. Age and gender estimation of unfiltered faces. IEEE Trans. Inf. Forensics Secur., 9(12):2170-2179, 2014. 5, 7 +[7] Frontex. Best practice technical guidelines for automated border control (abc) systems, 2015. 5 +[8] Biying Fu, Cong Chen, Olaf Henniger, and Naser Damer. A deep insight into measuring face image utility with general and face-specific image quality metrics. pages 1121-1130, 2022. 1, 6 +[9] Ross B. Girshick. Fast R-CNN. In 2015 IEEE International Conference on Computer Vision, ICCV 2015, Santiago, Chile, December 7-13, 2015, pages 1440-1448. IEEE Computer Society, 2015. 4 +[10] P. Grother, M. Ngan A. Hom, and K. Hanaoka. Ongoing face recognition vendor test (frvt) part 5: Face image quality assessment (4th draft). In National Institute of Standards and Technology. Tech. Rep., Sep. 2021. 1, 5 +[11] P. Grother and E. Tabassi. Performance of biometric quality measures. IEEE Trans. on Pattern Analysis and Machine Intelligence, 29(4):531-543, Apr. 2007. 5 +[12] Yandong Guo, Lei Zhang, Yuxiao Hu, Xiaodong He, and Jianfeng Gao. Ms-celeb-Im: A dataset and benchmark for large-scale face recognition. In Bastian Leibe, Jiri Matas, Nicu Sebe, and Max Welling, editors, Computer Vision - ECCV 2016 - 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part III, volume 9907 of Lecture Notes in Computer Science, pages 87-102. Springer, 2016. 4, 5 +[13] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In 2016 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2016, Las Vegas, NV, USA, June 27-30, 2016, pages 770-778. IEEE Computer Society, 2016. 4, 5 + +[14] Javier Hernandez-Ortega, Javier Galbally, Julian Fierrez, and Laurent Beslay. Biometric quality: Review and application to face recognition with faceqnet. CoRR, abs/2006.03298, 2020. 1, 2, 6, 7 +[15] Javier Hernandez-Ortega, Javier Galbally, Julian Fierrez, Rudolf Haraksim, and Laurent Beslay. Faceqnet: Quality assessment for face recognition based on deep learning. In 2019 International Conference on Biometrics, ICB 2019, Crete, Greece, June 4-7, 2019, pages 1-8. IEEE, 2019. 1, 2, 6, 7 +[16] Gary B. Huang, Manu Ramesh, Tamara Berg, and Erik Learned-Miller. Labeled faces in the wild: A database for studying face recognition in unconstrained environments. Technical Report 07-49, University of Massachusetts, Amherst, October 2007. 5, 7 +[17] Yuge Huang, Yuhan Wang, Ying Tai, Xiaoming Liu, Pengcheng Shen, Shaoxin Li, Jilin Li, and Feiyue Huang. Curricularface: Adaptive curriculum learning loss for deep face recognition. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2020, Seattle, WA, USA, June 13-19, 2020, pages 5900-5909. Computer Vision Foundation / IEEE, 2020. 3, 5, 6, 7, 8 +[18] ISO/IEC JTC1 SC17 WG3. Portrait Quality - Reference Facial Images for MRTD. International Civil Aviation Organization, 2018. 2 +[19] ISO/IEC JTC1 SC37 Biometrics. ISO/IEC 29794-1:2016 Information technology - Biometric sample quality - Part 1: Framework. International Organization for Standardization, 2016. 1 +[20] ISO/IEC JTC1 SC37 Biometrics. ISO/IEC 2382-37:2017 Information technology - Vocabulary - Part 37: Biometrics. International Organization for Standardization, 2017. 1 +[21] ISO/IEC JTC1 SC37 Biometrics. ISO/IEC 19795-1:2021 Information technology — Biometric performance testing and reporting — Part 1: Principles and framework. International Organization for Standardization, 2021. 5 +[22] Thorsten Joachims. Optimizing search engines using click-through data. In Proceedings of the Eighth ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, July 23-26, 2002, Edmonton, Alberta, Canada, pages 133-142. ACM, 2002. 2 +[23] Martin Knoche, Stefan Hörmann, and Gerhard Rigoll. Cross-quality LFW: A database for analyzing cross-resolution image face recognition in unconstrained environments. In 16th IEEE International Conference on Automatic Face and Gesture Recognition, FG 2021, Jodhpur, India, December 15-18, 2021, pages 1-5. IEEE, 2021. 5, 7 +[24] Weiyang Liu, Yandong Wen, Zhiding Yu, Ming Li, Bhiksha Raj, and Le Song. Spherical face: Deep hypersphere embedding for face recognition. In 2017 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2017, Honolulu, HI, USA, July 21-26, 2017, pages 6738-6746. IEEE Computer Society, 2017. 3 +[25] Weiyang Liu, Yandong Wen, Zhiding Yu, Ming Li, Bhiksha Raj, and Le Song. SphereFace: Deep hypersphere embedding for face recognition. In Proc. of the IEEE Conf. on Computer Vision and Pattern Recognition, pages 212-220, 2017. 3 + +[26] Xialei Liu, Joost van de Weijer, and Andrew D. Bagdanov. Rankiaq: Learning from rankings for no-reference image quality assessment. In IEEE International Conference on Computer Vision, ICCV 2017, Venice, Italy, October 22-29, 2017, pages 1040-1049. IEEE Computer Society, 2017. 1, 6, 7 +[27] Brianna Maze, Jocelyn C. Adams, James A. Duncan, Nathan D. Kalka, Tim Miller, Charles Otto, Anil K. Jain, W. Tyler Niggel, Janet Anderson, Jordan Cheney, and Patrick Grother. IARPA janus benchmark - C: face dataset and protocol. In 2018 International Conference on Biometrics, ICB 2018, Gold Coast, Australia, February 20-23, 2018, pages 158-165. IEEE, 2018. 5, 7, 8 +[28] Qiang Meng, Shichao Zhao, Zhida Huang, and Feng Zhou. Magface: A universal representation for face recognition and quality assessment. In IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2021, virtual, June 19-25, 2021, pages 14225-14234. Computer Vision Foundation / IEEE, 2021. 1, 2, 3, 5, 6, 7, 8 +[29] Anish Mittal, Anush Krishna Moorthy, and Alan Conrad Bovik. No-reference image quality assessment in the spatial domain. IEEE Trans. Image Process., 21(12):4695-4708, 2012. 1, 6, 7 +[30] Anish Mittal, Rajiv Soundararajan, and Alan C. Bovik. Making a "completely blind" image quality analyzer. IEEE Signal Process. Lett., 20(3):209-212, 2013. 1 +[31] Stylianos Moschoglou, Athanasios Papaioannou, Christos Sagonas, Jiankang Deng, Irene Kotsia, and Stefanos Zafeiriou. Agedb: The first manually collected, in-the-wild age database. In 2017 IEEE CVPRW, CVPR Workshops 2017, Honolulu, HI, USA, July 21-26, 2017, pages 1997-2005. IEEE Computer Society, 2017. 5, 7 +[32] Fu-Zhao Ou, Xingyu Chen, Ruixin Zhang, Yuge Huang, Shaoxin Li, Jilin Li, Yong Li, Liujuan Cao, and Yuan-Gen Wang. SDD-FIQA: unsupervised face image quality assessment with similarity distribution distance. In IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2021, virtual, June 19-25, 2021, pages 7670-7679. Computer Vision Foundation / IEEE, 2021. 1, 2, 5, 6, 7 +[33] Soumyadip Sengupta, Jun-Cheng Chen, Carlos Domingo Castillo, Vishal M. Patel, Rama Chellappa, and David W. Jacobs. Frontal to profile face verification in the wild. In 2016 IEEE Winter Conference on Applications of Computer Vision, WACV 2016, Lake Placid, NY, USA, March 7-10, 2016, pages 1-9. IEEE Computer Society, 2016. 5, 7 +[34] Yichun Shi and Anil K. Jain. Probabilistic face embeddings. In 2019 IEEE/CVF International Conference on Computer Vision, ICCV 2019, Seoul, Korea (South), October 27 - November 2, 2019, pages 6901-6910. IEEE, 2019. 1, 2, 5, 6, 7 +[35] Philipp Terhörst, Jan Niklas Kolf, Naser Damer, Florian Kirchbuchner, and Arjan Kuijper. Face quality estimation and its correlation to demographic and non-demographic bias in face recognition. In 2020 IEEE International Joint Conference on Biometrics, IJCB 2020, Houston, TX, USA, September 28 - October 1, 2020, pages 1-11. IEEE, 2020. 1 +[36] Philipp Terhörst, Jan Niklas Kolf, Naser Damer, Florian Kirchbuchner, and Arjan Kuijper. SER-FIQ: unsupervised + +estimation of face image quality based on stochastic embedding robustness. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2020, Seattle, WA, USA, June 13-19, 2020, pages 5650-5659. Computer Vision Foundation / IEEE, 2020. 1, 2, 6, 7 +[37] Hao Wang, Yitong Wang, Zheng Zhou, Xing Ji, Dihong Gong, Jingchao Zhou, Zhifeng Li, and Wei Liu. Cosface: Large margin cosine loss for deep face recognition. In 2018 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2018, Salt Lake City, UT, USA, June 18-22, 2018, pages 5265-5274. IEEE Computer Society, 2018. 3 +[38] Weidi Xie, Jeffrey Byrne, and Andrew Zisserman. Inducing predictive uncertainty estimation for face verification. In 31st British Machine Vision Conference 2020, BMVC 2020, Virtual Event, UK, September 7-10, 2020. BMVA Press, 2020. 1, 2 +[39] Dong Yi, Zhen Lei, Shengcai Liao, and Stan Z. Li. Learning face representation from scratch. CoRR, abs/1411.7923, 2014. 2, 4, 5, 6 +[40] T. Zheng and W. Deng. Cross-posed lfw: A database for studying cross-posed face recognition in unconstrained environments. Technical Report 18-01, Beijing University of Posts and Telecommunications, February 2018. 5, 7 +[41] Tianyue Zheng, Weihong Deng, and Jiani Hu. Cross-age LFW: A database for studying cross-age face recognition in unconstrained environments. CoRR, abs/1708.08197, 2017. 5, 7 \ No newline at end of file diff --git a/2023/CR-FIQA_ Face Image Quality Assessment by Learning Sample Relative Classifiability/images.zip b/2023/CR-FIQA_ Face Image Quality Assessment by Learning Sample Relative Classifiability/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..ac3656aef7dc24cd49ac7f48ac9991423b5dd2fc --- /dev/null +++ b/2023/CR-FIQA_ Face Image Quality Assessment by Learning Sample Relative Classifiability/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:82b9345a79788741afc30042a5892783f4da3c4524986a8a421316a82626fe20 +size 971199 diff --git a/2023/CR-FIQA_ Face Image Quality Assessment by Learning Sample Relative Classifiability/layout.json b/2023/CR-FIQA_ Face Image Quality Assessment by Learning Sample Relative Classifiability/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..65e5e432494ec9fb0bbfd23ad04d1b53f183e0d7 --- /dev/null +++ b/2023/CR-FIQA_ Face Image Quality Assessment by Learning Sample Relative Classifiability/layout.json @@ -0,0 +1,7276 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 73, + 102, + 521, + 139 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 73, + 102, + 521, + 139 + ], + "spans": [ + { + "bbox": [ + 73, + 102, + 521, + 139 + ], + "type": "text", + "content": "CR-FIQA: Face Image Quality Assessment by Learning Sample Relative Classifiability" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 99, + 160, + 494, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 99, + 160, + 494, + 217 + ], + "spans": [ + { + "bbox": [ + 99, + 160, + 494, + 217 + ], + "type": "text", + "content": "Fadi Boutros1, Meiling Fang1,2, Marcel Klemt1, Biying Fu1, Naser Damer1,2 \n1Fraunhofer Institute for Computer Graphics Research IGD, Darmstadt, Germany \n2Department of Computer Science, TU Darmstadt, Darmstadt, Germany \nEmail: fadi.boutros@igd.fraunhofer.de" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 143, + 244, + 192, + 258 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 244, + 192, + 258 + ], + "spans": [ + { + "bbox": [ + 143, + 244, + 192, + 258 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 270, + 290, + 485 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 270, + 290, + 485 + ], + "spans": [ + { + "bbox": [ + 46, + 270, + 290, + 485 + ], + "type": "text", + "content": "Face image quality assessment (FIQA) estimates the utility of the captured image in achieving reliable and accurate recognition performance. This work proposes a novel FIQA method, CR-FIQA, that estimates the face image quality of a sample by learning to predict its relative classifiability. This classifiability is measured based on the allocation of the training sample feature representation in angular space with respect to its class center and the nearest negative class center. We experimentally illustrate the correlation between the face image quality and the sample relative classifiability. As such property is only observable for the training dataset, we propose to learn this property by probing internal network observations during the training process and utilizing it to predict the quality of unseen samples. Through extensive evaluation experiments on eight benchmarks and four face recognition models, we demonstrate the superiority of our proposed CR-FIQA over state-of-the-art (SOTA) FIQA algorithms." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 491, + 128, + 502 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 491, + 128, + 502 + ], + "spans": [ + { + "bbox": [ + 47, + 491, + 128, + 502 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 505, + 287, + 564 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 505, + 287, + 564 + ], + "spans": [ + { + "bbox": [ + 46, + 505, + 287, + 564 + ], + "type": "text", + "content": "Face image utility indicates the utility (value) of an image to face recognition (FR) algorithms [1, 19]. This utility is measured with a scalar, namely the face image quality (FIQ) score, following the definition in ISO/IEC 2382-37 [20] and the FR Vendor Test (FRVT) for FIQA [10]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 564, + 288, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 564, + 288, + 696 + ], + "spans": [ + { + "bbox": [ + 46, + 564, + 288, + 696 + ], + "type": "text", + "content": "As FIQA measures the face utility to FR algorithm, it does not necessary reflects, and does not aim at measuring, the perceived image quality, e.g. a profile face image can be of high perceived quality but of low utility to FR algorithm [35]. Assessing this perceived image quality has been addressed in the literature by general image quality assessment (IQA) methods [26, 29, 30] and is different than assessing the utility of an the image for FR. This is reflected by FIQA methods [28, 32, 36] significantly outperforming IQA methods [26, 29, 30] in measuring the utility [19] of face images in FR, as demonstrated in [8, 28, 36]." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 246, + 546, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 246, + 546, + 389 + ], + "spans": [ + { + "bbox": [ + 306, + 246, + 546, + 389 + ], + "type": "text", + "content": "SOTA FIQA methods focused either on creating concepts to label the training data with FIQ scores and then learn a regression problem [14, 15, 32], or on developing a link between face embedding properties under certain scenarios and the FIQ [28, 34, 36]. Generally, the second approach led to better FIQA performances with most works mentioning the error-prone labeling of the ground truth quality in the first research direction as a possible reason [28, 36]. However, in the second category, transferring the information in network embeddings into an FIQ score is not a learnable process, but rather a form of statistical analysis, which might not be optimal." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 391, + 547, + 643 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 391, + 547, + 643 + ], + "spans": [ + { + "bbox": [ + 304, + 391, + 547, + 643 + ], + "type": "text", + "content": "This paper proposes a novel learning paradigm to assess FIQ, namely the CR-FIQA. Our concept is based on learning to predict the classifiability of FR training samples by probing internal network observations that point to the relative proximity of these samples to their class centers and negative class centers. This regression is learned simultaneously with a conventional FR training process that minimizes the distance between the training samples and their class centers. Linking the properties that cause high/low classifiability of a training sample to the properties leading to high/low FIQ, we can use our CR-FIQA to predict the FIQ of any given sample. We empirically prove the theorized link between classifiability (Section 3.3) and FIQ and conduct thorough ablation studies on key aspects of our CR-FIQA design (Section 5). The proposed CR-FIQA is evaluated on eight benchmarks along with SOTA FIQAs. The reported results on four FR models demonstrate the superiority of our proposed CR-FIQA over SOTA methods and the stability of its performance across different FR models. An overview of the proposed CR-FIQA is presented in Figure 1 and will be clarified in detail in this paper." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 649, + 391, + 662 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 649, + 391, + 662 + ], + "spans": [ + { + "bbox": [ + 306, + 649, + 391, + 662 + ], + "type": "text", + "content": "2. Related work" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 665, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 665, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 305, + 665, + 547, + 715 + ], + "type": "text", + "content": "The recent SOTA FIQA approaches can be roughly grouped into two main categories. The first are approaches that learn a straight forward regressions problem to assess a FIQ score [1,14,15,32,38]. The second category uses prop" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 702, + 223, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 702, + 223, + 712 + ], + "spans": [ + { + "bbox": [ + 58, + 702, + 223, + 712 + ], + "type": "text", + "content": "1https://github.com/fdbtrs/CR-FIQA" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "5836" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 111, + 71, + 483, + 230 + ], + "blocks": [ + { + "bbox": [ + 111, + 71, + 483, + 230 + ], + "lines": [ + { + "bbox": [ + 111, + 71, + 483, + 230 + ], + "spans": [ + { + "bbox": [ + 111, + 71, + 483, + 230 + ], + "type": "image", + "image_path": "113af72237b96da2e5c4dbc25fd3f98375e9d2f01344258c80c376e07ffa2f6e.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 239, + 547, + 327 + ], + "lines": [ + { + "bbox": [ + 46, + 239, + 547, + 327 + ], + "spans": [ + { + "bbox": [ + 46, + 239, + 547, + 327 + ], + "type": "text", + "content": "Figure 1. An overview of our CR-FIQA training paradigm. We propose to simultaneously learn to optimize the class center " + }, + { + "bbox": [ + 46, + 239, + 547, + 327 + ], + "type": "inline_equation", + "content": "(\\mathcal{L}_{Arc})" + }, + { + "bbox": [ + 46, + 239, + 547, + 327 + ], + "type": "text", + "content": ", while learning to predict an internal network observation i.e. the allocation of the feature representation of sample " + }, + { + "bbox": [ + 46, + 239, + 547, + 327 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 46, + 239, + 547, + 327 + ], + "type": "text", + "content": " in feature space, with respect to its class center " + }, + { + "bbox": [ + 46, + 239, + 547, + 327 + ], + "type": "inline_equation", + "content": "w_{1}" + }, + { + "bbox": [ + 46, + 239, + 547, + 327 + ], + "type": "text", + "content": " and nearest negative class center " + }, + { + "bbox": [ + 46, + 239, + 547, + 327 + ], + "type": "inline_equation", + "content": "w_{2}" + }, + { + "bbox": [ + 46, + 239, + 547, + 327 + ], + "type": "inline_equation", + "content": "(\\mathcal{L}_{\\mathcal{CR}})" + }, + { + "bbox": [ + 46, + 239, + 547, + 327 + ], + "type": "text", + "content": ". The figure in the red rectangle illustrates the angle between two samples " + }, + { + "bbox": [ + 46, + 239, + 547, + 327 + ], + "type": "inline_equation", + "content": "x_{1}" + }, + { + "bbox": [ + 46, + 239, + 547, + 327 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 239, + 547, + 327 + ], + "type": "inline_equation", + "content": "x_{2}" + }, + { + "bbox": [ + 46, + 239, + 547, + 327 + ], + "type": "text", + "content": " (belong to identity 1) and their class center " + }, + { + "bbox": [ + 46, + 239, + 547, + 327 + ], + "type": "inline_equation", + "content": "w_{1}" + }, + { + "bbox": [ + 46, + 239, + 547, + 327 + ], + "type": "text", + "content": ". The plot on the right of the figure shows the distribution of the cosine similarity between training samples and their class centers (CCS) and nearest negative class centers (NNCCS) obtained from ResNet-50 trained on CASIA-WebFace [39]. The example images on the top-right of this plot are of high CCS values and the ones on the top-left are of low CCS values (notice the correspondence to perceived quality). These samples are selected from CASIA-WebFace [39]. During the testing mode, the classification layer is removed and the output of the regression layer is used to predict the FIQA of testing samples." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 334, + 289, + 706 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 334, + 289, + 706 + ], + "spans": [ + { + "bbox": [ + 46, + 334, + 289, + 706 + ], + "type": "text", + "content": "erties of the FR model responses to face samples to estimate the sample quality without explicitly learning a typical supervised regression that requires quality labels [28, 34, 36]. In the first category, the innovation focused on creating the FIQ labels for training. These quality labels included human-labeled quality labels [1], the FR genuine comparison score between a sample and an ICAO [18] compliant sample [14, 15], the FR comparison score involving the labeled sample (assumed to have the lower quality in the comparison pair) [38], and the Wasserstein distance between a randomly selected genuine and imposter FR comparisons with the labeled sample [32]. These solutions generally trained a regression network to predict the quality label, using both, trained-from-scratch networks in some cases [38], and pre-trained FR networks in other cases [14, 15, 32]. A slightly different approach, however also based on learning from labels, focuses on learning to predict the sample quality as a rank [22] based on FR performance-based training rank labels of a set of databases [4]. In the second category, the innovation was rather focused on linking face embedding properties under certain scenarios to the FIQ, without the explicit need for quality-labeled data. In [36], the assessed sample is passed through an FR network multiple times, each with a different random dropout pattern. The robustness of the resulting embeddings, represented by the sigmoid of the negative mean of the Euclidean distances between the embeddings, is considered the FIQ score. In [28], the FIQ score is calculated as the magnitude of the sample embedding. This is based on training the FR model using a loss that adapts the penalty margin loss based on this magnitude, and thus links the closeness of a sample to its class" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 304, + 334, + 547, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 334, + 547, + 441 + ], + "spans": [ + { + "bbox": [ + 304, + 334, + 547, + 441 + ], + "type": "text", + "content": "center to the unnormalized embedding magnitude. While in [34], the solution produces both, an FR embedding and a gaussian variance (uncertainty) vector, from a face sample. The inverse of harmonic mean of the uncertainty vector is considered as the FIQ score. Our CR-FIQA learns a regression problem to estimate the FIQ score, however, unlike previous works, without relying on preset labels, but rather learn a dynamic internal network observations (during training) that point out sample classifiability." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 306, + 447, + 373, + 460 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 447, + 373, + 460 + ], + "spans": [ + { + "bbox": [ + 306, + 447, + 373, + 460 + ], + "type": "text", + "content": "3. Approach" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 462, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 462, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 462, + 547, + 715 + ], + "type": "text", + "content": "This section presents our proposed Certainty Ratio Face Image Quality Assessment (CR-FIQA) approach, which inspects internal network observations to learn to predict the sample relative classifiability. This classifiability prediction is then used to estimate the FIQ. An overview of the proposed CR-FIQA approach is presented in Figure 1. During the training phase of an FR model, the model can conveniently push the high-quality samples close to their class center and relatively far from other class centers. Conversely, the FR is not able to push, to the same degree, low-quality samples to their class center, and thus they will remain relatively farther from their class center than the high-quality ones. Based on this assumption, we theorize our approach by stating that the properties that cause a face sample to lay relatively closer to its class center during training are the ones that make it a high-quality sample, and vice versa. Therefore, learning to predict such properties in any given sample would lead to learning to assess this sample quality. To learn to perform such assessment, our training paradigm targets learning internal network observations that evolve during the FR training phase, where these observa" + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5837" + } + ] + } + ], + "index": 6 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 156 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 156 + ], + "type": "text", + "content": "tions act as a training objective. The predictions of such training paradigm can be simply stated as answering a question: if a given sample was hypothetically part of the FR model training (which it is not), how relatively close would it be to its class center? Answering this question would give us an indication of this sample quality as will be shown in detail in this paper." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 156, + 286, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 156, + 286, + 240 + ], + "spans": [ + { + "bbox": [ + 46, + 156, + 286, + 240 + ], + "type": "text", + "content": "In the rest of this section, We formalize and empirically rationalize our proposed CR-FIQA approach and its components. To do that, we start by shortly revisiting angular margin penalty-based softmax loss utilized to optimize the class centers of the FR model. Then present a detailed description of our proposed CR-FIQA concept and the associated training paradigm." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 243, + 286, + 255 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 243, + 286, + 255 + ], + "spans": [ + { + "bbox": [ + 47, + 243, + 286, + 255 + ], + "type": "text", + "content": "3.1. Revisiting Margin Penalty-based Softmax Loss" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 258, + 286, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 258, + 286, + 413 + ], + "spans": [ + { + "bbox": [ + 46, + 258, + 286, + 413 + ], + "type": "text", + "content": "Angular margin penalty-based softmax is a widely used loss function for training FR models [3,5,17,28]. It extends over softmax loss by deploying angular penalty margin on the angle between the deep features and their corresponding weights. Margin penalty-based softmax loss aims to push the decision boundary of softmax, and thus enhance intraclass compactness and inter-class discrepancy. From this family of loss functions, this work utilizes ArcFace loss [5] to optimize the distance between the training samples and their class center. Our choice of ArcFace loss is based on the SOTA performance achieved by ResNet-100 network trained with ArcFace on mainstream benchmarks [5]. Formally, ArcFace loss is defined as follows:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 415, + 286, + 467 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 415, + 286, + 467 + ], + "spans": [ + { + "bbox": [ + 48, + 415, + 286, + 467 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {A r c} = \\frac {1}{N} \\sum_ {i \\in N} - \\log \\frac {e ^ {s \\left(\\cos \\left(\\theta_ {y _ {i}} + m\\right)\\right)}}{e ^ {s \\left(\\cos \\left(\\theta_ {y _ {i}} + m\\right)\\right)} + \\sum_ {j = 1 , j \\neq y _ {i}} ^ {C} e ^ {s \\left(\\cos \\left(\\theta_ {j}\\right)\\right)}}, \\tag {1}", + "image_path": "a784e9f2548ae13420df539248175c8a6f520f3457c0ba6009708128cb744d7a.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 468, + 286, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 468, + 286, + 635 + ], + "spans": [ + { + "bbox": [ + 46, + 468, + 286, + 635 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 468, + 286, + 635 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 468, + 286, + 635 + ], + "type": "text", + "content": " is the batch size, " + }, + { + "bbox": [ + 46, + 468, + 286, + 635 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 46, + 468, + 286, + 635 + ], + "type": "text", + "content": " is the number of classes (identities), " + }, + { + "bbox": [ + 46, + 468, + 286, + 635 + ], + "type": "inline_equation", + "content": "y_{i}" + }, + { + "bbox": [ + 46, + 468, + 286, + 635 + ], + "type": "text", + "content": " is the class label of sample " + }, + { + "bbox": [ + 46, + 468, + 286, + 635 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 468, + 286, + 635 + ], + "type": "text", + "content": " (in range " + }, + { + "bbox": [ + 46, + 468, + 286, + 635 + ], + "type": "inline_equation", + "content": "[1,C])" + }, + { + "bbox": [ + 46, + 468, + 286, + 635 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 468, + 286, + 635 + ], + "type": "inline_equation", + "content": "\\theta_{y_i}" + }, + { + "bbox": [ + 46, + 468, + 286, + 635 + ], + "type": "text", + "content": " is the angle between the features " + }, + { + "bbox": [ + 46, + 468, + 286, + 635 + ], + "type": "inline_equation", + "content": "x_{i}" + }, + { + "bbox": [ + 46, + 468, + 286, + 635 + ], + "type": "text", + "content": " and the " + }, + { + "bbox": [ + 46, + 468, + 286, + 635 + ], + "type": "inline_equation", + "content": "y_{i}" + }, + { + "bbox": [ + 46, + 468, + 286, + 635 + ], + "type": "text", + "content": "-th class center " + }, + { + "bbox": [ + 46, + 468, + 286, + 635 + ], + "type": "inline_equation", + "content": "w_{y_i}" + }, + { + "bbox": [ + 46, + 468, + 286, + 635 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 46, + 468, + 286, + 635 + ], + "type": "inline_equation", + "content": "x_{i} \\in R^{d}" + }, + { + "bbox": [ + 46, + 468, + 286, + 635 + ], + "type": "text", + "content": " is the deep feature embedding of the last fully connected layer of size " + }, + { + "bbox": [ + 46, + 468, + 286, + 635 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 46, + 468, + 286, + 635 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 46, + 468, + 286, + 635 + ], + "type": "inline_equation", + "content": "w_{y_i}" + }, + { + "bbox": [ + 46, + 468, + 286, + 635 + ], + "type": "text", + "content": " is the " + }, + { + "bbox": [ + 46, + 468, + 286, + 635 + ], + "type": "inline_equation", + "content": "y_{i}" + }, + { + "bbox": [ + 46, + 468, + 286, + 635 + ], + "type": "text", + "content": "-th column of weights " + }, + { + "bbox": [ + 46, + 468, + 286, + 635 + ], + "type": "inline_equation", + "content": "W \\in R_C^d" + }, + { + "bbox": [ + 46, + 468, + 286, + 635 + ], + "type": "text", + "content": " of the classification layer. " + }, + { + "bbox": [ + 46, + 468, + 286, + 635 + ], + "type": "inline_equation", + "content": "\\theta_{y_i}" + }, + { + "bbox": [ + 46, + 468, + 286, + 635 + ], + "type": "text", + "content": " is defined as " + }, + { + "bbox": [ + 46, + 468, + 286, + 635 + ], + "type": "inline_equation", + "content": "x_{i} w_{y_{i}}^{T} = \\| x_{i} \\| \\| w_{y_{i}} \\| \\cos (\\theta_{y_{i}})" + }, + { + "bbox": [ + 46, + 468, + 286, + 635 + ], + "type": "text", + "content": " [24]. The weights and the feature norms are fixed to " + }, + { + "bbox": [ + 46, + 468, + 286, + 635 + ], + "type": "inline_equation", + "content": "\\| w_{y_i} \\| = 1" + }, + { + "bbox": [ + 46, + 468, + 286, + 635 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 468, + 286, + 635 + ], + "type": "inline_equation", + "content": "\\| x_{i} \\| = 1" + }, + { + "bbox": [ + 46, + 468, + 286, + 635 + ], + "type": "text", + "content": ", respectively, using " + }, + { + "bbox": [ + 46, + 468, + 286, + 635 + ], + "type": "inline_equation", + "content": "l_2" + }, + { + "bbox": [ + 46, + 468, + 286, + 635 + ], + "type": "text", + "content": " normalization as defined in [24, 37]. The decision boundary, in this case, depends on the angle cosine between " + }, + { + "bbox": [ + 46, + 468, + 286, + 635 + ], + "type": "inline_equation", + "content": "x_{i}" + }, + { + "bbox": [ + 46, + 468, + 286, + 635 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 468, + 286, + 635 + ], + "type": "inline_equation", + "content": "w_{y_i}" + }, + { + "bbox": [ + 46, + 468, + 286, + 635 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 46, + 468, + 286, + 635 + ], + "type": "inline_equation", + "content": "m > 0" + }, + { + "bbox": [ + 46, + 468, + 286, + 635 + ], + "type": "text", + "content": " is an additive angular margin proposed by ArcFace [5] to enhance the intra-class compactness and inter-class discrepancy. Lastly, " + }, + { + "bbox": [ + 46, + 468, + 286, + 635 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 46, + 468, + 286, + 635 + ], + "type": "text", + "content": " is the scaling parameter [37]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 638, + 143, + 651 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 638, + 143, + 651 + ], + "spans": [ + { + "bbox": [ + 47, + 638, + 143, + 651 + ], + "type": "text", + "content": "3.2. Certainty Ratio" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 653, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 286, + 713 + ], + "type": "text", + "content": "In this section, we formulate and empirically rationalize the main concepts that build our FIQA solution. We derive our Certainty Ratio (CR) to estimate the sample relative classifiability. Additionally, we experimentally illustrate the strong relationship between our CR measure and FIQ." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "text", + "content": "Certainty Ratio During the FR model training phase, the model is trained to enhance the separability between the classes (identities) by pushing each sample " + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "inline_equation", + "content": "x_{i}" + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "text", + "content": " to be close to its class center " + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "inline_equation", + "content": "w_{y_i}" + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "text", + "content": " and far from the other (negative) class centers " + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "inline_equation", + "content": "w_{j}, j \\neq y_{i}" + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "text", + "content": ". Based on this, we first define the Class Center Angular Similarity (CCS) as the proximity between " + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "inline_equation", + "content": "x_{i}" + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "text", + "content": " and its class center " + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "inline_equation", + "content": "w_{y_i}" + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "text", + "content": ", as follows:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 383, + 159, + 545, + 172 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 383, + 159, + 545, + 172 + ], + "spans": [ + { + "bbox": [ + 383, + 159, + 545, + 172 + ], + "type": "interline_equation", + "content": "C C S _ {x _ {i}} = \\cos \\left(\\theta_ {y _ {i}}\\right), \\tag {2}", + "image_path": "d220d236527fb50a1c1db9f6f7945c8bfd915484f3f2c94002145c2af77227dc.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 182, + 545, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 182, + 545, + 266 + ], + "spans": [ + { + "bbox": [ + 304, + 182, + 545, + 266 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 182, + 545, + 266 + ], + "type": "inline_equation", + "content": "\\theta_{y_i}" + }, + { + "bbox": [ + 304, + 182, + 545, + 266 + ], + "type": "text", + "content": " is the angle between " + }, + { + "bbox": [ + 304, + 182, + 545, + 266 + ], + "type": "inline_equation", + "content": "x_i" + }, + { + "bbox": [ + 304, + 182, + 545, + 266 + ], + "type": "text", + "content": " and its class center " + }, + { + "bbox": [ + 304, + 182, + 545, + 266 + ], + "type": "inline_equation", + "content": "w_{y_i}" + }, + { + "bbox": [ + 304, + 182, + 545, + 266 + ], + "type": "text", + "content": ", where the weights of the last fully connected layer of the FR model trained with softmax loss are considered as the centers for each class [5,25]. Then, we define the Closest Nearest Negative Class Center Angular Similarity (NNCCS) as proximity between " + }, + { + "bbox": [ + 304, + 182, + 545, + 266 + ], + "type": "inline_equation", + "content": "x_i" + }, + { + "bbox": [ + 304, + 182, + 545, + 266 + ], + "type": "text", + "content": " and the nearest negative class center " + }, + { + "bbox": [ + 304, + 182, + 545, + 266 + ], + "type": "inline_equation", + "content": "w_j" + }, + { + "bbox": [ + 304, + 182, + 545, + 266 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 182, + 545, + 266 + ], + "type": "inline_equation", + "content": "j \\neq y_i" + }, + { + "bbox": [ + 304, + 182, + 545, + 266 + ], + "type": "text", + "content": ". Formally, NNCCS is defined as follows:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 356, + 269, + 545, + 292 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 356, + 269, + 545, + 292 + ], + "spans": [ + { + "bbox": [ + 356, + 269, + 545, + 292 + ], + "type": "interline_equation", + "content": "N N C C S _ {x _ {i}} = \\max _ {j = 1, j \\neq y _ {i}} ^ {C} (\\cos \\left(\\theta_ {j}\\right)), \\tag {3}", + "image_path": "067dd79c65e242d85c3987dbb8884b3245479289de68f23585028481fbb677ac.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 300, + 545, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 300, + 545, + 430 + ], + "spans": [ + { + "bbox": [ + 304, + 300, + 545, + 430 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 300, + 545, + 430 + ], + "type": "inline_equation", + "content": "\\theta_{j}" + }, + { + "bbox": [ + 304, + 300, + 545, + 430 + ], + "type": "text", + "content": " is the angle between " + }, + { + "bbox": [ + 304, + 300, + 545, + 430 + ], + "type": "inline_equation", + "content": "x_{i}" + }, + { + "bbox": [ + 304, + 300, + 545, + 430 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 300, + 545, + 430 + ], + "type": "inline_equation", + "content": "w_{j}" + }, + { + "bbox": [ + 304, + 300, + 545, + 430 + ], + "type": "text", + "content": ". As we theorize, when the FR model converges, the high-quality samples are pushed closer to their class centers (high CCS) in relation to their distance to neighbouring negative class centers (low NNCCS). However, low-quality samples can not be pushed as close to their class centers. A sample able to achieve high CCS with respect to NNCCS is a sample easily correctly classified during training, and thus is relatively highly classifiable. We thus measure this relative classifiability by the ratio of CCS to NNCS, which we note as the Certainty Ratio (CR), as follows:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 358, + 432, + 545, + 459 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 358, + 432, + 545, + 459 + ], + "spans": [ + { + "bbox": [ + 358, + 432, + 545, + 459 + ], + "type": "interline_equation", + "content": "C R _ {x _ {i}} = \\frac {C C S _ {x _ {i}}}{N N C C S _ {x _ {i}} + (1 + \\epsilon)}, \\tag {4}", + "image_path": "51971d7f9f7e0a82340d30deac3999d55b6d905823351cfd5b6d66ffa6d6e112.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 466, + 545, + 622 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 466, + 545, + 622 + ], + "spans": [ + { + "bbox": [ + 304, + 466, + 545, + 622 + ], + "type": "text", + "content": "where the " + }, + { + "bbox": [ + 304, + 466, + 545, + 622 + ], + "type": "inline_equation", + "content": "1 + \\epsilon" + }, + { + "bbox": [ + 304, + 466, + 545, + 622 + ], + "type": "text", + "content": " term is added to insure a positive above zero denominator, i.e. shift the NNCCS value range from " + }, + { + "bbox": [ + 304, + 466, + 545, + 622 + ], + "type": "inline_equation", + "content": "[-1, +1]" + }, + { + "bbox": [ + 304, + 466, + 545, + 622 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 304, + 466, + 545, + 622 + ], + "type": "inline_equation", + "content": "[\\epsilon, 2 + \\epsilon]" + }, + { + "bbox": [ + 304, + 466, + 545, + 622 + ], + "type": "text", + "content": ". This ensures that the CR of a sample with a lower NNCCS is relatively higher than a sample with a higher NNCCS, given the same CCS, i.e. NNCCS regulates the CCS value in relation to neighbouring classes. The " + }, + { + "bbox": [ + 304, + 466, + 545, + 622 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 304, + 466, + 545, + 622 + ], + "type": "text", + "content": " is set to " + }, + { + "bbox": [ + 304, + 466, + 545, + 622 + ], + "type": "inline_equation", + "content": "1e - 9" + }, + { + "bbox": [ + 304, + 466, + 545, + 622 + ], + "type": "text", + "content": " in our experiments. The optimal CR is obtained when the CCS is approaching the maximum cosine similarity value (+1) and the NNCCS is approaching the minimum cosine similarity value (-1), i.e. the training sample is capable of being pushed to its class center, and far away from the closest negative class center, and thus it is highly classifiable." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 626, + 488, + 639 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 626, + 488, + 639 + ], + "spans": [ + { + "bbox": [ + 306, + 626, + 488, + 639 + ], + "type": "text", + "content": "3.3. Relation between the CR and FIQ" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "type": "text", + "content": "Here, we empirically prove the theorized relationship between the CR and FIQ (defined earlier as image utility). Namely, we want to answer: if the CR values achieved by training samples of an FR model were used as FIQ, would they behave as expected from an optimal FIQ? If yes, then the face image properties leading to high/low CR" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5838" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 289, + 360 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 289, + 360 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 289, + 360 + ], + "type": "text", + "content": "do also theoretically lead to high/low FIQ. To answer this question we conducted an experiment on a ResNet-50 [13] FR model trained on CASIA-WebFace [39] with ArcFace loss [5] (noted as R50(CASIA)). Specifically, we calculate the CR, CCS, and NNCCS values from the trained model for all samples in the training dataset (0.5M images of 10K identities). An insight on the resulting CCS and NNCCS values (CR being a derivative measure) is given as value distributions in Figure 1, showing that these measures vary between different samples. Furthermore, based on the calculated scores, we plot Error vs. Reject Curves (ERC) (described in Section 4) to demonstrate the relationship between the CR, as an FIQ measure, and FR performance. To calculate the FR performance in the ERC curve, we extract the feature embedding of CASIA-WebFace [39] using a ResNet-100 model [13] trained on MS1M-V2 [5, 12] with ArcFace (noted as R100(MS1M-V2)). We utilize a different model (trained on a different database) to extract the embedding (R100(MS1M-V2)) than the one used to calculate CR, CCS, and NNCCS (R50(CASIA)) to provide fair evaluation where the FR performance is evaluated on unseen data. Then, we perform " + }, + { + "bbox": [ + 47, + 72, + 289, + 360 + ], + "type": "inline_equation", + "content": "n : n" + }, + { + "bbox": [ + 47, + 72, + 289, + 360 + ], + "type": "text", + "content": " comparisons between all samples of CASIA-WebFace using feature embedding obtained from the R100(MS1M-V2)." + } + ] + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 59, + 367, + 159, + 424 + ], + "blocks": [ + { + "bbox": [ + 59, + 367, + 159, + 424 + ], + "lines": [ + { + "bbox": [ + 59, + 367, + 159, + 424 + ], + "spans": [ + { + "bbox": [ + 59, + 367, + 159, + 424 + ], + "type": "image", + "image_path": "f127b852bbb10049bcd2c891ff2f7b7cc4b4eb23f68e608b52856ec8d7617c78.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 106, + 426, + 117, + 434 + ], + "lines": [ + { + "bbox": [ + 106, + 426, + 117, + 434 + ], + "spans": [ + { + "bbox": [ + 106, + 426, + 117, + 434 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 167, + 367, + 267, + 423 + ], + "blocks": [ + { + "bbox": [ + 167, + 367, + 267, + 423 + ], + "lines": [ + { + "bbox": [ + 167, + 367, + 267, + 423 + ], + "spans": [ + { + "bbox": [ + 167, + 367, + 267, + 423 + ], + "type": "image", + "image_path": "bc38a9ae33aced9bb1c910c20058afdc20a61b75cf6cbfab187e61b6df40bf26.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 215, + 426, + 227, + 434 + ], + "lines": [ + { + "bbox": [ + 215, + 426, + 227, + 434 + ], + "spans": [ + { + "bbox": [ + 215, + 426, + 227, + 434 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 436, + 287, + 491 + ], + "lines": [ + { + "bbox": [ + 46, + 436, + 287, + 491 + ], + "spans": [ + { + "bbox": [ + 46, + 436, + 287, + 491 + ], + "type": "text", + "content": "Figure 2. ERCs showing the verification performance as False None Match Rate (FNMR) at False Match Rate (FMR) of 1e-3 (2a) and 1e-4 (2b) with CCS, NNCCS and CR as FIQ vs. rejection ratio. This ERC plots show the effectiveness of rejecting samples with the lowest CCS and CR on the performance" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 497, + 287, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 497, + 287, + 545 + ], + "spans": [ + { + "bbox": [ + 46, + 497, + 287, + 545 + ], + "type": "text", + "content": "Figure 2 presents the ERC of CR, CCS, and NNCCS experimentally used as FIQ. An FIQ measure would cause the ERC to drop as rapidly as possible when rejecting a larger fraction of low-quality samples (moving to the right)." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 545, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 545, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 545, + 288, + 713 + ], + "type": "text", + "content": "It can be clearly noticed in Figure 2 that the CCS and CR do behave as we would expect from a good performing FIQ, as the verification error value drops rapidly when rejecting low quality (low CCS and CR) samples. It can be also observed that the CR does that more steadily when compared to CCS. This points out that adding the scaling term NNCCS in CR calculation can enhance the representation of the CCS as an FIQ measure, which will be clearer later when we experimentally evaluate our CR-FIQA approach in Section 5. As expected, the NNCCS measure by itself does not strongly act as an FIQ measure would, demonstrated by the relatively flat ERC in Figure 2, as it only considers the distance to the nearest negative class. This empirical evaluation does provide a confirming answer to the previ" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "text", + "content": "ously stated question by affirming that the CR does act as expected from an FIQ measure and thus, theoretically, one can strongly link the image properties that cause high/low CR in the FR training data to these causing high/low FIQ." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 126, + 510, + 139 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 126, + 510, + 139 + ], + "spans": [ + { + "bbox": [ + 305, + 126, + 510, + 139 + ], + "type": "text", + "content": "3.4. Quality Estimation Training Paradigm" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 142, + 545, + 668 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 142, + 545, + 668 + ], + "spans": [ + { + "bbox": [ + 304, + 142, + 545, + 668 + ], + "type": "text", + "content": "In the previous section, we proved that the CR does behave as an FIQ would, and thus, it can also relate to image properties that dictate FIQ. However, the CR measure is only observable for samples in the FR training dataset, where the class centers are known. In a real case scenario, the FIQ measure should be assessed to any single image, i.e. unseen evaluation data. Considering this, and in an effort to predict what the CR value would be for a given sample if hypothetically it was part of the FR training, we propose to simultaneously learn to predict the CR from the training dataset while optimizing the class centers (typical FR training) during the training phase, i.e. the CR-FIQA model. To enable this, we add a single regression layer to the FR model. The input of the regression layer is a feature embedding " + }, + { + "bbox": [ + 304, + 142, + 545, + 668 + ], + "type": "inline_equation", + "content": "x_{i}" + }, + { + "bbox": [ + 304, + 142, + 545, + 668 + ], + "type": "text", + "content": " and the output is an estimation of the CR. The output of this regression layer is used later to predict the FIQ score of the unseen sample, e.g. from the evaluation dataset. Thus, we capture the properties that make the CR high/low to predict the FIQ of any given sample. Towards this goal, during the training phase, the model (in Figure 1) has two learning objectives: a) It is trained to optimize the distance between the samples and the class centers using ArcFace loss defined in Equation 1. b) It is trained to predict the internal network observation, CR, using Smooth L1-Loss [9] applied between the output of the regression layer (P) and the CR calculated as in Equation 5. Smooth L1-loss can be interpreted as a combination of L1 and L2-losses by defining a threshold " + }, + { + "bbox": [ + 304, + 142, + 545, + 668 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 304, + 142, + 545, + 668 + ], + "type": "text", + "content": " that changes between them [9]. Our choice for smooth-l1 loss is based on: 1) It is less sensitive to outliers than l2. The derivative of L2 loss increases when the difference between the prediction and ground-truth label is increased, making the derivative of loss values large at the early stage of the training, leading to unstable training. Additionally, L2 loss can easily generate gradient explosion [9] when there are outliers in the training data. 2) L1 loss can lead to stable training. However, the absolute values of the difference between prediction and ground truth are small, especially in the later stage of the training. Therefore, the model accuracy can hardly be improved at a later stage of the training as the loss function will fluctuate around a stable value. Combining L1 and L2 as in Smooth L1-loss avoids gradient explosion, which might be caused by L2 and facilitates better convergence than L1. The loss leading to the second objective is then given as:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 671, + 547, + 713 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 671, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 671, + 547, + 713 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {C R} = \\frac {1}{N} \\sum_ {i \\in N} \\left\\{ \\begin{array}{l l} \\frac {0 . 5 \\times \\left(C R _ {x _ {i}} - P _ {i}\\right) ^ {2}}{\\beta} & \\text {i f} \\left| C R _ {x _ {i}} - P _ {i} \\right| < \\beta \\\\ \\left| C R _ {x _ {i}} - P _ {i} \\right| - 0. 5 \\times \\beta & \\text {o t h e r w i s e} \\end{array} \\right. \\tag {5}", + "image_path": "18bb6ab24e8853e22c84f4a7ca4a621bfeeb476627fb337496e7a14e2fe8f036.jpg" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5839" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 55, + 70, + 175, + 171 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 175, + 171 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 175, + 171 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 175, + 171 + ], + "type": "image", + "image_path": "c14e758581959881b3fd5bc9341c6344a0bd6d5f4843358172bff43491fe7809.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 109, + 174, + 120, + 182 + ], + "lines": [ + { + "bbox": [ + 109, + 174, + 120, + 182 + ], + "spans": [ + { + "bbox": [ + 109, + 174, + 120, + 182 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 177, + 70, + 296, + 170 + ], + "blocks": [ + { + "bbox": [ + 177, + 70, + 296, + 170 + ], + "lines": [ + { + "bbox": [ + 177, + 70, + 296, + 170 + ], + "spans": [ + { + "bbox": [ + 177, + 70, + 296, + 170 + ], + "type": "image", + "image_path": "c0a0ffe667186e01fc06bdb62efb807dc28b8ac45759bbbdd4e3eed889c96c18.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 230, + 174, + 242, + 182 + ], + "lines": [ + { + "bbox": [ + 230, + 174, + 242, + 182 + ], + "spans": [ + { + "bbox": [ + 230, + 174, + 242, + 182 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 298, + 70, + 417, + 171 + ], + "blocks": [ + { + "bbox": [ + 298, + 70, + 417, + 171 + ], + "lines": [ + { + "bbox": [ + 298, + 70, + 417, + 171 + ], + "spans": [ + { + "bbox": [ + 298, + 70, + 417, + 171 + ], + "type": "image", + "image_path": "7ca45a0c251b02023d85b195cf69060a7183d277e6ce53f7a6ceffe1c590a47e.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 352, + 174, + 363, + 182 + ], + "lines": [ + { + "bbox": [ + 352, + 174, + 363, + 182 + ], + "spans": [ + { + "bbox": [ + 352, + 174, + 363, + 182 + ], + "type": "text", + "content": "(c)" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 184, + 547, + 218 + ], + "lines": [ + { + "bbox": [ + 46, + 184, + 547, + 218 + ], + "spans": [ + { + "bbox": [ + 46, + 184, + 547, + 218 + ], + "type": "text", + "content": "Figure 3. ERC comparison between CR-FIQA(S), CCS-FIQA(S), CR-FIQA(S) (On top) and CCS-FIQA(S) (On top). The plots show the effect of rejecting samples of lowest quality, on the verification error (FNMR at FMR1e-3). CR-FIQA(S) and CCS-FIQA(S) outperformed the on-top solutions, and CR-FIQA(S) performs generally better than CCS-FIQA(S) (curve decays faster with more rejected samples)" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 419, + 72, + 536, + 171 + ], + "blocks": [ + { + "bbox": [ + 419, + 72, + 536, + 171 + ], + "lines": [ + { + "bbox": [ + 419, + 72, + 536, + 171 + ], + "spans": [ + { + "bbox": [ + 419, + 72, + 536, + 171 + ], + "type": "image", + "image_path": "f78a781dae2143838300b454c28e485eb6a63d039a537a4192f615f032204c64.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 472, + 174, + 483, + 182 + ], + "lines": [ + { + "bbox": [ + 472, + 174, + 483, + 182 + ], + "spans": [ + { + "bbox": [ + 472, + 174, + 483, + 182 + ], + "type": "text", + "content": "(d)" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 224, + 288, + 248 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 224, + 288, + 248 + ], + "spans": [ + { + "bbox": [ + 46, + 224, + 288, + 248 + ], + "type": "text", + "content": "The final loss combining both objectives for training our CR-FIQA model is defined as follows:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 118, + 251, + 287, + 264 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 118, + 251, + 287, + 264 + ], + "spans": [ + { + "bbox": [ + 118, + 251, + 287, + 264 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\mathcal {L} _ {A r c} + \\lambda \\times \\mathcal {L} _ {C R}, \\tag {6}", + "image_path": "676b78f7c524457aea705d6c576de344e9a39de49dad73fbf7af101afd0f654e.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 273, + 287, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 273, + 287, + 357 + ], + "spans": [ + { + "bbox": [ + 46, + 273, + 287, + 357 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 273, + 287, + 357 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 46, + 273, + 287, + 357 + ], + "type": "text", + "content": " is a hyper-parameter used to control the balance between the two losses. At the beginning of model training, the value range of " + }, + { + "bbox": [ + 46, + 273, + 287, + 357 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{CR}" + }, + { + "bbox": [ + 46, + 273, + 287, + 357 + ], + "type": "text", + "content": " is very small (" + }, + { + "bbox": [ + 46, + 273, + 287, + 357 + ], + "type": "inline_equation", + "content": "\\leq 2" + }, + { + "bbox": [ + 46, + 273, + 287, + 357 + ], + "type": "text", + "content": ") in comparison to " + }, + { + "bbox": [ + 46, + 273, + 287, + 357 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{Arc} (\\sim 45)" + }, + { + "bbox": [ + 46, + 273, + 287, + 357 + ], + "type": "text", + "content": ". Setting " + }, + { + "bbox": [ + 46, + 273, + 287, + 357 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 46, + 273, + 287, + 357 + ], + "type": "text", + "content": " to a small value, the model will only focus on " + }, + { + "bbox": [ + 46, + 273, + 287, + 357 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{Arc}" + }, + { + "bbox": [ + 46, + 273, + 287, + 357 + ], + "type": "text", + "content": ". Besides, setting " + }, + { + "bbox": [ + 46, + 273, + 287, + 357 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 46, + 273, + 287, + 357 + ], + "type": "text", + "content": " to a large value, i.e. " + }, + { + "bbox": [ + 46, + 273, + 287, + 357 + ], + "type": "inline_equation", + "content": ">10" + }, + { + "bbox": [ + 46, + 273, + 287, + 357 + ], + "type": "text", + "content": ", we observed that the model did not converge. Therefore, we set " + }, + { + "bbox": [ + 46, + 273, + 287, + 357 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 46, + 273, + 287, + 357 + ], + "type": "text", + "content": " to 10 in all the experiments in this paper." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 47, + 361, + 165, + 375 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 361, + 165, + 375 + ], + "spans": [ + { + "bbox": [ + 47, + 361, + 165, + 375 + ], + "type": "text", + "content": "4. Experimental Setup" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 46, + 379, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 379, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 379, + 288, + 713 + ], + "type": "text", + "content": "Implementation Details We demonstrate our proposed CR-FIQA under two protocols (small and large) based on the training dataset and the training model architecture. We utilize widely used architectures in the SOTA FR solutions, ResNet100 and ResNet50 [13], both modified as described in Section 3.4. For the small protocol, we utilize ResNet50 and the CASIA-WebFace [39] training data (noted as CR-FIQA(S)) and for the large protocol, we utilize ResNet100 and the MS1MV2 [5, 12] training data (noted as CR-FIQA(L)). The MS1MV2 is a refined version of the MS-Celeb-1M [12] by [5] containing 5.8M images of 85K identities. The CASIA-WebFace contains 0.5m images of 10K identities [39]. We follow the ArcFace training setting [5] to set the scale parameter " + }, + { + "bbox": [ + 46, + 379, + 288, + 713 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 46, + 379, + 288, + 713 + ], + "type": "text", + "content": " to 64 and the margin " + }, + { + "bbox": [ + 46, + 379, + 288, + 713 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 46, + 379, + 288, + 713 + ], + "type": "text", + "content": " to 0.5. We set the mini-batch size to 512. All models are trained with Stochastic Gradient Descent (SGD) optimizer with an initial learning rate of 1e-1. During the training, we use random horizontal flipping with a probability of 0.5 for data augmentation. We set the momentum to 0.9 and the weight decay to 5e-4. For CR-FIQA(S), the learning rate is divided by 10 at 20K and at 28K training iterations, following [5]. The training is stopped after 32K iterations. For CR-FIQA(L), the learning rate is divided by 10 at 100K and 160K training iterations, following [5]. The training is stopped after 180K iterations. All the images in evaluation and training datasets are aligned and cropped to " + }, + { + "bbox": [ + 46, + 379, + 288, + 713 + ], + "type": "inline_equation", + "content": "112 \\times 112" + }, + { + "bbox": [ + 46, + 379, + 288, + 713 + ], + "type": "text", + "content": ", as described in [5]. All the training and testing images are normalized to have pixel values between -1 and 1. Both" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 224, + 529, + 237 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 224, + 529, + 237 + ], + "spans": [ + { + "bbox": [ + 305, + 224, + 529, + 237 + ], + "type": "text", + "content": "models are trained using the loss defined in Equation 6." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 239, + 545, + 347 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 239, + 545, + 347 + ], + "spans": [ + { + "bbox": [ + 304, + 239, + 545, + 347 + ], + "type": "text", + "content": "Evaluation Benchmarks We reported the achieved results on eight different benchmarks: Labeled Faces in the Wild (LFW) [16], AgeDB-30 [31], Celebrities in Frontal-Profile in the Wild (CFP-FP) [33], Cross-age LFW (CALFW) [41], Adience [6], Cross-Pose LFW (CPLFW) [40], Cross-Quality LFW (XQLFW) [23], IARPA Janus Benchmark-C (IJB-C) [27]. These benchmarks are chosen to provide a wide comparison to SOTA FIQA algorithms and give an insight into the CR-FIQA generalizability." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 349, + 546, + 505 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 349, + 546, + 505 + ], + "spans": [ + { + "bbox": [ + 304, + 349, + 546, + 505 + ], + "type": "text", + "content": "Evaluation Metric We evaluate the FIQA by plotting ERCs [11]. The ERC is a widely used representation of the FIQA performance [10, 11] by demonstrating the effect of rejecting a fraction face images, of the lowest quality, on face verification performance in terms of False None Match Rate [21] (FNMR) at a specific threshold calculated at fixed False Match Rate [21] (FMR). The ERC curves for all benchmarks are plotted at two fixed FMRs, 1e-3 (as recommended for border control operations by Frontex [7]) and 1e-4 (the latter is provided in the supplementary material). We also report the Area under the Curve (AUC) of the ERC, to provide a quantitative aggregate measure of verification performance across all rejection ratios." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 507, + 547, + 627 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 507, + 547, + 627 + ], + "spans": [ + { + "bbox": [ + 304, + 507, + 547, + 627 + ], + "type": "text", + "content": "Additionally, motivated by evaluating the FIQ as a weighting term for face embedding [32, 34], we follow the IJB-C 1:1 mixed verification benchmark [27] by weighting the frames such that all frames belonging to the same subject within a video have a combined weight equal to a single still image as described in IJB-C benchmark [27]. We do that by using the CR-FIQA quality scores as well as all SOTA methods. We report the verification performance of IJB-C as true acceptance rates (TAR) at false acceptance rates (FAR) of 1e-4, 1e-5, and 1e-6, as defined in [27]." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 630, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 630, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 630, + 547, + 713 + ], + "type": "text", + "content": "Face Recognition Models We utilize four different SOTA FR models to report the verification performance at different quality rejection rate to inspect the generalizability of FIQA over FR solutions. The FR models are ArcFace [5], ElasticFace (ElasticFace-Arc) [3], MagFace [28], and CurricularFace [17]. All models process " + }, + { + "bbox": [ + 304, + 630, + 547, + 713 + ], + "type": "inline_equation", + "content": "112 \\times 112" + }, + { + "bbox": [ + 304, + 630, + 547, + 713 + ], + "type": "text", + "content": " aligned and cropped image to produce 512-D feature em" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5840" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 204 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 204 + ], + "type": "text", + "content": "bedding. We used the officially released pretrained ResNet-100 models trained on MS1MV2 released by the four FR solutions. Although, the presented solution in this paper does not aim, and is not presented as, a solution to extract face embeddings, but rather an FIQA solution, we opted to evaluate CR-FIQA(L) backbone as a FR model on mainstream FR benchmarks for sake of providing complete experiment evaluation and probe the possibility of simultaneously using it as both FIQA and FR model. The evaluation results of CR-FIQA(L) backbone as a FR model are provided in the supplementary material." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 204, + 289, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 204, + 289, + 300 + ], + "spans": [ + { + "bbox": [ + 46, + 204, + 289, + 300 + ], + "type": "text", + "content": "Baseline We compare our CR-FIQA approach with nine quality assessment methods. Three are general IQA methods that have been proven in [8] to correlate well to face utility i.e. BRISQUE [29], RankIQA [26], and DeepIQA [2], and six are SOTA face-specific FIQA methods, namely RankIQ [4], PFE [34], SER-FIQ [36], FaceQnet (v1 [14]) [14, 15], MagFace [28], and SDD-FIQA [32], all as officially released in the respective works." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 304, + 148, + 316 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 304, + 148, + 316 + ], + "spans": [ + { + "bbox": [ + 47, + 304, + 148, + 316 + ], + "type": "text", + "content": "5. Ablation Studies" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 319, + 287, + 342 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 319, + 287, + 342 + ], + "spans": [ + { + "bbox": [ + 46, + 319, + 287, + 342 + ], + "type": "text", + "content": "This section provides experimental proof of the two main design choices in CR-FIQA." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 343, + 289, + 534 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 343, + 289, + 534 + ], + "spans": [ + { + "bbox": [ + 46, + 343, + 289, + 534 + ], + "type": "text", + "content": "Does CR-FIQA benefit from the NNCCS scaling term? To answer this, we conducted additional experiments using ResNet-50 model trained on CASIA-WebFace [39] using the experimental setup described in Section 4. This model is noted as CCS-FIAQ(S). The only difference from CR-FIAQ(S) is that the CCS-FIAQ(S) is trained to learn CCS (instead of CR) by replacing the " + }, + { + "bbox": [ + 46, + 343, + 289, + 534 + ], + "type": "inline_equation", + "content": "CR_{x_i}" + }, + { + "bbox": [ + 46, + 343, + 289, + 534 + ], + "type": "text", + "content": " in Equation 5 with " + }, + { + "bbox": [ + 46, + 343, + 289, + 534 + ], + "type": "inline_equation", + "content": "CCS_{x_i}" + }, + { + "bbox": [ + 46, + 343, + 289, + 534 + ], + "type": "text", + "content": ", thus neglecting the NNCCS scaling term in the equation. Figure 3 presents the ERCs along with AUC using CR-FIAQ(S) and CCS-FIAQ(S) on Adience, AgeDb-30, CALLFW and CRF-FP. The verification error, FNMR at FMR1e-3, is calculated using ArcFace FR model (described in Section 4). The ERCs and AUC values show that the reduction in the error is more evident for CR-FIAQ(S) than CCS-FIQA(S). Thus, adding the scaling term NNCCS in CR enhanced the performance of the FIQA." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 534, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 534, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 534, + 289, + 715 + ], + "type": "text", + "content": "Dose the simultaneous learning in CR-FIQA lead to better performance in comparison to on-the-top learning? We consider the possibility of learning to estimate CR after finalizing the FR training, in comparison to the simultaneously learning in CR-FIQA. We conducted two additional experiments using pretrained ResNet-50 trained with ArcFace loss [5] on CASIA-WebFace [39]. Specifically, we add an additional single regression layer to this pretrained model. We freeze the weights of the pretrained model and train only the regression layer to learn the internal network observation of the pretrained model using only the " + }, + { + "bbox": [ + 46, + 534, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathcal{CR}}" + }, + { + "bbox": [ + 46, + 534, + 289, + 715 + ], + "type": "text", + "content": " (Equation 5). Using this setting, we present two instances, CR-FIQA(S) (On top) and CCS-FIQA(S) (On top), that learned to predict CR and CCS, respectively. Each of CR-FIQA(S) (On top) and CCS-FIQA(S) (On top) is fine" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 72, + 547, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 300 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 300 + ], + "type": "text", + "content": "tuned for 32K iteration with an initial learning rate of 0.01. The learning rate is divided by 10 at 20K and 28K training iterations, similarly to CR-FIAQ(S) and CCS-FIAQ(S). The results (ERCs and AUCs) of these models are compared to CR-FIQA(S) and CCS-FIQA(S) in Figure 3. The ERCs in Figure 3 presented the evaluation on Adience, AgeDb-30, CALFW and CFP-FP using the ArcFace FR model. The ERCs and AUCs in Figure 3 show that both CR-FIQA(S) and CCS-FIQA(S) lead to stronger reductions in the error than the CR-FIQA(S) (On top) and the CCS-FIQA(S) (On top), when rejecting low-quality samples. This supports our training paradigm that simultaneously learns the internal network observation, CR, while optimizing the class centers. This can be related to the step-wise convergence towards the final CR value in the simultaneously training. For both ablation study questions, ERCs and AUC for all the remaining benchmarks and FR models (mentioned in Section 4) lead to similar conclusions and are provided in the supplementary material." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 304, + 433, + 316 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 304, + 433, + 316 + ], + "spans": [ + { + "bbox": [ + 305, + 304, + 433, + 316 + ], + "type": "text", + "content": "6. Result and Discussion" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 319, + 547, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 319, + 547, + 498 + ], + "spans": [ + { + "bbox": [ + 304, + 319, + 547, + 498 + ], + "type": "text", + "content": "All CR-FIQA performances reported in this paper are obtained under cross-model settings. The proposed CR-FIQA is used only to predict FIQ and not to extract feature representation of face images. None of the utilized FR models (ArcFace [5], ElasticFace [3], MagFace [28], and CurricularFace [17]) is trained with our paradigm. Instead, we trained a separate model for CR-FIQA and used the official pretrained FR models (as described in Section 4) for feature extraction. The verification performances as AUC at FMR1e-3 and FMR1e-4 are presented in Table 1. The visual verification performances as ERC curves (Figure 4) using ArcFace and ElasticFace FR models are reported at FMR1e-3. The ERC curves at FMR1e-4 and for MagFace and CurricularFace FR models at FMR1e-3 are provided in the supplementary material." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 498, + 548, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 498, + 548, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 498, + 548, + 714 + ], + "type": "text", + "content": "The ERC curves (Figure 4) and the AUC values (Table 1) show that our proposed CR-FIQA(S) and CR-FIQA(L) outperformed the SOTA methods by significant margins in almost all settings. Observing the results on IJB-C, Adience, CFP-FP, CALFW, and CPLFW at FMR1e-3 and FMR1e-4 (Figure 4 and Table 1), our proposed CR-FIQA outperformed all SOTA methods on all the considered FR models. On the AgeDB-30 benchmark, our proposed CR-FIQA ranked first in five out of eight settings and second in the other three settings (Table 1). On the LFW benchmark, our proposed CR-FIQA ranked behind the MagFace and the RankIQ. This is the only case that our models did not outperform all SOTA methods. However, it can be noticed from the ERC curves in Figure 4 that none of the SOTA methods were able to achieve stable behavior (smoothly decaying curve) on LFW. The main reason for such unstable ERC behavior on LFW is that the FR performance on LFW is nearly saturated (all models achieved above " + }, + { + "bbox": [ + 304, + 498, + 548, + 714 + ], + "type": "inline_equation", + "content": "99.80\\%" + }, + { + "bbox": [ + 304, + 498, + 548, + 714 + ], + "type": "text", + "content": " accu" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "5841" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 67, + 116, + 528, + 450 + ], + "blocks": [ + { + "bbox": [ + 47, + 70, + 545, + 115 + ], + "lines": [ + { + "bbox": [ + 47, + 70, + 545, + 115 + ], + "spans": [ + { + "bbox": [ + 47, + 70, + 545, + 115 + ], + "type": "text", + "content": "Table 1. The AUCs of ERC achieved by our CR-FIQA and the SOTA methods under different experimental settings. CR-FIQA achieved the best performance (lowest AUC) in almost all settings. On XQLFW, the SER-FIQ (marked with *) is used for the sample selection of the XQLFW benchmark. The best result for each experimental setting is in bold and the second-ranked one is in italic. The notions of " + }, + { + "bbox": [ + 47, + 70, + 545, + 115 + ], + "type": "inline_equation", + "content": "1e - 3" + }, + { + "bbox": [ + 47, + 70, + 545, + 115 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 70, + 545, + 115 + ], + "type": "inline_equation", + "content": "1e - 4" + }, + { + "bbox": [ + 47, + 70, + 545, + 115 + ], + "type": "text", + "content": " indicate the value of the fixed FMR at which the ERC curve (FNMR vs. reject) was calculated at" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 67, + 116, + 528, + 450 + ], + "lines": [ + { + "bbox": [ + 67, + 116, + 528, + 450 + ], + "spans": [ + { + "bbox": [ + 67, + 116, + 528, + 450 + ], + "type": "table", + "html": "
FRMethodAdience [6]AgeDB-30 [31]CFP-FP [33]LFW [16]CALFW [41]CPLFW [40]XQLFW [23]IJB-C [27]
1e-31e-41e-31e-41e-31e-41e-31e-41e-31e-41e-31e-41e-31e-41e-31e-4
ArcFace [5]QABRISQUE [29]0.05650.12850.04000.05850.03430.04330.00430.00490.07550.08130.25580.30370.66800.71220.03810.0656
RankIQA [26]0.04000.09330.03720.05230.03010.03840.00390.00450.08460.09150.24370.29690.65840.70390.03850.0640
DeepIQA [2]0.05680.13720.04030.05230.02380.02920.00490.00560.07930.08500.23090.28560.59580.64580.03830.0640
FIQARankIQ [4]0.03530.08730.03220.04200.01520.02600.00180.00240.06080.06720.06330.08480.27890.33320.02270.0342
PFE [34]0.02120.04280.01720.02260.00920.01290.00230.00280.06470.06810.04500.06380.23020.27100.01760.0248
SER-FIQ [36]0.02230.04340.01670.02230.00650.01030.00230.00280.05950.06270.03890.05840.1812*0.2295*0.01610.0241
FaceQnet [14, 15]0.03460.07340.01970.02450.02400.02730.00220.00270.07740.08220.15040.17510.58290.61360.02700.0376
MagFace [28]0.02070.04250.01560.01980.00730.01050.00160.00210.05680.06020.04920.06420.40220.46360.01710.0254
SDD-FIQA [32]0.02480.05620.01860.02060.01220.01930.00210.00270.06410.06980.05170.06700.30900.35610.01860.0270
CR-FIQA(S)(Our)0.02410.05170.01440.01870.00900.01450.00200.00250.05210.05540.03910.05670.23770.27400.01710.0250
CR-FIQA(L)(Our)0.02040.03530.01590.01890.00500.00820.00230.00290.06160.06320.03600.05150.20840.24410.01380.0207
ElasticFace [3]QABRISQUE [29]0.06440.11840.03750.04030.02810.03720.00340.00470.07260.07470.26410.46880.63430.69640.03570.0621
RankIQA [26]0.04330.08620.03740.04360.02690.03180.00330.00450.08100.08350.23250.43060.61890.68560.03660.0599
DeepIQA [2]0.06450.12030.03840.04110.01910.02560.00430.00560.07560.07720.24010.45410.54000.58320.03790.0590
FIQARankIQ [4]0.04000.07770.03090.03370.01490.01800.00130.00200.05980.06140.05810.07270.24680.27760.02260.0334
PFE [34]0.02220.03810.01630.01720.00880.01130.00180.00250.06280.06430.04190.08950.21120.24360.01710.0247
SER-FIQ [36]0.02400.04170.01630.01790.00610.00850.00210.00280.05740.05900.03870.05130.1576*0.1868*0.01560.0235
FaceQnet [14, 15]0.03690.06670.01940.02070.02270.02470.00210.00260.07630.07770.14200.28800.55490.58440.02630.0370
MagFace [28]0.02250.03850.01500.01580.00690.00950.00140.00210.05530.05630.04740.05970.39730.42820.01660.0243
SDD-FIQA [32]0.02770.05120.01870.02000.00980.01180.00190.00270.06240.06380.04930.06340.30520.35620.01830.0266
CR-FIQA(S)(Our)0.02570.04650.01460.01600.00700.00960.00150.00220.05090.05220.03830.05020.20930.28350.01670.0244
CR-FIQA(L)(Our)0.02140.03570.01490.01590.00450.00650.00180.00250.05940.06080.03500.04620.17980.20600.01350.0203
MagFace [28]QABRISQUE [29]0.05940.13080.04420.07990.04220.05890.00430.00580.07580.07880.46490.68090.69110.72290.04620.0787
RankIQA [26]0.04070.08890.03700.06810.03690.05430.00410.00560.08290.08570.32510.64750.67060.70460.04610.0750
DeepIQA [2]0.05710.13020.04170.07210.03220.05450.00480.00590.07870.08090.36720.66320.61620.65190.04740.0765
FIQARankIQ [4]0.03590.08370.03610.05310.02130.03320.00790.00270.06020.06290.06590.16420.30760.34750.02690.0383
PFE [34]0.02150.04230.01920.03170.01070.01380.00230.00290.06400.06520.04490.14350.26150.29260.02000.0283
SER-FIQ [36]0.02330.04510.01850.02930.00800.01390.00250.00330.05900.06070.03970.52180.60160.62100.03050.0422
FaceQnet [14, 15]0.03650.07200.02170.03140.02710.03510.00220.00270.07630.07730.29880.52180.64750.71880.19550.1979
MagFace [28]0.02120.04170.01590.02470.00850.01290.00170.00220.05620.05780.05060.88870.44780.49000.01950.0279
SDD-FIQA [32]0.02530.05620.02160.03050.01460.02010.00210.06430.06570.05250.11880.34040.39280.39280.02150.0307
CR-FIQA(S)(Our)0.02440.05070.01650.02340.01020.01210.00200.00280.05160.05280.04090.08400.26700.33360.01980.0284
CR-FIQA(L)(Our)0.02110.03720.01740.02350.00620.00800.00230.00280.06140.06280.03740.06790.23690.28390.01620.0236
CurricularFace [17]QABRISQUE [29]0.05020.10950.04330.04910.03230.03570.00410.00470.07550.07840.27090.50570.61460.63360.03620.0589
RankIQA [26]0.03590.07520.03940.05100.02980.03560.00390.00450.88060.86560.23460.46540.56440.62120.13610.1558
DeepIQA [2]0.04920.10700.04070.04760.02270.02780.00500.00560.7640.7860.24880.49610.51650.55260.13760.1558
FIQARankIQ [4]0.03140.07150.03650.04170.01860.02490.00180.00240.59590.66451515*1515*1515*1515*1515*1515*
PFE [34]0.01980.03650.01970.02270.01030.01340.00240.00280.63630.65751515*1515*1515*1515*1515*1515*
SER-FIQ [36]0.02110.03810.01670.01930.00740.01110.00250.03561518*1519*1518*1519*1519*1519*1519*1519*
FaceQnet [14, 15]0.03260.06260.02210.02670.02260.02740.00220.02771518*1519*1518*1519*1519*1519*1519*1519*
MagFace [28]0.02000.03640.01670.01950.00780.01110.00160.00211516*1517*1518*1519*1519*1519*1519*1519*
SDD-FIQA [32]0.02300.04620.02190.02540.01380.01850.00210.06271517*1518*1518*1519*1519*1519*1519*1519*
CR-FIQA(S)(Our)0.02270.04460.01560.01980.00970.01480.00200.02551513*1514*1514*1515*1515*1515*1515*1515*
CR-FIQA(L)(Our)0.01980.03360.01620.02000.00540.08800.00230.02991516*1517*1518*1519*1519*1519*1519*1519*
", + "image_path": "a83dde4923042bfae4c611ef0b68d1a56fc0c639dc7c63ba3adc4eb9b92e81a7.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 64, + 468, + 525, + 582 + ], + "blocks": [ + { + "bbox": [ + 52, + 456, + 541, + 468 + ], + "lines": [ + { + "bbox": [ + 52, + 456, + 541, + 468 + ], + "spans": [ + { + "bbox": [ + 52, + 456, + 541, + 468 + ], + "type": "text", + "content": "Table 2. Verification performance on the IJB-C (1:1 mixed verification) [27]. CR-FIQA outperformed SOTA methods under all settings" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 64, + 468, + 525, + 582 + ], + "lines": [ + { + "bbox": [ + 64, + 468, + 525, + 582 + ], + "spans": [ + { + "bbox": [ + 64, + 468, + 525, + 582 + ], + "type": "table", + "html": "
Quality Estimation1:1 mixed Verification: TAR (%) at
ArcFace [5]ElasticFace [3]MagFace [28]CurricularFace [17]
FAR=1e-6FAR=1e-5FAR=1e-4FAR=1e-6FAR=1e-5FAR=1e-4FAR=1e-6FAR=1e-5FAR=1e-4FAR=1e-6FAR=1e-5FAR=1e-4
-89.8594.4796.2889.1594.5496.4985.6793.0896.6590.4694.8996.58
IAQBRISQUE [29]86.6593.6295.9885.6893.5195.6581.1190.6494.8288.1693.9896.29
RankIQA [26]86.3793.6195.8386.7193.4696.0080.7890.7594.8688.1694.1196.22
DeepIQA [2]81.9791.6494.6778.9391.5994.8173.5386.3492.9082.6592.0495.00
FIQARankIQ [4]88.7894.4296.2088.8894.6496.4585.6392.6695.7090.0094.9396.53
PFE [34]89.5094.5196.3189.1094.6796.5184.9392.4495.6090.3695.0496.54
SER-FIQ [36]89.7494.6596.3290.0594.7996.5786.0293.3595.8090.6695.1196.58
FaceQNet [14, 15]87.8794.0496.1286.2694.0996.2582.9190.5695.0389.6194.6596.36
MagFace [28]89.4994.4196.2289.3794.6996.4685.7592.7195.5490.3495.0296.50
SDD-FIQA [32]89.3994.6196.3488.0794.8296.4984.6992.8395.7389.9195.1296.63
CR-FIQA(S)(Our)89.5994.7896.3590.3094.9796.6386.4593.4895.9590.8295.1396.64
CR-FIQA(L)(Our)90.1694.7596.3690.0094.9296.5887.1293.6795.9090.7995.1296.58
", + "image_path": "bccaa1745a6c4cd0aba0c843c803d8fbe3c42f672e186bd9926fea5ef137c70a.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 594, + 286, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 594, + 286, + 616 + ], + "spans": [ + { + "bbox": [ + 47, + 594, + 286, + 616 + ], + "type": "text", + "content": "racy [3,5,17,28]), leaving very few samples causing errors and thus lowering the significance of the measured FNMR." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 617, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 617, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 617, + 287, + 713 + ], + "type": "text", + "content": "The XQLFW benchmark is derived from LFW to contain pairs with a maximum difference in quality. The XQLFW images are chosen based on BRISQUE [29] and SER-FIQ [23] quality scores to be either extremely high or low. The use of SER-FIQ [36] in this selection might give a biased edge for SER-FIQ on this benchmark. On XQLFW, our CR-FIQA achieved very close performance to the selection method (SER-FIQ) and is far ahead of all other SOTA" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 594, + 545, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 545, + 641 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 545, + 641 + ], + "type": "text", + "content": "methods. Lastly, our proposed CR-FIQA(S) achieved very comparable performance to our CR-FIQA(L), pointing out the robustness of our approach, regardless of the training database and architecture complexity." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "type": "text", + "content": "Table 2 presents the verification performance on the IJB-C 1:1 mixed verification benchmark [27] using quality scores as an embedding weighting term (as defined in [27]) under different experimental settings. For each of the FR models, we report in the first row the evaluation result of the corresponding FR model as defined in the protocol [27] and" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "5842" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 66, + 70, + 180, + 155 + ], + "blocks": [ + { + "bbox": [ + 66, + 70, + 180, + 155 + ], + "lines": [ + { + "bbox": [ + 66, + 70, + 180, + 155 + ], + "spans": [ + { + "bbox": [ + 66, + 70, + 180, + 155 + ], + "type": "image", + "image_path": "a82f22ad1d52348cae670ed95b1d433880540117ff65dfeac6f27bed43eab300.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 181, + 70, + 294, + 156 + ], + "blocks": [ + { + "bbox": [ + 181, + 70, + 294, + 156 + ], + "lines": [ + { + "bbox": [ + 181, + 70, + 294, + 156 + ], + "spans": [ + { + "bbox": [ + 181, + 70, + 294, + 156 + ], + "type": "image", + "image_path": "35fd13f5e6b524ca30ce69a4ca87e66bb352cb00e02ffce4323012f9e319af6b.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 299, + 70, + 413, + 156 + ], + "blocks": [ + { + "bbox": [ + 299, + 70, + 413, + 156 + ], + "lines": [ + { + "bbox": [ + 299, + 70, + 413, + 156 + ], + "spans": [ + { + "bbox": [ + 299, + 70, + 413, + 156 + ], + "type": "image", + "image_path": "dd3f2ff2716f695cb1b20f5a21e8994a64b92decdde2434660c1580bf37c4872.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 413, + 71, + 527, + 156 + ], + "blocks": [ + { + "bbox": [ + 413, + 71, + 527, + 156 + ], + "lines": [ + { + "bbox": [ + 413, + 71, + 527, + 156 + ], + "spans": [ + { + "bbox": [ + 413, + 71, + 527, + 156 + ], + "type": "image", + "image_path": "eeab048702330d160cf068de1e61c9c8c719c70d4a5234bb2c1a337493f81ec3.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 65, + 157, + 180, + 242 + ], + "blocks": [ + { + "bbox": [ + 65, + 157, + 180, + 242 + ], + "lines": [ + { + "bbox": [ + 65, + 157, + 180, + 242 + ], + "spans": [ + { + "bbox": [ + 65, + 157, + 180, + 242 + ], + "type": "image", + "image_path": "fadcbf964e12e0dbe9bcadd5f6854b07c15d34449b1b5e0006c456b408d746ae.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 181, + 157, + 294, + 242 + ], + "blocks": [ + { + "bbox": [ + 181, + 157, + 294, + 242 + ], + "lines": [ + { + "bbox": [ + 181, + 157, + 294, + 242 + ], + "spans": [ + { + "bbox": [ + 181, + 157, + 294, + 242 + ], + "type": "image", + "image_path": "7741785b04c7e29ac5e32d5519d423666998ec255da71cdf7abe0cd322fb48f3.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 299, + 157, + 413, + 242 + ], + "blocks": [ + { + "bbox": [ + 299, + 157, + 413, + 242 + ], + "lines": [ + { + "bbox": [ + 299, + 157, + 413, + 242 + ], + "spans": [ + { + "bbox": [ + 299, + 157, + 413, + 242 + ], + "type": "image", + "image_path": "8ab76e9fddc9a173b612786ef91a3f8c3ec9eabb5edfcfa96acb622fa9884c91.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 413, + 157, + 527, + 242 + ], + "blocks": [ + { + "bbox": [ + 413, + 157, + 527, + 242 + ], + "lines": [ + { + "bbox": [ + 413, + 157, + 527, + 242 + ], + "spans": [ + { + "bbox": [ + 413, + 157, + 527, + 242 + ], + "type": "image", + "image_path": "e2e3e85d37ad943874fb5cf960272f0d698373ca90379092b9d3c4a0fa897358.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 65, + 243, + 181, + 328 + ], + "blocks": [ + { + "bbox": [ + 65, + 243, + 181, + 328 + ], + "lines": [ + { + "bbox": [ + 65, + 243, + 181, + 328 + ], + "spans": [ + { + "bbox": [ + 65, + 243, + 181, + 328 + ], + "type": "image", + "image_path": "771b35430170d9df9908ce8d99f810923a68162605e46e89191629d401bdf2a7.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 181, + 243, + 294, + 328 + ], + "blocks": [ + { + "bbox": [ + 181, + 243, + 294, + 328 + ], + "lines": [ + { + "bbox": [ + 181, + 243, + 294, + 328 + ], + "spans": [ + { + "bbox": [ + 181, + 243, + 294, + 328 + ], + "type": "image", + "image_path": "80a3098b1c61c8fa07b1fe66a09f99bd622791f47301a62ec369cf59f48fbed5.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 299, + 243, + 413, + 327 + ], + "blocks": [ + { + "bbox": [ + 299, + 243, + 413, + 327 + ], + "lines": [ + { + "bbox": [ + 299, + 243, + 413, + 327 + ], + "spans": [ + { + "bbox": [ + 299, + 243, + 413, + 327 + ], + "type": "image", + "image_path": "b74275bebd231178c2913a54158c46f8c1aafdcd4751a5c33b91115c833e7f16.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 413, + 243, + 527, + 327 + ], + "blocks": [ + { + "bbox": [ + 413, + 243, + 527, + 327 + ], + "lines": [ + { + "bbox": [ + 413, + 243, + 527, + 327 + ], + "spans": [ + { + "bbox": [ + 413, + 243, + 527, + 327 + ], + "type": "image", + "image_path": "754b9d7570e68dac2860bd33ad61da7fb0e41879677bb01c05e17149e0f8f39e.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 65, + 329, + 181, + 415 + ], + "blocks": [ + { + "bbox": [ + 65, + 329, + 181, + 415 + ], + "lines": [ + { + "bbox": [ + 65, + 329, + 181, + 415 + ], + "spans": [ + { + "bbox": [ + 65, + 329, + 181, + 415 + ], + "type": "image", + "image_path": "fd3370c640db7d3890be16d27af380026231c669c8b5ff0e18eee497f6a5dac8.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 109, + 417, + 481, + 424 + ], + "lines": [ + { + "bbox": [ + 109, + 417, + 481, + 424 + ], + "spans": [ + { + "bbox": [ + 109, + 417, + 481, + 424 + ], + "type": "text", + "content": "BRIQUE RankIQ DeepIQ " + }, + { + "bbox": [ + 109, + 417, + 481, + 424 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 109, + 417, + 481, + 424 + ], + "type": "text", + "content": " RankIQ PFE SER-FIQ FaceQnet MagFace SDD-FIQA CR-FIQA(S) Our) CR-FIQA(L) Our" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 45, + 427, + 545, + 483 + ], + "lines": [ + { + "bbox": [ + 45, + 427, + 545, + 483 + ], + "spans": [ + { + "bbox": [ + 45, + 427, + 545, + 483 + ], + "type": "text", + "content": "Figure 4. ERC (FNMR at FMR1e-3 vs reject) curves for all evaluated benchmarks using ArcFace and ElasticFace FR models corresponding to Table 1 results. The visual evaluation, ERC curves, using MagFace and CurricularFace FR models are provided in the supplementary material. The proposed CR-FIQA(L) and CR-FIQA(S) are marked with solid blue and red lines, respectively. CR-FIQA leads to lower verification error, when rejecting a fraction of images, of the lowest quality, in comparison to SOTA methods (faster decaying curve) under most experimental settings." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 181, + 329, + 294, + 415 + ], + "blocks": [ + { + "bbox": [ + 181, + 329, + 294, + 415 + ], + "lines": [ + { + "bbox": [ + 181, + 329, + 294, + 415 + ], + "spans": [ + { + "bbox": [ + 181, + 329, + 294, + 415 + ], + "type": "image", + "image_path": "c51273b871eb43e818f7674904a61d55ce6efead9e5ec90af238d4680af0f9b2.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 299, + 329, + 413, + 415 + ], + "blocks": [ + { + "bbox": [ + 299, + 329, + 413, + 415 + ], + "lines": [ + { + "bbox": [ + 299, + 329, + 413, + 415 + ], + "spans": [ + { + "bbox": [ + 299, + 329, + 413, + 415 + ], + "type": "image", + "image_path": "4574db913f7abfee3cf4c2214a57e182c99d16898a22d7961980aa2737ed2843.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 413, + 329, + 527, + 415 + ], + "blocks": [ + { + "bbox": [ + 413, + 329, + 527, + 415 + ], + "lines": [ + { + "bbox": [ + 413, + 329, + 527, + 415 + ], + "spans": [ + { + "bbox": [ + 413, + 329, + 527, + 415 + ], + "type": "image", + "image_path": "e0e060cd3b1e3f8ff022c16b61b1dd512397d2ca7e17851a36ebbc283c3e656c.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "bbox": [ + 46, + 486, + 288, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 486, + 288, + 594 + ], + "spans": [ + { + "bbox": [ + 46, + 486, + 288, + 594 + ], + "type": "text", + "content": "the corresponding released evaluation scripts [3, 5, 17, 28], i.e. without considering the FIQ. Our proposed CR-FIQA significantly leads to higher verification performance than all evaluated SOTA methods, when the quality score is used as an embedding weighting term (Table 2). This achievement is observable under all experimental settings (Table 2). Another outcome of this evaluation is that the integration of CR-FIQA leads to SOTA verification performance on one of the most challenging FR benchmarks, IJB-C [27]." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 47, + 599, + 119, + 611 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 599, + 119, + 611 + ], + "spans": [ + { + "bbox": [ + 47, + 599, + 119, + 611 + ], + "type": "text", + "content": "7. Conclusion" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 46, + 617, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 617, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 617, + 288, + 715 + ], + "type": "text", + "content": "In this work, we propose the CR-FIQA approach that probes the relative classifiability of training samples of the FR model and utilize this observation to learn to predict the utility of any given sample in achieving an accurate FR performance. We experimentally prove the theorized relationship between the sample relative classifiability and FIQ and build on that towards our CR-FIQA. The CR-FIQA training paradigm simultaneously learns to optimize the class center" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 486, + 545, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 486, + 545, + 617 + ], + "spans": [ + { + "bbox": [ + 304, + 486, + 545, + 617 + ], + "type": "text", + "content": "while learning to predict sample relative classifiability. The presented ablation studies and the extensive experimental results prove the effectiveness of the proposed CR-FIQA approach, and its design choices, as an FIQ method. The reported results demonstrated that our proposed CR-FIQA outperformed SOTA methods repeatedly across multiple FR models and on multiple benchmarks, including ones with a large age gap (AgeDb-30, Adience, CALFW), large quality difference (XQLFW), large pose variation (CPLFW, CFPFP), and extremely large-scale and challenging FR benchmarks (IJB-C)." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 304, + 617, + 546, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 617, + 546, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 617, + 546, + 714 + ], + "type": "text", + "content": "Acknowledgment This research work has been funded by the German Federal Ministry of Education and Research and the Hessian Ministry of Higher Education, Research, Science and the Arts within their joint support of the National Research Center for Applied Cybersecurity ATHENE. This work has been partially funded by the German Federal Ministry of Education and Research (BMBF) through the Software Campus Project." + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5843" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 123 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 123 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 123 + ], + "type": "text", + "content": "[1] Lacey Best-Rowden and Anil K. Jain. Learning face image quality from human assessments. IEEE Trans. Inf. Forensics Secur., 13(12):3064–3077, 2018. 1, 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 125, + 288, + 178 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 125, + 288, + 178 + ], + "spans": [ + { + "bbox": [ + 53, + 125, + 288, + 178 + ], + "type": "text", + "content": "[2] Sebastian Bosse, Dominique Maniry, Klaus-Robert Müller, Thomas Wiegand, and Wojciech Samek. Deep neural networks for no-reference and full-reference image quality assessment. IEEE Trans. Image Process., 27(1):206-219, 2018. 6, 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 180, + 287, + 245 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 180, + 287, + 245 + ], + "spans": [ + { + "bbox": [ + 53, + 180, + 287, + 245 + ], + "type": "text", + "content": "[3] Fadi Boutros, Naser Damer, Florian Kirchbuchner, and Arjan Kuijper. Elasticface: Elastic margin loss for deep face recognition. In IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, CVPR Workshops 2022, New Orleans, LA, USA, June 19-20, 2022, pages 1577-1586. IEEE, 2022. 3, 5, 6, 7, 8" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 247, + 286, + 279 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 247, + 286, + 279 + ], + "spans": [ + { + "bbox": [ + 53, + 247, + 286, + 279 + ], + "type": "text", + "content": "[4] Jiansheng Chen, Yu Deng, Gaocheng Bai, and Guangda Su. Face image quality assessment based on learning to rank. IEEE Signal Process. Lett., 22(1):90-94, 2015. 2, 6, 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 280, + 286, + 345 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 280, + 286, + 345 + ], + "spans": [ + { + "bbox": [ + 53, + 280, + 286, + 345 + ], + "type": "text", + "content": "[5] Jiankang Deng, Jia Guo, Niannan Xue, and Stefanos Zafeiriou. Arcface: Additive angular margin loss for deep face recognition. In IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2019, Long Beach, CA, USA, June 16-20, 2019, pages 4690-4699. Computer Vision Foundation / IEEE, 2019. 3, 4, 5, 6, 7, 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 346, + 286, + 379 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 346, + 286, + 379 + ], + "spans": [ + { + "bbox": [ + 53, + 346, + 286, + 379 + ], + "type": "text", + "content": "[6] Eran Eidinger, Roee Enbar, and Tal Hassner. Age and gender estimation of unfiltered faces. IEEE Trans. Inf. Forensics Secur., 9(12):2170-2179, 2014. 5, 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 380, + 286, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 380, + 286, + 402 + ], + "spans": [ + { + "bbox": [ + 53, + 380, + 286, + 402 + ], + "type": "text", + "content": "[7] Frontex. Best practice technical guidelines for automated border control (abc) systems, 2015. 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 403, + 286, + 446 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 403, + 286, + 446 + ], + "spans": [ + { + "bbox": [ + 53, + 403, + 286, + 446 + ], + "type": "text", + "content": "[8] Biying Fu, Cong Chen, Olaf Henniger, and Naser Damer. A deep insight into measuring face image utility with general and face-specific image quality metrics. pages 1121-1130, 2022. 1, 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 447, + 286, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 447, + 286, + 491 + ], + "spans": [ + { + "bbox": [ + 53, + 447, + 286, + 491 + ], + "type": "text", + "content": "[9] Ross B. Girshick. Fast R-CNN. In 2015 IEEE International Conference on Computer Vision, ICCV 2015, Santiago, Chile, December 7-13, 2015, pages 1440-1448. IEEE Computer Society, 2015. 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 492, + 286, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 492, + 286, + 536 + ], + "spans": [ + { + "bbox": [ + 48, + 492, + 286, + 536 + ], + "type": "text", + "content": "[10] P. Grother, M. Ngan A. Hom, and K. Hanaoka. Ongoing face recognition vendor test (frvt) part 5: Face image quality assessment (4th draft). In National Institute of Standards and Technology. Tech. Rep., Sep. 2021. 1, 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 537, + 286, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 537, + 286, + 568 + ], + "spans": [ + { + "bbox": [ + 48, + 537, + 286, + 568 + ], + "type": "text", + "content": "[11] P. Grother and E. Tabassi. Performance of biometric quality measures. IEEE Trans. on Pattern Analysis and Machine Intelligence, 29(4):531-543, Apr. 2007. 5" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 570, + 286, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 570, + 286, + 658 + ], + "spans": [ + { + "bbox": [ + 48, + 570, + 286, + 658 + ], + "type": "text", + "content": "[12] Yandong Guo, Lei Zhang, Yuxiao Hu, Xiaodong He, and Jianfeng Gao. Ms-celeb-Im: A dataset and benchmark for large-scale face recognition. In Bastian Leibe, Jiri Matas, Nicu Sebe, and Max Welling, editors, Computer Vision - ECCV 2016 - 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part III, volume 9907 of Lecture Notes in Computer Science, pages 87-102. Springer, 2016. 4, 5" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 658, + 286, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 286, + 713 + ], + "type": "text", + "content": "[13] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In 2016 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2016, Las Vegas, NV, USA, June 27-30, 2016, pages 770-778. IEEE Computer Society, 2016. 4, 5" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "text", + "content": "[14] Javier Hernandez-Ortega, Javier Galbally, Julian Fierrez, and Laurent Beslay. Biometric quality: Review and application to face recognition with faceqnet. CoRR, abs/2006.03298, 2020. 1, 2, 6, 7" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 118, + 545, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 545, + 183 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 545, + 183 + ], + "type": "text", + "content": "[15] Javier Hernandez-Ortega, Javier Galbally, Julian Fierrez, Rudolf Haraksim, and Laurent Beslay. Faceqnet: Quality assessment for face recognition based on deep learning. In 2019 International Conference on Biometrics, ICB 2019, Crete, Greece, June 4-7, 2019, pages 1-8. IEEE, 2019. 1, 2, 6, 7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 184, + 545, + 238 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 184, + 545, + 238 + ], + "spans": [ + { + "bbox": [ + 308, + 184, + 545, + 238 + ], + "type": "text", + "content": "[16] Gary B. Huang, Manu Ramesh, Tamara Berg, and Erik Learned-Miller. Labeled faces in the wild: A database for studying face recognition in unconstrained environments. Technical Report 07-49, University of Massachusetts, Amherst, October 2007. 5, 7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 239, + 545, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 239, + 545, + 315 + ], + "spans": [ + { + "bbox": [ + 308, + 239, + 545, + 315 + ], + "type": "text", + "content": "[17] Yuge Huang, Yuhan Wang, Ying Tai, Xiaoming Liu, Pengcheng Shen, Shaoxin Li, Jilin Li, and Feiyue Huang. Curricularface: Adaptive curriculum learning loss for deep face recognition. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2020, Seattle, WA, USA, June 13-19, 2020, pages 5900-5909. Computer Vision Foundation / IEEE, 2020. 3, 5, 6, 7, 8" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 316, + 545, + 347 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 316, + 545, + 347 + ], + "spans": [ + { + "bbox": [ + 308, + 316, + 545, + 347 + ], + "type": "text", + "content": "[18] ISO/IEC JTC1 SC17 WG3. Portrait Quality - Reference Facial Images for MRTD. International Civil Aviation Organization, 2018. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 349, + 545, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 349, + 545, + 392 + ], + "spans": [ + { + "bbox": [ + 308, + 349, + 545, + 392 + ], + "type": "text", + "content": "[19] ISO/IEC JTC1 SC37 Biometrics. ISO/IEC 29794-1:2016 Information technology - Biometric sample quality - Part 1: Framework. International Organization for Standardization, 2016. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 393, + 545, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 393, + 545, + 426 + ], + "spans": [ + { + "bbox": [ + 308, + 393, + 545, + 426 + ], + "type": "text", + "content": "[20] ISO/IEC JTC1 SC37 Biometrics. ISO/IEC 2382-37:2017 Information technology - Vocabulary - Part 37: Biometrics. International Organization for Standardization, 2017. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 426, + 545, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 426, + 545, + 470 + ], + "spans": [ + { + "bbox": [ + 308, + 426, + 545, + 470 + ], + "type": "text", + "content": "[21] ISO/IEC JTC1 SC37 Biometrics. ISO/IEC 19795-1:2021 Information technology — Biometric performance testing and reporting — Part 1: Principles and framework. International Organization for Standardization, 2021. 5" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 471, + 545, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 471, + 545, + 525 + ], + "spans": [ + { + "bbox": [ + 308, + 471, + 545, + 525 + ], + "type": "text", + "content": "[22] Thorsten Joachims. Optimizing search engines using click-through data. In Proceedings of the Eighth ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, July 23-26, 2002, Edmonton, Alberta, Canada, pages 133-142. ACM, 2002. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 526, + 545, + 592 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 526, + 545, + 592 + ], + "spans": [ + { + "bbox": [ + 308, + 526, + 545, + 592 + ], + "type": "text", + "content": "[23] Martin Knoche, Stefan Hörmann, and Gerhard Rigoll. Cross-quality LFW: A database for analyzing cross-resolution image face recognition in unconstrained environments. In 16th IEEE International Conference on Automatic Face and Gesture Recognition, FG 2021, Jodhpur, India, December 15-18, 2021, pages 1-5. IEEE, 2021. 5, 7" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 592, + 545, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 592, + 545, + 658 + ], + "spans": [ + { + "bbox": [ + 308, + 592, + 545, + 658 + ], + "type": "text", + "content": "[24] Weiyang Liu, Yandong Wen, Zhiding Yu, Ming Li, Bhiksha Raj, and Le Song. Spherical face: Deep hypersphere embedding for face recognition. In 2017 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2017, Honolulu, HI, USA, July 21-26, 2017, pages 6738-6746. IEEE Computer Society, 2017. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 658, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 658, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 308, + 658, + 545, + 712 + ], + "type": "text", + "content": "[25] Weiyang Liu, Yandong Wen, Zhiding Yu, Ming Li, Bhiksha Raj, and Le Song. SphereFace: Deep hypersphere embedding for face recognition. In Proc. of the IEEE Conf. on Computer Vision and Pattern Recognition, pages 212-220, 2017. 3" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "5844" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 137 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 137 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 137 + ], + "type": "text", + "content": "[26] Xialei Liu, Joost van de Weijer, and Andrew D. Bagdanov. Rankiaq: Learning from rankings for no-reference image quality assessment. In IEEE International Conference on Computer Vision, ICCV 2017, Venice, Italy, October 22-29, 2017, pages 1040-1049. IEEE Computer Society, 2017. 1, 6, 7" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 140, + 288, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 140, + 288, + 215 + ], + "spans": [ + { + "bbox": [ + 48, + 140, + 288, + 215 + ], + "type": "text", + "content": "[27] Brianna Maze, Jocelyn C. Adams, James A. Duncan, Nathan D. Kalka, Tim Miller, Charles Otto, Anil K. Jain, W. Tyler Niggel, Janet Anderson, Jordan Cheney, and Patrick Grother. IARPA janus benchmark - C: face dataset and protocol. In 2018 International Conference on Biometrics, ICB 2018, Gold Coast, Australia, February 20-23, 2018, pages 158-165. IEEE, 2018. 5, 7, 8" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 217, + 288, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 217, + 288, + 281 + ], + "spans": [ + { + "bbox": [ + 48, + 217, + 288, + 281 + ], + "type": "text", + "content": "[28] Qiang Meng, Shichao Zhao, Zhida Huang, and Feng Zhou. Magface: A universal representation for face recognition and quality assessment. In IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2021, virtual, June 19-25, 2021, pages 14225-14234. Computer Vision Foundation / IEEE, 2021. 1, 2, 3, 5, 6, 7, 8" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 283, + 287, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 283, + 287, + 326 + ], + "spans": [ + { + "bbox": [ + 48, + 283, + 287, + 326 + ], + "type": "text", + "content": "[29] Anish Mittal, Anush Krishna Moorthy, and Alan Conrad Bovik. No-reference image quality assessment in the spatial domain. IEEE Trans. Image Process., 21(12):4695-4708, 2012. 1, 6, 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 327, + 287, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 327, + 287, + 360 + ], + "spans": [ + { + "bbox": [ + 48, + 327, + 287, + 360 + ], + "type": "text", + "content": "[30] Anish Mittal, Rajiv Soundararajan, and Alan C. Bovik. Making a \"completely blind\" image quality analyzer. IEEE Signal Process. Lett., 20(3):209-212, 2013. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 361, + 287, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 361, + 287, + 426 + ], + "spans": [ + { + "bbox": [ + 48, + 361, + 287, + 426 + ], + "type": "text", + "content": "[31] Stylianos Moschoglou, Athanasios Papaioannou, Christos Sagonas, Jiankang Deng, Irene Kotsia, and Stefanos Zafeiriou. Agedb: The first manually collected, in-the-wild age database. In 2017 IEEE CVPRW, CVPR Workshops 2017, Honolulu, HI, USA, July 21-26, 2017, pages 1997-2005. IEEE Computer Society, 2017. 5, 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 426, + 287, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 426, + 287, + 503 + ], + "spans": [ + { + "bbox": [ + 48, + 426, + 287, + 503 + ], + "type": "text", + "content": "[32] Fu-Zhao Ou, Xingyu Chen, Ruixin Zhang, Yuge Huang, Shaoxin Li, Jilin Li, Yong Li, Liujuan Cao, and Yuan-Gen Wang. SDD-FIQA: unsupervised face image quality assessment with similarity distribution distance. In IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2021, virtual, June 19-25, 2021, pages 7670-7679. Computer Vision Foundation / IEEE, 2021. 1, 2, 5, 6, 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 504, + 287, + 570 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 504, + 287, + 570 + ], + "spans": [ + { + "bbox": [ + 48, + 504, + 287, + 570 + ], + "type": "text", + "content": "[33] Soumyadip Sengupta, Jun-Cheng Chen, Carlos Domingo Castillo, Vishal M. Patel, Rama Chellappa, and David W. Jacobs. Frontal to profile face verification in the wild. In 2016 IEEE Winter Conference on Applications of Computer Vision, WACV 2016, Lake Placid, NY, USA, March 7-10, 2016, pages 1-9. IEEE Computer Society, 2016. 5, 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 571, + 287, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 571, + 287, + 624 + ], + "spans": [ + { + "bbox": [ + 48, + 571, + 287, + 624 + ], + "type": "text", + "content": "[34] Yichun Shi and Anil K. Jain. Probabilistic face embeddings. In 2019 IEEE/CVF International Conference on Computer Vision, ICCV 2019, Seoul, Korea (South), October 27 - November 2, 2019, pages 6901-6910. IEEE, 2019. 1, 2, 5, 6, 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 625, + 287, + 691 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 625, + 287, + 691 + ], + "spans": [ + { + "bbox": [ + 48, + 625, + 287, + 691 + ], + "type": "text", + "content": "[35] Philipp Terhörst, Jan Niklas Kolf, Naser Damer, Florian Kirchbuchner, and Arjan Kuijper. Face quality estimation and its correlation to demographic and non-demographic bias in face recognition. In 2020 IEEE International Joint Conference on Biometrics, IJCB 2020, Houston, TX, USA, September 28 - October 1, 2020, pages 1-11. IEEE, 2020. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 692, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 692, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 692, + 287, + 713 + ], + "type": "text", + "content": "[36] Philipp Terhörst, Jan Niklas Kolf, Naser Damer, Florian Kirchbuchner, and Arjan Kuijper. SER-FIQ: unsupervised" + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 373 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 128 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 128 + ], + "type": "text", + "content": "estimation of face image quality based on stochastic embedding robustness. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2020, Seattle, WA, USA, June 13-19, 2020, pages 5650-5659. Computer Vision Foundation / IEEE, 2020. 1, 2, 6, 7" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 307, + 129, + 545, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 129, + 545, + 195 + ], + "spans": [ + { + "bbox": [ + 307, + 129, + 545, + 195 + ], + "type": "text", + "content": "[37] Hao Wang, Yitong Wang, Zheng Zhou, Xing Ji, Dihong Gong, Jingchao Zhou, Zhifeng Li, and Wei Liu. Cosface: Large margin cosine loss for deep face recognition. In 2018 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2018, Salt Lake City, UT, USA, June 18-22, 2018, pages 5265-5274. IEEE Computer Society, 2018. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 307, + 196, + 545, + 250 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 196, + 545, + 250 + ], + "spans": [ + { + "bbox": [ + 307, + 196, + 545, + 250 + ], + "type": "text", + "content": "[38] Weidi Xie, Jeffrey Byrne, and Andrew Zisserman. Inducing predictive uncertainty estimation for face verification. In 31st British Machine Vision Conference 2020, BMVC 2020, Virtual Event, UK, September 7-10, 2020. BMVA Press, 2020. 1, 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 251, + 545, + 283 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 251, + 545, + 283 + ], + "spans": [ + { + "bbox": [ + 307, + 251, + 545, + 283 + ], + "type": "text", + "content": "[39] Dong Yi, Zhen Lei, Shengcai Liao, and Stan Z. Li. Learning face representation from scratch. CoRR, abs/1411.7923, 2014. 2, 4, 5, 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 285, + 545, + 329 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 285, + 545, + 329 + ], + "spans": [ + { + "bbox": [ + 308, + 285, + 545, + 329 + ], + "type": "text", + "content": "[40] T. Zheng and W. Deng. Cross-posed lfw: A database for studying cross-posed face recognition in unconstrained environments. Technical Report 18-01, Beijing University of Posts and Telecommunications, February 2018. 5, 7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 331, + 545, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 331, + 545, + 373 + ], + "spans": [ + { + "bbox": [ + 308, + 331, + 545, + 373 + ], + "type": "text", + "content": "[41] Tianyue Zheng, Weihong Deng, and Jiani Hu. Cross-age LFW: A database for studying cross-age face recognition in unconstrained environments. CoRR, abs/1708.08197, 2017. 5, 7" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5845" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/CRAFT_ Concept Recursive Activation FacTorization for Explainability/6e462b05-7931-4758-8402-5b3e432aab49_content_list.json b/2023/CRAFT_ Concept Recursive Activation FacTorization for Explainability/6e462b05-7931-4758-8402-5b3e432aab49_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..1e8282ef3b25957e8f546252594df4eeb80037b5 --- /dev/null +++ b/2023/CRAFT_ Concept Recursive Activation FacTorization for Explainability/6e462b05-7931-4758-8402-5b3e432aab49_content_list.json @@ -0,0 +1,1596 @@ +[ + { + "type": "text", + "text": "CRAFT: Concept Recursive Activation FacTorization for Explainability", + "text_level": 1, + "bbox": [ + 120, + 130, + 843, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Thomas Fel $^{1,3,5*}$ Agustin Picard $^{3,6*}$ Louis Bethune $^{3*}$ Thibaut Boissin $^{3,4*}$ David Vigouroux $^{3,4}$ Julien Colin $^{1,3}$ Rémi Cadène $^{1,2}$ Thomas Serre $^{1,3}$", + "text_level": 1, + "bbox": [ + 189, + 161, + 790, + 222 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Carney Institute for Brain Science, Brown University, USA $^{2}$ Sorbonne Université, CNRS, France", + "bbox": [ + 158, + 227, + 808, + 243 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3Artificial and Natural Intelligence Toulouse Institute, Université de Toulouse, France", + "bbox": [ + 202, + 244, + 764, + 261 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "4 Institut de Recherche Technologique Saint-Exupery, France", + "bbox": [ + 284, + 263, + 686, + 279 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{5}$ Innovation & Research Division, SNCF, $^{6}$ Scalian", + "bbox": [ + 305, + 280, + 656, + 296 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/d56fc2459ec3e516eb51004c50cc917e487af368e9f0e0748d6313be63eb1089.jpg", + "image_caption": [ + "Figure 1. The \"Man on the Moon\" incorrectly classified as a \"shovel\" by an ImageNet-trained ResNet50. Heatmap generated by a classic attribution method [55] (left) vs. concept attribution maps generated with the proposed CRAFT approach (right) which highlights the two most influential concepts that drove the ResNet50's decision along with their corresponding locations. CRAFT suggests that the neural net arrived at its decision because it identified the concept of \"dirt\" $\\bullet$ commonly found in members of the image class \"shovel\" and the concept of \"ski pants\" $\\bullet$ typically worn by people clearing snow from their driveway with a shovel instead the correct concept of astronaut's pants (which was probably never seen during training)." + ], + "image_footnote": [], + "bbox": [ + 156, + 306, + 395, + 489 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/cfaae2cfa00822a5aaf09c97ac20222f3dbdded4a740b519378efb4996e10350.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 405, + 306, + 622, + 489 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/6835fc9574e013856aa964e78ddce99ff8752d0d9cf9e6ea644f2ba6d1955fc0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 624, + 309, + 671, + 347 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/5483d089ef7a9bae1eb123914f9294f97bb3a9f8d11f0d49f0c90775826474f6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 681, + 311, + 728, + 347 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/0b85a981ba4c88245dbfa77a2854edaea9ef873e88e992ad1b8b4d7a1d18dd60.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 738, + 311, + 784, + 347 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/50b2c5017faf5a3ce4184951b8c86b5099a4b2c9796403fb3dd7c7793669b945.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 625, + 352, + 671, + 390 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/ac323bfa9cd0fa3a16accf3ae3a9b9b7691b25760f3c52c28032f39284837361.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 681, + 352, + 728, + 390 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/5ea86c83b0c1d9ec6a497be15e05d4d4f8f8bdd75cbc977e2f2ea1ac10ce1a1d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 738, + 352, + 784, + 390 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/16db1080dad5ebbf1b0dec5c82e8be9894b5f6f7a6d600f44a6a199860ae7d28.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 622, + 406, + 671, + 444 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/93b31b3e45f8d1330b6dd828eacfff7ed168ba2fae6aa85c955c36c9887ba996.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 676, + 407, + 723, + 443 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/68174e964303eabc967e664f324f7940387368b35bbb24afc6ac725c0e2a529c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 735, + 407, + 782, + 444 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/aea7dc080165edc0f2e8ba45de0d4027c82715e57bc8b26fba4b9e2dbf260592.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 622, + 449, + 671, + 484 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/b2604fe305ebdb7f4d7a78839207afd070c12f4646315e3b2de8b16c4c7a35a3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 679, + 449, + 723, + 484 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/87798a3af8a411c63c8d751d755be2c6b5dda99166e85ca07541112978aa008b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 735, + 449, + 782, + 487 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 607, + 313, + 622 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Attribution methods, which employ heatmaps to identify the most influential regions of an image that impact model decisions, have gained widespread popularity as a type of explainability method. However, recent research has exposed the limited practical value of these methods, attributed in part to their narrow focus on the most prominent regions of an image - revealing \"where\" the model looks, but failing to elucidate \"what\" the model sees in those areas. In this work, we try to fill in this gap with CRAFT - a novel approach to identify both \"what\" and \"where\" by generating concept-based explanations. We introduce 3 new ingredients to the automatic concept extraction literature: (i) a recursive strategy to detect and decompose concepts across layers, (ii) a novel method for a more faithful estimation of concept importance using Sobol indices, and (iii) the use of implicit differentiation to unlock Concept Attribution Maps.", + "bbox": [ + 75, + 627, + 472, + 868 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We conduct both human and computer vision experiments to demonstrate the benefits of the proposed approach.", + "bbox": [ + 76, + 869, + 470, + 901 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We show that the proposed concept importance estimation technique is more faithful to the model than previous methods. When evaluating the usefulness of the method for human experimenters on a human-centered utility benchmark, we find that our approach significantly improves on two of the three test scenarios.", + "bbox": [ + 496, + 608, + 893, + 699 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 500, + 710, + 632, + 726 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Interpreting the decisions of modern machine learning models such as neural networks remains a major challenge. Given the ever-increasing range of machine learning applications, the need for robust and reliable explainability methods continues to grow [11, 35]. Recently enacted European laws (including the General Data Protection Regulation (GDPR) [37] and the European AI act [43]) require the assessment of explainable decisions, especially those made by algorithms.", + "bbox": [ + 496, + 733, + 893, + 869 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In order to try to meet this growing need, an array of explainability methods have already been proposed [14, 50,", + "bbox": [ + 498, + 869, + 893, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "2711", + "bbox": [ + 480, + 924, + 513, + 936 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/052520db0224c7b7815264d63e73a1ac961132927e72f15c833437870d919206.jpg", + "image_caption": [ + "Figure 2. CRAFT results for the prediction \"chain saw\". First, our method uses Non-Negative Matrix Factorization (NMF) to extract the most relevant concepts used by the network (ResNet50V2) from the train set (ILSVRC2012 [10]). The global influence of these concepts on the predictions is then measured using Sobol indices (right panel). Finally, the method provides local explanations through concept attribution maps (heatmaps associated with a concept, and computed using grad-CAM by backpropagating through the NMF concept values with implicit differentiation). Besides, concepts can be interpreted by looking at crops that maximize the NMF coefficients. For the class \"chain saw\", the detected concepts seem to be: $\\bullet$ the chainsaw engine, $\\bullet$ the saw blade, $\\bullet$ the human head, $\\bullet$ the vegetation, $\\bullet$ the jeans and $\\bullet$ the tree trunk." + ], + "image_footnote": [], + "bbox": [ + 189, + 95, + 568, + 257 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/61a7202338e1d1ecb53d1059ce55768474216d7dc8d5ca828866fbf4c3093902.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 581, + 107, + 782, + 241 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "59, 61, 62, 66, 69, 71, 76]. One of the main classes of methods called attribution methods yields heatmaps that indicate the importance of individual pixels for driving a model's decision. However, these methods exhibit critical limitations [1, 27, 63, 65], as they have been shown to fail – or only marginally help – in recent human-centered benchmarks [8, 26, 41, 49, 60, 64]. It has been suggested that their limitations stem from the fact that they are only capable of explaining where in an image are the pixels that are critical to the decision but they cannot tell what visual features are actually driving decisions at these locations. In other words, they show where the model looks but not what it sees. For example, in the scenario depicted in Fig. 1, where an ImageNet-trained ResNet mistakenly identifies an image as containing a shovel, the attribution map displayed on the left fails to explain the reasoning behind this misclassification.", + "bbox": [ + 75, + 375, + 472, + 628 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "A recent approach has sought to move past attribution methods [40] by using so-called \"concepts\" to communicate information to users on how a model works. The goal is to find human-interpretable concepts in the activation space of a neural network. Although the approach exhibited potential, its practicality is significantly restricted due to the need for prior knowledge of pertinent concepts in its original formulation and, more critically, the requirement for a labeled dataset of such concepts. Several lines of work have focused on trying to automate the concept discovery process based only on the training dataset and without explicit human supervision. The most prominent of these techniques, ACE [24], uses a combination of segmentation and clustering techniques but requires heuristics to remove outliers. However, ACE provides a proof of concept that it might be possible to discover concepts automatically and at scale – without additional labeling or human supervision. Nevertheless, the approach suffers several limitations: by", + "bbox": [ + 75, + 628, + 472, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "construction, each image segment can only belong to a single cluster, a layer has to be selected by the user to be used to retrieve the relevant concepts, and the amount of information lost during the outlier rejection phase can be a cause of concern. More recently, Zhang et al. [77] proposes to leverage matrix decompositions on internal feature maps to discover concepts.", + "bbox": [ + 496, + 375, + 893, + 479 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Here, we try to fill these gaps with a novel method called CRAFT which uses Non-Negative Matrix Factorization (NMF) [46] for concept discovery. In contrast to other concept-based explanation methods, our approach provides an explicit link between their global and local explanations (Fig. 2) and identifies the relevant layer(s) to use to represent individual concepts (Fig. 3). Our main contributions can be described as follows:", + "bbox": [ + 496, + 482, + 893, + 603 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "(i) A novel approach for the automated extraction of high-level concepts learned by deep neural networks. We validate its practical utility to users with human psychophysics experiments.", + "(ii) A recursive procedure to automatically identify concepts and sub-concepts at the right level of granularity – starting with our decomposition at the top of the model and working our way upstream. We validate the benefit of this approach with human psychophysics experiments showing that (i) the decomposition of a concept yields more coherent sub-concepts and (ii) that the groups of points formed by these sub-concepts are more refined and appear meaningful to humans.", + "(iii) A novel technique to quantify the importance of individual concepts for a model's prediction using Sobol indices [34, 67, 68] – a technique borrowed from Sensitivity Analysis.", + "(iv) The first concept-based explainability method which produces concept attribution maps by backpropagating con" + ], + "bbox": [ + 496, + 606, + 893, + 901 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2712", + "bbox": [ + 480, + 924, + 514, + 936 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "cept scores into the pixel space by leveraging the implicit function theorem in order to localize the pixels associated with the concept of a given input image. This effectively opens up the toolbox of both white-box [15, 59, 62, 66, 69, 71, 78] and black-box [14, 47, 55, 57] explainability methods to derive concept-wise attribution maps.", + "bbox": [ + 75, + 90, + 470, + 183 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 76, + 196, + 218, + 212 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Attribution methods Attribution methods are widely used as post-hoc explainability techniques to determine the input variables that contribute to a model's prediction by generating importance maps, such as the ones shown in Fig.1. The first attribution method, Saliency, introduced in [62], generates a heatmap by utilizing the gradient of a given classification score with respect to the pixels. This method was later improved upon in the context of deep convolutional networks for classification in subsequent studies, such as [66, 69, 71, 76]. However, the image gradient only reflects the model's operation within an infinitesimal neighborhood around an input, which can yield misleading importance estimates [22] since gradients of large vision models are notoriously noisy [66]. That is why several methods leverage perturbations on the input image to probe the model and create importance maps that indicate the most crucial areas for the decision, such as Rise [55], Sobol [14], or more recently HSIC [50].", + "bbox": [ + 75, + 223, + 470, + 494 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Unfortunately, a severe limitation of these approaches – apart from the fact that they only show the “where” – is that they are subject to confirmation bias: while they may appear to offer useful explanations to a user, sometimes these explanations are actually incorrect [1, 23, 65]. These limitations raise questions about their usefulness, as recent research has shown by using human-centered experiments to evaluate the utility of attribution [8, 26, 41, 49, 60].", + "bbox": [ + 75, + 496, + 470, + 617 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In particular, in [8], a protocol is proposed to measure the usefulness of explanations, corresponding to how much they help users identify rules driving a model's predictions (correct or incorrect) that transfer to unseen data - using the concept of meta-predictor (also called simulatability) [11, 18, 39]. The main idea is to train users to predict the output of the system using a small set of images along with associated model predictions and corresponding explanations. A method that performs well on this benchmark is said useful, as it helps users better predict the output of the model by providing meaningful information about the internal functioning of the model. This framework being agnostic to the type of explainability method, we have chosen to use it in Section 4 in order to compare CRAFT with attribution methods.", + "bbox": [ + 75, + 618, + 470, + 843 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Concepts-based methods Kim et al. [40] introduced a method aimed at providing explanations that go beyond attribution-based approaches by measuring the impact of", + "bbox": [ + 76, + 854, + 470, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "pre-selected concepts on a model's outputs. Although this method appears more interpretable to human users than standard attribution techniques, it requires a database of images describing the relevant concepts to be manually curated. Ghorbani et al. [24] further extended the approach to extract concepts without the need for human supervision. The approach, called ACE [24], uses a segmentation scheme on images, that belong to an image class of interest. The authors leveraged the intermediate activations of a neural network for specific image segments. These segments were resized to the appropriate input size and filled with a baseline value. The resulting activations were clustered to produce prototypes, which they referred to as \"concepts\". However, some concepts contained background segments, leading to the inclusion of uninteresting and outlier concepts. To address this, the authors implemented a post-processing cleanup step to remove these concepts, including those that were present in only one image of the class and were not representative. While this improved the interpretability of their explanations to human subjects, the use of a baseline value filled around the segments could introduce biases in the explanations [28, 31, 42, 70].", + "bbox": [ + 496, + 90, + 893, + 422 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Zhang et al. [77] developed a solution to the unsupervised concept discovery problem by using matrix factorizations in the latent spaces of neural networks. However, one major drawback of this method is that it operates at the level of convolutional kernels, leading to the discovery of localized concepts. For example, the concept of \"grass\" at the bottom of the image is considered to be distinct from the concept of \"grass\" at the top of the image.", + "bbox": [ + 496, + 422, + 893, + 544 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Overview of the method", + "text_level": 1, + "bbox": [ + 500, + 550, + 723, + 566 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we first describe our concept activations factorization method. Below we highlight the main differences with related work. We then proceed to introduce the three novel ingredients that make up CRAFT: (1) a method to recursively decompose concepts into sub-concepts, (2) a method to better estimate the importance of extracted concepts, and (3) a method to use any attribution method to create concept attribution maps, using implicit differentiation [4, 25, 44].", + "bbox": [ + 496, + 575, + 893, + 712 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Notations In this work, we consider a general supervised learning setting, where $(\\pmb{x}_1,\\dots,\\pmb{x}_n)\\in \\mathcal{X}^n\\subseteq \\mathbb{R}^{n\\times d}$ are $n$ inputs images and $(y_{1},\\ldots ,y_{n})\\in \\mathcal{V}^{n}$ their associated labels. We are given a (machine-learnt) black-box predictor $\\pmb {f}:\\mathcal{X}\\rightarrow \\mathcal{Y}$ , which at some test input $\\pmb{x}$ predicts the output $\\pmb {f}(\\pmb {x})$ . Without loss of generality, we establish that $\\pmb{f}$ is a neural network that can be decomposed into two distinct components. The first component is a function $\\pmb{g}$ that maps the input to an intermediate state, and the second component is $\\pmb{h}$ , which takes this intermediate state to the output, such that $\\pmb {f}(\\pmb {x}) = (\\pmb {h}\\circ \\pmb {g})(\\pmb {x})$ . In this context, $\\pmb {g}(\\pmb {x})\\subseteq \\mathbb{R}^{p}$ represents the intermediate activations of $\\pmb{x}$ within the network.", + "bbox": [ + 496, + 719, + 895, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "2713", + "bbox": [ + 480, + 925, + 514, + 936 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/a2a3380f552a846acecb3712c8a3860bdbf91e6323fd03082069489064b1d72b.jpg", + "image_caption": [ + "Figure 3. (1) Neural collapse (amalgamation). A classifier needs to be able to linearly separate classes by the final layer. It is commonly assumed that in order to achieve this, image activations from the same class get progressively \"merged\" such that these image activations converge to a one-hot vector associated with the class at the level of the logits layer. In practice, this means that different concepts get ultimately blended together along the way. (2) Recursive process. When a concept is not understood (e.g., $\\mathcal{C}$ ), we propose to decompose it into multiple sub-concepts (e.g., $\\mathcal{C}_1, \\mathcal{C}_2, \\mathcal{C}_3$ ) using the activations from an earlier layer to overcome the aforementioned neural collapse issue. (3) Example of recursive concept decomposition using CRAFT on the ImageNet class \"parachute\"." + ], + "image_footnote": [], + "bbox": [ + 125, + 90, + 290, + 200 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/4c64d3cbc2ea37f63d43a79059e87b7c1fbd58b8e3470ea5ac5cd0c4a99fb751.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 292, + 90, + 460, + 199 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/50fe647394038ab8f1cc171d224b5a1f7b125578c3fba4e4e52aae750a13caa9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 462, + 90, + 846, + 210 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Further, we will assume non-negative activations: $g(\\boldsymbol{x}) \\geq 0$ . In particular, this assumption is verified by any architecture that utilizes $ReLU$ , but any non-negative activation function works.", + "bbox": [ + 75, + 321, + 470, + 383 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1. Concept activation factorization.", + "text_level": 1, + "bbox": [ + 76, + 393, + 364, + 409 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We use Non-negative matrix factorization to identify a basis for concepts based on a network's activations (Fig.4). Inspired by the approach taken in ACE [24], we will use image sub-regions to try to identify coherent concepts.", + "bbox": [ + 75, + 416, + 470, + 477 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The first step involves gathering a set of images that one wishes to explain, such as the dataset, in order to generate associated concepts. In our examples, to explain a specific class $y \\in \\mathcal{V}$ , we selected the set of points $\\mathcal{C}$ from the dataset for which the model's predictions matched a specific class $\\mathcal{C} = \\{\\pmb{x}_i : \\pmb{f}(\\pmb{x}_i) = y, 1 \\leq i \\leq n\\}$ . It is important to emphasize that this choice is significant. The goal is not to understand how humans labeled the data, but rather to comprehend the model itself. By only selecting correctly classified images, important biases and failure cases may be missed, preventing a complete understanding of our model.", + "bbox": [ + 75, + 478, + 468, + 643 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Now that we have defined our set of images, we will proceed with selecting sub-regions of those images to identify specific concepts within a localized context. It has been observed that the implementation of segmentation masks suggested in ACE can lead to the introduction of artifacts due to the associated inpainting with a baseline value. In contrast, our proposed method takes advantage of the prevalent use of modern data augmentation techniques such as randaugment, mixup, and cutmix during the training of current models. These techniques involve the current practice of models being trained on image crops, which enables us to leverage a straightforward crop and resize function denoted by $\\pi (\\cdot)$ to create sub-regions (illustrated in Fig.4). By applying $\\pi$ function to each image in the set $\\mathcal{C}$ , we obtain an auxiliary dataset $\\mathbf{X}\\in \\mathbb{R}^{n\\times d}$ such that each entries $\\mathbf{X}_i = \\pi (\\mathbf{x}_i)$ is an image crop.", + "bbox": [ + 75, + 643, + 470, + 885 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To discover the concept basis, we start by obtaining the", + "bbox": [ + 96, + 885, + 470, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "activations for the random crops $\\mathbf{A} = g(\\mathbf{X})\\in \\mathbb{R}^{n\\times p}$ . In the case where $\\pmb{f}$ is a convolutional neural network, a global average pooling is applied to the activations.", + "bbox": [ + 496, + 321, + 890, + 368 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We are now ready to apply Non-negative Matrix Factorization (NMF) to decompose positive activations $\\mathbf{A}$ into a product of non-negative, low-rank matrices $\\mathbf{U} \\in \\mathbb{R}^{n \\times r}$ and $\\mathbf{W} \\in \\mathbb{R}^{p \\times r}$ by solving:", + "bbox": [ + 496, + 368, + 893, + 428 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\left(\\mathbf {U}, \\mathbf {W}\\right) = \\underset {\\mathbf {U} \\geq 0, \\mathbf {W} \\geq 0} {\\arg \\min } \\frac {1}{2} \\| \\mathbf {A} - \\mathbf {U W} ^ {\\mathsf {T}} \\| _ {F} ^ {2}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 558, + 428, + 890, + 460 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $||\\cdot ||_F$ denotes the Frobenius norm.", + "bbox": [ + 496, + 462, + 779, + 477 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "This decomposition of our activations $\\mathbf{A}$ yields two matrices: $\\mathbf{W}$ containing our Concept Activation Vectors (CAVs) and $\\mathbf{U}$ that redefines the data points in our dataset according to this new basis. Moreover, this decomposition in this new basis has some interesting properties that go beyond the simple low-rank factorization - since $r \\ll \\min(n, p)$ . First, NMF can be understood as the joint learning of a dictionary of Concept Activation Vectors - called a \"concept bank\" in Fig. 4 - that maps a $\\mathbb{R}^p$ basis onto $\\mathbb{R}^r$ , and $\\mathbf{U}$ the coefficients of the vectors $\\mathbf{A}$ expressed in this new basis. The minimization of the reconstruction error $\\frac{1}{2} \\| \\mathbf{A} - \\mathbf{U} \\mathbf{W} \\|_F^2$ ensures that the new basis contains (mostly) relevant concepts. Intuitively, the non-negativity constraints $\\mathbf{U} \\geq 0$ , $\\mathbf{W} \\geq 0$ encourage $(i)$ $\\mathbf{W}$ to be sparse (useful for creating disentangled concepts), $(ii)$ $\\mathbf{U}$ to be sparse (convenient for selecting a minimal set of useful concepts) and $(iii)$ missing data to be imputed [56], which corresponds to the sparsity pattern of post-ReLU activations $\\mathbf{A}$ .", + "bbox": [ + 496, + 478, + 893, + 750 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "It is worth noting that each input $\\mathbf{x}_i$ can be expressed as a linear combination of concepts denoted as $\\mathbf{A}_i = \\sum_{j=1}^{r} \\mathbf{U}_{(i,j)} \\mathbf{W}_j^{\\top}$ . This approach is advantageous because it allows us to interpret each input as a composition of the underlying concepts. Furthermore, the strict positivity of each term - NMF is working over the anti-negative semiring, - enhances the interpretability of the decomposition. Another interesting interpretation could be that each input is represented as a superposition of concepts [13].", + "bbox": [ + 496, + 750, + 893, + 885 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "While other methods in the literature solve a similar", + "bbox": [ + 517, + 885, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "2714", + "bbox": [ + 480, + 924, + 514, + 936 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "problem (such as low-rank factorization using SVD or ICA), the NMF is both fast and effective and is known to yield concepts that are meaningful to humans [20,74,77]. Finally, once the concept bank $\\mathbf{W}$ has been precomputed, we can associate the concept coefficients $\\pmb{u}$ to any new input $\\pmb{x}$ (e.g., a full image) by solving the underlying Non-Negative Least Squares (NNLS) problem $\\min_{\\pmb{u} \\geq 0} \\frac{1}{2} \\| \\pmb{g}(\\pmb{x}) - \\pmb{u} \\mathbf{W}^{\\top} \\|_F^2$ , and therefore recover its decomposition in the concept basis.", + "bbox": [ + 75, + 90, + 472, + 212 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In essence, the core of our method can be summarized as follows: using a set of images, the idea is to re-interpret their embedding at a given layer as a composition of concepts that humans can easily understand. In the next section, we show how one can recursively apply concept activation factorizations to preceding layer for an image containing a previously computed concept.", + "bbox": [ + 75, + 215, + 472, + 321 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2. Ingredient 1: A pinch of recursivity", + "text_level": 1, + "bbox": [ + 76, + 338, + 387, + 356 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "One of the most apparent issues in previous work [24,77] is the need for choosing a priori a layer at which the activation maps are computed. This choice will critically affect the concepts that are identified because certain concepts get amalgamated [53] into one at different layers of the neural network, resulting in incoherent and indecipherable clusters, as illustrated in Fig. 3. We posit that this can be solved by iteratively applying our decomposition at different layer depths, and for the concepts that remain difficult to understand, by looking for their sub-concepts in earlier layers by isolating the images that contain them. This allows us to build hierarchies of concepts for each class.", + "bbox": [ + 75, + 364, + 472, + 547 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We offer a simple solution consisting of reapplying our method to a concept by performing a second step of concept activation factorization on a set of images that contain the concept $\\mathcal{C}$ in order to refine it and create sub-concepts (e.g., decompose $\\mathcal{C}$ into $\\{\\mathcal{C}_1, \\mathcal{C}_2, \\mathcal{C}_3\\}$ ) see Fig. 3 for an illustrative example. Note that we generalize current methods in the sense that taking images $(x_1, \\dots, x_n)$ that are clustered in the logits layer (belonging to the same class) and decomposing them in a previous layer - as done in [24, 77] - is a valid recursive step. For a more general case, let us assume that a set of images that contain a common concept is obtained using the first step of concept activation factorization.", + "bbox": [ + 75, + 550, + 472, + 733 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We will then take a subset of the auxiliary dataset points to refine any concept $j$ . To do this, we select the subset of points that contain the concept $\\mathcal{C}_j = \\{\\mathbf{X}_i : \\mathbf{U}_{(i,j)} > \\lambda_j, 1 \\leq i \\leq n\\}$ , where $\\lambda_j$ is the 90th percentile of the values of the concept $\\mathbf{U}_{(1,j)}, \\ldots, \\mathbf{U}_{(n,j)}$ . In other words, the 10% of images that activate the concept $j$ the most are selected for further refinement into sub-concepts. Given this new set of points, we can then re-apply the Concept Matrix Factorization method to an earlier layer $g'(\\cdot)$ to obtain the sub-concepts decomposition from the initial concept - as illustrated in Fig.3.", + "bbox": [ + 75, + 734, + 472, + 902 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Ingredient 2: A dash of sensitivity analysis", + "text_level": 1, + "bbox": [ + 498, + 90, + 862, + 107 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "A major concern with concept extraction methods is that concepts that make sense to humans are not necessarily the same as those being used by a model to classify images. In order to prevent such confirmation bias during our concept analysis phase, a faithful estimate the overall importance of the extracted concepts is crucial. Kim et al. [40] proposed an importance estimator based on directional derivatives: the partial derivative of the model output with respect to the vector of concepts. While this measure is theoretically grounded, it relies on the same principle as gradient-based methods, and thus, suffers from the same pitfalls: neural network models have noisy gradients [66, 71]. Hence, the farther the chosen layer is from the output, the noisier the directional derivative score will be.", + "bbox": [ + 496, + 114, + 893, + 325 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Since we essentially want to know which concept has the greatest effect on the output of the model, it is natural to consider the field of sensitivity analysis [9,33,34,67,68]. In this section, we briefly recall the classic \"total Sobol indices\" and how to apply them to our problem. The complete derivation of the Sobol-Hoeffding decomposition is presented in Section D of the supplementary materials. Formally, a natural way to estimate the importance of a concept $i$ is to measure the fluctuations of the model's output $h(\\mathbf{U}\\mathbf{W}^{\\top})$ in response to meaningful perturbations of the concept coefficient $\\mathbf{U}_{(1,i)},\\ldots ,\\mathbf{U}_{(n,i)}$ . Concretely, we will use perturbation masks $\\mathbf{M} = (M_1,\\dots,M_r)\\sim \\mathcal{U}([0,1]^r)$ , here an i.i.d sequence of real-valued random variables, we introduce a concept fluctuation to reconstruct a perturbed activation $\\tilde{\\mathbf{A}} = (\\mathbf{U}\\odot \\mathbf{M})\\mathbf{W}^{\\top}$ where $\\odot$ denote the Hadamard product (e.g., the masks can be used to remove a concept by setting its value to zero). We can then propagate this perturbed activation to the model output $\\mathbf{Y} = h(\\tilde{\\mathbf{A}})$ . Simply put, removing or applying perturbation of an important concept will result in a substantial variation in the output, whereas an unused concept will have minimal effect on the output.", + "bbox": [ + 496, + 327, + 895, + 657 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Finally, we can capture the importance that a concept might have as a main effect - along with its interactions with other concepts - on the model's output by calculating the expected variance that would remain if all the concepts except the $i$ were to be fixed. This yields the general definition of the total Sobol indices.", + "bbox": [ + 496, + 659, + 893, + 750 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Definition 3.1 (Total Sobol indices). The total Sobol index $S_{i}^{T}$ , which measures the contribution of a concept $i$ as well as its interactions of any order with any other concepts to the model output variance, is given by:", + "bbox": [ + 498, + 758, + 893, + 820 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {S} _ {i} ^ {T} = \\frac {\\mathbb {E} _ {\\boldsymbol {M} _ {\\sim i}} (\\mathbb {V} _ {M _ {i}} (\\boldsymbol {h} ((\\boldsymbol {U} \\odot \\boldsymbol {M}) \\mathbf {W} ^ {\\top}) | \\boldsymbol {M} _ {\\sim i}))}{\\mathbb {V} (\\boldsymbol {h} ((\\boldsymbol {U} \\odot \\boldsymbol {M}) \\mathbf {W} ^ {\\top}))}. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 540, + 827, + 893, + 863 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In practice, this index can be calculated very efficiently [36, 48, 52, 58, 72], more details on the Quasi", + "bbox": [ + 498, + 869, + 895, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "2715", + "bbox": [ + 480, + 925, + 514, + 936 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/a39ce196d5f54782047aaccaa4dc624346401fd6aec3c6c6250a0abc4d16b1ab.jpg", + "image_caption": [ + "Figure 4. Overview of CRAFT. Starting from a set of crops $\\mathbf{X}$ containing a concept $\\mathcal{C}$ (e.g., crops images of the class \"parachute\"), we compute activations $g(\\mathbf{X})$ corresponding to an intermediate layer from a neural network for random image crops. We then factorize these activations into two lower-rank matrices, $(\\mathbf{U}, \\mathbf{W})$ . $\\mathbf{W}$ is what we call a \"concept bank\" and is a new basis used to express the activations, while $\\mathbf{U}$ corresponds to the corresponding coefficients in this new basis. We then extend the method with 3 new ingredients: (1) recursivity - by proposing to re-decompose a concept (e.g., take a new set of images containing $\\mathcal{C}_1$ ) at an earlier layer, (2) a better importance estimation using Sobol indices and (3) an approach to leverage implicit differentiation to generate concept attribution maps to localize concepts in an image." + ], + "image_footnote": [], + "bbox": [ + 148, + 97, + 830, + 244 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Monte Carlo sampling and the estimator used are left in appendix D.", + "bbox": [ + 75, + 364, + 470, + 397 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.4. Ingredient 3: A smidgen of implicit differentiation", + "text_level": 1, + "bbox": [ + 76, + 407, + 465, + 436 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Attribution methods are useful for determining the regions deemed important by a model for its decision, but they lack information about what exactly triggered it. We have seen that we can already extract this information from the matrices $\\mathbf{U}$ and $\\mathbf{W}$ , but as it is, we do not know in what part of an image a given concept is represented. In this section, we will show how we can leverage attribution methods (forward and backward modes) to find where a concept is located in the input image (see Fig. 2). Forward attribution methods do not rely on any gradient computation as they only use inference processes, whereas backward methods require back-propagating through a network's layers. By application of the chain rule, computing $\\partial \\mathbf{U} / \\partial \\mathbf{X}$ requires access to $\\partial \\mathbf{U} / \\partial \\mathbf{A}$ .", + "bbox": [ + 75, + 439, + 470, + 648 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To do so, one could be tempted to solve the linear system $\\mathbf{U}\\mathbf{W}^{\\top} = \\mathbf{A}$ . However, this problem is ill-posed since $\\mathbf{W}^{\\top}$ is low rank. A standard approach is to calculate the Moore-Penrose pseudo-inverse $(\\mathbf{W}^{\\top})^{\\dagger}$ , which solves rank deficient systems by looking at the minimum norm solution [2]. In practice, $(\\mathbf{W}^{\\top})^{\\dagger}$ is computed with the Singular Value Decomposition (SVD) of $\\mathbf{W}^{\\top}$ . Unfortunately, SVD is also the solution to the unstructured minimization of $\\frac{1}{2}\\|\\mathbf{A} - \\mathbf{U}\\mathbf{W}^{\\top}\\|_F^2$ by the Eckart-Young-Mirsky theorem [12]. Hence, the non-negativity constraints of the NMF are ignored, which prevents such approaches from succeeding. Other issues stem from the fact that the $\\mathbf{U},\\mathbf{W}$ decomposition is generally not unique.", + "bbox": [ + 75, + 650, + 470, + 845 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Our third contribution consists of tackling this problem to allow the use of attribution methods, i.e., concept attribution maps, by proposing a strategy to differentiate through the NMF block.", + "bbox": [ + 75, + 847, + 470, + 906 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implicit differentiation of NMF block The NMF problem 1 is NP-hard [73], and it is not convex with respect to the input pair $(\\mathbf{U},\\mathbf{W})$ . However, fixing the value of one of the two factors and optimizing the other turns the NMF formulation into a pair of Non-Negative Least Squares (NNLS) problems, which are convex. This ensures that alternating minimization (a standard approach for NMF) of $(\\mathbf{U},\\mathbf{W})$ factors will eventually reach a local minimum. Each of this alternating NNLS problems fulfills the Karush-Kuhn-Tucker (KKT) conditions [38,45], which can be encoded in the so-called optimality function $\\mathbf{F}$ from [4], see Eq. 9 Appendix C.2. The implicit function theorem [25] allows us to use implicit differentiation [3,25,44] to efficiently compute the Jacobians $\\partial \\mathbf{U} / \\partial \\mathbf{A}$ and $\\partial \\mathbf{W} / \\partial \\mathbf{A}$ without requiring to back-propagate through each of the iterations of the NMF solver:", + "bbox": [ + 496, + 364, + 893, + 603 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\frac {\\partial (\\mathbf {U} , \\mathbf {W} , \\bar {\\mathbf {U}} , \\bar {\\mathbf {W}})}{\\partial \\mathbf {A}} = - (\\partial_ {1} \\mathbf {F}) ^ {- 1} \\partial_ {2} \\mathbf {F}. \\qquad (3)\n$$\n", + "text_format": "latex", + "bbox": [ + 573, + 599, + 893, + 633 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "However, this requires the dual variables $\\bar{\\mathbf{U}}$ and $\\bar{\\mathbf{W}}$ , which are not computed in scikit-learn's [54] popular implementation*. Consequently, we leverage the work of [32] and we re-implement our own solver with Jaxopt [4] based on ADMM [5], a GPU friendly algorithm (see Appendix C.2).", + "bbox": [ + 496, + 640, + 893, + 717 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Concretely, given our concepts bank $\\mathbf{W}$ , the concept attribution maps of a new input $\\pmb{x}$ are calculated by solving the NNLS problem $\\min_{\\mathbf{U} \\geq 0} \\frac{1}{2} \\| \\pmb{g}(\\pmb{x}) - \\mathbf{U} \\mathbf{W}^{\\top} \\|_F^2$ . The implicit differentiation of the NMF block $\\partial \\mathbf{U} / \\partial \\mathbf{A}$ is integrated into the classic back-propagation to obtain $\\partial \\mathbf{U} / \\partial \\pmb{x}$ . Most interestingly, this technical advance enables the use of all white-box explainability methods [59, 62, 66, 69, 71, 78] to generate concept-wise attribution maps and trace the part of an image that triggered the detection of the concept by the network. Additionally, it is even possible to employ black-box", + "bbox": [ + 496, + 717, + 895, + 868 + ], + "page_idx": 5 + }, + { + "type": "page_footnote", + "text": "*Scikit-learn uses a block coordinate descent algorithm [7, 17], with a randomized SVD initialization.", + "bbox": [ + 500, + 875, + 893, + 898 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "2716", + "bbox": [ + 480, + 925, + 514, + 936 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/9e37786f4a8d60e36f6b7046af4bc29705db12b2d17e4c6a39426b11ef62903b.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Husky vs. WolfLeaves"Kit Fox" vs "Red Fox"
Session n°123Utility123Utility123
Baseline55.766.262.970.176.878.658.862.258.8
Control53.361.061.40.9572.078.080.21.0260.759.248.5
AttributionsSaliency [62]53.969.673.31.0683.288.782.41.1361.760.258.2
Integ.-Grad. [71]67.472.873.21.1582.582.585.31.1159.458.358.3
SmoothGrad [66]68.775.378.01.2083.085.786.31.1350.355.061.4
GradCAM [59]77.685.784.11.3481.983.582.41.1054.452.554.1
Occlusion [76]71.075.778.11.2278.886.182.91.1051.060.255.1
Grad.-Input [61]65.863.367.91.0676.582.979.51.0550.057.662.6
ConceptsACE [24]68.871.472.71.1579.873.882.11.0548.446.546.1
CRAFTCO (ours)82.487.085.11.3878.885.589.41.1255.549.553.3
CRAFT (ours)90.697.395.51.5386.286.685.51.1556.550.649.4
", + "bbox": [ + 119, + 88, + 854, + 305 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 1. Utility scores on 3 datasets from [8]. Their Utility benchmark evaluates how well explanations help users identify general rules driving classifications that readily transfer to unseen instances. At training time, users are asked to infer rules driving the decisions of the model given a set of images, and their associated predictions and explanations. At test time, the Utility metric measures the accuracy of users at predicting the model decision on novel images averaged over 3 sessions, and normalized by the baseline accuracy of users trained without explanations. The higher the Utility score, the more useful the explanation, and the more crucial the information provided is for understanding –and thus predicting the model's output– on novel samples. CRAFTCO stands for “CRAFT Concept Only” and designates an experimental condition where only global concepts are given to users, without local explanations (i.e., the concept attribution maps). The first and second best results above the baseline are in bold and underlined, respectively.", + "bbox": [ + 75, + 316, + 895, + 429 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "methods [14, 47, 55, 57] since it only amounts to solving an NNLS problem.", + "bbox": [ + 75, + 441, + 468, + 472 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4. Experimental evaluation", + "text_level": 1, + "bbox": [ + 76, + 481, + 307, + 498 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In order to evaluate the interest and the benefits brought by CRAFT, we start in Section 4.1 by assessing the practical utility of the method on a human-centered benchmark composed of 3 XAI scenarios.", + "bbox": [ + 75, + 506, + 470, + 566 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "After demonstrating the usefulness of the method using these human experiments, we independently validate the 3 proposed ingredients. First, we provide evidence that recursivity allows refining concepts, making them more meaningful to humans using two additional human experiments in Section 4.2. Next, we evaluate our new Sobol estimator and show quantitatively that it provides a more faithful assessment of concept importance in Section 4.3. Finally, we run an ablation experiment that measures the interest of local explanations based on concept attribution maps coupled with global explanations. Additional experiments, including a sanity check and an example of deep dreams applied on the concept bank, as well as many other examples of local explanations for randomly picked images from ILSVRC2012, are included in Section B of the supplementary materials. We leave the discussion on the limitations of this method and on the broader impact in appendix A.", + "bbox": [ + 75, + 566, + 472, + 825 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.1. Utility Evaluation", + "text_level": 1, + "bbox": [ + 76, + 832, + 250, + 848 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "As emphasized by Doshi-Velez et al. [11], the goal of XAI should be to develop methods that help a user better understand the behavior of deep neural network models. An", + "bbox": [ + 75, + 854, + 470, + 900 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "instantiation of this idea was recently introduced by Colin & Fel et al. [8] who described an experimental framework to quantitatively measure the practical usefulness of explainability methods in real-world scenarios. For their initial setup, these authors recruited $n = 1$ , 150 online participants (evaluated over 8 unique conditions and 3 AI scenarios) – making it the largest benchmark to date in XAI. Here, we follow their framework rigorously to allow for the robust evaluation of the utility of our proposed CRAFT method and the related ACE. The 3 representative real-world scenarios are: (1) identifying bias in an AI system (using Husky vs Wolf dataset from [57]), (2) characterizing the visual strategy that are too difficult for an untrained non-expert human observer (using the Paleobotanical dataset from [75]), (3) understanding complex failure cases (using ImageNet “Red fox” vs “Kit fox” binary classification). Using this benchmark, we evaluate CRAFT, ACE, as well as CRAFT with only the global concepts (CRAFTCO) to allow for a fair comparison with ACE. To the best of our knowledge, we are the first to systematically evaluate concept-based methods against attribution methods.", + "bbox": [ + 496, + 441, + 893, + 758 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Results are shown in Table 1 and demonstrate the benefit of CRAFT, which achieves higher scores than all of the attribution methods tested as well as ACE in the first two scenarios. To date, no method appears to exceed the baseline on the third scenario suggesting that additional work is required. We also note that, in the first two scenarios, CRAFTCO is one of the best-performing methods and it always outperforms ACE – meaning that even without the local explanation of the concept attribution maps, CRAFT", + "bbox": [ + 496, + 763, + 895, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "2717", + "bbox": [ + 480, + 925, + 514, + 936 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/20516a634912ecd4577feb6f45635f1a652596a5e85a8f6132459343b7079681.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
IntruderExperts (n = 36)Laymen (n = 37)
Acc. Concept70.19%61.08%
Acc. Sub-Concept74.81% (p = 0.18)67.03% (p = 0.043)
Binary choice
Sub-Concept76.1% (p < 0.001)74.95% (p < 0.001)
Odds Ratios3.532.99
", + "bbox": [ + 81, + 88, + 467, + 212 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 2. Results from the psychophysics experiments to validate the recursivity ingredient.", + "bbox": [ + 75, + 222, + 470, + 251 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "largely outperforms ACE. Examples of concepts produced by CRAFT are shown in the Appendix E.1.", + "bbox": [ + 75, + 260, + 468, + 292 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.2. Validation of Recursivity", + "text_level": 1, + "bbox": [ + 76, + 301, + 303, + 319 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To evaluate the meaningfulness of the extracted high-level concepts, we performed psychophysics experiments with human subjects, whom we asked to answer a survey in two phases. Furthermore, we distinguished two different audiences: on the one hand, experts in machine learning, and on the other hand, people with no particular knowledge of computer vision. Both groups of participants were volunteers and did not receive any monetary compensation. Some examples of the developed interface are available the appendix E. It is important to note that this experiment was carried out independently from the utility evaluation and thus it was setup differently.", + "bbox": [ + 75, + 325, + 470, + 507 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Intruder detection experiment First, we ask users to identify the intruder out of a series of five image crops belonging to a certain class, with the odd one being taken from a different concept but still from the same class. Then, we compare the results of this intruder detection with another intruder detection, this time, using a concept (e.g., $\\mathcal{C}_1$ ) coming from a layer $l$ and one of its sub-concepts (e.g., $\\mathcal{C}_{12}$ in Fig.3) extracted using our recursive method. If the concept (or sub-concept) is coherent, then it should be easy for the users to find the intruder. Table 2 summarizes our results, showing that indeed both concepts and sub-concepts are coherent, and that recursivity can lead to a slightly higher understanding of the generated concepts (significant for non-experts, but not for experts) and might suggest a way to make concepts more interpretable.", + "bbox": [ + 75, + 508, + 470, + 734 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Binary choice experiment In order to test the improvement of coherence of the sub-concept generated by recursivity with respect to the larger parent concept, we showed participants an image crop belonging to both a subcluster and a parent cluster (e.g., $\\pi(\\boldsymbol{x}) \\in \\mathcal{C}_{11} \\subset \\mathcal{C}_1$ ) and asked them which of the two clusters (i.e., $\\mathcal{C}_{11}$ or $\\mathcal{C}_1$ ) seemed to accommodate the image the best. If our hypothesis is correct, then the concept refinement brought by recursivity should help form more coherent clusters. The results in Table 2 are satisfying since in both the expert and non-expert groups, the participants chose the sub-cluster more than $74\\%$ of the", + "bbox": [ + 75, + 734, + 470, + 901 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "time. We measure the significance of our results by fitting a binomial logistic regression to our data, and we find that both groups are more likely to choose the sub-concept cluster (at a $p < 0.001$ ).", + "bbox": [ + 496, + 90, + 893, + 151 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.3. Fidelity analysis", + "text_level": 1, + "bbox": [ + 498, + 160, + 661, + 176 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We propose to simultaneously verify that identified concepts are faithful to the model and that the concept importance estimator performs better than that used in TCAV [40] by using the fidelity metrics introduced in [24, 77]. These metrics are similar to the ones used for attribution methods, which consist of studying the change of the logit score when removing/adding pixels considered important. Here, we do not introduce these perturbations in the pixel space but in the concept space: once $\\mathbf{U}$ and $\\mathbf{W}$ are computed, we reconstruct the matrix $\\mathbf{A} \\approx \\mathbf{U}\\mathbf{W}^{\\top}$ using only the most important concept (or removing the most important concept for deletion) and compute the resulting change in the output of the model. As can be seen from Fig. 5, ranking the extracted concepts using Sobol's importance score results in steeper curves than when they are sorted by their TCAV scores. We confirm that these results generalize with other matrix factorization techniques (PCA, ICA, RCA) in Section F of the Appendix.", + "bbox": [ + 496, + 183, + 893, + 455 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/394c34741cab892729b7195e9895bbc32d1d6b184bbcbf102995c9a3b3e59bc5.jpg", + "image_caption": [ + "Figure 5. (Left) Deletion curves (lower is better). (Right) Insertion curves (higher is better). For both the deletion or insertion metrics, Sobol indices lead to better estimates (calculated on $>100\\mathrm{K}$ images) of important concepts." + ], + "image_footnote": [], + "bbox": [ + 501, + 468, + 702, + 590 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/cec4981062c213d3bf377157effc90255f65310f1d8320ce2b194a1c57bc6484.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 707, + 468, + 890, + 590 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 662, + 619, + 678 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We presented CRAFT, a method to automatically extract understandable concepts from deep networks, which provides insights into the model's decision-making process. Our approach uses a recursive formulation to identify concepts at the correct level of granularity, a novel method for measuring concept importance, and implicit differentiation to generate concept attribution maps. We conducted psychophysics experiments to validate the approach's usefulness and meaningfulness to human experimenters. We anticipate that our work will inspire future research in concept-based explainability methods.", + "bbox": [ + 496, + 686, + 893, + 854 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "2718", + "bbox": [ + 480, + 925, + 514, + 936 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Julius Adebayo, Justin Gilmer, Michael Muelly, Ian Goodfellow, Moritz Hardt, and Been Kim. Sanity checks for saliency maps. Advances in neural information processing systems, 31, 2018. 2, 3, 24", + "[2] João Carlos Alves Barata and Mahir Saleh Hussein. The moore-penrose pseudoinverse: A tutorial review of the theory. Brazilian Journal of Physics, 42(1):146-165, 2012. 6", + "[3] Bradley M Bell and James V Burke. Algorithmic differentiation of implicit functions and optimal values. In Advances in Automatic Differentiation, pages 67-77. Springer, 2008. 6, 19", + "[4] Mathieu Blondel, Quentin Berthet, Marco Cuturi, Roy Frostig, Stephan Hoyer, Felipe Llinares-López, Fabian Pedregosa, and Jean-Philippe Vert. Efficient and modular implicit differentiation. arXiv preprint arXiv:2105.15183, 2021. 3, 6, 19", + "[5] Stephen Boyd, Neal Parikh, Eric Chu, Borja Peleato, Jonathan Eckstein, et al. Distributed optimization and statistical learning via the alternating direction method of multipliers. Foundations and Trends® in Machine learning, 3(1):1-122, 2011. 6, 13, 18, 19", + "[6] Chaofan Chen, Oscar Li, Daniel Tao, Alina Barnett, Cynthia Rudin, and Jonathan K Su. This looks like that: deep learning for interpretable image recognition. Advances in neural information processing systems, 32, 2019. 13", + "[7] Andrzej Cichocki and Anh-Huy Phan. Fast local algorithms for large scale nonnegative matrix and tensor factorizations. IEICE transactions on fundamentals of electronics, communications and computer sciences, 92(3):708-721, 2009. 6, 19", + "[8] Julien Colin, Thomas Fel, Rémi Cadène, and Thomas Serre. What i cannot predict, i do not understand: A human-centered evaluation framework for explainability methods. Advances in Neural Information Processing Systems (NeurIPS), 2021. 2, 3, 7, 21", + "[9] RI Cukier, CM Fortuin, Kurt E Shuler, AG Petschek, and J Ho Schaibly. Study of the sensitivity of coupled reaction systems to uncertainties in rate coefficients. i theory. The Journal of chemical physics, 59(8):3873-3878, 1973. 5", + "[10] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 2, 14, 23", + "[11] Finale Doshi-Velez and Been Kim. Towards a rigorous science of interpretable machine learning. ArXiv e-print, 2017. 1, 3, 7", + "[12] Carl Eckart and Gale Young. The approximation of one matrix by another of lower rank. Psychometrika, 1(3):211-218, 1936. 6", + "[13] Nelson Elhage, Tristan Hume, Catherine Olsson, Nicholas Schiefer, Tom Henighan, Shauna Kravec, Zac Hatfield-Dodds, Robert Lasenby, Dawn Drain, Carol Chen, Roger Grosse, Sam McCandlish, Jared Kaplan, Dario Amodei, Martin Wattenberg, and Christopher Olah. Toy models of superposition. Transformer Circuits Thread, 2022. 4" + ], + "bbox": [ + 78, + 114, + 470, + 891 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[14] Thomas Fel, Rémi Cadène, Mathieu Chalvidal, Matthieu Cord, David Vigouroux, and Thomas Serre. Look at the variance! efficient black-box explanations with sobol-based sensitivity analysis. Advances in Neural Information Processing Systems, 34, 2021. 1, 3, 7, 19", + "[15] Thomas Fel, Mélanie Ducoffe, David Vigouroux, Rémi Cadène, Mikael Capelle, Claire Nicodème, and Thomas Serre. Don't lie to me! robust and efficient explainability with verified perturbation analysis. arXiv preprint arXiv:2202.07728, 2022.3", + "[16] Thomas Fel, Lucas Hervier, David Vigouroux, Antonin Poche, Justin Plakoo, Remi Cadene, Mathieu Chalvidal, Julien Colin, Thibaut Boissin, Louis Béthune, Agustin Picard, Claire Nicodeme, Laurent Gardes, Gregory Flandin, and Thomas Serre. Xplique: A deep learning explainability toolbox. Workshop, Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 16", + "[17] Cédric Févotte and Jérôme Idier. Algorithms for nonnegative matrix factorization with the $\\beta$ -divergence. Neural computation, 23(9):2421-2456, 2011. 6, 19", + "[18] Ruth C Fong and Andrea Vedaldi. Interpretable explanations of black boxes by meaningful perturbation. In Proceedings of the IEEE international conference on computer vision, pages 3429-3437, 2017. 3", + "[19] Ruth C Fong and Andrea Vedaldi. Interpretable explanations of black boxes by meaningful perturbation. In Proceedings of the IEEE international conference on computer vision, pages 3429-3437, 2017. 13", + "[20] Xiao Fu, Kejun Huang, Nicholas D Sidiropoulos, and Wing-Kin Ma. Nonnegative matrix factorization for signal and data analytics: Identifiability, algorithms, and applications. IEEE Signal Process. Mag., 36(2):59-80, 2019. 5", + "[21] Yunhao Ge, Yao Xiao, Zhi Xu, Meng Zheng, Srikrishna Karanam, Terrence Chen, Laurent Itti, and Ziyan Wu. A peek into the reasoning of neural networks: Interpreting with structural visual concepts. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2195-2204, 2021. 13", + "[22] Sahara Ghalebikesabi, Lucile Ter-Minassian, Karla DiazOrdaz, and Chris C Holmes. On locality of local explanation models. Advances in Neural Information Processing Systems (NeurIPS), 2021. 3", + "[23] Amirata Ghorbani, Abubakar Abid, and James Zou. Interpretation of neural networks is fragile. In Proceedings of the AAAI Conference on Artificial Intelligence (AAAI), 2017. 3", + "[24] Amirata Ghorbani, James Wexler, James Y Zou, and Been Kim. Towards automatic concept-based explanations. Advances in Neural Information Processing Systems, 32, 2019, 2, 3, 4, 5, 7, 8, 13, 14, 23", + "[25] Andreas Griewank and Andrea Walther. Evaluating derivatives: principles and techniques of algorithmic differentiation. SIAM, 2008. 3, 6, 19", + "[26] Peter Hase and Mohit Bansal. Evaluating explainable ai: Which algorithmic explanations help users predict model behavior? In Proceedings of the Annual Meeting of the Association for Computational Linguistics (ACL Short Papers), 2020. 2, 3" + ], + "bbox": [ + 501, + 92, + 893, + 907 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "2719", + "bbox": [ + 482, + 926, + 514, + 936 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[27] Peter Hase, Harry Xie, and Mohit Bansal. The out-of-distribution problem in explainability and search methods for feature importance explanations. Advances in Neural Information Processing Systems (NeurIPS), 2021. 2", + "[28] Johannes Haug, Stefan Zürn, Peter El-Jiz, and Gjergji Kasneci. On baselines for local feature attributions. arXiv preprint arXiv:2101.00905, 2021. 3", + "[29] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Identity mappings in deep residual networks. In European conference on computer vision, pages 630-645. Springer, 2016. 23", + "[30] Magnus R Hestenes and Eduard Stiefel. Methods of conjugate gradients for solving. Journal of research of the National Bureau of Standards, 49(6):409, 1952. 18, 19", + "[31] Cheng-Yu Hsieh, Chih-Kuan Yeh, Xuanqing Liu, Pradeep Ravikumar, Seungyeon Kim, Sanjiv Kumar, and Cho-Jui Hsieh. Evaluations and methods for explanation through robustness analysis. In Proceedings of the International Conference on Learning Representations (ICLR), 2021. 3", + "[32] Kejun Huang, Nicholas D Sidiropoulos, and Athanasios P Liavas. A flexible and efficient algorithmic framework for constrained matrix and tensor factorization. IEEE Transactions on Signal Processing, 64(19):5052-5065, 2016. 6, 18", + "[33] Marouane Il Idrissi, Vincent Chabridon, and Bertrand Iooss. Developments and applications of shapley effects to reliability-oriented sensitivity analysis with correlated inputs. Environmental Modelling & Software, 2021. 5", + "[34] Bertrand Iooss and Paul Lemaitre. A review on global sensitivity analysis methods. In Uncertainty management in simulation-optimization of complex systems, pages 101-122. Springer, 2015. 2, 5", + "[35] Alon Jacovi, Ana Marasovic, Tim Miller, and Yoav Goldberg. Formalizing trust in artificial intelligence: Prerequisites, causes and goals of human trust in ai. In Proceedings of the 2021 ACM conference on fairness, accountability, and transparency, pages 624-635, 2021. 1", + "[36] Alexandre Janon, Thierry Klein, Agnes Lagnoux, Maëlle Nodet, and Clémentine Prieur. Asymptotic normality and efficiency of two sobol index estimators. *ESAIM: Probability and Statistics*, 18:342–364, 2014. 5, 20", + "[37] Margot E Kaminski and Jennifer M Urban. The right to contest ai. Columbia Law Review, 121(7):1957-2048, 2021. 1", + "[38] William Karush. Minima of functions of several variables with inequalities as side constraints. M. Sc. Dissertation. Dept. of Mathematics, Univ. of Chicago, 1939. 6, 18", + "[39] Been Kim, Rajiv Khanna, and Oluwasanmi O Koyejo. Examples are not enough, learn to criticize! criticism for interpretability. Advances in Neural Information Processing Systems (NeurIPS), 2016. 3", + "[40] Been Kim, Martin Wattenberg, Justin Gilmer, Carrie Cai, James Wexler, Fernanda Viegas, et al. Interpretability beyond feature attribution: Quantitative testing with concept activation vectors (tcav). In International conference on machine learning, pages 2668-2677. PMLR, 2018. 2, 3, 5, 8, 16, 23", + "[41] Sunnie S. Y. Kim, Nicole Meister, Vikram V. Ramaswamy, Ruth Fong, and Olga Russakovsky. HIVE: Evaluating the human interpretability of visual explanations. In European Conference on Computer Vision (ECCV), 2022. 2, 3" + ], + "bbox": [ + 78, + 90, + 470, + 906 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[42] Pieter-Jan Kindermans, Sara Hooker, Julius Adebayo, Maximilian Alber, Kristof T. Schütt, Sven Dähne, Dumitru Erhan, and Been Kim. The (Un)reliability of Saliency Methods, pages 267-280. Springer International Publishing, Cham, 2019. 3", + "[43] Mauritz Kop. Eu artificial intelligence act: The European approach to ai. In Stanford - Vienna Transatlantic Technology Law Forum, Transatlantic Antitrust and IPR Developments, Stanford University, Issue No. 2/2021. https://law.stanford.edu/publications/eu-artificial-intelligence-act-the-european-approach-to-ai/. Stanford-Vienna Transatlantic Technology Law Forum, Transatlantic Antitrust, 2021. 1", + "[44] Steven George Krantz and Harold R Parks. The implicit function theorem: history, theory, and applications. Springer Science & Business Media, 2002. 3, 6, 19", + "[45] Harold W Kuhn and Albert W Tucker. Nonlinear programming proceedings of the second berkeley symposium on mathematical statistics and probability. Neyman, pages 481-492, 1951. 6, 18", + "[46] Daniel D Lee and H Sebastian Seung. Learning the parts of objects by non-negative matrix factorization. Nature, 401(6755):788-791, 1999. 2", + "[47] Scott M Lundberg and Su-In Lee. A unified approach to interpreting model predictions. Advances in neural information processing systems, 30, 2017. 3, 7", + "[48] Amandine Marrel, Bertrand Iooss, Beatrice Laurent, and Olivier Roustant. Calculations of sobol indices for the gaussian process metamodel. Reliability Engineering & System Safety, 94(3):742-751, 2009. 5", + "[49] Giang Nguyen, Daeyoung Kim, and Anh Nguyen. The effectiveness of feature attribution methods and its correlation with automatic evaluation scores. Advances in Neural Information Processing Systems (NeurIPS), 2021. 2, 3", + "[50] Paul Novello, Thomas Fel, and David Vigouroux. Making sense of dependence: Efficient black-box explanations using dependence measure. In Advances in Neural Information Processing Systems (NeurIPS), 2022. 1, 3", + "[51] Chris Olah, Alexander Mordvintsev, and Ludwig Schubert. Feature visualization. Distill, 2017. https://distill.pub/2017/feature-visualization. 13, 16", + "[52] Art B Owen. Better estimation of small sobol'sensitivity indices. ACM Transactions on Modeling and Computer Simulation (TOMACS), 23(2):1-17, 2013. 5", + "[53] Vardan Papyan, XY Han, and David L Donoho. Prevalence of neural collapse during the terminal phase of deep learning training. Proceedings of the National Academy of Sciences, 117(40):24652-24663, 2020. 5", + "[54] Fabian Pedregosa, Gael Varoquaux, Alexandre Gramfort, Vincent Michel, Bertrand Thirion, Olivier Grisel, Mathieu Blondel, Peter Prettenhofer, Ron Weiss, Vincent Dubourg, et al. Scikit-learn: Machine learning in python. Journal of machine learning research, 12(Oct):2825–2830, 2011. 6", + "[55] Vitali Petsiuk, Abir Das, and Kate Saenko. Rise: Randomized input sampling for explanation of black-box models. arXiv preprint arXiv:1806.07421, 2018. 1, 3, 7, 19, 23" + ], + "bbox": [ + 501, + 92, + 893, + 883 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "2720", + "bbox": [ + 482, + 926, + 514, + 936 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[56] Bin Ren, Laurent Pueyo, Christine Chen, Élodie Choquet, John H Debes, Gaspard Duchêne, François Ménard, and Marshall D Perrin. Using data imputation for signal separation in high-contrast imaging. The Astrophysical Journal, 892(2):74, 2020. 4", + "[57] Marco Tulio Ribeiro, Sameer Singh, and Carlos Guestrin. \" why should i trust you?\" explaining the predictions of any classifier. In Proceedings of the 22nd ACM SIGKDD international conference on knowledge discovery and data mining, pages 1135-1144, 2016. 3, 7, 19", + "[58] Andrea Saltelli, Paola Annoni, Ivano Azzini, Francesca Campolongo, Marco Ratto, and Stefano Tarantola. Variance based sensitivity analysis of model output. design and estimator for the total sensitivity index. Computer physics communications, 181(2):259-270, 2010. 5", + "[59] Ramprasaath R Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra. Grad-cam: Visual explanations from deep networks via gradient-based localization. In Proceedings of the IEEE international conference on computer vision, pages 618-626, 2017. 1, 3, 6, 7", + "[60] Hua Shen and Ting-Hao Huang. How useful are the machine-generated interpretations to general users? a human evaluation on guessing the incorrectly predicted labels. In Proceedings of the AAAI Conference on Human Computation and Crowdsourcing, volume 8, pages 168-172, 2020. 2, 3", + "[61] Avanti Shrikumar, Peyton Greenside, and Anshul Kundaje. Learning important features through propagating activation differences. In Proceedings of the International Conference on Machine Learning (ICML), 2017. 1, 7", + "[62] Karen Simonyan, Andrea Vedaldi, and Andrew Zisserman. Deep inside convolutional networks: Visualising image classification models and saliency maps. In In Workshop at International Conference on Learning Representations. CiteSeer, 2014. 1, 3, 6, 7", + "[63] Leon Sixt, Maximilian Granz, and Tim Landgraf. When explanations lie: Why many modified bp attributions fail. In Proceedings of the International Conference on Machine Learning (ICML), 2020. 2", + "[64] Leon Sixt, Martin Schuessler, Oana-Iuliana Popescu, Philipp Weiß, and Tim Landgraf. Do users benefit from interpretable vision? a user study, baseline, and dataset. Proceedings of the International Conference on Learning Representations (ICLR), 2022. 2", + "[65] Dylan Slack, Sophie Hilgard, Emily Jia, Sameer Singh, and Himabindu Lakkaraju. Fooling lime and shap: Adversarial attacks on post hoc explanation methods. In Proceedings of the AAAI/ACM Conference on AI, Ethics, and Society, pages 180-186, 2020. 2, 3" + ], + "bbox": [ + 78, + 90, + 470, + 781 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[66] Daniel Smilkov, Nikhil Thorat, Been Kim, Fernanda Viégas, and Martin Wattenberg. Smoothgrad: removing noise by adding noise. arXiv preprint arXiv:1706.03825, 2017. 1, 3, 5, 6, 7", + "[67] Ilya M Sobol. Sensitivity analysis for non-linear mathematical models. Mathematical modelling and computational experiment, 1:407-414, 1993. 2, 5, 20", + "[68] Ilya M Sobol. Global sensitivity indices for nonlinear mathematical models and their monte carlo estimates. Mathematics and computers in simulation, 55(1-3):271-280, 2001. 2, 5", + "[69] Jost Tobias Springenberg, Alexey Dosovitskiy, Thomas Brox, and Martin Riedmiller. Striving for simplicity: The all convolutional net. arXiv preprint arXiv:1412.6806, 2014. 1, 3, 6", + "[70] Pascal Sturmfels, Scott Lundberg, and Su-In Lee. Visualizing the impact of feature attribution baselines. Distill, 2020. 3, 13", + "[71] Mukund Sundararajan, Ankur Taly, and Qiqi Yan. Axiomatic attribution for deep networks. In International conference on machine learning, pages 3319-3328. PMLR, 2017. 1, 3, 5, 6, 7", + "[72] Stefano Tarantola, Debora Gatelli, and Thierry Alex Mara. Random balance designs for the estimation of first order global sensitivity indices. Reliability Engineering & System Safety, 91(6):717-727, 2006. 5", + "[73] Stephen A Vavasis. On the complexity of nonnegative matrix factorization. SIAM Journal on Optimization, 20(3):1364-1377, 2010. 6", + "[74] Yu-Xiong Wang and Yu-Jin Zhang. Nonnegative matrix factorization: A comprehensive review. IEEE Transactions on Knowledge and Data Engineering, 25(6):1336-1353, 2013. 5", + "[75] Peter Wilf, Shengping Zhang, Sharat Chikkerur, Stefan A Little, Scott L Wing, and Thomas Serre. Computer vision cracks the leaf code. Proceedings of the National Academy of Sciences, 113(12):3305-3310, 2016. 7", + "[76] Matthew D Zeiler and Rob Fergus. Visualizing and understanding convolutional networks. In European conference on computer vision, pages 818-833. Springer, 2014. 1, 3, 7", + "[77] Ruihan Zhang, Prashan Madumal, Tim Miller, Krista A Ehinger, and Benjamin IP Rubinstein. Invertible concept-based explanations for cnn models with non-negative concept activation vectors. arXiv preprint arXiv:2006.15417, 2020. 2, 3, 5, 8, 23", + "[78] Bolei Zhou, Aditya Khosla, Agata Lapedriza, Aude Oliva, and Antonio Torralba. Learning deep features for discriminative localization. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2921-2929, 2016. 3, 6" + ], + "bbox": [ + 501, + 92, + 893, + 782 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "2721", + "bbox": [ + 482, + 926, + 513, + 936 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2023/CRAFT_ Concept Recursive Activation FacTorization for Explainability/6e462b05-7931-4758-8402-5b3e432aab49_model.json b/2023/CRAFT_ Concept Recursive Activation FacTorization for Explainability/6e462b05-7931-4758-8402-5b3e432aab49_model.json new file mode 100644 index 0000000000000000000000000000000000000000..a92a12f70225dc3669b26cab66f2a5d557346c0d --- /dev/null +++ b/2023/CRAFT_ Concept Recursive Activation FacTorization for Explainability/6e462b05-7931-4758-8402-5b3e432aab49_model.json @@ -0,0 +1,2400 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.121, + 0.131, + 0.844, + 0.153 + ], + "angle": 0, + "content": "CRAFT: Concept Recursive Activation FacTorization for Explainability" + }, + { + "type": "title", + "bbox": [ + 0.19, + 0.162, + 0.791, + 0.223 + ], + "angle": 0, + "content": "Thomas Fel\\(^{1,3,5*}\\) Agustin Picard\\(^{3,6*}\\) Louis Bethune\\(^{3*}\\) Thibaut Boissin\\(^{3,4*}\\) David Vigouroux\\(^{3,4}\\) Julien Colin\\(^{1,3}\\) Rémi Cadène\\(^{1,2}\\) Thomas Serre\\(^{1,3}\\)" + }, + { + "type": "text", + "bbox": [ + 0.16, + 0.228, + 0.81, + 0.244 + ], + "angle": 0, + "content": "\\(^{1}\\)Carney Institute for Brain Science, Brown University, USA \\(^{2}\\)Sorbonne Université, CNRS, France" + }, + { + "type": "text", + "bbox": [ + 0.204, + 0.246, + 0.766, + 0.262 + ], + "angle": 0, + "content": "3Artificial and Natural Intelligence Toulouse Institute, Université de Toulouse, France" + }, + { + "type": "text", + "bbox": [ + 0.285, + 0.264, + 0.687, + 0.28 + ], + "angle": 0, + "content": "4 Institut de Recherche Technologique Saint-Exupery, France" + }, + { + "type": "text", + "bbox": [ + 0.307, + 0.281, + 0.657, + 0.297 + ], + "angle": 0, + "content": "\\(^{5}\\) Innovation & Research Division, SNCF, \\(^{6}\\) Scalian" + }, + { + "type": "image", + "bbox": [ + 0.158, + 0.308, + 0.396, + 0.491 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.406, + 0.308, + 0.624, + 0.49 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.625, + 0.31, + 0.673, + 0.348 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.683, + 0.312, + 0.73, + 0.348 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.74, + 0.312, + 0.785, + 0.348 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.626, + 0.353, + 0.673, + 0.391 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.683, + 0.353, + 0.729, + 0.391 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.74, + 0.353, + 0.785, + 0.391 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.624, + 0.407, + 0.672, + 0.445 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.678, + 0.408, + 0.725, + 0.444 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.736, + 0.408, + 0.784, + 0.445 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.624, + 0.45, + 0.672, + 0.486 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.68, + 0.45, + 0.725, + 0.486 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.736, + 0.45, + 0.784, + 0.488 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.506, + 0.895, + 0.59 + ], + "angle": 0, + "content": "Figure 1. The \"Man on the Moon\" incorrectly classified as a \"shovel\" by an ImageNet-trained ResNet50. Heatmap generated by a classic attribution method [55] (left) vs. concept attribution maps generated with the proposed CRAFT approach (right) which highlights the two most influential concepts that drove the ResNet50's decision along with their corresponding locations. CRAFT suggests that the neural net arrived at its decision because it identified the concept of \"dirt\" \\(\\bullet\\) commonly found in members of the image class \"shovel\" and the concept of \"ski pants\" \\(\\bullet\\) typically worn by people clearing snow from their driveway with a shovel instead the correct concept of astronaut's pants (which was probably never seen during training)." + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.608, + 0.314, + 0.623 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.628, + 0.473, + 0.869 + ], + "angle": 0, + "content": "Attribution methods, which employ heatmaps to identify the most influential regions of an image that impact model decisions, have gained widespread popularity as a type of explainability method. However, recent research has exposed the limited practical value of these methods, attributed in part to their narrow focus on the most prominent regions of an image - revealing \"where\" the model looks, but failing to elucidate \"what\" the model sees in those areas. In this work, we try to fill in this gap with CRAFT - a novel approach to identify both \"what\" and \"where\" by generating concept-based explanations. We introduce 3 new ingredients to the automatic concept extraction literature: (i) a recursive strategy to detect and decompose concepts across layers, (ii) a novel method for a more faithful estimation of concept importance using Sobol indices, and (iii) the use of implicit differentiation to unlock Concept Attribution Maps." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.472, + 0.902 + ], + "angle": 0, + "content": "We conduct both human and computer vision experiments to demonstrate the benefits of the proposed approach." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.609, + 0.895, + 0.7 + ], + "angle": 0, + "content": "We show that the proposed concept importance estimation technique is more faithful to the model than previous methods. When evaluating the usefulness of the method for human experimenters on a human-centered utility benchmark, we find that our approach significantly improves on two of the three test scenarios." + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.711, + 0.633, + 0.727 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.734, + 0.895, + 0.87 + ], + "angle": 0, + "content": "Interpreting the decisions of modern machine learning models such as neural networks remains a major challenge. Given the ever-increasing range of machine learning applications, the need for robust and reliable explainability methods continues to grow [11, 35]. Recently enacted European laws (including the General Data Protection Regulation (GDPR) [37] and the European AI act [43]) require the assessment of explainable decisions, especially those made by algorithms." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.871, + 0.895, + 0.902 + ], + "angle": 0, + "content": "In order to try to meet this growing need, an array of explainability methods have already been proposed [14, 50," + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.925, + 0.514, + 0.937 + ], + "angle": 0, + "content": "2711" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.191, + 0.096, + 0.57, + 0.258 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.582, + 0.108, + 0.783, + 0.242 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.075, + 0.265, + 0.892, + 0.363 + ], + "angle": 0, + "content": "Figure 2. CRAFT results for the prediction \"chain saw\". First, our method uses Non-Negative Matrix Factorization (NMF) to extract the most relevant concepts used by the network (ResNet50V2) from the train set (ILSVRC2012 [10]). The global influence of these concepts on the predictions is then measured using Sobol indices (right panel). Finally, the method provides local explanations through concept attribution maps (heatmaps associated with a concept, and computed using grad-CAM by backpropagating through the NMF concept values with implicit differentiation). Besides, concepts can be interpreted by looking at crops that maximize the NMF coefficients. For the class \"chain saw\", the detected concepts seem to be: \\(\\bullet\\) the chainsaw engine, \\(\\bullet\\) the saw blade, \\(\\bullet\\) the human head, \\(\\bullet\\) the vegetation, \\(\\bullet\\) the jeans and \\(\\bullet\\) the tree trunk." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.375, + 0.473, + 0.629 + ], + "angle": 0, + "content": "59, 61, 62, 66, 69, 71, 76]. One of the main classes of methods called attribution methods yields heatmaps that indicate the importance of individual pixels for driving a model's decision. However, these methods exhibit critical limitations [1, 27, 63, 65], as they have been shown to fail – or only marginally help – in recent human-centered benchmarks [8, 26, 41, 49, 60, 64]. It has been suggested that their limitations stem from the fact that they are only capable of explaining where in an image are the pixels that are critical to the decision but they cannot tell what visual features are actually driving decisions at these locations. In other words, they show where the model looks but not what it sees. For example, in the scenario depicted in Fig. 1, where an ImageNet-trained ResNet mistakenly identifies an image as containing a shovel, the attribution map displayed on the left fails to explain the reasoning behind this misclassification." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.63, + 0.473, + 0.902 + ], + "angle": 0, + "content": "A recent approach has sought to move past attribution methods [40] by using so-called \"concepts\" to communicate information to users on how a model works. The goal is to find human-interpretable concepts in the activation space of a neural network. Although the approach exhibited potential, its practicality is significantly restricted due to the need for prior knowledge of pertinent concepts in its original formulation and, more critically, the requirement for a labeled dataset of such concepts. Several lines of work have focused on trying to automate the concept discovery process based only on the training dataset and without explicit human supervision. The most prominent of these techniques, ACE [24], uses a combination of segmentation and clustering techniques but requires heuristics to remove outliers. However, ACE provides a proof of concept that it might be possible to discover concepts automatically and at scale – without additional labeling or human supervision. Nevertheless, the approach suffers several limitations: by" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.375, + 0.895, + 0.481 + ], + "angle": 0, + "content": "construction, each image segment can only belong to a single cluster, a layer has to be selected by the user to be used to retrieve the relevant concepts, and the amount of information lost during the outlier rejection phase can be a cause of concern. More recently, Zhang et al. [77] proposes to leverage matrix decompositions on internal feature maps to discover concepts." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.483, + 0.895, + 0.604 + ], + "angle": 0, + "content": "Here, we try to fill these gaps with a novel method called CRAFT which uses Non-Negative Matrix Factorization (NMF) [46] for concept discovery. In contrast to other concept-based explanation methods, our approach provides an explicit link between their global and local explanations (Fig. 2) and identifies the relevant layer(s) to use to represent individual concepts (Fig. 3). Our main contributions can be described as follows:" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.607, + 0.894, + 0.667 + ], + "angle": 0, + "content": "(i) A novel approach for the automated extraction of high-level concepts learned by deep neural networks. We validate its practical utility to users with human psychophysics experiments." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.669, + 0.895, + 0.805 + ], + "angle": 0, + "content": "(ii) A recursive procedure to automatically identify concepts and sub-concepts at the right level of granularity – starting with our decomposition at the top of the model and working our way upstream. We validate the benefit of this approach with human psychophysics experiments showing that (i) the decomposition of a concept yields more coherent sub-concepts and (ii) that the groups of points formed by these sub-concepts are more refined and appear meaningful to humans." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.808, + 0.895, + 0.868 + ], + "angle": 0, + "content": "(iii) A novel technique to quantify the importance of individual concepts for a model's prediction using Sobol indices [34, 67, 68] – a technique borrowed from Sensitivity Analysis." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.871, + 0.895, + 0.902 + ], + "angle": 0, + "content": "(iv) The first concept-based explainability method which produces concept attribution maps by backpropagating con" + }, + { + "type": "list", + "bbox": [ + 0.498, + 0.607, + 0.895, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.925, + 0.516, + 0.937 + ], + "angle": 0, + "content": "2712" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.471, + 0.184 + ], + "angle": 0, + "content": "cept scores into the pixel space by leveraging the implicit function theorem in order to localize the pixels associated with the concept of a given input image. This effectively opens up the toolbox of both white-box [15, 59, 62, 66, 69, 71, 78] and black-box [14, 47, 55, 57] explainability methods to derive concept-wise attribution maps." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.198, + 0.22, + 0.213 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.224, + 0.471, + 0.495 + ], + "angle": 0, + "content": "Attribution methods Attribution methods are widely used as post-hoc explainability techniques to determine the input variables that contribute to a model's prediction by generating importance maps, such as the ones shown in Fig.1. The first attribution method, Saliency, introduced in [62], generates a heatmap by utilizing the gradient of a given classification score with respect to the pixels. This method was later improved upon in the context of deep convolutional networks for classification in subsequent studies, such as [66, 69, 71, 76]. However, the image gradient only reflects the model's operation within an infinitesimal neighborhood around an input, which can yield misleading importance estimates [22] since gradients of large vision models are notoriously noisy [66]. That is why several methods leverage perturbations on the input image to probe the model and create importance maps that indicate the most crucial areas for the decision, such as Rise [55], Sobol [14], or more recently HSIC [50]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.497, + 0.472, + 0.618 + ], + "angle": 0, + "content": "Unfortunately, a severe limitation of these approaches – apart from the fact that they only show the “where” – is that they are subject to confirmation bias: while they may appear to offer useful explanations to a user, sometimes these explanations are actually incorrect [1, 23, 65]. These limitations raise questions about their usefulness, as recent research has shown by using human-centered experiments to evaluate the utility of attribution [8, 26, 41, 49, 60]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.619, + 0.472, + 0.844 + ], + "angle": 0, + "content": "In particular, in [8], a protocol is proposed to measure the usefulness of explanations, corresponding to how much they help users identify rules driving a model's predictions (correct or incorrect) that transfer to unseen data - using the concept of meta-predictor (also called simulatability) [11, 18, 39]. The main idea is to train users to predict the output of the system using a small set of images along with associated model predictions and corresponding explanations. A method that performs well on this benchmark is said useful, as it helps users better predict the output of the model by providing meaningful information about the internal functioning of the model. This framework being agnostic to the type of explainability method, we have chosen to use it in Section 4 in order to compare CRAFT with attribution methods." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.856, + 0.471, + 0.901 + ], + "angle": 0, + "content": "Concepts-based methods Kim et al. [40] introduced a method aimed at providing explanations that go beyond attribution-based approaches by measuring the impact of" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.895, + 0.424 + ], + "angle": 0, + "content": "pre-selected concepts on a model's outputs. Although this method appears more interpretable to human users than standard attribution techniques, it requires a database of images describing the relevant concepts to be manually curated. Ghorbani et al. [24] further extended the approach to extract concepts without the need for human supervision. The approach, called ACE [24], uses a segmentation scheme on images, that belong to an image class of interest. The authors leveraged the intermediate activations of a neural network for specific image segments. These segments were resized to the appropriate input size and filled with a baseline value. The resulting activations were clustered to produce prototypes, which they referred to as \"concepts\". However, some concepts contained background segments, leading to the inclusion of uninteresting and outlier concepts. To address this, the authors implemented a post-processing cleanup step to remove these concepts, including those that were present in only one image of the class and were not representative. While this improved the interpretability of their explanations to human subjects, the use of a baseline value filled around the segments could introduce biases in the explanations [28, 31, 42, 70]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.424, + 0.895, + 0.545 + ], + "angle": 0, + "content": "Zhang et al. [77] developed a solution to the unsupervised concept discovery problem by using matrix factorizations in the latent spaces of neural networks. However, one major drawback of this method is that it operates at the level of convolutional kernels, leading to the discovery of localized concepts. For example, the concept of \"grass\" at the bottom of the image is considered to be distinct from the concept of \"grass\" at the top of the image." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.551, + 0.725, + 0.567 + ], + "angle": 0, + "content": "3. Overview of the method" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.577, + 0.895, + 0.713 + ], + "angle": 0, + "content": "In this section, we first describe our concept activations factorization method. Below we highlight the main differences with related work. We then proceed to introduce the three novel ingredients that make up CRAFT: (1) a method to recursively decompose concepts into sub-concepts, (2) a method to better estimate the importance of extracted concepts, and (3) a method to use any attribution method to create concept attribution maps, using implicit differentiation [4, 25, 44]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.897, + 0.901 + ], + "angle": 0, + "content": "Notations In this work, we consider a general supervised learning setting, where \\((\\pmb{x}_1,\\dots,\\pmb{x}_n)\\in \\mathcal{X}^n\\subseteq \\mathbb{R}^{n\\times d}\\) are \\(n\\) inputs images and \\((y_{1},\\ldots ,y_{n})\\in \\mathcal{V}^{n}\\) their associated labels. We are given a (machine-learnt) black-box predictor \\(\\pmb {f}:\\mathcal{X}\\rightarrow \\mathcal{Y}\\), which at some test input \\(\\pmb{x}\\) predicts the output \\(\\pmb {f}(\\pmb {x})\\). Without loss of generality, we establish that \\(\\pmb{f}\\) is a neural network that can be decomposed into two distinct components. The first component is a function \\(\\pmb{g}\\) that maps the input to an intermediate state, and the second component is \\(\\pmb{h}\\), which takes this intermediate state to the output, such that \\(\\pmb {f}(\\pmb {x}) = (\\pmb {h}\\circ \\pmb {g})(\\pmb {x})\\). In this context, \\(\\pmb {g}(\\pmb {x})\\subseteq \\mathbb{R}^{p}\\) represents the intermediate activations of \\(\\pmb{x}\\) within the network." + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.926, + 0.516, + 0.938 + ], + "angle": 0, + "content": "2713" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.126, + 0.092, + 0.291, + 0.202 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.293, + 0.092, + 0.461, + 0.2 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.463, + 0.092, + 0.847, + 0.212 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.225, + 0.895, + 0.312 + ], + "angle": 0, + "content": "Figure 3. (1) Neural collapse (amalgamation). A classifier needs to be able to linearly separate classes by the final layer. It is commonly assumed that in order to achieve this, image activations from the same class get progressively \"merged\" such that these image activations converge to a one-hot vector associated with the class at the level of the logits layer. In practice, this means that different concepts get ultimately blended together along the way. (2) Recursive process. When a concept is not understood (e.g., \\(\\mathcal{C}\\)), we propose to decompose it into multiple sub-concepts (e.g., \\(\\mathcal{C}_1, \\mathcal{C}_2, \\mathcal{C}_3\\)) using the activations from an earlier layer to overcome the aforementioned neural collapse issue. (3) Example of recursive concept decomposition using CRAFT on the ImageNet class \"parachute\"." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.323, + 0.471, + 0.385 + ], + "angle": 0, + "content": "Further, we will assume non-negative activations: \\( g(\\boldsymbol{x}) \\geq 0 \\). In particular, this assumption is verified by any architecture that utilizes \\( ReLU \\), but any non-negative activation function works." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.394, + 0.366, + 0.41 + ], + "angle": 0, + "content": "3.1. Concept activation factorization." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.417, + 0.471, + 0.478 + ], + "angle": 0, + "content": "We use Non-negative matrix factorization to identify a basis for concepts based on a network's activations (Fig.4). Inspired by the approach taken in ACE [24], we will use image sub-regions to try to identify coherent concepts." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.479, + 0.47, + 0.644 + ], + "angle": 0, + "content": "The first step involves gathering a set of images that one wishes to explain, such as the dataset, in order to generate associated concepts. In our examples, to explain a specific class \\( y \\in \\mathcal{V} \\), we selected the set of points \\( \\mathcal{C} \\) from the dataset for which the model's predictions matched a specific class \\( \\mathcal{C} = \\{\\pmb{x}_i : \\pmb{f}(\\pmb{x}_i) = y, 1 \\leq i \\leq n\\} \\). It is important to emphasize that this choice is significant. The goal is not to understand how humans labeled the data, but rather to comprehend the model itself. By only selecting correctly classified images, important biases and failure cases may be missed, preventing a complete understanding of our model." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.645, + 0.471, + 0.886 + ], + "angle": 0, + "content": "Now that we have defined our set of images, we will proceed with selecting sub-regions of those images to identify specific concepts within a localized context. It has been observed that the implementation of segmentation masks suggested in ACE can lead to the introduction of artifacts due to the associated inpainting with a baseline value. In contrast, our proposed method takes advantage of the prevalent use of modern data augmentation techniques such as randaugment, mixup, and cutmix during the training of current models. These techniques involve the current practice of models being trained on image crops, which enables us to leverage a straightforward crop and resize function denoted by \\(\\pi (\\cdot)\\) to create sub-regions (illustrated in Fig.4). By applying \\(\\pi\\) function to each image in the set \\(\\mathcal{C}\\), we obtain an auxiliary dataset \\(\\mathbf{X}\\in \\mathbb{R}^{n\\times d}\\) such that each entries \\(\\mathbf{X}_i = \\pi (\\mathbf{x}_i)\\) is an image crop." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.886, + 0.471, + 0.902 + ], + "angle": 0, + "content": "To discover the concept basis, we start by obtaining the" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.323, + 0.892, + 0.369 + ], + "angle": 0, + "content": "activations for the random crops \\(\\mathbf{A} = g(\\mathbf{X})\\in \\mathbb{R}^{n\\times p}\\). In the case where \\(\\pmb{f}\\) is a convolutional neural network, a global average pooling is applied to the activations." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.369, + 0.894, + 0.429 + ], + "angle": 0, + "content": "We are now ready to apply Non-negative Matrix Factorization (NMF) to decompose positive activations \\(\\mathbf{A}\\) into a product of non-negative, low-rank matrices \\(\\mathbf{U} \\in \\mathbb{R}^{n \\times r}\\) and \\(\\mathbf{W} \\in \\mathbb{R}^{p \\times r}\\) by solving:" + }, + { + "type": "equation", + "bbox": [ + 0.56, + 0.429, + 0.892, + 0.461 + ], + "angle": 0, + "content": "\\[\n\\left(\\mathbf {U}, \\mathbf {W}\\right) = \\underset {\\mathbf {U} \\geq 0, \\mathbf {W} \\geq 0} {\\arg \\min } \\frac {1}{2} \\| \\mathbf {A} - \\mathbf {U W} ^ {\\mathsf {T}} \\| _ {F} ^ {2}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.463, + 0.78, + 0.478 + ], + "angle": 0, + "content": "where \\(||\\cdot ||_F\\) denotes the Frobenius norm." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.479, + 0.895, + 0.75 + ], + "angle": 0, + "content": "This decomposition of our activations \\(\\mathbf{A}\\) yields two matrices: \\(\\mathbf{W}\\) containing our Concept Activation Vectors (CAVs) and \\(\\mathbf{U}\\) that redefines the data points in our dataset according to this new basis. Moreover, this decomposition in this new basis has some interesting properties that go beyond the simple low-rank factorization - since \\(r \\ll \\min(n, p)\\). First, NMF can be understood as the joint learning of a dictionary of Concept Activation Vectors - called a \"concept bank\" in Fig. 4 - that maps a \\(\\mathbb{R}^p\\) basis onto \\(\\mathbb{R}^r\\), and \\(\\mathbf{U}\\) the coefficients of the vectors \\(\\mathbf{A}\\) expressed in this new basis. The minimization of the reconstruction error \\(\\frac{1}{2} \\| \\mathbf{A} - \\mathbf{U} \\mathbf{W} \\|_F^2\\) ensures that the new basis contains (mostly) relevant concepts. Intuitively, the non-negativity constraints \\(\\mathbf{U} \\geq 0\\), \\(\\mathbf{W} \\geq 0\\) encourage \\((i)\\) \\(\\mathbf{W}\\) to be sparse (useful for creating disentangled concepts), \\((ii)\\) \\(\\mathbf{U}\\) to be sparse (convenient for selecting a minimal set of useful concepts) and \\((iii)\\) missing data to be imputed [56], which corresponds to the sparsity pattern of post-ReLU activations \\(\\mathbf{A}\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.895, + 0.886 + ], + "angle": 0, + "content": "It is worth noting that each input \\( \\mathbf{x}_i \\) can be expressed as a linear combination of concepts denoted as \\( \\mathbf{A}_i = \\sum_{j=1}^{r} \\mathbf{U}_{(i,j)} \\mathbf{W}_j^{\\top} \\). This approach is advantageous because it allows us to interpret each input as a composition of the underlying concepts. Furthermore, the strict positivity of each term - NMF is working over the anti-negative semiring, - enhances the interpretability of the decomposition. Another interesting interpretation could be that each input is represented as a superposition of concepts [13]." + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.886, + 0.892, + 0.901 + ], + "angle": 0, + "content": "While other methods in the literature solve a similar" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.925, + 0.516, + 0.937 + ], + "angle": 0, + "content": "2714" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.213 + ], + "angle": 0, + "content": "problem (such as low-rank factorization using SVD or ICA), the NMF is both fast and effective and is known to yield concepts that are meaningful to humans [20,74,77]. Finally, once the concept bank \\(\\mathbf{W}\\) has been precomputed, we can associate the concept coefficients \\(\\pmb{u}\\) to any new input \\(\\pmb{x}\\) (e.g., a full image) by solving the underlying Non-Negative Least Squares (NNLS) problem \\(\\min_{\\pmb{u} \\geq 0} \\frac{1}{2} \\| \\pmb{g}(\\pmb{x}) - \\pmb{u} \\mathbf{W}^{\\top} \\|_F^2\\), and therefore recover its decomposition in the concept basis." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.217, + 0.473, + 0.323 + ], + "angle": 0, + "content": "In essence, the core of our method can be summarized as follows: using a set of images, the idea is to re-interpret their embedding at a given layer as a composition of concepts that humans can easily understand. In the next section, we show how one can recursively apply concept activation factorizations to preceding layer for an image containing a previously computed concept." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.339, + 0.388, + 0.357 + ], + "angle": 0, + "content": "3.2. Ingredient 1: A pinch of recursivity" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.366, + 0.473, + 0.548 + ], + "angle": 0, + "content": "One of the most apparent issues in previous work [24,77] is the need for choosing a priori a layer at which the activation maps are computed. This choice will critically affect the concepts that are identified because certain concepts get amalgamated [53] into one at different layers of the neural network, resulting in incoherent and indecipherable clusters, as illustrated in Fig. 3. We posit that this can be solved by iteratively applying our decomposition at different layer depths, and for the concepts that remain difficult to understand, by looking for their sub-concepts in earlier layers by isolating the images that contain them. This allows us to build hierarchies of concepts for each class." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.551, + 0.473, + 0.734 + ], + "angle": 0, + "content": "We offer a simple solution consisting of reapplying our method to a concept by performing a second step of concept activation factorization on a set of images that contain the concept \\(\\mathcal{C}\\) in order to refine it and create sub-concepts (e.g., decompose \\(\\mathcal{C}\\) into \\(\\{\\mathcal{C}_1, \\mathcal{C}_2, \\mathcal{C}_3\\}\\)) see Fig. 3 for an illustrative example. Note that we generalize current methods in the sense that taking images \\((x_1, \\dots, x_n)\\) that are clustered in the logits layer (belonging to the same class) and decomposing them in a previous layer - as done in [24, 77] - is a valid recursive step. For a more general case, let us assume that a set of images that contain a common concept is obtained using the first step of concept activation factorization." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.736, + 0.473, + 0.903 + ], + "angle": 0, + "content": "We will then take a subset of the auxiliary dataset points to refine any concept \\( j \\). To do this, we select the subset of points that contain the concept \\( \\mathcal{C}_j = \\{\\mathbf{X}_i : \\mathbf{U}_{(i,j)} > \\lambda_j, 1 \\leq i \\leq n\\} \\), where \\( \\lambda_j \\) is the 90th percentile of the values of the concept \\( \\mathbf{U}_{(1,j)}, \\ldots, \\mathbf{U}_{(n,j)} \\). In other words, the 10% of images that activate the concept \\( j \\) the most are selected for further refinement into sub-concepts. Given this new set of points, we can then re-apply the Concept Matrix Factorization method to an earlier layer \\( g'(\\cdot) \\) to obtain the sub-concepts decomposition from the initial concept - as illustrated in Fig.3." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.091, + 0.864, + 0.108 + ], + "angle": 0, + "content": "3.3. Ingredient 2: A dash of sensitivity analysis" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.115, + 0.895, + 0.326 + ], + "angle": 0, + "content": "A major concern with concept extraction methods is that concepts that make sense to humans are not necessarily the same as those being used by a model to classify images. In order to prevent such confirmation bias during our concept analysis phase, a faithful estimate the overall importance of the extracted concepts is crucial. Kim et al. [40] proposed an importance estimator based on directional derivatives: the partial derivative of the model output with respect to the vector of concepts. While this measure is theoretically grounded, it relies on the same principle as gradient-based methods, and thus, suffers from the same pitfalls: neural network models have noisy gradients [66, 71]. Hence, the farther the chosen layer is from the output, the noisier the directional derivative score will be." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.328, + 0.897, + 0.659 + ], + "angle": 0, + "content": "Since we essentially want to know which concept has the greatest effect on the output of the model, it is natural to consider the field of sensitivity analysis [9,33,34,67,68]. In this section, we briefly recall the classic \"total Sobol indices\" and how to apply them to our problem. The complete derivation of the Sobol-Hoeffding decomposition is presented in Section D of the supplementary materials. Formally, a natural way to estimate the importance of a concept \\( i \\) is to measure the fluctuations of the model's output \\( h(\\mathbf{U}\\mathbf{W}^{\\top}) \\) in response to meaningful perturbations of the concept coefficient \\( \\mathbf{U}_{(1,i)},\\ldots ,\\mathbf{U}_{(n,i)} \\). Concretely, we will use perturbation masks \\( \\mathbf{M} = (M_1,\\dots,M_r)\\sim \\mathcal{U}([0,1]^r) \\), here an i.i.d sequence of real-valued random variables, we introduce a concept fluctuation to reconstruct a perturbed activation \\( \\tilde{\\mathbf{A}} = (\\mathbf{U}\\odot \\mathbf{M})\\mathbf{W}^{\\top} \\) where \\( \\odot \\) denote the Hadamard product (e.g., the masks can be used to remove a concept by setting its value to zero). We can then propagate this perturbed activation to the model output \\( \\mathbf{Y} = h(\\tilde{\\mathbf{A}}) \\). Simply put, removing or applying perturbation of an important concept will result in a substantial variation in the output, whereas an unused concept will have minimal effect on the output." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.66, + 0.895, + 0.751 + ], + "angle": 0, + "content": "Finally, we can capture the importance that a concept might have as a main effect - along with its interactions with other concepts - on the model's output by calculating the expected variance that would remain if all the concepts except the \\(i\\) were to be fixed. This yields the general definition of the total Sobol indices." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.759, + 0.895, + 0.821 + ], + "angle": 0, + "content": "Definition 3.1 (Total Sobol indices). The total Sobol index \\( S_{i}^{T} \\), which measures the contribution of a concept \\( i \\) as well as its interactions of any order with any other concepts to the model output variance, is given by:" + }, + { + "type": "equation", + "bbox": [ + 0.541, + 0.828, + 0.895, + 0.864 + ], + "angle": 0, + "content": "\\[\n\\mathcal {S} _ {i} ^ {T} = \\frac {\\mathbb {E} _ {\\boldsymbol {M} _ {\\sim i}} (\\mathbb {V} _ {M _ {i}} (\\boldsymbol {h} ((\\boldsymbol {U} \\odot \\boldsymbol {M}) \\mathbf {W} ^ {\\top}) | \\boldsymbol {M} _ {\\sim i}))}{\\mathbb {V} (\\boldsymbol {h} ((\\boldsymbol {U} \\odot \\boldsymbol {M}) \\mathbf {W} ^ {\\top}))}. \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.871, + 0.897, + 0.903 + ], + "angle": 0, + "content": "In practice, this index can be calculated very efficiently [36, 48, 52, 58, 72], more details on the Quasi" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.926, + 0.516, + 0.938 + ], + "angle": 0, + "content": "2715" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.149, + 0.098, + 0.831, + 0.245 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.075, + 0.254, + 0.895, + 0.352 + ], + "angle": 0, + "content": "Figure 4. Overview of CRAFT. Starting from a set of crops \\(\\mathbf{X}\\) containing a concept \\(\\mathcal{C}\\) (e.g., crops images of the class \"parachute\"), we compute activations \\(g(\\mathbf{X})\\) corresponding to an intermediate layer from a neural network for random image crops. We then factorize these activations into two lower-rank matrices, \\((\\mathbf{U}, \\mathbf{W})\\). \\(\\mathbf{W}\\) is what we call a \"concept bank\" and is a new basis used to express the activations, while \\(\\mathbf{U}\\) corresponds to the corresponding coefficients in this new basis. We then extend the method with 3 new ingredients: (1) recursivity - by proposing to re-decompose a concept (e.g., take a new set of images containing \\(\\mathcal{C}_1\\)) at an earlier layer, (2) a better importance estimation using Sobol indices and (3) an approach to leverage implicit differentiation to generate concept attribution maps to localize concepts in an image." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.366, + 0.471, + 0.398 + ], + "angle": 0, + "content": "Monte Carlo sampling and the estimator used are left in appendix D." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.408, + 0.466, + 0.438 + ], + "angle": 0, + "content": "3.4. Ingredient 3: A smidgen of implicit differentiation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.44, + 0.471, + 0.65 + ], + "angle": 0, + "content": "Attribution methods are useful for determining the regions deemed important by a model for its decision, but they lack information about what exactly triggered it. We have seen that we can already extract this information from the matrices \\(\\mathbf{U}\\) and \\(\\mathbf{W}\\), but as it is, we do not know in what part of an image a given concept is represented. In this section, we will show how we can leverage attribution methods (forward and backward modes) to find where a concept is located in the input image (see Fig. 2). Forward attribution methods do not rely on any gradient computation as they only use inference processes, whereas backward methods require back-propagating through a network's layers. By application of the chain rule, computing \\(\\partial \\mathbf{U} / \\partial \\mathbf{X}\\) requires access to \\(\\partial \\mathbf{U} / \\partial \\mathbf{A}\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.651, + 0.472, + 0.847 + ], + "angle": 0, + "content": "To do so, one could be tempted to solve the linear system \\(\\mathbf{U}\\mathbf{W}^{\\top} = \\mathbf{A}\\). However, this problem is ill-posed since \\(\\mathbf{W}^{\\top}\\) is low rank. A standard approach is to calculate the Moore-Penrose pseudo-inverse \\((\\mathbf{W}^{\\top})^{\\dagger}\\), which solves rank deficient systems by looking at the minimum norm solution [2]. In practice, \\((\\mathbf{W}^{\\top})^{\\dagger}\\) is computed with the Singular Value Decomposition (SVD) of \\(\\mathbf{W}^{\\top}\\). Unfortunately, SVD is also the solution to the unstructured minimization of \\(\\frac{1}{2}\\|\\mathbf{A} - \\mathbf{U}\\mathbf{W}^{\\top}\\|_F^2\\) by the Eckart-Young-Mirsky theorem [12]. Hence, the non-negativity constraints of the NMF are ignored, which prevents such approaches from succeeding. Other issues stem from the fact that the \\(\\mathbf{U},\\mathbf{W}\\) decomposition is generally not unique." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.848, + 0.472, + 0.907 + ], + "angle": 0, + "content": "Our third contribution consists of tackling this problem to allow the use of attribution methods, i.e., concept attribution maps, by proposing a strategy to differentiate through the NMF block." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.366, + 0.895, + 0.604 + ], + "angle": 0, + "content": "Implicit differentiation of NMF block The NMF problem 1 is NP-hard [73], and it is not convex with respect to the input pair \\((\\mathbf{U},\\mathbf{W})\\). However, fixing the value of one of the two factors and optimizing the other turns the NMF formulation into a pair of Non-Negative Least Squares (NNLS) problems, which are convex. This ensures that alternating minimization (a standard approach for NMF) of \\((\\mathbf{U},\\mathbf{W})\\) factors will eventually reach a local minimum. Each of this alternating NNLS problems fulfills the Karush-Kuhn-Tucker (KKT) conditions [38,45], which can be encoded in the so-called optimality function \\(\\mathbf{F}\\) from [4], see Eq. 9 Appendix C.2. The implicit function theorem [25] allows us to use implicit differentiation [3,25,44] to efficiently compute the Jacobians \\(\\partial \\mathbf{U} / \\partial \\mathbf{A}\\) and \\(\\partial \\mathbf{W} / \\partial \\mathbf{A}\\) without requiring to back-propagate through each of the iterations of the NMF solver:" + }, + { + "type": "equation", + "bbox": [ + 0.574, + 0.601, + 0.895, + 0.634 + ], + "angle": 0, + "content": "\\[\n\\frac {\\partial (\\mathbf {U} , \\mathbf {W} , \\bar {\\mathbf {U}} , \\bar {\\mathbf {W}})}{\\partial \\mathbf {A}} = - (\\partial_ {1} \\mathbf {F}) ^ {- 1} \\partial_ {2} \\mathbf {F}. \\qquad (3)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.641, + 0.895, + 0.718 + ], + "angle": 0, + "content": "However, this requires the dual variables \\(\\bar{\\mathbf{U}}\\) and \\(\\bar{\\mathbf{W}}\\), which are not computed in scikit-learn's [54] popular implementation*. Consequently, we leverage the work of [32] and we re-implement our own solver with Jaxopt [4] based on ADMM [5], a GPU friendly algorithm (see Appendix C.2)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.718, + 0.896, + 0.869 + ], + "angle": 0, + "content": "Concretely, given our concepts bank \\(\\mathbf{W}\\), the concept attribution maps of a new input \\(\\pmb{x}\\) are calculated by solving the NNLS problem \\(\\min_{\\mathbf{U} \\geq 0} \\frac{1}{2} \\| \\pmb{g}(\\pmb{x}) - \\mathbf{U} \\mathbf{W}^{\\top} \\|_F^2\\). The implicit differentiation of the NMF block \\(\\partial \\mathbf{U} / \\partial \\mathbf{A}\\) is integrated into the classic back-propagation to obtain \\(\\partial \\mathbf{U} / \\partial \\pmb{x}\\). Most interestingly, this technical advance enables the use of all white-box explainability methods [59, 62, 66, 69, 71, 78] to generate concept-wise attribution maps and trace the part of an image that triggered the detection of the concept by the network. Additionally, it is even possible to employ black-box" + }, + { + "type": "page_footnote", + "bbox": [ + 0.5, + 0.875, + 0.894, + 0.9 + ], + "angle": 0, + "content": "*Scikit-learn uses a block coordinate descent algorithm [7, 17], with a randomized SVD initialization." + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.926, + 0.516, + 0.938 + ], + "angle": 0, + "content": "2716" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.12, + 0.089, + 0.855, + 0.306 + ], + "angle": 0, + "content": "
Husky vs. WolfLeaves"Kit Fox" vs "Red Fox"
Session n°123Utility123Utility123
Baseline55.766.262.970.176.878.658.862.258.8
Control53.361.061.40.9572.078.080.21.0260.759.248.5
AttributionsSaliency [62]53.969.673.31.0683.288.782.41.1361.760.258.2
Integ.-Grad. [71]67.472.873.21.1582.582.585.31.1159.458.358.3
SmoothGrad [66]68.775.378.01.2083.085.786.31.1350.355.061.4
GradCAM [59]77.685.784.11.3481.983.582.41.1054.452.554.1
Occlusion [76]71.075.778.11.2278.886.182.91.1051.060.255.1
Grad.-Input [61]65.863.367.91.0676.582.979.51.0550.057.662.6
ConceptsACE [24]68.871.472.71.1579.873.882.11.0548.446.546.1
CRAFTCO (ours)82.487.085.11.3878.885.589.41.1255.549.553.3
CRAFT (ours)90.697.395.51.5386.286.685.51.1556.550.649.4
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.317, + 0.896, + 0.43 + ], + "angle": 0, + "content": "Table 1. Utility scores on 3 datasets from [8]. Their Utility benchmark evaluates how well explanations help users identify general rules driving classifications that readily transfer to unseen instances. At training time, users are asked to infer rules driving the decisions of the model given a set of images, and their associated predictions and explanations. At test time, the Utility metric measures the accuracy of users at predicting the model decision on novel images averaged over 3 sessions, and normalized by the baseline accuracy of users trained without explanations. The higher the Utility score, the more useful the explanation, and the more crucial the information provided is for understanding –and thus predicting the model's output– on novel samples. CRAFTCO stands for “CRAFT Concept Only” and designates an experimental condition where only global concepts are given to users, without local explanations (i.e., the concept attribution maps). The first and second best results above the baseline are in bold and underlined, respectively." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.443, + 0.47, + 0.473 + ], + "angle": 0, + "content": "methods [14, 47, 55, 57] since it only amounts to solving an NNLS problem." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.482, + 0.308, + 0.499 + ], + "angle": 0, + "content": "4. Experimental evaluation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.507, + 0.471, + 0.568 + ], + "angle": 0, + "content": "In order to evaluate the interest and the benefits brought by CRAFT, we start in Section 4.1 by assessing the practical utility of the method on a human-centered benchmark composed of 3 XAI scenarios." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.568, + 0.473, + 0.826 + ], + "angle": 0, + "content": "After demonstrating the usefulness of the method using these human experiments, we independently validate the 3 proposed ingredients. First, we provide evidence that recursivity allows refining concepts, making them more meaningful to humans using two additional human experiments in Section 4.2. Next, we evaluate our new Sobol estimator and show quantitatively that it provides a more faithful assessment of concept importance in Section 4.3. Finally, we run an ablation experiment that measures the interest of local explanations based on concept attribution maps coupled with global explanations. Additional experiments, including a sanity check and an example of deep dreams applied on the concept bank, as well as many other examples of local explanations for randomly picked images from ILSVRC2012, are included in Section B of the supplementary materials. We leave the discussion on the limitations of this method and on the broader impact in appendix A." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.833, + 0.251, + 0.849 + ], + "angle": 0, + "content": "4.1. Utility Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.471, + 0.901 + ], + "angle": 0, + "content": "As emphasized by Doshi-Velez et al. [11], the goal of XAI should be to develop methods that help a user better understand the behavior of deep neural network models. An" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.443, + 0.895, + 0.76 + ], + "angle": 0, + "content": "instantiation of this idea was recently introduced by Colin & Fel et al. [8] who described an experimental framework to quantitatively measure the practical usefulness of explainability methods in real-world scenarios. For their initial setup, these authors recruited \\( n = 1 \\), 150 online participants (evaluated over 8 unique conditions and 3 AI scenarios) – making it the largest benchmark to date in XAI. Here, we follow their framework rigorously to allow for the robust evaluation of the utility of our proposed CRAFT method and the related ACE. The 3 representative real-world scenarios are: (1) identifying bias in an AI system (using Husky vs Wolf dataset from [57]), (2) characterizing the visual strategy that are too difficult for an untrained non-expert human observer (using the Paleobotanical dataset from [75]), (3) understanding complex failure cases (using ImageNet “Red fox” vs “Kit fox” binary classification). Using this benchmark, we evaluate CRAFT, ACE, as well as CRAFT with only the global concepts (CRAFTCO) to allow for a fair comparison with ACE. To the best of our knowledge, we are the first to systematically evaluate concept-based methods against attribution methods." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.896, + 0.903 + ], + "angle": 0, + "content": "Results are shown in Table 1 and demonstrate the benefit of CRAFT, which achieves higher scores than all of the attribution methods tested as well as ACE in the first two scenarios. To date, no method appears to exceed the baseline on the third scenario suggesting that additional work is required. We also note that, in the first two scenarios, CRAFTCO is one of the best-performing methods and it always outperforms ACE – meaning that even without the local explanation of the concept attribution maps, CRAFT" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.926, + 0.516, + 0.938 + ], + "angle": 0, + "content": "2717" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.082, + 0.089, + 0.468, + 0.213 + ], + "angle": 0, + "content": "
IntruderExperts (n = 36)Laymen (n = 37)
Acc. Concept70.19%61.08%
Acc. Sub-Concept74.81% (p = 0.18)67.03% (p = 0.043)
Binary choice
Sub-Concept76.1% (p < 0.001)74.95% (p < 0.001)
Odds Ratios3.532.99
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.223, + 0.472, + 0.252 + ], + "angle": 0, + "content": "Table 2. Results from the psychophysics experiments to validate the recursivity ingredient." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.261, + 0.47, + 0.293 + ], + "angle": 0, + "content": "largely outperforms ACE. Examples of concepts produced by CRAFT are shown in the Appendix E.1." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.303, + 0.304, + 0.32 + ], + "angle": 0, + "content": "4.2. Validation of Recursivity" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.327, + 0.471, + 0.508 + ], + "angle": 0, + "content": "To evaluate the meaningfulness of the extracted high-level concepts, we performed psychophysics experiments with human subjects, whom we asked to answer a survey in two phases. Furthermore, we distinguished two different audiences: on the one hand, experts in machine learning, and on the other hand, people with no particular knowledge of computer vision. Both groups of participants were volunteers and did not receive any monetary compensation. Some examples of the developed interface are available the appendix E. It is important to note that this experiment was carried out independently from the utility evaluation and thus it was setup differently." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.509, + 0.471, + 0.735 + ], + "angle": 0, + "content": "Intruder detection experiment First, we ask users to identify the intruder out of a series of five image crops belonging to a certain class, with the odd one being taken from a different concept but still from the same class. Then, we compare the results of this intruder detection with another intruder detection, this time, using a concept (e.g., \\(\\mathcal{C}_1\\)) coming from a layer \\(l\\) and one of its sub-concepts (e.g., \\(\\mathcal{C}_{12}\\) in Fig.3) extracted using our recursive method. If the concept (or sub-concept) is coherent, then it should be easy for the users to find the intruder. Table 2 summarizes our results, showing that indeed both concepts and sub-concepts are coherent, and that recursivity can lead to a slightly higher understanding of the generated concepts (significant for non-experts, but not for experts) and might suggest a way to make concepts more interpretable." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.735, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Binary choice experiment In order to test the improvement of coherence of the sub-concept generated by recursivity with respect to the larger parent concept, we showed participants an image crop belonging to both a subcluster and a parent cluster (e.g., \\(\\pi(\\boldsymbol{x}) \\in \\mathcal{C}_{11} \\subset \\mathcal{C}_1\\)) and asked them which of the two clusters (i.e., \\(\\mathcal{C}_{11}\\) or \\(\\mathcal{C}_1\\)) seemed to accommodate the image the best. If our hypothesis is correct, then the concept refinement brought by recursivity should help form more coherent clusters. The results in Table 2 are satisfying since in both the expert and non-expert groups, the participants chose the sub-cluster more than \\(74\\%\\) of the" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.894, + 0.152 + ], + "angle": 0, + "content": "time. We measure the significance of our results by fitting a binomial logistic regression to our data, and we find that both groups are more likely to choose the sub-concept cluster (at a \\( p < 0.001 \\))." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.161, + 0.663, + 0.178 + ], + "angle": 0, + "content": "4.3. Fidelity analysis" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.184, + 0.895, + 0.457 + ], + "angle": 0, + "content": "We propose to simultaneously verify that identified concepts are faithful to the model and that the concept importance estimator performs better than that used in TCAV [40] by using the fidelity metrics introduced in [24, 77]. These metrics are similar to the ones used for attribution methods, which consist of studying the change of the logit score when removing/adding pixels considered important. Here, we do not introduce these perturbations in the pixel space but in the concept space: once \\(\\mathbf{U}\\) and \\(\\mathbf{W}\\) are computed, we reconstruct the matrix \\(\\mathbf{A} \\approx \\mathbf{U}\\mathbf{W}^{\\top}\\) using only the most important concept (or removing the most important concept for deletion) and compute the resulting change in the output of the model. As can be seen from Fig. 5, ranking the extracted concepts using Sobol's importance score results in steeper curves than when they are sorted by their TCAV scores. We confirm that these results generalize with other matrix factorization techniques (PCA, ICA, RCA) in Section F of the Appendix." + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.469, + 0.703, + 0.591 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.708, + 0.469, + 0.892, + 0.591 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.602, + 0.895, + 0.658 + ], + "angle": 0, + "content": "Figure 5. (Left) Deletion curves (lower is better). (Right) Insertion curves (higher is better). For both the deletion or insertion metrics, Sobol indices lead to better estimates (calculated on \\(>100\\mathrm{K}\\) images) of important concepts." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.663, + 0.62, + 0.679 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.688, + 0.895, + 0.855 + ], + "angle": 0, + "content": "We presented CRAFT, a method to automatically extract understandable concepts from deep networks, which provides insights into the model's decision-making process. Our approach uses a recursive formulation to identify concepts at the correct level of granularity, a novel method for measuring concept importance, and implicit differentiation to generate concept attribution maps. We conducted psychophysics experiments to validate the approach's usefulness and meaningfulness to human experimenters. We anticipate that our work will inspire future research in concept-based explainability methods." + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.926, + 0.516, + 0.938 + ], + "angle": 0, + "content": "2718" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.472, + 0.169 + ], + "angle": 0, + "content": "[1] Julius Adebayo, Justin Gilmer, Michael Muelly, Ian Goodfellow, Moritz Hardt, and Been Kim. Sanity checks for saliency maps. Advances in neural information processing systems, 31, 2018. 2, 3, 24" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.172, + 0.472, + 0.213 + ], + "angle": 0, + "content": "[2] João Carlos Alves Barata and Mahir Saleh Hussein. The moore-penrose pseudoinverse: A tutorial review of the theory. Brazilian Journal of Physics, 42(1):146-165, 2012. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.215, + 0.472, + 0.268 + ], + "angle": 0, + "content": "[3] Bradley M Bell and James V Burke. Algorithmic differentiation of implicit functions and optimal values. In Advances in Automatic Differentiation, pages 67-77. Springer, 2008. 6, 19" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.271, + 0.472, + 0.34 + ], + "angle": 0, + "content": "[4] Mathieu Blondel, Quentin Berthet, Marco Cuturi, Roy Frostig, Stephan Hoyer, Felipe Llinares-López, Fabian Pedregosa, and Jean-Philippe Vert. Efficient and modular implicit differentiation. arXiv preprint arXiv:2105.15183, 2021. 3, 6, 19" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.342, + 0.472, + 0.41 + ], + "angle": 0, + "content": "[5] Stephen Boyd, Neal Parikh, Eric Chu, Borja Peleato, Jonathan Eckstein, et al. Distributed optimization and statistical learning via the alternating direction method of multipliers. Foundations and Trends® in Machine learning, 3(1):1-122, 2011. 6, 13, 18, 19" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.412, + 0.472, + 0.467 + ], + "angle": 0, + "content": "[6] Chaofan Chen, Oscar Li, Daniel Tao, Alina Barnett, Cynthia Rudin, and Jonathan K Su. This looks like that: deep learning for interpretable image recognition. Advances in neural information processing systems, 32, 2019. 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.469, + 0.472, + 0.536 + ], + "angle": 0, + "content": "[7] Andrzej Cichocki and Anh-Huy Phan. Fast local algorithms for large scale nonnegative matrix and tensor factorizations. IEICE transactions on fundamentals of electronics, communications and computer sciences, 92(3):708-721, 2009. 6, 19" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.539, + 0.472, + 0.608 + ], + "angle": 0, + "content": "[8] Julien Colin, Thomas Fel, Rémi Cadène, and Thomas Serre. What i cannot predict, i do not understand: A human-centered evaluation framework for explainability methods. Advances in Neural Information Processing Systems (NeurIPS), 2021. 2, 3, 7, 21" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.61, + 0.472, + 0.665 + ], + "angle": 0, + "content": "[9] RI Cukier, CM Fortuin, Kurt E Shuler, AG Petschek, and J Ho Schaibly. Study of the sensitivity of coupled reaction systems to uncertainties in rate coefficients. i theory. The Journal of chemical physics, 59(8):3873-3878, 1973. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.667, + 0.472, + 0.722 + ], + "angle": 0, + "content": "[10] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 2, 14, 23" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.723, + 0.472, + 0.763 + ], + "angle": 0, + "content": "[11] Finale Doshi-Velez and Been Kim. Towards a rigorous science of interpretable machine learning. ArXiv e-print, 2017. 1, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.766, + 0.472, + 0.806 + ], + "angle": 0, + "content": "[12] Carl Eckart and Gale Young. The approximation of one matrix by another of lower rank. Psychometrika, 1(3):211-218, 1936. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.808, + 0.472, + 0.892 + ], + "angle": 0, + "content": "[13] Nelson Elhage, Tristan Hume, Catherine Olsson, Nicholas Schiefer, Tom Henighan, Shauna Kravec, Zac Hatfield-Dodds, Robert Lasenby, Dawn Drain, Carol Chen, Roger Grosse, Sam McCandlish, Jared Kaplan, Dario Amodei, Martin Wattenberg, and Christopher Olah. Toy models of superposition. Transformer Circuits Thread, 2022. 4" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.472, + 0.892 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.162 + ], + "angle": 0, + "content": "[14] Thomas Fel, Rémi Cadène, Mathieu Chalvidal, Matthieu Cord, David Vigouroux, and Thomas Serre. Look at the variance! efficient black-box explanations with sobol-based sensitivity analysis. Advances in Neural Information Processing Systems, 34, 2021. 1, 3, 7, 19" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.164, + 0.894, + 0.232 + ], + "angle": 0, + "content": "[15] Thomas Fel, Mélanie Ducoffe, David Vigouroux, Rémi Cadène, Mikael Capelle, Claire Nicodème, and Thomas Serre. Don't lie to me! robust and efficient explainability with verified perturbation analysis. arXiv preprint arXiv:2202.07728, 2022.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.234, + 0.894, + 0.343 + ], + "angle": 0, + "content": "[16] Thomas Fel, Lucas Hervier, David Vigouroux, Antonin Poche, Justin Plakoo, Remi Cadene, Mathieu Chalvidal, Julien Colin, Thibaut Boissin, Louis Béthune, Agustin Picard, Claire Nicodeme, Laurent Gardes, Gregory Flandin, and Thomas Serre. Xplique: A deep learning explainability toolbox. Workshop, Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 16" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.346, + 0.894, + 0.386 + ], + "angle": 0, + "content": "[17] Cédric Févotte and Jérôme Idier. Algorithms for nonnegative matrix factorization with the \\(\\beta\\)-divergence. Neural computation, 23(9):2421-2456, 2011. 6, 19" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.388, + 0.893, + 0.443 + ], + "angle": 0, + "content": "[18] Ruth C Fong and Andrea Vedaldi. Interpretable explanations of black boxes by meaningful perturbation. In Proceedings of the IEEE international conference on computer vision, pages 3429-3437, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.446, + 0.894, + 0.5 + ], + "angle": 0, + "content": "[19] Ruth C Fong and Andrea Vedaldi. Interpretable explanations of black boxes by meaningful perturbation. In Proceedings of the IEEE international conference on computer vision, pages 3429-3437, 2017. 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.502, + 0.894, + 0.557 + ], + "angle": 0, + "content": "[20] Xiao Fu, Kejun Huang, Nicholas D Sidiropoulos, and Wing-Kin Ma. Nonnegative matrix factorization for signal and data analytics: Identifiability, algorithms, and applications. IEEE Signal Process. Mag., 36(2):59-80, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.559, + 0.894, + 0.641 + ], + "angle": 0, + "content": "[21] Yunhao Ge, Yao Xiao, Zhi Xu, Meng Zheng, Srikrishna Karanam, Terrence Chen, Laurent Itti, and Ziyan Wu. A peek into the reasoning of neural networks: Interpreting with structural visual concepts. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2195-2204, 2021. 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.643, + 0.894, + 0.697 + ], + "angle": 0, + "content": "[22] Sahara Ghalebikesabi, Lucile Ter-Minassian, Karla DiazOrdaz, and Chris C Holmes. On locality of local explanation models. Advances in Neural Information Processing Systems (NeurIPS), 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.7, + 0.894, + 0.74 + ], + "angle": 0, + "content": "[23] Amirata Ghorbani, Abubakar Abid, and James Zou. Interpretation of neural networks is fragile. In Proceedings of the AAAI Conference on Artificial Intelligence (AAAI), 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.742, + 0.894, + 0.796 + ], + "angle": 0, + "content": "[24] Amirata Ghorbani, James Wexler, James Y Zou, and Been Kim. Towards automatic concept-based explanations. Advances in Neural Information Processing Systems, 32, 2019, 2, 3, 4, 5, 7, 8, 13, 14, 23" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.799, + 0.894, + 0.839 + ], + "angle": 0, + "content": "[25] Andreas Griewank and Andrea Walther. Evaluating derivatives: principles and techniques of algorithmic differentiation. SIAM, 2008. 3, 6, 19" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.841, + 0.894, + 0.909 + ], + "angle": 0, + "content": "[26] Peter Hase and Mohit Bansal. Evaluating explainable ai: Which algorithmic explanations help users predict model behavior? In Proceedings of the Annual Meeting of the Association for Computational Linguistics (ACL Short Papers), 2020. 2, 3" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.909 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.927, + 0.516, + 0.938 + ], + "angle": 0, + "content": "2719" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.147 + ], + "angle": 0, + "content": "[27] Peter Hase, Harry Xie, and Mohit Bansal. The out-of-distribution problem in explainability and search methods for feature importance explanations. Advances in Neural Information Processing Systems (NeurIPS), 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.149, + 0.472, + 0.189 + ], + "angle": 0, + "content": "[28] Johannes Haug, Stefan Zürn, Peter El-Jiz, and Gjergji Kasneci. On baselines for local feature attributions. arXiv preprint arXiv:2101.00905, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.19, + 0.472, + 0.243 + ], + "angle": 0, + "content": "[29] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Identity mappings in deep residual networks. In European conference on computer vision, pages 630-645. Springer, 2016. 23" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.246, + 0.472, + 0.286 + ], + "angle": 0, + "content": "[30] Magnus R Hestenes and Eduard Stiefel. Methods of conjugate gradients for solving. Journal of research of the National Bureau of Standards, 49(6):409, 1952. 18, 19" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.287, + 0.472, + 0.355 + ], + "angle": 0, + "content": "[31] Cheng-Yu Hsieh, Chih-Kuan Yeh, Xuanqing Liu, Pradeep Ravikumar, Seungyeon Kim, Sanjiv Kumar, and Cho-Jui Hsieh. Evaluations and methods for explanation through robustness analysis. In Proceedings of the International Conference on Learning Representations (ICLR), 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.357, + 0.472, + 0.411 + ], + "angle": 0, + "content": "[32] Kejun Huang, Nicholas D Sidiropoulos, and Athanasios P Liavas. A flexible and efficient algorithmic framework for constrained matrix and tensor factorization. IEEE Transactions on Signal Processing, 64(19):5052-5065, 2016. 6, 18" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.412, + 0.472, + 0.466 + ], + "angle": 0, + "content": "[33] Marouane Il Idrissi, Vincent Chabridon, and Bertrand Iooss. Developments and applications of shapley effects to reliability-oriented sensitivity analysis with correlated inputs. Environmental Modelling & Software, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.467, + 0.472, + 0.522 + ], + "angle": 0, + "content": "[34] Bertrand Iooss and Paul Lemaitre. A review on global sensitivity analysis methods. In Uncertainty management in simulation-optimization of complex systems, pages 101-122. Springer, 2015. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.523, + 0.472, + 0.59 + ], + "angle": 0, + "content": "[35] Alon Jacovi, Ana Marasovic, Tim Miller, and Yoav Goldberg. Formalizing trust in artificial intelligence: Prerequisites, causes and goals of human trust in ai. In Proceedings of the 2021 ACM conference on fairness, accountability, and transparency, pages 624-635, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.592, + 0.472, + 0.646 + ], + "angle": 0, + "content": "[36] Alexandre Janon, Thierry Klein, Agnes Lagnoux, Maëlle Nodet, and Clémentine Prieur. Asymptotic normality and efficiency of two sobol index estimators. *ESAIM: Probability and Statistics*, 18:342–364, 2014. 5, 20" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.647, + 0.472, + 0.674 + ], + "angle": 0, + "content": "[37] Margot E Kaminski and Jennifer M Urban. The right to contest ai. Columbia Law Review, 121(7):1957-2048, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.675, + 0.472, + 0.716 + ], + "angle": 0, + "content": "[38] William Karush. Minima of functions of several variables with inequalities as side constraints. M. Sc. Dissertation. Dept. of Mathematics, Univ. of Chicago, 1939. 6, 18" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.717, + 0.472, + 0.77 + ], + "angle": 0, + "content": "[39] Been Kim, Rajiv Khanna, and Oluwasanmi O Koyejo. Examples are not enough, learn to criticize! criticism for interpretability. Advances in Neural Information Processing Systems (NeurIPS), 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.772, + 0.472, + 0.853 + ], + "angle": 0, + "content": "[40] Been Kim, Martin Wattenberg, Justin Gilmer, Carrie Cai, James Wexler, Fernanda Viegas, et al. Interpretability beyond feature attribution: Quantitative testing with concept activation vectors (tcav). In International conference on machine learning, pages 2668-2677. PMLR, 2018. 2, 3, 5, 8, 16, 23" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.855, + 0.472, + 0.907 + ], + "angle": 0, + "content": "[41] Sunnie S. Y. Kim, Nicole Meister, Vikram V. Ramaswamy, Ruth Fong, and Olga Russakovsky. HIVE: Evaluating the human interpretability of visual explanations. In European Conference on Computer Vision (ECCV), 2022. 2, 3" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.907 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.16 + ], + "angle": 0, + "content": "[42] Pieter-Jan Kindermans, Sara Hooker, Julius Adebayo, Maximilian Alber, Kristof T. Schütt, Sven Dähne, Dumitru Erhan, and Been Kim. The (Un)reliability of Saliency Methods, pages 267-280. Springer International Publishing, Cham, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.164, + 0.894, + 0.272 + ], + "angle": 0, + "content": "[43] Mauritz Kop. Eu artificial intelligence act: The European approach to ai. In Stanford - Vienna Transatlantic Technology Law Forum, Transatlantic Antitrust and IPR Developments, Stanford University, Issue No. 2/2021. https://law.stanford.edu/publications/eu-artificial-intelligence-act-the-european-approach-to-ai/. Stanford-Vienna Transatlantic Technology Law Forum, Transatlantic Antitrust, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.276, + 0.894, + 0.316 + ], + "angle": 0, + "content": "[44] Steven George Krantz and Harold R Parks. The implicit function theorem: history, theory, and applications. Springer Science & Business Media, 2002. 3, 6, 19" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.318, + 0.894, + 0.372 + ], + "angle": 0, + "content": "[45] Harold W Kuhn and Albert W Tucker. Nonlinear programming proceedings of the second berkeley symposium on mathematical statistics and probability. Neyman, pages 481-492, 1951. 6, 18" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.375, + 0.894, + 0.415 + ], + "angle": 0, + "content": "[46] Daniel D Lee and H Sebastian Seung. Learning the parts of objects by non-negative matrix factorization. Nature, 401(6755):788-791, 1999. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.418, + 0.894, + 0.459 + ], + "angle": 0, + "content": "[47] Scott M Lundberg and Su-In Lee. A unified approach to interpreting model predictions. Advances in neural information processing systems, 30, 2017. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.461, + 0.894, + 0.514 + ], + "angle": 0, + "content": "[48] Amandine Marrel, Bertrand Iooss, Beatrice Laurent, and Olivier Roustant. Calculations of sobol indices for the gaussian process metamodel. Reliability Engineering & System Safety, 94(3):742-751, 2009. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.517, + 0.894, + 0.572 + ], + "angle": 0, + "content": "[49] Giang Nguyen, Daeyoung Kim, and Anh Nguyen. The effectiveness of feature attribution methods and its correlation with automatic evaluation scores. Advances in Neural Information Processing Systems (NeurIPS), 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.574, + 0.894, + 0.628 + ], + "angle": 0, + "content": "[50] Paul Novello, Thomas Fel, and David Vigouroux. Making sense of dependence: Efficient black-box explanations using dependence measure. In Advances in Neural Information Processing Systems (NeurIPS), 2022. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.63, + 0.894, + 0.671 + ], + "angle": 0, + "content": "[51] Chris Olah, Alexander Mordvintsev, and Ludwig Schubert. Feature visualization. Distill, 2017. https://distill.pub/2017/feature-visualization. 13, 16" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.673, + 0.894, + 0.713 + ], + "angle": 0, + "content": "[52] Art B Owen. Better estimation of small sobol'sensitivity indices. ACM Transactions on Modeling and Computer Simulation (TOMACS), 23(2):1-17, 2013. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.716, + 0.894, + 0.77 + ], + "angle": 0, + "content": "[53] Vardan Papyan, XY Han, and David L Donoho. Prevalence of neural collapse during the terminal phase of deep learning training. Proceedings of the National Academy of Sciences, 117(40):24652-24663, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.772, + 0.894, + 0.841 + ], + "angle": 0, + "content": "[54] Fabian Pedregosa, Gael Varoquaux, Alexandre Gramfort, Vincent Michel, Bertrand Thirion, Olivier Grisel, Mathieu Blondel, Peter Prettenhofer, Ron Weiss, Vincent Dubourg, et al. Scikit-learn: Machine learning in python. Journal of machine learning research, 12(Oct):2825–2830, 2011. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.843, + 0.894, + 0.884 + ], + "angle": 0, + "content": "[55] Vitali Petsiuk, Abir Das, and Kate Saenko. Rise: Randomized input sampling for explanation of black-box models. arXiv preprint arXiv:1806.07421, 2018. 1, 3, 7, 19, 23" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.884 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.927, + 0.516, + 0.938 + ], + "angle": 0, + "content": "2720" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.161 + ], + "angle": 0, + "content": "[56] Bin Ren, Laurent Pueyo, Christine Chen, Élodie Choquet, John H Debes, Gaspard Duchêne, François Ménard, and Marshall D Perrin. Using data imputation for signal separation in high-contrast imaging. The Astrophysical Journal, 892(2):74, 2020. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.163, + 0.472, + 0.233 + ], + "angle": 0, + "content": "[57] Marco Tulio Ribeiro, Sameer Singh, and Carlos Guestrin. \" why should i trust you?\" explaining the predictions of any classifier. In Proceedings of the 22nd ACM SIGKDD international conference on knowledge discovery and data mining, pages 1135-1144, 2016. 3, 7, 19" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.234, + 0.472, + 0.302 + ], + "angle": 0, + "content": "[58] Andrea Saltelli, Paola Annoni, Ivano Azzini, Francesca Campolongo, Marco Ratto, and Stefano Tarantola. Variance based sensitivity analysis of model output. design and estimator for the total sensitivity index. Computer physics communications, 181(2):259-270, 2010. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.304, + 0.472, + 0.387 + ], + "angle": 0, + "content": "[59] Ramprasaath R Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra. Grad-cam: Visual explanations from deep networks via gradient-based localization. In Proceedings of the IEEE international conference on computer vision, pages 618-626, 2017. 1, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.388, + 0.472, + 0.457 + ], + "angle": 0, + "content": "[60] Hua Shen and Ting-Hao Huang. How useful are the machine-generated interpretations to general users? a human evaluation on guessing the incorrectly predicted labels. In Proceedings of the AAAI Conference on Human Computation and Crowdsourcing, volume 8, pages 168-172, 2020. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.459, + 0.472, + 0.514 + ], + "angle": 0, + "content": "[61] Avanti Shrikumar, Peyton Greenside, and Anshul Kundaje. Learning important features through propagating activation differences. In Proceedings of the International Conference on Machine Learning (ICML), 2017. 1, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.516, + 0.472, + 0.584 + ], + "angle": 0, + "content": "[62] Karen Simonyan, Andrea Vedaldi, and Andrew Zisserman. Deep inside convolutional networks: Visualising image classification models and saliency maps. In In Workshop at International Conference on Learning Representations. CiteSeer, 2014. 1, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.585, + 0.472, + 0.642 + ], + "angle": 0, + "content": "[63] Leon Sixt, Maximilian Granz, and Tim Landgraf. When explanations lie: Why many modified bp attributions fail. In Proceedings of the International Conference on Machine Learning (ICML), 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.643, + 0.472, + 0.711 + ], + "angle": 0, + "content": "[64] Leon Sixt, Martin Schuessler, Oana-Iuliana Popescu, Philipp Weiß, and Tim Landgraf. Do users benefit from interpretable vision? a user study, baseline, and dataset. Proceedings of the International Conference on Learning Representations (ICLR), 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.713, + 0.472, + 0.782 + ], + "angle": 0, + "content": "[65] Dylan Slack, Sophie Hilgard, Emily Jia, Sameer Singh, and Himabindu Lakkaraju. Fooling lime and shap: Adversarial attacks on post hoc explanation methods. In Proceedings of the AAAI/ACM Conference on AI, Ethics, and Society, pages 180-186, 2020. 2, 3" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.782 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.147 + ], + "angle": 0, + "content": "[66] Daniel Smilkov, Nikhil Thorat, Been Kim, Fernanda Viégas, and Martin Wattenberg. Smoothgrad: removing noise by adding noise. arXiv preprint arXiv:1706.03825, 2017. 1, 3, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.149, + 0.894, + 0.19 + ], + "angle": 0, + "content": "[67] Ilya M Sobol. Sensitivity analysis for non-linear mathematical models. Mathematical modelling and computational experiment, 1:407-414, 1993. 2, 5, 20" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.19, + 0.894, + 0.232 + ], + "angle": 0, + "content": "[68] Ilya M Sobol. Global sensitivity indices for nonlinear mathematical models and their monte carlo estimates. Mathematics and computers in simulation, 55(1-3):271-280, 2001. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.233, + 0.894, + 0.288 + ], + "angle": 0, + "content": "[69] Jost Tobias Springenberg, Alexey Dosovitskiy, Thomas Brox, and Martin Riedmiller. Striving for simplicity: The all convolutional net. arXiv preprint arXiv:1412.6806, 2014. 1, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.289, + 0.894, + 0.329 + ], + "angle": 0, + "content": "[70] Pascal Sturmfels, Scott Lundberg, and Su-In Lee. Visualizing the impact of feature attribution baselines. Distill, 2020. 3, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.331, + 0.894, + 0.387 + ], + "angle": 0, + "content": "[71] Mukund Sundararajan, Ankur Taly, and Qiqi Yan. Axiomatic attribution for deep networks. In International conference on machine learning, pages 3319-3328. PMLR, 2017. 1, 3, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.388, + 0.894, + 0.444 + ], + "angle": 0, + "content": "[72] Stefano Tarantola, Debora Gatelli, and Thierry Alex Mara. Random balance designs for the estimation of first order global sensitivity indices. Reliability Engineering & System Safety, 91(6):717-727, 2006. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.445, + 0.894, + 0.486 + ], + "angle": 0, + "content": "[73] Stephen A Vavasis. On the complexity of nonnegative matrix factorization. SIAM Journal on Optimization, 20(3):1364-1377, 2010. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.488, + 0.894, + 0.543 + ], + "angle": 0, + "content": "[74] Yu-Xiong Wang and Yu-Jin Zhang. Nonnegative matrix factorization: A comprehensive review. IEEE Transactions on Knowledge and Data Engineering, 25(6):1336-1353, 2013. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.545, + 0.894, + 0.6 + ], + "angle": 0, + "content": "[75] Peter Wilf, Shengping Zhang, Sharat Chikkerur, Stefan A Little, Scott L Wing, and Thomas Serre. Computer vision cracks the leaf code. Proceedings of the National Academy of Sciences, 113(12):3305-3310, 2016. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.601, + 0.894, + 0.643 + ], + "angle": 0, + "content": "[76] Matthew D Zeiler and Rob Fergus. Visualizing and understanding convolutional networks. In European conference on computer vision, pages 818-833. Springer, 2014. 1, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.644, + 0.894, + 0.712 + ], + "angle": 0, + "content": "[77] Ruihan Zhang, Prashan Madumal, Tim Miller, Krista A Ehinger, and Benjamin IP Rubinstein. Invertible concept-based explanations for cnn models with non-negative concept activation vectors. arXiv preprint arXiv:2006.15417, 2020. 2, 3, 5, 8, 23" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.713, + 0.894, + 0.783 + ], + "angle": 0, + "content": "[78] Bolei Zhou, Aditya Khosla, Agata Lapedriza, Aude Oliva, and Antonio Torralba. Learning deep features for discriminative localization. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2921-2929, 2016. 3, 6" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.783 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.927, + 0.514, + 0.938 + ], + "angle": 0, + "content": "2721" + } + ] +] \ No newline at end of file diff --git a/2023/CRAFT_ Concept Recursive Activation FacTorization for Explainability/6e462b05-7931-4758-8402-5b3e432aab49_origin.pdf b/2023/CRAFT_ Concept Recursive Activation FacTorization for Explainability/6e462b05-7931-4758-8402-5b3e432aab49_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..4167b912b7a62d053828d4c46052a00e3c10bd3d --- /dev/null +++ b/2023/CRAFT_ Concept Recursive Activation FacTorization for Explainability/6e462b05-7931-4758-8402-5b3e432aab49_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:78a93cfb08909c25ead9f46282f3dab9c99ab9c3cf715f9780126ad338da2e81 +size 4345348 diff --git a/2023/CRAFT_ Concept Recursive Activation FacTorization for Explainability/full.md b/2023/CRAFT_ Concept Recursive Activation FacTorization for Explainability/full.md new file mode 100644 index 0000000000000000000000000000000000000000..c1c8bf4e6cf44ff44978a76c118025cf5eb576bb --- /dev/null +++ b/2023/CRAFT_ Concept Recursive Activation FacTorization for Explainability/full.md @@ -0,0 +1,315 @@ +# CRAFT: Concept Recursive Activation FacTorization for Explainability + +# Thomas Fel $^{1,3,5*}$ Agustin Picard $^{3,6*}$ Louis Bethune $^{3*}$ Thibaut Boissin $^{3,4*}$ David Vigouroux $^{3,4}$ Julien Colin $^{1,3}$ Rémi Cadène $^{1,2}$ Thomas Serre $^{1,3}$ + +$^{1}$ Carney Institute for Brain Science, Brown University, USA $^{2}$ Sorbonne Université, CNRS, France + +3Artificial and Natural Intelligence Toulouse Institute, Université de Toulouse, France + +4 Institut de Recherche Technologique Saint-Exupery, France + +$^{5}$ Innovation & Research Division, SNCF, $^{6}$ Scalian + +![](images/d56fc2459ec3e516eb51004c50cc917e487af368e9f0e0748d6313be63eb1089.jpg) +Figure 1. The "Man on the Moon" incorrectly classified as a "shovel" by an ImageNet-trained ResNet50. Heatmap generated by a classic attribution method [55] (left) vs. concept attribution maps generated with the proposed CRAFT approach (right) which highlights the two most influential concepts that drove the ResNet50's decision along with their corresponding locations. CRAFT suggests that the neural net arrived at its decision because it identified the concept of "dirt" $\bullet$ commonly found in members of the image class "shovel" and the concept of "ski pants" $\bullet$ typically worn by people clearing snow from their driveway with a shovel instead the correct concept of astronaut's pants (which was probably never seen during training). + +![](images/cfaae2cfa00822a5aaf09c97ac20222f3dbdded4a740b519378efb4996e10350.jpg) + +![](images/6835fc9574e013856aa964e78ddce99ff8752d0d9cf9e6ea644f2ba6d1955fc0.jpg) + +![](images/5483d089ef7a9bae1eb123914f9294f97bb3a9f8d11f0d49f0c90775826474f6.jpg) + +![](images/0b85a981ba4c88245dbfa77a2854edaea9ef873e88e992ad1b8b4d7a1d18dd60.jpg) + +![](images/50b2c5017faf5a3ce4184951b8c86b5099a4b2c9796403fb3dd7c7793669b945.jpg) + +![](images/ac323bfa9cd0fa3a16accf3ae3a9b9b7691b25760f3c52c28032f39284837361.jpg) + +![](images/5ea86c83b0c1d9ec6a497be15e05d4d4f8f8bdd75cbc977e2f2ea1ac10ce1a1d.jpg) + +![](images/16db1080dad5ebbf1b0dec5c82e8be9894b5f6f7a6d600f44a6a199860ae7d28.jpg) + +![](images/93b31b3e45f8d1330b6dd828eacfff7ed168ba2fae6aa85c955c36c9887ba996.jpg) + +![](images/68174e964303eabc967e664f324f7940387368b35bbb24afc6ac725c0e2a529c.jpg) + +![](images/aea7dc080165edc0f2e8ba45de0d4027c82715e57bc8b26fba4b9e2dbf260592.jpg) + +![](images/b2604fe305ebdb7f4d7a78839207afd070c12f4646315e3b2de8b16c4c7a35a3.jpg) + +![](images/87798a3af8a411c63c8d751d755be2c6b5dda99166e85ca07541112978aa008b.jpg) + +# Abstract + +Attribution methods, which employ heatmaps to identify the most influential regions of an image that impact model decisions, have gained widespread popularity as a type of explainability method. However, recent research has exposed the limited practical value of these methods, attributed in part to their narrow focus on the most prominent regions of an image - revealing "where" the model looks, but failing to elucidate "what" the model sees in those areas. In this work, we try to fill in this gap with CRAFT - a novel approach to identify both "what" and "where" by generating concept-based explanations. We introduce 3 new ingredients to the automatic concept extraction literature: (i) a recursive strategy to detect and decompose concepts across layers, (ii) a novel method for a more faithful estimation of concept importance using Sobol indices, and (iii) the use of implicit differentiation to unlock Concept Attribution Maps. + +We conduct both human and computer vision experiments to demonstrate the benefits of the proposed approach. + +We show that the proposed concept importance estimation technique is more faithful to the model than previous methods. When evaluating the usefulness of the method for human experimenters on a human-centered utility benchmark, we find that our approach significantly improves on two of the three test scenarios. + +# 1. Introduction + +Interpreting the decisions of modern machine learning models such as neural networks remains a major challenge. Given the ever-increasing range of machine learning applications, the need for robust and reliable explainability methods continues to grow [11, 35]. Recently enacted European laws (including the General Data Protection Regulation (GDPR) [37] and the European AI act [43]) require the assessment of explainable decisions, especially those made by algorithms. + +In order to try to meet this growing need, an array of explainability methods have already been proposed [14, 50, + +![](images/052520db0224c7b7815264d63e73a1ac961132927e72f15c833437870d919206.jpg) +Figure 2. CRAFT results for the prediction "chain saw". First, our method uses Non-Negative Matrix Factorization (NMF) to extract the most relevant concepts used by the network (ResNet50V2) from the train set (ILSVRC2012 [10]). The global influence of these concepts on the predictions is then measured using Sobol indices (right panel). Finally, the method provides local explanations through concept attribution maps (heatmaps associated with a concept, and computed using grad-CAM by backpropagating through the NMF concept values with implicit differentiation). Besides, concepts can be interpreted by looking at crops that maximize the NMF coefficients. For the class "chain saw", the detected concepts seem to be: $\bullet$ the chainsaw engine, $\bullet$ the saw blade, $\bullet$ the human head, $\bullet$ the vegetation, $\bullet$ the jeans and $\bullet$ the tree trunk. + +![](images/61a7202338e1d1ecb53d1059ce55768474216d7dc8d5ca828866fbf4c3093902.jpg) + +59, 61, 62, 66, 69, 71, 76]. One of the main classes of methods called attribution methods yields heatmaps that indicate the importance of individual pixels for driving a model's decision. However, these methods exhibit critical limitations [1, 27, 63, 65], as they have been shown to fail – or only marginally help – in recent human-centered benchmarks [8, 26, 41, 49, 60, 64]. It has been suggested that their limitations stem from the fact that they are only capable of explaining where in an image are the pixels that are critical to the decision but they cannot tell what visual features are actually driving decisions at these locations. In other words, they show where the model looks but not what it sees. For example, in the scenario depicted in Fig. 1, where an ImageNet-trained ResNet mistakenly identifies an image as containing a shovel, the attribution map displayed on the left fails to explain the reasoning behind this misclassification. + +A recent approach has sought to move past attribution methods [40] by using so-called "concepts" to communicate information to users on how a model works. The goal is to find human-interpretable concepts in the activation space of a neural network. Although the approach exhibited potential, its practicality is significantly restricted due to the need for prior knowledge of pertinent concepts in its original formulation and, more critically, the requirement for a labeled dataset of such concepts. Several lines of work have focused on trying to automate the concept discovery process based only on the training dataset and without explicit human supervision. The most prominent of these techniques, ACE [24], uses a combination of segmentation and clustering techniques but requires heuristics to remove outliers. However, ACE provides a proof of concept that it might be possible to discover concepts automatically and at scale – without additional labeling or human supervision. Nevertheless, the approach suffers several limitations: by + +construction, each image segment can only belong to a single cluster, a layer has to be selected by the user to be used to retrieve the relevant concepts, and the amount of information lost during the outlier rejection phase can be a cause of concern. More recently, Zhang et al. [77] proposes to leverage matrix decompositions on internal feature maps to discover concepts. + +Here, we try to fill these gaps with a novel method called CRAFT which uses Non-Negative Matrix Factorization (NMF) [46] for concept discovery. In contrast to other concept-based explanation methods, our approach provides an explicit link between their global and local explanations (Fig. 2) and identifies the relevant layer(s) to use to represent individual concepts (Fig. 3). Our main contributions can be described as follows: + +(i) A novel approach for the automated extraction of high-level concepts learned by deep neural networks. We validate its practical utility to users with human psychophysics experiments. +(ii) A recursive procedure to automatically identify concepts and sub-concepts at the right level of granularity – starting with our decomposition at the top of the model and working our way upstream. We validate the benefit of this approach with human psychophysics experiments showing that (i) the decomposition of a concept yields more coherent sub-concepts and (ii) that the groups of points formed by these sub-concepts are more refined and appear meaningful to humans. +(iii) A novel technique to quantify the importance of individual concepts for a model's prediction using Sobol indices [34, 67, 68] – a technique borrowed from Sensitivity Analysis. +(iv) The first concept-based explainability method which produces concept attribution maps by backpropagating con + +cept scores into the pixel space by leveraging the implicit function theorem in order to localize the pixels associated with the concept of a given input image. This effectively opens up the toolbox of both white-box [15, 59, 62, 66, 69, 71, 78] and black-box [14, 47, 55, 57] explainability methods to derive concept-wise attribution maps. + +# 2. Related Work + +Attribution methods Attribution methods are widely used as post-hoc explainability techniques to determine the input variables that contribute to a model's prediction by generating importance maps, such as the ones shown in Fig.1. The first attribution method, Saliency, introduced in [62], generates a heatmap by utilizing the gradient of a given classification score with respect to the pixels. This method was later improved upon in the context of deep convolutional networks for classification in subsequent studies, such as [66, 69, 71, 76]. However, the image gradient only reflects the model's operation within an infinitesimal neighborhood around an input, which can yield misleading importance estimates [22] since gradients of large vision models are notoriously noisy [66]. That is why several methods leverage perturbations on the input image to probe the model and create importance maps that indicate the most crucial areas for the decision, such as Rise [55], Sobol [14], or more recently HSIC [50]. + +Unfortunately, a severe limitation of these approaches – apart from the fact that they only show the “where” – is that they are subject to confirmation bias: while they may appear to offer useful explanations to a user, sometimes these explanations are actually incorrect [1, 23, 65]. These limitations raise questions about their usefulness, as recent research has shown by using human-centered experiments to evaluate the utility of attribution [8, 26, 41, 49, 60]. + +In particular, in [8], a protocol is proposed to measure the usefulness of explanations, corresponding to how much they help users identify rules driving a model's predictions (correct or incorrect) that transfer to unseen data - using the concept of meta-predictor (also called simulatability) [11, 18, 39]. The main idea is to train users to predict the output of the system using a small set of images along with associated model predictions and corresponding explanations. A method that performs well on this benchmark is said useful, as it helps users better predict the output of the model by providing meaningful information about the internal functioning of the model. This framework being agnostic to the type of explainability method, we have chosen to use it in Section 4 in order to compare CRAFT with attribution methods. + +Concepts-based methods Kim et al. [40] introduced a method aimed at providing explanations that go beyond attribution-based approaches by measuring the impact of + +pre-selected concepts on a model's outputs. Although this method appears more interpretable to human users than standard attribution techniques, it requires a database of images describing the relevant concepts to be manually curated. Ghorbani et al. [24] further extended the approach to extract concepts without the need for human supervision. The approach, called ACE [24], uses a segmentation scheme on images, that belong to an image class of interest. The authors leveraged the intermediate activations of a neural network for specific image segments. These segments were resized to the appropriate input size and filled with a baseline value. The resulting activations were clustered to produce prototypes, which they referred to as "concepts". However, some concepts contained background segments, leading to the inclusion of uninteresting and outlier concepts. To address this, the authors implemented a post-processing cleanup step to remove these concepts, including those that were present in only one image of the class and were not representative. While this improved the interpretability of their explanations to human subjects, the use of a baseline value filled around the segments could introduce biases in the explanations [28, 31, 42, 70]. + +Zhang et al. [77] developed a solution to the unsupervised concept discovery problem by using matrix factorizations in the latent spaces of neural networks. However, one major drawback of this method is that it operates at the level of convolutional kernels, leading to the discovery of localized concepts. For example, the concept of "grass" at the bottom of the image is considered to be distinct from the concept of "grass" at the top of the image. + +# 3. Overview of the method + +In this section, we first describe our concept activations factorization method. Below we highlight the main differences with related work. We then proceed to introduce the three novel ingredients that make up CRAFT: (1) a method to recursively decompose concepts into sub-concepts, (2) a method to better estimate the importance of extracted concepts, and (3) a method to use any attribution method to create concept attribution maps, using implicit differentiation [4, 25, 44]. + +Notations In this work, we consider a general supervised learning setting, where $(\pmb{x}_1,\dots,\pmb{x}_n)\in \mathcal{X}^n\subseteq \mathbb{R}^{n\times d}$ are $n$ inputs images and $(y_{1},\ldots ,y_{n})\in \mathcal{V}^{n}$ their associated labels. We are given a (machine-learnt) black-box predictor $\pmb {f}:\mathcal{X}\rightarrow \mathcal{Y}$ , which at some test input $\pmb{x}$ predicts the output $\pmb {f}(\pmb {x})$ . Without loss of generality, we establish that $\pmb{f}$ is a neural network that can be decomposed into two distinct components. The first component is a function $\pmb{g}$ that maps the input to an intermediate state, and the second component is $\pmb{h}$ , which takes this intermediate state to the output, such that $\pmb {f}(\pmb {x}) = (\pmb {h}\circ \pmb {g})(\pmb {x})$ . In this context, $\pmb {g}(\pmb {x})\subseteq \mathbb{R}^{p}$ represents the intermediate activations of $\pmb{x}$ within the network. + +![](images/a2a3380f552a846acecb3712c8a3860bdbf91e6323fd03082069489064b1d72b.jpg) +Figure 3. (1) Neural collapse (amalgamation). A classifier needs to be able to linearly separate classes by the final layer. It is commonly assumed that in order to achieve this, image activations from the same class get progressively "merged" such that these image activations converge to a one-hot vector associated with the class at the level of the logits layer. In practice, this means that different concepts get ultimately blended together along the way. (2) Recursive process. When a concept is not understood (e.g., $\mathcal{C}$ ), we propose to decompose it into multiple sub-concepts (e.g., $\mathcal{C}_1, \mathcal{C}_2, \mathcal{C}_3$ ) using the activations from an earlier layer to overcome the aforementioned neural collapse issue. (3) Example of recursive concept decomposition using CRAFT on the ImageNet class "parachute". + +![](images/4c64d3cbc2ea37f63d43a79059e87b7c1fbd58b8e3470ea5ac5cd0c4a99fb751.jpg) + +![](images/50fe647394038ab8f1cc171d224b5a1f7b125578c3fba4e4e52aae750a13caa9.jpg) + +Further, we will assume non-negative activations: $g(\boldsymbol{x}) \geq 0$ . In particular, this assumption is verified by any architecture that utilizes $ReLU$ , but any non-negative activation function works. + +# 3.1. Concept activation factorization. + +We use Non-negative matrix factorization to identify a basis for concepts based on a network's activations (Fig.4). Inspired by the approach taken in ACE [24], we will use image sub-regions to try to identify coherent concepts. + +The first step involves gathering a set of images that one wishes to explain, such as the dataset, in order to generate associated concepts. In our examples, to explain a specific class $y \in \mathcal{V}$ , we selected the set of points $\mathcal{C}$ from the dataset for which the model's predictions matched a specific class $\mathcal{C} = \{\pmb{x}_i : \pmb{f}(\pmb{x}_i) = y, 1 \leq i \leq n\}$ . It is important to emphasize that this choice is significant. The goal is not to understand how humans labeled the data, but rather to comprehend the model itself. By only selecting correctly classified images, important biases and failure cases may be missed, preventing a complete understanding of our model. + +Now that we have defined our set of images, we will proceed with selecting sub-regions of those images to identify specific concepts within a localized context. It has been observed that the implementation of segmentation masks suggested in ACE can lead to the introduction of artifacts due to the associated inpainting with a baseline value. In contrast, our proposed method takes advantage of the prevalent use of modern data augmentation techniques such as randaugment, mixup, and cutmix during the training of current models. These techniques involve the current practice of models being trained on image crops, which enables us to leverage a straightforward crop and resize function denoted by $\pi (\cdot)$ to create sub-regions (illustrated in Fig.4). By applying $\pi$ function to each image in the set $\mathcal{C}$ , we obtain an auxiliary dataset $\mathbf{X}\in \mathbb{R}^{n\times d}$ such that each entries $\mathbf{X}_i = \pi (\mathbf{x}_i)$ is an image crop. + +To discover the concept basis, we start by obtaining the + +activations for the random crops $\mathbf{A} = g(\mathbf{X})\in \mathbb{R}^{n\times p}$ . In the case where $\pmb{f}$ is a convolutional neural network, a global average pooling is applied to the activations. + +We are now ready to apply Non-negative Matrix Factorization (NMF) to decompose positive activations $\mathbf{A}$ into a product of non-negative, low-rank matrices $\mathbf{U} \in \mathbb{R}^{n \times r}$ and $\mathbf{W} \in \mathbb{R}^{p \times r}$ by solving: + +$$ +\left(\mathbf {U}, \mathbf {W}\right) = \underset {\mathbf {U} \geq 0, \mathbf {W} \geq 0} {\arg \min } \frac {1}{2} \| \mathbf {A} - \mathbf {U W} ^ {\mathsf {T}} \| _ {F} ^ {2}, \tag {1} +$$ + +where $||\cdot ||_F$ denotes the Frobenius norm. + +This decomposition of our activations $\mathbf{A}$ yields two matrices: $\mathbf{W}$ containing our Concept Activation Vectors (CAVs) and $\mathbf{U}$ that redefines the data points in our dataset according to this new basis. Moreover, this decomposition in this new basis has some interesting properties that go beyond the simple low-rank factorization - since $r \ll \min(n, p)$ . First, NMF can be understood as the joint learning of a dictionary of Concept Activation Vectors - called a "concept bank" in Fig. 4 - that maps a $\mathbb{R}^p$ basis onto $\mathbb{R}^r$ , and $\mathbf{U}$ the coefficients of the vectors $\mathbf{A}$ expressed in this new basis. The minimization of the reconstruction error $\frac{1}{2} \| \mathbf{A} - \mathbf{U} \mathbf{W} \|_F^2$ ensures that the new basis contains (mostly) relevant concepts. Intuitively, the non-negativity constraints $\mathbf{U} \geq 0$ , $\mathbf{W} \geq 0$ encourage $(i)$ $\mathbf{W}$ to be sparse (useful for creating disentangled concepts), $(ii)$ $\mathbf{U}$ to be sparse (convenient for selecting a minimal set of useful concepts) and $(iii)$ missing data to be imputed [56], which corresponds to the sparsity pattern of post-ReLU activations $\mathbf{A}$ . + +It is worth noting that each input $\mathbf{x}_i$ can be expressed as a linear combination of concepts denoted as $\mathbf{A}_i = \sum_{j=1}^{r} \mathbf{U}_{(i,j)} \mathbf{W}_j^{\top}$ . This approach is advantageous because it allows us to interpret each input as a composition of the underlying concepts. Furthermore, the strict positivity of each term - NMF is working over the anti-negative semiring, - enhances the interpretability of the decomposition. Another interesting interpretation could be that each input is represented as a superposition of concepts [13]. + +While other methods in the literature solve a similar + +problem (such as low-rank factorization using SVD or ICA), the NMF is both fast and effective and is known to yield concepts that are meaningful to humans [20,74,77]. Finally, once the concept bank $\mathbf{W}$ has been precomputed, we can associate the concept coefficients $\pmb{u}$ to any new input $\pmb{x}$ (e.g., a full image) by solving the underlying Non-Negative Least Squares (NNLS) problem $\min_{\pmb{u} \geq 0} \frac{1}{2} \| \pmb{g}(\pmb{x}) - \pmb{u} \mathbf{W}^{\top} \|_F^2$ , and therefore recover its decomposition in the concept basis. + +In essence, the core of our method can be summarized as follows: using a set of images, the idea is to re-interpret their embedding at a given layer as a composition of concepts that humans can easily understand. In the next section, we show how one can recursively apply concept activation factorizations to preceding layer for an image containing a previously computed concept. + +# 3.2. Ingredient 1: A pinch of recursivity + +One of the most apparent issues in previous work [24,77] is the need for choosing a priori a layer at which the activation maps are computed. This choice will critically affect the concepts that are identified because certain concepts get amalgamated [53] into one at different layers of the neural network, resulting in incoherent and indecipherable clusters, as illustrated in Fig. 3. We posit that this can be solved by iteratively applying our decomposition at different layer depths, and for the concepts that remain difficult to understand, by looking for their sub-concepts in earlier layers by isolating the images that contain them. This allows us to build hierarchies of concepts for each class. + +We offer a simple solution consisting of reapplying our method to a concept by performing a second step of concept activation factorization on a set of images that contain the concept $\mathcal{C}$ in order to refine it and create sub-concepts (e.g., decompose $\mathcal{C}$ into $\{\mathcal{C}_1, \mathcal{C}_2, \mathcal{C}_3\}$ ) see Fig. 3 for an illustrative example. Note that we generalize current methods in the sense that taking images $(x_1, \dots, x_n)$ that are clustered in the logits layer (belonging to the same class) and decomposing them in a previous layer - as done in [24, 77] - is a valid recursive step. For a more general case, let us assume that a set of images that contain a common concept is obtained using the first step of concept activation factorization. + +We will then take a subset of the auxiliary dataset points to refine any concept $j$ . To do this, we select the subset of points that contain the concept $\mathcal{C}_j = \{\mathbf{X}_i : \mathbf{U}_{(i,j)} > \lambda_j, 1 \leq i \leq n\}$ , where $\lambda_j$ is the 90th percentile of the values of the concept $\mathbf{U}_{(1,j)}, \ldots, \mathbf{U}_{(n,j)}$ . In other words, the 10% of images that activate the concept $j$ the most are selected for further refinement into sub-concepts. Given this new set of points, we can then re-apply the Concept Matrix Factorization method to an earlier layer $g'(\cdot)$ to obtain the sub-concepts decomposition from the initial concept - as illustrated in Fig.3. + +# 3.3. Ingredient 2: A dash of sensitivity analysis + +A major concern with concept extraction methods is that concepts that make sense to humans are not necessarily the same as those being used by a model to classify images. In order to prevent such confirmation bias during our concept analysis phase, a faithful estimate the overall importance of the extracted concepts is crucial. Kim et al. [40] proposed an importance estimator based on directional derivatives: the partial derivative of the model output with respect to the vector of concepts. While this measure is theoretically grounded, it relies on the same principle as gradient-based methods, and thus, suffers from the same pitfalls: neural network models have noisy gradients [66, 71]. Hence, the farther the chosen layer is from the output, the noisier the directional derivative score will be. + +Since we essentially want to know which concept has the greatest effect on the output of the model, it is natural to consider the field of sensitivity analysis [9,33,34,67,68]. In this section, we briefly recall the classic "total Sobol indices" and how to apply them to our problem. The complete derivation of the Sobol-Hoeffding decomposition is presented in Section D of the supplementary materials. Formally, a natural way to estimate the importance of a concept $i$ is to measure the fluctuations of the model's output $h(\mathbf{U}\mathbf{W}^{\top})$ in response to meaningful perturbations of the concept coefficient $\mathbf{U}_{(1,i)},\ldots ,\mathbf{U}_{(n,i)}$ . Concretely, we will use perturbation masks $\mathbf{M} = (M_1,\dots,M_r)\sim \mathcal{U}([0,1]^r)$ , here an i.i.d sequence of real-valued random variables, we introduce a concept fluctuation to reconstruct a perturbed activation $\tilde{\mathbf{A}} = (\mathbf{U}\odot \mathbf{M})\mathbf{W}^{\top}$ where $\odot$ denote the Hadamard product (e.g., the masks can be used to remove a concept by setting its value to zero). We can then propagate this perturbed activation to the model output $\mathbf{Y} = h(\tilde{\mathbf{A}})$ . Simply put, removing or applying perturbation of an important concept will result in a substantial variation in the output, whereas an unused concept will have minimal effect on the output. + +Finally, we can capture the importance that a concept might have as a main effect - along with its interactions with other concepts - on the model's output by calculating the expected variance that would remain if all the concepts except the $i$ were to be fixed. This yields the general definition of the total Sobol indices. + +Definition 3.1 (Total Sobol indices). The total Sobol index $S_{i}^{T}$ , which measures the contribution of a concept $i$ as well as its interactions of any order with any other concepts to the model output variance, is given by: + +$$ +\mathcal {S} _ {i} ^ {T} = \frac {\mathbb {E} _ {\boldsymbol {M} _ {\sim i}} (\mathbb {V} _ {M _ {i}} (\boldsymbol {h} ((\boldsymbol {U} \odot \boldsymbol {M}) \mathbf {W} ^ {\top}) | \boldsymbol {M} _ {\sim i}))}{\mathbb {V} (\boldsymbol {h} ((\boldsymbol {U} \odot \boldsymbol {M}) \mathbf {W} ^ {\top}))}. \tag {2} +$$ + +In practice, this index can be calculated very efficiently [36, 48, 52, 58, 72], more details on the Quasi + +![](images/a39ce196d5f54782047aaccaa4dc624346401fd6aec3c6c6250a0abc4d16b1ab.jpg) +Figure 4. Overview of CRAFT. Starting from a set of crops $\mathbf{X}$ containing a concept $\mathcal{C}$ (e.g., crops images of the class "parachute"), we compute activations $g(\mathbf{X})$ corresponding to an intermediate layer from a neural network for random image crops. We then factorize these activations into two lower-rank matrices, $(\mathbf{U}, \mathbf{W})$ . $\mathbf{W}$ is what we call a "concept bank" and is a new basis used to express the activations, while $\mathbf{U}$ corresponds to the corresponding coefficients in this new basis. We then extend the method with 3 new ingredients: (1) recursivity - by proposing to re-decompose a concept (e.g., take a new set of images containing $\mathcal{C}_1$ ) at an earlier layer, (2) a better importance estimation using Sobol indices and (3) an approach to leverage implicit differentiation to generate concept attribution maps to localize concepts in an image. + +Monte Carlo sampling and the estimator used are left in appendix D. + +# 3.4. Ingredient 3: A smidgen of implicit differentiation + +Attribution methods are useful for determining the regions deemed important by a model for its decision, but they lack information about what exactly triggered it. We have seen that we can already extract this information from the matrices $\mathbf{U}$ and $\mathbf{W}$ , but as it is, we do not know in what part of an image a given concept is represented. In this section, we will show how we can leverage attribution methods (forward and backward modes) to find where a concept is located in the input image (see Fig. 2). Forward attribution methods do not rely on any gradient computation as they only use inference processes, whereas backward methods require back-propagating through a network's layers. By application of the chain rule, computing $\partial \mathbf{U} / \partial \mathbf{X}$ requires access to $\partial \mathbf{U} / \partial \mathbf{A}$ . + +To do so, one could be tempted to solve the linear system $\mathbf{U}\mathbf{W}^{\top} = \mathbf{A}$ . However, this problem is ill-posed since $\mathbf{W}^{\top}$ is low rank. A standard approach is to calculate the Moore-Penrose pseudo-inverse $(\mathbf{W}^{\top})^{\dagger}$ , which solves rank deficient systems by looking at the minimum norm solution [2]. In practice, $(\mathbf{W}^{\top})^{\dagger}$ is computed with the Singular Value Decomposition (SVD) of $\mathbf{W}^{\top}$ . Unfortunately, SVD is also the solution to the unstructured minimization of $\frac{1}{2}\|\mathbf{A} - \mathbf{U}\mathbf{W}^{\top}\|_F^2$ by the Eckart-Young-Mirsky theorem [12]. Hence, the non-negativity constraints of the NMF are ignored, which prevents such approaches from succeeding. Other issues stem from the fact that the $\mathbf{U},\mathbf{W}$ decomposition is generally not unique. + +Our third contribution consists of tackling this problem to allow the use of attribution methods, i.e., concept attribution maps, by proposing a strategy to differentiate through the NMF block. + +Implicit differentiation of NMF block The NMF problem 1 is NP-hard [73], and it is not convex with respect to the input pair $(\mathbf{U},\mathbf{W})$ . However, fixing the value of one of the two factors and optimizing the other turns the NMF formulation into a pair of Non-Negative Least Squares (NNLS) problems, which are convex. This ensures that alternating minimization (a standard approach for NMF) of $(\mathbf{U},\mathbf{W})$ factors will eventually reach a local minimum. Each of this alternating NNLS problems fulfills the Karush-Kuhn-Tucker (KKT) conditions [38,45], which can be encoded in the so-called optimality function $\mathbf{F}$ from [4], see Eq. 9 Appendix C.2. The implicit function theorem [25] allows us to use implicit differentiation [3,25,44] to efficiently compute the Jacobians $\partial \mathbf{U} / \partial \mathbf{A}$ and $\partial \mathbf{W} / \partial \mathbf{A}$ without requiring to back-propagate through each of the iterations of the NMF solver: + +$$ +\frac {\partial (\mathbf {U} , \mathbf {W} , \bar {\mathbf {U}} , \bar {\mathbf {W}})}{\partial \mathbf {A}} = - (\partial_ {1} \mathbf {F}) ^ {- 1} \partial_ {2} \mathbf {F}. \qquad (3) +$$ + +However, this requires the dual variables $\bar{\mathbf{U}}$ and $\bar{\mathbf{W}}$ , which are not computed in scikit-learn's [54] popular implementation*. Consequently, we leverage the work of [32] and we re-implement our own solver with Jaxopt [4] based on ADMM [5], a GPU friendly algorithm (see Appendix C.2). + +Concretely, given our concepts bank $\mathbf{W}$ , the concept attribution maps of a new input $\pmb{x}$ are calculated by solving the NNLS problem $\min_{\mathbf{U} \geq 0} \frac{1}{2} \| \pmb{g}(\pmb{x}) - \mathbf{U} \mathbf{W}^{\top} \|_F^2$ . The implicit differentiation of the NMF block $\partial \mathbf{U} / \partial \mathbf{A}$ is integrated into the classic back-propagation to obtain $\partial \mathbf{U} / \partial \pmb{x}$ . Most interestingly, this technical advance enables the use of all white-box explainability methods [59, 62, 66, 69, 71, 78] to generate concept-wise attribution maps and trace the part of an image that triggered the detection of the concept by the network. Additionally, it is even possible to employ black-box + +
Husky vs. WolfLeaves"Kit Fox" vs "Red Fox"
Session n°123Utility123Utility123
Baseline55.766.262.970.176.878.658.862.258.8
Control53.361.061.40.9572.078.080.21.0260.759.248.5
AttributionsSaliency [62]53.969.673.31.0683.288.782.41.1361.760.258.2
Integ.-Grad. [71]67.472.873.21.1582.582.585.31.1159.458.358.3
SmoothGrad [66]68.775.378.01.2083.085.786.31.1350.355.061.4
GradCAM [59]77.685.784.11.3481.983.582.41.1054.452.554.1
Occlusion [76]71.075.778.11.2278.886.182.91.1051.060.255.1
Grad.-Input [61]65.863.367.91.0676.582.979.51.0550.057.662.6
ConceptsACE [24]68.871.472.71.1579.873.882.11.0548.446.546.1
CRAFTCO (ours)82.487.085.11.3878.885.589.41.1255.549.553.3
CRAFT (ours)90.697.395.51.5386.286.685.51.1556.550.649.4
+ +Table 1. Utility scores on 3 datasets from [8]. Their Utility benchmark evaluates how well explanations help users identify general rules driving classifications that readily transfer to unseen instances. At training time, users are asked to infer rules driving the decisions of the model given a set of images, and their associated predictions and explanations. At test time, the Utility metric measures the accuracy of users at predicting the model decision on novel images averaged over 3 sessions, and normalized by the baseline accuracy of users trained without explanations. The higher the Utility score, the more useful the explanation, and the more crucial the information provided is for understanding –and thus predicting the model's output– on novel samples. CRAFTCO stands for “CRAFT Concept Only” and designates an experimental condition where only global concepts are given to users, without local explanations (i.e., the concept attribution maps). The first and second best results above the baseline are in bold and underlined, respectively. + +methods [14, 47, 55, 57] since it only amounts to solving an NNLS problem. + +# 4. Experimental evaluation + +In order to evaluate the interest and the benefits brought by CRAFT, we start in Section 4.1 by assessing the practical utility of the method on a human-centered benchmark composed of 3 XAI scenarios. + +After demonstrating the usefulness of the method using these human experiments, we independently validate the 3 proposed ingredients. First, we provide evidence that recursivity allows refining concepts, making them more meaningful to humans using two additional human experiments in Section 4.2. Next, we evaluate our new Sobol estimator and show quantitatively that it provides a more faithful assessment of concept importance in Section 4.3. Finally, we run an ablation experiment that measures the interest of local explanations based on concept attribution maps coupled with global explanations. Additional experiments, including a sanity check and an example of deep dreams applied on the concept bank, as well as many other examples of local explanations for randomly picked images from ILSVRC2012, are included in Section B of the supplementary materials. We leave the discussion on the limitations of this method and on the broader impact in appendix A. + +# 4.1. Utility Evaluation + +As emphasized by Doshi-Velez et al. [11], the goal of XAI should be to develop methods that help a user better understand the behavior of deep neural network models. An + +instantiation of this idea was recently introduced by Colin & Fel et al. [8] who described an experimental framework to quantitatively measure the practical usefulness of explainability methods in real-world scenarios. For their initial setup, these authors recruited $n = 1$ , 150 online participants (evaluated over 8 unique conditions and 3 AI scenarios) – making it the largest benchmark to date in XAI. Here, we follow their framework rigorously to allow for the robust evaluation of the utility of our proposed CRAFT method and the related ACE. The 3 representative real-world scenarios are: (1) identifying bias in an AI system (using Husky vs Wolf dataset from [57]), (2) characterizing the visual strategy that are too difficult for an untrained non-expert human observer (using the Paleobotanical dataset from [75]), (3) understanding complex failure cases (using ImageNet “Red fox” vs “Kit fox” binary classification). Using this benchmark, we evaluate CRAFT, ACE, as well as CRAFT with only the global concepts (CRAFTCO) to allow for a fair comparison with ACE. To the best of our knowledge, we are the first to systematically evaluate concept-based methods against attribution methods. + +Results are shown in Table 1 and demonstrate the benefit of CRAFT, which achieves higher scores than all of the attribution methods tested as well as ACE in the first two scenarios. To date, no method appears to exceed the baseline on the third scenario suggesting that additional work is required. We also note that, in the first two scenarios, CRAFTCO is one of the best-performing methods and it always outperforms ACE – meaning that even without the local explanation of the concept attribution maps, CRAFT + +
IntruderExperts (n = 36)Laymen (n = 37)
Acc. Concept70.19%61.08%
Acc. Sub-Concept74.81% (p = 0.18)67.03% (p = 0.043)
Binary choice
Sub-Concept76.1% (p < 0.001)74.95% (p < 0.001)
Odds Ratios3.532.99
+ +Table 2. Results from the psychophysics experiments to validate the recursivity ingredient. + +largely outperforms ACE. Examples of concepts produced by CRAFT are shown in the Appendix E.1. + +# 4.2. Validation of Recursivity + +To evaluate the meaningfulness of the extracted high-level concepts, we performed psychophysics experiments with human subjects, whom we asked to answer a survey in two phases. Furthermore, we distinguished two different audiences: on the one hand, experts in machine learning, and on the other hand, people with no particular knowledge of computer vision. Both groups of participants were volunteers and did not receive any monetary compensation. Some examples of the developed interface are available the appendix E. It is important to note that this experiment was carried out independently from the utility evaluation and thus it was setup differently. + +Intruder detection experiment First, we ask users to identify the intruder out of a series of five image crops belonging to a certain class, with the odd one being taken from a different concept but still from the same class. Then, we compare the results of this intruder detection with another intruder detection, this time, using a concept (e.g., $\mathcal{C}_1$ ) coming from a layer $l$ and one of its sub-concepts (e.g., $\mathcal{C}_{12}$ in Fig.3) extracted using our recursive method. If the concept (or sub-concept) is coherent, then it should be easy for the users to find the intruder. Table 2 summarizes our results, showing that indeed both concepts and sub-concepts are coherent, and that recursivity can lead to a slightly higher understanding of the generated concepts (significant for non-experts, but not for experts) and might suggest a way to make concepts more interpretable. + +Binary choice experiment In order to test the improvement of coherence of the sub-concept generated by recursivity with respect to the larger parent concept, we showed participants an image crop belonging to both a subcluster and a parent cluster (e.g., $\pi(\boldsymbol{x}) \in \mathcal{C}_{11} \subset \mathcal{C}_1$ ) and asked them which of the two clusters (i.e., $\mathcal{C}_{11}$ or $\mathcal{C}_1$ ) seemed to accommodate the image the best. If our hypothesis is correct, then the concept refinement brought by recursivity should help form more coherent clusters. The results in Table 2 are satisfying since in both the expert and non-expert groups, the participants chose the sub-cluster more than $74\%$ of the + +time. We measure the significance of our results by fitting a binomial logistic regression to our data, and we find that both groups are more likely to choose the sub-concept cluster (at a $p < 0.001$ ). + +# 4.3. Fidelity analysis + +We propose to simultaneously verify that identified concepts are faithful to the model and that the concept importance estimator performs better than that used in TCAV [40] by using the fidelity metrics introduced in [24, 77]. These metrics are similar to the ones used for attribution methods, which consist of studying the change of the logit score when removing/adding pixels considered important. Here, we do not introduce these perturbations in the pixel space but in the concept space: once $\mathbf{U}$ and $\mathbf{W}$ are computed, we reconstruct the matrix $\mathbf{A} \approx \mathbf{U}\mathbf{W}^{\top}$ using only the most important concept (or removing the most important concept for deletion) and compute the resulting change in the output of the model. As can be seen from Fig. 5, ranking the extracted concepts using Sobol's importance score results in steeper curves than when they are sorted by their TCAV scores. We confirm that these results generalize with other matrix factorization techniques (PCA, ICA, RCA) in Section F of the Appendix. + +![](images/394c34741cab892729b7195e9895bbc32d1d6b184bbcbf102995c9a3b3e59bc5.jpg) +Figure 5. (Left) Deletion curves (lower is better). (Right) Insertion curves (higher is better). For both the deletion or insertion metrics, Sobol indices lead to better estimates (calculated on $>100\mathrm{K}$ images) of important concepts. + +![](images/cec4981062c213d3bf377157effc90255f65310f1d8320ce2b194a1c57bc6484.jpg) + +# 5. Conclusion + +We presented CRAFT, a method to automatically extract understandable concepts from deep networks, which provides insights into the model's decision-making process. Our approach uses a recursive formulation to identify concepts at the correct level of granularity, a novel method for measuring concept importance, and implicit differentiation to generate concept attribution maps. We conducted psychophysics experiments to validate the approach's usefulness and meaningfulness to human experimenters. We anticipate that our work will inspire future research in concept-based explainability methods. + +# References + +[1] Julius Adebayo, Justin Gilmer, Michael Muelly, Ian Goodfellow, Moritz Hardt, and Been Kim. Sanity checks for saliency maps. Advances in neural information processing systems, 31, 2018. 2, 3, 24 +[2] João Carlos Alves Barata and Mahir Saleh Hussein. The moore-penrose pseudoinverse: A tutorial review of the theory. Brazilian Journal of Physics, 42(1):146-165, 2012. 6 +[3] Bradley M Bell and James V Burke. Algorithmic differentiation of implicit functions and optimal values. In Advances in Automatic Differentiation, pages 67-77. Springer, 2008. 6, 19 +[4] Mathieu Blondel, Quentin Berthet, Marco Cuturi, Roy Frostig, Stephan Hoyer, Felipe Llinares-López, Fabian Pedregosa, and Jean-Philippe Vert. Efficient and modular implicit differentiation. arXiv preprint arXiv:2105.15183, 2021. 3, 6, 19 +[5] Stephen Boyd, Neal Parikh, Eric Chu, Borja Peleato, Jonathan Eckstein, et al. Distributed optimization and statistical learning via the alternating direction method of multipliers. Foundations and Trends® in Machine learning, 3(1):1-122, 2011. 6, 13, 18, 19 +[6] Chaofan Chen, Oscar Li, Daniel Tao, Alina Barnett, Cynthia Rudin, and Jonathan K Su. This looks like that: deep learning for interpretable image recognition. Advances in neural information processing systems, 32, 2019. 13 +[7] Andrzej Cichocki and Anh-Huy Phan. Fast local algorithms for large scale nonnegative matrix and tensor factorizations. IEICE transactions on fundamentals of electronics, communications and computer sciences, 92(3):708-721, 2009. 6, 19 +[8] Julien Colin, Thomas Fel, Rémi Cadène, and Thomas Serre. What i cannot predict, i do not understand: A human-centered evaluation framework for explainability methods. Advances in Neural Information Processing Systems (NeurIPS), 2021. 2, 3, 7, 21 +[9] RI Cukier, CM Fortuin, Kurt E Shuler, AG Petschek, and J Ho Schaibly. Study of the sensitivity of coupled reaction systems to uncertainties in rate coefficients. i theory. The Journal of chemical physics, 59(8):3873-3878, 1973. 5 +[10] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 2, 14, 23 +[11] Finale Doshi-Velez and Been Kim. Towards a rigorous science of interpretable machine learning. ArXiv e-print, 2017. 1, 3, 7 +[12] Carl Eckart and Gale Young. The approximation of one matrix by another of lower rank. Psychometrika, 1(3):211-218, 1936. 6 +[13] Nelson Elhage, Tristan Hume, Catherine Olsson, Nicholas Schiefer, Tom Henighan, Shauna Kravec, Zac Hatfield-Dodds, Robert Lasenby, Dawn Drain, Carol Chen, Roger Grosse, Sam McCandlish, Jared Kaplan, Dario Amodei, Martin Wattenberg, and Christopher Olah. Toy models of superposition. Transformer Circuits Thread, 2022. 4 + +[14] Thomas Fel, Rémi Cadène, Mathieu Chalvidal, Matthieu Cord, David Vigouroux, and Thomas Serre. Look at the variance! efficient black-box explanations with sobol-based sensitivity analysis. Advances in Neural Information Processing Systems, 34, 2021. 1, 3, 7, 19 +[15] Thomas Fel, Mélanie Ducoffe, David Vigouroux, Rémi Cadène, Mikael Capelle, Claire Nicodème, and Thomas Serre. Don't lie to me! robust and efficient explainability with verified perturbation analysis. arXiv preprint arXiv:2202.07728, 2022.3 +[16] Thomas Fel, Lucas Hervier, David Vigouroux, Antonin Poche, Justin Plakoo, Remi Cadene, Mathieu Chalvidal, Julien Colin, Thibaut Boissin, Louis Béthune, Agustin Picard, Claire Nicodeme, Laurent Gardes, Gregory Flandin, and Thomas Serre. Xplique: A deep learning explainability toolbox. Workshop, Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 16 +[17] Cédric Févotte and Jérôme Idier. Algorithms for nonnegative matrix factorization with the $\beta$ -divergence. Neural computation, 23(9):2421-2456, 2011. 6, 19 +[18] Ruth C Fong and Andrea Vedaldi. Interpretable explanations of black boxes by meaningful perturbation. In Proceedings of the IEEE international conference on computer vision, pages 3429-3437, 2017. 3 +[19] Ruth C Fong and Andrea Vedaldi. Interpretable explanations of black boxes by meaningful perturbation. In Proceedings of the IEEE international conference on computer vision, pages 3429-3437, 2017. 13 +[20] Xiao Fu, Kejun Huang, Nicholas D Sidiropoulos, and Wing-Kin Ma. Nonnegative matrix factorization for signal and data analytics: Identifiability, algorithms, and applications. IEEE Signal Process. Mag., 36(2):59-80, 2019. 5 +[21] Yunhao Ge, Yao Xiao, Zhi Xu, Meng Zheng, Srikrishna Karanam, Terrence Chen, Laurent Itti, and Ziyan Wu. A peek into the reasoning of neural networks: Interpreting with structural visual concepts. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2195-2204, 2021. 13 +[22] Sahara Ghalebikesabi, Lucile Ter-Minassian, Karla DiazOrdaz, and Chris C Holmes. On locality of local explanation models. Advances in Neural Information Processing Systems (NeurIPS), 2021. 3 +[23] Amirata Ghorbani, Abubakar Abid, and James Zou. Interpretation of neural networks is fragile. In Proceedings of the AAAI Conference on Artificial Intelligence (AAAI), 2017. 3 +[24] Amirata Ghorbani, James Wexler, James Y Zou, and Been Kim. Towards automatic concept-based explanations. Advances in Neural Information Processing Systems, 32, 2019, 2, 3, 4, 5, 7, 8, 13, 14, 23 +[25] Andreas Griewank and Andrea Walther. Evaluating derivatives: principles and techniques of algorithmic differentiation. SIAM, 2008. 3, 6, 19 +[26] Peter Hase and Mohit Bansal. Evaluating explainable ai: Which algorithmic explanations help users predict model behavior? In Proceedings of the Annual Meeting of the Association for Computational Linguistics (ACL Short Papers), 2020. 2, 3 + +[27] Peter Hase, Harry Xie, and Mohit Bansal. The out-of-distribution problem in explainability and search methods for feature importance explanations. Advances in Neural Information Processing Systems (NeurIPS), 2021. 2 +[28] Johannes Haug, Stefan Zürn, Peter El-Jiz, and Gjergji Kasneci. On baselines for local feature attributions. arXiv preprint arXiv:2101.00905, 2021. 3 +[29] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Identity mappings in deep residual networks. In European conference on computer vision, pages 630-645. Springer, 2016. 23 +[30] Magnus R Hestenes and Eduard Stiefel. Methods of conjugate gradients for solving. Journal of research of the National Bureau of Standards, 49(6):409, 1952. 18, 19 +[31] Cheng-Yu Hsieh, Chih-Kuan Yeh, Xuanqing Liu, Pradeep Ravikumar, Seungyeon Kim, Sanjiv Kumar, and Cho-Jui Hsieh. Evaluations and methods for explanation through robustness analysis. In Proceedings of the International Conference on Learning Representations (ICLR), 2021. 3 +[32] Kejun Huang, Nicholas D Sidiropoulos, and Athanasios P Liavas. A flexible and efficient algorithmic framework for constrained matrix and tensor factorization. IEEE Transactions on Signal Processing, 64(19):5052-5065, 2016. 6, 18 +[33] Marouane Il Idrissi, Vincent Chabridon, and Bertrand Iooss. Developments and applications of shapley effects to reliability-oriented sensitivity analysis with correlated inputs. Environmental Modelling & Software, 2021. 5 +[34] Bertrand Iooss and Paul Lemaitre. A review on global sensitivity analysis methods. In Uncertainty management in simulation-optimization of complex systems, pages 101-122. Springer, 2015. 2, 5 +[35] Alon Jacovi, Ana Marasovic, Tim Miller, and Yoav Goldberg. Formalizing trust in artificial intelligence: Prerequisites, causes and goals of human trust in ai. In Proceedings of the 2021 ACM conference on fairness, accountability, and transparency, pages 624-635, 2021. 1 +[36] Alexandre Janon, Thierry Klein, Agnes Lagnoux, Maëlle Nodet, and Clémentine Prieur. Asymptotic normality and efficiency of two sobol index estimators. *ESAIM: Probability and Statistics*, 18:342–364, 2014. 5, 20 +[37] Margot E Kaminski and Jennifer M Urban. The right to contest ai. Columbia Law Review, 121(7):1957-2048, 2021. 1 +[38] William Karush. Minima of functions of several variables with inequalities as side constraints. M. Sc. Dissertation. Dept. of Mathematics, Univ. of Chicago, 1939. 6, 18 +[39] Been Kim, Rajiv Khanna, and Oluwasanmi O Koyejo. Examples are not enough, learn to criticize! criticism for interpretability. Advances in Neural Information Processing Systems (NeurIPS), 2016. 3 +[40] Been Kim, Martin Wattenberg, Justin Gilmer, Carrie Cai, James Wexler, Fernanda Viegas, et al. Interpretability beyond feature attribution: Quantitative testing with concept activation vectors (tcav). In International conference on machine learning, pages 2668-2677. PMLR, 2018. 2, 3, 5, 8, 16, 23 +[41] Sunnie S. Y. Kim, Nicole Meister, Vikram V. Ramaswamy, Ruth Fong, and Olga Russakovsky. HIVE: Evaluating the human interpretability of visual explanations. In European Conference on Computer Vision (ECCV), 2022. 2, 3 + +[42] Pieter-Jan Kindermans, Sara Hooker, Julius Adebayo, Maximilian Alber, Kristof T. Schütt, Sven Dähne, Dumitru Erhan, and Been Kim. The (Un)reliability of Saliency Methods, pages 267-280. Springer International Publishing, Cham, 2019. 3 +[43] Mauritz Kop. Eu artificial intelligence act: The European approach to ai. In Stanford - Vienna Transatlantic Technology Law Forum, Transatlantic Antitrust and IPR Developments, Stanford University, Issue No. 2/2021. https://law.stanford.edu/publications/eu-artificial-intelligence-act-the-european-approach-to-ai/. Stanford-Vienna Transatlantic Technology Law Forum, Transatlantic Antitrust, 2021. 1 +[44] Steven George Krantz and Harold R Parks. The implicit function theorem: history, theory, and applications. Springer Science & Business Media, 2002. 3, 6, 19 +[45] Harold W Kuhn and Albert W Tucker. Nonlinear programming proceedings of the second berkeley symposium on mathematical statistics and probability. Neyman, pages 481-492, 1951. 6, 18 +[46] Daniel D Lee and H Sebastian Seung. Learning the parts of objects by non-negative matrix factorization. Nature, 401(6755):788-791, 1999. 2 +[47] Scott M Lundberg and Su-In Lee. A unified approach to interpreting model predictions. Advances in neural information processing systems, 30, 2017. 3, 7 +[48] Amandine Marrel, Bertrand Iooss, Beatrice Laurent, and Olivier Roustant. Calculations of sobol indices for the gaussian process metamodel. Reliability Engineering & System Safety, 94(3):742-751, 2009. 5 +[49] Giang Nguyen, Daeyoung Kim, and Anh Nguyen. The effectiveness of feature attribution methods and its correlation with automatic evaluation scores. Advances in Neural Information Processing Systems (NeurIPS), 2021. 2, 3 +[50] Paul Novello, Thomas Fel, and David Vigouroux. Making sense of dependence: Efficient black-box explanations using dependence measure. In Advances in Neural Information Processing Systems (NeurIPS), 2022. 1, 3 +[51] Chris Olah, Alexander Mordvintsev, and Ludwig Schubert. Feature visualization. Distill, 2017. https://distill.pub/2017/feature-visualization. 13, 16 +[52] Art B Owen. Better estimation of small sobol'sensitivity indices. ACM Transactions on Modeling and Computer Simulation (TOMACS), 23(2):1-17, 2013. 5 +[53] Vardan Papyan, XY Han, and David L Donoho. Prevalence of neural collapse during the terminal phase of deep learning training. Proceedings of the National Academy of Sciences, 117(40):24652-24663, 2020. 5 +[54] Fabian Pedregosa, Gael Varoquaux, Alexandre Gramfort, Vincent Michel, Bertrand Thirion, Olivier Grisel, Mathieu Blondel, Peter Prettenhofer, Ron Weiss, Vincent Dubourg, et al. Scikit-learn: Machine learning in python. Journal of machine learning research, 12(Oct):2825–2830, 2011. 6 +[55] Vitali Petsiuk, Abir Das, and Kate Saenko. Rise: Randomized input sampling for explanation of black-box models. arXiv preprint arXiv:1806.07421, 2018. 1, 3, 7, 19, 23 + +[56] Bin Ren, Laurent Pueyo, Christine Chen, Élodie Choquet, John H Debes, Gaspard Duchêne, François Ménard, and Marshall D Perrin. Using data imputation for signal separation in high-contrast imaging. The Astrophysical Journal, 892(2):74, 2020. 4 +[57] Marco Tulio Ribeiro, Sameer Singh, and Carlos Guestrin. " why should i trust you?" explaining the predictions of any classifier. In Proceedings of the 22nd ACM SIGKDD international conference on knowledge discovery and data mining, pages 1135-1144, 2016. 3, 7, 19 +[58] Andrea Saltelli, Paola Annoni, Ivano Azzini, Francesca Campolongo, Marco Ratto, and Stefano Tarantola. Variance based sensitivity analysis of model output. design and estimator for the total sensitivity index. Computer physics communications, 181(2):259-270, 2010. 5 +[59] Ramprasaath R Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra. Grad-cam: Visual explanations from deep networks via gradient-based localization. In Proceedings of the IEEE international conference on computer vision, pages 618-626, 2017. 1, 3, 6, 7 +[60] Hua Shen and Ting-Hao Huang. How useful are the machine-generated interpretations to general users? a human evaluation on guessing the incorrectly predicted labels. In Proceedings of the AAAI Conference on Human Computation and Crowdsourcing, volume 8, pages 168-172, 2020. 2, 3 +[61] Avanti Shrikumar, Peyton Greenside, and Anshul Kundaje. Learning important features through propagating activation differences. In Proceedings of the International Conference on Machine Learning (ICML), 2017. 1, 7 +[62] Karen Simonyan, Andrea Vedaldi, and Andrew Zisserman. Deep inside convolutional networks: Visualising image classification models and saliency maps. In In Workshop at International Conference on Learning Representations. CiteSeer, 2014. 1, 3, 6, 7 +[63] Leon Sixt, Maximilian Granz, and Tim Landgraf. When explanations lie: Why many modified bp attributions fail. In Proceedings of the International Conference on Machine Learning (ICML), 2020. 2 +[64] Leon Sixt, Martin Schuessler, Oana-Iuliana Popescu, Philipp Weiß, and Tim Landgraf. Do users benefit from interpretable vision? a user study, baseline, and dataset. Proceedings of the International Conference on Learning Representations (ICLR), 2022. 2 +[65] Dylan Slack, Sophie Hilgard, Emily Jia, Sameer Singh, and Himabindu Lakkaraju. Fooling lime and shap: Adversarial attacks on post hoc explanation methods. In Proceedings of the AAAI/ACM Conference on AI, Ethics, and Society, pages 180-186, 2020. 2, 3 + +[66] Daniel Smilkov, Nikhil Thorat, Been Kim, Fernanda Viégas, and Martin Wattenberg. Smoothgrad: removing noise by adding noise. arXiv preprint arXiv:1706.03825, 2017. 1, 3, 5, 6, 7 +[67] Ilya M Sobol. Sensitivity analysis for non-linear mathematical models. Mathematical modelling and computational experiment, 1:407-414, 1993. 2, 5, 20 +[68] Ilya M Sobol. Global sensitivity indices for nonlinear mathematical models and their monte carlo estimates. Mathematics and computers in simulation, 55(1-3):271-280, 2001. 2, 5 +[69] Jost Tobias Springenberg, Alexey Dosovitskiy, Thomas Brox, and Martin Riedmiller. Striving for simplicity: The all convolutional net. arXiv preprint arXiv:1412.6806, 2014. 1, 3, 6 +[70] Pascal Sturmfels, Scott Lundberg, and Su-In Lee. Visualizing the impact of feature attribution baselines. Distill, 2020. 3, 13 +[71] Mukund Sundararajan, Ankur Taly, and Qiqi Yan. Axiomatic attribution for deep networks. In International conference on machine learning, pages 3319-3328. PMLR, 2017. 1, 3, 5, 6, 7 +[72] Stefano Tarantola, Debora Gatelli, and Thierry Alex Mara. Random balance designs for the estimation of first order global sensitivity indices. Reliability Engineering & System Safety, 91(6):717-727, 2006. 5 +[73] Stephen A Vavasis. On the complexity of nonnegative matrix factorization. SIAM Journal on Optimization, 20(3):1364-1377, 2010. 6 +[74] Yu-Xiong Wang and Yu-Jin Zhang. Nonnegative matrix factorization: A comprehensive review. IEEE Transactions on Knowledge and Data Engineering, 25(6):1336-1353, 2013. 5 +[75] Peter Wilf, Shengping Zhang, Sharat Chikkerur, Stefan A Little, Scott L Wing, and Thomas Serre. Computer vision cracks the leaf code. Proceedings of the National Academy of Sciences, 113(12):3305-3310, 2016. 7 +[76] Matthew D Zeiler and Rob Fergus. Visualizing and understanding convolutional networks. In European conference on computer vision, pages 818-833. Springer, 2014. 1, 3, 7 +[77] Ruihan Zhang, Prashan Madumal, Tim Miller, Krista A Ehinger, and Benjamin IP Rubinstein. Invertible concept-based explanations for cnn models with non-negative concept activation vectors. arXiv preprint arXiv:2006.15417, 2020. 2, 3, 5, 8, 23 +[78] Bolei Zhou, Aditya Khosla, Agata Lapedriza, Aude Oliva, and Antonio Torralba. Learning deep features for discriminative localization. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2921-2929, 2016. 3, 6 \ No newline at end of file diff --git a/2023/CRAFT_ Concept Recursive Activation FacTorization for Explainability/images.zip b/2023/CRAFT_ Concept Recursive Activation FacTorization for Explainability/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..b3ea47d1e308a309e97dcf5da86ed8696d0415cc --- /dev/null +++ b/2023/CRAFT_ Concept Recursive Activation FacTorization for Explainability/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:99303376c239a5bdbbae7ea737e2a4668b47ec01ddfc2ce5ca623debeef8370c +size 385397 diff --git a/2023/CRAFT_ Concept Recursive Activation FacTorization for Explainability/layout.json b/2023/CRAFT_ Concept Recursive Activation FacTorization for Explainability/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..e160dbf095c65a3e3bc33d797b2393ece38d5714 --- /dev/null +++ b/2023/CRAFT_ Concept Recursive Activation FacTorization for Explainability/layout.json @@ -0,0 +1,10105 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 74, + 103, + 516, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 74, + 103, + 516, + 121 + ], + "spans": [ + { + "bbox": [ + 74, + 103, + 516, + 121 + ], + "type": "text", + "content": "CRAFT: Concept Recursive Activation FacTorization for Explainability" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 116, + 128, + 484, + 176 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 116, + 128, + 484, + 176 + ], + "spans": [ + { + "bbox": [ + 116, + 128, + 484, + 176 + ], + "type": "text", + "content": "Thomas Fel" + }, + { + "bbox": [ + 116, + 128, + 484, + 176 + ], + "type": "inline_equation", + "content": "^{1,3,5*}" + }, + { + "bbox": [ + 116, + 128, + 484, + 176 + ], + "type": "text", + "content": " Agustin Picard" + }, + { + "bbox": [ + 116, + 128, + 484, + 176 + ], + "type": "inline_equation", + "content": "^{3,6*}" + }, + { + "bbox": [ + 116, + 128, + 484, + 176 + ], + "type": "text", + "content": " Louis Bethune" + }, + { + "bbox": [ + 116, + 128, + 484, + 176 + ], + "type": "inline_equation", + "content": "^{3*}" + }, + { + "bbox": [ + 116, + 128, + 484, + 176 + ], + "type": "text", + "content": " Thibaut Boissin" + }, + { + "bbox": [ + 116, + 128, + 484, + 176 + ], + "type": "inline_equation", + "content": "^{3,4*}" + }, + { + "bbox": [ + 116, + 128, + 484, + 176 + ], + "type": "text", + "content": " David Vigouroux" + }, + { + "bbox": [ + 116, + 128, + 484, + 176 + ], + "type": "inline_equation", + "content": "^{3,4}" + }, + { + "bbox": [ + 116, + 128, + 484, + 176 + ], + "type": "text", + "content": " Julien Colin" + }, + { + "bbox": [ + 116, + 128, + 484, + 176 + ], + "type": "inline_equation", + "content": "^{1,3}" + }, + { + "bbox": [ + 116, + 128, + 484, + 176 + ], + "type": "text", + "content": " Rémi Cadène" + }, + { + "bbox": [ + 116, + 128, + 484, + 176 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 116, + 128, + 484, + 176 + ], + "type": "text", + "content": " Thomas Serre" + }, + { + "bbox": [ + 116, + 128, + 484, + 176 + ], + "type": "inline_equation", + "content": "^{1,3}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 97, + 180, + 495, + 193 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 97, + 180, + 495, + 193 + ], + "spans": [ + { + "bbox": [ + 97, + 180, + 495, + 193 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 97, + 180, + 495, + 193 + ], + "type": "text", + "content": "Carney Institute for Brain Science, Brown University, USA " + }, + { + "bbox": [ + 97, + 180, + 495, + 193 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 97, + 180, + 495, + 193 + ], + "type": "text", + "content": "Sorbonne Université, CNRS, France" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 124, + 194, + 468, + 207 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 124, + 194, + 468, + 207 + ], + "spans": [ + { + "bbox": [ + 124, + 194, + 468, + 207 + ], + "type": "text", + "content": "3Artificial and Natural Intelligence Toulouse Institute, Université de Toulouse, France" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 174, + 209, + 420, + 221 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 174, + 209, + 420, + 221 + ], + "spans": [ + { + "bbox": [ + 174, + 209, + 420, + 221 + ], + "type": "text", + "content": "4 Institut de Recherche Technologique Saint-Exupery, France" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 187, + 222, + 402, + 235 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 187, + 222, + 402, + 235 + ], + "spans": [ + { + "bbox": [ + 187, + 222, + 402, + 235 + ], + "type": "inline_equation", + "content": "^{5}" + }, + { + "bbox": [ + 187, + 222, + 402, + 235 + ], + "type": "text", + "content": " Innovation & Research Division, SNCF, " + }, + { + "bbox": [ + 187, + 222, + 402, + 235 + ], + "type": "inline_equation", + "content": "^{6}" + }, + { + "bbox": [ + 187, + 222, + 402, + 235 + ], + "type": "text", + "content": " Scalian" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 96, + 243, + 242, + 388 + ], + "blocks": [ + { + "bbox": [ + 96, + 243, + 242, + 388 + ], + "lines": [ + { + "bbox": [ + 96, + 243, + 242, + 388 + ], + "spans": [ + { + "bbox": [ + 96, + 243, + 242, + 388 + ], + "type": "image", + "image_path": "d56fc2459ec3e516eb51004c50cc917e487af368e9f0e0748d6313be63eb1089.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 400, + 547, + 467 + ], + "lines": [ + { + "bbox": [ + 46, + 400, + 547, + 467 + ], + "spans": [ + { + "bbox": [ + 46, + 400, + 547, + 467 + ], + "type": "text", + "content": "Figure 1. The \"Man on the Moon\" incorrectly classified as a \"shovel\" by an ImageNet-trained ResNet50. Heatmap generated by a classic attribution method [55] (left) vs. concept attribution maps generated with the proposed CRAFT approach (right) which highlights the two most influential concepts that drove the ResNet50's decision along with their corresponding locations. CRAFT suggests that the neural net arrived at its decision because it identified the concept of \"dirt\" " + }, + { + "bbox": [ + 46, + 400, + 547, + 467 + ], + "type": "inline_equation", + "content": "\\bullet" + }, + { + "bbox": [ + 46, + 400, + 547, + 467 + ], + "type": "text", + "content": " commonly found in members of the image class \"shovel\" and the concept of \"ski pants\" " + }, + { + "bbox": [ + 46, + 400, + 547, + 467 + ], + "type": "inline_equation", + "content": "\\bullet" + }, + { + "bbox": [ + 46, + 400, + 547, + 467 + ], + "type": "text", + "content": " typically worn by people clearing snow from their driveway with a shovel instead the correct concept of astronaut's pants (which was probably never seen during training)." + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 248, + 243, + 381, + 388 + ], + "blocks": [ + { + "bbox": [ + 248, + 243, + 381, + 388 + ], + "lines": [ + { + "bbox": [ + 248, + 243, + 381, + 388 + ], + "spans": [ + { + "bbox": [ + 248, + 243, + 381, + 388 + ], + "type": "image", + "image_path": "cfaae2cfa00822a5aaf09c97ac20222f3dbdded4a740b519378efb4996e10350.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 382, + 245, + 411, + 275 + ], + "blocks": [ + { + "bbox": [ + 382, + 245, + 411, + 275 + ], + "lines": [ + { + "bbox": [ + 382, + 245, + 411, + 275 + ], + "spans": [ + { + "bbox": [ + 382, + 245, + 411, + 275 + ], + "type": "image", + "image_path": "6835fc9574e013856aa964e78ddce99ff8752d0d9cf9e6ea644f2ba6d1955fc0.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 417, + 247, + 446, + 275 + ], + "blocks": [ + { + "bbox": [ + 417, + 247, + 446, + 275 + ], + "lines": [ + { + "bbox": [ + 417, + 247, + 446, + 275 + ], + "spans": [ + { + "bbox": [ + 417, + 247, + 446, + 275 + ], + "type": "image", + "image_path": "5483d089ef7a9bae1eb123914f9294f97bb3a9f8d11f0d49f0c90775826474f6.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 452, + 247, + 480, + 275 + ], + "blocks": [ + { + "bbox": [ + 452, + 247, + 480, + 275 + ], + "lines": [ + { + "bbox": [ + 452, + 247, + 480, + 275 + ], + "spans": [ + { + "bbox": [ + 452, + 247, + 480, + 275 + ], + "type": "image", + "image_path": "0b85a981ba4c88245dbfa77a2854edaea9ef873e88e992ad1b8b4d7a1d18dd60.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 383, + 279, + 411, + 309 + ], + "blocks": [ + { + "bbox": [ + 383, + 279, + 411, + 309 + ], + "lines": [ + { + "bbox": [ + 383, + 279, + 411, + 309 + ], + "spans": [ + { + "bbox": [ + 383, + 279, + 411, + 309 + ], + "type": "image", + "image_path": "50b2c5017faf5a3ce4184951b8c86b5099a4b2c9796403fb3dd7c7793669b945.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 417, + 279, + 446, + 309 + ], + "blocks": [ + { + "bbox": [ + 417, + 279, + 446, + 309 + ], + "lines": [ + { + "bbox": [ + 417, + 279, + 446, + 309 + ], + "spans": [ + { + "bbox": [ + 417, + 279, + 446, + 309 + ], + "type": "image", + "image_path": "ac323bfa9cd0fa3a16accf3ae3a9b9b7691b25760f3c52c28032f39284837361.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 452, + 279, + 480, + 309 + ], + "blocks": [ + { + "bbox": [ + 452, + 279, + 480, + 309 + ], + "lines": [ + { + "bbox": [ + 452, + 279, + 480, + 309 + ], + "spans": [ + { + "bbox": [ + 452, + 279, + 480, + 309 + ], + "type": "image", + "image_path": "5ea86c83b0c1d9ec6a497be15e05d4d4f8f8bdd75cbc977e2f2ea1ac10ce1a1d.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 381, + 322, + 411, + 352 + ], + "blocks": [ + { + "bbox": [ + 381, + 322, + 411, + 352 + ], + "lines": [ + { + "bbox": [ + 381, + 322, + 411, + 352 + ], + "spans": [ + { + "bbox": [ + 381, + 322, + 411, + 352 + ], + "type": "image", + "image_path": "16db1080dad5ebbf1b0dec5c82e8be9894b5f6f7a6d600f44a6a199860ae7d28.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 414, + 323, + 443, + 351 + ], + "blocks": [ + { + "bbox": [ + 414, + 323, + 443, + 351 + ], + "lines": [ + { + "bbox": [ + 414, + 323, + 443, + 351 + ], + "spans": [ + { + "bbox": [ + 414, + 323, + 443, + 351 + ], + "type": "image", + "image_path": "93b31b3e45f8d1330b6dd828eacfff7ed168ba2fae6aa85c955c36c9887ba996.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 450, + 323, + 479, + 352 + ], + "blocks": [ + { + "bbox": [ + 450, + 323, + 479, + 352 + ], + "lines": [ + { + "bbox": [ + 450, + 323, + 479, + 352 + ], + "spans": [ + { + "bbox": [ + 450, + 323, + 479, + 352 + ], + "type": "image", + "image_path": "68174e964303eabc967e664f324f7940387368b35bbb24afc6ac725c0e2a529c.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 381, + 356, + 411, + 384 + ], + "blocks": [ + { + "bbox": [ + 381, + 356, + 411, + 384 + ], + "lines": [ + { + "bbox": [ + 381, + 356, + 411, + 384 + ], + "spans": [ + { + "bbox": [ + 381, + 356, + 411, + 384 + ], + "type": "image", + "image_path": "aea7dc080165edc0f2e8ba45de0d4027c82715e57bc8b26fba4b9e2dbf260592.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 416, + 356, + 443, + 384 + ], + "blocks": [ + { + "bbox": [ + 416, + 356, + 443, + 384 + ], + "lines": [ + { + "bbox": [ + 416, + 356, + 443, + 384 + ], + "spans": [ + { + "bbox": [ + 416, + 356, + 443, + 384 + ], + "type": "image", + "image_path": "b2604fe305ebdb7f4d7a78839207afd070c12f4646315e3b2de8b16c4c7a35a3.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 450, + 356, + 479, + 386 + ], + "blocks": [ + { + "bbox": [ + 450, + 356, + 479, + 386 + ], + "lines": [ + { + "bbox": [ + 450, + 356, + 479, + 386 + ], + "spans": [ + { + "bbox": [ + 450, + 356, + 479, + 386 + ], + "type": "image", + "image_path": "87798a3af8a411c63c8d751d755be2c6b5dda99166e85ca07541112978aa008b.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + } + ], + "index": 21 + }, + { + "bbox": [ + 143, + 481, + 192, + 493 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 481, + 192, + 493 + ], + "spans": [ + { + "bbox": [ + 143, + 481, + 192, + 493 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 46, + 497, + 289, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 497, + 289, + 688 + ], + "spans": [ + { + "bbox": [ + 46, + 497, + 289, + 688 + ], + "type": "text", + "content": "Attribution methods, which employ heatmaps to identify the most influential regions of an image that impact model decisions, have gained widespread popularity as a type of explainability method. However, recent research has exposed the limited practical value of these methods, attributed in part to their narrow focus on the most prominent regions of an image - revealing \"where\" the model looks, but failing to elucidate \"what\" the model sees in those areas. In this work, we try to fill in this gap with CRAFT - a novel approach to identify both \"what\" and \"where\" by generating concept-based explanations. We introduce 3 new ingredients to the automatic concept extraction literature: (i) a recursive strategy to detect and decompose concepts across layers, (ii) a novel method for a more faithful estimation of concept importance using Sobol indices, and (iii) the use of implicit differentiation to unlock Concept Attribution Maps." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 47, + 689, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 288, + 714 + ], + "type": "text", + "content": "We conduct both human and computer vision experiments to demonstrate the benefits of the proposed approach." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 304, + 482, + 547, + 554 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 482, + 547, + 554 + ], + "spans": [ + { + "bbox": [ + 304, + 482, + 547, + 554 + ], + "type": "text", + "content": "We show that the proposed concept importance estimation technique is more faithful to the model than previous methods. When evaluating the usefulness of the method for human experimenters on a human-centered utility benchmark, we find that our approach significantly improves on two of the three test scenarios." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 306, + 563, + 387, + 575 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 563, + 387, + 575 + ], + "spans": [ + { + "bbox": [ + 306, + 563, + 387, + 575 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 304, + 581, + 547, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 581, + 547, + 689 + ], + "spans": [ + { + "bbox": [ + 304, + 581, + 547, + 689 + ], + "type": "text", + "content": "Interpreting the decisions of modern machine learning models such as neural networks remains a major challenge. Given the ever-increasing range of machine learning applications, the need for robust and reliable explainability methods continues to grow [11, 35]. Recently enacted European laws (including the General Data Protection Regulation (GDPR) [37] and the European AI act [43]) require the assessment of explainable decisions, especially those made by algorithms." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 305, + 689, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 689, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 305, + 689, + 547, + 714 + ], + "type": "text", + "content": "In order to try to meet this growing need, an array of explainability methods have already been proposed [14, 50," + } + ] + } + ], + "index": 29 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 294, + 732, + 314, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 732, + 314, + 742 + ], + "spans": [ + { + "bbox": [ + 294, + 732, + 314, + 742 + ], + "type": "text", + "content": "2711" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 116, + 76, + 348, + 204 + ], + "blocks": [ + { + "bbox": [ + 116, + 76, + 348, + 204 + ], + "lines": [ + { + "bbox": [ + 116, + 76, + 348, + 204 + ], + "spans": [ + { + "bbox": [ + 116, + 76, + 348, + 204 + ], + "type": "image", + "image_path": "052520db0224c7b7815264d63e73a1ac961132927e72f15c833437870d919206.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 45, + 209, + 545, + 287 + ], + "lines": [ + { + "bbox": [ + 45, + 209, + 545, + 287 + ], + "spans": [ + { + "bbox": [ + 45, + 209, + 545, + 287 + ], + "type": "text", + "content": "Figure 2. CRAFT results for the prediction \"chain saw\". First, our method uses Non-Negative Matrix Factorization (NMF) to extract the most relevant concepts used by the network (ResNet50V2) from the train set (ILSVRC2012 [10]). The global influence of these concepts on the predictions is then measured using Sobol indices (right panel). Finally, the method provides local explanations through concept attribution maps (heatmaps associated with a concept, and computed using grad-CAM by backpropagating through the NMF concept values with implicit differentiation). Besides, concepts can be interpreted by looking at crops that maximize the NMF coefficients. For the class \"chain saw\", the detected concepts seem to be: " + }, + { + "bbox": [ + 45, + 209, + 545, + 287 + ], + "type": "inline_equation", + "content": "\\bullet" + }, + { + "bbox": [ + 45, + 209, + 545, + 287 + ], + "type": "text", + "content": " the chainsaw engine, " + }, + { + "bbox": [ + 45, + 209, + 545, + 287 + ], + "type": "inline_equation", + "content": "\\bullet" + }, + { + "bbox": [ + 45, + 209, + 545, + 287 + ], + "type": "text", + "content": " the saw blade, " + }, + { + "bbox": [ + 45, + 209, + 545, + 287 + ], + "type": "inline_equation", + "content": "\\bullet" + }, + { + "bbox": [ + 45, + 209, + 545, + 287 + ], + "type": "text", + "content": " the human head, " + }, + { + "bbox": [ + 45, + 209, + 545, + 287 + ], + "type": "inline_equation", + "content": "\\bullet" + }, + { + "bbox": [ + 45, + 209, + 545, + 287 + ], + "type": "text", + "content": " the vegetation, " + }, + { + "bbox": [ + 45, + 209, + 545, + 287 + ], + "type": "inline_equation", + "content": "\\bullet" + }, + { + "bbox": [ + 45, + 209, + 545, + 287 + ], + "type": "text", + "content": " the jeans and " + }, + { + "bbox": [ + 45, + 209, + 545, + 287 + ], + "type": "inline_equation", + "content": "\\bullet" + }, + { + "bbox": [ + 45, + 209, + 545, + 287 + ], + "type": "text", + "content": " the tree trunk." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 356, + 85, + 479, + 191 + ], + "blocks": [ + { + "bbox": [ + 356, + 85, + 479, + 191 + ], + "lines": [ + { + "bbox": [ + 356, + 85, + 479, + 191 + ], + "spans": [ + { + "bbox": [ + 356, + 85, + 479, + 191 + ], + "type": "image", + "image_path": "61a7202338e1d1ecb53d1059ce55768474216d7dc8d5ca828866fbf4c3093902.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 297, + 289, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 297, + 289, + 498 + ], + "spans": [ + { + "bbox": [ + 46, + 297, + 289, + 498 + ], + "type": "text", + "content": "59, 61, 62, 66, 69, 71, 76]. One of the main classes of methods called attribution methods yields heatmaps that indicate the importance of individual pixels for driving a model's decision. However, these methods exhibit critical limitations [1, 27, 63, 65], as they have been shown to fail – or only marginally help – in recent human-centered benchmarks [8, 26, 41, 49, 60, 64]. It has been suggested that their limitations stem from the fact that they are only capable of explaining where in an image are the pixels that are critical to the decision but they cannot tell what visual features are actually driving decisions at these locations. In other words, they show where the model looks but not what it sees. For example, in the scenario depicted in Fig. 1, where an ImageNet-trained ResNet mistakenly identifies an image as containing a shovel, the attribution map displayed on the left fails to explain the reasoning behind this misclassification." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 498, + 289, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 498, + 289, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 498, + 289, + 714 + ], + "type": "text", + "content": "A recent approach has sought to move past attribution methods [40] by using so-called \"concepts\" to communicate information to users on how a model works. The goal is to find human-interpretable concepts in the activation space of a neural network. Although the approach exhibited potential, its practicality is significantly restricted due to the need for prior knowledge of pertinent concepts in its original formulation and, more critically, the requirement for a labeled dataset of such concepts. Several lines of work have focused on trying to automate the concept discovery process based only on the training dataset and without explicit human supervision. The most prominent of these techniques, ACE [24], uses a combination of segmentation and clustering techniques but requires heuristics to remove outliers. However, ACE provides a proof of concept that it might be possible to discover concepts automatically and at scale – without additional labeling or human supervision. Nevertheless, the approach suffers several limitations: by" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 297, + 547, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 297, + 547, + 380 + ], + "spans": [ + { + "bbox": [ + 304, + 297, + 547, + 380 + ], + "type": "text", + "content": "construction, each image segment can only belong to a single cluster, a layer has to be selected by the user to be used to retrieve the relevant concepts, and the amount of information lost during the outlier rejection phase can be a cause of concern. More recently, Zhang et al. [77] proposes to leverage matrix decompositions on internal feature maps to discover concepts." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 382, + 547, + 478 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 382, + 547, + 478 + ], + "spans": [ + { + "bbox": [ + 304, + 382, + 547, + 478 + ], + "type": "text", + "content": "Here, we try to fill these gaps with a novel method called CRAFT which uses Non-Negative Matrix Factorization (NMF) [46] for concept discovery. In contrast to other concept-based explanation methods, our approach provides an explicit link between their global and local explanations (Fig. 2) and identifies the relevant layer(s) to use to represent individual concepts (Fig. 3). Our main contributions can be described as follows:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 480, + 547, + 714 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 304, + 480, + 547, + 528 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 480, + 547, + 528 + ], + "spans": [ + { + "bbox": [ + 304, + 480, + 547, + 528 + ], + "type": "text", + "content": "(i) A novel approach for the automated extraction of high-level concepts learned by deep neural networks. We validate its practical utility to users with human psychophysics experiments." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 529, + 547, + 637 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 529, + 547, + 637 + ], + "spans": [ + { + "bbox": [ + 304, + 529, + 547, + 637 + ], + "type": "text", + "content": "(ii) A recursive procedure to automatically identify concepts and sub-concepts at the right level of granularity – starting with our decomposition at the top of the model and working our way upstream. We validate the benefit of this approach with human psychophysics experiments showing that (i) the decomposition of a concept yields more coherent sub-concepts and (ii) that the groups of points formed by these sub-concepts are more refined and appear meaningful to humans." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 639, + 547, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 639, + 547, + 687 + ], + "spans": [ + { + "bbox": [ + 304, + 639, + 547, + 687 + ], + "type": "text", + "content": "(iii) A novel technique to quantify the importance of individual concepts for a model's prediction using Sobol indices [34, 67, 68] – a technique borrowed from Sensitivity Analysis." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 689, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 689, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 689, + 547, + 714 + ], + "type": "text", + "content": "(iv) The first concept-based explainability method which produces concept attribution maps by backpropagating con" + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 732, + 315, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 732, + 315, + 742 + ], + "spans": [ + { + "bbox": [ + 294, + 732, + 315, + 742 + ], + "type": "text", + "content": "2712" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 288, + 145 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 288, + 145 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 288, + 145 + ], + "type": "text", + "content": "cept scores into the pixel space by leveraging the implicit function theorem in order to localize the pixels associated with the concept of a given input image. This effectively opens up the toolbox of both white-box [15, 59, 62, 66, 69, 71, 78] and black-box [14, 47, 55, 57] explainability methods to derive concept-wise attribution maps." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 156, + 134, + 168 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 156, + 134, + 168 + ], + "spans": [ + { + "bbox": [ + 47, + 156, + 134, + 168 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 177, + 288, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 177, + 288, + 392 + ], + "spans": [ + { + "bbox": [ + 46, + 177, + 288, + 392 + ], + "type": "text", + "content": "Attribution methods Attribution methods are widely used as post-hoc explainability techniques to determine the input variables that contribute to a model's prediction by generating importance maps, such as the ones shown in Fig.1. The first attribution method, Saliency, introduced in [62], generates a heatmap by utilizing the gradient of a given classification score with respect to the pixels. This method was later improved upon in the context of deep convolutional networks for classification in subsequent studies, such as [66, 69, 71, 76]. However, the image gradient only reflects the model's operation within an infinitesimal neighborhood around an input, which can yield misleading importance estimates [22] since gradients of large vision models are notoriously noisy [66]. That is why several methods leverage perturbations on the input image to probe the model and create importance maps that indicate the most crucial areas for the decision, such as Rise [55], Sobol [14], or more recently HSIC [50]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 393, + 288, + 489 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 393, + 288, + 489 + ], + "spans": [ + { + "bbox": [ + 46, + 393, + 288, + 489 + ], + "type": "text", + "content": "Unfortunately, a severe limitation of these approaches – apart from the fact that they only show the “where” – is that they are subject to confirmation bias: while they may appear to offer useful explanations to a user, sometimes these explanations are actually incorrect [1, 23, 65]. These limitations raise questions about their usefulness, as recent research has shown by using human-centered experiments to evaluate the utility of attribution [8, 26, 41, 49, 60]." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 490, + 288, + 668 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 490, + 288, + 668 + ], + "spans": [ + { + "bbox": [ + 46, + 490, + 288, + 668 + ], + "type": "text", + "content": "In particular, in [8], a protocol is proposed to measure the usefulness of explanations, corresponding to how much they help users identify rules driving a model's predictions (correct or incorrect) that transfer to unseen data - using the concept of meta-predictor (also called simulatability) [11, 18, 39]. The main idea is to train users to predict the output of the system using a small set of images along with associated model predictions and corresponding explanations. A method that performs well on this benchmark is said useful, as it helps users better predict the output of the model by providing meaningful information about the internal functioning of the model. This framework being agnostic to the type of explainability method, we have chosen to use it in Section 4 in order to compare CRAFT with attribution methods." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 677, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 677, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 677, + 288, + 713 + ], + "type": "text", + "content": "Concepts-based methods Kim et al. [40] introduced a method aimed at providing explanations that go beyond attribution-based approaches by measuring the impact of" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 72, + 547, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 335 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 335 + ], + "type": "text", + "content": "pre-selected concepts on a model's outputs. Although this method appears more interpretable to human users than standard attribution techniques, it requires a database of images describing the relevant concepts to be manually curated. Ghorbani et al. [24] further extended the approach to extract concepts without the need for human supervision. The approach, called ACE [24], uses a segmentation scheme on images, that belong to an image class of interest. The authors leveraged the intermediate activations of a neural network for specific image segments. These segments were resized to the appropriate input size and filled with a baseline value. The resulting activations were clustered to produce prototypes, which they referred to as \"concepts\". However, some concepts contained background segments, leading to the inclusion of uninteresting and outlier concepts. To address this, the authors implemented a post-processing cleanup step to remove these concepts, including those that were present in only one image of the class and were not representative. While this improved the interpretability of their explanations to human subjects, the use of a baseline value filled around the segments could introduce biases in the explanations [28, 31, 42, 70]." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 335, + 547, + 431 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 335, + 547, + 431 + ], + "spans": [ + { + "bbox": [ + 304, + 335, + 547, + 431 + ], + "type": "text", + "content": "Zhang et al. [77] developed a solution to the unsupervised concept discovery problem by using matrix factorizations in the latent spaces of neural networks. However, one major drawback of this method is that it operates at the level of convolutional kernels, leading to the discovery of localized concepts. For example, the concept of \"grass\" at the bottom of the image is considered to be distinct from the concept of \"grass\" at the top of the image." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 436, + 443, + 449 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 436, + 443, + 449 + ], + "spans": [ + { + "bbox": [ + 306, + 436, + 443, + 449 + ], + "type": "text", + "content": "3. Overview of the method" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 456, + 547, + 564 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 456, + 547, + 564 + ], + "spans": [ + { + "bbox": [ + 304, + 456, + 547, + 564 + ], + "type": "text", + "content": "In this section, we first describe our concept activations factorization method. Below we highlight the main differences with related work. We then proceed to introduce the three novel ingredients that make up CRAFT: (1) a method to recursively decompose concepts into sub-concepts, (2) a method to better estimate the importance of extracted concepts, and (3) a method to use any attribution method to create concept attribution maps, using implicit differentiation [4, 25, 44]." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 570, + 548, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 548, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 548, + 713 + ], + "type": "text", + "content": "Notations In this work, we consider a general supervised learning setting, where " + }, + { + "bbox": [ + 304, + 570, + 548, + 713 + ], + "type": "inline_equation", + "content": "(\\pmb{x}_1,\\dots,\\pmb{x}_n)\\in \\mathcal{X}^n\\subseteq \\mathbb{R}^{n\\times d}" + }, + { + "bbox": [ + 304, + 570, + 548, + 713 + ], + "type": "text", + "content": " are " + }, + { + "bbox": [ + 304, + 570, + 548, + 713 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 304, + 570, + 548, + 713 + ], + "type": "text", + "content": " inputs images and " + }, + { + "bbox": [ + 304, + 570, + 548, + 713 + ], + "type": "inline_equation", + "content": "(y_{1},\\ldots ,y_{n})\\in \\mathcal{V}^{n}" + }, + { + "bbox": [ + 304, + 570, + 548, + 713 + ], + "type": "text", + "content": " their associated labels. We are given a (machine-learnt) black-box predictor " + }, + { + "bbox": [ + 304, + 570, + 548, + 713 + ], + "type": "inline_equation", + "content": "\\pmb {f}:\\mathcal{X}\\rightarrow \\mathcal{Y}" + }, + { + "bbox": [ + 304, + 570, + 548, + 713 + ], + "type": "text", + "content": ", which at some test input " + }, + { + "bbox": [ + 304, + 570, + 548, + 713 + ], + "type": "inline_equation", + "content": "\\pmb{x}" + }, + { + "bbox": [ + 304, + 570, + 548, + 713 + ], + "type": "text", + "content": " predicts the output " + }, + { + "bbox": [ + 304, + 570, + 548, + 713 + ], + "type": "inline_equation", + "content": "\\pmb {f}(\\pmb {x})" + }, + { + "bbox": [ + 304, + 570, + 548, + 713 + ], + "type": "text", + "content": ". Without loss of generality, we establish that " + }, + { + "bbox": [ + 304, + 570, + 548, + 713 + ], + "type": "inline_equation", + "content": "\\pmb{f}" + }, + { + "bbox": [ + 304, + 570, + 548, + 713 + ], + "type": "text", + "content": " is a neural network that can be decomposed into two distinct components. The first component is a function " + }, + { + "bbox": [ + 304, + 570, + 548, + 713 + ], + "type": "inline_equation", + "content": "\\pmb{g}" + }, + { + "bbox": [ + 304, + 570, + 548, + 713 + ], + "type": "text", + "content": " that maps the input to an intermediate state, and the second component is " + }, + { + "bbox": [ + 304, + 570, + 548, + 713 + ], + "type": "inline_equation", + "content": "\\pmb{h}" + }, + { + "bbox": [ + 304, + 570, + 548, + 713 + ], + "type": "text", + "content": ", which takes this intermediate state to the output, such that " + }, + { + "bbox": [ + 304, + 570, + 548, + 713 + ], + "type": "inline_equation", + "content": "\\pmb {f}(\\pmb {x}) = (\\pmb {h}\\circ \\pmb {g})(\\pmb {x})" + }, + { + "bbox": [ + 304, + 570, + 548, + 713 + ], + "type": "text", + "content": ". In this context, " + }, + { + "bbox": [ + 304, + 570, + 548, + 713 + ], + "type": "inline_equation", + "content": "\\pmb {g}(\\pmb {x})\\subseteq \\mathbb{R}^{p}" + }, + { + "bbox": [ + 304, + 570, + 548, + 713 + ], + "type": "text", + "content": " represents the intermediate activations of " + }, + { + "bbox": [ + 304, + 570, + 548, + 713 + ], + "type": "inline_equation", + "content": "\\pmb{x}" + }, + { + "bbox": [ + 304, + 570, + 548, + 713 + ], + "type": "text", + "content": " within the network." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 733, + 315, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 733, + 315, + 742 + ], + "spans": [ + { + "bbox": [ + 294, + 733, + 315, + 742 + ], + "type": "text", + "content": "2713" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 77, + 72, + 178, + 159 + ], + "blocks": [ + { + "bbox": [ + 77, + 72, + 178, + 159 + ], + "lines": [ + { + "bbox": [ + 77, + 72, + 178, + 159 + ], + "spans": [ + { + "bbox": [ + 77, + 72, + 178, + 159 + ], + "type": "image", + "image_path": "a2a3380f552a846acecb3712c8a3860bdbf91e6323fd03082069489064b1d72b.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 178, + 547, + 247 + ], + "lines": [ + { + "bbox": [ + 46, + 178, + 547, + 247 + ], + "spans": [ + { + "bbox": [ + 46, + 178, + 547, + 247 + ], + "type": "text", + "content": "Figure 3. (1) Neural collapse (amalgamation). A classifier needs to be able to linearly separate classes by the final layer. It is commonly assumed that in order to achieve this, image activations from the same class get progressively \"merged\" such that these image activations converge to a one-hot vector associated with the class at the level of the logits layer. In practice, this means that different concepts get ultimately blended together along the way. (2) Recursive process. When a concept is not understood (e.g., " + }, + { + "bbox": [ + 46, + 178, + 547, + 247 + ], + "type": "inline_equation", + "content": "\\mathcal{C}" + }, + { + "bbox": [ + 46, + 178, + 547, + 247 + ], + "type": "text", + "content": "), we propose to decompose it into multiple sub-concepts (e.g., " + }, + { + "bbox": [ + 46, + 178, + 547, + 247 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_1, \\mathcal{C}_2, \\mathcal{C}_3" + }, + { + "bbox": [ + 46, + 178, + 547, + 247 + ], + "type": "text", + "content": ") using the activations from an earlier layer to overcome the aforementioned neural collapse issue. (3) Example of recursive concept decomposition using CRAFT on the ImageNet class \"parachute\"." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 179, + 72, + 282, + 158 + ], + "blocks": [ + { + "bbox": [ + 179, + 72, + 282, + 158 + ], + "lines": [ + { + "bbox": [ + 179, + 72, + 282, + 158 + ], + "spans": [ + { + "bbox": [ + 179, + 72, + 282, + 158 + ], + "type": "image", + "image_path": "4c64d3cbc2ea37f63d43a79059e87b7c1fbd58b8e3470ea5ac5cd0c4a99fb751.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 283, + 72, + 518, + 167 + ], + "blocks": [ + { + "bbox": [ + 283, + 72, + 518, + 167 + ], + "lines": [ + { + "bbox": [ + 283, + 72, + 518, + 167 + ], + "spans": [ + { + "bbox": [ + 283, + 72, + 518, + 167 + ], + "type": "image", + "image_path": "50fe647394038ab8f1cc171d224b5a1f7b125578c3fba4e4e52aae750a13caa9.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 255, + 288, + 304 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 255, + 288, + 304 + ], + "spans": [ + { + "bbox": [ + 46, + 255, + 288, + 304 + ], + "type": "text", + "content": "Further, we will assume non-negative activations: " + }, + { + "bbox": [ + 46, + 255, + 288, + 304 + ], + "type": "inline_equation", + "content": "g(\\boldsymbol{x}) \\geq 0" + }, + { + "bbox": [ + 46, + 255, + 288, + 304 + ], + "type": "text", + "content": ". In particular, this assumption is verified by any architecture that utilizes " + }, + { + "bbox": [ + 46, + 255, + 288, + 304 + ], + "type": "inline_equation", + "content": "ReLU" + }, + { + "bbox": [ + 46, + 255, + 288, + 304 + ], + "type": "text", + "content": ", but any non-negative activation function works." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 312, + 223, + 324 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 312, + 223, + 324 + ], + "spans": [ + { + "bbox": [ + 47, + 312, + 223, + 324 + ], + "type": "text", + "content": "3.1. Concept activation factorization." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 330, + 288, + 378 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 330, + 288, + 378 + ], + "spans": [ + { + "bbox": [ + 46, + 330, + 288, + 378 + ], + "type": "text", + "content": "We use Non-negative matrix factorization to identify a basis for concepts based on a network's activations (Fig.4). Inspired by the approach taken in ACE [24], we will use image sub-regions to try to identify coherent concepts." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 379, + 287, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 379, + 287, + 510 + ], + "spans": [ + { + "bbox": [ + 46, + 379, + 287, + 510 + ], + "type": "text", + "content": "The first step involves gathering a set of images that one wishes to explain, such as the dataset, in order to generate associated concepts. In our examples, to explain a specific class " + }, + { + "bbox": [ + 46, + 379, + 287, + 510 + ], + "type": "inline_equation", + "content": "y \\in \\mathcal{V}" + }, + { + "bbox": [ + 46, + 379, + 287, + 510 + ], + "type": "text", + "content": ", we selected the set of points " + }, + { + "bbox": [ + 46, + 379, + 287, + 510 + ], + "type": "inline_equation", + "content": "\\mathcal{C}" + }, + { + "bbox": [ + 46, + 379, + 287, + 510 + ], + "type": "text", + "content": " from the dataset for which the model's predictions matched a specific class " + }, + { + "bbox": [ + 46, + 379, + 287, + 510 + ], + "type": "inline_equation", + "content": "\\mathcal{C} = \\{\\pmb{x}_i : \\pmb{f}(\\pmb{x}_i) = y, 1 \\leq i \\leq n\\}" + }, + { + "bbox": [ + 46, + 379, + 287, + 510 + ], + "type": "text", + "content": ". It is important to emphasize that this choice is significant. The goal is not to understand how humans labeled the data, but rather to comprehend the model itself. By only selecting correctly classified images, important biases and failure cases may be missed, preventing a complete understanding of our model." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 510, + 288, + 701 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 510, + 288, + 701 + ], + "spans": [ + { + "bbox": [ + 46, + 510, + 288, + 701 + ], + "type": "text", + "content": "Now that we have defined our set of images, we will proceed with selecting sub-regions of those images to identify specific concepts within a localized context. It has been observed that the implementation of segmentation masks suggested in ACE can lead to the introduction of artifacts due to the associated inpainting with a baseline value. In contrast, our proposed method takes advantage of the prevalent use of modern data augmentation techniques such as randaugment, mixup, and cutmix during the training of current models. These techniques involve the current practice of models being trained on image crops, which enables us to leverage a straightforward crop and resize function denoted by " + }, + { + "bbox": [ + 46, + 510, + 288, + 701 + ], + "type": "inline_equation", + "content": "\\pi (\\cdot)" + }, + { + "bbox": [ + 46, + 510, + 288, + 701 + ], + "type": "text", + "content": " to create sub-regions (illustrated in Fig.4). By applying " + }, + { + "bbox": [ + 46, + 510, + 288, + 701 + ], + "type": "inline_equation", + "content": "\\pi" + }, + { + "bbox": [ + 46, + 510, + 288, + 701 + ], + "type": "text", + "content": " function to each image in the set " + }, + { + "bbox": [ + 46, + 510, + 288, + 701 + ], + "type": "inline_equation", + "content": "\\mathcal{C}" + }, + { + "bbox": [ + 46, + 510, + 288, + 701 + ], + "type": "text", + "content": ", we obtain an auxiliary dataset " + }, + { + "bbox": [ + 46, + 510, + 288, + 701 + ], + "type": "inline_equation", + "content": "\\mathbf{X}\\in \\mathbb{R}^{n\\times d}" + }, + { + "bbox": [ + 46, + 510, + 288, + 701 + ], + "type": "text", + "content": " such that each entries " + }, + { + "bbox": [ + 46, + 510, + 288, + 701 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_i = \\pi (\\mathbf{x}_i)" + }, + { + "bbox": [ + 46, + 510, + 288, + 701 + ], + "type": "text", + "content": " is an image crop." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 59, + 701, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 701, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 59, + 701, + 288, + 714 + ], + "type": "text", + "content": "To discover the concept basis, we start by obtaining the" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 255, + 545, + 292 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 255, + 545, + 292 + ], + "spans": [ + { + "bbox": [ + 304, + 255, + 545, + 292 + ], + "type": "text", + "content": "activations for the random crops " + }, + { + "bbox": [ + 304, + 255, + 545, + 292 + ], + "type": "inline_equation", + "content": "\\mathbf{A} = g(\\mathbf{X})\\in \\mathbb{R}^{n\\times p}" + }, + { + "bbox": [ + 304, + 255, + 545, + 292 + ], + "type": "text", + "content": ". In the case where " + }, + { + "bbox": [ + 304, + 255, + 545, + 292 + ], + "type": "inline_equation", + "content": "\\pmb{f}" + }, + { + "bbox": [ + 304, + 255, + 545, + 292 + ], + "type": "text", + "content": " is a convolutional neural network, a global average pooling is applied to the activations." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 292, + 547, + 339 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 292, + 547, + 339 + ], + "spans": [ + { + "bbox": [ + 304, + 292, + 547, + 339 + ], + "type": "text", + "content": "We are now ready to apply Non-negative Matrix Factorization (NMF) to decompose positive activations " + }, + { + "bbox": [ + 304, + 292, + 547, + 339 + ], + "type": "inline_equation", + "content": "\\mathbf{A}" + }, + { + "bbox": [ + 304, + 292, + 547, + 339 + ], + "type": "text", + "content": " into a product of non-negative, low-rank matrices " + }, + { + "bbox": [ + 304, + 292, + 547, + 339 + ], + "type": "inline_equation", + "content": "\\mathbf{U} \\in \\mathbb{R}^{n \\times r}" + }, + { + "bbox": [ + 304, + 292, + 547, + 339 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 292, + 547, + 339 + ], + "type": "inline_equation", + "content": "\\mathbf{W} \\in \\mathbb{R}^{p \\times r}" + }, + { + "bbox": [ + 304, + 292, + 547, + 339 + ], + "type": "text", + "content": " by solving:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 342, + 339, + 545, + 365 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 342, + 339, + 545, + 365 + ], + "spans": [ + { + "bbox": [ + 342, + 339, + 545, + 365 + ], + "type": "interline_equation", + "content": "\\left(\\mathbf {U}, \\mathbf {W}\\right) = \\underset {\\mathbf {U} \\geq 0, \\mathbf {W} \\geq 0} {\\arg \\min } \\frac {1}{2} \\| \\mathbf {A} - \\mathbf {U W} ^ {\\mathsf {T}} \\| _ {F} ^ {2}, \\tag {1}", + "image_path": "c8dfcaec790396b185b6b6d58e41b072e967a8854df8d2c711c12f5800057669.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 366, + 477, + 378 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 366, + 477, + 378 + ], + "spans": [ + { + "bbox": [ + 304, + 366, + 477, + 378 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 366, + 477, + 378 + ], + "type": "inline_equation", + "content": "||\\cdot ||_F" + }, + { + "bbox": [ + 304, + 366, + 477, + 378 + ], + "type": "text", + "content": " denotes the Frobenius norm." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 379, + 547, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 379, + 547, + 594 + ], + "spans": [ + { + "bbox": [ + 304, + 379, + 547, + 594 + ], + "type": "text", + "content": "This decomposition of our activations " + }, + { + "bbox": [ + 304, + 379, + 547, + 594 + ], + "type": "inline_equation", + "content": "\\mathbf{A}" + }, + { + "bbox": [ + 304, + 379, + 547, + 594 + ], + "type": "text", + "content": " yields two matrices: " + }, + { + "bbox": [ + 304, + 379, + 547, + 594 + ], + "type": "inline_equation", + "content": "\\mathbf{W}" + }, + { + "bbox": [ + 304, + 379, + 547, + 594 + ], + "type": "text", + "content": " containing our Concept Activation Vectors (CAVs) and " + }, + { + "bbox": [ + 304, + 379, + 547, + 594 + ], + "type": "inline_equation", + "content": "\\mathbf{U}" + }, + { + "bbox": [ + 304, + 379, + 547, + 594 + ], + "type": "text", + "content": " that redefines the data points in our dataset according to this new basis. Moreover, this decomposition in this new basis has some interesting properties that go beyond the simple low-rank factorization - since " + }, + { + "bbox": [ + 304, + 379, + 547, + 594 + ], + "type": "inline_equation", + "content": "r \\ll \\min(n, p)" + }, + { + "bbox": [ + 304, + 379, + 547, + 594 + ], + "type": "text", + "content": ". First, NMF can be understood as the joint learning of a dictionary of Concept Activation Vectors - called a \"concept bank\" in Fig. 4 - that maps a " + }, + { + "bbox": [ + 304, + 379, + 547, + 594 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^p" + }, + { + "bbox": [ + 304, + 379, + 547, + 594 + ], + "type": "text", + "content": " basis onto " + }, + { + "bbox": [ + 304, + 379, + 547, + 594 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^r" + }, + { + "bbox": [ + 304, + 379, + 547, + 594 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 379, + 547, + 594 + ], + "type": "inline_equation", + "content": "\\mathbf{U}" + }, + { + "bbox": [ + 304, + 379, + 547, + 594 + ], + "type": "text", + "content": " the coefficients of the vectors " + }, + { + "bbox": [ + 304, + 379, + 547, + 594 + ], + "type": "inline_equation", + "content": "\\mathbf{A}" + }, + { + "bbox": [ + 304, + 379, + 547, + 594 + ], + "type": "text", + "content": " expressed in this new basis. The minimization of the reconstruction error " + }, + { + "bbox": [ + 304, + 379, + 547, + 594 + ], + "type": "inline_equation", + "content": "\\frac{1}{2} \\| \\mathbf{A} - \\mathbf{U} \\mathbf{W} \\|_F^2" + }, + { + "bbox": [ + 304, + 379, + 547, + 594 + ], + "type": "text", + "content": " ensures that the new basis contains (mostly) relevant concepts. Intuitively, the non-negativity constraints " + }, + { + "bbox": [ + 304, + 379, + 547, + 594 + ], + "type": "inline_equation", + "content": "\\mathbf{U} \\geq 0" + }, + { + "bbox": [ + 304, + 379, + 547, + 594 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 379, + 547, + 594 + ], + "type": "inline_equation", + "content": "\\mathbf{W} \\geq 0" + }, + { + "bbox": [ + 304, + 379, + 547, + 594 + ], + "type": "text", + "content": " encourage " + }, + { + "bbox": [ + 304, + 379, + 547, + 594 + ], + "type": "inline_equation", + "content": "(i)" + }, + { + "bbox": [ + 304, + 379, + 547, + 594 + ], + "type": "inline_equation", + "content": "\\mathbf{W}" + }, + { + "bbox": [ + 304, + 379, + 547, + 594 + ], + "type": "text", + "content": " to be sparse (useful for creating disentangled concepts), " + }, + { + "bbox": [ + 304, + 379, + 547, + 594 + ], + "type": "inline_equation", + "content": "(ii)" + }, + { + "bbox": [ + 304, + 379, + 547, + 594 + ], + "type": "inline_equation", + "content": "\\mathbf{U}" + }, + { + "bbox": [ + 304, + 379, + 547, + 594 + ], + "type": "text", + "content": " to be sparse (convenient for selecting a minimal set of useful concepts) and " + }, + { + "bbox": [ + 304, + 379, + 547, + 594 + ], + "type": "inline_equation", + "content": "(iii)" + }, + { + "bbox": [ + 304, + 379, + 547, + 594 + ], + "type": "text", + "content": " missing data to be imputed [56], which corresponds to the sparsity pattern of post-ReLU activations " + }, + { + "bbox": [ + 304, + 379, + 547, + 594 + ], + "type": "inline_equation", + "content": "\\mathbf{A}" + }, + { + "bbox": [ + 304, + 379, + 547, + 594 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 594, + 547, + 701 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 547, + 701 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 547, + 701 + ], + "type": "text", + "content": "It is worth noting that each input " + }, + { + "bbox": [ + 304, + 594, + 547, + 701 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_i" + }, + { + "bbox": [ + 304, + 594, + 547, + 701 + ], + "type": "text", + "content": " can be expressed as a linear combination of concepts denoted as " + }, + { + "bbox": [ + 304, + 594, + 547, + 701 + ], + "type": "inline_equation", + "content": "\\mathbf{A}_i = \\sum_{j=1}^{r} \\mathbf{U}_{(i,j)} \\mathbf{W}_j^{\\top}" + }, + { + "bbox": [ + 304, + 594, + 547, + 701 + ], + "type": "text", + "content": ". This approach is advantageous because it allows us to interpret each input as a composition of the underlying concepts. Furthermore, the strict positivity of each term - NMF is working over the anti-negative semiring, - enhances the interpretability of the decomposition. Another interesting interpretation could be that each input is represented as a superposition of concepts [13]." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 701, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 701, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 317, + 701, + 545, + 713 + ], + "type": "text", + "content": "While other methods in the literature solve a similar" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 732, + 315, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 732, + 315, + 742 + ], + "spans": [ + { + "bbox": [ + 294, + 732, + 315, + 742 + ], + "type": "text", + "content": "2714" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 168 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 168 + ], + "type": "text", + "content": "problem (such as low-rank factorization using SVD or ICA), the NMF is both fast and effective and is known to yield concepts that are meaningful to humans [20,74,77]. Finally, once the concept bank " + }, + { + "bbox": [ + 46, + 72, + 289, + 168 + ], + "type": "inline_equation", + "content": "\\mathbf{W}" + }, + { + "bbox": [ + 46, + 72, + 289, + 168 + ], + "type": "text", + "content": " has been precomputed, we can associate the concept coefficients " + }, + { + "bbox": [ + 46, + 72, + 289, + 168 + ], + "type": "inline_equation", + "content": "\\pmb{u}" + }, + { + "bbox": [ + 46, + 72, + 289, + 168 + ], + "type": "text", + "content": " to any new input " + }, + { + "bbox": [ + 46, + 72, + 289, + 168 + ], + "type": "inline_equation", + "content": "\\pmb{x}" + }, + { + "bbox": [ + 46, + 72, + 289, + 168 + ], + "type": "text", + "content": " (e.g., a full image) by solving the underlying Non-Negative Least Squares (NNLS) problem " + }, + { + "bbox": [ + 46, + 72, + 289, + 168 + ], + "type": "inline_equation", + "content": "\\min_{\\pmb{u} \\geq 0} \\frac{1}{2} \\| \\pmb{g}(\\pmb{x}) - \\pmb{u} \\mathbf{W}^{\\top} \\|_F^2" + }, + { + "bbox": [ + 46, + 72, + 289, + 168 + ], + "type": "text", + "content": ", and therefore recover its decomposition in the concept basis." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 171, + 289, + 255 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 171, + 289, + 255 + ], + "spans": [ + { + "bbox": [ + 46, + 171, + 289, + 255 + ], + "type": "text", + "content": "In essence, the core of our method can be summarized as follows: using a set of images, the idea is to re-interpret their embedding at a given layer as a composition of concepts that humans can easily understand. In the next section, we show how one can recursively apply concept activation factorizations to preceding layer for an image containing a previously computed concept." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 268, + 237, + 282 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 268, + 237, + 282 + ], + "spans": [ + { + "bbox": [ + 47, + 268, + 237, + 282 + ], + "type": "text", + "content": "3.2. Ingredient 1: A pinch of recursivity" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 289, + 289, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 289, + 289, + 434 + ], + "spans": [ + { + "bbox": [ + 46, + 289, + 289, + 434 + ], + "type": "text", + "content": "One of the most apparent issues in previous work [24,77] is the need for choosing a priori a layer at which the activation maps are computed. This choice will critically affect the concepts that are identified because certain concepts get amalgamated [53] into one at different layers of the neural network, resulting in incoherent and indecipherable clusters, as illustrated in Fig. 3. We posit that this can be solved by iteratively applying our decomposition at different layer depths, and for the concepts that remain difficult to understand, by looking for their sub-concepts in earlier layers by isolating the images that contain them. This allows us to build hierarchies of concepts for each class." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 436, + 289, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 436, + 289, + 581 + ], + "spans": [ + { + "bbox": [ + 46, + 436, + 289, + 581 + ], + "type": "text", + "content": "We offer a simple solution consisting of reapplying our method to a concept by performing a second step of concept activation factorization on a set of images that contain the concept " + }, + { + "bbox": [ + 46, + 436, + 289, + 581 + ], + "type": "inline_equation", + "content": "\\mathcal{C}" + }, + { + "bbox": [ + 46, + 436, + 289, + 581 + ], + "type": "text", + "content": " in order to refine it and create sub-concepts (e.g., decompose " + }, + { + "bbox": [ + 46, + 436, + 289, + 581 + ], + "type": "inline_equation", + "content": "\\mathcal{C}" + }, + { + "bbox": [ + 46, + 436, + 289, + 581 + ], + "type": "text", + "content": " into " + }, + { + "bbox": [ + 46, + 436, + 289, + 581 + ], + "type": "inline_equation", + "content": "\\{\\mathcal{C}_1, \\mathcal{C}_2, \\mathcal{C}_3\\}" + }, + { + "bbox": [ + 46, + 436, + 289, + 581 + ], + "type": "text", + "content": ") see Fig. 3 for an illustrative example. Note that we generalize current methods in the sense that taking images " + }, + { + "bbox": [ + 46, + 436, + 289, + 581 + ], + "type": "inline_equation", + "content": "(x_1, \\dots, x_n)" + }, + { + "bbox": [ + 46, + 436, + 289, + 581 + ], + "type": "text", + "content": " that are clustered in the logits layer (belonging to the same class) and decomposing them in a previous layer - as done in [24, 77] - is a valid recursive step. For a more general case, let us assume that a set of images that contain a common concept is obtained using the first step of concept activation factorization." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "text", + "content": "We will then take a subset of the auxiliary dataset points to refine any concept " + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "text", + "content": ". To do this, we select the subset of points that contain the concept " + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_j = \\{\\mathbf{X}_i : \\mathbf{U}_{(i,j)} > \\lambda_j, 1 \\leq i \\leq n\\}" + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\lambda_j" + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "text", + "content": " is the 90th percentile of the values of the concept " + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{U}_{(1,j)}, \\ldots, \\mathbf{U}_{(n,j)}" + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "text", + "content": ". In other words, the 10% of images that activate the concept " + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "text", + "content": " the most are selected for further refinement into sub-concepts. Given this new set of points, we can then re-apply the Concept Matrix Factorization method to an earlier layer " + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "inline_equation", + "content": "g'(\\cdot)" + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "text", + "content": " to obtain the sub-concepts decomposition from the initial concept - as illustrated in Fig.3." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 305, + 72, + 528, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 528, + 85 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 528, + 85 + ], + "type": "text", + "content": "3.3. Ingredient 2: A dash of sensitivity analysis" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 91, + 547, + 258 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 91, + 547, + 258 + ], + "spans": [ + { + "bbox": [ + 304, + 91, + 547, + 258 + ], + "type": "text", + "content": "A major concern with concept extraction methods is that concepts that make sense to humans are not necessarily the same as those being used by a model to classify images. In order to prevent such confirmation bias during our concept analysis phase, a faithful estimate the overall importance of the extracted concepts is crucial. Kim et al. [40] proposed an importance estimator based on directional derivatives: the partial derivative of the model output with respect to the vector of concepts. While this measure is theoretically grounded, it relies on the same principle as gradient-based methods, and thus, suffers from the same pitfalls: neural network models have noisy gradients [66, 71]. Hence, the farther the chosen layer is from the output, the noisier the directional derivative score will be." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 259, + 548, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 259, + 548, + 521 + ], + "spans": [ + { + "bbox": [ + 304, + 259, + 548, + 521 + ], + "type": "text", + "content": "Since we essentially want to know which concept has the greatest effect on the output of the model, it is natural to consider the field of sensitivity analysis [9,33,34,67,68]. In this section, we briefly recall the classic \"total Sobol indices\" and how to apply them to our problem. The complete derivation of the Sobol-Hoeffding decomposition is presented in Section D of the supplementary materials. Formally, a natural way to estimate the importance of a concept " + }, + { + "bbox": [ + 304, + 259, + 548, + 521 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 259, + 548, + 521 + ], + "type": "text", + "content": " is to measure the fluctuations of the model's output " + }, + { + "bbox": [ + 304, + 259, + 548, + 521 + ], + "type": "inline_equation", + "content": "h(\\mathbf{U}\\mathbf{W}^{\\top})" + }, + { + "bbox": [ + 304, + 259, + 548, + 521 + ], + "type": "text", + "content": " in response to meaningful perturbations of the concept coefficient " + }, + { + "bbox": [ + 304, + 259, + 548, + 521 + ], + "type": "inline_equation", + "content": "\\mathbf{U}_{(1,i)},\\ldots ,\\mathbf{U}_{(n,i)}" + }, + { + "bbox": [ + 304, + 259, + 548, + 521 + ], + "type": "text", + "content": ". Concretely, we will use perturbation masks " + }, + { + "bbox": [ + 304, + 259, + 548, + 521 + ], + "type": "inline_equation", + "content": "\\mathbf{M} = (M_1,\\dots,M_r)\\sim \\mathcal{U}([0,1]^r)" + }, + { + "bbox": [ + 304, + 259, + 548, + 521 + ], + "type": "text", + "content": ", here an i.i.d sequence of real-valued random variables, we introduce a concept fluctuation to reconstruct a perturbed activation " + }, + { + "bbox": [ + 304, + 259, + 548, + 521 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{A}} = (\\mathbf{U}\\odot \\mathbf{M})\\mathbf{W}^{\\top}" + }, + { + "bbox": [ + 304, + 259, + 548, + 521 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 304, + 259, + 548, + 521 + ], + "type": "inline_equation", + "content": "\\odot" + }, + { + "bbox": [ + 304, + 259, + 548, + 521 + ], + "type": "text", + "content": " denote the Hadamard product (e.g., the masks can be used to remove a concept by setting its value to zero). We can then propagate this perturbed activation to the model output " + }, + { + "bbox": [ + 304, + 259, + 548, + 521 + ], + "type": "inline_equation", + "content": "\\mathbf{Y} = h(\\tilde{\\mathbf{A}})" + }, + { + "bbox": [ + 304, + 259, + 548, + 521 + ], + "type": "text", + "content": ". Simply put, removing or applying perturbation of an important concept will result in a substantial variation in the output, whereas an unused concept will have minimal effect on the output." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 522, + 547, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 522, + 547, + 594 + ], + "spans": [ + { + "bbox": [ + 304, + 522, + 547, + 594 + ], + "type": "text", + "content": "Finally, we can capture the importance that a concept might have as a main effect - along with its interactions with other concepts - on the model's output by calculating the expected variance that would remain if all the concepts except the " + }, + { + "bbox": [ + 304, + 522, + 547, + 594 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 522, + 547, + 594 + ], + "type": "text", + "content": " were to be fixed. This yields the general definition of the total Sobol indices." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 601, + 547, + 650 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 601, + 547, + 650 + ], + "spans": [ + { + "bbox": [ + 305, + 601, + 547, + 650 + ], + "type": "text", + "content": "Definition 3.1 (Total Sobol indices). The total Sobol index " + }, + { + "bbox": [ + 305, + 601, + 547, + 650 + ], + "type": "inline_equation", + "content": "S_{i}^{T}" + }, + { + "bbox": [ + 305, + 601, + 547, + 650 + ], + "type": "text", + "content": ", which measures the contribution of a concept " + }, + { + "bbox": [ + 305, + 601, + 547, + 650 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 305, + 601, + 547, + 650 + ], + "type": "text", + "content": " as well as its interactions of any order with any other concepts to the model output variance, is given by:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 331, + 655, + 547, + 684 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 331, + 655, + 547, + 684 + ], + "spans": [ + { + "bbox": [ + 331, + 655, + 547, + 684 + ], + "type": "interline_equation", + "content": "\\mathcal {S} _ {i} ^ {T} = \\frac {\\mathbb {E} _ {\\boldsymbol {M} _ {\\sim i}} (\\mathbb {V} _ {M _ {i}} (\\boldsymbol {h} ((\\boldsymbol {U} \\odot \\boldsymbol {M}) \\mathbf {W} ^ {\\top}) | \\boldsymbol {M} _ {\\sim i}))}{\\mathbb {V} (\\boldsymbol {h} ((\\boldsymbol {U} \\odot \\boldsymbol {M}) \\mathbf {W} ^ {\\top}))}. \\tag {2}", + "image_path": "5ecee13324f4bdfd5ff13f88c6a8e458b072557e4e1612ff464c452db526fb1c.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 689, + 548, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 689, + 548, + 715 + ], + "spans": [ + { + "bbox": [ + 305, + 689, + 548, + 715 + ], + "type": "text", + "content": "In practice, this index can be calculated very efficiently [36, 48, 52, 58, 72], more details on the Quasi" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 733, + 315, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 733, + 315, + 742 + ], + "spans": [ + { + "bbox": [ + 294, + 733, + 315, + 742 + ], + "type": "text", + "content": "2715" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 91, + 77, + 508, + 194 + ], + "blocks": [ + { + "bbox": [ + 91, + 77, + 508, + 194 + ], + "lines": [ + { + "bbox": [ + 91, + 77, + 508, + 194 + ], + "spans": [ + { + "bbox": [ + 91, + 77, + 508, + 194 + ], + "type": "image", + "image_path": "a39ce196d5f54782047aaccaa4dc624346401fd6aec3c6c6250a0abc4d16b1ab.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 45, + 201, + 547, + 278 + ], + "lines": [ + { + "bbox": [ + 45, + 201, + 547, + 278 + ], + "spans": [ + { + "bbox": [ + 45, + 201, + 547, + 278 + ], + "type": "text", + "content": "Figure 4. Overview of CRAFT. Starting from a set of crops " + }, + { + "bbox": [ + 45, + 201, + 547, + 278 + ], + "type": "inline_equation", + "content": "\\mathbf{X}" + }, + { + "bbox": [ + 45, + 201, + 547, + 278 + ], + "type": "text", + "content": " containing a concept " + }, + { + "bbox": [ + 45, + 201, + 547, + 278 + ], + "type": "inline_equation", + "content": "\\mathcal{C}" + }, + { + "bbox": [ + 45, + 201, + 547, + 278 + ], + "type": "text", + "content": " (e.g., crops images of the class \"parachute\"), we compute activations " + }, + { + "bbox": [ + 45, + 201, + 547, + 278 + ], + "type": "inline_equation", + "content": "g(\\mathbf{X})" + }, + { + "bbox": [ + 45, + 201, + 547, + 278 + ], + "type": "text", + "content": " corresponding to an intermediate layer from a neural network for random image crops. We then factorize these activations into two lower-rank matrices, " + }, + { + "bbox": [ + 45, + 201, + 547, + 278 + ], + "type": "inline_equation", + "content": "(\\mathbf{U}, \\mathbf{W})" + }, + { + "bbox": [ + 45, + 201, + 547, + 278 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 45, + 201, + 547, + 278 + ], + "type": "inline_equation", + "content": "\\mathbf{W}" + }, + { + "bbox": [ + 45, + 201, + 547, + 278 + ], + "type": "text", + "content": " is what we call a \"concept bank\" and is a new basis used to express the activations, while " + }, + { + "bbox": [ + 45, + 201, + 547, + 278 + ], + "type": "inline_equation", + "content": "\\mathbf{U}" + }, + { + "bbox": [ + 45, + 201, + 547, + 278 + ], + "type": "text", + "content": " corresponds to the corresponding coefficients in this new basis. We then extend the method with 3 new ingredients: (1) recursivity - by proposing to re-decompose a concept (e.g., take a new set of images containing " + }, + { + "bbox": [ + 45, + 201, + 547, + 278 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_1" + }, + { + "bbox": [ + 45, + 201, + 547, + 278 + ], + "type": "text", + "content": ") at an earlier layer, (2) a better importance estimation using Sobol indices and (3) an approach to leverage implicit differentiation to generate concept attribution maps to localize concepts in an image." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 289, + 288, + 315 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 289, + 288, + 315 + ], + "spans": [ + { + "bbox": [ + 46, + 289, + 288, + 315 + ], + "type": "text", + "content": "Monte Carlo sampling and the estimator used are left in appendix D." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 323, + 285, + 346 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 323, + 285, + 346 + ], + "spans": [ + { + "bbox": [ + 47, + 323, + 285, + 346 + ], + "type": "text", + "content": "3.4. Ingredient 3: A smidgen of implicit differentiation" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 348, + 288, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 348, + 288, + 514 + ], + "spans": [ + { + "bbox": [ + 46, + 348, + 288, + 514 + ], + "type": "text", + "content": "Attribution methods are useful for determining the regions deemed important by a model for its decision, but they lack information about what exactly triggered it. We have seen that we can already extract this information from the matrices " + }, + { + "bbox": [ + 46, + 348, + 288, + 514 + ], + "type": "inline_equation", + "content": "\\mathbf{U}" + }, + { + "bbox": [ + 46, + 348, + 288, + 514 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 348, + 288, + 514 + ], + "type": "inline_equation", + "content": "\\mathbf{W}" + }, + { + "bbox": [ + 46, + 348, + 288, + 514 + ], + "type": "text", + "content": ", but as it is, we do not know in what part of an image a given concept is represented. In this section, we will show how we can leverage attribution methods (forward and backward modes) to find where a concept is located in the input image (see Fig. 2). Forward attribution methods do not rely on any gradient computation as they only use inference processes, whereas backward methods require back-propagating through a network's layers. By application of the chain rule, computing " + }, + { + "bbox": [ + 46, + 348, + 288, + 514 + ], + "type": "inline_equation", + "content": "\\partial \\mathbf{U} / \\partial \\mathbf{X}" + }, + { + "bbox": [ + 46, + 348, + 288, + 514 + ], + "type": "text", + "content": " requires access to " + }, + { + "bbox": [ + 46, + 348, + 288, + 514 + ], + "type": "inline_equation", + "content": "\\partial \\mathbf{U} / \\partial \\mathbf{A}" + }, + { + "bbox": [ + 46, + 348, + 288, + 514 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 515, + 288, + 670 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 515, + 288, + 670 + ], + "spans": [ + { + "bbox": [ + 46, + 515, + 288, + 670 + ], + "type": "text", + "content": "To do so, one could be tempted to solve the linear system " + }, + { + "bbox": [ + 46, + 515, + 288, + 670 + ], + "type": "inline_equation", + "content": "\\mathbf{U}\\mathbf{W}^{\\top} = \\mathbf{A}" + }, + { + "bbox": [ + 46, + 515, + 288, + 670 + ], + "type": "text", + "content": ". However, this problem is ill-posed since " + }, + { + "bbox": [ + 46, + 515, + 288, + 670 + ], + "type": "inline_equation", + "content": "\\mathbf{W}^{\\top}" + }, + { + "bbox": [ + 46, + 515, + 288, + 670 + ], + "type": "text", + "content": " is low rank. A standard approach is to calculate the Moore-Penrose pseudo-inverse " + }, + { + "bbox": [ + 46, + 515, + 288, + 670 + ], + "type": "inline_equation", + "content": "(\\mathbf{W}^{\\top})^{\\dagger}" + }, + { + "bbox": [ + 46, + 515, + 288, + 670 + ], + "type": "text", + "content": ", which solves rank deficient systems by looking at the minimum norm solution [2]. In practice, " + }, + { + "bbox": [ + 46, + 515, + 288, + 670 + ], + "type": "inline_equation", + "content": "(\\mathbf{W}^{\\top})^{\\dagger}" + }, + { + "bbox": [ + 46, + 515, + 288, + 670 + ], + "type": "text", + "content": " is computed with the Singular Value Decomposition (SVD) of " + }, + { + "bbox": [ + 46, + 515, + 288, + 670 + ], + "type": "inline_equation", + "content": "\\mathbf{W}^{\\top}" + }, + { + "bbox": [ + 46, + 515, + 288, + 670 + ], + "type": "text", + "content": ". Unfortunately, SVD is also the solution to the unstructured minimization of " + }, + { + "bbox": [ + 46, + 515, + 288, + 670 + ], + "type": "inline_equation", + "content": "\\frac{1}{2}\\|\\mathbf{A} - \\mathbf{U}\\mathbf{W}^{\\top}\\|_F^2" + }, + { + "bbox": [ + 46, + 515, + 288, + 670 + ], + "type": "text", + "content": " by the Eckart-Young-Mirsky theorem [12]. Hence, the non-negativity constraints of the NMF are ignored, which prevents such approaches from succeeding. Other issues stem from the fact that the " + }, + { + "bbox": [ + 46, + 515, + 288, + 670 + ], + "type": "inline_equation", + "content": "\\mathbf{U},\\mathbf{W}" + }, + { + "bbox": [ + 46, + 515, + 288, + 670 + ], + "type": "text", + "content": " decomposition is generally not unique." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 671, + 288, + 718 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 671, + 288, + 718 + ], + "spans": [ + { + "bbox": [ + 46, + 671, + 288, + 718 + ], + "type": "text", + "content": "Our third contribution consists of tackling this problem to allow the use of attribution methods, i.e., concept attribution maps, by proposing a strategy to differentiate through the NMF block." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 289, + 547, + 478 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 289, + 547, + 478 + ], + "spans": [ + { + "bbox": [ + 304, + 289, + 547, + 478 + ], + "type": "text", + "content": "Implicit differentiation of NMF block The NMF problem 1 is NP-hard [73], and it is not convex with respect to the input pair " + }, + { + "bbox": [ + 304, + 289, + 547, + 478 + ], + "type": "inline_equation", + "content": "(\\mathbf{U},\\mathbf{W})" + }, + { + "bbox": [ + 304, + 289, + 547, + 478 + ], + "type": "text", + "content": ". However, fixing the value of one of the two factors and optimizing the other turns the NMF formulation into a pair of Non-Negative Least Squares (NNLS) problems, which are convex. This ensures that alternating minimization (a standard approach for NMF) of " + }, + { + "bbox": [ + 304, + 289, + 547, + 478 + ], + "type": "inline_equation", + "content": "(\\mathbf{U},\\mathbf{W})" + }, + { + "bbox": [ + 304, + 289, + 547, + 478 + ], + "type": "text", + "content": " factors will eventually reach a local minimum. Each of this alternating NNLS problems fulfills the Karush-Kuhn-Tucker (KKT) conditions [38,45], which can be encoded in the so-called optimality function " + }, + { + "bbox": [ + 304, + 289, + 547, + 478 + ], + "type": "inline_equation", + "content": "\\mathbf{F}" + }, + { + "bbox": [ + 304, + 289, + 547, + 478 + ], + "type": "text", + "content": " from [4], see Eq. 9 Appendix C.2. The implicit function theorem [25] allows us to use implicit differentiation [3,25,44] to efficiently compute the Jacobians " + }, + { + "bbox": [ + 304, + 289, + 547, + 478 + ], + "type": "inline_equation", + "content": "\\partial \\mathbf{U} / \\partial \\mathbf{A}" + }, + { + "bbox": [ + 304, + 289, + 547, + 478 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 289, + 547, + 478 + ], + "type": "inline_equation", + "content": "\\partial \\mathbf{W} / \\partial \\mathbf{A}" + }, + { + "bbox": [ + 304, + 289, + 547, + 478 + ], + "type": "text", + "content": " without requiring to back-propagate through each of the iterations of the NMF solver:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 351, + 475, + 547, + 502 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 351, + 475, + 547, + 502 + ], + "spans": [ + { + "bbox": [ + 351, + 475, + 547, + 502 + ], + "type": "interline_equation", + "content": "\\frac {\\partial (\\mathbf {U} , \\mathbf {W} , \\bar {\\mathbf {U}} , \\bar {\\mathbf {W}})}{\\partial \\mathbf {A}} = - (\\partial_ {1} \\mathbf {F}) ^ {- 1} \\partial_ {2} \\mathbf {F}. \\qquad (3)", + "image_path": "3f75b2f0109c33e60de27b57062b7ce7048ab0d23b7d6ddbea47b3a225f54e52.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 507, + 547, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 507, + 547, + 568 + ], + "spans": [ + { + "bbox": [ + 304, + 507, + 547, + 568 + ], + "type": "text", + "content": "However, this requires the dual variables " + }, + { + "bbox": [ + 304, + 507, + 547, + 568 + ], + "type": "inline_equation", + "content": "\\bar{\\mathbf{U}}" + }, + { + "bbox": [ + 304, + 507, + 547, + 568 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 507, + 547, + 568 + ], + "type": "inline_equation", + "content": "\\bar{\\mathbf{W}}" + }, + { + "bbox": [ + 304, + 507, + 547, + 568 + ], + "type": "text", + "content": ", which are not computed in scikit-learn's [54] popular implementation*. Consequently, we leverage the work of [32] and we re-implement our own solver with Jaxopt [4] based on ADMM [5], a GPU friendly algorithm (see Appendix C.2)." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 568, + 548, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 568, + 548, + 688 + ], + "spans": [ + { + "bbox": [ + 304, + 568, + 548, + 688 + ], + "type": "text", + "content": "Concretely, given our concepts bank " + }, + { + "bbox": [ + 304, + 568, + 548, + 688 + ], + "type": "inline_equation", + "content": "\\mathbf{W}" + }, + { + "bbox": [ + 304, + 568, + 548, + 688 + ], + "type": "text", + "content": ", the concept attribution maps of a new input " + }, + { + "bbox": [ + 304, + 568, + 548, + 688 + ], + "type": "inline_equation", + "content": "\\pmb{x}" + }, + { + "bbox": [ + 304, + 568, + 548, + 688 + ], + "type": "text", + "content": " are calculated by solving the NNLS problem " + }, + { + "bbox": [ + 304, + 568, + 548, + 688 + ], + "type": "inline_equation", + "content": "\\min_{\\mathbf{U} \\geq 0} \\frac{1}{2} \\| \\pmb{g}(\\pmb{x}) - \\mathbf{U} \\mathbf{W}^{\\top} \\|_F^2" + }, + { + "bbox": [ + 304, + 568, + 548, + 688 + ], + "type": "text", + "content": ". The implicit differentiation of the NMF block " + }, + { + "bbox": [ + 304, + 568, + 548, + 688 + ], + "type": "inline_equation", + "content": "\\partial \\mathbf{U} / \\partial \\mathbf{A}" + }, + { + "bbox": [ + 304, + 568, + 548, + 688 + ], + "type": "text", + "content": " is integrated into the classic back-propagation to obtain " + }, + { + "bbox": [ + 304, + 568, + 548, + 688 + ], + "type": "inline_equation", + "content": "\\partial \\mathbf{U} / \\partial \\pmb{x}" + }, + { + "bbox": [ + 304, + 568, + 548, + 688 + ], + "type": "text", + "content": ". Most interestingly, this technical advance enables the use of all white-box explainability methods [59, 62, 66, 69, 71, 78] to generate concept-wise attribution maps and trace the part of an image that triggered the detection of the concept by the network. Additionally, it is even possible to employ black-box" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 306, + 693, + 547, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 693, + 547, + 712 + ], + "spans": [ + { + "bbox": [ + 306, + 693, + 547, + 712 + ], + "type": "text", + "content": "*Scikit-learn uses a block coordinate descent algorithm [7, 17], with a randomized SVD initialization." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 294, + 733, + 315, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 733, + 315, + 742 + ], + "spans": [ + { + "bbox": [ + 294, + 733, + 315, + 742 + ], + "type": "text", + "content": "2716" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 73, + 70, + 523, + 242 + ], + "blocks": [ + { + "bbox": [ + 73, + 70, + 523, + 242 + ], + "lines": [ + { + "bbox": [ + 73, + 70, + 523, + 242 + ], + "spans": [ + { + "bbox": [ + 73, + 70, + 523, + 242 + ], + "type": "table", + "html": "
Husky vs. WolfLeaves"Kit Fox" vs "Red Fox"
Session n°123Utility123Utility123
Baseline55.766.262.970.176.878.658.862.258.8
Control53.361.061.40.9572.078.080.21.0260.759.248.5
AttributionsSaliency [62]53.969.673.31.0683.288.782.41.1361.760.258.2
Integ.-Grad. [71]67.472.873.21.1582.582.585.31.1159.458.358.3
SmoothGrad [66]68.775.378.01.2083.085.786.31.1350.355.061.4
GradCAM [59]77.685.784.11.3481.983.582.41.1054.452.554.1
Occlusion [76]71.075.778.11.2278.886.182.91.1051.060.255.1
Grad.-Input [61]65.863.367.91.0676.582.979.51.0550.057.662.6
ConceptsACE [24]68.871.472.71.1579.873.882.11.0548.446.546.1
CRAFTCO (ours)82.487.085.11.3878.885.589.41.1255.549.553.3
CRAFT (ours)90.697.395.51.5386.286.685.51.1556.550.649.4
", + "image_path": "9e37786f4a8d60e36f6b7046af4bc29705db12b2d17e4c6a39426b11ef62903b.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 251, + 548, + 340 + ], + "lines": [ + { + "bbox": [ + 46, + 251, + 548, + 340 + ], + "spans": [ + { + "bbox": [ + 46, + 251, + 548, + 340 + ], + "type": "text", + "content": "Table 1. Utility scores on 3 datasets from [8]. Their Utility benchmark evaluates how well explanations help users identify general rules driving classifications that readily transfer to unseen instances. At training time, users are asked to infer rules driving the decisions of the model given a set of images, and their associated predictions and explanations. At test time, the Utility metric measures the accuracy of users at predicting the model decision on novel images averaged over 3 sessions, and normalized by the baseline accuracy of users trained without explanations. The higher the Utility score, the more useful the explanation, and the more crucial the information provided is for understanding –and thus predicting the model's output– on novel samples. CRAFTCO stands for “CRAFT Concept Only” and designates an experimental condition where only global concepts are given to users, without local explanations (i.e., the concept attribution maps). The first and second best results above the baseline are in bold and underlined, respectively." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 350, + 287, + 374 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 350, + 287, + 374 + ], + "spans": [ + { + "bbox": [ + 46, + 350, + 287, + 374 + ], + "type": "text", + "content": "methods [14, 47, 55, 57] since it only amounts to solving an NNLS problem." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 381, + 188, + 395 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 381, + 188, + 395 + ], + "spans": [ + { + "bbox": [ + 47, + 381, + 188, + 395 + ], + "type": "text", + "content": "4. Experimental evaluation" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 401, + 288, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 401, + 288, + 449 + ], + "spans": [ + { + "bbox": [ + 46, + 401, + 288, + 449 + ], + "type": "text", + "content": "In order to evaluate the interest and the benefits brought by CRAFT, we start in Section 4.1 by assessing the practical utility of the method on a human-centered benchmark composed of 3 XAI scenarios." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 449, + 289, + 654 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 449, + 289, + 654 + ], + "spans": [ + { + "bbox": [ + 46, + 449, + 289, + 654 + ], + "type": "text", + "content": "After demonstrating the usefulness of the method using these human experiments, we independently validate the 3 proposed ingredients. First, we provide evidence that recursivity allows refining concepts, making them more meaningful to humans using two additional human experiments in Section 4.2. Next, we evaluate our new Sobol estimator and show quantitatively that it provides a more faithful assessment of concept importance in Section 4.3. Finally, we run an ablation experiment that measures the interest of local explanations based on concept attribution maps coupled with global explanations. Additional experiments, including a sanity check and an example of deep dreams applied on the concept bank, as well as many other examples of local explanations for randomly picked images from ILSVRC2012, are included in Section B of the supplementary materials. We leave the discussion on the limitations of this method and on the broader impact in appendix A." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 659, + 153, + 672 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 659, + 153, + 672 + ], + "spans": [ + { + "bbox": [ + 47, + 659, + 153, + 672 + ], + "type": "text", + "content": "4.1. Utility Evaluation" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 677, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 288, + 713 + ], + "type": "text", + "content": "As emphasized by Doshi-Velez et al. [11], the goal of XAI should be to develop methods that help a user better understand the behavior of deep neural network models. An" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 350, + 547, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 350, + 547, + 601 + ], + "spans": [ + { + "bbox": [ + 304, + 350, + 547, + 601 + ], + "type": "text", + "content": "instantiation of this idea was recently introduced by Colin & Fel et al. [8] who described an experimental framework to quantitatively measure the practical usefulness of explainability methods in real-world scenarios. For their initial setup, these authors recruited " + }, + { + "bbox": [ + 304, + 350, + 547, + 601 + ], + "type": "inline_equation", + "content": "n = 1" + }, + { + "bbox": [ + 304, + 350, + 547, + 601 + ], + "type": "text", + "content": ", 150 online participants (evaluated over 8 unique conditions and 3 AI scenarios) – making it the largest benchmark to date in XAI. Here, we follow their framework rigorously to allow for the robust evaluation of the utility of our proposed CRAFT method and the related ACE. The 3 representative real-world scenarios are: (1) identifying bias in an AI system (using Husky vs Wolf dataset from [57]), (2) characterizing the visual strategy that are too difficult for an untrained non-expert human observer (using the Paleobotanical dataset from [75]), (3) understanding complex failure cases (using ImageNet “Red fox” vs “Kit fox” binary classification). Using this benchmark, we evaluate CRAFT, ACE, as well as CRAFT with only the global concepts (CRAFTCO) to allow for a fair comparison with ACE. To the best of our knowledge, we are the first to systematically evaluate concept-based methods against attribution methods." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 605, + 548, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 548, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 548, + 715 + ], + "type": "text", + "content": "Results are shown in Table 1 and demonstrate the benefit of CRAFT, which achieves higher scores than all of the attribution methods tested as well as ACE in the first two scenarios. To date, no method appears to exceed the baseline on the third scenario suggesting that additional work is required. We also note that, in the first two scenarios, CRAFTCO is one of the best-performing methods and it always outperforms ACE – meaning that even without the local explanation of the concept attribution maps, CRAFT" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 733, + 315, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 733, + 315, + 742 + ], + "spans": [ + { + "bbox": [ + 294, + 733, + 315, + 742 + ], + "type": "text", + "content": "2717" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 50, + 70, + 286, + 168 + ], + "blocks": [ + { + "bbox": [ + 50, + 70, + 286, + 168 + ], + "lines": [ + { + "bbox": [ + 50, + 70, + 286, + 168 + ], + "spans": [ + { + "bbox": [ + 50, + 70, + 286, + 168 + ], + "type": "table", + "html": "
IntruderExperts (n = 36)Laymen (n = 37)
Acc. Concept70.19%61.08%
Acc. Sub-Concept74.81% (p = 0.18)67.03% (p = 0.043)
Binary choice
Sub-Concept76.1% (p < 0.001)74.95% (p < 0.001)
Odds Ratios3.532.99
", + "image_path": "20516a634912ecd4577feb6f45635f1a652596a5e85a8f6132459343b7079681.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 176, + 288, + 199 + ], + "lines": [ + { + "bbox": [ + 46, + 176, + 288, + 199 + ], + "spans": [ + { + "bbox": [ + 46, + 176, + 288, + 199 + ], + "type": "text", + "content": "Table 2. Results from the psychophysics experiments to validate the recursivity ingredient." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 206, + 287, + 232 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 206, + 287, + 232 + ], + "spans": [ + { + "bbox": [ + 46, + 206, + 287, + 232 + ], + "type": "text", + "content": "largely outperforms ACE. Examples of concepts produced by CRAFT are shown in the Appendix E.1." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 239, + 186, + 253 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 239, + 186, + 253 + ], + "spans": [ + { + "bbox": [ + 47, + 239, + 186, + 253 + ], + "type": "text", + "content": "4.2. Validation of Recursivity" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 258, + 288, + 402 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 258, + 288, + 402 + ], + "spans": [ + { + "bbox": [ + 46, + 258, + 288, + 402 + ], + "type": "text", + "content": "To evaluate the meaningfulness of the extracted high-level concepts, we performed psychophysics experiments with human subjects, whom we asked to answer a survey in two phases. Furthermore, we distinguished two different audiences: on the one hand, experts in machine learning, and on the other hand, people with no particular knowledge of computer vision. Both groups of participants were volunteers and did not receive any monetary compensation. Some examples of the developed interface are available the appendix E. It is important to note that this experiment was carried out independently from the utility evaluation and thus it was setup differently." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 403, + 288, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 403, + 288, + 582 + ], + "spans": [ + { + "bbox": [ + 46, + 403, + 288, + 582 + ], + "type": "text", + "content": "Intruder detection experiment First, we ask users to identify the intruder out of a series of five image crops belonging to a certain class, with the odd one being taken from a different concept but still from the same class. Then, we compare the results of this intruder detection with another intruder detection, this time, using a concept (e.g., " + }, + { + "bbox": [ + 46, + 403, + 288, + 582 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_1" + }, + { + "bbox": [ + 46, + 403, + 288, + 582 + ], + "type": "text", + "content": ") coming from a layer " + }, + { + "bbox": [ + 46, + 403, + 288, + 582 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 46, + 403, + 288, + 582 + ], + "type": "text", + "content": " and one of its sub-concepts (e.g., " + }, + { + "bbox": [ + 46, + 403, + 288, + 582 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_{12}" + }, + { + "bbox": [ + 46, + 403, + 288, + 582 + ], + "type": "text", + "content": " in Fig.3) extracted using our recursive method. If the concept (or sub-concept) is coherent, then it should be easy for the users to find the intruder. Table 2 summarizes our results, showing that indeed both concepts and sub-concepts are coherent, and that recursivity can lead to a slightly higher understanding of the generated concepts (significant for non-experts, but not for experts) and might suggest a way to make concepts more interpretable." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 582, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 582, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 582, + 288, + 714 + ], + "type": "text", + "content": "Binary choice experiment In order to test the improvement of coherence of the sub-concept generated by recursivity with respect to the larger parent concept, we showed participants an image crop belonging to both a subcluster and a parent cluster (e.g., " + }, + { + "bbox": [ + 46, + 582, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\pi(\\boldsymbol{x}) \\in \\mathcal{C}_{11} \\subset \\mathcal{C}_1" + }, + { + "bbox": [ + 46, + 582, + 288, + 714 + ], + "type": "text", + "content": ") and asked them which of the two clusters (i.e., " + }, + { + "bbox": [ + 46, + 582, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_{11}" + }, + { + "bbox": [ + 46, + 582, + 288, + 714 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 46, + 582, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_1" + }, + { + "bbox": [ + 46, + 582, + 288, + 714 + ], + "type": "text", + "content": ") seemed to accommodate the image the best. If our hypothesis is correct, then the concept refinement brought by recursivity should help form more coherent clusters. The results in Table 2 are satisfying since in both the expert and non-expert groups, the participants chose the sub-cluster more than " + }, + { + "bbox": [ + 46, + 582, + 288, + 714 + ], + "type": "inline_equation", + "content": "74\\%" + }, + { + "bbox": [ + 46, + 582, + 288, + 714 + ], + "type": "text", + "content": " of the" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 72, + 547, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 120 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 120 + ], + "type": "text", + "content": "time. We measure the significance of our results by fitting a binomial logistic regression to our data, and we find that both groups are more likely to choose the sub-concept cluster (at a " + }, + { + "bbox": [ + 304, + 72, + 547, + 120 + ], + "type": "inline_equation", + "content": "p < 0.001" + }, + { + "bbox": [ + 304, + 72, + 547, + 120 + ], + "type": "text", + "content": ")." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 127, + 405, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 127, + 405, + 140 + ], + "spans": [ + { + "bbox": [ + 305, + 127, + 405, + 140 + ], + "type": "text", + "content": "4.3. Fidelity analysis" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 145, + 547, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 145, + 547, + 361 + ], + "spans": [ + { + "bbox": [ + 304, + 145, + 547, + 361 + ], + "type": "text", + "content": "We propose to simultaneously verify that identified concepts are faithful to the model and that the concept importance estimator performs better than that used in TCAV [40] by using the fidelity metrics introduced in [24, 77]. These metrics are similar to the ones used for attribution methods, which consist of studying the change of the logit score when removing/adding pixels considered important. Here, we do not introduce these perturbations in the pixel space but in the concept space: once " + }, + { + "bbox": [ + 304, + 145, + 547, + 361 + ], + "type": "inline_equation", + "content": "\\mathbf{U}" + }, + { + "bbox": [ + 304, + 145, + 547, + 361 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 145, + 547, + 361 + ], + "type": "inline_equation", + "content": "\\mathbf{W}" + }, + { + "bbox": [ + 304, + 145, + 547, + 361 + ], + "type": "text", + "content": " are computed, we reconstruct the matrix " + }, + { + "bbox": [ + 304, + 145, + 547, + 361 + ], + "type": "inline_equation", + "content": "\\mathbf{A} \\approx \\mathbf{U}\\mathbf{W}^{\\top}" + }, + { + "bbox": [ + 304, + 145, + 547, + 361 + ], + "type": "text", + "content": " using only the most important concept (or removing the most important concept for deletion) and compute the resulting change in the output of the model. As can be seen from Fig. 5, ranking the extracted concepts using Sobol's importance score results in steeper curves than when they are sorted by their TCAV scores. We confirm that these results generalize with other matrix factorization techniques (PCA, ICA, RCA) in Section F of the Appendix." + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 307, + 371, + 430, + 468 + ], + "blocks": [ + { + "bbox": [ + 307, + 371, + 430, + 468 + ], + "lines": [ + { + "bbox": [ + 307, + 371, + 430, + 468 + ], + "spans": [ + { + "bbox": [ + 307, + 371, + 430, + 468 + ], + "type": "image", + "image_path": "394c34741cab892729b7195e9895bbc32d1d6b184bbcbf102995c9a3b3e59bc5.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 476, + 547, + 521 + ], + "lines": [ + { + "bbox": [ + 304, + 476, + 547, + 521 + ], + "spans": [ + { + "bbox": [ + 304, + 476, + 547, + 521 + ], + "type": "text", + "content": "Figure 5. (Left) Deletion curves (lower is better). (Right) Insertion curves (higher is better). For both the deletion or insertion metrics, Sobol indices lead to better estimates (calculated on " + }, + { + "bbox": [ + 304, + 476, + 547, + 521 + ], + "type": "inline_equation", + "content": ">100\\mathrm{K}" + }, + { + "bbox": [ + 304, + 476, + 547, + 521 + ], + "type": "text", + "content": " images) of important concepts." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 433, + 371, + 545, + 468 + ], + "blocks": [ + { + "bbox": [ + 433, + 371, + 545, + 468 + ], + "lines": [ + { + "bbox": [ + 433, + 371, + 545, + 468 + ], + "spans": [ + { + "bbox": [ + 433, + 371, + 545, + 468 + ], + "type": "image", + "image_path": "cec4981062c213d3bf377157effc90255f65310f1d8320ce2b194a1c57bc6484.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 525, + 379, + 537 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 525, + 379, + 537 + ], + "spans": [ + { + "bbox": [ + 306, + 525, + 379, + 537 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 544, + 547, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 544, + 547, + 677 + ], + "spans": [ + { + "bbox": [ + 304, + 544, + 547, + 677 + ], + "type": "text", + "content": "We presented CRAFT, a method to automatically extract understandable concepts from deep networks, which provides insights into the model's decision-making process. Our approach uses a recursive formulation to identify concepts at the correct level of granularity, a novel method for measuring concept importance, and implicit differentiation to generate concept attribution maps. We conducted psychophysics experiments to validate the approach's usefulness and meaningfulness to human experimenters. We anticipate that our work will inspire future research in concept-based explainability methods." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 733, + 315, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 733, + 315, + 742 + ], + "spans": [ + { + "bbox": [ + 294, + 733, + 315, + 742 + ], + "type": "text", + "content": "2718" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 706 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 53, + 91, + 288, + 133 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 288, + 133 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 288, + 133 + ], + "type": "text", + "content": "[1] Julius Adebayo, Justin Gilmer, Michael Muelly, Ian Goodfellow, Moritz Hardt, and Been Kim. Sanity checks for saliency maps. Advances in neural information processing systems, 31, 2018. 2, 3, 24" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 136, + 288, + 168 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 136, + 288, + 168 + ], + "spans": [ + { + "bbox": [ + 53, + 136, + 288, + 168 + ], + "type": "text", + "content": "[2] João Carlos Alves Barata and Mahir Saleh Hussein. The moore-penrose pseudoinverse: A tutorial review of the theory. Brazilian Journal of Physics, 42(1):146-165, 2012. 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 170, + 288, + 212 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 170, + 288, + 212 + ], + "spans": [ + { + "bbox": [ + 53, + 170, + 288, + 212 + ], + "type": "text", + "content": "[3] Bradley M Bell and James V Burke. Algorithmic differentiation of implicit functions and optimal values. In Advances in Automatic Differentiation, pages 67-77. Springer, 2008. 6, 19" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 214, + 288, + 269 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 214, + 288, + 269 + ], + "spans": [ + { + "bbox": [ + 53, + 214, + 288, + 269 + ], + "type": "text", + "content": "[4] Mathieu Blondel, Quentin Berthet, Marco Cuturi, Roy Frostig, Stephan Hoyer, Felipe Llinares-López, Fabian Pedregosa, and Jean-Philippe Vert. Efficient and modular implicit differentiation. arXiv preprint arXiv:2105.15183, 2021. 3, 6, 19" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 270, + 288, + 324 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 270, + 288, + 324 + ], + "spans": [ + { + "bbox": [ + 53, + 270, + 288, + 324 + ], + "type": "text", + "content": "[5] Stephen Boyd, Neal Parikh, Eric Chu, Borja Peleato, Jonathan Eckstein, et al. Distributed optimization and statistical learning via the alternating direction method of multipliers. Foundations and Trends® in Machine learning, 3(1):1-122, 2011. 6, 13, 18, 19" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 326, + 288, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 326, + 288, + 369 + ], + "spans": [ + { + "bbox": [ + 53, + 326, + 288, + 369 + ], + "type": "text", + "content": "[6] Chaofan Chen, Oscar Li, Daniel Tao, Alina Barnett, Cynthia Rudin, and Jonathan K Su. This looks like that: deep learning for interpretable image recognition. Advances in neural information processing systems, 32, 2019. 13" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 371, + 288, + 424 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 371, + 288, + 424 + ], + "spans": [ + { + "bbox": [ + 53, + 371, + 288, + 424 + ], + "type": "text", + "content": "[7] Andrzej Cichocki and Anh-Huy Phan. Fast local algorithms for large scale nonnegative matrix and tensor factorizations. IEICE transactions on fundamentals of electronics, communications and computer sciences, 92(3):708-721, 2009. 6, 19" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 426, + 288, + 481 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 426, + 288, + 481 + ], + "spans": [ + { + "bbox": [ + 53, + 426, + 288, + 481 + ], + "type": "text", + "content": "[8] Julien Colin, Thomas Fel, Rémi Cadène, and Thomas Serre. What i cannot predict, i do not understand: A human-centered evaluation framework for explainability methods. Advances in Neural Information Processing Systems (NeurIPS), 2021. 2, 3, 7, 21" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 483, + 288, + 526 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 483, + 288, + 526 + ], + "spans": [ + { + "bbox": [ + 53, + 483, + 288, + 526 + ], + "type": "text", + "content": "[9] RI Cukier, CM Fortuin, Kurt E Shuler, AG Petschek, and J Ho Schaibly. Study of the sensitivity of coupled reaction systems to uncertainties in rate coefficients. i theory. The Journal of chemical physics, 59(8):3873-3878, 1973. 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 528, + 288, + 571 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 528, + 288, + 571 + ], + "spans": [ + { + "bbox": [ + 48, + 528, + 288, + 571 + ], + "type": "text", + "content": "[10] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 2, 14, 23" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 572, + 288, + 604 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 572, + 288, + 604 + ], + "spans": [ + { + "bbox": [ + 48, + 572, + 288, + 604 + ], + "type": "text", + "content": "[11] Finale Doshi-Velez and Been Kim. Towards a rigorous science of interpretable machine learning. ArXiv e-print, 2017. 1, 3, 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 606, + 288, + 638 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 606, + 288, + 638 + ], + "spans": [ + { + "bbox": [ + 48, + 606, + 288, + 638 + ], + "type": "text", + "content": "[12] Carl Eckart and Gale Young. The approximation of one matrix by another of lower rank. Psychometrika, 1(3):211-218, 1936. 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 639, + 288, + 706 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 639, + 288, + 706 + ], + "spans": [ + { + "bbox": [ + 48, + 639, + 288, + 706 + ], + "type": "text", + "content": "[13] Nelson Elhage, Tristan Hume, Catherine Olsson, Nicholas Schiefer, Tom Henighan, Shauna Kravec, Zac Hatfield-Dodds, Robert Lasenby, Dawn Drain, Carol Chen, Roger Grosse, Sam McCandlish, Jared Kaplan, Dario Amodei, Martin Wattenberg, and Christopher Olah. Toy models of superposition. Transformer Circuits Thread, 2022. 4" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 719 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 307, + 73, + 547, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 547, + 128 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 547, + 128 + ], + "type": "text", + "content": "[14] Thomas Fel, Rémi Cadène, Mathieu Chalvidal, Matthieu Cord, David Vigouroux, and Thomas Serre. Look at the variance! efficient black-box explanations with sobol-based sensitivity analysis. Advances in Neural Information Processing Systems, 34, 2021. 1, 3, 7, 19" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 129, + 547, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 129, + 547, + 183 + ], + "spans": [ + { + "bbox": [ + 307, + 129, + 547, + 183 + ], + "type": "text", + "content": "[15] Thomas Fel, Mélanie Ducoffe, David Vigouroux, Rémi Cadène, Mikael Capelle, Claire Nicodème, and Thomas Serre. Don't lie to me! robust and efficient explainability with verified perturbation analysis. arXiv preprint arXiv:2202.07728, 2022.3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 185, + 547, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 185, + 547, + 271 + ], + "spans": [ + { + "bbox": [ + 308, + 185, + 547, + 271 + ], + "type": "text", + "content": "[16] Thomas Fel, Lucas Hervier, David Vigouroux, Antonin Poche, Justin Plakoo, Remi Cadene, Mathieu Chalvidal, Julien Colin, Thibaut Boissin, Louis Béthune, Agustin Picard, Claire Nicodeme, Laurent Gardes, Gregory Flandin, and Thomas Serre. Xplique: A deep learning explainability toolbox. Workshop, Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 16" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 274, + 547, + 305 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 274, + 547, + 305 + ], + "spans": [ + { + "bbox": [ + 308, + 274, + 547, + 305 + ], + "type": "text", + "content": "[17] Cédric Févotte and Jérôme Idier. Algorithms for nonnegative matrix factorization with the " + }, + { + "bbox": [ + 308, + 274, + 547, + 305 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 308, + 274, + 547, + 305 + ], + "type": "text", + "content": "-divergence. Neural computation, 23(9):2421-2456, 2011. 6, 19" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 307, + 546, + 350 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 307, + 546, + 350 + ], + "spans": [ + { + "bbox": [ + 308, + 307, + 546, + 350 + ], + "type": "text", + "content": "[18] Ruth C Fong and Andrea Vedaldi. Interpretable explanations of black boxes by meaningful perturbation. In Proceedings of the IEEE international conference on computer vision, pages 3429-3437, 2017. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 353, + 547, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 353, + 547, + 396 + ], + "spans": [ + { + "bbox": [ + 308, + 353, + 547, + 396 + ], + "type": "text", + "content": "[19] Ruth C Fong and Andrea Vedaldi. Interpretable explanations of black boxes by meaningful perturbation. In Proceedings of the IEEE international conference on computer vision, pages 3429-3437, 2017. 13" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 397, + 547, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 397, + 547, + 441 + ], + "spans": [ + { + "bbox": [ + 308, + 397, + 547, + 441 + ], + "type": "text", + "content": "[20] Xiao Fu, Kejun Huang, Nicholas D Sidiropoulos, and Wing-Kin Ma. Nonnegative matrix factorization for signal and data analytics: Identifiability, algorithms, and applications. IEEE Signal Process. Mag., 36(2):59-80, 2019. 5" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 442, + 547, + 507 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 442, + 547, + 507 + ], + "spans": [ + { + "bbox": [ + 308, + 442, + 547, + 507 + ], + "type": "text", + "content": "[21] Yunhao Ge, Yao Xiao, Zhi Xu, Meng Zheng, Srikrishna Karanam, Terrence Chen, Laurent Itti, and Ziyan Wu. A peek into the reasoning of neural networks: Interpreting with structural visual concepts. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2195-2204, 2021. 13" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 509, + 547, + 552 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 509, + 547, + 552 + ], + "spans": [ + { + "bbox": [ + 308, + 509, + 547, + 552 + ], + "type": "text", + "content": "[22] Sahara Ghalebikesabi, Lucile Ter-Minassian, Karla DiazOrdaz, and Chris C Holmes. On locality of local explanation models. Advances in Neural Information Processing Systems (NeurIPS), 2021. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 554, + 547, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 554, + 547, + 586 + ], + "spans": [ + { + "bbox": [ + 308, + 554, + 547, + 586 + ], + "type": "text", + "content": "[23] Amirata Ghorbani, Abubakar Abid, and James Zou. Interpretation of neural networks is fragile. In Proceedings of the AAAI Conference on Artificial Intelligence (AAAI), 2017. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 587, + 547, + 630 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 587, + 547, + 630 + ], + "spans": [ + { + "bbox": [ + 308, + 587, + 547, + 630 + ], + "type": "text", + "content": "[24] Amirata Ghorbani, James Wexler, James Y Zou, and Been Kim. Towards automatic concept-based explanations. Advances in Neural Information Processing Systems, 32, 2019, 2, 3, 4, 5, 7, 8, 13, 14, 23" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 632, + 547, + 664 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 632, + 547, + 664 + ], + "spans": [ + { + "bbox": [ + 308, + 632, + 547, + 664 + ], + "type": "text", + "content": "[25] Andreas Griewank and Andrea Walther. Evaluating derivatives: principles and techniques of algorithmic differentiation. SIAM, 2008. 3, 6, 19" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 666, + 547, + 719 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 666, + 547, + 719 + ], + "spans": [ + { + "bbox": [ + 308, + 666, + 547, + 719 + ], + "type": "text", + "content": "[26] Peter Hase and Mohit Bansal. Evaluating explainable ai: Which algorithmic explanations help users predict model behavior? In Proceedings of the Annual Meeting of the Association for Computational Linguistics (ACL Short Papers), 2020. 2, 3" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 734, + 315, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 734, + 315, + 742 + ], + "spans": [ + { + "bbox": [ + 295, + 734, + 315, + 742 + ], + "type": "text", + "content": "2719" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 718 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 288, + 116 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 288, + 116 + ], + "type": "text", + "content": "[27] Peter Hase, Harry Xie, and Mohit Bansal. The out-of-distribution problem in explainability and search methods for feature importance explanations. Advances in Neural Information Processing Systems (NeurIPS), 2021. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 118, + 288, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 118, + 288, + 149 + ], + "spans": [ + { + "bbox": [ + 49, + 118, + 288, + 149 + ], + "type": "text", + "content": "[28] Johannes Haug, Stefan Zürn, Peter El-Jiz, and Gjergji Kasneci. On baselines for local feature attributions. arXiv preprint arXiv:2101.00905, 2021. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 150, + 288, + 192 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 150, + 288, + 192 + ], + "spans": [ + { + "bbox": [ + 49, + 150, + 288, + 192 + ], + "type": "text", + "content": "[29] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Identity mappings in deep residual networks. In European conference on computer vision, pages 630-645. Springer, 2016. 23" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 194, + 288, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 194, + 288, + 226 + ], + "spans": [ + { + "bbox": [ + 49, + 194, + 288, + 226 + ], + "type": "text", + "content": "[30] Magnus R Hestenes and Eduard Stiefel. Methods of conjugate gradients for solving. Journal of research of the National Bureau of Standards, 49(6):409, 1952. 18, 19" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 49, + 227, + 288, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 227, + 288, + 281 + ], + "spans": [ + { + "bbox": [ + 49, + 227, + 288, + 281 + ], + "type": "text", + "content": "[31] Cheng-Yu Hsieh, Chih-Kuan Yeh, Xuanqing Liu, Pradeep Ravikumar, Seungyeon Kim, Sanjiv Kumar, and Cho-Jui Hsieh. Evaluations and methods for explanation through robustness analysis. In Proceedings of the International Conference on Learning Representations (ICLR), 2021. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 49, + 282, + 288, + 325 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 282, + 288, + 325 + ], + "spans": [ + { + "bbox": [ + 49, + 282, + 288, + 325 + ], + "type": "text", + "content": "[32] Kejun Huang, Nicholas D Sidiropoulos, and Athanasios P Liavas. A flexible and efficient algorithmic framework for constrained matrix and tensor factorization. IEEE Transactions on Signal Processing, 64(19):5052-5065, 2016. 6, 18" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 49, + 326, + 288, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 326, + 288, + 369 + ], + "spans": [ + { + "bbox": [ + 49, + 326, + 288, + 369 + ], + "type": "text", + "content": "[33] Marouane Il Idrissi, Vincent Chabridon, and Bertrand Iooss. Developments and applications of shapley effects to reliability-oriented sensitivity analysis with correlated inputs. Environmental Modelling & Software, 2021. 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 49, + 369, + 288, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 369, + 288, + 413 + ], + "spans": [ + { + "bbox": [ + 49, + 369, + 288, + 413 + ], + "type": "text", + "content": "[34] Bertrand Iooss and Paul Lemaitre. A review on global sensitivity analysis methods. In Uncertainty management in simulation-optimization of complex systems, pages 101-122. Springer, 2015. 2, 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 49, + 414, + 288, + 467 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 414, + 288, + 467 + ], + "spans": [ + { + "bbox": [ + 49, + 414, + 288, + 467 + ], + "type": "text", + "content": "[35] Alon Jacovi, Ana Marasovic, Tim Miller, and Yoav Goldberg. Formalizing trust in artificial intelligence: Prerequisites, causes and goals of human trust in ai. In Proceedings of the 2021 ACM conference on fairness, accountability, and transparency, pages 624-635, 2021. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 49, + 468, + 288, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 468, + 288, + 511 + ], + "spans": [ + { + "bbox": [ + 49, + 468, + 288, + 511 + ], + "type": "text", + "content": "[36] Alexandre Janon, Thierry Klein, Agnes Lagnoux, Maëlle Nodet, and Clémentine Prieur. Asymptotic normality and efficiency of two sobol index estimators. *ESAIM: Probability and Statistics*, 18:342–364, 2014. 5, 20" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 49, + 512, + 288, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 512, + 288, + 533 + ], + "spans": [ + { + "bbox": [ + 49, + 512, + 288, + 533 + ], + "type": "text", + "content": "[37] Margot E Kaminski and Jennifer M Urban. The right to contest ai. Columbia Law Review, 121(7):1957-2048, 2021. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 49, + 534, + 288, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 534, + 288, + 567 + ], + "spans": [ + { + "bbox": [ + 49, + 534, + 288, + 567 + ], + "type": "text", + "content": "[38] William Karush. Minima of functions of several variables with inequalities as side constraints. M. Sc. Dissertation. Dept. of Mathematics, Univ. of Chicago, 1939. 6, 18" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 49, + 567, + 288, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 567, + 288, + 609 + ], + "spans": [ + { + "bbox": [ + 49, + 567, + 288, + 609 + ], + "type": "text", + "content": "[39] Been Kim, Rajiv Khanna, and Oluwasanmi O Koyejo. Examples are not enough, learn to criticize! criticism for interpretability. Advances in Neural Information Processing Systems (NeurIPS), 2016. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 49, + 611, + 288, + 675 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 611, + 288, + 675 + ], + "spans": [ + { + "bbox": [ + 49, + 611, + 288, + 675 + ], + "type": "text", + "content": "[40] Been Kim, Martin Wattenberg, Justin Gilmer, Carrie Cai, James Wexler, Fernanda Viegas, et al. Interpretability beyond feature attribution: Quantitative testing with concept activation vectors (tcav). In International conference on machine learning, pages 2668-2677. PMLR, 2018. 2, 3, 5, 8, 16, 23" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 49, + 677, + 288, + 718 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 677, + 288, + 718 + ], + "spans": [ + { + "bbox": [ + 49, + 677, + 288, + 718 + ], + "type": "text", + "content": "[41] Sunnie S. Y. Kim, Nicole Meister, Vikram V. Ramaswamy, Ruth Fong, and Olga Russakovsky. HIVE: Evaluating the human interpretability of visual explanations. In European Conference on Computer Vision (ECCV), 2022. 2, 3" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 700 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 307, + 73, + 547, + 126 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 547, + 126 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 547, + 126 + ], + "type": "text", + "content": "[42] Pieter-Jan Kindermans, Sara Hooker, Julius Adebayo, Maximilian Alber, Kristof T. Schütt, Sven Dähne, Dumitru Erhan, and Been Kim. The (Un)reliability of Saliency Methods, pages 267-280. Springer International Publishing, Cham, 2019. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 129, + 547, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 129, + 547, + 215 + ], + "spans": [ + { + "bbox": [ + 307, + 129, + 547, + 215 + ], + "type": "text", + "content": "[43] Mauritz Kop. Eu artificial intelligence act: The European approach to ai. In Stanford - Vienna Transatlantic Technology Law Forum, Transatlantic Antitrust and IPR Developments, Stanford University, Issue No. 2/2021. https://law.stanford.edu/publications/eu-artificial-intelligence-act-the-european-approach-to-ai/. Stanford-Vienna Transatlantic Technology Law Forum, Transatlantic Antitrust, 2021. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 218, + 547, + 250 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 218, + 547, + 250 + ], + "spans": [ + { + "bbox": [ + 307, + 218, + 547, + 250 + ], + "type": "text", + "content": "[44] Steven George Krantz and Harold R Parks. The implicit function theorem: history, theory, and applications. Springer Science & Business Media, 2002. 3, 6, 19" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 251, + 547, + 294 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 251, + 547, + 294 + ], + "spans": [ + { + "bbox": [ + 307, + 251, + 547, + 294 + ], + "type": "text", + "content": "[45] Harold W Kuhn and Albert W Tucker. Nonlinear programming proceedings of the second berkeley symposium on mathematical statistics and probability. Neyman, pages 481-492, 1951. 6, 18" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 297, + 547, + 328 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 297, + 547, + 328 + ], + "spans": [ + { + "bbox": [ + 307, + 297, + 547, + 328 + ], + "type": "text", + "content": "[46] Daniel D Lee and H Sebastian Seung. Learning the parts of objects by non-negative matrix factorization. Nature, 401(6755):788-791, 1999. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 331, + 547, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 331, + 547, + 363 + ], + "spans": [ + { + "bbox": [ + 307, + 331, + 547, + 363 + ], + "type": "text", + "content": "[47] Scott M Lundberg and Su-In Lee. A unified approach to interpreting model predictions. Advances in neural information processing systems, 30, 2017. 3, 7" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 365, + 547, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 365, + 547, + 407 + ], + "spans": [ + { + "bbox": [ + 307, + 365, + 547, + 407 + ], + "type": "text", + "content": "[48] Amandine Marrel, Bertrand Iooss, Beatrice Laurent, and Olivier Roustant. Calculations of sobol indices for the gaussian process metamodel. Reliability Engineering & System Safety, 94(3):742-751, 2009. 5" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 409, + 547, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 409, + 547, + 453 + ], + "spans": [ + { + "bbox": [ + 307, + 409, + 547, + 453 + ], + "type": "text", + "content": "[49] Giang Nguyen, Daeyoung Kim, and Anh Nguyen. The effectiveness of feature attribution methods and its correlation with automatic evaluation scores. Advances in Neural Information Processing Systems (NeurIPS), 2021. 2, 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 454, + 547, + 497 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 454, + 547, + 497 + ], + "spans": [ + { + "bbox": [ + 307, + 454, + 547, + 497 + ], + "type": "text", + "content": "[50] Paul Novello, Thomas Fel, and David Vigouroux. Making sense of dependence: Efficient black-box explanations using dependence measure. In Advances in Neural Information Processing Systems (NeurIPS), 2022. 1, 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 498, + 547, + 531 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 498, + 547, + 531 + ], + "spans": [ + { + "bbox": [ + 307, + 498, + 547, + 531 + ], + "type": "text", + "content": "[51] Chris Olah, Alexander Mordvintsev, and Ludwig Schubert. Feature visualization. Distill, 2017. https://distill.pub/2017/feature-visualization. 13, 16" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 533, + 547, + 564 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 533, + 547, + 564 + ], + "spans": [ + { + "bbox": [ + 307, + 533, + 547, + 564 + ], + "type": "text", + "content": "[52] Art B Owen. Better estimation of small sobol'sensitivity indices. ACM Transactions on Modeling and Computer Simulation (TOMACS), 23(2):1-17, 2013. 5" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 567, + 547, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 567, + 547, + 609 + ], + "spans": [ + { + "bbox": [ + 307, + 567, + 547, + 609 + ], + "type": "text", + "content": "[53] Vardan Papyan, XY Han, and David L Donoho. Prevalence of neural collapse during the terminal phase of deep learning training. Proceedings of the National Academy of Sciences, 117(40):24652-24663, 2020. 5" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 611, + 547, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 611, + 547, + 666 + ], + "spans": [ + { + "bbox": [ + 307, + 611, + 547, + 666 + ], + "type": "text", + "content": "[54] Fabian Pedregosa, Gael Varoquaux, Alexandre Gramfort, Vincent Michel, Bertrand Thirion, Olivier Grisel, Mathieu Blondel, Peter Prettenhofer, Ron Weiss, Vincent Dubourg, et al. Scikit-learn: Machine learning in python. Journal of machine learning research, 12(Oct):2825–2830, 2011. 6" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 667, + 547, + 700 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 667, + 547, + 700 + ], + "spans": [ + { + "bbox": [ + 307, + 667, + 547, + 700 + ], + "type": "text", + "content": "[55] Vitali Petsiuk, Abir Das, and Kate Saenko. Rise: Randomized input sampling for explanation of black-box models. arXiv preprint arXiv:1806.07421, 2018. 1, 3, 7, 19, 23" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 734, + 315, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 734, + 315, + 742 + ], + "spans": [ + { + "bbox": [ + 295, + 734, + 315, + 742 + ], + "type": "text", + "content": "2720" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 619 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 288, + 127 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 288, + 127 + ], + "type": "text", + "content": "[56] Bin Ren, Laurent Pueyo, Christine Chen, Élodie Choquet, John H Debes, Gaspard Duchêne, François Ménard, and Marshall D Perrin. Using data imputation for signal separation in high-contrast imaging. The Astrophysical Journal, 892(2):74, 2020. 4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 129, + 288, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 129, + 288, + 184 + ], + "spans": [ + { + "bbox": [ + 48, + 129, + 288, + 184 + ], + "type": "text", + "content": "[57] Marco Tulio Ribeiro, Sameer Singh, and Carlos Guestrin. \" why should i trust you?\" explaining the predictions of any classifier. In Proceedings of the 22nd ACM SIGKDD international conference on knowledge discovery and data mining, pages 1135-1144, 2016. 3, 7, 19" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 185, + 288, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 185, + 288, + 239 + ], + "spans": [ + { + "bbox": [ + 48, + 185, + 288, + 239 + ], + "type": "text", + "content": "[58] Andrea Saltelli, Paola Annoni, Ivano Azzini, Francesca Campolongo, Marco Ratto, and Stefano Tarantola. Variance based sensitivity analysis of model output. design and estimator for the total sensitivity index. Computer physics communications, 181(2):259-270, 2010. 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 240, + 288, + 306 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 240, + 288, + 306 + ], + "spans": [ + { + "bbox": [ + 48, + 240, + 288, + 306 + ], + "type": "text", + "content": "[59] Ramprasaath R Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra. Grad-cam: Visual explanations from deep networks via gradient-based localization. In Proceedings of the IEEE international conference on computer vision, pages 618-626, 2017. 1, 3, 6, 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 307, + 288, + 361 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 307, + 288, + 361 + ], + "spans": [ + { + "bbox": [ + 48, + 307, + 288, + 361 + ], + "type": "text", + "content": "[60] Hua Shen and Ting-Hao Huang. How useful are the machine-generated interpretations to general users? a human evaluation on guessing the incorrectly predicted labels. In Proceedings of the AAAI Conference on Human Computation and Crowdsourcing, volume 8, pages 168-172, 2020. 2, 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 363, + 288, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 363, + 288, + 407 + ], + "spans": [ + { + "bbox": [ + 48, + 363, + 288, + 407 + ], + "type": "text", + "content": "[61] Avanti Shrikumar, Peyton Greenside, and Anshul Kundaje. Learning important features through propagating activation differences. In Proceedings of the International Conference on Machine Learning (ICML), 2017. 1, 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 408, + 288, + 462 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 408, + 288, + 462 + ], + "spans": [ + { + "bbox": [ + 48, + 408, + 288, + 462 + ], + "type": "text", + "content": "[62] Karen Simonyan, Andrea Vedaldi, and Andrew Zisserman. Deep inside convolutional networks: Visualising image classification models and saliency maps. In In Workshop at International Conference on Learning Representations. CiteSeer, 2014. 1, 3, 6, 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 463, + 288, + 508 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 463, + 288, + 508 + ], + "spans": [ + { + "bbox": [ + 48, + 463, + 288, + 508 + ], + "type": "text", + "content": "[63] Leon Sixt, Maximilian Granz, and Tim Landgraf. When explanations lie: Why many modified bp attributions fail. In Proceedings of the International Conference on Machine Learning (ICML), 2020. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 509, + 288, + 563 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 509, + 288, + 563 + ], + "spans": [ + { + "bbox": [ + 48, + 509, + 288, + 563 + ], + "type": "text", + "content": "[64] Leon Sixt, Martin Schuessler, Oana-Iuliana Popescu, Philipp Weiß, and Tim Landgraf. Do users benefit from interpretable vision? a user study, baseline, and dataset. Proceedings of the International Conference on Learning Representations (ICLR), 2022. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 564, + 288, + 619 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 564, + 288, + 619 + ], + "spans": [ + { + "bbox": [ + 48, + 564, + 288, + 619 + ], + "type": "text", + "content": "[65] Dylan Slack, Sophie Hilgard, Emily Jia, Sameer Singh, and Himabindu Lakkaraju. Fooling lime and shap: Adversarial attacks on post hoc explanation methods. In Proceedings of the AAAI/ACM Conference on AI, Ethics, and Society, pages 180-186, 2020. 2, 3" + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 620 + ], + "type": "list", + "angle": 0, + "index": 24, + "blocks": [ + { + "bbox": [ + 307, + 73, + 547, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 547, + 116 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 547, + 116 + ], + "type": "text", + "content": "[66] Daniel Smilkov, Nikhil Thorat, Been Kim, Fernanda Viégas, and Martin Wattenberg. Smoothgrad: removing noise by adding noise. arXiv preprint arXiv:1706.03825, 2017. 1, 3, 5, 6, 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 307, + 118, + 547, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 547, + 150 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 547, + 150 + ], + "type": "text", + "content": "[67] Ilya M Sobol. Sensitivity analysis for non-linear mathematical models. Mathematical modelling and computational experiment, 1:407-414, 1993. 2, 5, 20" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 307, + 150, + 547, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 150, + 547, + 183 + ], + "spans": [ + { + "bbox": [ + 307, + 150, + 547, + 183 + ], + "type": "text", + "content": "[68] Ilya M Sobol. Global sensitivity indices for nonlinear mathematical models and their monte carlo estimates. Mathematics and computers in simulation, 55(1-3):271-280, 2001. 2, 5" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 308, + 184, + 547, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 184, + 547, + 228 + ], + "spans": [ + { + "bbox": [ + 308, + 184, + 547, + 228 + ], + "type": "text", + "content": "[69] Jost Tobias Springenberg, Alexey Dosovitskiy, Thomas Brox, and Martin Riedmiller. Striving for simplicity: The all convolutional net. arXiv preprint arXiv:1412.6806, 2014. 1, 3, 6" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 308, + 228, + 547, + 260 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 228, + 547, + 260 + ], + "spans": [ + { + "bbox": [ + 308, + 228, + 547, + 260 + ], + "type": "text", + "content": "[70] Pascal Sturmfels, Scott Lundberg, and Su-In Lee. Visualizing the impact of feature attribution baselines. Distill, 2020. 3, 13" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 262, + 547, + 306 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 262, + 547, + 306 + ], + "spans": [ + { + "bbox": [ + 308, + 262, + 547, + 306 + ], + "type": "text", + "content": "[71] Mukund Sundararajan, Ankur Taly, and Qiqi Yan. Axiomatic attribution for deep networks. In International conference on machine learning, pages 3319-3328. PMLR, 2017. 1, 3, 5, 6, 7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 307, + 547, + 351 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 307, + 547, + 351 + ], + "spans": [ + { + "bbox": [ + 308, + 307, + 547, + 351 + ], + "type": "text", + "content": "[72] Stefano Tarantola, Debora Gatelli, and Thierry Alex Mara. Random balance designs for the estimation of first order global sensitivity indices. Reliability Engineering & System Safety, 91(6):717-727, 2006. 5" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 352, + 547, + 384 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 352, + 547, + 384 + ], + "spans": [ + { + "bbox": [ + 308, + 352, + 547, + 384 + ], + "type": "text", + "content": "[73] Stephen A Vavasis. On the complexity of nonnegative matrix factorization. SIAM Journal on Optimization, 20(3):1364-1377, 2010. 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 386, + 547, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 386, + 547, + 430 + ], + "spans": [ + { + "bbox": [ + 308, + 386, + 547, + 430 + ], + "type": "text", + "content": "[74] Yu-Xiong Wang and Yu-Jin Zhang. Nonnegative matrix factorization: A comprehensive review. IEEE Transactions on Knowledge and Data Engineering, 25(6):1336-1353, 2013. 5" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 431, + 547, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 431, + 547, + 475 + ], + "spans": [ + { + "bbox": [ + 308, + 431, + 547, + 475 + ], + "type": "text", + "content": "[75] Peter Wilf, Shengping Zhang, Sharat Chikkerur, Stefan A Little, Scott L Wing, and Thomas Serre. Computer vision cracks the leaf code. Proceedings of the National Academy of Sciences, 113(12):3305-3310, 2016. 7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 475, + 547, + 509 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 475, + 547, + 509 + ], + "spans": [ + { + "bbox": [ + 308, + 475, + 547, + 509 + ], + "type": "text", + "content": "[76] Matthew D Zeiler and Rob Fergus. Visualizing and understanding convolutional networks. In European conference on computer vision, pages 818-833. Springer, 2014. 1, 3, 7" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 510, + 547, + 563 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 510, + 547, + 563 + ], + "spans": [ + { + "bbox": [ + 308, + 510, + 547, + 563 + ], + "type": "text", + "content": "[77] Ruihan Zhang, Prashan Madumal, Tim Miller, Krista A Ehinger, and Benjamin IP Rubinstein. Invertible concept-based explanations for cnn models with non-negative concept activation vectors. arXiv preprint arXiv:2006.15417, 2020. 2, 3, 5, 8, 23" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 564, + 547, + 620 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 564, + 547, + 620 + ], + "spans": [ + { + "bbox": [ + 308, + 564, + 547, + 620 + ], + "type": "text", + "content": "[78] Bolei Zhou, Aditya Khosla, Agata Lapedriza, Aude Oliva, and Antonio Torralba. Learning deep features for discriminative localization. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2921-2929, 2016. 3, 6" + } + ] + } + ], + "index": 23 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 734, + 314, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 734, + 314, + 742 + ], + "spans": [ + { + "bbox": [ + 295, + 734, + 314, + 742 + ], + "type": "text", + "content": "2721" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/CREPE_ Can Vision-Language Foundation Models Reason Compositionally_/6b09bf56-aeff-4b97-8111-a915e425b709_content_list.json b/2023/CREPE_ Can Vision-Language Foundation Models Reason Compositionally_/6b09bf56-aeff-4b97-8111-a915e425b709_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..10e819e56453394d5cc0e8c7d9eb028aad89d316 --- /dev/null +++ b/2023/CREPE_ Can Vision-Language Foundation Models Reason Compositionally_/6b09bf56-aeff-4b97-8111-a915e425b709_content_list.json @@ -0,0 +1,1372 @@ +[ + { + "type": "image", + "img_path": "images/a9c996fb36116783728bd08036147b3e393278ffa081301e60aa8db52c0df423.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 84, + 126, + 107, + 147 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "CREPE: Can Vision-Language Foundation Models Reason Compositionally?", + "text_level": 1, + "bbox": [ + 109, + 128, + 883, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Zixian Ma $^{1*}$ , Jerry Hong $^{1*}$ , Mustafa Omer Gul $^{2*}$ , Mona Gandhi $^{3}$ , Irenea Gao $^{1}$ , Ranjay Krishna $^{4}$ , Stanford University $^{1}$ , Cornell University $^{2}$ , University of Pennsylvania $^{3}$ , University of Washington $^{4}$", + "bbox": [ + 94, + 179, + 872, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{zixianma, jerryhong, irena}@cs.stanford.edu mog29@cornell.edu mona09@seas.upenn.edu ranjay@cs.washington.edu", + "bbox": [ + 109, + 219, + 851, + 251 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 286, + 313, + 301 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "A fundamental characteristic common to both human vision and natural language is their compositional nature. Yet, despite the performance gains contributed by large vision and language pretraining, we find that—across 7 architectures trained with 4 algorithms on massive datasets—they struggle at compositionality. To arrive at this conclusion, we introduce a new compositionality evaluation benchmark, CREPE, which measures two important aspects of compositionality identified by cognitive science literature: systematicity and productivity. To measure systematicity, CREPE consists of a test dataset containing over $370K$ image-text pairs and three different seen-unseen splits. The three splits are designed to test models trained on three popular training datasets: CC-12M, YFCC-15M, and LAION-400M. We also generate $325K$ , $316K$ , and $309K$ hard negative captions for a subset of the pairs. To test productivity, CREPE contains $17K$ image-text pairs with nine different complexities plus $278K$ hard negative captions with atomic, swapping and negation foils. The datasets are generated by repurposing the Visual Genome scene graphs and region descriptions and applying handcrafted templates and GPT-3. For systematicity, we find that model performance decreases consistently when novel compositions dominate the retrieval set, with Recall@1 dropping by up to $9\\%$ . For productivity, models' retrieval success decays as complexity increases, frequently nearing random chance at high complexity. These results hold regardless of model and training dataset size.", + "bbox": [ + 75, + 318, + 473, + 728 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 757, + 209, + 773 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Compositionality, the understanding that \"the meaning of the whole is a function of the meanings of its parts\" [11], is held to be a key characteristic of human intelligence. In language, the whole is a sentence, made up of words. In vision, the whole is a scene, made up of parts like objects, their attributes, and their relationships [31, 35].", + "bbox": [ + 75, + 784, + 470, + 875 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/642384e546156cd84a8f39013a8be14bd43d1f63dadd53c9dd7323fe5ade1c91.jpg", + "image_caption": [ + "Figure 1. We introduce CREPE, a benchmark to evaluate whether vision-language foundation models demonstrate two fundamental aspects of compositionality: systematicity and productivity. To evaluate systematicity, CREPE utilizes Visual Genome and introduces three new test datasets for the three popular pretraining datasets: CC-12M, YFCC-15M, and LAION-400M. These enable evaluating models' abilities to systematically generalize their understanding to seen compounds, unseen compounds, and even unseen atoms. To evaluate productivity, CREPE introduces examples of nine complexities, with three types of hard negatives for each." + ], + "image_footnote": [], + "bbox": [ + 519, + 287, + 687, + 465 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/264f24940f3c1e1214081c6bd61d27dcfc06cd0d89c6209edb018b689942bb80.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 705, + 287, + 875, + 465 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Through compositional reasoning, humans can understand new scenes and generate complex sentences by combining known parts [6, 27, 30]. Despite compositionality's importance, there are no large-scale benchmarks directly evaluating whether vision-language models can reason compositionally. These models are pretrained using large-scale image-capture datasets [62, 64, 74], and are already widely applied for tasks that benefit from compositional reasoning, including retrieval, text-to-image generation, and open-vocabulary classification [10, 57, 60]. Especially as such models become ubiquitous \"foundations\" for other models [5], it is critical to understand their compositional abilities.", + "bbox": [ + 496, + 656, + 893, + 837 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Previous work has evaluated these models using image-text retrieval [32,56,82]. However, the retrieval datasets used either do not provide controlled sets of negatives [45, 74] or study narrow negatives which vary along a single axis", + "bbox": [ + 496, + 839, + 893, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Equal contribution", + "bbox": [ + 96, + 886, + 202, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "10910", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/78892989831eda5ec895483bb36d684450ec83896c6acbcf645c83f505bc112e.jpg", + "image_caption": [ + "Figure 2. An overview of the systematicity retrieval set generation process. First, a model's image-caption training set is parsed to identify what atoms and compounds the model has seen. Then, an evaluation set is divided into three compositional splits according to whether the model has seen all the compounds (Seen Compounds), only all the atoms of the caption (Unseen Compounds), or neither (Unseen Atoms). Finally, hard negative captions HN-ATOM and HN-COMP are generated for the hard negatives retrieval set $\\mathcal{D}_{test}^{HN}$ ." + ], + "image_footnote": [], + "bbox": [ + 89, + 85, + 861, + 282 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "(e.g. permuted word orders or single word substitutions as negative captions) [21, 51, 65, 75]. Further, these analyses have also not studied how retrieval performance varies when generalizing to unseen compositional combinations, or to combinations of increased complexity.", + "bbox": [ + 75, + 378, + 470, + 455 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We introduce CREPE (Compositional REpresentation Evaluation): a new large-scale benchmark to evaluate two aspects of compositionality: systematicity and productivity (Figure 1). Systematicity measures how well a model is able to represent seen versus unseen atoms and their compositions. Productivity studies how well a model can comprehend an unbounded set of increasingly complex expressions. CREPE uses Visual Genome's scene graph representation as the compositionality language [35] and constructs evaluation datasets using its annotations. To test systematicity, we parse the captions in three popular training datasets, CC-12M [8], YFCC-15M [74], and LAION-400M [62], to identify atoms (objects, relations, or attributes) and compounds (combinations of atoms) present in each dataset. For each training set, we curate corresponding test sets containing $385K$ , $385K$ and $373K$ image-text pairs respectively, with splits checking generalization to seen compounds, unseen compounds, and unseen atoms. To test productivity, CREPE contains $17K$ image-text pairs split across nine levels of complexity, as defined by the number of atoms present in the text. Examples across all datasets are paired with various hard negative types to ensure the legitimacy of our conclusions.", + "bbox": [ + 75, + 455, + 473, + 792 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our experiments—across 7 architectures trained with 4 training algorithms on massive datasets—find that vision-language models struggle at compositionality, with both systematicity and productivity. We present six key findings: first, our systematicity experiments find that models' performance consistently drops between seen and unseen compositions; second, we observe larger drops for models trained on", + "bbox": [ + 75, + 795, + 470, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "LAION-400M (up to a $9\\%$ decrease in Recall@1); third, our productivity experiments indicate that retrieval performance degrades with increased caption complexity; fourth, we find no clear trend relating training dataset size to models' compositional reasoning; fifth, model size also has no impact; finally, models' zero-shot ImageNet classification accuracy correlates only with their absolute retrieval performance on the systematicity dataset but not systematic generalization to unseen compounds or to productivity.", + "bbox": [ + 496, + 378, + 893, + 513 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 531, + 640, + 547 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our work lies within the field of evaluating foundation models. Specifically, we measure visio-linguistic compositionality. To do so, we create a retrieval benchmark with hard negatives.", + "bbox": [ + 496, + 558, + 893, + 618 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Contrastive Image-Text Pretraining. The recently released contrastively trained CLIP model [56] has catalyzed a wide array of work at the intersection of Computer Vision and Natural Language Processing. Since its release, CLIP has enabled several tasks, ranging from semantic segmentation to image captioning, many of which have remarkable zero-shot capability [12, 16, 38, 56, 71, 73]. CLIP has been used as a loss function within image synthesis applications [29, 44, 46, 54, 79, 83], acted as an automated evaluation metric [22, 52], used successfully as a feature extractor for various vision and language tasks [66], and incorporated into architectures for various tasks including dense prediction and video summarization [43, 50, 55, 58, 67, 68]. This success has also encouraged the design of other contrastive vision and language pretraining algorithms for image [15, 18, 40–42, 48, 69, 80, 81] and video domains [39, 76, 78]. Our work evaluates how well", + "bbox": [ + 496, + 619, + 895, + 862 + ], + "page_idx": 1 + }, + { + "type": "page_footnote", + "text": "1We release our datasets, and code to generate and evaluate on our test sets at https://github.com/RAIVNLab/CREPE.", + "bbox": [ + 500, + 875, + 893, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "10911", + "bbox": [ + 480, + 944, + 517, + 957 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "such contrastively trained models capture a fundamental property present in human vision and language: compositionality.", + "bbox": [ + 76, + 90, + 470, + 136 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Compositionality. Compositionality allows us to comprehend an infinite number of scenes and utterances [37]. For an AI model, compositionality would not only allow for systematic, combinatorial generalization, but would also confer benefits such as controllability [5]. This promise prompted a wealth of work on both designing [2, 23, 25] and evaluating [17, 19, 26, 36, 72] compositional models. In our work, we focus on two aspects of compositionality: systematicity and productivity. While there is a plethora of benchmarks for systematic generalization within Computer Vision [3, 4, 19, 33] and Machine Learning [34, 36, 59], the subject has been almost unexplored for vision-language models, largely due to lack of benchmarks complementary to the different largescale training datasets. To address this, CREPE provides a benchmark with three different datasets to evaluate the compositional generalization of vision-language models. Productivity, on the other hand, has been studied only for specialized tasks [19] or toy domains [27, 36, 59]. CREPE evaluates productivity by using an image-text retrieval task featuring captions of varying compositional complexity.", + "bbox": [ + 76, + 141, + 470, + 444 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Evaluation with hard negatives. Like us, past work evaluating models has commonly designed tasks featuring hard negatives to isolate particular model capabilities while overcoming the limitations of prior evaluation tasks. Using atomic foils that replace an atom in the image or text with a distractor has been the most common strategy [4,9,21,24,51,53,65]. Notably, Park et al. [53] targets verbs and person entities in videos; COVR [4] studies question answering with distractor images; FALSE [51] targets linguistic phenomena such as existence, cardinality and the recognition of actions and spatial relationships. Another strategy has been to swap atoms within a caption to test whether models behave akin to a bag-of-words [1, 51, 75]. In particular, Winoground [75] introduces a set of 800 human edited negatives to evaluate compositionality; it is the closest related work to us. We complement Winoground by scaling it up by three orders of magnitude, by decomposing compositionality into systematicity and productivity, and by studying a variety of different types of hard negatives.", + "bbox": [ + 76, + 448, + 472, + 734 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Compositional evaluation", + "text_level": 1, + "bbox": [ + 76, + 763, + 313, + 781 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The following section builds from the formally vacuous principle of compositionality to a well-defined evaluation scheme [27]. First, we establish the syntax and semantics of the composed language (Section 3.1). Then, we define expected behaviors from a model that achieves comprehension of said language (3.2, 3.3). Finally, we establish how to empirically measure those behaviors via retrieval (3.4).", + "bbox": [ + 76, + 794, + 470, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Compositional language of visual concepts", + "text_level": 1, + "bbox": [ + 500, + 90, + 861, + 107 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To evaluate vision-language models, we find that a compositional language consisting of scene graph visual concepts to be an appropriate foundation [35]. Accordingly, an atom $A$ is defined as a singular visual concept, corresponding to a single scene graph node. Atoms are subtyped into objects $A_{o}$ , relationships $A_{r}$ , and attributes $A_{a}$ . A compound $C$ is defined as a primitive composition of multiple atoms, which corresponds to connections between scene graph nodes. Visual concepts admit two compound types: the attachment of attribute to objects (\"black dog\") $C_{ao}$ , and the attachment of two objects via a relationship (\"man hugs child\") $C_{oro}$ .", + "bbox": [ + 496, + 113, + 893, + 280 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The composition of these compounds form subgraphs $S$ , which can be translated to natural language captions $T$ . Conversely, captions $T$ derived from image-text datasets $\\mathcal{D}$ can be parsed to become scene graphs $S$ . This extensible language is capable of capturing a number of linguistic phenomena identified in existing literature [51, 72], including the existence of concepts (\"a photo with flowers\"), spatial relationships (\"a grill on the left of a staircase\"), action relationships (\"a person throwing a frisbee\"), prepositional attachment (\"A bird with green wings\"), and negation (\"There are no trucks on the road\"). Furthermore, while this study focuses on visual concepts, scene graphs featuring common-sense relationships or other more abstract concepts can be designed; therefore, our methodology is widely applicable [61].", + "bbox": [ + 496, + 280, + 893, + 491 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Systematicity", + "text_level": 1, + "bbox": [ + 500, + 500, + 638, + 515 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "With our compositional language in place, we now define two dimensions of compositionality—systematicity and productivity—which we adapt to vision-language representations. Systematicity evaluates a model's ability to systematically recombine seen atoms in compounds. Concretely, let $\\mathrm{SEEN}(A,D)$ denote if an atom is seen in a training dataset $\\mathcal{D}$ , namely $\\exists (I,S) \\in \\mathcal{D}: A \\in S$ , and $\\mathrm{SEEN}(C,D)$ denote if a compound is seen in a dataset $\\mathcal{D}$ , namely $\\exists (I,S) \\in \\mathcal{D}: C \\subseteq S$ . To evaluate systematicity, we define three compositional splits: Seen Compounds (SC), Unseen Compounds (UC) and Unseen Atoms (UA). SC is the split where all compounds (and thus all atoms) of every caption have been seen in the training dataset, i.e. $\\mathcal{D}_{\\mathrm{SC}} = \\{(I,S) \\in \\mathcal{D}_{test} | \\forall C \\subseteq S : \\mathrm{SEEN}(C,\\mathcal{D}_{train})\\}$ . UC is the split where, for each caption, all atoms have been seen but at least one compound has NOT, i.e. $\\mathcal{D}_{UC} = \\{(I,S) \\in \\mathcal{D}_{test} | (\\forall A \\in S : \\mathrm{SEEN}(A,\\mathcal{D}_{train}) \\land (\\exists C \\subseteq S : \\neg \\mathrm{Seen}(C,\\mathcal{D}_{train}))\\}$ . UA is the split where each caption contains at least one atom that has NOT been seen, i.e. $\\mathcal{D}_{UA} = \\{(I,S) \\in \\mathcal{D}_{test} | \\exists A \\in S : \\neg \\mathrm{SEEN}(A,\\mathcal{D}_{train})\\}$ .", + "bbox": [ + 496, + 522, + 893, + 825 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.3. Productivity", + "text_level": 1, + "bbox": [ + 500, + 832, + 633, + 848 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Productivity refers to a capacity to comprehend an unbounded set of expressions. Since the set of atoms in any dataset is finite, a reasonable substitute for testing unbounded", + "bbox": [ + 496, + 854, + 893, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "10912", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/c58dde3537b407e65e2764fcc4f7e9e0d34d1cd97d4f5d4f67c359b050929c7c.jpg", + "image_caption": [ + "Figure 3. An overview of the productivity retrieval set generation process. By performing random walks on the scene graphs of an evaluation dataset, we generate subgraphs of various complexities. Then, for complexities $n \\in \\{4, .5, .12\\}$ and three hard negative types, we populate the retrieval set $\\mathcal{D}_{test}^{HN}$ by generating a ground truth caption for each $n$ -subgraph and hard negatives for each caption." + ], + "image_footnote": [], + "bbox": [ + 89, + 85, + 877, + 265 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/242c2e1577768db32adbccecf3d2fe3b90986baa585a81b12dded08ba4e1c8ac.jpg", + "table_caption": [ + "Table 1. We summarize the sizes of the eight evaluation datasets we create for systematicity and productivity evaluation." + ], + "table_footnote": [], + "table_body": "
SystematicityProductivity
\\(\\mathcal{D}_{test}^{RAW}\\) (# of image-text pairs)\\(\\mathcal{D}_{test}^{HN}\\) (# of texts)\\(\\mathcal{D}_{test}^{RAW}\\)\\(\\mathcal{D}_{test}^{HN}\\)
Training dataCC-12MYFCC-15MLAION-400MCC-12MYFCC-15MLAION-400MAnyAny
Dataset size385,777385,777373,703325,523316,668309,34217,553278,730
", + "bbox": [ + 89, + 359, + 890, + 435 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "comprehension is testing comprehension over increasingly complex scenes. Now, an image $I$ does not have a notion of complexity, since it is theoretically infinitely describable; on the other hand, we can define a notion of complexity for a caption $T$ : the number of atoms in its corresponding scene graph $|S_T|$ . Therefore, a productive vision-language model should be able to match a given image to the correct corresponding caption, regardless of that caption's complexity. To evaluate productivity, we define a range of productivity complexity (in our case, $n = 4, 5, \\ldots, 12$ ). We need splits of the evaluation dataset based on these complexities, where image-text pairs in a given split have a fixed complexity $n$ , and evaluate a model's performance over each split.", + "bbox": [ + 75, + 462, + 473, + 657 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.4. Compositional evaluation via retrieval", + "text_level": 1, + "bbox": [ + 76, + 666, + 406, + 681 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We evaluate compositional reasoning using zero-shot image-to-text and text-to-image retrieval. This formulation probes the representation space as directly as possible and is already the most common evaluation method for vision-language foundation models [56]. Theoretically, any existing image-text dataset can be used as retrieval sets for our evaluation. However, one challenging limitation in existing datasets renders the metrics evaluated on them inaccurate. Consider using an image query of a \"plant inside a yellow vase on top of a black television.\" Retrieving unintended alternative positives (e.g. \"a black television\") is not necessarily incor", + "bbox": [ + 75, + 689, + 470, + 854 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "rect. Similarly, if no other texts in the retrieval set contain a \"plant\" and a \"television\", retrieving the correct text doesn't suggest that the model comprehends the image. Ideally, to properly evaluate a model, the retrieval dataset should contain hard negatives for every query. A hard negative is a caption that does not faithfully represent the corresponding image, and differs from the ground truth caption by some minimal atomic shift. A example hard negative for the query above is \"man inside a yellow vase on top of a black television.\" By erring in a single, granular syntactic or semantic fashion, hard negatives allow for variations in retrieval performance to be attributable to a specific failure mode of a model's compositional comprehension (see Appendix). We address this need for a new benchmark dataset to evaluate the systematicity and productivity of vision-language models.", + "bbox": [ + 496, + 462, + 893, + 689 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. CREPE: a large-scale benchmark for vision-language compositionality", + "text_level": 1, + "bbox": [ + 498, + 703, + 890, + 741 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "There are several challenges to creating image-text retrieval datasets that evaluate compositional systematicity and productivity. For systematicity, the primary challenge lies in parsing the training dataset for seen atoms and compounds in order to split the data into the three compositional splits. For productivity, the major challenge is generating imagedataset pairs across different text complexities for the retrieval sets. For both datasets, it is crucial to enumerate different types of hard negatives, and to design an automated hard negative generator which ensures the incorrectness of the", + "bbox": [ + 496, + 750, + 893, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_footnote", + "text": "2By avoiding captions with redundant objects (\"... a lamb and a lamb and...\" and abstract modifiers (\"there are many lampposts\"), we ensure atom count is tightly coupled with caption complexity.", + "bbox": [ + 75, + 862, + 470, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "10913", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "negatives it generates. We detail our methods for tackling these challenges for future efforts that attempt to create similar benchmarks for other training datasets.", + "bbox": [ + 76, + 90, + 470, + 137 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Creating systematicity datasets", + "text_level": 1, + "bbox": [ + 76, + 155, + 352, + 171 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To create the three systematicity splits—SC, SA, UA—we parse a given training dataset $\\mathcal{D}$ into its constituent atoms and compounds, filter low-quality data, and generate hard negatives (Figure 2).", + "bbox": [ + 76, + 181, + 468, + 242 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Parsing a dataset into atoms and compounds Since we utilize the scene graph representation as our compositional language, we use the Stanford Scene Graph Parser [63,77] to parse texts in $\\mathcal{D}_{train}$ into their corresponding scene graphs with objects, attributes and relationships. Since the parser only parses for objects and relationships, we further extract the attributes from the text via spaCy's natural language processing parser by identifying adjective part-of-speech tags. These connected objects, attributes, and relationships constitute our seen atoms and compounds. Similarly, we parse a given $\\mathcal{D}_{test}$ and divide all the image-text pairs into the three splits based on the presence of unseen atoms and/or compounds in the parsed training set. Details on the quality of the scene graph parser can be found in the Appendix.", + "bbox": [ + 75, + 244, + 470, + 455 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Filtering low-quality data We perform the following filtering steps on the image-text pairs in all splits: we only keep region crops which have an area greater than or equal to $40K$ pixels, occupy at least $10\\%$ of the whole image, and whose width-to-height ratio is between 0.5-2.0. We only include text which have at least 2 atoms and 1 compound and de-duplicate text using their corresponding scene graphs.", + "bbox": [ + 75, + 458, + 470, + 566 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Generating hard negatives We introduce two types of hard negatives: HN-ATOM and HN-COMP. HN-ATOM replaces $A_{a}$ , $A_{o}$ , or $A_{r}$ in the text with an atomic foil. For example, for the caption \"a grill on top of the porch\", one HN-ATOM can be \"a grill underneath the porch\", where the $A_{r}$ \"on top of\" is replaced by \"underneath\". Since captions and scene graphs are not exhaustive, this replacement must be done carefully. For example, if a dog is white and furry, but only \"white\" is annotated, replacing the atom \"white\" with \"furry\" would result in a correct caption. To minimize errors, we employ WordNet [49] to pick replacement atoms that are either antonyms (\"black dog\") or share the same grand-hypernym (\"pink dog\") with respect to the original atom. Furthermore, we use BERT to select the most sensical negatives for each ground truth caption [13,51]. HN-COMP concatenates two compound foils where each contains an atomic foil. For instance, one HN-COMP of the caption \"a pink car\" can be \"a blue car and a pink toy\", where \"blue\" and \"toy\" are the atomic foils in the two compounds foils \"blue car\" and \"pink toy\". We only generate negatives for one-compound examples for systematicity evaluation, as productivity covers complex captions with more atoms.", + "bbox": [ + 75, + 568, + 472, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Creating productivity datasets", + "text_level": 1, + "bbox": [ + 500, + 90, + 769, + 107 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We first generate ground truth captions for scene graphs of varying complexity, filter for data quality, and then generate hard negatives for each example (Figure 3).", + "bbox": [ + 498, + 113, + 893, + 157 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Generating captions We systematically generate captions of different atom counts for each image. Given a scene graph, we perform a random walk of length $n$ through the graph to generate a subgraph. Each subgraph corresponds to a specific region of the image, determined by the union of the bounding boxes of the subgraph atoms. We filter out low-quality regions using the same process as systematicity with additional dedduplication on patches that overlap by $\\geq 75\\%$ . For simple subgraphs ( $n = 4$ ), we produce captions using handcrafted templates. For larger subgraphs ( $n \\geq 5$ ), we leverage GPT-3 [7] (text-davinci-002) to generate captions based on a text description of the scene graph, which lists all objects and relationships. We prompt GPT-3 using 5 manually written captions per complexity, filtering out captions where GPT-3 errs and omits atoms from the subgraph during generation (see more details in Appendix).", + "bbox": [ + 496, + 159, + 893, + 400 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Generating hard negatives For productivity, we employ three hard negatives types (HN-ATOM from systematicity, HN-SWAP, and HN-NEG) corresponding to three hypothesized model error modes. First, as a caption's complexity increases, a model may begin to ignore individual atoms. HN-ATOM randomly selects an atom from the caption and replaces it with an incorrect atom. Second, as a caption's complexity increases, a model may treat captions as \"bags of words\", ignoring syntactic connections built out of word order. A swap hard negative (HN-SWAP) accordingly permutes atoms of the same subtype in a caption. This hard negative is similar to Winoground [75], but in the context of varying caption complexity. On top of Wordnet, we use entailment with RoBERTa to further filter errant HN-SWAP hard negatives [47]. Finally, as a caption's complexity increases, a model may begin to lose comprehension of negations. A negation hard negative (HN-NEG) either negates the entire caption or a specific atom. Refer to the Appendix for details on generating HN-SWAP and HN-NEG.", + "bbox": [ + 496, + 400, + 893, + 686 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.3. The final benchmark datasets", + "text_level": 1, + "bbox": [ + 500, + 696, + 764, + 710 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For both productivity and systematicity, we generate two test datasets: $\\mathcal{D}_{test}^{HN}$ , which contains image-ground truth text pairs along with all generated hard negatives, and $D_{test}^{RAW}$ , which contains only image-ground truth text pairs. To measure the data quality, we randomly sample $2\\%$ of productivity ground truth captions generated by GPT-3 and $1\\%$ of the queries in the productivity and systematicity $\\mathcal{D}_{test}^{HN}$ sets for manual human verification. We assign 2 annotators to each set and measure both generated quality and intra-annotator agreement. $87.9\\%$ of sampled productivity ground truth captions generated by GPT-3 are rated as faithful to the image, with an average pairwise annotator agreement of", + "bbox": [ + 496, + 719, + 893, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "10914", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/ed36c6a8819305e748103f78ece790008792765f4fdd142efd834a9cff626ecd.jpg", + "image_caption": [ + "UC vs. SC Recall@1 in Systematicity Hard Negative Retrieval Set" + ], + "image_footnote": [], + "bbox": [ + 81, + 103, + 295, + 246 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/2621a403d99f578ed9f00e8ee9db686a9d364a27bd02d80ffc9f61299891aeed.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 305, + 103, + 514, + 246 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/e0a0d07ee20f91e37fadf885a6f76c1e160233bdbd168e786d74f060721bd78f.jpg", + "image_caption": [ + "Figure 4. We plot models' recall@1 on the Seen Compounds vs. Unseen Compounds split of the systematicity retrieval set with hard negatives HN-ATOM, HN-COMP and both types. We observe a consistent drop in models' performance from the SC to UC split when the hard negative set consists of both HN-ATOM and HN-COMP or HN-ATOM only." + ], + "image_footnote": [], + "bbox": [ + 522, + 103, + 883, + 247 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/15f88ae173b78505f0baced8169ba753fb9391b0ba48e585b2868b727d2f448a.jpg", + "image_caption": [ + "Figure 5. Productivity Analysis. We plot models' Recall@1 on the hard negatives retrieval set against complexity, averaged across all models pretrained on all three training datasets. We find that models' ability to retrieve the ground-truth degrades as complexity increases." + ], + "image_footnote": [], + "bbox": [ + 161, + 316, + 403, + 484 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/828c834819114f045306472fb4615b1148643cfe9e6d839c3b2327547c069ec4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 406, + 316, + 604, + 484 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/7d7e86b7d68791de34e7909e6e86fefa9cecd38e116ecd7b3197df03406cc2bc.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 609, + 318, + 810, + 484 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "88.8%. 83.7% of productivity and 86.0% of systematicity hard negatives were rated as genuine negatives (i.e. made factually incorrect statements about the image), with pairwise annotator agreements of 84.3% and 83.7% respectively.", + "bbox": [ + 75, + 551, + 472, + 613 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 76, + 630, + 209, + 647 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We present our experimental setup and results with six takeaways. First, our systematicity experiments show performance decreases consistently on compounds unseen in training. Second, the greatest drop between splits occurs for models trained on LAION-400M. Third, our productivity results reveal models' retrieval performance decays with increasing complexity. Fourth, we find that dataset size has no impact on compositionality. Fifth, we find no clear trend relating model size to compositionality. Finally, models' zero-shot ImageNet classification accuracy correlates with retrieval performance on the systematicity dataset but not systematic generalization to the UC split or productivity.", + "bbox": [ + 75, + 657, + 468, + 838 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Datasets. We utilize Visual Genome to create our test datasets. For systematicity, image patches and corresponding spelling-corrected region descriptions are used. We provide three different splits for $\\mathcal{D}_{test}^{HN}$ , for three training datasets:", + "bbox": [ + 75, + 839, + 470, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "CC-12M, YFCC-15M and LAION-400M. For productivity, Visual Genome's image-scene graph pairs are used to create captions and hard negatives for $\\mathcal{D}_{test}^{RAW}$ and $\\mathcal{D}_{test}^{HN}$ (Table 1). Models. We firstly evaluate seven vision-language models pretrained with contrastive loss [70] across three commonly used image-text datasets: Conceptual Captions 12M (CC-12M) [8], a subset of the YFCC100M dataset (YFCC-15M) [56, 74] and LAION-400M [62]. We limit our evaluation to models openly released in the OpenCLIP repository [28] for systematicity evaluation. These include ResNet (RN) [20] and Vision Transformer (ViT) [14] encoders of different sizes: RN50, RN101, ViT-B-16, ViT-B-16-plus-240, ViT-B-32 and ViT-L-14. Additionally, since productivity evaluation is not restricted to models that were trained on publicly released datasets, we conduct productivity evaluation on other foundation vision-language models as well. Specifically, we consider OpenAI's CLIP [56] with ResNet and ViT backbones, CyCLIP [18] (a variant of CLIP introducing auxiliary losses that regularize the gap in similarity scores between mismatched pairs, trained on Conceptual Captions 3M [64] with a ResNet-50 [20] backbone), ALBEF [41] (additionally trained with a masked language modeling and image-text matching loss) and FLAVA [69] (which", + "bbox": [ + 496, + 551, + 895, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "10915", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/e5a2090aee51b8259ee05d7ddcbd12458a75b88f96749bb4501142027638bf9a.jpg", + "image_caption": [ + "Figure 6. Productivity Analysis on Additional Foundation Vision-language Models. We plot models' Recall@1 on the productivity hard negatives retrieval set against complexity, where OpenAI CLIP's performance is averaged across five models RN50, RN101, ViT-B-16, ViT-B-32 and ViT-L-14. We find that all models' retrieval performance decreases as complexity increases in both the HN-ATOM and HN-SWAP retrieval sets. For the HN-NEG set, all models except for CLIP either drop in performance or remain at random chance." + ], + "image_footnote": [], + "bbox": [ + 161, + 88, + 816, + 250 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/aae2bbb24f2aa7dfb837046d5a26df655857613ca4224f7bc0e1a7b78debef63.jpg", + "image_caption": [ + "Figure 7. A plot showing the correlation between zero-shot top-1 accuracy on ImageNet and Recall@1 on CREPE's systematicity hard-negative sets. We observe a strong correlation, with an $R^2$ score of 0.9914 for the SC split and 0.9534 for the UC split." + ], + "image_footnote": [], + "bbox": [ + 78, + 342, + 460, + 486 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "further adds unimodal losses for image and text domains). Retrieval. For $\\mathcal{D}_{test}^{HN}$ , we perform image-to-text retrieval and stratify results by split and hard negative type. For systematicity, the splits are $SC$ , $UC$ , and $UA$ ; for productivity, the splits are by caption complexity $n$ (denoted $\\mathcal{D}_{test}^{HN,n}$ ). Each retrieval task is between one image and its ground truth caption plus $h$ hard negatives of a single type (see Appendix). We adopt commonly used retrieval metrics Recall@1, 3, 5 and Average Recall@K. For $\\mathcal{D}_{test}^{RAW}$ , retrieval experiments are described in the Appendix.", + "bbox": [ + 75, + 580, + 467, + 734 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.1. Systematicity evaluation", + "text_level": 1, + "bbox": [ + 76, + 741, + 302, + 758 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Model performance on the $\\mathcal{D}_{test}^{HN}$ dataset for systematicity decreases monotonically when compounds are unseen. We first observe a monotonic decrease in recall@1 from the Seen Compounds to the Unseen Compounds split on the systematicity $\\mathcal{D}_{test}^{HN}$ set consisting of both HN-ATOM and HN-COMP (Figure 4 left). This drop is relatively small $(1 - 5\\%)$ for the CC-12M and YFCC-15M trained models and the most pronounced for models trained on the largest dataset LAION-400M [62], with the decrease reaching $7.9\\%$", + "bbox": [ + 75, + 763, + 472, + 901 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "for the ViT-B-32 model. However, CC-12M and YFCC-15M models also significantly underperform LAION-400M models in general, meaning that small drops between sets may be due to overall poor performance rather than improved systematic generalization. In comparison, human oracle experiments generalize with $100\\%$ accuracy to $\\mathcal{D}_{test}^{HN}$ .", + "bbox": [ + 496, + 340, + 893, + 434 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Similar to the overall results, there is also a consistent discrepancy between the SC and UC split on the $\\mathcal{D}_{test}^{HN}$ subset consisting of HN-ATOM only (Figure 4 center). This drop is consistently smaller (3 - 6%) for models trained on CC-12M and YFCC-15M, but pronounced (6% or higher, reaching 9.4% drop for ViT-B-32) for LAION-400M models.", + "bbox": [ + 496, + 434, + 893, + 525 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "On the HN-COMP subset (Figure 4 right), we find little to no difference in performance between the $SC$ and $UC$ split. We hypothesize that this is due to the lower difficulty of the HN-COMP hard negatives, as they introduce more foils to the caption, are always longer than the ground truth, and thus offer more opportunities for the model to correctly distinguish the ground truth. This hypothesis is supported by the fact that Recall@1 values on HN-COMP are similar or higher than the ones on HN-ATOM even though the HN-COMP retrieval set size is larger than that of HN-ATOM.", + "bbox": [ + 496, + 527, + 895, + 679 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.2. Productivity evaluation", + "text_level": 1, + "bbox": [ + 500, + 694, + 718, + 709 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Models' performance decreases with complexity on HN-ATOM and HN-SWAP negatives. At small complexities such as $n = 4$ , we observe that model retrieval quality is well above random chance (Figure 5). However, as caption complexity increases, we observe a steady decrease in performance, nearing random chance for HN-SWAP and dipping below it for HN-ATOM negatives. Similarly, we find that the same downward trend persists for other vision-language foundation models (Figure 6). Importantly, the downward trend occurs for FLAVA and ALBEF even though their training set contains Visual Genome images. We note that for HN-NEG negatives, the OpenAI CLIP models do not adhere", + "bbox": [ + 496, + 719, + 895, + 901 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "10916", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/39d5d0840bc78533ebbca857036d61b7ca5b8948ee453f276cbe05f9bda4d0ca.jpg", + "image_caption": [ + "Figure 8. A plot showing the correlation between zero-shot top-1 accuracy on ImageNet and Recall@1 on CREPE's productivity hard negative sets for complexities of 4, 8 and 12. Overall, we find no correlation between ImageNet accuracy and Recall@1 on our productivity sets. The strongest correlations are $R^2$ scores of 0.284 for HN-NEG negatives on $n = 12$ and of 0.222 for HN-ATOM negatives on $n = 8$ ." + ], + "image_footnote": [], + "bbox": [ + 109, + 89, + 851, + 256 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "to the downward trend, achieving their lowest scores for the lowest complexity. Their performances on higher complexities, however, show great variation. In short, our conclusion is that vision-language foundation models struggle with productivity. Our results on models released by OpenCLIP [28] as well as other vision-language foundation models demonstrate the challenge of differentiating between atomic and swapping foils is exacerbated by caption complexity.", + "bbox": [ + 75, + 339, + 472, + 460 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We see no effect of dataset size on productivity. We do not observe a clear advantage for larger pretraining datasets in our productivity evaluation. For atomic and swapping foils, we see similar performance for models trained on the three datasets, with slightly worse performance on atomic foils for the CC-12M trained models. However, on negation hard negatives (Figure 5), we see variable performance across training sets, with CC-12M models outperforming larger models trained on larger datasets YFCC and LAION.", + "bbox": [ + 75, + 460, + 472, + 595 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.3. Effect of model size", + "text_level": 1, + "bbox": [ + 76, + 604, + 264, + 619 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We find no trends relating compositionality to model size. Overall, we note that the LAION trained models (which are both larger models and trained on larger datasets) achieve significantly better absolute performances than smaller models. However, model's systematicity and productivity remain indifferent to the size of the model itself (Figures 4 and 5).", + "bbox": [ + 75, + 627, + 470, + 718 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.4. Correlation with ImageNet performance", + "text_level": 1, + "bbox": [ + 76, + 726, + 423, + 742 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We find that zero-shot ImageNet top-1 accuracy strongly correlates with models' Recall@1 on the systematicity retrieval set. Specifically, we acquire $R^2$ scores of 0.984 and 0.877 for the $SC$ and $UC$ splits respectively (Figure 7). However, this correlation does not imply that models' zero-shot ImageNet performance correlates with systematic generalization, which is instead indicated by small or no difference between the $SC$ and $UC$ splits. On our productivity dataset, we do not observe such a strong correlation, where the highest $R^2$ score is 0.284 for HN-NEG negatives on a complexity", + "bbox": [ + 75, + 750, + 472, + 902 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "of $n = 12$ (Figure 8). As such, we can infer that successful zero-shot performance on ImageNet does not necessarily lead to better performance on our productivity sets.", + "bbox": [ + 496, + 339, + 893, + 385 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Discussion", + "text_level": 1, + "bbox": [ + 500, + 398, + 612, + 414 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Limitations. First, although our data validation protocols verified our generated hard negatives for productivity as high-quality, approximately $70\\%$ of HN-SWAP and of HN-NEG negatives were rated as correct. While this does not invalidate our key productivity result, this noise is a limitation of CREPE and could hinder future evaluations once foundation models begin performing better. Second, our evaluation only covers a limited set of vision-language foundation models that were trained with contrastive loss. Additionally, given the computational requirements associated with training a foundation model, our experiments centered around model architectures that were already available publicly. We hope that future foundation models are evaluated with our publicly available CREPE benchmark. Third, while we observe text-to-image and image-to-text retrieval to have similar trends for our systematicity experiments, we lack text-to-image datasets with hard negatives. Future work can explore mechanisms to generate counterfactual negative images.", + "bbox": [ + 496, + 422, + 893, + 696 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Conclusion. We present CREPE, a collection of text-to-image and image-to-text retrieval datasets for evaluating pretrained vision-language models' systematicity and productivity. We demonstrate that models struggle with compositionality along both axes, with performance drops across different compositional splits and increasing complexity. We expect that CREPE will provide a more systematic evaluation to benchmark the emergence of compositionality as future models improve. Finally, researchers can leverage our hard-negative generation process to create training batches with hard negatives to incentivize vision-language compositionality.", + "bbox": [ + 496, + 712, + 895, + 896 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "10917", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 90, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Arjun Akula, Spandana Gella, Yaser Al-Onaizan, Song-Chun Zhu, and Siva Reddy. Words aren't enough, their order matters: On the robustness of grounding visual referring expressions. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 6555-6565, Online, July 2020. Association for Computational Linguistics.", + "[2] Jacob Andreas, Marcus Rohrbach, Trevor Darrell, and Dan Klein. Neural module networks, 2015.", + "[3] Dzmitry Bahdanau, Harm de Vries, Timothy J O'Donnell, Shikhar Murty, Philippe Beaudoin, Yoshua Bengio, and Aaron Courville. Closure: Assessing systematic generalization of clevr models. arXiv preprint arXiv:1912.05783, 2019.", + "[4] Ben Bogin, Shivanshu Gupta, Matt Gardner, and Jonathan Berant. COVR: A test-bed for visually grounded compositional generalization with real images. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 9824-9846, Online and Punta Cana, Dominican Republic, Nov. 2021. Association for Computational Linguistics.", + "[5] Rishi Bommasani, Drew A Hudson, Ehsan Adeli, Russ Altman, Simran Arora, Sydney von Arx, Michael S Bernstein, Jeannette Bohg, Antoine Bosselut, Emma Brunskill, et al. On the opportunities and risks of foundation models. arXiv preprint arXiv:2108.07258, 2021.", + "[6] Léon Bottou. From machine learning to machine reasoning. Machine learning, 94(2):133-149, 2014.", + "[7] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. Advances in neural information processing systems, 33:1877-1901, 2020.", + "[8] Soravit Changpinyo, Piyush Sharma, Nan Ding, and Radu Soricut. Conceptual 12m: Pushing web-scale image-text pretraining to recognize long-tail visual concepts. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3558-3568, 2021.", + "[9] Zhenfang Chen, Peng Wang, Lin Ma, Kwan-Yee K. Wong, and Qi Wu. Cops-ref: A new dataset and task on compositional referring expression comprehension, 2020.", + "[10] Colin Conwell and Tomer Ullman. Testing relational understanding in text-guided image generation. arXiv preprint arXiv:2208.00005, 2022.", + "[11] MJ Cresswell. Logics and languages. 1973.", + "[12] Yuchen Cui, Scott Niekum, Abhinav Gupta, Vikash Kumar, and Aravind Rajeswaran. Can foundation models perform zero-shot task specification for robot manipulation? In Roya Firoozi, Negar Mehr, Esen Yel, Rika Antonova, Jeannette Bohg, Mac Schwager, and Mykel Kochenderfer, editors, Proceedings of The 4th Annual Learning for Dynamics and Control Conference, volume 168 of Proceedings of Machine Learning Research, pages 893-905. PMLR, 23-24 Jun 2022.", + "[13] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional" + ], + "bbox": [ + 78, + 114, + 472, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018.", + "[14] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations, 2021.", + "[15] Jiali Duan, Liquen Chen, Son Tran, Jinyu Yang, Yi Xu, Belinda Zeng, and Trishul Chilimbi. Multi-modal alignment using representation codebook, 2022.", + "[16] Sepideh Esmaeilpour, Bing Liu, Eric Robertson, and Lei Shu. Zero-shot out-of-distribution detection based on the pre-trained model clip, 2021.", + "[17] Mona Gandhi, Mustafa O. Gul, Eva Prakash, Madeleine Grunde-McLaughlin, Ranjay Krishna, and Maneesh Agrawala. Measuring compositional consistency for video question answering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022.", + "[18] Shashank Goel, Hritik Bansal, Sumit Bhatia, Ryan A. Rossi, Vishwa Vinay, and Aditya Grover. Cyclip: Cyclic contrastive language-image pretraining, 2022.", + "[19] Madeleine Grunde-McLaughlin, Ranjay Krishna, and Maneesh Agrawala. Agqa: A benchmark for compositional spatio-temporal reasoning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2021.", + "[20] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016.", + "[21] Lisa Anne Hendricks and Aida Nematzadeh. Probing image-language transformers for verb understanding, 2021.", + "[22] Jack Hessel, Ari Holtzman, Maxwell Forbes, Ronan Le Bras, and Yejin Choi. CLIPScore: A reference-free evaluation metric for image captioning. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 7514-7528, Online and Punta Cana, Dominican Republic, Nov. 2021. Association for Computational Linguistics.", + "[23] Irina Higgins, Nicolas Sonnerat, Loic Matthey, Arka Pal, Christopher P Burgess, Matko Bosnjak, Murray Shanahan, Matthew Botvinick, Demis Hassabis, and Alexander Lerchner. Scan: Learning hierarchical compositional visual concepts. arXiv preprint arXiv:1707.03389, 2017.", + "[24] Hexiang Hu, Ishan Misra, and Laurens van der Maaten. Evaluating text-to-image matching using binary image selection (bison), 2019.", + "[25] Drew A. Hudson and Christopher D. Manning. Compositional attention networks for machine reasoning, 2018.", + "[26] Drew A Hudson and Christopher D Manning. Gqa: A new dataset for real-world visual reasoning and compositional question answering. Conference on Computer Vision and Pattern Recognition (CVPR), 2019.", + "[27] Dieuwke Hupkes, Verna Dankers, Mathijs Mul, and Elia Bruni. Compositionality decomposed: How do neural networks generalise? Journal of Artificial Intelligence Research, 67:757-795, 2020." + ], + "bbox": [ + 501, + 92, + 893, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "10918", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[28] Gabriel Ilharco, Mitchell Wortsman, Ross Wightman, Cade Gordon, Nicholas Carlini, Rohan Taori, Achal Dave, Vaishaal Shankar, Hongseok Namkoong, John Miller, Hannaneh Hajishirzi, Ali Farhadi, and Ludwig Schmidt. OpenCLIP, 7 2021.", + "[29] Ajay Jain, Matthew Tancik, and Pieter Abbeel. Putting nerf on a diet: Semantically consistent few-shot view synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 5885-5894, October 2021.", + "[30] Theo MV Janssen and Barbara H Partee. Compositionality. In Handbook of logic and language, pages 417-473. Elsevier, 1997.", + "[31] Jingwei Ji, Ranjay Krishna, Li Fei-Fei, and Juan Carlos Niebles. Action genome: Actions as compositions of spatiotemporal scene graphs. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10236–10247, 2020.", + "[32] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In International Conference on Machine Learning, pages 4904-4916. PMLR, 2021.", + "[33] Justin Johnson, Bharath Hariharan, Laurens van der Maaten, Li Fei-Fei, C. Lawrence Zitnick, and Ross Girshick. Clevr: A diagnostic dataset for compositional language and elementary visual reasoning, 2016.", + "[34] Daniel Keysers, Nathanael Scharli, Nathan Scales, Hylke Buisman, Daniel Furrer, Sergii Kashubin, Nikola Momchev, Danila Sinopalnikov, Lukasz Stafiniak, Tibor Tihon, Dmitry Tsarkov, Xiao Wang, Marc van Zee, and Olivier Bousquet. Measuring compositional generalization: A comprehensive method on realistic data. In International Conference on Learning Representations, 2020.", + "[35] Ranjay Krishna, Yuke Zhu, Oliver Groth, Justin Johnson, Kenji Hata, Joshua Kravitz, Stephanie Chen, Yannis Kalantidis, Li-Jia Li, David A Shamma, Michael Bernstein, and Li Fei-Fei. Visual genome: Connecting language and vision using crowdsourced dense image annotations. 2016.", + "[36] Brenden Lake and Marco Baroni. Generalization without systematicity: On the compositional skills of sequence-to-sequence recurrent networks. In International conference on machine learning, pages 2873-2882. PMLR, 2018.", + "[37] Brenden M. Lake, Tomer D. Ullman, Joshua B. Tenenbaum, and Samuel J. Gershman. Building machines that learn and think like people. Behavioral and Brain Sciences, 40:e253, 2017.", + "[38] Boyi Li, Kilian Q Weinberger, Serge Belongie, Vladlen Koltun, and Rene Ranftl. Language-driven semantic segmentation. In International Conference on Learning Representations, 2022.", + "[39] Dongxu Li, Junnan Li, Hongdong Li, Juan Carlos Niebles, and Steven C.H. Hoi. Align and prompt: Video-and-language pre-training with entity prompts. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4943-4953, 2022." + ], + "bbox": [ + 78, + 90, + 470, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[40] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In ICML, 2022.", + "[41] Junnan Li, Ramprasaath R. Selvaraju, Akhilesh Deepak Gotmare, Shafiq Joty, Caiming Xiong, and Steven Hoi. Align before fuse: Vision and language representation learning with momentum distillation. In A. Beygelzimer, Y. Dauphin, P. Liang, and J. Wortman Vaughan, editors, Advances in Neural Information Processing Systems, 2021.", + "[42] Wei Li, Can Gao, Guocheng Niu, Xinyan Xiao, Hao Liu, Jiachen Liu, Hua Wu, and Haifeng Wang. UNIMO-2: End-to-end unified vision-language grounded learning. In Findings of the Association for Computational Linguistics: ACL 2022, pages 3187–3201, Dublin, Ireland, May 2022. Association for Computational Linguistics.", + "[43] Yehao Li, Yingwei Pan, Ting Yao, and Tao Mei. Comprehending and ordering semantics for image captioning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022.", + "[44] Zhiheng Li, Martin Renqiang Min, Kai Li, and Chenliang Xu. Stylet2i: Toward compositional and high-fidelity text-to-image synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022.", + "[45] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In European conference on computer vision, pages 740-755. Springer, 2014.", + "[46] Xihui Liu, Dong Huk Park, Samaneh Azadi, Gong Zhang, Arman Chopikyan, Yuxiao Hu, Humphrey Shi, Anna Rohrbach, and Trevor Darrell. More control for free! image synthesis with semantic diffusion guidance, 2021.", + "[47] Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettle-moyer, and Veselin Stoyanov. Roberta: A robustly optimized bert pretraining approach. arXiv preprint arXiv:1907.11692, 2019.", + "[48] Haoyu Lu, Nanyi Fei, Yuqi Huo, Yizhao Gao, Zhiwu Lu, and Ji-Rong Wen. Cots: Collaborative two-stream vision-language pre-training model for cross-modal retrieval, 2022.", + "[49] George A Miller. Wordnet: a lexical database for english. Communications of the ACM, 38(11):39-41, 1995.", + "[50] Medhini Narasimhan, Anna Rohrbach, and Trevor Darrell. Clip-it! language-guided video summarization. 2021.", + "[51] Letitia Parcalabescu, Michele Cafagna, Lilitta Muradjan, Anette Frank, Iacer Calixto, and Albert Gatt. VALID: A task-independent benchmark for vision and language models centered on linguistic phenomena. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 8253-8280, Dublin, Ireland, May 2022. Association for Computational Linguistics.", + "[52] Dong Huk Park, Samaneh Azadi, Xihui Liu, Trevor Darrell, and Anna Rohrbach. Benchmark for compositional text-to-image synthesis. In Thirty-fifth Conference on Neural Information Processing Systems Datasets and Benchmarks Track (Round 1), 2021." + ], + "bbox": [ + 501, + 92, + 893, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10919", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[53] Jae Sung Park, Sheng Shen, Ali Farhadi, Trevor Darrell, Yejin Choi, and Anna Rohrbach. Exposing the limits of video-text models through contrast sets. In Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 3574-3586, Seattle, United States, July 2022. Association for Computational Linguistics.", + "[54] Or Patashnik, Zongze Wu, Eli Shechtman, Daniel Cohen-Or, and Dani Lischinski. Styleclip: Text-driven manipulation of stylegan imagery. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 2085–2094, October 2021.", + "[55] Suzanne Petryk, Lisa Dunlap, Keyan Nasseri, Joseph Gonzalez, Trevor Darrell, and Anna Rohrbach. On guiding visual attention with language specification. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 18071-18081, 2022.", + "[56] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763. PMLR, 2021.", + "[57] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 2022.", + "[58] Yongming Rao, Wenliang Zhao, Guangyi Chen, Yansong Tang, Zheng Zhu, Guan Huang, Jie Zhou, and Jiwen Lu. Denseclip: Language-guided dense prediction with context-aware prompting. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022.", + "[59] Laura Ruis, Jacob Andreas, Marco Baroni, Diane Bouchacourt, and Brenden M. Lake. A benchmark for systematic generalization in grounded language understanding. In Proceedings of the 34th International Conference on Neural Information Processing Systems, NIPS'20, Red Hook, NY, USA, 2020. Curran Associates Inc.", + "[60] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily Denton, Seyed Kamyar Seyed Ghasemipour, Burcu Karagol Ayan, S Sara Mahdavi, Rapha Gontijo Lopes, et al. Photorealistic text-to-image diffusion models with deep language understanding. arXiv preprint arXiv:2205.11487, 2022.", + "[61] Maarten Sap, Ronan Le Bras, Emily Allaway, Chandra Bhagavatula, Nicholas Lourie, Hannah Rashkin, Brendan Roof, Noah A. Smith, and Yejin Choi. ATOMIC: an atlas of machine commonsense for if-then reasoning. CoRR, abs/1811.00146, 2018.", + "[62] Christoph Schuhmann, Richard Vencu, Romain Beaumont, Robert Kaczmarczyk, Clayton Mullis, Aarush Katta, Theo Coombes, Jenia Jitsev, and Aran Komatsuzaki. Laion-400m: Open dataset of clip-filtered 400 million image-text pairs. arXiv preprint arXiv:2111.02114, 2021.", + "[63] Sebastian Schuster, Ranjay Krishna, Angel Chang, Li Fei-Fei, and Christopher D Manning. Generating semantically precise scene graphs from textual descriptions for improved image" + ], + "bbox": [ + 78, + 90, + 470, + 901 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "retrieval. In Proceedings of the fourth workshop on vision and language, pages 70-80, 2015.", + "[64] Piyush Sharma, Nan Ding, Sebastian Goodman, and Radu Soricut. Conceptual captions: A cleaned, hypernymed, image alt-text dataset for automatic image captioning. In Proceedings of ACL, 2018.", + "[65] Ravi Shekhar, Sandro Pezzelle, Yauhen Klimovich, Aurélie Herbelot, Moin Nabi, Enver Sangineto, and Raffaella Bernardi. FOIL it! find one mismatch between image and language caption. In Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 255-265, Vancouver, Canada, July 2017. Association for Computational Linguistics.", + "[66] Sheng Shen, Liunian Harold Li, Hao Tan, Mohit Bansal, Anna Rohrbach, Kai-Wei Chang, Zhewei Yao, and Kurt Keutzer. How much can CLIP benefit vision-and-language tasks? In International Conference on Learning Representations, 2022.", + "[67] Hengcan Shi, Munawar Hayat, Yicheng Wu, and Jianfei Cai. Proposalclip: Unsupervised open-category object proposal generation via exploiting clip cues, 2022.", + "[68] Mohit Shridhar, Lucas Manuelli, and Dieter Fox. CIiport: What and where pathways for robotic manipulation. In Proceedings of the 5th Conference on Robot Learning (CoRL), 2021.", + "[69] Amanpreet Singh, Ronghang Hu, Vedanuj Goswami, Guillaume Couairon, Wojciech Galuba, Marcus Rohrbach, and Douwe Kiela. Flava: A foundational language and vision alignment model, 2021.", + "[70] Kihyuk Sohn. Improved deep metric learning with multiclass n-pair loss objective. Advances in neural information processing systems, 29, 2016.", + "[71] Sanjay Subramanian, Will Merrill, Trevor Darrell, Matt Gardner, Sameer Singh, and Anna Rohrbach. Reclip: A strong zero-shot baseline for referring expression comprehension. arXiv preprint arXiv:2204.05991, 2022.", + "[72] Alane Suhr, Stephanie Zhou, Ally Zhang, Iris Zhang, Huajun Bai, and Yoav Artzi. A corpus for reasoning about natural language grounded in photographs. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 6418-6428, Florence, Italy, July 2019. Association for Computational Linguistics.", + "[73] Yoad Tewel, Yoav Shalev, Idan Schwartz, and Lior Wolf. Zero-shot image-to-text generation for visual-semantic arithmetic. arXiv preprint arXiv:2111.14447, 2021.", + "[74] Bart Thomee, David A Shamma, Gerald Friedland, Benjamin Elizalde, Karl Ni, Douglas Poland, Damian Borth, and Li-Jia Li. Yfcc100m: The new data in multimedia research. Communications of the ACM, 59(2):64-73, 2016.", + "[75] Tristan Thrush, Ryan Jiang, Max Bartolo, Amanpreet Singh, Adina Williams, Douwe Kiela, and Candace Ross. Winoground: Probing vision and language models for visio-linguistic compositionality. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5238–5248, 2022.", + "[76] Mengmeng Wang, Jiazheng Xing, and Yong Liu. Actionclip: A new paradigm for video action recognition, 2021.", + "[77] Hao Wu, Jiayuan Mao, Yufeng Zhang, Yuning Jiang, Lei Li, Weiwei Sun, and Wei-Ying Ma. Unified visual-semantic" + ], + "bbox": [ + 503, + 92, + 893, + 900 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "10920", + "bbox": [ + 480, + 945, + 519, + 955 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "embeddings: Bridging vision and language with structured meaning representations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6609-6618, 2019.", + "[78] Hu Xu, Gargi Ghosh, Po-Yao Huang, Dmytro Okhonko, Armen Aghajanyan, Florian Metze, Luke Zettlemoyer, and Christoph Feichtenhofer. VideoCLIP: Contrastive pretraining for zero-shot video-text understanding. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 6787-6800, Online and Punta Cana, Dominican Republic, Nov. 2021. Association for Computational Linguistics.", + "[79] Zipeng Xu, Tianwei Lin, Hao Tang, Fu Li, Dongliang He, Nicu Sebe, Radu Timofte, Luc Van Gool, and Errui Ding. Predict, prevent, and evaluate: Disentangled text-driven image manipulation empowered by pre-trained vision-language model. arXiv preprint arXiv:2111.13333, 2021.", + "[80] Jinyu Yang, Jiali Duan, Son Tran, Yi Xu, Sampath Chanda, Liquun Chen, Belinda Zeng, Trishul Chilimbi, and Junzhou Huang. Vision-language pre-training with triple contrastive learning. 2022.", + "[81] Jiahui Yu, Zirui Wang, Vijay Vasudevan, Legg Yeung, Mojtaba Seyedhosseini, and Yonghui Wu. Coca: Contrastive captioners are image-text foundation models. Transactions on Machine Learning Research, 2022.", + "[82] Pengchuan Zhang, Xiujun Li, Xiaowei Hu, Jianwei Yang, Lei Zhang, Lijuan Wang, Yejin Choi, and Jianfeng Gao. Vinvl: Revisiting visual representations in vision-language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5579-5588, June 2021.", + "[83] Yufan Zhou, Ruiyi Zhang, Changyou Chen, Chunyuan Li, Chris Tensmeyer, Tong Yu, Jiuming Gu, Jinhui Xu, and Tong Sun. Lafite: Towards language-free training for text-to-image generation. arXiv preprint arXiv:2111.13792, 2021." + ], + "bbox": [ + 78, + 90, + 470, + 584 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "10921", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 11 + } +] \ No newline at end of file diff --git a/2023/CREPE_ Can Vision-Language Foundation Models Reason Compositionally_/6b09bf56-aeff-4b97-8111-a915e425b709_model.json b/2023/CREPE_ Can Vision-Language Foundation Models Reason Compositionally_/6b09bf56-aeff-4b97-8111-a915e425b709_model.json new file mode 100644 index 0000000000000000000000000000000000000000..d913cc5be93bc305e2ed1ba2f9b862bfecd398ff --- /dev/null +++ b/2023/CREPE_ Can Vision-Language Foundation Models Reason Compositionally_/6b09bf56-aeff-4b97-8111-a915e425b709_model.json @@ -0,0 +1,2281 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "image", + "bbox": [ + 0.086, + 0.127, + 0.108, + 0.148 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.11, + 0.13, + 0.885, + 0.154 + ], + "angle": 0, + "content": "CREPE: Can Vision-Language Foundation Models Reason Compositionally?" + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.18, + 0.873, + 0.217 + ], + "angle": 0, + "content": "Zixian Ma\\(^{1*}\\), Jerry Hong\\(^{1*}\\), Mustafa Omer Gul\\(^{2*}\\), Mona Gandhi\\(^{3}\\), Irenea Gao\\(^{1}\\), Ranjay Krishna\\(^{4}\\), Stanford University\\(^{1}\\), Cornell University\\(^{2}\\), University of Pennsylvania\\(^{3}\\), University of Washington\\(^{4}\\)" + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.22, + 0.852, + 0.252 + ], + "angle": 0, + "content": "{zixianma, jerryhong, irena}@cs.stanford.edu mog29@cornell.edu mona09@seas.upenn.edu ranjay@cs.washington.edu" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.287, + 0.314, + 0.303 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.319, + 0.474, + 0.729 + ], + "angle": 0, + "content": "A fundamental characteristic common to both human vision and natural language is their compositional nature. Yet, despite the performance gains contributed by large vision and language pretraining, we find that—across 7 architectures trained with 4 algorithms on massive datasets—they struggle at compositionality. To arrive at this conclusion, we introduce a new compositionality evaluation benchmark, CREPE, which measures two important aspects of compositionality identified by cognitive science literature: systematicity and productivity. To measure systematicity, CREPE consists of a test dataset containing over \\(370K\\) image-text pairs and three different seen-unseen splits. The three splits are designed to test models trained on three popular training datasets: CC-12M, YFCC-15M, and LAION-400M. We also generate \\(325K\\), \\(316K\\), and \\(309K\\) hard negative captions for a subset of the pairs. To test productivity, CREPE contains \\(17K\\) image-text pairs with nine different complexities plus \\(278K\\) hard negative captions with atomic, swapping and negation foils. The datasets are generated by repurposing the Visual Genome scene graphs and region descriptions and applying handcrafted templates and GPT-3. For systematicity, we find that model performance decreases consistently when novel compositions dominate the retrieval set, with Recall@1 dropping by up to \\(9\\%\\). For productivity, models' retrieval success decays as complexity increases, frequently nearing random chance at high complexity. These results hold regardless of model and training dataset size." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.758, + 0.21, + 0.774 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.785, + 0.472, + 0.876 + ], + "angle": 0, + "content": "Compositionality, the understanding that \"the meaning of the whole is a function of the meanings of its parts\" [11], is held to be a key characteristic of human intelligence. In language, the whole is a sentence, made up of words. In vision, the whole is a scene, made up of parts like objects, their attributes, and their relationships [31, 35]." + }, + { + "type": "image", + "bbox": [ + 0.52, + 0.288, + 0.688, + 0.467 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.706, + 0.289, + 0.876, + 0.467 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.485, + 0.896, + 0.626 + ], + "angle": 0, + "content": "Figure 1. We introduce CREPE, a benchmark to evaluate whether vision-language foundation models demonstrate two fundamental aspects of compositionality: systematicity and productivity. To evaluate systematicity, CREPE utilizes Visual Genome and introduces three new test datasets for the three popular pretraining datasets: CC-12M, YFCC-15M, and LAION-400M. These enable evaluating models' abilities to systematically generalize their understanding to seen compounds, unseen compounds, and even unseen atoms. To evaluate productivity, CREPE introduces examples of nine complexities, with three types of hard negatives for each." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.657, + 0.895, + 0.838 + ], + "angle": 0, + "content": "Through compositional reasoning, humans can understand new scenes and generate complex sentences by combining known parts [6, 27, 30]. Despite compositionality's importance, there are no large-scale benchmarks directly evaluating whether vision-language models can reason compositionally. These models are pretrained using large-scale image-capture datasets [62, 64, 74], and are already widely applied for tasks that benefit from compositional reasoning, including retrieval, text-to-image generation, and open-vocabulary classification [10, 57, 60]. Especially as such models become ubiquitous \"foundations\" for other models [5], it is critical to understand their compositional abilities." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.895, + 0.902 + ], + "angle": 0, + "content": "Previous work has evaluated these models using image-text retrieval [32,56,82]. However, the retrieval datasets used either do not provide controlled sets of negatives [45, 74] or study narrow negatives which vary along a single axis" + }, + { + "type": "page_footnote", + "bbox": [ + 0.097, + 0.887, + 0.203, + 0.9 + ], + "angle": 0, + "content": "*Equal contribution" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "10910" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.091, + 0.086, + 0.862, + 0.284 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.296, + 0.897, + 0.354 + ], + "angle": 0, + "content": "Figure 2. An overview of the systematicity retrieval set generation process. First, a model's image-caption training set is parsed to identify what atoms and compounds the model has seen. Then, an evaluation set is divided into three compositional splits according to whether the model has seen all the compounds (Seen Compounds), only all the atoms of the caption (Unseen Compounds), or neither (Unseen Atoms). Finally, hard negative captions HN-ATOM and HN-COMP are generated for the hard negatives retrieval set \\(\\mathcal{D}_{test}^{HN}\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.379, + 0.471, + 0.456 + ], + "angle": 0, + "content": "(e.g. permuted word orders or single word substitutions as negative captions) [21, 51, 65, 75]. Further, these analyses have also not studied how retrieval performance varies when generalizing to unseen compositional combinations, or to combinations of increased complexity." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.457, + 0.474, + 0.793 + ], + "angle": 0, + "content": "We introduce CREPE (Compositional REpresentation Evaluation): a new large-scale benchmark to evaluate two aspects of compositionality: systematicity and productivity (Figure 1). Systematicity measures how well a model is able to represent seen versus unseen atoms and their compositions. Productivity studies how well a model can comprehend an unbounded set of increasingly complex expressions. CREPE uses Visual Genome's scene graph representation as the compositionality language [35] and constructs evaluation datasets using its annotations. To test systematicity, we parse the captions in three popular training datasets, CC-12M [8], YFCC-15M [74], and LAION-400M [62], to identify atoms (objects, relations, or attributes) and compounds (combinations of atoms) present in each dataset. For each training set, we curate corresponding test sets containing \\(385K\\), \\(385K\\) and \\(373K\\) image-text pairs respectively, with splits checking generalization to seen compounds, unseen compounds, and unseen atoms. To test productivity, CREPE contains \\(17K\\) image-text pairs split across nine levels of complexity, as defined by the number of atoms present in the text. Examples across all datasets are paired with various hard negative types to ensure the legitimacy of our conclusions." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.796, + 0.472, + 0.903 + ], + "angle": 0, + "content": "Our experiments—across 7 architectures trained with 4 training algorithms on massive datasets—find that vision-language models struggle at compositionality, with both systematicity and productivity. We present six key findings: first, our systematicity experiments find that models' performance consistently drops between seen and unseen compositions; second, we observe larger drops for models trained on" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.379, + 0.895, + 0.515 + ], + "angle": 0, + "content": "LAION-400M (up to a \\(9\\%\\) decrease in Recall@1); third, our productivity experiments indicate that retrieval performance degrades with increased caption complexity; fourth, we find no clear trend relating training dataset size to models' compositional reasoning; fifth, model size also has no impact; finally, models' zero-shot ImageNet classification accuracy correlates only with their absolute retrieval performance on the systematicity dataset but not systematic generalization to unseen compounds or to productivity." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.532, + 0.642, + 0.548 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.559, + 0.895, + 0.619 + ], + "angle": 0, + "content": "Our work lies within the field of evaluating foundation models. Specifically, we measure visio-linguistic compositionality. To do so, we create a retrieval benchmark with hard negatives." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.621, + 0.896, + 0.863 + ], + "angle": 0, + "content": "Contrastive Image-Text Pretraining. The recently released contrastively trained CLIP model [56] has catalyzed a wide array of work at the intersection of Computer Vision and Natural Language Processing. Since its release, CLIP has enabled several tasks, ranging from semantic segmentation to image captioning, many of which have remarkable zero-shot capability [12, 16, 38, 56, 71, 73]. CLIP has been used as a loss function within image synthesis applications [29, 44, 46, 54, 79, 83], acted as an automated evaluation metric [22, 52], used successfully as a feature extractor for various vision and language tasks [66], and incorporated into architectures for various tasks including dense prediction and video summarization [43, 50, 55, 58, 67, 68]. This success has also encouraged the design of other contrastive vision and language pretraining algorithms for image [15, 18, 40–42, 48, 69, 80, 81] and video domains [39, 76, 78]. Our work evaluates how well" + }, + { + "type": "page_footnote", + "bbox": [ + 0.5, + 0.875, + 0.895, + 0.901 + ], + "angle": 0, + "content": "1We release our datasets, and code to generate and evaluate on our test sets at https://github.com/RAIVNLab/CREPE." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.958 + ], + "angle": 0, + "content": "10911" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.471, + 0.137 + ], + "angle": 0, + "content": "such contrastively trained models capture a fundamental property present in human vision and language: compositionality." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.142, + 0.472, + 0.445 + ], + "angle": 0, + "content": "Compositionality. Compositionality allows us to comprehend an infinite number of scenes and utterances [37]. For an AI model, compositionality would not only allow for systematic, combinatorial generalization, but would also confer benefits such as controllability [5]. This promise prompted a wealth of work on both designing [2, 23, 25] and evaluating [17, 19, 26, 36, 72] compositional models. In our work, we focus on two aspects of compositionality: systematicity and productivity. While there is a plethora of benchmarks for systematic generalization within Computer Vision [3, 4, 19, 33] and Machine Learning [34, 36, 59], the subject has been almost unexplored for vision-language models, largely due to lack of benchmarks complementary to the different largescale training datasets. To address this, CREPE provides a benchmark with three different datasets to evaluate the compositional generalization of vision-language models. Productivity, on the other hand, has been studied only for specialized tasks [19] or toy domains [27, 36, 59]. CREPE evaluates productivity by using an image-text retrieval task featuring captions of varying compositional complexity." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.449, + 0.473, + 0.736 + ], + "angle": 0, + "content": "Evaluation with hard negatives. Like us, past work evaluating models has commonly designed tasks featuring hard negatives to isolate particular model capabilities while overcoming the limitations of prior evaluation tasks. Using atomic foils that replace an atom in the image or text with a distractor has been the most common strategy [4,9,21,24,51,53,65]. Notably, Park et al. [53] targets verbs and person entities in videos; COVR [4] studies question answering with distractor images; FALSE [51] targets linguistic phenomena such as existence, cardinality and the recognition of actions and spatial relationships. Another strategy has been to swap atoms within a caption to test whether models behave akin to a bag-of-words [1, 51, 75]. In particular, Winoground [75] introduces a set of 800 human edited negatives to evaluate compositionality; it is the closest related work to us. We complement Winoground by scaling it up by three orders of magnitude, by decomposing compositionality into systematicity and productivity, and by studying a variety of different types of hard negatives." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.765, + 0.315, + 0.782 + ], + "angle": 0, + "content": "3. Compositional evaluation" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.795, + 0.471, + 0.901 + ], + "angle": 0, + "content": "The following section builds from the formally vacuous principle of compositionality to a well-defined evaluation scheme [27]. First, we establish the syntax and semantics of the composed language (Section 3.1). Then, we define expected behaviors from a model that achieves comprehension of said language (3.2, 3.3). Finally, we establish how to empirically measure those behaviors via retrieval (3.4)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.092, + 0.862, + 0.108 + ], + "angle": 0, + "content": "3.1. Compositional language of visual concepts" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.114, + 0.895, + 0.281 + ], + "angle": 0, + "content": "To evaluate vision-language models, we find that a compositional language consisting of scene graph visual concepts to be an appropriate foundation [35]. Accordingly, an atom \\( A \\) is defined as a singular visual concept, corresponding to a single scene graph node. Atoms are subtyped into objects \\( A_{o} \\), relationships \\( A_{r} \\), and attributes \\( A_{a} \\). A compound \\( C \\) is defined as a primitive composition of multiple atoms, which corresponds to connections between scene graph nodes. Visual concepts admit two compound types: the attachment of attribute to objects (\"black dog\") \\( C_{ao} \\), and the attachment of two objects via a relationship (\"man hugs child\") \\( C_{oro} \\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.281, + 0.895, + 0.492 + ], + "angle": 0, + "content": "The composition of these compounds form subgraphs \\( S \\), which can be translated to natural language captions \\( T \\). Conversely, captions \\( T \\) derived from image-text datasets \\( \\mathcal{D} \\) can be parsed to become scene graphs \\( S \\). This extensible language is capable of capturing a number of linguistic phenomena identified in existing literature [51, 72], including the existence of concepts (\"a photo with flowers\"), spatial relationships (\"a grill on the left of a staircase\"), action relationships (\"a person throwing a frisbee\"), prepositional attachment (\"A bird with green wings\"), and negation (\"There are no trucks on the road\"). Furthermore, while this study focuses on visual concepts, scene graphs featuring common-sense relationships or other more abstract concepts can be designed; therefore, our methodology is widely applicable [61]." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.5, + 0.64, + 0.516 + ], + "angle": 0, + "content": "3.2. Systematicity" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.523, + 0.895, + 0.826 + ], + "angle": 0, + "content": "With our compositional language in place, we now define two dimensions of compositionality—systematicity and productivity—which we adapt to vision-language representations. Systematicity evaluates a model's ability to systematically recombine seen atoms in compounds. Concretely, let \\(\\mathrm{SEEN}(A,D)\\) denote if an atom is seen in a training dataset \\(\\mathcal{D}\\), namely \\(\\exists (I,S) \\in \\mathcal{D}: A \\in S\\), and \\(\\mathrm{SEEN}(C,D)\\) denote if a compound is seen in a dataset \\(\\mathcal{D}\\), namely \\(\\exists (I,S) \\in \\mathcal{D}: C \\subseteq S\\). To evaluate systematicity, we define three compositional splits: Seen Compounds (SC), Unseen Compounds (UC) and Unseen Atoms (UA). SC is the split where all compounds (and thus all atoms) of every caption have been seen in the training dataset, i.e. \\(\\mathcal{D}_{\\mathrm{SC}} = \\{(I,S) \\in \\mathcal{D}_{test} | \\forall C \\subseteq S : \\mathrm{SEEN}(C,\\mathcal{D}_{train})\\}\\). UC is the split where, for each caption, all atoms have been seen but at least one compound has NOT, i.e. \\(\\mathcal{D}_{UC} = \\{(I,S) \\in \\mathcal{D}_{test} | (\\forall A \\in S : \\mathrm{SEEN}(A,\\mathcal{D}_{train}) \\land (\\exists C \\subseteq S : \\neg \\mathrm{Seen}(C,\\mathcal{D}_{train}))\\}\\). UA is the split where each caption contains at least one atom that has NOT been seen, i.e. \\(\\mathcal{D}_{UA} = \\{(I,S) \\in \\mathcal{D}_{test} | \\exists A \\in S : \\neg \\mathrm{SEEN}(A,\\mathcal{D}_{train})\\}\\)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.833, + 0.634, + 0.849 + ], + "angle": 0, + "content": "3.3. Productivity" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.895, + 0.901 + ], + "angle": 0, + "content": "Productivity refers to a capacity to comprehend an unbounded set of expressions. Since the set of atoms in any dataset is finite, a reasonable substitute for testing unbounded" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "10912" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.09, + 0.086, + 0.878, + 0.266 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.279, + 0.895, + 0.322 + ], + "angle": 0, + "content": "Figure 3. An overview of the productivity retrieval set generation process. By performing random walks on the scene graphs of an evaluation dataset, we generate subgraphs of various complexities. Then, for complexities \\( n \\in \\{4, .5, .12\\} \\) and three hard negative types, we populate the retrieval set \\( \\mathcal{D}_{test}^{HN} \\) by generating a ground truth caption for each \\( n \\)-subgraph and hard negatives for each caption." + }, + { + "type": "table_caption", + "bbox": [ + 0.128, + 0.336, + 0.84, + 0.351 + ], + "angle": 0, + "content": "Table 1. We summarize the sizes of the eight evaluation datasets we create for systematicity and productivity evaluation." + }, + { + "type": "table", + "bbox": [ + 0.09, + 0.361, + 0.891, + 0.436 + ], + "angle": 0, + "content": "
SystematicityProductivity
\\(\\mathcal{D}_{test}^{RAW}\\) (# of image-text pairs)\\(\\mathcal{D}_{test}^{HN}\\) (# of texts)\\(\\mathcal{D}_{test}^{RAW}\\)\\(\\mathcal{D}_{test}^{HN}\\)
Training dataCC-12MYFCC-15MLAION-400MCC-12MYFCC-15MLAION-400MAnyAny
Dataset size385,777385,777373,703325,523316,668309,34217,553278,730
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.463, + 0.474, + 0.659 + ], + "angle": 0, + "content": "comprehension is testing comprehension over increasingly complex scenes. Now, an image \\( I \\) does not have a notion of complexity, since it is theoretically infinitely describable; on the other hand, we can define a notion of complexity for a caption \\( T \\): the number of atoms in its corresponding scene graph \\( |S_T| \\). Therefore, a productive vision-language model should be able to match a given image to the correct corresponding caption, regardless of that caption's complexity. To evaluate productivity, we define a range of productivity complexity (in our case, \\( n = 4, 5, \\ldots, 12 \\)). We need splits of the evaluation dataset based on these complexities, where image-text pairs in a given split have a fixed complexity \\( n \\), and evaluate a model's performance over each split." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.667, + 0.407, + 0.683 + ], + "angle": 0, + "content": "3.4. Compositional evaluation via retrieval" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.69, + 0.472, + 0.856 + ], + "angle": 0, + "content": "We evaluate compositional reasoning using zero-shot image-to-text and text-to-image retrieval. This formulation probes the representation space as directly as possible and is already the most common evaluation method for vision-language foundation models [56]. Theoretically, any existing image-text dataset can be used as retrieval sets for our evaluation. However, one challenging limitation in existing datasets renders the metrics evaluated on them inaccurate. Consider using an image query of a \"plant inside a yellow vase on top of a black television.\" Retrieving unintended alternative positives (e.g. \"a black television\") is not necessarily incor" + }, + { + "type": "page_footnote", + "bbox": [ + 0.076, + 0.863, + 0.471, + 0.902 + ], + "angle": 0, + "content": "2By avoiding captions with redundant objects (\"... a lamb and a lamb and...\" and abstract modifiers (\"there are many lampposts\"), we ensure atom count is tightly coupled with caption complexity." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.463, + 0.895, + 0.69 + ], + "angle": 0, + "content": "rect. Similarly, if no other texts in the retrieval set contain a \"plant\" and a \"television\", retrieving the correct text doesn't suggest that the model comprehends the image. Ideally, to properly evaluate a model, the retrieval dataset should contain hard negatives for every query. A hard negative is a caption that does not faithfully represent the corresponding image, and differs from the ground truth caption by some minimal atomic shift. A example hard negative for the query above is \"man inside a yellow vase on top of a black television.\" By erring in a single, granular syntactic or semantic fashion, hard negatives allow for variations in retrieval performance to be attributable to a specific failure mode of a model's compositional comprehension (see Appendix). We address this need for a new benchmark dataset to evaluate the systematicity and productivity of vision-language models." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.704, + 0.892, + 0.742 + ], + "angle": 0, + "content": "4. CREPE: a large-scale benchmark for vision-language compositionality" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.895, + 0.901 + ], + "angle": 0, + "content": "There are several challenges to creating image-text retrieval datasets that evaluate compositional systematicity and productivity. For systematicity, the primary challenge lies in parsing the training dataset for seen atoms and compounds in order to split the data into the three compositional splits. For productivity, the major challenge is generating imagedataset pairs across different text complexities for the retrieval sets. For both datasets, it is crucial to enumerate different types of hard negatives, and to design an automated hard negative generator which ensures the incorrectness of the" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "10913" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.472, + 0.138 + ], + "angle": 0, + "content": "negatives it generates. We detail our methods for tackling these challenges for future efforts that attempt to create similar benchmarks for other training datasets." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.156, + 0.353, + 0.172 + ], + "angle": 0, + "content": "4.1. Creating systematicity datasets" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.182, + 0.47, + 0.243 + ], + "angle": 0, + "content": "To create the three systematicity splits—SC, SA, UA—we parse a given training dataset \\(\\mathcal{D}\\) into its constituent atoms and compounds, filter low-quality data, and generate hard negatives (Figure 2)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.245, + 0.471, + 0.457 + ], + "angle": 0, + "content": "Parsing a dataset into atoms and compounds Since we utilize the scene graph representation as our compositional language, we use the Stanford Scene Graph Parser [63,77] to parse texts in \\( \\mathcal{D}_{train} \\) into their corresponding scene graphs with objects, attributes and relationships. Since the parser only parses for objects and relationships, we further extract the attributes from the text via spaCy's natural language processing parser by identifying adjective part-of-speech tags. These connected objects, attributes, and relationships constitute our seen atoms and compounds. Similarly, we parse a given \\( \\mathcal{D}_{test} \\) and divide all the image-text pairs into the three splits based on the presence of unseen atoms and/or compounds in the parsed training set. Details on the quality of the scene graph parser can be found in the Appendix." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.459, + 0.472, + 0.567 + ], + "angle": 0, + "content": "Filtering low-quality data We perform the following filtering steps on the image-text pairs in all splits: we only keep region crops which have an area greater than or equal to \\(40K\\) pixels, occupy at least \\(10\\%\\) of the whole image, and whose width-to-height ratio is between 0.5-2.0. We only include text which have at least 2 atoms and 1 compound and de-duplicate text using their corresponding scene graphs." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.569, + 0.473, + 0.902 + ], + "angle": 0, + "content": "Generating hard negatives We introduce two types of hard negatives: HN-ATOM and HN-COMP. HN-ATOM replaces \\(A_{a}\\), \\(A_{o}\\), or \\(A_{r}\\) in the text with an atomic foil. For example, for the caption \"a grill on top of the porch\", one HN-ATOM can be \"a grill underneath the porch\", where the \\(A_{r}\\) \"on top of\" is replaced by \"underneath\". Since captions and scene graphs are not exhaustive, this replacement must be done carefully. For example, if a dog is white and furry, but only \"white\" is annotated, replacing the atom \"white\" with \"furry\" would result in a correct caption. To minimize errors, we employ WordNet [49] to pick replacement atoms that are either antonyms (\"black dog\") or share the same grand-hypernym (\"pink dog\") with respect to the original atom. Furthermore, we use BERT to select the most sensical negatives for each ground truth caption [13,51]. HN-COMP concatenates two compound foils where each contains an atomic foil. For instance, one HN-COMP of the caption \"a pink car\" can be \"a blue car and a pink toy\", where \"blue\" and \"toy\" are the atomic foils in the two compounds foils \"blue car\" and \"pink toy\". We only generate negatives for one-compound examples for systematicity evaluation, as productivity covers complex captions with more atoms." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.092, + 0.771, + 0.108 + ], + "angle": 0, + "content": "4.2. Creating productivity datasets" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.114, + 0.894, + 0.159 + ], + "angle": 0, + "content": "We first generate ground truth captions for scene graphs of varying complexity, filter for data quality, and then generate hard negatives for each example (Figure 3)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.16, + 0.895, + 0.401 + ], + "angle": 0, + "content": "Generating captions We systematically generate captions of different atom counts for each image. Given a scene graph, we perform a random walk of length \\( n \\) through the graph to generate a subgraph. Each subgraph corresponds to a specific region of the image, determined by the union of the bounding boxes of the subgraph atoms. We filter out low-quality regions using the same process as systematicity with additional dedduplication on patches that overlap by \\( \\geq 75\\% \\). For simple subgraphs (\\( n = 4 \\)), we produce captions using handcrafted templates. For larger subgraphs (\\( n \\geq 5 \\)), we leverage GPT-3 [7] (text-davinci-002) to generate captions based on a text description of the scene graph, which lists all objects and relationships. We prompt GPT-3 using 5 manually written captions per complexity, filtering out captions where GPT-3 errs and omits atoms from the subgraph during generation (see more details in Appendix)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.401, + 0.895, + 0.688 + ], + "angle": 0, + "content": "Generating hard negatives For productivity, we employ three hard negatives types (HN-ATOM from systematicity, HN-SWAP, and HN-NEG) corresponding to three hypothesized model error modes. First, as a caption's complexity increases, a model may begin to ignore individual atoms. HN-ATOM randomly selects an atom from the caption and replaces it with an incorrect atom. Second, as a caption's complexity increases, a model may treat captions as \"bags of words\", ignoring syntactic connections built out of word order. A swap hard negative (HN-SWAP) accordingly permutes atoms of the same subtype in a caption. This hard negative is similar to Winoground [75], but in the context of varying caption complexity. On top of Wordnet, we use entailment with RoBERTa to further filter errant HN-SWAP hard negatives [47]. Finally, as a caption's complexity increases, a model may begin to lose comprehension of negations. A negation hard negative (HN-NEG) either negates the entire caption or a specific atom. Refer to the Appendix for details on generating HN-SWAP and HN-NEG." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.697, + 0.765, + 0.712 + ], + "angle": 0, + "content": "4.3. The final benchmark datasets" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.895, + 0.901 + ], + "angle": 0, + "content": "For both productivity and systematicity, we generate two test datasets: \\(\\mathcal{D}_{test}^{HN}\\), which contains image-ground truth text pairs along with all generated hard negatives, and \\(D_{test}^{RAW}\\), which contains only image-ground truth text pairs. To measure the data quality, we randomly sample \\(2\\%\\) of productivity ground truth captions generated by GPT-3 and \\(1\\%\\) of the queries in the productivity and systematicity \\(\\mathcal{D}_{test}^{HN}\\) sets for manual human verification. We assign 2 annotators to each set and measure both generated quality and intra-annotator agreement. \\(87.9\\%\\) of sampled productivity ground truth captions generated by GPT-3 are rated as faithful to the image, with an average pairwise annotator agreement of" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "10914" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.232, + 0.091, + 0.59, + 0.103 + ], + "angle": 0, + "content": "UC vs. SC Recall@1 in Systematicity Hard Negative Retrieval Set" + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.104, + 0.297, + 0.247 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.306, + 0.104, + 0.515, + 0.247 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.524, + 0.104, + 0.885, + 0.248 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.075, + 0.26, + 0.895, + 0.304 + ], + "angle": 0, + "content": "Figure 4. We plot models' recall@1 on the Seen Compounds vs. Unseen Compounds split of the systematicity retrieval set with hard negatives HN-ATOM, HN-COMP and both types. We observe a consistent drop in models' performance from the SC to UC split when the hard negative set consists of both HN-ATOM and HN-COMP or HN-ATOM only." + }, + { + "type": "image", + "bbox": [ + 0.162, + 0.318, + 0.405, + 0.485 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.408, + 0.318, + 0.605, + 0.485 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.61, + 0.319, + 0.811, + 0.485 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.498, + 0.893, + 0.527 + ], + "angle": 0, + "content": "Figure 5. Productivity Analysis. We plot models' Recall@1 on the hard negatives retrieval set against complexity, averaged across all models pretrained on all three training datasets. We find that models' ability to retrieve the ground-truth degrades as complexity increases." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.552, + 0.473, + 0.614 + ], + "angle": 0, + "content": "88.8%. 83.7% of productivity and 86.0% of systematicity hard negatives were rated as genuine negatives (i.e. made factually incorrect statements about the image), with pairwise annotator agreements of 84.3% and 83.7% respectively." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.631, + 0.21, + 0.648 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.658, + 0.47, + 0.839 + ], + "angle": 0, + "content": "We present our experimental setup and results with six takeaways. First, our systematicity experiments show performance decreases consistently on compounds unseen in training. Second, the greatest drop between splits occurs for models trained on LAION-400M. Third, our productivity results reveal models' retrieval performance decays with increasing complexity. Fourth, we find that dataset size has no impact on compositionality. Fifth, we find no clear trend relating model size to compositionality. Finally, models' zero-shot ImageNet classification accuracy correlates with retrieval performance on the systematicity dataset but not systematic generalization to the UC split or productivity." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Datasets. We utilize Visual Genome to create our test datasets. For systematicity, image patches and corresponding spelling-corrected region descriptions are used. We provide three different splits for \\(\\mathcal{D}_{test}^{HN}\\), for three training datasets:" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.552, + 0.896, + 0.903 + ], + "angle": 0, + "content": "CC-12M, YFCC-15M and LAION-400M. For productivity, Visual Genome's image-scene graph pairs are used to create captions and hard negatives for \\(\\mathcal{D}_{test}^{RAW}\\) and \\(\\mathcal{D}_{test}^{HN}\\) (Table 1). Models. We firstly evaluate seven vision-language models pretrained with contrastive loss [70] across three commonly used image-text datasets: Conceptual Captions 12M (CC-12M) [8], a subset of the YFCC100M dataset (YFCC-15M) [56, 74] and LAION-400M [62]. We limit our evaluation to models openly released in the OpenCLIP repository [28] for systematicity evaluation. These include ResNet (RN) [20] and Vision Transformer (ViT) [14] encoders of different sizes: RN50, RN101, ViT-B-16, ViT-B-16-plus-240, ViT-B-32 and ViT-L-14. Additionally, since productivity evaluation is not restricted to models that were trained on publicly released datasets, we conduct productivity evaluation on other foundation vision-language models as well. Specifically, we consider OpenAI's CLIP [56] with ResNet and ViT backbones, CyCLIP [18] (a variant of CLIP introducing auxiliary losses that regularize the gap in similarity scores between mismatched pairs, trained on Conceptual Captions 3M [64] with a ResNet-50 [20] backbone), ALBEF [41] (additionally trained with a masked language modeling and image-text matching loss) and FLAVA [69] (which" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "10915" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.162, + 0.089, + 0.817, + 0.251 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.26, + 0.895, + 0.318 + ], + "angle": 0, + "content": "Figure 6. Productivity Analysis on Additional Foundation Vision-language Models. We plot models' Recall@1 on the productivity hard negatives retrieval set against complexity, where OpenAI CLIP's performance is averaged across five models RN50, RN101, ViT-B-16, ViT-B-32 and ViT-L-14. We find that all models' retrieval performance decreases as complexity increases in both the HN-ATOM and HN-SWAP retrieval sets. For the HN-NEG set, all models except for CLIP either drop in performance or remain at random chance." + }, + { + "type": "image", + "bbox": [ + 0.08, + 0.343, + 0.462, + 0.487 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.5, + 0.472, + 0.556 + ], + "angle": 0, + "content": "Figure 7. A plot showing the correlation between zero-shot top-1 accuracy on ImageNet and Recall@1 on CREPE's systematicity hard-negative sets. We observe a strong correlation, with an \\( R^2 \\) score of 0.9914 for the SC split and 0.9534 for the UC split." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.582, + 0.468, + 0.735 + ], + "angle": 0, + "content": "further adds unimodal losses for image and text domains). Retrieval. For \\(\\mathcal{D}_{test}^{HN}\\), we perform image-to-text retrieval and stratify results by split and hard negative type. For systematicity, the splits are \\(SC\\), \\(UC\\), and \\(UA\\); for productivity, the splits are by caption complexity \\(n\\) (denoted \\(\\mathcal{D}_{test}^{HN,n}\\)). Each retrieval task is between one image and its ground truth caption plus \\(h\\) hard negatives of a single type (see Appendix). We adopt commonly used retrieval metrics Recall@1, 3, 5 and Average Recall@K. For \\(\\mathcal{D}_{test}^{RAW}\\), retrieval experiments are described in the Appendix." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.742, + 0.303, + 0.759 + ], + "angle": 0, + "content": "5.1. Systematicity evaluation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.764, + 0.473, + 0.902 + ], + "angle": 0, + "content": "Model performance on the \\(\\mathcal{D}_{test}^{HN}\\) dataset for systematicity decreases monotonically when compounds are unseen. We first observe a monotonic decrease in recall@1 from the Seen Compounds to the Unseen Compounds split on the systematicity \\(\\mathcal{D}_{test}^{HN}\\) set consisting of both HN-ATOM and HN-COMP (Figure 4 left). This drop is relatively small \\((1 - 5\\%)\\) for the CC-12M and YFCC-15M trained models and the most pronounced for models trained on the largest dataset LAION-400M [62], with the decrease reaching \\(7.9\\%\\)" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.342, + 0.895, + 0.435 + ], + "angle": 0, + "content": "for the ViT-B-32 model. However, CC-12M and YFCC-15M models also significantly underperform LAION-400M models in general, meaning that small drops between sets may be due to overall poor performance rather than improved systematic generalization. In comparison, human oracle experiments generalize with \\(100\\%\\) accuracy to \\(\\mathcal{D}_{test}^{HN}\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.435, + 0.895, + 0.526 + ], + "angle": 0, + "content": "Similar to the overall results, there is also a consistent discrepancy between the SC and UC split on the \\( \\mathcal{D}_{test}^{HN} \\) subset consisting of HN-ATOM only (Figure 4 center). This drop is consistently smaller (3 - 6%) for models trained on CC-12M and YFCC-15M, but pronounced (6% or higher, reaching 9.4% drop for ViT-B-32) for LAION-400M models." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.528, + 0.896, + 0.68 + ], + "angle": 0, + "content": "On the HN-COMP subset (Figure 4 right), we find little to no difference in performance between the \\(SC\\) and \\(UC\\) split. We hypothesize that this is due to the lower difficulty of the HN-COMP hard negatives, as they introduce more foils to the caption, are always longer than the ground truth, and thus offer more opportunities for the model to correctly distinguish the ground truth. This hypothesis is supported by the fact that Recall@1 values on HN-COMP are similar or higher than the ones on HN-ATOM even though the HN-COMP retrieval set size is larger than that of HN-ATOM." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.695, + 0.719, + 0.71 + ], + "angle": 0, + "content": "5.2. Productivity evaluation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.897, + 0.902 + ], + "angle": 0, + "content": "Models' performance decreases with complexity on HN-ATOM and HN-SWAP negatives. At small complexities such as \\( n = 4 \\), we observe that model retrieval quality is well above random chance (Figure 5). However, as caption complexity increases, we observe a steady decrease in performance, nearing random chance for HN-SWAP and dipping below it for HN-ATOM negatives. Similarly, we find that the same downward trend persists for other vision-language foundation models (Figure 6). Importantly, the downward trend occurs for FLAVA and ALBEF even though their training set contains Visual Genome images. We note that for HN-NEG negatives, the OpenAI CLIP models do not adhere" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "10916" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.111, + 0.09, + 0.852, + 0.257 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.271, + 0.895, + 0.315 + ], + "angle": 0, + "content": "Figure 8. A plot showing the correlation between zero-shot top-1 accuracy on ImageNet and Recall@1 on CREPE's productivity hard negative sets for complexities of 4, 8 and 12. Overall, we find no correlation between ImageNet accuracy and Recall@1 on our productivity sets. The strongest correlations are \\( R^2 \\) scores of 0.284 for HN-NEG negatives on \\( n = 12 \\) and of 0.222 for HN-ATOM negatives on \\( n = 8 \\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.34, + 0.473, + 0.461 + ], + "angle": 0, + "content": "to the downward trend, achieving their lowest scores for the lowest complexity. Their performances on higher complexities, however, show great variation. In short, our conclusion is that vision-language foundation models struggle with productivity. Our results on models released by OpenCLIP [28] as well as other vision-language foundation models demonstrate the challenge of differentiating between atomic and swapping foils is exacerbated by caption complexity." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.461, + 0.473, + 0.597 + ], + "angle": 0, + "content": "We see no effect of dataset size on productivity. We do not observe a clear advantage for larger pretraining datasets in our productivity evaluation. For atomic and swapping foils, we see similar performance for models trained on the three datasets, with slightly worse performance on atomic foils for the CC-12M trained models. However, on negation hard negatives (Figure 5), we see variable performance across training sets, with CC-12M models outperforming larger models trained on larger datasets YFCC and LAION." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.605, + 0.265, + 0.62 + ], + "angle": 0, + "content": "5.3. Effect of model size" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.628, + 0.472, + 0.719 + ], + "angle": 0, + "content": "We find no trends relating compositionality to model size. Overall, we note that the LAION trained models (which are both larger models and trained on larger datasets) achieve significantly better absolute performances than smaller models. However, model's systematicity and productivity remain indifferent to the size of the model itself (Figures 4 and 5)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.727, + 0.424, + 0.743 + ], + "angle": 0, + "content": "5.4. Correlation with ImageNet performance" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.473, + 0.903 + ], + "angle": 0, + "content": "We find that zero-shot ImageNet top-1 accuracy strongly correlates with models' Recall@1 on the systematicity retrieval set. Specifically, we acquire \\( R^2 \\) scores of 0.984 and 0.877 for the \\( SC \\) and \\( UC \\) splits respectively (Figure 7). However, this correlation does not imply that models' zero-shot ImageNet performance correlates with systematic generalization, which is instead indicated by small or no difference between the \\( SC \\) and \\( UC \\) splits. On our productivity dataset, we do not observe such a strong correlation, where the highest \\( R^2 \\) score is 0.284 for HN-NEG negatives on a complexity" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.34, + 0.895, + 0.386 + ], + "angle": 0, + "content": "of \\( n = 12 \\) (Figure 8). As such, we can infer that successful zero-shot performance on ImageNet does not necessarily lead to better performance on our productivity sets." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.399, + 0.614, + 0.415 + ], + "angle": 0, + "content": "6. Discussion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.424, + 0.895, + 0.697 + ], + "angle": 0, + "content": "Limitations. First, although our data validation protocols verified our generated hard negatives for productivity as high-quality, approximately \\(70\\%\\) of HN-SWAP and of HN-NEG negatives were rated as correct. While this does not invalidate our key productivity result, this noise is a limitation of CREPE and could hinder future evaluations once foundation models begin performing better. Second, our evaluation only covers a limited set of vision-language foundation models that were trained with contrastive loss. Additionally, given the computational requirements associated with training a foundation model, our experiments centered around model architectures that were already available publicly. We hope that future foundation models are evaluated with our publicly available CREPE benchmark. Third, while we observe text-to-image and image-to-text retrieval to have similar trends for our systematicity experiments, we lack text-to-image datasets with hard negatives. Future work can explore mechanisms to generate counterfactual negative images." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.713, + 0.897, + 0.897 + ], + "angle": 0, + "content": "Conclusion. We present CREPE, a collection of text-to-image and image-to-text retrieval datasets for evaluating pretrained vision-language models' systematicity and productivity. We demonstrate that models struggle with compositionality along both axes, with performance drops across different compositional splits and increasing complexity. We expect that CREPE will provide a more systematic evaluation to benchmark the emergence of compositionality as future models improve. Finally, researchers can leverage our hard-negative generation process to create training batches with hard negatives to incentivize vision-language compositionality." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "10917" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.091, + 0.174, + 0.105 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.116, + 0.473, + 0.212 + ], + "angle": 0, + "content": "[1] Arjun Akula, Spandana Gella, Yaser Al-Onaizan, Song-Chun Zhu, and Siva Reddy. Words aren't enough, their order matters: On the robustness of grounding visual referring expressions. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 6555-6565, Online, July 2020. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.215, + 0.47, + 0.242 + ], + "angle": 0, + "content": "[2] Jacob Andreas, Marcus Rohrbach, Trevor Darrell, and Dan Klein. Neural module networks, 2015." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.245, + 0.472, + 0.313 + ], + "angle": 0, + "content": "[3] Dzmitry Bahdanau, Harm de Vries, Timothy J O'Donnell, Shikhar Murty, Philippe Beaudoin, Yoshua Bengio, and Aaron Courville. Closure: Assessing systematic generalization of clevr models. arXiv preprint arXiv:1912.05783, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.316, + 0.472, + 0.413 + ], + "angle": 0, + "content": "[4] Ben Bogin, Shivanshu Gupta, Matt Gardner, and Jonathan Berant. COVR: A test-bed for visually grounded compositional generalization with real images. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 9824-9846, Online and Punta Cana, Dominican Republic, Nov. 2021. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.415, + 0.472, + 0.484 + ], + "angle": 0, + "content": "[5] Rishi Bommasani, Drew A Hudson, Ehsan Adeli, Russ Altman, Simran Arora, Sydney von Arx, Michael S Bernstein, Jeannette Bohg, Antoine Bosselut, Emma Brunskill, et al. On the opportunities and risks of foundation models. arXiv preprint arXiv:2108.07258, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.486, + 0.472, + 0.514 + ], + "angle": 0, + "content": "[6] Léon Bottou. From machine learning to machine reasoning. Machine learning, 94(2):133-149, 2014." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.515, + 0.472, + 0.584 + ], + "angle": 0, + "content": "[7] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. Advances in neural information processing systems, 33:1877-1901, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.587, + 0.472, + 0.657 + ], + "angle": 0, + "content": "[8] Soravit Changpinyo, Piyush Sharma, Nan Ding, and Radu Soricut. Conceptual 12m: Pushing web-scale image-text pretraining to recognize long-tail visual concepts. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3558-3568, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.658, + 0.472, + 0.699 + ], + "angle": 0, + "content": "[9] Zhenfang Chen, Peng Wang, Lin Ma, Kwan-Yee K. Wong, and Qi Wu. Cops-ref: A new dataset and task on compositional referring expression comprehension, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.701, + 0.472, + 0.742 + ], + "angle": 0, + "content": "[10] Colin Conwell and Tomer Ullman. Testing relational understanding in text-guided image generation. arXiv preprint arXiv:2208.00005, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.745, + 0.373, + 0.759 + ], + "angle": 0, + "content": "[11] MJ Cresswell. Logics and languages. 1973." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.761, + 0.472, + 0.871 + ], + "angle": 0, + "content": "[12] Yuchen Cui, Scott Niekum, Abhinav Gupta, Vikash Kumar, and Aravind Rajeswaran. Can foundation models perform zero-shot task specification for robot manipulation? In Roya Firoozi, Negar Mehr, Esen Yel, Rika Antonova, Jeannette Bohg, Mac Schwager, and Mykel Kochenderfer, editors, Proceedings of The 4th Annual Learning for Dynamics and Control Conference, volume 168 of Proceedings of Machine Learning Research, pages 893-905. PMLR, 23-24 Jun 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.472, + 0.901 + ], + "angle": 0, + "content": "[13] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.116, + 0.473, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.119 + ], + "angle": 0, + "content": "transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.121, + 0.895, + 0.217 + ], + "angle": 0, + "content": "[14] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.218, + 0.892, + 0.259 + ], + "angle": 0, + "content": "[15] Jiali Duan, Liquen Chen, Son Tran, Jinyu Yang, Yi Xu, Belinda Zeng, and Trishul Chilimbi. Multi-modal alignment using representation codebook, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.261, + 0.892, + 0.302 + ], + "angle": 0, + "content": "[16] Sepideh Esmaeilpour, Bing Liu, Eric Robertson, and Lei Shu. Zero-shot out-of-distribution detection based on the pre-trained model clip, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.302, + 0.892, + 0.371 + ], + "angle": 0, + "content": "[17] Mona Gandhi, Mustafa O. Gul, Eva Prakash, Madeleine Grunde-McLaughlin, Ranjay Krishna, and Maneesh Agrawala. Measuring compositional consistency for video question answering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.372, + 0.892, + 0.413 + ], + "angle": 0, + "content": "[18] Shashank Goel, Hritik Bansal, Sumit Bhatia, Ryan A. Rossi, Vishwa Vinay, and Aditya Grover. Cyclip: Cyclic contrastive language-image pretraining, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.414, + 0.892, + 0.481 + ], + "angle": 0, + "content": "[19] Madeleine Grunde-McLaughlin, Ranjay Krishna, and Maneesh Agrawala. Agqa: A benchmark for compositional spatio-temporal reasoning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.484, + 0.892, + 0.538 + ], + "angle": 0, + "content": "[20] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.539, + 0.892, + 0.566 + ], + "angle": 0, + "content": "[21] Lisa Anne Hendricks and Aida Nematzadeh. Probing image-language transformers for verb understanding, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.567, + 0.892, + 0.65 + ], + "angle": 0, + "content": "[22] Jack Hessel, Ari Holtzman, Maxwell Forbes, Ronan Le Bras, and Yejin Choi. CLIPScore: A reference-free evaluation metric for image captioning. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 7514-7528, Online and Punta Cana, Dominican Republic, Nov. 2021. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.651, + 0.892, + 0.719 + ], + "angle": 0, + "content": "[23] Irina Higgins, Nicolas Sonnerat, Loic Matthey, Arka Pal, Christopher P Burgess, Matko Bosnjak, Murray Shanahan, Matthew Botvinick, Demis Hassabis, and Alexander Lerchner. Scan: Learning hierarchical compositional visual concepts. arXiv preprint arXiv:1707.03389, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.72, + 0.892, + 0.761 + ], + "angle": 0, + "content": "[24] Hexiang Hu, Ishan Misra, and Laurens van der Maaten. Evaluating text-to-image matching using binary image selection (bison), 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.761, + 0.892, + 0.79 + ], + "angle": 0, + "content": "[25] Drew A. Hudson and Christopher D. Manning. Compositional attention networks for machine reasoning, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.791, + 0.892, + 0.845 + ], + "angle": 0, + "content": "[26] Drew A Hudson and Christopher D Manning. Gqa: A new dataset for real-world visual reasoning and compositional question answering. Conference on Computer Vision and Pattern Recognition (CVPR), 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.846, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[27] Dieuwke Hupkes, Verna Dankers, Mathijs Mul, and Elia Bruni. Compositionality decomposed: How do neural networks generalise? Journal of Artificial Intelligence Research, 67:757-795, 2020." + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.895, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "10918" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.161 + ], + "angle": 0, + "content": "[28] Gabriel Ilharco, Mitchell Wortsman, Ross Wightman, Cade Gordon, Nicholas Carlini, Rohan Taori, Achal Dave, Vaishaal Shankar, Hongseok Namkoong, John Miller, Hannaneh Hajishirzi, Ali Farhadi, and Ludwig Schmidt. OpenCLIP, 7 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.165, + 0.472, + 0.221 + ], + "angle": 0, + "content": "[29] Ajay Jain, Matthew Tancik, and Pieter Abbeel. Putting nerf on a diet: Semantically consistent few-shot view synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 5885-5894, October 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.223, + 0.472, + 0.264 + ], + "angle": 0, + "content": "[30] Theo MV Janssen and Barbara H Partee. Compositionality. In Handbook of logic and language, pages 417-473. Elsevier, 1997." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.268, + 0.472, + 0.337 + ], + "angle": 0, + "content": "[31] Jingwei Ji, Ranjay Krishna, Li Fei-Fei, and Juan Carlos Niebles. Action genome: Actions as compositions of spatiotemporal scene graphs. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10236–10247, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.34, + 0.472, + 0.422 + ], + "angle": 0, + "content": "[32] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In International Conference on Machine Learning, pages 4904-4916. PMLR, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.426, + 0.472, + 0.481 + ], + "angle": 0, + "content": "[33] Justin Johnson, Bharath Hariharan, Laurens van der Maaten, Li Fei-Fei, C. Lawrence Zitnick, and Ross Girshick. Clevr: A diagnostic dataset for compositional language and elementary visual reasoning, 2016." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.484, + 0.472, + 0.581 + ], + "angle": 0, + "content": "[34] Daniel Keysers, Nathanael Scharli, Nathan Scales, Hylke Buisman, Daniel Furrer, Sergii Kashubin, Nikola Momchev, Danila Sinopalnikov, Lukasz Stafiniak, Tibor Tihon, Dmitry Tsarkov, Xiao Wang, Marc van Zee, and Olivier Bousquet. Measuring compositional generalization: A comprehensive method on realistic data. In International Conference on Learning Representations, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.584, + 0.472, + 0.655 + ], + "angle": 0, + "content": "[35] Ranjay Krishna, Yuke Zhu, Oliver Groth, Justin Johnson, Kenji Hata, Joshua Kravitz, Stephanie Chen, Yannis Kalantidis, Li-Jia Li, David A Shamma, Michael Bernstein, and Li Fei-Fei. Visual genome: Connecting language and vision using crowdsourced dense image annotations. 2016." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.657, + 0.472, + 0.712 + ], + "angle": 0, + "content": "[36] Brenden Lake and Marco Baroni. Generalization without systematicity: On the compositional skills of sequence-to-sequence recurrent networks. In International conference on machine learning, pages 2873-2882. PMLR, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.715, + 0.472, + 0.769 + ], + "angle": 0, + "content": "[37] Brenden M. Lake, Tomer D. Ullman, Joshua B. Tenenbaum, and Samuel J. Gershman. Building machines that learn and think like people. Behavioral and Brain Sciences, 40:e253, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.773, + 0.472, + 0.828 + ], + "angle": 0, + "content": "[38] Boyi Li, Kilian Q Weinberger, Serge Belongie, Vladlen Koltun, and Rene Ranftl. Language-driven semantic segmentation. In International Conference on Learning Representations, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.832, + 0.472, + 0.901 + ], + "angle": 0, + "content": "[39] Dongxu Li, Junnan Li, Hongdong Li, Juan Carlos Niebles, and Steven C.H. Hoi. Align and prompt: Video-and-language pre-training with entity prompts. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4943-4953, 2022." + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.135 + ], + "angle": 0, + "content": "[40] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In ICML, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.137, + 0.894, + 0.219 + ], + "angle": 0, + "content": "[41] Junnan Li, Ramprasaath R. Selvaraju, Akhilesh Deepak Gotmare, Shafiq Joty, Caiming Xiong, and Steven Hoi. Align before fuse: Vision and language representation learning with momentum distillation. In A. Beygelzimer, Y. Dauphin, P. Liang, and J. Wortman Vaughan, editors, Advances in Neural Information Processing Systems, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.221, + 0.894, + 0.304 + ], + "angle": 0, + "content": "[42] Wei Li, Can Gao, Guocheng Niu, Xinyan Xiao, Hao Liu, Jiachen Liu, Hua Wu, and Haifeng Wang. UNIMO-2: End-to-end unified vision-language grounded learning. In Findings of the Association for Computational Linguistics: ACL 2022, pages 3187–3201, Dublin, Ireland, May 2022. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.305, + 0.894, + 0.36 + ], + "angle": 0, + "content": "[43] Yehao Li, Yingwei Pan, Ting Yao, and Tao Mei. Comprehending and ordering semantics for image captioning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.362, + 0.894, + 0.417 + ], + "angle": 0, + "content": "[44] Zhiheng Li, Martin Renqiang Min, Kai Li, and Chenliang Xu. Stylet2i: Toward compositional and high-fidelity text-to-image synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.419, + 0.894, + 0.488 + ], + "angle": 0, + "content": "[45] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In European conference on computer vision, pages 740-755. Springer, 2014." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.49, + 0.894, + 0.545 + ], + "angle": 0, + "content": "[46] Xihui Liu, Dong Huk Park, Samaneh Azadi, Gong Zhang, Arman Chopikyan, Yuxiao Hu, Humphrey Shi, Anna Rohrbach, and Trevor Darrell. More control for free! image synthesis with semantic diffusion guidance, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.547, + 0.894, + 0.614 + ], + "angle": 0, + "content": "[47] Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettle-moyer, and Veselin Stoyanov. Roberta: A robustly optimized bert pretraining approach. arXiv preprint arXiv:1907.11692, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.617, + 0.894, + 0.659 + ], + "angle": 0, + "content": "[48] Haoyu Lu, Nanyi Fei, Yuqi Huo, Yizhao Gao, Zhiwu Lu, and Ji-Rong Wen. Cots: Collaborative two-stream vision-language pre-training model for cross-modal retrieval, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.661, + 0.894, + 0.688 + ], + "angle": 0, + "content": "[49] George A Miller. Wordnet: a lexical database for english. Communications of the ACM, 38(11):39-41, 1995." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.69, + 0.894, + 0.718 + ], + "angle": 0, + "content": "[50] Medhini Narasimhan, Anna Rohrbach, and Trevor Darrell. Clip-it! language-guided video summarization. 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.719, + 0.894, + 0.829 + ], + "angle": 0, + "content": "[51] Letitia Parcalabescu, Michele Cafagna, Lilitta Muradjan, Anette Frank, Iacer Calixto, and Albert Gatt. VALID: A task-independent benchmark for vision and language models centered on linguistic phenomena. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 8253-8280, Dublin, Ireland, May 2022. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.832, + 0.894, + 0.901 + ], + "angle": 0, + "content": "[52] Dong Huk Park, Samaneh Azadi, Xihui Liu, Trevor Darrell, and Anna Rohrbach. Benchmark for compositional text-to-image synthesis. In Thirty-fifth Conference on Neural Information Processing Systems Datasets and Benchmarks Track (Round 1), 2021." + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "10919" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.189 + ], + "angle": 0, + "content": "[53] Jae Sung Park, Sheng Shen, Ali Farhadi, Trevor Darrell, Yejin Choi, and Anna Rohrbach. Exposing the limits of video-text models through contrast sets. In Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 3574-3586, Seattle, United States, July 2022. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.191, + 0.472, + 0.26 + ], + "angle": 0, + "content": "[54] Or Patashnik, Zongze Wu, Eli Shechtman, Daniel Cohen-Or, and Dani Lischinski. Styleclip: Text-driven manipulation of stylegan imagery. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 2085–2094, October 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.263, + 0.472, + 0.332 + ], + "angle": 0, + "content": "[55] Suzanne Petryk, Lisa Dunlap, Keyan Nasseri, Joseph Gonzalez, Trevor Darrell, and Anna Rohrbach. On guiding visual attention with language specification. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 18071-18081, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.334, + 0.472, + 0.416 + ], + "angle": 0, + "content": "[56] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763. PMLR, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.419, + 0.472, + 0.473 + ], + "angle": 0, + "content": "[57] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.476, + 0.472, + 0.546 + ], + "angle": 0, + "content": "[58] Yongming Rao, Wenliang Zhao, Guangyi Chen, Yansong Tang, Zheng Zhu, Guan Huang, Jie Zhou, and Jiwen Lu. Denseclip: Language-guided dense prediction with context-aware prompting. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.548, + 0.472, + 0.63 + ], + "angle": 0, + "content": "[59] Laura Ruis, Jacob Andreas, Marco Baroni, Diane Bouchacourt, and Brenden M. Lake. A benchmark for systematic generalization in grounded language understanding. In Proceedings of the 34th International Conference on Neural Information Processing Systems, NIPS'20, Red Hook, NY, USA, 2020. Curran Associates Inc." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.632, + 0.472, + 0.714 + ], + "angle": 0, + "content": "[60] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily Denton, Seyed Kamyar Seyed Ghasemipour, Burcu Karagol Ayan, S Sara Mahdavi, Rapha Gontijo Lopes, et al. Photorealistic text-to-image diffusion models with deep language understanding. arXiv preprint arXiv:2205.11487, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.717, + 0.472, + 0.785 + ], + "angle": 0, + "content": "[61] Maarten Sap, Ronan Le Bras, Emily Allaway, Chandra Bhagavatula, Nicholas Lourie, Hannah Rashkin, Brendan Roof, Noah A. Smith, and Yejin Choi. ATOMIC: an atlas of machine commonsense for if-then reasoning. CoRR, abs/1811.00146, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.788, + 0.472, + 0.857 + ], + "angle": 0, + "content": "[62] Christoph Schuhmann, Richard Vencu, Romain Beaumont, Robert Kaczmarczyk, Clayton Mullis, Aarush Katta, Theo Coombes, Jenia Jitsev, and Aran Komatsuzaki. Laion-400m: Open dataset of clip-filtered 400 million image-text pairs. arXiv preprint arXiv:2111.02114, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.859, + 0.472, + 0.902 + ], + "angle": 0, + "content": "[63] Sebastian Schuster, Ranjay Krishna, Angel Chang, Li Fei-Fei, and Christopher D Manning. Generating semantically precise scene graphs from textual descriptions for improved image" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "retrieval. In Proceedings of the fourth workshop on vision and language, pages 70-80, 2015." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.121, + 0.894, + 0.176 + ], + "angle": 0, + "content": "[64] Piyush Sharma, Nan Ding, Sebastian Goodman, and Radu Soricut. Conceptual captions: A cleaned, hypernymed, image alt-text dataset for automatic image captioning. In Proceedings of ACL, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.178, + 0.894, + 0.273 + ], + "angle": 0, + "content": "[65] Ravi Shekhar, Sandro Pezzelle, Yauhen Klimovich, Aurélie Herbelot, Moin Nabi, Enver Sangineto, and Raffaella Bernardi. FOIL it! find one mismatch between image and language caption. In Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 255-265, Vancouver, Canada, July 2017. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.274, + 0.894, + 0.328 + ], + "angle": 0, + "content": "[66] Sheng Shen, Liunian Harold Li, Hao Tan, Mohit Bansal, Anna Rohrbach, Kai-Wei Chang, Zhewei Yao, and Kurt Keutzer. How much can CLIP benefit vision-and-language tasks? In International Conference on Learning Representations, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.33, + 0.894, + 0.371 + ], + "angle": 0, + "content": "[67] Hengcan Shi, Munawar Hayat, Yicheng Wu, and Jianfei Cai. Proposalclip: Unsupervised open-category object proposal generation via exploiting clip cues, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.372, + 0.894, + 0.425 + ], + "angle": 0, + "content": "[68] Mohit Shridhar, Lucas Manuelli, and Dieter Fox. CIiport: What and where pathways for robotic manipulation. In Proceedings of the 5th Conference on Robot Learning (CoRL), 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.427, + 0.894, + 0.482 + ], + "angle": 0, + "content": "[69] Amanpreet Singh, Ronghang Hu, Vedanuj Goswami, Guillaume Couairon, Wojciech Galuba, Marcus Rohrbach, and Douwe Kiela. Flava: A foundational language and vision alignment model, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.484, + 0.894, + 0.524 + ], + "angle": 0, + "content": "[70] Kihyuk Sohn. Improved deep metric learning with multiclass n-pair loss objective. Advances in neural information processing systems, 29, 2016." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.525, + 0.894, + 0.58 + ], + "angle": 0, + "content": "[71] Sanjay Subramanian, Will Merrill, Trevor Darrell, Matt Gardner, Sameer Singh, and Anna Rohrbach. Reclip: A strong zero-shot baseline for referring expression comprehension. arXiv preprint arXiv:2204.05991, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.581, + 0.894, + 0.663 + ], + "angle": 0, + "content": "[72] Alane Suhr, Stephanie Zhou, Ally Zhang, Iris Zhang, Huajun Bai, and Yoav Artzi. A corpus for reasoning about natural language grounded in photographs. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 6418-6428, Florence, Italy, July 2019. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.664, + 0.894, + 0.705 + ], + "angle": 0, + "content": "[73] Yoad Tewel, Yoav Shalev, Idan Schwartz, and Lior Wolf. Zero-shot image-to-text generation for visual-semantic arithmetic. arXiv preprint arXiv:2111.14447, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.706, + 0.894, + 0.761 + ], + "angle": 0, + "content": "[74] Bart Thomee, David A Shamma, Gerald Friedland, Benjamin Elizalde, Karl Ni, Douglas Poland, Damian Borth, and Li-Jia Li. Yfcc100m: The new data in multimedia research. Communications of the ACM, 59(2):64-73, 2016." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.762, + 0.894, + 0.844 + ], + "angle": 0, + "content": "[75] Tristan Thrush, Ryan Jiang, Max Bartolo, Amanpreet Singh, Adina Williams, Douwe Kiela, and Candace Ross. Winoground: Probing vision and language models for visio-linguistic compositionality. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5238–5248, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.845, + 0.894, + 0.873 + ], + "angle": 0, + "content": "[76] Mengmeng Wang, Jiazheng Xing, and Yong Liu. Actionclip: A new paradigm for video action recognition, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.874, + 0.894, + 0.901 + ], + "angle": 0, + "content": "[77] Hao Wu, Jiayuan Mao, Yufeng Zhang, Yuning Jiang, Lei Li, Weiwei Sun, and Wei-Ying Ma. Unified visual-semantic" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.894, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "10920" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.11, + 0.092, + 0.47, + 0.147 + ], + "angle": 0, + "content": "embeddings: Bridging vision and language with structured meaning representations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6609-6618, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.15, + 0.472, + 0.26 + ], + "angle": 0, + "content": "[78] Hu Xu, Gargi Ghosh, Po-Yao Huang, Dmytro Okhonko, Armen Aghajanyan, Florian Metze, Luke Zettlemoyer, and Christoph Feichtenhofer. VideoCLIP: Contrastive pretraining for zero-shot video-text understanding. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 6787-6800, Online and Punta Cana, Dominican Republic, Nov. 2021. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.262, + 0.472, + 0.33 + ], + "angle": 0, + "content": "[79] Zipeng Xu, Tianwei Lin, Hao Tang, Fu Li, Dongliang He, Nicu Sebe, Radu Timofte, Luc Van Gool, and Errui Ding. Predict, prevent, and evaluate: Disentangled text-driven image manipulation empowered by pre-trained vision-language model. arXiv preprint arXiv:2111.13333, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.332, + 0.472, + 0.387 + ], + "angle": 0, + "content": "[80] Jinyu Yang, Jiali Duan, Son Tran, Yi Xu, Sampath Chanda, Liquun Chen, Belinda Zeng, Trishul Chilimbi, and Junzhou Huang. Vision-language pre-training with triple contrastive learning. 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.389, + 0.472, + 0.443 + ], + "angle": 0, + "content": "[81] Jiahui Yu, Zirui Wang, Vijay Vasudevan, Legg Yeung, Mojtaba Seyedhosseini, and Yonghui Wu. Coca: Contrastive captioners are image-text foundation models. Transactions on Machine Learning Research, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.446, + 0.472, + 0.527 + ], + "angle": 0, + "content": "[82] Pengchuan Zhang, Xiujun Li, Xiaowei Hu, Jianwei Yang, Lei Zhang, Lijuan Wang, Yejin Choi, and Jianfeng Gao. Vinvl: Revisiting visual representations in vision-language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5579-5588, June 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.53, + 0.472, + 0.585 + ], + "angle": 0, + "content": "[83] Yufan Zhou, Ruiyi Zhang, Changyou Chen, Chunyuan Li, Chris Tensmeyer, Tong Yu, Jiuming Gu, Jinhui Xu, and Tong Sun. Lafite: Towards language-free training for text-to-image generation. arXiv preprint arXiv:2111.13792, 2021." + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.585 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.518, + 0.957 + ], + "angle": 0, + "content": "10921" + } + ] +] \ No newline at end of file diff --git a/2023/CREPE_ Can Vision-Language Foundation Models Reason Compositionally_/6b09bf56-aeff-4b97-8111-a915e425b709_origin.pdf b/2023/CREPE_ Can Vision-Language Foundation Models Reason Compositionally_/6b09bf56-aeff-4b97-8111-a915e425b709_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..ae1e7805ab89550cda2ced921ce09a3058eebab2 --- /dev/null +++ b/2023/CREPE_ Can Vision-Language Foundation Models Reason Compositionally_/6b09bf56-aeff-4b97-8111-a915e425b709_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4ab103120213f13fb5f301914cfbed33d0183b55c3f2ef93fac4ea0475c63621 +size 6461268 diff --git a/2023/CREPE_ Can Vision-Language Foundation Models Reason Compositionally_/full.md b/2023/CREPE_ Can Vision-Language Foundation Models Reason Compositionally_/full.md new file mode 100644 index 0000000000000000000000000000000000000000..d6b36e7a9ae1d071960ed1b222454823afb37f28 --- /dev/null +++ b/2023/CREPE_ Can Vision-Language Foundation Models Reason Compositionally_/full.md @@ -0,0 +1,273 @@ +![](images/a9c996fb36116783728bd08036147b3e393278ffa081301e60aa8db52c0df423.jpg) + +# CREPE: Can Vision-Language Foundation Models Reason Compositionally? + +Zixian Ma $^{1*}$ , Jerry Hong $^{1*}$ , Mustafa Omer Gul $^{2*}$ , Mona Gandhi $^{3}$ , Irenea Gao $^{1}$ , Ranjay Krishna $^{4}$ , Stanford University $^{1}$ , Cornell University $^{2}$ , University of Pennsylvania $^{3}$ , University of Washington $^{4}$ + +{zixianma, jerryhong, irena}@cs.stanford.edu mog29@cornell.edu mona09@seas.upenn.edu ranjay@cs.washington.edu + +# Abstract + +A fundamental characteristic common to both human vision and natural language is their compositional nature. Yet, despite the performance gains contributed by large vision and language pretraining, we find that—across 7 architectures trained with 4 algorithms on massive datasets—they struggle at compositionality. To arrive at this conclusion, we introduce a new compositionality evaluation benchmark, CREPE, which measures two important aspects of compositionality identified by cognitive science literature: systematicity and productivity. To measure systematicity, CREPE consists of a test dataset containing over $370K$ image-text pairs and three different seen-unseen splits. The three splits are designed to test models trained on three popular training datasets: CC-12M, YFCC-15M, and LAION-400M. We also generate $325K$ , $316K$ , and $309K$ hard negative captions for a subset of the pairs. To test productivity, CREPE contains $17K$ image-text pairs with nine different complexities plus $278K$ hard negative captions with atomic, swapping and negation foils. The datasets are generated by repurposing the Visual Genome scene graphs and region descriptions and applying handcrafted templates and GPT-3. For systematicity, we find that model performance decreases consistently when novel compositions dominate the retrieval set, with Recall@1 dropping by up to $9\%$ . For productivity, models' retrieval success decays as complexity increases, frequently nearing random chance at high complexity. These results hold regardless of model and training dataset size. + +# 1. Introduction + +Compositionality, the understanding that "the meaning of the whole is a function of the meanings of its parts" [11], is held to be a key characteristic of human intelligence. In language, the whole is a sentence, made up of words. In vision, the whole is a scene, made up of parts like objects, their attributes, and their relationships [31, 35]. + +![](images/642384e546156cd84a8f39013a8be14bd43d1f63dadd53c9dd7323fe5ade1c91.jpg) +Figure 1. We introduce CREPE, a benchmark to evaluate whether vision-language foundation models demonstrate two fundamental aspects of compositionality: systematicity and productivity. To evaluate systematicity, CREPE utilizes Visual Genome and introduces three new test datasets for the three popular pretraining datasets: CC-12M, YFCC-15M, and LAION-400M. These enable evaluating models' abilities to systematically generalize their understanding to seen compounds, unseen compounds, and even unseen atoms. To evaluate productivity, CREPE introduces examples of nine complexities, with three types of hard negatives for each. + +![](images/264f24940f3c1e1214081c6bd61d27dcfc06cd0d89c6209edb018b689942bb80.jpg) + +Through compositional reasoning, humans can understand new scenes and generate complex sentences by combining known parts [6, 27, 30]. Despite compositionality's importance, there are no large-scale benchmarks directly evaluating whether vision-language models can reason compositionally. These models are pretrained using large-scale image-capture datasets [62, 64, 74], and are already widely applied for tasks that benefit from compositional reasoning, including retrieval, text-to-image generation, and open-vocabulary classification [10, 57, 60]. Especially as such models become ubiquitous "foundations" for other models [5], it is critical to understand their compositional abilities. + +Previous work has evaluated these models using image-text retrieval [32,56,82]. However, the retrieval datasets used either do not provide controlled sets of negatives [45, 74] or study narrow negatives which vary along a single axis + +![](images/78892989831eda5ec895483bb36d684450ec83896c6acbcf645c83f505bc112e.jpg) +Figure 2. An overview of the systematicity retrieval set generation process. First, a model's image-caption training set is parsed to identify what atoms and compounds the model has seen. Then, an evaluation set is divided into three compositional splits according to whether the model has seen all the compounds (Seen Compounds), only all the atoms of the caption (Unseen Compounds), or neither (Unseen Atoms). Finally, hard negative captions HN-ATOM and HN-COMP are generated for the hard negatives retrieval set $\mathcal{D}_{test}^{HN}$ . + +(e.g. permuted word orders or single word substitutions as negative captions) [21, 51, 65, 75]. Further, these analyses have also not studied how retrieval performance varies when generalizing to unseen compositional combinations, or to combinations of increased complexity. + +We introduce CREPE (Compositional REpresentation Evaluation): a new large-scale benchmark to evaluate two aspects of compositionality: systematicity and productivity (Figure 1). Systematicity measures how well a model is able to represent seen versus unseen atoms and their compositions. Productivity studies how well a model can comprehend an unbounded set of increasingly complex expressions. CREPE uses Visual Genome's scene graph representation as the compositionality language [35] and constructs evaluation datasets using its annotations. To test systematicity, we parse the captions in three popular training datasets, CC-12M [8], YFCC-15M [74], and LAION-400M [62], to identify atoms (objects, relations, or attributes) and compounds (combinations of atoms) present in each dataset. For each training set, we curate corresponding test sets containing $385K$ , $385K$ and $373K$ image-text pairs respectively, with splits checking generalization to seen compounds, unseen compounds, and unseen atoms. To test productivity, CREPE contains $17K$ image-text pairs split across nine levels of complexity, as defined by the number of atoms present in the text. Examples across all datasets are paired with various hard negative types to ensure the legitimacy of our conclusions. + +Our experiments—across 7 architectures trained with 4 training algorithms on massive datasets—find that vision-language models struggle at compositionality, with both systematicity and productivity. We present six key findings: first, our systematicity experiments find that models' performance consistently drops between seen and unseen compositions; second, we observe larger drops for models trained on + +LAION-400M (up to a $9\%$ decrease in Recall@1); third, our productivity experiments indicate that retrieval performance degrades with increased caption complexity; fourth, we find no clear trend relating training dataset size to models' compositional reasoning; fifth, model size also has no impact; finally, models' zero-shot ImageNet classification accuracy correlates only with their absolute retrieval performance on the systematicity dataset but not systematic generalization to unseen compounds or to productivity. + +# 2. Related Work + +Our work lies within the field of evaluating foundation models. Specifically, we measure visio-linguistic compositionality. To do so, we create a retrieval benchmark with hard negatives. + +Contrastive Image-Text Pretraining. The recently released contrastively trained CLIP model [56] has catalyzed a wide array of work at the intersection of Computer Vision and Natural Language Processing. Since its release, CLIP has enabled several tasks, ranging from semantic segmentation to image captioning, many of which have remarkable zero-shot capability [12, 16, 38, 56, 71, 73]. CLIP has been used as a loss function within image synthesis applications [29, 44, 46, 54, 79, 83], acted as an automated evaluation metric [22, 52], used successfully as a feature extractor for various vision and language tasks [66], and incorporated into architectures for various tasks including dense prediction and video summarization [43, 50, 55, 58, 67, 68]. This success has also encouraged the design of other contrastive vision and language pretraining algorithms for image [15, 18, 40–42, 48, 69, 80, 81] and video domains [39, 76, 78]. Our work evaluates how well + +such contrastively trained models capture a fundamental property present in human vision and language: compositionality. + +Compositionality. Compositionality allows us to comprehend an infinite number of scenes and utterances [37]. For an AI model, compositionality would not only allow for systematic, combinatorial generalization, but would also confer benefits such as controllability [5]. This promise prompted a wealth of work on both designing [2, 23, 25] and evaluating [17, 19, 26, 36, 72] compositional models. In our work, we focus on two aspects of compositionality: systematicity and productivity. While there is a plethora of benchmarks for systematic generalization within Computer Vision [3, 4, 19, 33] and Machine Learning [34, 36, 59], the subject has been almost unexplored for vision-language models, largely due to lack of benchmarks complementary to the different largescale training datasets. To address this, CREPE provides a benchmark with three different datasets to evaluate the compositional generalization of vision-language models. Productivity, on the other hand, has been studied only for specialized tasks [19] or toy domains [27, 36, 59]. CREPE evaluates productivity by using an image-text retrieval task featuring captions of varying compositional complexity. + +Evaluation with hard negatives. Like us, past work evaluating models has commonly designed tasks featuring hard negatives to isolate particular model capabilities while overcoming the limitations of prior evaluation tasks. Using atomic foils that replace an atom in the image or text with a distractor has been the most common strategy [4,9,21,24,51,53,65]. Notably, Park et al. [53] targets verbs and person entities in videos; COVR [4] studies question answering with distractor images; FALSE [51] targets linguistic phenomena such as existence, cardinality and the recognition of actions and spatial relationships. Another strategy has been to swap atoms within a caption to test whether models behave akin to a bag-of-words [1, 51, 75]. In particular, Winoground [75] introduces a set of 800 human edited negatives to evaluate compositionality; it is the closest related work to us. We complement Winoground by scaling it up by three orders of magnitude, by decomposing compositionality into systematicity and productivity, and by studying a variety of different types of hard negatives. + +# 3. Compositional evaluation + +The following section builds from the formally vacuous principle of compositionality to a well-defined evaluation scheme [27]. First, we establish the syntax and semantics of the composed language (Section 3.1). Then, we define expected behaviors from a model that achieves comprehension of said language (3.2, 3.3). Finally, we establish how to empirically measure those behaviors via retrieval (3.4). + +# 3.1. Compositional language of visual concepts + +To evaluate vision-language models, we find that a compositional language consisting of scene graph visual concepts to be an appropriate foundation [35]. Accordingly, an atom $A$ is defined as a singular visual concept, corresponding to a single scene graph node. Atoms are subtyped into objects $A_{o}$ , relationships $A_{r}$ , and attributes $A_{a}$ . A compound $C$ is defined as a primitive composition of multiple atoms, which corresponds to connections between scene graph nodes. Visual concepts admit two compound types: the attachment of attribute to objects ("black dog") $C_{ao}$ , and the attachment of two objects via a relationship ("man hugs child") $C_{oro}$ . + +The composition of these compounds form subgraphs $S$ , which can be translated to natural language captions $T$ . Conversely, captions $T$ derived from image-text datasets $\mathcal{D}$ can be parsed to become scene graphs $S$ . This extensible language is capable of capturing a number of linguistic phenomena identified in existing literature [51, 72], including the existence of concepts ("a photo with flowers"), spatial relationships ("a grill on the left of a staircase"), action relationships ("a person throwing a frisbee"), prepositional attachment ("A bird with green wings"), and negation ("There are no trucks on the road"). Furthermore, while this study focuses on visual concepts, scene graphs featuring common-sense relationships or other more abstract concepts can be designed; therefore, our methodology is widely applicable [61]. + +# 3.2. Systematicity + +With our compositional language in place, we now define two dimensions of compositionality—systematicity and productivity—which we adapt to vision-language representations. Systematicity evaluates a model's ability to systematically recombine seen atoms in compounds. Concretely, let $\mathrm{SEEN}(A,D)$ denote if an atom is seen in a training dataset $\mathcal{D}$ , namely $\exists (I,S) \in \mathcal{D}: A \in S$ , and $\mathrm{SEEN}(C,D)$ denote if a compound is seen in a dataset $\mathcal{D}$ , namely $\exists (I,S) \in \mathcal{D}: C \subseteq S$ . To evaluate systematicity, we define three compositional splits: Seen Compounds (SC), Unseen Compounds (UC) and Unseen Atoms (UA). SC is the split where all compounds (and thus all atoms) of every caption have been seen in the training dataset, i.e. $\mathcal{D}_{\mathrm{SC}} = \{(I,S) \in \mathcal{D}_{test} | \forall C \subseteq S : \mathrm{SEEN}(C,\mathcal{D}_{train})\}$ . UC is the split where, for each caption, all atoms have been seen but at least one compound has NOT, i.e. $\mathcal{D}_{UC} = \{(I,S) \in \mathcal{D}_{test} | (\forall A \in S : \mathrm{SEEN}(A,\mathcal{D}_{train}) \land (\exists C \subseteq S : \neg \mathrm{Seen}(C,\mathcal{D}_{train}))\}$ . UA is the split where each caption contains at least one atom that has NOT been seen, i.e. $\mathcal{D}_{UA} = \{(I,S) \in \mathcal{D}_{test} | \exists A \in S : \neg \mathrm{SEEN}(A,\mathcal{D}_{train})\}$ . + +# 3.3. Productivity + +Productivity refers to a capacity to comprehend an unbounded set of expressions. Since the set of atoms in any dataset is finite, a reasonable substitute for testing unbounded + +![](images/c58dde3537b407e65e2764fcc4f7e9e0d34d1cd97d4f5d4f67c359b050929c7c.jpg) +Figure 3. An overview of the productivity retrieval set generation process. By performing random walks on the scene graphs of an evaluation dataset, we generate subgraphs of various complexities. Then, for complexities $n \in \{4, .5, .12\}$ and three hard negative types, we populate the retrieval set $\mathcal{D}_{test}^{HN}$ by generating a ground truth caption for each $n$ -subgraph and hard negatives for each caption. + +Table 1. We summarize the sizes of the eight evaluation datasets we create for systematicity and productivity evaluation. + +
SystematicityProductivity
\(\mathcal{D}_{test}^{RAW}\) (# of image-text pairs)\(\mathcal{D}_{test}^{HN}\) (# of texts)\(\mathcal{D}_{test}^{RAW}\)\(\mathcal{D}_{test}^{HN}\)
Training dataCC-12MYFCC-15MLAION-400MCC-12MYFCC-15MLAION-400MAnyAny
Dataset size385,777385,777373,703325,523316,668309,34217,553278,730
+ +comprehension is testing comprehension over increasingly complex scenes. Now, an image $I$ does not have a notion of complexity, since it is theoretically infinitely describable; on the other hand, we can define a notion of complexity for a caption $T$ : the number of atoms in its corresponding scene graph $|S_T|$ . Therefore, a productive vision-language model should be able to match a given image to the correct corresponding caption, regardless of that caption's complexity. To evaluate productivity, we define a range of productivity complexity (in our case, $n = 4, 5, \ldots, 12$ ). We need splits of the evaluation dataset based on these complexities, where image-text pairs in a given split have a fixed complexity $n$ , and evaluate a model's performance over each split. + +# 3.4. Compositional evaluation via retrieval + +We evaluate compositional reasoning using zero-shot image-to-text and text-to-image retrieval. This formulation probes the representation space as directly as possible and is already the most common evaluation method for vision-language foundation models [56]. Theoretically, any existing image-text dataset can be used as retrieval sets for our evaluation. However, one challenging limitation in existing datasets renders the metrics evaluated on them inaccurate. Consider using an image query of a "plant inside a yellow vase on top of a black television." Retrieving unintended alternative positives (e.g. "a black television") is not necessarily incor + +rect. Similarly, if no other texts in the retrieval set contain a "plant" and a "television", retrieving the correct text doesn't suggest that the model comprehends the image. Ideally, to properly evaluate a model, the retrieval dataset should contain hard negatives for every query. A hard negative is a caption that does not faithfully represent the corresponding image, and differs from the ground truth caption by some minimal atomic shift. A example hard negative for the query above is "man inside a yellow vase on top of a black television." By erring in a single, granular syntactic or semantic fashion, hard negatives allow for variations in retrieval performance to be attributable to a specific failure mode of a model's compositional comprehension (see Appendix). We address this need for a new benchmark dataset to evaluate the systematicity and productivity of vision-language models. + +# 4. CREPE: a large-scale benchmark for vision-language compositionality + +There are several challenges to creating image-text retrieval datasets that evaluate compositional systematicity and productivity. For systematicity, the primary challenge lies in parsing the training dataset for seen atoms and compounds in order to split the data into the three compositional splits. For productivity, the major challenge is generating imagedataset pairs across different text complexities for the retrieval sets. For both datasets, it is crucial to enumerate different types of hard negatives, and to design an automated hard negative generator which ensures the incorrectness of the + +negatives it generates. We detail our methods for tackling these challenges for future efforts that attempt to create similar benchmarks for other training datasets. + +# 4.1. Creating systematicity datasets + +To create the three systematicity splits—SC, SA, UA—we parse a given training dataset $\mathcal{D}$ into its constituent atoms and compounds, filter low-quality data, and generate hard negatives (Figure 2). + +Parsing a dataset into atoms and compounds Since we utilize the scene graph representation as our compositional language, we use the Stanford Scene Graph Parser [63,77] to parse texts in $\mathcal{D}_{train}$ into their corresponding scene graphs with objects, attributes and relationships. Since the parser only parses for objects and relationships, we further extract the attributes from the text via spaCy's natural language processing parser by identifying adjective part-of-speech tags. These connected objects, attributes, and relationships constitute our seen atoms and compounds. Similarly, we parse a given $\mathcal{D}_{test}$ and divide all the image-text pairs into the three splits based on the presence of unseen atoms and/or compounds in the parsed training set. Details on the quality of the scene graph parser can be found in the Appendix. + +Filtering low-quality data We perform the following filtering steps on the image-text pairs in all splits: we only keep region crops which have an area greater than or equal to $40K$ pixels, occupy at least $10\%$ of the whole image, and whose width-to-height ratio is between 0.5-2.0. We only include text which have at least 2 atoms and 1 compound and de-duplicate text using their corresponding scene graphs. + +Generating hard negatives We introduce two types of hard negatives: HN-ATOM and HN-COMP. HN-ATOM replaces $A_{a}$ , $A_{o}$ , or $A_{r}$ in the text with an atomic foil. For example, for the caption "a grill on top of the porch", one HN-ATOM can be "a grill underneath the porch", where the $A_{r}$ "on top of" is replaced by "underneath". Since captions and scene graphs are not exhaustive, this replacement must be done carefully. For example, if a dog is white and furry, but only "white" is annotated, replacing the atom "white" with "furry" would result in a correct caption. To minimize errors, we employ WordNet [49] to pick replacement atoms that are either antonyms ("black dog") or share the same grand-hypernym ("pink dog") with respect to the original atom. Furthermore, we use BERT to select the most sensical negatives for each ground truth caption [13,51]. HN-COMP concatenates two compound foils where each contains an atomic foil. For instance, one HN-COMP of the caption "a pink car" can be "a blue car and a pink toy", where "blue" and "toy" are the atomic foils in the two compounds foils "blue car" and "pink toy". We only generate negatives for one-compound examples for systematicity evaluation, as productivity covers complex captions with more atoms. + +# 4.2. Creating productivity datasets + +We first generate ground truth captions for scene graphs of varying complexity, filter for data quality, and then generate hard negatives for each example (Figure 3). + +Generating captions We systematically generate captions of different atom counts for each image. Given a scene graph, we perform a random walk of length $n$ through the graph to generate a subgraph. Each subgraph corresponds to a specific region of the image, determined by the union of the bounding boxes of the subgraph atoms. We filter out low-quality regions using the same process as systematicity with additional dedduplication on patches that overlap by $\geq 75\%$ . For simple subgraphs ( $n = 4$ ), we produce captions using handcrafted templates. For larger subgraphs ( $n \geq 5$ ), we leverage GPT-3 [7] (text-davinci-002) to generate captions based on a text description of the scene graph, which lists all objects and relationships. We prompt GPT-3 using 5 manually written captions per complexity, filtering out captions where GPT-3 errs and omits atoms from the subgraph during generation (see more details in Appendix). + +Generating hard negatives For productivity, we employ three hard negatives types (HN-ATOM from systematicity, HN-SWAP, and HN-NEG) corresponding to three hypothesized model error modes. First, as a caption's complexity increases, a model may begin to ignore individual atoms. HN-ATOM randomly selects an atom from the caption and replaces it with an incorrect atom. Second, as a caption's complexity increases, a model may treat captions as "bags of words", ignoring syntactic connections built out of word order. A swap hard negative (HN-SWAP) accordingly permutes atoms of the same subtype in a caption. This hard negative is similar to Winoground [75], but in the context of varying caption complexity. On top of Wordnet, we use entailment with RoBERTa to further filter errant HN-SWAP hard negatives [47]. Finally, as a caption's complexity increases, a model may begin to lose comprehension of negations. A negation hard negative (HN-NEG) either negates the entire caption or a specific atom. Refer to the Appendix for details on generating HN-SWAP and HN-NEG. + +# 4.3. The final benchmark datasets + +For both productivity and systematicity, we generate two test datasets: $\mathcal{D}_{test}^{HN}$ , which contains image-ground truth text pairs along with all generated hard negatives, and $D_{test}^{RAW}$ , which contains only image-ground truth text pairs. To measure the data quality, we randomly sample $2\%$ of productivity ground truth captions generated by GPT-3 and $1\%$ of the queries in the productivity and systematicity $\mathcal{D}_{test}^{HN}$ sets for manual human verification. We assign 2 annotators to each set and measure both generated quality and intra-annotator agreement. $87.9\%$ of sampled productivity ground truth captions generated by GPT-3 are rated as faithful to the image, with an average pairwise annotator agreement of + +![](images/ed36c6a8819305e748103f78ece790008792765f4fdd142efd834a9cff626ecd.jpg) +UC vs. SC Recall@1 in Systematicity Hard Negative Retrieval Set + +![](images/2621a403d99f578ed9f00e8ee9db686a9d364a27bd02d80ffc9f61299891aeed.jpg) + +![](images/e0a0d07ee20f91e37fadf885a6f76c1e160233bdbd168e786d74f060721bd78f.jpg) +Figure 4. We plot models' recall@1 on the Seen Compounds vs. Unseen Compounds split of the systematicity retrieval set with hard negatives HN-ATOM, HN-COMP and both types. We observe a consistent drop in models' performance from the SC to UC split when the hard negative set consists of both HN-ATOM and HN-COMP or HN-ATOM only. + +![](images/15f88ae173b78505f0baced8169ba753fb9391b0ba48e585b2868b727d2f448a.jpg) +Figure 5. Productivity Analysis. We plot models' Recall@1 on the hard negatives retrieval set against complexity, averaged across all models pretrained on all three training datasets. We find that models' ability to retrieve the ground-truth degrades as complexity increases. + +![](images/828c834819114f045306472fb4615b1148643cfe9e6d839c3b2327547c069ec4.jpg) + +![](images/7d7e86b7d68791de34e7909e6e86fefa9cecd38e116ecd7b3197df03406cc2bc.jpg) + +88.8%. 83.7% of productivity and 86.0% of systematicity hard negatives were rated as genuine negatives (i.e. made factually incorrect statements about the image), with pairwise annotator agreements of 84.3% and 83.7% respectively. + +# 5. Experiments + +We present our experimental setup and results with six takeaways. First, our systematicity experiments show performance decreases consistently on compounds unseen in training. Second, the greatest drop between splits occurs for models trained on LAION-400M. Third, our productivity results reveal models' retrieval performance decays with increasing complexity. Fourth, we find that dataset size has no impact on compositionality. Fifth, we find no clear trend relating model size to compositionality. Finally, models' zero-shot ImageNet classification accuracy correlates with retrieval performance on the systematicity dataset but not systematic generalization to the UC split or productivity. + +Datasets. We utilize Visual Genome to create our test datasets. For systematicity, image patches and corresponding spelling-corrected region descriptions are used. We provide three different splits for $\mathcal{D}_{test}^{HN}$ , for three training datasets: + +CC-12M, YFCC-15M and LAION-400M. For productivity, Visual Genome's image-scene graph pairs are used to create captions and hard negatives for $\mathcal{D}_{test}^{RAW}$ and $\mathcal{D}_{test}^{HN}$ (Table 1). Models. We firstly evaluate seven vision-language models pretrained with contrastive loss [70] across three commonly used image-text datasets: Conceptual Captions 12M (CC-12M) [8], a subset of the YFCC100M dataset (YFCC-15M) [56, 74] and LAION-400M [62]. We limit our evaluation to models openly released in the OpenCLIP repository [28] for systematicity evaluation. These include ResNet (RN) [20] and Vision Transformer (ViT) [14] encoders of different sizes: RN50, RN101, ViT-B-16, ViT-B-16-plus-240, ViT-B-32 and ViT-L-14. Additionally, since productivity evaluation is not restricted to models that were trained on publicly released datasets, we conduct productivity evaluation on other foundation vision-language models as well. Specifically, we consider OpenAI's CLIP [56] with ResNet and ViT backbones, CyCLIP [18] (a variant of CLIP introducing auxiliary losses that regularize the gap in similarity scores between mismatched pairs, trained on Conceptual Captions 3M [64] with a ResNet-50 [20] backbone), ALBEF [41] (additionally trained with a masked language modeling and image-text matching loss) and FLAVA [69] (which + +![](images/e5a2090aee51b8259ee05d7ddcbd12458a75b88f96749bb4501142027638bf9a.jpg) +Figure 6. Productivity Analysis on Additional Foundation Vision-language Models. We plot models' Recall@1 on the productivity hard negatives retrieval set against complexity, where OpenAI CLIP's performance is averaged across five models RN50, RN101, ViT-B-16, ViT-B-32 and ViT-L-14. We find that all models' retrieval performance decreases as complexity increases in both the HN-ATOM and HN-SWAP retrieval sets. For the HN-NEG set, all models except for CLIP either drop in performance or remain at random chance. + +![](images/aae2bbb24f2aa7dfb837046d5a26df655857613ca4224f7bc0e1a7b78debef63.jpg) +Figure 7. A plot showing the correlation between zero-shot top-1 accuracy on ImageNet and Recall@1 on CREPE's systematicity hard-negative sets. We observe a strong correlation, with an $R^2$ score of 0.9914 for the SC split and 0.9534 for the UC split. + +further adds unimodal losses for image and text domains). Retrieval. For $\mathcal{D}_{test}^{HN}$ , we perform image-to-text retrieval and stratify results by split and hard negative type. For systematicity, the splits are $SC$ , $UC$ , and $UA$ ; for productivity, the splits are by caption complexity $n$ (denoted $\mathcal{D}_{test}^{HN,n}$ ). Each retrieval task is between one image and its ground truth caption plus $h$ hard negatives of a single type (see Appendix). We adopt commonly used retrieval metrics Recall@1, 3, 5 and Average Recall@K. For $\mathcal{D}_{test}^{RAW}$ , retrieval experiments are described in the Appendix. + +# 5.1. Systematicity evaluation + +Model performance on the $\mathcal{D}_{test}^{HN}$ dataset for systematicity decreases monotonically when compounds are unseen. We first observe a monotonic decrease in recall@1 from the Seen Compounds to the Unseen Compounds split on the systematicity $\mathcal{D}_{test}^{HN}$ set consisting of both HN-ATOM and HN-COMP (Figure 4 left). This drop is relatively small $(1 - 5\%)$ for the CC-12M and YFCC-15M trained models and the most pronounced for models trained on the largest dataset LAION-400M [62], with the decrease reaching $7.9\%$ + +for the ViT-B-32 model. However, CC-12M and YFCC-15M models also significantly underperform LAION-400M models in general, meaning that small drops between sets may be due to overall poor performance rather than improved systematic generalization. In comparison, human oracle experiments generalize with $100\%$ accuracy to $\mathcal{D}_{test}^{HN}$ . + +Similar to the overall results, there is also a consistent discrepancy between the SC and UC split on the $\mathcal{D}_{test}^{HN}$ subset consisting of HN-ATOM only (Figure 4 center). This drop is consistently smaller (3 - 6%) for models trained on CC-12M and YFCC-15M, but pronounced (6% or higher, reaching 9.4% drop for ViT-B-32) for LAION-400M models. + +On the HN-COMP subset (Figure 4 right), we find little to no difference in performance between the $SC$ and $UC$ split. We hypothesize that this is due to the lower difficulty of the HN-COMP hard negatives, as they introduce more foils to the caption, are always longer than the ground truth, and thus offer more opportunities for the model to correctly distinguish the ground truth. This hypothesis is supported by the fact that Recall@1 values on HN-COMP are similar or higher than the ones on HN-ATOM even though the HN-COMP retrieval set size is larger than that of HN-ATOM. + +# 5.2. Productivity evaluation + +Models' performance decreases with complexity on HN-ATOM and HN-SWAP negatives. At small complexities such as $n = 4$ , we observe that model retrieval quality is well above random chance (Figure 5). However, as caption complexity increases, we observe a steady decrease in performance, nearing random chance for HN-SWAP and dipping below it for HN-ATOM negatives. Similarly, we find that the same downward trend persists for other vision-language foundation models (Figure 6). Importantly, the downward trend occurs for FLAVA and ALBEF even though their training set contains Visual Genome images. We note that for HN-NEG negatives, the OpenAI CLIP models do not adhere + +![](images/39d5d0840bc78533ebbca857036d61b7ca5b8948ee453f276cbe05f9bda4d0ca.jpg) +Figure 8. A plot showing the correlation between zero-shot top-1 accuracy on ImageNet and Recall@1 on CREPE's productivity hard negative sets for complexities of 4, 8 and 12. Overall, we find no correlation between ImageNet accuracy and Recall@1 on our productivity sets. The strongest correlations are $R^2$ scores of 0.284 for HN-NEG negatives on $n = 12$ and of 0.222 for HN-ATOM negatives on $n = 8$ . + +to the downward trend, achieving their lowest scores for the lowest complexity. Their performances on higher complexities, however, show great variation. In short, our conclusion is that vision-language foundation models struggle with productivity. Our results on models released by OpenCLIP [28] as well as other vision-language foundation models demonstrate the challenge of differentiating between atomic and swapping foils is exacerbated by caption complexity. + +We see no effect of dataset size on productivity. We do not observe a clear advantage for larger pretraining datasets in our productivity evaluation. For atomic and swapping foils, we see similar performance for models trained on the three datasets, with slightly worse performance on atomic foils for the CC-12M trained models. However, on negation hard negatives (Figure 5), we see variable performance across training sets, with CC-12M models outperforming larger models trained on larger datasets YFCC and LAION. + +# 5.3. Effect of model size + +We find no trends relating compositionality to model size. Overall, we note that the LAION trained models (which are both larger models and trained on larger datasets) achieve significantly better absolute performances than smaller models. However, model's systematicity and productivity remain indifferent to the size of the model itself (Figures 4 and 5). + +# 5.4. Correlation with ImageNet performance + +We find that zero-shot ImageNet top-1 accuracy strongly correlates with models' Recall@1 on the systematicity retrieval set. Specifically, we acquire $R^2$ scores of 0.984 and 0.877 for the $SC$ and $UC$ splits respectively (Figure 7). However, this correlation does not imply that models' zero-shot ImageNet performance correlates with systematic generalization, which is instead indicated by small or no difference between the $SC$ and $UC$ splits. On our productivity dataset, we do not observe such a strong correlation, where the highest $R^2$ score is 0.284 for HN-NEG negatives on a complexity + +of $n = 12$ (Figure 8). As such, we can infer that successful zero-shot performance on ImageNet does not necessarily lead to better performance on our productivity sets. + +# 6. Discussion + +Limitations. First, although our data validation protocols verified our generated hard negatives for productivity as high-quality, approximately $70\%$ of HN-SWAP and of HN-NEG negatives were rated as correct. While this does not invalidate our key productivity result, this noise is a limitation of CREPE and could hinder future evaluations once foundation models begin performing better. Second, our evaluation only covers a limited set of vision-language foundation models that were trained with contrastive loss. Additionally, given the computational requirements associated with training a foundation model, our experiments centered around model architectures that were already available publicly. We hope that future foundation models are evaluated with our publicly available CREPE benchmark. Third, while we observe text-to-image and image-to-text retrieval to have similar trends for our systematicity experiments, we lack text-to-image datasets with hard negatives. Future work can explore mechanisms to generate counterfactual negative images. + +Conclusion. We present CREPE, a collection of text-to-image and image-to-text retrieval datasets for evaluating pretrained vision-language models' systematicity and productivity. We demonstrate that models struggle with compositionality along both axes, with performance drops across different compositional splits and increasing complexity. We expect that CREPE will provide a more systematic evaluation to benchmark the emergence of compositionality as future models improve. Finally, researchers can leverage our hard-negative generation process to create training batches with hard negatives to incentivize vision-language compositionality. + +# References + +[1] Arjun Akula, Spandana Gella, Yaser Al-Onaizan, Song-Chun Zhu, and Siva Reddy. Words aren't enough, their order matters: On the robustness of grounding visual referring expressions. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 6555-6565, Online, July 2020. Association for Computational Linguistics. +[2] Jacob Andreas, Marcus Rohrbach, Trevor Darrell, and Dan Klein. Neural module networks, 2015. +[3] Dzmitry Bahdanau, Harm de Vries, Timothy J O'Donnell, Shikhar Murty, Philippe Beaudoin, Yoshua Bengio, and Aaron Courville. Closure: Assessing systematic generalization of clevr models. arXiv preprint arXiv:1912.05783, 2019. +[4] Ben Bogin, Shivanshu Gupta, Matt Gardner, and Jonathan Berant. COVR: A test-bed for visually grounded compositional generalization with real images. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 9824-9846, Online and Punta Cana, Dominican Republic, Nov. 2021. Association for Computational Linguistics. +[5] Rishi Bommasani, Drew A Hudson, Ehsan Adeli, Russ Altman, Simran Arora, Sydney von Arx, Michael S Bernstein, Jeannette Bohg, Antoine Bosselut, Emma Brunskill, et al. On the opportunities and risks of foundation models. arXiv preprint arXiv:2108.07258, 2021. +[6] Léon Bottou. From machine learning to machine reasoning. Machine learning, 94(2):133-149, 2014. +[7] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. Advances in neural information processing systems, 33:1877-1901, 2020. +[8] Soravit Changpinyo, Piyush Sharma, Nan Ding, and Radu Soricut. Conceptual 12m: Pushing web-scale image-text pretraining to recognize long-tail visual concepts. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3558-3568, 2021. +[9] Zhenfang Chen, Peng Wang, Lin Ma, Kwan-Yee K. Wong, and Qi Wu. Cops-ref: A new dataset and task on compositional referring expression comprehension, 2020. +[10] Colin Conwell and Tomer Ullman. Testing relational understanding in text-guided image generation. arXiv preprint arXiv:2208.00005, 2022. +[11] MJ Cresswell. Logics and languages. 1973. +[12] Yuchen Cui, Scott Niekum, Abhinav Gupta, Vikash Kumar, and Aravind Rajeswaran. Can foundation models perform zero-shot task specification for robot manipulation? In Roya Firoozi, Negar Mehr, Esen Yel, Rika Antonova, Jeannette Bohg, Mac Schwager, and Mykel Kochenderfer, editors, Proceedings of The 4th Annual Learning for Dynamics and Control Conference, volume 168 of Proceedings of Machine Learning Research, pages 893-905. PMLR, 23-24 Jun 2022. +[13] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional + +transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018. +[14] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations, 2021. +[15] Jiali Duan, Liquen Chen, Son Tran, Jinyu Yang, Yi Xu, Belinda Zeng, and Trishul Chilimbi. Multi-modal alignment using representation codebook, 2022. +[16] Sepideh Esmaeilpour, Bing Liu, Eric Robertson, and Lei Shu. Zero-shot out-of-distribution detection based on the pre-trained model clip, 2021. +[17] Mona Gandhi, Mustafa O. Gul, Eva Prakash, Madeleine Grunde-McLaughlin, Ranjay Krishna, and Maneesh Agrawala. Measuring compositional consistency for video question answering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022. +[18] Shashank Goel, Hritik Bansal, Sumit Bhatia, Ryan A. Rossi, Vishwa Vinay, and Aditya Grover. Cyclip: Cyclic contrastive language-image pretraining, 2022. +[19] Madeleine Grunde-McLaughlin, Ranjay Krishna, and Maneesh Agrawala. Agqa: A benchmark for compositional spatio-temporal reasoning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2021. +[20] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. +[21] Lisa Anne Hendricks and Aida Nematzadeh. Probing image-language transformers for verb understanding, 2021. +[22] Jack Hessel, Ari Holtzman, Maxwell Forbes, Ronan Le Bras, and Yejin Choi. CLIPScore: A reference-free evaluation metric for image captioning. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 7514-7528, Online and Punta Cana, Dominican Republic, Nov. 2021. Association for Computational Linguistics. +[23] Irina Higgins, Nicolas Sonnerat, Loic Matthey, Arka Pal, Christopher P Burgess, Matko Bosnjak, Murray Shanahan, Matthew Botvinick, Demis Hassabis, and Alexander Lerchner. Scan: Learning hierarchical compositional visual concepts. arXiv preprint arXiv:1707.03389, 2017. +[24] Hexiang Hu, Ishan Misra, and Laurens van der Maaten. Evaluating text-to-image matching using binary image selection (bison), 2019. +[25] Drew A. Hudson and Christopher D. Manning. Compositional attention networks for machine reasoning, 2018. +[26] Drew A Hudson and Christopher D Manning. Gqa: A new dataset for real-world visual reasoning and compositional question answering. Conference on Computer Vision and Pattern Recognition (CVPR), 2019. +[27] Dieuwke Hupkes, Verna Dankers, Mathijs Mul, and Elia Bruni. Compositionality decomposed: How do neural networks generalise? Journal of Artificial Intelligence Research, 67:757-795, 2020. + +[28] Gabriel Ilharco, Mitchell Wortsman, Ross Wightman, Cade Gordon, Nicholas Carlini, Rohan Taori, Achal Dave, Vaishaal Shankar, Hongseok Namkoong, John Miller, Hannaneh Hajishirzi, Ali Farhadi, and Ludwig Schmidt. OpenCLIP, 7 2021. +[29] Ajay Jain, Matthew Tancik, and Pieter Abbeel. Putting nerf on a diet: Semantically consistent few-shot view synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 5885-5894, October 2021. +[30] Theo MV Janssen and Barbara H Partee. Compositionality. In Handbook of logic and language, pages 417-473. Elsevier, 1997. +[31] Jingwei Ji, Ranjay Krishna, Li Fei-Fei, and Juan Carlos Niebles. Action genome: Actions as compositions of spatiotemporal scene graphs. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10236–10247, 2020. +[32] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In International Conference on Machine Learning, pages 4904-4916. PMLR, 2021. +[33] Justin Johnson, Bharath Hariharan, Laurens van der Maaten, Li Fei-Fei, C. Lawrence Zitnick, and Ross Girshick. Clevr: A diagnostic dataset for compositional language and elementary visual reasoning, 2016. +[34] Daniel Keysers, Nathanael Scharli, Nathan Scales, Hylke Buisman, Daniel Furrer, Sergii Kashubin, Nikola Momchev, Danila Sinopalnikov, Lukasz Stafiniak, Tibor Tihon, Dmitry Tsarkov, Xiao Wang, Marc van Zee, and Olivier Bousquet. Measuring compositional generalization: A comprehensive method on realistic data. In International Conference on Learning Representations, 2020. +[35] Ranjay Krishna, Yuke Zhu, Oliver Groth, Justin Johnson, Kenji Hata, Joshua Kravitz, Stephanie Chen, Yannis Kalantidis, Li-Jia Li, David A Shamma, Michael Bernstein, and Li Fei-Fei. Visual genome: Connecting language and vision using crowdsourced dense image annotations. 2016. +[36] Brenden Lake and Marco Baroni. Generalization without systematicity: On the compositional skills of sequence-to-sequence recurrent networks. In International conference on machine learning, pages 2873-2882. PMLR, 2018. +[37] Brenden M. Lake, Tomer D. Ullman, Joshua B. Tenenbaum, and Samuel J. Gershman. Building machines that learn and think like people. Behavioral and Brain Sciences, 40:e253, 2017. +[38] Boyi Li, Kilian Q Weinberger, Serge Belongie, Vladlen Koltun, and Rene Ranftl. Language-driven semantic segmentation. In International Conference on Learning Representations, 2022. +[39] Dongxu Li, Junnan Li, Hongdong Li, Juan Carlos Niebles, and Steven C.H. Hoi. Align and prompt: Video-and-language pre-training with entity prompts. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4943-4953, 2022. + +[40] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In ICML, 2022. +[41] Junnan Li, Ramprasaath R. Selvaraju, Akhilesh Deepak Gotmare, Shafiq Joty, Caiming Xiong, and Steven Hoi. Align before fuse: Vision and language representation learning with momentum distillation. In A. Beygelzimer, Y. Dauphin, P. Liang, and J. Wortman Vaughan, editors, Advances in Neural Information Processing Systems, 2021. +[42] Wei Li, Can Gao, Guocheng Niu, Xinyan Xiao, Hao Liu, Jiachen Liu, Hua Wu, and Haifeng Wang. UNIMO-2: End-to-end unified vision-language grounded learning. In Findings of the Association for Computational Linguistics: ACL 2022, pages 3187–3201, Dublin, Ireland, May 2022. Association for Computational Linguistics. +[43] Yehao Li, Yingwei Pan, Ting Yao, and Tao Mei. Comprehending and ordering semantics for image captioning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022. +[44] Zhiheng Li, Martin Renqiang Min, Kai Li, and Chenliang Xu. Stylet2i: Toward compositional and high-fidelity text-to-image synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. +[45] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In European conference on computer vision, pages 740-755. Springer, 2014. +[46] Xihui Liu, Dong Huk Park, Samaneh Azadi, Gong Zhang, Arman Chopikyan, Yuxiao Hu, Humphrey Shi, Anna Rohrbach, and Trevor Darrell. More control for free! image synthesis with semantic diffusion guidance, 2021. +[47] Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettle-moyer, and Veselin Stoyanov. Roberta: A robustly optimized bert pretraining approach. arXiv preprint arXiv:1907.11692, 2019. +[48] Haoyu Lu, Nanyi Fei, Yuqi Huo, Yizhao Gao, Zhiwu Lu, and Ji-Rong Wen. Cots: Collaborative two-stream vision-language pre-training model for cross-modal retrieval, 2022. +[49] George A Miller. Wordnet: a lexical database for english. Communications of the ACM, 38(11):39-41, 1995. +[50] Medhini Narasimhan, Anna Rohrbach, and Trevor Darrell. Clip-it! language-guided video summarization. 2021. +[51] Letitia Parcalabescu, Michele Cafagna, Lilitta Muradjan, Anette Frank, Iacer Calixto, and Albert Gatt. VALID: A task-independent benchmark for vision and language models centered on linguistic phenomena. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 8253-8280, Dublin, Ireland, May 2022. Association for Computational Linguistics. +[52] Dong Huk Park, Samaneh Azadi, Xihui Liu, Trevor Darrell, and Anna Rohrbach. Benchmark for compositional text-to-image synthesis. In Thirty-fifth Conference on Neural Information Processing Systems Datasets and Benchmarks Track (Round 1), 2021. + +[53] Jae Sung Park, Sheng Shen, Ali Farhadi, Trevor Darrell, Yejin Choi, and Anna Rohrbach. Exposing the limits of video-text models through contrast sets. In Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 3574-3586, Seattle, United States, July 2022. Association for Computational Linguistics. +[54] Or Patashnik, Zongze Wu, Eli Shechtman, Daniel Cohen-Or, and Dani Lischinski. Styleclip: Text-driven manipulation of stylegan imagery. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 2085–2094, October 2021. +[55] Suzanne Petryk, Lisa Dunlap, Keyan Nasseri, Joseph Gonzalez, Trevor Darrell, and Anna Rohrbach. On guiding visual attention with language specification. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 18071-18081, 2022. +[56] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. +[57] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 2022. +[58] Yongming Rao, Wenliang Zhao, Guangyi Chen, Yansong Tang, Zheng Zhu, Guan Huang, Jie Zhou, and Jiwen Lu. Denseclip: Language-guided dense prediction with context-aware prompting. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. +[59] Laura Ruis, Jacob Andreas, Marco Baroni, Diane Bouchacourt, and Brenden M. Lake. A benchmark for systematic generalization in grounded language understanding. In Proceedings of the 34th International Conference on Neural Information Processing Systems, NIPS'20, Red Hook, NY, USA, 2020. Curran Associates Inc. +[60] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily Denton, Seyed Kamyar Seyed Ghasemipour, Burcu Karagol Ayan, S Sara Mahdavi, Rapha Gontijo Lopes, et al. Photorealistic text-to-image diffusion models with deep language understanding. arXiv preprint arXiv:2205.11487, 2022. +[61] Maarten Sap, Ronan Le Bras, Emily Allaway, Chandra Bhagavatula, Nicholas Lourie, Hannah Rashkin, Brendan Roof, Noah A. Smith, and Yejin Choi. ATOMIC: an atlas of machine commonsense for if-then reasoning. CoRR, abs/1811.00146, 2018. +[62] Christoph Schuhmann, Richard Vencu, Romain Beaumont, Robert Kaczmarczyk, Clayton Mullis, Aarush Katta, Theo Coombes, Jenia Jitsev, and Aran Komatsuzaki. Laion-400m: Open dataset of clip-filtered 400 million image-text pairs. arXiv preprint arXiv:2111.02114, 2021. +[63] Sebastian Schuster, Ranjay Krishna, Angel Chang, Li Fei-Fei, and Christopher D Manning. Generating semantically precise scene graphs from textual descriptions for improved image + +retrieval. In Proceedings of the fourth workshop on vision and language, pages 70-80, 2015. +[64] Piyush Sharma, Nan Ding, Sebastian Goodman, and Radu Soricut. Conceptual captions: A cleaned, hypernymed, image alt-text dataset for automatic image captioning. In Proceedings of ACL, 2018. +[65] Ravi Shekhar, Sandro Pezzelle, Yauhen Klimovich, Aurélie Herbelot, Moin Nabi, Enver Sangineto, and Raffaella Bernardi. FOIL it! find one mismatch between image and language caption. In Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 255-265, Vancouver, Canada, July 2017. Association for Computational Linguistics. +[66] Sheng Shen, Liunian Harold Li, Hao Tan, Mohit Bansal, Anna Rohrbach, Kai-Wei Chang, Zhewei Yao, and Kurt Keutzer. How much can CLIP benefit vision-and-language tasks? In International Conference on Learning Representations, 2022. +[67] Hengcan Shi, Munawar Hayat, Yicheng Wu, and Jianfei Cai. Proposalclip: Unsupervised open-category object proposal generation via exploiting clip cues, 2022. +[68] Mohit Shridhar, Lucas Manuelli, and Dieter Fox. CIiport: What and where pathways for robotic manipulation. In Proceedings of the 5th Conference on Robot Learning (CoRL), 2021. +[69] Amanpreet Singh, Ronghang Hu, Vedanuj Goswami, Guillaume Couairon, Wojciech Galuba, Marcus Rohrbach, and Douwe Kiela. Flava: A foundational language and vision alignment model, 2021. +[70] Kihyuk Sohn. Improved deep metric learning with multiclass n-pair loss objective. Advances in neural information processing systems, 29, 2016. +[71] Sanjay Subramanian, Will Merrill, Trevor Darrell, Matt Gardner, Sameer Singh, and Anna Rohrbach. Reclip: A strong zero-shot baseline for referring expression comprehension. arXiv preprint arXiv:2204.05991, 2022. +[72] Alane Suhr, Stephanie Zhou, Ally Zhang, Iris Zhang, Huajun Bai, and Yoav Artzi. A corpus for reasoning about natural language grounded in photographs. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 6418-6428, Florence, Italy, July 2019. Association for Computational Linguistics. +[73] Yoad Tewel, Yoav Shalev, Idan Schwartz, and Lior Wolf. Zero-shot image-to-text generation for visual-semantic arithmetic. arXiv preprint arXiv:2111.14447, 2021. +[74] Bart Thomee, David A Shamma, Gerald Friedland, Benjamin Elizalde, Karl Ni, Douglas Poland, Damian Borth, and Li-Jia Li. Yfcc100m: The new data in multimedia research. Communications of the ACM, 59(2):64-73, 2016. +[75] Tristan Thrush, Ryan Jiang, Max Bartolo, Amanpreet Singh, Adina Williams, Douwe Kiela, and Candace Ross. Winoground: Probing vision and language models for visio-linguistic compositionality. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5238–5248, 2022. +[76] Mengmeng Wang, Jiazheng Xing, and Yong Liu. Actionclip: A new paradigm for video action recognition, 2021. +[77] Hao Wu, Jiayuan Mao, Yufeng Zhang, Yuning Jiang, Lei Li, Weiwei Sun, and Wei-Ying Ma. Unified visual-semantic + +embeddings: Bridging vision and language with structured meaning representations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6609-6618, 2019. +[78] Hu Xu, Gargi Ghosh, Po-Yao Huang, Dmytro Okhonko, Armen Aghajanyan, Florian Metze, Luke Zettlemoyer, and Christoph Feichtenhofer. VideoCLIP: Contrastive pretraining for zero-shot video-text understanding. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 6787-6800, Online and Punta Cana, Dominican Republic, Nov. 2021. Association for Computational Linguistics. +[79] Zipeng Xu, Tianwei Lin, Hao Tang, Fu Li, Dongliang He, Nicu Sebe, Radu Timofte, Luc Van Gool, and Errui Ding. Predict, prevent, and evaluate: Disentangled text-driven image manipulation empowered by pre-trained vision-language model. arXiv preprint arXiv:2111.13333, 2021. +[80] Jinyu Yang, Jiali Duan, Son Tran, Yi Xu, Sampath Chanda, Liquun Chen, Belinda Zeng, Trishul Chilimbi, and Junzhou Huang. Vision-language pre-training with triple contrastive learning. 2022. +[81] Jiahui Yu, Zirui Wang, Vijay Vasudevan, Legg Yeung, Mojtaba Seyedhosseini, and Yonghui Wu. Coca: Contrastive captioners are image-text foundation models. Transactions on Machine Learning Research, 2022. +[82] Pengchuan Zhang, Xiujun Li, Xiaowei Hu, Jianwei Yang, Lei Zhang, Lijuan Wang, Yejin Choi, and Jianfeng Gao. Vinvl: Revisiting visual representations in vision-language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5579-5588, June 2021. +[83] Yufan Zhou, Ruiyi Zhang, Changyou Chen, Chunyuan Li, Chris Tensmeyer, Tong Yu, Jiuming Gu, Jinhui Xu, and Tong Sun. Lafite: Towards language-free training for text-to-image generation. arXiv preprint arXiv:2111.13792, 2021. \ No newline at end of file diff --git a/2023/CREPE_ Can Vision-Language Foundation Models Reason Compositionally_/images.zip b/2023/CREPE_ Can Vision-Language Foundation Models Reason Compositionally_/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..a6ff20334b149eedb7387a781ae27ec1c1cb3c4b --- /dev/null +++ b/2023/CREPE_ Can Vision-Language Foundation Models Reason Compositionally_/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:15ddbbc8d6b8abc4ebef42f3c31a760336ac12bdb137429af016ac7fde24da2d +size 410542 diff --git a/2023/CREPE_ Can Vision-Language Foundation Models Reason Compositionally_/layout.json b/2023/CREPE_ Can Vision-Language Foundation Models Reason Compositionally_/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..1c692fc148fc0a1ac48269ee741d74d6d44b02dc --- /dev/null +++ b/2023/CREPE_ Can Vision-Language Foundation Models Reason Compositionally_/layout.json @@ -0,0 +1,9011 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 52, + 100, + 66, + 117 + ], + "blocks": [ + { + "bbox": [ + 52, + 100, + 66, + 117 + ], + "lines": [ + { + "bbox": [ + 52, + 100, + 66, + 117 + ], + "spans": [ + { + "bbox": [ + 52, + 100, + 66, + 117 + ], + "type": "image", + "image_path": "a9c996fb36116783728bd08036147b3e393278ffa081301e60aa8db52c0df423.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 67, + 102, + 541, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 102, + 541, + 121 + ], + "spans": [ + { + "bbox": [ + 67, + 102, + 541, + 121 + ], + "type": "text", + "content": "CREPE: Can Vision-Language Foundation Models Reason Compositionally?" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 58, + 142, + 534, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 142, + 534, + 171 + ], + "spans": [ + { + "bbox": [ + 58, + 142, + 534, + 171 + ], + "type": "text", + "content": "Zixian Ma" + }, + { + "bbox": [ + 58, + 142, + 534, + 171 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 58, + 142, + 534, + 171 + ], + "type": "text", + "content": ", Jerry Hong" + }, + { + "bbox": [ + 58, + 142, + 534, + 171 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 58, + 142, + 534, + 171 + ], + "type": "text", + "content": ", Mustafa Omer Gul" + }, + { + "bbox": [ + 58, + 142, + 534, + 171 + ], + "type": "inline_equation", + "content": "^{2*}" + }, + { + "bbox": [ + 58, + 142, + 534, + 171 + ], + "type": "text", + "content": ", Mona Gandhi" + }, + { + "bbox": [ + 58, + 142, + 534, + 171 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 58, + 142, + 534, + 171 + ], + "type": "text", + "content": ", Irenea Gao" + }, + { + "bbox": [ + 58, + 142, + 534, + 171 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 58, + 142, + 534, + 171 + ], + "type": "text", + "content": ", Ranjay Krishna" + }, + { + "bbox": [ + 58, + 142, + 534, + 171 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 58, + 142, + 534, + 171 + ], + "type": "text", + "content": ", Stanford University" + }, + { + "bbox": [ + 58, + 142, + 534, + 171 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 58, + 142, + 534, + 171 + ], + "type": "text", + "content": ", Cornell University" + }, + { + "bbox": [ + 58, + 142, + 534, + 171 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 58, + 142, + 534, + 171 + ], + "type": "text", + "content": ", University of Pennsylvania" + }, + { + "bbox": [ + 58, + 142, + 534, + 171 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 58, + 142, + 534, + 171 + ], + "type": "text", + "content": ", University of Washington" + }, + { + "bbox": [ + 58, + 142, + 534, + 171 + ], + "type": "inline_equation", + "content": "^{4}" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 67, + 174, + 521, + 199 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 174, + 521, + 199 + ], + "spans": [ + { + "bbox": [ + 67, + 174, + 521, + 199 + ], + "type": "text", + "content": "{zixianma, jerryhong, irena}@cs.stanford.edu mog29@cornell.edu mona09@seas.upenn.edu ranjay@cs.washington.edu" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 143, + 227, + 192, + 239 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 227, + 192, + 239 + ], + "spans": [ + { + "bbox": [ + 143, + 227, + 192, + 239 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 252, + 290, + 577 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 252, + 290, + 577 + ], + "spans": [ + { + "bbox": [ + 46, + 252, + 290, + 577 + ], + "type": "text", + "content": "A fundamental characteristic common to both human vision and natural language is their compositional nature. Yet, despite the performance gains contributed by large vision and language pretraining, we find that—across 7 architectures trained with 4 algorithms on massive datasets—they struggle at compositionality. To arrive at this conclusion, we introduce a new compositionality evaluation benchmark, CREPE, which measures two important aspects of compositionality identified by cognitive science literature: systematicity and productivity. To measure systematicity, CREPE consists of a test dataset containing over " + }, + { + "bbox": [ + 46, + 252, + 290, + 577 + ], + "type": "inline_equation", + "content": "370K" + }, + { + "bbox": [ + 46, + 252, + 290, + 577 + ], + "type": "text", + "content": " image-text pairs and three different seen-unseen splits. The three splits are designed to test models trained on three popular training datasets: CC-12M, YFCC-15M, and LAION-400M. We also generate " + }, + { + "bbox": [ + 46, + 252, + 290, + 577 + ], + "type": "inline_equation", + "content": "325K" + }, + { + "bbox": [ + 46, + 252, + 290, + 577 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 252, + 290, + 577 + ], + "type": "inline_equation", + "content": "316K" + }, + { + "bbox": [ + 46, + 252, + 290, + 577 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 252, + 290, + 577 + ], + "type": "inline_equation", + "content": "309K" + }, + { + "bbox": [ + 46, + 252, + 290, + 577 + ], + "type": "text", + "content": " hard negative captions for a subset of the pairs. To test productivity, CREPE contains " + }, + { + "bbox": [ + 46, + 252, + 290, + 577 + ], + "type": "inline_equation", + "content": "17K" + }, + { + "bbox": [ + 46, + 252, + 290, + 577 + ], + "type": "text", + "content": " image-text pairs with nine different complexities plus " + }, + { + "bbox": [ + 46, + 252, + 290, + 577 + ], + "type": "inline_equation", + "content": "278K" + }, + { + "bbox": [ + 46, + 252, + 290, + 577 + ], + "type": "text", + "content": " hard negative captions with atomic, swapping and negation foils. The datasets are generated by repurposing the Visual Genome scene graphs and region descriptions and applying handcrafted templates and GPT-3. For systematicity, we find that model performance decreases consistently when novel compositions dominate the retrieval set, with Recall@1 dropping by up to " + }, + { + "bbox": [ + 46, + 252, + 290, + 577 + ], + "type": "inline_equation", + "content": "9\\%" + }, + { + "bbox": [ + 46, + 252, + 290, + 577 + ], + "type": "text", + "content": ". For productivity, models' retrieval success decays as complexity increases, frequently nearing random chance at high complexity. These results hold regardless of model and training dataset size." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 600, + 128, + 613 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 600, + 128, + 613 + ], + "spans": [ + { + "bbox": [ + 47, + 600, + 128, + 613 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 621, + 288, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 621, + 288, + 693 + ], + "spans": [ + { + "bbox": [ + 46, + 621, + 288, + 693 + ], + "type": "text", + "content": "Compositionality, the understanding that \"the meaning of the whole is a function of the meanings of its parts\" [11], is held to be a key characteristic of human intelligence. In language, the whole is a sentence, made up of words. In vision, the whole is a scene, made up of parts like objects, their attributes, and their relationships [31, 35]." + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 318, + 228, + 421, + 369 + ], + "blocks": [ + { + "bbox": [ + 318, + 228, + 421, + 369 + ], + "lines": [ + { + "bbox": [ + 318, + 228, + 421, + 369 + ], + "spans": [ + { + "bbox": [ + 318, + 228, + 421, + 369 + ], + "type": "image", + "image_path": "642384e546156cd84a8f39013a8be14bd43d1f63dadd53c9dd7323fe5ade1c91.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 384, + 548, + 495 + ], + "lines": [ + { + "bbox": [ + 304, + 384, + 548, + 495 + ], + "spans": [ + { + "bbox": [ + 304, + 384, + 548, + 495 + ], + "type": "text", + "content": "Figure 1. We introduce CREPE, a benchmark to evaluate whether vision-language foundation models demonstrate two fundamental aspects of compositionality: systematicity and productivity. To evaluate systematicity, CREPE utilizes Visual Genome and introduces three new test datasets for the three popular pretraining datasets: CC-12M, YFCC-15M, and LAION-400M. These enable evaluating models' abilities to systematically generalize their understanding to seen compounds, unseen compounds, and even unseen atoms. To evaluate productivity, CREPE introduces examples of nine complexities, with three types of hard negatives for each." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 432, + 228, + 536, + 369 + ], + "blocks": [ + { + "bbox": [ + 432, + 228, + 536, + 369 + ], + "lines": [ + { + "bbox": [ + 432, + 228, + 536, + 369 + ], + "spans": [ + { + "bbox": [ + 432, + 228, + 536, + 369 + ], + "type": "image", + "image_path": "264f24940f3c1e1214081c6bd61d27dcfc06cd0d89c6209edb018b689942bb80.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 520, + 547, + 663 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 520, + 547, + 663 + ], + "spans": [ + { + "bbox": [ + 304, + 520, + 547, + 663 + ], + "type": "text", + "content": "Through compositional reasoning, humans can understand new scenes and generate complex sentences by combining known parts [6, 27, 30]. Despite compositionality's importance, there are no large-scale benchmarks directly evaluating whether vision-language models can reason compositionally. These models are pretrained using large-scale image-capture datasets [62, 64, 74], and are already widely applied for tasks that benefit from compositional reasoning, including retrieval, text-to-image generation, and open-vocabulary classification [10, 57, 60]. Especially as such models become ubiquitous \"foundations\" for other models [5], it is critical to understand their compositional abilities." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 665, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 547, + 714 + ], + "type": "text", + "content": "Previous work has evaluated these models using image-text retrieval [32,56,82]. However, the retrieval datasets used either do not provide controlled sets of negatives [45, 74] or study narrow negatives which vary along a single axis" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 59, + 702, + 124, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 702, + 124, + 712 + ], + "spans": [ + { + "bbox": [ + 59, + 702, + 124, + 712 + ], + "type": "text", + "content": "*Equal contribution" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "10910" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 55, + 68, + 527, + 224 + ], + "blocks": [ + { + "bbox": [ + 55, + 68, + 527, + 224 + ], + "lines": [ + { + "bbox": [ + 55, + 68, + 527, + 224 + ], + "spans": [ + { + "bbox": [ + 55, + 68, + 527, + 224 + ], + "type": "image", + "image_path": "78892989831eda5ec895483bb36d684450ec83896c6acbcf645c83f505bc112e.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 234, + 548, + 280 + ], + "lines": [ + { + "bbox": [ + 46, + 234, + 548, + 280 + ], + "spans": [ + { + "bbox": [ + 46, + 234, + 548, + 280 + ], + "type": "text", + "content": "Figure 2. An overview of the systematicity retrieval set generation process. First, a model's image-caption training set is parsed to identify what atoms and compounds the model has seen. Then, an evaluation set is divided into three compositional splits according to whether the model has seen all the compounds (Seen Compounds), only all the atoms of the caption (Unseen Compounds), or neither (Unseen Atoms). Finally, hard negative captions HN-ATOM and HN-COMP are generated for the hard negatives retrieval set " + }, + { + "bbox": [ + 46, + 234, + 548, + 280 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{test}^{HN}" + }, + { + "bbox": [ + 46, + 234, + 548, + 280 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 300, + 288, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 300, + 288, + 361 + ], + "spans": [ + { + "bbox": [ + 46, + 300, + 288, + 361 + ], + "type": "text", + "content": "(e.g. permuted word orders or single word substitutions as negative captions) [21, 51, 65, 75]. Further, these analyses have also not studied how retrieval performance varies when generalizing to unseen compositional combinations, or to combinations of increased complexity." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 361, + 290, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 361, + 290, + 628 + ], + "spans": [ + { + "bbox": [ + 46, + 361, + 290, + 628 + ], + "type": "text", + "content": "We introduce CREPE (Compositional REpresentation Evaluation): a new large-scale benchmark to evaluate two aspects of compositionality: systematicity and productivity (Figure 1). Systematicity measures how well a model is able to represent seen versus unseen atoms and their compositions. Productivity studies how well a model can comprehend an unbounded set of increasingly complex expressions. CREPE uses Visual Genome's scene graph representation as the compositionality language [35] and constructs evaluation datasets using its annotations. To test systematicity, we parse the captions in three popular training datasets, CC-12M [8], YFCC-15M [74], and LAION-400M [62], to identify atoms (objects, relations, or attributes) and compounds (combinations of atoms) present in each dataset. For each training set, we curate corresponding test sets containing " + }, + { + "bbox": [ + 46, + 361, + 290, + 628 + ], + "type": "inline_equation", + "content": "385K" + }, + { + "bbox": [ + 46, + 361, + 290, + 628 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 361, + 290, + 628 + ], + "type": "inline_equation", + "content": "385K" + }, + { + "bbox": [ + 46, + 361, + 290, + 628 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 361, + 290, + 628 + ], + "type": "inline_equation", + "content": "373K" + }, + { + "bbox": [ + 46, + 361, + 290, + 628 + ], + "type": "text", + "content": " image-text pairs respectively, with splits checking generalization to seen compounds, unseen compounds, and unseen atoms. To test productivity, CREPE contains " + }, + { + "bbox": [ + 46, + 361, + 290, + 628 + ], + "type": "inline_equation", + "content": "17K" + }, + { + "bbox": [ + 46, + 361, + 290, + 628 + ], + "type": "text", + "content": " image-text pairs split across nine levels of complexity, as defined by the number of atoms present in the text. Examples across all datasets are paired with various hard negative types to ensure the legitimacy of our conclusions." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 630, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 630, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 630, + 288, + 715 + ], + "type": "text", + "content": "Our experiments—across 7 architectures trained with 4 training algorithms on massive datasets—find that vision-language models struggle at compositionality, with both systematicity and productivity. We present six key findings: first, our systematicity experiments find that models' performance consistently drops between seen and unseen compositions; second, we observe larger drops for models trained on" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 300, + 547, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 300, + 547, + 407 + ], + "spans": [ + { + "bbox": [ + 304, + 300, + 547, + 407 + ], + "type": "text", + "content": "LAION-400M (up to a " + }, + { + "bbox": [ + 304, + 300, + 547, + 407 + ], + "type": "inline_equation", + "content": "9\\%" + }, + { + "bbox": [ + 304, + 300, + 547, + 407 + ], + "type": "text", + "content": " decrease in Recall@1); third, our productivity experiments indicate that retrieval performance degrades with increased caption complexity; fourth, we find no clear trend relating training dataset size to models' compositional reasoning; fifth, model size also has no impact; finally, models' zero-shot ImageNet classification accuracy correlates only with their absolute retrieval performance on the systematicity dataset but not systematic generalization to unseen compounds or to productivity." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 421, + 392, + 434 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 421, + 392, + 434 + ], + "spans": [ + { + "bbox": [ + 306, + 421, + 392, + 434 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 442, + 547, + 490 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 442, + 547, + 490 + ], + "spans": [ + { + "bbox": [ + 304, + 442, + 547, + 490 + ], + "type": "text", + "content": "Our work lies within the field of evaluating foundation models. Specifically, we measure visio-linguistic compositionality. To do so, we create a retrieval benchmark with hard negatives." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 491, + 548, + 683 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 491, + 548, + 683 + ], + "spans": [ + { + "bbox": [ + 304, + 491, + 548, + 683 + ], + "type": "text", + "content": "Contrastive Image-Text Pretraining. The recently released contrastively trained CLIP model [56] has catalyzed a wide array of work at the intersection of Computer Vision and Natural Language Processing. Since its release, CLIP has enabled several tasks, ranging from semantic segmentation to image captioning, many of which have remarkable zero-shot capability [12, 16, 38, 56, 71, 73]. CLIP has been used as a loss function within image synthesis applications [29, 44, 46, 54, 79, 83], acted as an automated evaluation metric [22, 52], used successfully as a feature extractor for various vision and language tasks [66], and incorporated into architectures for various tasks including dense prediction and video summarization [43, 50, 55, 58, 67, 68]. This success has also encouraged the design of other contrastive vision and language pretraining algorithms for image [15, 18, 40–42, 48, 69, 80, 81] and video domains [39, 76, 78]. Our work evaluates how well" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 306, + 693, + 547, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 693, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 693, + 547, + 713 + ], + "type": "text", + "content": "1We release our datasets, and code to generate and evaluate on our test sets at https://github.com/RAIVNLab/CREPE." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 294, + 748, + 317, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 758 + ], + "type": "text", + "content": "10911" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 288, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 288, + 108 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 288, + 108 + ], + "type": "text", + "content": "such contrastively trained models capture a fundamental property present in human vision and language: compositionality." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 112, + 288, + 352 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 112, + 288, + 352 + ], + "spans": [ + { + "bbox": [ + 47, + 112, + 288, + 352 + ], + "type": "text", + "content": "Compositionality. Compositionality allows us to comprehend an infinite number of scenes and utterances [37]. For an AI model, compositionality would not only allow for systematic, combinatorial generalization, but would also confer benefits such as controllability [5]. This promise prompted a wealth of work on both designing [2, 23, 25] and evaluating [17, 19, 26, 36, 72] compositional models. In our work, we focus on two aspects of compositionality: systematicity and productivity. While there is a plethora of benchmarks for systematic generalization within Computer Vision [3, 4, 19, 33] and Machine Learning [34, 36, 59], the subject has been almost unexplored for vision-language models, largely due to lack of benchmarks complementary to the different largescale training datasets. To address this, CREPE provides a benchmark with three different datasets to evaluate the compositional generalization of vision-language models. Productivity, on the other hand, has been studied only for specialized tasks [19] or toy domains [27, 36, 59]. CREPE evaluates productivity by using an image-text retrieval task featuring captions of varying compositional complexity." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 355, + 289, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 355, + 289, + 582 + ], + "spans": [ + { + "bbox": [ + 47, + 355, + 289, + 582 + ], + "type": "text", + "content": "Evaluation with hard negatives. Like us, past work evaluating models has commonly designed tasks featuring hard negatives to isolate particular model capabilities while overcoming the limitations of prior evaluation tasks. Using atomic foils that replace an atom in the image or text with a distractor has been the most common strategy [4,9,21,24,51,53,65]. Notably, Park et al. [53] targets verbs and person entities in videos; COVR [4] studies question answering with distractor images; FALSE [51] targets linguistic phenomena such as existence, cardinality and the recognition of actions and spatial relationships. Another strategy has been to swap atoms within a caption to test whether models behave akin to a bag-of-words [1, 51, 75]. In particular, Winoground [75] introduces a set of 800 human edited negatives to evaluate compositionality; it is the closest related work to us. We complement Winoground by scaling it up by three orders of magnitude, by decomposing compositionality into systematicity and productivity, and by studying a variety of different types of hard negatives." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 605, + 192, + 619 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 605, + 192, + 619 + ], + "spans": [ + { + "bbox": [ + 47, + 605, + 192, + 619 + ], + "type": "text", + "content": "3. Compositional evaluation" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 629, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 629, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 629, + 288, + 713 + ], + "type": "text", + "content": "The following section builds from the formally vacuous principle of compositionality to a well-defined evaluation scheme [27]. First, we establish the syntax and semantics of the composed language (Section 3.1). Then, we define expected behaviors from a model that achieves comprehension of said language (3.2, 3.3). Finally, we establish how to empirically measure those behaviors via retrieval (3.4)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 306, + 72, + 527, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 527, + 85 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 527, + 85 + ], + "type": "text", + "content": "3.1. Compositional language of visual concepts" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 90, + 547, + 222 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 90, + 547, + 222 + ], + "spans": [ + { + "bbox": [ + 304, + 90, + 547, + 222 + ], + "type": "text", + "content": "To evaluate vision-language models, we find that a compositional language consisting of scene graph visual concepts to be an appropriate foundation [35]. Accordingly, an atom " + }, + { + "bbox": [ + 304, + 90, + 547, + 222 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 304, + 90, + 547, + 222 + ], + "type": "text", + "content": " is defined as a singular visual concept, corresponding to a single scene graph node. Atoms are subtyped into objects " + }, + { + "bbox": [ + 304, + 90, + 547, + 222 + ], + "type": "inline_equation", + "content": "A_{o}" + }, + { + "bbox": [ + 304, + 90, + 547, + 222 + ], + "type": "text", + "content": ", relationships " + }, + { + "bbox": [ + 304, + 90, + 547, + 222 + ], + "type": "inline_equation", + "content": "A_{r}" + }, + { + "bbox": [ + 304, + 90, + 547, + 222 + ], + "type": "text", + "content": ", and attributes " + }, + { + "bbox": [ + 304, + 90, + 547, + 222 + ], + "type": "inline_equation", + "content": "A_{a}" + }, + { + "bbox": [ + 304, + 90, + 547, + 222 + ], + "type": "text", + "content": ". A compound " + }, + { + "bbox": [ + 304, + 90, + 547, + 222 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 304, + 90, + 547, + 222 + ], + "type": "text", + "content": " is defined as a primitive composition of multiple atoms, which corresponds to connections between scene graph nodes. Visual concepts admit two compound types: the attachment of attribute to objects (\"black dog\") " + }, + { + "bbox": [ + 304, + 90, + 547, + 222 + ], + "type": "inline_equation", + "content": "C_{ao}" + }, + { + "bbox": [ + 304, + 90, + 547, + 222 + ], + "type": "text", + "content": ", and the attachment of two objects via a relationship (\"man hugs child\") " + }, + { + "bbox": [ + 304, + 90, + 547, + 222 + ], + "type": "inline_equation", + "content": "C_{oro}" + }, + { + "bbox": [ + 304, + 90, + 547, + 222 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 222, + 547, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 222, + 547, + 389 + ], + "spans": [ + { + "bbox": [ + 304, + 222, + 547, + 389 + ], + "type": "text", + "content": "The composition of these compounds form subgraphs " + }, + { + "bbox": [ + 304, + 222, + 547, + 389 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 304, + 222, + 547, + 389 + ], + "type": "text", + "content": ", which can be translated to natural language captions " + }, + { + "bbox": [ + 304, + 222, + 547, + 389 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 304, + 222, + 547, + 389 + ], + "type": "text", + "content": ". Conversely, captions " + }, + { + "bbox": [ + 304, + 222, + 547, + 389 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 304, + 222, + 547, + 389 + ], + "type": "text", + "content": " derived from image-text datasets " + }, + { + "bbox": [ + 304, + 222, + 547, + 389 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 304, + 222, + 547, + 389 + ], + "type": "text", + "content": " can be parsed to become scene graphs " + }, + { + "bbox": [ + 304, + 222, + 547, + 389 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 304, + 222, + 547, + 389 + ], + "type": "text", + "content": ". This extensible language is capable of capturing a number of linguistic phenomena identified in existing literature [51, 72], including the existence of concepts (\"a photo with flowers\"), spatial relationships (\"a grill on the left of a staircase\"), action relationships (\"a person throwing a frisbee\"), prepositional attachment (\"A bird with green wings\"), and negation (\"There are no trucks on the road\"). Furthermore, while this study focuses on visual concepts, scene graphs featuring common-sense relationships or other more abstract concepts can be designed; therefore, our methodology is widely applicable [61]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 396, + 391, + 408 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 396, + 391, + 408 + ], + "spans": [ + { + "bbox": [ + 306, + 396, + 391, + 408 + ], + "type": "text", + "content": "3.2. Systematicity" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 414, + 547, + 654 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 414, + 547, + 654 + ], + "spans": [ + { + "bbox": [ + 304, + 414, + 547, + 654 + ], + "type": "text", + "content": "With our compositional language in place, we now define two dimensions of compositionality—systematicity and productivity—which we adapt to vision-language representations. Systematicity evaluates a model's ability to systematically recombine seen atoms in compounds. Concretely, let " + }, + { + "bbox": [ + 304, + 414, + 547, + 654 + ], + "type": "inline_equation", + "content": "\\mathrm{SEEN}(A,D)" + }, + { + "bbox": [ + 304, + 414, + 547, + 654 + ], + "type": "text", + "content": " denote if an atom is seen in a training dataset " + }, + { + "bbox": [ + 304, + 414, + 547, + 654 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 304, + 414, + 547, + 654 + ], + "type": "text", + "content": ", namely " + }, + { + "bbox": [ + 304, + 414, + 547, + 654 + ], + "type": "inline_equation", + "content": "\\exists (I,S) \\in \\mathcal{D}: A \\in S" + }, + { + "bbox": [ + 304, + 414, + 547, + 654 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 414, + 547, + 654 + ], + "type": "inline_equation", + "content": "\\mathrm{SEEN}(C,D)" + }, + { + "bbox": [ + 304, + 414, + 547, + 654 + ], + "type": "text", + "content": " denote if a compound is seen in a dataset " + }, + { + "bbox": [ + 304, + 414, + 547, + 654 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 304, + 414, + 547, + 654 + ], + "type": "text", + "content": ", namely " + }, + { + "bbox": [ + 304, + 414, + 547, + 654 + ], + "type": "inline_equation", + "content": "\\exists (I,S) \\in \\mathcal{D}: C \\subseteq S" + }, + { + "bbox": [ + 304, + 414, + 547, + 654 + ], + "type": "text", + "content": ". To evaluate systematicity, we define three compositional splits: Seen Compounds (SC), Unseen Compounds (UC) and Unseen Atoms (UA). SC is the split where all compounds (and thus all atoms) of every caption have been seen in the training dataset, i.e. " + }, + { + "bbox": [ + 304, + 414, + 547, + 654 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{\\mathrm{SC}} = \\{(I,S) \\in \\mathcal{D}_{test} | \\forall C \\subseteq S : \\mathrm{SEEN}(C,\\mathcal{D}_{train})\\}" + }, + { + "bbox": [ + 304, + 414, + 547, + 654 + ], + "type": "text", + "content": ". UC is the split where, for each caption, all atoms have been seen but at least one compound has NOT, i.e. " + }, + { + "bbox": [ + 304, + 414, + 547, + 654 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{UC} = \\{(I,S) \\in \\mathcal{D}_{test} | (\\forall A \\in S : \\mathrm{SEEN}(A,\\mathcal{D}_{train}) \\land (\\exists C \\subseteq S : \\neg \\mathrm{Seen}(C,\\mathcal{D}_{train}))\\}" + }, + { + "bbox": [ + 304, + 414, + 547, + 654 + ], + "type": "text", + "content": ". UA is the split where each caption contains at least one atom that has NOT been seen, i.e. " + }, + { + "bbox": [ + 304, + 414, + 547, + 654 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{UA} = \\{(I,S) \\in \\mathcal{D}_{test} | \\exists A \\in S : \\neg \\mathrm{SEEN}(A,\\mathcal{D}_{train})\\}" + }, + { + "bbox": [ + 304, + 414, + 547, + 654 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 659, + 388, + 672 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 659, + 388, + 672 + ], + "spans": [ + { + "bbox": [ + 306, + 659, + 388, + 672 + ], + "type": "text", + "content": "3.3. Productivity" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 677, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 547, + 713 + ], + "type": "text", + "content": "Productivity refers to a capacity to comprehend an unbounded set of expressions. Since the set of atoms in any dataset is finite, a reasonable substitute for testing unbounded" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "10912" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 55, + 68, + 537, + 210 + ], + "blocks": [ + { + "bbox": [ + 55, + 68, + 537, + 210 + ], + "lines": [ + { + "bbox": [ + 55, + 68, + 537, + 210 + ], + "spans": [ + { + "bbox": [ + 55, + 68, + 537, + 210 + ], + "type": "image", + "image_path": "c58dde3537b407e65e2764fcc4f7e9e0d34d1cd97d4f5d4f67c359b050929c7c.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 220, + 547, + 255 + ], + "lines": [ + { + "bbox": [ + 46, + 220, + 547, + 255 + ], + "spans": [ + { + "bbox": [ + 46, + 220, + 547, + 255 + ], + "type": "text", + "content": "Figure 3. An overview of the productivity retrieval set generation process. By performing random walks on the scene graphs of an evaluation dataset, we generate subgraphs of various complexities. Then, for complexities " + }, + { + "bbox": [ + 46, + 220, + 547, + 255 + ], + "type": "inline_equation", + "content": "n \\in \\{4, .5, .12\\}" + }, + { + "bbox": [ + 46, + 220, + 547, + 255 + ], + "type": "text", + "content": " and three hard negative types, we populate the retrieval set " + }, + { + "bbox": [ + 46, + 220, + 547, + 255 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{test}^{HN}" + }, + { + "bbox": [ + 46, + 220, + 547, + 255 + ], + "type": "text", + "content": " by generating a ground truth caption for each " + }, + { + "bbox": [ + 46, + 220, + 547, + 255 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 46, + 220, + 547, + 255 + ], + "type": "text", + "content": "-subgraph and hard negatives for each caption." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 55, + 285, + 545, + 345 + ], + "blocks": [ + { + "bbox": [ + 78, + 266, + 514, + 277 + ], + "lines": [ + { + "bbox": [ + 78, + 266, + 514, + 277 + ], + "spans": [ + { + "bbox": [ + 78, + 266, + 514, + 277 + ], + "type": "text", + "content": "Table 1. We summarize the sizes of the eight evaluation datasets we create for systematicity and productivity evaluation." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 55, + 285, + 545, + 345 + ], + "lines": [ + { + "bbox": [ + 55, + 285, + 545, + 345 + ], + "spans": [ + { + "bbox": [ + 55, + 285, + 545, + 345 + ], + "type": "table", + "html": "
SystematicityProductivity
\\(\\mathcal{D}_{test}^{RAW}\\) (# of image-text pairs)\\(\\mathcal{D}_{test}^{HN}\\) (# of texts)\\(\\mathcal{D}_{test}^{RAW}\\)\\(\\mathcal{D}_{test}^{HN}\\)
Training dataCC-12MYFCC-15MLAION-400MCC-12MYFCC-15MLAION-400MAnyAny
Dataset size385,777385,777373,703325,523316,668309,34217,553278,730
", + "image_path": "242c2e1577768db32adbccecf3d2fe3b90986baa585a81b12dded08ba4e1c8ac.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 366, + 290, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 366, + 290, + 521 + ], + "spans": [ + { + "bbox": [ + 46, + 366, + 290, + 521 + ], + "type": "text", + "content": "comprehension is testing comprehension over increasingly complex scenes. Now, an image " + }, + { + "bbox": [ + 46, + 366, + 290, + 521 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 46, + 366, + 290, + 521 + ], + "type": "text", + "content": " does not have a notion of complexity, since it is theoretically infinitely describable; on the other hand, we can define a notion of complexity for a caption " + }, + { + "bbox": [ + 46, + 366, + 290, + 521 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 46, + 366, + 290, + 521 + ], + "type": "text", + "content": ": the number of atoms in its corresponding scene graph " + }, + { + "bbox": [ + 46, + 366, + 290, + 521 + ], + "type": "inline_equation", + "content": "|S_T|" + }, + { + "bbox": [ + 46, + 366, + 290, + 521 + ], + "type": "text", + "content": ". Therefore, a productive vision-language model should be able to match a given image to the correct corresponding caption, regardless of that caption's complexity. To evaluate productivity, we define a range of productivity complexity (in our case, " + }, + { + "bbox": [ + 46, + 366, + 290, + 521 + ], + "type": "inline_equation", + "content": "n = 4, 5, \\ldots, 12" + }, + { + "bbox": [ + 46, + 366, + 290, + 521 + ], + "type": "text", + "content": "). We need splits of the evaluation dataset based on these complexities, where image-text pairs in a given split have a fixed complexity " + }, + { + "bbox": [ + 46, + 366, + 290, + 521 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 46, + 366, + 290, + 521 + ], + "type": "text", + "content": ", and evaluate a model's performance over each split." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 528, + 249, + 540 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 528, + 249, + 540 + ], + "spans": [ + { + "bbox": [ + 47, + 528, + 249, + 540 + ], + "type": "text", + "content": "3.4. Compositional evaluation via retrieval" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 546, + 288, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 546, + 288, + 677 + ], + "spans": [ + { + "bbox": [ + 46, + 546, + 288, + 677 + ], + "type": "text", + "content": "We evaluate compositional reasoning using zero-shot image-to-text and text-to-image retrieval. This formulation probes the representation space as directly as possible and is already the most common evaluation method for vision-language foundation models [56]. Theoretically, any existing image-text dataset can be used as retrieval sets for our evaluation. However, one challenging limitation in existing datasets renders the metrics evaluated on them inaccurate. Consider using an image query of a \"plant inside a yellow vase on top of a black television.\" Retrieving unintended alternative positives (e.g. \"a black television\") is not necessarily incor" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 366, + 547, + 546 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 366, + 547, + 546 + ], + "spans": [ + { + "bbox": [ + 304, + 366, + 547, + 546 + ], + "type": "text", + "content": "rect. Similarly, if no other texts in the retrieval set contain a \"plant\" and a \"television\", retrieving the correct text doesn't suggest that the model comprehends the image. Ideally, to properly evaluate a model, the retrieval dataset should contain hard negatives for every query. A hard negative is a caption that does not faithfully represent the corresponding image, and differs from the ground truth caption by some minimal atomic shift. A example hard negative for the query above is \"man inside a yellow vase on top of a black television.\" By erring in a single, granular syntactic or semantic fashion, hard negatives allow for variations in retrieval performance to be attributable to a specific failure mode of a model's compositional comprehension (see Appendix). We address this need for a new benchmark dataset to evaluate the systematicity and productivity of vision-language models." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 557, + 545, + 587 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 557, + 545, + 587 + ], + "spans": [ + { + "bbox": [ + 305, + 557, + 545, + 587 + ], + "type": "text", + "content": "4. CREPE: a large-scale benchmark for vision-language compositionality" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 594, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 547, + 713 + ], + "type": "text", + "content": "There are several challenges to creating image-text retrieval datasets that evaluate compositional systematicity and productivity. For systematicity, the primary challenge lies in parsing the training dataset for seen atoms and compounds in order to split the data into the three compositional splits. For productivity, the major challenge is generating imagedataset pairs across different text complexities for the retrieval sets. For both datasets, it is crucial to enumerate different types of hard negatives, and to design an automated hard negative generator which ensures the incorrectness of the" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 46, + 683, + 288, + 714 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 683, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 683, + 288, + 714 + ], + "type": "text", + "content": "2By avoiding captions with redundant objects (\"... a lamb and a lamb and...\" and abstract modifiers (\"there are many lampposts\"), we ensure atom count is tightly coupled with caption complexity." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10913" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 288, + 109 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 288, + 109 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 288, + 109 + ], + "type": "text", + "content": "negatives it generates. We detail our methods for tackling these challenges for future efforts that attempt to create similar benchmarks for other training datasets." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 123, + 216, + 136 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 123, + 216, + 136 + ], + "spans": [ + { + "bbox": [ + 47, + 123, + 216, + 136 + ], + "type": "text", + "content": "4.1. Creating systematicity datasets" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 144, + 287, + 192 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 144, + 287, + 192 + ], + "spans": [ + { + "bbox": [ + 47, + 144, + 287, + 192 + ], + "type": "text", + "content": "To create the three systematicity splits—SC, SA, UA—we parse a given training dataset " + }, + { + "bbox": [ + 47, + 144, + 287, + 192 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 47, + 144, + 287, + 192 + ], + "type": "text", + "content": " into its constituent atoms and compounds, filter low-quality data, and generate hard negatives (Figure 2)." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 194, + 288, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 194, + 288, + 361 + ], + "spans": [ + { + "bbox": [ + 46, + 194, + 288, + 361 + ], + "type": "text", + "content": "Parsing a dataset into atoms and compounds Since we utilize the scene graph representation as our compositional language, we use the Stanford Scene Graph Parser [63,77] to parse texts in " + }, + { + "bbox": [ + 46, + 194, + 288, + 361 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{train}" + }, + { + "bbox": [ + 46, + 194, + 288, + 361 + ], + "type": "text", + "content": " into their corresponding scene graphs with objects, attributes and relationships. Since the parser only parses for objects and relationships, we further extract the attributes from the text via spaCy's natural language processing parser by identifying adjective part-of-speech tags. These connected objects, attributes, and relationships constitute our seen atoms and compounds. Similarly, we parse a given " + }, + { + "bbox": [ + 46, + 194, + 288, + 361 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{test}" + }, + { + "bbox": [ + 46, + 194, + 288, + 361 + ], + "type": "text", + "content": " and divide all the image-text pairs into the three splits based on the presence of unseen atoms and/or compounds in the parsed training set. Details on the quality of the scene graph parser can be found in the Appendix." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 363, + 288, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 363, + 288, + 449 + ], + "spans": [ + { + "bbox": [ + 46, + 363, + 288, + 449 + ], + "type": "text", + "content": "Filtering low-quality data We perform the following filtering steps on the image-text pairs in all splits: we only keep region crops which have an area greater than or equal to " + }, + { + "bbox": [ + 46, + 363, + 288, + 449 + ], + "type": "inline_equation", + "content": "40K" + }, + { + "bbox": [ + 46, + 363, + 288, + 449 + ], + "type": "text", + "content": " pixels, occupy at least " + }, + { + "bbox": [ + 46, + 363, + 288, + 449 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 46, + 363, + 288, + 449 + ], + "type": "text", + "content": " of the whole image, and whose width-to-height ratio is between 0.5-2.0. We only include text which have at least 2 atoms and 1 compound and de-duplicate text using their corresponding scene graphs." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 450, + 289, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 450, + 289, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 450, + 289, + 714 + ], + "type": "text", + "content": "Generating hard negatives We introduce two types of hard negatives: HN-ATOM and HN-COMP. HN-ATOM replaces " + }, + { + "bbox": [ + 46, + 450, + 289, + 714 + ], + "type": "inline_equation", + "content": "A_{a}" + }, + { + "bbox": [ + 46, + 450, + 289, + 714 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 450, + 289, + 714 + ], + "type": "inline_equation", + "content": "A_{o}" + }, + { + "bbox": [ + 46, + 450, + 289, + 714 + ], + "type": "text", + "content": ", or " + }, + { + "bbox": [ + 46, + 450, + 289, + 714 + ], + "type": "inline_equation", + "content": "A_{r}" + }, + { + "bbox": [ + 46, + 450, + 289, + 714 + ], + "type": "text", + "content": " in the text with an atomic foil. For example, for the caption \"a grill on top of the porch\", one HN-ATOM can be \"a grill underneath the porch\", where the " + }, + { + "bbox": [ + 46, + 450, + 289, + 714 + ], + "type": "inline_equation", + "content": "A_{r}" + }, + { + "bbox": [ + 46, + 450, + 289, + 714 + ], + "type": "text", + "content": " \"on top of\" is replaced by \"underneath\". Since captions and scene graphs are not exhaustive, this replacement must be done carefully. For example, if a dog is white and furry, but only \"white\" is annotated, replacing the atom \"white\" with \"furry\" would result in a correct caption. To minimize errors, we employ WordNet [49] to pick replacement atoms that are either antonyms (\"black dog\") or share the same grand-hypernym (\"pink dog\") with respect to the original atom. Furthermore, we use BERT to select the most sensical negatives for each ground truth caption [13,51]. HN-COMP concatenates two compound foils where each contains an atomic foil. For instance, one HN-COMP of the caption \"a pink car\" can be \"a blue car and a pink toy\", where \"blue\" and \"toy\" are the atomic foils in the two compounds foils \"blue car\" and \"pink toy\". We only generate negatives for one-compound examples for systematicity evaluation, as productivity covers complex captions with more atoms." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 72, + 471, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 471, + 85 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 471, + 85 + ], + "type": "text", + "content": "4.2. Creating productivity datasets" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 90, + 547, + 125 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 90, + 547, + 125 + ], + "spans": [ + { + "bbox": [ + 305, + 90, + 547, + 125 + ], + "type": "text", + "content": "We first generate ground truth captions for scene graphs of varying complexity, filter for data quality, and then generate hard negatives for each example (Figure 3)." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 126, + 547, + 317 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 126, + 547, + 317 + ], + "spans": [ + { + "bbox": [ + 304, + 126, + 547, + 317 + ], + "type": "text", + "content": "Generating captions We systematically generate captions of different atom counts for each image. Given a scene graph, we perform a random walk of length " + }, + { + "bbox": [ + 304, + 126, + 547, + 317 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 304, + 126, + 547, + 317 + ], + "type": "text", + "content": " through the graph to generate a subgraph. Each subgraph corresponds to a specific region of the image, determined by the union of the bounding boxes of the subgraph atoms. We filter out low-quality regions using the same process as systematicity with additional dedduplication on patches that overlap by " + }, + { + "bbox": [ + 304, + 126, + 547, + 317 + ], + "type": "inline_equation", + "content": "\\geq 75\\%" + }, + { + "bbox": [ + 304, + 126, + 547, + 317 + ], + "type": "text", + "content": ". For simple subgraphs (" + }, + { + "bbox": [ + 304, + 126, + 547, + 317 + ], + "type": "inline_equation", + "content": "n = 4" + }, + { + "bbox": [ + 304, + 126, + 547, + 317 + ], + "type": "text", + "content": "), we produce captions using handcrafted templates. For larger subgraphs (" + }, + { + "bbox": [ + 304, + 126, + 547, + 317 + ], + "type": "inline_equation", + "content": "n \\geq 5" + }, + { + "bbox": [ + 304, + 126, + 547, + 317 + ], + "type": "text", + "content": "), we leverage GPT-3 [7] (text-davinci-002) to generate captions based on a text description of the scene graph, which lists all objects and relationships. We prompt GPT-3 using 5 manually written captions per complexity, filtering out captions where GPT-3 errs and omits atoms from the subgraph during generation (see more details in Appendix)." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 317, + 547, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 317, + 547, + 544 + ], + "spans": [ + { + "bbox": [ + 304, + 317, + 547, + 544 + ], + "type": "text", + "content": "Generating hard negatives For productivity, we employ three hard negatives types (HN-ATOM from systematicity, HN-SWAP, and HN-NEG) corresponding to three hypothesized model error modes. First, as a caption's complexity increases, a model may begin to ignore individual atoms. HN-ATOM randomly selects an atom from the caption and replaces it with an incorrect atom. Second, as a caption's complexity increases, a model may treat captions as \"bags of words\", ignoring syntactic connections built out of word order. A swap hard negative (HN-SWAP) accordingly permutes atoms of the same subtype in a caption. This hard negative is similar to Winoground [75], but in the context of varying caption complexity. On top of Wordnet, we use entailment with RoBERTa to further filter errant HN-SWAP hard negatives [47]. Finally, as a caption's complexity increases, a model may begin to lose comprehension of negations. A negation hard negative (HN-NEG) either negates the entire caption or a specific atom. Refer to the Appendix for details on generating HN-SWAP and HN-NEG." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 552, + 468, + 563 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 552, + 468, + 563 + ], + "spans": [ + { + "bbox": [ + 306, + 552, + 468, + 563 + ], + "type": "text", + "content": "4.3. The final benchmark datasets" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 570, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 547, + 713 + ], + "type": "text", + "content": "For both productivity and systematicity, we generate two test datasets: " + }, + { + "bbox": [ + 304, + 570, + 547, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{test}^{HN}" + }, + { + "bbox": [ + 304, + 570, + 547, + 713 + ], + "type": "text", + "content": ", which contains image-ground truth text pairs along with all generated hard negatives, and " + }, + { + "bbox": [ + 304, + 570, + 547, + 713 + ], + "type": "inline_equation", + "content": "D_{test}^{RAW}" + }, + { + "bbox": [ + 304, + 570, + 547, + 713 + ], + "type": "text", + "content": ", which contains only image-ground truth text pairs. To measure the data quality, we randomly sample " + }, + { + "bbox": [ + 304, + 570, + 547, + 713 + ], + "type": "inline_equation", + "content": "2\\%" + }, + { + "bbox": [ + 304, + 570, + 547, + 713 + ], + "type": "text", + "content": " of productivity ground truth captions generated by GPT-3 and " + }, + { + "bbox": [ + 304, + 570, + 547, + 713 + ], + "type": "inline_equation", + "content": "1\\%" + }, + { + "bbox": [ + 304, + 570, + 547, + 713 + ], + "type": "text", + "content": " of the queries in the productivity and systematicity " + }, + { + "bbox": [ + 304, + 570, + 547, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{test}^{HN}" + }, + { + "bbox": [ + 304, + 570, + 547, + 713 + ], + "type": "text", + "content": " sets for manual human verification. We assign 2 annotators to each set and measure both generated quality and intra-annotator agreement. " + }, + { + "bbox": [ + 304, + 570, + 547, + 713 + ], + "type": "inline_equation", + "content": "87.9\\%" + }, + { + "bbox": [ + 304, + 570, + 547, + 713 + ], + "type": "text", + "content": " of sampled productivity ground truth captions generated by GPT-3 are rated as faithful to the image, with an average pairwise annotator agreement of" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "10914" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 82, + 181, + 195 + ], + "blocks": [ + { + "bbox": [ + 141, + 72, + 361, + 81 + ], + "lines": [ + { + "bbox": [ + 141, + 72, + 361, + 81 + ], + "spans": [ + { + "bbox": [ + 141, + 72, + 361, + 81 + ], + "type": "text", + "content": "UC vs. SC Recall@1 in Systematicity Hard Negative Retrieval Set" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 50, + 82, + 181, + 195 + ], + "lines": [ + { + "bbox": [ + 50, + 82, + 181, + 195 + ], + "spans": [ + { + "bbox": [ + 50, + 82, + 181, + 195 + ], + "type": "image", + "image_path": "ed36c6a8819305e748103f78ece790008792765f4fdd142efd834a9cff626ecd.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 187, + 82, + 315, + 195 + ], + "blocks": [ + { + "bbox": [ + 187, + 82, + 315, + 195 + ], + "lines": [ + { + "bbox": [ + 187, + 82, + 315, + 195 + ], + "spans": [ + { + "bbox": [ + 187, + 82, + 315, + 195 + ], + "type": "image", + "image_path": "2621a403d99f578ed9f00e8ee9db686a9d364a27bd02d80ffc9f61299891aeed.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 320, + 82, + 541, + 196 + ], + "blocks": [ + { + "bbox": [ + 320, + 82, + 541, + 196 + ], + "lines": [ + { + "bbox": [ + 320, + 82, + 541, + 196 + ], + "spans": [ + { + "bbox": [ + 320, + 82, + 541, + 196 + ], + "type": "image", + "image_path": "e0a0d07ee20f91e37fadf885a6f76c1e160233bdbd168e786d74f060721bd78f.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 45, + 205, + 547, + 240 + ], + "lines": [ + { + "bbox": [ + 45, + 205, + 547, + 240 + ], + "spans": [ + { + "bbox": [ + 45, + 205, + 547, + 240 + ], + "type": "text", + "content": "Figure 4. We plot models' recall@1 on the Seen Compounds vs. Unseen Compounds split of the systematicity retrieval set with hard negatives HN-ATOM, HN-COMP and both types. We observe a consistent drop in models' performance from the SC to UC split when the hard negative set consists of both HN-ATOM and HN-COMP or HN-ATOM only." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 99, + 251, + 247, + 384 + ], + "blocks": [ + { + "bbox": [ + 99, + 251, + 247, + 384 + ], + "lines": [ + { + "bbox": [ + 99, + 251, + 247, + 384 + ], + "spans": [ + { + "bbox": [ + 99, + 251, + 247, + 384 + ], + "type": "image", + "image_path": "15f88ae173b78505f0baced8169ba753fb9391b0ba48e585b2868b727d2f448a.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 394, + 546, + 417 + ], + "lines": [ + { + "bbox": [ + 46, + 394, + 546, + 417 + ], + "spans": [ + { + "bbox": [ + 46, + 394, + 546, + 417 + ], + "type": "text", + "content": "Figure 5. Productivity Analysis. We plot models' Recall@1 on the hard negatives retrieval set against complexity, averaged across all models pretrained on all three training datasets. We find that models' ability to retrieve the ground-truth degrades as complexity increases." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 249, + 251, + 370, + 384 + ], + "blocks": [ + { + "bbox": [ + 249, + 251, + 370, + 384 + ], + "lines": [ + { + "bbox": [ + 249, + 251, + 370, + 384 + ], + "spans": [ + { + "bbox": [ + 249, + 251, + 370, + 384 + ], + "type": "image", + "image_path": "828c834819114f045306472fb4615b1148643cfe9e6d839c3b2327547c069ec4.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 373, + 252, + 496, + 384 + ], + "blocks": [ + { + "bbox": [ + 373, + 252, + 496, + 384 + ], + "lines": [ + { + "bbox": [ + 373, + 252, + 496, + 384 + ], + "spans": [ + { + "bbox": [ + 373, + 252, + 496, + 384 + ], + "type": "image", + "image_path": "7d7e86b7d68791de34e7909e6e86fefa9cecd38e116ecd7b3197df03406cc2bc.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 437, + 289, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 437, + 289, + 486 + ], + "spans": [ + { + "bbox": [ + 46, + 437, + 289, + 486 + ], + "type": "text", + "content": "88.8%. 83.7% of productivity and 86.0% of systematicity hard negatives were rated as genuine negatives (i.e. made factually incorrect statements about the image), with pairwise annotator agreements of 84.3% and 83.7% respectively." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 499, + 128, + 513 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 499, + 128, + 513 + ], + "spans": [ + { + "bbox": [ + 47, + 499, + 128, + 513 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 521, + 287, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 521, + 287, + 664 + ], + "spans": [ + { + "bbox": [ + 46, + 521, + 287, + 664 + ], + "type": "text", + "content": "We present our experimental setup and results with six takeaways. First, our systematicity experiments show performance decreases consistently on compounds unseen in training. Second, the greatest drop between splits occurs for models trained on LAION-400M. Third, our productivity results reveal models' retrieval performance decays with increasing complexity. Fourth, we find that dataset size has no impact on compositionality. Fifth, we find no clear trend relating model size to compositionality. Finally, models' zero-shot ImageNet classification accuracy correlates with retrieval performance on the systematicity dataset but not systematic generalization to the UC split or productivity." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "type": "text", + "content": "Datasets. We utilize Visual Genome to create our test datasets. For systematicity, image patches and corresponding spelling-corrected region descriptions are used. We provide three different splits for " + }, + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{test}^{HN}" + }, + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "type": "text", + "content": ", for three training datasets:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 437, + 548, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 437, + 548, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 437, + 548, + 715 + ], + "type": "text", + "content": "CC-12M, YFCC-15M and LAION-400M. For productivity, Visual Genome's image-scene graph pairs are used to create captions and hard negatives for " + }, + { + "bbox": [ + 304, + 437, + 548, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{test}^{RAW}" + }, + { + "bbox": [ + 304, + 437, + 548, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 437, + 548, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{test}^{HN}" + }, + { + "bbox": [ + 304, + 437, + 548, + 715 + ], + "type": "text", + "content": " (Table 1). Models. We firstly evaluate seven vision-language models pretrained with contrastive loss [70] across three commonly used image-text datasets: Conceptual Captions 12M (CC-12M) [8], a subset of the YFCC100M dataset (YFCC-15M) [56, 74] and LAION-400M [62]. We limit our evaluation to models openly released in the OpenCLIP repository [28] for systematicity evaluation. These include ResNet (RN) [20] and Vision Transformer (ViT) [14] encoders of different sizes: RN50, RN101, ViT-B-16, ViT-B-16-plus-240, ViT-B-32 and ViT-L-14. Additionally, since productivity evaluation is not restricted to models that were trained on publicly released datasets, we conduct productivity evaluation on other foundation vision-language models as well. Specifically, we consider OpenAI's CLIP [56] with ResNet and ViT backbones, CyCLIP [18] (a variant of CLIP introducing auxiliary losses that regularize the gap in similarity scores between mismatched pairs, trained on Conceptual Captions 3M [64] with a ResNet-50 [20] backbone), ALBEF [41] (additionally trained with a masked language modeling and image-text matching loss) and FLAVA [69] (which" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10915" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 99, + 70, + 500, + 198 + ], + "blocks": [ + { + "bbox": [ + 99, + 70, + 500, + 198 + ], + "lines": [ + { + "bbox": [ + 99, + 70, + 500, + 198 + ], + "spans": [ + { + "bbox": [ + 99, + 70, + 500, + 198 + ], + "type": "image", + "image_path": "e5a2090aee51b8259ee05d7ddcbd12458a75b88f96749bb4501142027638bf9a.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 205, + 547, + 251 + ], + "lines": [ + { + "bbox": [ + 46, + 205, + 547, + 251 + ], + "spans": [ + { + "bbox": [ + 46, + 205, + 547, + 251 + ], + "type": "text", + "content": "Figure 6. Productivity Analysis on Additional Foundation Vision-language Models. We plot models' Recall@1 on the productivity hard negatives retrieval set against complexity, where OpenAI CLIP's performance is averaged across five models RN50, RN101, ViT-B-16, ViT-B-32 and ViT-L-14. We find that all models' retrieval performance decreases as complexity increases in both the HN-ATOM and HN-SWAP retrieval sets. For the HN-NEG set, all models except for CLIP either drop in performance or remain at random chance." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 48, + 271, + 282, + 385 + ], + "blocks": [ + { + "bbox": [ + 48, + 271, + 282, + 385 + ], + "lines": [ + { + "bbox": [ + 48, + 271, + 282, + 385 + ], + "spans": [ + { + "bbox": [ + 48, + 271, + 282, + 385 + ], + "type": "image", + "image_path": "aae2bbb24f2aa7dfb837046d5a26df655857613ca4224f7bc0e1a7b78debef63.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 396, + 288, + 440 + ], + "lines": [ + { + "bbox": [ + 46, + 396, + 288, + 440 + ], + "spans": [ + { + "bbox": [ + 46, + 396, + 288, + 440 + ], + "type": "text", + "content": "Figure 7. A plot showing the correlation between zero-shot top-1 accuracy on ImageNet and Recall@1 on CREPE's systematicity hard-negative sets. We observe a strong correlation, with an " + }, + { + "bbox": [ + 46, + 396, + 288, + 440 + ], + "type": "inline_equation", + "content": "R^2" + }, + { + "bbox": [ + 46, + 396, + 288, + 440 + ], + "type": "text", + "content": " score of 0.9914 for the SC split and 0.9534 for the UC split." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 460, + 286, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 460, + 286, + 582 + ], + "spans": [ + { + "bbox": [ + 46, + 460, + 286, + 582 + ], + "type": "text", + "content": "further adds unimodal losses for image and text domains). Retrieval. For " + }, + { + "bbox": [ + 46, + 460, + 286, + 582 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{test}^{HN}" + }, + { + "bbox": [ + 46, + 460, + 286, + 582 + ], + "type": "text", + "content": ", we perform image-to-text retrieval and stratify results by split and hard negative type. For systematicity, the splits are " + }, + { + "bbox": [ + 46, + 460, + 286, + 582 + ], + "type": "inline_equation", + "content": "SC" + }, + { + "bbox": [ + 46, + 460, + 286, + 582 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 460, + 286, + 582 + ], + "type": "inline_equation", + "content": "UC" + }, + { + "bbox": [ + 46, + 460, + 286, + 582 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 460, + 286, + 582 + ], + "type": "inline_equation", + "content": "UA" + }, + { + "bbox": [ + 46, + 460, + 286, + 582 + ], + "type": "text", + "content": "; for productivity, the splits are by caption complexity " + }, + { + "bbox": [ + 46, + 460, + 286, + 582 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 46, + 460, + 286, + 582 + ], + "type": "text", + "content": " (denoted " + }, + { + "bbox": [ + 46, + 460, + 286, + 582 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{test}^{HN,n}" + }, + { + "bbox": [ + 46, + 460, + 286, + 582 + ], + "type": "text", + "content": "). Each retrieval task is between one image and its ground truth caption plus " + }, + { + "bbox": [ + 46, + 460, + 286, + 582 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 46, + 460, + 286, + 582 + ], + "type": "text", + "content": " hard negatives of a single type (see Appendix). We adopt commonly used retrieval metrics Recall@1, 3, 5 and Average Recall@K. For " + }, + { + "bbox": [ + 46, + 460, + 286, + 582 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{test}^{RAW}" + }, + { + "bbox": [ + 46, + 460, + 286, + 582 + ], + "type": "text", + "content": ", retrieval experiments are described in the Appendix." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 587, + 185, + 601 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 587, + 185, + 601 + ], + "spans": [ + { + "bbox": [ + 47, + 587, + 185, + 601 + ], + "type": "text", + "content": "5.1. Systematicity evaluation" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 605, + 289, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 289, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 289, + 714 + ], + "type": "text", + "content": "Model performance on the " + }, + { + "bbox": [ + 46, + 605, + 289, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{test}^{HN}" + }, + { + "bbox": [ + 46, + 605, + 289, + 714 + ], + "type": "text", + "content": " dataset for systematicity decreases monotonically when compounds are unseen. We first observe a monotonic decrease in recall@1 from the Seen Compounds to the Unseen Compounds split on the systematicity " + }, + { + "bbox": [ + 46, + 605, + 289, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{test}^{HN}" + }, + { + "bbox": [ + 46, + 605, + 289, + 714 + ], + "type": "text", + "content": " set consisting of both HN-ATOM and HN-COMP (Figure 4 left). This drop is relatively small " + }, + { + "bbox": [ + 46, + 605, + 289, + 714 + ], + "type": "inline_equation", + "content": "(1 - 5\\%)" + }, + { + "bbox": [ + 46, + 605, + 289, + 714 + ], + "type": "text", + "content": " for the CC-12M and YFCC-15M trained models and the most pronounced for models trained on the largest dataset LAION-400M [62], with the decrease reaching " + }, + { + "bbox": [ + 46, + 605, + 289, + 714 + ], + "type": "inline_equation", + "content": "7.9\\%" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 270, + 547, + 344 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 270, + 547, + 344 + ], + "spans": [ + { + "bbox": [ + 304, + 270, + 547, + 344 + ], + "type": "text", + "content": "for the ViT-B-32 model. However, CC-12M and YFCC-15M models also significantly underperform LAION-400M models in general, meaning that small drops between sets may be due to overall poor performance rather than improved systematic generalization. In comparison, human oracle experiments generalize with " + }, + { + "bbox": [ + 304, + 270, + 547, + 344 + ], + "type": "inline_equation", + "content": "100\\%" + }, + { + "bbox": [ + 304, + 270, + 547, + 344 + ], + "type": "text", + "content": " accuracy to " + }, + { + "bbox": [ + 304, + 270, + 547, + 344 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{test}^{HN}" + }, + { + "bbox": [ + 304, + 270, + 547, + 344 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 344, + 547, + 416 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 344, + 547, + 416 + ], + "spans": [ + { + "bbox": [ + 304, + 344, + 547, + 416 + ], + "type": "text", + "content": "Similar to the overall results, there is also a consistent discrepancy between the SC and UC split on the " + }, + { + "bbox": [ + 304, + 344, + 547, + 416 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{test}^{HN}" + }, + { + "bbox": [ + 304, + 344, + 547, + 416 + ], + "type": "text", + "content": " subset consisting of HN-ATOM only (Figure 4 center). This drop is consistently smaller (3 - 6%) for models trained on CC-12M and YFCC-15M, but pronounced (6% or higher, reaching 9.4% drop for ViT-B-32) for LAION-400M models." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 418, + 548, + 538 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 418, + 548, + 538 + ], + "spans": [ + { + "bbox": [ + 304, + 418, + 548, + 538 + ], + "type": "text", + "content": "On the HN-COMP subset (Figure 4 right), we find little to no difference in performance between the " + }, + { + "bbox": [ + 304, + 418, + 548, + 538 + ], + "type": "inline_equation", + "content": "SC" + }, + { + "bbox": [ + 304, + 418, + 548, + 538 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 418, + 548, + 538 + ], + "type": "inline_equation", + "content": "UC" + }, + { + "bbox": [ + 304, + 418, + 548, + 538 + ], + "type": "text", + "content": " split. We hypothesize that this is due to the lower difficulty of the HN-COMP hard negatives, as they introduce more foils to the caption, are always longer than the ground truth, and thus offer more opportunities for the model to correctly distinguish the ground truth. This hypothesis is supported by the fact that Recall@1 values on HN-COMP are similar or higher than the ones on HN-ATOM even though the HN-COMP retrieval set size is larger than that of HN-ATOM." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 550, + 440, + 562 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 550, + 440, + 562 + ], + "spans": [ + { + "bbox": [ + 306, + 550, + 440, + 562 + ], + "type": "text", + "content": "5.2. Productivity evaluation" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 570, + 548, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 548, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 548, + 714 + ], + "type": "text", + "content": "Models' performance decreases with complexity on HN-ATOM and HN-SWAP negatives. At small complexities such as " + }, + { + "bbox": [ + 304, + 570, + 548, + 714 + ], + "type": "inline_equation", + "content": "n = 4" + }, + { + "bbox": [ + 304, + 570, + 548, + 714 + ], + "type": "text", + "content": ", we observe that model retrieval quality is well above random chance (Figure 5). However, as caption complexity increases, we observe a steady decrease in performance, nearing random chance for HN-SWAP and dipping below it for HN-ATOM negatives. Similarly, we find that the same downward trend persists for other vision-language foundation models (Figure 6). Importantly, the downward trend occurs for FLAVA and ALBEF even though their training set contains Visual Genome images. We note that for HN-NEG negatives, the OpenAI CLIP models do not adhere" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "10916" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 67, + 71, + 521, + 203 + ], + "blocks": [ + { + "bbox": [ + 67, + 71, + 521, + 203 + ], + "lines": [ + { + "bbox": [ + 67, + 71, + 521, + 203 + ], + "spans": [ + { + "bbox": [ + 67, + 71, + 521, + 203 + ], + "type": "image", + "image_path": "39d5d0840bc78533ebbca857036d61b7ca5b8948ee453f276cbe05f9bda4d0ca.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 214, + 547, + 249 + ], + "lines": [ + { + "bbox": [ + 46, + 214, + 547, + 249 + ], + "spans": [ + { + "bbox": [ + 46, + 214, + 547, + 249 + ], + "type": "text", + "content": "Figure 8. A plot showing the correlation between zero-shot top-1 accuracy on ImageNet and Recall@1 on CREPE's productivity hard negative sets for complexities of 4, 8 and 12. Overall, we find no correlation between ImageNet accuracy and Recall@1 on our productivity sets. The strongest correlations are " + }, + { + "bbox": [ + 46, + 214, + 547, + 249 + ], + "type": "inline_equation", + "content": "R^2" + }, + { + "bbox": [ + 46, + 214, + 547, + 249 + ], + "type": "text", + "content": " scores of 0.284 for HN-NEG negatives on " + }, + { + "bbox": [ + 46, + 214, + 547, + 249 + ], + "type": "inline_equation", + "content": "n = 12" + }, + { + "bbox": [ + 46, + 214, + 547, + 249 + ], + "type": "text", + "content": " and of 0.222 for HN-ATOM negatives on " + }, + { + "bbox": [ + 46, + 214, + 547, + 249 + ], + "type": "inline_equation", + "content": "n = 8" + }, + { + "bbox": [ + 46, + 214, + 547, + 249 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 269, + 289, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 269, + 289, + 365 + ], + "spans": [ + { + "bbox": [ + 46, + 269, + 289, + 365 + ], + "type": "text", + "content": "to the downward trend, achieving their lowest scores for the lowest complexity. Their performances on higher complexities, however, show great variation. In short, our conclusion is that vision-language foundation models struggle with productivity. Our results on models released by OpenCLIP [28] as well as other vision-language foundation models demonstrate the challenge of differentiating between atomic and swapping foils is exacerbated by caption complexity." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 365, + 289, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 365, + 289, + 472 + ], + "spans": [ + { + "bbox": [ + 46, + 365, + 289, + 472 + ], + "type": "text", + "content": "We see no effect of dataset size on productivity. We do not observe a clear advantage for larger pretraining datasets in our productivity evaluation. For atomic and swapping foils, we see similar performance for models trained on the three datasets, with slightly worse performance on atomic foils for the CC-12M trained models. However, on negation hard negatives (Figure 5), we see variable performance across training sets, with CC-12M models outperforming larger models trained on larger datasets YFCC and LAION." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 479, + 162, + 491 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 479, + 162, + 491 + ], + "spans": [ + { + "bbox": [ + 47, + 479, + 162, + 491 + ], + "type": "text", + "content": "5.3. Effect of model size" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 497, + 288, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 497, + 288, + 569 + ], + "spans": [ + { + "bbox": [ + 46, + 497, + 288, + 569 + ], + "type": "text", + "content": "We find no trends relating compositionality to model size. Overall, we note that the LAION trained models (which are both larger models and trained on larger datasets) achieve significantly better absolute performances than smaller models. However, model's systematicity and productivity remain indifferent to the size of the model itself (Figures 4 and 5)." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 575, + 259, + 588 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 575, + 259, + 588 + ], + "spans": [ + { + "bbox": [ + 47, + 575, + 259, + 588 + ], + "type": "text", + "content": "5.4. Correlation with ImageNet performance" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 594, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 289, + 715 + ], + "type": "text", + "content": "We find that zero-shot ImageNet top-1 accuracy strongly correlates with models' Recall@1 on the systematicity retrieval set. Specifically, we acquire " + }, + { + "bbox": [ + 46, + 594, + 289, + 715 + ], + "type": "inline_equation", + "content": "R^2" + }, + { + "bbox": [ + 46, + 594, + 289, + 715 + ], + "type": "text", + "content": " scores of 0.984 and 0.877 for the " + }, + { + "bbox": [ + 46, + 594, + 289, + 715 + ], + "type": "inline_equation", + "content": "SC" + }, + { + "bbox": [ + 46, + 594, + 289, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 594, + 289, + 715 + ], + "type": "inline_equation", + "content": "UC" + }, + { + "bbox": [ + 46, + 594, + 289, + 715 + ], + "type": "text", + "content": " splits respectively (Figure 7). However, this correlation does not imply that models' zero-shot ImageNet performance correlates with systematic generalization, which is instead indicated by small or no difference between the " + }, + { + "bbox": [ + 46, + 594, + 289, + 715 + ], + "type": "inline_equation", + "content": "SC" + }, + { + "bbox": [ + 46, + 594, + 289, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 594, + 289, + 715 + ], + "type": "inline_equation", + "content": "UC" + }, + { + "bbox": [ + 46, + 594, + 289, + 715 + ], + "type": "text", + "content": " splits. On our productivity dataset, we do not observe such a strong correlation, where the highest " + }, + { + "bbox": [ + 46, + 594, + 289, + 715 + ], + "type": "inline_equation", + "content": "R^2" + }, + { + "bbox": [ + 46, + 594, + 289, + 715 + ], + "type": "text", + "content": " score is 0.284 for HN-NEG negatives on a complexity" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 269, + 547, + 305 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 269, + 547, + 305 + ], + "spans": [ + { + "bbox": [ + 304, + 269, + 547, + 305 + ], + "type": "text", + "content": "of " + }, + { + "bbox": [ + 304, + 269, + 547, + 305 + ], + "type": "inline_equation", + "content": "n = 12" + }, + { + "bbox": [ + 304, + 269, + 547, + 305 + ], + "type": "text", + "content": " (Figure 8). As such, we can infer that successful zero-shot performance on ImageNet does not necessarily lead to better performance on our productivity sets." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 316, + 375, + 328 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 316, + 375, + 328 + ], + "spans": [ + { + "bbox": [ + 306, + 316, + 375, + 328 + ], + "type": "text", + "content": "6. Discussion" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 335, + 547, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 335, + 547, + 552 + ], + "spans": [ + { + "bbox": [ + 304, + 335, + 547, + 552 + ], + "type": "text", + "content": "Limitations. First, although our data validation protocols verified our generated hard negatives for productivity as high-quality, approximately " + }, + { + "bbox": [ + 304, + 335, + 547, + 552 + ], + "type": "inline_equation", + "content": "70\\%" + }, + { + "bbox": [ + 304, + 335, + 547, + 552 + ], + "type": "text", + "content": " of HN-SWAP and of HN-NEG negatives were rated as correct. While this does not invalidate our key productivity result, this noise is a limitation of CREPE and could hinder future evaluations once foundation models begin performing better. Second, our evaluation only covers a limited set of vision-language foundation models that were trained with contrastive loss. Additionally, given the computational requirements associated with training a foundation model, our experiments centered around model architectures that were already available publicly. We hope that future foundation models are evaluated with our publicly available CREPE benchmark. Third, while we observe text-to-image and image-to-text retrieval to have similar trends for our systematicity experiments, we lack text-to-image datasets with hard negatives. Future work can explore mechanisms to generate counterfactual negative images." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 564, + 548, + 710 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 564, + 548, + 710 + ], + "spans": [ + { + "bbox": [ + 304, + 564, + 548, + 710 + ], + "type": "text", + "content": "Conclusion. We present CREPE, a collection of text-to-image and image-to-text retrieval datasets for evaluating pretrained vision-language models' systematicity and productivity. We demonstrate that models struggle with compositionality along both axes, with performance drops across different compositional splits and increasing complexity. We expect that CREPE will provide a more systematic evaluation to benchmark the emergence of compositionality as future models improve. Finally, researchers can leverage our hard-negative generation process to create training batches with hard negatives to incentivize vision-language compositionality." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "10917" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 289, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 53, + 91, + 289, + 167 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 289, + 167 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 289, + 167 + ], + "type": "text", + "content": "[1] Arjun Akula, Spandana Gella, Yaser Al-Onaizan, Song-Chun Zhu, and Siva Reddy. Words aren't enough, their order matters: On the robustness of grounding visual referring expressions. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 6555-6565, Online, July 2020. Association for Computational Linguistics." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 170, + 287, + 191 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 170, + 287, + 191 + ], + "spans": [ + { + "bbox": [ + 53, + 170, + 287, + 191 + ], + "type": "text", + "content": "[2] Jacob Andreas, Marcus Rohrbach, Trevor Darrell, and Dan Klein. Neural module networks, 2015." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 194, + 288, + 247 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 194, + 288, + 247 + ], + "spans": [ + { + "bbox": [ + 53, + 194, + 288, + 247 + ], + "type": "text", + "content": "[3] Dzmitry Bahdanau, Harm de Vries, Timothy J O'Donnell, Shikhar Murty, Philippe Beaudoin, Yoshua Bengio, and Aaron Courville. Closure: Assessing systematic generalization of clevr models. arXiv preprint arXiv:1912.05783, 2019." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 250, + 288, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 250, + 288, + 327 + ], + "spans": [ + { + "bbox": [ + 53, + 250, + 288, + 327 + ], + "type": "text", + "content": "[4] Ben Bogin, Shivanshu Gupta, Matt Gardner, and Jonathan Berant. COVR: A test-bed for visually grounded compositional generalization with real images. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 9824-9846, Online and Punta Cana, Dominican Republic, Nov. 2021. Association for Computational Linguistics." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 328, + 288, + 383 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 328, + 288, + 383 + ], + "spans": [ + { + "bbox": [ + 53, + 328, + 288, + 383 + ], + "type": "text", + "content": "[5] Rishi Bommasani, Drew A Hudson, Ehsan Adeli, Russ Altman, Simran Arora, Sydney von Arx, Michael S Bernstein, Jeannette Bohg, Antoine Bosselut, Emma Brunskill, et al. On the opportunities and risks of foundation models. arXiv preprint arXiv:2108.07258, 2021." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 384, + 288, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 384, + 288, + 407 + ], + "spans": [ + { + "bbox": [ + 53, + 384, + 288, + 407 + ], + "type": "text", + "content": "[6] Léon Bottou. From machine learning to machine reasoning. Machine learning, 94(2):133-149, 2014." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 407, + 288, + 462 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 407, + 288, + 462 + ], + "spans": [ + { + "bbox": [ + 53, + 407, + 288, + 462 + ], + "type": "text", + "content": "[7] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. Advances in neural information processing systems, 33:1877-1901, 2020." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 464, + 288, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 464, + 288, + 520 + ], + "spans": [ + { + "bbox": [ + 53, + 464, + 288, + 520 + ], + "type": "text", + "content": "[8] Soravit Changpinyo, Piyush Sharma, Nan Ding, and Radu Soricut. Conceptual 12m: Pushing web-scale image-text pretraining to recognize long-tail visual concepts. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3558-3568, 2021." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 521, + 288, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 521, + 288, + 553 + ], + "spans": [ + { + "bbox": [ + 53, + 521, + 288, + 553 + ], + "type": "text", + "content": "[9] Zhenfang Chen, Peng Wang, Lin Ma, Kwan-Yee K. Wong, and Qi Wu. Cops-ref: A new dataset and task on compositional referring expression comprehension, 2020." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 555, + 288, + 587 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 555, + 288, + 587 + ], + "spans": [ + { + "bbox": [ + 48, + 555, + 288, + 587 + ], + "type": "text", + "content": "[10] Colin Conwell and Tomer Ullman. Testing relational understanding in text-guided image generation. arXiv preprint arXiv:2208.00005, 2022." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 590, + 228, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 590, + 228, + 601 + ], + "spans": [ + { + "bbox": [ + 48, + 590, + 228, + 601 + ], + "type": "text", + "content": "[11] MJ Cresswell. Logics and languages. 1973." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 602, + 288, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 602, + 288, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 602, + 288, + 689 + ], + "type": "text", + "content": "[12] Yuchen Cui, Scott Niekum, Abhinav Gupta, Vikash Kumar, and Aravind Rajeswaran. Can foundation models perform zero-shot task specification for robot manipulation? In Roya Firoozi, Negar Mehr, Esen Yel, Rika Antonova, Jeannette Bohg, Mac Schwager, and Mykel Kochenderfer, editors, Proceedings of The 4th Annual Learning for Dynamics and Control Conference, volume 168 of Proceedings of Machine Learning Research, pages 893-905. PMLR, 23-24 Jun 2022." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 691, + 288, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 288, + 713 + ], + "type": "text", + "content": "[13] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 712 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 94 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 94 + ], + "type": "text", + "content": "transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 95, + 547, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 95, + 547, + 171 + ], + "spans": [ + { + "bbox": [ + 307, + 95, + 547, + 171 + ], + "type": "text", + "content": "[14] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations, 2021." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 172, + 545, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 172, + 545, + 205 + ], + "spans": [ + { + "bbox": [ + 307, + 172, + 545, + 205 + ], + "type": "text", + "content": "[15] Jiali Duan, Liquen Chen, Son Tran, Jinyu Yang, Yi Xu, Belinda Zeng, and Trishul Chilimbi. Multi-modal alignment using representation codebook, 2022." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 206, + 545, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 206, + 545, + 239 + ], + "spans": [ + { + "bbox": [ + 307, + 206, + 545, + 239 + ], + "type": "text", + "content": "[16] Sepideh Esmaeilpour, Bing Liu, Eric Robertson, and Lei Shu. Zero-shot out-of-distribution detection based on the pre-trained model clip, 2021." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 239, + 545, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 239, + 545, + 293 + ], + "spans": [ + { + "bbox": [ + 307, + 239, + 545, + 293 + ], + "type": "text", + "content": "[17] Mona Gandhi, Mustafa O. Gul, Eva Prakash, Madeleine Grunde-McLaughlin, Ranjay Krishna, and Maneesh Agrawala. Measuring compositional consistency for video question answering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 294, + 545, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 294, + 545, + 327 + ], + "spans": [ + { + "bbox": [ + 307, + 294, + 545, + 327 + ], + "type": "text", + "content": "[18] Shashank Goel, Hritik Bansal, Sumit Bhatia, Ryan A. Rossi, Vishwa Vinay, and Aditya Grover. Cyclip: Cyclic contrastive language-image pretraining, 2022." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 327, + 545, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 327, + 545, + 380 + ], + "spans": [ + { + "bbox": [ + 307, + 327, + 545, + 380 + ], + "type": "text", + "content": "[19] Madeleine Grunde-McLaughlin, Ranjay Krishna, and Maneesh Agrawala. Agqa: A benchmark for compositional spatio-temporal reasoning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2021." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 383, + 545, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 383, + 545, + 426 + ], + "spans": [ + { + "bbox": [ + 307, + 383, + 545, + 426 + ], + "type": "text", + "content": "[20] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 426, + 545, + 448 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 426, + 545, + 448 + ], + "spans": [ + { + "bbox": [ + 307, + 426, + 545, + 448 + ], + "type": "text", + "content": "[21] Lisa Anne Hendricks and Aida Nematzadeh. Probing image-language transformers for verb understanding, 2021." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 449, + 545, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 449, + 545, + 514 + ], + "spans": [ + { + "bbox": [ + 307, + 449, + 545, + 514 + ], + "type": "text", + "content": "[22] Jack Hessel, Ari Holtzman, Maxwell Forbes, Ronan Le Bras, and Yejin Choi. CLIPScore: A reference-free evaluation metric for image captioning. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 7514-7528, Online and Punta Cana, Dominican Republic, Nov. 2021. Association for Computational Linguistics." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 515, + 545, + 569 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 515, + 545, + 569 + ], + "spans": [ + { + "bbox": [ + 307, + 515, + 545, + 569 + ], + "type": "text", + "content": "[23] Irina Higgins, Nicolas Sonnerat, Loic Matthey, Arka Pal, Christopher P Burgess, Matko Bosnjak, Murray Shanahan, Matthew Botvinick, Demis Hassabis, and Alexander Lerchner. Scan: Learning hierarchical compositional visual concepts. arXiv preprint arXiv:1707.03389, 2017." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 570, + 545, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 570, + 545, + 602 + ], + "spans": [ + { + "bbox": [ + 307, + 570, + 545, + 602 + ], + "type": "text", + "content": "[24] Hexiang Hu, Ishan Misra, and Laurens van der Maaten. Evaluating text-to-image matching using binary image selection (bison), 2019." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 602, + 545, + 625 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 602, + 545, + 625 + ], + "spans": [ + { + "bbox": [ + 307, + 602, + 545, + 625 + ], + "type": "text", + "content": "[25] Drew A. Hudson and Christopher D. Manning. Compositional attention networks for machine reasoning, 2018." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 626, + 545, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 626, + 545, + 669 + ], + "spans": [ + { + "bbox": [ + 307, + 626, + 545, + 669 + ], + "type": "text", + "content": "[26] Drew A Hudson and Christopher D Manning. Gqa: A new dataset for real-world visual reasoning and compositional question answering. Conference on Computer Vision and Pattern Recognition (CVPR), 2019." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 670, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 670, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 307, + 670, + 545, + 712 + ], + "type": "text", + "content": "[27] Dieuwke Hupkes, Verna Dankers, Mathijs Mul, and Elia Bruni. Compositionality decomposed: How do neural networks generalise? Journal of Artificial Intelligence Research, 67:757-795, 2020." + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "10918" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 288, + 127 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 288, + 127 + ], + "type": "text", + "content": "[28] Gabriel Ilharco, Mitchell Wortsman, Ross Wightman, Cade Gordon, Nicholas Carlini, Rohan Taori, Achal Dave, Vaishaal Shankar, Hongseok Namkoong, John Miller, Hannaneh Hajishirzi, Ali Farhadi, and Ludwig Schmidt. OpenCLIP, 7 2021." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 130, + 288, + 175 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 130, + 288, + 175 + ], + "spans": [ + { + "bbox": [ + 48, + 130, + 288, + 175 + ], + "type": "text", + "content": "[29] Ajay Jain, Matthew Tancik, and Pieter Abbeel. Putting nerf on a diet: Semantically consistent few-shot view synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 5885-5894, October 2021." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 176, + 288, + 209 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 176, + 288, + 209 + ], + "spans": [ + { + "bbox": [ + 48, + 176, + 288, + 209 + ], + "type": "text", + "content": "[30] Theo MV Janssen and Barbara H Partee. Compositionality. In Handbook of logic and language, pages 417-473. Elsevier, 1997." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 212, + 288, + 266 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 212, + 288, + 266 + ], + "spans": [ + { + "bbox": [ + 48, + 212, + 288, + 266 + ], + "type": "text", + "content": "[31] Jingwei Ji, Ranjay Krishna, Li Fei-Fei, and Juan Carlos Niebles. Action genome: Actions as compositions of spatiotemporal scene graphs. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10236–10247, 2020." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 269, + 288, + 334 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 269, + 288, + 334 + ], + "spans": [ + { + "bbox": [ + 48, + 269, + 288, + 334 + ], + "type": "text", + "content": "[32] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In International Conference on Machine Learning, pages 4904-4916. PMLR, 2021." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 337, + 288, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 337, + 288, + 380 + ], + "spans": [ + { + "bbox": [ + 48, + 337, + 288, + 380 + ], + "type": "text", + "content": "[33] Justin Johnson, Bharath Hariharan, Laurens van der Maaten, Li Fei-Fei, C. Lawrence Zitnick, and Ross Girshick. Clevr: A diagnostic dataset for compositional language and elementary visual reasoning, 2016." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 383, + 288, + 460 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 383, + 288, + 460 + ], + "spans": [ + { + "bbox": [ + 48, + 383, + 288, + 460 + ], + "type": "text", + "content": "[34] Daniel Keysers, Nathanael Scharli, Nathan Scales, Hylke Buisman, Daniel Furrer, Sergii Kashubin, Nikola Momchev, Danila Sinopalnikov, Lukasz Stafiniak, Tibor Tihon, Dmitry Tsarkov, Xiao Wang, Marc van Zee, and Olivier Bousquet. Measuring compositional generalization: A comprehensive method on realistic data. In International Conference on Learning Representations, 2020." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 462, + 288, + 518 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 462, + 288, + 518 + ], + "spans": [ + { + "bbox": [ + 48, + 462, + 288, + 518 + ], + "type": "text", + "content": "[35] Ranjay Krishna, Yuke Zhu, Oliver Groth, Justin Johnson, Kenji Hata, Joshua Kravitz, Stephanie Chen, Yannis Kalantidis, Li-Jia Li, David A Shamma, Michael Bernstein, and Li Fei-Fei. Visual genome: Connecting language and vision using crowdsourced dense image annotations. 2016." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 520, + 288, + 563 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 520, + 288, + 563 + ], + "spans": [ + { + "bbox": [ + 48, + 520, + 288, + 563 + ], + "type": "text", + "content": "[36] Brenden Lake and Marco Baroni. Generalization without systematicity: On the compositional skills of sequence-to-sequence recurrent networks. In International conference on machine learning, pages 2873-2882. PMLR, 2018." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 566, + 288, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 566, + 288, + 609 + ], + "spans": [ + { + "bbox": [ + 48, + 566, + 288, + 609 + ], + "type": "text", + "content": "[37] Brenden M. Lake, Tomer D. Ullman, Joshua B. Tenenbaum, and Samuel J. Gershman. Building machines that learn and think like people. Behavioral and Brain Sciences, 40:e253, 2017." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 612, + 288, + 655 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 612, + 288, + 655 + ], + "spans": [ + { + "bbox": [ + 48, + 612, + 288, + 655 + ], + "type": "text", + "content": "[38] Boyi Li, Kilian Q Weinberger, Serge Belongie, Vladlen Koltun, and Rene Ranftl. Language-driven semantic segmentation. In International Conference on Learning Representations, 2022." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 658, + 288, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 288, + 713 + ], + "type": "text", + "content": "[39] Dongxu Li, Junnan Li, Hongdong Li, Juan Carlos Niebles, and Steven C.H. Hoi. Align and prompt: Video-and-language pre-training with entity prompts. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4943-4953, 2022." + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 713 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 307, + 73, + 547, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 547, + 106 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 547, + 106 + ], + "type": "text", + "content": "[40] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In ICML, 2022." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 307, + 108, + 547, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 108, + 547, + 173 + ], + "spans": [ + { + "bbox": [ + 307, + 108, + 547, + 173 + ], + "type": "text", + "content": "[41] Junnan Li, Ramprasaath R. Selvaraju, Akhilesh Deepak Gotmare, Shafiq Joty, Caiming Xiong, and Steven Hoi. Align before fuse: Vision and language representation learning with momentum distillation. In A. Beygelzimer, Y. Dauphin, P. Liang, and J. Wortman Vaughan, editors, Advances in Neural Information Processing Systems, 2021." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 175, + 547, + 240 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 175, + 547, + 240 + ], + "spans": [ + { + "bbox": [ + 307, + 175, + 547, + 240 + ], + "type": "text", + "content": "[42] Wei Li, Can Gao, Guocheng Niu, Xinyan Xiao, Hao Liu, Jiachen Liu, Hua Wu, and Haifeng Wang. UNIMO-2: End-to-end unified vision-language grounded learning. In Findings of the Association for Computational Linguistics: ACL 2022, pages 3187–3201, Dublin, Ireland, May 2022. Association for Computational Linguistics." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 241, + 547, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 241, + 547, + 285 + ], + "spans": [ + { + "bbox": [ + 307, + 241, + 547, + 285 + ], + "type": "text", + "content": "[43] Yehao Li, Yingwei Pan, Ting Yao, and Tao Mei. Comprehending and ordering semantics for image captioning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 286, + 547, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 286, + 547, + 330 + ], + "spans": [ + { + "bbox": [ + 308, + 286, + 547, + 330 + ], + "type": "text", + "content": "[44] Zhiheng Li, Martin Renqiang Min, Kai Li, and Chenliang Xu. Stylet2i: Toward compositional and high-fidelity text-to-image synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 331, + 547, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 331, + 547, + 386 + ], + "spans": [ + { + "bbox": [ + 308, + 331, + 547, + 386 + ], + "type": "text", + "content": "[45] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In European conference on computer vision, pages 740-755. Springer, 2014." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 388, + 547, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 388, + 547, + 431 + ], + "spans": [ + { + "bbox": [ + 308, + 388, + 547, + 431 + ], + "type": "text", + "content": "[46] Xihui Liu, Dong Huk Park, Samaneh Azadi, Gong Zhang, Arman Chopikyan, Yuxiao Hu, Humphrey Shi, Anna Rohrbach, and Trevor Darrell. More control for free! image synthesis with semantic diffusion guidance, 2021." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 433, + 547, + 486 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 433, + 547, + 486 + ], + "spans": [ + { + "bbox": [ + 308, + 433, + 547, + 486 + ], + "type": "text", + "content": "[47] Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettle-moyer, and Veselin Stoyanov. Roberta: A robustly optimized bert pretraining approach. arXiv preprint arXiv:1907.11692, 2019." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 488, + 547, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 488, + 547, + 521 + ], + "spans": [ + { + "bbox": [ + 308, + 488, + 547, + 521 + ], + "type": "text", + "content": "[48] Haoyu Lu, Nanyi Fei, Yuqi Huo, Yizhao Gao, Zhiwu Lu, and Ji-Rong Wen. Cots: Collaborative two-stream vision-language pre-training model for cross-modal retrieval, 2022." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 523, + 547, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 523, + 547, + 544 + ], + "spans": [ + { + "bbox": [ + 308, + 523, + 547, + 544 + ], + "type": "text", + "content": "[49] George A Miller. Wordnet: a lexical database for english. Communications of the ACM, 38(11):39-41, 1995." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 546, + 547, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 546, + 547, + 568 + ], + "spans": [ + { + "bbox": [ + 308, + 546, + 547, + 568 + ], + "type": "text", + "content": "[50] Medhini Narasimhan, Anna Rohrbach, and Trevor Darrell. Clip-it! language-guided video summarization. 2021." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 569, + 547, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 569, + 547, + 656 + ], + "spans": [ + { + "bbox": [ + 308, + 569, + 547, + 656 + ], + "type": "text", + "content": "[51] Letitia Parcalabescu, Michele Cafagna, Lilitta Muradjan, Anette Frank, Iacer Calixto, and Albert Gatt. VALID: A task-independent benchmark for vision and language models centered on linguistic phenomena. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 8253-8280, Dublin, Ireland, May 2022. Association for Computational Linguistics." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 658, + 547, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 658, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 658, + 547, + 713 + ], + "type": "text", + "content": "[52] Dong Huk Park, Samaneh Azadi, Xihui Liu, Trevor Darrell, and Anna Rohrbach. Benchmark for compositional text-to-image synthesis. In Thirty-fifth Conference on Neural Information Processing Systems Datasets and Benchmarks Track (Round 1), 2021." + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "10919" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 714 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 288, + 149 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 288, + 149 + ], + "type": "text", + "content": "[53] Jae Sung Park, Sheng Shen, Ali Farhadi, Trevor Darrell, Yejin Choi, and Anna Rohrbach. Exposing the limits of video-text models through contrast sets. In Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 3574-3586, Seattle, United States, July 2022. Association for Computational Linguistics." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 151, + 288, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 151, + 288, + 205 + ], + "spans": [ + { + "bbox": [ + 48, + 151, + 288, + 205 + ], + "type": "text", + "content": "[54] Or Patashnik, Zongze Wu, Eli Shechtman, Daniel Cohen-Or, and Dani Lischinski. Styleclip: Text-driven manipulation of stylegan imagery. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 2085–2094, October 2021." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 208, + 288, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 208, + 288, + 262 + ], + "spans": [ + { + "bbox": [ + 48, + 208, + 288, + 262 + ], + "type": "text", + "content": "[55] Suzanne Petryk, Lisa Dunlap, Keyan Nasseri, Joseph Gonzalez, Trevor Darrell, and Anna Rohrbach. On guiding visual attention with language specification. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 18071-18081, 2022." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 264, + 288, + 329 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 264, + 288, + 329 + ], + "spans": [ + { + "bbox": [ + 48, + 264, + 288, + 329 + ], + "type": "text", + "content": "[56] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763. PMLR, 2021." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 331, + 288, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 331, + 288, + 374 + ], + "spans": [ + { + "bbox": [ + 48, + 331, + 288, + 374 + ], + "type": "text", + "content": "[57] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 2022." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 376, + 288, + 432 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 376, + 288, + 432 + ], + "spans": [ + { + "bbox": [ + 48, + 376, + 288, + 432 + ], + "type": "text", + "content": "[58] Yongming Rao, Wenliang Zhao, Guangyi Chen, Yansong Tang, Zheng Zhu, Guan Huang, Jie Zhou, and Jiwen Lu. Denseclip: Language-guided dense prediction with context-aware prompting. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 434, + 288, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 434, + 288, + 498 + ], + "spans": [ + { + "bbox": [ + 48, + 434, + 288, + 498 + ], + "type": "text", + "content": "[59] Laura Ruis, Jacob Andreas, Marco Baroni, Diane Bouchacourt, and Brenden M. Lake. A benchmark for systematic generalization in grounded language understanding. In Proceedings of the 34th International Conference on Neural Information Processing Systems, NIPS'20, Red Hook, NY, USA, 2020. Curran Associates Inc." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 500, + 288, + 565 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 500, + 288, + 565 + ], + "spans": [ + { + "bbox": [ + 48, + 500, + 288, + 565 + ], + "type": "text", + "content": "[60] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily Denton, Seyed Kamyar Seyed Ghasemipour, Burcu Karagol Ayan, S Sara Mahdavi, Rapha Gontijo Lopes, et al. Photorealistic text-to-image diffusion models with deep language understanding. arXiv preprint arXiv:2205.11487, 2022." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 567, + 288, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 567, + 288, + 621 + ], + "spans": [ + { + "bbox": [ + 48, + 567, + 288, + 621 + ], + "type": "text", + "content": "[61] Maarten Sap, Ronan Le Bras, Emily Allaway, Chandra Bhagavatula, Nicholas Lourie, Hannah Rashkin, Brendan Roof, Noah A. Smith, and Yejin Choi. ATOMIC: an atlas of machine commonsense for if-then reasoning. CoRR, abs/1811.00146, 2018." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 624, + 288, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 624, + 288, + 678 + ], + "spans": [ + { + "bbox": [ + 48, + 624, + 288, + 678 + ], + "type": "text", + "content": "[62] Christoph Schuhmann, Richard Vencu, Romain Beaumont, Robert Kaczmarczyk, Clayton Mullis, Aarush Katta, Theo Coombes, Jenia Jitsev, and Aran Komatsuzaki. Laion-400m: Open dataset of clip-filtered 400 million image-text pairs. arXiv preprint arXiv:2111.02114, 2021." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 680, + 288, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 680, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 48, + 680, + 288, + 714 + ], + "type": "text", + "content": "[63] Sebastian Schuster, Ranjay Krishna, Angel Chang, Li Fei-Fei, and Christopher D Manning. Generating semantically precise scene graphs from textual descriptions for improved image" + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 547, + 713 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "type": "text", + "content": "retrieval. In Proceedings of the fourth workshop on vision and language, pages 70-80, 2015." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 308, + 95, + 547, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 95, + 547, + 139 + ], + "spans": [ + { + "bbox": [ + 308, + 95, + 547, + 139 + ], + "type": "text", + "content": "[64] Piyush Sharma, Nan Ding, Sebastian Goodman, and Radu Soricut. Conceptual captions: A cleaned, hypernymed, image alt-text dataset for automatic image captioning. In Proceedings of ACL, 2018." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 308, + 140, + 547, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 140, + 547, + 216 + ], + "spans": [ + { + "bbox": [ + 308, + 140, + 547, + 216 + ], + "type": "text", + "content": "[65] Ravi Shekhar, Sandro Pezzelle, Yauhen Klimovich, Aurélie Herbelot, Moin Nabi, Enver Sangineto, and Raffaella Bernardi. FOIL it! find one mismatch between image and language caption. In Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 255-265, Vancouver, Canada, July 2017. Association for Computational Linguistics." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 308, + 217, + 547, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 217, + 547, + 259 + ], + "spans": [ + { + "bbox": [ + 308, + 217, + 547, + 259 + ], + "type": "text", + "content": "[66] Sheng Shen, Liunian Harold Li, Hao Tan, Mohit Bansal, Anna Rohrbach, Kai-Wei Chang, Zhewei Yao, and Kurt Keutzer. How much can CLIP benefit vision-and-language tasks? In International Conference on Learning Representations, 2022." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 261, + 547, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 261, + 547, + 293 + ], + "spans": [ + { + "bbox": [ + 308, + 261, + 547, + 293 + ], + "type": "text", + "content": "[67] Hengcan Shi, Munawar Hayat, Yicheng Wu, and Jianfei Cai. Proposalclip: Unsupervised open-category object proposal generation via exploiting clip cues, 2022." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 294, + 547, + 336 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 294, + 547, + 336 + ], + "spans": [ + { + "bbox": [ + 308, + 294, + 547, + 336 + ], + "type": "text", + "content": "[68] Mohit Shridhar, Lucas Manuelli, and Dieter Fox. CIiport: What and where pathways for robotic manipulation. In Proceedings of the 5th Conference on Robot Learning (CoRL), 2021." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 338, + 547, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 338, + 547, + 381 + ], + "spans": [ + { + "bbox": [ + 308, + 338, + 547, + 381 + ], + "type": "text", + "content": "[69] Amanpreet Singh, Ronghang Hu, Vedanuj Goswami, Guillaume Couairon, Wojciech Galuba, Marcus Rohrbach, and Douwe Kiela. Flava: A foundational language and vision alignment model, 2021." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 383, + 547, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 383, + 547, + 415 + ], + "spans": [ + { + "bbox": [ + 308, + 383, + 547, + 415 + ], + "type": "text", + "content": "[70] Kihyuk Sohn. Improved deep metric learning with multiclass n-pair loss objective. Advances in neural information processing systems, 29, 2016." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 415, + 547, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 415, + 547, + 459 + ], + "spans": [ + { + "bbox": [ + 308, + 415, + 547, + 459 + ], + "type": "text", + "content": "[71] Sanjay Subramanian, Will Merrill, Trevor Darrell, Matt Gardner, Sameer Singh, and Anna Rohrbach. Reclip: A strong zero-shot baseline for referring expression comprehension. arXiv preprint arXiv:2204.05991, 2022." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 460, + 547, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 460, + 547, + 525 + ], + "spans": [ + { + "bbox": [ + 308, + 460, + 547, + 525 + ], + "type": "text", + "content": "[72] Alane Suhr, Stephanie Zhou, Ally Zhang, Iris Zhang, Huajun Bai, and Yoav Artzi. A corpus for reasoning about natural language grounded in photographs. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 6418-6428, Florence, Italy, July 2019. Association for Computational Linguistics." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 525, + 547, + 558 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 525, + 547, + 558 + ], + "spans": [ + { + "bbox": [ + 308, + 525, + 547, + 558 + ], + "type": "text", + "content": "[73] Yoad Tewel, Yoav Shalev, Idan Schwartz, and Lior Wolf. Zero-shot image-to-text generation for visual-semantic arithmetic. arXiv preprint arXiv:2111.14447, 2021." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 559, + 547, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 559, + 547, + 602 + ], + "spans": [ + { + "bbox": [ + 308, + 559, + 547, + 602 + ], + "type": "text", + "content": "[74] Bart Thomee, David A Shamma, Gerald Friedland, Benjamin Elizalde, Karl Ni, Douglas Poland, Damian Borth, and Li-Jia Li. Yfcc100m: The new data in multimedia research. Communications of the ACM, 59(2):64-73, 2016." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 603, + 547, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 603, + 547, + 668 + ], + "spans": [ + { + "bbox": [ + 308, + 603, + 547, + 668 + ], + "type": "text", + "content": "[75] Tristan Thrush, Ryan Jiang, Max Bartolo, Amanpreet Singh, Adina Williams, Douwe Kiela, and Candace Ross. Winoground: Probing vision and language models for visio-linguistic compositionality. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5238–5248, 2022." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 669, + 547, + 691 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 669, + 547, + 691 + ], + "spans": [ + { + "bbox": [ + 308, + 669, + 547, + 691 + ], + "type": "text", + "content": "[76] Mengmeng Wang, Jiazheng Xing, and Yong Liu. Actionclip: A new paradigm for video action recognition, 2021." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 692, + 547, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 692, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 692, + 547, + 713 + ], + "type": "text", + "content": "[77] Hao Wu, Jiayuan Mao, Yufeng Zhang, Yuning Jiang, Lei Li, Weiwei Sun, and Wei-Ying Ma. Unified visual-semantic" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "text", + "content": "10920" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 463 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 67, + 72, + 287, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 72, + 287, + 116 + ], + "spans": [ + { + "bbox": [ + 67, + 72, + 287, + 116 + ], + "type": "text", + "content": "embeddings: Bridging vision and language with structured meaning representations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6609-6618, 2019." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 118, + 288, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 118, + 288, + 205 + ], + "spans": [ + { + "bbox": [ + 48, + 118, + 288, + 205 + ], + "type": "text", + "content": "[78] Hu Xu, Gargi Ghosh, Po-Yao Huang, Dmytro Okhonko, Armen Aghajanyan, Florian Metze, Luke Zettlemoyer, and Christoph Feichtenhofer. VideoCLIP: Contrastive pretraining for zero-shot video-text understanding. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 6787-6800, Online and Punta Cana, Dominican Republic, Nov. 2021. Association for Computational Linguistics." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 207, + 288, + 261 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 207, + 288, + 261 + ], + "spans": [ + { + "bbox": [ + 48, + 207, + 288, + 261 + ], + "type": "text", + "content": "[79] Zipeng Xu, Tianwei Lin, Hao Tang, Fu Li, Dongliang He, Nicu Sebe, Radu Timofte, Luc Van Gool, and Errui Ding. Predict, prevent, and evaluate: Disentangled text-driven image manipulation empowered by pre-trained vision-language model. arXiv preprint arXiv:2111.13333, 2021." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 262, + 288, + 306 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 262, + 288, + 306 + ], + "spans": [ + { + "bbox": [ + 48, + 262, + 288, + 306 + ], + "type": "text", + "content": "[80] Jinyu Yang, Jiali Duan, Son Tran, Yi Xu, Sampath Chanda, Liquun Chen, Belinda Zeng, Trishul Chilimbi, and Junzhou Huang. Vision-language pre-training with triple contrastive learning. 2022." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 308, + 288, + 350 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 308, + 288, + 350 + ], + "spans": [ + { + "bbox": [ + 48, + 308, + 288, + 350 + ], + "type": "text", + "content": "[81] Jiahui Yu, Zirui Wang, Vijay Vasudevan, Legg Yeung, Mojtaba Seyedhosseini, and Yonghui Wu. Coca: Contrastive captioners are image-text foundation models. Transactions on Machine Learning Research, 2022." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 353, + 288, + 417 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 353, + 288, + 417 + ], + "spans": [ + { + "bbox": [ + 48, + 353, + 288, + 417 + ], + "type": "text", + "content": "[82] Pengchuan Zhang, Xiujun Li, Xiaowei Hu, Jianwei Yang, Lei Zhang, Lijuan Wang, Yejin Choi, and Jianfeng Gao. Vinvl: Revisiting visual representations in vision-language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5579-5588, June 2021." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 419, + 288, + 463 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 419, + 288, + 463 + ], + "spans": [ + { + "bbox": [ + 48, + 419, + 288, + 463 + ], + "type": "text", + "content": "[83] Yufan Zhou, Ruiyi Zhang, Changyou Chen, Chunyuan Li, Chris Tensmeyer, Tong Yu, Jiuming Gu, Jinhui Xu, and Tong Sun. Lafite: Towards language-free training for text-to-image generation. arXiv preprint arXiv:2111.13792, 2021." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "10921" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/CUDA_ Convolution-Based Unlearnable Datasets/e0179aca-27a9-460d-b895-5e7378133f26_content_list.json b/2023/CUDA_ Convolution-Based Unlearnable Datasets/e0179aca-27a9-460d-b895-5e7378133f26_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..cebe31781cdf7631745ffe096b40ed243caf6c04 --- /dev/null +++ b/2023/CUDA_ Convolution-Based Unlearnable Datasets/e0179aca-27a9-460d-b895-5e7378133f26_content_list.json @@ -0,0 +1,1425 @@ +[ + { + "type": "text", + "text": "CUDA: Convolution-based Unlearnable Datasets", + "text_level": 1, + "bbox": [ + 238, + 130, + 732, + 150 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Vinu Sankar Sadasivan \nUniversity of Maryland", + "bbox": [ + 122, + 181, + 310, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "vinu@umd.edu", + "bbox": [ + 161, + 220, + 271, + 232 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Mahdi Soltanolkotabi \nUniversity of Southern California", + "bbox": [ + 351, + 181, + 617, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "soltanol@usc.edu", + "bbox": [ + 411, + 220, + 557, + 232 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Soheil Feizi \nUniversity of Maryland", + "bbox": [ + 658, + 181, + 846, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "sfeizi@cs.umd.edu", + "bbox": [ + 676, + 220, + 828, + 232 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 268, + 313, + 284 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Large-scale training of modern deep learning models heavily relies on publicly available data on the web. This potentially unauthorized usage of online data leads to concerns regarding data privacy. Recent works aim to make unlearnable data for deep learning models by adding small, specially designed noises to tackle this issue. However, these methods are vulnerable to adversarial training (AT) and/or are computationally heavy. In this work, we propose a novel, model-free, Convolution-based Unlearnable DAtaset (CUDA) generation technique. CUDA is generated using controlled class-wise convolutions with filters that are randomly generated via a private key. CUDA encourages the network to learn the relation between filters and labels rather than informative features for classifying the clean data. We develop some theoretical analysis demonstrating that CUDA can successfully poison Gaussian mixture data by reducing the clean data performance of the optimal Bayes classifier. We also empirically demonstrate the effectiveness of CUDA with various datasets (CIFAR-10, CIFAR100, ImageNet-100, and Tiny-ImageNet), and architectures (ResNet-18, VGG-16, Wide ResNet-34-10, DenseNet-121, DeIT, EfficientNetV2-S, and MobileNetV2). Our experiments show that CUDA is robust to various data augmentations and training approaches such as smoothing, AT with different budgets, transfer learning, and fine-tuning. For instance, training a ResNet-18 on ImageNet-100 CUDA achieves only $8.96\\%$ , $40.08\\%$ , and $20.58\\%$ clean test accuracies with empirical risk minimization (ERM), $L_{\\infty}$ AT, and $L_{2}$ AT, respectively. Here, ERM on the clean training data achieves a clean test accuracy of $80.66\\%$ . CUDA exhibits unlearnability effect with ERM even when only a fraction of the training dataset is perturbed. Furthermore, we also show that CUDA is robust to adaptive defenses designed specifically to break it.", + "bbox": [ + 76, + 300, + 473, + 816 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 844, + 209, + 859 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Modern deep learning training frameworks heavily depend on large-scale datasets for achieving high accuracy.", + "bbox": [ + 76, + 869, + 468, + 902 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "This encourages deep learning practitioners to scrape data from the web for data collection [8, 31, 38, 41]. Since a lot of the data is publicly available online, sometime this scraping of data is unauthorized. For instance, a recent article [15] discloses that a private company trained a commercial face recognition system using over three billion facial images collected from the internet without any user consent. Although such massive data can significantly boost the performance of deep learning models, it raises serious concerns about data privacy and security.", + "bbox": [ + 496, + 268, + 893, + 421 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To prevent the unauthorized usage of personal data, a series of recent papers [10, 17, 53] propose to poison data with additive noise. The idea is to make datasets unlearnable for deep learning models by ensuring that they learn the correspondence between noises and labels. Thereby, they do not learn much useful information about the clean data, significantly degrading their clean test accuracy. However, in recent works [11, 17, 47], these unlearnability methods are shown to be vulnerable to adversarial training (AT) frameworks [34]. Motivated by this problem, Fu et al. [11] developed Robust Error-Minimization (REM) noises to make unlearnable data that is protected from AT. While the authors show the effectiveness of REM in multiple scenarios, we demonstrate that these methods are still not robust against different data augmentations or training settings (see Section 3.2). Furthermore, current unlearnability frameworks [10, 11, 17, 53] are model-dependent and require expensive optimization steps on deep learning models to obtain the additive noises. They also need to train the deep learning models from scratch to obtain noises for each new data set.", + "bbox": [ + 496, + 422, + 895, + 724 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this paper, we propose a novel Convolution-based Unlearnable DAataset (CUDA) generation technique. We address limitations of existing unlearnable data generation techniques in Section 3.2 and motivate our CUDA technique in Section 3.3. For generating CUDA, an attacker randomly generates different convolutional filters for each class in the dataset using a private key or seed value. These filters are used to perform controlled class-wise convolutions on the clean training dataset to obtain CUDA. As we", + "bbox": [ + 496, + 724, + 893, + 861 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "Code available here: https://github.com/vinusankars/Convolution-based-Unlearnability", + "bbox": [ + 500, + 875, + 890, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "3862", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/3c62eaf7fff8338381c2b75a38ca730daa8f02c74e3fd14983becd10a92a993c.jpg", + "image_caption": [ + "Figure 1. CIFAR-10 CUDA images from each of the 10 classes convolved using convolutional filters generated with blur parameter $p_b$ . As seen in the plot, a higher $p_b$ value results in better unlearnability (lower clean test accuracy), but increased blurring. CIFAR-10 CUDA does not break with ERM for all the three $p_b$ values. In Section 5.2, we propose Deconvolution-based Adversarial Training (DAT) that we specifically design to break CUDA. DAT adversarially learns class-wise filters to deconvolve CUDA images. However, CIFAR-10 CUDA with $p_b \\in \\{0.3, 0.5\\}$ does not break with DAT. Therefore, we select $p_b = 0.3$ (lesser blurring) for CIFAR-10 CUDA as discussed in Section 5.1. More details on DAT in supplementary material." + ], + "image_footnote": [], + "bbox": [ + 148, + 95, + 648, + 226 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/fe3c7da284d700e4944361e4bb0eedd4e99e2f82fea68330f5d6b146ab36bfa9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 666, + 92, + 821, + 229 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "describe in Sections 3.3 and 5.2, CUDA generation performs controlled convolutions using a blur parameter $p_b$ to ensure that the semantics of the dataset are preserved (see Figure 1). CUDA generation with a lower blurring parameter $p_b$ adds less perceptible noises to clean samples. A network trained by a defender on CUDA is encouraged to learn the shortcut relation between class-wise convolutional filters and labels rather than useful features for classifying the clean data. Since the seed value for generating the filters are private, its not possible for the defender to obtain clean data from CUDA alone. Additionally, CUDA exhibits unlearnability effect with ERM even when only a fraction of the training dataset is perturbed (see Section 5). While the existing unlearnability works use additive noises, CUDA generation technique enjoys the advantage of introducing multiplicative noises in the Fourier domain due to the convolution theorem (since convolution of signals is the same as element-wise multiplication in the Fourier domain). This lets CUDA generation add higher amounts of noise in the image space, specifically along the edges in images, and makes it resilient to AT with small additive noise budgets. In Figure 2, with the help of t-SNE plots [49], we also find that the noises added by CUDA generation is not linearly separable while they are linearly separable for the existing works on unlearnability [52].", + "bbox": [ + 75, + 364, + 472, + 739 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In Section 4, we theoretically show that CUDA generation can successfully poison Gaussian mixture data by degrading the clean data accuracy of the optimal Bayes classifier. We state our result informally below while the formal version is presented in Theorem 2.", + "bbox": [ + 75, + 741, + 468, + 816 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Theorem 1 (Informal) Let $\\mathcal{D}$ denote a Gaussian mixture data with two modes, $P_{\\mathcal{D}}$ denote the optimal Bayes classifier trained on $\\mathcal{D}$ , and $\\tau_{\\mathcal{D}}(P_{\\mathcal{D}})$ denote the accuracy of the classifier $P_{\\mathcal{D}}$ on $\\mathcal{D}$ . Then, under some assumptions, for every clean data $\\mathcal{D}$ , there is a CUDA $\\tilde{\\mathcal{D}}$ such that", + "bbox": [ + 73, + 824, + 470, + 898 + ], + "page_idx": 1 + }, + { + "type": "equation", + "text": "\n$$\n\\tau_ {\\mathcal {D}} (P _ {\\tilde {\\mathcal {D}}}) < \\tau_ {\\mathcal {D}} (P _ {\\mathcal {D}}).\n$$\n", + "text_format": "latex", + "bbox": [ + 500, + 364, + 635, + 380 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Furthermore, our empirical experiments in Section 5 demonstrate the effectiveness of CUDA under various training scenarios such as ERM with various augmentations and regularizations, AT with different budgets, randomized smoothing [6, 22, 27], transfer learning, and fine-tuning. For instance, training a ResNet-18 on CIFAR-10 CUDA achieves only $18.48\\%$ , $44.4\\%$ , and $51.14\\%$ clean test accuracies with ERM, $L_{\\infty}$ AT, and $L_{2}$ AT, respectively (see Figure 2). Here, ERM on the clean training data achieves a clean test accuracy of $94.66\\%$ . In addition, we also design adaptive defenses to investigate if CUDA breaks with random or adversarial defense mechanisms. We find that CUDA is robust to the adaptive defenses that we specifically design to break it.", + "bbox": [ + 496, + 392, + 890, + 603 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 500, + 617, + 648, + 633 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our CUDA generation technique is intimately related with adversarial and poisoning attacks. We first discuss some of this literature and then explain their relation with CUDA generation.", + "bbox": [ + 496, + 643, + 890, + 703 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Adversarial attacks. Adversarial examples are specially designed examples that can fool deep learning models at test time [4, 12, 23, 24, 45]. The adversary crafts these examples by adding error-maximizing noises to the clean data. Even slightly perturbed data can serve as adversarial examples. AT is a training framework proposed to make deep learning models robust to adversarial examples [19,25,34,51,57]. AT is a min-max optimization problem where the model is trained to minimize loss on adversarial examples that have the maximum loss.", + "bbox": [ + 496, + 704, + 890, + 853 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Poisoning attacks. In data poisoning, an attacker aims to hurt the deep learning model's performance by perturbing the training data [2, 20, 28, 32, 42, 50]. The backdoor attack", + "bbox": [ + 496, + 854, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "3863", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/0551ddd300a3b3d91e6650f1b8561e44da0383cc955cd589d1c107c0388fdfef.jpg", + "image_caption": [ + "Figure 2. In the left column, clean test accuracies of ResNet-18 on clean CIFAR-10 and CIFAR-10 CUDA are shown. The lower the test accuracy, the higher the effectiveness of unlearnability is. We show that CUDA is robust to various training settings such as ERM, $L_{2}$ AT, and $L_{\\infty}$ AT. The error bars represent the standard deviation of test accuracy over 5 independent trials. CUDA is robust to the randomness in its data generation process. In the right column, t-SNEs of the noises generated by REM (top) and CUDA (bottom) for the first 3 classes of CIFAR-10 are plotted." + ], + "image_footnote": [], + "bbox": [ + 220, + 89, + 576, + 257 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/a9c5a7c417aaecaa0f53bc946ffaaa6517a6c3243ee2335214b3699ec499c80f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 583, + 89, + 751, + 253 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "is a special type of poisoning attack where a trigger pattern is injected into clean data at training time [5, 29, 33, 36]. The model trained on this data would misclassify an image with a trigger pattern at test time. Gu et al. [13] and Li et al. [30] use perceptible amounts of noises similar to CUDA for data poisoning. However, backdoor attacks do not affect the performance of the model on clean data [1, 5, 42].", + "bbox": [ + 75, + 369, + 468, + 474 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Recent literature utilize data poisoning to protect data from being used for model training without authorization. Yuan and Wu [53] use neural tangent kernels [18] to generate clean label attacks that can hurt the generalization of deep learning models. Huang et al. [17] show that error-minimizing noise addition can serve as a poisoning technique. Fowl et al. [10] show that error-maximizing noises as well can make strong poison attacks. However, all these poisoning techniques do not offer data protection with AT [47, 52]. Fu et al. [11] proposes a min-min-max optimization technique to generate poisoned data that offers better unlearnability effects with AT.", + "bbox": [ + 75, + 476, + 470, + 656 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Convolution-based Ulearnable DAtaset (CUDA)", + "text_level": 1, + "bbox": [ + 76, + 671, + 468, + 705 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, first we give some preliminaries about unlearnability. Then we discuss the limitations of the existing unlearnability methods. Finally, we propose our CUDA generation technique.", + "bbox": [ + 75, + 714, + 468, + 776 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Preliminaries", + "text_level": 1, + "bbox": [ + 76, + 786, + 218, + 800 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Let $\\{(x_i, y_i)\\}_{i=1}^n \\sim \\mathcal{D}^n$ be the clean training dataset where $\\mathcal{D}$ is the clean data distribution, $x_i \\in \\mathcal{X} \\subset \\mathbb{R}^d$ are the samples, and $y_i \\in \\mathcal{Y}$ are the corresponding labels. Suppose a network is given as $f_\\theta: \\mathcal{X} \\to \\mathcal{Y}$ where $\\theta \\in \\Theta$ is the network parameter and $\\ell: \\mathcal{Y} \\times \\mathcal{Y} \\to \\mathbb{R}$ is the loss function. ERM trains a network using a minimization prob", + "bbox": [ + 75, + 809, + 470, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "lem of the form: $\\min_{\\theta} \\frac{1}{n} \\sum_{i=1}^{n} \\ell(f_{\\theta}(\\pmb{x}_i), y_i)$ . Standard $L_p$ -based AT (for $p \\in \\mathbb{R}^+$ ) solves the following min-max problem: $\\min_{\\theta} \\frac{1}{n} \\sum_{i=1}^{n} \\max_{\\|\\pmb{\\delta}_i\\|_p \\leq \\rho_a} \\ell(f_{\\theta}(\\pmb{x}_i + \\pmb{\\delta}_i), y_i)$ where $\\rho_a$ is the adversarial perturbation radius.", + "bbox": [ + 496, + 368, + 890, + 430 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We will use $\\tau_{\\mathcal{D}}(\\theta)$ , with $\\theta$ the clean model parameter, to denote the clean test accuracy of a model trained on the clean training dataset (i.e., clean model accuracy). In unlearnable dataset generation, an attacker uses an algorithm $\\mathcal{A}:\\mathcal{X}\\rightarrow \\mathcal{X}$ to generate an unlearnable dataset $\\{(\\tilde{\\boldsymbol{x}}_i = \\mathcal{A}(\\boldsymbol {x}_i),y_i)\\}_{i = 1}^n\\sim \\tilde{\\mathcal{D}}^n$ from the clean training data. Here, the attacker assumes access to the full clean training dataset. Moreover, the attacker cannot modify the unlearnable dataset once it is released publicly. A defender trains using the unlearnable dataset to obtain a network $f_{\\tilde{\\theta}}$ . The objective of the attacker is to design an unlearnable dataset such that the defender's model trained on the unlearnable data achieves a clean test accuracy (i.e., unlearnable model accuracy) worse than the clean model accuracy i.e. $\\tau_{\\mathcal{D}}(\\tilde{\\theta})\\ll \\tau_{\\mathcal{D}}(\\theta)$ .", + "bbox": [ + 496, + 430, + 892, + 657 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Limitations of existing works", + "text_level": 1, + "bbox": [ + 500, + 664, + 759, + 681 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Fu et al. [11] show that the previous unlearnability methods including Error-Minimization (EM) [17], Targeted Adversarial Poisoning (TAP) [10], and Neural Tangent Generalization Attack (NTGA) [53] are vulnerable to AT. Hence, they propose a Robust Error-Minimization (REM) [11] method that exploits a min-min-max optimization procedure to generate unlearnable noises. REM first trains a noise generator $f_{\\theta}^{\\prime}$ on data points $\\{(x_i,y_i)\\}_{i = 1}^n$ over a loss function $\\ell$ as follows", + "bbox": [ + 496, + 686, + 890, + 821 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\min _ {\\theta} \\frac {1}{n} \\sum_ {i = 1} ^ {n} \\min _ {\\| \\delta_ {i} ^ {u} \\| \\leq \\rho_ {u}} \\mathbb {E} _ {t \\sim T} \\\\ \\max _ {\\| \\boldsymbol {\\delta} _ {i} ^ {a} \\| \\leq \\rho_ {a}} \\ell \\left(f _ {\\theta} ^ {\\prime} \\left(t \\left(\\boldsymbol {x} _ {i} + \\boldsymbol {\\delta} _ {i} ^ {u}\\right) + \\boldsymbol {\\delta} _ {i} ^ {a}\\right), y\\right). \\tag {1} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 557, + 829, + 890, + 897 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3864", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "table", + "img_path": "images/103ce9779f82f2d79e51779c6f83ecf7ce50dbf8e261e18e59c5816a1ccc451e.jpg", + "table_caption": [ + "Table 1. Time taken for generating unlearnable noises for various datasets using different unlearnability techniques." + ], + "table_footnote": [], + "table_body": "
DatasetEMTAPNTGAREMCUDA
CIFAR-100.4 hr0.5 hr5.2 hrs22.6 hrs10.8 s
CIFAR-1000.4 hr0.5 hr5.2 hrs22.6 hrs15.5 s
ImageNet-1003.9 hrs5.2 hrs14.6 hrs51.2 hrs0.15 hr
", + "bbox": [ + 78, + 128, + 468, + 193 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Here, $T$ is a distribution over a set of transformations $\\{t:$ $\\mathcal{X}\\to \\mathcal{X}\\}$ , $\\rho_{u}$ is the defensive perturbation radius, and $\\rho_{a}$ controls the protection level of REM against AT. After training the noise generator, an unlearnable example $(\\tilde{\\pmb{x}},y)$ is generated via", + "bbox": [ + 76, + 218, + 467, + 294 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\tilde{\\boldsymbol{x}} = \\boldsymbol {x} + \\operatorname *{arg min}_{\\| \\boldsymbol{\\delta}^{u}\\| \\leq \\rho_{u}}\\mathbb{E}_{t\\sim T} \\\\ \\max _ {\\| \\boldsymbol {\\delta} ^ {a} \\| \\leq \\rho_ {a}} \\ell \\left(f _ {\\theta} ^ {\\prime} \\left(t \\left(\\boldsymbol {x} + \\boldsymbol {\\delta} ^ {u}\\right) + \\boldsymbol {\\delta} ^ {a}\\right), y\\right) \\tag {2} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 125, + 305, + 467, + 359 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "First, note that REM is computationally expensive since it needs to generate unlearnability noises through solving optimization equation 2. Moreover, the existing techniques are model-dependent and they require gradient-based training with a neural network to generate unlearnable data. They also require neural network training from scratch for every dataset that is to be made unlearnable. Table 1 shows the amount of time required to generate various unlearnable datasets using NVIDIA® Tesla V100 GPU and 10 CPU cores. CUDA generation is significantly faster than the existing methods since it uses a model-free approach (no training required). Furthermore, REM is sensitive to hyperparameters and norm-budgets of AT since they generate noises with fixed $L_{\\infty}$ norm budgets. For instance, a ResNet-18 trained on clean CIFAR-10 dataset achieves a clean test data accuracy of $94.66\\%$ . $L_{\\infty}$ AT with perturbation radius $\\rho_{a} = 4 / 255$ on REM CIFAR-10 data (generated using $\\rho_{u} = 8 / 255$ and $\\rho_{a} = 4 / 255$ ) achieves only a clean test accuracy of $48.16\\%$ . However, $L_{\\infty}$ AT with perturbation radius $\\rho_{a} = 8 / 255$ and $L_{2}$ AT with perturbation radius $\\rho_{a} = 0.75$ can achieve a clean test accuracy of $78.71\\%$ and $79.65\\%$ , respectively, on the same REM data. We also find that ERM with a ResNet-18 on grayscaled REM CIFAR-10 images can achieve a high test accuracy of $70.76\\%$ on the grayscaled CIFAR-10 test data. This shows that REM relies upon the color space for poisoning clean data. Fu et al. [11] also show that REM noise generated using ResNet-18 is not transferable to DenseNet-121.", + "bbox": [ + 76, + 369, + 468, + 791 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Our method: CUDA", + "text_level": 1, + "bbox": [ + 76, + 801, + 271, + 816 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The major limitations of the previous works are that they are vulnerable to AT, and computationally expensive. We think the major reason for the former limitation is the usage of small additive noises for unlearnability. AT is designed to train in the presence of such additive noises. In", + "bbox": [ + 76, + 825, + 467, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "creasing the budget of the amount of additive noises for unlearnability might destroy the semantics of the images while perturbing them. The latter limitation arises from the fact that these methods are model-dependent and they require multilevel optimizations. Hence, we are motivated to design a compute-efficient unlearnability method that is robust to AT. CUDA technique can perform convolutions to add larger amounts of noises to clean images without destroying its semantics. This can help CUDA to be robust against AT. CUDA uses randomly generated convolution filters for blurring images from each class. This makes a model trained on CUDA to learn shortcut relations between filters and labels. We empirically support these claims in Section 5.2. Additionally, randomly generating the filters makes our technique model-free. Since the keys for generating the filters are private, it is not possible to reverse the blurring effect in CUDA without having access to the corresponding clean images. Hence, we assume that the data publisher deletes the clean images after perturbing them. Moreover, CUDA technique is a novel non-additive noise based poisoning attack that needs to be studied.", + "bbox": [ + 496, + 90, + 890, + 407 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "CUDA uses convolutional filters $s_i \\in \\mathbb{R}^{k \\times k}$ for each class $i \\in [1, K]$ . A random parameter, out of the $k^2$ parameters, in each of the filters is set to have a value of 1. The rest of the filter parameters are randomly initialized from a uniform distribution $\\mathcal{U}(0, p_b)$ using a private seed where $p_b$ is the blur parameter. Blur parameter controls the level of blurring that occurs when an image $x \\in [0, 1]^{d_1 \\times d_2 \\times d_3}$ is convolved with a filter $s_i$ . Here, $d_1, d_2$ , and $d_3$ are the height, width, and number of channels of the image, respectively. For example, a CIFAR-10 image has a dimension of $32 \\times 32 \\times 3$ . The higher $p_b$ is, the higher the blurring effect is. Let $\\hat{x} = x * s_i$ where $x$ belongs to class $i$ . The CUDA data point for $x$ is given by $\\hat{x} = \\hat{x} / \\mathrm{MAX}(\\hat{x})$ . Rescaling is performed to make sure that the CUDA image pixels lie between 0 and 1. We find that the unlearnability effect gets stronger with larger $p_b$ and $k$ values (see supplementary material for details).", + "bbox": [ + 496, + 407, + 890, + 664 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Theory for CUDA", + "text_level": 1, + "bbox": [ + 500, + 679, + 674, + 696 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this section, we define a binary classification setup similar to [19, 35] to theoretically analyze CUDA. Let $\\mathcal{D}$ be a clean dataset modelled by an isotropic Gaussian mixture model given by $\\mathcal{N}(y\\pmb{\\mu}, I)$ , where $y \\in \\{\\pm 1\\}$ is the class label, $\\pmb{\\mu} \\in \\mathbb{R}^d$ , and $I \\in \\mathbb{R}^{d \\times d}$ is the identity matrix. We defer the proofs for all the lemmas and theorem to the supplementary material. The Bayes optimal decision boundary for classifying this Gaussian mixture model is as follows.", + "bbox": [ + 496, + 705, + 890, + 825 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Lemma 1 The Bayes optimal decision boundary for classifying $\\mathcal{D}$ is given by $P(\\pmb{x}) \\equiv \\pmb{\\mu}^{\\top} \\pmb{x} = 0$ . The accuracy of the decision boundary $P$ on the clean dataset $\\mathcal{D}$ (i.e. $\\tau_{\\mathcal{D}}(P)$ ) is equal to $\\phi(\\|\\pmb{\\mu}\\|_2)$ . Here, $\\phi(.)$ represents the CDF of the", + "bbox": [ + 496, + 840, + 890, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "3865", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/c03e7e9239abc12ac49cd209af107cff469828f26ce520550c6089288a1847c4.jpg", + "image_caption": [ + "(a) $\\| \\pmb {\\mu}\\| _2 = 1.0$" + ], + "image_footnote": [], + "bbox": [ + 250, + 88, + 398, + 261 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/f652421a0fd1624ff8771573c4ac561d5b841f9e5159fe0cbe40f0159d577972.jpg", + "image_caption": [ + "(b) $\\| \\pmb{\\mu} \\|_2 = 2.5$", + "Figure 3. We plot the contour plots for clean test accuracy of the CUDA Bayes classifier $\\tilde{P}$ (given as $\\tau_{\\mathcal{D}}(\\tilde{P})$ ) for varying $a_{y}$ parameters where $y \\in \\{\\pm 1\\}$ in Figures 3(a-c). Gaussian mixture model $\\mathcal{N}(y\\mu, I)$ denotes the clean data. In the top row, we use Lemmas 1 & 4 to empirically generate the plots. In the bottom row, we plot the theoretical upper bound of $\\tau_{\\mathcal{D}}(\\tilde{P})$ we obtain using Theorem 2." + ], + "image_footnote": [], + "bbox": [ + 411, + 88, + 558, + 261 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/d105371da4d0456bc3918acd25f6d10a0c36cb54bfb91fe493cf2672ceb5fc0a.jpg", + "image_caption": [ + "(c) $\\| \\pmb{\\mu}\\| _2 = 5.0$" + ], + "image_footnote": [], + "bbox": [ + 571, + 88, + 720, + 260 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "standard normal distribution.", + "text_level": 1, + "bbox": [ + 76, + 359, + 272, + 373 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Now, to obtain $\\tilde{\\mathcal{D}}$ (i.e., CUDA), we perform class-wise 1D convolutions. Here, we consider a class of 1D convolutional filters with kernel size 3 of the form $f_{a} = [a,1,a]$ where $a\\in [0,0.5]$ . The convolution of signal $\\pmb {x}\\in \\mathbb{R}^{d}$ with filter $\\pmb{f}_{a}$ using stride 1 (denoted by $\\pmb {x}*\\pmb{f}_{a}$ ), can be treated as a matrix operation $Ax$ . Here, $A\\in \\mathbb{R}^{d\\times d}$ is a tri-diagonal Toeplitz matrix denoted by $T(d;a,1,a)$ . For CUDA generation, we use class-wise convolution matrices $A_{y} = T(d;a_{y},1,a_{y})$ to perturb a clean data point $(x,y)\\sim \\mathcal{D}$ to an unlearnable data point $(A_yx,y)$ , where $a_{y}\\in [0,0.5]$ for $y\\in \\{\\pm 1\\}$ . Note that in CUDA, the labels remain the same. Next we show that such a perturbed dataset (i.e. CUDA) can be represented as a Gaussian mixture model.", + "bbox": [ + 75, + 385, + 472, + 597 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Lemma 2 A CUDA generated from clean data distribution $\\mathcal{D}$ and the $A_{y}$ 's defined above can be modelled as a Gaussian mixture model with the distribution given by $\\tilde{\\mathcal{D}} = \\mathcal{N}(yA_y\\pmb {\\mu},A_y^2)$ .", + "bbox": [ + 75, + 608, + 468, + 672 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To characterize the decision boundary for CUDA, we need to use some properties of Toeplitz matrices from Noschese et al. [37] given in the following Lemma 3.", + "bbox": [ + 75, + 681, + 468, + 727 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Lemma 3 Any tri-diagonal Toeplitz matrix $A = T(d; a, 1, a) \\in \\mathbb{R}^{d \\times d}$ with $a \\in [0, 0.5]$ can be diagonalized as $A = QDQ$ where $Q_{i,j} = \\left(\\frac{2}{d + 1}\\right)^{1/2} \\sin \\left(\\frac{ij\\pi}{d + 1}\\right)$ and $D$ is a diagonal matrix with $D_{i,i} = 1 + 2a \\cos \\left(\\frac{i\\pi}{d + 1}\\right)$ for $1 \\leq i, j \\leq d$ .", + "bbox": [ + 73, + 737, + 467, + 820 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Next we use Lemma 3 to show that the Bayes optimal decision boundary for classifying $\\tilde{\\mathcal{D}}$ is a quadratic plane.", + "bbox": [ + 75, + 830, + 467, + 861 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Lemma 4 Let $A_{-1} = T(d; a_{-1}, 1, a_{-1})$ and $A_1 = T(d; a_1, 1, a_1)$ . The Bayes optimal decision boundary for", + "bbox": [ + 75, + 869, + 468, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "classifying $\\tilde{\\mathcal{D}}$ is given by $\\tilde{P} (\\pmb {x})\\equiv \\pmb{x}^{\\top}A\\pmb {x} + \\pmb{b}^{\\top}\\pmb {x} + c = 0,$ where $A = A_{-1}^{-2} - A_1^{-2},\\pmb {b} = 2(A_{-1}^{-1} + A_{1}^{-1})\\pmb {\\mu},$ and $c = 2\\sum_{i = 1}^{d}\\left[\\ln (1 + 2a_{-1}\\cos (\\frac{i\\pi}{d + 1})) - \\ln (1 + 2a_{1}\\cos (\\frac{i\\pi}{d + 1}))\\right].$", + "bbox": [ + 498, + 357, + 892, + 412 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Now we state a lemma regarding the tail of Gaussian quadratic forms which plays a crucial role in our main result.", + "bbox": [ + 498, + 417, + 890, + 462 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Lemma 5 Let $\\| \\cdot \\|$ denote the operator norm, $\\| \\cdot \\|_2$ denote the vector 2-norm, $z \\sim \\mathcal{N}(\\mathbf{0}, I)$ , and $Z = z^\\top A z + z^\\top b + c$ where $A = Q \\Lambda Q^\\top$ . Using Chernoff bound, for any $t \\geq 0$ and $\\gamma \\in \\mathbb{R}$", + "bbox": [ + 498, + 472, + 892, + 534 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathbb {P} \\{Z - \\mathbb {E} Z \\geq \\gamma \\} \\leq \\\\ \\frac {\\exp \\left\\{\\frac {- t}{4 \\| \\boldsymbol {b} \\| _ {2} ^ {\\frac {2}{2}} \\| \\boldsymbol {\\Lambda} \\|} - t (\\gamma + \\operatorname {T r} (\\boldsymbol {\\Lambda}) + \\| \\boldsymbol {b} \\| _ {2}) \\right\\}}{| I - 2 t \\boldsymbol {\\Lambda} | ^ {\\frac {1}{2}}}. \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 500, + 560, + 890, + 619 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Lemma 5 allows us to provide an upper bound for the accuracy of the unlearnable decision boundary $\\tilde{P}$ on the clean dataset $\\mathcal{D}$ , given as $\\tau_{\\mathcal{D}}(\\tilde{P})$ , in Theorem 2 below.", + "bbox": [ + 498, + 628, + 890, + 675 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Theorem 2 (Main result) Let $A = Q\\Lambda Q = A_{-1}^{-2} - A_{1}^{-2}$ . For any non-negative constants $t_1$ and $t_2$ , the accuracy of the unlearnable decision boundary $\\tilde{P}$ on the clean dataset $\\mathcal{D}$ can be upper-bounded as", + "bbox": [ + 498, + 686, + 893, + 748 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\tau_ {\\mathcal {D}} (\\tilde {P}) \\leq \\frac {\\exp \\left[ t _ {1} \\left(\\boldsymbol {b} ^ {\\top} \\boldsymbol {\\mu} + \\boldsymbol {\\mu} ^ {\\top} A \\boldsymbol {\\mu} + c - \\frac {1}{4 \\| 2 A \\boldsymbol {\\mu} + \\boldsymbol {b} \\| _ {2} \\| \\Lambda \\|}\\right) \\right]}{2 | I - 2 t _ {1} \\Lambda | ^ {\\frac {1}{2}}} \\\\ + \\frac {\\exp \\left[ t _ {2} \\left(\\boldsymbol {b} ^ {\\top} \\boldsymbol {\\mu} - \\boldsymbol {\\mu} ^ {\\top} A \\boldsymbol {\\mu} - c - \\frac {1}{4 \\| 2 A \\boldsymbol {\\mu} - \\boldsymbol {b} \\| _ {2} \\| \\Lambda \\|}\\right) \\right]}{2 | I - 2 t _ {2} \\Lambda | ^ {\\frac {1}{2}}} \\\\ := p _ {1} + p _ {2}. \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 503, + 757, + 883, + 857 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Furthermore, if $\\pmb{\\mu}^{\\top}A\\pmb{\\mu} + \\pmb{b}^{\\top}\\pmb{\\mu} + c + \\mathrm{Tr}(A) + \\| 2A\\pmb{\\mu} + \\pmb{b}\\| < 0$ we have $\\tau_{\\mathcal{D}}(\\tilde{P})\\leq \\frac{1}{2} (p_1 + 1) < 1$ . Also, if $-\\pmb{\\mu}^{\\top}A\\pmb{\\mu} + \\pmb{b}^{\\top}\\pmb{\\mu}-$", + "bbox": [ + 500, + 868, + 890, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "3866", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "$c - \\operatorname {Tr}(A) + \\| 2A\\pmb {\\mu} - \\pmb {b}\\| < 0$ we have $\\tau_{\\mathcal{D}}(\\tilde{P})\\leq \\frac{1}{2} (1 + p_2) < 1.$ Moreover, for any $\\pmb {\\mu}\\neq \\mathbf{0}$ and $a_{-1}\\in [0,0.5]$ , $\\exists a_{1}$ such that $\\tau_{\\mathcal{D}}(\\tilde{P}) < \\tau_{\\mathcal{D}}(P)$", + "bbox": [ + 76, + 89, + 468, + 137 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Poisoning is effective only if the accuracy of the unlearnable model $\\tilde{P}$ is less than that of the clean model $P$ on the clean dataset $\\mathcal{D}$ , that is, $\\tau_{\\mathcal{D}}(\\tilde{P}) < \\tau_{\\mathcal{D}}(P)$ . To satisfy this condition, we need to carefully select $a_{y}$ 's. In Theorem 2, we formally state this condition1. Theorem 2 shows that CUDA can effectively poison when there are two distinct modes in the clean Gaussian mixture data model. We validate our theoretical claim through empirical analysis as well (see Figure 3). Details for the analysis is given in supplementary material. We find that our upper bound for the clean test accuracy of CUDA classifiers are consistent with our empirical analysis. In our experiments, we also find that the unlearnability effect is stronger with a larger blur parameter. This effect is evident from Figure 3 where we find that it is likely to get a lower $\\tau_{\\mathcal{D}}(\\tilde{P})$ with higher $a_{y}$ values. These results are consistent with our experimental results in Section 5 with CIFAR-10, CIFAR-100, and ImageNet-100 datasets.", + "bbox": [ + 75, + 148, + 472, + 421 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 76, + 436, + 209, + 454 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this section, we first discuss our experimental setup. More details on the setup is deferred to supplementary material. We then show the robustness of CUDA generation with various datasets and architectures. We also run various experiments to analyze the effectiveness of CUDA under different training techniques (ERM, AT with varying budgets, randomized smoothing, transfer learning, and finetuning) and augmentation techniques (mixup [56], grayscale, random blurring, cutout [9], cutmix [54], autoaugment [7], and orthogonal regularization [3]). Finally, we also design adaptive defenses to test the robustness of CUDA. One might think that CUDA filters can be obtained by adversarially training them with the data. We show that CUDA is robust to such adaptive defenses that we design.", + "bbox": [ + 75, + 462, + 472, + 674 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1. Experimental setup", + "text_level": 1, + "bbox": [ + 76, + 683, + 264, + 700 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Datasets. We use three image classification datasets - CIFAR-10, CIFAR-100, [21] and ImageNet-100 (a subset of ImageNet made of the first 100 classes) [39]. We use the data augmentation techniques such as random flipping, cropping, and resizing [43].", + "bbox": [ + 75, + 707, + 468, + 782 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Architectures. We use ResNet-18 [14], VGG-16 [44], Wide ResNet-34-10 [55], and DenseNet-121 [16]. We train the networks with hyperparameters used in Fu et al. [11].", + "bbox": [ + 75, + 782, + 468, + 829 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "1We note that the conditions $\\pmb{\\mu}^{\\top}A\\pmb{\\mu} + \\pmb{b}^{\\top}\\pmb{\\mu} + c + \\mathrm{Tr}(A) + \\|2A\\pmb{\\mu} + \\pmb{b}\\|_{2} < 0$ or $-\\pmb{\\mu}^{\\top}A\\pmb{\\mu} + \\pmb{b}^{\\top}\\pmb{\\mu} - c - \\mathrm{Tr}(A) + \\|2A\\pmb{\\mu} - \\pmb{b}\\|_{2} < 0$ can always be satisfied by picking a sufficiently large $\\pmb{\\mu}$ in the direction of an eigenvector corresponding to a negative or positive eigenvalue of $A$ (note that $A$ has negative and positive eigenvalues).", + "bbox": [ + 75, + 837, + 468, + 901 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/82f8998bf785a0ace91873b717e04a2a437310fa1bdbc7fb6dbdce80dfbdaa8f.jpg", + "table_caption": [ + "Table 2. Test accuracy (\\%) of ResNet-18 trained on various unlearnable datasets. We use $L_{\\infty}$ AT with budget $\\rho_{a} = 4 / 255$ ." + ], + "table_footnote": [], + "table_body": "
DatasetTraining methodCleanUnlearnability method
EMTAPNTGAREMCUDA
CIFAR-10ERM94.6613.2022.5116.2727.0918.48
AT89.5188.6288.0288.9648.1644.40
CIFAR-100ERM76.271.6013.753.2210.1412.69
AT64.5063.4362.3962.4427.1034.34
ImageNet-100ERM80.661.269.108.4213.748.96
AT66.6263.4063.5663.0641.6638.68
", + "bbox": [ + 506, + 128, + 887, + 242 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/12a453c77b2ec42bae866f62683c3fab0a805612482ff4be802fb3133dd74ad5.jpg", + "table_caption": [ + "Table 3. Test accuracy $(\\%)$ of network architectures trained on various CIFAR-10 unlearnable datasets with $L_{\\infty}$ AT $(\\rho_{a} = 4 / 255)$ ." + ], + "table_footnote": [], + "table_body": "
ModelCleanUnlearnability method
EMTAPNTGAREMCUDA
ResNet-1889.5188.6288.0288.9648.1644.40
VGG-1687.5186.4886.2786.6565.2342.98
Wide ResNet-34-1091.2190.0590.2389.9548.3953.02
DenseNet-12183.2782.4481.7280.7381.4845.95
", + "bbox": [ + 506, + 296, + 887, + 397 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Previous works mainly employ ResNet-18 for most of their evaluations. Additional experiments on Tiny-ImageNet [26], DeIT [48], EfficientNetV2-S [46], and MobileNetV2 [40] are provided in the supplementary material.", + "bbox": [ + 498, + 422, + 890, + 484 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "CUDA generation. We use filters of size $k = 3$ and blur parameter $p_b = 0.3$ for both CIFAR-10 and CIFAR-100 datasets. ImageNet-100 is a higher dimensional $224 \\times 224 \\times 3$ image dataset when compared to the $32 \\times 32 \\times 3$ dimensional CIFAR datasets. Hence, we use larger filters of size $k = 9$ with $p_b = 0.06$ for ImageNet-100. These hyperparameters are chosen such that the CUDA images are not perceptibly highly perturbed and give good unlearnability effect (see plot in Figure 1). In supplementary material, we show the results of training on CUDA with different hyperparameters for data generation.", + "bbox": [ + 496, + 484, + 892, + 651 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Baselines. We compare CUDA generation technique with four state-of-the-art unlearnability methods - REM [11], EM [17], TAP [10], and NTGA [53]. We adopt the results reported in [11] since we use the same hyperparameters for training. For REM, we select hyperparameters $\\rho_{u} = 8 / 255$ and $\\rho_{a} = 4 / 255$ , the highest radii values in [11]. For comparing unlearnable methods, we look at the clean test accuracy. The lower the test accuracy, the better the unlearnability method is. As mentioned in the supplementary material, we use publicly released official codebases for reproducing the baselines using their default hyperparameters.", + "bbox": [ + 496, + 652, + 893, + 834 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.2. Effectiveness of CUDA", + "text_level": 1, + "bbox": [ + 500, + 845, + 710, + 861 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Why does CUDA work? In order to measure how much CUDA technique's blurring affects the dataset's quality, we", + "bbox": [ + 500, + 869, + 890, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "3867", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/b4c64526cc6177916ff8a34f965b400fc170077a1092d9a7defbc2aee0944cc3.jpg", + "table_caption": [ + "Table 4. Test accuracy $(\\%)$ of ResNet-18 with CUDA under various training settings." + ], + "table_footnote": [], + "table_body": "
DatasetCleanTraining methodCUDA (ours)
CIFAR-1094.66ERM18.48
AT L∞ (ρa = 4/255)44.40
AT L∞ (ρa = 8/255)32.85
AT L∞ (ρa = 16/255)19.32
AT L2 (ρa = 0.25)39.05
AT L2 (ρa = 0.50)51.19
AT L2 (ρa = 0.75)51.14
CIFAR-10076.27ERM12.69
AT L∞ (ρa = 4/255)34.34
AT L∞ (ρa = 8/255)30.00
AT L2 (ρa = 0.75)36.90
ImageNet-10080.66ERM8.96
AT L∞ (ρa = 4/255)38.68
AT L∞ (ρa = 8/255)40.08
AT L2 (ρa = 0.75)20.58
", + "bbox": [ + 80, + 128, + 467, + 319 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "compare class-wise blurring (CUDA technique) against universal blurring where a single convolutional filter is used for blurring all the images. We keep the filter generation parameters fixed ( $p_b = 0.3$ and $k = 3$ ) for both CUDA class-wise blurring and universal blurring. ResNet-18 trained with clean CIFAR-10, universally blurred CIFAR-10, and CIFAR-10 CUDA achieve clean test accuracies of $94.66\\%$ , $90.47\\%$ , and $18.48\\%$ , respectively. This suggests that our controlled blurring does not obscure the semantics of the dataset. Hence, the significant drop in the clean test accuracy introduced by CUDA is most likely due to the usage of class-wise filters. This suggests that a model trained on CUDA learns the relation between the class-wise convolution filters and their corresponding labels. Therefore, during test time when this convolution effect is absent, the CUDA trained model fails to make correct classifications. Furthermore, the model trained on CUDA achieves an accuracy of $99.91\\%$ on the CIFAR-10 CUDA testset. This strongly supports our claim that the CUDA model learns to classify images based on the convolutional filters used to blur them. In addition, if we permute the class-wise filters for blurring the test set (i.e., blurring class 1 images with class 2 filters, class 2 images with class 3 filters, and so on), we get a very low accuracy of $2.53\\%$ on this test set. Further details are provided in the supplementary material. Finally, we note that real-world datasets might also contain blurred images due to various factors such as motion blurring, weather conditions, issues with the camera, etc. Hence, detecting if a blurred image is poisoned might not always be possible. However, one might argue that it is possible to detect if an entire dataset is blurred. Interestingly, later in this section we show that CUDA technique exhibits unlearnability effect even when only a fraction of the training dataset is poisoned.", + "bbox": [ + 75, + 351, + 472, + 866 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Different datasets. We first compare the effectiveness of CUDA with different datasets using ERM and $L_{\\infty}$ AT with", + "bbox": [ + 76, + 869, + 470, + 902 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "$\\rho_{a} = 4 / 255$ . We use ResNet-18 for the experiments. The results are shown in Table 2. These results show that EM, TAP, and NTGA are not robust to AT. However, both CUDA and REM are successful. Here, our method CUDA outperforms REM with CIFAR-10 and ImageNet-100 datasets. Smartly designed additive noise in AT helps in achieving better generalization than ERM on the unlearnable datasets. This experiment thus demonstrates that ERM and AT are not good choices for training with CUDA and REM dataset.", + "bbox": [ + 496, + 90, + 890, + 227 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Different models. Next we compare the effectiveness of CUDA using various deep learning architectures with $L_{\\infty}$ AT ( $\\rho_{a} = 4 / 255$ ). We use CIFAR-10 for the experiments. The results are shown in Table 3. As we see in the table, CUDA is effective with all the five network architectures. However, REM is not seen to be transferrable with DenseNet-121.", + "bbox": [ + 496, + 229, + 890, + 335 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Robustness to different training settings. In Section 3.2, we show that REM is sensitive to the training settings. REM generated using $L_{\\infty}$ radii budgets of $\\rho_u = 8 / 255$ and $\\rho_a = 4 / 255$ for CIFAR-10 breaks with $L_{\\infty}$ AT $(\\rho_a = 8 / 255)$ and $L_2$ AT $(\\rho_a = 0.75)$ to get test accuracy of $78.71\\%$ and $79.65\\%$ , respectively. Hence, we run experiments to check the robustness of CUDA with various AT norm budgets. The results are shown in Table 4. As we see in the table, CUDA is robust to ERM, $L_{\\infty}$ , and $L_2$ AT settings with varying training budgets. Impressively, the highest test accuracy achieved with training on CIFAR-10 CUDA, CIFAR-100 CUDA, and ImageNet-100 CUDA are as low as $51.19\\%$ , $36.90\\%$ , and $40.08\\%$ , respectively. We also find that using a pre-trained ResNet-18 with CIFAR-10 CUDA only achieves clean test accuracy of $42.42\\%$ and $48.22\\%$ with fine-tuning the full network and a newly trained final layer, respectively (details are deferred to the supplementary material).", + "bbox": [ + 496, + 337, + 892, + 609 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Different protection percentages. In Section 3.1, we assume that the attacker has access to the full clean training data. However, in real life settings, this might not be always possible. Hence, we train ResNet-18 on a mix of CIFAR-10 CUDA and clean CIFAR-10 training datasets to evaluate the effectiveness of poisoning with varying data protection percentages. Protection percentage denotes the percentage of the training data that is poisoned.", + "bbox": [ + 496, + 611, + 890, + 733 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We show the results in Table 5. In the table, the \"Mixed\" column denotes the clean test accuracy of a model trained using a mix of both clean and poisoned data. The \"Clean\" column denotes the clean test accuracy of a model trained only using the clean subset of the training data. In the last row of Table 5, we provide the results for CUDA trained using ERM with varying data protection percentages. The remainder of the rows provide the results for the $L_{\\infty}$ AT $(\\rho_{a} = 4 / 255)$ scenario. For example, CUDA trained with ERM using an $80\\%$ clean training data partition achieves a test accuracy of $93.75\\%$ . Adding the $20\\%$ CUDA data", + "bbox": [ + 496, + 734, + 893, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "3868", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/98e511d21bab45a25028cf7f17abcb863a07019773528faa978ae92d278c35a8.jpg", + "table_caption": [ + "Table 5. Test accuracy (\\%) of ResNet-18 on CIFAR-10 with different data protection percentages. The last row shows the results for CUDA with ERM setting. The rest of the rows show results for unlearnability methods trained in the $L_{\\infty}$ AT ( $\\rho_a = 4/255$ ) setting." + ], + "table_footnote": [], + "table_body": "
Unlearnability methodData Protection Percentage
0%20% MixedClean40% MixedClean60% MixedClean80% MixedClean100%
EM89.5189.6088.1789.4086.7689.4985.0789.1079.4188.62
TAP89.0188.6688.4088.0488.02
NTGA89.5689.3589.2289.1788.96
REM89.6089.3489.6188.0948.16
CUDA (ours)88.5487.2486.0384.3444.40
CUDA + ERM94.6693.2893.7591.3492.5689.9189.7785.6184.3018.48
", + "bbox": [ + 171, + 127, + 803, + 268 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "partition to the training dataset drops the test accuracy of the model to $93.28\\%$ . Results from the last row of Table 5 show that CUDA with varying data protection percentages is effective with the ERM setting. However, with the AT scenario, the unlearnability techniques are not as successful as with ERM with varying data protection percentages. Nevertheless, it is interesting to note that CUDA technique performs better than all the other unlearnability techniques with AT.", + "bbox": [ + 75, + 290, + 472, + 425 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Robustness to smoothing and data augmentations. Cohen et al. [6] proposes randomized smoothing which is a provable adversarial defense in $L_{2}$ norm. Since CUDA does not use additive noise, CUDA is robust to randomized smoothing. A smoothed ResNet-18 (with a noise level of 0.5) with CIFAR-10 CUDA achieves only a clean test accuracy of $43.85\\%$ . In Section 3.2, we see that REM breaks with grayscaleing. While ResNet-18 training using grayscaleed REM achieves $70.76\\%$ test accuracy, grayscaleed CUDA only achieves $20.12\\%$ test accuracy on the clean grayscaleed CIFAR-10 test data. This shows that CUDA technique exhibits the desirable property of not relying upon the color space for its attack. CIFAR-10 CUDA training achieves only $25.53\\%$ , $25.80\\%$ , $26.93\\%$ , $34.09\\%$ , and $50.72\\%$ clean test accuracy with mixup, cutout, cutmix, autoaugment, and orthogonal regularization, respectively (see supplementary material for more details).", + "bbox": [ + 75, + 428, + 472, + 685 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Adaptive defenses for CUDA. Here, we first investigate the effect of training CIFAR-10 CUDA with random blurring augmentations using ResNet-18. Each batch of the CUDA training data is convolved with random $3 \\times 3$ filters of varying blur parameters $p_b'$ . With $p_b'$ values of 0.1 and 0.3, CUDA training achieves lower test accuracy $9.2\\%$ and $13.37\\%$ , respectively. This shows that the noise added by CUDA technique is robust to random convolution augmentations.", + "bbox": [ + 75, + 686, + 470, + 821 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "One may think that CUDA technique can be broken by learning the private filters from the data. We test this idea by training deconvolution filters to find if we can reverse the blurring effect in CUDA with adversarially trained filters. We use a novel Deconvolution-based Adversarial Training", + "bbox": [ + 75, + 824, + 470, + 902 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "(DAT) technique that is similar to AT (check supplementary material for details). While the adversarial step in AT learns sample-wise error-maximizing additive noises, the adversarial step in DAT learns class-wise error-maximizing deconvolution filters. We train DAT with filters of varying sizes (3, 5, and 7) on CIFAR-10 CUDA using ResNet-18. The filter parameters are constrained within a finite range to make sure that the images do not get distorted with the adversarial step similar to projection in projected gradient descent. We find that CUDA is robust to DAT. DAT using filters of size 3, 5, and 7 with CUDA achieves only test accuracy of $39.05\\%$ , $46.21\\%$ , and $38.48\\%$ , respectively. DAT is not successful against CUDA since we can not invert convolutions without the knowledge of the private filters or clean images corresponding to CUDA.", + "bbox": [ + 496, + 290, + 893, + 518 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Limitations and future directions. The unlearnability effect of CUDA can be defended if some fraction of the clean data and its corresponding CUDA images are leaked. The defender must also be able to detect if all samples in the dataset are poisoned. However, our work assumes a setup where the filters remain private. For example, a data publisher could simply publish their CUDA images and delete the clean images permanently to prevent this scenario. As discussed in this section, CUDA as well as other prior works do not perform well with different protection percentages with AT. Improving this can be an interesting research direction. We believe that extending CUDA technique to other domains such as tabular and text data is also an interesting future direction. It would also be interesting to see theoretical analysis of CUDA considering more complex setups.", + "bbox": [ + 496, + 522, + 893, + 750 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgement", + "text_level": 1, + "bbox": [ + 500, + 779, + 660, + 797 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This project was supported in part by Meta grant 23010098, NSF CAREER AWARD 1942230, HR001119S0026 (GARD), ONR YIP award N00014-22-1-2271, Army Grant No. W911NF2120076, a capital one grant, NIST 60NANB20D134 and the NSF award CCF2212458.", + "bbox": [ + 496, + 810, + 893, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "3869", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Mauro Barni, Kassem Kallas, and Benedetta Tondi. A new backdoor attack in cnns by training set corruption without label poisoning. In 2019 IEEE International Conference on Image Processing (ICIP), pages 101-105. IEEE, 2019. 3", + "[2] Battista Biggio, Blaine Nelson, and Pavel Laskov. Poisoning attacks against support vector machines. arXiv preprint arXiv:1206.6389, 2012. 2", + "[3] Andrew Brock, Theodore Lim, James M Ritchie, and Nick Weston. Neural photo editing with introspective adversarial networks. arXiv preprint arXiv:1609.07093, 2016. 6", + "[4] Nicholas Carlini and David Wagner. Towards evaluating the robustness of neural networks. In 2017 IEEE symposium on security and privacy (sp), pages 39-57. IEEE, 2017. 2", + "[5] Xinyun Chen, Chang Liu, Bo Li, Kimberly Lu, and Dawn Song. Targeted backdoor attacks on deep learning systems using data poisoning. arXiv preprint arXiv:1712.05526, 2017.3", + "[6] Jeremy Cohen, Elan Rosenfeld, and Zico Kolter. Certified adversarial robustness via randomized smoothing. In Kamalika Chaudhuri and Ruslan Salakhutdinov, editors, Proceedings of the 36th International Conference on Machine Learning, volume 97 of Proceedings of Machine Learning Research, pages 1310-1320. PMLR, 09-15 Jun 2019. 2, 8", + "[7] Ekin D Cubuk, Barret Zoph, Dandelion Mane, Vijay Vasudevan, and Quoc V Le. Autoaugment: Learning augmentation policies from data. arXiv preprint arXiv:1805.09501, 2018. 6", + "[8] Karan Desai and Justin Johnson. Virtex: Learning visual representations from textual annotations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11162-11173, 2021. 1", + "[9] Terrance DeVries and Graham W Taylor. Improved regularization of convolutional neural networks with cutout. arXiv preprint arXiv:1708.04552, 2017. 6", + "[10] Liam Fowl, Micah Goldblum, Ping-yeh Chiang, Jonas Geiping, Wojtek Czaja, and Tom Goldstein. Adversarial examples make strong poisons. arXiv preprint arXiv:2106.10807, 2021. 1, 3, 6", + "[11] Shaopeng Fu, Fengxiang He, Yang Liu, Li Shen, and Dacheng Tao. Robust unlearnable examples: Protecting data privacy against adversarial learning. In International Conference on Learning Representations, 2021. 1, 3, 4, 6", + "[12] Ian J Goodfellow, Jonathon Shlens, and Christian Szegedy. Explaining and harnessing adversarial examples. arXiv preprint arXiv:1412.6572, 2014. 2", + "[13] Tianyu Gu, Brendan Dolan-Gavitt, and Siddharth Garg. Badnets: Identifying vulnerabilities in the machine learning model supply chain. arXiv preprint arXiv:1708.06733, 2017. 3", + "[14] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. CoRR, abs/1512.03385, 2015. 6", + "[15] Kashmir Hill. The secretive company that might end privacy as we know it. In *Ethics of Data and Analytics*, pages 170–177. Auerbach Publications, 2020. 1" + ], + "bbox": [ + 78, + 114, + 470, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[16] Gao Huang, Zhuang Liu, and Kilian Q. Weinberger. Densely connected convolutional networks. CoRR, abs/1608.06993, 2016. 6", + "[17] Hanxun Huang, Xingjun Ma, Sarah Monazam Erfani, James Bailey, and Yisen Wang. Unlearnable examples: Making personal data unexploitable. arXiv preprint arXiv:2101.04898, 2021. 1, 3, 6", + "[18] Arthur Jacot, Franck Gabriel, and Clément Hongler. Neural tangent kernel: Convergence and generalization in neural networks. Advances in neural information processing systems, 31, 2018. 3", + "[19] Adel Javanmard and Mahdi Soltanolkotabi. Precise statistical analysis of classification accuracies for adversarial training. ArXiv, abs/2010.11213, 2020. 2, 4", + "[20] Pang Wei Koh and Percy Liang. Understanding black-box predictions via influence functions. In International conference on machine learning, pages 1885-1894. PMLR, 2017. 2", + "[21] Alex Krizhevsky. Learning multiple layers of features from tiny images. Technical report, 2009. 6", + "[22] Aounon Kumar, Alexander Levine, Soheil Feizi, and Tom Goldstein. Certifying confidence via randomized smoothing. CoRR, abs/2009.08061, 2020. 2", + "[23] Alexey Kurakin, Ian J Goodfellow, and Samy Bengio. Adversarial examples in the physical world. In Artificial intelligence safety and security, pages 99-112. Chapman and Hall/CRC, 2018. 2", + "[24] Cassidy Laidlaw and Soheil Feizi. Functional adversarial attacks. In H. Wallach, H. Larochelle, A. Beygelzimer, F. d'Alché-Buc, E. Fox, and R. Garnett, editors, Advances in Neural Information Processing Systems, volume 32. Curran Associates, Inc., 2019. 2", + "[25] Cassidy Laidlaw, Sahil Singla, and Soheil Feizi. Perceptual adversarial robustness: Defense against unseen threat models. CoRR, abs/2006.12655, 2020. 2", + "[26] Ya Le and Xuan Yang. Tiny imagenet visual recognition challenge. CS 231N, 7(7):3, 2015. 6", + "[27] Alexander Levine and Soheil Feizi. Deep partition aggregation: Provable defense against general poisoning attacks. CoRR, abs/2006.14768, 2020. 2", + "[28] Alexander Levine and Soheil Feizi. Deep partition aggregation: Provable defense against general poisoning attacks. CoRR, abs/2006.14768, 2020. 2", + "[29] Yige Li, Xixiang Lyu, Nodens Koren, Lingjuan Lyu, Bo Li, and Xingjun Ma. Neural attention distillation: Erasing backdoor triggers from deep neural networks. arXiv preprint arXiv:2101.05930, 2021. 3", + "[30] Yiming Li, Tongqing Zhai, Baoyuan Wu, Yong Jiang, Zhifeng Li, and Shutao Xia. Rethinking the trigger of backdoor attack. arXiv preprint arXiv:2004.04692, 2020. 3", + "[31] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In European conference on computer vision, pages 740-755. Springer, 2014. 1", + "[32] Sijia Liu, Songtao Lu, Xiangyi Chen, Yao Feng, Kaidi Xu, Abdullah Al-Dujaili, Mingyi Hong, and Una-May O'Reilly." + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "3870", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Min-max optimization without gradients: Convergence and applications to black-box evasion and poisoning attacks. In International Conference on Machine Learning, pages 6282-6293. PMLR, 2020. 2", + "[33] Yunfei Liu, Xingjun Ma, James Bailey, and Feng Lu. Reflection backdoor: A natural backdoor attack on deep neural networks. In European Conference on Computer Vision, pages 182-199. Springer, 2020. 3", + "[34] Aleksander Madry, Aleksandar Makelov, Ludwig Schmidt, Dimitris Tsipras, and Adrian Vladu. Towards deep learning models resistant to adversarial attacks. arXiv preprint arXiv:1706.06083, 2017. 1, 2", + "[35] Yifei Min, Lin Chen, and Amin Karbasi. The curious case of adversarially robust models: More data can help, double descend, or hurt generalization. In Uncertainty in Artificial Intelligence, pages 129-139. PMLR, 2021. 4", + "[36] Tuan Anh Nguyen and Anh Tran. Input-aware dynamic backdoor attack. Advances in Neural Information Processing Systems, 33:3454-3464, 2020. 3", + "[37] Silvia Noschese, Lionello Pasquini, and Lothar Reichel. Tridiagonal toeplitz matrices: Properties and novel applications. 2006. 5", + "[38] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 1", + "[39] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, Alexander C. Berg, and Li Fei-Fei. ImageNet Large Scale Visual Recognition Challenge. International Journal of Computer Vision (IJCV), 115(3):211-252, 2015. 6", + "[40] M. Sandler, A. Howard, M. Zhu, A. Zhmoginov, and L.C. Chen. Mobilenetv2: Inverted residuals and linear bottle-necks. In CVPR, 2018. 6", + "[41] Mert Bulent Sariyildiz, Julien Perez, and Diane Larlus. Learning visual representations with caption annotations. In European Conference on Computer Vision, pages 153-170. Springer, 2020. 1", + "[42] Ali Shafahi, W Ronny Huang, Mahyar Najibi, Octavian Suciu, Christoph Studer, Tudor Dumitras, and Tom Goldstein. Poison frogs! targeted clean-label poisoning attacks on neural networks. Advances in neural information processing systems, 31, 2018. 2, 3", + "[43] Connor Shorten and Taghi Khoshgoftaar. A survey on image data augmentation for deep learning. Journal of Big Data, 6, 07 2019. 6", + "[44] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. CoRR, abs/1409.1556, 2015. 6", + "[45] Christian Szegedy, Wojciech Zaremba, Ilya Sutskever, Joan Bruna, Dumitru Erhan, Ian Goodfellow, and Rob Fergus. Intriguing properties of neural networks. arXiv preprint arXiv:1312.6199, 2013. 2", + "[46] M. Tan and Q. Le. Efficientnetv2: Smaller models and faster training. In ICML, 2021. 6" + ], + "bbox": [ + 78, + 90, + 470, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[47] Lue Tao, Lei Feng, Jinfeng Yi, Sheng-Jun Huang, and Songcan Chen. Better safe than sorry: Preventing delusive adversaries with adversarial training. Advances in Neural Information Processing Systems, 34, 2021. 1, 3", + "[48] H. Touvron, M. Cord, M. Douze, F. Massa, A. Sablayrolles, and H. Jegou. Training data-efficient image transformers & distillation through attention. In ICML, 2021. 6", + "[49] Laurens van der Maaten and Geoffrey Hinton. Visualizing data using t-sne. Journal of Machine Learning Research, 9(86):2579-2605, 2008. 2", + "[50] Wenxiao Wang, Alexander Levine, and Soheil Feizi. Improved certified defenses against data poisoning with (deterministic) finite aggregation. CoRR, abs/2202.02628, 2022. 2", + "[51] Yisen Wang, Xingjun Ma, James Bailey, Jinfeng Yi, Bowen Zhou, and Quanquan Gu. On the convergence and robustness of adversarial training. arXiv preprint arXiv:2112.08304, 2021. 2", + "[52] Da Yu, Huishuai Zhang, Wei Chen, Jian Yin, and Tie-Yan Liu. Availability attacks create shortcuts. In Proceedings of the 28th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, KDD '22, page 2367-2376, New York, NY, USA, 2022. Association for Computing Machinery. 2, 3", + "[53] Chia-Hung Yuan and Shan-Hung Wu. Neural tangent generalization attacks. In International Conference on Machine Learning, pages 12230-12240. PMLR, 2021. 1, 3, 6", + "[54] Sangdoo Yun, Dongyoon Han, Seong Joon Oh, Sanghyuk Chun, Junsuk Choe, and Youngjoon Yoo. Cutmix: Regularization strategy to train strong classifiers with localizable features. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6023-6032, 2019. 6", + "[55] Sergey Zagoruyko and Nikos Komodakis. Wide residual networks. In Edwin R. Hancock Richard C. Wilson and William A. P. Smith, editors, Proceedings of the British Machine Vision Conference (BMVC), pages 87.1-87.12. BMVA Press, September 2016. 6", + "[56] Hongyi Zhang, Moustapha Cisse, Yann N. Dauphin, and David Lopez-Paz. mixup: Beyond empirical risk minimization. CoRR, abs/1710.09412, 2017. 6", + "[57] Hongyang Zhang, Yaodong Yu, Jiantao Jiao, Eric Xing, Laurent El Ghaoui, and Michael Jordan. Theoretically principled trade-off between robustness and accuracy. In International conference on machine learning, pages 7472-7482. PMLR, 2019. 2" + ], + "bbox": [ + 501, + 92, + 893, + 712 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "3871", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2023/CUDA_ Convolution-Based Unlearnable Datasets/e0179aca-27a9-460d-b895-5e7378133f26_model.json b/2023/CUDA_ Convolution-Based Unlearnable Datasets/e0179aca-27a9-460d-b895-5e7378133f26_model.json new file mode 100644 index 0000000000000000000000000000000000000000..78646f7f5c489aeb04b042889f8bf55feb287681 --- /dev/null +++ b/2023/CUDA_ Convolution-Based Unlearnable Datasets/e0179aca-27a9-460d-b895-5e7378133f26_model.json @@ -0,0 +1,2068 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.239, + 0.131, + 0.733, + 0.151 + ], + "angle": 0, + "content": "CUDA: Convolution-based Unlearnable Datasets" + }, + { + "type": "text", + "bbox": [ + 0.124, + 0.182, + 0.311, + 0.216 + ], + "angle": 0, + "content": "Vinu Sankar Sadasivan \nUniversity of Maryland" + }, + { + "type": "text", + "bbox": [ + 0.162, + 0.221, + 0.272, + 0.233 + ], + "angle": 0, + "content": "vinu@umd.edu" + }, + { + "type": "text", + "bbox": [ + 0.352, + 0.182, + 0.619, + 0.216 + ], + "angle": 0, + "content": "Mahdi Soltanolkotabi \nUniversity of Southern California" + }, + { + "type": "text", + "bbox": [ + 0.413, + 0.221, + 0.558, + 0.233 + ], + "angle": 0, + "content": "soltanol@usc.edu" + }, + { + "type": "text", + "bbox": [ + 0.66, + 0.182, + 0.848, + 0.216 + ], + "angle": 0, + "content": "Soheil Feizi \nUniversity of Maryland" + }, + { + "type": "text", + "bbox": [ + 0.677, + 0.221, + 0.83, + 0.233 + ], + "angle": 0, + "content": "sfeizi@cs.umd.edu" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.269, + 0.314, + 0.285 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.301, + 0.474, + 0.817 + ], + "angle": 0, + "content": "Large-scale training of modern deep learning models heavily relies on publicly available data on the web. This potentially unauthorized usage of online data leads to concerns regarding data privacy. Recent works aim to make unlearnable data for deep learning models by adding small, specially designed noises to tackle this issue. However, these methods are vulnerable to adversarial training (AT) and/or are computationally heavy. In this work, we propose a novel, model-free, Convolution-based Unlearnable DAtaset (CUDA) generation technique. CUDA is generated using controlled class-wise convolutions with filters that are randomly generated via a private key. CUDA encourages the network to learn the relation between filters and labels rather than informative features for classifying the clean data. We develop some theoretical analysis demonstrating that CUDA can successfully poison Gaussian mixture data by reducing the clean data performance of the optimal Bayes classifier. We also empirically demonstrate the effectiveness of CUDA with various datasets (CIFAR-10, CIFAR100, ImageNet-100, and Tiny-ImageNet), and architectures (ResNet-18, VGG-16, Wide ResNet-34-10, DenseNet-121, DeIT, EfficientNetV2-S, and MobileNetV2). Our experiments show that CUDA is robust to various data augmentations and training approaches such as smoothing, AT with different budgets, transfer learning, and fine-tuning. For instance, training a ResNet-18 on ImageNet-100 CUDA achieves only \\(8.96\\%\\), \\(40.08\\%\\), and \\(20.58\\%\\) clean test accuracies with empirical risk minimization (ERM), \\(L_{\\infty}\\) AT, and \\(L_{2}\\) AT, respectively. Here, ERM on the clean training data achieves a clean test accuracy of \\(80.66\\%\\). CUDA exhibits unlearnability effect with ERM even when only a fraction of the training dataset is perturbed. Furthermore, we also show that CUDA is robust to adaptive defenses designed specifically to break it." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.845, + 0.21, + 0.86 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.47, + 0.903 + ], + "angle": 0, + "content": "Modern deep learning training frameworks heavily depend on large-scale datasets for achieving high accuracy." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.27, + 0.895, + 0.422 + ], + "angle": 0, + "content": "This encourages deep learning practitioners to scrape data from the web for data collection [8, 31, 38, 41]. Since a lot of the data is publicly available online, sometime this scraping of data is unauthorized. For instance, a recent article [15] discloses that a private company trained a commercial face recognition system using over three billion facial images collected from the internet without any user consent. Although such massive data can significantly boost the performance of deep learning models, it raises serious concerns about data privacy and security." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.423, + 0.896, + 0.725 + ], + "angle": 0, + "content": "To prevent the unauthorized usage of personal data, a series of recent papers [10, 17, 53] propose to poison data with additive noise. The idea is to make datasets unlearnable for deep learning models by ensuring that they learn the correspondence between noises and labels. Thereby, they do not learn much useful information about the clean data, significantly degrading their clean test accuracy. However, in recent works [11, 17, 47], these unlearnability methods are shown to be vulnerable to adversarial training (AT) frameworks [34]. Motivated by this problem, Fu et al. [11] developed Robust Error-Minimization (REM) noises to make unlearnable data that is protected from AT. While the authors show the effectiveness of REM in multiple scenarios, we demonstrate that these methods are still not robust against different data augmentations or training settings (see Section 3.2). Furthermore, current unlearnability frameworks [10, 11, 17, 53] are model-dependent and require expensive optimization steps on deep learning models to obtain the additive noises. They also need to train the deep learning models from scratch to obtain noises for each new data set." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.726, + 0.895, + 0.862 + ], + "angle": 0, + "content": "In this paper, we propose a novel Convolution-based Unlearnable DAataset (CUDA) generation technique. We address limitations of existing unlearnable data generation techniques in Section 3.2 and motivate our CUDA technique in Section 3.3. For generating CUDA, an attacker randomly generates different convolutional filters for each class in the dataset using a private key or seed value. These filters are used to perform controlled class-wise convolutions on the clean training dataset to obtain CUDA. As we" + }, + { + "type": "page_footnote", + "bbox": [ + 0.5, + 0.875, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Code available here: https://github.com/vinusankars/Convolution-based-Unlearnability" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3862" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.15, + 0.096, + 0.65, + 0.227 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.668, + 0.093, + 0.822, + 0.231 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.254, + 0.892, + 0.338 + ], + "angle": 0, + "content": "Figure 1. CIFAR-10 CUDA images from each of the 10 classes convolved using convolutional filters generated with blur parameter \\( p_b \\). As seen in the plot, a higher \\( p_b \\) value results in better unlearnability (lower clean test accuracy), but increased blurring. CIFAR-10 CUDA does not break with ERM for all the three \\( p_b \\) values. In Section 5.2, we propose Deconvolution-based Adversarial Training (DAT) that we specifically design to break CUDA. DAT adversarially learns class-wise filters to deconvolve CUDA images. However, CIFAR-10 CUDA with \\( p_b \\in \\{0.3, 0.5\\} \\) does not break with DAT. Therefore, we select \\( p_b = 0.3 \\) (lesser blurring) for CIFAR-10 CUDA as discussed in Section 5.1. More details on DAT in supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.365, + 0.473, + 0.741 + ], + "angle": 0, + "content": "describe in Sections 3.3 and 5.2, CUDA generation performs controlled convolutions using a blur parameter \\( p_b \\) to ensure that the semantics of the dataset are preserved (see Figure 1). CUDA generation with a lower blurring parameter \\( p_b \\) adds less perceptible noises to clean samples. A network trained by a defender on CUDA is encouraged to learn the shortcut relation between class-wise convolutional filters and labels rather than useful features for classifying the clean data. Since the seed value for generating the filters are private, its not possible for the defender to obtain clean data from CUDA alone. Additionally, CUDA exhibits unlearnability effect with ERM even when only a fraction of the training dataset is perturbed (see Section 5). While the existing unlearnability works use additive noises, CUDA generation technique enjoys the advantage of introducing multiplicative noises in the Fourier domain due to the convolution theorem (since convolution of signals is the same as element-wise multiplication in the Fourier domain). This lets CUDA generation add higher amounts of noise in the image space, specifically along the edges in images, and makes it resilient to AT with small additive noise budgets. In Figure 2, with the help of t-SNE plots [49], we also find that the noises added by CUDA generation is not linearly separable while they are linearly separable for the existing works on unlearnability [52]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.742, + 0.47, + 0.817 + ], + "angle": 0, + "content": "In Section 4, we theoretically show that CUDA generation can successfully poison Gaussian mixture data by degrading the clean data accuracy of the optimal Bayes classifier. We state our result informally below while the formal version is presented in Theorem 2." + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.825, + 0.471, + 0.9 + ], + "angle": 0, + "content": "Theorem 1 (Informal) Let \\(\\mathcal{D}\\) denote a Gaussian mixture data with two modes, \\(P_{\\mathcal{D}}\\) denote the optimal Bayes classifier trained on \\(\\mathcal{D}\\), and \\(\\tau_{\\mathcal{D}}(P_{\\mathcal{D}})\\) denote the accuracy of the classifier \\(P_{\\mathcal{D}}\\) on \\(\\mathcal{D}\\). Then, under some assumptions, for every clean data \\(\\mathcal{D}\\), there is a CUDA \\(\\tilde{\\mathcal{D}}\\) such that" + }, + { + "type": "equation", + "bbox": [ + 0.5, + 0.365, + 0.637, + 0.381 + ], + "angle": 0, + "content": "\\[\n\\tau_ {\\mathcal {D}} (P _ {\\tilde {\\mathcal {D}}}) < \\tau_ {\\mathcal {D}} (P _ {\\mathcal {D}}).\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.393, + 0.892, + 0.604 + ], + "angle": 0, + "content": "Furthermore, our empirical experiments in Section 5 demonstrate the effectiveness of CUDA under various training scenarios such as ERM with various augmentations and regularizations, AT with different budgets, randomized smoothing [6, 22, 27], transfer learning, and fine-tuning. For instance, training a ResNet-18 on CIFAR-10 CUDA achieves only \\(18.48\\%\\), \\(44.4\\%\\), and \\(51.14\\%\\) clean test accuracies with ERM, \\(L_{\\infty}\\) AT, and \\(L_{2}\\) AT, respectively (see Figure 2). Here, ERM on the clean training data achieves a clean test accuracy of \\(94.66\\%\\). In addition, we also design adaptive defenses to investigate if CUDA breaks with random or adversarial defense mechanisms. We find that CUDA is robust to the adaptive defenses that we specifically design to break it." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.618, + 0.65, + 0.634 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.644, + 0.892, + 0.704 + ], + "angle": 0, + "content": "Our CUDA generation technique is intimately related with adversarial and poisoning attacks. We first discuss some of this literature and then explain their relation with CUDA generation." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.705, + 0.892, + 0.854 + ], + "angle": 0, + "content": "Adversarial attacks. Adversarial examples are specially designed examples that can fool deep learning models at test time [4, 12, 23, 24, 45]. The adversary crafts these examples by adding error-maximizing noises to the clean data. Even slightly perturbed data can serve as adversarial examples. AT is a training framework proposed to make deep learning models robust to adversarial examples [19,25,34,51,57]. AT is a min-max optimization problem where the model is trained to minimize loss on adversarial examples that have the maximum loss." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Poisoning attacks. In data poisoning, an attacker aims to hurt the deep learning model's performance by perturbing the training data [2, 20, 28, 32, 42, 50]. The backdoor attack" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3863" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.221, + 0.09, + 0.578, + 0.258 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.584, + 0.09, + 0.752, + 0.255 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.075, + 0.273, + 0.895, + 0.345 + ], + "angle": 0, + "content": "Figure 2. In the left column, clean test accuracies of ResNet-18 on clean CIFAR-10 and CIFAR-10 CUDA are shown. The lower the test accuracy, the higher the effectiveness of unlearnability is. We show that CUDA is robust to various training settings such as ERM, \\( L_{2} \\) AT, and \\( L_{\\infty} \\) AT. The error bars represent the standard deviation of test accuracy over 5 independent trials. CUDA is robust to the randomness in its data generation process. In the right column, t-SNEs of the noises generated by REM (top) and CUDA (bottom) for the first 3 classes of CIFAR-10 are plotted." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.37, + 0.47, + 0.476 + ], + "angle": 0, + "content": "is a special type of poisoning attack where a trigger pattern is injected into clean data at training time [5, 29, 33, 36]. The model trained on this data would misclassify an image with a trigger pattern at test time. Gu et al. [13] and Li et al. [30] use perceptible amounts of noises similar to CUDA for data poisoning. However, backdoor attacks do not affect the performance of the model on clean data [1, 5, 42]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.477, + 0.471, + 0.657 + ], + "angle": 0, + "content": "Recent literature utilize data poisoning to protect data from being used for model training without authorization. Yuan and Wu [53] use neural tangent kernels [18] to generate clean label attacks that can hurt the generalization of deep learning models. Huang et al. [17] show that error-minimizing noise addition can serve as a poisoning technique. Fowl et al. [10] show that error-maximizing noises as well can make strong poison attacks. However, all these poisoning techniques do not offer data protection with AT [47, 52]. Fu et al. [11] proposes a min-min-max optimization technique to generate poisoned data that offers better unlearnability effects with AT." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.672, + 0.47, + 0.707 + ], + "angle": 0, + "content": "3. Convolution-based Ulearnable DAtaset (CUDA)" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.715, + 0.47, + 0.777 + ], + "angle": 0, + "content": "In this section, first we give some preliminaries about unlearnability. Then we discuss the limitations of the existing unlearnability methods. Finally, we propose our CUDA generation technique." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.787, + 0.219, + 0.801 + ], + "angle": 0, + "content": "3.1. Preliminaries" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.81, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Let \\(\\{(x_i, y_i)\\}_{i=1}^n \\sim \\mathcal{D}^n\\) be the clean training dataset where \\(\\mathcal{D}\\) is the clean data distribution, \\(x_i \\in \\mathcal{X} \\subset \\mathbb{R}^d\\) are the samples, and \\(y_i \\in \\mathcal{Y}\\) are the corresponding labels. Suppose a network is given as \\(f_\\theta: \\mathcal{X} \\to \\mathcal{Y}\\) where \\(\\theta \\in \\Theta\\) is the network parameter and \\(\\ell: \\mathcal{Y} \\times \\mathcal{Y} \\to \\mathbb{R}\\) is the loss function. ERM trains a network using a minimization prob" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.369, + 0.892, + 0.431 + ], + "angle": 0, + "content": "lem of the form: \\(\\min_{\\theta} \\frac{1}{n} \\sum_{i=1}^{n} \\ell(f_{\\theta}(\\pmb{x}_i), y_i)\\). Standard \\(L_p\\)-based AT (for \\(p \\in \\mathbb{R}^+\\)) solves the following min-max problem: \\(\\min_{\\theta} \\frac{1}{n} \\sum_{i=1}^{n} \\max_{\\|\\pmb{\\delta}_i\\|_p \\leq \\rho_a} \\ell(f_{\\theta}(\\pmb{x}_i + \\pmb{\\delta}_i), y_i)\\) where \\(\\rho_a\\) is the adversarial perturbation radius." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.431, + 0.893, + 0.658 + ], + "angle": 0, + "content": "We will use \\(\\tau_{\\mathcal{D}}(\\theta)\\), with \\(\\theta\\) the clean model parameter, to denote the clean test accuracy of a model trained on the clean training dataset (i.e., clean model accuracy). In unlearnable dataset generation, an attacker uses an algorithm \\(\\mathcal{A}:\\mathcal{X}\\rightarrow \\mathcal{X}\\) to generate an unlearnable dataset \\(\\{(\\tilde{\\boldsymbol{x}}_i = \\mathcal{A}(\\boldsymbol {x}_i),y_i)\\}_{i = 1}^n\\sim \\tilde{\\mathcal{D}}^n\\) from the clean training data. Here, the attacker assumes access to the full clean training dataset. Moreover, the attacker cannot modify the unlearnable dataset once it is released publicly. A defender trains using the unlearnable dataset to obtain a network \\(f_{\\tilde{\\theta}}\\). The objective of the attacker is to design an unlearnable dataset such that the defender's model trained on the unlearnable data achieves a clean test accuracy (i.e., unlearnable model accuracy) worse than the clean model accuracy i.e. \\(\\tau_{\\mathcal{D}}(\\tilde{\\theta})\\ll \\tau_{\\mathcal{D}}(\\theta)\\)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.665, + 0.76, + 0.682 + ], + "angle": 0, + "content": "3.2. Limitations of existing works" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.688, + 0.892, + 0.822 + ], + "angle": 0, + "content": "Fu et al. [11] show that the previous unlearnability methods including Error-Minimization (EM) [17], Targeted Adversarial Poisoning (TAP) [10], and Neural Tangent Generalization Attack (NTGA) [53] are vulnerable to AT. Hence, they propose a Robust Error-Minimization (REM) [11] method that exploits a min-min-max optimization procedure to generate unlearnable noises. REM first trains a noise generator \\( f_{\\theta}^{\\prime} \\) on data points \\( \\{(x_i,y_i)\\}_{i = 1}^n \\) over a loss function \\( \\ell \\) as follows" + }, + { + "type": "equation", + "bbox": [ + 0.558, + 0.83, + 0.892, + 0.898 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\min _ {\\theta} \\frac {1}{n} \\sum_ {i = 1} ^ {n} \\min _ {\\| \\delta_ {i} ^ {u} \\| \\leq \\rho_ {u}} \\mathbb {E} _ {t \\sim T} \\\\ \\max _ {\\| \\boldsymbol {\\delta} _ {i} ^ {a} \\| \\leq \\rho_ {a}} \\ell \\left(f _ {\\theta} ^ {\\prime} \\left(t \\left(\\boldsymbol {x} _ {i} + \\boldsymbol {\\delta} _ {i} ^ {u}\\right) + \\boldsymbol {\\delta} _ {i} ^ {a}\\right), y\\right). \\tag {1} \\\\ \\end{array}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3864" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.09, + 0.47, + 0.118 + ], + "angle": 0, + "content": "Table 1. Time taken for generating unlearnable noises for various datasets using different unlearnability techniques." + }, + { + "type": "table", + "bbox": [ + 0.08, + 0.129, + 0.47, + 0.194 + ], + "angle": 0, + "content": "
DatasetEMTAPNTGAREMCUDA
CIFAR-100.4 hr0.5 hr5.2 hrs22.6 hrs10.8 s
CIFAR-1000.4 hr0.5 hr5.2 hrs22.6 hrs15.5 s
ImageNet-1003.9 hrs5.2 hrs14.6 hrs51.2 hrs0.15 hr
" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.219, + 0.468, + 0.295 + ], + "angle": 0, + "content": "Here, \\(T\\) is a distribution over a set of transformations \\(\\{t:\\) \\(\\mathcal{X}\\to \\mathcal{X}\\}\\), \\(\\rho_{u}\\) is the defensive perturbation radius, and \\(\\rho_{a}\\) controls the protection level of REM against AT. After training the noise generator, an unlearnable example \\((\\tilde{\\pmb{x}},y)\\) is generated via" + }, + { + "type": "equation", + "bbox": [ + 0.126, + 0.306, + 0.468, + 0.36 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\tilde{\\boldsymbol{x}} = \\boldsymbol {x} + \\operatorname *{arg min}_{\\| \\boldsymbol{\\delta}^{u}\\| \\leq \\rho_{u}}\\mathbb{E}_{t\\sim T} \\\\ \\max _ {\\| \\boldsymbol {\\delta} ^ {a} \\| \\leq \\rho_ {a}} \\ell \\left(f _ {\\theta} ^ {\\prime} \\left(t \\left(\\boldsymbol {x} + \\boldsymbol {\\delta} ^ {u}\\right) + \\boldsymbol {\\delta} ^ {a}\\right), y\\right) \\tag {2} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.371, + 0.47, + 0.792 + ], + "angle": 0, + "content": "First, note that REM is computationally expensive since it needs to generate unlearnability noises through solving optimization equation 2. Moreover, the existing techniques are model-dependent and they require gradient-based training with a neural network to generate unlearnable data. They also require neural network training from scratch for every dataset that is to be made unlearnable. Table 1 shows the amount of time required to generate various unlearnable datasets using NVIDIA® Tesla V100 GPU and 10 CPU cores. CUDA generation is significantly faster than the existing methods since it uses a model-free approach (no training required). Furthermore, REM is sensitive to hyperparameters and norm-budgets of AT since they generate noises with fixed \\( L_{\\infty} \\) norm budgets. For instance, a ResNet-18 trained on clean CIFAR-10 dataset achieves a clean test data accuracy of \\( 94.66\\% \\). \\( L_{\\infty} \\) AT with perturbation radius \\( \\rho_{a} = 4 / 255 \\) on REM CIFAR-10 data (generated using \\( \\rho_{u} = 8 / 255 \\) and \\( \\rho_{a} = 4 / 255 \\)) achieves only a clean test accuracy of \\( 48.16\\% \\). However, \\( L_{\\infty} \\) AT with perturbation radius \\( \\rho_{a} = 8 / 255 \\) and \\( L_{2} \\) AT with perturbation radius \\( \\rho_{a} = 0.75 \\) can achieve a clean test accuracy of \\( 78.71\\% \\) and \\( 79.65\\% \\), respectively, on the same REM data. We also find that ERM with a ResNet-18 on grayscaled REM CIFAR-10 images can achieve a high test accuracy of \\( 70.76\\% \\) on the grayscaled CIFAR-10 test data. This shows that REM relies upon the color space for poisoning clean data. Fu et al. [11] also show that REM noise generated using ResNet-18 is not transferable to DenseNet-121." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.802, + 0.272, + 0.817 + ], + "angle": 0, + "content": "3.3. Our method: CUDA" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.826, + 0.468, + 0.901 + ], + "angle": 0, + "content": "The major limitations of the previous works are that they are vulnerable to AT, and computationally expensive. We think the major reason for the former limitation is the usage of small additive noises for unlearnability. AT is designed to train in the presence of such additive noises. In" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.408 + ], + "angle": 0, + "content": "creasing the budget of the amount of additive noises for unlearnability might destroy the semantics of the images while perturbing them. The latter limitation arises from the fact that these methods are model-dependent and they require multilevel optimizations. Hence, we are motivated to design a compute-efficient unlearnability method that is robust to AT. CUDA technique can perform convolutions to add larger amounts of noises to clean images without destroying its semantics. This can help CUDA to be robust against AT. CUDA uses randomly generated convolution filters for blurring images from each class. This makes a model trained on CUDA to learn shortcut relations between filters and labels. We empirically support these claims in Section 5.2. Additionally, randomly generating the filters makes our technique model-free. Since the keys for generating the filters are private, it is not possible to reverse the blurring effect in CUDA without having access to the corresponding clean images. Hence, we assume that the data publisher deletes the clean images after perturbing them. Moreover, CUDA technique is a novel non-additive noise based poisoning attack that needs to be studied." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.408, + 0.892, + 0.665 + ], + "angle": 0, + "content": "CUDA uses convolutional filters \\( s_i \\in \\mathbb{R}^{k \\times k} \\) for each class \\( i \\in [1, K] \\). A random parameter, out of the \\( k^2 \\) parameters, in each of the filters is set to have a value of 1. The rest of the filter parameters are randomly initialized from a uniform distribution \\( \\mathcal{U}(0, p_b) \\) using a private seed where \\( p_b \\) is the blur parameter. Blur parameter controls the level of blurring that occurs when an image \\( x \\in [0, 1]^{d_1 \\times d_2 \\times d_3} \\) is convolved with a filter \\( s_i \\). Here, \\( d_1, d_2 \\), and \\( d_3 \\) are the height, width, and number of channels of the image, respectively. For example, a CIFAR-10 image has a dimension of \\( 32 \\times 32 \\times 3 \\). The higher \\( p_b \\) is, the higher the blurring effect is. Let \\( \\hat{x} = x * s_i \\) where \\( x \\) belongs to class \\( i \\). The CUDA data point for \\( x \\) is given by \\( \\hat{x} = \\hat{x} / \\mathrm{MAX}(\\hat{x}) \\). Rescaling is performed to make sure that the CUDA image pixels lie between 0 and 1. We find that the unlearnability effect gets stronger with larger \\( p_b \\) and \\( k \\) values (see supplementary material for details)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.68, + 0.676, + 0.697 + ], + "angle": 0, + "content": "4. Theory for CUDA" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.706, + 0.892, + 0.827 + ], + "angle": 0, + "content": "In this section, we define a binary classification setup similar to [19, 35] to theoretically analyze CUDA. Let \\(\\mathcal{D}\\) be a clean dataset modelled by an isotropic Gaussian mixture model given by \\(\\mathcal{N}(y\\pmb{\\mu}, I)\\), where \\(y \\in \\{\\pm 1\\}\\) is the class label, \\(\\pmb{\\mu} \\in \\mathbb{R}^d\\), and \\(I \\in \\mathbb{R}^{d \\times d}\\) is the identity matrix. We defer the proofs for all the lemmas and theorem to the supplementary material. The Bayes optimal decision boundary for classifying this Gaussian mixture model is as follows." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.841, + 0.892, + 0.903 + ], + "angle": 0, + "content": "Lemma 1 The Bayes optimal decision boundary for classifying \\(\\mathcal{D}\\) is given by \\(P(\\pmb{x}) \\equiv \\pmb{\\mu}^{\\top} \\pmb{x} = 0\\). The accuracy of the decision boundary \\(P\\) on the clean dataset \\(\\mathcal{D}\\) (i.e. \\(\\tau_{\\mathcal{D}}(P)\\)) is equal to \\(\\phi(\\|\\pmb{\\mu}\\|_2)\\). Here, \\(\\phi(.)\\) represents the CDF of the" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3865" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.251, + 0.089, + 0.4, + 0.262 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.28, + 0.265, + 0.369, + 0.28 + ], + "angle": 0, + "content": "(a) \\(\\| \\pmb {\\mu}\\| _2 = 1.0\\)" + }, + { + "type": "image", + "bbox": [ + 0.412, + 0.089, + 0.56, + 0.262 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.441, + 0.266, + 0.53, + 0.28 + ], + "angle": 0, + "content": "(b) \\( \\| \\pmb{\\mu} \\|_2 = 2.5 \\)" + }, + { + "type": "image", + "bbox": [ + 0.572, + 0.089, + 0.721, + 0.261 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.603, + 0.266, + 0.692, + 0.28 + ], + "angle": 0, + "content": "(c) \\(\\| \\pmb{\\mu}\\| _2 = 5.0\\)" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.291, + 0.895, + 0.335 + ], + "angle": 0, + "content": "Figure 3. We plot the contour plots for clean test accuracy of the CUDA Bayes classifier \\(\\tilde{P}\\) (given as \\(\\tau_{\\mathcal{D}}(\\tilde{P})\\)) for varying \\(a_{y}\\) parameters where \\(y \\in \\{\\pm 1\\}\\) in Figures 3(a-c). Gaussian mixture model \\(\\mathcal{N}(y\\mu, I)\\) denotes the clean data. In the top row, we use Lemmas 1 & 4 to empirically generate the plots. In the bottom row, we plot the theoretical upper bound of \\(\\tau_{\\mathcal{D}}(\\tilde{P})\\) we obtain using Theorem 2." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.36, + 0.274, + 0.374 + ], + "angle": 0, + "content": "standard normal distribution." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.386, + 0.473, + 0.598 + ], + "angle": 0, + "content": "Now, to obtain \\(\\tilde{\\mathcal{D}}\\) (i.e., CUDA), we perform class-wise 1D convolutions. Here, we consider a class of 1D convolutional filters with kernel size 3 of the form \\(f_{a} = [a,1,a]\\) where \\(a\\in [0,0.5]\\). The convolution of signal \\(\\pmb {x}\\in \\mathbb{R}^{d}\\) with filter \\(\\pmb{f}_{a}\\) using stride 1 (denoted by \\(\\pmb {x}*\\pmb{f}_{a}\\)), can be treated as a matrix operation \\(Ax\\). Here, \\(A\\in \\mathbb{R}^{d\\times d}\\) is a tri-diagonal Toeplitz matrix denoted by \\(T(d;a,1,a)\\). For CUDA generation, we use class-wise convolution matrices \\(A_{y} = T(d;a_{y},1,a_{y})\\) to perturb a clean data point \\((x,y)\\sim \\mathcal{D}\\) to an unlearnable data point \\((A_yx,y)\\), where \\(a_{y}\\in [0,0.5]\\) for \\(y\\in \\{\\pm 1\\}\\). Note that in CUDA, the labels remain the same. Next we show that such a perturbed dataset (i.e. CUDA) can be represented as a Gaussian mixture model." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.609, + 0.47, + 0.673 + ], + "angle": 0, + "content": "Lemma 2 A CUDA generated from clean data distribution \\(\\mathcal{D}\\) and the \\(A_{y}\\)'s defined above can be modelled as a Gaussian mixture model with the distribution given by \\(\\tilde{\\mathcal{D}} = \\mathcal{N}(yA_y\\pmb {\\mu},A_y^2)\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.682, + 0.47, + 0.728 + ], + "angle": 0, + "content": "To characterize the decision boundary for CUDA, we need to use some properties of Toeplitz matrices from Noschese et al. [37] given in the following Lemma 3." + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.738, + 0.468, + 0.821 + ], + "angle": 0, + "content": "Lemma 3 Any tri-diagonal Toeplitz matrix \\( A = T(d; a, 1, a) \\in \\mathbb{R}^{d \\times d} \\) with \\( a \\in [0, 0.5] \\) can be diagonalized as \\( A = QDQ \\) where \\( Q_{i,j} = \\left(\\frac{2}{d + 1}\\right)^{1/2} \\sin \\left(\\frac{ij\\pi}{d + 1}\\right) \\) and \\( D \\) is a diagonal matrix with \\( D_{i,i} = 1 + 2a \\cos \\left(\\frac{i\\pi}{d + 1}\\right) \\) for \\( 1 \\leq i, j \\leq d \\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.831, + 0.468, + 0.862 + ], + "angle": 0, + "content": "Next we use Lemma 3 to show that the Bayes optimal decision boundary for classifying \\(\\tilde{\\mathcal{D}}\\) is a quadratic plane." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.87, + 0.47, + 0.902 + ], + "angle": 0, + "content": "Lemma 4 Let \\( A_{-1} = T(d; a_{-1}, 1, a_{-1}) \\) and \\( A_1 = T(d; a_1, 1, a_1) \\). The Bayes optimal decision boundary for" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.358, + 0.893, + 0.413 + ], + "angle": 0, + "content": "classifying \\(\\tilde{\\mathcal{D}}\\) is given by \\(\\tilde{P} (\\pmb {x})\\equiv \\pmb{x}^{\\top}A\\pmb {x} + \\pmb{b}^{\\top}\\pmb {x} + c = 0,\\) where \\(A = A_{-1}^{-2} - A_1^{-2},\\pmb {b} = 2(A_{-1}^{-1} + A_{1}^{-1})\\pmb {\\mu},\\) and \\(c = 2\\sum_{i = 1}^{d}\\left[\\ln (1 + 2a_{-1}\\cos (\\frac{i\\pi}{d + 1})) - \\ln (1 + 2a_{1}\\cos (\\frac{i\\pi}{d + 1}))\\right].\\)" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.418, + 0.892, + 0.463 + ], + "angle": 0, + "content": "Now we state a lemma regarding the tail of Gaussian quadratic forms which plays a crucial role in our main result." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.473, + 0.893, + 0.535 + ], + "angle": 0, + "content": "Lemma 5 Let \\(\\| \\cdot \\|\\) denote the operator norm, \\(\\| \\cdot \\|_2\\) denote the vector 2-norm, \\(z \\sim \\mathcal{N}(\\mathbf{0}, I)\\), and \\(Z = z^\\top A z + z^\\top b + c\\) where \\(A = Q \\Lambda Q^\\top\\). Using Chernoff bound, for any \\(t \\geq 0\\) and \\(\\gamma \\in \\mathbb{R}\\)" + }, + { + "type": "equation", + "bbox": [ + 0.5, + 0.561, + 0.891, + 0.621 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathbb {P} \\{Z - \\mathbb {E} Z \\geq \\gamma \\} \\leq \\\\ \\frac {\\exp \\left\\{\\frac {- t}{4 \\| \\boldsymbol {b} \\| _ {2} ^ {\\frac {2}{2}} \\| \\boldsymbol {\\Lambda} \\|} - t (\\gamma + \\operatorname {T r} (\\boldsymbol {\\Lambda}) + \\| \\boldsymbol {b} \\| _ {2}) \\right\\}}{| I - 2 t \\boldsymbol {\\Lambda} | ^ {\\frac {1}{2}}}. \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.63, + 0.892, + 0.676 + ], + "angle": 0, + "content": "Lemma 5 allows us to provide an upper bound for the accuracy of the unlearnable decision boundary \\(\\tilde{P}\\) on the clean dataset \\(\\mathcal{D}\\), given as \\(\\tau_{\\mathcal{D}}(\\tilde{P})\\), in Theorem 2 below." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.687, + 0.894, + 0.749 + ], + "angle": 0, + "content": "Theorem 2 (Main result) Let \\( A = Q\\Lambda Q = A_{-1}^{-2} - A_{1}^{-2} \\). For any non-negative constants \\( t_1 \\) and \\( t_2 \\), the accuracy of the unlearnable decision boundary \\( \\tilde{P} \\) on the clean dataset \\( \\mathcal{D} \\) can be upper-bounded as" + }, + { + "type": "equation", + "bbox": [ + 0.504, + 0.758, + 0.885, + 0.858 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\tau_ {\\mathcal {D}} (\\tilde {P}) \\leq \\frac {\\exp \\left[ t _ {1} \\left(\\boldsymbol {b} ^ {\\top} \\boldsymbol {\\mu} + \\boldsymbol {\\mu} ^ {\\top} A \\boldsymbol {\\mu} + c - \\frac {1}{4 \\| 2 A \\boldsymbol {\\mu} + \\boldsymbol {b} \\| _ {2} \\| \\Lambda \\|}\\right) \\right]}{2 | I - 2 t _ {1} \\Lambda | ^ {\\frac {1}{2}}} \\\\ + \\frac {\\exp \\left[ t _ {2} \\left(\\boldsymbol {b} ^ {\\top} \\boldsymbol {\\mu} - \\boldsymbol {\\mu} ^ {\\top} A \\boldsymbol {\\mu} - c - \\frac {1}{4 \\| 2 A \\boldsymbol {\\mu} - \\boldsymbol {b} \\| _ {2} \\| \\Lambda \\|}\\right) \\right]}{2 | I - 2 t _ {2} \\Lambda | ^ {\\frac {1}{2}}} \\\\ := p _ {1} + p _ {2}. \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.869, + 0.892, + 0.903 + ], + "angle": 0, + "content": "Furthermore, if \\(\\pmb{\\mu}^{\\top}A\\pmb{\\mu} + \\pmb{b}^{\\top}\\pmb{\\mu} + c + \\mathrm{Tr}(A) + \\| 2A\\pmb{\\mu} + \\pmb{b}\\| < 0\\) we have \\(\\tau_{\\mathcal{D}}(\\tilde{P})\\leq \\frac{1}{2} (p_1 + 1) < 1\\). Also, if \\(-\\pmb{\\mu}^{\\top}A\\pmb{\\mu} + \\pmb{b}^{\\top}\\pmb{\\mu}-\\)" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3866" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.09, + 0.47, + 0.138 + ], + "angle": 0, + "content": "\\(c - \\operatorname {Tr}(A) + \\| 2A\\pmb {\\mu} - \\pmb {b}\\| < 0\\) we have \\(\\tau_{\\mathcal{D}}(\\tilde{P})\\leq \\frac{1}{2} (1 + p_2) < 1.\\) Moreover, for any \\(\\pmb {\\mu}\\neq \\mathbf{0}\\) and \\(a_{-1}\\in [0,0.5]\\) , \\(\\exists a_{1}\\) such that \\(\\tau_{\\mathcal{D}}(\\tilde{P}) < \\tau_{\\mathcal{D}}(P)\\)" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.15, + 0.473, + 0.422 + ], + "angle": 0, + "content": "Poisoning is effective only if the accuracy of the unlearnable model \\(\\tilde{P}\\) is less than that of the clean model \\(P\\) on the clean dataset \\(\\mathcal{D}\\), that is, \\(\\tau_{\\mathcal{D}}(\\tilde{P}) < \\tau_{\\mathcal{D}}(P)\\). To satisfy this condition, we need to carefully select \\(a_{y}\\)'s. In Theorem 2, we formally state this condition1. Theorem 2 shows that CUDA can effectively poison when there are two distinct modes in the clean Gaussian mixture data model. We validate our theoretical claim through empirical analysis as well (see Figure 3). Details for the analysis is given in supplementary material. We find that our upper bound for the clean test accuracy of CUDA classifiers are consistent with our empirical analysis. In our experiments, we also find that the unlearnability effect is stronger with a larger blur parameter. This effect is evident from Figure 3 where we find that it is likely to get a lower \\(\\tau_{\\mathcal{D}}(\\tilde{P})\\) with higher \\(a_{y}\\) values. These results are consistent with our experimental results in Section 5 with CIFAR-10, CIFAR-100, and ImageNet-100 datasets." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.437, + 0.21, + 0.455 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.463, + 0.473, + 0.675 + ], + "angle": 0, + "content": "In this section, we first discuss our experimental setup. More details on the setup is deferred to supplementary material. We then show the robustness of CUDA generation with various datasets and architectures. We also run various experiments to analyze the effectiveness of CUDA under different training techniques (ERM, AT with varying budgets, randomized smoothing, transfer learning, and finetuning) and augmentation techniques (mixup [56], grayscale, random blurring, cutout [9], cutmix [54], autoaugment [7], and orthogonal regularization [3]). Finally, we also design adaptive defenses to test the robustness of CUDA. One might think that CUDA filters can be obtained by adversarially training them with the data. We show that CUDA is robust to such adaptive defenses that we design." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.684, + 0.265, + 0.701 + ], + "angle": 0, + "content": "5.1. Experimental setup" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.708, + 0.47, + 0.784 + ], + "angle": 0, + "content": "Datasets. We use three image classification datasets - CIFAR-10, CIFAR-100, [21] and ImageNet-100 (a subset of ImageNet made of the first 100 classes) [39]. We use the data augmentation techniques such as random flipping, cropping, and resizing [43]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.784, + 0.47, + 0.83 + ], + "angle": 0, + "content": "Architectures. We use ResNet-18 [14], VGG-16 [44], Wide ResNet-34-10 [55], and DenseNet-121 [16]. We train the networks with hyperparameters used in Fu et al. [11]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.838, + 0.47, + 0.902 + ], + "angle": 0, + "content": "1We note that the conditions \\(\\pmb{\\mu}^{\\top}A\\pmb{\\mu} + \\pmb{b}^{\\top}\\pmb{\\mu} + c + \\mathrm{Tr}(A) + \\|2A\\pmb{\\mu} + \\pmb{b}\\|_{2} < 0\\) or \\(-\\pmb{\\mu}^{\\top}A\\pmb{\\mu} + \\pmb{b}^{\\top}\\pmb{\\mu} - c - \\mathrm{Tr}(A) + \\|2A\\pmb{\\mu} - \\pmb{b}\\|_{2} < 0\\) can always be satisfied by picking a sufficiently large \\(\\pmb{\\mu}\\) in the direction of an eigenvector corresponding to a negative or positive eigenvalue of \\(A\\) (note that \\(A\\) has negative and positive eigenvalues)." + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.089, + 0.892, + 0.118 + ], + "angle": 0, + "content": "Table 2. Test accuracy (\\%) of ResNet-18 trained on various unlearnable datasets. We use \\( L_{\\infty} \\) AT with budget \\( \\rho_{a} = 4 / 255 \\)." + }, + { + "type": "table", + "bbox": [ + 0.507, + 0.129, + 0.888, + 0.243 + ], + "angle": 0, + "content": "
DatasetTraining methodCleanUnlearnability method
EMTAPNTGAREMCUDA
CIFAR-10ERM94.6613.2022.5116.2727.0918.48
AT89.5188.6288.0288.9648.1644.40
CIFAR-100ERM76.271.6013.753.2210.1412.69
AT64.5063.4362.3962.4427.1034.34
ImageNet-100ERM80.661.269.108.4213.748.96
AT66.6263.4063.5663.0641.6638.68
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.258, + 0.892, + 0.287 + ], + "angle": 0, + "content": "Table 3. Test accuracy \\((\\%)\\) of network architectures trained on various CIFAR-10 unlearnable datasets with \\(L_{\\infty}\\) AT \\((\\rho_{a} = 4 / 255)\\)." + }, + { + "type": "table", + "bbox": [ + 0.508, + 0.297, + 0.888, + 0.398 + ], + "angle": 0, + "content": "
ModelCleanUnlearnability method
EMTAPNTGAREMCUDA
ResNet-1889.5188.6288.0288.9648.1644.40
VGG-1687.5186.4886.2786.6565.2342.98
Wide ResNet-34-1091.2190.0590.2389.9548.3953.02
DenseNet-12183.2782.4481.7280.7381.4845.95
" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.424, + 0.892, + 0.485 + ], + "angle": 0, + "content": "Previous works mainly employ ResNet-18 for most of their evaluations. Additional experiments on Tiny-ImageNet [26], DeIT [48], EfficientNetV2-S [46], and MobileNetV2 [40] are provided in the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.486, + 0.893, + 0.652 + ], + "angle": 0, + "content": "CUDA generation. We use filters of size \\( k = 3 \\) and blur parameter \\( p_b = 0.3 \\) for both CIFAR-10 and CIFAR-100 datasets. ImageNet-100 is a higher dimensional \\( 224 \\times 224 \\times 3 \\) image dataset when compared to the \\( 32 \\times 32 \\times 3 \\) dimensional CIFAR datasets. Hence, we use larger filters of size \\( k = 9 \\) with \\( p_b = 0.06 \\) for ImageNet-100. These hyperparameters are chosen such that the CUDA images are not perceptibly highly perturbed and give good unlearnability effect (see plot in Figure 1). In supplementary material, we show the results of training on CUDA with different hyperparameters for data generation." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.653, + 0.894, + 0.835 + ], + "angle": 0, + "content": "Baselines. We compare CUDA generation technique with four state-of-the-art unlearnability methods - REM [11], EM [17], TAP [10], and NTGA [53]. We adopt the results reported in [11] since we use the same hyperparameters for training. For REM, we select hyperparameters \\(\\rho_{u} = 8 / 255\\) and \\(\\rho_{a} = 4 / 255\\), the highest radii values in [11]. For comparing unlearnable methods, we look at the clean test accuracy. The lower the test accuracy, the better the unlearnability method is. As mentioned in the supplementary material, we use publicly released official codebases for reproducing the baselines using their default hyperparameters." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.846, + 0.712, + 0.862 + ], + "angle": 0, + "content": "5.2. Effectiveness of CUDA" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.892, + 0.902 + ], + "angle": 0, + "content": "Why does CUDA work? In order to measure how much CUDA technique's blurring affects the dataset's quality, we" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3867" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.089, + 0.47, + 0.118 + ], + "angle": 0, + "content": "Table 4. Test accuracy \\((\\%)\\) of ResNet-18 with CUDA under various training settings." + }, + { + "type": "table", + "bbox": [ + 0.081, + 0.129, + 0.468, + 0.32 + ], + "angle": 0, + "content": "
DatasetCleanTraining methodCUDA (ours)
CIFAR-1094.66ERM18.48
AT L∞ (ρa = 4/255)44.40
AT L∞ (ρa = 8/255)32.85
AT L∞ (ρa = 16/255)19.32
AT L2 (ρa = 0.25)39.05
AT L2 (ρa = 0.50)51.19
AT L2 (ρa = 0.75)51.14
CIFAR-10076.27ERM12.69
AT L∞ (ρa = 4/255)34.34
AT L∞ (ρa = 8/255)30.00
AT L2 (ρa = 0.75)36.90
ImageNet-10080.66ERM8.96
AT L∞ (ρa = 4/255)38.68
AT L∞ (ρa = 8/255)40.08
AT L2 (ρa = 0.75)20.58
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.352, + 0.473, + 0.867 + ], + "angle": 0, + "content": "compare class-wise blurring (CUDA technique) against universal blurring where a single convolutional filter is used for blurring all the images. We keep the filter generation parameters fixed (\\(p_b = 0.3\\) and \\(k = 3\\)) for both CUDA class-wise blurring and universal blurring. ResNet-18 trained with clean CIFAR-10, universally blurred CIFAR-10, and CIFAR-10 CUDA achieve clean test accuracies of \\(94.66\\%\\), \\(90.47\\%\\), and \\(18.48\\%\\), respectively. This suggests that our controlled blurring does not obscure the semantics of the dataset. Hence, the significant drop in the clean test accuracy introduced by CUDA is most likely due to the usage of class-wise filters. This suggests that a model trained on CUDA learns the relation between the class-wise convolution filters and their corresponding labels. Therefore, during test time when this convolution effect is absent, the CUDA trained model fails to make correct classifications. Furthermore, the model trained on CUDA achieves an accuracy of \\(99.91\\%\\) on the CIFAR-10 CUDA testset. This strongly supports our claim that the CUDA model learns to classify images based on the convolutional filters used to blur them. In addition, if we permute the class-wise filters for blurring the test set (i.e., blurring class 1 images with class 2 filters, class 2 images with class 3 filters, and so on), we get a very low accuracy of \\(2.53\\%\\) on this test set. Further details are provided in the supplementary material. Finally, we note that real-world datasets might also contain blurred images due to various factors such as motion blurring, weather conditions, issues with the camera, etc. Hence, detecting if a blurred image is poisoned might not always be possible. However, one might argue that it is possible to detect if an entire dataset is blurred. Interestingly, later in this section we show that CUDA technique exhibits unlearnability effect even when only a fraction of the training dataset is poisoned." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.472, + 0.903 + ], + "angle": 0, + "content": "Different datasets. We first compare the effectiveness of CUDA with different datasets using ERM and \\( L_{\\infty} \\) AT with" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.228 + ], + "angle": 0, + "content": "\\(\\rho_{a} = 4 / 255\\). We use ResNet-18 for the experiments. The results are shown in Table 2. These results show that EM, TAP, and NTGA are not robust to AT. However, both CUDA and REM are successful. Here, our method CUDA outperforms REM with CIFAR-10 and ImageNet-100 datasets. Smartly designed additive noise in AT helps in achieving better generalization than ERM on the unlearnable datasets. This experiment thus demonstrates that ERM and AT are not good choices for training with CUDA and REM dataset." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.23, + 0.892, + 0.336 + ], + "angle": 0, + "content": "Different models. Next we compare the effectiveness of CUDA using various deep learning architectures with \\( L_{\\infty} \\) AT (\\( \\rho_{a} = 4 / 255 \\)). We use CIFAR-10 for the experiments. The results are shown in Table 3. As we see in the table, CUDA is effective with all the five network architectures. However, REM is not seen to be transferrable with DenseNet-121." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.338, + 0.893, + 0.61 + ], + "angle": 0, + "content": "Robustness to different training settings. In Section 3.2, we show that REM is sensitive to the training settings. REM generated using \\( L_{\\infty} \\) radii budgets of \\( \\rho_u = 8 / 255 \\) and \\( \\rho_a = 4 / 255 \\) for CIFAR-10 breaks with \\( L_{\\infty} \\) AT \\( (\\rho_a = 8 / 255) \\) and \\( L_2 \\) AT \\( (\\rho_a = 0.75) \\) to get test accuracy of \\( 78.71\\% \\) and \\( 79.65\\% \\), respectively. Hence, we run experiments to check the robustness of CUDA with various AT norm budgets. The results are shown in Table 4. As we see in the table, CUDA is robust to ERM, \\( L_{\\infty} \\), and \\( L_2 \\) AT settings with varying training budgets. Impressively, the highest test accuracy achieved with training on CIFAR-10 CUDA, CIFAR-100 CUDA, and ImageNet-100 CUDA are as low as \\( 51.19\\% \\), \\( 36.90\\% \\), and \\( 40.08\\% \\), respectively. We also find that using a pre-trained ResNet-18 with CIFAR-10 CUDA only achieves clean test accuracy of \\( 42.42\\% \\) and \\( 48.22\\% \\) with fine-tuning the full network and a newly trained final layer, respectively (details are deferred to the supplementary material)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.612, + 0.892, + 0.734 + ], + "angle": 0, + "content": "Different protection percentages. In Section 3.1, we assume that the attacker has access to the full clean training data. However, in real life settings, this might not be always possible. Hence, we train ResNet-18 on a mix of CIFAR-10 CUDA and clean CIFAR-10 training datasets to evaluate the effectiveness of poisoning with varying data protection percentages. Protection percentage denotes the percentage of the training data that is poisoned." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.735, + 0.895, + 0.903 + ], + "angle": 0, + "content": "We show the results in Table 5. In the table, the \"Mixed\" column denotes the clean test accuracy of a model trained using a mix of both clean and poisoned data. The \"Clean\" column denotes the clean test accuracy of a model trained only using the clean subset of the training data. In the last row of Table 5, we provide the results for CUDA trained using ERM with varying data protection percentages. The remainder of the rows provide the results for the \\( L_{\\infty} \\) AT \\( (\\rho_{a} = 4 / 255) \\) scenario. For example, CUDA trained with ERM using an \\( 80\\% \\) clean training data partition achieves a test accuracy of \\( 93.75\\% \\). Adding the \\( 20\\% \\) CUDA data" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "3868" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.089, + 0.894, + 0.12 + ], + "angle": 0, + "content": "Table 5. Test accuracy (\\%) of ResNet-18 on CIFAR-10 with different data protection percentages. The last row shows the results for CUDA with ERM setting. The rest of the rows show results for unlearnability methods trained in the \\( L_{\\infty} \\) AT (\\( \\rho_a = 4/255 \\)) setting." + }, + { + "type": "table", + "bbox": [ + 0.172, + 0.128, + 0.805, + 0.269 + ], + "angle": 0, + "content": "
Unlearnability methodData Protection Percentage
0%20% MixedClean40% MixedClean60% MixedClean80% MixedClean100%
EM89.5189.6088.1789.4086.7689.4985.0789.1079.4188.62
TAP89.0188.6688.4088.0488.02
NTGA89.5689.3589.2289.1788.96
REM89.6089.3489.6188.0948.16
CUDA (ours)88.5487.2486.0384.3444.40
CUDA + ERM94.6693.2893.7591.3492.5689.9189.7785.6184.3018.48
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.291, + 0.473, + 0.426 + ], + "angle": 0, + "content": "partition to the training dataset drops the test accuracy of the model to \\(93.28\\%\\). Results from the last row of Table 5 show that CUDA with varying data protection percentages is effective with the ERM setting. However, with the AT scenario, the unlearnability techniques are not as successful as with ERM with varying data protection percentages. Nevertheless, it is interesting to note that CUDA technique performs better than all the other unlearnability techniques with AT." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.429, + 0.473, + 0.686 + ], + "angle": 0, + "content": "Robustness to smoothing and data augmentations. Cohen et al. [6] proposes randomized smoothing which is a provable adversarial defense in \\( L_{2} \\) norm. Since CUDA does not use additive noise, CUDA is robust to randomized smoothing. A smoothed ResNet-18 (with a noise level of 0.5) with CIFAR-10 CUDA achieves only a clean test accuracy of \\( 43.85\\% \\). In Section 3.2, we see that REM breaks with grayscaleing. While ResNet-18 training using grayscaleed REM achieves \\( 70.76\\% \\) test accuracy, grayscaleed CUDA only achieves \\( 20.12\\% \\) test accuracy on the clean grayscaleed CIFAR-10 test data. This shows that CUDA technique exhibits the desirable property of not relying upon the color space for its attack. CIFAR-10 CUDA training achieves only \\( 25.53\\% \\), \\( 25.80\\% \\), \\( 26.93\\% \\), \\( 34.09\\% \\), and \\( 50.72\\% \\) clean test accuracy with mixup, cutout, cutmix, autoaugment, and orthogonal regularization, respectively (see supplementary material for more details)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.688, + 0.472, + 0.822 + ], + "angle": 0, + "content": "Adaptive defenses for CUDA. Here, we first investigate the effect of training CIFAR-10 CUDA with random blurring augmentations using ResNet-18. Each batch of the CUDA training data is convolved with random \\(3 \\times 3\\) filters of varying blur parameters \\(p_b'\\). With \\(p_b'\\) values of 0.1 and 0.3, CUDA training achieves lower test accuracy \\(9.2\\%\\) and \\(13.37\\%\\), respectively. This shows that the noise added by CUDA technique is robust to random convolution augmentations." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.472, + 0.903 + ], + "angle": 0, + "content": "One may think that CUDA technique can be broken by learning the private filters from the data. We test this idea by training deconvolution filters to find if we can reverse the blurring effect in CUDA with adversarially trained filters. We use a novel Deconvolution-based Adversarial Training" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.291, + 0.895, + 0.519 + ], + "angle": 0, + "content": "(DAT) technique that is similar to AT (check supplementary material for details). While the adversarial step in AT learns sample-wise error-maximizing additive noises, the adversarial step in DAT learns class-wise error-maximizing deconvolution filters. We train DAT with filters of varying sizes (3, 5, and 7) on CIFAR-10 CUDA using ResNet-18. The filter parameters are constrained within a finite range to make sure that the images do not get distorted with the adversarial step similar to projection in projected gradient descent. We find that CUDA is robust to DAT. DAT using filters of size 3, 5, and 7 with CUDA achieves only test accuracy of \\(39.05\\%\\), \\(46.21\\%\\), and \\(38.48\\%\\), respectively. DAT is not successful against CUDA since we can not invert convolutions without the knowledge of the private filters or clean images corresponding to CUDA." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.523, + 0.895, + 0.751 + ], + "angle": 0, + "content": "Limitations and future directions. The unlearnability effect of CUDA can be defended if some fraction of the clean data and its corresponding CUDA images are leaked. The defender must also be able to detect if all samples in the dataset are poisoned. However, our work assumes a setup where the filters remain private. For example, a data publisher could simply publish their CUDA images and delete the clean images permanently to prevent this scenario. As discussed in this section, CUDA as well as other prior works do not perform well with different protection percentages with AT. Improving this can be an interesting research direction. We believe that extending CUDA technique to other domains such as tabular and text data is also an interesting future direction. It would also be interesting to see theoretical analysis of CUDA considering more complex setups." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.78, + 0.661, + 0.798 + ], + "angle": 0, + "content": "Acknowledgement" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.811, + 0.895, + 0.901 + ], + "angle": 0, + "content": "This project was supported in part by Meta grant 23010098, NSF CAREER AWARD 1942230, HR001119S0026 (GARD), ONR YIP award N00014-22-1-2271, Army Grant No. W911NF2120076, a capital one grant, NIST 60NANB20D134 and the NSF award CCF2212458." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3869" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.116, + 0.47, + 0.171 + ], + "angle": 0, + "content": "[1] Mauro Barni, Kassem Kallas, and Benedetta Tondi. A new backdoor attack in cnns by training set corruption without label poisoning. In 2019 IEEE International Conference on Image Processing (ICIP), pages 101-105. IEEE, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.173, + 0.471, + 0.214 + ], + "angle": 0, + "content": "[2] Battista Biggio, Blaine Nelson, and Pavel Laskov. Poisoning attacks against support vector machines. arXiv preprint arXiv:1206.6389, 2012. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.216, + 0.471, + 0.257 + ], + "angle": 0, + "content": "[3] Andrew Brock, Theodore Lim, James M Ritchie, and Nick Weston. Neural photo editing with introspective adversarial networks. arXiv preprint arXiv:1609.07093, 2016. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.259, + 0.471, + 0.3 + ], + "angle": 0, + "content": "[4] Nicholas Carlini and David Wagner. Towards evaluating the robustness of neural networks. In 2017 IEEE symposium on security and privacy (sp), pages 39-57. IEEE, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.302, + 0.471, + 0.357 + ], + "angle": 0, + "content": "[5] Xinyun Chen, Chang Liu, Bo Li, Kimberly Lu, and Dawn Song. Targeted backdoor attacks on deep learning systems using data poisoning. arXiv preprint arXiv:1712.05526, 2017.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.36, + 0.471, + 0.442 + ], + "angle": 0, + "content": "[6] Jeremy Cohen, Elan Rosenfeld, and Zico Kolter. Certified adversarial robustness via randomized smoothing. In Kamalika Chaudhuri and Ruslan Salakhutdinov, editors, Proceedings of the 36th International Conference on Machine Learning, volume 97 of Proceedings of Machine Learning Research, pages 1310-1320. PMLR, 09-15 Jun 2019. 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.444, + 0.471, + 0.498 + ], + "angle": 0, + "content": "[7] Ekin D Cubuk, Barret Zoph, Dandelion Mane, Vijay Vasudevan, and Quoc V Le. Autoaugment: Learning augmentation policies from data. arXiv preprint arXiv:1805.09501, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.502, + 0.472, + 0.557 + ], + "angle": 0, + "content": "[8] Karan Desai and Justin Johnson. Virtex: Learning visual representations from textual annotations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11162-11173, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.559, + 0.471, + 0.599 + ], + "angle": 0, + "content": "[9] Terrance DeVries and Graham W Taylor. Improved regularization of convolutional neural networks with cutout. arXiv preprint arXiv:1708.04552, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.602, + 0.47, + 0.656 + ], + "angle": 0, + "content": "[10] Liam Fowl, Micah Goldblum, Ping-yeh Chiang, Jonas Geiping, Wojtek Czaja, and Tom Goldstein. Adversarial examples make strong poisons. arXiv preprint arXiv:2106.10807, 2021. 1, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.659, + 0.47, + 0.714 + ], + "angle": 0, + "content": "[11] Shaopeng Fu, Fengxiang He, Yang Liu, Li Shen, and Dacheng Tao. Robust unlearnable examples: Protecting data privacy against adversarial learning. In International Conference on Learning Representations, 2021. 1, 3, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.716, + 0.47, + 0.757 + ], + "angle": 0, + "content": "[12] Ian J Goodfellow, Jonathon Shlens, and Christian Szegedy. Explaining and harnessing adversarial examples. arXiv preprint arXiv:1412.6572, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.759, + 0.47, + 0.813 + ], + "angle": 0, + "content": "[13] Tianyu Gu, Brendan Dolan-Gavitt, and Siddharth Garg. Badnets: Identifying vulnerabilities in the machine learning model supply chain. arXiv preprint arXiv:1708.06733, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.816, + 0.47, + 0.857 + ], + "angle": 0, + "content": "[14] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. CoRR, abs/1512.03385, 2015. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.859, + 0.47, + 0.9 + ], + "angle": 0, + "content": "[15] Kashmir Hill. The secretive company that might end privacy as we know it. In *Ethics of Data and Analytics*, pages 170–177. Auerbach Publications, 2020. 1" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.116, + 0.472, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.133 + ], + "angle": 0, + "content": "[16] Gao Huang, Zhuang Liu, and Kilian Q. Weinberger. Densely connected convolutional networks. CoRR, abs/1608.06993, 2016. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.136, + 0.892, + 0.19 + ], + "angle": 0, + "content": "[17] Hanxun Huang, Xingjun Ma, Sarah Monazam Erfani, James Bailey, and Yisen Wang. Unlearnable examples: Making personal data unexploitable. arXiv preprint arXiv:2101.04898, 2021. 1, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.193, + 0.892, + 0.246 + ], + "angle": 0, + "content": "[18] Arthur Jacot, Franck Gabriel, and Clément Hongler. Neural tangent kernel: Convergence and generalization in neural networks. Advances in neural information processing systems, 31, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.249, + 0.892, + 0.289 + ], + "angle": 0, + "content": "[19] Adel Javanmard and Mahdi Soltanolkotabi. Precise statistical analysis of classification accuracies for adversarial training. ArXiv, abs/2010.11213, 2020. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.292, + 0.892, + 0.345 + ], + "angle": 0, + "content": "[20] Pang Wei Koh and Percy Liang. Understanding black-box predictions via influence functions. In International conference on machine learning, pages 1885-1894. PMLR, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.348, + 0.892, + 0.375 + ], + "angle": 0, + "content": "[21] Alex Krizhevsky. Learning multiple layers of features from tiny images. Technical report, 2009. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.377, + 0.892, + 0.417 + ], + "angle": 0, + "content": "[22] Aounon Kumar, Alexander Levine, Soheil Feizi, and Tom Goldstein. Certifying confidence via randomized smoothing. CoRR, abs/2009.08061, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.42, + 0.892, + 0.473 + ], + "angle": 0, + "content": "[23] Alexey Kurakin, Ian J Goodfellow, and Samy Bengio. Adversarial examples in the physical world. In Artificial intelligence safety and security, pages 99-112. Chapman and Hall/CRC, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.476, + 0.892, + 0.544 + ], + "angle": 0, + "content": "[24] Cassidy Laidlaw and Soheil Feizi. Functional adversarial attacks. In H. Wallach, H. Larochelle, A. Beygelzimer, F. d'Alché-Buc, E. Fox, and R. Garnett, editors, Advances in Neural Information Processing Systems, volume 32. Curran Associates, Inc., 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.547, + 0.892, + 0.587 + ], + "angle": 0, + "content": "[25] Cassidy Laidlaw, Sahil Singla, and Soheil Feizi. Perceptual adversarial robustness: Defense against unseen threat models. CoRR, abs/2006.12655, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.589, + 0.892, + 0.616 + ], + "angle": 0, + "content": "[26] Ya Le and Xuan Yang. Tiny imagenet visual recognition challenge. CS 231N, 7(7):3, 2015. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.618, + 0.892, + 0.658 + ], + "angle": 0, + "content": "[27] Alexander Levine and Soheil Feizi. Deep partition aggregation: Provable defense against general poisoning attacks. CoRR, abs/2006.14768, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.661, + 0.892, + 0.701 + ], + "angle": 0, + "content": "[28] Alexander Levine and Soheil Feizi. Deep partition aggregation: Provable defense against general poisoning attacks. CoRR, abs/2006.14768, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.704, + 0.892, + 0.757 + ], + "angle": 0, + "content": "[29] Yige Li, Xixiang Lyu, Nodens Koren, Lingjuan Lyu, Bo Li, and Xingjun Ma. Neural attention distillation: Erasing backdoor triggers from deep neural networks. arXiv preprint arXiv:2101.05930, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.76, + 0.892, + 0.801 + ], + "angle": 0, + "content": "[30] Yiming Li, Tongqing Zhai, Baoyuan Wu, Yong Jiang, Zhifeng Li, and Shutao Xia. Rethinking the trigger of backdoor attack. arXiv preprint arXiv:2004.04692, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.804, + 0.892, + 0.871 + ], + "angle": 0, + "content": "[31] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In European conference on computer vision, pages 740-755. Springer, 2014. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.874, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[32] Sijia Liu, Songtao Lu, Xiangyi Chen, Yao Feng, Kaidi Xu, Abdullah Al-Dujaili, Mingyi Hong, and Una-May O'Reilly." + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "3870" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.092, + 0.47, + 0.147 + ], + "angle": 0, + "content": "Min-max optimization without gradients: Convergence and applications to black-box evasion and poisoning attacks. In International Conference on Machine Learning, pages 6282-6293. PMLR, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.149, + 0.47, + 0.204 + ], + "angle": 0, + "content": "[33] Yunfei Liu, Xingjun Ma, James Bailey, and Feng Lu. Reflection backdoor: A natural backdoor attack on deep neural networks. In European Conference on Computer Vision, pages 182-199. Springer, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.205, + 0.471, + 0.259 + ], + "angle": 0, + "content": "[34] Aleksander Madry, Aleksandar Makelov, Ludwig Schmidt, Dimitris Tsipras, and Adrian Vladu. Towards deep learning models resistant to adversarial attacks. arXiv preprint arXiv:1706.06083, 2017. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.261, + 0.471, + 0.315 + ], + "angle": 0, + "content": "[35] Yifei Min, Lin Chen, and Amin Karbasi. The curious case of adversarially robust models: More data can help, double descend, or hurt generalization. In Uncertainty in Artificial Intelligence, pages 129-139. PMLR, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.317, + 0.471, + 0.357 + ], + "angle": 0, + "content": "[36] Tuan Anh Nguyen and Anh Tran. Input-aware dynamic backdoor attack. Advances in Neural Information Processing Systems, 33:3454-3464, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.359, + 0.47, + 0.398 + ], + "angle": 0, + "content": "[37] Silvia Noschese, Lionello Pasquini, and Lothar Reichel. Tridiagonal toeplitz matrices: Properties and novel applications. 2006. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.401, + 0.471, + 0.482 + ], + "angle": 0, + "content": "[38] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.484, + 0.471, + 0.566 + ], + "angle": 0, + "content": "[39] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, Alexander C. Berg, and Li Fei-Fei. ImageNet Large Scale Visual Recognition Challenge. International Journal of Computer Vision (IJCV), 115(3):211-252, 2015. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.567, + 0.47, + 0.607 + ], + "angle": 0, + "content": "[40] M. Sandler, A. Howard, M. Zhu, A. Zhmoginov, and L.C. Chen. Mobilenetv2: Inverted residuals and linear bottle-necks. In CVPR, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.609, + 0.47, + 0.663 + ], + "angle": 0, + "content": "[41] Mert Bulent Sariyildiz, Julien Perez, and Diane Larlus. Learning visual representations with caption annotations. In European Conference on Computer Vision, pages 153-170. Springer, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.665, + 0.47, + 0.733 + ], + "angle": 0, + "content": "[42] Ali Shafahi, W Ronny Huang, Mahyar Najibi, Octavian Suciu, Christoph Studer, Tudor Dumitras, and Tom Goldstein. Poison frogs! targeted clean-label poisoning attacks on neural networks. Advances in neural information processing systems, 31, 2018. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.735, + 0.47, + 0.773 + ], + "angle": 0, + "content": "[43] Connor Shorten and Taghi Khoshgoftaar. A survey on image data augmentation for deep learning. Journal of Big Data, 6, 07 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.776, + 0.47, + 0.816 + ], + "angle": 0, + "content": "[44] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. CoRR, abs/1409.1556, 2015. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.818, + 0.47, + 0.871 + ], + "angle": 0, + "content": "[45] Christian Szegedy, Wojciech Zaremba, Ilya Sutskever, Joan Bruna, Dumitru Erhan, Ian Goodfellow, and Rob Fergus. Intriguing properties of neural networks. arXiv preprint arXiv:1312.6199, 2013. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.47, + 0.901 + ], + "angle": 0, + "content": "[46] M. Tan and Q. Le. Efficientnetv2: Smaller models and faster training. In ICML, 2021. 6" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.471, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.147 + ], + "angle": 0, + "content": "[47] Lue Tao, Lei Feng, Jinfeng Yi, Sheng-Jun Huang, and Songcan Chen. Better safe than sorry: Preventing delusive adversaries with adversarial training. Advances in Neural Information Processing Systems, 34, 2021. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.149, + 0.892, + 0.19 + ], + "angle": 0, + "content": "[48] H. Touvron, M. Cord, M. Douze, F. Massa, A. Sablayrolles, and H. Jegou. Training data-efficient image transformers & distillation through attention. In ICML, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.192, + 0.892, + 0.233 + ], + "angle": 0, + "content": "[49] Laurens van der Maaten and Geoffrey Hinton. Visualizing data using t-sne. Journal of Machine Learning Research, 9(86):2579-2605, 2008. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.235, + 0.892, + 0.288 + ], + "angle": 0, + "content": "[50] Wenxiao Wang, Alexander Levine, and Soheil Feizi. Improved certified defenses against data poisoning with (deterministic) finite aggregation. CoRR, abs/2202.02628, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.291, + 0.892, + 0.345 + ], + "angle": 0, + "content": "[51] Yisen Wang, Xingjun Ma, James Bailey, Jinfeng Yi, Bowen Zhou, and Quanquan Gu. On the convergence and robustness of adversarial training. arXiv preprint arXiv:2112.08304, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.348, + 0.894, + 0.417 + ], + "angle": 0, + "content": "[52] Da Yu, Huishuai Zhang, Wei Chen, Jian Yin, and Tie-Yan Liu. Availability attacks create shortcuts. In Proceedings of the 28th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, KDD '22, page 2367-2376, New York, NY, USA, 2022. Association for Computing Machinery. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.419, + 0.892, + 0.46 + ], + "angle": 0, + "content": "[53] Chia-Hung Yuan and Shan-Hung Wu. Neural tangent generalization attacks. In International Conference on Machine Learning, pages 12230-12240. PMLR, 2021. 1, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.462, + 0.892, + 0.531 + ], + "angle": 0, + "content": "[54] Sangdoo Yun, Dongyoon Han, Seong Joon Oh, Sanghyuk Chun, Junsuk Choe, and Youngjoon Yoo. Cutmix: Regularization strategy to train strong classifiers with localizable features. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6023-6032, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.532, + 0.892, + 0.6 + ], + "angle": 0, + "content": "[55] Sergey Zagoruyko and Nikos Komodakis. Wide residual networks. In Edwin R. Hancock Richard C. Wilson and William A. P. Smith, editors, Proceedings of the British Machine Vision Conference (BMVC), pages 87.1-87.12. BMVA Press, September 2016. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.602, + 0.892, + 0.643 + ], + "angle": 0, + "content": "[56] Hongyi Zhang, Moustapha Cisse, Yann N. Dauphin, and David Lopez-Paz. mixup: Beyond empirical risk minimization. CoRR, abs/1710.09412, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.645, + 0.892, + 0.713 + ], + "angle": 0, + "content": "[57] Hongyang Zhang, Yaodong Yu, Jiantao Jiao, Eric Xing, Laurent El Ghaoui, and Michael Jordan. Theoretically principled trade-off between robustness and accuracy. In International conference on machine learning, pages 7472-7482. PMLR, 2019. 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.713 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.956 + ], + "angle": 0, + "content": "3871" + } + ] +] \ No newline at end of file diff --git a/2023/CUDA_ Convolution-Based Unlearnable Datasets/e0179aca-27a9-460d-b895-5e7378133f26_origin.pdf b/2023/CUDA_ Convolution-Based Unlearnable Datasets/e0179aca-27a9-460d-b895-5e7378133f26_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..11e3252fec527ea48db17339a24025cb7d6b43ca --- /dev/null +++ b/2023/CUDA_ Convolution-Based Unlearnable Datasets/e0179aca-27a9-460d-b895-5e7378133f26_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:29f22cc138340614d7a17f69367bfd6b6d3535dc0aaad5b5197e275a47f47989 +size 4555504 diff --git a/2023/CUDA_ Convolution-Based Unlearnable Datasets/full.md b/2023/CUDA_ Convolution-Based Unlearnable Datasets/full.md new file mode 100644 index 0000000000000000000000000000000000000000..ab0cee27619d4235e66b335db859a5e416151c2e --- /dev/null +++ b/2023/CUDA_ Convolution-Based Unlearnable Datasets/full.md @@ -0,0 +1,290 @@ +# CUDA: Convolution-based Unlearnable Datasets + +Vinu Sankar Sadasivan +University of Maryland + +vinu@umd.edu + +Mahdi Soltanolkotabi +University of Southern California + +soltanol@usc.edu + +Soheil Feizi +University of Maryland + +sfeizi@cs.umd.edu + +# Abstract + +Large-scale training of modern deep learning models heavily relies on publicly available data on the web. This potentially unauthorized usage of online data leads to concerns regarding data privacy. Recent works aim to make unlearnable data for deep learning models by adding small, specially designed noises to tackle this issue. However, these methods are vulnerable to adversarial training (AT) and/or are computationally heavy. In this work, we propose a novel, model-free, Convolution-based Unlearnable DAtaset (CUDA) generation technique. CUDA is generated using controlled class-wise convolutions with filters that are randomly generated via a private key. CUDA encourages the network to learn the relation between filters and labels rather than informative features for classifying the clean data. We develop some theoretical analysis demonstrating that CUDA can successfully poison Gaussian mixture data by reducing the clean data performance of the optimal Bayes classifier. We also empirically demonstrate the effectiveness of CUDA with various datasets (CIFAR-10, CIFAR100, ImageNet-100, and Tiny-ImageNet), and architectures (ResNet-18, VGG-16, Wide ResNet-34-10, DenseNet-121, DeIT, EfficientNetV2-S, and MobileNetV2). Our experiments show that CUDA is robust to various data augmentations and training approaches such as smoothing, AT with different budgets, transfer learning, and fine-tuning. For instance, training a ResNet-18 on ImageNet-100 CUDA achieves only $8.96\%$ , $40.08\%$ , and $20.58\%$ clean test accuracies with empirical risk minimization (ERM), $L_{\infty}$ AT, and $L_{2}$ AT, respectively. Here, ERM on the clean training data achieves a clean test accuracy of $80.66\%$ . CUDA exhibits unlearnability effect with ERM even when only a fraction of the training dataset is perturbed. Furthermore, we also show that CUDA is robust to adaptive defenses designed specifically to break it. + +# 1. Introduction + +Modern deep learning training frameworks heavily depend on large-scale datasets for achieving high accuracy. + +This encourages deep learning practitioners to scrape data from the web for data collection [8, 31, 38, 41]. Since a lot of the data is publicly available online, sometime this scraping of data is unauthorized. For instance, a recent article [15] discloses that a private company trained a commercial face recognition system using over three billion facial images collected from the internet without any user consent. Although such massive data can significantly boost the performance of deep learning models, it raises serious concerns about data privacy and security. + +To prevent the unauthorized usage of personal data, a series of recent papers [10, 17, 53] propose to poison data with additive noise. The idea is to make datasets unlearnable for deep learning models by ensuring that they learn the correspondence between noises and labels. Thereby, they do not learn much useful information about the clean data, significantly degrading their clean test accuracy. However, in recent works [11, 17, 47], these unlearnability methods are shown to be vulnerable to adversarial training (AT) frameworks [34]. Motivated by this problem, Fu et al. [11] developed Robust Error-Minimization (REM) noises to make unlearnable data that is protected from AT. While the authors show the effectiveness of REM in multiple scenarios, we demonstrate that these methods are still not robust against different data augmentations or training settings (see Section 3.2). Furthermore, current unlearnability frameworks [10, 11, 17, 53] are model-dependent and require expensive optimization steps on deep learning models to obtain the additive noises. They also need to train the deep learning models from scratch to obtain noises for each new data set. + +In this paper, we propose a novel Convolution-based Unlearnable DAataset (CUDA) generation technique. We address limitations of existing unlearnable data generation techniques in Section 3.2 and motivate our CUDA technique in Section 3.3. For generating CUDA, an attacker randomly generates different convolutional filters for each class in the dataset using a private key or seed value. These filters are used to perform controlled class-wise convolutions on the clean training dataset to obtain CUDA. As we + +![](images/3c62eaf7fff8338381c2b75a38ca730daa8f02c74e3fd14983becd10a92a993c.jpg) +Figure 1. CIFAR-10 CUDA images from each of the 10 classes convolved using convolutional filters generated with blur parameter $p_b$ . As seen in the plot, a higher $p_b$ value results in better unlearnability (lower clean test accuracy), but increased blurring. CIFAR-10 CUDA does not break with ERM for all the three $p_b$ values. In Section 5.2, we propose Deconvolution-based Adversarial Training (DAT) that we specifically design to break CUDA. DAT adversarially learns class-wise filters to deconvolve CUDA images. However, CIFAR-10 CUDA with $p_b \in \{0.3, 0.5\}$ does not break with DAT. Therefore, we select $p_b = 0.3$ (lesser blurring) for CIFAR-10 CUDA as discussed in Section 5.1. More details on DAT in supplementary material. + +![](images/fe3c7da284d700e4944361e4bb0eedd4e99e2f82fea68330f5d6b146ab36bfa9.jpg) + +describe in Sections 3.3 and 5.2, CUDA generation performs controlled convolutions using a blur parameter $p_b$ to ensure that the semantics of the dataset are preserved (see Figure 1). CUDA generation with a lower blurring parameter $p_b$ adds less perceptible noises to clean samples. A network trained by a defender on CUDA is encouraged to learn the shortcut relation between class-wise convolutional filters and labels rather than useful features for classifying the clean data. Since the seed value for generating the filters are private, its not possible for the defender to obtain clean data from CUDA alone. Additionally, CUDA exhibits unlearnability effect with ERM even when only a fraction of the training dataset is perturbed (see Section 5). While the existing unlearnability works use additive noises, CUDA generation technique enjoys the advantage of introducing multiplicative noises in the Fourier domain due to the convolution theorem (since convolution of signals is the same as element-wise multiplication in the Fourier domain). This lets CUDA generation add higher amounts of noise in the image space, specifically along the edges in images, and makes it resilient to AT with small additive noise budgets. In Figure 2, with the help of t-SNE plots [49], we also find that the noises added by CUDA generation is not linearly separable while they are linearly separable for the existing works on unlearnability [52]. + +In Section 4, we theoretically show that CUDA generation can successfully poison Gaussian mixture data by degrading the clean data accuracy of the optimal Bayes classifier. We state our result informally below while the formal version is presented in Theorem 2. + +Theorem 1 (Informal) Let $\mathcal{D}$ denote a Gaussian mixture data with two modes, $P_{\mathcal{D}}$ denote the optimal Bayes classifier trained on $\mathcal{D}$ , and $\tau_{\mathcal{D}}(P_{\mathcal{D}})$ denote the accuracy of the classifier $P_{\mathcal{D}}$ on $\mathcal{D}$ . Then, under some assumptions, for every clean data $\mathcal{D}$ , there is a CUDA $\tilde{\mathcal{D}}$ such that + +$$ +\tau_ {\mathcal {D}} (P _ {\tilde {\mathcal {D}}}) < \tau_ {\mathcal {D}} (P _ {\mathcal {D}}). +$$ + +Furthermore, our empirical experiments in Section 5 demonstrate the effectiveness of CUDA under various training scenarios such as ERM with various augmentations and regularizations, AT with different budgets, randomized smoothing [6, 22, 27], transfer learning, and fine-tuning. For instance, training a ResNet-18 on CIFAR-10 CUDA achieves only $18.48\%$ , $44.4\%$ , and $51.14\%$ clean test accuracies with ERM, $L_{\infty}$ AT, and $L_{2}$ AT, respectively (see Figure 2). Here, ERM on the clean training data achieves a clean test accuracy of $94.66\%$ . In addition, we also design adaptive defenses to investigate if CUDA breaks with random or adversarial defense mechanisms. We find that CUDA is robust to the adaptive defenses that we specifically design to break it. + +# 2. Related Works + +Our CUDA generation technique is intimately related with adversarial and poisoning attacks. We first discuss some of this literature and then explain their relation with CUDA generation. + +Adversarial attacks. Adversarial examples are specially designed examples that can fool deep learning models at test time [4, 12, 23, 24, 45]. The adversary crafts these examples by adding error-maximizing noises to the clean data. Even slightly perturbed data can serve as adversarial examples. AT is a training framework proposed to make deep learning models robust to adversarial examples [19,25,34,51,57]. AT is a min-max optimization problem where the model is trained to minimize loss on adversarial examples that have the maximum loss. + +Poisoning attacks. In data poisoning, an attacker aims to hurt the deep learning model's performance by perturbing the training data [2, 20, 28, 32, 42, 50]. The backdoor attack + +![](images/0551ddd300a3b3d91e6650f1b8561e44da0383cc955cd589d1c107c0388fdfef.jpg) +Figure 2. In the left column, clean test accuracies of ResNet-18 on clean CIFAR-10 and CIFAR-10 CUDA are shown. The lower the test accuracy, the higher the effectiveness of unlearnability is. We show that CUDA is robust to various training settings such as ERM, $L_{2}$ AT, and $L_{\infty}$ AT. The error bars represent the standard deviation of test accuracy over 5 independent trials. CUDA is robust to the randomness in its data generation process. In the right column, t-SNEs of the noises generated by REM (top) and CUDA (bottom) for the first 3 classes of CIFAR-10 are plotted. + +![](images/a9c5a7c417aaecaa0f53bc946ffaaa6517a6c3243ee2335214b3699ec499c80f.jpg) + +is a special type of poisoning attack where a trigger pattern is injected into clean data at training time [5, 29, 33, 36]. The model trained on this data would misclassify an image with a trigger pattern at test time. Gu et al. [13] and Li et al. [30] use perceptible amounts of noises similar to CUDA for data poisoning. However, backdoor attacks do not affect the performance of the model on clean data [1, 5, 42]. + +Recent literature utilize data poisoning to protect data from being used for model training without authorization. Yuan and Wu [53] use neural tangent kernels [18] to generate clean label attacks that can hurt the generalization of deep learning models. Huang et al. [17] show that error-minimizing noise addition can serve as a poisoning technique. Fowl et al. [10] show that error-maximizing noises as well can make strong poison attacks. However, all these poisoning techniques do not offer data protection with AT [47, 52]. Fu et al. [11] proposes a min-min-max optimization technique to generate poisoned data that offers better unlearnability effects with AT. + +# 3. Convolution-based Ulearnable DAtaset (CUDA) + +In this section, first we give some preliminaries about unlearnability. Then we discuss the limitations of the existing unlearnability methods. Finally, we propose our CUDA generation technique. + +# 3.1. Preliminaries + +Let $\{(x_i, y_i)\}_{i=1}^n \sim \mathcal{D}^n$ be the clean training dataset where $\mathcal{D}$ is the clean data distribution, $x_i \in \mathcal{X} \subset \mathbb{R}^d$ are the samples, and $y_i \in \mathcal{Y}$ are the corresponding labels. Suppose a network is given as $f_\theta: \mathcal{X} \to \mathcal{Y}$ where $\theta \in \Theta$ is the network parameter and $\ell: \mathcal{Y} \times \mathcal{Y} \to \mathbb{R}$ is the loss function. ERM trains a network using a minimization prob + +lem of the form: $\min_{\theta} \frac{1}{n} \sum_{i=1}^{n} \ell(f_{\theta}(\pmb{x}_i), y_i)$ . Standard $L_p$ -based AT (for $p \in \mathbb{R}^+$ ) solves the following min-max problem: $\min_{\theta} \frac{1}{n} \sum_{i=1}^{n} \max_{\|\pmb{\delta}_i\|_p \leq \rho_a} \ell(f_{\theta}(\pmb{x}_i + \pmb{\delta}_i), y_i)$ where $\rho_a$ is the adversarial perturbation radius. + +We will use $\tau_{\mathcal{D}}(\theta)$ , with $\theta$ the clean model parameter, to denote the clean test accuracy of a model trained on the clean training dataset (i.e., clean model accuracy). In unlearnable dataset generation, an attacker uses an algorithm $\mathcal{A}:\mathcal{X}\rightarrow \mathcal{X}$ to generate an unlearnable dataset $\{(\tilde{\boldsymbol{x}}_i = \mathcal{A}(\boldsymbol {x}_i),y_i)\}_{i = 1}^n\sim \tilde{\mathcal{D}}^n$ from the clean training data. Here, the attacker assumes access to the full clean training dataset. Moreover, the attacker cannot modify the unlearnable dataset once it is released publicly. A defender trains using the unlearnable dataset to obtain a network $f_{\tilde{\theta}}$ . The objective of the attacker is to design an unlearnable dataset such that the defender's model trained on the unlearnable data achieves a clean test accuracy (i.e., unlearnable model accuracy) worse than the clean model accuracy i.e. $\tau_{\mathcal{D}}(\tilde{\theta})\ll \tau_{\mathcal{D}}(\theta)$ . + +# 3.2. Limitations of existing works + +Fu et al. [11] show that the previous unlearnability methods including Error-Minimization (EM) [17], Targeted Adversarial Poisoning (TAP) [10], and Neural Tangent Generalization Attack (NTGA) [53] are vulnerable to AT. Hence, they propose a Robust Error-Minimization (REM) [11] method that exploits a min-min-max optimization procedure to generate unlearnable noises. REM first trains a noise generator $f_{\theta}^{\prime}$ on data points $\{(x_i,y_i)\}_{i = 1}^n$ over a loss function $\ell$ as follows + +$$ +\begin{array}{l} \min _ {\theta} \frac {1}{n} \sum_ {i = 1} ^ {n} \min _ {\| \delta_ {i} ^ {u} \| \leq \rho_ {u}} \mathbb {E} _ {t \sim T} \\ \max _ {\| \boldsymbol {\delta} _ {i} ^ {a} \| \leq \rho_ {a}} \ell \left(f _ {\theta} ^ {\prime} \left(t \left(\boldsymbol {x} _ {i} + \boldsymbol {\delta} _ {i} ^ {u}\right) + \boldsymbol {\delta} _ {i} ^ {a}\right), y\right). \tag {1} \\ \end{array} +$$ + +Table 1. Time taken for generating unlearnable noises for various datasets using different unlearnability techniques. + +
DatasetEMTAPNTGAREMCUDA
CIFAR-100.4 hr0.5 hr5.2 hrs22.6 hrs10.8 s
CIFAR-1000.4 hr0.5 hr5.2 hrs22.6 hrs15.5 s
ImageNet-1003.9 hrs5.2 hrs14.6 hrs51.2 hrs0.15 hr
+ +Here, $T$ is a distribution over a set of transformations $\{t:$ $\mathcal{X}\to \mathcal{X}\}$ , $\rho_{u}$ is the defensive perturbation radius, and $\rho_{a}$ controls the protection level of REM against AT. After training the noise generator, an unlearnable example $(\tilde{\pmb{x}},y)$ is generated via + +$$ +\begin{array}{l} \tilde{\boldsymbol{x}} = \boldsymbol {x} + \operatorname *{arg min}_{\| \boldsymbol{\delta}^{u}\| \leq \rho_{u}}\mathbb{E}_{t\sim T} \\ \max _ {\| \boldsymbol {\delta} ^ {a} \| \leq \rho_ {a}} \ell \left(f _ {\theta} ^ {\prime} \left(t \left(\boldsymbol {x} + \boldsymbol {\delta} ^ {u}\right) + \boldsymbol {\delta} ^ {a}\right), y\right) \tag {2} \\ \end{array} +$$ + +First, note that REM is computationally expensive since it needs to generate unlearnability noises through solving optimization equation 2. Moreover, the existing techniques are model-dependent and they require gradient-based training with a neural network to generate unlearnable data. They also require neural network training from scratch for every dataset that is to be made unlearnable. Table 1 shows the amount of time required to generate various unlearnable datasets using NVIDIA® Tesla V100 GPU and 10 CPU cores. CUDA generation is significantly faster than the existing methods since it uses a model-free approach (no training required). Furthermore, REM is sensitive to hyperparameters and norm-budgets of AT since they generate noises with fixed $L_{\infty}$ norm budgets. For instance, a ResNet-18 trained on clean CIFAR-10 dataset achieves a clean test data accuracy of $94.66\%$ . $L_{\infty}$ AT with perturbation radius $\rho_{a} = 4 / 255$ on REM CIFAR-10 data (generated using $\rho_{u} = 8 / 255$ and $\rho_{a} = 4 / 255$ ) achieves only a clean test accuracy of $48.16\%$ . However, $L_{\infty}$ AT with perturbation radius $\rho_{a} = 8 / 255$ and $L_{2}$ AT with perturbation radius $\rho_{a} = 0.75$ can achieve a clean test accuracy of $78.71\%$ and $79.65\%$ , respectively, on the same REM data. We also find that ERM with a ResNet-18 on grayscaled REM CIFAR-10 images can achieve a high test accuracy of $70.76\%$ on the grayscaled CIFAR-10 test data. This shows that REM relies upon the color space for poisoning clean data. Fu et al. [11] also show that REM noise generated using ResNet-18 is not transferable to DenseNet-121. + +# 3.3. Our method: CUDA + +The major limitations of the previous works are that they are vulnerable to AT, and computationally expensive. We think the major reason for the former limitation is the usage of small additive noises for unlearnability. AT is designed to train in the presence of such additive noises. In + +creasing the budget of the amount of additive noises for unlearnability might destroy the semantics of the images while perturbing them. The latter limitation arises from the fact that these methods are model-dependent and they require multilevel optimizations. Hence, we are motivated to design a compute-efficient unlearnability method that is robust to AT. CUDA technique can perform convolutions to add larger amounts of noises to clean images without destroying its semantics. This can help CUDA to be robust against AT. CUDA uses randomly generated convolution filters for blurring images from each class. This makes a model trained on CUDA to learn shortcut relations between filters and labels. We empirically support these claims in Section 5.2. Additionally, randomly generating the filters makes our technique model-free. Since the keys for generating the filters are private, it is not possible to reverse the blurring effect in CUDA without having access to the corresponding clean images. Hence, we assume that the data publisher deletes the clean images after perturbing them. Moreover, CUDA technique is a novel non-additive noise based poisoning attack that needs to be studied. + +CUDA uses convolutional filters $s_i \in \mathbb{R}^{k \times k}$ for each class $i \in [1, K]$ . A random parameter, out of the $k^2$ parameters, in each of the filters is set to have a value of 1. The rest of the filter parameters are randomly initialized from a uniform distribution $\mathcal{U}(0, p_b)$ using a private seed where $p_b$ is the blur parameter. Blur parameter controls the level of blurring that occurs when an image $x \in [0, 1]^{d_1 \times d_2 \times d_3}$ is convolved with a filter $s_i$ . Here, $d_1, d_2$ , and $d_3$ are the height, width, and number of channels of the image, respectively. For example, a CIFAR-10 image has a dimension of $32 \times 32 \times 3$ . The higher $p_b$ is, the higher the blurring effect is. Let $\hat{x} = x * s_i$ where $x$ belongs to class $i$ . The CUDA data point for $x$ is given by $\hat{x} = \hat{x} / \mathrm{MAX}(\hat{x})$ . Rescaling is performed to make sure that the CUDA image pixels lie between 0 and 1. We find that the unlearnability effect gets stronger with larger $p_b$ and $k$ values (see supplementary material for details). + +# 4. Theory for CUDA + +In this section, we define a binary classification setup similar to [19, 35] to theoretically analyze CUDA. Let $\mathcal{D}$ be a clean dataset modelled by an isotropic Gaussian mixture model given by $\mathcal{N}(y\pmb{\mu}, I)$ , where $y \in \{\pm 1\}$ is the class label, $\pmb{\mu} \in \mathbb{R}^d$ , and $I \in \mathbb{R}^{d \times d}$ is the identity matrix. We defer the proofs for all the lemmas and theorem to the supplementary material. The Bayes optimal decision boundary for classifying this Gaussian mixture model is as follows. + +Lemma 1 The Bayes optimal decision boundary for classifying $\mathcal{D}$ is given by $P(\pmb{x}) \equiv \pmb{\mu}^{\top} \pmb{x} = 0$ . The accuracy of the decision boundary $P$ on the clean dataset $\mathcal{D}$ (i.e. $\tau_{\mathcal{D}}(P)$ ) is equal to $\phi(\|\pmb{\mu}\|_2)$ . Here, $\phi(.)$ represents the CDF of the + +![](images/c03e7e9239abc12ac49cd209af107cff469828f26ce520550c6089288a1847c4.jpg) +(a) $\| \pmb {\mu}\| _2 = 1.0$ + +![](images/f652421a0fd1624ff8771573c4ac561d5b841f9e5159fe0cbe40f0159d577972.jpg) +(b) $\| \pmb{\mu} \|_2 = 2.5$ +Figure 3. We plot the contour plots for clean test accuracy of the CUDA Bayes classifier $\tilde{P}$ (given as $\tau_{\mathcal{D}}(\tilde{P})$ ) for varying $a_{y}$ parameters where $y \in \{\pm 1\}$ in Figures 3(a-c). Gaussian mixture model $\mathcal{N}(y\mu, I)$ denotes the clean data. In the top row, we use Lemmas 1 & 4 to empirically generate the plots. In the bottom row, we plot the theoretical upper bound of $\tau_{\mathcal{D}}(\tilde{P})$ we obtain using Theorem 2. + +![](images/d105371da4d0456bc3918acd25f6d10a0c36cb54bfb91fe493cf2672ceb5fc0a.jpg) +(c) $\| \pmb{\mu}\| _2 = 5.0$ + +# standard normal distribution. + +Now, to obtain $\tilde{\mathcal{D}}$ (i.e., CUDA), we perform class-wise 1D convolutions. Here, we consider a class of 1D convolutional filters with kernel size 3 of the form $f_{a} = [a,1,a]$ where $a\in [0,0.5]$ . The convolution of signal $\pmb {x}\in \mathbb{R}^{d}$ with filter $\pmb{f}_{a}$ using stride 1 (denoted by $\pmb {x}*\pmb{f}_{a}$ ), can be treated as a matrix operation $Ax$ . Here, $A\in \mathbb{R}^{d\times d}$ is a tri-diagonal Toeplitz matrix denoted by $T(d;a,1,a)$ . For CUDA generation, we use class-wise convolution matrices $A_{y} = T(d;a_{y},1,a_{y})$ to perturb a clean data point $(x,y)\sim \mathcal{D}$ to an unlearnable data point $(A_yx,y)$ , where $a_{y}\in [0,0.5]$ for $y\in \{\pm 1\}$ . Note that in CUDA, the labels remain the same. Next we show that such a perturbed dataset (i.e. CUDA) can be represented as a Gaussian mixture model. + +Lemma 2 A CUDA generated from clean data distribution $\mathcal{D}$ and the $A_{y}$ 's defined above can be modelled as a Gaussian mixture model with the distribution given by $\tilde{\mathcal{D}} = \mathcal{N}(yA_y\pmb {\mu},A_y^2)$ . + +To characterize the decision boundary for CUDA, we need to use some properties of Toeplitz matrices from Noschese et al. [37] given in the following Lemma 3. + +Lemma 3 Any tri-diagonal Toeplitz matrix $A = T(d; a, 1, a) \in \mathbb{R}^{d \times d}$ with $a \in [0, 0.5]$ can be diagonalized as $A = QDQ$ where $Q_{i,j} = \left(\frac{2}{d + 1}\right)^{1/2} \sin \left(\frac{ij\pi}{d + 1}\right)$ and $D$ is a diagonal matrix with $D_{i,i} = 1 + 2a \cos \left(\frac{i\pi}{d + 1}\right)$ for $1 \leq i, j \leq d$ . + +Next we use Lemma 3 to show that the Bayes optimal decision boundary for classifying $\tilde{\mathcal{D}}$ is a quadratic plane. + +Lemma 4 Let $A_{-1} = T(d; a_{-1}, 1, a_{-1})$ and $A_1 = T(d; a_1, 1, a_1)$ . The Bayes optimal decision boundary for + +classifying $\tilde{\mathcal{D}}$ is given by $\tilde{P} (\pmb {x})\equiv \pmb{x}^{\top}A\pmb {x} + \pmb{b}^{\top}\pmb {x} + c = 0,$ where $A = A_{-1}^{-2} - A_1^{-2},\pmb {b} = 2(A_{-1}^{-1} + A_{1}^{-1})\pmb {\mu},$ and $c = 2\sum_{i = 1}^{d}\left[\ln (1 + 2a_{-1}\cos (\frac{i\pi}{d + 1})) - \ln (1 + 2a_{1}\cos (\frac{i\pi}{d + 1}))\right].$ + +Now we state a lemma regarding the tail of Gaussian quadratic forms which plays a crucial role in our main result. + +Lemma 5 Let $\| \cdot \|$ denote the operator norm, $\| \cdot \|_2$ denote the vector 2-norm, $z \sim \mathcal{N}(\mathbf{0}, I)$ , and $Z = z^\top A z + z^\top b + c$ where $A = Q \Lambda Q^\top$ . Using Chernoff bound, for any $t \geq 0$ and $\gamma \in \mathbb{R}$ + +$$ +\begin{array}{l} \mathbb {P} \{Z - \mathbb {E} Z \geq \gamma \} \leq \\ \frac {\exp \left\{\frac {- t}{4 \| \boldsymbol {b} \| _ {2} ^ {\frac {2}{2}} \| \boldsymbol {\Lambda} \|} - t (\gamma + \operatorname {T r} (\boldsymbol {\Lambda}) + \| \boldsymbol {b} \| _ {2}) \right\}}{| I - 2 t \boldsymbol {\Lambda} | ^ {\frac {1}{2}}}. \\ \end{array} +$$ + +Lemma 5 allows us to provide an upper bound for the accuracy of the unlearnable decision boundary $\tilde{P}$ on the clean dataset $\mathcal{D}$ , given as $\tau_{\mathcal{D}}(\tilde{P})$ , in Theorem 2 below. + +Theorem 2 (Main result) Let $A = Q\Lambda Q = A_{-1}^{-2} - A_{1}^{-2}$ . For any non-negative constants $t_1$ and $t_2$ , the accuracy of the unlearnable decision boundary $\tilde{P}$ on the clean dataset $\mathcal{D}$ can be upper-bounded as + +$$ +\begin{array}{l} \tau_ {\mathcal {D}} (\tilde {P}) \leq \frac {\exp \left[ t _ {1} \left(\boldsymbol {b} ^ {\top} \boldsymbol {\mu} + \boldsymbol {\mu} ^ {\top} A \boldsymbol {\mu} + c - \frac {1}{4 \| 2 A \boldsymbol {\mu} + \boldsymbol {b} \| _ {2} \| \Lambda \|}\right) \right]}{2 | I - 2 t _ {1} \Lambda | ^ {\frac {1}{2}}} \\ + \frac {\exp \left[ t _ {2} \left(\boldsymbol {b} ^ {\top} \boldsymbol {\mu} - \boldsymbol {\mu} ^ {\top} A \boldsymbol {\mu} - c - \frac {1}{4 \| 2 A \boldsymbol {\mu} - \boldsymbol {b} \| _ {2} \| \Lambda \|}\right) \right]}{2 | I - 2 t _ {2} \Lambda | ^ {\frac {1}{2}}} \\ := p _ {1} + p _ {2}. \\ \end{array} +$$ + +Furthermore, if $\pmb{\mu}^{\top}A\pmb{\mu} + \pmb{b}^{\top}\pmb{\mu} + c + \mathrm{Tr}(A) + \| 2A\pmb{\mu} + \pmb{b}\| < 0$ we have $\tau_{\mathcal{D}}(\tilde{P})\leq \frac{1}{2} (p_1 + 1) < 1$ . Also, if $-\pmb{\mu}^{\top}A\pmb{\mu} + \pmb{b}^{\top}\pmb{\mu}-$ + +$c - \operatorname {Tr}(A) + \| 2A\pmb {\mu} - \pmb {b}\| < 0$ we have $\tau_{\mathcal{D}}(\tilde{P})\leq \frac{1}{2} (1 + p_2) < 1.$ Moreover, for any $\pmb {\mu}\neq \mathbf{0}$ and $a_{-1}\in [0,0.5]$ , $\exists a_{1}$ such that $\tau_{\mathcal{D}}(\tilde{P}) < \tau_{\mathcal{D}}(P)$ + +Poisoning is effective only if the accuracy of the unlearnable model $\tilde{P}$ is less than that of the clean model $P$ on the clean dataset $\mathcal{D}$ , that is, $\tau_{\mathcal{D}}(\tilde{P}) < \tau_{\mathcal{D}}(P)$ . To satisfy this condition, we need to carefully select $a_{y}$ 's. In Theorem 2, we formally state this condition1. Theorem 2 shows that CUDA can effectively poison when there are two distinct modes in the clean Gaussian mixture data model. We validate our theoretical claim through empirical analysis as well (see Figure 3). Details for the analysis is given in supplementary material. We find that our upper bound for the clean test accuracy of CUDA classifiers are consistent with our empirical analysis. In our experiments, we also find that the unlearnability effect is stronger with a larger blur parameter. This effect is evident from Figure 3 where we find that it is likely to get a lower $\tau_{\mathcal{D}}(\tilde{P})$ with higher $a_{y}$ values. These results are consistent with our experimental results in Section 5 with CIFAR-10, CIFAR-100, and ImageNet-100 datasets. + +# 5. Experiments + +In this section, we first discuss our experimental setup. More details on the setup is deferred to supplementary material. We then show the robustness of CUDA generation with various datasets and architectures. We also run various experiments to analyze the effectiveness of CUDA under different training techniques (ERM, AT with varying budgets, randomized smoothing, transfer learning, and finetuning) and augmentation techniques (mixup [56], grayscale, random blurring, cutout [9], cutmix [54], autoaugment [7], and orthogonal regularization [3]). Finally, we also design adaptive defenses to test the robustness of CUDA. One might think that CUDA filters can be obtained by adversarially training them with the data. We show that CUDA is robust to such adaptive defenses that we design. + +# 5.1. Experimental setup + +Datasets. We use three image classification datasets - CIFAR-10, CIFAR-100, [21] and ImageNet-100 (a subset of ImageNet made of the first 100 classes) [39]. We use the data augmentation techniques such as random flipping, cropping, and resizing [43]. + +Architectures. We use ResNet-18 [14], VGG-16 [44], Wide ResNet-34-10 [55], and DenseNet-121 [16]. We train the networks with hyperparameters used in Fu et al. [11]. + +1We note that the conditions $\pmb{\mu}^{\top}A\pmb{\mu} + \pmb{b}^{\top}\pmb{\mu} + c + \mathrm{Tr}(A) + \|2A\pmb{\mu} + \pmb{b}\|_{2} < 0$ or $-\pmb{\mu}^{\top}A\pmb{\mu} + \pmb{b}^{\top}\pmb{\mu} - c - \mathrm{Tr}(A) + \|2A\pmb{\mu} - \pmb{b}\|_{2} < 0$ can always be satisfied by picking a sufficiently large $\pmb{\mu}$ in the direction of an eigenvector corresponding to a negative or positive eigenvalue of $A$ (note that $A$ has negative and positive eigenvalues). + +Table 2. Test accuracy (\%) of ResNet-18 trained on various unlearnable datasets. We use $L_{\infty}$ AT with budget $\rho_{a} = 4 / 255$ . + +
DatasetTraining methodCleanUnlearnability method
EMTAPNTGAREMCUDA
CIFAR-10ERM94.6613.2022.5116.2727.0918.48
AT89.5188.6288.0288.9648.1644.40
CIFAR-100ERM76.271.6013.753.2210.1412.69
AT64.5063.4362.3962.4427.1034.34
ImageNet-100ERM80.661.269.108.4213.748.96
AT66.6263.4063.5663.0641.6638.68
+ +Table 3. Test accuracy $(\%)$ of network architectures trained on various CIFAR-10 unlearnable datasets with $L_{\infty}$ AT $(\rho_{a} = 4 / 255)$ . + +
ModelCleanUnlearnability method
EMTAPNTGAREMCUDA
ResNet-1889.5188.6288.0288.9648.1644.40
VGG-1687.5186.4886.2786.6565.2342.98
Wide ResNet-34-1091.2190.0590.2389.9548.3953.02
DenseNet-12183.2782.4481.7280.7381.4845.95
+ +Previous works mainly employ ResNet-18 for most of their evaluations. Additional experiments on Tiny-ImageNet [26], DeIT [48], EfficientNetV2-S [46], and MobileNetV2 [40] are provided in the supplementary material. + +CUDA generation. We use filters of size $k = 3$ and blur parameter $p_b = 0.3$ for both CIFAR-10 and CIFAR-100 datasets. ImageNet-100 is a higher dimensional $224 \times 224 \times 3$ image dataset when compared to the $32 \times 32 \times 3$ dimensional CIFAR datasets. Hence, we use larger filters of size $k = 9$ with $p_b = 0.06$ for ImageNet-100. These hyperparameters are chosen such that the CUDA images are not perceptibly highly perturbed and give good unlearnability effect (see plot in Figure 1). In supplementary material, we show the results of training on CUDA with different hyperparameters for data generation. + +Baselines. We compare CUDA generation technique with four state-of-the-art unlearnability methods - REM [11], EM [17], TAP [10], and NTGA [53]. We adopt the results reported in [11] since we use the same hyperparameters for training. For REM, we select hyperparameters $\rho_{u} = 8 / 255$ and $\rho_{a} = 4 / 255$ , the highest radii values in [11]. For comparing unlearnable methods, we look at the clean test accuracy. The lower the test accuracy, the better the unlearnability method is. As mentioned in the supplementary material, we use publicly released official codebases for reproducing the baselines using their default hyperparameters. + +# 5.2. Effectiveness of CUDA + +Why does CUDA work? In order to measure how much CUDA technique's blurring affects the dataset's quality, we + +Table 4. Test accuracy $(\%)$ of ResNet-18 with CUDA under various training settings. + +
DatasetCleanTraining methodCUDA (ours)
CIFAR-1094.66ERM18.48
AT L∞ (ρa = 4/255)44.40
AT L∞ (ρa = 8/255)32.85
AT L∞ (ρa = 16/255)19.32
AT L2 (ρa = 0.25)39.05
AT L2 (ρa = 0.50)51.19
AT L2 (ρa = 0.75)51.14
CIFAR-10076.27ERM12.69
AT L∞ (ρa = 4/255)34.34
AT L∞ (ρa = 8/255)30.00
AT L2 (ρa = 0.75)36.90
ImageNet-10080.66ERM8.96
AT L∞ (ρa = 4/255)38.68
AT L∞ (ρa = 8/255)40.08
AT L2 (ρa = 0.75)20.58
+ +compare class-wise blurring (CUDA technique) against universal blurring where a single convolutional filter is used for blurring all the images. We keep the filter generation parameters fixed ( $p_b = 0.3$ and $k = 3$ ) for both CUDA class-wise blurring and universal blurring. ResNet-18 trained with clean CIFAR-10, universally blurred CIFAR-10, and CIFAR-10 CUDA achieve clean test accuracies of $94.66\%$ , $90.47\%$ , and $18.48\%$ , respectively. This suggests that our controlled blurring does not obscure the semantics of the dataset. Hence, the significant drop in the clean test accuracy introduced by CUDA is most likely due to the usage of class-wise filters. This suggests that a model trained on CUDA learns the relation between the class-wise convolution filters and their corresponding labels. Therefore, during test time when this convolution effect is absent, the CUDA trained model fails to make correct classifications. Furthermore, the model trained on CUDA achieves an accuracy of $99.91\%$ on the CIFAR-10 CUDA testset. This strongly supports our claim that the CUDA model learns to classify images based on the convolutional filters used to blur them. In addition, if we permute the class-wise filters for blurring the test set (i.e., blurring class 1 images with class 2 filters, class 2 images with class 3 filters, and so on), we get a very low accuracy of $2.53\%$ on this test set. Further details are provided in the supplementary material. Finally, we note that real-world datasets might also contain blurred images due to various factors such as motion blurring, weather conditions, issues with the camera, etc. Hence, detecting if a blurred image is poisoned might not always be possible. However, one might argue that it is possible to detect if an entire dataset is blurred. Interestingly, later in this section we show that CUDA technique exhibits unlearnability effect even when only a fraction of the training dataset is poisoned. + +Different datasets. We first compare the effectiveness of CUDA with different datasets using ERM and $L_{\infty}$ AT with + +$\rho_{a} = 4 / 255$ . We use ResNet-18 for the experiments. The results are shown in Table 2. These results show that EM, TAP, and NTGA are not robust to AT. However, both CUDA and REM are successful. Here, our method CUDA outperforms REM with CIFAR-10 and ImageNet-100 datasets. Smartly designed additive noise in AT helps in achieving better generalization than ERM on the unlearnable datasets. This experiment thus demonstrates that ERM and AT are not good choices for training with CUDA and REM dataset. + +Different models. Next we compare the effectiveness of CUDA using various deep learning architectures with $L_{\infty}$ AT ( $\rho_{a} = 4 / 255$ ). We use CIFAR-10 for the experiments. The results are shown in Table 3. As we see in the table, CUDA is effective with all the five network architectures. However, REM is not seen to be transferrable with DenseNet-121. + +Robustness to different training settings. In Section 3.2, we show that REM is sensitive to the training settings. REM generated using $L_{\infty}$ radii budgets of $\rho_u = 8 / 255$ and $\rho_a = 4 / 255$ for CIFAR-10 breaks with $L_{\infty}$ AT $(\rho_a = 8 / 255)$ and $L_2$ AT $(\rho_a = 0.75)$ to get test accuracy of $78.71\%$ and $79.65\%$ , respectively. Hence, we run experiments to check the robustness of CUDA with various AT norm budgets. The results are shown in Table 4. As we see in the table, CUDA is robust to ERM, $L_{\infty}$ , and $L_2$ AT settings with varying training budgets. Impressively, the highest test accuracy achieved with training on CIFAR-10 CUDA, CIFAR-100 CUDA, and ImageNet-100 CUDA are as low as $51.19\%$ , $36.90\%$ , and $40.08\%$ , respectively. We also find that using a pre-trained ResNet-18 with CIFAR-10 CUDA only achieves clean test accuracy of $42.42\%$ and $48.22\%$ with fine-tuning the full network and a newly trained final layer, respectively (details are deferred to the supplementary material). + +Different protection percentages. In Section 3.1, we assume that the attacker has access to the full clean training data. However, in real life settings, this might not be always possible. Hence, we train ResNet-18 on a mix of CIFAR-10 CUDA and clean CIFAR-10 training datasets to evaluate the effectiveness of poisoning with varying data protection percentages. Protection percentage denotes the percentage of the training data that is poisoned. + +We show the results in Table 5. In the table, the "Mixed" column denotes the clean test accuracy of a model trained using a mix of both clean and poisoned data. The "Clean" column denotes the clean test accuracy of a model trained only using the clean subset of the training data. In the last row of Table 5, we provide the results for CUDA trained using ERM with varying data protection percentages. The remainder of the rows provide the results for the $L_{\infty}$ AT $(\rho_{a} = 4 / 255)$ scenario. For example, CUDA trained with ERM using an $80\%$ clean training data partition achieves a test accuracy of $93.75\%$ . Adding the $20\%$ CUDA data + +Table 5. Test accuracy (\%) of ResNet-18 on CIFAR-10 with different data protection percentages. The last row shows the results for CUDA with ERM setting. The rest of the rows show results for unlearnability methods trained in the $L_{\infty}$ AT ( $\rho_a = 4/255$ ) setting. + +
Unlearnability methodData Protection Percentage
0%20% MixedClean40% MixedClean60% MixedClean80% MixedClean100%
EM89.5189.6088.1789.4086.7689.4985.0789.1079.4188.62
TAP89.0188.6688.4088.0488.02
NTGA89.5689.3589.2289.1788.96
REM89.6089.3489.6188.0948.16
CUDA (ours)88.5487.2486.0384.3444.40
CUDA + ERM94.6693.2893.7591.3492.5689.9189.7785.6184.3018.48
+ +partition to the training dataset drops the test accuracy of the model to $93.28\%$ . Results from the last row of Table 5 show that CUDA with varying data protection percentages is effective with the ERM setting. However, with the AT scenario, the unlearnability techniques are not as successful as with ERM with varying data protection percentages. Nevertheless, it is interesting to note that CUDA technique performs better than all the other unlearnability techniques with AT. + +Robustness to smoothing and data augmentations. Cohen et al. [6] proposes randomized smoothing which is a provable adversarial defense in $L_{2}$ norm. Since CUDA does not use additive noise, CUDA is robust to randomized smoothing. A smoothed ResNet-18 (with a noise level of 0.5) with CIFAR-10 CUDA achieves only a clean test accuracy of $43.85\%$ . In Section 3.2, we see that REM breaks with grayscaleing. While ResNet-18 training using grayscaleed REM achieves $70.76\%$ test accuracy, grayscaleed CUDA only achieves $20.12\%$ test accuracy on the clean grayscaleed CIFAR-10 test data. This shows that CUDA technique exhibits the desirable property of not relying upon the color space for its attack. CIFAR-10 CUDA training achieves only $25.53\%$ , $25.80\%$ , $26.93\%$ , $34.09\%$ , and $50.72\%$ clean test accuracy with mixup, cutout, cutmix, autoaugment, and orthogonal regularization, respectively (see supplementary material for more details). + +Adaptive defenses for CUDA. Here, we first investigate the effect of training CIFAR-10 CUDA with random blurring augmentations using ResNet-18. Each batch of the CUDA training data is convolved with random $3 \times 3$ filters of varying blur parameters $p_b'$ . With $p_b'$ values of 0.1 and 0.3, CUDA training achieves lower test accuracy $9.2\%$ and $13.37\%$ , respectively. This shows that the noise added by CUDA technique is robust to random convolution augmentations. + +One may think that CUDA technique can be broken by learning the private filters from the data. We test this idea by training deconvolution filters to find if we can reverse the blurring effect in CUDA with adversarially trained filters. We use a novel Deconvolution-based Adversarial Training + +(DAT) technique that is similar to AT (check supplementary material for details). While the adversarial step in AT learns sample-wise error-maximizing additive noises, the adversarial step in DAT learns class-wise error-maximizing deconvolution filters. We train DAT with filters of varying sizes (3, 5, and 7) on CIFAR-10 CUDA using ResNet-18. The filter parameters are constrained within a finite range to make sure that the images do not get distorted with the adversarial step similar to projection in projected gradient descent. We find that CUDA is robust to DAT. DAT using filters of size 3, 5, and 7 with CUDA achieves only test accuracy of $39.05\%$ , $46.21\%$ , and $38.48\%$ , respectively. DAT is not successful against CUDA since we can not invert convolutions without the knowledge of the private filters or clean images corresponding to CUDA. + +Limitations and future directions. The unlearnability effect of CUDA can be defended if some fraction of the clean data and its corresponding CUDA images are leaked. The defender must also be able to detect if all samples in the dataset are poisoned. However, our work assumes a setup where the filters remain private. For example, a data publisher could simply publish their CUDA images and delete the clean images permanently to prevent this scenario. As discussed in this section, CUDA as well as other prior works do not perform well with different protection percentages with AT. Improving this can be an interesting research direction. We believe that extending CUDA technique to other domains such as tabular and text data is also an interesting future direction. It would also be interesting to see theoretical analysis of CUDA considering more complex setups. + +# Acknowledgement + +This project was supported in part by Meta grant 23010098, NSF CAREER AWARD 1942230, HR001119S0026 (GARD), ONR YIP award N00014-22-1-2271, Army Grant No. W911NF2120076, a capital one grant, NIST 60NANB20D134 and the NSF award CCF2212458. + +# References + +[1] Mauro Barni, Kassem Kallas, and Benedetta Tondi. A new backdoor attack in cnns by training set corruption without label poisoning. In 2019 IEEE International Conference on Image Processing (ICIP), pages 101-105. IEEE, 2019. 3 +[2] Battista Biggio, Blaine Nelson, and Pavel Laskov. Poisoning attacks against support vector machines. arXiv preprint arXiv:1206.6389, 2012. 2 +[3] Andrew Brock, Theodore Lim, James M Ritchie, and Nick Weston. Neural photo editing with introspective adversarial networks. arXiv preprint arXiv:1609.07093, 2016. 6 +[4] Nicholas Carlini and David Wagner. Towards evaluating the robustness of neural networks. In 2017 IEEE symposium on security and privacy (sp), pages 39-57. IEEE, 2017. 2 +[5] Xinyun Chen, Chang Liu, Bo Li, Kimberly Lu, and Dawn Song. Targeted backdoor attacks on deep learning systems using data poisoning. arXiv preprint arXiv:1712.05526, 2017.3 +[6] Jeremy Cohen, Elan Rosenfeld, and Zico Kolter. Certified adversarial robustness via randomized smoothing. In Kamalika Chaudhuri and Ruslan Salakhutdinov, editors, Proceedings of the 36th International Conference on Machine Learning, volume 97 of Proceedings of Machine Learning Research, pages 1310-1320. PMLR, 09-15 Jun 2019. 2, 8 +[7] Ekin D Cubuk, Barret Zoph, Dandelion Mane, Vijay Vasudevan, and Quoc V Le. Autoaugment: Learning augmentation policies from data. arXiv preprint arXiv:1805.09501, 2018. 6 +[8] Karan Desai and Justin Johnson. Virtex: Learning visual representations from textual annotations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11162-11173, 2021. 1 +[9] Terrance DeVries and Graham W Taylor. Improved regularization of convolutional neural networks with cutout. arXiv preprint arXiv:1708.04552, 2017. 6 +[10] Liam Fowl, Micah Goldblum, Ping-yeh Chiang, Jonas Geiping, Wojtek Czaja, and Tom Goldstein. Adversarial examples make strong poisons. arXiv preprint arXiv:2106.10807, 2021. 1, 3, 6 +[11] Shaopeng Fu, Fengxiang He, Yang Liu, Li Shen, and Dacheng Tao. Robust unlearnable examples: Protecting data privacy against adversarial learning. In International Conference on Learning Representations, 2021. 1, 3, 4, 6 +[12] Ian J Goodfellow, Jonathon Shlens, and Christian Szegedy. Explaining and harnessing adversarial examples. arXiv preprint arXiv:1412.6572, 2014. 2 +[13] Tianyu Gu, Brendan Dolan-Gavitt, and Siddharth Garg. Badnets: Identifying vulnerabilities in the machine learning model supply chain. arXiv preprint arXiv:1708.06733, 2017. 3 +[14] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. CoRR, abs/1512.03385, 2015. 6 +[15] Kashmir Hill. The secretive company that might end privacy as we know it. In *Ethics of Data and Analytics*, pages 170–177. Auerbach Publications, 2020. 1 + +[16] Gao Huang, Zhuang Liu, and Kilian Q. Weinberger. Densely connected convolutional networks. CoRR, abs/1608.06993, 2016. 6 +[17] Hanxun Huang, Xingjun Ma, Sarah Monazam Erfani, James Bailey, and Yisen Wang. Unlearnable examples: Making personal data unexploitable. arXiv preprint arXiv:2101.04898, 2021. 1, 3, 6 +[18] Arthur Jacot, Franck Gabriel, and Clément Hongler. Neural tangent kernel: Convergence and generalization in neural networks. Advances in neural information processing systems, 31, 2018. 3 +[19] Adel Javanmard and Mahdi Soltanolkotabi. Precise statistical analysis of classification accuracies for adversarial training. ArXiv, abs/2010.11213, 2020. 2, 4 +[20] Pang Wei Koh and Percy Liang. Understanding black-box predictions via influence functions. In International conference on machine learning, pages 1885-1894. PMLR, 2017. 2 +[21] Alex Krizhevsky. Learning multiple layers of features from tiny images. Technical report, 2009. 6 +[22] Aounon Kumar, Alexander Levine, Soheil Feizi, and Tom Goldstein. Certifying confidence via randomized smoothing. CoRR, abs/2009.08061, 2020. 2 +[23] Alexey Kurakin, Ian J Goodfellow, and Samy Bengio. Adversarial examples in the physical world. In Artificial intelligence safety and security, pages 99-112. Chapman and Hall/CRC, 2018. 2 +[24] Cassidy Laidlaw and Soheil Feizi. Functional adversarial attacks. In H. Wallach, H. Larochelle, A. Beygelzimer, F. d'Alché-Buc, E. Fox, and R. Garnett, editors, Advances in Neural Information Processing Systems, volume 32. Curran Associates, Inc., 2019. 2 +[25] Cassidy Laidlaw, Sahil Singla, and Soheil Feizi. Perceptual adversarial robustness: Defense against unseen threat models. CoRR, abs/2006.12655, 2020. 2 +[26] Ya Le and Xuan Yang. Tiny imagenet visual recognition challenge. CS 231N, 7(7):3, 2015. 6 +[27] Alexander Levine and Soheil Feizi. Deep partition aggregation: Provable defense against general poisoning attacks. CoRR, abs/2006.14768, 2020. 2 +[28] Alexander Levine and Soheil Feizi. Deep partition aggregation: Provable defense against general poisoning attacks. CoRR, abs/2006.14768, 2020. 2 +[29] Yige Li, Xixiang Lyu, Nodens Koren, Lingjuan Lyu, Bo Li, and Xingjun Ma. Neural attention distillation: Erasing backdoor triggers from deep neural networks. arXiv preprint arXiv:2101.05930, 2021. 3 +[30] Yiming Li, Tongqing Zhai, Baoyuan Wu, Yong Jiang, Zhifeng Li, and Shutao Xia. Rethinking the trigger of backdoor attack. arXiv preprint arXiv:2004.04692, 2020. 3 +[31] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In European conference on computer vision, pages 740-755. Springer, 2014. 1 +[32] Sijia Liu, Songtao Lu, Xiangyi Chen, Yao Feng, Kaidi Xu, Abdullah Al-Dujaili, Mingyi Hong, and Una-May O'Reilly. + +Min-max optimization without gradients: Convergence and applications to black-box evasion and poisoning attacks. In International Conference on Machine Learning, pages 6282-6293. PMLR, 2020. 2 +[33] Yunfei Liu, Xingjun Ma, James Bailey, and Feng Lu. Reflection backdoor: A natural backdoor attack on deep neural networks. In European Conference on Computer Vision, pages 182-199. Springer, 2020. 3 +[34] Aleksander Madry, Aleksandar Makelov, Ludwig Schmidt, Dimitris Tsipras, and Adrian Vladu. Towards deep learning models resistant to adversarial attacks. arXiv preprint arXiv:1706.06083, 2017. 1, 2 +[35] Yifei Min, Lin Chen, and Amin Karbasi. The curious case of adversarially robust models: More data can help, double descend, or hurt generalization. In Uncertainty in Artificial Intelligence, pages 129-139. PMLR, 2021. 4 +[36] Tuan Anh Nguyen and Anh Tran. Input-aware dynamic backdoor attack. Advances in Neural Information Processing Systems, 33:3454-3464, 2020. 3 +[37] Silvia Noschese, Lionello Pasquini, and Lothar Reichel. Tridiagonal toeplitz matrices: Properties and novel applications. 2006. 5 +[38] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 1 +[39] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, Alexander C. Berg, and Li Fei-Fei. ImageNet Large Scale Visual Recognition Challenge. International Journal of Computer Vision (IJCV), 115(3):211-252, 2015. 6 +[40] M. Sandler, A. Howard, M. Zhu, A. Zhmoginov, and L.C. Chen. Mobilenetv2: Inverted residuals and linear bottle-necks. In CVPR, 2018. 6 +[41] Mert Bulent Sariyildiz, Julien Perez, and Diane Larlus. Learning visual representations with caption annotations. In European Conference on Computer Vision, pages 153-170. Springer, 2020. 1 +[42] Ali Shafahi, W Ronny Huang, Mahyar Najibi, Octavian Suciu, Christoph Studer, Tudor Dumitras, and Tom Goldstein. Poison frogs! targeted clean-label poisoning attacks on neural networks. Advances in neural information processing systems, 31, 2018. 2, 3 +[43] Connor Shorten and Taghi Khoshgoftaar. A survey on image data augmentation for deep learning. Journal of Big Data, 6, 07 2019. 6 +[44] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. CoRR, abs/1409.1556, 2015. 6 +[45] Christian Szegedy, Wojciech Zaremba, Ilya Sutskever, Joan Bruna, Dumitru Erhan, Ian Goodfellow, and Rob Fergus. Intriguing properties of neural networks. arXiv preprint arXiv:1312.6199, 2013. 2 +[46] M. Tan and Q. Le. Efficientnetv2: Smaller models and faster training. In ICML, 2021. 6 + +[47] Lue Tao, Lei Feng, Jinfeng Yi, Sheng-Jun Huang, and Songcan Chen. Better safe than sorry: Preventing delusive adversaries with adversarial training. Advances in Neural Information Processing Systems, 34, 2021. 1, 3 +[48] H. Touvron, M. Cord, M. Douze, F. Massa, A. Sablayrolles, and H. Jegou. Training data-efficient image transformers & distillation through attention. In ICML, 2021. 6 +[49] Laurens van der Maaten and Geoffrey Hinton. Visualizing data using t-sne. Journal of Machine Learning Research, 9(86):2579-2605, 2008. 2 +[50] Wenxiao Wang, Alexander Levine, and Soheil Feizi. Improved certified defenses against data poisoning with (deterministic) finite aggregation. CoRR, abs/2202.02628, 2022. 2 +[51] Yisen Wang, Xingjun Ma, James Bailey, Jinfeng Yi, Bowen Zhou, and Quanquan Gu. On the convergence and robustness of adversarial training. arXiv preprint arXiv:2112.08304, 2021. 2 +[52] Da Yu, Huishuai Zhang, Wei Chen, Jian Yin, and Tie-Yan Liu. Availability attacks create shortcuts. In Proceedings of the 28th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, KDD '22, page 2367-2376, New York, NY, USA, 2022. Association for Computing Machinery. 2, 3 +[53] Chia-Hung Yuan and Shan-Hung Wu. Neural tangent generalization attacks. In International Conference on Machine Learning, pages 12230-12240. PMLR, 2021. 1, 3, 6 +[54] Sangdoo Yun, Dongyoon Han, Seong Joon Oh, Sanghyuk Chun, Junsuk Choe, and Youngjoon Yoo. Cutmix: Regularization strategy to train strong classifiers with localizable features. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6023-6032, 2019. 6 +[55] Sergey Zagoruyko and Nikos Komodakis. Wide residual networks. In Edwin R. Hancock Richard C. Wilson and William A. P. Smith, editors, Proceedings of the British Machine Vision Conference (BMVC), pages 87.1-87.12. BMVA Press, September 2016. 6 +[56] Hongyi Zhang, Moustapha Cisse, Yann N. Dauphin, and David Lopez-Paz. mixup: Beyond empirical risk minimization. CoRR, abs/1710.09412, 2017. 6 +[57] Hongyang Zhang, Yaodong Yu, Jiantao Jiao, Eric Xing, Laurent El Ghaoui, and Michael Jordan. Theoretically principled trade-off between robustness and accuracy. In International conference on machine learning, pages 7472-7482. PMLR, 2019. 2 \ No newline at end of file diff --git a/2023/CUDA_ Convolution-Based Unlearnable Datasets/images.zip b/2023/CUDA_ Convolution-Based Unlearnable Datasets/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..eb96c361ed58d9a659f26bd63d2d89c70d7fbdf0 --- /dev/null +++ b/2023/CUDA_ Convolution-Based Unlearnable Datasets/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:38cd9e4a766103c2fc51cd245a2e48d83b5db9959f8066af05d1bf162da04e7b +size 316251 diff --git a/2023/CUDA_ Convolution-Based Unlearnable Datasets/layout.json b/2023/CUDA_ Convolution-Based Unlearnable Datasets/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..ac6973768da5adf1fa9d78233e3f35ff7bda656f --- /dev/null +++ b/2023/CUDA_ Convolution-Based Unlearnable Datasets/layout.json @@ -0,0 +1,11072 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 146, + 103, + 448, + 119 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 103, + 448, + 119 + ], + "spans": [ + { + "bbox": [ + 146, + 103, + 448, + 119 + ], + "type": "text", + "content": "CUDA: Convolution-based Unlearnable Datasets" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 75, + 144, + 190, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 144, + 190, + 171 + ], + "spans": [ + { + "bbox": [ + 75, + 144, + 190, + 171 + ], + "type": "text", + "content": "Vinu Sankar Sadasivan \nUniversity of Maryland" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 99, + 175, + 166, + 184 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 99, + 175, + 166, + 184 + ], + "spans": [ + { + "bbox": [ + 99, + 175, + 166, + 184 + ], + "type": "text", + "content": "vinu@umd.edu" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 215, + 144, + 378, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 215, + 144, + 378, + 171 + ], + "spans": [ + { + "bbox": [ + 215, + 144, + 378, + 171 + ], + "type": "text", + "content": "Mahdi Soltanolkotabi \nUniversity of Southern California" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 252, + 175, + 341, + 184 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 252, + 175, + 341, + 184 + ], + "spans": [ + { + "bbox": [ + 252, + 175, + 341, + 184 + ], + "type": "text", + "content": "soltanol@usc.edu" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 403, + 144, + 518, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 403, + 144, + 518, + 171 + ], + "spans": [ + { + "bbox": [ + 403, + 144, + 518, + 171 + ], + "type": "text", + "content": "Soheil Feizi \nUniversity of Maryland" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 414, + 175, + 507, + 184 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 414, + 175, + 507, + 184 + ], + "spans": [ + { + "bbox": [ + 414, + 175, + 507, + 184 + ], + "type": "text", + "content": "sfeizi@cs.umd.edu" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "spans": [ + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 238, + 290, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 238, + 290, + 647 + ], + "spans": [ + { + "bbox": [ + 47, + 238, + 290, + 647 + ], + "type": "text", + "content": "Large-scale training of modern deep learning models heavily relies on publicly available data on the web. This potentially unauthorized usage of online data leads to concerns regarding data privacy. Recent works aim to make unlearnable data for deep learning models by adding small, specially designed noises to tackle this issue. However, these methods are vulnerable to adversarial training (AT) and/or are computationally heavy. In this work, we propose a novel, model-free, Convolution-based Unlearnable DAtaset (CUDA) generation technique. CUDA is generated using controlled class-wise convolutions with filters that are randomly generated via a private key. CUDA encourages the network to learn the relation between filters and labels rather than informative features for classifying the clean data. We develop some theoretical analysis demonstrating that CUDA can successfully poison Gaussian mixture data by reducing the clean data performance of the optimal Bayes classifier. We also empirically demonstrate the effectiveness of CUDA with various datasets (CIFAR-10, CIFAR100, ImageNet-100, and Tiny-ImageNet), and architectures (ResNet-18, VGG-16, Wide ResNet-34-10, DenseNet-121, DeIT, EfficientNetV2-S, and MobileNetV2). Our experiments show that CUDA is robust to various data augmentations and training approaches such as smoothing, AT with different budgets, transfer learning, and fine-tuning. For instance, training a ResNet-18 on ImageNet-100 CUDA achieves only " + }, + { + "bbox": [ + 47, + 238, + 290, + 647 + ], + "type": "inline_equation", + "content": "8.96\\%" + }, + { + "bbox": [ + 47, + 238, + 290, + 647 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 47, + 238, + 290, + 647 + ], + "type": "inline_equation", + "content": "40.08\\%" + }, + { + "bbox": [ + 47, + 238, + 290, + 647 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 47, + 238, + 290, + 647 + ], + "type": "inline_equation", + "content": "20.58\\%" + }, + { + "bbox": [ + 47, + 238, + 290, + 647 + ], + "type": "text", + "content": " clean test accuracies with empirical risk minimization (ERM), " + }, + { + "bbox": [ + 47, + 238, + 290, + 647 + ], + "type": "inline_equation", + "content": "L_{\\infty}" + }, + { + "bbox": [ + 47, + 238, + 290, + 647 + ], + "type": "text", + "content": " AT, and " + }, + { + "bbox": [ + 47, + 238, + 290, + 647 + ], + "type": "inline_equation", + "content": "L_{2}" + }, + { + "bbox": [ + 47, + 238, + 290, + 647 + ], + "type": "text", + "content": " AT, respectively. Here, ERM on the clean training data achieves a clean test accuracy of " + }, + { + "bbox": [ + 47, + 238, + 290, + 647 + ], + "type": "inline_equation", + "content": "80.66\\%" + }, + { + "bbox": [ + 47, + 238, + 290, + 647 + ], + "type": "text", + "content": ". CUDA exhibits unlearnability effect with ERM even when only a fraction of the training dataset is perturbed. Furthermore, we also show that CUDA is robust to adaptive defenses designed specifically to break it." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 669, + 128, + 681 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 669, + 128, + 681 + ], + "spans": [ + { + "bbox": [ + 47, + 669, + 128, + 681 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 47, + 689, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 287, + 715 + ], + "type": "text", + "content": "Modern deep learning training frameworks heavily depend on large-scale datasets for achieving high accuracy." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 213, + 547, + 334 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 213, + 547, + 334 + ], + "spans": [ + { + "bbox": [ + 304, + 213, + 547, + 334 + ], + "type": "text", + "content": "This encourages deep learning practitioners to scrape data from the web for data collection [8, 31, 38, 41]. Since a lot of the data is publicly available online, sometime this scraping of data is unauthorized. For instance, a recent article [15] discloses that a private company trained a commercial face recognition system using over three billion facial images collected from the internet without any user consent. Although such massive data can significantly boost the performance of deep learning models, it raises serious concerns about data privacy and security." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 335, + 548, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 335, + 548, + 574 + ], + "spans": [ + { + "bbox": [ + 304, + 335, + 548, + 574 + ], + "type": "text", + "content": "To prevent the unauthorized usage of personal data, a series of recent papers [10, 17, 53] propose to poison data with additive noise. The idea is to make datasets unlearnable for deep learning models by ensuring that they learn the correspondence between noises and labels. Thereby, they do not learn much useful information about the clean data, significantly degrading their clean test accuracy. However, in recent works [11, 17, 47], these unlearnability methods are shown to be vulnerable to adversarial training (AT) frameworks [34]. Motivated by this problem, Fu et al. [11] developed Robust Error-Minimization (REM) noises to make unlearnable data that is protected from AT. While the authors show the effectiveness of REM in multiple scenarios, we demonstrate that these methods are still not robust against different data augmentations or training settings (see Section 3.2). Furthermore, current unlearnability frameworks [10, 11, 17, 53] are model-dependent and require expensive optimization steps on deep learning models to obtain the additive noises. They also need to train the deep learning models from scratch to obtain noises for each new data set." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 574, + 547, + 682 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 574, + 547, + 682 + ], + "spans": [ + { + "bbox": [ + 304, + 574, + 547, + 682 + ], + "type": "text", + "content": "In this paper, we propose a novel Convolution-based Unlearnable DAataset (CUDA) generation technique. We address limitations of existing unlearnable data generation techniques in Section 3.2 and motivate our CUDA technique in Section 3.3. For generating CUDA, an attacker randomly generates different convolutional filters for each class in the dataset using a private key or seed value. These filters are used to perform controlled class-wise convolutions on the clean training dataset to obtain CUDA. As we" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 306, + 693, + 545, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 693, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 693, + 545, + 713 + ], + "type": "text", + "content": "Code available here: https://github.com/vinusankars/Convolution-based-Unlearnability" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3862" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 91, + 76, + 397, + 179 + ], + "blocks": [ + { + "bbox": [ + 91, + 76, + 397, + 179 + ], + "lines": [ + { + "bbox": [ + 91, + 76, + 397, + 179 + ], + "spans": [ + { + "bbox": [ + 91, + 76, + 397, + 179 + ], + "type": "image", + "image_path": "3c62eaf7fff8338381c2b75a38ca730daa8f02c74e3fd14983becd10a92a993c.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 201, + 545, + 267 + ], + "lines": [ + { + "bbox": [ + 46, + 201, + 545, + 267 + ], + "spans": [ + { + "bbox": [ + 46, + 201, + 545, + 267 + ], + "type": "text", + "content": "Figure 1. CIFAR-10 CUDA images from each of the 10 classes convolved using convolutional filters generated with blur parameter " + }, + { + "bbox": [ + 46, + 201, + 545, + 267 + ], + "type": "inline_equation", + "content": "p_b" + }, + { + "bbox": [ + 46, + 201, + 545, + 267 + ], + "type": "text", + "content": ". As seen in the plot, a higher " + }, + { + "bbox": [ + 46, + 201, + 545, + 267 + ], + "type": "inline_equation", + "content": "p_b" + }, + { + "bbox": [ + 46, + 201, + 545, + 267 + ], + "type": "text", + "content": " value results in better unlearnability (lower clean test accuracy), but increased blurring. CIFAR-10 CUDA does not break with ERM for all the three " + }, + { + "bbox": [ + 46, + 201, + 545, + 267 + ], + "type": "inline_equation", + "content": "p_b" + }, + { + "bbox": [ + 46, + 201, + 545, + 267 + ], + "type": "text", + "content": " values. In Section 5.2, we propose Deconvolution-based Adversarial Training (DAT) that we specifically design to break CUDA. DAT adversarially learns class-wise filters to deconvolve CUDA images. However, CIFAR-10 CUDA with " + }, + { + "bbox": [ + 46, + 201, + 545, + 267 + ], + "type": "inline_equation", + "content": "p_b \\in \\{0.3, 0.5\\}" + }, + { + "bbox": [ + 46, + 201, + 545, + 267 + ], + "type": "text", + "content": " does not break with DAT. Therefore, we select " + }, + { + "bbox": [ + 46, + 201, + 545, + 267 + ], + "type": "inline_equation", + "content": "p_b = 0.3" + }, + { + "bbox": [ + 46, + 201, + 545, + 267 + ], + "type": "text", + "content": " (lesser blurring) for CIFAR-10 CUDA as discussed in Section 5.1. More details on DAT in supplementary material." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 408, + 73, + 503, + 182 + ], + "blocks": [ + { + "bbox": [ + 408, + 73, + 503, + 182 + ], + "lines": [ + { + "bbox": [ + 408, + 73, + 503, + 182 + ], + "spans": [ + { + "bbox": [ + 408, + 73, + 503, + 182 + ], + "type": "image", + "image_path": "fe3c7da284d700e4944361e4bb0eedd4e99e2f82fea68330f5d6b146ab36bfa9.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 289, + 289, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 289, + 289, + 586 + ], + "spans": [ + { + "bbox": [ + 46, + 289, + 289, + 586 + ], + "type": "text", + "content": "describe in Sections 3.3 and 5.2, CUDA generation performs controlled convolutions using a blur parameter " + }, + { + "bbox": [ + 46, + 289, + 289, + 586 + ], + "type": "inline_equation", + "content": "p_b" + }, + { + "bbox": [ + 46, + 289, + 289, + 586 + ], + "type": "text", + "content": " to ensure that the semantics of the dataset are preserved (see Figure 1). CUDA generation with a lower blurring parameter " + }, + { + "bbox": [ + 46, + 289, + 289, + 586 + ], + "type": "inline_equation", + "content": "p_b" + }, + { + "bbox": [ + 46, + 289, + 289, + 586 + ], + "type": "text", + "content": " adds less perceptible noises to clean samples. A network trained by a defender on CUDA is encouraged to learn the shortcut relation between class-wise convolutional filters and labels rather than useful features for classifying the clean data. Since the seed value for generating the filters are private, its not possible for the defender to obtain clean data from CUDA alone. Additionally, CUDA exhibits unlearnability effect with ERM even when only a fraction of the training dataset is perturbed (see Section 5). While the existing unlearnability works use additive noises, CUDA generation technique enjoys the advantage of introducing multiplicative noises in the Fourier domain due to the convolution theorem (since convolution of signals is the same as element-wise multiplication in the Fourier domain). This lets CUDA generation add higher amounts of noise in the image space, specifically along the edges in images, and makes it resilient to AT with small additive noise budgets. In Figure 2, with the help of t-SNE plots [49], we also find that the noises added by CUDA generation is not linearly separable while they are linearly separable for the existing works on unlearnability [52]." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 587, + 287, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 587, + 287, + 647 + ], + "spans": [ + { + "bbox": [ + 46, + 587, + 287, + 647 + ], + "type": "text", + "content": "In Section 4, we theoretically show that CUDA generation can successfully poison Gaussian mixture data by degrading the clean data accuracy of the optimal Bayes classifier. We state our result informally below while the formal version is presented in Theorem 2." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 45, + 653, + 288, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 653, + 288, + 712 + ], + "spans": [ + { + "bbox": [ + 45, + 653, + 288, + 712 + ], + "type": "text", + "content": "Theorem 1 (Informal) Let " + }, + { + "bbox": [ + 45, + 653, + 288, + 712 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 45, + 653, + 288, + 712 + ], + "type": "text", + "content": " denote a Gaussian mixture data with two modes, " + }, + { + "bbox": [ + 45, + 653, + 288, + 712 + ], + "type": "inline_equation", + "content": "P_{\\mathcal{D}}" + }, + { + "bbox": [ + 45, + 653, + 288, + 712 + ], + "type": "text", + "content": " denote the optimal Bayes classifier trained on " + }, + { + "bbox": [ + 45, + 653, + 288, + 712 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 45, + 653, + 288, + 712 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 45, + 653, + 288, + 712 + ], + "type": "inline_equation", + "content": "\\tau_{\\mathcal{D}}(P_{\\mathcal{D}})" + }, + { + "bbox": [ + 45, + 653, + 288, + 712 + ], + "type": "text", + "content": " denote the accuracy of the classifier " + }, + { + "bbox": [ + 45, + 653, + 288, + 712 + ], + "type": "inline_equation", + "content": "P_{\\mathcal{D}}" + }, + { + "bbox": [ + 45, + 653, + 288, + 712 + ], + "type": "text", + "content": " on " + }, + { + "bbox": [ + 45, + 653, + 288, + 712 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 45, + 653, + 288, + 712 + ], + "type": "text", + "content": ". Then, under some assumptions, for every clean data " + }, + { + "bbox": [ + 45, + 653, + 288, + 712 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 45, + 653, + 288, + 712 + ], + "type": "text", + "content": ", there is a CUDA " + }, + { + "bbox": [ + 45, + 653, + 288, + 712 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathcal{D}}" + }, + { + "bbox": [ + 45, + 653, + 288, + 712 + ], + "type": "text", + "content": " such that" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 289, + 389, + 301 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 289, + 389, + 301 + ], + "spans": [ + { + "bbox": [ + 306, + 289, + 389, + 301 + ], + "type": "interline_equation", + "content": "\\tau_ {\\mathcal {D}} (P _ {\\tilde {\\mathcal {D}}}) < \\tau_ {\\mathcal {D}} (P _ {\\mathcal {D}}).", + "image_path": "5883658e8fa38eaad9dde2a33964469f03a6510add22b0afa3c5afb1cc8ca17f.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 311, + 545, + 478 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 311, + 545, + 478 + ], + "spans": [ + { + "bbox": [ + 304, + 311, + 545, + 478 + ], + "type": "text", + "content": "Furthermore, our empirical experiments in Section 5 demonstrate the effectiveness of CUDA under various training scenarios such as ERM with various augmentations and regularizations, AT with different budgets, randomized smoothing [6, 22, 27], transfer learning, and fine-tuning. For instance, training a ResNet-18 on CIFAR-10 CUDA achieves only " + }, + { + "bbox": [ + 304, + 311, + 545, + 478 + ], + "type": "inline_equation", + "content": "18.48\\%" + }, + { + "bbox": [ + 304, + 311, + 545, + 478 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 311, + 545, + 478 + ], + "type": "inline_equation", + "content": "44.4\\%" + }, + { + "bbox": [ + 304, + 311, + 545, + 478 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 311, + 545, + 478 + ], + "type": "inline_equation", + "content": "51.14\\%" + }, + { + "bbox": [ + 304, + 311, + 545, + 478 + ], + "type": "text", + "content": " clean test accuracies with ERM, " + }, + { + "bbox": [ + 304, + 311, + 545, + 478 + ], + "type": "inline_equation", + "content": "L_{\\infty}" + }, + { + "bbox": [ + 304, + 311, + 545, + 478 + ], + "type": "text", + "content": " AT, and " + }, + { + "bbox": [ + 304, + 311, + 545, + 478 + ], + "type": "inline_equation", + "content": "L_{2}" + }, + { + "bbox": [ + 304, + 311, + 545, + 478 + ], + "type": "text", + "content": " AT, respectively (see Figure 2). Here, ERM on the clean training data achieves a clean test accuracy of " + }, + { + "bbox": [ + 304, + 311, + 545, + 478 + ], + "type": "inline_equation", + "content": "94.66\\%" + }, + { + "bbox": [ + 304, + 311, + 545, + 478 + ], + "type": "text", + "content": ". In addition, we also design adaptive defenses to investigate if CUDA breaks with random or adversarial defense mechanisms. We find that CUDA is robust to the adaptive defenses that we specifically design to break it." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 489, + 397, + 502 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 489, + 397, + 502 + ], + "spans": [ + { + "bbox": [ + 306, + 489, + 397, + 502 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 510, + 545, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 510, + 545, + 557 + ], + "spans": [ + { + "bbox": [ + 304, + 510, + 545, + 557 + ], + "type": "text", + "content": "Our CUDA generation technique is intimately related with adversarial and poisoning attacks. We first discuss some of this literature and then explain their relation with CUDA generation." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 558, + 545, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 558, + 545, + 676 + ], + "spans": [ + { + "bbox": [ + 304, + 558, + 545, + 676 + ], + "type": "text", + "content": "Adversarial attacks. Adversarial examples are specially designed examples that can fool deep learning models at test time [4, 12, 23, 24, 45]. The adversary crafts these examples by adding error-maximizing noises to the clean data. Even slightly perturbed data can serve as adversarial examples. AT is a training framework proposed to make deep learning models robust to adversarial examples [19,25,34,51,57]. AT is a min-max optimization problem where the model is trained to minimize loss on adversarial examples that have the maximum loss." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "content": "Poisoning attacks. In data poisoning, an attacker aims to hurt the deep learning model's performance by perturbing the training data [2, 20, 28, 32, 42, 50]. The backdoor attack" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3863" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 135, + 71, + 353, + 204 + ], + "blocks": [ + { + "bbox": [ + 135, + 71, + 353, + 204 + ], + "lines": [ + { + "bbox": [ + 135, + 71, + 353, + 204 + ], + "spans": [ + { + "bbox": [ + 135, + 71, + 353, + 204 + ], + "type": "image", + "image_path": "0551ddd300a3b3d91e6650f1b8561e44da0383cc955cd589d1c107c0388fdfef.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 45, + 216, + 547, + 273 + ], + "lines": [ + { + "bbox": [ + 45, + 216, + 547, + 273 + ], + "spans": [ + { + "bbox": [ + 45, + 216, + 547, + 273 + ], + "type": "text", + "content": "Figure 2. In the left column, clean test accuracies of ResNet-18 on clean CIFAR-10 and CIFAR-10 CUDA are shown. The lower the test accuracy, the higher the effectiveness of unlearnability is. We show that CUDA is robust to various training settings such as ERM, " + }, + { + "bbox": [ + 45, + 216, + 547, + 273 + ], + "type": "inline_equation", + "content": "L_{2}" + }, + { + "bbox": [ + 45, + 216, + 547, + 273 + ], + "type": "text", + "content": " AT, and " + }, + { + "bbox": [ + 45, + 216, + 547, + 273 + ], + "type": "inline_equation", + "content": "L_{\\infty}" + }, + { + "bbox": [ + 45, + 216, + 547, + 273 + ], + "type": "text", + "content": " AT. The error bars represent the standard deviation of test accuracy over 5 independent trials. CUDA is robust to the randomness in its data generation process. In the right column, t-SNEs of the noises generated by REM (top) and CUDA (bottom) for the first 3 classes of CIFAR-10 are plotted." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 357, + 71, + 460, + 201 + ], + "blocks": [ + { + "bbox": [ + 357, + 71, + 460, + 201 + ], + "lines": [ + { + "bbox": [ + 357, + 71, + 460, + 201 + ], + "spans": [ + { + "bbox": [ + 357, + 71, + 460, + 201 + ], + "type": "image", + "image_path": "a9c5a7c417aaecaa0f53bc946ffaaa6517a6c3243ee2335214b3699ec499c80f.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 293, + 287, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 293, + 287, + 376 + ], + "spans": [ + { + "bbox": [ + 46, + 293, + 287, + 376 + ], + "type": "text", + "content": "is a special type of poisoning attack where a trigger pattern is injected into clean data at training time [5, 29, 33, 36]. The model trained on this data would misclassify an image with a trigger pattern at test time. Gu et al. [13] and Li et al. [30] use perceptible amounts of noises similar to CUDA for data poisoning. However, backdoor attacks do not affect the performance of the model on clean data [1, 5, 42]." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 377, + 288, + 520 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 377, + 288, + 520 + ], + "spans": [ + { + "bbox": [ + 46, + 377, + 288, + 520 + ], + "type": "text", + "content": "Recent literature utilize data poisoning to protect data from being used for model training without authorization. Yuan and Wu [53] use neural tangent kernels [18] to generate clean label attacks that can hurt the generalization of deep learning models. Huang et al. [17] show that error-minimizing noise addition can serve as a poisoning technique. Fowl et al. [10] show that error-maximizing noises as well can make strong poison attacks. However, all these poisoning techniques do not offer data protection with AT [47, 52]. Fu et al. [11] proposes a min-min-max optimization technique to generate poisoned data that offers better unlearnability effects with AT." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 532, + 287, + 559 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 532, + 287, + 559 + ], + "spans": [ + { + "bbox": [ + 47, + 532, + 287, + 559 + ], + "type": "text", + "content": "3. Convolution-based Ulearnable DAtaset (CUDA)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 566, + 287, + 615 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 566, + 287, + 615 + ], + "spans": [ + { + "bbox": [ + 46, + 566, + 287, + 615 + ], + "type": "text", + "content": "In this section, first we give some preliminaries about unlearnability. Then we discuss the limitations of the existing unlearnability methods. Finally, we propose our CUDA generation technique." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 623, + 134, + 634 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 623, + 134, + 634 + ], + "spans": [ + { + "bbox": [ + 47, + 623, + 134, + 634 + ], + "type": "text", + "content": "3.1. Preliminaries" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 641, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 641, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 641, + 288, + 714 + ], + "type": "text", + "content": "Let " + }, + { + "bbox": [ + 46, + 641, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\{(x_i, y_i)\\}_{i=1}^n \\sim \\mathcal{D}^n" + }, + { + "bbox": [ + 46, + 641, + 288, + 714 + ], + "type": "text", + "content": " be the clean training dataset where " + }, + { + "bbox": [ + 46, + 641, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 46, + 641, + 288, + 714 + ], + "type": "text", + "content": " is the clean data distribution, " + }, + { + "bbox": [ + 46, + 641, + 288, + 714 + ], + "type": "inline_equation", + "content": "x_i \\in \\mathcal{X} \\subset \\mathbb{R}^d" + }, + { + "bbox": [ + 46, + 641, + 288, + 714 + ], + "type": "text", + "content": " are the samples, and " + }, + { + "bbox": [ + 46, + 641, + 288, + 714 + ], + "type": "inline_equation", + "content": "y_i \\in \\mathcal{Y}" + }, + { + "bbox": [ + 46, + 641, + 288, + 714 + ], + "type": "text", + "content": " are the corresponding labels. Suppose a network is given as " + }, + { + "bbox": [ + 46, + 641, + 288, + 714 + ], + "type": "inline_equation", + "content": "f_\\theta: \\mathcal{X} \\to \\mathcal{Y}" + }, + { + "bbox": [ + 46, + 641, + 288, + 714 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 46, + 641, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\theta \\in \\Theta" + }, + { + "bbox": [ + 46, + 641, + 288, + 714 + ], + "type": "text", + "content": " is the network parameter and " + }, + { + "bbox": [ + 46, + 641, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\ell: \\mathcal{Y} \\times \\mathcal{Y} \\to \\mathbb{R}" + }, + { + "bbox": [ + 46, + 641, + 288, + 714 + ], + "type": "text", + "content": " is the loss function. ERM trains a network using a minimization prob" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 292, + 545, + 341 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 292, + 545, + 341 + ], + "spans": [ + { + "bbox": [ + 304, + 292, + 545, + 341 + ], + "type": "text", + "content": "lem of the form: " + }, + { + "bbox": [ + 304, + 292, + 545, + 341 + ], + "type": "inline_equation", + "content": "\\min_{\\theta} \\frac{1}{n} \\sum_{i=1}^{n} \\ell(f_{\\theta}(\\pmb{x}_i), y_i)" + }, + { + "bbox": [ + 304, + 292, + 545, + 341 + ], + "type": "text", + "content": ". Standard " + }, + { + "bbox": [ + 304, + 292, + 545, + 341 + ], + "type": "inline_equation", + "content": "L_p" + }, + { + "bbox": [ + 304, + 292, + 545, + 341 + ], + "type": "text", + "content": "-based AT (for " + }, + { + "bbox": [ + 304, + 292, + 545, + 341 + ], + "type": "inline_equation", + "content": "p \\in \\mathbb{R}^+" + }, + { + "bbox": [ + 304, + 292, + 545, + 341 + ], + "type": "text", + "content": ") solves the following min-max problem: " + }, + { + "bbox": [ + 304, + 292, + 545, + 341 + ], + "type": "inline_equation", + "content": "\\min_{\\theta} \\frac{1}{n} \\sum_{i=1}^{n} \\max_{\\|\\pmb{\\delta}_i\\|_p \\leq \\rho_a} \\ell(f_{\\theta}(\\pmb{x}_i + \\pmb{\\delta}_i), y_i)" + }, + { + "bbox": [ + 304, + 292, + 545, + 341 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 304, + 292, + 545, + 341 + ], + "type": "inline_equation", + "content": "\\rho_a" + }, + { + "bbox": [ + 304, + 292, + 545, + 341 + ], + "type": "text", + "content": " is the adversarial perturbation radius." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 341, + 546, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 341, + 546, + 521 + ], + "spans": [ + { + "bbox": [ + 304, + 341, + 546, + 521 + ], + "type": "text", + "content": "We will use " + }, + { + "bbox": [ + 304, + 341, + 546, + 521 + ], + "type": "inline_equation", + "content": "\\tau_{\\mathcal{D}}(\\theta)" + }, + { + "bbox": [ + 304, + 341, + 546, + 521 + ], + "type": "text", + "content": ", with " + }, + { + "bbox": [ + 304, + 341, + 546, + 521 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 304, + 341, + 546, + 521 + ], + "type": "text", + "content": " the clean model parameter, to denote the clean test accuracy of a model trained on the clean training dataset (i.e., clean model accuracy). In unlearnable dataset generation, an attacker uses an algorithm " + }, + { + "bbox": [ + 304, + 341, + 546, + 521 + ], + "type": "inline_equation", + "content": "\\mathcal{A}:\\mathcal{X}\\rightarrow \\mathcal{X}" + }, + { + "bbox": [ + 304, + 341, + 546, + 521 + ], + "type": "text", + "content": " to generate an unlearnable dataset " + }, + { + "bbox": [ + 304, + 341, + 546, + 521 + ], + "type": "inline_equation", + "content": "\\{(\\tilde{\\boldsymbol{x}}_i = \\mathcal{A}(\\boldsymbol {x}_i),y_i)\\}_{i = 1}^n\\sim \\tilde{\\mathcal{D}}^n" + }, + { + "bbox": [ + 304, + 341, + 546, + 521 + ], + "type": "text", + "content": " from the clean training data. Here, the attacker assumes access to the full clean training dataset. Moreover, the attacker cannot modify the unlearnable dataset once it is released publicly. A defender trains using the unlearnable dataset to obtain a network " + }, + { + "bbox": [ + 304, + 341, + 546, + 521 + ], + "type": "inline_equation", + "content": "f_{\\tilde{\\theta}}" + }, + { + "bbox": [ + 304, + 341, + 546, + 521 + ], + "type": "text", + "content": ". The objective of the attacker is to design an unlearnable dataset such that the defender's model trained on the unlearnable data achieves a clean test accuracy (i.e., unlearnable model accuracy) worse than the clean model accuracy i.e. " + }, + { + "bbox": [ + 304, + 341, + 546, + 521 + ], + "type": "inline_equation", + "content": "\\tau_{\\mathcal{D}}(\\tilde{\\theta})\\ll \\tau_{\\mathcal{D}}(\\theta)" + }, + { + "bbox": [ + 304, + 341, + 546, + 521 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 526, + 465, + 540 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 526, + 465, + 540 + ], + "spans": [ + { + "bbox": [ + 306, + 526, + 465, + 540 + ], + "type": "text", + "content": "3.2. Limitations of existing works" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 544, + 545, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 544, + 545, + 651 + ], + "spans": [ + { + "bbox": [ + 304, + 544, + 545, + 651 + ], + "type": "text", + "content": "Fu et al. [11] show that the previous unlearnability methods including Error-Minimization (EM) [17], Targeted Adversarial Poisoning (TAP) [10], and Neural Tangent Generalization Attack (NTGA) [53] are vulnerable to AT. Hence, they propose a Robust Error-Minimization (REM) [11] method that exploits a min-min-max optimization procedure to generate unlearnable noises. REM first trains a noise generator " + }, + { + "bbox": [ + 304, + 544, + 545, + 651 + ], + "type": "inline_equation", + "content": "f_{\\theta}^{\\prime}" + }, + { + "bbox": [ + 304, + 544, + 545, + 651 + ], + "type": "text", + "content": " on data points " + }, + { + "bbox": [ + 304, + 544, + 545, + 651 + ], + "type": "inline_equation", + "content": "\\{(x_i,y_i)\\}_{i = 1}^n" + }, + { + "bbox": [ + 304, + 544, + 545, + 651 + ], + "type": "text", + "content": " over a loss function " + }, + { + "bbox": [ + 304, + 544, + 545, + 651 + ], + "type": "inline_equation", + "content": "\\ell" + }, + { + "bbox": [ + 304, + 544, + 545, + 651 + ], + "type": "text", + "content": " as follows" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 341, + 657, + 545, + 711 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 341, + 657, + 545, + 711 + ], + "spans": [ + { + "bbox": [ + 341, + 657, + 545, + 711 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\min _ {\\theta} \\frac {1}{n} \\sum_ {i = 1} ^ {n} \\min _ {\\| \\delta_ {i} ^ {u} \\| \\leq \\rho_ {u}} \\mathbb {E} _ {t \\sim T} \\\\ \\max _ {\\| \\boldsymbol {\\delta} _ {i} ^ {a} \\| \\leq \\rho_ {a}} \\ell \\left(f _ {\\theta} ^ {\\prime} \\left(t \\left(\\boldsymbol {x} _ {i} + \\boldsymbol {\\delta} _ {i} ^ {u}\\right) + \\boldsymbol {\\delta} _ {i} ^ {a}\\right), y\\right). \\tag {1} \\\\ \\end{array}", + "image_path": "f7a885838cbbff9d8d0e9a2eb4340f591ded7972e1c8998a741e343e52785a5e.jpg" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3864" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 48, + 102, + 287, + 153 + ], + "blocks": [ + { + "bbox": [ + 47, + 71, + 287, + 93 + ], + "lines": [ + { + "bbox": [ + 47, + 71, + 287, + 93 + ], + "spans": [ + { + "bbox": [ + 47, + 71, + 287, + 93 + ], + "type": "text", + "content": "Table 1. Time taken for generating unlearnable noises for various datasets using different unlearnability techniques." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 48, + 102, + 287, + 153 + ], + "lines": [ + { + "bbox": [ + 48, + 102, + 287, + 153 + ], + "spans": [ + { + "bbox": [ + 48, + 102, + 287, + 153 + ], + "type": "table", + "html": "
DatasetEMTAPNTGAREMCUDA
CIFAR-100.4 hr0.5 hr5.2 hrs22.6 hrs10.8 s
CIFAR-1000.4 hr0.5 hr5.2 hrs22.6 hrs15.5 s
ImageNet-1003.9 hrs5.2 hrs14.6 hrs51.2 hrs0.15 hr
", + "image_path": "103ce9779f82f2d79e51779c6f83ecf7ce50dbf8e261e18e59c5816a1ccc451e.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 173, + 286, + 233 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 173, + 286, + 233 + ], + "spans": [ + { + "bbox": [ + 47, + 173, + 286, + 233 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 47, + 173, + 286, + 233 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 47, + 173, + 286, + 233 + ], + "type": "text", + "content": " is a distribution over a set of transformations " + }, + { + "bbox": [ + 47, + 173, + 286, + 233 + ], + "type": "inline_equation", + "content": "\\{t:" + }, + { + "bbox": [ + 47, + 173, + 286, + 233 + ], + "type": "inline_equation", + "content": "\\mathcal{X}\\to \\mathcal{X}\\}" + }, + { + "bbox": [ + 47, + 173, + 286, + 233 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 47, + 173, + 286, + 233 + ], + "type": "inline_equation", + "content": "\\rho_{u}" + }, + { + "bbox": [ + 47, + 173, + 286, + 233 + ], + "type": "text", + "content": " is the defensive perturbation radius, and " + }, + { + "bbox": [ + 47, + 173, + 286, + 233 + ], + "type": "inline_equation", + "content": "\\rho_{a}" + }, + { + "bbox": [ + 47, + 173, + 286, + 233 + ], + "type": "text", + "content": " controls the protection level of REM against AT. After training the noise generator, an unlearnable example " + }, + { + "bbox": [ + 47, + 173, + 286, + 233 + ], + "type": "inline_equation", + "content": "(\\tilde{\\pmb{x}},y)" + }, + { + "bbox": [ + 47, + 173, + 286, + 233 + ], + "type": "text", + "content": " is generated via" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 77, + 242, + 286, + 285 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 242, + 286, + 285 + ], + "spans": [ + { + "bbox": [ + 77, + 242, + 286, + 285 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\tilde{\\boldsymbol{x}} = \\boldsymbol {x} + \\operatorname *{arg min}_{\\| \\boldsymbol{\\delta}^{u}\\| \\leq \\rho_{u}}\\mathbb{E}_{t\\sim T} \\\\ \\max _ {\\| \\boldsymbol {\\delta} ^ {a} \\| \\leq \\rho_ {a}} \\ell \\left(f _ {\\theta} ^ {\\prime} \\left(t \\left(\\boldsymbol {x} + \\boldsymbol {\\delta} ^ {u}\\right) + \\boldsymbol {\\delta} ^ {a}\\right), y\\right) \\tag {2} \\\\ \\end{array}", + "image_path": "e3863adea4f80c07ce9c4bb0d1e07bde9d37c8b7f137640c606d7c3ca76cc62d.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 293, + 287, + 627 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 293, + 287, + 627 + ], + "spans": [ + { + "bbox": [ + 47, + 293, + 287, + 627 + ], + "type": "text", + "content": "First, note that REM is computationally expensive since it needs to generate unlearnability noises through solving optimization equation 2. Moreover, the existing techniques are model-dependent and they require gradient-based training with a neural network to generate unlearnable data. They also require neural network training from scratch for every dataset that is to be made unlearnable. Table 1 shows the amount of time required to generate various unlearnable datasets using NVIDIA® Tesla V100 GPU and 10 CPU cores. CUDA generation is significantly faster than the existing methods since it uses a model-free approach (no training required). Furthermore, REM is sensitive to hyperparameters and norm-budgets of AT since they generate noises with fixed " + }, + { + "bbox": [ + 47, + 293, + 287, + 627 + ], + "type": "inline_equation", + "content": "L_{\\infty}" + }, + { + "bbox": [ + 47, + 293, + 287, + 627 + ], + "type": "text", + "content": " norm budgets. For instance, a ResNet-18 trained on clean CIFAR-10 dataset achieves a clean test data accuracy of " + }, + { + "bbox": [ + 47, + 293, + 287, + 627 + ], + "type": "inline_equation", + "content": "94.66\\%" + }, + { + "bbox": [ + 47, + 293, + 287, + 627 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 47, + 293, + 287, + 627 + ], + "type": "inline_equation", + "content": "L_{\\infty}" + }, + { + "bbox": [ + 47, + 293, + 287, + 627 + ], + "type": "text", + "content": " AT with perturbation radius " + }, + { + "bbox": [ + 47, + 293, + 287, + 627 + ], + "type": "inline_equation", + "content": "\\rho_{a} = 4 / 255" + }, + { + "bbox": [ + 47, + 293, + 287, + 627 + ], + "type": "text", + "content": " on REM CIFAR-10 data (generated using " + }, + { + "bbox": [ + 47, + 293, + 287, + 627 + ], + "type": "inline_equation", + "content": "\\rho_{u} = 8 / 255" + }, + { + "bbox": [ + 47, + 293, + 287, + 627 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 293, + 287, + 627 + ], + "type": "inline_equation", + "content": "\\rho_{a} = 4 / 255" + }, + { + "bbox": [ + 47, + 293, + 287, + 627 + ], + "type": "text", + "content": ") achieves only a clean test accuracy of " + }, + { + "bbox": [ + 47, + 293, + 287, + 627 + ], + "type": "inline_equation", + "content": "48.16\\%" + }, + { + "bbox": [ + 47, + 293, + 287, + 627 + ], + "type": "text", + "content": ". However, " + }, + { + "bbox": [ + 47, + 293, + 287, + 627 + ], + "type": "inline_equation", + "content": "L_{\\infty}" + }, + { + "bbox": [ + 47, + 293, + 287, + 627 + ], + "type": "text", + "content": " AT with perturbation radius " + }, + { + "bbox": [ + 47, + 293, + 287, + 627 + ], + "type": "inline_equation", + "content": "\\rho_{a} = 8 / 255" + }, + { + "bbox": [ + 47, + 293, + 287, + 627 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 293, + 287, + 627 + ], + "type": "inline_equation", + "content": "L_{2}" + }, + { + "bbox": [ + 47, + 293, + 287, + 627 + ], + "type": "text", + "content": " AT with perturbation radius " + }, + { + "bbox": [ + 47, + 293, + 287, + 627 + ], + "type": "inline_equation", + "content": "\\rho_{a} = 0.75" + }, + { + "bbox": [ + 47, + 293, + 287, + 627 + ], + "type": "text", + "content": " can achieve a clean test accuracy of " + }, + { + "bbox": [ + 47, + 293, + 287, + 627 + ], + "type": "inline_equation", + "content": "78.71\\%" + }, + { + "bbox": [ + 47, + 293, + 287, + 627 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 293, + 287, + 627 + ], + "type": "inline_equation", + "content": "79.65\\%" + }, + { + "bbox": [ + 47, + 293, + 287, + 627 + ], + "type": "text", + "content": ", respectively, on the same REM data. We also find that ERM with a ResNet-18 on grayscaled REM CIFAR-10 images can achieve a high test accuracy of " + }, + { + "bbox": [ + 47, + 293, + 287, + 627 + ], + "type": "inline_equation", + "content": "70.76\\%" + }, + { + "bbox": [ + 47, + 293, + 287, + 627 + ], + "type": "text", + "content": " on the grayscaled CIFAR-10 test data. This shows that REM relies upon the color space for poisoning clean data. Fu et al. [11] also show that REM noise generated using ResNet-18 is not transferable to DenseNet-121." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 635, + 166, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 635, + 166, + 647 + ], + "spans": [ + { + "bbox": [ + 47, + 635, + 166, + 647 + ], + "type": "text", + "content": "3.3. Our method: CUDA" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 654, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 654, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 654, + 286, + 713 + ], + "type": "text", + "content": "The major limitations of the previous works are that they are vulnerable to AT, and computationally expensive. We think the major reason for the former limitation is the usage of small additive noises for unlearnability. AT is designed to train in the presence of such additive noises. In" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 72, + 545, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 323 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 323 + ], + "type": "text", + "content": "creasing the budget of the amount of additive noises for unlearnability might destroy the semantics of the images while perturbing them. The latter limitation arises from the fact that these methods are model-dependent and they require multilevel optimizations. Hence, we are motivated to design a compute-efficient unlearnability method that is robust to AT. CUDA technique can perform convolutions to add larger amounts of noises to clean images without destroying its semantics. This can help CUDA to be robust against AT. CUDA uses randomly generated convolution filters for blurring images from each class. This makes a model trained on CUDA to learn shortcut relations between filters and labels. We empirically support these claims in Section 5.2. Additionally, randomly generating the filters makes our technique model-free. Since the keys for generating the filters are private, it is not possible to reverse the blurring effect in CUDA without having access to the corresponding clean images. Hence, we assume that the data publisher deletes the clean images after perturbing them. Moreover, CUDA technique is a novel non-additive noise based poisoning attack that needs to be studied." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 323, + 545, + 526 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 323, + 545, + 526 + ], + "spans": [ + { + "bbox": [ + 304, + 323, + 545, + 526 + ], + "type": "text", + "content": "CUDA uses convolutional filters " + }, + { + "bbox": [ + 304, + 323, + 545, + 526 + ], + "type": "inline_equation", + "content": "s_i \\in \\mathbb{R}^{k \\times k}" + }, + { + "bbox": [ + 304, + 323, + 545, + 526 + ], + "type": "text", + "content": " for each class " + }, + { + "bbox": [ + 304, + 323, + 545, + 526 + ], + "type": "inline_equation", + "content": "i \\in [1, K]" + }, + { + "bbox": [ + 304, + 323, + 545, + 526 + ], + "type": "text", + "content": ". A random parameter, out of the " + }, + { + "bbox": [ + 304, + 323, + 545, + 526 + ], + "type": "inline_equation", + "content": "k^2" + }, + { + "bbox": [ + 304, + 323, + 545, + 526 + ], + "type": "text", + "content": " parameters, in each of the filters is set to have a value of 1. The rest of the filter parameters are randomly initialized from a uniform distribution " + }, + { + "bbox": [ + 304, + 323, + 545, + 526 + ], + "type": "inline_equation", + "content": "\\mathcal{U}(0, p_b)" + }, + { + "bbox": [ + 304, + 323, + 545, + 526 + ], + "type": "text", + "content": " using a private seed where " + }, + { + "bbox": [ + 304, + 323, + 545, + 526 + ], + "type": "inline_equation", + "content": "p_b" + }, + { + "bbox": [ + 304, + 323, + 545, + 526 + ], + "type": "text", + "content": " is the blur parameter. Blur parameter controls the level of blurring that occurs when an image " + }, + { + "bbox": [ + 304, + 323, + 545, + 526 + ], + "type": "inline_equation", + "content": "x \\in [0, 1]^{d_1 \\times d_2 \\times d_3}" + }, + { + "bbox": [ + 304, + 323, + 545, + 526 + ], + "type": "text", + "content": " is convolved with a filter " + }, + { + "bbox": [ + 304, + 323, + 545, + 526 + ], + "type": "inline_equation", + "content": "s_i" + }, + { + "bbox": [ + 304, + 323, + 545, + 526 + ], + "type": "text", + "content": ". Here, " + }, + { + "bbox": [ + 304, + 323, + 545, + 526 + ], + "type": "inline_equation", + "content": "d_1, d_2" + }, + { + "bbox": [ + 304, + 323, + 545, + 526 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 323, + 545, + 526 + ], + "type": "inline_equation", + "content": "d_3" + }, + { + "bbox": [ + 304, + 323, + 545, + 526 + ], + "type": "text", + "content": " are the height, width, and number of channels of the image, respectively. For example, a CIFAR-10 image has a dimension of " + }, + { + "bbox": [ + 304, + 323, + 545, + 526 + ], + "type": "inline_equation", + "content": "32 \\times 32 \\times 3" + }, + { + "bbox": [ + 304, + 323, + 545, + 526 + ], + "type": "text", + "content": ". The higher " + }, + { + "bbox": [ + 304, + 323, + 545, + 526 + ], + "type": "inline_equation", + "content": "p_b" + }, + { + "bbox": [ + 304, + 323, + 545, + 526 + ], + "type": "text", + "content": " is, the higher the blurring effect is. Let " + }, + { + "bbox": [ + 304, + 323, + 545, + 526 + ], + "type": "inline_equation", + "content": "\\hat{x} = x * s_i" + }, + { + "bbox": [ + 304, + 323, + 545, + 526 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 304, + 323, + 545, + 526 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 304, + 323, + 545, + 526 + ], + "type": "text", + "content": " belongs to class " + }, + { + "bbox": [ + 304, + 323, + 545, + 526 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 323, + 545, + 526 + ], + "type": "text", + "content": ". The CUDA data point for " + }, + { + "bbox": [ + 304, + 323, + 545, + 526 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 304, + 323, + 545, + 526 + ], + "type": "text", + "content": " is given by " + }, + { + "bbox": [ + 304, + 323, + 545, + 526 + ], + "type": "inline_equation", + "content": "\\hat{x} = \\hat{x} / \\mathrm{MAX}(\\hat{x})" + }, + { + "bbox": [ + 304, + 323, + 545, + 526 + ], + "type": "text", + "content": ". Rescaling is performed to make sure that the CUDA image pixels lie between 0 and 1. We find that the unlearnability effect gets stronger with larger " + }, + { + "bbox": [ + 304, + 323, + 545, + 526 + ], + "type": "inline_equation", + "content": "p_b" + }, + { + "bbox": [ + 304, + 323, + 545, + 526 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 323, + 545, + 526 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 323, + 545, + 526 + ], + "type": "text", + "content": " values (see supplementary material for details)." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 538, + 413, + 552 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 538, + 413, + 552 + ], + "spans": [ + { + "bbox": [ + 306, + 538, + 413, + 552 + ], + "type": "text", + "content": "4. Theory for CUDA" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 559, + 545, + 654 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 559, + 545, + 654 + ], + "spans": [ + { + "bbox": [ + 304, + 559, + 545, + 654 + ], + "type": "text", + "content": "In this section, we define a binary classification setup similar to [19, 35] to theoretically analyze CUDA. Let " + }, + { + "bbox": [ + 304, + 559, + 545, + 654 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 304, + 559, + 545, + 654 + ], + "type": "text", + "content": " be a clean dataset modelled by an isotropic Gaussian mixture model given by " + }, + { + "bbox": [ + 304, + 559, + 545, + 654 + ], + "type": "inline_equation", + "content": "\\mathcal{N}(y\\pmb{\\mu}, I)" + }, + { + "bbox": [ + 304, + 559, + 545, + 654 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 559, + 545, + 654 + ], + "type": "inline_equation", + "content": "y \\in \\{\\pm 1\\}" + }, + { + "bbox": [ + 304, + 559, + 545, + 654 + ], + "type": "text", + "content": " is the class label, " + }, + { + "bbox": [ + 304, + 559, + 545, + 654 + ], + "type": "inline_equation", + "content": "\\pmb{\\mu} \\in \\mathbb{R}^d" + }, + { + "bbox": [ + 304, + 559, + 545, + 654 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 559, + 545, + 654 + ], + "type": "inline_equation", + "content": "I \\in \\mathbb{R}^{d \\times d}" + }, + { + "bbox": [ + 304, + 559, + 545, + 654 + ], + "type": "text", + "content": " is the identity matrix. We defer the proofs for all the lemmas and theorem to the supplementary material. The Bayes optimal decision boundary for classifying this Gaussian mixture model is as follows." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 666, + 545, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 666, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 666, + 545, + 715 + ], + "type": "text", + "content": "Lemma 1 The Bayes optimal decision boundary for classifying " + }, + { + "bbox": [ + 304, + 666, + 545, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 304, + 666, + 545, + 715 + ], + "type": "text", + "content": " is given by " + }, + { + "bbox": [ + 304, + 666, + 545, + 715 + ], + "type": "inline_equation", + "content": "P(\\pmb{x}) \\equiv \\pmb{\\mu}^{\\top} \\pmb{x} = 0" + }, + { + "bbox": [ + 304, + 666, + 545, + 715 + ], + "type": "text", + "content": ". The accuracy of the decision boundary " + }, + { + "bbox": [ + 304, + 666, + 545, + 715 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 304, + 666, + 545, + 715 + ], + "type": "text", + "content": " on the clean dataset " + }, + { + "bbox": [ + 304, + 666, + 545, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 304, + 666, + 545, + 715 + ], + "type": "text", + "content": " (i.e. " + }, + { + "bbox": [ + 304, + 666, + 545, + 715 + ], + "type": "inline_equation", + "content": "\\tau_{\\mathcal{D}}(P)" + }, + { + "bbox": [ + 304, + 666, + 545, + 715 + ], + "type": "text", + "content": ") is equal to " + }, + { + "bbox": [ + 304, + 666, + 545, + 715 + ], + "type": "inline_equation", + "content": "\\phi(\\|\\pmb{\\mu}\\|_2)" + }, + { + "bbox": [ + 304, + 666, + 545, + 715 + ], + "type": "text", + "content": ". Here, " + }, + { + "bbox": [ + 304, + 666, + 545, + 715 + ], + "type": "inline_equation", + "content": "\\phi(.)" + }, + { + "bbox": [ + 304, + 666, + 545, + 715 + ], + "type": "text", + "content": " represents the CDF of the" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3865" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 153, + 70, + 244, + 207 + ], + "blocks": [ + { + "bbox": [ + 153, + 70, + 244, + 207 + ], + "lines": [ + { + "bbox": [ + 153, + 70, + 244, + 207 + ], + "spans": [ + { + "bbox": [ + 153, + 70, + 244, + 207 + ], + "type": "image", + "image_path": "c03e7e9239abc12ac49cd209af107cff469828f26ce520550c6089288a1847c4.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 171, + 209, + 225, + 221 + ], + "lines": [ + { + "bbox": [ + 171, + 209, + 225, + 221 + ], + "spans": [ + { + "bbox": [ + 171, + 209, + 225, + 221 + ], + "type": "text", + "content": "(a) " + }, + { + "bbox": [ + 171, + 209, + 225, + 221 + ], + "type": "inline_equation", + "content": "\\| \\pmb {\\mu}\\| _2 = 1.0" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 252, + 70, + 342, + 207 + ], + "blocks": [ + { + "bbox": [ + 252, + 70, + 342, + 207 + ], + "lines": [ + { + "bbox": [ + 252, + 70, + 342, + 207 + ], + "spans": [ + { + "bbox": [ + 252, + 70, + 342, + 207 + ], + "type": "image", + "image_path": "f652421a0fd1624ff8771573c4ac561d5b841f9e5159fe0cbe40f0159d577972.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 269, + 210, + 324, + 221 + ], + "lines": [ + { + "bbox": [ + 269, + 210, + 324, + 221 + ], + "spans": [ + { + "bbox": [ + 269, + 210, + 324, + 221 + ], + "type": "text", + "content": "(b) " + }, + { + "bbox": [ + 269, + 210, + 324, + 221 + ], + "type": "inline_equation", + "content": "\\| \\pmb{\\mu} \\|_2 = 2.5" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 230, + 547, + 265 + ], + "lines": [ + { + "bbox": [ + 46, + 230, + 547, + 265 + ], + "spans": [ + { + "bbox": [ + 46, + 230, + 547, + 265 + ], + "type": "text", + "content": "Figure 3. We plot the contour plots for clean test accuracy of the CUDA Bayes classifier " + }, + { + "bbox": [ + 46, + 230, + 547, + 265 + ], + "type": "inline_equation", + "content": "\\tilde{P}" + }, + { + "bbox": [ + 46, + 230, + 547, + 265 + ], + "type": "text", + "content": " (given as " + }, + { + "bbox": [ + 46, + 230, + 547, + 265 + ], + "type": "inline_equation", + "content": "\\tau_{\\mathcal{D}}(\\tilde{P})" + }, + { + "bbox": [ + 46, + 230, + 547, + 265 + ], + "type": "text", + "content": ") for varying " + }, + { + "bbox": [ + 46, + 230, + 547, + 265 + ], + "type": "inline_equation", + "content": "a_{y}" + }, + { + "bbox": [ + 46, + 230, + 547, + 265 + ], + "type": "text", + "content": " parameters where " + }, + { + "bbox": [ + 46, + 230, + 547, + 265 + ], + "type": "inline_equation", + "content": "y \\in \\{\\pm 1\\}" + }, + { + "bbox": [ + 46, + 230, + 547, + 265 + ], + "type": "text", + "content": " in Figures 3(a-c). Gaussian mixture model " + }, + { + "bbox": [ + 46, + 230, + 547, + 265 + ], + "type": "inline_equation", + "content": "\\mathcal{N}(y\\mu, I)" + }, + { + "bbox": [ + 46, + 230, + 547, + 265 + ], + "type": "text", + "content": " denotes the clean data. In the top row, we use Lemmas 1 & 4 to empirically generate the plots. In the bottom row, we plot the theoretical upper bound of " + }, + { + "bbox": [ + 46, + 230, + 547, + 265 + ], + "type": "inline_equation", + "content": "\\tau_{\\mathcal{D}}(\\tilde{P})" + }, + { + "bbox": [ + 46, + 230, + 547, + 265 + ], + "type": "text", + "content": " we obtain using Theorem 2." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 350, + 70, + 441, + 206 + ], + "blocks": [ + { + "bbox": [ + 350, + 70, + 441, + 206 + ], + "lines": [ + { + "bbox": [ + 350, + 70, + 441, + 206 + ], + "spans": [ + { + "bbox": [ + 350, + 70, + 441, + 206 + ], + "type": "image", + "image_path": "d105371da4d0456bc3918acd25f6d10a0c36cb54bfb91fe493cf2672ceb5fc0a.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 369, + 210, + 423, + 221 + ], + "lines": [ + { + "bbox": [ + 369, + 210, + 423, + 221 + ], + "spans": [ + { + "bbox": [ + 369, + 210, + 423, + 221 + ], + "type": "text", + "content": "(c) " + }, + { + "bbox": [ + 369, + 210, + 423, + 221 + ], + "type": "inline_equation", + "content": "\\| \\pmb{\\mu}\\| _2 = 5.0" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 285, + 167, + 296 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 285, + 167, + 296 + ], + "spans": [ + { + "bbox": [ + 47, + 285, + 167, + 296 + ], + "type": "text", + "content": "standard normal distribution." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 305, + 289, + 473 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 305, + 289, + 473 + ], + "spans": [ + { + "bbox": [ + 46, + 305, + 289, + 473 + ], + "type": "text", + "content": "Now, to obtain " + }, + { + "bbox": [ + 46, + 305, + 289, + 473 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathcal{D}}" + }, + { + "bbox": [ + 46, + 305, + 289, + 473 + ], + "type": "text", + "content": " (i.e., CUDA), we perform class-wise 1D convolutions. Here, we consider a class of 1D convolutional filters with kernel size 3 of the form " + }, + { + "bbox": [ + 46, + 305, + 289, + 473 + ], + "type": "inline_equation", + "content": "f_{a} = [a,1,a]" + }, + { + "bbox": [ + 46, + 305, + 289, + 473 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 46, + 305, + 289, + 473 + ], + "type": "inline_equation", + "content": "a\\in [0,0.5]" + }, + { + "bbox": [ + 46, + 305, + 289, + 473 + ], + "type": "text", + "content": ". The convolution of signal " + }, + { + "bbox": [ + 46, + 305, + 289, + 473 + ], + "type": "inline_equation", + "content": "\\pmb {x}\\in \\mathbb{R}^{d}" + }, + { + "bbox": [ + 46, + 305, + 289, + 473 + ], + "type": "text", + "content": " with filter " + }, + { + "bbox": [ + 46, + 305, + 289, + 473 + ], + "type": "inline_equation", + "content": "\\pmb{f}_{a}" + }, + { + "bbox": [ + 46, + 305, + 289, + 473 + ], + "type": "text", + "content": " using stride 1 (denoted by " + }, + { + "bbox": [ + 46, + 305, + 289, + 473 + ], + "type": "inline_equation", + "content": "\\pmb {x}*\\pmb{f}_{a}" + }, + { + "bbox": [ + 46, + 305, + 289, + 473 + ], + "type": "text", + "content": "), can be treated as a matrix operation " + }, + { + "bbox": [ + 46, + 305, + 289, + 473 + ], + "type": "inline_equation", + "content": "Ax" + }, + { + "bbox": [ + 46, + 305, + 289, + 473 + ], + "type": "text", + "content": ". Here, " + }, + { + "bbox": [ + 46, + 305, + 289, + 473 + ], + "type": "inline_equation", + "content": "A\\in \\mathbb{R}^{d\\times d}" + }, + { + "bbox": [ + 46, + 305, + 289, + 473 + ], + "type": "text", + "content": " is a tri-diagonal Toeplitz matrix denoted by " + }, + { + "bbox": [ + 46, + 305, + 289, + 473 + ], + "type": "inline_equation", + "content": "T(d;a,1,a)" + }, + { + "bbox": [ + 46, + 305, + 289, + 473 + ], + "type": "text", + "content": ". For CUDA generation, we use class-wise convolution matrices " + }, + { + "bbox": [ + 46, + 305, + 289, + 473 + ], + "type": "inline_equation", + "content": "A_{y} = T(d;a_{y},1,a_{y})" + }, + { + "bbox": [ + 46, + 305, + 289, + 473 + ], + "type": "text", + "content": " to perturb a clean data point " + }, + { + "bbox": [ + 46, + 305, + 289, + 473 + ], + "type": "inline_equation", + "content": "(x,y)\\sim \\mathcal{D}" + }, + { + "bbox": [ + 46, + 305, + 289, + 473 + ], + "type": "text", + "content": " to an unlearnable data point " + }, + { + "bbox": [ + 46, + 305, + 289, + 473 + ], + "type": "inline_equation", + "content": "(A_yx,y)" + }, + { + "bbox": [ + 46, + 305, + 289, + 473 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 305, + 289, + 473 + ], + "type": "inline_equation", + "content": "a_{y}\\in [0,0.5]" + }, + { + "bbox": [ + 46, + 305, + 289, + 473 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 46, + 305, + 289, + 473 + ], + "type": "inline_equation", + "content": "y\\in \\{\\pm 1\\}" + }, + { + "bbox": [ + 46, + 305, + 289, + 473 + ], + "type": "text", + "content": ". Note that in CUDA, the labels remain the same. Next we show that such a perturbed dataset (i.e. CUDA) can be represented as a Gaussian mixture model." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 482, + 287, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 482, + 287, + 533 + ], + "spans": [ + { + "bbox": [ + 46, + 482, + 287, + 533 + ], + "type": "text", + "content": "Lemma 2 A CUDA generated from clean data distribution " + }, + { + "bbox": [ + 46, + 482, + 287, + 533 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 46, + 482, + 287, + 533 + ], + "type": "text", + "content": " and the " + }, + { + "bbox": [ + 46, + 482, + 287, + 533 + ], + "type": "inline_equation", + "content": "A_{y}" + }, + { + "bbox": [ + 46, + 482, + 287, + 533 + ], + "type": "text", + "content": "'s defined above can be modelled as a Gaussian mixture model with the distribution given by " + }, + { + "bbox": [ + 46, + 482, + 287, + 533 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathcal{D}} = \\mathcal{N}(yA_y\\pmb {\\mu},A_y^2)" + }, + { + "bbox": [ + 46, + 482, + 287, + 533 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 540, + 287, + 576 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 540, + 287, + 576 + ], + "spans": [ + { + "bbox": [ + 46, + 540, + 287, + 576 + ], + "type": "text", + "content": "To characterize the decision boundary for CUDA, we need to use some properties of Toeplitz matrices from Noschese et al. [37] given in the following Lemma 3." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 45, + 584, + 286, + 650 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 584, + 286, + 650 + ], + "spans": [ + { + "bbox": [ + 45, + 584, + 286, + 650 + ], + "type": "text", + "content": "Lemma 3 Any tri-diagonal Toeplitz matrix " + }, + { + "bbox": [ + 45, + 584, + 286, + 650 + ], + "type": "inline_equation", + "content": "A = T(d; a, 1, a) \\in \\mathbb{R}^{d \\times d}" + }, + { + "bbox": [ + 45, + 584, + 286, + 650 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 45, + 584, + 286, + 650 + ], + "type": "inline_equation", + "content": "a \\in [0, 0.5]" + }, + { + "bbox": [ + 45, + 584, + 286, + 650 + ], + "type": "text", + "content": " can be diagonalized as " + }, + { + "bbox": [ + 45, + 584, + 286, + 650 + ], + "type": "inline_equation", + "content": "A = QDQ" + }, + { + "bbox": [ + 45, + 584, + 286, + 650 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 45, + 584, + 286, + 650 + ], + "type": "inline_equation", + "content": "Q_{i,j} = \\left(\\frac{2}{d + 1}\\right)^{1/2} \\sin \\left(\\frac{ij\\pi}{d + 1}\\right)" + }, + { + "bbox": [ + 45, + 584, + 286, + 650 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 45, + 584, + 286, + 650 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 45, + 584, + 286, + 650 + ], + "type": "text", + "content": " is a diagonal matrix with " + }, + { + "bbox": [ + 45, + 584, + 286, + 650 + ], + "type": "inline_equation", + "content": "D_{i,i} = 1 + 2a \\cos \\left(\\frac{i\\pi}{d + 1}\\right)" + }, + { + "bbox": [ + 45, + 584, + 286, + 650 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 45, + 584, + 286, + 650 + ], + "type": "inline_equation", + "content": "1 \\leq i, j \\leq d" + }, + { + "bbox": [ + 45, + 584, + 286, + 650 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 658, + 286, + 682 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 658, + 286, + 682 + ], + "spans": [ + { + "bbox": [ + 46, + 658, + 286, + 682 + ], + "type": "text", + "content": "Next we use Lemma 3 to show that the Bayes optimal decision boundary for classifying " + }, + { + "bbox": [ + 46, + 658, + 286, + 682 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathcal{D}}" + }, + { + "bbox": [ + 46, + 658, + 286, + 682 + ], + "type": "text", + "content": " is a quadratic plane." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 46, + 689, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 689, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 689, + 287, + 714 + ], + "type": "text", + "content": "Lemma 4 Let " + }, + { + "bbox": [ + 46, + 689, + 287, + 714 + ], + "type": "inline_equation", + "content": "A_{-1} = T(d; a_{-1}, 1, a_{-1})" + }, + { + "bbox": [ + 46, + 689, + 287, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 689, + 287, + 714 + ], + "type": "inline_equation", + "content": "A_1 = T(d; a_1, 1, a_1)" + }, + { + "bbox": [ + 46, + 689, + 287, + 714 + ], + "type": "text", + "content": ". The Bayes optimal decision boundary for" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 283, + 546, + 327 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 283, + 546, + 327 + ], + "spans": [ + { + "bbox": [ + 305, + 283, + 546, + 327 + ], + "type": "text", + "content": "classifying " + }, + { + "bbox": [ + 305, + 283, + 546, + 327 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathcal{D}}" + }, + { + "bbox": [ + 305, + 283, + 546, + 327 + ], + "type": "text", + "content": " is given by " + }, + { + "bbox": [ + 305, + 283, + 546, + 327 + ], + "type": "inline_equation", + "content": "\\tilde{P} (\\pmb {x})\\equiv \\pmb{x}^{\\top}A\\pmb {x} + \\pmb{b}^{\\top}\\pmb {x} + c = 0," + }, + { + "bbox": [ + 305, + 283, + 546, + 327 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 305, + 283, + 546, + 327 + ], + "type": "inline_equation", + "content": "A = A_{-1}^{-2} - A_1^{-2},\\pmb {b} = 2(A_{-1}^{-1} + A_{1}^{-1})\\pmb {\\mu}," + }, + { + "bbox": [ + 305, + 283, + 546, + 327 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 283, + 546, + 327 + ], + "type": "inline_equation", + "content": "c = 2\\sum_{i = 1}^{d}\\left[\\ln (1 + 2a_{-1}\\cos (\\frac{i\\pi}{d + 1})) - \\ln (1 + 2a_{1}\\cos (\\frac{i\\pi}{d + 1}))\\right]." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 331, + 545, + 366 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 331, + 545, + 366 + ], + "spans": [ + { + "bbox": [ + 305, + 331, + 545, + 366 + ], + "type": "text", + "content": "Now we state a lemma regarding the tail of Gaussian quadratic forms which plays a crucial role in our main result." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 305, + 374, + 546, + 423 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 374, + 546, + 423 + ], + "spans": [ + { + "bbox": [ + 305, + 374, + 546, + 423 + ], + "type": "text", + "content": "Lemma 5 Let " + }, + { + "bbox": [ + 305, + 374, + 546, + 423 + ], + "type": "inline_equation", + "content": "\\| \\cdot \\|" + }, + { + "bbox": [ + 305, + 374, + 546, + 423 + ], + "type": "text", + "content": " denote the operator norm, " + }, + { + "bbox": [ + 305, + 374, + 546, + 423 + ], + "type": "inline_equation", + "content": "\\| \\cdot \\|_2" + }, + { + "bbox": [ + 305, + 374, + 546, + 423 + ], + "type": "text", + "content": " denote the vector 2-norm, " + }, + { + "bbox": [ + 305, + 374, + 546, + 423 + ], + "type": "inline_equation", + "content": "z \\sim \\mathcal{N}(\\mathbf{0}, I)" + }, + { + "bbox": [ + 305, + 374, + 546, + 423 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 305, + 374, + 546, + 423 + ], + "type": "inline_equation", + "content": "Z = z^\\top A z + z^\\top b + c" + }, + { + "bbox": [ + 305, + 374, + 546, + 423 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 305, + 374, + 546, + 423 + ], + "type": "inline_equation", + "content": "A = Q \\Lambda Q^\\top" + }, + { + "bbox": [ + 305, + 374, + 546, + 423 + ], + "type": "text", + "content": ". Using Chernoff bound, for any " + }, + { + "bbox": [ + 305, + 374, + 546, + 423 + ], + "type": "inline_equation", + "content": "t \\geq 0" + }, + { + "bbox": [ + 305, + 374, + 546, + 423 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 374, + 546, + 423 + ], + "type": "inline_equation", + "content": "\\gamma \\in \\mathbb{R}" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 444, + 545, + 491 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 444, + 545, + 491 + ], + "spans": [ + { + "bbox": [ + 306, + 444, + 545, + 491 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathbb {P} \\{Z - \\mathbb {E} Z \\geq \\gamma \\} \\leq \\\\ \\frac {\\exp \\left\\{\\frac {- t}{4 \\| \\boldsymbol {b} \\| _ {2} ^ {\\frac {2}{2}} \\| \\boldsymbol {\\Lambda} \\|} - t (\\gamma + \\operatorname {T r} (\\boldsymbol {\\Lambda}) + \\| \\boldsymbol {b} \\| _ {2}) \\right\\}}{| I - 2 t \\boldsymbol {\\Lambda} | ^ {\\frac {1}{2}}}. \\\\ \\end{array}", + "image_path": "cb0c1f0c9a41599ee31f1ce086d4978ee2f03a603f21b6b307e1cebac5a6e94a.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 305, + 498, + 545, + 535 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 498, + 545, + 535 + ], + "spans": [ + { + "bbox": [ + 305, + 498, + 545, + 535 + ], + "type": "text", + "content": "Lemma 5 allows us to provide an upper bound for the accuracy of the unlearnable decision boundary " + }, + { + "bbox": [ + 305, + 498, + 545, + 535 + ], + "type": "inline_equation", + "content": "\\tilde{P}" + }, + { + "bbox": [ + 305, + 498, + 545, + 535 + ], + "type": "text", + "content": " on the clean dataset " + }, + { + "bbox": [ + 305, + 498, + 545, + 535 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 305, + 498, + 545, + 535 + ], + "type": "text", + "content": ", given as " + }, + { + "bbox": [ + 305, + 498, + 545, + 535 + ], + "type": "inline_equation", + "content": "\\tau_{\\mathcal{D}}(\\tilde{P})" + }, + { + "bbox": [ + 305, + 498, + 545, + 535 + ], + "type": "text", + "content": ", in Theorem 2 below." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 305, + 544, + 547, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 544, + 547, + 593 + ], + "spans": [ + { + "bbox": [ + 305, + 544, + 547, + 593 + ], + "type": "text", + "content": "Theorem 2 (Main result) Let " + }, + { + "bbox": [ + 305, + 544, + 547, + 593 + ], + "type": "inline_equation", + "content": "A = Q\\Lambda Q = A_{-1}^{-2} - A_{1}^{-2}" + }, + { + "bbox": [ + 305, + 544, + 547, + 593 + ], + "type": "text", + "content": ". For any non-negative constants " + }, + { + "bbox": [ + 305, + 544, + 547, + 593 + ], + "type": "inline_equation", + "content": "t_1" + }, + { + "bbox": [ + 305, + 544, + 547, + 593 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 544, + 547, + 593 + ], + "type": "inline_equation", + "content": "t_2" + }, + { + "bbox": [ + 305, + 544, + 547, + 593 + ], + "type": "text", + "content": ", the accuracy of the unlearnable decision boundary " + }, + { + "bbox": [ + 305, + 544, + 547, + 593 + ], + "type": "inline_equation", + "content": "\\tilde{P}" + }, + { + "bbox": [ + 305, + 544, + 547, + 593 + ], + "type": "text", + "content": " on the clean dataset " + }, + { + "bbox": [ + 305, + 544, + 547, + 593 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 305, + 544, + 547, + 593 + ], + "type": "text", + "content": " can be upper-bounded as" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 600, + 541, + 679 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 600, + 541, + 679 + ], + "spans": [ + { + "bbox": [ + 308, + 600, + 541, + 679 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\tau_ {\\mathcal {D}} (\\tilde {P}) \\leq \\frac {\\exp \\left[ t _ {1} \\left(\\boldsymbol {b} ^ {\\top} \\boldsymbol {\\mu} + \\boldsymbol {\\mu} ^ {\\top} A \\boldsymbol {\\mu} + c - \\frac {1}{4 \\| 2 A \\boldsymbol {\\mu} + \\boldsymbol {b} \\| _ {2} \\| \\Lambda \\|}\\right) \\right]}{2 | I - 2 t _ {1} \\Lambda | ^ {\\frac {1}{2}}} \\\\ + \\frac {\\exp \\left[ t _ {2} \\left(\\boldsymbol {b} ^ {\\top} \\boldsymbol {\\mu} - \\boldsymbol {\\mu} ^ {\\top} A \\boldsymbol {\\mu} - c - \\frac {1}{4 \\| 2 A \\boldsymbol {\\mu} - \\boldsymbol {b} \\| _ {2} \\| \\Lambda \\|}\\right) \\right]}{2 | I - 2 t _ {2} \\Lambda | ^ {\\frac {1}{2}}} \\\\ := p _ {1} + p _ {2}. \\\\ \\end{array}", + "image_path": "ff2fa3ba9b9d1930e5b1e5d264d248b32264d083604bf8ac8ccaf81e50f41574.jpg" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 306, + 688, + 545, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 688, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 306, + 688, + 545, + 715 + ], + "type": "text", + "content": "Furthermore, if " + }, + { + "bbox": [ + 306, + 688, + 545, + 715 + ], + "type": "inline_equation", + "content": "\\pmb{\\mu}^{\\top}A\\pmb{\\mu} + \\pmb{b}^{\\top}\\pmb{\\mu} + c + \\mathrm{Tr}(A) + \\| 2A\\pmb{\\mu} + \\pmb{b}\\| < 0" + }, + { + "bbox": [ + 306, + 688, + 545, + 715 + ], + "type": "text", + "content": " we have " + }, + { + "bbox": [ + 306, + 688, + 545, + 715 + ], + "type": "inline_equation", + "content": "\\tau_{\\mathcal{D}}(\\tilde{P})\\leq \\frac{1}{2} (p_1 + 1) < 1" + }, + { + "bbox": [ + 306, + 688, + 545, + 715 + ], + "type": "text", + "content": ". Also, if " + }, + { + "bbox": [ + 306, + 688, + 545, + 715 + ], + "type": "inline_equation", + "content": "-\\pmb{\\mu}^{\\top}A\\pmb{\\mu} + \\pmb{b}^{\\top}\\pmb{\\mu}-" + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3866" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 71, + 287, + 109 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 71, + 287, + 109 + ], + "spans": [ + { + "bbox": [ + 47, + 71, + 287, + 109 + ], + "type": "inline_equation", + "content": "c - \\operatorname {Tr}(A) + \\| 2A\\pmb {\\mu} - \\pmb {b}\\| < 0" + }, + { + "bbox": [ + 47, + 71, + 287, + 109 + ], + "type": "text", + "content": " we have " + }, + { + "bbox": [ + 47, + 71, + 287, + 109 + ], + "type": "inline_equation", + "content": "\\tau_{\\mathcal{D}}(\\tilde{P})\\leq \\frac{1}{2} (1 + p_2) < 1." + }, + { + "bbox": [ + 47, + 71, + 287, + 109 + ], + "type": "text", + "content": " Moreover, for any " + }, + { + "bbox": [ + 47, + 71, + 287, + 109 + ], + "type": "inline_equation", + "content": "\\pmb {\\mu}\\neq \\mathbf{0}" + }, + { + "bbox": [ + 47, + 71, + 287, + 109 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 71, + 287, + 109 + ], + "type": "inline_equation", + "content": "a_{-1}\\in [0,0.5]" + }, + { + "bbox": [ + 47, + 71, + 287, + 109 + ], + "type": "text", + "content": " , " + }, + { + "bbox": [ + 47, + 71, + 287, + 109 + ], + "type": "inline_equation", + "content": "\\exists a_{1}" + }, + { + "bbox": [ + 47, + 71, + 287, + 109 + ], + "type": "text", + "content": " such that " + }, + { + "bbox": [ + 47, + 71, + 287, + 109 + ], + "type": "inline_equation", + "content": "\\tau_{\\mathcal{D}}(\\tilde{P}) < \\tau_{\\mathcal{D}}(P)" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 118, + 289, + 334 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 118, + 289, + 334 + ], + "spans": [ + { + "bbox": [ + 46, + 118, + 289, + 334 + ], + "type": "text", + "content": "Poisoning is effective only if the accuracy of the unlearnable model " + }, + { + "bbox": [ + 46, + 118, + 289, + 334 + ], + "type": "inline_equation", + "content": "\\tilde{P}" + }, + { + "bbox": [ + 46, + 118, + 289, + 334 + ], + "type": "text", + "content": " is less than that of the clean model " + }, + { + "bbox": [ + 46, + 118, + 289, + 334 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 46, + 118, + 289, + 334 + ], + "type": "text", + "content": " on the clean dataset " + }, + { + "bbox": [ + 46, + 118, + 289, + 334 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 46, + 118, + 289, + 334 + ], + "type": "text", + "content": ", that is, " + }, + { + "bbox": [ + 46, + 118, + 289, + 334 + ], + "type": "inline_equation", + "content": "\\tau_{\\mathcal{D}}(\\tilde{P}) < \\tau_{\\mathcal{D}}(P)" + }, + { + "bbox": [ + 46, + 118, + 289, + 334 + ], + "type": "text", + "content": ". To satisfy this condition, we need to carefully select " + }, + { + "bbox": [ + 46, + 118, + 289, + 334 + ], + "type": "inline_equation", + "content": "a_{y}" + }, + { + "bbox": [ + 46, + 118, + 289, + 334 + ], + "type": "text", + "content": "'s. In Theorem 2, we formally state this condition1. Theorem 2 shows that CUDA can effectively poison when there are two distinct modes in the clean Gaussian mixture data model. We validate our theoretical claim through empirical analysis as well (see Figure 3). Details for the analysis is given in supplementary material. We find that our upper bound for the clean test accuracy of CUDA classifiers are consistent with our empirical analysis. In our experiments, we also find that the unlearnability effect is stronger with a larger blur parameter. This effect is evident from Figure 3 where we find that it is likely to get a lower " + }, + { + "bbox": [ + 46, + 118, + 289, + 334 + ], + "type": "inline_equation", + "content": "\\tau_{\\mathcal{D}}(\\tilde{P})" + }, + { + "bbox": [ + 46, + 118, + 289, + 334 + ], + "type": "text", + "content": " with higher " + }, + { + "bbox": [ + 46, + 118, + 289, + 334 + ], + "type": "inline_equation", + "content": "a_{y}" + }, + { + "bbox": [ + 46, + 118, + 289, + 334 + ], + "type": "text", + "content": " values. These results are consistent with our experimental results in Section 5 with CIFAR-10, CIFAR-100, and ImageNet-100 datasets." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 346, + 128, + 360 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 346, + 128, + 360 + ], + "spans": [ + { + "bbox": [ + 47, + 346, + 128, + 360 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 366, + 289, + 534 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 366, + 289, + 534 + ], + "spans": [ + { + "bbox": [ + 46, + 366, + 289, + 534 + ], + "type": "text", + "content": "In this section, we first discuss our experimental setup. More details on the setup is deferred to supplementary material. We then show the robustness of CUDA generation with various datasets and architectures. We also run various experiments to analyze the effectiveness of CUDA under different training techniques (ERM, AT with varying budgets, randomized smoothing, transfer learning, and finetuning) and augmentation techniques (mixup [56], grayscale, random blurring, cutout [9], cutmix [54], autoaugment [7], and orthogonal regularization [3]). Finally, we also design adaptive defenses to test the robustness of CUDA. One might think that CUDA filters can be obtained by adversarially training them with the data. We show that CUDA is robust to such adaptive defenses that we design." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 541, + 162, + 555 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 541, + 162, + 555 + ], + "spans": [ + { + "bbox": [ + 47, + 541, + 162, + 555 + ], + "type": "text", + "content": "5.1. Experimental setup" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 560, + 287, + 620 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 560, + 287, + 620 + ], + "spans": [ + { + "bbox": [ + 46, + 560, + 287, + 620 + ], + "type": "text", + "content": "Datasets. We use three image classification datasets - CIFAR-10, CIFAR-100, [21] and ImageNet-100 (a subset of ImageNet made of the first 100 classes) [39]. We use the data augmentation techniques such as random flipping, cropping, and resizing [43]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 620, + 287, + 657 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 620, + 287, + 657 + ], + "spans": [ + { + "bbox": [ + 46, + 620, + 287, + 657 + ], + "type": "text", + "content": "Architectures. We use ResNet-18 [14], VGG-16 [44], Wide ResNet-34-10 [55], and DenseNet-121 [16]. We train the networks with hyperparameters used in Fu et al. [11]." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 663, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 663, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 663, + 287, + 714 + ], + "type": "text", + "content": "1We note that the conditions " + }, + { + "bbox": [ + 46, + 663, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\pmb{\\mu}^{\\top}A\\pmb{\\mu} + \\pmb{b}^{\\top}\\pmb{\\mu} + c + \\mathrm{Tr}(A) + \\|2A\\pmb{\\mu} + \\pmb{b}\\|_{2} < 0" + }, + { + "bbox": [ + 46, + 663, + 287, + 714 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 46, + 663, + 287, + 714 + ], + "type": "inline_equation", + "content": "-\\pmb{\\mu}^{\\top}A\\pmb{\\mu} + \\pmb{b}^{\\top}\\pmb{\\mu} - c - \\mathrm{Tr}(A) + \\|2A\\pmb{\\mu} - \\pmb{b}\\|_{2} < 0" + }, + { + "bbox": [ + 46, + 663, + 287, + 714 + ], + "type": "text", + "content": " can always be satisfied by picking a sufficiently large " + }, + { + "bbox": [ + 46, + 663, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\pmb{\\mu}" + }, + { + "bbox": [ + 46, + 663, + 287, + 714 + ], + "type": "text", + "content": " in the direction of an eigenvector corresponding to a negative or positive eigenvalue of " + }, + { + "bbox": [ + 46, + 663, + 287, + 714 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 46, + 663, + 287, + 714 + ], + "type": "text", + "content": " (note that " + }, + { + "bbox": [ + 46, + 663, + 287, + 714 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 46, + 663, + 287, + 714 + ], + "type": "text", + "content": " has negative and positive eigenvalues)." + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 310, + 102, + 543, + 192 + ], + "blocks": [ + { + "bbox": [ + 305, + 70, + 545, + 93 + ], + "lines": [ + { + "bbox": [ + 305, + 70, + 545, + 93 + ], + "spans": [ + { + "bbox": [ + 305, + 70, + 545, + 93 + ], + "type": "text", + "content": "Table 2. Test accuracy (\\%) of ResNet-18 trained on various unlearnable datasets. We use " + }, + { + "bbox": [ + 305, + 70, + 545, + 93 + ], + "type": "inline_equation", + "content": "L_{\\infty}" + }, + { + "bbox": [ + 305, + 70, + 545, + 93 + ], + "type": "text", + "content": " AT with budget " + }, + { + "bbox": [ + 305, + 70, + 545, + 93 + ], + "type": "inline_equation", + "content": "\\rho_{a} = 4 / 255" + }, + { + "bbox": [ + 305, + 70, + 545, + 93 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 310, + 102, + 543, + 192 + ], + "lines": [ + { + "bbox": [ + 310, + 102, + 543, + 192 + ], + "spans": [ + { + "bbox": [ + 310, + 102, + 543, + 192 + ], + "type": "table", + "html": "
DatasetTraining methodCleanUnlearnability method
EMTAPNTGAREMCUDA
CIFAR-10ERM94.6613.2022.5116.2727.0918.48
AT89.5188.6288.0288.9648.1644.40
CIFAR-100ERM76.271.6013.753.2210.1412.69
AT64.5063.4362.3962.4427.1034.34
ImageNet-100ERM80.661.269.108.4213.748.96
AT66.6263.4063.5663.0641.6638.68
", + "image_path": "82f8998bf785a0ace91873b717e04a2a437310fa1bdbc7fb6dbdce80dfbdaa8f.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 310, + 235, + 543, + 315 + ], + "blocks": [ + { + "bbox": [ + 305, + 204, + 545, + 227 + ], + "lines": [ + { + "bbox": [ + 305, + 204, + 545, + 227 + ], + "spans": [ + { + "bbox": [ + 305, + 204, + 545, + 227 + ], + "type": "text", + "content": "Table 3. Test accuracy " + }, + { + "bbox": [ + 305, + 204, + 545, + 227 + ], + "type": "inline_equation", + "content": "(\\%)" + }, + { + "bbox": [ + 305, + 204, + 545, + 227 + ], + "type": "text", + "content": " of network architectures trained on various CIFAR-10 unlearnable datasets with " + }, + { + "bbox": [ + 305, + 204, + 545, + 227 + ], + "type": "inline_equation", + "content": "L_{\\infty}" + }, + { + "bbox": [ + 305, + 204, + 545, + 227 + ], + "type": "text", + "content": " AT " + }, + { + "bbox": [ + 305, + 204, + 545, + 227 + ], + "type": "inline_equation", + "content": "(\\rho_{a} = 4 / 255)" + }, + { + "bbox": [ + 305, + 204, + 545, + 227 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 310, + 235, + 543, + 315 + ], + "lines": [ + { + "bbox": [ + 310, + 235, + 543, + 315 + ], + "spans": [ + { + "bbox": [ + 310, + 235, + 543, + 315 + ], + "type": "table", + "html": "
ModelCleanUnlearnability method
EMTAPNTGAREMCUDA
ResNet-1889.5188.6288.0288.9648.1644.40
VGG-1687.5186.4886.2786.6565.2342.98
Wide ResNet-34-1091.2190.0590.2389.9548.3953.02
DenseNet-12183.2782.4481.7280.7381.4845.95
", + "image_path": "12a453c77b2ec42bae866f62683c3fab0a805612482ff4be802fb3133dd74ad5.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 335, + 545, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 335, + 545, + 384 + ], + "spans": [ + { + "bbox": [ + 305, + 335, + 545, + 384 + ], + "type": "text", + "content": "Previous works mainly employ ResNet-18 for most of their evaluations. Additional experiments on Tiny-ImageNet [26], DeIT [48], EfficientNetV2-S [46], and MobileNetV2 [40] are provided in the supplementary material." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 384, + 546, + 516 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 384, + 546, + 516 + ], + "spans": [ + { + "bbox": [ + 304, + 384, + 546, + 516 + ], + "type": "text", + "content": "CUDA generation. We use filters of size " + }, + { + "bbox": [ + 304, + 384, + 546, + 516 + ], + "type": "inline_equation", + "content": "k = 3" + }, + { + "bbox": [ + 304, + 384, + 546, + 516 + ], + "type": "text", + "content": " and blur parameter " + }, + { + "bbox": [ + 304, + 384, + 546, + 516 + ], + "type": "inline_equation", + "content": "p_b = 0.3" + }, + { + "bbox": [ + 304, + 384, + 546, + 516 + ], + "type": "text", + "content": " for both CIFAR-10 and CIFAR-100 datasets. ImageNet-100 is a higher dimensional " + }, + { + "bbox": [ + 304, + 384, + 546, + 516 + ], + "type": "inline_equation", + "content": "224 \\times 224 \\times 3" + }, + { + "bbox": [ + 304, + 384, + 546, + 516 + ], + "type": "text", + "content": " image dataset when compared to the " + }, + { + "bbox": [ + 304, + 384, + 546, + 516 + ], + "type": "inline_equation", + "content": "32 \\times 32 \\times 3" + }, + { + "bbox": [ + 304, + 384, + 546, + 516 + ], + "type": "text", + "content": " dimensional CIFAR datasets. Hence, we use larger filters of size " + }, + { + "bbox": [ + 304, + 384, + 546, + 516 + ], + "type": "inline_equation", + "content": "k = 9" + }, + { + "bbox": [ + 304, + 384, + 546, + 516 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 304, + 384, + 546, + 516 + ], + "type": "inline_equation", + "content": "p_b = 0.06" + }, + { + "bbox": [ + 304, + 384, + 546, + 516 + ], + "type": "text", + "content": " for ImageNet-100. These hyperparameters are chosen such that the CUDA images are not perceptibly highly perturbed and give good unlearnability effect (see plot in Figure 1). In supplementary material, we show the results of training on CUDA with different hyperparameters for data generation." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 517, + 547, + 661 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 517, + 547, + 661 + ], + "spans": [ + { + "bbox": [ + 304, + 517, + 547, + 661 + ], + "type": "text", + "content": "Baselines. We compare CUDA generation technique with four state-of-the-art unlearnability methods - REM [11], EM [17], TAP [10], and NTGA [53]. We adopt the results reported in [11] since we use the same hyperparameters for training. For REM, we select hyperparameters " + }, + { + "bbox": [ + 304, + 517, + 547, + 661 + ], + "type": "inline_equation", + "content": "\\rho_{u} = 8 / 255" + }, + { + "bbox": [ + 304, + 517, + 547, + 661 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 517, + 547, + 661 + ], + "type": "inline_equation", + "content": "\\rho_{a} = 4 / 255" + }, + { + "bbox": [ + 304, + 517, + 547, + 661 + ], + "type": "text", + "content": ", the highest radii values in [11]. For comparing unlearnable methods, we look at the clean test accuracy. The lower the test accuracy, the better the unlearnability method is. As mentioned in the supplementary material, we use publicly released official codebases for reproducing the baselines using their default hyperparameters." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 670, + 435, + 682 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 670, + 435, + 682 + ], + "spans": [ + { + "bbox": [ + 306, + 670, + 435, + 682 + ], + "type": "text", + "content": "5.2. Effectiveness of CUDA" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 689, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 545, + 714 + ], + "type": "text", + "content": "Why does CUDA work? In order to measure how much CUDA technique's blurring affects the dataset's quality, we" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3867" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 49, + 102, + 286, + 253 + ], + "blocks": [ + { + "bbox": [ + 47, + 70, + 287, + 93 + ], + "lines": [ + { + "bbox": [ + 47, + 70, + 287, + 93 + ], + "spans": [ + { + "bbox": [ + 47, + 70, + 287, + 93 + ], + "type": "text", + "content": "Table 4. Test accuracy " + }, + { + "bbox": [ + 47, + 70, + 287, + 93 + ], + "type": "inline_equation", + "content": "(\\%)" + }, + { + "bbox": [ + 47, + 70, + 287, + 93 + ], + "type": "text", + "content": " of ResNet-18 with CUDA under various training settings." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 49, + 102, + 286, + 253 + ], + "lines": [ + { + "bbox": [ + 49, + 102, + 286, + 253 + ], + "spans": [ + { + "bbox": [ + 49, + 102, + 286, + 253 + ], + "type": "table", + "html": "
DatasetCleanTraining methodCUDA (ours)
CIFAR-1094.66ERM18.48
AT L∞ (ρa = 4/255)44.40
AT L∞ (ρa = 8/255)32.85
AT L∞ (ρa = 16/255)19.32
AT L2 (ρa = 0.25)39.05
AT L2 (ρa = 0.50)51.19
AT L2 (ρa = 0.75)51.14
CIFAR-10076.27ERM12.69
AT L∞ (ρa = 4/255)34.34
AT L∞ (ρa = 8/255)30.00
AT L2 (ρa = 0.75)36.90
ImageNet-10080.66ERM8.96
AT L∞ (ρa = 4/255)38.68
AT L∞ (ρa = 8/255)40.08
AT L2 (ρa = 0.75)20.58
", + "image_path": "b4c64526cc6177916ff8a34f965b400fc170077a1092d9a7defbc2aee0944cc3.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 278, + 289, + 686 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 278, + 289, + 686 + ], + "spans": [ + { + "bbox": [ + 46, + 278, + 289, + 686 + ], + "type": "text", + "content": "compare class-wise blurring (CUDA technique) against universal blurring where a single convolutional filter is used for blurring all the images. We keep the filter generation parameters fixed (" + }, + { + "bbox": [ + 46, + 278, + 289, + 686 + ], + "type": "inline_equation", + "content": "p_b = 0.3" + }, + { + "bbox": [ + 46, + 278, + 289, + 686 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 278, + 289, + 686 + ], + "type": "inline_equation", + "content": "k = 3" + }, + { + "bbox": [ + 46, + 278, + 289, + 686 + ], + "type": "text", + "content": ") for both CUDA class-wise blurring and universal blurring. ResNet-18 trained with clean CIFAR-10, universally blurred CIFAR-10, and CIFAR-10 CUDA achieve clean test accuracies of " + }, + { + "bbox": [ + 46, + 278, + 289, + 686 + ], + "type": "inline_equation", + "content": "94.66\\%" + }, + { + "bbox": [ + 46, + 278, + 289, + 686 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 278, + 289, + 686 + ], + "type": "inline_equation", + "content": "90.47\\%" + }, + { + "bbox": [ + 46, + 278, + 289, + 686 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 278, + 289, + 686 + ], + "type": "inline_equation", + "content": "18.48\\%" + }, + { + "bbox": [ + 46, + 278, + 289, + 686 + ], + "type": "text", + "content": ", respectively. This suggests that our controlled blurring does not obscure the semantics of the dataset. Hence, the significant drop in the clean test accuracy introduced by CUDA is most likely due to the usage of class-wise filters. This suggests that a model trained on CUDA learns the relation between the class-wise convolution filters and their corresponding labels. Therefore, during test time when this convolution effect is absent, the CUDA trained model fails to make correct classifications. Furthermore, the model trained on CUDA achieves an accuracy of " + }, + { + "bbox": [ + 46, + 278, + 289, + 686 + ], + "type": "inline_equation", + "content": "99.91\\%" + }, + { + "bbox": [ + 46, + 278, + 289, + 686 + ], + "type": "text", + "content": " on the CIFAR-10 CUDA testset. This strongly supports our claim that the CUDA model learns to classify images based on the convolutional filters used to blur them. In addition, if we permute the class-wise filters for blurring the test set (i.e., blurring class 1 images with class 2 filters, class 2 images with class 3 filters, and so on), we get a very low accuracy of " + }, + { + "bbox": [ + 46, + 278, + 289, + 686 + ], + "type": "inline_equation", + "content": "2.53\\%" + }, + { + "bbox": [ + 46, + 278, + 289, + 686 + ], + "type": "text", + "content": " on this test set. Further details are provided in the supplementary material. Finally, we note that real-world datasets might also contain blurred images due to various factors such as motion blurring, weather conditions, issues with the camera, etc. Hence, detecting if a blurred image is poisoned might not always be possible. However, one might argue that it is possible to detect if an entire dataset is blurred. Interestingly, later in this section we show that CUDA technique exhibits unlearnability effect even when only a fraction of the training dataset is poisoned." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 689, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 288, + 715 + ], + "type": "text", + "content": "Different datasets. We first compare the effectiveness of CUDA with different datasets using ERM and " + }, + { + "bbox": [ + 47, + 689, + 288, + 715 + ], + "type": "inline_equation", + "content": "L_{\\infty}" + }, + { + "bbox": [ + 47, + 689, + 288, + 715 + ], + "type": "text", + "content": " AT with" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 304, + 72, + 545, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 180 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 180 + ], + "type": "inline_equation", + "content": "\\rho_{a} = 4 / 255" + }, + { + "bbox": [ + 304, + 72, + 545, + 180 + ], + "type": "text", + "content": ". We use ResNet-18 for the experiments. The results are shown in Table 2. These results show that EM, TAP, and NTGA are not robust to AT. However, both CUDA and REM are successful. Here, our method CUDA outperforms REM with CIFAR-10 and ImageNet-100 datasets. Smartly designed additive noise in AT helps in achieving better generalization than ERM on the unlearnable datasets. This experiment thus demonstrates that ERM and AT are not good choices for training with CUDA and REM dataset." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 182, + 545, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 182, + 545, + 266 + ], + "spans": [ + { + "bbox": [ + 304, + 182, + 545, + 266 + ], + "type": "text", + "content": "Different models. Next we compare the effectiveness of CUDA using various deep learning architectures with " + }, + { + "bbox": [ + 304, + 182, + 545, + 266 + ], + "type": "inline_equation", + "content": "L_{\\infty}" + }, + { + "bbox": [ + 304, + 182, + 545, + 266 + ], + "type": "text", + "content": " AT (" + }, + { + "bbox": [ + 304, + 182, + 545, + 266 + ], + "type": "inline_equation", + "content": "\\rho_{a} = 4 / 255" + }, + { + "bbox": [ + 304, + 182, + 545, + 266 + ], + "type": "text", + "content": "). We use CIFAR-10 for the experiments. The results are shown in Table 3. As we see in the table, CUDA is effective with all the five network architectures. However, REM is not seen to be transferrable with DenseNet-121." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 267, + 546, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 267, + 546, + 483 + ], + "spans": [ + { + "bbox": [ + 304, + 267, + 546, + 483 + ], + "type": "text", + "content": "Robustness to different training settings. In Section 3.2, we show that REM is sensitive to the training settings. REM generated using " + }, + { + "bbox": [ + 304, + 267, + 546, + 483 + ], + "type": "inline_equation", + "content": "L_{\\infty}" + }, + { + "bbox": [ + 304, + 267, + 546, + 483 + ], + "type": "text", + "content": " radii budgets of " + }, + { + "bbox": [ + 304, + 267, + 546, + 483 + ], + "type": "inline_equation", + "content": "\\rho_u = 8 / 255" + }, + { + "bbox": [ + 304, + 267, + 546, + 483 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 267, + 546, + 483 + ], + "type": "inline_equation", + "content": "\\rho_a = 4 / 255" + }, + { + "bbox": [ + 304, + 267, + 546, + 483 + ], + "type": "text", + "content": " for CIFAR-10 breaks with " + }, + { + "bbox": [ + 304, + 267, + 546, + 483 + ], + "type": "inline_equation", + "content": "L_{\\infty}" + }, + { + "bbox": [ + 304, + 267, + 546, + 483 + ], + "type": "text", + "content": " AT " + }, + { + "bbox": [ + 304, + 267, + 546, + 483 + ], + "type": "inline_equation", + "content": "(\\rho_a = 8 / 255)" + }, + { + "bbox": [ + 304, + 267, + 546, + 483 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 267, + 546, + 483 + ], + "type": "inline_equation", + "content": "L_2" + }, + { + "bbox": [ + 304, + 267, + 546, + 483 + ], + "type": "text", + "content": " AT " + }, + { + "bbox": [ + 304, + 267, + 546, + 483 + ], + "type": "inline_equation", + "content": "(\\rho_a = 0.75)" + }, + { + "bbox": [ + 304, + 267, + 546, + 483 + ], + "type": "text", + "content": " to get test accuracy of " + }, + { + "bbox": [ + 304, + 267, + 546, + 483 + ], + "type": "inline_equation", + "content": "78.71\\%" + }, + { + "bbox": [ + 304, + 267, + 546, + 483 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 267, + 546, + 483 + ], + "type": "inline_equation", + "content": "79.65\\%" + }, + { + "bbox": [ + 304, + 267, + 546, + 483 + ], + "type": "text", + "content": ", respectively. Hence, we run experiments to check the robustness of CUDA with various AT norm budgets. The results are shown in Table 4. As we see in the table, CUDA is robust to ERM, " + }, + { + "bbox": [ + 304, + 267, + 546, + 483 + ], + "type": "inline_equation", + "content": "L_{\\infty}" + }, + { + "bbox": [ + 304, + 267, + 546, + 483 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 267, + 546, + 483 + ], + "type": "inline_equation", + "content": "L_2" + }, + { + "bbox": [ + 304, + 267, + 546, + 483 + ], + "type": "text", + "content": " AT settings with varying training budgets. Impressively, the highest test accuracy achieved with training on CIFAR-10 CUDA, CIFAR-100 CUDA, and ImageNet-100 CUDA are as low as " + }, + { + "bbox": [ + 304, + 267, + 546, + 483 + ], + "type": "inline_equation", + "content": "51.19\\%" + }, + { + "bbox": [ + 304, + 267, + 546, + 483 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 267, + 546, + 483 + ], + "type": "inline_equation", + "content": "36.90\\%" + }, + { + "bbox": [ + 304, + 267, + 546, + 483 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 267, + 546, + 483 + ], + "type": "inline_equation", + "content": "40.08\\%" + }, + { + "bbox": [ + 304, + 267, + 546, + 483 + ], + "type": "text", + "content": ", respectively. We also find that using a pre-trained ResNet-18 with CIFAR-10 CUDA only achieves clean test accuracy of " + }, + { + "bbox": [ + 304, + 267, + 546, + 483 + ], + "type": "inline_equation", + "content": "42.42\\%" + }, + { + "bbox": [ + 304, + 267, + 546, + 483 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 267, + 546, + 483 + ], + "type": "inline_equation", + "content": "48.22\\%" + }, + { + "bbox": [ + 304, + 267, + 546, + 483 + ], + "type": "text", + "content": " with fine-tuning the full network and a newly trained final layer, respectively (details are deferred to the supplementary material)." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 484, + 545, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 484, + 545, + 581 + ], + "spans": [ + { + "bbox": [ + 304, + 484, + 545, + 581 + ], + "type": "text", + "content": "Different protection percentages. In Section 3.1, we assume that the attacker has access to the full clean training data. However, in real life settings, this might not be always possible. Hence, we train ResNet-18 on a mix of CIFAR-10 CUDA and clean CIFAR-10 training datasets to evaluate the effectiveness of poisoning with varying data protection percentages. Protection percentage denotes the percentage of the training data that is poisoned." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 582, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 582, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 582, + 547, + 715 + ], + "type": "text", + "content": "We show the results in Table 5. In the table, the \"Mixed\" column denotes the clean test accuracy of a model trained using a mix of both clean and poisoned data. The \"Clean\" column denotes the clean test accuracy of a model trained only using the clean subset of the training data. In the last row of Table 5, we provide the results for CUDA trained using ERM with varying data protection percentages. The remainder of the rows provide the results for the " + }, + { + "bbox": [ + 304, + 582, + 547, + 715 + ], + "type": "inline_equation", + "content": "L_{\\infty}" + }, + { + "bbox": [ + 304, + 582, + 547, + 715 + ], + "type": "text", + "content": " AT " + }, + { + "bbox": [ + 304, + 582, + 547, + 715 + ], + "type": "inline_equation", + "content": "(\\rho_{a} = 4 / 255)" + }, + { + "bbox": [ + 304, + 582, + 547, + 715 + ], + "type": "text", + "content": " scenario. For example, CUDA trained with ERM using an " + }, + { + "bbox": [ + 304, + 582, + 547, + 715 + ], + "type": "inline_equation", + "content": "80\\%" + }, + { + "bbox": [ + 304, + 582, + 547, + 715 + ], + "type": "text", + "content": " clean training data partition achieves a test accuracy of " + }, + { + "bbox": [ + 304, + 582, + 547, + 715 + ], + "type": "inline_equation", + "content": "93.75\\%" + }, + { + "bbox": [ + 304, + 582, + 547, + 715 + ], + "type": "text", + "content": ". Adding the " + }, + { + "bbox": [ + 304, + 582, + 547, + 715 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 304, + 582, + 547, + 715 + ], + "type": "text", + "content": " CUDA data" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3868" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 105, + 101, + 492, + 213 + ], + "blocks": [ + { + "bbox": [ + 46, + 70, + 547, + 95 + ], + "lines": [ + { + "bbox": [ + 46, + 70, + 547, + 95 + ], + "spans": [ + { + "bbox": [ + 46, + 70, + 547, + 95 + ], + "type": "text", + "content": "Table 5. Test accuracy (\\%) of ResNet-18 on CIFAR-10 with different data protection percentages. The last row shows the results for CUDA with ERM setting. The rest of the rows show results for unlearnability methods trained in the " + }, + { + "bbox": [ + 46, + 70, + 547, + 95 + ], + "type": "inline_equation", + "content": "L_{\\infty}" + }, + { + "bbox": [ + 46, + 70, + 547, + 95 + ], + "type": "text", + "content": " AT (" + }, + { + "bbox": [ + 46, + 70, + 547, + 95 + ], + "type": "inline_equation", + "content": "\\rho_a = 4/255" + }, + { + "bbox": [ + 46, + 70, + 547, + 95 + ], + "type": "text", + "content": ") setting." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 105, + 101, + 492, + 213 + ], + "lines": [ + { + "bbox": [ + 105, + 101, + 492, + 213 + ], + "spans": [ + { + "bbox": [ + 105, + 101, + 492, + 213 + ], + "type": "table", + "html": "
Unlearnability methodData Protection Percentage
0%20% MixedClean40% MixedClean60% MixedClean80% MixedClean100%
EM89.5189.6088.1789.4086.7689.4985.0789.1079.4188.62
TAP89.0188.6688.4088.0488.02
NTGA89.5689.3589.2289.1788.96
REM89.6089.3489.6188.0948.16
CUDA (ours)88.5487.2486.0384.3444.40
CUDA + ERM94.6693.2893.7591.3492.5689.9189.7785.6184.3018.48
", + "image_path": "98e511d21bab45a25028cf7f17abcb863a07019773528faa978ae92d278c35a8.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 230, + 289, + 337 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 230, + 289, + 337 + ], + "spans": [ + { + "bbox": [ + 46, + 230, + 289, + 337 + ], + "type": "text", + "content": "partition to the training dataset drops the test accuracy of the model to " + }, + { + "bbox": [ + 46, + 230, + 289, + 337 + ], + "type": "inline_equation", + "content": "93.28\\%" + }, + { + "bbox": [ + 46, + 230, + 289, + 337 + ], + "type": "text", + "content": ". Results from the last row of Table 5 show that CUDA with varying data protection percentages is effective with the ERM setting. However, with the AT scenario, the unlearnability techniques are not as successful as with ERM with varying data protection percentages. Nevertheless, it is interesting to note that CUDA technique performs better than all the other unlearnability techniques with AT." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 339, + 289, + 543 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 339, + 289, + 543 + ], + "spans": [ + { + "bbox": [ + 46, + 339, + 289, + 543 + ], + "type": "text", + "content": "Robustness to smoothing and data augmentations. Cohen et al. [6] proposes randomized smoothing which is a provable adversarial defense in " + }, + { + "bbox": [ + 46, + 339, + 289, + 543 + ], + "type": "inline_equation", + "content": "L_{2}" + }, + { + "bbox": [ + 46, + 339, + 289, + 543 + ], + "type": "text", + "content": " norm. Since CUDA does not use additive noise, CUDA is robust to randomized smoothing. A smoothed ResNet-18 (with a noise level of 0.5) with CIFAR-10 CUDA achieves only a clean test accuracy of " + }, + { + "bbox": [ + 46, + 339, + 289, + 543 + ], + "type": "inline_equation", + "content": "43.85\\%" + }, + { + "bbox": [ + 46, + 339, + 289, + 543 + ], + "type": "text", + "content": ". In Section 3.2, we see that REM breaks with grayscaleing. While ResNet-18 training using grayscaleed REM achieves " + }, + { + "bbox": [ + 46, + 339, + 289, + 543 + ], + "type": "inline_equation", + "content": "70.76\\%" + }, + { + "bbox": [ + 46, + 339, + 289, + 543 + ], + "type": "text", + "content": " test accuracy, grayscaleed CUDA only achieves " + }, + { + "bbox": [ + 46, + 339, + 289, + 543 + ], + "type": "inline_equation", + "content": "20.12\\%" + }, + { + "bbox": [ + 46, + 339, + 289, + 543 + ], + "type": "text", + "content": " test accuracy on the clean grayscaleed CIFAR-10 test data. This shows that CUDA technique exhibits the desirable property of not relying upon the color space for its attack. CIFAR-10 CUDA training achieves only " + }, + { + "bbox": [ + 46, + 339, + 289, + 543 + ], + "type": "inline_equation", + "content": "25.53\\%" + }, + { + "bbox": [ + 46, + 339, + 289, + 543 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 339, + 289, + 543 + ], + "type": "inline_equation", + "content": "25.80\\%" + }, + { + "bbox": [ + 46, + 339, + 289, + 543 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 339, + 289, + 543 + ], + "type": "inline_equation", + "content": "26.93\\%" + }, + { + "bbox": [ + 46, + 339, + 289, + 543 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 339, + 289, + 543 + ], + "type": "inline_equation", + "content": "34.09\\%" + }, + { + "bbox": [ + 46, + 339, + 289, + 543 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 339, + 289, + 543 + ], + "type": "inline_equation", + "content": "50.72\\%" + }, + { + "bbox": [ + 46, + 339, + 289, + 543 + ], + "type": "text", + "content": " clean test accuracy with mixup, cutout, cutmix, autoaugment, and orthogonal regularization, respectively (see supplementary material for more details)." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 544, + 288, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 544, + 288, + 651 + ], + "spans": [ + { + "bbox": [ + 46, + 544, + 288, + 651 + ], + "type": "text", + "content": "Adaptive defenses for CUDA. Here, we first investigate the effect of training CIFAR-10 CUDA with random blurring augmentations using ResNet-18. Each batch of the CUDA training data is convolved with random " + }, + { + "bbox": [ + 46, + 544, + 288, + 651 + ], + "type": "inline_equation", + "content": "3 \\times 3" + }, + { + "bbox": [ + 46, + 544, + 288, + 651 + ], + "type": "text", + "content": " filters of varying blur parameters " + }, + { + "bbox": [ + 46, + 544, + 288, + 651 + ], + "type": "inline_equation", + "content": "p_b'" + }, + { + "bbox": [ + 46, + 544, + 288, + 651 + ], + "type": "text", + "content": ". With " + }, + { + "bbox": [ + 46, + 544, + 288, + 651 + ], + "type": "inline_equation", + "content": "p_b'" + }, + { + "bbox": [ + 46, + 544, + 288, + 651 + ], + "type": "text", + "content": " values of 0.1 and 0.3, CUDA training achieves lower test accuracy " + }, + { + "bbox": [ + 46, + 544, + 288, + 651 + ], + "type": "inline_equation", + "content": "9.2\\%" + }, + { + "bbox": [ + 46, + 544, + 288, + 651 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 544, + 288, + 651 + ], + "type": "inline_equation", + "content": "13.37\\%" + }, + { + "bbox": [ + 46, + 544, + 288, + 651 + ], + "type": "text", + "content": ", respectively. This shows that the noise added by CUDA technique is robust to random convolution augmentations." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "type": "text", + "content": "One may think that CUDA technique can be broken by learning the private filters from the data. We test this idea by training deconvolution filters to find if we can reverse the blurring effect in CUDA with adversarially trained filters. We use a novel Deconvolution-based Adversarial Training" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 230, + 547, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 230, + 547, + 411 + ], + "spans": [ + { + "bbox": [ + 304, + 230, + 547, + 411 + ], + "type": "text", + "content": "(DAT) technique that is similar to AT (check supplementary material for details). While the adversarial step in AT learns sample-wise error-maximizing additive noises, the adversarial step in DAT learns class-wise error-maximizing deconvolution filters. We train DAT with filters of varying sizes (3, 5, and 7) on CIFAR-10 CUDA using ResNet-18. The filter parameters are constrained within a finite range to make sure that the images do not get distorted with the adversarial step similar to projection in projected gradient descent. We find that CUDA is robust to DAT. DAT using filters of size 3, 5, and 7 with CUDA achieves only test accuracy of " + }, + { + "bbox": [ + 304, + 230, + 547, + 411 + ], + "type": "inline_equation", + "content": "39.05\\%" + }, + { + "bbox": [ + 304, + 230, + 547, + 411 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 230, + 547, + 411 + ], + "type": "inline_equation", + "content": "46.21\\%" + }, + { + "bbox": [ + 304, + 230, + 547, + 411 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 230, + 547, + 411 + ], + "type": "inline_equation", + "content": "38.48\\%" + }, + { + "bbox": [ + 304, + 230, + 547, + 411 + ], + "type": "text", + "content": ", respectively. DAT is not successful against CUDA since we can not invert convolutions without the knowledge of the private filters or clean images corresponding to CUDA." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 414, + 547, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 414, + 547, + 594 + ], + "spans": [ + { + "bbox": [ + 304, + 414, + 547, + 594 + ], + "type": "text", + "content": "Limitations and future directions. The unlearnability effect of CUDA can be defended if some fraction of the clean data and its corresponding CUDA images are leaked. The defender must also be able to detect if all samples in the dataset are poisoned. However, our work assumes a setup where the filters remain private. For example, a data publisher could simply publish their CUDA images and delete the clean images permanently to prevent this scenario. As discussed in this section, CUDA as well as other prior works do not perform well with different protection percentages with AT. Improving this can be an interesting research direction. We believe that extending CUDA technique to other domains such as tabular and text data is also an interesting future direction. It would also be interesting to see theoretical analysis of CUDA considering more complex setups." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 617, + 404, + 632 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 617, + 404, + 632 + ], + "spans": [ + { + "bbox": [ + 306, + 617, + 404, + 632 + ], + "type": "text", + "content": "Acknowledgement" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 642, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 642, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 642, + 547, + 713 + ], + "type": "text", + "content": "This project was supported in part by Meta grant 23010098, NSF CAREER AWARD 1942230, HR001119S0026 (GARD), ONR YIP award N00014-22-1-2271, Army Grant No. W911NF2120076, a capital one grant, NIST 60NANB20D134 and the NSF award CCF2212458." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3869" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 135 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 135 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 135 + ], + "type": "text", + "content": "[1] Mauro Barni, Kassem Kallas, and Benedetta Tondi. A new backdoor attack in cnns by training set corruption without label poisoning. In 2019 IEEE International Conference on Image Processing (ICIP), pages 101-105. IEEE, 2019. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 137, + 288, + 169 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 137, + 288, + 169 + ], + "spans": [ + { + "bbox": [ + 53, + 137, + 288, + 169 + ], + "type": "text", + "content": "[2] Battista Biggio, Blaine Nelson, and Pavel Laskov. Poisoning attacks against support vector machines. arXiv preprint arXiv:1206.6389, 2012. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 171, + 288, + 203 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 171, + 288, + 203 + ], + "spans": [ + { + "bbox": [ + 53, + 171, + 288, + 203 + ], + "type": "text", + "content": "[3] Andrew Brock, Theodore Lim, James M Ritchie, and Nick Weston. Neural photo editing with introspective adversarial networks. arXiv preprint arXiv:1609.07093, 2016. 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 205, + 288, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 205, + 288, + 237 + ], + "spans": [ + { + "bbox": [ + 53, + 205, + 288, + 237 + ], + "type": "text", + "content": "[4] Nicholas Carlini and David Wagner. Towards evaluating the robustness of neural networks. In 2017 IEEE symposium on security and privacy (sp), pages 39-57. IEEE, 2017. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 239, + 288, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 239, + 288, + 282 + ], + "spans": [ + { + "bbox": [ + 53, + 239, + 288, + 282 + ], + "type": "text", + "content": "[5] Xinyun Chen, Chang Liu, Bo Li, Kimberly Lu, and Dawn Song. Targeted backdoor attacks on deep learning systems using data poisoning. arXiv preprint arXiv:1712.05526, 2017.3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 285, + 288, + 350 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 285, + 288, + 350 + ], + "spans": [ + { + "bbox": [ + 53, + 285, + 288, + 350 + ], + "type": "text", + "content": "[6] Jeremy Cohen, Elan Rosenfeld, and Zico Kolter. Certified adversarial robustness via randomized smoothing. In Kamalika Chaudhuri and Ruslan Salakhutdinov, editors, Proceedings of the 36th International Conference on Machine Learning, volume 97 of Proceedings of Machine Learning Research, pages 1310-1320. PMLR, 09-15 Jun 2019. 2, 8" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 351, + 288, + 394 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 351, + 288, + 394 + ], + "spans": [ + { + "bbox": [ + 53, + 351, + 288, + 394 + ], + "type": "text", + "content": "[7] Ekin D Cubuk, Barret Zoph, Dandelion Mane, Vijay Vasudevan, and Quoc V Le. Autoaugment: Learning augmentation policies from data. arXiv preprint arXiv:1805.09501, 2018. 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 397, + 288, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 397, + 288, + 441 + ], + "spans": [ + { + "bbox": [ + 53, + 397, + 288, + 441 + ], + "type": "text", + "content": "[8] Karan Desai and Justin Johnson. Virtex: Learning visual representations from textual annotations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11162-11173, 2021. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 442, + 288, + 474 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 442, + 288, + 474 + ], + "spans": [ + { + "bbox": [ + 53, + 442, + 288, + 474 + ], + "type": "text", + "content": "[9] Terrance DeVries and Graham W Taylor. Improved regularization of convolutional neural networks with cutout. arXiv preprint arXiv:1708.04552, 2017. 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 476, + 287, + 519 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 476, + 287, + 519 + ], + "spans": [ + { + "bbox": [ + 48, + 476, + 287, + 519 + ], + "type": "text", + "content": "[10] Liam Fowl, Micah Goldblum, Ping-yeh Chiang, Jonas Geiping, Wojtek Czaja, and Tom Goldstein. Adversarial examples make strong poisons. arXiv preprint arXiv:2106.10807, 2021. 1, 3, 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 521, + 287, + 565 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 521, + 287, + 565 + ], + "spans": [ + { + "bbox": [ + 48, + 521, + 287, + 565 + ], + "type": "text", + "content": "[11] Shaopeng Fu, Fengxiang He, Yang Liu, Li Shen, and Dacheng Tao. Robust unlearnable examples: Protecting data privacy against adversarial learning. In International Conference on Learning Representations, 2021. 1, 3, 4, 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 567, + 287, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 567, + 287, + 599 + ], + "spans": [ + { + "bbox": [ + 48, + 567, + 287, + 599 + ], + "type": "text", + "content": "[12] Ian J Goodfellow, Jonathon Shlens, and Christian Szegedy. Explaining and harnessing adversarial examples. arXiv preprint arXiv:1412.6572, 2014. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 601, + 287, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 601, + 287, + 643 + ], + "spans": [ + { + "bbox": [ + 48, + 601, + 287, + 643 + ], + "type": "text", + "content": "[13] Tianyu Gu, Brendan Dolan-Gavitt, and Siddharth Garg. Badnets: Identifying vulnerabilities in the machine learning model supply chain. arXiv preprint arXiv:1708.06733, 2017. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 646, + 287, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 646, + 287, + 678 + ], + "spans": [ + { + "bbox": [ + 48, + 646, + 287, + 678 + ], + "type": "text", + "content": "[14] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. CoRR, abs/1512.03385, 2015. 6" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 680, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 680, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 680, + 287, + 712 + ], + "type": "text", + "content": "[15] Kashmir Hill. The secretive company that might end privacy as we know it. In *Ethics of Data and Analytics*, pages 170–177. Auerbach Publications, 2020. 1" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 34, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 105 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 105 + ], + "type": "text", + "content": "[16] Gao Huang, Zhuang Liu, and Kilian Q. Weinberger. Densely connected convolutional networks. CoRR, abs/1608.06993, 2016. 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 107, + 545, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 107, + 545, + 150 + ], + "spans": [ + { + "bbox": [ + 307, + 107, + 545, + 150 + ], + "type": "text", + "content": "[17] Hanxun Huang, Xingjun Ma, Sarah Monazam Erfani, James Bailey, and Yisen Wang. Unlearnable examples: Making personal data unexploitable. arXiv preprint arXiv:2101.04898, 2021. 1, 3, 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 152, + 545, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 152, + 545, + 194 + ], + "spans": [ + { + "bbox": [ + 307, + 152, + 545, + 194 + ], + "type": "text", + "content": "[18] Arthur Jacot, Franck Gabriel, and Clément Hongler. Neural tangent kernel: Convergence and generalization in neural networks. Advances in neural information processing systems, 31, 2018. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 197, + 545, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 197, + 545, + 228 + ], + "spans": [ + { + "bbox": [ + 307, + 197, + 545, + 228 + ], + "type": "text", + "content": "[19] Adel Javanmard and Mahdi Soltanolkotabi. Precise statistical analysis of classification accuracies for adversarial training. ArXiv, abs/2010.11213, 2020. 2, 4" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 231, + 545, + 273 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 231, + 545, + 273 + ], + "spans": [ + { + "bbox": [ + 307, + 231, + 545, + 273 + ], + "type": "text", + "content": "[20] Pang Wei Koh and Percy Liang. Understanding black-box predictions via influence functions. In International conference on machine learning, pages 1885-1894. PMLR, 2017. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 275, + 545, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 275, + 545, + 297 + ], + "spans": [ + { + "bbox": [ + 307, + 275, + 545, + 297 + ], + "type": "text", + "content": "[21] Alex Krizhevsky. Learning multiple layers of features from tiny images. Technical report, 2009. 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 298, + 545, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 298, + 545, + 330 + ], + "spans": [ + { + "bbox": [ + 307, + 298, + 545, + 330 + ], + "type": "text", + "content": "[22] Aounon Kumar, Alexander Levine, Soheil Feizi, and Tom Goldstein. Certifying confidence via randomized smoothing. CoRR, abs/2009.08061, 2020. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 332, + 545, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 332, + 545, + 374 + ], + "spans": [ + { + "bbox": [ + 307, + 332, + 545, + 374 + ], + "type": "text", + "content": "[23] Alexey Kurakin, Ian J Goodfellow, and Samy Bengio. Adversarial examples in the physical world. In Artificial intelligence safety and security, pages 99-112. Chapman and Hall/CRC, 2018. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 376, + 545, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 376, + 545, + 430 + ], + "spans": [ + { + "bbox": [ + 307, + 376, + 545, + 430 + ], + "type": "text", + "content": "[24] Cassidy Laidlaw and Soheil Feizi. Functional adversarial attacks. In H. Wallach, H. Larochelle, A. Beygelzimer, F. d'Alché-Buc, E. Fox, and R. Garnett, editors, Advances in Neural Information Processing Systems, volume 32. Curran Associates, Inc., 2019. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 433, + 545, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 433, + 545, + 464 + ], + "spans": [ + { + "bbox": [ + 307, + 433, + 545, + 464 + ], + "type": "text", + "content": "[25] Cassidy Laidlaw, Sahil Singla, and Soheil Feizi. Perceptual adversarial robustness: Defense against unseen threat models. CoRR, abs/2006.12655, 2020. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 466, + 545, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 466, + 545, + 487 + ], + "spans": [ + { + "bbox": [ + 307, + 466, + 545, + 487 + ], + "type": "text", + "content": "[26] Ya Le and Xuan Yang. Tiny imagenet visual recognition challenge. CS 231N, 7(7):3, 2015. 6" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 489, + 545, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 489, + 545, + 521 + ], + "spans": [ + { + "bbox": [ + 307, + 489, + 545, + 521 + ], + "type": "text", + "content": "[27] Alexander Levine and Soheil Feizi. Deep partition aggregation: Provable defense against general poisoning attacks. CoRR, abs/2006.14768, 2020. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 523, + 545, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 523, + 545, + 555 + ], + "spans": [ + { + "bbox": [ + 307, + 523, + 545, + 555 + ], + "type": "text", + "content": "[28] Alexander Levine and Soheil Feizi. Deep partition aggregation: Provable defense against general poisoning attacks. CoRR, abs/2006.14768, 2020. 2" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 307, + 557, + 545, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 557, + 545, + 599 + ], + "spans": [ + { + "bbox": [ + 307, + 557, + 545, + 599 + ], + "type": "text", + "content": "[29] Yige Li, Xixiang Lyu, Nodens Koren, Lingjuan Lyu, Bo Li, and Xingjun Ma. Neural attention distillation: Erasing backdoor triggers from deep neural networks. arXiv preprint arXiv:2101.05930, 2021. 3" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 307, + 601, + 545, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 601, + 545, + 634 + ], + "spans": [ + { + "bbox": [ + 307, + 601, + 545, + 634 + ], + "type": "text", + "content": "[30] Yiming Li, Tongqing Zhai, Baoyuan Wu, Yong Jiang, Zhifeng Li, and Shutao Xia. Rethinking the trigger of backdoor attack. arXiv preprint arXiv:2004.04692, 2020. 3" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 307, + 636, + 545, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 636, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 307, + 636, + 545, + 689 + ], + "type": "text", + "content": "[31] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In European conference on computer vision, pages 740-755. Springer, 2014. 1" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 307, + 692, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 692, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 692, + 545, + 713 + ], + "type": "text", + "content": "[32] Sijia Liu, Songtao Lu, Xiangyi Chen, Yao Feng, Kaidi Xu, Abdullah Al-Dujaili, Mingyi Hong, and Una-May O'Reilly." + } + ] + } + ], + "index": 33 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3870" + } + ] + } + ], + "index": 35 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 66, + 72, + 287, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 72, + 287, + 116 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 287, + 116 + ], + "type": "text", + "content": "Min-max optimization without gradients: Convergence and applications to black-box evasion and poisoning attacks. In International Conference on Machine Learning, pages 6282-6293. PMLR, 2020. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 118, + 287, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 118, + 287, + 161 + ], + "spans": [ + { + "bbox": [ + 48, + 118, + 287, + 161 + ], + "type": "text", + "content": "[33] Yunfei Liu, Xingjun Ma, James Bailey, and Feng Lu. Reflection backdoor: A natural backdoor attack on deep neural networks. In European Conference on Computer Vision, pages 182-199. Springer, 2020. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 162, + 288, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 162, + 288, + 205 + ], + "spans": [ + { + "bbox": [ + 48, + 162, + 288, + 205 + ], + "type": "text", + "content": "[34] Aleksander Madry, Aleksandar Makelov, Ludwig Schmidt, Dimitris Tsipras, and Adrian Vladu. Towards deep learning models resistant to adversarial attacks. arXiv preprint arXiv:1706.06083, 2017. 1, 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 206, + 288, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 206, + 288, + 249 + ], + "spans": [ + { + "bbox": [ + 48, + 206, + 288, + 249 + ], + "type": "text", + "content": "[35] Yifei Min, Lin Chen, and Amin Karbasi. The curious case of adversarially robust models: More data can help, double descend, or hurt generalization. In Uncertainty in Artificial Intelligence, pages 129-139. PMLR, 2021. 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 251, + 288, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 251, + 288, + 282 + ], + "spans": [ + { + "bbox": [ + 48, + 251, + 288, + 282 + ], + "type": "text", + "content": "[36] Tuan Anh Nguyen and Anh Tran. Input-aware dynamic backdoor attack. Advances in Neural Information Processing Systems, 33:3454-3464, 2020. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 284, + 287, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 284, + 287, + 315 + ], + "spans": [ + { + "bbox": [ + 48, + 284, + 287, + 315 + ], + "type": "text", + "content": "[37] Silvia Noschese, Lionello Pasquini, and Lothar Reichel. Tridiagonal toeplitz matrices: Properties and novel applications. 2006. 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 317, + 288, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 317, + 288, + 381 + ], + "spans": [ + { + "bbox": [ + 48, + 317, + 288, + 381 + ], + "type": "text", + "content": "[38] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 383, + 288, + 448 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 383, + 288, + 448 + ], + "spans": [ + { + "bbox": [ + 48, + 383, + 288, + 448 + ], + "type": "text", + "content": "[39] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, Alexander C. Berg, and Li Fei-Fei. ImageNet Large Scale Visual Recognition Challenge. International Journal of Computer Vision (IJCV), 115(3):211-252, 2015. 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 449, + 287, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 449, + 287, + 480 + ], + "spans": [ + { + "bbox": [ + 48, + 449, + 287, + 480 + ], + "type": "text", + "content": "[40] M. Sandler, A. Howard, M. Zhu, A. Zhmoginov, and L.C. Chen. Mobilenetv2: Inverted residuals and linear bottle-necks. In CVPR, 2018. 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 482, + 287, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 482, + 287, + 525 + ], + "spans": [ + { + "bbox": [ + 48, + 482, + 287, + 525 + ], + "type": "text", + "content": "[41] Mert Bulent Sariyildiz, Julien Perez, and Diane Larlus. Learning visual representations with caption annotations. In European Conference on Computer Vision, pages 153-170. Springer, 2020. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 526, + 287, + 580 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 526, + 287, + 580 + ], + "spans": [ + { + "bbox": [ + 48, + 526, + 287, + 580 + ], + "type": "text", + "content": "[42] Ali Shafahi, W Ronny Huang, Mahyar Najibi, Octavian Suciu, Christoph Studer, Tudor Dumitras, and Tom Goldstein. Poison frogs! targeted clean-label poisoning attacks on neural networks. Advances in neural information processing systems, 31, 2018. 2, 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 582, + 287, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 582, + 287, + 612 + ], + "spans": [ + { + "bbox": [ + 48, + 582, + 287, + 612 + ], + "type": "text", + "content": "[43] Connor Shorten and Taghi Khoshgoftaar. A survey on image data augmentation for deep learning. Journal of Big Data, 6, 07 2019. 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 614, + 287, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 614, + 287, + 646 + ], + "spans": [ + { + "bbox": [ + 48, + 614, + 287, + 646 + ], + "type": "text", + "content": "[44] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. CoRR, abs/1409.1556, 2015. 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 647, + 287, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 647, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 647, + 287, + 689 + ], + "type": "text", + "content": "[45] Christian Szegedy, Wojciech Zaremba, Ilya Sutskever, Joan Bruna, Dumitru Erhan, Ian Goodfellow, and Rob Fergus. Intriguing properties of neural networks. arXiv preprint arXiv:1312.6199, 2013. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "type": "text", + "content": "[46] M. Tan and Q. Le. Efficientnetv2: Smaller models and faster training. In ICML, 2021. 6" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 564 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "text", + "content": "[47] Lue Tao, Lei Feng, Jinfeng Yi, Sheng-Jun Huang, and Songcan Chen. Better safe than sorry: Preventing delusive adversaries with adversarial training. Advances in Neural Information Processing Systems, 34, 2021. 1, 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 118, + 545, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 545, + 150 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 545, + 150 + ], + "type": "text", + "content": "[48] H. Touvron, M. Cord, M. Douze, F. Massa, A. Sablayrolles, and H. Jegou. Training data-efficient image transformers & distillation through attention. In ICML, 2021. 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 152, + 545, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 152, + 545, + 184 + ], + "spans": [ + { + "bbox": [ + 307, + 152, + 545, + 184 + ], + "type": "text", + "content": "[49] Laurens van der Maaten and Geoffrey Hinton. Visualizing data using t-sne. Journal of Machine Learning Research, 9(86):2579-2605, 2008. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 186, + 545, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 186, + 545, + 228 + ], + "spans": [ + { + "bbox": [ + 307, + 186, + 545, + 228 + ], + "type": "text", + "content": "[50] Wenxiao Wang, Alexander Levine, and Soheil Feizi. Improved certified defenses against data poisoning with (deterministic) finite aggregation. CoRR, abs/2202.02628, 2022. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 230, + 545, + 273 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 230, + 545, + 273 + ], + "spans": [ + { + "bbox": [ + 307, + 230, + 545, + 273 + ], + "type": "text", + "content": "[51] Yisen Wang, Xingjun Ma, James Bailey, Jinfeng Yi, Bowen Zhou, and Quanquan Gu. On the convergence and robustness of adversarial training. arXiv preprint arXiv:2112.08304, 2021. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 275, + 547, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 275, + 547, + 330 + ], + "spans": [ + { + "bbox": [ + 307, + 275, + 547, + 330 + ], + "type": "text", + "content": "[52] Da Yu, Huishuai Zhang, Wei Chen, Jian Yin, and Tie-Yan Liu. Availability attacks create shortcuts. In Proceedings of the 28th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, KDD '22, page 2367-2376, New York, NY, USA, 2022. Association for Computing Machinery. 2, 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 331, + 545, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 331, + 545, + 364 + ], + "spans": [ + { + "bbox": [ + 307, + 331, + 545, + 364 + ], + "type": "text", + "content": "[53] Chia-Hung Yuan and Shan-Hung Wu. Neural tangent generalization attacks. In International Conference on Machine Learning, pages 12230-12240. PMLR, 2021. 1, 3, 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 365, + 545, + 420 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 365, + 545, + 420 + ], + "spans": [ + { + "bbox": [ + 307, + 365, + 545, + 420 + ], + "type": "text", + "content": "[54] Sangdoo Yun, Dongyoon Han, Seong Joon Oh, Sanghyuk Chun, Junsuk Choe, and Youngjoon Yoo. Cutmix: Regularization strategy to train strong classifiers with localizable features. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6023-6032, 2019. 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 421, + 545, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 421, + 545, + 475 + ], + "spans": [ + { + "bbox": [ + 307, + 421, + 545, + 475 + ], + "type": "text", + "content": "[55] Sergey Zagoruyko and Nikos Komodakis. Wide residual networks. In Edwin R. Hancock Richard C. Wilson and William A. P. Smith, editors, Proceedings of the British Machine Vision Conference (BMVC), pages 87.1-87.12. BMVA Press, September 2016. 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 476, + 545, + 509 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 476, + 545, + 509 + ], + "spans": [ + { + "bbox": [ + 307, + 476, + 545, + 509 + ], + "type": "text", + "content": "[56] Hongyi Zhang, Moustapha Cisse, Yann N. Dauphin, and David Lopez-Paz. mixup: Beyond empirical risk minimization. CoRR, abs/1710.09412, 2017. 6" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 510, + 545, + 564 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 510, + 545, + 564 + ], + "spans": [ + { + "bbox": [ + 307, + 510, + 545, + 564 + ], + "type": "text", + "content": "[57] Hongyang Zhang, Yaodong Yu, Jiantao Jiao, Eric Xing, Laurent El Ghaoui, and Michael Jordan. Theoretically principled trade-off between robustness and accuracy. In International conference on machine learning, pages 7472-7482. PMLR, 2019. 2" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "3871" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/CUF_ Continuous Upsampling Filters/5b0cd1d3-2b2b-407b-9bec-e263533bd51d_content_list.json b/2023/CUF_ Continuous Upsampling Filters/5b0cd1d3-2b2b-407b-9bec-e263533bd51d_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..8730219b9c93a28181763043c14f08ea90dcd41e --- /dev/null +++ b/2023/CUF_ Continuous Upsampling Filters/5b0cd1d3-2b2b-407b-9bec-e263533bd51d_content_list.json @@ -0,0 +1,2023 @@ +[ + { + "type": "text", + "text": "CUF: Continuous Upsampling Filters", + "text_level": 1, + "bbox": [ + 294, + 161, + 673, + 181 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Cristina N. Vasconcelos Cengiz Oztireli Mark Matthews", + "bbox": [ + 240, + 207, + 727, + 223 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Milad Hashemi Kevin Swersky Andrea Tagliasacchi", + "bbox": [ + 258, + 224, + 709, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Google Research", + "bbox": [ + 411, + 254, + 549, + 270 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/5e5adc538e942a4d00d574aba78eb0bad1b5e60eb13a5bd807a8bc06cde65a15.jpg", + "image_caption": [ + "Memory Consumption" + ], + "image_footnote": [], + "bbox": [ + 117, + 318, + 467, + 450 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/498b52f40a7682f7ba7ddc53672eca71dc97ed84987bf2488a3858a8ee7c327c.jpg", + "image_caption": [ + "Super-resolution Accuracy", + "Figure 1. We report (left) memory and FLOPs for upsampling an $256 \\times 256$ image by different scale factors $(2 \\times, 3 \\times, 4 \\times)$ using integer scale upsamplers (Sub-Pixel Convolution [31] and CIF-instantiated) and arbitrary-scale upsamplers (Meta-SR [16], LIIF [6], LTE [19] and CIF), but the same encoder backbone (EDSR-baseline [21]); and (right) the relationship between each upsampler FLOPS and PSNR performance on DIV2k dataset. Our arbitrary scale model is significantly lighter than other methods in the same class (i.e. continuous super-res). Further, when instantiated for integer scale factors, our upsampler is even-more efficient than sub-pixel convolutions [31]." + ], + "image_footnote": [], + "bbox": [ + 490, + 317, + 838, + 450 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 530, + 312, + 543 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Neural fields have rapidly been adopted for representing 3D signals, but their application to more classical 2D image-processing has been relatively limited. In this paper, we consider one of the most important operations in image processing: upsampling. In deep learning, learnable upsampling layers have extensively been used for single image super-resolution. We propose to parameterize upsampling kernels as neural fields. This parameterization leads to a compact architecture that obtains a 40-fold reduction in the number of parameters when compared with competing arbitrary-scale super-resolution architectures. When upsampling images of size $256 \\times 256$ we show that our architecture is $2x - 10x$ more efficient than competing arbitrary-scale super-resolution architectures, and more efficient than sub-pixel convolutions when instantiated to a single-scale model. In the general setting, these gains grow polynomially with the square of the target scale. We validate our method on standard benchmarks showing such efficiency gains can be achieved without sacrifices in super-resolution performance. https://cuf-paper.github.io", + "bbox": [ + 75, + 576, + 472, + 853 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 500, + 530, + 630, + 543 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Neural-fields represent signals with coordinate-based neural-networks. They have found application in a multitude of areas including 3D reconstruction [24], novel-view synthesis [32], convolutions [14], and many others [37].", + "bbox": [ + 497, + 553, + 890, + 609 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recent research has investigated the use of neural fields in the context of single image super-resolution [6, 19]. These models are based on multi-layer perceptrons conditioned on latent representation produced by encoders. While such architectures allow for continuous-scale superresolution, they require the execution of a conditional neural field for every pixel at the target resolution, making them unsuitable in applications with limited computational resources. Further, such a large use of resources is not justified by a increase in performance compared to classical convolutional architectures such as sub-pixel convolutions [31]. In summary, neural fields have not yet found widespread adoption as classical solutions are ① trivial to implement and ② more efficient. As they generally perform compa", + "bbox": [ + 497, + 611, + 892, + 805 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 95, + 8, + 169, + 46 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation.", + "bbox": [ + 228, + 5, + 818, + 21 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "Except for this watermark, it is identical to the accepted version;", + "bbox": [ + 319, + 21, + 727, + 34 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 287, + 34, + 761, + 48 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "1 Thereon we assume single image when talking about super-resolution.", + "bbox": [ + 499, + 820, + 890, + 832 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "2These encoders were originally proposed for classical super-resolution applications, and include both convolutional [8, 21, 40] as well as attentional [7, 20, 39] architectures.", + "bbox": [ + 499, + 832, + 890, + 865 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "9999", + "bbox": [ + 468, + 887, + 500, + 897 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "rably, their usage is not justified in light of these points; see Figure 1. In this paper, we focus on overcoming these limitations, while noting that (regressive) super-resolution performance is in the saturation regime (i.e. further improvements in image quality seem unlikely without relying on generative modeling [30], and small improvements in PSNR do not necessarily correlate with image quality).", + "bbox": [ + 75, + 125, + 468, + 223 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our driving hypothesis is that super-resolution convolutional filters are highly correlated - both spatially, as well as across scales. Hence, representing such filters in the latent space of a conditional neural field can effectively capture and compress such correlations. Given that neural fields encode continuous functions within neural networks, we call our filters Continuous Upsampling Filters (CUFs). While neural fields have proven effective in parameterizing 3D convolutions [5,14,36], we demonstrate that there are very significant savings to be had in the parameterization of 2D convolutions.", + "bbox": [ + 75, + 223, + 468, + 374 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In implementing continuous upsampling filters, we draw inspiration from sub-pixel convolutions [31], and realize them via depth-wise convolutions. This not only makes CUFs significantly more efficient than competing continuous super-resolution architectures, but when instantiated to single-image super-resolution they are, surprisingly, even more efficient than \"ad-hoc\" sub-pixel convolutions with same number of input and output channels; see Figure 1.", + "bbox": [ + 75, + 375, + 468, + 486 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Contributions. We investigate the use of neural fields as a parameterization of convolutional upsampling layers in super-resolution architectures, and show how:", + "bbox": [ + 75, + 490, + 468, + 532 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- The continuity of neural-fields leads to training of compact convolutional architecture for continuous super-resolution (similar super-resolution PSNR but with flop reduction that grows polynomially with the square of the target scale.).", + "- Instantiating our continuous kernels into their discrete counterpart leads to an efficient inference pipeline, even more efficient in number of operations per target pixel than ad-hoc sub-pixel convolutions with same number of input and output channels.", + "- Discrete cosine transforms can be used as an efficient replacement for Fourier bases in the implementation of positional encoding for neural fields.", + "- These gains do not hinder reconstruction performance (PSNR), by carefully validating our method on standard benchmarks, and thoroughly abating our design choices." + ], + "bbox": [ + 77, + 532, + 468, + 754 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related work", + "text_level": 1, + "bbox": [ + 75, + 762, + 213, + 777 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "A thorough coverage of single-image super-resolution can be found in the following surveys [2, 3, 35]; in what follows, we provide an overview of classical techniques for super-resolution based on deep learning. As our model is based on neural-fields, we also point the reader to a survey in this topic [37], and below we discuss the existing works", + "bbox": [ + 75, + 783, + 468, + 866 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/9a45162d0acb5af7b4168ea6c9d54d846460fa724d116bbee1664940ae4fea2c.jpg", + "image_caption": [ + "Figure 2. Previous Architectures - A visual comparison of the super-resolution architectures of Meta-SR [16], local implicit image function LIIF [6], and local texture estimators (LTE) [19], shown in an isotropic scaling configuration. Components inside the $2 \\times 2$ Ensemble box are computed over the four closest pixels to compute a weighted average. The number of encoder channels, $C_e$ , is dependent on the encoder chosen." + ], + "image_footnote": [], + "bbox": [ + 500, + 110, + 894, + 274 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "for super-resolution based on neural fields.", + "bbox": [ + 499, + 382, + 781, + 396 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Single scale. There are two main frameworks for super-res, which mainly differ in the placement of upsampling operators within the architecture.", + "bbox": [ + 497, + 402, + 890, + 442 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- In the pre-upsampling framework [8], an initial upscaled image is obtained using a non-trainable upsampler (e.g. bicubic), which is then post-processed by a neural network. This enables arbitrary size/scaling, but can introduce side effects such as noise amplification and blurring.", + "- In the post-upsampling framework [31], an encoder that preserves the input spatial resolution is followed by a shallow upsampling component." + ], + "bbox": [ + 500, + 444, + 890, + 555 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Note the computational complexity of pre-upsampling is significant, as these model operate directly at the target resolution. Consequently, post-upsampling is the de-facto mainstream approach [9, 20, 21, 34, 39, 40]. For this reason, we investigate the impact of an implicit upsampler in postupsampling frameworks, and cover a diverse set of architectures ranging from large [22], to extremely lightweight models [10].", + "bbox": [ + 497, + 555, + 890, + 666 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Neural Fields. The application of neural fields to superresolution [16] has enabled the design of architectures that support multiple target scales, as well as non-integer super-resolution scales.", + "bbox": [ + 497, + 672, + 890, + 726 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "- In meta super-resolution MetaSR [16], a hyper-network is conditioned on the target scale factor and on the output pixel relative coordinates (ie. the difference between source and target grids covering the same normalized global space). Their model consists of: (1) transforming the input image using an encoder into deep features at the source resolution; (2) upsampling the deep features into the target space (using nearest-neighbors); (3) using the hyper-network to produce a filter set per target relative position and scale; (4) processing the enlarged target", + "bbox": [ + 500, + 728, + 890, + 866 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "10000", + "bbox": [ + 465, + 929, + 505, + 941 + ], + "page_idx": 1 + }, + { + "type": "table", + "img_path": "images/54995636e4a109efe7e975633994d5adfe0d5ddbe3affa9d26dbd514e24e9ab0.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Hyper NetworkUpsampler
input#layers // #neuronsoutputparams (K)inputlayersneuronsparams (K)
Meta-SR [16][δs(x), 1/s]2 // 25664 × 9 × 3445Eθconvolution--
LIIF [6]----[Εθ, δs(x), 1/s]5 dense256347
LTE [19]----ha(Εθ){sin, cos}{δs(x)π(hb(Εθ) + hc(1/s))}4 dense256494
Ours:[Π(δs(x)),Π(s),Π(k)]4 // 32646Eθd-w conv. + 2 dense644
", + "bbox": [ + 78, + 122, + 895, + 181 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Table 1. Super-resolution with neural-fields: An overview of the main characteristics of neural field based architectures. Square brackets represent concatenation. In LTE [19], $h_a, h_b$ are $3 \\times 3$ convolutional layers and $h_c$ is a linear layer.", + "bbox": [ + 75, + 184, + 894, + 212 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/be3372693e5c8fc6e98b078c9b12aca104531ad914a3891667513b6798354c9e.jpg", + "image_caption": [ + "Figure 3. Architecture - Continuous Upsampling Filters (CUF) are based on a hyper-network. For each pixel position in the highres space, the hyper-network is computed $K^2$ times, once for each position $i,j$ of the upsampling kernel (batched in practice). This can be reduced when upsampling with integer scales $s$ , as is common in practice. The Hyper-network output is aggregated into a tensor that is dot multiplied against the unfolded nearest-neighbor interpolation, computing a depth-wise convolution with per-pixel filters. The number of encoder channels, $C_e$ , is dependent on the encoder used. In our experiments, we set the number of hypernetwork channels, $C_h = 32$ ." + ], + "image_footnote": [], + "bbox": [ + 77, + 220, + 472, + 423 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "resolution feature map into the final output using the corresponding weights.", + "bbox": [ + 90, + 604, + 468, + 631 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- In local implicit image function LIIF [6], the author eliminates the use of a hyper-network (MetaSR, step 3) by extending the sampled target resolution features (step 2) with extra channels representing the relative coordinates and target scale. Their combined features are further processed through a stack of five dense layers inspired by the multi-layer perceptron (MLP) architecture typically adopted as the end-to-end model in neural-fields formulations.", + "- In local texture estimator LTE [19] the authors enhance the features fed as input to the MLP layers. In their model, the output of the encoder is processed by three extra trainable layers associated with the amplitude, frequency and phase of sin/cosine waves. The resulting projection in the target resolution layer is used as the input to an MLP that stacks 4 fully-connected layers. In order to further improve the results they also adopt a global skip" + ], + "bbox": [ + 77, + 631, + 470, + 866 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "connection with a bilinear up-scaled version of the input around the full model, such that the deep model focus on the computation of the residual between to the closed form approximation and the final result.", + "bbox": [ + 512, + 223, + 892, + 277 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "A visualization of these architectures can be found in Figure 2. In practice, LIIF and LTE also increase the number of feature channels as their MLP layers have more neurons (256) than the number of channels produced as the encoder output (64 for encoders such as EDSR [21] and RDN [40]). In contrast to these previous works, our upsampler:", + "bbox": [ + 497, + 279, + 892, + 362 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Moves the computational burden of arbitrary upsampling into the continuous upsampler operator while reducing the operations performed in the target resolution space, preserving the number of channels produced by the deep encoder (Table 1).", + "- Requires fewer operations than the single-scale sub-pixel convolution in the most typical application case of upsampling an image by an integer scale factor, and producing results of comparable quality.", + "- Adopt a neural-fields formulation that focuses the sensitivity of our hyper-network to fine-grained changes in the representation of scale and relative position (reducing spectral bias)." + ], + "bbox": [ + 500, + 363, + 892, + 542 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 500, + 554, + 589, + 568 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Given a target scale $s$ , we aim to produce an upscaled image of size $sH \\times sW \\times 3$ given an input image of size $H \\times W \\times 3$ . Within the context of super-resolution, our core contribution is the introduction of a novel learnable upsampling layer. The upsampling layer can be interpreted as a decoder in a classical encoder-decoder architecture. We briefly review our encoder architecture in Section 3.1, and detail our decoder in Section 3.2. Note that our analysis focuses on the decoder, as a variety of encoders can be used. In Section 3.4, we then perform a conceptual comparison between our architecture, shown in Figure 3, and others based on neural fields, which are visually summarized in Figure 2.", + "bbox": [ + 497, + 576, + 892, + 756 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Training. Our network with parameters $\\theta$ is trained to regress a high resolution image $\\tilde{\\mathbf{I}}^s$ matching the ground truth $\\tilde{\\mathbf{I}}_{\\mathrm{gt}}^s$ at random positions $\\mathbf{x}$ given the low resolution input image $\\mathbf{I}$ :", + "bbox": [ + 497, + 761, + 892, + 818 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\underset {\\boldsymbol {\\theta}} {\\arg \\min} \\mathbb {E} _ {\\mathbf {I}} \\mathbb {E} _ {s} \\mathbb {E} _ {\\mathbf {x}} \\| \\tilde {\\mathbf {I}} ^ {s} (\\mathbf {x}; \\mathbf {I}, \\boldsymbol {\\theta}) - \\tilde {\\mathbf {I}} _ {\\mathrm {g l}} ^ {s} (\\mathbf {x}) \\| _ {2} ^ {1} \\qquad (1)\n$$\n", + "text_format": "latex", + "bbox": [ + 554, + 828, + 890, + 852 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "10001", + "bbox": [ + 465, + 929, + 504, + 942 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/a96e23ecf22c6927002922b940f84b654a448f063839424303ac92c15ed3ac5b.jpg", + "image_caption": [ + "Figure 4. Continuous super-resolution - Our architecture directly allows for an arbitrary and continuous choice of the superresolution factor. Conversely, traditional methods perform this indirectly: they need to upsample to an integer factor, and then consequently downsample to the target resolution. More examples are available in the appendix. Note that $\\times 4.5$ and $\\times 5$ are upsampling factors not available in the training data." + ], + "image_footnote": [], + "bbox": [ + 77, + 129, + 463, + 179 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/5cad7d8416adf878db3d614d692fa1f109362e6cb1b270a1f953906f800fdfc5.jpg", + "image_caption": [ + "all filters" + ], + "image_footnote": [], + "bbox": [ + 80, + 271, + 208, + 362 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/2e6fdc1a9cf64f469a5a1659ef1e0e9d695bfc25c3f180f6ba5c9018d01b850f.jpg", + "image_caption": [ + "filter #2 - continuous" + ], + "image_footnote": [], + "bbox": [ + 213, + 271, + 344, + 362 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/272b8837f562fca0ad5791ee96b8cab0237d77ffe09112f927ed64c4f4dd9d5c.jpg", + "image_caption": [ + "filter #2 - discretized", + "Figure 5. Continuous filters - CUF filters are continuous functions stored in neural fields. We visualize our filters by densely sampling each on the $[0,1]^2$ domain to highlight their continuous nature (left). Close-up of an exemplar filter (middle). Filters are discretized by sampling at pixel centers (right), allowing them to be used in classical fixed-scale super-res architectures." + ], + "image_footnote": [], + "bbox": [ + 346, + 271, + 477, + 362 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1.Encoder", + "text_level": 1, + "bbox": [ + 75, + 465, + 179, + 478 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Towards this objective, we first process our input image with an encoder producing $C$ -dimensional features:", + "bbox": [ + 75, + 486, + 468, + 514 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {E} _ {\\boldsymbol {\\theta}}: \\mathbb {R} ^ {H \\times W \\times 3} \\rightarrow \\mathbb {R} ^ {H \\times W \\times C} \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 173, + 517, + 468, + 535 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "and then unfold a $k \\times k$ spatial neighborhood of $C$ -dimensional features into a tensor of $C \\times k^2$ channels. Finally, we upsample the encoded image using nearest-neighbour interpolation. We define this step as:", + "bbox": [ + 75, + 541, + 468, + 597 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {U}: \\mathbb {R} ^ {H \\times W \\times C} \\rightarrow \\mathbb {R} ^ {s H \\times s W \\times C \\times k ^ {2}} \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 156, + 602, + 468, + 619 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The unfolding part of this procedure leads to feature maps that will allow us to implement depth-wise, spatial convolutions as dot products, as is typically done for low-level neural network implementations. We define this whole feature extraction procedure as:", + "bbox": [ + 75, + 625, + 468, + 694 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {F} = \\mathcal {U} \\left(\\mathcal {E} _ {\\theta} (\\mathbf {I})\\right) \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 221, + 701, + 468, + 716 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Decoder – “Continuous Upsampling Filters”", + "text_level": 1, + "bbox": [ + 75, + 726, + 453, + 742 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We draw our inspiration from classical sub-pixel convolutions [31], and combine it with recent research that apply neural fields for continuous super-resolution [6, 19]. Differently from these architectures, we achieve this objective by instantiating upsampling kernels $\\kappa$ via hypernetworks [13]. Our super-resolution network is algebraically expressed as:", + "bbox": [ + 75, + 749, + 468, + 845 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {\\mathbf {I}} ^ {s} (\\mathbf {x}; \\mathbf {I}, \\boldsymbol {\\theta}) = \\mathcal {D} _ {\\boldsymbol {\\theta}} \\left(\\mathcal {K} _ {\\boldsymbol {\\theta}} \\left(\\delta^ {s} (\\mathbf {x}); s\\right) \\cdot \\mathbf {F} (\\lfloor \\mathbf {x} / s \\rfloor)\\right) \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 126, + 851, + 468, + 866 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathbf{x}$ indexes pixel locations in the target resolution, and $\\delta^s (\\mathbf{x}) = \\mathrm{mod}(\\mathbf{x},s) / s$ renders our convolutional filters translation invariant, and the dot product implements a spatial convolution. The network $\\mathcal{D}$ is a point-wise layer that maps (super-resolved) features back to RGB values:", + "bbox": [ + 497, + 125, + 890, + 195 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {D} _ {\\theta}: \\mathbb {R} ^ {C} \\rightarrow \\mathbb {R} ^ {3} \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 640, + 206, + 890, + 222 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "This is applied to a grid of coordinates of size $\\mathbb{R}^{sH\\times sW\\times C}$ to get an image of size $\\mathbb{R}^{sH\\times sW\\times 3}$ . The coordinate $\\mathbf{x}$ is continuous, hence $\\kappa_{\\theta}$ is a neural field parameterization of a convolutional kernel mapping (continuous) spatial offsets and (continuous) scales to convolutional weights:", + "bbox": [ + 497, + 226, + 890, + 298 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {K} _ {\\boldsymbol {\\theta}}: [ 0, 1 ] ^ {2} \\times \\mathbb {R} ^ {+} \\rightarrow \\mathbb {R} ^ {C \\times K ^ {2}} \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 594, + 304, + 890, + 321 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Continuous kernel indexing. We can take our continuous formulation a step further by introducing a continuous parametrization of its kernel indexes $k_{i}, k_{j}$ indexing our $K \\times K$ convolution weight entries. Thus, the hyper-network representing the convolution field becomes:", + "bbox": [ + 497, + 336, + 890, + 404 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathcal {K}} _ {\\boldsymbol {\\theta}}: [ 0, 1 ] ^ {2} \\times \\{0, \\dots , K - 1 \\} ^ {2} \\times \\mathbb {R} ^ {+} \\rightarrow \\mathbb {R} ^ {C} \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 544, + 412, + 890, + 429 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "and $\\kappa_{\\theta}$ can then be constructed by 9 invocations to $\\hat{\\kappa}_{\\theta}$ in the case of a $K = 3$ and $K^2$ invocations in the general setting. We use $K = 3$ throughout the paper in order to facilitate a direct comparison to pre-existing baseline models.", + "bbox": [ + 497, + 435, + 890, + 491 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Equation (7) and Equation (8) are both valid ways of generating spatial kernels. The first can be seen as a multi-headed hypernetwork, while the second uses input conditioning to generate the kernel values. We use the latter in our experiments, as the layers of nonlinearities provide additional expressiveness compared to the linear transformation used in the multi-headed version. A comparison between the two is presented in Appendix D.", + "bbox": [ + 497, + 492, + 890, + 603 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Positional encoding", + "text_level": 1, + "bbox": [ + 499, + 610, + 685, + 625 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Naively representing signals with MLPs leads to \"spectral bias\" [26]: the overall difficulty these networks have in representing high-frequency signals. Hence we implement $\\kappa_{\\theta}$ as:", + "bbox": [ + 497, + 631, + 890, + 686 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {K} _ {\\boldsymbol {\\theta}} \\left(\\delta^ {s} (\\mathbf {x}), s, k\\right) = \\operatorname {M L P} \\left(\\Pi \\left(\\delta^ {s} (\\mathbf {x})\\right), \\Pi (s), \\Pi (k)\\right) \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 526, + 695, + 890, + 711 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where we apply positional encoding $\\Pi (\\cdot)$ to the MLP inputs. Several variants of positional encoding exist, including Fourier [24] and random Fourier [32] variants. However, as we strive for efficiency, we take inspiration from classical signal processing (e.g. JPEG) and instead employ a cosine-only transformation. Given a scalar $f_{n}$ sampled uniformly in the range $[0,f_{\\mathrm{max}}]$ , a scalar quantity $z\\in [0,1]$ is encoded as the (sorted) vector:", + "bbox": [ + 497, + 718, + 890, + 828 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\Pi (z) = \\left\\{\\cos \\left(\\frac {(2 z + 1) f _ {n} \\pi}{2}\\right) \\right\\} _ {n = 1} ^ {N} \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 569, + 834, + 890, + 870 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "10002", + "bbox": [ + 465, + 929, + 505, + 941 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/2fba32e02ee3228e31402839f325558c4cdbc3453b2a4233023dadf04dc10df5.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
EncoderUpsamplerEns.Set5Set14BSD100Urban100
×2×3×4×6×2×3×4×6×2×3×4×6×2×3×4×6
RDNfixed-scale [40]38.2434.7132.47-34.0130.5728.81-32.3429.2627.72-32.8928.8026.61-
MetaSR [16]38.2234.6332.3829.0433.9830.5428.7826.5132.3329.2627.7125.9032.9228.8226.5523.99
CUF (ours)38.2334.7232.5429.2533.9930.5828.8626.7032.3529.2927.7625.9933.0128.9126.7524.23
fixed-scale [40]+geo38.3034.7832.61-34.1030.6728.92-32.4029.3327.75-33.0929.0026.82-
LIIF [6]+loc38.1734.6832.5029.1533.9730.5328.8026.6432.3229.2627.7425.9832.8728.8226.6824.20
LTE [19]+loc38.2334.7232.6129.3234.0930.5828.8826.7132.3629.3027.7726.0133.0428.9726.8124.28
CUF (ours)+geo38.2834.8032.6329.2734.0830.6528.9226.7432.3929.3327.8026.0333.1629.0526.8724.32
SwinIRfixed-scale [31]38.3534.8932.72-34.1430.7728.94-32.4429.3727.83-33.4029.2927.07-
MetaSR [16]38.2634.7732.4729.0934.1430.6628.8526.5832.3929.3127.7525.9433.2929.1226.7624.16
CUF (ours)38.3434.8832.8029.5334.2930.7929.0226.8532.4529.3827.8526.0933.5429.4527.2424.62
fixed-scale [31]+geo38.3834.9532.81-34.2430.8329.02-32.4729.4127.87-33.5129.4227.21-
LIIF [6]+loc38.2834.8732.7329.4634.1430.7528.9826.8232.3929.3427.8426.0733.3629.3327.1524.59
LTE+1c [19]+loc38.3334.8932.8129.5034.2530.8029.0626.8632.4429.3927.8626.0933.5029.4127.2424.62
CUF (ours)++geo38.3834.9232.8329.5734.3330.8429.0526.9132.4729.4027.8826.1133.6529.5527.3224.69
", + "bbox": [ + 80, + 122, + 892, + 268 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/a01376d3c3a21d7430b94108a0c282c84338fdaa94634342621578786db937a5.jpg", + "table_caption": [ + "Table 2. Out-of-domain evaluation - performance on datasets not seen during training. Metrics computed on the luminance channel in the YCbCr color space, following previous work. Models using self-ensemble are marked with $+_{geo}$ and models using local self-ensemble are marked with $+_{loc}$ . Legend: best $2^{\\text{nd}}$ best" + ], + "table_footnote": [], + "table_body": "
Multi-scale up-sampling methods - DIV2k
EncoderUpsamplerEns.seen scalesunseen scales
×2×3×4×6×12
Bicubic31.0128.2226.6624.8222.27
EDSR-b.Sub-pixel conv.34.6930.9428.97--
MetaSR34.6430.9328.9226.6123.55
CUF (ours)34.7030.9929.0126.7623.73
Sub-pixel conv.+geo34.7831.0329.06--
LIIF+loc34.6730.9629.0026.7523.71
LTE+loc34.7231.0229.0426.8123.78
CUF (ours)+geo34.7931.0729.0926.8223.78
RDNSub-pixel conv.35.0131.2229.20--
MetaSR35.0031.2729.2526.8823.73
CUF (ours)35.0331.3129.3227.0323.94
Sub-pixel conv.+geo35.1031.3329.31--
LIIF+loc34.9931.2629.2726.9923.89
LTE+loc35.0431.3229.3327.0423.95
CUF (ours)+geo35.1131.3929.3927.0923.99
SwinIRSub-pixel conv.35.2831.4729.40--
MetaSR35.1531.4029.3326.9423.80
CUF (ours)35.2631.5229.5227.1924.07
Sub-pixel conv.+geo35.3331.5229.44--
LIIF+loc35.1731.4629.4627.1524.02
LTE+loc35.2431.5029.5127.2024.09
CUF (ours)+geo35.3131.5629.5627.2324.10
", + "bbox": [ + 80, + 324, + 467, + 551 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Table 3. In domain evaluation: tests on DIV2K's validation subset [25] on scales seen $(2\\times -4\\times)$ and unseen during training $(6\\times -30\\times)$ . Metrics taken on the RGB space. Results taken using geometric self-ensemble [21] are marked with $^+$ geo' and local-ensemble [6] with $^+$ loc'. Legend: best $2^{\\mathrm{nd}}$ best", + "bbox": [ + 75, + 555, + 468, + 624 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "By eliminating the imaginary component from the Fourier basis, we show in Section 4.3 how this reduces the number of trainable parameters of the first hyper-network layer by half, without affecting super-resolution quality.", + "bbox": [ + 75, + 636, + 468, + 693 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Analysis", + "text_level": 1, + "bbox": [ + 77, + 700, + 179, + 714 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We now perform a conceptual comparison of our network architecture to the commonly used Sub-Pixel Convolution, as well as several super-resolution approaches based on neural-fields. We further note how continuous upsampling filters can be instantiated to generate filters compatible with Sub-Pixel Convolution.", + "bbox": [ + 75, + 722, + 468, + 805 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Comparison to Sub-Pixel Convolution [31]. Sub-Pixel Convolution is the most commonly used super-resolution operator, especially when efficiency is critical. This is achieved by initially employing an expansion convolution", + "bbox": [ + 75, + 809, + 470, + 866 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/e64749d874d3c55d0d1c9f97deffbb2d76bb2892733a83adcebcc8c6302af96a.jpg", + "image_caption": [ + "Figure 6. A simple illustration of periodic shuffling used in SubPixel Convolution." + ], + "image_footnote": [], + "bbox": [ + 598, + 323, + 796, + 399 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "that generates a feature map with $s^2 N_{out}$ channels, where $N_{out}$ is a hyper parameter defining the target number of channels of the Sub-Pixel Convolution operation. Next, the periodic shuffling operator re-arranges the channels to produce a higher-resolution feature map (see Figure 6):", + "bbox": [ + 497, + 437, + 892, + 508 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {P}: \\mathbb {R} ^ {H \\times W \\times C} \\rightarrow \\mathbb {R} ^ {s H \\times s W \\times C / s ^ {2}} \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 581, + 511, + 890, + 530 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Similar to (6), whenever $N_{out}$ is chosen as different from the number of output colors, the resulting feature map is further projected by point-wise convolutions. Often, $N_{out}$ is taken as the same as the number of channels as produced by the encoder $\\mathcal{E}$ for architectures adopting Sub-Pixel Convolution and targeting high quality results [20,21,40]. We note that the Sub-Pixel Convolution design based on the expansion convolution does not enforce spatial correlations between neighboring pixels (here we refer to neighborhood in the output, high resolution, domain) and need to be learned from training data. Conversely, these correlations are inherently captured by our continuous convolutional filters.", + "bbox": [ + 497, + 536, + 892, + 703 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Further, in the integer scale setup where hyper-networks can be pre-instantiated, CUFs are computationally more efficient than sub-pixel convolutions whenever $K^2 + N_{in} + N_{out} < N_{out} \\times K^2$ (in the worst case scenario) as the cost associated with pre-computing the weights for a given integer up-sampling scale can be neglected when compared to the operations performed on the image grid (Appendix E).", + "bbox": [ + 497, + 703, + 890, + 801 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Comparisons to LIIF [6] and LTE [19]. In comparison to these works, we make the main network shallower (1 depth-wise continuous convolution and 2 dense layers) in order to", + "bbox": [ + 497, + 805, + 892, + 847 + ], + "page_idx": 4 + }, + { + "type": "page_footnote", + "text": "also called pixel-shuffle (PyTorch) or depth-to-space (TensorFlow).", + "bbox": [ + 515, + 853, + 877, + 866 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "10003", + "bbox": [ + 465, + 929, + 504, + 941 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "reduce the number of layers that operate at the target spatial resolution and at the encoder's channel resolution. At the same time, our upsampler head is considerably lighter than previous arbitrary scale methods, not only by the use of a depth-wise convolution but also by keeping the same number of channels produced by the features encoder in the main network ( $C_e = 64$ , which results in dense layers $16 \\times$ cheaper than its LIIF and LTE counterparts using 256 neurons). When performing non-integer upsampling, the costs with processing the hyper-network layers are proportional to the target image resolution, but still smaller than a single dense layer of LIFF and LTE heads, due to the adoption of a reduced number of channels (as $C_h = 32$ , each of CUF's hypernetwork dense layers is $64 \\times$ cheaper than LIIF and LTE layers using 256 neurons).", + "bbox": [ + 75, + 125, + 472, + 333 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Instantiating CUFs. At inference time, when targeting an integer upscaling factor $s$ , the hyper-network representing $\\mathcal{K}$ can be queried to retrieve the weights corresponding to $s^2$ relative subpixel positions as an initialization step during pre-processing. The retrieved weights are re-used across all pixels taking advantage of the existent periodicity. Thus, in the CUF-instantiated architecture the continuous kernel is replaced at test time with a discrete depth-wise convolution, followed by a pixel shuffling operation, in contrast with the unfolding operator used in the regular, fully continuous CUF described in this paper. Otherwise, the architecture remains the same. The costs associated with the hyper network at initialization can be neglected as $s^2 << sH \\times sW$ . In this setting, our model becomes as efficient as a sub-pixel-convolution architecture, while retaining the aforementioned continuous modeling properties at training time.", + "bbox": [ + 75, + 339, + 472, + 574 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Results", + "text_level": 1, + "bbox": [ + 75, + 587, + 164, + 600 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this section we describe our experimental setup in terms of backbone encoders and training/validation dataset, perform careful comparisons to the state-of-the-art (Section 4.1), and additional evaluation towards the implementation of lightweight super-res architectures (Section 4.2). We conclude by performing a thorough ablation (Section 4.3).", + "bbox": [ + 75, + 608, + 468, + 692 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Encoders.", + "text_level": 1, + "bbox": [ + 77, + 697, + 149, + 709 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We apply CUFs to a variety of encoders, to both show its generality, as well as its performance gains in a variety of settings.", + "bbox": [ + 75, + 712, + 468, + 752 + ], + "page_idx": 5 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- A state-of-the-art encoder for super-res named SwinIR [20] based on Swin-transformers [22].", + "- Well-known convolution-based encoders, namely EDSR-baseline [21] and Residual Dense Networks (RDN) [40].", + "- Architectures for lightweight and extremely lightweight inference, respectively SwinT-lightweight [20] and ABPN [10]." + ], + "bbox": [ + 77, + 755, + 468, + 851 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 4 contrasts the ablated encoders' composition and", + "bbox": [ + 75, + 852, + 468, + 866 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/04d3137e4d0c90bb047e2c7728221a49f1314054a84443403ea7934198123f11.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodCompositionParameters (in M)
encoderspc. 4×
RDNConvolutions22.000.30
SwinTConv. and Self-Attention11.600.30
EDSR-baselineConvolutions1.200.30
SwinT-lightweightConv. and Self-Attention0.900.03
ABPNConvolutions0.030.03
", + "bbox": [ + 504, + 122, + 889, + 214 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 4. Encoders - Our experiments use CUF layers in combination with various encoders differing in composition and size. Parameter counts reflect the encoder size and the upsampling head using a Sub-Pixel Convolution (spc) targeting $4 \\times$ upsampling.", + "bbox": [ + 497, + 218, + 890, + 270 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "size. Hyperparameter settings can be found in the appendix.", + "bbox": [ + 497, + 282, + 890, + 296 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Datasets. All models are trained using the DIV2K dataset training subset (800k images), introduced at the NTIRE 2017 Super Resolution Challenge [25]. Similarly to previous works, each image is randomly cropped 20 times per epoch and augmented with random flips, and 90 degrees rotations. We report peak signal-to-noise ratio (PSNR) results on the DIV2K validation set (100 images) $^4$ , in the RGB space. To understand the generalizability of our experimental results, we test our models on additional well known datasets (Set5 [4] - 5 images; Set14 [38] - 14 images; B100 [23] - 100 images; and Urban100 [17] - 100 images), in which, following previous works, PSNR is measured in the YCbCr luminance channel.", + "bbox": [ + 497, + 301, + 892, + 480 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Ensembling. Whenever remarked by $^{+}_{\\mathrm{geo}}$ , results incorporate the (commonly adopted) Geometric Self-ensemble [21], where the results of rotated versions of the input images are averaged at test time. LIIF and LTE results adopt Local Self-ensemble $(^{+}_{\\mathrm{loc}}$ ), where the result of applying the upsampler at 4 shifted grid points is averaged, resulting in a $\\times 4$ increase in computational complexity at training/test time. It may also introduce confounding factors to the optimization process that invalidate direct comparison to the fixed-scale upsamplers, as during backpropagation the encoder gradients are averaged. In Appendix B we combine these models.", + "bbox": [ + 497, + 486, + 890, + 652 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. State-of-the-art comparisons", + "text_level": 1, + "bbox": [ + 497, + 661, + 756, + 675 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 2 and Table 3 show the quantitative comparisons, while qualitative results can be observed in Figure 4, Figure 7 and Appendix F as well as on our supplementary website https://cuf-paper.github.io/. Table 3 contains in-domain results: models trained on DIV2k and tested on DIV2k (train/test split), while Table 2 presents out-of-domain results: models are trained on DIV2k, but tested on different datasets. Note that upsampling ratios larger than $\\times 4$ are not provided at training time. The models adopting sub-pixel convolutions are fixed-scale models, that is, one model is trained per target scale and thus do not", + "bbox": [ + 497, + 682, + 890, + 834 + ], + "page_idx": 5 + }, + { + "type": "page_footnote", + "text": "4The models we compare to also do not report results on the test-set, which is no longer publicly available.", + "bbox": [ + 499, + 843, + 890, + 866 + ], + "page_idx": 5 + }, + { + "type": "footer", + "text": "10004", + "bbox": [ + 465, + 929, + 505, + 941 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/4586c8a3fd80a808e0b05221866617bec8c08f87663c8ba05ecd8b9f2dec16b8.jpg", + "image_caption": [ + "Urban100 - #088" + ], + "image_footnote": [], + "bbox": [ + 97, + 122, + 211, + 173 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/9e58391643991332fe4c329c0477c87267164c8bf4026a6daa07fbb3b998d566.jpg", + "image_caption": [ + "LR" + ], + "image_footnote": [], + "bbox": [ + 213, + 123, + 324, + 174 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/3b426c19a5f7f535597822b7c85e804457a1fe9190335fed7d35b4e6186efc3a.jpg", + "image_caption": [ + "GT" + ], + "image_footnote": [], + "bbox": [ + 327, + 124, + 440, + 174 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/c3c0031b00bf782a9266f44315c1319367590822520f5091bf3c5210cfe1467b.jpg", + "image_caption": [ + "SwinIR + LIIF × 4" + ], + "image_footnote": [], + "bbox": [ + 97, + 182, + 211, + 233 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/3489f2945e0e53a2fc32604cb84c519932f56a4391e4b97683ed1586da5d6e7d.jpg", + "image_caption": [ + "SwinIR + LTE × 4" + ], + "image_footnote": [], + "bbox": [ + 213, + 182, + 324, + 233 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/3002623944c9c75fcf9cf89766f4bcae9509a6357a4ff777aab2d150774f47ef.jpg", + "image_caption": [ + "SwinIR + CUF × 4" + ], + "image_footnote": [], + "bbox": [ + 327, + 182, + 440, + 233 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/0231013cff2cc662bb784a6f52051ba45c6603dfb518f79458d4bf4649d26a36.jpg", + "image_caption": [ + "DIV2k - #853" + ], + "image_footnote": [], + "bbox": [ + 97, + 243, + 211, + 293 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b14bb37fa3c7c5e119d28d89350297f149cea13f2cd3e41bc65b34cf1aa666c3.jpg", + "image_caption": [ + "LR" + ], + "image_footnote": [], + "bbox": [ + 213, + 243, + 324, + 293 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/732f5b867c3ac7f4bb15e239019a7e49fa0acb48d9797a949f0e04c124f7835f.jpg", + "image_caption": [ + "GT" + ], + "image_footnote": [], + "bbox": [ + 327, + 243, + 440, + 293 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/1b961892fd03dfcdd62a1a80bb8179219db93b92446baf3d6e06843c45cae315.jpg", + "image_caption": [ + "RDN + LIIF × 8" + ], + "image_footnote": [], + "bbox": [ + 97, + 302, + 211, + 352 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/41cedf49c99746d1b1cdc083e689ded0c336299f748cb3e145cfa13fd7d0c819.jpg", + "image_caption": [ + "RDN+LTE×8" + ], + "image_footnote": [], + "bbox": [ + 213, + 302, + 324, + 352 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a845e23a0ef19ad4f1cf9defee4a993dab0cd79e60d0843f6750ff55b3e255e6.jpg", + "image_caption": [ + "RDN+CUF×8" + ], + "image_footnote": [], + "bbox": [ + 327, + 302, + 440, + 352 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/9308183792fdc1bbce259fe7dfb0f4f3ed5e06699023a8e4bf38fd71835603d0.jpg", + "image_caption": [ + "Urban100 - #030" + ], + "image_footnote": [], + "bbox": [ + 97, + 361, + 211, + 410 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a912ae22bde445f82745922ded934e03c7ce57cec0e86958df726d100d43d4a1.jpg", + "image_caption": [ + "LR" + ], + "image_footnote": [], + "bbox": [ + 213, + 361, + 324, + 410 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/0f546039c6d66d558b33a3e7ef670331c80ac329e000d15184683c24bde54012.jpg", + "image_caption": [ + "GT" + ], + "image_footnote": [], + "bbox": [ + 327, + 361, + 440, + 412 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/23d15cc28ecf17fb84b27930ca53c31806084e0d2d98e1a59b2fe41f854199cf.jpg", + "image_caption": [ + "EDSR + LIF × 4" + ], + "image_footnote": [], + "bbox": [ + 97, + 421, + 211, + 470 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/000f5e5546064e7592b68cd292b4e5d796b3d1b65449ffad0ab27b59ebfe3cdb.jpg", + "image_caption": [ + "EDSR + LTE × 4" + ], + "image_footnote": [], + "bbox": [ + 213, + 421, + 324, + 470 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/2f603dc4fc57483bcddf2a1a09630fc3c1ee252b353a16d7b584cbe3f2a4ed57.jpg", + "image_caption": [ + "EDSR + CUF × 4" + ], + "image_footnote": [], + "bbox": [ + 327, + 421, + 440, + 470 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/41369710361a31dac81168f56ea1109c5d3aca11a2b02c3536eaa8c8a2eb3bcd.jpg", + "image_caption": [ + "Figure 7. Qualitative evaluation - CUF combined with different encoders and upsampling scales.", + "Figure 8. Residuals - between ground truth and (from left to right) results from LIIF, LTE and CUF shown in Fig 7." + ], + "image_footnote": [], + "bbox": [ + 95, + 511, + 210, + 564 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/09137e00bd551d75b0d426599bebaab0afe01fbceeaef04d8cbfee5f3fa3671f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 215, + 511, + 329, + 564 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/057e4f837f989a538cc20928506cd0c73e11a4e1b5c46c3726585ba596eadb7a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 334, + 511, + 450, + 564 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "generalize to unseen or fractional scales. On the remaining rows we present results of arbitrary-scale upsamplers (MetaSR [16], LIIF [6], LTE [19]), where a single model is trained and tested across different upsampling ratios. Our method is the first arbitrary-scale method to match (or surpass) fixed-scale performance across all encoders in both single-pass inference and self-ensemble scenarios.", + "bbox": [ + 75, + 606, + 467, + 704 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/248a3659190d6dfb9cfbd087c8a6bc04d25f006912b7f65497f0eb30cfff7bf1.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
UpsamplerUses Periodicity\\( \\times 2 \\)\\( \\times 3 \\)\\( \\times 4 \\)
WCGPUWCGPUWCGPU
EDSR-baseline encoder
Sub-pixel conv. **✓10.96.411.07.311.69.0
CUF inst. (ours) **✓11.16.511.27.611.09.1
LIF75.574.0164.6163.1295.0293.4
LTE54.845.9101.792.5167.0157.8
CUF (ours) - k as input26.718.852.444.081.573.0
CUF (ours) - k as output20.212.227.219.837.630.0
ABPN encoder
Sub-pixel conv. **✓1.90.61.90.61.90.7
CUF inst. (ours) **✓2.10.72.11.12.11.5
", + "bbox": [ + 78, + 712, + 467, + 820 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 5. Upsampling $(256^{2})$ on a A100 GPU. Wall clock (wc) (i.e., considering both CPU and GPU execution) and GPU time in ms averaged over $1k$ runs.", + "bbox": [ + 75, + 823, + 467, + 862 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/07f5c6bfac8c1d35f9c972ab3023874823312fa45b26dfd9e262507cad3e1f1f.jpg", + "image_caption": [ + "ABPN + Continuous Upsampler", + "Figure 9. Architectures (lightweight) - The original ABPN [11] single scale architecture (left) and our arbitrary-scale version using the implicit upsampler (right). Note how the two consecutive $3 \\times 3$ native convolutions are exchanged for a $3 \\times 3$ depth-wise implicit convolution followed by point-wise layers." + ], + "image_footnote": [], + "bbox": [ + 519, + 122, + 850, + 321 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. Lightweight super-resolution", + "text_level": 1, + "bbox": [ + 499, + 400, + 756, + 416 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The encoders and upsamplers considered so far are not designed to be run on devices with limited memory and power. Running super-resolution algorithms on mobile devices comes with unique challenges due to limited RAM and non-efficient support for certain common operations [18]. The proposed CUF-based upsampler, in combination with a mobile-compatible encoder, can be effectively used for upscaling on mobile devices.", + "bbox": [ + 497, + 422, + 890, + 530 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To test the effectiveness of the CUFs, we adopt a lightweight version of SwinIR (0.9M parameters) [20] as well as one of the best performing super-lightweight architectures for mobile devices (30K parameters) [10] as the encoder. For the latter, the original architecture (Figure 9, left) consists of a lightweight encoder and an upscaling module consisting of a series of convolutions, ReLU's, and a standard periodic shuffle layer. We replace this upscaling module with a CUF-based one (Figure 9, right). With instantiated CUFs, this means simply introducing a depthwise convolution followed by pointwise layers, where the weights are output by the learned CUFs.", + "bbox": [ + 497, + 533, + 890, + 698 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 6 illustrates that CUFs provide the same level of accuracy while enabling continuous scale and efficient inference. To our knowledge, this is the first mobile-friendly continuous scale super-resolution architecture.", + "bbox": [ + 497, + 698, + 890, + 752 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Ablations", + "text_level": 1, + "bbox": [ + 499, + 762, + 610, + 775 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We run several ablation experiments in order to analyze the effects of various model choices. Here, we detail our choice of using DCT for positional encoding vs Fourier features, as well as an investigation of the redundancy of CUF filters versus subpixel convolution. An analysis on the various conditioning factors of our neural fields (positional en", + "bbox": [ + 497, + 783, + 890, + 866 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "10005", + "bbox": [ + 465, + 929, + 504, + 941 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/3fef5d493e7bd622eb96fd2fee14e7c7b772d4ee3b754eea364b89ddeea7332b.jpg", + "table_caption": [ + "Multi scale up-sampling methods" + ], + "table_footnote": [], + "table_body": "
EncoderUpsamplerSet5Set14BSD100Urban100
x2x3x4x6x2x3x4x6x2x3x4x6x2x3x4x6
SwintIR-light [20]Sub-Pixel Conv.38.1134.5432.35-33.8330.5328.76-32.2729.1927.68-32.5628.5826.48-
CUF (ours)38.1734.6232.4629.0933.9330.5728.8326.6132.3029.2327.7125.9532.7328.6826.5524.11
ABPN [10]Sub-Pixel Conv.37.3033.3931.11-32.9029.6827.97-31.6728.6427.14-30.3726.7724.96-
CUF (ours)37.3533.5531.3128.0632.9629.7428.0425.8631.7228.6927.1925.5230.5326.9225.0523.01
", + "bbox": [ + 80, + 131, + 889, + 192 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/06e88f4a47fcb3751ab05af46da0739cc85bef84c16c741ab2649c5ae2b31f51.jpg", + "table_caption": [ + "Table 6. Lightweight super-res - Our CUFs enable continuous super-resolution with lightweight (SwinIR-lightweight [20]) and mobile-friendly (ABPN [10]) backbones without quality loss." + ], + "table_footnote": [], + "table_body": "
2×3×4×
Fourier30.4526.8525.00
DCT30.4426.8625.00
", + "bbox": [ + 157, + 225, + 389, + 276 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 7. Cosine basis - matching results but using half of sinusoidal projections and neurons at the hyper-network input. PSNR from Urban-100 dataset using a ABPN encoder.", + "bbox": [ + 75, + 280, + 468, + 319 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "coding, kernel indices, scale, and target subpixel) can be found in the Supplementary Material.", + "bbox": [ + 75, + 346, + 468, + 375 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Positional encoding. A large body of work uses a Fourier basis for the positional encodings, e.g. [24, 33]. However, we opt to use a DCT basis as this requires half of the sinusoidal projections and neurons at the hyper-network first layer. This trick has been used in other domains, such as random approximations of stationary kernel functions [27].", + "bbox": [ + 75, + 381, + 468, + 465 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 7 shows the PSNR, as obtained using a DFT versus a DCT basis, and verify that we do not observe loss in performance, while at the same time saving half of the computation associated with positional encoding operations.", + "bbox": [ + 75, + 466, + 468, + 522 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Redundancy of filters. As noted in Section 3.4, sub-pixel convolution does not enforce spatial correlation between neighboring weights in its convolutional filters, but has to learn them from data. CUF on the other hand generates the filters with a hypernetwork, which intrinsically builds on smoothness due to its functional form. Here, we test this hypothesis on the filters of CUF and sub-pixel convolutions. For this analysis, we use an EDSR encoder with $C_e = 64$ , $K = 3 \\times 3$ , and $s = 3$ . Subpixel convolution maps $C_e$ features to $s^2 C_e$ features, and each group of $s^2$ output features are rearranged to form the upscaled feature space. We therefore consider each group of $s^2$ filters of size $K \\times K \\times C_e$ and do an eigenvalue analysis to determine the compressibility of these filters. This is done by forming $C_e$ separate matrix", + "bbox": [ + 75, + 528, + 468, + 723 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/7df9f14026dccd79d65aa4f295e919f21d391a9292f961e8579a336ce214c660.jpg", + "image_caption": [ + "(a) Subpixel Convolutions" + ], + "image_footnote": [], + "bbox": [ + 80, + 735, + 272, + 818 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/9847b4ff361a573ff20a3fc325b213fb34a6b6ab58e942ec7dec555bd6e2fb11.jpg", + "image_caption": [ + "(b) CUF", + "Figure 10. CUF depthwise continuous filters are low rank by construction. Subpixel-convolution fails to find low rank filters." + ], + "image_footnote": [], + "bbox": [ + 277, + 735, + 467, + 819 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "ces of size $s^2 \\times C_e K^2$ and performing PCA on each of them separately, forming 64 separate sets of eigenvalues. We plot this distribution for trained and untrained sub-pixel layers in Figure 10a. A faster decay means more redundancy, as more variance is explained by fewer components. In this case, training adds significant structure to the filters.", + "bbox": [ + 497, + 226, + 890, + 311 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Conversely, each output channel in CUF is computed by $s^2$ filters of size $K \\times K$ . These are not directly comparable to subpixel convolution, but we perform an analogous eigenvalue analysis on trained and untrained CUFs. As shown in Figure 10b, unlike subpixel convolution, CUFs actually become less redundant with training, suggesting that the initial hypernetwork already imposes a very smooth prior on the filters. In other words, CUF filters incorporate spatial correlations from the very beginning due to their functional form.", + "bbox": [ + 497, + 315, + 892, + 453 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusions", + "text_level": 1, + "bbox": [ + 499, + 477, + 626, + 491 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We propose CUF, a computationally efficient modeling of continuous upsampling filters as neural fields. Different from previous arbitrary-scale architectures, the performance gains obtained by our upsampler are not due to increased capacity but rather to a more efficient use of model parameters. A single hyper-network supports filter adaptation across scales while at the same time CUF's upsampling head has fewer parameters than previous arbitrary-scale heads, as well as fewer parameters than a single SubPixel Convolution head at $4 \\times$ upsampling. CUF is the first arbitrary scale model that can be effectively used for upscaling on mobile devices, an area previously dominated by single scale Sub-Pixel Convolution architectures. CUF's architecture is quite general. A natural direction of future work is to investigate its use for architectures with multiple upsampling layers (and the reuse of filters across scales), for different image generation settings such as GANs [12] and diffusion models [15]. Neural fields have done particularly well on higher-dimensional signals. Another line of future work is the use of CUF for efficient super-resolution of video data. Continuous convolution filters are an exciting and increasingly studied area of research [28,29] with many applications as an alternative to traditional convolutions.", + "bbox": [ + 497, + 502, + 890, + 820 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Overall, the presented results aligned with low computational resources suggest CUF offers a practical modeling of continuous filters.", + "bbox": [ + 497, + 825, + 892, + 865 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "10006", + "bbox": [ + 465, + 929, + 505, + 941 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 124, + 173, + 138 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Manoj Alwani, Han Chen, Michael Ferdman, and Peter Milder. Fused-layer cnn accelerators. 10 2016. 13", + "[2] Saeed Anwar, Salman Khan, and Nick Barnes. A deep journey into super-resolution: A survey. ACM Computing Surveys (ACMCSUR), 53(3), May 2020. 2", + "[3] Syed Muhammad Arsalan Bashir, Yi Wang, Mahrukh Khan, and Yilong Niu. A comprehensive review of deep learning-based single image super-resolution. PeerJ Computer Science, 7:e621, Jul 2021. 2", + "[4] Marco Bevilacqua, Aline Roumy, Christine Guillemot, and Marie line Alberi Morel. Low-complexity single-image super-resolution based on nonnegative neighbor embedding. In Proceedings of the British Machine Vision Conference, pages 135.1-135.10. BMVA Press, 2012. 6", + "[5] Alexandre Boulch. Convpoint: Continuous convolutions for point cloud processing. Computers & Graphics, 2020. 2", + "[6] Yinbo Chen, Sifei Liu, and Xiaolong Wang. Learning continuous image representation with local implicit image function. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8628-8638, 2021. 1, 2, 3, 4, 5, 7, 11", + "[7] Tao Dai, Jianrui Cai, Yongbing Zhang, Shu-Tao Xia, and Lei Zhang. Second-order attention network for single image super-resolution. In 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11057-11066, 2019. 1", + "[8] Chao Dong, Chen Change Loy, Kaiming He, and Xiaou Tang. Learning a deep convolutional network for image super-resolution. In David Fleet, Tomas Pajdla, Bernt Schiele, and Tinne Tuytelaars, editors, Computer Vision – ECCV 2014, pages 184–199, Cham, 2014. Springer International Publishing. 1, 2", + "[9] Chao Dong, Chen Change Loy, Kaiming He, and Xiaou Tang. Image super-resolution using deep convolutional networks. IEEE Trans. Pattern Anal. Mach. Intell., 38(2):295-307, feb 2016. 2", + "[10] Zongcai Du, Jie Liu, Jie Tang, and Gangshan Wu. Anchors-based plain net for mobile image super-resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, pages 2494-2502, June 2021. 2, 6, 7, 8, 11", + "[11] Z. Du, J. Liu, J. Tang, and G. Wu. Anchor-based plain net for mobile image super-resolution. In 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), pages 2494-2502, Los Alamitos, CA, USA, jun 2021. IEEE Computer Society. 7", + "[12] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial networks. Communications of the ACM, 63(11):139-144, 2020. 8", + "[13] David Ha, Andrew Dai, and Quoc V Le. Hypernetworks. arXiv preprint arXiv:1609.09106, 2016. 4", + "[14] Pedro Hermosilla, Tobias Ritschel, Pere-Pau Vázquez, Álvar Vinacua, and Timo Ropinski. Monte carlo convolution for learning on non-uniformly sampled point clouds. ACM Trans. Graph., 37(6), dec 2018. 1, 2" + ], + "bbox": [ + 78, + 147, + 470, + 866 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[15] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in Neural Information Processing Systems, 33:6840-6851, 2020. 8", + "[16] Xuecai Hu, Haoyuan Mu, Xiangyu Zhang, Zilei Wang, Tieniu Tan, and Jian Sun. Meta-sr: A magnification-arbitrary network for super-resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2019. 1, 2, 3, 5, 7", + "[17] Jia-Bin Huang, Abhishek Singh, and Narendra Ahuja. Single image super-resolution from transformed self-exemplars. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 5197-5206, 2015. 6, 11", + "[18] A. Ignatov, R. Timofte, M. Denna, A. Younes, A. Lek, M. Ayazoglu, J. Liu, Z. Du, J. Guo, X. Zhou, H. Jia, Y. Yan, Z. Zhang, Y. Chen, Y. Peng, Y. Lin, X. Zhang, H. Zeng, K. Zeng, P. Li, Z. Liu, S. Xue, and S. Wang. Real-time quantized image super-resolution on mobile npus, mobile ai 2021 challenge: Report. In 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), pages 2525-2534, Los Alamitos, CA, USA, jun 2021. IEEE Computer Society. 7", + "[19] Kyong Hwan Jin Jaewon Lee. Local texture estimator for implicit representation function. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022. 1, 2, 3, 4, 5, 7, 11", + "[20] Jingyun Liang, Jiezhang Cao, Guolei Sun, Kai Zhang, Luc Van Gool, and Radu Timofte. Swinir: Image restoration using swim transformer. In IEEE International Conference on Computer Vision Workshops, 2021. 1, 2, 5, 6, 7, 8, 11, 12", + "[21] Bee Lim, Sanghyun Son, Heewon Kim, Seungjun Nah, and Young Mu Lee. Enhanced deep residual networks for single image super-resolution. In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, July 2017. 1, 2, 3, 5, 6, 11, 12", + "[22] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 2, 6", + "[23] D. Martin, C. Fowlkes, D. Tal, and J. Malik. A database of human segmented natural images and its application to evaluating segmentation algorithms and measuring ecological statistics. In Computer Vision, 2001. ICCV 2001. Proceedings. Eighth IEEE International Conference on, volume 2, pages 416-423, 2001. 6", + "[24] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In ECCV, 2020. 1, 4, 8", + "[25] et al. Radu Timofte. Ntire 2017 challenge on single image super-resolution: Methods and results. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition Workshops, CVPRW 2017, pages 1110-1121, 2017. 5, 6, 11", + "[26] Nasim Rahman, Aristide Baratin, Devansh Arpit, Felix Draxler, Min Lin, Fred Hamprecht, Yoshua Bengio, and Aaron Courville. On the spectral bias of neural networks." + ], + "bbox": [ + 502, + 125, + 892, + 866 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "10007", + "bbox": [ + 465, + 929, + 504, + 941 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "In Kamalika Chaudhuri and Ruslan Salakhutdinov, editors, Proceedings of the 36th International Conference on Machine Learning, volume 97 of Proceedings of Machine Learning Research, pages 5301-5310. PMLR, 09-15 Jun 2019. 4", + "[27] Ali Rahimi and Benjamin Recht. Random features for large-scale kernel machines. Advances in neural information processing systems, 20, 2007. 8", + "[28] David W. Romero, Robert-Jan Bruintjes, Jakub Mikelaj Tomczak, Erik J. Bekkers, Mark Hoogendoorn, and Jan van Gemert. Flexconv: Continuous kernel convolutions with differentiable kernel sizes. In The Tenth International Conference on Learning Representations, ICLR 2022, Virtual Event, April 25-29, 2022. OpenReview.net, 2022. 8", + "[29] David W. Romero, Anna Kuzina, Erik J. Bekkers, Jakub Mikolaj Tomczak, and Mark Hoogendoorn. Ckconv: Continuous kernel convolution for sequential data. In The Tenth International Conference on Learning Representations, ICLR 2022, Virtual Event, April 25-29, 2022. OpenReview.net, 2022. 8", + "[30] Chitwan Sahara, Jonathan Ho, William Chan, Tim Salimans, David J Fleet, and Mohammad Norouzi. Image superresolution via iterative refinement. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2022. 2", + "[31] Wenzhe Shi, Jose Caballero, Ferenc Huszar, Johannes Totz, Andrew P. Aitken, Rob Bishop, Daniel Rueckert, and Zehan Wang. Real-time single image and video super-resolution using an efficient sub-pixel convolutional neural network. In 2016 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2016, Las Vegas, NV, USA, June 27-30, 2016, pages 1874-1883. IEEE Computer Society, 2016. 1, 2, 4, 5", + "[32] Matthew Tancik, Pratul P. Srinivasan, Ben Mildenhall, Sara Fridovich-Keil, Nithin Raghavan, Utkarsh Singhal, Ravi Ramamoorthi, Jonathan T. Barron, and Ren Ng. Fourier features let networks learn high frequency functions in low dimensional domains. NeurIPS, 2020. 1, 4", + "[33] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 8", + "[34] Xintao Wang, Ke Yu, Shixiang Wu, Jinjin Gu, Yihao Liu, Chao Dong, Yu Qiao, and Chen Change Loy. Esrgan: Enhanced super-resolution generative adversarial networks. In Laura Leal-Taixe and Stefan Roth, editors, Computer Vision - ECCV 2018 Workshops, pages 63-79, Cham, 2019. Springer International Publishing. 2", + "[35] Z. Wang, J. Chen, and S. H. Hoi. Deep learning for image super-resolution: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence, 43(10):3365-3387, oct 2021. 2", + "[36] Wenxuan Wu, Zhongang Qi, and Fuxin Li. Pointconv: Deep convolutional networks on 3d point clouds. In IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2019, Long Beach, CA, USA, June 16-20, 2019, pages 9621-9630. Computer Vision Foundation / IEEE, 2019. 2", + "[37] Yiheng Xie, Towaki Takikawa, Shunsuke Saito, Or Litany, Shiqin Yan, Numair Khan, Federico Tombari, James Tomp" + ], + "bbox": [ + 78, + 125, + 470, + 866 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "kin, Vincent Sitzmann, and Srinath Sridhar. Neural fields in visual computing and beyond, 2021. 1, 2", + "[38] Roman Zeyde, Michael Elad, and Matan Protter. On single image scale-up using sparse-representations. In Jean-Daniel Boissonnat, Patrick Chenin, Albert Cohen, Christian Gout, Tom Lyche, Marie-Laurence Mazure, and Larry L. Schumaker, editors, Curves and Surfaces, volume 6920 of Lecture Notes in Computer Science, pages 711-730. Springer, 2010. 6", + "[39] Yulun Zhang, Kunpeng Li, Kai Li, Lichen Wang, Bineng Zhong, and Yun Fu. Image super-resolution using very deep residual channel attention networks. In Vittorio Ferrari, Martial Hebert, Cristian Sminchisescu, and Yair Weiss, editors, Computer Vision - ECCV 2018, pages 294-310, Cham, 2018. Springer International Publishing. 1, 2", + "[40] Yulun Zhang, Yapeng Tian, Yu Kong, Bineng Zhong, and Yun Fu. Residual dense network for image restoration. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 43(7):2480-2495, 2021. 1, 2, 3, 5, 6, 11, 12" + ], + "bbox": [ + 502, + 125, + 892, + 371 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10008", + "bbox": [ + 465, + 929, + 504, + 941 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2023/CUF_ Continuous Upsampling Filters/5b0cd1d3-2b2b-407b-9bec-e263533bd51d_model.json b/2023/CUF_ Continuous Upsampling Filters/5b0cd1d3-2b2b-407b-9bec-e263533bd51d_model.json new file mode 100644 index 0000000000000000000000000000000000000000..ca8a279e2ec8d328f848e9a8106cd3e693d95dab --- /dev/null +++ b/2023/CUF_ Continuous Upsampling Filters/5b0cd1d3-2b2b-407b-9bec-e263533bd51d_model.json @@ -0,0 +1,2838 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.096, + 0.009, + 0.171, + 0.047 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.23, + 0.006, + 0.819, + 0.022 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation." + }, + { + "type": "header", + "bbox": [ + 0.32, + 0.022, + 0.729, + 0.035 + ], + "angle": 0, + "content": "Except for this watermark, it is identical to the accepted version;" + }, + { + "type": "header", + "bbox": [ + 0.288, + 0.035, + 0.762, + 0.049 + ], + "angle": 0, + "content": "the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.295, + 0.162, + 0.675, + 0.182 + ], + "angle": 0, + "content": "CUF: Continuous Upsampling Filters" + }, + { + "type": "text", + "bbox": [ + 0.242, + 0.208, + 0.728, + 0.224 + ], + "angle": 0, + "content": "Cristina N. Vasconcelos Cengiz Oztireli Mark Matthews" + }, + { + "type": "text", + "bbox": [ + 0.259, + 0.225, + 0.71, + 0.241 + ], + "angle": 0, + "content": "Milad Hashemi Kevin Swersky Andrea Tagliasacchi" + }, + { + "type": "text", + "bbox": [ + 0.413, + 0.255, + 0.551, + 0.271 + ], + "angle": 0, + "content": "Google Research" + }, + { + "type": "image_caption", + "bbox": [ + 0.22, + 0.303, + 0.396, + 0.318 + ], + "angle": 0, + "content": "Memory Consumption" + }, + { + "type": "image", + "bbox": [ + 0.119, + 0.319, + 0.468, + 0.451 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.574, + 0.303, + 0.779, + 0.318 + ], + "angle": 0, + "content": "Super-resolution Accuracy" + }, + { + "type": "image", + "bbox": [ + 0.492, + 0.318, + 0.84, + 0.451 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.456, + 0.893, + 0.52 + ], + "angle": 0, + "content": "Figure 1. We report (left) memory and FLOPs for upsampling an \\(256 \\times 256\\) image by different scale factors \\((2 \\times, 3 \\times, 4 \\times)\\) using integer scale upsamplers (Sub-Pixel Convolution [31] and CIF-instantiated) and arbitrary-scale upsamplers (Meta-SR [16], LIIF [6], LTE [19] and CIF), but the same encoder backbone (EDSR-baseline [21]); and (right) the relationship between each upsampler FLOPS and PSNR performance on DIV2k dataset. Our arbitrary scale model is significantly lighter than other methods in the same class (i.e. continuous super-res). Further, when instantiated for integer scale factors, our upsampler is even-more efficient than sub-pixel convolutions [31]." + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.532, + 0.313, + 0.545 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.577, + 0.473, + 0.854 + ], + "angle": 0, + "content": "Neural fields have rapidly been adopted for representing 3D signals, but their application to more classical 2D image-processing has been relatively limited. In this paper, we consider one of the most important operations in image processing: upsampling. In deep learning, learnable upsampling layers have extensively been used for single image super-resolution. We propose to parameterize upsampling kernels as neural fields. This parameterization leads to a compact architecture that obtains a 40-fold reduction in the number of parameters when compared with competing arbitrary-scale super-resolution architectures. When upsampling images of size \\(256 \\times 256\\) we show that our architecture is \\(2x - 10x\\) more efficient than competing arbitrary-scale super-resolution architectures, and more efficient than sub-pixel convolutions when instantiated to a single-scale model. In the general setting, these gains grow polynomially with the square of the target scale. We validate our method on standard benchmarks showing such efficiency gains can be achieved without sacrifices in super-resolution performance. https://cuf-paper.github.io" + }, + { + "type": "title", + "bbox": [ + 0.502, + 0.532, + 0.631, + 0.545 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.554, + 0.892, + 0.61 + ], + "angle": 0, + "content": "Neural-fields represent signals with coordinate-based neural-networks. They have found application in a multitude of areas including 3D reconstruction [24], novel-view synthesis [32], convolutions [14], and many others [37]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.612, + 0.893, + 0.806 + ], + "angle": 0, + "content": "Recent research has investigated the use of neural fields in the context of single image super-resolution [6, 19]. These models are based on multi-layer perceptrons conditioned on latent representation produced by encoders. While such architectures allow for continuous-scale superresolution, they require the execution of a conditional neural field for every pixel at the target resolution, making them unsuitable in applications with limited computational resources. Further, such a large use of resources is not justified by a increase in performance compared to classical convolutional architectures such as sub-pixel convolutions [31]. In summary, neural fields have not yet found widespread adoption as classical solutions are ① trivial to implement and ② more efficient. As they generally perform compa" + }, + { + "type": "page_footnote", + "bbox": [ + 0.5, + 0.821, + 0.892, + 0.833 + ], + "angle": 0, + "content": "1 Thereon we assume single image when talking about super-resolution." + }, + { + "type": "page_footnote", + "bbox": [ + 0.5, + 0.833, + 0.892, + 0.866 + ], + "angle": 0, + "content": "2These encoders were originally proposed for classical super-resolution applications, and include both convolutional [8, 21, 40] as well as attentional [7, 20, 39] architectures." + }, + { + "type": "list", + "bbox": [ + 0.5, + 0.821, + 0.892, + 0.866 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.469, + 0.888, + 0.501, + 0.898 + ], + "angle": 0, + "content": "9999" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.127, + 0.469, + 0.224 + ], + "angle": 0, + "content": "rably, their usage is not justified in light of these points; see Figure 1. In this paper, we focus on overcoming these limitations, while noting that (regressive) super-resolution performance is in the saturation regime (i.e. further improvements in image quality seem unlikely without relying on generative modeling [30], and small improvements in PSNR do not necessarily correlate with image quality)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.224, + 0.469, + 0.375 + ], + "angle": 0, + "content": "Our driving hypothesis is that super-resolution convolutional filters are highly correlated - both spatially, as well as across scales. Hence, representing such filters in the latent space of a conditional neural field can effectively capture and compress such correlations. Given that neural fields encode continuous functions within neural networks, we call our filters Continuous Upsampling Filters (CUFs). While neural fields have proven effective in parameterizing 3D convolutions [5,14,36], we demonstrate that there are very significant savings to be had in the parameterization of 2D convolutions." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.376, + 0.469, + 0.487 + ], + "angle": 0, + "content": "In implementing continuous upsampling filters, we draw inspiration from sub-pixel convolutions [31], and realize them via depth-wise convolutions. This not only makes CUFs significantly more efficient than competing continuous super-resolution architectures, but when instantiated to single-image super-resolution they are, surprisingly, even more efficient than \"ad-hoc\" sub-pixel convolutions with same number of input and output channels; see Figure 1." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.491, + 0.469, + 0.533 + ], + "angle": 0, + "content": "Contributions. We investigate the use of neural fields as a parameterization of convolutional upsampling layers in super-resolution architectures, and show how:" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.533, + 0.469, + 0.602 + ], + "angle": 0, + "content": "- The continuity of neural-fields leads to training of compact convolutional architecture for continuous super-resolution (similar super-resolution PSNR but with flop reduction that grows polynomially with the square of the target scale.)." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.602, + 0.469, + 0.671 + ], + "angle": 0, + "content": "- Instantiating our continuous kernels into their discrete counterpart leads to an efficient inference pipeline, even more efficient in number of operations per target pixel than ad-hoc sub-pixel convolutions with same number of input and output channels." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.671, + 0.469, + 0.712 + ], + "angle": 0, + "content": "- Discrete cosine transforms can be used as an efficient replacement for Fourier bases in the implementation of positional encoding for neural fields." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.712, + 0.469, + 0.755 + ], + "angle": 0, + "content": "- These gains do not hinder reconstruction performance (PSNR), by carefully validating our method on standard benchmarks, and thoroughly abating our design choices." + }, + { + "type": "list", + "bbox": [ + 0.078, + 0.533, + 0.469, + 0.755 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.763, + 0.215, + 0.779 + ], + "angle": 0, + "content": "2. Related work" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.784, + 0.469, + 0.867 + ], + "angle": 0, + "content": "A thorough coverage of single-image super-resolution can be found in the following surveys [2, 3, 35]; in what follows, we provide an overview of classical techniques for super-resolution based on deep learning. As our model is based on neural-fields, we also point the reader to a survey in this topic [37], and below we discuss the existing works" + }, + { + "type": "image", + "bbox": [ + 0.501, + 0.111, + 0.895, + 0.275 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.279, + 0.892, + 0.369 + ], + "angle": 0, + "content": "Figure 2. Previous Architectures - A visual comparison of the super-resolution architectures of Meta-SR [16], local implicit image function LIIF [6], and local texture estimators (LTE) [19], shown in an isotropic scaling configuration. Components inside the \\(2 \\times 2\\) Ensemble box are computed over the four closest pixels to compute a weighted average. The number of encoder channels, \\(C_e\\), is dependent on the encoder chosen." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.383, + 0.783, + 0.397 + ], + "angle": 0, + "content": "for super-resolution based on neural fields." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.403, + 0.891, + 0.444 + ], + "angle": 0, + "content": "Single scale. There are two main frameworks for super-res, which mainly differ in the placement of upsampling operators within the architecture." + }, + { + "type": "text", + "bbox": [ + 0.501, + 0.445, + 0.892, + 0.514 + ], + "angle": 0, + "content": "- In the pre-upsampling framework [8], an initial upscaled image is obtained using a non-trainable upsampler (e.g. bicubic), which is then post-processed by a neural network. This enables arbitrary size/scaling, but can introduce side effects such as noise amplification and blurring." + }, + { + "type": "text", + "bbox": [ + 0.501, + 0.514, + 0.892, + 0.556 + ], + "angle": 0, + "content": "- In the post-upsampling framework [31], an encoder that preserves the input spatial resolution is followed by a shallow upsampling component." + }, + { + "type": "list", + "bbox": [ + 0.501, + 0.445, + 0.892, + 0.556 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.556, + 0.892, + 0.667 + ], + "angle": 0, + "content": "Note the computational complexity of pre-upsampling is significant, as these model operate directly at the target resolution. Consequently, post-upsampling is the de-facto mainstream approach [9, 20, 21, 34, 39, 40]. For this reason, we investigate the impact of an implicit upsampler in postupsampling frameworks, and cover a diverse set of architectures ranging from large [22], to extremely lightweight models [10]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.673, + 0.892, + 0.728 + ], + "angle": 0, + "content": "Neural Fields. The application of neural fields to superresolution [16] has enabled the design of architectures that support multiple target scales, as well as non-integer super-resolution scales." + }, + { + "type": "text", + "bbox": [ + 0.501, + 0.729, + 0.892, + 0.868 + ], + "angle": 0, + "content": "- In meta super-resolution MetaSR [16], a hyper-network is conditioned on the target scale factor and on the output pixel relative coordinates (ie. the difference between source and target grids covering the same normalized global space). Their model consists of: (1) transforming the input image using an encoder into deep features at the source resolution; (2) upsampling the deep features into the target space (using nearest-neighbors); (3) using the hyper-network to produce a filter set per target relative position and scale; (4) processing the enlarged target" + }, + { + "type": "page_number", + "bbox": [ + 0.466, + 0.931, + 0.506, + 0.942 + ], + "angle": 0, + "content": "10000" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.079, + 0.123, + 0.896, + 0.182 + ], + "angle": 0, + "content": "
Hyper NetworkUpsampler
input#layers // #neuronsoutputparams (K)inputlayersneuronsparams (K)
Meta-SR [16][δs(x), 1/s]2 // 25664 × 9 × 3445Eθconvolution--
LIIF [6]----[Εθ, δs(x), 1/s]5 dense256347
LTE [19]----ha(Εθ){sin, cos}{δs(x)π(hb(Εθ) + hc(1/s))}4 dense256494
Ours:[Π(δs(x)),Π(s),Π(k)]4 // 32646Eθd-w conv. + 2 dense644
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.185, + 0.895, + 0.213 + ], + "angle": 0, + "content": "Table 1. Super-resolution with neural-fields: An overview of the main characteristics of neural field based architectures. Square brackets represent concatenation. In LTE [19], \\( h_a, h_b \\) are \\( 3 \\times 3 \\) convolutional layers and \\( h_c \\) is a linear layer." + }, + { + "type": "image", + "bbox": [ + 0.078, + 0.221, + 0.473, + 0.424 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.439, + 0.473, + 0.581 + ], + "angle": 0, + "content": "Figure 3. Architecture - Continuous Upsampling Filters (CUF) are based on a hyper-network. For each pixel position in the highres space, the hyper-network is computed \\( K^2 \\) times, once for each position \\( i,j \\) of the upsampling kernel (batched in practice). This can be reduced when upsampling with integer scales \\( s \\), as is common in practice. The Hyper-network output is aggregated into a tensor that is dot multiplied against the unfolded nearest-neighbor interpolation, computing a depth-wise convolution with per-pixel filters. The number of encoder channels, \\( C_e \\), is dependent on the encoder used. In our experiments, we set the number of hypernetwork channels, \\( C_h = 32 \\)." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.605, + 0.47, + 0.633 + ], + "angle": 0, + "content": "resolution feature map into the final output using the corresponding weights." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.632, + 0.47, + 0.754 + ], + "angle": 0, + "content": "- In local implicit image function LIIF [6], the author eliminates the use of a hyper-network (MetaSR, step 3) by extending the sampled target resolution features (step 2) with extra channels representing the relative coordinates and target scale. Their combined features are further processed through a stack of five dense layers inspired by the multi-layer perceptron (MLP) architecture typically adopted as the end-to-end model in neural-fields formulations." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.757, + 0.471, + 0.868 + ], + "angle": 0, + "content": "- In local texture estimator LTE [19] the authors enhance the features fed as input to the MLP layers. In their model, the output of the encoder is processed by three extra trainable layers associated with the amplitude, frequency and phase of sin/cosine waves. The resulting projection in the target resolution layer is used as the input to an MLP that stacks 4 fully-connected layers. In order to further improve the results they also adopt a global skip" + }, + { + "type": "list", + "bbox": [ + 0.078, + 0.632, + 0.471, + 0.868 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.224, + 0.893, + 0.279 + ], + "angle": 0, + "content": "connection with a bilinear up-scaled version of the input around the full model, such that the deep model focus on the computation of the residual between to the closed form approximation and the final result." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.28, + 0.893, + 0.363 + ], + "angle": 0, + "content": "A visualization of these architectures can be found in Figure 2. In practice, LIIF and LTE also increase the number of feature channels as their MLP layers have more neurons (256) than the number of channels produced as the encoder output (64 for encoders such as EDSR [21] and RDN [40]). In contrast to these previous works, our upsampler:" + }, + { + "type": "text", + "bbox": [ + 0.501, + 0.364, + 0.893, + 0.432 + ], + "angle": 0, + "content": "- Moves the computational burden of arbitrary upsampling into the continuous upsampler operator while reducing the operations performed in the target resolution space, preserving the number of channels produced by the deep encoder (Table 1)." + }, + { + "type": "text", + "bbox": [ + 0.501, + 0.433, + 0.892, + 0.488 + ], + "angle": 0, + "content": "- Requires fewer operations than the single-scale sub-pixel convolution in the most typical application case of upsampling an image by an integer scale factor, and producing results of comparable quality." + }, + { + "type": "text", + "bbox": [ + 0.501, + 0.488, + 0.892, + 0.543 + ], + "angle": 0, + "content": "- Adopt a neural-fields formulation that focuses the sensitivity of our hyper-network to fine-grained changes in the representation of scale and relative position (reducing spectral bias)." + }, + { + "type": "list", + "bbox": [ + 0.501, + 0.364, + 0.893, + 0.543 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.555, + 0.591, + 0.569 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.577, + 0.893, + 0.757 + ], + "angle": 0, + "content": "Given a target scale \\( s \\), we aim to produce an upscaled image of size \\( sH \\times sW \\times 3 \\) given an input image of size \\( H \\times W \\times 3 \\). Within the context of super-resolution, our core contribution is the introduction of a novel learnable upsampling layer. The upsampling layer can be interpreted as a decoder in a classical encoder-decoder architecture. We briefly review our encoder architecture in Section 3.1, and detail our decoder in Section 3.2. Note that our analysis focuses on the decoder, as a variety of encoders can be used. In Section 3.4, we then perform a conceptual comparison between our architecture, shown in Figure 3, and others based on neural fields, which are visually summarized in Figure 2." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.762, + 0.894, + 0.819 + ], + "angle": 0, + "content": "Training. Our network with parameters \\(\\theta\\) is trained to regress a high resolution image \\(\\tilde{\\mathbf{I}}^s\\) matching the ground truth \\(\\tilde{\\mathbf{I}}_{\\mathrm{gt}}^s\\) at random positions \\(\\mathbf{x}\\) given the low resolution input image \\(\\mathbf{I}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.555, + 0.829, + 0.892, + 0.853 + ], + "angle": 0, + "content": "\\[\n\\underset {\\boldsymbol {\\theta}} {\\arg \\min} \\mathbb {E} _ {\\mathbf {I}} \\mathbb {E} _ {s} \\mathbb {E} _ {\\mathbf {x}} \\| \\tilde {\\mathbf {I}} ^ {s} (\\mathbf {x}; \\mathbf {I}, \\boldsymbol {\\theta}) - \\tilde {\\mathbf {I}} _ {\\mathrm {g l}} ^ {s} (\\mathbf {x}) \\| _ {2} ^ {1} \\qquad (1)\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.466, + 0.931, + 0.505, + 0.943 + ], + "angle": 0, + "content": "10001" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.078, + 0.13, + 0.465, + 0.18 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.182, + 0.47, + 0.271 + ], + "angle": 0, + "content": "Figure 4. Continuous super-resolution - Our architecture directly allows for an arbitrary and continuous choice of the superresolution factor. Conversely, traditional methods perform this indirectly: they need to upsample to an integer factor, and then consequently downsample to the target resolution. More examples are available in the appendix. Note that \\( \\times 4.5 \\) and \\( \\times 5 \\) are upsampling factors not available in the training data." + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.272, + 0.21, + 0.363 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.125, + 0.366, + 0.168, + 0.375 + ], + "angle": 0, + "content": "all filters" + }, + { + "type": "image", + "bbox": [ + 0.214, + 0.272, + 0.345, + 0.363 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.23, + 0.365, + 0.329, + 0.375 + ], + "angle": 0, + "content": "filter #2 - continuous" + }, + { + "type": "image", + "bbox": [ + 0.347, + 0.272, + 0.478, + 0.363 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.363, + 0.365, + 0.462, + 0.375 + ], + "angle": 0, + "content": "filter #2 - discretized" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.381, + 0.47, + 0.457 + ], + "angle": 0, + "content": "Figure 5. Continuous filters - CUF filters are continuous functions stored in neural fields. We visualize our filters by densely sampling each on the \\([0,1]^2\\) domain to highlight their continuous nature (left). Close-up of an exemplar filter (middle). Filters are discretized by sampling at pixel centers (right), allowing them to be used in classical fixed-scale super-res architectures." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.466, + 0.18, + 0.479 + ], + "angle": 0, + "content": "3.1.Encoder" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.487, + 0.47, + 0.515 + ], + "angle": 0, + "content": "Towards this objective, we first process our input image with an encoder producing \\( C \\)-dimensional features:" + }, + { + "type": "equation", + "bbox": [ + 0.174, + 0.519, + 0.469, + 0.536 + ], + "angle": 0, + "content": "\\[\n\\mathcal {E} _ {\\boldsymbol {\\theta}}: \\mathbb {R} ^ {H \\times W \\times 3} \\rightarrow \\mathbb {R} ^ {H \\times W \\times C} \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.542, + 0.47, + 0.598 + ], + "angle": 0, + "content": "and then unfold a \\(k \\times k\\) spatial neighborhood of \\(C\\)-dimensional features into a tensor of \\(C \\times k^2\\) channels. Finally, we upsample the encoded image using nearest-neighbour interpolation. We define this step as:" + }, + { + "type": "equation", + "bbox": [ + 0.157, + 0.603, + 0.469, + 0.62 + ], + "angle": 0, + "content": "\\[\n\\mathcal {U}: \\mathbb {R} ^ {H \\times W \\times C} \\rightarrow \\mathbb {R} ^ {s H \\times s W \\times C \\times k ^ {2}} \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.627, + 0.469, + 0.695 + ], + "angle": 0, + "content": "The unfolding part of this procedure leads to feature maps that will allow us to implement depth-wise, spatial convolutions as dot products, as is typically done for low-level neural network implementations. We define this whole feature extraction procedure as:" + }, + { + "type": "equation", + "bbox": [ + 0.222, + 0.702, + 0.469, + 0.717 + ], + "angle": 0, + "content": "\\[\n\\mathbf {F} = \\mathcal {U} \\left(\\mathcal {E} _ {\\theta} (\\mathbf {I})\\right) \\tag {4}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.728, + 0.454, + 0.743 + ], + "angle": 0, + "content": "3.2. Decoder – “Continuous Upsampling Filters”" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.469, + 0.846 + ], + "angle": 0, + "content": "We draw our inspiration from classical sub-pixel convolutions [31], and combine it with recent research that apply neural fields for continuous super-resolution [6, 19]. Differently from these architectures, we achieve this objective by instantiating upsampling kernels \\(\\kappa\\) via hypernetworks [13]. Our super-resolution network is algebraically expressed as:" + }, + { + "type": "equation", + "bbox": [ + 0.127, + 0.852, + 0.469, + 0.868 + ], + "angle": 0, + "content": "\\[\n\\tilde {\\mathbf {I}} ^ {s} (\\mathbf {x}; \\mathbf {I}, \\boldsymbol {\\theta}) = \\mathcal {D} _ {\\boldsymbol {\\theta}} \\left(\\mathcal {K} _ {\\boldsymbol {\\theta}} \\left(\\delta^ {s} (\\mathbf {x}); s\\right) \\cdot \\mathbf {F} (\\lfloor \\mathbf {x} / s \\rfloor)\\right) \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.127, + 0.892, + 0.196 + ], + "angle": 0, + "content": "where \\(\\mathbf{x}\\) indexes pixel locations in the target resolution, and \\(\\delta^s (\\mathbf{x}) = \\mathrm{mod}(\\mathbf{x},s) / s\\) renders our convolutional filters translation invariant, and the dot product implements a spatial convolution. The network \\(\\mathcal{D}\\) is a point-wise layer that maps (super-resolved) features back to RGB values:" + }, + { + "type": "equation", + "bbox": [ + 0.642, + 0.207, + 0.892, + 0.223 + ], + "angle": 0, + "content": "\\[\n\\mathcal {D} _ {\\theta}: \\mathbb {R} ^ {C} \\rightarrow \\mathbb {R} ^ {3} \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.228, + 0.892, + 0.299 + ], + "angle": 0, + "content": "This is applied to a grid of coordinates of size \\(\\mathbb{R}^{sH\\times sW\\times C}\\) to get an image of size \\(\\mathbb{R}^{sH\\times sW\\times 3}\\). The coordinate \\(\\mathbf{x}\\) is continuous, hence \\(\\kappa_{\\theta}\\) is a neural field parameterization of a convolutional kernel mapping (continuous) spatial offsets and (continuous) scales to convolutional weights:" + }, + { + "type": "equation", + "bbox": [ + 0.596, + 0.305, + 0.892, + 0.323 + ], + "angle": 0, + "content": "\\[\n\\mathcal {K} _ {\\boldsymbol {\\theta}}: [ 0, 1 ] ^ {2} \\times \\mathbb {R} ^ {+} \\rightarrow \\mathbb {R} ^ {C \\times K ^ {2}} \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.337, + 0.892, + 0.406 + ], + "angle": 0, + "content": "Continuous kernel indexing. We can take our continuous formulation a step further by introducing a continuous parametrization of its kernel indexes \\( k_{i}, k_{j} \\) indexing our \\( K \\times K \\) convolution weight entries. Thus, the hyper-network representing the convolution field becomes:" + }, + { + "type": "equation", + "bbox": [ + 0.545, + 0.413, + 0.892, + 0.43 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathcal {K}} _ {\\boldsymbol {\\theta}}: [ 0, 1 ] ^ {2} \\times \\{0, \\dots , K - 1 \\} ^ {2} \\times \\mathbb {R} ^ {+} \\rightarrow \\mathbb {R} ^ {C} \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.437, + 0.892, + 0.492 + ], + "angle": 0, + "content": "and \\(\\kappa_{\\theta}\\) can then be constructed by 9 invocations to \\(\\hat{\\kappa}_{\\theta}\\) in the case of a \\(K = 3\\) and \\(K^2\\) invocations in the general setting. We use \\(K = 3\\) throughout the paper in order to facilitate a direct comparison to pre-existing baseline models." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.493, + 0.892, + 0.604 + ], + "angle": 0, + "content": "Equation (7) and Equation (8) are both valid ways of generating spatial kernels. The first can be seen as a multi-headed hypernetwork, while the second uses input conditioning to generate the kernel values. We use the latter in our experiments, as the layers of nonlinearities provide additional expressiveness compared to the linear transformation used in the multi-headed version. A comparison between the two is presented in Appendix D." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.611, + 0.687, + 0.626 + ], + "angle": 0, + "content": "3.3. Positional encoding" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.632, + 0.892, + 0.687 + ], + "angle": 0, + "content": "Naively representing signals with MLPs leads to \"spectral bias\" [26]: the overall difficulty these networks have in representing high-frequency signals. Hence we implement \\(\\kappa_{\\theta}\\) as:" + }, + { + "type": "equation", + "bbox": [ + 0.527, + 0.696, + 0.892, + 0.712 + ], + "angle": 0, + "content": "\\[\n\\mathcal {K} _ {\\boldsymbol {\\theta}} \\left(\\delta^ {s} (\\mathbf {x}), s, k\\right) = \\operatorname {M L P} \\left(\\Pi \\left(\\delta^ {s} (\\mathbf {x})\\right), \\Pi (s), \\Pi (k)\\right) \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.719, + 0.892, + 0.829 + ], + "angle": 0, + "content": "where we apply positional encoding \\(\\Pi (\\cdot)\\) to the MLP inputs. Several variants of positional encoding exist, including Fourier [24] and random Fourier [32] variants. However, as we strive for efficiency, we take inspiration from classical signal processing (e.g. JPEG) and instead employ a cosine-only transformation. Given a scalar \\(f_{n}\\) sampled uniformly in the range \\([0,f_{\\mathrm{max}}]\\), a scalar quantity \\(z\\in [0,1]\\) is encoded as the (sorted) vector:" + }, + { + "type": "equation", + "bbox": [ + 0.571, + 0.836, + 0.892, + 0.871 + ], + "angle": 0, + "content": "\\[\n\\Pi (z) = \\left\\{\\cos \\left(\\frac {(2 z + 1) f _ {n} \\pi}{2}\\right) \\right\\} _ {n = 1} ^ {N} \\tag {10}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.466, + 0.931, + 0.506, + 0.942 + ], + "angle": 0, + "content": "10002" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.082, + 0.123, + 0.894, + 0.269 + ], + "angle": 0, + "content": "
EncoderUpsamplerEns.Set5Set14BSD100Urban100
×2×3×4×6×2×3×4×6×2×3×4×6×2×3×4×6
RDNfixed-scale [40]38.2434.7132.47-34.0130.5728.81-32.3429.2627.72-32.8928.8026.61-
MetaSR [16]38.2234.6332.3829.0433.9830.5428.7826.5132.3329.2627.7125.9032.9228.8226.5523.99
CUF (ours)38.2334.7232.5429.2533.9930.5828.8626.7032.3529.2927.7625.9933.0128.9126.7524.23
fixed-scale [40]+geo38.3034.7832.61-34.1030.6728.92-32.4029.3327.75-33.0929.0026.82-
LIIF [6]+loc38.1734.6832.5029.1533.9730.5328.8026.6432.3229.2627.7425.9832.8728.8226.6824.20
LTE [19]+loc38.2334.7232.6129.3234.0930.5828.8826.7132.3629.3027.7726.0133.0428.9726.8124.28
CUF (ours)+geo38.2834.8032.6329.2734.0830.6528.9226.7432.3929.3327.8026.0333.1629.0526.8724.32
SwinIRfixed-scale [31]38.3534.8932.72-34.1430.7728.94-32.4429.3727.83-33.4029.2927.07-
MetaSR [16]38.2634.7732.4729.0934.1430.6628.8526.5832.3929.3127.7525.9433.2929.1226.7624.16
CUF (ours)38.3434.8832.8029.5334.2930.7929.0226.8532.4529.3827.8526.0933.5429.4527.2424.62
fixed-scale [31]+geo38.3834.9532.81-34.2430.8329.02-32.4729.4127.87-33.5129.4227.21-
LIIF [6]+loc38.2834.8732.7329.4634.1430.7528.9826.8232.3929.3427.8426.0733.3629.3327.1524.59
LTE+1c [19]+loc38.3334.8932.8129.5034.2530.8029.0626.8632.4429.3927.8626.0933.5029.4127.2424.62
CUF (ours)++geo38.3834.9232.8329.5734.3330.8429.0526.9132.4729.4027.8826.1133.6529.5527.3224.69
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.273, + 0.893, + 0.316 + ], + "angle": 0, + "content": "Table 2. Out-of-domain evaluation - performance on datasets not seen during training. Metrics computed on the luminance channel in the YCbCr color space, following previous work. Models using self-ensemble are marked with \\( +_{geo} \\) and models using local self-ensemble are marked with \\( +_{loc} \\). Legend: best \\( 2^{\\text{nd}} \\) best" + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.325, + 0.468, + 0.552 + ], + "angle": 0, + "content": "
Multi-scale up-sampling methods - DIV2k
EncoderUpsamplerEns.seen scalesunseen scales
×2×3×4×6×12
Bicubic31.0128.2226.6624.8222.27
EDSR-b.Sub-pixel conv.34.6930.9428.97--
MetaSR34.6430.9328.9226.6123.55
CUF (ours)34.7030.9929.0126.7623.73
Sub-pixel conv.+geo34.7831.0329.06--
LIIF+loc34.6730.9629.0026.7523.71
LTE+loc34.7231.0229.0426.8123.78
CUF (ours)+geo34.7931.0729.0926.8223.78
RDNSub-pixel conv.35.0131.2229.20--
MetaSR35.0031.2729.2526.8823.73
CUF (ours)35.0331.3129.3227.0323.94
Sub-pixel conv.+geo35.1031.3329.31--
LIIF+loc34.9931.2629.2726.9923.89
LTE+loc35.0431.3229.3327.0423.95
CUF (ours)+geo35.1131.3929.3927.0923.99
SwinIRSub-pixel conv.35.2831.4729.40--
MetaSR35.1531.4029.3326.9423.80
CUF (ours)35.2631.5229.5227.1924.07
Sub-pixel conv.+geo35.3331.5229.44--
LIIF+loc35.1731.4629.4627.1524.02
LTE+loc35.2431.5029.5127.2024.09
CUF (ours)+geo35.3131.5629.5627.2324.10
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.556, + 0.47, + 0.625 + ], + "angle": 0, + "content": "Table 3. In domain evaluation: tests on DIV2K's validation subset [25] on scales seen \\((2\\times -4\\times)\\) and unseen during training \\((6\\times -30\\times)\\). Metrics taken on the RGB space. Results taken using geometric self-ensemble [21] are marked with \\(^+\\) geo' and local-ensemble [6] with \\(^+\\) loc'. Legend: best \\(2^{\\mathrm{nd}}\\) best" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.637, + 0.47, + 0.694 + ], + "angle": 0, + "content": "By eliminating the imaginary component from the Fourier basis, we show in Section 4.3 how this reduces the number of trainable parameters of the first hyper-network layer by half, without affecting super-resolution quality." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.701, + 0.181, + 0.716 + ], + "angle": 0, + "content": "3.4. Analysis" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.723, + 0.47, + 0.806 + ], + "angle": 0, + "content": "We now perform a conceptual comparison of our network architecture to the commonly used Sub-Pixel Convolution, as well as several super-resolution approaches based on neural-fields. We further note how continuous upsampling filters can be instantiated to generate filters compatible with Sub-Pixel Convolution." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.811, + 0.471, + 0.868 + ], + "angle": 0, + "content": "Comparison to Sub-Pixel Convolution [31]. Sub-Pixel Convolution is the most commonly used super-resolution operator, especially when efficiency is critical. This is achieved by initially employing an expansion convolution" + }, + { + "type": "image", + "bbox": [ + 0.599, + 0.324, + 0.797, + 0.4 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.403, + 0.892, + 0.428 + ], + "angle": 0, + "content": "Figure 6. A simple illustration of periodic shuffling used in SubPixel Convolution." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.438, + 0.893, + 0.509 + ], + "angle": 0, + "content": "that generates a feature map with \\( s^2 N_{out} \\) channels, where \\( N_{out} \\) is a hyper parameter defining the target number of channels of the Sub-Pixel Convolution operation. Next, the periodic shuffling operator re-arranges the channels to produce a higher-resolution feature map (see Figure 6):" + }, + { + "type": "equation", + "bbox": [ + 0.583, + 0.513, + 0.891, + 0.531 + ], + "angle": 0, + "content": "\\[\n\\mathcal {P}: \\mathbb {R} ^ {H \\times W \\times C} \\rightarrow \\mathbb {R} ^ {s H \\times s W \\times C / s ^ {2}} \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.538, + 0.893, + 0.704 + ], + "angle": 0, + "content": "Similar to (6), whenever \\( N_{out} \\) is chosen as different from the number of output colors, the resulting feature map is further projected by point-wise convolutions. Often, \\( N_{out} \\) is taken as the same as the number of channels as produced by the encoder \\( \\mathcal{E} \\) for architectures adopting Sub-Pixel Convolution and targeting high quality results [20,21,40]. We note that the Sub-Pixel Convolution design based on the expansion convolution does not enforce spatial correlations between neighboring pixels (here we refer to neighborhood in the output, high resolution, domain) and need to be learned from training data. Conversely, these correlations are inherently captured by our continuous convolutional filters." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.704, + 0.892, + 0.802 + ], + "angle": 0, + "content": "Further, in the integer scale setup where hyper-networks can be pre-instantiated, CUFs are computationally more efficient than sub-pixel convolutions whenever \\( K^2 + N_{in} + N_{out} < N_{out} \\times K^2 \\) (in the worst case scenario) as the cost associated with pre-computing the weights for a given integer up-sampling scale can be neglected when compared to the operations performed on the image grid (Appendix E)." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.806, + 0.894, + 0.849 + ], + "angle": 0, + "content": "Comparisons to LIIF [6] and LTE [19]. In comparison to these works, we make the main network shallower (1 depth-wise continuous convolution and 2 dense layers) in order to" + }, + { + "type": "page_footnote", + "bbox": [ + 0.516, + 0.854, + 0.878, + 0.867 + ], + "angle": 0, + "content": "also called pixel-shuffle (PyTorch) or depth-to-space (TensorFlow)." + }, + { + "type": "page_number", + "bbox": [ + 0.466, + 0.931, + 0.505, + 0.942 + ], + "angle": 0, + "content": "10003" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.126, + 0.473, + 0.334 + ], + "angle": 0, + "content": "reduce the number of layers that operate at the target spatial resolution and at the encoder's channel resolution. At the same time, our upsampler head is considerably lighter than previous arbitrary scale methods, not only by the use of a depth-wise convolution but also by keeping the same number of channels produced by the features encoder in the main network (\\( C_e = 64 \\), which results in dense layers \\( 16 \\times \\) cheaper than its LIIF and LTE counterparts using 256 neurons). When performing non-integer upsampling, the costs with processing the hyper-network layers are proportional to the target image resolution, but still smaller than a single dense layer of LIFF and LTE heads, due to the adoption of a reduced number of channels (as \\( C_h = 32 \\), each of CUF's hypernetwork dense layers is \\( 64 \\times \\) cheaper than LIIF and LTE layers using 256 neurons)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.34, + 0.473, + 0.576 + ], + "angle": 0, + "content": "Instantiating CUFs. At inference time, when targeting an integer upscaling factor \\( s \\), the hyper-network representing \\( \\mathcal{K} \\) can be queried to retrieve the weights corresponding to \\( s^2 \\) relative subpixel positions as an initialization step during pre-processing. The retrieved weights are re-used across all pixels taking advantage of the existent periodicity. Thus, in the CUF-instantiated architecture the continuous kernel is replaced at test time with a discrete depth-wise convolution, followed by a pixel shuffling operation, in contrast with the unfolding operator used in the regular, fully continuous CUF described in this paper. Otherwise, the architecture remains the same. The costs associated with the hyper network at initialization can be neglected as \\( s^2 << sH \\times sW \\). In this setting, our model becomes as efficient as a sub-pixel-convolution architecture, while retaining the aforementioned continuous modeling properties at training time." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.588, + 0.165, + 0.602 + ], + "angle": 0, + "content": "4. Results" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.609, + 0.47, + 0.693 + ], + "angle": 0, + "content": "In this section we describe our experimental setup in terms of backbone encoders and training/validation dataset, perform careful comparisons to the state-of-the-art (Section 4.1), and additional evaluation towards the implementation of lightweight super-res architectures (Section 4.2). We conclude by performing a thorough ablation (Section 4.3)." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.698, + 0.151, + 0.71 + ], + "angle": 0, + "content": "Encoders." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.713, + 0.47, + 0.754 + ], + "angle": 0, + "content": "We apply CUFs to a variety of encoders, to both show its generality, as well as its performance gains in a variety of settings." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.756, + 0.469, + 0.783 + ], + "angle": 0, + "content": "- A state-of-the-art encoder for super-res named SwinIR [20] based on Swin-transformers [22]." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.783, + 0.469, + 0.81 + ], + "angle": 0, + "content": "- Well-known convolution-based encoders, namely EDSR-baseline [21] and Residual Dense Networks (RDN) [40]." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.811, + 0.469, + 0.852 + ], + "angle": 0, + "content": "- Architectures for lightweight and extremely lightweight inference, respectively SwinT-lightweight [20] and ABPN [10]." + }, + { + "type": "list", + "bbox": [ + 0.078, + 0.756, + 0.469, + 0.852 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.853, + 0.47, + 0.867 + ], + "angle": 0, + "content": "Table 4 contrasts the ablated encoders' composition and" + }, + { + "type": "table", + "bbox": [ + 0.505, + 0.123, + 0.89, + 0.215 + ], + "angle": 0, + "content": "
MethodCompositionParameters (in M)
encoderspc. 4×
RDNConvolutions22.000.30
SwinTConv. and Self-Attention11.600.30
EDSR-baselineConvolutions1.200.30
SwinT-lightweightConv. and Self-Attention0.900.03
ABPNConvolutions0.030.03
" + }, + { + "type": "table_caption", + "bbox": [ + 0.498, + 0.219, + 0.892, + 0.271 + ], + "angle": 0, + "content": "Table 4. Encoders - Our experiments use CUF layers in combination with various encoders differing in composition and size. Parameter counts reflect the encoder size and the upsampling head using a Sub-Pixel Convolution (spc) targeting \\(4 \\times\\) upsampling." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.283, + 0.892, + 0.298 + ], + "angle": 0, + "content": "size. Hyperparameter settings can be found in the appendix." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.302, + 0.893, + 0.482 + ], + "angle": 0, + "content": "Datasets. All models are trained using the DIV2K dataset training subset (800k images), introduced at the NTIRE 2017 Super Resolution Challenge [25]. Similarly to previous works, each image is randomly cropped 20 times per epoch and augmented with random flips, and 90 degrees rotations. We report peak signal-to-noise ratio (PSNR) results on the DIV2K validation set (100 images)\\(^4\\), in the RGB space. To understand the generalizability of our experimental results, we test our models on additional well known datasets (Set5 [4] - 5 images; Set14 [38] - 14 images; B100 [23] - 100 images; and Urban100 [17] - 100 images), in which, following previous works, PSNR is measured in the YCbCr luminance channel." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.488, + 0.892, + 0.653 + ], + "angle": 0, + "content": "Ensembling. Whenever remarked by \\( ^{+}_{\\mathrm{geo}} \\), results incorporate the (commonly adopted) Geometric Self-ensemble [21], where the results of rotated versions of the input images are averaged at test time. LIIF and LTE results adopt Local Self-ensemble \\( (^{+}_{\\mathrm{loc}} \\) ), where the result of applying the upsampler at 4 shifted grid points is averaged, resulting in a \\( \\times 4 \\) increase in computational complexity at training/test time. It may also introduce confounding factors to the optimization process that invalidate direct comparison to the fixed-scale upsamplers, as during backpropagation the encoder gradients are averaged. In Appendix B we combine these models." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.662, + 0.757, + 0.676 + ], + "angle": 0, + "content": "4.1. State-of-the-art comparisons" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.683, + 0.892, + 0.835 + ], + "angle": 0, + "content": "Table 2 and Table 3 show the quantitative comparisons, while qualitative results can be observed in Figure 4, Figure 7 and Appendix F as well as on our supplementary website https://cuf-paper.github.io/. Table 3 contains in-domain results: models trained on DIV2k and tested on DIV2k (train/test split), while Table 2 presents out-of-domain results: models are trained on DIV2k, but tested on different datasets. Note that upsampling ratios larger than \\(\\times 4\\) are not provided at training time. The models adopting sub-pixel convolutions are fixed-scale models, that is, one model is trained per target scale and thus do not" + }, + { + "type": "page_footnote", + "bbox": [ + 0.5, + 0.844, + 0.892, + 0.867 + ], + "angle": 0, + "content": "4The models we compare to also do not report results on the test-set, which is no longer publicly available." + }, + { + "type": "footer", + "bbox": [ + 0.466, + 0.931, + 0.506, + 0.942 + ], + "angle": 0, + "content": "10004" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.099, + 0.123, + 0.212, + 0.174 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.116, + 0.174, + 0.196, + 0.182 + ], + "angle": 0, + "content": "Urban100 - #088" + }, + { + "type": "image", + "bbox": [ + 0.214, + 0.124, + 0.326, + 0.175 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.262, + 0.175, + 0.278, + 0.182 + ], + "angle": 0, + "content": "LR" + }, + { + "type": "image", + "bbox": [ + 0.328, + 0.125, + 0.442, + 0.175 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.376, + 0.175, + 0.393, + 0.182 + ], + "angle": 0, + "content": "GT" + }, + { + "type": "image", + "bbox": [ + 0.099, + 0.184, + 0.212, + 0.234 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.111, + 0.235, + 0.202, + 0.242 + ], + "angle": 0, + "content": "SwinIR + LIIF × 4" + }, + { + "type": "image", + "bbox": [ + 0.214, + 0.184, + 0.326, + 0.234 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.225, + 0.235, + 0.316, + 0.242 + ], + "angle": 0, + "content": "SwinIR + LTE × 4" + }, + { + "type": "image", + "bbox": [ + 0.328, + 0.184, + 0.442, + 0.234 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.338, + 0.235, + 0.432, + 0.242 + ], + "angle": 0, + "content": "SwinIR + CUF × 4" + }, + { + "type": "image", + "bbox": [ + 0.099, + 0.244, + 0.212, + 0.294 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.124, + 0.294, + 0.187, + 0.301 + ], + "angle": 0, + "content": "DIV2k - #853" + }, + { + "type": "image", + "bbox": [ + 0.214, + 0.244, + 0.326, + 0.294 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.264, + 0.294, + 0.278, + 0.301 + ], + "angle": 0, + "content": "LR" + }, + { + "type": "image", + "bbox": [ + 0.328, + 0.244, + 0.442, + 0.294 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.377, + 0.294, + 0.393, + 0.301 + ], + "angle": 0, + "content": "GT" + }, + { + "type": "image", + "bbox": [ + 0.099, + 0.303, + 0.212, + 0.353 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.116, + 0.353, + 0.196, + 0.36 + ], + "angle": 0, + "content": "RDN + LIIF × 8" + }, + { + "type": "image", + "bbox": [ + 0.214, + 0.303, + 0.326, + 0.353 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.231, + 0.353, + 0.31, + 0.36 + ], + "angle": 0, + "content": "RDN+LTE×8" + }, + { + "type": "image", + "bbox": [ + 0.328, + 0.303, + 0.442, + 0.353 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.344, + 0.353, + 0.425, + 0.36 + ], + "angle": 0, + "content": "RDN+CUF×8" + }, + { + "type": "image", + "bbox": [ + 0.099, + 0.362, + 0.212, + 0.412 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.116, + 0.413, + 0.195, + 0.42 + ], + "angle": 0, + "content": "Urban100 - #030" + }, + { + "type": "image", + "bbox": [ + 0.214, + 0.362, + 0.326, + 0.412 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.263, + 0.413, + 0.277, + 0.419 + ], + "angle": 0, + "content": "LR" + }, + { + "type": "image", + "bbox": [ + 0.328, + 0.362, + 0.442, + 0.413 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.377, + 0.414, + 0.392, + 0.42 + ], + "angle": 0, + "content": "GT" + }, + { + "type": "image", + "bbox": [ + 0.099, + 0.422, + 0.212, + 0.471 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.113, + 0.472, + 0.199, + 0.479 + ], + "angle": 0, + "content": "EDSR + LIF × 4" + }, + { + "type": "image", + "bbox": [ + 0.214, + 0.422, + 0.326, + 0.471 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.229, + 0.472, + 0.313, + 0.479 + ], + "angle": 0, + "content": "EDSR + LTE × 4" + }, + { + "type": "image", + "bbox": [ + 0.328, + 0.422, + 0.442, + 0.471 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.342, + 0.472, + 0.429, + 0.479 + ], + "angle": 0, + "content": "EDSR + CUF × 4" + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.482, + 0.468, + 0.507 + ], + "angle": 0, + "content": "Figure 7. Qualitative evaluation - CUF combined with different encoders and upsampling scales." + }, + { + "type": "image", + "bbox": [ + 0.096, + 0.513, + 0.211, + 0.565 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.216, + 0.513, + 0.331, + 0.565 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.336, + 0.513, + 0.452, + 0.565 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.571, + 0.468, + 0.597 + ], + "angle": 0, + "content": "Figure 8. Residuals - between ground truth and (from left to right) results from LIIF, LTE and CUF shown in Fig 7." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.608, + 0.468, + 0.705 + ], + "angle": 0, + "content": "generalize to unseen or fractional scales. On the remaining rows we present results of arbitrary-scale upsamplers (MetaSR [16], LIIF [6], LTE [19]), where a single model is trained and tested across different upsampling ratios. Our method is the first arbitrary-scale method to match (or surpass) fixed-scale performance across all encoders in both single-pass inference and self-ensemble scenarios." + }, + { + "type": "table", + "bbox": [ + 0.08, + 0.713, + 0.468, + 0.821 + ], + "angle": 0, + "content": "
UpsamplerUses Periodicity\\( \\times 2 \\)\\( \\times 3 \\)\\( \\times 4 \\)
WCGPUWCGPUWCGPU
EDSR-baseline encoder
Sub-pixel conv. **✓10.96.411.07.311.69.0
CUF inst. (ours) **✓11.16.511.27.611.09.1
LIF75.574.0164.6163.1295.0293.4
LTE54.845.9101.792.5167.0157.8
CUF (ours) - k as input26.718.852.444.081.573.0
CUF (ours) - k as output20.212.227.219.837.630.0
ABPN encoder
Sub-pixel conv. **✓1.90.61.90.61.90.7
CUF inst. (ours) **✓2.10.72.11.12.11.5
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.824, + 0.468, + 0.863 + ], + "angle": 0, + "content": "Table 5. Upsampling \\((256^{2})\\) on a A100 GPU. Wall clock (wc) (i.e., considering both CPU and GPU execution) and GPU time in ms averaged over \\(1k\\) runs." + }, + { + "type": "image", + "bbox": [ + 0.52, + 0.123, + 0.852, + 0.322 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.686, + 0.31, + 0.872, + 0.322 + ], + "angle": 0, + "content": "ABPN + Continuous Upsampler" + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.326, + 0.892, + 0.39 + ], + "angle": 0, + "content": "Figure 9. Architectures (lightweight) - The original ABPN [11] single scale architecture (left) and our arbitrary-scale version using the implicit upsampler (right). Note how the two consecutive \\(3 \\times 3\\) native convolutions are exchanged for a \\(3 \\times 3\\) depth-wise implicit convolution followed by point-wise layers." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.401, + 0.757, + 0.417 + ], + "angle": 0, + "content": "4.2. Lightweight super-resolution" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.423, + 0.892, + 0.532 + ], + "angle": 0, + "content": "The encoders and upsamplers considered so far are not designed to be run on devices with limited memory and power. Running super-resolution algorithms on mobile devices comes with unique challenges due to limited RAM and non-efficient support for certain common operations [18]. The proposed CUF-based upsampler, in combination with a mobile-compatible encoder, can be effectively used for upscaling on mobile devices." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.534, + 0.892, + 0.699 + ], + "angle": 0, + "content": "To test the effectiveness of the CUFs, we adopt a lightweight version of SwinIR (0.9M parameters) [20] as well as one of the best performing super-lightweight architectures for mobile devices (30K parameters) [10] as the encoder. For the latter, the original architecture (Figure 9, left) consists of a lightweight encoder and an upscaling module consisting of a series of convolutions, ReLU's, and a standard periodic shuffle layer. We replace this upscaling module with a CUF-based one (Figure 9, right). With instantiated CUFs, this means simply introducing a depthwise convolution followed by pointwise layers, where the weights are output by the learned CUFs." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.699, + 0.892, + 0.754 + ], + "angle": 0, + "content": "Table 6 illustrates that CUFs provide the same level of accuracy while enabling continuous scale and efficient inference. To our knowledge, this is the first mobile-friendly continuous scale super-resolution architecture." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.763, + 0.611, + 0.776 + ], + "angle": 0, + "content": "4.3. Ablations" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.784, + 0.892, + 0.867 + ], + "angle": 0, + "content": "We run several ablation experiments in order to analyze the effects of various model choices. Here, we detail our choice of using DCT for positional encoding vs Fourier features, as well as an investigation of the redundancy of CUF filters versus subpixel convolution. An analysis on the various conditioning factors of our neural fields (positional en" + }, + { + "type": "page_number", + "bbox": [ + 0.466, + 0.931, + 0.505, + 0.942 + ], + "angle": 0, + "content": "10005" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.415, + 0.124, + 0.556, + 0.133 + ], + "angle": 0, + "content": "Multi scale up-sampling methods" + }, + { + "type": "table", + "bbox": [ + 0.081, + 0.133, + 0.89, + 0.193 + ], + "angle": 0, + "content": "
EncoderUpsamplerSet5Set14BSD100Urban100
x2x3x4x6x2x3x4x6x2x3x4x6x2x3x4x6
SwintIR-light [20]Sub-Pixel Conv.38.1134.5432.35-33.8330.5328.76-32.2729.1927.68-32.5628.5826.48-
CUF (ours)38.1734.6232.4629.0933.9330.5728.8326.6132.3029.2327.7125.9532.7328.6826.5524.11
ABPN [10]Sub-Pixel Conv.37.3033.3931.11-32.9029.6827.97-31.6728.6427.14-30.3726.7724.96-
CUF (ours)37.3533.5531.3128.0632.9629.7428.0425.8631.7228.6927.1925.5230.5326.9225.0523.01
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.196, + 0.89, + 0.223 + ], + "angle": 0, + "content": "Table 6. Lightweight super-res - Our CUFs enable continuous super-resolution with lightweight (SwinIR-lightweight [20]) and mobile-friendly (ABPN [10]) backbones without quality loss." + }, + { + "type": "table", + "bbox": [ + 0.158, + 0.226, + 0.39, + 0.277 + ], + "angle": 0, + "content": "
2×3×4×
Fourier30.4526.8525.00
DCT30.4426.8625.00
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.281, + 0.469, + 0.32 + ], + "angle": 0, + "content": "Table 7. Cosine basis - matching results but using half of sinusoidal projections and neurons at the hyper-network input. PSNR from Urban-100 dataset using a ABPN encoder." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.347, + 0.469, + 0.376 + ], + "angle": 0, + "content": "coding, kernel indices, scale, and target subpixel) can be found in the Supplementary Material." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.382, + 0.469, + 0.466 + ], + "angle": 0, + "content": "Positional encoding. A large body of work uses a Fourier basis for the positional encodings, e.g. [24, 33]. However, we opt to use a DCT basis as this requires half of the sinusoidal projections and neurons at the hyper-network first layer. This trick has been used in other domains, such as random approximations of stationary kernel functions [27]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.467, + 0.469, + 0.523 + ], + "angle": 0, + "content": "Table 7 shows the PSNR, as obtained using a DFT versus a DCT basis, and verify that we do not observe loss in performance, while at the same time saving half of the computation associated with positional encoding operations." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.529, + 0.469, + 0.724 + ], + "angle": 0, + "content": "Redundancy of filters. As noted in Section 3.4, sub-pixel convolution does not enforce spatial correlation between neighboring weights in its convolutional filters, but has to learn them from data. CUF on the other hand generates the filters with a hypernetwork, which intrinsically builds on smoothness due to its functional form. Here, we test this hypothesis on the filters of CUF and sub-pixel convolutions. For this analysis, we use an EDSR encoder with \\(C_e = 64\\), \\(K = 3 \\times 3\\), and \\(s = 3\\). Subpixel convolution maps \\(C_e\\) features to \\(s^2 C_e\\) features, and each group of \\(s^2\\) output features are rearranged to form the upscaled feature space. We therefore consider each group of \\(s^2\\) filters of size \\(K \\times K \\times C_e\\) and do an eigenvalue analysis to determine the compressibility of these filters. This is done by forming \\(C_e\\) separate matrix" + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.736, + 0.273, + 0.819 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.116, + 0.822, + 0.238, + 0.832 + ], + "angle": 0, + "content": "(a) Subpixel Convolutions" + }, + { + "type": "image", + "bbox": [ + 0.278, + 0.736, + 0.468, + 0.82 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.35, + 0.822, + 0.391, + 0.832 + ], + "angle": 0, + "content": "(b) CUF" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.837, + 0.469, + 0.863 + ], + "angle": 0, + "content": "Figure 10. CUF depthwise continuous filters are low rank by construction. Subpixel-convolution fails to find low rank filters." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.228, + 0.892, + 0.312 + ], + "angle": 0, + "content": "ces of size \\( s^2 \\times C_e K^2 \\) and performing PCA on each of them separately, forming 64 separate sets of eigenvalues. We plot this distribution for trained and untrained sub-pixel layers in Figure 10a. A faster decay means more redundancy, as more variance is explained by fewer components. In this case, training adds significant structure to the filters." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.316, + 0.893, + 0.454 + ], + "angle": 0, + "content": "Conversely, each output channel in CUF is computed by \\( s^2 \\) filters of size \\( K \\times K \\). These are not directly comparable to subpixel convolution, but we perform an analogous eigenvalue analysis on trained and untrained CUFs. As shown in Figure 10b, unlike subpixel convolution, CUFs actually become less redundant with training, suggesting that the initial hypernetwork already imposes a very smooth prior on the filters. In other words, CUF filters incorporate spatial correlations from the very beginning due to their functional form." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.478, + 0.627, + 0.492 + ], + "angle": 0, + "content": "5. Conclusions" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.503, + 0.892, + 0.821 + ], + "angle": 0, + "content": "We propose CUF, a computationally efficient modeling of continuous upsampling filters as neural fields. Different from previous arbitrary-scale architectures, the performance gains obtained by our upsampler are not due to increased capacity but rather to a more efficient use of model parameters. A single hyper-network supports filter adaptation across scales while at the same time CUF's upsampling head has fewer parameters than previous arbitrary-scale heads, as well as fewer parameters than a single SubPixel Convolution head at \\(4 \\times\\) upsampling. CUF is the first arbitrary scale model that can be effectively used for upscaling on mobile devices, an area previously dominated by single scale Sub-Pixel Convolution architectures. CUF's architecture is quite general. A natural direction of future work is to investigate its use for architectures with multiple upsampling layers (and the reuse of filters across scales), for different image generation settings such as GANs [12] and diffusion models [15]. Neural fields have done particularly well on higher-dimensional signals. Another line of future work is the use of CUF for efficient super-resolution of video data. Continuous convolution filters are an exciting and increasingly studied area of research [28,29] with many applications as an alternative to traditional convolutions." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.826, + 0.893, + 0.866 + ], + "angle": 0, + "content": "Overall, the presented results aligned with low computational resources suggest CUF offers a practical modeling of continuous filters." + }, + { + "type": "page_number", + "bbox": [ + 0.466, + 0.931, + 0.506, + 0.942 + ], + "angle": 0, + "content": "10006" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.125, + 0.174, + 0.139 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.148, + 0.47, + 0.173 + ], + "angle": 0, + "content": "[1] Manoj Alwani, Han Chen, Michael Ferdman, and Peter Milder. Fused-layer cnn accelerators. 10 2016. 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.175, + 0.47, + 0.212 + ], + "angle": 0, + "content": "[2] Saeed Anwar, Salman Khan, and Nick Barnes. A deep journey into super-resolution: A survey. ACM Computing Surveys (ACMCSUR), 53(3), May 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.213, + 0.471, + 0.263 + ], + "angle": 0, + "content": "[3] Syed Muhammad Arsalan Bashir, Yi Wang, Mahrukh Khan, and Yilong Niu. A comprehensive review of deep learning-based single image super-resolution. PeerJ Computer Science, 7:e621, Jul 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.264, + 0.471, + 0.327 + ], + "angle": 0, + "content": "[4] Marco Bevilacqua, Aline Roumy, Christine Guillemot, and Marie line Alberi Morel. Low-complexity single-image super-resolution based on nonnegative neighbor embedding. In Proceedings of the British Machine Vision Conference, pages 135.1-135.10. BMVA Press, 2012. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.329, + 0.471, + 0.354 + ], + "angle": 0, + "content": "[5] Alexandre Boulch. Convpoint: Continuous convolutions for point cloud processing. Computers & Graphics, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.355, + 0.47, + 0.417 + ], + "angle": 0, + "content": "[6] Yinbo Chen, Sifei Liu, and Xiaolong Wang. Learning continuous image representation with local implicit image function. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8628-8638, 2021. 1, 2, 3, 4, 5, 7, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.418, + 0.471, + 0.481 + ], + "angle": 0, + "content": "[7] Tao Dai, Jianrui Cai, Yongbing Zhang, Shu-Tao Xia, and Lei Zhang. Second-order attention network for single image super-resolution. In 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11057-11066, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.482, + 0.471, + 0.558 + ], + "angle": 0, + "content": "[8] Chao Dong, Chen Change Loy, Kaiming He, and Xiaou Tang. Learning a deep convolutional network for image super-resolution. In David Fleet, Tomas Pajdla, Bernt Schiele, and Tinne Tuytelaars, editors, Computer Vision – ECCV 2014, pages 184–199, Cham, 2014. Springer International Publishing. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.559, + 0.471, + 0.609 + ], + "angle": 0, + "content": "[9] Chao Dong, Chen Change Loy, Kaiming He, and Xiaou Tang. Image super-resolution using deep convolutional networks. IEEE Trans. Pattern Anal. Mach. Intell., 38(2):295-307, feb 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.611, + 0.47, + 0.673 + ], + "angle": 0, + "content": "[10] Zongcai Du, Jie Liu, Jie Tang, and Gangshan Wu. Anchors-based plain net for mobile image super-resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, pages 2494-2502, June 2021. 2, 6, 7, 8, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.675, + 0.47, + 0.737 + ], + "angle": 0, + "content": "[11] Z. Du, J. Liu, J. Tang, and G. Wu. Anchor-based plain net for mobile image super-resolution. In 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), pages 2494-2502, Los Alamitos, CA, USA, jun 2021. IEEE Computer Society. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.738, + 0.47, + 0.789 + ], + "angle": 0, + "content": "[12] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial networks. Communications of the ACM, 63(11):139-144, 2020. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.79, + 0.469, + 0.815 + ], + "angle": 0, + "content": "[13] David Ha, Andrew Dai, and Quoc V Le. Hypernetworks. arXiv preprint arXiv:1609.09106, 2016. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.816, + 0.47, + 0.867 + ], + "angle": 0, + "content": "[14] Pedro Hermosilla, Tobias Ritschel, Pere-Pau Vázquez, Álvar Vinacua, and Timo Ropinski. Monte carlo convolution for learning on non-uniformly sampled point clouds. ACM Trans. Graph., 37(6), dec 2018. 1, 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.148, + 0.471, + 0.867 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.127, + 0.892, + 0.166 + ], + "angle": 0, + "content": "[15] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in Neural Information Processing Systems, 33:6840-6851, 2020. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.168, + 0.894, + 0.23 + ], + "angle": 0, + "content": "[16] Xuecai Hu, Haoyuan Mu, Xiangyu Zhang, Zilei Wang, Tieniu Tan, and Jian Sun. Meta-sr: A magnification-arbitrary network for super-resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2019. 1, 2, 3, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.232, + 0.892, + 0.282 + ], + "angle": 0, + "content": "[17] Jia-Bin Huang, Abhishek Singh, and Narendra Ahuja. Single image super-resolution from transformed self-exemplars. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 5197-5206, 2015. 6, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.284, + 0.892, + 0.398 + ], + "angle": 0, + "content": "[18] A. Ignatov, R. Timofte, M. Denna, A. Younes, A. Lek, M. Ayazoglu, J. Liu, Z. Du, J. Guo, X. Zhou, H. Jia, Y. Yan, Z. Zhang, Y. Chen, Y. Peng, Y. Lin, X. Zhang, H. Zeng, K. Zeng, P. Li, Z. Liu, S. Xue, and S. Wang. Real-time quantized image super-resolution on mobile npus, mobile ai 2021 challenge: Report. In 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), pages 2525-2534, Los Alamitos, CA, USA, jun 2021. IEEE Computer Society. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.4, + 0.892, + 0.45 + ], + "angle": 0, + "content": "[19] Kyong Hwan Jin Jaewon Lee. Local texture estimator for implicit representation function. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022. 1, 2, 3, 4, 5, 7, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.452, + 0.892, + 0.503 + ], + "angle": 0, + "content": "[20] Jingyun Liang, Jiezhang Cao, Guolei Sun, Kai Zhang, Luc Van Gool, and Radu Timofte. Swinir: Image restoration using swim transformer. In IEEE International Conference on Computer Vision Workshops, 2021. 1, 2, 5, 6, 7, 8, 11, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.504, + 0.892, + 0.568 + ], + "angle": 0, + "content": "[21] Bee Lim, Sanghyun Son, Heewon Kim, Seungjun Nah, and Young Mu Lee. Enhanced deep residual networks for single image super-resolution. In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, July 2017. 1, 2, 3, 5, 6, 11, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.569, + 0.892, + 0.632 + ], + "angle": 0, + "content": "[22] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.634, + 0.892, + 0.71 + ], + "angle": 0, + "content": "[23] D. Martin, C. Fowlkes, D. Tal, and J. Malik. A database of human segmented natural images and its application to evaluating segmentation algorithms and measuring ecological statistics. In Computer Vision, 2001. ICCV 2001. Proceedings. Eighth IEEE International Conference on, volume 2, pages 416-423, 2001. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.711, + 0.892, + 0.761 + ], + "angle": 0, + "content": "[24] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In ECCV, 2020. 1, 4, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.763, + 0.892, + 0.827 + ], + "angle": 0, + "content": "[25] et al. Radu Timofte. Ntire 2017 challenge on single image super-resolution: Methods and results. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition Workshops, CVPRW 2017, pages 1110-1121, 2017. 5, 6, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.828, + 0.892, + 0.867 + ], + "angle": 0, + "content": "[26] Nasim Rahman, Aristide Baratin, Devansh Arpit, Felix Draxler, Min Lin, Fred Hamprecht, Yoshua Bengio, and Aaron Courville. On the spectral bias of neural networks." + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.127, + 0.894, + 0.867 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.467, + 0.931, + 0.505, + 0.942 + ], + "angle": 0, + "content": "10007" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.127, + 0.47, + 0.19 + ], + "angle": 0, + "content": "In Kamalika Chaudhuri and Ruslan Salakhutdinov, editors, Proceedings of the 36th International Conference on Machine Learning, volume 97 of Proceedings of Machine Learning Research, pages 5301-5310. PMLR, 09-15 Jun 2019. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.191, + 0.47, + 0.228 + ], + "angle": 0, + "content": "[27] Ali Rahimi and Benjamin Recht. Random features for large-scale kernel machines. Advances in neural information processing systems, 20, 2007. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.23, + 0.471, + 0.306 + ], + "angle": 0, + "content": "[28] David W. Romero, Robert-Jan Bruintjes, Jakub Mikelaj Tomczak, Erik J. Bekkers, Mark Hoogendoorn, and Jan van Gemert. Flexconv: Continuous kernel convolutions with differentiable kernel sizes. In The Tenth International Conference on Learning Representations, ICLR 2022, Virtual Event, April 25-29, 2022. OpenReview.net, 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.307, + 0.471, + 0.382 + ], + "angle": 0, + "content": "[29] David W. Romero, Anna Kuzina, Erik J. Bekkers, Jakub Mikolaj Tomczak, and Mark Hoogendoorn. Ckconv: Continuous kernel convolution for sequential data. In The Tenth International Conference on Learning Representations, ICLR 2022, Virtual Event, April 25-29, 2022. OpenReview.net, 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.383, + 0.471, + 0.433 + ], + "angle": 0, + "content": "[30] Chitwan Sahara, Jonathan Ho, William Chan, Tim Salimans, David J Fleet, and Mohammad Norouzi. Image superresolution via iterative refinement. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.434, + 0.471, + 0.534 + ], + "angle": 0, + "content": "[31] Wenzhe Shi, Jose Caballero, Ferenc Huszar, Johannes Totz, Andrew P. Aitken, Rob Bishop, Daniel Rueckert, and Zehan Wang. Real-time single image and video super-resolution using an efficient sub-pixel convolutional neural network. In 2016 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2016, Las Vegas, NV, USA, June 27-30, 2016, pages 1874-1883. IEEE Computer Society, 2016. 1, 2, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.535, + 0.471, + 0.598 + ], + "angle": 0, + "content": "[32] Matthew Tancik, Pratul P. Srinivasan, Ben Mildenhall, Sara Fridovich-Keil, Nithin Raghavan, Utkarsh Singhal, Ravi Ramamoorthi, Jonathan T. Barron, and Ren Ng. Fourier features let networks learn high frequency functions in low dimensional domains. NeurIPS, 2020. 1, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.599, + 0.471, + 0.65 + ], + "angle": 0, + "content": "[33] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.65, + 0.471, + 0.726 + ], + "angle": 0, + "content": "[34] Xintao Wang, Ke Yu, Shixiang Wu, Jinjin Gu, Yihao Liu, Chao Dong, Yu Qiao, and Chen Change Loy. Esrgan: Enhanced super-resolution generative adversarial networks. In Laura Leal-Taixe and Stefan Roth, editors, Computer Vision - ECCV 2018 Workshops, pages 63-79, Cham, 2019. Springer International Publishing. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.727, + 0.471, + 0.776 + ], + "angle": 0, + "content": "[35] Z. Wang, J. Chen, and S. H. Hoi. Deep learning for image super-resolution: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence, 43(10):3365-3387, oct 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.778, + 0.471, + 0.841 + ], + "angle": 0, + "content": "[36] Wenxuan Wu, Zhongang Qi, and Fuxin Li. Pointconv: Deep convolutional networks on 3d point clouds. In IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2019, Long Beach, CA, USA, June 16-20, 2019, pages 9621-9630. Computer Vision Foundation / IEEE, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.842, + 0.471, + 0.868 + ], + "angle": 0, + "content": "[37] Yiheng Xie, Towaki Takikawa, Shunsuke Saito, Or Litany, Shiqin Yan, Numair Khan, Federico Tombari, James Tomp" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.127, + 0.471, + 0.868 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.127, + 0.892, + 0.153 + ], + "angle": 0, + "content": "kin, Vincent Sitzmann, and Srinath Sridhar. Neural fields in visual computing and beyond, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.155, + 0.894, + 0.242 + ], + "angle": 0, + "content": "[38] Roman Zeyde, Michael Elad, and Matan Protter. On single image scale-up using sparse-representations. In Jean-Daniel Boissonnat, Patrick Chenin, Albert Cohen, Christian Gout, Tom Lyche, Marie-Laurence Mazure, and Larry L. Schumaker, editors, Curves and Surfaces, volume 6920 of Lecture Notes in Computer Science, pages 711-730. Springer, 2010. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.244, + 0.894, + 0.32 + ], + "angle": 0, + "content": "[39] Yulun Zhang, Kunpeng Li, Kai Li, Lichen Wang, Bineng Zhong, and Yun Fu. Image super-resolution using very deep residual channel attention networks. In Vittorio Ferrari, Martial Hebert, Cristian Sminchisescu, and Yair Weiss, editors, Computer Vision - ECCV 2018, pages 294-310, Cham, 2018. Springer International Publishing. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.321, + 0.894, + 0.372 + ], + "angle": 0, + "content": "[40] Yulun Zhang, Yapeng Tian, Yu Kong, Bineng Zhong, and Yun Fu. Residual dense network for image restoration. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 43(7):2480-2495, 2021. 1, 2, 3, 5, 6, 11, 12" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.127, + 0.894, + 0.372 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.467, + 0.931, + 0.505, + 0.942 + ], + "angle": 0, + "content": "10008" + } + ] +] \ No newline at end of file diff --git a/2023/CUF_ Continuous Upsampling Filters/5b0cd1d3-2b2b-407b-9bec-e263533bd51d_origin.pdf b/2023/CUF_ Continuous Upsampling Filters/5b0cd1d3-2b2b-407b-9bec-e263533bd51d_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..5c1394f7bdab03165e8d7277b6297a4ba6b95b65 --- /dev/null +++ b/2023/CUF_ Continuous Upsampling Filters/5b0cd1d3-2b2b-407b-9bec-e263533bd51d_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:360607c49534c8642084ba21b2941d2d5529293729280e6e46d5edea13dd9b56 +size 3609585 diff --git a/2023/CUF_ Continuous Upsampling Filters/full.md b/2023/CUF_ Continuous Upsampling Filters/full.md new file mode 100644 index 0000000000000000000000000000000000000000..cc47124131a040a0b4511abdce1299df5435b4b1 --- /dev/null +++ b/2023/CUF_ Continuous Upsampling Filters/full.md @@ -0,0 +1,394 @@ +# CUF: Continuous Upsampling Filters + +Cristina N. Vasconcelos Cengiz Oztireli Mark Matthews + +Milad Hashemi Kevin Swersky Andrea Tagliasacchi + +Google Research + +![](images/5e5adc538e942a4d00d574aba78eb0bad1b5e60eb13a5bd807a8bc06cde65a15.jpg) +Memory Consumption + +![](images/498b52f40a7682f7ba7ddc53672eca71dc97ed84987bf2488a3858a8ee7c327c.jpg) +Super-resolution Accuracy +Figure 1. We report (left) memory and FLOPs for upsampling an $256 \times 256$ image by different scale factors $(2 \times, 3 \times, 4 \times)$ using integer scale upsamplers (Sub-Pixel Convolution [31] and CIF-instantiated) and arbitrary-scale upsamplers (Meta-SR [16], LIIF [6], LTE [19] and CIF), but the same encoder backbone (EDSR-baseline [21]); and (right) the relationship between each upsampler FLOPS and PSNR performance on DIV2k dataset. Our arbitrary scale model is significantly lighter than other methods in the same class (i.e. continuous super-res). Further, when instantiated for integer scale factors, our upsampler is even-more efficient than sub-pixel convolutions [31]. + +# Abstract + +Neural fields have rapidly been adopted for representing 3D signals, but their application to more classical 2D image-processing has been relatively limited. In this paper, we consider one of the most important operations in image processing: upsampling. In deep learning, learnable upsampling layers have extensively been used for single image super-resolution. We propose to parameterize upsampling kernels as neural fields. This parameterization leads to a compact architecture that obtains a 40-fold reduction in the number of parameters when compared with competing arbitrary-scale super-resolution architectures. When upsampling images of size $256 \times 256$ we show that our architecture is $2x - 10x$ more efficient than competing arbitrary-scale super-resolution architectures, and more efficient than sub-pixel convolutions when instantiated to a single-scale model. In the general setting, these gains grow polynomially with the square of the target scale. We validate our method on standard benchmarks showing such efficiency gains can be achieved without sacrifices in super-resolution performance. https://cuf-paper.github.io + +# 1. Introduction + +Neural-fields represent signals with coordinate-based neural-networks. They have found application in a multitude of areas including 3D reconstruction [24], novel-view synthesis [32], convolutions [14], and many others [37]. + +Recent research has investigated the use of neural fields in the context of single image super-resolution [6, 19]. These models are based on multi-layer perceptrons conditioned on latent representation produced by encoders. While such architectures allow for continuous-scale superresolution, they require the execution of a conditional neural field for every pixel at the target resolution, making them unsuitable in applications with limited computational resources. Further, such a large use of resources is not justified by a increase in performance compared to classical convolutional architectures such as sub-pixel convolutions [31]. In summary, neural fields have not yet found widespread adoption as classical solutions are ① trivial to implement and ② more efficient. As they generally perform compa + +rably, their usage is not justified in light of these points; see Figure 1. In this paper, we focus on overcoming these limitations, while noting that (regressive) super-resolution performance is in the saturation regime (i.e. further improvements in image quality seem unlikely without relying on generative modeling [30], and small improvements in PSNR do not necessarily correlate with image quality). + +Our driving hypothesis is that super-resolution convolutional filters are highly correlated - both spatially, as well as across scales. Hence, representing such filters in the latent space of a conditional neural field can effectively capture and compress such correlations. Given that neural fields encode continuous functions within neural networks, we call our filters Continuous Upsampling Filters (CUFs). While neural fields have proven effective in parameterizing 3D convolutions [5,14,36], we demonstrate that there are very significant savings to be had in the parameterization of 2D convolutions. + +In implementing continuous upsampling filters, we draw inspiration from sub-pixel convolutions [31], and realize them via depth-wise convolutions. This not only makes CUFs significantly more efficient than competing continuous super-resolution architectures, but when instantiated to single-image super-resolution they are, surprisingly, even more efficient than "ad-hoc" sub-pixel convolutions with same number of input and output channels; see Figure 1. + +Contributions. We investigate the use of neural fields as a parameterization of convolutional upsampling layers in super-resolution architectures, and show how: + +- The continuity of neural-fields leads to training of compact convolutional architecture for continuous super-resolution (similar super-resolution PSNR but with flop reduction that grows polynomially with the square of the target scale.). +- Instantiating our continuous kernels into their discrete counterpart leads to an efficient inference pipeline, even more efficient in number of operations per target pixel than ad-hoc sub-pixel convolutions with same number of input and output channels. +- Discrete cosine transforms can be used as an efficient replacement for Fourier bases in the implementation of positional encoding for neural fields. +- These gains do not hinder reconstruction performance (PSNR), by carefully validating our method on standard benchmarks, and thoroughly abating our design choices. + +# 2. Related work + +A thorough coverage of single-image super-resolution can be found in the following surveys [2, 3, 35]; in what follows, we provide an overview of classical techniques for super-resolution based on deep learning. As our model is based on neural-fields, we also point the reader to a survey in this topic [37], and below we discuss the existing works + +![](images/9a45162d0acb5af7b4168ea6c9d54d846460fa724d116bbee1664940ae4fea2c.jpg) +Figure 2. Previous Architectures - A visual comparison of the super-resolution architectures of Meta-SR [16], local implicit image function LIIF [6], and local texture estimators (LTE) [19], shown in an isotropic scaling configuration. Components inside the $2 \times 2$ Ensemble box are computed over the four closest pixels to compute a weighted average. The number of encoder channels, $C_e$ , is dependent on the encoder chosen. + +for super-resolution based on neural fields. + +Single scale. There are two main frameworks for super-res, which mainly differ in the placement of upsampling operators within the architecture. + +- In the pre-upsampling framework [8], an initial upscaled image is obtained using a non-trainable upsampler (e.g. bicubic), which is then post-processed by a neural network. This enables arbitrary size/scaling, but can introduce side effects such as noise amplification and blurring. +- In the post-upsampling framework [31], an encoder that preserves the input spatial resolution is followed by a shallow upsampling component. + +Note the computational complexity of pre-upsampling is significant, as these model operate directly at the target resolution. Consequently, post-upsampling is the de-facto mainstream approach [9, 20, 21, 34, 39, 40]. For this reason, we investigate the impact of an implicit upsampler in postupsampling frameworks, and cover a diverse set of architectures ranging from large [22], to extremely lightweight models [10]. + +Neural Fields. The application of neural fields to superresolution [16] has enabled the design of architectures that support multiple target scales, as well as non-integer super-resolution scales. + +- In meta super-resolution MetaSR [16], a hyper-network is conditioned on the target scale factor and on the output pixel relative coordinates (ie. the difference between source and target grids covering the same normalized global space). Their model consists of: (1) transforming the input image using an encoder into deep features at the source resolution; (2) upsampling the deep features into the target space (using nearest-neighbors); (3) using the hyper-network to produce a filter set per target relative position and scale; (4) processing the enlarged target + +
Hyper NetworkUpsampler
input#layers // #neuronsoutputparams (K)inputlayersneuronsparams (K)
Meta-SR [16][δs(x), 1/s]2 // 25664 × 9 × 3445Eθconvolution--
LIIF [6]----[Εθ, δs(x), 1/s]5 dense256347
LTE [19]----ha(Εθ){sin, cos}{δs(x)π(hb(Εθ) + hc(1/s))}4 dense256494
Ours:[Π(δs(x)),Π(s),Π(k)]4 // 32646Eθd-w conv. + 2 dense644
+ +Table 1. Super-resolution with neural-fields: An overview of the main characteristics of neural field based architectures. Square brackets represent concatenation. In LTE [19], $h_a, h_b$ are $3 \times 3$ convolutional layers and $h_c$ is a linear layer. + +![](images/be3372693e5c8fc6e98b078c9b12aca104531ad914a3891667513b6798354c9e.jpg) +Figure 3. Architecture - Continuous Upsampling Filters (CUF) are based on a hyper-network. For each pixel position in the highres space, the hyper-network is computed $K^2$ times, once for each position $i,j$ of the upsampling kernel (batched in practice). This can be reduced when upsampling with integer scales $s$ , as is common in practice. The Hyper-network output is aggregated into a tensor that is dot multiplied against the unfolded nearest-neighbor interpolation, computing a depth-wise convolution with per-pixel filters. The number of encoder channels, $C_e$ , is dependent on the encoder used. In our experiments, we set the number of hypernetwork channels, $C_h = 32$ . + +resolution feature map into the final output using the corresponding weights. + +- In local implicit image function LIIF [6], the author eliminates the use of a hyper-network (MetaSR, step 3) by extending the sampled target resolution features (step 2) with extra channels representing the relative coordinates and target scale. Their combined features are further processed through a stack of five dense layers inspired by the multi-layer perceptron (MLP) architecture typically adopted as the end-to-end model in neural-fields formulations. +- In local texture estimator LTE [19] the authors enhance the features fed as input to the MLP layers. In their model, the output of the encoder is processed by three extra trainable layers associated with the amplitude, frequency and phase of sin/cosine waves. The resulting projection in the target resolution layer is used as the input to an MLP that stacks 4 fully-connected layers. In order to further improve the results they also adopt a global skip + +connection with a bilinear up-scaled version of the input around the full model, such that the deep model focus on the computation of the residual between to the closed form approximation and the final result. + +A visualization of these architectures can be found in Figure 2. In practice, LIIF and LTE also increase the number of feature channels as their MLP layers have more neurons (256) than the number of channels produced as the encoder output (64 for encoders such as EDSR [21] and RDN [40]). In contrast to these previous works, our upsampler: + +- Moves the computational burden of arbitrary upsampling into the continuous upsampler operator while reducing the operations performed in the target resolution space, preserving the number of channels produced by the deep encoder (Table 1). +- Requires fewer operations than the single-scale sub-pixel convolution in the most typical application case of upsampling an image by an integer scale factor, and producing results of comparable quality. +- Adopt a neural-fields formulation that focuses the sensitivity of our hyper-network to fine-grained changes in the representation of scale and relative position (reducing spectral bias). + +# 3. Method + +Given a target scale $s$ , we aim to produce an upscaled image of size $sH \times sW \times 3$ given an input image of size $H \times W \times 3$ . Within the context of super-resolution, our core contribution is the introduction of a novel learnable upsampling layer. The upsampling layer can be interpreted as a decoder in a classical encoder-decoder architecture. We briefly review our encoder architecture in Section 3.1, and detail our decoder in Section 3.2. Note that our analysis focuses on the decoder, as a variety of encoders can be used. In Section 3.4, we then perform a conceptual comparison between our architecture, shown in Figure 3, and others based on neural fields, which are visually summarized in Figure 2. + +Training. Our network with parameters $\theta$ is trained to regress a high resolution image $\tilde{\mathbf{I}}^s$ matching the ground truth $\tilde{\mathbf{I}}_{\mathrm{gt}}^s$ at random positions $\mathbf{x}$ given the low resolution input image $\mathbf{I}$ : + +$$ +\underset {\boldsymbol {\theta}} {\arg \min} \mathbb {E} _ {\mathbf {I}} \mathbb {E} _ {s} \mathbb {E} _ {\mathbf {x}} \| \tilde {\mathbf {I}} ^ {s} (\mathbf {x}; \mathbf {I}, \boldsymbol {\theta}) - \tilde {\mathbf {I}} _ {\mathrm {g l}} ^ {s} (\mathbf {x}) \| _ {2} ^ {1} \qquad (1) +$$ + +![](images/a96e23ecf22c6927002922b940f84b654a448f063839424303ac92c15ed3ac5b.jpg) +Figure 4. Continuous super-resolution - Our architecture directly allows for an arbitrary and continuous choice of the superresolution factor. Conversely, traditional methods perform this indirectly: they need to upsample to an integer factor, and then consequently downsample to the target resolution. More examples are available in the appendix. Note that $\times 4.5$ and $\times 5$ are upsampling factors not available in the training data. + +![](images/5cad7d8416adf878db3d614d692fa1f109362e6cb1b270a1f953906f800fdfc5.jpg) +all filters + +![](images/2e6fdc1a9cf64f469a5a1659ef1e0e9d695bfc25c3f180f6ba5c9018d01b850f.jpg) +filter #2 - continuous + +![](images/272b8837f562fca0ad5791ee96b8cab0237d77ffe09112f927ed64c4f4dd9d5c.jpg) +filter #2 - discretized +Figure 5. Continuous filters - CUF filters are continuous functions stored in neural fields. We visualize our filters by densely sampling each on the $[0,1]^2$ domain to highlight their continuous nature (left). Close-up of an exemplar filter (middle). Filters are discretized by sampling at pixel centers (right), allowing them to be used in classical fixed-scale super-res architectures. + +# 3.1.Encoder + +Towards this objective, we first process our input image with an encoder producing $C$ -dimensional features: + +$$ +\mathcal {E} _ {\boldsymbol {\theta}}: \mathbb {R} ^ {H \times W \times 3} \rightarrow \mathbb {R} ^ {H \times W \times C} \tag {2} +$$ + +and then unfold a $k \times k$ spatial neighborhood of $C$ -dimensional features into a tensor of $C \times k^2$ channels. Finally, we upsample the encoded image using nearest-neighbour interpolation. We define this step as: + +$$ +\mathcal {U}: \mathbb {R} ^ {H \times W \times C} \rightarrow \mathbb {R} ^ {s H \times s W \times C \times k ^ {2}} \tag {3} +$$ + +The unfolding part of this procedure leads to feature maps that will allow us to implement depth-wise, spatial convolutions as dot products, as is typically done for low-level neural network implementations. We define this whole feature extraction procedure as: + +$$ +\mathbf {F} = \mathcal {U} \left(\mathcal {E} _ {\theta} (\mathbf {I})\right) \tag {4} +$$ + +# 3.2. Decoder – “Continuous Upsampling Filters” + +We draw our inspiration from classical sub-pixel convolutions [31], and combine it with recent research that apply neural fields for continuous super-resolution [6, 19]. Differently from these architectures, we achieve this objective by instantiating upsampling kernels $\kappa$ via hypernetworks [13]. Our super-resolution network is algebraically expressed as: + +$$ +\tilde {\mathbf {I}} ^ {s} (\mathbf {x}; \mathbf {I}, \boldsymbol {\theta}) = \mathcal {D} _ {\boldsymbol {\theta}} \left(\mathcal {K} _ {\boldsymbol {\theta}} \left(\delta^ {s} (\mathbf {x}); s\right) \cdot \mathbf {F} (\lfloor \mathbf {x} / s \rfloor)\right) \tag {5} +$$ + +where $\mathbf{x}$ indexes pixel locations in the target resolution, and $\delta^s (\mathbf{x}) = \mathrm{mod}(\mathbf{x},s) / s$ renders our convolutional filters translation invariant, and the dot product implements a spatial convolution. The network $\mathcal{D}$ is a point-wise layer that maps (super-resolved) features back to RGB values: + +$$ +\mathcal {D} _ {\theta}: \mathbb {R} ^ {C} \rightarrow \mathbb {R} ^ {3} \tag {6} +$$ + +This is applied to a grid of coordinates of size $\mathbb{R}^{sH\times sW\times C}$ to get an image of size $\mathbb{R}^{sH\times sW\times 3}$ . The coordinate $\mathbf{x}$ is continuous, hence $\kappa_{\theta}$ is a neural field parameterization of a convolutional kernel mapping (continuous) spatial offsets and (continuous) scales to convolutional weights: + +$$ +\mathcal {K} _ {\boldsymbol {\theta}}: [ 0, 1 ] ^ {2} \times \mathbb {R} ^ {+} \rightarrow \mathbb {R} ^ {C \times K ^ {2}} \tag {7} +$$ + +Continuous kernel indexing. We can take our continuous formulation a step further by introducing a continuous parametrization of its kernel indexes $k_{i}, k_{j}$ indexing our $K \times K$ convolution weight entries. Thus, the hyper-network representing the convolution field becomes: + +$$ +\hat {\mathcal {K}} _ {\boldsymbol {\theta}}: [ 0, 1 ] ^ {2} \times \{0, \dots , K - 1 \} ^ {2} \times \mathbb {R} ^ {+} \rightarrow \mathbb {R} ^ {C} \tag {8} +$$ + +and $\kappa_{\theta}$ can then be constructed by 9 invocations to $\hat{\kappa}_{\theta}$ in the case of a $K = 3$ and $K^2$ invocations in the general setting. We use $K = 3$ throughout the paper in order to facilitate a direct comparison to pre-existing baseline models. + +Equation (7) and Equation (8) are both valid ways of generating spatial kernels. The first can be seen as a multi-headed hypernetwork, while the second uses input conditioning to generate the kernel values. We use the latter in our experiments, as the layers of nonlinearities provide additional expressiveness compared to the linear transformation used in the multi-headed version. A comparison between the two is presented in Appendix D. + +# 3.3. Positional encoding + +Naively representing signals with MLPs leads to "spectral bias" [26]: the overall difficulty these networks have in representing high-frequency signals. Hence we implement $\kappa_{\theta}$ as: + +$$ +\mathcal {K} _ {\boldsymbol {\theta}} \left(\delta^ {s} (\mathbf {x}), s, k\right) = \operatorname {M L P} \left(\Pi \left(\delta^ {s} (\mathbf {x})\right), \Pi (s), \Pi (k)\right) \tag {9} +$$ + +where we apply positional encoding $\Pi (\cdot)$ to the MLP inputs. Several variants of positional encoding exist, including Fourier [24] and random Fourier [32] variants. However, as we strive for efficiency, we take inspiration from classical signal processing (e.g. JPEG) and instead employ a cosine-only transformation. Given a scalar $f_{n}$ sampled uniformly in the range $[0,f_{\mathrm{max}}]$ , a scalar quantity $z\in [0,1]$ is encoded as the (sorted) vector: + +$$ +\Pi (z) = \left\{\cos \left(\frac {(2 z + 1) f _ {n} \pi}{2}\right) \right\} _ {n = 1} ^ {N} \tag {10} +$$ + +
EncoderUpsamplerEns.Set5Set14BSD100Urban100
×2×3×4×6×2×3×4×6×2×3×4×6×2×3×4×6
RDNfixed-scale [40]38.2434.7132.47-34.0130.5728.81-32.3429.2627.72-32.8928.8026.61-
MetaSR [16]38.2234.6332.3829.0433.9830.5428.7826.5132.3329.2627.7125.9032.9228.8226.5523.99
CUF (ours)38.2334.7232.5429.2533.9930.5828.8626.7032.3529.2927.7625.9933.0128.9126.7524.23
fixed-scale [40]+geo38.3034.7832.61-34.1030.6728.92-32.4029.3327.75-33.0929.0026.82-
LIIF [6]+loc38.1734.6832.5029.1533.9730.5328.8026.6432.3229.2627.7425.9832.8728.8226.6824.20
LTE [19]+loc38.2334.7232.6129.3234.0930.5828.8826.7132.3629.3027.7726.0133.0428.9726.8124.28
CUF (ours)+geo38.2834.8032.6329.2734.0830.6528.9226.7432.3929.3327.8026.0333.1629.0526.8724.32
SwinIRfixed-scale [31]38.3534.8932.72-34.1430.7728.94-32.4429.3727.83-33.4029.2927.07-
MetaSR [16]38.2634.7732.4729.0934.1430.6628.8526.5832.3929.3127.7525.9433.2929.1226.7624.16
CUF (ours)38.3434.8832.8029.5334.2930.7929.0226.8532.4529.3827.8526.0933.5429.4527.2424.62
fixed-scale [31]+geo38.3834.9532.81-34.2430.8329.02-32.4729.4127.87-33.5129.4227.21-
LIIF [6]+loc38.2834.8732.7329.4634.1430.7528.9826.8232.3929.3427.8426.0733.3629.3327.1524.59
LTE+1c [19]+loc38.3334.8932.8129.5034.2530.8029.0626.8632.4429.3927.8626.0933.5029.4127.2424.62
CUF (ours)++geo38.3834.9232.8329.5734.3330.8429.0526.9132.4729.4027.8826.1133.6529.5527.3224.69
+ +Table 2. Out-of-domain evaluation - performance on datasets not seen during training. Metrics computed on the luminance channel in the YCbCr color space, following previous work. Models using self-ensemble are marked with $+_{geo}$ and models using local self-ensemble are marked with $+_{loc}$ . Legend: best $2^{\text{nd}}$ best + +
Multi-scale up-sampling methods - DIV2k
EncoderUpsamplerEns.seen scalesunseen scales
×2×3×4×6×12
Bicubic31.0128.2226.6624.8222.27
EDSR-b.Sub-pixel conv.34.6930.9428.97--
MetaSR34.6430.9328.9226.6123.55
CUF (ours)34.7030.9929.0126.7623.73
Sub-pixel conv.+geo34.7831.0329.06--
LIIF+loc34.6730.9629.0026.7523.71
LTE+loc34.7231.0229.0426.8123.78
CUF (ours)+geo34.7931.0729.0926.8223.78
RDNSub-pixel conv.35.0131.2229.20--
MetaSR35.0031.2729.2526.8823.73
CUF (ours)35.0331.3129.3227.0323.94
Sub-pixel conv.+geo35.1031.3329.31--
LIIF+loc34.9931.2629.2726.9923.89
LTE+loc35.0431.3229.3327.0423.95
CUF (ours)+geo35.1131.3929.3927.0923.99
SwinIRSub-pixel conv.35.2831.4729.40--
MetaSR35.1531.4029.3326.9423.80
CUF (ours)35.2631.5229.5227.1924.07
Sub-pixel conv.+geo35.3331.5229.44--
LIIF+loc35.1731.4629.4627.1524.02
LTE+loc35.2431.5029.5127.2024.09
CUF (ours)+geo35.3131.5629.5627.2324.10
+ +Table 3. In domain evaluation: tests on DIV2K's validation subset [25] on scales seen $(2\times -4\times)$ and unseen during training $(6\times -30\times)$ . Metrics taken on the RGB space. Results taken using geometric self-ensemble [21] are marked with $^+$ geo' and local-ensemble [6] with $^+$ loc'. Legend: best $2^{\mathrm{nd}}$ best + +By eliminating the imaginary component from the Fourier basis, we show in Section 4.3 how this reduces the number of trainable parameters of the first hyper-network layer by half, without affecting super-resolution quality. + +# 3.4. Analysis + +We now perform a conceptual comparison of our network architecture to the commonly used Sub-Pixel Convolution, as well as several super-resolution approaches based on neural-fields. We further note how continuous upsampling filters can be instantiated to generate filters compatible with Sub-Pixel Convolution. + +Comparison to Sub-Pixel Convolution [31]. Sub-Pixel Convolution is the most commonly used super-resolution operator, especially when efficiency is critical. This is achieved by initially employing an expansion convolution + +![](images/e64749d874d3c55d0d1c9f97deffbb2d76bb2892733a83adcebcc8c6302af96a.jpg) +Figure 6. A simple illustration of periodic shuffling used in SubPixel Convolution. + +that generates a feature map with $s^2 N_{out}$ channels, where $N_{out}$ is a hyper parameter defining the target number of channels of the Sub-Pixel Convolution operation. Next, the periodic shuffling operator re-arranges the channels to produce a higher-resolution feature map (see Figure 6): + +$$ +\mathcal {P}: \mathbb {R} ^ {H \times W \times C} \rightarrow \mathbb {R} ^ {s H \times s W \times C / s ^ {2}} \tag {11} +$$ + +Similar to (6), whenever $N_{out}$ is chosen as different from the number of output colors, the resulting feature map is further projected by point-wise convolutions. Often, $N_{out}$ is taken as the same as the number of channels as produced by the encoder $\mathcal{E}$ for architectures adopting Sub-Pixel Convolution and targeting high quality results [20,21,40]. We note that the Sub-Pixel Convolution design based on the expansion convolution does not enforce spatial correlations between neighboring pixels (here we refer to neighborhood in the output, high resolution, domain) and need to be learned from training data. Conversely, these correlations are inherently captured by our continuous convolutional filters. + +Further, in the integer scale setup where hyper-networks can be pre-instantiated, CUFs are computationally more efficient than sub-pixel convolutions whenever $K^2 + N_{in} + N_{out} < N_{out} \times K^2$ (in the worst case scenario) as the cost associated with pre-computing the weights for a given integer up-sampling scale can be neglected when compared to the operations performed on the image grid (Appendix E). + +Comparisons to LIIF [6] and LTE [19]. In comparison to these works, we make the main network shallower (1 depth-wise continuous convolution and 2 dense layers) in order to + +reduce the number of layers that operate at the target spatial resolution and at the encoder's channel resolution. At the same time, our upsampler head is considerably lighter than previous arbitrary scale methods, not only by the use of a depth-wise convolution but also by keeping the same number of channels produced by the features encoder in the main network ( $C_e = 64$ , which results in dense layers $16 \times$ cheaper than its LIIF and LTE counterparts using 256 neurons). When performing non-integer upsampling, the costs with processing the hyper-network layers are proportional to the target image resolution, but still smaller than a single dense layer of LIFF and LTE heads, due to the adoption of a reduced number of channels (as $C_h = 32$ , each of CUF's hypernetwork dense layers is $64 \times$ cheaper than LIIF and LTE layers using 256 neurons). + +Instantiating CUFs. At inference time, when targeting an integer upscaling factor $s$ , the hyper-network representing $\mathcal{K}$ can be queried to retrieve the weights corresponding to $s^2$ relative subpixel positions as an initialization step during pre-processing. The retrieved weights are re-used across all pixels taking advantage of the existent periodicity. Thus, in the CUF-instantiated architecture the continuous kernel is replaced at test time with a discrete depth-wise convolution, followed by a pixel shuffling operation, in contrast with the unfolding operator used in the regular, fully continuous CUF described in this paper. Otherwise, the architecture remains the same. The costs associated with the hyper network at initialization can be neglected as $s^2 << sH \times sW$ . In this setting, our model becomes as efficient as a sub-pixel-convolution architecture, while retaining the aforementioned continuous modeling properties at training time. + +# 4. Results + +In this section we describe our experimental setup in terms of backbone encoders and training/validation dataset, perform careful comparisons to the state-of-the-art (Section 4.1), and additional evaluation towards the implementation of lightweight super-res architectures (Section 4.2). We conclude by performing a thorough ablation (Section 4.3). + +# Encoders. + +We apply CUFs to a variety of encoders, to both show its generality, as well as its performance gains in a variety of settings. + +- A state-of-the-art encoder for super-res named SwinIR [20] based on Swin-transformers [22]. +- Well-known convolution-based encoders, namely EDSR-baseline [21] and Residual Dense Networks (RDN) [40]. +- Architectures for lightweight and extremely lightweight inference, respectively SwinT-lightweight [20] and ABPN [10]. + +Table 4 contrasts the ablated encoders' composition and + +
MethodCompositionParameters (in M)
encoderspc. 4×
RDNConvolutions22.000.30
SwinTConv. and Self-Attention11.600.30
EDSR-baselineConvolutions1.200.30
SwinT-lightweightConv. and Self-Attention0.900.03
ABPNConvolutions0.030.03
+ +Table 4. Encoders - Our experiments use CUF layers in combination with various encoders differing in composition and size. Parameter counts reflect the encoder size and the upsampling head using a Sub-Pixel Convolution (spc) targeting $4 \times$ upsampling. + +size. Hyperparameter settings can be found in the appendix. + +Datasets. All models are trained using the DIV2K dataset training subset (800k images), introduced at the NTIRE 2017 Super Resolution Challenge [25]. Similarly to previous works, each image is randomly cropped 20 times per epoch and augmented with random flips, and 90 degrees rotations. We report peak signal-to-noise ratio (PSNR) results on the DIV2K validation set (100 images) $^4$ , in the RGB space. To understand the generalizability of our experimental results, we test our models on additional well known datasets (Set5 [4] - 5 images; Set14 [38] - 14 images; B100 [23] - 100 images; and Urban100 [17] - 100 images), in which, following previous works, PSNR is measured in the YCbCr luminance channel. + +Ensembling. Whenever remarked by $^{+}_{\mathrm{geo}}$ , results incorporate the (commonly adopted) Geometric Self-ensemble [21], where the results of rotated versions of the input images are averaged at test time. LIIF and LTE results adopt Local Self-ensemble $(^{+}_{\mathrm{loc}}$ ), where the result of applying the upsampler at 4 shifted grid points is averaged, resulting in a $\times 4$ increase in computational complexity at training/test time. It may also introduce confounding factors to the optimization process that invalidate direct comparison to the fixed-scale upsamplers, as during backpropagation the encoder gradients are averaged. In Appendix B we combine these models. + +# 4.1. State-of-the-art comparisons + +Table 2 and Table 3 show the quantitative comparisons, while qualitative results can be observed in Figure 4, Figure 7 and Appendix F as well as on our supplementary website https://cuf-paper.github.io/. Table 3 contains in-domain results: models trained on DIV2k and tested on DIV2k (train/test split), while Table 2 presents out-of-domain results: models are trained on DIV2k, but tested on different datasets. Note that upsampling ratios larger than $\times 4$ are not provided at training time. The models adopting sub-pixel convolutions are fixed-scale models, that is, one model is trained per target scale and thus do not + +![](images/4586c8a3fd80a808e0b05221866617bec8c08f87663c8ba05ecd8b9f2dec16b8.jpg) +Urban100 - #088 + +![](images/9e58391643991332fe4c329c0477c87267164c8bf4026a6daa07fbb3b998d566.jpg) +LR + +![](images/3b426c19a5f7f535597822b7c85e804457a1fe9190335fed7d35b4e6186efc3a.jpg) +GT + +![](images/c3c0031b00bf782a9266f44315c1319367590822520f5091bf3c5210cfe1467b.jpg) +SwinIR + LIIF × 4 + +![](images/3489f2945e0e53a2fc32604cb84c519932f56a4391e4b97683ed1586da5d6e7d.jpg) +SwinIR + LTE × 4 + +![](images/3002623944c9c75fcf9cf89766f4bcae9509a6357a4ff777aab2d150774f47ef.jpg) +SwinIR + CUF × 4 + +![](images/0231013cff2cc662bb784a6f52051ba45c6603dfb518f79458d4bf4649d26a36.jpg) +DIV2k - #853 + +![](images/b14bb37fa3c7c5e119d28d89350297f149cea13f2cd3e41bc65b34cf1aa666c3.jpg) +LR + +![](images/732f5b867c3ac7f4bb15e239019a7e49fa0acb48d9797a949f0e04c124f7835f.jpg) +GT + +![](images/1b961892fd03dfcdd62a1a80bb8179219db93b92446baf3d6e06843c45cae315.jpg) +RDN + LIIF × 8 + +![](images/41cedf49c99746d1b1cdc083e689ded0c336299f748cb3e145cfa13fd7d0c819.jpg) +RDN+LTE×8 + +![](images/a845e23a0ef19ad4f1cf9defee4a993dab0cd79e60d0843f6750ff55b3e255e6.jpg) +RDN+CUF×8 + +![](images/9308183792fdc1bbce259fe7dfb0f4f3ed5e06699023a8e4bf38fd71835603d0.jpg) +Urban100 - #030 + +![](images/a912ae22bde445f82745922ded934e03c7ce57cec0e86958df726d100d43d4a1.jpg) +LR + +![](images/0f546039c6d66d558b33a3e7ef670331c80ac329e000d15184683c24bde54012.jpg) +GT + +![](images/23d15cc28ecf17fb84b27930ca53c31806084e0d2d98e1a59b2fe41f854199cf.jpg) +EDSR + LIF × 4 + +![](images/000f5e5546064e7592b68cd292b4e5d796b3d1b65449ffad0ab27b59ebfe3cdb.jpg) +EDSR + LTE × 4 + +![](images/2f603dc4fc57483bcddf2a1a09630fc3c1ee252b353a16d7b584cbe3f2a4ed57.jpg) +EDSR + CUF × 4 + +![](images/41369710361a31dac81168f56ea1109c5d3aca11a2b02c3536eaa8c8a2eb3bcd.jpg) +Figure 7. Qualitative evaluation - CUF combined with different encoders and upsampling scales. +Figure 8. Residuals - between ground truth and (from left to right) results from LIIF, LTE and CUF shown in Fig 7. + +![](images/09137e00bd551d75b0d426599bebaab0afe01fbceeaef04d8cbfee5f3fa3671f.jpg) + +![](images/057e4f837f989a538cc20928506cd0c73e11a4e1b5c46c3726585ba596eadb7a.jpg) + +generalize to unseen or fractional scales. On the remaining rows we present results of arbitrary-scale upsamplers (MetaSR [16], LIIF [6], LTE [19]), where a single model is trained and tested across different upsampling ratios. Our method is the first arbitrary-scale method to match (or surpass) fixed-scale performance across all encoders in both single-pass inference and self-ensemble scenarios. + +
UpsamplerUses Periodicity\( \times 2 \)\( \times 3 \)\( \times 4 \)
WCGPUWCGPUWCGPU
EDSR-baseline encoder
Sub-pixel conv. **✓10.96.411.07.311.69.0
CUF inst. (ours) **✓11.16.511.27.611.09.1
LIF75.574.0164.6163.1295.0293.4
LTE54.845.9101.792.5167.0157.8
CUF (ours) - k as input26.718.852.444.081.573.0
CUF (ours) - k as output20.212.227.219.837.630.0
ABPN encoder
Sub-pixel conv. **✓1.90.61.90.61.90.7
CUF inst. (ours) **✓2.10.72.11.12.11.5
+ +Table 5. Upsampling $(256^{2})$ on a A100 GPU. Wall clock (wc) (i.e., considering both CPU and GPU execution) and GPU time in ms averaged over $1k$ runs. + +![](images/07f5c6bfac8c1d35f9c972ab3023874823312fa45b26dfd9e262507cad3e1f1f.jpg) +ABPN + Continuous Upsampler +Figure 9. Architectures (lightweight) - The original ABPN [11] single scale architecture (left) and our arbitrary-scale version using the implicit upsampler (right). Note how the two consecutive $3 \times 3$ native convolutions are exchanged for a $3 \times 3$ depth-wise implicit convolution followed by point-wise layers. + +# 4.2. Lightweight super-resolution + +The encoders and upsamplers considered so far are not designed to be run on devices with limited memory and power. Running super-resolution algorithms on mobile devices comes with unique challenges due to limited RAM and non-efficient support for certain common operations [18]. The proposed CUF-based upsampler, in combination with a mobile-compatible encoder, can be effectively used for upscaling on mobile devices. + +To test the effectiveness of the CUFs, we adopt a lightweight version of SwinIR (0.9M parameters) [20] as well as one of the best performing super-lightweight architectures for mobile devices (30K parameters) [10] as the encoder. For the latter, the original architecture (Figure 9, left) consists of a lightweight encoder and an upscaling module consisting of a series of convolutions, ReLU's, and a standard periodic shuffle layer. We replace this upscaling module with a CUF-based one (Figure 9, right). With instantiated CUFs, this means simply introducing a depthwise convolution followed by pointwise layers, where the weights are output by the learned CUFs. + +Table 6 illustrates that CUFs provide the same level of accuracy while enabling continuous scale and efficient inference. To our knowledge, this is the first mobile-friendly continuous scale super-resolution architecture. + +# 4.3. Ablations + +We run several ablation experiments in order to analyze the effects of various model choices. Here, we detail our choice of using DCT for positional encoding vs Fourier features, as well as an investigation of the redundancy of CUF filters versus subpixel convolution. An analysis on the various conditioning factors of our neural fields (positional en + +Multi scale up-sampling methods + +
EncoderUpsamplerSet5Set14BSD100Urban100
x2x3x4x6x2x3x4x6x2x3x4x6x2x3x4x6
SwintIR-light [20]Sub-Pixel Conv.38.1134.5432.35-33.8330.5328.76-32.2729.1927.68-32.5628.5826.48-
CUF (ours)38.1734.6232.4629.0933.9330.5728.8326.6132.3029.2327.7125.9532.7328.6826.5524.11
ABPN [10]Sub-Pixel Conv.37.3033.3931.11-32.9029.6827.97-31.6728.6427.14-30.3726.7724.96-
CUF (ours)37.3533.5531.3128.0632.9629.7428.0425.8631.7228.6927.1925.5230.5326.9225.0523.01
+ +Table 6. Lightweight super-res - Our CUFs enable continuous super-resolution with lightweight (SwinIR-lightweight [20]) and mobile-friendly (ABPN [10]) backbones without quality loss. + +
2×3×4×
Fourier30.4526.8525.00
DCT30.4426.8625.00
+ +Table 7. Cosine basis - matching results but using half of sinusoidal projections and neurons at the hyper-network input. PSNR from Urban-100 dataset using a ABPN encoder. + +coding, kernel indices, scale, and target subpixel) can be found in the Supplementary Material. + +Positional encoding. A large body of work uses a Fourier basis for the positional encodings, e.g. [24, 33]. However, we opt to use a DCT basis as this requires half of the sinusoidal projections and neurons at the hyper-network first layer. This trick has been used in other domains, such as random approximations of stationary kernel functions [27]. + +Table 7 shows the PSNR, as obtained using a DFT versus a DCT basis, and verify that we do not observe loss in performance, while at the same time saving half of the computation associated with positional encoding operations. + +Redundancy of filters. As noted in Section 3.4, sub-pixel convolution does not enforce spatial correlation between neighboring weights in its convolutional filters, but has to learn them from data. CUF on the other hand generates the filters with a hypernetwork, which intrinsically builds on smoothness due to its functional form. Here, we test this hypothesis on the filters of CUF and sub-pixel convolutions. For this analysis, we use an EDSR encoder with $C_e = 64$ , $K = 3 \times 3$ , and $s = 3$ . Subpixel convolution maps $C_e$ features to $s^2 C_e$ features, and each group of $s^2$ output features are rearranged to form the upscaled feature space. We therefore consider each group of $s^2$ filters of size $K \times K \times C_e$ and do an eigenvalue analysis to determine the compressibility of these filters. This is done by forming $C_e$ separate matrix + +![](images/7df9f14026dccd79d65aa4f295e919f21d391a9292f961e8579a336ce214c660.jpg) +(a) Subpixel Convolutions + +![](images/9847b4ff361a573ff20a3fc325b213fb34a6b6ab58e942ec7dec555bd6e2fb11.jpg) +(b) CUF +Figure 10. CUF depthwise continuous filters are low rank by construction. Subpixel-convolution fails to find low rank filters. + +ces of size $s^2 \times C_e K^2$ and performing PCA on each of them separately, forming 64 separate sets of eigenvalues. We plot this distribution for trained and untrained sub-pixel layers in Figure 10a. A faster decay means more redundancy, as more variance is explained by fewer components. In this case, training adds significant structure to the filters. + +Conversely, each output channel in CUF is computed by $s^2$ filters of size $K \times K$ . These are not directly comparable to subpixel convolution, but we perform an analogous eigenvalue analysis on trained and untrained CUFs. As shown in Figure 10b, unlike subpixel convolution, CUFs actually become less redundant with training, suggesting that the initial hypernetwork already imposes a very smooth prior on the filters. In other words, CUF filters incorporate spatial correlations from the very beginning due to their functional form. + +# 5. Conclusions + +We propose CUF, a computationally efficient modeling of continuous upsampling filters as neural fields. Different from previous arbitrary-scale architectures, the performance gains obtained by our upsampler are not due to increased capacity but rather to a more efficient use of model parameters. A single hyper-network supports filter adaptation across scales while at the same time CUF's upsampling head has fewer parameters than previous arbitrary-scale heads, as well as fewer parameters than a single SubPixel Convolution head at $4 \times$ upsampling. CUF is the first arbitrary scale model that can be effectively used for upscaling on mobile devices, an area previously dominated by single scale Sub-Pixel Convolution architectures. CUF's architecture is quite general. A natural direction of future work is to investigate its use for architectures with multiple upsampling layers (and the reuse of filters across scales), for different image generation settings such as GANs [12] and diffusion models [15]. Neural fields have done particularly well on higher-dimensional signals. Another line of future work is the use of CUF for efficient super-resolution of video data. Continuous convolution filters are an exciting and increasingly studied area of research [28,29] with many applications as an alternative to traditional convolutions. + +Overall, the presented results aligned with low computational resources suggest CUF offers a practical modeling of continuous filters. + +# References + +[1] Manoj Alwani, Han Chen, Michael Ferdman, and Peter Milder. Fused-layer cnn accelerators. 10 2016. 13 +[2] Saeed Anwar, Salman Khan, and Nick Barnes. A deep journey into super-resolution: A survey. ACM Computing Surveys (ACMCSUR), 53(3), May 2020. 2 +[3] Syed Muhammad Arsalan Bashir, Yi Wang, Mahrukh Khan, and Yilong Niu. A comprehensive review of deep learning-based single image super-resolution. PeerJ Computer Science, 7:e621, Jul 2021. 2 +[4] Marco Bevilacqua, Aline Roumy, Christine Guillemot, and Marie line Alberi Morel. Low-complexity single-image super-resolution based on nonnegative neighbor embedding. In Proceedings of the British Machine Vision Conference, pages 135.1-135.10. BMVA Press, 2012. 6 +[5] Alexandre Boulch. Convpoint: Continuous convolutions for point cloud processing. Computers & Graphics, 2020. 2 +[6] Yinbo Chen, Sifei Liu, and Xiaolong Wang. Learning continuous image representation with local implicit image function. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8628-8638, 2021. 1, 2, 3, 4, 5, 7, 11 +[7] Tao Dai, Jianrui Cai, Yongbing Zhang, Shu-Tao Xia, and Lei Zhang. Second-order attention network for single image super-resolution. In 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11057-11066, 2019. 1 +[8] Chao Dong, Chen Change Loy, Kaiming He, and Xiaou Tang. Learning a deep convolutional network for image super-resolution. In David Fleet, Tomas Pajdla, Bernt Schiele, and Tinne Tuytelaars, editors, Computer Vision – ECCV 2014, pages 184–199, Cham, 2014. Springer International Publishing. 1, 2 +[9] Chao Dong, Chen Change Loy, Kaiming He, and Xiaou Tang. Image super-resolution using deep convolutional networks. IEEE Trans. Pattern Anal. Mach. Intell., 38(2):295-307, feb 2016. 2 +[10] Zongcai Du, Jie Liu, Jie Tang, and Gangshan Wu. Anchors-based plain net for mobile image super-resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, pages 2494-2502, June 2021. 2, 6, 7, 8, 11 +[11] Z. Du, J. Liu, J. Tang, and G. Wu. Anchor-based plain net for mobile image super-resolution. In 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), pages 2494-2502, Los Alamitos, CA, USA, jun 2021. IEEE Computer Society. 7 +[12] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial networks. Communications of the ACM, 63(11):139-144, 2020. 8 +[13] David Ha, Andrew Dai, and Quoc V Le. Hypernetworks. arXiv preprint arXiv:1609.09106, 2016. 4 +[14] Pedro Hermosilla, Tobias Ritschel, Pere-Pau Vázquez, Álvar Vinacua, and Timo Ropinski. Monte carlo convolution for learning on non-uniformly sampled point clouds. ACM Trans. Graph., 37(6), dec 2018. 1, 2 + +[15] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in Neural Information Processing Systems, 33:6840-6851, 2020. 8 +[16] Xuecai Hu, Haoyuan Mu, Xiangyu Zhang, Zilei Wang, Tieniu Tan, and Jian Sun. Meta-sr: A magnification-arbitrary network for super-resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2019. 1, 2, 3, 5, 7 +[17] Jia-Bin Huang, Abhishek Singh, and Narendra Ahuja. Single image super-resolution from transformed self-exemplars. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 5197-5206, 2015. 6, 11 +[18] A. Ignatov, R. Timofte, M. Denna, A. Younes, A. Lek, M. Ayazoglu, J. Liu, Z. Du, J. Guo, X. Zhou, H. Jia, Y. Yan, Z. Zhang, Y. Chen, Y. Peng, Y. Lin, X. Zhang, H. Zeng, K. Zeng, P. Li, Z. Liu, S. Xue, and S. Wang. Real-time quantized image super-resolution on mobile npus, mobile ai 2021 challenge: Report. In 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), pages 2525-2534, Los Alamitos, CA, USA, jun 2021. IEEE Computer Society. 7 +[19] Kyong Hwan Jin Jaewon Lee. Local texture estimator for implicit representation function. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022. 1, 2, 3, 4, 5, 7, 11 +[20] Jingyun Liang, Jiezhang Cao, Guolei Sun, Kai Zhang, Luc Van Gool, and Radu Timofte. Swinir: Image restoration using swim transformer. In IEEE International Conference on Computer Vision Workshops, 2021. 1, 2, 5, 6, 7, 8, 11, 12 +[21] Bee Lim, Sanghyun Son, Heewon Kim, Seungjun Nah, and Young Mu Lee. Enhanced deep residual networks for single image super-resolution. In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, July 2017. 1, 2, 3, 5, 6, 11, 12 +[22] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 2, 6 +[23] D. Martin, C. Fowlkes, D. Tal, and J. Malik. A database of human segmented natural images and its application to evaluating segmentation algorithms and measuring ecological statistics. In Computer Vision, 2001. ICCV 2001. Proceedings. Eighth IEEE International Conference on, volume 2, pages 416-423, 2001. 6 +[24] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In ECCV, 2020. 1, 4, 8 +[25] et al. Radu Timofte. Ntire 2017 challenge on single image super-resolution: Methods and results. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition Workshops, CVPRW 2017, pages 1110-1121, 2017. 5, 6, 11 +[26] Nasim Rahman, Aristide Baratin, Devansh Arpit, Felix Draxler, Min Lin, Fred Hamprecht, Yoshua Bengio, and Aaron Courville. On the spectral bias of neural networks. + +In Kamalika Chaudhuri and Ruslan Salakhutdinov, editors, Proceedings of the 36th International Conference on Machine Learning, volume 97 of Proceedings of Machine Learning Research, pages 5301-5310. PMLR, 09-15 Jun 2019. 4 +[27] Ali Rahimi and Benjamin Recht. Random features for large-scale kernel machines. Advances in neural information processing systems, 20, 2007. 8 +[28] David W. Romero, Robert-Jan Bruintjes, Jakub Mikelaj Tomczak, Erik J. Bekkers, Mark Hoogendoorn, and Jan van Gemert. Flexconv: Continuous kernel convolutions with differentiable kernel sizes. In The Tenth International Conference on Learning Representations, ICLR 2022, Virtual Event, April 25-29, 2022. OpenReview.net, 2022. 8 +[29] David W. Romero, Anna Kuzina, Erik J. Bekkers, Jakub Mikolaj Tomczak, and Mark Hoogendoorn. Ckconv: Continuous kernel convolution for sequential data. In The Tenth International Conference on Learning Representations, ICLR 2022, Virtual Event, April 25-29, 2022. OpenReview.net, 2022. 8 +[30] Chitwan Sahara, Jonathan Ho, William Chan, Tim Salimans, David J Fleet, and Mohammad Norouzi. Image superresolution via iterative refinement. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2022. 2 +[31] Wenzhe Shi, Jose Caballero, Ferenc Huszar, Johannes Totz, Andrew P. Aitken, Rob Bishop, Daniel Rueckert, and Zehan Wang. Real-time single image and video super-resolution using an efficient sub-pixel convolutional neural network. In 2016 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2016, Las Vegas, NV, USA, June 27-30, 2016, pages 1874-1883. IEEE Computer Society, 2016. 1, 2, 4, 5 +[32] Matthew Tancik, Pratul P. Srinivasan, Ben Mildenhall, Sara Fridovich-Keil, Nithin Raghavan, Utkarsh Singhal, Ravi Ramamoorthi, Jonathan T. Barron, and Ren Ng. Fourier features let networks learn high frequency functions in low dimensional domains. NeurIPS, 2020. 1, 4 +[33] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 8 +[34] Xintao Wang, Ke Yu, Shixiang Wu, Jinjin Gu, Yihao Liu, Chao Dong, Yu Qiao, and Chen Change Loy. Esrgan: Enhanced super-resolution generative adversarial networks. In Laura Leal-Taixe and Stefan Roth, editors, Computer Vision - ECCV 2018 Workshops, pages 63-79, Cham, 2019. Springer International Publishing. 2 +[35] Z. Wang, J. Chen, and S. H. Hoi. Deep learning for image super-resolution: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence, 43(10):3365-3387, oct 2021. 2 +[36] Wenxuan Wu, Zhongang Qi, and Fuxin Li. Pointconv: Deep convolutional networks on 3d point clouds. In IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2019, Long Beach, CA, USA, June 16-20, 2019, pages 9621-9630. Computer Vision Foundation / IEEE, 2019. 2 +[37] Yiheng Xie, Towaki Takikawa, Shunsuke Saito, Or Litany, Shiqin Yan, Numair Khan, Federico Tombari, James Tomp + +kin, Vincent Sitzmann, and Srinath Sridhar. Neural fields in visual computing and beyond, 2021. 1, 2 +[38] Roman Zeyde, Michael Elad, and Matan Protter. On single image scale-up using sparse-representations. In Jean-Daniel Boissonnat, Patrick Chenin, Albert Cohen, Christian Gout, Tom Lyche, Marie-Laurence Mazure, and Larry L. Schumaker, editors, Curves and Surfaces, volume 6920 of Lecture Notes in Computer Science, pages 711-730. Springer, 2010. 6 +[39] Yulun Zhang, Kunpeng Li, Kai Li, Lichen Wang, Bineng Zhong, and Yun Fu. Image super-resolution using very deep residual channel attention networks. In Vittorio Ferrari, Martial Hebert, Cristian Sminchisescu, and Yair Weiss, editors, Computer Vision - ECCV 2018, pages 294-310, Cham, 2018. Springer International Publishing. 1, 2 +[40] Yulun Zhang, Yapeng Tian, Yu Kong, Bineng Zhong, and Yun Fu. Residual dense network for image restoration. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 43(7):2480-2495, 2021. 1, 2, 3, 5, 6, 11, 12 \ No newline at end of file diff --git a/2023/CUF_ Continuous Upsampling Filters/images.zip b/2023/CUF_ Continuous Upsampling Filters/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..337ab7c9a69fd02ac2f44571028ecb4b680b5aa8 --- /dev/null +++ b/2023/CUF_ Continuous Upsampling Filters/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:016e1dbf9c5d5f5b8f1f2604da9f0f3699d9f46b66d20b877e347be14efeff23 +size 713178 diff --git a/2023/CUF_ Continuous Upsampling Filters/layout.json b/2023/CUF_ Continuous Upsampling Filters/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..ba1fc0e3823e4cb7ac36c3f2882053b5c46b5d30 --- /dev/null +++ b/2023/CUF_ Continuous Upsampling Filters/layout.json @@ -0,0 +1,10738 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 175, + 136, + 401, + 153 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 175, + 136, + 401, + 153 + ], + "spans": [ + { + "bbox": [ + 175, + 136, + 401, + 153 + ], + "type": "text", + "content": "CUF: Continuous Upsampling Filters" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 143, + 175, + 433, + 188 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 175, + 433, + 188 + ], + "spans": [ + { + "bbox": [ + 143, + 175, + 433, + 188 + ], + "type": "text", + "content": "Cristina N. Vasconcelos Cengiz Oztireli Mark Matthews" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 154, + 189, + 422, + 202 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 154, + 189, + 422, + 202 + ], + "spans": [ + { + "bbox": [ + 154, + 189, + 422, + 202 + ], + "type": "text", + "content": "Milad Hashemi Kevin Swersky Andrea Tagliasacchi" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 245, + 214, + 327, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 245, + 214, + 327, + 228 + ], + "spans": [ + { + "bbox": [ + 245, + 214, + 327, + 228 + ], + "type": "text", + "content": "Google Research" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 70, + 268, + 278, + 379 + ], + "blocks": [ + { + "bbox": [ + 130, + 255, + 235, + 267 + ], + "lines": [ + { + "bbox": [ + 130, + 255, + 235, + 267 + ], + "spans": [ + { + "bbox": [ + 130, + 255, + 235, + 267 + ], + "type": "text", + "content": "Memory Consumption" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 70, + 268, + 278, + 379 + ], + "lines": [ + { + "bbox": [ + 70, + 268, + 278, + 379 + ], + "spans": [ + { + "bbox": [ + 70, + 268, + 278, + 379 + ], + "type": "image", + "image_path": "5e5adc538e942a4d00d574aba78eb0bad1b5e60eb13a5bd807a8bc06cde65a15.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 292, + 267, + 499, + 379 + ], + "blocks": [ + { + "bbox": [ + 341, + 255, + 463, + 267 + ], + "lines": [ + { + "bbox": [ + 341, + 255, + 463, + 267 + ], + "spans": [ + { + "bbox": [ + 341, + 255, + 463, + 267 + ], + "type": "text", + "content": "Super-resolution Accuracy" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 292, + 267, + 499, + 379 + ], + "lines": [ + { + "bbox": [ + 292, + 267, + 499, + 379 + ], + "spans": [ + { + "bbox": [ + 292, + 267, + 499, + 379 + ], + "type": "image", + "image_path": "498b52f40a7682f7ba7ddc53672eca71dc97ed84987bf2488a3858a8ee7c327c.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 45, + 383, + 531, + 437 + ], + "lines": [ + { + "bbox": [ + 45, + 383, + 531, + 437 + ], + "spans": [ + { + "bbox": [ + 45, + 383, + 531, + 437 + ], + "type": "text", + "content": "Figure 1. We report (left) memory and FLOPs for upsampling an " + }, + { + "bbox": [ + 45, + 383, + 531, + 437 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 45, + 383, + 531, + 437 + ], + "type": "text", + "content": " image by different scale factors " + }, + { + "bbox": [ + 45, + 383, + 531, + 437 + ], + "type": "inline_equation", + "content": "(2 \\times, 3 \\times, 4 \\times)" + }, + { + "bbox": [ + 45, + 383, + 531, + 437 + ], + "type": "text", + "content": " using integer scale upsamplers (Sub-Pixel Convolution [31] and CIF-instantiated) and arbitrary-scale upsamplers (Meta-SR [16], LIIF [6], LTE [19] and CIF), but the same encoder backbone (EDSR-baseline [21]); and (right) the relationship between each upsampler FLOPS and PSNR performance on DIV2k dataset. Our arbitrary scale model is significantly lighter than other methods in the same class (i.e. continuous super-res). Further, when instantiated for integer scale factors, our upsampler is even-more efficient than sub-pixel convolutions [31]." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 139, + 447, + 186, + 458 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 139, + 447, + 186, + 458 + ], + "spans": [ + { + "bbox": [ + 139, + 447, + 186, + 458 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 45, + 485, + 281, + 719 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 485, + 281, + 719 + ], + "spans": [ + { + "bbox": [ + 45, + 485, + 281, + 719 + ], + "type": "text", + "content": "Neural fields have rapidly been adopted for representing 3D signals, but their application to more classical 2D image-processing has been relatively limited. In this paper, we consider one of the most important operations in image processing: upsampling. In deep learning, learnable upsampling layers have extensively been used for single image super-resolution. We propose to parameterize upsampling kernels as neural fields. This parameterization leads to a compact architecture that obtains a 40-fold reduction in the number of parameters when compared with competing arbitrary-scale super-resolution architectures. When upsampling images of size " + }, + { + "bbox": [ + 45, + 485, + 281, + 719 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 45, + 485, + 281, + 719 + ], + "type": "text", + "content": " we show that our architecture is " + }, + { + "bbox": [ + 45, + 485, + 281, + 719 + ], + "type": "inline_equation", + "content": "2x - 10x" + }, + { + "bbox": [ + 45, + 485, + 281, + 719 + ], + "type": "text", + "content": " more efficient than competing arbitrary-scale super-resolution architectures, and more efficient than sub-pixel convolutions when instantiated to a single-scale model. In the general setting, these gains grow polynomially with the square of the target scale. We validate our method on standard benchmarks showing such efficiency gains can be achieved without sacrifices in super-resolution performance. https://cuf-paper.github.io" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 298, + 447, + 375, + 458 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 447, + 375, + 458 + ], + "spans": [ + { + "bbox": [ + 298, + 447, + 375, + 458 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 296, + 466, + 530, + 513 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 466, + 530, + 513 + ], + "spans": [ + { + "bbox": [ + 296, + 466, + 530, + 513 + ], + "type": "text", + "content": "Neural-fields represent signals with coordinate-based neural-networks. They have found application in a multitude of areas including 3D reconstruction [24], novel-view synthesis [32], convolutions [14], and many others [37]." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 296, + 515, + 531, + 678 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 515, + 531, + 678 + ], + "spans": [ + { + "bbox": [ + 296, + 515, + 531, + 678 + ], + "type": "text", + "content": "Recent research has investigated the use of neural fields in the context of single image super-resolution [6, 19]. These models are based on multi-layer perceptrons conditioned on latent representation produced by encoders. While such architectures allow for continuous-scale superresolution, they require the execution of a conditional neural field for every pixel at the target resolution, making them unsuitable in applications with limited computational resources. Further, such a large use of resources is not justified by a increase in performance compared to classical convolutional architectures such as sub-pixel convolutions [31]. In summary, neural fields have not yet found widespread adoption as classical solutions are ① trivial to implement and ② more efficient. As they generally perform compa" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 57, + 7, + 101, + 39 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 7, + 101, + 39 + ], + "spans": [ + { + "bbox": [ + 57, + 7, + 101, + 39 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 136, + 5, + 487, + 18 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 5, + 487, + 18 + ], + "spans": [ + { + "bbox": [ + 136, + 5, + 487, + 18 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 190, + 18, + 433, + 29 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 190, + 18, + 433, + 29 + ], + "spans": [ + { + "bbox": [ + 190, + 18, + 433, + 29 + ], + "type": "text", + "content": "Except for this watermark, it is identical to the accepted version;" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 171, + 29, + 453, + 41 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 171, + 29, + 453, + 41 + ], + "spans": [ + { + "bbox": [ + 171, + 29, + 453, + 41 + ], + "type": "text", + "content": "the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 297, + 691, + 530, + 701 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 691, + 530, + 701 + ], + "spans": [ + { + "bbox": [ + 297, + 691, + 530, + 701 + ], + "type": "text", + "content": "1 Thereon we assume single image when talking about super-resolution." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 297, + 701, + 530, + 729 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 701, + 530, + 729 + ], + "spans": [ + { + "bbox": [ + 297, + 701, + 530, + 729 + ], + "type": "text", + "content": "2These encoders were originally proposed for classical super-resolution applications, and include both convolutional [8, 21, 40] as well as attentional [7, 20, 39] architectures." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 279, + 747, + 298, + 756 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 279, + 747, + 298, + 756 + ], + "spans": [ + { + "bbox": [ + 279, + 747, + 298, + 756 + ], + "type": "text", + "content": "9999" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 595, + 842 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 45, + 106, + 279, + 188 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 106, + 279, + 188 + ], + "spans": [ + { + "bbox": [ + 45, + 106, + 279, + 188 + ], + "type": "text", + "content": "rably, their usage is not justified in light of these points; see Figure 1. In this paper, we focus on overcoming these limitations, while noting that (regressive) super-resolution performance is in the saturation regime (i.e. further improvements in image quality seem unlikely without relying on generative modeling [30], and small improvements in PSNR do not necessarily correlate with image quality)." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 45, + 188, + 279, + 315 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 188, + 279, + 315 + ], + "spans": [ + { + "bbox": [ + 45, + 188, + 279, + 315 + ], + "type": "text", + "content": "Our driving hypothesis is that super-resolution convolutional filters are highly correlated - both spatially, as well as across scales. Hence, representing such filters in the latent space of a conditional neural field can effectively capture and compress such correlations. Given that neural fields encode continuous functions within neural networks, we call our filters Continuous Upsampling Filters (CUFs). While neural fields have proven effective in parameterizing 3D convolutions [5,14,36], we demonstrate that there are very significant savings to be had in the parameterization of 2D convolutions." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 45, + 316, + 279, + 410 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 316, + 279, + 410 + ], + "spans": [ + { + "bbox": [ + 45, + 316, + 279, + 410 + ], + "type": "text", + "content": "In implementing continuous upsampling filters, we draw inspiration from sub-pixel convolutions [31], and realize them via depth-wise convolutions. This not only makes CUFs significantly more efficient than competing continuous super-resolution architectures, but when instantiated to single-image super-resolution they are, surprisingly, even more efficient than \"ad-hoc\" sub-pixel convolutions with same number of input and output channels; see Figure 1." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 45, + 413, + 279, + 448 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 413, + 279, + 448 + ], + "spans": [ + { + "bbox": [ + 45, + 413, + 279, + 448 + ], + "type": "text", + "content": "Contributions. We investigate the use of neural fields as a parameterization of convolutional upsampling layers in super-resolution architectures, and show how:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 448, + 279, + 635 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 46, + 448, + 279, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 448, + 279, + 506 + ], + "spans": [ + { + "bbox": [ + 46, + 448, + 279, + 506 + ], + "type": "text", + "content": "- The continuity of neural-fields leads to training of compact convolutional architecture for continuous super-resolution (similar super-resolution PSNR but with flop reduction that grows polynomially with the square of the target scale.)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 506, + 279, + 564 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 506, + 279, + 564 + ], + "spans": [ + { + "bbox": [ + 46, + 506, + 279, + 564 + ], + "type": "text", + "content": "- Instantiating our continuous kernels into their discrete counterpart leads to an efficient inference pipeline, even more efficient in number of operations per target pixel than ad-hoc sub-pixel convolutions with same number of input and output channels." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 564, + 279, + 599 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 564, + 279, + 599 + ], + "spans": [ + { + "bbox": [ + 46, + 564, + 279, + 599 + ], + "type": "text", + "content": "- Discrete cosine transforms can be used as an efficient replacement for Fourier bases in the implementation of positional encoding for neural fields." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 599, + 279, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 599, + 279, + 635 + ], + "spans": [ + { + "bbox": [ + 46, + 599, + 279, + 635 + ], + "type": "text", + "content": "- These gains do not hinder reconstruction performance (PSNR), by carefully validating our method on standard benchmarks, and thoroughly abating our design choices." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 45, + 642, + 127, + 655 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 642, + 127, + 655 + ], + "spans": [ + { + "bbox": [ + 45, + 642, + 127, + 655 + ], + "type": "text", + "content": "2. Related work" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 45, + 660, + 279, + 730 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 660, + 279, + 730 + ], + "spans": [ + { + "bbox": [ + 45, + 660, + 279, + 730 + ], + "type": "text", + "content": "A thorough coverage of single-image super-resolution can be found in the following surveys [2, 3, 35]; in what follows, we provide an overview of classical techniques for super-resolution based on deep learning. As our model is based on neural-fields, we also point the reader to a survey in this topic [37], and below we discuss the existing works" + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 298, + 93, + 532, + 231 + ], + "blocks": [ + { + "bbox": [ + 298, + 93, + 532, + 231 + ], + "lines": [ + { + "bbox": [ + 298, + 93, + 532, + 231 + ], + "spans": [ + { + "bbox": [ + 298, + 93, + 532, + 231 + ], + "type": "image", + "image_path": "9a45162d0acb5af7b4168ea6c9d54d846460fa724d116bbee1664940ae4fea2c.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 296, + 234, + 530, + 310 + ], + "lines": [ + { + "bbox": [ + 296, + 234, + 530, + 310 + ], + "spans": [ + { + "bbox": [ + 296, + 234, + 530, + 310 + ], + "type": "text", + "content": "Figure 2. Previous Architectures - A visual comparison of the super-resolution architectures of Meta-SR [16], local implicit image function LIIF [6], and local texture estimators (LTE) [19], shown in an isotropic scaling configuration. Components inside the " + }, + { + "bbox": [ + 296, + 234, + 530, + 310 + ], + "type": "inline_equation", + "content": "2 \\times 2" + }, + { + "bbox": [ + 296, + 234, + 530, + 310 + ], + "type": "text", + "content": " Ensemble box are computed over the four closest pixels to compute a weighted average. The number of encoder channels, " + }, + { + "bbox": [ + 296, + 234, + 530, + 310 + ], + "type": "inline_equation", + "content": "C_e" + }, + { + "bbox": [ + 296, + 234, + 530, + 310 + ], + "type": "text", + "content": ", is dependent on the encoder chosen." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 297, + 322, + 465, + 334 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 322, + 465, + 334 + ], + "spans": [ + { + "bbox": [ + 297, + 322, + 465, + 334 + ], + "type": "text", + "content": "for super-resolution based on neural fields." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 296, + 339, + 530, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 339, + 530, + 373 + ], + "spans": [ + { + "bbox": [ + 296, + 339, + 530, + 373 + ], + "type": "text", + "content": "Single scale. There are two main frameworks for super-res, which mainly differ in the placement of upsampling operators within the architecture." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 298, + 374, + 530, + 468 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 298, + 374, + 530, + 432 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 374, + 530, + 432 + ], + "spans": [ + { + "bbox": [ + 298, + 374, + 530, + 432 + ], + "type": "text", + "content": "- In the pre-upsampling framework [8], an initial upscaled image is obtained using a non-trainable upsampler (e.g. bicubic), which is then post-processed by a neural network. This enables arbitrary size/scaling, but can introduce side effects such as noise amplification and blurring." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 298, + 432, + 530, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 432, + 530, + 468 + ], + "spans": [ + { + "bbox": [ + 298, + 432, + 530, + 468 + ], + "type": "text", + "content": "- In the post-upsampling framework [31], an encoder that preserves the input spatial resolution is followed by a shallow upsampling component." + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 296, + 468, + 530, + 561 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 468, + 530, + 561 + ], + "spans": [ + { + "bbox": [ + 296, + 468, + 530, + 561 + ], + "type": "text", + "content": "Note the computational complexity of pre-upsampling is significant, as these model operate directly at the target resolution. Consequently, post-upsampling is the de-facto mainstream approach [9, 20, 21, 34, 39, 40]. For this reason, we investigate the impact of an implicit upsampler in postupsampling frameworks, and cover a diverse set of architectures ranging from large [22], to extremely lightweight models [10]." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 296, + 566, + 530, + 612 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 566, + 530, + 612 + ], + "spans": [ + { + "bbox": [ + 296, + 566, + 530, + 612 + ], + "type": "text", + "content": "Neural Fields. The application of neural fields to superresolution [16] has enabled the design of architectures that support multiple target scales, as well as non-integer super-resolution scales." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 298, + 613, + 530, + 730 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 613, + 530, + 730 + ], + "spans": [ + { + "bbox": [ + 298, + 613, + 530, + 730 + ], + "type": "text", + "content": "- In meta super-resolution MetaSR [16], a hyper-network is conditioned on the target scale factor and on the output pixel relative coordinates (ie. the difference between source and target grids covering the same normalized global space). Their model consists of: (1) transforming the input image using an encoder into deep features at the source resolution; (2) upsampling the deep features into the target space (using nearest-neighbors); (3) using the hyper-network to produce a filter set per target relative position and scale; (4) processing the enlarged target" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 277, + 783, + 301, + 793 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 277, + 783, + 301, + 793 + ], + "spans": [ + { + "bbox": [ + 277, + 783, + 301, + 793 + ], + "type": "text", + "content": "10000" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 595, + 842 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 47, + 103, + 533, + 153 + ], + "blocks": [ + { + "bbox": [ + 47, + 103, + 533, + 153 + ], + "lines": [ + { + "bbox": [ + 47, + 103, + 533, + 153 + ], + "spans": [ + { + "bbox": [ + 47, + 103, + 533, + 153 + ], + "type": "table", + "html": "
Hyper NetworkUpsampler
input#layers // #neuronsoutputparams (K)inputlayersneuronsparams (K)
Meta-SR [16][δs(x), 1/s]2 // 25664 × 9 × 3445Eθconvolution--
LIIF [6]----[Εθ, δs(x), 1/s]5 dense256347
LTE [19]----ha(Εθ){sin, cos}{δs(x)π(hb(Εθ) + hc(1/s))}4 dense256494
Ours:[Π(δs(x)),Π(s),Π(k)]4 // 32646Eθd-w conv. + 2 dense644
", + "image_path": "54995636e4a109efe7e975633994d5adfe0d5ddbe3affa9d26dbd514e24e9ab0.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 45, + 155, + 532, + 179 + ], + "lines": [ + { + "bbox": [ + 45, + 155, + 532, + 179 + ], + "spans": [ + { + "bbox": [ + 45, + 155, + 532, + 179 + ], + "type": "text", + "content": "Table 1. Super-resolution with neural-fields: An overview of the main characteristics of neural field based architectures. Square brackets represent concatenation. In LTE [19], " + }, + { + "bbox": [ + 45, + 155, + 532, + 179 + ], + "type": "inline_equation", + "content": "h_a, h_b" + }, + { + "bbox": [ + 45, + 155, + 532, + 179 + ], + "type": "text", + "content": " are " + }, + { + "bbox": [ + 45, + 155, + 532, + 179 + ], + "type": "inline_equation", + "content": "3 \\times 3" + }, + { + "bbox": [ + 45, + 155, + 532, + 179 + ], + "type": "text", + "content": " convolutional layers and " + }, + { + "bbox": [ + 45, + 155, + 532, + 179 + ], + "type": "inline_equation", + "content": "h_c" + }, + { + "bbox": [ + 45, + 155, + 532, + 179 + ], + "type": "text", + "content": " is a linear layer." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 46, + 186, + 281, + 357 + ], + "blocks": [ + { + "bbox": [ + 46, + 186, + 281, + 357 + ], + "lines": [ + { + "bbox": [ + 46, + 186, + 281, + 357 + ], + "spans": [ + { + "bbox": [ + 46, + 186, + 281, + 357 + ], + "type": "image", + "image_path": "be3372693e5c8fc6e98b078c9b12aca104531ad914a3891667513b6798354c9e.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 45, + 369, + 281, + 489 + ], + "lines": [ + { + "bbox": [ + 45, + 369, + 281, + 489 + ], + "spans": [ + { + "bbox": [ + 45, + 369, + 281, + 489 + ], + "type": "text", + "content": "Figure 3. Architecture - Continuous Upsampling Filters (CUF) are based on a hyper-network. For each pixel position in the highres space, the hyper-network is computed " + }, + { + "bbox": [ + 45, + 369, + 281, + 489 + ], + "type": "inline_equation", + "content": "K^2" + }, + { + "bbox": [ + 45, + 369, + 281, + 489 + ], + "type": "text", + "content": " times, once for each position " + }, + { + "bbox": [ + 45, + 369, + 281, + 489 + ], + "type": "inline_equation", + "content": "i,j" + }, + { + "bbox": [ + 45, + 369, + 281, + 489 + ], + "type": "text", + "content": " of the upsampling kernel (batched in practice). This can be reduced when upsampling with integer scales " + }, + { + "bbox": [ + 45, + 369, + 281, + 489 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 45, + 369, + 281, + 489 + ], + "type": "text", + "content": ", as is common in practice. The Hyper-network output is aggregated into a tensor that is dot multiplied against the unfolded nearest-neighbor interpolation, computing a depth-wise convolution with per-pixel filters. The number of encoder channels, " + }, + { + "bbox": [ + 45, + 369, + 281, + 489 + ], + "type": "inline_equation", + "content": "C_e" + }, + { + "bbox": [ + 45, + 369, + 281, + 489 + ], + "type": "text", + "content": ", is dependent on the encoder used. In our experiments, we set the number of hypernetwork channels, " + }, + { + "bbox": [ + 45, + 369, + 281, + 489 + ], + "type": "inline_equation", + "content": "C_h = 32" + }, + { + "bbox": [ + 45, + 369, + 281, + 489 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 509, + 279, + 532 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 509, + 279, + 532 + ], + "spans": [ + { + "bbox": [ + 54, + 509, + 279, + 532 + ], + "type": "text", + "content": "resolution feature map into the final output using the corresponding weights." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 532, + 280, + 730 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 46, + 532, + 279, + 634 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 532, + 279, + 634 + ], + "spans": [ + { + "bbox": [ + 46, + 532, + 279, + 634 + ], + "type": "text", + "content": "- In local implicit image function LIIF [6], the author eliminates the use of a hyper-network (MetaSR, step 3) by extending the sampled target resolution features (step 2) with extra channels representing the relative coordinates and target scale. Their combined features are further processed through a stack of five dense layers inspired by the multi-layer perceptron (MLP) architecture typically adopted as the end-to-end model in neural-fields formulations." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 637, + 280, + 730 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 637, + 280, + 730 + ], + "spans": [ + { + "bbox": [ + 46, + 637, + 280, + 730 + ], + "type": "text", + "content": "- In local texture estimator LTE [19] the authors enhance the features fed as input to the MLP layers. In their model, the output of the encoder is processed by three extra trainable layers associated with the amplitude, frequency and phase of sin/cosine waves. The resulting projection in the target resolution layer is used as the input to an MLP that stacks 4 fully-connected layers. In order to further improve the results they also adopt a global skip" + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 305, + 188, + 531, + 234 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 188, + 531, + 234 + ], + "spans": [ + { + "bbox": [ + 305, + 188, + 531, + 234 + ], + "type": "text", + "content": "connection with a bilinear up-scaled version of the input around the full model, such that the deep model focus on the computation of the residual between to the closed form approximation and the final result." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 296, + 235, + 531, + 305 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 235, + 531, + 305 + ], + "spans": [ + { + "bbox": [ + 296, + 235, + 531, + 305 + ], + "type": "text", + "content": "A visualization of these architectures can be found in Figure 2. In practice, LIIF and LTE also increase the number of feature channels as their MLP layers have more neurons (256) than the number of channels produced as the encoder output (64 for encoders such as EDSR [21] and RDN [40]). In contrast to these previous works, our upsampler:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 298, + 306, + 531, + 457 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 298, + 306, + 531, + 363 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 306, + 531, + 363 + ], + "spans": [ + { + "bbox": [ + 298, + 306, + 531, + 363 + ], + "type": "text", + "content": "- Moves the computational burden of arbitrary upsampling into the continuous upsampler operator while reducing the operations performed in the target resolution space, preserving the number of channels produced by the deep encoder (Table 1)." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 298, + 364, + 530, + 410 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 364, + 530, + 410 + ], + "spans": [ + { + "bbox": [ + 298, + 364, + 530, + 410 + ], + "type": "text", + "content": "- Requires fewer operations than the single-scale sub-pixel convolution in the most typical application case of upsampling an image by an integer scale factor, and producing results of comparable quality." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 298, + 410, + 530, + 457 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 410, + 530, + 457 + ], + "spans": [ + { + "bbox": [ + 298, + 410, + 530, + 457 + ], + "type": "text", + "content": "- Adopt a neural-fields formulation that focuses the sensitivity of our hyper-network to fine-grained changes in the representation of scale and relative position (reducing spectral bias)." + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 298, + 467, + 351, + 479 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 467, + 351, + 479 + ], + "spans": [ + { + "bbox": [ + 298, + 467, + 351, + 479 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 296, + 485, + 531, + 637 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 485, + 531, + 637 + ], + "spans": [ + { + "bbox": [ + 296, + 485, + 531, + 637 + ], + "type": "text", + "content": "Given a target scale " + }, + { + "bbox": [ + 296, + 485, + 531, + 637 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 296, + 485, + 531, + 637 + ], + "type": "text", + "content": ", we aim to produce an upscaled image of size " + }, + { + "bbox": [ + 296, + 485, + 531, + 637 + ], + "type": "inline_equation", + "content": "sH \\times sW \\times 3" + }, + { + "bbox": [ + 296, + 485, + 531, + 637 + ], + "type": "text", + "content": " given an input image of size " + }, + { + "bbox": [ + 296, + 485, + 531, + 637 + ], + "type": "inline_equation", + "content": "H \\times W \\times 3" + }, + { + "bbox": [ + 296, + 485, + 531, + 637 + ], + "type": "text", + "content": ". Within the context of super-resolution, our core contribution is the introduction of a novel learnable upsampling layer. The upsampling layer can be interpreted as a decoder in a classical encoder-decoder architecture. We briefly review our encoder architecture in Section 3.1, and detail our decoder in Section 3.2. Note that our analysis focuses on the decoder, as a variety of encoders can be used. In Section 3.4, we then perform a conceptual comparison between our architecture, shown in Figure 3, and others based on neural fields, which are visually summarized in Figure 2." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 296, + 641, + 531, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 641, + 531, + 689 + ], + "spans": [ + { + "bbox": [ + 296, + 641, + 531, + 689 + ], + "type": "text", + "content": "Training. Our network with parameters " + }, + { + "bbox": [ + 296, + 641, + 531, + 689 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 296, + 641, + 531, + 689 + ], + "type": "text", + "content": " is trained to regress a high resolution image " + }, + { + "bbox": [ + 296, + 641, + 531, + 689 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{I}}^s" + }, + { + "bbox": [ + 296, + 641, + 531, + 689 + ], + "type": "text", + "content": " matching the ground truth " + }, + { + "bbox": [ + 296, + 641, + 531, + 689 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{I}}_{\\mathrm{gt}}^s" + }, + { + "bbox": [ + 296, + 641, + 531, + 689 + ], + "type": "text", + "content": " at random positions " + }, + { + "bbox": [ + 296, + 641, + 531, + 689 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 296, + 641, + 531, + 689 + ], + "type": "text", + "content": " given the low resolution input image " + }, + { + "bbox": [ + 296, + 641, + 531, + 689 + ], + "type": "inline_equation", + "content": "\\mathbf{I}" + }, + { + "bbox": [ + 296, + 641, + 531, + 689 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 330, + 698, + 530, + 718 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 330, + 698, + 530, + 718 + ], + "spans": [ + { + "bbox": [ + 330, + 698, + 530, + 718 + ], + "type": "interline_equation", + "content": "\\underset {\\boldsymbol {\\theta}} {\\arg \\min} \\mathbb {E} _ {\\mathbf {I}} \\mathbb {E} _ {s} \\mathbb {E} _ {\\mathbf {x}} \\| \\tilde {\\mathbf {I}} ^ {s} (\\mathbf {x}; \\mathbf {I}, \\boldsymbol {\\theta}) - \\tilde {\\mathbf {I}} _ {\\mathrm {g l}} ^ {s} (\\mathbf {x}) \\| _ {2} ^ {1} \\qquad (1)", + "image_path": "111ac1af77bdce220c91befae7b6ca50522782a7206bf69c49806ed5bea3a75b.jpg" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 277, + 783, + 300, + 794 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 277, + 783, + 300, + 794 + ], + "spans": [ + { + "bbox": [ + 277, + 783, + 300, + 794 + ], + "type": "text", + "content": "10001" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 595, + 842 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 46, + 109, + 276, + 151 + ], + "blocks": [ + { + "bbox": [ + 46, + 109, + 276, + 151 + ], + "lines": [ + { + "bbox": [ + 46, + 109, + 276, + 151 + ], + "spans": [ + { + "bbox": [ + 46, + 109, + 276, + 151 + ], + "type": "image", + "image_path": "a96e23ecf22c6927002922b940f84b654a448f063839424303ac92c15ed3ac5b.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 45, + 153, + 279, + 228 + ], + "lines": [ + { + "bbox": [ + 45, + 153, + 279, + 228 + ], + "spans": [ + { + "bbox": [ + 45, + 153, + 279, + 228 + ], + "type": "text", + "content": "Figure 4. Continuous super-resolution - Our architecture directly allows for an arbitrary and continuous choice of the superresolution factor. Conversely, traditional methods perform this indirectly: they need to upsample to an integer factor, and then consequently downsample to the target resolution. More examples are available in the appendix. Note that " + }, + { + "bbox": [ + 45, + 153, + 279, + 228 + ], + "type": "inline_equation", + "content": "\\times 4.5" + }, + { + "bbox": [ + 45, + 153, + 279, + 228 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 45, + 153, + 279, + 228 + ], + "type": "inline_equation", + "content": "\\times 5" + }, + { + "bbox": [ + 45, + 153, + 279, + 228 + ], + "type": "text", + "content": " are upsampling factors not available in the training data." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 48, + 229, + 124, + 305 + ], + "blocks": [ + { + "bbox": [ + 48, + 229, + 124, + 305 + ], + "lines": [ + { + "bbox": [ + 48, + 229, + 124, + 305 + ], + "spans": [ + { + "bbox": [ + 48, + 229, + 124, + 305 + ], + "type": "image", + "image_path": "5cad7d8416adf878db3d614d692fa1f109362e6cb1b270a1f953906f800fdfc5.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 74, + 308, + 99, + 315 + ], + "lines": [ + { + "bbox": [ + 74, + 308, + 99, + 315 + ], + "spans": [ + { + "bbox": [ + 74, + 308, + 99, + 315 + ], + "type": "text", + "content": "all filters" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 127, + 229, + 205, + 305 + ], + "blocks": [ + { + "bbox": [ + 127, + 229, + 205, + 305 + ], + "lines": [ + { + "bbox": [ + 127, + 229, + 205, + 305 + ], + "spans": [ + { + "bbox": [ + 127, + 229, + 205, + 305 + ], + "type": "image", + "image_path": "2e6fdc1a9cf64f469a5a1659ef1e0e9d695bfc25c3f180f6ba5c9018d01b850f.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 136, + 307, + 195, + 315 + ], + "lines": [ + { + "bbox": [ + 136, + 307, + 195, + 315 + ], + "spans": [ + { + "bbox": [ + 136, + 307, + 195, + 315 + ], + "type": "text", + "content": "filter #2 - continuous" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 206, + 229, + 284, + 305 + ], + "blocks": [ + { + "bbox": [ + 206, + 229, + 284, + 305 + ], + "lines": [ + { + "bbox": [ + 206, + 229, + 284, + 305 + ], + "spans": [ + { + "bbox": [ + 206, + 229, + 284, + 305 + ], + "type": "image", + "image_path": "272b8837f562fca0ad5791ee96b8cab0237d77ffe09112f927ed64c4f4dd9d5c.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 215, + 307, + 274, + 315 + ], + "lines": [ + { + "bbox": [ + 215, + 307, + 274, + 315 + ], + "spans": [ + { + "bbox": [ + 215, + 307, + 274, + 315 + ], + "type": "text", + "content": "filter #2 - discretized" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 45, + 320, + 279, + 384 + ], + "lines": [ + { + "bbox": [ + 45, + 320, + 279, + 384 + ], + "spans": [ + { + "bbox": [ + 45, + 320, + 279, + 384 + ], + "type": "text", + "content": "Figure 5. Continuous filters - CUF filters are continuous functions stored in neural fields. We visualize our filters by densely sampling each on the " + }, + { + "bbox": [ + 45, + 320, + 279, + 384 + ], + "type": "inline_equation", + "content": "[0,1]^2" + }, + { + "bbox": [ + 45, + 320, + 279, + 384 + ], + "type": "text", + "content": " domain to highlight their continuous nature (left). Close-up of an exemplar filter (middle). Filters are discretized by sampling at pixel centers (right), allowing them to be used in classical fixed-scale super-res architectures." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 45, + 392, + 107, + 403 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 392, + 107, + 403 + ], + "spans": [ + { + "bbox": [ + 45, + 392, + 107, + 403 + ], + "type": "text", + "content": "3.1.Encoder" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 45, + 410, + 279, + 433 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 410, + 279, + 433 + ], + "spans": [ + { + "bbox": [ + 45, + 410, + 279, + 433 + ], + "type": "text", + "content": "Towards this objective, we first process our input image with an encoder producing " + }, + { + "bbox": [ + 45, + 410, + 279, + 433 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 45, + 410, + 279, + 433 + ], + "type": "text", + "content": "-dimensional features:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 103, + 436, + 279, + 451 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 103, + 436, + 279, + 451 + ], + "spans": [ + { + "bbox": [ + 103, + 436, + 279, + 451 + ], + "type": "interline_equation", + "content": "\\mathcal {E} _ {\\boldsymbol {\\theta}}: \\mathbb {R} ^ {H \\times W \\times 3} \\rightarrow \\mathbb {R} ^ {H \\times W \\times C} \\tag {2}", + "image_path": "66babf9b8e352b6fad78cccad59700a008a3df601a848668ff3a468f74c416e7.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 45, + 456, + 279, + 503 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 456, + 279, + 503 + ], + "spans": [ + { + "bbox": [ + 45, + 456, + 279, + 503 + ], + "type": "text", + "content": "and then unfold a " + }, + { + "bbox": [ + 45, + 456, + 279, + 503 + ], + "type": "inline_equation", + "content": "k \\times k" + }, + { + "bbox": [ + 45, + 456, + 279, + 503 + ], + "type": "text", + "content": " spatial neighborhood of " + }, + { + "bbox": [ + 45, + 456, + 279, + 503 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 45, + 456, + 279, + 503 + ], + "type": "text", + "content": "-dimensional features into a tensor of " + }, + { + "bbox": [ + 45, + 456, + 279, + 503 + ], + "type": "inline_equation", + "content": "C \\times k^2" + }, + { + "bbox": [ + 45, + 456, + 279, + 503 + ], + "type": "text", + "content": " channels. Finally, we upsample the encoded image using nearest-neighbour interpolation. We define this step as:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 93, + 507, + 279, + 522 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 93, + 507, + 279, + 522 + ], + "spans": [ + { + "bbox": [ + 93, + 507, + 279, + 522 + ], + "type": "interline_equation", + "content": "\\mathcal {U}: \\mathbb {R} ^ {H \\times W \\times C} \\rightarrow \\mathbb {R} ^ {s H \\times s W \\times C \\times k ^ {2}} \\tag {3}", + "image_path": "2297e38a4607566e02186f708a462ed4398d314e9039438eda0c7b40254415be.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 45, + 527, + 279, + 585 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 527, + 279, + 585 + ], + "spans": [ + { + "bbox": [ + 45, + 527, + 279, + 585 + ], + "type": "text", + "content": "The unfolding part of this procedure leads to feature maps that will allow us to implement depth-wise, spatial convolutions as dot products, as is typically done for low-level neural network implementations. We define this whole feature extraction procedure as:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 591, + 279, + 603 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 591, + 279, + 603 + ], + "spans": [ + { + "bbox": [ + 132, + 591, + 279, + 603 + ], + "type": "interline_equation", + "content": "\\mathbf {F} = \\mathcal {U} \\left(\\mathcal {E} _ {\\theta} (\\mathbf {I})\\right) \\tag {4}", + "image_path": "267db0610ecead261186e34245ea25a9b552beac9ec55ada045dd8b7e26596e3.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 45, + 612, + 270, + 625 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 612, + 270, + 625 + ], + "spans": [ + { + "bbox": [ + 45, + 612, + 270, + 625 + ], + "type": "text", + "content": "3.2. Decoder – “Continuous Upsampling Filters”" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 45, + 631, + 279, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 631, + 279, + 712 + ], + "spans": [ + { + "bbox": [ + 45, + 631, + 279, + 712 + ], + "type": "text", + "content": "We draw our inspiration from classical sub-pixel convolutions [31], and combine it with recent research that apply neural fields for continuous super-resolution [6, 19]. Differently from these architectures, we achieve this objective by instantiating upsampling kernels " + }, + { + "bbox": [ + 45, + 631, + 279, + 712 + ], + "type": "inline_equation", + "content": "\\kappa" + }, + { + "bbox": [ + 45, + 631, + 279, + 712 + ], + "type": "text", + "content": " via hypernetworks [13]. Our super-resolution network is algebraically expressed as:" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 75, + 717, + 279, + 730 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 717, + 279, + 730 + ], + "spans": [ + { + "bbox": [ + 75, + 717, + 279, + 730 + ], + "type": "interline_equation", + "content": "\\tilde {\\mathbf {I}} ^ {s} (\\mathbf {x}; \\mathbf {I}, \\boldsymbol {\\theta}) = \\mathcal {D} _ {\\boldsymbol {\\theta}} \\left(\\mathcal {K} _ {\\boldsymbol {\\theta}} \\left(\\delta^ {s} (\\mathbf {x}); s\\right) \\cdot \\mathbf {F} (\\lfloor \\mathbf {x} / s \\rfloor)\\right) \\tag {5}", + "image_path": "7df2be6e823a80905ce5bbca39f4f89da69f5022ec45b31d36dc1120aeceff07.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 296, + 106, + 530, + 165 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 106, + 530, + 165 + ], + "spans": [ + { + "bbox": [ + 296, + 106, + 530, + 165 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 296, + 106, + 530, + 165 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 296, + 106, + 530, + 165 + ], + "type": "text", + "content": " indexes pixel locations in the target resolution, and " + }, + { + "bbox": [ + 296, + 106, + 530, + 165 + ], + "type": "inline_equation", + "content": "\\delta^s (\\mathbf{x}) = \\mathrm{mod}(\\mathbf{x},s) / s" + }, + { + "bbox": [ + 296, + 106, + 530, + 165 + ], + "type": "text", + "content": " renders our convolutional filters translation invariant, and the dot product implements a spatial convolution. The network " + }, + { + "bbox": [ + 296, + 106, + 530, + 165 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 296, + 106, + 530, + 165 + ], + "type": "text", + "content": " is a point-wise layer that maps (super-resolved) features back to RGB values:" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 381, + 174, + 530, + 187 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 381, + 174, + 530, + 187 + ], + "spans": [ + { + "bbox": [ + 381, + 174, + 530, + 187 + ], + "type": "interline_equation", + "content": "\\mathcal {D} _ {\\theta}: \\mathbb {R} ^ {C} \\rightarrow \\mathbb {R} ^ {3} \\tag {6}", + "image_path": "29763518ff2187c89114079bf63afc882d8c076ed78724f0fa938190d829fab7.jpg" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 296, + 191, + 530, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 191, + 530, + 251 + ], + "spans": [ + { + "bbox": [ + 296, + 191, + 530, + 251 + ], + "type": "text", + "content": "This is applied to a grid of coordinates of size " + }, + { + "bbox": [ + 296, + 191, + 530, + 251 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^{sH\\times sW\\times C}" + }, + { + "bbox": [ + 296, + 191, + 530, + 251 + ], + "type": "text", + "content": " to get an image of size " + }, + { + "bbox": [ + 296, + 191, + 530, + 251 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^{sH\\times sW\\times 3}" + }, + { + "bbox": [ + 296, + 191, + 530, + 251 + ], + "type": "text", + "content": ". The coordinate " + }, + { + "bbox": [ + 296, + 191, + 530, + 251 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 296, + 191, + 530, + 251 + ], + "type": "text", + "content": " is continuous, hence " + }, + { + "bbox": [ + 296, + 191, + 530, + 251 + ], + "type": "inline_equation", + "content": "\\kappa_{\\theta}" + }, + { + "bbox": [ + 296, + 191, + 530, + 251 + ], + "type": "text", + "content": " is a neural field parameterization of a convolutional kernel mapping (continuous) spatial offsets and (continuous) scales to convolutional weights:" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 354, + 256, + 530, + 271 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 354, + 256, + 530, + 271 + ], + "spans": [ + { + "bbox": [ + 354, + 256, + 530, + 271 + ], + "type": "interline_equation", + "content": "\\mathcal {K} _ {\\boldsymbol {\\theta}}: [ 0, 1 ] ^ {2} \\times \\mathbb {R} ^ {+} \\rightarrow \\mathbb {R} ^ {C \\times K ^ {2}} \\tag {7}", + "image_path": "d9cf01cc2111bc406ac139b22cd08efd97221b774ec851faf2687456ab7d6f63.jpg" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 296, + 283, + 530, + 341 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 283, + 530, + 341 + ], + "spans": [ + { + "bbox": [ + 296, + 283, + 530, + 341 + ], + "type": "text", + "content": "Continuous kernel indexing. We can take our continuous formulation a step further by introducing a continuous parametrization of its kernel indexes " + }, + { + "bbox": [ + 296, + 283, + 530, + 341 + ], + "type": "inline_equation", + "content": "k_{i}, k_{j}" + }, + { + "bbox": [ + 296, + 283, + 530, + 341 + ], + "type": "text", + "content": " indexing our " + }, + { + "bbox": [ + 296, + 283, + 530, + 341 + ], + "type": "inline_equation", + "content": "K \\times K" + }, + { + "bbox": [ + 296, + 283, + 530, + 341 + ], + "type": "text", + "content": " convolution weight entries. Thus, the hyper-network representing the convolution field becomes:" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 324, + 347, + 530, + 362 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 324, + 347, + 530, + 362 + ], + "spans": [ + { + "bbox": [ + 324, + 347, + 530, + 362 + ], + "type": "interline_equation", + "content": "\\hat {\\mathcal {K}} _ {\\boldsymbol {\\theta}}: [ 0, 1 ] ^ {2} \\times \\{0, \\dots , K - 1 \\} ^ {2} \\times \\mathbb {R} ^ {+} \\rightarrow \\mathbb {R} ^ {C} \\tag {8}", + "image_path": "6ded56b2c6b75d724dbecd89e07dc49f9cb65b8bf218864dbc32b9d45e8c0bc5.jpg" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 296, + 367, + 530, + 414 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 367, + 530, + 414 + ], + "spans": [ + { + "bbox": [ + 296, + 367, + 530, + 414 + ], + "type": "text", + "content": "and " + }, + { + "bbox": [ + 296, + 367, + 530, + 414 + ], + "type": "inline_equation", + "content": "\\kappa_{\\theta}" + }, + { + "bbox": [ + 296, + 367, + 530, + 414 + ], + "type": "text", + "content": " can then be constructed by 9 invocations to " + }, + { + "bbox": [ + 296, + 367, + 530, + 414 + ], + "type": "inline_equation", + "content": "\\hat{\\kappa}_{\\theta}" + }, + { + "bbox": [ + 296, + 367, + 530, + 414 + ], + "type": "text", + "content": " in the case of a " + }, + { + "bbox": [ + 296, + 367, + 530, + 414 + ], + "type": "inline_equation", + "content": "K = 3" + }, + { + "bbox": [ + 296, + 367, + 530, + 414 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 296, + 367, + 530, + 414 + ], + "type": "inline_equation", + "content": "K^2" + }, + { + "bbox": [ + 296, + 367, + 530, + 414 + ], + "type": "text", + "content": " invocations in the general setting. We use " + }, + { + "bbox": [ + 296, + 367, + 530, + 414 + ], + "type": "inline_equation", + "content": "K = 3" + }, + { + "bbox": [ + 296, + 367, + 530, + 414 + ], + "type": "text", + "content": " throughout the paper in order to facilitate a direct comparison to pre-existing baseline models." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 296, + 415, + 530, + 508 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 415, + 530, + 508 + ], + "spans": [ + { + "bbox": [ + 296, + 415, + 530, + 508 + ], + "type": "text", + "content": "Equation (7) and Equation (8) are both valid ways of generating spatial kernels. The first can be seen as a multi-headed hypernetwork, while the second uses input conditioning to generate the kernel values. We use the latter in our experiments, as the layers of nonlinearities provide additional expressiveness compared to the linear transformation used in the multi-headed version. A comparison between the two is presented in Appendix D." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 297, + 514, + 408, + 527 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 514, + 408, + 527 + ], + "spans": [ + { + "bbox": [ + 297, + 514, + 408, + 527 + ], + "type": "text", + "content": "3.3. Positional encoding" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 296, + 532, + 530, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 532, + 530, + 578 + ], + "spans": [ + { + "bbox": [ + 296, + 532, + 530, + 578 + ], + "type": "text", + "content": "Naively representing signals with MLPs leads to \"spectral bias\" [26]: the overall difficulty these networks have in representing high-frequency signals. Hence we implement " + }, + { + "bbox": [ + 296, + 532, + 530, + 578 + ], + "type": "inline_equation", + "content": "\\kappa_{\\theta}" + }, + { + "bbox": [ + 296, + 532, + 530, + 578 + ], + "type": "text", + "content": " as:" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 313, + 586, + 530, + 599 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 586, + 530, + 599 + ], + "spans": [ + { + "bbox": [ + 313, + 586, + 530, + 599 + ], + "type": "interline_equation", + "content": "\\mathcal {K} _ {\\boldsymbol {\\theta}} \\left(\\delta^ {s} (\\mathbf {x}), s, k\\right) = \\operatorname {M L P} \\left(\\Pi \\left(\\delta^ {s} (\\mathbf {x})\\right), \\Pi (s), \\Pi (k)\\right) \\tag {9}", + "image_path": "c52284af8a99d0453d04b85fddbc4e66b09c85681b571a79f5a5203766c8593b.jpg" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 296, + 605, + 530, + 698 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 605, + 530, + 698 + ], + "spans": [ + { + "bbox": [ + 296, + 605, + 530, + 698 + ], + "type": "text", + "content": "where we apply positional encoding " + }, + { + "bbox": [ + 296, + 605, + 530, + 698 + ], + "type": "inline_equation", + "content": "\\Pi (\\cdot)" + }, + { + "bbox": [ + 296, + 605, + 530, + 698 + ], + "type": "text", + "content": " to the MLP inputs. Several variants of positional encoding exist, including Fourier [24] and random Fourier [32] variants. However, as we strive for efficiency, we take inspiration from classical signal processing (e.g. JPEG) and instead employ a cosine-only transformation. Given a scalar " + }, + { + "bbox": [ + 296, + 605, + 530, + 698 + ], + "type": "inline_equation", + "content": "f_{n}" + }, + { + "bbox": [ + 296, + 605, + 530, + 698 + ], + "type": "text", + "content": " sampled uniformly in the range " + }, + { + "bbox": [ + 296, + 605, + 530, + 698 + ], + "type": "inline_equation", + "content": "[0,f_{\\mathrm{max}}]" + }, + { + "bbox": [ + 296, + 605, + 530, + 698 + ], + "type": "text", + "content": ", a scalar quantity " + }, + { + "bbox": [ + 296, + 605, + 530, + 698 + ], + "type": "inline_equation", + "content": "z\\in [0,1]" + }, + { + "bbox": [ + 296, + 605, + 530, + 698 + ], + "type": "text", + "content": " is encoded as the (sorted) vector:" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 339, + 703, + 530, + 733 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 339, + 703, + 530, + 733 + ], + "spans": [ + { + "bbox": [ + 339, + 703, + 530, + 733 + ], + "type": "interline_equation", + "content": "\\Pi (z) = \\left\\{\\cos \\left(\\frac {(2 z + 1) f _ {n} \\pi}{2}\\right) \\right\\} _ {n = 1} ^ {N} \\tag {10}", + "image_path": "ae43d9b6ff762288c51f8e4232371fb5ab780a8842f68647aee3440b4d0e6ce2.jpg" + } + ] + } + ], + "index": 31 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 277, + 783, + 301, + 793 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 277, + 783, + 301, + 793 + ], + "spans": [ + { + "bbox": [ + 277, + 783, + 301, + 793 + ], + "type": "text", + "content": "10002" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 595, + 842 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 48, + 103, + 531, + 226 + ], + "blocks": [ + { + "bbox": [ + 48, + 103, + 531, + 226 + ], + "lines": [ + { + "bbox": [ + 48, + 103, + 531, + 226 + ], + "spans": [ + { + "bbox": [ + 48, + 103, + 531, + 226 + ], + "type": "table", + "html": "
EncoderUpsamplerEns.Set5Set14BSD100Urban100
×2×3×4×6×2×3×4×6×2×3×4×6×2×3×4×6
RDNfixed-scale [40]38.2434.7132.47-34.0130.5728.81-32.3429.2627.72-32.8928.8026.61-
MetaSR [16]38.2234.6332.3829.0433.9830.5428.7826.5132.3329.2627.7125.9032.9228.8226.5523.99
CUF (ours)38.2334.7232.5429.2533.9930.5828.8626.7032.3529.2927.7625.9933.0128.9126.7524.23
fixed-scale [40]+geo38.3034.7832.61-34.1030.6728.92-32.4029.3327.75-33.0929.0026.82-
LIIF [6]+loc38.1734.6832.5029.1533.9730.5328.8026.6432.3229.2627.7425.9832.8728.8226.6824.20
LTE [19]+loc38.2334.7232.6129.3234.0930.5828.8826.7132.3629.3027.7726.0133.0428.9726.8124.28
CUF (ours)+geo38.2834.8032.6329.2734.0830.6528.9226.7432.3929.3327.8026.0333.1629.0526.8724.32
SwinIRfixed-scale [31]38.3534.8932.72-34.1430.7728.94-32.4429.3727.83-33.4029.2927.07-
MetaSR [16]38.2634.7732.4729.0934.1430.6628.8526.5832.3929.3127.7525.9433.2929.1226.7624.16
CUF (ours)38.3434.8832.8029.5334.2930.7929.0226.8532.4529.3827.8526.0933.5429.4527.2424.62
fixed-scale [31]+geo38.3834.9532.81-34.2430.8329.02-32.4729.4127.87-33.5129.4227.21-
LIIF [6]+loc38.2834.8732.7329.4634.1430.7528.9826.8232.3929.3427.8426.0733.3629.3327.1524.59
LTE+1c [19]+loc38.3334.8932.8129.5034.2530.8029.0626.8632.4429.3927.8626.0933.5029.4127.2424.62
CUF (ours)++geo38.3834.9232.8329.5734.3330.8429.0526.9132.4729.4027.8826.1133.6529.5527.3224.69
", + "image_path": "2fba32e02ee3228e31402839f325558c4cdbc3453b2a4233023dadf04dc10df5.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 48, + 273, + 278, + 464 + ], + "blocks": [ + { + "bbox": [ + 45, + 229, + 531, + 266 + ], + "lines": [ + { + "bbox": [ + 45, + 229, + 531, + 266 + ], + "spans": [ + { + "bbox": [ + 45, + 229, + 531, + 266 + ], + "type": "text", + "content": "Table 2. Out-of-domain evaluation - performance on datasets not seen during training. Metrics computed on the luminance channel in the YCbCr color space, following previous work. Models using self-ensemble are marked with " + }, + { + "bbox": [ + 45, + 229, + 531, + 266 + ], + "type": "inline_equation", + "content": "+_{geo}" + }, + { + "bbox": [ + 45, + 229, + 531, + 266 + ], + "type": "text", + "content": " and models using local self-ensemble are marked with " + }, + { + "bbox": [ + 45, + 229, + 531, + 266 + ], + "type": "inline_equation", + "content": "+_{loc}" + }, + { + "bbox": [ + 45, + 229, + 531, + 266 + ], + "type": "text", + "content": ". Legend: best " + }, + { + "bbox": [ + 45, + 229, + 531, + 266 + ], + "type": "inline_equation", + "content": "2^{\\text{nd}}" + }, + { + "bbox": [ + 45, + 229, + 531, + 266 + ], + "type": "text", + "content": " best" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 48, + 273, + 278, + 464 + ], + "lines": [ + { + "bbox": [ + 48, + 273, + 278, + 464 + ], + "spans": [ + { + "bbox": [ + 48, + 273, + 278, + 464 + ], + "type": "table", + "html": "
Multi-scale up-sampling methods - DIV2k
EncoderUpsamplerEns.seen scalesunseen scales
×2×3×4×6×12
Bicubic31.0128.2226.6624.8222.27
EDSR-b.Sub-pixel conv.34.6930.9428.97--
MetaSR34.6430.9328.9226.6123.55
CUF (ours)34.7030.9929.0126.7623.73
Sub-pixel conv.+geo34.7831.0329.06--
LIIF+loc34.6730.9629.0026.7523.71
LTE+loc34.7231.0229.0426.8123.78
CUF (ours)+geo34.7931.0729.0926.8223.78
RDNSub-pixel conv.35.0131.2229.20--
MetaSR35.0031.2729.2526.8823.73
CUF (ours)35.0331.3129.3227.0323.94
Sub-pixel conv.+geo35.1031.3329.31--
LIIF+loc34.9931.2629.2726.9923.89
LTE+loc35.0431.3229.3327.0423.95
CUF (ours)+geo35.1131.3929.3927.0923.99
SwinIRSub-pixel conv.35.2831.4729.40--
MetaSR35.1531.4029.3326.9423.80
CUF (ours)35.2631.5229.5227.1924.07
Sub-pixel conv.+geo35.3331.5229.44--
LIIF+loc35.1731.4629.4627.1524.02
LTE+loc35.2431.5029.5127.2024.09
CUF (ours)+geo35.3131.5629.5627.2324.10
", + "image_path": "a01376d3c3a21d7430b94108a0c282c84338fdaa94634342621578786db937a5.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 45, + 468, + 279, + 526 + ], + "lines": [ + { + "bbox": [ + 45, + 468, + 279, + 526 + ], + "spans": [ + { + "bbox": [ + 45, + 468, + 279, + 526 + ], + "type": "text", + "content": "Table 3. In domain evaluation: tests on DIV2K's validation subset [25] on scales seen " + }, + { + "bbox": [ + 45, + 468, + 279, + 526 + ], + "type": "inline_equation", + "content": "(2\\times -4\\times)" + }, + { + "bbox": [ + 45, + 468, + 279, + 526 + ], + "type": "text", + "content": " and unseen during training " + }, + { + "bbox": [ + 45, + 468, + 279, + 526 + ], + "type": "inline_equation", + "content": "(6\\times -30\\times)" + }, + { + "bbox": [ + 45, + 468, + 279, + 526 + ], + "type": "text", + "content": ". Metrics taken on the RGB space. Results taken using geometric self-ensemble [21] are marked with " + }, + { + "bbox": [ + 45, + 468, + 279, + 526 + ], + "type": "inline_equation", + "content": "^+" + }, + { + "bbox": [ + 45, + 468, + 279, + 526 + ], + "type": "text", + "content": " geo' and local-ensemble [6] with " + }, + { + "bbox": [ + 45, + 468, + 279, + 526 + ], + "type": "inline_equation", + "content": "^+" + }, + { + "bbox": [ + 45, + 468, + 279, + 526 + ], + "type": "text", + "content": " loc'. Legend: best " + }, + { + "bbox": [ + 45, + 468, + 279, + 526 + ], + "type": "inline_equation", + "content": "2^{\\mathrm{nd}}" + }, + { + "bbox": [ + 45, + 468, + 279, + 526 + ], + "type": "text", + "content": " best" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 45, + 536, + 279, + 584 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 536, + 279, + 584 + ], + "spans": [ + { + "bbox": [ + 45, + 536, + 279, + 584 + ], + "type": "text", + "content": "By eliminating the imaginary component from the Fourier basis, we show in Section 4.3 how this reduces the number of trainable parameters of the first hyper-network layer by half, without affecting super-resolution quality." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 590, + 107, + 602 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 590, + 107, + 602 + ], + "spans": [ + { + "bbox": [ + 46, + 590, + 107, + 602 + ], + "type": "text", + "content": "3.4. Analysis" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 45, + 608, + 279, + 678 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 608, + 279, + 678 + ], + "spans": [ + { + "bbox": [ + 45, + 608, + 279, + 678 + ], + "type": "text", + "content": "We now perform a conceptual comparison of our network architecture to the commonly used Sub-Pixel Convolution, as well as several super-resolution approaches based on neural-fields. We further note how continuous upsampling filters can be instantiated to generate filters compatible with Sub-Pixel Convolution." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 45, + 682, + 280, + 730 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 682, + 280, + 730 + ], + "spans": [ + { + "bbox": [ + 45, + 682, + 280, + 730 + ], + "type": "text", + "content": "Comparison to Sub-Pixel Convolution [31]. Sub-Pixel Convolution is the most commonly used super-resolution operator, especially when efficiency is critical. This is achieved by initially employing an expansion convolution" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 356, + 272, + 474, + 336 + ], + "blocks": [ + { + "bbox": [ + 356, + 272, + 474, + 336 + ], + "lines": [ + { + "bbox": [ + 356, + 272, + 474, + 336 + ], + "spans": [ + { + "bbox": [ + 356, + 272, + 474, + 336 + ], + "type": "image", + "image_path": "e64749d874d3c55d0d1c9f97deffbb2d76bb2892733a83adcebcc8c6302af96a.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 296, + 339, + 530, + 360 + ], + "lines": [ + { + "bbox": [ + 296, + 339, + 530, + 360 + ], + "spans": [ + { + "bbox": [ + 296, + 339, + 530, + 360 + ], + "type": "text", + "content": "Figure 6. A simple illustration of periodic shuffling used in SubPixel Convolution." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 296, + 368, + 531, + 428 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 368, + 531, + 428 + ], + "spans": [ + { + "bbox": [ + 296, + 368, + 531, + 428 + ], + "type": "text", + "content": "that generates a feature map with " + }, + { + "bbox": [ + 296, + 368, + 531, + 428 + ], + "type": "inline_equation", + "content": "s^2 N_{out}" + }, + { + "bbox": [ + 296, + 368, + 531, + 428 + ], + "type": "text", + "content": " channels, where " + }, + { + "bbox": [ + 296, + 368, + 531, + 428 + ], + "type": "inline_equation", + "content": "N_{out}" + }, + { + "bbox": [ + 296, + 368, + 531, + 428 + ], + "type": "text", + "content": " is a hyper parameter defining the target number of channels of the Sub-Pixel Convolution operation. Next, the periodic shuffling operator re-arranges the channels to produce a higher-resolution feature map (see Figure 6):" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 346, + 431, + 530, + 447 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 346, + 431, + 530, + 447 + ], + "spans": [ + { + "bbox": [ + 346, + 431, + 530, + 447 + ], + "type": "interline_equation", + "content": "\\mathcal {P}: \\mathbb {R} ^ {H \\times W \\times C} \\rightarrow \\mathbb {R} ^ {s H \\times s W \\times C / s ^ {2}} \\tag {11}", + "image_path": "ba44772cbbbb982d8c2a2a3e5d4ab11fc01dac71c4450c7c11045b3db8e10f40.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 296, + 452, + 531, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 452, + 531, + 592 + ], + "spans": [ + { + "bbox": [ + 296, + 452, + 531, + 592 + ], + "type": "text", + "content": "Similar to (6), whenever " + }, + { + "bbox": [ + 296, + 452, + 531, + 592 + ], + "type": "inline_equation", + "content": "N_{out}" + }, + { + "bbox": [ + 296, + 452, + 531, + 592 + ], + "type": "text", + "content": " is chosen as different from the number of output colors, the resulting feature map is further projected by point-wise convolutions. Often, " + }, + { + "bbox": [ + 296, + 452, + 531, + 592 + ], + "type": "inline_equation", + "content": "N_{out}" + }, + { + "bbox": [ + 296, + 452, + 531, + 592 + ], + "type": "text", + "content": " is taken as the same as the number of channels as produced by the encoder " + }, + { + "bbox": [ + 296, + 452, + 531, + 592 + ], + "type": "inline_equation", + "content": "\\mathcal{E}" + }, + { + "bbox": [ + 296, + 452, + 531, + 592 + ], + "type": "text", + "content": " for architectures adopting Sub-Pixel Convolution and targeting high quality results [20,21,40]. We note that the Sub-Pixel Convolution design based on the expansion convolution does not enforce spatial correlations between neighboring pixels (here we refer to neighborhood in the output, high resolution, domain) and need to be learned from training data. Conversely, these correlations are inherently captured by our continuous convolutional filters." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 296, + 592, + 530, + 675 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 592, + 530, + 675 + ], + "spans": [ + { + "bbox": [ + 296, + 592, + 530, + 675 + ], + "type": "text", + "content": "Further, in the integer scale setup where hyper-networks can be pre-instantiated, CUFs are computationally more efficient than sub-pixel convolutions whenever " + }, + { + "bbox": [ + 296, + 592, + 530, + 675 + ], + "type": "inline_equation", + "content": "K^2 + N_{in} + N_{out} < N_{out} \\times K^2" + }, + { + "bbox": [ + 296, + 592, + 530, + 675 + ], + "type": "text", + "content": " (in the worst case scenario) as the cost associated with pre-computing the weights for a given integer up-sampling scale can be neglected when compared to the operations performed on the image grid (Appendix E)." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 296, + 678, + 531, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 678, + 531, + 714 + ], + "spans": [ + { + "bbox": [ + 296, + 678, + 531, + 714 + ], + "type": "text", + "content": "Comparisons to LIIF [6] and LTE [19]. In comparison to these works, we make the main network shallower (1 depth-wise continuous convolution and 2 dense layers) in order to" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 307, + 719, + 522, + 730 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 719, + 522, + 730 + ], + "spans": [ + { + "bbox": [ + 307, + 719, + 522, + 730 + ], + "type": "text", + "content": "also called pixel-shuffle (PyTorch) or depth-to-space (TensorFlow)." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 277, + 783, + 300, + 793 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 277, + 783, + 300, + 793 + ], + "spans": [ + { + "bbox": [ + 277, + 783, + 300, + 793 + ], + "type": "text", + "content": "10003" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 595, + 842 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 45, + 106, + 281, + 281 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 106, + 281, + 281 + ], + "spans": [ + { + "bbox": [ + 45, + 106, + 281, + 281 + ], + "type": "text", + "content": "reduce the number of layers that operate at the target spatial resolution and at the encoder's channel resolution. At the same time, our upsampler head is considerably lighter than previous arbitrary scale methods, not only by the use of a depth-wise convolution but also by keeping the same number of channels produced by the features encoder in the main network (" + }, + { + "bbox": [ + 45, + 106, + 281, + 281 + ], + "type": "inline_equation", + "content": "C_e = 64" + }, + { + "bbox": [ + 45, + 106, + 281, + 281 + ], + "type": "text", + "content": ", which results in dense layers " + }, + { + "bbox": [ + 45, + 106, + 281, + 281 + ], + "type": "inline_equation", + "content": "16 \\times" + }, + { + "bbox": [ + 45, + 106, + 281, + 281 + ], + "type": "text", + "content": " cheaper than its LIIF and LTE counterparts using 256 neurons). When performing non-integer upsampling, the costs with processing the hyper-network layers are proportional to the target image resolution, but still smaller than a single dense layer of LIFF and LTE heads, due to the adoption of a reduced number of channels (as " + }, + { + "bbox": [ + 45, + 106, + 281, + 281 + ], + "type": "inline_equation", + "content": "C_h = 32" + }, + { + "bbox": [ + 45, + 106, + 281, + 281 + ], + "type": "text", + "content": ", each of CUF's hypernetwork dense layers is " + }, + { + "bbox": [ + 45, + 106, + 281, + 281 + ], + "type": "inline_equation", + "content": "64 \\times" + }, + { + "bbox": [ + 45, + 106, + 281, + 281 + ], + "type": "text", + "content": " cheaper than LIIF and LTE layers using 256 neurons)." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 45, + 286, + 281, + 484 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 286, + 281, + 484 + ], + "spans": [ + { + "bbox": [ + 45, + 286, + 281, + 484 + ], + "type": "text", + "content": "Instantiating CUFs. At inference time, when targeting an integer upscaling factor " + }, + { + "bbox": [ + 45, + 286, + 281, + 484 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 45, + 286, + 281, + 484 + ], + "type": "text", + "content": ", the hyper-network representing " + }, + { + "bbox": [ + 45, + 286, + 281, + 484 + ], + "type": "inline_equation", + "content": "\\mathcal{K}" + }, + { + "bbox": [ + 45, + 286, + 281, + 484 + ], + "type": "text", + "content": " can be queried to retrieve the weights corresponding to " + }, + { + "bbox": [ + 45, + 286, + 281, + 484 + ], + "type": "inline_equation", + "content": "s^2" + }, + { + "bbox": [ + 45, + 286, + 281, + 484 + ], + "type": "text", + "content": " relative subpixel positions as an initialization step during pre-processing. The retrieved weights are re-used across all pixels taking advantage of the existent periodicity. Thus, in the CUF-instantiated architecture the continuous kernel is replaced at test time with a discrete depth-wise convolution, followed by a pixel shuffling operation, in contrast with the unfolding operator used in the regular, fully continuous CUF described in this paper. Otherwise, the architecture remains the same. The costs associated with the hyper network at initialization can be neglected as " + }, + { + "bbox": [ + 45, + 286, + 281, + 484 + ], + "type": "inline_equation", + "content": "s^2 << sH \\times sW" + }, + { + "bbox": [ + 45, + 286, + 281, + 484 + ], + "type": "text", + "content": ". In this setting, our model becomes as efficient as a sub-pixel-convolution architecture, while retaining the aforementioned continuous modeling properties at training time." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 45, + 495, + 98, + 506 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 495, + 98, + 506 + ], + "spans": [ + { + "bbox": [ + 45, + 495, + 98, + 506 + ], + "type": "text", + "content": "4. Results" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 45, + 512, + 279, + 583 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 512, + 279, + 583 + ], + "spans": [ + { + "bbox": [ + 45, + 512, + 279, + 583 + ], + "type": "text", + "content": "In this section we describe our experimental setup in terms of backbone encoders and training/validation dataset, perform careful comparisons to the state-of-the-art (Section 4.1), and additional evaluation towards the implementation of lightweight super-res architectures (Section 4.2). We conclude by performing a thorough ablation (Section 4.3)." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 587, + 89, + 597 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 587, + 89, + 597 + ], + "spans": [ + { + "bbox": [ + 46, + 587, + 89, + 597 + ], + "type": "text", + "content": "Encoders." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 45, + 600, + 279, + 634 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 600, + 279, + 634 + ], + "spans": [ + { + "bbox": [ + 45, + 600, + 279, + 634 + ], + "type": "text", + "content": "We apply CUFs to a variety of encoders, to both show its generality, as well as its performance gains in a variety of settings." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 636, + 279, + 717 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 46, + 636, + 279, + 659 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 636, + 279, + 659 + ], + "spans": [ + { + "bbox": [ + 46, + 636, + 279, + 659 + ], + "type": "text", + "content": "- A state-of-the-art encoder for super-res named SwinIR [20] based on Swin-transformers [22]." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 659, + 279, + 682 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 659, + 279, + 682 + ], + "spans": [ + { + "bbox": [ + 46, + 659, + 279, + 682 + ], + "type": "text", + "content": "- Well-known convolution-based encoders, namely EDSR-baseline [21] and Residual Dense Networks (RDN) [40]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 682, + 279, + 717 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 682, + 279, + 717 + ], + "spans": [ + { + "bbox": [ + 46, + 682, + 279, + 717 + ], + "type": "text", + "content": "- Architectures for lightweight and extremely lightweight inference, respectively SwinT-lightweight [20] and ABPN [10]." + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 45, + 718, + 279, + 730 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 718, + 279, + 730 + ], + "spans": [ + { + "bbox": [ + 45, + 718, + 279, + 730 + ], + "type": "text", + "content": "Table 4 contrasts the ablated encoders' composition and" + } + ] + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 300, + 103, + 529, + 181 + ], + "blocks": [ + { + "bbox": [ + 300, + 103, + 529, + 181 + ], + "lines": [ + { + "bbox": [ + 300, + 103, + 529, + 181 + ], + "spans": [ + { + "bbox": [ + 300, + 103, + 529, + 181 + ], + "type": "table", + "html": "
MethodCompositionParameters (in M)
encoderspc. 4×
RDNConvolutions22.000.30
SwinTConv. and Self-Attention11.600.30
EDSR-baselineConvolutions1.200.30
SwinT-lightweightConv. and Self-Attention0.900.03
ABPNConvolutions0.030.03
", + "image_path": "04d3137e4d0c90bb047e2c7728221a49f1314054a84443403ea7934198123f11.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 296, + 184, + 530, + 228 + ], + "lines": [ + { + "bbox": [ + 296, + 184, + 530, + 228 + ], + "spans": [ + { + "bbox": [ + 296, + 184, + 530, + 228 + ], + "type": "text", + "content": "Table 4. Encoders - Our experiments use CUF layers in combination with various encoders differing in composition and size. Parameter counts reflect the encoder size and the upsampling head using a Sub-Pixel Convolution (spc) targeting " + }, + { + "bbox": [ + 296, + 184, + 530, + 228 + ], + "type": "inline_equation", + "content": "4 \\times" + }, + { + "bbox": [ + 296, + 184, + 530, + 228 + ], + "type": "text", + "content": " upsampling." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 296, + 238, + 530, + 250 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 238, + 530, + 250 + ], + "spans": [ + { + "bbox": [ + 296, + 238, + 530, + 250 + ], + "type": "text", + "content": "size. Hyperparameter settings can be found in the appendix." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 296, + 254, + 531, + 405 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 254, + 531, + 405 + ], + "spans": [ + { + "bbox": [ + 296, + 254, + 531, + 405 + ], + "type": "text", + "content": "Datasets. All models are trained using the DIV2K dataset training subset (800k images), introduced at the NTIRE 2017 Super Resolution Challenge [25]. Similarly to previous works, each image is randomly cropped 20 times per epoch and augmented with random flips, and 90 degrees rotations. We report peak signal-to-noise ratio (PSNR) results on the DIV2K validation set (100 images)" + }, + { + "bbox": [ + 296, + 254, + 531, + 405 + ], + "type": "inline_equation", + "content": "^4" + }, + { + "bbox": [ + 296, + 254, + 531, + 405 + ], + "type": "text", + "content": ", in the RGB space. To understand the generalizability of our experimental results, we test our models on additional well known datasets (Set5 [4] - 5 images; Set14 [38] - 14 images; B100 [23] - 100 images; and Urban100 [17] - 100 images), in which, following previous works, PSNR is measured in the YCbCr luminance channel." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 296, + 410, + 530, + 549 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 410, + 530, + 549 + ], + "spans": [ + { + "bbox": [ + 296, + 410, + 530, + 549 + ], + "type": "text", + "content": "Ensembling. Whenever remarked by " + }, + { + "bbox": [ + 296, + 410, + 530, + 549 + ], + "type": "inline_equation", + "content": "^{+}_{\\mathrm{geo}}" + }, + { + "bbox": [ + 296, + 410, + 530, + 549 + ], + "type": "text", + "content": ", results incorporate the (commonly adopted) Geometric Self-ensemble [21], where the results of rotated versions of the input images are averaged at test time. LIIF and LTE results adopt Local Self-ensemble " + }, + { + "bbox": [ + 296, + 410, + 530, + 549 + ], + "type": "inline_equation", + "content": "(^{+}_{\\mathrm{loc}}" + }, + { + "bbox": [ + 296, + 410, + 530, + 549 + ], + "type": "text", + "content": " ), where the result of applying the upsampler at 4 shifted grid points is averaged, resulting in a " + }, + { + "bbox": [ + 296, + 410, + 530, + 549 + ], + "type": "inline_equation", + "content": "\\times 4" + }, + { + "bbox": [ + 296, + 410, + 530, + 549 + ], + "type": "text", + "content": " increase in computational complexity at training/test time. It may also introduce confounding factors to the optimization process that invalidate direct comparison to the fixed-scale upsamplers, as during backpropagation the encoder gradients are averaged. In Appendix B we combine these models." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 296, + 557, + 450, + 569 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 557, + 450, + 569 + ], + "spans": [ + { + "bbox": [ + 296, + 557, + 450, + 569 + ], + "type": "text", + "content": "4.1. State-of-the-art comparisons" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 296, + 575, + 530, + 703 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 575, + 530, + 703 + ], + "spans": [ + { + "bbox": [ + 296, + 575, + 530, + 703 + ], + "type": "text", + "content": "Table 2 and Table 3 show the quantitative comparisons, while qualitative results can be observed in Figure 4, Figure 7 and Appendix F as well as on our supplementary website https://cuf-paper.github.io/. Table 3 contains in-domain results: models trained on DIV2k and tested on DIV2k (train/test split), while Table 2 presents out-of-domain results: models are trained on DIV2k, but tested on different datasets. Note that upsampling ratios larger than " + }, + { + "bbox": [ + 296, + 575, + 530, + 703 + ], + "type": "inline_equation", + "content": "\\times 4" + }, + { + "bbox": [ + 296, + 575, + 530, + 703 + ], + "type": "text", + "content": " are not provided at training time. The models adopting sub-pixel convolutions are fixed-scale models, that is, one model is trained per target scale and thus do not" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 710, + 530, + 730 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 710, + 530, + 730 + ], + "spans": [ + { + "bbox": [ + 297, + 710, + 530, + 730 + ], + "type": "text", + "content": "4The models we compare to also do not report results on the test-set, which is no longer publicly available." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 277, + 783, + 301, + 793 + ], + "type": "footer", + "angle": 0, + "lines": [ + { + "bbox": [ + 277, + 783, + 301, + 793 + ], + "spans": [ + { + "bbox": [ + 277, + 783, + 301, + 793 + ], + "type": "text", + "content": "10004" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 595, + 842 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 103, + 126, + 146 + ], + "blocks": [ + { + "bbox": [ + 58, + 103, + 126, + 146 + ], + "lines": [ + { + "bbox": [ + 58, + 103, + 126, + 146 + ], + "spans": [ + { + "bbox": [ + 58, + 103, + 126, + 146 + ], + "type": "image", + "image_path": "4586c8a3fd80a808e0b05221866617bec8c08f87663c8ba05ecd8b9f2dec16b8.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 69, + 146, + 116, + 153 + ], + "lines": [ + { + "bbox": [ + 69, + 146, + 116, + 153 + ], + "spans": [ + { + "bbox": [ + 69, + 146, + 116, + 153 + ], + "type": "text", + "content": "Urban100 - #088" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 127, + 104, + 193, + 147 + ], + "blocks": [ + { + "bbox": [ + 127, + 104, + 193, + 147 + ], + "lines": [ + { + "bbox": [ + 127, + 104, + 193, + 147 + ], + "spans": [ + { + "bbox": [ + 127, + 104, + 193, + 147 + ], + "type": "image", + "image_path": "9e58391643991332fe4c329c0477c87267164c8bf4026a6daa07fbb3b998d566.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 155, + 147, + 165, + 153 + ], + "lines": [ + { + "bbox": [ + 155, + 147, + 165, + 153 + ], + "spans": [ + { + "bbox": [ + 155, + 147, + 165, + 153 + ], + "type": "text", + "content": "LR" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 195, + 105, + 262, + 147 + ], + "blocks": [ + { + "bbox": [ + 195, + 105, + 262, + 147 + ], + "lines": [ + { + "bbox": [ + 195, + 105, + 262, + 147 + ], + "spans": [ + { + "bbox": [ + 195, + 105, + 262, + 147 + ], + "type": "image", + "image_path": "3b426c19a5f7f535597822b7c85e804457a1fe9190335fed7d35b4e6186efc3a.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 223, + 147, + 233, + 153 + ], + "lines": [ + { + "bbox": [ + 223, + 147, + 233, + 153 + ], + "spans": [ + { + "bbox": [ + 223, + 147, + 233, + 153 + ], + "type": "text", + "content": "GT" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 58, + 154, + 126, + 197 + ], + "blocks": [ + { + "bbox": [ + 58, + 154, + 126, + 197 + ], + "lines": [ + { + "bbox": [ + 58, + 154, + 126, + 197 + ], + "spans": [ + { + "bbox": [ + 58, + 154, + 126, + 197 + ], + "type": "image", + "image_path": "c3c0031b00bf782a9266f44315c1319367590822520f5091bf3c5210cfe1467b.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 66, + 197, + 120, + 203 + ], + "lines": [ + { + "bbox": [ + 66, + 197, + 120, + 203 + ], + "spans": [ + { + "bbox": [ + 66, + 197, + 120, + 203 + ], + "type": "text", + "content": "SwinIR + LIIF × 4" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 127, + 154, + 193, + 197 + ], + "blocks": [ + { + "bbox": [ + 127, + 154, + 193, + 197 + ], + "lines": [ + { + "bbox": [ + 127, + 154, + 193, + 197 + ], + "spans": [ + { + "bbox": [ + 127, + 154, + 193, + 197 + ], + "type": "image", + "image_path": "3489f2945e0e53a2fc32604cb84c519932f56a4391e4b97683ed1586da5d6e7d.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 133, + 197, + 188, + 203 + ], + "lines": [ + { + "bbox": [ + 133, + 197, + 188, + 203 + ], + "spans": [ + { + "bbox": [ + 133, + 197, + 188, + 203 + ], + "type": "text", + "content": "SwinIR + LTE × 4" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 195, + 154, + 262, + 197 + ], + "blocks": [ + { + "bbox": [ + 195, + 154, + 262, + 197 + ], + "lines": [ + { + "bbox": [ + 195, + 154, + 262, + 197 + ], + "spans": [ + { + "bbox": [ + 195, + 154, + 262, + 197 + ], + "type": "image", + "image_path": "3002623944c9c75fcf9cf89766f4bcae9509a6357a4ff777aab2d150774f47ef.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 201, + 197, + 257, + 203 + ], + "lines": [ + { + "bbox": [ + 201, + 197, + 257, + 203 + ], + "spans": [ + { + "bbox": [ + 201, + 197, + 257, + 203 + ], + "type": "text", + "content": "SwinIR + CUF × 4" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 58, + 205, + 126, + 247 + ], + "blocks": [ + { + "bbox": [ + 58, + 205, + 126, + 247 + ], + "lines": [ + { + "bbox": [ + 58, + 205, + 126, + 247 + ], + "spans": [ + { + "bbox": [ + 58, + 205, + 126, + 247 + ], + "type": "image", + "image_path": "0231013cff2cc662bb784a6f52051ba45c6603dfb518f79458d4bf4649d26a36.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 73, + 247, + 111, + 253 + ], + "lines": [ + { + "bbox": [ + 73, + 247, + 111, + 253 + ], + "spans": [ + { + "bbox": [ + 73, + 247, + 111, + 253 + ], + "type": "text", + "content": "DIV2k - #853" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 127, + 205, + 193, + 247 + ], + "blocks": [ + { + "bbox": [ + 127, + 205, + 193, + 247 + ], + "lines": [ + { + "bbox": [ + 127, + 205, + 193, + 247 + ], + "spans": [ + { + "bbox": [ + 127, + 205, + 193, + 247 + ], + "type": "image", + "image_path": "b14bb37fa3c7c5e119d28d89350297f149cea13f2cd3e41bc65b34cf1aa666c3.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 157, + 247, + 165, + 253 + ], + "lines": [ + { + "bbox": [ + 157, + 247, + 165, + 253 + ], + "spans": [ + { + "bbox": [ + 157, + 247, + 165, + 253 + ], + "type": "text", + "content": "LR" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 195, + 205, + 262, + 247 + ], + "blocks": [ + { + "bbox": [ + 195, + 205, + 262, + 247 + ], + "lines": [ + { + "bbox": [ + 195, + 205, + 262, + 247 + ], + "spans": [ + { + "bbox": [ + 195, + 205, + 262, + 247 + ], + "type": "image", + "image_path": "732f5b867c3ac7f4bb15e239019a7e49fa0acb48d9797a949f0e04c124f7835f.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 224, + 247, + 233, + 253 + ], + "lines": [ + { + "bbox": [ + 224, + 247, + 233, + 253 + ], + "spans": [ + { + "bbox": [ + 224, + 247, + 233, + 253 + ], + "type": "text", + "content": "GT" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 58, + 255, + 126, + 297 + ], + "blocks": [ + { + "bbox": [ + 58, + 255, + 126, + 297 + ], + "lines": [ + { + "bbox": [ + 58, + 255, + 126, + 297 + ], + "spans": [ + { + "bbox": [ + 58, + 255, + 126, + 297 + ], + "type": "image", + "image_path": "1b961892fd03dfcdd62a1a80bb8179219db93b92446baf3d6e06843c45cae315.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 69, + 297, + 116, + 303 + ], + "lines": [ + { + "bbox": [ + 69, + 297, + 116, + 303 + ], + "spans": [ + { + "bbox": [ + 69, + 297, + 116, + 303 + ], + "type": "text", + "content": "RDN + LIIF × 8" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 127, + 255, + 193, + 297 + ], + "blocks": [ + { + "bbox": [ + 127, + 255, + 193, + 297 + ], + "lines": [ + { + "bbox": [ + 127, + 255, + 193, + 297 + ], + "spans": [ + { + "bbox": [ + 127, + 255, + 193, + 297 + ], + "type": "image", + "image_path": "41cedf49c99746d1b1cdc083e689ded0c336299f748cb3e145cfa13fd7d0c819.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 137, + 297, + 184, + 303 + ], + "lines": [ + { + "bbox": [ + 137, + 297, + 184, + 303 + ], + "spans": [ + { + "bbox": [ + 137, + 297, + 184, + 303 + ], + "type": "text", + "content": "RDN+LTE×8" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 195, + 255, + 262, + 297 + ], + "blocks": [ + { + "bbox": [ + 195, + 255, + 262, + 297 + ], + "lines": [ + { + "bbox": [ + 195, + 255, + 262, + 297 + ], + "spans": [ + { + "bbox": [ + 195, + 255, + 262, + 297 + ], + "type": "image", + "image_path": "a845e23a0ef19ad4f1cf9defee4a993dab0cd79e60d0843f6750ff55b3e255e6.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 204, + 297, + 252, + 303 + ], + "lines": [ + { + "bbox": [ + 204, + 297, + 252, + 303 + ], + "spans": [ + { + "bbox": [ + 204, + 297, + 252, + 303 + ], + "type": "text", + "content": "RDN+CUF×8" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 58, + 304, + 126, + 346 + ], + "blocks": [ + { + "bbox": [ + 58, + 304, + 126, + 346 + ], + "lines": [ + { + "bbox": [ + 58, + 304, + 126, + 346 + ], + "spans": [ + { + "bbox": [ + 58, + 304, + 126, + 346 + ], + "type": "image", + "image_path": "9308183792fdc1bbce259fe7dfb0f4f3ed5e06699023a8e4bf38fd71835603d0.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 69, + 347, + 116, + 353 + ], + "lines": [ + { + "bbox": [ + 69, + 347, + 116, + 353 + ], + "spans": [ + { + "bbox": [ + 69, + 347, + 116, + 353 + ], + "type": "text", + "content": "Urban100 - #030" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 127, + 304, + 193, + 346 + ], + "blocks": [ + { + "bbox": [ + 127, + 304, + 193, + 346 + ], + "lines": [ + { + "bbox": [ + 127, + 304, + 193, + 346 + ], + "spans": [ + { + "bbox": [ + 127, + 304, + 193, + 346 + ], + "type": "image", + "image_path": "a912ae22bde445f82745922ded934e03c7ce57cec0e86958df726d100d43d4a1.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 156, + 347, + 164, + 352 + ], + "lines": [ + { + "bbox": [ + 156, + 347, + 164, + 352 + ], + "spans": [ + { + "bbox": [ + 156, + 347, + 164, + 352 + ], + "type": "text", + "content": "LR" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_caption" + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 195, + 304, + 262, + 347 + ], + "blocks": [ + { + "bbox": [ + 195, + 304, + 262, + 347 + ], + "lines": [ + { + "bbox": [ + 195, + 304, + 262, + 347 + ], + "spans": [ + { + "bbox": [ + 195, + 304, + 262, + 347 + ], + "type": "image", + "image_path": "0f546039c6d66d558b33a3e7ef670331c80ac329e000d15184683c24bde54012.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 224, + 348, + 233, + 353 + ], + "lines": [ + { + "bbox": [ + 224, + 348, + 233, + 353 + ], + "spans": [ + { + "bbox": [ + 224, + 348, + 233, + 353 + ], + "type": "text", + "content": "GT" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_caption" + } + ], + "index": 28 + }, + { + "type": "image", + "bbox": [ + 58, + 355, + 126, + 396 + ], + "blocks": [ + { + "bbox": [ + 58, + 355, + 126, + 396 + ], + "lines": [ + { + "bbox": [ + 58, + 355, + 126, + 396 + ], + "spans": [ + { + "bbox": [ + 58, + 355, + 126, + 396 + ], + "type": "image", + "image_path": "23d15cc28ecf17fb84b27930ca53c31806084e0d2d98e1a59b2fe41f854199cf.jpg" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 67, + 397, + 118, + 403 + ], + "lines": [ + { + "bbox": [ + 67, + 397, + 118, + 403 + ], + "spans": [ + { + "bbox": [ + 67, + 397, + 118, + 403 + ], + "type": "text", + "content": "EDSR + LIF × 4" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_caption" + } + ], + "index": 30 + }, + { + "type": "image", + "bbox": [ + 127, + 355, + 193, + 396 + ], + "blocks": [ + { + "bbox": [ + 127, + 355, + 193, + 396 + ], + "lines": [ + { + "bbox": [ + 127, + 355, + 193, + 396 + ], + "spans": [ + { + "bbox": [ + 127, + 355, + 193, + 396 + ], + "type": "image", + "image_path": "000f5e5546064e7592b68cd292b4e5d796b3d1b65449ffad0ab27b59ebfe3cdb.jpg" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 136, + 397, + 186, + 403 + ], + "lines": [ + { + "bbox": [ + 136, + 397, + 186, + 403 + ], + "spans": [ + { + "bbox": [ + 136, + 397, + 186, + 403 + ], + "type": "text", + "content": "EDSR + LTE × 4" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_caption" + } + ], + "index": 32 + }, + { + "type": "image", + "bbox": [ + 195, + 355, + 262, + 396 + ], + "blocks": [ + { + "bbox": [ + 195, + 355, + 262, + 396 + ], + "lines": [ + { + "bbox": [ + 195, + 355, + 262, + 396 + ], + "spans": [ + { + "bbox": [ + 195, + 355, + 262, + 396 + ], + "type": "image", + "image_path": "2f603dc4fc57483bcddf2a1a09630fc3c1ee252b353a16d7b584cbe3f2a4ed57.jpg" + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 203, + 397, + 255, + 403 + ], + "lines": [ + { + "bbox": [ + 203, + 397, + 255, + 403 + ], + "spans": [ + { + "bbox": [ + 203, + 397, + 255, + 403 + ], + "type": "text", + "content": "EDSR + CUF × 4" + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "image_caption" + } + ], + "index": 34 + }, + { + "type": "image", + "bbox": [ + 57, + 431, + 125, + 475 + ], + "blocks": [ + { + "bbox": [ + 45, + 405, + 278, + 426 + ], + "lines": [ + { + "bbox": [ + 45, + 405, + 278, + 426 + ], + "spans": [ + { + "bbox": [ + 45, + 405, + 278, + 426 + ], + "type": "text", + "content": "Figure 7. Qualitative evaluation - CUF combined with different encoders and upsampling scales." + } + ] + } + ], + "index": 36, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 57, + 431, + 125, + 475 + ], + "lines": [ + { + "bbox": [ + 57, + 431, + 125, + 475 + ], + "spans": [ + { + "bbox": [ + 57, + 431, + 125, + 475 + ], + "type": "image", + "image_path": "41369710361a31dac81168f56ea1109c5d3aca11a2b02c3536eaa8c8a2eb3bcd.jpg" + } + ] + } + ], + "index": 37, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 45, + 480, + 278, + 502 + ], + "lines": [ + { + "bbox": [ + 45, + 480, + 278, + 502 + ], + "spans": [ + { + "bbox": [ + 45, + 480, + 278, + 502 + ], + "type": "text", + "content": "Figure 8. Residuals - between ground truth and (from left to right) results from LIIF, LTE and CUF shown in Fig 7." + } + ] + } + ], + "index": 40, + "angle": 0, + "type": "image_caption" + } + ], + "index": 37 + }, + { + "type": "image", + "bbox": [ + 128, + 431, + 196, + 475 + ], + "blocks": [ + { + "bbox": [ + 128, + 431, + 196, + 475 + ], + "lines": [ + { + "bbox": [ + 128, + 431, + 196, + 475 + ], + "spans": [ + { + "bbox": [ + 128, + 431, + 196, + 475 + ], + "type": "image", + "image_path": "09137e00bd551d75b0d426599bebaab0afe01fbceeaef04d8cbfee5f3fa3671f.jpg" + } + ] + } + ], + "index": 38, + "angle": 0, + "type": "image_body" + } + ], + "index": 38 + }, + { + "type": "image", + "bbox": [ + 199, + 431, + 268, + 475 + ], + "blocks": [ + { + "bbox": [ + 199, + 431, + 268, + 475 + ], + "lines": [ + { + "bbox": [ + 199, + 431, + 268, + 475 + ], + "spans": [ + { + "bbox": [ + 199, + 431, + 268, + 475 + ], + "type": "image", + "image_path": "057e4f837f989a538cc20928506cd0c73e11a4e1b5c46c3726585ba596eadb7a.jpg" + } + ] + } + ], + "index": 39, + "angle": 0, + "type": "image_body" + } + ], + "index": 39 + }, + { + "bbox": [ + 45, + 511, + 278, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 511, + 278, + 593 + ], + "spans": [ + { + "bbox": [ + 45, + 511, + 278, + 593 + ], + "type": "text", + "content": "generalize to unseen or fractional scales. On the remaining rows we present results of arbitrary-scale upsamplers (MetaSR [16], LIIF [6], LTE [19]), where a single model is trained and tested across different upsampling ratios. Our method is the first arbitrary-scale method to match (or surpass) fixed-scale performance across all encoders in both single-pass inference and self-ensemble scenarios." + } + ] + } + ], + "index": 41 + }, + { + "type": "table", + "bbox": [ + 47, + 600, + 278, + 691 + ], + "blocks": [ + { + "bbox": [ + 47, + 600, + 278, + 691 + ], + "lines": [ + { + "bbox": [ + 47, + 600, + 278, + 691 + ], + "spans": [ + { + "bbox": [ + 47, + 600, + 278, + 691 + ], + "type": "table", + "html": "
UpsamplerUses Periodicity\\( \\times 2 \\)\\( \\times 3 \\)\\( \\times 4 \\)
WCGPUWCGPUWCGPU
EDSR-baseline encoder
Sub-pixel conv. **✓10.96.411.07.311.69.0
CUF inst. (ours) **✓11.16.511.27.611.09.1
LIF75.574.0164.6163.1295.0293.4
LTE54.845.9101.792.5167.0157.8
CUF (ours) - k as input26.718.852.444.081.573.0
CUF (ours) - k as output20.212.227.219.837.630.0
ABPN encoder
Sub-pixel conv. **✓1.90.61.90.61.90.7
CUF inst. (ours) **✓2.10.72.11.12.11.5
", + "image_path": "248a3659190d6dfb9cfbd087c8a6bc04d25f006912b7f65497f0eb30cfff7bf1.jpg" + } + ] + } + ], + "index": 42, + "angle": 0, + "type": "table_body" + } + ], + "index": 42 + }, + { + "bbox": [ + 45, + 693, + 278, + 726 + ], + "lines": [ + { + "bbox": [ + 45, + 693, + 278, + 726 + ], + "spans": [ + { + "bbox": [ + 45, + 693, + 278, + 726 + ], + "type": "text", + "content": "Table 5. Upsampling " + }, + { + "bbox": [ + 45, + 693, + 278, + 726 + ], + "type": "inline_equation", + "content": "(256^{2})" + }, + { + "bbox": [ + 45, + 693, + 278, + 726 + ], + "type": "text", + "content": " on a A100 GPU. Wall clock (wc) (i.e., considering both CPU and GPU execution) and GPU time in ms averaged over " + }, + { + "bbox": [ + 45, + 693, + 278, + 726 + ], + "type": "inline_equation", + "content": "1k" + }, + { + "bbox": [ + 45, + 693, + 278, + 726 + ], + "type": "text", + "content": " runs." + } + ] + } + ], + "index": 43, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 309, + 103, + 506, + 271 + ], + "blocks": [ + { + "bbox": [ + 309, + 103, + 506, + 271 + ], + "lines": [ + { + "bbox": [ + 309, + 103, + 506, + 271 + ], + "spans": [ + { + "bbox": [ + 309, + 103, + 506, + 271 + ], + "type": "image", + "image_path": "07f5c6bfac8c1d35f9c972ab3023874823312fa45b26dfd9e262507cad3e1f1f.jpg" + } + ] + } + ], + "index": 44, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 408, + 261, + 518, + 271 + ], + "lines": [ + { + "bbox": [ + 408, + 261, + 518, + 271 + ], + "spans": [ + { + "bbox": [ + 408, + 261, + 518, + 271 + ], + "type": "text", + "content": "ABPN + Continuous Upsampler" + } + ] + } + ], + "index": 45, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 296, + 274, + 530, + 328 + ], + "lines": [ + { + "bbox": [ + 296, + 274, + 530, + 328 + ], + "spans": [ + { + "bbox": [ + 296, + 274, + 530, + 328 + ], + "type": "text", + "content": "Figure 9. Architectures (lightweight) - The original ABPN [11] single scale architecture (left) and our arbitrary-scale version using the implicit upsampler (right). Note how the two consecutive " + }, + { + "bbox": [ + 296, + 274, + 530, + 328 + ], + "type": "inline_equation", + "content": "3 \\times 3" + }, + { + "bbox": [ + 296, + 274, + 530, + 328 + ], + "type": "text", + "content": " native convolutions are exchanged for a " + }, + { + "bbox": [ + 296, + 274, + 530, + 328 + ], + "type": "inline_equation", + "content": "3 \\times 3" + }, + { + "bbox": [ + 296, + 274, + 530, + 328 + ], + "type": "text", + "content": " depth-wise implicit convolution followed by point-wise layers." + } + ] + } + ], + "index": 46, + "angle": 0, + "type": "image_caption" + } + ], + "index": 44 + }, + { + "bbox": [ + 297, + 337, + 450, + 351 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 337, + 450, + 351 + ], + "spans": [ + { + "bbox": [ + 297, + 337, + 450, + 351 + ], + "type": "text", + "content": "4.2. Lightweight super-resolution" + } + ] + } + ], + "index": 47 + }, + { + "bbox": [ + 296, + 356, + 530, + 447 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 356, + 530, + 447 + ], + "spans": [ + { + "bbox": [ + 296, + 356, + 530, + 447 + ], + "type": "text", + "content": "The encoders and upsamplers considered so far are not designed to be run on devices with limited memory and power. Running super-resolution algorithms on mobile devices comes with unique challenges due to limited RAM and non-efficient support for certain common operations [18]. The proposed CUF-based upsampler, in combination with a mobile-compatible encoder, can be effectively used for upscaling on mobile devices." + } + ] + } + ], + "index": 48 + }, + { + "bbox": [ + 296, + 449, + 530, + 588 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 449, + 530, + 588 + ], + "spans": [ + { + "bbox": [ + 296, + 449, + 530, + 588 + ], + "type": "text", + "content": "To test the effectiveness of the CUFs, we adopt a lightweight version of SwinIR (0.9M parameters) [20] as well as one of the best performing super-lightweight architectures for mobile devices (30K parameters) [10] as the encoder. For the latter, the original architecture (Figure 9, left) consists of a lightweight encoder and an upscaling module consisting of a series of convolutions, ReLU's, and a standard periodic shuffle layer. We replace this upscaling module with a CUF-based one (Figure 9, right). With instantiated CUFs, this means simply introducing a depthwise convolution followed by pointwise layers, where the weights are output by the learned CUFs." + } + ] + } + ], + "index": 49 + }, + { + "bbox": [ + 296, + 588, + 530, + 634 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 588, + 530, + 634 + ], + "spans": [ + { + "bbox": [ + 296, + 588, + 530, + 634 + ], + "type": "text", + "content": "Table 6 illustrates that CUFs provide the same level of accuracy while enabling continuous scale and efficient inference. To our knowledge, this is the first mobile-friendly continuous scale super-resolution architecture." + } + ] + } + ], + "index": 50 + }, + { + "bbox": [ + 297, + 642, + 363, + 653 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 642, + 363, + 653 + ], + "spans": [ + { + "bbox": [ + 297, + 642, + 363, + 653 + ], + "type": "text", + "content": "4.3. Ablations" + } + ] + } + ], + "index": 51 + }, + { + "bbox": [ + 296, + 660, + 530, + 730 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 660, + 530, + 730 + ], + "spans": [ + { + "bbox": [ + 296, + 660, + 530, + 730 + ], + "type": "text", + "content": "We run several ablation experiments in order to analyze the effects of various model choices. Here, we detail our choice of using DCT for positional encoding vs Fourier features, as well as an investigation of the redundancy of CUF filters versus subpixel convolution. An analysis on the various conditioning factors of our neural fields (positional en" + } + ] + } + ], + "index": 52 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 277, + 783, + 300, + 793 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 277, + 783, + 300, + 793 + ], + "spans": [ + { + "bbox": [ + 277, + 783, + 300, + 793 + ], + "type": "text", + "content": "10005" + } + ] + } + ], + "index": 53 + } + ], + "page_size": [ + 595, + 842 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 48, + 111, + 529, + 162 + ], + "blocks": [ + { + "bbox": [ + 246, + 104, + 330, + 111 + ], + "lines": [ + { + "bbox": [ + 246, + 104, + 330, + 111 + ], + "spans": [ + { + "bbox": [ + 246, + 104, + 330, + 111 + ], + "type": "text", + "content": "Multi scale up-sampling methods" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 48, + 111, + 529, + 162 + ], + "lines": [ + { + "bbox": [ + 48, + 111, + 529, + 162 + ], + "spans": [ + { + "bbox": [ + 48, + 111, + 529, + 162 + ], + "type": "table", + "html": "
EncoderUpsamplerSet5Set14BSD100Urban100
x2x3x4x6x2x3x4x6x2x3x4x6x2x3x4x6
SwintIR-light [20]Sub-Pixel Conv.38.1134.5432.35-33.8330.5328.76-32.2729.1927.68-32.5628.5826.48-
CUF (ours)38.1734.6232.4629.0933.9330.5728.8326.6132.3029.2327.7125.9532.7328.6826.5524.11
ABPN [10]Sub-Pixel Conv.37.3033.3931.11-32.9029.6827.97-31.6728.6427.14-30.3726.7724.96-
CUF (ours)37.3533.5531.3128.0632.9629.7428.0425.8631.7228.6927.1925.5230.5326.9225.0523.01
", + "image_path": "3fef5d493e7bd622eb96fd2fee14e7c7b772d4ee3b754eea364b89ddeea7332b.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 94, + 190, + 232, + 233 + ], + "blocks": [ + { + "bbox": [ + 45, + 165, + 529, + 187 + ], + "lines": [ + { + "bbox": [ + 45, + 165, + 529, + 187 + ], + "spans": [ + { + "bbox": [ + 45, + 165, + 529, + 187 + ], + "type": "text", + "content": "Table 6. Lightweight super-res - Our CUFs enable continuous super-resolution with lightweight (SwinIR-lightweight [20]) and mobile-friendly (ABPN [10]) backbones without quality loss." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 94, + 190, + 232, + 233 + ], + "lines": [ + { + "bbox": [ + 94, + 190, + 232, + 233 + ], + "spans": [ + { + "bbox": [ + 94, + 190, + 232, + 233 + ], + "type": "table", + "html": "
2×3×4×
Fourier30.4526.8525.00
DCT30.4426.8625.00
", + "image_path": "06e88f4a47fcb3751ab05af46da0739cc85bef84c16c741ab2649c5ae2b31f51.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 45, + 236, + 279, + 269 + ], + "lines": [ + { + "bbox": [ + 45, + 236, + 279, + 269 + ], + "spans": [ + { + "bbox": [ + 45, + 236, + 279, + 269 + ], + "type": "text", + "content": "Table 7. Cosine basis - matching results but using half of sinusoidal projections and neurons at the hyper-network input. PSNR from Urban-100 dataset using a ABPN encoder." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 45, + 292, + 279, + 316 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 292, + 279, + 316 + ], + "spans": [ + { + "bbox": [ + 45, + 292, + 279, + 316 + ], + "type": "text", + "content": "coding, kernel indices, scale, and target subpixel) can be found in the Supplementary Material." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 45, + 321, + 279, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 321, + 279, + 392 + ], + "spans": [ + { + "bbox": [ + 45, + 321, + 279, + 392 + ], + "type": "text", + "content": "Positional encoding. A large body of work uses a Fourier basis for the positional encodings, e.g. [24, 33]. However, we opt to use a DCT basis as this requires half of the sinusoidal projections and neurons at the hyper-network first layer. This trick has been used in other domains, such as random approximations of stationary kernel functions [27]." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 45, + 393, + 279, + 440 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 393, + 279, + 440 + ], + "spans": [ + { + "bbox": [ + 45, + 393, + 279, + 440 + ], + "type": "text", + "content": "Table 7 shows the PSNR, as obtained using a DFT versus a DCT basis, and verify that we do not observe loss in performance, while at the same time saving half of the computation associated with positional encoding operations." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 45, + 445, + 279, + 609 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 445, + 279, + 609 + ], + "spans": [ + { + "bbox": [ + 45, + 445, + 279, + 609 + ], + "type": "text", + "content": "Redundancy of filters. As noted in Section 3.4, sub-pixel convolution does not enforce spatial correlation between neighboring weights in its convolutional filters, but has to learn them from data. CUF on the other hand generates the filters with a hypernetwork, which intrinsically builds on smoothness due to its functional form. Here, we test this hypothesis on the filters of CUF and sub-pixel convolutions. For this analysis, we use an EDSR encoder with " + }, + { + "bbox": [ + 45, + 445, + 279, + 609 + ], + "type": "inline_equation", + "content": "C_e = 64" + }, + { + "bbox": [ + 45, + 445, + 279, + 609 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 45, + 445, + 279, + 609 + ], + "type": "inline_equation", + "content": "K = 3 \\times 3" + }, + { + "bbox": [ + 45, + 445, + 279, + 609 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 45, + 445, + 279, + 609 + ], + "type": "inline_equation", + "content": "s = 3" + }, + { + "bbox": [ + 45, + 445, + 279, + 609 + ], + "type": "text", + "content": ". Subpixel convolution maps " + }, + { + "bbox": [ + 45, + 445, + 279, + 609 + ], + "type": "inline_equation", + "content": "C_e" + }, + { + "bbox": [ + 45, + 445, + 279, + 609 + ], + "type": "text", + "content": " features to " + }, + { + "bbox": [ + 45, + 445, + 279, + 609 + ], + "type": "inline_equation", + "content": "s^2 C_e" + }, + { + "bbox": [ + 45, + 445, + 279, + 609 + ], + "type": "text", + "content": " features, and each group of " + }, + { + "bbox": [ + 45, + 445, + 279, + 609 + ], + "type": "inline_equation", + "content": "s^2" + }, + { + "bbox": [ + 45, + 445, + 279, + 609 + ], + "type": "text", + "content": " output features are rearranged to form the upscaled feature space. We therefore consider each group of " + }, + { + "bbox": [ + 45, + 445, + 279, + 609 + ], + "type": "inline_equation", + "content": "s^2" + }, + { + "bbox": [ + 45, + 445, + 279, + 609 + ], + "type": "text", + "content": " filters of size " + }, + { + "bbox": [ + 45, + 445, + 279, + 609 + ], + "type": "inline_equation", + "content": "K \\times K \\times C_e" + }, + { + "bbox": [ + 45, + 445, + 279, + 609 + ], + "type": "text", + "content": " and do an eigenvalue analysis to determine the compressibility of these filters. This is done by forming " + }, + { + "bbox": [ + 45, + 445, + 279, + 609 + ], + "type": "inline_equation", + "content": "C_e" + }, + { + "bbox": [ + 45, + 445, + 279, + 609 + ], + "type": "text", + "content": " separate matrix" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 48, + 619, + 162, + 689 + ], + "blocks": [ + { + "bbox": [ + 48, + 619, + 162, + 689 + ], + "lines": [ + { + "bbox": [ + 48, + 619, + 162, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 619, + 162, + 689 + ], + "type": "image", + "image_path": "7df9f14026dccd79d65aa4f295e919f21d391a9292f961e8579a336ce214c660.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 69, + 692, + 141, + 700 + ], + "lines": [ + { + "bbox": [ + 69, + 692, + 141, + 700 + ], + "spans": [ + { + "bbox": [ + 69, + 692, + 141, + 700 + ], + "type": "text", + "content": "(a) Subpixel Convolutions" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 165, + 619, + 278, + 690 + ], + "blocks": [ + { + "bbox": [ + 165, + 619, + 278, + 690 + ], + "lines": [ + { + "bbox": [ + 165, + 619, + 278, + 690 + ], + "spans": [ + { + "bbox": [ + 165, + 619, + 278, + 690 + ], + "type": "image", + "image_path": "9847b4ff361a573ff20a3fc325b213fb34a6b6ab58e942ec7dec555bd6e2fb11.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 208, + 692, + 232, + 700 + ], + "lines": [ + { + "bbox": [ + 208, + 692, + 232, + 700 + ], + "spans": [ + { + "bbox": [ + 208, + 692, + 232, + 700 + ], + "type": "text", + "content": "(b) CUF" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 45, + 704, + 279, + 726 + ], + "lines": [ + { + "bbox": [ + 45, + 704, + 279, + 726 + ], + "spans": [ + { + "bbox": [ + 45, + 704, + 279, + 726 + ], + "type": "text", + "content": "Figure 10. CUF depthwise continuous filters are low rank by construction. Subpixel-convolution fails to find low rank filters." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 296, + 191, + 530, + 262 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 191, + 530, + 262 + ], + "spans": [ + { + "bbox": [ + 296, + 191, + 530, + 262 + ], + "type": "text", + "content": "ces of size " + }, + { + "bbox": [ + 296, + 191, + 530, + 262 + ], + "type": "inline_equation", + "content": "s^2 \\times C_e K^2" + }, + { + "bbox": [ + 296, + 191, + 530, + 262 + ], + "type": "text", + "content": " and performing PCA on each of them separately, forming 64 separate sets of eigenvalues. We plot this distribution for trained and untrained sub-pixel layers in Figure 10a. A faster decay means more redundancy, as more variance is explained by fewer components. In this case, training adds significant structure to the filters." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 296, + 266, + 531, + 382 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 266, + 531, + 382 + ], + "spans": [ + { + "bbox": [ + 296, + 266, + 531, + 382 + ], + "type": "text", + "content": "Conversely, each output channel in CUF is computed by " + }, + { + "bbox": [ + 296, + 266, + 531, + 382 + ], + "type": "inline_equation", + "content": "s^2" + }, + { + "bbox": [ + 296, + 266, + 531, + 382 + ], + "type": "text", + "content": " filters of size " + }, + { + "bbox": [ + 296, + 266, + 531, + 382 + ], + "type": "inline_equation", + "content": "K \\times K" + }, + { + "bbox": [ + 296, + 266, + 531, + 382 + ], + "type": "text", + "content": ". These are not directly comparable to subpixel convolution, but we perform an analogous eigenvalue analysis on trained and untrained CUFs. As shown in Figure 10b, unlike subpixel convolution, CUFs actually become less redundant with training, suggesting that the initial hypernetwork already imposes a very smooth prior on the filters. In other words, CUF filters incorporate spatial correlations from the very beginning due to their functional form." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 297, + 402, + 373, + 414 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 402, + 373, + 414 + ], + "spans": [ + { + "bbox": [ + 297, + 402, + 373, + 414 + ], + "type": "text", + "content": "5. Conclusions" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 296, + 423, + 530, + 691 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 423, + 530, + 691 + ], + "spans": [ + { + "bbox": [ + 296, + 423, + 530, + 691 + ], + "type": "text", + "content": "We propose CUF, a computationally efficient modeling of continuous upsampling filters as neural fields. Different from previous arbitrary-scale architectures, the performance gains obtained by our upsampler are not due to increased capacity but rather to a more efficient use of model parameters. A single hyper-network supports filter adaptation across scales while at the same time CUF's upsampling head has fewer parameters than previous arbitrary-scale heads, as well as fewer parameters than a single SubPixel Convolution head at " + }, + { + "bbox": [ + 296, + 423, + 530, + 691 + ], + "type": "inline_equation", + "content": "4 \\times" + }, + { + "bbox": [ + 296, + 423, + 530, + 691 + ], + "type": "text", + "content": " upsampling. CUF is the first arbitrary scale model that can be effectively used for upscaling on mobile devices, an area previously dominated by single scale Sub-Pixel Convolution architectures. CUF's architecture is quite general. A natural direction of future work is to investigate its use for architectures with multiple upsampling layers (and the reuse of filters across scales), for different image generation settings such as GANs [12] and diffusion models [15]. Neural fields have done particularly well on higher-dimensional signals. Another line of future work is the use of CUF for efficient super-resolution of video data. Continuous convolution filters are an exciting and increasingly studied area of research [28,29] with many applications as an alternative to traditional convolutions." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 296, + 695, + 531, + 729 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 695, + 531, + 729 + ], + "spans": [ + { + "bbox": [ + 296, + 695, + 531, + 729 + ], + "type": "text", + "content": "Overall, the presented results aligned with low computational resources suggest CUF offers a practical modeling of continuous filters." + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 277, + 783, + 301, + 793 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 277, + 783, + 301, + 793 + ], + "spans": [ + { + "bbox": [ + 277, + 783, + 301, + 793 + ], + "type": "text", + "content": "10006" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 595, + 842 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 105, + 103, + 117 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 105, + 103, + 117 + ], + "spans": [ + { + "bbox": [ + 47, + 105, + 103, + 117 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 124, + 280, + 730 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 51, + 124, + 279, + 145 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 51, + 124, + 279, + 145 + ], + "spans": [ + { + "bbox": [ + 51, + 124, + 279, + 145 + ], + "type": "text", + "content": "[1] Manoj Alwani, Han Chen, Michael Ferdman, and Peter Milder. Fused-layer cnn accelerators. 10 2016. 13" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 51, + 147, + 279, + 178 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 51, + 147, + 279, + 178 + ], + "spans": [ + { + "bbox": [ + 51, + 147, + 279, + 178 + ], + "type": "text", + "content": "[2] Saeed Anwar, Salman Khan, and Nick Barnes. A deep journey into super-resolution: A survey. ACM Computing Surveys (ACMCSUR), 53(3), May 2020. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 51, + 179, + 280, + 221 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 51, + 179, + 280, + 221 + ], + "spans": [ + { + "bbox": [ + 51, + 179, + 280, + 221 + ], + "type": "text", + "content": "[3] Syed Muhammad Arsalan Bashir, Yi Wang, Mahrukh Khan, and Yilong Niu. A comprehensive review of deep learning-based single image super-resolution. PeerJ Computer Science, 7:e621, Jul 2021. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 51, + 222, + 280, + 275 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 51, + 222, + 280, + 275 + ], + "spans": [ + { + "bbox": [ + 51, + 222, + 280, + 275 + ], + "type": "text", + "content": "[4] Marco Bevilacqua, Aline Roumy, Christine Guillemot, and Marie line Alberi Morel. Low-complexity single-image super-resolution based on nonnegative neighbor embedding. In Proceedings of the British Machine Vision Conference, pages 135.1-135.10. BMVA Press, 2012. 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 51, + 277, + 280, + 298 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 51, + 277, + 280, + 298 + ], + "spans": [ + { + "bbox": [ + 51, + 277, + 280, + 298 + ], + "type": "text", + "content": "[5] Alexandre Boulch. Convpoint: Continuous convolutions for point cloud processing. Computers & Graphics, 2020. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 51, + 298, + 279, + 351 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 51, + 298, + 279, + 351 + ], + "spans": [ + { + "bbox": [ + 51, + 298, + 279, + 351 + ], + "type": "text", + "content": "[6] Yinbo Chen, Sifei Liu, and Xiaolong Wang. Learning continuous image representation with local implicit image function. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8628-8638, 2021. 1, 2, 3, 4, 5, 7, 11" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 51, + 351, + 280, + 405 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 51, + 351, + 280, + 405 + ], + "spans": [ + { + "bbox": [ + 51, + 351, + 280, + 405 + ], + "type": "text", + "content": "[7] Tao Dai, Jianrui Cai, Yongbing Zhang, Shu-Tao Xia, and Lei Zhang. Second-order attention network for single image super-resolution. In 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11057-11066, 2019. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 51, + 405, + 280, + 469 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 51, + 405, + 280, + 469 + ], + "spans": [ + { + "bbox": [ + 51, + 405, + 280, + 469 + ], + "type": "text", + "content": "[8] Chao Dong, Chen Change Loy, Kaiming He, and Xiaou Tang. Learning a deep convolutional network for image super-resolution. In David Fleet, Tomas Pajdla, Bernt Schiele, and Tinne Tuytelaars, editors, Computer Vision – ECCV 2014, pages 184–199, Cham, 2014. Springer International Publishing. 1, 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 51, + 470, + 280, + 512 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 51, + 470, + 280, + 512 + ], + "spans": [ + { + "bbox": [ + 51, + 470, + 280, + 512 + ], + "type": "text", + "content": "[9] Chao Dong, Chen Change Loy, Kaiming He, and Xiaou Tang. Image super-resolution using deep convolutional networks. IEEE Trans. Pattern Anal. Mach. Intell., 38(2):295-307, feb 2016. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 514, + 279, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 514, + 279, + 566 + ], + "spans": [ + { + "bbox": [ + 47, + 514, + 279, + 566 + ], + "type": "text", + "content": "[10] Zongcai Du, Jie Liu, Jie Tang, and Gangshan Wu. Anchors-based plain net for mobile image super-resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, pages 2494-2502, June 2021. 2, 6, 7, 8, 11" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 568, + 279, + 620 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 568, + 279, + 620 + ], + "spans": [ + { + "bbox": [ + 47, + 568, + 279, + 620 + ], + "type": "text", + "content": "[11] Z. Du, J. Liu, J. Tang, and G. Wu. Anchor-based plain net for mobile image super-resolution. In 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), pages 2494-2502, Los Alamitos, CA, USA, jun 2021. IEEE Computer Society. 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 47, + 621, + 279, + 664 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 621, + 279, + 664 + ], + "spans": [ + { + "bbox": [ + 47, + 621, + 279, + 664 + ], + "type": "text", + "content": "[12] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial networks. Communications of the ACM, 63(11):139-144, 2020. 8" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 47, + 665, + 279, + 686 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 665, + 279, + 686 + ], + "spans": [ + { + "bbox": [ + 47, + 665, + 279, + 686 + ], + "type": "text", + "content": "[13] David Ha, Andrew Dai, and Quoc V Le. Hypernetworks. arXiv preprint arXiv:1609.09106, 2016. 4" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 47, + 687, + 279, + 730 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 687, + 279, + 730 + ], + "spans": [ + { + "bbox": [ + 47, + 687, + 279, + 730 + ], + "type": "text", + "content": "[14] Pedro Hermosilla, Tobias Ritschel, Pere-Pau Vázquez, Álvar Vinacua, and Timo Ropinski. Monte carlo convolution for learning on non-uniformly sampled point clouds. ACM Trans. Graph., 37(6), dec 2018. 1, 2" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 299, + 106, + 531, + 730 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 299, + 106, + 530, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 106, + 530, + 139 + ], + "spans": [ + { + "bbox": [ + 299, + 106, + 530, + 139 + ], + "type": "text", + "content": "[15] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in Neural Information Processing Systems, 33:6840-6851, 2020. 8" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 299, + 141, + 531, + 193 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 141, + 531, + 193 + ], + "spans": [ + { + "bbox": [ + 299, + 141, + 531, + 193 + ], + "type": "text", + "content": "[16] Xuecai Hu, Haoyuan Mu, Xiangyu Zhang, Zilei Wang, Tieniu Tan, and Jian Sun. Meta-sr: A magnification-arbitrary network for super-resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2019. 1, 2, 3, 5, 7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 299, + 195, + 530, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 195, + 530, + 237 + ], + "spans": [ + { + "bbox": [ + 299, + 195, + 530, + 237 + ], + "type": "text", + "content": "[17] Jia-Bin Huang, Abhishek Singh, and Narendra Ahuja. Single image super-resolution from transformed self-exemplars. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 5197-5206, 2015. 6, 11" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 299, + 239, + 530, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 239, + 530, + 335 + ], + "spans": [ + { + "bbox": [ + 299, + 239, + 530, + 335 + ], + "type": "text", + "content": "[18] A. Ignatov, R. Timofte, M. Denna, A. Younes, A. Lek, M. Ayazoglu, J. Liu, Z. Du, J. Guo, X. Zhou, H. Jia, Y. Yan, Z. Zhang, Y. Chen, Y. Peng, Y. Lin, X. Zhang, H. Zeng, K. Zeng, P. Li, Z. Liu, S. Xue, and S. Wang. Real-time quantized image super-resolution on mobile npus, mobile ai 2021 challenge: Report. In 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), pages 2525-2534, Los Alamitos, CA, USA, jun 2021. IEEE Computer Society. 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 299, + 336, + 530, + 378 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 336, + 530, + 378 + ], + "spans": [ + { + "bbox": [ + 299, + 336, + 530, + 378 + ], + "type": "text", + "content": "[19] Kyong Hwan Jin Jaewon Lee. Local texture estimator for implicit representation function. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022. 1, 2, 3, 4, 5, 7, 11" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 299, + 380, + 530, + 423 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 380, + 530, + 423 + ], + "spans": [ + { + "bbox": [ + 299, + 380, + 530, + 423 + ], + "type": "text", + "content": "[20] Jingyun Liang, Jiezhang Cao, Guolei Sun, Kai Zhang, Luc Van Gool, and Radu Timofte. Swinir: Image restoration using swim transformer. In IEEE International Conference on Computer Vision Workshops, 2021. 1, 2, 5, 6, 7, 8, 11, 12" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 299, + 424, + 530, + 478 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 424, + 530, + 478 + ], + "spans": [ + { + "bbox": [ + 299, + 424, + 530, + 478 + ], + "type": "text", + "content": "[21] Bee Lim, Sanghyun Son, Heewon Kim, Seungjun Nah, and Young Mu Lee. Enhanced deep residual networks for single image super-resolution. In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, July 2017. 1, 2, 3, 5, 6, 11, 12" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 299, + 479, + 530, + 532 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 479, + 530, + 532 + ], + "spans": [ + { + "bbox": [ + 299, + 479, + 530, + 532 + ], + "type": "text", + "content": "[22] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 2, 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 299, + 533, + 530, + 597 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 533, + 530, + 597 + ], + "spans": [ + { + "bbox": [ + 299, + 533, + 530, + 597 + ], + "type": "text", + "content": "[23] D. Martin, C. Fowlkes, D. Tal, and J. Malik. A database of human segmented natural images and its application to evaluating segmentation algorithms and measuring ecological statistics. In Computer Vision, 2001. ICCV 2001. Proceedings. Eighth IEEE International Conference on, volume 2, pages 416-423, 2001. 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 299, + 598, + 530, + 640 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 598, + 530, + 640 + ], + "spans": [ + { + "bbox": [ + 299, + 598, + 530, + 640 + ], + "type": "text", + "content": "[24] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In ECCV, 2020. 1, 4, 8" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 299, + 642, + 530, + 696 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 642, + 530, + 696 + ], + "spans": [ + { + "bbox": [ + 299, + 642, + 530, + 696 + ], + "type": "text", + "content": "[25] et al. Radu Timofte. Ntire 2017 challenge on single image super-resolution: Methods and results. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition Workshops, CVPRW 2017, pages 1110-1121, 2017. 5, 6, 11" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 299, + 697, + 530, + 730 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 697, + 530, + 730 + ], + "spans": [ + { + "bbox": [ + 299, + 697, + 530, + 730 + ], + "type": "text", + "content": "[26] Nasim Rahman, Aristide Baratin, Devansh Arpit, Felix Draxler, Min Lin, Fred Hamprecht, Yoshua Bengio, and Aaron Courville. On the spectral bias of neural networks." + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 277, + 783, + 300, + 793 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 277, + 783, + 300, + 793 + ], + "spans": [ + { + "bbox": [ + 277, + 783, + 300, + 793 + ], + "type": "text", + "content": "10007" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 595, + 842 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 106, + 280, + 730 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 64, + 106, + 279, + 159 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 106, + 279, + 159 + ], + "spans": [ + { + "bbox": [ + 64, + 106, + 279, + 159 + ], + "type": "text", + "content": "In Kamalika Chaudhuri and Ruslan Salakhutdinov, editors, Proceedings of the 36th International Conference on Machine Learning, volume 97 of Proceedings of Machine Learning Research, pages 5301-5310. PMLR, 09-15 Jun 2019. 4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 160, + 279, + 191 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 160, + 279, + 191 + ], + "spans": [ + { + "bbox": [ + 47, + 160, + 279, + 191 + ], + "type": "text", + "content": "[27] Ali Rahimi and Benjamin Recht. Random features for large-scale kernel machines. Advances in neural information processing systems, 20, 2007. 8" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 193, + 280, + 257 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 193, + 280, + 257 + ], + "spans": [ + { + "bbox": [ + 47, + 193, + 280, + 257 + ], + "type": "text", + "content": "[28] David W. Romero, Robert-Jan Bruintjes, Jakub Mikelaj Tomczak, Erik J. Bekkers, Mark Hoogendoorn, and Jan van Gemert. Flexconv: Continuous kernel convolutions with differentiable kernel sizes. In The Tenth International Conference on Learning Representations, ICLR 2022, Virtual Event, April 25-29, 2022. OpenReview.net, 2022. 8" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 258, + 280, + 321 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 258, + 280, + 321 + ], + "spans": [ + { + "bbox": [ + 47, + 258, + 280, + 321 + ], + "type": "text", + "content": "[29] David W. Romero, Anna Kuzina, Erik J. Bekkers, Jakub Mikolaj Tomczak, and Mark Hoogendoorn. Ckconv: Continuous kernel convolution for sequential data. In The Tenth International Conference on Learning Representations, ICLR 2022, Virtual Event, April 25-29, 2022. OpenReview.net, 2022. 8" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 322, + 280, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 322, + 280, + 364 + ], + "spans": [ + { + "bbox": [ + 47, + 322, + 280, + 364 + ], + "type": "text", + "content": "[30] Chitwan Sahara, Jonathan Ho, William Chan, Tim Salimans, David J Fleet, and Mohammad Norouzi. Image superresolution via iterative refinement. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2022. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 365, + 280, + 449 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 365, + 280, + 449 + ], + "spans": [ + { + "bbox": [ + 47, + 365, + 280, + 449 + ], + "type": "text", + "content": "[31] Wenzhe Shi, Jose Caballero, Ferenc Huszar, Johannes Totz, Andrew P. Aitken, Rob Bishop, Daniel Rueckert, and Zehan Wang. Real-time single image and video super-resolution using an efficient sub-pixel convolutional neural network. In 2016 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2016, Las Vegas, NV, USA, June 27-30, 2016, pages 1874-1883. IEEE Computer Society, 2016. 1, 2, 4, 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 450, + 280, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 450, + 280, + 503 + ], + "spans": [ + { + "bbox": [ + 47, + 450, + 280, + 503 + ], + "type": "text", + "content": "[32] Matthew Tancik, Pratul P. Srinivasan, Ben Mildenhall, Sara Fridovich-Keil, Nithin Raghavan, Utkarsh Singhal, Ravi Ramamoorthi, Jonathan T. Barron, and Ren Ng. Fourier features let networks learn high frequency functions in low dimensional domains. NeurIPS, 2020. 1, 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 504, + 280, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 504, + 280, + 547 + ], + "spans": [ + { + "bbox": [ + 47, + 504, + 280, + 547 + ], + "type": "text", + "content": "[33] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 8" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 547, + 280, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 547, + 280, + 611 + ], + "spans": [ + { + "bbox": [ + 47, + 547, + 280, + 611 + ], + "type": "text", + "content": "[34] Xintao Wang, Ke Yu, Shixiang Wu, Jinjin Gu, Yihao Liu, Chao Dong, Yu Qiao, and Chen Change Loy. Esrgan: Enhanced super-resolution generative adversarial networks. In Laura Leal-Taixe and Stefan Roth, editors, Computer Vision - ECCV 2018 Workshops, pages 63-79, Cham, 2019. Springer International Publishing. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 612, + 280, + 653 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 612, + 280, + 653 + ], + "spans": [ + { + "bbox": [ + 47, + 612, + 280, + 653 + ], + "type": "text", + "content": "[35] Z. Wang, J. Chen, and S. H. Hoi. Deep learning for image super-resolution: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence, 43(10):3365-3387, oct 2021. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 655, + 280, + 708 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 655, + 280, + 708 + ], + "spans": [ + { + "bbox": [ + 47, + 655, + 280, + 708 + ], + "type": "text", + "content": "[36] Wenxuan Wu, Zhongang Qi, and Fuxin Li. Pointconv: Deep convolutional networks on 3d point clouds. In IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2019, Long Beach, CA, USA, June 16-20, 2019, pages 9621-9630. Computer Vision Foundation / IEEE, 2019. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 708, + 280, + 730 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 708, + 280, + 730 + ], + "spans": [ + { + "bbox": [ + 47, + 708, + 280, + 730 + ], + "type": "text", + "content": "[37] Yiheng Xie, Towaki Takikawa, Shunsuke Saito, Or Litany, Shiqin Yan, Numair Khan, Federico Tombari, James Tomp" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 299, + 106, + 531, + 313 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 318, + 106, + 530, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 106, + 530, + 128 + ], + "spans": [ + { + "bbox": [ + 318, + 106, + 530, + 128 + ], + "type": "text", + "content": "kin, Vincent Sitzmann, and Srinath Sridhar. Neural fields in visual computing and beyond, 2021. 1, 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 299, + 130, + 531, + 203 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 130, + 531, + 203 + ], + "spans": [ + { + "bbox": [ + 299, + 130, + 531, + 203 + ], + "type": "text", + "content": "[38] Roman Zeyde, Michael Elad, and Matan Protter. On single image scale-up using sparse-representations. In Jean-Daniel Boissonnat, Patrick Chenin, Albert Cohen, Christian Gout, Tom Lyche, Marie-Laurence Mazure, and Larry L. Schumaker, editors, Curves and Surfaces, volume 6920 of Lecture Notes in Computer Science, pages 711-730. Springer, 2010. 6" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 299, + 205, + 531, + 269 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 205, + 531, + 269 + ], + "spans": [ + { + "bbox": [ + 299, + 205, + 531, + 269 + ], + "type": "text", + "content": "[39] Yulun Zhang, Kunpeng Li, Kai Li, Lichen Wang, Bineng Zhong, and Yun Fu. Image super-resolution using very deep residual channel attention networks. In Vittorio Ferrari, Martial Hebert, Cristian Sminchisescu, and Yair Weiss, editors, Computer Vision - ECCV 2018, pages 294-310, Cham, 2018. Springer International Publishing. 1, 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 299, + 270, + 531, + 313 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 270, + 531, + 313 + ], + "spans": [ + { + "bbox": [ + 299, + 270, + 531, + 313 + ], + "type": "text", + "content": "[40] Yulun Zhang, Yapeng Tian, Yu Kong, Bineng Zhong, and Yun Fu. Residual dense network for image restoration. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 43(7):2480-2495, 2021. 1, 2, 3, 5, 6, 11, 12" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 277, + 783, + 300, + 793 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 277, + 783, + 300, + 793 + ], + "spans": [ + { + "bbox": [ + 277, + 783, + 300, + 793 + ], + "type": "text", + "content": "10008" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 595, + 842 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/CVT-SLR_ Contrastive Visual-Textual Transformation for Sign Language Recognition With Variational Alignment/012722bf-9201-40ff-b0cf-52287586b76f_content_list.json b/2023/CVT-SLR_ Contrastive Visual-Textual Transformation for Sign Language Recognition With Variational Alignment/012722bf-9201-40ff-b0cf-52287586b76f_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..d9d061ea506b9c06beaea5b678b3e76c2c8945f9 --- /dev/null +++ b/2023/CVT-SLR_ Contrastive Visual-Textual Transformation for Sign Language Recognition With Variational Alignment/012722bf-9201-40ff-b0cf-52287586b76f_content_list.json @@ -0,0 +1,1596 @@ +[ + { + "type": "text", + "text": "CVT-SLR: Contrastive Visual-Textual Transformation for Sign Language Recognition with Variational Alignment", + "text_level": 1, + "bbox": [ + 114, + 128, + 854, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Jiangbin Zheng $^{1}$ , Yile Wang $^{1,2}$ , Cheng Tan $^{1}$ , Siyuan Li $^{1}$ , Ge Wang $^{1}$ , Jun Xia $^{1}$ , Yidong Chen $^{3}$ , Stan Z. Li $^{1*}$", + "bbox": [ + 250, + 202, + 718, + 238 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ AI Lab, Research Center for Industries of the Future, Westlake University $^{2}$ Institute for AI Industry Research (AIR), Tsinghua University $^{3}$ School of Informatics, Xiamen University", + "bbox": [ + 191, + 239, + 779, + 292 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{zhengjiangbin, wangyile, tancheng, lisiyuan, wangge, xiajun, Stan.ZQ.Li}@westlake.edu.cn \nydchen@xmu.edu.cn", + "bbox": [ + 117, + 294, + 849, + 325 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 361, + 313, + 378 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Sign language recognition (SLR) is a weakly supervised task that annotates sign videos as textual glosses. Recent studies show that insufficient training caused by the lack of large-scale available sign datasets becomes the main bottleneck for SLR. Most SLR works thereby adopt pretrained visual modules and develop two mainstream solutions. The multi-stream architectures extend multi-cue visual features, yielding the current SOTA performances but requiring complex designs and might introduce potential noise. Alternatively, the advanced single-cue SLR frameworks using explicit cross-modal alignment between visual and textual modalities are simple and effective, potentially competitive with the multi-cue framework. In this work, we propose a novel contrastive visual-textual transformation for SLR, CVT-SLR, to fully explore the pretrained knowledge of both the visual and language modalities. Based on the single-cue cross-modal alignment framework, we propose a variational autoencoder (VAE) for pretrained contextual knowledge while introducing the complete pretrained language module. The VAE implicitly aligns visual and textual modalities while benefiting from pretrained contextual knowledge as the traditional contextual module. Meanwhile, a contrastive cross-modal alignment algorithm is designed to explicitly enhance the consistency constraints. Extensive experiments on public datasets (PHOENIX-2014 and PHOENIX-2014T) demonstrate that our proposed CVT-SLR consistently outperforms existing single-cue methods and even outperforms SOTA multi-cue methods. The source codes and models are available at https://github.com/binbinjiang/CVT-SLR", + "bbox": [ + 76, + 393, + 472, + 848 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 501, + 361, + 630, + 378 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "As a special visual natural language, sign language is the primary communication medium of the deaf community [19]. With the progress of deep learning [1, 17, 25, 39, 42], sign language recognition (SLR) has emerged as a multimodal task that aims to annotate sign videos into textual sign glosses. However, a significant dilemma of SLR is the lack of publicly available sign language datasets. For example, the most commonly-used PHOENIX-2014 [23] and PHOENIX-2014T [2] datasets only include about 10K pairs of sign videos and gloss annotations, which are far from training a robust SLR system with full supervision as typical vision-language cross-modal tasks [34]. Therefore, data limitation that may easily lead to insufficient training or overfitting problems is the main bottleneck of SLR tasks.", + "bbox": [ + 496, + 392, + 893, + 603 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The development of weakly supervised SLR has witnessed most of the improvement efforts focus on the visual module (e.g., CNN) [9, 10, 15, 29, 32, 33]. Transferring pretrained visual networks from general domains of human actions becomes a consensus to alleviate the low-resource limitation. The mainstream multi-stream SLR framework extends the pretrained visual module with multi-cue visual information [3, 22, 24, 43, 48, 50], including global features and regional features such as hands and faces in independent streams. The theoretical support for this approach comes from sign language linguistics, where sign language utilizes multiple complementary channels (e.g., hand shapes, facial expressions) to convey information [3]. The multi-cue mechanism essentially exploits hard attention to key information, yielding the current SOTA performances. However, the multi-cue framework is complex (e.g., cropping multiple regions, requiring more parameters), and the fusion of multiple streams might introduce additional potential noise.", + "bbox": [ + 496, + 608, + 895, + 881 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Another mainstream advanced solution is the single-cue", + "bbox": [ + 519, + 885, + 890, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author.", + "bbox": [ + 94, + 886, + 220, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "23141", + "bbox": [ + 478, + 924, + 517, + 936 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/1f9e04cf2819b24831d3b261aa24eea37fa88956f77e5bc41c72e100fbe9581f.jpg", + "image_caption": [ + "Figure 1. (a) An advanced single-cue SLR framework with explicit cross-modal alignment; (b) Our proposed single-cue SLR framework with explicit cross-modal alignment and implicit autoencoder alignment. Both frameworks use pretrained visual features. But our framework uses the autoencoder module to replace the mainstream contextual module, which not only includes the functions of the contextual module but also can introduce complete pretrained language knowledge and implicit cross-modal alignment. To maximize the preservation of the complete pretrained language parameters and migrated visual features, a video-gloss adapter is introduced." + ], + "image_footnote": [], + "bbox": [ + 116, + 93, + 851, + 313 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "cross-modal alignment framework [15, 28], which consists of a pretrained visual module followed by a contextual module (e.g., RNN, LSTM, Transformer) and a Connectionist Temporal Classification (CTC) [14] based alignment module for gloss generation, as shown in Figure 1 (a). Explicit cross-modal alignment constraints further improve feature interactions [15, 28, 38], which could be treated as a kind of consistency between two different modalities [50] facilitating the visual module learn long-term temporal information from contextual module [13, 37]. The cross-modal alignment framework is simple and effective, potentially competitive with the multi-cue framework. Despite the advanced performance of complex multi-cue architectures with pretrained visual modules, the cross-modal consistency is a more elegant design for practical usage. It also implies the potential of prior contextual linguistic knowledge, which has been overlooked by existing SLR works.", + "bbox": [ + 75, + 414, + 472, + 672 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this work, we propose a novel contrastive visual-textual transformation framework for SLR, called CVT-SLR, to fully explore the pretrained knowledge of both the visual and language modalities, as shown in Figure 1 (b). Based on the single-cue cross-modal alignment framework, CVT-SLR keeps the pretrained visual module but replaces the traditional contextual module with a variational autoencoder (VAE). Since a full encoder-decoder architecture is used, the VAE is responsible for learning pretrained contextual knowledge based on a pseudo-translation task while introducing the complete pretrained language module. In addition, the VAE maintains the consistency of input and output modalities due to the form of an autoencoder, playing an implicit cross-modal alignment role. Furthermore, inspired by contrastive learning [4-6,34], we introduce a contrastive", + "bbox": [ + 75, + 674, + 472, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "alignment algorithm that focuses on both positive and negative samples to enhance explicit cross-modal consistency constraints. Extensive quantitative experiments conducted on the public datasets PHOENIX-2014 and PHOENIX-2014T demonstrate the advance of the proposed CVT-SLR framework. Through ablation study and qualitative analysis, we also verify the effectiveness of introducing pretrained language knowledge and the new consistency constraint mechanism.", + "bbox": [ + 496, + 414, + 893, + 549 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our main contributions are as follows:", + "bbox": [ + 517, + 551, + 774, + 565 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- A novel visual-textual transformation-based SLR framework is proposed, which introduces fully pretrained language knowledge for the first time and provides new approaches for other cross-modal tasks.", + "- New alignment methods are proposed for cross-modal consistency constraints: a) exploiting the special properties of the autoencoder to implicitly align visual and textual modalities; b) introducing an explicit contrastive cross-modal alignment method.", + "- The proposed single-cue SLR framework not only outperforms existing single-cue baselines by a large margin but even surpasses SOTA multi-cue baselines." + ], + "bbox": [ + 517, + 568, + 890, + 757 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 768, + 640, + 784 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "SLR tasks are generally divided into three categories [19]: finger-spelling recognition [18, 30, 31], isolated word recognition [26, 40, 41, 49] and continuous sign sentence recognition [8, 11, 12, 29, 33]. In the early days, SLR works mainly focused on lexical-level tasks, such as fingerspelling recognition and isolated word recognition. Nowadays, the more practical continuous sign sentence recogni", + "bbox": [ + 496, + 794, + 893, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "23142", + "bbox": [ + 478, + 924, + 519, + 936 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "tion task has become mainstream sign language research. In this work, the mentioned SLRs refer specifically to continuous sign sentence recognition.", + "bbox": [ + 75, + 90, + 468, + 135 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The recent SLR works [8, 11, 15, 29] are summarized based on these three aspects: feature extraction, recognition, and alignment. Mostly, the components of feature extraction are composed of the visual module and the contextual module. The visual module encodes short-term spatial information, while the contextual module encodes long-term context information. Based on the extracted features, the classifiers can get a posterior probability for recognition. And the alignment module is required to find the proper alignment between clips and glosses to ensure an accurate training procedure.", + "bbox": [ + 75, + 136, + 468, + 301 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "A few studies extend the common SLR framework by incorporating multi-cue information [22,24,43,44,48]. The multi-cue features mainly include information from hand shapes, facial expressions, mouths, and poses. The current SOTA SLR works are based on multi-cue mechanisms, such as $\\mathrm{C}^2\\mathrm{SLR}$ [50]. Recently, some works show that explicitly enforcing the consistency between the visual and textual modules can also get comparable performance, although only single-cue features are considered. For example, VAC [28] treats the visual and textual modules as the student and teacher, respectively, and achieves knowledge distillation to align the visual and the textual modalities. Similarly, SMKD [15] achieves knowledge transfer by further sharing classifiers between different modalities. Note that SMKD is the current SOTA single-cue method.", + "bbox": [ + 75, + 301, + 468, + 527 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "It is worth mentioning that neural sign language translation [2,3,7,46] is another popular sign language task. As a distinction, the neural sign language translation task aims to convert sign videos into equivalent spoken language translations, further focusing on semantic relationships and adjusting word order [2], while the SLR task only annotates sign videos into gloss sequences in the frame order.", + "bbox": [ + 75, + 529, + 468, + 635 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Methods", + "text_level": 1, + "bbox": [ + 76, + 647, + 174, + 662 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "As shown in Figure 2, our training pipeline is roughly divided into two steps. Step 1 is to pretrain a VAE network as the textual module. And step 2 is to transfer the off-the-shelf visual module (usually a publicly available CNN trained on Kinetics/ImageNet) and the pretrained textual module from step 1 into the CVT-SLR framework. A video-gloss adapter is used to bridge these two pretrained modules.", + "bbox": [ + 75, + 672, + 468, + 777 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Textual Model Pretraining", + "text_level": 1, + "bbox": [ + 76, + 786, + 316, + 801 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our VAE model is an asymmetric encoder-decoder architecture, where the encoder consists of self-attention layers, and the decoder consists of Bi-LSTM layers, as shown in the top part of Figure 2. To pretrain the VAE for our proposed SLR framework solely based on the gloss data from the PHOENIX-14 training set (excluding the development", + "bbox": [ + 75, + 809, + 468, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "set and testing set), we construct a pseudo-translation task, Gloss2Gloss, which enables the VAE to learn the contextual semantic knowledge. Formally, it is assumed that the input gloss of the encoder is denoted as $G$ , and the generative gloss of the decoder is denoted as $G^{*}$ .", + "bbox": [ + 496, + 90, + 890, + 166 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Posterior Self-attention Encoder. We employ the neural approximation, and reparameterization [21, 35] techniques to approximate the posterior self-attention encoder for generating the posterior distribution $q_{\\phi}(z|G)$ , where $\\phi$ denotes the variational parameters of the encoder. The $q_{\\phi}(z|G)$ is a diagonal Gaussian distribution $N(\\mu, \\mathrm{diag}(\\sigma^2))$ , and two single-layer MLPs are used to parameterize the mean $\\mu$ and the variance $\\sigma^2$ , respectively, as:", + "bbox": [ + 496, + 167, + 890, + 287 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nz = \\mu + \\sigma \\odot \\epsilon , \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 642, + 296, + 890, + 311 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\epsilon$ is a standard Gaussian noise, and $\\odot$ denotes an element-wise product.", + "bbox": [ + 496, + 319, + 890, + 348 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Generative Bi-LSTM Decoder. Given the continuous latent variable $z$ , then the translation target $G^{*}$ is generated from $z$ as:", + "bbox": [ + 496, + 349, + 890, + 392 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\np _ {\\theta} \\left(G ^ {*}, z\\right) = p _ {\\theta} \\left(G ^ {*} \\mid z\\right) p _ {\\theta} (z), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 596, + 393, + 890, + 411 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\theta$ denotes the parameters of the generative Bi-LSTM decoder, and $p_{\\theta}(G^{*}|z)$ is the conditional distribution that models the generation procedure estimated via the decoder.", + "bbox": [ + 496, + 414, + 890, + 460 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Variational Optimization. Both the generative model $p_{\\theta}(G^{*},z)$ and the posterior inference model $q_{\\phi}(z|G)$ are incorporated tightly in an end-to-end manner, and we can apply a standard back-propagation to optimize the gradient of the variational lower bound:", + "bbox": [ + 496, + 460, + 890, + 534 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\begin{array}{l} \\mathcal {L} _ {\\mathrm {V A E}} (\\theta , \\phi ; G) = - \\mathrm {K L} (q _ {\\phi} (z | G) | | p _ {\\theta} (z)) \\\\ \\vdots \\end{array} \\tag {3} \\\\ + \\mathbb {E} _ {q _ {\\phi} (z | G)} \\left[ \\log p _ {\\theta} \\left(G ^ {*} \\mid z\\right) \\right] \\leq \\log p _ {\\theta} \\left(G ^ {*}\\right), \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 562, + 541, + 890, + 575 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where KL denotes the Kullback-Leibler divergence and the prior $p_{\\theta}(z)$ denotes a standard Gaussian distribution. Practically, the first KL term of Eq. (3) can be written as:", + "bbox": [ + 496, + 582, + 890, + 628 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {K L}} = - \\frac {1}{2} \\left(\\log \\sigma^ {2} - \\sigma^ {2} - \\mu^ {2} + 1\\right). \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 576, + 635, + 890, + 664 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Since our main target is a Gloss2Gloss translation task, we employ a cross-entropy loss to maximize the generative gloss probability as the second term of Eq. (3):", + "bbox": [ + 496, + 669, + 890, + 715 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {g l o s s 2 g l o s s}} = \\text {C r o s s E n t r o p y} (G, G ^ {*}). \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 576, + 723, + 890, + 739 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Therefore, the overall optimization objective of textual model pretraining is a joint loss of $\\mathcal{L}_{\\mathrm{KL}}$ and $\\mathcal{L}_{\\mathrm{gloss2gloss}}$ as:", + "bbox": [ + 496, + 746, + 890, + 777 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {V A E}} = \\mathcal {L} _ {\\mathrm {K L}} + \\mathcal {L} _ {\\text {g l o s s 2 g l o s s}}. \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 606, + 785, + 890, + 801 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Training CVT-SLR Framework", + "text_level": 1, + "bbox": [ + 500, + 808, + 779, + 824 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2.1 Training Pipeline", + "text_level": 1, + "bbox": [ + 500, + 830, + 674, + 847 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In the CVT-SLR framework, we sequentially concatenate the pretrained visual CNN module and the pretrained textual VAE module through a video-gloss adapter, as shown", + "bbox": [ + 496, + 854, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "23143", + "bbox": [ + 478, + 925, + 517, + 936 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/c02c1185002b9db6f4cfe6b1db04e336f1ed3f17399395a2f054100091fb0252.jpg", + "image_caption": [ + "Figure 2. Overview of proposed CVT-SLR framework. Our training pipeline is divided into two steps. Step 1: Pretrain the VAE module with the self-supervised gloss-to-gloss pseudo-translation task. Step 2: Train the CVT-SLR based on the VAE model from Step 1." + ], + "image_footnote": [], + "bbox": [ + 122, + 89, + 851, + 333 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "in Figure 2. The pretrained ResNet18 model [17] is selected as the frame-wise feature extractor, which is trained on large-scale human action datasets. For the VAE module, although Gaussian noise is introduced in the reparameterization process in step 1, we do not add any noise after migrating the VAE to the SLR framework in step 2. In this way, the reparameterization module of the VAE degenerates into two common linear layers, i.e., $\\epsilon$ is removed in Eq. (1).", + "bbox": [ + 75, + 387, + 468, + 507 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Formally, given a sign video with $T$ RGB frames $\\mathcal{F} = \\{\\mathcal{F}_t\\}_{t=1}^T \\in \\mathbb{R}^{T \\times H \\times W \\times 3}$ , the visual module, which is composed of a stack of 2D-CNN layers and a global average pooling layer, first extracts spatial features $v = \\{v_t\\}_{t=1}^T \\in \\mathbb{R}^{T \\times d}$ . Then the textual module further extracts temporal features $s = \\{s_t\\}_{t=1}^T \\in \\mathbb{R}^{T \\times d}$ . Finally, the alignment module utilizes CTC to compute the conditional probability of the gloss label sequence $p(\\mathcal{G}|\\mathcal{F})$ based on temporal features $s$ , where $\\mathcal{G} = \\{\\mathcal{G}_i\\}_{i=1}^N$ is a generated sign gloss sequence with $N$ glosses. A shared classifier is used between the visual module and the textual module to generate spatial and temporal probability, respectively. And the proposed contrastive cross-modal alignment algorithm further enhances the consistency across different modalities.", + "bbox": [ + 76, + 508, + 470, + 719 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2.2 Video-Gloss Adapter", + "text_level": 1, + "bbox": [ + 76, + 739, + 277, + 755 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To maximize the preservation of both pretrained visual and textual knowledge, we use a video-gloss adapter to connect spatial-temporal features, as shown in Figure 2. The videogloss adapter is simply implemented as a fully-connected MLP with two hidden layers. We take the CNN output probabilities as the input of the adapter where the final fully connected layer is initialized using the weights of the pretrained word embedding from the pretrained VAE [7].", + "bbox": [ + 75, + 763, + 468, + 883 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Since both the input and output are in the same textual", + "bbox": [ + 96, + 885, + 468, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "form during the textual model pretraining in step 1, the VAE plays an implicit cross-modal alignment role after being transferred to the CVT-SLR framework in step 2, which is close to the effect of consistency. Therefore, the videogloss adapter is essentially a mediator for activating the consistency between the spatial and temporal modalities.", + "bbox": [ + 496, + 387, + 890, + 479 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2.3 CTC Alignment", + "text_level": 1, + "bbox": [ + 500, + 494, + 666, + 511 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Due to the spatial-temporal nature of sign videos, glosses have a one-to-many mapping to video frames but share the same ordering. Unlike spoken language texts, the continuous gloss annotations are chronologically consistent with the sign videos. Concretely, for an input sign video $\\mathcal{F}$ and the corresponding generated gloss sequence $\\mathcal{G}$ , we use CTC to compute $p(\\mathcal{G}|\\mathcal{F})$ by marginalizing over all possible $\\mathcal{F}$ to $\\mathcal{G}$ alignments:", + "bbox": [ + 496, + 518, + 890, + 640 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\np (\\mathcal {G} | \\mathcal {F}) = \\sum_ {\\pi \\in \\mathcal {B}} p (\\pi | \\mathcal {F}), \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 614, + 648, + 890, + 680 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\pi$ denotes a path and $\\mathcal{B}$ is the set of all viable paths corresponding to $\\mathcal{G}$ . The probability $p(\\pi|\\mathcal{F})$ is computed by the textual module. The CTC loss is then formulated as:", + "bbox": [ + 496, + 688, + 890, + 733 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {c t c}} = - \\log p (\\mathcal {G} ^ {*} | \\mathcal {F}), \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 619, + 742, + 890, + 760 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathcal{G}^*$ is the ground truth gloss sequence.", + "bbox": [ + 500, + 768, + 795, + 784 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2.4 Contrastive Alignment", + "text_level": 1, + "bbox": [ + 500, + 801, + 712, + 816 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Aligning both visual and textual modalities help improve SLR [28]. To enforce alignment constraints across modalities, we propose a contrastive alignment loss. In addition, shared classifiers are also used as an improvement technique between visual and textual features [15].", + "bbox": [ + 496, + 824, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "23144", + "bbox": [ + 478, + 924, + 519, + 936 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/53ea506c1a71204fae10bf2082c2d115a4d0aebe97563b77ddb4040551ef849a.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
GroupsModelsDev (%)Test (%)Cues
DEL/INSWERDEL/INSWER
Group 1SubUNet [9]14.6/4.040.814.3/4.040.7video
Staged-Opt [9]13.7/7.339.412.2/7.538.7video
Align-iOpt [33]12.6/2.637.113.0/2.536.7video
DPD+TEM [47]9.5/3.235.69.3/3.134.5video
Re-Sign [24]-27.1-26.8video
SFL [29]7.9/6.526.27.5/6.326.8video
DNF [11]7.8/3.523.87.8/3.424.4video
FCN [8]-23.7-23.9video
VAC [28]7.9/2.521.28.4/2.622.3video
CMA [32]7.3/2.721.37.3/2.421.9video
SFL [29]10.3/4.124.910.4/3.625.3video
VL-SLT [7]-21.9-22.5video
SMKD [15]6.8/2.520.86.3/2.321.0video
Group 2DNF [11]7.3/3.323.16.7/3.322.9video+optical flow
STMC [48]7.7/3.421.17.4/2.620.7video+hand+face+pose
C2SLR [50]-20.5-20.4video+keypoints
Group 3Ours1 (w/o VAE+Contra)7.1/3.021.17.3/2.921.4video
Ours2 (w/ VAE)6.5/2.420.26.3/2.220.3video
Ours3 (w/ Contra)6.7/2.720.46.4/2.520.7video
Ours4 (w/ VAE+Contra)6.4/2.619.86.1/2.320.1video
", + "bbox": [ + 130, + 88, + 836, + 449 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Table 1. Performance comparison (%) on PHOENIX-14 dataset. DEL/INS: deletion error and insertion error. The best results and SOTA baseline for each group are marked as bold and underlined.", + "bbox": [ + 75, + 455, + 890, + 484 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In previous SLR works, cross-modal alignment only focuses on positive samples [15, 48]. Inspired by contrastive learning [5, 16, 45], we construct both positive and negative samples in the same mini-batch and implement a contrastive cross-modal alignment method to ensure that similar features are closer while different features are farther apart. Given that the normalized spatial features from CNN as $S_{\\mathrm{logits}} \\in \\mathbb{R}^{B \\times T \\times d}$ , and the normalized temporal features from VAE as $\\mathcal{V}_{\\mathrm{logits}} \\in \\mathbb{R}^{B \\times T \\times d}$ , where $B$ denotes the number of samples. Then we compute pair matrices firstly as:", + "bbox": [ + 75, + 507, + 468, + 659 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {S} 2 \\mathcal {V} _ {\\text {p a i r}} = \\mathcal {S} _ {\\text {l o g i t s}} \\times \\mathcal {V} _ {\\text {l o g i t s}} ^ {T}, \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 184, + 666, + 468, + 686 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {V} 2 \\mathcal {S} _ {\\text {p a i r}} = \\mathcal {V} _ {\\log_ {1 5}} \\times \\mathcal {S} _ {\\log_ {1 5}} ^ {T}, \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 187, + 694, + 468, + 714 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Taking $S2\\mathcal{V}_{\\mathrm{pair}} \\in \\mathbb{R}^{B \\times B}$ pair matrix (denoted as $\\mathcal{P}$ ) as an example, $\\mathcal{P}[i,j](0 \\leq i,j < B)$ represents the similarity value between the visual feature of the $i$ -th batch and the textual feature of $j$ -th batch. The visual and textual features from the same input instances are positive samples, and the features from other different input instances are negative samples. Hence, in the matrix $\\mathcal{P}$ , the diagonal similarity values are from positive sample pairs, and the rest values are from negative sample pairs. To make the loss computation differentiable and easy, we convert the computation of positive and negative pairs into a binary classification task [34]. The value $i$ corresponding to the $i$ -th row in the matrix $\\mathcal{P}$", + "bbox": [ + 75, + 719, + 468, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "is a positive label. Therefore, the corresponding labels of $\\mathcal{P}$ are Labels $= \\{0,1,\\dots ,i,\\dots ,B\\}$ , then the contrastive alignment loss related to $S2\\mathcal{V}_{\\mathrm{pair}}$ as:", + "bbox": [ + 498, + 507, + 890, + 555 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {a l i g n} _ {S}} = \\text {C r o s s E n t r o p y} (\\text {S o f t m a x} (S 2 \\mathcal {V} \\text {p a i r}), \\text {L a b e l s}). \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 501, + 561, + 890, + 580 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Similarly, the $\\mathcal{V}2S_{\\mathrm{pair}}$ temporal-to-spatial alignment loss as:", + "bbox": [ + 500, + 588, + 890, + 604 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {a l i g n} _ {\\nu}} = \\text {C r o s s E n t r o p y} (\\text {S o f t m a x} (\\mathcal {V} 2 S _ {\\text {p a i r}}), \\text {L a b e l s}). \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 509, + 613, + 890, + 631 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Therefore, the complete contrastive alignment loss as:", + "bbox": [ + 500, + 638, + 854, + 654 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {a l i g n}} = \\frac {1}{2} \\left(\\mathcal {L} _ {\\text {a l i g n} _ {S}} + \\mathcal {L} _ {\\text {a l i g n} _ {V}}\\right). \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 596, + 662, + 890, + 691 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Overall, the objective of CVT-SLR is jointly optimized CTC loss $\\mathcal{L}_{\\mathrm{ctc}}$ and contrastive alignment loss $\\mathcal{L}_{\\mathrm{align}}$ as:", + "bbox": [ + 498, + 698, + 890, + 729 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {C V T - S L R}} = \\mathcal {L} _ {\\mathrm {c t c}} + \\mathcal {L} _ {\\mathrm {a l i g n}}. \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 609, + 739, + 890, + 755 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 500, + 777, + 630, + 794 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Settings", + "text_level": 1, + "bbox": [ + 500, + 801, + 596, + 816 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Datasets. We use PHOENIX-2014 [23] and PHOENIX-2014T [2] as datasets, where RGB videos and their corresponding annotations are provided. PHOENIX-2014 is a German SLR dataset about weather forecasts with a vocabulary size of 1,081, which is divided into three parts: 5,672", + "bbox": [ + 498, + 824, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "23145", + "bbox": [ + 478, + 925, + 517, + 936 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/7894c2879a29d10b70aea43897d2bb45125eadba8b0e94f349e987db46766525.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
GroupsModelsWERCues
Dev(%)Test(%)
Group 1SFL [29]25.126.1video
CNN+LSTM+HMM [22]24.526.5video
SLT [3]24.924.6video
FCN [8]23.325.1video
SMKD [15]20.822.4video
Group 2CNN+LSTM+HMM [22]24.525.4video+mouth
CNN+LSTM+HMM [22]22.124.1video+mouth+hand
SLT [3]24.624.5video+text
STMC [48]19.621.0video+hand+face+pose
C²SLR [50]20.220.4video+keypoints
Group 3Ours1 (w/o VAE+Contra)21.822.0video
Ours2 (w/ VAE)20.120.4video
Ours3 (w/ Contra)21.021.5video
Ours4 (w/ VAE+Contra)19.420.3video
", + "bbox": [ + 204, + 83, + 763, + 347 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 2. Performance comparison (%) on PHOENIX14T dataset. The best results and SOTA baseline for each group are marked as **bold** and **underlined**, respectively.", + "bbox": [ + 75, + 349, + 893, + 380 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "instances for training, 540 for validation, and 629 for testing, respectively. And PHOENIX-2014T is an extension of PHOENIX-2014 with a vocabulary size of 1,085, which is also divided into three parts: 7,096 instances for training, 519 for validation, and 642 for testing, respectively.", + "bbox": [ + 75, + 396, + 470, + 470 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation Metric. Word error rate (WER) is the most widely-used metric to evaluate the SLR performance, which measures the number of necessary insertions, substitutions, and deletions in the recognized sentences when compared to the reference sentences. WER is an edit distance, indicating the least number of operations of substitutions (#sub), insertions (#ins), and deletions (#del) to transform the predicted sentences into the reference sequences (#reference):", + "bbox": [ + 75, + 472, + 468, + 592 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {W E R} = \\frac {\\# \\mathrm {d e l} + \\# \\mathrm {i n s} + \\# \\mathrm {s u b}}{\\# \\text {r e f e r e n c e}}. \\tag {15}\n$$\n", + "text_format": "latex", + "bbox": [ + 165, + 597, + 468, + 627 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation Details. The videos are resized to $256 \\times 256$ and then cropped to $224 \\times 224$ . During SLR training, we use random crop and horizontal flips (50%) for data augmentation. During SLR testing, we only adopt center cropping. Beam Search is used for gloss generation. Our visual module has the same configuration as [15]. The VAE module contains two multi-head self-attention layers with 4 heads and two Bi-LSTM layers with $2 \\times 512$ dimensional hidden states sequentially. Adam [20] optimizer is used to pretrain our VAE network with a batch size of 16, an initial learning rate of 1e-4. And AdamW [27] optimizer is used to train our SLR framework with a batch size of 8, a learning rate of 1e-4, and a weight decay of 1e-4. And 1 NVIDIA A100 80GB GPU is used.", + "bbox": [ + 75, + 628, + 470, + 839 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Main Results", + "text_level": 1, + "bbox": [ + 76, + 847, + 217, + 861 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1 and Table 2 show our comparisons with two-group different baselines, where Group 1 includes", + "bbox": [ + 75, + 869, + 470, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "single-cue baselines and Group 2 includes multi-cue baselines. Oursnumber denote our trained CVT-SLR framework, which have only two variables, \"VAE\" and \"Contra\". Note that \"w/o (w/)\" VAE\" denotes the VAE module in CVT-SLR framework without (with) transferred pretrained parameters, and \"w/o (w/)\" Contra\" denotes training CVT-SLR framework without (with) contrastive alignment loss.", + "bbox": [ + 496, + 396, + 890, + 501 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation on PHOENIX-14. Table 1 shows our methods compared to other methods on PHOENIX-14. The optimalWER scores of our single-cue CVT-SLR $(\\mathrm{Ours}_4)$ on the development set and testing set are $19.8\\%$ and $20.1\\%$ respectively. And $\\mathrm{Ours_4}$ achieves the best performance among video-based only/single-cue methods in Group 1 (Ours vs. SMKD). Furthermore, while no extra cues are used, our CVT-SLR achieves the best result among the models trained with extra cues (i.e., multi-cue methods) in Group 2 (Ours vs. $\\mathrm{C}^2\\mathrm{SLR}$ ).", + "bbox": [ + 496, + 503, + 892, + 654 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "From within-group comparisons in Group 3, we find that the case, $\\mathrm{Ours}_1$ , which neither migrates the pretrained VAE parameters nor uses the contrastive alignment constraint, performs the worst. Nevertheless, relative to other groups, it still performs well. Based on $\\mathrm{Ours}_1$ , by only configuring the pretrained VAE parameters $(\\mathrm{Ours}_2)$ or the contrastive alignment constraints $(\\mathrm{Ours}_3)$ , the CVT-SLR framework can also gain gains, and the pretrained VAE brings more obvious gains $(+0.9\\% / +1.1\\%$ vs. $+0.7\\% / +0.7\\%$ ). Overall, Group 3 demonstrates that both the prior language knowledge from the pretrained textual module and the contrastive constraints through cross-modal alignment play important enhancement roles for our SLR framework.", + "bbox": [ + 496, + 656, + 893, + 851 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation on PHOENIX-14T. In Table 2, we evaluate the CVT-SLR on PHOENIX-14T. We can see that our optimal model $(\\mathrm{Ours}_4)$ in this dataset also achieves the best per", + "bbox": [ + 496, + 854, + 893, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "23146", + "bbox": [ + 478, + 925, + 519, + 936 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/6f57670ce6478fdea0424f69a8d7b7432835ae4f2a0c08da4452fc8cb0e0f6d6.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
#Pre-VisualPre-TextualWER
Dev(%)Test(%)
1××89.288.5
2×✓30.831.4
3✓×20.420.7
4✓✓19.820.1
", + "bbox": [ + 93, + 85, + 450, + 193 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/66ef0c1f1f733154a12f660b3537c8a8dd5f079daf0693f882196a3429e6aa71.jpg", + "table_caption": [ + "Table 3. Ablation study of pretrained modules on PHOENIX-14 (Pre-Visual: pretrained visual module, Pre-Language: pretrained textual module). The best results are bolded." + ], + "table_footnote": [], + "table_body": "
#CTC WeightAlign WeightWER
Dev(%)Test(%)
11.0020.220.3
21.01.020.020.2
31.010.019.820.1
41.0100.020.120.6
51.01000.020.520.9
", + "bbox": [ + 91, + 251, + 450, + 367 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 4. Ablation study of varied weights of the CTC loss and the alignment loss on PHOENIX-14. The best results are bolded.", + "bbox": [ + 76, + 375, + 467, + 402 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "formance (development set: $19.4\\%$ and testing set: $20.3\\%$ ) with video-based only/single-cue information compared to Group 1 $(\\text{Ours}_4$ vs. SMKD). It is worth mentioning that the SOTA single-cue method (SMKD) fails to surpass the multi-cue baselines in their work on PHOENIX-14T. And our CVT-SLR makes up for their shortcomings while outperforming the SOTA multi-cue baselines in Group 2 on this dataset $(\\text{Ours}_4$ vs. $\\mathbf{C}^2\\mathbf{SLR})$ .", + "bbox": [ + 75, + 417, + 467, + 537 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Compared within Group 3 on the PHOENIX-14T, the observed conclusions are more or less the same as on the PHOENIX-14, i.e., both the pretrained textual module and the proposed contrastive alignment constraint have a beneficial effect on our CVT-SLR framework. But the pretrained textual module gain on PHOENIX-14T is more significant than that on PHOENIX-14 $(+1.7\\% / + 1.6\\%$ vs. $+0.9\\% / + 1.1\\%)$ , indicating that prior language knowledge is more beneficial for this dataset.", + "bbox": [ + 75, + 539, + 467, + 674 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Ablation Study", + "text_level": 1, + "bbox": [ + 76, + 684, + 228, + 700 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We study the effectiveness of the proposed pretrained modules and alignment loss weights in our framework.", + "bbox": [ + 76, + 707, + 467, + 738 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3.1 Pretrained Modules", + "text_level": 1, + "bbox": [ + 76, + 756, + 269, + 768 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To illustrate the importance of the pretrained modules in our CVT-SLR framework, we conduct the ablation studies of both the pretrained visual module (CNN module) and the pretrained textual module (VAE module) on PHOENIX-14, as verified in Table 3.", + "bbox": [ + 75, + 779, + 467, + 853 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Overall, both pretrained modules are useful. Comparing #1 and #3, we find that pretrained visual features account for a larger determinant of performance, which is consistent", + "bbox": [ + 75, + 854, + 467, + 900 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e44ef830258e9dfb75d6ada8135a918bcd7474f7e6d6593933ae1b12d988b997.jpg", + "image_caption": [ + "Figure 3. The training loss curves corresponding to the ablation study of pretrained modules in Table 3." + ], + "image_footnote": [], + "bbox": [ + 509, + 84, + 877, + 296 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "with our reason for retaining the pretrained visual knowledge. Surprisingly, comparing #1 and #2, we find that without pretrained parameters of the visual module, i.e., using only the pretrained textual module, our CVT-SLR performance is also greatly improved. Notably, this is a new observation that has not been explored in previous SLR works extensively. These comparisons demonstrate that the pretrained VAE module has learned sufficient language knowledge to improve the SLR task. Comparing #3 with #4 further confirms this point.", + "bbox": [ + 496, + 344, + 890, + 494 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To verify the effectiveness of the pretrained modules for the training convergence speed, training loss curves corresponding to Table 3 are shown in Figure 3. Specifically, comparing #1 and #2 (or #3 and #4), configuring a pretrained textual module can promote faster convergence and reach a stable state earlier. Similarly, comparing #1 and #3 (or #2 and #4), a pretrained visual module can accelerate convergence speed compared to the pretrained textual module. As can be seen from the loss curves, the degrees of convergence are positively correlated with the final evaluation performances of WER scores.", + "bbox": [ + 496, + 496, + 890, + 661 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3.2 Alignment Loss Weights", + "text_level": 1, + "bbox": [ + 500, + 680, + 723, + 696 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Training CVT-SLR to recognize glosses is optimized jointly by CTC loss and contrastive alignment loss. To explore the effect of loss weights, we conduct an ablation study of varied weights of the CTC loss and the proposed contrastive alignment loss, as shown in Table 4. For simplicity, we set the weights of CTC loss to a fixed value of 1.0. In Table 4, #1, without contrastive alignment loss (weight=0), performs worse than the models with contrastive alignment in an optimal configuration (#2, #3), which also illustrates the gain effect of the proposed contrastive alignment. As shown in #2-#5 in Table 4, we keep increasing the weight of the alignment loss from 1.0 to 1000.0. As the weight of the contrastive loss increases, the performance starts to rise, reach", + "bbox": [ + 496, + 703, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "23147", + "bbox": [ + 478, + 925, + 517, + 936 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/7a58421c59f84ac7013abff7a0f97ba68c1d80800e8eb8bbaeb29a6a51158b59.jpg", + "image_caption": [ + "Figure 4. Four examples with cross-modal alignment matrices (left), saliency maps (middle), and generated glosses (right) on the PHOENIX-14 test set. Cross-modal alignment matrices show the alignment between visual and textual features, while saliency maps show the highest activation regions focused on gestures." + ], + "image_footnote": [], + "bbox": [ + 84, + 84, + 883, + 309 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "ing a peak when the weight value is 10.0. As the weight increases and the value is greater than 10.0, the performance starts to decline instead. This ablation study indicates the important role of the proposed contrastive alignment. By balancing the weights between CTC loss and contrastive alignment loss, the model can achieve optimal performance.", + "bbox": [ + 75, + 378, + 472, + 470 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4. Qualitative Analysis", + "text_level": 1, + "bbox": [ + 76, + 478, + 269, + 494 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We randomly select four examples from the test set of PHOENIX-14 for qualitative analysis. Each example displays an alignment matrix, a series of saliency maps, and generated glosses in Figure 4. An alignment matrix shows alignment relationships between visual and textual features, while a saliency map shows the highest activations.", + "bbox": [ + 75, + 501, + 468, + 590 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Visualization on Cross-modal Alignment Matrices. From these four alignment matrices in Figure 4, we can see that the reference and inferred glosses are almost the same with quite low WER scores, and the performances are relatively perfect. Hence, the highlighted areas of the alignment matrices are concentrated near the diagonal, which means that our contrastive alignment algorithm plays a crucial role in aligning visual and textual features.", + "bbox": [ + 75, + 592, + 468, + 712 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Visualization on Saliency Maps. We visualize the key parts of the sign video frames in focus by using GradCAM [36] as shown in Figure 4. The results show that, for well-performing inferences, significant spatial features are mainly focused on the face and hand regions in the saliency maps. This observation is consistent with the sign linguistic knowledge [3, 19], where sign language conveys information mainly relying on hand shapes and facial expressions.", + "bbox": [ + 75, + 712, + 468, + 834 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 76, + 844, + 194, + 859 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this work, we propose a novel contrastive visual-textual transformation framework for SLR, called CVT-", + "bbox": [ + 76, + 869, + 468, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "SLR, which further alleviates insufficient training by fully exploring the pretrained knowledge of both the visual and language modalities. We introduce prior language knowledge into the single-cue CVT-SLR for the first time via a VAE-based pretrained textual module. Furthermore, we propose new methods for cross-modal consistency constraints. One method takes advantage of the properties of the autoencoder to implicitly align the visual and textual modalities. While another method introduces a contrastive cross-modal alignment for explicit consistency. To prove the proposed CVT-SLR framework, extensive quantitative experiments and qualitative analysis are conducted, showing that our single-cue CVT-SLR not only outperforms single-cue baselines by a large margin but also surpasses SOTA multi-cue methods.", + "bbox": [ + 496, + 378, + 890, + 606 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Overall, we hope our CVT-SLR can inspire the SLR community to mine pretrained modules from new aspects and design a more efficient SLR framework for practical usage. In future work, we plan to introduce a large-scale pretrained language model to guide textual module learning and further explore the potential of prior contextual linguistic knowledge for SLR tasks.", + "bbox": [ + 496, + 612, + 890, + 717 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgement", + "text_level": 1, + "bbox": [ + 500, + 750, + 660, + 766 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We thank the anonymous reviewers for their constructive and helpful reviews. This work was supported by the National Key R&D Program of China (Project No. 2022ZD0115100), the National Natural Science Foundation of China (Project No. U21A20427), the Competitive Research Fund (Project No. WU2022A009) from the Westlake Center for Synthetic Biology and Integrated Bioengineering.", + "bbox": [ + 496, + 780, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "23148", + "bbox": [ + 478, + 925, + 519, + 936 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Hangbo Bao, Li Dong, and Furu Wei. Beit: Bert pre-training of image transformers. In ICLR, 2022. 1", + "[2] Necati Cihan Camgoz, Simon Hadfield, Oscar Koller, Hermann Ney, and Richard Bowden. Neural sign language translation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7784-7793, 2018. 1, 3, 5", + "[3] Necati Cihan Camgoz, Oscar Koller, Simon Hadfield, and Richard Bowden. Sign language transformers: Joint end-to-end sign language recognition and translation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10023-10033, 2020. 1, 3, 6, 8", + "[4] Mathilde Caron, Ishan Misra, Julien Mairal, Priya Goyal, Piotr Bojanowski, and Armand Joulin. Unsupervised learning of visual features by contrasting cluster assignments. Advances in Neural Information Processing Systems, 33:9912-9924, 2020. 2", + "[5] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International conference on machine learning, pages 1597-1607. PMLR, 2020. 2, 5", + "[6] Ting Chen, Simon Kornblith, Kevin Swersky, Mohammad Norouzi, and Geoffrey E Hinton. Big self-supervised models are strong semi-supervised learners. Advances in neural information processing systems, 33:22243-22255, 2020. 2", + "[7] Yutong Chen, Fangyun Wei, Xiao Sun, Zhirong Wu, and Stephen Lin. A simple multi-modality transfer learning baseline for sign language translation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5120-5130, 2022. 3, 4, 5", + "[8] Ka Leong Cheng, Zhaoyang Yang, Qifeng Chen, and Yu-Wing Tai. Fully convolutional networks for continuous sign language recognition. In European Conference on Computer Vision, pages 697-714. Springer, 2020. 2, 3, 5, 6", + "[9] Necati Cihan Camgoz, Simon Hadfield, Oscar Koller, and Richard Bowden. Subunets: End-to-end hand shape and continuous sign language recognition. In Proceedings of the IEEE international conference on computer vision, pages 3056-3065, 2017. 1, 5", + "[10] Runpeng Cui, Hu Liu, and Changshui Zhang. Recurrent convolutional neural networks for continuous sign language recognition by staged optimization. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7361-7369, 2017. 1", + "[11] Runpeng Cui, Hu Liu, and Changshui Zhang. A deep neural framework for continuous sign language recognition by iterative training. IEEE Transactions on Multimedia, 21(7):1880-1891, 2019. 2, 3, 5", + "[12] Cleison Correia de Amorim, David Macêdo, and Cleber Zanchettin. Spatial-temporal graph convolutional networks for sign language recognition. In International Conference on Artificial Neural Networks, pages 646-657. Springer, 2019. 2", + "[13] Zhangyang Gao, Cheng Tan, Lirong Wu, and Stan Z Li. Simvp: Simpler yet better video prediction. In Proceedings" + ], + "bbox": [ + 78, + 114, + 468, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3170-3180, 2022. 2", + "[14] Alex Graves, Santiago Fernández, Faustino Gomez, and Jürgen Schmidhuber. Connectionist temporal classification: labelling unsegmented sequence data with recurrent neural networks. In Proceedings of the 23rd international conference on Machine learning, pages 369-376, 2006. 2", + "[15] Aiming Hao, Yuecong Min, and Xilin Chen. Self-mutual distillation learning for continuous sign language recognition. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11303-11312, 2021. 1, 2, 3, 4, 5, 6", + "[16] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9729-9738, 2020. 5", + "[17] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 1, 4", + "[18] Yi Ji, Chunping Liu, Shengrong Gong, and Weidong Cheng. 3d hand gesture coding for sign language learning. In 2016 International Conference on Virtual Reality and Visualization (ICVRV), pages 407-410. IEEE, 2016. 2", + "[19] Suhail Muhammad Kamal, Yidong Chen, Shaozi Li, Xiaodong Shi, and Jiangbin Zheng. Technical approaches to Chinese sign language processing: a review. IEEE Access, 7:96926-96935, 2019. 1, 2, 8", + "[20] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6", + "[21] Diederik P Kingma and Max Welling. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114, 2013. 3", + "[22] Oscar Koller, Necati Cihan Camgoz, Hermann Ney, and Richard Bowden. Weakly supervised learning with multistream cnn-lstm-hmms to discover sequential parallelism in sign language videos. IEEE transactions on pattern analysis and machine intelligence, 42(9):2306-2320, 2019. 1, 3, 6", + "[23] Oscar Koller, Jens Forster, and Hermann Ney. Continuous sign language recognition: Towards large vocabulary statistical recognition systems handling multiple signers. Computer Vision and Image Understanding, 141:108-125, 2015. 1, 5", + "[24] Oscar Koller, Sepehr Zargaran, and Hermann Ney. Re-sign: Re-aligned end-to-end sequence modelling with deep recurrent cnn-hmms. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4297-4305, 2017. 1, 3, 5", + "[25] Siyuan Li, Zedong Wang, Zicheng Liu, Cheng Tan, Haitao Lin, Di Wu, Zhiyuan Chen, Jiangbin Zheng, and Stan Z Li. Efficient multi-order gated aggregation network. arXiv preprint arXiv:2211.03295, 2022. 1", + "[26] Zhi-jie Liang, Sheng-bin Liao, and Bing-zhang Hu. 3d convolutional neural networks for dynamic sign language recognition. The Computer Journal, 61(11):1724-1736, 2018. 2", + "[27] Ilya Loshchilov and Frank Hutter. Fixing weight decay regularization in adam. 2018. 6" + ], + "bbox": [ + 501, + 92, + 890, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "23149", + "bbox": [ + 478, + 925, + 519, + 936 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[28] Yuecong Min, Aiming Hao, Xiujuan Chai, and Xilin Chen. Visual alignment constraint for continuous sign language recognition. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11542-11551, 2021. 2, 3, 4, 5", + "[29] Zhe Niu and Brian Mak. Stochastic fine-grained labeling of multi-state sign glosses for continuous sign language recognition. In European Conference on Computer Vision, pages 172-186. Springer, 2020. 1, 2, 3, 5, 6", + "[30] Tse-Yu Pan, Li-Yun Lo, Chung-Wei Yeh, Jhe-Wei Li, Hou-Tim Liu, and Min-Chun Hu. Sign language recognition in complex background scene based on adaptive skin colour modelling and support vector machine. International Journal of Big Data Intelligence, 5(1-2):21-30, 2018. 2", + "[31] Jinye Peng, Li Yulong, et al. Chinese sign language recognition based on gray-level co-occurrence matrix and other multi-features fusion. In 2009 4th IEEE Conference on Industrial Electronics and Applications, pages 1569–1572. IEEE, 2009. 2", + "[32] Junfu Pu, Wengang Zhou, Hezhen Hu, and Houqiang Li. Boosting continuous sign language recognition via cross modality augmentation. In Proceedings of the 28th ACM International Conference on Multimedia, pages 1497-1505, 2020. 1, 5", + "[33] Junfu Pu, Wengang Zhou, and Houqiang Li. Iterative alignment network for continuous sign language recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4165-4174, 2019. 1, 2, 5", + "[34] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 1, 2, 5", + "[35] Danilo Jimenez Rezende, Shakir Mohamed, and Daan Wierstra. Stochastic backpropagation and approximate inference in deep generative models. In International conference on machine learning, pages 1278-1286. PMLR, 2014. 3", + "[36] Ramprasaath R Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra. Grad-cam: Visual explanations from deep networks via gradient-based localization. In Proceedings of the IEEE international conference on computer vision, pages 618-626, 2017. 8", + "[37] Cheng Tan, Zhangyang Gao, Siyuan Li, Yongjie Xu, and Stan Z Li. Temporal attention unit: Towards efficient spatiotemporal predictive learning. arXiv preprint arXiv:2206.12126, 2022. 2", + "[38] Cheng Tan, Jun Xia, Lirong Wu, and Stan Z Li. Co-learning: Learning from noisy labels with self-supervision. In Proceedings of the 29th ACM International Conference on Multimedia, pages 1405-1413, 2021. 2", + "[39] Yiqi Tong, Jiangbin Zheng, Hongkang Zhu, Yidong Chen, and Xiaodong Shi. A document-level neural machine translation model with dynamic caching guided by theme-rheme information. In Proceedings of the 28th International Con-" + ], + "bbox": [ + 78, + 90, + 470, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "ference on Computational Linguistics, pages 4385-4395, 2020.1", + "[40] Hanjie Wang, Xiujuan Chai, Xiaopeng Hong, Guoying Zhao, and Xilin Chen. Isolated sign language recognition with grassmann covariance matrices. ACM Transactions on Accessible Computing (TACCESS), 8(4):1-21, 2016. 2", + "[41] Hanjie Wang, Xiujuan Chai, Yu Zhou, and Xin Chen. Fast sign language recognition benefited from low rank approximation. In 2015 11th IEEE International Conference and Workshops on Automatic Face and Gesture Recognition (FG), volume 1, pages 1-6. IEEE, 2015. 2", + "[42] Chong Wu, Jiangbin Zheng, Zhenan Feng, Houwang Zhang, Le Zhang, Jiawang Cao, and Hong Yan. Fuzzy slic: Fuzzy simple linear iterative clustering. IEEE Transactions on Circuits and Systems for Video Technology, 31(6):2114-2124, 2020. 1", + "[43] Jiangbin Zheng, Yidong Chen, Chong Wu, Xiaodong Shi, and Suhail Muhammad Kamal. Enhancing neural sign language translation by highlighting the facial expression information. Neurocomputing, 464:462-472, 2021. 1, 3", + "[44] Jiangbin Zheng, Siyuan Li, Cheng Tan, Chong Wu, Yidong Chen, and Stan Z Li. Leveraging graph-based cross-modal information fusion for neural sign language translation. arXiv preprint arXiv:2211.00526, 2022. 3", + "[45] Jiangbin Zheng, Yile Wang, Ge Wang, Jun Xia, Yufei Huang, Guojiang Zhao, Yue Zhang, and Stan Z Li. Using context-to-vector with graph retrofitting to improve word embeddings. arXiv preprint arXiv:2210.16848, 2022. 5", + "[46] Jiangbin Zheng, Zheng Zhao, Min Chen, Jing Chen, Chong Wu, Yidong Chen, Xiaodong Shi, and Yiqi Tong. An improved sign language translation model with explainable adaptations for processing long sign sentences. Computational Intelligence and Neuroscience, 2020, 2020. 3", + "[47] Hao Zhou, Wengang Zhou, and Houqiang Li. Dynamic pseudo label decoding for continuous sign language recognition. In 2019 IEEE International Conference on Multimedia and Expo (ICME), pages 1282-1287. IEEE, 2019. 5", + "[48] Hao Zhou, Wengang Zhou, Yun Zhou, and Houqiang Li. Spatial-temporal multi-cue network for continuous sign language recognition. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 34, pages 13009-13016, 2020. 1, 3, 5, 6", + "[49] Yongjie Zhuang, Bo Lv, Xinjun Sheng, and Xiangyang Zhu. Towards Chinese sign language recognition using surface electromyography and accelerometers. In 2017 24th International Conference on Mechatronics and Machine Vision in Practice (M2VIP), pages 1-5. IEEE, 2017. 2", + "[50] Ronglai Zuo and Brian Mak. C2slr: Consistency-enhanced continuous sign language recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5131-5140, 2022. 1, 2, 3, 5, 6" + ], + "bbox": [ + 501, + 92, + 893, + 811 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "23150", + "bbox": [ + 478, + 926, + 519, + 936 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2023/CVT-SLR_ Contrastive Visual-Textual Transformation for Sign Language Recognition With Variational Alignment/012722bf-9201-40ff-b0cf-52287586b76f_model.json b/2023/CVT-SLR_ Contrastive Visual-Textual Transformation for Sign Language Recognition With Variational Alignment/012722bf-9201-40ff-b0cf-52287586b76f_model.json new file mode 100644 index 0000000000000000000000000000000000000000..d94e774118089ac6f384da676d5f14c2d5d28582 --- /dev/null +++ b/2023/CVT-SLR_ Contrastive Visual-Textual Transformation for Sign Language Recognition With Variational Alignment/012722bf-9201-40ff-b0cf-52287586b76f_model.json @@ -0,0 +1,2145 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.0, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.115, + 0.13, + 0.856, + 0.177 + ], + "angle": 0, + "content": "CVT-SLR: Contrastive Visual-Textual Transformation for Sign Language Recognition with Variational Alignment" + }, + { + "type": "text", + "bbox": [ + 0.25, + 0.203, + 0.72, + 0.239 + ], + "angle": 0, + "content": "Jiangbin Zheng\\(^{1}\\), Yile Wang\\(^{1,2}\\), Cheng Tan\\(^{1}\\), Siyuan Li\\(^{1}\\), Ge Wang\\(^{1}\\), Jun Xia\\(^{1}\\), Yidong Chen\\(^{3}\\), Stan Z. Li\\(^{1*}\\)" + }, + { + "type": "text", + "bbox": [ + 0.192, + 0.24, + 0.78, + 0.294 + ], + "angle": 0, + "content": "\\(^{1}\\)AI Lab, Research Center for Industries of the Future, Westlake University \\(^{2}\\)Institute for AI Industry Research (AIR), Tsinghua University \\(^{3}\\)School of Informatics, Xiamen University" + }, + { + "type": "text", + "bbox": [ + 0.119, + 0.295, + 0.851, + 0.327 + ], + "angle": 0, + "content": "{zhengjiangbin, wangyile, tancheng, lisiyuan, wangge, xiajun, Stan.ZQ.Li}@westlake.edu.cn \nydchen@xmu.edu.cn" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.362, + 0.314, + 0.379 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.395, + 0.473, + 0.849 + ], + "angle": 0, + "content": "Sign language recognition (SLR) is a weakly supervised task that annotates sign videos as textual glosses. Recent studies show that insufficient training caused by the lack of large-scale available sign datasets becomes the main bottleneck for SLR. Most SLR works thereby adopt pretrained visual modules and develop two mainstream solutions. The multi-stream architectures extend multi-cue visual features, yielding the current SOTA performances but requiring complex designs and might introduce potential noise. Alternatively, the advanced single-cue SLR frameworks using explicit cross-modal alignment between visual and textual modalities are simple and effective, potentially competitive with the multi-cue framework. In this work, we propose a novel contrastive visual-textual transformation for SLR, CVT-SLR, to fully explore the pretrained knowledge of both the visual and language modalities. Based on the single-cue cross-modal alignment framework, we propose a variational autoencoder (VAE) for pretrained contextual knowledge while introducing the complete pretrained language module. The VAE implicitly aligns visual and textual modalities while benefiting from pretrained contextual knowledge as the traditional contextual module. Meanwhile, a contrastive cross-modal alignment algorithm is designed to explicitly enhance the consistency constraints. Extensive experiments on public datasets (PHOENIX-2014 and PHOENIX-2014T) demonstrate that our proposed CVT-SLR consistently outperforms existing single-cue methods and even outperforms SOTA multi-cue methods. The source codes and models are available at https://github.com/binbinjiang/CVT-SLR" + }, + { + "type": "title", + "bbox": [ + 0.502, + 0.362, + 0.631, + 0.379 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.393, + 0.895, + 0.604 + ], + "angle": 0, + "content": "As a special visual natural language, sign language is the primary communication medium of the deaf community [19]. With the progress of deep learning [1, 17, 25, 39, 42], sign language recognition (SLR) has emerged as a multimodal task that aims to annotate sign videos into textual sign glosses. However, a significant dilemma of SLR is the lack of publicly available sign language datasets. For example, the most commonly-used PHOENIX-2014 [23] and PHOENIX-2014T [2] datasets only include about 10K pairs of sign videos and gloss annotations, which are far from training a robust SLR system with full supervision as typical vision-language cross-modal tasks [34]. Therefore, data limitation that may easily lead to insufficient training or overfitting problems is the main bottleneck of SLR tasks." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.609, + 0.896, + 0.882 + ], + "angle": 0, + "content": "The development of weakly supervised SLR has witnessed most of the improvement efforts focus on the visual module (e.g., CNN) [9, 10, 15, 29, 32, 33]. Transferring pretrained visual networks from general domains of human actions becomes a consensus to alleviate the low-resource limitation. The mainstream multi-stream SLR framework extends the pretrained visual module with multi-cue visual information [3, 22, 24, 43, 48, 50], including global features and regional features such as hands and faces in independent streams. The theoretical support for this approach comes from sign language linguistics, where sign language utilizes multiple complementary channels (e.g., hand shapes, facial expressions) to convey information [3]. The multi-cue mechanism essentially exploits hard attention to key information, yielding the current SOTA performances. However, the multi-cue framework is complex (e.g., cropping multiple regions, requiring more parameters), and the fusion of multiple streams might introduce additional potential noise." + }, + { + "type": "text", + "bbox": [ + 0.52, + 0.886, + 0.892, + 0.902 + ], + "angle": 0, + "content": "Another mainstream advanced solution is the single-cue" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.887, + 0.222, + 0.9 + ], + "angle": 0, + "content": "*Corresponding author." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.925, + 0.518, + 0.937 + ], + "angle": 0, + "content": "23141" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.117, + 0.094, + 0.852, + 0.314 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.319, + 0.895, + 0.39 + ], + "angle": 0, + "content": "Figure 1. (a) An advanced single-cue SLR framework with explicit cross-modal alignment; (b) Our proposed single-cue SLR framework with explicit cross-modal alignment and implicit autoencoder alignment. Both frameworks use pretrained visual features. But our framework uses the autoencoder module to replace the mainstream contextual module, which not only includes the functions of the contextual module but also can introduce complete pretrained language knowledge and implicit cross-modal alignment. To maximize the preservation of the complete pretrained language parameters and migrated visual features, a video-gloss adapter is introduced." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.415, + 0.473, + 0.673 + ], + "angle": 0, + "content": "cross-modal alignment framework [15, 28], which consists of a pretrained visual module followed by a contextual module (e.g., RNN, LSTM, Transformer) and a Connectionist Temporal Classification (CTC) [14] based alignment module for gloss generation, as shown in Figure 1 (a). Explicit cross-modal alignment constraints further improve feature interactions [15, 28, 38], which could be treated as a kind of consistency between two different modalities [50] facilitating the visual module learn long-term temporal information from contextual module [13, 37]. The cross-modal alignment framework is simple and effective, potentially competitive with the multi-cue framework. Despite the advanced performance of complex multi-cue architectures with pretrained visual modules, the cross-modal consistency is a more elegant design for practical usage. It also implies the potential of prior contextual linguistic knowledge, which has been overlooked by existing SLR works." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.675, + 0.473, + 0.903 + ], + "angle": 0, + "content": "In this work, we propose a novel contrastive visual-textual transformation framework for SLR, called CVT-SLR, to fully explore the pretrained knowledge of both the visual and language modalities, as shown in Figure 1 (b). Based on the single-cue cross-modal alignment framework, CVT-SLR keeps the pretrained visual module but replaces the traditional contextual module with a variational autoencoder (VAE). Since a full encoder-decoder architecture is used, the VAE is responsible for learning pretrained contextual knowledge based on a pseudo-translation task while introducing the complete pretrained language module. In addition, the VAE maintains the consistency of input and output modalities due to the form of an autoencoder, playing an implicit cross-modal alignment role. Furthermore, inspired by contrastive learning [4-6,34], we introduce a contrastive" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.415, + 0.895, + 0.55 + ], + "angle": 0, + "content": "alignment algorithm that focuses on both positive and negative samples to enhance explicit cross-modal consistency constraints. Extensive quantitative experiments conducted on the public datasets PHOENIX-2014 and PHOENIX-2014T demonstrate the advance of the proposed CVT-SLR framework. Through ablation study and qualitative analysis, we also verify the effectiveness of introducing pretrained language knowledge and the new consistency constraint mechanism." + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.552, + 0.776, + 0.566 + ], + "angle": 0, + "content": "Our main contributions are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.569, + 0.892, + 0.63 + ], + "angle": 0, + "content": "- A novel visual-textual transformation-based SLR framework is proposed, which introduces fully pretrained language knowledge for the first time and provides new approaches for other cross-modal tasks." + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.633, + 0.892, + 0.709 + ], + "angle": 0, + "content": "- New alignment methods are proposed for cross-modal consistency constraints: a) exploiting the special properties of the autoencoder to implicitly align visual and textual modalities; b) introducing an explicit contrastive cross-modal alignment method." + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.712, + 0.891, + 0.758 + ], + "angle": 0, + "content": "- The proposed single-cue SLR framework not only outperforms existing single-cue baselines by a large margin but even surpasses SOTA multi-cue baselines." + }, + { + "type": "list", + "bbox": [ + 0.519, + 0.569, + 0.892, + 0.758 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.77, + 0.642, + 0.785 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.795, + 0.895, + 0.903 + ], + "angle": 0, + "content": "SLR tasks are generally divided into three categories [19]: finger-spelling recognition [18, 30, 31], isolated word recognition [26, 40, 41, 49] and continuous sign sentence recognition [8, 11, 12, 29, 33]. In the early days, SLR works mainly focused on lexical-level tasks, such as fingerspelling recognition and isolated word recognition. Nowadays, the more practical continuous sign sentence recogni" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.925, + 0.521, + 0.937 + ], + "angle": 0, + "content": "23142" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.136 + ], + "angle": 0, + "content": "tion task has become mainstream sign language research. In this work, the mentioned SLRs refer specifically to continuous sign sentence recognition." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.137, + 0.47, + 0.302 + ], + "angle": 0, + "content": "The recent SLR works [8, 11, 15, 29] are summarized based on these three aspects: feature extraction, recognition, and alignment. Mostly, the components of feature extraction are composed of the visual module and the contextual module. The visual module encodes short-term spatial information, while the contextual module encodes long-term context information. Based on the extracted features, the classifiers can get a posterior probability for recognition. And the alignment module is required to find the proper alignment between clips and glosses to ensure an accurate training procedure." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.303, + 0.469, + 0.529 + ], + "angle": 0, + "content": "A few studies extend the common SLR framework by incorporating multi-cue information [22,24,43,44,48]. The multi-cue features mainly include information from hand shapes, facial expressions, mouths, and poses. The current SOTA SLR works are based on multi-cue mechanisms, such as \\(\\mathrm{C}^2\\mathrm{SLR}\\) [50]. Recently, some works show that explicitly enforcing the consistency between the visual and textual modules can also get comparable performance, although only single-cue features are considered. For example, VAC [28] treats the visual and textual modules as the student and teacher, respectively, and achieves knowledge distillation to align the visual and the textual modalities. Similarly, SMKD [15] achieves knowledge transfer by further sharing classifiers between different modalities. Note that SMKD is the current SOTA single-cue method." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.53, + 0.47, + 0.636 + ], + "angle": 0, + "content": "It is worth mentioning that neural sign language translation [2,3,7,46] is another popular sign language task. As a distinction, the neural sign language translation task aims to convert sign videos into equivalent spoken language translations, further focusing on semantic relationships and adjusting word order [2], while the SLR task only annotates sign videos into gloss sequences in the frame order." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.648, + 0.176, + 0.663 + ], + "angle": 0, + "content": "3. Methods" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.673, + 0.47, + 0.779 + ], + "angle": 0, + "content": "As shown in Figure 2, our training pipeline is roughly divided into two steps. Step 1 is to pretrain a VAE network as the textual module. And step 2 is to transfer the off-the-shelf visual module (usually a publicly available CNN trained on Kinetics/ImageNet) and the pretrained textual module from step 1 into the CVT-SLR framework. A video-gloss adapter is used to bridge these two pretrained modules." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.787, + 0.318, + 0.803 + ], + "angle": 0, + "content": "3.1. Textual Model Pretraining" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.81, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Our VAE model is an asymmetric encoder-decoder architecture, where the encoder consists of self-attention layers, and the decoder consists of Bi-LSTM layers, as shown in the top part of Figure 2. To pretrain the VAE for our proposed SLR framework solely based on the gloss data from the PHOENIX-14 training set (excluding the development" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.167 + ], + "angle": 0, + "content": "set and testing set), we construct a pseudo-translation task, Gloss2Gloss, which enables the VAE to learn the contextual semantic knowledge. Formally, it is assumed that the input gloss of the encoder is denoted as \\( G \\), and the generative gloss of the decoder is denoted as \\( G^{*} \\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.168, + 0.892, + 0.289 + ], + "angle": 0, + "content": "Posterior Self-attention Encoder. We employ the neural approximation, and reparameterization [21, 35] techniques to approximate the posterior self-attention encoder for generating the posterior distribution \\( q_{\\phi}(z|G) \\), where \\( \\phi \\) denotes the variational parameters of the encoder. The \\( q_{\\phi}(z|G) \\) is a diagonal Gaussian distribution \\( N(\\mu, \\mathrm{diag}(\\sigma^2)) \\), and two single-layer MLPs are used to parameterize the mean \\( \\mu \\) and the variance \\( \\sigma^2 \\), respectively, as:" + }, + { + "type": "equation", + "bbox": [ + 0.643, + 0.297, + 0.892, + 0.312 + ], + "angle": 0, + "content": "\\[\nz = \\mu + \\sigma \\odot \\epsilon , \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.32, + 0.892, + 0.349 + ], + "angle": 0, + "content": "where \\(\\epsilon\\) is a standard Gaussian noise, and \\(\\odot\\) denotes an element-wise product." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.35, + 0.892, + 0.393 + ], + "angle": 0, + "content": "Generative Bi-LSTM Decoder. Given the continuous latent variable \\(z\\), then the translation target \\(G^{*}\\) is generated from \\(z\\) as:" + }, + { + "type": "equation", + "bbox": [ + 0.598, + 0.395, + 0.891, + 0.412 + ], + "angle": 0, + "content": "\\[\np _ {\\theta} \\left(G ^ {*}, z\\right) = p _ {\\theta} \\left(G ^ {*} \\mid z\\right) p _ {\\theta} (z), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.415, + 0.892, + 0.461 + ], + "angle": 0, + "content": "where \\(\\theta\\) denotes the parameters of the generative Bi-LSTM decoder, and \\(p_{\\theta}(G^{*}|z)\\) is the conditional distribution that models the generation procedure estimated via the decoder." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.461, + 0.892, + 0.535 + ], + "angle": 0, + "content": "Variational Optimization. Both the generative model \\( p_{\\theta}(G^{*},z) \\) and the posterior inference model \\( q_{\\phi}(z|G) \\) are incorporated tightly in an end-to-end manner, and we can apply a standard back-propagation to optimize the gradient of the variational lower bound:" + }, + { + "type": "equation", + "bbox": [ + 0.563, + 0.542, + 0.892, + 0.577 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\begin{array}{l} \\mathcal {L} _ {\\mathrm {V A E}} (\\theta , \\phi ; G) = - \\mathrm {K L} (q _ {\\phi} (z | G) | | p _ {\\theta} (z)) \\\\ \\vdots \\end{array} \\tag {3} \\\\ + \\mathbb {E} _ {q _ {\\phi} (z | G)} \\left[ \\log p _ {\\theta} \\left(G ^ {*} \\mid z\\right) \\right] \\leq \\log p _ {\\theta} \\left(G ^ {*}\\right), \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.583, + 0.892, + 0.63 + ], + "angle": 0, + "content": "where KL denotes the Kullback-Leibler divergence and the prior \\(p_{\\theta}(z)\\) denotes a standard Gaussian distribution. Practically, the first KL term of Eq. (3) can be written as:" + }, + { + "type": "equation", + "bbox": [ + 0.577, + 0.636, + 0.892, + 0.665 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {K L}} = - \\frac {1}{2} \\left(\\log \\sigma^ {2} - \\sigma^ {2} - \\mu^ {2} + 1\\right). \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.67, + 0.892, + 0.716 + ], + "angle": 0, + "content": "Since our main target is a Gloss2Gloss translation task, we employ a cross-entropy loss to maximize the generative gloss probability as the second term of Eq. (3):" + }, + { + "type": "equation", + "bbox": [ + 0.577, + 0.724, + 0.892, + 0.74 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {g l o s s 2 g l o s s}} = \\text {C r o s s E n t r o p y} (G, G ^ {*}). \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.747, + 0.892, + 0.779 + ], + "angle": 0, + "content": "Therefore, the overall optimization objective of textual model pretraining is a joint loss of \\(\\mathcal{L}_{\\mathrm{KL}}\\) and \\(\\mathcal{L}_{\\mathrm{gloss2gloss}}\\) as:" + }, + { + "type": "equation", + "bbox": [ + 0.607, + 0.786, + 0.892, + 0.802 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {V A E}} = \\mathcal {L} _ {\\mathrm {K L}} + \\mathcal {L} _ {\\text {g l o s s 2 g l o s s}}. \\tag {6}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.809, + 0.78, + 0.825 + ], + "angle": 0, + "content": "3.2. Training CVT-SLR Framework" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.832, + 0.676, + 0.848 + ], + "angle": 0, + "content": "3.2.1 Training Pipeline" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.892, + 0.901 + ], + "angle": 0, + "content": "In the CVT-SLR framework, we sequentially concatenate the pretrained visual CNN module and the pretrained textual VAE module through a video-gloss adapter, as shown" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.926, + 0.519, + 0.938 + ], + "angle": 0, + "content": "23143" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.123, + 0.09, + 0.852, + 0.334 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.339, + 0.895, + 0.369 + ], + "angle": 0, + "content": "Figure 2. Overview of proposed CVT-SLR framework. Our training pipeline is divided into two steps. Step 1: Pretrain the VAE module with the self-supervised gloss-to-gloss pseudo-translation task. Step 2: Train the CVT-SLR based on the VAE model from Step 1." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.388, + 0.47, + 0.508 + ], + "angle": 0, + "content": "in Figure 2. The pretrained ResNet18 model [17] is selected as the frame-wise feature extractor, which is trained on large-scale human action datasets. For the VAE module, although Gaussian noise is introduced in the reparameterization process in step 1, we do not add any noise after migrating the VAE to the SLR framework in step 2. In this way, the reparameterization module of the VAE degenerates into two common linear layers, i.e., \\(\\epsilon\\) is removed in Eq. (1)." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.509, + 0.471, + 0.72 + ], + "angle": 0, + "content": "Formally, given a sign video with \\(T\\) RGB frames \\(\\mathcal{F} = \\{\\mathcal{F}_t\\}_{t=1}^T \\in \\mathbb{R}^{T \\times H \\times W \\times 3}\\), the visual module, which is composed of a stack of 2D-CNN layers and a global average pooling layer, first extracts spatial features \\(v = \\{v_t\\}_{t=1}^T \\in \\mathbb{R}^{T \\times d}\\). Then the textual module further extracts temporal features \\(s = \\{s_t\\}_{t=1}^T \\in \\mathbb{R}^{T \\times d}\\). Finally, the alignment module utilizes CTC to compute the conditional probability of the gloss label sequence \\(p(\\mathcal{G}|\\mathcal{F})\\) based on temporal features \\(s\\), where \\(\\mathcal{G} = \\{\\mathcal{G}_i\\}_{i=1}^N\\) is a generated sign gloss sequence with \\(N\\) glosses. A shared classifier is used between the visual module and the textual module to generate spatial and temporal probability, respectively. And the proposed contrastive cross-modal alignment algorithm further enhances the consistency across different modalities." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.74, + 0.279, + 0.756 + ], + "angle": 0, + "content": "3.2.2 Video-Gloss Adapter" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.47, + 0.885 + ], + "angle": 0, + "content": "To maximize the preservation of both pretrained visual and textual knowledge, we use a video-gloss adapter to connect spatial-temporal features, as shown in Figure 2. The videogloss adapter is simply implemented as a fully-connected MLP with two hidden layers. We take the CNN output probabilities as the input of the adapter where the final fully connected layer is initialized using the weights of the pretrained word embedding from the pretrained VAE [7]." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.886, + 0.47, + 0.901 + ], + "angle": 0, + "content": "Since both the input and output are in the same textual" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.388, + 0.892, + 0.48 + ], + "angle": 0, + "content": "form during the textual model pretraining in step 1, the VAE plays an implicit cross-modal alignment role after being transferred to the CVT-SLR framework in step 2, which is close to the effect of consistency. Therefore, the videogloss adapter is essentially a mediator for activating the consistency between the spatial and temporal modalities." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.496, + 0.667, + 0.512 + ], + "angle": 0, + "content": "3.2.3 CTC Alignment" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.52, + 0.892, + 0.641 + ], + "angle": 0, + "content": "Due to the spatial-temporal nature of sign videos, glosses have a one-to-many mapping to video frames but share the same ordering. Unlike spoken language texts, the continuous gloss annotations are chronologically consistent with the sign videos. Concretely, for an input sign video \\(\\mathcal{F}\\) and the corresponding generated gloss sequence \\(\\mathcal{G}\\), we use CTC to compute \\(p(\\mathcal{G}|\\mathcal{F})\\) by marginalizing over all possible \\(\\mathcal{F}\\) to \\(\\mathcal{G}\\) alignments:" + }, + { + "type": "equation", + "bbox": [ + 0.616, + 0.649, + 0.892, + 0.681 + ], + "angle": 0, + "content": "\\[\np (\\mathcal {G} | \\mathcal {F}) = \\sum_ {\\pi \\in \\mathcal {B}} p (\\pi | \\mathcal {F}), \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.689, + 0.891, + 0.734 + ], + "angle": 0, + "content": "where \\(\\pi\\) denotes a path and \\(\\mathcal{B}\\) is the set of all viable paths corresponding to \\(\\mathcal{G}\\). The probability \\(p(\\pi|\\mathcal{F})\\) is computed by the textual module. The CTC loss is then formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.62, + 0.743, + 0.892, + 0.761 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {c t c}} = - \\log p (\\mathcal {G} ^ {*} | \\mathcal {F}), \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.769, + 0.797, + 0.785 + ], + "angle": 0, + "content": "where \\(\\mathcal{G}^*\\) is the ground truth gloss sequence." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.802, + 0.714, + 0.817 + ], + "angle": 0, + "content": "3.2.4 Contrastive Alignment" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.825, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Aligning both visual and textual modalities help improve SLR [28]. To enforce alignment constraints across modalities, we propose a contrastive alignment loss. In addition, shared classifiers are also used as an improvement technique between visual and textual features [15]." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.925, + 0.521, + 0.937 + ], + "angle": 0, + "content": "23144" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.131, + 0.089, + 0.837, + 0.45 + ], + "angle": 0, + "content": "
GroupsModelsDev (%)Test (%)Cues
DEL/INSWERDEL/INSWER
Group 1SubUNet [9]14.6/4.040.814.3/4.040.7video
Staged-Opt [9]13.7/7.339.412.2/7.538.7video
Align-iOpt [33]12.6/2.637.113.0/2.536.7video
DPD+TEM [47]9.5/3.235.69.3/3.134.5video
Re-Sign [24]-27.1-26.8video
SFL [29]7.9/6.526.27.5/6.326.8video
DNF [11]7.8/3.523.87.8/3.424.4video
FCN [8]-23.7-23.9video
VAC [28]7.9/2.521.28.4/2.622.3video
CMA [32]7.3/2.721.37.3/2.421.9video
SFL [29]10.3/4.124.910.4/3.625.3video
VL-SLT [7]-21.9-22.5video
SMKD [15]6.8/2.520.86.3/2.321.0video
Group 2DNF [11]7.3/3.323.16.7/3.322.9video+optical flow
STMC [48]7.7/3.421.17.4/2.620.7video+hand+face+pose
C2SLR [50]-20.5-20.4video+keypoints
Group 3Ours1 (w/o VAE+Contra)7.1/3.021.17.3/2.921.4video
Ours2 (w/ VAE)6.5/2.420.26.3/2.220.3video
Ours3 (w/ Contra)6.7/2.720.46.4/2.520.7video
Ours4 (w/ VAE+Contra)6.4/2.619.86.1/2.320.1video
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.457, + 0.892, + 0.486 + ], + "angle": 0, + "content": "Table 1. Performance comparison (%) on PHOENIX-14 dataset. DEL/INS: deletion error and insertion error. The best results and SOTA baseline for each group are marked as bold and underlined." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.508, + 0.47, + 0.66 + ], + "angle": 0, + "content": "In previous SLR works, cross-modal alignment only focuses on positive samples [15, 48]. Inspired by contrastive learning [5, 16, 45], we construct both positive and negative samples in the same mini-batch and implement a contrastive cross-modal alignment method to ensure that similar features are closer while different features are farther apart. Given that the normalized spatial features from CNN as \\( S_{\\mathrm{logits}} \\in \\mathbb{R}^{B \\times T \\times d} \\), and the normalized temporal features from VAE as \\( \\mathcal{V}_{\\mathrm{logits}} \\in \\mathbb{R}^{B \\times T \\times d} \\), where \\( B \\) denotes the number of samples. Then we compute pair matrices firstly as:" + }, + { + "type": "equation", + "bbox": [ + 0.186, + 0.667, + 0.469, + 0.687 + ], + "angle": 0, + "content": "\\[\n\\mathcal {S} 2 \\mathcal {V} _ {\\text {p a i r}} = \\mathcal {S} _ {\\text {l o g i t s}} \\times \\mathcal {V} _ {\\text {l o g i t s}} ^ {T}, \\tag {9}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.188, + 0.695, + 0.469, + 0.715 + ], + "angle": 0, + "content": "\\[\n\\mathcal {V} 2 \\mathcal {S} _ {\\text {p a i r}} = \\mathcal {V} _ {\\log_ {1 5}} \\times \\mathcal {S} _ {\\log_ {1 5}} ^ {T}, \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.72, + 0.47, + 0.901 + ], + "angle": 0, + "content": "Taking \\( S2\\mathcal{V}_{\\mathrm{pair}} \\in \\mathbb{R}^{B \\times B} \\) pair matrix (denoted as \\( \\mathcal{P} \\)) as an example, \\( \\mathcal{P}[i,j](0 \\leq i,j < B) \\) represents the similarity value between the visual feature of the \\( i \\)-th batch and the textual feature of \\( j \\)-th batch. The visual and textual features from the same input instances are positive samples, and the features from other different input instances are negative samples. Hence, in the matrix \\( \\mathcal{P} \\), the diagonal similarity values are from positive sample pairs, and the rest values are from negative sample pairs. To make the loss computation differentiable and easy, we convert the computation of positive and negative pairs into a binary classification task [34]. The value \\( i \\) corresponding to the \\( i \\)-th row in the matrix \\( \\mathcal{P} \\)" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.508, + 0.892, + 0.556 + ], + "angle": 0, + "content": "is a positive label. Therefore, the corresponding labels of \\(\\mathcal{P}\\) are Labels \\(= \\{0,1,\\dots ,i,\\dots ,B\\}\\), then the contrastive alignment loss related to \\(S2\\mathcal{V}_{\\mathrm{pair}}\\) as:" + }, + { + "type": "equation", + "bbox": [ + 0.502, + 0.563, + 0.892, + 0.581 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {a l i g n} _ {S}} = \\text {C r o s s E n t r o p y} (\\text {S o f t m a x} (S 2 \\mathcal {V} \\text {p a i r}), \\text {L a b e l s}). \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.589, + 0.891, + 0.605 + ], + "angle": 0, + "content": "Similarly, the \\(\\mathcal{V}2S_{\\mathrm{pair}}\\) temporal-to-spatial alignment loss as:" + }, + { + "type": "equation", + "bbox": [ + 0.51, + 0.614, + 0.891, + 0.632 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {a l i g n} _ {\\nu}} = \\text {C r o s s E n t r o p y} (\\text {S o f t m a x} (\\mathcal {V} 2 S _ {\\text {p a i r}}), \\text {L a b e l s}). \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.64, + 0.856, + 0.655 + ], + "angle": 0, + "content": "Therefore, the complete contrastive alignment loss as:" + }, + { + "type": "equation", + "bbox": [ + 0.597, + 0.663, + 0.891, + 0.692 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {a l i g n}} = \\frac {1}{2} \\left(\\mathcal {L} _ {\\text {a l i g n} _ {S}} + \\mathcal {L} _ {\\text {a l i g n} _ {V}}\\right). \\tag {13}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.699, + 0.891, + 0.731 + ], + "angle": 0, + "content": "Overall, the objective of CVT-SLR is jointly optimized CTC loss \\(\\mathcal{L}_{\\mathrm{ctc}}\\) and contrastive alignment loss \\(\\mathcal{L}_{\\mathrm{align}}\\) as:" + }, + { + "type": "equation", + "bbox": [ + 0.611, + 0.74, + 0.891, + 0.756 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {C V T - S L R}} = \\mathcal {L} _ {\\mathrm {c t c}} + \\mathcal {L} _ {\\mathrm {a l i g n}}. \\tag {14}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.778, + 0.631, + 0.795 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.802, + 0.598, + 0.818 + ], + "angle": 0, + "content": "4.1. Settings" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.825, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Datasets. We use PHOENIX-2014 [23] and PHOENIX-2014T [2] as datasets, where RGB videos and their corresponding annotations are provided. PHOENIX-2014 is a German SLR dataset about weather forecasts with a vocabulary size of 1,081, which is divided into three parts: 5,672" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.926, + 0.518, + 0.938 + ], + "angle": 0, + "content": "23145" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.205, + 0.084, + 0.764, + 0.348 + ], + "angle": 0, + "content": "
GroupsModelsWERCues
Dev(%)Test(%)
Group 1SFL [29]25.126.1video
CNN+LSTM+HMM [22]24.526.5video
SLT [3]24.924.6video
FCN [8]23.325.1video
SMKD [15]20.822.4video
Group 2CNN+LSTM+HMM [22]24.525.4video+mouth
CNN+LSTM+HMM [22]22.124.1video+mouth+hand
SLT [3]24.624.5video+text
STMC [48]19.621.0video+hand+face+pose
C²SLR [50]20.220.4video+keypoints
Group 3Ours1 (w/o VAE+Contra)21.822.0video
Ours2 (w/ VAE)20.120.4video
Ours3 (w/ Contra)21.021.5video
Ours4 (w/ VAE+Contra)19.420.3video
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.351, + 0.895, + 0.381 + ], + "angle": 0, + "content": "Table 2. Performance comparison (%) on PHOENIX14T dataset. The best results and SOTA baseline for each group are marked as **bold** and **underlined**, respectively." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.397, + 0.471, + 0.472 + ], + "angle": 0, + "content": "instances for training, 540 for validation, and 629 for testing, respectively. And PHOENIX-2014T is an extension of PHOENIX-2014 with a vocabulary size of 1,085, which is also divided into three parts: 7,096 instances for training, 519 for validation, and 642 for testing, respectively." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.473, + 0.47, + 0.593 + ], + "angle": 0, + "content": "Evaluation Metric. Word error rate (WER) is the most widely-used metric to evaluate the SLR performance, which measures the number of necessary insertions, substitutions, and deletions in the recognized sentences when compared to the reference sentences. WER is an edit distance, indicating the least number of operations of substitutions (#sub), insertions (#ins), and deletions (#del) to transform the predicted sentences into the reference sequences (#reference):" + }, + { + "type": "equation", + "bbox": [ + 0.166, + 0.598, + 0.47, + 0.628 + ], + "angle": 0, + "content": "\\[\n\\mathrm {W E R} = \\frac {\\# \\mathrm {d e l} + \\# \\mathrm {i n s} + \\# \\mathrm {s u b}}{\\# \\text {r e f e r e n c e}}. \\tag {15}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.63, + 0.471, + 0.84 + ], + "angle": 0, + "content": "Implementation Details. The videos are resized to \\(256 \\times 256\\) and then cropped to \\(224 \\times 224\\). During SLR training, we use random crop and horizontal flips (50%) for data augmentation. During SLR testing, we only adopt center cropping. Beam Search is used for gloss generation. Our visual module has the same configuration as [15]. The VAE module contains two multi-head self-attention layers with 4 heads and two Bi-LSTM layers with \\(2 \\times 512\\) dimensional hidden states sequentially. Adam [20] optimizer is used to pretrain our VAE network with a batch size of 16, an initial learning rate of 1e-4. And AdamW [27] optimizer is used to train our SLR framework with a batch size of 8, a learning rate of 1e-4, and a weight decay of 1e-4. And 1 NVIDIA A100 80GB GPU is used." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.848, + 0.218, + 0.862 + ], + "angle": 0, + "content": "4.2. Main Results" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.871, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Table 1 and Table 2 show our comparisons with two-group different baselines, where Group 1 includes" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.397, + 0.892, + 0.502 + ], + "angle": 0, + "content": "single-cue baselines and Group 2 includes multi-cue baselines. Oursnumber denote our trained CVT-SLR framework, which have only two variables, \"VAE\" and \"Contra\". Note that \"w/o (w/)\" VAE\" denotes the VAE module in CVT-SLR framework without (with) transferred pretrained parameters, and \"w/o (w/)\" Contra\" denotes training CVT-SLR framework without (with) contrastive alignment loss." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.505, + 0.893, + 0.655 + ], + "angle": 0, + "content": "Evaluation on PHOENIX-14. Table 1 shows our methods compared to other methods on PHOENIX-14. The optimalWER scores of our single-cue CVT-SLR \\((\\mathrm{Ours}_4)\\) on the development set and testing set are \\(19.8\\%\\) and \\(20.1\\%\\) respectively. And \\(\\mathrm{Ours_4}\\) achieves the best performance among video-based only/single-cue methods in Group 1 (Ours vs. SMKD). Furthermore, while no extra cues are used, our CVT-SLR achieves the best result among the models trained with extra cues (i.e., multi-cue methods) in Group 2 (Ours vs. \\(\\mathrm{C}^2\\mathrm{SLR}\\))." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.657, + 0.894, + 0.852 + ], + "angle": 0, + "content": "From within-group comparisons in Group 3, we find that the case, \\( \\mathrm{Ours}_1 \\), which neither migrates the pretrained VAE parameters nor uses the contrastive alignment constraint, performs the worst. Nevertheless, relative to other groups, it still performs well. Based on \\( \\mathrm{Ours}_1 \\), by only configuring the pretrained VAE parameters \\( (\\mathrm{Ours}_2) \\) or the contrastive alignment constraints \\( (\\mathrm{Ours}_3) \\), the CVT-SLR framework can also gain gains, and the pretrained VAE brings more obvious gains \\( (+0.9\\% / +1.1\\% \\) vs. \\( +0.7\\% / +0.7\\% \\)). Overall, Group 3 demonstrates that both the prior language knowledge from the pretrained textual module and the contrastive constraints through cross-modal alignment play important enhancement roles for our SLR framework." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.894, + 0.901 + ], + "angle": 0, + "content": "Evaluation on PHOENIX-14T. In Table 2, we evaluate the CVT-SLR on PHOENIX-14T. We can see that our optimal model \\((\\mathrm{Ours}_4)\\) in this dataset also achieves the best per" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.926, + 0.521, + 0.938 + ], + "angle": 0, + "content": "23146" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.094, + 0.086, + 0.451, + 0.194 + ], + "angle": 0, + "content": "
#Pre-VisualPre-TextualWER
Dev(%)Test(%)
1××89.288.5
2×✓30.831.4
3✓×20.420.7
4✓✓19.820.1
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.2, + 0.47, + 0.242 + ], + "angle": 0, + "content": "Table 3. Ablation study of pretrained modules on PHOENIX-14 (Pre-Visual: pretrained visual module, Pre-Language: pretrained textual module). The best results are bolded." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.252, + 0.451, + 0.368 + ], + "angle": 0, + "content": "
#CTC WeightAlign WeightWER
Dev(%)Test(%)
11.0020.220.3
21.01.020.020.2
31.010.019.820.1
41.0100.020.120.6
51.01000.020.520.9
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.375, + 0.468, + 0.403 + ], + "angle": 0, + "content": "Table 4. Ablation study of varied weights of the CTC loss and the alignment loss on PHOENIX-14. The best results are bolded." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.419, + 0.468, + 0.539 + ], + "angle": 0, + "content": "formance (development set: \\(19.4\\%\\) and testing set: \\(20.3\\%\\)) with video-based only/single-cue information compared to Group 1 \\((\\text{Ours}_4\\) vs. SMKD). It is worth mentioning that the SOTA single-cue method (SMKD) fails to surpass the multi-cue baselines in their work on PHOENIX-14T. And our CVT-SLR makes up for their shortcomings while outperforming the SOTA multi-cue baselines in Group 2 on this dataset \\((\\text{Ours}_4\\) vs. \\(\\mathbf{C}^2\\mathbf{SLR})\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.54, + 0.468, + 0.675 + ], + "angle": 0, + "content": "Compared within Group 3 on the PHOENIX-14T, the observed conclusions are more or less the same as on the PHOENIX-14, i.e., both the pretrained textual module and the proposed contrastive alignment constraint have a beneficial effect on our CVT-SLR framework. But the pretrained textual module gain on PHOENIX-14T is more significant than that on PHOENIX-14 \\((+1.7\\% / + 1.6\\%\\) vs. \\(+0.9\\% / + 1.1\\%)\\), indicating that prior language knowledge is more beneficial for this dataset." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.685, + 0.229, + 0.701 + ], + "angle": 0, + "content": "4.3. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.708, + 0.468, + 0.739 + ], + "angle": 0, + "content": "We study the effectiveness of the proposed pretrained modules and alignment loss weights in our framework." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.757, + 0.27, + 0.77 + ], + "angle": 0, + "content": "4.3.1 Pretrained Modules" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.78, + 0.468, + 0.854 + ], + "angle": 0, + "content": "To illustrate the importance of the pretrained modules in our CVT-SLR framework, we conduct the ablation studies of both the pretrained visual module (CNN module) and the pretrained textual module (VAE module) on PHOENIX-14, as verified in Table 3." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.468, + 0.901 + ], + "angle": 0, + "content": "Overall, both pretrained modules are useful. Comparing #1 and #3, we find that pretrained visual features account for a larger determinant of performance, which is consistent" + }, + { + "type": "image", + "bbox": [ + 0.51, + 0.085, + 0.878, + 0.297 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.301, + 0.892, + 0.33 + ], + "angle": 0, + "content": "Figure 3. The training loss curves corresponding to the ablation study of pretrained modules in Table 3." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.345, + 0.892, + 0.496 + ], + "angle": 0, + "content": "with our reason for retaining the pretrained visual knowledge. Surprisingly, comparing #1 and #2, we find that without pretrained parameters of the visual module, i.e., using only the pretrained textual module, our CVT-SLR performance is also greatly improved. Notably, this is a new observation that has not been explored in previous SLR works extensively. These comparisons demonstrate that the pretrained VAE module has learned sufficient language knowledge to improve the SLR task. Comparing #3 with #4 further confirms this point." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.497, + 0.892, + 0.662 + ], + "angle": 0, + "content": "To verify the effectiveness of the pretrained modules for the training convergence speed, training loss curves corresponding to Table 3 are shown in Figure 3. Specifically, comparing #1 and #2 (or #3 and #4), configuring a pretrained textual module can promote faster convergence and reach a stable state earlier. Similarly, comparing #1 and #3 (or #2 and #4), a pretrained visual module can accelerate convergence speed compared to the pretrained textual module. As can be seen from the loss curves, the degrees of convergence are positively correlated with the final evaluation performances of WER scores." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.681, + 0.724, + 0.697 + ], + "angle": 0, + "content": "4.3.2 Alignment Loss Weights" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.704, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Training CVT-SLR to recognize glosses is optimized jointly by CTC loss and contrastive alignment loss. To explore the effect of loss weights, we conduct an ablation study of varied weights of the CTC loss and the proposed contrastive alignment loss, as shown in Table 4. For simplicity, we set the weights of CTC loss to a fixed value of 1.0. In Table 4, #1, without contrastive alignment loss (weight=0), performs worse than the models with contrastive alignment in an optimal configuration (#2, #3), which also illustrates the gain effect of the proposed contrastive alignment. As shown in #2-#5 in Table 4, we keep increasing the weight of the alignment loss from 1.0 to 1000.0. As the weight of the contrastive loss increases, the performance starts to rise, reach" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.926, + 0.519, + 0.938 + ], + "angle": 0, + "content": "23147" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.086, + 0.085, + 0.885, + 0.31 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.317, + 0.895, + 0.361 + ], + "angle": 0, + "content": "Figure 4. Four examples with cross-modal alignment matrices (left), saliency maps (middle), and generated glosses (right) on the PHOENIX-14 test set. Cross-modal alignment matrices show the alignment between visual and textual features, while saliency maps show the highest activation regions focused on gestures." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.38, + 0.473, + 0.472 + ], + "angle": 0, + "content": "ing a peak when the weight value is 10.0. As the weight increases and the value is greater than 10.0, the performance starts to decline instead. This ablation study indicates the important role of the proposed contrastive alignment. By balancing the weights between CTC loss and contrastive alignment loss, the model can achieve optimal performance." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.479, + 0.27, + 0.495 + ], + "angle": 0, + "content": "4.4. Qualitative Analysis" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.502, + 0.47, + 0.592 + ], + "angle": 0, + "content": "We randomly select four examples from the test set of PHOENIX-14 for qualitative analysis. Each example displays an alignment matrix, a series of saliency maps, and generated glosses in Figure 4. An alignment matrix shows alignment relationships between visual and textual features, while a saliency map shows the highest activations." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.593, + 0.47, + 0.713 + ], + "angle": 0, + "content": "Visualization on Cross-modal Alignment Matrices. From these four alignment matrices in Figure 4, we can see that the reference and inferred glosses are almost the same with quite low WER scores, and the performances are relatively perfect. Hence, the highlighted areas of the alignment matrices are concentrated near the diagonal, which means that our contrastive alignment algorithm plays a crucial role in aligning visual and textual features." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.713, + 0.47, + 0.835 + ], + "angle": 0, + "content": "Visualization on Saliency Maps. We visualize the key parts of the sign video frames in focus by using GradCAM [36] as shown in Figure 4. The results show that, for well-performing inferences, significant spatial features are mainly focused on the face and hand regions in the saliency maps. This observation is consistent with the sign linguistic knowledge [3, 19], where sign language conveys information mainly relying on hand shapes and facial expressions." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.845, + 0.196, + 0.861 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.47, + 0.901 + ], + "angle": 0, + "content": "In this work, we propose a novel contrastive visual-textual transformation framework for SLR, called CVT-" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.38, + 0.892, + 0.607 + ], + "angle": 0, + "content": "SLR, which further alleviates insufficient training by fully exploring the pretrained knowledge of both the visual and language modalities. We introduce prior language knowledge into the single-cue CVT-SLR for the first time via a VAE-based pretrained textual module. Furthermore, we propose new methods for cross-modal consistency constraints. One method takes advantage of the properties of the autoencoder to implicitly align the visual and textual modalities. While another method introduces a contrastive cross-modal alignment for explicit consistency. To prove the proposed CVT-SLR framework, extensive quantitative experiments and qualitative analysis are conducted, showing that our single-cue CVT-SLR not only outperforms single-cue baselines by a large margin but also surpasses SOTA multi-cue methods." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.613, + 0.892, + 0.718 + ], + "angle": 0, + "content": "Overall, we hope our CVT-SLR can inspire the SLR community to mine pretrained modules from new aspects and design a more efficient SLR framework for practical usage. In future work, we plan to introduce a large-scale pretrained language model to guide textual module learning and further explore the potential of prior contextual linguistic knowledge for SLR tasks." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.75, + 0.661, + 0.767 + ], + "angle": 0, + "content": "Acknowledgement" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.781, + 0.892, + 0.901 + ], + "angle": 0, + "content": "We thank the anonymous reviewers for their constructive and helpful reviews. This work was supported by the National Key R&D Program of China (Project No. 2022ZD0115100), the National Natural Science Foundation of China (Project No. U21A20427), the Competitive Research Fund (Project No. WU2022A009) from the Westlake Center for Synthetic Biology and Integrated Bioengineering." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.926, + 0.521, + 0.938 + ], + "angle": 0, + "content": "23148" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.116, + 0.468, + 0.143 + ], + "angle": 0, + "content": "[1] Hangbo Bao, Li Dong, and Furu Wei. Beit: Bert pre-training of image transformers. In ICLR, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.145, + 0.469, + 0.214 + ], + "angle": 0, + "content": "[2] Necati Cihan Camgoz, Simon Hadfield, Oscar Koller, Hermann Ney, and Richard Bowden. Neural sign language translation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7784-7793, 2018. 1, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.217, + 0.469, + 0.286 + ], + "angle": 0, + "content": "[3] Necati Cihan Camgoz, Oscar Koller, Simon Hadfield, and Richard Bowden. Sign language transformers: Joint end-to-end sign language recognition and translation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10023-10033, 2020. 1, 3, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.288, + 0.469, + 0.356 + ], + "angle": 0, + "content": "[4] Mathilde Caron, Ishan Misra, Julien Mairal, Priya Goyal, Piotr Bojanowski, and Armand Joulin. Unsupervised learning of visual features by contrasting cluster assignments. Advances in Neural Information Processing Systems, 33:9912-9924, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.359, + 0.469, + 0.414 + ], + "angle": 0, + "content": "[5] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International conference on machine learning, pages 1597-1607. PMLR, 2020. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.417, + 0.469, + 0.471 + ], + "angle": 0, + "content": "[6] Ting Chen, Simon Kornblith, Kevin Swersky, Mohammad Norouzi, and Geoffrey E Hinton. Big self-supervised models are strong semi-supervised learners. Advances in neural information processing systems, 33:22243-22255, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.474, + 0.47, + 0.543 + ], + "angle": 0, + "content": "[7] Yutong Chen, Fangyun Wei, Xiao Sun, Zhirong Wu, and Stephen Lin. A simple multi-modality transfer learning baseline for sign language translation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5120-5130, 2022. 3, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.546, + 0.469, + 0.6 + ], + "angle": 0, + "content": "[8] Ka Leong Cheng, Zhaoyang Yang, Qifeng Chen, and Yu-Wing Tai. Fully convolutional networks for continuous sign language recognition. In European Conference on Computer Vision, pages 697-714. Springer, 2020. 2, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.603, + 0.47, + 0.67 + ], + "angle": 0, + "content": "[9] Necati Cihan Camgoz, Simon Hadfield, Oscar Koller, and Richard Bowden. Subunets: End-to-end hand shape and continuous sign language recognition. In Proceedings of the IEEE international conference on computer vision, pages 3056-3065, 2017. 1, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.673, + 0.469, + 0.742 + ], + "angle": 0, + "content": "[10] Runpeng Cui, Hu Liu, and Changshui Zhang. Recurrent convolutional neural networks for continuous sign language recognition by staged optimization. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7361-7369, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.745, + 0.469, + 0.799 + ], + "angle": 0, + "content": "[11] Runpeng Cui, Hu Liu, and Changshui Zhang. A deep neural framework for continuous sign language recognition by iterative training. IEEE Transactions on Multimedia, 21(7):1880-1891, 2019. 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.802, + 0.469, + 0.87 + ], + "angle": 0, + "content": "[12] Cleison Correia de Amorim, David Macêdo, and Cleber Zanchettin. Spatial-temporal graph convolutional networks for sign language recognition. In International Conference on Artificial Neural Networks, pages 646-657. Springer, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[13] Zhangyang Gao, Cheng Tan, Lirong Wu, and Stan Z Li. Simvp: Simpler yet better video prediction. In Proceedings" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.116, + 0.47, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3170-3180, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.122, + 0.892, + 0.19 + ], + "angle": 0, + "content": "[14] Alex Graves, Santiago Fernández, Faustino Gomez, and Jürgen Schmidhuber. Connectionist temporal classification: labelling unsegmented sequence data with recurrent neural networks. In Proceedings of the 23rd international conference on Machine learning, pages 369-376, 2006. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.193, + 0.892, + 0.26 + ], + "angle": 0, + "content": "[15] Aiming Hao, Yuecong Min, and Xilin Chen. Self-mutual distillation learning for continuous sign language recognition. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11303-11312, 2021. 1, 2, 3, 4, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.263, + 0.892, + 0.331 + ], + "angle": 0, + "content": "[16] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9729-9738, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.334, + 0.892, + 0.388 + ], + "angle": 0, + "content": "[17] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 1, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.391, + 0.892, + 0.445 + ], + "angle": 0, + "content": "[18] Yi Ji, Chunping Liu, Shengrong Gong, and Weidong Cheng. 3d hand gesture coding for sign language learning. In 2016 International Conference on Virtual Reality and Visualization (ICVRV), pages 407-410. IEEE, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.447, + 0.892, + 0.501 + ], + "angle": 0, + "content": "[19] Suhail Muhammad Kamal, Yidong Chen, Shaozi Li, Xiaodong Shi, and Jiangbin Zheng. Technical approaches to Chinese sign language processing: a review. IEEE Access, 7:96926-96935, 2019. 1, 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.504, + 0.892, + 0.544 + ], + "angle": 0, + "content": "[20] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.547, + 0.892, + 0.574 + ], + "angle": 0, + "content": "[21] Diederik P Kingma and Max Welling. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114, 2013. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.576, + 0.892, + 0.644 + ], + "angle": 0, + "content": "[22] Oscar Koller, Necati Cihan Camgoz, Hermann Ney, and Richard Bowden. Weakly supervised learning with multistream cnn-lstm-hmms to discover sequential parallelism in sign language videos. IEEE transactions on pattern analysis and machine intelligence, 42(9):2306-2320, 2019. 1, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.647, + 0.892, + 0.701 + ], + "angle": 0, + "content": "[23] Oscar Koller, Jens Forster, and Hermann Ney. Continuous sign language recognition: Towards large vocabulary statistical recognition systems handling multiple signers. Computer Vision and Image Understanding, 141:108-125, 2015. 1, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.704, + 0.892, + 0.77 + ], + "angle": 0, + "content": "[24] Oscar Koller, Sepehr Zargaran, and Hermann Ney. Re-sign: Re-aligned end-to-end sequence modelling with deep recurrent cnn-hmms. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4297-4305, 2017. 1, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.774, + 0.892, + 0.829 + ], + "angle": 0, + "content": "[25] Siyuan Li, Zedong Wang, Zicheng Liu, Cheng Tan, Haitao Lin, Di Wu, Zhiyuan Chen, Jiangbin Zheng, and Stan Z Li. Efficient multi-order gated aggregation network. arXiv preprint arXiv:2211.03295, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.831, + 0.892, + 0.871 + ], + "angle": 0, + "content": "[26] Zhi-jie Liang, Sheng-bin Liao, and Bing-zhang Hu. 3d convolutional neural networks for dynamic sign language recognition. The Computer Journal, 61(11):1724-1736, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.874, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[27] Ilya Loshchilov and Frank Hutter. Fixing weight decay regularization in adam. 2018. 6" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.926, + 0.521, + 0.938 + ], + "angle": 0, + "content": "23149" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.161 + ], + "angle": 0, + "content": "[28] Yuecong Min, Aiming Hao, Xiujuan Chai, and Xilin Chen. Visual alignment constraint for continuous sign language recognition. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11542-11551, 2021. 2, 3, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.164, + 0.472, + 0.22 + ], + "angle": 0, + "content": "[29] Zhe Niu and Brian Mak. Stochastic fine-grained labeling of multi-state sign glosses for continuous sign language recognition. In European Conference on Computer Vision, pages 172-186. Springer, 2020. 1, 2, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.221, + 0.471, + 0.29 + ], + "angle": 0, + "content": "[30] Tse-Yu Pan, Li-Yun Lo, Chung-Wei Yeh, Jhe-Wei Li, Hou-Tim Liu, and Min-Chun Hu. Sign language recognition in complex background scene based on adaptive skin colour modelling and support vector machine. International Journal of Big Data Intelligence, 5(1-2):21-30, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.292, + 0.471, + 0.36 + ], + "angle": 0, + "content": "[31] Jinye Peng, Li Yulong, et al. Chinese sign language recognition based on gray-level co-occurrence matrix and other multi-features fusion. In 2009 4th IEEE Conference on Industrial Electronics and Applications, pages 1569–1572. IEEE, 2009. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.362, + 0.47, + 0.431 + ], + "angle": 0, + "content": "[32] Junfu Pu, Wengang Zhou, Hezhen Hu, and Houqiang Li. Boosting continuous sign language recognition via cross modality augmentation. In Proceedings of the 28th ACM International Conference on Multimedia, pages 1497-1505, 2020. 1, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.433, + 0.47, + 0.502 + ], + "angle": 0, + "content": "[33] Junfu Pu, Wengang Zhou, and Houqiang Li. Iterative alignment network for continuous sign language recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4165-4174, 2019. 1, 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.504, + 0.47, + 0.588 + ], + "angle": 0, + "content": "[34] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 1, 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.589, + 0.47, + 0.645 + ], + "angle": 0, + "content": "[35] Danilo Jimenez Rezende, Shakir Mohamed, and Daan Wierstra. Stochastic backpropagation and approximate inference in deep generative models. In International conference on machine learning, pages 1278-1286. PMLR, 2014. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.647, + 0.47, + 0.728 + ], + "angle": 0, + "content": "[36] Ramprasaath R Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra. Grad-cam: Visual explanations from deep networks via gradient-based localization. In Proceedings of the IEEE international conference on computer vision, pages 618-626, 2017. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.731, + 0.47, + 0.786 + ], + "angle": 0, + "content": "[37] Cheng Tan, Zhangyang Gao, Siyuan Li, Yongjie Xu, and Stan Z Li. Temporal attention unit: Towards efficient spatiotemporal predictive learning. arXiv preprint arXiv:2206.12126, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.788, + 0.47, + 0.844 + ], + "angle": 0, + "content": "[38] Cheng Tan, Jun Xia, Lirong Wu, and Stan Z Li. Co-learning: Learning from noisy labels with self-supervision. In Proceedings of the 29th ACM International Conference on Multimedia, pages 1405-1413, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.845, + 0.47, + 0.901 + ], + "angle": 0, + "content": "[39] Yiqi Tong, Jiangbin Zheng, Hongkang Zhu, Yidong Chen, and Xiaodong Shi. A document-level neural machine translation model with dynamic caching guided by theme-rheme information. In Proceedings of the 28th International Con-" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "ference on Computational Linguistics, pages 4385-4395, 2020.1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.122, + 0.892, + 0.178 + ], + "angle": 0, + "content": "[40] Hanjie Wang, Xiujuan Chai, Xiaopeng Hong, Guoying Zhao, and Xilin Chen. Isolated sign language recognition with grassmann covariance matrices. ACM Transactions on Accessible Computing (TACCESS), 8(4):1-21, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.179, + 0.892, + 0.248 + ], + "angle": 0, + "content": "[41] Hanjie Wang, Xiujuan Chai, Yu Zhou, and Xin Chen. Fast sign language recognition benefited from low rank approximation. In 2015 11th IEEE International Conference and Workshops on Automatic Face and Gesture Recognition (FG), volume 1, pages 1-6. IEEE, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.249, + 0.892, + 0.317 + ], + "angle": 0, + "content": "[42] Chong Wu, Jiangbin Zheng, Zhenan Feng, Houwang Zhang, Le Zhang, Jiawang Cao, and Hong Yan. Fuzzy slic: Fuzzy simple linear iterative clustering. IEEE Transactions on Circuits and Systems for Video Technology, 31(6):2114-2124, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.319, + 0.892, + 0.374 + ], + "angle": 0, + "content": "[43] Jiangbin Zheng, Yidong Chen, Chong Wu, Xiaodong Shi, and Suhail Muhammad Kamal. Enhancing neural sign language translation by highlighting the facial expression information. Neurocomputing, 464:462-472, 2021. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.375, + 0.892, + 0.432 + ], + "angle": 0, + "content": "[44] Jiangbin Zheng, Siyuan Li, Cheng Tan, Chong Wu, Yidong Chen, and Stan Z Li. Leveraging graph-based cross-modal information fusion for neural sign language translation. arXiv preprint arXiv:2211.00526, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.433, + 0.892, + 0.488 + ], + "angle": 0, + "content": "[45] Jiangbin Zheng, Yile Wang, Ge Wang, Jun Xia, Yufei Huang, Guojiang Zhao, Yue Zhang, and Stan Z Li. Using context-to-vector with graph retrofitting to improve word embeddings. arXiv preprint arXiv:2210.16848, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.489, + 0.892, + 0.559 + ], + "angle": 0, + "content": "[46] Jiangbin Zheng, Zheng Zhao, Min Chen, Jing Chen, Chong Wu, Yidong Chen, Xiaodong Shi, and Yiqi Tong. An improved sign language translation model with explainable adaptations for processing long sign sentences. Computational Intelligence and Neuroscience, 2020, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.56, + 0.892, + 0.615 + ], + "angle": 0, + "content": "[47] Hao Zhou, Wengang Zhou, and Houqiang Li. Dynamic pseudo label decoding for continuous sign language recognition. In 2019 IEEE International Conference on Multimedia and Expo (ICME), pages 1282-1287. IEEE, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.616, + 0.892, + 0.685 + ], + "angle": 0, + "content": "[48] Hao Zhou, Wengang Zhou, Yun Zhou, and Houqiang Li. Spatial-temporal multi-cue network for continuous sign language recognition. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 34, pages 13009-13016, 2020. 1, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.687, + 0.892, + 0.756 + ], + "angle": 0, + "content": "[49] Yongjie Zhuang, Bo Lv, Xinjun Sheng, and Xiangyang Zhu. Towards Chinese sign language recognition using surface electromyography and accelerometers. In 2017 24th International Conference on Mechatronics and Machine Vision in Practice (M2VIP), pages 1-5. IEEE, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.757, + 0.894, + 0.813 + ], + "angle": 0, + "content": "[50] Ronglai Zuo and Brian Mak. C2slr: Consistency-enhanced continuous sign language recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5131-5140, 2022. 1, 2, 3, 5, 6" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.813 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.927, + 0.52, + 0.938 + ], + "angle": 0, + "content": "23150" + } + ] +] \ No newline at end of file diff --git a/2023/CVT-SLR_ Contrastive Visual-Textual Transformation for Sign Language Recognition With Variational Alignment/012722bf-9201-40ff-b0cf-52287586b76f_origin.pdf b/2023/CVT-SLR_ Contrastive Visual-Textual Transformation for Sign Language Recognition With Variational Alignment/012722bf-9201-40ff-b0cf-52287586b76f_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..cfda017bf10ef599d42894976dac21b13d746c8b --- /dev/null +++ b/2023/CVT-SLR_ Contrastive Visual-Textual Transformation for Sign Language Recognition With Variational Alignment/012722bf-9201-40ff-b0cf-52287586b76f_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:11838774e670c7fcabf2bbacb9b81575b419b729ab0e4768e06d3abbf944934c +size 759252 diff --git a/2023/CVT-SLR_ Contrastive Visual-Textual Transformation for Sign Language Recognition With Variational Alignment/full.md b/2023/CVT-SLR_ Contrastive Visual-Textual Transformation for Sign Language Recognition With Variational Alignment/full.md new file mode 100644 index 0000000000000000000000000000000000000000..e13928682c856190d92380a8fa06f220fdd05a3f --- /dev/null +++ b/2023/CVT-SLR_ Contrastive Visual-Textual Transformation for Sign Language Recognition With Variational Alignment/full.md @@ -0,0 +1,326 @@ +# CVT-SLR: Contrastive Visual-Textual Transformation for Sign Language Recognition with Variational Alignment + +Jiangbin Zheng $^{1}$ , Yile Wang $^{1,2}$ , Cheng Tan $^{1}$ , Siyuan Li $^{1}$ , Ge Wang $^{1}$ , Jun Xia $^{1}$ , Yidong Chen $^{3}$ , Stan Z. Li $^{1*}$ + +$^{1}$ AI Lab, Research Center for Industries of the Future, Westlake University $^{2}$ Institute for AI Industry Research (AIR), Tsinghua University $^{3}$ School of Informatics, Xiamen University + +{zhengjiangbin, wangyile, tancheng, lisiyuan, wangge, xiajun, Stan.ZQ.Li}@westlake.edu.cn +ydchen@xmu.edu.cn + +# Abstract + +Sign language recognition (SLR) is a weakly supervised task that annotates sign videos as textual glosses. Recent studies show that insufficient training caused by the lack of large-scale available sign datasets becomes the main bottleneck for SLR. Most SLR works thereby adopt pretrained visual modules and develop two mainstream solutions. The multi-stream architectures extend multi-cue visual features, yielding the current SOTA performances but requiring complex designs and might introduce potential noise. Alternatively, the advanced single-cue SLR frameworks using explicit cross-modal alignment between visual and textual modalities are simple and effective, potentially competitive with the multi-cue framework. In this work, we propose a novel contrastive visual-textual transformation for SLR, CVT-SLR, to fully explore the pretrained knowledge of both the visual and language modalities. Based on the single-cue cross-modal alignment framework, we propose a variational autoencoder (VAE) for pretrained contextual knowledge while introducing the complete pretrained language module. The VAE implicitly aligns visual and textual modalities while benefiting from pretrained contextual knowledge as the traditional contextual module. Meanwhile, a contrastive cross-modal alignment algorithm is designed to explicitly enhance the consistency constraints. Extensive experiments on public datasets (PHOENIX-2014 and PHOENIX-2014T) demonstrate that our proposed CVT-SLR consistently outperforms existing single-cue methods and even outperforms SOTA multi-cue methods. The source codes and models are available at https://github.com/binbinjiang/CVT-SLR + +# 1. Introduction + +As a special visual natural language, sign language is the primary communication medium of the deaf community [19]. With the progress of deep learning [1, 17, 25, 39, 42], sign language recognition (SLR) has emerged as a multimodal task that aims to annotate sign videos into textual sign glosses. However, a significant dilemma of SLR is the lack of publicly available sign language datasets. For example, the most commonly-used PHOENIX-2014 [23] and PHOENIX-2014T [2] datasets only include about 10K pairs of sign videos and gloss annotations, which are far from training a robust SLR system with full supervision as typical vision-language cross-modal tasks [34]. Therefore, data limitation that may easily lead to insufficient training or overfitting problems is the main bottleneck of SLR tasks. + +The development of weakly supervised SLR has witnessed most of the improvement efforts focus on the visual module (e.g., CNN) [9, 10, 15, 29, 32, 33]. Transferring pretrained visual networks from general domains of human actions becomes a consensus to alleviate the low-resource limitation. The mainstream multi-stream SLR framework extends the pretrained visual module with multi-cue visual information [3, 22, 24, 43, 48, 50], including global features and regional features such as hands and faces in independent streams. The theoretical support for this approach comes from sign language linguistics, where sign language utilizes multiple complementary channels (e.g., hand shapes, facial expressions) to convey information [3]. The multi-cue mechanism essentially exploits hard attention to key information, yielding the current SOTA performances. However, the multi-cue framework is complex (e.g., cropping multiple regions, requiring more parameters), and the fusion of multiple streams might introduce additional potential noise. + +Another mainstream advanced solution is the single-cue + +![](images/1f9e04cf2819b24831d3b261aa24eea37fa88956f77e5bc41c72e100fbe9581f.jpg) +Figure 1. (a) An advanced single-cue SLR framework with explicit cross-modal alignment; (b) Our proposed single-cue SLR framework with explicit cross-modal alignment and implicit autoencoder alignment. Both frameworks use pretrained visual features. But our framework uses the autoencoder module to replace the mainstream contextual module, which not only includes the functions of the contextual module but also can introduce complete pretrained language knowledge and implicit cross-modal alignment. To maximize the preservation of the complete pretrained language parameters and migrated visual features, a video-gloss adapter is introduced. + +cross-modal alignment framework [15, 28], which consists of a pretrained visual module followed by a contextual module (e.g., RNN, LSTM, Transformer) and a Connectionist Temporal Classification (CTC) [14] based alignment module for gloss generation, as shown in Figure 1 (a). Explicit cross-modal alignment constraints further improve feature interactions [15, 28, 38], which could be treated as a kind of consistency between two different modalities [50] facilitating the visual module learn long-term temporal information from contextual module [13, 37]. The cross-modal alignment framework is simple and effective, potentially competitive with the multi-cue framework. Despite the advanced performance of complex multi-cue architectures with pretrained visual modules, the cross-modal consistency is a more elegant design for practical usage. It also implies the potential of prior contextual linguistic knowledge, which has been overlooked by existing SLR works. + +In this work, we propose a novel contrastive visual-textual transformation framework for SLR, called CVT-SLR, to fully explore the pretrained knowledge of both the visual and language modalities, as shown in Figure 1 (b). Based on the single-cue cross-modal alignment framework, CVT-SLR keeps the pretrained visual module but replaces the traditional contextual module with a variational autoencoder (VAE). Since a full encoder-decoder architecture is used, the VAE is responsible for learning pretrained contextual knowledge based on a pseudo-translation task while introducing the complete pretrained language module. In addition, the VAE maintains the consistency of input and output modalities due to the form of an autoencoder, playing an implicit cross-modal alignment role. Furthermore, inspired by contrastive learning [4-6,34], we introduce a contrastive + +alignment algorithm that focuses on both positive and negative samples to enhance explicit cross-modal consistency constraints. Extensive quantitative experiments conducted on the public datasets PHOENIX-2014 and PHOENIX-2014T demonstrate the advance of the proposed CVT-SLR framework. Through ablation study and qualitative analysis, we also verify the effectiveness of introducing pretrained language knowledge and the new consistency constraint mechanism. + +Our main contributions are as follows: + +- A novel visual-textual transformation-based SLR framework is proposed, which introduces fully pretrained language knowledge for the first time and provides new approaches for other cross-modal tasks. +- New alignment methods are proposed for cross-modal consistency constraints: a) exploiting the special properties of the autoencoder to implicitly align visual and textual modalities; b) introducing an explicit contrastive cross-modal alignment method. +- The proposed single-cue SLR framework not only outperforms existing single-cue baselines by a large margin but even surpasses SOTA multi-cue baselines. + +# 2. Related Work + +SLR tasks are generally divided into three categories [19]: finger-spelling recognition [18, 30, 31], isolated word recognition [26, 40, 41, 49] and continuous sign sentence recognition [8, 11, 12, 29, 33]. In the early days, SLR works mainly focused on lexical-level tasks, such as fingerspelling recognition and isolated word recognition. Nowadays, the more practical continuous sign sentence recogni + +tion task has become mainstream sign language research. In this work, the mentioned SLRs refer specifically to continuous sign sentence recognition. + +The recent SLR works [8, 11, 15, 29] are summarized based on these three aspects: feature extraction, recognition, and alignment. Mostly, the components of feature extraction are composed of the visual module and the contextual module. The visual module encodes short-term spatial information, while the contextual module encodes long-term context information. Based on the extracted features, the classifiers can get a posterior probability for recognition. And the alignment module is required to find the proper alignment between clips and glosses to ensure an accurate training procedure. + +A few studies extend the common SLR framework by incorporating multi-cue information [22,24,43,44,48]. The multi-cue features mainly include information from hand shapes, facial expressions, mouths, and poses. The current SOTA SLR works are based on multi-cue mechanisms, such as $\mathrm{C}^2\mathrm{SLR}$ [50]. Recently, some works show that explicitly enforcing the consistency between the visual and textual modules can also get comparable performance, although only single-cue features are considered. For example, VAC [28] treats the visual and textual modules as the student and teacher, respectively, and achieves knowledge distillation to align the visual and the textual modalities. Similarly, SMKD [15] achieves knowledge transfer by further sharing classifiers between different modalities. Note that SMKD is the current SOTA single-cue method. + +It is worth mentioning that neural sign language translation [2,3,7,46] is another popular sign language task. As a distinction, the neural sign language translation task aims to convert sign videos into equivalent spoken language translations, further focusing on semantic relationships and adjusting word order [2], while the SLR task only annotates sign videos into gloss sequences in the frame order. + +# 3. Methods + +As shown in Figure 2, our training pipeline is roughly divided into two steps. Step 1 is to pretrain a VAE network as the textual module. And step 2 is to transfer the off-the-shelf visual module (usually a publicly available CNN trained on Kinetics/ImageNet) and the pretrained textual module from step 1 into the CVT-SLR framework. A video-gloss adapter is used to bridge these two pretrained modules. + +# 3.1. Textual Model Pretraining + +Our VAE model is an asymmetric encoder-decoder architecture, where the encoder consists of self-attention layers, and the decoder consists of Bi-LSTM layers, as shown in the top part of Figure 2. To pretrain the VAE for our proposed SLR framework solely based on the gloss data from the PHOENIX-14 training set (excluding the development + +set and testing set), we construct a pseudo-translation task, Gloss2Gloss, which enables the VAE to learn the contextual semantic knowledge. Formally, it is assumed that the input gloss of the encoder is denoted as $G$ , and the generative gloss of the decoder is denoted as $G^{*}$ . + +Posterior Self-attention Encoder. We employ the neural approximation, and reparameterization [21, 35] techniques to approximate the posterior self-attention encoder for generating the posterior distribution $q_{\phi}(z|G)$ , where $\phi$ denotes the variational parameters of the encoder. The $q_{\phi}(z|G)$ is a diagonal Gaussian distribution $N(\mu, \mathrm{diag}(\sigma^2))$ , and two single-layer MLPs are used to parameterize the mean $\mu$ and the variance $\sigma^2$ , respectively, as: + +$$ +z = \mu + \sigma \odot \epsilon , \tag {1} +$$ + +where $\epsilon$ is a standard Gaussian noise, and $\odot$ denotes an element-wise product. + +Generative Bi-LSTM Decoder. Given the continuous latent variable $z$ , then the translation target $G^{*}$ is generated from $z$ as: + +$$ +p _ {\theta} \left(G ^ {*}, z\right) = p _ {\theta} \left(G ^ {*} \mid z\right) p _ {\theta} (z), \tag {2} +$$ + +where $\theta$ denotes the parameters of the generative Bi-LSTM decoder, and $p_{\theta}(G^{*}|z)$ is the conditional distribution that models the generation procedure estimated via the decoder. + +Variational Optimization. Both the generative model $p_{\theta}(G^{*},z)$ and the posterior inference model $q_{\phi}(z|G)$ are incorporated tightly in an end-to-end manner, and we can apply a standard back-propagation to optimize the gradient of the variational lower bound: + +$$ +\begin{array}{l} \begin{array}{l} \mathcal {L} _ {\mathrm {V A E}} (\theta , \phi ; G) = - \mathrm {K L} (q _ {\phi} (z | G) | | p _ {\theta} (z)) \\ \vdots \end{array} \tag {3} \\ + \mathbb {E} _ {q _ {\phi} (z | G)} \left[ \log p _ {\theta} \left(G ^ {*} \mid z\right) \right] \leq \log p _ {\theta} \left(G ^ {*}\right), \\ \end{array} +$$ + +where KL denotes the Kullback-Leibler divergence and the prior $p_{\theta}(z)$ denotes a standard Gaussian distribution. Practically, the first KL term of Eq. (3) can be written as: + +$$ +\mathcal {L} _ {\mathrm {K L}} = - \frac {1}{2} \left(\log \sigma^ {2} - \sigma^ {2} - \mu^ {2} + 1\right). \tag {4} +$$ + +Since our main target is a Gloss2Gloss translation task, we employ a cross-entropy loss to maximize the generative gloss probability as the second term of Eq. (3): + +$$ +\mathcal {L} _ {\text {g l o s s 2 g l o s s}} = \text {C r o s s E n t r o p y} (G, G ^ {*}). \tag {5} +$$ + +Therefore, the overall optimization objective of textual model pretraining is a joint loss of $\mathcal{L}_{\mathrm{KL}}$ and $\mathcal{L}_{\mathrm{gloss2gloss}}$ as: + +$$ +\mathcal {L} _ {\mathrm {V A E}} = \mathcal {L} _ {\mathrm {K L}} + \mathcal {L} _ {\text {g l o s s 2 g l o s s}}. \tag {6} +$$ + +# 3.2. Training CVT-SLR Framework + +# 3.2.1 Training Pipeline + +In the CVT-SLR framework, we sequentially concatenate the pretrained visual CNN module and the pretrained textual VAE module through a video-gloss adapter, as shown + +![](images/c02c1185002b9db6f4cfe6b1db04e336f1ed3f17399395a2f054100091fb0252.jpg) +Figure 2. Overview of proposed CVT-SLR framework. Our training pipeline is divided into two steps. Step 1: Pretrain the VAE module with the self-supervised gloss-to-gloss pseudo-translation task. Step 2: Train the CVT-SLR based on the VAE model from Step 1. + +in Figure 2. The pretrained ResNet18 model [17] is selected as the frame-wise feature extractor, which is trained on large-scale human action datasets. For the VAE module, although Gaussian noise is introduced in the reparameterization process in step 1, we do not add any noise after migrating the VAE to the SLR framework in step 2. In this way, the reparameterization module of the VAE degenerates into two common linear layers, i.e., $\epsilon$ is removed in Eq. (1). + +Formally, given a sign video with $T$ RGB frames $\mathcal{F} = \{\mathcal{F}_t\}_{t=1}^T \in \mathbb{R}^{T \times H \times W \times 3}$ , the visual module, which is composed of a stack of 2D-CNN layers and a global average pooling layer, first extracts spatial features $v = \{v_t\}_{t=1}^T \in \mathbb{R}^{T \times d}$ . Then the textual module further extracts temporal features $s = \{s_t\}_{t=1}^T \in \mathbb{R}^{T \times d}$ . Finally, the alignment module utilizes CTC to compute the conditional probability of the gloss label sequence $p(\mathcal{G}|\mathcal{F})$ based on temporal features $s$ , where $\mathcal{G} = \{\mathcal{G}_i\}_{i=1}^N$ is a generated sign gloss sequence with $N$ glosses. A shared classifier is used between the visual module and the textual module to generate spatial and temporal probability, respectively. And the proposed contrastive cross-modal alignment algorithm further enhances the consistency across different modalities. + +# 3.2.2 Video-Gloss Adapter + +To maximize the preservation of both pretrained visual and textual knowledge, we use a video-gloss adapter to connect spatial-temporal features, as shown in Figure 2. The videogloss adapter is simply implemented as a fully-connected MLP with two hidden layers. We take the CNN output probabilities as the input of the adapter where the final fully connected layer is initialized using the weights of the pretrained word embedding from the pretrained VAE [7]. + +Since both the input and output are in the same textual + +form during the textual model pretraining in step 1, the VAE plays an implicit cross-modal alignment role after being transferred to the CVT-SLR framework in step 2, which is close to the effect of consistency. Therefore, the videogloss adapter is essentially a mediator for activating the consistency between the spatial and temporal modalities. + +# 3.2.3 CTC Alignment + +Due to the spatial-temporal nature of sign videos, glosses have a one-to-many mapping to video frames but share the same ordering. Unlike spoken language texts, the continuous gloss annotations are chronologically consistent with the sign videos. Concretely, for an input sign video $\mathcal{F}$ and the corresponding generated gloss sequence $\mathcal{G}$ , we use CTC to compute $p(\mathcal{G}|\mathcal{F})$ by marginalizing over all possible $\mathcal{F}$ to $\mathcal{G}$ alignments: + +$$ +p (\mathcal {G} | \mathcal {F}) = \sum_ {\pi \in \mathcal {B}} p (\pi | \mathcal {F}), \tag {7} +$$ + +where $\pi$ denotes a path and $\mathcal{B}$ is the set of all viable paths corresponding to $\mathcal{G}$ . The probability $p(\pi|\mathcal{F})$ is computed by the textual module. The CTC loss is then formulated as: + +$$ +\mathcal {L} _ {\mathrm {c t c}} = - \log p (\mathcal {G} ^ {*} | \mathcal {F}), \tag {8} +$$ + +where $\mathcal{G}^*$ is the ground truth gloss sequence. + +# 3.2.4 Contrastive Alignment + +Aligning both visual and textual modalities help improve SLR [28]. To enforce alignment constraints across modalities, we propose a contrastive alignment loss. In addition, shared classifiers are also used as an improvement technique between visual and textual features [15]. + +
GroupsModelsDev (%)Test (%)Cues
DEL/INSWERDEL/INSWER
Group 1SubUNet [9]14.6/4.040.814.3/4.040.7video
Staged-Opt [9]13.7/7.339.412.2/7.538.7video
Align-iOpt [33]12.6/2.637.113.0/2.536.7video
DPD+TEM [47]9.5/3.235.69.3/3.134.5video
Re-Sign [24]-27.1-26.8video
SFL [29]7.9/6.526.27.5/6.326.8video
DNF [11]7.8/3.523.87.8/3.424.4video
FCN [8]-23.7-23.9video
VAC [28]7.9/2.521.28.4/2.622.3video
CMA [32]7.3/2.721.37.3/2.421.9video
SFL [29]10.3/4.124.910.4/3.625.3video
VL-SLT [7]-21.9-22.5video
SMKD [15]6.8/2.520.86.3/2.321.0video
Group 2DNF [11]7.3/3.323.16.7/3.322.9video+optical flow
STMC [48]7.7/3.421.17.4/2.620.7video+hand+face+pose
C2SLR [50]-20.5-20.4video+keypoints
Group 3Ours1 (w/o VAE+Contra)7.1/3.021.17.3/2.921.4video
Ours2 (w/ VAE)6.5/2.420.26.3/2.220.3video
Ours3 (w/ Contra)6.7/2.720.46.4/2.520.7video
Ours4 (w/ VAE+Contra)6.4/2.619.86.1/2.320.1video
+ +Table 1. Performance comparison (%) on PHOENIX-14 dataset. DEL/INS: deletion error and insertion error. The best results and SOTA baseline for each group are marked as bold and underlined. + +In previous SLR works, cross-modal alignment only focuses on positive samples [15, 48]. Inspired by contrastive learning [5, 16, 45], we construct both positive and negative samples in the same mini-batch and implement a contrastive cross-modal alignment method to ensure that similar features are closer while different features are farther apart. Given that the normalized spatial features from CNN as $S_{\mathrm{logits}} \in \mathbb{R}^{B \times T \times d}$ , and the normalized temporal features from VAE as $\mathcal{V}_{\mathrm{logits}} \in \mathbb{R}^{B \times T \times d}$ , where $B$ denotes the number of samples. Then we compute pair matrices firstly as: + +$$ +\mathcal {S} 2 \mathcal {V} _ {\text {p a i r}} = \mathcal {S} _ {\text {l o g i t s}} \times \mathcal {V} _ {\text {l o g i t s}} ^ {T}, \tag {9} +$$ + +$$ +\mathcal {V} 2 \mathcal {S} _ {\text {p a i r}} = \mathcal {V} _ {\log_ {1 5}} \times \mathcal {S} _ {\log_ {1 5}} ^ {T}, \tag {10} +$$ + +Taking $S2\mathcal{V}_{\mathrm{pair}} \in \mathbb{R}^{B \times B}$ pair matrix (denoted as $\mathcal{P}$ ) as an example, $\mathcal{P}[i,j](0 \leq i,j < B)$ represents the similarity value between the visual feature of the $i$ -th batch and the textual feature of $j$ -th batch. The visual and textual features from the same input instances are positive samples, and the features from other different input instances are negative samples. Hence, in the matrix $\mathcal{P}$ , the diagonal similarity values are from positive sample pairs, and the rest values are from negative sample pairs. To make the loss computation differentiable and easy, we convert the computation of positive and negative pairs into a binary classification task [34]. The value $i$ corresponding to the $i$ -th row in the matrix $\mathcal{P}$ + +is a positive label. Therefore, the corresponding labels of $\mathcal{P}$ are Labels $= \{0,1,\dots ,i,\dots ,B\}$ , then the contrastive alignment loss related to $S2\mathcal{V}_{\mathrm{pair}}$ as: + +$$ +\mathcal {L} _ {\text {a l i g n} _ {S}} = \text {C r o s s E n t r o p y} (\text {S o f t m a x} (S 2 \mathcal {V} \text {p a i r}), \text {L a b e l s}). \tag {11} +$$ + +Similarly, the $\mathcal{V}2S_{\mathrm{pair}}$ temporal-to-spatial alignment loss as: + +$$ +\mathcal {L} _ {\text {a l i g n} _ {\nu}} = \text {C r o s s E n t r o p y} (\text {S o f t m a x} (\mathcal {V} 2 S _ {\text {p a i r}}), \text {L a b e l s}). \tag {12} +$$ + +Therefore, the complete contrastive alignment loss as: + +$$ +\mathcal {L} _ {\text {a l i g n}} = \frac {1}{2} \left(\mathcal {L} _ {\text {a l i g n} _ {S}} + \mathcal {L} _ {\text {a l i g n} _ {V}}\right). \tag {13} +$$ + +Overall, the objective of CVT-SLR is jointly optimized CTC loss $\mathcal{L}_{\mathrm{ctc}}$ and contrastive alignment loss $\mathcal{L}_{\mathrm{align}}$ as: + +$$ +\mathcal {L} _ {\mathrm {C V T - S L R}} = \mathcal {L} _ {\mathrm {c t c}} + \mathcal {L} _ {\mathrm {a l i g n}}. \tag {14} +$$ + +# 4. Experiments + +# 4.1. Settings + +Datasets. We use PHOENIX-2014 [23] and PHOENIX-2014T [2] as datasets, where RGB videos and their corresponding annotations are provided. PHOENIX-2014 is a German SLR dataset about weather forecasts with a vocabulary size of 1,081, which is divided into three parts: 5,672 + +
GroupsModelsWERCues
Dev(%)Test(%)
Group 1SFL [29]25.126.1video
CNN+LSTM+HMM [22]24.526.5video
SLT [3]24.924.6video
FCN [8]23.325.1video
SMKD [15]20.822.4video
Group 2CNN+LSTM+HMM [22]24.525.4video+mouth
CNN+LSTM+HMM [22]22.124.1video+mouth+hand
SLT [3]24.624.5video+text
STMC [48]19.621.0video+hand+face+pose
C²SLR [50]20.220.4video+keypoints
Group 3Ours1 (w/o VAE+Contra)21.822.0video
Ours2 (w/ VAE)20.120.4video
Ours3 (w/ Contra)21.021.5video
Ours4 (w/ VAE+Contra)19.420.3video
+ +Table 2. Performance comparison (%) on PHOENIX14T dataset. The best results and SOTA baseline for each group are marked as **bold** and **underlined**, respectively. + +instances for training, 540 for validation, and 629 for testing, respectively. And PHOENIX-2014T is an extension of PHOENIX-2014 with a vocabulary size of 1,085, which is also divided into three parts: 7,096 instances for training, 519 for validation, and 642 for testing, respectively. + +Evaluation Metric. Word error rate (WER) is the most widely-used metric to evaluate the SLR performance, which measures the number of necessary insertions, substitutions, and deletions in the recognized sentences when compared to the reference sentences. WER is an edit distance, indicating the least number of operations of substitutions (#sub), insertions (#ins), and deletions (#del) to transform the predicted sentences into the reference sequences (#reference): + +$$ +\mathrm {W E R} = \frac {\# \mathrm {d e l} + \# \mathrm {i n s} + \# \mathrm {s u b}}{\# \text {r e f e r e n c e}}. \tag {15} +$$ + +Implementation Details. The videos are resized to $256 \times 256$ and then cropped to $224 \times 224$ . During SLR training, we use random crop and horizontal flips (50%) for data augmentation. During SLR testing, we only adopt center cropping. Beam Search is used for gloss generation. Our visual module has the same configuration as [15]. The VAE module contains two multi-head self-attention layers with 4 heads and two Bi-LSTM layers with $2 \times 512$ dimensional hidden states sequentially. Adam [20] optimizer is used to pretrain our VAE network with a batch size of 16, an initial learning rate of 1e-4. And AdamW [27] optimizer is used to train our SLR framework with a batch size of 8, a learning rate of 1e-4, and a weight decay of 1e-4. And 1 NVIDIA A100 80GB GPU is used. + +# 4.2. Main Results + +Table 1 and Table 2 show our comparisons with two-group different baselines, where Group 1 includes + +single-cue baselines and Group 2 includes multi-cue baselines. Oursnumber denote our trained CVT-SLR framework, which have only two variables, "VAE" and "Contra". Note that "w/o (w/)" VAE" denotes the VAE module in CVT-SLR framework without (with) transferred pretrained parameters, and "w/o (w/)" Contra" denotes training CVT-SLR framework without (with) contrastive alignment loss. + +Evaluation on PHOENIX-14. Table 1 shows our methods compared to other methods on PHOENIX-14. The optimalWER scores of our single-cue CVT-SLR $(\mathrm{Ours}_4)$ on the development set and testing set are $19.8\%$ and $20.1\%$ respectively. And $\mathrm{Ours_4}$ achieves the best performance among video-based only/single-cue methods in Group 1 (Ours vs. SMKD). Furthermore, while no extra cues are used, our CVT-SLR achieves the best result among the models trained with extra cues (i.e., multi-cue methods) in Group 2 (Ours vs. $\mathrm{C}^2\mathrm{SLR}$ ). + +From within-group comparisons in Group 3, we find that the case, $\mathrm{Ours}_1$ , which neither migrates the pretrained VAE parameters nor uses the contrastive alignment constraint, performs the worst. Nevertheless, relative to other groups, it still performs well. Based on $\mathrm{Ours}_1$ , by only configuring the pretrained VAE parameters $(\mathrm{Ours}_2)$ or the contrastive alignment constraints $(\mathrm{Ours}_3)$ , the CVT-SLR framework can also gain gains, and the pretrained VAE brings more obvious gains $(+0.9\% / +1.1\%$ vs. $+0.7\% / +0.7\%$ ). Overall, Group 3 demonstrates that both the prior language knowledge from the pretrained textual module and the contrastive constraints through cross-modal alignment play important enhancement roles for our SLR framework. + +Evaluation on PHOENIX-14T. In Table 2, we evaluate the CVT-SLR on PHOENIX-14T. We can see that our optimal model $(\mathrm{Ours}_4)$ in this dataset also achieves the best per + +
#Pre-VisualPre-TextualWER
Dev(%)Test(%)
1××89.288.5
2×✓30.831.4
3✓×20.420.7
4✓✓19.820.1
+ +Table 3. Ablation study of pretrained modules on PHOENIX-14 (Pre-Visual: pretrained visual module, Pre-Language: pretrained textual module). The best results are bolded. + +
#CTC WeightAlign WeightWER
Dev(%)Test(%)
11.0020.220.3
21.01.020.020.2
31.010.019.820.1
41.0100.020.120.6
51.01000.020.520.9
+ +Table 4. Ablation study of varied weights of the CTC loss and the alignment loss on PHOENIX-14. The best results are bolded. + +formance (development set: $19.4\%$ and testing set: $20.3\%$ ) with video-based only/single-cue information compared to Group 1 $(\text{Ours}_4$ vs. SMKD). It is worth mentioning that the SOTA single-cue method (SMKD) fails to surpass the multi-cue baselines in their work on PHOENIX-14T. And our CVT-SLR makes up for their shortcomings while outperforming the SOTA multi-cue baselines in Group 2 on this dataset $(\text{Ours}_4$ vs. $\mathbf{C}^2\mathbf{SLR})$ . + +Compared within Group 3 on the PHOENIX-14T, the observed conclusions are more or less the same as on the PHOENIX-14, i.e., both the pretrained textual module and the proposed contrastive alignment constraint have a beneficial effect on our CVT-SLR framework. But the pretrained textual module gain on PHOENIX-14T is more significant than that on PHOENIX-14 $(+1.7\% / + 1.6\%$ vs. $+0.9\% / + 1.1\%)$ , indicating that prior language knowledge is more beneficial for this dataset. + +# 4.3. Ablation Study + +We study the effectiveness of the proposed pretrained modules and alignment loss weights in our framework. + +# 4.3.1 Pretrained Modules + +To illustrate the importance of the pretrained modules in our CVT-SLR framework, we conduct the ablation studies of both the pretrained visual module (CNN module) and the pretrained textual module (VAE module) on PHOENIX-14, as verified in Table 3. + +Overall, both pretrained modules are useful. Comparing #1 and #3, we find that pretrained visual features account for a larger determinant of performance, which is consistent + +![](images/e44ef830258e9dfb75d6ada8135a918bcd7474f7e6d6593933ae1b12d988b997.jpg) +Figure 3. The training loss curves corresponding to the ablation study of pretrained modules in Table 3. + +with our reason for retaining the pretrained visual knowledge. Surprisingly, comparing #1 and #2, we find that without pretrained parameters of the visual module, i.e., using only the pretrained textual module, our CVT-SLR performance is also greatly improved. Notably, this is a new observation that has not been explored in previous SLR works extensively. These comparisons demonstrate that the pretrained VAE module has learned sufficient language knowledge to improve the SLR task. Comparing #3 with #4 further confirms this point. + +To verify the effectiveness of the pretrained modules for the training convergence speed, training loss curves corresponding to Table 3 are shown in Figure 3. Specifically, comparing #1 and #2 (or #3 and #4), configuring a pretrained textual module can promote faster convergence and reach a stable state earlier. Similarly, comparing #1 and #3 (or #2 and #4), a pretrained visual module can accelerate convergence speed compared to the pretrained textual module. As can be seen from the loss curves, the degrees of convergence are positively correlated with the final evaluation performances of WER scores. + +# 4.3.2 Alignment Loss Weights + +Training CVT-SLR to recognize glosses is optimized jointly by CTC loss and contrastive alignment loss. To explore the effect of loss weights, we conduct an ablation study of varied weights of the CTC loss and the proposed contrastive alignment loss, as shown in Table 4. For simplicity, we set the weights of CTC loss to a fixed value of 1.0. In Table 4, #1, without contrastive alignment loss (weight=0), performs worse than the models with contrastive alignment in an optimal configuration (#2, #3), which also illustrates the gain effect of the proposed contrastive alignment. As shown in #2-#5 in Table 4, we keep increasing the weight of the alignment loss from 1.0 to 1000.0. As the weight of the contrastive loss increases, the performance starts to rise, reach + +![](images/7a58421c59f84ac7013abff7a0f97ba68c1d80800e8eb8bbaeb29a6a51158b59.jpg) +Figure 4. Four examples with cross-modal alignment matrices (left), saliency maps (middle), and generated glosses (right) on the PHOENIX-14 test set. Cross-modal alignment matrices show the alignment between visual and textual features, while saliency maps show the highest activation regions focused on gestures. + +ing a peak when the weight value is 10.0. As the weight increases and the value is greater than 10.0, the performance starts to decline instead. This ablation study indicates the important role of the proposed contrastive alignment. By balancing the weights between CTC loss and contrastive alignment loss, the model can achieve optimal performance. + +# 4.4. Qualitative Analysis + +We randomly select four examples from the test set of PHOENIX-14 for qualitative analysis. Each example displays an alignment matrix, a series of saliency maps, and generated glosses in Figure 4. An alignment matrix shows alignment relationships between visual and textual features, while a saliency map shows the highest activations. + +Visualization on Cross-modal Alignment Matrices. From these four alignment matrices in Figure 4, we can see that the reference and inferred glosses are almost the same with quite low WER scores, and the performances are relatively perfect. Hence, the highlighted areas of the alignment matrices are concentrated near the diagonal, which means that our contrastive alignment algorithm plays a crucial role in aligning visual and textual features. + +Visualization on Saliency Maps. We visualize the key parts of the sign video frames in focus by using GradCAM [36] as shown in Figure 4. The results show that, for well-performing inferences, significant spatial features are mainly focused on the face and hand regions in the saliency maps. This observation is consistent with the sign linguistic knowledge [3, 19], where sign language conveys information mainly relying on hand shapes and facial expressions. + +# 5. Conclusion + +In this work, we propose a novel contrastive visual-textual transformation framework for SLR, called CVT- + +SLR, which further alleviates insufficient training by fully exploring the pretrained knowledge of both the visual and language modalities. We introduce prior language knowledge into the single-cue CVT-SLR for the first time via a VAE-based pretrained textual module. Furthermore, we propose new methods for cross-modal consistency constraints. One method takes advantage of the properties of the autoencoder to implicitly align the visual and textual modalities. While another method introduces a contrastive cross-modal alignment for explicit consistency. To prove the proposed CVT-SLR framework, extensive quantitative experiments and qualitative analysis are conducted, showing that our single-cue CVT-SLR not only outperforms single-cue baselines by a large margin but also surpasses SOTA multi-cue methods. + +Overall, we hope our CVT-SLR can inspire the SLR community to mine pretrained modules from new aspects and design a more efficient SLR framework for practical usage. In future work, we plan to introduce a large-scale pretrained language model to guide textual module learning and further explore the potential of prior contextual linguistic knowledge for SLR tasks. + +# Acknowledgement + +We thank the anonymous reviewers for their constructive and helpful reviews. This work was supported by the National Key R&D Program of China (Project No. 2022ZD0115100), the National Natural Science Foundation of China (Project No. U21A20427), the Competitive Research Fund (Project No. WU2022A009) from the Westlake Center for Synthetic Biology and Integrated Bioengineering. + +# References + +[1] Hangbo Bao, Li Dong, and Furu Wei. Beit: Bert pre-training of image transformers. In ICLR, 2022. 1 +[2] Necati Cihan Camgoz, Simon Hadfield, Oscar Koller, Hermann Ney, and Richard Bowden. Neural sign language translation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7784-7793, 2018. 1, 3, 5 +[3] Necati Cihan Camgoz, Oscar Koller, Simon Hadfield, and Richard Bowden. Sign language transformers: Joint end-to-end sign language recognition and translation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10023-10033, 2020. 1, 3, 6, 8 +[4] Mathilde Caron, Ishan Misra, Julien Mairal, Priya Goyal, Piotr Bojanowski, and Armand Joulin. Unsupervised learning of visual features by contrasting cluster assignments. Advances in Neural Information Processing Systems, 33:9912-9924, 2020. 2 +[5] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International conference on machine learning, pages 1597-1607. PMLR, 2020. 2, 5 +[6] Ting Chen, Simon Kornblith, Kevin Swersky, Mohammad Norouzi, and Geoffrey E Hinton. Big self-supervised models are strong semi-supervised learners. Advances in neural information processing systems, 33:22243-22255, 2020. 2 +[7] Yutong Chen, Fangyun Wei, Xiao Sun, Zhirong Wu, and Stephen Lin. A simple multi-modality transfer learning baseline for sign language translation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5120-5130, 2022. 3, 4, 5 +[8] Ka Leong Cheng, Zhaoyang Yang, Qifeng Chen, and Yu-Wing Tai. Fully convolutional networks for continuous sign language recognition. In European Conference on Computer Vision, pages 697-714. Springer, 2020. 2, 3, 5, 6 +[9] Necati Cihan Camgoz, Simon Hadfield, Oscar Koller, and Richard Bowden. Subunets: End-to-end hand shape and continuous sign language recognition. In Proceedings of the IEEE international conference on computer vision, pages 3056-3065, 2017. 1, 5 +[10] Runpeng Cui, Hu Liu, and Changshui Zhang. Recurrent convolutional neural networks for continuous sign language recognition by staged optimization. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7361-7369, 2017. 1 +[11] Runpeng Cui, Hu Liu, and Changshui Zhang. A deep neural framework for continuous sign language recognition by iterative training. IEEE Transactions on Multimedia, 21(7):1880-1891, 2019. 2, 3, 5 +[12] Cleison Correia de Amorim, David Macêdo, and Cleber Zanchettin. Spatial-temporal graph convolutional networks for sign language recognition. In International Conference on Artificial Neural Networks, pages 646-657. Springer, 2019. 2 +[13] Zhangyang Gao, Cheng Tan, Lirong Wu, and Stan Z Li. Simvp: Simpler yet better video prediction. In Proceedings + +of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3170-3180, 2022. 2 +[14] Alex Graves, Santiago Fernández, Faustino Gomez, and Jürgen Schmidhuber. Connectionist temporal classification: labelling unsegmented sequence data with recurrent neural networks. In Proceedings of the 23rd international conference on Machine learning, pages 369-376, 2006. 2 +[15] Aiming Hao, Yuecong Min, and Xilin Chen. Self-mutual distillation learning for continuous sign language recognition. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11303-11312, 2021. 1, 2, 3, 4, 5, 6 +[16] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9729-9738, 2020. 5 +[17] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 1, 4 +[18] Yi Ji, Chunping Liu, Shengrong Gong, and Weidong Cheng. 3d hand gesture coding for sign language learning. In 2016 International Conference on Virtual Reality and Visualization (ICVRV), pages 407-410. IEEE, 2016. 2 +[19] Suhail Muhammad Kamal, Yidong Chen, Shaozi Li, Xiaodong Shi, and Jiangbin Zheng. Technical approaches to Chinese sign language processing: a review. IEEE Access, 7:96926-96935, 2019. 1, 2, 8 +[20] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6 +[21] Diederik P Kingma and Max Welling. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114, 2013. 3 +[22] Oscar Koller, Necati Cihan Camgoz, Hermann Ney, and Richard Bowden. Weakly supervised learning with multistream cnn-lstm-hmms to discover sequential parallelism in sign language videos. IEEE transactions on pattern analysis and machine intelligence, 42(9):2306-2320, 2019. 1, 3, 6 +[23] Oscar Koller, Jens Forster, and Hermann Ney. Continuous sign language recognition: Towards large vocabulary statistical recognition systems handling multiple signers. Computer Vision and Image Understanding, 141:108-125, 2015. 1, 5 +[24] Oscar Koller, Sepehr Zargaran, and Hermann Ney. Re-sign: Re-aligned end-to-end sequence modelling with deep recurrent cnn-hmms. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4297-4305, 2017. 1, 3, 5 +[25] Siyuan Li, Zedong Wang, Zicheng Liu, Cheng Tan, Haitao Lin, Di Wu, Zhiyuan Chen, Jiangbin Zheng, and Stan Z Li. Efficient multi-order gated aggregation network. arXiv preprint arXiv:2211.03295, 2022. 1 +[26] Zhi-jie Liang, Sheng-bin Liao, and Bing-zhang Hu. 3d convolutional neural networks for dynamic sign language recognition. The Computer Journal, 61(11):1724-1736, 2018. 2 +[27] Ilya Loshchilov and Frank Hutter. Fixing weight decay regularization in adam. 2018. 6 + +[28] Yuecong Min, Aiming Hao, Xiujuan Chai, and Xilin Chen. Visual alignment constraint for continuous sign language recognition. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11542-11551, 2021. 2, 3, 4, 5 +[29] Zhe Niu and Brian Mak. Stochastic fine-grained labeling of multi-state sign glosses for continuous sign language recognition. In European Conference on Computer Vision, pages 172-186. Springer, 2020. 1, 2, 3, 5, 6 +[30] Tse-Yu Pan, Li-Yun Lo, Chung-Wei Yeh, Jhe-Wei Li, Hou-Tim Liu, and Min-Chun Hu. Sign language recognition in complex background scene based on adaptive skin colour modelling and support vector machine. International Journal of Big Data Intelligence, 5(1-2):21-30, 2018. 2 +[31] Jinye Peng, Li Yulong, et al. Chinese sign language recognition based on gray-level co-occurrence matrix and other multi-features fusion. In 2009 4th IEEE Conference on Industrial Electronics and Applications, pages 1569–1572. IEEE, 2009. 2 +[32] Junfu Pu, Wengang Zhou, Hezhen Hu, and Houqiang Li. Boosting continuous sign language recognition via cross modality augmentation. In Proceedings of the 28th ACM International Conference on Multimedia, pages 1497-1505, 2020. 1, 5 +[33] Junfu Pu, Wengang Zhou, and Houqiang Li. Iterative alignment network for continuous sign language recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4165-4174, 2019. 1, 2, 5 +[34] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 1, 2, 5 +[35] Danilo Jimenez Rezende, Shakir Mohamed, and Daan Wierstra. Stochastic backpropagation and approximate inference in deep generative models. In International conference on machine learning, pages 1278-1286. PMLR, 2014. 3 +[36] Ramprasaath R Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra. Grad-cam: Visual explanations from deep networks via gradient-based localization. In Proceedings of the IEEE international conference on computer vision, pages 618-626, 2017. 8 +[37] Cheng Tan, Zhangyang Gao, Siyuan Li, Yongjie Xu, and Stan Z Li. Temporal attention unit: Towards efficient spatiotemporal predictive learning. arXiv preprint arXiv:2206.12126, 2022. 2 +[38] Cheng Tan, Jun Xia, Lirong Wu, and Stan Z Li. Co-learning: Learning from noisy labels with self-supervision. In Proceedings of the 29th ACM International Conference on Multimedia, pages 1405-1413, 2021. 2 +[39] Yiqi Tong, Jiangbin Zheng, Hongkang Zhu, Yidong Chen, and Xiaodong Shi. A document-level neural machine translation model with dynamic caching guided by theme-rheme information. In Proceedings of the 28th International Con- + +ference on Computational Linguistics, pages 4385-4395, 2020.1 +[40] Hanjie Wang, Xiujuan Chai, Xiaopeng Hong, Guoying Zhao, and Xilin Chen. Isolated sign language recognition with grassmann covariance matrices. ACM Transactions on Accessible Computing (TACCESS), 8(4):1-21, 2016. 2 +[41] Hanjie Wang, Xiujuan Chai, Yu Zhou, and Xin Chen. Fast sign language recognition benefited from low rank approximation. In 2015 11th IEEE International Conference and Workshops on Automatic Face and Gesture Recognition (FG), volume 1, pages 1-6. IEEE, 2015. 2 +[42] Chong Wu, Jiangbin Zheng, Zhenan Feng, Houwang Zhang, Le Zhang, Jiawang Cao, and Hong Yan. Fuzzy slic: Fuzzy simple linear iterative clustering. IEEE Transactions on Circuits and Systems for Video Technology, 31(6):2114-2124, 2020. 1 +[43] Jiangbin Zheng, Yidong Chen, Chong Wu, Xiaodong Shi, and Suhail Muhammad Kamal. Enhancing neural sign language translation by highlighting the facial expression information. Neurocomputing, 464:462-472, 2021. 1, 3 +[44] Jiangbin Zheng, Siyuan Li, Cheng Tan, Chong Wu, Yidong Chen, and Stan Z Li. Leveraging graph-based cross-modal information fusion for neural sign language translation. arXiv preprint arXiv:2211.00526, 2022. 3 +[45] Jiangbin Zheng, Yile Wang, Ge Wang, Jun Xia, Yufei Huang, Guojiang Zhao, Yue Zhang, and Stan Z Li. Using context-to-vector with graph retrofitting to improve word embeddings. arXiv preprint arXiv:2210.16848, 2022. 5 +[46] Jiangbin Zheng, Zheng Zhao, Min Chen, Jing Chen, Chong Wu, Yidong Chen, Xiaodong Shi, and Yiqi Tong. An improved sign language translation model with explainable adaptations for processing long sign sentences. Computational Intelligence and Neuroscience, 2020, 2020. 3 +[47] Hao Zhou, Wengang Zhou, and Houqiang Li. Dynamic pseudo label decoding for continuous sign language recognition. In 2019 IEEE International Conference on Multimedia and Expo (ICME), pages 1282-1287. IEEE, 2019. 5 +[48] Hao Zhou, Wengang Zhou, Yun Zhou, and Houqiang Li. Spatial-temporal multi-cue network for continuous sign language recognition. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 34, pages 13009-13016, 2020. 1, 3, 5, 6 +[49] Yongjie Zhuang, Bo Lv, Xinjun Sheng, and Xiangyang Zhu. Towards Chinese sign language recognition using surface electromyography and accelerometers. In 2017 24th International Conference on Mechatronics and Machine Vision in Practice (M2VIP), pages 1-5. IEEE, 2017. 2 +[50] Ronglai Zuo and Brian Mak. C2slr: Consistency-enhanced continuous sign language recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5131-5140, 2022. 1, 2, 3, 5, 6 \ No newline at end of file diff --git a/2023/CVT-SLR_ Contrastive Visual-Textual Transformation for Sign Language Recognition With Variational Alignment/images.zip b/2023/CVT-SLR_ Contrastive Visual-Textual Transformation for Sign Language Recognition With Variational Alignment/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..f3b8bdc42f659d9bbb5e70c2868764122fdb5443 --- /dev/null +++ b/2023/CVT-SLR_ Contrastive Visual-Textual Transformation for Sign Language Recognition With Variational Alignment/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6494d6337403c6ee21083bcab54ff0254283c985021193c0d856d6c32039bd65 +size 540702 diff --git a/2023/CVT-SLR_ Contrastive Visual-Textual Transformation for Sign Language Recognition With Variational Alignment/layout.json b/2023/CVT-SLR_ Contrastive Visual-Textual Transformation for Sign Language Recognition With Variational Alignment/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..199a616af413eddcf6df139368b16bf01cca7432 --- /dev/null +++ b/2023/CVT-SLR_ Contrastive Visual-Textual Transformation for Sign Language Recognition With Variational Alignment/layout.json @@ -0,0 +1,8295 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 70, + 102, + 523, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 102, + 523, + 140 + ], + "spans": [ + { + "bbox": [ + 70, + 102, + 523, + 140 + ], + "type": "text", + "content": "CVT-SLR: Contrastive Visual-Textual Transformation for Sign Language Recognition with Variational Alignment" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 153, + 160, + 440, + 189 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 153, + 160, + 440, + 189 + ], + "spans": [ + { + "bbox": [ + 153, + 160, + 440, + 189 + ], + "type": "text", + "content": "Jiangbin Zheng" + }, + { + "bbox": [ + 153, + 160, + 440, + 189 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 153, + 160, + 440, + 189 + ], + "type": "text", + "content": ", Yile Wang" + }, + { + "bbox": [ + 153, + 160, + 440, + 189 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 153, + 160, + 440, + 189 + ], + "type": "text", + "content": ", Cheng Tan" + }, + { + "bbox": [ + 153, + 160, + 440, + 189 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 153, + 160, + 440, + 189 + ], + "type": "text", + "content": ", Siyuan Li" + }, + { + "bbox": [ + 153, + 160, + 440, + 189 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 153, + 160, + 440, + 189 + ], + "type": "text", + "content": ", Ge Wang" + }, + { + "bbox": [ + 153, + 160, + 440, + 189 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 153, + 160, + 440, + 189 + ], + "type": "text", + "content": ", Jun Xia" + }, + { + "bbox": [ + 153, + 160, + 440, + 189 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 153, + 160, + 440, + 189 + ], + "type": "text", + "content": ", Yidong Chen" + }, + { + "bbox": [ + 153, + 160, + 440, + 189 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 153, + 160, + 440, + 189 + ], + "type": "text", + "content": ", Stan Z. Li" + }, + { + "bbox": [ + 153, + 160, + 440, + 189 + ], + "type": "inline_equation", + "content": "^{1*}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 117, + 190, + 477, + 232 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 117, + 190, + 477, + 232 + ], + "spans": [ + { + "bbox": [ + 117, + 190, + 477, + 232 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 117, + 190, + 477, + 232 + ], + "type": "text", + "content": "AI Lab, Research Center for Industries of the Future, Westlake University " + }, + { + "bbox": [ + 117, + 190, + 477, + 232 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 117, + 190, + 477, + 232 + ], + "type": "text", + "content": "Institute for AI Industry Research (AIR), Tsinghua University " + }, + { + "bbox": [ + 117, + 190, + 477, + 232 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 117, + 190, + 477, + 232 + ], + "type": "text", + "content": "School of Informatics, Xiamen University" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 72, + 233, + 520, + 258 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 72, + 233, + 520, + 258 + ], + "spans": [ + { + "bbox": [ + 72, + 233, + 520, + 258 + ], + "type": "text", + "content": "{zhengjiangbin, wangyile, tancheng, lisiyuan, wangge, xiajun, Stan.ZQ.Li}@westlake.edu.cn \nydchen@xmu.edu.cn" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 143, + 286, + 192, + 300 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 286, + 192, + 300 + ], + "spans": [ + { + "bbox": [ + 143, + 286, + 192, + 300 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 312, + 289, + 672 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 312, + 289, + 672 + ], + "spans": [ + { + "bbox": [ + 47, + 312, + 289, + 672 + ], + "type": "text", + "content": "Sign language recognition (SLR) is a weakly supervised task that annotates sign videos as textual glosses. Recent studies show that insufficient training caused by the lack of large-scale available sign datasets becomes the main bottleneck for SLR. Most SLR works thereby adopt pretrained visual modules and develop two mainstream solutions. The multi-stream architectures extend multi-cue visual features, yielding the current SOTA performances but requiring complex designs and might introduce potential noise. Alternatively, the advanced single-cue SLR frameworks using explicit cross-modal alignment between visual and textual modalities are simple and effective, potentially competitive with the multi-cue framework. In this work, we propose a novel contrastive visual-textual transformation for SLR, CVT-SLR, to fully explore the pretrained knowledge of both the visual and language modalities. Based on the single-cue cross-modal alignment framework, we propose a variational autoencoder (VAE) for pretrained contextual knowledge while introducing the complete pretrained language module. The VAE implicitly aligns visual and textual modalities while benefiting from pretrained contextual knowledge as the traditional contextual module. Meanwhile, a contrastive cross-modal alignment algorithm is designed to explicitly enhance the consistency constraints. Extensive experiments on public datasets (PHOENIX-2014 and PHOENIX-2014T) demonstrate that our proposed CVT-SLR consistently outperforms existing single-cue methods and even outperforms SOTA multi-cue methods. The source codes and models are available at https://github.com/binbinjiang/CVT-SLR" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 307, + 286, + 386, + 300 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 286, + 386, + 300 + ], + "spans": [ + { + "bbox": [ + 307, + 286, + 386, + 300 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 311, + 547, + 478 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 311, + 547, + 478 + ], + "spans": [ + { + "bbox": [ + 304, + 311, + 547, + 478 + ], + "type": "text", + "content": "As a special visual natural language, sign language is the primary communication medium of the deaf community [19]. With the progress of deep learning [1, 17, 25, 39, 42], sign language recognition (SLR) has emerged as a multimodal task that aims to annotate sign videos into textual sign glosses. However, a significant dilemma of SLR is the lack of publicly available sign language datasets. For example, the most commonly-used PHOENIX-2014 [23] and PHOENIX-2014T [2] datasets only include about 10K pairs of sign videos and gloss annotations, which are far from training a robust SLR system with full supervision as typical vision-language cross-modal tasks [34]. Therefore, data limitation that may easily lead to insufficient training or overfitting problems is the main bottleneck of SLR tasks." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 482, + 548, + 698 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 482, + 548, + 698 + ], + "spans": [ + { + "bbox": [ + 304, + 482, + 548, + 698 + ], + "type": "text", + "content": "The development of weakly supervised SLR has witnessed most of the improvement efforts focus on the visual module (e.g., CNN) [9, 10, 15, 29, 32, 33]. Transferring pretrained visual networks from general domains of human actions becomes a consensus to alleviate the low-resource limitation. The mainstream multi-stream SLR framework extends the pretrained visual module with multi-cue visual information [3, 22, 24, 43, 48, 50], including global features and regional features such as hands and faces in independent streams. The theoretical support for this approach comes from sign language linguistics, where sign language utilizes multiple complementary channels (e.g., hand shapes, facial expressions) to convey information [3]. The multi-cue mechanism essentially exploits hard attention to key information, yielding the current SOTA performances. However, the multi-cue framework is complex (e.g., cropping multiple regions, requiring more parameters), and the fusion of multiple streams might introduce additional potential noise." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 318, + 701, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 701, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 318, + 701, + 545, + 714 + ], + "type": "text", + "content": "Another mainstream advanced solution is the single-cue" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 702, + 135, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 702, + 135, + 712 + ], + "spans": [ + { + "bbox": [ + 58, + 702, + 135, + 712 + ], + "type": "text", + "content": "*Corresponding author." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 293, + 732, + 317, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 732, + 317, + 742 + ], + "spans": [ + { + "bbox": [ + 293, + 732, + 317, + 742 + ], + "type": "text", + "content": "23141" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 71, + 74, + 521, + 248 + ], + "blocks": [ + { + "bbox": [ + 71, + 74, + 521, + 248 + ], + "lines": [ + { + "bbox": [ + 71, + 74, + 521, + 248 + ], + "spans": [ + { + "bbox": [ + 71, + 74, + 521, + 248 + ], + "type": "image", + "image_path": "1f9e04cf2819b24831d3b261aa24eea37fa88956f77e5bc41c72e100fbe9581f.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 252, + 547, + 308 + ], + "lines": [ + { + "bbox": [ + 46, + 252, + 547, + 308 + ], + "spans": [ + { + "bbox": [ + 46, + 252, + 547, + 308 + ], + "type": "text", + "content": "Figure 1. (a) An advanced single-cue SLR framework with explicit cross-modal alignment; (b) Our proposed single-cue SLR framework with explicit cross-modal alignment and implicit autoencoder alignment. Both frameworks use pretrained visual features. But our framework uses the autoencoder module to replace the mainstream contextual module, which not only includes the functions of the contextual module but also can introduce complete pretrained language knowledge and implicit cross-modal alignment. To maximize the preservation of the complete pretrained language parameters and migrated visual features, a video-gloss adapter is introduced." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 328, + 289, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 328, + 289, + 533 + ], + "spans": [ + { + "bbox": [ + 46, + 328, + 289, + 533 + ], + "type": "text", + "content": "cross-modal alignment framework [15, 28], which consists of a pretrained visual module followed by a contextual module (e.g., RNN, LSTM, Transformer) and a Connectionist Temporal Classification (CTC) [14] based alignment module for gloss generation, as shown in Figure 1 (a). Explicit cross-modal alignment constraints further improve feature interactions [15, 28, 38], which could be treated as a kind of consistency between two different modalities [50] facilitating the visual module learn long-term temporal information from contextual module [13, 37]. The cross-modal alignment framework is simple and effective, potentially competitive with the multi-cue framework. Despite the advanced performance of complex multi-cue architectures with pretrained visual modules, the cross-modal consistency is a more elegant design for practical usage. It also implies the potential of prior contextual linguistic knowledge, which has been overlooked by existing SLR works." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 534, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 534, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 534, + 289, + 715 + ], + "type": "text", + "content": "In this work, we propose a novel contrastive visual-textual transformation framework for SLR, called CVT-SLR, to fully explore the pretrained knowledge of both the visual and language modalities, as shown in Figure 1 (b). Based on the single-cue cross-modal alignment framework, CVT-SLR keeps the pretrained visual module but replaces the traditional contextual module with a variational autoencoder (VAE). Since a full encoder-decoder architecture is used, the VAE is responsible for learning pretrained contextual knowledge based on a pseudo-translation task while introducing the complete pretrained language module. In addition, the VAE maintains the consistency of input and output modalities due to the form of an autoencoder, playing an implicit cross-modal alignment role. Furthermore, inspired by contrastive learning [4-6,34], we introduce a contrastive" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 304, + 328, + 547, + 435 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 328, + 547, + 435 + ], + "spans": [ + { + "bbox": [ + 304, + 328, + 547, + 435 + ], + "type": "text", + "content": "alignment algorithm that focuses on both positive and negative samples to enhance explicit cross-modal consistency constraints. Extensive quantitative experiments conducted on the public datasets PHOENIX-2014 and PHOENIX-2014T demonstrate the advance of the proposed CVT-SLR framework. Through ablation study and qualitative analysis, we also verify the effectiveness of introducing pretrained language knowledge and the new consistency constraint mechanism." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 317, + 437, + 474, + 448 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 437, + 474, + 448 + ], + "spans": [ + { + "bbox": [ + 317, + 437, + 474, + 448 + ], + "type": "text", + "content": "Our main contributions are as follows:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 317, + 450, + 545, + 600 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 317, + 450, + 545, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 450, + 545, + 498 + ], + "spans": [ + { + "bbox": [ + 317, + 450, + 545, + 498 + ], + "type": "text", + "content": "- A novel visual-textual transformation-based SLR framework is proposed, which introduces fully pretrained language knowledge for the first time and provides new approaches for other cross-modal tasks." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 317, + 501, + 545, + 561 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 501, + 545, + 561 + ], + "spans": [ + { + "bbox": [ + 317, + 501, + 545, + 561 + ], + "type": "text", + "content": "- New alignment methods are proposed for cross-modal consistency constraints: a) exploiting the special properties of the autoencoder to implicitly align visual and textual modalities; b) introducing an explicit contrastive cross-modal alignment method." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 317, + 563, + 545, + 600 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 563, + 545, + 600 + ], + "spans": [ + { + "bbox": [ + 317, + 563, + 545, + 600 + ], + "type": "text", + "content": "- The proposed single-cue SLR framework not only outperforms existing single-cue baselines by a large margin but even surpasses SOTA multi-cue baselines." + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 609, + 392, + 621 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 609, + 392, + 621 + ], + "spans": [ + { + "bbox": [ + 306, + 609, + 392, + 621 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 629, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 629, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 629, + 547, + 715 + ], + "type": "text", + "content": "SLR tasks are generally divided into three categories [19]: finger-spelling recognition [18, 30, 31], isolated word recognition [26, 40, 41, 49] and continuous sign sentence recognition [8, 11, 12, 29, 33]. In the early days, SLR works mainly focused on lexical-level tasks, such as fingerspelling recognition and isolated word recognition. Nowadays, the more practical continuous sign sentence recogni" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 732, + 318, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 732, + 318, + 742 + ], + "spans": [ + { + "bbox": [ + 293, + 732, + 318, + 742 + ], + "type": "text", + "content": "23142" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 107 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 107 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 107 + ], + "type": "text", + "content": "tion task has become mainstream sign language research. In this work, the mentioned SLRs refer specifically to continuous sign sentence recognition." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 108, + 287, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 108, + 287, + 239 + ], + "spans": [ + { + "bbox": [ + 46, + 108, + 287, + 239 + ], + "type": "text", + "content": "The recent SLR works [8, 11, 15, 29] are summarized based on these three aspects: feature extraction, recognition, and alignment. Mostly, the components of feature extraction are composed of the visual module and the contextual module. The visual module encodes short-term spatial information, while the contextual module encodes long-term context information. Based on the extracted features, the classifiers can get a posterior probability for recognition. And the alignment module is required to find the proper alignment between clips and glosses to ensure an accurate training procedure." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 239, + 287, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 239, + 287, + 418 + ], + "spans": [ + { + "bbox": [ + 46, + 239, + 287, + 418 + ], + "type": "text", + "content": "A few studies extend the common SLR framework by incorporating multi-cue information [22,24,43,44,48]. The multi-cue features mainly include information from hand shapes, facial expressions, mouths, and poses. The current SOTA SLR works are based on multi-cue mechanisms, such as " + }, + { + "bbox": [ + 46, + 239, + 287, + 418 + ], + "type": "inline_equation", + "content": "\\mathrm{C}^2\\mathrm{SLR}" + }, + { + "bbox": [ + 46, + 239, + 287, + 418 + ], + "type": "text", + "content": " [50]. Recently, some works show that explicitly enforcing the consistency between the visual and textual modules can also get comparable performance, although only single-cue features are considered. For example, VAC [28] treats the visual and textual modules as the student and teacher, respectively, and achieves knowledge distillation to align the visual and the textual modalities. Similarly, SMKD [15] achieves knowledge transfer by further sharing classifiers between different modalities. Note that SMKD is the current SOTA single-cue method." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 419, + 287, + 503 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 419, + 287, + 503 + ], + "spans": [ + { + "bbox": [ + 46, + 419, + 287, + 503 + ], + "type": "text", + "content": "It is worth mentioning that neural sign language translation [2,3,7,46] is another popular sign language task. As a distinction, the neural sign language translation task aims to convert sign videos into equivalent spoken language translations, further focusing on semantic relationships and adjusting word order [2], while the SLR task only annotates sign videos into gloss sequences in the frame order." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 513, + 107, + 525 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 513, + 107, + 525 + ], + "spans": [ + { + "bbox": [ + 47, + 513, + 107, + 525 + ], + "type": "text", + "content": "3. Methods" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 533, + 287, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 533, + 287, + 616 + ], + "spans": [ + { + "bbox": [ + 46, + 533, + 287, + 616 + ], + "type": "text", + "content": "As shown in Figure 2, our training pipeline is roughly divided into two steps. Step 1 is to pretrain a VAE network as the textual module. And step 2 is to transfer the off-the-shelf visual module (usually a publicly available CNN trained on Kinetics/ImageNet) and the pretrained textual module from step 1 into the CVT-SLR framework. A video-gloss adapter is used to bridge these two pretrained modules." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 623, + 194, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 623, + 194, + 635 + ], + "spans": [ + { + "bbox": [ + 47, + 623, + 194, + 635 + ], + "type": "text", + "content": "3.1. Textual Model Pretraining" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 641, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 641, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 641, + 287, + 713 + ], + "type": "text", + "content": "Our VAE model is an asymmetric encoder-decoder architecture, where the encoder consists of self-attention layers, and the decoder consists of Bi-LSTM layers, as shown in the top part of Figure 2. To pretrain the VAE for our proposed SLR framework solely based on the gloss data from the PHOENIX-14 training set (excluding the development" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 72, + 545, + 132 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 132 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 132 + ], + "type": "text", + "content": "set and testing set), we construct a pseudo-translation task, Gloss2Gloss, which enables the VAE to learn the contextual semantic knowledge. Formally, it is assumed that the input gloss of the encoder is denoted as " + }, + { + "bbox": [ + 304, + 72, + 545, + 132 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 304, + 72, + 545, + 132 + ], + "type": "text", + "content": ", and the generative gloss of the decoder is denoted as " + }, + { + "bbox": [ + 304, + 72, + 545, + 132 + ], + "type": "inline_equation", + "content": "G^{*}" + }, + { + "bbox": [ + 304, + 72, + 545, + 132 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 133, + 545, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 133, + 545, + 228 + ], + "spans": [ + { + "bbox": [ + 304, + 133, + 545, + 228 + ], + "type": "text", + "content": "Posterior Self-attention Encoder. We employ the neural approximation, and reparameterization [21, 35] techniques to approximate the posterior self-attention encoder for generating the posterior distribution " + }, + { + "bbox": [ + 304, + 133, + 545, + 228 + ], + "type": "inline_equation", + "content": "q_{\\phi}(z|G)" + }, + { + "bbox": [ + 304, + 133, + 545, + 228 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 133, + 545, + 228 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 304, + 133, + 545, + 228 + ], + "type": "text", + "content": " denotes the variational parameters of the encoder. The " + }, + { + "bbox": [ + 304, + 133, + 545, + 228 + ], + "type": "inline_equation", + "content": "q_{\\phi}(z|G)" + }, + { + "bbox": [ + 304, + 133, + 545, + 228 + ], + "type": "text", + "content": " is a diagonal Gaussian distribution " + }, + { + "bbox": [ + 304, + 133, + 545, + 228 + ], + "type": "inline_equation", + "content": "N(\\mu, \\mathrm{diag}(\\sigma^2))" + }, + { + "bbox": [ + 304, + 133, + 545, + 228 + ], + "type": "text", + "content": ", and two single-layer MLPs are used to parameterize the mean " + }, + { + "bbox": [ + 304, + 133, + 545, + 228 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 304, + 133, + 545, + 228 + ], + "type": "text", + "content": " and the variance " + }, + { + "bbox": [ + 304, + 133, + 545, + 228 + ], + "type": "inline_equation", + "content": "\\sigma^2" + }, + { + "bbox": [ + 304, + 133, + 545, + 228 + ], + "type": "text", + "content": ", respectively, as:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 393, + 235, + 545, + 247 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 393, + 235, + 545, + 247 + ], + "spans": [ + { + "bbox": [ + 393, + 235, + 545, + 247 + ], + "type": "interline_equation", + "content": "z = \\mu + \\sigma \\odot \\epsilon , \\tag {1}", + "image_path": "e7c4a451d09b9ac4daae45989bb864f29ebf1126624930614bb2d132b56799c1.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 253, + 545, + 276 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 253, + 545, + 276 + ], + "spans": [ + { + "bbox": [ + 304, + 253, + 545, + 276 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 253, + 545, + 276 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 304, + 253, + 545, + 276 + ], + "type": "text", + "content": " is a standard Gaussian noise, and " + }, + { + "bbox": [ + 304, + 253, + 545, + 276 + ], + "type": "inline_equation", + "content": "\\odot" + }, + { + "bbox": [ + 304, + 253, + 545, + 276 + ], + "type": "text", + "content": " denotes an element-wise product." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 277, + 545, + 311 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 277, + 545, + 311 + ], + "spans": [ + { + "bbox": [ + 304, + 277, + 545, + 311 + ], + "type": "text", + "content": "Generative Bi-LSTM Decoder. Given the continuous latent variable " + }, + { + "bbox": [ + 304, + 277, + 545, + 311 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 304, + 277, + 545, + 311 + ], + "type": "text", + "content": ", then the translation target " + }, + { + "bbox": [ + 304, + 277, + 545, + 311 + ], + "type": "inline_equation", + "content": "G^{*}" + }, + { + "bbox": [ + 304, + 277, + 545, + 311 + ], + "type": "text", + "content": " is generated from " + }, + { + "bbox": [ + 304, + 277, + 545, + 311 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 304, + 277, + 545, + 311 + ], + "type": "text", + "content": " as:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 365, + 312, + 545, + 326 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 365, + 312, + 545, + 326 + ], + "spans": [ + { + "bbox": [ + 365, + 312, + 545, + 326 + ], + "type": "interline_equation", + "content": "p _ {\\theta} \\left(G ^ {*}, z\\right) = p _ {\\theta} \\left(G ^ {*} \\mid z\\right) p _ {\\theta} (z), \\tag {2}", + "image_path": "51097ab12f6ace184597df586febd777489275dab4cb37038b71c160e8f24fae.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 328, + 545, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 328, + 545, + 365 + ], + "spans": [ + { + "bbox": [ + 304, + 328, + 545, + 365 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 328, + 545, + 365 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 304, + 328, + 545, + 365 + ], + "type": "text", + "content": " denotes the parameters of the generative Bi-LSTM decoder, and " + }, + { + "bbox": [ + 304, + 328, + 545, + 365 + ], + "type": "inline_equation", + "content": "p_{\\theta}(G^{*}|z)" + }, + { + "bbox": [ + 304, + 328, + 545, + 365 + ], + "type": "text", + "content": " is the conditional distribution that models the generation procedure estimated via the decoder." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 365, + 545, + 423 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 365, + 545, + 423 + ], + "spans": [ + { + "bbox": [ + 304, + 365, + 545, + 423 + ], + "type": "text", + "content": "Variational Optimization. Both the generative model " + }, + { + "bbox": [ + 304, + 365, + 545, + 423 + ], + "type": "inline_equation", + "content": "p_{\\theta}(G^{*},z)" + }, + { + "bbox": [ + 304, + 365, + 545, + 423 + ], + "type": "text", + "content": " and the posterior inference model " + }, + { + "bbox": [ + 304, + 365, + 545, + 423 + ], + "type": "inline_equation", + "content": "q_{\\phi}(z|G)" + }, + { + "bbox": [ + 304, + 365, + 545, + 423 + ], + "type": "text", + "content": " are incorporated tightly in an end-to-end manner, and we can apply a standard back-propagation to optimize the gradient of the variational lower bound:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 344, + 429, + 545, + 456 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 344, + 429, + 545, + 456 + ], + "spans": [ + { + "bbox": [ + 344, + 429, + 545, + 456 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\begin{array}{l} \\mathcal {L} _ {\\mathrm {V A E}} (\\theta , \\phi ; G) = - \\mathrm {K L} (q _ {\\phi} (z | G) | | p _ {\\theta} (z)) \\\\ \\vdots \\end{array} \\tag {3} \\\\ + \\mathbb {E} _ {q _ {\\phi} (z | G)} \\left[ \\log p _ {\\theta} \\left(G ^ {*} \\mid z\\right) \\right] \\leq \\log p _ {\\theta} \\left(G ^ {*}\\right), \\\\ \\end{array}", + "image_path": "e5692d49bc8397a054d6979703ab7adb7513ce84e8236184fad4854df68d1936.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 461, + 545, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 461, + 545, + 498 + ], + "spans": [ + { + "bbox": [ + 304, + 461, + 545, + 498 + ], + "type": "text", + "content": "where KL denotes the Kullback-Leibler divergence and the prior " + }, + { + "bbox": [ + 304, + 461, + 545, + 498 + ], + "type": "inline_equation", + "content": "p_{\\theta}(z)" + }, + { + "bbox": [ + 304, + 461, + 545, + 498 + ], + "type": "text", + "content": " denotes a standard Gaussian distribution. Practically, the first KL term of Eq. (3) can be written as:" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 353, + 503, + 545, + 526 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 353, + 503, + 545, + 526 + ], + "spans": [ + { + "bbox": [ + 353, + 503, + 545, + 526 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {K L}} = - \\frac {1}{2} \\left(\\log \\sigma^ {2} - \\sigma^ {2} - \\mu^ {2} + 1\\right). \\tag {4}", + "image_path": "c939426dbee1dda619dc5fd1748e18001e5d5c45c45dcfc9c57797288bf6e002.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 530, + 545, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 530, + 545, + 567 + ], + "spans": [ + { + "bbox": [ + 304, + 530, + 545, + 567 + ], + "type": "text", + "content": "Since our main target is a Gloss2Gloss translation task, we employ a cross-entropy loss to maximize the generative gloss probability as the second term of Eq. (3):" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 353, + 573, + 545, + 586 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 353, + 573, + 545, + 586 + ], + "spans": [ + { + "bbox": [ + 353, + 573, + 545, + 586 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {g l o s s 2 g l o s s}} = \\text {C r o s s E n t r o p y} (G, G ^ {*}). \\tag {5}", + "image_path": "7dcfcb6d0730e3af2341c13139bbc5ceab077e57cb3993ee59a42155cf14159a.jpg" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 591, + 545, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 591, + 545, + 616 + ], + "spans": [ + { + "bbox": [ + 304, + 591, + 545, + 616 + ], + "type": "text", + "content": "Therefore, the overall optimization objective of textual model pretraining is a joint loss of " + }, + { + "bbox": [ + 304, + 591, + 545, + 616 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{KL}}" + }, + { + "bbox": [ + 304, + 591, + 545, + 616 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 591, + 545, + 616 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{gloss2gloss}}" + }, + { + "bbox": [ + 304, + 591, + 545, + 616 + ], + "type": "text", + "content": " as:" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 371, + 622, + 545, + 635 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 371, + 622, + 545, + 635 + ], + "spans": [ + { + "bbox": [ + 371, + 622, + 545, + 635 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {V A E}} = \\mathcal {L} _ {\\mathrm {K L}} + \\mathcal {L} _ {\\text {g l o s s 2 g l o s s}}. \\tag {6}", + "image_path": "9179a016b334e3cf1552909e9cc6e323121356f8047f1b25b8ae387f1cd154bb.jpg" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 306, + 640, + 477, + 653 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 640, + 477, + 653 + ], + "spans": [ + { + "bbox": [ + 306, + 640, + 477, + 653 + ], + "type": "text", + "content": "3.2. Training CVT-SLR Framework" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 306, + 658, + 413, + 671 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 658, + 413, + 671 + ], + "spans": [ + { + "bbox": [ + 306, + 658, + 413, + 671 + ], + "type": "text", + "content": "3.2.1 Training Pipeline" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "content": "In the CVT-SLR framework, we sequentially concatenate the pretrained visual CNN module and the pretrained textual VAE module through a video-gloss adapter, as shown" + } + ] + } + ], + "index": 25 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 733, + 317, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 733, + 317, + 742 + ], + "spans": [ + { + "bbox": [ + 293, + 733, + 317, + 742 + ], + "type": "text", + "content": "23143" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 75, + 71, + 521, + 264 + ], + "blocks": [ + { + "bbox": [ + 75, + 71, + 521, + 264 + ], + "lines": [ + { + "bbox": [ + 75, + 71, + 521, + 264 + ], + "spans": [ + { + "bbox": [ + 75, + 71, + 521, + 264 + ], + "type": "image", + "image_path": "c02c1185002b9db6f4cfe6b1db04e336f1ed3f17399395a2f054100091fb0252.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 268, + 547, + 292 + ], + "lines": [ + { + "bbox": [ + 46, + 268, + 547, + 292 + ], + "spans": [ + { + "bbox": [ + 46, + 268, + 547, + 292 + ], + "type": "text", + "content": "Figure 2. Overview of proposed CVT-SLR framework. Our training pipeline is divided into two steps. Step 1: Pretrain the VAE module with the self-supervised gloss-to-gloss pseudo-translation task. Step 2: Train the CVT-SLR based on the VAE model from Step 1." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 307, + 287, + 402 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 307, + 287, + 402 + ], + "spans": [ + { + "bbox": [ + 46, + 307, + 287, + 402 + ], + "type": "text", + "content": "in Figure 2. The pretrained ResNet18 model [17] is selected as the frame-wise feature extractor, which is trained on large-scale human action datasets. For the VAE module, although Gaussian noise is introduced in the reparameterization process in step 1, we do not add any noise after migrating the VAE to the SLR framework in step 2. In this way, the reparameterization module of the VAE degenerates into two common linear layers, i.e., " + }, + { + "bbox": [ + 46, + 307, + 287, + 402 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 46, + 307, + 287, + 402 + ], + "type": "text", + "content": " is removed in Eq. (1)." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 403, + 288, + 570 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 403, + 288, + 570 + ], + "spans": [ + { + "bbox": [ + 47, + 403, + 288, + 570 + ], + "type": "text", + "content": "Formally, given a sign video with " + }, + { + "bbox": [ + 47, + 403, + 288, + 570 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 47, + 403, + 288, + 570 + ], + "type": "text", + "content": " RGB frames " + }, + { + "bbox": [ + 47, + 403, + 288, + 570 + ], + "type": "inline_equation", + "content": "\\mathcal{F} = \\{\\mathcal{F}_t\\}_{t=1}^T \\in \\mathbb{R}^{T \\times H \\times W \\times 3}" + }, + { + "bbox": [ + 47, + 403, + 288, + 570 + ], + "type": "text", + "content": ", the visual module, which is composed of a stack of 2D-CNN layers and a global average pooling layer, first extracts spatial features " + }, + { + "bbox": [ + 47, + 403, + 288, + 570 + ], + "type": "inline_equation", + "content": "v = \\{v_t\\}_{t=1}^T \\in \\mathbb{R}^{T \\times d}" + }, + { + "bbox": [ + 47, + 403, + 288, + 570 + ], + "type": "text", + "content": ". Then the textual module further extracts temporal features " + }, + { + "bbox": [ + 47, + 403, + 288, + 570 + ], + "type": "inline_equation", + "content": "s = \\{s_t\\}_{t=1}^T \\in \\mathbb{R}^{T \\times d}" + }, + { + "bbox": [ + 47, + 403, + 288, + 570 + ], + "type": "text", + "content": ". Finally, the alignment module utilizes CTC to compute the conditional probability of the gloss label sequence " + }, + { + "bbox": [ + 47, + 403, + 288, + 570 + ], + "type": "inline_equation", + "content": "p(\\mathcal{G}|\\mathcal{F})" + }, + { + "bbox": [ + 47, + 403, + 288, + 570 + ], + "type": "text", + "content": " based on temporal features " + }, + { + "bbox": [ + 47, + 403, + 288, + 570 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 47, + 403, + 288, + 570 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 47, + 403, + 288, + 570 + ], + "type": "inline_equation", + "content": "\\mathcal{G} = \\{\\mathcal{G}_i\\}_{i=1}^N" + }, + { + "bbox": [ + 47, + 403, + 288, + 570 + ], + "type": "text", + "content": " is a generated sign gloss sequence with " + }, + { + "bbox": [ + 47, + 403, + 288, + 570 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 47, + 403, + 288, + 570 + ], + "type": "text", + "content": " glosses. A shared classifier is used between the visual module and the textual module to generate spatial and temporal probability, respectively. And the proposed contrastive cross-modal alignment algorithm further enhances the consistency across different modalities." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 586, + 170, + 598 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 586, + 170, + 598 + ], + "spans": [ + { + "bbox": [ + 47, + 586, + 170, + 598 + ], + "type": "text", + "content": "3.2.2 Video-Gloss Adapter" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 605, + 287, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 287, + 700 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 287, + 700 + ], + "type": "text", + "content": "To maximize the preservation of both pretrained visual and textual knowledge, we use a video-gloss adapter to connect spatial-temporal features, as shown in Figure 2. The videogloss adapter is simply implemented as a fully-connected MLP with two hidden layers. We take the CNN output probabilities as the input of the adapter where the final fully connected layer is initialized using the weights of the pretrained word embedding from the pretrained VAE [7]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 59, + 701, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 701, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 59, + 701, + 287, + 713 + ], + "type": "text", + "content": "Since both the input and output are in the same textual" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 307, + 545, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 307, + 545, + 380 + ], + "spans": [ + { + "bbox": [ + 304, + 307, + 545, + 380 + ], + "type": "text", + "content": "form during the textual model pretraining in step 1, the VAE plays an implicit cross-modal alignment role after being transferred to the CVT-SLR framework in step 2, which is close to the effect of consistency. Therefore, the videogloss adapter is essentially a mediator for activating the consistency between the spatial and temporal modalities." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 392, + 408, + 405 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 392, + 408, + 405 + ], + "spans": [ + { + "bbox": [ + 306, + 392, + 408, + 405 + ], + "type": "text", + "content": "3.2.3 CTC Alignment" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 411, + 545, + 507 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 411, + 545, + 507 + ], + "spans": [ + { + "bbox": [ + 304, + 411, + 545, + 507 + ], + "type": "text", + "content": "Due to the spatial-temporal nature of sign videos, glosses have a one-to-many mapping to video frames but share the same ordering. Unlike spoken language texts, the continuous gloss annotations are chronologically consistent with the sign videos. Concretely, for an input sign video " + }, + { + "bbox": [ + 304, + 411, + 545, + 507 + ], + "type": "inline_equation", + "content": "\\mathcal{F}" + }, + { + "bbox": [ + 304, + 411, + 545, + 507 + ], + "type": "text", + "content": " and the corresponding generated gloss sequence " + }, + { + "bbox": [ + 304, + 411, + 545, + 507 + ], + "type": "inline_equation", + "content": "\\mathcal{G}" + }, + { + "bbox": [ + 304, + 411, + 545, + 507 + ], + "type": "text", + "content": ", we use CTC to compute " + }, + { + "bbox": [ + 304, + 411, + 545, + 507 + ], + "type": "inline_equation", + "content": "p(\\mathcal{G}|\\mathcal{F})" + }, + { + "bbox": [ + 304, + 411, + 545, + 507 + ], + "type": "text", + "content": " by marginalizing over all possible " + }, + { + "bbox": [ + 304, + 411, + 545, + 507 + ], + "type": "inline_equation", + "content": "\\mathcal{F}" + }, + { + "bbox": [ + 304, + 411, + 545, + 507 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 304, + 411, + 545, + 507 + ], + "type": "inline_equation", + "content": "\\mathcal{G}" + }, + { + "bbox": [ + 304, + 411, + 545, + 507 + ], + "type": "text", + "content": " alignments:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 376, + 514, + 545, + 539 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 376, + 514, + 545, + 539 + ], + "spans": [ + { + "bbox": [ + 376, + 514, + 545, + 539 + ], + "type": "interline_equation", + "content": "p (\\mathcal {G} | \\mathcal {F}) = \\sum_ {\\pi \\in \\mathcal {B}} p (\\pi | \\mathcal {F}), \\tag {7}", + "image_path": "25f957dc589114db7555b52b3955e7fb516ebca20567f286945babc053ec1ff0.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 545, + 545, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 545, + 545, + 581 + ], + "spans": [ + { + "bbox": [ + 304, + 545, + 545, + 581 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 545, + 545, + 581 + ], + "type": "inline_equation", + "content": "\\pi" + }, + { + "bbox": [ + 304, + 545, + 545, + 581 + ], + "type": "text", + "content": " denotes a path and " + }, + { + "bbox": [ + 304, + 545, + 545, + 581 + ], + "type": "inline_equation", + "content": "\\mathcal{B}" + }, + { + "bbox": [ + 304, + 545, + 545, + 581 + ], + "type": "text", + "content": " is the set of all viable paths corresponding to " + }, + { + "bbox": [ + 304, + 545, + 545, + 581 + ], + "type": "inline_equation", + "content": "\\mathcal{G}" + }, + { + "bbox": [ + 304, + 545, + 545, + 581 + ], + "type": "text", + "content": ". The probability " + }, + { + "bbox": [ + 304, + 545, + 545, + 581 + ], + "type": "inline_equation", + "content": "p(\\pi|\\mathcal{F})" + }, + { + "bbox": [ + 304, + 545, + 545, + 581 + ], + "type": "text", + "content": " is computed by the textual module. The CTC loss is then formulated as:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 379, + 588, + 545, + 602 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 379, + 588, + 545, + 602 + ], + "spans": [ + { + "bbox": [ + 379, + 588, + 545, + 602 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {c t c}} = - \\log p (\\mathcal {G} ^ {*} | \\mathcal {F}), \\tag {8}", + "image_path": "28cb6e3ebe1fa5715ce5ba35e7351a6be18dd5cb71256b7c38486de18824e85d.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 609, + 487, + 621 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 609, + 487, + 621 + ], + "spans": [ + { + "bbox": [ + 306, + 609, + 487, + 621 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 306, + 609, + 487, + 621 + ], + "type": "inline_equation", + "content": "\\mathcal{G}^*" + }, + { + "bbox": [ + 306, + 609, + 487, + 621 + ], + "type": "text", + "content": " is the ground truth gloss sequence." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 635, + 436, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 635, + 436, + 647 + ], + "spans": [ + { + "bbox": [ + 306, + 635, + 436, + 647 + ], + "type": "text", + "content": "3.2.4 Contrastive Alignment" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "type": "text", + "content": "Aligning both visual and textual modalities help improve SLR [28]. To enforce alignment constraints across modalities, we propose a contrastive alignment loss. In addition, shared classifiers are also used as an improvement technique between visual and textual features [15]." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 732, + 318, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 732, + 318, + 742 + ], + "spans": [ + { + "bbox": [ + 293, + 732, + 318, + 742 + ], + "type": "text", + "content": "23144" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 80, + 70, + 512, + 356 + ], + "blocks": [ + { + "bbox": [ + 80, + 70, + 512, + 356 + ], + "lines": [ + { + "bbox": [ + 80, + 70, + 512, + 356 + ], + "spans": [ + { + "bbox": [ + 80, + 70, + 512, + 356 + ], + "type": "table", + "html": "
GroupsModelsDev (%)Test (%)Cues
DEL/INSWERDEL/INSWER
Group 1SubUNet [9]14.6/4.040.814.3/4.040.7video
Staged-Opt [9]13.7/7.339.412.2/7.538.7video
Align-iOpt [33]12.6/2.637.113.0/2.536.7video
DPD+TEM [47]9.5/3.235.69.3/3.134.5video
Re-Sign [24]-27.1-26.8video
SFL [29]7.9/6.526.27.5/6.326.8video
DNF [11]7.8/3.523.87.8/3.424.4video
FCN [8]-23.7-23.9video
VAC [28]7.9/2.521.28.4/2.622.3video
CMA [32]7.3/2.721.37.3/2.421.9video
SFL [29]10.3/4.124.910.4/3.625.3video
VL-SLT [7]-21.9-22.5video
SMKD [15]6.8/2.520.86.3/2.321.0video
Group 2DNF [11]7.3/3.323.16.7/3.322.9video+optical flow
STMC [48]7.7/3.421.17.4/2.620.7video+hand+face+pose
C2SLR [50]-20.5-20.4video+keypoints
Group 3Ours1 (w/o VAE+Contra)7.1/3.021.17.3/2.921.4video
Ours2 (w/ VAE)6.5/2.420.26.3/2.220.3video
Ours3 (w/ Contra)6.7/2.720.46.4/2.520.7video
Ours4 (w/ VAE+Contra)6.4/2.619.86.1/2.320.1video
", + "image_path": "53ea506c1a71204fae10bf2082c2d115a4d0aebe97563b77ddb4040551ef849a.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 361, + 545, + 384 + ], + "lines": [ + { + "bbox": [ + 46, + 361, + 545, + 384 + ], + "spans": [ + { + "bbox": [ + 46, + 361, + 545, + 384 + ], + "type": "text", + "content": "Table 1. Performance comparison (%) on PHOENIX-14 dataset. DEL/INS: deletion error and insertion error. The best results and SOTA baseline for each group are marked as bold and underlined." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 402, + 287, + 522 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 402, + 287, + 522 + ], + "spans": [ + { + "bbox": [ + 46, + 402, + 287, + 522 + ], + "type": "text", + "content": "In previous SLR works, cross-modal alignment only focuses on positive samples [15, 48]. Inspired by contrastive learning [5, 16, 45], we construct both positive and negative samples in the same mini-batch and implement a contrastive cross-modal alignment method to ensure that similar features are closer while different features are farther apart. Given that the normalized spatial features from CNN as " + }, + { + "bbox": [ + 46, + 402, + 287, + 522 + ], + "type": "inline_equation", + "content": "S_{\\mathrm{logits}} \\in \\mathbb{R}^{B \\times T \\times d}" + }, + { + "bbox": [ + 46, + 402, + 287, + 522 + ], + "type": "text", + "content": ", and the normalized temporal features from VAE as " + }, + { + "bbox": [ + 46, + 402, + 287, + 522 + ], + "type": "inline_equation", + "content": "\\mathcal{V}_{\\mathrm{logits}} \\in \\mathbb{R}^{B \\times T \\times d}" + }, + { + "bbox": [ + 46, + 402, + 287, + 522 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 402, + 287, + 522 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 46, + 402, + 287, + 522 + ], + "type": "text", + "content": " denotes the number of samples. Then we compute pair matrices firstly as:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 113, + 528, + 287, + 544 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 113, + 528, + 287, + 544 + ], + "spans": [ + { + "bbox": [ + 113, + 528, + 287, + 544 + ], + "type": "interline_equation", + "content": "\\mathcal {S} 2 \\mathcal {V} _ {\\text {p a i r}} = \\mathcal {S} _ {\\text {l o g i t s}} \\times \\mathcal {V} _ {\\text {l o g i t s}} ^ {T}, \\tag {9}", + "image_path": "1977a1751f55ce51e6dc7d525f1af08776436e1f9663f0a179d86d8722c462e0.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 115, + 550, + 287, + 566 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 115, + 550, + 287, + 566 + ], + "spans": [ + { + "bbox": [ + 115, + 550, + 287, + 566 + ], + "type": "interline_equation", + "content": "\\mathcal {V} 2 \\mathcal {S} _ {\\text {p a i r}} = \\mathcal {V} _ {\\log_ {1 5}} \\times \\mathcal {S} _ {\\log_ {1 5}} ^ {T}, \\tag {10}", + "image_path": "35257429f09b7862fe79a0a4953b9448f548ee770160debe4829f23954576637.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 570, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 570, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 570, + 287, + 713 + ], + "type": "text", + "content": "Taking " + }, + { + "bbox": [ + 46, + 570, + 287, + 713 + ], + "type": "inline_equation", + "content": "S2\\mathcal{V}_{\\mathrm{pair}} \\in \\mathbb{R}^{B \\times B}" + }, + { + "bbox": [ + 46, + 570, + 287, + 713 + ], + "type": "text", + "content": " pair matrix (denoted as " + }, + { + "bbox": [ + 46, + 570, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 46, + 570, + 287, + 713 + ], + "type": "text", + "content": ") as an example, " + }, + { + "bbox": [ + 46, + 570, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{P}[i,j](0 \\leq i,j < B)" + }, + { + "bbox": [ + 46, + 570, + 287, + 713 + ], + "type": "text", + "content": " represents the similarity value between the visual feature of the " + }, + { + "bbox": [ + 46, + 570, + 287, + 713 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 570, + 287, + 713 + ], + "type": "text", + "content": "-th batch and the textual feature of " + }, + { + "bbox": [ + 46, + 570, + 287, + 713 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 46, + 570, + 287, + 713 + ], + "type": "text", + "content": "-th batch. The visual and textual features from the same input instances are positive samples, and the features from other different input instances are negative samples. Hence, in the matrix " + }, + { + "bbox": [ + 46, + 570, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 46, + 570, + 287, + 713 + ], + "type": "text", + "content": ", the diagonal similarity values are from positive sample pairs, and the rest values are from negative sample pairs. To make the loss computation differentiable and easy, we convert the computation of positive and negative pairs into a binary classification task [34]. The value " + }, + { + "bbox": [ + 46, + 570, + 287, + 713 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 570, + 287, + 713 + ], + "type": "text", + "content": " corresponding to the " + }, + { + "bbox": [ + 46, + 570, + 287, + 713 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 570, + 287, + 713 + ], + "type": "text", + "content": "-th row in the matrix " + }, + { + "bbox": [ + 46, + 570, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 305, + 402, + 545, + 440 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 402, + 545, + 440 + ], + "spans": [ + { + "bbox": [ + 305, + 402, + 545, + 440 + ], + "type": "text", + "content": "is a positive label. Therefore, the corresponding labels of " + }, + { + "bbox": [ + 305, + 402, + 545, + 440 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 305, + 402, + 545, + 440 + ], + "type": "text", + "content": " are Labels " + }, + { + "bbox": [ + 305, + 402, + 545, + 440 + ], + "type": "inline_equation", + "content": "= \\{0,1,\\dots ,i,\\dots ,B\\}" + }, + { + "bbox": [ + 305, + 402, + 545, + 440 + ], + "type": "text", + "content": ", then the contrastive alignment loss related to " + }, + { + "bbox": [ + 305, + 402, + 545, + 440 + ], + "type": "inline_equation", + "content": "S2\\mathcal{V}_{\\mathrm{pair}}" + }, + { + "bbox": [ + 305, + 402, + 545, + 440 + ], + "type": "text", + "content": " as:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 307, + 445, + 545, + 460 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 445, + 545, + 460 + ], + "spans": [ + { + "bbox": [ + 307, + 445, + 545, + 460 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {a l i g n} _ {S}} = \\text {C r o s s E n t r o p y} (\\text {S o f t m a x} (S 2 \\mathcal {V} \\text {p a i r}), \\text {L a b e l s}). \\tag {11}", + "image_path": "d1faf1b3a6ad2f83836fa245fc00f87b4ffb8e296313907d4a74c8334698222d.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 466, + 545, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 466, + 545, + 479 + ], + "spans": [ + { + "bbox": [ + 306, + 466, + 545, + 479 + ], + "type": "text", + "content": "Similarly, the " + }, + { + "bbox": [ + 306, + 466, + 545, + 479 + ], + "type": "inline_equation", + "content": "\\mathcal{V}2S_{\\mathrm{pair}}" + }, + { + "bbox": [ + 306, + 466, + 545, + 479 + ], + "type": "text", + "content": " temporal-to-spatial alignment loss as:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 312, + 486, + 545, + 500 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 486, + 545, + 500 + ], + "spans": [ + { + "bbox": [ + 312, + 486, + 545, + 500 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {a l i g n} _ {\\nu}} = \\text {C r o s s E n t r o p y} (\\text {S o f t m a x} (\\mathcal {V} 2 S _ {\\text {p a i r}}), \\text {L a b e l s}). \\tag {12}", + "image_path": "b529d127c08dfc3df8178ca77dd1b0eee0cd7463c7c3b3bae153315bd8ad17fa.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 506, + 523, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 506, + 523, + 518 + ], + "spans": [ + { + "bbox": [ + 306, + 506, + 523, + 518 + ], + "type": "text", + "content": "Therefore, the complete contrastive alignment loss as:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 365, + 525, + 545, + 548 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 365, + 525, + 545, + 548 + ], + "spans": [ + { + "bbox": [ + 365, + 525, + 545, + 548 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {a l i g n}} = \\frac {1}{2} \\left(\\mathcal {L} _ {\\text {a l i g n} _ {S}} + \\mathcal {L} _ {\\text {a l i g n} _ {V}}\\right). \\tag {13}", + "image_path": "f40292264022609ccbc8e46197509040ec4ed125020226c09fbe554abce09f11.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 553, + 545, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 553, + 545, + 578 + ], + "spans": [ + { + "bbox": [ + 305, + 553, + 545, + 578 + ], + "type": "text", + "content": "Overall, the objective of CVT-SLR is jointly optimized CTC loss " + }, + { + "bbox": [ + 305, + 553, + 545, + 578 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{ctc}}" + }, + { + "bbox": [ + 305, + 553, + 545, + 578 + ], + "type": "text", + "content": " and contrastive alignment loss " + }, + { + "bbox": [ + 305, + 553, + 545, + 578 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{align}}" + }, + { + "bbox": [ + 305, + 553, + 545, + 578 + ], + "type": "text", + "content": " as:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 373, + 586, + 545, + 598 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 373, + 586, + 545, + 598 + ], + "spans": [ + { + "bbox": [ + 373, + 586, + 545, + 598 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {C V T - S L R}} = \\mathcal {L} _ {\\mathrm {c t c}} + \\mathcal {L} _ {\\mathrm {a l i g n}}. \\tag {14}", + "image_path": "863d2d1780353b980db8ad8736a9715178267b98773b501862d53cab96b1ae23.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 616, + 386, + 629 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 616, + 386, + 629 + ], + "spans": [ + { + "bbox": [ + 306, + 616, + 386, + 629 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 635, + 365, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 635, + 365, + 647 + ], + "spans": [ + { + "bbox": [ + 306, + 635, + 365, + 647 + ], + "type": "text", + "content": "4.1. Settings" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 305, + 653, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 653, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 653, + 545, + 713 + ], + "type": "text", + "content": "Datasets. We use PHOENIX-2014 [23] and PHOENIX-2014T [2] as datasets, where RGB videos and their corresponding annotations are provided. PHOENIX-2014 is a German SLR dataset about weather forecasts with a vocabulary size of 1,081, which is divided into three parts: 5,672" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 733, + 317, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 733, + 317, + 742 + ], + "spans": [ + { + "bbox": [ + 293, + 733, + 317, + 742 + ], + "type": "text", + "content": "23145" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 125, + 66, + 467, + 275 + ], + "blocks": [ + { + "bbox": [ + 125, + 66, + 467, + 275 + ], + "lines": [ + { + "bbox": [ + 125, + 66, + 467, + 275 + ], + "spans": [ + { + "bbox": [ + 125, + 66, + 467, + 275 + ], + "type": "table", + "html": "
GroupsModelsWERCues
Dev(%)Test(%)
Group 1SFL [29]25.126.1video
CNN+LSTM+HMM [22]24.526.5video
SLT [3]24.924.6video
FCN [8]23.325.1video
SMKD [15]20.822.4video
Group 2CNN+LSTM+HMM [22]24.525.4video+mouth
CNN+LSTM+HMM [22]22.124.1video+mouth+hand
SLT [3]24.624.5video+text
STMC [48]19.621.0video+hand+face+pose
C²SLR [50]20.220.4video+keypoints
Group 3Ours1 (w/o VAE+Contra)21.822.0video
Ours2 (w/ VAE)20.120.4video
Ours3 (w/ Contra)21.021.5video
Ours4 (w/ VAE+Contra)19.420.3video
", + "image_path": "7894c2879a29d10b70aea43897d2bb45125eadba8b0e94f349e987db46766525.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 277, + 547, + 301 + ], + "lines": [ + { + "bbox": [ + 46, + 277, + 547, + 301 + ], + "spans": [ + { + "bbox": [ + 46, + 277, + 547, + 301 + ], + "type": "text", + "content": "Table 2. Performance comparison (%) on PHOENIX14T dataset. The best results and SOTA baseline for each group are marked as **bold** and **underlined**, respectively." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 314, + 288, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 314, + 288, + 373 + ], + "spans": [ + { + "bbox": [ + 46, + 314, + 288, + 373 + ], + "type": "text", + "content": "instances for training, 540 for validation, and 629 for testing, respectively. And PHOENIX-2014T is an extension of PHOENIX-2014 with a vocabulary size of 1,085, which is also divided into three parts: 7,096 instances for training, 519 for validation, and 642 for testing, respectively." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 374, + 287, + 469 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 374, + 287, + 469 + ], + "spans": [ + { + "bbox": [ + 46, + 374, + 287, + 469 + ], + "type": "text", + "content": "Evaluation Metric. Word error rate (WER) is the most widely-used metric to evaluate the SLR performance, which measures the number of necessary insertions, substitutions, and deletions in the recognized sentences when compared to the reference sentences. WER is an edit distance, indicating the least number of operations of substitutions (#sub), insertions (#ins), and deletions (#del) to transform the predicted sentences into the reference sequences (#reference):" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 101, + 473, + 287, + 497 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 101, + 473, + 287, + 497 + ], + "spans": [ + { + "bbox": [ + 101, + 473, + 287, + 497 + ], + "type": "interline_equation", + "content": "\\mathrm {W E R} = \\frac {\\# \\mathrm {d e l} + \\# \\mathrm {i n s} + \\# \\mathrm {s u b}}{\\# \\text {r e f e r e n c e}}. \\tag {15}", + "image_path": "937508fa42270eea6b26023c0ae40f01ed5fde9bc8cf29c699813d34b851b947.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 498, + 288, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 498, + 288, + 665 + ], + "spans": [ + { + "bbox": [ + 46, + 498, + 288, + 665 + ], + "type": "text", + "content": "Implementation Details. The videos are resized to " + }, + { + "bbox": [ + 46, + 498, + 288, + 665 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 46, + 498, + 288, + 665 + ], + "type": "text", + "content": " and then cropped to " + }, + { + "bbox": [ + 46, + 498, + 288, + 665 + ], + "type": "inline_equation", + "content": "224 \\times 224" + }, + { + "bbox": [ + 46, + 498, + 288, + 665 + ], + "type": "text", + "content": ". During SLR training, we use random crop and horizontal flips (50%) for data augmentation. During SLR testing, we only adopt center cropping. Beam Search is used for gloss generation. Our visual module has the same configuration as [15]. The VAE module contains two multi-head self-attention layers with 4 heads and two Bi-LSTM layers with " + }, + { + "bbox": [ + 46, + 498, + 288, + 665 + ], + "type": "inline_equation", + "content": "2 \\times 512" + }, + { + "bbox": [ + 46, + 498, + 288, + 665 + ], + "type": "text", + "content": " dimensional hidden states sequentially. Adam [20] optimizer is used to pretrain our VAE network with a batch size of 16, an initial learning rate of 1e-4. And AdamW [27] optimizer is used to train our SLR framework with a batch size of 8, a learning rate of 1e-4, and a weight decay of 1e-4. And 1 NVIDIA A100 80GB GPU is used." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 671, + 133, + 682 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 671, + 133, + 682 + ], + "spans": [ + { + "bbox": [ + 47, + 671, + 133, + 682 + ], + "type": "text", + "content": "4.2. Main Results" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 689, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 689, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 689, + 288, + 715 + ], + "type": "text", + "content": "Table 1 and Table 2 show our comparisons with two-group different baselines, where Group 1 includes" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 314, + 545, + 397 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 314, + 545, + 397 + ], + "spans": [ + { + "bbox": [ + 304, + 314, + 545, + 397 + ], + "type": "text", + "content": "single-cue baselines and Group 2 includes multi-cue baselines. Oursnumber denote our trained CVT-SLR framework, which have only two variables, \"VAE\" and \"Contra\". Note that \"w/o (w/)\" VAE\" denotes the VAE module in CVT-SLR framework without (with) transferred pretrained parameters, and \"w/o (w/)\" Contra\" denotes training CVT-SLR framework without (with) contrastive alignment loss." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 399, + 546, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 399, + 546, + 518 + ], + "spans": [ + { + "bbox": [ + 304, + 399, + 546, + 518 + ], + "type": "text", + "content": "Evaluation on PHOENIX-14. Table 1 shows our methods compared to other methods on PHOENIX-14. The optimalWER scores of our single-cue CVT-SLR " + }, + { + "bbox": [ + 304, + 399, + 546, + 518 + ], + "type": "inline_equation", + "content": "(\\mathrm{Ours}_4)" + }, + { + "bbox": [ + 304, + 399, + 546, + 518 + ], + "type": "text", + "content": " on the development set and testing set are " + }, + { + "bbox": [ + 304, + 399, + 546, + 518 + ], + "type": "inline_equation", + "content": "19.8\\%" + }, + { + "bbox": [ + 304, + 399, + 546, + 518 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 399, + 546, + 518 + ], + "type": "inline_equation", + "content": "20.1\\%" + }, + { + "bbox": [ + 304, + 399, + 546, + 518 + ], + "type": "text", + "content": " respectively. And " + }, + { + "bbox": [ + 304, + 399, + 546, + 518 + ], + "type": "inline_equation", + "content": "\\mathrm{Ours_4}" + }, + { + "bbox": [ + 304, + 399, + 546, + 518 + ], + "type": "text", + "content": " achieves the best performance among video-based only/single-cue methods in Group 1 (Ours vs. SMKD). Furthermore, while no extra cues are used, our CVT-SLR achieves the best result among the models trained with extra cues (i.e., multi-cue methods) in Group 2 (Ours vs. " + }, + { + "bbox": [ + 304, + 399, + 546, + 518 + ], + "type": "inline_equation", + "content": "\\mathrm{C}^2\\mathrm{SLR}" + }, + { + "bbox": [ + 304, + 399, + 546, + 518 + ], + "type": "text", + "content": ")." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 520, + 547, + 674 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 520, + 547, + 674 + ], + "spans": [ + { + "bbox": [ + 304, + 520, + 547, + 674 + ], + "type": "text", + "content": "From within-group comparisons in Group 3, we find that the case, " + }, + { + "bbox": [ + 304, + 520, + 547, + 674 + ], + "type": "inline_equation", + "content": "\\mathrm{Ours}_1" + }, + { + "bbox": [ + 304, + 520, + 547, + 674 + ], + "type": "text", + "content": ", which neither migrates the pretrained VAE parameters nor uses the contrastive alignment constraint, performs the worst. Nevertheless, relative to other groups, it still performs well. Based on " + }, + { + "bbox": [ + 304, + 520, + 547, + 674 + ], + "type": "inline_equation", + "content": "\\mathrm{Ours}_1" + }, + { + "bbox": [ + 304, + 520, + 547, + 674 + ], + "type": "text", + "content": ", by only configuring the pretrained VAE parameters " + }, + { + "bbox": [ + 304, + 520, + 547, + 674 + ], + "type": "inline_equation", + "content": "(\\mathrm{Ours}_2)" + }, + { + "bbox": [ + 304, + 520, + 547, + 674 + ], + "type": "text", + "content": " or the contrastive alignment constraints " + }, + { + "bbox": [ + 304, + 520, + 547, + 674 + ], + "type": "inline_equation", + "content": "(\\mathrm{Ours}_3)" + }, + { + "bbox": [ + 304, + 520, + 547, + 674 + ], + "type": "text", + "content": ", the CVT-SLR framework can also gain gains, and the pretrained VAE brings more obvious gains " + }, + { + "bbox": [ + 304, + 520, + 547, + 674 + ], + "type": "inline_equation", + "content": "(+0.9\\% / +1.1\\%" + }, + { + "bbox": [ + 304, + 520, + 547, + 674 + ], + "type": "text", + "content": " vs. " + }, + { + "bbox": [ + 304, + 520, + 547, + 674 + ], + "type": "inline_equation", + "content": "+0.7\\% / +0.7\\%" + }, + { + "bbox": [ + 304, + 520, + 547, + 674 + ], + "type": "text", + "content": "). Overall, Group 3 demonstrates that both the prior language knowledge from the pretrained textual module and the contrastive constraints through cross-modal alignment play important enhancement roles for our SLR framework." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 677, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 547, + 713 + ], + "type": "text", + "content": "Evaluation on PHOENIX-14T. In Table 2, we evaluate the CVT-SLR on PHOENIX-14T. We can see that our optimal model " + }, + { + "bbox": [ + 304, + 677, + 547, + 713 + ], + "type": "inline_equation", + "content": "(\\mathrm{Ours}_4)" + }, + { + "bbox": [ + 304, + 677, + 547, + 713 + ], + "type": "text", + "content": " in this dataset also achieves the best per" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 733, + 318, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 733, + 318, + 742 + ], + "spans": [ + { + "bbox": [ + 293, + 733, + 318, + 742 + ], + "type": "text", + "content": "23146" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 57, + 68, + 276, + 153 + ], + "blocks": [ + { + "bbox": [ + 57, + 68, + 276, + 153 + ], + "lines": [ + { + "bbox": [ + 57, + 68, + 276, + 153 + ], + "spans": [ + { + "bbox": [ + 57, + 68, + 276, + 153 + ], + "type": "table", + "html": "
#Pre-VisualPre-TextualWER
Dev(%)Test(%)
1××89.288.5
2×✓30.831.4
3✓×20.420.7
4✓✓19.820.1
", + "image_path": "6f57670ce6478fdea0424f69a8d7b7432835ae4f2a0c08da4452fc8cb0e0f6d6.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 56, + 199, + 276, + 291 + ], + "blocks": [ + { + "bbox": [ + 47, + 158, + 287, + 191 + ], + "lines": [ + { + "bbox": [ + 47, + 158, + 287, + 191 + ], + "spans": [ + { + "bbox": [ + 47, + 158, + 287, + 191 + ], + "type": "text", + "content": "Table 3. Ablation study of pretrained modules on PHOENIX-14 (Pre-Visual: pretrained visual module, Pre-Language: pretrained textual module). The best results are bolded." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 199, + 276, + 291 + ], + "lines": [ + { + "bbox": [ + 56, + 199, + 276, + 291 + ], + "spans": [ + { + "bbox": [ + 56, + 199, + 276, + 291 + ], + "type": "table", + "html": "
#CTC WeightAlign WeightWER
Dev(%)Test(%)
11.0020.220.3
21.01.020.020.2
31.010.019.820.1
41.0100.020.120.6
51.01000.020.520.9
", + "image_path": "66ef0c1f1f733154a12f660b3537c8a8dd5f079daf0693f882196a3429e6aa71.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 297, + 286, + 319 + ], + "lines": [ + { + "bbox": [ + 47, + 297, + 286, + 319 + ], + "spans": [ + { + "bbox": [ + 47, + 297, + 286, + 319 + ], + "type": "text", + "content": "Table 4. Ablation study of varied weights of the CTC loss and the alignment loss on PHOENIX-14. The best results are bolded." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 331, + 286, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 331, + 286, + 426 + ], + "spans": [ + { + "bbox": [ + 46, + 331, + 286, + 426 + ], + "type": "text", + "content": "formance (development set: " + }, + { + "bbox": [ + 46, + 331, + 286, + 426 + ], + "type": "inline_equation", + "content": "19.4\\%" + }, + { + "bbox": [ + 46, + 331, + 286, + 426 + ], + "type": "text", + "content": " and testing set: " + }, + { + "bbox": [ + 46, + 331, + 286, + 426 + ], + "type": "inline_equation", + "content": "20.3\\%" + }, + { + "bbox": [ + 46, + 331, + 286, + 426 + ], + "type": "text", + "content": ") with video-based only/single-cue information compared to Group 1 " + }, + { + "bbox": [ + 46, + 331, + 286, + 426 + ], + "type": "inline_equation", + "content": "(\\text{Ours}_4" + }, + { + "bbox": [ + 46, + 331, + 286, + 426 + ], + "type": "text", + "content": " vs. SMKD). It is worth mentioning that the SOTA single-cue method (SMKD) fails to surpass the multi-cue baselines in their work on PHOENIX-14T. And our CVT-SLR makes up for their shortcomings while outperforming the SOTA multi-cue baselines in Group 2 on this dataset " + }, + { + "bbox": [ + 46, + 331, + 286, + 426 + ], + "type": "inline_equation", + "content": "(\\text{Ours}_4" + }, + { + "bbox": [ + 46, + 331, + 286, + 426 + ], + "type": "text", + "content": " vs. " + }, + { + "bbox": [ + 46, + 331, + 286, + 426 + ], + "type": "inline_equation", + "content": "\\mathbf{C}^2\\mathbf{SLR})" + }, + { + "bbox": [ + 46, + 331, + 286, + 426 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 427, + 286, + 534 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 427, + 286, + 534 + ], + "spans": [ + { + "bbox": [ + 46, + 427, + 286, + 534 + ], + "type": "text", + "content": "Compared within Group 3 on the PHOENIX-14T, the observed conclusions are more or less the same as on the PHOENIX-14, i.e., both the pretrained textual module and the proposed contrastive alignment constraint have a beneficial effect on our CVT-SLR framework. But the pretrained textual module gain on PHOENIX-14T is more significant than that on PHOENIX-14 " + }, + { + "bbox": [ + 46, + 427, + 286, + 534 + ], + "type": "inline_equation", + "content": "(+1.7\\% / + 1.6\\%" + }, + { + "bbox": [ + 46, + 427, + 286, + 534 + ], + "type": "text", + "content": " vs. " + }, + { + "bbox": [ + 46, + 427, + 286, + 534 + ], + "type": "inline_equation", + "content": "+0.9\\% / + 1.1\\%)" + }, + { + "bbox": [ + 46, + 427, + 286, + 534 + ], + "type": "text", + "content": ", indicating that prior language knowledge is more beneficial for this dataset." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 542, + 140, + 555 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 542, + 140, + 555 + ], + "spans": [ + { + "bbox": [ + 47, + 542, + 140, + 555 + ], + "type": "text", + "content": "4.3. Ablation Study" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 560, + 286, + 585 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 560, + 286, + 585 + ], + "spans": [ + { + "bbox": [ + 47, + 560, + 286, + 585 + ], + "type": "text", + "content": "We study the effectiveness of the proposed pretrained modules and alignment loss weights in our framework." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 599, + 165, + 609 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 599, + 165, + 609 + ], + "spans": [ + { + "bbox": [ + 47, + 599, + 165, + 609 + ], + "type": "text", + "content": "4.3.1 Pretrained Modules" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 617, + 286, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 617, + 286, + 676 + ], + "spans": [ + { + "bbox": [ + 46, + 617, + 286, + 676 + ], + "type": "text", + "content": "To illustrate the importance of the pretrained modules in our CVT-SLR framework, we conduct the ablation studies of both the pretrained visual module (CNN module) and the pretrained textual module (VAE module) on PHOENIX-14, as verified in Table 3." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 677, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 286, + 713 + ], + "type": "text", + "content": "Overall, both pretrained modules are useful. Comparing #1 and #3, we find that pretrained visual features account for a larger determinant of performance, which is consistent" + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 312, + 67, + 537, + 235 + ], + "blocks": [ + { + "bbox": [ + 312, + 67, + 537, + 235 + ], + "lines": [ + { + "bbox": [ + 312, + 67, + 537, + 235 + ], + "spans": [ + { + "bbox": [ + 312, + 67, + 537, + 235 + ], + "type": "image", + "image_path": "e44ef830258e9dfb75d6ada8135a918bcd7474f7e6d6593933ae1b12d988b997.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 238, + 545, + 261 + ], + "lines": [ + { + "bbox": [ + 305, + 238, + 545, + 261 + ], + "spans": [ + { + "bbox": [ + 305, + 238, + 545, + 261 + ], + "type": "text", + "content": "Figure 3. The training loss curves corresponding to the ablation study of pretrained modules in Table 3." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 273, + 545, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 273, + 545, + 392 + ], + "spans": [ + { + "bbox": [ + 304, + 273, + 545, + 392 + ], + "type": "text", + "content": "with our reason for retaining the pretrained visual knowledge. Surprisingly, comparing #1 and #2, we find that without pretrained parameters of the visual module, i.e., using only the pretrained textual module, our CVT-SLR performance is also greatly improved. Notably, this is a new observation that has not been explored in previous SLR works extensively. These comparisons demonstrate that the pretrained VAE module has learned sufficient language knowledge to improve the SLR task. Comparing #3 with #4 further confirms this point." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 393, + 545, + 524 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 393, + 545, + 524 + ], + "spans": [ + { + "bbox": [ + 304, + 393, + 545, + 524 + ], + "type": "text", + "content": "To verify the effectiveness of the pretrained modules for the training convergence speed, training loss curves corresponding to Table 3 are shown in Figure 3. Specifically, comparing #1 and #2 (or #3 and #4), configuring a pretrained textual module can promote faster convergence and reach a stable state earlier. Similarly, comparing #1 and #3 (or #2 and #4), a pretrained visual module can accelerate convergence speed compared to the pretrained textual module. As can be seen from the loss curves, the degrees of convergence are positively correlated with the final evaluation performances of WER scores." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 539, + 443, + 552 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 539, + 443, + 552 + ], + "spans": [ + { + "bbox": [ + 306, + 539, + 443, + 552 + ], + "type": "text", + "content": "4.3.2 Alignment Loss Weights" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 557, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 557, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 557, + 545, + 713 + ], + "type": "text", + "content": "Training CVT-SLR to recognize glosses is optimized jointly by CTC loss and contrastive alignment loss. To explore the effect of loss weights, we conduct an ablation study of varied weights of the CTC loss and the proposed contrastive alignment loss, as shown in Table 4. For simplicity, we set the weights of CTC loss to a fixed value of 1.0. In Table 4, #1, without contrastive alignment loss (weight=0), performs worse than the models with contrastive alignment in an optimal configuration (#2, #3), which also illustrates the gain effect of the proposed contrastive alignment. As shown in #2-#5 in Table 4, we keep increasing the weight of the alignment loss from 1.0 to 1000.0. As the weight of the contrastive loss increases, the performance starts to rise, reach" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 733, + 317, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 733, + 317, + 742 + ], + "spans": [ + { + "bbox": [ + 293, + 733, + 317, + 742 + ], + "type": "text", + "content": "23147" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 52, + 67, + 541, + 245 + ], + "blocks": [ + { + "bbox": [ + 52, + 67, + 541, + 245 + ], + "lines": [ + { + "bbox": [ + 52, + 67, + 541, + 245 + ], + "spans": [ + { + "bbox": [ + 52, + 67, + 541, + 245 + ], + "type": "image", + "image_path": "7a58421c59f84ac7013abff7a0f97ba68c1d80800e8eb8bbaeb29a6a51158b59.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 251, + 547, + 285 + ], + "lines": [ + { + "bbox": [ + 46, + 251, + 547, + 285 + ], + "spans": [ + { + "bbox": [ + 46, + 251, + 547, + 285 + ], + "type": "text", + "content": "Figure 4. Four examples with cross-modal alignment matrices (left), saliency maps (middle), and generated glosses (right) on the PHOENIX-14 test set. Cross-modal alignment matrices show the alignment between visual and textual features, while saliency maps show the highest activation regions focused on gestures." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 300, + 289, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 300, + 289, + 373 + ], + "spans": [ + { + "bbox": [ + 46, + 300, + 289, + 373 + ], + "type": "text", + "content": "ing a peak when the weight value is 10.0. As the weight increases and the value is greater than 10.0, the performance starts to decline instead. This ablation study indicates the important role of the proposed contrastive alignment. By balancing the weights between CTC loss and contrastive alignment loss, the model can achieve optimal performance." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 379, + 165, + 392 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 379, + 165, + 392 + ], + "spans": [ + { + "bbox": [ + 47, + 379, + 165, + 392 + ], + "type": "text", + "content": "4.4. Qualitative Analysis" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 397, + 287, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 397, + 287, + 468 + ], + "spans": [ + { + "bbox": [ + 46, + 397, + 287, + 468 + ], + "type": "text", + "content": "We randomly select four examples from the test set of PHOENIX-14 for qualitative analysis. Each example displays an alignment matrix, a series of saliency maps, and generated glosses in Figure 4. An alignment matrix shows alignment relationships between visual and textual features, while a saliency map shows the highest activations." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 469, + 287, + 564 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 469, + 287, + 564 + ], + "spans": [ + { + "bbox": [ + 46, + 469, + 287, + 564 + ], + "type": "text", + "content": "Visualization on Cross-modal Alignment Matrices. From these four alignment matrices in Figure 4, we can see that the reference and inferred glosses are almost the same with quite low WER scores, and the performances are relatively perfect. Hence, the highlighted areas of the alignment matrices are concentrated near the diagonal, which means that our contrastive alignment algorithm plays a crucial role in aligning visual and textual features." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 564, + 287, + 661 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 564, + 287, + 661 + ], + "spans": [ + { + "bbox": [ + 46, + 564, + 287, + 661 + ], + "type": "text", + "content": "Visualization on Saliency Maps. We visualize the key parts of the sign video frames in focus by using GradCAM [36] as shown in Figure 4. The results show that, for well-performing inferences, significant spatial features are mainly focused on the face and hand regions in the saliency maps. This observation is consistent with the sign linguistic knowledge [3, 19], where sign language conveys information mainly relying on hand shapes and facial expressions." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 669, + 119, + 681 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 669, + 119, + 681 + ], + "spans": [ + { + "bbox": [ + 47, + 669, + 119, + 681 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "content": "In this work, we propose a novel contrastive visual-textual transformation framework for SLR, called CVT-" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 300, + 545, + 480 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 300, + 545, + 480 + ], + "spans": [ + { + "bbox": [ + 304, + 300, + 545, + 480 + ], + "type": "text", + "content": "SLR, which further alleviates insufficient training by fully exploring the pretrained knowledge of both the visual and language modalities. We introduce prior language knowledge into the single-cue CVT-SLR for the first time via a VAE-based pretrained textual module. Furthermore, we propose new methods for cross-modal consistency constraints. One method takes advantage of the properties of the autoencoder to implicitly align the visual and textual modalities. While another method introduces a contrastive cross-modal alignment for explicit consistency. To prove the proposed CVT-SLR framework, extensive quantitative experiments and qualitative analysis are conducted, showing that our single-cue CVT-SLR not only outperforms single-cue baselines by a large margin but also surpasses SOTA multi-cue methods." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 485, + 545, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 485, + 545, + 568 + ], + "spans": [ + { + "bbox": [ + 304, + 485, + 545, + 568 + ], + "type": "text", + "content": "Overall, we hope our CVT-SLR can inspire the SLR community to mine pretrained modules from new aspects and design a more efficient SLR framework for practical usage. In future work, we plan to introduce a large-scale pretrained language model to guide textual module learning and further explore the potential of prior contextual linguistic knowledge for SLR tasks." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 594, + 404, + 607 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 594, + 404, + 607 + ], + "spans": [ + { + "bbox": [ + 306, + 594, + 404, + 607 + ], + "type": "text", + "content": "Acknowledgement" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 618, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 618, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 618, + 545, + 713 + ], + "type": "text", + "content": "We thank the anonymous reviewers for their constructive and helpful reviews. This work was supported by the National Key R&D Program of China (Project No. 2022ZD0115100), the National Natural Science Foundation of China (Project No. U21A20427), the Competitive Research Fund (Project No. WU2022A009) from the Westlake Center for Synthetic Biology and Integrated Bioengineering." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 733, + 318, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 733, + 318, + 742 + ], + "spans": [ + { + "bbox": [ + 293, + 733, + 318, + 742 + ], + "type": "text", + "content": "23148" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 53, + 91, + 286, + 113 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 286, + 113 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 286, + 113 + ], + "type": "text", + "content": "[1] Hangbo Bao, Li Dong, and Furu Wei. Beit: Bert pre-training of image transformers. In ICLR, 2022. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 114, + 287, + 169 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 114, + 287, + 169 + ], + "spans": [ + { + "bbox": [ + 53, + 114, + 287, + 169 + ], + "type": "text", + "content": "[2] Necati Cihan Camgoz, Simon Hadfield, Oscar Koller, Hermann Ney, and Richard Bowden. Neural sign language translation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7784-7793, 2018. 1, 3, 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 171, + 287, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 171, + 287, + 226 + ], + "spans": [ + { + "bbox": [ + 53, + 171, + 287, + 226 + ], + "type": "text", + "content": "[3] Necati Cihan Camgoz, Oscar Koller, Simon Hadfield, and Richard Bowden. Sign language transformers: Joint end-to-end sign language recognition and translation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10023-10033, 2020. 1, 3, 6, 8" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 228, + 287, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 228, + 287, + 281 + ], + "spans": [ + { + "bbox": [ + 53, + 228, + 287, + 281 + ], + "type": "text", + "content": "[4] Mathilde Caron, Ishan Misra, Julien Mairal, Priya Goyal, Piotr Bojanowski, and Armand Joulin. Unsupervised learning of visual features by contrasting cluster assignments. Advances in Neural Information Processing Systems, 33:9912-9924, 2020. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 284, + 287, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 284, + 287, + 327 + ], + "spans": [ + { + "bbox": [ + 53, + 284, + 287, + 327 + ], + "type": "text", + "content": "[5] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International conference on machine learning, pages 1597-1607. PMLR, 2020. 2, 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 330, + 287, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 330, + 287, + 373 + ], + "spans": [ + { + "bbox": [ + 53, + 330, + 287, + 373 + ], + "type": "text", + "content": "[6] Ting Chen, Simon Kornblith, Kevin Swersky, Mohammad Norouzi, and Geoffrey E Hinton. Big self-supervised models are strong semi-supervised learners. Advances in neural information processing systems, 33:22243-22255, 2020. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 375, + 287, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 375, + 287, + 430 + ], + "spans": [ + { + "bbox": [ + 53, + 375, + 287, + 430 + ], + "type": "text", + "content": "[7] Yutong Chen, Fangyun Wei, Xiao Sun, Zhirong Wu, and Stephen Lin. A simple multi-modality transfer learning baseline for sign language translation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5120-5130, 2022. 3, 4, 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 432, + 287, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 432, + 287, + 475 + ], + "spans": [ + { + "bbox": [ + 53, + 432, + 287, + 475 + ], + "type": "text", + "content": "[8] Ka Leong Cheng, Zhaoyang Yang, Qifeng Chen, and Yu-Wing Tai. Fully convolutional networks for continuous sign language recognition. In European Conference on Computer Vision, pages 697-714. Springer, 2020. 2, 3, 5, 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 477, + 287, + 530 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 477, + 287, + 530 + ], + "spans": [ + { + "bbox": [ + 53, + 477, + 287, + 530 + ], + "type": "text", + "content": "[9] Necati Cihan Camgoz, Simon Hadfield, Oscar Koller, and Richard Bowden. Subunets: End-to-end hand shape and continuous sign language recognition. In Proceedings of the IEEE international conference on computer vision, pages 3056-3065, 2017. 1, 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 533, + 287, + 587 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 533, + 287, + 587 + ], + "spans": [ + { + "bbox": [ + 48, + 533, + 287, + 587 + ], + "type": "text", + "content": "[10] Runpeng Cui, Hu Liu, and Changshui Zhang. Recurrent convolutional neural networks for continuous sign language recognition by staged optimization. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7361-7369, 2017. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 590, + 287, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 590, + 287, + 632 + ], + "spans": [ + { + "bbox": [ + 48, + 590, + 287, + 632 + ], + "type": "text", + "content": "[11] Runpeng Cui, Hu Liu, and Changshui Zhang. A deep neural framework for continuous sign language recognition by iterative training. IEEE Transactions on Multimedia, 21(7):1880-1891, 2019. 2, 3, 5" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 635, + 287, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 635, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 635, + 287, + 689 + ], + "type": "text", + "content": "[12] Cleison Correia de Amorim, David Macêdo, and Cleber Zanchettin. Spatial-temporal graph convolutional networks for sign language recognition. In International Conference on Artificial Neural Networks, pages 646-657. Springer, 2019. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "type": "text", + "content": "[13] Zhangyang Gao, Cheng Tan, Lirong Wu, and Stan Z Li. Simvp: Simpler yet better video prediction. In Proceedings" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "text", + "content": "of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3170-3180, 2022. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 96, + 545, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 96, + 545, + 150 + ], + "spans": [ + { + "bbox": [ + 308, + 96, + 545, + 150 + ], + "type": "text", + "content": "[14] Alex Graves, Santiago Fernández, Faustino Gomez, and Jürgen Schmidhuber. Connectionist temporal classification: labelling unsegmented sequence data with recurrent neural networks. In Proceedings of the 23rd international conference on Machine learning, pages 369-376, 2006. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 152, + 545, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 152, + 545, + 205 + ], + "spans": [ + { + "bbox": [ + 307, + 152, + 545, + 205 + ], + "type": "text", + "content": "[15] Aiming Hao, Yuecong Min, and Xilin Chen. Self-mutual distillation learning for continuous sign language recognition. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11303-11312, 2021. 1, 2, 3, 4, 5, 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 208, + 545, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 208, + 545, + 262 + ], + "spans": [ + { + "bbox": [ + 308, + 208, + 545, + 262 + ], + "type": "text", + "content": "[16] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9729-9738, 2020. 5" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 264, + 545, + 307 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 264, + 545, + 307 + ], + "spans": [ + { + "bbox": [ + 308, + 264, + 545, + 307 + ], + "type": "text", + "content": "[17] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 1, 4" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 309, + 545, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 309, + 545, + 352 + ], + "spans": [ + { + "bbox": [ + 308, + 309, + 545, + 352 + ], + "type": "text", + "content": "[18] Yi Ji, Chunping Liu, Shengrong Gong, and Weidong Cheng. 3d hand gesture coding for sign language learning. In 2016 International Conference on Virtual Reality and Visualization (ICVRV), pages 407-410. IEEE, 2016. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 354, + 545, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 354, + 545, + 396 + ], + "spans": [ + { + "bbox": [ + 308, + 354, + 545, + 396 + ], + "type": "text", + "content": "[19] Suhail Muhammad Kamal, Yidong Chen, Shaozi Li, Xiaodong Shi, and Jiangbin Zheng. Technical approaches to Chinese sign language processing: a review. IEEE Access, 7:96926-96935, 2019. 1, 2, 8" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 399, + 545, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 399, + 545, + 430 + ], + "spans": [ + { + "bbox": [ + 308, + 399, + 545, + 430 + ], + "type": "text", + "content": "[20] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 433, + 545, + 454 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 433, + 545, + 454 + ], + "spans": [ + { + "bbox": [ + 308, + 433, + 545, + 454 + ], + "type": "text", + "content": "[21] Diederik P Kingma and Max Welling. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114, 2013. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 456, + 545, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 456, + 545, + 510 + ], + "spans": [ + { + "bbox": [ + 308, + 456, + 545, + 510 + ], + "type": "text", + "content": "[22] Oscar Koller, Necati Cihan Camgoz, Hermann Ney, and Richard Bowden. Weakly supervised learning with multistream cnn-lstm-hmms to discover sequential parallelism in sign language videos. IEEE transactions on pattern analysis and machine intelligence, 42(9):2306-2320, 2019. 1, 3, 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 512, + 545, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 512, + 545, + 555 + ], + "spans": [ + { + "bbox": [ + 308, + 512, + 545, + 555 + ], + "type": "text", + "content": "[23] Oscar Koller, Jens Forster, and Hermann Ney. Continuous sign language recognition: Towards large vocabulary statistical recognition systems handling multiple signers. Computer Vision and Image Understanding, 141:108-125, 2015. 1, 5" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 557, + 545, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 557, + 545, + 609 + ], + "spans": [ + { + "bbox": [ + 308, + 557, + 545, + 609 + ], + "type": "text", + "content": "[24] Oscar Koller, Sepehr Zargaran, and Hermann Ney. Re-sign: Re-aligned end-to-end sequence modelling with deep recurrent cnn-hmms. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4297-4305, 2017. 1, 3, 5" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 613, + 545, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 613, + 545, + 656 + ], + "spans": [ + { + "bbox": [ + 308, + 613, + 545, + 656 + ], + "type": "text", + "content": "[25] Siyuan Li, Zedong Wang, Zicheng Liu, Cheng Tan, Haitao Lin, Di Wu, Zhiyuan Chen, Jiangbin Zheng, and Stan Z Li. Efficient multi-order gated aggregation network. arXiv preprint arXiv:2211.03295, 2022. 1" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 658, + 545, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 658, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 308, + 658, + 545, + 689 + ], + "type": "text", + "content": "[26] Zhi-jie Liang, Sheng-bin Liao, and Bing-zhang Hu. 3d convolutional neural networks for dynamic sign language recognition. The Computer Journal, 61(11):1724-1736, 2018. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 308, + 692, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 692, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 308, + 692, + 545, + 712 + ], + "type": "text", + "content": "[27] Ilya Loshchilov and Frank Hutter. Fixing weight decay regularization in adam. 2018. 6" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 733, + 318, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 733, + 318, + 742 + ], + "spans": [ + { + "bbox": [ + 293, + 733, + 318, + 742 + ], + "type": "text", + "content": "23149" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "text", + "content": "[28] Yuecong Min, Aiming Hao, Xiujuan Chai, and Xilin Chen. Visual alignment constraint for continuous sign language recognition. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11542-11551, 2021. 2, 3, 4, 5" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 129, + 288, + 174 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 129, + 288, + 174 + ], + "spans": [ + { + "bbox": [ + 48, + 129, + 288, + 174 + ], + "type": "text", + "content": "[29] Zhe Niu and Brian Mak. Stochastic fine-grained labeling of multi-state sign glosses for continuous sign language recognition. In European Conference on Computer Vision, pages 172-186. Springer, 2020. 1, 2, 3, 5, 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 175, + 288, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 175, + 288, + 229 + ], + "spans": [ + { + "bbox": [ + 48, + 175, + 288, + 229 + ], + "type": "text", + "content": "[30] Tse-Yu Pan, Li-Yun Lo, Chung-Wei Yeh, Jhe-Wei Li, Hou-Tim Liu, and Min-Chun Hu. Sign language recognition in complex background scene based on adaptive skin colour modelling and support vector machine. International Journal of Big Data Intelligence, 5(1-2):21-30, 2018. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 231, + 288, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 231, + 288, + 285 + ], + "spans": [ + { + "bbox": [ + 48, + 231, + 288, + 285 + ], + "type": "text", + "content": "[31] Jinye Peng, Li Yulong, et al. Chinese sign language recognition based on gray-level co-occurrence matrix and other multi-features fusion. In 2009 4th IEEE Conference on Industrial Electronics and Applications, pages 1569–1572. IEEE, 2009. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 286, + 287, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 286, + 287, + 341 + ], + "spans": [ + { + "bbox": [ + 48, + 286, + 287, + 341 + ], + "type": "text", + "content": "[32] Junfu Pu, Wengang Zhou, Hezhen Hu, and Houqiang Li. Boosting continuous sign language recognition via cross modality augmentation. In Proceedings of the 28th ACM International Conference on Multimedia, pages 1497-1505, 2020. 1, 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 342, + 287, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 342, + 287, + 397 + ], + "spans": [ + { + "bbox": [ + 48, + 342, + 287, + 397 + ], + "type": "text", + "content": "[33] Junfu Pu, Wengang Zhou, and Houqiang Li. Iterative alignment network for continuous sign language recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4165-4174, 2019. 1, 2, 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 399, + 287, + 465 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 399, + 287, + 465 + ], + "spans": [ + { + "bbox": [ + 48, + 399, + 287, + 465 + ], + "type": "text", + "content": "[34] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 1, 2, 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 466, + 287, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 466, + 287, + 510 + ], + "spans": [ + { + "bbox": [ + 48, + 466, + 287, + 510 + ], + "type": "text", + "content": "[35] Danilo Jimenez Rezende, Shakir Mohamed, and Daan Wierstra. Stochastic backpropagation and approximate inference in deep generative models. In International conference on machine learning, pages 1278-1286. PMLR, 2014. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 512, + 287, + 576 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 512, + 287, + 576 + ], + "spans": [ + { + "bbox": [ + 48, + 512, + 287, + 576 + ], + "type": "text", + "content": "[36] Ramprasaath R Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra. Grad-cam: Visual explanations from deep networks via gradient-based localization. In Proceedings of the IEEE international conference on computer vision, pages 618-626, 2017. 8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 578, + 287, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 578, + 287, + 622 + ], + "spans": [ + { + "bbox": [ + 48, + 578, + 287, + 622 + ], + "type": "text", + "content": "[37] Cheng Tan, Zhangyang Gao, Siyuan Li, Yongjie Xu, and Stan Z Li. Temporal attention unit: Towards efficient spatiotemporal predictive learning. arXiv preprint arXiv:2206.12126, 2022. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 624, + 287, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 624, + 287, + 668 + ], + "spans": [ + { + "bbox": [ + 48, + 624, + 287, + 668 + ], + "type": "text", + "content": "[38] Cheng Tan, Jun Xia, Lirong Wu, and Stan Z Li. Co-learning: Learning from noisy labels with self-supervision. In Proceedings of the 29th ACM International Conference on Multimedia, pages 1405-1413, 2021. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 669, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 669, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 669, + 287, + 713 + ], + "type": "text", + "content": "[39] Yiqi Tong, Jiangbin Zheng, Hongkang Zhu, Yidong Chen, and Xiaodong Shi. A document-level neural machine translation model with dynamic caching guided by theme-rheme information. In Proceedings of the 28th International Con-" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 643 + ], + "type": "list", + "angle": 0, + "index": 25, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "text", + "content": "ference on Computational Linguistics, pages 4385-4395, 2020.1" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 307, + 96, + 545, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 96, + 545, + 140 + ], + "spans": [ + { + "bbox": [ + 307, + 96, + 545, + 140 + ], + "type": "text", + "content": "[40] Hanjie Wang, Xiujuan Chai, Xiaopeng Hong, Guoying Zhao, and Xilin Chen. Isolated sign language recognition with grassmann covariance matrices. ACM Transactions on Accessible Computing (TACCESS), 8(4):1-21, 2016. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 141, + 545, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 141, + 545, + 196 + ], + "spans": [ + { + "bbox": [ + 307, + 141, + 545, + 196 + ], + "type": "text", + "content": "[41] Hanjie Wang, Xiujuan Chai, Yu Zhou, and Xin Chen. Fast sign language recognition benefited from low rank approximation. In 2015 11th IEEE International Conference and Workshops on Automatic Face and Gesture Recognition (FG), volume 1, pages 1-6. IEEE, 2015. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 197, + 545, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 197, + 545, + 251 + ], + "spans": [ + { + "bbox": [ + 307, + 197, + 545, + 251 + ], + "type": "text", + "content": "[42] Chong Wu, Jiangbin Zheng, Zhenan Feng, Houwang Zhang, Le Zhang, Jiawang Cao, and Hong Yan. Fuzzy slic: Fuzzy simple linear iterative clustering. IEEE Transactions on Circuits and Systems for Video Technology, 31(6):2114-2124, 2020. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 252, + 545, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 252, + 545, + 296 + ], + "spans": [ + { + "bbox": [ + 307, + 252, + 545, + 296 + ], + "type": "text", + "content": "[43] Jiangbin Zheng, Yidong Chen, Chong Wu, Xiaodong Shi, and Suhail Muhammad Kamal. Enhancing neural sign language translation by highlighting the facial expression information. Neurocomputing, 464:462-472, 2021. 1, 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 297, + 545, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 297, + 545, + 342 + ], + "spans": [ + { + "bbox": [ + 307, + 297, + 545, + 342 + ], + "type": "text", + "content": "[44] Jiangbin Zheng, Siyuan Li, Cheng Tan, Chong Wu, Yidong Chen, and Stan Z Li. Leveraging graph-based cross-modal information fusion for neural sign language translation. arXiv preprint arXiv:2211.00526, 2022. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 342, + 545, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 342, + 545, + 386 + ], + "spans": [ + { + "bbox": [ + 307, + 342, + 545, + 386 + ], + "type": "text", + "content": "[45] Jiangbin Zheng, Yile Wang, Ge Wang, Jun Xia, Yufei Huang, Guojiang Zhao, Yue Zhang, and Stan Z Li. Using context-to-vector with graph retrofitting to improve word embeddings. arXiv preprint arXiv:2210.16848, 2022. 5" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 387, + 545, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 387, + 545, + 442 + ], + "spans": [ + { + "bbox": [ + 307, + 387, + 545, + 442 + ], + "type": "text", + "content": "[46] Jiangbin Zheng, Zheng Zhao, Min Chen, Jing Chen, Chong Wu, Yidong Chen, Xiaodong Shi, and Yiqi Tong. An improved sign language translation model with explainable adaptations for processing long sign sentences. Computational Intelligence and Neuroscience, 2020, 2020. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 443, + 545, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 443, + 545, + 487 + ], + "spans": [ + { + "bbox": [ + 307, + 443, + 545, + 487 + ], + "type": "text", + "content": "[47] Hao Zhou, Wengang Zhou, and Houqiang Li. Dynamic pseudo label decoding for continuous sign language recognition. In 2019 IEEE International Conference on Multimedia and Expo (ICME), pages 1282-1287. IEEE, 2019. 5" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 487, + 545, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 487, + 545, + 542 + ], + "spans": [ + { + "bbox": [ + 307, + 487, + 545, + 542 + ], + "type": "text", + "content": "[48] Hao Zhou, Wengang Zhou, Yun Zhou, and Houqiang Li. Spatial-temporal multi-cue network for continuous sign language recognition. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 34, pages 13009-13016, 2020. 1, 3, 5, 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 544, + 545, + 598 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 544, + 545, + 598 + ], + "spans": [ + { + "bbox": [ + 307, + 544, + 545, + 598 + ], + "type": "text", + "content": "[49] Yongjie Zhuang, Bo Lv, Xinjun Sheng, and Xiangyang Zhu. Towards Chinese sign language recognition using surface electromyography and accelerometers. In 2017 24th International Conference on Mechatronics and Machine Vision in Practice (M2VIP), pages 1-5. IEEE, 2017. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 599, + 547, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 599, + 547, + 643 + ], + "spans": [ + { + "bbox": [ + 307, + 599, + 547, + 643 + ], + "type": "text", + "content": "[50] Ronglai Zuo and Brian Mak. C2slr: Consistency-enhanced continuous sign language recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5131-5140, 2022. 1, 2, 3, 5, 6" + } + ] + } + ], + "index": 24 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 734, + 318, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 734, + 318, + 742 + ], + "spans": [ + { + "bbox": [ + 293, + 734, + 318, + 742 + ], + "type": "text", + "content": "23150" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/CXTrack_ Improving 3D Point Cloud Tracking With Contextual Information/c7b22be7-4a5c-4780-8a5a-1db50c1d60af_content_list.json b/2023/CXTrack_ Improving 3D Point Cloud Tracking With Contextual Information/c7b22be7-4a5c-4780-8a5a-1db50c1d60af_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..af2369b06c8b8703977eee4361e2e6bd59874229 --- /dev/null +++ b/2023/CXTrack_ Improving 3D Point Cloud Tracking With Contextual Information/c7b22be7-4a5c-4780-8a5a-1db50c1d60af_content_list.json @@ -0,0 +1,1621 @@ +[ + { + "type": "text", + "text": "CXTrack: Improving 3D Point Cloud Tracking with Contextual Information", + "text_level": 1, + "bbox": [ + 99, + 130, + 870, + 151 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Tian-Xing Xu $^{1}$ Yuan-Chen Guo $^{1}$ Yu-Kun Lai $^{2}$ Song-Hai Zhang $^{1*}$", + "bbox": [ + 200, + 179, + 772, + 199 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Tsinghua University, China $^{2}$ Cardiff University, United Kingdom", + "bbox": [ + 209, + 198, + 761, + 217 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}\\{xutx21@mails., guocycl9@mails., shz@}tsinghua.edu.cn$ $^{2}$ LaiY4@cardiff.ac.uk", + "bbox": [ + 156, + 218, + 808, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 268, + 313, + 284 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3D single object tracking plays an essential role in many applications, such as autonomous driving. It remains a challenging problem due to the large appearance variation and the sparsity of points caused by occlusion and limited sensor capabilities. Therefore, contextual information across two consecutive frames is crucial for effective object tracking. However, points containing such useful information are often overlooked and cropped out in existing methods, leading to insufficient use of important contextual knowledge. To address this issue, we propose CXTrack, a novel transformer-based network for 3D object tracking, which exploits ConteXtual information to improve the tracking results. Specifically, we design a target-centric transformer network that directly takes point features from two consecutive frames and the previous bounding box as input to explore contextual information and implicitly propagate target cues. To achieve accurate localization for objects of all sizes, we propose a transformer-based localization head with a novel center embedding module to distinguish the target from distractors. Extensive experiments on three large-scale datasets, KITTI, nuScenes and Waymo Open Dataset, show that CXTrack achieves state-of-the-art tracking performance while running at 34 FPS.", + "bbox": [ + 73, + 301, + 473, + 650 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 680, + 209, + 696 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Single Object Tracking (SOT) has been a fundamental task in computer vision for decades, aiming to keep track of a specific target across a video sequence, given only its initial status. In recent years, with the development of 3D data acquisition devices, it has drawn increasing attention for using point clouds to solve various vision tasks such as object detection [7, 12, 14, 15, 18] and object tracking [20, 29, 31-33]. In particular, much progress has been made on point cloud-based object tracking for its huge potential in applications such as autonomous driving [11, 30]. However, it remains challenging due to the large appearance variation", + "bbox": [ + 75, + 705, + 468, + 875 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/77c8f1b7e853edc39d17d9d5dfaa6d87a85734be4e4583ceef1649bc2a212abf.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 504, + 265, + 890, + 368 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/7a1c931e1bcc9e0e957af9c111cdb8ef9d5c78f9ea73ca221e92b24c3e776ae1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 503, + 371, + 890, + 472 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/fc593d0d67219ed4cf6f62564c4def8a65a2a62ff521c9e74d0b28369e62bb15.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 504, + 474, + 890, + 575 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/d05a41c5882a19cec4a751cff59ff44cb4bc1ec2b25308a70ec07439396b86be.jpg", + "image_caption": [ + "Figure 1. Comparison of various 3D SOT paradigms. Previous methods crop the target from the frames to specify the region of interest, which largely overlook contextual information around the target. On the contrary, our proposed CXTrack fully exploits contextual information to improve the tracking results." + ], + "image_footnote": [], + "bbox": [ + 506, + 578, + 890, + 678 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "of the target and the sparsity of 3D point clouds caused by occlusion and limited sensor resolution.", + "bbox": [ + 496, + 762, + 890, + 792 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Existing 3D point cloud-based SOT methods can be categorized into three main paradigms, namely SC3D, P2B and motion-centric, as shown in Fig. 1. As a pioneering work, SC3D [6] crops the target from the previous frame, and compares the target template with a potentially large number of candidate patches generated from the current frame, which consumes much time. To address the effi", + "bbox": [ + 496, + 794, + 893, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* corresponding author", + "bbox": [ + 94, + 887, + 217, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "1084", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ciency problem, P2B [20] takes the cropped target template from the previous frame as well as the complete search area in the current frame as input, propagates target cues into the search area and then adopts a 3D region proposal network [18] to predict the current bounding box. P2B reaches a balance between performance and speed. Therefore many follow-up works adopt the same paradigm [3, 8, 9, 22, 29, 31, 33]. However, both SC3D and P2D paradigms overlook the contextual information across two consecutive frames and rely entirely on the appearance of the target. As mentioned in previous work [32], these methods are sensitive to appearance variation caused by occlusions and tend to drift towards intra-class distractors. To this end, M2-Track [32] introduces a novel motion-centric paradigm, which directly takes point clouds from two frames without cropping as input, and then segments the target points from their surroundings. After that, these points are cropped and the current bounding box is estimated by explicitly modeling motion between the two frames. Hence, the motion-centric paradigm still works on cropped patches that lack contextual information in later localization. In short, none of these methods could fully utilize the contextual information around the target to predict the current bounding box, which may degrade tracking performance due to the existence of large appearance variation and widespread distractors.", + "bbox": [ + 76, + 90, + 472, + 484 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address the above concerns, we propose a novel transformer-based tracker named CXTrack for 3D SOT, which exploits contextual information across two consecutive frames to improve the tracking performance. As shown in Fig. 1, different from paradigms commonly adopted by previous methods, CXTrack directly takes point clouds from the two consecutive frames as input, specifies the target of interest with the previous bounding box and predicts the current bounding box without any cropping, largely preserving contextual information. We first embed local geometric information of the two point clouds into point features using a shared backbone network. Then we integrate the targetness information into the point features according to the previous bounding box and adopt a target-centric transformer to propagate the target cues into the current frame while exploring contextual information in the surroundings of the target. After that, the enhanced point features are fed into a novel localization head named X-RPN to obtain the final target proposals. Specifically, X-RPN adopts a local transformer [25] to model point feature interactions within the target, which achieves a better balance between handling small and large objects compared with other localization heads. To distinguish the target from distractors, we incorporate a novel center embedding module into X-RPN, which embeds the relative target motion between two frames for explicit motion modeling. Extensive experiments on three popular tracking datasets demonstrate", + "bbox": [ + 76, + 493, + 472, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "that CXTrack significantly outperforms the current state-of-the-art methods by a large margin while running at real-time (34 FPS) on a single NVIDIA RTX3090 GPU.", + "bbox": [ + 496, + 90, + 890, + 136 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In short, our contributions can be summarized as: (1) a new paradigm for the real-time 3D SOT task, which fully exploits contextual information across consecutive frames to improve the tracking accuracy; (2) CXTrack: a transformer-based tracker that employs a target-centric transformer architecture to propagate targetness information and exploit contextual information; and (3) X-RPN: a localization head that is robust to intra-class distractors and achieves a good balance between small and large targets.", + "bbox": [ + 496, + 136, + 893, + 273 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 286, + 640, + 301 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Early methods [13, 17, 23] for the 3D SOT task mainly focus on RGB-D information and tend to adopt 2D Siamese networks used in 2D object tracking with additional depth maps. However, the changes in illumination and appearance may degrade the performance of these RGB-D methods. As a pioneering work in this area, SC3D [6] crops the target from the previous frame with the previous bounding box, and then computes the cosine similarity between the target template and a series of 3D target proposals sampled from the current frame using a Siamese backbone. The pipeline relies on heuristic sampling, which is very time-consuming.", + "bbox": [ + 496, + 311, + 893, + 477 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address these issues, P2B [20] develops an end-to-end framework, which first employs a shared backbone to embed local geometry into point features, and then propagates target cues from the target template to the search area in the current frame. Finally, it adopts VoteNet [18] to generate 3D proposals and selects the proposal with the highest score as the target. P2B [20] reaches a balance between performance and efficiency, and many works follow the same paradigm. MLVSNet [29] aggregates information at multiple levels for more effective target localization. BAT [31] introduces a box-aware feature fusion module to enhance the correlation learning between the target template and the search area. V2B [8] proposes a voxel-to-BEV (Bird's Eye View) target localization network, which projects the point features into a dense BEV feature map to tackle the sparsity of point clouds. Inspired by the success of transformers [25], LTTR [3] adopts a transformer-based architecture to fuse features from two branches and propagate target cues. PTT [22] integrates a transformer module into the P2B architecture to refine point features. PTTR [33] introduces Point Relation Transformer for feature fusion and a light-weight Prediction Refinement Module for coarse-to-fine localization. ST-Net [9] develops an iterative coarse-to-fine correlation network for robust correlation learning.", + "bbox": [ + 496, + 478, + 893, + 839 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Although achieving promising results, the aforementioned methods crop the target from the previous frame using the given bounding box. This overlook of contextual information across two frames makes these methods sensitive", + "bbox": [ + 496, + 839, + 893, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "1085", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "to appearance variations caused by commonly occurred occlusions and thus the results tend to drift towards intra-class distractors, as mentioned in M2-Track [32]. To this end, M2-Track introduces a motion-centric paradigm to handle the 3D SOT problem, which directly takes the point clouds from two consecutive frames as input without cropping. It first localizes the target in the two frames by target segmentation, and then adopts PointNet [19] to predict the relative target motion from the cropped target area that lacks contextual information. M2-Track could not fully utilize local geometric and contextual information for prediction, which may hinder precise bounding box regression.", + "bbox": [ + 75, + 90, + 472, + 273 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 76, + 282, + 169, + 299 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Problem Definition", + "text_level": 1, + "bbox": [ + 76, + 308, + 261, + 324 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Given the initial state of the target, single object tracking aims to localize the target in a dynamic scene frame by frame. The initial state in the first frame is given as the 3D bounding box of the target, which can be parameterized by its center coordinates $(x,y,z)$ , orientation angle $\\theta$ (around the up-axis, which is sufficient for most tracked objects staying on the ground) and sizes along each axis $(w,l,h)$ . Since the tracking target has little change in size across frames even for non-rigid objects, we assume constant target size and only regress the translation offset $(\\Delta x,\\Delta y,\\Delta z)$ and the rotation angle $(\\Delta \\theta)$ between two consecutive frames to simplify the tracking task. By applying the translation and rotation to the 3D bounding box $\\mathcal{B}_{t - 1}\\in \\mathbb{R}^7$ in the previous frame, we can compute the 3D bounding box $\\mathcal{B}_t\\in \\mathbb{R}^7$ to localize the target in the current frame.", + "bbox": [ + 75, + 332, + 472, + 556 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Suppose the point clouds in two consecutive frames are denoted as $\\mathcal{P}_{t - 1}\\in \\mathbb{R}^{\\dot{N}_{t - 1}\\times 3}$ and $\\mathcal{P}_t\\in \\mathbb{R}^{\\dot{N}_t\\times 3}$ , respectively, where $\\dot{N}_{t - 1}$ and $\\dot{N}_t$ are the numbers of points in the point clouds. We follow M2-Track [32] and encode the 3D bounding box $\\mathcal{B}_{t - 1}$ as a targetness mask $\\dot{\\mathcal{M}}_{t - 1} = (m_{t - 1}^{1},m_{t - 1}^{2},\\dots ,m_{t - 1}^{\\dot{N}_{t - 1}})\\in \\mathbb{R}^{\\dot{N}_{t - 1}}$ to indicate the target position, where the mask $m_{t - 1}^{i}$ for the $i$ -th point $p_{t - 1}^{i}$ is defined as", + "bbox": [ + 75, + 559, + 472, + 683 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nm _ {t - 1} ^ {i} = \\left\\{ \\begin{array}{l l} 0 & p _ {t - 1} ^ {i} \\text {n o t i n} \\mathcal {B} _ {t - 1} \\\\ 1 & p _ {t - 1} ^ {i} \\text {i n} \\mathcal {B} _ {t - 1} \\end{array} \\right. \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 163, + 688, + 468, + 720 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Thus, the 3D SOT task can be formalized as learning the following mapping", + "bbox": [ + 75, + 726, + 468, + 758 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {F} \\left(\\mathcal {P} _ {t - 1}, \\dot {\\mathcal {M}} _ {t - 1}, \\mathcal {P} _ {t}\\right) \\mapsto (\\Delta x, \\Delta y, \\Delta z, \\Delta \\theta) \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 129, + 763, + 468, + 781 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Overview of CXTrack", + "text_level": 1, + "bbox": [ + 76, + 786, + 282, + 801 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Following Eq. (2), we propose a network named CX-Track to improve tracking accuracy by fully exploiting contextual information across frames, and the overall design is illustrated in Fig. 2. We first apply a hierarchical feature learning architecture as the shared backbone to embed local geometric features of the point clouds into point", + "bbox": [ + 75, + 809, + 472, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "features. We use $N_{t - 1}$ and $N_{t}$ to denote the numbers of point features extracted by the backbone. For convenience of calculation, we create a targetness mask $\\dot{\\mathcal{M}}_t$ and fill it with 0.5 as it is unknown. We then concatenate the point features and targetness masks of the two frames to get $\\mathcal{X} = \\mathcal{X}_{t - 1}\\oplus \\mathcal{X}_t\\in \\mathbb{R}^{N\\times C}$ and $\\mathcal{M} = \\mathcal{M}_{t - 1}\\oplus \\mathcal{M}_t\\in \\mathbb{R}^{N\\times 1}$ where $N = N_{t - 1} + N_{t}$ , $\\mathcal{M}_{t - 1}$ and $\\mathcal{M}_t$ are masks corresponding to point features, and extracted from $\\dot{\\mathcal{M}}_{t - 1}$ and $\\dot{\\mathcal{M}}_t$ , and $C$ is the number of channels for point features. We employ the target-centric transformer (Sec. 3.3) to integrate the targetness mask information into point features while exploring the contextual information across frames. Finally, we propose a novel localization network, named X-RPN (Sec. 3.4), to obtain the target proposals. The proposal with the highest targetness score is verified as the result.", + "bbox": [ + 496, + 90, + 893, + 319 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.3. Target-Centric Transformer", + "text_level": 1, + "bbox": [ + 498, + 332, + 754, + 349 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Target-Centric Transformer aims to enhance the point features using the contextual information around the target while propagating the target cues from the previous frame to the current frame. It is composed of $N_{L} = 4$ identical layers stacked in series. Given the point features $\\mathcal{X}^{(k - 1)}\\in \\mathbb{R}^{N\\times C}$ and the targetness mask $\\mathcal{M}^{(k - 1)}$ from the $(k - 1)$ -th layer as input $(\\mathcal{M}^{(0)} = \\mathcal{M}$ and $\\mathcal{X}^{(0)} = \\mathcal{X})$ , the $k$ -th layer first models the interactions between any two points while integrating the targetness mask into point features using a modified self-attention operation, and then adopts Multi-Layer-Perceptrons (MLPs) to compute the new point features $\\mathcal{X}^{(k)}$ as well as the refined targetness mask $\\mathcal{M}^{(k)}$ . Thus, the predicted targetness mask will be consistently refined layer by layer. Moreover, we found it beneficial to add an auxiliary loss by predicting a potential target center for each point via Hough voting, so each layer also applies a shared MLP to generate the potential target center $C^{(k)}\\in \\mathbb{R}^{N\\times 3}$ .", + "bbox": [ + 496, + 358, + 893, + 616 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Formally, we first employ layer normalization [1] $\\mathrm{LN}(\\cdot)$ before the self-attention mechanism [25] following the design of 3DETR [15], which can be written as", + "bbox": [ + 496, + 617, + 893, + 662 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\bar {X} = \\mathrm {L N} \\left(\\mathcal {X} ^ {(k - 1)}\\right) \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 632, + 676, + 890, + 695 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Then, we add positional encodings (PE) of the coordinates to the normalized point features before feeding them into the self-attention operation", + "bbox": [ + 496, + 710, + 893, + 756 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nX _ {Q} = X _ {K} = \\bar {X} + \\mathrm {P E} \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 617, + 770, + 890, + 787 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nX _ {V} = \\bar {X} \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 666, + 791, + 890, + 809 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "It is worth noting that we only adopt PE for the query and key branches, therefore each refined point feature is constrained to focus more on local geometry instead of its associated absolute position. Subsequently, the transformer layer employs a global self-attention operation to model the", + "bbox": [ + 496, + 825, + 893, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "1086", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/7ea30e35c661f2e919c71b0854d6d3650642d285775be33f81fd7d3ada7aa47f.jpg", + "image_caption": [ + "Figure 2. The overall architecture of CXTrack. Given two consecutive point clouds and the 3D bounding box in the previous frame, CXTrack first embeds the local geometry into point features using the backbone. Then, CXTrack employs the target-centric transformer to explore contextual information across two frames and propagate the target cues to the current frame. Finally, the enhanced features are fed into a novel localization network named X-RPN to obtain high-quality proposals for verification." + ], + "image_footnote": [], + "bbox": [ + 76, + 85, + 888, + 231 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/80369015367aba9cb344d87501ad2babdbf0830d30fff22e886393a47cbc9065.jpg", + "image_caption": [ + "Figure 3. Comparison of various transformer layers to fuse the targetness mask and point features. We introduce three types of target-centric transformer layers, namely Vanilla, Semi-Dropout and Gated layer to integrate the targetness mask information into the point features while modeling intra-frame and inter-frame feature relationships." + ], + "image_footnote": [], + "bbox": [ + 76, + 301, + 893, + 474 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "relationships between point features, formulated as", + "bbox": [ + 76, + 530, + 416, + 545 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {M H A} \\left(X _ {Q}, X _ {K}, X _ {V}\\right) = \\operatorname {C o n c a t} \\left(\\text {h e a d} _ {1}, \\dots \\text {h e a d} _ {h}\\right) W ^ {O} \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 84, + 556, + 468, + 574 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\text {w h e r e} \\operatorname {h e a d} _ {i} = \\operatorname {A t t n} \\left(X _ {Q} W _ {i} ^ {Q}, X _ {K} W _ {i} ^ {K}, \\bar {X} W _ {i} ^ {V}\\right), \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 142, + 577, + 468, + 609 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {A t t n} (Q, K, V) = \\operatorname {s o f t m a x} \\left(\\frac {Q K ^ {T}}{\\sqrt {d _ {k}}}\\right) V \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 130, + 613, + 468, + 648 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Here, MHA indicates a multi-head attention, where the attention is applied in $h$ subspaces before concatenation. The projections are implemented by parameter matrices $W_{i}^{Q}\\in \\mathbb{R}^{C\\times d_{k}}$ , $W_{i}^{K}\\in \\mathbb{R}^{C\\times d_{k}}$ , $W_{i}^{V}\\in \\mathbb{R}^{C\\times d_{v}}$ and $W_{i}^{O}\\in \\mathbb{R}^{hd_{v}\\times C}$ , where $i$ indicates the $i$ -th subspace. The self-attention sublayer can be written as", + "bbox": [ + 76, + 657, + 468, + 748 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\widehat {X} = \\mathcal {X} ^ {(k - 1)} + \\operatorname {D r o p o u t} \\left(\\mathrm {M H A} \\left(X _ {Q}, X _ {K}, X _ {V}\\right)\\right) \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 104, + 758, + 468, + 777 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In addition to the self-attention sublayer, each transformer layer also contains a fully connected feed-forward network to refine the point features. The final output of the $k$ -th transformer layer is given by", + "bbox": [ + 76, + 789, + 470, + 849 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {X} ^ {(k)} = \\widehat {X} + \\operatorname {D r o p o u t} (\\operatorname {F F N} (\\operatorname {L N} (\\widehat {X}))), \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 163, + 859, + 468, + 878 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\text {w h e r e} \\operatorname {F F N} (x) = \\max \\left(0, x W _ {1} + b _ {1}\\right) W _ {2} + b _ {2}. \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 96, + 880, + 468, + 896 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To integrate the targetness mask information into point features, we need to modify the classic transformer layer. We introduce three types of modified transformer layers in Fig. 3, namely Vanilla, Semi-Dropout and Gated layer.", + "bbox": [ + 498, + 530, + 890, + 590 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Vanilla. We project the input $\\mathcal{M}^{(k - 1)}$ to mask embedding $\\mathrm{ME}\\in \\mathbb{R}^{N\\times C}$ using a linear transformation. Following the design of positional encoding, we simply add ME to the input token embedding $X_{V}$ , which re-formulates Eq. (5) as", + "bbox": [ + 498, + 590, + 890, + 652 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nX _ {V} = \\bar {X} + \\mathrm {M E} \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 640, + 667, + 890, + 683 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Semi-Dropout. Notably, the targetness mask information can only flow across layers along the attention path. For small objects which only have a few points to track, applying dropout to the mask embedding may discard the targetness information and lead to performance degradation. To this end, we separate the self-attention mechanism into a feature branch and a mask branch with shared attention weights, while only applying dropout to the refined point features. As shown in Fig. 3b, the self-attention sublayer in Eq. (9) is re-formulated as", + "bbox": [ + 496, + 693, + 892, + 844 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\widehat {X} = \\mathcal {X} ^ {(k - 1)} + \\operatorname {D r o p o u t} (\\mathrm {M H A} (X _ {Q}, X _ {K}, \\overline {{X}})) \\\\ + \\mathrm {M H A} \\left(X _ {Q}, X _ {K}, \\mathrm {M E}\\right) \\tag {13} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 514, + 858, + 890, + 896 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "1087", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/4a939d898bf5efa9aa6bdb33f8bbe36dc10ce6615a2e25b37222d34b0a53616f.jpg", + "image_caption": [ + "Figure 4. The overall architecture of X-RPN. X-RPN adopts a local transformer to model point feature interaction within the target and aggregate local clues. It also incorporates a center embedding mechanism which embeds the relative target motion between two frames to distinguish the target from distractors." + ], + "image_footnote": [], + "bbox": [ + 99, + 88, + 444, + 294 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Gated. Inspired by the design of TrDimp [26], we introduce a gated mechanism into the self-attention sublayer to integrate the mask information. It has two parallel branches, namely mask transformation and feature transformation. For mask transformation, we first obtain the feature mask $\\overline{M} \\in \\mathbb{R}^{N \\times C}$ by repeating the input point-wise mask $\\mathcal{M}^{(k - 1)} \\in \\mathbb{R}^{N \\times 1}$ for $C$ times. Then we can propagate the targetness cues to the current frame via adopting self-attention on the mask feature. The transformed mask serves as the gate matrix for the point features", + "bbox": [ + 75, + 376, + 468, + 526 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\widehat {X} _ {m} = \\mathrm {L N} (\\mathrm {M H A} (X _ {Q}, X _ {K}, \\bar {M}) \\otimes \\bar {X}). \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 142, + 536, + 468, + 555 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For feature transformation, we first mask the point features to suppress feature activation in background areas, and then employ self-attention with a residual connection to model the relationships between features", + "bbox": [ + 75, + 565, + 468, + 625 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\widehat {X} _ {f} = \\mathrm {L N} (\\mathrm {M H A} (X _ {Q}, X _ {K}, \\bar {M} \\otimes \\bar {X}) + \\bar {X}). \\tag {15}\n$$\n", + "text_format": "latex", + "bbox": [ + 114, + 633, + 468, + 652 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As illustrated in Fig. 3c, we sum and normalize the output features $\\widehat{X}_m$ and $\\widehat{X}_f$ from the two branches. Eq. (9) can be re-formulated as", + "bbox": [ + 75, + 662, + 468, + 708 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\widehat {X} = \\operatorname {L N} \\left(\\widehat {X} _ {f} + \\widehat {X} _ {m}\\right). \\tag {16}\n$$\n", + "text_format": "latex", + "bbox": [ + 199, + 717, + 468, + 736 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Among the above three layers, we observe significant performance gain from using Semi-Dropout target-centric transformer layers (Sec. 4.3). Thus CXTrack employs Semi-Dropout layers to integrate targetness information while exploring contextual information across frames.", + "bbox": [ + 75, + 747, + 468, + 823 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. X-RPN", + "text_level": 1, + "bbox": [ + 76, + 830, + 169, + 845 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Previous works [20] indicate that individual point features can only capture limited local information, which may not be sufficient for precise bounding box regression. Thus", + "bbox": [ + 75, + 854, + 468, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "we develop a simple yet effective localization network, named X-RPN, which extends RPN [18] using local transformer and center embedding, as shown in Fig. 4. Different from previous works [8,20], X-RPN aggregates local clues from point features without downsampling or voxelization, thus avoiding information loss and reaching a good balance between handling large and small objects. Our intuition is that each point should only interact with points belonging to the same object to suppress irrelevant information. Given the point features $\\mathcal{X}^{(N_L)}$ , targetness mask $\\mathcal{M}^{(N_L)}$ and target center $\\mathcal{C}^{(N_L)}$ output by the target centric transformer, we first split them along the spatial dimension and only feed those belonging to the current frame into X-RPN, which is denoted as $\\widetilde{\\mathcal{X}}\\in \\mathbb{R}^{N_t\\times C}$ , $\\widetilde{\\mathcal{C}}\\in \\mathbb{R}^{N_t\\times 3}$ and $\\widetilde{\\mathcal{M}}\\in \\mathbb{R}^{N_t\\times 1}$ , respectively. X-RPN first computes the neighborhood $\\mathcal{N}(p_i)$ for each point $p_i$ using its potential target center $c_{i}$", + "bbox": [ + 496, + 90, + 890, + 334 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {N} \\left(p _ {i}\\right) = \\left\\{p _ {j} \\mid | | c _ {i} - c _ {j} | | _ {2} < r \\right\\} \\tag {17}\n$$\n", + "text_format": "latex", + "bbox": [ + 584, + 343, + 890, + 369 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Here $r$ is a hyperparameter indicating the size of the neighborhood. Then X-RPN adopts the transformer architecture mentioned in Sec. 3.3 to aggregate local information, where each point only interacts with its neighborhood points to suppress noise. We remove the feed-forward network in the transformer layer because we observe that one layer is sufficient to generate high quality target proposals.", + "bbox": [ + 496, + 377, + 890, + 483 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To deal with intra-class distractors which are widespread in the scenes [32] especially for pedestrian tracking, we propose to combine the potential center information with the targetness mask. Our intuition lies in two folds. First, the tracking target keeps similar local geometry across two frames. Second, if the duration between two consecutive frames is sufficiently short, the displacement of the target is small. Therefore, we construct a Gaussian proposal-wise mask $\\mathcal{M}_c$ to indicate the magnitude of the displacement of each proposal. Formally, for each point $p_i$ with the predicted target center $c_i$ , the mask value $m_i^c \\in \\mathcal{M}_c$ is", + "bbox": [ + 496, + 483, + 890, + 650 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nm _ {i} ^ {c} = \\exp \\left(- \\frac {\\left| \\left| c _ {i} - \\bar {c} \\right| \\right| _ {2} ^ {2}}{2 \\sigma^ {2}}\\right) \\tag {18}\n$$\n", + "text_format": "latex", + "bbox": [ + 611, + 659, + 890, + 690 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\overline{c} \\in \\mathbb{R}^3$ is the target center in the previous frame and $\\sigma$ is a learnable or fixed scaling factor. We embed the target center mask $\\mathcal{M}_c$ into the center embedding matrix $\\mathrm{CE} \\in \\mathbb{R}^{N \\times C}$ using a linear transformation, and equally combine the mask embedding and the center embedding.", + "bbox": [ + 496, + 699, + 890, + 776 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.5. Loss Functions", + "text_level": 1, + "bbox": [ + 500, + 784, + 651, + 799 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For the prediction $\\mathcal{M}^{(k)}$ given by the $k$ -th transformer layer, we adopt a standard cross entropy loss $\\mathcal{L}_{\\mathrm{cm}}^{(i)}$ . As for the potential target centers, we observe that it is difficult to regress precise centers for non-rigid objects such as pedestrians. Hence the predicted centers $\\mathcal{C}^{(k)}$ are supervised by $L_{2}$ loss for non-rigid objects, and by Huber loss [21] for", + "bbox": [ + 496, + 806, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "1088", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "rigid objects. For the target center regression loss $\\mathcal{L}_{\\mathrm{cc}}^{(i)}$ , only points in the ground truth bounding box are supervised.", + "bbox": [ + 75, + 89, + 467, + 121 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Following previous works [20], proposals with predicted centers near the target center $(< 0.3\\mathrm{m})$ are considered as positives and those far away $(>0.6\\mathrm{m})$ are considered as negatives. Others are left unsupervised. The predicted targetness mask is supervised via standard cross-entropy loss $\\mathcal{L}_{\\mathrm{rm}}$ and only the bounding box parameters of positive predictions are supervised by Huber (Smooth- $L_{1}$ ) loss $\\mathcal{L}_{\\mathrm{box}}$ .", + "bbox": [ + 75, + 121, + 467, + 227 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The overall loss is the weighted combination of the above loss terms", + "bbox": [ + 75, + 227, + 467, + 255 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\gamma_ {1} \\sum_ {i = 1} ^ {N _ {L}} \\mathcal {L} _ {\\mathrm {c m}} ^ {(i)} + \\gamma_ {2} \\sum_ {i = 1} ^ {N _ {L}} \\mathcal {L} _ {\\mathrm {c c}} ^ {(i)} + \\gamma_ {3} \\mathcal {L} _ {\\mathrm {r m}} + \\mathcal {L} _ {\\mathrm {b o x}} \\tag {19}\n$$\n", + "text_format": "latex", + "bbox": [ + 102, + 262, + 467, + 303 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\gamma_{1},\\gamma_{2}$ and $\\gamma_{3}$ are hyper-parameters. We empirically set $\\gamma_{1} = 0.2$ $\\gamma_{2} = 1.0$ $\\gamma_{3} = 1.5$ for rigid objects and $\\gamma_{1} = 0.2$ $\\gamma_{2} = 10.0$ $\\gamma_{3} = 1.0$ for non-rigid objects.", + "bbox": [ + 75, + 311, + 467, + 357 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 75, + 368, + 207, + 385 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Settings", + "text_level": 1, + "bbox": [ + 76, + 392, + 174, + 407 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Datasets. We compare CXTrack with previous state of the arts on three large-scale datasets: KITTI [5], nuScenes [2] and Waymo Open Dataset (WOD) [24]. KITTI contains 21 training video sequences and 29 test sequences. We follow previous work [6] and split the training sequences into three parts, 0-16 for training, 17-18 for validation and 19-20 for testing. For nuScenes, we use its validation split to evaluate our model, which contains 150 scenes. For WOD, we follow LiDAR-SOT [16] to evaluate our method, dividing it into three splits according to the sparsity of point clouds.", + "bbox": [ + 75, + 415, + 467, + 566 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation Details. We adopt DGCNN [28] as the backbone network to extract local geometric information. In the X-RPN, we initialize the scaling parameter $\\sigma^2 = 10$ . Notably, we empirically fix $\\sigma$ as a hyper-parameter for pedestrians and cyclists, and set it as a learnable parameter for cars and vans, since they may have larger motions. More details are provided in the supplementary material.", + "bbox": [ + 75, + 566, + 467, + 672 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation Metrics. We use Success and Precision defined in one pass evaluation [10] as evaluation metrics. Success denotes the Area Under Curve (AUC) for the plot showing the ratio of frames where the Intersection Over Union (IOU) between the predicted and ground-truth bounding boxes is larger than a threshold, ranging from 0 to 1. Precision is defined as the AUC of the plot showing the ratio of frames where the distance between their centers is within a threshold, from 0 to 2 meters.", + "bbox": [ + 75, + 672, + 467, + 806 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Comparison with State of the Arts", + "text_level": 1, + "bbox": [ + 76, + 816, + 375, + 832 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We make comprehensive comparisons on the KITTI dataset with previous state-of-the-art methods, including SC3D [6], P2B [20], 3DSiamRPN [4], LTTR [3], MLVS-Net [29], BAT [31], PTT [22], V2B [8], PTTR [33],", + "bbox": [ + 75, + 839, + 467, + 900 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/c4e090f2647fe6f303e79d7107991734181ce316d7e511b6544db5663944e157.jpg", + "table_caption": [ + "Table 1. Comparisons with the state-of-the-art methods on KITTI dataset. \"Mean\" is the average result weighted by frame numbers. \"Blue\" and \"Bold\" denote previous and current best performance, respectively. Success/Precision are used for evaluation." + ], + "table_footnote": [], + "table_body": "
MethodCar (6424)Pedestrian (6088)Van (1248)Cyclist (308)Mean (14068)
SC3D41.3/57.918.2/37.840.4/47.041.5/70.431.2/48.5
P2B56.2/72.828.7/49.640.8/48.432.1/44.742.4/60.0
3DSiamRPN58.2/76.235.2/56.245.7/52.936.2/49.046.7/64.9
LTPR65.0/77.133.2/56.835.8/45.666.2/89.948.7/65.8
MLVSNet56.0/74.034.1/61.152.0/61.434.3/44.545.7/66.7
BAT60.5/77.742.1/70.152.4/67.033.7/45.451.2/72.8
PTT67.8/81.844.9/72.043.6/52.537.2/47.355.1/74.2
V2B70.5/81.348.3/73.550.1/58.040.8/49.758.4/75.2
PTTR65.2/77.450.9/81.652.5/61.865.1/90.557.9/78.1
STNet72.1/84.049.9/77.258.0/70.673.5/93.761.3/80.1
M2-Track65.5/80.861.5/88.253.8/70.773.2/93.562.9/83.4
CXTrack69.1/81.667.0/91.560.0/71.874.2/94.367.5/85.3
Improvement↓3.0/↓2.4↑5.5/↑3.3↑2.0/↑1.1↑0.7/↑0.6↑4.6/↑1.9
", + "bbox": [ + 506, + 147, + 887, + 335 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/6faee4123ed18230c85f19925750a624d6b630c55a423bd4856637e9d1cf8937.jpg", + "table_caption": [ + "Table 2. Robustness under scenes that contain intra-class distractors on KITTI Pedestrian category." + ], + "table_footnote": [], + "table_body": "
MethodAll(6088)Distractor-Only(3917)Improvement
PTTR50.9/81.644.3/70.0↓6.6/↓11.6
STNet49.9/77.235.1/58.5↓14.8/↓18.7
M2-Track61.5/88.258.0/88.4↓3.5/↑0.2
CXTrack67.0/91.566.1/91.3↓0.9/↓0.3
", + "bbox": [ + 535, + 378, + 857, + 441 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "STNet [9] and M2-Track [32]. As illustrated in Tab. 1, CXTrack surpasses previous state-of-the-art methods, with a significant improvement of average Success and Precision. Notably, our method achieves the best performance under all categories, except for the Car, where voxel-based STNet [9] surpasses us by a minor margin (72.1/84.0 v.s. 69.1/81.6). Most vehicles have simple shapes and limited rotation angles, which fit well in voxels. We argue that voxelization provides a strong shape prior, thereby leading to performance gain for large objects with simple shapes. The lack of distractors for cars also makes our improvement over previous methods insignificant. However, our method has a significant improvement (67.0/91.5 v.s. 49.9/77.2) on the Pedestrian category. We claim that this stems from our special design to handle distractors and our better preservation for contextual information. Besides, compared with M2-Track [32], CXTrack obtains consistent performance gains on all categories especially on the Success metric, which demonstrates the importance of local geometry and contextual information. For further analysis on the impact of intraclass distractors, we manually pick out scenes that contain Pedestrian distractors from the KITTI test split and then evaluate different methods on these scenes. As shown in Tab. 2, both M2-Track and CXTrack are robust to distractors, while CXTrack can make more accurate predictions.", + "bbox": [ + 496, + 459, + 890, + 837 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To verify the generalization ability of our method, we follow previous methods [8, 9] and test the KITTI pretrained model on nuScenes and WOD. The comparison results on WOD are shown in Tab. 3. It can be seen that our", + "bbox": [ + 496, + 839, + 890, + 898 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "1089", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/e73309839e5472be3298e0f456e5823fe24c8f2dc9affa05eb572ae7084cffd0.jpg", + "table_caption": [ + "Table 3. Comparison with state of the arts on Waymo Open Dataset." + ], + "table_footnote": [], + "table_body": "
MethodVehicle(185731)Pedestrian(241752)Mean(427483)
EasyMediumHardMeanEasyMediumHardMean
P2B57.1/65.452.0/60.747.9/58.552.6/61.718.1/30.817.8/30.017.7/29.317.9/30.133.0/43.8
BAT61.0/68.353.3/60.948.9/57.854.7/62.719.3/32.617.8/29.817.2/28.318.2/30.334.1/44.4
V2B64.5/71.555.1/63.252.0/62.057.6/65.927.9/43.922.5/36.220.1/33.123.7/37.938.4/50.1
STNet65.9/72.757.5/66.054.6/64.759.7/68.029.2/45.324.7/38.222.2/35.825.5/39.940.4/52.1
CXTrack63.9/71.154.2/62.752.1/63.757.1/66.135.4/55.329.7/47.926.3/44.430.7/49.442.2/56.7
Improvement↓2.0/↓1.6↓3.3/↓3.3↓3.5/↓1.0↓2.6/↓1.9↑6.2/↑10.0↑5.0/↑9.7↑4.1/↑8.6↑5.2/↑9.5↑1.8/↑4.6
", + "bbox": [ + 101, + 103, + 867, + 212 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/3b98402537dbb5dc1aecf6cdc84181d292b8036ac0e42164c365f7887d7da724.jpg", + "table_caption": [ + "Table 4. Comparison with state of the arts on nuScenes." + ], + "table_footnote": [], + "table_body": "
MethodCar (15578)Pedestrian (8019)Van (3710)Cyclist (501)Mean (27808)
SC3D25.0/27.114.2/16.225.7/21.917.0/18.221.8/23.1
P2B27.0/29.215.9/22.021.5/16.220.0/26.422.9/25.3
BAT22.5/24.117.3/24.519.3/15.817.0/18.820.5/23.0
V2B31.3/35.117.3/23.421.7/16.722.2/19.125.8/29.0
STNet32.2/36.119.1/27.222.3/16.821.2/29.226.9/30.8
CXTrack29.6/33.420.4/32.927.6/20.818.5/26.826.5/31.5
Improvement↓2.6/↓2.7↑1.3/↑5.7↑1.9/↓1.1↓3.7/↓2.4↓0.4/↑0.7
", + "bbox": [ + 84, + 234, + 464, + 351 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/563202c21b5f3313f0759e028a694c8ec496c6100ddf5e3862be36e6b56b51ad.jpg", + "table_caption": [ + "Table 5. Efficiency analysis of different components." + ], + "table_footnote": [], + "table_body": "
ComponentFLOPs#ParamsInfer Speed
backbone3.18G1.3M8.5ms
transformer1.28G14.7M10.9ms
X-RPN0.17G2.3M3.0ms
pre/postprocess--6.8ms
CXTrack4.63G18.3M29.2ms(34FPS)
", + "bbox": [ + 120, + 377, + 431, + 452 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "method outperforms others in terms of average Success and Precision with a clear margin. Notably, KITTI and WOD data are captured by 64-beam LiDARs, while nuScenes data are captured by 32-beam LiDARs. Thus it is more challenging to generalize the pretrained model on the nuScenes dataset. As shown in Tab. 4, our method achieves comparable performance on the nuScenes dataset. In short, CXTrack not only achieves a good balance between small objects and large objects, but also generalizes well to unseen scenes.", + "bbox": [ + 75, + 468, + 468, + 604 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We also visualize the tracking results for qualitative comparisons. As shown in Fig. 5, CXTrack achieves good accuracy in scenes with both sparse and dense point clouds on both categories. In the sparse cases (left), previous methods drift towards intra-class distractors due to large appearance variations caused by occlusions, while only our method keeps track of the target, thanks to the sufficient use of contextual information. In the dense cases (right), our method can track the target more accurately than M2-Track by leveraging local geometric information.", + "bbox": [ + 75, + 604, + 468, + 755 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We report the efficiency of different components in Tab. 5. It can be observed that the target-centric transformer is the bottleneck of CXTrack during inference. We can replace the vanilla self-attention in CXTrack with linear attention such as linformer [27] for further speedup.", + "bbox": [ + 75, + 757, + 468, + 834 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Ablation Studies", + "text_level": 1, + "bbox": [ + 76, + 845, + 241, + 859 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To validate the effectiveness of several design choices in CXTrack, we conduct ablation studies on the KITTI dataset.", + "bbox": [ + 76, + 869, + 468, + 898 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/7ced26cf11fc6659f499ceb74280347fde288ea4bc6850c51f00a4003f4de29c.jpg", + "table_caption": [ + "Table 6. Ablation studies of different components of the target-centric transformer. \"Cx\" refers to contextual information, \"M\" refers to the cascaded targetness mask prediction and \"C\" refers to the auxiliary target center regression branches." + ], + "table_footnote": [ + "†: unstable training process" + ], + "table_body": "
CxMCCarPedestrianVanCyclistMean
✓62.5/74.260.6/87.058.3/71.472.0/93.361.5/79.9
✓✓67.4/80.263.9/89.057.8/70.872.7/93.865.1/83.5
✓✓59.7/73.6†51.8/81.659.9/71.571.7/93.256.6/77.3
✓✓✓69.1/81.667.0/91.560.0/71.874.2/94.367.5/85.3
", + "bbox": [ + 503, + 280, + 888, + 339 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/7786ec22a88631829f9abe3d5a1b393a12b911df0596ceae13701d252064ea8b.jpg", + "table_caption": [ + "Table 7. Ablation studies of different transformer layers on KITTI. \"V\" refers to the vanilla transformer layer and \"G\" refers to the gated transformer layer. \"S\" represents the semi-dropout transformer layer which is adopted in our proposed CXTrack." + ], + "table_footnote": [], + "table_body": "
CarPedestrianVanCyclistMean
V68.8/80.462.9/87.857.2/69.672.7/94.265.3/82.9
G64.8/76.964.7/91.156.2/70.570.6/93.464.1/82.8
S69.1/81.667.0/91.560.0/71.874.2/94.367.5/85.3
", + "bbox": [ + 503, + 416, + 888, + 472 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Components of Target-Centric Transformer. Tab. 6 presents ablation studies of different components of transformer to gain a better understanding of its designs. We crop the input point cloud $\\mathcal{P}_{t-1}$ using $\\mathcal{B}_{t-1}$ to ablate contextual information in the previous frame. We can observe significant performance drop when not using contextual information, especially on Car and Pedestrian. For Car, it suffers from heavy occlusions (Fig. 5), while pedestrian distractors are widespread in the scene. We also find that removing context leads to unstable training on Car. We presume that the lack of supervised signals to tell the model what not to attend may confuse the model and introduce noise in training. For the cascaded targetness mask prediction and auxiliary target center regression, removing either of them leads to an obvious decline on terms of average metrics. We argue that the auxiliary regression loss can increase the feature similarities of points belonging to the same object.", + "bbox": [ + 496, + 492, + 890, + 748 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Target-Centric Transformer Layer. Tab. 7 shows the impact of different target-centric transformer layers. Semi-Dropout achieves better performance than Vanilla, especially on Pedestrian. Small objects often consist of fewer points, hence applying dropout directly on the targetness information in training may confuse the network and lead to sub-optimal results. Gated relies entirely on the predicted targetness mask to modulate the amount of exposure for input features, which may suffer from information loss when the targetness mask is not accurate enough.", + "bbox": [ + 496, + 750, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "1090", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/8a18fb599e340f05b4de5b2f3d9521f4b98e9a2ff6806d2c7b64df51f9321571.jpg", + "image_caption": [ + "Figure 5. Visualization results. Left: Sparse cases in KITTI. Right: Dense cases in KITTI." + ], + "image_footnote": [], + "bbox": [ + 83, + 89, + 488, + 277 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/134c9d88ef8e452ba93f35a66f819d6f2db382d8f0404d66f1235534f795f4bc.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 491, + 90, + 883, + 277 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/5764e63c17277e8db8af717426f089092c5df9c2354ca1754b1190aca3d72f39.jpg", + "table_caption": [ + "Table 8. Ablation studies of various localization heads on KITTI. \"X-RPN\\C\" indicates our proposed localization head X-RPN without center embedding." + ], + "table_footnote": [], + "table_body": "
HeadCarPedestrianVanCyclistMean
PRM [33]66.5/77.462.2/86.852.9/64.972.5/93.863.6/80.7
RPN [18]64.1/76.959.8/88.355.0/65.668.2/92.461.5/81.2
V2B [8]70.5/82.660.1/86.758.0/69.870.5/93.364.9/83.5
X-RPN\\C67.8/80.365.5/89.559.9/72.172.6/94.166.2/83.9
X-RPN69.1/81.667.0/91.560.0/71.874.2/94.367.5/85.3
", + "bbox": [ + 80, + 343, + 467, + 417 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/ad227c0d17dd9257df9a43ab0d9999566ccb1b92022a95c44183282c8c4dcd90.jpg", + "image_caption": [ + "Figure 6. Representative examples of attention maps in the transformer. Target-centric transformer attends to objects that have similar geometry." + ], + "image_footnote": [], + "bbox": [ + 78, + 428, + 263, + 523 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/1b440cac86f7e678c254783772b24d356a32f8bc26dc4f5f790924c4de3ffd14.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 266, + 428, + 467, + 523 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/3da16ceb93d9623f917fbcd52dab6e3372d8d696310ef5ca5971bdf92a49e19d.jpg", + "image_caption": [ + "Figure 7. Visualization of ablation study. Center embedding can benefit object tracking in challenging scenes with distractors." + ], + "image_footnote": [], + "bbox": [ + 78, + 577, + 468, + 676 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "X-RPN. We replace X-RPN with other alternatives [8, 20, 33] and report the comparison results in Tab. 8. Although the V2B head achieves better performance than X-RPN on the Car category, it fails to track small objects such as pedestrians effectively due to intra-class distractors and inevitable information loss brought in by voxelization. It is also worth noting that we observe a performance drop without center embedding, especially on the Pedestrian category, for which distractors are more commonly seen. To explore the effectiveness of the center embedding, we visualize the attention map of the last transformer layer in Fig. 6. We observe that the transformer alone can attend to regions with similar ge", + "bbox": [ + 75, + 719, + 468, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "ometry to the target, but fails to distinguish the target from distractors. As shown in Fig. 7, with the help of the center embedding, the network precisely keeps track of the target. In short, X-RPN achieves a good balance between large and small objects, and effectively alleviates the distractor problem.", + "bbox": [ + 496, + 301, + 890, + 390 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4. Failure Cases", + "text_level": 1, + "bbox": [ + 500, + 401, + 638, + 416 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Although CXTrack is robust to intra-class distractors, it fails to predict accurate orientation of the target when the point clouds are too sparse to capture informative local geometry or when large appearance variations occur, as shown in Fig. 7. Besides, the center embedding directly encodes the displacement of target center into features, so our model may suffer from performance degradation if trained with $2\\mathrm{Hz}$ data and tested with $10\\mathrm{Hz}$ data because the scale of the displacement differs significantly.", + "bbox": [ + 496, + 424, + 892, + 561 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 573, + 617, + 588 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We revisit existing paradigms for the 3D SOT task and propose a new paradigm to fully exploit contextual information across frames, which is largely overlooked by previous methods. Following this paradigm, we design a novel transformer-based network named CXTrack, which employs a target-centric transformer to explore contextual information and model intra-frame and inter-frame feature relationships. We also introduce a localization head named X-RPN to obtain high-quality proposals for objects of all sizes, as well as a center embedding module to distinguish the target from distractors. Extensive experiments show that CXTrack significantly outperforms previous state-of-the-arts, and is robust to distractors. We hope our work can promote further exploitations in this task by showing the necessity to explore contextual information for more robust predictions.", + "bbox": [ + 496, + 598, + 890, + 824 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgment The authors thank Jiahui Huang for his discussions. This work was supported by the Natural Science Foundation of China (Project Number 61832016), and Tsinghua-Tencent Joint Laboratory for Internet Innovation Technology.", + "bbox": [ + 496, + 825, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "1091", + "bbox": [ + 483, + 944, + 513, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 174, + 106 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Jimmy Lei Ba, Jamie Ryan Kiros, and Geoffrey E Hinton. Layer normalization. arXiv preprint arXiv:1607.06450, 2016.3", + "[2] Holger Caesar, Varun Bankiti, Alex H Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuScenes: A multimodal dataset for autonomous driving. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11621-11631, 2020. 6", + "[3] Yubo Cui, Zheng Fang, Jiayao Shan, Zuoxu Gu, and Sifan Zhou. 3D object tracking with transformer. arXiv preprint arXiv:2110.14921, 2021. 2, 6", + "[4] Zheng Fang, Sifan Zhou, Yubo Cui, and Sebastian Scherer. 3D-SiamRPN: An end-to-end learning method for real-time 3D single object tracking using raw point cloud. IEEE Sensors Journal, 21(4):4995-5011, 2020. 6", + "[5] Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we ready for autonomous driving? the KITTI vision benchmark suite. In IEEE Conference on Computer Vision and Pattern Recognition, pages 3354-3361. IEEE, 2012. 6", + "[6] Silvio Giancola, Jesus Zarzar, and Bernard Ghanem. Leveraging shape completion for 3D Siamese tracking. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1359-1368, 2019. 1, 2, 6", + "[7] Chenhang He, Ruihuang Li, Shuai Li, and Lei Zhang. Voxel set transformer: A set-to-set approach to 3D object detection from point clouds. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8417-8427, 2022. 1", + "[8] Le Hui, Lingpeng Wang, Mingmei Cheng, Jin Xie, and Jian Yang. 3D Siamese voxel-to-BEV tracker for sparse point clouds. Advances in Neural Information Processing Systems, 34:28714-28727, 2021. 2, 5, 6, 8", + "[9] Le Hui, Lingpeng Wang, Linghua Tang, Kaihao Lan, Jin Xie, and Jian Yang. 3D Siamese transformer network for single object tracking on point clouds. In Proceedings of the European Conference on Computer Vision, Part II, pages 293-310. Springer, 2022. 2, 6", + "[10] Matej Kristan, Jiri Matas, Aleš Leonardis, Tomáš Vojíř, Roman Pflugfelder, Gustavo Fernández, Georg Nebehay, Fatih Porikli, and Luka Čehovin. A novel performance evaluation methodology for single-target trackers. IEEE Transactions on Pattern Analysis and Machine Intelligence, 38(11):2137-2155, 2016. 6", + "[11] H Kuang Chiu, A Prioletti, J Li, and J Bohg. Probabilistic 3D multi-object tracking for autonomous driving. ArXiv, vol. abs/2001.05673, 2020. 1", + "[12] Ze Liu, Zheng Zhang, Yue Cao, Han Hu, and Xin Tong. Group-free 3D object detection via transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2949-2958, 2021. 1", + "[13] Matthias Luber, Luciano Spinello, and Kai O Arras. People tracking in RGB-D data with on-line boosted target models. In 2011 IEEE/RSJ International Conference on Intelligent Robots and Systems, pages 3844-3849. IEEE, 2011. 2" + ], + "bbox": [ + 78, + 114, + 470, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[14] Jiageng Mao, Yujing Xue, Minzhe Niu, Haoyue Bai, Jiashi Feng, Xiaodan Liang, Hang Xu, and Chunjing Xu. Voxel transformer for 3D object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3164-3173, 2021. 1", + "[15] Ishan Misra, Rohit Girdhar, and Armand Joulin. An end-to-end transformer model for 3D object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2906-2917, 2021. 1, 3", + "[16] Ziqi Pang, Zhichao Li, and Naiyan Wang. Model-free vehicle tracking and state estimation in point cloud sequences. In IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 8075-8082, 2021. 6", + "[17] Alessandro Pieropan, Niklas Bergstrom, Masatoshi Ishikawa, and Hedvig Kjellstrom. Robust 3D tracking of unknown objects. In IEEE International Conference on Robotics and Automation (ICRA), pages 2410-2417. IEEE, 2015. 2", + "[18] Charles R Qi, Or Litany, Kaiming He, and Leonidas J Guibas. Deep Hough voting for 3D object detection in point clouds. In proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9277-9286, 2019. 1, 2, 5, 8", + "[19] Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. PointNet: Deep learning on point sets for 3D classification and segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 652-660, 2017. 3", + "[20] Haozhe Qi, Chen Feng, Zhiguo Cao, Feng Zhao, and Yang Xiao. P2B: Point-to-box network for 3D object tracking in point clouds. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6329-6338, 2020. 1, 2, 5, 6, 8", + "[21] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster R-CNN: Towards real-time object detection with region proposal networks. Advances in Neural Information Processing Systems, 28, 2015. 5", + "[22] Jiayao Shan, Sifan Zhou, Zheng Fang, and Yubo Cui. PTT: Point-track-transformer module for 3D single object tracking in point clouds. In IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 1310-1316. IEEE, 2021. 2, 6", + "[23] Luciano Spinello, Kai Arras, Rudolph Triebel, and Roland Siegwart. A layered approach to people detection in 3D range data. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 24, pages 1625-1630, 2010. 2", + "[24] Pei Sun, Henrik Kretzschmar, Xerxes Dotiwalla, Aurelien Chouard, Vijaysai Patnaik, Paul Tsui, James Guo, Yin Zhou, Yuning Chai, Benjamin Caine, et al. Scalability in perception for autonomous driving: Waymo open dataset. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2446-2454, 2020. 6", + "[25] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in Neural Information Processing Systems, 30, 2017. 2, 3", + "[26] Ning Wang, Wengang Zhou, Jie Wang, and Houqiang Li. Transformer meets tracker: Exploiting temporal context for" + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "1092", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "robust visual tracking. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1571-1580, 2021. 5", + "[27] Sinong Wang, Belinda Z Li, Madian Khabsa, Han Fang, and Hao Ma. Linformer: Self-attention with linear complexity. arXiv preprint arXiv:2006.04768, 2020. 7", + "[28] Yue Wang, Yongbin Sun, Ziwei Liu, Sanjay E Sarma, Michael M Bronstein, and Justin M Solomon. Dynamic graph CNN for learning on point clouds. ACM Transactions on Graphics (TOG), 38(5):1-12, 2019. 6", + "[29] Zhoutao Wang, Qian Xie, Yu-Kun Lai, Jing Wu, Kun Long, and Jun Wang. MLVSNet: Multi-level voting Siamese network for 3D visual tracking. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3101-3110, 2021. 1, 2, 6", + "[30] Tianwei Yin, Xingyi Zhou, and Philipp Krahenbuhl. Center-based 3D object detection and tracking. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11784-11793, 2021. 1", + "[31] Chaoda Zheng, Xu Yan, Jiantao Gao, Weibing Zhao, Wei Zhang, Zhen Li, and Shuguang Cui. Box-aware feature enhancement for single object tracking on point clouds. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13199-13208, 2021. 1, 2, 6", + "[32] Chaoda Zheng, Xu Yan, Haiming Zhang, Baoyuan Wang, Shenghui Cheng, Shuguang Cui, and Zhen Li. Beyond 3D Siamese tracking: A motion-centric paradigm for 3D single object tracking in point clouds. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8111-8120, 2022. 1, 2, 3, 5, 6", + "[33] Changqing Zhou, Zhipeng Luo, Yueru Luo, Tianrui Liu, Liang Pan, Zhongang Cai, Haiyu Zhao, and Shijian Lu. PTTR: Relational 3D point cloud object tracking with transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8531-8540, 2022. 1, 2, 6, 8" + ], + "bbox": [ + 78, + 90, + 468, + 598 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "1093", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2023/CXTrack_ Improving 3D Point Cloud Tracking With Contextual Information/c7b22be7-4a5c-4780-8a5a-1db50c1d60af_model.json b/2023/CXTrack_ Improving 3D Point Cloud Tracking With Contextual Information/c7b22be7-4a5c-4780-8a5a-1db50c1d60af_model.json new file mode 100644 index 0000000000000000000000000000000000000000..2edce443c0a6da78dbfe877d82fd29484ab9fbc4 --- /dev/null +++ b/2023/CXTrack_ Improving 3D Point Cloud Tracking With Contextual Information/c7b22be7-4a5c-4780-8a5a-1db50c1d60af_model.json @@ -0,0 +1,2035 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.1, + 0.131, + 0.872, + 0.152 + ], + "angle": 0, + "content": "CXTrack: Improving 3D Point Cloud Tracking with Contextual Information" + }, + { + "type": "text", + "bbox": [ + 0.202, + 0.18, + 0.774, + 0.2 + ], + "angle": 0, + "content": "Tian-Xing Xu\\(^{1}\\) Yuan-Chen Guo\\(^{1}\\) Yu-Kun Lai\\(^{2}\\) Song-Hai Zhang\\(^{1*}\\)" + }, + { + "type": "text", + "bbox": [ + 0.21, + 0.199, + 0.762, + 0.218 + ], + "angle": 0, + "content": "\\(^{1}\\) Tsinghua University, China \\(^{2}\\) Cardiff University, United Kingdom" + }, + { + "type": "text", + "bbox": [ + 0.157, + 0.219, + 0.81, + 0.234 + ], + "angle": 0, + "content": "\\(^{1}\\{xutx21@mails., guocycl9@mails., shz@}tsinghua.edu.cn\\) \\(^{2}\\)LaiY4@cardiff.ac.uk" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.269, + 0.314, + 0.285 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.302, + 0.474, + 0.651 + ], + "angle": 0, + "content": "3D single object tracking plays an essential role in many applications, such as autonomous driving. It remains a challenging problem due to the large appearance variation and the sparsity of points caused by occlusion and limited sensor capabilities. Therefore, contextual information across two consecutive frames is crucial for effective object tracking. However, points containing such useful information are often overlooked and cropped out in existing methods, leading to insufficient use of important contextual knowledge. To address this issue, we propose CXTrack, a novel transformer-based network for 3D object tracking, which exploits ConteXtual information to improve the tracking results. Specifically, we design a target-centric transformer network that directly takes point features from two consecutive frames and the previous bounding box as input to explore contextual information and implicitly propagate target cues. To achieve accurate localization for objects of all sizes, we propose a transformer-based localization head with a novel center embedding module to distinguish the target from distractors. Extensive experiments on three large-scale datasets, KITTI, nuScenes and Waymo Open Dataset, show that CXTrack achieves state-of-the-art tracking performance while running at 34 FPS." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.681, + 0.21, + 0.697 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.707, + 0.47, + 0.875 + ], + "angle": 0, + "content": "Single Object Tracking (SOT) has been a fundamental task in computer vision for decades, aiming to keep track of a specific target across a video sequence, given only its initial status. In recent years, with the development of 3D data acquisition devices, it has drawn increasing attention for using point clouds to solve various vision tasks such as object detection [7, 12, 14, 15, 18] and object tracking [20, 29, 31-33]. In particular, much progress has been made on point cloud-based object tracking for its huge potential in applications such as autonomous driving [11, 30]. However, it remains challenging due to the large appearance variation" + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.266, + 0.892, + 0.369 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.504, + 0.372, + 0.891, + 0.473 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.506, + 0.475, + 0.891, + 0.577 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.508, + 0.579, + 0.891, + 0.679 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.689, + 0.894, + 0.76 + ], + "angle": 0, + "content": "Figure 1. Comparison of various 3D SOT paradigms. Previous methods crop the target from the frames to specify the region of interest, which largely overlook contextual information around the target. On the contrary, our proposed CXTrack fully exploits contextual information to improve the tracking results." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.763, + 0.892, + 0.793 + ], + "angle": 0, + "content": "of the target and the sparsity of 3D point clouds caused by occlusion and limited sensor resolution." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.795, + 0.895, + 0.901 + ], + "angle": 0, + "content": "Existing 3D point cloud-based SOT methods can be categorized into three main paradigms, namely SC3D, P2B and motion-centric, as shown in Fig. 1. As a pioneering work, SC3D [6] crops the target from the previous frame, and compares the target template with a potentially large number of candidate patches generated from the current frame, which consumes much time. To address the effi" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.888, + 0.218, + 0.901 + ], + "angle": 0, + "content": "* corresponding author" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1084" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.473, + 0.485 + ], + "angle": 0, + "content": "ciency problem, P2B [20] takes the cropped target template from the previous frame as well as the complete search area in the current frame as input, propagates target cues into the search area and then adopts a 3D region proposal network [18] to predict the current bounding box. P2B reaches a balance between performance and speed. Therefore many follow-up works adopt the same paradigm [3, 8, 9, 22, 29, 31, 33]. However, both SC3D and P2D paradigms overlook the contextual information across two consecutive frames and rely entirely on the appearance of the target. As mentioned in previous work [32], these methods are sensitive to appearance variation caused by occlusions and tend to drift towards intra-class distractors. To this end, M2-Track [32] introduces a novel motion-centric paradigm, which directly takes point clouds from two frames without cropping as input, and then segments the target points from their surroundings. After that, these points are cropped and the current bounding box is estimated by explicitly modeling motion between the two frames. Hence, the motion-centric paradigm still works on cropped patches that lack contextual information in later localization. In short, none of these methods could fully utilize the contextual information around the target to predict the current bounding box, which may degrade tracking performance due to the existence of large appearance variation and widespread distractors." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.494, + 0.473, + 0.902 + ], + "angle": 0, + "content": "To address the above concerns, we propose a novel transformer-based tracker named CXTrack for 3D SOT, which exploits contextual information across two consecutive frames to improve the tracking performance. As shown in Fig. 1, different from paradigms commonly adopted by previous methods, CXTrack directly takes point clouds from the two consecutive frames as input, specifies the target of interest with the previous bounding box and predicts the current bounding box without any cropping, largely preserving contextual information. We first embed local geometric information of the two point clouds into point features using a shared backbone network. Then we integrate the targetness information into the point features according to the previous bounding box and adopt a target-centric transformer to propagate the target cues into the current frame while exploring contextual information in the surroundings of the target. After that, the enhanced point features are fed into a novel localization head named X-RPN to obtain the final target proposals. Specifically, X-RPN adopts a local transformer [25] to model point feature interactions within the target, which achieves a better balance between handling small and large objects compared with other localization heads. To distinguish the target from distractors, we incorporate a novel center embedding module into X-RPN, which embeds the relative target motion between two frames for explicit motion modeling. Extensive experiments on three popular tracking datasets demonstrate" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.137 + ], + "angle": 0, + "content": "that CXTrack significantly outperforms the current state-of-the-art methods by a large margin while running at real-time (34 FPS) on a single NVIDIA RTX3090 GPU." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.137, + 0.895, + 0.274 + ], + "angle": 0, + "content": "In short, our contributions can be summarized as: (1) a new paradigm for the real-time 3D SOT task, which fully exploits contextual information across consecutive frames to improve the tracking accuracy; (2) CXTrack: a transformer-based tracker that employs a target-centric transformer architecture to propagate targetness information and exploit contextual information; and (3) X-RPN: a localization head that is robust to intra-class distractors and achieves a good balance between small and large targets." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.287, + 0.642, + 0.303 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.312, + 0.895, + 0.478 + ], + "angle": 0, + "content": "Early methods [13, 17, 23] for the 3D SOT task mainly focus on RGB-D information and tend to adopt 2D Siamese networks used in 2D object tracking with additional depth maps. However, the changes in illumination and appearance may degrade the performance of these RGB-D methods. As a pioneering work in this area, SC3D [6] crops the target from the previous frame with the previous bounding box, and then computes the cosine similarity between the target template and a series of 3D target proposals sampled from the current frame using a Siamese backbone. The pipeline relies on heuristic sampling, which is very time-consuming." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.479, + 0.895, + 0.84 + ], + "angle": 0, + "content": "To address these issues, P2B [20] develops an end-to-end framework, which first employs a shared backbone to embed local geometry into point features, and then propagates target cues from the target template to the search area in the current frame. Finally, it adopts VoteNet [18] to generate 3D proposals and selects the proposal with the highest score as the target. P2B [20] reaches a balance between performance and efficiency, and many works follow the same paradigm. MLVSNet [29] aggregates information at multiple levels for more effective target localization. BAT [31] introduces a box-aware feature fusion module to enhance the correlation learning between the target template and the search area. V2B [8] proposes a voxel-to-BEV (Bird's Eye View) target localization network, which projects the point features into a dense BEV feature map to tackle the sparsity of point clouds. Inspired by the success of transformers [25], LTTR [3] adopts a transformer-based architecture to fuse features from two branches and propagate target cues. PTT [22] integrates a transformer module into the P2B architecture to refine point features. PTTR [33] introduces Point Relation Transformer for feature fusion and a light-weight Prediction Refinement Module for coarse-to-fine localization. ST-Net [9] develops an iterative coarse-to-fine correlation network for robust correlation learning." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.895, + 0.901 + ], + "angle": 0, + "content": "Although achieving promising results, the aforementioned methods crop the target from the previous frame using the given bounding box. This overlook of contextual information across two frames makes these methods sensitive" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1085" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.274 + ], + "angle": 0, + "content": "to appearance variations caused by commonly occurred occlusions and thus the results tend to drift towards intra-class distractors, as mentioned in M2-Track [32]. To this end, M2-Track introduces a motion-centric paradigm to handle the 3D SOT problem, which directly takes the point clouds from two consecutive frames as input without cropping. It first localizes the target in the two frames by target segmentation, and then adopts PointNet [19] to predict the relative target motion from the cropped target area that lacks contextual information. M2-Track could not fully utilize local geometric and contextual information for prediction, which may hinder precise bounding box regression." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.284, + 0.17, + 0.3 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.309, + 0.262, + 0.325 + ], + "angle": 0, + "content": "3.1. Problem Definition" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.333, + 0.473, + 0.558 + ], + "angle": 0, + "content": "Given the initial state of the target, single object tracking aims to localize the target in a dynamic scene frame by frame. The initial state in the first frame is given as the 3D bounding box of the target, which can be parameterized by its center coordinates \\((x,y,z)\\), orientation angle \\(\\theta\\) (around the up-axis, which is sufficient for most tracked objects staying on the ground) and sizes along each axis \\((w,l,h)\\). Since the tracking target has little change in size across frames even for non-rigid objects, we assume constant target size and only regress the translation offset \\((\\Delta x,\\Delta y,\\Delta z)\\) and the rotation angle \\((\\Delta \\theta)\\) between two consecutive frames to simplify the tracking task. By applying the translation and rotation to the 3D bounding box \\(\\mathcal{B}_{t - 1}\\in \\mathbb{R}^7\\) in the previous frame, we can compute the 3D bounding box \\(\\mathcal{B}_t\\in \\mathbb{R}^7\\) to localize the target in the current frame." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.56, + 0.473, + 0.684 + ], + "angle": 0, + "content": "Suppose the point clouds in two consecutive frames are denoted as \\(\\mathcal{P}_{t - 1}\\in \\mathbb{R}^{\\dot{N}_{t - 1}\\times 3}\\) and \\(\\mathcal{P}_t\\in \\mathbb{R}^{\\dot{N}_t\\times 3}\\), respectively, where \\(\\dot{N}_{t - 1}\\) and \\(\\dot{N}_t\\) are the numbers of points in the point clouds. We follow M2-Track [32] and encode the 3D bounding box \\(\\mathcal{B}_{t - 1}\\) as a targetness mask \\(\\dot{\\mathcal{M}}_{t - 1} = (m_{t - 1}^{1},m_{t - 1}^{2},\\dots ,m_{t - 1}^{\\dot{N}_{t - 1}})\\in \\mathbb{R}^{\\dot{N}_{t - 1}}\\) to indicate the target position, where the mask \\(m_{t - 1}^{i}\\) for the \\(i\\)-th point \\(p_{t - 1}^{i}\\) is defined as" + }, + { + "type": "equation", + "bbox": [ + 0.165, + 0.689, + 0.47, + 0.722 + ], + "angle": 0, + "content": "\\[\nm _ {t - 1} ^ {i} = \\left\\{ \\begin{array}{l l} 0 & p _ {t - 1} ^ {i} \\text {n o t i n} \\mathcal {B} _ {t - 1} \\\\ 1 & p _ {t - 1} ^ {i} \\text {i n} \\mathcal {B} _ {t - 1} \\end{array} \\right. \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.727, + 0.47, + 0.759 + ], + "angle": 0, + "content": "Thus, the 3D SOT task can be formalized as learning the following mapping" + }, + { + "type": "equation", + "bbox": [ + 0.13, + 0.764, + 0.469, + 0.782 + ], + "angle": 0, + "content": "\\[\n\\mathcal {F} \\left(\\mathcal {P} _ {t - 1}, \\dot {\\mathcal {M}} _ {t - 1}, \\mathcal {P} _ {t}\\right) \\mapsto (\\Delta x, \\Delta y, \\Delta z, \\Delta \\theta) \\tag {2}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.787, + 0.284, + 0.803 + ], + "angle": 0, + "content": "3.2. Overview of CXTrack" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.81, + 0.473, + 0.903 + ], + "angle": 0, + "content": "Following Eq. (2), we propose a network named CX-Track to improve tracking accuracy by fully exploiting contextual information across frames, and the overall design is illustrated in Fig. 2. We first apply a hierarchical feature learning architecture as the shared backbone to embed local geometric features of the point clouds into point" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.895, + 0.32 + ], + "angle": 0, + "content": "features. We use \\(N_{t - 1}\\) and \\(N_{t}\\) to denote the numbers of point features extracted by the backbone. For convenience of calculation, we create a targetness mask \\(\\dot{\\mathcal{M}}_t\\) and fill it with 0.5 as it is unknown. We then concatenate the point features and targetness masks of the two frames to get \\(\\mathcal{X} = \\mathcal{X}_{t - 1}\\oplus \\mathcal{X}_t\\in \\mathbb{R}^{N\\times C}\\) and \\(\\mathcal{M} = \\mathcal{M}_{t - 1}\\oplus \\mathcal{M}_t\\in \\mathbb{R}^{N\\times 1}\\) where \\(N = N_{t - 1} + N_{t}\\), \\(\\mathcal{M}_{t - 1}\\) and \\(\\mathcal{M}_t\\) are masks corresponding to point features, and extracted from \\(\\dot{\\mathcal{M}}_{t - 1}\\) and \\(\\dot{\\mathcal{M}}_t\\), and \\(C\\) is the number of channels for point features. We employ the target-centric transformer (Sec. 3.3) to integrate the targetness mask information into point features while exploring the contextual information across frames. Finally, we propose a novel localization network, named X-RPN (Sec. 3.4), to obtain the target proposals. The proposal with the highest targetness score is verified as the result." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.333, + 0.755, + 0.35 + ], + "angle": 0, + "content": "3.3. Target-Centric Transformer" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.359, + 0.895, + 0.617 + ], + "angle": 0, + "content": "Target-Centric Transformer aims to enhance the point features using the contextual information around the target while propagating the target cues from the previous frame to the current frame. It is composed of \\( N_{L} = 4 \\) identical layers stacked in series. Given the point features \\( \\mathcal{X}^{(k - 1)}\\in \\mathbb{R}^{N\\times C} \\) and the targetness mask \\( \\mathcal{M}^{(k - 1)} \\) from the \\( (k - 1) \\)-th layer as input \\( (\\mathcal{M}^{(0)} = \\mathcal{M} \\) and \\( \\mathcal{X}^{(0)} = \\mathcal{X}) \\), the \\( k \\)-th layer first models the interactions between any two points while integrating the targetness mask into point features using a modified self-attention operation, and then adopts Multi-Layer-Perceptrons (MLPs) to compute the new point features \\( \\mathcal{X}^{(k)} \\) as well as the refined targetness mask \\( \\mathcal{M}^{(k)} \\). Thus, the predicted targetness mask will be consistently refined layer by layer. Moreover, we found it beneficial to add an auxiliary loss by predicting a potential target center for each point via Hough voting, so each layer also applies a shared MLP to generate the potential target center \\( C^{(k)}\\in \\mathbb{R}^{N\\times 3} \\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.618, + 0.894, + 0.664 + ], + "angle": 0, + "content": "Formally, we first employ layer normalization [1] \\(\\mathrm{LN}(\\cdot)\\) before the self-attention mechanism [25] following the design of 3DETR [15], which can be written as" + }, + { + "type": "equation", + "bbox": [ + 0.633, + 0.677, + 0.892, + 0.696 + ], + "angle": 0, + "content": "\\[\n\\bar {X} = \\mathrm {L N} \\left(\\mathcal {X} ^ {(k - 1)}\\right) \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.711, + 0.894, + 0.757 + ], + "angle": 0, + "content": "Then, we add positional encodings (PE) of the coordinates to the normalized point features before feeding them into the self-attention operation" + }, + { + "type": "equation", + "bbox": [ + 0.619, + 0.771, + 0.891, + 0.789 + ], + "angle": 0, + "content": "\\[\nX _ {Q} = X _ {K} = \\bar {X} + \\mathrm {P E} \\tag {4}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.668, + 0.792, + 0.891, + 0.81 + ], + "angle": 0, + "content": "\\[\nX _ {V} = \\bar {X} \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.826, + 0.894, + 0.903 + ], + "angle": 0, + "content": "It is worth noting that we only adopt PE for the query and key branches, therefore each refined point feature is constrained to focus more on local geometry instead of its associated absolute position. Subsequently, the transformer layer employs a global self-attention operation to model the" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1086" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.078, + 0.087, + 0.889, + 0.232 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.075, + 0.24, + 0.896, + 0.298 + ], + "angle": 0, + "content": "Figure 2. The overall architecture of CXTrack. Given two consecutive point clouds and the 3D bounding box in the previous frame, CXTrack first embeds the local geometry into point features using the backbone. Then, CXTrack employs the target-centric transformer to explore contextual information across two frames and propagate the target cues to the current frame. Finally, the enhanced features are fed into a novel localization network named X-RPN to obtain high-quality proposals for verification." + }, + { + "type": "image", + "bbox": [ + 0.078, + 0.302, + 0.894, + 0.476 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.075, + 0.48, + 0.894, + 0.522 + ], + "angle": 0, + "content": "Figure 3. Comparison of various transformer layers to fuse the targetness mask and point features. We introduce three types of target-centric transformer layers, namely Vanilla, Semi-Dropout and Gated layer to integrate the targetness mask information into the point features while modeling intra-frame and inter-frame feature relationships." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.531, + 0.418, + 0.546 + ], + "angle": 0, + "content": "relationships between point features, formulated as" + }, + { + "type": "equation", + "bbox": [ + 0.086, + 0.557, + 0.469, + 0.575 + ], + "angle": 0, + "content": "\\[\n\\mathrm {M H A} \\left(X _ {Q}, X _ {K}, X _ {V}\\right) = \\operatorname {C o n c a t} \\left(\\text {h e a d} _ {1}, \\dots \\text {h e a d} _ {h}\\right) W ^ {O} \\tag {6}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.143, + 0.578, + 0.469, + 0.61 + ], + "angle": 0, + "content": "\\[\n\\text {w h e r e} \\operatorname {h e a d} _ {i} = \\operatorname {A t t n} \\left(X _ {Q} W _ {i} ^ {Q}, X _ {K} W _ {i} ^ {K}, \\bar {X} W _ {i} ^ {V}\\right), \\tag {7}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.132, + 0.614, + 0.469, + 0.649 + ], + "angle": 0, + "content": "\\[\n\\operatorname {A t t n} (Q, K, V) = \\operatorname {s o f t m a x} \\left(\\frac {Q K ^ {T}}{\\sqrt {d _ {k}}}\\right) V \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.658, + 0.47, + 0.749 + ], + "angle": 0, + "content": "Here, MHA indicates a multi-head attention, where the attention is applied in \\(h\\) subspaces before concatenation. The projections are implemented by parameter matrices \\(W_{i}^{Q}\\in \\mathbb{R}^{C\\times d_{k}}\\), \\(W_{i}^{K}\\in \\mathbb{R}^{C\\times d_{k}}\\), \\(W_{i}^{V}\\in \\mathbb{R}^{C\\times d_{v}}\\) and \\(W_{i}^{O}\\in \\mathbb{R}^{hd_{v}\\times C}\\), where \\(i\\) indicates the \\(i\\)-th subspace. The self-attention sublayer can be written as" + }, + { + "type": "equation", + "bbox": [ + 0.105, + 0.759, + 0.47, + 0.779 + ], + "angle": 0, + "content": "\\[\n\\widehat {X} = \\mathcal {X} ^ {(k - 1)} + \\operatorname {D r o p o u t} \\left(\\mathrm {M H A} \\left(X _ {Q}, X _ {K}, X _ {V}\\right)\\right) \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.79, + 0.471, + 0.851 + ], + "angle": 0, + "content": "In addition to the self-attention sublayer, each transformer layer also contains a fully connected feed-forward network to refine the point features. The final output of the \\(k\\)-th transformer layer is given by" + }, + { + "type": "equation", + "bbox": [ + 0.165, + 0.86, + 0.469, + 0.879 + ], + "angle": 0, + "content": "\\[\n\\mathcal {X} ^ {(k)} = \\widehat {X} + \\operatorname {D r o p o u t} (\\operatorname {F F N} (\\operatorname {L N} (\\widehat {X}))), \\tag {10}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.097, + 0.881, + 0.469, + 0.897 + ], + "angle": 0, + "content": "\\[\n\\text {w h e r e} \\operatorname {F F N} (x) = \\max \\left(0, x W _ {1} + b _ {1}\\right) W _ {2} + b _ {2}. \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.531, + 0.892, + 0.591 + ], + "angle": 0, + "content": "To integrate the targetness mask information into point features, we need to modify the classic transformer layer. We introduce three types of modified transformer layers in Fig. 3, namely Vanilla, Semi-Dropout and Gated layer." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.592, + 0.892, + 0.654 + ], + "angle": 0, + "content": "Vanilla. We project the input \\(\\mathcal{M}^{(k - 1)}\\) to mask embedding \\(\\mathrm{ME}\\in \\mathbb{R}^{N\\times C}\\) using a linear transformation. Following the design of positional encoding, we simply add ME to the input token embedding \\(X_{V}\\), which re-formulates Eq. (5) as" + }, + { + "type": "equation", + "bbox": [ + 0.641, + 0.668, + 0.891, + 0.684 + ], + "angle": 0, + "content": "\\[\nX _ {V} = \\bar {X} + \\mathrm {M E} \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.694, + 0.893, + 0.845 + ], + "angle": 0, + "content": "Semi-Dropout. Notably, the targetness mask information can only flow across layers along the attention path. For small objects which only have a few points to track, applying dropout to the mask embedding may discard the targetness information and lead to performance degradation. To this end, we separate the self-attention mechanism into a feature branch and a mask branch with shared attention weights, while only applying dropout to the refined point features. As shown in Fig. 3b, the self-attention sublayer in Eq. (9) is re-formulated as" + }, + { + "type": "equation", + "bbox": [ + 0.516, + 0.859, + 0.891, + 0.897 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\widehat {X} = \\mathcal {X} ^ {(k - 1)} + \\operatorname {D r o p o u t} (\\mathrm {M H A} (X _ {Q}, X _ {K}, \\overline {{X}})) \\\\ + \\mathrm {M H A} \\left(X _ {Q}, X _ {K}, \\mathrm {M E}\\right) \\tag {13} \\\\ \\end{array}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1087" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.1, + 0.089, + 0.445, + 0.295 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.299, + 0.471, + 0.37 + ], + "angle": 0, + "content": "Figure 4. The overall architecture of X-RPN. X-RPN adopts a local transformer to model point feature interaction within the target and aggregate local clues. It also incorporates a center embedding mechanism which embeds the relative target motion between two frames to distinguish the target from distractors." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.377, + 0.47, + 0.527 + ], + "angle": 0, + "content": "Gated. Inspired by the design of TrDimp [26], we introduce a gated mechanism into the self-attention sublayer to integrate the mask information. It has two parallel branches, namely mask transformation and feature transformation. For mask transformation, we first obtain the feature mask \\(\\overline{M} \\in \\mathbb{R}^{N \\times C}\\) by repeating the input point-wise mask \\(\\mathcal{M}^{(k - 1)} \\in \\mathbb{R}^{N \\times 1}\\) for \\(C\\) times. Then we can propagate the targetness cues to the current frame via adopting self-attention on the mask feature. The transformed mask serves as the gate matrix for the point features" + }, + { + "type": "equation", + "bbox": [ + 0.143, + 0.537, + 0.469, + 0.556 + ], + "angle": 0, + "content": "\\[\n\\widehat {X} _ {m} = \\mathrm {L N} (\\mathrm {M H A} (X _ {Q}, X _ {K}, \\bar {M}) \\otimes \\bar {X}). \\tag {14}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.566, + 0.469, + 0.626 + ], + "angle": 0, + "content": "For feature transformation, we first mask the point features to suppress feature activation in background areas, and then employ self-attention with a residual connection to model the relationships between features" + }, + { + "type": "equation", + "bbox": [ + 0.115, + 0.635, + 0.469, + 0.654 + ], + "angle": 0, + "content": "\\[\n\\widehat {X} _ {f} = \\mathrm {L N} (\\mathrm {M H A} (X _ {Q}, X _ {K}, \\bar {M} \\otimes \\bar {X}) + \\bar {X}). \\tag {15}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.664, + 0.469, + 0.709 + ], + "angle": 0, + "content": "As illustrated in Fig. 3c, we sum and normalize the output features \\(\\widehat{X}_m\\) and \\(\\widehat{X}_f\\) from the two branches. Eq. (9) can be re-formulated as" + }, + { + "type": "equation", + "bbox": [ + 0.2, + 0.718, + 0.469, + 0.737 + ], + "angle": 0, + "content": "\\[\n\\widehat {X} = \\operatorname {L N} \\left(\\widehat {X} _ {f} + \\widehat {X} _ {m}\\right). \\tag {16}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.748, + 0.469, + 0.824 + ], + "angle": 0, + "content": "Among the above three layers, we observe significant performance gain from using Semi-Dropout target-centric transformer layers (Sec. 4.3). Thus CXTrack employs Semi-Dropout layers to integrate targetness information while exploring contextual information across frames." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.832, + 0.17, + 0.847 + ], + "angle": 0, + "content": "3.4. X-RPN" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Previous works [20] indicate that individual point features can only capture limited local information, which may not be sufficient for precise bounding box regression. Thus" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.335 + ], + "angle": 0, + "content": "we develop a simple yet effective localization network, named X-RPN, which extends RPN [18] using local transformer and center embedding, as shown in Fig. 4. Different from previous works [8,20], X-RPN aggregates local clues from point features without downsampling or voxelization, thus avoiding information loss and reaching a good balance between handling large and small objects. Our intuition is that each point should only interact with points belonging to the same object to suppress irrelevant information. Given the point features \\(\\mathcal{X}^{(N_L)}\\), targetness mask \\(\\mathcal{M}^{(N_L)}\\) and target center \\(\\mathcal{C}^{(N_L)}\\) output by the target centric transformer, we first split them along the spatial dimension and only feed those belonging to the current frame into X-RPN, which is denoted as \\(\\widetilde{\\mathcal{X}}\\in \\mathbb{R}^{N_t\\times C}\\), \\(\\widetilde{\\mathcal{C}}\\in \\mathbb{R}^{N_t\\times 3}\\) and \\(\\widetilde{\\mathcal{M}}\\in \\mathbb{R}^{N_t\\times 1}\\), respectively. X-RPN first computes the neighborhood \\(\\mathcal{N}(p_i)\\) for each point \\(p_i\\) using its potential target center \\(c_{i}\\)" + }, + { + "type": "equation", + "bbox": [ + 0.586, + 0.344, + 0.892, + 0.37 + ], + "angle": 0, + "content": "\\[\n\\mathcal {N} \\left(p _ {i}\\right) = \\left\\{p _ {j} \\mid | | c _ {i} - c _ {j} | | _ {2} < r \\right\\} \\tag {17}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.378, + 0.891, + 0.484 + ], + "angle": 0, + "content": "Here \\( r \\) is a hyperparameter indicating the size of the neighborhood. Then X-RPN adopts the transformer architecture mentioned in Sec. 3.3 to aggregate local information, where each point only interacts with its neighborhood points to suppress noise. We remove the feed-forward network in the transformer layer because we observe that one layer is sufficient to generate high quality target proposals." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.484, + 0.892, + 0.651 + ], + "angle": 0, + "content": "To deal with intra-class distractors which are widespread in the scenes [32] especially for pedestrian tracking, we propose to combine the potential center information with the targetness mask. Our intuition lies in two folds. First, the tracking target keeps similar local geometry across two frames. Second, if the duration between two consecutive frames is sufficiently short, the displacement of the target is small. Therefore, we construct a Gaussian proposal-wise mask \\(\\mathcal{M}_c\\) to indicate the magnitude of the displacement of each proposal. Formally, for each point \\(p_i\\) with the predicted target center \\(c_i\\), the mask value \\(m_i^c \\in \\mathcal{M}_c\\) is" + }, + { + "type": "equation", + "bbox": [ + 0.612, + 0.66, + 0.891, + 0.691 + ], + "angle": 0, + "content": "\\[\nm _ {i} ^ {c} = \\exp \\left(- \\frac {\\left| \\left| c _ {i} - \\bar {c} \\right| \\right| _ {2} ^ {2}}{2 \\sigma^ {2}}\\right) \\tag {18}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.7, + 0.892, + 0.777 + ], + "angle": 0, + "content": "where \\(\\overline{c} \\in \\mathbb{R}^3\\) is the target center in the previous frame and \\(\\sigma\\) is a learnable or fixed scaling factor. We embed the target center mask \\(\\mathcal{M}_c\\) into the center embedding matrix \\(\\mathrm{CE} \\in \\mathbb{R}^{N \\times C}\\) using a linear transformation, and equally combine the mask embedding and the center embedding." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.785, + 0.652, + 0.8 + ], + "angle": 0, + "content": "3.5. Loss Functions" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.808, + 0.892, + 0.901 + ], + "angle": 0, + "content": "For the prediction \\(\\mathcal{M}^{(k)}\\) given by the \\(k\\)-th transformer layer, we adopt a standard cross entropy loss \\(\\mathcal{L}_{\\mathrm{cm}}^{(i)}\\). As for the potential target centers, we observe that it is difficult to regress precise centers for non-rigid objects such as pedestrians. Hence the predicted centers \\(\\mathcal{C}^{(k)}\\) are supervised by \\(L_{2}\\) loss for non-rigid objects, and by Huber loss [21] for" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1088" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.09, + 0.468, + 0.122 + ], + "angle": 0, + "content": "rigid objects. For the target center regression loss \\(\\mathcal{L}_{\\mathrm{cc}}^{(i)}\\), only points in the ground truth bounding box are supervised." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.122, + 0.468, + 0.228 + ], + "angle": 0, + "content": "Following previous works [20], proposals with predicted centers near the target center \\((< 0.3\\mathrm{m})\\) are considered as positives and those far away \\((>0.6\\mathrm{m})\\) are considered as negatives. Others are left unsupervised. The predicted targetness mask is supervised via standard cross-entropy loss \\(\\mathcal{L}_{\\mathrm{rm}}\\) and only the bounding box parameters of positive predictions are supervised by Huber (Smooth-\\(L_{1}\\)) loss \\(\\mathcal{L}_{\\mathrm{box}}\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.228, + 0.468, + 0.256 + ], + "angle": 0, + "content": "The overall loss is the weighted combination of the above loss terms" + }, + { + "type": "equation", + "bbox": [ + 0.104, + 0.263, + 0.468, + 0.304 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\gamma_ {1} \\sum_ {i = 1} ^ {N _ {L}} \\mathcal {L} _ {\\mathrm {c m}} ^ {(i)} + \\gamma_ {2} \\sum_ {i = 1} ^ {N _ {L}} \\mathcal {L} _ {\\mathrm {c c}} ^ {(i)} + \\gamma_ {3} \\mathcal {L} _ {\\mathrm {r m}} + \\mathcal {L} _ {\\mathrm {b o x}} \\tag {19}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.312, + 0.468, + 0.358 + ], + "angle": 0, + "content": "where \\(\\gamma_{1},\\gamma_{2}\\) and \\(\\gamma_{3}\\) are hyper-parameters. We empirically set \\(\\gamma_{1} = 0.2\\) \\(\\gamma_{2} = 1.0\\) \\(\\gamma_{3} = 1.5\\) for rigid objects and \\(\\gamma_{1} = 0.2\\) \\(\\gamma_{2} = 10.0\\) \\(\\gamma_{3} = 1.0\\) for non-rigid objects." + }, + { + "type": "title", + "bbox": [ + 0.076, + 0.369, + 0.208, + 0.386 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.393, + 0.175, + 0.409 + ], + "angle": 0, + "content": "4.1. Settings" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.416, + 0.468, + 0.568 + ], + "angle": 0, + "content": "Datasets. We compare CXTrack with previous state of the arts on three large-scale datasets: KITTI [5], nuScenes [2] and Waymo Open Dataset (WOD) [24]. KITTI contains 21 training video sequences and 29 test sequences. We follow previous work [6] and split the training sequences into three parts, 0-16 for training, 17-18 for validation and 19-20 for testing. For nuScenes, we use its validation split to evaluate our model, which contains 150 scenes. For WOD, we follow LiDAR-SOT [16] to evaluate our method, dividing it into three splits according to the sparsity of point clouds." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.568, + 0.468, + 0.673 + ], + "angle": 0, + "content": "Implementation Details. We adopt DGCNN [28] as the backbone network to extract local geometric information. In the X-RPN, we initialize the scaling parameter \\(\\sigma^2 = 10\\). Notably, we empirically fix \\(\\sigma\\) as a hyper-parameter for pedestrians and cyclists, and set it as a learnable parameter for cars and vans, since they may have larger motions. More details are provided in the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.673, + 0.468, + 0.808 + ], + "angle": 0, + "content": "Evaluation Metrics. We use Success and Precision defined in one pass evaluation [10] as evaluation metrics. Success denotes the Area Under Curve (AUC) for the plot showing the ratio of frames where the Intersection Over Union (IOU) between the predicted and ground-truth bounding boxes is larger than a threshold, ranging from 0 to 1. Precision is defined as the AUC of the plot showing the ratio of frames where the distance between their centers is within a threshold, from 0 to 2 meters." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.818, + 0.377, + 0.833 + ], + "angle": 0, + "content": "4.2. Comparison with State of the Arts" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.468, + 0.901 + ], + "angle": 0, + "content": "We make comprehensive comparisons on the KITTI dataset with previous state-of-the-art methods, including SC3D [6], P2B [20], 3DSiamRPN [4], LTTR [3], MLVS-Net [29], BAT [31], PTT [22], V2B [8], PTTR [33]," + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.09, + 0.892, + 0.145 + ], + "angle": 0, + "content": "Table 1. Comparisons with the state-of-the-art methods on KITTI dataset. \"Mean\" is the average result weighted by frame numbers. \"Blue\" and \"Bold\" denote previous and current best performance, respectively. Success/Precision are used for evaluation." + }, + { + "type": "table", + "bbox": [ + 0.508, + 0.148, + 0.888, + 0.337 + ], + "angle": 0, + "content": "
MethodCar (6424)Pedestrian (6088)Van (1248)Cyclist (308)Mean (14068)
SC3D41.3/57.918.2/37.840.4/47.041.5/70.431.2/48.5
P2B56.2/72.828.7/49.640.8/48.432.1/44.742.4/60.0
3DSiamRPN58.2/76.235.2/56.245.7/52.936.2/49.046.7/64.9
LTPR65.0/77.133.2/56.835.8/45.666.2/89.948.7/65.8
MLVSNet56.0/74.034.1/61.152.0/61.434.3/44.545.7/66.7
BAT60.5/77.742.1/70.152.4/67.033.7/45.451.2/72.8
PTT67.8/81.844.9/72.043.6/52.537.2/47.355.1/74.2
V2B70.5/81.348.3/73.550.1/58.040.8/49.758.4/75.2
PTTR65.2/77.450.9/81.652.5/61.865.1/90.557.9/78.1
STNet72.1/84.049.9/77.258.0/70.673.5/93.761.3/80.1
M2-Track65.5/80.861.5/88.253.8/70.773.2/93.562.9/83.4
CXTrack69.1/81.667.0/91.560.0/71.874.2/94.367.5/85.3
Improvement↓3.0/↓2.4↑5.5/↑3.3↑2.0/↑1.1↑0.7/↑0.6↑4.6/↑1.9
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.35, + 0.891, + 0.377 + ], + "angle": 0, + "content": "Table 2. Robustness under scenes that contain intra-class distractors on KITTI Pedestrian category." + }, + { + "type": "table", + "bbox": [ + 0.537, + 0.379, + 0.859, + 0.442 + ], + "angle": 0, + "content": "
MethodAll(6088)Distractor-Only(3917)Improvement
PTTR50.9/81.644.3/70.0↓6.6/↓11.6
STNet49.9/77.235.1/58.5↓14.8/↓18.7
M2-Track61.5/88.258.0/88.4↓3.5/↑0.2
CXTrack67.0/91.566.1/91.3↓0.9/↓0.3
" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.46, + 0.892, + 0.838 + ], + "angle": 0, + "content": "STNet [9] and M2-Track [32]. As illustrated in Tab. 1, CXTrack surpasses previous state-of-the-art methods, with a significant improvement of average Success and Precision. Notably, our method achieves the best performance under all categories, except for the Car, where voxel-based STNet [9] surpasses us by a minor margin (72.1/84.0 v.s. 69.1/81.6). Most vehicles have simple shapes and limited rotation angles, which fit well in voxels. We argue that voxelization provides a strong shape prior, thereby leading to performance gain for large objects with simple shapes. The lack of distractors for cars also makes our improvement over previous methods insignificant. However, our method has a significant improvement (67.0/91.5 v.s. 49.9/77.2) on the Pedestrian category. We claim that this stems from our special design to handle distractors and our better preservation for contextual information. Besides, compared with M2-Track [32], CXTrack obtains consistent performance gains on all categories especially on the Success metric, which demonstrates the importance of local geometry and contextual information. For further analysis on the impact of intraclass distractors, we manually pick out scenes that contain Pedestrian distractors from the KITTI test split and then evaluate different methods on these scenes. As shown in Tab. 2, both M2-Track and CXTrack are robust to distractors, while CXTrack can make more accurate predictions." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.892, + 0.899 + ], + "angle": 0, + "content": "To verify the generalization ability of our method, we follow previous methods [8, 9] and test the KITTI pretrained model on nuScenes and WOD. The comparison results on WOD are shown in Tab. 3. It can be seen that our" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "1089" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.271, + 0.09, + 0.7, + 0.103 + ], + "angle": 0, + "content": "Table 3. Comparison with state of the arts on Waymo Open Dataset." + }, + { + "type": "table", + "bbox": [ + 0.102, + 0.104, + 0.868, + 0.213 + ], + "angle": 0, + "content": "
MethodVehicle(185731)Pedestrian(241752)Mean(427483)
EasyMediumHardMeanEasyMediumHardMean
P2B57.1/65.452.0/60.747.9/58.552.6/61.718.1/30.817.8/30.017.7/29.317.9/30.133.0/43.8
BAT61.0/68.353.3/60.948.9/57.854.7/62.719.3/32.617.8/29.817.2/28.318.2/30.334.1/44.4
V2B64.5/71.555.1/63.252.0/62.057.6/65.927.9/43.922.5/36.220.1/33.123.7/37.938.4/50.1
STNet65.9/72.757.5/66.054.6/64.759.7/68.029.2/45.324.7/38.222.2/35.825.5/39.940.4/52.1
CXTrack63.9/71.154.2/62.752.1/63.757.1/66.135.4/55.329.7/47.926.3/44.430.7/49.442.2/56.7
Improvement↓2.0/↓1.6↓3.3/↓3.3↓3.5/↓1.0↓2.6/↓1.9↑6.2/↑10.0↑5.0/↑9.7↑4.1/↑8.6↑5.2/↑9.5↑1.8/↑4.6
" + }, + { + "type": "table_caption", + "bbox": [ + 0.099, + 0.222, + 0.448, + 0.234 + ], + "angle": 0, + "content": "Table 4. Comparison with state of the arts on nuScenes." + }, + { + "type": "table", + "bbox": [ + 0.086, + 0.236, + 0.465, + 0.352 + ], + "angle": 0, + "content": "
MethodCar (15578)Pedestrian (8019)Van (3710)Cyclist (501)Mean (27808)
SC3D25.0/27.114.2/16.225.7/21.917.0/18.221.8/23.1
P2B27.0/29.215.9/22.021.5/16.220.0/26.422.9/25.3
BAT22.5/24.117.3/24.519.3/15.817.0/18.820.5/23.0
V2B31.3/35.117.3/23.421.7/16.722.2/19.125.8/29.0
STNet32.2/36.119.1/27.222.3/16.821.2/29.226.9/30.8
CXTrack29.6/33.420.4/32.927.6/20.818.5/26.826.5/31.5
Improvement↓2.6/↓2.7↑1.3/↑5.7↑1.9/↓1.1↓3.7/↓2.4↓0.4/↑0.7
" + }, + { + "type": "table_caption", + "bbox": [ + 0.108, + 0.363, + 0.438, + 0.376 + ], + "angle": 0, + "content": "Table 5. Efficiency analysis of different components." + }, + { + "type": "table", + "bbox": [ + 0.121, + 0.378, + 0.432, + 0.453 + ], + "angle": 0, + "content": "
ComponentFLOPs#ParamsInfer Speed
backbone3.18G1.3M8.5ms
transformer1.28G14.7M10.9ms
X-RPN0.17G2.3M3.0ms
pre/postprocess--6.8ms
CXTrack4.63G18.3M29.2ms(34FPS)
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.469, + 0.47, + 0.605 + ], + "angle": 0, + "content": "method outperforms others in terms of average Success and Precision with a clear margin. Notably, KITTI and WOD data are captured by 64-beam LiDARs, while nuScenes data are captured by 32-beam LiDARs. Thus it is more challenging to generalize the pretrained model on the nuScenes dataset. As shown in Tab. 4, our method achieves comparable performance on the nuScenes dataset. In short, CXTrack not only achieves a good balance between small objects and large objects, but also generalizes well to unseen scenes." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.606, + 0.469, + 0.756 + ], + "angle": 0, + "content": "We also visualize the tracking results for qualitative comparisons. As shown in Fig. 5, CXTrack achieves good accuracy in scenes with both sparse and dense point clouds on both categories. In the sparse cases (left), previous methods drift towards intra-class distractors due to large appearance variations caused by occlusions, while only our method keeps track of the target, thanks to the sufficient use of contextual information. In the dense cases (right), our method can track the target more accurately than M2-Track by leveraging local geometric information." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.758, + 0.47, + 0.835 + ], + "angle": 0, + "content": "We report the efficiency of different components in Tab. 5. It can be observed that the target-centric transformer is the bottleneck of CXTrack during inference. We can replace the vanilla self-attention in CXTrack with linear attention such as linformer [27] for further speedup." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.846, + 0.242, + 0.861 + ], + "angle": 0, + "content": "4.3. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.469, + 0.9 + ], + "angle": 0, + "content": "To validate the effectiveness of several design choices in CXTrack, we conduct ablation studies on the KITTI dataset." + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.221, + 0.892, + 0.276 + ], + "angle": 0, + "content": "Table 6. Ablation studies of different components of the target-centric transformer. \"Cx\" refers to contextual information, \"M\" refers to the cascaded targetness mask prediction and \"C\" refers to the auxiliary target center regression branches." + }, + { + "type": "table", + "bbox": [ + 0.504, + 0.281, + 0.889, + 0.34 + ], + "angle": 0, + "content": "
CxMCCarPedestrianVanCyclistMean
✓62.5/74.260.6/87.058.3/71.472.0/93.361.5/79.9
✓✓67.4/80.263.9/89.057.8/70.872.7/93.865.1/83.5
✓✓59.7/73.6†51.8/81.659.9/71.571.7/93.256.6/77.3
✓✓✓69.1/81.667.0/91.560.0/71.874.2/94.367.5/85.3
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.623, + 0.34, + 0.771, + 0.351 + ], + "angle": 0, + "content": "†: unstable training process" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.356, + 0.892, + 0.412 + ], + "angle": 0, + "content": "Table 7. Ablation studies of different transformer layers on KITTI. \"V\" refers to the vanilla transformer layer and \"G\" refers to the gated transformer layer. \"S\" represents the semi-dropout transformer layer which is adopted in our proposed CXTrack." + }, + { + "type": "table", + "bbox": [ + 0.504, + 0.417, + 0.889, + 0.473 + ], + "angle": 0, + "content": "
CarPedestrianVanCyclistMean
V68.8/80.462.9/87.857.2/69.672.7/94.265.3/82.9
G64.8/76.964.7/91.156.2/70.570.6/93.464.1/82.8
S69.1/81.667.0/91.560.0/71.874.2/94.367.5/85.3
" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.493, + 0.892, + 0.749 + ], + "angle": 0, + "content": "Components of Target-Centric Transformer. Tab. 6 presents ablation studies of different components of transformer to gain a better understanding of its designs. We crop the input point cloud \\(\\mathcal{P}_{t-1}\\) using \\(\\mathcal{B}_{t-1}\\) to ablate contextual information in the previous frame. We can observe significant performance drop when not using contextual information, especially on Car and Pedestrian. For Car, it suffers from heavy occlusions (Fig. 5), while pedestrian distractors are widespread in the scene. We also find that removing context leads to unstable training on Car. We presume that the lack of supervised signals to tell the model what not to attend may confuse the model and introduce noise in training. For the cascaded targetness mask prediction and auxiliary target center regression, removing either of them leads to an obvious decline on terms of average metrics. We argue that the auxiliary regression loss can increase the feature similarities of points belonging to the same object." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Target-Centric Transformer Layer. Tab. 7 shows the impact of different target-centric transformer layers. Semi-Dropout achieves better performance than Vanilla, especially on Pedestrian. Small objects often consist of fewer points, hence applying dropout directly on the targetness information in training may confuse the network and lead to sub-optimal results. Gated relies entirely on the predicted targetness mask to modulate the amount of exposure for input features, which may suffer from information loss when the targetness mask is not accurate enough." + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1090" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.084, + 0.09, + 0.49, + 0.278 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.493, + 0.091, + 0.885, + 0.278 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.212, + 0.28, + 0.757, + 0.295 + ], + "angle": 0, + "content": "Figure 5. Visualization results. Left: Sparse cases in KITTI. Right: Dense cases in KITTI." + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.301, + 0.471, + 0.342 + ], + "angle": 0, + "content": "Table 8. Ablation studies of various localization heads on KITTI. \"X-RPN\\C\" indicates our proposed localization head X-RPN without center embedding." + }, + { + "type": "table", + "bbox": [ + 0.081, + 0.344, + 0.468, + 0.418 + ], + "angle": 0, + "content": "
HeadCarPedestrianVanCyclistMean
PRM [33]66.5/77.462.2/86.852.9/64.972.5/93.863.6/80.7
RPN [18]64.1/76.959.8/88.355.0/65.668.2/92.461.5/81.2
V2B [8]70.5/82.660.1/86.758.0/69.870.5/93.364.9/83.5
X-RPN\\C67.8/80.365.5/89.559.9/72.172.6/94.166.2/83.9
X-RPN69.1/81.667.0/91.560.0/71.874.2/94.367.5/85.3
" + }, + { + "type": "image", + "bbox": [ + 0.08, + 0.429, + 0.264, + 0.524 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.267, + 0.429, + 0.468, + 0.524 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.525, + 0.47, + 0.567 + ], + "angle": 0, + "content": "Figure 6. Representative examples of attention maps in the transformer. Target-centric transformer attends to objects that have similar geometry." + }, + { + "type": "image", + "bbox": [ + 0.08, + 0.578, + 0.469, + 0.677 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.679, + 0.469, + 0.708 + ], + "angle": 0, + "content": "Figure 7. Visualization of ablation study. Center embedding can benefit object tracking in challenging scenes with distractors." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.72, + 0.47, + 0.901 + ], + "angle": 0, + "content": "X-RPN. We replace X-RPN with other alternatives [8, 20, 33] and report the comparison results in Tab. 8. Although the V2B head achieves better performance than X-RPN on the Car category, it fails to track small objects such as pedestrians effectively due to intra-class distractors and inevitable information loss brought in by voxelization. It is also worth noting that we observe a performance drop without center embedding, especially on the Pedestrian category, for which distractors are more commonly seen. To explore the effectiveness of the center embedding, we visualize the attention map of the last transformer layer in Fig. 6. We observe that the transformer alone can attend to regions with similar ge" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.303, + 0.892, + 0.391 + ], + "angle": 0, + "content": "ometry to the target, but fails to distinguish the target from distractors. As shown in Fig. 7, with the help of the center embedding, the network precisely keeps track of the target. In short, X-RPN achieves a good balance between large and small objects, and effectively alleviates the distractor problem." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.402, + 0.64, + 0.417 + ], + "angle": 0, + "content": "4.4. Failure Cases" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.425, + 0.893, + 0.562 + ], + "angle": 0, + "content": "Although CXTrack is robust to intra-class distractors, it fails to predict accurate orientation of the target when the point clouds are too sparse to capture informative local geometry or when large appearance variations occur, as shown in Fig. 7. Besides, the center embedding directly encodes the displacement of target center into features, so our model may suffer from performance degradation if trained with \\(2\\mathrm{Hz}\\) data and tested with \\(10\\mathrm{Hz}\\) data because the scale of the displacement differs significantly." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.574, + 0.619, + 0.589 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.599, + 0.892, + 0.825 + ], + "angle": 0, + "content": "We revisit existing paradigms for the 3D SOT task and propose a new paradigm to fully exploit contextual information across frames, which is largely overlooked by previous methods. Following this paradigm, we design a novel transformer-based network named CXTrack, which employs a target-centric transformer to explore contextual information and model intra-frame and inter-frame feature relationships. We also introduce a localization head named X-RPN to obtain high-quality proposals for objects of all sizes, as well as a center embedding module to distinguish the target from distractors. Extensive experiments show that CXTrack significantly outperforms previous state-of-the-arts, and is robust to distractors. We hope our work can promote further exploitations in this task by showing the necessity to explore contextual information for more robust predictions." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.826, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Acknowledgment The authors thank Jiahui Huang for his discussions. This work was supported by the Natural Science Foundation of China (Project Number 61832016), and Tsinghua-Tencent Joint Laboratory for Internet Innovation Technology." + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "1091" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.175, + 0.107 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.116, + 0.47, + 0.156 + ], + "angle": 0, + "content": "[1] Jimmy Lei Ba, Jamie Ryan Kiros, and Geoffrey E Hinton. Layer normalization. arXiv preprint arXiv:1607.06450, 2016.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.16, + 0.472, + 0.243 + ], + "angle": 0, + "content": "[2] Holger Caesar, Varun Bankiti, Alex H Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuScenes: A multimodal dataset for autonomous driving. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11621-11631, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.245, + 0.471, + 0.286 + ], + "angle": 0, + "content": "[3] Yubo Cui, Zheng Fang, Jiayao Shan, Zuoxu Gu, and Sifan Zhou. 3D object tracking with transformer. arXiv preprint arXiv:2110.14921, 2021. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.288, + 0.471, + 0.343 + ], + "angle": 0, + "content": "[4] Zheng Fang, Sifan Zhou, Yubo Cui, and Sebastian Scherer. 3D-SiamRPN: An end-to-end learning method for real-time 3D single object tracking using raw point cloud. IEEE Sensors Journal, 21(4):4995-5011, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.346, + 0.471, + 0.401 + ], + "angle": 0, + "content": "[5] Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we ready for autonomous driving? the KITTI vision benchmark suite. In IEEE Conference on Computer Vision and Pattern Recognition, pages 3354-3361. IEEE, 2012. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.403, + 0.471, + 0.459 + ], + "angle": 0, + "content": "[6] Silvio Giancola, Jesus Zarzar, and Bernard Ghanem. Leveraging shape completion for 3D Siamese tracking. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1359-1368, 2019. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.46, + 0.471, + 0.528 + ], + "angle": 0, + "content": "[7] Chenhang He, Ruihuang Li, Shuai Li, and Lei Zhang. Voxel set transformer: A set-to-set approach to 3D object detection from point clouds. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8417-8427, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.531, + 0.471, + 0.586 + ], + "angle": 0, + "content": "[8] Le Hui, Lingpeng Wang, Mingmei Cheng, Jin Xie, and Jian Yang. 3D Siamese voxel-to-BEV tracker for sparse point clouds. Advances in Neural Information Processing Systems, 34:28714-28727, 2021. 2, 5, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.589, + 0.471, + 0.657 + ], + "angle": 0, + "content": "[9] Le Hui, Lingpeng Wang, Linghua Tang, Kaihao Lan, Jin Xie, and Jian Yang. 3D Siamese transformer network for single object tracking on point clouds. In Proceedings of the European Conference on Computer Vision, Part II, pages 293-310. Springer, 2022. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.66, + 0.47, + 0.741 + ], + "angle": 0, + "content": "[10] Matej Kristan, Jiri Matas, Aleš Leonardis, Tomáš Vojíř, Roman Pflugfelder, Gustavo Fernández, Georg Nebehay, Fatih Porikli, and Luka Čehovin. A novel performance evaluation methodology for single-target trackers. IEEE Transactions on Pattern Analysis and Machine Intelligence, 38(11):2137-2155, 2016. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.745, + 0.47, + 0.786 + ], + "angle": 0, + "content": "[11] H Kuang Chiu, A Prioletti, J Li, and J Bohg. Probabilistic 3D multi-object tracking for autonomous driving. ArXiv, vol. abs/2001.05673, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.789, + 0.47, + 0.843 + ], + "angle": 0, + "content": "[12] Ze Liu, Zheng Zhang, Yue Cao, Han Hu, and Xin Tong. Group-free 3D object detection via transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2949-2958, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.846, + 0.47, + 0.901 + ], + "angle": 0, + "content": "[13] Matthias Luber, Luciano Spinello, and Kai O Arras. People tracking in RGB-D data with on-line boosted target models. In 2011 IEEE/RSJ International Conference on Intelligent Robots and Systems, pages 3844-3849. IEEE, 2011. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.116, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.162 + ], + "angle": 0, + "content": "[14] Jiageng Mao, Yujing Xue, Minzhe Niu, Haoyue Bai, Jiashi Feng, Xiaodan Liang, Hang Xu, and Chunjing Xu. Voxel transformer for 3D object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3164-3173, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.163, + 0.892, + 0.218 + ], + "angle": 0, + "content": "[15] Ishan Misra, Rohit Girdhar, and Armand Joulin. An end-to-end transformer model for 3D object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2906-2917, 2021. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.219, + 0.892, + 0.273 + ], + "angle": 0, + "content": "[16] Ziqi Pang, Zhichao Li, and Naiyan Wang. Model-free vehicle tracking and state estimation in point cloud sequences. In IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 8075-8082, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.274, + 0.892, + 0.342 + ], + "angle": 0, + "content": "[17] Alessandro Pieropan, Niklas Bergstrom, Masatoshi Ishikawa, and Hedvig Kjellstrom. Robust 3D tracking of unknown objects. In IEEE International Conference on Robotics and Automation (ICRA), pages 2410-2417. IEEE, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.344, + 0.892, + 0.412 + ], + "angle": 0, + "content": "[18] Charles R Qi, Or Litany, Kaiming He, and Leonidas J Guibas. Deep Hough voting for 3D object detection in point clouds. In proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9277-9286, 2019. 1, 2, 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.413, + 0.892, + 0.481 + ], + "angle": 0, + "content": "[19] Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. PointNet: Deep learning on point sets for 3D classification and segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 652-660, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.483, + 0.892, + 0.551 + ], + "angle": 0, + "content": "[20] Haozhe Qi, Chen Feng, Zhiguo Cao, Feng Zhao, and Yang Xiao. P2B: Point-to-box network for 3D object tracking in point clouds. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6329-6338, 2020. 1, 2, 5, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.553, + 0.892, + 0.607 + ], + "angle": 0, + "content": "[21] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster R-CNN: Towards real-time object detection with region proposal networks. Advances in Neural Information Processing Systems, 28, 2015. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.609, + 0.892, + 0.676 + ], + "angle": 0, + "content": "[22] Jiayao Shan, Sifan Zhou, Zheng Fang, and Yubo Cui. PTT: Point-track-transformer module for 3D single object tracking in point clouds. In IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 1310-1316. IEEE, 2021. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.678, + 0.892, + 0.734 + ], + "angle": 0, + "content": "[23] Luciano Spinello, Kai Arras, Rudolph Triebel, and Roland Siegwart. A layered approach to people detection in 3D range data. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 24, pages 1625-1630, 2010. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.735, + 0.892, + 0.816 + ], + "angle": 0, + "content": "[24] Pei Sun, Henrik Kretzschmar, Xerxes Dotiwalla, Aurelien Chouard, Vijaysai Patnaik, Paul Tsui, James Guo, Yin Zhou, Yuning Chai, Benjamin Caine, et al. Scalability in perception for autonomous driving: Waymo open dataset. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2446-2454, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.818, + 0.892, + 0.872 + ], + "angle": 0, + "content": "[25] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in Neural Information Processing Systems, 30, 2017. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.874, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[26] Ning Wang, Wengang Zhou, Jie Wang, and Houqiang Li. Transformer meets tracker: Exploiting temporal context for" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1092" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.092, + 0.47, + 0.133 + ], + "angle": 0, + "content": "robust visual tracking. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1571-1580, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.136, + 0.47, + 0.177 + ], + "angle": 0, + "content": "[27] Sinong Wang, Belinda Z Li, Madian Khabsa, Han Fang, and Hao Ma. Linformer: Self-attention with linear complexity. arXiv preprint arXiv:2006.04768, 2020. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.179, + 0.469, + 0.233 + ], + "angle": 0, + "content": "[28] Yue Wang, Yongbin Sun, Ziwei Liu, Sanjay E Sarma, Michael M Bronstein, and Justin M Solomon. Dynamic graph CNN for learning on point clouds. ACM Transactions on Graphics (TOG), 38(5):1-12, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.235, + 0.469, + 0.303 + ], + "angle": 0, + "content": "[29] Zhoutao Wang, Qian Xie, Yu-Kun Lai, Jing Wu, Kun Long, and Jun Wang. MLVSNet: Multi-level voting Siamese network for 3D visual tracking. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3101-3110, 2021. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.305, + 0.47, + 0.36 + ], + "angle": 0, + "content": "[30] Tianwei Yin, Xingyi Zhou, and Philipp Krahenbuhl. Center-based 3D object detection and tracking. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11784-11793, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.362, + 0.469, + 0.431 + ], + "angle": 0, + "content": "[31] Chaoda Zheng, Xu Yan, Jiantao Gao, Weibing Zhao, Wei Zhang, Zhen Li, and Shuguang Cui. Box-aware feature enhancement for single object tracking on point clouds. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13199-13208, 2021. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.433, + 0.47, + 0.515 + ], + "angle": 0, + "content": "[32] Chaoda Zheng, Xu Yan, Haiming Zhang, Baoyuan Wang, Shenghui Cheng, Shuguang Cui, and Zhen Li. Beyond 3D Siamese tracking: A motion-centric paradigm for 3D single object tracking in point clouds. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8111-8120, 2022. 1, 2, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.517, + 0.469, + 0.599 + ], + "angle": 0, + "content": "[33] Changqing Zhou, Zhipeng Luo, Yueru Luo, Tianrui Liu, Liang Pan, Zhongang Cai, Haiyu Zhao, and Shijian Lu. PTTR: Relational 3D point cloud object tracking with transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8531-8540, 2022. 1, 2, 6, 8" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.599 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1093" + } + ] +] \ No newline at end of file diff --git a/2023/CXTrack_ Improving 3D Point Cloud Tracking With Contextual Information/c7b22be7-4a5c-4780-8a5a-1db50c1d60af_origin.pdf b/2023/CXTrack_ Improving 3D Point Cloud Tracking With Contextual Information/c7b22be7-4a5c-4780-8a5a-1db50c1d60af_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..f7f69ee3cfad5dad126fb47a7b20bf3e2db61101 --- /dev/null +++ b/2023/CXTrack_ Improving 3D Point Cloud Tracking With Contextual Information/c7b22be7-4a5c-4780-8a5a-1db50c1d60af_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:affb267ced0b7692a3cc468bd4c3bfe7229a2f630e5c1065bced8d1b2d465b53 +size 1544728 diff --git a/2023/CXTrack_ Improving 3D Point Cloud Tracking With Contextual Information/full.md b/2023/CXTrack_ Improving 3D Point Cloud Tracking With Contextual Information/full.md new file mode 100644 index 0000000000000000000000000000000000000000..7b9a5dac11067aeae1e82708401ccb3e9ac2fa23 --- /dev/null +++ b/2023/CXTrack_ Improving 3D Point Cloud Tracking With Contextual Information/full.md @@ -0,0 +1,333 @@ +# CXTrack: Improving 3D Point Cloud Tracking with Contextual Information + +Tian-Xing Xu $^{1}$ Yuan-Chen Guo $^{1}$ Yu-Kun Lai $^{2}$ Song-Hai Zhang $^{1*}$ + +$^{1}$ Tsinghua University, China $^{2}$ Cardiff University, United Kingdom + +$^{1}\{xutx21@mails., guocycl9@mails., shz@}tsinghua.edu.cn$ $^{2}$ LaiY4@cardiff.ac.uk + +# Abstract + +3D single object tracking plays an essential role in many applications, such as autonomous driving. It remains a challenging problem due to the large appearance variation and the sparsity of points caused by occlusion and limited sensor capabilities. Therefore, contextual information across two consecutive frames is crucial for effective object tracking. However, points containing such useful information are often overlooked and cropped out in existing methods, leading to insufficient use of important contextual knowledge. To address this issue, we propose CXTrack, a novel transformer-based network for 3D object tracking, which exploits ConteXtual information to improve the tracking results. Specifically, we design a target-centric transformer network that directly takes point features from two consecutive frames and the previous bounding box as input to explore contextual information and implicitly propagate target cues. To achieve accurate localization for objects of all sizes, we propose a transformer-based localization head with a novel center embedding module to distinguish the target from distractors. Extensive experiments on three large-scale datasets, KITTI, nuScenes and Waymo Open Dataset, show that CXTrack achieves state-of-the-art tracking performance while running at 34 FPS. + +# 1. Introduction + +Single Object Tracking (SOT) has been a fundamental task in computer vision for decades, aiming to keep track of a specific target across a video sequence, given only its initial status. In recent years, with the development of 3D data acquisition devices, it has drawn increasing attention for using point clouds to solve various vision tasks such as object detection [7, 12, 14, 15, 18] and object tracking [20, 29, 31-33]. In particular, much progress has been made on point cloud-based object tracking for its huge potential in applications such as autonomous driving [11, 30]. However, it remains challenging due to the large appearance variation + +![](images/77c8f1b7e853edc39d17d9d5dfaa6d87a85734be4e4583ceef1649bc2a212abf.jpg) + +![](images/7a1c931e1bcc9e0e957af9c111cdb8ef9d5c78f9ea73ca221e92b24c3e776ae1.jpg) + +![](images/fc593d0d67219ed4cf6f62564c4def8a65a2a62ff521c9e74d0b28369e62bb15.jpg) + +![](images/d05a41c5882a19cec4a751cff59ff44cb4bc1ec2b25308a70ec07439396b86be.jpg) +Figure 1. Comparison of various 3D SOT paradigms. Previous methods crop the target from the frames to specify the region of interest, which largely overlook contextual information around the target. On the contrary, our proposed CXTrack fully exploits contextual information to improve the tracking results. + +of the target and the sparsity of 3D point clouds caused by occlusion and limited sensor resolution. + +Existing 3D point cloud-based SOT methods can be categorized into three main paradigms, namely SC3D, P2B and motion-centric, as shown in Fig. 1. As a pioneering work, SC3D [6] crops the target from the previous frame, and compares the target template with a potentially large number of candidate patches generated from the current frame, which consumes much time. To address the effi + +ciency problem, P2B [20] takes the cropped target template from the previous frame as well as the complete search area in the current frame as input, propagates target cues into the search area and then adopts a 3D region proposal network [18] to predict the current bounding box. P2B reaches a balance between performance and speed. Therefore many follow-up works adopt the same paradigm [3, 8, 9, 22, 29, 31, 33]. However, both SC3D and P2D paradigms overlook the contextual information across two consecutive frames and rely entirely on the appearance of the target. As mentioned in previous work [32], these methods are sensitive to appearance variation caused by occlusions and tend to drift towards intra-class distractors. To this end, M2-Track [32] introduces a novel motion-centric paradigm, which directly takes point clouds from two frames without cropping as input, and then segments the target points from their surroundings. After that, these points are cropped and the current bounding box is estimated by explicitly modeling motion between the two frames. Hence, the motion-centric paradigm still works on cropped patches that lack contextual information in later localization. In short, none of these methods could fully utilize the contextual information around the target to predict the current bounding box, which may degrade tracking performance due to the existence of large appearance variation and widespread distractors. + +To address the above concerns, we propose a novel transformer-based tracker named CXTrack for 3D SOT, which exploits contextual information across two consecutive frames to improve the tracking performance. As shown in Fig. 1, different from paradigms commonly adopted by previous methods, CXTrack directly takes point clouds from the two consecutive frames as input, specifies the target of interest with the previous bounding box and predicts the current bounding box without any cropping, largely preserving contextual information. We first embed local geometric information of the two point clouds into point features using a shared backbone network. Then we integrate the targetness information into the point features according to the previous bounding box and adopt a target-centric transformer to propagate the target cues into the current frame while exploring contextual information in the surroundings of the target. After that, the enhanced point features are fed into a novel localization head named X-RPN to obtain the final target proposals. Specifically, X-RPN adopts a local transformer [25] to model point feature interactions within the target, which achieves a better balance between handling small and large objects compared with other localization heads. To distinguish the target from distractors, we incorporate a novel center embedding module into X-RPN, which embeds the relative target motion between two frames for explicit motion modeling. Extensive experiments on three popular tracking datasets demonstrate + +that CXTrack significantly outperforms the current state-of-the-art methods by a large margin while running at real-time (34 FPS) on a single NVIDIA RTX3090 GPU. + +In short, our contributions can be summarized as: (1) a new paradigm for the real-time 3D SOT task, which fully exploits contextual information across consecutive frames to improve the tracking accuracy; (2) CXTrack: a transformer-based tracker that employs a target-centric transformer architecture to propagate targetness information and exploit contextual information; and (3) X-RPN: a localization head that is robust to intra-class distractors and achieves a good balance between small and large targets. + +# 2. Related Work + +Early methods [13, 17, 23] for the 3D SOT task mainly focus on RGB-D information and tend to adopt 2D Siamese networks used in 2D object tracking with additional depth maps. However, the changes in illumination and appearance may degrade the performance of these RGB-D methods. As a pioneering work in this area, SC3D [6] crops the target from the previous frame with the previous bounding box, and then computes the cosine similarity between the target template and a series of 3D target proposals sampled from the current frame using a Siamese backbone. The pipeline relies on heuristic sampling, which is very time-consuming. + +To address these issues, P2B [20] develops an end-to-end framework, which first employs a shared backbone to embed local geometry into point features, and then propagates target cues from the target template to the search area in the current frame. Finally, it adopts VoteNet [18] to generate 3D proposals and selects the proposal with the highest score as the target. P2B [20] reaches a balance between performance and efficiency, and many works follow the same paradigm. MLVSNet [29] aggregates information at multiple levels for more effective target localization. BAT [31] introduces a box-aware feature fusion module to enhance the correlation learning between the target template and the search area. V2B [8] proposes a voxel-to-BEV (Bird's Eye View) target localization network, which projects the point features into a dense BEV feature map to tackle the sparsity of point clouds. Inspired by the success of transformers [25], LTTR [3] adopts a transformer-based architecture to fuse features from two branches and propagate target cues. PTT [22] integrates a transformer module into the P2B architecture to refine point features. PTTR [33] introduces Point Relation Transformer for feature fusion and a light-weight Prediction Refinement Module for coarse-to-fine localization. ST-Net [9] develops an iterative coarse-to-fine correlation network for robust correlation learning. + +Although achieving promising results, the aforementioned methods crop the target from the previous frame using the given bounding box. This overlook of contextual information across two frames makes these methods sensitive + +to appearance variations caused by commonly occurred occlusions and thus the results tend to drift towards intra-class distractors, as mentioned in M2-Track [32]. To this end, M2-Track introduces a motion-centric paradigm to handle the 3D SOT problem, which directly takes the point clouds from two consecutive frames as input without cropping. It first localizes the target in the two frames by target segmentation, and then adopts PointNet [19] to predict the relative target motion from the cropped target area that lacks contextual information. M2-Track could not fully utilize local geometric and contextual information for prediction, which may hinder precise bounding box regression. + +# 3. Method + +# 3.1. Problem Definition + +Given the initial state of the target, single object tracking aims to localize the target in a dynamic scene frame by frame. The initial state in the first frame is given as the 3D bounding box of the target, which can be parameterized by its center coordinates $(x,y,z)$ , orientation angle $\theta$ (around the up-axis, which is sufficient for most tracked objects staying on the ground) and sizes along each axis $(w,l,h)$ . Since the tracking target has little change in size across frames even for non-rigid objects, we assume constant target size and only regress the translation offset $(\Delta x,\Delta y,\Delta z)$ and the rotation angle $(\Delta \theta)$ between two consecutive frames to simplify the tracking task. By applying the translation and rotation to the 3D bounding box $\mathcal{B}_{t - 1}\in \mathbb{R}^7$ in the previous frame, we can compute the 3D bounding box $\mathcal{B}_t\in \mathbb{R}^7$ to localize the target in the current frame. + +Suppose the point clouds in two consecutive frames are denoted as $\mathcal{P}_{t - 1}\in \mathbb{R}^{\dot{N}_{t - 1}\times 3}$ and $\mathcal{P}_t\in \mathbb{R}^{\dot{N}_t\times 3}$ , respectively, where $\dot{N}_{t - 1}$ and $\dot{N}_t$ are the numbers of points in the point clouds. We follow M2-Track [32] and encode the 3D bounding box $\mathcal{B}_{t - 1}$ as a targetness mask $\dot{\mathcal{M}}_{t - 1} = (m_{t - 1}^{1},m_{t - 1}^{2},\dots ,m_{t - 1}^{\dot{N}_{t - 1}})\in \mathbb{R}^{\dot{N}_{t - 1}}$ to indicate the target position, where the mask $m_{t - 1}^{i}$ for the $i$ -th point $p_{t - 1}^{i}$ is defined as + +$$ +m _ {t - 1} ^ {i} = \left\{ \begin{array}{l l} 0 & p _ {t - 1} ^ {i} \text {n o t i n} \mathcal {B} _ {t - 1} \\ 1 & p _ {t - 1} ^ {i} \text {i n} \mathcal {B} _ {t - 1} \end{array} \right. \tag {1} +$$ + +Thus, the 3D SOT task can be formalized as learning the following mapping + +$$ +\mathcal {F} \left(\mathcal {P} _ {t - 1}, \dot {\mathcal {M}} _ {t - 1}, \mathcal {P} _ {t}\right) \mapsto (\Delta x, \Delta y, \Delta z, \Delta \theta) \tag {2} +$$ + +# 3.2. Overview of CXTrack + +Following Eq. (2), we propose a network named CX-Track to improve tracking accuracy by fully exploiting contextual information across frames, and the overall design is illustrated in Fig. 2. We first apply a hierarchical feature learning architecture as the shared backbone to embed local geometric features of the point clouds into point + +features. We use $N_{t - 1}$ and $N_{t}$ to denote the numbers of point features extracted by the backbone. For convenience of calculation, we create a targetness mask $\dot{\mathcal{M}}_t$ and fill it with 0.5 as it is unknown. We then concatenate the point features and targetness masks of the two frames to get $\mathcal{X} = \mathcal{X}_{t - 1}\oplus \mathcal{X}_t\in \mathbb{R}^{N\times C}$ and $\mathcal{M} = \mathcal{M}_{t - 1}\oplus \mathcal{M}_t\in \mathbb{R}^{N\times 1}$ where $N = N_{t - 1} + N_{t}$ , $\mathcal{M}_{t - 1}$ and $\mathcal{M}_t$ are masks corresponding to point features, and extracted from $\dot{\mathcal{M}}_{t - 1}$ and $\dot{\mathcal{M}}_t$ , and $C$ is the number of channels for point features. We employ the target-centric transformer (Sec. 3.3) to integrate the targetness mask information into point features while exploring the contextual information across frames. Finally, we propose a novel localization network, named X-RPN (Sec. 3.4), to obtain the target proposals. The proposal with the highest targetness score is verified as the result. + +# 3.3. Target-Centric Transformer + +Target-Centric Transformer aims to enhance the point features using the contextual information around the target while propagating the target cues from the previous frame to the current frame. It is composed of $N_{L} = 4$ identical layers stacked in series. Given the point features $\mathcal{X}^{(k - 1)}\in \mathbb{R}^{N\times C}$ and the targetness mask $\mathcal{M}^{(k - 1)}$ from the $(k - 1)$ -th layer as input $(\mathcal{M}^{(0)} = \mathcal{M}$ and $\mathcal{X}^{(0)} = \mathcal{X})$ , the $k$ -th layer first models the interactions between any two points while integrating the targetness mask into point features using a modified self-attention operation, and then adopts Multi-Layer-Perceptrons (MLPs) to compute the new point features $\mathcal{X}^{(k)}$ as well as the refined targetness mask $\mathcal{M}^{(k)}$ . Thus, the predicted targetness mask will be consistently refined layer by layer. Moreover, we found it beneficial to add an auxiliary loss by predicting a potential target center for each point via Hough voting, so each layer also applies a shared MLP to generate the potential target center $C^{(k)}\in \mathbb{R}^{N\times 3}$ . + +Formally, we first employ layer normalization [1] $\mathrm{LN}(\cdot)$ before the self-attention mechanism [25] following the design of 3DETR [15], which can be written as + +$$ +\bar {X} = \mathrm {L N} \left(\mathcal {X} ^ {(k - 1)}\right) \tag {3} +$$ + +Then, we add positional encodings (PE) of the coordinates to the normalized point features before feeding them into the self-attention operation + +$$ +X _ {Q} = X _ {K} = \bar {X} + \mathrm {P E} \tag {4} +$$ + +$$ +X _ {V} = \bar {X} \tag {5} +$$ + +It is worth noting that we only adopt PE for the query and key branches, therefore each refined point feature is constrained to focus more on local geometry instead of its associated absolute position. Subsequently, the transformer layer employs a global self-attention operation to model the + +![](images/7ea30e35c661f2e919c71b0854d6d3650642d285775be33f81fd7d3ada7aa47f.jpg) +Figure 2. The overall architecture of CXTrack. Given two consecutive point clouds and the 3D bounding box in the previous frame, CXTrack first embeds the local geometry into point features using the backbone. Then, CXTrack employs the target-centric transformer to explore contextual information across two frames and propagate the target cues to the current frame. Finally, the enhanced features are fed into a novel localization network named X-RPN to obtain high-quality proposals for verification. + +![](images/80369015367aba9cb344d87501ad2babdbf0830d30fff22e886393a47cbc9065.jpg) +Figure 3. Comparison of various transformer layers to fuse the targetness mask and point features. We introduce three types of target-centric transformer layers, namely Vanilla, Semi-Dropout and Gated layer to integrate the targetness mask information into the point features while modeling intra-frame and inter-frame feature relationships. + +relationships between point features, formulated as + +$$ +\mathrm {M H A} \left(X _ {Q}, X _ {K}, X _ {V}\right) = \operatorname {C o n c a t} \left(\text {h e a d} _ {1}, \dots \text {h e a d} _ {h}\right) W ^ {O} \tag {6} +$$ + +$$ +\text {w h e r e} \operatorname {h e a d} _ {i} = \operatorname {A t t n} \left(X _ {Q} W _ {i} ^ {Q}, X _ {K} W _ {i} ^ {K}, \bar {X} W _ {i} ^ {V}\right), \tag {7} +$$ + +$$ +\operatorname {A t t n} (Q, K, V) = \operatorname {s o f t m a x} \left(\frac {Q K ^ {T}}{\sqrt {d _ {k}}}\right) V \tag {8} +$$ + +Here, MHA indicates a multi-head attention, where the attention is applied in $h$ subspaces before concatenation. The projections are implemented by parameter matrices $W_{i}^{Q}\in \mathbb{R}^{C\times d_{k}}$ , $W_{i}^{K}\in \mathbb{R}^{C\times d_{k}}$ , $W_{i}^{V}\in \mathbb{R}^{C\times d_{v}}$ and $W_{i}^{O}\in \mathbb{R}^{hd_{v}\times C}$ , where $i$ indicates the $i$ -th subspace. The self-attention sublayer can be written as + +$$ +\widehat {X} = \mathcal {X} ^ {(k - 1)} + \operatorname {D r o p o u t} \left(\mathrm {M H A} \left(X _ {Q}, X _ {K}, X _ {V}\right)\right) \tag {9} +$$ + +In addition to the self-attention sublayer, each transformer layer also contains a fully connected feed-forward network to refine the point features. The final output of the $k$ -th transformer layer is given by + +$$ +\mathcal {X} ^ {(k)} = \widehat {X} + \operatorname {D r o p o u t} (\operatorname {F F N} (\operatorname {L N} (\widehat {X}))), \tag {10} +$$ + +$$ +\text {w h e r e} \operatorname {F F N} (x) = \max \left(0, x W _ {1} + b _ {1}\right) W _ {2} + b _ {2}. \tag {11} +$$ + +To integrate the targetness mask information into point features, we need to modify the classic transformer layer. We introduce three types of modified transformer layers in Fig. 3, namely Vanilla, Semi-Dropout and Gated layer. + +Vanilla. We project the input $\mathcal{M}^{(k - 1)}$ to mask embedding $\mathrm{ME}\in \mathbb{R}^{N\times C}$ using a linear transformation. Following the design of positional encoding, we simply add ME to the input token embedding $X_{V}$ , which re-formulates Eq. (5) as + +$$ +X _ {V} = \bar {X} + \mathrm {M E} \tag {12} +$$ + +Semi-Dropout. Notably, the targetness mask information can only flow across layers along the attention path. For small objects which only have a few points to track, applying dropout to the mask embedding may discard the targetness information and lead to performance degradation. To this end, we separate the self-attention mechanism into a feature branch and a mask branch with shared attention weights, while only applying dropout to the refined point features. As shown in Fig. 3b, the self-attention sublayer in Eq. (9) is re-formulated as + +$$ +\begin{array}{l} \widehat {X} = \mathcal {X} ^ {(k - 1)} + \operatorname {D r o p o u t} (\mathrm {M H A} (X _ {Q}, X _ {K}, \overline {{X}})) \\ + \mathrm {M H A} \left(X _ {Q}, X _ {K}, \mathrm {M E}\right) \tag {13} \\ \end{array} +$$ + +![](images/4a939d898bf5efa9aa6bdb33f8bbe36dc10ce6615a2e25b37222d34b0a53616f.jpg) +Figure 4. The overall architecture of X-RPN. X-RPN adopts a local transformer to model point feature interaction within the target and aggregate local clues. It also incorporates a center embedding mechanism which embeds the relative target motion between two frames to distinguish the target from distractors. + +Gated. Inspired by the design of TrDimp [26], we introduce a gated mechanism into the self-attention sublayer to integrate the mask information. It has two parallel branches, namely mask transformation and feature transformation. For mask transformation, we first obtain the feature mask $\overline{M} \in \mathbb{R}^{N \times C}$ by repeating the input point-wise mask $\mathcal{M}^{(k - 1)} \in \mathbb{R}^{N \times 1}$ for $C$ times. Then we can propagate the targetness cues to the current frame via adopting self-attention on the mask feature. The transformed mask serves as the gate matrix for the point features + +$$ +\widehat {X} _ {m} = \mathrm {L N} (\mathrm {M H A} (X _ {Q}, X _ {K}, \bar {M}) \otimes \bar {X}). \tag {14} +$$ + +For feature transformation, we first mask the point features to suppress feature activation in background areas, and then employ self-attention with a residual connection to model the relationships between features + +$$ +\widehat {X} _ {f} = \mathrm {L N} (\mathrm {M H A} (X _ {Q}, X _ {K}, \bar {M} \otimes \bar {X}) + \bar {X}). \tag {15} +$$ + +As illustrated in Fig. 3c, we sum and normalize the output features $\widehat{X}_m$ and $\widehat{X}_f$ from the two branches. Eq. (9) can be re-formulated as + +$$ +\widehat {X} = \operatorname {L N} \left(\widehat {X} _ {f} + \widehat {X} _ {m}\right). \tag {16} +$$ + +Among the above three layers, we observe significant performance gain from using Semi-Dropout target-centric transformer layers (Sec. 4.3). Thus CXTrack employs Semi-Dropout layers to integrate targetness information while exploring contextual information across frames. + +# 3.4. X-RPN + +Previous works [20] indicate that individual point features can only capture limited local information, which may not be sufficient for precise bounding box regression. Thus + +we develop a simple yet effective localization network, named X-RPN, which extends RPN [18] using local transformer and center embedding, as shown in Fig. 4. Different from previous works [8,20], X-RPN aggregates local clues from point features without downsampling or voxelization, thus avoiding information loss and reaching a good balance between handling large and small objects. Our intuition is that each point should only interact with points belonging to the same object to suppress irrelevant information. Given the point features $\mathcal{X}^{(N_L)}$ , targetness mask $\mathcal{M}^{(N_L)}$ and target center $\mathcal{C}^{(N_L)}$ output by the target centric transformer, we first split them along the spatial dimension and only feed those belonging to the current frame into X-RPN, which is denoted as $\widetilde{\mathcal{X}}\in \mathbb{R}^{N_t\times C}$ , $\widetilde{\mathcal{C}}\in \mathbb{R}^{N_t\times 3}$ and $\widetilde{\mathcal{M}}\in \mathbb{R}^{N_t\times 1}$ , respectively. X-RPN first computes the neighborhood $\mathcal{N}(p_i)$ for each point $p_i$ using its potential target center $c_{i}$ + +$$ +\mathcal {N} \left(p _ {i}\right) = \left\{p _ {j} \mid | | c _ {i} - c _ {j} | | _ {2} < r \right\} \tag {17} +$$ + +Here $r$ is a hyperparameter indicating the size of the neighborhood. Then X-RPN adopts the transformer architecture mentioned in Sec. 3.3 to aggregate local information, where each point only interacts with its neighborhood points to suppress noise. We remove the feed-forward network in the transformer layer because we observe that one layer is sufficient to generate high quality target proposals. + +To deal with intra-class distractors which are widespread in the scenes [32] especially for pedestrian tracking, we propose to combine the potential center information with the targetness mask. Our intuition lies in two folds. First, the tracking target keeps similar local geometry across two frames. Second, if the duration between two consecutive frames is sufficiently short, the displacement of the target is small. Therefore, we construct a Gaussian proposal-wise mask $\mathcal{M}_c$ to indicate the magnitude of the displacement of each proposal. Formally, for each point $p_i$ with the predicted target center $c_i$ , the mask value $m_i^c \in \mathcal{M}_c$ is + +$$ +m _ {i} ^ {c} = \exp \left(- \frac {\left| \left| c _ {i} - \bar {c} \right| \right| _ {2} ^ {2}}{2 \sigma^ {2}}\right) \tag {18} +$$ + +where $\overline{c} \in \mathbb{R}^3$ is the target center in the previous frame and $\sigma$ is a learnable or fixed scaling factor. We embed the target center mask $\mathcal{M}_c$ into the center embedding matrix $\mathrm{CE} \in \mathbb{R}^{N \times C}$ using a linear transformation, and equally combine the mask embedding and the center embedding. + +# 3.5. Loss Functions + +For the prediction $\mathcal{M}^{(k)}$ given by the $k$ -th transformer layer, we adopt a standard cross entropy loss $\mathcal{L}_{\mathrm{cm}}^{(i)}$ . As for the potential target centers, we observe that it is difficult to regress precise centers for non-rigid objects such as pedestrians. Hence the predicted centers $\mathcal{C}^{(k)}$ are supervised by $L_{2}$ loss for non-rigid objects, and by Huber loss [21] for + +rigid objects. For the target center regression loss $\mathcal{L}_{\mathrm{cc}}^{(i)}$ , only points in the ground truth bounding box are supervised. + +Following previous works [20], proposals with predicted centers near the target center $(< 0.3\mathrm{m})$ are considered as positives and those far away $(>0.6\mathrm{m})$ are considered as negatives. Others are left unsupervised. The predicted targetness mask is supervised via standard cross-entropy loss $\mathcal{L}_{\mathrm{rm}}$ and only the bounding box parameters of positive predictions are supervised by Huber (Smooth- $L_{1}$ ) loss $\mathcal{L}_{\mathrm{box}}$ . + +The overall loss is the weighted combination of the above loss terms + +$$ +\mathcal {L} = \gamma_ {1} \sum_ {i = 1} ^ {N _ {L}} \mathcal {L} _ {\mathrm {c m}} ^ {(i)} + \gamma_ {2} \sum_ {i = 1} ^ {N _ {L}} \mathcal {L} _ {\mathrm {c c}} ^ {(i)} + \gamma_ {3} \mathcal {L} _ {\mathrm {r m}} + \mathcal {L} _ {\mathrm {b o x}} \tag {19} +$$ + +where $\gamma_{1},\gamma_{2}$ and $\gamma_{3}$ are hyper-parameters. We empirically set $\gamma_{1} = 0.2$ $\gamma_{2} = 1.0$ $\gamma_{3} = 1.5$ for rigid objects and $\gamma_{1} = 0.2$ $\gamma_{2} = 10.0$ $\gamma_{3} = 1.0$ for non-rigid objects. + +# 4. Experiments + +# 4.1. Settings + +Datasets. We compare CXTrack with previous state of the arts on three large-scale datasets: KITTI [5], nuScenes [2] and Waymo Open Dataset (WOD) [24]. KITTI contains 21 training video sequences and 29 test sequences. We follow previous work [6] and split the training sequences into three parts, 0-16 for training, 17-18 for validation and 19-20 for testing. For nuScenes, we use its validation split to evaluate our model, which contains 150 scenes. For WOD, we follow LiDAR-SOT [16] to evaluate our method, dividing it into three splits according to the sparsity of point clouds. + +Implementation Details. We adopt DGCNN [28] as the backbone network to extract local geometric information. In the X-RPN, we initialize the scaling parameter $\sigma^2 = 10$ . Notably, we empirically fix $\sigma$ as a hyper-parameter for pedestrians and cyclists, and set it as a learnable parameter for cars and vans, since they may have larger motions. More details are provided in the supplementary material. + +Evaluation Metrics. We use Success and Precision defined in one pass evaluation [10] as evaluation metrics. Success denotes the Area Under Curve (AUC) for the plot showing the ratio of frames where the Intersection Over Union (IOU) between the predicted and ground-truth bounding boxes is larger than a threshold, ranging from 0 to 1. Precision is defined as the AUC of the plot showing the ratio of frames where the distance between their centers is within a threshold, from 0 to 2 meters. + +# 4.2. Comparison with State of the Arts + +We make comprehensive comparisons on the KITTI dataset with previous state-of-the-art methods, including SC3D [6], P2B [20], 3DSiamRPN [4], LTTR [3], MLVS-Net [29], BAT [31], PTT [22], V2B [8], PTTR [33], + +Table 1. Comparisons with the state-of-the-art methods on KITTI dataset. "Mean" is the average result weighted by frame numbers. "Blue" and "Bold" denote previous and current best performance, respectively. Success/Precision are used for evaluation. + +
MethodCar (6424)Pedestrian (6088)Van (1248)Cyclist (308)Mean (14068)
SC3D41.3/57.918.2/37.840.4/47.041.5/70.431.2/48.5
P2B56.2/72.828.7/49.640.8/48.432.1/44.742.4/60.0
3DSiamRPN58.2/76.235.2/56.245.7/52.936.2/49.046.7/64.9
LTPR65.0/77.133.2/56.835.8/45.666.2/89.948.7/65.8
MLVSNet56.0/74.034.1/61.152.0/61.434.3/44.545.7/66.7
BAT60.5/77.742.1/70.152.4/67.033.7/45.451.2/72.8
PTT67.8/81.844.9/72.043.6/52.537.2/47.355.1/74.2
V2B70.5/81.348.3/73.550.1/58.040.8/49.758.4/75.2
PTTR65.2/77.450.9/81.652.5/61.865.1/90.557.9/78.1
STNet72.1/84.049.9/77.258.0/70.673.5/93.761.3/80.1
M2-Track65.5/80.861.5/88.253.8/70.773.2/93.562.9/83.4
CXTrack69.1/81.667.0/91.560.0/71.874.2/94.367.5/85.3
Improvement↓3.0/↓2.4↑5.5/↑3.3↑2.0/↑1.1↑0.7/↑0.6↑4.6/↑1.9
+ +Table 2. Robustness under scenes that contain intra-class distractors on KITTI Pedestrian category. + +
MethodAll(6088)Distractor-Only(3917)Improvement
PTTR50.9/81.644.3/70.0↓6.6/↓11.6
STNet49.9/77.235.1/58.5↓14.8/↓18.7
M2-Track61.5/88.258.0/88.4↓3.5/↑0.2
CXTrack67.0/91.566.1/91.3↓0.9/↓0.3
+ +STNet [9] and M2-Track [32]. As illustrated in Tab. 1, CXTrack surpasses previous state-of-the-art methods, with a significant improvement of average Success and Precision. Notably, our method achieves the best performance under all categories, except for the Car, where voxel-based STNet [9] surpasses us by a minor margin (72.1/84.0 v.s. 69.1/81.6). Most vehicles have simple shapes and limited rotation angles, which fit well in voxels. We argue that voxelization provides a strong shape prior, thereby leading to performance gain for large objects with simple shapes. The lack of distractors for cars also makes our improvement over previous methods insignificant. However, our method has a significant improvement (67.0/91.5 v.s. 49.9/77.2) on the Pedestrian category. We claim that this stems from our special design to handle distractors and our better preservation for contextual information. Besides, compared with M2-Track [32], CXTrack obtains consistent performance gains on all categories especially on the Success metric, which demonstrates the importance of local geometry and contextual information. For further analysis on the impact of intraclass distractors, we manually pick out scenes that contain Pedestrian distractors from the KITTI test split and then evaluate different methods on these scenes. As shown in Tab. 2, both M2-Track and CXTrack are robust to distractors, while CXTrack can make more accurate predictions. + +To verify the generalization ability of our method, we follow previous methods [8, 9] and test the KITTI pretrained model on nuScenes and WOD. The comparison results on WOD are shown in Tab. 3. It can be seen that our + +Table 3. Comparison with state of the arts on Waymo Open Dataset. + +
MethodVehicle(185731)Pedestrian(241752)Mean(427483)
EasyMediumHardMeanEasyMediumHardMean
P2B57.1/65.452.0/60.747.9/58.552.6/61.718.1/30.817.8/30.017.7/29.317.9/30.133.0/43.8
BAT61.0/68.353.3/60.948.9/57.854.7/62.719.3/32.617.8/29.817.2/28.318.2/30.334.1/44.4
V2B64.5/71.555.1/63.252.0/62.057.6/65.927.9/43.922.5/36.220.1/33.123.7/37.938.4/50.1
STNet65.9/72.757.5/66.054.6/64.759.7/68.029.2/45.324.7/38.222.2/35.825.5/39.940.4/52.1
CXTrack63.9/71.154.2/62.752.1/63.757.1/66.135.4/55.329.7/47.926.3/44.430.7/49.442.2/56.7
Improvement↓2.0/↓1.6↓3.3/↓3.3↓3.5/↓1.0↓2.6/↓1.9↑6.2/↑10.0↑5.0/↑9.7↑4.1/↑8.6↑5.2/↑9.5↑1.8/↑4.6
+ +Table 4. Comparison with state of the arts on nuScenes. + +
MethodCar (15578)Pedestrian (8019)Van (3710)Cyclist (501)Mean (27808)
SC3D25.0/27.114.2/16.225.7/21.917.0/18.221.8/23.1
P2B27.0/29.215.9/22.021.5/16.220.0/26.422.9/25.3
BAT22.5/24.117.3/24.519.3/15.817.0/18.820.5/23.0
V2B31.3/35.117.3/23.421.7/16.722.2/19.125.8/29.0
STNet32.2/36.119.1/27.222.3/16.821.2/29.226.9/30.8
CXTrack29.6/33.420.4/32.927.6/20.818.5/26.826.5/31.5
Improvement↓2.6/↓2.7↑1.3/↑5.7↑1.9/↓1.1↓3.7/↓2.4↓0.4/↑0.7
+ +Table 5. Efficiency analysis of different components. + +
ComponentFLOPs#ParamsInfer Speed
backbone3.18G1.3M8.5ms
transformer1.28G14.7M10.9ms
X-RPN0.17G2.3M3.0ms
pre/postprocess--6.8ms
CXTrack4.63G18.3M29.2ms(34FPS)
+ +method outperforms others in terms of average Success and Precision with a clear margin. Notably, KITTI and WOD data are captured by 64-beam LiDARs, while nuScenes data are captured by 32-beam LiDARs. Thus it is more challenging to generalize the pretrained model on the nuScenes dataset. As shown in Tab. 4, our method achieves comparable performance on the nuScenes dataset. In short, CXTrack not only achieves a good balance between small objects and large objects, but also generalizes well to unseen scenes. + +We also visualize the tracking results for qualitative comparisons. As shown in Fig. 5, CXTrack achieves good accuracy in scenes with both sparse and dense point clouds on both categories. In the sparse cases (left), previous methods drift towards intra-class distractors due to large appearance variations caused by occlusions, while only our method keeps track of the target, thanks to the sufficient use of contextual information. In the dense cases (right), our method can track the target more accurately than M2-Track by leveraging local geometric information. + +We report the efficiency of different components in Tab. 5. It can be observed that the target-centric transformer is the bottleneck of CXTrack during inference. We can replace the vanilla self-attention in CXTrack with linear attention such as linformer [27] for further speedup. + +# 4.3. Ablation Studies + +To validate the effectiveness of several design choices in CXTrack, we conduct ablation studies on the KITTI dataset. + +Table 6. Ablation studies of different components of the target-centric transformer. "Cx" refers to contextual information, "M" refers to the cascaded targetness mask prediction and "C" refers to the auxiliary target center regression branches. + +
CxMCCarPedestrianVanCyclistMean
✓62.5/74.260.6/87.058.3/71.472.0/93.361.5/79.9
✓✓67.4/80.263.9/89.057.8/70.872.7/93.865.1/83.5
✓✓59.7/73.6†51.8/81.659.9/71.571.7/93.256.6/77.3
✓✓✓69.1/81.667.0/91.560.0/71.874.2/94.367.5/85.3
+ +†: unstable training process + +Table 7. Ablation studies of different transformer layers on KITTI. "V" refers to the vanilla transformer layer and "G" refers to the gated transformer layer. "S" represents the semi-dropout transformer layer which is adopted in our proposed CXTrack. + +
CarPedestrianVanCyclistMean
V68.8/80.462.9/87.857.2/69.672.7/94.265.3/82.9
G64.8/76.964.7/91.156.2/70.570.6/93.464.1/82.8
S69.1/81.667.0/91.560.0/71.874.2/94.367.5/85.3
+ +Components of Target-Centric Transformer. Tab. 6 presents ablation studies of different components of transformer to gain a better understanding of its designs. We crop the input point cloud $\mathcal{P}_{t-1}$ using $\mathcal{B}_{t-1}$ to ablate contextual information in the previous frame. We can observe significant performance drop when not using contextual information, especially on Car and Pedestrian. For Car, it suffers from heavy occlusions (Fig. 5), while pedestrian distractors are widespread in the scene. We also find that removing context leads to unstable training on Car. We presume that the lack of supervised signals to tell the model what not to attend may confuse the model and introduce noise in training. For the cascaded targetness mask prediction and auxiliary target center regression, removing either of them leads to an obvious decline on terms of average metrics. We argue that the auxiliary regression loss can increase the feature similarities of points belonging to the same object. + +Target-Centric Transformer Layer. Tab. 7 shows the impact of different target-centric transformer layers. Semi-Dropout achieves better performance than Vanilla, especially on Pedestrian. Small objects often consist of fewer points, hence applying dropout directly on the targetness information in training may confuse the network and lead to sub-optimal results. Gated relies entirely on the predicted targetness mask to modulate the amount of exposure for input features, which may suffer from information loss when the targetness mask is not accurate enough. + +![](images/8a18fb599e340f05b4de5b2f3d9521f4b98e9a2ff6806d2c7b64df51f9321571.jpg) +Figure 5. Visualization results. Left: Sparse cases in KITTI. Right: Dense cases in KITTI. + +![](images/134c9d88ef8e452ba93f35a66f819d6f2db382d8f0404d66f1235534f795f4bc.jpg) + +Table 8. Ablation studies of various localization heads on KITTI. "X-RPN\C" indicates our proposed localization head X-RPN without center embedding. + +
HeadCarPedestrianVanCyclistMean
PRM [33]66.5/77.462.2/86.852.9/64.972.5/93.863.6/80.7
RPN [18]64.1/76.959.8/88.355.0/65.668.2/92.461.5/81.2
V2B [8]70.5/82.660.1/86.758.0/69.870.5/93.364.9/83.5
X-RPN\C67.8/80.365.5/89.559.9/72.172.6/94.166.2/83.9
X-RPN69.1/81.667.0/91.560.0/71.874.2/94.367.5/85.3
+ +![](images/ad227c0d17dd9257df9a43ab0d9999566ccb1b92022a95c44183282c8c4dcd90.jpg) +Figure 6. Representative examples of attention maps in the transformer. Target-centric transformer attends to objects that have similar geometry. + +![](images/1b440cac86f7e678c254783772b24d356a32f8bc26dc4f5f790924c4de3ffd14.jpg) + +![](images/3da16ceb93d9623f917fbcd52dab6e3372d8d696310ef5ca5971bdf92a49e19d.jpg) +Figure 7. Visualization of ablation study. Center embedding can benefit object tracking in challenging scenes with distractors. + +X-RPN. We replace X-RPN with other alternatives [8, 20, 33] and report the comparison results in Tab. 8. Although the V2B head achieves better performance than X-RPN on the Car category, it fails to track small objects such as pedestrians effectively due to intra-class distractors and inevitable information loss brought in by voxelization. It is also worth noting that we observe a performance drop without center embedding, especially on the Pedestrian category, for which distractors are more commonly seen. To explore the effectiveness of the center embedding, we visualize the attention map of the last transformer layer in Fig. 6. We observe that the transformer alone can attend to regions with similar ge + +ometry to the target, but fails to distinguish the target from distractors. As shown in Fig. 7, with the help of the center embedding, the network precisely keeps track of the target. In short, X-RPN achieves a good balance between large and small objects, and effectively alleviates the distractor problem. + +# 4.4. Failure Cases + +Although CXTrack is robust to intra-class distractors, it fails to predict accurate orientation of the target when the point clouds are too sparse to capture informative local geometry or when large appearance variations occur, as shown in Fig. 7. Besides, the center embedding directly encodes the displacement of target center into features, so our model may suffer from performance degradation if trained with $2\mathrm{Hz}$ data and tested with $10\mathrm{Hz}$ data because the scale of the displacement differs significantly. + +# 5. Conclusion + +We revisit existing paradigms for the 3D SOT task and propose a new paradigm to fully exploit contextual information across frames, which is largely overlooked by previous methods. Following this paradigm, we design a novel transformer-based network named CXTrack, which employs a target-centric transformer to explore contextual information and model intra-frame and inter-frame feature relationships. We also introduce a localization head named X-RPN to obtain high-quality proposals for objects of all sizes, as well as a center embedding module to distinguish the target from distractors. Extensive experiments show that CXTrack significantly outperforms previous state-of-the-arts, and is robust to distractors. We hope our work can promote further exploitations in this task by showing the necessity to explore contextual information for more robust predictions. + +Acknowledgment The authors thank Jiahui Huang for his discussions. This work was supported by the Natural Science Foundation of China (Project Number 61832016), and Tsinghua-Tencent Joint Laboratory for Internet Innovation Technology. + +# References + +[1] Jimmy Lei Ba, Jamie Ryan Kiros, and Geoffrey E Hinton. Layer normalization. arXiv preprint arXiv:1607.06450, 2016.3 +[2] Holger Caesar, Varun Bankiti, Alex H Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuScenes: A multimodal dataset for autonomous driving. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11621-11631, 2020. 6 +[3] Yubo Cui, Zheng Fang, Jiayao Shan, Zuoxu Gu, and Sifan Zhou. 3D object tracking with transformer. arXiv preprint arXiv:2110.14921, 2021. 2, 6 +[4] Zheng Fang, Sifan Zhou, Yubo Cui, and Sebastian Scherer. 3D-SiamRPN: An end-to-end learning method for real-time 3D single object tracking using raw point cloud. IEEE Sensors Journal, 21(4):4995-5011, 2020. 6 +[5] Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we ready for autonomous driving? the KITTI vision benchmark suite. In IEEE Conference on Computer Vision and Pattern Recognition, pages 3354-3361. IEEE, 2012. 6 +[6] Silvio Giancola, Jesus Zarzar, and Bernard Ghanem. Leveraging shape completion for 3D Siamese tracking. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1359-1368, 2019. 1, 2, 6 +[7] Chenhang He, Ruihuang Li, Shuai Li, and Lei Zhang. Voxel set transformer: A set-to-set approach to 3D object detection from point clouds. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8417-8427, 2022. 1 +[8] Le Hui, Lingpeng Wang, Mingmei Cheng, Jin Xie, and Jian Yang. 3D Siamese voxel-to-BEV tracker for sparse point clouds. Advances in Neural Information Processing Systems, 34:28714-28727, 2021. 2, 5, 6, 8 +[9] Le Hui, Lingpeng Wang, Linghua Tang, Kaihao Lan, Jin Xie, and Jian Yang. 3D Siamese transformer network for single object tracking on point clouds. In Proceedings of the European Conference on Computer Vision, Part II, pages 293-310. Springer, 2022. 2, 6 +[10] Matej Kristan, Jiri Matas, Aleš Leonardis, Tomáš Vojíř, Roman Pflugfelder, Gustavo Fernández, Georg Nebehay, Fatih Porikli, and Luka Čehovin. A novel performance evaluation methodology for single-target trackers. IEEE Transactions on Pattern Analysis and Machine Intelligence, 38(11):2137-2155, 2016. 6 +[11] H Kuang Chiu, A Prioletti, J Li, and J Bohg. Probabilistic 3D multi-object tracking for autonomous driving. ArXiv, vol. abs/2001.05673, 2020. 1 +[12] Ze Liu, Zheng Zhang, Yue Cao, Han Hu, and Xin Tong. Group-free 3D object detection via transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2949-2958, 2021. 1 +[13] Matthias Luber, Luciano Spinello, and Kai O Arras. People tracking in RGB-D data with on-line boosted target models. In 2011 IEEE/RSJ International Conference on Intelligent Robots and Systems, pages 3844-3849. IEEE, 2011. 2 + +[14] Jiageng Mao, Yujing Xue, Minzhe Niu, Haoyue Bai, Jiashi Feng, Xiaodan Liang, Hang Xu, and Chunjing Xu. Voxel transformer for 3D object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3164-3173, 2021. 1 +[15] Ishan Misra, Rohit Girdhar, and Armand Joulin. An end-to-end transformer model for 3D object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2906-2917, 2021. 1, 3 +[16] Ziqi Pang, Zhichao Li, and Naiyan Wang. Model-free vehicle tracking and state estimation in point cloud sequences. In IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 8075-8082, 2021. 6 +[17] Alessandro Pieropan, Niklas Bergstrom, Masatoshi Ishikawa, and Hedvig Kjellstrom. Robust 3D tracking of unknown objects. In IEEE International Conference on Robotics and Automation (ICRA), pages 2410-2417. IEEE, 2015. 2 +[18] Charles R Qi, Or Litany, Kaiming He, and Leonidas J Guibas. Deep Hough voting for 3D object detection in point clouds. In proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9277-9286, 2019. 1, 2, 5, 8 +[19] Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. PointNet: Deep learning on point sets for 3D classification and segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 652-660, 2017. 3 +[20] Haozhe Qi, Chen Feng, Zhiguo Cao, Feng Zhao, and Yang Xiao. P2B: Point-to-box network for 3D object tracking in point clouds. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6329-6338, 2020. 1, 2, 5, 6, 8 +[21] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster R-CNN: Towards real-time object detection with region proposal networks. Advances in Neural Information Processing Systems, 28, 2015. 5 +[22] Jiayao Shan, Sifan Zhou, Zheng Fang, and Yubo Cui. PTT: Point-track-transformer module for 3D single object tracking in point clouds. In IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 1310-1316. IEEE, 2021. 2, 6 +[23] Luciano Spinello, Kai Arras, Rudolph Triebel, and Roland Siegwart. A layered approach to people detection in 3D range data. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 24, pages 1625-1630, 2010. 2 +[24] Pei Sun, Henrik Kretzschmar, Xerxes Dotiwalla, Aurelien Chouard, Vijaysai Patnaik, Paul Tsui, James Guo, Yin Zhou, Yuning Chai, Benjamin Caine, et al. Scalability in perception for autonomous driving: Waymo open dataset. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2446-2454, 2020. 6 +[25] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in Neural Information Processing Systems, 30, 2017. 2, 3 +[26] Ning Wang, Wengang Zhou, Jie Wang, and Houqiang Li. Transformer meets tracker: Exploiting temporal context for + +robust visual tracking. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1571-1580, 2021. 5 +[27] Sinong Wang, Belinda Z Li, Madian Khabsa, Han Fang, and Hao Ma. Linformer: Self-attention with linear complexity. arXiv preprint arXiv:2006.04768, 2020. 7 +[28] Yue Wang, Yongbin Sun, Ziwei Liu, Sanjay E Sarma, Michael M Bronstein, and Justin M Solomon. Dynamic graph CNN for learning on point clouds. ACM Transactions on Graphics (TOG), 38(5):1-12, 2019. 6 +[29] Zhoutao Wang, Qian Xie, Yu-Kun Lai, Jing Wu, Kun Long, and Jun Wang. MLVSNet: Multi-level voting Siamese network for 3D visual tracking. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3101-3110, 2021. 1, 2, 6 +[30] Tianwei Yin, Xingyi Zhou, and Philipp Krahenbuhl. Center-based 3D object detection and tracking. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11784-11793, 2021. 1 +[31] Chaoda Zheng, Xu Yan, Jiantao Gao, Weibing Zhao, Wei Zhang, Zhen Li, and Shuguang Cui. Box-aware feature enhancement for single object tracking on point clouds. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13199-13208, 2021. 1, 2, 6 +[32] Chaoda Zheng, Xu Yan, Haiming Zhang, Baoyuan Wang, Shenghui Cheng, Shuguang Cui, and Zhen Li. Beyond 3D Siamese tracking: A motion-centric paradigm for 3D single object tracking in point clouds. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8111-8120, 2022. 1, 2, 3, 5, 6 +[33] Changqing Zhou, Zhipeng Luo, Yueru Luo, Tianrui Liu, Liang Pan, Zhongang Cai, Haiyu Zhao, and Shijian Lu. PTTR: Relational 3D point cloud object tracking with transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8531-8540, 2022. 1, 2, 6, 8 \ No newline at end of file diff --git a/2023/CXTrack_ Improving 3D Point Cloud Tracking With Contextual Information/images.zip b/2023/CXTrack_ Improving 3D Point Cloud Tracking With Contextual Information/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..74ba16873433a878fca6ab9723f972e3e21bca23 --- /dev/null +++ b/2023/CXTrack_ Improving 3D Point Cloud Tracking With Contextual Information/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f22245badf2086ce3565a3de6466f7a24ab09c1048e4fe2c3e65297b1624a0e4 +size 784096 diff --git a/2023/CXTrack_ Improving 3D Point Cloud Tracking With Contextual Information/layout.json b/2023/CXTrack_ Improving 3D Point Cloud Tracking With Contextual Information/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..3b559337774714d8bf09dcc0ab4a2df0dedcb92d --- /dev/null +++ b/2023/CXTrack_ Improving 3D Point Cloud Tracking With Contextual Information/layout.json @@ -0,0 +1,8375 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 61, + 103, + 533, + 120 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 103, + 533, + 120 + ], + "spans": [ + { + "bbox": [ + 61, + 103, + 533, + 120 + ], + "type": "text", + "content": "CXTrack: Improving 3D Point Cloud Tracking with Contextual Information" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 123, + 142, + 473, + 158 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 123, + 142, + 473, + 158 + ], + "spans": [ + { + "bbox": [ + 123, + 142, + 473, + 158 + ], + "type": "text", + "content": "Tian-Xing Xu" + }, + { + "bbox": [ + 123, + 142, + 473, + 158 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 123, + 142, + 473, + 158 + ], + "type": "text", + "content": " Yuan-Chen Guo" + }, + { + "bbox": [ + 123, + 142, + 473, + 158 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 123, + 142, + 473, + 158 + ], + "type": "text", + "content": " Yu-Kun Lai" + }, + { + "bbox": [ + 123, + 142, + 473, + 158 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 123, + 142, + 473, + 158 + ], + "type": "text", + "content": " Song-Hai Zhang" + }, + { + "bbox": [ + 123, + 142, + 473, + 158 + ], + "type": "inline_equation", + "content": "^{1*}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 128, + 157, + 466, + 172 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 128, + 157, + 466, + 172 + ], + "spans": [ + { + "bbox": [ + 128, + 157, + 466, + 172 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 128, + 157, + 466, + 172 + ], + "type": "text", + "content": " Tsinghua University, China " + }, + { + "bbox": [ + 128, + 157, + 466, + 172 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 128, + 157, + 466, + 172 + ], + "type": "text", + "content": " Cardiff University, United Kingdom" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 96, + 173, + 495, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 96, + 173, + 495, + 185 + ], + "spans": [ + { + "bbox": [ + 96, + 173, + 495, + 185 + ], + "type": "inline_equation", + "content": "^{1}\\{xutx21@mails., guocycl9@mails., shz@}tsinghua.edu.cn" + }, + { + "bbox": [ + 96, + 173, + 495, + 185 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 96, + 173, + 495, + 185 + ], + "type": "text", + "content": "LaiY4@cardiff.ac.uk" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "spans": [ + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 45, + 239, + 290, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 239, + 290, + 515 + ], + "spans": [ + { + "bbox": [ + 45, + 239, + 290, + 515 + ], + "type": "text", + "content": "3D single object tracking plays an essential role in many applications, such as autonomous driving. It remains a challenging problem due to the large appearance variation and the sparsity of points caused by occlusion and limited sensor capabilities. Therefore, contextual information across two consecutive frames is crucial for effective object tracking. However, points containing such useful information are often overlooked and cropped out in existing methods, leading to insufficient use of important contextual knowledge. To address this issue, we propose CXTrack, a novel transformer-based network for 3D object tracking, which exploits ConteXtual information to improve the tracking results. Specifically, we design a target-centric transformer network that directly takes point features from two consecutive frames and the previous bounding box as input to explore contextual information and implicitly propagate target cues. To achieve accurate localization for objects of all sizes, we propose a transformer-based localization head with a novel center embedding module to distinguish the target from distractors. Extensive experiments on three large-scale datasets, KITTI, nuScenes and Waymo Open Dataset, show that CXTrack achieves state-of-the-art tracking performance while running at 34 FPS." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 539, + 128, + 552 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 539, + 128, + 552 + ], + "spans": [ + { + "bbox": [ + 47, + 539, + 128, + 552 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 559, + 287, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 559, + 287, + 693 + ], + "spans": [ + { + "bbox": [ + 46, + 559, + 287, + 693 + ], + "type": "text", + "content": "Single Object Tracking (SOT) has been a fundamental task in computer vision for decades, aiming to keep track of a specific target across a video sequence, given only its initial status. In recent years, with the development of 3D data acquisition devices, it has drawn increasing attention for using point clouds to solve various vision tasks such as object detection [7, 12, 14, 15, 18] and object tracking [20, 29, 31-33]. In particular, much progress has been made on point cloud-based object tracking for its huge potential in applications such as autonomous driving [11, 30]. However, it remains challenging due to the large appearance variation" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 309, + 210, + 545, + 292 + ], + "blocks": [ + { + "bbox": [ + 309, + 210, + 545, + 292 + ], + "lines": [ + { + "bbox": [ + 309, + 210, + 545, + 292 + ], + "spans": [ + { + "bbox": [ + 309, + 210, + 545, + 292 + ], + "type": "image", + "image_path": "77c8f1b7e853edc39d17d9d5dfaa6d87a85734be4e4583ceef1649bc2a212abf.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 308, + 294, + 545, + 374 + ], + "blocks": [ + { + "bbox": [ + 308, + 294, + 545, + 374 + ], + "lines": [ + { + "bbox": [ + 308, + 294, + 545, + 374 + ], + "spans": [ + { + "bbox": [ + 308, + 294, + 545, + 374 + ], + "type": "image", + "image_path": "7a1c931e1bcc9e0e957af9c111cdb8ef9d5c78f9ea73ca221e92b24c3e776ae1.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 309, + 376, + 545, + 456 + ], + "blocks": [ + { + "bbox": [ + 309, + 376, + 545, + 456 + ], + "lines": [ + { + "bbox": [ + 309, + 376, + 545, + 456 + ], + "spans": [ + { + "bbox": [ + 309, + 376, + 545, + 456 + ], + "type": "image", + "image_path": "fc593d0d67219ed4cf6f62564c4def8a65a2a62ff521c9e74d0b28369e62bb15.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 310, + 458, + 545, + 537 + ], + "blocks": [ + { + "bbox": [ + 310, + 458, + 545, + 537 + ], + "lines": [ + { + "bbox": [ + 310, + 458, + 545, + 537 + ], + "spans": [ + { + "bbox": [ + 310, + 458, + 545, + 537 + ], + "type": "image", + "image_path": "d05a41c5882a19cec4a751cff59ff44cb4bc1ec2b25308a70ec07439396b86be.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 545, + 547, + 601 + ], + "lines": [ + { + "bbox": [ + 304, + 545, + 547, + 601 + ], + "spans": [ + { + "bbox": [ + 304, + 545, + 547, + 601 + ], + "type": "text", + "content": "Figure 1. Comparison of various 3D SOT paradigms. Previous methods crop the target from the frames to specify the region of interest, which largely overlook contextual information around the target. On the contrary, our proposed CXTrack fully exploits contextual information to improve the tracking results." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 604, + 545, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 604, + 545, + 628 + ], + "spans": [ + { + "bbox": [ + 304, + 604, + 545, + 628 + ], + "type": "text", + "content": "of the target and the sparsity of 3D point clouds caused by occlusion and limited sensor resolution." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 629, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 629, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 629, + 547, + 713 + ], + "type": "text", + "content": "Existing 3D point cloud-based SOT methods can be categorized into three main paradigms, namely SC3D, P2B and motion-centric, as shown in Fig. 1. As a pioneering work, SC3D [6] crops the target from the previous frame, and compares the target template with a potentially large number of candidate patches generated from the current frame, which consumes much time. To address the effi" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 703, + 133, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 703, + 133, + 713 + ], + "spans": [ + { + "bbox": [ + 58, + 703, + 133, + 713 + ], + "type": "text", + "content": "* corresponding author" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1084" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 289, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 289, + 384 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 289, + 384 + ], + "type": "text", + "content": "ciency problem, P2B [20] takes the cropped target template from the previous frame as well as the complete search area in the current frame as input, propagates target cues into the search area and then adopts a 3D region proposal network [18] to predict the current bounding box. P2B reaches a balance between performance and speed. Therefore many follow-up works adopt the same paradigm [3, 8, 9, 22, 29, 31, 33]. However, both SC3D and P2D paradigms overlook the contextual information across two consecutive frames and rely entirely on the appearance of the target. As mentioned in previous work [32], these methods are sensitive to appearance variation caused by occlusions and tend to drift towards intra-class distractors. To this end, M2-Track [32] introduces a novel motion-centric paradigm, which directly takes point clouds from two frames without cropping as input, and then segments the target points from their surroundings. After that, these points are cropped and the current bounding box is estimated by explicitly modeling motion between the two frames. Hence, the motion-centric paradigm still works on cropped patches that lack contextual information in later localization. In short, none of these methods could fully utilize the contextual information around the target to predict the current bounding box, which may degrade tracking performance due to the existence of large appearance variation and widespread distractors." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 391, + 289, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 391, + 289, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 391, + 289, + 714 + ], + "type": "text", + "content": "To address the above concerns, we propose a novel transformer-based tracker named CXTrack for 3D SOT, which exploits contextual information across two consecutive frames to improve the tracking performance. As shown in Fig. 1, different from paradigms commonly adopted by previous methods, CXTrack directly takes point clouds from the two consecutive frames as input, specifies the target of interest with the previous bounding box and predicts the current bounding box without any cropping, largely preserving contextual information. We first embed local geometric information of the two point clouds into point features using a shared backbone network. Then we integrate the targetness information into the point features according to the previous bounding box and adopt a target-centric transformer to propagate the target cues into the current frame while exploring contextual information in the surroundings of the target. After that, the enhanced point features are fed into a novel localization head named X-RPN to obtain the final target proposals. Specifically, X-RPN adopts a local transformer [25] to model point feature interactions within the target, which achieves a better balance between handling small and large objects compared with other localization heads. To distinguish the target from distractors, we incorporate a novel center embedding module into X-RPN, which embeds the relative target motion between two frames for explicit motion modeling. Extensive experiments on three popular tracking datasets demonstrate" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 304, + 72, + 545, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 108 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 108 + ], + "type": "text", + "content": "that CXTrack significantly outperforms the current state-of-the-art methods by a large margin while running at real-time (34 FPS) on a single NVIDIA RTX3090 GPU." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 304, + 108, + 547, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 108, + 547, + 217 + ], + "spans": [ + { + "bbox": [ + 304, + 108, + 547, + 217 + ], + "type": "text", + "content": "In short, our contributions can be summarized as: (1) a new paradigm for the real-time 3D SOT task, which fully exploits contextual information across consecutive frames to improve the tracking accuracy; (2) CXTrack: a transformer-based tracker that employs a target-centric transformer architecture to propagate targetness information and exploit contextual information; and (3) X-RPN: a localization head that is robust to intra-class distractors and achieves a good balance between small and large targets." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 306, + 227, + 392, + 239 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 227, + 392, + 239 + ], + "spans": [ + { + "bbox": [ + 306, + 227, + 392, + 239 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 247, + 547, + 378 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 247, + 547, + 378 + ], + "spans": [ + { + "bbox": [ + 304, + 247, + 547, + 378 + ], + "type": "text", + "content": "Early methods [13, 17, 23] for the 3D SOT task mainly focus on RGB-D information and tend to adopt 2D Siamese networks used in 2D object tracking with additional depth maps. However, the changes in illumination and appearance may degrade the performance of these RGB-D methods. As a pioneering work in this area, SC3D [6] crops the target from the previous frame with the previous bounding box, and then computes the cosine similarity between the target template and a series of 3D target proposals sampled from the current frame using a Siamese backbone. The pipeline relies on heuristic sampling, which is very time-consuming." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 379, + 547, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 379, + 547, + 665 + ], + "spans": [ + { + "bbox": [ + 304, + 379, + 547, + 665 + ], + "type": "text", + "content": "To address these issues, P2B [20] develops an end-to-end framework, which first employs a shared backbone to embed local geometry into point features, and then propagates target cues from the target template to the search area in the current frame. Finally, it adopts VoteNet [18] to generate 3D proposals and selects the proposal with the highest score as the target. P2B [20] reaches a balance between performance and efficiency, and many works follow the same paradigm. MLVSNet [29] aggregates information at multiple levels for more effective target localization. BAT [31] introduces a box-aware feature fusion module to enhance the correlation learning between the target template and the search area. V2B [8] proposes a voxel-to-BEV (Bird's Eye View) target localization network, which projects the point features into a dense BEV feature map to tackle the sparsity of point clouds. Inspired by the success of transformers [25], LTTR [3] adopts a transformer-based architecture to fuse features from two branches and propagate target cues. PTT [22] integrates a transformer module into the P2B architecture to refine point features. PTTR [33] introduces Point Relation Transformer for feature fusion and a light-weight Prediction Refinement Module for coarse-to-fine localization. ST-Net [9] develops an iterative coarse-to-fine correlation network for robust correlation learning." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "type": "text", + "content": "Although achieving promising results, the aforementioned methods crop the target from the previous frame using the given bounding box. This overlook of contextual information across two frames makes these methods sensitive" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1085" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 217 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 217 + ], + "type": "text", + "content": "to appearance variations caused by commonly occurred occlusions and thus the results tend to drift towards intra-class distractors, as mentioned in M2-Track [32]. To this end, M2-Track introduces a motion-centric paradigm to handle the 3D SOT problem, which directly takes the point clouds from two consecutive frames as input without cropping. It first localizes the target in the two frames by target segmentation, and then adopts PointNet [19] to predict the relative target motion from the cropped target area that lacks contextual information. M2-Track could not fully utilize local geometric and contextual information for prediction, which may hinder precise bounding box regression." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 224, + 104, + 237 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 224, + 104, + 237 + ], + "spans": [ + { + "bbox": [ + 47, + 224, + 104, + 237 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 244, + 160, + 257 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 244, + 160, + 257 + ], + "spans": [ + { + "bbox": [ + 47, + 244, + 160, + 257 + ], + "type": "text", + "content": "3.1. Problem Definition" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 263, + 289, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 263, + 289, + 441 + ], + "spans": [ + { + "bbox": [ + 46, + 263, + 289, + 441 + ], + "type": "text", + "content": "Given the initial state of the target, single object tracking aims to localize the target in a dynamic scene frame by frame. The initial state in the first frame is given as the 3D bounding box of the target, which can be parameterized by its center coordinates " + }, + { + "bbox": [ + 46, + 263, + 289, + 441 + ], + "type": "inline_equation", + "content": "(x,y,z)" + }, + { + "bbox": [ + 46, + 263, + 289, + 441 + ], + "type": "text", + "content": ", orientation angle " + }, + { + "bbox": [ + 46, + 263, + 289, + 441 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 46, + 263, + 289, + 441 + ], + "type": "text", + "content": " (around the up-axis, which is sufficient for most tracked objects staying on the ground) and sizes along each axis " + }, + { + "bbox": [ + 46, + 263, + 289, + 441 + ], + "type": "inline_equation", + "content": "(w,l,h)" + }, + { + "bbox": [ + 46, + 263, + 289, + 441 + ], + "type": "text", + "content": ". Since the tracking target has little change in size across frames even for non-rigid objects, we assume constant target size and only regress the translation offset " + }, + { + "bbox": [ + 46, + 263, + 289, + 441 + ], + "type": "inline_equation", + "content": "(\\Delta x,\\Delta y,\\Delta z)" + }, + { + "bbox": [ + 46, + 263, + 289, + 441 + ], + "type": "text", + "content": " and the rotation angle " + }, + { + "bbox": [ + 46, + 263, + 289, + 441 + ], + "type": "inline_equation", + "content": "(\\Delta \\theta)" + }, + { + "bbox": [ + 46, + 263, + 289, + 441 + ], + "type": "text", + "content": " between two consecutive frames to simplify the tracking task. By applying the translation and rotation to the 3D bounding box " + }, + { + "bbox": [ + 46, + 263, + 289, + 441 + ], + "type": "inline_equation", + "content": "\\mathcal{B}_{t - 1}\\in \\mathbb{R}^7" + }, + { + "bbox": [ + 46, + 263, + 289, + 441 + ], + "type": "text", + "content": " in the previous frame, we can compute the 3D bounding box " + }, + { + "bbox": [ + 46, + 263, + 289, + 441 + ], + "type": "inline_equation", + "content": "\\mathcal{B}_t\\in \\mathbb{R}^7" + }, + { + "bbox": [ + 46, + 263, + 289, + 441 + ], + "type": "text", + "content": " to localize the target in the current frame." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 443, + 289, + 541 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 443, + 289, + 541 + ], + "spans": [ + { + "bbox": [ + 46, + 443, + 289, + 541 + ], + "type": "text", + "content": "Suppose the point clouds in two consecutive frames are denoted as " + }, + { + "bbox": [ + 46, + 443, + 289, + 541 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_{t - 1}\\in \\mathbb{R}^{\\dot{N}_{t - 1}\\times 3}" + }, + { + "bbox": [ + 46, + 443, + 289, + 541 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 443, + 289, + 541 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_t\\in \\mathbb{R}^{\\dot{N}_t\\times 3}" + }, + { + "bbox": [ + 46, + 443, + 289, + 541 + ], + "type": "text", + "content": ", respectively, where " + }, + { + "bbox": [ + 46, + 443, + 289, + 541 + ], + "type": "inline_equation", + "content": "\\dot{N}_{t - 1}" + }, + { + "bbox": [ + 46, + 443, + 289, + 541 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 443, + 289, + 541 + ], + "type": "inline_equation", + "content": "\\dot{N}_t" + }, + { + "bbox": [ + 46, + 443, + 289, + 541 + ], + "type": "text", + "content": " are the numbers of points in the point clouds. We follow M2-Track [32] and encode the 3D bounding box " + }, + { + "bbox": [ + 46, + 443, + 289, + 541 + ], + "type": "inline_equation", + "content": "\\mathcal{B}_{t - 1}" + }, + { + "bbox": [ + 46, + 443, + 289, + 541 + ], + "type": "text", + "content": " as a targetness mask " + }, + { + "bbox": [ + 46, + 443, + 289, + 541 + ], + "type": "inline_equation", + "content": "\\dot{\\mathcal{M}}_{t - 1} = (m_{t - 1}^{1},m_{t - 1}^{2},\\dots ,m_{t - 1}^{\\dot{N}_{t - 1}})\\in \\mathbb{R}^{\\dot{N}_{t - 1}}" + }, + { + "bbox": [ + 46, + 443, + 289, + 541 + ], + "type": "text", + "content": " to indicate the target position, where the mask " + }, + { + "bbox": [ + 46, + 443, + 289, + 541 + ], + "type": "inline_equation", + "content": "m_{t - 1}^{i}" + }, + { + "bbox": [ + 46, + 443, + 289, + 541 + ], + "type": "text", + "content": " for the " + }, + { + "bbox": [ + 46, + 443, + 289, + 541 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 443, + 289, + 541 + ], + "type": "text", + "content": "-th point " + }, + { + "bbox": [ + 46, + 443, + 289, + 541 + ], + "type": "inline_equation", + "content": "p_{t - 1}^{i}" + }, + { + "bbox": [ + 46, + 443, + 289, + 541 + ], + "type": "text", + "content": " is defined as" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 100, + 545, + 287, + 571 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 100, + 545, + 287, + 571 + ], + "spans": [ + { + "bbox": [ + 100, + 545, + 287, + 571 + ], + "type": "interline_equation", + "content": "m _ {t - 1} ^ {i} = \\left\\{ \\begin{array}{l l} 0 & p _ {t - 1} ^ {i} \\text {n o t i n} \\mathcal {B} _ {t - 1} \\\\ 1 & p _ {t - 1} ^ {i} \\text {i n} \\mathcal {B} _ {t - 1} \\end{array} \\right. \\tag {1}", + "image_path": "59529899d8954965543e6023459ab1e12829c540110b5ffa455aa2d2605757d0.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 575, + 287, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 575, + 287, + 601 + ], + "spans": [ + { + "bbox": [ + 46, + 575, + 287, + 601 + ], + "type": "text", + "content": "Thus, the 3D SOT task can be formalized as learning the following mapping" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 79, + 605, + 287, + 619 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 79, + 605, + 287, + 619 + ], + "spans": [ + { + "bbox": [ + 79, + 605, + 287, + 619 + ], + "type": "interline_equation", + "content": "\\mathcal {F} \\left(\\mathcal {P} _ {t - 1}, \\dot {\\mathcal {M}} _ {t - 1}, \\mathcal {P} _ {t}\\right) \\mapsto (\\Delta x, \\Delta y, \\Delta z, \\Delta \\theta) \\tag {2}", + "image_path": "d560b85ae051e3c4a17dda985ac905bf7090a20add0d830d52b78506edc7d710.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 623, + 173, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 623, + 173, + 635 + ], + "spans": [ + { + "bbox": [ + 47, + 623, + 173, + 635 + ], + "type": "text", + "content": "3.2. Overview of CXTrack" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 641, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 641, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 641, + 289, + 715 + ], + "type": "text", + "content": "Following Eq. (2), we propose a network named CX-Track to improve tracking accuracy by fully exploiting contextual information across frames, and the overall design is illustrated in Fig. 2. We first apply a hierarchical feature learning architecture as the shared backbone to embed local geometric features of the point clouds into point" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 72, + 547, + 253 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 253 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 253 + ], + "type": "text", + "content": "features. We use " + }, + { + "bbox": [ + 304, + 72, + 547, + 253 + ], + "type": "inline_equation", + "content": "N_{t - 1}" + }, + { + "bbox": [ + 304, + 72, + 547, + 253 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 72, + 547, + 253 + ], + "type": "inline_equation", + "content": "N_{t}" + }, + { + "bbox": [ + 304, + 72, + 547, + 253 + ], + "type": "text", + "content": " to denote the numbers of point features extracted by the backbone. For convenience of calculation, we create a targetness mask " + }, + { + "bbox": [ + 304, + 72, + 547, + 253 + ], + "type": "inline_equation", + "content": "\\dot{\\mathcal{M}}_t" + }, + { + "bbox": [ + 304, + 72, + 547, + 253 + ], + "type": "text", + "content": " and fill it with 0.5 as it is unknown. We then concatenate the point features and targetness masks of the two frames to get " + }, + { + "bbox": [ + 304, + 72, + 547, + 253 + ], + "type": "inline_equation", + "content": "\\mathcal{X} = \\mathcal{X}_{t - 1}\\oplus \\mathcal{X}_t\\in \\mathbb{R}^{N\\times C}" + }, + { + "bbox": [ + 304, + 72, + 547, + 253 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 72, + 547, + 253 + ], + "type": "inline_equation", + "content": "\\mathcal{M} = \\mathcal{M}_{t - 1}\\oplus \\mathcal{M}_t\\in \\mathbb{R}^{N\\times 1}" + }, + { + "bbox": [ + 304, + 72, + 547, + 253 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 304, + 72, + 547, + 253 + ], + "type": "inline_equation", + "content": "N = N_{t - 1} + N_{t}" + }, + { + "bbox": [ + 304, + 72, + 547, + 253 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 72, + 547, + 253 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_{t - 1}" + }, + { + "bbox": [ + 304, + 72, + 547, + 253 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 72, + 547, + 253 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_t" + }, + { + "bbox": [ + 304, + 72, + 547, + 253 + ], + "type": "text", + "content": " are masks corresponding to point features, and extracted from " + }, + { + "bbox": [ + 304, + 72, + 547, + 253 + ], + "type": "inline_equation", + "content": "\\dot{\\mathcal{M}}_{t - 1}" + }, + { + "bbox": [ + 304, + 72, + 547, + 253 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 72, + 547, + 253 + ], + "type": "inline_equation", + "content": "\\dot{\\mathcal{M}}_t" + }, + { + "bbox": [ + 304, + 72, + 547, + 253 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 72, + 547, + 253 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 304, + 72, + 547, + 253 + ], + "type": "text", + "content": " is the number of channels for point features. We employ the target-centric transformer (Sec. 3.3) to integrate the targetness mask information into point features while exploring the contextual information across frames. Finally, we propose a novel localization network, named X-RPN (Sec. 3.4), to obtain the target proposals. The proposal with the highest targetness score is verified as the result." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 263, + 462, + 277 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 263, + 462, + 277 + ], + "spans": [ + { + "bbox": [ + 305, + 263, + 462, + 277 + ], + "type": "text", + "content": "3.3. Target-Centric Transformer" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 284, + 547, + 488 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 284, + 547, + 488 + ], + "spans": [ + { + "bbox": [ + 304, + 284, + 547, + 488 + ], + "type": "text", + "content": "Target-Centric Transformer aims to enhance the point features using the contextual information around the target while propagating the target cues from the previous frame to the current frame. It is composed of " + }, + { + "bbox": [ + 304, + 284, + 547, + 488 + ], + "type": "inline_equation", + "content": "N_{L} = 4" + }, + { + "bbox": [ + 304, + 284, + 547, + 488 + ], + "type": "text", + "content": " identical layers stacked in series. Given the point features " + }, + { + "bbox": [ + 304, + 284, + 547, + 488 + ], + "type": "inline_equation", + "content": "\\mathcal{X}^{(k - 1)}\\in \\mathbb{R}^{N\\times C}" + }, + { + "bbox": [ + 304, + 284, + 547, + 488 + ], + "type": "text", + "content": " and the targetness mask " + }, + { + "bbox": [ + 304, + 284, + 547, + 488 + ], + "type": "inline_equation", + "content": "\\mathcal{M}^{(k - 1)}" + }, + { + "bbox": [ + 304, + 284, + 547, + 488 + ], + "type": "text", + "content": " from the " + }, + { + "bbox": [ + 304, + 284, + 547, + 488 + ], + "type": "inline_equation", + "content": "(k - 1)" + }, + { + "bbox": [ + 304, + 284, + 547, + 488 + ], + "type": "text", + "content": "-th layer as input " + }, + { + "bbox": [ + 304, + 284, + 547, + 488 + ], + "type": "inline_equation", + "content": "(\\mathcal{M}^{(0)} = \\mathcal{M}" + }, + { + "bbox": [ + 304, + 284, + 547, + 488 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 284, + 547, + 488 + ], + "type": "inline_equation", + "content": "\\mathcal{X}^{(0)} = \\mathcal{X})" + }, + { + "bbox": [ + 304, + 284, + 547, + 488 + ], + "type": "text", + "content": ", the " + }, + { + "bbox": [ + 304, + 284, + 547, + 488 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 284, + 547, + 488 + ], + "type": "text", + "content": "-th layer first models the interactions between any two points while integrating the targetness mask into point features using a modified self-attention operation, and then adopts Multi-Layer-Perceptrons (MLPs) to compute the new point features " + }, + { + "bbox": [ + 304, + 284, + 547, + 488 + ], + "type": "inline_equation", + "content": "\\mathcal{X}^{(k)}" + }, + { + "bbox": [ + 304, + 284, + 547, + 488 + ], + "type": "text", + "content": " as well as the refined targetness mask " + }, + { + "bbox": [ + 304, + 284, + 547, + 488 + ], + "type": "inline_equation", + "content": "\\mathcal{M}^{(k)}" + }, + { + "bbox": [ + 304, + 284, + 547, + 488 + ], + "type": "text", + "content": ". Thus, the predicted targetness mask will be consistently refined layer by layer. Moreover, we found it beneficial to add an auxiliary loss by predicting a potential target center for each point via Hough voting, so each layer also applies a shared MLP to generate the potential target center " + }, + { + "bbox": [ + 304, + 284, + 547, + 488 + ], + "type": "inline_equation", + "content": "C^{(k)}\\in \\mathbb{R}^{N\\times 3}" + }, + { + "bbox": [ + 304, + 284, + 547, + 488 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 489, + 547, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 489, + 547, + 525 + ], + "spans": [ + { + "bbox": [ + 304, + 489, + 547, + 525 + ], + "type": "text", + "content": "Formally, we first employ layer normalization [1] " + }, + { + "bbox": [ + 304, + 489, + 547, + 525 + ], + "type": "inline_equation", + "content": "\\mathrm{LN}(\\cdot)" + }, + { + "bbox": [ + 304, + 489, + 547, + 525 + ], + "type": "text", + "content": " before the self-attention mechanism [25] following the design of 3DETR [15], which can be written as" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 387, + 536, + 545, + 551 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 387, + 536, + 545, + 551 + ], + "spans": [ + { + "bbox": [ + 387, + 536, + 545, + 551 + ], + "type": "interline_equation", + "content": "\\bar {X} = \\mathrm {L N} \\left(\\mathcal {X} ^ {(k - 1)}\\right) \\tag {3}", + "image_path": "ce49692fe36752b1bff0f1eb08f962a613259bc4e6277b0e9ad7d555a63b7d43.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 563, + 547, + 599 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 563, + 547, + 599 + ], + "spans": [ + { + "bbox": [ + 304, + 563, + 547, + 599 + ], + "type": "text", + "content": "Then, we add positional encodings (PE) of the coordinates to the normalized point features before feeding them into the self-attention operation" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 378, + 610, + 545, + 624 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 378, + 610, + 545, + 624 + ], + "spans": [ + { + "bbox": [ + 378, + 610, + 545, + 624 + ], + "type": "interline_equation", + "content": "X _ {Q} = X _ {K} = \\bar {X} + \\mathrm {P E} \\tag {4}", + "image_path": "7d8ddaa94b63ca5bb66fa9871b0632c8760e43cbb6197ebcb3ff6c2bf627d00a.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 408, + 627, + 545, + 641 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 408, + 627, + 545, + 641 + ], + "spans": [ + { + "bbox": [ + 408, + 627, + 545, + 641 + ], + "type": "interline_equation", + "content": "X _ {V} = \\bar {X} \\tag {5}", + "image_path": "44fb583e61d0380af4daa7023f98693813a0109553985fb8bb881893d4c3fabf.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 654, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 547, + 715 + ], + "type": "text", + "content": "It is worth noting that we only adopt PE for the query and key branches, therefore each refined point feature is constrained to focus more on local geometry instead of its associated absolute position. Subsequently, the transformer layer employs a global self-attention operation to model the" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1086" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 47, + 68, + 544, + 183 + ], + "blocks": [ + { + "bbox": [ + 47, + 68, + 544, + 183 + ], + "lines": [ + { + "bbox": [ + 47, + 68, + 544, + 183 + ], + "spans": [ + { + "bbox": [ + 47, + 68, + 544, + 183 + ], + "type": "image", + "image_path": "7ea30e35c661f2e919c71b0854d6d3650642d285775be33f81fd7d3ada7aa47f.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 45, + 190, + 548, + 236 + ], + "lines": [ + { + "bbox": [ + 45, + 190, + 548, + 236 + ], + "spans": [ + { + "bbox": [ + 45, + 190, + 548, + 236 + ], + "type": "text", + "content": "Figure 2. The overall architecture of CXTrack. Given two consecutive point clouds and the 3D bounding box in the previous frame, CXTrack first embeds the local geometry into point features using the backbone. Then, CXTrack employs the target-centric transformer to explore contextual information across two frames and propagate the target cues to the current frame. Finally, the enhanced features are fed into a novel localization network named X-RPN to obtain high-quality proposals for verification." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 47, + 239, + 547, + 376 + ], + "blocks": [ + { + "bbox": [ + 47, + 239, + 547, + 376 + ], + "lines": [ + { + "bbox": [ + 47, + 239, + 547, + 376 + ], + "spans": [ + { + "bbox": [ + 47, + 239, + 547, + 376 + ], + "type": "image", + "image_path": "80369015367aba9cb344d87501ad2babdbf0830d30fff22e886393a47cbc9065.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 45, + 380, + 547, + 413 + ], + "lines": [ + { + "bbox": [ + 45, + 380, + 547, + 413 + ], + "spans": [ + { + "bbox": [ + 45, + 380, + 547, + 413 + ], + "type": "text", + "content": "Figure 3. Comparison of various transformer layers to fuse the targetness mask and point features. We introduce three types of target-centric transformer layers, namely Vanilla, Semi-Dropout and Gated layer to integrate the targetness mask information into the point features while modeling intra-frame and inter-frame feature relationships." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 420, + 255, + 432 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 420, + 255, + 432 + ], + "spans": [ + { + "bbox": [ + 47, + 420, + 255, + 432 + ], + "type": "text", + "content": "relationships between point features, formulated as" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 52, + 441, + 287, + 455 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 52, + 441, + 287, + 455 + ], + "spans": [ + { + "bbox": [ + 52, + 441, + 287, + 455 + ], + "type": "interline_equation", + "content": "\\mathrm {M H A} \\left(X _ {Q}, X _ {K}, X _ {V}\\right) = \\operatorname {C o n c a t} \\left(\\text {h e a d} _ {1}, \\dots \\text {h e a d} _ {h}\\right) W ^ {O} \\tag {6}", + "image_path": "9b6c325652e31ed3a0d6b26fa488a9630ed34ba0ffb1600965c7ddc20f7b50da.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 87, + 457, + 287, + 483 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 87, + 457, + 287, + 483 + ], + "spans": [ + { + "bbox": [ + 87, + 457, + 287, + 483 + ], + "type": "interline_equation", + "content": "\\text {w h e r e} \\operatorname {h e a d} _ {i} = \\operatorname {A t t n} \\left(X _ {Q} W _ {i} ^ {Q}, X _ {K} W _ {i} ^ {K}, \\bar {X} W _ {i} ^ {V}\\right), \\tag {7}", + "image_path": "32239c6e5bfe2188ad17285b7a9232690eb17f8dc7fb75e56508febc558472c6.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 80, + 486, + 287, + 514 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 80, + 486, + 287, + 514 + ], + "spans": [ + { + "bbox": [ + 80, + 486, + 287, + 514 + ], + "type": "interline_equation", + "content": "\\operatorname {A t t n} (Q, K, V) = \\operatorname {s o f t m a x} \\left(\\frac {Q K ^ {T}}{\\sqrt {d _ {k}}}\\right) V \\tag {8}", + "image_path": "720ed49295b5e0627bac1b644319878c0841c6eb17556095b388e6b30af1b099.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 521, + 287, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 521, + 287, + 593 + ], + "spans": [ + { + "bbox": [ + 47, + 521, + 287, + 593 + ], + "type": "text", + "content": "Here, MHA indicates a multi-head attention, where the attention is applied in " + }, + { + "bbox": [ + 47, + 521, + 287, + 593 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 47, + 521, + 287, + 593 + ], + "type": "text", + "content": " subspaces before concatenation. The projections are implemented by parameter matrices " + }, + { + "bbox": [ + 47, + 521, + 287, + 593 + ], + "type": "inline_equation", + "content": "W_{i}^{Q}\\in \\mathbb{R}^{C\\times d_{k}}" + }, + { + "bbox": [ + 47, + 521, + 287, + 593 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 47, + 521, + 287, + 593 + ], + "type": "inline_equation", + "content": "W_{i}^{K}\\in \\mathbb{R}^{C\\times d_{k}}" + }, + { + "bbox": [ + 47, + 521, + 287, + 593 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 47, + 521, + 287, + 593 + ], + "type": "inline_equation", + "content": "W_{i}^{V}\\in \\mathbb{R}^{C\\times d_{v}}" + }, + { + "bbox": [ + 47, + 521, + 287, + 593 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 521, + 287, + 593 + ], + "type": "inline_equation", + "content": "W_{i}^{O}\\in \\mathbb{R}^{hd_{v}\\times C}" + }, + { + "bbox": [ + 47, + 521, + 287, + 593 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 47, + 521, + 287, + 593 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 47, + 521, + 287, + 593 + ], + "type": "text", + "content": " indicates the " + }, + { + "bbox": [ + 47, + 521, + 287, + 593 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 47, + 521, + 287, + 593 + ], + "type": "text", + "content": "-th subspace. The self-attention sublayer can be written as" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 64, + 601, + 287, + 616 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 601, + 287, + 616 + ], + "spans": [ + { + "bbox": [ + 64, + 601, + 287, + 616 + ], + "type": "interline_equation", + "content": "\\widehat {X} = \\mathcal {X} ^ {(k - 1)} + \\operatorname {D r o p o u t} \\left(\\mathrm {M H A} \\left(X _ {Q}, X _ {K}, X _ {V}\\right)\\right) \\tag {9}", + "image_path": "e44ff07581f016b148920ffac9bccbb0cfcfe9d998dcd1e9a9631342e1804b6d.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 625, + 288, + 673 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 625, + 288, + 673 + ], + "spans": [ + { + "bbox": [ + 47, + 625, + 288, + 673 + ], + "type": "text", + "content": "In addition to the self-attention sublayer, each transformer layer also contains a fully connected feed-forward network to refine the point features. The final output of the " + }, + { + "bbox": [ + 47, + 625, + 288, + 673 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 47, + 625, + 288, + 673 + ], + "type": "text", + "content": "-th transformer layer is given by" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 100, + 681, + 287, + 696 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 100, + 681, + 287, + 696 + ], + "spans": [ + { + "bbox": [ + 100, + 681, + 287, + 696 + ], + "type": "interline_equation", + "content": "\\mathcal {X} ^ {(k)} = \\widehat {X} + \\operatorname {D r o p o u t} (\\operatorname {F F N} (\\operatorname {L N} (\\widehat {X}))), \\tag {10}", + "image_path": "67299c4057ff2651e3e9abb3acb5b637d130d8632d79d673241641904dd886d1.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 59, + 697, + 287, + 710 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 697, + 287, + 710 + ], + "spans": [ + { + "bbox": [ + 59, + 697, + 287, + 710 + ], + "type": "interline_equation", + "content": "\\text {w h e r e} \\operatorname {F F N} (x) = \\max \\left(0, x W _ {1} + b _ {1}\\right) W _ {2} + b _ {2}. \\tag {11}", + "image_path": "a1f35e05b933a00d5781e7a6b181c6f04164c6100b2f0fcb9077273069ddc22e.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 420, + 545, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 420, + 545, + 468 + ], + "spans": [ + { + "bbox": [ + 305, + 420, + 545, + 468 + ], + "type": "text", + "content": "To integrate the targetness mask information into point features, we need to modify the classic transformer layer. We introduce three types of modified transformer layers in Fig. 3, namely Vanilla, Semi-Dropout and Gated layer." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 468, + 545, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 468, + 545, + 517 + ], + "spans": [ + { + "bbox": [ + 305, + 468, + 545, + 517 + ], + "type": "text", + "content": "Vanilla. We project the input " + }, + { + "bbox": [ + 305, + 468, + 545, + 517 + ], + "type": "inline_equation", + "content": "\\mathcal{M}^{(k - 1)}" + }, + { + "bbox": [ + 305, + 468, + 545, + 517 + ], + "type": "text", + "content": " to mask embedding " + }, + { + "bbox": [ + 305, + 468, + 545, + 517 + ], + "type": "inline_equation", + "content": "\\mathrm{ME}\\in \\mathbb{R}^{N\\times C}" + }, + { + "bbox": [ + 305, + 468, + 545, + 517 + ], + "type": "text", + "content": " using a linear transformation. Following the design of positional encoding, we simply add ME to the input token embedding " + }, + { + "bbox": [ + 305, + 468, + 545, + 517 + ], + "type": "inline_equation", + "content": "X_{V}" + }, + { + "bbox": [ + 305, + 468, + 545, + 517 + ], + "type": "text", + "content": ", which re-formulates Eq. (5) as" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 392, + 529, + 545, + 541 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 392, + 529, + 545, + 541 + ], + "spans": [ + { + "bbox": [ + 392, + 529, + 545, + 541 + ], + "type": "interline_equation", + "content": "X _ {V} = \\bar {X} + \\mathrm {M E} \\tag {12}", + "image_path": "e65b1e4fd59733a218f3db55dba8b6fcbcb1d140885f14c841498188f150357b.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 549, + 546, + 669 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 549, + 546, + 669 + ], + "spans": [ + { + "bbox": [ + 304, + 549, + 546, + 669 + ], + "type": "text", + "content": "Semi-Dropout. Notably, the targetness mask information can only flow across layers along the attention path. For small objects which only have a few points to track, applying dropout to the mask embedding may discard the targetness information and lead to performance degradation. To this end, we separate the self-attention mechanism into a feature branch and a mask branch with shared attention weights, while only applying dropout to the refined point features. As shown in Fig. 3b, the self-attention sublayer in Eq. (9) is re-formulated as" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 315, + 680, + 545, + 710 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 680, + 545, + 710 + ], + "spans": [ + { + "bbox": [ + 315, + 680, + 545, + 710 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\widehat {X} = \\mathcal {X} ^ {(k - 1)} + \\operatorname {D r o p o u t} (\\mathrm {M H A} (X _ {Q}, X _ {K}, \\overline {{X}})) \\\\ + \\mathrm {M H A} \\left(X _ {Q}, X _ {K}, \\mathrm {M E}\\right) \\tag {13} \\\\ \\end{array}", + "image_path": "331735028c5421a7da939bf7ae7599a3676e0c58753046f95197fb565b589698.jpg" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1087" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 61, + 70, + 272, + 233 + ], + "blocks": [ + { + "bbox": [ + 61, + 70, + 272, + 233 + ], + "lines": [ + { + "bbox": [ + 61, + 70, + 272, + 233 + ], + "spans": [ + { + "bbox": [ + 61, + 70, + 272, + 233 + ], + "type": "image", + "image_path": "4a939d898bf5efa9aa6bdb33f8bbe36dc10ce6615a2e25b37222d34b0a53616f.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 236, + 288, + 293 + ], + "lines": [ + { + "bbox": [ + 46, + 236, + 288, + 293 + ], + "spans": [ + { + "bbox": [ + 46, + 236, + 288, + 293 + ], + "type": "text", + "content": "Figure 4. The overall architecture of X-RPN. X-RPN adopts a local transformer to model point feature interaction within the target and aggregate local clues. It also incorporates a center embedding mechanism which embeds the relative target motion between two frames to distinguish the target from distractors." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 298, + 287, + 417 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 298, + 287, + 417 + ], + "spans": [ + { + "bbox": [ + 46, + 298, + 287, + 417 + ], + "type": "text", + "content": "Gated. Inspired by the design of TrDimp [26], we introduce a gated mechanism into the self-attention sublayer to integrate the mask information. It has two parallel branches, namely mask transformation and feature transformation. For mask transformation, we first obtain the feature mask " + }, + { + "bbox": [ + 46, + 298, + 287, + 417 + ], + "type": "inline_equation", + "content": "\\overline{M} \\in \\mathbb{R}^{N \\times C}" + }, + { + "bbox": [ + 46, + 298, + 287, + 417 + ], + "type": "text", + "content": " by repeating the input point-wise mask " + }, + { + "bbox": [ + 46, + 298, + 287, + 417 + ], + "type": "inline_equation", + "content": "\\mathcal{M}^{(k - 1)} \\in \\mathbb{R}^{N \\times 1}" + }, + { + "bbox": [ + 46, + 298, + 287, + 417 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 46, + 298, + 287, + 417 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 46, + 298, + 287, + 417 + ], + "type": "text", + "content": " times. Then we can propagate the targetness cues to the current frame via adopting self-attention on the mask feature. The transformed mask serves as the gate matrix for the point features" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 87, + 425, + 287, + 440 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 87, + 425, + 287, + 440 + ], + "spans": [ + { + "bbox": [ + 87, + 425, + 287, + 440 + ], + "type": "interline_equation", + "content": "\\widehat {X} _ {m} = \\mathrm {L N} (\\mathrm {M H A} (X _ {Q}, X _ {K}, \\bar {M}) \\otimes \\bar {X}). \\tag {14}", + "image_path": "568aff33ebbe87faabcb5a4f9e0912da7daf15c8258efedb0e89add0549979ab.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 448, + 287, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 448, + 287, + 495 + ], + "spans": [ + { + "bbox": [ + 46, + 448, + 287, + 495 + ], + "type": "text", + "content": "For feature transformation, we first mask the point features to suppress feature activation in background areas, and then employ self-attention with a residual connection to model the relationships between features" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 70, + 502, + 287, + 517 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 502, + 287, + 517 + ], + "spans": [ + { + "bbox": [ + 70, + 502, + 287, + 517 + ], + "type": "interline_equation", + "content": "\\widehat {X} _ {f} = \\mathrm {L N} (\\mathrm {M H A} (X _ {Q}, X _ {K}, \\bar {M} \\otimes \\bar {X}) + \\bar {X}). \\tag {15}", + "image_path": "9fee1f0cc58e68bdb28a9615b33776aaf684debea9d43217044d0450bcf226c9.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 525, + 287, + 561 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 525, + 287, + 561 + ], + "spans": [ + { + "bbox": [ + 46, + 525, + 287, + 561 + ], + "type": "text", + "content": "As illustrated in Fig. 3c, we sum and normalize the output features " + }, + { + "bbox": [ + 46, + 525, + 287, + 561 + ], + "type": "inline_equation", + "content": "\\widehat{X}_m" + }, + { + "bbox": [ + 46, + 525, + 287, + 561 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 525, + 287, + 561 + ], + "type": "inline_equation", + "content": "\\widehat{X}_f" + }, + { + "bbox": [ + 46, + 525, + 287, + 561 + ], + "type": "text", + "content": " from the two branches. Eq. (9) can be re-formulated as" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 122, + 568, + 287, + 583 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 122, + 568, + 287, + 583 + ], + "spans": [ + { + "bbox": [ + 122, + 568, + 287, + 583 + ], + "type": "interline_equation", + "content": "\\widehat {X} = \\operatorname {L N} \\left(\\widehat {X} _ {f} + \\widehat {X} _ {m}\\right). \\tag {16}", + "image_path": "4cf9df68bdc3efffe3096f95e1fe861df47f99219826837deb20c3dc42c34d2b.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 592, + 287, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 592, + 287, + 652 + ], + "spans": [ + { + "bbox": [ + 46, + 592, + 287, + 652 + ], + "type": "text", + "content": "Among the above three layers, we observe significant performance gain from using Semi-Dropout target-centric transformer layers (Sec. 4.3). Thus CXTrack employs Semi-Dropout layers to integrate targetness information while exploring contextual information across frames." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 658, + 104, + 670 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 658, + 104, + 670 + ], + "spans": [ + { + "bbox": [ + 47, + 658, + 104, + 670 + ], + "type": "text", + "content": "3.4. X-RPN" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "text", + "content": "Previous works [20] indicate that individual point features can only capture limited local information, which may not be sufficient for precise bounding box regression. Thus" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 72, + 545, + 265 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 265 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 265 + ], + "type": "text", + "content": "we develop a simple yet effective localization network, named X-RPN, which extends RPN [18] using local transformer and center embedding, as shown in Fig. 4. Different from previous works [8,20], X-RPN aggregates local clues from point features without downsampling or voxelization, thus avoiding information loss and reaching a good balance between handling large and small objects. Our intuition is that each point should only interact with points belonging to the same object to suppress irrelevant information. Given the point features " + }, + { + "bbox": [ + 304, + 72, + 545, + 265 + ], + "type": "inline_equation", + "content": "\\mathcal{X}^{(N_L)}" + }, + { + "bbox": [ + 304, + 72, + 545, + 265 + ], + "type": "text", + "content": ", targetness mask " + }, + { + "bbox": [ + 304, + 72, + 545, + 265 + ], + "type": "inline_equation", + "content": "\\mathcal{M}^{(N_L)}" + }, + { + "bbox": [ + 304, + 72, + 545, + 265 + ], + "type": "text", + "content": " and target center " + }, + { + "bbox": [ + 304, + 72, + 545, + 265 + ], + "type": "inline_equation", + "content": "\\mathcal{C}^{(N_L)}" + }, + { + "bbox": [ + 304, + 72, + 545, + 265 + ], + "type": "text", + "content": " output by the target centric transformer, we first split them along the spatial dimension and only feed those belonging to the current frame into X-RPN, which is denoted as " + }, + { + "bbox": [ + 304, + 72, + 545, + 265 + ], + "type": "inline_equation", + "content": "\\widetilde{\\mathcal{X}}\\in \\mathbb{R}^{N_t\\times C}" + }, + { + "bbox": [ + 304, + 72, + 545, + 265 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 72, + 545, + 265 + ], + "type": "inline_equation", + "content": "\\widetilde{\\mathcal{C}}\\in \\mathbb{R}^{N_t\\times 3}" + }, + { + "bbox": [ + 304, + 72, + 545, + 265 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 72, + 545, + 265 + ], + "type": "inline_equation", + "content": "\\widetilde{\\mathcal{M}}\\in \\mathbb{R}^{N_t\\times 1}" + }, + { + "bbox": [ + 304, + 72, + 545, + 265 + ], + "type": "text", + "content": ", respectively. X-RPN first computes the neighborhood " + }, + { + "bbox": [ + 304, + 72, + 545, + 265 + ], + "type": "inline_equation", + "content": "\\mathcal{N}(p_i)" + }, + { + "bbox": [ + 304, + 72, + 545, + 265 + ], + "type": "text", + "content": " for each point " + }, + { + "bbox": [ + 304, + 72, + 545, + 265 + ], + "type": "inline_equation", + "content": "p_i" + }, + { + "bbox": [ + 304, + 72, + 545, + 265 + ], + "type": "text", + "content": " using its potential target center " + }, + { + "bbox": [ + 304, + 72, + 545, + 265 + ], + "type": "inline_equation", + "content": "c_{i}" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 358, + 272, + 545, + 293 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 358, + 272, + 545, + 293 + ], + "spans": [ + { + "bbox": [ + 358, + 272, + 545, + 293 + ], + "type": "interline_equation", + "content": "\\mathcal {N} \\left(p _ {i}\\right) = \\left\\{p _ {j} \\mid | | c _ {i} - c _ {j} | | _ {2} < r \\right\\} \\tag {17}", + "image_path": "a20f26d7b63c294c5ef127b8a15964a065c885aaa351d9098b50fe143bdf430e.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 299, + 545, + 383 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 299, + 545, + 383 + ], + "spans": [ + { + "bbox": [ + 304, + 299, + 545, + 383 + ], + "type": "text", + "content": "Here " + }, + { + "bbox": [ + 304, + 299, + 545, + 383 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 304, + 299, + 545, + 383 + ], + "type": "text", + "content": " is a hyperparameter indicating the size of the neighborhood. Then X-RPN adopts the transformer architecture mentioned in Sec. 3.3 to aggregate local information, where each point only interacts with its neighborhood points to suppress noise. We remove the feed-forward network in the transformer layer because we observe that one layer is sufficient to generate high quality target proposals." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 383, + 545, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 383, + 545, + 515 + ], + "spans": [ + { + "bbox": [ + 304, + 383, + 545, + 515 + ], + "type": "text", + "content": "To deal with intra-class distractors which are widespread in the scenes [32] especially for pedestrian tracking, we propose to combine the potential center information with the targetness mask. Our intuition lies in two folds. First, the tracking target keeps similar local geometry across two frames. Second, if the duration between two consecutive frames is sufficiently short, the displacement of the target is small. Therefore, we construct a Gaussian proposal-wise mask " + }, + { + "bbox": [ + 304, + 383, + 545, + 515 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_c" + }, + { + "bbox": [ + 304, + 383, + 545, + 515 + ], + "type": "text", + "content": " to indicate the magnitude of the displacement of each proposal. Formally, for each point " + }, + { + "bbox": [ + 304, + 383, + 545, + 515 + ], + "type": "inline_equation", + "content": "p_i" + }, + { + "bbox": [ + 304, + 383, + 545, + 515 + ], + "type": "text", + "content": " with the predicted target center " + }, + { + "bbox": [ + 304, + 383, + 545, + 515 + ], + "type": "inline_equation", + "content": "c_i" + }, + { + "bbox": [ + 304, + 383, + 545, + 515 + ], + "type": "text", + "content": ", the mask value " + }, + { + "bbox": [ + 304, + 383, + 545, + 515 + ], + "type": "inline_equation", + "content": "m_i^c \\in \\mathcal{M}_c" + }, + { + "bbox": [ + 304, + 383, + 545, + 515 + ], + "type": "text", + "content": " is" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 374, + 522, + 545, + 547 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 374, + 522, + 545, + 547 + ], + "spans": [ + { + "bbox": [ + 374, + 522, + 545, + 547 + ], + "type": "interline_equation", + "content": "m _ {i} ^ {c} = \\exp \\left(- \\frac {\\left| \\left| c _ {i} - \\bar {c} \\right| \\right| _ {2} ^ {2}}{2 \\sigma^ {2}}\\right) \\tag {18}", + "image_path": "c2c8d072e694a0f8a933e1ab57d5db36e381283f6251d99785de065bc2655484.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 554, + 545, + 615 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 554, + 545, + 615 + ], + "spans": [ + { + "bbox": [ + 304, + 554, + 545, + 615 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 554, + 545, + 615 + ], + "type": "inline_equation", + "content": "\\overline{c} \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 304, + 554, + 545, + 615 + ], + "type": "text", + "content": " is the target center in the previous frame and " + }, + { + "bbox": [ + 304, + 554, + 545, + 615 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 304, + 554, + 545, + 615 + ], + "type": "text", + "content": " is a learnable or fixed scaling factor. We embed the target center mask " + }, + { + "bbox": [ + 304, + 554, + 545, + 615 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_c" + }, + { + "bbox": [ + 304, + 554, + 545, + 615 + ], + "type": "text", + "content": " into the center embedding matrix " + }, + { + "bbox": [ + 304, + 554, + 545, + 615 + ], + "type": "inline_equation", + "content": "\\mathrm{CE} \\in \\mathbb{R}^{N \\times C}" + }, + { + "bbox": [ + 304, + 554, + 545, + 615 + ], + "type": "text", + "content": " using a linear transformation, and equally combine the mask embedding and the center embedding." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 621, + 399, + 633 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 621, + 399, + 633 + ], + "spans": [ + { + "bbox": [ + 306, + 621, + 399, + 633 + ], + "type": "text", + "content": "3.5. Loss Functions" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 639, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 639, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 639, + 545, + 713 + ], + "type": "text", + "content": "For the prediction " + }, + { + "bbox": [ + 304, + 639, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{M}^{(k)}" + }, + { + "bbox": [ + 304, + 639, + 545, + 713 + ], + "type": "text", + "content": " given by the " + }, + { + "bbox": [ + 304, + 639, + 545, + 713 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 639, + 545, + 713 + ], + "type": "text", + "content": "-th transformer layer, we adopt a standard cross entropy loss " + }, + { + "bbox": [ + 304, + 639, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{cm}}^{(i)}" + }, + { + "bbox": [ + 304, + 639, + 545, + 713 + ], + "type": "text", + "content": ". As for the potential target centers, we observe that it is difficult to regress precise centers for non-rigid objects such as pedestrians. Hence the predicted centers " + }, + { + "bbox": [ + 304, + 639, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{C}^{(k)}" + }, + { + "bbox": [ + 304, + 639, + 545, + 713 + ], + "type": "text", + "content": " are supervised by " + }, + { + "bbox": [ + 304, + 639, + 545, + 713 + ], + "type": "inline_equation", + "content": "L_{2}" + }, + { + "bbox": [ + 304, + 639, + 545, + 713 + ], + "type": "text", + "content": " loss for non-rigid objects, and by Huber loss [21] for" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1088" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 71, + 286, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 71, + 286, + 96 + ], + "spans": [ + { + "bbox": [ + 46, + 71, + 286, + 96 + ], + "type": "text", + "content": "rigid objects. For the target center regression loss " + }, + { + "bbox": [ + 46, + 71, + 286, + 96 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{cc}}^{(i)}" + }, + { + "bbox": [ + 46, + 71, + 286, + 96 + ], + "type": "text", + "content": ", only points in the ground truth bounding box are supervised." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 96, + 286, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 96, + 286, + 180 + ], + "spans": [ + { + "bbox": [ + 46, + 96, + 286, + 180 + ], + "type": "text", + "content": "Following previous works [20], proposals with predicted centers near the target center " + }, + { + "bbox": [ + 46, + 96, + 286, + 180 + ], + "type": "inline_equation", + "content": "(< 0.3\\mathrm{m})" + }, + { + "bbox": [ + 46, + 96, + 286, + 180 + ], + "type": "text", + "content": " are considered as positives and those far away " + }, + { + "bbox": [ + 46, + 96, + 286, + 180 + ], + "type": "inline_equation", + "content": "(>0.6\\mathrm{m})" + }, + { + "bbox": [ + 46, + 96, + 286, + 180 + ], + "type": "text", + "content": " are considered as negatives. Others are left unsupervised. The predicted targetness mask is supervised via standard cross-entropy loss " + }, + { + "bbox": [ + 46, + 96, + 286, + 180 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{rm}}" + }, + { + "bbox": [ + 46, + 96, + 286, + 180 + ], + "type": "text", + "content": " and only the bounding box parameters of positive predictions are supervised by Huber (Smooth-" + }, + { + "bbox": [ + 46, + 96, + 286, + 180 + ], + "type": "inline_equation", + "content": "L_{1}" + }, + { + "bbox": [ + 46, + 96, + 286, + 180 + ], + "type": "text", + "content": ") loss " + }, + { + "bbox": [ + 46, + 96, + 286, + 180 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{box}}" + }, + { + "bbox": [ + 46, + 96, + 286, + 180 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 180, + 286, + 202 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 180, + 286, + 202 + ], + "spans": [ + { + "bbox": [ + 46, + 180, + 286, + 202 + ], + "type": "text", + "content": "The overall loss is the weighted combination of the above loss terms" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 63, + 208, + 286, + 240 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 208, + 286, + 240 + ], + "spans": [ + { + "bbox": [ + 63, + 208, + 286, + 240 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\gamma_ {1} \\sum_ {i = 1} ^ {N _ {L}} \\mathcal {L} _ {\\mathrm {c m}} ^ {(i)} + \\gamma_ {2} \\sum_ {i = 1} ^ {N _ {L}} \\mathcal {L} _ {\\mathrm {c c}} ^ {(i)} + \\gamma_ {3} \\mathcal {L} _ {\\mathrm {r m}} + \\mathcal {L} _ {\\mathrm {b o x}} \\tag {19}", + "image_path": "a06963343ec4a382a8d8ec007b9e94bcdaa5ee52ef43b7e2dba251b63d3c3b2b.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 247, + 286, + 283 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 247, + 286, + 283 + ], + "spans": [ + { + "bbox": [ + 46, + 247, + 286, + 283 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 247, + 286, + 283 + ], + "type": "inline_equation", + "content": "\\gamma_{1},\\gamma_{2}" + }, + { + "bbox": [ + 46, + 247, + 286, + 283 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 247, + 286, + 283 + ], + "type": "inline_equation", + "content": "\\gamma_{3}" + }, + { + "bbox": [ + 46, + 247, + 286, + 283 + ], + "type": "text", + "content": " are hyper-parameters. We empirically set " + }, + { + "bbox": [ + 46, + 247, + 286, + 283 + ], + "type": "inline_equation", + "content": "\\gamma_{1} = 0.2" + }, + { + "bbox": [ + 46, + 247, + 286, + 283 + ], + "type": "inline_equation", + "content": "\\gamma_{2} = 1.0" + }, + { + "bbox": [ + 46, + 247, + 286, + 283 + ], + "type": "inline_equation", + "content": "\\gamma_{3} = 1.5" + }, + { + "bbox": [ + 46, + 247, + 286, + 283 + ], + "type": "text", + "content": " for rigid objects and " + }, + { + "bbox": [ + 46, + 247, + 286, + 283 + ], + "type": "inline_equation", + "content": "\\gamma_{1} = 0.2" + }, + { + "bbox": [ + 46, + 247, + 286, + 283 + ], + "type": "inline_equation", + "content": "\\gamma_{2} = 10.0" + }, + { + "bbox": [ + 46, + 247, + 286, + 283 + ], + "type": "inline_equation", + "content": "\\gamma_{3} = 1.0" + }, + { + "bbox": [ + 46, + 247, + 286, + 283 + ], + "type": "text", + "content": " for non-rigid objects." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 292, + 127, + 305 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 292, + 127, + 305 + ], + "spans": [ + { + "bbox": [ + 46, + 292, + 127, + 305 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 311, + 107, + 323 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 311, + 107, + 323 + ], + "spans": [ + { + "bbox": [ + 47, + 311, + 107, + 323 + ], + "type": "text", + "content": "4.1. Settings" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 329, + 286, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 329, + 286, + 449 + ], + "spans": [ + { + "bbox": [ + 46, + 329, + 286, + 449 + ], + "type": "text", + "content": "Datasets. We compare CXTrack with previous state of the arts on three large-scale datasets: KITTI [5], nuScenes [2] and Waymo Open Dataset (WOD) [24]. KITTI contains 21 training video sequences and 29 test sequences. We follow previous work [6] and split the training sequences into three parts, 0-16 for training, 17-18 for validation and 19-20 for testing. For nuScenes, we use its validation split to evaluate our model, which contains 150 scenes. For WOD, we follow LiDAR-SOT [16] to evaluate our method, dividing it into three splits according to the sparsity of point clouds." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 449, + 286, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 449, + 286, + 533 + ], + "spans": [ + { + "bbox": [ + 46, + 449, + 286, + 533 + ], + "type": "text", + "content": "Implementation Details. We adopt DGCNN [28] as the backbone network to extract local geometric information. In the X-RPN, we initialize the scaling parameter " + }, + { + "bbox": [ + 46, + 449, + 286, + 533 + ], + "type": "inline_equation", + "content": "\\sigma^2 = 10" + }, + { + "bbox": [ + 46, + 449, + 286, + 533 + ], + "type": "text", + "content": ". Notably, we empirically fix " + }, + { + "bbox": [ + 46, + 449, + 286, + 533 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 46, + 449, + 286, + 533 + ], + "type": "text", + "content": " as a hyper-parameter for pedestrians and cyclists, and set it as a learnable parameter for cars and vans, since they may have larger motions. More details are provided in the supplementary material." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 533, + 286, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 533, + 286, + 639 + ], + "spans": [ + { + "bbox": [ + 46, + 533, + 286, + 639 + ], + "type": "text", + "content": "Evaluation Metrics. We use Success and Precision defined in one pass evaluation [10] as evaluation metrics. Success denotes the Area Under Curve (AUC) for the plot showing the ratio of frames where the Intersection Over Union (IOU) between the predicted and ground-truth bounding boxes is larger than a threshold, ranging from 0 to 1. Precision is defined as the AUC of the plot showing the ratio of frames where the distance between their centers is within a threshold, from 0 to 2 meters." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 647, + 230, + 659 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 647, + 230, + 659 + ], + "spans": [ + { + "bbox": [ + 47, + 647, + 230, + 659 + ], + "type": "text", + "content": "4.2. Comparison with State of the Arts" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 665, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 286, + 713 + ], + "type": "text", + "content": "We make comprehensive comparisons on the KITTI dataset with previous state-of-the-art methods, including SC3D [6], P2B [20], 3DSiamRPN [4], LTTR [3], MLVS-Net [29], BAT [31], PTT [22], V2B [8], PTTR [33]," + } + ] + } + ], + "index": 11 + }, + { + "type": "table", + "bbox": [ + 310, + 117, + 543, + 266 + ], + "blocks": [ + { + "bbox": [ + 305, + 71, + 545, + 114 + ], + "lines": [ + { + "bbox": [ + 305, + 71, + 545, + 114 + ], + "spans": [ + { + "bbox": [ + 305, + 71, + 545, + 114 + ], + "type": "text", + "content": "Table 1. Comparisons with the state-of-the-art methods on KITTI dataset. \"Mean\" is the average result weighted by frame numbers. \"Blue\" and \"Bold\" denote previous and current best performance, respectively. Success/Precision are used for evaluation." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 310, + 117, + 543, + 266 + ], + "lines": [ + { + "bbox": [ + 310, + 117, + 543, + 266 + ], + "spans": [ + { + "bbox": [ + 310, + 117, + 543, + 266 + ], + "type": "table", + "html": "
MethodCar (6424)Pedestrian (6088)Van (1248)Cyclist (308)Mean (14068)
SC3D41.3/57.918.2/37.840.4/47.041.5/70.431.2/48.5
P2B56.2/72.828.7/49.640.8/48.432.1/44.742.4/60.0
3DSiamRPN58.2/76.235.2/56.245.7/52.936.2/49.046.7/64.9
LTPR65.0/77.133.2/56.835.8/45.666.2/89.948.7/65.8
MLVSNet56.0/74.034.1/61.152.0/61.434.3/44.545.7/66.7
BAT60.5/77.742.1/70.152.4/67.033.7/45.451.2/72.8
PTT67.8/81.844.9/72.043.6/52.537.2/47.355.1/74.2
V2B70.5/81.348.3/73.550.1/58.040.8/49.758.4/75.2
PTTR65.2/77.450.9/81.652.5/61.865.1/90.557.9/78.1
STNet72.1/84.049.9/77.258.0/70.673.5/93.761.3/80.1
M2-Track65.5/80.861.5/88.253.8/70.773.2/93.562.9/83.4
CXTrack69.1/81.667.0/91.560.0/71.874.2/94.367.5/85.3
Improvement↓3.0/↓2.4↑5.5/↑3.3↑2.0/↑1.1↑0.7/↑0.6↑4.6/↑1.9
", + "image_path": "c4e090f2647fe6f303e79d7107991734181ce316d7e511b6544db5663944e157.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_body" + } + ], + "index": 13 + }, + { + "type": "table", + "bbox": [ + 328, + 300, + 525, + 350 + ], + "blocks": [ + { + "bbox": [ + 305, + 277, + 545, + 298 + ], + "lines": [ + { + "bbox": [ + 305, + 277, + 545, + 298 + ], + "spans": [ + { + "bbox": [ + 305, + 277, + 545, + 298 + ], + "type": "text", + "content": "Table 2. Robustness under scenes that contain intra-class distractors on KITTI Pedestrian category." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 328, + 300, + 525, + 350 + ], + "lines": [ + { + "bbox": [ + 328, + 300, + 525, + 350 + ], + "spans": [ + { + "bbox": [ + 328, + 300, + 525, + 350 + ], + "type": "table", + "html": "
MethodAll(6088)Distractor-Only(3917)Improvement
PTTR50.9/81.644.3/70.0↓6.6/↓11.6
STNet49.9/77.235.1/58.5↓14.8/↓18.7
M2-Track61.5/88.258.0/88.4↓3.5/↑0.2
CXTrack67.0/91.566.1/91.3↓0.9/↓0.3
", + "image_path": "6faee4123ed18230c85f19925750a624d6b630c55a423bd4856637e9d1cf8937.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "table_body" + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 364, + 545, + 663 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 364, + 545, + 663 + ], + "spans": [ + { + "bbox": [ + 304, + 364, + 545, + 663 + ], + "type": "text", + "content": "STNet [9] and M2-Track [32]. As illustrated in Tab. 1, CXTrack surpasses previous state-of-the-art methods, with a significant improvement of average Success and Precision. Notably, our method achieves the best performance under all categories, except for the Car, where voxel-based STNet [9] surpasses us by a minor margin (72.1/84.0 v.s. 69.1/81.6). Most vehicles have simple shapes and limited rotation angles, which fit well in voxels. We argue that voxelization provides a strong shape prior, thereby leading to performance gain for large objects with simple shapes. The lack of distractors for cars also makes our improvement over previous methods insignificant. However, our method has a significant improvement (67.0/91.5 v.s. 49.9/77.2) on the Pedestrian category. We claim that this stems from our special design to handle distractors and our better preservation for contextual information. Besides, compared with M2-Track [32], CXTrack obtains consistent performance gains on all categories especially on the Success metric, which demonstrates the importance of local geometry and contextual information. For further analysis on the impact of intraclass distractors, we manually pick out scenes that contain Pedestrian distractors from the KITTI test split and then evaluate different methods on these scenes. As shown in Tab. 2, both M2-Track and CXTrack are robust to distractors, while CXTrack can make more accurate predictions." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 665, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 545, + 712 + ], + "type": "text", + "content": "To verify the generalization ability of our method, we follow previous methods [8, 9] and test the KITTI pretrained model on nuScenes and WOD. The comparison results on WOD are shown in Tab. 3. It can be seen that our" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1089" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 62, + 82, + 531, + 168 + ], + "blocks": [ + { + "bbox": [ + 165, + 71, + 428, + 81 + ], + "lines": [ + { + "bbox": [ + 165, + 71, + 428, + 81 + ], + "spans": [ + { + "bbox": [ + 165, + 71, + 428, + 81 + ], + "type": "text", + "content": "Table 3. Comparison with state of the arts on Waymo Open Dataset." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 62, + 82, + 531, + 168 + ], + "lines": [ + { + "bbox": [ + 62, + 82, + 531, + 168 + ], + "spans": [ + { + "bbox": [ + 62, + 82, + 531, + 168 + ], + "type": "table", + "html": "
MethodVehicle(185731)Pedestrian(241752)Mean(427483)
EasyMediumHardMeanEasyMediumHardMean
P2B57.1/65.452.0/60.747.9/58.552.6/61.718.1/30.817.8/30.017.7/29.317.9/30.133.0/43.8
BAT61.0/68.353.3/60.948.9/57.854.7/62.719.3/32.617.8/29.817.2/28.318.2/30.334.1/44.4
V2B64.5/71.555.1/63.252.0/62.057.6/65.927.9/43.922.5/36.220.1/33.123.7/37.938.4/50.1
STNet65.9/72.757.5/66.054.6/64.759.7/68.029.2/45.324.7/38.222.2/35.825.5/39.940.4/52.1
CXTrack63.9/71.154.2/62.752.1/63.757.1/66.135.4/55.329.7/47.926.3/44.430.7/49.442.2/56.7
Improvement↓2.0/↓1.6↓3.3/↓3.3↓3.5/↓1.0↓2.6/↓1.9↑6.2/↑10.0↑5.0/↑9.7↑4.1/↑8.6↑5.2/↑9.5↑1.8/↑4.6
", + "image_path": "e73309839e5472be3298e0f456e5823fe24c8f2dc9affa05eb572ae7084cffd0.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 52, + 186, + 284, + 278 + ], + "blocks": [ + { + "bbox": [ + 60, + 175, + 274, + 185 + ], + "lines": [ + { + "bbox": [ + 60, + 175, + 274, + 185 + ], + "spans": [ + { + "bbox": [ + 60, + 175, + 274, + 185 + ], + "type": "text", + "content": "Table 4. Comparison with state of the arts on nuScenes." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 52, + 186, + 284, + 278 + ], + "lines": [ + { + "bbox": [ + 52, + 186, + 284, + 278 + ], + "spans": [ + { + "bbox": [ + 52, + 186, + 284, + 278 + ], + "type": "table", + "html": "
MethodCar (15578)Pedestrian (8019)Van (3710)Cyclist (501)Mean (27808)
SC3D25.0/27.114.2/16.225.7/21.917.0/18.221.8/23.1
P2B27.0/29.215.9/22.021.5/16.220.0/26.422.9/25.3
BAT22.5/24.117.3/24.519.3/15.817.0/18.820.5/23.0
V2B31.3/35.117.3/23.421.7/16.722.2/19.125.8/29.0
STNet32.2/36.119.1/27.222.3/16.821.2/29.226.9/30.8
CXTrack29.6/33.420.4/32.927.6/20.818.5/26.826.5/31.5
Improvement↓2.6/↓2.7↑1.3/↑5.7↑1.9/↓1.1↓3.7/↓2.4↓0.4/↑0.7
", + "image_path": "3b98402537dbb5dc1aecf6cdc84181d292b8036ac0e42164c365f7887d7da724.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 74, + 299, + 264, + 358 + ], + "blocks": [ + { + "bbox": [ + 66, + 287, + 268, + 297 + ], + "lines": [ + { + "bbox": [ + 66, + 287, + 268, + 297 + ], + "spans": [ + { + "bbox": [ + 66, + 287, + 268, + 297 + ], + "type": "text", + "content": "Table 5. Efficiency analysis of different components." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 74, + 299, + 264, + 358 + ], + "lines": [ + { + "bbox": [ + 74, + 299, + 264, + 358 + ], + "spans": [ + { + "bbox": [ + 74, + 299, + 264, + 358 + ], + "type": "table", + "html": "
ComponentFLOPs#ParamsInfer Speed
backbone3.18G1.3M8.5ms
transformer1.28G14.7M10.9ms
X-RPN0.17G2.3M3.0ms
pre/postprocess--6.8ms
CXTrack4.63G18.3M29.2ms(34FPS)
", + "image_path": "563202c21b5f3313f0759e028a694c8ec496c6100ddf5e3862be36e6b56b51ad.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 371, + 287, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 371, + 287, + 479 + ], + "spans": [ + { + "bbox": [ + 46, + 371, + 287, + 479 + ], + "type": "text", + "content": "method outperforms others in terms of average Success and Precision with a clear margin. Notably, KITTI and WOD data are captured by 64-beam LiDARs, while nuScenes data are captured by 32-beam LiDARs. Thus it is more challenging to generalize the pretrained model on the nuScenes dataset. As shown in Tab. 4, our method achieves comparable performance on the nuScenes dataset. In short, CXTrack not only achieves a good balance between small objects and large objects, but also generalizes well to unseen scenes." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 479, + 287, + 598 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 479, + 287, + 598 + ], + "spans": [ + { + "bbox": [ + 46, + 479, + 287, + 598 + ], + "type": "text", + "content": "We also visualize the tracking results for qualitative comparisons. As shown in Fig. 5, CXTrack achieves good accuracy in scenes with both sparse and dense point clouds on both categories. In the sparse cases (left), previous methods drift towards intra-class distractors due to large appearance variations caused by occlusions, while only our method keeps track of the target, thanks to the sufficient use of contextual information. In the dense cases (right), our method can track the target more accurately than M2-Track by leveraging local geometric information." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 600, + 287, + 661 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 600, + 287, + 661 + ], + "spans": [ + { + "bbox": [ + 46, + 600, + 287, + 661 + ], + "type": "text", + "content": "We report the efficiency of different components in Tab. 5. It can be observed that the target-centric transformer is the bottleneck of CXTrack during inference. We can replace the vanilla self-attention in CXTrack with linear attention such as linformer [27] for further speedup." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 670, + 148, + 681 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 670, + 148, + 681 + ], + "spans": [ + { + "bbox": [ + 47, + 670, + 148, + 681 + ], + "type": "text", + "content": "4.3. Ablation Studies" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 689, + 287, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 287, + 712 + ], + "type": "text", + "content": "To validate the effectiveness of several design choices in CXTrack, we conduct ablation studies on the KITTI dataset." + } + ] + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 308, + 222, + 544, + 269 + ], + "blocks": [ + { + "bbox": [ + 305, + 175, + 545, + 218 + ], + "lines": [ + { + "bbox": [ + 305, + 175, + 545, + 218 + ], + "spans": [ + { + "bbox": [ + 305, + 175, + 545, + 218 + ], + "type": "text", + "content": "Table 6. Ablation studies of different components of the target-centric transformer. \"Cx\" refers to contextual information, \"M\" refers to the cascaded targetness mask prediction and \"C\" refers to the auxiliary target center regression branches." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 308, + 222, + 544, + 269 + ], + "lines": [ + { + "bbox": [ + 308, + 222, + 544, + 269 + ], + "spans": [ + { + "bbox": [ + 308, + 222, + 544, + 269 + ], + "type": "table", + "html": "
CxMCCarPedestrianVanCyclistMean
✓62.5/74.260.6/87.058.3/71.472.0/93.361.5/79.9
✓✓67.4/80.263.9/89.057.8/70.872.7/93.865.1/83.5
✓✓59.7/73.6†51.8/81.659.9/71.571.7/93.256.6/77.3
✓✓✓69.1/81.667.0/91.560.0/71.874.2/94.367.5/85.3
", + "image_path": "7ced26cf11fc6659f499ceb74280347fde288ea4bc6850c51f00a4003f4de29c.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 381, + 269, + 471, + 277 + ], + "lines": [ + { + "bbox": [ + 381, + 269, + 471, + 277 + ], + "spans": [ + { + "bbox": [ + 381, + 269, + 471, + 277 + ], + "type": "text", + "content": "†: unstable training process" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 12 + }, + { + "type": "table", + "bbox": [ + 308, + 330, + 544, + 374 + ], + "blocks": [ + { + "bbox": [ + 305, + 281, + 545, + 326 + ], + "lines": [ + { + "bbox": [ + 305, + 281, + 545, + 326 + ], + "spans": [ + { + "bbox": [ + 305, + 281, + 545, + 326 + ], + "type": "text", + "content": "Table 7. Ablation studies of different transformer layers on KITTI. \"V\" refers to the vanilla transformer layer and \"G\" refers to the gated transformer layer. \"S\" represents the semi-dropout transformer layer which is adopted in our proposed CXTrack." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 308, + 330, + 544, + 374 + ], + "lines": [ + { + "bbox": [ + 308, + 330, + 544, + 374 + ], + "spans": [ + { + "bbox": [ + 308, + 330, + 544, + 374 + ], + "type": "table", + "html": "
CarPedestrianVanCyclistMean
V68.8/80.462.9/87.857.2/69.672.7/94.265.3/82.9
G64.8/76.964.7/91.156.2/70.570.6/93.464.1/82.8
S69.1/81.667.0/91.560.0/71.874.2/94.367.5/85.3
", + "image_path": "7786ec22a88631829f9abe3d5a1b393a12b911df0596ceae13701d252064ea8b.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "table_body" + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 390, + 545, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 390, + 545, + 593 + ], + "spans": [ + { + "bbox": [ + 304, + 390, + 545, + 593 + ], + "type": "text", + "content": "Components of Target-Centric Transformer. Tab. 6 presents ablation studies of different components of transformer to gain a better understanding of its designs. We crop the input point cloud " + }, + { + "bbox": [ + 304, + 390, + 545, + 593 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_{t-1}" + }, + { + "bbox": [ + 304, + 390, + 545, + 593 + ], + "type": "text", + "content": " using " + }, + { + "bbox": [ + 304, + 390, + 545, + 593 + ], + "type": "inline_equation", + "content": "\\mathcal{B}_{t-1}" + }, + { + "bbox": [ + 304, + 390, + 545, + 593 + ], + "type": "text", + "content": " to ablate contextual information in the previous frame. We can observe significant performance drop when not using contextual information, especially on Car and Pedestrian. For Car, it suffers from heavy occlusions (Fig. 5), while pedestrian distractors are widespread in the scene. We also find that removing context leads to unstable training on Car. We presume that the lack of supervised signals to tell the model what not to attend may confuse the model and introduce noise in training. For the cascaded targetness mask prediction and auxiliary target center regression, removing either of them leads to an obvious decline on terms of average metrics. We argue that the auxiliary regression loss can increase the feature similarities of points belonging to the same object." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "text", + "content": "Target-Centric Transformer Layer. Tab. 7 shows the impact of different target-centric transformer layers. Semi-Dropout achieves better performance than Vanilla, especially on Pedestrian. Small objects often consist of fewer points, hence applying dropout directly on the targetness information in training may confuse the network and lead to sub-optimal results. Gated relies entirely on the predicted targetness mask to modulate the amount of exposure for input features, which may suffer from information loss when the targetness mask is not accurate enough." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1090" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 51, + 71, + 299, + 220 + ], + "blocks": [ + { + "bbox": [ + 51, + 71, + 299, + 220 + ], + "lines": [ + { + "bbox": [ + 51, + 71, + 299, + 220 + ], + "spans": [ + { + "bbox": [ + 51, + 71, + 299, + 220 + ], + "type": "image", + "image_path": "8a18fb599e340f05b4de5b2f3d9521f4b98e9a2ff6806d2c7b64df51f9321571.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 129, + 221, + 463, + 233 + ], + "lines": [ + { + "bbox": [ + 129, + 221, + 463, + 233 + ], + "spans": [ + { + "bbox": [ + 129, + 221, + 463, + 233 + ], + "type": "text", + "content": "Figure 5. Visualization results. Left: Sparse cases in KITTI. Right: Dense cases in KITTI." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 301, + 72, + 541, + 220 + ], + "blocks": [ + { + "bbox": [ + 301, + 72, + 541, + 220 + ], + "lines": [ + { + "bbox": [ + 301, + 72, + 541, + 220 + ], + "spans": [ + { + "bbox": [ + 301, + 72, + 541, + 220 + ], + "type": "image", + "image_path": "134c9d88ef8e452ba93f35a66f819d6f2db382d8f0404d66f1235534f795f4bc.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 49, + 272, + 286, + 331 + ], + "blocks": [ + { + "bbox": [ + 47, + 238, + 288, + 270 + ], + "lines": [ + { + "bbox": [ + 47, + 238, + 288, + 270 + ], + "spans": [ + { + "bbox": [ + 47, + 238, + 288, + 270 + ], + "type": "text", + "content": "Table 8. Ablation studies of various localization heads on KITTI. \"X-RPN\\C\" indicates our proposed localization head X-RPN without center embedding." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 49, + 272, + 286, + 331 + ], + "lines": [ + { + "bbox": [ + 49, + 272, + 286, + 331 + ], + "spans": [ + { + "bbox": [ + 49, + 272, + 286, + 331 + ], + "type": "table", + "html": "
HeadCarPedestrianVanCyclistMean
PRM [33]66.5/77.462.2/86.852.9/64.972.5/93.863.6/80.7
RPN [18]64.1/76.959.8/88.355.0/65.668.2/92.461.5/81.2
V2B [8]70.5/82.660.1/86.758.0/69.870.5/93.364.9/83.5
X-RPN\\C67.8/80.365.5/89.559.9/72.172.6/94.166.2/83.9
X-RPN69.1/81.667.0/91.560.0/71.874.2/94.367.5/85.3
", + "image_path": "5764e63c17277e8db8af717426f089092c5df9c2354ca1754b1190aca3d72f39.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 48, + 339, + 161, + 415 + ], + "blocks": [ + { + "bbox": [ + 48, + 339, + 161, + 415 + ], + "lines": [ + { + "bbox": [ + 48, + 339, + 161, + 415 + ], + "spans": [ + { + "bbox": [ + 48, + 339, + 161, + 415 + ], + "type": "image", + "image_path": "ad227c0d17dd9257df9a43ab0d9999566ccb1b92022a95c44183282c8c4dcd90.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 47, + 415, + 287, + 449 + ], + "lines": [ + { + "bbox": [ + 47, + 415, + 287, + 449 + ], + "spans": [ + { + "bbox": [ + 47, + 415, + 287, + 449 + ], + "type": "text", + "content": "Figure 6. Representative examples of attention maps in the transformer. Target-centric transformer attends to objects that have similar geometry." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 163, + 339, + 286, + 415 + ], + "blocks": [ + { + "bbox": [ + 163, + 339, + 286, + 415 + ], + "lines": [ + { + "bbox": [ + 163, + 339, + 286, + 415 + ], + "spans": [ + { + "bbox": [ + 163, + 339, + 286, + 415 + ], + "type": "image", + "image_path": "1b440cac86f7e678c254783772b24d356a32f8bc26dc4f5f790924c4de3ffd14.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 48, + 457, + 287, + 536 + ], + "blocks": [ + { + "bbox": [ + 48, + 457, + 287, + 536 + ], + "lines": [ + { + "bbox": [ + 48, + 457, + 287, + 536 + ], + "spans": [ + { + "bbox": [ + 48, + 457, + 287, + 536 + ], + "type": "image", + "image_path": "3da16ceb93d9623f917fbcd52dab6e3372d8d696310ef5ca5971bdf92a49e19d.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 47, + 537, + 287, + 560 + ], + "lines": [ + { + "bbox": [ + 47, + 537, + 287, + 560 + ], + "spans": [ + { + "bbox": [ + 47, + 537, + 287, + 560 + ], + "type": "text", + "content": "Figure 7. Visualization of ablation study. Center embedding can benefit object tracking in challenging scenes with distractors." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 570, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 570, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 570, + 287, + 713 + ], + "type": "text", + "content": "X-RPN. We replace X-RPN with other alternatives [8, 20, 33] and report the comparison results in Tab. 8. Although the V2B head achieves better performance than X-RPN on the Car category, it fails to track small objects such as pedestrians effectively due to intra-class distractors and inevitable information loss brought in by voxelization. It is also worth noting that we observe a performance drop without center embedding, especially on the Pedestrian category, for which distractors are more commonly seen. To explore the effectiveness of the center embedding, we visualize the attention map of the last transformer layer in Fig. 6. We observe that the transformer alone can attend to regions with similar ge" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 239, + 545, + 309 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 239, + 545, + 309 + ], + "spans": [ + { + "bbox": [ + 304, + 239, + 545, + 309 + ], + "type": "text", + "content": "ometry to the target, but fails to distinguish the target from distractors. As shown in Fig. 7, with the help of the center embedding, the network precisely keeps track of the target. In short, X-RPN achieves a good balance between large and small objects, and effectively alleviates the distractor problem." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 318, + 391, + 330 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 318, + 391, + 330 + ], + "spans": [ + { + "bbox": [ + 306, + 318, + 391, + 330 + ], + "type": "text", + "content": "4.4. Failure Cases" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 336, + 546, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 336, + 546, + 445 + ], + "spans": [ + { + "bbox": [ + 304, + 336, + 546, + 445 + ], + "type": "text", + "content": "Although CXTrack is robust to intra-class distractors, it fails to predict accurate orientation of the target when the point clouds are too sparse to capture informative local geometry or when large appearance variations occur, as shown in Fig. 7. Besides, the center embedding directly encodes the displacement of target center into features, so our model may suffer from performance degradation if trained with " + }, + { + "bbox": [ + 304, + 336, + 546, + 445 + ], + "type": "inline_equation", + "content": "2\\mathrm{Hz}" + }, + { + "bbox": [ + 304, + 336, + 546, + 445 + ], + "type": "text", + "content": " data and tested with " + }, + { + "bbox": [ + 304, + 336, + 546, + 445 + ], + "type": "inline_equation", + "content": "10\\mathrm{Hz}" + }, + { + "bbox": [ + 304, + 336, + 546, + 445 + ], + "type": "text", + "content": " data because the scale of the displacement differs significantly." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 454, + 378, + 466 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 454, + 378, + 466 + ], + "spans": [ + { + "bbox": [ + 306, + 454, + 378, + 466 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 474, + 545, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 474, + 545, + 653 + ], + "spans": [ + { + "bbox": [ + 304, + 474, + 545, + 653 + ], + "type": "text", + "content": "We revisit existing paradigms for the 3D SOT task and propose a new paradigm to fully exploit contextual information across frames, which is largely overlooked by previous methods. Following this paradigm, we design a novel transformer-based network named CXTrack, which employs a target-centric transformer to explore contextual information and model intra-frame and inter-frame feature relationships. We also introduce a localization head named X-RPN to obtain high-quality proposals for objects of all sizes, as well as a center embedding module to distinguish the target from distractors. Extensive experiments show that CXTrack significantly outperforms previous state-of-the-arts, and is robust to distractors. We hope our work can promote further exploitations in this task by showing the necessity to explore contextual information for more robust predictions." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "content": "Acknowledgment The authors thank Jiahui Huang for his discussions. This work was supported by the Natural Science Foundation of China (Project Number 61832016), and Tsinghua-Tencent Joint Laboratory for Internet Innovation Technology." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 314, + 757 + ], + "type": "text", + "content": "1091" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 123 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 123 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 123 + ], + "type": "text", + "content": "[1] Jimmy Lei Ba, Jamie Ryan Kiros, and Geoffrey E Hinton. Layer normalization. arXiv preprint arXiv:1607.06450, 2016.3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 126, + 288, + 192 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 126, + 288, + 192 + ], + "spans": [ + { + "bbox": [ + 53, + 126, + 288, + 192 + ], + "type": "text", + "content": "[2] Holger Caesar, Varun Bankiti, Alex H Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuScenes: A multimodal dataset for autonomous driving. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11621-11631, 2020. 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 194, + 288, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 194, + 288, + 226 + ], + "spans": [ + { + "bbox": [ + 53, + 194, + 288, + 226 + ], + "type": "text", + "content": "[3] Yubo Cui, Zheng Fang, Jiayao Shan, Zuoxu Gu, and Sifan Zhou. 3D object tracking with transformer. arXiv preprint arXiv:2110.14921, 2021. 2, 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 228, + 288, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 228, + 288, + 271 + ], + "spans": [ + { + "bbox": [ + 53, + 228, + 288, + 271 + ], + "type": "text", + "content": "[4] Zheng Fang, Sifan Zhou, Yubo Cui, and Sebastian Scherer. 3D-SiamRPN: An end-to-end learning method for real-time 3D single object tracking using raw point cloud. IEEE Sensors Journal, 21(4):4995-5011, 2020. 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 274, + 288, + 317 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 274, + 288, + 317 + ], + "spans": [ + { + "bbox": [ + 53, + 274, + 288, + 317 + ], + "type": "text", + "content": "[5] Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we ready for autonomous driving? the KITTI vision benchmark suite. In IEEE Conference on Computer Vision and Pattern Recognition, pages 3354-3361. IEEE, 2012. 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 319, + 288, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 319, + 288, + 363 + ], + "spans": [ + { + "bbox": [ + 53, + 319, + 288, + 363 + ], + "type": "text", + "content": "[6] Silvio Giancola, Jesus Zarzar, and Bernard Ghanem. Leveraging shape completion for 3D Siamese tracking. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1359-1368, 2019. 1, 2, 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 364, + 288, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 364, + 288, + 418 + ], + "spans": [ + { + "bbox": [ + 53, + 364, + 288, + 418 + ], + "type": "text", + "content": "[7] Chenhang He, Ruihuang Li, Shuai Li, and Lei Zhang. Voxel set transformer: A set-to-set approach to 3D object detection from point clouds. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8417-8427, 2022. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 420, + 288, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 420, + 288, + 464 + ], + "spans": [ + { + "bbox": [ + 53, + 420, + 288, + 464 + ], + "type": "text", + "content": "[8] Le Hui, Lingpeng Wang, Mingmei Cheng, Jin Xie, and Jian Yang. 3D Siamese voxel-to-BEV tracker for sparse point clouds. Advances in Neural Information Processing Systems, 34:28714-28727, 2021. 2, 5, 6, 8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 466, + 288, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 466, + 288, + 520 + ], + "spans": [ + { + "bbox": [ + 53, + 466, + 288, + 520 + ], + "type": "text", + "content": "[9] Le Hui, Lingpeng Wang, Linghua Tang, Kaihao Lan, Jin Xie, and Jian Yang. 3D Siamese transformer network for single object tracking on point clouds. In Proceedings of the European Conference on Computer Vision, Part II, pages 293-310. Springer, 2022. 2, 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 522, + 287, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 522, + 287, + 586 + ], + "spans": [ + { + "bbox": [ + 48, + 522, + 287, + 586 + ], + "type": "text", + "content": "[10] Matej Kristan, Jiri Matas, Aleš Leonardis, Tomáš Vojíř, Roman Pflugfelder, Gustavo Fernández, Georg Nebehay, Fatih Porikli, and Luka Čehovin. A novel performance evaluation methodology for single-target trackers. IEEE Transactions on Pattern Analysis and Machine Intelligence, 38(11):2137-2155, 2016. 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 590, + 287, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 590, + 287, + 622 + ], + "spans": [ + { + "bbox": [ + 48, + 590, + 287, + 622 + ], + "type": "text", + "content": "[11] H Kuang Chiu, A Prioletti, J Li, and J Bohg. Probabilistic 3D multi-object tracking for autonomous driving. ArXiv, vol. abs/2001.05673, 2020. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 624, + 287, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 624, + 287, + 667 + ], + "spans": [ + { + "bbox": [ + 48, + 624, + 287, + 667 + ], + "type": "text", + "content": "[12] Ze Liu, Zheng Zhang, Yue Cao, Han Hu, and Xin Tong. Group-free 3D object detection via transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2949-2958, 2021. 1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 670, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 670, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 670, + 287, + 713 + ], + "type": "text", + "content": "[13] Matthias Luber, Luciano Spinello, and Kai O Arras. People tracking in RGB-D data with on-line boosted target models. In 2011 IEEE/RSJ International Conference on Intelligent Robots and Systems, pages 3844-3849. IEEE, 2011. 2" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 128 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 128 + ], + "type": "text", + "content": "[14] Jiageng Mao, Yujing Xue, Minzhe Niu, Haoyue Bai, Jiashi Feng, Xiaodan Liang, Hang Xu, and Chunjing Xu. Voxel transformer for 3D object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3164-3173, 2021. 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 129, + 545, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 129, + 545, + 172 + ], + "spans": [ + { + "bbox": [ + 307, + 129, + 545, + 172 + ], + "type": "text", + "content": "[15] Ishan Misra, Rohit Girdhar, and Armand Joulin. An end-to-end transformer model for 3D object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2906-2917, 2021. 1, 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 173, + 545, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 173, + 545, + 216 + ], + "spans": [ + { + "bbox": [ + 307, + 173, + 545, + 216 + ], + "type": "text", + "content": "[16] Ziqi Pang, Zhichao Li, and Naiyan Wang. Model-free vehicle tracking and state estimation in point cloud sequences. In IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 8075-8082, 2021. 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 217, + 545, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 217, + 545, + 270 + ], + "spans": [ + { + "bbox": [ + 307, + 217, + 545, + 270 + ], + "type": "text", + "content": "[17] Alessandro Pieropan, Niklas Bergstrom, Masatoshi Ishikawa, and Hedvig Kjellstrom. Robust 3D tracking of unknown objects. In IEEE International Conference on Robotics and Automation (ICRA), pages 2410-2417. IEEE, 2015. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 272, + 545, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 272, + 545, + 326 + ], + "spans": [ + { + "bbox": [ + 307, + 272, + 545, + 326 + ], + "type": "text", + "content": "[18] Charles R Qi, Or Litany, Kaiming He, and Leonidas J Guibas. Deep Hough voting for 3D object detection in point clouds. In proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9277-9286, 2019. 1, 2, 5, 8" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 327, + 545, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 327, + 545, + 380 + ], + "spans": [ + { + "bbox": [ + 307, + 327, + 545, + 380 + ], + "type": "text", + "content": "[19] Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. PointNet: Deep learning on point sets for 3D classification and segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 652-660, 2017. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 382, + 545, + 436 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 382, + 545, + 436 + ], + "spans": [ + { + "bbox": [ + 307, + 382, + 545, + 436 + ], + "type": "text", + "content": "[20] Haozhe Qi, Chen Feng, Zhiguo Cao, Feng Zhao, and Yang Xiao. P2B: Point-to-box network for 3D object tracking in point clouds. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6329-6338, 2020. 1, 2, 5, 6, 8" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 437, + 545, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 437, + 545, + 480 + ], + "spans": [ + { + "bbox": [ + 307, + 437, + 545, + 480 + ], + "type": "text", + "content": "[21] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster R-CNN: Towards real-time object detection with region proposal networks. Advances in Neural Information Processing Systems, 28, 2015. 5" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 482, + 545, + 535 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 482, + 545, + 535 + ], + "spans": [ + { + "bbox": [ + 307, + 482, + 545, + 535 + ], + "type": "text", + "content": "[22] Jiayao Shan, Sifan Zhou, Zheng Fang, and Yubo Cui. PTT: Point-track-transformer module for 3D single object tracking in point clouds. In IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 1310-1316. IEEE, 2021. 2, 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 536, + 545, + 581 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 536, + 545, + 581 + ], + "spans": [ + { + "bbox": [ + 307, + 536, + 545, + 581 + ], + "type": "text", + "content": "[23] Luciano Spinello, Kai Arras, Rudolph Triebel, and Roland Siegwart. A layered approach to people detection in 3D range data. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 24, pages 1625-1630, 2010. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 582, + 545, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 582, + 545, + 646 + ], + "spans": [ + { + "bbox": [ + 307, + 582, + 545, + 646 + ], + "type": "text", + "content": "[24] Pei Sun, Henrik Kretzschmar, Xerxes Dotiwalla, Aurelien Chouard, Vijaysai Patnaik, Paul Tsui, James Guo, Yin Zhou, Yuning Chai, Benjamin Caine, et al. Scalability in perception for autonomous driving: Waymo open dataset. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2446-2454, 2020. 6" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 647, + 545, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 647, + 545, + 690 + ], + "spans": [ + { + "bbox": [ + 307, + 647, + 545, + 690 + ], + "type": "text", + "content": "[25] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in Neural Information Processing Systems, 30, 2017. 2, 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 692, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 692, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 692, + 545, + 713 + ], + "type": "text", + "content": "[26] Ning Wang, Wengang Zhou, Jie Wang, and Houqiang Li. Transformer meets tracker: Exploiting temporal context for" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1092" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 474 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 66, + 72, + 287, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 72, + 287, + 105 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 287, + 105 + ], + "type": "text", + "content": "robust visual tracking. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1571-1580, 2021. 5" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 107, + 287, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 107, + 287, + 140 + ], + "spans": [ + { + "bbox": [ + 48, + 107, + 287, + 140 + ], + "type": "text", + "content": "[27] Sinong Wang, Belinda Z Li, Madian Khabsa, Han Fang, and Hao Ma. Linformer: Self-attention with linear complexity. arXiv preprint arXiv:2006.04768, 2020. 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 141, + 287, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 141, + 287, + 184 + ], + "spans": [ + { + "bbox": [ + 48, + 141, + 287, + 184 + ], + "type": "text", + "content": "[28] Yue Wang, Yongbin Sun, Ziwei Liu, Sanjay E Sarma, Michael M Bronstein, and Justin M Solomon. Dynamic graph CNN for learning on point clouds. ACM Transactions on Graphics (TOG), 38(5):1-12, 2019. 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 186, + 287, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 186, + 287, + 239 + ], + "spans": [ + { + "bbox": [ + 48, + 186, + 287, + 239 + ], + "type": "text", + "content": "[29] Zhoutao Wang, Qian Xie, Yu-Kun Lai, Jing Wu, Kun Long, and Jun Wang. MLVSNet: Multi-level voting Siamese network for 3D visual tracking. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3101-3110, 2021. 1, 2, 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 241, + 287, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 241, + 287, + 285 + ], + "spans": [ + { + "bbox": [ + 48, + 241, + 287, + 285 + ], + "type": "text", + "content": "[30] Tianwei Yin, Xingyi Zhou, and Philipp Krahenbuhl. Center-based 3D object detection and tracking. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11784-11793, 2021. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 286, + 287, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 286, + 287, + 341 + ], + "spans": [ + { + "bbox": [ + 48, + 286, + 287, + 341 + ], + "type": "text", + "content": "[31] Chaoda Zheng, Xu Yan, Jiantao Gao, Weibing Zhao, Wei Zhang, Zhen Li, and Shuguang Cui. Box-aware feature enhancement for single object tracking on point clouds. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13199-13208, 2021. 1, 2, 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 342, + 287, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 342, + 287, + 407 + ], + "spans": [ + { + "bbox": [ + 48, + 342, + 287, + 407 + ], + "type": "text", + "content": "[32] Chaoda Zheng, Xu Yan, Haiming Zhang, Baoyuan Wang, Shenghui Cheng, Shuguang Cui, and Zhen Li. Beyond 3D Siamese tracking: A motion-centric paradigm for 3D single object tracking in point clouds. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8111-8120, 2022. 1, 2, 3, 5, 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 409, + 287, + 474 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 409, + 287, + 474 + ], + "spans": [ + { + "bbox": [ + 48, + 409, + 287, + 474 + ], + "type": "text", + "content": "[33] Changqing Zhou, Zhipeng Luo, Yueru Luo, Tianrui Liu, Liang Pan, Zhongang Cai, Haiyu Zhao, and Shijian Lu. PTTR: Relational 3D point cloud object tracking with transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8531-8540, 2022. 1, 2, 6, 8" + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1093" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/CaPriDe Learning_ Confidential and Private Decentralized Learning Based on Encryption-Friendly Distillation Loss/ca732fe2-4161-4d55-be24-51fc46dde867_content_list.json b/2023/CaPriDe Learning_ Confidential and Private Decentralized Learning Based on Encryption-Friendly Distillation Loss/ca732fe2-4161-4d55-be24-51fc46dde867_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..2861f9ab8d744839b17ad7d6e09617e05a5f8772 --- /dev/null +++ b/2023/CaPriDe Learning_ Confidential and Private Decentralized Learning Based on Encryption-Friendly Distillation Loss/ca732fe2-4161-4d55-be24-51fc46dde867_content_list.json @@ -0,0 +1,1457 @@ +[ + { + "type": "text", + "text": "CaPriDe Learning: Confidential and Private Decentralized Learning based on Encryption-friendly Distillation Loss", + "text_level": 1, + "bbox": [ + 91, + 128, + 879, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Nurbek Tastan Karthik Nandakumar \nMohamed bin Zayed University of Artificial Intelligence, UAE \n{nurbek.tastan,karthik.nandakumar}@mbzuai.ac.ae", + "bbox": [ + 235, + 203, + 733, + 256 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 291, + 313, + 306 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Large volumes of data required to train accurate deep neural networks (DNNs) are seldom available with any single entity. Often, privacy concerns prevent entities from sharing data with each other or with a third-party learning service provider. While cross-silo federated learning (FL) allows collaborative learning of large DNNs without sharing the data itself, most existing cross-silo FL algorithms have an unacceptable utility-privacy trade-off. In this work, we propose a framework called Confidential and Private Decentralized (CaPriDe) learning, which optimally leverages the power of fully homomorphic encryption (FHE) to enable collaborative learning without compromising on the confidentiality and privacy of data. In CaPriDe learning, participating entities release their private data in an encrypted form allowing other participants to perform inference in the encrypted domain. The crux of CaPriDe learning is mutual knowledge distillation between multiple local models through a novel distillation loss, which is an approximation of the Kullback-Leibler (KL) divergence between the local predictions and encrypted inferences of other participants on the same data that can be computed in the encrypted domain. Extensive experiments on three datasets show that CaPriDe learning can improve the accuracy of local models without any central coordination, provide strong guarantees of data confidentiality and privacy, and has the ability to handle statistical heterogeneity. Constraints on the model architecture (arising from the need to be FHE-friendly), limited scalability, and computational complexity of encrypted domain inference are the main limitations of the proposed approach. The code can be found at https://github.com/tnurbek/capride-learning.", + "bbox": [ + 75, + 325, + 472, + 792 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 828, + 209, + 843 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Rapid strides have been made in machine learning (ML) (in particular, deep learning) over the past decade. However, in many important application domains such as health-", + "bbox": [ + 75, + 854, + 468, + 902 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "care and finance, the absence of large, centralized datasets is a significant obstacle to realizing the full benefits of deep learning algorithms. Data in these applications often resides in silos and is governed by strict regulations (e.g., HIPAA, GDPR, etc.) because of its privacy sensitive nature [22]. Competing business interests of data owners and the lack of appropriate incentives for data sharing further accentuate the problem. To overcome these issues, there is a need for collaborative learning algorithms that ideally satisfy the following requirements [15]: (i) confidentiality - no sharing of raw data, (ii) privacy - minimal leakage of information via the knowledge exchange mechanism (e.g., gradients or predictions), (iii) utility - gain in accuracy (over the individual models) resulting from the collaboration, even in the presence of statistical heterogeneity, (iv) efficiency - minimize computational complexity and communication burden, (v) robustness - handle unintentional failures and attacks emanating from malicious entities, and (vi) fairness - utility should be proportional to the individual contributions.", + "bbox": [ + 496, + 292, + 893, + 580 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Federated learning (FL) [15] is a special case of collaborative learning, which works under the orchestration of a central server. FL allows multiple entities to collaboratively solve a ML problem by sharing of focused updates (e.g., gradients), instead of raw data. Specifically, cross-silo FL (typically between 2-100 participants) has been touted as a promising solution to address the data fragmentation problem in finance [8] and healthcare [16,23]. Most prior cross-silo FL algorithms assume that all the parties are collectively training a single model with a common architecture, which is too restrictive in practice. Furthermore, knowledge exchange usually happens through sharing of gradients or model parameters. Recent gradient inversion attacks [9, 14] demonstrate that it is indeed possible to recover high fidelity information from individual gradient updates, thus violating the privacy requirement. While differential privacy (DP) [24], secure multi-party computation (MPC) [26], and trusted execution environment (TEE) [21] have been proposed as potential remedies to safeguard privacy in FL, none of the existing solutions offer an acceptable privacy-utility trade-off with reasonable efficiency. As", + "bbox": [ + 496, + 583, + 893, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "8084", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "noted in [5], sharing of high dimensional gradients/model parameters is a fundamental privacy limitation of standard FL methods, which cannot be addressed by simply wrapping around a DP, MPC, or TEE solution.", + "bbox": [ + 75, + 90, + 468, + 151 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Confidential and private collaborative learning (CaPC) [7] is the only method that claims to achieve both confidentiality and privacy in a collaborative setting. In CaPC learning, each participant is able to query other parties via a private inference protocol based on 2-party computation (2PC). However, the answering parties cannot directly reveal the inference logits to the querying party because it leaks information about their local models (e.g., through model extraction attacks). To circumvent this problem, differential privacy, private aggregation of teacher ensembles, and secure argmax computation through a central privacy guardian (PG) are employed to output only the predicted label to the querying party. A drawback of the CaPC approach is that it achieves all privacy guarantees using the help of a semi-trusted PG. Moreover, the use of differential privacy reduces the performance of FL algorithms, especially if strong privacy guarantees are required. Finally, the use of MPC requires multiple rounds of communication between the parties for each query. All other approaches that have been proposed to achieve knowledge transfer through distillation [18] require a non-private labeled/unlabeled/synthetic dataset that can be shared among participants.", + "bbox": [ + 75, + 155, + 472, + 488 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this work, we propose a new protocol called Confidential and Private Decentralized (CaPriDe) learning, where participants learn from each other in a collaborative setting while preserving their confidentiality and privacy (see Figure 1). Unlike the 2PC protocols used in CaPC, we leverage fully homomorphic encryption (FHE) to enable participants to publish their encrypted data. Since the published data is encrypted using the data owner's public key and only the owner can decrypt the data (using the secret key), confidentiality is preserved. However, the CaPriDe framework allows other collaborators to perform encrypted inference on the published (encrypted) data by applying their own local model. Mutual knowledge distillation (KD) [13] between the data owner's local model and the local models of the collaborators is used to transfer knowledge between models and ensure a collaborative gain. KD is typically achieved through minimizing the distillation loss between the student and teacher responses (logits). Since the collaborators in CaPriDe learning make predictions on encrypted data, a loss function that can be computed without any decryption is required. Hence, we propose a new distillation loss, which is an approximation of the KL divergence that can be securely computed. Only an encrypted value of the distillation loss aggregated over an entire batch is sent back to the data owner, which ensures strong privacy. Data owners can decrypt this aggregate loss value and use it for updating their local models. Our contributions are:", + "bbox": [ + 75, + 493, + 472, + 901 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We introduce the CaPriDe learning framework, which exploits FHE-based encrypted inference and knowledge distillation to achieve confidential and private collaborative learning without any central orchestration and any need for non-private shared data.", + "- To enable CaPriDe learning, we propose an encryption-friendly distillation loss that estimates the approximate KL divergence between two model predictions and design a protocol to securely compute this loss in the encrypted domain.", + "- We conduct extensive experiments to show that CaPriDe learning enables participants to achieve collaborative gain, even in the non-iid setting. To prove feasibility, we implement encrypted inference using the Tile Tensors library with a FHE-friendly model." + ], + "bbox": [ + 517, + 90, + 890, + 342 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Background", + "text_level": 1, + "bbox": [ + 500, + 358, + 627, + 375 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Federated learning (FL) was first proposed in [20] as a distributed ML algorithm that does not use user data directly. Subsequent works attempted to address various challenges in FL, including privacy [7, 19], communication efficiency [12], convergence, catastrophic forgetting [25], and secure aggregation [3]. Though several privacy-enhancing solutions have been proposed, most of them fail to provide a satisfactory balance between accuracy and privacy. Only recently, some promising solutions such as CaPC have been proposed [7]. These advancements have been made by jettisoning the conventional gradient sharing approach and borrowing ideas from the knowledge distillation literature [13]. Specifically, it has been shown that it is possible for an ML model to benefit from predictions made by other models, by aligning the local predictions with those obtained from external parties. Deep mutual learning (DML) [27] is an important work in this direction, where models learn collaboratively and teach each other throughout the training process. Compared to the distillation by a pre-defined static teacher network, DML achieves better performance, and mutually learned models outperform independently trained models irrespective of model capacity. Knowledge Distillation via Collaborative Learning (KDCL) [11], Cronus [5], and Ensemble Distillation [18] also follow a similar idea.", + "bbox": [ + 496, + 383, + 890, + 746 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Fully Homomorphic Encryption", + "text_level": 1, + "bbox": [ + 500, + 757, + 782, + 773 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Let $\\mathcal{E}(x)$ denote the encrypted value of $x$ using the public key $\\kappa$ . The encryption scheme is called homomorphic over an operation $\\oplus$ if it supports the following equation:", + "bbox": [ + 496, + 781, + 890, + 827 + ], + "page_idx": 1 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {E} \\left(x _ {1} \\oplus x _ {2}\\right) = \\mathcal {E} \\left(x _ {1}\\right) \\otimes \\mathcal {E} \\left(x _ {2}\\right), \\forall x _ {1}, x _ {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 565, + 840, + 823, + 856 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "where operators $\\otimes$ and $\\oplus$ need not be the same. The scheme is fully homomorphic if for some arbitrary function", + "bbox": [ + 496, + 869, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "8085", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/2c55aebc2280679ccc1d83c904c96b3cbc3fbb805415b830bcdb37fb1a6a6dcd.jpg", + "image_caption": [ + "Figure 1. Illustration of confidential and private decentralized (CaPriDe) learning framework for $K$ participants, showing the learning process for only participant $\\mathcal{P}_1$ . Here, $\\mathcal{D}_1 = (X_1, Y_1) = \\{x_{j,1}, y_{j,1}\\}_{j=1}^{N_1}$ is the local data of $\\mathcal{P}_1$ , $\\mathcal{E}(X_1)$ denotes the collection of encrypted unlabeled samples of $\\mathcal{P}_1$ , and $\\pmb{p}_{(1,l)}$ and $\\pmb{z}_{(1,l)}$ are the prediction probabilities and logits obtained by applying model $\\mathcal{M}_{\\theta_l}$ of participant $\\mathcal{P}_l$ to $X_1$ . Black dashed line represents exchange of encrypted data between participants in each round, and blue dashed line denotes a single transfer at the beginning of the protocol. $\\mathcal{L}_{CE}$ and $\\mathcal{L}_{DL}$ represent the cross-entropy and distillation losses, respectively." + ], + "image_footnote": [], + "bbox": [ + 158, + 93, + 812, + 315 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "$f(x_{1},x_{2},\\ldots ,x_{n})$ , there exists a function $g$ and decryption method $^d$ such that:", + "bbox": [ + 76, + 428, + 470, + 458 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nf \\left(x _ {1}, x _ {2}, \\dots , x _ {n}\\right) = d \\left(g \\left(\\mathcal {E} \\left(x _ {1}\\right), \\mathcal {E} \\left(x _ {2}\\right), \\dots , \\mathcal {E} \\left(x _ {n}\\right)\\right)\\right),\n$$\n", + "text_format": "latex", + "bbox": [ + 102, + 465, + 442, + 484 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $d$ uses secret key $\\tilde{\\kappa}$ to decrypt the given ciphertext value. Gentry [10] first constructed somewhat homomorphic encryption (SWHE) and later developed bootstrapping of ciphertexts to reduce noise. Among the various FHE schemes such as BGV, BFV, and CKKS [6], only CKKS works with real numbers and approximate computations.", + "bbox": [ + 75, + 491, + 468, + 583 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. Preliminaries", + "text_level": 1, + "bbox": [ + 76, + 589, + 217, + 604 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Typically, a supervised classifier $\\mathcal{M}$ is mapping from the input space $\\pmb{x} \\in \\mathcal{X} \\subset \\mathbb{R}^n$ to the label space $y \\in \\mathcal{Y} = \\{1,2,\\dots,M\\}$ , where $M$ is the number of classes. For convenience, we assume that $\\mathcal{M}: \\mathcal{X} \\to \\mathcal{Z}$ , where $\\mathcal{Z}$ denotes the logits space ( $\\mathcal{Z} \\subset \\mathbb{R}^M$ ). In other words, $\\mathcal{M}$ maps the input $\\pmb{x}$ to a $M$ -dimensional logits vector $\\pmb{z} = [z^{1},z^{2},\\dots,z^{M}]$ . A softmax function $(\\sigma)$ can be applied to the logits vector $\\pmb{z}$ to obtain the probability distribution $\\pmb{p} = \\sigma(\\pmb{z}) = [p^{1},p^{2},\\dots,p^{M}]$ over $M$ labels, i.e.,", + "bbox": [ + 75, + 612, + 468, + 750 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\np ^ {m} = \\frac {\\exp (z ^ {m} / T)}{C}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 204, + 756, + 468, + 786 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $T$ is the softmax temperature parameter and $C = \\sum_{i=1}^{M} \\exp(z^i / T)$ is a normalization constant that ensures $\\sum_{m=1}^{M} p^m = 1$ . Here, $p^m$ denotes the probability that input $\\mathbf{x}$ belongs to class $m, m = 1, 2, \\dots, M$ . Given two $V$ -dimensional vectors $\\mathbf{v}_1, \\mathbf{v}_2 \\in \\mathbb{R}^V$ , let $(\\mathbf{v}_1 \\cdot \\mathbf{v}_2)$ denote the dot product, $(\\mathbf{v}_1 \\odot \\mathbf{v}_2)$ denote the element-wise product, and $f(\\mathbf{v}_1)$ represent the vector obtained by applying", + "bbox": [ + 75, + 792, + 470, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "the function $f$ to each element in $\\pmb{v}_1$ . The cross entropy loss of model $\\mathcal{M}$ on a sample $(x, y)$ is defined as:", + "bbox": [ + 498, + 428, + 890, + 459 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nL _ {C E} (\\boldsymbol {p}, y) = - \\mathbf {1} _ {y} \\cdot l o g (\\boldsymbol {p}),\n$$\n", + "text_format": "latex", + "bbox": [ + 601, + 465, + 790, + 483 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\pmb{p} = \\sigma(\\mathcal{M}(\\pmb{x}))$ and $\\mathbf{1}_y$ is an indicator vector whose $y^{th}$ element is 1 and other elements are 0. The KL divergence between two distributions $\\pmb{p}_1$ and $\\pmb{p}_2$ is given by:", + "bbox": [ + 498, + 489, + 890, + 536 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nD _ {K L} \\left(\\boldsymbol {p} _ {2} \\| \\boldsymbol {p} _ {1}\\right) = \\sum_ {m = 1} ^ {M} p _ {2} ^ {m} \\log \\frac {p _ {2} ^ {m}}{p _ {1} ^ {m}} = \\left(\\boldsymbol {p} _ {2} \\cdot \\log \\boldsymbol {p} _ {2}\\right) - \\left(\\boldsymbol {p} _ {2} \\cdot \\log \\boldsymbol {p} _ {1}\\right). \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 500, + 542, + 908, + 595 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The L2 distance between two logit vectors $z_{1}$ and $z_{2}$ is:", + "bbox": [ + 500, + 597, + 870, + 612 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nD _ {L 2} \\left(\\boldsymbol {z} _ {2}, \\boldsymbol {z} _ {1}\\right) = \\sum_ {m = 1} ^ {M} \\left(z _ {2} ^ {m} - z _ {1} ^ {m}\\right) ^ {2} = \\left(\\boldsymbol {z} _ {2} - \\boldsymbol {z} _ {1}\\right) \\cdot \\left(\\boldsymbol {z} _ {2} - \\boldsymbol {z} _ {1}\\right).\n$$\n", + "text_format": "latex", + "bbox": [ + 500, + 618, + 890, + 659 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Proposed Framework", + "text_level": 1, + "bbox": [ + 500, + 666, + 705, + 683 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Problem Statement", + "text_level": 1, + "bbox": [ + 500, + 690, + 684, + 705 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Suppose that there are $K$ participants $\\mathcal{P}_1, \\mathcal{P}_2, \\dots, \\mathcal{P}_K$ . Each party $\\mathcal{P}_k, k = 1, 2, \\dots, K$ has its own local training dataset $\\mathcal{D}_k = \\{\\pmb{x}_{j,k}, y_{j,k}\\}_{j=1}^{N_k}$ of size $N_k$ using which it learns its own local ML model $\\mathcal{M}_{\\theta_k}$ , where $\\mathcal{M}$ denotes the architecture and $\\theta$ represents the parameter set. We also assume that each party $\\mathcal{P}_k$ has its own validation set $\\tilde{\\mathcal{D}}_k = \\{\\tilde{\\pmb{x}}_{j,k}, \\tilde{y}_{j,k}\\}_{j=1}^{\\tilde{N}_k}$ of size $\\tilde{N}_k$ and the validation accuracy of model $\\mathcal{M}_{\\theta_k}$ on $\\tilde{\\mathcal{D}}_k$ is $\\mathcal{A}_k$ . The goal of each participant is to collaborate with other parties in order to improve", + "bbox": [ + 496, + 713, + 890, + 856 + ], + "page_idx": 2 + }, + { + "type": "page_footnote", + "text": "1For simplicity of notation, we use the same symbol $\\mathcal{M}$ to denote the model architecture of all the $K$ participants. The proposed framework does not require all the parties to have the same model architecture.", + "bbox": [ + 498, + 862, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "8086", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "the accuracy of its local model, i.e., obtain a better model $\\mathcal{M}_{\\theta_k^*}$ that has a higher validation accuracy $\\mathcal{A}_k^*$ on $\\tilde{\\mathcal{D}}_k$ compared to $\\mathcal{A}_k$ , without compromising on the confidentiality and privacy of its local data $\\mathcal{D}_k$ . In other words, no participant $\\mathcal{P}_l, l \\neq k$ can access $\\mathcal{D}_k$ in its plaintext form or learn any private information about $\\mathcal{D}_k$ such as the $x_{j,k}$ and $y_{j,k}$ values. The utility of the framework is measured as the average collaboration gain, i.e., $U = \\frac{1}{K} \\sum_{k=1}^{K} (\\mathcal{A}_k^* - \\mathcal{A}_k)$ .", + "bbox": [ + 75, + 90, + 472, + 217 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. CaPriDe Learning Framework Description", + "text_level": 1, + "bbox": [ + 76, + 220, + 444, + 238 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Let $\\mathcal{E}(\\pmb{X}_k) = \\{\\mathcal{E}(\\pmb{x}_{j,k})\\}_{j=1}^{N_k}$ be the collection of encrypted input samples of participant $\\mathcal{P}_k$ . Here, the encryption is based on a FHE scheme with public key $\\kappa_k$ of $\\mathcal{P}_k$ . In CaPriDe learning, each party initiates the collaboration by publishing $\\mathcal{E}(\\pmb{X}_k)$ to the other participants. We now describe how the learning proceeds at participant $\\mathcal{P}_k$ . Let $\\pmb{z}_{j,(k,l)}$ and $\\pmb{p}_{j,(k,l)}$ denote the logits vector and probability distribution obtained when model $\\mathcal{M}_{\\theta_l}$ of $\\mathcal{P}_l$ is applied on the input sample $\\pmb{x}_{j,k}$ , i.e., the $j^{th}$ training sample of $\\mathcal{P}_k$ . Assume that we are at round $(t+1), 0 \\leq t < \\tau$ , of the collaboration and the current model of $\\mathcal{P}_k$ is $\\mathcal{M}_{\\theta_k^t}$ . $\\mathcal{P}_k$ performs one forward pass on its own training set $\\mathcal{D}_k$ to obtain the predictions $\\{\\pmb{p}_{j,(k,k)}\\}_{j=1}^{N_k}$ . The cross-entropy loss of party $\\mathcal{P}_k$ can be obtained as:", + "bbox": [ + 75, + 243, + 473, + 455 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {C E _ {k}} = \\sum_ {j = 1} ^ {N _ {k}} L _ {C E} \\left(\\boldsymbol {p} _ {j, (k, k)}, y _ {j, k}\\right). \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 153, + 463, + 468, + 505 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To enable knowledge transfer between participants, we make use of the knowledge distillation (KD) approach [13], where a student model learns to mimic the predictions of a teacher model. In CaPriDe learning, there is no designated teacher model, and mutual KD between multiple peer models is used for knowledge transfer. Let $\\mathcal{L}_{DL_{(k,l)}}$ denote the pairwise distillation loss between predictions of $\\mathcal{P}_k$ and $\\mathcal{P}_l$ , which is defined as:", + "bbox": [ + 75, + 511, + 468, + 632 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {D L _ {(k, l)}} = \\sum_ {j = 1} ^ {N _ {k}} D _ {D L} \\left(\\boldsymbol {p} _ {j, (k, k)}, \\boldsymbol {p} _ {j, (k, l)}\\right), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 135, + 636, + 468, + 679 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $D_{DL}(\\pmb{p}_1, \\pmb{p}_2)$ denotes the mimic distance between two predictions $\\pmb{p}_1$ and $\\pmb{p}_2$ . The total distillation loss for $\\mathcal{P}_k$ can be computed as:", + "bbox": [ + 75, + 686, + 468, + 731 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {D L _ {k}} = \\frac {1}{K - 1} \\sum_ {l = 1, l \\neq k} ^ {K} \\mathcal {L} _ {D L _ {(k, l)}}. \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 160, + 737, + 468, + 780 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Therefore, the overall loss of $\\mathcal{P}_k$ is given by:", + "bbox": [ + 76, + 786, + 372, + 801 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {k} = \\mathcal {L} _ {C E _ {k}} + \\lambda_ {k} \\mathcal {L} _ {D L _ {k}}, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 189, + 809, + 468, + 825 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\lambda_{k}$ is the weighting factor between the cross-entropy and distillation losses for $\\mathcal{P}_k$ . The model of participant $\\mathcal{P}_k$ is then updated through a backward pass as follows:", + "bbox": [ + 75, + 832, + 468, + 878 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\theta_ {k} ^ {t + 1} = \\theta_ {k} ^ {t} - \\alpha_ {k} \\nabla_ {\\theta_ {k} ^ {t}} \\mathcal {L} _ {k}, \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 189, + 883, + 468, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\alpha_{k}$ is the learning rate of $\\mathcal{P}_k$", + "bbox": [ + 500, + 90, + 736, + 106 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Encryption-friendly Distillation Loss", + "text_level": 1, + "bbox": [ + 500, + 114, + 818, + 132 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The key missing component in the above CaPriDe learning framework is the computation $D_{DL}$ in Eq. 4. For participant $\\mathcal{P}_k$ , its logit vectors $\\pmb{z}_{.,(k,k)}$ and predictions $\\pmb{p}_{.,(k,k)}$ are available in the plaintext domain. However, since participant $\\mathcal{P}_l$ receives the input samples of $\\mathcal{P}_k$ in encrypted form, it can perform inference in the encrypted domain and obtain only $\\mathcal{E}\\left(\\pmb{z}_{.,(k,l)}\\right)$ , where $\\mathcal{E}\\left(\\pmb{z}_{.,(k,l)}\\right) = \\mathcal{M}_{\\theta_l}\\left(\\mathcal{E}(\\pmb{x}_{.,k})\\right)$ . It is not straightforward to obtain $\\mathcal{E}\\left(\\pmb{p}_{.,(k,l)}\\right)$ from $\\mathcal{E}\\left(\\pmb{z}_{.,(k,l)}\\right)$ because it involves computing a non-linear softmax function in the encrypted domain, which is a challenging task in the FHE domain and consumes a large multiplicative depth. Even if this softmax operation is somehow implemented, it is not possible to send $\\mathcal{E}\\left(\\pmb{p}_{.,(k,l)}\\right)$ back to $\\mathcal{P}_k$ , who can decrypt it and obtain $\\pmb{p}_{.,(k,l)}$ . Since $\\mathcal{P}_k$ also has knowledge of $\\pmb{x}_{.,k}$ , it becomes feasible for $\\mathcal{P}_k$ to carry out a model extraction attack based on $(\\pmb{x}_{.,k},\\pmb{p}_{.,(k,l)})$ pairs to glean information about model $\\mathcal{M}_{\\theta_l}$ . This violates the privacy requirements of participant $\\mathcal{P}_l$ . Hence, there is a need for a loss function that can be computed in the encrypted domain.", + "bbox": [ + 496, + 138, + 893, + 439 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "A straightforward solution is to compute the L2 distance between logit vectors $\\mathbf{z}_{.,(k,k)}$ and $\\mathbf{z}_{.,(k,l)}$ . In this case, participant $\\mathcal{P}_k$ can publish $\\mathcal{E}\\left(\\mathbf{z}_{.,(k,k)}\\right)$ and the pairwise distillation loss $\\mathcal{L}_{DL_{(k,l)}}$ can be computed by participant $\\mathcal{P}_l$ as:", + "bbox": [ + 496, + 440, + 892, + 503 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\left. \\mathcal {E} \\left(D _ {D L} \\left(\\boldsymbol {p} _ {j, (k, k)}, \\boldsymbol {p} _ {j, (k, l)}\\right)\\right) \\equiv \\mathcal {E} \\left(D _ {L 2} \\left(\\boldsymbol {z} _ {j, (k, k)}, \\boldsymbol {z} _ {j, (k, l)}\\right)\\right) \\right. \\\\ = \\left(\\mathcal {E} \\left(\\boldsymbol {z} _ {j, (k, k)}\\right) - \\mathcal {E} \\left(\\boldsymbol {z} _ {j, (k, l)}\\right)\\right) \\cdot \\left(\\mathcal {E} \\left(\\boldsymbol {z} _ {j, (k, k)}\\right) - \\mathcal {E} \\left(\\boldsymbol {z} _ {j, (k, l)}\\right)\\right) \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 501, + 513, + 900, + 561 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "However, our experiments indicate the above simple solution based on L2 distance does not lead to high utility $U$ . On the other hand, we observed that using the KL divergence between predictions of $\\mathcal{P}_k$ and $\\mathcal{P}_l$ almost always leads to better utility. But, $D_{KL}\\left(\\pmb{p}_{.,(k,k)}\\| \\pmb{p}_{.,(k,l)}\\right)$ cannot be directly computed in the encrypted domain because participant $\\mathcal{P}_l$ only has access to the encrypted logits $\\mathcal{E}\\left(z_{.,(k,l)}\\right)$ .", + "bbox": [ + 496, + 571, + 890, + 685 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Fortunately, this problem can be solved by approximating the KL divergence. Combining Eqs. 1 and 2, we get:", + "bbox": [ + 496, + 684, + 890, + 715 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nD _ {K L} \\left(\\boldsymbol {p} _ {2} \\right\\| \\boldsymbol {p} _ {1}) = \\sum_ {m = 1} ^ {M} p _ {2} ^ {m} \\log p _ {2} ^ {m} - \\sum_ {m = 1} ^ {M} p _ {2} ^ {m} \\left(z _ {1} ^ {m} / T\\right) + M \\log C. \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 500, + 724, + 906, + 779 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Here, the first term can be denoted as $(\\pmb{p}_2 \\cdot \\log \\pmb{p}_2)$ and measures the entropy (uncertainty) in $\\pmb{p}_2$ . The second term measures the divergence (unnormized cross-entropy) between $\\pmb{p}_2$ and $\\pmb{p}_1$ and can be expressed as $(\\pmb{p}_2 \\cdot \\pmb{z}_1 / T)$ . The final term $M \\log C$ is just a normalization constant that does not affect the relative ordering of the class predictions. We observed that this normalization term does not appear to play a major role in model training. Hence, we ignore this value", + "bbox": [ + 496, + 780, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "8087", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "code", + "sub_type": "algorithm", + "code_caption": [ + "Algorithm 1 CaPRiDe learning algorithm" + ], + "code_body": "Input: Number of participants $K$ , weight parameter $\\lambda_{k}$ , learning rate $\\alpha_{k}$ for each participant \nAssumption: Each participant has local training set $\\mathcal{D}_k = \\{\\pmb {x}_{j,k},y_{j,k}\\}_{j = [1,N_k]}$ of size $N_{k}$ , a fully homomorphic encryption function $\\mathcal{E}$ with public key $\\kappa_{k}$ , a decryption function $d$ with secret key $\\tilde{\\kappa}_k$ \nfor each participant $k = 1,2,\\dots ,K$ do Train initial model $\\mathcal{M}_{\\theta_k^0}$ based on local data $\\mathcal{D}_k$ Publish encrypted data $\\{\\mathcal{E}(\\pmb{x}_{j,k})\\}_{j = [1,N_k]}$ \nend for \nfor each round $t = 1,2,\\dots ,\\tau$ do \nfor each participant $k = 1,2,\\dots ,K$ do $\\{\\pmb {p}_{j,(k,k)}\\}_{j = [1,N_k]}\\gets \\{\\sigma (\\mathcal{M}_{\\theta_k^{t - 1}}(\\pmb {x}_{j,k}))\\}_{j = [1,N_k]}$ $\\mathcal{L}_{CE_k}\\gets$ Compute cross-entropy loss of participant $k$ using Equation 3 \n $\\hat{\\mathcal{L}}_{KL_k}\\gets$ Compute local uncertainty of participant $k$ using Equation 11 \nPublish encrypted inferences $\\{\\mathcal{E}(\\pmb {p}_{j,(k,k)})\\}_{j = [1,N_k]}$ \nfor each participant $l = 1,2,\\dots ,K,l\\neq k$ do $\\{\\mathcal{E}(\\pmb {z}_{j,(k,l)})\\}_{j = [1,N_k]}\\gets \\{\\mathcal{M}_{\\theta_l^{t - 1}}(\\mathcal{E}(\\pmb {x}_{j,k}))\\}_{j = [1,N_k]}$ (Perform inference on encrypted data) \n $\\mathcal{E}(\\tilde{\\mathcal{L}}_{KL_{(k,l)}})\\gets$ Compute pairwise divergence in encrypted domain using Equation 12 \nSend $\\mathcal{E}(\\tilde{\\mathcal{L}}_{KL_{(k,l)}})$ to participant $k$ $\\mathcal{L}_{DL_{(k,l)}}\\gets$ Compute pairwise distillation loss between participants $k$ and $l$ using Equation 10 \nend for \n $\\mathcal{L}_{DL_k}\\gets$ Compute distillation loss of participant $k$ using Equation 5 \n $\\mathcal{L}_k\\gets$ Compute total loss of participant $k$ using Equation 6 \n $\\theta_k^t\\gets$ Update local model of participant $k$ using Equation 7 \nend for \nend for", + "bbox": [ + 78, + 108, + 890, + 483 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "and define the approximate KL divergence is defined as:", + "bbox": [ + 76, + 513, + 450, + 527 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {D} _ {K L} \\left(\\boldsymbol {p} _ {2} \\| \\boldsymbol {p} _ {1}\\right) = \\left(\\boldsymbol {p} _ {2} \\cdot \\log \\boldsymbol {p} _ {2}\\right) - \\left(\\boldsymbol {p} _ {2} \\cdot \\boldsymbol {z} _ {1} / T\\right). \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 120, + 537, + 468, + 555 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Going back to Eq. 4, the pairwise distillation loss based on the above approximate KL divergence can be easily computed in the encrypted domain as:", + "bbox": [ + 75, + 565, + 468, + 609 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\left. \\mathcal {E} \\left(D _ {D L} \\left(\\boldsymbol {p} _ {j, (k, k)}, \\boldsymbol {p} _ {j, (k, l)}\\right)\\right) \\equiv \\mathcal {E} \\left(\\hat {D} _ {K L} \\left(\\boldsymbol {p} _ {j, (k, k)} | | \\boldsymbol {p} _ {j, (k, l)}\\right)\\right) \\right. \\\\ = \\mathcal {E} \\left(\\boldsymbol {p} _ {j, (k, k)} \\cdot \\log \\boldsymbol {p} _ {j, (k, k)}\\right) - \\mathcal {E} \\left(\\boldsymbol {p} _ {j, (k, k)}\\right) \\cdot \\mathcal {E} \\left(\\boldsymbol {z} _ {j, (k, l)}\\right) / T \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 76, + 618, + 483, + 672 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The above distillation loss based on the approximate KL divergence is encryption-friendly because it involves only linear operations and scaling by a constant plaintext value $T$ . In practice, our goal is to compute $\\mathcal{L}_{DL_{(k,l)}}$ at participant $\\mathcal{P}_k$ . Therefore, Eq. 4 can be further reformulated as:", + "bbox": [ + 75, + 681, + 468, + 756 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {D L _ {(k, l)}} = \\hat {\\mathcal {L}} _ {K L _ {k}} - d \\left(\\mathcal {E} \\left(\\tilde {\\mathcal {L}} _ {K L _ {(k, l)}}\\right)\\right), \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 138, + 765, + 468, + 792 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where", + "bbox": [ + 76, + 801, + 120, + 813 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathcal {L}} _ {K L _ {k}} = \\sum_ {j = 1} ^ {N _ {k}} \\left(\\boldsymbol {p} _ {j, (k, k)} \\cdot \\log \\boldsymbol {p} _ {j, (k, k)}\\right), \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 153, + 811, + 468, + 854 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {E} \\left(\\tilde {\\mathcal {L}} _ {K L _ {(k, l)}}\\right) = \\sum_ {j = 1} ^ {N _ {k}} \\mathcal {E} \\left(\\boldsymbol {p} _ {j, (k, k)}\\right) \\cdot \\mathcal {E} \\left(\\boldsymbol {z} _ {j, (k, l)}\\right) / T. \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 91, + 861, + 468, + 902 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Since $\\mathcal{P}_k$ has access to $\\pmb{p}_{j,(k,k)}$ , it is straightforward to compute $\\hat{\\mathcal{L}}_{KL_k}$ in plaintext space. If participant $\\mathcal{P}_k$ can publish $\\left\\{\\mathcal{E}(\\pmb{p}_{j,(k,k)})\\right\\}_{j=1}^{N_k}$ in each round, $\\mathcal{P}_l$ can compute $\\mathcal{E}\\left(\\tilde{\\mathcal{L}}_{KL(k,l)}\\right)$ in the encrypted domain because it already has access to $\\left\\{\\mathcal{E}\\left(z_{j,(k,l)}\\right)\\right\\}_{j=1}^{N_k}$ . The encrypted value of $\\tilde{\\mathcal{L}}_{KL(k,l)}$ is sent back to participant $\\mathcal{P}_k$ , who decrypts it to compute the overall loss via Eqs. 10, 5 and 6.", + "bbox": [ + 496, + 512, + 890, + 652 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Privacy Analysis: In the CaPriDe learning framework, the only quantity that can potentially leak private information is the pairwise distillation loss between participants $\\tilde{\\mathcal{L}}_{KL_{(k,l)}}$ , which has already been aggregated over all the $N_{k}$ samples. A curious participant $\\mathcal{P}_k$ attempting to reconstruct the logit values $z_{j,(k,l)}^m$ of participant $\\mathcal{P}_l$ needs to dis-aggregate this sum and compute per-sample loss, which is infeasible in the absence of other constraints. Hence, the proposed CaPriDe learning framework is privacy-preserving. It must be emphasized that even gradient inversion attacks [9, 14] are significantly harder when the batch size is large (above 16) due to the same reason. The proposed method also has other advantages that enhance privacy: (i) A curious participant cannot query its collaborator adaptively. The encrypted samples are published once at the beginning and do not change. (ii) The responding participant can choose a", + "bbox": [ + 496, + 657, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "8088", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "random subset of samples in each round to compute the aggregate loss. This can improve efficiency as well as provide a privacy benefit by eliminating correlation across multiple rounds. (iii) We can incorporate differential privacy (DP) by adding noise to the aggregate pairwise distillation loss. Since the framework allows exchange of only encrypted input samples and encrypted predictions, which never get decrypted during the execution of the protocol, the CaPriDe learning approach also preserves data confidentiality.", + "bbox": [ + 75, + 90, + 472, + 227 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 76, + 242, + 209, + 258 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Datasets", + "text_level": 1, + "bbox": [ + 76, + 267, + 179, + 282 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "CIFAR-10 [17] is a dataset consisting of 60000 $32 \\times 32$ RGB images from 10 classes, with 6000 images per class. It has 50000 training and 10000 test samples.", + "bbox": [ + 75, + 290, + 468, + 335 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "CIFAR-100 dataset [17] is similar to the CIFAR-10 dataset, but it has 100 classes containing 600 samples each. Training and evaluation sets are split in the same way as in CIFAR-10. CIFAR-100 dataset is used with data augmentation to account for the problem complexity. The data augmentations employed include random rotation (up to 15 degrees), random crop, and horizontal flip.", + "bbox": [ + 75, + 335, + 468, + 443 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "HAM10000 is a multi-class imbalanced dataset comprising of 10,015 dermoscopic images from diverse populations with 7 categories of pigmented lesions. We randomly perform an $85\\% / 15\\%$ split for training/ testing.", + "bbox": [ + 75, + 443, + 468, + 503 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Experimental Setup", + "text_level": 1, + "bbox": [ + 76, + 513, + 266, + 530 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We compare the CaPriDe learning framework against the following baselines: vanilla FL based on FedAvg (FedAvg) [20], FL with local differential privacy (FedAvg + DP) [24], and deep mutual learning (DML) [27]. We use ResNet-18 architecture for all our experiments. Stochastic Gradient Descent (SGD) algorithm with momentum is used as the optimizer. For all methods, 25 epochs of local training happens before 75 rounds of collaboration. Momentum is set to 0.9, the initial learning rate is set to $\\alpha_{k} = 0.1$ and the learning rate decays in rounds 50 and 75 by a factor $\\gamma = 0.1$ . The batch size is set to 128 for CIFAR-10 and CIFAR-100 and 32 for HAM10000. The weight factor $\\lambda_{k}$ between cross-entropy and distillation losses is set to 50 for CIFAR-10 and CIFAR-100 and 20 for HAM10000. All algorithms are implemented using PyTorch and experiments are carried out on a single NVIDIA RTX A6000 GPU. Additionally, in CaPriDe and DML, the logits are scaled by a temperature parameter $T = 5.0$ . Finally, for FedAvg + DP, the Gaussian noise mechanism is used with $\\sigma_{g} = 0.01$ . To implement an FHE-compatible version of ResNet-18, we use the Tile Tensors framework [1,2], which in turn relies on low-level FHE libraries such as HElib and SEAL. For encrypted inference, ReLU is replaced with a polynomial approximation of Gaussian Error Linear Unit", + "bbox": [ + 75, + 537, + 472, + 902 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/43a7b76ab875c0ebbfaa5638e55f6c51a369f325bf764b75d75e67e3419db4c8.jpg", + "image_caption": [ + "Figure 2. Accuracy trend of two $(K = 2)$ collaboratively trained models on different datasets (rows) and for different partition strategies (columns)." + ], + "image_footnote": [], + "bbox": [ + 506, + 90, + 888, + 369 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n(g (x) = 0. 0 7 1 1 + 0. 5 x + 0. 2 5 7 6 x ^ {2} - 0. 0 1 2 8 x ^ {4}).\n$$\n", + "text_format": "latex", + "bbox": [ + 500, + 452, + 830, + 468 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We consider three types of strategies to partition the training dataset among the participants: (i) Homogeneous: each participant has an equal number of samples per class, (ii) Heterogeneous: each participant has an unequal number of samples determined randomly (both total number and number of samples per class), (iii) Non-overlapping class distribution: (denoted as \"no class overlap\") each participant has samples from a non-overlapping subset of classes (e.g., for $K = 2$ , participants get even and odd classes). While homogeneous partitioning ensures that data distributions of participants are iid, heterogeneous split simulates typical non-iid scenarios, and the no class overlap case corresponds to extreme statistical heterogeneity.", + "bbox": [ + 496, + 469, + 890, + 666 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Collaborative Learning Results", + "text_level": 1, + "bbox": [ + 500, + 675, + 777, + 691 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To begin with, we establish the utility (collaborative gain) of the proposed CaPriDe learning framework for the case of $K = 2$ participants on all three datasets and all three partition strategies. The results of these experiments are shown in Figure 2. In each case, the collaboration gain of CaPriDe learning is positive and the actual collaboration gain values can be found in Table 1. We make the following observations based on these results.", + "bbox": [ + 496, + 699, + 890, + 819 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "- CaPriDe learning works not only in the iid setting but also equally well in the two non iid settings. This", + "bbox": [ + 517, + 834, + 890, + 864 + ], + "page_idx": 5 + }, + { + "type": "page_footnote", + "text": "2Since HAM10000 already has a massive class imbalance and an odd number of class, we omit the no class overlap setting for this dataset.", + "bbox": [ + 500, + 875, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "8089", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/56d7237cbf3bee473ba0218f864896f783fcc45292f809ea4f8533990ac36947.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
DatasetSettingKFedAvgFedAvg+DPCaPriDe
CIFAR-10Homogeneous23.46-20.541.58
CIFAR-10Homogeneous59.22-11.83.55
CIFAR-10Homogeneous1015.442.585.52
CIFAR-10Heterogeneous26.12-17.040.93
CIFAR-10Heterogeneous513.78-7.845.06
CIFAR-10Heterogeneous1021.815.696.20
CIFAR-10No class overlap229.8511.088.16
CIFAR-100Homogeneous27.08-30.974.19
CIFAR-100Homogeneous522.68-13.819.10
CIFAR-100Homogeneous1026.10-0.1111.69
CIFAR-100Heterogeneous29.59-28.896.28
CIFAR-100Heterogeneous522.50-13.269.23
CIFAR-100Heterogeneous1034.383.8610.68
CIFAR-100No class overlap219.405.147.75
HAM10000Homogeneous20.970.481.81
HAM10000Heterogeneous21.93-1.211.67
", + "bbox": [ + 78, + 88, + 467, + 327 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 1. Collaborative gain (\\%) on all datasets and all partition strategies for different number of participants $K$ using different algorithms (columns 4-6). DML results are not reported here because there is negligible difference between CaPriDe and DML.", + "bbox": [ + 75, + 337, + 470, + 393 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "demonstrates that knowledge distillation is reasonably effective even under extreme statistical heterogeneity.", + "bbox": [ + 107, + 421, + 467, + 452 + ], + "page_idx": 6 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- There is a negligible difference between the performance of DML (which uses true KL divergence) and CaPriDe learning (which uses approximate KL divergence) in all the settings. This clearly justifies the dropping of the normalization (third) term in Eq. 8 to obtain the encryption-friendly distillation loss in Eq. 9. While DML offers no confidentiality or privacy, CaPriDe learning offers both without any utility cost.", + "- In almost all the settings, there is a huge difference in the collaboration gain between FedAvg and CaPriDe learning. This is unsurprising because FedAvg involves sharing and aggregation of full model parameters leading to high utility at a significant cost to privacy. On the other hand, FedAvg + DP provides strong privacy, but this comes with a significant degradation in utility. Hence, it can be argued that CaPriDe learning achieves a good trade-off between privacy and utility compared to standard FL algorithms." + ], + "bbox": [ + 94, + 464, + 468, + 750 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Next, we evaluate the utility of the proposed method with the different number of participants $K = 2,5,10$ on CIFAR-10 and CIFAR-100 datasets based on the homogeneous partition (iid) setting. From Figure 3 and Table 1 we observe that the collaboration gain increases with the number of participants. This is to be expected because the training samples per participant become less when there are more participants, consequently resulting in lower individual accuracy.", + "bbox": [ + 75, + 763, + 470, + 902 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/08b0e21db30a0721f446a7ba76a23281ac45917998144ad7fc2ec990142e0562.jpg", + "image_caption": [ + "Figure 3. Accuracy with different number of participants $K$ based on the homogeneous partition (iid) setting." + ], + "image_footnote": [], + "bbox": [ + 506, + 90, + 888, + 281 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a92183316b47486ba36b76855ea902e5a985c9b32356968e9760a3d7000f5c81.jpg", + "image_caption": [ + "(a) Homogeneous" + ], + "image_footnote": [], + "bbox": [ + 501, + 339, + 630, + 434 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/0efbfc2e767cf394f01e1fc1c1ba384a62e69cfbceae82c5bd82ed4a8b8ff32b.jpg", + "image_caption": [ + "(b) Heterogeneous" + ], + "image_footnote": [], + "bbox": [ + 632, + 339, + 759, + 434 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d385c9c34c991f29411068436b1d8ca28973a1c6a365138fe76c88d8a3deff76.jpg", + "image_caption": [ + "(c) No overlap" + ], + "image_footnote": [], + "bbox": [ + 763, + 339, + 890, + 434 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d6a14f28d4d29f9e9313aea1d6fa5fa0f6a697790d9e6354e97fd7099b8a5fc9.jpg", + "image_caption": [ + "Figure 4. Per-class accuracy of CaPriDe learning evaluated on CIFAR-10 averaged across 5 participants. Dashed lines represent mean accuracy and error bars indicate standard deviation.", + "(a) Homogeneous", + "Figure 5. Per-participant accuracy of CaPriDe learning evaluated on CIFAR-100 dataset." + ], + "image_footnote": [], + "bbox": [ + 501, + 521, + 629, + 618 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d578d0c4b1510d5adfd0edbfab2897f8c6ed6d25b5ffa42683a90e928c362176.jpg", + "image_caption": [ + "(b) Heterogeneous" + ], + "image_footnote": [], + "bbox": [ + 632, + 523, + 758, + 618 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/4e4fc757f199692b37ad1218748a2bf0cf5b388b180730e0331ca7dfc8c3ddb3.jpg", + "image_caption": [ + "(c) No overlap" + ], + "image_footnote": [], + "bbox": [ + 763, + 523, + 890, + 618 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Furthermore, we evaluate the class-wise and participant-wise accuracy of our CaPriDe learning algorithm with $K = 5$ participants. Figure 4 shows class-wise accuracy averaged across 5 participants for each setting. From this figure, it is clear that almost all the classes benefit from collaboration, and the per-class accuracy of the individual models after collaboration has a lower standard deviation compared to the case without collaboration, especially in homogeneous and heterogeneous settings. Thus, in these two settings, the individual models become similar to each other after collaboration proving the successful transfer of knowledge between participants. Note that Figure 4 displays the average top-1 accuracy of the participants. The collaboration gains", + "bbox": [ + 496, + 704, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "8090", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "in the no class overlap case are significantly higher when top-5 accuracy is considered. Finally, Figure 5 illustrates participant-wise accuracy over all classes for each setting. It is obvious that each participant has a positive collaboration gain, which is the eventual goal that we set out to achieve. Moreover, all the 5 participant models have a similar accuracy at the end of the collaboration, once again demonstrating good knowledge transfer. Other results such as the sensitivity of the proposed method to weight factor $\\lambda_{k}$ , the sensitivity of FedAvg + DP to noise variance, and the comparison between approximate KL divergence and L2 loss are included in the supplementary material.", + "bbox": [ + 78, + 90, + 467, + 271 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4. FHE Implementation Results", + "text_level": 1, + "bbox": [ + 78, + 282, + 336, + 297 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "For FHE implementation using the Tile Tensors framework, Table 2 summarizes the information about all three datasets at the same security level of 128 bits. Two critical metrics to be kept in mind are the time required for encrypted inference and the memory consumed by a ciphertext. While encrypted inference time determines the computational complexity and latency involved in each round, the ciphertext size determines the communication burden. Note that encrypted inference can be performed on a persample basis or a batch of samples [4]. The single sample inference mode is preferred for machine learning as-a-service applications, where low latency is a critical requirement. However, the batched inference is better suited for the CaPriDe learning framework, where only the overall time required to complete inference on all the samples matter.", + "bbox": [ + 78, + 306, + 467, + 532 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/fa950b2ec76fc5803e10412bd103b1b7c9943a9b4dec79f034cc26ec85a38405.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Security LevelCIFAR-10CIFAR-100HAM10000
128128128
Number of slots163841638416384
Time taken to encrypt one sample90 ms103 ms619 ms
Ciphertext size of one sample29.101 KB29.152 KB1.359 MB
Time taken to encrypt a batch of 32 samples1.31 s1.26 s15.57 s
Encrypted inference of a batch of 32 samples110.21 s112.09 s896.12 s
", + "bbox": [ + 78, + 545, + 467, + 648 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 2. FHE configuration parameters and memory and computational requirements for the chosen datasets based on ResNet-18 architecture.", + "bbox": [ + 78, + 660, + 467, + 700 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The encrypted inference time for a batch of 32 samples is approximately 110s for CIFAR-10 and CIFAR-100 and 900s for HAM10000. Thus, if the local dataset of each participant has 1024 training samples, it will take around 1 hour (8 hours) for encrypted inference per participant and then computation of the pairwise distillation loss in each communication round for CIFAR-10 and CIFAR-100 (HAM10000). While this is a significant computational load, it is less of an issue in the cross-silo collaborative learning setting, especially in sectors such as finance and healthcare, where data confidentiality and privacy as well as accuracy gain are more critical. Moreover, the ciphertext", + "bbox": [ + 78, + 719, + 467, + 898 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "size for one sample is approximately, 30KB for CIFAR-10 and CIFAR-100 and 1.36MB for HAM10000. For the same local dataset size of 1024, the encrypted dataset will consume 30MB for CIFAR-10 and CIFAR-100 and 1.36GB for HAM10000. Fortunately, the encrypted dataset is published and needs to be downloaded only once by each participant at the beginning of the protocol. During every communication round, only the encrypted predictions and pairwise losses are exchanged. Depending on the local dataset size, the number of classes, and the ciphertext packing strategy used, these intermediate values can be packed into at most a few tens of ciphertexts (note that this does not depend on the image size). Hence, the per round communication burden of CaPriDe learning framework is relatively small.", + "bbox": [ + 501, + 90, + 890, + 301 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Limitations and Potential Solutions: Note that all the above complexity computations are based on the assumption of $K = 2$ participants. As $K$ increases, these requirements will increase linearly for each participant bringing the scalability of the system into question. Hence, the vanilla CaPriDe learning framework is more appropriate for a small network of participants. However, there are ways to improve scalability borrowing on ideas from cross-device FL. For example, instead of computing the pairwise distillation loss with respect to each participant, a random subset of participants can be sampled in each round for distillation. Similarly, the computational complexity for each participant can be significantly reduced by performing encrypted inference only on a randomly selected subset of encrypted samples. Furthermore, these actions can be determined at the local level without any central orchestration. The need to have FHE-friendly models for encrypted inference is a key limitation, which precludes the use of very deep models within the proposed framework. One potential solution is to employ two models at each participant: one deep model with large capacity and another smaller FHE-friendly model for encrypted inference on other participants' data and perform an additional mutual distillation between these two models after each round.", + "bbox": [ + 503, + 305, + 890, + 666 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Summary", + "text_level": 1, + "bbox": [ + 503, + 691, + 604, + 707 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Collaborative learning without compromising on data confidentiality and privacy is a valuable tool in applications with stringent data privacy requirements. We have proposed a new framework called CaPriDe learning based on inference in the FHE domain to achieve this goal. Knowledge distillation is used to exchange useful information between models, and a new distillation loss based on approximate KL divergence has been proposed to enable secure loss computation in the encrypted domain. While the proposed approach has shown promising results, its efficiency can be further improved. Other vulnerabilities such as poisoning attacks will also be addressed in the future.", + "bbox": [ + 503, + 719, + 890, + 898 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8091", + "bbox": [ + 483, + 946, + 511, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Ehud Aharoni, Allon Adir, Moran Baruch, Nir Drucker, Gilad Ezov, Ariel Farkash, Lev Greenberg, Rami Masalha, Guy Moshkowish, Dov Murik, Hayim Shaul, and Omri Soceanu. HeLayers: A Tile Tensors Framework for Large Neural Networks on Encrypted Data. arXiv:2011.01805v2, 2021. 6", + "[2] Ehud Aharoni, Nir Drucker, Gilad Ezov, Hayim Shaul, and Omri Soceanu. Complex encoded tile tensors: Accelerating encrypted analytics. In IEEE S&P 20 (5), pages 35-43, 2022. 6", + "[3] Keith Bonawitz, Vladimir Ivanov, Ben Kreuter, Antonio Marcedone, H. Brendan McMahan, Sarvar Patel, Daniel Ramage, Aaron Segal, and Karn Seth. Practical secure aggregation for privacy preserving machine learning. ACM Conference on Computer and Communications Security (ACM CCS), 2017. 2", + "[4] Alon Brutzkus, Ran Gilad-Bachrach, and Oren Elisha. Low latency privacy preserving inference. In International Conference on Machine Learning, pages 812-821, 2019. 8", + "[5] Hongyan Chang, Virat Shejwalkar, Reza Shokri, and Amir Houmansadr. Cronus: Robust and heterogeneous collaborative learning with black-box knowledge transfer. In arXiv preprint arXiv:1912.11279, 2019. 2", + "[6] Jung Hee Cheon, Andrey Kim, Miran Kim, and Yongsoo Song. Homomorphic Encryption for Arithmetic of Approximate Numbers. 2016. 3", + "[7] Christopher A. Choquette-Choo, Natalie Dullerud, Adam Dziedzic, Yunxiang Zhang, Somesh Jha, Nicolas Papernot, and Xiao. Wang. CaPC Learning: Confidential and Private Collaborative Learning. Proceedings of ICLR, 2021. 2", + "[8] Deshpande, Nikhil and Shiffman, Gary M. Federated learning: The new thing in ai/ml for detecting financial crimes and managing risk. https://morningconsult.com/opinions/federated-learning-the-new-thing-in-ai-ml-for-detecting-financial-crimes-and-managing-risk/, 2021.1", + "[9] Jonas Geiping, Hartmut Bauermeister, Hannah Dröge, and Michael. Moeller. Inverting gradients-How easy is it to break privacy in federated learning? Advances in Neural Information Processing Systems (NeurIPS), 33, 2020. 1, 5", + "[10] Craig. Gentry. Fully Homomorphic Encryption. PhD thesis, 2009. 3", + "[11] Qiushan Guo, Xinjiang Wang, Yichao Wu, Zhipeng Yu, Ding Liang, Xiaolin Hu, and Ping. Luo. Online Knowledge Distillation via Collaborative Learning. Proceedings of CVPR, 2020. 2", + "[12] Jenny Hamer, Mehryar Mohri, and Ananda Theertha Suresh. FedBoost: A Communication-Efficient Algorithm for Federated Learning. Proceedings of the 37th International Conference on Machine Learning, PMLR, 119:3973-3983, 2020. 2", + "[13] Geoffrey Hinton, Oriol Vinyls, and Jeff Dean. Distilling the Knowledge in a Neural Network. NeurIPS, 2014. 2, 4", + "[14] Jiwnoo Jeon, Kangwook Lee, Sewoong Oh, and Jungseul. Ok. Gradient Inversion with Generative Image Prior. Advances in Neural Information Processing Systems (NeurIPS), 34, 2021. 1, 5" + ], + "bbox": [ + 78, + 114, + 470, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[15] Peter Kairouz, H. Brendan McMahan, Brendan Avent, Aurélien Bellet, and Mehdi. Bennis. Advances and Open Problems in Federated Learning. Now Publishers, 2021. 1", + "[16] Georgios A. Kaissis, Marcus R. Makowski, Daniel Rückert, and Rickmer F. Braren. Secure, privacy-preserving and federated machine learning in medical imaging. Nature Machine Intelligence, 2:305-311, 2020. 1", + "[17] Alex Krizhevsky. Gradient-based learning applied to document recognition. Technical report, University of Toronto, 2009. 6", + "[18] Tao Lin, Lingjing Kong, Sebastian U Stich, and Martin Jaggi. Ensemble distillation for robust model fusion in federated learning. In Advances in Neural Information Processing Systems, volume 33, pages 2351-2363, 2020. 2", + "[19] Chuan Ma, Jun Li, Ming Ding, Howard H. Yang, Feng Shu, Tony Q. S. Quek, and Tony Q. S. Poor. On safeguarding privacy and security in the framework of federated learning. IEEE Network, pages 242-248, 2020. 2", + "[20] H. Brendan McMahan, Eider Moore, Daniel Ramage, Seth Hampson, and Blaise Agüera y. Arcas. Communication-Efficient Learning of Deep Networks from Decentralized Data. Proceedings of the 20th International Conference on Artificial Intelligence and Statistics (AISTATS), 2016. 2, 6", + "[21] Fan Mo, Hamed Haddadi, Kleomenis Katevas, Eduard Marin, Diego Perino, and Nicolas Kourtellis. PPFL: Privacy-Preserving Federated Learning with Trusted Execution Environments. In Proceedings of the 19th Annual International Conference on Mobile Systems, Applications, and Services (MobiSys), page 94-108, 2021. 1", + "[22] Willem G van Panhuis, Proma Paul, Claudia Emerson, John Grefenstette, Richard Wilder, Abraham J Herbst, and Donald S Heymann, David Burke. A systematic review of barriers to data sharing in public health. BMC Public Health, 2014. 1", + "[23] Micah J. Sheller, Brandon Edwards, G. Anthony Reina, Jason Martin, Sarthak Pati, Aikaterini Kotrotsou, Mikhail Milchenko, Weilin Xu, Daniel Marcus, Rivka R. Colen, and Spyridon. Bakas. Federated learning in medicine: facilitating multi-institutional collaborations without sharing patient data. Nature Scientific Reports, 10, 2020. 1", + "[24] Kang Wei, Jun Li, Ming Ding, Chuan Ma, Howard H. Yang, and Farhad. Farokhi. Federated Learning With Differential Privacy: Algorithms and Performance Analysis. IEEE Transactions on Information Forensics and Security, 15:3454-3469, 2020. 1, 6", + "[25] Jaehong Yoon, Wonyong Jeong, Giwoong Lee, Eunho Yang, and Sung Ju. Hwang. Federated Continual Learning with Weighted Inter-client Transfer. Proceedings of the 37th International Conference on Machine Learning, PMLR, 139:12073-12086, 2021. 2", + "[26] Chengliang Zhang, Suyi Li, Junzhe Xia, Wei Wang, Feng Yan, and Yang Liu. BatchCrypt: efficient homomorphic encryption for cross-silo federated learning. Proceedings of the USENIX Conference, page 493-506, 2020. 1", + "[27] Ying Zhang, Tao Xiang, Timothy M. Hospedales, and Huchuan Lu. Deep Mutual Learning. CVPR, 2017. 2, 6" + ], + "bbox": [ + 501, + 92, + 890, + 882 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "8092", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + } +] \ No newline at end of file diff --git a/2023/CaPriDe Learning_ Confidential and Private Decentralized Learning Based on Encryption-Friendly Distillation Loss/ca732fe2-4161-4d55-be24-51fc46dde867_model.json b/2023/CaPriDe Learning_ Confidential and Private Decentralized Learning Based on Encryption-Friendly Distillation Loss/ca732fe2-4161-4d55-be24-51fc46dde867_model.json new file mode 100644 index 0000000000000000000000000000000000000000..8dda5f5619070478b00b7e01da8d9a4551d7cd7c --- /dev/null +++ b/2023/CaPriDe Learning_ Confidential and Private Decentralized Learning Based on Encryption-Friendly Distillation Loss/ca732fe2-4161-4d55-be24-51fc46dde867_model.json @@ -0,0 +1,1835 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.13, + 0.88, + 0.177 + ], + "angle": 0, + "content": "CaPriDe Learning: Confidential and Private Decentralized Learning based on Encryption-friendly Distillation Loss" + }, + { + "type": "text", + "bbox": [ + 0.236, + 0.204, + 0.734, + 0.257 + ], + "angle": 0, + "content": "Nurbek Tastan Karthik Nandakumar \nMohamed bin Zayed University of Artificial Intelligence, UAE \n{nurbek.tastan,karthik.nandakumar}@mbzuai.ac.ae" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.292, + 0.314, + 0.308 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.326, + 0.473, + 0.794 + ], + "angle": 0, + "content": "Large volumes of data required to train accurate deep neural networks (DNNs) are seldom available with any single entity. Often, privacy concerns prevent entities from sharing data with each other or with a third-party learning service provider. While cross-silo federated learning (FL) allows collaborative learning of large DNNs without sharing the data itself, most existing cross-silo FL algorithms have an unacceptable utility-privacy trade-off. In this work, we propose a framework called Confidential and Private Decentralized (CaPriDe) learning, which optimally leverages the power of fully homomorphic encryption (FHE) to enable collaborative learning without compromising on the confidentiality and privacy of data. In CaPriDe learning, participating entities release their private data in an encrypted form allowing other participants to perform inference in the encrypted domain. The crux of CaPriDe learning is mutual knowledge distillation between multiple local models through a novel distillation loss, which is an approximation of the Kullback-Leibler (KL) divergence between the local predictions and encrypted inferences of other participants on the same data that can be computed in the encrypted domain. Extensive experiments on three datasets show that CaPriDe learning can improve the accuracy of local models without any central coordination, provide strong guarantees of data confidentiality and privacy, and has the ability to handle statistical heterogeneity. Constraints on the model architecture (arising from the need to be FHE-friendly), limited scalability, and computational complexity of encrypted domain inference are the main limitations of the proposed approach. The code can be found at https://github.com/tnurbek/capride-learning." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.829, + 0.21, + 0.844 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.47, + 0.903 + ], + "angle": 0, + "content": "Rapid strides have been made in machine learning (ML) (in particular, deep learning) over the past decade. However, in many important application domains such as health-" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.293, + 0.895, + 0.581 + ], + "angle": 0, + "content": "care and finance, the absence of large, centralized datasets is a significant obstacle to realizing the full benefits of deep learning algorithms. Data in these applications often resides in silos and is governed by strict regulations (e.g., HIPAA, GDPR, etc.) because of its privacy sensitive nature [22]. Competing business interests of data owners and the lack of appropriate incentives for data sharing further accentuate the problem. To overcome these issues, there is a need for collaborative learning algorithms that ideally satisfy the following requirements [15]: (i) confidentiality - no sharing of raw data, (ii) privacy - minimal leakage of information via the knowledge exchange mechanism (e.g., gradients or predictions), (iii) utility - gain in accuracy (over the individual models) resulting from the collaboration, even in the presence of statistical heterogeneity, (iv) efficiency - minimize computational complexity and communication burden, (v) robustness - handle unintentional failures and attacks emanating from malicious entities, and (vi) fairness - utility should be proportional to the individual contributions." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.584, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Federated learning (FL) [15] is a special case of collaborative learning, which works under the orchestration of a central server. FL allows multiple entities to collaboratively solve a ML problem by sharing of focused updates (e.g., gradients), instead of raw data. Specifically, cross-silo FL (typically between 2-100 participants) has been touted as a promising solution to address the data fragmentation problem in finance [8] and healthcare [16,23]. Most prior cross-silo FL algorithms assume that all the parties are collectively training a single model with a common architecture, which is too restrictive in practice. Furthermore, knowledge exchange usually happens through sharing of gradients or model parameters. Recent gradient inversion attacks [9, 14] demonstrate that it is indeed possible to recover high fidelity information from individual gradient updates, thus violating the privacy requirement. While differential privacy (DP) [24], secure multi-party computation (MPC) [26], and trusted execution environment (TEE) [21] have been proposed as potential remedies to safeguard privacy in FL, none of the existing solutions offer an acceptable privacy-utility trade-off with reasonable efficiency. As" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8084" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.152 + ], + "angle": 0, + "content": "noted in [5], sharing of high dimensional gradients/model parameters is a fundamental privacy limitation of standard FL methods, which cannot be addressed by simply wrapping around a DP, MPC, or TEE solution." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.156, + 0.473, + 0.489 + ], + "angle": 0, + "content": "Confidential and private collaborative learning (CaPC) [7] is the only method that claims to achieve both confidentiality and privacy in a collaborative setting. In CaPC learning, each participant is able to query other parties via a private inference protocol based on 2-party computation (2PC). However, the answering parties cannot directly reveal the inference logits to the querying party because it leaks information about their local models (e.g., through model extraction attacks). To circumvent this problem, differential privacy, private aggregation of teacher ensembles, and secure argmax computation through a central privacy guardian (PG) are employed to output only the predicted label to the querying party. A drawback of the CaPC approach is that it achieves all privacy guarantees using the help of a semi-trusted PG. Moreover, the use of differential privacy reduces the performance of FL algorithms, especially if strong privacy guarantees are required. Finally, the use of MPC requires multiple rounds of communication between the parties for each query. All other approaches that have been proposed to achieve knowledge transfer through distillation [18] require a non-private labeled/unlabeled/synthetic dataset that can be shared among participants." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.494, + 0.473, + 0.902 + ], + "angle": 0, + "content": "In this work, we propose a new protocol called Confidential and Private Decentralized (CaPriDe) learning, where participants learn from each other in a collaborative setting while preserving their confidentiality and privacy (see Figure 1). Unlike the 2PC protocols used in CaPC, we leverage fully homomorphic encryption (FHE) to enable participants to publish their encrypted data. Since the published data is encrypted using the data owner's public key and only the owner can decrypt the data (using the secret key), confidentiality is preserved. However, the CaPriDe framework allows other collaborators to perform encrypted inference on the published (encrypted) data by applying their own local model. Mutual knowledge distillation (KD) [13] between the data owner's local model and the local models of the collaborators is used to transfer knowledge between models and ensure a collaborative gain. KD is typically achieved through minimizing the distillation loss between the student and teacher responses (logits). Since the collaborators in CaPriDe learning make predictions on encrypted data, a loss function that can be computed without any decryption is required. Hence, we propose a new distillation loss, which is an approximation of the KL divergence that can be securely computed. Only an encrypted value of the distillation loss aggregated over an entire batch is sent back to the data owner, which ensures strong privacy. Data owners can decrypt this aggregate loss value and use it for updating their local models. Our contributions are:" + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.092, + 0.892, + 0.168 + ], + "angle": 0, + "content": "- We introduce the CaPriDe learning framework, which exploits FHE-based encrypted inference and knowledge distillation to achieve confidential and private collaborative learning without any central orchestration and any need for non-private shared data." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.18, + 0.892, + 0.255 + ], + "angle": 0, + "content": "- To enable CaPriDe learning, we propose an encryption-friendly distillation loss that estimates the approximate KL divergence between two model predictions and design a protocol to securely compute this loss in the encrypted domain." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.268, + 0.892, + 0.343 + ], + "angle": 0, + "content": "- We conduct extensive experiments to show that CaPriDe learning enables participants to achieve collaborative gain, even in the non-iid setting. To prove feasibility, we implement encrypted inference using the Tile Tensors library with a FHE-friendly model." + }, + { + "type": "list", + "bbox": [ + 0.518, + 0.092, + 0.892, + 0.343 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.359, + 0.628, + 0.376 + ], + "angle": 0, + "content": "2. Background" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.385, + 0.892, + 0.747 + ], + "angle": 0, + "content": "Federated learning (FL) was first proposed in [20] as a distributed ML algorithm that does not use user data directly. Subsequent works attempted to address various challenges in FL, including privacy [7, 19], communication efficiency [12], convergence, catastrophic forgetting [25], and secure aggregation [3]. Though several privacy-enhancing solutions have been proposed, most of them fail to provide a satisfactory balance between accuracy and privacy. Only recently, some promising solutions such as CaPC have been proposed [7]. These advancements have been made by jettisoning the conventional gradient sharing approach and borrowing ideas from the knowledge distillation literature [13]. Specifically, it has been shown that it is possible for an ML model to benefit from predictions made by other models, by aligning the local predictions with those obtained from external parties. Deep mutual learning (DML) [27] is an important work in this direction, where models learn collaboratively and teach each other throughout the training process. Compared to the distillation by a pre-defined static teacher network, DML achieves better performance, and mutually learned models outperform independently trained models irrespective of model capacity. Knowledge Distillation via Collaborative Learning (KDCL) [11], Cronus [5], and Ensemble Distillation [18] also follow a similar idea." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.758, + 0.783, + 0.775 + ], + "angle": 0, + "content": "2.1. Fully Homomorphic Encryption" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.782, + 0.892, + 0.828 + ], + "angle": 0, + "content": "Let \\(\\mathcal{E}(x)\\) denote the encrypted value of \\(x\\) using the public key \\(\\kappa\\). The encryption scheme is called homomorphic over an operation \\(\\oplus\\) if it supports the following equation:" + }, + { + "type": "equation", + "bbox": [ + 0.566, + 0.841, + 0.825, + 0.857 + ], + "angle": 0, + "content": "\\[\n\\mathcal {E} \\left(x _ {1} \\oplus x _ {2}\\right) = \\mathcal {E} \\left(x _ {1}\\right) \\otimes \\mathcal {E} \\left(x _ {2}\\right), \\forall x _ {1}, x _ {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.871, + 0.892, + 0.901 + ], + "angle": 0, + "content": "where operators \\(\\otimes\\) and \\(\\oplus\\) need not be the same. The scheme is fully homomorphic if for some arbitrary function" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8085" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.16, + 0.094, + 0.813, + 0.316 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.332, + 0.895, + 0.404 + ], + "angle": 0, + "content": "Figure 1. Illustration of confidential and private decentralized (CaPriDe) learning framework for \\( K \\) participants, showing the learning process for only participant \\( \\mathcal{P}_1 \\). Here, \\( \\mathcal{D}_1 = (X_1, Y_1) = \\{x_{j,1}, y_{j,1}\\}_{j=1}^{N_1} \\) is the local data of \\( \\mathcal{P}_1 \\), \\( \\mathcal{E}(X_1) \\) denotes the collection of encrypted unlabeled samples of \\( \\mathcal{P}_1 \\), and \\( \\pmb{p}_{(1,l)} \\) and \\( \\pmb{z}_{(1,l)} \\) are the prediction probabilities and logits obtained by applying model \\( \\mathcal{M}_{\\theta_l} \\) of participant \\( \\mathcal{P}_l \\) to \\( X_1 \\). Black dashed line represents exchange of encrypted data between participants in each round, and blue dashed line denotes a single transfer at the beginning of the protocol. \\( \\mathcal{L}_{CE} \\) and \\( \\mathcal{L}_{DL} \\) represent the cross-entropy and distillation losses, respectively." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.429, + 0.471, + 0.459 + ], + "angle": 0, + "content": "\\(f(x_{1},x_{2},\\ldots ,x_{n})\\) , there exists a function \\(g\\) and decryption method \\(^d\\) such that:" + }, + { + "type": "equation", + "bbox": [ + 0.104, + 0.467, + 0.443, + 0.485 + ], + "angle": 0, + "content": "\\[\nf \\left(x _ {1}, x _ {2}, \\dots , x _ {n}\\right) = d \\left(g \\left(\\mathcal {E} \\left(x _ {1}\\right), \\mathcal {E} \\left(x _ {2}\\right), \\dots , \\mathcal {E} \\left(x _ {n}\\right)\\right)\\right),\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.492, + 0.47, + 0.584 + ], + "angle": 0, + "content": "where \\(d\\) uses secret key \\(\\tilde{\\kappa}\\) to decrypt the given ciphertext value. Gentry [10] first constructed somewhat homomorphic encryption (SWHE) and later developed bootstrapping of ciphertexts to reduce noise. Among the various FHE schemes such as BGV, BFV, and CKKS [6], only CKKS works with real numbers and approximate computations." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.59, + 0.218, + 0.605 + ], + "angle": 0, + "content": "2.2. Preliminaries" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.613, + 0.47, + 0.751 + ], + "angle": 0, + "content": "Typically, a supervised classifier \\(\\mathcal{M}\\) is mapping from the input space \\(\\pmb{x} \\in \\mathcal{X} \\subset \\mathbb{R}^n\\) to the label space \\(y \\in \\mathcal{Y} = \\{1,2,\\dots,M\\}\\), where \\(M\\) is the number of classes. For convenience, we assume that \\(\\mathcal{M}: \\mathcal{X} \\to \\mathcal{Z}\\), where \\(\\mathcal{Z}\\) denotes the logits space (\\(\\mathcal{Z} \\subset \\mathbb{R}^M\\)). In other words, \\(\\mathcal{M}\\) maps the input \\(\\pmb{x}\\) to a \\(M\\)-dimensional logits vector \\(\\pmb{z} = [z^{1},z^{2},\\dots,z^{M}]\\). A softmax function \\((\\sigma)\\) can be applied to the logits vector \\(\\pmb{z}\\) to obtain the probability distribution \\(\\pmb{p} = \\sigma(\\pmb{z}) = [p^{1},p^{2},\\dots,p^{M}]\\) over \\(M\\) labels, i.e.," + }, + { + "type": "equation", + "bbox": [ + 0.205, + 0.757, + 0.469, + 0.787 + ], + "angle": 0, + "content": "\\[\np ^ {m} = \\frac {\\exp (z ^ {m} / T)}{C}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.793, + 0.471, + 0.904 + ], + "angle": 0, + "content": "where \\(T\\) is the softmax temperature parameter and \\(C = \\sum_{i=1}^{M} \\exp(z^i / T)\\) is a normalization constant that ensures \\(\\sum_{m=1}^{M} p^m = 1\\). Here, \\(p^m\\) denotes the probability that input \\(\\mathbf{x}\\) belongs to class \\(m, m = 1, 2, \\dots, M\\). Given two \\(V\\)-dimensional vectors \\(\\mathbf{v}_1, \\mathbf{v}_2 \\in \\mathbb{R}^V\\), let \\((\\mathbf{v}_1 \\cdot \\mathbf{v}_2)\\) denote the dot product, \\((\\mathbf{v}_1 \\odot \\mathbf{v}_2)\\) denote the element-wise product, and \\(f(\\mathbf{v}_1)\\) represent the vector obtained by applying" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.429, + 0.892, + 0.46 + ], + "angle": 0, + "content": "the function \\( f \\) to each element in \\( \\pmb{v}_1 \\). The cross entropy loss of model \\( \\mathcal{M} \\) on a sample \\( (x, y) \\) is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.602, + 0.467, + 0.791, + 0.484 + ], + "angle": 0, + "content": "\\[\nL _ {C E} (\\boldsymbol {p}, y) = - \\mathbf {1} _ {y} \\cdot l o g (\\boldsymbol {p}),\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.49, + 0.892, + 0.537 + ], + "angle": 0, + "content": "where \\(\\pmb{p} = \\sigma(\\mathcal{M}(\\pmb{x}))\\) and \\(\\mathbf{1}_y\\) is an indicator vector whose \\(y^{th}\\) element is 1 and other elements are 0. The KL divergence between two distributions \\(\\pmb{p}_1\\) and \\(\\pmb{p}_2\\) is given by:" + }, + { + "type": "equation", + "bbox": [ + 0.5, + 0.543, + 0.909, + 0.596 + ], + "angle": 0, + "content": "\\[\nD _ {K L} \\left(\\boldsymbol {p} _ {2} \\| \\boldsymbol {p} _ {1}\\right) = \\sum_ {m = 1} ^ {M} p _ {2} ^ {m} \\log \\frac {p _ {2} ^ {m}}{p _ {1} ^ {m}} = \\left(\\boldsymbol {p} _ {2} \\cdot \\log \\boldsymbol {p} _ {2}\\right) - \\left(\\boldsymbol {p} _ {2} \\cdot \\log \\boldsymbol {p} _ {1}\\right). \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.598, + 0.871, + 0.613 + ], + "angle": 0, + "content": "The L2 distance between two logit vectors \\( z_{1} \\) and \\( z_{2} \\) is:" + }, + { + "type": "equation", + "bbox": [ + 0.5, + 0.619, + 0.892, + 0.66 + ], + "angle": 0, + "content": "\\[\nD _ {L 2} \\left(\\boldsymbol {z} _ {2}, \\boldsymbol {z} _ {1}\\right) = \\sum_ {m = 1} ^ {M} \\left(z _ {2} ^ {m} - z _ {1} ^ {m}\\right) ^ {2} = \\left(\\boldsymbol {z} _ {2} - \\boldsymbol {z} _ {1}\\right) \\cdot \\left(\\boldsymbol {z} _ {2} - \\boldsymbol {z} _ {1}\\right).\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.667, + 0.707, + 0.684 + ], + "angle": 0, + "content": "3. Proposed Framework" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.691, + 0.686, + 0.706 + ], + "angle": 0, + "content": "3.1. Problem Statement" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.714, + 0.892, + 0.857 + ], + "angle": 0, + "content": "Suppose that there are \\(K\\) participants \\(\\mathcal{P}_1, \\mathcal{P}_2, \\dots, \\mathcal{P}_K\\). Each party \\(\\mathcal{P}_k, k = 1, 2, \\dots, K\\) has its own local training dataset \\(\\mathcal{D}_k = \\{\\pmb{x}_{j,k}, y_{j,k}\\}_{j=1}^{N_k}\\) of size \\(N_k\\) using which it learns its own local ML model \\(\\mathcal{M}_{\\theta_k}\\), where \\(\\mathcal{M}\\) denotes the architecture and \\(\\theta\\) represents the parameter set. We also assume that each party \\(\\mathcal{P}_k\\) has its own validation set \\(\\tilde{\\mathcal{D}}_k = \\{\\tilde{\\pmb{x}}_{j,k}, \\tilde{y}_{j,k}\\}_{j=1}^{\\tilde{N}_k}\\) of size \\(\\tilde{N}_k\\) and the validation accuracy of model \\(\\mathcal{M}_{\\theta_k}\\) on \\(\\tilde{\\mathcal{D}}_k\\) is \\(\\mathcal{A}_k\\). The goal of each participant is to collaborate with other parties in order to improve" + }, + { + "type": "page_footnote", + "bbox": [ + 0.499, + 0.863, + 0.892, + 0.901 + ], + "angle": 0, + "content": "1For simplicity of notation, we use the same symbol \\(\\mathcal{M}\\) to denote the model architecture of all the \\(K\\) participants. The proposed framework does not require all the parties to have the same model architecture." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8086" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.218 + ], + "angle": 0, + "content": "the accuracy of its local model, i.e., obtain a better model \\(\\mathcal{M}_{\\theta_k^*}\\) that has a higher validation accuracy \\(\\mathcal{A}_k^*\\) on \\(\\tilde{\\mathcal{D}}_k\\) compared to \\(\\mathcal{A}_k\\), without compromising on the confidentiality and privacy of its local data \\(\\mathcal{D}_k\\). In other words, no participant \\(\\mathcal{P}_l, l \\neq k\\) can access \\(\\mathcal{D}_k\\) in its plaintext form or learn any private information about \\(\\mathcal{D}_k\\) such as the \\(x_{j,k}\\) and \\(y_{j,k}\\) values. The utility of the framework is measured as the average collaboration gain, i.e., \\(U = \\frac{1}{K} \\sum_{k=1}^{K} (\\mathcal{A}_k^* - \\mathcal{A}_k)\\)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.222, + 0.445, + 0.239 + ], + "angle": 0, + "content": "3.2. CaPriDe Learning Framework Description" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.244, + 0.474, + 0.457 + ], + "angle": 0, + "content": "Let \\(\\mathcal{E}(\\pmb{X}_k) = \\{\\mathcal{E}(\\pmb{x}_{j,k})\\}_{j=1}^{N_k}\\) be the collection of encrypted input samples of participant \\(\\mathcal{P}_k\\). Here, the encryption is based on a FHE scheme with public key \\(\\kappa_k\\) of \\(\\mathcal{P}_k\\). In CaPriDe learning, each party initiates the collaboration by publishing \\(\\mathcal{E}(\\pmb{X}_k)\\) to the other participants. We now describe how the learning proceeds at participant \\(\\mathcal{P}_k\\). Let \\(\\pmb{z}_{j,(k,l)}\\) and \\(\\pmb{p}_{j,(k,l)}\\) denote the logits vector and probability distribution obtained when model \\(\\mathcal{M}_{\\theta_l}\\) of \\(\\mathcal{P}_l\\) is applied on the input sample \\(\\pmb{x}_{j,k}\\), i.e., the \\(j^{th}\\) training sample of \\(\\mathcal{P}_k\\). Assume that we are at round \\((t+1), 0 \\leq t < \\tau\\), of the collaboration and the current model of \\(\\mathcal{P}_k\\) is \\(\\mathcal{M}_{\\theta_k^t}\\). \\(\\mathcal{P}_k\\) performs one forward pass on its own training set \\(\\mathcal{D}_k\\) to obtain the predictions \\(\\{\\pmb{p}_{j,(k,k)}\\}_{j=1}^{N_k}\\). The cross-entropy loss of party \\(\\mathcal{P}_k\\) can be obtained as:" + }, + { + "type": "equation", + "bbox": [ + 0.155, + 0.464, + 0.47, + 0.506 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {C E _ {k}} = \\sum_ {j = 1} ^ {N _ {k}} L _ {C E} \\left(\\boldsymbol {p} _ {j, (k, k)}, y _ {j, k}\\right). \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.512, + 0.47, + 0.633 + ], + "angle": 0, + "content": "To enable knowledge transfer between participants, we make use of the knowledge distillation (KD) approach [13], where a student model learns to mimic the predictions of a teacher model. In CaPriDe learning, there is no designated teacher model, and mutual KD between multiple peer models is used for knowledge transfer. Let \\(\\mathcal{L}_{DL_{(k,l)}}\\) denote the pairwise distillation loss between predictions of \\(\\mathcal{P}_k\\) and \\(\\mathcal{P}_l\\), which is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.136, + 0.637, + 0.47, + 0.68 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {D L _ {(k, l)}} = \\sum_ {j = 1} ^ {N _ {k}} D _ {D L} \\left(\\boldsymbol {p} _ {j, (k, k)}, \\boldsymbol {p} _ {j, (k, l)}\\right), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.687, + 0.469, + 0.732 + ], + "angle": 0, + "content": "where \\(D_{DL}(\\pmb{p}_1, \\pmb{p}_2)\\) denotes the mimic distance between two predictions \\(\\pmb{p}_1\\) and \\(\\pmb{p}_2\\). The total distillation loss for \\(\\mathcal{P}_k\\) can be computed as:" + }, + { + "type": "equation", + "bbox": [ + 0.161, + 0.738, + 0.469, + 0.781 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {D L _ {k}} = \\frac {1}{K - 1} \\sum_ {l = 1, l \\neq k} ^ {K} \\mathcal {L} _ {D L _ {(k, l)}}. \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.787, + 0.374, + 0.803 + ], + "angle": 0, + "content": "Therefore, the overall loss of \\(\\mathcal{P}_k\\) is given by:" + }, + { + "type": "equation", + "bbox": [ + 0.19, + 0.81, + 0.469, + 0.827 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {k} = \\mathcal {L} _ {C E _ {k}} + \\lambda_ {k} \\mathcal {L} _ {D L _ {k}}, \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.833, + 0.469, + 0.879 + ], + "angle": 0, + "content": "where \\(\\lambda_{k}\\) is the weighting factor between the cross-entropy and distillation losses for \\(\\mathcal{P}_k\\). The model of participant \\(\\mathcal{P}_k\\) is then updated through a backward pass as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.191, + 0.884, + 0.469, + 0.904 + ], + "angle": 0, + "content": "\\[\n\\theta_ {k} ^ {t + 1} = \\theta_ {k} ^ {t} - \\alpha_ {k} \\nabla_ {\\theta_ {k} ^ {t}} \\mathcal {L} _ {k}, \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.092, + 0.738, + 0.107 + ], + "angle": 0, + "content": "where \\(\\alpha_{k}\\) is the learning rate of \\(\\mathcal{P}_k\\)" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.116, + 0.819, + 0.133 + ], + "angle": 0, + "content": "3.3. Encryption-friendly Distillation Loss" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.139, + 0.894, + 0.44 + ], + "angle": 0, + "content": "The key missing component in the above CaPriDe learning framework is the computation \\(D_{DL}\\) in Eq. 4. For participant \\(\\mathcal{P}_k\\), its logit vectors \\(\\pmb{z}_{.,(k,k)}\\) and predictions \\(\\pmb{p}_{.,(k,k)}\\) are available in the plaintext domain. However, since participant \\(\\mathcal{P}_l\\) receives the input samples of \\(\\mathcal{P}_k\\) in encrypted form, it can perform inference in the encrypted domain and obtain only \\(\\mathcal{E}\\left(\\pmb{z}_{.,(k,l)}\\right)\\), where \\(\\mathcal{E}\\left(\\pmb{z}_{.,(k,l)}\\right) = \\mathcal{M}_{\\theta_l}\\left(\\mathcal{E}(\\pmb{x}_{.,k})\\right)\\). It is not straightforward to obtain \\(\\mathcal{E}\\left(\\pmb{p}_{.,(k,l)}\\right)\\) from \\(\\mathcal{E}\\left(\\pmb{z}_{.,(k,l)}\\right)\\) because it involves computing a non-linear softmax function in the encrypted domain, which is a challenging task in the FHE domain and consumes a large multiplicative depth. Even if this softmax operation is somehow implemented, it is not possible to send \\(\\mathcal{E}\\left(\\pmb{p}_{.,(k,l)}\\right)\\) back to \\(\\mathcal{P}_k\\), who can decrypt it and obtain \\(\\pmb{p}_{.,(k,l)}\\). Since \\(\\mathcal{P}_k\\) also has knowledge of \\(\\pmb{x}_{.,k}\\), it becomes feasible for \\(\\mathcal{P}_k\\) to carry out a model extraction attack based on \\((\\pmb{x}_{.,k},\\pmb{p}_{.,(k,l)})\\) pairs to glean information about model \\(\\mathcal{M}_{\\theta_l}\\). This violates the privacy requirements of participant \\(\\mathcal{P}_l\\). Hence, there is a need for a loss function that can be computed in the encrypted domain." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.441, + 0.893, + 0.505 + ], + "angle": 0, + "content": "A straightforward solution is to compute the L2 distance between logit vectors \\( \\mathbf{z}_{.,(k,k)} \\) and \\( \\mathbf{z}_{.,(k,l)} \\). In this case, participant \\( \\mathcal{P}_k \\) can publish \\( \\mathcal{E}\\left(\\mathbf{z}_{.,(k,k)}\\right) \\) and the pairwise distillation loss \\( \\mathcal{L}_{DL_{(k,l)}} \\) can be computed by participant \\( \\mathcal{P}_l \\) as:" + }, + { + "type": "equation", + "bbox": [ + 0.502, + 0.515, + 0.901, + 0.563 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\left. \\mathcal {E} \\left(D _ {D L} \\left(\\boldsymbol {p} _ {j, (k, k)}, \\boldsymbol {p} _ {j, (k, l)}\\right)\\right) \\equiv \\mathcal {E} \\left(D _ {L 2} \\left(\\boldsymbol {z} _ {j, (k, k)}, \\boldsymbol {z} _ {j, (k, l)}\\right)\\right) \\right. \\\\ = \\left(\\mathcal {E} \\left(\\boldsymbol {z} _ {j, (k, k)}\\right) - \\mathcal {E} \\left(\\boldsymbol {z} _ {j, (k, l)}\\right)\\right) \\cdot \\left(\\mathcal {E} \\left(\\boldsymbol {z} _ {j, (k, k)}\\right) - \\mathcal {E} \\left(\\boldsymbol {z} _ {j, (k, l)}\\right)\\right) \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.573, + 0.892, + 0.686 + ], + "angle": 0, + "content": "However, our experiments indicate the above simple solution based on L2 distance does not lead to high utility \\(U\\). On the other hand, we observed that using the KL divergence between predictions of \\(\\mathcal{P}_k\\) and \\(\\mathcal{P}_l\\) almost always leads to better utility. But, \\(D_{KL}\\left(\\pmb{p}_{.,(k,k)}\\| \\pmb{p}_{.,(k,l)}\\right)\\) cannot be directly computed in the encrypted domain because participant \\(\\mathcal{P}_l\\) only has access to the encrypted logits \\(\\mathcal{E}\\left(z_{.,(k,l)}\\right)\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.685, + 0.892, + 0.716 + ], + "angle": 0, + "content": "Fortunately, this problem can be solved by approximating the KL divergence. Combining Eqs. 1 and 2, we get:" + }, + { + "type": "equation", + "bbox": [ + 0.5, + 0.726, + 0.907, + 0.78 + ], + "angle": 0, + "content": "\\[\nD _ {K L} \\left(\\boldsymbol {p} _ {2} \\right\\| \\boldsymbol {p} _ {1}) = \\sum_ {m = 1} ^ {M} p _ {2} ^ {m} \\log p _ {2} ^ {m} - \\sum_ {m = 1} ^ {M} p _ {2} ^ {m} \\left(z _ {1} ^ {m} / T\\right) + M \\log C. \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.781, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Here, the first term can be denoted as \\((\\pmb{p}_2 \\cdot \\log \\pmb{p}_2)\\) and measures the entropy (uncertainty) in \\(\\pmb{p}_2\\). The second term measures the divergence (unnormized cross-entropy) between \\(\\pmb{p}_2\\) and \\(\\pmb{p}_1\\) and can be expressed as \\((\\pmb{p}_2 \\cdot \\pmb{z}_1 / T)\\). The final term \\(M \\log C\\) is just a normalization constant that does not affect the relative ordering of the class predictions. We observed that this normalization term does not appear to play a major role in model training. Hence, we ignore this value" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8087" + } + ], + [ + { + "type": "code_caption", + "bbox": [ + 0.081, + 0.092, + 0.357, + 0.107 + ], + "angle": 0, + "content": "Algorithm 1 CaPRiDe learning algorithm" + }, + { + "type": "algorithm", + "bbox": [ + 0.08, + 0.109, + 0.892, + 0.484 + ], + "angle": 0, + "content": "Input: Number of participants \\(K\\) , weight parameter \\(\\lambda_{k}\\) , learning rate \\(\\alpha_{k}\\) for each participant \nAssumption: Each participant has local training set \\(\\mathcal{D}_k = \\{\\pmb {x}_{j,k},y_{j,k}\\}_{j = [1,N_k]}\\) of size \\(N_{k}\\) , a fully homomorphic encryption function \\(\\mathcal{E}\\) with public key \\(\\kappa_{k}\\) , a decryption function \\(d\\) with secret key \\(\\tilde{\\kappa}_k\\) \nfor each participant \\(k = 1,2,\\dots ,K\\) do Train initial model \\(\\mathcal{M}_{\\theta_k^0}\\) based on local data \\(\\mathcal{D}_k\\) Publish encrypted data \\(\\{\\mathcal{E}(\\pmb{x}_{j,k})\\}_{j = [1,N_k]}\\) \nend for \nfor each round \\(t = 1,2,\\dots ,\\tau\\) do \nfor each participant \\(k = 1,2,\\dots ,K\\) do \\(\\{\\pmb {p}_{j,(k,k)}\\}_{j = [1,N_k]}\\gets \\{\\sigma (\\mathcal{M}_{\\theta_k^{t - 1}}(\\pmb {x}_{j,k}))\\}_{j = [1,N_k]}\\) \n\\(\\mathcal{L}_{CE_k}\\gets\\) Compute cross-entropy loss of participant \\(k\\) using Equation 3 \n\\(\\hat{\\mathcal{L}}_{KL_k}\\gets\\) Compute local uncertainty of participant \\(k\\) using Equation 11 \nPublish encrypted inferences \\(\\{\\mathcal{E}(\\pmb {p}_{j,(k,k)})\\}_{j = [1,N_k]}\\) \nfor each participant \\(l = 1,2,\\dots ,K,l\\neq k\\) do \\(\\{\\mathcal{E}(\\pmb {z}_{j,(k,l)})\\}_{j = [1,N_k]}\\gets \\{\\mathcal{M}_{\\theta_l^{t - 1}}(\\mathcal{E}(\\pmb {x}_{j,k}))\\}_{j = [1,N_k]}\\) (Perform inference on encrypted data) \n\\(\\mathcal{E}(\\tilde{\\mathcal{L}}_{KL_{(k,l)}})\\gets\\) Compute pairwise divergence in encrypted domain using Equation 12 \nSend \\(\\mathcal{E}(\\tilde{\\mathcal{L}}_{KL_{(k,l)}})\\) to participant \\(k\\) \n\\(\\mathcal{L}_{DL_{(k,l)}}\\gets\\) Compute pairwise distillation loss between participants \\(k\\) and \\(l\\) using Equation 10 \nend for \n\\(\\mathcal{L}_{DL_k}\\gets\\) Compute distillation loss of participant \\(k\\) using Equation 5 \n\\(\\mathcal{L}_k\\gets\\) Compute total loss of participant \\(k\\) using Equation 6 \n\\(\\theta_k^t\\gets\\) Update local model of participant \\(k\\) using Equation 7 \nend for \nend for" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.514, + 0.452, + 0.529 + ], + "angle": 0, + "content": "and define the approximate KL divergence is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.121, + 0.538, + 0.469, + 0.556 + ], + "angle": 0, + "content": "\\[\n\\hat {D} _ {K L} \\left(\\boldsymbol {p} _ {2} \\| \\boldsymbol {p} _ {1}\\right) = \\left(\\boldsymbol {p} _ {2} \\cdot \\log \\boldsymbol {p} _ {2}\\right) - \\left(\\boldsymbol {p} _ {2} \\cdot \\boldsymbol {z} _ {1} / T\\right). \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.566, + 0.469, + 0.611 + ], + "angle": 0, + "content": "Going back to Eq. 4, the pairwise distillation loss based on the above approximate KL divergence can be easily computed in the encrypted domain as:" + }, + { + "type": "equation", + "bbox": [ + 0.077, + 0.619, + 0.485, + 0.674 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\left. \\mathcal {E} \\left(D _ {D L} \\left(\\boldsymbol {p} _ {j, (k, k)}, \\boldsymbol {p} _ {j, (k, l)}\\right)\\right) \\equiv \\mathcal {E} \\left(\\hat {D} _ {K L} \\left(\\boldsymbol {p} _ {j, (k, k)} | | \\boldsymbol {p} _ {j, (k, l)}\\right)\\right) \\right. \\\\ = \\mathcal {E} \\left(\\boldsymbol {p} _ {j, (k, k)} \\cdot \\log \\boldsymbol {p} _ {j, (k, k)}\\right) - \\mathcal {E} \\left(\\boldsymbol {p} _ {j, (k, k)}\\right) \\cdot \\mathcal {E} \\left(\\boldsymbol {z} _ {j, (k, l)}\\right) / T \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.682, + 0.469, + 0.757 + ], + "angle": 0, + "content": "The above distillation loss based on the approximate KL divergence is encryption-friendly because it involves only linear operations and scaling by a constant plaintext value \\(T\\). In practice, our goal is to compute \\(\\mathcal{L}_{DL_{(k,l)}}\\) at participant \\(\\mathcal{P}_k\\). Therefore, Eq. 4 can be further reformulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.139, + 0.766, + 0.469, + 0.793 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {D L _ {(k, l)}} = \\hat {\\mathcal {L}} _ {K L _ {k}} - d \\left(\\mathcal {E} \\left(\\tilde {\\mathcal {L}} _ {K L _ {(k, l)}}\\right)\\right), \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.802, + 0.122, + 0.814 + ], + "angle": 0, + "content": "where" + }, + { + "type": "equation", + "bbox": [ + 0.154, + 0.812, + 0.469, + 0.855 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathcal {L}} _ {K L _ {k}} = \\sum_ {j = 1} ^ {N _ {k}} \\left(\\boldsymbol {p} _ {j, (k, k)} \\cdot \\log \\boldsymbol {p} _ {j, (k, k)}\\right), \\tag {11}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.093, + 0.862, + 0.469, + 0.904 + ], + "angle": 0, + "content": "\\[\n\\mathcal {E} \\left(\\tilde {\\mathcal {L}} _ {K L _ {(k, l)}}\\right) = \\sum_ {j = 1} ^ {N _ {k}} \\mathcal {E} \\left(\\boldsymbol {p} _ {j, (k, k)}\\right) \\cdot \\mathcal {E} \\left(\\boldsymbol {z} _ {j, (k, l)}\\right) / T. \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.513, + 0.892, + 0.653 + ], + "angle": 0, + "content": "Since \\(\\mathcal{P}_k\\) has access to \\(\\pmb{p}_{j,(k,k)}\\), it is straightforward to compute \\(\\hat{\\mathcal{L}}_{KL_k}\\) in plaintext space. If participant \\(\\mathcal{P}_k\\) can publish \\(\\left\\{\\mathcal{E}(\\pmb{p}_{j,(k,k)})\\right\\}_{j=1}^{N_k}\\) in each round, \\(\\mathcal{P}_l\\) can compute \\(\\mathcal{E}\\left(\\tilde{\\mathcal{L}}_{KL(k,l)}\\right)\\) in the encrypted domain because it already has access to \\(\\left\\{\\mathcal{E}\\left(z_{j,(k,l)}\\right)\\right\\}_{j=1}^{N_k}\\). The encrypted value of \\(\\tilde{\\mathcal{L}}_{KL(k,l)}\\) is sent back to participant \\(\\mathcal{P}_k\\), who decrypts it to compute the overall loss via Eqs. 10, 5 and 6." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.659, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Privacy Analysis: In the CaPriDe learning framework, the only quantity that can potentially leak private information is the pairwise distillation loss between participants \\(\\tilde{\\mathcal{L}}_{KL_{(k,l)}}\\), which has already been aggregated over all the \\(N_{k}\\) samples. A curious participant \\(\\mathcal{P}_k\\) attempting to reconstruct the logit values \\(z_{j,(k,l)}^m\\) of participant \\(\\mathcal{P}_l\\) needs to dis-aggregate this sum and compute per-sample loss, which is infeasible in the absence of other constraints. Hence, the proposed CaPriDe learning framework is privacy-preserving. It must be emphasized that even gradient inversion attacks [9, 14] are significantly harder when the batch size is large (above 16) due to the same reason. The proposed method also has other advantages that enhance privacy: (i) A curious participant cannot query its collaborator adaptively. The encrypted samples are published once at the beginning and do not change. (ii) The responding participant can choose a" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.515, + 0.956 + ], + "angle": 0, + "content": "8088" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.228 + ], + "angle": 0, + "content": "random subset of samples in each round to compute the aggregate loss. This can improve efficiency as well as provide a privacy benefit by eliminating correlation across multiple rounds. (iii) We can incorporate differential privacy (DP) by adding noise to the aggregate pairwise distillation loss. Since the framework allows exchange of only encrypted input samples and encrypted predictions, which never get decrypted during the execution of the protocol, the CaPriDe learning approach also preserves data confidentiality." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.243, + 0.21, + 0.26 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.268, + 0.181, + 0.283 + ], + "angle": 0, + "content": "4.1. Datasets" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.291, + 0.47, + 0.337 + ], + "angle": 0, + "content": "CIFAR-10 [17] is a dataset consisting of 60000 \\(32 \\times 32\\) RGB images from 10 classes, with 6000 images per class. It has 50000 training and 10000 test samples." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.337, + 0.469, + 0.444 + ], + "angle": 0, + "content": "CIFAR-100 dataset [17] is similar to the CIFAR-10 dataset, but it has 100 classes containing 600 samples each. Training and evaluation sets are split in the same way as in CIFAR-10. CIFAR-100 dataset is used with data augmentation to account for the problem complexity. The data augmentations employed include random rotation (up to 15 degrees), random crop, and horizontal flip." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.444, + 0.469, + 0.505 + ], + "angle": 0, + "content": "HAM10000 is a multi-class imbalanced dataset comprising of 10,015 dermoscopic images from diverse populations with 7 categories of pigmented lesions. We randomly perform an \\(85\\% / 15\\%\\) split for training/ testing." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.515, + 0.267, + 0.531 + ], + "angle": 0, + "content": "4.2. Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.539, + 0.473, + 0.903 + ], + "angle": 0, + "content": "We compare the CaPriDe learning framework against the following baselines: vanilla FL based on FedAvg (FedAvg) [20], FL with local differential privacy (FedAvg + DP) [24], and deep mutual learning (DML) [27]. We use ResNet-18 architecture for all our experiments. Stochastic Gradient Descent (SGD) algorithm with momentum is used as the optimizer. For all methods, 25 epochs of local training happens before 75 rounds of collaboration. Momentum is set to 0.9, the initial learning rate is set to \\(\\alpha_{k} = 0.1\\) and the learning rate decays in rounds 50 and 75 by a factor \\(\\gamma = 0.1\\). The batch size is set to 128 for CIFAR-10 and CIFAR-100 and 32 for HAM10000. The weight factor \\(\\lambda_{k}\\) between cross-entropy and distillation losses is set to 50 for CIFAR-10 and CIFAR-100 and 20 for HAM10000. All algorithms are implemented using PyTorch and experiments are carried out on a single NVIDIA RTX A6000 GPU. Additionally, in CaPriDe and DML, the logits are scaled by a temperature parameter \\(T = 5.0\\). Finally, for FedAvg + DP, the Gaussian noise mechanism is used with \\(\\sigma_{g} = 0.01\\). To implement an FHE-compatible version of ResNet-18, we use the Tile Tensors framework [1,2], which in turn relies on low-level FHE libraries such as HElib and SEAL. For encrypted inference, ReLU is replaced with a polynomial approximation of Gaussian Error Linear Unit" + }, + { + "type": "image", + "bbox": [ + 0.508, + 0.091, + 0.889, + 0.371 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.384, + 0.892, + 0.426 + ], + "angle": 0, + "content": "Figure 2. Accuracy trend of two \\((K = 2)\\) collaboratively trained models on different datasets (rows) and for different partition strategies (columns)." + }, + { + "type": "equation", + "bbox": [ + 0.5, + 0.453, + 0.831, + 0.469 + ], + "angle": 0, + "content": "\\[\n(g (x) = 0. 0 7 1 1 + 0. 5 x + 0. 2 5 7 6 x ^ {2} - 0. 0 1 2 8 x ^ {4}).\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.47, + 0.892, + 0.667 + ], + "angle": 0, + "content": "We consider three types of strategies to partition the training dataset among the participants: (i) Homogeneous: each participant has an equal number of samples per class, (ii) Heterogeneous: each participant has an unequal number of samples determined randomly (both total number and number of samples per class), (iii) Non-overlapping class distribution: (denoted as \"no class overlap\") each participant has samples from a non-overlapping subset of classes (e.g., for \\( K = 2 \\), participants get even and odd classes). While homogeneous partitioning ensures that data distributions of participants are iid, heterogeneous split simulates typical non-iid scenarios, and the no class overlap case corresponds to extreme statistical heterogeneity." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.676, + 0.779, + 0.692 + ], + "angle": 0, + "content": "4.3. Collaborative Learning Results" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.7, + 0.892, + 0.82 + ], + "angle": 0, + "content": "To begin with, we establish the utility (collaborative gain) of the proposed CaPriDe learning framework for the case of \\( K = 2 \\) participants on all three datasets and all three partition strategies. The results of these experiments are shown in Figure 2. In each case, the collaboration gain of CaPriDe learning is positive and the actual collaboration gain values can be found in Table 1. We make the following observations based on these results." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.835, + 0.892, + 0.865 + ], + "angle": 0, + "content": "- CaPriDe learning works not only in the iid setting but also equally well in the two non iid settings. This" + }, + { + "type": "page_footnote", + "bbox": [ + 0.5, + 0.875, + 0.892, + 0.901 + ], + "angle": 0, + "content": "2Since HAM10000 already has a massive class imbalance and an odd number of class, we omit the no class overlap setting for this dataset." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8089" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.08, + 0.089, + 0.468, + 0.328 + ], + "angle": 0, + "content": "
DatasetSettingKFedAvgFedAvg+DPCaPriDe
CIFAR-10Homogeneous23.46-20.541.58
CIFAR-10Homogeneous59.22-11.83.55
CIFAR-10Homogeneous1015.442.585.52
CIFAR-10Heterogeneous26.12-17.040.93
CIFAR-10Heterogeneous513.78-7.845.06
CIFAR-10Heterogeneous1021.815.696.20
CIFAR-10No class overlap229.8511.088.16
CIFAR-100Homogeneous27.08-30.974.19
CIFAR-100Homogeneous522.68-13.819.10
CIFAR-100Homogeneous1026.10-0.1111.69
CIFAR-100Heterogeneous29.59-28.896.28
CIFAR-100Heterogeneous522.50-13.269.23
CIFAR-100Heterogeneous1034.383.8610.68
CIFAR-100No class overlap219.405.147.75
HAM10000Homogeneous20.970.481.81
HAM10000Heterogeneous21.93-1.211.67
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.338, + 0.471, + 0.395 + ], + "angle": 0, + "content": "Table 1. Collaborative gain (\\%) on all datasets and all partition strategies for different number of participants \\( K \\) using different algorithms (columns 4-6). DML results are not reported here because there is negligible difference between CaPriDe and DML." + }, + { + "type": "text", + "bbox": [ + 0.109, + 0.422, + 0.468, + 0.453 + ], + "angle": 0, + "content": "demonstrates that knowledge distillation is reasonably effective even under extreme statistical heterogeneity." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.465, + 0.47, + 0.587 + ], + "angle": 0, + "content": "- There is a negligible difference between the performance of DML (which uses true KL divergence) and CaPriDe learning (which uses approximate KL divergence) in all the settings. This clearly justifies the dropping of the normalization (third) term in Eq. 8 to obtain the encryption-friendly distillation loss in Eq. 9. While DML offers no confidentiality or privacy, CaPriDe learning offers both without any utility cost." + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.599, + 0.47, + 0.751 + ], + "angle": 0, + "content": "- In almost all the settings, there is a huge difference in the collaboration gain between FedAvg and CaPriDe learning. This is unsurprising because FedAvg involves sharing and aggregation of full model parameters leading to high utility at a significant cost to privacy. On the other hand, FedAvg + DP provides strong privacy, but this comes with a significant degradation in utility. Hence, it can be argued that CaPriDe learning achieves a good trade-off between privacy and utility compared to standard FL algorithms." + }, + { + "type": "list", + "bbox": [ + 0.096, + 0.465, + 0.47, + 0.751 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Next, we evaluate the utility of the proposed method with the different number of participants \\( K = 2,5,10 \\) on CIFAR-10 and CIFAR-100 datasets based on the homogeneous partition (iid) setting. From Figure 3 and Table 1 we observe that the collaboration gain increases with the number of participants. This is to be expected because the training samples per participant become less when there are more participants, consequently resulting in lower individual accuracy." + }, + { + "type": "image", + "bbox": [ + 0.507, + 0.091, + 0.889, + 0.282 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.294, + 0.892, + 0.323 + ], + "angle": 0, + "content": "Figure 3. Accuracy with different number of participants \\( K \\) based on the homogeneous partition (iid) setting." + }, + { + "type": "image", + "bbox": [ + 0.502, + 0.34, + 0.631, + 0.435 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.518, + 0.438, + 0.614, + 0.451 + ], + "angle": 0, + "content": "(a) Homogeneous" + }, + { + "type": "image", + "bbox": [ + 0.633, + 0.34, + 0.761, + 0.435 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.647, + 0.438, + 0.746, + 0.451 + ], + "angle": 0, + "content": "(b) Heterogeneous" + }, + { + "type": "image", + "bbox": [ + 0.764, + 0.34, + 0.892, + 0.435 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.788, + 0.438, + 0.866, + 0.451 + ], + "angle": 0, + "content": "(c) No overlap" + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.463, + 0.892, + 0.505 + ], + "angle": 0, + "content": "Figure 4. Per-class accuracy of CaPriDe learning evaluated on CIFAR-10 averaged across 5 participants. Dashed lines represent mean accuracy and error bars indicate standard deviation." + }, + { + "type": "image", + "bbox": [ + 0.502, + 0.522, + 0.63, + 0.619 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.518, + 0.623, + 0.614, + 0.635 + ], + "angle": 0, + "content": "(a) Homogeneous" + }, + { + "type": "image", + "bbox": [ + 0.633, + 0.524, + 0.759, + 0.619 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.647, + 0.623, + 0.746, + 0.635 + ], + "angle": 0, + "content": "(b) Heterogeneous" + }, + { + "type": "image", + "bbox": [ + 0.764, + 0.524, + 0.891, + 0.619 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.788, + 0.623, + 0.866, + 0.635 + ], + "angle": 0, + "content": "(c) No overlap" + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.647, + 0.892, + 0.674 + ], + "angle": 0, + "content": "Figure 5. Per-participant accuracy of CaPriDe learning evaluated on CIFAR-100 dataset." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.705, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Furthermore, we evaluate the class-wise and participant-wise accuracy of our CaPriDe learning algorithm with \\( K = 5 \\) participants. Figure 4 shows class-wise accuracy averaged across 5 participants for each setting. From this figure, it is clear that almost all the classes benefit from collaboration, and the per-class accuracy of the individual models after collaboration has a lower standard deviation compared to the case without collaboration, especially in homogeneous and heterogeneous settings. Thus, in these two settings, the individual models become similar to each other after collaboration proving the successful transfer of knowledge between participants. Note that Figure 4 displays the average top-1 accuracy of the participants. The collaboration gains" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8090" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.08, + 0.092, + 0.468, + 0.272 + ], + "angle": 0, + "content": "in the no class overlap case are significantly higher when top-5 accuracy is considered. Finally, Figure 5 illustrates participant-wise accuracy over all classes for each setting. It is obvious that each participant has a positive collaboration gain, which is the eventual goal that we set out to achieve. Moreover, all the 5 participant models have a similar accuracy at the end of the collaboration, once again demonstrating good knowledge transfer. Other results such as the sensitivity of the proposed method to weight factor \\(\\lambda_{k}\\), the sensitivity of FedAvg + DP to noise variance, and the comparison between approximate KL divergence and L2 loss are included in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.08, + 0.284, + 0.337, + 0.299 + ], + "angle": 0, + "content": "4.4. FHE Implementation Results" + }, + { + "type": "text", + "bbox": [ + 0.08, + 0.308, + 0.468, + 0.533 + ], + "angle": 0, + "content": "For FHE implementation using the Tile Tensors framework, Table 2 summarizes the information about all three datasets at the same security level of 128 bits. Two critical metrics to be kept in mind are the time required for encrypted inference and the memory consumed by a ciphertext. While encrypted inference time determines the computational complexity and latency involved in each round, the ciphertext size determines the communication burden. Note that encrypted inference can be performed on a persample basis or a batch of samples [4]. The single sample inference mode is preferred for machine learning as-a-service applications, where low latency is a critical requirement. However, the batched inference is better suited for the CaPriDe learning framework, where only the overall time required to complete inference on all the samples matter." + }, + { + "type": "table", + "bbox": [ + 0.079, + 0.546, + 0.468, + 0.649 + ], + "angle": 0, + "content": "
Security LevelCIFAR-10CIFAR-100HAM10000
128128128
Number of slots163841638416384
Time taken to encrypt one sample90 ms103 ms619 ms
Ciphertext size of one sample29.101 KB29.152 KB1.359 MB
Time taken to encrypt a batch of 32 samples1.31 s1.26 s15.57 s
Encrypted inference of a batch of 32 samples110.21 s112.09 s896.12 s
" + }, + { + "type": "table_caption", + "bbox": [ + 0.08, + 0.661, + 0.468, + 0.701 + ], + "angle": 0, + "content": "Table 2. FHE configuration parameters and memory and computational requirements for the chosen datasets based on ResNet-18 architecture." + }, + { + "type": "text", + "bbox": [ + 0.08, + 0.72, + 0.468, + 0.9 + ], + "angle": 0, + "content": "The encrypted inference time for a batch of 32 samples is approximately 110s for CIFAR-10 and CIFAR-100 and 900s for HAM10000. Thus, if the local dataset of each participant has 1024 training samples, it will take around 1 hour (8 hours) for encrypted inference per participant and then computation of the pairwise distillation loss in each communication round for CIFAR-10 and CIFAR-100 (HAM10000). While this is a significant computational load, it is less of an issue in the cross-silo collaborative learning setting, especially in sectors such as finance and healthcare, where data confidentiality and privacy as well as accuracy gain are more critical. Moreover, the ciphertext" + }, + { + "type": "text", + "bbox": [ + 0.503, + 0.092, + 0.891, + 0.302 + ], + "angle": 0, + "content": "size for one sample is approximately, 30KB for CIFAR-10 and CIFAR-100 and 1.36MB for HAM10000. For the same local dataset size of 1024, the encrypted dataset will consume 30MB for CIFAR-10 and CIFAR-100 and 1.36GB for HAM10000. Fortunately, the encrypted dataset is published and needs to be downloaded only once by each participant at the beginning of the protocol. During every communication round, only the encrypted predictions and pairwise losses are exchanged. Depending on the local dataset size, the number of classes, and the ciphertext packing strategy used, these intermediate values can be packed into at most a few tens of ciphertexts (note that this does not depend on the image size). Hence, the per round communication burden of CaPriDe learning framework is relatively small." + }, + { + "type": "text", + "bbox": [ + 0.504, + 0.306, + 0.891, + 0.667 + ], + "angle": 0, + "content": "Limitations and Potential Solutions: Note that all the above complexity computations are based on the assumption of \\( K = 2 \\) participants. As \\( K \\) increases, these requirements will increase linearly for each participant bringing the scalability of the system into question. Hence, the vanilla CaPriDe learning framework is more appropriate for a small network of participants. However, there are ways to improve scalability borrowing on ideas from cross-device FL. For example, instead of computing the pairwise distillation loss with respect to each participant, a random subset of participants can be sampled in each round for distillation. Similarly, the computational complexity for each participant can be significantly reduced by performing encrypted inference only on a randomly selected subset of encrypted samples. Furthermore, these actions can be determined at the local level without any central orchestration. The need to have FHE-friendly models for encrypted inference is a key limitation, which precludes the use of very deep models within the proposed framework. One potential solution is to employ two models at each participant: one deep model with large capacity and another smaller FHE-friendly model for encrypted inference on other participants' data and perform an additional mutual distillation between these two models after each round." + }, + { + "type": "title", + "bbox": [ + 0.504, + 0.692, + 0.606, + 0.708 + ], + "angle": 0, + "content": "5. Summary" + }, + { + "type": "text", + "bbox": [ + 0.504, + 0.72, + 0.891, + 0.9 + ], + "angle": 0, + "content": "Collaborative learning without compromising on data confidentiality and privacy is a valuable tool in applications with stringent data privacy requirements. We have proposed a new framework called CaPriDe learning based on inference in the FHE domain to achieve this goal. Knowledge distillation is used to exchange useful information between models, and a new distillation loss based on approximate KL divergence has been proposed to enable secure loss computation in the encrypted domain. While the proposed approach has shown promising results, its efficiency can be further improved. Other vulnerabilities such as poisoning attacks will also be addressed in the future." + }, + { + "type": "page_number", + "bbox": [ + 0.485, + 0.947, + 0.513, + 0.957 + ], + "angle": 0, + "content": "8091" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.116, + 0.47, + 0.186 + ], + "angle": 0, + "content": "[1] Ehud Aharoni, Allon Adir, Moran Baruch, Nir Drucker, Gilad Ezov, Ariel Farkash, Lev Greenberg, Rami Masalha, Guy Moshkowish, Dov Murik, Hayim Shaul, and Omri Soceanu. HeLayers: A Tile Tensors Framework for Large Neural Networks on Encrypted Data. arXiv:2011.01805v2, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.187, + 0.472, + 0.242 + ], + "angle": 0, + "content": "[2] Ehud Aharoni, Nir Drucker, Gilad Ezov, Hayim Shaul, and Omri Soceanu. Complex encoded tile tensors: Accelerating encrypted analytics. In IEEE S&P 20 (5), pages 35-43, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.245, + 0.472, + 0.327 + ], + "angle": 0, + "content": "[3] Keith Bonawitz, Vladimir Ivanov, Ben Kreuter, Antonio Marcedone, H. Brendan McMahan, Sarvar Patel, Daniel Ramage, Aaron Segal, and Karn Seth. Practical secure aggregation for privacy preserving machine learning. ACM Conference on Computer and Communications Security (ACM CCS), 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.329, + 0.472, + 0.371 + ], + "angle": 0, + "content": "[4] Alon Brutzkus, Ran Gilad-Bachrach, and Oren Elisha. Low latency privacy preserving inference. In International Conference on Machine Learning, pages 812-821, 2019. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.373, + 0.472, + 0.428 + ], + "angle": 0, + "content": "[5] Hongyan Chang, Virat Shejwalkar, Reza Shokri, and Amir Houmansadr. Cronus: Robust and heterogeneous collaborative learning with black-box knowledge transfer. In arXiv preprint arXiv:1912.11279, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.43, + 0.472, + 0.47 + ], + "angle": 0, + "content": "[6] Jung Hee Cheon, Andrey Kim, Miran Kim, and Yongsoo Song. Homomorphic Encryption for Arithmetic of Approximate Numbers. 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.473, + 0.472, + 0.528 + ], + "angle": 0, + "content": "[7] Christopher A. Choquette-Choo, Natalie Dullerud, Adam Dziedzic, Yunxiang Zhang, Somesh Jha, Nicolas Papernot, and Xiao. Wang. CaPC Learning: Confidential and Private Collaborative Learning. Proceedings of ICLR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.53, + 0.472, + 0.613 + ], + "angle": 0, + "content": "[8] Deshpande, Nikhil and Shiffman, Gary M. Federated learning: The new thing in ai/ml for detecting financial crimes and managing risk. https://morningconsult.com/opinions/federated-learning-the-new-thing-in-ai-ml-for-detecting-financial-crimes-and-managing-risk/, 2021.1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.615, + 0.472, + 0.671 + ], + "angle": 0, + "content": "[9] Jonas Geiping, Hartmut Bauermeister, Hannah Dröge, and Michael. Moeller. Inverting gradients-How easy is it to break privacy in federated learning? Advances in Neural Information Processing Systems (NeurIPS), 33, 2020. 1, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.672, + 0.47, + 0.699 + ], + "angle": 0, + "content": "[10] Craig. Gentry. Fully Homomorphic Encryption. PhD thesis, 2009. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.701, + 0.47, + 0.756 + ], + "angle": 0, + "content": "[11] Qiushan Guo, Xinjiang Wang, Yichao Wu, Zhipeng Yu, Ding Liang, Xiaolin Hu, and Ping. Luo. Online Knowledge Distillation via Collaborative Learning. Proceedings of CVPR, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.758, + 0.47, + 0.814 + ], + "angle": 0, + "content": "[12] Jenny Hamer, Mehryar Mohri, and Ananda Theertha Suresh. FedBoost: A Communication-Efficient Algorithm for Federated Learning. Proceedings of the 37th International Conference on Machine Learning, PMLR, 119:3973-3983, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.816, + 0.47, + 0.844 + ], + "angle": 0, + "content": "[13] Geoffrey Hinton, Oriol Vinyls, and Jeff Dean. Distilling the Knowledge in a Neural Network. NeurIPS, 2014. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.846, + 0.47, + 0.9 + ], + "angle": 0, + "content": "[14] Jiwnoo Jeon, Kangwook Lee, Sewoong Oh, and Jungseul. Ok. Gradient Inversion with Generative Image Prior. Advances in Neural Information Processing Systems (NeurIPS), 34, 2021. 1, 5" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.116, + 0.472, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.135 + ], + "angle": 0, + "content": "[15] Peter Kairouz, H. Brendan McMahan, Brendan Avent, Aurélien Bellet, and Mehdi. Bennis. Advances and Open Problems in Federated Learning. Now Publishers, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.137, + 0.892, + 0.19 + ], + "angle": 0, + "content": "[16] Georgios A. Kaissis, Marcus R. Makowski, Daniel Rückert, and Rickmer F. Braren. Secure, privacy-preserving and federated machine learning in medical imaging. Nature Machine Intelligence, 2:305-311, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.192, + 0.892, + 0.233 + ], + "angle": 0, + "content": "[17] Alex Krizhevsky. Gradient-based learning applied to document recognition. Technical report, University of Toronto, 2009. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.235, + 0.892, + 0.29 + ], + "angle": 0, + "content": "[18] Tao Lin, Lingjing Kong, Sebastian U Stich, and Martin Jaggi. Ensemble distillation for robust model fusion in federated learning. In Advances in Neural Information Processing Systems, volume 33, pages 2351-2363, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.292, + 0.892, + 0.347 + ], + "angle": 0, + "content": "[19] Chuan Ma, Jun Li, Ming Ding, Howard H. Yang, Feng Shu, Tony Q. S. Quek, and Tony Q. S. Poor. On safeguarding privacy and security in the framework of federated learning. IEEE Network, pages 242-248, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.349, + 0.892, + 0.417 + ], + "angle": 0, + "content": "[20] H. Brendan McMahan, Eider Moore, Daniel Ramage, Seth Hampson, and Blaise Agüera y. Arcas. Communication-Efficient Learning of Deep Networks from Decentralized Data. Proceedings of the 20th International Conference on Artificial Intelligence and Statistics (AISTATS), 2016. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.419, + 0.892, + 0.502 + ], + "angle": 0, + "content": "[21] Fan Mo, Hamed Haddadi, Kleomenis Katevas, Eduard Marin, Diego Perino, and Nicolas Kourtellis. PPFL: Privacy-Preserving Federated Learning with Trusted Execution Environments. In Proceedings of the 19th Annual International Conference on Mobile Systems, Applications, and Services (MobiSys), page 94-108, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.504, + 0.892, + 0.571 + ], + "angle": 0, + "content": "[22] Willem G van Panhuis, Proma Paul, Claudia Emerson, John Grefenstette, Richard Wilder, Abraham J Herbst, and Donald S Heymann, David Burke. A systematic review of barriers to data sharing in public health. BMC Public Health, 2014. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.574, + 0.892, + 0.657 + ], + "angle": 0, + "content": "[23] Micah J. Sheller, Brandon Edwards, G. Anthony Reina, Jason Martin, Sarthak Pati, Aikaterini Kotrotsou, Mikhail Milchenko, Weilin Xu, Daniel Marcus, Rivka R. Colen, and Spyridon. Bakas. Federated learning in medicine: facilitating multi-institutional collaborations without sharing patient data. Nature Scientific Reports, 10, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.658, + 0.892, + 0.727 + ], + "angle": 0, + "content": "[24] Kang Wei, Jun Li, Ming Ding, Chuan Ma, Howard H. Yang, and Farhad. Farokhi. Federated Learning With Differential Privacy: Algorithms and Performance Analysis. IEEE Transactions on Information Forensics and Security, 15:3454-3469, 2020. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.728, + 0.892, + 0.797 + ], + "angle": 0, + "content": "[25] Jaehong Yoon, Wonyong Jeong, Giwoong Lee, Eunho Yang, and Sung Ju. Hwang. Federated Continual Learning with Weighted Inter-client Transfer. Proceedings of the 37th International Conference on Machine Learning, PMLR, 139:12073-12086, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.799, + 0.892, + 0.854 + ], + "angle": 0, + "content": "[26] Chengliang Zhang, Suyi Li, Junzhe Xia, Wei Wang, Feng Yan, and Yang Liu. BatchCrypt: efficient homomorphic encryption for cross-silo federated learning. Proceedings of the USENIX Conference, page 493-506, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.855, + 0.892, + 0.883 + ], + "angle": 0, + "content": "[27] Ying Zhang, Tao Xiang, Timothy M. Hospedales, and Huchuan Lu. Deep Mutual Learning. CVPR, 2017. 2, 6" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.883 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8092" + } + ] +] \ No newline at end of file diff --git a/2023/CaPriDe Learning_ Confidential and Private Decentralized Learning Based on Encryption-Friendly Distillation Loss/ca732fe2-4161-4d55-be24-51fc46dde867_origin.pdf b/2023/CaPriDe Learning_ Confidential and Private Decentralized Learning Based on Encryption-Friendly Distillation Loss/ca732fe2-4161-4d55-be24-51fc46dde867_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..2746e3356032ee439cf7778f493bb4133eb7e437 --- /dev/null +++ b/2023/CaPriDe Learning_ Confidential and Private Decentralized Learning Based on Encryption-Friendly Distillation Loss/ca732fe2-4161-4d55-be24-51fc46dde867_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:62b922db3e57acad2ac990efb1465e2b5c0e2395f30c7ac2c8488fe9e12e8a4f +size 1006803 diff --git a/2023/CaPriDe Learning_ Confidential and Private Decentralized Learning Based on Encryption-Friendly Distillation Loss/full.md b/2023/CaPriDe Learning_ Confidential and Private Decentralized Learning Based on Encryption-Friendly Distillation Loss/full.md new file mode 100644 index 0000000000000000000000000000000000000000..8cc9caabb8a61f0f1e247964916316dfaab9af69 --- /dev/null +++ b/2023/CaPriDe Learning_ Confidential and Private Decentralized Learning Based on Encryption-Friendly Distillation Loss/full.md @@ -0,0 +1,311 @@ +# CaPriDe Learning: Confidential and Private Decentralized Learning based on Encryption-friendly Distillation Loss + +Nurbek Tastan Karthik Nandakumar +Mohamed bin Zayed University of Artificial Intelligence, UAE +{nurbek.tastan,karthik.nandakumar}@mbzuai.ac.ae + +# Abstract + +Large volumes of data required to train accurate deep neural networks (DNNs) are seldom available with any single entity. Often, privacy concerns prevent entities from sharing data with each other or with a third-party learning service provider. While cross-silo federated learning (FL) allows collaborative learning of large DNNs without sharing the data itself, most existing cross-silo FL algorithms have an unacceptable utility-privacy trade-off. In this work, we propose a framework called Confidential and Private Decentralized (CaPriDe) learning, which optimally leverages the power of fully homomorphic encryption (FHE) to enable collaborative learning without compromising on the confidentiality and privacy of data. In CaPriDe learning, participating entities release their private data in an encrypted form allowing other participants to perform inference in the encrypted domain. The crux of CaPriDe learning is mutual knowledge distillation between multiple local models through a novel distillation loss, which is an approximation of the Kullback-Leibler (KL) divergence between the local predictions and encrypted inferences of other participants on the same data that can be computed in the encrypted domain. Extensive experiments on three datasets show that CaPriDe learning can improve the accuracy of local models without any central coordination, provide strong guarantees of data confidentiality and privacy, and has the ability to handle statistical heterogeneity. Constraints on the model architecture (arising from the need to be FHE-friendly), limited scalability, and computational complexity of encrypted domain inference are the main limitations of the proposed approach. The code can be found at https://github.com/tnurbek/capride-learning. + +# 1. Introduction + +Rapid strides have been made in machine learning (ML) (in particular, deep learning) over the past decade. However, in many important application domains such as health- + +care and finance, the absence of large, centralized datasets is a significant obstacle to realizing the full benefits of deep learning algorithms. Data in these applications often resides in silos and is governed by strict regulations (e.g., HIPAA, GDPR, etc.) because of its privacy sensitive nature [22]. Competing business interests of data owners and the lack of appropriate incentives for data sharing further accentuate the problem. To overcome these issues, there is a need for collaborative learning algorithms that ideally satisfy the following requirements [15]: (i) confidentiality - no sharing of raw data, (ii) privacy - minimal leakage of information via the knowledge exchange mechanism (e.g., gradients or predictions), (iii) utility - gain in accuracy (over the individual models) resulting from the collaboration, even in the presence of statistical heterogeneity, (iv) efficiency - minimize computational complexity and communication burden, (v) robustness - handle unintentional failures and attacks emanating from malicious entities, and (vi) fairness - utility should be proportional to the individual contributions. + +Federated learning (FL) [15] is a special case of collaborative learning, which works under the orchestration of a central server. FL allows multiple entities to collaboratively solve a ML problem by sharing of focused updates (e.g., gradients), instead of raw data. Specifically, cross-silo FL (typically between 2-100 participants) has been touted as a promising solution to address the data fragmentation problem in finance [8] and healthcare [16,23]. Most prior cross-silo FL algorithms assume that all the parties are collectively training a single model with a common architecture, which is too restrictive in practice. Furthermore, knowledge exchange usually happens through sharing of gradients or model parameters. Recent gradient inversion attacks [9, 14] demonstrate that it is indeed possible to recover high fidelity information from individual gradient updates, thus violating the privacy requirement. While differential privacy (DP) [24], secure multi-party computation (MPC) [26], and trusted execution environment (TEE) [21] have been proposed as potential remedies to safeguard privacy in FL, none of the existing solutions offer an acceptable privacy-utility trade-off with reasonable efficiency. As + +noted in [5], sharing of high dimensional gradients/model parameters is a fundamental privacy limitation of standard FL methods, which cannot be addressed by simply wrapping around a DP, MPC, or TEE solution. + +Confidential and private collaborative learning (CaPC) [7] is the only method that claims to achieve both confidentiality and privacy in a collaborative setting. In CaPC learning, each participant is able to query other parties via a private inference protocol based on 2-party computation (2PC). However, the answering parties cannot directly reveal the inference logits to the querying party because it leaks information about their local models (e.g., through model extraction attacks). To circumvent this problem, differential privacy, private aggregation of teacher ensembles, and secure argmax computation through a central privacy guardian (PG) are employed to output only the predicted label to the querying party. A drawback of the CaPC approach is that it achieves all privacy guarantees using the help of a semi-trusted PG. Moreover, the use of differential privacy reduces the performance of FL algorithms, especially if strong privacy guarantees are required. Finally, the use of MPC requires multiple rounds of communication between the parties for each query. All other approaches that have been proposed to achieve knowledge transfer through distillation [18] require a non-private labeled/unlabeled/synthetic dataset that can be shared among participants. + +In this work, we propose a new protocol called Confidential and Private Decentralized (CaPriDe) learning, where participants learn from each other in a collaborative setting while preserving their confidentiality and privacy (see Figure 1). Unlike the 2PC protocols used in CaPC, we leverage fully homomorphic encryption (FHE) to enable participants to publish their encrypted data. Since the published data is encrypted using the data owner's public key and only the owner can decrypt the data (using the secret key), confidentiality is preserved. However, the CaPriDe framework allows other collaborators to perform encrypted inference on the published (encrypted) data by applying their own local model. Mutual knowledge distillation (KD) [13] between the data owner's local model and the local models of the collaborators is used to transfer knowledge between models and ensure a collaborative gain. KD is typically achieved through minimizing the distillation loss between the student and teacher responses (logits). Since the collaborators in CaPriDe learning make predictions on encrypted data, a loss function that can be computed without any decryption is required. Hence, we propose a new distillation loss, which is an approximation of the KL divergence that can be securely computed. Only an encrypted value of the distillation loss aggregated over an entire batch is sent back to the data owner, which ensures strong privacy. Data owners can decrypt this aggregate loss value and use it for updating their local models. Our contributions are: + +- We introduce the CaPriDe learning framework, which exploits FHE-based encrypted inference and knowledge distillation to achieve confidential and private collaborative learning without any central orchestration and any need for non-private shared data. +- To enable CaPriDe learning, we propose an encryption-friendly distillation loss that estimates the approximate KL divergence between two model predictions and design a protocol to securely compute this loss in the encrypted domain. +- We conduct extensive experiments to show that CaPriDe learning enables participants to achieve collaborative gain, even in the non-iid setting. To prove feasibility, we implement encrypted inference using the Tile Tensors library with a FHE-friendly model. + +# 2. Background + +Federated learning (FL) was first proposed in [20] as a distributed ML algorithm that does not use user data directly. Subsequent works attempted to address various challenges in FL, including privacy [7, 19], communication efficiency [12], convergence, catastrophic forgetting [25], and secure aggregation [3]. Though several privacy-enhancing solutions have been proposed, most of them fail to provide a satisfactory balance between accuracy and privacy. Only recently, some promising solutions such as CaPC have been proposed [7]. These advancements have been made by jettisoning the conventional gradient sharing approach and borrowing ideas from the knowledge distillation literature [13]. Specifically, it has been shown that it is possible for an ML model to benefit from predictions made by other models, by aligning the local predictions with those obtained from external parties. Deep mutual learning (DML) [27] is an important work in this direction, where models learn collaboratively and teach each other throughout the training process. Compared to the distillation by a pre-defined static teacher network, DML achieves better performance, and mutually learned models outperform independently trained models irrespective of model capacity. Knowledge Distillation via Collaborative Learning (KDCL) [11], Cronus [5], and Ensemble Distillation [18] also follow a similar idea. + +# 2.1. Fully Homomorphic Encryption + +Let $\mathcal{E}(x)$ denote the encrypted value of $x$ using the public key $\kappa$ . The encryption scheme is called homomorphic over an operation $\oplus$ if it supports the following equation: + +$$ +\mathcal {E} \left(x _ {1} \oplus x _ {2}\right) = \mathcal {E} \left(x _ {1}\right) \otimes \mathcal {E} \left(x _ {2}\right), \forall x _ {1}, x _ {2} +$$ + +where operators $\otimes$ and $\oplus$ need not be the same. The scheme is fully homomorphic if for some arbitrary function + +![](images/2c55aebc2280679ccc1d83c904c96b3cbc3fbb805415b830bcdb37fb1a6a6dcd.jpg) +Figure 1. Illustration of confidential and private decentralized (CaPriDe) learning framework for $K$ participants, showing the learning process for only participant $\mathcal{P}_1$ . Here, $\mathcal{D}_1 = (X_1, Y_1) = \{x_{j,1}, y_{j,1}\}_{j=1}^{N_1}$ is the local data of $\mathcal{P}_1$ , $\mathcal{E}(X_1)$ denotes the collection of encrypted unlabeled samples of $\mathcal{P}_1$ , and $\pmb{p}_{(1,l)}$ and $\pmb{z}_{(1,l)}$ are the prediction probabilities and logits obtained by applying model $\mathcal{M}_{\theta_l}$ of participant $\mathcal{P}_l$ to $X_1$ . Black dashed line represents exchange of encrypted data between participants in each round, and blue dashed line denotes a single transfer at the beginning of the protocol. $\mathcal{L}_{CE}$ and $\mathcal{L}_{DL}$ represent the cross-entropy and distillation losses, respectively. + +$f(x_{1},x_{2},\ldots ,x_{n})$ , there exists a function $g$ and decryption method $^d$ such that: + +$$ +f \left(x _ {1}, x _ {2}, \dots , x _ {n}\right) = d \left(g \left(\mathcal {E} \left(x _ {1}\right), \mathcal {E} \left(x _ {2}\right), \dots , \mathcal {E} \left(x _ {n}\right)\right)\right), +$$ + +where $d$ uses secret key $\tilde{\kappa}$ to decrypt the given ciphertext value. Gentry [10] first constructed somewhat homomorphic encryption (SWHE) and later developed bootstrapping of ciphertexts to reduce noise. Among the various FHE schemes such as BGV, BFV, and CKKS [6], only CKKS works with real numbers and approximate computations. + +# 2.2. Preliminaries + +Typically, a supervised classifier $\mathcal{M}$ is mapping from the input space $\pmb{x} \in \mathcal{X} \subset \mathbb{R}^n$ to the label space $y \in \mathcal{Y} = \{1,2,\dots,M\}$ , where $M$ is the number of classes. For convenience, we assume that $\mathcal{M}: \mathcal{X} \to \mathcal{Z}$ , where $\mathcal{Z}$ denotes the logits space ( $\mathcal{Z} \subset \mathbb{R}^M$ ). In other words, $\mathcal{M}$ maps the input $\pmb{x}$ to a $M$ -dimensional logits vector $\pmb{z} = [z^{1},z^{2},\dots,z^{M}]$ . A softmax function $(\sigma)$ can be applied to the logits vector $\pmb{z}$ to obtain the probability distribution $\pmb{p} = \sigma(\pmb{z}) = [p^{1},p^{2},\dots,p^{M}]$ over $M$ labels, i.e., + +$$ +p ^ {m} = \frac {\exp (z ^ {m} / T)}{C}, \tag {1} +$$ + +where $T$ is the softmax temperature parameter and $C = \sum_{i=1}^{M} \exp(z^i / T)$ is a normalization constant that ensures $\sum_{m=1}^{M} p^m = 1$ . Here, $p^m$ denotes the probability that input $\mathbf{x}$ belongs to class $m, m = 1, 2, \dots, M$ . Given two $V$ -dimensional vectors $\mathbf{v}_1, \mathbf{v}_2 \in \mathbb{R}^V$ , let $(\mathbf{v}_1 \cdot \mathbf{v}_2)$ denote the dot product, $(\mathbf{v}_1 \odot \mathbf{v}_2)$ denote the element-wise product, and $f(\mathbf{v}_1)$ represent the vector obtained by applying + +the function $f$ to each element in $\pmb{v}_1$ . The cross entropy loss of model $\mathcal{M}$ on a sample $(x, y)$ is defined as: + +$$ +L _ {C E} (\boldsymbol {p}, y) = - \mathbf {1} _ {y} \cdot l o g (\boldsymbol {p}), +$$ + +where $\pmb{p} = \sigma(\mathcal{M}(\pmb{x}))$ and $\mathbf{1}_y$ is an indicator vector whose $y^{th}$ element is 1 and other elements are 0. The KL divergence between two distributions $\pmb{p}_1$ and $\pmb{p}_2$ is given by: + +$$ +D _ {K L} \left(\boldsymbol {p} _ {2} \| \boldsymbol {p} _ {1}\right) = \sum_ {m = 1} ^ {M} p _ {2} ^ {m} \log \frac {p _ {2} ^ {m}}{p _ {1} ^ {m}} = \left(\boldsymbol {p} _ {2} \cdot \log \boldsymbol {p} _ {2}\right) - \left(\boldsymbol {p} _ {2} \cdot \log \boldsymbol {p} _ {1}\right). \tag {2} +$$ + +The L2 distance between two logit vectors $z_{1}$ and $z_{2}$ is: + +$$ +D _ {L 2} \left(\boldsymbol {z} _ {2}, \boldsymbol {z} _ {1}\right) = \sum_ {m = 1} ^ {M} \left(z _ {2} ^ {m} - z _ {1} ^ {m}\right) ^ {2} = \left(\boldsymbol {z} _ {2} - \boldsymbol {z} _ {1}\right) \cdot \left(\boldsymbol {z} _ {2} - \boldsymbol {z} _ {1}\right). +$$ + +# 3. Proposed Framework + +# 3.1. Problem Statement + +Suppose that there are $K$ participants $\mathcal{P}_1, \mathcal{P}_2, \dots, \mathcal{P}_K$ . Each party $\mathcal{P}_k, k = 1, 2, \dots, K$ has its own local training dataset $\mathcal{D}_k = \{\pmb{x}_{j,k}, y_{j,k}\}_{j=1}^{N_k}$ of size $N_k$ using which it learns its own local ML model $\mathcal{M}_{\theta_k}$ , where $\mathcal{M}$ denotes the architecture and $\theta$ represents the parameter set. We also assume that each party $\mathcal{P}_k$ has its own validation set $\tilde{\mathcal{D}}_k = \{\tilde{\pmb{x}}_{j,k}, \tilde{y}_{j,k}\}_{j=1}^{\tilde{N}_k}$ of size $\tilde{N}_k$ and the validation accuracy of model $\mathcal{M}_{\theta_k}$ on $\tilde{\mathcal{D}}_k$ is $\mathcal{A}_k$ . The goal of each participant is to collaborate with other parties in order to improve + +the accuracy of its local model, i.e., obtain a better model $\mathcal{M}_{\theta_k^*}$ that has a higher validation accuracy $\mathcal{A}_k^*$ on $\tilde{\mathcal{D}}_k$ compared to $\mathcal{A}_k$ , without compromising on the confidentiality and privacy of its local data $\mathcal{D}_k$ . In other words, no participant $\mathcal{P}_l, l \neq k$ can access $\mathcal{D}_k$ in its plaintext form or learn any private information about $\mathcal{D}_k$ such as the $x_{j,k}$ and $y_{j,k}$ values. The utility of the framework is measured as the average collaboration gain, i.e., $U = \frac{1}{K} \sum_{k=1}^{K} (\mathcal{A}_k^* - \mathcal{A}_k)$ . + +# 3.2. CaPriDe Learning Framework Description + +Let $\mathcal{E}(\pmb{X}_k) = \{\mathcal{E}(\pmb{x}_{j,k})\}_{j=1}^{N_k}$ be the collection of encrypted input samples of participant $\mathcal{P}_k$ . Here, the encryption is based on a FHE scheme with public key $\kappa_k$ of $\mathcal{P}_k$ . In CaPriDe learning, each party initiates the collaboration by publishing $\mathcal{E}(\pmb{X}_k)$ to the other participants. We now describe how the learning proceeds at participant $\mathcal{P}_k$ . Let $\pmb{z}_{j,(k,l)}$ and $\pmb{p}_{j,(k,l)}$ denote the logits vector and probability distribution obtained when model $\mathcal{M}_{\theta_l}$ of $\mathcal{P}_l$ is applied on the input sample $\pmb{x}_{j,k}$ , i.e., the $j^{th}$ training sample of $\mathcal{P}_k$ . Assume that we are at round $(t+1), 0 \leq t < \tau$ , of the collaboration and the current model of $\mathcal{P}_k$ is $\mathcal{M}_{\theta_k^t}$ . $\mathcal{P}_k$ performs one forward pass on its own training set $\mathcal{D}_k$ to obtain the predictions $\{\pmb{p}_{j,(k,k)}\}_{j=1}^{N_k}$ . The cross-entropy loss of party $\mathcal{P}_k$ can be obtained as: + +$$ +\mathcal {L} _ {C E _ {k}} = \sum_ {j = 1} ^ {N _ {k}} L _ {C E} \left(\boldsymbol {p} _ {j, (k, k)}, y _ {j, k}\right). \tag {3} +$$ + +To enable knowledge transfer between participants, we make use of the knowledge distillation (KD) approach [13], where a student model learns to mimic the predictions of a teacher model. In CaPriDe learning, there is no designated teacher model, and mutual KD between multiple peer models is used for knowledge transfer. Let $\mathcal{L}_{DL_{(k,l)}}$ denote the pairwise distillation loss between predictions of $\mathcal{P}_k$ and $\mathcal{P}_l$ , which is defined as: + +$$ +\mathcal {L} _ {D L _ {(k, l)}} = \sum_ {j = 1} ^ {N _ {k}} D _ {D L} \left(\boldsymbol {p} _ {j, (k, k)}, \boldsymbol {p} _ {j, (k, l)}\right), \tag {4} +$$ + +where $D_{DL}(\pmb{p}_1, \pmb{p}_2)$ denotes the mimic distance between two predictions $\pmb{p}_1$ and $\pmb{p}_2$ . The total distillation loss for $\mathcal{P}_k$ can be computed as: + +$$ +\mathcal {L} _ {D L _ {k}} = \frac {1}{K - 1} \sum_ {l = 1, l \neq k} ^ {K} \mathcal {L} _ {D L _ {(k, l)}}. \tag {5} +$$ + +Therefore, the overall loss of $\mathcal{P}_k$ is given by: + +$$ +\mathcal {L} _ {k} = \mathcal {L} _ {C E _ {k}} + \lambda_ {k} \mathcal {L} _ {D L _ {k}}, \tag {6} +$$ + +where $\lambda_{k}$ is the weighting factor between the cross-entropy and distillation losses for $\mathcal{P}_k$ . The model of participant $\mathcal{P}_k$ is then updated through a backward pass as follows: + +$$ +\theta_ {k} ^ {t + 1} = \theta_ {k} ^ {t} - \alpha_ {k} \nabla_ {\theta_ {k} ^ {t}} \mathcal {L} _ {k}, \tag {7} +$$ + +where $\alpha_{k}$ is the learning rate of $\mathcal{P}_k$ + +# 3.3. Encryption-friendly Distillation Loss + +The key missing component in the above CaPriDe learning framework is the computation $D_{DL}$ in Eq. 4. For participant $\mathcal{P}_k$ , its logit vectors $\pmb{z}_{.,(k,k)}$ and predictions $\pmb{p}_{.,(k,k)}$ are available in the plaintext domain. However, since participant $\mathcal{P}_l$ receives the input samples of $\mathcal{P}_k$ in encrypted form, it can perform inference in the encrypted domain and obtain only $\mathcal{E}\left(\pmb{z}_{.,(k,l)}\right)$ , where $\mathcal{E}\left(\pmb{z}_{.,(k,l)}\right) = \mathcal{M}_{\theta_l}\left(\mathcal{E}(\pmb{x}_{.,k})\right)$ . It is not straightforward to obtain $\mathcal{E}\left(\pmb{p}_{.,(k,l)}\right)$ from $\mathcal{E}\left(\pmb{z}_{.,(k,l)}\right)$ because it involves computing a non-linear softmax function in the encrypted domain, which is a challenging task in the FHE domain and consumes a large multiplicative depth. Even if this softmax operation is somehow implemented, it is not possible to send $\mathcal{E}\left(\pmb{p}_{.,(k,l)}\right)$ back to $\mathcal{P}_k$ , who can decrypt it and obtain $\pmb{p}_{.,(k,l)}$ . Since $\mathcal{P}_k$ also has knowledge of $\pmb{x}_{.,k}$ , it becomes feasible for $\mathcal{P}_k$ to carry out a model extraction attack based on $(\pmb{x}_{.,k},\pmb{p}_{.,(k,l)})$ pairs to glean information about model $\mathcal{M}_{\theta_l}$ . This violates the privacy requirements of participant $\mathcal{P}_l$ . Hence, there is a need for a loss function that can be computed in the encrypted domain. + +A straightforward solution is to compute the L2 distance between logit vectors $\mathbf{z}_{.,(k,k)}$ and $\mathbf{z}_{.,(k,l)}$ . In this case, participant $\mathcal{P}_k$ can publish $\mathcal{E}\left(\mathbf{z}_{.,(k,k)}\right)$ and the pairwise distillation loss $\mathcal{L}_{DL_{(k,l)}}$ can be computed by participant $\mathcal{P}_l$ as: + +$$ +\begin{array}{l} \left. \mathcal {E} \left(D _ {D L} \left(\boldsymbol {p} _ {j, (k, k)}, \boldsymbol {p} _ {j, (k, l)}\right)\right) \equiv \mathcal {E} \left(D _ {L 2} \left(\boldsymbol {z} _ {j, (k, k)}, \boldsymbol {z} _ {j, (k, l)}\right)\right) \right. \\ = \left(\mathcal {E} \left(\boldsymbol {z} _ {j, (k, k)}\right) - \mathcal {E} \left(\boldsymbol {z} _ {j, (k, l)}\right)\right) \cdot \left(\mathcal {E} \left(\boldsymbol {z} _ {j, (k, k)}\right) - \mathcal {E} \left(\boldsymbol {z} _ {j, (k, l)}\right)\right) \\ \end{array} +$$ + +However, our experiments indicate the above simple solution based on L2 distance does not lead to high utility $U$ . On the other hand, we observed that using the KL divergence between predictions of $\mathcal{P}_k$ and $\mathcal{P}_l$ almost always leads to better utility. But, $D_{KL}\left(\pmb{p}_{.,(k,k)}\| \pmb{p}_{.,(k,l)}\right)$ cannot be directly computed in the encrypted domain because participant $\mathcal{P}_l$ only has access to the encrypted logits $\mathcal{E}\left(z_{.,(k,l)}\right)$ . + +Fortunately, this problem can be solved by approximating the KL divergence. Combining Eqs. 1 and 2, we get: + +$$ +D _ {K L} \left(\boldsymbol {p} _ {2} \right\| \boldsymbol {p} _ {1}) = \sum_ {m = 1} ^ {M} p _ {2} ^ {m} \log p _ {2} ^ {m} - \sum_ {m = 1} ^ {M} p _ {2} ^ {m} \left(z _ {1} ^ {m} / T\right) + M \log C. \tag {8} +$$ + +Here, the first term can be denoted as $(\pmb{p}_2 \cdot \log \pmb{p}_2)$ and measures the entropy (uncertainty) in $\pmb{p}_2$ . The second term measures the divergence (unnormized cross-entropy) between $\pmb{p}_2$ and $\pmb{p}_1$ and can be expressed as $(\pmb{p}_2 \cdot \pmb{z}_1 / T)$ . The final term $M \log C$ is just a normalization constant that does not affect the relative ordering of the class predictions. We observed that this normalization term does not appear to play a major role in model training. Hence, we ignore this value + +Algorithm 1 CaPRiDe learning algorithm +Input: Number of participants $K$ , weight parameter $\lambda_{k}$ , learning rate $\alpha_{k}$ for each participant +Assumption: Each participant has local training set $\mathcal{D}_k = \{\pmb {x}_{j,k},y_{j,k}\}_{j = [1,N_k]}$ of size $N_{k}$ , a fully homomorphic encryption function $\mathcal{E}$ with public key $\kappa_{k}$ , a decryption function $d$ with secret key $\tilde{\kappa}_k$ +for each participant $k = 1,2,\dots ,K$ do Train initial model $\mathcal{M}_{\theta_k^0}$ based on local data $\mathcal{D}_k$ Publish encrypted data $\{\mathcal{E}(\pmb{x}_{j,k})\}_{j = [1,N_k]}$ +end for +for each round $t = 1,2,\dots ,\tau$ do +for each participant $k = 1,2,\dots ,K$ do $\{\pmb {p}_{j,(k,k)}\}_{j = [1,N_k]}\gets \{\sigma (\mathcal{M}_{\theta_k^{t - 1}}(\pmb {x}_{j,k}))\}_{j = [1,N_k]}$ $\mathcal{L}_{CE_k}\gets$ Compute cross-entropy loss of participant $k$ using Equation 3 + $\hat{\mathcal{L}}_{KL_k}\gets$ Compute local uncertainty of participant $k$ using Equation 11 +Publish encrypted inferences $\{\mathcal{E}(\pmb {p}_{j,(k,k)})\}_{j = [1,N_k]}$ +for each participant $l = 1,2,\dots ,K,l\neq k$ do $\{\mathcal{E}(\pmb {z}_{j,(k,l)})\}_{j = [1,N_k]}\gets \{\mathcal{M}_{\theta_l^{t - 1}}(\mathcal{E}(\pmb {x}_{j,k}))\}_{j = [1,N_k]}$ (Perform inference on encrypted data) + $\mathcal{E}(\tilde{\mathcal{L}}_{KL_{(k,l)}})\gets$ Compute pairwise divergence in encrypted domain using Equation 12 +Send $\mathcal{E}(\tilde{\mathcal{L}}_{KL_{(k,l)}})$ to participant $k$ $\mathcal{L}_{DL_{(k,l)}}\gets$ Compute pairwise distillation loss between participants $k$ and $l$ using Equation 10 +end for + $\mathcal{L}_{DL_k}\gets$ Compute distillation loss of participant $k$ using Equation 5 + $\mathcal{L}_k\gets$ Compute total loss of participant $k$ using Equation 6 + $\theta_k^t\gets$ Update local model of participant $k$ using Equation 7 +end for +end for + +and define the approximate KL divergence is defined as: + +$$ +\hat {D} _ {K L} \left(\boldsymbol {p} _ {2} \| \boldsymbol {p} _ {1}\right) = \left(\boldsymbol {p} _ {2} \cdot \log \boldsymbol {p} _ {2}\right) - \left(\boldsymbol {p} _ {2} \cdot \boldsymbol {z} _ {1} / T\right). \tag {9} +$$ + +Going back to Eq. 4, the pairwise distillation loss based on the above approximate KL divergence can be easily computed in the encrypted domain as: + +$$ +\begin{array}{l} \left. \mathcal {E} \left(D _ {D L} \left(\boldsymbol {p} _ {j, (k, k)}, \boldsymbol {p} _ {j, (k, l)}\right)\right) \equiv \mathcal {E} \left(\hat {D} _ {K L} \left(\boldsymbol {p} _ {j, (k, k)} | | \boldsymbol {p} _ {j, (k, l)}\right)\right) \right. \\ = \mathcal {E} \left(\boldsymbol {p} _ {j, (k, k)} \cdot \log \boldsymbol {p} _ {j, (k, k)}\right) - \mathcal {E} \left(\boldsymbol {p} _ {j, (k, k)}\right) \cdot \mathcal {E} \left(\boldsymbol {z} _ {j, (k, l)}\right) / T \\ \end{array} +$$ + +The above distillation loss based on the approximate KL divergence is encryption-friendly because it involves only linear operations and scaling by a constant plaintext value $T$ . In practice, our goal is to compute $\mathcal{L}_{DL_{(k,l)}}$ at participant $\mathcal{P}_k$ . Therefore, Eq. 4 can be further reformulated as: + +$$ +\mathcal {L} _ {D L _ {(k, l)}} = \hat {\mathcal {L}} _ {K L _ {k}} - d \left(\mathcal {E} \left(\tilde {\mathcal {L}} _ {K L _ {(k, l)}}\right)\right), \tag {10} +$$ + +where + +$$ +\hat {\mathcal {L}} _ {K L _ {k}} = \sum_ {j = 1} ^ {N _ {k}} \left(\boldsymbol {p} _ {j, (k, k)} \cdot \log \boldsymbol {p} _ {j, (k, k)}\right), \tag {11} +$$ + +$$ +\mathcal {E} \left(\tilde {\mathcal {L}} _ {K L _ {(k, l)}}\right) = \sum_ {j = 1} ^ {N _ {k}} \mathcal {E} \left(\boldsymbol {p} _ {j, (k, k)}\right) \cdot \mathcal {E} \left(\boldsymbol {z} _ {j, (k, l)}\right) / T. \tag {12} +$$ + +Since $\mathcal{P}_k$ has access to $\pmb{p}_{j,(k,k)}$ , it is straightforward to compute $\hat{\mathcal{L}}_{KL_k}$ in plaintext space. If participant $\mathcal{P}_k$ can publish $\left\{\mathcal{E}(\pmb{p}_{j,(k,k)})\right\}_{j=1}^{N_k}$ in each round, $\mathcal{P}_l$ can compute $\mathcal{E}\left(\tilde{\mathcal{L}}_{KL(k,l)}\right)$ in the encrypted domain because it already has access to $\left\{\mathcal{E}\left(z_{j,(k,l)}\right)\right\}_{j=1}^{N_k}$ . The encrypted value of $\tilde{\mathcal{L}}_{KL(k,l)}$ is sent back to participant $\mathcal{P}_k$ , who decrypts it to compute the overall loss via Eqs. 10, 5 and 6. + +Privacy Analysis: In the CaPriDe learning framework, the only quantity that can potentially leak private information is the pairwise distillation loss between participants $\tilde{\mathcal{L}}_{KL_{(k,l)}}$ , which has already been aggregated over all the $N_{k}$ samples. A curious participant $\mathcal{P}_k$ attempting to reconstruct the logit values $z_{j,(k,l)}^m$ of participant $\mathcal{P}_l$ needs to dis-aggregate this sum and compute per-sample loss, which is infeasible in the absence of other constraints. Hence, the proposed CaPriDe learning framework is privacy-preserving. It must be emphasized that even gradient inversion attacks [9, 14] are significantly harder when the batch size is large (above 16) due to the same reason. The proposed method also has other advantages that enhance privacy: (i) A curious participant cannot query its collaborator adaptively. The encrypted samples are published once at the beginning and do not change. (ii) The responding participant can choose a + +random subset of samples in each round to compute the aggregate loss. This can improve efficiency as well as provide a privacy benefit by eliminating correlation across multiple rounds. (iii) We can incorporate differential privacy (DP) by adding noise to the aggregate pairwise distillation loss. Since the framework allows exchange of only encrypted input samples and encrypted predictions, which never get decrypted during the execution of the protocol, the CaPriDe learning approach also preserves data confidentiality. + +# 4. Experiments + +# 4.1. Datasets + +CIFAR-10 [17] is a dataset consisting of 60000 $32 \times 32$ RGB images from 10 classes, with 6000 images per class. It has 50000 training and 10000 test samples. + +CIFAR-100 dataset [17] is similar to the CIFAR-10 dataset, but it has 100 classes containing 600 samples each. Training and evaluation sets are split in the same way as in CIFAR-10. CIFAR-100 dataset is used with data augmentation to account for the problem complexity. The data augmentations employed include random rotation (up to 15 degrees), random crop, and horizontal flip. + +HAM10000 is a multi-class imbalanced dataset comprising of 10,015 dermoscopic images from diverse populations with 7 categories of pigmented lesions. We randomly perform an $85\% / 15\%$ split for training/ testing. + +# 4.2. Experimental Setup + +We compare the CaPriDe learning framework against the following baselines: vanilla FL based on FedAvg (FedAvg) [20], FL with local differential privacy (FedAvg + DP) [24], and deep mutual learning (DML) [27]. We use ResNet-18 architecture for all our experiments. Stochastic Gradient Descent (SGD) algorithm with momentum is used as the optimizer. For all methods, 25 epochs of local training happens before 75 rounds of collaboration. Momentum is set to 0.9, the initial learning rate is set to $\alpha_{k} = 0.1$ and the learning rate decays in rounds 50 and 75 by a factor $\gamma = 0.1$ . The batch size is set to 128 for CIFAR-10 and CIFAR-100 and 32 for HAM10000. The weight factor $\lambda_{k}$ between cross-entropy and distillation losses is set to 50 for CIFAR-10 and CIFAR-100 and 20 for HAM10000. All algorithms are implemented using PyTorch and experiments are carried out on a single NVIDIA RTX A6000 GPU. Additionally, in CaPriDe and DML, the logits are scaled by a temperature parameter $T = 5.0$ . Finally, for FedAvg + DP, the Gaussian noise mechanism is used with $\sigma_{g} = 0.01$ . To implement an FHE-compatible version of ResNet-18, we use the Tile Tensors framework [1,2], which in turn relies on low-level FHE libraries such as HElib and SEAL. For encrypted inference, ReLU is replaced with a polynomial approximation of Gaussian Error Linear Unit + +![](images/43a7b76ab875c0ebbfaa5638e55f6c51a369f325bf764b75d75e67e3419db4c8.jpg) +Figure 2. Accuracy trend of two $(K = 2)$ collaboratively trained models on different datasets (rows) and for different partition strategies (columns). + +$$ +(g (x) = 0. 0 7 1 1 + 0. 5 x + 0. 2 5 7 6 x ^ {2} - 0. 0 1 2 8 x ^ {4}). +$$ + +We consider three types of strategies to partition the training dataset among the participants: (i) Homogeneous: each participant has an equal number of samples per class, (ii) Heterogeneous: each participant has an unequal number of samples determined randomly (both total number and number of samples per class), (iii) Non-overlapping class distribution: (denoted as "no class overlap") each participant has samples from a non-overlapping subset of classes (e.g., for $K = 2$ , participants get even and odd classes). While homogeneous partitioning ensures that data distributions of participants are iid, heterogeneous split simulates typical non-iid scenarios, and the no class overlap case corresponds to extreme statistical heterogeneity. + +# 4.3. Collaborative Learning Results + +To begin with, we establish the utility (collaborative gain) of the proposed CaPriDe learning framework for the case of $K = 2$ participants on all three datasets and all three partition strategies. The results of these experiments are shown in Figure 2. In each case, the collaboration gain of CaPriDe learning is positive and the actual collaboration gain values can be found in Table 1. We make the following observations based on these results. + +- CaPriDe learning works not only in the iid setting but also equally well in the two non iid settings. This + +
DatasetSettingKFedAvgFedAvg+DPCaPriDe
CIFAR-10Homogeneous23.46-20.541.58
CIFAR-10Homogeneous59.22-11.83.55
CIFAR-10Homogeneous1015.442.585.52
CIFAR-10Heterogeneous26.12-17.040.93
CIFAR-10Heterogeneous513.78-7.845.06
CIFAR-10Heterogeneous1021.815.696.20
CIFAR-10No class overlap229.8511.088.16
CIFAR-100Homogeneous27.08-30.974.19
CIFAR-100Homogeneous522.68-13.819.10
CIFAR-100Homogeneous1026.10-0.1111.69
CIFAR-100Heterogeneous29.59-28.896.28
CIFAR-100Heterogeneous522.50-13.269.23
CIFAR-100Heterogeneous1034.383.8610.68
CIFAR-100No class overlap219.405.147.75
HAM10000Homogeneous20.970.481.81
HAM10000Heterogeneous21.93-1.211.67
+ +Table 1. Collaborative gain (\%) on all datasets and all partition strategies for different number of participants $K$ using different algorithms (columns 4-6). DML results are not reported here because there is negligible difference between CaPriDe and DML. + +demonstrates that knowledge distillation is reasonably effective even under extreme statistical heterogeneity. + +- There is a negligible difference between the performance of DML (which uses true KL divergence) and CaPriDe learning (which uses approximate KL divergence) in all the settings. This clearly justifies the dropping of the normalization (third) term in Eq. 8 to obtain the encryption-friendly distillation loss in Eq. 9. While DML offers no confidentiality or privacy, CaPriDe learning offers both without any utility cost. +- In almost all the settings, there is a huge difference in the collaboration gain between FedAvg and CaPriDe learning. This is unsurprising because FedAvg involves sharing and aggregation of full model parameters leading to high utility at a significant cost to privacy. On the other hand, FedAvg + DP provides strong privacy, but this comes with a significant degradation in utility. Hence, it can be argued that CaPriDe learning achieves a good trade-off between privacy and utility compared to standard FL algorithms. + +Next, we evaluate the utility of the proposed method with the different number of participants $K = 2,5,10$ on CIFAR-10 and CIFAR-100 datasets based on the homogeneous partition (iid) setting. From Figure 3 and Table 1 we observe that the collaboration gain increases with the number of participants. This is to be expected because the training samples per participant become less when there are more participants, consequently resulting in lower individual accuracy. + +![](images/08b0e21db30a0721f446a7ba76a23281ac45917998144ad7fc2ec990142e0562.jpg) +Figure 3. Accuracy with different number of participants $K$ based on the homogeneous partition (iid) setting. + +![](images/a92183316b47486ba36b76855ea902e5a985c9b32356968e9760a3d7000f5c81.jpg) +(a) Homogeneous + +![](images/0efbfc2e767cf394f01e1fc1c1ba384a62e69cfbceae82c5bd82ed4a8b8ff32b.jpg) +(b) Heterogeneous + +![](images/d385c9c34c991f29411068436b1d8ca28973a1c6a365138fe76c88d8a3deff76.jpg) +(c) No overlap + +![](images/d6a14f28d4d29f9e9313aea1d6fa5fa0f6a697790d9e6354e97fd7099b8a5fc9.jpg) +Figure 4. Per-class accuracy of CaPriDe learning evaluated on CIFAR-10 averaged across 5 participants. Dashed lines represent mean accuracy and error bars indicate standard deviation. +(a) Homogeneous +Figure 5. Per-participant accuracy of CaPriDe learning evaluated on CIFAR-100 dataset. + +![](images/d578d0c4b1510d5adfd0edbfab2897f8c6ed6d25b5ffa42683a90e928c362176.jpg) +(b) Heterogeneous + +![](images/4e4fc757f199692b37ad1218748a2bf0cf5b388b180730e0331ca7dfc8c3ddb3.jpg) +(c) No overlap + +Furthermore, we evaluate the class-wise and participant-wise accuracy of our CaPriDe learning algorithm with $K = 5$ participants. Figure 4 shows class-wise accuracy averaged across 5 participants for each setting. From this figure, it is clear that almost all the classes benefit from collaboration, and the per-class accuracy of the individual models after collaboration has a lower standard deviation compared to the case without collaboration, especially in homogeneous and heterogeneous settings. Thus, in these two settings, the individual models become similar to each other after collaboration proving the successful transfer of knowledge between participants. Note that Figure 4 displays the average top-1 accuracy of the participants. The collaboration gains + +in the no class overlap case are significantly higher when top-5 accuracy is considered. Finally, Figure 5 illustrates participant-wise accuracy over all classes for each setting. It is obvious that each participant has a positive collaboration gain, which is the eventual goal that we set out to achieve. Moreover, all the 5 participant models have a similar accuracy at the end of the collaboration, once again demonstrating good knowledge transfer. Other results such as the sensitivity of the proposed method to weight factor $\lambda_{k}$ , the sensitivity of FedAvg + DP to noise variance, and the comparison between approximate KL divergence and L2 loss are included in the supplementary material. + +# 4.4. FHE Implementation Results + +For FHE implementation using the Tile Tensors framework, Table 2 summarizes the information about all three datasets at the same security level of 128 bits. Two critical metrics to be kept in mind are the time required for encrypted inference and the memory consumed by a ciphertext. While encrypted inference time determines the computational complexity and latency involved in each round, the ciphertext size determines the communication burden. Note that encrypted inference can be performed on a persample basis or a batch of samples [4]. The single sample inference mode is preferred for machine learning as-a-service applications, where low latency is a critical requirement. However, the batched inference is better suited for the CaPriDe learning framework, where only the overall time required to complete inference on all the samples matter. + +
Security LevelCIFAR-10CIFAR-100HAM10000
128128128
Number of slots163841638416384
Time taken to encrypt one sample90 ms103 ms619 ms
Ciphertext size of one sample29.101 KB29.152 KB1.359 MB
Time taken to encrypt a batch of 32 samples1.31 s1.26 s15.57 s
Encrypted inference of a batch of 32 samples110.21 s112.09 s896.12 s
+ +Table 2. FHE configuration parameters and memory and computational requirements for the chosen datasets based on ResNet-18 architecture. + +The encrypted inference time for a batch of 32 samples is approximately 110s for CIFAR-10 and CIFAR-100 and 900s for HAM10000. Thus, if the local dataset of each participant has 1024 training samples, it will take around 1 hour (8 hours) for encrypted inference per participant and then computation of the pairwise distillation loss in each communication round for CIFAR-10 and CIFAR-100 (HAM10000). While this is a significant computational load, it is less of an issue in the cross-silo collaborative learning setting, especially in sectors such as finance and healthcare, where data confidentiality and privacy as well as accuracy gain are more critical. Moreover, the ciphertext + +size for one sample is approximately, 30KB for CIFAR-10 and CIFAR-100 and 1.36MB for HAM10000. For the same local dataset size of 1024, the encrypted dataset will consume 30MB for CIFAR-10 and CIFAR-100 and 1.36GB for HAM10000. Fortunately, the encrypted dataset is published and needs to be downloaded only once by each participant at the beginning of the protocol. During every communication round, only the encrypted predictions and pairwise losses are exchanged. Depending on the local dataset size, the number of classes, and the ciphertext packing strategy used, these intermediate values can be packed into at most a few tens of ciphertexts (note that this does not depend on the image size). Hence, the per round communication burden of CaPriDe learning framework is relatively small. + +Limitations and Potential Solutions: Note that all the above complexity computations are based on the assumption of $K = 2$ participants. As $K$ increases, these requirements will increase linearly for each participant bringing the scalability of the system into question. Hence, the vanilla CaPriDe learning framework is more appropriate for a small network of participants. However, there are ways to improve scalability borrowing on ideas from cross-device FL. For example, instead of computing the pairwise distillation loss with respect to each participant, a random subset of participants can be sampled in each round for distillation. Similarly, the computational complexity for each participant can be significantly reduced by performing encrypted inference only on a randomly selected subset of encrypted samples. Furthermore, these actions can be determined at the local level without any central orchestration. The need to have FHE-friendly models for encrypted inference is a key limitation, which precludes the use of very deep models within the proposed framework. One potential solution is to employ two models at each participant: one deep model with large capacity and another smaller FHE-friendly model for encrypted inference on other participants' data and perform an additional mutual distillation between these two models after each round. + +# 5. Summary + +Collaborative learning without compromising on data confidentiality and privacy is a valuable tool in applications with stringent data privacy requirements. We have proposed a new framework called CaPriDe learning based on inference in the FHE domain to achieve this goal. Knowledge distillation is used to exchange useful information between models, and a new distillation loss based on approximate KL divergence has been proposed to enable secure loss computation in the encrypted domain. While the proposed approach has shown promising results, its efficiency can be further improved. Other vulnerabilities such as poisoning attacks will also be addressed in the future. + +# References + +[1] Ehud Aharoni, Allon Adir, Moran Baruch, Nir Drucker, Gilad Ezov, Ariel Farkash, Lev Greenberg, Rami Masalha, Guy Moshkowish, Dov Murik, Hayim Shaul, and Omri Soceanu. HeLayers: A Tile Tensors Framework for Large Neural Networks on Encrypted Data. arXiv:2011.01805v2, 2021. 6 +[2] Ehud Aharoni, Nir Drucker, Gilad Ezov, Hayim Shaul, and Omri Soceanu. Complex encoded tile tensors: Accelerating encrypted analytics. In IEEE S&P 20 (5), pages 35-43, 2022. 6 +[3] Keith Bonawitz, Vladimir Ivanov, Ben Kreuter, Antonio Marcedone, H. Brendan McMahan, Sarvar Patel, Daniel Ramage, Aaron Segal, and Karn Seth. Practical secure aggregation for privacy preserving machine learning. ACM Conference on Computer and Communications Security (ACM CCS), 2017. 2 +[4] Alon Brutzkus, Ran Gilad-Bachrach, and Oren Elisha. Low latency privacy preserving inference. In International Conference on Machine Learning, pages 812-821, 2019. 8 +[5] Hongyan Chang, Virat Shejwalkar, Reza Shokri, and Amir Houmansadr. Cronus: Robust and heterogeneous collaborative learning with black-box knowledge transfer. In arXiv preprint arXiv:1912.11279, 2019. 2 +[6] Jung Hee Cheon, Andrey Kim, Miran Kim, and Yongsoo Song. Homomorphic Encryption for Arithmetic of Approximate Numbers. 2016. 3 +[7] Christopher A. Choquette-Choo, Natalie Dullerud, Adam Dziedzic, Yunxiang Zhang, Somesh Jha, Nicolas Papernot, and Xiao. Wang. CaPC Learning: Confidential and Private Collaborative Learning. Proceedings of ICLR, 2021. 2 +[8] Deshpande, Nikhil and Shiffman, Gary M. Federated learning: The new thing in ai/ml for detecting financial crimes and managing risk. https://morningconsult.com/opinions/federated-learning-the-new-thing-in-ai-ml-for-detecting-financial-crimes-and-managing-risk/, 2021.1 +[9] Jonas Geiping, Hartmut Bauermeister, Hannah Dröge, and Michael. Moeller. Inverting gradients-How easy is it to break privacy in federated learning? Advances in Neural Information Processing Systems (NeurIPS), 33, 2020. 1, 5 +[10] Craig. Gentry. Fully Homomorphic Encryption. PhD thesis, 2009. 3 +[11] Qiushan Guo, Xinjiang Wang, Yichao Wu, Zhipeng Yu, Ding Liang, Xiaolin Hu, and Ping. Luo. Online Knowledge Distillation via Collaborative Learning. Proceedings of CVPR, 2020. 2 +[12] Jenny Hamer, Mehryar Mohri, and Ananda Theertha Suresh. FedBoost: A Communication-Efficient Algorithm for Federated Learning. Proceedings of the 37th International Conference on Machine Learning, PMLR, 119:3973-3983, 2020. 2 +[13] Geoffrey Hinton, Oriol Vinyls, and Jeff Dean. Distilling the Knowledge in a Neural Network. NeurIPS, 2014. 2, 4 +[14] Jiwnoo Jeon, Kangwook Lee, Sewoong Oh, and Jungseul. Ok. Gradient Inversion with Generative Image Prior. Advances in Neural Information Processing Systems (NeurIPS), 34, 2021. 1, 5 + +[15] Peter Kairouz, H. Brendan McMahan, Brendan Avent, Aurélien Bellet, and Mehdi. Bennis. Advances and Open Problems in Federated Learning. Now Publishers, 2021. 1 +[16] Georgios A. Kaissis, Marcus R. Makowski, Daniel Rückert, and Rickmer F. Braren. Secure, privacy-preserving and federated machine learning in medical imaging. Nature Machine Intelligence, 2:305-311, 2020. 1 +[17] Alex Krizhevsky. Gradient-based learning applied to document recognition. Technical report, University of Toronto, 2009. 6 +[18] Tao Lin, Lingjing Kong, Sebastian U Stich, and Martin Jaggi. Ensemble distillation for robust model fusion in federated learning. In Advances in Neural Information Processing Systems, volume 33, pages 2351-2363, 2020. 2 +[19] Chuan Ma, Jun Li, Ming Ding, Howard H. Yang, Feng Shu, Tony Q. S. Quek, and Tony Q. S. Poor. On safeguarding privacy and security in the framework of federated learning. IEEE Network, pages 242-248, 2020. 2 +[20] H. Brendan McMahan, Eider Moore, Daniel Ramage, Seth Hampson, and Blaise Agüera y. Arcas. Communication-Efficient Learning of Deep Networks from Decentralized Data. Proceedings of the 20th International Conference on Artificial Intelligence and Statistics (AISTATS), 2016. 2, 6 +[21] Fan Mo, Hamed Haddadi, Kleomenis Katevas, Eduard Marin, Diego Perino, and Nicolas Kourtellis. PPFL: Privacy-Preserving Federated Learning with Trusted Execution Environments. In Proceedings of the 19th Annual International Conference on Mobile Systems, Applications, and Services (MobiSys), page 94-108, 2021. 1 +[22] Willem G van Panhuis, Proma Paul, Claudia Emerson, John Grefenstette, Richard Wilder, Abraham J Herbst, and Donald S Heymann, David Burke. A systematic review of barriers to data sharing in public health. BMC Public Health, 2014. 1 +[23] Micah J. Sheller, Brandon Edwards, G. Anthony Reina, Jason Martin, Sarthak Pati, Aikaterini Kotrotsou, Mikhail Milchenko, Weilin Xu, Daniel Marcus, Rivka R. Colen, and Spyridon. Bakas. Federated learning in medicine: facilitating multi-institutional collaborations without sharing patient data. Nature Scientific Reports, 10, 2020. 1 +[24] Kang Wei, Jun Li, Ming Ding, Chuan Ma, Howard H. Yang, and Farhad. Farokhi. Federated Learning With Differential Privacy: Algorithms and Performance Analysis. IEEE Transactions on Information Forensics and Security, 15:3454-3469, 2020. 1, 6 +[25] Jaehong Yoon, Wonyong Jeong, Giwoong Lee, Eunho Yang, and Sung Ju. Hwang. Federated Continual Learning with Weighted Inter-client Transfer. Proceedings of the 37th International Conference on Machine Learning, PMLR, 139:12073-12086, 2021. 2 +[26] Chengliang Zhang, Suyi Li, Junzhe Xia, Wei Wang, Feng Yan, and Yang Liu. BatchCrypt: efficient homomorphic encryption for cross-silo federated learning. Proceedings of the USENIX Conference, page 493-506, 2020. 1 +[27] Ying Zhang, Tao Xiang, Timothy M. Hospedales, and Huchuan Lu. Deep Mutual Learning. CVPR, 2017. 2, 6 \ No newline at end of file diff --git a/2023/CaPriDe Learning_ Confidential and Private Decentralized Learning Based on Encryption-Friendly Distillation Loss/images.zip b/2023/CaPriDe Learning_ Confidential and Private Decentralized Learning Based on Encryption-Friendly Distillation Loss/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..683c4a9fdc5c23c742fd7afe64fb79d5260fda98 --- /dev/null +++ b/2023/CaPriDe Learning_ Confidential and Private Decentralized Learning Based on Encryption-Friendly Distillation Loss/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8a0f04df8f81060959fc94749be2a46c693e6eedd3e6412e1f8835bcbdd7e34c +size 410384 diff --git a/2023/CaPriDe Learning_ Confidential and Private Decentralized Learning Based on Encryption-Friendly Distillation Loss/layout.json b/2023/CaPriDe Learning_ Confidential and Private Decentralized Learning Based on Encryption-Friendly Distillation Loss/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..903bd50e20aaf839d0da7fe6a6eb5d510fa6fd5f --- /dev/null +++ b/2023/CaPriDe Learning_ Confidential and Private Decentralized Learning Based on Encryption-Friendly Distillation Loss/layout.json @@ -0,0 +1,10160 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 56, + 102, + 538, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 102, + 538, + 140 + ], + "spans": [ + { + "bbox": [ + 56, + 102, + 538, + 140 + ], + "type": "text", + "content": "CaPriDe Learning: Confidential and Private Decentralized Learning based on Encryption-friendly Distillation Loss" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 144, + 161, + 449, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 144, + 161, + 449, + 203 + ], + "spans": [ + { + "bbox": [ + 144, + 161, + 449, + 203 + ], + "type": "text", + "content": "Nurbek Tastan Karthik Nandakumar \nMohamed bin Zayed University of Artificial Intelligence, UAE \n{nurbek.tastan,karthik.nandakumar}@mbzuai.ac.ae" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 143, + 231, + 192, + 243 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 231, + 192, + 243 + ], + "spans": [ + { + "bbox": [ + 143, + 231, + 192, + 243 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 258, + 289, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 258, + 289, + 628 + ], + "spans": [ + { + "bbox": [ + 46, + 258, + 289, + 628 + ], + "type": "text", + "content": "Large volumes of data required to train accurate deep neural networks (DNNs) are seldom available with any single entity. Often, privacy concerns prevent entities from sharing data with each other or with a third-party learning service provider. While cross-silo federated learning (FL) allows collaborative learning of large DNNs without sharing the data itself, most existing cross-silo FL algorithms have an unacceptable utility-privacy trade-off. In this work, we propose a framework called Confidential and Private Decentralized (CaPriDe) learning, which optimally leverages the power of fully homomorphic encryption (FHE) to enable collaborative learning without compromising on the confidentiality and privacy of data. In CaPriDe learning, participating entities release their private data in an encrypted form allowing other participants to perform inference in the encrypted domain. The crux of CaPriDe learning is mutual knowledge distillation between multiple local models through a novel distillation loss, which is an approximation of the Kullback-Leibler (KL) divergence between the local predictions and encrypted inferences of other participants on the same data that can be computed in the encrypted domain. Extensive experiments on three datasets show that CaPriDe learning can improve the accuracy of local models without any central coordination, provide strong guarantees of data confidentiality and privacy, and has the ability to handle statistical heterogeneity. Constraints on the model architecture (arising from the need to be FHE-friendly), limited scalability, and computational complexity of encrypted domain inference are the main limitations of the proposed approach. The code can be found at https://github.com/tnurbek/capride-learning." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 656, + 128, + 668 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 656, + 128, + 668 + ], + "spans": [ + { + "bbox": [ + 47, + 656, + 128, + 668 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 677, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 287, + 715 + ], + "type": "text", + "content": "Rapid strides have been made in machine learning (ML) (in particular, deep learning) over the past decade. However, in many important application domains such as health-" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 232, + 547, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 232, + 547, + 460 + ], + "spans": [ + { + "bbox": [ + 304, + 232, + 547, + 460 + ], + "type": "text", + "content": "care and finance, the absence of large, centralized datasets is a significant obstacle to realizing the full benefits of deep learning algorithms. Data in these applications often resides in silos and is governed by strict regulations (e.g., HIPAA, GDPR, etc.) because of its privacy sensitive nature [22]. Competing business interests of data owners and the lack of appropriate incentives for data sharing further accentuate the problem. To overcome these issues, there is a need for collaborative learning algorithms that ideally satisfy the following requirements [15]: (i) confidentiality - no sharing of raw data, (ii) privacy - minimal leakage of information via the knowledge exchange mechanism (e.g., gradients or predictions), (iii) utility - gain in accuracy (over the individual models) resulting from the collaboration, even in the presence of statistical heterogeneity, (iv) efficiency - minimize computational complexity and communication burden, (v) robustness - handle unintentional failures and attacks emanating from malicious entities, and (vi) fairness - utility should be proportional to the individual contributions." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 462, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 462, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 462, + 547, + 715 + ], + "type": "text", + "content": "Federated learning (FL) [15] is a special case of collaborative learning, which works under the orchestration of a central server. FL allows multiple entities to collaboratively solve a ML problem by sharing of focused updates (e.g., gradients), instead of raw data. Specifically, cross-silo FL (typically between 2-100 participants) has been touted as a promising solution to address the data fragmentation problem in finance [8] and healthcare [16,23]. Most prior cross-silo FL algorithms assume that all the parties are collectively training a single model with a common architecture, which is too restrictive in practice. Furthermore, knowledge exchange usually happens through sharing of gradients or model parameters. Recent gradient inversion attacks [9, 14] demonstrate that it is indeed possible to recover high fidelity information from individual gradient updates, thus violating the privacy requirement. While differential privacy (DP) [24], secure multi-party computation (MPC) [26], and trusted execution environment (TEE) [21] have been proposed as potential remedies to safeguard privacy in FL, none of the existing solutions offer an acceptable privacy-utility trade-off with reasonable efficiency. As" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8084" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "text", + "content": "noted in [5], sharing of high dimensional gradients/model parameters is a fundamental privacy limitation of standard FL methods, which cannot be addressed by simply wrapping around a DP, MPC, or TEE solution." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 123, + 289, + 387 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 123, + 289, + 387 + ], + "spans": [ + { + "bbox": [ + 46, + 123, + 289, + 387 + ], + "type": "text", + "content": "Confidential and private collaborative learning (CaPC) [7] is the only method that claims to achieve both confidentiality and privacy in a collaborative setting. In CaPC learning, each participant is able to query other parties via a private inference protocol based on 2-party computation (2PC). However, the answering parties cannot directly reveal the inference logits to the querying party because it leaks information about their local models (e.g., through model extraction attacks). To circumvent this problem, differential privacy, private aggregation of teacher ensembles, and secure argmax computation through a central privacy guardian (PG) are employed to output only the predicted label to the querying party. A drawback of the CaPC approach is that it achieves all privacy guarantees using the help of a semi-trusted PG. Moreover, the use of differential privacy reduces the performance of FL algorithms, especially if strong privacy guarantees are required. Finally, the use of MPC requires multiple rounds of communication between the parties for each query. All other approaches that have been proposed to achieve knowledge transfer through distillation [18] require a non-private labeled/unlabeled/synthetic dataset that can be shared among participants." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 391, + 289, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 391, + 289, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 391, + 289, + 714 + ], + "type": "text", + "content": "In this work, we propose a new protocol called Confidential and Private Decentralized (CaPriDe) learning, where participants learn from each other in a collaborative setting while preserving their confidentiality and privacy (see Figure 1). Unlike the 2PC protocols used in CaPC, we leverage fully homomorphic encryption (FHE) to enable participants to publish their encrypted data. Since the published data is encrypted using the data owner's public key and only the owner can decrypt the data (using the secret key), confidentiality is preserved. However, the CaPriDe framework allows other collaborators to perform encrypted inference on the published (encrypted) data by applying their own local model. Mutual knowledge distillation (KD) [13] between the data owner's local model and the local models of the collaborators is used to transfer knowledge between models and ensure a collaborative gain. KD is typically achieved through minimizing the distillation loss between the student and teacher responses (logits). Since the collaborators in CaPriDe learning make predictions on encrypted data, a loss function that can be computed without any decryption is required. Hence, we propose a new distillation loss, which is an approximation of the KL divergence that can be securely computed. Only an encrypted value of the distillation loss aggregated over an entire batch is sent back to the data owner, which ensures strong privacy. Data owners can decrypt this aggregate loss value and use it for updating their local models. Our contributions are:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 317, + 72, + 545, + 271 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 317, + 72, + 545, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 72, + 545, + 133 + ], + "spans": [ + { + "bbox": [ + 317, + 72, + 545, + 133 + ], + "type": "text", + "content": "- We introduce the CaPriDe learning framework, which exploits FHE-based encrypted inference and knowledge distillation to achieve confidential and private collaborative learning without any central orchestration and any need for non-private shared data." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 317, + 142, + 545, + 201 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 142, + 545, + 201 + ], + "spans": [ + { + "bbox": [ + 317, + 142, + 545, + 201 + ], + "type": "text", + "content": "- To enable CaPriDe learning, we propose an encryption-friendly distillation loss that estimates the approximate KL divergence between two model predictions and design a protocol to securely compute this loss in the encrypted domain." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 317, + 212, + 545, + 271 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 212, + 545, + 271 + ], + "spans": [ + { + "bbox": [ + 317, + 212, + 545, + 271 + ], + "type": "text", + "content": "- We conduct extensive experiments to show that CaPriDe learning enables participants to achieve collaborative gain, even in the non-iid setting. To prove feasibility, we implement encrypted inference using the Tile Tensors library with a FHE-friendly model." + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 284, + 384, + 297 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 284, + 384, + 297 + ], + "spans": [ + { + "bbox": [ + 306, + 284, + 384, + 297 + ], + "type": "text", + "content": "2. Background" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 304, + 545, + 591 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 304, + 545, + 591 + ], + "spans": [ + { + "bbox": [ + 304, + 304, + 545, + 591 + ], + "type": "text", + "content": "Federated learning (FL) was first proposed in [20] as a distributed ML algorithm that does not use user data directly. Subsequent works attempted to address various challenges in FL, including privacy [7, 19], communication efficiency [12], convergence, catastrophic forgetting [25], and secure aggregation [3]. Though several privacy-enhancing solutions have been proposed, most of them fail to provide a satisfactory balance between accuracy and privacy. Only recently, some promising solutions such as CaPC have been proposed [7]. These advancements have been made by jettisoning the conventional gradient sharing approach and borrowing ideas from the knowledge distillation literature [13]. Specifically, it has been shown that it is possible for an ML model to benefit from predictions made by other models, by aligning the local predictions with those obtained from external parties. Deep mutual learning (DML) [27] is an important work in this direction, where models learn collaboratively and teach each other throughout the training process. Compared to the distillation by a pre-defined static teacher network, DML achieves better performance, and mutually learned models outperform independently trained models irrespective of model capacity. Knowledge Distillation via Collaborative Learning (KDCL) [11], Cronus [5], and Ensemble Distillation [18] also follow a similar idea." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 600, + 479, + 613 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 600, + 479, + 613 + ], + "spans": [ + { + "bbox": [ + 306, + 600, + 479, + 613 + ], + "type": "text", + "content": "2.1. Fully Homomorphic Encryption" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 619, + 545, + 655 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 619, + 545, + 655 + ], + "spans": [ + { + "bbox": [ + 304, + 619, + 545, + 655 + ], + "type": "text", + "content": "Let " + }, + { + "bbox": [ + 304, + 619, + 545, + 655 + ], + "type": "inline_equation", + "content": "\\mathcal{E}(x)" + }, + { + "bbox": [ + 304, + 619, + 545, + 655 + ], + "type": "text", + "content": " denote the encrypted value of " + }, + { + "bbox": [ + 304, + 619, + 545, + 655 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 304, + 619, + 545, + 655 + ], + "type": "text", + "content": " using the public key " + }, + { + "bbox": [ + 304, + 619, + 545, + 655 + ], + "type": "inline_equation", + "content": "\\kappa" + }, + { + "bbox": [ + 304, + 619, + 545, + 655 + ], + "type": "text", + "content": ". The encryption scheme is called homomorphic over an operation " + }, + { + "bbox": [ + 304, + 619, + 545, + 655 + ], + "type": "inline_equation", + "content": "\\oplus" + }, + { + "bbox": [ + 304, + 619, + 545, + 655 + ], + "type": "text", + "content": " if it supports the following equation:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 346, + 666, + 504, + 678 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 346, + 666, + 504, + 678 + ], + "spans": [ + { + "bbox": [ + 346, + 666, + 504, + 678 + ], + "type": "interline_equation", + "content": "\\mathcal {E} \\left(x _ {1} \\oplus x _ {2}\\right) = \\mathcal {E} \\left(x _ {1}\\right) \\otimes \\mathcal {E} \\left(x _ {2}\\right), \\forall x _ {1}, x _ {2}", + "image_path": "87559f84da69f6c49d8a89a5fd5c6a4f7720682164bcdfbfe28cca3f20323177.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 689, + 545, + 713 + ], + "type": "text", + "content": "where operators " + }, + { + "bbox": [ + 304, + 689, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\otimes" + }, + { + "bbox": [ + 304, + 689, + 545, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 689, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\oplus" + }, + { + "bbox": [ + 304, + 689, + 545, + 713 + ], + "type": "text", + "content": " need not be the same. The scheme is fully homomorphic if for some arbitrary function" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8085" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 97, + 74, + 497, + 250 + ], + "blocks": [ + { + "bbox": [ + 97, + 74, + 497, + 250 + ], + "lines": [ + { + "bbox": [ + 97, + 74, + 497, + 250 + ], + "spans": [ + { + "bbox": [ + 97, + 74, + 497, + 250 + ], + "type": "image", + "image_path": "2c55aebc2280679ccc1d83c904c96b3cbc3fbb805415b830bcdb37fb1a6a6dcd.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 262, + 547, + 319 + ], + "lines": [ + { + "bbox": [ + 46, + 262, + 547, + 319 + ], + "spans": [ + { + "bbox": [ + 46, + 262, + 547, + 319 + ], + "type": "text", + "content": "Figure 1. Illustration of confidential and private decentralized (CaPriDe) learning framework for " + }, + { + "bbox": [ + 46, + 262, + 547, + 319 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 46, + 262, + 547, + 319 + ], + "type": "text", + "content": " participants, showing the learning process for only participant " + }, + { + "bbox": [ + 46, + 262, + 547, + 319 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_1" + }, + { + "bbox": [ + 46, + 262, + 547, + 319 + ], + "type": "text", + "content": ". Here, " + }, + { + "bbox": [ + 46, + 262, + 547, + 319 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_1 = (X_1, Y_1) = \\{x_{j,1}, y_{j,1}\\}_{j=1}^{N_1}" + }, + { + "bbox": [ + 46, + 262, + 547, + 319 + ], + "type": "text", + "content": " is the local data of " + }, + { + "bbox": [ + 46, + 262, + 547, + 319 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_1" + }, + { + "bbox": [ + 46, + 262, + 547, + 319 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 262, + 547, + 319 + ], + "type": "inline_equation", + "content": "\\mathcal{E}(X_1)" + }, + { + "bbox": [ + 46, + 262, + 547, + 319 + ], + "type": "text", + "content": " denotes the collection of encrypted unlabeled samples of " + }, + { + "bbox": [ + 46, + 262, + 547, + 319 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_1" + }, + { + "bbox": [ + 46, + 262, + 547, + 319 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 262, + 547, + 319 + ], + "type": "inline_equation", + "content": "\\pmb{p}_{(1,l)}" + }, + { + "bbox": [ + 46, + 262, + 547, + 319 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 262, + 547, + 319 + ], + "type": "inline_equation", + "content": "\\pmb{z}_{(1,l)}" + }, + { + "bbox": [ + 46, + 262, + 547, + 319 + ], + "type": "text", + "content": " are the prediction probabilities and logits obtained by applying model " + }, + { + "bbox": [ + 46, + 262, + 547, + 319 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_{\\theta_l}" + }, + { + "bbox": [ + 46, + 262, + 547, + 319 + ], + "type": "text", + "content": " of participant " + }, + { + "bbox": [ + 46, + 262, + 547, + 319 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_l" + }, + { + "bbox": [ + 46, + 262, + 547, + 319 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 46, + 262, + 547, + 319 + ], + "type": "inline_equation", + "content": "X_1" + }, + { + "bbox": [ + 46, + 262, + 547, + 319 + ], + "type": "text", + "content": ". Black dashed line represents exchange of encrypted data between participants in each round, and blue dashed line denotes a single transfer at the beginning of the protocol. " + }, + { + "bbox": [ + 46, + 262, + 547, + 319 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{CE}" + }, + { + "bbox": [ + 46, + 262, + 547, + 319 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 262, + 547, + 319 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{DL}" + }, + { + "bbox": [ + 46, + 262, + 547, + 319 + ], + "type": "text", + "content": " represent the cross-entropy and distillation losses, respectively." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 339, + 288, + 363 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 339, + 288, + 363 + ], + "spans": [ + { + "bbox": [ + 47, + 339, + 288, + 363 + ], + "type": "inline_equation", + "content": "f(x_{1},x_{2},\\ldots ,x_{n})" + }, + { + "bbox": [ + 47, + 339, + 288, + 363 + ], + "type": "text", + "content": " , there exists a function " + }, + { + "bbox": [ + 47, + 339, + 288, + 363 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 47, + 339, + 288, + 363 + ], + "type": "text", + "content": " and decryption method " + }, + { + "bbox": [ + 47, + 339, + 288, + 363 + ], + "type": "inline_equation", + "content": "^d" + }, + { + "bbox": [ + 47, + 339, + 288, + 363 + ], + "type": "text", + "content": " such that:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 63, + 369, + 271, + 384 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 369, + 271, + 384 + ], + "spans": [ + { + "bbox": [ + 63, + 369, + 271, + 384 + ], + "type": "interline_equation", + "content": "f \\left(x _ {1}, x _ {2}, \\dots , x _ {n}\\right) = d \\left(g \\left(\\mathcal {E} \\left(x _ {1}\\right), \\mathcal {E} \\left(x _ {2}\\right), \\dots , \\mathcal {E} \\left(x _ {n}\\right)\\right)\\right),", + "image_path": "8e0b0b8edfa875c81bfabe2ae59c536a407e132beb278ae89ac81b14365c3fa2.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 389, + 287, + 462 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 389, + 287, + 462 + ], + "spans": [ + { + "bbox": [ + 46, + 389, + 287, + 462 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 389, + 287, + 462 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 46, + 389, + 287, + 462 + ], + "type": "text", + "content": " uses secret key " + }, + { + "bbox": [ + 46, + 389, + 287, + 462 + ], + "type": "inline_equation", + "content": "\\tilde{\\kappa}" + }, + { + "bbox": [ + 46, + 389, + 287, + 462 + ], + "type": "text", + "content": " to decrypt the given ciphertext value. Gentry [10] first constructed somewhat homomorphic encryption (SWHE) and later developed bootstrapping of ciphertexts to reduce noise. Among the various FHE schemes such as BGV, BFV, and CKKS [6], only CKKS works with real numbers and approximate computations." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 467, + 133, + 479 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 467, + 133, + 479 + ], + "spans": [ + { + "bbox": [ + 47, + 467, + 133, + 479 + ], + "type": "text", + "content": "2.2. Preliminaries" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 485, + 287, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 485, + 287, + 594 + ], + "spans": [ + { + "bbox": [ + 46, + 485, + 287, + 594 + ], + "type": "text", + "content": "Typically, a supervised classifier " + }, + { + "bbox": [ + 46, + 485, + 287, + 594 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 46, + 485, + 287, + 594 + ], + "type": "text", + "content": " is mapping from the input space " + }, + { + "bbox": [ + 46, + 485, + 287, + 594 + ], + "type": "inline_equation", + "content": "\\pmb{x} \\in \\mathcal{X} \\subset \\mathbb{R}^n" + }, + { + "bbox": [ + 46, + 485, + 287, + 594 + ], + "type": "text", + "content": " to the label space " + }, + { + "bbox": [ + 46, + 485, + 287, + 594 + ], + "type": "inline_equation", + "content": "y \\in \\mathcal{Y} = \\{1,2,\\dots,M\\}" + }, + { + "bbox": [ + 46, + 485, + 287, + 594 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 485, + 287, + 594 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 46, + 485, + 287, + 594 + ], + "type": "text", + "content": " is the number of classes. For convenience, we assume that " + }, + { + "bbox": [ + 46, + 485, + 287, + 594 + ], + "type": "inline_equation", + "content": "\\mathcal{M}: \\mathcal{X} \\to \\mathcal{Z}" + }, + { + "bbox": [ + 46, + 485, + 287, + 594 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 485, + 287, + 594 + ], + "type": "inline_equation", + "content": "\\mathcal{Z}" + }, + { + "bbox": [ + 46, + 485, + 287, + 594 + ], + "type": "text", + "content": " denotes the logits space (" + }, + { + "bbox": [ + 46, + 485, + 287, + 594 + ], + "type": "inline_equation", + "content": "\\mathcal{Z} \\subset \\mathbb{R}^M" + }, + { + "bbox": [ + 46, + 485, + 287, + 594 + ], + "type": "text", + "content": "). In other words, " + }, + { + "bbox": [ + 46, + 485, + 287, + 594 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 46, + 485, + 287, + 594 + ], + "type": "text", + "content": " maps the input " + }, + { + "bbox": [ + 46, + 485, + 287, + 594 + ], + "type": "inline_equation", + "content": "\\pmb{x}" + }, + { + "bbox": [ + 46, + 485, + 287, + 594 + ], + "type": "text", + "content": " to a " + }, + { + "bbox": [ + 46, + 485, + 287, + 594 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 46, + 485, + 287, + 594 + ], + "type": "text", + "content": "-dimensional logits vector " + }, + { + "bbox": [ + 46, + 485, + 287, + 594 + ], + "type": "inline_equation", + "content": "\\pmb{z} = [z^{1},z^{2},\\dots,z^{M}]" + }, + { + "bbox": [ + 46, + 485, + 287, + 594 + ], + "type": "text", + "content": ". A softmax function " + }, + { + "bbox": [ + 46, + 485, + 287, + 594 + ], + "type": "inline_equation", + "content": "(\\sigma)" + }, + { + "bbox": [ + 46, + 485, + 287, + 594 + ], + "type": "text", + "content": " can be applied to the logits vector " + }, + { + "bbox": [ + 46, + 485, + 287, + 594 + ], + "type": "inline_equation", + "content": "\\pmb{z}" + }, + { + "bbox": [ + 46, + 485, + 287, + 594 + ], + "type": "text", + "content": " to obtain the probability distribution " + }, + { + "bbox": [ + 46, + 485, + 287, + 594 + ], + "type": "inline_equation", + "content": "\\pmb{p} = \\sigma(\\pmb{z}) = [p^{1},p^{2},\\dots,p^{M}]" + }, + { + "bbox": [ + 46, + 485, + 287, + 594 + ], + "type": "text", + "content": " over " + }, + { + "bbox": [ + 46, + 485, + 287, + 594 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 46, + 485, + 287, + 594 + ], + "type": "text", + "content": " labels, i.e.," + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 125, + 599, + 287, + 623 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 125, + 599, + 287, + 623 + ], + "spans": [ + { + "bbox": [ + 125, + 599, + 287, + 623 + ], + "type": "interline_equation", + "content": "p ^ {m} = \\frac {\\exp (z ^ {m} / T)}{C}, \\tag {1}", + "image_path": "f7a0dd4e5369b6da003a3f2e468c7d48c7124bea744d2f7878136f995c465ee5.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 628, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 628, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 628, + 288, + 715 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 628, + 288, + 715 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 46, + 628, + 288, + 715 + ], + "type": "text", + "content": " is the softmax temperature parameter and " + }, + { + "bbox": [ + 46, + 628, + 288, + 715 + ], + "type": "inline_equation", + "content": "C = \\sum_{i=1}^{M} \\exp(z^i / T)" + }, + { + "bbox": [ + 46, + 628, + 288, + 715 + ], + "type": "text", + "content": " is a normalization constant that ensures " + }, + { + "bbox": [ + 46, + 628, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\sum_{m=1}^{M} p^m = 1" + }, + { + "bbox": [ + 46, + 628, + 288, + 715 + ], + "type": "text", + "content": ". Here, " + }, + { + "bbox": [ + 46, + 628, + 288, + 715 + ], + "type": "inline_equation", + "content": "p^m" + }, + { + "bbox": [ + 46, + 628, + 288, + 715 + ], + "type": "text", + "content": " denotes the probability that input " + }, + { + "bbox": [ + 46, + 628, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 46, + 628, + 288, + 715 + ], + "type": "text", + "content": " belongs to class " + }, + { + "bbox": [ + 46, + 628, + 288, + 715 + ], + "type": "inline_equation", + "content": "m, m = 1, 2, \\dots, M" + }, + { + "bbox": [ + 46, + 628, + 288, + 715 + ], + "type": "text", + "content": ". Given two " + }, + { + "bbox": [ + 46, + 628, + 288, + 715 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 46, + 628, + 288, + 715 + ], + "type": "text", + "content": "-dimensional vectors " + }, + { + "bbox": [ + 46, + 628, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_1, \\mathbf{v}_2 \\in \\mathbb{R}^V" + }, + { + "bbox": [ + 46, + 628, + 288, + 715 + ], + "type": "text", + "content": ", let " + }, + { + "bbox": [ + 46, + 628, + 288, + 715 + ], + "type": "inline_equation", + "content": "(\\mathbf{v}_1 \\cdot \\mathbf{v}_2)" + }, + { + "bbox": [ + 46, + 628, + 288, + 715 + ], + "type": "text", + "content": " denote the dot product, " + }, + { + "bbox": [ + 46, + 628, + 288, + 715 + ], + "type": "inline_equation", + "content": "(\\mathbf{v}_1 \\odot \\mathbf{v}_2)" + }, + { + "bbox": [ + 46, + 628, + 288, + 715 + ], + "type": "text", + "content": " denote the element-wise product, and " + }, + { + "bbox": [ + 46, + 628, + 288, + 715 + ], + "type": "inline_equation", + "content": "f(\\mathbf{v}_1)" + }, + { + "bbox": [ + 46, + 628, + 288, + 715 + ], + "type": "text", + "content": " represent the vector obtained by applying" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 339, + 545, + 364 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 339, + 545, + 364 + ], + "spans": [ + { + "bbox": [ + 305, + 339, + 545, + 364 + ], + "type": "text", + "content": "the function " + }, + { + "bbox": [ + 305, + 339, + 545, + 364 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 305, + 339, + 545, + 364 + ], + "type": "text", + "content": " to each element in " + }, + { + "bbox": [ + 305, + 339, + 545, + 364 + ], + "type": "inline_equation", + "content": "\\pmb{v}_1" + }, + { + "bbox": [ + 305, + 339, + 545, + 364 + ], + "type": "text", + "content": ". The cross entropy loss of model " + }, + { + "bbox": [ + 305, + 339, + 545, + 364 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 305, + 339, + 545, + 364 + ], + "type": "text", + "content": " on a sample " + }, + { + "bbox": [ + 305, + 339, + 545, + 364 + ], + "type": "inline_equation", + "content": "(x, y)" + }, + { + "bbox": [ + 305, + 339, + 545, + 364 + ], + "type": "text", + "content": " is defined as:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 368, + 369, + 484, + 383 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 368, + 369, + 484, + 383 + ], + "spans": [ + { + "bbox": [ + 368, + 369, + 484, + 383 + ], + "type": "interline_equation", + "content": "L _ {C E} (\\boldsymbol {p}, y) = - \\mathbf {1} _ {y} \\cdot l o g (\\boldsymbol {p}),", + "image_path": "a6586d9e71085694f5ee57cbfb271211dace07323df245661549f99fdabc935d.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 388, + 545, + 425 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 388, + 545, + 425 + ], + "spans": [ + { + "bbox": [ + 305, + 388, + 545, + 425 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 388, + 545, + 425 + ], + "type": "inline_equation", + "content": "\\pmb{p} = \\sigma(\\mathcal{M}(\\pmb{x}))" + }, + { + "bbox": [ + 305, + 388, + 545, + 425 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 388, + 545, + 425 + ], + "type": "inline_equation", + "content": "\\mathbf{1}_y" + }, + { + "bbox": [ + 305, + 388, + 545, + 425 + ], + "type": "text", + "content": " is an indicator vector whose " + }, + { + "bbox": [ + 305, + 388, + 545, + 425 + ], + "type": "inline_equation", + "content": "y^{th}" + }, + { + "bbox": [ + 305, + 388, + 545, + 425 + ], + "type": "text", + "content": " element is 1 and other elements are 0. The KL divergence between two distributions " + }, + { + "bbox": [ + 305, + 388, + 545, + 425 + ], + "type": "inline_equation", + "content": "\\pmb{p}_1" + }, + { + "bbox": [ + 305, + 388, + 545, + 425 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 388, + 545, + 425 + ], + "type": "inline_equation", + "content": "\\pmb{p}_2" + }, + { + "bbox": [ + 305, + 388, + 545, + 425 + ], + "type": "text", + "content": " is given by:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 430, + 556, + 472 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 430, + 556, + 472 + ], + "spans": [ + { + "bbox": [ + 306, + 430, + 556, + 472 + ], + "type": "interline_equation", + "content": "D _ {K L} \\left(\\boldsymbol {p} _ {2} \\| \\boldsymbol {p} _ {1}\\right) = \\sum_ {m = 1} ^ {M} p _ {2} ^ {m} \\log \\frac {p _ {2} ^ {m}}{p _ {1} ^ {m}} = \\left(\\boldsymbol {p} _ {2} \\cdot \\log \\boldsymbol {p} _ {2}\\right) - \\left(\\boldsymbol {p} _ {2} \\cdot \\log \\boldsymbol {p} _ {1}\\right). \\tag {2}", + "image_path": "80c5ee2b1b63daa06da892936b5f89adcb26919d5716285a60d1974f20a9a26f.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 473, + 533, + 485 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 473, + 533, + 485 + ], + "spans": [ + { + "bbox": [ + 306, + 473, + 533, + 485 + ], + "type": "text", + "content": "The L2 distance between two logit vectors " + }, + { + "bbox": [ + 306, + 473, + 533, + 485 + ], + "type": "inline_equation", + "content": "z_{1}" + }, + { + "bbox": [ + 306, + 473, + 533, + 485 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 306, + 473, + 533, + 485 + ], + "type": "inline_equation", + "content": "z_{2}" + }, + { + "bbox": [ + 306, + 473, + 533, + 485 + ], + "type": "text", + "content": " is:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 490, + 545, + 522 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 490, + 545, + 522 + ], + "spans": [ + { + "bbox": [ + 306, + 490, + 545, + 522 + ], + "type": "interline_equation", + "content": "D _ {L 2} \\left(\\boldsymbol {z} _ {2}, \\boldsymbol {z} _ {1}\\right) = \\sum_ {m = 1} ^ {M} \\left(z _ {2} ^ {m} - z _ {1} ^ {m}\\right) ^ {2} = \\left(\\boldsymbol {z} _ {2} - \\boldsymbol {z} _ {1}\\right) \\cdot \\left(\\boldsymbol {z} _ {2} - \\boldsymbol {z} _ {1}\\right).", + "image_path": "6e50a330af2ad733151a2db5ebe758b4110faa0d59cf94a4346016bd9d5c3450.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 528, + 432, + 541 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 528, + 432, + 541 + ], + "spans": [ + { + "bbox": [ + 306, + 528, + 432, + 541 + ], + "type": "text", + "content": "3. Proposed Framework" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 547, + 419, + 559 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 547, + 419, + 559 + ], + "spans": [ + { + "bbox": [ + 306, + 547, + 419, + 559 + ], + "type": "text", + "content": "3.1. Problem Statement" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 565, + 545, + 678 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 565, + 545, + 678 + ], + "spans": [ + { + "bbox": [ + 304, + 565, + 545, + 678 + ], + "type": "text", + "content": "Suppose that there are " + }, + { + "bbox": [ + 304, + 565, + 545, + 678 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 304, + 565, + 545, + 678 + ], + "type": "text", + "content": " participants " + }, + { + "bbox": [ + 304, + 565, + 545, + 678 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_1, \\mathcal{P}_2, \\dots, \\mathcal{P}_K" + }, + { + "bbox": [ + 304, + 565, + 545, + 678 + ], + "type": "text", + "content": ". Each party " + }, + { + "bbox": [ + 304, + 565, + 545, + 678 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_k, k = 1, 2, \\dots, K" + }, + { + "bbox": [ + 304, + 565, + 545, + 678 + ], + "type": "text", + "content": " has its own local training dataset " + }, + { + "bbox": [ + 304, + 565, + 545, + 678 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_k = \\{\\pmb{x}_{j,k}, y_{j,k}\\}_{j=1}^{N_k}" + }, + { + "bbox": [ + 304, + 565, + 545, + 678 + ], + "type": "text", + "content": " of size " + }, + { + "bbox": [ + 304, + 565, + 545, + 678 + ], + "type": "inline_equation", + "content": "N_k" + }, + { + "bbox": [ + 304, + 565, + 545, + 678 + ], + "type": "text", + "content": " using which it learns its own local ML model " + }, + { + "bbox": [ + 304, + 565, + 545, + 678 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_{\\theta_k}" + }, + { + "bbox": [ + 304, + 565, + 545, + 678 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 565, + 545, + 678 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 304, + 565, + 545, + 678 + ], + "type": "text", + "content": " denotes the architecture and " + }, + { + "bbox": [ + 304, + 565, + 545, + 678 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 304, + 565, + 545, + 678 + ], + "type": "text", + "content": " represents the parameter set. We also assume that each party " + }, + { + "bbox": [ + 304, + 565, + 545, + 678 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_k" + }, + { + "bbox": [ + 304, + 565, + 545, + 678 + ], + "type": "text", + "content": " has its own validation set " + }, + { + "bbox": [ + 304, + 565, + 545, + 678 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathcal{D}}_k = \\{\\tilde{\\pmb{x}}_{j,k}, \\tilde{y}_{j,k}\\}_{j=1}^{\\tilde{N}_k}" + }, + { + "bbox": [ + 304, + 565, + 545, + 678 + ], + "type": "text", + "content": " of size " + }, + { + "bbox": [ + 304, + 565, + 545, + 678 + ], + "type": "inline_equation", + "content": "\\tilde{N}_k" + }, + { + "bbox": [ + 304, + 565, + 545, + 678 + ], + "type": "text", + "content": " and the validation accuracy of model " + }, + { + "bbox": [ + 304, + 565, + 545, + 678 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_{\\theta_k}" + }, + { + "bbox": [ + 304, + 565, + 545, + 678 + ], + "type": "text", + "content": " on " + }, + { + "bbox": [ + 304, + 565, + 545, + 678 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathcal{D}}_k" + }, + { + "bbox": [ + 304, + 565, + 545, + 678 + ], + "type": "text", + "content": " is " + }, + { + "bbox": [ + 304, + 565, + 545, + 678 + ], + "type": "inline_equation", + "content": "\\mathcal{A}_k" + }, + { + "bbox": [ + 304, + 565, + 545, + 678 + ], + "type": "text", + "content": ". The goal of each participant is to collaborate with other parties in order to improve" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 305, + 683, + 545, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 683, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 683, + 545, + 713 + ], + "type": "text", + "content": "1For simplicity of notation, we use the same symbol " + }, + { + "bbox": [ + 305, + 683, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 305, + 683, + 545, + 713 + ], + "type": "text", + "content": " to denote the model architecture of all the " + }, + { + "bbox": [ + 305, + 683, + 545, + 713 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 305, + 683, + 545, + 713 + ], + "type": "text", + "content": " participants. The proposed framework does not require all the parties to have the same model architecture." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8086" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 172 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 172 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 172 + ], + "type": "text", + "content": "the accuracy of its local model, i.e., obtain a better model " + }, + { + "bbox": [ + 46, + 72, + 289, + 172 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_{\\theta_k^*}" + }, + { + "bbox": [ + 46, + 72, + 289, + 172 + ], + "type": "text", + "content": " that has a higher validation accuracy " + }, + { + "bbox": [ + 46, + 72, + 289, + 172 + ], + "type": "inline_equation", + "content": "\\mathcal{A}_k^*" + }, + { + "bbox": [ + 46, + 72, + 289, + 172 + ], + "type": "text", + "content": " on " + }, + { + "bbox": [ + 46, + 72, + 289, + 172 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathcal{D}}_k" + }, + { + "bbox": [ + 46, + 72, + 289, + 172 + ], + "type": "text", + "content": " compared to " + }, + { + "bbox": [ + 46, + 72, + 289, + 172 + ], + "type": "inline_equation", + "content": "\\mathcal{A}_k" + }, + { + "bbox": [ + 46, + 72, + 289, + 172 + ], + "type": "text", + "content": ", without compromising on the confidentiality and privacy of its local data " + }, + { + "bbox": [ + 46, + 72, + 289, + 172 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_k" + }, + { + "bbox": [ + 46, + 72, + 289, + 172 + ], + "type": "text", + "content": ". In other words, no participant " + }, + { + "bbox": [ + 46, + 72, + 289, + 172 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_l, l \\neq k" + }, + { + "bbox": [ + 46, + 72, + 289, + 172 + ], + "type": "text", + "content": " can access " + }, + { + "bbox": [ + 46, + 72, + 289, + 172 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_k" + }, + { + "bbox": [ + 46, + 72, + 289, + 172 + ], + "type": "text", + "content": " in its plaintext form or learn any private information about " + }, + { + "bbox": [ + 46, + 72, + 289, + 172 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_k" + }, + { + "bbox": [ + 46, + 72, + 289, + 172 + ], + "type": "text", + "content": " such as the " + }, + { + "bbox": [ + 46, + 72, + 289, + 172 + ], + "type": "inline_equation", + "content": "x_{j,k}" + }, + { + "bbox": [ + 46, + 72, + 289, + 172 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 72, + 289, + 172 + ], + "type": "inline_equation", + "content": "y_{j,k}" + }, + { + "bbox": [ + 46, + 72, + 289, + 172 + ], + "type": "text", + "content": " values. The utility of the framework is measured as the average collaboration gain, i.e., " + }, + { + "bbox": [ + 46, + 72, + 289, + 172 + ], + "type": "inline_equation", + "content": "U = \\frac{1}{K} \\sum_{k=1}^{K} (\\mathcal{A}_k^* - \\mathcal{A}_k)" + }, + { + "bbox": [ + 46, + 72, + 289, + 172 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 175, + 272, + 189 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 175, + 272, + 189 + ], + "spans": [ + { + "bbox": [ + 47, + 175, + 272, + 189 + ], + "type": "text", + "content": "3.2. CaPriDe Learning Framework Description" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 193, + 290, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 193, + 290, + 361 + ], + "spans": [ + { + "bbox": [ + 46, + 193, + 290, + 361 + ], + "type": "text", + "content": "Let " + }, + { + "bbox": [ + 46, + 193, + 290, + 361 + ], + "type": "inline_equation", + "content": "\\mathcal{E}(\\pmb{X}_k) = \\{\\mathcal{E}(\\pmb{x}_{j,k})\\}_{j=1}^{N_k}" + }, + { + "bbox": [ + 46, + 193, + 290, + 361 + ], + "type": "text", + "content": " be the collection of encrypted input samples of participant " + }, + { + "bbox": [ + 46, + 193, + 290, + 361 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_k" + }, + { + "bbox": [ + 46, + 193, + 290, + 361 + ], + "type": "text", + "content": ". Here, the encryption is based on a FHE scheme with public key " + }, + { + "bbox": [ + 46, + 193, + 290, + 361 + ], + "type": "inline_equation", + "content": "\\kappa_k" + }, + { + "bbox": [ + 46, + 193, + 290, + 361 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 46, + 193, + 290, + 361 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_k" + }, + { + "bbox": [ + 46, + 193, + 290, + 361 + ], + "type": "text", + "content": ". In CaPriDe learning, each party initiates the collaboration by publishing " + }, + { + "bbox": [ + 46, + 193, + 290, + 361 + ], + "type": "inline_equation", + "content": "\\mathcal{E}(\\pmb{X}_k)" + }, + { + "bbox": [ + 46, + 193, + 290, + 361 + ], + "type": "text", + "content": " to the other participants. We now describe how the learning proceeds at participant " + }, + { + "bbox": [ + 46, + 193, + 290, + 361 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_k" + }, + { + "bbox": [ + 46, + 193, + 290, + 361 + ], + "type": "text", + "content": ". Let " + }, + { + "bbox": [ + 46, + 193, + 290, + 361 + ], + "type": "inline_equation", + "content": "\\pmb{z}_{j,(k,l)}" + }, + { + "bbox": [ + 46, + 193, + 290, + 361 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 193, + 290, + 361 + ], + "type": "inline_equation", + "content": "\\pmb{p}_{j,(k,l)}" + }, + { + "bbox": [ + 46, + 193, + 290, + 361 + ], + "type": "text", + "content": " denote the logits vector and probability distribution obtained when model " + }, + { + "bbox": [ + 46, + 193, + 290, + 361 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_{\\theta_l}" + }, + { + "bbox": [ + 46, + 193, + 290, + 361 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 46, + 193, + 290, + 361 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_l" + }, + { + "bbox": [ + 46, + 193, + 290, + 361 + ], + "type": "text", + "content": " is applied on the input sample " + }, + { + "bbox": [ + 46, + 193, + 290, + 361 + ], + "type": "inline_equation", + "content": "\\pmb{x}_{j,k}" + }, + { + "bbox": [ + 46, + 193, + 290, + 361 + ], + "type": "text", + "content": ", i.e., the " + }, + { + "bbox": [ + 46, + 193, + 290, + 361 + ], + "type": "inline_equation", + "content": "j^{th}" + }, + { + "bbox": [ + 46, + 193, + 290, + 361 + ], + "type": "text", + "content": " training sample of " + }, + { + "bbox": [ + 46, + 193, + 290, + 361 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_k" + }, + { + "bbox": [ + 46, + 193, + 290, + 361 + ], + "type": "text", + "content": ". Assume that we are at round " + }, + { + "bbox": [ + 46, + 193, + 290, + 361 + ], + "type": "inline_equation", + "content": "(t+1), 0 \\leq t < \\tau" + }, + { + "bbox": [ + 46, + 193, + 290, + 361 + ], + "type": "text", + "content": ", of the collaboration and the current model of " + }, + { + "bbox": [ + 46, + 193, + 290, + 361 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_k" + }, + { + "bbox": [ + 46, + 193, + 290, + 361 + ], + "type": "text", + "content": " is " + }, + { + "bbox": [ + 46, + 193, + 290, + 361 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_{\\theta_k^t}" + }, + { + "bbox": [ + 46, + 193, + 290, + 361 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 46, + 193, + 290, + 361 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_k" + }, + { + "bbox": [ + 46, + 193, + 290, + 361 + ], + "type": "text", + "content": " performs one forward pass on its own training set " + }, + { + "bbox": [ + 46, + 193, + 290, + 361 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_k" + }, + { + "bbox": [ + 46, + 193, + 290, + 361 + ], + "type": "text", + "content": " to obtain the predictions " + }, + { + "bbox": [ + 46, + 193, + 290, + 361 + ], + "type": "inline_equation", + "content": "\\{\\pmb{p}_{j,(k,k)}\\}_{j=1}^{N_k}" + }, + { + "bbox": [ + 46, + 193, + 290, + 361 + ], + "type": "text", + "content": ". The cross-entropy loss of party " + }, + { + "bbox": [ + 46, + 193, + 290, + 361 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_k" + }, + { + "bbox": [ + 46, + 193, + 290, + 361 + ], + "type": "text", + "content": " can be obtained as:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 94, + 367, + 287, + 400 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 94, + 367, + 287, + 400 + ], + "spans": [ + { + "bbox": [ + 94, + 367, + 287, + 400 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {C E _ {k}} = \\sum_ {j = 1} ^ {N _ {k}} L _ {C E} \\left(\\boldsymbol {p} _ {j, (k, k)}, y _ {j, k}\\right). \\tag {3}", + "image_path": "b606e230fe6089c215d6e25c92a5bd562157b027263abf67928f1380bd349892.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 405, + 287, + 501 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 405, + 287, + 501 + ], + "spans": [ + { + "bbox": [ + 46, + 405, + 287, + 501 + ], + "type": "text", + "content": "To enable knowledge transfer between participants, we make use of the knowledge distillation (KD) approach [13], where a student model learns to mimic the predictions of a teacher model. In CaPriDe learning, there is no designated teacher model, and mutual KD between multiple peer models is used for knowledge transfer. Let " + }, + { + "bbox": [ + 46, + 405, + 287, + 501 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{DL_{(k,l)}}" + }, + { + "bbox": [ + 46, + 405, + 287, + 501 + ], + "type": "text", + "content": " denote the pairwise distillation loss between predictions of " + }, + { + "bbox": [ + 46, + 405, + 287, + 501 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_k" + }, + { + "bbox": [ + 46, + 405, + 287, + 501 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 405, + 287, + 501 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_l" + }, + { + "bbox": [ + 46, + 405, + 287, + 501 + ], + "type": "text", + "content": ", which is defined as:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 83, + 504, + 287, + 538 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 83, + 504, + 287, + 538 + ], + "spans": [ + { + "bbox": [ + 83, + 504, + 287, + 538 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {D L _ {(k, l)}} = \\sum_ {j = 1} ^ {N _ {k}} D _ {D L} \\left(\\boldsymbol {p} _ {j, (k, k)}, \\boldsymbol {p} _ {j, (k, l)}\\right), \\tag {4}", + "image_path": "32c5281178734afb1bac99486778228314bff952f731b0337077e77bb5f0045d.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 544, + 287, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 544, + 287, + 579 + ], + "spans": [ + { + "bbox": [ + 46, + 544, + 287, + 579 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 544, + 287, + 579 + ], + "type": "inline_equation", + "content": "D_{DL}(\\pmb{p}_1, \\pmb{p}_2)" + }, + { + "bbox": [ + 46, + 544, + 287, + 579 + ], + "type": "text", + "content": " denotes the mimic distance between two predictions " + }, + { + "bbox": [ + 46, + 544, + 287, + 579 + ], + "type": "inline_equation", + "content": "\\pmb{p}_1" + }, + { + "bbox": [ + 46, + 544, + 287, + 579 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 544, + 287, + 579 + ], + "type": "inline_equation", + "content": "\\pmb{p}_2" + }, + { + "bbox": [ + 46, + 544, + 287, + 579 + ], + "type": "text", + "content": ". The total distillation loss for " + }, + { + "bbox": [ + 46, + 544, + 287, + 579 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_k" + }, + { + "bbox": [ + 46, + 544, + 287, + 579 + ], + "type": "text", + "content": " can be computed as:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 98, + 584, + 287, + 618 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 98, + 584, + 287, + 618 + ], + "spans": [ + { + "bbox": [ + 98, + 584, + 287, + 618 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {D L _ {k}} = \\frac {1}{K - 1} \\sum_ {l = 1, l \\neq k} ^ {K} \\mathcal {L} _ {D L _ {(k, l)}}. \\tag {5}", + "image_path": "f477fe191cb42dd92609926c26c41459c4d834f0f68bedacb2c628bf6b088d7a.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 623, + 228, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 623, + 228, + 635 + ], + "spans": [ + { + "bbox": [ + 47, + 623, + 228, + 635 + ], + "type": "text", + "content": "Therefore, the overall loss of " + }, + { + "bbox": [ + 47, + 623, + 228, + 635 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_k" + }, + { + "bbox": [ + 47, + 623, + 228, + 635 + ], + "type": "text", + "content": " is given by:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 116, + 641, + 287, + 654 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 116, + 641, + 287, + 654 + ], + "spans": [ + { + "bbox": [ + 116, + 641, + 287, + 654 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {k} = \\mathcal {L} _ {C E _ {k}} + \\lambda_ {k} \\mathcal {L} _ {D L _ {k}}, \\tag {6}", + "image_path": "4352b687118e58baa4fbd5f89253ae5fbe8c3e85910a6bc435283dcb7c8a10d9.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 659, + 287, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 659, + 287, + 696 + ], + "spans": [ + { + "bbox": [ + 46, + 659, + 287, + 696 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 659, + 287, + 696 + ], + "type": "inline_equation", + "content": "\\lambda_{k}" + }, + { + "bbox": [ + 46, + 659, + 287, + 696 + ], + "type": "text", + "content": " is the weighting factor between the cross-entropy and distillation losses for " + }, + { + "bbox": [ + 46, + 659, + 287, + 696 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_k" + }, + { + "bbox": [ + 46, + 659, + 287, + 696 + ], + "type": "text", + "content": ". The model of participant " + }, + { + "bbox": [ + 46, + 659, + 287, + 696 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_k" + }, + { + "bbox": [ + 46, + 659, + 287, + 696 + ], + "type": "text", + "content": " is then updated through a backward pass as follows:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 116, + 700, + 287, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 116, + 700, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 116, + 700, + 287, + 715 + ], + "type": "interline_equation", + "content": "\\theta_ {k} ^ {t + 1} = \\theta_ {k} ^ {t} - \\alpha_ {k} \\nabla_ {\\theta_ {k} ^ {t}} \\mathcal {L} _ {k}, \\tag {7}", + "image_path": "4672e135727e3252487ac3e88e24569ab55b9f3d3e226e6ee189bee52f6cc17c.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 72, + 451, + 84 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 451, + 84 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 451, + 84 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 306, + 72, + 451, + 84 + ], + "type": "inline_equation", + "content": "\\alpha_{k}" + }, + { + "bbox": [ + 306, + 72, + 451, + 84 + ], + "type": "text", + "content": " is the learning rate of " + }, + { + "bbox": [ + 306, + 72, + 451, + 84 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_k" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 91, + 501, + 105 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 91, + 501, + 105 + ], + "spans": [ + { + "bbox": [ + 306, + 91, + 501, + 105 + ], + "type": "text", + "content": "3.3. Encryption-friendly Distillation Loss" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 110, + 547, + 348 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 110, + 547, + 348 + ], + "spans": [ + { + "bbox": [ + 304, + 110, + 547, + 348 + ], + "type": "text", + "content": "The key missing component in the above CaPriDe learning framework is the computation " + }, + { + "bbox": [ + 304, + 110, + 547, + 348 + ], + "type": "inline_equation", + "content": "D_{DL}" + }, + { + "bbox": [ + 304, + 110, + 547, + 348 + ], + "type": "text", + "content": " in Eq. 4. For participant " + }, + { + "bbox": [ + 304, + 110, + 547, + 348 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_k" + }, + { + "bbox": [ + 304, + 110, + 547, + 348 + ], + "type": "text", + "content": ", its logit vectors " + }, + { + "bbox": [ + 304, + 110, + 547, + 348 + ], + "type": "inline_equation", + "content": "\\pmb{z}_{.,(k,k)}" + }, + { + "bbox": [ + 304, + 110, + 547, + 348 + ], + "type": "text", + "content": " and predictions " + }, + { + "bbox": [ + 304, + 110, + 547, + 348 + ], + "type": "inline_equation", + "content": "\\pmb{p}_{.,(k,k)}" + }, + { + "bbox": [ + 304, + 110, + 547, + 348 + ], + "type": "text", + "content": " are available in the plaintext domain. However, since participant " + }, + { + "bbox": [ + 304, + 110, + 547, + 348 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_l" + }, + { + "bbox": [ + 304, + 110, + 547, + 348 + ], + "type": "text", + "content": " receives the input samples of " + }, + { + "bbox": [ + 304, + 110, + 547, + 348 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_k" + }, + { + "bbox": [ + 304, + 110, + 547, + 348 + ], + "type": "text", + "content": " in encrypted form, it can perform inference in the encrypted domain and obtain only " + }, + { + "bbox": [ + 304, + 110, + 547, + 348 + ], + "type": "inline_equation", + "content": "\\mathcal{E}\\left(\\pmb{z}_{.,(k,l)}\\right)" + }, + { + "bbox": [ + 304, + 110, + 547, + 348 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 110, + 547, + 348 + ], + "type": "inline_equation", + "content": "\\mathcal{E}\\left(\\pmb{z}_{.,(k,l)}\\right) = \\mathcal{M}_{\\theta_l}\\left(\\mathcal{E}(\\pmb{x}_{.,k})\\right)" + }, + { + "bbox": [ + 304, + 110, + 547, + 348 + ], + "type": "text", + "content": ". It is not straightforward to obtain " + }, + { + "bbox": [ + 304, + 110, + 547, + 348 + ], + "type": "inline_equation", + "content": "\\mathcal{E}\\left(\\pmb{p}_{.,(k,l)}\\right)" + }, + { + "bbox": [ + 304, + 110, + 547, + 348 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 304, + 110, + 547, + 348 + ], + "type": "inline_equation", + "content": "\\mathcal{E}\\left(\\pmb{z}_{.,(k,l)}\\right)" + }, + { + "bbox": [ + 304, + 110, + 547, + 348 + ], + "type": "text", + "content": " because it involves computing a non-linear softmax function in the encrypted domain, which is a challenging task in the FHE domain and consumes a large multiplicative depth. Even if this softmax operation is somehow implemented, it is not possible to send " + }, + { + "bbox": [ + 304, + 110, + 547, + 348 + ], + "type": "inline_equation", + "content": "\\mathcal{E}\\left(\\pmb{p}_{.,(k,l)}\\right)" + }, + { + "bbox": [ + 304, + 110, + 547, + 348 + ], + "type": "text", + "content": " back to " + }, + { + "bbox": [ + 304, + 110, + 547, + 348 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_k" + }, + { + "bbox": [ + 304, + 110, + 547, + 348 + ], + "type": "text", + "content": ", who can decrypt it and obtain " + }, + { + "bbox": [ + 304, + 110, + 547, + 348 + ], + "type": "inline_equation", + "content": "\\pmb{p}_{.,(k,l)}" + }, + { + "bbox": [ + 304, + 110, + 547, + 348 + ], + "type": "text", + "content": ". Since " + }, + { + "bbox": [ + 304, + 110, + 547, + 348 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_k" + }, + { + "bbox": [ + 304, + 110, + 547, + 348 + ], + "type": "text", + "content": " also has knowledge of " + }, + { + "bbox": [ + 304, + 110, + 547, + 348 + ], + "type": "inline_equation", + "content": "\\pmb{x}_{.,k}" + }, + { + "bbox": [ + 304, + 110, + 547, + 348 + ], + "type": "text", + "content": ", it becomes feasible for " + }, + { + "bbox": [ + 304, + 110, + 547, + 348 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_k" + }, + { + "bbox": [ + 304, + 110, + 547, + 348 + ], + "type": "text", + "content": " to carry out a model extraction attack based on " + }, + { + "bbox": [ + 304, + 110, + 547, + 348 + ], + "type": "inline_equation", + "content": "(\\pmb{x}_{.,k},\\pmb{p}_{.,(k,l)})" + }, + { + "bbox": [ + 304, + 110, + 547, + 348 + ], + "type": "text", + "content": " pairs to glean information about model " + }, + { + "bbox": [ + 304, + 110, + 547, + 348 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_{\\theta_l}" + }, + { + "bbox": [ + 304, + 110, + 547, + 348 + ], + "type": "text", + "content": ". This violates the privacy requirements of participant " + }, + { + "bbox": [ + 304, + 110, + 547, + 348 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_l" + }, + { + "bbox": [ + 304, + 110, + 547, + 348 + ], + "type": "text", + "content": ". Hence, there is a need for a loss function that can be computed in the encrypted domain." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 349, + 546, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 349, + 546, + 399 + ], + "spans": [ + { + "bbox": [ + 304, + 349, + 546, + 399 + ], + "type": "text", + "content": "A straightforward solution is to compute the L2 distance between logit vectors " + }, + { + "bbox": [ + 304, + 349, + 546, + 399 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_{.,(k,k)}" + }, + { + "bbox": [ + 304, + 349, + 546, + 399 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 349, + 546, + 399 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_{.,(k,l)}" + }, + { + "bbox": [ + 304, + 349, + 546, + 399 + ], + "type": "text", + "content": ". In this case, participant " + }, + { + "bbox": [ + 304, + 349, + 546, + 399 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_k" + }, + { + "bbox": [ + 304, + 349, + 546, + 399 + ], + "type": "text", + "content": " can publish " + }, + { + "bbox": [ + 304, + 349, + 546, + 399 + ], + "type": "inline_equation", + "content": "\\mathcal{E}\\left(\\mathbf{z}_{.,(k,k)}\\right)" + }, + { + "bbox": [ + 304, + 349, + 546, + 399 + ], + "type": "text", + "content": " and the pairwise distillation loss " + }, + { + "bbox": [ + 304, + 349, + 546, + 399 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{DL_{(k,l)}}" + }, + { + "bbox": [ + 304, + 349, + 546, + 399 + ], + "type": "text", + "content": " can be computed by participant " + }, + { + "bbox": [ + 304, + 349, + 546, + 399 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_l" + }, + { + "bbox": [ + 304, + 349, + 546, + 399 + ], + "type": "text", + "content": " as:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 407, + 551, + 445 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 407, + 551, + 445 + ], + "spans": [ + { + "bbox": [ + 307, + 407, + 551, + 445 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\left. \\mathcal {E} \\left(D _ {D L} \\left(\\boldsymbol {p} _ {j, (k, k)}, \\boldsymbol {p} _ {j, (k, l)}\\right)\\right) \\equiv \\mathcal {E} \\left(D _ {L 2} \\left(\\boldsymbol {z} _ {j, (k, k)}, \\boldsymbol {z} _ {j, (k, l)}\\right)\\right) \\right. \\\\ = \\left(\\mathcal {E} \\left(\\boldsymbol {z} _ {j, (k, k)}\\right) - \\mathcal {E} \\left(\\boldsymbol {z} _ {j, (k, l)}\\right)\\right) \\cdot \\left(\\mathcal {E} \\left(\\boldsymbol {z} _ {j, (k, k)}\\right) - \\mathcal {E} \\left(\\boldsymbol {z} _ {j, (k, l)}\\right)\\right) \\\\ \\end{array}", + "image_path": "d1fb084f797d25238b85ae692f52f06df370c81e75602be3b6d5bbd4fed7a1d3.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 453, + 545, + 543 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 453, + 545, + 543 + ], + "spans": [ + { + "bbox": [ + 304, + 453, + 545, + 543 + ], + "type": "text", + "content": "However, our experiments indicate the above simple solution based on L2 distance does not lead to high utility " + }, + { + "bbox": [ + 304, + 453, + 545, + 543 + ], + "type": "inline_equation", + "content": "U" + }, + { + "bbox": [ + 304, + 453, + 545, + 543 + ], + "type": "text", + "content": ". On the other hand, we observed that using the KL divergence between predictions of " + }, + { + "bbox": [ + 304, + 453, + 545, + 543 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_k" + }, + { + "bbox": [ + 304, + 453, + 545, + 543 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 453, + 545, + 543 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_l" + }, + { + "bbox": [ + 304, + 453, + 545, + 543 + ], + "type": "text", + "content": " almost always leads to better utility. But, " + }, + { + "bbox": [ + 304, + 453, + 545, + 543 + ], + "type": "inline_equation", + "content": "D_{KL}\\left(\\pmb{p}_{.,(k,k)}\\| \\pmb{p}_{.,(k,l)}\\right)" + }, + { + "bbox": [ + 304, + 453, + 545, + 543 + ], + "type": "text", + "content": " cannot be directly computed in the encrypted domain because participant " + }, + { + "bbox": [ + 304, + 453, + 545, + 543 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_l" + }, + { + "bbox": [ + 304, + 453, + 545, + 543 + ], + "type": "text", + "content": " only has access to the encrypted logits " + }, + { + "bbox": [ + 304, + 453, + 545, + 543 + ], + "type": "inline_equation", + "content": "\\mathcal{E}\\left(z_{.,(k,l)}\\right)" + }, + { + "bbox": [ + 304, + 453, + 545, + 543 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 542, + 545, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 542, + 545, + 567 + ], + "spans": [ + { + "bbox": [ + 304, + 542, + 545, + 567 + ], + "type": "text", + "content": "Fortunately, this problem can be solved by approximating the KL divergence. Combining Eqs. 1 and 2, we get:" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 306, + 574, + 555, + 617 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 574, + 555, + 617 + ], + "spans": [ + { + "bbox": [ + 306, + 574, + 555, + 617 + ], + "type": "interline_equation", + "content": "D _ {K L} \\left(\\boldsymbol {p} _ {2} \\right\\| \\boldsymbol {p} _ {1}) = \\sum_ {m = 1} ^ {M} p _ {2} ^ {m} \\log p _ {2} ^ {m} - \\sum_ {m = 1} ^ {M} p _ {2} ^ {m} \\left(z _ {1} ^ {m} / T\\right) + M \\log C. \\tag {8}", + "image_path": "7db40b65cf0fd88cdbd1b26d93330b713de06659a3950a043f3f2bf3d6c33e58.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 618, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 618, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 618, + 545, + 713 + ], + "type": "text", + "content": "Here, the first term can be denoted as " + }, + { + "bbox": [ + 304, + 618, + 545, + 713 + ], + "type": "inline_equation", + "content": "(\\pmb{p}_2 \\cdot \\log \\pmb{p}_2)" + }, + { + "bbox": [ + 304, + 618, + 545, + 713 + ], + "type": "text", + "content": " and measures the entropy (uncertainty) in " + }, + { + "bbox": [ + 304, + 618, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\pmb{p}_2" + }, + { + "bbox": [ + 304, + 618, + 545, + 713 + ], + "type": "text", + "content": ". The second term measures the divergence (unnormized cross-entropy) between " + }, + { + "bbox": [ + 304, + 618, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\pmb{p}_2" + }, + { + "bbox": [ + 304, + 618, + 545, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 618, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\pmb{p}_1" + }, + { + "bbox": [ + 304, + 618, + 545, + 713 + ], + "type": "text", + "content": " and can be expressed as " + }, + { + "bbox": [ + 304, + 618, + 545, + 713 + ], + "type": "inline_equation", + "content": "(\\pmb{p}_2 \\cdot \\pmb{z}_1 / T)" + }, + { + "bbox": [ + 304, + 618, + 545, + 713 + ], + "type": "text", + "content": ". The final term " + }, + { + "bbox": [ + 304, + 618, + 545, + 713 + ], + "type": "inline_equation", + "content": "M \\log C" + }, + { + "bbox": [ + 304, + 618, + 545, + 713 + ], + "type": "text", + "content": " is just a normalization constant that does not affect the relative ordering of the class predictions. We observed that this normalization term does not appear to play a major role in model training. Hence, we ignore this value" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8087" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "code", + "bbox": [ + 48, + 86, + 545, + 383 + ], + "blocks": [ + { + "bbox": [ + 49, + 72, + 218, + 84 + ], + "lines": [ + { + "bbox": [ + 49, + 72, + 218, + 84 + ], + "spans": [ + { + "bbox": [ + 49, + 72, + 218, + 84 + ], + "type": "text", + "content": "Algorithm 1 CaPRiDe learning algorithm" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "code_caption" + }, + { + "bbox": [ + 48, + 86, + 545, + 383 + ], + "lines": [ + { + "bbox": [ + 48, + 86, + 545, + 383 + ], + "spans": [ + { + "bbox": [ + 48, + 86, + 545, + 383 + ], + "type": "text", + "content": "Input: Number of participants " + }, + { + "bbox": [ + 48, + 86, + 545, + 383 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 48, + 86, + 545, + 383 + ], + "type": "text", + "content": " , weight parameter " + }, + { + "bbox": [ + 48, + 86, + 545, + 383 + ], + "type": "inline_equation", + "content": "\\lambda_{k}" + }, + { + "bbox": [ + 48, + 86, + 545, + 383 + ], + "type": "text", + "content": " , learning rate " + }, + { + "bbox": [ + 48, + 86, + 545, + 383 + ], + "type": "inline_equation", + "content": "\\alpha_{k}" + }, + { + "bbox": [ + 48, + 86, + 545, + 383 + ], + "type": "text", + "content": " for each participant \nAssumption: Each participant has local training set " + }, + { + "bbox": [ + 48, + 86, + 545, + 383 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_k = \\{\\pmb {x}_{j,k},y_{j,k}\\}_{j = [1,N_k]}" + }, + { + "bbox": [ + 48, + 86, + 545, + 383 + ], + "type": "text", + "content": " of size " + }, + { + "bbox": [ + 48, + 86, + 545, + 383 + ], + "type": "inline_equation", + "content": "N_{k}" + }, + { + "bbox": [ + 48, + 86, + 545, + 383 + ], + "type": "text", + "content": " , a fully homomorphic encryption function " + }, + { + "bbox": [ + 48, + 86, + 545, + 383 + ], + "type": "inline_equation", + "content": "\\mathcal{E}" + }, + { + "bbox": [ + 48, + 86, + 545, + 383 + ], + "type": "text", + "content": " with public key " + }, + { + "bbox": [ + 48, + 86, + 545, + 383 + ], + "type": "inline_equation", + "content": "\\kappa_{k}" + }, + { + "bbox": [ + 48, + 86, + 545, + 383 + ], + "type": "text", + "content": " , a decryption function " + }, + { + "bbox": [ + 48, + 86, + 545, + 383 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 48, + 86, + 545, + 383 + ], + "type": "text", + "content": " with secret key " + }, + { + "bbox": [ + 48, + 86, + 545, + 383 + ], + "type": "inline_equation", + "content": "\\tilde{\\kappa}_k" + }, + { + "bbox": [ + 48, + 86, + 545, + 383 + ], + "type": "text", + "content": " \nfor each participant " + }, + { + "bbox": [ + 48, + 86, + 545, + 383 + ], + "type": "inline_equation", + "content": "k = 1,2,\\dots ,K" + }, + { + "bbox": [ + 48, + 86, + 545, + 383 + ], + "type": "text", + "content": " do Train initial model " + }, + { + "bbox": [ + 48, + 86, + 545, + 383 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_{\\theta_k^0}" + }, + { + "bbox": [ + 48, + 86, + 545, + 383 + ], + "type": "text", + "content": " based on local data " + }, + { + "bbox": [ + 48, + 86, + 545, + 383 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_k" + }, + { + "bbox": [ + 48, + 86, + 545, + 383 + ], + "type": "text", + "content": " Publish encrypted data " + }, + { + "bbox": [ + 48, + 86, + 545, + 383 + ], + "type": "inline_equation", + "content": "\\{\\mathcal{E}(\\pmb{x}_{j,k})\\}_{j = [1,N_k]}" + }, + { + "bbox": [ + 48, + 86, + 545, + 383 + ], + "type": "text", + "content": " \nend for \nfor each round " + }, + { + "bbox": [ + 48, + 86, + 545, + 383 + ], + "type": "inline_equation", + "content": "t = 1,2,\\dots ,\\tau" + }, + { + "bbox": [ + 48, + 86, + 545, + 383 + ], + "type": "text", + "content": " do \nfor each participant " + }, + { + "bbox": [ + 48, + 86, + 545, + 383 + ], + "type": "inline_equation", + "content": "k = 1,2,\\dots ,K" + }, + { + "bbox": [ + 48, + 86, + 545, + 383 + ], + "type": "text", + "content": " do " + }, + { + "bbox": [ + 48, + 86, + 545, + 383 + ], + "type": "inline_equation", + "content": "\\{\\pmb {p}_{j,(k,k)}\\}_{j = [1,N_k]}\\gets \\{\\sigma (\\mathcal{M}_{\\theta_k^{t - 1}}(\\pmb {x}_{j,k}))\\}_{j = [1,N_k]}" + }, + { + "bbox": [ + 48, + 86, + 545, + 383 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{CE_k}\\gets" + }, + { + "bbox": [ + 48, + 86, + 545, + 383 + ], + "type": "text", + "content": " Compute cross-entropy loss of participant " + }, + { + "bbox": [ + 48, + 86, + 545, + 383 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 48, + 86, + 545, + 383 + ], + "type": "text", + "content": " using Equation 3 \n" + }, + { + "bbox": [ + 48, + 86, + 545, + 383 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{L}}_{KL_k}\\gets" + }, + { + "bbox": [ + 48, + 86, + 545, + 383 + ], + "type": "text", + "content": " Compute local uncertainty of participant " + }, + { + "bbox": [ + 48, + 86, + 545, + 383 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 48, + 86, + 545, + 383 + ], + "type": "text", + "content": " using Equation 11 \nPublish encrypted inferences " + }, + { + "bbox": [ + 48, + 86, + 545, + 383 + ], + "type": "inline_equation", + "content": "\\{\\mathcal{E}(\\pmb {p}_{j,(k,k)})\\}_{j = [1,N_k]}" + }, + { + "bbox": [ + 48, + 86, + 545, + 383 + ], + "type": "text", + "content": " \nfor each participant " + }, + { + "bbox": [ + 48, + 86, + 545, + 383 + ], + "type": "inline_equation", + "content": "l = 1,2,\\dots ,K,l\\neq k" + }, + { + "bbox": [ + 48, + 86, + 545, + 383 + ], + "type": "text", + "content": " do " + }, + { + "bbox": [ + 48, + 86, + 545, + 383 + ], + "type": "inline_equation", + "content": "\\{\\mathcal{E}(\\pmb {z}_{j,(k,l)})\\}_{j = [1,N_k]}\\gets \\{\\mathcal{M}_{\\theta_l^{t - 1}}(\\mathcal{E}(\\pmb {x}_{j,k}))\\}_{j = [1,N_k]}" + }, + { + "bbox": [ + 48, + 86, + 545, + 383 + ], + "type": "text", + "content": " (Perform inference on encrypted data) \n" + }, + { + "bbox": [ + 48, + 86, + 545, + 383 + ], + "type": "inline_equation", + "content": "\\mathcal{E}(\\tilde{\\mathcal{L}}_{KL_{(k,l)}})\\gets" + }, + { + "bbox": [ + 48, + 86, + 545, + 383 + ], + "type": "text", + "content": " Compute pairwise divergence in encrypted domain using Equation 12 \nSend " + }, + { + "bbox": [ + 48, + 86, + 545, + 383 + ], + "type": "inline_equation", + "content": "\\mathcal{E}(\\tilde{\\mathcal{L}}_{KL_{(k,l)}})" + }, + { + "bbox": [ + 48, + 86, + 545, + 383 + ], + "type": "text", + "content": " to participant " + }, + { + "bbox": [ + 48, + 86, + 545, + 383 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 48, + 86, + 545, + 383 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{DL_{(k,l)}}\\gets" + }, + { + "bbox": [ + 48, + 86, + 545, + 383 + ], + "type": "text", + "content": " Compute pairwise distillation loss between participants " + }, + { + "bbox": [ + 48, + 86, + 545, + 383 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 48, + 86, + 545, + 383 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 48, + 86, + 545, + 383 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 48, + 86, + 545, + 383 + ], + "type": "text", + "content": " using Equation 10 \nend for \n" + }, + { + "bbox": [ + 48, + 86, + 545, + 383 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{DL_k}\\gets" + }, + { + "bbox": [ + 48, + 86, + 545, + 383 + ], + "type": "text", + "content": " Compute distillation loss of participant " + }, + { + "bbox": [ + 48, + 86, + 545, + 383 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 48, + 86, + 545, + 383 + ], + "type": "text", + "content": " using Equation 5 \n" + }, + { + "bbox": [ + 48, + 86, + 545, + 383 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_k\\gets" + }, + { + "bbox": [ + 48, + 86, + 545, + 383 + ], + "type": "text", + "content": " Compute total loss of participant " + }, + { + "bbox": [ + 48, + 86, + 545, + 383 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 48, + 86, + 545, + 383 + ], + "type": "text", + "content": " using Equation 6 \n" + }, + { + "bbox": [ + 48, + 86, + 545, + 383 + ], + "type": "inline_equation", + "content": "\\theta_k^t\\gets" + }, + { + "bbox": [ + 48, + 86, + 545, + 383 + ], + "type": "text", + "content": " Update local model of participant " + }, + { + "bbox": [ + 48, + 86, + 545, + 383 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 48, + 86, + 545, + 383 + ], + "type": "text", + "content": " using Equation 7 \nend for \nend for" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "code_body" + } + ], + "index": 1, + "sub_type": "algorithm" + }, + { + "bbox": [ + 47, + 407, + 276, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 407, + 276, + 418 + ], + "spans": [ + { + "bbox": [ + 47, + 407, + 276, + 418 + ], + "type": "text", + "content": "and define the approximate KL divergence is defined as:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 74, + 426, + 287, + 440 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 74, + 426, + 287, + 440 + ], + "spans": [ + { + "bbox": [ + 74, + 426, + 287, + 440 + ], + "type": "interline_equation", + "content": "\\hat {D} _ {K L} \\left(\\boldsymbol {p} _ {2} \\| \\boldsymbol {p} _ {1}\\right) = \\left(\\boldsymbol {p} _ {2} \\cdot \\log \\boldsymbol {p} _ {2}\\right) - \\left(\\boldsymbol {p} _ {2} \\cdot \\boldsymbol {z} _ {1} / T\\right). \\tag {9}", + "image_path": "a5fe1ab18004f79d3183e53686f492e55cb15fe197d3ff5bd8982098a76a2a5f.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 448, + 287, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 448, + 287, + 483 + ], + "spans": [ + { + "bbox": [ + 46, + 448, + 287, + 483 + ], + "type": "text", + "content": "Going back to Eq. 4, the pairwise distillation loss based on the above approximate KL divergence can be easily computed in the encrypted domain as:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 490, + 296, + 533 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 490, + 296, + 533 + ], + "spans": [ + { + "bbox": [ + 47, + 490, + 296, + 533 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\left. \\mathcal {E} \\left(D _ {D L} \\left(\\boldsymbol {p} _ {j, (k, k)}, \\boldsymbol {p} _ {j, (k, l)}\\right)\\right) \\equiv \\mathcal {E} \\left(\\hat {D} _ {K L} \\left(\\boldsymbol {p} _ {j, (k, k)} | | \\boldsymbol {p} _ {j, (k, l)}\\right)\\right) \\right. \\\\ = \\mathcal {E} \\left(\\boldsymbol {p} _ {j, (k, k)} \\cdot \\log \\boldsymbol {p} _ {j, (k, k)}\\right) - \\mathcal {E} \\left(\\boldsymbol {p} _ {j, (k, k)}\\right) \\cdot \\mathcal {E} \\left(\\boldsymbol {z} _ {j, (k, l)}\\right) / T \\\\ \\end{array}", + "image_path": "ecfb6c84273afb26cac0cbad23ccdd62458393a29d3df50f92ff2d24607e881e.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 540, + 287, + 599 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 540, + 287, + 599 + ], + "spans": [ + { + "bbox": [ + 46, + 540, + 287, + 599 + ], + "type": "text", + "content": "The above distillation loss based on the approximate KL divergence is encryption-friendly because it involves only linear operations and scaling by a constant plaintext value " + }, + { + "bbox": [ + 46, + 540, + 287, + 599 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 46, + 540, + 287, + 599 + ], + "type": "text", + "content": ". In practice, our goal is to compute " + }, + { + "bbox": [ + 46, + 540, + 287, + 599 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{DL_{(k,l)}}" + }, + { + "bbox": [ + 46, + 540, + 287, + 599 + ], + "type": "text", + "content": " at participant " + }, + { + "bbox": [ + 46, + 540, + 287, + 599 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_k" + }, + { + "bbox": [ + 46, + 540, + 287, + 599 + ], + "type": "text", + "content": ". Therefore, Eq. 4 can be further reformulated as:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 85, + 606, + 287, + 628 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 85, + 606, + 287, + 628 + ], + "spans": [ + { + "bbox": [ + 85, + 606, + 287, + 628 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {D L _ {(k, l)}} = \\hat {\\mathcal {L}} _ {K L _ {k}} - d \\left(\\mathcal {E} \\left(\\tilde {\\mathcal {L}} _ {K L _ {(k, l)}}\\right)\\right), \\tag {10}", + "image_path": "d8106edfc4b4de2893ebc7adf9e22fd3b6d6b65a3ae2df2dda56f6a7d2c19920.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 635, + 74, + 644 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 635, + 74, + 644 + ], + "spans": [ + { + "bbox": [ + 47, + 635, + 74, + 644 + ], + "type": "text", + "content": "where" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 94, + 643, + 287, + 677 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 94, + 643, + 287, + 677 + ], + "spans": [ + { + "bbox": [ + 94, + 643, + 287, + 677 + ], + "type": "interline_equation", + "content": "\\hat {\\mathcal {L}} _ {K L _ {k}} = \\sum_ {j = 1} ^ {N _ {k}} \\left(\\boldsymbol {p} _ {j, (k, k)} \\cdot \\log \\boldsymbol {p} _ {j, (k, k)}\\right), \\tag {11}", + "image_path": "34a71f703c1f4b565a05e358910a52a651a8aac9774db9f9d8610709edc95e38.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 682, + 287, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 682, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 56, + 682, + 287, + 715 + ], + "type": "interline_equation", + "content": "\\mathcal {E} \\left(\\tilde {\\mathcal {L}} _ {K L _ {(k, l)}}\\right) = \\sum_ {j = 1} ^ {N _ {k}} \\mathcal {E} \\left(\\boldsymbol {p} _ {j, (k, k)}\\right) \\cdot \\mathcal {E} \\left(\\boldsymbol {z} _ {j, (k, l)}\\right) / T. \\tag {12}", + "image_path": "6fa8649fd616e5dc643e797e8ee550a2bbaceb630c1d7dc1292df736c7713ad1.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 406, + 545, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 406, + 545, + 517 + ], + "spans": [ + { + "bbox": [ + 304, + 406, + 545, + 517 + ], + "type": "text", + "content": "Since " + }, + { + "bbox": [ + 304, + 406, + 545, + 517 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_k" + }, + { + "bbox": [ + 304, + 406, + 545, + 517 + ], + "type": "text", + "content": " has access to " + }, + { + "bbox": [ + 304, + 406, + 545, + 517 + ], + "type": "inline_equation", + "content": "\\pmb{p}_{j,(k,k)}" + }, + { + "bbox": [ + 304, + 406, + 545, + 517 + ], + "type": "text", + "content": ", it is straightforward to compute " + }, + { + "bbox": [ + 304, + 406, + 545, + 517 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{L}}_{KL_k}" + }, + { + "bbox": [ + 304, + 406, + 545, + 517 + ], + "type": "text", + "content": " in plaintext space. If participant " + }, + { + "bbox": [ + 304, + 406, + 545, + 517 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_k" + }, + { + "bbox": [ + 304, + 406, + 545, + 517 + ], + "type": "text", + "content": " can publish " + }, + { + "bbox": [ + 304, + 406, + 545, + 517 + ], + "type": "inline_equation", + "content": "\\left\\{\\mathcal{E}(\\pmb{p}_{j,(k,k)})\\right\\}_{j=1}^{N_k}" + }, + { + "bbox": [ + 304, + 406, + 545, + 517 + ], + "type": "text", + "content": " in each round, " + }, + { + "bbox": [ + 304, + 406, + 545, + 517 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_l" + }, + { + "bbox": [ + 304, + 406, + 545, + 517 + ], + "type": "text", + "content": " can compute " + }, + { + "bbox": [ + 304, + 406, + 545, + 517 + ], + "type": "inline_equation", + "content": "\\mathcal{E}\\left(\\tilde{\\mathcal{L}}_{KL(k,l)}\\right)" + }, + { + "bbox": [ + 304, + 406, + 545, + 517 + ], + "type": "text", + "content": " in the encrypted domain because it already has access to " + }, + { + "bbox": [ + 304, + 406, + 545, + 517 + ], + "type": "inline_equation", + "content": "\\left\\{\\mathcal{E}\\left(z_{j,(k,l)}\\right)\\right\\}_{j=1}^{N_k}" + }, + { + "bbox": [ + 304, + 406, + 545, + 517 + ], + "type": "text", + "content": ". The encrypted value of " + }, + { + "bbox": [ + 304, + 406, + 545, + 517 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathcal{L}}_{KL(k,l)}" + }, + { + "bbox": [ + 304, + 406, + 545, + 517 + ], + "type": "text", + "content": " is sent back to participant " + }, + { + "bbox": [ + 304, + 406, + 545, + 517 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_k" + }, + { + "bbox": [ + 304, + 406, + 545, + 517 + ], + "type": "text", + "content": ", who decrypts it to compute the overall loss via Eqs. 10, 5 and 6." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 521, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 521, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 521, + 545, + 713 + ], + "type": "text", + "content": "Privacy Analysis: In the CaPriDe learning framework, the only quantity that can potentially leak private information is the pairwise distillation loss between participants " + }, + { + "bbox": [ + 304, + 521, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathcal{L}}_{KL_{(k,l)}}" + }, + { + "bbox": [ + 304, + 521, + 545, + 713 + ], + "type": "text", + "content": ", which has already been aggregated over all the " + }, + { + "bbox": [ + 304, + 521, + 545, + 713 + ], + "type": "inline_equation", + "content": "N_{k}" + }, + { + "bbox": [ + 304, + 521, + 545, + 713 + ], + "type": "text", + "content": " samples. A curious participant " + }, + { + "bbox": [ + 304, + 521, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_k" + }, + { + "bbox": [ + 304, + 521, + 545, + 713 + ], + "type": "text", + "content": " attempting to reconstruct the logit values " + }, + { + "bbox": [ + 304, + 521, + 545, + 713 + ], + "type": "inline_equation", + "content": "z_{j,(k,l)}^m" + }, + { + "bbox": [ + 304, + 521, + 545, + 713 + ], + "type": "text", + "content": " of participant " + }, + { + "bbox": [ + 304, + 521, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_l" + }, + { + "bbox": [ + 304, + 521, + 545, + 713 + ], + "type": "text", + "content": " needs to dis-aggregate this sum and compute per-sample loss, which is infeasible in the absence of other constraints. Hence, the proposed CaPriDe learning framework is privacy-preserving. It must be emphasized that even gradient inversion attacks [9, 14] are significantly harder when the batch size is large (above 16) due to the same reason. The proposed method also has other advantages that enhance privacy: (i) A curious participant cannot query its collaborator adaptively. The encrypted samples are published once at the beginning and do not change. (ii) The responding participant can choose a" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8088" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "type": "text", + "content": "random subset of samples in each round to compute the aggregate loss. This can improve efficiency as well as provide a privacy benefit by eliminating correlation across multiple rounds. (iii) We can incorporate differential privacy (DP) by adding noise to the aggregate pairwise distillation loss. Since the framework allows exchange of only encrypted input samples and encrypted predictions, which never get decrypted during the execution of the protocol, the CaPriDe learning approach also preserves data confidentiality." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 192, + 128, + 205 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 192, + 128, + 205 + ], + "spans": [ + { + "bbox": [ + 47, + 192, + 128, + 205 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 212, + 110, + 224 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 212, + 110, + 224 + ], + "spans": [ + { + "bbox": [ + 47, + 212, + 110, + 224 + ], + "type": "text", + "content": "4.1. Datasets" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 230, + 287, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 230, + 287, + 266 + ], + "spans": [ + { + "bbox": [ + 46, + 230, + 287, + 266 + ], + "type": "text", + "content": "CIFAR-10 [17] is a dataset consisting of 60000 " + }, + { + "bbox": [ + 46, + 230, + 287, + 266 + ], + "type": "inline_equation", + "content": "32 \\times 32" + }, + { + "bbox": [ + 46, + 230, + 287, + 266 + ], + "type": "text", + "content": " RGB images from 10 classes, with 6000 images per class. It has 50000 training and 10000 test samples." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 266, + 287, + 351 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 266, + 287, + 351 + ], + "spans": [ + { + "bbox": [ + 46, + 266, + 287, + 351 + ], + "type": "text", + "content": "CIFAR-100 dataset [17] is similar to the CIFAR-10 dataset, but it has 100 classes containing 600 samples each. Training and evaluation sets are split in the same way as in CIFAR-10. CIFAR-100 dataset is used with data augmentation to account for the problem complexity. The data augmentations employed include random rotation (up to 15 degrees), random crop, and horizontal flip." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 351, + 287, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 351, + 287, + 399 + ], + "spans": [ + { + "bbox": [ + 46, + 351, + 287, + 399 + ], + "type": "text", + "content": "HAM10000 is a multi-class imbalanced dataset comprising of 10,015 dermoscopic images from diverse populations with 7 categories of pigmented lesions. We randomly perform an " + }, + { + "bbox": [ + 46, + 351, + 287, + 399 + ], + "type": "inline_equation", + "content": "85\\% / 15\\%" + }, + { + "bbox": [ + 46, + 351, + 287, + 399 + ], + "type": "text", + "content": " split for training/ testing." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 407, + 163, + 420 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 407, + 163, + 420 + ], + "spans": [ + { + "bbox": [ + 47, + 407, + 163, + 420 + ], + "type": "text", + "content": "4.2. Experimental Setup" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 426, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 426, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 426, + 289, + 715 + ], + "type": "text", + "content": "We compare the CaPriDe learning framework against the following baselines: vanilla FL based on FedAvg (FedAvg) [20], FL with local differential privacy (FedAvg + DP) [24], and deep mutual learning (DML) [27]. We use ResNet-18 architecture for all our experiments. Stochastic Gradient Descent (SGD) algorithm with momentum is used as the optimizer. For all methods, 25 epochs of local training happens before 75 rounds of collaboration. Momentum is set to 0.9, the initial learning rate is set to " + }, + { + "bbox": [ + 46, + 426, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\alpha_{k} = 0.1" + }, + { + "bbox": [ + 46, + 426, + 289, + 715 + ], + "type": "text", + "content": " and the learning rate decays in rounds 50 and 75 by a factor " + }, + { + "bbox": [ + 46, + 426, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\gamma = 0.1" + }, + { + "bbox": [ + 46, + 426, + 289, + 715 + ], + "type": "text", + "content": ". The batch size is set to 128 for CIFAR-10 and CIFAR-100 and 32 for HAM10000. The weight factor " + }, + { + "bbox": [ + 46, + 426, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\lambda_{k}" + }, + { + "bbox": [ + 46, + 426, + 289, + 715 + ], + "type": "text", + "content": " between cross-entropy and distillation losses is set to 50 for CIFAR-10 and CIFAR-100 and 20 for HAM10000. All algorithms are implemented using PyTorch and experiments are carried out on a single NVIDIA RTX A6000 GPU. Additionally, in CaPriDe and DML, the logits are scaled by a temperature parameter " + }, + { + "bbox": [ + 46, + 426, + 289, + 715 + ], + "type": "inline_equation", + "content": "T = 5.0" + }, + { + "bbox": [ + 46, + 426, + 289, + 715 + ], + "type": "text", + "content": ". Finally, for FedAvg + DP, the Gaussian noise mechanism is used with " + }, + { + "bbox": [ + 46, + 426, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\sigma_{g} = 0.01" + }, + { + "bbox": [ + 46, + 426, + 289, + 715 + ], + "type": "text", + "content": ". To implement an FHE-compatible version of ResNet-18, we use the Tile Tensors framework [1,2], which in turn relies on low-level FHE libraries such as HElib and SEAL. For encrypted inference, ReLU is replaced with a polynomial approximation of Gaussian Error Linear Unit" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 310, + 72, + 544, + 293 + ], + "blocks": [ + { + "bbox": [ + 310, + 72, + 544, + 293 + ], + "lines": [ + { + "bbox": [ + 310, + 72, + 544, + 293 + ], + "spans": [ + { + "bbox": [ + 310, + 72, + 544, + 293 + ], + "type": "image", + "image_path": "43a7b76ab875c0ebbfaa5638e55f6c51a369f325bf764b75d75e67e3419db4c8.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 304, + 545, + 337 + ], + "lines": [ + { + "bbox": [ + 305, + 304, + 545, + 337 + ], + "spans": [ + { + "bbox": [ + 305, + 304, + 545, + 337 + ], + "type": "text", + "content": "Figure 2. Accuracy trend of two " + }, + { + "bbox": [ + 305, + 304, + 545, + 337 + ], + "type": "inline_equation", + "content": "(K = 2)" + }, + { + "bbox": [ + 305, + 304, + 545, + 337 + ], + "type": "text", + "content": " collaboratively trained models on different datasets (rows) and for different partition strategies (columns)." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 358, + 508, + 371 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 358, + 508, + 371 + ], + "spans": [ + { + "bbox": [ + 306, + 358, + 508, + 371 + ], + "type": "interline_equation", + "content": "(g (x) = 0. 0 7 1 1 + 0. 5 x + 0. 2 5 7 6 x ^ {2} - 0. 0 1 2 8 x ^ {4}).", + "image_path": "2d4b5b0764cc37aecd50958852bd373d925e57de86e247b580621c1c6ca422d9.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 372, + 545, + 528 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 372, + 545, + 528 + ], + "spans": [ + { + "bbox": [ + 304, + 372, + 545, + 528 + ], + "type": "text", + "content": "We consider three types of strategies to partition the training dataset among the participants: (i) Homogeneous: each participant has an equal number of samples per class, (ii) Heterogeneous: each participant has an unequal number of samples determined randomly (both total number and number of samples per class), (iii) Non-overlapping class distribution: (denoted as \"no class overlap\") each participant has samples from a non-overlapping subset of classes (e.g., for " + }, + { + "bbox": [ + 304, + 372, + 545, + 528 + ], + "type": "inline_equation", + "content": "K = 2" + }, + { + "bbox": [ + 304, + 372, + 545, + 528 + ], + "type": "text", + "content": ", participants get even and odd classes). While homogeneous partitioning ensures that data distributions of participants are iid, heterogeneous split simulates typical non-iid scenarios, and the no class overlap case corresponds to extreme statistical heterogeneity." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 535, + 476, + 548 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 535, + 476, + 548 + ], + "spans": [ + { + "bbox": [ + 306, + 535, + 476, + 548 + ], + "type": "text", + "content": "4.3. Collaborative Learning Results" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 554, + 545, + 649 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 554, + 545, + 649 + ], + "spans": [ + { + "bbox": [ + 304, + 554, + 545, + 649 + ], + "type": "text", + "content": "To begin with, we establish the utility (collaborative gain) of the proposed CaPriDe learning framework for the case of " + }, + { + "bbox": [ + 304, + 554, + 545, + 649 + ], + "type": "inline_equation", + "content": "K = 2" + }, + { + "bbox": [ + 304, + 554, + 545, + 649 + ], + "type": "text", + "content": " participants on all three datasets and all three partition strategies. The results of these experiments are shown in Figure 2. In each case, the collaboration gain of CaPriDe learning is positive and the actual collaboration gain values can be found in Table 1. We make the following observations based on these results." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 317, + 661, + 545, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 661, + 545, + 685 + ], + "spans": [ + { + "bbox": [ + 317, + 661, + 545, + 685 + ], + "type": "text", + "content": "- CaPriDe learning works not only in the iid setting but also equally well in the two non iid settings. This" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 306, + 693, + 545, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 693, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 693, + 545, + 713 + ], + "type": "text", + "content": "2Since HAM10000 already has a massive class imbalance and an odd number of class, we omit the no class overlap setting for this dataset." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8089" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 48, + 70, + 286, + 259 + ], + "blocks": [ + { + "bbox": [ + 48, + 70, + 286, + 259 + ], + "lines": [ + { + "bbox": [ + 48, + 70, + 286, + 259 + ], + "spans": [ + { + "bbox": [ + 48, + 70, + 286, + 259 + ], + "type": "table", + "html": "
DatasetSettingKFedAvgFedAvg+DPCaPriDe
CIFAR-10Homogeneous23.46-20.541.58
CIFAR-10Homogeneous59.22-11.83.55
CIFAR-10Homogeneous1015.442.585.52
CIFAR-10Heterogeneous26.12-17.040.93
CIFAR-10Heterogeneous513.78-7.845.06
CIFAR-10Heterogeneous1021.815.696.20
CIFAR-10No class overlap229.8511.088.16
CIFAR-100Homogeneous27.08-30.974.19
CIFAR-100Homogeneous522.68-13.819.10
CIFAR-100Homogeneous1026.10-0.1111.69
CIFAR-100Heterogeneous29.59-28.896.28
CIFAR-100Heterogeneous522.50-13.269.23
CIFAR-100Heterogeneous1034.383.8610.68
CIFAR-100No class overlap219.405.147.75
HAM10000Homogeneous20.970.481.81
HAM10000Heterogeneous21.93-1.211.67
", + "image_path": "56d7237cbf3bee473ba0218f864896f783fcc45292f809ea4f8533990ac36947.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 267, + 288, + 312 + ], + "lines": [ + { + "bbox": [ + 46, + 267, + 288, + 312 + ], + "spans": [ + { + "bbox": [ + 46, + 267, + 288, + 312 + ], + "type": "text", + "content": "Table 1. Collaborative gain (\\%) on all datasets and all partition strategies for different number of participants " + }, + { + "bbox": [ + 46, + 267, + 288, + 312 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 46, + 267, + 288, + 312 + ], + "type": "text", + "content": " using different algorithms (columns 4-6). DML results are not reported here because there is negligible difference between CaPriDe and DML." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 66, + 334, + 286, + 358 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 334, + 286, + 358 + ], + "spans": [ + { + "bbox": [ + 66, + 334, + 286, + 358 + ], + "type": "text", + "content": "demonstrates that knowledge distillation is reasonably effective even under extreme statistical heterogeneity." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 58, + 368, + 287, + 594 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 59, + 368, + 287, + 464 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 368, + 287, + 464 + ], + "spans": [ + { + "bbox": [ + 59, + 368, + 287, + 464 + ], + "type": "text", + "content": "- There is a negligible difference between the performance of DML (which uses true KL divergence) and CaPriDe learning (which uses approximate KL divergence) in all the settings. This clearly justifies the dropping of the normalization (third) term in Eq. 8 to obtain the encryption-friendly distillation loss in Eq. 9. While DML offers no confidentiality or privacy, CaPriDe learning offers both without any utility cost." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 58, + 474, + 287, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 474, + 287, + 594 + ], + "spans": [ + { + "bbox": [ + 58, + 474, + 287, + 594 + ], + "type": "text", + "content": "- In almost all the settings, there is a huge difference in the collaboration gain between FedAvg and CaPriDe learning. This is unsurprising because FedAvg involves sharing and aggregation of full model parameters leading to high utility at a significant cost to privacy. On the other hand, FedAvg + DP provides strong privacy, but this comes with a significant degradation in utility. Hence, it can be argued that CaPriDe learning achieves a good trade-off between privacy and utility compared to standard FL algorithms." + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "text", + "content": "Next, we evaluate the utility of the proposed method with the different number of participants " + }, + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "inline_equation", + "content": "K = 2,5,10" + }, + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "text", + "content": " on CIFAR-10 and CIFAR-100 datasets based on the homogeneous partition (iid) setting. From Figure 3 and Table 1 we observe that the collaboration gain increases with the number of participants. This is to be expected because the training samples per participant become less when there are more participants, consequently resulting in lower individual accuracy." + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 310, + 72, + 544, + 223 + ], + "blocks": [ + { + "bbox": [ + 310, + 72, + 544, + 223 + ], + "lines": [ + { + "bbox": [ + 310, + 72, + 544, + 223 + ], + "spans": [ + { + "bbox": [ + 310, + 72, + 544, + 223 + ], + "type": "image", + "image_path": "08b0e21db30a0721f446a7ba76a23281ac45917998144ad7fc2ec990142e0562.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 232, + 545, + 255 + ], + "lines": [ + { + "bbox": [ + 305, + 232, + 545, + 255 + ], + "spans": [ + { + "bbox": [ + 305, + 232, + 545, + 255 + ], + "type": "text", + "content": "Figure 3. Accuracy with different number of participants " + }, + { + "bbox": [ + 305, + 232, + 545, + 255 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 305, + 232, + 545, + 255 + ], + "type": "text", + "content": " based on the homogeneous partition (iid) setting." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 307, + 269, + 386, + 344 + ], + "blocks": [ + { + "bbox": [ + 307, + 269, + 386, + 344 + ], + "lines": [ + { + "bbox": [ + 307, + 269, + 386, + 344 + ], + "spans": [ + { + "bbox": [ + 307, + 269, + 386, + 344 + ], + "type": "image", + "image_path": "a92183316b47486ba36b76855ea902e5a985c9b32356968e9760a3d7000f5c81.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 317, + 346, + 375, + 357 + ], + "lines": [ + { + "bbox": [ + 317, + 346, + 375, + 357 + ], + "spans": [ + { + "bbox": [ + 317, + 346, + 375, + 357 + ], + "type": "text", + "content": "(a) Homogeneous" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 387, + 269, + 465, + 344 + ], + "blocks": [ + { + "bbox": [ + 387, + 269, + 465, + 344 + ], + "lines": [ + { + "bbox": [ + 387, + 269, + 465, + 344 + ], + "spans": [ + { + "bbox": [ + 387, + 269, + 465, + 344 + ], + "type": "image", + "image_path": "0efbfc2e767cf394f01e1fc1c1ba384a62e69cfbceae82c5bd82ed4a8b8ff32b.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 395, + 346, + 456, + 357 + ], + "lines": [ + { + "bbox": [ + 395, + 346, + 456, + 357 + ], + "spans": [ + { + "bbox": [ + 395, + 346, + 456, + 357 + ], + "type": "text", + "content": "(b) Heterogeneous" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 467, + 269, + 545, + 344 + ], + "blocks": [ + { + "bbox": [ + 467, + 269, + 545, + 344 + ], + "lines": [ + { + "bbox": [ + 467, + 269, + 545, + 344 + ], + "spans": [ + { + "bbox": [ + 467, + 269, + 545, + 344 + ], + "type": "image", + "image_path": "d385c9c34c991f29411068436b1d8ca28973a1c6a365138fe76c88d8a3deff76.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 482, + 346, + 529, + 357 + ], + "lines": [ + { + "bbox": [ + 482, + 346, + 529, + 357 + ], + "spans": [ + { + "bbox": [ + 482, + 346, + 529, + 357 + ], + "type": "text", + "content": "(c) No overlap" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 307, + 413, + 385, + 490 + ], + "blocks": [ + { + "bbox": [ + 305, + 366, + 545, + 399 + ], + "lines": [ + { + "bbox": [ + 305, + 366, + 545, + 399 + ], + "spans": [ + { + "bbox": [ + 305, + 366, + 545, + 399 + ], + "type": "text", + "content": "Figure 4. Per-class accuracy of CaPriDe learning evaluated on CIFAR-10 averaged across 5 participants. Dashed lines represent mean accuracy and error bars indicate standard deviation." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 307, + 413, + 385, + 490 + ], + "lines": [ + { + "bbox": [ + 307, + 413, + 385, + 490 + ], + "spans": [ + { + "bbox": [ + 307, + 413, + 385, + 490 + ], + "type": "image", + "image_path": "d6a14f28d4d29f9e9313aea1d6fa5fa0f6a697790d9e6354e97fd7099b8a5fc9.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 317, + 493, + 375, + 502 + ], + "lines": [ + { + "bbox": [ + 317, + 493, + 375, + 502 + ], + "spans": [ + { + "bbox": [ + 317, + 493, + 375, + 502 + ], + "type": "text", + "content": "(a) Homogeneous" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 305, + 512, + 545, + 533 + ], + "lines": [ + { + "bbox": [ + 305, + 512, + 545, + 533 + ], + "spans": [ + { + "bbox": [ + 305, + 512, + 545, + 533 + ], + "type": "text", + "content": "Figure 5. Per-participant accuracy of CaPriDe learning evaluated on CIFAR-100 dataset." + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 387, + 415, + 464, + 490 + ], + "blocks": [ + { + "bbox": [ + 387, + 415, + 464, + 490 + ], + "lines": [ + { + "bbox": [ + 387, + 415, + 464, + 490 + ], + "spans": [ + { + "bbox": [ + 387, + 415, + 464, + 490 + ], + "type": "image", + "image_path": "d578d0c4b1510d5adfd0edbfab2897f8c6ed6d25b5ffa42683a90e928c362176.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 395, + 493, + 456, + 502 + ], + "lines": [ + { + "bbox": [ + 395, + 493, + 456, + 502 + ], + "spans": [ + { + "bbox": [ + 395, + 493, + 456, + 502 + ], + "type": "text", + "content": "(b) Heterogeneous" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 467, + 415, + 545, + 490 + ], + "blocks": [ + { + "bbox": [ + 467, + 415, + 545, + 490 + ], + "lines": [ + { + "bbox": [ + 467, + 415, + 545, + 490 + ], + "spans": [ + { + "bbox": [ + 467, + 415, + 545, + 490 + ], + "type": "image", + "image_path": "4e4fc757f199692b37ad1218748a2bf0cf5b388b180730e0331ca7dfc8c3ddb3.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 482, + 493, + 529, + 502 + ], + "lines": [ + { + "bbox": [ + 482, + 493, + 529, + 502 + ], + "spans": [ + { + "bbox": [ + 482, + 493, + 529, + 502 + ], + "type": "text", + "content": "(c) No overlap" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 558, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 558, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 558, + 545, + 713 + ], + "type": "text", + "content": "Furthermore, we evaluate the class-wise and participant-wise accuracy of our CaPriDe learning algorithm with " + }, + { + "bbox": [ + 304, + 558, + 545, + 713 + ], + "type": "inline_equation", + "content": "K = 5" + }, + { + "bbox": [ + 304, + 558, + 545, + 713 + ], + "type": "text", + "content": " participants. Figure 4 shows class-wise accuracy averaged across 5 participants for each setting. From this figure, it is clear that almost all the classes benefit from collaboration, and the per-class accuracy of the individual models after collaboration has a lower standard deviation compared to the case without collaboration, especially in homogeneous and heterogeneous settings. Thus, in these two settings, the individual models become similar to each other after collaboration proving the successful transfer of knowledge between participants. Note that Figure 4 displays the average top-1 accuracy of the participants. The collaboration gains" + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8090" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 286, + 215 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 286, + 215 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 286, + 215 + ], + "type": "text", + "content": "in the no class overlap case are significantly higher when top-5 accuracy is considered. Finally, Figure 5 illustrates participant-wise accuracy over all classes for each setting. It is obvious that each participant has a positive collaboration gain, which is the eventual goal that we set out to achieve. Moreover, all the 5 participant models have a similar accuracy at the end of the collaboration, once again demonstrating good knowledge transfer. Other results such as the sensitivity of the proposed method to weight factor " + }, + { + "bbox": [ + 48, + 72, + 286, + 215 + ], + "type": "inline_equation", + "content": "\\lambda_{k}" + }, + { + "bbox": [ + 48, + 72, + 286, + 215 + ], + "type": "text", + "content": ", the sensitivity of FedAvg + DP to noise variance, and the comparison between approximate KL divergence and L2 loss are included in the supplementary material." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 224, + 206, + 236 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 224, + 206, + 236 + ], + "spans": [ + { + "bbox": [ + 48, + 224, + 206, + 236 + ], + "type": "text", + "content": "4.4. FHE Implementation Results" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 243, + 286, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 243, + 286, + 422 + ], + "spans": [ + { + "bbox": [ + 48, + 243, + 286, + 422 + ], + "type": "text", + "content": "For FHE implementation using the Tile Tensors framework, Table 2 summarizes the information about all three datasets at the same security level of 128 bits. Two critical metrics to be kept in mind are the time required for encrypted inference and the memory consumed by a ciphertext. While encrypted inference time determines the computational complexity and latency involved in each round, the ciphertext size determines the communication burden. Note that encrypted inference can be performed on a persample basis or a batch of samples [4]. The single sample inference mode is preferred for machine learning as-a-service applications, where low latency is a critical requirement. However, the batched inference is better suited for the CaPriDe learning framework, where only the overall time required to complete inference on all the samples matter." + } + ] + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 48, + 432, + 286, + 514 + ], + "blocks": [ + { + "bbox": [ + 48, + 432, + 286, + 514 + ], + "lines": [ + { + "bbox": [ + 48, + 432, + 286, + 514 + ], + "spans": [ + { + "bbox": [ + 48, + 432, + 286, + 514 + ], + "type": "table", + "html": "
Security LevelCIFAR-10CIFAR-100HAM10000
128128128
Number of slots163841638416384
Time taken to encrypt one sample90 ms103 ms619 ms
Ciphertext size of one sample29.101 KB29.152 KB1.359 MB
Time taken to encrypt a batch of 32 samples1.31 s1.26 s15.57 s
Encrypted inference of a batch of 32 samples110.21 s112.09 s896.12 s
", + "image_path": "fa950b2ec76fc5803e10412bd103b1b7c9943a9b4dec79f034cc26ec85a38405.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 523, + 286, + 555 + ], + "lines": [ + { + "bbox": [ + 48, + 523, + 286, + 555 + ], + "spans": [ + { + "bbox": [ + 48, + 523, + 286, + 555 + ], + "type": "text", + "content": "Table 2. FHE configuration parameters and memory and computational requirements for the chosen datasets based on ResNet-18 architecture." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 48, + 570, + 286, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 570, + 286, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 570, + 286, + 712 + ], + "type": "text", + "content": "The encrypted inference time for a batch of 32 samples is approximately 110s for CIFAR-10 and CIFAR-100 and 900s for HAM10000. Thus, if the local dataset of each participant has 1024 training samples, it will take around 1 hour (8 hours) for encrypted inference per participant and then computation of the pairwise distillation loss in each communication round for CIFAR-10 and CIFAR-100 (HAM10000). While this is a significant computational load, it is less of an issue in the cross-silo collaborative learning setting, especially in sectors such as finance and healthcare, where data confidentiality and privacy as well as accuracy gain are more critical. Moreover, the ciphertext" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 307, + 72, + 545, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 72, + 545, + 239 + ], + "spans": [ + { + "bbox": [ + 307, + 72, + 545, + 239 + ], + "type": "text", + "content": "size for one sample is approximately, 30KB for CIFAR-10 and CIFAR-100 and 1.36MB for HAM10000. For the same local dataset size of 1024, the encrypted dataset will consume 30MB for CIFAR-10 and CIFAR-100 and 1.36GB for HAM10000. Fortunately, the encrypted dataset is published and needs to be downloaded only once by each participant at the beginning of the protocol. During every communication round, only the encrypted predictions and pairwise losses are exchanged. Depending on the local dataset size, the number of classes, and the ciphertext packing strategy used, these intermediate values can be packed into at most a few tens of ciphertexts (note that this does not depend on the image size). Hence, the per round communication burden of CaPriDe learning framework is relatively small." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 308, + 242, + 545, + 528 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 242, + 545, + 528 + ], + "spans": [ + { + "bbox": [ + 308, + 242, + 545, + 528 + ], + "type": "text", + "content": "Limitations and Potential Solutions: Note that all the above complexity computations are based on the assumption of " + }, + { + "bbox": [ + 308, + 242, + 545, + 528 + ], + "type": "inline_equation", + "content": "K = 2" + }, + { + "bbox": [ + 308, + 242, + 545, + 528 + ], + "type": "text", + "content": " participants. As " + }, + { + "bbox": [ + 308, + 242, + 545, + 528 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 308, + 242, + 545, + 528 + ], + "type": "text", + "content": " increases, these requirements will increase linearly for each participant bringing the scalability of the system into question. Hence, the vanilla CaPriDe learning framework is more appropriate for a small network of participants. However, there are ways to improve scalability borrowing on ideas from cross-device FL. For example, instead of computing the pairwise distillation loss with respect to each participant, a random subset of participants can be sampled in each round for distillation. Similarly, the computational complexity for each participant can be significantly reduced by performing encrypted inference only on a randomly selected subset of encrypted samples. Furthermore, these actions can be determined at the local level without any central orchestration. The need to have FHE-friendly models for encrypted inference is a key limitation, which precludes the use of very deep models within the proposed framework. One potential solution is to employ two models at each participant: one deep model with large capacity and another smaller FHE-friendly model for encrypted inference on other participants' data and perform an additional mutual distillation between these two models after each round." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 308, + 548, + 370, + 560 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 548, + 370, + 560 + ], + "spans": [ + { + "bbox": [ + 308, + 548, + 370, + 560 + ], + "type": "text", + "content": "5. Summary" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 308, + 570, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 570, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 308, + 570, + 545, + 712 + ], + "type": "text", + "content": "Collaborative learning without compromising on data confidentiality and privacy is a valuable tool in applications with stringent data privacy requirements. We have proposed a new framework called CaPriDe learning based on inference in the FHE domain to achieve this goal. Knowledge distillation is used to exchange useful information between models, and a new distillation loss based on approximate KL divergence has been proposed to enable secure loss computation in the encrypted domain. While the proposed approach has shown promising results, its efficiency can be further improved. Other vulnerabilities such as poisoning attacks will also be addressed in the future." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 750, + 313, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 750, + 313, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 750, + 313, + 757 + ], + "type": "text", + "content": "8091" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 147 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 147 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 147 + ], + "type": "text", + "content": "[1] Ehud Aharoni, Allon Adir, Moran Baruch, Nir Drucker, Gilad Ezov, Ariel Farkash, Lev Greenberg, Rami Masalha, Guy Moshkowish, Dov Murik, Hayim Shaul, and Omri Soceanu. HeLayers: A Tile Tensors Framework for Large Neural Networks on Encrypted Data. arXiv:2011.01805v2, 2021. 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 148, + 288, + 191 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 148, + 288, + 191 + ], + "spans": [ + { + "bbox": [ + 53, + 148, + 288, + 191 + ], + "type": "text", + "content": "[2] Ehud Aharoni, Nir Drucker, Gilad Ezov, Hayim Shaul, and Omri Soceanu. Complex encoded tile tensors: Accelerating encrypted analytics. In IEEE S&P 20 (5), pages 35-43, 2022. 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 194, + 288, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 194, + 288, + 258 + ], + "spans": [ + { + "bbox": [ + 53, + 194, + 288, + 258 + ], + "type": "text", + "content": "[3] Keith Bonawitz, Vladimir Ivanov, Ben Kreuter, Antonio Marcedone, H. Brendan McMahan, Sarvar Patel, Daniel Ramage, Aaron Segal, and Karn Seth. Practical secure aggregation for privacy preserving machine learning. ACM Conference on Computer and Communications Security (ACM CCS), 2017. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 260, + 288, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 260, + 288, + 293 + ], + "spans": [ + { + "bbox": [ + 53, + 260, + 288, + 293 + ], + "type": "text", + "content": "[4] Alon Brutzkus, Ran Gilad-Bachrach, and Oren Elisha. Low latency privacy preserving inference. In International Conference on Machine Learning, pages 812-821, 2019. 8" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 295, + 288, + 338 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 295, + 288, + 338 + ], + "spans": [ + { + "bbox": [ + 53, + 295, + 288, + 338 + ], + "type": "text", + "content": "[5] Hongyan Chang, Virat Shejwalkar, Reza Shokri, and Amir Houmansadr. Cronus: Robust and heterogeneous collaborative learning with black-box knowledge transfer. In arXiv preprint arXiv:1912.11279, 2019. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 340, + 288, + 372 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 340, + 288, + 372 + ], + "spans": [ + { + "bbox": [ + 53, + 340, + 288, + 372 + ], + "type": "text", + "content": "[6] Jung Hee Cheon, Andrey Kim, Miran Kim, and Yongsoo Song. Homomorphic Encryption for Arithmetic of Approximate Numbers. 2016. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 374, + 288, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 374, + 288, + 418 + ], + "spans": [ + { + "bbox": [ + 53, + 374, + 288, + 418 + ], + "type": "text", + "content": "[7] Christopher A. Choquette-Choo, Natalie Dullerud, Adam Dziedzic, Yunxiang Zhang, Somesh Jha, Nicolas Papernot, and Xiao. Wang. CaPC Learning: Confidential and Private Collaborative Learning. Proceedings of ICLR, 2021. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 419, + 288, + 485 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 419, + 288, + 485 + ], + "spans": [ + { + "bbox": [ + 53, + 419, + 288, + 485 + ], + "type": "text", + "content": "[8] Deshpande, Nikhil and Shiffman, Gary M. Federated learning: The new thing in ai/ml for detecting financial crimes and managing risk. https://morningconsult.com/opinions/federated-learning-the-new-thing-in-ai-ml-for-detecting-financial-crimes-and-managing-risk/, 2021.1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 487, + 288, + 531 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 487, + 288, + 531 + ], + "spans": [ + { + "bbox": [ + 53, + 487, + 288, + 531 + ], + "type": "text", + "content": "[9] Jonas Geiping, Hartmut Bauermeister, Hannah Dröge, and Michael. Moeller. Inverting gradients-How easy is it to break privacy in federated learning? Advances in Neural Information Processing Systems (NeurIPS), 33, 2020. 1, 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 532, + 287, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 532, + 287, + 553 + ], + "spans": [ + { + "bbox": [ + 48, + 532, + 287, + 553 + ], + "type": "text", + "content": "[10] Craig. Gentry. Fully Homomorphic Encryption. PhD thesis, 2009. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 555, + 287, + 598 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 555, + 287, + 598 + ], + "spans": [ + { + "bbox": [ + 48, + 555, + 287, + 598 + ], + "type": "text", + "content": "[11] Qiushan Guo, Xinjiang Wang, Yichao Wu, Zhipeng Yu, Ding Liang, Xiaolin Hu, and Ping. Luo. Online Knowledge Distillation via Collaborative Learning. Proceedings of CVPR, 2020. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 600, + 287, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 600, + 287, + 644 + ], + "spans": [ + { + "bbox": [ + 48, + 600, + 287, + 644 + ], + "type": "text", + "content": "[12] Jenny Hamer, Mehryar Mohri, and Ananda Theertha Suresh. FedBoost: A Communication-Efficient Algorithm for Federated Learning. Proceedings of the 37th International Conference on Machine Learning, PMLR, 119:3973-3983, 2020. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 646, + 287, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 646, + 287, + 668 + ], + "spans": [ + { + "bbox": [ + 48, + 646, + 287, + 668 + ], + "type": "text", + "content": "[13] Geoffrey Hinton, Oriol Vinyls, and Jeff Dean. Distilling the Knowledge in a Neural Network. NeurIPS, 2014. 2, 4" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "type": "text", + "content": "[14] Jiwnoo Jeon, Kangwook Lee, Sewoong Oh, and Jungseul. Ok. Gradient Inversion with Generative Image Prior. Advances in Neural Information Processing Systems (NeurIPS), 34, 2021. 1, 5" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 699 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "text", + "content": "[15] Peter Kairouz, H. Brendan McMahan, Brendan Avent, Aurélien Bellet, and Mehdi. Bennis. Advances and Open Problems in Federated Learning. Now Publishers, 2021. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 108, + 545, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 108, + 545, + 150 + ], + "spans": [ + { + "bbox": [ + 307, + 108, + 545, + 150 + ], + "type": "text", + "content": "[16] Georgios A. Kaissis, Marcus R. Makowski, Daniel Rückert, and Rickmer F. Braren. Secure, privacy-preserving and federated machine learning in medical imaging. Nature Machine Intelligence, 2:305-311, 2020. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 152, + 545, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 152, + 545, + 184 + ], + "spans": [ + { + "bbox": [ + 307, + 152, + 545, + 184 + ], + "type": "text", + "content": "[17] Alex Krizhevsky. Gradient-based learning applied to document recognition. Technical report, University of Toronto, 2009. 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 186, + 545, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 186, + 545, + 229 + ], + "spans": [ + { + "bbox": [ + 307, + 186, + 545, + 229 + ], + "type": "text", + "content": "[18] Tao Lin, Lingjing Kong, Sebastian U Stich, and Martin Jaggi. Ensemble distillation for robust model fusion in federated learning. In Advances in Neural Information Processing Systems, volume 33, pages 2351-2363, 2020. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 231, + 545, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 231, + 545, + 274 + ], + "spans": [ + { + "bbox": [ + 307, + 231, + 545, + 274 + ], + "type": "text", + "content": "[19] Chuan Ma, Jun Li, Ming Ding, Howard H. Yang, Feng Shu, Tony Q. S. Quek, and Tony Q. S. Poor. On safeguarding privacy and security in the framework of federated learning. IEEE Network, pages 242-248, 2020. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 276, + 545, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 276, + 545, + 330 + ], + "spans": [ + { + "bbox": [ + 307, + 276, + 545, + 330 + ], + "type": "text", + "content": "[20] H. Brendan McMahan, Eider Moore, Daniel Ramage, Seth Hampson, and Blaise Agüera y. Arcas. Communication-Efficient Learning of Deep Networks from Decentralized Data. Proceedings of the 20th International Conference on Artificial Intelligence and Statistics (AISTATS), 2016. 2, 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 331, + 545, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 331, + 545, + 397 + ], + "spans": [ + { + "bbox": [ + 307, + 331, + 545, + 397 + ], + "type": "text", + "content": "[21] Fan Mo, Hamed Haddadi, Kleomenis Katevas, Eduard Marin, Diego Perino, and Nicolas Kourtellis. PPFL: Privacy-Preserving Federated Learning with Trusted Execution Environments. In Proceedings of the 19th Annual International Conference on Mobile Systems, Applications, and Services (MobiSys), page 94-108, 2021. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 399, + 545, + 452 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 399, + 545, + 452 + ], + "spans": [ + { + "bbox": [ + 307, + 399, + 545, + 452 + ], + "type": "text", + "content": "[22] Willem G van Panhuis, Proma Paul, Claudia Emerson, John Grefenstette, Richard Wilder, Abraham J Herbst, and Donald S Heymann, David Burke. A systematic review of barriers to data sharing in public health. BMC Public Health, 2014. 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 454, + 545, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 454, + 545, + 520 + ], + "spans": [ + { + "bbox": [ + 307, + 454, + 545, + 520 + ], + "type": "text", + "content": "[23] Micah J. Sheller, Brandon Edwards, G. Anthony Reina, Jason Martin, Sarthak Pati, Aikaterini Kotrotsou, Mikhail Milchenko, Weilin Xu, Daniel Marcus, Rivka R. Colen, and Spyridon. Bakas. Federated learning in medicine: facilitating multi-institutional collaborations without sharing patient data. Nature Scientific Reports, 10, 2020. 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 521, + 545, + 575 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 521, + 545, + 575 + ], + "spans": [ + { + "bbox": [ + 307, + 521, + 545, + 575 + ], + "type": "text", + "content": "[24] Kang Wei, Jun Li, Ming Ding, Chuan Ma, Howard H. Yang, and Farhad. Farokhi. Federated Learning With Differential Privacy: Algorithms and Performance Analysis. IEEE Transactions on Information Forensics and Security, 15:3454-3469, 2020. 1, 6" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 576, + 545, + 631 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 576, + 545, + 631 + ], + "spans": [ + { + "bbox": [ + 307, + 576, + 545, + 631 + ], + "type": "text", + "content": "[25] Jaehong Yoon, Wonyong Jeong, Giwoong Lee, Eunho Yang, and Sung Ju. Hwang. Federated Continual Learning with Weighted Inter-client Transfer. Proceedings of the 37th International Conference on Machine Learning, PMLR, 139:12073-12086, 2021. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 632, + 545, + 676 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 632, + 545, + 676 + ], + "spans": [ + { + "bbox": [ + 307, + 632, + 545, + 676 + ], + "type": "text", + "content": "[26] Chengliang Zhang, Suyi Li, Junzhe Xia, Wei Wang, Feng Yan, and Yang Liu. BatchCrypt: efficient homomorphic encryption for cross-silo federated learning. Proceedings of the USENIX Conference, page 493-506, 2020. 1" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 677, + 545, + 699 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 677, + 545, + 699 + ], + "spans": [ + { + "bbox": [ + 307, + 677, + 545, + 699 + ], + "type": "text", + "content": "[27] Ying Zhang, Tao Xiang, Timothy M. Hospedales, and Huchuan Lu. Deep Mutual Learning. CVPR, 2017. 2, 6" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8092" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/CafeBoost_ Causal Feature Boost To Eliminate Task-Induced Bias for Class Incremental Learning/6f2564b9-d9c3-4c4f-b871-b0003952027b_content_list.json b/2023/CafeBoost_ Causal Feature Boost To Eliminate Task-Induced Bias for Class Incremental Learning/6f2564b9-d9c3-4c4f-b871-b0003952027b_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..1bf2e1d8fc4fbd860b6d9c6faa1a3e742ef35224 --- /dev/null +++ b/2023/CafeBoost_ Causal Feature Boost To Eliminate Task-Induced Bias for Class Incremental Learning/6f2564b9-d9c3-4c4f-b871-b0003952027b_content_list.json @@ -0,0 +1,2026 @@ +[ + { + "type": "text", + "text": "CafeBoost: Causal Feature Boost to Eliminate Task-Induced Bias for Class Incremental Learning", + "text_level": 1, + "bbox": [ + 109, + 128, + 861, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Benliu Qiu", + "bbox": [ + 232, + 203, + 325, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Hongliang Li*", + "bbox": [ + 364, + 203, + 480, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Haitao Wen", + "bbox": [ + 509, + 203, + 604, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Heqian Qiu", + "bbox": [ + 640, + 203, + 733, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Lanxiao Wang*", + "bbox": [ + 233, + 222, + 357, + 238 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Fanman Meng", + "bbox": [ + 387, + 222, + 504, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Qingbo Wu", + "bbox": [ + 539, + 222, + 633, + 238 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Lili Pan", + "bbox": [ + 669, + 222, + 733, + 237 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "University of Electronic Science and Technology of China, Chengdu, China", + "bbox": [ + 186, + 239, + 782, + 256 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{qbenliu, haitaowen, hqqiu, lanxiao.wang}@std.uestc.edu.cn", + "bbox": [ + 227, + 258, + 740, + 273 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{h1li, fmmeng, qbwu}@uestc.edu.cn panlili8255@gmail.com", + "bbox": [ + 228, + 277, + 736, + 292 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 325, + 313, + 342 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Continual learning requires a model to incrementally learn a sequence of tasks and aims to predict well on all the learned tasks so far, which notoriously suffers from the catastrophic forgetting problem. In this paper, we find a new type of bias appearing in continual learning, coined as task-induced bias. We place continual learning into a causal framework, based on which we find the task-induced bias is reduced naturally by two underlying mechanisms in task and domain incremental learning. However, these mechanisms do not exist in class incremental learning (CIL), in which each task contains a unique subset of classes. To eliminate the task-induced bias in CIL, we devise a causal intervention operation so as to cut off the causal path that causes the task-induced bias, and then implement it as a causal debias module that transforms biased features into unbiased ones. In addition, we propose a training pipeline to incorporate the novel module into existing methods and jointly optimize the entire architecture. Our overall approach does not rely on data replay, and is simple and convenient to plug into existing methods. Extensive empirical study on CIFAR-100 and ImageNet shows that our approach can improve accuracy and reduce forgetting of well-established methods by a large margin.", + "bbox": [ + 75, + 347, + 473, + 694 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 709, + 209, + 724 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Deep learning has been widely applied in many fields like computer vision, social media analytics, natural language processing, etc. The standard paradigm of deep learning is to train models on a prepared dataset, of which all data are accessible in the whole training stage. However, in most real-world applications, a sequence of tasks arrives incrementally, which requires models to learn continuously from a new task, namely continual learning [27]. Despite fine-tuning achieves this goal somewhat, such naive method forgets obtained knowledge of previous tasks after learning", + "bbox": [ + 75, + 727, + 468, + 878 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/611a0457922a71cbe8aa8c7ad14ff5e4cf77871ed1591514252685cb139a7c54.jpg", + "image_caption": [ + "(a) Illustration of CIL" + ], + "image_footnote": [], + "bbox": [ + 511, + 321, + 864, + 414 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/22da54b8de582da85f61a1c55395d23cc884fa1b07d0a717f8cbebab9c3e8dd7.jpg", + "image_caption": [ + "(b) Task-induced bias between class features" + ], + "image_footnote": [], + "bbox": [ + 514, + 425, + 879, + 489 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/001c7710a5d3ca6c868622763ceebbd82037903c1918bd266a6171e7d2a197fb.jpg", + "image_caption": [ + "(c) Task-induced bias between images and labels", + "Figure 1. Illustration of CIL, task-induced bias and their detrimental effects. (a) Illustration of CIL. (b) Task-induced bias between class features. Two classes in a task are correlated by a task identifier. (c) Task-induced bias between images and labels. Images and labels in a task are correlated by a task identifier. In (b)(c), the first column describes perspectives of task-induced bias, the second shows the underlying causal effect causing task-induced bias, and the last illustrates adverse influence of task-induced bias." + ], + "image_footnote": [], + "bbox": [ + 509, + 500, + 867, + 570 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "a new task, dubbed as catastrophic forgetting [18].", + "bbox": [ + 500, + 704, + 834, + 718 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Much effort has been devoted as a remedy for alleviating this problem in literature [2, 3, 11-14, 16, 33]. One general direction is to preserve knowledge representations of previous tasks by knowledge distillation [10] and by storing a few observed exemplars for replay [22]. [11] discovers the class imbalance between old classes and new ones is a key challenge of large-scale applications in class incremental learning (CIL), in which classes of a new task never appear in the previous tasks, as illustrated in Fig. 1 (a). [2,7,31,35] all follow this direction and propose effective methods to overcome the class imbalance. However, these methods ignore bias problems in continual learning. In this work, we", + "bbox": [ + 496, + 719, + 892, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding authors.", + "bbox": [ + 94, + 886, + 227, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "16016", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "consider the problem of CIL in a novel perspective of bias. We think that catastrophic forgetting is the reflection of a new-type bias, coined as task-induced bias.", + "bbox": [ + 76, + 90, + 468, + 136 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Deep neural networks are easily misled by language bias [1] and vision bias [6, 9, 29], which are both derived from spurious correlations of two or more items in different positions of a sentence or an image. In contrast to these two bias, the task-induced bias occurs across different images in a common task. In CIL, data of specific classes are accessible only in a certain task if replay buffer is not applied. A deep neural network is capable of learning correlations among images of different categories or between images and labels as possible as it can, which includes a few spurious correlations, namely biases, leading to a limited generalization. On the one hand, a portion of these image correlations can be shared beneficially in this task, but these benefits cannot share among different tasks. For example, in Fig. 1 (b), color is sufficient to discriminate elephants and leopards in a specific task since elephants are grey while leopards are orange, and thus deep models have no motivation to capture more complex characteristic, like stripe. This is detrimental to the feature extractor because it captures a few details that cannot generalize among all tasks. On the other hand, as shown in Fig. 1 (c), a classifier will be biased to new classes in the current task [11, 31, 35], because the classes in the training data are all the new classes and to predict new classes is more possible to be correct.", + "bbox": [ + 76, + 137, + 470, + 500 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To better understand the task-induced bias in continual learning, we model three continual learning scenarios [27] into a common causal inference framework, which merely contains three necessary variables: image, label, and task identifier. Thanks to the framework, we discover that the task-induced bias is minor in task incremental and domain incremental learning due to innate mechanisms that cut off an inappropriate causal path, which partially explain why the two scenarios are easier to tackle than CIL [27]. For the existence of the causal path in CIL, we seek an active causal intervention to close the path. To implement the intervention, we devise a simple but effective causal debias module, which is established on attention mechanisms [28]. Subsequently, we incorporate this module into existing CIL architectures and propose a training pipeline to optimize it. Our overall approach transforms causally biased features into unbiased ones, and hence we refer to it as Casual Feature Boost (CafeBoost). Finally, we systematically compare accuracy and forgetting of well-established methods with ours in various settings. Experimental results show that CafeBoost yields significant and consistent performance boost for existing methods in CIFAR-100 and ImageNet by a large margin $(0.67\\% \\sim 12.08\\%$ on accuracy).", + "bbox": [ + 76, + 500, + 468, + 847 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 76, + 851, + 218, + 867 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Continual learning. Existing CIL works mostly aim to maintain knowledge representations of previous tasks, in-", + "bbox": [ + 76, + 869, + 468, + 900 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/78f4c3a1b646b817171e8d8de5ba734e13f29b73a66eaabb932008587fc81506.jpg", + "image_caption": [ + "Figure 2. Overview of Casual Feature Boost (CafeBoost). Green blocks are learnable, whereas white blocks are fixed. Conv block is updated with a smaller learning rate than Causal block. The training objectives including distillation loss, margin ranking loss, etc., are omitted in this figure, since they depend on baselines. Whether to use replay buffer $\\mathcal{M}$ depends on specific baselines." + ], + "image_footnote": [], + "bbox": [ + 514, + 95, + 879, + 305 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "including replay-based methods and distillation-based methods. Replay-based methods [22, 25, 26] store a few examples to represent previous data and replay the stored data in the current training stage. The purpose of relay-based methods is to select the most representative samples of the entire training dataset. iCaRL [22] uses herding strategy [30] to choose typical exemplars for future training. [25] calculates example influence of stability and plasticity for every example in validation sets, based on which it selects examples for replay. Distillation-based methods [3,11,23,24] focus on transferring learned knowledge of a previous model to the current one. LUCIR [11] combines cosine normalization, less-forget constraint, and inter-class separation to mitigate the forgetting caused by class imbalance. GeoDL [24] utilizes geodesic path for knowledge distillation which considers the gradual change between subspaces of tasks. CSCCT [3] encourages features of each class to locate in an appropriate position of the feature space, and utilizes inter-class relationships to transfer inter-class knowledge positively or negatively. CwD [23] scatters uniformly representations of each class in the initial task, which significantly benefits performance in subsequent tasks. These two types of methods are usually incorporated together to achieve better performance.", + "bbox": [ + 496, + 401, + 892, + 763 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Spurious correlation. Spurious correlation is also known as bias. [19] has extensively investigated the bias problem in machine learning. Many methods [1,9,29] have been proposed to eliminate bias in their specific applications. [1] introduces counterfactual in the training to eliminate bias in visual question answering datasets. [9] proposes a new Equalizer model to force a model look at a person instead of inferring a person's gender by context cues. [29] applies causal intervention to alleviate vision bias", + "bbox": [ + 496, + 763, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "16017", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "among detected objects in an image.", + "bbox": [ + 76, + 90, + 318, + 104 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Causal inference. Causal inference [21] aims to establish reasonable causal graphs for specific applications, then analyze problems and propose solutions based on the established causal graph. Causal intervention and counterfactual process are two powerful tools of causal inference. These two tools have been ubiquitously applied in multiple applications, such as out-of-distribution recognition [17], visual question answering [1], image captioning [34], knowledge distillation [6], etc. Recently, [12] builds a causal view to analyze the cause of forgetting in class incremental learning, and designs an anti-forgetting technique so as to distill causal effect of data without replay.", + "bbox": [ + 75, + 108, + 470, + 289 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Preliminary", + "text_level": 1, + "bbox": [ + 76, + 301, + 202, + 318 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Continual learning scenarios", + "text_level": 1, + "bbox": [ + 76, + 327, + 333, + 343 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Continual learning can be categorized into three scenarios [27] in the light of whether a task identifier is given in a test phase and whether a model is required to infer which task it meets. If the task identifier is available during both training and test phases, this scenario is task incremental. If the task identifier is accessible in training phase but it is only necessary for a model to solve the task at hand, this scenario is domain incremental. Similar to the domain incremental scenario, the class incremental scenario doesn't provide task identifiers in test phase, but it requires a model to infer the current task identifier. Generally speaking, the class incremental scenario is the most difficult among the three, whereas the task incremental scenario is the easiest [27].", + "bbox": [ + 75, + 353, + 468, + 547 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To formulate continual learning, we assume a sequence of $T$ tasks is given as $\\{\\mathcal{T}_0,\\mathcal{T}_1,\\dots,\\mathcal{T}_{T - 1}\\}$ , each with data $\\mathcal{D}_t,t = 0,1,\\ldots ,T - 1$ . Every sample in $\\mathcal{D}_t$ is a tuple of an image $\\mathbf{x}_i$ and a corresponding label $y_{i}$ , and hence we can denote the data of task $t$ as $\\mathcal{D}_t = \\{(\\mathbf{x}_i,y_i)_{i = 1}^{N_t}\\}$ where $N_{t}$ is the number of samples, $\\mathbf{x}_i\\in \\mathcal{X}_t$ and $y_{i}\\in \\mathcal{Y}_{t}$ . Replay-based continual learning methods require a memory buffer $\\mathcal{M}$ to store partial data of previous tasks. For the task $\\mathcal{T}_0$ , only data $\\mathcal{D}_0$ can be used for training, but for the task $\\mathcal{T}_t,t > 0$ , we can train models using data $\\mathcal{M}\\cup \\mathcal{D}_t$ . For the task incremental scenario, task identifier $\\mathcal{T}$ is given as an input to models in test phase but not for the other two scenarios.", + "bbox": [ + 75, + 551, + 468, + 733 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "After the above formulation, we elucidate distinctions of three incremental scenarios in a formal style [4]. Let $P(\\cdot)$ denote a marginal distribution. The common ground of three scenarios is $P(\\mathcal{X}_t) \\neq P(\\mathcal{X}_{t + 1})$ for all $t$ . Their distinctions lie in label set $\\mathcal{Y}_t$ and its marginal distribution $P(\\mathcal{Y}_t)$ . For task incremental learning, $\\mathcal{Y}_t \\neq \\mathcal{Y}_{t + 1}$ . For domain incremental learning, $\\mathcal{Y}_t = \\mathcal{Y}_{t + 1}$ and $P(\\mathcal{Y}_t) = P(\\mathcal{Y}_{t + 1})$ . These two scenarios have no requirement on inferring task identifiers. Conversely, the class incremental learning that defines $\\mathcal{Y}_t \\subset \\mathcal{Y}_{t + 1}$ and $P(\\mathcal{Y}_t) \\neq P(\\mathcal{Y}_{t + 1})$ , requires models to infer which label set is used for an arbitrary task.", + "bbox": [ + 75, + 734, + 470, + 901 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/f08dc029d1598dc6b2016c5217f8a1b49735b2e71e782590799b520563fc7740.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 542, + 88, + 638, + 162 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/48cfaa10052886a2460f9cc3f076be9842e0fe7d12daf515695bc285c43b6246.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 651, + 90, + 743, + 161 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/1b55d7506b2fb981f50787d5f8ee3cfa13698623a426e647e1022a83ec410f0c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 759, + 90, + 852, + 161 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/533436adbcd9fc65d1aa5fe309363ae21e3f3f0da36b2b5fbd7303d97f695ffc.jpg", + "image_caption": [ + "Figure 3. Causal graphs of supervised learning and continual learning. Gray nodes means unconditioned variables. (a) Causal graphs of supervised learning in training and test phase. (b) Causal graphs of continual learning. The red task variable depends on specific continual scenarios. (c) Causal path between images and labels in task incremental learning. The path $X \\gets T \\to Y$ is impeded because variable $T$ is conditioned. (d) Causal path between images and labels in domain incremental learning. The path $X \\gets T \\to Y$ is cut off because of $T \\nrightarrow Y$ . (e) Causal path between images and labels in class incremental learning. The path $X \\gets T \\to Y$ causes task-induced bias." + ], + "image_footnote": [], + "bbox": [ + 540, + 164, + 637, + 233 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/134c16f3195c1b63fac8a792eb263ee83fbe35bd61b7905d52b3a8d4ff1f7e74.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 650, + 164, + 748, + 232 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/ec6fb4e487aa0966b00f3427d943d74afcc567e0db56dbd8d8d0a687938a3e6a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 758, + 164, + 856, + 232 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Causal graphs in continual learning", + "text_level": 1, + "bbox": [ + 500, + 397, + 808, + 412 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "A causal graph is a directed acyclic graph, of which nodes denote variables and directed edges denote the causality between two variables. We treat images, labels and task identifiers as nodes in our causal graphs. Although only three nodes are incorporated into our causal graph, it is powerful to reveal underlying characteristics of continual learning as we will show in Sec. 4.", + "bbox": [ + 496, + 416, + 890, + 521 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We denote an image as $X$ , a label as $Y$ and a task identifier as $T$ in Fig. 3. The edges are listed below: $X \\to Y$ means label $Y$ is deduced from image $X$ ; $T \\to X$ denotes that task identifier $T$ have a causal influence to image $X$ ; $T \\to Y$ denotes task identifier $T$ causally affect label $Y$ .", + "bbox": [ + 496, + 522, + 890, + 597 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "For supervised learning in Fig. 3 (a), its causal graph only contains an image $X$ , a label $Y$ and an causal path $X \\to Y$ that represents a model to capture the causality between two variables. In training phase of supervised learning, both image $X$ and label $Y$ are accessible and we utilize them to learn the causal effect into the model. But in test phase, an image $X$ and the causal path are known and $Y$ is the label to be inferred. For three continual learning scenarios, images, labels and task identifiers are all given in training phase. As a result, they have the same causal graph for training in Fig. 3 (b). We illustrate their distinctions in test phase in Fig. 3 (b). Due to the involvement of the task identifier $T$ , $X$ may have an indirect effect on $Y$ through the causal path $X \\gets T \\to Y$ , solely dependent on $T$ . We will explain it in Sec. 4. See Appendix for additional explanation of causal graphs.", + "bbox": [ + 496, + 598, + 892, + 839 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4. Methodology", + "text_level": 1, + "bbox": [ + 500, + 845, + 635, + 861 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we introduce the task-induced bias through the causal framework, a causal intervention for", + "bbox": [ + 500, + 869, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "16018", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "CIL, its corresponding implementation, and the entire pipeline of our approach.", + "bbox": [ + 75, + 90, + 468, + 119 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1. Task-induced bias in three continual scenarios", + "text_level": 1, + "bbox": [ + 76, + 122, + 467, + 137 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "According to the causal graphs, we can explain why task incremental learning is the most effortless scenario among the three and why class incremental learning is the most challenging. We think that this phenomenon is caused by task-induced bias.", + "bbox": [ + 75, + 138, + 467, + 212 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In Fig. 3 (b), the causal path $X \\leftarrow T \\rightarrow Y$ is of a fork structure, in which two kid nodes are independent conditioned on the parent. [21] points out that if $T$ is unconditioned, the causal influence of $X$ to $Y$ exists, causing the task-induced bias. We give a simple example to help understand this statement. A causal graph lighter $\\leftarrow$ smoking $\\rightarrow$ cancer means that a smoking person has a possibility to bring a lighter and get a lung cancer. Whether a person brings a lighter is usually independent of getting a lung cancer. Nevertheless, a smoker is highly possible to bring a lighter and to get a lung cancer, which leads to observe a correlation between lighter and cancer. If we can control the confounder smoking and collect data about non-smoking persons who usually bring lighters, then we can draw a correct conclusion that lighter is independent on cancer.", + "bbox": [ + 75, + 213, + 467, + 452 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Fortunately, mechanisms to eliminate the task-induced bias naturally exist under setups of task incremental learning and domain incremental learning. In task incremental learning shown in Fig. 3 (c), the value of parent node (task identifier) is given, hence the indirected effect of $X$ on $Y$ is 0 and the task-induced bias between $X$ and $Y$ disappears due to the characteristic of the fork structure. In domain incremental learning, the label set remains the same along task identifiers, which is implied in $\\mathcal{Y}_t = \\mathcal{Y}_{t + 1}$ and $P(\\mathcal{Y}_t) = P(\\mathcal{Y}_{t + 1})$ . For this reason, the causal edge $T\\to Y$ is cut as shown in Fig. 3 (d), resulting in disappearance of the task-induced bias. However, in the CIL, the causal path $X\\gets T\\to Y$ cannot be cut by the task identifiers since they are inaccessible, or by the independence between $T$ and $Y$ like domain incremental learning. In order to eliminate the task-induced bias in CIL, we seek advanced causal techniques among which causal intervention [21] is a powerful one. It is noteworthy that our analysis indicates that a key objective of CIL is to cut the causal path $X\\gets T\\to Y$ and we can cut it not only by conditioning variable $T$ , but also by cutting $X\\gets T$ or $T\\rightarrow Y$ .", + "bbox": [ + 75, + 454, + 468, + 772 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.2. Causal debias", + "text_level": 1, + "bbox": [ + 76, + 776, + 220, + 790 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The overall causal graph of CIL is illustrated in Fig. 4 (a), which considers two facets of task-induced bias in Fig. 1. The causal effect of $X_{1}$ and $X_{2}$ on $Y$ in the original causal graph can be formulated by Bayes' rule:", + "bbox": [ + 75, + 792, + 467, + 853 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nP (Y | X _ {1}, X _ {2}) = \\sum_ {t} P (Y | \\underline {{X}} _ {1}, X _ {2}, t) P (t | \\underline {{X}} _ {1}, X _ {2}) \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 94, + 872, + 468, + 902 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/99e227d7b07a2ac02be42df53f19a6061d2893037227a7e0b519c1dacdd8a455.jpg", + "image_caption": [ + "(c) Simplified causal graph", + "(d) Causal intervention", + "(e) An implementation", + "Figure 4. Causal intervention for class incremental learning. To approximate the task identifier, we utilize a class feature dictionary $D_{t}$ recording information about tasks." + ], + "image_footnote": [], + "bbox": [ + 517, + 97, + 880, + 258 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We can treat the two variables $X_{1}$ and $X_{2}$ as one as shown in Fig. 4 (b), because they have the same context in the causal graph. As a result, we obtain a simplified causal graph in Fig. 4 (c), of which the causal effect of $X$ on $Y$ is reformulated as:", + "bbox": [ + 498, + 333, + 890, + 406 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nP (Y | X) = \\sum_ {t} P (Y | X, t) P (t | X). \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 576, + 422, + 890, + 454 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We have shown in Fig. 3 (e) that the causal path $X \\gets T \\to Y$ is the cause of task-induced bias. If we can cut this causal path, the task-induced bias will be eliminated. We opt to cut the $X \\gets T$ using the causal intervention, as illustrated in Fig. 4 (d). The causal path $X \\gets T$ is introduced through the term $P(t|X)$ in Eq. 2. Following [21], the direct causal effect of $X$ on $Y$ is denoted by $P(Y|do(X))$ , where $do(X)$ means a causal intervention on the single variable $X$ . Because the variable $T$ blocks all the backdoor path in the causal graph, we can obtain the expression of $P(Y|do(X))$ according to the backdoor criterion [21] as:", + "bbox": [ + 498, + 470, + 892, + 638 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nP (Y | d o (X)) = \\sum_ {t} P (Y | X, t) P (t) \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 573, + 654, + 890, + 684 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Compared with the old term $P(t|X)$ , the new term $P(t)$ is no longer changed with $X$ , equating to cut the causal edge $X \\gets T$ , as shown in Fig. 4 (d).", + "bbox": [ + 498, + 700, + 890, + 746 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Based on the intervened causal graph in Fig. 4 (d) in CIL, we devise a causal debias module to implement the Equation 3. Since the task identifier is not accessible in the test phase, we propose to use a feature dictionary $D_{t}$ that contains information about tasks to approximate the unobserved confounder $T$ as shown in Fig. 4 (e). The task information is retained in class labels of images for CIL, because class label sets of different tasks cannot overlap with each other and thus task identifiers can be inferred from classes. Specifically, given an image $\\mathbf{x}$ , a label $y$ and a task identifier", + "bbox": [ + 496, + 750, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "16019", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/50c829668428a825fcb6ab579a26e7a39876e9fdd5f438cfa2c4dc6c129684f8.jpg", + "image_caption": [ + "Figure 5. Illustration of causal intervention operation. This operation borrows features of classes in previous tasks, and puts them to features of new classes in the current task. The causal intervention is implied in the \"borrow\" and \"put\" operations, because the two operations mean that the features containing task information are directly controlled by ourselves." + ], + "image_footnote": [], + "bbox": [ + 86, + 95, + 460, + 220 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "$t$ , Equation 3 is implemented as", + "bbox": [ + 76, + 328, + 290, + 342 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} P (Y | d o (X)) = \\sum_ {t} P (y | \\mathbf {x}, t) P (t) \\\\ = \\sum_ {\\boldsymbol {d} _ {t}} P (y | \\boldsymbol {x}, \\boldsymbol {d} _ {t}) P (\\boldsymbol {d} _ {t}), \\tag {4} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 143, + 348, + 468, + 415 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\pmb{x}$ is the feature of the image $\\mathbf{x}$ , $d_{t}$ is sampled from $\\pmb{D}_{t}$ . If the last layer of a classifier is a softmax layer, we get $P(y|\\pmb{x},\\pmb{d}_t) = \\mathrm{Softmax}(f_y(\\pmb{x},\\pmb{d}_t))$ , where $f_{y}(\\cdot)$ is a similarity function. In a nutshell, the implementation of Equation 3 is newly formulated as:", + "bbox": [ + 76, + 421, + 468, + 496 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nP (Y | d o (X)) := \\mathbb {E} _ {\\boldsymbol {d} _ {t}} \\left[ \\operatorname {S o f t m a x} \\left(f _ {y} (\\boldsymbol {x}, \\boldsymbol {d} _ {t})\\right) \\right]. \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 129, + 503, + 468, + 520 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The above expectation can be approximated by NWGM [32] as follows:", + "bbox": [ + 75, + 527, + 468, + 556 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbb {E} _ {\\boldsymbol {d} _ {t}} \\left[ \\operatorname {S o f t m a x} \\left(f _ {y} \\left(\\boldsymbol {x}, \\boldsymbol {d} _ {t}\\right)\\right) \\right] \\approx \\operatorname {S o f t m a x} \\left(\\mathbb {E} _ {\\boldsymbol {d} _ {t}} \\left[ f _ {y} \\left(\\boldsymbol {x}, \\boldsymbol {d} _ {t}\\right) \\right]\\right). \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 91, + 564, + 468, + 594 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To implement our causal intervention operation in Fig. 5, we devise a simple attention-based mechanism with ignorable number of parameters:", + "bbox": [ + 75, + 595, + 468, + 640 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nf _ {y} (\\boldsymbol {x}, \\boldsymbol {d} _ {t}) = \\boldsymbol {W} _ {0} ^ {\\top} \\operatorname {C a t} \\left(\\boldsymbol {W} _ {1} ^ {\\top} \\boldsymbol {x}, \\boldsymbol {W} _ {2} ^ {\\top} g _ {y} (\\boldsymbol {d} _ {t})\\right), \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 127, + 645, + 468, + 665 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $W_{1}$ and $W_{2}$ both are fully connected layers and $\\operatorname{Cat}(\\cdot)$ is the concatenation operator. Due to the additive property of expectation, the $\\mathbb{E}_{\\boldsymbol{d}_t}[f_y(\\boldsymbol{x}, \\boldsymbol{d}_t)]$ can be rewritten as:", + "bbox": [ + 75, + 670, + 468, + 729 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbb {E} _ {\\boldsymbol {d} _ {t}} \\left[ f _ {y} (\\boldsymbol {x}, \\boldsymbol {d} _ {t}) \\right] = \\boldsymbol {W} _ {\\boldsymbol {0}} ^ {\\top} \\operatorname {C a t} \\left(\\boldsymbol {W} _ {\\boldsymbol {1}} ^ {\\top} \\boldsymbol {x}, \\boldsymbol {W} _ {\\boldsymbol {2}} ^ {\\top} \\mathbb {E} _ {\\boldsymbol {d} _ {t}} \\left[ g _ {y} \\left(\\boldsymbol {d} _ {t}\\right) \\right]\\right). \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 93, + 734, + 468, + 767 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We devise $\\mathbb{E}_{\\boldsymbol{d}_t}[g_y(\\boldsymbol{d}_t)]$ utilizing a scaled dot-product attention [28], which is calculated as follows:", + "bbox": [ + 75, + 768, + 468, + 797 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nQ = W _ {3} ^ {\\top} x, K = W _ {4} ^ {\\top} D _ {t}, V = D _ {t},\n$$\n", + "text_format": "latex", + "bbox": [ + 143, + 803, + 400, + 821 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {A} = \\operatorname {S o f t m a x} \\left(\\boldsymbol {Q} ^ {\\top} \\boldsymbol {K} / \\sqrt {\\sigma}\\right), \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 147, + 825, + 468, + 842 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbb {E} _ {\\boldsymbol {d} _ {t}} \\left[ g _ {y} (\\boldsymbol {d} _ {t}) \\right] = \\boldsymbol {V} ^ {\\top} \\boldsymbol {A},\n$$\n", + "text_format": "latex", + "bbox": [ + 148, + 845, + 297, + 864 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $W_{3}$ and $W_{4}$ both are fully connected layers, and $\\sigma$ is the output dimension of them.", + "bbox": [ + 75, + 869, + 468, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Algorithm 1 CafeBoost training pipeline.", + "bbox": [ + 500, + 90, + 777, + 106 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Input: Training data $\\mathcal{D}_0, \\dots, \\mathcal{D}_T$ , empty dictionary $D_0$ , empty example memory $\\mathcal{M}$ ,", + "bbox": [ + 500, + 109, + 890, + 140 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "1: Initialize parameters of feature extractor $\\psi_0$ , causal bias module $\\phi_0$ (Eq. 8, 9) and classifier $\\theta_0$", + "bbox": [ + 509, + 141, + 890, + 170 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "2: // First task training stage.", + "bbox": [ + 509, + 171, + 709, + 185 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3: Train $\\psi_0, \\theta_0$ on $\\mathcal{D}_0$ by minimizing $\\mathcal{L}_{CE}$", + "bbox": [ + 509, + 186, + 795, + 200 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4: Extract example features on $\\mathcal{D}_0$ , using $\\psi_0$", + "bbox": [ + 509, + 202, + 805, + 215 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5: Store class mean features on $\\mathcal{D}_0$ to dictionary $D_0$", + "bbox": [ + 509, + 217, + 854, + 229 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "6: // Causal debias module training stage.", + "bbox": [ + 509, + 231, + 790, + 246 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "7: Train $\\phi_0, \\theta_0$ on $\\mathcal{D}_0$ by minimizing $\\mathcal{L}_{CE}$", + "bbox": [ + 509, + 247, + 794, + 261 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "8: Extract example features on $\\mathcal{D}_0$ , using $\\psi_0$ and $\\phi_0$", + "bbox": [ + 509, + 262, + 854, + 276 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "9: if example memory $\\mathcal{M}$ is used then", + "bbox": [ + 509, + 277, + 766, + 290 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "10: Update example memory $\\mathcal{M}$", + "bbox": [ + 509, + 292, + 736, + 306 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "11: end if", + "bbox": [ + 509, + 306, + 571, + 319 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "12: // Class incremental training stage.", + "bbox": [ + 504, + 321, + 764, + 335 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "13: for $t$ in $\\{1, \\dots, T\\}$ do", + "bbox": [ + 504, + 337, + 669, + 351 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "14: Add a new classifier $\\theta_t$", + "bbox": [ + 504, + 352, + 700, + 364 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "15: Train $\\psi_t, \\phi_t$ and $\\theta_t$ on $\\mathcal{D}_t \\cup \\mathcal{M}$ (or $\\mathcal{D}_t$ ) by Eq. 10", + "bbox": [ + 504, + 367, + 877, + 382 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "16: Extract example features on $\\mathcal{D}_t$ , using $\\psi_t$ and $\\phi_t$", + "bbox": [ + 504, + 383, + 867, + 397 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "17: if example memory $\\mathcal{M}$ is used then", + "bbox": [ + 504, + 398, + 782, + 411 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "18: Update example memory $\\mathcal{M}$", + "bbox": [ + 504, + 412, + 753, + 426 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "19: Update dictionary $D_{t}$ by Eq. 11", + "bbox": [ + 504, + 428, + 772, + 441 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "20: end if", + "bbox": [ + 504, + 443, + 589, + 455 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "21: end for", + "bbox": [ + 504, + 458, + 581, + 469 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Output: Final feature extractor $\\psi_T$ , final causal debias module $\\phi_T$ , and all classifiers $\\{\\pmb{\\theta}_0, \\dots, \\pmb{\\theta}_T\\}$", + "bbox": [ + 504, + 472, + 890, + 503 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.3. Causal Feature Boost", + "text_level": 1, + "bbox": [ + 500, + 537, + 700, + 551 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this section, we describe how the causal debias module is added into the continual learning pipeline, and the training strategies to enable the overall architecture. Additionally, we propose a momentum method to update the dictionary so as to catch more task information.", + "bbox": [ + 496, + 563, + 890, + 638 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Architecture. Our causal debias module can be added between the feature extractor and the classifier of existing continual learning architectures. In Fig. 2, given an image as input, the CNN backbone (e.g. ResNet [8]) will extract a feature map. Then, we feed this feature map and a pre-collected dictionary into our causal debias module, and achieve a newly-updated feature map which is forwarded to the classifier to get a class label. After a new task comes, we add one more classifier for new classes and add no other layers.", + "bbox": [ + 496, + 641, + 890, + 792 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Training pipeline. Because the dictionary used in the causal debias module need to be filled with average features of classes, we train a feature extractor and a classifier by using data of the initial phase firstly. Then, we fix the feature extractor and run an extra phase before the incremental phases to train the newly-added causal debias module and the classifier together. In the remaining phases,", + "bbox": [ + 496, + 795, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "16020", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/3a125f140b603fdee2ee2bc374b608b40354ad8c66aceebbf1f52fc921e144e5.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodCIFAR-100ImageNet-SubImageNet-Full
Average accuracy(%)5102551025510
LwF†[16]49.5946.9845.5153.6247.6444.3244.3538.90
BiC†[31]59.3654.2050.0070.0764.9657.7362.6558.72
iCaRL [22]62.9556.9251.3065.0460.8254.5651.5046.89
LUCIR [11]63.5657.2050.2970.8468.3261.4464.4561.57
GeoDL [24]67.0860.9653.9871.2770.2464.8665.2364.46
CwD [23]66.8462.6656.2867.3664.0960.52--
CSCCT [3]66.8959.9152.3271.3367.9161.52--
iCaRL+CafeBoost69.4164.7061.6070.5367.0463.4053.4648.46
LUCIR+CafeBoost70.1066.5662.3771.5170.0367.7466.8864.77
Forgetting rate(%)5102551025510
LwF†[16]43.3643.5841.6655.3257.0055.1248.7047.94
BiC†[31]31.4232.5034.6027.0431.0437.8825.0628.34
iCaRL [22]22.9525.3025.7418.7021.3525.0926.0333.76
LUCIR [11]23.3026.4029.2131.8833.4835.4024.0827.29
GeoDL [24]16.8321.8324.918.689.0711.9711.0312.81
CwD [23]17.7026.8026.9111.1617.5325.41--
iCaRL+CafeBoost8.5112.9116.0812.2014.2318.5225.1928.24
LUCIR+CafeBoost6.999.0311.758.167.509.108.6511.73
", + "bbox": [ + 189, + 88, + 785, + 377 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1. The average incremental accuracy and the forgetting rate on CIFAR-100, ImageNet-Sub and ImageNet-Full. The number of incremental phases are 5, 10, and 25. The number of exemplars per class is 20. An ideal continual learning method should have high average incremental accuracy and low forgetting rate. Results of models with $\\dagger$ are taken from [24].", + "bbox": [ + 75, + 378, + 893, + 420 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "the causal debias module are always learnable, whereas the classifiers of old tasks are fixed. Our causal debias module can be used in iCaRL [22] and LUCIR [11] as a plugin. Our overall approach borrows their corresponding training objectives. Therefore, the training objective of our approach (CafeBoost) is formulated as:", + "bbox": [ + 75, + 435, + 470, + 525 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {\\theta} \\left[ L _ {c e} (x, y, \\theta) + \\eta L _ {*} (\\theta) \\right], \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 176, + 537, + 468, + 560 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $L_{*}(\\theta)$ denotes specific losses of the corresponding base methods, e.g., distillation loss for iCaRL, margin ranking loss and others for LUCIR. Algorithm 1 details out the CafeBoost training pipeline.", + "bbox": [ + 75, + 571, + 468, + 632 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Momentum dictionary updating. Our approach exploits an external dictionary to store mean features of each class in the initial phase. We can exploit a momentum updating strategy to renew features in the dictionary, if replay buffer is used. The new dictionary at task $t$ is updated as", + "bbox": [ + 75, + 632, + 468, + 708 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\pmb {D} _ {t} = \\alpha \\pmb {D} _ {t - 1} + (1 - \\alpha) [ \\mathbf {f} _ {1} ^ {t - 1}, \\dots , \\mathbf {f} _ {c} ^ {t - 1}, \\dots , \\mathbf {f} _ {C} ^ {t - 1} ], \\quad (1 1)\n$$\n", + "text_format": "latex", + "bbox": [ + 96, + 718, + 468, + 738 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\alpha$ controls the updating rate, $\\mathbf{f}_c$ represents a mean feature of class $c$ , and $C$ is the number of classes.", + "bbox": [ + 75, + 748, + 468, + 777 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 76, + 791, + 209, + 809 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1. Settings", + "text_level": 1, + "bbox": [ + 76, + 816, + 176, + 832 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Datasets & Protocols. We evaluate our approach on two commonly-used datasets: CIFAR-100 [15] and ImageNet [5]. CIFAR-100 contains 100 classes and each class has 500 images for training and 100 images for test, with", + "bbox": [ + 75, + 839, + 470, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "each image of size $32 \\times 32$ . ImageNet with about 1.3 million images of size $224 \\times 224$ is divided into two CIL settings: ImageNet-Sub that owns 100 classes and ImageNet-Full that owns 1000 classes. We follow the protocol used in many recent works [11,23,24]. Models are firstly trained on half classes, i.e., 50 classes for CIFAR100 and ImageNet-Sub, and 500 for ImageNet-Full. Subsequently, we start incremental learning phases with adding a fixed number of unobserved classes in each phase. We choose three different numbers of incremental phases: 5, 10, and 25.", + "bbox": [ + 496, + 435, + 893, + 585 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation Metrics. The two commonly-used metrics to evaluate performance of a CIL model are 1) the average incremental accuracy [22] $\\mathcal{A}_T$ , and 2) the average forgetting $\\mathcal{F}_T$ . The average incremental accuracy and the average forgetting after $T$ tasks are defined as", + "bbox": [ + 496, + 585, + 893, + 662 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {A} _ {T} = \\frac {1}{T} \\sum_ {t = 1} ^ {T} A _ {t}, \\mathcal {F} _ {T} = \\frac {1}{T - 1} \\sum_ {t = 2} ^ {T} F _ {t},\n$$\n", + "text_format": "latex", + "bbox": [ + 566, + 667, + 825, + 710 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $a_{i,j}$ denotes the accuracy on task $i$ after the model learning task $j$ , $A_{t} = \\frac{1}{t}\\sum_{i=1}^{t}a_{i,t}$ , and $F_{t} = \\frac{1}{t-1}\\sum_{j=2}^{t}\\max_{i}(a_{i,t}-a_{j,t})$ with $i\\in\\uparrow,\\dots,j-1$ . We report top-1 accuracy.", + "bbox": [ + 496, + 715, + 890, + 780 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation details. All methods are implemented with PyTorch [20]. We use the ResNet18 for CIFAR-100 and ImageNet following [23]. For all experiments, SGD optimizer and the batch size 128 are used. For CIFAR-100 and ImageNet, we train models for 160 epochs and 90 epochs, respectively. For CIFAR-100, our results are averaged over three runs with different orders generated by random seed 1993. For ImageNet, the first order with random seed 1993", + "bbox": [ + 496, + 780, + 893, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "16021", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/40b0fadf920b754d05c327f6107da3150598c29de32d14a4b4417b23fc602a8a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 236, + 90, + 730, + 106 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/038c345e9eac4d2e9ffac27d24f8ecbbe3df915a08537de4c4cd5793fd408a72.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 151, + 116, + 370, + 250 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/de3b6a1e3f8156fff399204403ac09a3261f3c9645b06d4e15a5d85cff4d7431.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 383, + 116, + 591, + 250 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/40b992f1e8937c5bffe9dce24273eeaf673b0c865ac47d7656eea329e1912352.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 604, + 117, + 810, + 250 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/8a414de9bd3f211a56d8c2d5c4435c695155f9c581d0fe25649d3039ae1fce0f.jpg", + "image_caption": [ + "Figure 6. Results on CIFAR-100 (5, 10, and 25 tasks), ImageNet-Sub (5, 10, and 25 tasks). The horizontal axis shows the task identifier and the vertical axis is the corresponding average incremental accuracy." + ], + "image_footnote": [], + "bbox": [ + 151, + 251, + 370, + 391 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f31257b9d713d80071cb3cdc340df6d5c45d20d56c6bbd3c442c28b03fe35300.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 383, + 252, + 591, + 391 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/aa30277a053aa74a1688d2ceaa1d7a4a14e983758e3546cd3626acd04bb35f98.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 606, + 252, + 812, + 391 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "is conventionally used. In addition, we deploy CafeBoost to two baselines: iCaRL [22] and LUCIR [11] and compare it with LwF [16], BiC [31], GeoDL [24], CwD [23] and CSCCT [3] on all datasets. GeoDL, CwD and CSCCT are implemented based on LUCIR.", + "bbox": [ + 75, + 438, + 468, + 513 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.2. Evaluations", + "text_level": 1, + "bbox": [ + 76, + 527, + 205, + 542 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Results on CIFAR-100. Tab. 1 summarizes the experimental results of CIFAR-100 in three different numbers of incremental tasks. We observe that CafeBoost improves iCaRL baseline by $6.46\\%$ , $7.78\\%$ , $10.30\\%$ for 5, 10, and 25 tasks, respectively. After adding to LUCIR baseline, we observe a performance boost: $6.54\\%$ , $9.36\\%$ , $12.08\\%$ , for 5, 10, and 25 tasks. The corresponding average incremental accuracies reach high levels of $70.10\\%$ , $66.56\\%$ , and $62.37\\%$ . Furthermore, CafeBoost achieves lower forgetting rates after applying to the baselines. It can decrease the forgetting rates of iCaRL by $14.44\\%$ , $12.39\\%$ , and $9.66\\%$ and those of LUCIR by $16.31\\%$ , $17.37\\%$ , and $17.46\\%$ , for 5, 10, and 25 tasks. The first row of Fig. 6 shows CafeBoost improves performance of baselines by a large margin, achieves higher accuracies and prevents catastrophic forgetting better than the prior methods for CIL. See Appendix for more results about the forgetting rates.", + "bbox": [ + 75, + 551, + 468, + 808 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Results on ImageNet. Tab. 1 also reports experimental results of ImageNet-Sub and ImageNet-Full. On ImageNet-Sub, CafeBoost boosts average incremental accuracies by $0.67\\% \\sim 8.84\\%$ over iCaRL and LUCIR baselines, for 5, 20, and 25 tasks, and decreases the corresponding forgetting rates by $6.57\\% \\sim 26.30\\%$ . For example, CafeBoost", + "bbox": [ + 75, + 810, + 468, + 901 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "improves average incremental accuracy of LUCIR from $61.44\\%$ to $67.74\\%$ $(+6.30\\%)$ , and reduces forgetting rate from $35.40\\%$ to $9.10\\%$ $(-26.30\\%)$ . The second row of Fig. 6 further validates effectiveness of CafeBoost for all numbers of incremental phases. Additionally, we observe in Fig. 6 CafeBoost has almost no effect on the initial phase, but brings great boost in the following incremental phases. CafeBoost also enhances performance of iCaRL and LUCIR on ImageNet-Full, as shown in Tab. 1. We observe CafeBoost improves accuracies of iCaRL by $1.96\\%$ , $1.57\\%$ , and those of LUCIR by $2.43\\%$ , $3.20\\%$ for 5 and 10 incremental phases respectively.", + "bbox": [ + 496, + 438, + 892, + 619 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/409646a02e6def168a1da2f23f83d129fa4e5c90949f9960812446bb90f2d149.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
SB1020304050
10LUCIR59.1260.2661.7761.6262.91
+Ours55.8762.6166.2168.2269.70
↑-3.25+2.35+4.44+6.60+6.79
5LUCIR54.6256.2056.6557.2956.90
+Ours49.0258.5363.3965.7365.95
↑-5.60+2.33+6.74+8.44+9.05
", + "bbox": [ + 504, + 631, + 890, + 756 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 2. Ablation study on impact of number of classes in the initial phase. $B$ denotes number of classes in the initial phase. $S$ is the number of new classes per incremental step. Experiments are conducted on CIFAR-100 with 20 exemplars per class. The reported values (\\%) are average incremental accuracies. $\\uparrow$ denotes the increment of values.", + "bbox": [ + 496, + 757, + 890, + 839 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.3. Ablation Studies", + "text_level": 1, + "bbox": [ + 500, + 847, + 663, + 861 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In this section, we conduct experiments to study the effect on our approach of 1) number of classes in initial phase,", + "bbox": [ + 498, + 869, + 890, + 901 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "16022", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/e5f20c61c589ff9e450b61de1fdf89caa9bc427bec422b831703bc066dab2bff.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
SRLUCIR+Ours↑
10554.23±0.7864.80±0.36+10.57
1059.98±0.5468.30±0.31+8.32
2063.56±0.5570.10±0.39+6.54
5547.74±0.8462.23±0.67+14.49
1053.22±1.2864.08±0.77+10.86
2057.20±1.1466.56±0.80+9.36
", + "bbox": [ + 112, + 88, + 433, + 215 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/768a14a0cdba8bc17e05bcbad1dd15fdfe8396dc45a79504ad5ad210cb859b74.jpg", + "table_caption": [ + "Table 3. Ablation study on impact of number of exemplars. $R$ represents the number of exemplars per class. $S$ is the number of new classes per incremental step. Experiments are conducted on CIFAR-100. The reported values (%) are average incremental accuracies. $\\uparrow$ denotes the increment of values." + ], + "table_footnote": [], + "table_body": "
MethodsCIFAR-100ImageNet-Sub
AccuracyForgettingAccuracyForgetting
LUCIR40.10±1.2752.35±1.3160.4725.73
+Ours59.01±0.2223.31±0.7665.1518.05
", + "bbox": [ + 81, + 301, + 464, + 366 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "2) number of exemplars in memory, 3) different classifiers, and 4) coefficient of momentum dictionary update.", + "bbox": [ + 75, + 428, + 467, + 458 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Impact of number of classes in initial phase. We investigate CafeBoost with 10, 20, 30, 40, and 50 classes of initial phase in Tab. 2. CafeBoost can enhance performance when the number of classes in initial phase is larger than 10. The performance improvement increases from negative to positive, along the number for both $S = 10$ and $S = 5$ . When $B$ equates 10, our approach has a negative influence on the baseline. This is because characteristics of classes in the initial phase are insufficient to share for all classes in the dataset. However, when $B >= 20$ , our approach boosts the baseline by a large margin, which shows that it tolerates a broad range of the number $B$ .", + "bbox": [ + 75, + 460, + 467, + 640 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Impact of number of exemplars in memory. We vary the number of exemplars per class and show its effect in Tab. 3. Both the performance of baselines and CafeBoost improve along the increase of exemplar number per class, which is consistently observed for $S = 10$ and $S = 5$ . In addition, the improvement margin decreases along the increase of $R$ . For instance, $+10.57$ for $S = 10$ and $R = 5$ , whereas $+6.54$ for $S = 10$ and $R = 20$ . It is worth noting that CafeBoost does not rely on a replay buffer. Tab. 4 shows CafeBoost outperforms remarkably the baseline on accuracy and forgetting in CIFAR-100 and ImageNet-Sub.", + "bbox": [ + 75, + 642, + 467, + 808 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Impact of different classifiers. We also conduct experiments to research behaviors of CafeBoost combined with three different classifiers. We adopt the classifiers CNN, $k$ -NME, and AME, as in [11]. CNN classifies examples by exploiting one class prototype per class. $k$ -NME uses $k$ samples of each class to represent the class, while AME uti", + "bbox": [ + 75, + 809, + 467, + 900 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/717fff577a8b71b5bf6bd3e7698077ebe90c1a0c189ed8fd22791af3be08ebfe.jpg", + "table_caption": [ + "Table 4. Comparisons of average incremental accuracies and forgetting rates with no replay data. Experiments are conducted with 5 incremental steps. 20 exemplars are stored per class." + ], + "table_footnote": [], + "table_body": "
MethodiCaRLiCaRL+oursLUCIRLUCIR+ours
CNN553.47±0.1556.53±0.2663.56±0.5570.10±0.39
1049.30±0.4353.75±0.0357.20±1.1466.56±0.80
2545.59±0.7551.18±0.7850.29±1.1462.37±1.17
k-NME562.95±0.5169.40±0.2365.42±0.6769.77±0.35
1056.91±0.6464.70±0.4059.85±1.2066.53±0.72
2551.30±0.9261.6±0.9052.07±0.4362.64±1.30
AME565.02±0.6470.61±0.2966.76±0.8470.54±0.31
1058.27±0.6366.52±0.4661.54±1.3667.77±0.71
2552.69±0.9264.10±0.8954.84±1.7865.10±1.35
", + "bbox": [ + 504, + 88, + 898, + 229 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 5. The average incremental accuracy by using different classifiers (CNN, $k$ -NME, AME) with 20 exemplars per class. Values are evaluated on CIFAR-100 with 5, 10, and 25 incremental steps.", + "bbox": [ + 500, + 229, + 890, + 271 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "izes all samples. Tab. 5 shows our approach achieves the best performance when using AME classifier. Moreover, it is noteworthy that our approach enhances performance of baselines using all three classifiers by a large margin in various numbers of tasks.", + "bbox": [ + 498, + 287, + 890, + 359 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Sensitivity of momentum coefficient. Fig. 7 shows our approach is not sensitive to the momentum coefficient of dictionary updating for all three classifiers. We set $\\alpha$ to 0.9 for other experiments of our approach.", + "bbox": [ + 498, + 362, + 890, + 422 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/e44ae4c03db8c8a32c9050f240ef0ef953fcdcc0051c70a4ca7dc6d91b445f8c.jpg", + "image_caption": [ + "Figure 7. Experiments on sensitivity of momentum coefficient. Average accuracies are evaluated on CIFAR-100 with 5 incremental steps and 20 exemplars per class." + ], + "image_footnote": [], + "bbox": [ + 513, + 439, + 725, + 564 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusions", + "text_level": 1, + "bbox": [ + 500, + 579, + 625, + 594 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To the best of our knowledge, this work is the first to consider a new type of bias - task-induced bias in a causal perspective, which is derived from special settings of continual learning. We analyzed three continual scenarios in a causal framework, and found that the task-induced bias is particularly detrimental to CIL. Based on this framework, we designed a causal intervention operation and implemented it as a causal debias module by exploiting a powerful attention mechanism. Comprehensive experiments demonstrated our plugin method can boost performance of baselines and outperform several existing strong methods. In the future, we think it is a promising direction to regard the task-induced bias as a problem in CIL and design more exquisite approaches to eliminate it by cutting the corresponding causal path.", + "bbox": [ + 496, + 604, + 890, + 829 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments. This work was supported in part by National Key R&D Program of China (2021ZD0112001) and National Natural Science Foundation of China (No.61831005, 61971095 and 62271119).", + "bbox": [ + 498, + 830, + 890, + 886 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "16023", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 174, + 106 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Ehsan Abbasnejad, Damien Teney, Amin Parvaneh, Javen Shi, and Anton van den Hengel. Counterfactual vision and language learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10041—10051, 2020. 2, 3", + "[2] Hongjoon Ahn, Jihwan Kwak, Subin Lim, Hyeonsu Bang, Hyojun Kim, and Taesup Moon. SS-IL: Separated softmax for incremental learning. In Proceedings of the IEEE/CVF Conference on Computer Vision, pages 844–853, 2021. 1", + "[3] Arjun Ashok, K. J. Joseph, and Vineeth Balasubramanian. Class-incremental learning with cross-space clustering and controlled transfer. In European Conference on Computer Vision, 2022. 1, 2, 6, 7", + "[4] Matthias Delange, Rahaf Aljundi, Marc Masana, Sarah Parisot, Xu Jia, Ales Leonardis, Greg Slabaugh, and Tinne Tuytelaars. A continual learning survey: Defying forgetting in classification tasks. Transactions on Pattern Analysis and Machine Intelligence, 44(7):3366-3385, 2021. 3", + "[5] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. ImageNet: A large-scale hierarchical image database. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 248-255, 2009. 6", + "[6] Xiang Deng and Zhongfei Zhang. Comprehensive knowledge distillation with causal intervention. In Advances in Neural Information Processing Systems, pages 22158-22170, 2021. 2, 3", + "[7] Chen He, Ruiping Wang, and Xilin Chen. A tale of two CILs: The connections between class incremental learning and class imbalanced learning, and beyond. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshop, pages 3554-3564, 2021. 1", + "[8] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2016. 5", + "[9] Lisa Anne Hendricks, Kaylee Burns, Kate Saenko, Trevor Darrell, and Anna Rohrbach. Women also snowboard: Overcoming bias in captioning models. In European Conference on Computer Vision, 2018. 2", + "[10] Geoffrey Hinton, Oriol Vinyals, and Jeff Dean. Distilling the knowledge in a neural network. In ArXiv: 1503.02531, 2015. 1", + "[11] Saihui Hou, Xinyu Pan, Chen Change Loy, Zilei Wang, and Dahua Lin. Learning a unified classifier incrementally via rebalancing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 831-839, 2019. 1, 2, 6, 7, 8", + "[12] Xinting Hu, Kaihua Tang, Chunyan Miao, Xian-Sheng Hua, and Hanwang Zhang. Distilling causal effect of data in class-incremental learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3956-3965, 2021. 1, 3", + "[13] K. J. Joseph, Salman Khan, Fahad Shahbaz Khan, Rao Muhammad Anwer, and Vineeth N. Balasubramanian." + ], + "bbox": [ + 78, + 114, + 470, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Energy-based latent aligner for incremental learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7452-7461, 2022. 1", + "[14] Minsoo Kang, Jaeyoo Park, and Bohyung Han. Class incremental learning by knowledge distillation with adaptive feature consolidation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16071-16080, 2022. 1", + "[15] Alex Kriszhevsky. Learning multiple layers of features from tiny images. In Technical report, 2009. 6", + "[16] Zhizhong Li and Derek Hoiem. Learning without forgetting. Transactions on Pattern Analysis and Machine Intelligence, 40(12):2935-2947, 2018. 1, 6, 7", + "[17] Chengzhi Mao, Kevin Xia, James Wang, Hao Wang, Junfeng Yang, Elias Bareinboim, and Carl Vondrick. Causal transportability for visual recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7521-7531, 2022. 3", + "[18] Michael McCloskey and Neal J. Cohen. Catastrophic interference in connectionist networks: The sequential learning problem. In *Psychology of Learning and Motivation*, volume 24, pages 109–165. Academic Press, 1989. 1", + "[19] Ninareh Mehrabi, Fred Morstatter, Nripsuta Saxena, Kristina Lerman, and Aram Galstyan. A survey on bias and fairness in machine learning. ACM Computing Surveys, 54(6):1-35, 2021. 2", + "[20] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, Alban Desmaison, Andreas Köpf, Edward Yang, Zach DeVito, Martin Raison, Alykhan Tejani, Sasank Chilamkurthy, Benoit Steiner, Lu Fang, Junjie Bai, and Soumith Chintala. PyTorch: An imperative style, high-performance deep learning library. In Advances in Neural Information Processing Systems, 2019. 6", + "[21] Judea Pearl, Madelyn Glymour, and Nicholas P. Jewell. Causal Inference In Statistics: A primer. Wiley, 2016. 3, 4", + "[22] Sylvestre Alvise Rebuffi, Alexander Kolesnikov, Georg Sperl, and Christoph H. Lampert. iCaRL: Incremental classifier and representation learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5533-5542, 2017. 1, 2, 6, 7", + "[23] Yujun Shi, Kuangqi Zhou, Jian Liang, Zihang Jiang, Jiashi Feng, Philip H. S. Torr, Song Bai, and Vincent Y. F. Tan. Mimicking the oracle: An initial phase decorrelation approach for class incremental learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16722-16731, 2022. 2, 6, 7", + "[24] Christian Simon, Piotr Koniusz, and Mehrtash Harandi. On learning the geodesic path for incremental learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1591-1600, 2021. 2, 6, 7", + "[25] Qing Sun, Fan Lyu, Fanhua Shang, Wei Feng, and Liang Wan. Exploring example influence in continual learning. In Advances in Neural Information Processing Systems, 2022. 2" + ], + "bbox": [ + 503, + 92, + 890, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "16024", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[26] Rishabh Tiwari, Krishnateja Killamsetty, Rishabh Iyer, and Pradeep Shenoy. GCR: Gradient coreset based replay buffer selection for continual learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 99-108, 2022. 2", + "[27] Gido M van de Ven and Andreas S Tolias. Three scenarios for continual learning. In arXiv: 1904.07734, 2019. 1, 2, 3", + "[28] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Advances in Neural Information Processing Systems, page 5999-6009, 2017. 2, 5", + "[29] Tan Wang, Jianqiang Huang, Hanwang Zhang, and Qianru Sun. Visual commonsense R-CNN. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2020. 2", + "[30] Max Welling. Herding dynamical weights to learn. In International Conference on Machine Learning, 2009. 2", + "[31] Yue Wu, Yinpeng Chen, Lijuan Wang, Yuancheng Ye, Zicheng Liu, Yandong Guo, and Yun Fu. Large scale incremental learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 374-382, 2019. 1, 2, 6, 7", + "[32] Kelvin Xu, Jimmy Lei Ba, Ryan Kiros, Kyunghyun Cho, Aaron Courville, Ruslan Salakhutdinov, Richard S. Zemel, and Yoshua Bengio. Show, attend and tell: Neural image caption generation with visual attention. In International Conference on Machine Learning, pages 2048-2057, 2015. 5", + "[33] Shipeng Yan, Jiangwei Xie, and Xuming He. DER: Dynamically expandable representation for class incremental learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3013-3022, 2021. 1", + "[34] Xu Yang, Hanwang Zhang, and Jianfei Cai. Deconfounded image captioning: A causal retrospect. Transactions on Pattern Analysis and Machine Intelligence, 2021. 3", + "[35] Bowen Zhao, Xi Xiao, Guojun Gan, Bin Zhang, and Shu Tao Xia. Maintaining discrimination and fairness in class incremental learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13205-13214, 2020. 1, 2" + ], + "bbox": [ + 78, + 90, + 468, + 683 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "16025", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2023/CafeBoost_ Causal Feature Boost To Eliminate Task-Induced Bias for Class Incremental Learning/6f2564b9-d9c3-4c4f-b871-b0003952027b_model.json b/2023/CafeBoost_ Causal Feature Boost To Eliminate Task-Induced Bias for Class Incremental Learning/6f2564b9-d9c3-4c4f-b871-b0003952027b_model.json new file mode 100644 index 0000000000000000000000000000000000000000..cb5502670d8167a5802dc07a28600d1b5c4fe4d4 --- /dev/null +++ b/2023/CafeBoost_ Causal Feature Boost To Eliminate Task-Induced Bias for Class Incremental Learning/6f2564b9-d9c3-4c4f-b871-b0003952027b_model.json @@ -0,0 +1,2453 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.11, + 0.13, + 0.862, + 0.177 + ], + "angle": 0, + "content": "CafeBoost: Causal Feature Boost to Eliminate Task-Induced Bias for Class Incremental Learning" + }, + { + "type": "text", + "bbox": [ + 0.233, + 0.204, + 0.326, + 0.222 + ], + "angle": 0, + "content": "Benliu Qiu" + }, + { + "type": "text", + "bbox": [ + 0.366, + 0.204, + 0.481, + 0.222 + ], + "angle": 0, + "content": "Hongliang Li*" + }, + { + "type": "text", + "bbox": [ + 0.51, + 0.204, + 0.605, + 0.221 + ], + "angle": 0, + "content": "Haitao Wen" + }, + { + "type": "text", + "bbox": [ + 0.641, + 0.204, + 0.735, + 0.222 + ], + "angle": 0, + "content": "Heqian Qiu" + }, + { + "type": "text", + "bbox": [ + 0.235, + 0.223, + 0.359, + 0.239 + ], + "angle": 0, + "content": "Lanxiao Wang*" + }, + { + "type": "text", + "bbox": [ + 0.388, + 0.223, + 0.505, + 0.24 + ], + "angle": 0, + "content": "Fanman Meng" + }, + { + "type": "text", + "bbox": [ + 0.54, + 0.223, + 0.635, + 0.239 + ], + "angle": 0, + "content": "Qingbo Wu" + }, + { + "type": "text", + "bbox": [ + 0.67, + 0.223, + 0.735, + 0.238 + ], + "angle": 0, + "content": "Lili Pan" + }, + { + "type": "text", + "bbox": [ + 0.187, + 0.24, + 0.784, + 0.257 + ], + "angle": 0, + "content": "University of Electronic Science and Technology of China, Chengdu, China" + }, + { + "type": "text", + "bbox": [ + 0.228, + 0.26, + 0.741, + 0.275 + ], + "angle": 0, + "content": "{qbenliu, haitaowen, hqqiu, lanxiao.wang}@std.uestc.edu.cn" + }, + { + "type": "text", + "bbox": [ + 0.23, + 0.278, + 0.737, + 0.293 + ], + "angle": 0, + "content": "{h1li, fmmeng, qbwu}@uestc.edu.cn panlili8255@gmail.com" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.327, + 0.314, + 0.343 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.348, + 0.474, + 0.695 + ], + "angle": 0, + "content": "Continual learning requires a model to incrementally learn a sequence of tasks and aims to predict well on all the learned tasks so far, which notoriously suffers from the catastrophic forgetting problem. In this paper, we find a new type of bias appearing in continual learning, coined as task-induced bias. We place continual learning into a causal framework, based on which we find the task-induced bias is reduced naturally by two underlying mechanisms in task and domain incremental learning. However, these mechanisms do not exist in class incremental learning (CIL), in which each task contains a unique subset of classes. To eliminate the task-induced bias in CIL, we devise a causal intervention operation so as to cut off the causal path that causes the task-induced bias, and then implement it as a causal debias module that transforms biased features into unbiased ones. In addition, we propose a training pipeline to incorporate the novel module into existing methods and jointly optimize the entire architecture. Our overall approach does not rely on data replay, and is simple and convenient to plug into existing methods. Extensive empirical study on CIFAR-100 and ImageNet shows that our approach can improve accuracy and reduce forgetting of well-established methods by a large margin." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.71, + 0.21, + 0.725 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.728, + 0.47, + 0.879 + ], + "angle": 0, + "content": "Deep learning has been widely applied in many fields like computer vision, social media analytics, natural language processing, etc. The standard paradigm of deep learning is to train models on a prepared dataset, of which all data are accessible in the whole training stage. However, in most real-world applications, a sequence of tasks arrives incrementally, which requires models to learn continuously from a new task, namely continual learning [27]. Despite fine-tuning achieves this goal somewhat, such naive method forgets obtained knowledge of previous tasks after learning" + }, + { + "type": "image", + "bbox": [ + 0.513, + 0.322, + 0.866, + 0.415 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.635, + 0.411, + 0.736, + 0.422 + ], + "angle": 0, + "content": "(a) Illustration of CIL" + }, + { + "type": "image", + "bbox": [ + 0.515, + 0.426, + 0.88, + 0.49 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.58, + 0.491, + 0.802, + 0.5 + ], + "angle": 0, + "content": "(b) Task-induced bias between class features" + }, + { + "type": "image", + "bbox": [ + 0.51, + 0.501, + 0.868, + 0.571 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.574, + 0.572, + 0.816, + 0.583 + ], + "angle": 0, + "content": "(c) Task-induced bias between images and labels" + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.584, + 0.892, + 0.695 + ], + "angle": 0, + "content": "Figure 1. Illustration of CIL, task-induced bias and their detrimental effects. (a) Illustration of CIL. (b) Task-induced bias between class features. Two classes in a task are correlated by a task identifier. (c) Task-induced bias between images and labels. Images and labels in a task are correlated by a task identifier. In (b)(c), the first column describes perspectives of task-induced bias, the second shows the underlying causal effect causing task-induced bias, and the last illustrates adverse influence of task-induced bias." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.705, + 0.836, + 0.719 + ], + "angle": 0, + "content": "a new task, dubbed as catastrophic forgetting [18]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.893, + 0.901 + ], + "angle": 0, + "content": "Much effort has been devoted as a remedy for alleviating this problem in literature [2, 3, 11-14, 16, 33]. One general direction is to preserve knowledge representations of previous tasks by knowledge distillation [10] and by storing a few observed exemplars for replay [22]. [11] discovers the class imbalance between old classes and new ones is a key challenge of large-scale applications in class incremental learning (CIL), in which classes of a new task never appear in the previous tasks, as illustrated in Fig. 1 (a). [2,7,31,35] all follow this direction and propose effective methods to overcome the class imbalance. However, these methods ignore bias problems in continual learning. In this work, we" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.887, + 0.228, + 0.9 + ], + "angle": 0, + "content": "*Corresponding authors." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "16016" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.47, + 0.137 + ], + "angle": 0, + "content": "consider the problem of CIL in a novel perspective of bias. We think that catastrophic forgetting is the reflection of a new-type bias, coined as task-induced bias." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.138, + 0.471, + 0.5 + ], + "angle": 0, + "content": "Deep neural networks are easily misled by language bias [1] and vision bias [6, 9, 29], which are both derived from spurious correlations of two or more items in different positions of a sentence or an image. In contrast to these two bias, the task-induced bias occurs across different images in a common task. In CIL, data of specific classes are accessible only in a certain task if replay buffer is not applied. A deep neural network is capable of learning correlations among images of different categories or between images and labels as possible as it can, which includes a few spurious correlations, namely biases, leading to a limited generalization. On the one hand, a portion of these image correlations can be shared beneficially in this task, but these benefits cannot share among different tasks. For example, in Fig. 1 (b), color is sufficient to discriminate elephants and leopards in a specific task since elephants are grey while leopards are orange, and thus deep models have no motivation to capture more complex characteristic, like stripe. This is detrimental to the feature extractor because it captures a few details that cannot generalize among all tasks. On the other hand, as shown in Fig. 1 (c), a classifier will be biased to new classes in the current task [11, 31, 35], because the classes in the training data are all the new classes and to predict new classes is more possible to be correct." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.501, + 0.47, + 0.848 + ], + "angle": 0, + "content": "To better understand the task-induced bias in continual learning, we model three continual learning scenarios [27] into a common causal inference framework, which merely contains three necessary variables: image, label, and task identifier. Thanks to the framework, we discover that the task-induced bias is minor in task incremental and domain incremental learning due to innate mechanisms that cut off an inappropriate causal path, which partially explain why the two scenarios are easier to tackle than CIL [27]. For the existence of the causal path in CIL, we seek an active causal intervention to close the path. To implement the intervention, we devise a simple but effective causal debias module, which is established on attention mechanisms [28]. Subsequently, we incorporate this module into existing CIL architectures and propose a training pipeline to optimize it. Our overall approach transforms causally biased features into unbiased ones, and hence we refer to it as Casual Feature Boost (CafeBoost). Finally, we systematically compare accuracy and forgetting of well-established methods with ours in various settings. Experimental results show that CafeBoost yields significant and consistent performance boost for existing methods in CIFAR-100 and ImageNet by a large margin \\((0.67\\% \\sim 12.08\\%\\) on accuracy)." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.852, + 0.22, + 0.868 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.871, + 0.47, + 0.901 + ], + "angle": 0, + "content": "Continual learning. Existing CIL works mostly aim to maintain knowledge representations of previous tasks, in-" + }, + { + "type": "image", + "bbox": [ + 0.515, + 0.097, + 0.88, + 0.306 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.309, + 0.892, + 0.392 + ], + "angle": 0, + "content": "Figure 2. Overview of Casual Feature Boost (CafeBoost). Green blocks are learnable, whereas white blocks are fixed. Conv block is updated with a smaller learning rate than Causal block. The training objectives including distillation loss, margin ranking loss, etc., are omitted in this figure, since they depend on baselines. Whether to use replay buffer \\(\\mathcal{M}\\) depends on specific baselines." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.402, + 0.893, + 0.764 + ], + "angle": 0, + "content": "including replay-based methods and distillation-based methods. Replay-based methods [22, 25, 26] store a few examples to represent previous data and replay the stored data in the current training stage. The purpose of relay-based methods is to select the most representative samples of the entire training dataset. iCaRL [22] uses herding strategy [30] to choose typical exemplars for future training. [25] calculates example influence of stability and plasticity for every example in validation sets, based on which it selects examples for replay. Distillation-based methods [3,11,23,24] focus on transferring learned knowledge of a previous model to the current one. LUCIR [11] combines cosine normalization, less-forget constraint, and inter-class separation to mitigate the forgetting caused by class imbalance. GeoDL [24] utilizes geodesic path for knowledge distillation which considers the gradual change between subspaces of tasks. CSCCT [3] encourages features of each class to locate in an appropriate position of the feature space, and utilizes inter-class relationships to transfer inter-class knowledge positively or negatively. CwD [23] scatters uniformly representations of each class in the initial task, which significantly benefits performance in subsequent tasks. These two types of methods are usually incorporated together to achieve better performance." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Spurious correlation. Spurious correlation is also known as bias. [19] has extensively investigated the bias problem in machine learning. Many methods [1,9,29] have been proposed to eliminate bias in their specific applications. [1] introduces counterfactual in the training to eliminate bias in visual question answering datasets. [9] proposes a new Equalizer model to force a model look at a person instead of inferring a person's gender by context cues. [29] applies causal intervention to alleviate vision bias" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16017" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.32, + 0.106 + ], + "angle": 0, + "content": "among detected objects in an image." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.109, + 0.471, + 0.29 + ], + "angle": 0, + "content": "Causal inference. Causal inference [21] aims to establish reasonable causal graphs for specific applications, then analyze problems and propose solutions based on the established causal graph. Causal intervention and counterfactual process are two powerful tools of causal inference. These two tools have been ubiquitously applied in multiple applications, such as out-of-distribution recognition [17], visual question answering [1], image captioning [34], knowledge distillation [6], etc. Recently, [12] builds a causal view to analyze the cause of forgetting in class incremental learning, and designs an anti-forgetting technique so as to distill causal effect of data without replay." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.302, + 0.204, + 0.319 + ], + "angle": 0, + "content": "3. Preliminary" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.328, + 0.334, + 0.344 + ], + "angle": 0, + "content": "3.1. Continual learning scenarios" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.354, + 0.47, + 0.549 + ], + "angle": 0, + "content": "Continual learning can be categorized into three scenarios [27] in the light of whether a task identifier is given in a test phase and whether a model is required to infer which task it meets. If the task identifier is available during both training and test phases, this scenario is task incremental. If the task identifier is accessible in training phase but it is only necessary for a model to solve the task at hand, this scenario is domain incremental. Similar to the domain incremental scenario, the class incremental scenario doesn't provide task identifiers in test phase, but it requires a model to infer the current task identifier. Generally speaking, the class incremental scenario is the most difficult among the three, whereas the task incremental scenario is the easiest [27]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.552, + 0.47, + 0.734 + ], + "angle": 0, + "content": "To formulate continual learning, we assume a sequence of \\(T\\) tasks is given as \\(\\{\\mathcal{T}_0,\\mathcal{T}_1,\\dots,\\mathcal{T}_{T - 1}\\}\\), each with data \\(\\mathcal{D}_t,t = 0,1,\\ldots ,T - 1\\). Every sample in \\(\\mathcal{D}_t\\) is a tuple of an image \\(\\mathbf{x}_i\\) and a corresponding label \\(y_{i}\\), and hence we can denote the data of task \\(t\\) as \\(\\mathcal{D}_t = \\{(\\mathbf{x}_i,y_i)_{i = 1}^{N_t}\\}\\) where \\(N_{t}\\) is the number of samples, \\(\\mathbf{x}_i\\in \\mathcal{X}_t\\) and \\(y_{i}\\in \\mathcal{Y}_{t}\\). Replay-based continual learning methods require a memory buffer \\(\\mathcal{M}\\) to store partial data of previous tasks. For the task \\(\\mathcal{T}_0\\), only data \\(\\mathcal{D}_0\\) can be used for training, but for the task \\(\\mathcal{T}_t,t > 0\\), we can train models using data \\(\\mathcal{M}\\cup \\mathcal{D}_t\\). For the task incremental scenario, task identifier \\(\\mathcal{T}\\) is given as an input to models in test phase but not for the other two scenarios." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.735, + 0.471, + 0.902 + ], + "angle": 0, + "content": "After the above formulation, we elucidate distinctions of three incremental scenarios in a formal style [4]. Let \\( P(\\cdot) \\) denote a marginal distribution. The common ground of three scenarios is \\( P(\\mathcal{X}_t) \\neq P(\\mathcal{X}_{t + 1}) \\) for all \\( t \\). Their distinctions lie in label set \\( \\mathcal{Y}_t \\) and its marginal distribution \\( P(\\mathcal{Y}_t) \\). For task incremental learning, \\( \\mathcal{Y}_t \\neq \\mathcal{Y}_{t + 1} \\). For domain incremental learning, \\( \\mathcal{Y}_t = \\mathcal{Y}_{t + 1} \\) and \\( P(\\mathcal{Y}_t) = P(\\mathcal{Y}_{t + 1}) \\). These two scenarios have no requirement on inferring task identifiers. Conversely, the class incremental learning that defines \\( \\mathcal{Y}_t \\subset \\mathcal{Y}_{t + 1} \\) and \\( P(\\mathcal{Y}_t) \\neq P(\\mathcal{Y}_{t + 1}) \\), requires models to infer which label set is used for an arbitrary task." + }, + { + "type": "image", + "bbox": [ + 0.543, + 0.089, + 0.64, + 0.163 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.652, + 0.092, + 0.744, + 0.162 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.76, + 0.092, + 0.854, + 0.162 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.542, + 0.165, + 0.638, + 0.234 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.651, + 0.165, + 0.749, + 0.233 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.759, + 0.165, + 0.857, + 0.233 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.236, + 0.892, + 0.389 + ], + "angle": 0, + "content": "Figure 3. Causal graphs of supervised learning and continual learning. Gray nodes means unconditioned variables. (a) Causal graphs of supervised learning in training and test phase. (b) Causal graphs of continual learning. The red task variable depends on specific continual scenarios. (c) Causal path between images and labels in task incremental learning. The path \\( X \\gets T \\to Y \\) is impeded because variable \\( T \\) is conditioned. (d) Causal path between images and labels in domain incremental learning. The path \\( X \\gets T \\to Y \\) is cut off because of \\( T \\nrightarrow Y \\). (e) Causal path between images and labels in class incremental learning. The path \\( X \\gets T \\to Y \\) causes task-induced bias." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.398, + 0.81, + 0.414 + ], + "angle": 0, + "content": "3.2. Causal graphs in continual learning" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.417, + 0.892, + 0.522 + ], + "angle": 0, + "content": "A causal graph is a directed acyclic graph, of which nodes denote variables and directed edges denote the causality between two variables. We treat images, labels and task identifiers as nodes in our causal graphs. Although only three nodes are incorporated into our causal graph, it is powerful to reveal underlying characteristics of continual learning as we will show in Sec. 4." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.523, + 0.892, + 0.598 + ], + "angle": 0, + "content": "We denote an image as \\( X \\), a label as \\( Y \\) and a task identifier as \\( T \\) in Fig. 3. The edges are listed below: \\( X \\to Y \\) means label \\( Y \\) is deduced from image \\( X \\); \\( T \\to X \\) denotes that task identifier \\( T \\) have a causal influence to image \\( X \\); \\( T \\to Y \\) denotes task identifier \\( T \\) causally affect label \\( Y \\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.599, + 0.893, + 0.84 + ], + "angle": 0, + "content": "For supervised learning in Fig. 3 (a), its causal graph only contains an image \\( X \\), a label \\( Y \\) and an causal path \\( X \\to Y \\) that represents a model to capture the causality between two variables. In training phase of supervised learning, both image \\( X \\) and label \\( Y \\) are accessible and we utilize them to learn the causal effect into the model. But in test phase, an image \\( X \\) and the causal path are known and \\( Y \\) is the label to be inferred. For three continual learning scenarios, images, labels and task identifiers are all given in training phase. As a result, they have the same causal graph for training in Fig. 3 (b). We illustrate their distinctions in test phase in Fig. 3 (b). Due to the involvement of the task identifier \\( T \\), \\( X \\) may have an indirect effect on \\( Y \\) through the causal path \\( X \\gets T \\to Y \\), solely dependent on \\( T \\). We will explain it in Sec. 4. See Appendix for additional explanation of causal graphs." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.846, + 0.636, + 0.862 + ], + "angle": 0, + "content": "4. Methodology" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.892, + 0.901 + ], + "angle": 0, + "content": "In this section, we introduce the task-induced bias through the causal framework, a causal intervention for" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16018" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.121 + ], + "angle": 0, + "content": "CIL, its corresponding implementation, and the entire pipeline of our approach." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.123, + 0.468, + 0.138 + ], + "angle": 0, + "content": "4.1. Task-induced bias in three continual scenarios" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.139, + 0.468, + 0.213 + ], + "angle": 0, + "content": "According to the causal graphs, we can explain why task incremental learning is the most effortless scenario among the three and why class incremental learning is the most challenging. We think that this phenomenon is caused by task-induced bias." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.214, + 0.468, + 0.453 + ], + "angle": 0, + "content": "In Fig. 3 (b), the causal path \\( X \\leftarrow T \\rightarrow Y \\) is of a fork structure, in which two kid nodes are independent conditioned on the parent. [21] points out that if \\( T \\) is unconditioned, the causal influence of \\( X \\) to \\( Y \\) exists, causing the task-induced bias. We give a simple example to help understand this statement. A causal graph lighter \\(\\leftarrow\\) smoking \\(\\rightarrow\\) cancer means that a smoking person has a possibility to bring a lighter and get a lung cancer. Whether a person brings a lighter is usually independent of getting a lung cancer. Nevertheless, a smoker is highly possible to bring a lighter and to get a lung cancer, which leads to observe a correlation between lighter and cancer. If we can control the confounder smoking and collect data about non-smoking persons who usually bring lighters, then we can draw a correct conclusion that lighter is independent on cancer." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.455, + 0.47, + 0.773 + ], + "angle": 0, + "content": "Fortunately, mechanisms to eliminate the task-induced bias naturally exist under setups of task incremental learning and domain incremental learning. In task incremental learning shown in Fig. 3 (c), the value of parent node (task identifier) is given, hence the indirected effect of \\( X \\) on \\( Y \\) is 0 and the task-induced bias between \\( X \\) and \\( Y \\) disappears due to the characteristic of the fork structure. In domain incremental learning, the label set remains the same along task identifiers, which is implied in \\( \\mathcal{Y}_t = \\mathcal{Y}_{t + 1} \\) and \\( P(\\mathcal{Y}_t) = P(\\mathcal{Y}_{t + 1}) \\). For this reason, the causal edge \\( T\\to Y \\) is cut as shown in Fig. 3 (d), resulting in disappearance of the task-induced bias. However, in the CIL, the causal path \\( X\\gets T\\to Y \\) cannot be cut by the task identifiers since they are inaccessible, or by the independence between \\( T \\) and \\( Y \\) like domain incremental learning. In order to eliminate the task-induced bias in CIL, we seek advanced causal techniques among which causal intervention [21] is a powerful one. It is noteworthy that our analysis indicates that a key objective of CIL is to cut the causal path \\( X\\gets T\\to Y \\) and we can cut it not only by conditioning variable \\( T \\), but also by cutting \\( X\\gets T \\) or \\( T\\rightarrow Y \\)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.777, + 0.221, + 0.791 + ], + "angle": 0, + "content": "4.2. Causal debias" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.793, + 0.468, + 0.854 + ], + "angle": 0, + "content": "The overall causal graph of CIL is illustrated in Fig. 4 (a), which considers two facets of task-induced bias in Fig. 1. The causal effect of \\( X_{1} \\) and \\( X_{2} \\) on \\( Y \\) in the original causal graph can be formulated by Bayes' rule:" + }, + { + "type": "equation", + "bbox": [ + 0.096, + 0.873, + 0.469, + 0.904 + ], + "angle": 0, + "content": "\\[\nP (Y | X _ {1}, X _ {2}) = \\sum_ {t} P (Y | \\underline {{X}} _ {1}, X _ {2}, t) P (t | \\underline {{X}} _ {1}, X _ {2}) \\tag {1}\n\\]" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.098, + 0.882, + 0.26 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.516, + 0.26, + 0.641, + 0.27 + ], + "angle": 0, + "content": "(c) Simplified causal graph" + }, + { + "type": "image_caption", + "bbox": [ + 0.651, + 0.26, + 0.756, + 0.27 + ], + "angle": 0, + "content": "(d) Causal intervention" + }, + { + "type": "image_caption", + "bbox": [ + 0.772, + 0.26, + 0.877, + 0.27 + ], + "angle": 0, + "content": "(e) An implementation" + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.271, + 0.892, + 0.313 + ], + "angle": 0, + "content": "Figure 4. Causal intervention for class incremental learning. To approximate the task identifier, we utilize a class feature dictionary \\( D_{t} \\) recording information about tasks." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.334, + 0.892, + 0.407 + ], + "angle": 0, + "content": "We can treat the two variables \\( X_{1} \\) and \\( X_{2} \\) as one as shown in Fig. 4 (b), because they have the same context in the causal graph. As a result, we obtain a simplified causal graph in Fig. 4 (c), of which the causal effect of \\( X \\) on \\( Y \\) is reformulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.578, + 0.423, + 0.892, + 0.455 + ], + "angle": 0, + "content": "\\[\nP (Y | X) = \\sum_ {t} P (Y | X, t) P (t | X). \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.472, + 0.893, + 0.639 + ], + "angle": 0, + "content": "We have shown in Fig. 3 (e) that the causal path \\( X \\gets T \\to Y \\) is the cause of task-induced bias. If we can cut this causal path, the task-induced bias will be eliminated. We opt to cut the \\( X \\gets T \\) using the causal intervention, as illustrated in Fig. 4 (d). The causal path \\( X \\gets T \\) is introduced through the term \\( P(t|X) \\) in Eq. 2. Following [21], the direct causal effect of \\( X \\) on \\( Y \\) is denoted by \\( P(Y|do(X)) \\), where \\( do(X) \\) means a causal intervention on the single variable \\( X \\). Because the variable \\( T \\) blocks all the backdoor path in the causal graph, we can obtain the expression of \\( P(Y|do(X)) \\) according to the backdoor criterion [21] as:" + }, + { + "type": "equation", + "bbox": [ + 0.575, + 0.655, + 0.891, + 0.685 + ], + "angle": 0, + "content": "\\[\nP (Y | d o (X)) = \\sum_ {t} P (Y | X, t) P (t) \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.702, + 0.891, + 0.747 + ], + "angle": 0, + "content": "Compared with the old term \\( P(t|X) \\), the new term \\( P(t) \\) is no longer changed with \\( X \\), equating to cut the causal edge \\( X \\gets T \\), as shown in Fig. 4 (d)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Based on the intervened causal graph in Fig. 4 (d) in CIL, we devise a causal debias module to implement the Equation 3. Since the task identifier is not accessible in the test phase, we propose to use a feature dictionary \\( D_{t} \\) that contains information about tasks to approximate the unobserved confounder \\( T \\) as shown in Fig. 4 (e). The task information is retained in class labels of images for CIL, because class label sets of different tasks cannot overlap with each other and thus task identifiers can be inferred from classes. Specifically, given an image \\( \\mathbf{x} \\), a label \\( y \\) and a task identifier" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16019" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.088, + 0.096, + 0.462, + 0.221 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.223, + 0.471, + 0.306 + ], + "angle": 0, + "content": "Figure 5. Illustration of causal intervention operation. This operation borrows features of classes in previous tasks, and puts them to features of new classes in the current task. The causal intervention is implied in the \"borrow\" and \"put\" operations, because the two operations mean that the features containing task information are directly controlled by ourselves." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.329, + 0.291, + 0.343 + ], + "angle": 0, + "content": "\\(t\\), Equation 3 is implemented as" + }, + { + "type": "equation", + "bbox": [ + 0.144, + 0.349, + 0.469, + 0.416 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} P (Y | d o (X)) = \\sum_ {t} P (y | \\mathbf {x}, t) P (t) \\\\ = \\sum_ {\\boldsymbol {d} _ {t}} P (y | \\boldsymbol {x}, \\boldsymbol {d} _ {t}) P (\\boldsymbol {d} _ {t}), \\tag {4} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.422, + 0.47, + 0.497 + ], + "angle": 0, + "content": "where \\(\\pmb{x}\\) is the feature of the image \\(\\mathbf{x}\\), \\(d_{t}\\) is sampled from \\(\\pmb{D}_{t}\\). If the last layer of a classifier is a softmax layer, we get \\(P(y|\\pmb{x},\\pmb{d}_t) = \\mathrm{Softmax}(f_y(\\pmb{x},\\pmb{d}_t))\\), where \\(f_{y}(\\cdot)\\) is a similarity function. In a nutshell, the implementation of Equation 3 is newly formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.13, + 0.504, + 0.469, + 0.521 + ], + "angle": 0, + "content": "\\[\nP (Y | d o (X)) := \\mathbb {E} _ {\\boldsymbol {d} _ {t}} \\left[ \\operatorname {S o f t m a x} \\left(f _ {y} (\\boldsymbol {x}, \\boldsymbol {d} _ {t})\\right) \\right]. \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.528, + 0.469, + 0.557 + ], + "angle": 0, + "content": "The above expectation can be approximated by NWGM [32] as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.093, + 0.565, + 0.469, + 0.595 + ], + "angle": 0, + "content": "\\[\n\\mathbb {E} _ {\\boldsymbol {d} _ {t}} \\left[ \\operatorname {S o f t m a x} \\left(f _ {y} \\left(\\boldsymbol {x}, \\boldsymbol {d} _ {t}\\right)\\right) \\right] \\approx \\operatorname {S o f t m a x} \\left(\\mathbb {E} _ {\\boldsymbol {d} _ {t}} \\left[ f _ {y} \\left(\\boldsymbol {x}, \\boldsymbol {d} _ {t}\\right) \\right]\\right). \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.596, + 0.469, + 0.641 + ], + "angle": 0, + "content": "To implement our causal intervention operation in Fig. 5, we devise a simple attention-based mechanism with ignorable number of parameters:" + }, + { + "type": "equation", + "bbox": [ + 0.128, + 0.646, + 0.469, + 0.666 + ], + "angle": 0, + "content": "\\[\nf _ {y} (\\boldsymbol {x}, \\boldsymbol {d} _ {t}) = \\boldsymbol {W} _ {0} ^ {\\top} \\operatorname {C a t} \\left(\\boldsymbol {W} _ {1} ^ {\\top} \\boldsymbol {x}, \\boldsymbol {W} _ {2} ^ {\\top} g _ {y} (\\boldsymbol {d} _ {t})\\right), \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.671, + 0.469, + 0.731 + ], + "angle": 0, + "content": "where \\(W_{1}\\) and \\(W_{2}\\) both are fully connected layers and \\(\\operatorname{Cat}(\\cdot)\\) is the concatenation operator. Due to the additive property of expectation, the \\(\\mathbb{E}_{\\boldsymbol{d}_t}[f_y(\\boldsymbol{x}, \\boldsymbol{d}_t)]\\) can be rewritten as:" + }, + { + "type": "equation", + "bbox": [ + 0.094, + 0.736, + 0.469, + 0.768 + ], + "angle": 0, + "content": "\\[\n\\mathbb {E} _ {\\boldsymbol {d} _ {t}} \\left[ f _ {y} (\\boldsymbol {x}, \\boldsymbol {d} _ {t}) \\right] = \\boldsymbol {W} _ {\\boldsymbol {0}} ^ {\\top} \\operatorname {C a t} \\left(\\boldsymbol {W} _ {\\boldsymbol {1}} ^ {\\top} \\boldsymbol {x}, \\boldsymbol {W} _ {\\boldsymbol {2}} ^ {\\top} \\mathbb {E} _ {\\boldsymbol {d} _ {t}} \\left[ g _ {y} \\left(\\boldsymbol {d} _ {t}\\right) \\right]\\right). \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.769, + 0.469, + 0.799 + ], + "angle": 0, + "content": "We devise \\(\\mathbb{E}_{\\boldsymbol{d}_t}[g_y(\\boldsymbol{d}_t)]\\) utilizing a scaled dot-product attention [28], which is calculated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.144, + 0.804, + 0.401, + 0.823 + ], + "angle": 0, + "content": "\\[\nQ = W _ {3} ^ {\\top} x, K = W _ {4} ^ {\\top} D _ {t}, V = D _ {t},\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.148, + 0.826, + 0.469, + 0.843 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {A} = \\operatorname {S o f t m a x} \\left(\\boldsymbol {Q} ^ {\\top} \\boldsymbol {K} / \\sqrt {\\sigma}\\right), \\tag {9}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.15, + 0.846, + 0.298, + 0.866 + ], + "angle": 0, + "content": "\\[\n\\mathbb {E} _ {\\boldsymbol {d} _ {t}} \\left[ g _ {y} (\\boldsymbol {d} _ {t}) \\right] = \\boldsymbol {V} ^ {\\top} \\boldsymbol {A},\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.871, + 0.469, + 0.901 + ], + "angle": 0, + "content": "where \\( W_{3} \\) and \\( W_{4} \\) both are fully connected layers, and \\( \\sigma \\) is the output dimension of them." + }, + { + "type": "code_caption", + "bbox": [ + 0.501, + 0.091, + 0.778, + 0.107 + ], + "angle": 0, + "content": "Algorithm 1 CafeBoost training pipeline." + }, + { + "type": "text", + "bbox": [ + 0.501, + 0.111, + 0.892, + 0.141 + ], + "angle": 0, + "content": "Input: Training data \\(\\mathcal{D}_0, \\dots, \\mathcal{D}_T\\), empty dictionary \\(D_0\\), empty example memory \\(\\mathcal{M}\\)," + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.142, + 0.892, + 0.171 + ], + "angle": 0, + "content": "1: Initialize parameters of feature extractor \\(\\psi_0\\), causal bias module \\(\\phi_0\\) (Eq. 8, 9) and classifier \\(\\theta_0\\)" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.172, + 0.71, + 0.186 + ], + "angle": 0, + "content": "2: // First task training stage." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.187, + 0.796, + 0.202 + ], + "angle": 0, + "content": "3: Train \\(\\psi_0, \\theta_0\\) on \\(\\mathcal{D}_0\\) by minimizing \\(\\mathcal{L}_{CE}\\)" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.203, + 0.806, + 0.217 + ], + "angle": 0, + "content": "4: Extract example features on \\(\\mathcal{D}_0\\), using \\(\\psi_0\\)" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.218, + 0.855, + 0.231 + ], + "angle": 0, + "content": "5: Store class mean features on \\(\\mathcal{D}_0\\) to dictionary \\(D_0\\)" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.232, + 0.792, + 0.247 + ], + "angle": 0, + "content": "6: // Causal debias module training stage." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.248, + 0.795, + 0.262 + ], + "angle": 0, + "content": "7: Train \\(\\phi_0, \\theta_0\\) on \\(\\mathcal{D}_0\\) by minimizing \\(\\mathcal{L}_{CE}\\)" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.263, + 0.856, + 0.277 + ], + "angle": 0, + "content": "8: Extract example features on \\(\\mathcal{D}_0\\), using \\(\\psi_0\\) and \\(\\phi_0\\)" + }, + { + "type": "text", + "bbox": [ + 0.51, + 0.278, + 0.767, + 0.291 + ], + "angle": 0, + "content": "9: if example memory \\(\\mathcal{M}\\) is used then" + }, + { + "type": "text", + "bbox": [ + 0.51, + 0.293, + 0.737, + 0.307 + ], + "angle": 0, + "content": "10: Update example memory \\(\\mathcal{M}\\)" + }, + { + "type": "text", + "bbox": [ + 0.51, + 0.308, + 0.573, + 0.32 + ], + "angle": 0, + "content": "11: end if" + }, + { + "type": "text", + "bbox": [ + 0.506, + 0.322, + 0.765, + 0.337 + ], + "angle": 0, + "content": "12: // Class incremental training stage." + }, + { + "type": "text", + "bbox": [ + 0.506, + 0.338, + 0.671, + 0.352 + ], + "angle": 0, + "content": "13: for \\( t \\) in \\( \\{1, \\dots, T\\} \\) do" + }, + { + "type": "text", + "bbox": [ + 0.506, + 0.353, + 0.701, + 0.366 + ], + "angle": 0, + "content": "14: Add a new classifier \\(\\theta_t\\)" + }, + { + "type": "text", + "bbox": [ + 0.506, + 0.368, + 0.878, + 0.383 + ], + "angle": 0, + "content": "15: Train \\(\\psi_t, \\phi_t\\) and \\(\\theta_t\\) on \\(\\mathcal{D}_t \\cup \\mathcal{M}\\) (or \\(\\mathcal{D}_t\\)) by Eq. 10" + }, + { + "type": "text", + "bbox": [ + 0.506, + 0.384, + 0.868, + 0.398 + ], + "angle": 0, + "content": "16: Extract example features on \\(\\mathcal{D}_t\\), using \\(\\psi_t\\) and \\(\\phi_t\\)" + }, + { + "type": "text", + "bbox": [ + 0.506, + 0.399, + 0.784, + 0.412 + ], + "angle": 0, + "content": "17: if example memory \\(\\mathcal{M}\\) is used then" + }, + { + "type": "text", + "bbox": [ + 0.506, + 0.413, + 0.754, + 0.428 + ], + "angle": 0, + "content": "18: Update example memory \\(\\mathcal{M}\\)" + }, + { + "type": "text", + "bbox": [ + 0.506, + 0.429, + 0.774, + 0.443 + ], + "angle": 0, + "content": "19: Update dictionary \\(D_{t}\\) by Eq. 11" + }, + { + "type": "text", + "bbox": [ + 0.506, + 0.444, + 0.59, + 0.456 + ], + "angle": 0, + "content": "20: end if" + }, + { + "type": "text", + "bbox": [ + 0.505, + 0.459, + 0.583, + 0.47 + ], + "angle": 0, + "content": "21: end for" + }, + { + "type": "text", + "bbox": [ + 0.505, + 0.473, + 0.891, + 0.504 + ], + "angle": 0, + "content": "Output: Final feature extractor \\(\\psi_T\\), final causal debias module \\(\\phi_T\\), and all classifiers \\(\\{\\pmb{\\theta}_0, \\dots, \\pmb{\\theta}_T\\}\\)" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.538, + 0.701, + 0.553 + ], + "angle": 0, + "content": "4.3. Causal Feature Boost" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.564, + 0.892, + 0.639 + ], + "angle": 0, + "content": "In this section, we describe how the causal debias module is added into the continual learning pipeline, and the training strategies to enable the overall architecture. Additionally, we propose a momentum method to update the dictionary so as to catch more task information." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.642, + 0.892, + 0.793 + ], + "angle": 0, + "content": "Architecture. Our causal debias module can be added between the feature extractor and the classifier of existing continual learning architectures. In Fig. 2, given an image as input, the CNN backbone (e.g. ResNet [8]) will extract a feature map. Then, we feed this feature map and a pre-collected dictionary into our causal debias module, and achieve a newly-updated feature map which is forwarded to the classifier to get a class label. After a new task comes, we add one more classifier for new classes and add no other layers." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.796, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Training pipeline. Because the dictionary used in the causal debias module need to be filled with average features of classes, we train a feature extractor and a classifier by using data of the initial phase firstly. Then, we fix the feature extractor and run an extra phase before the incremental phases to train the newly-added causal debias module and the classifier together. In the remaining phases," + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16020" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.19, + 0.089, + 0.787, + 0.378 + ], + "angle": 0, + "content": "
MethodCIFAR-100ImageNet-SubImageNet-Full
Average accuracy(%)5102551025510
LwF†[16]49.5946.9845.5153.6247.6444.3244.3538.90
BiC†[31]59.3654.2050.0070.0764.9657.7362.6558.72
iCaRL [22]62.9556.9251.3065.0460.8254.5651.5046.89
LUCIR [11]63.5657.2050.2970.8468.3261.4464.4561.57
GeoDL [24]67.0860.9653.9871.2770.2464.8665.2364.46
CwD [23]66.8462.6656.2867.3664.0960.52--
CSCCT [3]66.8959.9152.3271.3367.9161.52--
iCaRL+CafeBoost69.4164.7061.6070.5367.0463.4053.4648.46
LUCIR+CafeBoost70.1066.5662.3771.5170.0367.7466.8864.77
Forgetting rate(%)5102551025510
LwF†[16]43.3643.5841.6655.3257.0055.1248.7047.94
BiC†[31]31.4232.5034.6027.0431.0437.8825.0628.34
iCaRL [22]22.9525.3025.7418.7021.3525.0926.0333.76
LUCIR [11]23.3026.4029.2131.8833.4835.4024.0827.29
GeoDL [24]16.8321.8324.918.689.0711.9711.0312.81
CwD [23]17.7026.8026.9111.1617.5325.41--
iCaRL+CafeBoost8.5112.9116.0812.2014.2318.5225.1928.24
LUCIR+CafeBoost6.999.0311.758.167.509.108.6511.73
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.379, + 0.895, + 0.421 + ], + "angle": 0, + "content": "Table 1. The average incremental accuracy and the forgetting rate on CIFAR-100, ImageNet-Sub and ImageNet-Full. The number of incremental phases are 5, 10, and 25. The number of exemplars per class is 20. An ideal continual learning method should have high average incremental accuracy and low forgetting rate. Results of models with \\(\\dagger\\) are taken from [24]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.436, + 0.471, + 0.526 + ], + "angle": 0, + "content": "the causal debias module are always learnable, whereas the classifiers of old tasks are fixed. Our causal debias module can be used in iCaRL [22] and LUCIR [11] as a plugin. Our overall approach borrows their corresponding training objectives. Therefore, the training objective of our approach (CafeBoost) is formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.177, + 0.538, + 0.47, + 0.561 + ], + "angle": 0, + "content": "\\[\n\\min _ {\\theta} \\left[ L _ {c e} (x, y, \\theta) + \\eta L _ {*} (\\theta) \\right], \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.572, + 0.47, + 0.633 + ], + "angle": 0, + "content": "where \\( L_{*}(\\theta) \\) denotes specific losses of the corresponding base methods, e.g., distillation loss for iCaRL, margin ranking loss and others for LUCIR. Algorithm 1 details out the CafeBoost training pipeline." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.633, + 0.47, + 0.709 + ], + "angle": 0, + "content": "Momentum dictionary updating. Our approach exploits an external dictionary to store mean features of each class in the initial phase. We can exploit a momentum updating strategy to renew features in the dictionary, if replay buffer is used. The new dictionary at task \\( t \\) is updated as" + }, + { + "type": "equation", + "bbox": [ + 0.097, + 0.719, + 0.47, + 0.739 + ], + "angle": 0, + "content": "\\[\n\\pmb {D} _ {t} = \\alpha \\pmb {D} _ {t - 1} + (1 - \\alpha) [ \\mathbf {f} _ {1} ^ {t - 1}, \\dots , \\mathbf {f} _ {c} ^ {t - 1}, \\dots , \\mathbf {f} _ {C} ^ {t - 1} ], \\quad (1 1)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.749, + 0.47, + 0.779 + ], + "angle": 0, + "content": "where \\(\\alpha\\) controls the updating rate, \\(\\mathbf{f}_c\\) represents a mean feature of class \\(c\\), and \\(C\\) is the number of classes." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.792, + 0.21, + 0.81 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.817, + 0.177, + 0.833 + ], + "angle": 0, + "content": "5.1. Settings" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Datasets & Protocols. We evaluate our approach on two commonly-used datasets: CIFAR-100 [15] and ImageNet [5]. CIFAR-100 contains 100 classes and each class has 500 images for training and 100 images for test, with" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.436, + 0.895, + 0.587 + ], + "angle": 0, + "content": "each image of size \\(32 \\times 32\\). ImageNet with about 1.3 million images of size \\(224 \\times 224\\) is divided into two CIL settings: ImageNet-Sub that owns 100 classes and ImageNet-Full that owns 1000 classes. We follow the protocol used in many recent works [11,23,24]. Models are firstly trained on half classes, i.e., 50 classes for CIFAR100 and ImageNet-Sub, and 500 for ImageNet-Full. Subsequently, we start incremental learning phases with adding a fixed number of unobserved classes in each phase. We choose three different numbers of incremental phases: 5, 10, and 25." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.587, + 0.895, + 0.663 + ], + "angle": 0, + "content": "Evaluation Metrics. The two commonly-used metrics to evaluate performance of a CIL model are 1) the average incremental accuracy [22] \\(\\mathcal{A}_T\\), and 2) the average forgetting \\(\\mathcal{F}_T\\). The average incremental accuracy and the average forgetting after \\(T\\) tasks are defined as" + }, + { + "type": "equation", + "bbox": [ + 0.567, + 0.669, + 0.826, + 0.711 + ], + "angle": 0, + "content": "\\[\n\\mathcal {A} _ {T} = \\frac {1}{T} \\sum_ {t = 1} ^ {T} A _ {t}, \\mathcal {F} _ {T} = \\frac {1}{T - 1} \\sum_ {t = 2} ^ {T} F _ {t},\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.716, + 0.892, + 0.781 + ], + "angle": 0, + "content": "where \\( a_{i,j} \\) denotes the accuracy on task \\( i \\) after the model learning task \\( j \\), \\( A_{t} = \\frac{1}{t}\\sum_{i=1}^{t}a_{i,t} \\), and \\( F_{t} = \\frac{1}{t-1}\\sum_{j=2}^{t}\\max_{i}(a_{i,t}-a_{j,t}) \\) with \\( i\\in\\uparrow,\\dots,j-1 \\). We report top-1 accuracy." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.781, + 0.895, + 0.901 + ], + "angle": 0, + "content": "Implementation details. All methods are implemented with PyTorch [20]. We use the ResNet18 for CIFAR-100 and ImageNet following [23]. For all experiments, SGD optimizer and the batch size 128 are used. For CIFAR-100 and ImageNet, we train models for 160 epochs and 90 epochs, respectively. For CIFAR-100, our results are averaged over three runs with different orders generated by random seed 1993. For ImageNet, the first order with random seed 1993" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "16021" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.237, + 0.092, + 0.731, + 0.107 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.153, + 0.117, + 0.371, + 0.25 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.384, + 0.117, + 0.592, + 0.25 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.606, + 0.118, + 0.812, + 0.251 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.153, + 0.252, + 0.371, + 0.392 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.385, + 0.253, + 0.592, + 0.392 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.607, + 0.253, + 0.813, + 0.392 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.395, + 0.893, + 0.424 + ], + "angle": 0, + "content": "Figure 6. Results on CIFAR-100 (5, 10, and 25 tasks), ImageNet-Sub (5, 10, and 25 tasks). The horizontal axis shows the task identifier and the vertical axis is the corresponding average incremental accuracy." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.439, + 0.47, + 0.515 + ], + "angle": 0, + "content": "is conventionally used. In addition, we deploy CafeBoost to two baselines: iCaRL [22] and LUCIR [11] and compare it with LwF [16], BiC [31], GeoDL [24], CwD [23] and CSCCT [3] on all datasets. GeoDL, CwD and CSCCT are implemented based on LUCIR." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.528, + 0.206, + 0.543 + ], + "angle": 0, + "content": "5.2. Evaluations" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.552, + 0.47, + 0.809 + ], + "angle": 0, + "content": "Results on CIFAR-100. Tab. 1 summarizes the experimental results of CIFAR-100 in three different numbers of incremental tasks. We observe that CafeBoost improves iCaRL baseline by \\(6.46\\%\\), \\(7.78\\%\\), \\(10.30\\%\\) for 5, 10, and 25 tasks, respectively. After adding to LUCIR baseline, we observe a performance boost: \\(6.54\\%\\), \\(9.36\\%\\), \\(12.08\\%\\), for 5, 10, and 25 tasks. The corresponding average incremental accuracies reach high levels of \\(70.10\\%\\), \\(66.56\\%\\), and \\(62.37\\%\\). Furthermore, CafeBoost achieves lower forgetting rates after applying to the baselines. It can decrease the forgetting rates of iCaRL by \\(14.44\\%\\), \\(12.39\\%\\), and \\(9.66\\%\\) and those of LUCIR by \\(16.31\\%\\), \\(17.37\\%\\), and \\(17.46\\%\\), for 5, 10, and 25 tasks. The first row of Fig. 6 shows CafeBoost improves performance of baselines by a large margin, achieves higher accuracies and prevents catastrophic forgetting better than the prior methods for CIL. See Appendix for more results about the forgetting rates." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.811, + 0.47, + 0.902 + ], + "angle": 0, + "content": "Results on ImageNet. Tab. 1 also reports experimental results of ImageNet-Sub and ImageNet-Full. On ImageNet-Sub, CafeBoost boosts average incremental accuracies by \\(0.67\\% \\sim 8.84\\%\\) over iCaRL and LUCIR baselines, for 5, 20, and 25 tasks, and decreases the corresponding forgetting rates by \\(6.57\\% \\sim 26.30\\%\\). For example, CafeBoost" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.439, + 0.893, + 0.621 + ], + "angle": 0, + "content": "improves average incremental accuracy of LUCIR from \\(61.44\\%\\) to \\(67.74\\%\\) \\((+6.30\\%)\\), and reduces forgetting rate from \\(35.40\\%\\) to \\(9.10\\%\\) \\((-26.30\\%)\\). The second row of Fig. 6 further validates effectiveness of CafeBoost for all numbers of incremental phases. Additionally, we observe in Fig. 6 CafeBoost has almost no effect on the initial phase, but brings great boost in the following incremental phases. CafeBoost also enhances performance of iCaRL and LUCIR on ImageNet-Full, as shown in Tab. 1. We observe CafeBoost improves accuracies of iCaRL by \\(1.96\\%\\), \\(1.57\\%\\), and those of LUCIR by \\(2.43\\%\\), \\(3.20\\%\\) for 5 and 10 incremental phases respectively." + }, + { + "type": "table", + "bbox": [ + 0.506, + 0.632, + 0.892, + 0.757 + ], + "angle": 0, + "content": "
SB1020304050
10LUCIR59.1260.2661.7761.6262.91
+Ours55.8762.6166.2168.2269.70
↑-3.25+2.35+4.44+6.60+6.79
5LUCIR54.6256.2056.6557.2956.90
+Ours49.0258.5363.3965.7365.95
↑-5.60+2.33+6.74+8.44+9.05
" + }, + { + "type": "table_caption", + "bbox": [ + 0.498, + 0.758, + 0.892, + 0.84 + ], + "angle": 0, + "content": "Table 2. Ablation study on impact of number of classes in the initial phase. \\( B \\) denotes number of classes in the initial phase. \\( S \\) is the number of new classes per incremental step. Experiments are conducted on CIFAR-100 with 20 exemplars per class. The reported values (\\%) are average incremental accuracies. \\( \\uparrow \\) denotes the increment of values." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.848, + 0.665, + 0.862 + ], + "angle": 0, + "content": "5.3. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.871, + 0.892, + 0.902 + ], + "angle": 0, + "content": "In this section, we conduct experiments to study the effect on our approach of 1) number of classes in initial phase," + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16022" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.114, + 0.089, + 0.434, + 0.216 + ], + "angle": 0, + "content": "
SRLUCIR+Ours↑
10554.23±0.7864.80±0.36+10.57
1059.98±0.5468.30±0.31+8.32
2063.56±0.5570.10±0.39+6.54
5547.74±0.8462.23±0.67+14.49
1053.22±1.2864.08±0.77+10.86
2057.20±1.1466.56±0.80+9.36
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.216, + 0.47, + 0.284 + ], + "angle": 0, + "content": "Table 3. Ablation study on impact of number of exemplars. \\( R \\) represents the number of exemplars per class. \\( S \\) is the number of new classes per incremental step. Experiments are conducted on CIFAR-100. The reported values (%) are average incremental accuracies. \\( \\uparrow \\) denotes the increment of values." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.302, + 0.465, + 0.367 + ], + "angle": 0, + "content": "
MethodsCIFAR-100ImageNet-Sub
AccuracyForgettingAccuracyForgetting
LUCIR40.10±1.2752.35±1.3160.4725.73
+Ours59.01±0.2223.31±0.7665.1518.05
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.368, + 0.468, + 0.41 + ], + "angle": 0, + "content": "Table 4. Comparisons of average incremental accuracies and forgetting rates with no replay data. Experiments are conducted with 5 incremental steps. 20 exemplars are stored per class." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.429, + 0.468, + 0.459 + ], + "angle": 0, + "content": "2) number of exemplars in memory, 3) different classifiers, and 4) coefficient of momentum dictionary update." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.461, + 0.468, + 0.641 + ], + "angle": 0, + "content": "Impact of number of classes in initial phase. We investigate CafeBoost with 10, 20, 30, 40, and 50 classes of initial phase in Tab. 2. CafeBoost can enhance performance when the number of classes in initial phase is larger than 10. The performance improvement increases from negative to positive, along the number for both \\( S = 10 \\) and \\( S = 5 \\). When \\( B \\) equates 10, our approach has a negative influence on the baseline. This is because characteristics of classes in the initial phase are insufficient to share for all classes in the dataset. However, when \\( B >= 20 \\), our approach boosts the baseline by a large margin, which shows that it tolerates a broad range of the number \\( B \\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.643, + 0.468, + 0.809 + ], + "angle": 0, + "content": "Impact of number of exemplars in memory. We vary the number of exemplars per class and show its effect in Tab. 3. Both the performance of baselines and CafeBoost improve along the increase of exemplar number per class, which is consistently observed for \\( S = 10 \\) and \\( S = 5 \\). In addition, the improvement margin decreases along the increase of \\( R \\). For instance, \\( +10.57 \\) for \\( S = 10 \\) and \\( R = 5 \\), whereas \\( +6.54 \\) for \\( S = 10 \\) and \\( R = 20 \\). It is worth noting that CafeBoost does not rely on a replay buffer. Tab. 4 shows CafeBoost outperforms remarkably the baseline on accuracy and forgetting in CIFAR-100 and ImageNet-Sub." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.81, + 0.468, + 0.901 + ], + "angle": 0, + "content": "Impact of different classifiers. We also conduct experiments to research behaviors of CafeBoost combined with three different classifiers. We adopt the classifiers CNN, \\(k\\)-NME, and AME, as in [11]. CNN classifies examples by exploiting one class prototype per class. \\(k\\)-NME uses \\(k\\) samples of each class to represent the class, while AME uti" + }, + { + "type": "table", + "bbox": [ + 0.505, + 0.089, + 0.9, + 0.23 + ], + "angle": 0, + "content": "
MethodiCaRLiCaRL+oursLUCIRLUCIR+ours
CNN553.47±0.1556.53±0.2663.56±0.5570.10±0.39
1049.30±0.4353.75±0.0357.20±1.1466.56±0.80
2545.59±0.7551.18±0.7850.29±1.1462.37±1.17
k-NME562.95±0.5169.40±0.2365.42±0.6769.77±0.35
1056.91±0.6464.70±0.4059.85±1.2066.53±0.72
2551.30±0.9261.6±0.9052.07±0.4362.64±1.30
AME565.02±0.6470.61±0.2966.76±0.8470.54±0.31
1058.27±0.6366.52±0.4661.54±1.3667.77±0.71
2552.69±0.9264.10±0.8954.84±1.7865.10±1.35
" + }, + { + "type": "table_caption", + "bbox": [ + 0.5, + 0.231, + 0.892, + 0.272 + ], + "angle": 0, + "content": "Table 5. The average incremental accuracy by using different classifiers (CNN, \\( k \\) -NME, AME) with 20 exemplars per class. Values are evaluated on CIFAR-100 with 5, 10, and 25 incremental steps." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.288, + 0.892, + 0.361 + ], + "angle": 0, + "content": "izes all samples. Tab. 5 shows our approach achieves the best performance when using AME classifier. Moreover, it is noteworthy that our approach enhances performance of baselines using all three classifiers by a large margin in various numbers of tasks." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.363, + 0.892, + 0.424 + ], + "angle": 0, + "content": "Sensitivity of momentum coefficient. Fig. 7 shows our approach is not sensitive to the momentum coefficient of dictionary updating for all three classifiers. We set \\(\\alpha\\) to 0.9 for other experiments of our approach." + }, + { + "type": "image", + "bbox": [ + 0.514, + 0.44, + 0.727, + 0.565 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.763, + 0.437, + 0.892, + 0.561 + ], + "angle": 0, + "content": "Figure 7. Experiments on sensitivity of momentum coefficient. Average accuracies are evaluated on CIFAR-100 with 5 incremental steps and 20 exemplars per class." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.58, + 0.627, + 0.595 + ], + "angle": 0, + "content": "6. Conclusions" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.605, + 0.892, + 0.83 + ], + "angle": 0, + "content": "To the best of our knowledge, this work is the first to consider a new type of bias - task-induced bias in a causal perspective, which is derived from special settings of continual learning. We analyzed three continual scenarios in a causal framework, and found that the task-induced bias is particularly detrimental to CIL. Based on this framework, we designed a causal intervention operation and implemented it as a causal debias module by exploiting a powerful attention mechanism. Comprehensive experiments demonstrated our plugin method can boost performance of baselines and outperform several existing strong methods. In the future, we think it is a promising direction to regard the task-induced bias as a problem in CIL and design more exquisite approaches to eliminate it by cutting the corresponding causal path." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.831, + 0.892, + 0.887 + ], + "angle": 0, + "content": "Acknowledgments. This work was supported in part by National Key R&D Program of China (2021ZD0112001) and National Natural Science Foundation of China (No.61831005, 61971095 and 62271119)." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.956 + ], + "angle": 0, + "content": "16023" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.079, + 0.09, + 0.176, + 0.107 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.116, + 0.47, + 0.185 + ], + "angle": 0, + "content": "[1] Ehsan Abbasnejad, Damien Teney, Amin Parvaneh, Javen Shi, and Anton van den Hengel. Counterfactual vision and language learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10041—10051, 2020. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.187, + 0.472, + 0.243 + ], + "angle": 0, + "content": "[2] Hongjoon Ahn, Jihwan Kwak, Subin Lim, Hyeonsu Bang, Hyojun Kim, and Taesup Moon. SS-IL: Separated softmax for incremental learning. In Proceedings of the IEEE/CVF Conference on Computer Vision, pages 844–853, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.245, + 0.472, + 0.3 + ], + "angle": 0, + "content": "[3] Arjun Ashok, K. J. Joseph, and Vineeth Balasubramanian. Class-incremental learning with cross-space clustering and controlled transfer. In European Conference on Computer Vision, 2022. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.302, + 0.472, + 0.371 + ], + "angle": 0, + "content": "[4] Matthias Delange, Rahaf Aljundi, Marc Masana, Sarah Parisot, Xu Jia, Ales Leonardis, Greg Slabaugh, and Tinne Tuytelaars. A continual learning survey: Defying forgetting in classification tasks. Transactions on Pattern Analysis and Machine Intelligence, 44(7):3366-3385, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.373, + 0.472, + 0.441 + ], + "angle": 0, + "content": "[5] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. ImageNet: A large-scale hierarchical image database. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 248-255, 2009. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.444, + 0.472, + 0.5 + ], + "angle": 0, + "content": "[6] Xiang Deng and Zhongfei Zhang. Comprehensive knowledge distillation with causal intervention. In Advances in Neural Information Processing Systems, pages 22158-22170, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.502, + 0.472, + 0.571 + ], + "angle": 0, + "content": "[7] Chen He, Ruiping Wang, and Xilin Chen. A tale of two CILs: The connections between class incremental learning and class imbalanced learning, and beyond. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshop, pages 3554-3564, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.573, + 0.472, + 0.628 + ], + "angle": 0, + "content": "[8] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2016. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.63, + 0.472, + 0.685 + ], + "angle": 0, + "content": "[9] Lisa Anne Hendricks, Kaylee Burns, Kate Saenko, Trevor Darrell, and Anna Rohrbach. Women also snowboard: Overcoming bias in captioning models. In European Conference on Computer Vision, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.687, + 0.472, + 0.728 + ], + "angle": 0, + "content": "[10] Geoffrey Hinton, Oriol Vinyals, and Jeff Dean. Distilling the knowledge in a neural network. In ArXiv: 1503.02531, 2015. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.73, + 0.472, + 0.8 + ], + "angle": 0, + "content": "[11] Saihui Hou, Xinyu Pan, Chen Change Loy, Zilei Wang, and Dahua Lin. Learning a unified classifier incrementally via rebalancing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 831-839, 2019. 1, 2, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.802, + 0.472, + 0.871 + ], + "angle": 0, + "content": "[12] Xinting Hu, Kaihua Tang, Chunyan Miao, Xian-Sheng Hua, and Hanwang Zhang. Distilling causal effect of data in class-incremental learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3956-3965, 2021. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.873, + 0.472, + 0.901 + ], + "angle": 0, + "content": "[13] K. J. Joseph, Salman Khan, Fahad Shahbaz Khan, Rao Muhammad Anwer, and Vineeth N. Balasubramanian." + }, + { + "type": "list", + "bbox": [ + 0.079, + 0.116, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.135 + ], + "angle": 0, + "content": "Energy-based latent aligner for incremental learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7452-7461, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.137, + 0.892, + 0.205 + ], + "angle": 0, + "content": "[14] Minsoo Kang, Jaeyoo Park, and Bohyung Han. Class incremental learning by knowledge distillation with adaptive feature consolidation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16071-16080, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.207, + 0.892, + 0.235 + ], + "angle": 0, + "content": "[15] Alex Kriszhevsky. Learning multiple layers of features from tiny images. In Technical report, 2009. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.236, + 0.892, + 0.277 + ], + "angle": 0, + "content": "[16] Zhizhong Li and Derek Hoiem. Learning without forgetting. Transactions on Pattern Analysis and Machine Intelligence, 40(12):2935-2947, 2018. 1, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.279, + 0.892, + 0.349 + ], + "angle": 0, + "content": "[17] Chengzhi Mao, Kevin Xia, James Wang, Hao Wang, Junfeng Yang, Elias Bareinboim, and Carl Vondrick. Causal transportability for visual recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7521-7531, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.35, + 0.892, + 0.406 + ], + "angle": 0, + "content": "[18] Michael McCloskey and Neal J. Cohen. Catastrophic interference in connectionist networks: The sequential learning problem. In *Psychology of Learning and Motivation*, volume 24, pages 109–165. Academic Press, 1989. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.407, + 0.892, + 0.462 + ], + "angle": 0, + "content": "[19] Ninareh Mehrabi, Fred Morstatter, Nripsuta Saxena, Kristina Lerman, and Aram Galstyan. A survey on bias and fairness in machine learning. ACM Computing Surveys, 54(6):1-35, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.464, + 0.892, + 0.588 + ], + "angle": 0, + "content": "[20] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, Alban Desmaison, Andreas Köpf, Edward Yang, Zach DeVito, Martin Raison, Alykhan Tejani, Sasank Chilamkurthy, Benoit Steiner, Lu Fang, Junjie Bai, and Soumith Chintala. PyTorch: An imperative style, high-performance deep learning library. In Advances in Neural Information Processing Systems, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.59, + 0.892, + 0.63 + ], + "angle": 0, + "content": "[21] Judea Pearl, Madelyn Glymour, and Nicholas P. Jewell. Causal Inference In Statistics: A primer. Wiley, 2016. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.632, + 0.892, + 0.702 + ], + "angle": 0, + "content": "[22] Sylvestre Alvise Rebuffi, Alexander Kolesnikov, Georg Sperl, and Christoph H. Lampert. iCaRL: Incremental classifier and representation learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5533-5542, 2017. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.704, + 0.892, + 0.788 + ], + "angle": 0, + "content": "[23] Yujun Shi, Kuangqi Zhou, Jian Liang, Zihang Jiang, Jiashi Feng, Philip H. S. Torr, Song Bai, and Vincent Y. F. Tan. Mimicking the oracle: An initial phase decorrelation approach for class incremental learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16722-16731, 2022. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.789, + 0.892, + 0.845 + ], + "angle": 0, + "content": "[24] Christian Simon, Piotr Koniusz, and Mehrtash Harandi. On learning the geodesic path for incremental learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1591-1600, 2021. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.846, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[25] Qing Sun, Fan Lyu, Fanhua Shang, Wei Feng, and Liang Wan. Exploring example influence in continual learning. In Advances in Neural Information Processing Systems, 2022. 2" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "16024" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.161 + ], + "angle": 0, + "content": "[26] Rishabh Tiwari, Krishnateja Killamsetty, Rishabh Iyer, and Pradeep Shenoy. GCR: Gradient coreset based replay buffer selection for continual learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 99-108, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.163, + 0.469, + 0.19 + ], + "angle": 0, + "content": "[27] Gido M van de Ven and Andreas S Tolias. Three scenarios for continual learning. In arXiv: 1904.07734, 2019. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.192, + 0.469, + 0.259 + ], + "angle": 0, + "content": "[28] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Advances in Neural Information Processing Systems, page 5999-6009, 2017. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.262, + 0.47, + 0.318 + ], + "angle": 0, + "content": "[29] Tan Wang, Jianqiang Huang, Hanwang Zhang, and Qianru Sun. Visual commonsense R-CNN. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.32, + 0.469, + 0.347 + ], + "angle": 0, + "content": "[30] Max Welling. Herding dynamical weights to learn. In International Conference on Machine Learning, 2009. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.349, + 0.469, + 0.416 + ], + "angle": 0, + "content": "[31] Yue Wu, Yinpeng Chen, Lijuan Wang, Yuancheng Ye, Zicheng Liu, Yandong Guo, and Yun Fu. Large scale incremental learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 374-382, 2019. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.419, + 0.469, + 0.5 + ], + "angle": 0, + "content": "[32] Kelvin Xu, Jimmy Lei Ba, Ryan Kiros, Kyunghyun Cho, Aaron Courville, Ruslan Salakhutdinov, Richard S. Zemel, and Yoshua Bengio. Show, attend and tell: Neural image caption generation with visual attention. In International Conference on Machine Learning, pages 2048-2057, 2015. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.503, + 0.469, + 0.571 + ], + "angle": 0, + "content": "[33] Shipeng Yan, Jiangwei Xie, and Xuming He. DER: Dynamically expandable representation for class incremental learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3013-3022, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.574, + 0.469, + 0.614 + ], + "angle": 0, + "content": "[34] Xu Yang, Hanwang Zhang, and Jianfei Cai. Deconfounded image captioning: A causal retrospect. Transactions on Pattern Analysis and Machine Intelligence, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.616, + 0.469, + 0.684 + ], + "angle": 0, + "content": "[35] Bowen Zhao, Xi Xiao, Guojun Gan, Bin Zhang, and Shu Tao Xia. Maintaining discrimination and fairness in class incremental learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13205-13214, 2020. 1, 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.684 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16025" + } + ] +] \ No newline at end of file diff --git a/2023/CafeBoost_ Causal Feature Boost To Eliminate Task-Induced Bias for Class Incremental Learning/6f2564b9-d9c3-4c4f-b871-b0003952027b_origin.pdf b/2023/CafeBoost_ Causal Feature Boost To Eliminate Task-Induced Bias for Class Incremental Learning/6f2564b9-d9c3-4c4f-b871-b0003952027b_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..72d5d112c435abade1063a1aa35a37360a638d1f --- /dev/null +++ b/2023/CafeBoost_ Causal Feature Boost To Eliminate Task-Induced Bias for Class Incremental Learning/6f2564b9-d9c3-4c4f-b871-b0003952027b_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:331e1ba9a9a83d718e509f6e1e47cad1f2895404fa046fc07fcab32f0d62f4a0 +size 925094 diff --git a/2023/CafeBoost_ Causal Feature Boost To Eliminate Task-Induced Bias for Class Incremental Learning/full.md b/2023/CafeBoost_ Causal Feature Boost To Eliminate Task-Induced Bias for Class Incremental Learning/full.md new file mode 100644 index 0000000000000000000000000000000000000000..04cc0aeff3795e1176768ab4de4816673530375e --- /dev/null +++ b/2023/CafeBoost_ Causal Feature Boost To Eliminate Task-Induced Bias for Class Incremental Learning/full.md @@ -0,0 +1,391 @@ +# CafeBoost: Causal Feature Boost to Eliminate Task-Induced Bias for Class Incremental Learning + +Benliu Qiu + +Hongliang Li* + +Haitao Wen + +Heqian Qiu + +Lanxiao Wang* + +Fanman Meng + +Qingbo Wu + +Lili Pan + +University of Electronic Science and Technology of China, Chengdu, China + +{qbenliu, haitaowen, hqqiu, lanxiao.wang}@std.uestc.edu.cn + +{h1li, fmmeng, qbwu}@uestc.edu.cn panlili8255@gmail.com + +# Abstract + +Continual learning requires a model to incrementally learn a sequence of tasks and aims to predict well on all the learned tasks so far, which notoriously suffers from the catastrophic forgetting problem. In this paper, we find a new type of bias appearing in continual learning, coined as task-induced bias. We place continual learning into a causal framework, based on which we find the task-induced bias is reduced naturally by two underlying mechanisms in task and domain incremental learning. However, these mechanisms do not exist in class incremental learning (CIL), in which each task contains a unique subset of classes. To eliminate the task-induced bias in CIL, we devise a causal intervention operation so as to cut off the causal path that causes the task-induced bias, and then implement it as a causal debias module that transforms biased features into unbiased ones. In addition, we propose a training pipeline to incorporate the novel module into existing methods and jointly optimize the entire architecture. Our overall approach does not rely on data replay, and is simple and convenient to plug into existing methods. Extensive empirical study on CIFAR-100 and ImageNet shows that our approach can improve accuracy and reduce forgetting of well-established methods by a large margin. + +# 1. Introduction + +Deep learning has been widely applied in many fields like computer vision, social media analytics, natural language processing, etc. The standard paradigm of deep learning is to train models on a prepared dataset, of which all data are accessible in the whole training stage. However, in most real-world applications, a sequence of tasks arrives incrementally, which requires models to learn continuously from a new task, namely continual learning [27]. Despite fine-tuning achieves this goal somewhat, such naive method forgets obtained knowledge of previous tasks after learning + +![](images/611a0457922a71cbe8aa8c7ad14ff5e4cf77871ed1591514252685cb139a7c54.jpg) +(a) Illustration of CIL + +![](images/22da54b8de582da85f61a1c55395d23cc884fa1b07d0a717f8cbebab9c3e8dd7.jpg) +(b) Task-induced bias between class features + +![](images/001c7710a5d3ca6c868622763ceebbd82037903c1918bd266a6171e7d2a197fb.jpg) +(c) Task-induced bias between images and labels +Figure 1. Illustration of CIL, task-induced bias and their detrimental effects. (a) Illustration of CIL. (b) Task-induced bias between class features. Two classes in a task are correlated by a task identifier. (c) Task-induced bias between images and labels. Images and labels in a task are correlated by a task identifier. In (b)(c), the first column describes perspectives of task-induced bias, the second shows the underlying causal effect causing task-induced bias, and the last illustrates adverse influence of task-induced bias. + +a new task, dubbed as catastrophic forgetting [18]. + +Much effort has been devoted as a remedy for alleviating this problem in literature [2, 3, 11-14, 16, 33]. One general direction is to preserve knowledge representations of previous tasks by knowledge distillation [10] and by storing a few observed exemplars for replay [22]. [11] discovers the class imbalance between old classes and new ones is a key challenge of large-scale applications in class incremental learning (CIL), in which classes of a new task never appear in the previous tasks, as illustrated in Fig. 1 (a). [2,7,31,35] all follow this direction and propose effective methods to overcome the class imbalance. However, these methods ignore bias problems in continual learning. In this work, we + +consider the problem of CIL in a novel perspective of bias. We think that catastrophic forgetting is the reflection of a new-type bias, coined as task-induced bias. + +Deep neural networks are easily misled by language bias [1] and vision bias [6, 9, 29], which are both derived from spurious correlations of two or more items in different positions of a sentence or an image. In contrast to these two bias, the task-induced bias occurs across different images in a common task. In CIL, data of specific classes are accessible only in a certain task if replay buffer is not applied. A deep neural network is capable of learning correlations among images of different categories or between images and labels as possible as it can, which includes a few spurious correlations, namely biases, leading to a limited generalization. On the one hand, a portion of these image correlations can be shared beneficially in this task, but these benefits cannot share among different tasks. For example, in Fig. 1 (b), color is sufficient to discriminate elephants and leopards in a specific task since elephants are grey while leopards are orange, and thus deep models have no motivation to capture more complex characteristic, like stripe. This is detrimental to the feature extractor because it captures a few details that cannot generalize among all tasks. On the other hand, as shown in Fig. 1 (c), a classifier will be biased to new classes in the current task [11, 31, 35], because the classes in the training data are all the new classes and to predict new classes is more possible to be correct. + +To better understand the task-induced bias in continual learning, we model three continual learning scenarios [27] into a common causal inference framework, which merely contains three necessary variables: image, label, and task identifier. Thanks to the framework, we discover that the task-induced bias is minor in task incremental and domain incremental learning due to innate mechanisms that cut off an inappropriate causal path, which partially explain why the two scenarios are easier to tackle than CIL [27]. For the existence of the causal path in CIL, we seek an active causal intervention to close the path. To implement the intervention, we devise a simple but effective causal debias module, which is established on attention mechanisms [28]. Subsequently, we incorporate this module into existing CIL architectures and propose a training pipeline to optimize it. Our overall approach transforms causally biased features into unbiased ones, and hence we refer to it as Casual Feature Boost (CafeBoost). Finally, we systematically compare accuracy and forgetting of well-established methods with ours in various settings. Experimental results show that CafeBoost yields significant and consistent performance boost for existing methods in CIFAR-100 and ImageNet by a large margin $(0.67\% \sim 12.08\%$ on accuracy). + +# 2. Related Work + +Continual learning. Existing CIL works mostly aim to maintain knowledge representations of previous tasks, in- + +![](images/78f4c3a1b646b817171e8d8de5ba734e13f29b73a66eaabb932008587fc81506.jpg) +Figure 2. Overview of Casual Feature Boost (CafeBoost). Green blocks are learnable, whereas white blocks are fixed. Conv block is updated with a smaller learning rate than Causal block. The training objectives including distillation loss, margin ranking loss, etc., are omitted in this figure, since they depend on baselines. Whether to use replay buffer $\mathcal{M}$ depends on specific baselines. + +including replay-based methods and distillation-based methods. Replay-based methods [22, 25, 26] store a few examples to represent previous data and replay the stored data in the current training stage. The purpose of relay-based methods is to select the most representative samples of the entire training dataset. iCaRL [22] uses herding strategy [30] to choose typical exemplars for future training. [25] calculates example influence of stability and plasticity for every example in validation sets, based on which it selects examples for replay. Distillation-based methods [3,11,23,24] focus on transferring learned knowledge of a previous model to the current one. LUCIR [11] combines cosine normalization, less-forget constraint, and inter-class separation to mitigate the forgetting caused by class imbalance. GeoDL [24] utilizes geodesic path for knowledge distillation which considers the gradual change between subspaces of tasks. CSCCT [3] encourages features of each class to locate in an appropriate position of the feature space, and utilizes inter-class relationships to transfer inter-class knowledge positively or negatively. CwD [23] scatters uniformly representations of each class in the initial task, which significantly benefits performance in subsequent tasks. These two types of methods are usually incorporated together to achieve better performance. + +Spurious correlation. Spurious correlation is also known as bias. [19] has extensively investigated the bias problem in machine learning. Many methods [1,9,29] have been proposed to eliminate bias in their specific applications. [1] introduces counterfactual in the training to eliminate bias in visual question answering datasets. [9] proposes a new Equalizer model to force a model look at a person instead of inferring a person's gender by context cues. [29] applies causal intervention to alleviate vision bias + +among detected objects in an image. + +Causal inference. Causal inference [21] aims to establish reasonable causal graphs for specific applications, then analyze problems and propose solutions based on the established causal graph. Causal intervention and counterfactual process are two powerful tools of causal inference. These two tools have been ubiquitously applied in multiple applications, such as out-of-distribution recognition [17], visual question answering [1], image captioning [34], knowledge distillation [6], etc. Recently, [12] builds a causal view to analyze the cause of forgetting in class incremental learning, and designs an anti-forgetting technique so as to distill causal effect of data without replay. + +# 3. Preliminary + +# 3.1. Continual learning scenarios + +Continual learning can be categorized into three scenarios [27] in the light of whether a task identifier is given in a test phase and whether a model is required to infer which task it meets. If the task identifier is available during both training and test phases, this scenario is task incremental. If the task identifier is accessible in training phase but it is only necessary for a model to solve the task at hand, this scenario is domain incremental. Similar to the domain incremental scenario, the class incremental scenario doesn't provide task identifiers in test phase, but it requires a model to infer the current task identifier. Generally speaking, the class incremental scenario is the most difficult among the three, whereas the task incremental scenario is the easiest [27]. + +To formulate continual learning, we assume a sequence of $T$ tasks is given as $\{\mathcal{T}_0,\mathcal{T}_1,\dots,\mathcal{T}_{T - 1}\}$ , each with data $\mathcal{D}_t,t = 0,1,\ldots ,T - 1$ . Every sample in $\mathcal{D}_t$ is a tuple of an image $\mathbf{x}_i$ and a corresponding label $y_{i}$ , and hence we can denote the data of task $t$ as $\mathcal{D}_t = \{(\mathbf{x}_i,y_i)_{i = 1}^{N_t}\}$ where $N_{t}$ is the number of samples, $\mathbf{x}_i\in \mathcal{X}_t$ and $y_{i}\in \mathcal{Y}_{t}$ . Replay-based continual learning methods require a memory buffer $\mathcal{M}$ to store partial data of previous tasks. For the task $\mathcal{T}_0$ , only data $\mathcal{D}_0$ can be used for training, but for the task $\mathcal{T}_t,t > 0$ , we can train models using data $\mathcal{M}\cup \mathcal{D}_t$ . For the task incremental scenario, task identifier $\mathcal{T}$ is given as an input to models in test phase but not for the other two scenarios. + +After the above formulation, we elucidate distinctions of three incremental scenarios in a formal style [4]. Let $P(\cdot)$ denote a marginal distribution. The common ground of three scenarios is $P(\mathcal{X}_t) \neq P(\mathcal{X}_{t + 1})$ for all $t$ . Their distinctions lie in label set $\mathcal{Y}_t$ and its marginal distribution $P(\mathcal{Y}_t)$ . For task incremental learning, $\mathcal{Y}_t \neq \mathcal{Y}_{t + 1}$ . For domain incremental learning, $\mathcal{Y}_t = \mathcal{Y}_{t + 1}$ and $P(\mathcal{Y}_t) = P(\mathcal{Y}_{t + 1})$ . These two scenarios have no requirement on inferring task identifiers. Conversely, the class incremental learning that defines $\mathcal{Y}_t \subset \mathcal{Y}_{t + 1}$ and $P(\mathcal{Y}_t) \neq P(\mathcal{Y}_{t + 1})$ , requires models to infer which label set is used for an arbitrary task. + +![](images/f08dc029d1598dc6b2016c5217f8a1b49735b2e71e782590799b520563fc7740.jpg) + +![](images/48cfaa10052886a2460f9cc3f076be9842e0fe7d12daf515695bc285c43b6246.jpg) + +![](images/1b55d7506b2fb981f50787d5f8ee3cfa13698623a426e647e1022a83ec410f0c.jpg) + +![](images/533436adbcd9fc65d1aa5fe309363ae21e3f3f0da36b2b5fbd7303d97f695ffc.jpg) +Figure 3. Causal graphs of supervised learning and continual learning. Gray nodes means unconditioned variables. (a) Causal graphs of supervised learning in training and test phase. (b) Causal graphs of continual learning. The red task variable depends on specific continual scenarios. (c) Causal path between images and labels in task incremental learning. The path $X \gets T \to Y$ is impeded because variable $T$ is conditioned. (d) Causal path between images and labels in domain incremental learning. The path $X \gets T \to Y$ is cut off because of $T \nrightarrow Y$ . (e) Causal path between images and labels in class incremental learning. The path $X \gets T \to Y$ causes task-induced bias. + +![](images/134c16f3195c1b63fac8a792eb263ee83fbe35bd61b7905d52b3a8d4ff1f7e74.jpg) + +![](images/ec6fb4e487aa0966b00f3427d943d74afcc567e0db56dbd8d8d0a687938a3e6a.jpg) + +# 3.2. Causal graphs in continual learning + +A causal graph is a directed acyclic graph, of which nodes denote variables and directed edges denote the causality between two variables. We treat images, labels and task identifiers as nodes in our causal graphs. Although only three nodes are incorporated into our causal graph, it is powerful to reveal underlying characteristics of continual learning as we will show in Sec. 4. + +We denote an image as $X$ , a label as $Y$ and a task identifier as $T$ in Fig. 3. The edges are listed below: $X \to Y$ means label $Y$ is deduced from image $X$ ; $T \to X$ denotes that task identifier $T$ have a causal influence to image $X$ ; $T \to Y$ denotes task identifier $T$ causally affect label $Y$ . + +For supervised learning in Fig. 3 (a), its causal graph only contains an image $X$ , a label $Y$ and an causal path $X \to Y$ that represents a model to capture the causality between two variables. In training phase of supervised learning, both image $X$ and label $Y$ are accessible and we utilize them to learn the causal effect into the model. But in test phase, an image $X$ and the causal path are known and $Y$ is the label to be inferred. For three continual learning scenarios, images, labels and task identifiers are all given in training phase. As a result, they have the same causal graph for training in Fig. 3 (b). We illustrate their distinctions in test phase in Fig. 3 (b). Due to the involvement of the task identifier $T$ , $X$ may have an indirect effect on $Y$ through the causal path $X \gets T \to Y$ , solely dependent on $T$ . We will explain it in Sec. 4. See Appendix for additional explanation of causal graphs. + +# 4. Methodology + +In this section, we introduce the task-induced bias through the causal framework, a causal intervention for + +CIL, its corresponding implementation, and the entire pipeline of our approach. + +# 4.1. Task-induced bias in three continual scenarios + +According to the causal graphs, we can explain why task incremental learning is the most effortless scenario among the three and why class incremental learning is the most challenging. We think that this phenomenon is caused by task-induced bias. + +In Fig. 3 (b), the causal path $X \leftarrow T \rightarrow Y$ is of a fork structure, in which two kid nodes are independent conditioned on the parent. [21] points out that if $T$ is unconditioned, the causal influence of $X$ to $Y$ exists, causing the task-induced bias. We give a simple example to help understand this statement. A causal graph lighter $\leftarrow$ smoking $\rightarrow$ cancer means that a smoking person has a possibility to bring a lighter and get a lung cancer. Whether a person brings a lighter is usually independent of getting a lung cancer. Nevertheless, a smoker is highly possible to bring a lighter and to get a lung cancer, which leads to observe a correlation between lighter and cancer. If we can control the confounder smoking and collect data about non-smoking persons who usually bring lighters, then we can draw a correct conclusion that lighter is independent on cancer. + +Fortunately, mechanisms to eliminate the task-induced bias naturally exist under setups of task incremental learning and domain incremental learning. In task incremental learning shown in Fig. 3 (c), the value of parent node (task identifier) is given, hence the indirected effect of $X$ on $Y$ is 0 and the task-induced bias between $X$ and $Y$ disappears due to the characteristic of the fork structure. In domain incremental learning, the label set remains the same along task identifiers, which is implied in $\mathcal{Y}_t = \mathcal{Y}_{t + 1}$ and $P(\mathcal{Y}_t) = P(\mathcal{Y}_{t + 1})$ . For this reason, the causal edge $T\to Y$ is cut as shown in Fig. 3 (d), resulting in disappearance of the task-induced bias. However, in the CIL, the causal path $X\gets T\to Y$ cannot be cut by the task identifiers since they are inaccessible, or by the independence between $T$ and $Y$ like domain incremental learning. In order to eliminate the task-induced bias in CIL, we seek advanced causal techniques among which causal intervention [21] is a powerful one. It is noteworthy that our analysis indicates that a key objective of CIL is to cut the causal path $X\gets T\to Y$ and we can cut it not only by conditioning variable $T$ , but also by cutting $X\gets T$ or $T\rightarrow Y$ . + +# 4.2. Causal debias + +The overall causal graph of CIL is illustrated in Fig. 4 (a), which considers two facets of task-induced bias in Fig. 1. The causal effect of $X_{1}$ and $X_{2}$ on $Y$ in the original causal graph can be formulated by Bayes' rule: + +$$ +P (Y | X _ {1}, X _ {2}) = \sum_ {t} P (Y | \underline {{X}} _ {1}, X _ {2}, t) P (t | \underline {{X}} _ {1}, X _ {2}) \tag {1} +$$ + +![](images/99e227d7b07a2ac02be42df53f19a6061d2893037227a7e0b519c1dacdd8a455.jpg) +(c) Simplified causal graph +(d) Causal intervention +(e) An implementation +Figure 4. Causal intervention for class incremental learning. To approximate the task identifier, we utilize a class feature dictionary $D_{t}$ recording information about tasks. + +We can treat the two variables $X_{1}$ and $X_{2}$ as one as shown in Fig. 4 (b), because they have the same context in the causal graph. As a result, we obtain a simplified causal graph in Fig. 4 (c), of which the causal effect of $X$ on $Y$ is reformulated as: + +$$ +P (Y | X) = \sum_ {t} P (Y | X, t) P (t | X). \tag {2} +$$ + +We have shown in Fig. 3 (e) that the causal path $X \gets T \to Y$ is the cause of task-induced bias. If we can cut this causal path, the task-induced bias will be eliminated. We opt to cut the $X \gets T$ using the causal intervention, as illustrated in Fig. 4 (d). The causal path $X \gets T$ is introduced through the term $P(t|X)$ in Eq. 2. Following [21], the direct causal effect of $X$ on $Y$ is denoted by $P(Y|do(X))$ , where $do(X)$ means a causal intervention on the single variable $X$ . Because the variable $T$ blocks all the backdoor path in the causal graph, we can obtain the expression of $P(Y|do(X))$ according to the backdoor criterion [21] as: + +$$ +P (Y | d o (X)) = \sum_ {t} P (Y | X, t) P (t) \tag {3} +$$ + +Compared with the old term $P(t|X)$ , the new term $P(t)$ is no longer changed with $X$ , equating to cut the causal edge $X \gets T$ , as shown in Fig. 4 (d). + +Based on the intervened causal graph in Fig. 4 (d) in CIL, we devise a causal debias module to implement the Equation 3. Since the task identifier is not accessible in the test phase, we propose to use a feature dictionary $D_{t}$ that contains information about tasks to approximate the unobserved confounder $T$ as shown in Fig. 4 (e). The task information is retained in class labels of images for CIL, because class label sets of different tasks cannot overlap with each other and thus task identifiers can be inferred from classes. Specifically, given an image $\mathbf{x}$ , a label $y$ and a task identifier + +![](images/50c829668428a825fcb6ab579a26e7a39876e9fdd5f438cfa2c4dc6c129684f8.jpg) +Figure 5. Illustration of causal intervention operation. This operation borrows features of classes in previous tasks, and puts them to features of new classes in the current task. The causal intervention is implied in the "borrow" and "put" operations, because the two operations mean that the features containing task information are directly controlled by ourselves. + +$t$ , Equation 3 is implemented as + +$$ +\begin{array}{l} P (Y | d o (X)) = \sum_ {t} P (y | \mathbf {x}, t) P (t) \\ = \sum_ {\boldsymbol {d} _ {t}} P (y | \boldsymbol {x}, \boldsymbol {d} _ {t}) P (\boldsymbol {d} _ {t}), \tag {4} \\ \end{array} +$$ + +where $\pmb{x}$ is the feature of the image $\mathbf{x}$ , $d_{t}$ is sampled from $\pmb{D}_{t}$ . If the last layer of a classifier is a softmax layer, we get $P(y|\pmb{x},\pmb{d}_t) = \mathrm{Softmax}(f_y(\pmb{x},\pmb{d}_t))$ , where $f_{y}(\cdot)$ is a similarity function. In a nutshell, the implementation of Equation 3 is newly formulated as: + +$$ +P (Y | d o (X)) := \mathbb {E} _ {\boldsymbol {d} _ {t}} \left[ \operatorname {S o f t m a x} \left(f _ {y} (\boldsymbol {x}, \boldsymbol {d} _ {t})\right) \right]. \tag {5} +$$ + +The above expectation can be approximated by NWGM [32] as follows: + +$$ +\mathbb {E} _ {\boldsymbol {d} _ {t}} \left[ \operatorname {S o f t m a x} \left(f _ {y} \left(\boldsymbol {x}, \boldsymbol {d} _ {t}\right)\right) \right] \approx \operatorname {S o f t m a x} \left(\mathbb {E} _ {\boldsymbol {d} _ {t}} \left[ f _ {y} \left(\boldsymbol {x}, \boldsymbol {d} _ {t}\right) \right]\right). \tag {6} +$$ + +To implement our causal intervention operation in Fig. 5, we devise a simple attention-based mechanism with ignorable number of parameters: + +$$ +f _ {y} (\boldsymbol {x}, \boldsymbol {d} _ {t}) = \boldsymbol {W} _ {0} ^ {\top} \operatorname {C a t} \left(\boldsymbol {W} _ {1} ^ {\top} \boldsymbol {x}, \boldsymbol {W} _ {2} ^ {\top} g _ {y} (\boldsymbol {d} _ {t})\right), \tag {7} +$$ + +where $W_{1}$ and $W_{2}$ both are fully connected layers and $\operatorname{Cat}(\cdot)$ is the concatenation operator. Due to the additive property of expectation, the $\mathbb{E}_{\boldsymbol{d}_t}[f_y(\boldsymbol{x}, \boldsymbol{d}_t)]$ can be rewritten as: + +$$ +\mathbb {E} _ {\boldsymbol {d} _ {t}} \left[ f _ {y} (\boldsymbol {x}, \boldsymbol {d} _ {t}) \right] = \boldsymbol {W} _ {\boldsymbol {0}} ^ {\top} \operatorname {C a t} \left(\boldsymbol {W} _ {\boldsymbol {1}} ^ {\top} \boldsymbol {x}, \boldsymbol {W} _ {\boldsymbol {2}} ^ {\top} \mathbb {E} _ {\boldsymbol {d} _ {t}} \left[ g _ {y} \left(\boldsymbol {d} _ {t}\right) \right]\right). \tag {8} +$$ + +We devise $\mathbb{E}_{\boldsymbol{d}_t}[g_y(\boldsymbol{d}_t)]$ utilizing a scaled dot-product attention [28], which is calculated as follows: + +$$ +Q = W _ {3} ^ {\top} x, K = W _ {4} ^ {\top} D _ {t}, V = D _ {t}, +$$ + +$$ +\boldsymbol {A} = \operatorname {S o f t m a x} \left(\boldsymbol {Q} ^ {\top} \boldsymbol {K} / \sqrt {\sigma}\right), \tag {9} +$$ + +$$ +\mathbb {E} _ {\boldsymbol {d} _ {t}} \left[ g _ {y} (\boldsymbol {d} _ {t}) \right] = \boldsymbol {V} ^ {\top} \boldsymbol {A}, +$$ + +where $W_{3}$ and $W_{4}$ both are fully connected layers, and $\sigma$ is the output dimension of them. + +Algorithm 1 CafeBoost training pipeline. + +Input: Training data $\mathcal{D}_0, \dots, \mathcal{D}_T$ , empty dictionary $D_0$ , empty example memory $\mathcal{M}$ , + +1: Initialize parameters of feature extractor $\psi_0$ , causal bias module $\phi_0$ (Eq. 8, 9) and classifier $\theta_0$ + +2: // First task training stage. + +3: Train $\psi_0, \theta_0$ on $\mathcal{D}_0$ by minimizing $\mathcal{L}_{CE}$ + +4: Extract example features on $\mathcal{D}_0$ , using $\psi_0$ + +5: Store class mean features on $\mathcal{D}_0$ to dictionary $D_0$ + +6: // Causal debias module training stage. + +7: Train $\phi_0, \theta_0$ on $\mathcal{D}_0$ by minimizing $\mathcal{L}_{CE}$ + +8: Extract example features on $\mathcal{D}_0$ , using $\psi_0$ and $\phi_0$ + +9: if example memory $\mathcal{M}$ is used then + +10: Update example memory $\mathcal{M}$ + +11: end if + +12: // Class incremental training stage. + +13: for $t$ in $\{1, \dots, T\}$ do + +14: Add a new classifier $\theta_t$ + +15: Train $\psi_t, \phi_t$ and $\theta_t$ on $\mathcal{D}_t \cup \mathcal{M}$ (or $\mathcal{D}_t$ ) by Eq. 10 + +16: Extract example features on $\mathcal{D}_t$ , using $\psi_t$ and $\phi_t$ + +17: if example memory $\mathcal{M}$ is used then + +18: Update example memory $\mathcal{M}$ + +19: Update dictionary $D_{t}$ by Eq. 11 + +20: end if + +21: end for + +Output: Final feature extractor $\psi_T$ , final causal debias module $\phi_T$ , and all classifiers $\{\pmb{\theta}_0, \dots, \pmb{\theta}_T\}$ + +# 4.3. Causal Feature Boost + +In this section, we describe how the causal debias module is added into the continual learning pipeline, and the training strategies to enable the overall architecture. Additionally, we propose a momentum method to update the dictionary so as to catch more task information. + +Architecture. Our causal debias module can be added between the feature extractor and the classifier of existing continual learning architectures. In Fig. 2, given an image as input, the CNN backbone (e.g. ResNet [8]) will extract a feature map. Then, we feed this feature map and a pre-collected dictionary into our causal debias module, and achieve a newly-updated feature map which is forwarded to the classifier to get a class label. After a new task comes, we add one more classifier for new classes and add no other layers. + +Training pipeline. Because the dictionary used in the causal debias module need to be filled with average features of classes, we train a feature extractor and a classifier by using data of the initial phase firstly. Then, we fix the feature extractor and run an extra phase before the incremental phases to train the newly-added causal debias module and the classifier together. In the remaining phases, + +
MethodCIFAR-100ImageNet-SubImageNet-Full
Average accuracy(%)5102551025510
LwF†[16]49.5946.9845.5153.6247.6444.3244.3538.90
BiC†[31]59.3654.2050.0070.0764.9657.7362.6558.72
iCaRL [22]62.9556.9251.3065.0460.8254.5651.5046.89
LUCIR [11]63.5657.2050.2970.8468.3261.4464.4561.57
GeoDL [24]67.0860.9653.9871.2770.2464.8665.2364.46
CwD [23]66.8462.6656.2867.3664.0960.52--
CSCCT [3]66.8959.9152.3271.3367.9161.52--
iCaRL+CafeBoost69.4164.7061.6070.5367.0463.4053.4648.46
LUCIR+CafeBoost70.1066.5662.3771.5170.0367.7466.8864.77
Forgetting rate(%)5102551025510
LwF†[16]43.3643.5841.6655.3257.0055.1248.7047.94
BiC†[31]31.4232.5034.6027.0431.0437.8825.0628.34
iCaRL [22]22.9525.3025.7418.7021.3525.0926.0333.76
LUCIR [11]23.3026.4029.2131.8833.4835.4024.0827.29
GeoDL [24]16.8321.8324.918.689.0711.9711.0312.81
CwD [23]17.7026.8026.9111.1617.5325.41--
iCaRL+CafeBoost8.5112.9116.0812.2014.2318.5225.1928.24
LUCIR+CafeBoost6.999.0311.758.167.509.108.6511.73
+ +Table 1. The average incremental accuracy and the forgetting rate on CIFAR-100, ImageNet-Sub and ImageNet-Full. The number of incremental phases are 5, 10, and 25. The number of exemplars per class is 20. An ideal continual learning method should have high average incremental accuracy and low forgetting rate. Results of models with $\dagger$ are taken from [24]. + +the causal debias module are always learnable, whereas the classifiers of old tasks are fixed. Our causal debias module can be used in iCaRL [22] and LUCIR [11] as a plugin. Our overall approach borrows their corresponding training objectives. Therefore, the training objective of our approach (CafeBoost) is formulated as: + +$$ +\min _ {\theta} \left[ L _ {c e} (x, y, \theta) + \eta L _ {*} (\theta) \right], \tag {10} +$$ + +where $L_{*}(\theta)$ denotes specific losses of the corresponding base methods, e.g., distillation loss for iCaRL, margin ranking loss and others for LUCIR. Algorithm 1 details out the CafeBoost training pipeline. + +Momentum dictionary updating. Our approach exploits an external dictionary to store mean features of each class in the initial phase. We can exploit a momentum updating strategy to renew features in the dictionary, if replay buffer is used. The new dictionary at task $t$ is updated as + +$$ +\pmb {D} _ {t} = \alpha \pmb {D} _ {t - 1} + (1 - \alpha) [ \mathbf {f} _ {1} ^ {t - 1}, \dots , \mathbf {f} _ {c} ^ {t - 1}, \dots , \mathbf {f} _ {C} ^ {t - 1} ], \quad (1 1) +$$ + +where $\alpha$ controls the updating rate, $\mathbf{f}_c$ represents a mean feature of class $c$ , and $C$ is the number of classes. + +# 5. Experiments + +# 5.1. Settings + +Datasets & Protocols. We evaluate our approach on two commonly-used datasets: CIFAR-100 [15] and ImageNet [5]. CIFAR-100 contains 100 classes and each class has 500 images for training and 100 images for test, with + +each image of size $32 \times 32$ . ImageNet with about 1.3 million images of size $224 \times 224$ is divided into two CIL settings: ImageNet-Sub that owns 100 classes and ImageNet-Full that owns 1000 classes. We follow the protocol used in many recent works [11,23,24]. Models are firstly trained on half classes, i.e., 50 classes for CIFAR100 and ImageNet-Sub, and 500 for ImageNet-Full. Subsequently, we start incremental learning phases with adding a fixed number of unobserved classes in each phase. We choose three different numbers of incremental phases: 5, 10, and 25. + +Evaluation Metrics. The two commonly-used metrics to evaluate performance of a CIL model are 1) the average incremental accuracy [22] $\mathcal{A}_T$ , and 2) the average forgetting $\mathcal{F}_T$ . The average incremental accuracy and the average forgetting after $T$ tasks are defined as + +$$ +\mathcal {A} _ {T} = \frac {1}{T} \sum_ {t = 1} ^ {T} A _ {t}, \mathcal {F} _ {T} = \frac {1}{T - 1} \sum_ {t = 2} ^ {T} F _ {t}, +$$ + +where $a_{i,j}$ denotes the accuracy on task $i$ after the model learning task $j$ , $A_{t} = \frac{1}{t}\sum_{i=1}^{t}a_{i,t}$ , and $F_{t} = \frac{1}{t-1}\sum_{j=2}^{t}\max_{i}(a_{i,t}-a_{j,t})$ with $i\in\uparrow,\dots,j-1$ . We report top-1 accuracy. + +Implementation details. All methods are implemented with PyTorch [20]. We use the ResNet18 for CIFAR-100 and ImageNet following [23]. For all experiments, SGD optimizer and the batch size 128 are used. For CIFAR-100 and ImageNet, we train models for 160 epochs and 90 epochs, respectively. For CIFAR-100, our results are averaged over three runs with different orders generated by random seed 1993. For ImageNet, the first order with random seed 1993 + +![](images/40b0fadf920b754d05c327f6107da3150598c29de32d14a4b4417b23fc602a8a.jpg) + +![](images/038c345e9eac4d2e9ffac27d24f8ecbbe3df915a08537de4c4cd5793fd408a72.jpg) + +![](images/de3b6a1e3f8156fff399204403ac09a3261f3c9645b06d4e15a5d85cff4d7431.jpg) + +![](images/40b992f1e8937c5bffe9dce24273eeaf673b0c865ac47d7656eea329e1912352.jpg) + +![](images/8a414de9bd3f211a56d8c2d5c4435c695155f9c581d0fe25649d3039ae1fce0f.jpg) +Figure 6. Results on CIFAR-100 (5, 10, and 25 tasks), ImageNet-Sub (5, 10, and 25 tasks). The horizontal axis shows the task identifier and the vertical axis is the corresponding average incremental accuracy. + +![](images/f31257b9d713d80071cb3cdc340df6d5c45d20d56c6bbd3c442c28b03fe35300.jpg) + +![](images/aa30277a053aa74a1688d2ceaa1d7a4a14e983758e3546cd3626acd04bb35f98.jpg) + +is conventionally used. In addition, we deploy CafeBoost to two baselines: iCaRL [22] and LUCIR [11] and compare it with LwF [16], BiC [31], GeoDL [24], CwD [23] and CSCCT [3] on all datasets. GeoDL, CwD and CSCCT are implemented based on LUCIR. + +# 5.2. Evaluations + +Results on CIFAR-100. Tab. 1 summarizes the experimental results of CIFAR-100 in three different numbers of incremental tasks. We observe that CafeBoost improves iCaRL baseline by $6.46\%$ , $7.78\%$ , $10.30\%$ for 5, 10, and 25 tasks, respectively. After adding to LUCIR baseline, we observe a performance boost: $6.54\%$ , $9.36\%$ , $12.08\%$ , for 5, 10, and 25 tasks. The corresponding average incremental accuracies reach high levels of $70.10\%$ , $66.56\%$ , and $62.37\%$ . Furthermore, CafeBoost achieves lower forgetting rates after applying to the baselines. It can decrease the forgetting rates of iCaRL by $14.44\%$ , $12.39\%$ , and $9.66\%$ and those of LUCIR by $16.31\%$ , $17.37\%$ , and $17.46\%$ , for 5, 10, and 25 tasks. The first row of Fig. 6 shows CafeBoost improves performance of baselines by a large margin, achieves higher accuracies and prevents catastrophic forgetting better than the prior methods for CIL. See Appendix for more results about the forgetting rates. + +Results on ImageNet. Tab. 1 also reports experimental results of ImageNet-Sub and ImageNet-Full. On ImageNet-Sub, CafeBoost boosts average incremental accuracies by $0.67\% \sim 8.84\%$ over iCaRL and LUCIR baselines, for 5, 20, and 25 tasks, and decreases the corresponding forgetting rates by $6.57\% \sim 26.30\%$ . For example, CafeBoost + +improves average incremental accuracy of LUCIR from $61.44\%$ to $67.74\%$ $(+6.30\%)$ , and reduces forgetting rate from $35.40\%$ to $9.10\%$ $(-26.30\%)$ . The second row of Fig. 6 further validates effectiveness of CafeBoost for all numbers of incremental phases. Additionally, we observe in Fig. 6 CafeBoost has almost no effect on the initial phase, but brings great boost in the following incremental phases. CafeBoost also enhances performance of iCaRL and LUCIR on ImageNet-Full, as shown in Tab. 1. We observe CafeBoost improves accuracies of iCaRL by $1.96\%$ , $1.57\%$ , and those of LUCIR by $2.43\%$ , $3.20\%$ for 5 and 10 incremental phases respectively. + +
SB1020304050
10LUCIR59.1260.2661.7761.6262.91
+Ours55.8762.6166.2168.2269.70
↑-3.25+2.35+4.44+6.60+6.79
5LUCIR54.6256.2056.6557.2956.90
+Ours49.0258.5363.3965.7365.95
↑-5.60+2.33+6.74+8.44+9.05
+ +Table 2. Ablation study on impact of number of classes in the initial phase. $B$ denotes number of classes in the initial phase. $S$ is the number of new classes per incremental step. Experiments are conducted on CIFAR-100 with 20 exemplars per class. The reported values (\%) are average incremental accuracies. $\uparrow$ denotes the increment of values. + +# 5.3. Ablation Studies + +In this section, we conduct experiments to study the effect on our approach of 1) number of classes in initial phase, + +
SRLUCIR+Ours↑
10554.23±0.7864.80±0.36+10.57
1059.98±0.5468.30±0.31+8.32
2063.56±0.5570.10±0.39+6.54
5547.74±0.8462.23±0.67+14.49
1053.22±1.2864.08±0.77+10.86
2057.20±1.1466.56±0.80+9.36
+ +Table 3. Ablation study on impact of number of exemplars. $R$ represents the number of exemplars per class. $S$ is the number of new classes per incremental step. Experiments are conducted on CIFAR-100. The reported values (%) are average incremental accuracies. $\uparrow$ denotes the increment of values. + +
MethodsCIFAR-100ImageNet-Sub
AccuracyForgettingAccuracyForgetting
LUCIR40.10±1.2752.35±1.3160.4725.73
+Ours59.01±0.2223.31±0.7665.1518.05
+ +2) number of exemplars in memory, 3) different classifiers, and 4) coefficient of momentum dictionary update. + +Impact of number of classes in initial phase. We investigate CafeBoost with 10, 20, 30, 40, and 50 classes of initial phase in Tab. 2. CafeBoost can enhance performance when the number of classes in initial phase is larger than 10. The performance improvement increases from negative to positive, along the number for both $S = 10$ and $S = 5$ . When $B$ equates 10, our approach has a negative influence on the baseline. This is because characteristics of classes in the initial phase are insufficient to share for all classes in the dataset. However, when $B >= 20$ , our approach boosts the baseline by a large margin, which shows that it tolerates a broad range of the number $B$ . + +Impact of number of exemplars in memory. We vary the number of exemplars per class and show its effect in Tab. 3. Both the performance of baselines and CafeBoost improve along the increase of exemplar number per class, which is consistently observed for $S = 10$ and $S = 5$ . In addition, the improvement margin decreases along the increase of $R$ . For instance, $+10.57$ for $S = 10$ and $R = 5$ , whereas $+6.54$ for $S = 10$ and $R = 20$ . It is worth noting that CafeBoost does not rely on a replay buffer. Tab. 4 shows CafeBoost outperforms remarkably the baseline on accuracy and forgetting in CIFAR-100 and ImageNet-Sub. + +Impact of different classifiers. We also conduct experiments to research behaviors of CafeBoost combined with three different classifiers. We adopt the classifiers CNN, $k$ -NME, and AME, as in [11]. CNN classifies examples by exploiting one class prototype per class. $k$ -NME uses $k$ samples of each class to represent the class, while AME uti + +Table 4. Comparisons of average incremental accuracies and forgetting rates with no replay data. Experiments are conducted with 5 incremental steps. 20 exemplars are stored per class. + +
MethodiCaRLiCaRL+oursLUCIRLUCIR+ours
CNN553.47±0.1556.53±0.2663.56±0.5570.10±0.39
1049.30±0.4353.75±0.0357.20±1.1466.56±0.80
2545.59±0.7551.18±0.7850.29±1.1462.37±1.17
k-NME562.95±0.5169.40±0.2365.42±0.6769.77±0.35
1056.91±0.6464.70±0.4059.85±1.2066.53±0.72
2551.30±0.9261.6±0.9052.07±0.4362.64±1.30
AME565.02±0.6470.61±0.2966.76±0.8470.54±0.31
1058.27±0.6366.52±0.4661.54±1.3667.77±0.71
2552.69±0.9264.10±0.8954.84±1.7865.10±1.35
+ +Table 5. The average incremental accuracy by using different classifiers (CNN, $k$ -NME, AME) with 20 exemplars per class. Values are evaluated on CIFAR-100 with 5, 10, and 25 incremental steps. + +izes all samples. Tab. 5 shows our approach achieves the best performance when using AME classifier. Moreover, it is noteworthy that our approach enhances performance of baselines using all three classifiers by a large margin in various numbers of tasks. + +Sensitivity of momentum coefficient. Fig. 7 shows our approach is not sensitive to the momentum coefficient of dictionary updating for all three classifiers. We set $\alpha$ to 0.9 for other experiments of our approach. + +![](images/e44ae4c03db8c8a32c9050f240ef0ef953fcdcc0051c70a4ca7dc6d91b445f8c.jpg) +Figure 7. Experiments on sensitivity of momentum coefficient. Average accuracies are evaluated on CIFAR-100 with 5 incremental steps and 20 exemplars per class. + +# 6. Conclusions + +To the best of our knowledge, this work is the first to consider a new type of bias - task-induced bias in a causal perspective, which is derived from special settings of continual learning. We analyzed three continual scenarios in a causal framework, and found that the task-induced bias is particularly detrimental to CIL. Based on this framework, we designed a causal intervention operation and implemented it as a causal debias module by exploiting a powerful attention mechanism. Comprehensive experiments demonstrated our plugin method can boost performance of baselines and outperform several existing strong methods. In the future, we think it is a promising direction to regard the task-induced bias as a problem in CIL and design more exquisite approaches to eliminate it by cutting the corresponding causal path. + +Acknowledgments. This work was supported in part by National Key R&D Program of China (2021ZD0112001) and National Natural Science Foundation of China (No.61831005, 61971095 and 62271119). + +# References + +[1] Ehsan Abbasnejad, Damien Teney, Amin Parvaneh, Javen Shi, and Anton van den Hengel. Counterfactual vision and language learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10041—10051, 2020. 2, 3 +[2] Hongjoon Ahn, Jihwan Kwak, Subin Lim, Hyeonsu Bang, Hyojun Kim, and Taesup Moon. SS-IL: Separated softmax for incremental learning. In Proceedings of the IEEE/CVF Conference on Computer Vision, pages 844–853, 2021. 1 +[3] Arjun Ashok, K. J. Joseph, and Vineeth Balasubramanian. Class-incremental learning with cross-space clustering and controlled transfer. In European Conference on Computer Vision, 2022. 1, 2, 6, 7 +[4] Matthias Delange, Rahaf Aljundi, Marc Masana, Sarah Parisot, Xu Jia, Ales Leonardis, Greg Slabaugh, and Tinne Tuytelaars. A continual learning survey: Defying forgetting in classification tasks. Transactions on Pattern Analysis and Machine Intelligence, 44(7):3366-3385, 2021. 3 +[5] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. ImageNet: A large-scale hierarchical image database. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 248-255, 2009. 6 +[6] Xiang Deng and Zhongfei Zhang. Comprehensive knowledge distillation with causal intervention. In Advances in Neural Information Processing Systems, pages 22158-22170, 2021. 2, 3 +[7] Chen He, Ruiping Wang, and Xilin Chen. A tale of two CILs: The connections between class incremental learning and class imbalanced learning, and beyond. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshop, pages 3554-3564, 2021. 1 +[8] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2016. 5 +[9] Lisa Anne Hendricks, Kaylee Burns, Kate Saenko, Trevor Darrell, and Anna Rohrbach. Women also snowboard: Overcoming bias in captioning models. In European Conference on Computer Vision, 2018. 2 +[10] Geoffrey Hinton, Oriol Vinyals, and Jeff Dean. Distilling the knowledge in a neural network. In ArXiv: 1503.02531, 2015. 1 +[11] Saihui Hou, Xinyu Pan, Chen Change Loy, Zilei Wang, and Dahua Lin. Learning a unified classifier incrementally via rebalancing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 831-839, 2019. 1, 2, 6, 7, 8 +[12] Xinting Hu, Kaihua Tang, Chunyan Miao, Xian-Sheng Hua, and Hanwang Zhang. Distilling causal effect of data in class-incremental learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3956-3965, 2021. 1, 3 +[13] K. J. Joseph, Salman Khan, Fahad Shahbaz Khan, Rao Muhammad Anwer, and Vineeth N. Balasubramanian. + +Energy-based latent aligner for incremental learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7452-7461, 2022. 1 +[14] Minsoo Kang, Jaeyoo Park, and Bohyung Han. Class incremental learning by knowledge distillation with adaptive feature consolidation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16071-16080, 2022. 1 +[15] Alex Kriszhevsky. Learning multiple layers of features from tiny images. In Technical report, 2009. 6 +[16] Zhizhong Li and Derek Hoiem. Learning without forgetting. Transactions on Pattern Analysis and Machine Intelligence, 40(12):2935-2947, 2018. 1, 6, 7 +[17] Chengzhi Mao, Kevin Xia, James Wang, Hao Wang, Junfeng Yang, Elias Bareinboim, and Carl Vondrick. Causal transportability for visual recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7521-7531, 2022. 3 +[18] Michael McCloskey and Neal J. Cohen. Catastrophic interference in connectionist networks: The sequential learning problem. In *Psychology of Learning and Motivation*, volume 24, pages 109–165. Academic Press, 1989. 1 +[19] Ninareh Mehrabi, Fred Morstatter, Nripsuta Saxena, Kristina Lerman, and Aram Galstyan. A survey on bias and fairness in machine learning. ACM Computing Surveys, 54(6):1-35, 2021. 2 +[20] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, Alban Desmaison, Andreas Köpf, Edward Yang, Zach DeVito, Martin Raison, Alykhan Tejani, Sasank Chilamkurthy, Benoit Steiner, Lu Fang, Junjie Bai, and Soumith Chintala. PyTorch: An imperative style, high-performance deep learning library. In Advances in Neural Information Processing Systems, 2019. 6 +[21] Judea Pearl, Madelyn Glymour, and Nicholas P. Jewell. Causal Inference In Statistics: A primer. Wiley, 2016. 3, 4 +[22] Sylvestre Alvise Rebuffi, Alexander Kolesnikov, Georg Sperl, and Christoph H. Lampert. iCaRL: Incremental classifier and representation learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5533-5542, 2017. 1, 2, 6, 7 +[23] Yujun Shi, Kuangqi Zhou, Jian Liang, Zihang Jiang, Jiashi Feng, Philip H. S. Torr, Song Bai, and Vincent Y. F. Tan. Mimicking the oracle: An initial phase decorrelation approach for class incremental learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16722-16731, 2022. 2, 6, 7 +[24] Christian Simon, Piotr Koniusz, and Mehrtash Harandi. On learning the geodesic path for incremental learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1591-1600, 2021. 2, 6, 7 +[25] Qing Sun, Fan Lyu, Fanhua Shang, Wei Feng, and Liang Wan. Exploring example influence in continual learning. In Advances in Neural Information Processing Systems, 2022. 2 + +[26] Rishabh Tiwari, Krishnateja Killamsetty, Rishabh Iyer, and Pradeep Shenoy. GCR: Gradient coreset based replay buffer selection for continual learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 99-108, 2022. 2 +[27] Gido M van de Ven and Andreas S Tolias. Three scenarios for continual learning. In arXiv: 1904.07734, 2019. 1, 2, 3 +[28] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Advances in Neural Information Processing Systems, page 5999-6009, 2017. 2, 5 +[29] Tan Wang, Jianqiang Huang, Hanwang Zhang, and Qianru Sun. Visual commonsense R-CNN. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2020. 2 +[30] Max Welling. Herding dynamical weights to learn. In International Conference on Machine Learning, 2009. 2 +[31] Yue Wu, Yinpeng Chen, Lijuan Wang, Yuancheng Ye, Zicheng Liu, Yandong Guo, and Yun Fu. Large scale incremental learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 374-382, 2019. 1, 2, 6, 7 +[32] Kelvin Xu, Jimmy Lei Ba, Ryan Kiros, Kyunghyun Cho, Aaron Courville, Ruslan Salakhutdinov, Richard S. Zemel, and Yoshua Bengio. Show, attend and tell: Neural image caption generation with visual attention. In International Conference on Machine Learning, pages 2048-2057, 2015. 5 +[33] Shipeng Yan, Jiangwei Xie, and Xuming He. DER: Dynamically expandable representation for class incremental learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3013-3022, 2021. 1 +[34] Xu Yang, Hanwang Zhang, and Jianfei Cai. Deconfounded image captioning: A causal retrospect. Transactions on Pattern Analysis and Machine Intelligence, 2021. 3 +[35] Bowen Zhao, Xi Xiao, Guojun Gan, Bin Zhang, and Shu Tao Xia. Maintaining discrimination and fairness in class incremental learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13205-13214, 2020. 1, 2 \ No newline at end of file diff --git a/2023/CafeBoost_ Causal Feature Boost To Eliminate Task-Induced Bias for Class Incremental Learning/images.zip b/2023/CafeBoost_ Causal Feature Boost To Eliminate Task-Induced Bias for Class Incremental Learning/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..dfbb760ff4c82c89b8db42d8ab1cffcdd53bdef5 --- /dev/null +++ b/2023/CafeBoost_ Causal Feature Boost To Eliminate Task-Induced Bias for Class Incremental Learning/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6240a77a4d19b2c9f7c54b8efd737b6b067f727f6d91b8a6f8b60c424afabea1 +size 634114 diff --git a/2023/CafeBoost_ Causal Feature Boost To Eliminate Task-Induced Bias for Class Incremental Learning/layout.json b/2023/CafeBoost_ Causal Feature Boost To Eliminate Task-Induced Bias for Class Incremental Learning/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..61fd25262ed7eddd7801cb77b9950a77360e7573 --- /dev/null +++ b/2023/CafeBoost_ Causal Feature Boost To Eliminate Task-Induced Bias for Class Incremental Learning/layout.json @@ -0,0 +1,12071 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 67, + 102, + 527, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 102, + 527, + 140 + ], + "spans": [ + { + "bbox": [ + 67, + 102, + 527, + 140 + ], + "type": "text", + "content": "CafeBoost: Causal Feature Boost to Eliminate Task-Induced Bias for Class Incremental Learning" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 142, + 161, + 199, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 142, + 161, + 199, + 175 + ], + "spans": [ + { + "bbox": [ + 142, + 161, + 199, + 175 + ], + "type": "text", + "content": "Benliu Qiu" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 223, + 161, + 294, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 223, + 161, + 294, + 175 + ], + "spans": [ + { + "bbox": [ + 223, + 161, + 294, + 175 + ], + "type": "text", + "content": "Hongliang Li*" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 312, + 161, + 370, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 161, + 370, + 175 + ], + "spans": [ + { + "bbox": [ + 312, + 161, + 370, + 175 + ], + "type": "text", + "content": "Haitao Wen" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 392, + 161, + 449, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 392, + 161, + 449, + 175 + ], + "spans": [ + { + "bbox": [ + 392, + 161, + 449, + 175 + ], + "type": "text", + "content": "Heqian Qiu" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 143, + 176, + 219, + 189 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 176, + 219, + 189 + ], + "spans": [ + { + "bbox": [ + 143, + 176, + 219, + 189 + ], + "type": "text", + "content": "Lanxiao Wang*" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 237, + 176, + 309, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 237, + 176, + 309, + 190 + ], + "spans": [ + { + "bbox": [ + 237, + 176, + 309, + 190 + ], + "type": "text", + "content": "Fanman Meng" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 330, + 176, + 388, + 189 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 330, + 176, + 388, + 189 + ], + "spans": [ + { + "bbox": [ + 330, + 176, + 388, + 189 + ], + "type": "text", + "content": "Qingbo Wu" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 410, + 176, + 449, + 188 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 410, + 176, + 449, + 188 + ], + "spans": [ + { + "bbox": [ + 410, + 176, + 449, + 188 + ], + "type": "text", + "content": "Lili Pan" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 114, + 190, + 479, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 114, + 190, + 479, + 203 + ], + "spans": [ + { + "bbox": [ + 114, + 190, + 479, + 203 + ], + "type": "text", + "content": "University of Electronic Science and Technology of China, Chengdu, China" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 139, + 205, + 453, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 139, + 205, + 453, + 217 + ], + "spans": [ + { + "bbox": [ + 139, + 205, + 453, + 217 + ], + "type": "text", + "content": "{qbenliu, haitaowen, hqqiu, lanxiao.wang}@std.uestc.edu.cn" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 140, + 220, + 451, + 232 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 140, + 220, + 451, + 232 + ], + "spans": [ + { + "bbox": [ + 140, + 220, + 451, + 232 + ], + "type": "text", + "content": "{h1li, fmmeng, qbwu}@uestc.edu.cn panlili8255@gmail.com" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 143, + 258, + 192, + 271 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 258, + 192, + 271 + ], + "spans": [ + { + "bbox": [ + 143, + 258, + 192, + 271 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 46, + 275, + 290, + 550 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 275, + 290, + 550 + ], + "spans": [ + { + "bbox": [ + 46, + 275, + 290, + 550 + ], + "type": "text", + "content": "Continual learning requires a model to incrementally learn a sequence of tasks and aims to predict well on all the learned tasks so far, which notoriously suffers from the catastrophic forgetting problem. In this paper, we find a new type of bias appearing in continual learning, coined as task-induced bias. We place continual learning into a causal framework, based on which we find the task-induced bias is reduced naturally by two underlying mechanisms in task and domain incremental learning. However, these mechanisms do not exist in class incremental learning (CIL), in which each task contains a unique subset of classes. To eliminate the task-induced bias in CIL, we devise a causal intervention operation so as to cut off the causal path that causes the task-induced bias, and then implement it as a causal debias module that transforms biased features into unbiased ones. In addition, we propose a training pipeline to incorporate the novel module into existing methods and jointly optimize the entire architecture. Our overall approach does not rely on data replay, and is simple and convenient to plug into existing methods. Extensive empirical study on CIFAR-100 and ImageNet shows that our approach can improve accuracy and reduce forgetting of well-established methods by a large margin." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 47, + 562, + 128, + 574 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 562, + 128, + 574 + ], + "spans": [ + { + "bbox": [ + 47, + 562, + 128, + 574 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 46, + 576, + 287, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 576, + 287, + 696 + ], + "spans": [ + { + "bbox": [ + 46, + 576, + 287, + 696 + ], + "type": "text", + "content": "Deep learning has been widely applied in many fields like computer vision, social media analytics, natural language processing, etc. The standard paradigm of deep learning is to train models on a prepared dataset, of which all data are accessible in the whole training stage. However, in most real-world applications, a sequence of tasks arrives incrementally, which requires models to learn continuously from a new task, namely continual learning [27]. Despite fine-tuning achieves this goal somewhat, such naive method forgets obtained knowledge of previous tasks after learning" + } + ] + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 313, + 255, + 529, + 328 + ], + "blocks": [ + { + "bbox": [ + 313, + 255, + 529, + 328 + ], + "lines": [ + { + "bbox": [ + 313, + 255, + 529, + 328 + ], + "spans": [ + { + "bbox": [ + 313, + 255, + 529, + 328 + ], + "type": "image", + "image_path": "611a0457922a71cbe8aa8c7ad14ff5e4cf77871ed1591514252685cb139a7c54.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 388, + 325, + 450, + 334 + ], + "lines": [ + { + "bbox": [ + 388, + 325, + 450, + 334 + ], + "spans": [ + { + "bbox": [ + 388, + 325, + 450, + 334 + ], + "type": "text", + "content": "(a) Illustration of CIL" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 315, + 337, + 538, + 388 + ], + "blocks": [ + { + "bbox": [ + 315, + 337, + 538, + 388 + ], + "lines": [ + { + "bbox": [ + 315, + 337, + 538, + 388 + ], + "spans": [ + { + "bbox": [ + 315, + 337, + 538, + 388 + ], + "type": "image", + "image_path": "22da54b8de582da85f61a1c55395d23cc884fa1b07d0a717f8cbebab9c3e8dd7.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 354, + 388, + 490, + 396 + ], + "lines": [ + { + "bbox": [ + 354, + 388, + 490, + 396 + ], + "spans": [ + { + "bbox": [ + 354, + 388, + 490, + 396 + ], + "type": "text", + "content": "(b) Task-induced bias between class features" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 312, + 396, + 531, + 452 + ], + "blocks": [ + { + "bbox": [ + 312, + 396, + 531, + 452 + ], + "lines": [ + { + "bbox": [ + 312, + 396, + 531, + 452 + ], + "spans": [ + { + "bbox": [ + 312, + 396, + 531, + 452 + ], + "type": "image", + "image_path": "001c7710a5d3ca6c868622763ceebbd82037903c1918bd266a6171e7d2a197fb.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 351, + 453, + 499, + 461 + ], + "lines": [ + { + "bbox": [ + 351, + 453, + 499, + 461 + ], + "spans": [ + { + "bbox": [ + 351, + 453, + 499, + 461 + ], + "type": "text", + "content": "(c) Task-induced bias between images and labels" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 304, + 462, + 545, + 550 + ], + "lines": [ + { + "bbox": [ + 304, + 462, + 545, + 550 + ], + "spans": [ + { + "bbox": [ + 304, + 462, + 545, + 550 + ], + "type": "text", + "content": "Figure 1. Illustration of CIL, task-induced bias and their detrimental effects. (a) Illustration of CIL. (b) Task-induced bias between class features. Two classes in a task are correlated by a task identifier. (c) Task-induced bias between images and labels. Images and labels in a task are correlated by a task identifier. In (b)(c), the first column describes perspectives of task-induced bias, the second shows the underlying causal effect causing task-induced bias, and the last illustrates adverse influence of task-induced bias." + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + } + ], + "index": 22 + }, + { + "bbox": [ + 306, + 558, + 511, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 558, + 511, + 569 + ], + "spans": [ + { + "bbox": [ + 306, + 558, + 511, + 569 + ], + "type": "text", + "content": "a new task, dubbed as catastrophic forgetting [18]." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 304, + 570, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 546, + 713 + ], + "type": "text", + "content": "Much effort has been devoted as a remedy for alleviating this problem in literature [2, 3, 11-14, 16, 33]. One general direction is to preserve knowledge representations of previous tasks by knowledge distillation [10] and by storing a few observed exemplars for replay [22]. [11] discovers the class imbalance between old classes and new ones is a key challenge of large-scale applications in class incremental learning (CIL), in which classes of a new task never appear in the previous tasks, as illustrated in Fig. 1 (a). [2,7,31,35] all follow this direction and propose effective methods to overcome the class imbalance. However, these methods ignore bias problems in continual learning. In this work, we" + } + ] + } + ], + "index": 26 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 702, + 139, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 702, + 139, + 712 + ], + "spans": [ + { + "bbox": [ + 58, + 702, + 139, + 712 + ], + "type": "text", + "content": "*Corresponding authors." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "16016" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 287, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 287, + 108 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 287, + 108 + ], + "type": "text", + "content": "consider the problem of CIL in a novel perspective of bias. We think that catastrophic forgetting is the reflection of a new-type bias, coined as task-induced bias." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 109, + 288, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 109, + 288, + 396 + ], + "spans": [ + { + "bbox": [ + 47, + 109, + 288, + 396 + ], + "type": "text", + "content": "Deep neural networks are easily misled by language bias [1] and vision bias [6, 9, 29], which are both derived from spurious correlations of two or more items in different positions of a sentence or an image. In contrast to these two bias, the task-induced bias occurs across different images in a common task. In CIL, data of specific classes are accessible only in a certain task if replay buffer is not applied. A deep neural network is capable of learning correlations among images of different categories or between images and labels as possible as it can, which includes a few spurious correlations, namely biases, leading to a limited generalization. On the one hand, a portion of these image correlations can be shared beneficially in this task, but these benefits cannot share among different tasks. For example, in Fig. 1 (b), color is sufficient to discriminate elephants and leopards in a specific task since elephants are grey while leopards are orange, and thus deep models have no motivation to capture more complex characteristic, like stripe. This is detrimental to the feature extractor because it captures a few details that cannot generalize among all tasks. On the other hand, as shown in Fig. 1 (c), a classifier will be biased to new classes in the current task [11, 31, 35], because the classes in the training data are all the new classes and to predict new classes is more possible to be correct." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 396, + 287, + 671 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 396, + 287, + 671 + ], + "spans": [ + { + "bbox": [ + 47, + 396, + 287, + 671 + ], + "type": "text", + "content": "To better understand the task-induced bias in continual learning, we model three continual learning scenarios [27] into a common causal inference framework, which merely contains three necessary variables: image, label, and task identifier. Thanks to the framework, we discover that the task-induced bias is minor in task incremental and domain incremental learning due to innate mechanisms that cut off an inappropriate causal path, which partially explain why the two scenarios are easier to tackle than CIL [27]. For the existence of the causal path in CIL, we seek an active causal intervention to close the path. To implement the intervention, we devise a simple but effective causal debias module, which is established on attention mechanisms [28]. Subsequently, we incorporate this module into existing CIL architectures and propose a training pipeline to optimize it. Our overall approach transforms causally biased features into unbiased ones, and hence we refer to it as Casual Feature Boost (CafeBoost). Finally, we systematically compare accuracy and forgetting of well-established methods with ours in various settings. Experimental results show that CafeBoost yields significant and consistent performance boost for existing methods in CIFAR-100 and ImageNet by a large margin " + }, + { + "bbox": [ + 47, + 396, + 287, + 671 + ], + "type": "inline_equation", + "content": "(0.67\\% \\sim 12.08\\%" + }, + { + "bbox": [ + 47, + 396, + 287, + 671 + ], + "type": "text", + "content": " on accuracy)." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 674, + 134, + 687 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 674, + 134, + 687 + ], + "spans": [ + { + "bbox": [ + 47, + 674, + 134, + 687 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "content": "Continual learning. Existing CIL works mostly aim to maintain knowledge representations of previous tasks, in-" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 315, + 76, + 538, + 242 + ], + "blocks": [ + { + "bbox": [ + 315, + 76, + 538, + 242 + ], + "lines": [ + { + "bbox": [ + 315, + 76, + 538, + 242 + ], + "spans": [ + { + "bbox": [ + 315, + 76, + 538, + 242 + ], + "type": "image", + "image_path": "78f4c3a1b646b817171e8d8de5ba734e13f29b73a66eaabb932008587fc81506.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 244, + 545, + 310 + ], + "lines": [ + { + "bbox": [ + 305, + 244, + 545, + 310 + ], + "spans": [ + { + "bbox": [ + 305, + 244, + 545, + 310 + ], + "type": "text", + "content": "Figure 2. Overview of Casual Feature Boost (CafeBoost). Green blocks are learnable, whereas white blocks are fixed. Conv block is updated with a smaller learning rate than Causal block. The training objectives including distillation loss, margin ranking loss, etc., are omitted in this figure, since they depend on baselines. Whether to use replay buffer " + }, + { + "bbox": [ + 305, + 244, + 545, + 310 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 305, + 244, + 545, + 310 + ], + "type": "text", + "content": " depends on specific baselines." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 318, + 546, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 318, + 546, + 605 + ], + "spans": [ + { + "bbox": [ + 304, + 318, + 546, + 605 + ], + "type": "text", + "content": "including replay-based methods and distillation-based methods. Replay-based methods [22, 25, 26] store a few examples to represent previous data and replay the stored data in the current training stage. The purpose of relay-based methods is to select the most representative samples of the entire training dataset. iCaRL [22] uses herding strategy [30] to choose typical exemplars for future training. [25] calculates example influence of stability and plasticity for every example in validation sets, based on which it selects examples for replay. Distillation-based methods [3,11,23,24] focus on transferring learned knowledge of a previous model to the current one. LUCIR [11] combines cosine normalization, less-forget constraint, and inter-class separation to mitigate the forgetting caused by class imbalance. GeoDL [24] utilizes geodesic path for knowledge distillation which considers the gradual change between subspaces of tasks. CSCCT [3] encourages features of each class to locate in an appropriate position of the feature space, and utilizes inter-class relationships to transfer inter-class knowledge positively or negatively. CwD [23] scatters uniformly representations of each class in the initial task, which significantly benefits performance in subsequent tasks. These two types of methods are usually incorporated together to achieve better performance." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": "Spurious correlation. Spurious correlation is also known as bias. [19] has extensively investigated the bias problem in machine learning. Many methods [1,9,29] have been proposed to eliminate bias in their specific applications. [1] introduces counterfactual in the training to eliminate bias in visual question answering datasets. [9] proposes a new Equalizer model to force a model look at a person instead of inferring a person's gender by context cues. [29] applies causal intervention to alleviate vision bias" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16017" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 195, + 83 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 195, + 83 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 195, + 83 + ], + "type": "text", + "content": "among detected objects in an image." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 86, + 288, + 229 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 86, + 288, + 229 + ], + "spans": [ + { + "bbox": [ + 46, + 86, + 288, + 229 + ], + "type": "text", + "content": "Causal inference. Causal inference [21] aims to establish reasonable causal graphs for specific applications, then analyze problems and propose solutions based on the established causal graph. Causal intervention and counterfactual process are two powerful tools of causal inference. These two tools have been ubiquitously applied in multiple applications, such as out-of-distribution recognition [17], visual question answering [1], image captioning [34], knowledge distillation [6], etc. Recently, [12] builds a causal view to analyze the cause of forgetting in class incremental learning, and designs an anti-forgetting technique so as to distill causal effect of data without replay." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 239, + 124, + 252 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 239, + 124, + 252 + ], + "spans": [ + { + "bbox": [ + 47, + 239, + 124, + 252 + ], + "type": "text", + "content": "3. Preliminary" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 259, + 204, + 272 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 259, + 204, + 272 + ], + "spans": [ + { + "bbox": [ + 47, + 259, + 204, + 272 + ], + "type": "text", + "content": "3.1. Continual learning scenarios" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 280, + 287, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 280, + 287, + 434 + ], + "spans": [ + { + "bbox": [ + 46, + 280, + 287, + 434 + ], + "type": "text", + "content": "Continual learning can be categorized into three scenarios [27] in the light of whether a task identifier is given in a test phase and whether a model is required to infer which task it meets. If the task identifier is available during both training and test phases, this scenario is task incremental. If the task identifier is accessible in training phase but it is only necessary for a model to solve the task at hand, this scenario is domain incremental. Similar to the domain incremental scenario, the class incremental scenario doesn't provide task identifiers in test phase, but it requires a model to infer the current task identifier. Generally speaking, the class incremental scenario is the most difficult among the three, whereas the task incremental scenario is the easiest [27]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 437, + 287, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 437, + 287, + 581 + ], + "spans": [ + { + "bbox": [ + 46, + 437, + 287, + 581 + ], + "type": "text", + "content": "To formulate continual learning, we assume a sequence of " + }, + { + "bbox": [ + 46, + 437, + 287, + 581 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 46, + 437, + 287, + 581 + ], + "type": "text", + "content": " tasks is given as " + }, + { + "bbox": [ + 46, + 437, + 287, + 581 + ], + "type": "inline_equation", + "content": "\\{\\mathcal{T}_0,\\mathcal{T}_1,\\dots,\\mathcal{T}_{T - 1}\\}" + }, + { + "bbox": [ + 46, + 437, + 287, + 581 + ], + "type": "text", + "content": ", each with data " + }, + { + "bbox": [ + 46, + 437, + 287, + 581 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_t,t = 0,1,\\ldots ,T - 1" + }, + { + "bbox": [ + 46, + 437, + 287, + 581 + ], + "type": "text", + "content": ". Every sample in " + }, + { + "bbox": [ + 46, + 437, + 287, + 581 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_t" + }, + { + "bbox": [ + 46, + 437, + 287, + 581 + ], + "type": "text", + "content": " is a tuple of an image " + }, + { + "bbox": [ + 46, + 437, + 287, + 581 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_i" + }, + { + "bbox": [ + 46, + 437, + 287, + 581 + ], + "type": "text", + "content": " and a corresponding label " + }, + { + "bbox": [ + 46, + 437, + 287, + 581 + ], + "type": "inline_equation", + "content": "y_{i}" + }, + { + "bbox": [ + 46, + 437, + 287, + 581 + ], + "type": "text", + "content": ", and hence we can denote the data of task " + }, + { + "bbox": [ + 46, + 437, + 287, + 581 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 437, + 287, + 581 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 46, + 437, + 287, + 581 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_t = \\{(\\mathbf{x}_i,y_i)_{i = 1}^{N_t}\\}" + }, + { + "bbox": [ + 46, + 437, + 287, + 581 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 46, + 437, + 287, + 581 + ], + "type": "inline_equation", + "content": "N_{t}" + }, + { + "bbox": [ + 46, + 437, + 287, + 581 + ], + "type": "text", + "content": " is the number of samples, " + }, + { + "bbox": [ + 46, + 437, + 287, + 581 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_i\\in \\mathcal{X}_t" + }, + { + "bbox": [ + 46, + 437, + 287, + 581 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 437, + 287, + 581 + ], + "type": "inline_equation", + "content": "y_{i}\\in \\mathcal{Y}_{t}" + }, + { + "bbox": [ + 46, + 437, + 287, + 581 + ], + "type": "text", + "content": ". Replay-based continual learning methods require a memory buffer " + }, + { + "bbox": [ + 46, + 437, + 287, + 581 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 46, + 437, + 287, + 581 + ], + "type": "text", + "content": " to store partial data of previous tasks. For the task " + }, + { + "bbox": [ + 46, + 437, + 287, + 581 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_0" + }, + { + "bbox": [ + 46, + 437, + 287, + 581 + ], + "type": "text", + "content": ", only data " + }, + { + "bbox": [ + 46, + 437, + 287, + 581 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_0" + }, + { + "bbox": [ + 46, + 437, + 287, + 581 + ], + "type": "text", + "content": " can be used for training, but for the task " + }, + { + "bbox": [ + 46, + 437, + 287, + 581 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_t,t > 0" + }, + { + "bbox": [ + 46, + 437, + 287, + 581 + ], + "type": "text", + "content": ", we can train models using data " + }, + { + "bbox": [ + 46, + 437, + 287, + 581 + ], + "type": "inline_equation", + "content": "\\mathcal{M}\\cup \\mathcal{D}_t" + }, + { + "bbox": [ + 46, + 437, + 287, + 581 + ], + "type": "text", + "content": ". For the task incremental scenario, task identifier " + }, + { + "bbox": [ + 46, + 437, + 287, + 581 + ], + "type": "inline_equation", + "content": "\\mathcal{T}" + }, + { + "bbox": [ + 46, + 437, + 287, + 581 + ], + "type": "text", + "content": " is given as an input to models in test phase but not for the other two scenarios." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 582, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 582, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 582, + 288, + 714 + ], + "type": "text", + "content": "After the above formulation, we elucidate distinctions of three incremental scenarios in a formal style [4]. Let " + }, + { + "bbox": [ + 46, + 582, + 288, + 714 + ], + "type": "inline_equation", + "content": "P(\\cdot)" + }, + { + "bbox": [ + 46, + 582, + 288, + 714 + ], + "type": "text", + "content": " denote a marginal distribution. The common ground of three scenarios is " + }, + { + "bbox": [ + 46, + 582, + 288, + 714 + ], + "type": "inline_equation", + "content": "P(\\mathcal{X}_t) \\neq P(\\mathcal{X}_{t + 1})" + }, + { + "bbox": [ + 46, + 582, + 288, + 714 + ], + "type": "text", + "content": " for all " + }, + { + "bbox": [ + 46, + 582, + 288, + 714 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 582, + 288, + 714 + ], + "type": "text", + "content": ". Their distinctions lie in label set " + }, + { + "bbox": [ + 46, + 582, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{Y}_t" + }, + { + "bbox": [ + 46, + 582, + 288, + 714 + ], + "type": "text", + "content": " and its marginal distribution " + }, + { + "bbox": [ + 46, + 582, + 288, + 714 + ], + "type": "inline_equation", + "content": "P(\\mathcal{Y}_t)" + }, + { + "bbox": [ + 46, + 582, + 288, + 714 + ], + "type": "text", + "content": ". For task incremental learning, " + }, + { + "bbox": [ + 46, + 582, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{Y}_t \\neq \\mathcal{Y}_{t + 1}" + }, + { + "bbox": [ + 46, + 582, + 288, + 714 + ], + "type": "text", + "content": ". For domain incremental learning, " + }, + { + "bbox": [ + 46, + 582, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{Y}_t = \\mathcal{Y}_{t + 1}" + }, + { + "bbox": [ + 46, + 582, + 288, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 582, + 288, + 714 + ], + "type": "inline_equation", + "content": "P(\\mathcal{Y}_t) = P(\\mathcal{Y}_{t + 1})" + }, + { + "bbox": [ + 46, + 582, + 288, + 714 + ], + "type": "text", + "content": ". These two scenarios have no requirement on inferring task identifiers. Conversely, the class incremental learning that defines " + }, + { + "bbox": [ + 46, + 582, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{Y}_t \\subset \\mathcal{Y}_{t + 1}" + }, + { + "bbox": [ + 46, + 582, + 288, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 582, + 288, + 714 + ], + "type": "inline_equation", + "content": "P(\\mathcal{Y}_t) \\neq P(\\mathcal{Y}_{t + 1})" + }, + { + "bbox": [ + 46, + 582, + 288, + 714 + ], + "type": "text", + "content": ", requires models to infer which label set is used for an arbitrary task." + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 332, + 70, + 391, + 129 + ], + "blocks": [ + { + "bbox": [ + 332, + 70, + 391, + 129 + ], + "lines": [ + { + "bbox": [ + 332, + 70, + 391, + 129 + ], + "spans": [ + { + "bbox": [ + 332, + 70, + 391, + 129 + ], + "type": "image", + "image_path": "f08dc029d1598dc6b2016c5217f8a1b49735b2e71e782590799b520563fc7740.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 399, + 72, + 455, + 128 + ], + "blocks": [ + { + "bbox": [ + 399, + 72, + 455, + 128 + ], + "lines": [ + { + "bbox": [ + 399, + 72, + 455, + 128 + ], + "spans": [ + { + "bbox": [ + 399, + 72, + 455, + 128 + ], + "type": "image", + "image_path": "48cfaa10052886a2460f9cc3f076be9842e0fe7d12daf515695bc285c43b6246.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 465, + 72, + 522, + 128 + ], + "blocks": [ + { + "bbox": [ + 465, + 72, + 522, + 128 + ], + "lines": [ + { + "bbox": [ + 465, + 72, + 522, + 128 + ], + "spans": [ + { + "bbox": [ + 465, + 72, + 522, + 128 + ], + "type": "image", + "image_path": "1b55d7506b2fb981f50787d5f8ee3cfa13698623a426e647e1022a83ec410f0c.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 331, + 130, + 390, + 185 + ], + "blocks": [ + { + "bbox": [ + 331, + 130, + 390, + 185 + ], + "lines": [ + { + "bbox": [ + 331, + 130, + 390, + 185 + ], + "spans": [ + { + "bbox": [ + 331, + 130, + 390, + 185 + ], + "type": "image", + "image_path": "533436adbcd9fc65d1aa5fe309363ae21e3f3f0da36b2b5fbd7303d97f695ffc.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 186, + 545, + 308 + ], + "lines": [ + { + "bbox": [ + 304, + 186, + 545, + 308 + ], + "spans": [ + { + "bbox": [ + 304, + 186, + 545, + 308 + ], + "type": "text", + "content": "Figure 3. Causal graphs of supervised learning and continual learning. Gray nodes means unconditioned variables. (a) Causal graphs of supervised learning in training and test phase. (b) Causal graphs of continual learning. The red task variable depends on specific continual scenarios. (c) Causal path between images and labels in task incremental learning. The path " + }, + { + "bbox": [ + 304, + 186, + 545, + 308 + ], + "type": "inline_equation", + "content": "X \\gets T \\to Y" + }, + { + "bbox": [ + 304, + 186, + 545, + 308 + ], + "type": "text", + "content": " is impeded because variable " + }, + { + "bbox": [ + 304, + 186, + 545, + 308 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 304, + 186, + 545, + 308 + ], + "type": "text", + "content": " is conditioned. (d) Causal path between images and labels in domain incremental learning. The path " + }, + { + "bbox": [ + 304, + 186, + 545, + 308 + ], + "type": "inline_equation", + "content": "X \\gets T \\to Y" + }, + { + "bbox": [ + 304, + 186, + 545, + 308 + ], + "type": "text", + "content": " is cut off because of " + }, + { + "bbox": [ + 304, + 186, + 545, + 308 + ], + "type": "inline_equation", + "content": "T \\nrightarrow Y" + }, + { + "bbox": [ + 304, + 186, + 545, + 308 + ], + "type": "text", + "content": ". (e) Causal path between images and labels in class incremental learning. The path " + }, + { + "bbox": [ + 304, + 186, + 545, + 308 + ], + "type": "inline_equation", + "content": "X \\gets T \\to Y" + }, + { + "bbox": [ + 304, + 186, + 545, + 308 + ], + "type": "text", + "content": " causes task-induced bias." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 398, + 130, + 458, + 184 + ], + "blocks": [ + { + "bbox": [ + 398, + 130, + 458, + 184 + ], + "lines": [ + { + "bbox": [ + 398, + 130, + 458, + 184 + ], + "spans": [ + { + "bbox": [ + 398, + 130, + 458, + 184 + ], + "type": "image", + "image_path": "134c16f3195c1b63fac8a792eb263ee83fbe35bd61b7905d52b3a8d4ff1f7e74.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 464, + 130, + 524, + 184 + ], + "blocks": [ + { + "bbox": [ + 464, + 130, + 524, + 184 + ], + "lines": [ + { + "bbox": [ + 464, + 130, + 524, + 184 + ], + "spans": [ + { + "bbox": [ + 464, + 130, + 524, + 184 + ], + "type": "image", + "image_path": "ec6fb4e487aa0966b00f3427d943d74afcc567e0db56dbd8d8d0a687938a3e6a.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 315, + 495, + 327 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 315, + 495, + 327 + ], + "spans": [ + { + "bbox": [ + 306, + 315, + 495, + 327 + ], + "type": "text", + "content": "3.2. Causal graphs in continual learning" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 330, + 545, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 330, + 545, + 413 + ], + "spans": [ + { + "bbox": [ + 304, + 330, + 545, + 413 + ], + "type": "text", + "content": "A causal graph is a directed acyclic graph, of which nodes denote variables and directed edges denote the causality between two variables. We treat images, labels and task identifiers as nodes in our causal graphs. Although only three nodes are incorporated into our causal graph, it is powerful to reveal underlying characteristics of continual learning as we will show in Sec. 4." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 414, + 545, + 473 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 414, + 545, + 473 + ], + "spans": [ + { + "bbox": [ + 304, + 414, + 545, + 473 + ], + "type": "text", + "content": "We denote an image as " + }, + { + "bbox": [ + 304, + 414, + 545, + 473 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 304, + 414, + 545, + 473 + ], + "type": "text", + "content": ", a label as " + }, + { + "bbox": [ + 304, + 414, + 545, + 473 + ], + "type": "inline_equation", + "content": "Y" + }, + { + "bbox": [ + 304, + 414, + 545, + 473 + ], + "type": "text", + "content": " and a task identifier as " + }, + { + "bbox": [ + 304, + 414, + 545, + 473 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 304, + 414, + 545, + 473 + ], + "type": "text", + "content": " in Fig. 3. The edges are listed below: " + }, + { + "bbox": [ + 304, + 414, + 545, + 473 + ], + "type": "inline_equation", + "content": "X \\to Y" + }, + { + "bbox": [ + 304, + 414, + 545, + 473 + ], + "type": "text", + "content": " means label " + }, + { + "bbox": [ + 304, + 414, + 545, + 473 + ], + "type": "inline_equation", + "content": "Y" + }, + { + "bbox": [ + 304, + 414, + 545, + 473 + ], + "type": "text", + "content": " is deduced from image " + }, + { + "bbox": [ + 304, + 414, + 545, + 473 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 304, + 414, + 545, + 473 + ], + "type": "text", + "content": "; " + }, + { + "bbox": [ + 304, + 414, + 545, + 473 + ], + "type": "inline_equation", + "content": "T \\to X" + }, + { + "bbox": [ + 304, + 414, + 545, + 473 + ], + "type": "text", + "content": " denotes that task identifier " + }, + { + "bbox": [ + 304, + 414, + 545, + 473 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 304, + 414, + 545, + 473 + ], + "type": "text", + "content": " have a causal influence to image " + }, + { + "bbox": [ + 304, + 414, + 545, + 473 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 304, + 414, + 545, + 473 + ], + "type": "text", + "content": "; " + }, + { + "bbox": [ + 304, + 414, + 545, + 473 + ], + "type": "inline_equation", + "content": "T \\to Y" + }, + { + "bbox": [ + 304, + 414, + 545, + 473 + ], + "type": "text", + "content": " denotes task identifier " + }, + { + "bbox": [ + 304, + 414, + 545, + 473 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 304, + 414, + 545, + 473 + ], + "type": "text", + "content": " causally affect label " + }, + { + "bbox": [ + 304, + 414, + 545, + 473 + ], + "type": "inline_equation", + "content": "Y" + }, + { + "bbox": [ + 304, + 414, + 545, + 473 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 474, + 546, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 474, + 546, + 665 + ], + "spans": [ + { + "bbox": [ + 304, + 474, + 546, + 665 + ], + "type": "text", + "content": "For supervised learning in Fig. 3 (a), its causal graph only contains an image " + }, + { + "bbox": [ + 304, + 474, + 546, + 665 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 304, + 474, + 546, + 665 + ], + "type": "text", + "content": ", a label " + }, + { + "bbox": [ + 304, + 474, + 546, + 665 + ], + "type": "inline_equation", + "content": "Y" + }, + { + "bbox": [ + 304, + 474, + 546, + 665 + ], + "type": "text", + "content": " and an causal path " + }, + { + "bbox": [ + 304, + 474, + 546, + 665 + ], + "type": "inline_equation", + "content": "X \\to Y" + }, + { + "bbox": [ + 304, + 474, + 546, + 665 + ], + "type": "text", + "content": " that represents a model to capture the causality between two variables. In training phase of supervised learning, both image " + }, + { + "bbox": [ + 304, + 474, + 546, + 665 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 304, + 474, + 546, + 665 + ], + "type": "text", + "content": " and label " + }, + { + "bbox": [ + 304, + 474, + 546, + 665 + ], + "type": "inline_equation", + "content": "Y" + }, + { + "bbox": [ + 304, + 474, + 546, + 665 + ], + "type": "text", + "content": " are accessible and we utilize them to learn the causal effect into the model. But in test phase, an image " + }, + { + "bbox": [ + 304, + 474, + 546, + 665 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 304, + 474, + 546, + 665 + ], + "type": "text", + "content": " and the causal path are known and " + }, + { + "bbox": [ + 304, + 474, + 546, + 665 + ], + "type": "inline_equation", + "content": "Y" + }, + { + "bbox": [ + 304, + 474, + 546, + 665 + ], + "type": "text", + "content": " is the label to be inferred. For three continual learning scenarios, images, labels and task identifiers are all given in training phase. As a result, they have the same causal graph for training in Fig. 3 (b). We illustrate their distinctions in test phase in Fig. 3 (b). Due to the involvement of the task identifier " + }, + { + "bbox": [ + 304, + 474, + 546, + 665 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 304, + 474, + 546, + 665 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 474, + 546, + 665 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 304, + 474, + 546, + 665 + ], + "type": "text", + "content": " may have an indirect effect on " + }, + { + "bbox": [ + 304, + 474, + 546, + 665 + ], + "type": "inline_equation", + "content": "Y" + }, + { + "bbox": [ + 304, + 474, + 546, + 665 + ], + "type": "text", + "content": " through the causal path " + }, + { + "bbox": [ + 304, + 474, + 546, + 665 + ], + "type": "inline_equation", + "content": "X \\gets T \\to Y" + }, + { + "bbox": [ + 304, + 474, + 546, + 665 + ], + "type": "text", + "content": ", solely dependent on " + }, + { + "bbox": [ + 304, + 474, + 546, + 665 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 304, + 474, + 546, + 665 + ], + "type": "text", + "content": ". We will explain it in Sec. 4. See Appendix for additional explanation of causal graphs." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 306, + 670, + 389, + 682 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 670, + 389, + 682 + ], + "spans": [ + { + "bbox": [ + 306, + 670, + 389, + 682 + ], + "type": "text", + "content": "4. Methodology" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "content": "In this section, we introduce the task-induced bias through the causal framework, a causal intervention for" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16018" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 95 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 95 + ], + "type": "text", + "content": "CIL, its corresponding implementation, and the entire pipeline of our approach." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 97, + 286, + 109 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 97, + 286, + 109 + ], + "spans": [ + { + "bbox": [ + 47, + 97, + 286, + 109 + ], + "type": "text", + "content": "4.1. Task-induced bias in three continual scenarios" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 110, + 286, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 110, + 286, + 168 + ], + "spans": [ + { + "bbox": [ + 46, + 110, + 286, + 168 + ], + "type": "text", + "content": "According to the causal graphs, we can explain why task incremental learning is the most effortless scenario among the three and why class incremental learning is the most challenging. We think that this phenomenon is caused by task-induced bias." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 169, + 286, + 358 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 169, + 286, + 358 + ], + "spans": [ + { + "bbox": [ + 46, + 169, + 286, + 358 + ], + "type": "text", + "content": "In Fig. 3 (b), the causal path " + }, + { + "bbox": [ + 46, + 169, + 286, + 358 + ], + "type": "inline_equation", + "content": "X \\leftarrow T \\rightarrow Y" + }, + { + "bbox": [ + 46, + 169, + 286, + 358 + ], + "type": "text", + "content": " is of a fork structure, in which two kid nodes are independent conditioned on the parent. [21] points out that if " + }, + { + "bbox": [ + 46, + 169, + 286, + 358 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 46, + 169, + 286, + 358 + ], + "type": "text", + "content": " is unconditioned, the causal influence of " + }, + { + "bbox": [ + 46, + 169, + 286, + 358 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 46, + 169, + 286, + 358 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 46, + 169, + 286, + 358 + ], + "type": "inline_equation", + "content": "Y" + }, + { + "bbox": [ + 46, + 169, + 286, + 358 + ], + "type": "text", + "content": " exists, causing the task-induced bias. We give a simple example to help understand this statement. A causal graph lighter " + }, + { + "bbox": [ + 46, + 169, + 286, + 358 + ], + "type": "inline_equation", + "content": "\\leftarrow" + }, + { + "bbox": [ + 46, + 169, + 286, + 358 + ], + "type": "text", + "content": " smoking " + }, + { + "bbox": [ + 46, + 169, + 286, + 358 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 46, + 169, + 286, + 358 + ], + "type": "text", + "content": " cancer means that a smoking person has a possibility to bring a lighter and get a lung cancer. Whether a person brings a lighter is usually independent of getting a lung cancer. Nevertheless, a smoker is highly possible to bring a lighter and to get a lung cancer, which leads to observe a correlation between lighter and cancer. If we can control the confounder smoking and collect data about non-smoking persons who usually bring lighters, then we can draw a correct conclusion that lighter is independent on cancer." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 360, + 287, + 612 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 360, + 287, + 612 + ], + "spans": [ + { + "bbox": [ + 46, + 360, + 287, + 612 + ], + "type": "text", + "content": "Fortunately, mechanisms to eliminate the task-induced bias naturally exist under setups of task incremental learning and domain incremental learning. In task incremental learning shown in Fig. 3 (c), the value of parent node (task identifier) is given, hence the indirected effect of " + }, + { + "bbox": [ + 46, + 360, + 287, + 612 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 46, + 360, + 287, + 612 + ], + "type": "text", + "content": " on " + }, + { + "bbox": [ + 46, + 360, + 287, + 612 + ], + "type": "inline_equation", + "content": "Y" + }, + { + "bbox": [ + 46, + 360, + 287, + 612 + ], + "type": "text", + "content": " is 0 and the task-induced bias between " + }, + { + "bbox": [ + 46, + 360, + 287, + 612 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 46, + 360, + 287, + 612 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 360, + 287, + 612 + ], + "type": "inline_equation", + "content": "Y" + }, + { + "bbox": [ + 46, + 360, + 287, + 612 + ], + "type": "text", + "content": " disappears due to the characteristic of the fork structure. In domain incremental learning, the label set remains the same along task identifiers, which is implied in " + }, + { + "bbox": [ + 46, + 360, + 287, + 612 + ], + "type": "inline_equation", + "content": "\\mathcal{Y}_t = \\mathcal{Y}_{t + 1}" + }, + { + "bbox": [ + 46, + 360, + 287, + 612 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 360, + 287, + 612 + ], + "type": "inline_equation", + "content": "P(\\mathcal{Y}_t) = P(\\mathcal{Y}_{t + 1})" + }, + { + "bbox": [ + 46, + 360, + 287, + 612 + ], + "type": "text", + "content": ". For this reason, the causal edge " + }, + { + "bbox": [ + 46, + 360, + 287, + 612 + ], + "type": "inline_equation", + "content": "T\\to Y" + }, + { + "bbox": [ + 46, + 360, + 287, + 612 + ], + "type": "text", + "content": " is cut as shown in Fig. 3 (d), resulting in disappearance of the task-induced bias. However, in the CIL, the causal path " + }, + { + "bbox": [ + 46, + 360, + 287, + 612 + ], + "type": "inline_equation", + "content": "X\\gets T\\to Y" + }, + { + "bbox": [ + 46, + 360, + 287, + 612 + ], + "type": "text", + "content": " cannot be cut by the task identifiers since they are inaccessible, or by the independence between " + }, + { + "bbox": [ + 46, + 360, + 287, + 612 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 46, + 360, + 287, + 612 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 360, + 287, + 612 + ], + "type": "inline_equation", + "content": "Y" + }, + { + "bbox": [ + 46, + 360, + 287, + 612 + ], + "type": "text", + "content": " like domain incremental learning. In order to eliminate the task-induced bias in CIL, we seek advanced causal techniques among which causal intervention [21] is a powerful one. It is noteworthy that our analysis indicates that a key objective of CIL is to cut the causal path " + }, + { + "bbox": [ + 46, + 360, + 287, + 612 + ], + "type": "inline_equation", + "content": "X\\gets T\\to Y" + }, + { + "bbox": [ + 46, + 360, + 287, + 612 + ], + "type": "text", + "content": " and we can cut it not only by conditioning variable " + }, + { + "bbox": [ + 46, + 360, + 287, + 612 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 46, + 360, + 287, + 612 + ], + "type": "text", + "content": ", but also by cutting " + }, + { + "bbox": [ + 46, + 360, + 287, + 612 + ], + "type": "inline_equation", + "content": "X\\gets T" + }, + { + "bbox": [ + 46, + 360, + 287, + 612 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 46, + 360, + 287, + 612 + ], + "type": "inline_equation", + "content": "T\\rightarrow Y" + }, + { + "bbox": [ + 46, + 360, + 287, + 612 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 615, + 135, + 626 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 615, + 135, + 626 + ], + "spans": [ + { + "bbox": [ + 47, + 615, + 135, + 626 + ], + "type": "text", + "content": "4.2. Causal debias" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 628, + 286, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 628, + 286, + 676 + ], + "spans": [ + { + "bbox": [ + 46, + 628, + 286, + 676 + ], + "type": "text", + "content": "The overall causal graph of CIL is illustrated in Fig. 4 (a), which considers two facets of task-induced bias in Fig. 1. The causal effect of " + }, + { + "bbox": [ + 46, + 628, + 286, + 676 + ], + "type": "inline_equation", + "content": "X_{1}" + }, + { + "bbox": [ + 46, + 628, + 286, + 676 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 628, + 286, + 676 + ], + "type": "inline_equation", + "content": "X_{2}" + }, + { + "bbox": [ + 46, + 628, + 286, + 676 + ], + "type": "text", + "content": " on " + }, + { + "bbox": [ + 46, + 628, + 286, + 676 + ], + "type": "inline_equation", + "content": "Y" + }, + { + "bbox": [ + 46, + 628, + 286, + 676 + ], + "type": "text", + "content": " in the original causal graph can be formulated by Bayes' rule:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 58, + 691, + 287, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 691, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 58, + 691, + 287, + 715 + ], + "type": "interline_equation", + "content": "P (Y | X _ {1}, X _ {2}) = \\sum_ {t} P (Y | \\underline {{X}} _ {1}, X _ {2}, t) P (t | \\underline {{X}} _ {1}, X _ {2}) \\tag {1}", + "image_path": "8e677bffebc317356897883b0daeb7dfceabcb79d3fb9363f2e775108286f762.jpg" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 317, + 77, + 539, + 205 + ], + "blocks": [ + { + "bbox": [ + 317, + 77, + 539, + 205 + ], + "lines": [ + { + "bbox": [ + 317, + 77, + 539, + 205 + ], + "spans": [ + { + "bbox": [ + 317, + 77, + 539, + 205 + ], + "type": "image", + "image_path": "99e227d7b07a2ac02be42df53f19a6061d2893037227a7e0b519c1dacdd8a455.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 315, + 205, + 392, + 213 + ], + "lines": [ + { + "bbox": [ + 315, + 205, + 392, + 213 + ], + "spans": [ + { + "bbox": [ + 315, + 205, + 392, + 213 + ], + "type": "text", + "content": "(c) Simplified causal graph" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 398, + 205, + 462, + 213 + ], + "lines": [ + { + "bbox": [ + 398, + 205, + 462, + 213 + ], + "spans": [ + { + "bbox": [ + 398, + 205, + 462, + 213 + ], + "type": "text", + "content": "(d) Causal intervention" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 472, + 205, + 536, + 213 + ], + "lines": [ + { + "bbox": [ + 472, + 205, + 536, + 213 + ], + "spans": [ + { + "bbox": [ + 472, + 205, + 536, + 213 + ], + "type": "text", + "content": "(e) An implementation" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 305, + 214, + 545, + 247 + ], + "lines": [ + { + "bbox": [ + 305, + 214, + 545, + 247 + ], + "spans": [ + { + "bbox": [ + 305, + 214, + 545, + 247 + ], + "type": "text", + "content": "Figure 4. Causal intervention for class incremental learning. To approximate the task identifier, we utilize a class feature dictionary " + }, + { + "bbox": [ + 305, + 214, + 545, + 247 + ], + "type": "inline_equation", + "content": "D_{t}" + }, + { + "bbox": [ + 305, + 214, + 545, + 247 + ], + "type": "text", + "content": " recording information about tasks." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 264, + 545, + 322 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 264, + 545, + 322 + ], + "spans": [ + { + "bbox": [ + 305, + 264, + 545, + 322 + ], + "type": "text", + "content": "We can treat the two variables " + }, + { + "bbox": [ + 305, + 264, + 545, + 322 + ], + "type": "inline_equation", + "content": "X_{1}" + }, + { + "bbox": [ + 305, + 264, + 545, + 322 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 264, + 545, + 322 + ], + "type": "inline_equation", + "content": "X_{2}" + }, + { + "bbox": [ + 305, + 264, + 545, + 322 + ], + "type": "text", + "content": " as one as shown in Fig. 4 (b), because they have the same context in the causal graph. As a result, we obtain a simplified causal graph in Fig. 4 (c), of which the causal effect of " + }, + { + "bbox": [ + 305, + 264, + 545, + 322 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 305, + 264, + 545, + 322 + ], + "type": "text", + "content": " on " + }, + { + "bbox": [ + 305, + 264, + 545, + 322 + ], + "type": "inline_equation", + "content": "Y" + }, + { + "bbox": [ + 305, + 264, + 545, + 322 + ], + "type": "text", + "content": " is reformulated as:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 353, + 335, + 545, + 360 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 353, + 335, + 545, + 360 + ], + "spans": [ + { + "bbox": [ + 353, + 335, + 545, + 360 + ], + "type": "interline_equation", + "content": "P (Y | X) = \\sum_ {t} P (Y | X, t) P (t | X). \\tag {2}", + "image_path": "d4479ee032af112ff225ac8a11a50456ee7294fbc3f803a5a1c990c73292c110.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 373, + 546, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 373, + 546, + 506 + ], + "spans": [ + { + "bbox": [ + 305, + 373, + 546, + 506 + ], + "type": "text", + "content": "We have shown in Fig. 3 (e) that the causal path " + }, + { + "bbox": [ + 305, + 373, + 546, + 506 + ], + "type": "inline_equation", + "content": "X \\gets T \\to Y" + }, + { + "bbox": [ + 305, + 373, + 546, + 506 + ], + "type": "text", + "content": " is the cause of task-induced bias. If we can cut this causal path, the task-induced bias will be eliminated. We opt to cut the " + }, + { + "bbox": [ + 305, + 373, + 546, + 506 + ], + "type": "inline_equation", + "content": "X \\gets T" + }, + { + "bbox": [ + 305, + 373, + 546, + 506 + ], + "type": "text", + "content": " using the causal intervention, as illustrated in Fig. 4 (d). The causal path " + }, + { + "bbox": [ + 305, + 373, + 546, + 506 + ], + "type": "inline_equation", + "content": "X \\gets T" + }, + { + "bbox": [ + 305, + 373, + 546, + 506 + ], + "type": "text", + "content": " is introduced through the term " + }, + { + "bbox": [ + 305, + 373, + 546, + 506 + ], + "type": "inline_equation", + "content": "P(t|X)" + }, + { + "bbox": [ + 305, + 373, + 546, + 506 + ], + "type": "text", + "content": " in Eq. 2. Following [21], the direct causal effect of " + }, + { + "bbox": [ + 305, + 373, + 546, + 506 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 305, + 373, + 546, + 506 + ], + "type": "text", + "content": " on " + }, + { + "bbox": [ + 305, + 373, + 546, + 506 + ], + "type": "inline_equation", + "content": "Y" + }, + { + "bbox": [ + 305, + 373, + 546, + 506 + ], + "type": "text", + "content": " is denoted by " + }, + { + "bbox": [ + 305, + 373, + 546, + 506 + ], + "type": "inline_equation", + "content": "P(Y|do(X))" + }, + { + "bbox": [ + 305, + 373, + 546, + 506 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 305, + 373, + 546, + 506 + ], + "type": "inline_equation", + "content": "do(X)" + }, + { + "bbox": [ + 305, + 373, + 546, + 506 + ], + "type": "text", + "content": " means a causal intervention on the single variable " + }, + { + "bbox": [ + 305, + 373, + 546, + 506 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 305, + 373, + 546, + 506 + ], + "type": "text", + "content": ". Because the variable " + }, + { + "bbox": [ + 305, + 373, + 546, + 506 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 305, + 373, + 546, + 506 + ], + "type": "text", + "content": " blocks all the backdoor path in the causal graph, we can obtain the expression of " + }, + { + "bbox": [ + 305, + 373, + 546, + 506 + ], + "type": "inline_equation", + "content": "P(Y|do(X))" + }, + { + "bbox": [ + 305, + 373, + 546, + 506 + ], + "type": "text", + "content": " according to the backdoor criterion [21] as:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 351, + 518, + 545, + 542 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 351, + 518, + 545, + 542 + ], + "spans": [ + { + "bbox": [ + 351, + 518, + 545, + 542 + ], + "type": "interline_equation", + "content": "P (Y | d o (X)) = \\sum_ {t} P (Y | X, t) P (t) \\tag {3}", + "image_path": "7d59253e9fae6c2666ff3396d68e32a3295dbfec2588e28431fae59afb47c3df.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 305, + 555, + 545, + 591 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 555, + 545, + 591 + ], + "spans": [ + { + "bbox": [ + 305, + 555, + 545, + 591 + ], + "type": "text", + "content": "Compared with the old term " + }, + { + "bbox": [ + 305, + 555, + 545, + 591 + ], + "type": "inline_equation", + "content": "P(t|X)" + }, + { + "bbox": [ + 305, + 555, + 545, + 591 + ], + "type": "text", + "content": ", the new term " + }, + { + "bbox": [ + 305, + 555, + 545, + 591 + ], + "type": "inline_equation", + "content": "P(t)" + }, + { + "bbox": [ + 305, + 555, + 545, + 591 + ], + "type": "text", + "content": " is no longer changed with " + }, + { + "bbox": [ + 305, + 555, + 545, + 591 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 305, + 555, + 545, + 591 + ], + "type": "text", + "content": ", equating to cut the causal edge " + }, + { + "bbox": [ + 305, + 555, + 545, + 591 + ], + "type": "inline_equation", + "content": "X \\gets T" + }, + { + "bbox": [ + 305, + 555, + 545, + 591 + ], + "type": "text", + "content": ", as shown in Fig. 4 (d)." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "text", + "content": "Based on the intervened causal graph in Fig. 4 (d) in CIL, we devise a causal debias module to implement the Equation 3. Since the task identifier is not accessible in the test phase, we propose to use a feature dictionary " + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "inline_equation", + "content": "D_{t}" + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "text", + "content": " that contains information about tasks to approximate the unobserved confounder " + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "text", + "content": " as shown in Fig. 4 (e). The task information is retained in class labels of images for CIL, because class label sets of different tasks cannot overlap with each other and thus task identifiers can be inferred from classes. Specifically, given an image " + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "text", + "content": ", a label " + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "text", + "content": " and a task identifier" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16019" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 53, + 76, + 282, + 175 + ], + "blocks": [ + { + "bbox": [ + 53, + 76, + 282, + 175 + ], + "lines": [ + { + "bbox": [ + 53, + 76, + 282, + 175 + ], + "spans": [ + { + "bbox": [ + 53, + 76, + 282, + 175 + ], + "type": "image", + "image_path": "50c829668428a825fcb6ab579a26e7a39876e9fdd5f438cfa2c4dc6c129684f8.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 176, + 288, + 242 + ], + "lines": [ + { + "bbox": [ + 46, + 176, + 288, + 242 + ], + "spans": [ + { + "bbox": [ + 46, + 176, + 288, + 242 + ], + "type": "text", + "content": "Figure 5. Illustration of causal intervention operation. This operation borrows features of classes in previous tasks, and puts them to features of new classes in the current task. The causal intervention is implied in the \"borrow\" and \"put\" operations, because the two operations mean that the features containing task information are directly controlled by ourselves." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 260, + 178, + 271 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 260, + 178, + 271 + ], + "spans": [ + { + "bbox": [ + 47, + 260, + 178, + 271 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 47, + 260, + 178, + 271 + ], + "type": "text", + "content": ", Equation 3 is implemented as" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 88, + 276, + 287, + 329 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 88, + 276, + 287, + 329 + ], + "spans": [ + { + "bbox": [ + 88, + 276, + 287, + 329 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} P (Y | d o (X)) = \\sum_ {t} P (y | \\mathbf {x}, t) P (t) \\\\ = \\sum_ {\\boldsymbol {d} _ {t}} P (y | \\boldsymbol {x}, \\boldsymbol {d} _ {t}) P (\\boldsymbol {d} _ {t}), \\tag {4} \\\\ \\end{array}", + "image_path": "231c7c9db5092f21c60320a8bc57148fbe53565a888d2e427a959e4d25aafcd8.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 334, + 287, + 393 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 334, + 287, + 393 + ], + "spans": [ + { + "bbox": [ + 47, + 334, + 287, + 393 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 334, + 287, + 393 + ], + "type": "inline_equation", + "content": "\\pmb{x}" + }, + { + "bbox": [ + 47, + 334, + 287, + 393 + ], + "type": "text", + "content": " is the feature of the image " + }, + { + "bbox": [ + 47, + 334, + 287, + 393 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 47, + 334, + 287, + 393 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 47, + 334, + 287, + 393 + ], + "type": "inline_equation", + "content": "d_{t}" + }, + { + "bbox": [ + 47, + 334, + 287, + 393 + ], + "type": "text", + "content": " is sampled from " + }, + { + "bbox": [ + 47, + 334, + 287, + 393 + ], + "type": "inline_equation", + "content": "\\pmb{D}_{t}" + }, + { + "bbox": [ + 47, + 334, + 287, + 393 + ], + "type": "text", + "content": ". If the last layer of a classifier is a softmax layer, we get " + }, + { + "bbox": [ + 47, + 334, + 287, + 393 + ], + "type": "inline_equation", + "content": "P(y|\\pmb{x},\\pmb{d}_t) = \\mathrm{Softmax}(f_y(\\pmb{x},\\pmb{d}_t))" + }, + { + "bbox": [ + 47, + 334, + 287, + 393 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 47, + 334, + 287, + 393 + ], + "type": "inline_equation", + "content": "f_{y}(\\cdot)" + }, + { + "bbox": [ + 47, + 334, + 287, + 393 + ], + "type": "text", + "content": " is a similarity function. In a nutshell, the implementation of Equation 3 is newly formulated as:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 79, + 399, + 287, + 412 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 79, + 399, + 287, + 412 + ], + "spans": [ + { + "bbox": [ + 79, + 399, + 287, + 412 + ], + "type": "interline_equation", + "content": "P (Y | d o (X)) := \\mathbb {E} _ {\\boldsymbol {d} _ {t}} \\left[ \\operatorname {S o f t m a x} \\left(f _ {y} (\\boldsymbol {x}, \\boldsymbol {d} _ {t})\\right) \\right]. \\tag {5}", + "image_path": "7808afb5be5a17a064c5bc262c2a2f5f0e9f896306156a173ee7db1ae107314b.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 418, + 287, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 418, + 287, + 441 + ], + "spans": [ + { + "bbox": [ + 46, + 418, + 287, + 441 + ], + "type": "text", + "content": "The above expectation can be approximated by NWGM [32] as follows:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 447, + 287, + 471 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 447, + 287, + 471 + ], + "spans": [ + { + "bbox": [ + 56, + 447, + 287, + 471 + ], + "type": "interline_equation", + "content": "\\mathbb {E} _ {\\boldsymbol {d} _ {t}} \\left[ \\operatorname {S o f t m a x} \\left(f _ {y} \\left(\\boldsymbol {x}, \\boldsymbol {d} _ {t}\\right)\\right) \\right] \\approx \\operatorname {S o f t m a x} \\left(\\mathbb {E} _ {\\boldsymbol {d} _ {t}} \\left[ f _ {y} \\left(\\boldsymbol {x}, \\boldsymbol {d} _ {t}\\right) \\right]\\right). \\tag {6}", + "image_path": "4f476ff9d95f6cd19dbefe67e1f39449534d590531ace236e0d5f089b525a778.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 472, + 287, + 507 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 472, + 287, + 507 + ], + "spans": [ + { + "bbox": [ + 46, + 472, + 287, + 507 + ], + "type": "text", + "content": "To implement our causal intervention operation in Fig. 5, we devise a simple attention-based mechanism with ignorable number of parameters:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 78, + 511, + 287, + 527 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 78, + 511, + 287, + 527 + ], + "spans": [ + { + "bbox": [ + 78, + 511, + 287, + 527 + ], + "type": "interline_equation", + "content": "f _ {y} (\\boldsymbol {x}, \\boldsymbol {d} _ {t}) = \\boldsymbol {W} _ {0} ^ {\\top} \\operatorname {C a t} \\left(\\boldsymbol {W} _ {1} ^ {\\top} \\boldsymbol {x}, \\boldsymbol {W} _ {2} ^ {\\top} g _ {y} (\\boldsymbol {d} _ {t})\\right), \\tag {7}", + "image_path": "97fb5786f919c000ee1f68ca2ab11587ae8de7dccf635c7f5b655ef798b41ebc.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 531, + 287, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 531, + 287, + 578 + ], + "spans": [ + { + "bbox": [ + 46, + 531, + 287, + 578 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 531, + 287, + 578 + ], + "type": "inline_equation", + "content": "W_{1}" + }, + { + "bbox": [ + 46, + 531, + 287, + 578 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 531, + 287, + 578 + ], + "type": "inline_equation", + "content": "W_{2}" + }, + { + "bbox": [ + 46, + 531, + 287, + 578 + ], + "type": "text", + "content": " both are fully connected layers and " + }, + { + "bbox": [ + 46, + 531, + 287, + 578 + ], + "type": "inline_equation", + "content": "\\operatorname{Cat}(\\cdot)" + }, + { + "bbox": [ + 46, + 531, + 287, + 578 + ], + "type": "text", + "content": " is the concatenation operator. Due to the additive property of expectation, the " + }, + { + "bbox": [ + 46, + 531, + 287, + 578 + ], + "type": "inline_equation", + "content": "\\mathbb{E}_{\\boldsymbol{d}_t}[f_y(\\boldsymbol{x}, \\boldsymbol{d}_t)]" + }, + { + "bbox": [ + 46, + 531, + 287, + 578 + ], + "type": "text", + "content": " can be rewritten as:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 582, + 287, + 608 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 582, + 287, + 608 + ], + "spans": [ + { + "bbox": [ + 57, + 582, + 287, + 608 + ], + "type": "interline_equation", + "content": "\\mathbb {E} _ {\\boldsymbol {d} _ {t}} \\left[ f _ {y} (\\boldsymbol {x}, \\boldsymbol {d} _ {t}) \\right] = \\boldsymbol {W} _ {\\boldsymbol {0}} ^ {\\top} \\operatorname {C a t} \\left(\\boldsymbol {W} _ {\\boldsymbol {1}} ^ {\\top} \\boldsymbol {x}, \\boldsymbol {W} _ {\\boldsymbol {2}} ^ {\\top} \\mathbb {E} _ {\\boldsymbol {d} _ {t}} \\left[ g _ {y} \\left(\\boldsymbol {d} _ {t}\\right) \\right]\\right). \\tag {8}", + "image_path": "92da7ea4d0c372cbfc2dcc9762951120022f331806c1c23a1253fc755c8a5268.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 609, + 287, + 632 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 609, + 287, + 632 + ], + "spans": [ + { + "bbox": [ + 46, + 609, + 287, + 632 + ], + "type": "text", + "content": "We devise " + }, + { + "bbox": [ + 46, + 609, + 287, + 632 + ], + "type": "inline_equation", + "content": "\\mathbb{E}_{\\boldsymbol{d}_t}[g_y(\\boldsymbol{d}_t)]" + }, + { + "bbox": [ + 46, + 609, + 287, + 632 + ], + "type": "text", + "content": " utilizing a scaled dot-product attention [28], which is calculated as follows:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 88, + 636, + 245, + 651 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 88, + 636, + 245, + 651 + ], + "spans": [ + { + "bbox": [ + 88, + 636, + 245, + 651 + ], + "type": "interline_equation", + "content": "Q = W _ {3} ^ {\\top} x, K = W _ {4} ^ {\\top} D _ {t}, V = D _ {t},", + "image_path": "3f24b1ed9c25b1d0f9863d91f40d037f007bbed1f78ae08ec148b027878ef663.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 90, + 654, + 287, + 667 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 90, + 654, + 287, + 667 + ], + "spans": [ + { + "bbox": [ + 90, + 654, + 287, + 667 + ], + "type": "interline_equation", + "content": "\\boldsymbol {A} = \\operatorname {S o f t m a x} \\left(\\boldsymbol {Q} ^ {\\top} \\boldsymbol {K} / \\sqrt {\\sigma}\\right), \\tag {9}", + "image_path": "626a886b48dcf350c4aef85bee751f7869f396e36588e68a86e8e7a72d402a47.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 91, + 670, + 182, + 685 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 670, + 182, + 685 + ], + "spans": [ + { + "bbox": [ + 91, + 670, + 182, + 685 + ], + "type": "interline_equation", + "content": "\\mathbb {E} _ {\\boldsymbol {d} _ {t}} \\left[ g _ {y} (\\boldsymbol {d} _ {t}) \\right] = \\boldsymbol {V} ^ {\\top} \\boldsymbol {A},", + "image_path": "a25021879300d95685fab21445500cd8daf5fa9e5d79c73c5739da13ce653291.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 46, + 689, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 689, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 689, + 287, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 689, + 287, + 713 + ], + "type": "inline_equation", + "content": "W_{3}" + }, + { + "bbox": [ + 46, + 689, + 287, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 689, + 287, + 713 + ], + "type": "inline_equation", + "content": "W_{4}" + }, + { + "bbox": [ + 46, + 689, + 287, + 713 + ], + "type": "text", + "content": " both are fully connected layers, and " + }, + { + "bbox": [ + 46, + 689, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 46, + 689, + 287, + 713 + ], + "type": "text", + "content": " is the output dimension of them." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 72, + 476, + 84 + ], + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 476, + 84 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 476, + 84 + ], + "type": "text", + "content": "Algorithm 1 CafeBoost training pipeline." + } + ] + } + ], + "index": 17, + "type": "text" + }, + { + "bbox": [ + 306, + 87, + 545, + 111 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 87, + 545, + 111 + ], + "spans": [ + { + "bbox": [ + 306, + 87, + 545, + 111 + ], + "type": "text", + "content": "Input: Training data " + }, + { + "bbox": [ + 306, + 87, + 545, + 111 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_0, \\dots, \\mathcal{D}_T" + }, + { + "bbox": [ + 306, + 87, + 545, + 111 + ], + "type": "text", + "content": ", empty dictionary " + }, + { + "bbox": [ + 306, + 87, + 545, + 111 + ], + "type": "inline_equation", + "content": "D_0" + }, + { + "bbox": [ + 306, + 87, + 545, + 111 + ], + "type": "text", + "content": ", empty example memory " + }, + { + "bbox": [ + 306, + 87, + 545, + 111 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 306, + 87, + 545, + 111 + ], + "type": "text", + "content": "," + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 312, + 112, + 545, + 135 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 112, + 545, + 135 + ], + "spans": [ + { + "bbox": [ + 312, + 112, + 545, + 135 + ], + "type": "text", + "content": "1: Initialize parameters of feature extractor " + }, + { + "bbox": [ + 312, + 112, + 545, + 135 + ], + "type": "inline_equation", + "content": "\\psi_0" + }, + { + "bbox": [ + 312, + 112, + 545, + 135 + ], + "type": "text", + "content": ", causal bias module " + }, + { + "bbox": [ + 312, + 112, + 545, + 135 + ], + "type": "inline_equation", + "content": "\\phi_0" + }, + { + "bbox": [ + 312, + 112, + 545, + 135 + ], + "type": "text", + "content": " (Eq. 8, 9) and classifier " + }, + { + "bbox": [ + 312, + 112, + 545, + 135 + ], + "type": "inline_equation", + "content": "\\theta_0" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 312, + 136, + 434, + 147 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 136, + 434, + 147 + ], + "spans": [ + { + "bbox": [ + 312, + 136, + 434, + 147 + ], + "type": "text", + "content": "2: // First task training stage." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 312, + 148, + 487, + 159 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 148, + 487, + 159 + ], + "spans": [ + { + "bbox": [ + 312, + 148, + 487, + 159 + ], + "type": "text", + "content": "3: Train " + }, + { + "bbox": [ + 312, + 148, + 487, + 159 + ], + "type": "inline_equation", + "content": "\\psi_0, \\theta_0" + }, + { + "bbox": [ + 312, + 148, + 487, + 159 + ], + "type": "text", + "content": " on " + }, + { + "bbox": [ + 312, + 148, + 487, + 159 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_0" + }, + { + "bbox": [ + 312, + 148, + 487, + 159 + ], + "type": "text", + "content": " by minimizing " + }, + { + "bbox": [ + 312, + 148, + 487, + 159 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{CE}" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 312, + 160, + 493, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 160, + 493, + 171 + ], + "spans": [ + { + "bbox": [ + 312, + 160, + 493, + 171 + ], + "type": "text", + "content": "4: Extract example features on " + }, + { + "bbox": [ + 312, + 160, + 493, + 171 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_0" + }, + { + "bbox": [ + 312, + 160, + 493, + 171 + ], + "type": "text", + "content": ", using " + }, + { + "bbox": [ + 312, + 160, + 493, + 171 + ], + "type": "inline_equation", + "content": "\\psi_0" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 312, + 172, + 523, + 182 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 172, + 523, + 182 + ], + "spans": [ + { + "bbox": [ + 312, + 172, + 523, + 182 + ], + "type": "text", + "content": "5: Store class mean features on " + }, + { + "bbox": [ + 312, + 172, + 523, + 182 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_0" + }, + { + "bbox": [ + 312, + 172, + 523, + 182 + ], + "type": "text", + "content": " to dictionary " + }, + { + "bbox": [ + 312, + 172, + 523, + 182 + ], + "type": "inline_equation", + "content": "D_0" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 312, + 183, + 484, + 195 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 183, + 484, + 195 + ], + "spans": [ + { + "bbox": [ + 312, + 183, + 484, + 195 + ], + "type": "text", + "content": "6: // Causal debias module training stage." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 312, + 196, + 486, + 207 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 196, + 486, + 207 + ], + "spans": [ + { + "bbox": [ + 312, + 196, + 486, + 207 + ], + "type": "text", + "content": "7: Train " + }, + { + "bbox": [ + 312, + 196, + 486, + 207 + ], + "type": "inline_equation", + "content": "\\phi_0, \\theta_0" + }, + { + "bbox": [ + 312, + 196, + 486, + 207 + ], + "type": "text", + "content": " on " + }, + { + "bbox": [ + 312, + 196, + 486, + 207 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_0" + }, + { + "bbox": [ + 312, + 196, + 486, + 207 + ], + "type": "text", + "content": " by minimizing " + }, + { + "bbox": [ + 312, + 196, + 486, + 207 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{CE}" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 312, + 208, + 523, + 219 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 208, + 523, + 219 + ], + "spans": [ + { + "bbox": [ + 312, + 208, + 523, + 219 + ], + "type": "text", + "content": "8: Extract example features on " + }, + { + "bbox": [ + 312, + 208, + 523, + 219 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_0" + }, + { + "bbox": [ + 312, + 208, + 523, + 219 + ], + "type": "text", + "content": ", using " + }, + { + "bbox": [ + 312, + 208, + 523, + 219 + ], + "type": "inline_equation", + "content": "\\psi_0" + }, + { + "bbox": [ + 312, + 208, + 523, + 219 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 312, + 208, + 523, + 219 + ], + "type": "inline_equation", + "content": "\\phi_0" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 312, + 220, + 469, + 230 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 220, + 469, + 230 + ], + "spans": [ + { + "bbox": [ + 312, + 220, + 469, + 230 + ], + "type": "text", + "content": "9: if example memory " + }, + { + "bbox": [ + 312, + 220, + 469, + 230 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 312, + 220, + 469, + 230 + ], + "type": "text", + "content": " is used then" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 312, + 232, + 451, + 243 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 232, + 451, + 243 + ], + "spans": [ + { + "bbox": [ + 312, + 232, + 451, + 243 + ], + "type": "text", + "content": "10: Update example memory " + }, + { + "bbox": [ + 312, + 232, + 451, + 243 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 312, + 243, + 350, + 253 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 243, + 350, + 253 + ], + "spans": [ + { + "bbox": [ + 312, + 243, + 350, + 253 + ], + "type": "text", + "content": "11: end if" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 309, + 255, + 468, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 309, + 255, + 468, + 266 + ], + "spans": [ + { + "bbox": [ + 309, + 255, + 468, + 266 + ], + "type": "text", + "content": "12: // Class incremental training stage." + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 309, + 267, + 410, + 278 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 309, + 267, + 410, + 278 + ], + "spans": [ + { + "bbox": [ + 309, + 267, + 410, + 278 + ], + "type": "text", + "content": "13: for " + }, + { + "bbox": [ + 309, + 267, + 410, + 278 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 309, + 267, + 410, + 278 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 309, + 267, + 410, + 278 + ], + "type": "inline_equation", + "content": "\\{1, \\dots, T\\}" + }, + { + "bbox": [ + 309, + 267, + 410, + 278 + ], + "type": "text", + "content": " do" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 309, + 279, + 429, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 309, + 279, + 429, + 289 + ], + "spans": [ + { + "bbox": [ + 309, + 279, + 429, + 289 + ], + "type": "text", + "content": "14: Add a new classifier " + }, + { + "bbox": [ + 309, + 279, + 429, + 289 + ], + "type": "inline_equation", + "content": "\\theta_t" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 309, + 291, + 537, + 303 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 309, + 291, + 537, + 303 + ], + "spans": [ + { + "bbox": [ + 309, + 291, + 537, + 303 + ], + "type": "text", + "content": "15: Train " + }, + { + "bbox": [ + 309, + 291, + 537, + 303 + ], + "type": "inline_equation", + "content": "\\psi_t, \\phi_t" + }, + { + "bbox": [ + 309, + 291, + 537, + 303 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 309, + 291, + 537, + 303 + ], + "type": "inline_equation", + "content": "\\theta_t" + }, + { + "bbox": [ + 309, + 291, + 537, + 303 + ], + "type": "text", + "content": " on " + }, + { + "bbox": [ + 309, + 291, + 537, + 303 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_t \\cup \\mathcal{M}" + }, + { + "bbox": [ + 309, + 291, + 537, + 303 + ], + "type": "text", + "content": " (or " + }, + { + "bbox": [ + 309, + 291, + 537, + 303 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_t" + }, + { + "bbox": [ + 309, + 291, + 537, + 303 + ], + "type": "text", + "content": ") by Eq. 10" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 309, + 304, + 531, + 315 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 309, + 304, + 531, + 315 + ], + "spans": [ + { + "bbox": [ + 309, + 304, + 531, + 315 + ], + "type": "text", + "content": "16: Extract example features on " + }, + { + "bbox": [ + 309, + 304, + 531, + 315 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_t" + }, + { + "bbox": [ + 309, + 304, + 531, + 315 + ], + "type": "text", + "content": ", using " + }, + { + "bbox": [ + 309, + 304, + 531, + 315 + ], + "type": "inline_equation", + "content": "\\psi_t" + }, + { + "bbox": [ + 309, + 304, + 531, + 315 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 309, + 304, + 531, + 315 + ], + "type": "inline_equation", + "content": "\\phi_t" + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 309, + 316, + 479, + 326 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 309, + 316, + 479, + 326 + ], + "spans": [ + { + "bbox": [ + 309, + 316, + 479, + 326 + ], + "type": "text", + "content": "17: if example memory " + }, + { + "bbox": [ + 309, + 316, + 479, + 326 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 309, + 316, + 479, + 326 + ], + "type": "text", + "content": " is used then" + } + ] + } + ], + "index": 35 + }, + { + "bbox": [ + 309, + 327, + 461, + 338 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 309, + 327, + 461, + 338 + ], + "spans": [ + { + "bbox": [ + 309, + 327, + 461, + 338 + ], + "type": "text", + "content": "18: Update example memory " + }, + { + "bbox": [ + 309, + 327, + 461, + 338 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + } + ] + } + ], + "index": 36 + }, + { + "bbox": [ + 309, + 339, + 473, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 309, + 339, + 473, + 350 + ], + "spans": [ + { + "bbox": [ + 309, + 339, + 473, + 350 + ], + "type": "text", + "content": "19: Update dictionary " + }, + { + "bbox": [ + 309, + 339, + 473, + 350 + ], + "type": "inline_equation", + "content": "D_{t}" + }, + { + "bbox": [ + 309, + 339, + 473, + 350 + ], + "type": "text", + "content": " by Eq. 11" + } + ] + } + ], + "index": 37 + }, + { + "bbox": [ + 309, + 351, + 361, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 309, + 351, + 361, + 361 + ], + "spans": [ + { + "bbox": [ + 309, + 351, + 361, + 361 + ], + "type": "text", + "content": "20: end if" + } + ] + } + ], + "index": 38 + }, + { + "bbox": [ + 309, + 363, + 356, + 372 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 309, + 363, + 356, + 372 + ], + "spans": [ + { + "bbox": [ + 309, + 363, + 356, + 372 + ], + "type": "text", + "content": "21: end for" + } + ] + } + ], + "index": 39 + }, + { + "bbox": [ + 309, + 374, + 545, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 309, + 374, + 545, + 399 + ], + "spans": [ + { + "bbox": [ + 309, + 374, + 545, + 399 + ], + "type": "text", + "content": "Output: Final feature extractor " + }, + { + "bbox": [ + 309, + 374, + 545, + 399 + ], + "type": "inline_equation", + "content": "\\psi_T" + }, + { + "bbox": [ + 309, + 374, + 545, + 399 + ], + "type": "text", + "content": ", final causal debias module " + }, + { + "bbox": [ + 309, + 374, + 545, + 399 + ], + "type": "inline_equation", + "content": "\\phi_T" + }, + { + "bbox": [ + 309, + 374, + 545, + 399 + ], + "type": "text", + "content": ", and all classifiers " + }, + { + "bbox": [ + 309, + 374, + 545, + 399 + ], + "type": "inline_equation", + "content": "\\{\\pmb{\\theta}_0, \\dots, \\pmb{\\theta}_T\\}" + } + ] + } + ], + "index": 40 + }, + { + "bbox": [ + 306, + 426, + 429, + 437 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 426, + 429, + 437 + ], + "spans": [ + { + "bbox": [ + 306, + 426, + 429, + 437 + ], + "type": "text", + "content": "4.3. Causal Feature Boost" + } + ] + } + ], + "index": 41 + }, + { + "bbox": [ + 304, + 446, + 545, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 446, + 545, + 506 + ], + "spans": [ + { + "bbox": [ + 304, + 446, + 545, + 506 + ], + "type": "text", + "content": "In this section, we describe how the causal debias module is added into the continual learning pipeline, and the training strategies to enable the overall architecture. Additionally, we propose a momentum method to update the dictionary so as to catch more task information." + } + ] + } + ], + "index": 42 + }, + { + "bbox": [ + 304, + 508, + 545, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 508, + 545, + 628 + ], + "spans": [ + { + "bbox": [ + 304, + 508, + 545, + 628 + ], + "type": "text", + "content": "Architecture. Our causal debias module can be added between the feature extractor and the classifier of existing continual learning architectures. In Fig. 2, given an image as input, the CNN backbone (e.g. ResNet [8]) will extract a feature map. Then, we feed this feature map and a pre-collected dictionary into our causal debias module, and achieve a newly-updated feature map which is forwarded to the classifier to get a class label. After a new task comes, we add one more classifier for new classes and add no other layers." + } + ] + } + ], + "index": 43 + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "content": "Training pipeline. Because the dictionary used in the causal debias module need to be filled with average features of classes, we train a feature extractor and a classifier by using data of the initial phase firstly. Then, we fix the feature extractor and run an extra phase before the incremental phases to train the newly-added causal debias module and the classifier together. In the remaining phases," + } + ] + } + ], + "index": 44 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16020" + } + ] + } + ], + "index": 45 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 116, + 70, + 481, + 299 + ], + "blocks": [ + { + "bbox": [ + 116, + 70, + 481, + 299 + ], + "lines": [ + { + "bbox": [ + 116, + 70, + 481, + 299 + ], + "spans": [ + { + "bbox": [ + 116, + 70, + 481, + 299 + ], + "type": "table", + "html": "
MethodCIFAR-100ImageNet-SubImageNet-Full
Average accuracy(%)5102551025510
LwF†[16]49.5946.9845.5153.6247.6444.3244.3538.90
BiC†[31]59.3654.2050.0070.0764.9657.7362.6558.72
iCaRL [22]62.9556.9251.3065.0460.8254.5651.5046.89
LUCIR [11]63.5657.2050.2970.8468.3261.4464.4561.57
GeoDL [24]67.0860.9653.9871.2770.2464.8665.2364.46
CwD [23]66.8462.6656.2867.3664.0960.52--
CSCCT [3]66.8959.9152.3271.3367.9161.52--
iCaRL+CafeBoost69.4164.7061.6070.5367.0463.4053.4648.46
LUCIR+CafeBoost70.1066.5662.3771.5170.0367.7466.8864.77
Forgetting rate(%)5102551025510
LwF†[16]43.3643.5841.6655.3257.0055.1248.7047.94
BiC†[31]31.4232.5034.6027.0431.0437.8825.0628.34
iCaRL [22]22.9525.3025.7418.7021.3525.0926.0333.76
LUCIR [11]23.3026.4029.2131.8833.4835.4024.0827.29
GeoDL [24]16.8321.8324.918.689.0711.9711.0312.81
CwD [23]17.7026.8026.9111.1617.5325.41--
iCaRL+CafeBoost8.5112.9116.0812.2014.2318.5225.1928.24
LUCIR+CafeBoost6.999.0311.758.167.509.108.6511.73
", + "image_path": "3a125f140b603fdee2ee2bc374b608b40354ad8c66aceebbf1f52fc921e144e5.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 300, + 547, + 333 + ], + "lines": [ + { + "bbox": [ + 46, + 300, + 547, + 333 + ], + "spans": [ + { + "bbox": [ + 46, + 300, + 547, + 333 + ], + "type": "text", + "content": "Table 1. The average incremental accuracy and the forgetting rate on CIFAR-100, ImageNet-Sub and ImageNet-Full. The number of incremental phases are 5, 10, and 25. The number of exemplars per class is 20. An ideal continual learning method should have high average incremental accuracy and low forgetting rate. Results of models with " + }, + { + "bbox": [ + 46, + 300, + 547, + 333 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 46, + 300, + 547, + 333 + ], + "type": "text", + "content": " are taken from [24]." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 345, + 288, + 416 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 345, + 288, + 416 + ], + "spans": [ + { + "bbox": [ + 46, + 345, + 288, + 416 + ], + "type": "text", + "content": "the causal debias module are always learnable, whereas the classifiers of old tasks are fixed. Our causal debias module can be used in iCaRL [22] and LUCIR [11] as a plugin. Our overall approach borrows their corresponding training objectives. Therefore, the training objective of our approach (CafeBoost) is formulated as:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 108, + 426, + 287, + 444 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 108, + 426, + 287, + 444 + ], + "spans": [ + { + "bbox": [ + 108, + 426, + 287, + 444 + ], + "type": "interline_equation", + "content": "\\min _ {\\theta} \\left[ L _ {c e} (x, y, \\theta) + \\eta L _ {*} (\\theta) \\right], \\tag {10}", + "image_path": "4799f7e2022fc64d12946c229e418ae62e1c1c9cf772e7a10742b7eb232effab.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 453, + 287, + 501 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 453, + 287, + 501 + ], + "spans": [ + { + "bbox": [ + 46, + 453, + 287, + 501 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 453, + 287, + 501 + ], + "type": "inline_equation", + "content": "L_{*}(\\theta)" + }, + { + "bbox": [ + 46, + 453, + 287, + 501 + ], + "type": "text", + "content": " denotes specific losses of the corresponding base methods, e.g., distillation loss for iCaRL, margin ranking loss and others for LUCIR. Algorithm 1 details out the CafeBoost training pipeline." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 501, + 287, + 561 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 501, + 287, + 561 + ], + "spans": [ + { + "bbox": [ + 46, + 501, + 287, + 561 + ], + "type": "text", + "content": "Momentum dictionary updating. Our approach exploits an external dictionary to store mean features of each class in the initial phase. We can exploit a momentum updating strategy to renew features in the dictionary, if replay buffer is used. The new dictionary at task " + }, + { + "bbox": [ + 46, + 501, + 287, + 561 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 501, + 287, + 561 + ], + "type": "text", + "content": " is updated as" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 59, + 569, + 287, + 585 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 569, + 287, + 585 + ], + "spans": [ + { + "bbox": [ + 59, + 569, + 287, + 585 + ], + "type": "interline_equation", + "content": "\\pmb {D} _ {t} = \\alpha \\pmb {D} _ {t - 1} + (1 - \\alpha) [ \\mathbf {f} _ {1} ^ {t - 1}, \\dots , \\mathbf {f} _ {c} ^ {t - 1}, \\dots , \\mathbf {f} _ {C} ^ {t - 1} ], \\quad (1 1)", + "image_path": "366b08fb9b3288f350bfacf1a1799665377390aeda91afba8cfbbc9daa32f537.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 593, + 287, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 593, + 287, + 616 + ], + "spans": [ + { + "bbox": [ + 46, + 593, + 287, + 616 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 593, + 287, + 616 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 46, + 593, + 287, + 616 + ], + "type": "text", + "content": " controls the updating rate, " + }, + { + "bbox": [ + 46, + 593, + 287, + 616 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_c" + }, + { + "bbox": [ + 46, + 593, + 287, + 616 + ], + "type": "text", + "content": " represents a mean feature of class " + }, + { + "bbox": [ + 46, + 593, + 287, + 616 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 46, + 593, + 287, + 616 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 593, + 287, + 616 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 46, + 593, + 287, + 616 + ], + "type": "text", + "content": " is the number of classes." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 627, + 128, + 641 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 627, + 128, + 641 + ], + "spans": [ + { + "bbox": [ + 47, + 627, + 128, + 641 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 647, + 108, + 659 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 647, + 108, + 659 + ], + "spans": [ + { + "bbox": [ + 47, + 647, + 108, + 659 + ], + "type": "text", + "content": "5.1. Settings" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 665, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 288, + 714 + ], + "type": "text", + "content": "Datasets & Protocols. We evaluate our approach on two commonly-used datasets: CIFAR-100 [15] and ImageNet [5]. CIFAR-100 contains 100 classes and each class has 500 images for training and 100 images for test, with" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 345, + 547, + 464 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 345, + 547, + 464 + ], + "spans": [ + { + "bbox": [ + 304, + 345, + 547, + 464 + ], + "type": "text", + "content": "each image of size " + }, + { + "bbox": [ + 304, + 345, + 547, + 464 + ], + "type": "inline_equation", + "content": "32 \\times 32" + }, + { + "bbox": [ + 304, + 345, + 547, + 464 + ], + "type": "text", + "content": ". ImageNet with about 1.3 million images of size " + }, + { + "bbox": [ + 304, + 345, + 547, + 464 + ], + "type": "inline_equation", + "content": "224 \\times 224" + }, + { + "bbox": [ + 304, + 345, + 547, + 464 + ], + "type": "text", + "content": " is divided into two CIL settings: ImageNet-Sub that owns 100 classes and ImageNet-Full that owns 1000 classes. We follow the protocol used in many recent works [11,23,24]. Models are firstly trained on half classes, i.e., 50 classes for CIFAR100 and ImageNet-Sub, and 500 for ImageNet-Full. Subsequently, we start incremental learning phases with adding a fixed number of unobserved classes in each phase. We choose three different numbers of incremental phases: 5, 10, and 25." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 464, + 547, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 464, + 547, + 525 + ], + "spans": [ + { + "bbox": [ + 304, + 464, + 547, + 525 + ], + "type": "text", + "content": "Evaluation Metrics. The two commonly-used metrics to evaluate performance of a CIL model are 1) the average incremental accuracy [22] " + }, + { + "bbox": [ + 304, + 464, + 547, + 525 + ], + "type": "inline_equation", + "content": "\\mathcal{A}_T" + }, + { + "bbox": [ + 304, + 464, + 547, + 525 + ], + "type": "text", + "content": ", and 2) the average forgetting " + }, + { + "bbox": [ + 304, + 464, + 547, + 525 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_T" + }, + { + "bbox": [ + 304, + 464, + 547, + 525 + ], + "type": "text", + "content": ". The average incremental accuracy and the average forgetting after " + }, + { + "bbox": [ + 304, + 464, + 547, + 525 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 304, + 464, + 547, + 525 + ], + "type": "text", + "content": " tasks are defined as" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 347, + 529, + 505, + 563 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 347, + 529, + 505, + 563 + ], + "spans": [ + { + "bbox": [ + 347, + 529, + 505, + 563 + ], + "type": "interline_equation", + "content": "\\mathcal {A} _ {T} = \\frac {1}{T} \\sum_ {t = 1} ^ {T} A _ {t}, \\mathcal {F} _ {T} = \\frac {1}{T - 1} \\sum_ {t = 2} ^ {T} F _ {t},", + "image_path": "5c9d9b4a762b523190aa08b144b25fe7c8df5d2190c1a2104a2ca92542d46f0f.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 567, + 545, + 618 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 567, + 545, + 618 + ], + "spans": [ + { + "bbox": [ + 304, + 567, + 545, + 618 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 567, + 545, + 618 + ], + "type": "inline_equation", + "content": "a_{i,j}" + }, + { + "bbox": [ + 304, + 567, + 545, + 618 + ], + "type": "text", + "content": " denotes the accuracy on task " + }, + { + "bbox": [ + 304, + 567, + 545, + 618 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 567, + 545, + 618 + ], + "type": "text", + "content": " after the model learning task " + }, + { + "bbox": [ + 304, + 567, + 545, + 618 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 304, + 567, + 545, + 618 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 567, + 545, + 618 + ], + "type": "inline_equation", + "content": "A_{t} = \\frac{1}{t}\\sum_{i=1}^{t}a_{i,t}" + }, + { + "bbox": [ + 304, + 567, + 545, + 618 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 567, + 545, + 618 + ], + "type": "inline_equation", + "content": "F_{t} = \\frac{1}{t-1}\\sum_{j=2}^{t}\\max_{i}(a_{i,t}-a_{j,t})" + }, + { + "bbox": [ + 304, + 567, + 545, + 618 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 304, + 567, + 545, + 618 + ], + "type": "inline_equation", + "content": "i\\in\\uparrow,\\dots,j-1" + }, + { + "bbox": [ + 304, + 567, + 545, + 618 + ], + "type": "text", + "content": ". We report top-1 accuracy." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 618, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 618, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 618, + 547, + 713 + ], + "type": "text", + "content": "Implementation details. All methods are implemented with PyTorch [20]. We use the ResNet18 for CIFAR-100 and ImageNet following [23]. For all experiments, SGD optimizer and the batch size 128 are used. For CIFAR-100 and ImageNet, we train models for 160 epochs and 90 epochs, respectively. For CIFAR-100, our results are averaged over three runs with different orders generated by random seed 1993. For ImageNet, the first order with random seed 1993" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16021" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 145, + 72, + 447, + 84 + ], + "blocks": [ + { + "bbox": [ + 145, + 72, + 447, + 84 + ], + "lines": [ + { + "bbox": [ + 145, + 72, + 447, + 84 + ], + "spans": [ + { + "bbox": [ + 145, + 72, + 447, + 84 + ], + "type": "image", + "image_path": "40b0fadf920b754d05c327f6107da3150598c29de32d14a4b4417b23fc602a8a.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 93, + 92, + 227, + 198 + ], + "blocks": [ + { + "bbox": [ + 93, + 92, + 227, + 198 + ], + "lines": [ + { + "bbox": [ + 93, + 92, + 227, + 198 + ], + "spans": [ + { + "bbox": [ + 93, + 92, + 227, + 198 + ], + "type": "image", + "image_path": "038c345e9eac4d2e9ffac27d24f8ecbbe3df915a08537de4c4cd5793fd408a72.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 235, + 92, + 362, + 198 + ], + "blocks": [ + { + "bbox": [ + 235, + 92, + 362, + 198 + ], + "lines": [ + { + "bbox": [ + 235, + 92, + 362, + 198 + ], + "spans": [ + { + "bbox": [ + 235, + 92, + 362, + 198 + ], + "type": "image", + "image_path": "de3b6a1e3f8156fff399204403ac09a3261f3c9645b06d4e15a5d85cff4d7431.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 370, + 93, + 496, + 198 + ], + "blocks": [ + { + "bbox": [ + 370, + 93, + 496, + 198 + ], + "lines": [ + { + "bbox": [ + 370, + 93, + 496, + 198 + ], + "spans": [ + { + "bbox": [ + 370, + 93, + 496, + 198 + ], + "type": "image", + "image_path": "40b992f1e8937c5bffe9dce24273eeaf673b0c865ac47d7656eea329e1912352.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 93, + 199, + 227, + 310 + ], + "blocks": [ + { + "bbox": [ + 93, + 199, + 227, + 310 + ], + "lines": [ + { + "bbox": [ + 93, + 199, + 227, + 310 + ], + "spans": [ + { + "bbox": [ + 93, + 199, + 227, + 310 + ], + "type": "image", + "image_path": "8a414de9bd3f211a56d8c2d5c4435c695155f9c581d0fe25649d3039ae1fce0f.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 312, + 546, + 335 + ], + "lines": [ + { + "bbox": [ + 46, + 312, + 546, + 335 + ], + "spans": [ + { + "bbox": [ + 46, + 312, + 546, + 335 + ], + "type": "text", + "content": "Figure 6. Results on CIFAR-100 (5, 10, and 25 tasks), ImageNet-Sub (5, 10, and 25 tasks). The horizontal axis shows the task identifier and the vertical axis is the corresponding average incremental accuracy." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 235, + 200, + 362, + 310 + ], + "blocks": [ + { + "bbox": [ + 235, + 200, + 362, + 310 + ], + "lines": [ + { + "bbox": [ + 235, + 200, + 362, + 310 + ], + "spans": [ + { + "bbox": [ + 235, + 200, + 362, + 310 + ], + "type": "image", + "image_path": "f31257b9d713d80071cb3cdc340df6d5c45d20d56c6bbd3c442c28b03fe35300.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 371, + 200, + 497, + 310 + ], + "blocks": [ + { + "bbox": [ + 371, + 200, + 497, + 310 + ], + "lines": [ + { + "bbox": [ + 371, + 200, + 497, + 310 + ], + "spans": [ + { + "bbox": [ + 371, + 200, + 497, + 310 + ], + "type": "image", + "image_path": "aa30277a053aa74a1688d2ceaa1d7a4a14e983758e3546cd3626acd04bb35f98.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 347, + 287, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 347, + 287, + 407 + ], + "spans": [ + { + "bbox": [ + 46, + 347, + 287, + 407 + ], + "type": "text", + "content": "is conventionally used. In addition, we deploy CafeBoost to two baselines: iCaRL [22] and LUCIR [11] and compare it with LwF [16], BiC [31], GeoDL [24], CwD [23] and CSCCT [3] on all datasets. GeoDL, CwD and CSCCT are implemented based on LUCIR." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 418, + 126, + 430 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 418, + 126, + 430 + ], + "spans": [ + { + "bbox": [ + 47, + 418, + 126, + 430 + ], + "type": "text", + "content": "5.2. Evaluations" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 437, + 287, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 437, + 287, + 640 + ], + "spans": [ + { + "bbox": [ + 46, + 437, + 287, + 640 + ], + "type": "text", + "content": "Results on CIFAR-100. Tab. 1 summarizes the experimental results of CIFAR-100 in three different numbers of incremental tasks. We observe that CafeBoost improves iCaRL baseline by " + }, + { + "bbox": [ + 46, + 437, + 287, + 640 + ], + "type": "inline_equation", + "content": "6.46\\%" + }, + { + "bbox": [ + 46, + 437, + 287, + 640 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 437, + 287, + 640 + ], + "type": "inline_equation", + "content": "7.78\\%" + }, + { + "bbox": [ + 46, + 437, + 287, + 640 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 437, + 287, + 640 + ], + "type": "inline_equation", + "content": "10.30\\%" + }, + { + "bbox": [ + 46, + 437, + 287, + 640 + ], + "type": "text", + "content": " for 5, 10, and 25 tasks, respectively. After adding to LUCIR baseline, we observe a performance boost: " + }, + { + "bbox": [ + 46, + 437, + 287, + 640 + ], + "type": "inline_equation", + "content": "6.54\\%" + }, + { + "bbox": [ + 46, + 437, + 287, + 640 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 437, + 287, + 640 + ], + "type": "inline_equation", + "content": "9.36\\%" + }, + { + "bbox": [ + 46, + 437, + 287, + 640 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 437, + 287, + 640 + ], + "type": "inline_equation", + "content": "12.08\\%" + }, + { + "bbox": [ + 46, + 437, + 287, + 640 + ], + "type": "text", + "content": ", for 5, 10, and 25 tasks. The corresponding average incremental accuracies reach high levels of " + }, + { + "bbox": [ + 46, + 437, + 287, + 640 + ], + "type": "inline_equation", + "content": "70.10\\%" + }, + { + "bbox": [ + 46, + 437, + 287, + 640 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 437, + 287, + 640 + ], + "type": "inline_equation", + "content": "66.56\\%" + }, + { + "bbox": [ + 46, + 437, + 287, + 640 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 437, + 287, + 640 + ], + "type": "inline_equation", + "content": "62.37\\%" + }, + { + "bbox": [ + 46, + 437, + 287, + 640 + ], + "type": "text", + "content": ". Furthermore, CafeBoost achieves lower forgetting rates after applying to the baselines. It can decrease the forgetting rates of iCaRL by " + }, + { + "bbox": [ + 46, + 437, + 287, + 640 + ], + "type": "inline_equation", + "content": "14.44\\%" + }, + { + "bbox": [ + 46, + 437, + 287, + 640 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 437, + 287, + 640 + ], + "type": "inline_equation", + "content": "12.39\\%" + }, + { + "bbox": [ + 46, + 437, + 287, + 640 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 437, + 287, + 640 + ], + "type": "inline_equation", + "content": "9.66\\%" + }, + { + "bbox": [ + 46, + 437, + 287, + 640 + ], + "type": "text", + "content": " and those of LUCIR by " + }, + { + "bbox": [ + 46, + 437, + 287, + 640 + ], + "type": "inline_equation", + "content": "16.31\\%" + }, + { + "bbox": [ + 46, + 437, + 287, + 640 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 437, + 287, + 640 + ], + "type": "inline_equation", + "content": "17.37\\%" + }, + { + "bbox": [ + 46, + 437, + 287, + 640 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 437, + 287, + 640 + ], + "type": "inline_equation", + "content": "17.46\\%" + }, + { + "bbox": [ + 46, + 437, + 287, + 640 + ], + "type": "text", + "content": ", for 5, 10, and 25 tasks. The first row of Fig. 6 shows CafeBoost improves performance of baselines by a large margin, achieves higher accuracies and prevents catastrophic forgetting better than the prior methods for CIL. See Appendix for more results about the forgetting rates." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 642, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 642, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 642, + 287, + 714 + ], + "type": "text", + "content": "Results on ImageNet. Tab. 1 also reports experimental results of ImageNet-Sub and ImageNet-Full. On ImageNet-Sub, CafeBoost boosts average incremental accuracies by " + }, + { + "bbox": [ + 46, + 642, + 287, + 714 + ], + "type": "inline_equation", + "content": "0.67\\% \\sim 8.84\\%" + }, + { + "bbox": [ + 46, + 642, + 287, + 714 + ], + "type": "text", + "content": " over iCaRL and LUCIR baselines, for 5, 20, and 25 tasks, and decreases the corresponding forgetting rates by " + }, + { + "bbox": [ + 46, + 642, + 287, + 714 + ], + "type": "inline_equation", + "content": "6.57\\% \\sim 26.30\\%" + }, + { + "bbox": [ + 46, + 642, + 287, + 714 + ], + "type": "text", + "content": ". For example, CafeBoost" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 347, + 546, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 347, + 546, + 491 + ], + "spans": [ + { + "bbox": [ + 304, + 347, + 546, + 491 + ], + "type": "text", + "content": "improves average incremental accuracy of LUCIR from " + }, + { + "bbox": [ + 304, + 347, + 546, + 491 + ], + "type": "inline_equation", + "content": "61.44\\%" + }, + { + "bbox": [ + 304, + 347, + 546, + 491 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 304, + 347, + 546, + 491 + ], + "type": "inline_equation", + "content": "67.74\\%" + }, + { + "bbox": [ + 304, + 347, + 546, + 491 + ], + "type": "inline_equation", + "content": "(+6.30\\%)" + }, + { + "bbox": [ + 304, + 347, + 546, + 491 + ], + "type": "text", + "content": ", and reduces forgetting rate from " + }, + { + "bbox": [ + 304, + 347, + 546, + 491 + ], + "type": "inline_equation", + "content": "35.40\\%" + }, + { + "bbox": [ + 304, + 347, + 546, + 491 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 304, + 347, + 546, + 491 + ], + "type": "inline_equation", + "content": "9.10\\%" + }, + { + "bbox": [ + 304, + 347, + 546, + 491 + ], + "type": "inline_equation", + "content": "(-26.30\\%)" + }, + { + "bbox": [ + 304, + 347, + 546, + 491 + ], + "type": "text", + "content": ". The second row of Fig. 6 further validates effectiveness of CafeBoost for all numbers of incremental phases. Additionally, we observe in Fig. 6 CafeBoost has almost no effect on the initial phase, but brings great boost in the following incremental phases. CafeBoost also enhances performance of iCaRL and LUCIR on ImageNet-Full, as shown in Tab. 1. We observe CafeBoost improves accuracies of iCaRL by " + }, + { + "bbox": [ + 304, + 347, + 546, + 491 + ], + "type": "inline_equation", + "content": "1.96\\%" + }, + { + "bbox": [ + 304, + 347, + 546, + 491 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 347, + 546, + 491 + ], + "type": "inline_equation", + "content": "1.57\\%" + }, + { + "bbox": [ + 304, + 347, + 546, + 491 + ], + "type": "text", + "content": ", and those of LUCIR by " + }, + { + "bbox": [ + 304, + 347, + 546, + 491 + ], + "type": "inline_equation", + "content": "2.43\\%" + }, + { + "bbox": [ + 304, + 347, + 546, + 491 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 347, + 546, + 491 + ], + "type": "inline_equation", + "content": "3.20\\%" + }, + { + "bbox": [ + 304, + 347, + 546, + 491 + ], + "type": "text", + "content": " for 5 and 10 incremental phases respectively." + } + ] + } + ], + "index": 12 + }, + { + "type": "table", + "bbox": [ + 309, + 500, + 545, + 599 + ], + "blocks": [ + { + "bbox": [ + 309, + 500, + 545, + 599 + ], + "lines": [ + { + "bbox": [ + 309, + 500, + 545, + 599 + ], + "spans": [ + { + "bbox": [ + 309, + 500, + 545, + 599 + ], + "type": "table", + "html": "
SB1020304050
10LUCIR59.1260.2661.7761.6262.91
+Ours55.8762.6166.2168.2269.70
↑-3.25+2.35+4.44+6.60+6.79
5LUCIR54.6256.2056.6557.2956.90
+Ours49.0258.5363.3965.7365.95
↑-5.60+2.33+6.74+8.44+9.05
", + "image_path": "409646a02e6def168a1da2f23f83d129fa4e5c90949f9960812446bb90f2d149.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_body" + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 600, + 545, + 665 + ], + "lines": [ + { + "bbox": [ + 304, + 600, + 545, + 665 + ], + "spans": [ + { + "bbox": [ + 304, + 600, + 545, + 665 + ], + "type": "text", + "content": "Table 2. Ablation study on impact of number of classes in the initial phase. " + }, + { + "bbox": [ + 304, + 600, + 545, + 665 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 304, + 600, + 545, + 665 + ], + "type": "text", + "content": " denotes number of classes in the initial phase. " + }, + { + "bbox": [ + 304, + 600, + 545, + 665 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 304, + 600, + 545, + 665 + ], + "type": "text", + "content": " is the number of new classes per incremental step. Experiments are conducted on CIFAR-100 with 20 exemplars per class. The reported values (\\%) are average incremental accuracies. " + }, + { + "bbox": [ + 304, + 600, + 545, + 665 + ], + "type": "inline_equation", + "content": "\\uparrow" + }, + { + "bbox": [ + 304, + 600, + 545, + 665 + ], + "type": "text", + "content": " denotes the increment of values." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 306, + 671, + 406, + 682 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 671, + 406, + 682 + ], + "spans": [ + { + "bbox": [ + 306, + 671, + 406, + 682 + ], + "type": "text", + "content": "5.3. Ablation Studies" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 305, + 689, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 689, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 305, + 689, + 545, + 714 + ], + "type": "text", + "content": "In this section, we conduct experiments to study the effect on our approach of 1) number of classes in initial phase," + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16022" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 69, + 70, + 265, + 171 + ], + "blocks": [ + { + "bbox": [ + 69, + 70, + 265, + 171 + ], + "lines": [ + { + "bbox": [ + 69, + 70, + 265, + 171 + ], + "spans": [ + { + "bbox": [ + 69, + 70, + 265, + 171 + ], + "type": "table", + "html": "
SRLUCIR+Ours↑
10554.23±0.7864.80±0.36+10.57
1059.98±0.5468.30±0.31+8.32
2063.56±0.5570.10±0.39+6.54
5547.74±0.8462.23±0.67+14.49
1053.22±1.2864.08±0.77+10.86
2057.20±1.1466.56±0.80+9.36
", + "image_path": "e5f20c61c589ff9e450b61de1fdf89caa9bc427bec422b831703bc066dab2bff.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 50, + 239, + 284, + 290 + ], + "blocks": [ + { + "bbox": [ + 46, + 171, + 287, + 224 + ], + "lines": [ + { + "bbox": [ + 46, + 171, + 287, + 224 + ], + "spans": [ + { + "bbox": [ + 46, + 171, + 287, + 224 + ], + "type": "text", + "content": "Table 3. Ablation study on impact of number of exemplars. " + }, + { + "bbox": [ + 46, + 171, + 287, + 224 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 46, + 171, + 287, + 224 + ], + "type": "text", + "content": " represents the number of exemplars per class. " + }, + { + "bbox": [ + 46, + 171, + 287, + 224 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 46, + 171, + 287, + 224 + ], + "type": "text", + "content": " is the number of new classes per incremental step. Experiments are conducted on CIFAR-100. The reported values (%) are average incremental accuracies. " + }, + { + "bbox": [ + 46, + 171, + 287, + 224 + ], + "type": "inline_equation", + "content": "\\uparrow" + }, + { + "bbox": [ + 46, + 171, + 287, + 224 + ], + "type": "text", + "content": " denotes the increment of values." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 50, + 239, + 284, + 290 + ], + "lines": [ + { + "bbox": [ + 50, + 239, + 284, + 290 + ], + "spans": [ + { + "bbox": [ + 50, + 239, + 284, + 290 + ], + "type": "table", + "html": "
MethodsCIFAR-100ImageNet-Sub
AccuracyForgettingAccuracyForgetting
LUCIR40.10±1.2752.35±1.3160.4725.73
+Ours59.01±0.2223.31±0.7665.1518.05
", + "image_path": "768a14a0cdba8bc17e05bcbad1dd15fdfe8396dc45a79504ad5ad210cb859b74.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 339, + 286, + 363 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 339, + 286, + 363 + ], + "spans": [ + { + "bbox": [ + 46, + 339, + 286, + 363 + ], + "type": "text", + "content": "2) number of exemplars in memory, 3) different classifiers, and 4) coefficient of momentum dictionary update." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 365, + 286, + 507 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 365, + 286, + 507 + ], + "spans": [ + { + "bbox": [ + 46, + 365, + 286, + 507 + ], + "type": "text", + "content": "Impact of number of classes in initial phase. We investigate CafeBoost with 10, 20, 30, 40, and 50 classes of initial phase in Tab. 2. CafeBoost can enhance performance when the number of classes in initial phase is larger than 10. The performance improvement increases from negative to positive, along the number for both " + }, + { + "bbox": [ + 46, + 365, + 286, + 507 + ], + "type": "inline_equation", + "content": "S = 10" + }, + { + "bbox": [ + 46, + 365, + 286, + 507 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 365, + 286, + 507 + ], + "type": "inline_equation", + "content": "S = 5" + }, + { + "bbox": [ + 46, + 365, + 286, + 507 + ], + "type": "text", + "content": ". When " + }, + { + "bbox": [ + 46, + 365, + 286, + 507 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 46, + 365, + 286, + 507 + ], + "type": "text", + "content": " equates 10, our approach has a negative influence on the baseline. This is because characteristics of classes in the initial phase are insufficient to share for all classes in the dataset. However, when " + }, + { + "bbox": [ + 46, + 365, + 286, + 507 + ], + "type": "inline_equation", + "content": "B >= 20" + }, + { + "bbox": [ + 46, + 365, + 286, + 507 + ], + "type": "text", + "content": ", our approach boosts the baseline by a large margin, which shows that it tolerates a broad range of the number " + }, + { + "bbox": [ + 46, + 365, + 286, + 507 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 46, + 365, + 286, + 507 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 509, + 286, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 509, + 286, + 640 + ], + "spans": [ + { + "bbox": [ + 46, + 509, + 286, + 640 + ], + "type": "text", + "content": "Impact of number of exemplars in memory. We vary the number of exemplars per class and show its effect in Tab. 3. Both the performance of baselines and CafeBoost improve along the increase of exemplar number per class, which is consistently observed for " + }, + { + "bbox": [ + 46, + 509, + 286, + 640 + ], + "type": "inline_equation", + "content": "S = 10" + }, + { + "bbox": [ + 46, + 509, + 286, + 640 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 509, + 286, + 640 + ], + "type": "inline_equation", + "content": "S = 5" + }, + { + "bbox": [ + 46, + 509, + 286, + 640 + ], + "type": "text", + "content": ". In addition, the improvement margin decreases along the increase of " + }, + { + "bbox": [ + 46, + 509, + 286, + 640 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 46, + 509, + 286, + 640 + ], + "type": "text", + "content": ". For instance, " + }, + { + "bbox": [ + 46, + 509, + 286, + 640 + ], + "type": "inline_equation", + "content": "+10.57" + }, + { + "bbox": [ + 46, + 509, + 286, + 640 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 46, + 509, + 286, + 640 + ], + "type": "inline_equation", + "content": "S = 10" + }, + { + "bbox": [ + 46, + 509, + 286, + 640 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 509, + 286, + 640 + ], + "type": "inline_equation", + "content": "R = 5" + }, + { + "bbox": [ + 46, + 509, + 286, + 640 + ], + "type": "text", + "content": ", whereas " + }, + { + "bbox": [ + 46, + 509, + 286, + 640 + ], + "type": "inline_equation", + "content": "+6.54" + }, + { + "bbox": [ + 46, + 509, + 286, + 640 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 46, + 509, + 286, + 640 + ], + "type": "inline_equation", + "content": "S = 10" + }, + { + "bbox": [ + 46, + 509, + 286, + 640 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 509, + 286, + 640 + ], + "type": "inline_equation", + "content": "R = 20" + }, + { + "bbox": [ + 46, + 509, + 286, + 640 + ], + "type": "text", + "content": ". It is worth noting that CafeBoost does not rely on a replay buffer. Tab. 4 shows CafeBoost outperforms remarkably the baseline on accuracy and forgetting in CIFAR-100 and ImageNet-Sub." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 641, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 641, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 641, + 286, + 713 + ], + "type": "text", + "content": "Impact of different classifiers. We also conduct experiments to research behaviors of CafeBoost combined with three different classifiers. We adopt the classifiers CNN, " + }, + { + "bbox": [ + 46, + 641, + 286, + 713 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 641, + 286, + 713 + ], + "type": "text", + "content": "-NME, and AME, as in [11]. CNN classifies examples by exploiting one class prototype per class. " + }, + { + "bbox": [ + 46, + 641, + 286, + 713 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 641, + 286, + 713 + ], + "type": "text", + "content": "-NME uses " + }, + { + "bbox": [ + 46, + 641, + 286, + 713 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 641, + 286, + 713 + ], + "type": "text", + "content": " samples of each class to represent the class, while AME uti" + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 309, + 70, + 550, + 182 + ], + "blocks": [ + { + "bbox": [ + 46, + 291, + 286, + 324 + ], + "lines": [ + { + "bbox": [ + 46, + 291, + 286, + 324 + ], + "spans": [ + { + "bbox": [ + 46, + 291, + 286, + 324 + ], + "type": "text", + "content": "Table 4. Comparisons of average incremental accuracies and forgetting rates with no replay data. Experiments are conducted with 5 incremental steps. 20 exemplars are stored per class." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 309, + 70, + 550, + 182 + ], + "lines": [ + { + "bbox": [ + 309, + 70, + 550, + 182 + ], + "spans": [ + { + "bbox": [ + 309, + 70, + 550, + 182 + ], + "type": "table", + "html": "
MethodiCaRLiCaRL+oursLUCIRLUCIR+ours
CNN553.47±0.1556.53±0.2663.56±0.5570.10±0.39
1049.30±0.4353.75±0.0357.20±1.1466.56±0.80
2545.59±0.7551.18±0.7850.29±1.1462.37±1.17
k-NME562.95±0.5169.40±0.2365.42±0.6769.77±0.35
1056.91±0.6464.70±0.4059.85±1.2066.53±0.72
2551.30±0.9261.6±0.9052.07±0.4362.64±1.30
AME565.02±0.6470.61±0.2966.76±0.8470.54±0.31
1058.27±0.6366.52±0.4661.54±1.3667.77±0.71
2552.69±0.9264.10±0.8954.84±1.7865.10±1.35
", + "image_path": "717fff577a8b71b5bf6bd3e7698077ebe90c1a0c189ed8fd22791af3be08ebfe.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 182, + 545, + 215 + ], + "lines": [ + { + "bbox": [ + 306, + 182, + 545, + 215 + ], + "spans": [ + { + "bbox": [ + 306, + 182, + 545, + 215 + ], + "type": "text", + "content": "Table 5. The average incremental accuracy by using different classifiers (CNN, " + }, + { + "bbox": [ + 306, + 182, + 545, + 215 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 306, + 182, + 545, + 215 + ], + "type": "text", + "content": " -NME, AME) with 20 exemplars per class. Values are evaluated on CIFAR-100 with 5, 10, and 25 incremental steps." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 305, + 228, + 545, + 285 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 228, + 545, + 285 + ], + "spans": [ + { + "bbox": [ + 305, + 228, + 545, + 285 + ], + "type": "text", + "content": "izes all samples. Tab. 5 shows our approach achieves the best performance when using AME classifier. Moreover, it is noteworthy that our approach enhances performance of baselines using all three classifiers by a large margin in various numbers of tasks." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 287, + 545, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 287, + 545, + 335 + ], + "spans": [ + { + "bbox": [ + 305, + 287, + 545, + 335 + ], + "type": "text", + "content": "Sensitivity of momentum coefficient. Fig. 7 shows our approach is not sensitive to the momentum coefficient of dictionary updating for all three classifiers. We set " + }, + { + "bbox": [ + 305, + 287, + 545, + 335 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 305, + 287, + 545, + 335 + ], + "type": "text", + "content": " to 0.9 for other experiments of our approach." + } + ] + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 314, + 348, + 444, + 447 + ], + "blocks": [ + { + "bbox": [ + 314, + 348, + 444, + 447 + ], + "lines": [ + { + "bbox": [ + 314, + 348, + 444, + 447 + ], + "spans": [ + { + "bbox": [ + 314, + 348, + 444, + 447 + ], + "type": "image", + "image_path": "e44ae4c03db8c8a32c9050f240ef0ef953fcdcc0051c70a4ca7dc6d91b445f8c.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 466, + 346, + 545, + 444 + ], + "lines": [ + { + "bbox": [ + 466, + 346, + 545, + 444 + ], + "spans": [ + { + "bbox": [ + 466, + 346, + 545, + 444 + ], + "type": "text", + "content": "Figure 7. Experiments on sensitivity of momentum coefficient. Average accuracies are evaluated on CIFAR-100 with 5 incremental steps and 20 exemplars per class." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 459, + 383, + 471 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 459, + 383, + 471 + ], + "spans": [ + { + "bbox": [ + 306, + 459, + 383, + 471 + ], + "type": "text", + "content": "6. Conclusions" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 479, + 545, + 657 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 479, + 545, + 657 + ], + "spans": [ + { + "bbox": [ + 304, + 479, + 545, + 657 + ], + "type": "text", + "content": "To the best of our knowledge, this work is the first to consider a new type of bias - task-induced bias in a causal perspective, which is derived from special settings of continual learning. We analyzed three continual scenarios in a causal framework, and found that the task-induced bias is particularly detrimental to CIL. Based on this framework, we designed a causal intervention operation and implemented it as a causal debias module by exploiting a powerful attention mechanism. Comprehensive experiments demonstrated our plugin method can boost performance of baselines and outperform several existing strong methods. In the future, we think it is a promising direction to regard the task-induced bias as a problem in CIL and design more exquisite approaches to eliminate it by cutting the corresponding causal path." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 305, + 658, + 545, + 702 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 658, + 545, + 702 + ], + "spans": [ + { + "bbox": [ + 305, + 658, + 545, + 702 + ], + "type": "text", + "content": "Acknowledgments. This work was supported in part by National Key R&D Program of China (2021ZD0112001) and National Natural Science Foundation of China (No.61831005, 61971095 and 62271119)." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16023" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 146 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 146 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 146 + ], + "type": "text", + "content": "[1] Ehsan Abbasnejad, Damien Teney, Amin Parvaneh, Javen Shi, and Anton van den Hengel. Counterfactual vision and language learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10041—10051, 2020. 2, 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 148, + 288, + 192 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 148, + 288, + 192 + ], + "spans": [ + { + "bbox": [ + 53, + 148, + 288, + 192 + ], + "type": "text", + "content": "[2] Hongjoon Ahn, Jihwan Kwak, Subin Lim, Hyeonsu Bang, Hyojun Kim, and Taesup Moon. SS-IL: Separated softmax for incremental learning. In Proceedings of the IEEE/CVF Conference on Computer Vision, pages 844–853, 2021. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 194, + 288, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 194, + 288, + 237 + ], + "spans": [ + { + "bbox": [ + 53, + 194, + 288, + 237 + ], + "type": "text", + "content": "[3] Arjun Ashok, K. J. Joseph, and Vineeth Balasubramanian. Class-incremental learning with cross-space clustering and controlled transfer. In European Conference on Computer Vision, 2022. 1, 2, 6, 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 239, + 288, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 239, + 288, + 293 + ], + "spans": [ + { + "bbox": [ + 53, + 239, + 288, + 293 + ], + "type": "text", + "content": "[4] Matthias Delange, Rahaf Aljundi, Marc Masana, Sarah Parisot, Xu Jia, Ales Leonardis, Greg Slabaugh, and Tinne Tuytelaars. A continual learning survey: Defying forgetting in classification tasks. Transactions on Pattern Analysis and Machine Intelligence, 44(7):3366-3385, 2021. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 295, + 288, + 349 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 295, + 288, + 349 + ], + "spans": [ + { + "bbox": [ + 53, + 295, + 288, + 349 + ], + "type": "text", + "content": "[5] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. ImageNet: A large-scale hierarchical image database. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 248-255, 2009. 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 351, + 288, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 351, + 288, + 396 + ], + "spans": [ + { + "bbox": [ + 53, + 351, + 288, + 396 + ], + "type": "text", + "content": "[6] Xiang Deng and Zhongfei Zhang. Comprehensive knowledge distillation with causal intervention. In Advances in Neural Information Processing Systems, pages 22158-22170, 2021. 2, 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 397, + 288, + 452 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 397, + 288, + 452 + ], + "spans": [ + { + "bbox": [ + 53, + 397, + 288, + 452 + ], + "type": "text", + "content": "[7] Chen He, Ruiping Wang, and Xilin Chen. A tale of two CILs: The connections between class incremental learning and class imbalanced learning, and beyond. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshop, pages 3554-3564, 2021. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 453, + 288, + 497 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 453, + 288, + 497 + ], + "spans": [ + { + "bbox": [ + 53, + 453, + 288, + 497 + ], + "type": "text", + "content": "[8] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2016. 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 498, + 288, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 498, + 288, + 542 + ], + "spans": [ + { + "bbox": [ + 53, + 498, + 288, + 542 + ], + "type": "text", + "content": "[9] Lisa Anne Hendricks, Kaylee Burns, Kate Saenko, Trevor Darrell, and Anna Rohrbach. Women also snowboard: Overcoming bias in captioning models. In European Conference on Computer Vision, 2018. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 544, + 288, + 576 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 544, + 288, + 576 + ], + "spans": [ + { + "bbox": [ + 48, + 544, + 288, + 576 + ], + "type": "text", + "content": "[10] Geoffrey Hinton, Oriol Vinyals, and Jeff Dean. Distilling the knowledge in a neural network. In ArXiv: 1503.02531, 2015. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 578, + 288, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 578, + 288, + 633 + ], + "spans": [ + { + "bbox": [ + 48, + 578, + 288, + 633 + ], + "type": "text", + "content": "[11] Saihui Hou, Xinyu Pan, Chen Change Loy, Zilei Wang, and Dahua Lin. Learning a unified classifier incrementally via rebalancing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 831-839, 2019. 1, 2, 6, 7, 8" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 635, + 288, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 635, + 288, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 635, + 288, + 689 + ], + "type": "text", + "content": "[12] Xinting Hu, Kaihua Tang, Chunyan Miao, Xian-Sheng Hua, and Hanwang Zhang. Distilling causal effect of data in class-incremental learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3956-3965, 2021. 1, 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 691, + 288, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 288, + 713 + ], + "type": "text", + "content": "[13] K. J. Joseph, Salman Khan, Fahad Shahbaz Khan, Rao Muhammad Anwer, and Vineeth N. Balasubramanian." + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 106 + ], + "type": "text", + "content": "Energy-based latent aligner for incremental learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7452-7461, 2022. 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 108, + 545, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 108, + 545, + 162 + ], + "spans": [ + { + "bbox": [ + 308, + 108, + 545, + 162 + ], + "type": "text", + "content": "[14] Minsoo Kang, Jaeyoo Park, and Bohyung Han. Class incremental learning by knowledge distillation with adaptive feature consolidation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16071-16080, 2022. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 163, + 545, + 186 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 163, + 545, + 186 + ], + "spans": [ + { + "bbox": [ + 308, + 163, + 545, + 186 + ], + "type": "text", + "content": "[15] Alex Kriszhevsky. Learning multiple layers of features from tiny images. In Technical report, 2009. 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 186, + 545, + 219 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 186, + 545, + 219 + ], + "spans": [ + { + "bbox": [ + 308, + 186, + 545, + 219 + ], + "type": "text", + "content": "[16] Zhizhong Li and Derek Hoiem. Learning without forgetting. Transactions on Pattern Analysis and Machine Intelligence, 40(12):2935-2947, 2018. 1, 6, 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 220, + 545, + 276 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 220, + 545, + 276 + ], + "spans": [ + { + "bbox": [ + 308, + 220, + 545, + 276 + ], + "type": "text", + "content": "[17] Chengzhi Mao, Kevin Xia, James Wang, Hao Wang, Junfeng Yang, Elias Bareinboim, and Carl Vondrick. Causal transportability for visual recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7521-7531, 2022. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 277, + 545, + 321 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 277, + 545, + 321 + ], + "spans": [ + { + "bbox": [ + 308, + 277, + 545, + 321 + ], + "type": "text", + "content": "[18] Michael McCloskey and Neal J. Cohen. Catastrophic interference in connectionist networks: The sequential learning problem. In *Psychology of Learning and Motivation*, volume 24, pages 109–165. Academic Press, 1989. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 322, + 545, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 322, + 545, + 365 + ], + "spans": [ + { + "bbox": [ + 308, + 322, + 545, + 365 + ], + "type": "text", + "content": "[19] Ninareh Mehrabi, Fred Morstatter, Nripsuta Saxena, Kristina Lerman, and Aram Galstyan. A survey on bias and fairness in machine learning. ACM Computing Surveys, 54(6):1-35, 2021. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 367, + 545, + 465 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 367, + 545, + 465 + ], + "spans": [ + { + "bbox": [ + 308, + 367, + 545, + 465 + ], + "type": "text", + "content": "[20] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, Alban Desmaison, Andreas Köpf, Edward Yang, Zach DeVito, Martin Raison, Alykhan Tejani, Sasank Chilamkurthy, Benoit Steiner, Lu Fang, Junjie Bai, and Soumith Chintala. PyTorch: An imperative style, high-performance deep learning library. In Advances in Neural Information Processing Systems, 2019. 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 467, + 545, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 467, + 545, + 498 + ], + "spans": [ + { + "bbox": [ + 308, + 467, + 545, + 498 + ], + "type": "text", + "content": "[21] Judea Pearl, Madelyn Glymour, and Nicholas P. Jewell. Causal Inference In Statistics: A primer. Wiley, 2016. 3, 4" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 500, + 545, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 500, + 545, + 555 + ], + "spans": [ + { + "bbox": [ + 308, + 500, + 545, + 555 + ], + "type": "text", + "content": "[22] Sylvestre Alvise Rebuffi, Alexander Kolesnikov, Georg Sperl, and Christoph H. Lampert. iCaRL: Incremental classifier and representation learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5533-5542, 2017. 1, 2, 6, 7" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 557, + 545, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 557, + 545, + 624 + ], + "spans": [ + { + "bbox": [ + 308, + 557, + 545, + 624 + ], + "type": "text", + "content": "[23] Yujun Shi, Kuangqi Zhou, Jian Liang, Zihang Jiang, Jiashi Feng, Philip H. S. Torr, Song Bai, and Vincent Y. F. Tan. Mimicking the oracle: An initial phase decorrelation approach for class incremental learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16722-16731, 2022. 2, 6, 7" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 624, + 545, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 624, + 545, + 669 + ], + "spans": [ + { + "bbox": [ + 308, + 624, + 545, + 669 + ], + "type": "text", + "content": "[24] Christian Simon, Piotr Koniusz, and Mehrtash Harandi. On learning the geodesic path for incremental learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1591-1600, 2021. 2, 6, 7" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 670, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 670, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 670, + 545, + 713 + ], + "type": "text", + "content": "[25] Qing Sun, Fan Lyu, Fanhua Shang, Wei Feng, and Liang Wan. Exploring example influence in continual learning. In Advances in Neural Information Processing Systems, 2022. 2" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "16024" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 541 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "text", + "content": "[26] Rishabh Tiwari, Krishnateja Killamsetty, Rishabh Iyer, and Pradeep Shenoy. GCR: Gradient coreset based replay buffer selection for continual learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 99-108, 2022. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 129, + 287, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 129, + 287, + 150 + ], + "spans": [ + { + "bbox": [ + 49, + 129, + 287, + 150 + ], + "type": "text", + "content": "[27] Gido M van de Ven and Andreas S Tolias. Three scenarios for continual learning. In arXiv: 1904.07734, 2019. 1, 2, 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 152, + 287, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 152, + 287, + 205 + ], + "spans": [ + { + "bbox": [ + 48, + 152, + 287, + 205 + ], + "type": "text", + "content": "[28] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Advances in Neural Information Processing Systems, page 5999-6009, 2017. 2, 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 207, + 287, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 207, + 287, + 251 + ], + "spans": [ + { + "bbox": [ + 48, + 207, + 287, + 251 + ], + "type": "text", + "content": "[29] Tan Wang, Jianqiang Huang, Hanwang Zhang, and Qianru Sun. Visual commonsense R-CNN. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2020. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 253, + 287, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 253, + 287, + 274 + ], + "spans": [ + { + "bbox": [ + 48, + 253, + 287, + 274 + ], + "type": "text", + "content": "[30] Max Welling. Herding dynamical weights to learn. In International Conference on Machine Learning, 2009. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 276, + 287, + 329 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 276, + 287, + 329 + ], + "spans": [ + { + "bbox": [ + 48, + 276, + 287, + 329 + ], + "type": "text", + "content": "[31] Yue Wu, Yinpeng Chen, Lijuan Wang, Yuancheng Ye, Zicheng Liu, Yandong Guo, and Yun Fu. Large scale incremental learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 374-382, 2019. 1, 2, 6, 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 331, + 287, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 331, + 287, + 396 + ], + "spans": [ + { + "bbox": [ + 48, + 331, + 287, + 396 + ], + "type": "text", + "content": "[32] Kelvin Xu, Jimmy Lei Ba, Ryan Kiros, Kyunghyun Cho, Aaron Courville, Ruslan Salakhutdinov, Richard S. Zemel, and Yoshua Bengio. Show, attend and tell: Neural image caption generation with visual attention. In International Conference on Machine Learning, pages 2048-2057, 2015. 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 398, + 287, + 452 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 398, + 287, + 452 + ], + "spans": [ + { + "bbox": [ + 48, + 398, + 287, + 452 + ], + "type": "text", + "content": "[33] Shipeng Yan, Jiangwei Xie, and Xuming He. DER: Dynamically expandable representation for class incremental learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3013-3022, 2021. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 454, + 287, + 486 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 454, + 287, + 486 + ], + "spans": [ + { + "bbox": [ + 48, + 454, + 287, + 486 + ], + "type": "text", + "content": "[34] Xu Yang, Hanwang Zhang, and Jianfei Cai. Deconfounded image captioning: A causal retrospect. Transactions on Pattern Analysis and Machine Intelligence, 2021. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 487, + 287, + 541 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 487, + 287, + 541 + ], + "spans": [ + { + "bbox": [ + 48, + 487, + 287, + 541 + ], + "type": "text", + "content": "[35] Bowen Zhao, Xi Xiao, Guojun Gan, Bin Zhang, and Shu Tao Xia. Maintaining discrimination and fairness in class incremental learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13205-13214, 2020. 1, 2" + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "16025" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Camouflaged Instance Segmentation via Explicit De-Camouflaging/9acac0a8-5352-4eef-91f5-74adb1a1e4b8_content_list.json b/2023/Camouflaged Instance Segmentation via Explicit De-Camouflaging/9acac0a8-5352-4eef-91f5-74adb1a1e4b8_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..d3411a27011c2fa035405ea57c9b9c17a1546235 --- /dev/null +++ b/2023/Camouflaged Instance Segmentation via Explicit De-Camouflaging/9acac0a8-5352-4eef-91f5-74adb1a1e4b8_content_list.json @@ -0,0 +1,2116 @@ +[ + { + "type": "text", + "text": "Camouflaged Instance Segmentation via Explicit De-camouflaging", + "text_level": 1, + "bbox": [ + 151, + 130, + 818, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Naisong Luo $^{1*}$ , Yuwen Pan $^{1*}$ , Rui Sun $^{1}$ , Tianzhu Zhang $^{1,2,3\\dagger}$ , Zhiwei Xiong $^{1,2}$ , Feng Wu $^{1,2}$ $^{1}$ University of Science and Technology of China \n $^{2}$ Institute of Artificial Intelligence, Hefei Comprehensive National Science Center \n $^{3}$ Deep Space Exploration Lab", + "bbox": [ + 127, + 179, + 839, + 251 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{lns6,panyw，issunrui}@mail.ustc.edu.cn，{tzzhang，zwxiong，fengwu}@ustc.edu.cn", + "bbox": [ + 129, + 253, + 831, + 268 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 304, + 312, + 319 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Camouflaged Instance Segmentation (CIS) aims at predicting the instance-level masks of camouflaged objects, which are usually the animals in the wild adapting their appearance to match the surroundings. Previous instance segmentation methods perform poorly on this task as they are easily disturbed by the deceptive camouflage. To address these challenges, we propose a novel De-camouflaging Network (DCNet) including a pixel-level camouflage decoupling module and an instance-level camouflage suppression module. The proposed DCNet enjoys several merits. First, the pixel-level camouflage decoupling module can extract camouflage characteristics based on the Fourier transformation. Then a difference attention mechanism is proposed to eliminate the camouflage characteristics while reserving target object characteristics in the pixel feature. Second, the instance-level camouflage suppression module can aggregate rich instance information from pixels by use of instance prototypes. To mitigate the effect of background noise during segmentation, we introduce some reliable reference points to build a more robust similarity measurement. With the aid of these two modules, our DCNet can effectively model de-camouflaging and achieve accurate segmentation for camouflaged instances. Extensive experimental results on two benchmarks demonstrate that our DCNet performs favorably against state-of-the-art CIS methods, e.g., with more than $5\\%$ performance gains on COD10K and NC4K datasets in average precision.", + "bbox": [ + 75, + 335, + 473, + 744 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 775, + 207, + 790 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In the field of biology, camouflage is defined as a strategy that animals use to adapt their body's appearance (e.g., color and pattern) to match their surroundings in order to achieve concealing and avoid being hunted by predators [37]. Cam-", + "bbox": [ + 76, + 801, + 468, + 863 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/66787f4e73dd444f87b544ef18db1f07f67db8c0b40d93c802038274aed215cc.jpg", + "image_caption": [ + "(a) De-camouflaging Process" + ], + "image_footnote": [], + "bbox": [ + 506, + 301, + 890, + 393 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/2aed844686c373604c0854ccf300539d523e2d5fd9c8d31f6b969f0b767e21b4.jpg", + "image_caption": [ + "(b) Pairwise Similarity" + ], + "image_footnote": [], + "bbox": [ + 511, + 414, + 622, + 521 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/8267c25d9a4ed979ecfd2339ef67b13f8d70e95a14554a94aea56fd96abacd4a.jpg", + "image_caption": [ + "(c) Reference Similarity (Ours)" + ], + "image_footnote": [], + "bbox": [ + 638, + 414, + 751, + 521 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/18279214b20cffc83e28d9e3eb508dc8d9c96ede24b00038523cec3bfa5ebb34.jpg", + "image_caption": [ + "Figure 1. Illustration of our motivation. (a) We propose to extract camouflage characteristics to model explicit de-camouflaging for CIS. (b) Pairwise similarity between prototype and pixel is always erroneous. (c) The prototype-pixel correlation is based on the prototype-reference and pixel-reference similarity distribution, which is more accurate." + ], + "image_footnote": [], + "bbox": [ + 759, + 417, + 879, + 506 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "outflaged Instance Segmentation (CIS) [22] aims at identifying the location and predicting instance-level masks of camouflaged objects, which has attracted more and more attention due to its widespread applications in medical image analysis [11, 44], search-and-rescue work [9] and recreational art [7], etc.", + "bbox": [ + 496, + 656, + 890, + 747 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Despite the tremendous progress in instance segmentation [14, 40, 41], there are few efforts [22, 33] working on Camouflaged Instance Segmentation, as CIS is a more challenging task where most camouflaged instances lack obvious contrast with the background, making general instance segmentation methods work poorly on this task. Recent CIS approaches [22, 33] are generally based on traditional instance segmentation models, either by naively fusing various general instance segmentation models [22] to get better representations, or by directly using global in", + "bbox": [ + 496, + 750, + 892, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 807, + 44 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Equal contribution", + "bbox": [ + 94, + 875, + 202, + 887 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "† Corresponding author", + "bbox": [ + 96, + 887, + 218, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "17918", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "teraction [33]. However, these approaches fail to explore the core of CIS: de-camouflaging, i.e., eliminating camouflage characteristics of the target object. Without explicitly de-camouflaging, the previous methods are easily disturbed by similar background. De-camouflaging is a challenging problem as no image-specific camouflage information is provided as supervision signals. Intuitively, humans have the ability to quickly recognise objects in a highly camouflaged scene. They first repeatedly discriminate the real target characteristics from the camouflage characteristics at the pixel level, and then aggregate the pixel information to discern the whole target instance from the background. This human visual mechanism motivates us to explore de-camouflaging strategy from the pixel level to the instance level in a progressive manner.", + "bbox": [ + 75, + 90, + 472, + 316 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In order to model de-camouflage at both the pixel and instance levels, a series of issues need to be considered. (1) The pixel level de-camouflage. In essence, each pixel feature of a camouflaged image contains the camouflage characteristics and the target object characteristics. Note that our goal is to remove camouflage characteristics and maintain target object information, thus a question naturally arises: How to decouple these two pieces of information at the pixel level? (2) The instance level segmentation. Based on the de-camouflaged pixels, we can naturally aggregate the pixel information and infer the mask of the target instance. Currently, the transformer-based models [5,6] have achieved leading performance for instance segmentation, where instance-specific prototypes are learned by constantly interacting with pixel features for final segmentation. However, directly applying the transformer to CIS is not trivial, as the prototypes would frequently absorb deceptive background information that has high similarity with the objects during the interaction, thus failing to discover desired targets accurately. As proved in Figure 1(b), the prototype fish is more similar to background pixels than target pixels in the camouflaged image. Therefore, we inevitably face another question: How to focus on camouflaged instances to achieve de-camouflaging?", + "bbox": [ + 75, + 321, + 472, + 685 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Motivated by the above discussions, we propose an end-to-end De-camouflaging Network (DCNet) by jointly modeling pixel-level camouflage decoupling and instance-level camouflage suppression for CIS. In the Pixel-level Camouflage Decoupling module (PCD), we focus on decoupling camouflage characteristics and target information fused in the pixel feature. First, we extract camouflage characteristics with the assistance of frequency domain information. The Fourier spectrum amplitude contains low-level statistics [34, 42] (e.g., color and texture of the environment, see Figure 5) that accords with the camouflage characteristics. Based on the obtained description of camouflage characteristics, we propose a novel difference attention mechanism to acquire de-camouflaged pixel fea", + "bbox": [ + 75, + 689, + 472, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "tures. In this mechanism, we calculate the discrepancy between features of the original image and camouflage characteristics (see Figure 1(a)), thereby decoupling the camouflage characteristics and valuable target information while filtering out the background interference. In the Instance-level Camouflage Suppression module (ICS), we aggregate the de-camouflaged pixels to achieve final segmentation and meanwhile mitigate the effect of background noise in prototype-pixel interactions. Specifically, we introduce a set of instance prototypes to capture each camouflaged instance through long-range context-aware interactions. To constrain the interactions to favor target pixels over background pixels, we design a novel reference attention mechanism, where we select de-camouflaged pixels with high contribution to prototypes as reference points (see Figure 1(c)). Then we calculate the similarity of prototype-reference and pixel-reference, respectively, thereby obtaining similarity distributions serving as soft multilabel to measure correlations between prototypes and pixels. Highly similar pixels and prototypes must have consistent similarity distributions. In this way, the soft multilabel-based correlation is more accurate than the normal pairwise correlation, as it benefits from consensus among reliable reference points with a global receptive field. As a result, the correlation noise brought by deceptive background can be suppressed and more effective prototypes that contain rich instance information can be obtained.", + "bbox": [ + 496, + 90, + 893, + 496 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The contributions of our method could be summarized as follows: (1) We propose a novel De-camouflaging Network (DCNet) by jointly modeling pixel-level camouflage decoupling and instance-level camouflage suppression for CIS. (2) We propose two effective designs in DCNet, i.e., difference attention mechanism and reference attention mechanism, which can highlight target information and suppress background interference. (3) Extensive experimental results on two benchmark datasets demonstrate the effectiveness of the proposed method, e.g., with more than $5\\%$ performance gains on COD10K and NC4K datasets in average precision.", + "bbox": [ + 496, + 500, + 893, + 666 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 676, + 640, + 694 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Since camouflaged instance segmentation is a new emerging task with very few papers, we mainly focus on introducing several lines of research in camouflaged object detection and instance segmentation.", + "bbox": [ + 496, + 703, + 893, + 763 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Camouflaged Object Detection", + "text_level": 1, + "bbox": [ + 500, + 771, + 772, + 787 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Camouflaged Object Detection (COD), aiming at localizing camouflaged objects from its background, has a long history in biology and art [10, 23]. Early works [18, 31, 36] in this field focus on distinguishing objects in their camouflaging environment by utilizing handcrafted features, such as texture, boundary, and intensity features. With the significant advances in deep learning (DL), DL-based meth", + "bbox": [ + 496, + 794, + 893, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "17919", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/0c152675a03c59e0981a88f2e0369956a2e5cf63050717d13b6c6bff5619036d.jpg", + "image_caption": [ + "Figure 2. Illustration of the proposed DCNet. Our framework mainly consists of two components: a pixel-level camouflage decoupling module and an instance-level camouflage suppression module. (1) The pixel-level camouflage decoupling module aims at extracting camouflaged characteristics and eliminating camouflaged characteristics while reserving target object characteristics to obtain de-camouflaged pixel features. (2) The instance-level camouflage suppression module is responsible for aggregating instance prototypes to achieve instance segmentation meanwhile suppressing the background noise in prototype-pixel interactions." + ], + "image_footnote": [], + "bbox": [ + 140, + 92, + 831, + 376 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "ods [16, 19, 29, 32, 43, 45, 46] have significantly improved the COD performance through end-to-end learning. Mei et al. [29] developed a distraction mining strategy to benefit the accurate segmentation of the camouflaged object. Pang et al. [32] proposed a mixed-scale triplet network, ZoomNet, to capture objects in complex scenes at different \"zoom\" scales. Zhong et al. [45] was the first to claim the COD task should go beyond the RGB domain and introduced frequency clues to better detect camouflaged objects. Despite the large progress in camouflaged object detection, camouflaged instance segmentation is a rarely studied task. In this paper, inspired by [34, 42, 45], we introduce frequency domain transform to model the camouflage characteristics and then achieve de-camouflaging for camouflaged instance segmentation. To the best of our knowledge, we are the first to explore explicit de-camouflaging for camouflaged instance segmentation.", + "bbox": [ + 75, + 486, + 472, + 743 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. Instance Segmentation", + "text_level": 1, + "bbox": [ + 76, + 755, + 290, + 771 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Instance segmentation is a challenging task as it requires both pixel-level and instance-level mask prediction. Existing instance segmentation approaches can be broadly separated into two paradigms: two-stage methods [14, 17, 21] and one-stage methods [1, 40, 41]. Two-stage methods generally utilize object detectors to generate proposal regions and then segment the mask of each detected object. Based on Faster R-CNN [35], Mask R-CNN [14] generates region-", + "bbox": [ + 75, + 780, + 468, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "of-interests (ROIs) in the first stage and utilizes an extra mask head to acquire the instance segmentation mask. The following works continue to improve the performance for this task in the two-stage manner. Recently, one-stage methods have emerged and driven a new trend in this field. YOLACT [1] combines the prototypes with the mask coefficients to produce instance masks without the detection step. SOLOv2 [41], evolved from SOLO [40], directly decouples the original mask prediction into kernel learning and feature learning to generate final instance segmentation results. Currently, the transformer-based methods [5, 6] introduce instance-specific prototypes to constantly interact with pixel features with the help of some attention mechanisms for final instance segmentation, achieving leading performance. However, the general instance segmentation methods can not directly apply to CIS, as the camouflaged objects present high similarity with the background. Therefore, we base on the one-stage transformer-based paradigm to design an end-to-end De-camouflaging Network, where we introduce reliable reference points to build accurate similarity measurement that is robust to deceptive backgrounds.", + "bbox": [ + 496, + 486, + 893, + 806 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Our Method", + "text_level": 1, + "bbox": [ + 500, + 827, + 630, + 842 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we first present the overall architecture of the proposed De-camouflage Network (DCNet), and then describe each module in detail.", + "bbox": [ + 498, + 854, + 893, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "17920", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Overall Architecture", + "text_level": 1, + "bbox": [ + 76, + 90, + 272, + 104 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our DCNet mainly consists of two modules. (1) The pixel-level camouflage decoupling module aims at extracting camouflage characteristics based on the amplitude information of Fourier spectrum, and eliminating camouflage characteristics via a difference attention mechanism. (2) The instance-level camouflage suppression module is responsible for learning instance prototypes from interactions with pixel features, through a reference attention mechanism to suppress the erroneous correlations caused by deceptive background.", + "bbox": [ + 75, + 113, + 468, + 263 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Pixel-level Camouflage Decoupling", + "text_level": 1, + "bbox": [ + 76, + 273, + 380, + 290 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In order to decouple camouflage characteristics and target information fused in the pixel feature, we first model the representation of camouflage characteristics by exploring Fourier frequency domain. Then we design a difference attention mechanism to remove camouflage characteristics while maintaining target information.", + "bbox": [ + 75, + 297, + 468, + 387 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Camouflage Characteristics Extraction. Generally, the camouflage characteristics are mainly comprised of color and texture information [10, 22]. On the other hand, in the frequency domain, it is known that the amplitude component of Fourier spectrum preserves low-level statistics information. Therefore, we can utilize the amplitude of Fourier spectrum to represent background information in the camouflaged image. Specifically, given a camouflaged image $x \\in \\mathbb{R}^{H \\times W \\times 3}$ , its Fourier transformation $\\mathcal{F}(x)$ is formulated as:", + "bbox": [ + 75, + 388, + 468, + 537 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {F} (x) _ {u, v} = \\sum_ {i = 0} ^ {H - 1} \\sum_ {j = 0} ^ {W - 1} x _ {i, j} e ^ {- J 2 \\pi \\left(\\frac {i}{H} u + \\frac {j}{W} v\\right)}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 130, + 547, + 468, + 589 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $J$ refers to the imaginary unit and each channel of the image is computed independently. The amplitude and phase components are then respectively expressed as:", + "bbox": [ + 76, + 599, + 468, + 646 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {A} (x) _ {u, v} = \\left[ R ^ {2} (x) _ {u, v} + I ^ {2} (x) _ {u, v} \\right] ^ {1 / 2}\n$$\n", + "text_format": "latex", + "bbox": [ + 145, + 657, + 398, + 678 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {P} (x) _ {u, v} = \\arctan \\left[ \\frac {I (x) _ {u , v}}{R (x) _ {u , v}} \\right], \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 147, + 678, + 468, + 714 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $R(x)$ and $I(x)$ represent the real and imaginary part of $\\mathcal{F}(x)$ , respectively. To represent the camouflaged background, we fix the phase to a constant $b$ (e.g., average value) and reconstruct the image with the amplitude information as", + "bbox": [ + 76, + 724, + 468, + 785 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {x} = \\mathcal {F} ^ {- 1} [ \\mathcal {A} (x) _ {u, v} e ^ {- J b} ], \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 187, + 795, + 468, + 813 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathcal{F}^{-1}$ indicates the inverse Fourier transformation. Then, we feed the reconstructed image $\\tilde{x}$ into a light-weight CNN (e.g., ResNet-18 [15]) with global average pooling layer to obtain the global camouflage characteristics $g$ of the original image.", + "bbox": [ + 76, + 824, + 468, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Difference Attention Mechanism. Given image features $\\mathbf{F} \\in \\mathbb{R}^{h \\times w \\times c}$ extracted from a backbone network (e.g., ResNet-50 [15]), we remove the camouflage characteristics $\\pmb{g}$ while retaining valuable target information by a novel difference attention mechanism. In specific, we first utilize two $1 \\times 1$ convolution layers to map $\\mathbf{F}$ and $\\pmb{g}$ to the same dimension $C$ , resulting in $\\hat{\\mathbf{F}} \\in \\mathbb{R}^{h \\times w \\times C}$ and $\\hat{\\pmb{g}} \\in \\mathbb{R}^C$ , respectively. The discrepancy between image features and camouflage characteristics indicates the saliency information of the target objects, helping to highlight the target information. Thus, we integrate the difference on all channels for each pixel to obtain a difference map $\\mathbf{D} \\in \\mathbb{R}^{h \\times w}$ :", + "bbox": [ + 496, + 90, + 890, + 272 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {D} _ {i, j} = \\sum_ {k = 1} ^ {C} \\left(\\hat {\\mathbf {F}} _ {i, j, k} - \\hat {\\boldsymbol {g}} _ {k}\\right) ^ {2}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 604, + 285, + 890, + 325 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $i,j$ , and $k$ are the index of height, width, and channel, respectively. By multiplying the original image features with the difference map $\\mathbf{D}$ , we can get the de-camouflaged pixel features $\\mathbf{F}^d$ as", + "bbox": [ + 496, + 338, + 890, + 398 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {F} ^ {d} = \\mathbf {F} \\circ \\mathbf {D}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 648, + 411, + 890, + 428 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\circ$ refers to the Hadamard product and $\\mathbf{D}$ is broadcasted to the same dimension as $\\mathbf{F}$ . In addition, we follow [33] to use a deformable self-attention layer [47] to further incorporate context information from other pixels.", + "bbox": [ + 496, + 441, + 890, + 502 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Fusion Layer. To acquire fine-grained target information for more accurate segmentation, we use multi-scale features from different stages of the backbone. Specifically, we first obtain the features from different stages of the backbone network, and then obtain the corresponding de-camouflaged features through the difference attention mechanism. Finally, we fuse the processed features with the FPN [24] network to generate the high-resolution pixel-level feature $\\mathbf{E}$ , which is used for subsequent instance segmentation.", + "bbox": [ + 496, + 503, + 890, + 652 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Instance-level Camouflage Suppression", + "text_level": 1, + "bbox": [ + 500, + 666, + 834, + 681 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In order to capture different camouflaged instances, we learn a set of instance prototypes $\\mathbf{P} = \\{\\pmb{p}^i\\}_{i=1}^N$ , and each $\\pmb{p}^i \\in \\mathbb{R}^{1 \\times L}$ is responsible for identifying whether each pixel belongs to a camouflaged instance. To eliminate the background interference for prototype learning, we design reference attention to aggregate target information.", + "bbox": [ + 496, + 689, + 890, + 779 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Reference Attention Mechanism. Given the decamouflaged pixel feature $\\mathbf{F}^d$ derived from the pixel-level camouflage decoupling module, we first obtain the queries $\\mathbf{Q} = [q_1;q_2;\\dots;q_N]$ from instance prototypes $\\mathbf{P}$ , keys $\\mathbf{K} = [k_{1};k_{2};\\dots;k_{hw}]$ and values $\\mathbf{V} = [v_{1};v_{2};\\dots;v_{hw}]$ from pixel features $\\mathbf{F}^d = [f_1;f_2;\\dots;f_{hw}]$ as", + "bbox": [ + 496, + 780, + 890, + 873 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {q} _ {i} = \\boldsymbol {p} _ {i} \\mathbf {W} ^ {q}, \\boldsymbol {k} _ {j} = \\boldsymbol {f} _ {j} \\mathbf {W} ^ {k}, \\boldsymbol {v} _ {j} = \\boldsymbol {f} _ {j} \\mathbf {W} ^ {v}, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 558, + 883, + 890, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "17921", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $i \\in 1,2,\\dots,N, j \\in 1,2,\\dots,hw$ , and $\\mathbf{W}^q \\in \\mathbb{R}^{L \\times L}$ , $\\mathbf{W}^k$ , $\\mathbf{W}^v \\in \\mathbb{R}^{c \\times L}$ are linear projections. With the prototype queries, we can compute the correlation between each query-key pair independently as", + "bbox": [ + 76, + 90, + 468, + 151 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\ns _ {i, j} = \\frac {d \\left(\\boldsymbol {q} _ {i} , \\boldsymbol {k} _ {j}\\right)}{\\sqrt {L}}, \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 210, + 157, + 468, + 190 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $d(\\cdot, \\cdot)$ denotes the distance metric which is the dot product similarity in [39], and $\\sqrt{L}$ is a scaling factor. As the camouflaged object presents high similarity with the surrounding backgrounds, the prototype-pixel correlation represented by $s_{i,j}$ is susceptible to background pixel interference, resulting in erroneous segmentation. Intuitively, humans are able to identify camouflaged instances by repeatedly comparing the target with the surrounding reference areas. Motivated by this, we first use an adaptive selection process to select some reliable pixels from de-camouflaged pixels $\\mathbf{F}^d$ as references. In specific, to measure the reliability of each pixel, we define and calculate the total contribution of each pixel with regard to all the prototypes, which is formulated as:", + "bbox": [ + 75, + 191, + 468, + 402 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nu _ {j} = \\sum_ {i = 1} ^ {N} s _ {i, j}, j \\in 1, 2, \\dots , h w. \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 169, + 407, + 468, + 446 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Then we select top- $K$ pixels with the largest correlations with the prototypes as reference points $\\mathbf{R} \\in \\mathbb{R}^{K \\times L}$ . In this way, we can pick out reliable reference points that are adaptive to the instance prototypes. Based on reference points, we first calculate the prototype-reference and pixel-reference similarity as", + "bbox": [ + 76, + 448, + 468, + 539 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\ns _ {i} ^ {q} = \\phi_ {1} \\left(\\boldsymbol {q} _ {i} \\mathbf {R} ^ {\\top}\\right), s _ {j} ^ {k} = \\phi_ {2} \\left(\\boldsymbol {k} _ {j} \\mathbf {R} ^ {\\top}\\right), \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 155, + 545, + 468, + 566 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\phi_{1}$ and $\\phi_{2}$ are feed-forward networks (FFN) containing two fully connected layers. $s_i^q$ and $s_j^k$ can serve as soft multilabel to measure the correlation between the $i$ -th prototype and the $j$ -th pixel:", + "bbox": [ + 76, + 571, + 468, + 632 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\ns _ {i, j} ^ {q k} = d \\left(\\boldsymbol {q} _ {i}, \\boldsymbol {k} _ {j}; \\mathbf {R}\\right) = s _ {i} ^ {q} \\left(s _ {j} ^ {k}\\right) ^ {\\top}. \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 166, + 638, + 468, + 659 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Compared to $s_{i,j}$ , $s_{i,j}^{qk}$ leverages the consensus among reliable reference points, thus suppressing the background noises and leading to fewer erroneous correlations. Given the correlation $s_{i,j}^{qk}$ , we can extract and purify target information from pixel features to update instance prototypes as", + "bbox": [ + 76, + 667, + 468, + 744 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {\\boldsymbol {p}} _ {i} = \\sum_ {j = 1} ^ {h w} a _ {i, j} \\boldsymbol {v} _ {j}, a _ {i, j} = \\frac {\\exp \\left(s _ {i , j} ^ {q k}\\right)}{\\sum_ {j = 1} ^ {h w} \\exp \\left(s _ {i , j} ^ {q k}\\right)}. \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 133, + 751, + 468, + 794 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Note that we have omitted the scaling factor and the multihead mechanism for notation simplicity.", + "bbox": [ + 76, + 801, + 467, + 830 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Given the high-resolution pixel-level feature embedding $\\mathbf{E}$ and the learned instance prototypes $\\tilde{\\pmb{p}}_i$ , we can obtain N mask predictions and according scores by", + "bbox": [ + 76, + 830, + 468, + 876 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nm _ {i} = \\operatorname {s i g m o i d} \\left(\\mathbf {E} \\tilde {\\boldsymbol {p}} _ {i} ^ {\\top}\\right), y _ {i} = h \\left(\\tilde {\\boldsymbol {p}} _ {i} \\mid \\sigma\\right), \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 143, + 883, + 468, + 902 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $h(\\cdot \\mid \\sigma)$ is a classifier parameterized by $\\sigma$ to predict the confidence score. Following [6], we match a ground truth label (including \"no object\") for each predicted segment by the Hungarian matching algorithm. For training, we use a cross-entropy loss to constrain the instance score $y_{i}$ , and a linear combination of a focal loss [25] and a dice loss [30] to constrain the mask prediction $m_{i}$ .", + "bbox": [ + 498, + 90, + 890, + 196 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiment", + "text_level": 1, + "bbox": [ + 500, + 212, + 624, + 228 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Experiment Setup", + "text_level": 1, + "bbox": [ + 500, + 237, + 674, + 253 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Datasets. We evaluate our method on two datasets: COD10K [10] and NC4K [28]. COD10K contains 5086 camouflaged images with high-quality instance-level annotations, which are divided into 3040 training images and 2026 testing images. NC4K contains 4121 test camouflaged images to evaluate the generalization ability of CIS models. Following the standard benchmark [33], we use the instance-level annotations in COD10K to train our DCNet and evaluate it on the test set of COD10K and NC4K.", + "bbox": [ + 498, + 260, + 890, + 396 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Evaluation Metrics. We adopt $\\mathrm{AP}_{50}$ , $\\mathrm{AP}_{75}$ and AP scores as evaluation metrics to quantify the effectiveness of our method. A true positive is counted if the intersection over union (IoU) between the ground truth and the segmentation is at least $50\\%$ or $75\\%$ , and the AP score is an overall metric that combines different IoU thresholds.", + "bbox": [ + 498, + 397, + 890, + 486 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Implementation Details. For fair comparisons, the feature extractor is implemented by ResNet-50 [15] which is pretrained on ImageNet [8]. During our training, our model is trained with batch size of 2, using the Adam optimizer [27] with an initial learning rate of 0.0001 for 90,000 iterations. We set the channel dimension $C = L = 256$ . In the instance-level camouflage suppression module, we set the number of prototypes $N$ as 10, and the number of reference points $K$ as 64, which turns out to generate the best segmentation during our experiments as shown in Figure 4.", + "bbox": [ + 498, + 488, + 890, + 638 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Comparison with State-of-the-arts", + "text_level": 1, + "bbox": [ + 500, + 650, + 799, + 666 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Quantitative results. Considering that CIS is a newly emerging task, we also compare some general instance segmentation methods besides the previous CIS methods and all the methods adopt ResNet-50 [15] backbone for fair comparisons. As demonstrated in Table 1, the proposed DCNet consistently outperforms the state-of-the-art methods by a large margin on both datasets. (1) On the COD10K [10], our method outperforms the previous best method OSFormer [33] by $4.3\\%$ in AP, indicating that our DCNet is able to acquire more accurate segmentation of camouflaged targets. (2) On the NC4K [28], Our model yields an accuracy of $52.8\\%$ in AP, making an obvious performance improvement by $7.0\\%$ in AP compared to the second best performing method Mask2Former [5]. Since the model is trained on the COD10K training set, the high", + "bbox": [ + 498, + 672, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "17922", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/619d16cc497f147a6af5273e487e1afbee725ad5410c2ac7af14b32b123df0da.jpg", + "table_caption": [ + "Table 1. Comparisons of existing CIS methods and general instance approaches on COD10K [10] and NC4K [28] testing set, where all the methods use ResNet-50 as feature extractor. The best results are shown in bold." + ], + "table_footnote": [], + "table_body": "
MethodsCOD10K-TestNC4K-TestParams(M)GFLOPs
AP\\( AP_{50} \\)\\( AP_{75} \\)AP\\( AP_{50} \\)\\( AP_{75} \\)
Mask R-CNN [14]25.055.520.427.758.622.743.9186.3
MS R-CNN [17]30.157.228.731.058.729.460.0198.5
Cascade R-CNN [2]25.356.121.329.560.824.871.7334.1
HTC [4]28.156.325.129.859.026.676.9331.7
BlendMask [3]28.256.425.227.756.724.235.8233.8
Mask Transfiner [20]28.756.326.429.456.727.244.3185.1
YOLACT [1]24.353.319.732.165.327.9--
CondInst [38]30.663.626.133.467.429.434.1200.1
QueryInst [12]28.560.123.133.066.729.4--
SOTR [13]27.958.724.129.361.025.663.1476.7
SOLOv2 [41]32.563.229.934.465.931.946.2318.7
MaskFormer [6]38.265.137.944.671.945.845.0174.2
Mask2Former [5]39.467.738.545.873.647.543.9241.0
OSFormer [33]41.071.140.842.572.542.346.6324.7
DCNet (ours)45.370.747.552.877.156.553.4207.0
", + "bbox": [ + 99, + 128, + 859, + 398 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/2eedba9b1a6066d056e2d8dfc90eb51b253ac1a95cc6dd10bf908de87eed8ab5.jpg", + "table_caption": [ + "Table 2. Comparison of the different components in pixel-level camouflage decoupling module in terms of AP scores." + ], + "table_footnote": [], + "table_body": "
COD10KNC4K
PCD (ours)45.352.8
-camouflage attribute extractor43.1(-2.2)48.1(-4.7)
-difference attention mechanism44.1(-1.2)51.3(-1.5)
-both 2 components above40.2(-5.1)46.4(-6.4)
", + "bbox": [ + 81, + 460, + 465, + 539 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "performance on NC4K indicates that our method also has better generalization ability. Notice that our method does not achieve the highest performance on $\\mathrm{AP}_{50}$ metric for the NC4K because different from the previous CIS method, we do not use additional Non-maximum suppression (NMS) post-processing to remove redundant predictions. For $\\mathrm{AP}_{50}$ , our more accurate segmentation results cannot increase the number of true positive (TP) under the $50\\%$ IoU threshold, while redundant predictions will increase the number of false positive (FP), thus our method has no obvious advantages. For $\\mathrm{AP}_{75}$ , our high precision segmentation will increase TP with a large margin. The overall metric AP reflects the superiority of our method over other methods. There is a similar trend on NC4K. In addition, the lower GFLOPs of our DCNet compared to state-of-the-art methods indicates that our model achieves high segmentation ability while maintaining high computation efficiency.", + "bbox": [ + 75, + 566, + 468, + 824 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Qualitative Results. As shown in Figure 3, our proposed method is capable of de-camouflaging by using the difference attention mechanism at the pixel level, and also able to identify the accurate locations of target camouflaged objects at the instance level. Compared to previous methods,", + "bbox": [ + 75, + 824, + 468, + 902 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/95be9cc39a0faaaa1bdebaad747ec42d68cb75eb2ffe1646e3cff71239856468.jpg", + "table_caption": [ + "Table 3. Comparisons of different attention mechanisms in the instance-level camouflage suppression module in terms of AP scores." + ], + "table_footnote": [], + "table_body": "
Attention MechanismCOD10KNC4K
cross-attention [39]42.449.5
masked attention [5]44.751.7
reference attention (ours)45.352.8
", + "bbox": [ + 531, + 474, + 857, + 536 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "our DCNet performs better at boundaries of camouflage instances (see the first 4 columns), suppresses distracting background regions (see $5^{th}$ to $6^{th}$ columns), and distinguishes multiple instances well (see the last 2 columns).", + "bbox": [ + 498, + 563, + 890, + 625 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Ablation Study", + "text_level": 1, + "bbox": [ + 500, + 633, + 653, + 651 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We conduct comprehensive ablation studies on COD10K and NC4K to verify the effectiveness of our modules.", + "bbox": [ + 498, + 657, + 890, + 686 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Effectiveness of the Pixel-level Camouflage Decoupling Module. The pixel-level camouflage decoupling module mainly consists of a camouflage attribute extractor and a difference attention mechanism. We validate the importance of each component by removing them one at a time. As shown in Table 2, we replace the proposed camouflage attribute extractor with a learnable vector, and the performance drops by an average of $3.5\\%$ across two datasets. It indicates that the Fourier spectrum amplitude is effective in modeling the camouflage characteristics, verifying that it is hard to learn de-camouflage without image-specific camouflage information supervision. Then we replace the difference attention mechanism with general subtraction operations. The reduced performance indicates that our pro", + "bbox": [ + 496, + 688, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "17923", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/f53f37466a24f47b59e089a7fb594e9755d4c6aec6c3a52f2428c7992f4dcccb.jpg", + "image_caption": [ + "Original image" + ], + "image_footnote": [], + "bbox": [ + 230, + 97, + 300, + 140 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/cc2b5806073f5e7dec46812d7dc1d5de022ed9090ec03822906887b55f861ae3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 307, + 97, + 377, + 140 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/aa62ac7220e556d24e0a874ae62fc6f1a708726f4bb978b10eac62cf9c85b379.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 385, + 97, + 455, + 138 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/988603c351cbf2895a772229e45782d3b2997c6323d9f64448363f36ab3e4ad5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 462, + 97, + 532, + 138 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d38ae4f9e9c4b6e1ffe9ae145f2d9f04165f93cfd477ca225262da6b0f0d379c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 540, + 97, + 611, + 138 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/fa0480c307b9c7548b0b0e97d22923df309fefb571aadf4a4a45668d7b288a98.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 617, + 97, + 689, + 138 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/45e9b494e3843707d4d4d85426bc4c2dfca555fb2392b3efc7505ae9ba17a454.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 696, + 97, + 766, + 138 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/384068335ec1bc22ac0215ca0d2e06e9942bd225a9ec630b89d1b0b4d4fe6c33.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 774, + 97, + 844, + 138 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/19776a8e98bdf84828fd51a0add9ec12917c7da9eeba01121d43ba1526b3256a.jpg", + "image_caption": [ + "GT" + ], + "image_footnote": [], + "bbox": [ + 230, + 145, + 300, + 189 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/4a61791df68fd000b715567e9829a21b7ee5ed297d7467ae1f7e153677956481.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 308, + 146, + 377, + 189 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/7ab5602b2131e8c9e153fee6a1579ecc7b6b920d98493356693995c2b302b92b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 385, + 146, + 455, + 189 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e564528de3fad60a33fe9de07f591c906022f58b251ca2faa60ae4744cf51bae.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 464, + 146, + 532, + 189 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d2e2c9545c62bcedbbfe0eb929a8e720ddb9db2ee910fcf9107e1e198d10b60e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 540, + 146, + 611, + 189 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/50e4dcb024d1f85e9fdd67e5eb3b577b9dc1feb1e9f7c53ca7883b4917da1cd1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 617, + 146, + 689, + 189 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/79376126f34aa029127fd3ce5605754ed4cdf4223f5d396e986e61c7aa33bce3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 696, + 146, + 766, + 189 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e1173728ada22a4838013b3852b45ebaa1e8fdfe9c3726be9dc8882250753765.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 774, + 146, + 844, + 189 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/c0e1d0c951d5130119bcae556b19774997840fde610df1d97d573d86cf653e9d.jpg", + "image_caption": [ + "MaskFormer" + ], + "image_footnote": [], + "bbox": [ + 230, + 195, + 300, + 238 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/db591762c71826ecaa7158c2b12bf95683049006dfebdb304c7e20371b6f4515.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 308, + 195, + 377, + 238 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d8f20f5e6d68073aa3cc050da8834f5a3e3aec5dd2f010e2ec059f69b27b4331.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 385, + 195, + 455, + 238 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/fe5ce7aaafc8662297caa99a7e5b6cb10d9c2e0dba72a5ff72f9437a1e4313cc.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 464, + 195, + 532, + 238 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/8c28960cd6739c9b272ef918b04a8340319234ea4eafe5bc68a854067c0026a1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 540, + 195, + 611, + 238 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/387c34e44e92582e9e53cc71016a1dd0564a869b596af22241cae2ea831d9fad.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 617, + 195, + 689, + 238 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/4362116f388be5ea0ff4b9b641d1388b18a2f9e4008b304b7764ee9dc3975d37.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 696, + 195, + 766, + 238 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/73b5df6e3e23054b0ae25a6f84887078d81c69e2b19f79724f117e1b3c8bce66.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 774, + 195, + 844, + 238 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/992f9eb9b56d378fed232f07294004d8d4a64f9e50ddee93fadad7d53120ce7a.jpg", + "image_caption": [ + "Mask2Former" + ], + "image_footnote": [], + "bbox": [ + 230, + 243, + 300, + 287 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e4c446a0e1b243d0ba4616df93f3fc619aea03f1115eca0b52481f68db755730.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 308, + 243, + 377, + 287 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/52ceae1732bbbf2969d17510fd189d136709c4327dd09a037753090f56946505.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 385, + 243, + 455, + 287 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/0ee51e6fb06d8fe662d524134080c2d8804bca6a13e85567d145f99ede169a97.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 464, + 243, + 532, + 286 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/8938e82d1c75c88eb2d4fff5ebae2a7f54789fdacb59b1c22fad9fd7b06949e5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 540, + 243, + 611, + 286 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d6d11c388168127d855e356382384276441265e12b25a8a8f31a01c9706eef98.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 617, + 243, + 689, + 286 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/338d66e27e360a3afeca42a0b0120f27d8b718e9bca7b3e8ceea22d3e14617ea.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 696, + 243, + 766, + 286 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ee5769b9c410b51a87be24f13381f1194620b63864b1a475a037c2b34e354d86.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 774, + 243, + 844, + 286 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/666cd1fd657c449bf2f31ec1108b8ffa3852d52648d2e20ecd5022e3c1763d18.jpg", + "image_caption": [ + "OSFormer" + ], + "image_footnote": [], + "bbox": [ + 230, + 292, + 300, + 335 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/5cb7de8580d9977f01b0adaf79a660a05c3a6ac29daa5742dbc2200f136aee71.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 308, + 292, + 377, + 335 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/021537d830b0c02cae219be65e3adaa7585fa476e0d3da773c93c9c4969d86d7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 385, + 292, + 455, + 335 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/55fa1b75b3ce068fdaa6436a0d3f9f75c35cea3c544e49fc30eaa020edcf2289.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 464, + 292, + 532, + 335 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a4ab53fc701fdbca6376516bcaf294ceae9bf200c883e8e9fa85b0c2c6886e11.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 540, + 292, + 611, + 335 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/2bda3a983111f15b722135ccaac65e512c88f0e4bb4e2d39131329fdb34a533d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 617, + 292, + 689, + 335 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/9051c92ca773f440375f4221f3ac1b771bad6d22f912f7e73ef76fc8028f5ba0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 696, + 292, + 766, + 335 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f59478999595bfed7d1d79302b3b8a82ca2729b857a336e97f61478a5214b71a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 774, + 292, + 844, + 335 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/9b36f8506c4dbe9fc3f03e3fd04d4bcdcddf6b87e40439ebc2ca55726135cc66.jpg", + "image_caption": [ + "DCNet(ours)" + ], + "image_footnote": [], + "bbox": [ + 230, + 340, + 300, + 383 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a371356dedced26ac5dd6fe7c380ff55ce48f7d2f9f95ffcea356fddf7a01ef6.jpg", + "image_caption": [ + "Figure 3. Qualitative results of DCNet compared with state-of-the-art methods. Different colored masks represent different instances." + ], + "image_footnote": [], + "bbox": [ + 308, + 340, + 377, + 383 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/3d6a4e5dc3d01faf1fd49cafce8039f219c4833f256bbe2f6ccc2973f2383142.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 385, + 340, + 455, + 383 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/7f795e4059003961d5fee101a687522949be5dbadb252c67f2a3432699780fac.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 464, + 340, + 532, + 383 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ad88d32e41c940cf5ae7d250161964e1bdbca807d485b16e9cfddbfc7d95bc40.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 540, + 340, + 611, + 383 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/5dfb3dadd3b9d42bd8480a69574718a5938f8dacc59d4eb040d8a1f25e8f9f33.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 617, + 340, + 689, + 383 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ee0272c1bb62fcfd5c28dd059da0e56ffdbcbbe9a7f60709c12aeda8d4f97f80.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 696, + 340, + 766, + 383 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/20300b7141a95ed2f7c01422085b07c13de7d9a89d8b79008aa86e8d6d01a06d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 774, + 340, + 844, + 383 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/7e5ed0fb43e9385b2a3f82c32f1257101c7adf9f04c629983a92f62632611913.jpg", + "table_caption": [ + "Table 4. Comparison of using different scales of features in terms of AP scores." + ], + "table_footnote": [], + "table_body": "
COD10KNC4K
single scale (1/32)45.252.5
single scale (1/16)44.851.6
single scale (1/8)45.152.2
multiple scales45.352.8
", + "bbox": [ + 117, + 465, + 423, + 551 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "posed mechanism can well model the discrepancy between image features and camouflage characteristics. Removing both of the components leads to huge performance degradation, showing that the combination of them is essential in improving the camouflaged object segmentation.", + "bbox": [ + 75, + 575, + 468, + 652 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Effectiveness of the Instance-level Camouflage Suppression Module. As shown in Table 3, the reference attention mechanism in our instance-level camouflage suppression module has the highest AP score on both datasets. Compared to vanilla cross-attention [39], our proposed reference points can effectively suppress the noise brought by background pixels in the prototype-pixel interaction, thus obtaining significant improvement for accurate segmentation. Meanwhile, our method is superior to existing cross-attention variants, such as the masked attention [5] in the segmentation field.", + "bbox": [ + 75, + 655, + 468, + 821 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Effectiveness of Hierarchical Features To explore decamouflage at multiple feature levels, we feed 3 scales of features (i.e., stride of 32, 16, and 8) into the difference attention mechanism, respectively. As shown in Table 4, low-resolution features exceed other scales. This may attribute", + "bbox": [ + 75, + 824, + 470, + 902 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/f5d27f8f0fbffe75ca8ace940cc2276151ca761184d0540b4ef4d94b23246df8.jpg", + "table_caption": [ + "Table 5. Comparisons with OSFormer [33] of different backbones on COD10K [10] and NC4K [28] in terms of AP scores." + ], + "table_footnote": [], + "table_body": "
MethodBackboneCOD10KNC4K
OSFormerR5041.042.5
DCNet (ours)45.352.8
OSFormerR10142.044.4
DCNet (ours)46.853.5
OSFormerSwin-T47.750.2
DCNet (ours)50.356.3
OSFormerSwin-S52.156.7
DCNet (ours)52.358.4
", + "bbox": [ + 542, + 465, + 846, + 611 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/bbf4c775b87a1d4b877c62d8215680e0c3e8a79611a80ab254b31c89474a0b8b.jpg", + "image_caption": [ + "Figure 4. Comparisons of performance with different numbers of prototypes and reference points on COD10K [10] and NC4K [28]." + ], + "image_footnote": [], + "bbox": [ + 501, + 622, + 691, + 718 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a9381e3e2a2f3d0974eb78b27e3b87aaf17951f8f8f1d381786999070e2e9310.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 697, + 622, + 890, + 718 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "to the retention of object semantic information. The highest performance in this experiment shows that the contribution of each scale is complementary to each other. Using single-scale features alone is not enough to model the target information, thus combining them together can further boost the performance.", + "bbox": [ + 496, + 777, + 890, + 868 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Compatibility for Feature Extractor. We equip our DC-Net with different feature extractor backbones, i.e., ResNet", + "bbox": [ + 498, + 869, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "17924", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b56eaac3c89497121214786ee85ee89972233eca8006d645d650bd0ac78b69f6.jpg", + "image_caption": [ + "Figure 5. Illustration of the amplitude images and difference maps in our proposed difference attention." + ], + "image_footnote": [], + "bbox": [ + 81, + 89, + 467, + 297 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "50 [15], ResNet-101 [15], Swin-T [26], and Swin-S [26]. All of them are pre-trained on ImageNet-1K [8]. As shown in Table 5, our method performs the best in all cases on both datasets, proving that our approach has tremendous potential for further improvement.", + "bbox": [ + 75, + 347, + 468, + 422 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Analysis of Hyperparameters. In the instance-level camouflage suppression module, prototypes play an important role in capturing instances by aggregating object pixels. As shown in Figure 4(a), we conduct quantitative experiments to analyze how many prototypes are suitable for instance segmentation. It can be seen that too many or too few prototypes will damage the AP metric. In fact, the number of prototypes $N$ should be larger than the count of instances in an image to avoid instance fusion error. While superabundant prototypes will cause many prototypes to match \"no object\" during training. We choose $N = 10$ by balancing the performance on both datasets. Besides, in the reference attention mechanism, the number of reference points $K$ influences the similarity distributions between prototypes and pixels. As shown in Figure 4(b), we can observe that the performance on NC4K continues to grow until $K = 64$ , which means that it is sufficient for building appropriate prototype-pixel correction.", + "bbox": [ + 75, + 424, + 472, + 698 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4. Visualizations", + "text_level": 1, + "bbox": [ + 76, + 708, + 220, + 724 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Visualization of Difference Map. As shown in Figure 5, the $2^{nd}$ row exhibits the amplitude images that are generated through Fourier transformation, which can be considered as camouflage characteristics. With the help of the difference attention mechanism, our network can remove the confusing background and localize the accurate salient areas (see $3^{rd}$ row), which turn out to be the correct regions of camouflaged instances.", + "bbox": [ + 75, + 732, + 468, + 853 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Visualization of Reference Attention. To explore the reference attention mechanism's ability to suppress the background noise, we make qualitative visualization to compare", + "bbox": [ + 75, + 854, + 470, + 902 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/36ae8d4fc95b96686bddad08067ee335e0d265ebd8cdfdf1931d64641717b526.jpg", + "image_caption": [ + "Figure 6. Visualizations of the activation map with different attention mechanisms." + ], + "image_footnote": [], + "bbox": [ + 504, + 90, + 890, + 290 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "different attention mechanisms. As shown in Figure 6 (d), with the proposed reference attention, the high response is clustered in the foreground area and the prototype pays little attention to the background area. In contrast, in Figure 6 (b), prototypes with the cross-attention [39] tend to absorb pixels from the non-target region. And in Figure 6 (c), with the masked attention [5], most background pixels are masked, while some foreground pixels are also suppressed, leading to sub-optimal performance. Thus, the reference mechanism plays an important role in mitigating the effect of background noise in prototype-pixel interactions, which helps to achieve more accurate segmentation.", + "bbox": [ + 496, + 349, + 893, + 532 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 560, + 617, + 575 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we propose an end-to-end Decamouflaging Network (DCNet) by jointly modeling pixel-level camouflage decoupling and instance-level camouflage suppression for camouflaged instance segmentation. We design a Pixel-level Camouflage Decoupling module to model and further eliminate the camouflage characteristics based on the frequency domain information. Besides, an Instance-level Camouflage Suppression module is proposed to achieve prototype-based instance segmentation and mitigate the effect of background noise by introducing reference points. Extensive experimental results on two benchmark datasets demonstrate the effectiveness of the proposed method.", + "bbox": [ + 496, + 584, + 890, + 780 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Acknowledgment", + "text_level": 1, + "bbox": [ + 500, + 809, + 669, + 827 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This work was partially supported by the National Nature Science Foundation of China (Grant 62022078, 62021001), National Defense Basic Scientific Research Program of China (Grant JCKY2021601B013).", + "bbox": [ + 498, + 839, + 893, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "17925", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Daniel Bolya, Chong Zhou, Fanyi Xiao, and Yong Jae Lee. Yolact: Real-time instance segmentation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9157-9166, 2019. 3, 6", + "[2] Zhaowei Cai and Nuno Vasconcelos. Cascade r-cnn: high quality object detection and instance segmentation. IEEE transactions on pattern analysis and machine intelligence, 43(5):1483-1498, 2019. 6", + "[3] Hao Chen, Kunyang Sun, Zhi Tian, Chunhua Shen, Yongming Huang, and Youliang Yan. Blendmask: Top-down meets bottom-up for instance segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8573-8581, 2020. 6", + "[4] Kai Chen, Jiangmiao Pang, Jiaqi Wang, Yu Xiong, Xiaoxiao Li, Shuyang Sun, Wansen Feng, Ziwei Liu, Jianping Shi, Wanli Ouyang, et al. Hybrid task cascade for instance segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4974-4983, 2019. 6", + "[5] Bowen Cheng, Ishan Misra, Alexander G Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1290–1299, 2022. 2, 3, 5, 6, 7, 8", + "[6] Bowen Cheng, Alex Schwing, and Alexander Kirillov. Per-pixel classification is not all you need for semantic segmentation. Advances in Neural Information Processing Systems, 34:17864-17875, 2021. 2, 3, 5, 6", + "[7] Hung-Kuo Chu, Wei-Hsin Hsu, Niloy J Mitra, Daniel Cohen-Or, Tien-Tsin Wong, and Tong-Yee Lee. Camouflage images. ACM Trans. Graph., 29(4):51-1, 2010. 1", + "[8] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 5, 8", + "[9] Deng-Ping Fan, Ge-Peng Ji, Ming-Ming Cheng, and Ling Shao. Concealed object detection. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2021. 1", + "[10] Deng-Ping Fan, Ge-Peng Ji, Guolei Sun, Ming-Ming Cheng, Jianbing Shen, and Ling Shao. Camouflaged object detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2777-2787, 2020. 2, 4, 5, 6, 7", + "[11] Deng-Ping Fan, Tao Zhou, Ge-Peng Ji, Yi Zhou, Geng Chen, Huazhu Fu, Jianbing Shen, and Ling Shao. Inf-net: Automatic Covid-19 lung infection segmentation from ct images. IEEE Transactions on Medical Imaging, 39(8):2626-2637, 2020. 1", + "[12] Yuxin Fang, Shusheng Yang, Xinggang Wang, Yu Li, Chen Fang, Ying Shan, Bin Feng, and Wenyu Liu. Instances as queries. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6910-6919, 2021. 6", + "[13] Ruohao Guo, Dantong Niu, Liao Qu, and Zhenbo Li. Sotr: Segmenting objects with transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7157-7166, 2021. 6" + ], + "bbox": [ + 78, + 114, + 470, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[14] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In Proceedings of the IEEE international conference on computer vision, pages 2961-2969, 2017. 1, 3, 6", + "[15] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 4, 5, 8", + "[16] Xiaobin Hu, Deng-Ping Fan, Xuebin Qin, Hang Dai, Wenqi Ren, Ying Tai, Chengjie Wang, and Ling Shao. High-resolution iterative feedback network for camouflaged object detection. arXiv preprint arXiv:2203.11624, 2022. 3", + "[17] Zhaojin Huang, Lichao Huang, Yongchao Gong, Chang Huang, and Xinggang Wang. Mask scoring r-cnn. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6409-6418, 2019. 3, 6", + "[18] Ivan Huerta, Daniel Rowe, Mikhail Mozerov, and Jordi González. Improving background subtraction based on a casuistry of colour-motion segmentation problems. In Iberian Conference on Pattern Recognition and Image Analysis, pages 475-482. Springer, 2007. 2", + "[19] Ge-Peng Ji, Deng-Ping Fan, Yu-Cheng Chou, Dengxin Dai, Alexander Liniger, and Luc Van Gool. Deep gradient learning for efficient camouflaged object detection. Machine Intelligence Research, 20(1):92-108, 2023. 3", + "[20] Lei Ke, Martin Danelljan, Xia Li, Yu-Wing Tai, Chi-Keung Tang, and Fisher Yu. Mask transfiner for high-quality instance segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4412–4421, 2022. 6", + "[21] Alexander Kirillov, Yuxin Wu, Kaiming He, and Ross Girshick. Pointrend: Image segmentation as rendering. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9799-9808, 2020. 3", + "[22] Trung-Nghia Le, Yubo Cao, Tan-Cong Nguyen, Minh-Quan Le, Khanh-Duy Nguyen, Thanh-Toan Do, Minh-Triet Tran, and Tam V Nguyen. Camouflaged instance segmentation in-the-wild: Dataset, method, and benchmark suite. IEEE Transactions on Image Processing, 31:287-300, 2021. 1, 4", + "[23] Trung-Nghia Le, Tam V Nguyen, Zhongliang Nie, MinhTriet Tran, and Akihiro Sugimoto. Anabranch network for camouflaged object segmentation. Computer Vision and Image Understanding, 184:45-56, 2019. 2", + "[24] Tsung-Yi Lin, Piotr Dólar, Ross Girshick, Kaiming He, Bharath Hariharan, and Serge Belongie. Feature pyramid networks for object detection. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2117-2125, 2017. 4", + "[25] Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dólar. Focal loss for dense object detection. In Proceedings of the IEEE international conference on computer vision, pages 2980-2988, 2017. 5", + "[26] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10012-10022, 2021. 8" + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "17926", + "bbox": [ + 480, + 945, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[27] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 5", + "[28] Yunqiu Lv, Jing Zhang, Yuchao Dai, Aixuan Li, Bowen Liu, Nick Barnes, and Deng-Ping Fan. Simultaneously localize, segment and rank the camouflaged objects. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11591-11601, 2021. 5, 6, 7", + "[29] Haiyang Mei, Ge-Peng Ji, Ziqi Wei, Xin Yang, Xiaopeng Wei, and Deng-Ping Fan. Camouflaged object segmentation with distraction mining. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8772-8781, 2021. 3", + "[30] Fausto Milletari, Nassir Navab, and Seyed-Ahmad Ahmadi. V-net: Fully convolutional neural networks for volumetric medical image segmentation. In 2016 fourth international conference on 3D vision (3DV), pages 565-571. IEEE, 2016. 5", + "[31] Yuxin Pan, Yiwang Chen, Qiang Fu, Ping Zhang, and Xin Xu. Study on the camouflaged target detection method based on 3d convexity. Modern Applied Science, 5(4):152, 2011. 2", + "[32] Youwei Pang, Xiaoqi Zhao, Tian-Zhu Xiang, Lihe Zhang, and Huchuan Lu. Zoom in and out: A mixed-scale triplet network for camouflaged object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2160-2170, 2022. 3", + "[33] Jialun Pei, Tianyang Cheng, Deng-Ping Fan, He Tang, Chuanbo Chen, and Luc Van Gool. Osformer: One-stage camouflaged instance segmentation with transformers. In European Conference on Computer Vision, pages 19-37. Springer, 2022. 1, 2, 4, 5, 6, 7", + "[34] Leon N Piotrowski and Fergus W Campbell. A demonstration of the visual importance and flexibility of spatial-frequency amplitude and phase. Perception, 11(3):337-346, 1982. 2, 3", + "[35] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. Advances in neural information processing systems, 28, 2015. 3", + "[36] P Sengottuvelan, Amitabh Wahi, and A Shanmugam. Performance of decamouflaging through exploratory image analysis. In 2008 First International Conference on Emerging Trends in Engineering and Technology, pages 6-10. IEEE, 2008. 2", + "[37] Sujit K Singh, Chitra A Dhawale, and Sanjay Misra. Survey of object detection methods in camouflaged image. IERI Procedia, 4:351-357, 2013. 1", + "[38] Zhi Tian, Chunhua Shen, and Hao Chen. Conditional convolutions for instance segmentation. In European conference on computer vision, pages 282-298. Springer, 2020. 6", + "[39] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 5, 6, 7, 8", + "[40] Xinlong Wang, Tao Kong, Chunhua Shen, Yuning Jiang, and Lei Li. Solo: Segmenting objects by locations. In European Conference on Computer Vision, pages 649-665. Springer, 2020. 1, 3" + ], + "bbox": [ + 78, + 90, + 468, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[41] Xinlong Wang, Rufeng Zhang, Tao Kong, Lei Li, and Chunhua Shen. Solov2: Dynamic and fast instance segmentation. Advances in Neural information processing systems, 33:17721-17732, 2020. 1, 3, 6", + "[42] Qinwei Xu, Ruipeng Zhang, Ya Zhang, Yanfeng Wang, and Qi Tian. A fourier-based framework for domain generalization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14383-14392, 2021. 2, 3", + "[43] Fan Yang, Qiang Zhai, Xin Li, Rui Huang, Ao Luo, Hong Cheng, and Deng-Ping Fan. Uncertainty-guided transformer reasoning for camouflaged object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4146-4155, 2021. 3", + "[44] Xiaoqi Zhao, Lihe Zhang, and Huchuan Lu. Automatic polyp segmentation via multi-scale subtraction network. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 120-130. Springer, 2021. 1", + "[45] Yijie Zhong, Bo Li, Lv Tang, Senyun Kuang, Shuang Wu, and Shouhong Ding. Detecting camouflaged object in frequency domain. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4504-4513, 2022. 3", + "[46] Tao Zhou, Yi Zhou, Chen Gong, Jian Yang, and Yu Zhang. Feature aggregation and propagation network for camouflaged object detection. IEEE Transactions on Image Processing, 31:7036-7047, 2022. 3", + "[47] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable detr: Deformable transformers for end-to-end object detection. arXiv preprint arXiv:2010.04159, 2020. 4" + ], + "bbox": [ + 501, + 90, + 890, + 541 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "17927", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2023/Camouflaged Instance Segmentation via Explicit De-Camouflaging/9acac0a8-5352-4eef-91f5-74adb1a1e4b8_model.json b/2023/Camouflaged Instance Segmentation via Explicit De-Camouflaging/9acac0a8-5352-4eef-91f5-74adb1a1e4b8_model.json new file mode 100644 index 0000000000000000000000000000000000000000..b2f585e79ad1c2ba398ac80d12974598ff469dca --- /dev/null +++ b/2023/Camouflaged Instance Segmentation via Explicit De-Camouflaging/9acac0a8-5352-4eef-91f5-74adb1a1e4b8_model.json @@ -0,0 +1,2629 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.808, + 0.045 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.152, + 0.131, + 0.819, + 0.154 + ], + "angle": 0, + "content": "Camouflaged Instance Segmentation via Explicit De-camouflaging" + }, + { + "type": "text", + "bbox": [ + 0.129, + 0.18, + 0.84, + 0.252 + ], + "angle": 0, + "content": "Naisong Luo\\(^{1*}\\), Yuwen Pan\\(^{1*}\\), Rui Sun\\(^{1}\\), Tianzhu Zhang\\(^{1,2,3\\dagger}\\), Zhiwei Xiong\\(^{1,2}\\), Feng Wu\\(^{1,2}\\) \n\\(^{1}\\)University of Science and Technology of China \n\\(^{2}\\)Institute of Artificial Intelligence, Hefei Comprehensive National Science Center \n\\(^{3}\\)Deep Space Exploration Lab" + }, + { + "type": "text", + "bbox": [ + 0.13, + 0.255, + 0.833, + 0.269 + ], + "angle": 0, + "content": "{lns6,panyw，issunrui}@mail.ustc.edu.cn，{tzzhang，zwxiong，fengwu}@ustc.edu.cn" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.305, + 0.313, + 0.32 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.337, + 0.474, + 0.745 + ], + "angle": 0, + "content": "Camouflaged Instance Segmentation (CIS) aims at predicting the instance-level masks of camouflaged objects, which are usually the animals in the wild adapting their appearance to match the surroundings. Previous instance segmentation methods perform poorly on this task as they are easily disturbed by the deceptive camouflage. To address these challenges, we propose a novel De-camouflaging Network (DCNet) including a pixel-level camouflage decoupling module and an instance-level camouflage suppression module. The proposed DCNet enjoys several merits. First, the pixel-level camouflage decoupling module can extract camouflage characteristics based on the Fourier transformation. Then a difference attention mechanism is proposed to eliminate the camouflage characteristics while reserving target object characteristics in the pixel feature. Second, the instance-level camouflage suppression module can aggregate rich instance information from pixels by use of instance prototypes. To mitigate the effect of background noise during segmentation, we introduce some reliable reference points to build a more robust similarity measurement. With the aid of these two modules, our DCNet can effectively model de-camouflaging and achieve accurate segmentation for camouflaged instances. Extensive experimental results on two benchmarks demonstrate that our DCNet performs favorably against state-of-the-art CIS methods, e.g., with more than \\(5\\%\\) performance gains on COD10K and NC4K datasets in average precision." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.776, + 0.208, + 0.791 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.802, + 0.47, + 0.864 + ], + "angle": 0, + "content": "In the field of biology, camouflage is defined as a strategy that animals use to adapt their body's appearance (e.g., color and pattern) to match their surroundings in order to achieve concealing and avoid being hunted by predators [37]. Cam-" + }, + { + "type": "image", + "bbox": [ + 0.508, + 0.303, + 0.891, + 0.395 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.634, + 0.397, + 0.758, + 0.407 + ], + "angle": 0, + "content": "(a) De-camouflaging Process" + }, + { + "type": "image", + "bbox": [ + 0.512, + 0.415, + 0.623, + 0.522 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.517, + 0.525, + 0.616, + 0.535 + ], + "angle": 0, + "content": "(b) Pairwise Similarity" + }, + { + "type": "image", + "bbox": [ + 0.64, + 0.415, + 0.752, + 0.522 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.691, + 0.525, + 0.825, + 0.535 + ], + "angle": 0, + "content": "(c) Reference Similarity (Ours)" + }, + { + "type": "image", + "bbox": [ + 0.761, + 0.419, + 0.88, + 0.507 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.551, + 0.892, + 0.633 + ], + "angle": 0, + "content": "Figure 1. Illustration of our motivation. (a) We propose to extract camouflage characteristics to model explicit de-camouflaging for CIS. (b) Pairwise similarity between prototype and pixel is always erroneous. (c) The prototype-pixel correlation is based on the prototype-reference and pixel-reference similarity distribution, which is more accurate." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.657, + 0.892, + 0.748 + ], + "angle": 0, + "content": "outflaged Instance Segmentation (CIS) [22] aims at identifying the location and predicting instance-level masks of camouflaged objects, which has attracted more and more attention due to its widespread applications in medical image analysis [11, 44], search-and-rescue work [9] and recreational art [7], etc." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.893, + 0.902 + ], + "angle": 0, + "content": "Despite the tremendous progress in instance segmentation [14, 40, 41], there are few efforts [22, 33] working on Camouflaged Instance Segmentation, as CIS is a more challenging task where most camouflaged instances lack obvious contrast with the background, making general instance segmentation methods work poorly on this task. Recent CIS approaches [22, 33] are generally based on traditional instance segmentation models, either by naively fusing various general instance segmentation models [22] to get better representations, or by directly using global in" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.875, + 0.204, + 0.888 + ], + "angle": 0, + "content": "*Equal contribution" + }, + { + "type": "page_footnote", + "bbox": [ + 0.098, + 0.888, + 0.22, + 0.9 + ], + "angle": 0, + "content": "† Corresponding author" + }, + { + "type": "list", + "bbox": [ + 0.096, + 0.875, + 0.22, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "17918" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.318 + ], + "angle": 0, + "content": "teraction [33]. However, these approaches fail to explore the core of CIS: de-camouflaging, i.e., eliminating camouflage characteristics of the target object. Without explicitly de-camouflaging, the previous methods are easily disturbed by similar background. De-camouflaging is a challenging problem as no image-specific camouflage information is provided as supervision signals. Intuitively, humans have the ability to quickly recognise objects in a highly camouflaged scene. They first repeatedly discriminate the real target characteristics from the camouflage characteristics at the pixel level, and then aggregate the pixel information to discern the whole target instance from the background. This human visual mechanism motivates us to explore de-camouflaging strategy from the pixel level to the instance level in a progressive manner." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.323, + 0.473, + 0.686 + ], + "angle": 0, + "content": "In order to model de-camouflage at both the pixel and instance levels, a series of issues need to be considered. (1) The pixel level de-camouflage. In essence, each pixel feature of a camouflaged image contains the camouflage characteristics and the target object characteristics. Note that our goal is to remove camouflage characteristics and maintain target object information, thus a question naturally arises: How to decouple these two pieces of information at the pixel level? (2) The instance level segmentation. Based on the de-camouflaged pixels, we can naturally aggregate the pixel information and infer the mask of the target instance. Currently, the transformer-based models [5,6] have achieved leading performance for instance segmentation, where instance-specific prototypes are learned by constantly interacting with pixel features for final segmentation. However, directly applying the transformer to CIS is not trivial, as the prototypes would frequently absorb deceptive background information that has high similarity with the objects during the interaction, thus failing to discover desired targets accurately. As proved in Figure 1(b), the prototype fish is more similar to background pixels than target pixels in the camouflaged image. Therefore, we inevitably face another question: How to focus on camouflaged instances to achieve de-camouflaging?" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.69, + 0.473, + 0.902 + ], + "angle": 0, + "content": "Motivated by the above discussions, we propose an end-to-end De-camouflaging Network (DCNet) by jointly modeling pixel-level camouflage decoupling and instance-level camouflage suppression for CIS. In the Pixel-level Camouflage Decoupling module (PCD), we focus on decoupling camouflage characteristics and target information fused in the pixel feature. First, we extract camouflage characteristics with the assistance of frequency domain information. The Fourier spectrum amplitude contains low-level statistics [34, 42] (e.g., color and texture of the environment, see Figure 5) that accords with the camouflage characteristics. Based on the obtained description of camouflage characteristics, we propose a novel difference attention mechanism to acquire de-camouflaged pixel fea" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.895, + 0.497 + ], + "angle": 0, + "content": "tures. In this mechanism, we calculate the discrepancy between features of the original image and camouflage characteristics (see Figure 1(a)), thereby decoupling the camouflage characteristics and valuable target information while filtering out the background interference. In the Instance-level Camouflage Suppression module (ICS), we aggregate the de-camouflaged pixels to achieve final segmentation and meanwhile mitigate the effect of background noise in prototype-pixel interactions. Specifically, we introduce a set of instance prototypes to capture each camouflaged instance through long-range context-aware interactions. To constrain the interactions to favor target pixels over background pixels, we design a novel reference attention mechanism, where we select de-camouflaged pixels with high contribution to prototypes as reference points (see Figure 1(c)). Then we calculate the similarity of prototype-reference and pixel-reference, respectively, thereby obtaining similarity distributions serving as soft multilabel to measure correlations between prototypes and pixels. Highly similar pixels and prototypes must have consistent similarity distributions. In this way, the soft multilabel-based correlation is more accurate than the normal pairwise correlation, as it benefits from consensus among reliable reference points with a global receptive field. As a result, the correlation noise brought by deceptive background can be suppressed and more effective prototypes that contain rich instance information can be obtained." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.5, + 0.895, + 0.667 + ], + "angle": 0, + "content": "The contributions of our method could be summarized as follows: (1) We propose a novel De-camouflaging Network (DCNet) by jointly modeling pixel-level camouflage decoupling and instance-level camouflage suppression for CIS. (2) We propose two effective designs in DCNet, i.e., difference attention mechanism and reference attention mechanism, which can highlight target information and suppress background interference. (3) Extensive experimental results on two benchmark datasets demonstrate the effectiveness of the proposed method, e.g., with more than \\(5\\%\\) performance gains on COD10K and NC4K datasets in average precision." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.678, + 0.642, + 0.695 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.704, + 0.894, + 0.765 + ], + "angle": 0, + "content": "Since camouflaged instance segmentation is a new emerging task with very few papers, we mainly focus on introducing several lines of research in camouflaged object detection and instance segmentation." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.772, + 0.773, + 0.789 + ], + "angle": 0, + "content": "2.1. Camouflaged Object Detection" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.795, + 0.894, + 0.903 + ], + "angle": 0, + "content": "Camouflaged Object Detection (COD), aiming at localizing camouflaged objects from its background, has a long history in biology and art [10, 23]. Early works [18, 31, 36] in this field focus on distinguishing objects in their camouflaging environment by utilizing handcrafted features, such as texture, boundary, and intensity features. With the significant advances in deep learning (DL), DL-based meth" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "17919" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.141, + 0.093, + 0.832, + 0.377 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.39, + 0.895, + 0.462 + ], + "angle": 0, + "content": "Figure 2. Illustration of the proposed DCNet. Our framework mainly consists of two components: a pixel-level camouflage decoupling module and an instance-level camouflage suppression module. (1) The pixel-level camouflage decoupling module aims at extracting camouflaged characteristics and eliminating camouflaged characteristics while reserving target object characteristics to obtain de-camouflaged pixel features. (2) The instance-level camouflage suppression module is responsible for aggregating instance prototypes to achieve instance segmentation meanwhile suppressing the background noise in prototype-pixel interactions." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.487, + 0.473, + 0.744 + ], + "angle": 0, + "content": "ods [16, 19, 29, 32, 43, 45, 46] have significantly improved the COD performance through end-to-end learning. Mei et al. [29] developed a distraction mining strategy to benefit the accurate segmentation of the camouflaged object. Pang et al. [32] proposed a mixed-scale triplet network, ZoomNet, to capture objects in complex scenes at different \"zoom\" scales. Zhong et al. [45] was the first to claim the COD task should go beyond the RGB domain and introduced frequency clues to better detect camouflaged objects. Despite the large progress in camouflaged object detection, camouflaged instance segmentation is a rarely studied task. In this paper, inspired by [34, 42, 45], we introduce frequency domain transform to model the camouflage characteristics and then achieve de-camouflaging for camouflaged instance segmentation. To the best of our knowledge, we are the first to explore explicit de-camouflaging for camouflaged instance segmentation." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.756, + 0.291, + 0.772 + ], + "angle": 0, + "content": "2.2. Instance Segmentation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.781, + 0.47, + 0.903 + ], + "angle": 0, + "content": "Instance segmentation is a challenging task as it requires both pixel-level and instance-level mask prediction. Existing instance segmentation approaches can be broadly separated into two paradigms: two-stage methods [14, 17, 21] and one-stage methods [1, 40, 41]. Two-stage methods generally utilize object detectors to generate proposal regions and then segment the mask of each detected object. Based on Faster R-CNN [35], Mask R-CNN [14] generates region-" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.487, + 0.895, + 0.807 + ], + "angle": 0, + "content": "of-interests (ROIs) in the first stage and utilizes an extra mask head to acquire the instance segmentation mask. The following works continue to improve the performance for this task in the two-stage manner. Recently, one-stage methods have emerged and driven a new trend in this field. YOLACT [1] combines the prototypes with the mask coefficients to produce instance masks without the detection step. SOLOv2 [41], evolved from SOLO [40], directly decouples the original mask prediction into kernel learning and feature learning to generate final instance segmentation results. Currently, the transformer-based methods [5, 6] introduce instance-specific prototypes to constantly interact with pixel features with the help of some attention mechanisms for final instance segmentation, achieving leading performance. However, the general instance segmentation methods can not directly apply to CIS, as the camouflaged objects present high similarity with the background. Therefore, we base on the one-stage transformer-based paradigm to design an end-to-end De-camouflaging Network, where we introduce reliable reference points to build accurate similarity measurement that is robust to deceptive backgrounds." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.828, + 0.631, + 0.843 + ], + "angle": 0, + "content": "3. Our Method" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.856, + 0.895, + 0.901 + ], + "angle": 0, + "content": "In this section, we first present the overall architecture of the proposed De-camouflage Network (DCNet), and then describe each module in detail." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "17920" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.078, + 0.091, + 0.274, + 0.105 + ], + "angle": 0, + "content": "3.1. Overall Architecture" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.114, + 0.47, + 0.265 + ], + "angle": 0, + "content": "Our DCNet mainly consists of two modules. (1) The pixel-level camouflage decoupling module aims at extracting camouflage characteristics based on the amplitude information of Fourier spectrum, and eliminating camouflage characteristics via a difference attention mechanism. (2) The instance-level camouflage suppression module is responsible for learning instance prototypes from interactions with pixel features, through a reference attention mechanism to suppress the erroneous correlations caused by deceptive background." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.275, + 0.381, + 0.291 + ], + "angle": 0, + "content": "3.2. Pixel-level Camouflage Decoupling" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.298, + 0.469, + 0.388 + ], + "angle": 0, + "content": "In order to decouple camouflage characteristics and target information fused in the pixel feature, we first model the representation of camouflage characteristics by exploring Fourier frequency domain. Then we design a difference attention mechanism to remove camouflage characteristics while maintaining target information." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.389, + 0.469, + 0.538 + ], + "angle": 0, + "content": "Camouflage Characteristics Extraction. Generally, the camouflage characteristics are mainly comprised of color and texture information [10, 22]. On the other hand, in the frequency domain, it is known that the amplitude component of Fourier spectrum preserves low-level statistics information. Therefore, we can utilize the amplitude of Fourier spectrum to represent background information in the camouflaged image. Specifically, given a camouflaged image \\( x \\in \\mathbb{R}^{H \\times W \\times 3} \\), its Fourier transformation \\( \\mathcal{F}(x) \\) is formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.131, + 0.548, + 0.469, + 0.59 + ], + "angle": 0, + "content": "\\[\n\\mathcal {F} (x) _ {u, v} = \\sum_ {i = 0} ^ {H - 1} \\sum_ {j = 0} ^ {W - 1} x _ {i, j} e ^ {- J 2 \\pi \\left(\\frac {i}{H} u + \\frac {j}{W} v\\right)}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.601, + 0.47, + 0.647 + ], + "angle": 0, + "content": "where \\( J \\) refers to the imaginary unit and each channel of the image is computed independently. The amplitude and phase components are then respectively expressed as:" + }, + { + "type": "equation", + "bbox": [ + 0.146, + 0.658, + 0.4, + 0.679 + ], + "angle": 0, + "content": "\\[\n\\mathcal {A} (x) _ {u, v} = \\left[ R ^ {2} (x) _ {u, v} + I ^ {2} (x) _ {u, v} \\right] ^ {1 / 2}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.148, + 0.679, + 0.469, + 0.715 + ], + "angle": 0, + "content": "\\[\n\\mathcal {P} (x) _ {u, v} = \\arctan \\left[ \\frac {I (x) _ {u , v}}{R (x) _ {u , v}} \\right], \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.725, + 0.47, + 0.786 + ], + "angle": 0, + "content": "where \\( R(x) \\) and \\( I(x) \\) represent the real and imaginary part of \\( \\mathcal{F}(x) \\), respectively. To represent the camouflaged background, we fix the phase to a constant \\( b \\) (e.g., average value) and reconstruct the image with the amplitude information as" + }, + { + "type": "equation", + "bbox": [ + 0.188, + 0.796, + 0.469, + 0.814 + ], + "angle": 0, + "content": "\\[\n\\tilde {x} = \\mathcal {F} ^ {- 1} [ \\mathcal {A} (x) _ {u, v} e ^ {- J b} ], \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.825, + 0.469, + 0.901 + ], + "angle": 0, + "content": "where \\(\\mathcal{F}^{-1}\\) indicates the inverse Fourier transformation. Then, we feed the reconstructed image \\(\\tilde{x}\\) into a light-weight CNN (e.g., ResNet-18 [15]) with global average pooling layer to obtain the global camouflage characteristics \\(g\\) of the original image." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.273 + ], + "angle": 0, + "content": "Difference Attention Mechanism. Given image features \\(\\mathbf{F} \\in \\mathbb{R}^{h \\times w \\times c}\\) extracted from a backbone network (e.g., ResNet-50 [15]), we remove the camouflage characteristics \\(\\pmb{g}\\) while retaining valuable target information by a novel difference attention mechanism. In specific, we first utilize two \\(1 \\times 1\\) convolution layers to map \\(\\mathbf{F}\\) and \\(\\pmb{g}\\) to the same dimension \\(C\\), resulting in \\(\\hat{\\mathbf{F}} \\in \\mathbb{R}^{h \\times w \\times C}\\) and \\(\\hat{\\pmb{g}} \\in \\mathbb{R}^C\\), respectively. The discrepancy between image features and camouflage characteristics indicates the saliency information of the target objects, helping to highlight the target information. Thus, we integrate the difference on all channels for each pixel to obtain a difference map \\(\\mathbf{D} \\in \\mathbb{R}^{h \\times w}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.605, + 0.286, + 0.892, + 0.327 + ], + "angle": 0, + "content": "\\[\n\\mathbf {D} _ {i, j} = \\sum_ {k = 1} ^ {C} \\left(\\hat {\\mathbf {F}} _ {i, j, k} - \\hat {\\boldsymbol {g}} _ {k}\\right) ^ {2}, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.339, + 0.892, + 0.4 + ], + "angle": 0, + "content": "where \\(i,j\\), and \\(k\\) are the index of height, width, and channel, respectively. By multiplying the original image features with the difference map \\(\\mathbf{D}\\), we can get the de-camouflaged pixel features \\(\\mathbf{F}^d\\) as" + }, + { + "type": "equation", + "bbox": [ + 0.65, + 0.412, + 0.892, + 0.429 + ], + "angle": 0, + "content": "\\[\n\\mathbf {F} ^ {d} = \\mathbf {F} \\circ \\mathbf {D}, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.443, + 0.891, + 0.503 + ], + "angle": 0, + "content": "where \\(\\circ\\) refers to the Hadamard product and \\(\\mathbf{D}\\) is broadcasted to the same dimension as \\(\\mathbf{F}\\). In addition, we follow [33] to use a deformable self-attention layer [47] to further incorporate context information from other pixels." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.504, + 0.892, + 0.653 + ], + "angle": 0, + "content": "Fusion Layer. To acquire fine-grained target information for more accurate segmentation, we use multi-scale features from different stages of the backbone. Specifically, we first obtain the features from different stages of the backbone network, and then obtain the corresponding de-camouflaged features through the difference attention mechanism. Finally, we fuse the processed features with the FPN [24] network to generate the high-resolution pixel-level feature \\(\\mathbf{E}\\), which is used for subsequent instance segmentation." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.667, + 0.836, + 0.683 + ], + "angle": 0, + "content": "3.3. Instance-level Camouflage Suppression" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.69, + 0.891, + 0.78 + ], + "angle": 0, + "content": "In order to capture different camouflaged instances, we learn a set of instance prototypes \\(\\mathbf{P} = \\{\\pmb{p}^i\\}_{i=1}^N\\), and each \\(\\pmb{p}^i \\in \\mathbb{R}^{1 \\times L}\\) is responsible for identifying whether each pixel belongs to a camouflaged instance. To eliminate the background interference for prototype learning, we design reference attention to aggregate target information." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.781, + 0.892, + 0.874 + ], + "angle": 0, + "content": "Reference Attention Mechanism. Given the decamouflaged pixel feature \\(\\mathbf{F}^d\\) derived from the pixel-level camouflage decoupling module, we first obtain the queries \\(\\mathbf{Q} = [q_1;q_2;\\dots;q_N]\\) from instance prototypes \\(\\mathbf{P}\\), keys \\(\\mathbf{K} = [k_{1};k_{2};\\dots;k_{hw}]\\) and values \\(\\mathbf{V} = [v_{1};v_{2};\\dots;v_{hw}]\\) from pixel features \\(\\mathbf{F}^d = [f_1;f_2;\\dots;f_{hw}]\\) as" + }, + { + "type": "equation", + "bbox": [ + 0.56, + 0.885, + 0.892, + 0.903 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {q} _ {i} = \\boldsymbol {p} _ {i} \\mathbf {W} ^ {q}, \\boldsymbol {k} _ {j} = \\boldsymbol {f} _ {j} \\mathbf {W} ^ {k}, \\boldsymbol {v} _ {j} = \\boldsymbol {f} _ {j} \\mathbf {W} ^ {v}, \\tag {6}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "17921" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.091, + 0.47, + 0.152 + ], + "angle": 0, + "content": "where \\(i \\in 1,2,\\dots,N, j \\in 1,2,\\dots,hw\\), and \\(\\mathbf{W}^q \\in \\mathbb{R}^{L \\times L}\\), \\(\\mathbf{W}^k\\), \\(\\mathbf{W}^v \\in \\mathbb{R}^{c \\times L}\\) are linear projections. With the prototype queries, we can compute the correlation between each query-key pair independently as" + }, + { + "type": "equation", + "bbox": [ + 0.212, + 0.159, + 0.469, + 0.191 + ], + "angle": 0, + "content": "\\[\ns _ {i, j} = \\frac {d \\left(\\boldsymbol {q} _ {i} , \\boldsymbol {k} _ {j}\\right)}{\\sqrt {L}}, \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.192, + 0.47, + 0.403 + ], + "angle": 0, + "content": "where \\(d(\\cdot, \\cdot)\\) denotes the distance metric which is the dot product similarity in [39], and \\(\\sqrt{L}\\) is a scaling factor. As the camouflaged object presents high similarity with the surrounding backgrounds, the prototype-pixel correlation represented by \\(s_{i,j}\\) is susceptible to background pixel interference, resulting in erroneous segmentation. Intuitively, humans are able to identify camouflaged instances by repeatedly comparing the target with the surrounding reference areas. Motivated by this, we first use an adaptive selection process to select some reliable pixels from de-camouflaged pixels \\(\\mathbf{F}^d\\) as references. In specific, to measure the reliability of each pixel, we define and calculate the total contribution of each pixel with regard to all the prototypes, which is formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.17, + 0.408, + 0.469, + 0.448 + ], + "angle": 0, + "content": "\\[\nu _ {j} = \\sum_ {i = 1} ^ {N} s _ {i, j}, j \\in 1, 2, \\dots , h w. \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.449, + 0.469, + 0.54 + ], + "angle": 0, + "content": "Then we select top-\\(K\\) pixels with the largest correlations with the prototypes as reference points \\(\\mathbf{R} \\in \\mathbb{R}^{K \\times L}\\). In this way, we can pick out reliable reference points that are adaptive to the instance prototypes. Based on reference points, we first calculate the prototype-reference and pixel-reference similarity as" + }, + { + "type": "equation", + "bbox": [ + 0.156, + 0.546, + 0.469, + 0.567 + ], + "angle": 0, + "content": "\\[\ns _ {i} ^ {q} = \\phi_ {1} \\left(\\boldsymbol {q} _ {i} \\mathbf {R} ^ {\\top}\\right), s _ {j} ^ {k} = \\phi_ {2} \\left(\\boldsymbol {k} _ {j} \\mathbf {R} ^ {\\top}\\right), \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.572, + 0.469, + 0.633 + ], + "angle": 0, + "content": "where \\(\\phi_{1}\\) and \\(\\phi_{2}\\) are feed-forward networks (FFN) containing two fully connected layers. \\(s_i^q\\) and \\(s_j^k\\) can serve as soft multilabel to measure the correlation between the \\(i\\)-th prototype and the \\(j\\)-th pixel:" + }, + { + "type": "equation", + "bbox": [ + 0.167, + 0.639, + 0.469, + 0.66 + ], + "angle": 0, + "content": "\\[\ns _ {i, j} ^ {q k} = d \\left(\\boldsymbol {q} _ {i}, \\boldsymbol {k} _ {j}; \\mathbf {R}\\right) = s _ {i} ^ {q} \\left(s _ {j} ^ {k}\\right) ^ {\\top}. \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.668, + 0.469, + 0.745 + ], + "angle": 0, + "content": "Compared to \\( s_{i,j} \\), \\( s_{i,j}^{qk} \\) leverages the consensus among reliable reference points, thus suppressing the background noises and leading to fewer erroneous correlations. Given the correlation \\( s_{i,j}^{qk} \\), we can extract and purify target information from pixel features to update instance prototypes as" + }, + { + "type": "equation", + "bbox": [ + 0.135, + 0.752, + 0.469, + 0.795 + ], + "angle": 0, + "content": "\\[\n\\tilde {\\boldsymbol {p}} _ {i} = \\sum_ {j = 1} ^ {h w} a _ {i, j} \\boldsymbol {v} _ {j}, a _ {i, j} = \\frac {\\exp \\left(s _ {i , j} ^ {q k}\\right)}{\\sum_ {j = 1} ^ {h w} \\exp \\left(s _ {i , j} ^ {q k}\\right)}. \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.802, + 0.468, + 0.832 + ], + "angle": 0, + "content": "Note that we have omitted the scaling factor and the multihead mechanism for notation simplicity." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.832, + 0.469, + 0.877 + ], + "angle": 0, + "content": "Given the high-resolution pixel-level feature embedding \\(\\mathbf{E}\\) and the learned instance prototypes \\(\\tilde{\\pmb{p}}_i\\), we can obtain N mask predictions and according scores by" + }, + { + "type": "equation", + "bbox": [ + 0.145, + 0.884, + 0.469, + 0.903 + ], + "angle": 0, + "content": "\\[\nm _ {i} = \\operatorname {s i g m o i d} \\left(\\mathbf {E} \\tilde {\\boldsymbol {p}} _ {i} ^ {\\top}\\right), y _ {i} = h \\left(\\tilde {\\boldsymbol {p}} _ {i} \\mid \\sigma\\right), \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.091, + 0.892, + 0.198 + ], + "angle": 0, + "content": "where \\( h(\\cdot \\mid \\sigma) \\) is a classifier parameterized by \\( \\sigma \\) to predict the confidence score. Following [6], we match a ground truth label (including \"no object\") for each predicted segment by the Hungarian matching algorithm. For training, we use a cross-entropy loss to constrain the instance score \\( y_{i} \\), and a linear combination of a focal loss [25] and a dice loss [30] to constrain the mask prediction \\( m_{i} \\)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.213, + 0.625, + 0.229 + ], + "angle": 0, + "content": "4. Experiment" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.238, + 0.676, + 0.254 + ], + "angle": 0, + "content": "4.1. Experiment Setup" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.261, + 0.892, + 0.397 + ], + "angle": 0, + "content": "Datasets. We evaluate our method on two datasets: COD10K [10] and NC4K [28]. COD10K contains 5086 camouflaged images with high-quality instance-level annotations, which are divided into 3040 training images and 2026 testing images. NC4K contains 4121 test camouflaged images to evaluate the generalization ability of CIS models. Following the standard benchmark [33], we use the instance-level annotations in COD10K to train our DCNet and evaluate it on the test set of COD10K and NC4K." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.398, + 0.892, + 0.487 + ], + "angle": 0, + "content": "Evaluation Metrics. We adopt \\(\\mathrm{AP}_{50}\\), \\(\\mathrm{AP}_{75}\\) and AP scores as evaluation metrics to quantify the effectiveness of our method. A true positive is counted if the intersection over union (IoU) between the ground truth and the segmentation is at least \\(50\\%\\) or \\(75\\%\\), and the AP score is an overall metric that combines different IoU thresholds." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.489, + 0.892, + 0.64 + ], + "angle": 0, + "content": "Implementation Details. For fair comparisons, the feature extractor is implemented by ResNet-50 [15] which is pretrained on ImageNet [8]. During our training, our model is trained with batch size of 2, using the Adam optimizer [27] with an initial learning rate of 0.0001 for 90,000 iterations. We set the channel dimension \\( C = L = 256 \\). In the instance-level camouflage suppression module, we set the number of prototypes \\( N \\) as 10, and the number of reference points \\( K \\) as 64, which turns out to generate the best segmentation during our experiments as shown in Figure 4." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.651, + 0.8, + 0.667 + ], + "angle": 0, + "content": "4.2. Comparison with State-of-the-arts" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.674, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Quantitative results. Considering that CIS is a newly emerging task, we also compare some general instance segmentation methods besides the previous CIS methods and all the methods adopt ResNet-50 [15] backbone for fair comparisons. As demonstrated in Table 1, the proposed DCNet consistently outperforms the state-of-the-art methods by a large margin on both datasets. (1) On the COD10K [10], our method outperforms the previous best method OSFormer [33] by \\(4.3\\%\\) in AP, indicating that our DCNet is able to acquire more accurate segmentation of camouflaged targets. (2) On the NC4K [28], Our model yields an accuracy of \\(52.8\\%\\) in AP, making an obvious performance improvement by \\(7.0\\%\\) in AP compared to the second best performing method Mask2Former [5]. Since the model is trained on the COD10K training set, the high" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "17922" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.089, + 0.892, + 0.119 + ], + "angle": 0, + "content": "Table 1. Comparisons of existing CIS methods and general instance approaches on COD10K [10] and NC4K [28] testing set, where all the methods use ResNet-50 as feature extractor. The best results are shown in bold." + }, + { + "type": "table", + "bbox": [ + 0.101, + 0.13, + 0.861, + 0.399 + ], + "angle": 0, + "content": "
MethodsCOD10K-TestNC4K-TestParams(M)GFLOPs
AP\\( AP_{50} \\)\\( AP_{75} \\)AP\\( AP_{50} \\)\\( AP_{75} \\)
Mask R-CNN [14]25.055.520.427.758.622.743.9186.3
MS R-CNN [17]30.157.228.731.058.729.460.0198.5
Cascade R-CNN [2]25.356.121.329.560.824.871.7334.1
HTC [4]28.156.325.129.859.026.676.9331.7
BlendMask [3]28.256.425.227.756.724.235.8233.8
Mask Transfiner [20]28.756.326.429.456.727.244.3185.1
YOLACT [1]24.353.319.732.165.327.9--
CondInst [38]30.663.626.133.467.429.434.1200.1
QueryInst [12]28.560.123.133.066.729.4--
SOTR [13]27.958.724.129.361.025.663.1476.7
SOLOv2 [41]32.563.229.934.465.931.946.2318.7
MaskFormer [6]38.265.137.944.671.945.845.0174.2
Mask2Former [5]39.467.738.545.873.647.543.9241.0
OSFormer [33]41.071.140.842.572.542.346.6324.7
DCNet (ours)45.370.747.552.877.156.553.4207.0
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.422, + 0.47, + 0.451 + ], + "angle": 0, + "content": "Table 2. Comparison of the different components in pixel-level camouflage decoupling module in terms of AP scores." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.462, + 0.466, + 0.54 + ], + "angle": 0, + "content": "
COD10KNC4K
PCD (ours)45.352.8
-camouflage attribute extractor43.1(-2.2)48.1(-4.7)
-difference attention mechanism44.1(-1.2)51.3(-1.5)
-both 2 components above40.2(-5.1)46.4(-6.4)
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.568, + 0.47, + 0.825 + ], + "angle": 0, + "content": "performance on NC4K indicates that our method also has better generalization ability. Notice that our method does not achieve the highest performance on \\(\\mathrm{AP}_{50}\\) metric for the NC4K because different from the previous CIS method, we do not use additional Non-maximum suppression (NMS) post-processing to remove redundant predictions. For \\(\\mathrm{AP}_{50}\\), our more accurate segmentation results cannot increase the number of true positive (TP) under the \\(50\\%\\) IoU threshold, while redundant predictions will increase the number of false positive (FP), thus our method has no obvious advantages. For \\(\\mathrm{AP}_{75}\\), our high precision segmentation will increase TP with a large margin. The overall metric AP reflects the superiority of our method over other methods. There is a similar trend on NC4K. In addition, the lower GFLOPs of our DCNet compared to state-of-the-art methods indicates that our model achieves high segmentation ability while maintaining high computation efficiency." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.47, + 0.903 + ], + "angle": 0, + "content": "Qualitative Results. As shown in Figure 3, our proposed method is capable of de-camouflaging by using the difference attention mechanism at the pixel level, and also able to identify the accurate locations of target camouflaged objects at the instance level. Compared to previous methods," + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.422, + 0.892, + 0.463 + ], + "angle": 0, + "content": "Table 3. Comparisons of different attention mechanisms in the instance-level camouflage suppression module in terms of AP scores." + }, + { + "type": "table", + "bbox": [ + 0.532, + 0.475, + 0.858, + 0.537 + ], + "angle": 0, + "content": "
Attention MechanismCOD10KNC4K
cross-attention [39]42.449.5
masked attention [5]44.751.7
reference attention (ours)45.352.8
" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.564, + 0.892, + 0.625 + ], + "angle": 0, + "content": "our DCNet performs better at boundaries of camouflage instances (see the first 4 columns), suppresses distracting background regions (see \\(5^{th}\\) to \\(6^{th}\\) columns), and distinguishes multiple instances well (see the last 2 columns)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.635, + 0.655, + 0.652 + ], + "angle": 0, + "content": "4.3. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.659, + 0.892, + 0.688 + ], + "angle": 0, + "content": "We conduct comprehensive ablation studies on COD10K and NC4K to verify the effectiveness of our modules." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.689, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Effectiveness of the Pixel-level Camouflage Decoupling Module. The pixel-level camouflage decoupling module mainly consists of a camouflage attribute extractor and a difference attention mechanism. We validate the importance of each component by removing them one at a time. As shown in Table 2, we replace the proposed camouflage attribute extractor with a learnable vector, and the performance drops by an average of \\(3.5\\%\\) across two datasets. It indicates that the Fourier spectrum amplitude is effective in modeling the camouflage characteristics, verifying that it is hard to learn de-camouflage without image-specific camouflage information supervision. Then we replace the difference attention mechanism with general subtraction operations. The reduced performance indicates that our pro" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "17923" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.123, + 0.113, + 0.216, + 0.127 + ], + "angle": 0, + "content": "Original image" + }, + { + "type": "image", + "bbox": [ + 0.231, + 0.098, + 0.301, + 0.141 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.308, + 0.098, + 0.378, + 0.141 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.387, + 0.098, + 0.456, + 0.14 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.464, + 0.098, + 0.534, + 0.14 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.542, + 0.098, + 0.612, + 0.14 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.619, + 0.098, + 0.69, + 0.14 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.697, + 0.098, + 0.767, + 0.14 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.775, + 0.098, + 0.845, + 0.14 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.162, + 0.162, + 0.184, + 0.173 + ], + "angle": 0, + "content": "GT" + }, + { + "type": "image", + "bbox": [ + 0.231, + 0.146, + 0.302, + 0.19 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.309, + 0.147, + 0.378, + 0.19 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.387, + 0.147, + 0.456, + 0.19 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.465, + 0.147, + 0.534, + 0.19 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.542, + 0.147, + 0.612, + 0.19 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.619, + 0.147, + 0.69, + 0.19 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.697, + 0.147, + 0.767, + 0.19 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.775, + 0.147, + 0.845, + 0.19 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.126, + 0.211, + 0.21, + 0.223 + ], + "angle": 0, + "content": "MaskFormer" + }, + { + "type": "image", + "bbox": [ + 0.231, + 0.196, + 0.302, + 0.239 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.309, + 0.196, + 0.378, + 0.239 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.387, + 0.196, + 0.456, + 0.239 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.465, + 0.196, + 0.534, + 0.239 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.542, + 0.196, + 0.612, + 0.239 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.619, + 0.196, + 0.69, + 0.239 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.697, + 0.196, + 0.767, + 0.239 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.775, + 0.196, + 0.845, + 0.239 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.125, + 0.26, + 0.214, + 0.272 + ], + "angle": 0, + "content": "Mask2Former" + }, + { + "type": "image", + "bbox": [ + 0.231, + 0.244, + 0.302, + 0.288 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.309, + 0.244, + 0.378, + 0.288 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.387, + 0.244, + 0.456, + 0.288 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.465, + 0.244, + 0.534, + 0.287 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.542, + 0.244, + 0.612, + 0.287 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.619, + 0.244, + 0.69, + 0.287 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.697, + 0.244, + 0.767, + 0.287 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.775, + 0.244, + 0.845, + 0.287 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.141, + 0.309, + 0.209, + 0.32 + ], + "angle": 0, + "content": "OSFormer" + }, + { + "type": "image", + "bbox": [ + 0.231, + 0.293, + 0.302, + 0.336 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.309, + 0.293, + 0.378, + 0.336 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.387, + 0.293, + 0.456, + 0.336 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.465, + 0.293, + 0.534, + 0.336 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.542, + 0.293, + 0.612, + 0.336 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.619, + 0.293, + 0.69, + 0.336 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.697, + 0.293, + 0.767, + 0.336 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.775, + 0.293, + 0.845, + 0.336 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.13, + 0.358, + 0.21, + 0.37 + ], + "angle": 0, + "content": "DCNet(ours)" + }, + { + "type": "image", + "bbox": [ + 0.231, + 0.342, + 0.302, + 0.384 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.309, + 0.342, + 0.378, + 0.384 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.387, + 0.342, + 0.456, + 0.384 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.465, + 0.342, + 0.534, + 0.384 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.542, + 0.342, + 0.612, + 0.384 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.619, + 0.342, + 0.69, + 0.384 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.697, + 0.342, + 0.767, + 0.384 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.775, + 0.342, + 0.845, + 0.384 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.399, + 0.878, + 0.414 + ], + "angle": 0, + "content": "Figure 3. Qualitative results of DCNet compared with state-of-the-art methods. Different colored masks represent different instances." + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.427, + 0.471, + 0.454 + ], + "angle": 0, + "content": "Table 4. Comparison of using different scales of features in terms of AP scores." + }, + { + "type": "table", + "bbox": [ + 0.118, + 0.467, + 0.424, + 0.552 + ], + "angle": 0, + "content": "
COD10KNC4K
single scale (1/32)45.252.5
single scale (1/16)44.851.6
single scale (1/8)45.152.2
multiple scales45.352.8
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.577, + 0.47, + 0.653 + ], + "angle": 0, + "content": "posed mechanism can well model the discrepancy between image features and camouflage characteristics. Removing both of the components leads to huge performance degradation, showing that the combination of them is essential in improving the camouflaged object segmentation." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.656, + 0.47, + 0.822 + ], + "angle": 0, + "content": "Effectiveness of the Instance-level Camouflage Suppression Module. As shown in Table 3, the reference attention mechanism in our instance-level camouflage suppression module has the highest AP score on both datasets. Compared to vanilla cross-attention [39], our proposed reference points can effectively suppress the noise brought by background pixels in the prototype-pixel interaction, thus obtaining significant improvement for accurate segmentation. Meanwhile, our method is superior to existing cross-attention variants, such as the masked attention [5] in the segmentation field." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Effectiveness of Hierarchical Features To explore decamouflage at multiple feature levels, we feed 3 scales of features (i.e., stride of 32, 16, and 8) into the difference attention mechanism, respectively. As shown in Table 4, low-resolution features exceed other scales. This may attribute" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.427, + 0.892, + 0.456 + ], + "angle": 0, + "content": "Table 5. Comparisons with OSFormer [33] of different backbones on COD10K [10] and NC4K [28] in terms of AP scores." + }, + { + "type": "table", + "bbox": [ + 0.543, + 0.467, + 0.848, + 0.612 + ], + "angle": 0, + "content": "
MethodBackboneCOD10KNC4K
OSFormerR5041.042.5
DCNet (ours)45.352.8
OSFormerR10142.044.4
DCNet (ours)46.853.5
OSFormerSwin-T47.750.2
DCNet (ours)50.356.3
OSFormerSwin-S52.156.7
DCNet (ours)52.358.4
" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.623, + 0.692, + 0.719 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.699, + 0.623, + 0.892, + 0.719 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.731, + 0.894, + 0.76 + ], + "angle": 0, + "content": "Figure 4. Comparisons of performance with different numbers of prototypes and reference points on COD10K [10] and NC4K [28]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.779, + 0.892, + 0.869 + ], + "angle": 0, + "content": "to the retention of object semantic information. The highest performance in this experiment shows that the contribution of each scale is complementary to each other. Using single-scale features alone is not enough to model the target information, thus combining them together can further boost the performance." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.871, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Compatibility for Feature Extractor. We equip our DC-Net with different feature extractor backbones, i.e., ResNet" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "17924" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.09, + 0.468, + 0.298 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.309, + 0.47, + 0.339 + ], + "angle": 0, + "content": "Figure 5. Illustration of the amplitude images and difference maps in our proposed difference attention." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.348, + 0.47, + 0.424 + ], + "angle": 0, + "content": "50 [15], ResNet-101 [15], Swin-T [26], and Swin-S [26]. All of them are pre-trained on ImageNet-1K [8]. As shown in Table 5, our method performs the best in all cases on both datasets, proving that our approach has tremendous potential for further improvement." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.425, + 0.473, + 0.699 + ], + "angle": 0, + "content": "Analysis of Hyperparameters. In the instance-level camouflage suppression module, prototypes play an important role in capturing instances by aggregating object pixels. As shown in Figure 4(a), we conduct quantitative experiments to analyze how many prototypes are suitable for instance segmentation. It can be seen that too many or too few prototypes will damage the AP metric. In fact, the number of prototypes \\( N \\) should be larger than the count of instances in an image to avoid instance fusion error. While superabundant prototypes will cause many prototypes to match \"no object\" during training. We choose \\( N = 10 \\) by balancing the performance on both datasets. Besides, in the reference attention mechanism, the number of reference points \\( K \\) influences the similarity distributions between prototypes and pixels. As shown in Figure 4(b), we can observe that the performance on NC4K continues to grow until \\( K = 64 \\), which means that it is sufficient for building appropriate prototype-pixel correction." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.709, + 0.221, + 0.725 + ], + "angle": 0, + "content": "4.4. Visualizations" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.733, + 0.47, + 0.854 + ], + "angle": 0, + "content": "Visualization of Difference Map. As shown in Figure 5, the \\(2^{nd}\\) row exhibits the amplitude images that are generated through Fourier transformation, which can be considered as camouflage characteristics. With the help of the difference attention mechanism, our network can remove the confusing background and localize the accurate salient areas (see \\(3^{rd}\\) row), which turn out to be the correct regions of camouflaged instances." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Visualization of Reference Attention. To explore the reference attention mechanism's ability to suppress the background noise, we make qualitative visualization to compare" + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.091, + 0.892, + 0.291 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.303, + 0.892, + 0.331 + ], + "angle": 0, + "content": "Figure 6. Visualizations of the activation map with different attention mechanisms." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.35, + 0.895, + 0.533 + ], + "angle": 0, + "content": "different attention mechanisms. As shown in Figure 6 (d), with the proposed reference attention, the high response is clustered in the foreground area and the prototype pays little attention to the background area. In contrast, in Figure 6 (b), prototypes with the cross-attention [39] tend to absorb pixels from the non-target region. And in Figure 6 (c), with the masked attention [5], most background pixels are masked, while some foreground pixels are also suppressed, leading to sub-optimal performance. Thus, the reference mechanism plays an important role in mitigating the effect of background noise in prototype-pixel interactions, which helps to achieve more accurate segmentation." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.561, + 0.619, + 0.577 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.585, + 0.892, + 0.781 + ], + "angle": 0, + "content": "In this paper, we propose an end-to-end Decamouflaging Network (DCNet) by jointly modeling pixel-level camouflage decoupling and instance-level camouflage suppression for camouflaged instance segmentation. We design a Pixel-level Camouflage Decoupling module to model and further eliminate the camouflage characteristics based on the frequency domain information. Besides, an Instance-level Camouflage Suppression module is proposed to achieve prototype-based instance segmentation and mitigate the effect of background noise by introducing reference points. Extensive experimental results on two benchmark datasets demonstrate the effectiveness of the proposed method." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.81, + 0.671, + 0.828 + ], + "angle": 0, + "content": "6. Acknowledgment" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.84, + 0.894, + 0.903 + ], + "angle": 0, + "content": "This work was partially supported by the National Nature Science Foundation of China (Grant 62022078, 62021001), National Defense Basic Scientific Research Program of China (Grant JCKY2021601B013)." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "17925" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.471, + 0.171 + ], + "angle": 0, + "content": "[1] Daniel Bolya, Chong Zhou, Fanyi Xiao, and Yong Jae Lee. Yolact: Real-time instance segmentation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9157-9166, 2019. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.172, + 0.472, + 0.226 + ], + "angle": 0, + "content": "[2] Zhaowei Cai and Nuno Vasconcelos. Cascade r-cnn: high quality object detection and instance segmentation. IEEE transactions on pattern analysis and machine intelligence, 43(5):1483-1498, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.228, + 0.471, + 0.297 + ], + "angle": 0, + "content": "[3] Hao Chen, Kunyang Sun, Zhi Tian, Chunhua Shen, Yongming Huang, and Youliang Yan. Blendmask: Top-down meets bottom-up for instance segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8573-8581, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.298, + 0.471, + 0.38 + ], + "angle": 0, + "content": "[4] Kai Chen, Jiangmiao Pang, Jiaqi Wang, Yu Xiong, Xiaoxiao Li, Shuyang Sun, Wansen Feng, Ziwei Liu, Jianping Shi, Wanli Ouyang, et al. Hybrid task cascade for instance segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4974-4983, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.382, + 0.471, + 0.451 + ], + "angle": 0, + "content": "[5] Bowen Cheng, Ishan Misra, Alexander G Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1290–1299, 2022. 2, 3, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.452, + 0.471, + 0.507 + ], + "angle": 0, + "content": "[6] Bowen Cheng, Alex Schwing, and Alexander Kirillov. Per-pixel classification is not all you need for semantic segmentation. Advances in Neural Information Processing Systems, 34:17864-17875, 2021. 2, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.508, + 0.471, + 0.55 + ], + "angle": 0, + "content": "[7] Hung-Kuo Chu, Wei-Hsin Hsu, Niloy J Mitra, Daniel Cohen-Or, Tien-Tsin Wong, and Tong-Yee Lee. Camouflage images. ACM Trans. Graph., 29(4):51-1, 2010. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.551, + 0.471, + 0.606 + ], + "angle": 0, + "content": "[8] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.607, + 0.471, + 0.648 + ], + "angle": 0, + "content": "[9] Deng-Ping Fan, Ge-Peng Ji, Ming-Ming Cheng, and Ling Shao. Concealed object detection. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.649, + 0.47, + 0.718 + ], + "angle": 0, + "content": "[10] Deng-Ping Fan, Ge-Peng Ji, Guolei Sun, Ming-Ming Cheng, Jianbing Shen, and Ling Shao. Camouflaged object detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2777-2787, 2020. 2, 4, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.719, + 0.47, + 0.788 + ], + "angle": 0, + "content": "[11] Deng-Ping Fan, Tao Zhou, Ge-Peng Ji, Yi Zhou, Geng Chen, Huazhu Fu, Jianbing Shen, and Ling Shao. Inf-net: Automatic Covid-19 lung infection segmentation from ct images. IEEE Transactions on Medical Imaging, 39(8):2626-2637, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.789, + 0.47, + 0.845 + ], + "angle": 0, + "content": "[12] Yuxin Fang, Shusheng Yang, Xinggang Wang, Yu Li, Chen Fang, Ying Shan, Bin Feng, and Wenyu Liu. Instances as queries. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6910-6919, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.846, + 0.47, + 0.901 + ], + "angle": 0, + "content": "[13] Ruohao Guo, Dantong Niu, Liao Qu, and Zhenbo Li. Sotr: Segmenting objects with transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7157-7166, 2021. 6" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.147 + ], + "angle": 0, + "content": "[14] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In Proceedings of the IEEE international conference on computer vision, pages 2961-2969, 2017. 1, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.15, + 0.892, + 0.205 + ], + "angle": 0, + "content": "[15] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 4, 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.207, + 0.892, + 0.262 + ], + "angle": 0, + "content": "[16] Xiaobin Hu, Deng-Ping Fan, Xuebin Qin, Hang Dai, Wenqi Ren, Ying Tai, Chengjie Wang, and Ling Shao. High-resolution iterative feedback network for camouflaged object detection. arXiv preprint arXiv:2203.11624, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.264, + 0.892, + 0.319 + ], + "angle": 0, + "content": "[17] Zhaojin Huang, Lichao Huang, Yongchao Gong, Chang Huang, and Xinggang Wang. Mask scoring r-cnn. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6409-6418, 2019. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.321, + 0.892, + 0.39 + ], + "angle": 0, + "content": "[18] Ivan Huerta, Daniel Rowe, Mikhail Mozerov, and Jordi González. Improving background subtraction based on a casuistry of colour-motion segmentation problems. In Iberian Conference on Pattern Recognition and Image Analysis, pages 475-482. Springer, 2007. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.392, + 0.892, + 0.447 + ], + "angle": 0, + "content": "[19] Ge-Peng Ji, Deng-Ping Fan, Yu-Cheng Chou, Dengxin Dai, Alexander Liniger, and Luc Van Gool. Deep gradient learning for efficient camouflaged object detection. Machine Intelligence Research, 20(1):92-108, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.449, + 0.892, + 0.517 + ], + "angle": 0, + "content": "[20] Lei Ke, Martin Danelljan, Xia Li, Yu-Wing Tai, Chi-Keung Tang, and Fisher Yu. Mask transfiner for high-quality instance segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4412–4421, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.519, + 0.892, + 0.574 + ], + "angle": 0, + "content": "[21] Alexander Kirillov, Yuxin Wu, Kaiming He, and Ross Girshick. Pointrend: Image segmentation as rendering. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9799-9808, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.576, + 0.892, + 0.646 + ], + "angle": 0, + "content": "[22] Trung-Nghia Le, Yubo Cao, Tan-Cong Nguyen, Minh-Quan Le, Khanh-Duy Nguyen, Thanh-Toan Do, Minh-Triet Tran, and Tam V Nguyen. Camouflaged instance segmentation in-the-wild: Dataset, method, and benchmark suite. IEEE Transactions on Image Processing, 31:287-300, 2021. 1, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.647, + 0.892, + 0.702 + ], + "angle": 0, + "content": "[23] Trung-Nghia Le, Tam V Nguyen, Zhongliang Nie, MinhTriet Tran, and Akihiro Sugimoto. Anabranch network for camouflaged object segmentation. Computer Vision and Image Understanding, 184:45-56, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.704, + 0.892, + 0.773 + ], + "angle": 0, + "content": "[24] Tsung-Yi Lin, Piotr Dólar, Ross Girshick, Kaiming He, Bharath Hariharan, and Serge Belongie. Feature pyramid networks for object detection. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2117-2125, 2017. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.775, + 0.892, + 0.83 + ], + "angle": 0, + "content": "[25] Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dólar. Focal loss for dense object detection. In Proceedings of the IEEE international conference on computer vision, pages 2980-2988, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.832, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[26] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10012-10022, 2021. 8" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "17926" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.468, + 0.12 + ], + "angle": 0, + "content": "[27] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.122, + 0.47, + 0.191 + ], + "angle": 0, + "content": "[28] Yunqiu Lv, Jing Zhang, Yuchao Dai, Aixuan Li, Bowen Liu, Nick Barnes, and Deng-Ping Fan. Simultaneously localize, segment and rank the camouflaged objects. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11591-11601, 2021. 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.192, + 0.468, + 0.262 + ], + "angle": 0, + "content": "[29] Haiyang Mei, Ge-Peng Ji, Ziqi Wei, Xin Yang, Xiaopeng Wei, and Deng-Ping Fan. Camouflaged object segmentation with distraction mining. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8772-8781, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.264, + 0.468, + 0.331 + ], + "angle": 0, + "content": "[30] Fausto Milletari, Nassir Navab, and Seyed-Ahmad Ahmadi. V-net: Fully convolutional neural networks for volumetric medical image segmentation. In 2016 fourth international conference on 3D vision (3DV), pages 565-571. IEEE, 2016. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.333, + 0.468, + 0.375 + ], + "angle": 0, + "content": "[31] Yuxin Pan, Yiwang Chen, Qiang Fu, Ping Zhang, and Xin Xu. Study on the camouflaged target detection method based on 3d convexity. Modern Applied Science, 5(4):152, 2011. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.377, + 0.468, + 0.446 + ], + "angle": 0, + "content": "[32] Youwei Pang, Xiaoqi Zhao, Tian-Zhu Xiang, Lihe Zhang, and Huchuan Lu. Zoom in and out: A mixed-scale triplet network for camouflaged object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2160-2170, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.448, + 0.468, + 0.517 + ], + "angle": 0, + "content": "[33] Jialun Pei, Tianyang Cheng, Deng-Ping Fan, He Tang, Chuanbo Chen, and Luc Van Gool. Osformer: One-stage camouflaged instance segmentation with transformers. In European Conference on Computer Vision, pages 19-37. Springer, 2022. 1, 2, 4, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.518, + 0.468, + 0.572 + ], + "angle": 0, + "content": "[34] Leon N Piotrowski and Fergus W Campbell. A demonstration of the visual importance and flexibility of spatial-frequency amplitude and phase. Perception, 11(3):337-346, 1982. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.575, + 0.468, + 0.63 + ], + "angle": 0, + "content": "[35] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. Advances in neural information processing systems, 28, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.632, + 0.468, + 0.7 + ], + "angle": 0, + "content": "[36] P Sengottuvelan, Amitabh Wahi, and A Shanmugam. Performance of decamouflaging through exploratory image analysis. In 2008 First International Conference on Emerging Trends in Engineering and Technology, pages 6-10. IEEE, 2008. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.702, + 0.468, + 0.743 + ], + "angle": 0, + "content": "[37] Sujit K Singh, Chitra A Dhawale, and Sanjay Misra. Survey of object detection methods in camouflaged image. IERI Procedia, 4:351-357, 2013. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.746, + 0.468, + 0.788 + ], + "angle": 0, + "content": "[38] Zhi Tian, Chunhua Shen, and Hao Chen. Conditional convolutions for instance segmentation. In European conference on computer vision, pages 282-298. Springer, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.789, + 0.468, + 0.844 + ], + "angle": 0, + "content": "[39] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.846, + 0.468, + 0.9 + ], + "angle": 0, + "content": "[40] Xinlong Wang, Tao Kong, Chunhua Shen, Yuning Jiang, and Lei Li. Solo: Segmenting objects by locations. In European Conference on Computer Vision, pages 649-665. Springer, 2020. 1, 3" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.092, + 0.892, + 0.147 + ], + "angle": 0, + "content": "[41] Xinlong Wang, Rufeng Zhang, Tao Kong, Lei Li, and Chunhua Shen. Solov2: Dynamic and fast instance segmentation. Advances in Neural information processing systems, 33:17721-17732, 2020. 1, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.149, + 0.892, + 0.217 + ], + "angle": 0, + "content": "[42] Qinwei Xu, Ruipeng Zhang, Ya Zhang, Yanfeng Wang, and Qi Tian. A fourier-based framework for domain generalization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14383-14392, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.22, + 0.892, + 0.289 + ], + "angle": 0, + "content": "[43] Fan Yang, Qiang Zhai, Xin Li, Rui Huang, Ao Luo, Hong Cheng, and Deng-Ping Fan. Uncertainty-guided transformer reasoning for camouflaged object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4146-4155, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.291, + 0.892, + 0.358 + ], + "angle": 0, + "content": "[44] Xiaoqi Zhao, Lihe Zhang, and Huchuan Lu. Automatic polyp segmentation via multi-scale subtraction network. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 120-130. Springer, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.361, + 0.892, + 0.429 + ], + "angle": 0, + "content": "[45] Yijie Zhong, Bo Li, Lv Tang, Senyun Kuang, Shuang Wu, and Shouhong Ding. Detecting camouflaged object in frequency domain. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4504-4513, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.431, + 0.892, + 0.486 + ], + "angle": 0, + "content": "[46] Tao Zhou, Yi Zhou, Chen Gong, Jian Yang, and Yu Zhang. Feature aggregation and propagation network for camouflaged object detection. IEEE Transactions on Image Processing, 31:7036-7047, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.488, + 0.892, + 0.542 + ], + "angle": 0, + "content": "[47] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable detr: Deformable transformers for end-to-end object detection. arXiv preprint arXiv:2010.04159, 2020. 4" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.092, + 0.892, + 0.542 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "17927" + } + ] +] \ No newline at end of file diff --git a/2023/Camouflaged Instance Segmentation via Explicit De-Camouflaging/9acac0a8-5352-4eef-91f5-74adb1a1e4b8_origin.pdf b/2023/Camouflaged Instance Segmentation via Explicit De-Camouflaging/9acac0a8-5352-4eef-91f5-74adb1a1e4b8_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..5bf9e24fe988dde8ccad9a71bd4daee8174c6a17 --- /dev/null +++ b/2023/Camouflaged Instance Segmentation via Explicit De-Camouflaging/9acac0a8-5352-4eef-91f5-74adb1a1e4b8_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:49beb0ef2c8b68fe7bf1406ee4c14f8501879e1b6bee03fec4e2f90e526f30cc +size 6217612 diff --git a/2023/Camouflaged Instance Segmentation via Explicit De-Camouflaging/full.md b/2023/Camouflaged Instance Segmentation via Explicit De-Camouflaging/full.md new file mode 100644 index 0000000000000000000000000000000000000000..5ab12f6ea290e5193724ff9744d96686d411656e --- /dev/null +++ b/2023/Camouflaged Instance Segmentation via Explicit De-Camouflaging/full.md @@ -0,0 +1,405 @@ +# Camouflaged Instance Segmentation via Explicit De-camouflaging + +Naisong Luo $^{1*}$ , Yuwen Pan $^{1*}$ , Rui Sun $^{1}$ , Tianzhu Zhang $^{1,2,3\dagger}$ , Zhiwei Xiong $^{1,2}$ , Feng Wu $^{1,2}$ $^{1}$ University of Science and Technology of China + $^{2}$ Institute of Artificial Intelligence, Hefei Comprehensive National Science Center + $^{3}$ Deep Space Exploration Lab + +{lns6,panyw，issunrui}@mail.ustc.edu.cn，{tzzhang，zwxiong，fengwu}@ustc.edu.cn + +# Abstract + +Camouflaged Instance Segmentation (CIS) aims at predicting the instance-level masks of camouflaged objects, which are usually the animals in the wild adapting their appearance to match the surroundings. Previous instance segmentation methods perform poorly on this task as they are easily disturbed by the deceptive camouflage. To address these challenges, we propose a novel De-camouflaging Network (DCNet) including a pixel-level camouflage decoupling module and an instance-level camouflage suppression module. The proposed DCNet enjoys several merits. First, the pixel-level camouflage decoupling module can extract camouflage characteristics based on the Fourier transformation. Then a difference attention mechanism is proposed to eliminate the camouflage characteristics while reserving target object characteristics in the pixel feature. Second, the instance-level camouflage suppression module can aggregate rich instance information from pixels by use of instance prototypes. To mitigate the effect of background noise during segmentation, we introduce some reliable reference points to build a more robust similarity measurement. With the aid of these two modules, our DCNet can effectively model de-camouflaging and achieve accurate segmentation for camouflaged instances. Extensive experimental results on two benchmarks demonstrate that our DCNet performs favorably against state-of-the-art CIS methods, e.g., with more than $5\%$ performance gains on COD10K and NC4K datasets in average precision. + +# 1. Introduction + +In the field of biology, camouflage is defined as a strategy that animals use to adapt their body's appearance (e.g., color and pattern) to match their surroundings in order to achieve concealing and avoid being hunted by predators [37]. Cam- + +![](images/66787f4e73dd444f87b544ef18db1f07f67db8c0b40d93c802038274aed215cc.jpg) +(a) De-camouflaging Process + +![](images/2aed844686c373604c0854ccf300539d523e2d5fd9c8d31f6b969f0b767e21b4.jpg) +(b) Pairwise Similarity + +![](images/8267c25d9a4ed979ecfd2339ef67b13f8d70e95a14554a94aea56fd96abacd4a.jpg) +(c) Reference Similarity (Ours) + +![](images/18279214b20cffc83e28d9e3eb508dc8d9c96ede24b00038523cec3bfa5ebb34.jpg) +Figure 1. Illustration of our motivation. (a) We propose to extract camouflage characteristics to model explicit de-camouflaging for CIS. (b) Pairwise similarity between prototype and pixel is always erroneous. (c) The prototype-pixel correlation is based on the prototype-reference and pixel-reference similarity distribution, which is more accurate. + +outflaged Instance Segmentation (CIS) [22] aims at identifying the location and predicting instance-level masks of camouflaged objects, which has attracted more and more attention due to its widespread applications in medical image analysis [11, 44], search-and-rescue work [9] and recreational art [7], etc. + +Despite the tremendous progress in instance segmentation [14, 40, 41], there are few efforts [22, 33] working on Camouflaged Instance Segmentation, as CIS is a more challenging task where most camouflaged instances lack obvious contrast with the background, making general instance segmentation methods work poorly on this task. Recent CIS approaches [22, 33] are generally based on traditional instance segmentation models, either by naively fusing various general instance segmentation models [22] to get better representations, or by directly using global in + +teraction [33]. However, these approaches fail to explore the core of CIS: de-camouflaging, i.e., eliminating camouflage characteristics of the target object. Without explicitly de-camouflaging, the previous methods are easily disturbed by similar background. De-camouflaging is a challenging problem as no image-specific camouflage information is provided as supervision signals. Intuitively, humans have the ability to quickly recognise objects in a highly camouflaged scene. They first repeatedly discriminate the real target characteristics from the camouflage characteristics at the pixel level, and then aggregate the pixel information to discern the whole target instance from the background. This human visual mechanism motivates us to explore de-camouflaging strategy from the pixel level to the instance level in a progressive manner. + +In order to model de-camouflage at both the pixel and instance levels, a series of issues need to be considered. (1) The pixel level de-camouflage. In essence, each pixel feature of a camouflaged image contains the camouflage characteristics and the target object characteristics. Note that our goal is to remove camouflage characteristics and maintain target object information, thus a question naturally arises: How to decouple these two pieces of information at the pixel level? (2) The instance level segmentation. Based on the de-camouflaged pixels, we can naturally aggregate the pixel information and infer the mask of the target instance. Currently, the transformer-based models [5,6] have achieved leading performance for instance segmentation, where instance-specific prototypes are learned by constantly interacting with pixel features for final segmentation. However, directly applying the transformer to CIS is not trivial, as the prototypes would frequently absorb deceptive background information that has high similarity with the objects during the interaction, thus failing to discover desired targets accurately. As proved in Figure 1(b), the prototype fish is more similar to background pixels than target pixels in the camouflaged image. Therefore, we inevitably face another question: How to focus on camouflaged instances to achieve de-camouflaging? + +Motivated by the above discussions, we propose an end-to-end De-camouflaging Network (DCNet) by jointly modeling pixel-level camouflage decoupling and instance-level camouflage suppression for CIS. In the Pixel-level Camouflage Decoupling module (PCD), we focus on decoupling camouflage characteristics and target information fused in the pixel feature. First, we extract camouflage characteristics with the assistance of frequency domain information. The Fourier spectrum amplitude contains low-level statistics [34, 42] (e.g., color and texture of the environment, see Figure 5) that accords with the camouflage characteristics. Based on the obtained description of camouflage characteristics, we propose a novel difference attention mechanism to acquire de-camouflaged pixel fea + +tures. In this mechanism, we calculate the discrepancy between features of the original image and camouflage characteristics (see Figure 1(a)), thereby decoupling the camouflage characteristics and valuable target information while filtering out the background interference. In the Instance-level Camouflage Suppression module (ICS), we aggregate the de-camouflaged pixels to achieve final segmentation and meanwhile mitigate the effect of background noise in prototype-pixel interactions. Specifically, we introduce a set of instance prototypes to capture each camouflaged instance through long-range context-aware interactions. To constrain the interactions to favor target pixels over background pixels, we design a novel reference attention mechanism, where we select de-camouflaged pixels with high contribution to prototypes as reference points (see Figure 1(c)). Then we calculate the similarity of prototype-reference and pixel-reference, respectively, thereby obtaining similarity distributions serving as soft multilabel to measure correlations between prototypes and pixels. Highly similar pixels and prototypes must have consistent similarity distributions. In this way, the soft multilabel-based correlation is more accurate than the normal pairwise correlation, as it benefits from consensus among reliable reference points with a global receptive field. As a result, the correlation noise brought by deceptive background can be suppressed and more effective prototypes that contain rich instance information can be obtained. + +The contributions of our method could be summarized as follows: (1) We propose a novel De-camouflaging Network (DCNet) by jointly modeling pixel-level camouflage decoupling and instance-level camouflage suppression for CIS. (2) We propose two effective designs in DCNet, i.e., difference attention mechanism and reference attention mechanism, which can highlight target information and suppress background interference. (3) Extensive experimental results on two benchmark datasets demonstrate the effectiveness of the proposed method, e.g., with more than $5\%$ performance gains on COD10K and NC4K datasets in average precision. + +# 2. Related Work + +Since camouflaged instance segmentation is a new emerging task with very few papers, we mainly focus on introducing several lines of research in camouflaged object detection and instance segmentation. + +# 2.1. Camouflaged Object Detection + +Camouflaged Object Detection (COD), aiming at localizing camouflaged objects from its background, has a long history in biology and art [10, 23]. Early works [18, 31, 36] in this field focus on distinguishing objects in their camouflaging environment by utilizing handcrafted features, such as texture, boundary, and intensity features. With the significant advances in deep learning (DL), DL-based meth + +![](images/0c152675a03c59e0981a88f2e0369956a2e5cf63050717d13b6c6bff5619036d.jpg) +Figure 2. Illustration of the proposed DCNet. Our framework mainly consists of two components: a pixel-level camouflage decoupling module and an instance-level camouflage suppression module. (1) The pixel-level camouflage decoupling module aims at extracting camouflaged characteristics and eliminating camouflaged characteristics while reserving target object characteristics to obtain de-camouflaged pixel features. (2) The instance-level camouflage suppression module is responsible for aggregating instance prototypes to achieve instance segmentation meanwhile suppressing the background noise in prototype-pixel interactions. + +ods [16, 19, 29, 32, 43, 45, 46] have significantly improved the COD performance through end-to-end learning. Mei et al. [29] developed a distraction mining strategy to benefit the accurate segmentation of the camouflaged object. Pang et al. [32] proposed a mixed-scale triplet network, ZoomNet, to capture objects in complex scenes at different "zoom" scales. Zhong et al. [45] was the first to claim the COD task should go beyond the RGB domain and introduced frequency clues to better detect camouflaged objects. Despite the large progress in camouflaged object detection, camouflaged instance segmentation is a rarely studied task. In this paper, inspired by [34, 42, 45], we introduce frequency domain transform to model the camouflage characteristics and then achieve de-camouflaging for camouflaged instance segmentation. To the best of our knowledge, we are the first to explore explicit de-camouflaging for camouflaged instance segmentation. + +# 2.2. Instance Segmentation + +Instance segmentation is a challenging task as it requires both pixel-level and instance-level mask prediction. Existing instance segmentation approaches can be broadly separated into two paradigms: two-stage methods [14, 17, 21] and one-stage methods [1, 40, 41]. Two-stage methods generally utilize object detectors to generate proposal regions and then segment the mask of each detected object. Based on Faster R-CNN [35], Mask R-CNN [14] generates region- + +of-interests (ROIs) in the first stage and utilizes an extra mask head to acquire the instance segmentation mask. The following works continue to improve the performance for this task in the two-stage manner. Recently, one-stage methods have emerged and driven a new trend in this field. YOLACT [1] combines the prototypes with the mask coefficients to produce instance masks without the detection step. SOLOv2 [41], evolved from SOLO [40], directly decouples the original mask prediction into kernel learning and feature learning to generate final instance segmentation results. Currently, the transformer-based methods [5, 6] introduce instance-specific prototypes to constantly interact with pixel features with the help of some attention mechanisms for final instance segmentation, achieving leading performance. However, the general instance segmentation methods can not directly apply to CIS, as the camouflaged objects present high similarity with the background. Therefore, we base on the one-stage transformer-based paradigm to design an end-to-end De-camouflaging Network, where we introduce reliable reference points to build accurate similarity measurement that is robust to deceptive backgrounds. + +# 3. Our Method + +In this section, we first present the overall architecture of the proposed De-camouflage Network (DCNet), and then describe each module in detail. + +# 3.1. Overall Architecture + +Our DCNet mainly consists of two modules. (1) The pixel-level camouflage decoupling module aims at extracting camouflage characteristics based on the amplitude information of Fourier spectrum, and eliminating camouflage characteristics via a difference attention mechanism. (2) The instance-level camouflage suppression module is responsible for learning instance prototypes from interactions with pixel features, through a reference attention mechanism to suppress the erroneous correlations caused by deceptive background. + +# 3.2. Pixel-level Camouflage Decoupling + +In order to decouple camouflage characteristics and target information fused in the pixel feature, we first model the representation of camouflage characteristics by exploring Fourier frequency domain. Then we design a difference attention mechanism to remove camouflage characteristics while maintaining target information. + +Camouflage Characteristics Extraction. Generally, the camouflage characteristics are mainly comprised of color and texture information [10, 22]. On the other hand, in the frequency domain, it is known that the amplitude component of Fourier spectrum preserves low-level statistics information. Therefore, we can utilize the amplitude of Fourier spectrum to represent background information in the camouflaged image. Specifically, given a camouflaged image $x \in \mathbb{R}^{H \times W \times 3}$ , its Fourier transformation $\mathcal{F}(x)$ is formulated as: + +$$ +\mathcal {F} (x) _ {u, v} = \sum_ {i = 0} ^ {H - 1} \sum_ {j = 0} ^ {W - 1} x _ {i, j} e ^ {- J 2 \pi \left(\frac {i}{H} u + \frac {j}{W} v\right)}, \tag {1} +$$ + +where $J$ refers to the imaginary unit and each channel of the image is computed independently. The amplitude and phase components are then respectively expressed as: + +$$ +\mathcal {A} (x) _ {u, v} = \left[ R ^ {2} (x) _ {u, v} + I ^ {2} (x) _ {u, v} \right] ^ {1 / 2} +$$ + +$$ +\mathcal {P} (x) _ {u, v} = \arctan \left[ \frac {I (x) _ {u , v}}{R (x) _ {u , v}} \right], \tag {2} +$$ + +where $R(x)$ and $I(x)$ represent the real and imaginary part of $\mathcal{F}(x)$ , respectively. To represent the camouflaged background, we fix the phase to a constant $b$ (e.g., average value) and reconstruct the image with the amplitude information as + +$$ +\tilde {x} = \mathcal {F} ^ {- 1} [ \mathcal {A} (x) _ {u, v} e ^ {- J b} ], \tag {3} +$$ + +where $\mathcal{F}^{-1}$ indicates the inverse Fourier transformation. Then, we feed the reconstructed image $\tilde{x}$ into a light-weight CNN (e.g., ResNet-18 [15]) with global average pooling layer to obtain the global camouflage characteristics $g$ of the original image. + +Difference Attention Mechanism. Given image features $\mathbf{F} \in \mathbb{R}^{h \times w \times c}$ extracted from a backbone network (e.g., ResNet-50 [15]), we remove the camouflage characteristics $\pmb{g}$ while retaining valuable target information by a novel difference attention mechanism. In specific, we first utilize two $1 \times 1$ convolution layers to map $\mathbf{F}$ and $\pmb{g}$ to the same dimension $C$ , resulting in $\hat{\mathbf{F}} \in \mathbb{R}^{h \times w \times C}$ and $\hat{\pmb{g}} \in \mathbb{R}^C$ , respectively. The discrepancy between image features and camouflage characteristics indicates the saliency information of the target objects, helping to highlight the target information. Thus, we integrate the difference on all channels for each pixel to obtain a difference map $\mathbf{D} \in \mathbb{R}^{h \times w}$ : + +$$ +\mathbf {D} _ {i, j} = \sum_ {k = 1} ^ {C} \left(\hat {\mathbf {F}} _ {i, j, k} - \hat {\boldsymbol {g}} _ {k}\right) ^ {2}, \tag {4} +$$ + +where $i,j$ , and $k$ are the index of height, width, and channel, respectively. By multiplying the original image features with the difference map $\mathbf{D}$ , we can get the de-camouflaged pixel features $\mathbf{F}^d$ as + +$$ +\mathbf {F} ^ {d} = \mathbf {F} \circ \mathbf {D}, \tag {5} +$$ + +where $\circ$ refers to the Hadamard product and $\mathbf{D}$ is broadcasted to the same dimension as $\mathbf{F}$ . In addition, we follow [33] to use a deformable self-attention layer [47] to further incorporate context information from other pixels. + +Fusion Layer. To acquire fine-grained target information for more accurate segmentation, we use multi-scale features from different stages of the backbone. Specifically, we first obtain the features from different stages of the backbone network, and then obtain the corresponding de-camouflaged features through the difference attention mechanism. Finally, we fuse the processed features with the FPN [24] network to generate the high-resolution pixel-level feature $\mathbf{E}$ , which is used for subsequent instance segmentation. + +# 3.3. Instance-level Camouflage Suppression + +In order to capture different camouflaged instances, we learn a set of instance prototypes $\mathbf{P} = \{\pmb{p}^i\}_{i=1}^N$ , and each $\pmb{p}^i \in \mathbb{R}^{1 \times L}$ is responsible for identifying whether each pixel belongs to a camouflaged instance. To eliminate the background interference for prototype learning, we design reference attention to aggregate target information. + +Reference Attention Mechanism. Given the decamouflaged pixel feature $\mathbf{F}^d$ derived from the pixel-level camouflage decoupling module, we first obtain the queries $\mathbf{Q} = [q_1;q_2;\dots;q_N]$ from instance prototypes $\mathbf{P}$ , keys $\mathbf{K} = [k_{1};k_{2};\dots;k_{hw}]$ and values $\mathbf{V} = [v_{1};v_{2};\dots;v_{hw}]$ from pixel features $\mathbf{F}^d = [f_1;f_2;\dots;f_{hw}]$ as + +$$ +\boldsymbol {q} _ {i} = \boldsymbol {p} _ {i} \mathbf {W} ^ {q}, \boldsymbol {k} _ {j} = \boldsymbol {f} _ {j} \mathbf {W} ^ {k}, \boldsymbol {v} _ {j} = \boldsymbol {f} _ {j} \mathbf {W} ^ {v}, \tag {6} +$$ + +where $i \in 1,2,\dots,N, j \in 1,2,\dots,hw$ , and $\mathbf{W}^q \in \mathbb{R}^{L \times L}$ , $\mathbf{W}^k$ , $\mathbf{W}^v \in \mathbb{R}^{c \times L}$ are linear projections. With the prototype queries, we can compute the correlation between each query-key pair independently as + +$$ +s _ {i, j} = \frac {d \left(\boldsymbol {q} _ {i} , \boldsymbol {k} _ {j}\right)}{\sqrt {L}}, \tag {7} +$$ + +where $d(\cdot, \cdot)$ denotes the distance metric which is the dot product similarity in [39], and $\sqrt{L}$ is a scaling factor. As the camouflaged object presents high similarity with the surrounding backgrounds, the prototype-pixel correlation represented by $s_{i,j}$ is susceptible to background pixel interference, resulting in erroneous segmentation. Intuitively, humans are able to identify camouflaged instances by repeatedly comparing the target with the surrounding reference areas. Motivated by this, we first use an adaptive selection process to select some reliable pixels from de-camouflaged pixels $\mathbf{F}^d$ as references. In specific, to measure the reliability of each pixel, we define and calculate the total contribution of each pixel with regard to all the prototypes, which is formulated as: + +$$ +u _ {j} = \sum_ {i = 1} ^ {N} s _ {i, j}, j \in 1, 2, \dots , h w. \tag {8} +$$ + +Then we select top- $K$ pixels with the largest correlations with the prototypes as reference points $\mathbf{R} \in \mathbb{R}^{K \times L}$ . In this way, we can pick out reliable reference points that are adaptive to the instance prototypes. Based on reference points, we first calculate the prototype-reference and pixel-reference similarity as + +$$ +s _ {i} ^ {q} = \phi_ {1} \left(\boldsymbol {q} _ {i} \mathbf {R} ^ {\top}\right), s _ {j} ^ {k} = \phi_ {2} \left(\boldsymbol {k} _ {j} \mathbf {R} ^ {\top}\right), \tag {9} +$$ + +where $\phi_{1}$ and $\phi_{2}$ are feed-forward networks (FFN) containing two fully connected layers. $s_i^q$ and $s_j^k$ can serve as soft multilabel to measure the correlation between the $i$ -th prototype and the $j$ -th pixel: + +$$ +s _ {i, j} ^ {q k} = d \left(\boldsymbol {q} _ {i}, \boldsymbol {k} _ {j}; \mathbf {R}\right) = s _ {i} ^ {q} \left(s _ {j} ^ {k}\right) ^ {\top}. \tag {10} +$$ + +Compared to $s_{i,j}$ , $s_{i,j}^{qk}$ leverages the consensus among reliable reference points, thus suppressing the background noises and leading to fewer erroneous correlations. Given the correlation $s_{i,j}^{qk}$ , we can extract and purify target information from pixel features to update instance prototypes as + +$$ +\tilde {\boldsymbol {p}} _ {i} = \sum_ {j = 1} ^ {h w} a _ {i, j} \boldsymbol {v} _ {j}, a _ {i, j} = \frac {\exp \left(s _ {i , j} ^ {q k}\right)}{\sum_ {j = 1} ^ {h w} \exp \left(s _ {i , j} ^ {q k}\right)}. \tag {11} +$$ + +Note that we have omitted the scaling factor and the multihead mechanism for notation simplicity. + +Given the high-resolution pixel-level feature embedding $\mathbf{E}$ and the learned instance prototypes $\tilde{\pmb{p}}_i$ , we can obtain N mask predictions and according scores by + +$$ +m _ {i} = \operatorname {s i g m o i d} \left(\mathbf {E} \tilde {\boldsymbol {p}} _ {i} ^ {\top}\right), y _ {i} = h \left(\tilde {\boldsymbol {p}} _ {i} \mid \sigma\right), \tag {12} +$$ + +where $h(\cdot \mid \sigma)$ is a classifier parameterized by $\sigma$ to predict the confidence score. Following [6], we match a ground truth label (including "no object") for each predicted segment by the Hungarian matching algorithm. For training, we use a cross-entropy loss to constrain the instance score $y_{i}$ , and a linear combination of a focal loss [25] and a dice loss [30] to constrain the mask prediction $m_{i}$ . + +# 4. Experiment + +# 4.1. Experiment Setup + +Datasets. We evaluate our method on two datasets: COD10K [10] and NC4K [28]. COD10K contains 5086 camouflaged images with high-quality instance-level annotations, which are divided into 3040 training images and 2026 testing images. NC4K contains 4121 test camouflaged images to evaluate the generalization ability of CIS models. Following the standard benchmark [33], we use the instance-level annotations in COD10K to train our DCNet and evaluate it on the test set of COD10K and NC4K. + +Evaluation Metrics. We adopt $\mathrm{AP}_{50}$ , $\mathrm{AP}_{75}$ and AP scores as evaluation metrics to quantify the effectiveness of our method. A true positive is counted if the intersection over union (IoU) between the ground truth and the segmentation is at least $50\%$ or $75\%$ , and the AP score is an overall metric that combines different IoU thresholds. + +Implementation Details. For fair comparisons, the feature extractor is implemented by ResNet-50 [15] which is pretrained on ImageNet [8]. During our training, our model is trained with batch size of 2, using the Adam optimizer [27] with an initial learning rate of 0.0001 for 90,000 iterations. We set the channel dimension $C = L = 256$ . In the instance-level camouflage suppression module, we set the number of prototypes $N$ as 10, and the number of reference points $K$ as 64, which turns out to generate the best segmentation during our experiments as shown in Figure 4. + +# 4.2. Comparison with State-of-the-arts + +Quantitative results. Considering that CIS is a newly emerging task, we also compare some general instance segmentation methods besides the previous CIS methods and all the methods adopt ResNet-50 [15] backbone for fair comparisons. As demonstrated in Table 1, the proposed DCNet consistently outperforms the state-of-the-art methods by a large margin on both datasets. (1) On the COD10K [10], our method outperforms the previous best method OSFormer [33] by $4.3\%$ in AP, indicating that our DCNet is able to acquire more accurate segmentation of camouflaged targets. (2) On the NC4K [28], Our model yields an accuracy of $52.8\%$ in AP, making an obvious performance improvement by $7.0\%$ in AP compared to the second best performing method Mask2Former [5]. Since the model is trained on the COD10K training set, the high + +Table 1. Comparisons of existing CIS methods and general instance approaches on COD10K [10] and NC4K [28] testing set, where all the methods use ResNet-50 as feature extractor. The best results are shown in bold. + +
MethodsCOD10K-TestNC4K-TestParams(M)GFLOPs
AP\( AP_{50} \)\( AP_{75} \)AP\( AP_{50} \)\( AP_{75} \)
Mask R-CNN [14]25.055.520.427.758.622.743.9186.3
MS R-CNN [17]30.157.228.731.058.729.460.0198.5
Cascade R-CNN [2]25.356.121.329.560.824.871.7334.1
HTC [4]28.156.325.129.859.026.676.9331.7
BlendMask [3]28.256.425.227.756.724.235.8233.8
Mask Transfiner [20]28.756.326.429.456.727.244.3185.1
YOLACT [1]24.353.319.732.165.327.9--
CondInst [38]30.663.626.133.467.429.434.1200.1
QueryInst [12]28.560.123.133.066.729.4--
SOTR [13]27.958.724.129.361.025.663.1476.7
SOLOv2 [41]32.563.229.934.465.931.946.2318.7
MaskFormer [6]38.265.137.944.671.945.845.0174.2
Mask2Former [5]39.467.738.545.873.647.543.9241.0
OSFormer [33]41.071.140.842.572.542.346.6324.7
DCNet (ours)45.370.747.552.877.156.553.4207.0
+ +Table 2. Comparison of the different components in pixel-level camouflage decoupling module in terms of AP scores. + +
COD10KNC4K
PCD (ours)45.352.8
-camouflage attribute extractor43.1(-2.2)48.1(-4.7)
-difference attention mechanism44.1(-1.2)51.3(-1.5)
-both 2 components above40.2(-5.1)46.4(-6.4)
+ +performance on NC4K indicates that our method also has better generalization ability. Notice that our method does not achieve the highest performance on $\mathrm{AP}_{50}$ metric for the NC4K because different from the previous CIS method, we do not use additional Non-maximum suppression (NMS) post-processing to remove redundant predictions. For $\mathrm{AP}_{50}$ , our more accurate segmentation results cannot increase the number of true positive (TP) under the $50\%$ IoU threshold, while redundant predictions will increase the number of false positive (FP), thus our method has no obvious advantages. For $\mathrm{AP}_{75}$ , our high precision segmentation will increase TP with a large margin. The overall metric AP reflects the superiority of our method over other methods. There is a similar trend on NC4K. In addition, the lower GFLOPs of our DCNet compared to state-of-the-art methods indicates that our model achieves high segmentation ability while maintaining high computation efficiency. + +Qualitative Results. As shown in Figure 3, our proposed method is capable of de-camouflaging by using the difference attention mechanism at the pixel level, and also able to identify the accurate locations of target camouflaged objects at the instance level. Compared to previous methods, + +Table 3. Comparisons of different attention mechanisms in the instance-level camouflage suppression module in terms of AP scores. + +
Attention MechanismCOD10KNC4K
cross-attention [39]42.449.5
masked attention [5]44.751.7
reference attention (ours)45.352.8
+ +our DCNet performs better at boundaries of camouflage instances (see the first 4 columns), suppresses distracting background regions (see $5^{th}$ to $6^{th}$ columns), and distinguishes multiple instances well (see the last 2 columns). + +# 4.3. Ablation Study + +We conduct comprehensive ablation studies on COD10K and NC4K to verify the effectiveness of our modules. + +Effectiveness of the Pixel-level Camouflage Decoupling Module. The pixel-level camouflage decoupling module mainly consists of a camouflage attribute extractor and a difference attention mechanism. We validate the importance of each component by removing them one at a time. As shown in Table 2, we replace the proposed camouflage attribute extractor with a learnable vector, and the performance drops by an average of $3.5\%$ across two datasets. It indicates that the Fourier spectrum amplitude is effective in modeling the camouflage characteristics, verifying that it is hard to learn de-camouflage without image-specific camouflage information supervision. Then we replace the difference attention mechanism with general subtraction operations. The reduced performance indicates that our pro + +![](images/f53f37466a24f47b59e089a7fb594e9755d4c6aec6c3a52f2428c7992f4dcccb.jpg) +Original image + +![](images/cc2b5806073f5e7dec46812d7dc1d5de022ed9090ec03822906887b55f861ae3.jpg) + +![](images/aa62ac7220e556d24e0a874ae62fc6f1a708726f4bb978b10eac62cf9c85b379.jpg) + +![](images/988603c351cbf2895a772229e45782d3b2997c6323d9f64448363f36ab3e4ad5.jpg) + +![](images/d38ae4f9e9c4b6e1ffe9ae145f2d9f04165f93cfd477ca225262da6b0f0d379c.jpg) + +![](images/fa0480c307b9c7548b0b0e97d22923df309fefb571aadf4a4a45668d7b288a98.jpg) + +![](images/45e9b494e3843707d4d4d85426bc4c2dfca555fb2392b3efc7505ae9ba17a454.jpg) + +![](images/384068335ec1bc22ac0215ca0d2e06e9942bd225a9ec630b89d1b0b4d4fe6c33.jpg) + +![](images/19776a8e98bdf84828fd51a0add9ec12917c7da9eeba01121d43ba1526b3256a.jpg) +GT + +![](images/4a61791df68fd000b715567e9829a21b7ee5ed297d7467ae1f7e153677956481.jpg) + +![](images/7ab5602b2131e8c9e153fee6a1579ecc7b6b920d98493356693995c2b302b92b.jpg) + +![](images/e564528de3fad60a33fe9de07f591c906022f58b251ca2faa60ae4744cf51bae.jpg) + +![](images/d2e2c9545c62bcedbbfe0eb929a8e720ddb9db2ee910fcf9107e1e198d10b60e.jpg) + +![](images/50e4dcb024d1f85e9fdd67e5eb3b577b9dc1feb1e9f7c53ca7883b4917da1cd1.jpg) + +![](images/79376126f34aa029127fd3ce5605754ed4cdf4223f5d396e986e61c7aa33bce3.jpg) + +![](images/e1173728ada22a4838013b3852b45ebaa1e8fdfe9c3726be9dc8882250753765.jpg) + +![](images/c0e1d0c951d5130119bcae556b19774997840fde610df1d97d573d86cf653e9d.jpg) +MaskFormer + +![](images/db591762c71826ecaa7158c2b12bf95683049006dfebdb304c7e20371b6f4515.jpg) + +![](images/d8f20f5e6d68073aa3cc050da8834f5a3e3aec5dd2f010e2ec059f69b27b4331.jpg) + +![](images/fe5ce7aaafc8662297caa99a7e5b6cb10d9c2e0dba72a5ff72f9437a1e4313cc.jpg) + +![](images/8c28960cd6739c9b272ef918b04a8340319234ea4eafe5bc68a854067c0026a1.jpg) + +![](images/387c34e44e92582e9e53cc71016a1dd0564a869b596af22241cae2ea831d9fad.jpg) + +![](images/4362116f388be5ea0ff4b9b641d1388b18a2f9e4008b304b7764ee9dc3975d37.jpg) + +![](images/73b5df6e3e23054b0ae25a6f84887078d81c69e2b19f79724f117e1b3c8bce66.jpg) + +![](images/992f9eb9b56d378fed232f07294004d8d4a64f9e50ddee93fadad7d53120ce7a.jpg) +Mask2Former + +![](images/e4c446a0e1b243d0ba4616df93f3fc619aea03f1115eca0b52481f68db755730.jpg) + +![](images/52ceae1732bbbf2969d17510fd189d136709c4327dd09a037753090f56946505.jpg) + +![](images/0ee51e6fb06d8fe662d524134080c2d8804bca6a13e85567d145f99ede169a97.jpg) + +![](images/8938e82d1c75c88eb2d4fff5ebae2a7f54789fdacb59b1c22fad9fd7b06949e5.jpg) + +![](images/d6d11c388168127d855e356382384276441265e12b25a8a8f31a01c9706eef98.jpg) + +![](images/338d66e27e360a3afeca42a0b0120f27d8b718e9bca7b3e8ceea22d3e14617ea.jpg) + +![](images/ee5769b9c410b51a87be24f13381f1194620b63864b1a475a037c2b34e354d86.jpg) + +![](images/666cd1fd657c449bf2f31ec1108b8ffa3852d52648d2e20ecd5022e3c1763d18.jpg) +OSFormer + +![](images/5cb7de8580d9977f01b0adaf79a660a05c3a6ac29daa5742dbc2200f136aee71.jpg) + +![](images/021537d830b0c02cae219be65e3adaa7585fa476e0d3da773c93c9c4969d86d7.jpg) + +![](images/55fa1b75b3ce068fdaa6436a0d3f9f75c35cea3c544e49fc30eaa020edcf2289.jpg) + +![](images/a4ab53fc701fdbca6376516bcaf294ceae9bf200c883e8e9fa85b0c2c6886e11.jpg) + +![](images/2bda3a983111f15b722135ccaac65e512c88f0e4bb4e2d39131329fdb34a533d.jpg) + +![](images/9051c92ca773f440375f4221f3ac1b771bad6d22f912f7e73ef76fc8028f5ba0.jpg) + +![](images/f59478999595bfed7d1d79302b3b8a82ca2729b857a336e97f61478a5214b71a.jpg) + +![](images/9b36f8506c4dbe9fc3f03e3fd04d4bcdcddf6b87e40439ebc2ca55726135cc66.jpg) +DCNet(ours) + +![](images/a371356dedced26ac5dd6fe7c380ff55ce48f7d2f9f95ffcea356fddf7a01ef6.jpg) +Figure 3. Qualitative results of DCNet compared with state-of-the-art methods. Different colored masks represent different instances. + +![](images/3d6a4e5dc3d01faf1fd49cafce8039f219c4833f256bbe2f6ccc2973f2383142.jpg) + +![](images/7f795e4059003961d5fee101a687522949be5dbadb252c67f2a3432699780fac.jpg) + +![](images/ad88d32e41c940cf5ae7d250161964e1bdbca807d485b16e9cfddbfc7d95bc40.jpg) + +![](images/5dfb3dadd3b9d42bd8480a69574718a5938f8dacc59d4eb040d8a1f25e8f9f33.jpg) + +![](images/ee0272c1bb62fcfd5c28dd059da0e56ffdbcbbe9a7f60709c12aeda8d4f97f80.jpg) + +![](images/20300b7141a95ed2f7c01422085b07c13de7d9a89d8b79008aa86e8d6d01a06d.jpg) + +Table 4. Comparison of using different scales of features in terms of AP scores. + +
COD10KNC4K
single scale (1/32)45.252.5
single scale (1/16)44.851.6
single scale (1/8)45.152.2
multiple scales45.352.8
+ +posed mechanism can well model the discrepancy between image features and camouflage characteristics. Removing both of the components leads to huge performance degradation, showing that the combination of them is essential in improving the camouflaged object segmentation. + +Effectiveness of the Instance-level Camouflage Suppression Module. As shown in Table 3, the reference attention mechanism in our instance-level camouflage suppression module has the highest AP score on both datasets. Compared to vanilla cross-attention [39], our proposed reference points can effectively suppress the noise brought by background pixels in the prototype-pixel interaction, thus obtaining significant improvement for accurate segmentation. Meanwhile, our method is superior to existing cross-attention variants, such as the masked attention [5] in the segmentation field. + +Effectiveness of Hierarchical Features To explore decamouflage at multiple feature levels, we feed 3 scales of features (i.e., stride of 32, 16, and 8) into the difference attention mechanism, respectively. As shown in Table 4, low-resolution features exceed other scales. This may attribute + +Table 5. Comparisons with OSFormer [33] of different backbones on COD10K [10] and NC4K [28] in terms of AP scores. + +
MethodBackboneCOD10KNC4K
OSFormerR5041.042.5
DCNet (ours)45.352.8
OSFormerR10142.044.4
DCNet (ours)46.853.5
OSFormerSwin-T47.750.2
DCNet (ours)50.356.3
OSFormerSwin-S52.156.7
DCNet (ours)52.358.4
+ +![](images/bbf4c775b87a1d4b877c62d8215680e0c3e8a79611a80ab254b31c89474a0b8b.jpg) +Figure 4. Comparisons of performance with different numbers of prototypes and reference points on COD10K [10] and NC4K [28]. + +![](images/a9381e3e2a2f3d0974eb78b27e3b87aaf17951f8f8f1d381786999070e2e9310.jpg) + +to the retention of object semantic information. The highest performance in this experiment shows that the contribution of each scale is complementary to each other. Using single-scale features alone is not enough to model the target information, thus combining them together can further boost the performance. + +Compatibility for Feature Extractor. We equip our DC-Net with different feature extractor backbones, i.e., ResNet + +![](images/b56eaac3c89497121214786ee85ee89972233eca8006d645d650bd0ac78b69f6.jpg) +Figure 5. Illustration of the amplitude images and difference maps in our proposed difference attention. + +50 [15], ResNet-101 [15], Swin-T [26], and Swin-S [26]. All of them are pre-trained on ImageNet-1K [8]. As shown in Table 5, our method performs the best in all cases on both datasets, proving that our approach has tremendous potential for further improvement. + +Analysis of Hyperparameters. In the instance-level camouflage suppression module, prototypes play an important role in capturing instances by aggregating object pixels. As shown in Figure 4(a), we conduct quantitative experiments to analyze how many prototypes are suitable for instance segmentation. It can be seen that too many or too few prototypes will damage the AP metric. In fact, the number of prototypes $N$ should be larger than the count of instances in an image to avoid instance fusion error. While superabundant prototypes will cause many prototypes to match "no object" during training. We choose $N = 10$ by balancing the performance on both datasets. Besides, in the reference attention mechanism, the number of reference points $K$ influences the similarity distributions between prototypes and pixels. As shown in Figure 4(b), we can observe that the performance on NC4K continues to grow until $K = 64$ , which means that it is sufficient for building appropriate prototype-pixel correction. + +# 4.4. Visualizations + +Visualization of Difference Map. As shown in Figure 5, the $2^{nd}$ row exhibits the amplitude images that are generated through Fourier transformation, which can be considered as camouflage characteristics. With the help of the difference attention mechanism, our network can remove the confusing background and localize the accurate salient areas (see $3^{rd}$ row), which turn out to be the correct regions of camouflaged instances. + +Visualization of Reference Attention. To explore the reference attention mechanism's ability to suppress the background noise, we make qualitative visualization to compare + +![](images/36ae8d4fc95b96686bddad08067ee335e0d265ebd8cdfdf1931d64641717b526.jpg) +Figure 6. Visualizations of the activation map with different attention mechanisms. + +different attention mechanisms. As shown in Figure 6 (d), with the proposed reference attention, the high response is clustered in the foreground area and the prototype pays little attention to the background area. In contrast, in Figure 6 (b), prototypes with the cross-attention [39] tend to absorb pixels from the non-target region. And in Figure 6 (c), with the masked attention [5], most background pixels are masked, while some foreground pixels are also suppressed, leading to sub-optimal performance. Thus, the reference mechanism plays an important role in mitigating the effect of background noise in prototype-pixel interactions, which helps to achieve more accurate segmentation. + +# 5. Conclusion + +In this paper, we propose an end-to-end Decamouflaging Network (DCNet) by jointly modeling pixel-level camouflage decoupling and instance-level camouflage suppression for camouflaged instance segmentation. We design a Pixel-level Camouflage Decoupling module to model and further eliminate the camouflage characteristics based on the frequency domain information. Besides, an Instance-level Camouflage Suppression module is proposed to achieve prototype-based instance segmentation and mitigate the effect of background noise by introducing reference points. Extensive experimental results on two benchmark datasets demonstrate the effectiveness of the proposed method. + +# 6. Acknowledgment + +This work was partially supported by the National Nature Science Foundation of China (Grant 62022078, 62021001), National Defense Basic Scientific Research Program of China (Grant JCKY2021601B013). + +# References + +[1] Daniel Bolya, Chong Zhou, Fanyi Xiao, and Yong Jae Lee. Yolact: Real-time instance segmentation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9157-9166, 2019. 3, 6 +[2] Zhaowei Cai and Nuno Vasconcelos. Cascade r-cnn: high quality object detection and instance segmentation. IEEE transactions on pattern analysis and machine intelligence, 43(5):1483-1498, 2019. 6 +[3] Hao Chen, Kunyang Sun, Zhi Tian, Chunhua Shen, Yongming Huang, and Youliang Yan. Blendmask: Top-down meets bottom-up for instance segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8573-8581, 2020. 6 +[4] Kai Chen, Jiangmiao Pang, Jiaqi Wang, Yu Xiong, Xiaoxiao Li, Shuyang Sun, Wansen Feng, Ziwei Liu, Jianping Shi, Wanli Ouyang, et al. Hybrid task cascade for instance segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4974-4983, 2019. 6 +[5] Bowen Cheng, Ishan Misra, Alexander G Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1290–1299, 2022. 2, 3, 5, 6, 7, 8 +[6] Bowen Cheng, Alex Schwing, and Alexander Kirillov. Per-pixel classification is not all you need for semantic segmentation. Advances in Neural Information Processing Systems, 34:17864-17875, 2021. 2, 3, 5, 6 +[7] Hung-Kuo Chu, Wei-Hsin Hsu, Niloy J Mitra, Daniel Cohen-Or, Tien-Tsin Wong, and Tong-Yee Lee. Camouflage images. ACM Trans. Graph., 29(4):51-1, 2010. 1 +[8] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 5, 8 +[9] Deng-Ping Fan, Ge-Peng Ji, Ming-Ming Cheng, and Ling Shao. Concealed object detection. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2021. 1 +[10] Deng-Ping Fan, Ge-Peng Ji, Guolei Sun, Ming-Ming Cheng, Jianbing Shen, and Ling Shao. Camouflaged object detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2777-2787, 2020. 2, 4, 5, 6, 7 +[11] Deng-Ping Fan, Tao Zhou, Ge-Peng Ji, Yi Zhou, Geng Chen, Huazhu Fu, Jianbing Shen, and Ling Shao. Inf-net: Automatic Covid-19 lung infection segmentation from ct images. IEEE Transactions on Medical Imaging, 39(8):2626-2637, 2020. 1 +[12] Yuxin Fang, Shusheng Yang, Xinggang Wang, Yu Li, Chen Fang, Ying Shan, Bin Feng, and Wenyu Liu. Instances as queries. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6910-6919, 2021. 6 +[13] Ruohao Guo, Dantong Niu, Liao Qu, and Zhenbo Li. Sotr: Segmenting objects with transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7157-7166, 2021. 6 + +[14] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In Proceedings of the IEEE international conference on computer vision, pages 2961-2969, 2017. 1, 3, 6 +[15] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 4, 5, 8 +[16] Xiaobin Hu, Deng-Ping Fan, Xuebin Qin, Hang Dai, Wenqi Ren, Ying Tai, Chengjie Wang, and Ling Shao. High-resolution iterative feedback network for camouflaged object detection. arXiv preprint arXiv:2203.11624, 2022. 3 +[17] Zhaojin Huang, Lichao Huang, Yongchao Gong, Chang Huang, and Xinggang Wang. Mask scoring r-cnn. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6409-6418, 2019. 3, 6 +[18] Ivan Huerta, Daniel Rowe, Mikhail Mozerov, and Jordi González. Improving background subtraction based on a casuistry of colour-motion segmentation problems. In Iberian Conference on Pattern Recognition and Image Analysis, pages 475-482. Springer, 2007. 2 +[19] Ge-Peng Ji, Deng-Ping Fan, Yu-Cheng Chou, Dengxin Dai, Alexander Liniger, and Luc Van Gool. Deep gradient learning for efficient camouflaged object detection. Machine Intelligence Research, 20(1):92-108, 2023. 3 +[20] Lei Ke, Martin Danelljan, Xia Li, Yu-Wing Tai, Chi-Keung Tang, and Fisher Yu. Mask transfiner for high-quality instance segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4412–4421, 2022. 6 +[21] Alexander Kirillov, Yuxin Wu, Kaiming He, and Ross Girshick. Pointrend: Image segmentation as rendering. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9799-9808, 2020. 3 +[22] Trung-Nghia Le, Yubo Cao, Tan-Cong Nguyen, Minh-Quan Le, Khanh-Duy Nguyen, Thanh-Toan Do, Minh-Triet Tran, and Tam V Nguyen. Camouflaged instance segmentation in-the-wild: Dataset, method, and benchmark suite. IEEE Transactions on Image Processing, 31:287-300, 2021. 1, 4 +[23] Trung-Nghia Le, Tam V Nguyen, Zhongliang Nie, MinhTriet Tran, and Akihiro Sugimoto. Anabranch network for camouflaged object segmentation. Computer Vision and Image Understanding, 184:45-56, 2019. 2 +[24] Tsung-Yi Lin, Piotr Dólar, Ross Girshick, Kaiming He, Bharath Hariharan, and Serge Belongie. Feature pyramid networks for object detection. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2117-2125, 2017. 4 +[25] Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dólar. Focal loss for dense object detection. In Proceedings of the IEEE international conference on computer vision, pages 2980-2988, 2017. 5 +[26] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10012-10022, 2021. 8 + +[27] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 5 +[28] Yunqiu Lv, Jing Zhang, Yuchao Dai, Aixuan Li, Bowen Liu, Nick Barnes, and Deng-Ping Fan. Simultaneously localize, segment and rank the camouflaged objects. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11591-11601, 2021. 5, 6, 7 +[29] Haiyang Mei, Ge-Peng Ji, Ziqi Wei, Xin Yang, Xiaopeng Wei, and Deng-Ping Fan. Camouflaged object segmentation with distraction mining. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8772-8781, 2021. 3 +[30] Fausto Milletari, Nassir Navab, and Seyed-Ahmad Ahmadi. V-net: Fully convolutional neural networks for volumetric medical image segmentation. In 2016 fourth international conference on 3D vision (3DV), pages 565-571. IEEE, 2016. 5 +[31] Yuxin Pan, Yiwang Chen, Qiang Fu, Ping Zhang, and Xin Xu. Study on the camouflaged target detection method based on 3d convexity. Modern Applied Science, 5(4):152, 2011. 2 +[32] Youwei Pang, Xiaoqi Zhao, Tian-Zhu Xiang, Lihe Zhang, and Huchuan Lu. Zoom in and out: A mixed-scale triplet network for camouflaged object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2160-2170, 2022. 3 +[33] Jialun Pei, Tianyang Cheng, Deng-Ping Fan, He Tang, Chuanbo Chen, and Luc Van Gool. Osformer: One-stage camouflaged instance segmentation with transformers. In European Conference on Computer Vision, pages 19-37. Springer, 2022. 1, 2, 4, 5, 6, 7 +[34] Leon N Piotrowski and Fergus W Campbell. A demonstration of the visual importance and flexibility of spatial-frequency amplitude and phase. Perception, 11(3):337-346, 1982. 2, 3 +[35] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. Advances in neural information processing systems, 28, 2015. 3 +[36] P Sengottuvelan, Amitabh Wahi, and A Shanmugam. Performance of decamouflaging through exploratory image analysis. In 2008 First International Conference on Emerging Trends in Engineering and Technology, pages 6-10. IEEE, 2008. 2 +[37] Sujit K Singh, Chitra A Dhawale, and Sanjay Misra. Survey of object detection methods in camouflaged image. IERI Procedia, 4:351-357, 2013. 1 +[38] Zhi Tian, Chunhua Shen, and Hao Chen. Conditional convolutions for instance segmentation. In European conference on computer vision, pages 282-298. Springer, 2020. 6 +[39] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 5, 6, 7, 8 +[40] Xinlong Wang, Tao Kong, Chunhua Shen, Yuning Jiang, and Lei Li. Solo: Segmenting objects by locations. In European Conference on Computer Vision, pages 649-665. Springer, 2020. 1, 3 + +[41] Xinlong Wang, Rufeng Zhang, Tao Kong, Lei Li, and Chunhua Shen. Solov2: Dynamic and fast instance segmentation. Advances in Neural information processing systems, 33:17721-17732, 2020. 1, 3, 6 +[42] Qinwei Xu, Ruipeng Zhang, Ya Zhang, Yanfeng Wang, and Qi Tian. A fourier-based framework for domain generalization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14383-14392, 2021. 2, 3 +[43] Fan Yang, Qiang Zhai, Xin Li, Rui Huang, Ao Luo, Hong Cheng, and Deng-Ping Fan. Uncertainty-guided transformer reasoning for camouflaged object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4146-4155, 2021. 3 +[44] Xiaoqi Zhao, Lihe Zhang, and Huchuan Lu. Automatic polyp segmentation via multi-scale subtraction network. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 120-130. Springer, 2021. 1 +[45] Yijie Zhong, Bo Li, Lv Tang, Senyun Kuang, Shuang Wu, and Shouhong Ding. Detecting camouflaged object in frequency domain. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4504-4513, 2022. 3 +[46] Tao Zhou, Yi Zhou, Chen Gong, Jian Yang, and Yu Zhang. Feature aggregation and propagation network for camouflaged object detection. IEEE Transactions on Image Processing, 31:7036-7047, 2022. 3 +[47] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable detr: Deformable transformers for end-to-end object detection. arXiv preprint arXiv:2010.04159, 2020. 4 \ No newline at end of file diff --git a/2023/Camouflaged Instance Segmentation via Explicit De-Camouflaging/images.zip b/2023/Camouflaged Instance Segmentation via Explicit De-Camouflaging/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..255bc7e8d3ffab7d7bad30bce8f29e068c115b8c --- /dev/null +++ b/2023/Camouflaged Instance Segmentation via Explicit De-Camouflaging/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f4f2d7ed03eadca31c701cb2ab198d8c0a76630c596fda929a7a54737691db5a +size 738373 diff --git a/2023/Camouflaged Instance Segmentation via Explicit De-Camouflaging/layout.json b/2023/Camouflaged Instance Segmentation via Explicit De-Camouflaging/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..de17c2c68a44db094f2c61046b9e6ba9b6a61343 --- /dev/null +++ b/2023/Camouflaged Instance Segmentation via Explicit De-Camouflaging/layout.json @@ -0,0 +1,10460 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 93, + 103, + 501, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 93, + 103, + 501, + 121 + ], + "spans": [ + { + "bbox": [ + 93, + 103, + 501, + 121 + ], + "type": "text", + "content": "Camouflaged Instance Segmentation via Explicit De-camouflaging" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 78, + 142, + 514, + 199 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 78, + 142, + 514, + 199 + ], + "spans": [ + { + "bbox": [ + 78, + 142, + 514, + 199 + ], + "type": "text", + "content": "Naisong Luo" + }, + { + "bbox": [ + 78, + 142, + 514, + 199 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 78, + 142, + 514, + 199 + ], + "type": "text", + "content": ", Yuwen Pan" + }, + { + "bbox": [ + 78, + 142, + 514, + 199 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 78, + 142, + 514, + 199 + ], + "type": "text", + "content": ", Rui Sun" + }, + { + "bbox": [ + 78, + 142, + 514, + 199 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 78, + 142, + 514, + 199 + ], + "type": "text", + "content": ", Tianzhu Zhang" + }, + { + "bbox": [ + 78, + 142, + 514, + 199 + ], + "type": "inline_equation", + "content": "^{1,2,3\\dagger}" + }, + { + "bbox": [ + 78, + 142, + 514, + 199 + ], + "type": "text", + "content": ", Zhiwei Xiong" + }, + { + "bbox": [ + 78, + 142, + 514, + 199 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 78, + 142, + 514, + 199 + ], + "type": "text", + "content": ", Feng Wu" + }, + { + "bbox": [ + 78, + 142, + 514, + 199 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 78, + 142, + 514, + 199 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 78, + 142, + 514, + 199 + ], + "type": "text", + "content": "University of Science and Technology of China \n" + }, + { + "bbox": [ + 78, + 142, + 514, + 199 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 78, + 142, + 514, + 199 + ], + "type": "text", + "content": "Institute of Artificial Intelligence, Hefei Comprehensive National Science Center \n" + }, + { + "bbox": [ + 78, + 142, + 514, + 199 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 78, + 142, + 514, + 199 + ], + "type": "text", + "content": "Deep Space Exploration Lab" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 79, + 201, + 509, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 79, + 201, + 509, + 213 + ], + "spans": [ + { + "bbox": [ + 79, + 201, + 509, + 213 + ], + "type": "text", + "content": "{lns6,panyw，issunrui}@mail.ustc.edu.cn，{tzzhang，zwxiong，fengwu}@ustc.edu.cn" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 143, + 241, + 191, + 253 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 241, + 191, + 253 + ], + "spans": [ + { + "bbox": [ + 143, + 241, + 191, + 253 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 266, + 290, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 266, + 290, + 590 + ], + "spans": [ + { + "bbox": [ + 46, + 266, + 290, + 590 + ], + "type": "text", + "content": "Camouflaged Instance Segmentation (CIS) aims at predicting the instance-level masks of camouflaged objects, which are usually the animals in the wild adapting their appearance to match the surroundings. Previous instance segmentation methods perform poorly on this task as they are easily disturbed by the deceptive camouflage. To address these challenges, we propose a novel De-camouflaging Network (DCNet) including a pixel-level camouflage decoupling module and an instance-level camouflage suppression module. The proposed DCNet enjoys several merits. First, the pixel-level camouflage decoupling module can extract camouflage characteristics based on the Fourier transformation. Then a difference attention mechanism is proposed to eliminate the camouflage characteristics while reserving target object characteristics in the pixel feature. Second, the instance-level camouflage suppression module can aggregate rich instance information from pixels by use of instance prototypes. To mitigate the effect of background noise during segmentation, we introduce some reliable reference points to build a more robust similarity measurement. With the aid of these two modules, our DCNet can effectively model de-camouflaging and achieve accurate segmentation for camouflaged instances. Extensive experimental results on two benchmarks demonstrate that our DCNet performs favorably against state-of-the-art CIS methods, e.g., with more than " + }, + { + "bbox": [ + 46, + 266, + 290, + 590 + ], + "type": "inline_equation", + "content": "5\\%" + }, + { + "bbox": [ + 46, + 266, + 290, + 590 + ], + "type": "text", + "content": " performance gains on COD10K and NC4K datasets in average precision." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 614, + 127, + 626 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 614, + 127, + 626 + ], + "spans": [ + { + "bbox": [ + 47, + 614, + 127, + 626 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 635, + 287, + 684 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 635, + 287, + 684 + ], + "spans": [ + { + "bbox": [ + 47, + 635, + 287, + 684 + ], + "type": "text", + "content": "In the field of biology, camouflage is defined as a strategy that animals use to adapt their body's appearance (e.g., color and pattern) to match their surroundings in order to achieve concealing and avoid being hunted by predators [37]. Cam-" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 310, + 239, + 545, + 312 + ], + "blocks": [ + { + "bbox": [ + 310, + 239, + 545, + 312 + ], + "lines": [ + { + "bbox": [ + 310, + 239, + 545, + 312 + ], + "spans": [ + { + "bbox": [ + 310, + 239, + 545, + 312 + ], + "type": "image", + "image_path": "66787f4e73dd444f87b544ef18db1f07f67db8c0b40d93c802038274aed215cc.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 388, + 314, + 463, + 322 + ], + "lines": [ + { + "bbox": [ + 388, + 314, + 463, + 322 + ], + "spans": [ + { + "bbox": [ + 388, + 314, + 463, + 322 + ], + "type": "text", + "content": "(a) De-camouflaging Process" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 313, + 328, + 381, + 413 + ], + "blocks": [ + { + "bbox": [ + 313, + 328, + 381, + 413 + ], + "lines": [ + { + "bbox": [ + 313, + 328, + 381, + 413 + ], + "spans": [ + { + "bbox": [ + 313, + 328, + 381, + 413 + ], + "type": "image", + "image_path": "2aed844686c373604c0854ccf300539d523e2d5fd9c8d31f6b969f0b767e21b4.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 316, + 415, + 376, + 423 + ], + "lines": [ + { + "bbox": [ + 316, + 415, + 376, + 423 + ], + "spans": [ + { + "bbox": [ + 316, + 415, + 376, + 423 + ], + "type": "text", + "content": "(b) Pairwise Similarity" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 391, + 328, + 460, + 413 + ], + "blocks": [ + { + "bbox": [ + 391, + 328, + 460, + 413 + ], + "lines": [ + { + "bbox": [ + 391, + 328, + 460, + 413 + ], + "spans": [ + { + "bbox": [ + 391, + 328, + 460, + 413 + ], + "type": "image", + "image_path": "8267c25d9a4ed979ecfd2339ef67b13f8d70e95a14554a94aea56fd96abacd4a.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 422, + 415, + 504, + 423 + ], + "lines": [ + { + "bbox": [ + 422, + 415, + 504, + 423 + ], + "spans": [ + { + "bbox": [ + 422, + 415, + 504, + 423 + ], + "type": "text", + "content": "(c) Reference Similarity (Ours)" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 465, + 331, + 538, + 401 + ], + "blocks": [ + { + "bbox": [ + 465, + 331, + 538, + 401 + ], + "lines": [ + { + "bbox": [ + 465, + 331, + 538, + 401 + ], + "spans": [ + { + "bbox": [ + 465, + 331, + 538, + 401 + ], + "type": "image", + "image_path": "18279214b20cffc83e28d9e3eb508dc8d9c96ede24b00038523cec3bfa5ebb34.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 436, + 545, + 501 + ], + "lines": [ + { + "bbox": [ + 304, + 436, + 545, + 501 + ], + "spans": [ + { + "bbox": [ + 304, + 436, + 545, + 501 + ], + "type": "text", + "content": "Figure 1. Illustration of our motivation. (a) We propose to extract camouflage characteristics to model explicit de-camouflaging for CIS. (b) Pairwise similarity between prototype and pixel is always erroneous. (c) The prototype-pixel correlation is based on the prototype-reference and pixel-reference similarity distribution, which is more accurate." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 520, + 545, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 520, + 545, + 592 + ], + "spans": [ + { + "bbox": [ + 304, + 520, + 545, + 592 + ], + "type": "text", + "content": "outflaged Instance Segmentation (CIS) [22] aims at identifying the location and predicting instance-level masks of camouflaged objects, which has attracted more and more attention due to its widespread applications in medical image analysis [11, 44], search-and-rescue work [9] and recreational art [7], etc." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 594, + 546, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 546, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 546, + 714 + ], + "type": "text", + "content": "Despite the tremendous progress in instance segmentation [14, 40, 41], there are few efforts [22, 33] working on Camouflaged Instance Segmentation, as CIS is a more challenging task where most camouflaged instances lack obvious contrast with the background, making general instance segmentation methods work poorly on this task. Recent CIS approaches [22, 33] are generally based on traditional instance segmentation models, either by naively fusing various general instance segmentation models [22] to get better representations, or by directly using global in" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 494, + 35 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 494, + 35 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 494, + 35 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 693, + 124, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 693, + 124, + 703 + ], + "spans": [ + { + "bbox": [ + 58, + 693, + 124, + 703 + ], + "type": "text", + "content": "*Equal contribution" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 59, + 703, + 134, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 703, + 134, + 712 + ], + "spans": [ + { + "bbox": [ + 59, + 703, + 134, + 712 + ], + "type": "text", + "content": "† Corresponding author" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17918" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 251 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 251 + ], + "type": "text", + "content": "teraction [33]. However, these approaches fail to explore the core of CIS: de-camouflaging, i.e., eliminating camouflage characteristics of the target object. Without explicitly de-camouflaging, the previous methods are easily disturbed by similar background. De-camouflaging is a challenging problem as no image-specific camouflage information is provided as supervision signals. Intuitively, humans have the ability to quickly recognise objects in a highly camouflaged scene. They first repeatedly discriminate the real target characteristics from the camouflage characteristics at the pixel level, and then aggregate the pixel information to discern the whole target instance from the background. This human visual mechanism motivates us to explore de-camouflaging strategy from the pixel level to the instance level in a progressive manner." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 255, + 289, + 543 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 255, + 289, + 543 + ], + "spans": [ + { + "bbox": [ + 46, + 255, + 289, + 543 + ], + "type": "text", + "content": "In order to model de-camouflage at both the pixel and instance levels, a series of issues need to be considered. (1) The pixel level de-camouflage. In essence, each pixel feature of a camouflaged image contains the camouflage characteristics and the target object characteristics. Note that our goal is to remove camouflage characteristics and maintain target object information, thus a question naturally arises: How to decouple these two pieces of information at the pixel level? (2) The instance level segmentation. Based on the de-camouflaged pixels, we can naturally aggregate the pixel information and infer the mask of the target instance. Currently, the transformer-based models [5,6] have achieved leading performance for instance segmentation, where instance-specific prototypes are learned by constantly interacting with pixel features for final segmentation. However, directly applying the transformer to CIS is not trivial, as the prototypes would frequently absorb deceptive background information that has high similarity with the objects during the interaction, thus failing to discover desired targets accurately. As proved in Figure 1(b), the prototype fish is more similar to background pixels than target pixels in the camouflaged image. Therefore, we inevitably face another question: How to focus on camouflaged instances to achieve de-camouflaging?" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 546, + 289, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 546, + 289, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 546, + 289, + 714 + ], + "type": "text", + "content": "Motivated by the above discussions, we propose an end-to-end De-camouflaging Network (DCNet) by jointly modeling pixel-level camouflage decoupling and instance-level camouflage suppression for CIS. In the Pixel-level Camouflage Decoupling module (PCD), we focus on decoupling camouflage characteristics and target information fused in the pixel feature. First, we extract camouflage characteristics with the assistance of frequency domain information. The Fourier spectrum amplitude contains low-level statistics [34, 42] (e.g., color and texture of the environment, see Figure 5) that accords with the camouflage characteristics. Based on the obtained description of camouflage characteristics, we propose a novel difference attention mechanism to acquire de-camouflaged pixel fea" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 304, + 72, + 547, + 393 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 393 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 393 + ], + "type": "text", + "content": "tures. In this mechanism, we calculate the discrepancy between features of the original image and camouflage characteristics (see Figure 1(a)), thereby decoupling the camouflage characteristics and valuable target information while filtering out the background interference. In the Instance-level Camouflage Suppression module (ICS), we aggregate the de-camouflaged pixels to achieve final segmentation and meanwhile mitigate the effect of background noise in prototype-pixel interactions. Specifically, we introduce a set of instance prototypes to capture each camouflaged instance through long-range context-aware interactions. To constrain the interactions to favor target pixels over background pixels, we design a novel reference attention mechanism, where we select de-camouflaged pixels with high contribution to prototypes as reference points (see Figure 1(c)). Then we calculate the similarity of prototype-reference and pixel-reference, respectively, thereby obtaining similarity distributions serving as soft multilabel to measure correlations between prototypes and pixels. Highly similar pixels and prototypes must have consistent similarity distributions. In this way, the soft multilabel-based correlation is more accurate than the normal pairwise correlation, as it benefits from consensus among reliable reference points with a global receptive field. As a result, the correlation noise brought by deceptive background can be suppressed and more effective prototypes that contain rich instance information can be obtained." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 304, + 396, + 547, + 528 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 396, + 547, + 528 + ], + "spans": [ + { + "bbox": [ + 304, + 396, + 547, + 528 + ], + "type": "text", + "content": "The contributions of our method could be summarized as follows: (1) We propose a novel De-camouflaging Network (DCNet) by jointly modeling pixel-level camouflage decoupling and instance-level camouflage suppression for CIS. (2) We propose two effective designs in DCNet, i.e., difference attention mechanism and reference attention mechanism, which can highlight target information and suppress background interference. (3) Extensive experimental results on two benchmark datasets demonstrate the effectiveness of the proposed method, e.g., with more than " + }, + { + "bbox": [ + 304, + 396, + 547, + 528 + ], + "type": "inline_equation", + "content": "5\\%" + }, + { + "bbox": [ + 304, + 396, + 547, + 528 + ], + "type": "text", + "content": " performance gains on COD10K and NC4K datasets in average precision." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 306, + 536, + 392, + 550 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 536, + 392, + 550 + ], + "spans": [ + { + "bbox": [ + 306, + 536, + 392, + 550 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 557, + 547, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 557, + 547, + 605 + ], + "spans": [ + { + "bbox": [ + 304, + 557, + 547, + 605 + ], + "type": "text", + "content": "Since camouflaged instance segmentation is a new emerging task with very few papers, we mainly focus on introducing several lines of research in camouflaged object detection and instance segmentation." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 611, + 473, + 624 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 611, + 473, + 624 + ], + "spans": [ + { + "bbox": [ + 306, + 611, + 473, + 624 + ], + "type": "text", + "content": "2.1. Camouflaged Object Detection" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 629, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 629, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 629, + 547, + 715 + ], + "type": "text", + "content": "Camouflaged Object Detection (COD), aiming at localizing camouflaged objects from its background, has a long history in biology and art [10, 23]. Early works [18, 31, 36] in this field focus on distinguishing objects in their camouflaging environment by utilizing handcrafted features, such as texture, boundary, and intensity features. With the significant advances in deep learning (DL), DL-based meth" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "17919" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 86, + 73, + 509, + 298 + ], + "blocks": [ + { + "bbox": [ + 86, + 73, + 509, + 298 + ], + "lines": [ + { + "bbox": [ + 86, + 73, + 509, + 298 + ], + "spans": [ + { + "bbox": [ + 86, + 73, + 509, + 298 + ], + "type": "image", + "image_path": "0c152675a03c59e0981a88f2e0369956a2e5cf63050717d13b6c6bff5619036d.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 308, + 547, + 365 + ], + "lines": [ + { + "bbox": [ + 46, + 308, + 547, + 365 + ], + "spans": [ + { + "bbox": [ + 46, + 308, + 547, + 365 + ], + "type": "text", + "content": "Figure 2. Illustration of the proposed DCNet. Our framework mainly consists of two components: a pixel-level camouflage decoupling module and an instance-level camouflage suppression module. (1) The pixel-level camouflage decoupling module aims at extracting camouflaged characteristics and eliminating camouflaged characteristics while reserving target object characteristics to obtain de-camouflaged pixel features. (2) The instance-level camouflage suppression module is responsible for aggregating instance prototypes to achieve instance segmentation meanwhile suppressing the background noise in prototype-pixel interactions." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 385, + 289, + 589 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 385, + 289, + 589 + ], + "spans": [ + { + "bbox": [ + 46, + 385, + 289, + 589 + ], + "type": "text", + "content": "ods [16, 19, 29, 32, 43, 45, 46] have significantly improved the COD performance through end-to-end learning. Mei et al. [29] developed a distraction mining strategy to benefit the accurate segmentation of the camouflaged object. Pang et al. [32] proposed a mixed-scale triplet network, ZoomNet, to capture objects in complex scenes at different \"zoom\" scales. Zhong et al. [45] was the first to claim the COD task should go beyond the RGB domain and introduced frequency clues to better detect camouflaged objects. Despite the large progress in camouflaged object detection, camouflaged instance segmentation is a rarely studied task. In this paper, inspired by [34, 42, 45], we introduce frequency domain transform to model the camouflage characteristics and then achieve de-camouflaging for camouflaged instance segmentation. To the best of our knowledge, we are the first to explore explicit de-camouflaging for camouflaged instance segmentation." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 598, + 178, + 611 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 598, + 178, + 611 + ], + "spans": [ + { + "bbox": [ + 47, + 598, + 178, + 611 + ], + "type": "text", + "content": "2.2. Instance Segmentation" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 618, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 618, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 618, + 287, + 715 + ], + "type": "text", + "content": "Instance segmentation is a challenging task as it requires both pixel-level and instance-level mask prediction. Existing instance segmentation approaches can be broadly separated into two paradigms: two-stage methods [14, 17, 21] and one-stage methods [1, 40, 41]. Two-stage methods generally utilize object detectors to generate proposal regions and then segment the mask of each detected object. Based on Faster R-CNN [35], Mask R-CNN [14] generates region-" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 385, + 547, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 385, + 547, + 639 + ], + "spans": [ + { + "bbox": [ + 304, + 385, + 547, + 639 + ], + "type": "text", + "content": "of-interests (ROIs) in the first stage and utilizes an extra mask head to acquire the instance segmentation mask. The following works continue to improve the performance for this task in the two-stage manner. Recently, one-stage methods have emerged and driven a new trend in this field. YOLACT [1] combines the prototypes with the mask coefficients to produce instance masks without the detection step. SOLOv2 [41], evolved from SOLO [40], directly decouples the original mask prediction into kernel learning and feature learning to generate final instance segmentation results. Currently, the transformer-based methods [5, 6] introduce instance-specific prototypes to constantly interact with pixel features with the help of some attention mechanisms for final instance segmentation, achieving leading performance. However, the general instance segmentation methods can not directly apply to CIS, as the camouflaged objects present high similarity with the background. Therefore, we base on the one-stage transformer-based paradigm to design an end-to-end De-camouflaging Network, where we introduce reliable reference points to build accurate similarity measurement that is robust to deceptive backgrounds." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 655, + 386, + 667 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 655, + 386, + 667 + ], + "spans": [ + { + "bbox": [ + 306, + 655, + 386, + 667 + ], + "type": "text", + "content": "3. Our Method" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 677, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 677, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 677, + 547, + 713 + ], + "type": "text", + "content": "In this section, we first present the overall architecture of the proposed De-camouflage Network (DCNet), and then describe each module in detail." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "17920" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 167, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 167, + 83 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 167, + 83 + ], + "type": "text", + "content": "3.1. Overall Architecture" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 90, + 287, + 209 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 90, + 287, + 209 + ], + "spans": [ + { + "bbox": [ + 46, + 90, + 287, + 209 + ], + "type": "text", + "content": "Our DCNet mainly consists of two modules. (1) The pixel-level camouflage decoupling module aims at extracting camouflage characteristics based on the amplitude information of Fourier spectrum, and eliminating camouflage characteristics via a difference attention mechanism. (2) The instance-level camouflage suppression module is responsible for learning instance prototypes from interactions with pixel features, through a reference attention mechanism to suppress the erroneous correlations caused by deceptive background." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 217, + 233, + 230 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 217, + 233, + 230 + ], + "spans": [ + { + "bbox": [ + 47, + 217, + 233, + 230 + ], + "type": "text", + "content": "3.2. Pixel-level Camouflage Decoupling" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 236, + 287, + 307 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 236, + 287, + 307 + ], + "spans": [ + { + "bbox": [ + 46, + 236, + 287, + 307 + ], + "type": "text", + "content": "In order to decouple camouflage characteristics and target information fused in the pixel feature, we first model the representation of camouflage characteristics by exploring Fourier frequency domain. Then we design a difference attention mechanism to remove camouflage characteristics while maintaining target information." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 308, + 287, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 308, + 287, + 426 + ], + "spans": [ + { + "bbox": [ + 46, + 308, + 287, + 426 + ], + "type": "text", + "content": "Camouflage Characteristics Extraction. Generally, the camouflage characteristics are mainly comprised of color and texture information [10, 22]. On the other hand, in the frequency domain, it is known that the amplitude component of Fourier spectrum preserves low-level statistics information. Therefore, we can utilize the amplitude of Fourier spectrum to represent background information in the camouflaged image. Specifically, given a camouflaged image " + }, + { + "bbox": [ + 46, + 308, + 287, + 426 + ], + "type": "inline_equation", + "content": "x \\in \\mathbb{R}^{H \\times W \\times 3}" + }, + { + "bbox": [ + 46, + 308, + 287, + 426 + ], + "type": "text", + "content": ", its Fourier transformation " + }, + { + "bbox": [ + 46, + 308, + 287, + 426 + ], + "type": "inline_equation", + "content": "\\mathcal{F}(x)" + }, + { + "bbox": [ + 46, + 308, + 287, + 426 + ], + "type": "text", + "content": " is formulated as:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 80, + 434, + 287, + 467 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 80, + 434, + 287, + 467 + ], + "spans": [ + { + "bbox": [ + 80, + 434, + 287, + 467 + ], + "type": "interline_equation", + "content": "\\mathcal {F} (x) _ {u, v} = \\sum_ {i = 0} ^ {H - 1} \\sum_ {j = 0} ^ {W - 1} x _ {i, j} e ^ {- J 2 \\pi \\left(\\frac {i}{H} u + \\frac {j}{W} v\\right)}, \\tag {1}", + "image_path": "2dfcf816082ed1115686e379916db8e0ae30dd74124547d1269b06920dfa3480.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 475, + 287, + 512 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 475, + 287, + 512 + ], + "spans": [ + { + "bbox": [ + 47, + 475, + 287, + 512 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 475, + 287, + 512 + ], + "type": "inline_equation", + "content": "J" + }, + { + "bbox": [ + 47, + 475, + 287, + 512 + ], + "type": "text", + "content": " refers to the imaginary unit and each channel of the image is computed independently. The amplitude and phase components are then respectively expressed as:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 89, + 521, + 244, + 537 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 521, + 244, + 537 + ], + "spans": [ + { + "bbox": [ + 89, + 521, + 244, + 537 + ], + "type": "interline_equation", + "content": "\\mathcal {A} (x) _ {u, v} = \\left[ R ^ {2} (x) _ {u, v} + I ^ {2} (x) _ {u, v} \\right] ^ {1 / 2}", + "image_path": "8f4958b7cb07618cf38f19815ef952eea5fca91cd243a074bf8dfd6eb042c3b8.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 90, + 537, + 287, + 566 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 90, + 537, + 287, + 566 + ], + "spans": [ + { + "bbox": [ + 90, + 537, + 287, + 566 + ], + "type": "interline_equation", + "content": "\\mathcal {P} (x) _ {u, v} = \\arctan \\left[ \\frac {I (x) _ {u , v}}{R (x) _ {u , v}} \\right], \\tag {2}", + "image_path": "004afe61286f4eaf8c9fb316a35b10de50ddebd6f2c19aeced8629a1834ed9da.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 574, + 287, + 622 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 574, + 287, + 622 + ], + "spans": [ + { + "bbox": [ + 47, + 574, + 287, + 622 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 574, + 287, + 622 + ], + "type": "inline_equation", + "content": "R(x)" + }, + { + "bbox": [ + 47, + 574, + 287, + 622 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 574, + 287, + 622 + ], + "type": "inline_equation", + "content": "I(x)" + }, + { + "bbox": [ + 47, + 574, + 287, + 622 + ], + "type": "text", + "content": " represent the real and imaginary part of " + }, + { + "bbox": [ + 47, + 574, + 287, + 622 + ], + "type": "inline_equation", + "content": "\\mathcal{F}(x)" + }, + { + "bbox": [ + 47, + 574, + 287, + 622 + ], + "type": "text", + "content": ", respectively. To represent the camouflaged background, we fix the phase to a constant " + }, + { + "bbox": [ + 47, + 574, + 287, + 622 + ], + "type": "inline_equation", + "content": "b" + }, + { + "bbox": [ + 47, + 574, + 287, + 622 + ], + "type": "text", + "content": " (e.g., average value) and reconstruct the image with the amplitude information as" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 115, + 630, + 287, + 644 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 115, + 630, + 287, + 644 + ], + "spans": [ + { + "bbox": [ + 115, + 630, + 287, + 644 + ], + "type": "interline_equation", + "content": "\\tilde {x} = \\mathcal {F} ^ {- 1} [ \\mathcal {A} (x) _ {u, v} e ^ {- J b} ], \\tag {3}", + "image_path": "ed5893fc1d92b30e0613de435a3c6ff53cf4b8174e37d5b8a68cc9d4ed223214.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 653, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 653, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 653, + 287, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 653, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{F}^{-1}" + }, + { + "bbox": [ + 47, + 653, + 287, + 713 + ], + "type": "text", + "content": " indicates the inverse Fourier transformation. Then, we feed the reconstructed image " + }, + { + "bbox": [ + 47, + 653, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\tilde{x}" + }, + { + "bbox": [ + 47, + 653, + 287, + 713 + ], + "type": "text", + "content": " into a light-weight CNN (e.g., ResNet-18 [15]) with global average pooling layer to obtain the global camouflage characteristics " + }, + { + "bbox": [ + 47, + 653, + 287, + 713 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 47, + 653, + 287, + 713 + ], + "type": "text", + "content": " of the original image." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 72, + 545, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 216 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 216 + ], + "type": "text", + "content": "Difference Attention Mechanism. Given image features " + }, + { + "bbox": [ + 304, + 72, + 545, + 216 + ], + "type": "inline_equation", + "content": "\\mathbf{F} \\in \\mathbb{R}^{h \\times w \\times c}" + }, + { + "bbox": [ + 304, + 72, + 545, + 216 + ], + "type": "text", + "content": " extracted from a backbone network (e.g., ResNet-50 [15]), we remove the camouflage characteristics " + }, + { + "bbox": [ + 304, + 72, + 545, + 216 + ], + "type": "inline_equation", + "content": "\\pmb{g}" + }, + { + "bbox": [ + 304, + 72, + 545, + 216 + ], + "type": "text", + "content": " while retaining valuable target information by a novel difference attention mechanism. In specific, we first utilize two " + }, + { + "bbox": [ + 304, + 72, + 545, + 216 + ], + "type": "inline_equation", + "content": "1 \\times 1" + }, + { + "bbox": [ + 304, + 72, + 545, + 216 + ], + "type": "text", + "content": " convolution layers to map " + }, + { + "bbox": [ + 304, + 72, + 545, + 216 + ], + "type": "inline_equation", + "content": "\\mathbf{F}" + }, + { + "bbox": [ + 304, + 72, + 545, + 216 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 72, + 545, + 216 + ], + "type": "inline_equation", + "content": "\\pmb{g}" + }, + { + "bbox": [ + 304, + 72, + 545, + 216 + ], + "type": "text", + "content": " to the same dimension " + }, + { + "bbox": [ + 304, + 72, + 545, + 216 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 304, + 72, + 545, + 216 + ], + "type": "text", + "content": ", resulting in " + }, + { + "bbox": [ + 304, + 72, + 545, + 216 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{F}} \\in \\mathbb{R}^{h \\times w \\times C}" + }, + { + "bbox": [ + 304, + 72, + 545, + 216 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 72, + 545, + 216 + ], + "type": "inline_equation", + "content": "\\hat{\\pmb{g}} \\in \\mathbb{R}^C" + }, + { + "bbox": [ + 304, + 72, + 545, + 216 + ], + "type": "text", + "content": ", respectively. The discrepancy between image features and camouflage characteristics indicates the saliency information of the target objects, helping to highlight the target information. Thus, we integrate the difference on all channels for each pixel to obtain a difference map " + }, + { + "bbox": [ + 304, + 72, + 545, + 216 + ], + "type": "inline_equation", + "content": "\\mathbf{D} \\in \\mathbb{R}^{h \\times w}" + }, + { + "bbox": [ + 304, + 72, + 545, + 216 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 370, + 226, + 545, + 258 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 370, + 226, + 545, + 258 + ], + "spans": [ + { + "bbox": [ + 370, + 226, + 545, + 258 + ], + "type": "interline_equation", + "content": "\\mathbf {D} _ {i, j} = \\sum_ {k = 1} ^ {C} \\left(\\hat {\\mathbf {F}} _ {i, j, k} - \\hat {\\boldsymbol {g}} _ {k}\\right) ^ {2}, \\tag {4}", + "image_path": "56d83f8396010333253f39dbf8389c622a177c85d065ac915fd3ab3036978612.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 268, + 545, + 316 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 268, + 545, + 316 + ], + "spans": [ + { + "bbox": [ + 304, + 268, + 545, + 316 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 268, + 545, + 316 + ], + "type": "inline_equation", + "content": "i,j" + }, + { + "bbox": [ + 304, + 268, + 545, + 316 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 268, + 545, + 316 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 268, + 545, + 316 + ], + "type": "text", + "content": " are the index of height, width, and channel, respectively. By multiplying the original image features with the difference map " + }, + { + "bbox": [ + 304, + 268, + 545, + 316 + ], + "type": "inline_equation", + "content": "\\mathbf{D}" + }, + { + "bbox": [ + 304, + 268, + 545, + 316 + ], + "type": "text", + "content": ", we can get the de-camouflaged pixel features " + }, + { + "bbox": [ + 304, + 268, + 545, + 316 + ], + "type": "inline_equation", + "content": "\\mathbf{F}^d" + }, + { + "bbox": [ + 304, + 268, + 545, + 316 + ], + "type": "text", + "content": " as" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 397, + 326, + 545, + 339 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 397, + 326, + 545, + 339 + ], + "spans": [ + { + "bbox": [ + 397, + 326, + 545, + 339 + ], + "type": "interline_equation", + "content": "\\mathbf {F} ^ {d} = \\mathbf {F} \\circ \\mathbf {D}, \\tag {5}", + "image_path": "1c6d457e1c9bb88d1865c7911e16039f8e2f9b17825c3d58b2bd3d8ed89b6c36.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 350, + 545, + 398 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 350, + 545, + 398 + ], + "spans": [ + { + "bbox": [ + 304, + 350, + 545, + 398 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 350, + 545, + 398 + ], + "type": "inline_equation", + "content": "\\circ" + }, + { + "bbox": [ + 304, + 350, + 545, + 398 + ], + "type": "text", + "content": " refers to the Hadamard product and " + }, + { + "bbox": [ + 304, + 350, + 545, + 398 + ], + "type": "inline_equation", + "content": "\\mathbf{D}" + }, + { + "bbox": [ + 304, + 350, + 545, + 398 + ], + "type": "text", + "content": " is broadcasted to the same dimension as " + }, + { + "bbox": [ + 304, + 350, + 545, + 398 + ], + "type": "inline_equation", + "content": "\\mathbf{F}" + }, + { + "bbox": [ + 304, + 350, + 545, + 398 + ], + "type": "text", + "content": ". In addition, we follow [33] to use a deformable self-attention layer [47] to further incorporate context information from other pixels." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 399, + 545, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 399, + 545, + 517 + ], + "spans": [ + { + "bbox": [ + 304, + 399, + 545, + 517 + ], + "type": "text", + "content": "Fusion Layer. To acquire fine-grained target information for more accurate segmentation, we use multi-scale features from different stages of the backbone. Specifically, we first obtain the features from different stages of the backbone network, and then obtain the corresponding de-camouflaged features through the difference attention mechanism. Finally, we fuse the processed features with the FPN [24] network to generate the high-resolution pixel-level feature " + }, + { + "bbox": [ + 304, + 399, + 545, + 517 + ], + "type": "inline_equation", + "content": "\\mathbf{E}" + }, + { + "bbox": [ + 304, + 399, + 545, + 517 + ], + "type": "text", + "content": ", which is used for subsequent instance segmentation." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 306, + 528, + 511, + 540 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 528, + 511, + 540 + ], + "spans": [ + { + "bbox": [ + 306, + 528, + 511, + 540 + ], + "type": "text", + "content": "3.3. Instance-level Camouflage Suppression" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 546, + 545, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 546, + 545, + 617 + ], + "spans": [ + { + "bbox": [ + 304, + 546, + 545, + 617 + ], + "type": "text", + "content": "In order to capture different camouflaged instances, we learn a set of instance prototypes " + }, + { + "bbox": [ + 304, + 546, + 545, + 617 + ], + "type": "inline_equation", + "content": "\\mathbf{P} = \\{\\pmb{p}^i\\}_{i=1}^N" + }, + { + "bbox": [ + 304, + 546, + 545, + 617 + ], + "type": "text", + "content": ", and each " + }, + { + "bbox": [ + 304, + 546, + 545, + 617 + ], + "type": "inline_equation", + "content": "\\pmb{p}^i \\in \\mathbb{R}^{1 \\times L}" + }, + { + "bbox": [ + 304, + 546, + 545, + 617 + ], + "type": "text", + "content": " is responsible for identifying whether each pixel belongs to a camouflaged instance. To eliminate the background interference for prototype learning, we design reference attention to aggregate target information." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 618, + 545, + 692 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 618, + 545, + 692 + ], + "spans": [ + { + "bbox": [ + 304, + 618, + 545, + 692 + ], + "type": "text", + "content": "Reference Attention Mechanism. Given the decamouflaged pixel feature " + }, + { + "bbox": [ + 304, + 618, + 545, + 692 + ], + "type": "inline_equation", + "content": "\\mathbf{F}^d" + }, + { + "bbox": [ + 304, + 618, + 545, + 692 + ], + "type": "text", + "content": " derived from the pixel-level camouflage decoupling module, we first obtain the queries " + }, + { + "bbox": [ + 304, + 618, + 545, + 692 + ], + "type": "inline_equation", + "content": "\\mathbf{Q} = [q_1;q_2;\\dots;q_N]" + }, + { + "bbox": [ + 304, + 618, + 545, + 692 + ], + "type": "text", + "content": " from instance prototypes " + }, + { + "bbox": [ + 304, + 618, + 545, + 692 + ], + "type": "inline_equation", + "content": "\\mathbf{P}" + }, + { + "bbox": [ + 304, + 618, + 545, + 692 + ], + "type": "text", + "content": ", keys " + }, + { + "bbox": [ + 304, + 618, + 545, + 692 + ], + "type": "inline_equation", + "content": "\\mathbf{K} = [k_{1};k_{2};\\dots;k_{hw}]" + }, + { + "bbox": [ + 304, + 618, + 545, + 692 + ], + "type": "text", + "content": " and values " + }, + { + "bbox": [ + 304, + 618, + 545, + 692 + ], + "type": "inline_equation", + "content": "\\mathbf{V} = [v_{1};v_{2};\\dots;v_{hw}]" + }, + { + "bbox": [ + 304, + 618, + 545, + 692 + ], + "type": "text", + "content": " from pixel features " + }, + { + "bbox": [ + 304, + 618, + 545, + 692 + ], + "type": "inline_equation", + "content": "\\mathbf{F}^d = [f_1;f_2;\\dots;f_{hw}]" + }, + { + "bbox": [ + 304, + 618, + 545, + 692 + ], + "type": "text", + "content": " as" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 342, + 700, + 545, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 342, + 700, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 342, + 700, + 545, + 715 + ], + "type": "interline_equation", + "content": "\\boldsymbol {q} _ {i} = \\boldsymbol {p} _ {i} \\mathbf {W} ^ {q}, \\boldsymbol {k} _ {j} = \\boldsymbol {f} _ {j} \\mathbf {W} ^ {k}, \\boldsymbol {v} _ {j} = \\boldsymbol {f} _ {j} \\mathbf {W} ^ {v}, \\tag {6}", + "image_path": "41583c3d291a3c0b50b990f60531216735ddf4391e591a04432e0bbbc4807348.jpg" + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17921" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 287, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 287, + 120 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 287, + 120 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 72, + 287, + 120 + ], + "type": "inline_equation", + "content": "i \\in 1,2,\\dots,N, j \\in 1,2,\\dots,hw" + }, + { + "bbox": [ + 47, + 72, + 287, + 120 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 47, + 72, + 287, + 120 + ], + "type": "inline_equation", + "content": "\\mathbf{W}^q \\in \\mathbb{R}^{L \\times L}" + }, + { + "bbox": [ + 47, + 72, + 287, + 120 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 47, + 72, + 287, + 120 + ], + "type": "inline_equation", + "content": "\\mathbf{W}^k" + }, + { + "bbox": [ + 47, + 72, + 287, + 120 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 47, + 72, + 287, + 120 + ], + "type": "inline_equation", + "content": "\\mathbf{W}^v \\in \\mathbb{R}^{c \\times L}" + }, + { + "bbox": [ + 47, + 72, + 287, + 120 + ], + "type": "text", + "content": " are linear projections. With the prototype queries, we can compute the correlation between each query-key pair independently as" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 129, + 125, + 287, + 151 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 129, + 125, + 287, + 151 + ], + "spans": [ + { + "bbox": [ + 129, + 125, + 287, + 151 + ], + "type": "interline_equation", + "content": "s _ {i, j} = \\frac {d \\left(\\boldsymbol {q} _ {i} , \\boldsymbol {k} _ {j}\\right)}{\\sqrt {L}}, \\tag {7}", + "image_path": "2fb64a5e7f4785c6e4a7e036f114af9eb81ec6fb8188dd2b244e4c2ba4e3819a.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 152, + 287, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 152, + 287, + 319 + ], + "spans": [ + { + "bbox": [ + 46, + 152, + 287, + 319 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 152, + 287, + 319 + ], + "type": "inline_equation", + "content": "d(\\cdot, \\cdot)" + }, + { + "bbox": [ + 46, + 152, + 287, + 319 + ], + "type": "text", + "content": " denotes the distance metric which is the dot product similarity in [39], and " + }, + { + "bbox": [ + 46, + 152, + 287, + 319 + ], + "type": "inline_equation", + "content": "\\sqrt{L}" + }, + { + "bbox": [ + 46, + 152, + 287, + 319 + ], + "type": "text", + "content": " is a scaling factor. As the camouflaged object presents high similarity with the surrounding backgrounds, the prototype-pixel correlation represented by " + }, + { + "bbox": [ + 46, + 152, + 287, + 319 + ], + "type": "inline_equation", + "content": "s_{i,j}" + }, + { + "bbox": [ + 46, + 152, + 287, + 319 + ], + "type": "text", + "content": " is susceptible to background pixel interference, resulting in erroneous segmentation. Intuitively, humans are able to identify camouflaged instances by repeatedly comparing the target with the surrounding reference areas. Motivated by this, we first use an adaptive selection process to select some reliable pixels from de-camouflaged pixels " + }, + { + "bbox": [ + 46, + 152, + 287, + 319 + ], + "type": "inline_equation", + "content": "\\mathbf{F}^d" + }, + { + "bbox": [ + 46, + 152, + 287, + 319 + ], + "type": "text", + "content": " as references. In specific, to measure the reliability of each pixel, we define and calculate the total contribution of each pixel with regard to all the prototypes, which is formulated as:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 104, + 323, + 287, + 354 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 104, + 323, + 287, + 354 + ], + "spans": [ + { + "bbox": [ + 104, + 323, + 287, + 354 + ], + "type": "interline_equation", + "content": "u _ {j} = \\sum_ {i = 1} ^ {N} s _ {i, j}, j \\in 1, 2, \\dots , h w. \\tag {8}", + "image_path": "4a1bb9ff35eefb905bc31acbe0280273d53c3010c8534586ee8c5c1b638b1ea2.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 355, + 287, + 427 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 355, + 287, + 427 + ], + "spans": [ + { + "bbox": [ + 47, + 355, + 287, + 427 + ], + "type": "text", + "content": "Then we select top-" + }, + { + "bbox": [ + 47, + 355, + 287, + 427 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 47, + 355, + 287, + 427 + ], + "type": "text", + "content": " pixels with the largest correlations with the prototypes as reference points " + }, + { + "bbox": [ + 47, + 355, + 287, + 427 + ], + "type": "inline_equation", + "content": "\\mathbf{R} \\in \\mathbb{R}^{K \\times L}" + }, + { + "bbox": [ + 47, + 355, + 287, + 427 + ], + "type": "text", + "content": ". In this way, we can pick out reliable reference points that are adaptive to the instance prototypes. Based on reference points, we first calculate the prototype-reference and pixel-reference similarity as" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 95, + 432, + 287, + 449 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 95, + 432, + 287, + 449 + ], + "spans": [ + { + "bbox": [ + 95, + 432, + 287, + 449 + ], + "type": "interline_equation", + "content": "s _ {i} ^ {q} = \\phi_ {1} \\left(\\boldsymbol {q} _ {i} \\mathbf {R} ^ {\\top}\\right), s _ {j} ^ {k} = \\phi_ {2} \\left(\\boldsymbol {k} _ {j} \\mathbf {R} ^ {\\top}\\right), \\tag {9}", + "image_path": "8f162dac3a8ac17e830936b24448898b98b083f51753b01ce2b331d95991f65e.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 453, + 287, + 501 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 453, + 287, + 501 + ], + "spans": [ + { + "bbox": [ + 47, + 453, + 287, + 501 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 453, + 287, + 501 + ], + "type": "inline_equation", + "content": "\\phi_{1}" + }, + { + "bbox": [ + 47, + 453, + 287, + 501 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 453, + 287, + 501 + ], + "type": "inline_equation", + "content": "\\phi_{2}" + }, + { + "bbox": [ + 47, + 453, + 287, + 501 + ], + "type": "text", + "content": " are feed-forward networks (FFN) containing two fully connected layers. " + }, + { + "bbox": [ + 47, + 453, + 287, + 501 + ], + "type": "inline_equation", + "content": "s_i^q" + }, + { + "bbox": [ + 47, + 453, + 287, + 501 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 453, + 287, + 501 + ], + "type": "inline_equation", + "content": "s_j^k" + }, + { + "bbox": [ + 47, + 453, + 287, + 501 + ], + "type": "text", + "content": " can serve as soft multilabel to measure the correlation between the " + }, + { + "bbox": [ + 47, + 453, + 287, + 501 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 47, + 453, + 287, + 501 + ], + "type": "text", + "content": "-th prototype and the " + }, + { + "bbox": [ + 47, + 453, + 287, + 501 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 47, + 453, + 287, + 501 + ], + "type": "text", + "content": "-th pixel:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 102, + 506, + 287, + 522 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 102, + 506, + 287, + 522 + ], + "spans": [ + { + "bbox": [ + 102, + 506, + 287, + 522 + ], + "type": "interline_equation", + "content": "s _ {i, j} ^ {q k} = d \\left(\\boldsymbol {q} _ {i}, \\boldsymbol {k} _ {j}; \\mathbf {R}\\right) = s _ {i} ^ {q} \\left(s _ {j} ^ {k}\\right) ^ {\\top}. \\tag {10}", + "image_path": "fa5fdad7704ea2d693f2a596c7d7c707f7377ad561638d1d737afc22ef96ffba.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 529, + 287, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 529, + 287, + 590 + ], + "spans": [ + { + "bbox": [ + 47, + 529, + 287, + 590 + ], + "type": "text", + "content": "Compared to " + }, + { + "bbox": [ + 47, + 529, + 287, + 590 + ], + "type": "inline_equation", + "content": "s_{i,j}" + }, + { + "bbox": [ + 47, + 529, + 287, + 590 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 47, + 529, + 287, + 590 + ], + "type": "inline_equation", + "content": "s_{i,j}^{qk}" + }, + { + "bbox": [ + 47, + 529, + 287, + 590 + ], + "type": "text", + "content": " leverages the consensus among reliable reference points, thus suppressing the background noises and leading to fewer erroneous correlations. Given the correlation " + }, + { + "bbox": [ + 47, + 529, + 287, + 590 + ], + "type": "inline_equation", + "content": "s_{i,j}^{qk}" + }, + { + "bbox": [ + 47, + 529, + 287, + 590 + ], + "type": "text", + "content": ", we can extract and purify target information from pixel features to update instance prototypes as" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 82, + 595, + 287, + 629 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 82, + 595, + 287, + 629 + ], + "spans": [ + { + "bbox": [ + 82, + 595, + 287, + 629 + ], + "type": "interline_equation", + "content": "\\tilde {\\boldsymbol {p}} _ {i} = \\sum_ {j = 1} ^ {h w} a _ {i, j} \\boldsymbol {v} _ {j}, a _ {i, j} = \\frac {\\exp \\left(s _ {i , j} ^ {q k}\\right)}{\\sum_ {j = 1} ^ {h w} \\exp \\left(s _ {i , j} ^ {q k}\\right)}. \\tag {11}", + "image_path": "d8d700c36875a2fddcbde976cc1f026137348a9d2dbc4a28fcacd3c931d52256.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 635, + 286, + 658 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 635, + 286, + 658 + ], + "spans": [ + { + "bbox": [ + 47, + 635, + 286, + 658 + ], + "type": "text", + "content": "Note that we have omitted the scaling factor and the multihead mechanism for notation simplicity." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 658, + 287, + 694 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 658, + 287, + 694 + ], + "spans": [ + { + "bbox": [ + 47, + 658, + 287, + 694 + ], + "type": "text", + "content": "Given the high-resolution pixel-level feature embedding " + }, + { + "bbox": [ + 47, + 658, + 287, + 694 + ], + "type": "inline_equation", + "content": "\\mathbf{E}" + }, + { + "bbox": [ + 47, + 658, + 287, + 694 + ], + "type": "text", + "content": " and the learned instance prototypes " + }, + { + "bbox": [ + 47, + 658, + 287, + 694 + ], + "type": "inline_equation", + "content": "\\tilde{\\pmb{p}}_i" + }, + { + "bbox": [ + 47, + 658, + 287, + 694 + ], + "type": "text", + "content": ", we can obtain N mask predictions and according scores by" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 88, + 700, + 287, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 88, + 700, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 88, + 700, + 287, + 715 + ], + "type": "interline_equation", + "content": "m _ {i} = \\operatorname {s i g m o i d} \\left(\\mathbf {E} \\tilde {\\boldsymbol {p}} _ {i} ^ {\\top}\\right), y _ {i} = h \\left(\\tilde {\\boldsymbol {p}} _ {i} \\mid \\sigma\\right), \\tag {12}", + "image_path": "e09313b0561364803a8fa4e9a9d539d7d015cb365410139bbf907e4779916b62.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 72, + 545, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 545, + 156 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 545, + 156 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 72, + 545, + 156 + ], + "type": "inline_equation", + "content": "h(\\cdot \\mid \\sigma)" + }, + { + "bbox": [ + 305, + 72, + 545, + 156 + ], + "type": "text", + "content": " is a classifier parameterized by " + }, + { + "bbox": [ + 305, + 72, + 545, + 156 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 305, + 72, + 545, + 156 + ], + "type": "text", + "content": " to predict the confidence score. Following [6], we match a ground truth label (including \"no object\") for each predicted segment by the Hungarian matching algorithm. For training, we use a cross-entropy loss to constrain the instance score " + }, + { + "bbox": [ + 305, + 72, + 545, + 156 + ], + "type": "inline_equation", + "content": "y_{i}" + }, + { + "bbox": [ + 305, + 72, + 545, + 156 + ], + "type": "text", + "content": ", and a linear combination of a focal loss [25] and a dice loss [30] to constrain the mask prediction " + }, + { + "bbox": [ + 305, + 72, + 545, + 156 + ], + "type": "inline_equation", + "content": "m_{i}" + }, + { + "bbox": [ + 305, + 72, + 545, + 156 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 168, + 382, + 181 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 168, + 382, + 181 + ], + "spans": [ + { + "bbox": [ + 306, + 168, + 382, + 181 + ], + "type": "text", + "content": "4. Experiment" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 188, + 413, + 201 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 188, + 413, + 201 + ], + "spans": [ + { + "bbox": [ + 306, + 188, + 413, + 201 + ], + "type": "text", + "content": "4.1. Experiment Setup" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 305, + 206, + 545, + 314 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 206, + 545, + 314 + ], + "spans": [ + { + "bbox": [ + 305, + 206, + 545, + 314 + ], + "type": "text", + "content": "Datasets. We evaluate our method on two datasets: COD10K [10] and NC4K [28]. COD10K contains 5086 camouflaged images with high-quality instance-level annotations, which are divided into 3040 training images and 2026 testing images. NC4K contains 4121 test camouflaged images to evaluate the generalization ability of CIS models. Following the standard benchmark [33], we use the instance-level annotations in COD10K to train our DCNet and evaluate it on the test set of COD10K and NC4K." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 305, + 315, + 545, + 385 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 315, + 545, + 385 + ], + "spans": [ + { + "bbox": [ + 305, + 315, + 545, + 385 + ], + "type": "text", + "content": "Evaluation Metrics. We adopt " + }, + { + "bbox": [ + 305, + 315, + 545, + 385 + ], + "type": "inline_equation", + "content": "\\mathrm{AP}_{50}" + }, + { + "bbox": [ + 305, + 315, + 545, + 385 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 305, + 315, + 545, + 385 + ], + "type": "inline_equation", + "content": "\\mathrm{AP}_{75}" + }, + { + "bbox": [ + 305, + 315, + 545, + 385 + ], + "type": "text", + "content": " and AP scores as evaluation metrics to quantify the effectiveness of our method. A true positive is counted if the intersection over union (IoU) between the ground truth and the segmentation is at least " + }, + { + "bbox": [ + 305, + 315, + 545, + 385 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 305, + 315, + 545, + 385 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 305, + 315, + 545, + 385 + ], + "type": "inline_equation", + "content": "75\\%" + }, + { + "bbox": [ + 305, + 315, + 545, + 385 + ], + "type": "text", + "content": ", and the AP score is an overall metric that combines different IoU thresholds." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 305, + 387, + 545, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 387, + 545, + 506 + ], + "spans": [ + { + "bbox": [ + 305, + 387, + 545, + 506 + ], + "type": "text", + "content": "Implementation Details. For fair comparisons, the feature extractor is implemented by ResNet-50 [15] which is pretrained on ImageNet [8]. During our training, our model is trained with batch size of 2, using the Adam optimizer [27] with an initial learning rate of 0.0001 for 90,000 iterations. We set the channel dimension " + }, + { + "bbox": [ + 305, + 387, + 545, + 506 + ], + "type": "inline_equation", + "content": "C = L = 256" + }, + { + "bbox": [ + 305, + 387, + 545, + 506 + ], + "type": "text", + "content": ". In the instance-level camouflage suppression module, we set the number of prototypes " + }, + { + "bbox": [ + 305, + 387, + 545, + 506 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 305, + 387, + 545, + 506 + ], + "type": "text", + "content": " as 10, and the number of reference points " + }, + { + "bbox": [ + 305, + 387, + 545, + 506 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 305, + 387, + 545, + 506 + ], + "type": "text", + "content": " as 64, which turns out to generate the best segmentation during our experiments as shown in Figure 4." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 306, + 515, + 489, + 528 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 515, + 489, + 528 + ], + "spans": [ + { + "bbox": [ + 306, + 515, + 489, + 528 + ], + "type": "text", + "content": "4.2. Comparison with State-of-the-arts" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 305, + 533, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 533, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 533, + 545, + 713 + ], + "type": "text", + "content": "Quantitative results. Considering that CIS is a newly emerging task, we also compare some general instance segmentation methods besides the previous CIS methods and all the methods adopt ResNet-50 [15] backbone for fair comparisons. As demonstrated in Table 1, the proposed DCNet consistently outperforms the state-of-the-art methods by a large margin on both datasets. (1) On the COD10K [10], our method outperforms the previous best method OSFormer [33] by " + }, + { + "bbox": [ + 305, + 533, + 545, + 713 + ], + "type": "inline_equation", + "content": "4.3\\%" + }, + { + "bbox": [ + 305, + 533, + 545, + 713 + ], + "type": "text", + "content": " in AP, indicating that our DCNet is able to acquire more accurate segmentation of camouflaged targets. (2) On the NC4K [28], Our model yields an accuracy of " + }, + { + "bbox": [ + 305, + 533, + 545, + 713 + ], + "type": "inline_equation", + "content": "52.8\\%" + }, + { + "bbox": [ + 305, + 533, + 545, + 713 + ], + "type": "text", + "content": " in AP, making an obvious performance improvement by " + }, + { + "bbox": [ + 305, + 533, + 545, + 713 + ], + "type": "inline_equation", + "content": "7.0\\%" + }, + { + "bbox": [ + 305, + 533, + 545, + 713 + ], + "type": "text", + "content": " in AP compared to the second best performing method Mask2Former [5]. Since the model is trained on the COD10K training set, the high" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17922" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 61, + 102, + 526, + 316 + ], + "blocks": [ + { + "bbox": [ + 47, + 70, + 545, + 94 + ], + "lines": [ + { + "bbox": [ + 47, + 70, + 545, + 94 + ], + "spans": [ + { + "bbox": [ + 47, + 70, + 545, + 94 + ], + "type": "text", + "content": "Table 1. Comparisons of existing CIS methods and general instance approaches on COD10K [10] and NC4K [28] testing set, where all the methods use ResNet-50 as feature extractor. The best results are shown in bold." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 61, + 102, + 526, + 316 + ], + "lines": [ + { + "bbox": [ + 61, + 102, + 526, + 316 + ], + "spans": [ + { + "bbox": [ + 61, + 102, + 526, + 316 + ], + "type": "table", + "html": "
MethodsCOD10K-TestNC4K-TestParams(M)GFLOPs
AP\\( AP_{50} \\)\\( AP_{75} \\)AP\\( AP_{50} \\)\\( AP_{75} \\)
Mask R-CNN [14]25.055.520.427.758.622.743.9186.3
MS R-CNN [17]30.157.228.731.058.729.460.0198.5
Cascade R-CNN [2]25.356.121.329.560.824.871.7334.1
HTC [4]28.156.325.129.859.026.676.9331.7
BlendMask [3]28.256.425.227.756.724.235.8233.8
Mask Transfiner [20]28.756.326.429.456.727.244.3185.1
YOLACT [1]24.353.319.732.165.327.9--
CondInst [38]30.663.626.133.467.429.434.1200.1
QueryInst [12]28.560.123.133.066.729.4--
SOTR [13]27.958.724.129.361.025.663.1476.7
SOLOv2 [41]32.563.229.934.465.931.946.2318.7
MaskFormer [6]38.265.137.944.671.945.845.0174.2
Mask2Former [5]39.467.738.545.873.647.543.9241.0
OSFormer [33]41.071.140.842.572.542.346.6324.7
DCNet (ours)45.370.747.552.877.156.553.4207.0
", + "image_path": "619d16cc497f147a6af5273e487e1afbee725ad5410c2ac7af14b32b123df0da.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 50, + 365, + 285, + 427 + ], + "blocks": [ + { + "bbox": [ + 47, + 334, + 287, + 357 + ], + "lines": [ + { + "bbox": [ + 47, + 334, + 287, + 357 + ], + "spans": [ + { + "bbox": [ + 47, + 334, + 287, + 357 + ], + "type": "text", + "content": "Table 2. Comparison of the different components in pixel-level camouflage decoupling module in terms of AP scores." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 50, + 365, + 285, + 427 + ], + "lines": [ + { + "bbox": [ + 50, + 365, + 285, + 427 + ], + "spans": [ + { + "bbox": [ + 50, + 365, + 285, + 427 + ], + "type": "table", + "html": "
COD10KNC4K
PCD (ours)45.352.8
-camouflage attribute extractor43.1(-2.2)48.1(-4.7)
-difference attention mechanism44.1(-1.2)51.3(-1.5)
-both 2 components above40.2(-5.1)46.4(-6.4)
", + "image_path": "2eedba9b1a6066d056e2d8dfc90eb51b253ac1a95cc6dd10bf908de87eed8ab5.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 449, + 287, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 449, + 287, + 653 + ], + "spans": [ + { + "bbox": [ + 46, + 449, + 287, + 653 + ], + "type": "text", + "content": "performance on NC4K indicates that our method also has better generalization ability. Notice that our method does not achieve the highest performance on " + }, + { + "bbox": [ + 46, + 449, + 287, + 653 + ], + "type": "inline_equation", + "content": "\\mathrm{AP}_{50}" + }, + { + "bbox": [ + 46, + 449, + 287, + 653 + ], + "type": "text", + "content": " metric for the NC4K because different from the previous CIS method, we do not use additional Non-maximum suppression (NMS) post-processing to remove redundant predictions. For " + }, + { + "bbox": [ + 46, + 449, + 287, + 653 + ], + "type": "inline_equation", + "content": "\\mathrm{AP}_{50}" + }, + { + "bbox": [ + 46, + 449, + 287, + 653 + ], + "type": "text", + "content": ", our more accurate segmentation results cannot increase the number of true positive (TP) under the " + }, + { + "bbox": [ + 46, + 449, + 287, + 653 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 46, + 449, + 287, + 653 + ], + "type": "text", + "content": " IoU threshold, while redundant predictions will increase the number of false positive (FP), thus our method has no obvious advantages. For " + }, + { + "bbox": [ + 46, + 449, + 287, + 653 + ], + "type": "inline_equation", + "content": "\\mathrm{AP}_{75}" + }, + { + "bbox": [ + 46, + 449, + 287, + 653 + ], + "type": "text", + "content": ", our high precision segmentation will increase TP with a large margin. The overall metric AP reflects the superiority of our method over other methods. There is a similar trend on NC4K. In addition, the lower GFLOPs of our DCNet compared to state-of-the-art methods indicates that our model achieves high segmentation ability while maintaining high computation efficiency." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 653, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 287, + 715 + ], + "type": "text", + "content": "Qualitative Results. As shown in Figure 3, our proposed method is capable of de-camouflaging by using the difference attention mechanism at the pixel level, and also able to identify the accurate locations of target camouflaged objects at the instance level. Compared to previous methods," + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 325, + 376, + 525, + 425 + ], + "blocks": [ + { + "bbox": [ + 305, + 334, + 545, + 366 + ], + "lines": [ + { + "bbox": [ + 305, + 334, + 545, + 366 + ], + "spans": [ + { + "bbox": [ + 305, + 334, + 545, + 366 + ], + "type": "text", + "content": "Table 3. Comparisons of different attention mechanisms in the instance-level camouflage suppression module in terms of AP scores." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 325, + 376, + 525, + 425 + ], + "lines": [ + { + "bbox": [ + 325, + 376, + 525, + 425 + ], + "spans": [ + { + "bbox": [ + 325, + 376, + 525, + 425 + ], + "type": "table", + "html": "
Attention MechanismCOD10KNC4K
cross-attention [39]42.449.5
masked attention [5]44.751.7
reference attention (ours)45.352.8
", + "image_path": "95be9cc39a0faaaa1bdebaad747ec42d68cb75eb2ffe1646e3cff71239856468.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 446, + 545, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 446, + 545, + 495 + ], + "spans": [ + { + "bbox": [ + 305, + 446, + 545, + 495 + ], + "type": "text", + "content": "our DCNet performs better at boundaries of camouflage instances (see the first 4 columns), suppresses distracting background regions (see " + }, + { + "bbox": [ + 305, + 446, + 545, + 495 + ], + "type": "inline_equation", + "content": "5^{th}" + }, + { + "bbox": [ + 305, + 446, + 545, + 495 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 305, + 446, + 545, + 495 + ], + "type": "inline_equation", + "content": "6^{th}" + }, + { + "bbox": [ + 305, + 446, + 545, + 495 + ], + "type": "text", + "content": " columns), and distinguishes multiple instances well (see the last 2 columns)." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 502, + 400, + 516 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 502, + 400, + 516 + ], + "spans": [ + { + "bbox": [ + 306, + 502, + 400, + 516 + ], + "type": "text", + "content": "4.3. Ablation Study" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 521, + 545, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 521, + 545, + 544 + ], + "spans": [ + { + "bbox": [ + 305, + 521, + 545, + 544 + ], + "type": "text", + "content": "We conduct comprehensive ablation studies on COD10K and NC4K to verify the effectiveness of our modules." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 545, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 545, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 545, + 545, + 713 + ], + "type": "text", + "content": "Effectiveness of the Pixel-level Camouflage Decoupling Module. The pixel-level camouflage decoupling module mainly consists of a camouflage attribute extractor and a difference attention mechanism. We validate the importance of each component by removing them one at a time. As shown in Table 2, we replace the proposed camouflage attribute extractor with a learnable vector, and the performance drops by an average of " + }, + { + "bbox": [ + 304, + 545, + 545, + 713 + ], + "type": "inline_equation", + "content": "3.5\\%" + }, + { + "bbox": [ + 304, + 545, + 545, + 713 + ], + "type": "text", + "content": " across two datasets. It indicates that the Fourier spectrum amplitude is effective in modeling the camouflage characteristics, verifying that it is hard to learn de-camouflage without image-specific camouflage information supervision. Then we replace the difference attention mechanism with general subtraction operations. The reduced performance indicates that our pro" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17923" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 141, + 77, + 184, + 111 + ], + "blocks": [ + { + "bbox": [ + 75, + 89, + 132, + 100 + ], + "lines": [ + { + "bbox": [ + 75, + 89, + 132, + 100 + ], + "spans": [ + { + "bbox": [ + 75, + 89, + 132, + 100 + ], + "type": "text", + "content": "Original image" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 141, + 77, + 184, + 111 + ], + "lines": [ + { + "bbox": [ + 141, + 77, + 184, + 111 + ], + "spans": [ + { + "bbox": [ + 141, + 77, + 184, + 111 + ], + "type": "image", + "image_path": "f53f37466a24f47b59e089a7fb594e9755d4c6aec6c3a52f2428c7992f4dcccb.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 188, + 77, + 231, + 111 + ], + "blocks": [ + { + "bbox": [ + 188, + 77, + 231, + 111 + ], + "lines": [ + { + "bbox": [ + 188, + 77, + 231, + 111 + ], + "spans": [ + { + "bbox": [ + 188, + 77, + 231, + 111 + ], + "type": "image", + "image_path": "cc2b5806073f5e7dec46812d7dc1d5de022ed9090ec03822906887b55f861ae3.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 236, + 77, + 279, + 110 + ], + "blocks": [ + { + "bbox": [ + 236, + 77, + 279, + 110 + ], + "lines": [ + { + "bbox": [ + 236, + 77, + 279, + 110 + ], + "spans": [ + { + "bbox": [ + 236, + 77, + 279, + 110 + ], + "type": "image", + "image_path": "aa62ac7220e556d24e0a874ae62fc6f1a708726f4bb978b10eac62cf9c85b379.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 283, + 77, + 326, + 110 + ], + "blocks": [ + { + "bbox": [ + 283, + 77, + 326, + 110 + ], + "lines": [ + { + "bbox": [ + 283, + 77, + 326, + 110 + ], + "spans": [ + { + "bbox": [ + 283, + 77, + 326, + 110 + ], + "type": "image", + "image_path": "988603c351cbf2895a772229e45782d3b2997c6323d9f64448363f36ab3e4ad5.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 331, + 77, + 374, + 110 + ], + "blocks": [ + { + "bbox": [ + 331, + 77, + 374, + 110 + ], + "lines": [ + { + "bbox": [ + 331, + 77, + 374, + 110 + ], + "spans": [ + { + "bbox": [ + 331, + 77, + 374, + 110 + ], + "type": "image", + "image_path": "d38ae4f9e9c4b6e1ffe9ae145f2d9f04165f93cfd477ca225262da6b0f0d379c.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 378, + 77, + 422, + 110 + ], + "blocks": [ + { + "bbox": [ + 378, + 77, + 422, + 110 + ], + "lines": [ + { + "bbox": [ + 378, + 77, + 422, + 110 + ], + "spans": [ + { + "bbox": [ + 378, + 77, + 422, + 110 + ], + "type": "image", + "image_path": "fa0480c307b9c7548b0b0e97d22923df309fefb571aadf4a4a45668d7b288a98.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 426, + 77, + 469, + 110 + ], + "blocks": [ + { + "bbox": [ + 426, + 77, + 469, + 110 + ], + "lines": [ + { + "bbox": [ + 426, + 77, + 469, + 110 + ], + "spans": [ + { + "bbox": [ + 426, + 77, + 469, + 110 + ], + "type": "image", + "image_path": "45e9b494e3843707d4d4d85426bc4c2dfca555fb2392b3efc7505ae9ba17a454.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 474, + 77, + 517, + 110 + ], + "blocks": [ + { + "bbox": [ + 474, + 77, + 517, + 110 + ], + "lines": [ + { + "bbox": [ + 474, + 77, + 517, + 110 + ], + "spans": [ + { + "bbox": [ + 474, + 77, + 517, + 110 + ], + "type": "image", + "image_path": "384068335ec1bc22ac0215ca0d2e06e9942bd225a9ec630b89d1b0b4d4fe6c33.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 141, + 115, + 184, + 150 + ], + "blocks": [ + { + "bbox": [ + 99, + 128, + 112, + 137 + ], + "lines": [ + { + "bbox": [ + 99, + 128, + 112, + 137 + ], + "spans": [ + { + "bbox": [ + 99, + 128, + 112, + 137 + ], + "type": "text", + "content": "GT" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 141, + 115, + 184, + 150 + ], + "lines": [ + { + "bbox": [ + 141, + 115, + 184, + 150 + ], + "spans": [ + { + "bbox": [ + 141, + 115, + 184, + 150 + ], + "type": "image", + "image_path": "19776a8e98bdf84828fd51a0add9ec12917c7da9eeba01121d43ba1526b3256a.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 189, + 116, + 231, + 150 + ], + "blocks": [ + { + "bbox": [ + 189, + 116, + 231, + 150 + ], + "lines": [ + { + "bbox": [ + 189, + 116, + 231, + 150 + ], + "spans": [ + { + "bbox": [ + 189, + 116, + 231, + 150 + ], + "type": "image", + "image_path": "4a61791df68fd000b715567e9829a21b7ee5ed297d7467ae1f7e153677956481.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 236, + 116, + 279, + 150 + ], + "blocks": [ + { + "bbox": [ + 236, + 116, + 279, + 150 + ], + "lines": [ + { + "bbox": [ + 236, + 116, + 279, + 150 + ], + "spans": [ + { + "bbox": [ + 236, + 116, + 279, + 150 + ], + "type": "image", + "image_path": "7ab5602b2131e8c9e153fee6a1579ecc7b6b920d98493356693995c2b302b92b.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 284, + 116, + 326, + 150 + ], + "blocks": [ + { + "bbox": [ + 284, + 116, + 326, + 150 + ], + "lines": [ + { + "bbox": [ + 284, + 116, + 326, + 150 + ], + "spans": [ + { + "bbox": [ + 284, + 116, + 326, + 150 + ], + "type": "image", + "image_path": "e564528de3fad60a33fe9de07f591c906022f58b251ca2faa60ae4744cf51bae.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 331, + 116, + 374, + 150 + ], + "blocks": [ + { + "bbox": [ + 331, + 116, + 374, + 150 + ], + "lines": [ + { + "bbox": [ + 331, + 116, + 374, + 150 + ], + "spans": [ + { + "bbox": [ + 331, + 116, + 374, + 150 + ], + "type": "image", + "image_path": "d2e2c9545c62bcedbbfe0eb929a8e720ddb9db2ee910fcf9107e1e198d10b60e.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 378, + 116, + 422, + 150 + ], + "blocks": [ + { + "bbox": [ + 378, + 116, + 422, + 150 + ], + "lines": [ + { + "bbox": [ + 378, + 116, + 422, + 150 + ], + "spans": [ + { + "bbox": [ + 378, + 116, + 422, + 150 + ], + "type": "image", + "image_path": "50e4dcb024d1f85e9fdd67e5eb3b577b9dc1feb1e9f7c53ca7883b4917da1cd1.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 426, + 116, + 469, + 150 + ], + "blocks": [ + { + "bbox": [ + 426, + 116, + 469, + 150 + ], + "lines": [ + { + "bbox": [ + 426, + 116, + 469, + 150 + ], + "spans": [ + { + "bbox": [ + 426, + 116, + 469, + 150 + ], + "type": "image", + "image_path": "79376126f34aa029127fd3ce5605754ed4cdf4223f5d396e986e61c7aa33bce3.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 474, + 116, + 517, + 150 + ], + "blocks": [ + { + "bbox": [ + 474, + 116, + 517, + 150 + ], + "lines": [ + { + "bbox": [ + 474, + 116, + 517, + 150 + ], + "spans": [ + { + "bbox": [ + 474, + 116, + 517, + 150 + ], + "type": "image", + "image_path": "e1173728ada22a4838013b3852b45ebaa1e8fdfe9c3726be9dc8882250753765.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 141, + 155, + 184, + 189 + ], + "blocks": [ + { + "bbox": [ + 77, + 167, + 128, + 176 + ], + "lines": [ + { + "bbox": [ + 77, + 167, + 128, + 176 + ], + "spans": [ + { + "bbox": [ + 77, + 167, + 128, + 176 + ], + "type": "text", + "content": "MaskFormer" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 141, + 155, + 184, + 189 + ], + "lines": [ + { + "bbox": [ + 141, + 155, + 184, + 189 + ], + "spans": [ + { + "bbox": [ + 141, + 155, + 184, + 189 + ], + "type": "image", + "image_path": "c0e1d0c951d5130119bcae556b19774997840fde610df1d97d573d86cf653e9d.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 189, + 155, + 231, + 189 + ], + "blocks": [ + { + "bbox": [ + 189, + 155, + 231, + 189 + ], + "lines": [ + { + "bbox": [ + 189, + 155, + 231, + 189 + ], + "spans": [ + { + "bbox": [ + 189, + 155, + 231, + 189 + ], + "type": "image", + "image_path": "db591762c71826ecaa7158c2b12bf95683049006dfebdb304c7e20371b6f4515.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 236, + 155, + 279, + 189 + ], + "blocks": [ + { + "bbox": [ + 236, + 155, + 279, + 189 + ], + "lines": [ + { + "bbox": [ + 236, + 155, + 279, + 189 + ], + "spans": [ + { + "bbox": [ + 236, + 155, + 279, + 189 + ], + "type": "image", + "image_path": "d8f20f5e6d68073aa3cc050da8834f5a3e3aec5dd2f010e2ec059f69b27b4331.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 284, + 155, + 326, + 189 + ], + "blocks": [ + { + "bbox": [ + 284, + 155, + 326, + 189 + ], + "lines": [ + { + "bbox": [ + 284, + 155, + 326, + 189 + ], + "spans": [ + { + "bbox": [ + 284, + 155, + 326, + 189 + ], + "type": "image", + "image_path": "fe5ce7aaafc8662297caa99a7e5b6cb10d9c2e0dba72a5ff72f9437a1e4313cc.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 331, + 155, + 374, + 189 + ], + "blocks": [ + { + "bbox": [ + 331, + 155, + 374, + 189 + ], + "lines": [ + { + "bbox": [ + 331, + 155, + 374, + 189 + ], + "spans": [ + { + "bbox": [ + 331, + 155, + 374, + 189 + ], + "type": "image", + "image_path": "8c28960cd6739c9b272ef918b04a8340319234ea4eafe5bc68a854067c0026a1.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 378, + 155, + 422, + 189 + ], + "blocks": [ + { + "bbox": [ + 378, + 155, + 422, + 189 + ], + "lines": [ + { + "bbox": [ + 378, + 155, + 422, + 189 + ], + "spans": [ + { + "bbox": [ + 378, + 155, + 422, + 189 + ], + "type": "image", + "image_path": "387c34e44e92582e9e53cc71016a1dd0564a869b596af22241cae2ea831d9fad.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 426, + 155, + 469, + 189 + ], + "blocks": [ + { + "bbox": [ + 426, + 155, + 469, + 189 + ], + "lines": [ + { + "bbox": [ + 426, + 155, + 469, + 189 + ], + "spans": [ + { + "bbox": [ + 426, + 155, + 469, + 189 + ], + "type": "image", + "image_path": "4362116f388be5ea0ff4b9b641d1388b18a2f9e4008b304b7764ee9dc3975d37.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 474, + 155, + 517, + 189 + ], + "blocks": [ + { + "bbox": [ + 474, + 155, + 517, + 189 + ], + "lines": [ + { + "bbox": [ + 474, + 155, + 517, + 189 + ], + "spans": [ + { + "bbox": [ + 474, + 155, + 517, + 189 + ], + "type": "image", + "image_path": "73b5df6e3e23054b0ae25a6f84887078d81c69e2b19f79724f117e1b3c8bce66.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 141, + 193, + 184, + 228 + ], + "blocks": [ + { + "bbox": [ + 76, + 205, + 130, + 215 + ], + "lines": [ + { + "bbox": [ + 76, + 205, + 130, + 215 + ], + "spans": [ + { + "bbox": [ + 76, + 205, + 130, + 215 + ], + "type": "text", + "content": "Mask2Former" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 141, + 193, + 184, + 228 + ], + "lines": [ + { + "bbox": [ + 141, + 193, + 184, + 228 + ], + "spans": [ + { + "bbox": [ + 141, + 193, + 184, + 228 + ], + "type": "image", + "image_path": "992f9eb9b56d378fed232f07294004d8d4a64f9e50ddee93fadad7d53120ce7a.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + } + ], + "index": 28 + }, + { + "type": "image", + "bbox": [ + 189, + 193, + 231, + 228 + ], + "blocks": [ + { + "bbox": [ + 189, + 193, + 231, + 228 + ], + "lines": [ + { + "bbox": [ + 189, + 193, + 231, + 228 + ], + "spans": [ + { + "bbox": [ + 189, + 193, + 231, + 228 + ], + "type": "image", + "image_path": "e4c446a0e1b243d0ba4616df93f3fc619aea03f1115eca0b52481f68db755730.jpg" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_body" + } + ], + "index": 29 + }, + { + "type": "image", + "bbox": [ + 236, + 193, + 279, + 228 + ], + "blocks": [ + { + "bbox": [ + 236, + 193, + 279, + 228 + ], + "lines": [ + { + "bbox": [ + 236, + 193, + 279, + 228 + ], + "spans": [ + { + "bbox": [ + 236, + 193, + 279, + 228 + ], + "type": "image", + "image_path": "52ceae1732bbbf2969d17510fd189d136709c4327dd09a037753090f56946505.jpg" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_body" + } + ], + "index": 30 + }, + { + "type": "image", + "bbox": [ + 284, + 193, + 326, + 227 + ], + "blocks": [ + { + "bbox": [ + 284, + 193, + 326, + 227 + ], + "lines": [ + { + "bbox": [ + 284, + 193, + 326, + 227 + ], + "spans": [ + { + "bbox": [ + 284, + 193, + 326, + 227 + ], + "type": "image", + "image_path": "0ee51e6fb06d8fe662d524134080c2d8804bca6a13e85567d145f99ede169a97.jpg" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_body" + } + ], + "index": 31 + }, + { + "type": "image", + "bbox": [ + 331, + 193, + 374, + 227 + ], + "blocks": [ + { + "bbox": [ + 331, + 193, + 374, + 227 + ], + "lines": [ + { + "bbox": [ + 331, + 193, + 374, + 227 + ], + "spans": [ + { + "bbox": [ + 331, + 193, + 374, + 227 + ], + "type": "image", + "image_path": "8938e82d1c75c88eb2d4fff5ebae2a7f54789fdacb59b1c22fad9fd7b06949e5.jpg" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_body" + } + ], + "index": 32 + }, + { + "type": "image", + "bbox": [ + 378, + 193, + 422, + 227 + ], + "blocks": [ + { + "bbox": [ + 378, + 193, + 422, + 227 + ], + "lines": [ + { + "bbox": [ + 378, + 193, + 422, + 227 + ], + "spans": [ + { + "bbox": [ + 378, + 193, + 422, + 227 + ], + "type": "image", + "image_path": "d6d11c388168127d855e356382384276441265e12b25a8a8f31a01c9706eef98.jpg" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_body" + } + ], + "index": 33 + }, + { + "type": "image", + "bbox": [ + 426, + 193, + 469, + 227 + ], + "blocks": [ + { + "bbox": [ + 426, + 193, + 469, + 227 + ], + "lines": [ + { + "bbox": [ + 426, + 193, + 469, + 227 + ], + "spans": [ + { + "bbox": [ + 426, + 193, + 469, + 227 + ], + "type": "image", + "image_path": "338d66e27e360a3afeca42a0b0120f27d8b718e9bca7b3e8ceea22d3e14617ea.jpg" + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_body" + } + ], + "index": 34 + }, + { + "type": "image", + "bbox": [ + 474, + 193, + 517, + 227 + ], + "blocks": [ + { + "bbox": [ + 474, + 193, + 517, + 227 + ], + "lines": [ + { + "bbox": [ + 474, + 193, + 517, + 227 + ], + "spans": [ + { + "bbox": [ + 474, + 193, + 517, + 227 + ], + "type": "image", + "image_path": "ee5769b9c410b51a87be24f13381f1194620b63864b1a475a037c2b34e354d86.jpg" + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "image_body" + } + ], + "index": 35 + }, + { + "type": "image", + "bbox": [ + 141, + 232, + 184, + 266 + ], + "blocks": [ + { + "bbox": [ + 86, + 244, + 127, + 253 + ], + "lines": [ + { + "bbox": [ + 86, + 244, + 127, + 253 + ], + "spans": [ + { + "bbox": [ + 86, + 244, + 127, + 253 + ], + "type": "text", + "content": "OSFormer" + } + ] + } + ], + "index": 36, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 141, + 232, + 184, + 266 + ], + "lines": [ + { + "bbox": [ + 141, + 232, + 184, + 266 + ], + "spans": [ + { + "bbox": [ + 141, + 232, + 184, + 266 + ], + "type": "image", + "image_path": "666cd1fd657c449bf2f31ec1108b8ffa3852d52648d2e20ecd5022e3c1763d18.jpg" + } + ] + } + ], + "index": 37, + "angle": 0, + "type": "image_body" + } + ], + "index": 37 + }, + { + "type": "image", + "bbox": [ + 189, + 232, + 231, + 266 + ], + "blocks": [ + { + "bbox": [ + 189, + 232, + 231, + 266 + ], + "lines": [ + { + "bbox": [ + 189, + 232, + 231, + 266 + ], + "spans": [ + { + "bbox": [ + 189, + 232, + 231, + 266 + ], + "type": "image", + "image_path": "5cb7de8580d9977f01b0adaf79a660a05c3a6ac29daa5742dbc2200f136aee71.jpg" + } + ] + } + ], + "index": 38, + "angle": 0, + "type": "image_body" + } + ], + "index": 38 + }, + { + "type": "image", + "bbox": [ + 236, + 232, + 279, + 266 + ], + "blocks": [ + { + "bbox": [ + 236, + 232, + 279, + 266 + ], + "lines": [ + { + "bbox": [ + 236, + 232, + 279, + 266 + ], + "spans": [ + { + "bbox": [ + 236, + 232, + 279, + 266 + ], + "type": "image", + "image_path": "021537d830b0c02cae219be65e3adaa7585fa476e0d3da773c93c9c4969d86d7.jpg" + } + ] + } + ], + "index": 39, + "angle": 0, + "type": "image_body" + } + ], + "index": 39 + }, + { + "type": "image", + "bbox": [ + 284, + 232, + 326, + 266 + ], + "blocks": [ + { + "bbox": [ + 284, + 232, + 326, + 266 + ], + "lines": [ + { + "bbox": [ + 284, + 232, + 326, + 266 + ], + "spans": [ + { + "bbox": [ + 284, + 232, + 326, + 266 + ], + "type": "image", + "image_path": "55fa1b75b3ce068fdaa6436a0d3f9f75c35cea3c544e49fc30eaa020edcf2289.jpg" + } + ] + } + ], + "index": 40, + "angle": 0, + "type": "image_body" + } + ], + "index": 40 + }, + { + "type": "image", + "bbox": [ + 331, + 232, + 374, + 266 + ], + "blocks": [ + { + "bbox": [ + 331, + 232, + 374, + 266 + ], + "lines": [ + { + "bbox": [ + 331, + 232, + 374, + 266 + ], + "spans": [ + { + "bbox": [ + 331, + 232, + 374, + 266 + ], + "type": "image", + "image_path": "a4ab53fc701fdbca6376516bcaf294ceae9bf200c883e8e9fa85b0c2c6886e11.jpg" + } + ] + } + ], + "index": 41, + "angle": 0, + "type": "image_body" + } + ], + "index": 41 + }, + { + "type": "image", + "bbox": [ + 378, + 232, + 422, + 266 + ], + "blocks": [ + { + "bbox": [ + 378, + 232, + 422, + 266 + ], + "lines": [ + { + "bbox": [ + 378, + 232, + 422, + 266 + ], + "spans": [ + { + "bbox": [ + 378, + 232, + 422, + 266 + ], + "type": "image", + "image_path": "2bda3a983111f15b722135ccaac65e512c88f0e4bb4e2d39131329fdb34a533d.jpg" + } + ] + } + ], + "index": 42, + "angle": 0, + "type": "image_body" + } + ], + "index": 42 + }, + { + "type": "image", + "bbox": [ + 426, + 232, + 469, + 266 + ], + "blocks": [ + { + "bbox": [ + 426, + 232, + 469, + 266 + ], + "lines": [ + { + "bbox": [ + 426, + 232, + 469, + 266 + ], + "spans": [ + { + "bbox": [ + 426, + 232, + 469, + 266 + ], + "type": "image", + "image_path": "9051c92ca773f440375f4221f3ac1b771bad6d22f912f7e73ef76fc8028f5ba0.jpg" + } + ] + } + ], + "index": 43, + "angle": 0, + "type": "image_body" + } + ], + "index": 43 + }, + { + "type": "image", + "bbox": [ + 474, + 232, + 517, + 266 + ], + "blocks": [ + { + "bbox": [ + 474, + 232, + 517, + 266 + ], + "lines": [ + { + "bbox": [ + 474, + 232, + 517, + 266 + ], + "spans": [ + { + "bbox": [ + 474, + 232, + 517, + 266 + ], + "type": "image", + "image_path": "f59478999595bfed7d1d79302b3b8a82ca2729b857a336e97f61478a5214b71a.jpg" + } + ] + } + ], + "index": 44, + "angle": 0, + "type": "image_body" + } + ], + "index": 44 + }, + { + "type": "image", + "bbox": [ + 141, + 270, + 184, + 304 + ], + "blocks": [ + { + "bbox": [ + 79, + 283, + 128, + 293 + ], + "lines": [ + { + "bbox": [ + 79, + 283, + 128, + 293 + ], + "spans": [ + { + "bbox": [ + 79, + 283, + 128, + 293 + ], + "type": "text", + "content": "DCNet(ours)" + } + ] + } + ], + "index": 45, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 141, + 270, + 184, + 304 + ], + "lines": [ + { + "bbox": [ + 141, + 270, + 184, + 304 + ], + "spans": [ + { + "bbox": [ + 141, + 270, + 184, + 304 + ], + "type": "image", + "image_path": "9b36f8506c4dbe9fc3f03e3fd04d4bcdcddf6b87e40439ebc2ca55726135cc66.jpg" + } + ] + } + ], + "index": 46, + "angle": 0, + "type": "image_body" + } + ], + "index": 46 + }, + { + "type": "image", + "bbox": [ + 189, + 270, + 231, + 304 + ], + "blocks": [ + { + "bbox": [ + 189, + 270, + 231, + 304 + ], + "lines": [ + { + "bbox": [ + 189, + 270, + 231, + 304 + ], + "spans": [ + { + "bbox": [ + 189, + 270, + 231, + 304 + ], + "type": "image", + "image_path": "a371356dedced26ac5dd6fe7c380ff55ce48f7d2f9f95ffcea356fddf7a01ef6.jpg" + } + ] + } + ], + "index": 47, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 316, + 537, + 327 + ], + "lines": [ + { + "bbox": [ + 55, + 316, + 537, + 327 + ], + "spans": [ + { + "bbox": [ + 55, + 316, + 537, + 327 + ], + "type": "text", + "content": "Figure 3. Qualitative results of DCNet compared with state-of-the-art methods. Different colored masks represent different instances." + } + ] + } + ], + "index": 54, + "angle": 0, + "type": "image_caption" + } + ], + "index": 47 + }, + { + "type": "image", + "bbox": [ + 236, + 270, + 279, + 304 + ], + "blocks": [ + { + "bbox": [ + 236, + 270, + 279, + 304 + ], + "lines": [ + { + "bbox": [ + 236, + 270, + 279, + 304 + ], + "spans": [ + { + "bbox": [ + 236, + 270, + 279, + 304 + ], + "type": "image", + "image_path": "3d6a4e5dc3d01faf1fd49cafce8039f219c4833f256bbe2f6ccc2973f2383142.jpg" + } + ] + } + ], + "index": 48, + "angle": 0, + "type": "image_body" + } + ], + "index": 48 + }, + { + "type": "image", + "bbox": [ + 284, + 270, + 326, + 304 + ], + "blocks": [ + { + "bbox": [ + 284, + 270, + 326, + 304 + ], + "lines": [ + { + "bbox": [ + 284, + 270, + 326, + 304 + ], + "spans": [ + { + "bbox": [ + 284, + 270, + 326, + 304 + ], + "type": "image", + "image_path": "7f795e4059003961d5fee101a687522949be5dbadb252c67f2a3432699780fac.jpg" + } + ] + } + ], + "index": 49, + "angle": 0, + "type": "image_body" + } + ], + "index": 49 + }, + { + "type": "image", + "bbox": [ + 331, + 270, + 374, + 304 + ], + "blocks": [ + { + "bbox": [ + 331, + 270, + 374, + 304 + ], + "lines": [ + { + "bbox": [ + 331, + 270, + 374, + 304 + ], + "spans": [ + { + "bbox": [ + 331, + 270, + 374, + 304 + ], + "type": "image", + "image_path": "ad88d32e41c940cf5ae7d250161964e1bdbca807d485b16e9cfddbfc7d95bc40.jpg" + } + ] + } + ], + "index": 50, + "angle": 0, + "type": "image_body" + } + ], + "index": 50 + }, + { + "type": "image", + "bbox": [ + 378, + 270, + 422, + 304 + ], + "blocks": [ + { + "bbox": [ + 378, + 270, + 422, + 304 + ], + "lines": [ + { + "bbox": [ + 378, + 270, + 422, + 304 + ], + "spans": [ + { + "bbox": [ + 378, + 270, + 422, + 304 + ], + "type": "image", + "image_path": "5dfb3dadd3b9d42bd8480a69574718a5938f8dacc59d4eb040d8a1f25e8f9f33.jpg" + } + ] + } + ], + "index": 51, + "angle": 0, + "type": "image_body" + } + ], + "index": 51 + }, + { + "type": "image", + "bbox": [ + 426, + 270, + 469, + 304 + ], + "blocks": [ + { + "bbox": [ + 426, + 270, + 469, + 304 + ], + "lines": [ + { + "bbox": [ + 426, + 270, + 469, + 304 + ], + "spans": [ + { + "bbox": [ + 426, + 270, + 469, + 304 + ], + "type": "image", + "image_path": "ee0272c1bb62fcfd5c28dd059da0e56ffdbcbbe9a7f60709c12aeda8d4f97f80.jpg" + } + ] + } + ], + "index": 52, + "angle": 0, + "type": "image_body" + } + ], + "index": 52 + }, + { + "type": "image", + "bbox": [ + 474, + 270, + 517, + 304 + ], + "blocks": [ + { + "bbox": [ + 474, + 270, + 517, + 304 + ], + "lines": [ + { + "bbox": [ + 474, + 270, + 517, + 304 + ], + "spans": [ + { + "bbox": [ + 474, + 270, + 517, + 304 + ], + "type": "image", + "image_path": "20300b7141a95ed2f7c01422085b07c13de7d9a89d8b79008aa86e8d6d01a06d.jpg" + } + ] + } + ], + "index": 53, + "angle": 0, + "type": "image_body" + } + ], + "index": 53 + }, + { + "type": "table", + "bbox": [ + 72, + 369, + 259, + 437 + ], + "blocks": [ + { + "bbox": [ + 46, + 338, + 288, + 359 + ], + "lines": [ + { + "bbox": [ + 46, + 338, + 288, + 359 + ], + "spans": [ + { + "bbox": [ + 46, + 338, + 288, + 359 + ], + "type": "text", + "content": "Table 4. Comparison of using different scales of features in terms of AP scores." + } + ] + } + ], + "index": 55, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 72, + 369, + 259, + 437 + ], + "lines": [ + { + "bbox": [ + 72, + 369, + 259, + 437 + ], + "spans": [ + { + "bbox": [ + 72, + 369, + 259, + 437 + ], + "type": "table", + "html": "
COD10KNC4K
single scale (1/32)45.252.5
single scale (1/16)44.851.6
single scale (1/8)45.152.2
multiple scales45.352.8
", + "image_path": "7e5ed0fb43e9385b2a3f82c32f1257101c7adf9f04c629983a92f62632611913.jpg" + } + ] + } + ], + "index": 56, + "angle": 0, + "type": "table_body" + } + ], + "index": 56 + }, + { + "bbox": [ + 46, + 456, + 287, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 456, + 287, + 517 + ], + "spans": [ + { + "bbox": [ + 46, + 456, + 287, + 517 + ], + "type": "text", + "content": "posed mechanism can well model the discrepancy between image features and camouflage characteristics. Removing both of the components leads to huge performance degradation, showing that the combination of them is essential in improving the camouflaged object segmentation." + } + ] + } + ], + "index": 57 + }, + { + "bbox": [ + 46, + 519, + 287, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 519, + 287, + 651 + ], + "spans": [ + { + "bbox": [ + 46, + 519, + 287, + 651 + ], + "type": "text", + "content": "Effectiveness of the Instance-level Camouflage Suppression Module. As shown in Table 3, the reference attention mechanism in our instance-level camouflage suppression module has the highest AP score on both datasets. Compared to vanilla cross-attention [39], our proposed reference points can effectively suppress the noise brought by background pixels in the prototype-pixel interaction, thus obtaining significant improvement for accurate segmentation. Meanwhile, our method is superior to existing cross-attention variants, such as the masked attention [5] in the segmentation field." + } + ] + } + ], + "index": 58 + }, + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "type": "text", + "content": "Effectiveness of Hierarchical Features To explore decamouflage at multiple feature levels, we feed 3 scales of features (i.e., stride of 32, 16, and 8) into the difference attention mechanism, respectively. As shown in Table 4, low-resolution features exceed other scales. This may attribute" + } + ] + } + ], + "index": 59 + }, + { + "type": "table", + "bbox": [ + 332, + 369, + 518, + 484 + ], + "blocks": [ + { + "bbox": [ + 305, + 338, + 545, + 361 + ], + "lines": [ + { + "bbox": [ + 305, + 338, + 545, + 361 + ], + "spans": [ + { + "bbox": [ + 305, + 338, + 545, + 361 + ], + "type": "text", + "content": "Table 5. Comparisons with OSFormer [33] of different backbones on COD10K [10] and NC4K [28] in terms of AP scores." + } + ] + } + ], + "index": 60, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 332, + 369, + 518, + 484 + ], + "lines": [ + { + "bbox": [ + 332, + 369, + 518, + 484 + ], + "spans": [ + { + "bbox": [ + 332, + 369, + 518, + 484 + ], + "type": "table", + "html": "
MethodBackboneCOD10KNC4K
OSFormerR5041.042.5
DCNet (ours)45.352.8
OSFormerR10142.044.4
DCNet (ours)46.853.5
OSFormerSwin-T47.750.2
DCNet (ours)50.356.3
OSFormerSwin-S52.156.7
DCNet (ours)52.358.4
", + "image_path": "f5d27f8f0fbffe75ca8ace940cc2276151ca761184d0540b4ef4d94b23246df8.jpg" + } + ] + } + ], + "index": 61, + "angle": 0, + "type": "table_body" + } + ], + "index": 61 + }, + { + "type": "image", + "bbox": [ + 307, + 493, + 423, + 569 + ], + "blocks": [ + { + "bbox": [ + 307, + 493, + 423, + 569 + ], + "lines": [ + { + "bbox": [ + 307, + 493, + 423, + 569 + ], + "spans": [ + { + "bbox": [ + 307, + 493, + 423, + 569 + ], + "type": "image", + "image_path": "bbf4c775b87a1d4b877c62d8215680e0c3e8a79611a80ab254b31c89474a0b8b.jpg" + } + ] + } + ], + "index": 62, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 578, + 547, + 601 + ], + "lines": [ + { + "bbox": [ + 305, + 578, + 547, + 601 + ], + "spans": [ + { + "bbox": [ + 305, + 578, + 547, + 601 + ], + "type": "text", + "content": "Figure 4. Comparisons of performance with different numbers of prototypes and reference points on COD10K [10] and NC4K [28]." + } + ] + } + ], + "index": 64, + "angle": 0, + "type": "image_caption" + } + ], + "index": 62 + }, + { + "type": "image", + "bbox": [ + 427, + 493, + 545, + 569 + ], + "blocks": [ + { + "bbox": [ + 427, + 493, + 545, + 569 + ], + "lines": [ + { + "bbox": [ + 427, + 493, + 545, + 569 + ], + "spans": [ + { + "bbox": [ + 427, + 493, + 545, + 569 + ], + "type": "image", + "image_path": "a9381e3e2a2f3d0974eb78b27e3b87aaf17951f8f8f1d381786999070e2e9310.jpg" + } + ] + } + ], + "index": 63, + "angle": 0, + "type": "image_body" + } + ], + "index": 63 + }, + { + "bbox": [ + 304, + 616, + 545, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 616, + 545, + 688 + ], + "spans": [ + { + "bbox": [ + 304, + 616, + 545, + 688 + ], + "type": "text", + "content": "to the retention of object semantic information. The highest performance in this experiment shows that the contribution of each scale is complementary to each other. Using single-scale features alone is not enough to model the target information, thus combining them together can further boost the performance." + } + ] + } + ], + "index": 65 + }, + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "text", + "content": "Compatibility for Feature Extractor. We equip our DC-Net with different feature extractor backbones, i.e., ResNet" + } + ] + } + ], + "index": 66 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17924" + } + ] + } + ], + "index": 67 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 71, + 286, + 236 + ], + "blocks": [ + { + "bbox": [ + 50, + 71, + 286, + 236 + ], + "lines": [ + { + "bbox": [ + 50, + 71, + 286, + 236 + ], + "spans": [ + { + "bbox": [ + 50, + 71, + 286, + 236 + ], + "type": "image", + "image_path": "b56eaac3c89497121214786ee85ee89972233eca8006d645d650bd0ac78b69f6.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 244, + 287, + 268 + ], + "lines": [ + { + "bbox": [ + 46, + 244, + 287, + 268 + ], + "spans": [ + { + "bbox": [ + 46, + 244, + 287, + 268 + ], + "type": "text", + "content": "Figure 5. Illustration of the amplitude images and difference maps in our proposed difference attention." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 275, + 287, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 275, + 287, + 335 + ], + "spans": [ + { + "bbox": [ + 46, + 275, + 287, + 335 + ], + "type": "text", + "content": "50 [15], ResNet-101 [15], Swin-T [26], and Swin-S [26]. All of them are pre-trained on ImageNet-1K [8]. As shown in Table 5, our method performs the best in all cases on both datasets, proving that our approach has tremendous potential for further improvement." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 336, + 289, + 553 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 336, + 289, + 553 + ], + "spans": [ + { + "bbox": [ + 46, + 336, + 289, + 553 + ], + "type": "text", + "content": "Analysis of Hyperparameters. In the instance-level camouflage suppression module, prototypes play an important role in capturing instances by aggregating object pixels. As shown in Figure 4(a), we conduct quantitative experiments to analyze how many prototypes are suitable for instance segmentation. It can be seen that too many or too few prototypes will damage the AP metric. In fact, the number of prototypes " + }, + { + "bbox": [ + 46, + 336, + 289, + 553 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 336, + 289, + 553 + ], + "type": "text", + "content": " should be larger than the count of instances in an image to avoid instance fusion error. While superabundant prototypes will cause many prototypes to match \"no object\" during training. We choose " + }, + { + "bbox": [ + 46, + 336, + 289, + 553 + ], + "type": "inline_equation", + "content": "N = 10" + }, + { + "bbox": [ + 46, + 336, + 289, + 553 + ], + "type": "text", + "content": " by balancing the performance on both datasets. Besides, in the reference attention mechanism, the number of reference points " + }, + { + "bbox": [ + 46, + 336, + 289, + 553 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 46, + 336, + 289, + 553 + ], + "type": "text", + "content": " influences the similarity distributions between prototypes and pixels. As shown in Figure 4(b), we can observe that the performance on NC4K continues to grow until " + }, + { + "bbox": [ + 46, + 336, + 289, + 553 + ], + "type": "inline_equation", + "content": "K = 64" + }, + { + "bbox": [ + 46, + 336, + 289, + 553 + ], + "type": "text", + "content": ", which means that it is sufficient for building appropriate prototype-pixel correction." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 561, + 135, + 574 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 561, + 135, + 574 + ], + "spans": [ + { + "bbox": [ + 47, + 561, + 135, + 574 + ], + "type": "text", + "content": "4.4. Visualizations" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 580, + 287, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 580, + 287, + 676 + ], + "spans": [ + { + "bbox": [ + 46, + 580, + 287, + 676 + ], + "type": "text", + "content": "Visualization of Difference Map. As shown in Figure 5, the " + }, + { + "bbox": [ + 46, + 580, + 287, + 676 + ], + "type": "inline_equation", + "content": "2^{nd}" + }, + { + "bbox": [ + 46, + 580, + 287, + 676 + ], + "type": "text", + "content": " row exhibits the amplitude images that are generated through Fourier transformation, which can be considered as camouflage characteristics. With the help of the difference attention mechanism, our network can remove the confusing background and localize the accurate salient areas (see " + }, + { + "bbox": [ + 46, + 580, + 287, + 676 + ], + "type": "inline_equation", + "content": "3^{rd}" + }, + { + "bbox": [ + 46, + 580, + 287, + 676 + ], + "type": "text", + "content": " row), which turn out to be the correct regions of camouflaged instances." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 677, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 288, + 715 + ], + "type": "text", + "content": "Visualization of Reference Attention. To explore the reference attention mechanism's ability to suppress the background noise, we make qualitative visualization to compare" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 309, + 72, + 545, + 230 + ], + "blocks": [ + { + "bbox": [ + 309, + 72, + 545, + 230 + ], + "lines": [ + { + "bbox": [ + 309, + 72, + 545, + 230 + ], + "spans": [ + { + "bbox": [ + 309, + 72, + 545, + 230 + ], + "type": "image", + "image_path": "36ae8d4fc95b96686bddad08067ee335e0d265ebd8cdfdf1931d64641717b526.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 239, + 545, + 262 + ], + "lines": [ + { + "bbox": [ + 305, + 239, + 545, + 262 + ], + "spans": [ + { + "bbox": [ + 305, + 239, + 545, + 262 + ], + "type": "text", + "content": "Figure 6. Visualizations of the activation map with different attention mechanisms." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 277, + 547, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 277, + 547, + 422 + ], + "spans": [ + { + "bbox": [ + 304, + 277, + 547, + 422 + ], + "type": "text", + "content": "different attention mechanisms. As shown in Figure 6 (d), with the proposed reference attention, the high response is clustered in the foreground area and the prototype pays little attention to the background area. In contrast, in Figure 6 (b), prototypes with the cross-attention [39] tend to absorb pixels from the non-target region. And in Figure 6 (c), with the masked attention [5], most background pixels are masked, while some foreground pixels are also suppressed, leading to sub-optimal performance. Thus, the reference mechanism plays an important role in mitigating the effect of background noise in prototype-pixel interactions, which helps to achieve more accurate segmentation." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 444, + 378, + 456 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 444, + 378, + 456 + ], + "spans": [ + { + "bbox": [ + 306, + 444, + 378, + 456 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 463, + 545, + 618 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 463, + 545, + 618 + ], + "spans": [ + { + "bbox": [ + 304, + 463, + 545, + 618 + ], + "type": "text", + "content": "In this paper, we propose an end-to-end Decamouflaging Network (DCNet) by jointly modeling pixel-level camouflage decoupling and instance-level camouflage suppression for camouflaged instance segmentation. We design a Pixel-level Camouflage Decoupling module to model and further eliminate the camouflage characteristics based on the frequency domain information. Besides, an Instance-level Camouflage Suppression module is proposed to achieve prototype-based instance segmentation and mitigate the effect of background noise by introducing reference points. Extensive experimental results on two benchmark datasets demonstrate the effectiveness of the proposed method." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 641, + 410, + 655 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 641, + 410, + 655 + ], + "spans": [ + { + "bbox": [ + 306, + 641, + 410, + 655 + ], + "type": "text", + "content": "6. Acknowledgment" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 665, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 665, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 305, + 665, + 547, + 715 + ], + "type": "text", + "content": "This work was partially supported by the National Nature Science Foundation of China (Grant 62022078, 62021001), National Defense Basic Scientific Research Program of China (Grant JCKY2021601B013)." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17925" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 53, + 91, + 288, + 135 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 288, + 135 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 288, + 135 + ], + "type": "text", + "content": "[1] Daniel Bolya, Chong Zhou, Fanyi Xiao, and Yong Jae Lee. Yolact: Real-time instance segmentation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9157-9166, 2019. 3, 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 136, + 288, + 178 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 136, + 288, + 178 + ], + "spans": [ + { + "bbox": [ + 53, + 136, + 288, + 178 + ], + "type": "text", + "content": "[2] Zhaowei Cai and Nuno Vasconcelos. Cascade r-cnn: high quality object detection and instance segmentation. IEEE transactions on pattern analysis and machine intelligence, 43(5):1483-1498, 2019. 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 180, + 288, + 235 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 180, + 288, + 235 + ], + "spans": [ + { + "bbox": [ + 53, + 180, + 288, + 235 + ], + "type": "text", + "content": "[3] Hao Chen, Kunyang Sun, Zhi Tian, Chunhua Shen, Yongming Huang, and Youliang Yan. Blendmask: Top-down meets bottom-up for instance segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8573-8581, 2020. 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 236, + 288, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 236, + 288, + 300 + ], + "spans": [ + { + "bbox": [ + 53, + 236, + 288, + 300 + ], + "type": "text", + "content": "[4] Kai Chen, Jiangmiao Pang, Jiaqi Wang, Yu Xiong, Xiaoxiao Li, Shuyang Sun, Wansen Feng, Ziwei Liu, Jianping Shi, Wanli Ouyang, et al. Hybrid task cascade for instance segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4974-4983, 2019. 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 302, + 288, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 302, + 288, + 357 + ], + "spans": [ + { + "bbox": [ + 53, + 302, + 288, + 357 + ], + "type": "text", + "content": "[5] Bowen Cheng, Ishan Misra, Alexander G Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1290–1299, 2022. 2, 3, 5, 6, 7, 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 357, + 288, + 401 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 357, + 288, + 401 + ], + "spans": [ + { + "bbox": [ + 53, + 357, + 288, + 401 + ], + "type": "text", + "content": "[6] Bowen Cheng, Alex Schwing, and Alexander Kirillov. Per-pixel classification is not all you need for semantic segmentation. Advances in Neural Information Processing Systems, 34:17864-17875, 2021. 2, 3, 5, 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 402, + 288, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 402, + 288, + 435 + ], + "spans": [ + { + "bbox": [ + 53, + 402, + 288, + 435 + ], + "type": "text", + "content": "[7] Hung-Kuo Chu, Wei-Hsin Hsu, Niloy J Mitra, Daniel Cohen-Or, Tien-Tsin Wong, and Tong-Yee Lee. Camouflage images. ACM Trans. Graph., 29(4):51-1, 2010. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 436, + 288, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 436, + 288, + 479 + ], + "spans": [ + { + "bbox": [ + 53, + 436, + 288, + 479 + ], + "type": "text", + "content": "[8] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 5, 8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 480, + 288, + 513 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 480, + 288, + 513 + ], + "spans": [ + { + "bbox": [ + 53, + 480, + 288, + 513 + ], + "type": "text", + "content": "[9] Deng-Ping Fan, Ge-Peng Ji, Ming-Ming Cheng, and Ling Shao. Concealed object detection. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2021. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 514, + 287, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 514, + 287, + 568 + ], + "spans": [ + { + "bbox": [ + 48, + 514, + 287, + 568 + ], + "type": "text", + "content": "[10] Deng-Ping Fan, Ge-Peng Ji, Guolei Sun, Ming-Ming Cheng, Jianbing Shen, and Ling Shao. Camouflaged object detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2777-2787, 2020. 2, 4, 5, 6, 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 569, + 287, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 569, + 287, + 624 + ], + "spans": [ + { + "bbox": [ + 48, + 569, + 287, + 624 + ], + "type": "text", + "content": "[11] Deng-Ping Fan, Tao Zhou, Ge-Peng Ji, Yi Zhou, Geng Chen, Huazhu Fu, Jianbing Shen, and Ling Shao. Inf-net: Automatic Covid-19 lung infection segmentation from ct images. IEEE Transactions on Medical Imaging, 39(8):2626-2637, 2020. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 624, + 287, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 624, + 287, + 669 + ], + "spans": [ + { + "bbox": [ + 48, + 624, + 287, + 669 + ], + "type": "text", + "content": "[12] Yuxin Fang, Shusheng Yang, Xinggang Wang, Yu Li, Chen Fang, Ying Shan, Bin Feng, and Wenyu Liu. Instances as queries. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6910-6919, 2021. 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 670, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 670, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 670, + 287, + 713 + ], + "type": "text", + "content": "[13] Ruohao Guo, Dantong Niu, Liao Qu, and Zhenbo Li. Sotr: Segmenting objects with transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7157-7166, 2021. 6" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "text", + "content": "[14] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In Proceedings of the IEEE international conference on computer vision, pages 2961-2969, 2017. 1, 3, 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 118, + 545, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 545, + 162 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 545, + 162 + ], + "type": "text", + "content": "[15] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 4, 5, 8" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 163, + 545, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 163, + 545, + 207 + ], + "spans": [ + { + "bbox": [ + 308, + 163, + 545, + 207 + ], + "type": "text", + "content": "[16] Xiaobin Hu, Deng-Ping Fan, Xuebin Qin, Hang Dai, Wenqi Ren, Ying Tai, Chengjie Wang, and Ling Shao. High-resolution iterative feedback network for camouflaged object detection. arXiv preprint arXiv:2203.11624, 2022. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 209, + 545, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 209, + 545, + 252 + ], + "spans": [ + { + "bbox": [ + 308, + 209, + 545, + 252 + ], + "type": "text", + "content": "[17] Zhaojin Huang, Lichao Huang, Yongchao Gong, Chang Huang, and Xinggang Wang. Mask scoring r-cnn. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6409-6418, 2019. 3, 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 254, + 545, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 254, + 545, + 308 + ], + "spans": [ + { + "bbox": [ + 308, + 254, + 545, + 308 + ], + "type": "text", + "content": "[18] Ivan Huerta, Daniel Rowe, Mikhail Mozerov, and Jordi González. Improving background subtraction based on a casuistry of colour-motion segmentation problems. In Iberian Conference on Pattern Recognition and Image Analysis, pages 475-482. Springer, 2007. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 310, + 545, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 310, + 545, + 354 + ], + "spans": [ + { + "bbox": [ + 308, + 310, + 545, + 354 + ], + "type": "text", + "content": "[19] Ge-Peng Ji, Deng-Ping Fan, Yu-Cheng Chou, Dengxin Dai, Alexander Liniger, and Luc Van Gool. Deep gradient learning for efficient camouflaged object detection. Machine Intelligence Research, 20(1):92-108, 2023. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 355, + 545, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 355, + 545, + 409 + ], + "spans": [ + { + "bbox": [ + 308, + 355, + 545, + 409 + ], + "type": "text", + "content": "[20] Lei Ke, Martin Danelljan, Xia Li, Yu-Wing Tai, Chi-Keung Tang, and Fisher Yu. Mask transfiner for high-quality instance segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4412–4421, 2022. 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 411, + 545, + 454 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 411, + 545, + 454 + ], + "spans": [ + { + "bbox": [ + 308, + 411, + 545, + 454 + ], + "type": "text", + "content": "[21] Alexander Kirillov, Yuxin Wu, Kaiming He, and Ross Girshick. Pointrend: Image segmentation as rendering. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9799-9808, 2020. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 456, + 545, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 456, + 545, + 511 + ], + "spans": [ + { + "bbox": [ + 308, + 456, + 545, + 511 + ], + "type": "text", + "content": "[22] Trung-Nghia Le, Yubo Cao, Tan-Cong Nguyen, Minh-Quan Le, Khanh-Duy Nguyen, Thanh-Toan Do, Minh-Triet Tran, and Tam V Nguyen. Camouflaged instance segmentation in-the-wild: Dataset, method, and benchmark suite. IEEE Transactions on Image Processing, 31:287-300, 2021. 1, 4" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 512, + 545, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 512, + 545, + 555 + ], + "spans": [ + { + "bbox": [ + 308, + 512, + 545, + 555 + ], + "type": "text", + "content": "[23] Trung-Nghia Le, Tam V Nguyen, Zhongliang Nie, MinhTriet Tran, and Akihiro Sugimoto. Anabranch network for camouflaged object segmentation. Computer Vision and Image Understanding, 184:45-56, 2019. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 557, + 545, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 557, + 545, + 612 + ], + "spans": [ + { + "bbox": [ + 308, + 557, + 545, + 612 + ], + "type": "text", + "content": "[24] Tsung-Yi Lin, Piotr Dólar, Ross Girshick, Kaiming He, Bharath Hariharan, and Serge Belongie. Feature pyramid networks for object detection. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2117-2125, 2017. 4" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 613, + 545, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 613, + 545, + 657 + ], + "spans": [ + { + "bbox": [ + 308, + 613, + 545, + 657 + ], + "type": "text", + "content": "[25] Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dólar. Focal loss for dense object detection. In Proceedings of the IEEE international conference on computer vision, pages 2980-2988, 2017. 5" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 658, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 658, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 658, + 545, + 713 + ], + "type": "text", + "content": "[26] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10012-10022, 2021. 8" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "text", + "content": "17926" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 48, + 72, + 286, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 286, + 95 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 286, + 95 + ], + "type": "text", + "content": "[27] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 5" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 96, + 287, + 151 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 96, + 287, + 151 + ], + "spans": [ + { + "bbox": [ + 48, + 96, + 287, + 151 + ], + "type": "text", + "content": "[28] Yunqiu Lv, Jing Zhang, Yuchao Dai, Aixuan Li, Bowen Liu, Nick Barnes, and Deng-Ping Fan. Simultaneously localize, segment and rank the camouflaged objects. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11591-11601, 2021. 5, 6, 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 152, + 286, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 152, + 286, + 207 + ], + "spans": [ + { + "bbox": [ + 48, + 152, + 286, + 207 + ], + "type": "text", + "content": "[29] Haiyang Mei, Ge-Peng Ji, Ziqi Wei, Xin Yang, Xiaopeng Wei, and Deng-Ping Fan. Camouflaged object segmentation with distraction mining. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8772-8781, 2021. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 209, + 286, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 209, + 286, + 262 + ], + "spans": [ + { + "bbox": [ + 48, + 209, + 286, + 262 + ], + "type": "text", + "content": "[30] Fausto Milletari, Nassir Navab, and Seyed-Ahmad Ahmadi. V-net: Fully convolutional neural networks for volumetric medical image segmentation. In 2016 fourth international conference on 3D vision (3DV), pages 565-571. IEEE, 2016. 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 263, + 286, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 263, + 286, + 297 + ], + "spans": [ + { + "bbox": [ + 48, + 263, + 286, + 297 + ], + "type": "text", + "content": "[31] Yuxin Pan, Yiwang Chen, Qiang Fu, Ping Zhang, and Xin Xu. Study on the camouflaged target detection method based on 3d convexity. Modern Applied Science, 5(4):152, 2011. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 298, + 286, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 298, + 286, + 353 + ], + "spans": [ + { + "bbox": [ + 48, + 298, + 286, + 353 + ], + "type": "text", + "content": "[32] Youwei Pang, Xiaoqi Zhao, Tian-Zhu Xiang, Lihe Zhang, and Huchuan Lu. Zoom in and out: A mixed-scale triplet network for camouflaged object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2160-2170, 2022. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 354, + 286, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 354, + 286, + 409 + ], + "spans": [ + { + "bbox": [ + 48, + 354, + 286, + 409 + ], + "type": "text", + "content": "[33] Jialun Pei, Tianyang Cheng, Deng-Ping Fan, He Tang, Chuanbo Chen, and Luc Van Gool. Osformer: One-stage camouflaged instance segmentation with transformers. In European Conference on Computer Vision, pages 19-37. Springer, 2022. 1, 2, 4, 5, 6, 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 410, + 286, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 410, + 286, + 453 + ], + "spans": [ + { + "bbox": [ + 48, + 410, + 286, + 453 + ], + "type": "text", + "content": "[34] Leon N Piotrowski and Fergus W Campbell. A demonstration of the visual importance and flexibility of spatial-frequency amplitude and phase. Perception, 11(3):337-346, 1982. 2, 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 455, + 286, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 455, + 286, + 498 + ], + "spans": [ + { + "bbox": [ + 48, + 455, + 286, + 498 + ], + "type": "text", + "content": "[35] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. Advances in neural information processing systems, 28, 2015. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 500, + 286, + 554 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 500, + 286, + 554 + ], + "spans": [ + { + "bbox": [ + 48, + 500, + 286, + 554 + ], + "type": "text", + "content": "[36] P Sengottuvelan, Amitabh Wahi, and A Shanmugam. Performance of decamouflaging through exploratory image analysis. In 2008 First International Conference on Emerging Trends in Engineering and Technology, pages 6-10. IEEE, 2008. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 555, + 286, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 555, + 286, + 588 + ], + "spans": [ + { + "bbox": [ + 48, + 555, + 286, + 588 + ], + "type": "text", + "content": "[37] Sujit K Singh, Chitra A Dhawale, and Sanjay Misra. Survey of object detection methods in camouflaged image. IERI Procedia, 4:351-357, 2013. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 590, + 286, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 590, + 286, + 624 + ], + "spans": [ + { + "bbox": [ + 48, + 590, + 286, + 624 + ], + "type": "text", + "content": "[38] Zhi Tian, Chunhua Shen, and Hao Chen. Conditional convolutions for instance segmentation. In European conference on computer vision, pages 282-298. Springer, 2020. 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 624, + 286, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 624, + 286, + 668 + ], + "spans": [ + { + "bbox": [ + 48, + 624, + 286, + 668 + ], + "type": "text", + "content": "[39] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 5, 6, 7, 8" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 670, + 286, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 670, + 286, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 670, + 286, + 712 + ], + "type": "text", + "content": "[40] Xinlong Wang, Tao Kong, Chunhua Shen, Yuning Jiang, and Lei Li. Solo: Segmenting objects by locations. In European Conference on Computer Vision, pages 649-665. Springer, 2020. 1, 3" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 72, + 545, + 429 + ], + "type": "list", + "angle": 0, + "index": 22, + "blocks": [ + { + "bbox": [ + 307, + 72, + 545, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 72, + 545, + 116 + ], + "spans": [ + { + "bbox": [ + 307, + 72, + 545, + 116 + ], + "type": "text", + "content": "[41] Xinlong Wang, Rufeng Zhang, Tao Kong, Lei Li, and Chunhua Shen. Solov2: Dynamic and fast instance segmentation. Advances in Neural information processing systems, 33:17721-17732, 2020. 1, 3, 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 118, + 545, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 545, + 171 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 545, + 171 + ], + "type": "text", + "content": "[42] Qinwei Xu, Ruipeng Zhang, Ya Zhang, Yanfeng Wang, and Qi Tian. A fourier-based framework for domain generalization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14383-14392, 2021. 2, 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 174, + 545, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 174, + 545, + 228 + ], + "spans": [ + { + "bbox": [ + 307, + 174, + 545, + 228 + ], + "type": "text", + "content": "[43] Fan Yang, Qiang Zhai, Xin Li, Rui Huang, Ao Luo, Hong Cheng, and Deng-Ping Fan. Uncertainty-guided transformer reasoning for camouflaged object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4146-4155, 2021. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 230, + 545, + 283 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 230, + 545, + 283 + ], + "spans": [ + { + "bbox": [ + 307, + 230, + 545, + 283 + ], + "type": "text", + "content": "[44] Xiaoqi Zhao, Lihe Zhang, and Huchuan Lu. Automatic polyp segmentation via multi-scale subtraction network. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 120-130. Springer, 2021. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 285, + 545, + 339 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 285, + 545, + 339 + ], + "spans": [ + { + "bbox": [ + 307, + 285, + 545, + 339 + ], + "type": "text", + "content": "[45] Yijie Zhong, Bo Li, Lv Tang, Senyun Kuang, Shuang Wu, and Shouhong Ding. Detecting camouflaged object in frequency domain. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4504-4513, 2022. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 341, + 545, + 384 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 341, + 545, + 384 + ], + "spans": [ + { + "bbox": [ + 307, + 341, + 545, + 384 + ], + "type": "text", + "content": "[46] Tao Zhou, Yi Zhou, Chen Gong, Jian Yang, and Yu Zhang. Feature aggregation and propagation network for camouflaged object detection. IEEE Transactions on Image Processing, 31:7036-7047, 2022. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 386, + 545, + 429 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 386, + 545, + 429 + ], + "spans": [ + { + "bbox": [ + 307, + 386, + 545, + 429 + ], + "type": "text", + "content": "[47] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable detr: Deformable transformers for end-to-end object detection. arXiv preprint arXiv:2010.04159, 2020. 4" + } + ] + } + ], + "index": 21 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17927" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Camouflaged Object Detection With Feature Decomposition and Edge Reconstruction/69e6313d-2f4d-49ec-aeb1-e472deb81070_content_list.json b/2023/Camouflaged Object Detection With Feature Decomposition and Edge Reconstruction/69e6313d-2f4d-49ec-aeb1-e472deb81070_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..cdc2b2f777f8a77e352bac5e1e819a7a522492d6 --- /dev/null +++ b/2023/Camouflaged Object Detection With Feature Decomposition and Edge Reconstruction/69e6313d-2f4d-49ec-aeb1-e472deb81070_content_list.json @@ -0,0 +1,1702 @@ +[ + { + "type": "text", + "text": "Camouflaged Object Detection with Feature Decomposition and Edge Reconstruction", + "text_level": 1, + "bbox": [ + 163, + 130, + 805, + 174 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Chunming He $^{1}$ , Kai Li $^{2*}$ , Yachao Zhang $^{1}$ , Longxiang Tang $^{1}$ , Yulun Zhang $^{3}$ , Zhenhua Guo $^{4}$ , Xiu Li $^{1*}$ $^{1}$ Shenzhen International Graduate School, Tsinghua University, $^{2}$ NEC Laboratories America, $^{3}$ ETH Zürich, $^{4}$ Tianyi Traffic Technology", + "bbox": [ + 86, + 202, + 890, + 256 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{chunminghe19990224,li.gml.kai,lloong.x,yulun100,cszguo}@gmail.com", + "bbox": [ + 194, + 258, + 777, + 273 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "yachaozhang@stu.xmu.edu.cn, li.xiu@sz.tsinghua.edu.cn", + "bbox": [ + 250, + 277, + 712, + 291 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 325, + 312, + 342 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Camouflaged object detection (COD) aims to address the tough issue of identifying camouflaged objects visually blended into the surrounding backgrounds. COD is a challenging task due to the intrinsic similarity of camouflaged objects with the background, as well as their ambiguous boundaries. Existing approaches to this problem have developed various techniques to mimic the human visual system. Albeit effective in many cases, these methods still struggle when camouflaged objects are so deceptive to the vision system. In this paper, we propose the FEature Decomposition and Edge Reconstruction (FEDER) model for COD. The FEDER model addresses the intrinsic similarity of foreground and background by decomposing the features into different frequency bands using learnable wavelets. It then focuses on the most informative bands to mine subtle cues that differentiate foreground and background. To achieve this, a frequency attention module and a guidance-based feature aggregation module are developed. To combat the ambiguous boundary problem, we propose to learn an auxiliary edge reconstruction task alongside the COD task. We design an ordinary differential equation-inspired edge reconstruction module that generates exact edges. By learning the auxiliary task in conjunction with the COD task, the FEDER model can generate precise prediction maps with accurate object boundaries. Experiments show that our FEDER model significantly outperforms state-of-the-art methods with cheaper computational and memory costs. The code will be available at https://github.com/ChunmingHe/FEDER.", + "bbox": [ + 76, + 358, + 473, + 796 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 78, + 823, + 207, + 838 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Camouflaged object detection (COD) aims to detect and segment objects \"seamlessly\" integrated into surrounding", + "bbox": [ + 76, + 848, + 468, + 878 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/2ceb23f8e240956ace12da1047172462b49289f379c43fd2be9387783d0522ab.jpg", + "image_caption": [ + "(a) Origin", + "(b) GT", + "(c) SegMaR", + "(d) FEDER", + "Figure 1. Results of SegMaR [14] and our method under the intrinsic similarity (IS) and edge disruption (ED) challenges. Our method better localizes the objects and produces clearer edges." + ], + "image_footnote": [], + "bbox": [ + 506, + 327, + 885, + 570 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "environments. COD is a challenging task as it needs to combat against excellent camouflage strategies, including background matching [41], disruptive coloration [33], etc., and distinguish the subtle differences between candidate objects and their backgrounds. Research in COD can simultaneously facilitate the development of visual perception for nuance discrimination and promote various valuable real-life applications, ranging from concealed defect detection [17] in industry to pest monitoring [35] in agriculture.", + "bbox": [ + 496, + 643, + 890, + 779 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "COD faces two main challenges. The first is the intrinsic similarity (IS) challenge, which occurs when camouflaged objects share similar colors and patterns with their backgrounds. This makes it difficult to even roughly localize those camouflaged objects. The second is the edge disruption (ED) challenge, which arises from extremely ambiguous object boundaries. Even if a rough localization is achieved, precise segmentation can barely be obtained.", + "bbox": [ + 496, + 779, + 890, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author.", + "bbox": [ + 94, + 886, + 220, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "22046", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To tackle these challenges, most existing works aim to develop models that mimic the human visual system [6,28]. However, since camouflage strategies are designed by prey to confuse the predator's visual system, and the intrinsic topological properties of candidate objects are not distinctive, such human perception-oriented attempts may struggle to identify subtle discriminative features and fail to effectively address the above challenges. For instance, as illustrated in Fig. 1, the state-of-the-art human perception-based COD method can only generate inaccurate prediction maps, such as the vague caddisfly and incomplete dog (Row 2 and 4), or even fail to detect camouflaged objects like the bird and snake (Row 1 and 3). Therefore, a better COD method should compensate for the \"flaw\" in human perception by emphasizing subtle discriminative features.", + "bbox": [ + 75, + 90, + 472, + 316 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Based on the biological study [41], camouflaged objects often employ various camouflage strategies to conceal their discriminative differences, which mainly exist in texture details and global information distribution, within surrounding environments. Such a study inspires us to cope with the COD task by decomposing the camouflage scenario into different parts. This allows for the disentanglement of various intricate connections, enabling each part to be separately handled to fully excavate subtle discriminative cues.", + "bbox": [ + 75, + 318, + 472, + 455 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "With this inspiration, we propose the FEature Decomposition and Edge Reconstruction (FEDER) model for the COD task, which compensates for the deficiencies of human perception by emphasizing subtle discriminative features and effectively addresses the IS and ED challenges. Specifically, to combat the intractable localization problem caused by the IS challenge, we design the deep wavelet-like decomposition (DWD) strategy, which decomposes the extracted features into different frequency bands using learnable wavelet-like modules. Then, we focus on the most informative bands by filtering out noteworthy parts where discriminative cues are most likely to exist by a novel frequency attention (FA) module. Moreover, a guidance-based feature aggregation (GFA) module is proposed to aggregate the multi-scale decomposed features with attentional guidance to further emphasize discriminative information.", + "bbox": [ + 75, + 455, + 472, + 696 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address the ambiguous boundary problem of the ED challenge, we propose learning an auxiliary edge reconstruction task to encourage the network to excavate edge details. We design the ordinary differential equation (ODE)-inspired edge reconstruction (OER) module to reconstruct accurate and complete edge prediction maps using a high-order ODE solver, specifically, the second-order Runge-Kutta. Incorporating this auxiliary task with the COD task can facilitate the generation of precise segmentation results with accurate object boundaries.", + "bbox": [ + 75, + 696, + 472, + 848 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our contributions are summarized as follows:", + "bbox": [ + 96, + 849, + 398, + 864 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "- We propose the FEature Decomposition and Edge Reconstruction (FEDER) model for the COD task. To", + "bbox": [ + 93, + 869, + 470, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "the best of our knowledge, we are the first to approach COD from a decomposition perspective.", + "bbox": [ + 529, + 90, + 890, + 121 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- To highlight the subtle discriminative features, we propose frequency attention modules to filter out the noteworthy parts of corresponding features and design the Guidance-based Feature Aggregation module to aggregate the multi-scale features with attentional guidance.", + "- We propose to learn an auxiliary edge reconstruction task along with the COD task to help generate precise segmentation maps with accurate object boundaries and design the ODE-inspired edge reconstruction module for complete edge prediction.", + "- The proposed FEDER significantly outperforms the state-of-the-art methods on four datasets by a large margin with cheaper computational and memory costs." + ], + "bbox": [ + 516, + 123, + 890, + 325 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 500, + 345, + 650, + 363 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Camouflaged object detection. Unlike existing object detection tasks, camouflaged object detection (COD) poses new challenges for mining subtle discriminative features under complex camouflage strategies [6, 11]. Early techniques utilized the hand-crafted operators for COD [11, 30], which were only applicable to camouflaged scenarios with simple backgrounds. Recent research has leveraged the huge capacity of deep learning to detect camouflaged objects in a learning manner [6, 14, 28]. Inspired by the hunting process of predators, SINet [6] designed a bio-inspired network to gradually search and locate the camouflaged object. PFNet [28] proposed the position module and focus module to imitate human identification with the distraction mining strategy. By simulating human behaviors in understanding complex scenarios, SegMaR [14] integrated segment, magnify and reiterate in a coarse-to-fine manner using the multistage strategy. However, these COD solutions mainly focus on mimicking biovision systems, which can be easily confused by complex camouflaged strategies and struggle to excavate the subtle discriminative features, thus failing to handle the IS and ED challenges (see Fig. 1). Unlike these human perception-oriented techniques, we first propose to address the COD task from a decomposition perspective by decomposing the extracted features into different frequency bands with learnable wavelets and filtering out the most informative bands to excavate those inconspicuous discriminative features, thus remedying the human visual deficiency and solving the IS challenge. To handle the ED challenge, we propose learning an auxiliary edge reconstruction task along with the COD task to facilitate the generation of precise segmentation results with clear object boundaries.", + "bbox": [ + 496, + 371, + 890, + 840 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Deep wavelet decomposition. Deep wavelet decomposition is an effective tool to decompose image/feature into various frequency components and has gained immense popularity in many domains, such as image restoration [15]", + "bbox": [ + 496, + 840, + 890, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "22047", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/e1db8a771e890aa816d72c7592408874967dd9d3e4bfbfbba41e6d5092702720.jpg", + "image_caption": [ + "Figure 2. Overview of the proposed FEDER model. CFE and SED are short for camouflaged feature encoder and segmentation-oriented edge-assisted decoder. HFA/LFA and GFA denote high/low frequency attention modules and the guidance-based feature aggregation module. RRS and OER indicate the reversible re-calibration segmentation module and the ODE-inspired edge reconstruction module." + ], + "image_footnote": [], + "bbox": [ + 81, + 88, + 890, + 395 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "and style transfer [49]. To handle the IS challenge, we introduce deep wavelet decomposition into the COD task. Furthermore, to better accommodate the COD data, we employ the learnable wavelets for deep adaptive feature decomposition, whose coefficients are updated following AWD [8].", + "bbox": [ + 75, + 460, + 468, + 537 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "ODE-inspired network. Researchers have established a relationship between ODE and neural networks. [46] first analyzed ResNet from the perspective of discrete ODE and [1] further extended ResNet to an ODE-inspired network architecture with a more accurate transmission. Since then, ODE-inspired networks are widely utilized in many fields, such as image dehazing [37] and machine translation [19]. In this paper, to accommodate the fine-grained property of the edge, we propose an ODE-inspired edge reconstruction module with the second-order Runge-Kutta and a weighted gate mechanism, aiming to generate more accurate boundaries. Furthermore, we apply the Hamiltonian system to our OER module to ensure the stability of edge reconstruction.", + "bbox": [ + 75, + 537, + 470, + 734 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Methodology", + "text_level": 1, + "bbox": [ + 76, + 753, + 210, + 768 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Given a camouflaged image, we first extract a cascade of features using the camouflaged feature encoder (CFE). We then perform a wavelet-like decomposition (DWD) on the features to decompose them into different frequency bands. We select the most informative bands, such as the high-frequency and low-frequency components, for further analysis. These informative bands are processed by the frequency attention (FA) module and guidance-based feature", + "bbox": [ + 75, + 780, + 468, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "aggregation (GFA) module to highlight the inconspicuous discriminative features. With the aggregated features, the segmentation-oriented edge-assisted decoder (SED) outputs both the segmentation map and the edge prediction map. Fig. 2 presents the framework of our FEDER model.", + "bbox": [ + 496, + 460, + 890, + 537 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Camouflaged Feature Encoder (CFE)", + "text_level": 1, + "bbox": [ + 498, + 547, + 823, + 564 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Following SINet V2 [4], the basic encoder $E$ adopts ResNet50 [10]/Res2Net50 [7] as its backbone. Given an image $I_{c}$ of size $W\\times H$ , the basic encoder $E$ generates a set of feature maps $\\{f_k\\}_{k = 0}^4$ with the resolution of $\\frac{H}{2^{k + 1}}\\times \\frac{W}{2^{k + 1}}$ , R-Net [6] is cascaded to transform $\\{f_k\\}_{k = 1}^4$ into a more informative and compact output, i.e., a series of 64-channel feature maps $\\{f_k^r\\}_{k = 1}^4$ . Additionally, the last feature map $f_{4}$ from the basic encoder $E$ is further fed into an efficient atrous spatial pyramid pooling (e-ASPP) $A_{e}$ [16] to enlarge the receptive field and fuse the multi-context information, resulting in $d_5^s = A_e(f_4)$ , where $d_5^s$ is a coarse segmentation result with the same spatial resolution as $f_{4}$ .", + "bbox": [ + 496, + 570, + 890, + 752 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Deep Wavelet-like Decomposition", + "text_level": 1, + "bbox": [ + 500, + 762, + 792, + 779 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2.1 Learnable Wavelet-like Decomposition", + "text_level": 1, + "bbox": [ + 500, + 786, + 823, + 801 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Camouflaged objects share a high intrinsic similarity with the background, which poses challenges for common feature extractors to mine the inconspicuous discriminative features, ultimately resulting in the suppression of segmentation performance. Based on the biological study [41], the discriminative features of COD mainly exist in the high-", + "bbox": [ + 496, + 809, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "22048", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/6d23091521cfbf166c2637d633254fb290ef33cf79932c2078161c5135671d3e.jpg", + "image_caption": [ + "Figure 3. Details of the proposed GFA, HFA, and LFA" + ], + "image_footnote": [], + "bbox": [ + 81, + 90, + 359, + 169 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/0197348a3e691a59e03b998b156935360fa977a868b609fd2185dcc4a5f125bd.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 362, + 90, + 635, + 167 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/7f6f84ff9fa0dff726a61ec5c2430ab3398798849268bc5678813f0ee5ea922c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 638, + 90, + 888, + 169 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "frequency (HF) component, e.g., texture and edge, and the low-frequency (LF) component, e.g., color and illumination. Inspired by this study, we propose to perform deep wavelet-like decomposition (DWD) on the extracted features $\\{f_k^r\\}_{k=1}^4$ and select the most informative HF and LF components for further refinement. We decompose $f_k^r$ as", + "bbox": [ + 75, + 203, + 468, + 294 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\left(f _ {k} ^ {r}\\right) _ {H F} = W _ {H F} \\left(f _ {k} ^ {r}\\right), \\left(f _ {k} ^ {r}\\right) _ {L F} = W _ {L F} \\left(f _ {k} ^ {r}\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 119, + 296, + 468, + 313 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $(f_k^r)_{HF}$ and $(f_k^r)_{LF}$ denote the HF and LF components of $f_k^r$ . $W_{HF}$ and $W_{LF}$ represent the learnable HF and LF filters with the coefficients updated following AWD [8] and initialized by Haar wavelet [40]. The learned wavelet-like transformer is expected to better cater to COD data than manually-designed wavelets [36, 42], thus further facilitating the extraction of inconspicuous discriminative features.", + "bbox": [ + 75, + 314, + 468, + 420 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2.2 Frequency Attention Modules", + "text_level": 1, + "bbox": [ + 76, + 452, + 339, + 467 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To extract discriminative information from the decomposed features, we propose a high-frequency attention (HFA) module and a low-frequency attention (LFA) module, corresponding to HF and LF bands, respectively. The detailed structures of the two modules are illustrated in Fig. 3.", + "bbox": [ + 75, + 479, + 468, + 556 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "High-frequency attention module. We design the HFA module to accentuate those texture-rich regions for subtle discriminative feature extraction. Following [21, 26], we first apply a residual block for texture preservation, consisting of a $3 \\times 3$ convolution layer, batch normalization (BN) [12], and ReLU. We then employ the joint attention module $J A(\\cdot)$ , which includes spatial attention [34] and channel attention [45], to highlight noteworthy parts in both spatial and channel domains. Therefore, given HF features $(f_k^r)_{HF}$ , the HF attention map $p_k^h$ is formulated as follows:", + "bbox": [ + 75, + 559, + 468, + 710 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\np _ {k} ^ {h} = J A \\left(R e s B \\left(\\left(f _ {k} ^ {r}\\right) _ {H F}\\right)\\right), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 171, + 713, + 468, + 729 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $ResB(\\cdot)$ denotes the residual block with BN.", + "bbox": [ + 75, + 731, + 423, + 746 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Low-frequency attention module. Low-frequency components focus more on global information, such as color distribution and illumination, which inevitably leads to inevitably existing redundant components and slight perturbations [43]. To handle those problems, we design a comprehensive normalization strategy to suppress the undesired artifacts and provide cleaner global information for attention calculation both at the instance level and channel dimension, which can highlight those abnormal regions from a global perspective. Specifically, this module takes the de", + "bbox": [ + 75, + 750, + 468, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "composed LF features $(f_k^r)_{LF}$ as the input and outputs", + "bbox": [ + 498, + 203, + 862, + 219 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\np _ {k} ^ {l} = J A \\left(P N \\left(R e s I N \\left(\\left(f _ {k} ^ {r}\\right) _ {L F}\\right)\\right)\\right), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 570, + 220, + 890, + 239 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $ResIN(\\cdot), PN(\\cdot)$ , and $JA(\\cdot)$ denote the instance normalization [44] constrained residual block, positional normalization [20], and joint attention, respectively.", + "bbox": [ + 496, + 239, + 890, + 286 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2.3 Guidance-based Feature Aggregation Module", + "text_level": 1, + "bbox": [ + 498, + 306, + 870, + 321 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As shown in Fig. 3, we propose a guidance-based feature aggregation (GFA) module to integrate the multi-scale decomposed features. Unlike existing heuristic-based feature aggregation strategies simply using concatenation [22, 24], GFA is specifically designed to address the key issue of COD, i.e., emphasizing the subtle discriminative features, by promoting inter-feature information interaction.", + "bbox": [ + 496, + 325, + 890, + 431 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Taking HF bands as an example, GFA generates the aggregated feature $\\{f_{k - 1}^h\\}_{k = 2}^4$ that combines deep semantic information of the low-resolution feature $(f_k^r)_{HF}$ (at a higher level) and the abundant spatial details of the highresolution feature $(f_{k - 1}^{r})_{HF}$ (at a lower level) with the guidance of the attention map $p_k^h$ . Therefore, the aggregated feature $f_{k - 1}^h$ can better highlight the subtle discriminative features. To extract the attention-guided semantic information, we first generate the down-sampled aggregated feature $f_{k - 1}^{dh}$ with the window-based linear model [23]:", + "bbox": [ + 496, + 433, + 892, + 585 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\left(f _ {k - 1} ^ {d h}\\right) _ {i} = \\sigma_ {w} d o w n \\left(\\left(f _ {k - 1} ^ {r}\\right) _ {H F}\\right) _ {i} + \\mu_ {w}, \\forall i \\in s _ {w}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 519, + 588, + 890, + 607 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $down(\\cdot), s_w$ , and $i$ are the down-sampling operation, local window, and pixel point $i$ . $\\{\\sigma_w, \\mu_w\\}$ are linear aggregation coefficients for the pixels in window $s_w$ , which can be acquired by optimizing the following objective function:", + "bbox": [ + 496, + 609, + 890, + 667 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {\\sigma_ {w}, \\mu_ {w}} \\sum_ {i \\in s _ {w}} \\left[ \\left(p _ {k} ^ {h}\\right) _ {i} ^ {2} \\left(\\left(f _ {k - 1} ^ {d h}\\right) _ {i} - \\left(\\left(f _ {k} ^ {r}\\right) _ {H F}\\right) _ {i}\\right) ^ {2} + \\epsilon \\sigma_ {w} ^ {2} \\right], \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 509, + 670, + 890, + 703 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\epsilon$ is a constraint value for $\\sigma_w$ . See Supplementary Material (Supp) for derivations and solutions of $\\{\\sigma_w, \\mu_w\\}$ .", + "bbox": [ + 496, + 705, + 890, + 734 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Considering pixel $i$ covered by multiple windows, we average those window-wise coefficients and get the specific aggregation coefficients $\\{\\sigma_i,\\mu_i\\}$ for pixel $i$ . By matrixing $\\{\\sigma_i,\\mu_i\\}$ into $\\{\\sigma_i,\\mu_i\\}$ , Eq. (4) can be rewritten as follows:", + "bbox": [ + 496, + 734, + 890, + 796 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nf _ {k - 1} ^ {d h} = \\sigma_ {i} \\odot d o w n \\left(\\left(f _ {k - 1} ^ {r}\\right) _ {H F}\\right) + \\boldsymbol {\\mu} _ {i}, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 566, + 797, + 890, + 814 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\odot$ is the Hadamard product. We then up-sample $\\{\\pmb {\\sigma}_i,\\pmb {\\mu}_i\\}$ as $\\{\\pmb {\\sigma}_h,\\pmb {\\mu}_h\\}$ and acquire the high-resolution ag-. ggregated feature $f_{k - 1}^{h}$ for enriching spatial details:", + "bbox": [ + 496, + 816, + 890, + 862 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} f _ {k - 1} ^ {h} = G F A \\left(\\left(f _ {k} ^ {r}\\right) _ {H F}, \\left(f _ {k - 1} ^ {r}\\right) _ {H F}, p _ {k} ^ {h}\\right), \\tag {7} \\\\ = \\sigma_ {h} \\odot \\left(f _ {k - 1} ^ {r}\\right) _ {H F} + \\boldsymbol {\\mu} _ {h}. \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 553, + 864, + 890, + 904 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "22049", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/866a33e0554a38702327cb1a77fe6198f9cf0428ebcc76212aff636e29623f83.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 80, + 89, + 263, + 167 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/51c635ff8eb5e70df58596dfe41a541b3c8582289e6dedbb546269f96fdde41e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 271, + 90, + 465, + 167 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/e906f30558141957c5ea9078f6ee271dbd79ad0f0545ea32cad7ca4d68cf224e.jpg", + "image_caption": [ + "Figure 4. Details of SED with RRS and OER modules. OER is an RK2-based module guaranteed with the Hamiltonian system." + ], + "image_footnote": [], + "bbox": [ + 80, + 172, + 261, + 252 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/8364ece33d2de787d7b9c493ef51c7e15fa6b70747e11a81ae71218ffa5a40e1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 271, + 169, + 467, + 252 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To iteratively acquire the aggregated features $\\left\\{f_{k - 1}^{h}\\right\\}_{k = 2}^{4}$ we redefine GFA module by replacing $(f_k^r)_HF$ with $f_{k}^{h}$ :", + "bbox": [ + 75, + 301, + 467, + 337 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nf _ {k - 1} ^ {h} = G F A \\left(f _ {k} ^ {h}, \\left(f _ {k - 1} ^ {r}\\right) _ {H F}, p _ {k} ^ {h}\\right), \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 150, + 337, + 467, + 357 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $p_k^h = JA(ResB(f_k^h))$ and $f_4^h = (f_4^r)_{HF}$ . Guaranteed with frequency-specific attention, our aggregated features can emphasize more discriminative features than others by combining abundant spatial details and deep semantic information, thus better catering to the COD task. The calculation of the aggregated LF features $f_{k - 1}^{l}$ are similar to $f_{k - 1}^{h}$ , which can be seen in Supp.", + "bbox": [ + 75, + 357, + 468, + 465 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Considering that bottom layers (at higher levels) focus more on HF details while top layers (at lower levels) care more about global information [38], we pass the aggregated HF/LF features into the bottom/top decoder layers along with the skip-connected encoded features $\\{f_k^r\\}_{k=1}^4$ . To balance performance and efficiency, the integrated features $\\{f_k^o\\}_{k=1}^4$ passed to the decoder are defined as:", + "bbox": [ + 75, + 465, + 468, + 571 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nf _ {1} ^ {o} = \\operatorname {c o n v 1} \\left(\\operatorname {c o n} \\left(f _ {1} ^ {r}, u p \\left(f _ {1} ^ {h}\\right)\\right)\\right), f _ {2} ^ {o} = f _ {2} ^ {r}, \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 125, + 574, + 468, + 598 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nf _ {3} ^ {o} = f _ {3} ^ {r}, f _ {4} ^ {o} = \\operatorname {c o n v 1} \\left(\\operatorname {c o n} \\left(f _ {4} ^ {r}, f _ {3} ^ {l}\\right)\\right),\n$$\n", + "text_format": "latex", + "bbox": [ + 127, + 595, + 380, + 613 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $up(\\cdot)$ and $con(\\cdot)$ denote the up-sampling operation and the concatenation operation. conv1 represents $1 \\times 1$ convolution, which is used for channel-level integration.", + "bbox": [ + 75, + 614, + 468, + 659 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Segmentation-oriented Edge-assisted Decoder", + "text_level": 1, + "bbox": [ + 76, + 670, + 465, + 686 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In the segmentation-oriented edge-assisted decoder (SED) $\\{D_k\\}_{k=1}^4$ , we propose to learn an auxiliary edge reconstruction task alongside the COD task to help generate precise segmentation maps with accurate boundaries. To be specific, as shown in Fig. 4, each decoder layer $D_k$ consists of a reversible re-calibration segmentation (RRS) module and an ODE-inspired edge reconstruction (OER) module.", + "bbox": [ + 75, + 695, + 468, + 800 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3.1 Reversible Re-calibration Segmentation Module", + "text_level": 1, + "bbox": [ + 76, + 819, + 465, + 834 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Due to complex camouflage, prediction maps inevitably have some ambiguous regions with low confidence, we adopt a reverse strategy to excavate cues from these low-confidence regions by reversing the attention, which", + "bbox": [ + 75, + 839, + 468, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "erases detected regions and amplifies the response in low-confidence regions, thus re-calibrating the misclassified regions. Specifically, we repeat the coarse segmentation map $\\{d_{k + 1}^s\\}_{k = 0}^3$ as a 64-dimension tensor, normalize it to [0, 1] with Sigmoid $S(\\cdot)$ , and reverse it by subtracting each element from 1. We then multiply the integrated feature $f_{k}^{o}$ with the reversed map and concatenate it with the edge feature $d_{k}^{fe}$ to obtain the segmentation result $d_{k}^{s}$ :", + "bbox": [ + 498, + 90, + 890, + 212 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nd _ {k} ^ {s} = \\operatorname {c o n v 3} \\left(\\operatorname {c o n} \\left(d _ {k} ^ {f e}, f _ {k} ^ {o} \\odot r v (S (r p \\left(d _ {k + 1} ^ {s}\\right))\\right)\\right), \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 501, + 213, + 890, + 239 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $rp(\\cdot)$ and $rv(\\cdot)$ denote repeat and reverse.", + "bbox": [ + 500, + 241, + 823, + 255 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3.2 ODE-inspired Edge Reconstruction Module", + "text_level": 1, + "bbox": [ + 498, + 277, + 857, + 292 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Existing methods tend to excavate edge information by incorporating certain priors within the residual network structure [13, 50]. However, either the intractable localization problem of the IS challenge or the ambiguous boundary problem of the ED challenge makes it difficult to design an appropriate edge prior for the COD task. In some cases, a biased prior can even reduce the segmentation performance.", + "bbox": [ + 496, + 299, + 890, + 405 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Therefore, instead of exploiting prior knowledge, we focus on proposing an edge-friendly network architecture, i.e., the ODE-inspired edge reconstruction (OER) module. Compared with the traditional residual network structure that can be seen as the first-order Euler discretization approximation of ODE with nonnegligible truncation errors [46], the proposed OER module employs a higher-order ODE solver, specifically, a second-order Runge-Kutta (RK2), to provide more accurate numerical solutions in edge information processing. This better accommodates the fine-grained property of edges and facilitates the complete edge reconstruction, thus addressing the ED challenge. To ensure the flexibility of our OER module, we replace the fixed trade-off parameter in the RK2 solver with a weighted gate mechanism $g_{w}$ with learnable coefficients. Given an input $e_t$ , where $e_t = conv1\\left(\\operatorname{con}\\left(f_k^o, rp\\left(d_{k+1}^s\\right)\\right)\\right)$ , the proposed OER module can be formulated as follows:", + "bbox": [ + 496, + 406, + 892, + 662 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\ne _ {t + 1} = e _ {t} + g _ {w} F _ {1} + \\left(1 - g _ {w}\\right) F _ {2},\n$$\n", + "text_format": "latex", + "bbox": [ + 562, + 665, + 792, + 681 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\ng _ {w} = S \\left(\\sigma_ {g} \\operatorname {c o n} (F 1, F 2) + \\mu_ {g}\\right), \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 563, + 684, + 890, + 700 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nF _ {1} = F \\left(e _ {t}, \\theta_ {t}\\right), F _ {2} = F \\left(e _ {t} + F _ {1}, \\theta_ {t}\\right),\n$$\n", + "text_format": "latex", + "bbox": [ + 563, + 703, + 825, + 718 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $e_{t + 1} = d_k^{fe}$ , $\\sigma_g$ and $\\mu_g$ are the learnable parameters in $g_w$ . $\\{F_i\\}_{i = 1}^2$ denotes the intermediate layers with shared parameters $\\theta_t$ for efficiency. Following [1], we set $F_i$ as a Conv-ReLU-Conv framework. To ensure the stability of OER, we apply the Hamiltonian system [9] to our OER module. By denoting Eq. (11) as $RK2(\\cdot)$ , the Hamiltonian-theory-guaranteed OER module is defined as follows:", + "bbox": [ + 496, + 722, + 890, + 828 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\ne _ {t + 1} = \\operatorname {c o n} \\left(e _ {t + 1} ^ {1}, e _ {t + 1} ^ {2}\\right), e _ {t + 1} ^ {1} = e _ {t} ^ {1} + R K 2 _ {1} \\left(e _ {t} ^ {2}\\right), \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 508, + 830, + 890, + 853 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\ne _ {t + 1} ^ {2} = e _ {t} ^ {2} - R K 2 _ {2} \\left(e _ {t} ^ {1} + R K 2 _ {1} \\left(e _ {t} ^ {2}\\right)\\right),\n$$\n", + "text_format": "latex", + "bbox": [ + 511, + 849, + 777, + 868 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $e_t$ is split to $e_t^1$ and $e_t^2$ in channel-wise. Note that the OER module in Eq. (12) is a reversible and sta", + "bbox": [ + 496, + 869, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "22050", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/313bb79684d6aa130b56af1e289bed344d0d630ad4a7b5fc7a68e463be8f6739.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodsPublicationsBackbonesCHAMELEON (76 images)CAMO (250 images)COD10K (2,026 images)NC4K (4,121 images)
M↓Fβ↑Eφ↑Sα↑M↓Fβ↑Eφ↑Sα↑M↓Fβ↑Eφ↑Sα↑
Common Setting: Single Input Scale and Single Stage
CPD [47]CVPR19ResNet500.0480.7750.8740.8570.1130.6750.7230.7160.0530.5780.7760.750
SINet [6]CVPR20ResNet500.0340.8230.9360.8720.0920.7120.8040.7450.0430.6670.8640.776
PFNet [28]CVPR21ResNet500.0330.8200.9310.8820.0850.7510.8410.7820.0400.6760.8770.800
MGL-R [50]CVPR21ResNet500.0310.8250.9170.8910.0880.7380.8120.7750.0350.6800.8510.814
LSR [25]CVPR21ResNet500.0300.8350.9350.8900.0800.7560.8380.7870.0370.6990.8800.804
UGTR [48]ICCV21ResNet500.0310.8050.9100.8880.0860.7470.8210.7840.0360.6700.8520.817
SLT-Net [2]CVPR22ResNet500.0300.8350.9400.8870.0820.7630.8480.7920.0360.6810.8750.804
SegMaR-1 [14]CVPR22ResNet500.0280.8280.9440.8920.0720.7720.8610.8050.0350.6990.8900.813
OSFormer [32]ECCV22ResNet500.0280.8360.9390.8910.0730.7670.8580.7990.0340.7010.8810.811
FEDER-R50—ResNet500.0280.8550.9470.8940.0690.7850.8730.8070.0320.7400.9000.823
SINet V2 [4]TPAMI22ResNet500.0300.8160.9420.8880.0700.7790.8820.8220.0370.6820.8870.815
BSA-Net [52]AAAI22ResNet500.0270.8510.9460.8950.0790.7680.8510.7960.0340.7230.8910.818
FEDER-R2N—ResNet500.0260.8560.9470.9030.0660.8070.8970.8360.0290.7480.9110.844
Other Setting: Multiple Input Scales (MIS)
ZoomNet [31]CVPR22ResNet500.0240.8580.9430.9020.0660.7920.8770.8200.0290.7400.8880.838
FEDER-MIS—ResNet500.0230.8690.9590.9060.0640.8010.8930.8270.0280.7560.9130.837
Other Setting: Multiple Stages (MS)
SegMaR-4 [14]CVPR22ResNet500.0250.8550.9550.9060.0710.7790.8650.8150.0330.7370.8960.833
FEDER-MS-4—ResNet500.0250.8740.9640.9070.0670.8090.8860.8220.0280.7520.9170.851
", + "bbox": [ + 83, + 89, + 893, + 362 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1. Quantitative comparisons of the proposed FEDER and other state-of-the-art methods on four benchmarks. SegMaR-1 and SegMaR-4 denote SegMaR at one stage and four stages. R50 and R2N indicate ResNet50 and Res2Net50. The best results are marked in bold. For ResNet50 backbone in the common setting, the best two results are in red and blue fonts.", + "bbox": [ + 75, + 363, + 892, + 405 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "ble block [9], which further promotes the edge reconstruction performance. In this case, the final edge predictions $\\{d_k^e\\}_{k = 1}^4$ can be acquired in the following manner:", + "bbox": [ + 76, + 412, + 468, + 459 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nd _ {k} ^ {e} = \\operatorname {c o n v 3} \\left(e _ {t + 1}\\right) = \\operatorname {c o n v 3} \\left(d _ {k} ^ {f e}\\right). \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 148, + 460, + 468, + 486 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.4. Loss Functions", + "text_level": 1, + "bbox": [ + 76, + 494, + 230, + 508 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The loss function of the proposed FEDER consists of two kinds of supervisions, namely the segmentation mask $GT_{s}$ and edge $GT_{e}$ of the camouflaged object, which correspond to the multi-scale segmentation maps $\\{d_k^s\\}_{k = 1}^5$ and the multi-scale object edges $\\{d_k^e\\}_{k = 1}^4$ . Following [4], we employ the weighted binary cross-entropy loss $L_{BCE}^{w}$ and weighted intersection-over-union loss $L_{IoU}^{w}$ for segmentation supervision, which focuses more on those hard pixels. For edge supervision, we use the dice loss $L_{dice}$ [29] to overcome the extreme imbalance between the positive and negative samples. Furthermore, to handle the multi-scale outputs, we up-sampling all the outputs to match the size of their corresponding ground truths during training. Therefore, the total loss of our FEDER is formulated as follows:", + "bbox": [ + 75, + 518, + 468, + 728 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} L _ {t} = \\sum_ {k = 1} ^ {5} \\frac {1}{2 ^ {k - 1}} \\left(L _ {B C E} ^ {w} \\left(d _ {k} ^ {s}, G T _ {s}\\right) + L _ {I o U} ^ {w} \\left(d _ {k} ^ {s}, G T _ {S}\\right)\\right) \\tag {14} \\\\ + \\sum_ {k = 1} ^ {4} \\frac {1}{2 ^ {k - 1}} L _ {d i c e} \\left(d _ {k} ^ {e}, G T _ {e}\\right). \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 73, + 729, + 468, + 813 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiment", + "text_level": 1, + "bbox": [ + 76, + 820, + 200, + 837 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Experimental Setup", + "text_level": 1, + "bbox": [ + 76, + 845, + 267, + 863 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Datasets. We used four widely-used COD datasets for evaluation, including CHAMELEON [39], CAMO [18],", + "bbox": [ + 75, + 869, + 468, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "COD10K [4], and NC4K [25]. CHAMELEON contains 76 hand-annotated images. CAMO has 1,250 camouflaged images with 1,000 training images and 250 testing images. Currently, COD10K is the largest COD benchmark, with 3,040 training images and 2,026 testing images. NC4K is a large-scale COD testing dataset, comprising 4,121 images. Following [4], we form the training set with 3,040 images from COD10K and 1,000 images from CAMO, while the remaining camouflaged images are used for testing.", + "bbox": [ + 496, + 412, + 890, + 547 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation metrics. Four commonly-used metrics are employed for COD task, including mean absolute error $M$ , adaptive F-measure $F_{\\beta}$ [27], mean E-measure $E_{\\phi}$ [5], and structure measure $S_{\\alpha}$ [3]. Larger $F_{\\beta}$ , $E_{\\phi}$ , $S_{\\alpha}$ , and smaller $M$ indicate better segmentation performance.", + "bbox": [ + 496, + 549, + 890, + 625 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation details. The proposed FEDER is implemented in PyTorch on two RTX3090TI GPUs and is optimized by the Adam with momentum terms (0.9, 0.999). Following the common setting [4, 6], our encoder (ResNet50 by default) is initialized with the model pre-trained on ImageNet [6]. In the training phase, the batch size is set to 36 and the learning rate is initialized to 0.0001, dividing by 10 every 80 epochs. For both training and inference phases, all images are resized as $384 \\times 384$ .", + "bbox": [ + 496, + 625, + 890, + 760 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Comparison with the State-of-the-arts", + "text_level": 1, + "bbox": [ + 498, + 771, + 828, + 787 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Quantitative analysis. We compare the proposed FEDER with 12 cutting-edge techniques in three different settings, including the common setting (single input scale and single stage) and two other settings (multiple input scales (MIS) and multiple stages (MS)). In the MIS and MS settings, the proposed FEDER follows the practices of ZoomNet [31] and SegMaR [14], where FEDER-MS-4 means", + "bbox": [ + 496, + 794, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "22051", + "bbox": [ + 478, + 944, + 517, + 957 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/368ed6bce6fca44bf494230f7a13b13d55ef9c0209b9ae769f594f1212645419.jpg", + "image_caption": [ + "Figure 5. Visual comparisons of the proposed FEDER and other nine state-of-the-art methods. Our method can generate more accurate prediction maps with clearer boundaries than other methods. Please zoom in for more details." + ], + "image_footnote": [], + "bbox": [ + 81, + 89, + 890, + 306 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b039425d32da85e97fb5cfb7a542cad0ae8a695d7c926159c176cc6ea880885d.jpg", + "image_caption": [ + "Figure 6. Performance-Params-FLOPs comparisons of some state-of-the-art deep learning-based COD methods on COD10K [4]." + ], + "image_footnote": [], + "bbox": [ + 78, + 361, + 467, + 569 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "complete COD task with our FEDER at four stages following SegMaR-4. For a fair comparison, prediction maps of the above techniques are directly segmented by their provided models with no modifications. Besides, all prediction maps are evaluated with the same code. As shown in Tab. 1, our method achieves the best results in all settings and backbones, which comprehensively demonstrates the superiority of our FEDER. Notably, $COD10K$ and $NC4K$ are the two most challenging datasets in terms of the number of images and segmentation difficulty. In the common setting, the proposed FEDER surpasses the second best results $3.6\\%$ on average over all metrics on $COD10K$ dataset and $2.6\\%$ on average on $NC4K$ dataset. Such a big margin further confirms the effectiveness of the proposed method, both with the deep wavelet-like feature decomposition strategy and the ODE-inspired edge reconstruction module.", + "bbox": [ + 75, + 625, + 467, + 864 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Qualitative analysis. Fig. 5 presents a visual comparison of our FEDER and other SOTAs. We select various typi", + "bbox": [ + 75, + 869, + 467, + 900 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "cal and challenging camouflaged images and arrange them in order of the camouflaged object size, from smallest to largest. Note that most of these images suffer from the IS or ED challenge, which can confuse existing COD techniques, resulting in mislocalization, ambiguous boundaries, etc. In contrast to those methods, our FEDER can overcome such challenges and generate more competitive prediction maps in the following aspects: (a) more accurate localization of small objects. For those small objects under the IS challenge, precise localization is a significant problem due to subtle differences and can confuse most existing methods. Thanks to our HFA, LFA, and GFA modules, our FEDER can emphasize the inconspicuous discriminative features and thus ensure more accurate camouflaged object localization (in Rows 1 and 2). (b) clearer edges on large objects. For those large objects, our prediction maps can achieve much clearer boundaries than others (see Rows 7 and 8), which mainly attributes to our joint training strategy of edge and segmentation and our edge-friendly OER module. (c) stronger suppression of redundant information. In the IS challenge and degraded imaging scenarios, the detection performance can be inevitably influenced by redundant information, such as background noise. However, the proposed deep decomposition strategy can suppress the redundant information by filtering out those components with the most discriminative information, namely, the HF and LF components. Thus, as depicted in Rows 3 and 4 in Fig. 5, FEDER can generate more accurate prediction maps.", + "bbox": [ + 496, + 363, + 890, + 786 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Efficiency analysis. We compare the performance, parameters, and FLOPs with other SOTAs on COD10K [4] in Fig. 6. As presented in Fig. 6, our proposed FEDER achieves the smallest FLOPs and parameters compared with other state-of-the-art deep learning-based COD techniques. Furthermore, our score in $F_{\\beta}$ is much higher than other methods and surpasses the second best one in $5.6\\%$ .", + "bbox": [ + 496, + 794, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "22052", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/6b379310846587918299ab0bae1c54d6b23582215e138cf61e3db7612dc4f4fa.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodsCOD10K (2026 images)NC4K (4121 images)
M↓Fβ↑Eφ↑Sα↑M↓Fβ↑Eφ↑Sα↑
(a)w/o DWD0.0350.6970.8610.7940.0490.7760.8730.817
(b)FW->LW0.0320.7310.8950.8220.0460.8110.8990.844
(c)w/o HFA0.0330.7280.8950.8170.0470.8050.8910.840
(d)w/o LFA0.0330.7240.8870.8160.0470.8030.8920.841
(e)UFA->GFA0.0330.7200.8940.8100.0470.8040.8900.838
(f)FEDER0.0320.7400.9000.8230.0450.8170.9050.846
", + "bbox": [ + 89, + 89, + 483, + 188 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/78b13002d9327170e05a7c5090799ef11d76d04b338afdc7d2df98140cbe72b4.jpg", + "table_caption": [ + "(a) Ablation study of DWD Component." + ], + "table_footnote": [], + "table_body": "
MethodsCOD10K (2026 images)NC4K (4121 images)
M↓Fβ↑Eφ↑Sα↑M↓Fβ↑Eφ↑Sα↑
(a)w/o OER0.0340.7140.8710.8040.0480.7800.8890.834
(b)FC->WGM0.0320.7310.8990.8170.0460.8140.8950.842
(c)w/o HS0.0320.7230.8850.8210.0460.7980.8920.840
(d)RB->RK20.0330.7220.8810.8110.0470.8020.8950.843
(e)RK4->RK20.0310.7420.9050.8290.0450.8160.9060.848
(f)FEDER0.0320.7400.9000.8230.0450.8170.9050.846
", + "bbox": [ + 490, + 89, + 885, + 188 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "(b) Ablation study of OER module.", + "bbox": [ + 589, + 190, + 777, + 202 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 2. Ablation study on $COD10k$ [4] and $NC4K$ [25], where w/o means without. (a) FW, LW, and UFA are short for fixed wavelet [40], learnable wavelet, and upsampling-based feature aggregation [51]. (b) FC, WGM, HS, RB, and RK4 are short for fixed coefficient, weighted gate mechanism, Hamiltonian system, residual block, and forth-order Runge-Kutta. The best results are marked in bold.", + "bbox": [ + 75, + 203, + 893, + 244 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/40688237d6af8db53c8358514051cbdb71c4b8ec2673fc9a246fcde553a3b2eb.jpg", + "image_caption": [ + "Figure 7. Visual verification of the effectiveness of the proposed components. B, D, and O denote baseline, DWD, and OER. Note that $\\mathrm{B + D + O}$ is our proposed FEDER." + ], + "image_footnote": [], + "bbox": [ + 81, + 262, + 465, + 468 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.3. Ablation Study", + "text_level": 1, + "bbox": [ + 76, + 530, + 230, + 546 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We conduct the ablation study on the two largest datasets, namely $COD10K$ and $NC4K$ .", + "bbox": [ + 75, + 553, + 468, + 583 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effect of DWD component. We demonstrate the effectiveness of our DWD component both in visual verification (see Fig. 7) and quantitative analysis (see Tab. 2a). In Fig. 7, the networks with DWD component (Rows 3 and 5) exhibit more accurate localizations and stronger redundant information suppression capacities. In addition, we present more detailed information in Tab. 2a to verify the validity of each part in the DWD component. As presented in Tab. 2a, we prove the superiority of the DWD component (in (a)), learnable wavelet-like decomposition strategy (in (b)), HFA/LFA module (in (c) and (d)), and GFA module (in (e)).", + "bbox": [ + 75, + 583, + 468, + 750 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effect of OER module. The efficacy of our OER module is demonstrated visually by Figs. 7 and 8. In Fig. 7, methods with the OER module can generate the prediction maps with clearer edges. Besides, Fig. 8 illustrates the advancement of our OER module in generating accurate and clear boundary information. We provide detailed information about the superiority of our OER module in Tab. 2b. Specifically, (b) and (c) verify the effectiveness of our weighted gate mechanism (learnable coefficient) and Hamiltonian system. We further compare the performance of the OER module with", + "bbox": [ + 75, + 750, + 470, + 901 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/3f0873bdd35e953d7658e995808541f94246ae4772f04327e94e2abc3cdf0dfd.jpg", + "image_caption": [ + "Figure 8. Visualizations of the reconstructed edge from the OER module, where the edges are marked in red for well-read." + ], + "image_footnote": [], + "bbox": [ + 504, + 263, + 887, + 416 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "different Runge-Kutta methods, i.e., RB (RK1), RK2, and RK4. Notably, the COD results with RK2 significantly outperform those with RB and are slightly lower than that with RK4. Therefore, we integrate RK2 into our OER module for a balance of performance and efficiency.", + "bbox": [ + 498, + 464, + 892, + 541 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusions", + "text_level": 1, + "bbox": [ + 500, + 554, + 627, + 569 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To address the IS and ED challenges, in this paper, we propose the FEDER model for COD. Specifically, we decompose the features into different frequency bands with learnable wavelets and filter out the most informative bands to excavate the subtle discriminative features with the HFA, LFA, and GFA modules, thereby solving the IS challenge. Besides, we propose to learn an auxiliary edge reconstruction task with our OER module to generate complete edges. Learning this auxiliary task along with the COD task thus facilitates the generation of precise segmentation results with accurate object boundaries, thus mitigating the ED challenge. Extensive experiments verify the superiority of our FEDER model in comparison with other SOTAs.", + "bbox": [ + 496, + 579, + 892, + 775 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements: This research is partly supported by the National Key R&D Program of China (Grant No. 2020AAA0108303), and Shenzhen Science and Technology Project (Grant No. JCYJ20200109143041798) & Shenzhen Stable Supporting Program (WDZC20200820200655001) & Shenzhen Key Laboratory of next generation interactive media innovative technology (Grant No. ZDSYS20210623092001004).", + "bbox": [ + 496, + 776, + 893, + 896 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "22053", + "bbox": [ + 478, + 944, + 517, + 957 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 174, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Bo Chang, Lili Meng, Eldad Haber, Lars Ruthotto, David Begert, and Elliot Holtham. Reversible architectures for arbitrarily deep residual neural networks. In AAAI, 2018. 3, 5", + "[2] Xuelian Cheng, Huan Xiong, Deng-Ping Fan, Yiran Zhong, Mehrtash Harandi, Tom Drummond, and Zongyuan Ge. Implicit motion handling for video camouflaged object detection. In CVPR, pages 13864-13873, 2022. 6", + "[3] Deng-Ping Fan, Ming-Ming Cheng, Yun Liu, Tao Li, and Ali Borji. Structure-measure: A new way to evaluate foreground maps. In ICCV, pages 4548-4557, 2017. 6", + "[4] Deng-Ping Fan, Ge-Peng Ji, Ming-Ming Cheng, and Ling Shao. Concealed object detection. IEEE Trans. Pattern Anal. Mach. Intell., 2021. 3, 6, 7, 8", + "[5] Deng-Ping Fan, Ge-Peng Ji, Xuebin Qin, and Ming-Ming Cheng. Cognitive vision inspired object segmentation metric and loss function. Sci. Sin. Inf., 6:6, 2021. 6", + "[6] Deng-Ping Fan, Ge-Peng Ji, Guolei Sun, Ming-Ming Cheng, Jianbing Shen, and Ling Shao. Camouflaged object detection. In CVPR, pages 2777-2787, 2020. 2, 3, 6", + "[7] Shang-Hua Gao, Ming-Ming Cheng, Kai Zhao, Xin-Yu Zhang, Ming-Hsuan Yang, and Philip Torr. Res2net: A new multi-scale backbone architecture. IEEE Trans. Pattern Anal. Mach. Intell., 43(2):652-662, 2019. 3", + "[8] Wooseok Ha, Chandan Singh, Francois Lanusse, Srigokul Upadhyayula, and Bin Yu. Adaptive wavelet distillation from neural networks through interpretations. NIPS, 34:20669-20682, 2021. 3, 4", + "[9] Eldad Haber and Lars Ruthotto. Stable architectures for deep neural networks. Inverse Probl, 34(1):014004, 2017. 5, 6", + "[10] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, pages 770-778, 2016. 3", + "[11] Jianqin Yin Yanbin Han Wendi Hou and Jinping Li. Detection of the mobile object with camouflage color under dynamic background based on optical flow. Procedia Eng., 15:2201-2205, 2011. 2", + "[12] Sergey Ioffe and Christian Szegedy. Batch normalization: Accelerating deep network training by reducing internal covariate shift. In ICML, pages 448-456. PMLR, 2015. 4", + "[13] Ge-Peng Ji, Lei Zhu, Mingchen Zhuge, and Keren Fu. Fast camouflaged object detection via edge-based reversible recalibration network. Pattern Recognit., 123:108414, 2022. 5", + "[14] Qi Jia, Shuilian Yao, Yu Liu, Xin Fan, Risheng Liu, and Zhongxuan Luo. Segment, magnify and reiterate: Detecting camouflaged objects the hard way. In CVPR, pages 4713-4722, 2022. 1, 2, 6", + "[15] Liming Jiang, Bo Dai, Wayne Wu, and Chen Change Loy. Focal frequency loss for image reconstruction and synthesis. In ICCV, pages 13919-13929, 2021. 2", + "[16] Zhanghan Ke, Jiayu Sun, Kaican Li, Qiong Yan, and Rynson WH Lau. Modnet: Real-time trimap-free portrait matting via objective decomposition. In AAAI, pages 1140-1147, 2022. 3", + "[17] Ajay Kumar. Computer-vision-based fabric defect detection: A survey. IEEE Trans. Ind. Electron., 55(1):348-363, 2008. 1" + ], + "bbox": [ + 78, + 114, + 468, + 914 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[18] Trung-Nghia Le, Tam V Nguyen, Zhongliang Nie, Minh-Triet Tran, and Akihiro Sugimoto. Anabranch network for camouflaged object segmentation. Comput. Vis. Image Underst., 184:45-56, 2019. 6", + "[19] Bei Li, Quan Du, Tao Zhou, Yi Jing, Shuhan Zhou, Xin Zeng, Tong Xiao, Jingbo Zhu, Xuebo Liu, and Min Zhang. Ode transformer: An ordinary differential equation-inspired model for sequence generation. In ACL, pages 8335-8351, 2022. 3", + "[20] Boyi Li, Felix Wu, Kilian Q Weinberger, and Serge Belongie. Positional normalization. NIPS, 32, 2019. 4", + "[21] Juncheng Li, Faming Fang, Kangfu Mei, and Guixu Zhang. Multi-scale residual network for image super-resolution. In ECCV, pages 517-532, 2018. 4", + "[22] Kai Li, Yulun Zhang, Kunpeng Li, and Yun Fu. Adversarial feature hallucination networks for few-shot learning. In CVPR, pages 13470-13479, 2020. 4", + "[23] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In ICCV, pages 10012-10022, 2021. 4", + "[24] Laetitia Loncan, Luis B De Almeida, José M Bioucas-Dias, Xavier Briottet, Jocelyn Chanussot, Nicolas Dobigeon, Sophie Fabre, Wenzhi Liao, Giorgio A Licciardi, Miguel Simoes, et al. Hyperspectral pansharpening: A review. IEEE Geosci. Remote Sens. Mag., 3(3):27-46, 2015. 4", + "[25] Yunqiu Lv, Jing Zhang, Yuchao Dai, Aixuan Li, Bowen Liu, Nick Barnes, and Deng-Ping Fan. Simultaneously localize, segment and rank the camouflaged objects. In CVPR, pages 11591-11601, 2021. 6, 8", + "[26] Yue Ma, Yali Wang, Yue Wu, Ziyu Lyu, Siran Chen, Xiu Li, and Yu Qiao. Visual knowledge graph for human action reasoning in videos. In ACM MM, pages 4132-4141, 2022. 4", + "[27] Ran Margolin, Lihi Zelnik-Manor, and Ayellet Tal. How to evaluate foreground maps? In CVPR, pages 248-255, 2014. 6", + "[28] Haiyang Mei, Ge-Peng Ji, Ziqi Wei, Xin Yang, Xiaopeng Wei, and Deng-Ping Fan. Camouflaged object segmentation with distraction mining. In CVPR, pages 8772-8781, 2021. 2, 6", + "[29] Fausto Milletari, Nassir Navab, and Seyed-Ahmad Ahmadi. V-net: Fully convolutional neural networks for volumetric medical image segmentation. In 3DV, pages 565-571. IEEE, 2016. 6", + "[30] Yuxin Pan, Yiwang Chen, Qiang Fu, Ping Zhang, and Xin Xu. Study on the camouflaged target detection method based on 3d convexity. Mod. Appl. Sci., 5(4):152, 2011. 2", + "[31] Youwei Pang, Xiaoqi Zhao, Tian-Zhu Xiang, Lihe Zhang, and Huchuan Lu. Zoom in and out: A mixed-scale triplet network for camouflaged object detection. In CVPR, pages 2160-2170, 2022. 6", + "[32] Jialun Pei, Tianyang Cheng, Deng-Ping Fan, He Tang, Chuanbo Chen, and Luc Van Gool. Osformer: One-stage camouflaged instance segmentation with transformers. In ECCV, pages 19-37. Springer, 2022. 6" + ], + "bbox": [ + 501, + 92, + 892, + 885 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "22054", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[33] Natasha Price, Samuel Green, Jolyon Troscianko, Tom Tre-genza, and Martin Stevens. Background matching and disruptive coloration as habitat-specific strategies for camouflage. Sci. Rep., 9(1):1-10, 2019. 1", + "[34] Xu Qin, Zhilin Wang, Yuanchao Bai, Xiaodong Xie, and Huizhu Jia. Ffa-net: Feature fusion attention network for single image dehazing. In AAAI, pages 11908-11915, 2020. 4", + "[35] Dan Jeric Arcega Rustia, Chien Erh Lin, Jui-Yung Chung, Yi-Ji Zhuang, Ju-Chun Hsu, and Ta-Te Lin. Application of an image and environmental sensor network for automated greenhouse insect pest monitoring. J. Asia Pac. Entomol., 23(1):17-28, 2020. 1", + "[36] Phil Sallee and Bruno Olshausen. Learning sparse multiscale image representations. NIPS, 15, 2002. 4", + "[37] Jiawei Shen, Zhuoyan Li, Lei Yu, Gui-Song Xia, and Wen Yang. Implicit euler ode networks for single-image dehazing. In CVPRW, pages 218-219, 2020. 3", + "[38] Chenyang Si, Weihao Yu, Pan Zhou, Yichen Zhou, Xinchao Wang, and Shuicheng Yan. Inception transformer. arXiv preprint arXiv:2205.12956, 2022. 5", + "[39] Przemysław Skurowski, Hassan Abdulameer, J Błaszczyk, Tomasz Depta, Adam Kornacki, and P Koziel. Animal camouflage analysis: Chameleon database. Unpublished manuscript, 2(6):7, 2018. 6", + "[40] Radomir S Stanković and Bogdan J Falkowski. The haar wavelet transform: its status and achievements. Comput. Electr. Eng., 29(1):25-44, 2003. 4, 8", + "[41] Martin Stevens and Sami Merilaita. Animal camouflage: current issues and new perspectives. Philos. Trans. R. Soc. Lond., B, Biol. Sci., 364(1516):423-427, 2009. 1, 2, 3", + "[42] Cheng Tai and E Weinan. Multiscale adaptive representation of signals: I. the basic framework. J. Mach. Learn. Res., 17(1):4875-4912, 2016. 4" + ], + "bbox": [ + 78, + 90, + 470, + 573 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[43] Christopher Torrence and Gilbert P Compo. A practical guide to wavelet analysis. Bull. Am. Meteorol. Soc., 79(1):61-78, 1998. 4", + "[44] Dmitry Ulyanov, Andrea Vedaldi, and Victor Lempitsky. Instance normalization: The missing ingredient for fast stylization. arXiv preprint arXiv:1607.08022, 2016. 4", + "[45] Qilong Wang, Banggu Wu, Pengfei Zhu, Peihua Li, Wangmeng Zuo, and Qinghua Hu. Eca-net: Efficient channel attention for deep convolutional neural networks. In CVPR, pages 13-19, 2020. 4", + "[46] E Weinan. A proposal on machine learning via dynamical systems. Commun. Math. Stat., 1(5):1-11, 2017. 3, 5", + "[47] Zhe Wu, Li Su, and Qingming Huang. Cascaded partial decoder for fast and accurate salient object detection. In CVPR, pages 3907-3916, 2019. 6", + "[48] Fan Yang, Qiang Zhai, Xin Li, Rui Huang, Ao Luo, Hong Cheng, and Deng-Ping Fan. Uncertainty-guided transformer reasoning for camouflaged object detection. In ICCV, pages 4146-4155, 2021. 6", + "[49] Jaejun Yoo, Youngjung Uh, Sanghyuk Chun, Byeongkyu Kang, and Jung-Woo Ha. Photorealistic style transfer via wavelet transforms. In ICCV, pages 9036-9045, 2019. 3", + "[50] Qiang Zhai, Xin Li, Fan Yang, Chenglizhao Chen, Hong Cheng, and Deng-Ping Fan. Mutual graph learning for camouflaged object detection. In CVPR, pages 12997-13007, 2021. 5, 6", + "[51] Yulun Zhang, Zhifei Zhang, Stephen DiVerdi, Zhaowen Wang, Jose Echevarria, and Yun Fu. Texture hallucination for large-factor painting super-resolution. In ECCV, pages 209-225. Springer, 2020. 8", + "[52] Hongwei Zhu, Peng Li, Haoran Xie, Xuefeng Yan, Dong Liang, Dapeng Chen, Mingqiang Wei, and Jing Qin. I can find you! boundary-guided separated attention network for camouflaged object detection. In AAAI, 2022. 6" + ], + "bbox": [ + 501, + 92, + 890, + 571 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "22055", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2023/Camouflaged Object Detection With Feature Decomposition and Edge Reconstruction/69e6313d-2f4d-49ec-aeb1-e472deb81070_model.json b/2023/Camouflaged Object Detection With Feature Decomposition and Edge Reconstruction/69e6313d-2f4d-49ec-aeb1-e472deb81070_model.json new file mode 100644 index 0000000000000000000000000000000000000000..9771765aecb3e1ca7c5511d568c969709f49ee82 --- /dev/null +++ b/2023/Camouflaged Object Detection With Feature Decomposition and Edge Reconstruction/69e6313d-2f4d-49ec-aeb1-e472deb81070_model.json @@ -0,0 +1,2310 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.165, + 0.131, + 0.806, + 0.175 + ], + "angle": 0, + "content": "Camouflaged Object Detection with Feature Decomposition and Edge Reconstruction" + }, + { + "type": "text", + "bbox": [ + 0.087, + 0.203, + 0.892, + 0.257 + ], + "angle": 0, + "content": "Chunming He\\(^{1}\\), Kai Li\\(^{2*}\\), Yachao Zhang\\(^{1}\\), Longxiang Tang\\(^{1}\\), Yulun Zhang\\(^{3}\\), Zhenhua Guo\\(^{4}\\), Xiu Li\\(^{1*}\\) \n\\(^{1}\\)Shenzhen International Graduate School, Tsinghua University, \\(^{2}\\)NEC Laboratories America, \\(^{3}\\)ETH Zürich, \\(^{4}\\)Tianyi Traffic Technology" + }, + { + "type": "text", + "bbox": [ + 0.195, + 0.26, + 0.778, + 0.275 + ], + "angle": 0, + "content": "{chunminghe19990224,li.gml.kai,lloong.x,yulun100,cszguo}@gmail.com" + }, + { + "type": "text", + "bbox": [ + 0.251, + 0.278, + 0.714, + 0.292 + ], + "angle": 0, + "content": "yachaozhang@stu.xmu.edu.cn, li.xiu@sz.tsinghua.edu.cn" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.327, + 0.313, + 0.343 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.359, + 0.474, + 0.797 + ], + "angle": 0, + "content": "Camouflaged object detection (COD) aims to address the tough issue of identifying camouflaged objects visually blended into the surrounding backgrounds. COD is a challenging task due to the intrinsic similarity of camouflaged objects with the background, as well as their ambiguous boundaries. Existing approaches to this problem have developed various techniques to mimic the human visual system. Albeit effective in many cases, these methods still struggle when camouflaged objects are so deceptive to the vision system. In this paper, we propose the FEature Decomposition and Edge Reconstruction (FEDER) model for COD. The FEDER model addresses the intrinsic similarity of foreground and background by decomposing the features into different frequency bands using learnable wavelets. It then focuses on the most informative bands to mine subtle cues that differentiate foreground and background. To achieve this, a frequency attention module and a guidance-based feature aggregation module are developed. To combat the ambiguous boundary problem, we propose to learn an auxiliary edge reconstruction task alongside the COD task. We design an ordinary differential equation-inspired edge reconstruction module that generates exact edges. By learning the auxiliary task in conjunction with the COD task, the FEDER model can generate precise prediction maps with accurate object boundaries. Experiments show that our FEDER model significantly outperforms state-of-the-art methods with cheaper computational and memory costs. The code will be available at https://github.com/ChunmingHe/FEDER." + }, + { + "type": "title", + "bbox": [ + 0.079, + 0.824, + 0.208, + 0.839 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.849, + 0.47, + 0.88 + ], + "angle": 0, + "content": "Camouflaged object detection (COD) aims to detect and segment objects \"seamlessly\" integrated into surrounding" + }, + { + "type": "image", + "bbox": [ + 0.507, + 0.328, + 0.887, + 0.571 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.55, + 0.571, + 0.604, + 0.583 + ], + "angle": 0, + "content": "(a) Origin" + }, + { + "type": "image_caption", + "bbox": [ + 0.646, + 0.571, + 0.683, + 0.581 + ], + "angle": 0, + "content": "(b) GT" + }, + { + "type": "image_caption", + "bbox": [ + 0.719, + 0.571, + 0.786, + 0.581 + ], + "angle": 0, + "content": "(c) SegMaR" + }, + { + "type": "image_caption", + "bbox": [ + 0.808, + 0.571, + 0.871, + 0.581 + ], + "angle": 0, + "content": "(d) FEDER" + }, + { + "type": "image_caption", + "bbox": [ + 0.5, + 0.585, + 0.892, + 0.628 + ], + "angle": 0, + "content": "Figure 1. Results of SegMaR [14] and our method under the intrinsic similarity (IS) and edge disruption (ED) challenges. Our method better localizes the objects and produces clearer edges." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.644, + 0.892, + 0.78 + ], + "angle": 0, + "content": "environments. COD is a challenging task as it needs to combat against excellent camouflage strategies, including background matching [41], disruptive coloration [33], etc., and distinguish the subtle differences between candidate objects and their backgrounds. Research in COD can simultaneously facilitate the development of visual perception for nuance discrimination and promote various valuable real-life applications, ranging from concealed defect detection [17] in industry to pest monitoring [35] in agriculture." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.78, + 0.892, + 0.901 + ], + "angle": 0, + "content": "COD faces two main challenges. The first is the intrinsic similarity (IS) challenge, which occurs when camouflaged objects share similar colors and patterns with their backgrounds. This makes it difficult to even roughly localize those camouflaged objects. The second is the edge disruption (ED) challenge, which arises from extremely ambiguous object boundaries. Even if a rough localization is achieved, precise segmentation can barely be obtained." + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.887, + 0.222, + 0.9 + ], + "angle": 0, + "content": "*Corresponding author." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "22046" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.318 + ], + "angle": 0, + "content": "To tackle these challenges, most existing works aim to develop models that mimic the human visual system [6,28]. However, since camouflage strategies are designed by prey to confuse the predator's visual system, and the intrinsic topological properties of candidate objects are not distinctive, such human perception-oriented attempts may struggle to identify subtle discriminative features and fail to effectively address the above challenges. For instance, as illustrated in Fig. 1, the state-of-the-art human perception-based COD method can only generate inaccurate prediction maps, such as the vague caddisfly and incomplete dog (Row 2 and 4), or even fail to detect camouflaged objects like the bird and snake (Row 1 and 3). Therefore, a better COD method should compensate for the \"flaw\" in human perception by emphasizing subtle discriminative features." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.319, + 0.473, + 0.456 + ], + "angle": 0, + "content": "Based on the biological study [41], camouflaged objects often employ various camouflage strategies to conceal their discriminative differences, which mainly exist in texture details and global information distribution, within surrounding environments. Such a study inspires us to cope with the COD task by decomposing the camouflage scenario into different parts. This allows for the disentanglement of various intricate connections, enabling each part to be separately handled to fully excavate subtle discriminative cues." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.456, + 0.473, + 0.698 + ], + "angle": 0, + "content": "With this inspiration, we propose the FEature Decomposition and Edge Reconstruction (FEDER) model for the COD task, which compensates for the deficiencies of human perception by emphasizing subtle discriminative features and effectively addresses the IS and ED challenges. Specifically, to combat the intractable localization problem caused by the IS challenge, we design the deep wavelet-like decomposition (DWD) strategy, which decomposes the extracted features into different frequency bands using learnable wavelet-like modules. Then, we focus on the most informative bands by filtering out noteworthy parts where discriminative cues are most likely to exist by a novel frequency attention (FA) module. Moreover, a guidance-based feature aggregation (GFA) module is proposed to aggregate the multi-scale decomposed features with attentional guidance to further emphasize discriminative information." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.698, + 0.473, + 0.849 + ], + "angle": 0, + "content": "To address the ambiguous boundary problem of the ED challenge, we propose learning an auxiliary edge reconstruction task to encourage the network to excavate edge details. We design the ordinary differential equation (ODE)-inspired edge reconstruction (OER) module to reconstruct accurate and complete edge prediction maps using a high-order ODE solver, specifically, the second-order Runge-Kutta. Incorporating this auxiliary task with the COD task can facilitate the generation of precise segmentation results with accurate object boundaries." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.851, + 0.4, + 0.865 + ], + "angle": 0, + "content": "Our contributions are summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.094, + 0.871, + 0.472, + 0.902 + ], + "angle": 0, + "content": "- We propose the FEature Decomposition and Edge Reconstruction (FEDER) model for the COD task. To" + }, + { + "type": "text", + "bbox": [ + 0.531, + 0.092, + 0.892, + 0.122 + ], + "angle": 0, + "content": "the best of our knowledge, we are the first to approach COD from a decomposition perspective." + }, + { + "type": "text", + "bbox": [ + 0.517, + 0.124, + 0.892, + 0.2 + ], + "angle": 0, + "content": "- To highlight the subtle discriminative features, we propose frequency attention modules to filter out the noteworthy parts of corresponding features and design the Guidance-based Feature Aggregation module to aggregate the multi-scale features with attentional guidance." + }, + { + "type": "text", + "bbox": [ + 0.517, + 0.204, + 0.892, + 0.279 + ], + "angle": 0, + "content": "- We propose to learn an auxiliary edge reconstruction task along with the COD task to help generate precise segmentation maps with accurate object boundaries and design the ODE-inspired edge reconstruction module for complete edge prediction." + }, + { + "type": "text", + "bbox": [ + 0.517, + 0.281, + 0.892, + 0.327 + ], + "angle": 0, + "content": "- The proposed FEDER significantly outperforms the state-of-the-art methods on four datasets by a large margin with cheaper computational and memory costs." + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.124, + 0.892, + 0.327 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.347, + 0.651, + 0.364 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.372, + 0.892, + 0.841 + ], + "angle": 0, + "content": "Camouflaged object detection. Unlike existing object detection tasks, camouflaged object detection (COD) poses new challenges for mining subtle discriminative features under complex camouflage strategies [6, 11]. Early techniques utilized the hand-crafted operators for COD [11, 30], which were only applicable to camouflaged scenarios with simple backgrounds. Recent research has leveraged the huge capacity of deep learning to detect camouflaged objects in a learning manner [6, 14, 28]. Inspired by the hunting process of predators, SINet [6] designed a bio-inspired network to gradually search and locate the camouflaged object. PFNet [28] proposed the position module and focus module to imitate human identification with the distraction mining strategy. By simulating human behaviors in understanding complex scenarios, SegMaR [14] integrated segment, magnify and reiterate in a coarse-to-fine manner using the multistage strategy. However, these COD solutions mainly focus on mimicking biovision systems, which can be easily confused by complex camouflaged strategies and struggle to excavate the subtle discriminative features, thus failing to handle the IS and ED challenges (see Fig. 1). Unlike these human perception-oriented techniques, we first propose to address the COD task from a decomposition perspective by decomposing the extracted features into different frequency bands with learnable wavelets and filtering out the most informative bands to excavate those inconspicuous discriminative features, thus remedying the human visual deficiency and solving the IS challenge. To handle the ED challenge, we propose learning an auxiliary edge reconstruction task along with the COD task to facilitate the generation of precise segmentation results with clear object boundaries." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.841, + 0.892, + 0.903 + ], + "angle": 0, + "content": "Deep wavelet decomposition. Deep wavelet decomposition is an effective tool to decompose image/feature into various frequency components and has gained immense popularity in many domains, such as image restoration [15]" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "22047" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.089, + 0.891, + 0.396 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.4, + 0.893, + 0.444 + ], + "angle": 0, + "content": "Figure 2. Overview of the proposed FEDER model. CFE and SED are short for camouflaged feature encoder and segmentation-oriented edge-assisted decoder. HFA/LFA and GFA denote high/low frequency attention modules and the guidance-based feature aggregation module. RRS and OER indicate the reversible re-calibration segmentation module and the ODE-inspired edge reconstruction module." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.462, + 0.47, + 0.538 + ], + "angle": 0, + "content": "and style transfer [49]. To handle the IS challenge, we introduce deep wavelet decomposition into the COD task. Furthermore, to better accommodate the COD data, we employ the learnable wavelets for deep adaptive feature decomposition, whose coefficients are updated following AWD [8]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.539, + 0.471, + 0.736 + ], + "angle": 0, + "content": "ODE-inspired network. Researchers have established a relationship between ODE and neural networks. [46] first analyzed ResNet from the perspective of discrete ODE and [1] further extended ResNet to an ODE-inspired network architecture with a more accurate transmission. Since then, ODE-inspired networks are widely utilized in many fields, such as image dehazing [37] and machine translation [19]. In this paper, to accommodate the fine-grained property of the edge, we propose an ODE-inspired edge reconstruction module with the second-order Runge-Kutta and a weighted gate mechanism, aiming to generate more accurate boundaries. Furthermore, we apply the Hamiltonian system to our OER module to ensure the stability of edge reconstruction." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.754, + 0.212, + 0.77 + ], + "angle": 0, + "content": "3. Methodology" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.781, + 0.47, + 0.903 + ], + "angle": 0, + "content": "Given a camouflaged image, we first extract a cascade of features using the camouflaged feature encoder (CFE). We then perform a wavelet-like decomposition (DWD) on the features to decompose them into different frequency bands. We select the most informative bands, such as the high-frequency and low-frequency components, for further analysis. These informative bands are processed by the frequency attention (FA) module and guidance-based feature" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.462, + 0.892, + 0.538 + ], + "angle": 0, + "content": "aggregation (GFA) module to highlight the inconspicuous discriminative features. With the aggregated features, the segmentation-oriented edge-assisted decoder (SED) outputs both the segmentation map and the edge prediction map. Fig. 2 presents the framework of our FEDER model." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.548, + 0.825, + 0.565 + ], + "angle": 0, + "content": "3.1. Camouflaged Feature Encoder (CFE)" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.571, + 0.892, + 0.753 + ], + "angle": 0, + "content": "Following SINet V2 [4], the basic encoder \\(E\\) adopts ResNet50 [10]/Res2Net50 [7] as its backbone. Given an image \\(I_{c}\\) of size \\(W\\times H\\), the basic encoder \\(E\\) generates a set of feature maps \\(\\{f_k\\}_{k = 0}^4\\) with the resolution of \\(\\frac{H}{2^{k + 1}}\\times \\frac{W}{2^{k + 1}}\\), R-Net [6] is cascaded to transform \\(\\{f_k\\}_{k = 1}^4\\) into a more informative and compact output, i.e., a series of 64-channel feature maps \\(\\{f_k^r\\}_{k = 1}^4\\). Additionally, the last feature map \\(f_{4}\\) from the basic encoder \\(E\\) is further fed into an efficient atrous spatial pyramid pooling (e-ASPP) \\(A_{e}\\) [16] to enlarge the receptive field and fuse the multi-context information, resulting in \\(d_5^s = A_e(f_4)\\), where \\(d_5^s\\) is a coarse segmentation result with the same spatial resolution as \\(f_{4}\\)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.763, + 0.794, + 0.78 + ], + "angle": 0, + "content": "3.2. Deep Wavelet-like Decomposition" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.787, + 0.824, + 0.802 + ], + "angle": 0, + "content": "3.2.1 Learnable Wavelet-like Decomposition" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.81, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Camouflaged objects share a high intrinsic similarity with the background, which poses challenges for common feature extractors to mine the inconspicuous discriminative features, ultimately resulting in the suppression of segmentation performance. Based on the biological study [41], the discriminative features of COD mainly exist in the high-" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "22048" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.092, + 0.36, + 0.17 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.364, + 0.092, + 0.637, + 0.169 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.64, + 0.092, + 0.89, + 0.17 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.321, + 0.174, + 0.647, + 0.188 + ], + "angle": 0, + "content": "Figure 3. Details of the proposed GFA, HFA, and LFA" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.204, + 0.47, + 0.295 + ], + "angle": 0, + "content": "frequency (HF) component, e.g., texture and edge, and the low-frequency (LF) component, e.g., color and illumination. Inspired by this study, we propose to perform deep wavelet-like decomposition (DWD) on the extracted features \\(\\{f_k^r\\}_{k=1}^4\\) and select the most informative HF and LF components for further refinement. We decompose \\(f_k^r\\) as" + }, + { + "type": "equation", + "bbox": [ + 0.12, + 0.297, + 0.469, + 0.314 + ], + "angle": 0, + "content": "\\[\n\\left(f _ {k} ^ {r}\\right) _ {H F} = W _ {H F} \\left(f _ {k} ^ {r}\\right), \\left(f _ {k} ^ {r}\\right) _ {L F} = W _ {L F} \\left(f _ {k} ^ {r}\\right), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.315, + 0.469, + 0.421 + ], + "angle": 0, + "content": "where \\((f_k^r)_{HF}\\) and \\((f_k^r)_{LF}\\) denote the HF and LF components of \\(f_k^r\\). \\(W_{HF}\\) and \\(W_{LF}\\) represent the learnable HF and LF filters with the coefficients updated following AWD [8] and initialized by Haar wavelet [40]. The learned wavelet-like transformer is expected to better cater to COD data than manually-designed wavelets [36, 42], thus further facilitating the extraction of inconspicuous discriminative features." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.453, + 0.34, + 0.468 + ], + "angle": 0, + "content": "3.2.2 Frequency Attention Modules" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.481, + 0.469, + 0.557 + ], + "angle": 0, + "content": "To extract discriminative information from the decomposed features, we propose a high-frequency attention (HFA) module and a low-frequency attention (LFA) module, corresponding to HF and LF bands, respectively. The detailed structures of the two modules are illustrated in Fig. 3." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.56, + 0.469, + 0.712 + ], + "angle": 0, + "content": "High-frequency attention module. We design the HFA module to accentuate those texture-rich regions for subtle discriminative feature extraction. Following [21, 26], we first apply a residual block for texture preservation, consisting of a \\(3 \\times 3\\) convolution layer, batch normalization (BN) [12], and ReLU. We then employ the joint attention module \\(J A(\\cdot)\\), which includes spatial attention [34] and channel attention [45], to highlight noteworthy parts in both spatial and channel domains. Therefore, given HF features \\((f_k^r)_{HF}\\), the HF attention map \\(p_k^h\\) is formulated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.173, + 0.714, + 0.469, + 0.731 + ], + "angle": 0, + "content": "\\[\np _ {k} ^ {h} = J A \\left(R e s B \\left(\\left(f _ {k} ^ {r}\\right) _ {H F}\\right)\\right), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.732, + 0.424, + 0.747 + ], + "angle": 0, + "content": "where \\( ResB(\\cdot) \\) denotes the residual block with BN." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.47, + 0.901 + ], + "angle": 0, + "content": "Low-frequency attention module. Low-frequency components focus more on global information, such as color distribution and illumination, which inevitably leads to inevitably existing redundant components and slight perturbations [43]. To handle those problems, we design a comprehensive normalization strategy to suppress the undesired artifacts and provide cleaner global information for attention calculation both at the instance level and channel dimension, which can highlight those abnormal regions from a global perspective. Specifically, this module takes the de" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.204, + 0.863, + 0.22 + ], + "angle": 0, + "content": "composed LF features \\((f_k^r)_{LF}\\) as the input and outputs" + }, + { + "type": "equation", + "bbox": [ + 0.571, + 0.222, + 0.892, + 0.24 + ], + "angle": 0, + "content": "\\[\np _ {k} ^ {l} = J A \\left(P N \\left(R e s I N \\left(\\left(f _ {k} ^ {r}\\right) _ {L F}\\right)\\right)\\right), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.241, + 0.892, + 0.287 + ], + "angle": 0, + "content": "where \\( ResIN(\\cdot), PN(\\cdot) \\), and \\( JA(\\cdot) \\) denote the instance normalization [44] constrained residual block, positional normalization [20], and joint attention, respectively." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.307, + 0.872, + 0.323 + ], + "angle": 0, + "content": "3.2.3 Guidance-based Feature Aggregation Module" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.327, + 0.892, + 0.433 + ], + "angle": 0, + "content": "As shown in Fig. 3, we propose a guidance-based feature aggregation (GFA) module to integrate the multi-scale decomposed features. Unlike existing heuristic-based feature aggregation strategies simply using concatenation [22, 24], GFA is specifically designed to address the key issue of COD, i.e., emphasizing the subtle discriminative features, by promoting inter-feature information interaction." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.434, + 0.893, + 0.587 + ], + "angle": 0, + "content": "Taking HF bands as an example, GFA generates the aggregated feature \\(\\{f_{k - 1}^h\\}_{k = 2}^4\\) that combines deep semantic information of the low-resolution feature \\((f_k^r)_{HF}\\) (at a higher level) and the abundant spatial details of the highresolution feature \\((f_{k - 1}^{r})_{HF}\\) (at a lower level) with the guidance of the attention map \\(p_k^h\\). Therefore, the aggregated feature \\(f_{k - 1}^h\\) can better highlight the subtle discriminative features. To extract the attention-guided semantic information, we first generate the down-sampled aggregated feature \\(f_{k - 1}^{dh}\\) with the window-based linear model [23]:" + }, + { + "type": "equation", + "bbox": [ + 0.52, + 0.589, + 0.892, + 0.608 + ], + "angle": 0, + "content": "\\[\n\\left(f _ {k - 1} ^ {d h}\\right) _ {i} = \\sigma_ {w} d o w n \\left(\\left(f _ {k - 1} ^ {r}\\right) _ {H F}\\right) _ {i} + \\mu_ {w}, \\forall i \\in s _ {w}, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.61, + 0.892, + 0.669 + ], + "angle": 0, + "content": "where \\(down(\\cdot), s_w\\), and \\(i\\) are the down-sampling operation, local window, and pixel point \\(i\\). \\(\\{\\sigma_w, \\mu_w\\}\\) are linear aggregation coefficients for the pixels in window \\(s_w\\), which can be acquired by optimizing the following objective function:" + }, + { + "type": "equation", + "bbox": [ + 0.51, + 0.671, + 0.891, + 0.704 + ], + "angle": 0, + "content": "\\[\n\\min _ {\\sigma_ {w}, \\mu_ {w}} \\sum_ {i \\in s _ {w}} \\left[ \\left(p _ {k} ^ {h}\\right) _ {i} ^ {2} \\left(\\left(f _ {k - 1} ^ {d h}\\right) _ {i} - \\left(\\left(f _ {k} ^ {r}\\right) _ {H F}\\right) _ {i}\\right) ^ {2} + \\epsilon \\sigma_ {w} ^ {2} \\right], \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.706, + 0.892, + 0.736 + ], + "angle": 0, + "content": "where \\(\\epsilon\\) is a constraint value for \\(\\sigma_w\\). See Supplementary Material (Supp) for derivations and solutions of \\(\\{\\sigma_w, \\mu_w\\}\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.736, + 0.892, + 0.797 + ], + "angle": 0, + "content": "Considering pixel \\(i\\) covered by multiple windows, we average those window-wise coefficients and get the specific aggregation coefficients \\(\\{\\sigma_i,\\mu_i\\}\\) for pixel \\(i\\). By matrixing \\(\\{\\sigma_i,\\mu_i\\}\\) into \\(\\{\\sigma_i,\\mu_i\\}\\), Eq. (4) can be rewritten as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.568, + 0.798, + 0.891, + 0.815 + ], + "angle": 0, + "content": "\\[\nf _ {k - 1} ^ {d h} = \\sigma_ {i} \\odot d o w n \\left(\\left(f _ {k - 1} ^ {r}\\right) _ {H F}\\right) + \\boldsymbol {\\mu} _ {i}, \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.817, + 0.892, + 0.863 + ], + "angle": 0, + "content": "where \\(\\odot\\) is the Hadamard product. We then up-sample \\(\\{\\pmb {\\sigma}_i,\\pmb {\\mu}_i\\}\\) as \\(\\{\\pmb {\\sigma}_h,\\pmb {\\mu}_h\\}\\) and acquire the high-resolution ag-. ggregated feature \\(f_{k - 1}^{h}\\) for enriching spatial details:" + }, + { + "type": "equation", + "bbox": [ + 0.555, + 0.865, + 0.891, + 0.905 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} f _ {k - 1} ^ {h} = G F A \\left(\\left(f _ {k} ^ {r}\\right) _ {H F}, \\left(f _ {k - 1} ^ {r}\\right) _ {H F}, p _ {k} ^ {h}\\right), \\tag {7} \\\\ = \\sigma_ {h} \\odot \\left(f _ {k - 1} ^ {r}\\right) _ {H F} + \\boldsymbol {\\mu} _ {h}. \\\\ \\end{array}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "22049" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.081, + 0.09, + 0.264, + 0.169 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.272, + 0.091, + 0.467, + 0.169 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.081, + 0.174, + 0.263, + 0.253 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.272, + 0.17, + 0.468, + 0.253 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.256, + 0.47, + 0.283 + ], + "angle": 0, + "content": "Figure 4. Details of SED with RRS and OER modules. OER is an RK2-based module guaranteed with the Hamiltonian system." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.303, + 0.468, + 0.338 + ], + "angle": 0, + "content": "To iteratively acquire the aggregated features \\(\\left\\{f_{k - 1}^{h}\\right\\}_{k = 2}^{4}\\) we redefine GFA module by replacing \\((f_k^r)_HF\\) with \\(f_{k}^{h}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.151, + 0.338, + 0.468, + 0.358 + ], + "angle": 0, + "content": "\\[\nf _ {k - 1} ^ {h} = G F A \\left(f _ {k} ^ {h}, \\left(f _ {k - 1} ^ {r}\\right) _ {H F}, p _ {k} ^ {h}\\right), \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.358, + 0.469, + 0.466 + ], + "angle": 0, + "content": "where \\( p_k^h = JA(ResB(f_k^h)) \\) and \\( f_4^h = (f_4^r)_{HF} \\). Guaranteed with frequency-specific attention, our aggregated features can emphasize more discriminative features than others by combining abundant spatial details and deep semantic information, thus better catering to the COD task. The calculation of the aggregated LF features \\( f_{k - 1}^{l} \\) are similar to \\( f_{k - 1}^{h} \\), which can be seen in Supp." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.467, + 0.469, + 0.573 + ], + "angle": 0, + "content": "Considering that bottom layers (at higher levels) focus more on HF details while top layers (at lower levels) care more about global information [38], we pass the aggregated HF/LF features into the bottom/top decoder layers along with the skip-connected encoded features \\(\\{f_k^r\\}_{k=1}^4\\). To balance performance and efficiency, the integrated features \\(\\{f_k^o\\}_{k=1}^4\\) passed to the decoder are defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.126, + 0.575, + 0.469, + 0.599 + ], + "angle": 0, + "content": "\\[\nf _ {1} ^ {o} = \\operatorname {c o n v 1} \\left(\\operatorname {c o n} \\left(f _ {1} ^ {r}, u p \\left(f _ {1} ^ {h}\\right)\\right)\\right), f _ {2} ^ {o} = f _ {2} ^ {r}, \\tag {9}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.129, + 0.596, + 0.381, + 0.614 + ], + "angle": 0, + "content": "\\[\nf _ {3} ^ {o} = f _ {3} ^ {r}, f _ {4} ^ {o} = \\operatorname {c o n v 1} \\left(\\operatorname {c o n} \\left(f _ {4} ^ {r}, f _ {3} ^ {l}\\right)\\right),\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.615, + 0.469, + 0.66 + ], + "angle": 0, + "content": "where \\(up(\\cdot)\\) and \\(con(\\cdot)\\) denote the up-sampling operation and the concatenation operation. conv1 represents \\(1 \\times 1\\) convolution, which is used for channel-level integration." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.671, + 0.466, + 0.687 + ], + "angle": 0, + "content": "3.3. Segmentation-oriented Edge-assisted Decoder" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.696, + 0.469, + 0.801 + ], + "angle": 0, + "content": "In the segmentation-oriented edge-assisted decoder (SED) \\(\\{D_k\\}_{k=1}^4\\), we propose to learn an auxiliary edge reconstruction task alongside the COD task to help generate precise segmentation maps with accurate boundaries. To be specific, as shown in Fig. 4, each decoder layer \\(D_k\\) consists of a reversible re-calibration segmentation (RRS) module and an ODE-inspired edge reconstruction (OER) module." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.82, + 0.466, + 0.835 + ], + "angle": 0, + "content": "3.3.1 Reversible Re-calibration Segmentation Module" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Due to complex camouflage, prediction maps inevitably have some ambiguous regions with low confidence, we adopt a reverse strategy to excavate cues from these low-confidence regions by reversing the attention, which" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.892, + 0.213 + ], + "angle": 0, + "content": "erases detected regions and amplifies the response in low-confidence regions, thus re-calibrating the misclassified regions. Specifically, we repeat the coarse segmentation map \\(\\{d_{k + 1}^s\\}_{k = 0}^3\\) as a 64-dimension tensor, normalize it to [0, 1] with Sigmoid \\(S(\\cdot)\\), and reverse it by subtracting each element from 1. We then multiply the integrated feature \\(f_{k}^{o}\\) with the reversed map and concatenate it with the edge feature \\(d_{k}^{fe}\\) to obtain the segmentation result \\(d_{k}^{s}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.502, + 0.214, + 0.892, + 0.24 + ], + "angle": 0, + "content": "\\[\nd _ {k} ^ {s} = \\operatorname {c o n v 3} \\left(\\operatorname {c o n} \\left(d _ {k} ^ {f e}, f _ {k} ^ {o} \\odot r v (S (r p \\left(d _ {k + 1} ^ {s}\\right))\\right)\\right), \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.501, + 0.242, + 0.825, + 0.256 + ], + "angle": 0, + "content": "where \\(rp(\\cdot)\\) and \\(rv(\\cdot)\\) denote repeat and reverse." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.278, + 0.859, + 0.293 + ], + "angle": 0, + "content": "3.3.2 ODE-inspired Edge Reconstruction Module" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.3, + 0.892, + 0.406 + ], + "angle": 0, + "content": "Existing methods tend to excavate edge information by incorporating certain priors within the residual network structure [13, 50]. However, either the intractable localization problem of the IS challenge or the ambiguous boundary problem of the ED challenge makes it difficult to design an appropriate edge prior for the COD task. In some cases, a biased prior can even reduce the segmentation performance." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.407, + 0.893, + 0.663 + ], + "angle": 0, + "content": "Therefore, instead of exploiting prior knowledge, we focus on proposing an edge-friendly network architecture, i.e., the ODE-inspired edge reconstruction (OER) module. Compared with the traditional residual network structure that can be seen as the first-order Euler discretization approximation of ODE with nonnegligible truncation errors [46], the proposed OER module employs a higher-order ODE solver, specifically, a second-order Runge-Kutta (RK2), to provide more accurate numerical solutions in edge information processing. This better accommodates the fine-grained property of edges and facilitates the complete edge reconstruction, thus addressing the ED challenge. To ensure the flexibility of our OER module, we replace the fixed trade-off parameter in the RK2 solver with a weighted gate mechanism \\( g_{w} \\) with learnable coefficients. Given an input \\( e_t \\), where \\( e_t = conv1\\left(\\operatorname{con}\\left(f_k^o, rp\\left(d_{k+1}^s\\right)\\right)\\right) \\), the proposed OER module can be formulated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.563, + 0.666, + 0.793, + 0.682 + ], + "angle": 0, + "content": "\\[\ne _ {t + 1} = e _ {t} + g _ {w} F _ {1} + \\left(1 - g _ {w}\\right) F _ {2},\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.565, + 0.685, + 0.891, + 0.701 + ], + "angle": 0, + "content": "\\[\ng _ {w} = S \\left(\\sigma_ {g} \\operatorname {c o n} (F 1, F 2) + \\mu_ {g}\\right), \\tag {11}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.565, + 0.704, + 0.826, + 0.719 + ], + "angle": 0, + "content": "\\[\nF _ {1} = F \\left(e _ {t}, \\theta_ {t}\\right), F _ {2} = F \\left(e _ {t} + F _ {1}, \\theta_ {t}\\right),\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.723, + 0.892, + 0.829 + ], + "angle": 0, + "content": "where \\(e_{t + 1} = d_k^{fe}\\), \\(\\sigma_g\\) and \\(\\mu_g\\) are the learnable parameters in \\(g_w\\). \\(\\{F_i\\}_{i = 1}^2\\) denotes the intermediate layers with shared parameters \\(\\theta_t\\) for efficiency. Following [1], we set \\(F_i\\) as a Conv-ReLU-Conv framework. To ensure the stability of OER, we apply the Hamiltonian system [9] to our OER module. By denoting Eq. (11) as \\(RK2(\\cdot)\\), the Hamiltonian-theory-guaranteed OER module is defined as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.509, + 0.831, + 0.891, + 0.854 + ], + "angle": 0, + "content": "\\[\ne _ {t + 1} = \\operatorname {c o n} \\left(e _ {t + 1} ^ {1}, e _ {t + 1} ^ {2}\\right), e _ {t + 1} ^ {1} = e _ {t} ^ {1} + R K 2 _ {1} \\left(e _ {t} ^ {2}\\right), \\tag {12}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.512, + 0.851, + 0.779, + 0.869 + ], + "angle": 0, + "content": "\\[\ne _ {t + 1} ^ {2} = e _ {t} ^ {2} - R K 2 _ {2} \\left(e _ {t} ^ {1} + R K 2 _ {1} \\left(e _ {t} ^ {2}\\right)\\right),\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.871, + 0.892, + 0.901 + ], + "angle": 0, + "content": "where \\(e_t\\) is split to \\(e_t^1\\) and \\(e_t^2\\) in channel-wise. Note that the OER module in Eq. (12) is a reversible and sta" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "22050" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.084, + 0.09, + 0.894, + 0.363 + ], + "angle": 0, + "content": "
MethodsPublicationsBackbonesCHAMELEON (76 images)CAMO (250 images)COD10K (2,026 images)NC4K (4,121 images)
M↓Fβ↑Eφ↑Sα↑M↓Fβ↑Eφ↑Sα↑M↓Fβ↑Eφ↑Sα↑
Common Setting: Single Input Scale and Single Stage
CPD [47]CVPR19ResNet500.0480.7750.8740.8570.1130.6750.7230.7160.0530.5780.7760.750
SINet [6]CVPR20ResNet500.0340.8230.9360.8720.0920.7120.8040.7450.0430.6670.8640.776
PFNet [28]CVPR21ResNet500.0330.8200.9310.8820.0850.7510.8410.7820.0400.6760.8770.800
MGL-R [50]CVPR21ResNet500.0310.8250.9170.8910.0880.7380.8120.7750.0350.6800.8510.814
LSR [25]CVPR21ResNet500.0300.8350.9350.8900.0800.7560.8380.7870.0370.6990.8800.804
UGTR [48]ICCV21ResNet500.0310.8050.9100.8880.0860.7470.8210.7840.0360.6700.8520.817
SLT-Net [2]CVPR22ResNet500.0300.8350.9400.8870.0820.7630.8480.7920.0360.6810.8750.804
SegMaR-1 [14]CVPR22ResNet500.0280.8280.9440.8920.0720.7720.8610.8050.0350.6990.8900.813
OSFormer [32]ECCV22ResNet500.0280.8360.9390.8910.0730.7670.8580.7990.0340.7010.8810.811
FEDER-R50—ResNet500.0280.8550.9470.8940.0690.7850.8730.8070.0320.7400.9000.823
SINet V2 [4]TPAMI22ResNet500.0300.8160.9420.8880.0700.7790.8820.8220.0370.6820.8870.815
BSA-Net [52]AAAI22ResNet500.0270.8510.9460.8950.0790.7680.8510.7960.0340.7230.8910.818
FEDER-R2N—ResNet500.0260.8560.9470.9030.0660.8070.8970.8360.0290.7480.9110.844
Other Setting: Multiple Input Scales (MIS)
ZoomNet [31]CVPR22ResNet500.0240.8580.9430.9020.0660.7920.8770.8200.0290.7400.8880.838
FEDER-MIS—ResNet500.0230.8690.9590.9060.0640.8010.8930.8270.0280.7560.9130.837
Other Setting: Multiple Stages (MS)
SegMaR-4 [14]CVPR22ResNet500.0250.8550.9550.9060.0710.7790.8650.8150.0330.7370.8960.833
FEDER-MS-4—ResNet500.0250.8740.9640.9070.0670.8090.8860.8220.0280.7520.9170.851
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.364, + 0.893, + 0.406 + ], + "angle": 0, + "content": "Table 1. Quantitative comparisons of the proposed FEDER and other state-of-the-art methods on four benchmarks. SegMaR-1 and SegMaR-4 denote SegMaR at one stage and four stages. R50 and R2N indicate ResNet50 and Res2Net50. The best results are marked in bold. For ResNet50 backbone in the common setting, the best two results are in red and blue fonts." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.413, + 0.47, + 0.46 + ], + "angle": 0, + "content": "ble block [9], which further promotes the edge reconstruction performance. In this case, the final edge predictions \\(\\{d_k^e\\}_{k = 1}^4\\) can be acquired in the following manner:" + }, + { + "type": "equation", + "bbox": [ + 0.15, + 0.461, + 0.47, + 0.487 + ], + "angle": 0, + "content": "\\[\nd _ {k} ^ {e} = \\operatorname {c o n v 3} \\left(e _ {t + 1}\\right) = \\operatorname {c o n v 3} \\left(d _ {k} ^ {f e}\\right). \\tag {13}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.495, + 0.231, + 0.51 + ], + "angle": 0, + "content": "3.4. Loss Functions" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.52, + 0.47, + 0.729 + ], + "angle": 0, + "content": "The loss function of the proposed FEDER consists of two kinds of supervisions, namely the segmentation mask \\(GT_{s}\\) and edge \\(GT_{e}\\) of the camouflaged object, which correspond to the multi-scale segmentation maps \\(\\{d_k^s\\}_{k = 1}^5\\) and the multi-scale object edges \\(\\{d_k^e\\}_{k = 1}^4\\). Following [4], we employ the weighted binary cross-entropy loss \\(L_{BCE}^{w}\\) and weighted intersection-over-union loss \\(L_{IoU}^{w}\\) for segmentation supervision, which focuses more on those hard pixels. For edge supervision, we use the dice loss \\(L_{dice}\\) [29] to overcome the extreme imbalance between the positive and negative samples. Furthermore, to handle the multi-scale outputs, we up-sampling all the outputs to match the size of their corresponding ground truths during training. Therefore, the total loss of our FEDER is formulated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.075, + 0.73, + 0.47, + 0.814 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} L _ {t} = \\sum_ {k = 1} ^ {5} \\frac {1}{2 ^ {k - 1}} \\left(L _ {B C E} ^ {w} \\left(d _ {k} ^ {s}, G T _ {s}\\right) + L _ {I o U} ^ {w} \\left(d _ {k} ^ {s}, G T _ {S}\\right)\\right) \\tag {14} \\\\ + \\sum_ {k = 1} ^ {4} \\frac {1}{2 ^ {k - 1}} L _ {d i c e} \\left(d _ {k} ^ {e}, G T _ {e}\\right). \\\\ \\end{array}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.821, + 0.202, + 0.838 + ], + "angle": 0, + "content": "4. Experiment" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.847, + 0.268, + 0.864 + ], + "angle": 0, + "content": "4.1. Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.871, + 0.47, + 0.902 + ], + "angle": 0, + "content": "Datasets. We used four widely-used COD datasets for evaluation, including CHAMELEON [39], CAMO [18]," + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.413, + 0.892, + 0.549 + ], + "angle": 0, + "content": "COD10K [4], and NC4K [25]. CHAMELEON contains 76 hand-annotated images. CAMO has 1,250 camouflaged images with 1,000 training images and 250 testing images. Currently, COD10K is the largest COD benchmark, with 3,040 training images and 2,026 testing images. NC4K is a large-scale COD testing dataset, comprising 4,121 images. Following [4], we form the training set with 3,040 images from COD10K and 1,000 images from CAMO, while the remaining camouflaged images are used for testing." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.55, + 0.892, + 0.626 + ], + "angle": 0, + "content": "Evaluation metrics. Four commonly-used metrics are employed for COD task, including mean absolute error \\( M \\), adaptive F-measure \\( F_{\\beta} \\) [27], mean E-measure \\( E_{\\phi} \\) [5], and structure measure \\( S_{\\alpha} \\) [3]. Larger \\( F_{\\beta} \\), \\( E_{\\phi} \\), \\( S_{\\alpha} \\), and smaller \\( M \\) indicate better segmentation performance." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.626, + 0.892, + 0.761 + ], + "angle": 0, + "content": "Implementation details. The proposed FEDER is implemented in PyTorch on two RTX3090TI GPUs and is optimized by the Adam with momentum terms (0.9, 0.999). Following the common setting [4, 6], our encoder (ResNet50 by default) is initialized with the model pre-trained on ImageNet [6]. In the training phase, the batch size is set to 36 and the learning rate is initialized to 0.0001, dividing by 10 every 80 epochs. For both training and inference phases, all images are resized as \\(384 \\times 384\\)." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.772, + 0.829, + 0.788 + ], + "angle": 0, + "content": "4.2. Comparison with the State-of-the-arts" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.795, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Quantitative analysis. We compare the proposed FEDER with 12 cutting-edge techniques in three different settings, including the common setting (single input scale and single stage) and two other settings (multiple input scales (MIS) and multiple stages (MS)). In the MIS and MS settings, the proposed FEDER follows the practices of ZoomNet [31] and SegMaR [14], where FEDER-MS-4 means" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.958 + ], + "angle": 0, + "content": "22051" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.09, + 0.891, + 0.307 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.31, + 0.893, + 0.337 + ], + "angle": 0, + "content": "Figure 5. Visual comparisons of the proposed FEDER and other nine state-of-the-art methods. Our method can generate more accurate prediction maps with clearer boundaries than other methods. Please zoom in for more details." + }, + { + "type": "image", + "bbox": [ + 0.08, + 0.362, + 0.468, + 0.57 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.572, + 0.468, + 0.599 + ], + "angle": 0, + "content": "Figure 6. Performance-Params-FLOPs comparisons of some state-of-the-art deep learning-based COD methods on COD10K [4]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.626, + 0.468, + 0.866 + ], + "angle": 0, + "content": "complete COD task with our FEDER at four stages following SegMaR-4. For a fair comparison, prediction maps of the above techniques are directly segmented by their provided models with no modifications. Besides, all prediction maps are evaluated with the same code. As shown in Tab. 1, our method achieves the best results in all settings and backbones, which comprehensively demonstrates the superiority of our FEDER. Notably, \\( COD10K \\) and \\( NC4K \\) are the two most challenging datasets in terms of the number of images and segmentation difficulty. In the common setting, the proposed FEDER surpasses the second best results \\( 3.6\\% \\) on average over all metrics on \\( COD10K \\) dataset and \\( 2.6\\% \\) on average on \\( NC4K \\) dataset. Such a big margin further confirms the effectiveness of the proposed method, both with the deep wavelet-like feature decomposition strategy and the ODE-inspired edge reconstruction module." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.871, + 0.468, + 0.901 + ], + "angle": 0, + "content": "Qualitative analysis. Fig. 5 presents a visual comparison of our FEDER and other SOTAs. We select various typi" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.364, + 0.892, + 0.787 + ], + "angle": 0, + "content": "cal and challenging camouflaged images and arrange them in order of the camouflaged object size, from smallest to largest. Note that most of these images suffer from the IS or ED challenge, which can confuse existing COD techniques, resulting in mislocalization, ambiguous boundaries, etc. In contrast to those methods, our FEDER can overcome such challenges and generate more competitive prediction maps in the following aspects: (a) more accurate localization of small objects. For those small objects under the IS challenge, precise localization is a significant problem due to subtle differences and can confuse most existing methods. Thanks to our HFA, LFA, and GFA modules, our FEDER can emphasize the inconspicuous discriminative features and thus ensure more accurate camouflaged object localization (in Rows 1 and 2). (b) clearer edges on large objects. For those large objects, our prediction maps can achieve much clearer boundaries than others (see Rows 7 and 8), which mainly attributes to our joint training strategy of edge and segmentation and our edge-friendly OER module. (c) stronger suppression of redundant information. In the IS challenge and degraded imaging scenarios, the detection performance can be inevitably influenced by redundant information, such as background noise. However, the proposed deep decomposition strategy can suppress the redundant information by filtering out those components with the most discriminative information, namely, the HF and LF components. Thus, as depicted in Rows 3 and 4 in Fig. 5, FEDER can generate more accurate prediction maps." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.795, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Efficiency analysis. We compare the performance, parameters, and FLOPs with other SOTAs on COD10K [4] in Fig. 6. As presented in Fig. 6, our proposed FEDER achieves the smallest FLOPs and parameters compared with other state-of-the-art deep learning-based COD techniques. Furthermore, our score in \\(F_{\\beta}\\) is much higher than other methods and surpasses the second best one in \\(5.6\\%\\)." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "22052" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.09, + 0.09, + 0.485, + 0.189 + ], + "angle": 0, + "content": "
MethodsCOD10K (2026 images)NC4K (4121 images)
M↓Fβ↑Eφ↑Sα↑M↓Fβ↑Eφ↑Sα↑
(a)w/o DWD0.0350.6970.8610.7940.0490.7760.8730.817
(b)FW->LW0.0320.7310.8950.8220.0460.8110.8990.844
(c)w/o HFA0.0330.7280.8950.8170.0470.8050.8910.840
(d)w/o LFA0.0330.7240.8870.8160.0470.8030.8920.841
(e)UFA->GFA0.0330.7200.8940.8100.0470.8040.8900.838
(f)FEDER0.0320.7400.9000.8230.0450.8170.9050.846
" + }, + { + "type": "table_caption", + "bbox": [ + 0.178, + 0.192, + 0.393, + 0.204 + ], + "angle": 0, + "content": "(a) Ablation study of DWD Component." + }, + { + "type": "table", + "bbox": [ + 0.491, + 0.09, + 0.887, + 0.189 + ], + "angle": 0, + "content": "
MethodsCOD10K (2026 images)NC4K (4121 images)
M↓Fβ↑Eφ↑Sα↑M↓Fβ↑Eφ↑Sα↑
(a)w/o OER0.0340.7140.8710.8040.0480.7800.8890.834
(b)FC->WGM0.0320.7310.8990.8170.0460.8140.8950.842
(c)w/o HS0.0320.7230.8850.8210.0460.7980.8920.840
(d)RB->RK20.0330.7220.8810.8110.0470.8020.8950.843
(e)RK4->RK20.0310.7420.9050.8290.0450.8160.9060.848
(f)FEDER0.0320.7400.9000.8230.0450.8170.9050.846
" + }, + { + "type": "table_caption", + "bbox": [ + 0.591, + 0.191, + 0.779, + 0.203 + ], + "angle": 0, + "content": "(b) Ablation study of OER module." + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.204, + 0.894, + 0.245 + ], + "angle": 0, + "content": "Table 2. Ablation study on \\( COD10k \\) [4] and \\( NC4K \\) [25], where w/o means without. (a) FW, LW, and UFA are short for fixed wavelet [40], learnable wavelet, and upsampling-based feature aggregation [51]. (b) FC, WGM, HS, RB, and RK4 are short for fixed coefficient, weighted gate mechanism, Hamiltonian system, residual block, and forth-order Runge-Kutta. The best results are marked in bold." + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.263, + 0.466, + 0.469 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.47, + 0.47, + 0.512 + ], + "angle": 0, + "content": "Figure 7. Visual verification of the effectiveness of the proposed components. B, D, and O denote baseline, DWD, and OER. Note that \\(\\mathrm{B + D + O}\\) is our proposed FEDER." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.531, + 0.232, + 0.547 + ], + "angle": 0, + "content": "4.3. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.554, + 0.47, + 0.584 + ], + "angle": 0, + "content": "We conduct the ablation study on the two largest datasets, namely \\( COD10K \\) and \\( NC4K \\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.584, + 0.47, + 0.75 + ], + "angle": 0, + "content": "Effect of DWD component. We demonstrate the effectiveness of our DWD component both in visual verification (see Fig. 7) and quantitative analysis (see Tab. 2a). In Fig. 7, the networks with DWD component (Rows 3 and 5) exhibit more accurate localizations and stronger redundant information suppression capacities. In addition, we present more detailed information in Tab. 2a to verify the validity of each part in the DWD component. As presented in Tab. 2a, we prove the superiority of the DWD component (in (a)), learnable wavelet-like decomposition strategy (in (b)), HFA/LFA module (in (c) and (d)), and GFA module (in (e))." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Effect of OER module. The efficacy of our OER module is demonstrated visually by Figs. 7 and 8. In Fig. 7, methods with the OER module can generate the prediction maps with clearer edges. Besides, Fig. 8 illustrates the advancement of our OER module in generating accurate and clear boundary information. We provide detailed information about the superiority of our OER module in Tab. 2b. Specifically, (b) and (c) verify the effectiveness of our weighted gate mechanism (learnable coefficient) and Hamiltonian system. We further compare the performance of the OER module with" + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.264, + 0.888, + 0.417 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.5, + 0.419, + 0.892, + 0.446 + ], + "angle": 0, + "content": "Figure 8. Visualizations of the reconstructed edge from the OER module, where the edges are marked in red for well-read." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.465, + 0.893, + 0.542 + ], + "angle": 0, + "content": "different Runge-Kutta methods, i.e., RB (RK1), RK2, and RK4. Notably, the COD results with RK2 significantly outperform those with RB and are slightly lower than that with RK4. Therefore, we integrate RK2 into our OER module for a balance of performance and efficiency." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.555, + 0.628, + 0.57 + ], + "angle": 0, + "content": "5. Conclusions" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.58, + 0.893, + 0.776 + ], + "angle": 0, + "content": "To address the IS and ED challenges, in this paper, we propose the FEDER model for COD. Specifically, we decompose the features into different frequency bands with learnable wavelets and filter out the most informative bands to excavate the subtle discriminative features with the HFA, LFA, and GFA modules, thereby solving the IS challenge. Besides, we propose to learn an auxiliary edge reconstruction task with our OER module to generate complete edges. Learning this auxiliary task along with the COD task thus facilitates the generation of precise segmentation results with accurate object boundaries, thus mitigating the ED challenge. Extensive experiments verify the superiority of our FEDER model in comparison with other SOTAs." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.777, + 0.894, + 0.897 + ], + "angle": 0, + "content": "Acknowledgements: This research is partly supported by the National Key R&D Program of China (Grant No. 2020AAA0108303), and Shenzhen Science and Technology Project (Grant No. JCYJ20200109143041798) & Shenzhen Stable Supporting Program (WDZC20200820200655001) & Shenzhen Key Laboratory of next generation interactive media innovative technology (Grant No. ZDSYS20210623092001004)." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.958 + ], + "angle": 0, + "content": "22053" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.175, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.169 + ], + "angle": 0, + "content": "[1] Bo Chang, Lili Meng, Eldad Haber, Lars Ruthotto, David Begert, and Elliot Holtham. Reversible architectures for arbitrarily deep residual neural networks. In AAAI, 2018. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.171, + 0.47, + 0.225 + ], + "angle": 0, + "content": "[2] Xuelian Cheng, Huan Xiong, Deng-Ping Fan, Yiran Zhong, Mehrtash Harandi, Tom Drummond, and Zongyuan Ge. Implicit motion handling for video camouflaged object detection. In CVPR, pages 13864-13873, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.226, + 0.47, + 0.268 + ], + "angle": 0, + "content": "[3] Deng-Ping Fan, Ming-Ming Cheng, Yun Liu, Tao Li, and Ali Borji. Structure-measure: A new way to evaluate foreground maps. In ICCV, pages 4548-4557, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.268, + 0.47, + 0.308 + ], + "angle": 0, + "content": "[4] Deng-Ping Fan, Ge-Peng Ji, Ming-Ming Cheng, and Ling Shao. Concealed object detection. IEEE Trans. Pattern Anal. Mach. Intell., 2021. 3, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.31, + 0.47, + 0.351 + ], + "angle": 0, + "content": "[5] Deng-Ping Fan, Ge-Peng Ji, Xuebin Qin, and Ming-Ming Cheng. Cognitive vision inspired object segmentation metric and loss function. Sci. Sin. Inf., 6:6, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.352, + 0.47, + 0.392 + ], + "angle": 0, + "content": "[6] Deng-Ping Fan, Ge-Peng Ji, Guolei Sun, Ming-Ming Cheng, Jianbing Shen, and Ling Shao. Camouflaged object detection. In CVPR, pages 2777-2787, 2020. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.393, + 0.47, + 0.447 + ], + "angle": 0, + "content": "[7] Shang-Hua Gao, Ming-Ming Cheng, Kai Zhao, Xin-Yu Zhang, Ming-Hsuan Yang, and Philip Torr. Res2net: A new multi-scale backbone architecture. IEEE Trans. Pattern Anal. Mach. Intell., 43(2):652-662, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.448, + 0.47, + 0.502 + ], + "angle": 0, + "content": "[8] Wooseok Ha, Chandan Singh, Francois Lanusse, Srigokul Upadhyayula, and Bin Yu. Adaptive wavelet distillation from neural networks through interpretations. NIPS, 34:20669-20682, 2021. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.504, + 0.47, + 0.53 + ], + "angle": 0, + "content": "[9] Eldad Haber and Lars Ruthotto. Stable architectures for deep neural networks. Inverse Probl, 34(1):014004, 2017. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.531, + 0.47, + 0.572 + ], + "angle": 0, + "content": "[10] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, pages 770-778, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.573, + 0.47, + 0.627 + ], + "angle": 0, + "content": "[11] Jianqin Yin Yanbin Han Wendi Hou and Jinping Li. Detection of the mobile object with camouflage color under dynamic background based on optical flow. Procedia Eng., 15:2201-2205, 2011. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.628, + 0.47, + 0.669 + ], + "angle": 0, + "content": "[12] Sergey Ioffe and Christian Szegedy. Batch normalization: Accelerating deep network training by reducing internal covariate shift. In ICML, pages 448-456. PMLR, 2015. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.669, + 0.47, + 0.723 + ], + "angle": 0, + "content": "[13] Ge-Peng Ji, Lei Zhu, Mingchen Zhuge, and Keren Fu. Fast camouflaged object detection via edge-based reversible recalibration network. Pattern Recognit., 123:108414, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.725, + 0.47, + 0.779 + ], + "angle": 0, + "content": "[14] Qi Jia, Shuilian Yao, Yu Liu, Xin Fan, Risheng Liu, and Zhongxuan Luo. Segment, magnify and reiterate: Detecting camouflaged objects the hard way. In CVPR, pages 4713-4722, 2022. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.78, + 0.47, + 0.821 + ], + "angle": 0, + "content": "[15] Liming Jiang, Bo Dai, Wayne Wu, and Chen Change Loy. Focal frequency loss for image reconstruction and synthesis. In ICCV, pages 13919-13929, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.822, + 0.47, + 0.875 + ], + "angle": 0, + "content": "[16] Zhanghan Ke, Jiayu Sun, Kaican Li, Qiong Yan, and Rynson WH Lau. Modnet: Real-time trimap-free portrait matting via objective decomposition. In AAAI, pages 1140-1147, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.876, + 0.47, + 0.915 + ], + "angle": 0, + "content": "[17] Ajay Kumar. Computer-vision-based fabric defect detection: A survey. IEEE Trans. Ind. Electron., 55(1):348-363, 2008. 1" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.47, + 0.915 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.893, + 0.147 + ], + "angle": 0, + "content": "[18] Trung-Nghia Le, Tam V Nguyen, Zhongliang Nie, Minh-Triet Tran, and Akihiro Sugimoto. Anabranch network for camouflaged object segmentation. Comput. Vis. Image Underst., 184:45-56, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.149, + 0.893, + 0.218 + ], + "angle": 0, + "content": "[19] Bei Li, Quan Du, Tao Zhou, Yi Jing, Shuhan Zhou, Xin Zeng, Tong Xiao, Jingbo Zhu, Xuebo Liu, and Min Zhang. Ode transformer: An ordinary differential equation-inspired model for sequence generation. In ACL, pages 8335-8351, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.22, + 0.893, + 0.248 + ], + "angle": 0, + "content": "[20] Boyi Li, Felix Wu, Kilian Q Weinberger, and Serge Belongie. Positional normalization. NIPS, 32, 2019. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.249, + 0.893, + 0.29 + ], + "angle": 0, + "content": "[21] Juncheng Li, Faming Fang, Kangfu Mei, and Guixu Zhang. Multi-scale residual network for image super-resolution. In ECCV, pages 517-532, 2018. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.292, + 0.893, + 0.333 + ], + "angle": 0, + "content": "[22] Kai Li, Yulun Zhang, Kunpeng Li, and Yun Fu. Adversarial feature hallucination networks for few-shot learning. In CVPR, pages 13470-13479, 2020. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.334, + 0.893, + 0.389 + ], + "angle": 0, + "content": "[23] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In ICCV, pages 10012-10022, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.391, + 0.893, + 0.46 + ], + "angle": 0, + "content": "[24] Laetitia Loncan, Luis B De Almeida, José M Bioucas-Dias, Xavier Briottet, Jocelyn Chanussot, Nicolas Dobigeon, Sophie Fabre, Wenzhi Liao, Giorgio A Licciardi, Miguel Simoes, et al. Hyperspectral pansharpening: A review. IEEE Geosci. Remote Sens. Mag., 3(3):27-46, 2015. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.461, + 0.893, + 0.516 + ], + "angle": 0, + "content": "[25] Yunqiu Lv, Jing Zhang, Yuchao Dai, Aixuan Li, Bowen Liu, Nick Barnes, and Deng-Ping Fan. Simultaneously localize, segment and rank the camouflaged objects. In CVPR, pages 11591-11601, 2021. 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.518, + 0.893, + 0.571 + ], + "angle": 0, + "content": "[26] Yue Ma, Yali Wang, Yue Wu, Ziyu Lyu, Siran Chen, Xiu Li, and Yu Qiao. Visual knowledge graph for human action reasoning in videos. In ACM MM, pages 4132-4141, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.575, + 0.893, + 0.614 + ], + "angle": 0, + "content": "[27] Ran Margolin, Lihi Zelnik-Manor, and Ayellet Tal. How to evaluate foreground maps? In CVPR, pages 248-255, 2014. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.617, + 0.893, + 0.672 + ], + "angle": 0, + "content": "[28] Haiyang Mei, Ge-Peng Ji, Ziqi Wei, Xin Yang, Xiaopeng Wei, and Deng-Ping Fan. Camouflaged object segmentation with distraction mining. In CVPR, pages 8772-8781, 2021. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.674, + 0.893, + 0.728 + ], + "angle": 0, + "content": "[29] Fausto Milletari, Nassir Navab, and Seyed-Ahmad Ahmadi. V-net: Fully convolutional neural networks for volumetric medical image segmentation. In 3DV, pages 565-571. IEEE, 2016. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.731, + 0.893, + 0.772 + ], + "angle": 0, + "content": "[30] Yuxin Pan, Yiwang Chen, Qiang Fu, Ping Zhang, and Xin Xu. Study on the camouflaged target detection method based on 3d convexity. Mod. Appl. Sci., 5(4):152, 2011. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.774, + 0.893, + 0.828 + ], + "angle": 0, + "content": "[31] Youwei Pang, Xiaoqi Zhao, Tian-Zhu Xiang, Lihe Zhang, and Huchuan Lu. Zoom in and out: A mixed-scale triplet network for camouflaged object detection. In CVPR, pages 2160-2170, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.83, + 0.893, + 0.886 + ], + "angle": 0, + "content": "[32] Jialun Pei, Tianyang Cheng, Deng-Ping Fan, He Tang, Chuanbo Chen, and Luc Van Gool. Osformer: One-stage camouflaged instance segmentation with transformers. In ECCV, pages 19-37. Springer, 2022. 6" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.893, + 0.886 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "22054" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.147 + ], + "angle": 0, + "content": "[33] Natasha Price, Samuel Green, Jolyon Troscianko, Tom Tre-genza, and Martin Stevens. Background matching and disruptive coloration as habitat-specific strategies for camouflage. Sci. Rep., 9(1):1-10, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.149, + 0.47, + 0.203 + ], + "angle": 0, + "content": "[34] Xu Qin, Zhilin Wang, Yuanchao Bai, Xiaodong Xie, and Huizhu Jia. Ffa-net: Feature fusion attention network for single image dehazing. In AAAI, pages 11908-11915, 2020. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.206, + 0.471, + 0.274 + ], + "angle": 0, + "content": "[35] Dan Jeric Arcega Rustia, Chien Erh Lin, Jui-Yung Chung, Yi-Ji Zhuang, Ju-Chun Hsu, and Ta-Te Lin. Application of an image and environmental sensor network for automated greenhouse insect pest monitoring. J. Asia Pac. Entomol., 23(1):17-28, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.277, + 0.469, + 0.303 + ], + "angle": 0, + "content": "[36] Phil Sallee and Bruno Olshausen. Learning sparse multiscale image representations. NIPS, 15, 2002. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.305, + 0.469, + 0.347 + ], + "angle": 0, + "content": "[37] Jiawei Shen, Zhuoyan Li, Lei Yu, Gui-Song Xia, and Wen Yang. Implicit euler ode networks for single-image dehazing. In CVPRW, pages 218-219, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.349, + 0.469, + 0.389 + ], + "angle": 0, + "content": "[38] Chenyang Si, Weihao Yu, Pan Zhou, Yichen Zhou, Xinchao Wang, and Shuicheng Yan. Inception transformer. arXiv preprint arXiv:2205.12956, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.391, + 0.469, + 0.446 + ], + "angle": 0, + "content": "[39] Przemysław Skurowski, Hassan Abdulameer, J Błaszczyk, Tomasz Depta, Adam Kornacki, and P Koziel. Animal camouflage analysis: Chameleon database. Unpublished manuscript, 2(6):7, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.448, + 0.469, + 0.489 + ], + "angle": 0, + "content": "[40] Radomir S Stanković and Bogdan J Falkowski. The haar wavelet transform: its status and achievements. Comput. Electr. Eng., 29(1):25-44, 2003. 4, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.491, + 0.469, + 0.532 + ], + "angle": 0, + "content": "[41] Martin Stevens and Sami Merilaita. Animal camouflage: current issues and new perspectives. Philos. Trans. R. Soc. Lond., B, Biol. Sci., 364(1516):423-427, 2009. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.533, + 0.469, + 0.574 + ], + "angle": 0, + "content": "[42] Cheng Tai and E Weinan. Multiscale adaptive representation of signals: I. the basic framework. J. Mach. Learn. Res., 17(1):4875-4912, 2016. 4" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.471, + 0.574 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.133 + ], + "angle": 0, + "content": "[43] Christopher Torrence and Gilbert P Compo. A practical guide to wavelet analysis. Bull. Am. Meteorol. Soc., 79(1):61-78, 1998. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.136, + 0.892, + 0.177 + ], + "angle": 0, + "content": "[44] Dmitry Ulyanov, Andrea Vedaldi, and Victor Lempitsky. Instance normalization: The missing ingredient for fast stylization. arXiv preprint arXiv:1607.08022, 2016. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.179, + 0.892, + 0.233 + ], + "angle": 0, + "content": "[45] Qilong Wang, Banggu Wu, Pengfei Zhu, Peihua Li, Wangmeng Zuo, and Qinghua Hu. Eca-net: Efficient channel attention for deep convolutional neural networks. In CVPR, pages 13-19, 2020. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.235, + 0.892, + 0.262 + ], + "angle": 0, + "content": "[46] E Weinan. A proposal on machine learning via dynamical systems. Commun. Math. Stat., 1(5):1-11, 2017. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.264, + 0.892, + 0.305 + ], + "angle": 0, + "content": "[47] Zhe Wu, Li Su, and Qingming Huang. Cascaded partial decoder for fast and accurate salient object detection. In CVPR, pages 3907-3916, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.307, + 0.892, + 0.36 + ], + "angle": 0, + "content": "[48] Fan Yang, Qiang Zhai, Xin Li, Rui Huang, Ao Luo, Hong Cheng, and Deng-Ping Fan. Uncertainty-guided transformer reasoning for camouflaged object detection. In ICCV, pages 4146-4155, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.363, + 0.892, + 0.404 + ], + "angle": 0, + "content": "[49] Jaejun Yoo, Youngjung Uh, Sanghyuk Chun, Byeongkyu Kang, and Jung-Woo Ha. Photorealistic style transfer via wavelet transforms. In ICCV, pages 9036-9045, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.405, + 0.892, + 0.458 + ], + "angle": 0, + "content": "[50] Qiang Zhai, Xin Li, Fan Yang, Chenglizhao Chen, Hong Cheng, and Deng-Ping Fan. Mutual graph learning for camouflaged object detection. In CVPR, pages 12997-13007, 2021. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.46, + 0.892, + 0.515 + ], + "angle": 0, + "content": "[51] Yulun Zhang, Zhifei Zhang, Stephen DiVerdi, Zhaowen Wang, Jose Echevarria, and Yun Fu. Texture hallucination for large-factor painting super-resolution. In ECCV, pages 209-225. Springer, 2020. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.517, + 0.892, + 0.572 + ], + "angle": 0, + "content": "[52] Hongwei Zhu, Peng Li, Haoran Xie, Xuefeng Yan, Dong Liang, Dapeng Chen, Mingqiang Wei, and Jing Qin. I can find you! boundary-guided separated attention network for camouflaged object detection. In AAAI, 2022. 6" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.572 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "22055" + } + ] +] \ No newline at end of file diff --git a/2023/Camouflaged Object Detection With Feature Decomposition and Edge Reconstruction/69e6313d-2f4d-49ec-aeb1-e472deb81070_origin.pdf b/2023/Camouflaged Object Detection With Feature Decomposition and Edge Reconstruction/69e6313d-2f4d-49ec-aeb1-e472deb81070_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..f0d45d66ed24ac00ac984aa088b12d34789e9f78 --- /dev/null +++ b/2023/Camouflaged Object Detection With Feature Decomposition and Edge Reconstruction/69e6313d-2f4d-49ec-aeb1-e472deb81070_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:af0c9243e8b9e8a383b7793079e608e61ae27b04473e98be9d8bd888db30a095 +size 4447661 diff --git a/2023/Camouflaged Object Detection With Feature Decomposition and Edge Reconstruction/full.md b/2023/Camouflaged Object Detection With Feature Decomposition and Edge Reconstruction/full.md new file mode 100644 index 0000000000000000000000000000000000000000..4770d3f49b2d19fe3987d19b57fc7b5f3202721e --- /dev/null +++ b/2023/Camouflaged Object Detection With Feature Decomposition and Edge Reconstruction/full.md @@ -0,0 +1,353 @@ +# Camouflaged Object Detection with Feature Decomposition and Edge Reconstruction + +Chunming He $^{1}$ , Kai Li $^{2*}$ , Yachao Zhang $^{1}$ , Longxiang Tang $^{1}$ , Yulun Zhang $^{3}$ , Zhenhua Guo $^{4}$ , Xiu Li $^{1*}$ $^{1}$ Shenzhen International Graduate School, Tsinghua University, $^{2}$ NEC Laboratories America, $^{3}$ ETH Zürich, $^{4}$ Tianyi Traffic Technology + +{chunminghe19990224,li.gml.kai,lloong.x,yulun100,cszguo}@gmail.com + +yachaozhang@stu.xmu.edu.cn, li.xiu@sz.tsinghua.edu.cn + +# Abstract + +Camouflaged object detection (COD) aims to address the tough issue of identifying camouflaged objects visually blended into the surrounding backgrounds. COD is a challenging task due to the intrinsic similarity of camouflaged objects with the background, as well as their ambiguous boundaries. Existing approaches to this problem have developed various techniques to mimic the human visual system. Albeit effective in many cases, these methods still struggle when camouflaged objects are so deceptive to the vision system. In this paper, we propose the FEature Decomposition and Edge Reconstruction (FEDER) model for COD. The FEDER model addresses the intrinsic similarity of foreground and background by decomposing the features into different frequency bands using learnable wavelets. It then focuses on the most informative bands to mine subtle cues that differentiate foreground and background. To achieve this, a frequency attention module and a guidance-based feature aggregation module are developed. To combat the ambiguous boundary problem, we propose to learn an auxiliary edge reconstruction task alongside the COD task. We design an ordinary differential equation-inspired edge reconstruction module that generates exact edges. By learning the auxiliary task in conjunction with the COD task, the FEDER model can generate precise prediction maps with accurate object boundaries. Experiments show that our FEDER model significantly outperforms state-of-the-art methods with cheaper computational and memory costs. The code will be available at https://github.com/ChunmingHe/FEDER. + +# 1. Introduction + +Camouflaged object detection (COD) aims to detect and segment objects "seamlessly" integrated into surrounding + +![](images/2ceb23f8e240956ace12da1047172462b49289f379c43fd2be9387783d0522ab.jpg) +(a) Origin +(b) GT +(c) SegMaR +(d) FEDER +Figure 1. Results of SegMaR [14] and our method under the intrinsic similarity (IS) and edge disruption (ED) challenges. Our method better localizes the objects and produces clearer edges. + +environments. COD is a challenging task as it needs to combat against excellent camouflage strategies, including background matching [41], disruptive coloration [33], etc., and distinguish the subtle differences between candidate objects and their backgrounds. Research in COD can simultaneously facilitate the development of visual perception for nuance discrimination and promote various valuable real-life applications, ranging from concealed defect detection [17] in industry to pest monitoring [35] in agriculture. + +COD faces two main challenges. The first is the intrinsic similarity (IS) challenge, which occurs when camouflaged objects share similar colors and patterns with their backgrounds. This makes it difficult to even roughly localize those camouflaged objects. The second is the edge disruption (ED) challenge, which arises from extremely ambiguous object boundaries. Even if a rough localization is achieved, precise segmentation can barely be obtained. + +To tackle these challenges, most existing works aim to develop models that mimic the human visual system [6,28]. However, since camouflage strategies are designed by prey to confuse the predator's visual system, and the intrinsic topological properties of candidate objects are not distinctive, such human perception-oriented attempts may struggle to identify subtle discriminative features and fail to effectively address the above challenges. For instance, as illustrated in Fig. 1, the state-of-the-art human perception-based COD method can only generate inaccurate prediction maps, such as the vague caddisfly and incomplete dog (Row 2 and 4), or even fail to detect camouflaged objects like the bird and snake (Row 1 and 3). Therefore, a better COD method should compensate for the "flaw" in human perception by emphasizing subtle discriminative features. + +Based on the biological study [41], camouflaged objects often employ various camouflage strategies to conceal their discriminative differences, which mainly exist in texture details and global information distribution, within surrounding environments. Such a study inspires us to cope with the COD task by decomposing the camouflage scenario into different parts. This allows for the disentanglement of various intricate connections, enabling each part to be separately handled to fully excavate subtle discriminative cues. + +With this inspiration, we propose the FEature Decomposition and Edge Reconstruction (FEDER) model for the COD task, which compensates for the deficiencies of human perception by emphasizing subtle discriminative features and effectively addresses the IS and ED challenges. Specifically, to combat the intractable localization problem caused by the IS challenge, we design the deep wavelet-like decomposition (DWD) strategy, which decomposes the extracted features into different frequency bands using learnable wavelet-like modules. Then, we focus on the most informative bands by filtering out noteworthy parts where discriminative cues are most likely to exist by a novel frequency attention (FA) module. Moreover, a guidance-based feature aggregation (GFA) module is proposed to aggregate the multi-scale decomposed features with attentional guidance to further emphasize discriminative information. + +To address the ambiguous boundary problem of the ED challenge, we propose learning an auxiliary edge reconstruction task to encourage the network to excavate edge details. We design the ordinary differential equation (ODE)-inspired edge reconstruction (OER) module to reconstruct accurate and complete edge prediction maps using a high-order ODE solver, specifically, the second-order Runge-Kutta. Incorporating this auxiliary task with the COD task can facilitate the generation of precise segmentation results with accurate object boundaries. + +Our contributions are summarized as follows: + +- We propose the FEature Decomposition and Edge Reconstruction (FEDER) model for the COD task. To + +the best of our knowledge, we are the first to approach COD from a decomposition perspective. + +- To highlight the subtle discriminative features, we propose frequency attention modules to filter out the noteworthy parts of corresponding features and design the Guidance-based Feature Aggregation module to aggregate the multi-scale features with attentional guidance. +- We propose to learn an auxiliary edge reconstruction task along with the COD task to help generate precise segmentation maps with accurate object boundaries and design the ODE-inspired edge reconstruction module for complete edge prediction. +- The proposed FEDER significantly outperforms the state-of-the-art methods on four datasets by a large margin with cheaper computational and memory costs. + +# 2. Related Works + +Camouflaged object detection. Unlike existing object detection tasks, camouflaged object detection (COD) poses new challenges for mining subtle discriminative features under complex camouflage strategies [6, 11]. Early techniques utilized the hand-crafted operators for COD [11, 30], which were only applicable to camouflaged scenarios with simple backgrounds. Recent research has leveraged the huge capacity of deep learning to detect camouflaged objects in a learning manner [6, 14, 28]. Inspired by the hunting process of predators, SINet [6] designed a bio-inspired network to gradually search and locate the camouflaged object. PFNet [28] proposed the position module and focus module to imitate human identification with the distraction mining strategy. By simulating human behaviors in understanding complex scenarios, SegMaR [14] integrated segment, magnify and reiterate in a coarse-to-fine manner using the multistage strategy. However, these COD solutions mainly focus on mimicking biovision systems, which can be easily confused by complex camouflaged strategies and struggle to excavate the subtle discriminative features, thus failing to handle the IS and ED challenges (see Fig. 1). Unlike these human perception-oriented techniques, we first propose to address the COD task from a decomposition perspective by decomposing the extracted features into different frequency bands with learnable wavelets and filtering out the most informative bands to excavate those inconspicuous discriminative features, thus remedying the human visual deficiency and solving the IS challenge. To handle the ED challenge, we propose learning an auxiliary edge reconstruction task along with the COD task to facilitate the generation of precise segmentation results with clear object boundaries. + +Deep wavelet decomposition. Deep wavelet decomposition is an effective tool to decompose image/feature into various frequency components and has gained immense popularity in many domains, such as image restoration [15] + +![](images/e1db8a771e890aa816d72c7592408874967dd9d3e4bfbfbba41e6d5092702720.jpg) +Figure 2. Overview of the proposed FEDER model. CFE and SED are short for camouflaged feature encoder and segmentation-oriented edge-assisted decoder. HFA/LFA and GFA denote high/low frequency attention modules and the guidance-based feature aggregation module. RRS and OER indicate the reversible re-calibration segmentation module and the ODE-inspired edge reconstruction module. + +and style transfer [49]. To handle the IS challenge, we introduce deep wavelet decomposition into the COD task. Furthermore, to better accommodate the COD data, we employ the learnable wavelets for deep adaptive feature decomposition, whose coefficients are updated following AWD [8]. + +ODE-inspired network. Researchers have established a relationship between ODE and neural networks. [46] first analyzed ResNet from the perspective of discrete ODE and [1] further extended ResNet to an ODE-inspired network architecture with a more accurate transmission. Since then, ODE-inspired networks are widely utilized in many fields, such as image dehazing [37] and machine translation [19]. In this paper, to accommodate the fine-grained property of the edge, we propose an ODE-inspired edge reconstruction module with the second-order Runge-Kutta and a weighted gate mechanism, aiming to generate more accurate boundaries. Furthermore, we apply the Hamiltonian system to our OER module to ensure the stability of edge reconstruction. + +# 3. Methodology + +Given a camouflaged image, we first extract a cascade of features using the camouflaged feature encoder (CFE). We then perform a wavelet-like decomposition (DWD) on the features to decompose them into different frequency bands. We select the most informative bands, such as the high-frequency and low-frequency components, for further analysis. These informative bands are processed by the frequency attention (FA) module and guidance-based feature + +aggregation (GFA) module to highlight the inconspicuous discriminative features. With the aggregated features, the segmentation-oriented edge-assisted decoder (SED) outputs both the segmentation map and the edge prediction map. Fig. 2 presents the framework of our FEDER model. + +# 3.1. Camouflaged Feature Encoder (CFE) + +Following SINet V2 [4], the basic encoder $E$ adopts ResNet50 [10]/Res2Net50 [7] as its backbone. Given an image $I_{c}$ of size $W\times H$ , the basic encoder $E$ generates a set of feature maps $\{f_k\}_{k = 0}^4$ with the resolution of $\frac{H}{2^{k + 1}}\times \frac{W}{2^{k + 1}}$ , R-Net [6] is cascaded to transform $\{f_k\}_{k = 1}^4$ into a more informative and compact output, i.e., a series of 64-channel feature maps $\{f_k^r\}_{k = 1}^4$ . Additionally, the last feature map $f_{4}$ from the basic encoder $E$ is further fed into an efficient atrous spatial pyramid pooling (e-ASPP) $A_{e}$ [16] to enlarge the receptive field and fuse the multi-context information, resulting in $d_5^s = A_e(f_4)$ , where $d_5^s$ is a coarse segmentation result with the same spatial resolution as $f_{4}$ . + +# 3.2. Deep Wavelet-like Decomposition + +# 3.2.1 Learnable Wavelet-like Decomposition + +Camouflaged objects share a high intrinsic similarity with the background, which poses challenges for common feature extractors to mine the inconspicuous discriminative features, ultimately resulting in the suppression of segmentation performance. Based on the biological study [41], the discriminative features of COD mainly exist in the high- + +![](images/6d23091521cfbf166c2637d633254fb290ef33cf79932c2078161c5135671d3e.jpg) +Figure 3. Details of the proposed GFA, HFA, and LFA + +![](images/0197348a3e691a59e03b998b156935360fa977a868b609fd2185dcc4a5f125bd.jpg) + +![](images/7f6f84ff9fa0dff726a61ec5c2430ab3398798849268bc5678813f0ee5ea922c.jpg) + +frequency (HF) component, e.g., texture and edge, and the low-frequency (LF) component, e.g., color and illumination. Inspired by this study, we propose to perform deep wavelet-like decomposition (DWD) on the extracted features $\{f_k^r\}_{k=1}^4$ and select the most informative HF and LF components for further refinement. We decompose $f_k^r$ as + +$$ +\left(f _ {k} ^ {r}\right) _ {H F} = W _ {H F} \left(f _ {k} ^ {r}\right), \left(f _ {k} ^ {r}\right) _ {L F} = W _ {L F} \left(f _ {k} ^ {r}\right), \tag {1} +$$ + +where $(f_k^r)_{HF}$ and $(f_k^r)_{LF}$ denote the HF and LF components of $f_k^r$ . $W_{HF}$ and $W_{LF}$ represent the learnable HF and LF filters with the coefficients updated following AWD [8] and initialized by Haar wavelet [40]. The learned wavelet-like transformer is expected to better cater to COD data than manually-designed wavelets [36, 42], thus further facilitating the extraction of inconspicuous discriminative features. + +# 3.2.2 Frequency Attention Modules + +To extract discriminative information from the decomposed features, we propose a high-frequency attention (HFA) module and a low-frequency attention (LFA) module, corresponding to HF and LF bands, respectively. The detailed structures of the two modules are illustrated in Fig. 3. + +High-frequency attention module. We design the HFA module to accentuate those texture-rich regions for subtle discriminative feature extraction. Following [21, 26], we first apply a residual block for texture preservation, consisting of a $3 \times 3$ convolution layer, batch normalization (BN) [12], and ReLU. We then employ the joint attention module $J A(\cdot)$ , which includes spatial attention [34] and channel attention [45], to highlight noteworthy parts in both spatial and channel domains. Therefore, given HF features $(f_k^r)_{HF}$ , the HF attention map $p_k^h$ is formulated as follows: + +$$ +p _ {k} ^ {h} = J A \left(R e s B \left(\left(f _ {k} ^ {r}\right) _ {H F}\right)\right), \tag {2} +$$ + +where $ResB(\cdot)$ denotes the residual block with BN. + +Low-frequency attention module. Low-frequency components focus more on global information, such as color distribution and illumination, which inevitably leads to inevitably existing redundant components and slight perturbations [43]. To handle those problems, we design a comprehensive normalization strategy to suppress the undesired artifacts and provide cleaner global information for attention calculation both at the instance level and channel dimension, which can highlight those abnormal regions from a global perspective. Specifically, this module takes the de + +composed LF features $(f_k^r)_{LF}$ as the input and outputs + +$$ +p _ {k} ^ {l} = J A \left(P N \left(R e s I N \left(\left(f _ {k} ^ {r}\right) _ {L F}\right)\right)\right), \tag {3} +$$ + +where $ResIN(\cdot), PN(\cdot)$ , and $JA(\cdot)$ denote the instance normalization [44] constrained residual block, positional normalization [20], and joint attention, respectively. + +# 3.2.3 Guidance-based Feature Aggregation Module + +As shown in Fig. 3, we propose a guidance-based feature aggregation (GFA) module to integrate the multi-scale decomposed features. Unlike existing heuristic-based feature aggregation strategies simply using concatenation [22, 24], GFA is specifically designed to address the key issue of COD, i.e., emphasizing the subtle discriminative features, by promoting inter-feature information interaction. + +Taking HF bands as an example, GFA generates the aggregated feature $\{f_{k - 1}^h\}_{k = 2}^4$ that combines deep semantic information of the low-resolution feature $(f_k^r)_{HF}$ (at a higher level) and the abundant spatial details of the highresolution feature $(f_{k - 1}^{r})_{HF}$ (at a lower level) with the guidance of the attention map $p_k^h$ . Therefore, the aggregated feature $f_{k - 1}^h$ can better highlight the subtle discriminative features. To extract the attention-guided semantic information, we first generate the down-sampled aggregated feature $f_{k - 1}^{dh}$ with the window-based linear model [23]: + +$$ +\left(f _ {k - 1} ^ {d h}\right) _ {i} = \sigma_ {w} d o w n \left(\left(f _ {k - 1} ^ {r}\right) _ {H F}\right) _ {i} + \mu_ {w}, \forall i \in s _ {w}, \tag {4} +$$ + +where $down(\cdot), s_w$ , and $i$ are the down-sampling operation, local window, and pixel point $i$ . $\{\sigma_w, \mu_w\}$ are linear aggregation coefficients for the pixels in window $s_w$ , which can be acquired by optimizing the following objective function: + +$$ +\min _ {\sigma_ {w}, \mu_ {w}} \sum_ {i \in s _ {w}} \left[ \left(p _ {k} ^ {h}\right) _ {i} ^ {2} \left(\left(f _ {k - 1} ^ {d h}\right) _ {i} - \left(\left(f _ {k} ^ {r}\right) _ {H F}\right) _ {i}\right) ^ {2} + \epsilon \sigma_ {w} ^ {2} \right], \tag {5} +$$ + +where $\epsilon$ is a constraint value for $\sigma_w$ . See Supplementary Material (Supp) for derivations and solutions of $\{\sigma_w, \mu_w\}$ . + +Considering pixel $i$ covered by multiple windows, we average those window-wise coefficients and get the specific aggregation coefficients $\{\sigma_i,\mu_i\}$ for pixel $i$ . By matrixing $\{\sigma_i,\mu_i\}$ into $\{\sigma_i,\mu_i\}$ , Eq. (4) can be rewritten as follows: + +$$ +f _ {k - 1} ^ {d h} = \sigma_ {i} \odot d o w n \left(\left(f _ {k - 1} ^ {r}\right) _ {H F}\right) + \boldsymbol {\mu} _ {i}, \tag {6} +$$ + +where $\odot$ is the Hadamard product. We then up-sample $\{\pmb {\sigma}_i,\pmb {\mu}_i\}$ as $\{\pmb {\sigma}_h,\pmb {\mu}_h\}$ and acquire the high-resolution ag-. ggregated feature $f_{k - 1}^{h}$ for enriching spatial details: + +$$ +\begin{array}{l} f _ {k - 1} ^ {h} = G F A \left(\left(f _ {k} ^ {r}\right) _ {H F}, \left(f _ {k - 1} ^ {r}\right) _ {H F}, p _ {k} ^ {h}\right), \tag {7} \\ = \sigma_ {h} \odot \left(f _ {k - 1} ^ {r}\right) _ {H F} + \boldsymbol {\mu} _ {h}. \\ \end{array} +$$ + +![](images/866a33e0554a38702327cb1a77fe6198f9cf0428ebcc76212aff636e29623f83.jpg) + +![](images/51c635ff8eb5e70df58596dfe41a541b3c8582289e6dedbb546269f96fdde41e.jpg) + +![](images/e906f30558141957c5ea9078f6ee271dbd79ad0f0545ea32cad7ca4d68cf224e.jpg) +Figure 4. Details of SED with RRS and OER modules. OER is an RK2-based module guaranteed with the Hamiltonian system. + +![](images/8364ece33d2de787d7b9c493ef51c7e15fa6b70747e11a81ae71218ffa5a40e1.jpg) + +To iteratively acquire the aggregated features $\left\{f_{k - 1}^{h}\right\}_{k = 2}^{4}$ we redefine GFA module by replacing $(f_k^r)_HF$ with $f_{k}^{h}$ : + +$$ +f _ {k - 1} ^ {h} = G F A \left(f _ {k} ^ {h}, \left(f _ {k - 1} ^ {r}\right) _ {H F}, p _ {k} ^ {h}\right), \tag {8} +$$ + +where $p_k^h = JA(ResB(f_k^h))$ and $f_4^h = (f_4^r)_{HF}$ . Guaranteed with frequency-specific attention, our aggregated features can emphasize more discriminative features than others by combining abundant spatial details and deep semantic information, thus better catering to the COD task. The calculation of the aggregated LF features $f_{k - 1}^{l}$ are similar to $f_{k - 1}^{h}$ , which can be seen in Supp. + +Considering that bottom layers (at higher levels) focus more on HF details while top layers (at lower levels) care more about global information [38], we pass the aggregated HF/LF features into the bottom/top decoder layers along with the skip-connected encoded features $\{f_k^r\}_{k=1}^4$ . To balance performance and efficiency, the integrated features $\{f_k^o\}_{k=1}^4$ passed to the decoder are defined as: + +$$ +f _ {1} ^ {o} = \operatorname {c o n v 1} \left(\operatorname {c o n} \left(f _ {1} ^ {r}, u p \left(f _ {1} ^ {h}\right)\right)\right), f _ {2} ^ {o} = f _ {2} ^ {r}, \tag {9} +$$ + +$$ +f _ {3} ^ {o} = f _ {3} ^ {r}, f _ {4} ^ {o} = \operatorname {c o n v 1} \left(\operatorname {c o n} \left(f _ {4} ^ {r}, f _ {3} ^ {l}\right)\right), +$$ + +where $up(\cdot)$ and $con(\cdot)$ denote the up-sampling operation and the concatenation operation. conv1 represents $1 \times 1$ convolution, which is used for channel-level integration. + +# 3.3. Segmentation-oriented Edge-assisted Decoder + +In the segmentation-oriented edge-assisted decoder (SED) $\{D_k\}_{k=1}^4$ , we propose to learn an auxiliary edge reconstruction task alongside the COD task to help generate precise segmentation maps with accurate boundaries. To be specific, as shown in Fig. 4, each decoder layer $D_k$ consists of a reversible re-calibration segmentation (RRS) module and an ODE-inspired edge reconstruction (OER) module. + +# 3.3.1 Reversible Re-calibration Segmentation Module + +Due to complex camouflage, prediction maps inevitably have some ambiguous regions with low confidence, we adopt a reverse strategy to excavate cues from these low-confidence regions by reversing the attention, which + +erases detected regions and amplifies the response in low-confidence regions, thus re-calibrating the misclassified regions. Specifically, we repeat the coarse segmentation map $\{d_{k + 1}^s\}_{k = 0}^3$ as a 64-dimension tensor, normalize it to [0, 1] with Sigmoid $S(\cdot)$ , and reverse it by subtracting each element from 1. We then multiply the integrated feature $f_{k}^{o}$ with the reversed map and concatenate it with the edge feature $d_{k}^{fe}$ to obtain the segmentation result $d_{k}^{s}$ : + +$$ +d _ {k} ^ {s} = \operatorname {c o n v 3} \left(\operatorname {c o n} \left(d _ {k} ^ {f e}, f _ {k} ^ {o} \odot r v (S (r p \left(d _ {k + 1} ^ {s}\right))\right)\right), \tag {10} +$$ + +where $rp(\cdot)$ and $rv(\cdot)$ denote repeat and reverse. + +# 3.3.2 ODE-inspired Edge Reconstruction Module + +Existing methods tend to excavate edge information by incorporating certain priors within the residual network structure [13, 50]. However, either the intractable localization problem of the IS challenge or the ambiguous boundary problem of the ED challenge makes it difficult to design an appropriate edge prior for the COD task. In some cases, a biased prior can even reduce the segmentation performance. + +Therefore, instead of exploiting prior knowledge, we focus on proposing an edge-friendly network architecture, i.e., the ODE-inspired edge reconstruction (OER) module. Compared with the traditional residual network structure that can be seen as the first-order Euler discretization approximation of ODE with nonnegligible truncation errors [46], the proposed OER module employs a higher-order ODE solver, specifically, a second-order Runge-Kutta (RK2), to provide more accurate numerical solutions in edge information processing. This better accommodates the fine-grained property of edges and facilitates the complete edge reconstruction, thus addressing the ED challenge. To ensure the flexibility of our OER module, we replace the fixed trade-off parameter in the RK2 solver with a weighted gate mechanism $g_{w}$ with learnable coefficients. Given an input $e_t$ , where $e_t = conv1\left(\operatorname{con}\left(f_k^o, rp\left(d_{k+1}^s\right)\right)\right)$ , the proposed OER module can be formulated as follows: + +$$ +e _ {t + 1} = e _ {t} + g _ {w} F _ {1} + \left(1 - g _ {w}\right) F _ {2}, +$$ + +$$ +g _ {w} = S \left(\sigma_ {g} \operatorname {c o n} (F 1, F 2) + \mu_ {g}\right), \tag {11} +$$ + +$$ +F _ {1} = F \left(e _ {t}, \theta_ {t}\right), F _ {2} = F \left(e _ {t} + F _ {1}, \theta_ {t}\right), +$$ + +where $e_{t + 1} = d_k^{fe}$ , $\sigma_g$ and $\mu_g$ are the learnable parameters in $g_w$ . $\{F_i\}_{i = 1}^2$ denotes the intermediate layers with shared parameters $\theta_t$ for efficiency. Following [1], we set $F_i$ as a Conv-ReLU-Conv framework. To ensure the stability of OER, we apply the Hamiltonian system [9] to our OER module. By denoting Eq. (11) as $RK2(\cdot)$ , the Hamiltonian-theory-guaranteed OER module is defined as follows: + +$$ +e _ {t + 1} = \operatorname {c o n} \left(e _ {t + 1} ^ {1}, e _ {t + 1} ^ {2}\right), e _ {t + 1} ^ {1} = e _ {t} ^ {1} + R K 2 _ {1} \left(e _ {t} ^ {2}\right), \tag {12} +$$ + +$$ +e _ {t + 1} ^ {2} = e _ {t} ^ {2} - R K 2 _ {2} \left(e _ {t} ^ {1} + R K 2 _ {1} \left(e _ {t} ^ {2}\right)\right), +$$ + +where $e_t$ is split to $e_t^1$ and $e_t^2$ in channel-wise. Note that the OER module in Eq. (12) is a reversible and sta + +
MethodsPublicationsBackbonesCHAMELEON (76 images)CAMO (250 images)COD10K (2,026 images)NC4K (4,121 images)
M↓Fβ↑Eφ↑Sα↑M↓Fβ↑Eφ↑Sα↑M↓Fβ↑Eφ↑Sα↑
Common Setting: Single Input Scale and Single Stage
CPD [47]CVPR19ResNet500.0480.7750.8740.8570.1130.6750.7230.7160.0530.5780.7760.750
SINet [6]CVPR20ResNet500.0340.8230.9360.8720.0920.7120.8040.7450.0430.6670.8640.776
PFNet [28]CVPR21ResNet500.0330.8200.9310.8820.0850.7510.8410.7820.0400.6760.8770.800
MGL-R [50]CVPR21ResNet500.0310.8250.9170.8910.0880.7380.8120.7750.0350.6800.8510.814
LSR [25]CVPR21ResNet500.0300.8350.9350.8900.0800.7560.8380.7870.0370.6990.8800.804
UGTR [48]ICCV21ResNet500.0310.8050.9100.8880.0860.7470.8210.7840.0360.6700.8520.817
SLT-Net [2]CVPR22ResNet500.0300.8350.9400.8870.0820.7630.8480.7920.0360.6810.8750.804
SegMaR-1 [14]CVPR22ResNet500.0280.8280.9440.8920.0720.7720.8610.8050.0350.6990.8900.813
OSFormer [32]ECCV22ResNet500.0280.8360.9390.8910.0730.7670.8580.7990.0340.7010.8810.811
FEDER-R50—ResNet500.0280.8550.9470.8940.0690.7850.8730.8070.0320.7400.9000.823
SINet V2 [4]TPAMI22ResNet500.0300.8160.9420.8880.0700.7790.8820.8220.0370.6820.8870.815
BSA-Net [52]AAAI22ResNet500.0270.8510.9460.8950.0790.7680.8510.7960.0340.7230.8910.818
FEDER-R2N—ResNet500.0260.8560.9470.9030.0660.8070.8970.8360.0290.7480.9110.844
Other Setting: Multiple Input Scales (MIS)
ZoomNet [31]CVPR22ResNet500.0240.8580.9430.9020.0660.7920.8770.8200.0290.7400.8880.838
FEDER-MIS—ResNet500.0230.8690.9590.9060.0640.8010.8930.8270.0280.7560.9130.837
Other Setting: Multiple Stages (MS)
SegMaR-4 [14]CVPR22ResNet500.0250.8550.9550.9060.0710.7790.8650.8150.0330.7370.8960.833
FEDER-MS-4—ResNet500.0250.8740.9640.9070.0670.8090.8860.8220.0280.7520.9170.851
+ +Table 1. Quantitative comparisons of the proposed FEDER and other state-of-the-art methods on four benchmarks. SegMaR-1 and SegMaR-4 denote SegMaR at one stage and four stages. R50 and R2N indicate ResNet50 and Res2Net50. The best results are marked in bold. For ResNet50 backbone in the common setting, the best two results are in red and blue fonts. + +ble block [9], which further promotes the edge reconstruction performance. In this case, the final edge predictions $\{d_k^e\}_{k = 1}^4$ can be acquired in the following manner: + +$$ +d _ {k} ^ {e} = \operatorname {c o n v 3} \left(e _ {t + 1}\right) = \operatorname {c o n v 3} \left(d _ {k} ^ {f e}\right). \tag {13} +$$ + +# 3.4. Loss Functions + +The loss function of the proposed FEDER consists of two kinds of supervisions, namely the segmentation mask $GT_{s}$ and edge $GT_{e}$ of the camouflaged object, which correspond to the multi-scale segmentation maps $\{d_k^s\}_{k = 1}^5$ and the multi-scale object edges $\{d_k^e\}_{k = 1}^4$ . Following [4], we employ the weighted binary cross-entropy loss $L_{BCE}^{w}$ and weighted intersection-over-union loss $L_{IoU}^{w}$ for segmentation supervision, which focuses more on those hard pixels. For edge supervision, we use the dice loss $L_{dice}$ [29] to overcome the extreme imbalance between the positive and negative samples. Furthermore, to handle the multi-scale outputs, we up-sampling all the outputs to match the size of their corresponding ground truths during training. Therefore, the total loss of our FEDER is formulated as follows: + +$$ +\begin{array}{l} L _ {t} = \sum_ {k = 1} ^ {5} \frac {1}{2 ^ {k - 1}} \left(L _ {B C E} ^ {w} \left(d _ {k} ^ {s}, G T _ {s}\right) + L _ {I o U} ^ {w} \left(d _ {k} ^ {s}, G T _ {S}\right)\right) \tag {14} \\ + \sum_ {k = 1} ^ {4} \frac {1}{2 ^ {k - 1}} L _ {d i c e} \left(d _ {k} ^ {e}, G T _ {e}\right). \\ \end{array} +$$ + +# 4. Experiment + +# 4.1. Experimental Setup + +Datasets. We used four widely-used COD datasets for evaluation, including CHAMELEON [39], CAMO [18], + +COD10K [4], and NC4K [25]. CHAMELEON contains 76 hand-annotated images. CAMO has 1,250 camouflaged images with 1,000 training images and 250 testing images. Currently, COD10K is the largest COD benchmark, with 3,040 training images and 2,026 testing images. NC4K is a large-scale COD testing dataset, comprising 4,121 images. Following [4], we form the training set with 3,040 images from COD10K and 1,000 images from CAMO, while the remaining camouflaged images are used for testing. + +Evaluation metrics. Four commonly-used metrics are employed for COD task, including mean absolute error $M$ , adaptive F-measure $F_{\beta}$ [27], mean E-measure $E_{\phi}$ [5], and structure measure $S_{\alpha}$ [3]. Larger $F_{\beta}$ , $E_{\phi}$ , $S_{\alpha}$ , and smaller $M$ indicate better segmentation performance. + +Implementation details. The proposed FEDER is implemented in PyTorch on two RTX3090TI GPUs and is optimized by the Adam with momentum terms (0.9, 0.999). Following the common setting [4, 6], our encoder (ResNet50 by default) is initialized with the model pre-trained on ImageNet [6]. In the training phase, the batch size is set to 36 and the learning rate is initialized to 0.0001, dividing by 10 every 80 epochs. For both training and inference phases, all images are resized as $384 \times 384$ . + +# 4.2. Comparison with the State-of-the-arts + +Quantitative analysis. We compare the proposed FEDER with 12 cutting-edge techniques in three different settings, including the common setting (single input scale and single stage) and two other settings (multiple input scales (MIS) and multiple stages (MS)). In the MIS and MS settings, the proposed FEDER follows the practices of ZoomNet [31] and SegMaR [14], where FEDER-MS-4 means + +![](images/368ed6bce6fca44bf494230f7a13b13d55ef9c0209b9ae769f594f1212645419.jpg) +Figure 5. Visual comparisons of the proposed FEDER and other nine state-of-the-art methods. Our method can generate more accurate prediction maps with clearer boundaries than other methods. Please zoom in for more details. + +![](images/b039425d32da85e97fb5cfb7a542cad0ae8a695d7c926159c176cc6ea880885d.jpg) +Figure 6. Performance-Params-FLOPs comparisons of some state-of-the-art deep learning-based COD methods on COD10K [4]. + +complete COD task with our FEDER at four stages following SegMaR-4. For a fair comparison, prediction maps of the above techniques are directly segmented by their provided models with no modifications. Besides, all prediction maps are evaluated with the same code. As shown in Tab. 1, our method achieves the best results in all settings and backbones, which comprehensively demonstrates the superiority of our FEDER. Notably, $COD10K$ and $NC4K$ are the two most challenging datasets in terms of the number of images and segmentation difficulty. In the common setting, the proposed FEDER surpasses the second best results $3.6\%$ on average over all metrics on $COD10K$ dataset and $2.6\%$ on average on $NC4K$ dataset. Such a big margin further confirms the effectiveness of the proposed method, both with the deep wavelet-like feature decomposition strategy and the ODE-inspired edge reconstruction module. + +Qualitative analysis. Fig. 5 presents a visual comparison of our FEDER and other SOTAs. We select various typi + +cal and challenging camouflaged images and arrange them in order of the camouflaged object size, from smallest to largest. Note that most of these images suffer from the IS or ED challenge, which can confuse existing COD techniques, resulting in mislocalization, ambiguous boundaries, etc. In contrast to those methods, our FEDER can overcome such challenges and generate more competitive prediction maps in the following aspects: (a) more accurate localization of small objects. For those small objects under the IS challenge, precise localization is a significant problem due to subtle differences and can confuse most existing methods. Thanks to our HFA, LFA, and GFA modules, our FEDER can emphasize the inconspicuous discriminative features and thus ensure more accurate camouflaged object localization (in Rows 1 and 2). (b) clearer edges on large objects. For those large objects, our prediction maps can achieve much clearer boundaries than others (see Rows 7 and 8), which mainly attributes to our joint training strategy of edge and segmentation and our edge-friendly OER module. (c) stronger suppression of redundant information. In the IS challenge and degraded imaging scenarios, the detection performance can be inevitably influenced by redundant information, such as background noise. However, the proposed deep decomposition strategy can suppress the redundant information by filtering out those components with the most discriminative information, namely, the HF and LF components. Thus, as depicted in Rows 3 and 4 in Fig. 5, FEDER can generate more accurate prediction maps. + +Efficiency analysis. We compare the performance, parameters, and FLOPs with other SOTAs on COD10K [4] in Fig. 6. As presented in Fig. 6, our proposed FEDER achieves the smallest FLOPs and parameters compared with other state-of-the-art deep learning-based COD techniques. Furthermore, our score in $F_{\beta}$ is much higher than other methods and surpasses the second best one in $5.6\%$ . + +
MethodsCOD10K (2026 images)NC4K (4121 images)
M↓Fβ↑Eφ↑Sα↑M↓Fβ↑Eφ↑Sα↑
(a)w/o DWD0.0350.6970.8610.7940.0490.7760.8730.817
(b)FW->LW0.0320.7310.8950.8220.0460.8110.8990.844
(c)w/o HFA0.0330.7280.8950.8170.0470.8050.8910.840
(d)w/o LFA0.0330.7240.8870.8160.0470.8030.8920.841
(e)UFA->GFA0.0330.7200.8940.8100.0470.8040.8900.838
(f)FEDER0.0320.7400.9000.8230.0450.8170.9050.846
+ +(a) Ablation study of DWD Component. + +
MethodsCOD10K (2026 images)NC4K (4121 images)
M↓Fβ↑Eφ↑Sα↑M↓Fβ↑Eφ↑Sα↑
(a)w/o OER0.0340.7140.8710.8040.0480.7800.8890.834
(b)FC->WGM0.0320.7310.8990.8170.0460.8140.8950.842
(c)w/o HS0.0320.7230.8850.8210.0460.7980.8920.840
(d)RB->RK20.0330.7220.8810.8110.0470.8020.8950.843
(e)RK4->RK20.0310.7420.9050.8290.0450.8160.9060.848
(f)FEDER0.0320.7400.9000.8230.0450.8170.9050.846
+ +(b) Ablation study of OER module. + +Table 2. Ablation study on $COD10k$ [4] and $NC4K$ [25], where w/o means without. (a) FW, LW, and UFA are short for fixed wavelet [40], learnable wavelet, and upsampling-based feature aggregation [51]. (b) FC, WGM, HS, RB, and RK4 are short for fixed coefficient, weighted gate mechanism, Hamiltonian system, residual block, and forth-order Runge-Kutta. The best results are marked in bold. + +![](images/40688237d6af8db53c8358514051cbdb71c4b8ec2673fc9a246fcde553a3b2eb.jpg) +Figure 7. Visual verification of the effectiveness of the proposed components. B, D, and O denote baseline, DWD, and OER. Note that $\mathrm{B + D + O}$ is our proposed FEDER. + +# 4.3. Ablation Study + +We conduct the ablation study on the two largest datasets, namely $COD10K$ and $NC4K$ . + +Effect of DWD component. We demonstrate the effectiveness of our DWD component both in visual verification (see Fig. 7) and quantitative analysis (see Tab. 2a). In Fig. 7, the networks with DWD component (Rows 3 and 5) exhibit more accurate localizations and stronger redundant information suppression capacities. In addition, we present more detailed information in Tab. 2a to verify the validity of each part in the DWD component. As presented in Tab. 2a, we prove the superiority of the DWD component (in (a)), learnable wavelet-like decomposition strategy (in (b)), HFA/LFA module (in (c) and (d)), and GFA module (in (e)). + +Effect of OER module. The efficacy of our OER module is demonstrated visually by Figs. 7 and 8. In Fig. 7, methods with the OER module can generate the prediction maps with clearer edges. Besides, Fig. 8 illustrates the advancement of our OER module in generating accurate and clear boundary information. We provide detailed information about the superiority of our OER module in Tab. 2b. Specifically, (b) and (c) verify the effectiveness of our weighted gate mechanism (learnable coefficient) and Hamiltonian system. We further compare the performance of the OER module with + +![](images/3f0873bdd35e953d7658e995808541f94246ae4772f04327e94e2abc3cdf0dfd.jpg) +Figure 8. Visualizations of the reconstructed edge from the OER module, where the edges are marked in red for well-read. + +different Runge-Kutta methods, i.e., RB (RK1), RK2, and RK4. Notably, the COD results with RK2 significantly outperform those with RB and are slightly lower than that with RK4. Therefore, we integrate RK2 into our OER module for a balance of performance and efficiency. + +# 5. Conclusions + +To address the IS and ED challenges, in this paper, we propose the FEDER model for COD. Specifically, we decompose the features into different frequency bands with learnable wavelets and filter out the most informative bands to excavate the subtle discriminative features with the HFA, LFA, and GFA modules, thereby solving the IS challenge. Besides, we propose to learn an auxiliary edge reconstruction task with our OER module to generate complete edges. Learning this auxiliary task along with the COD task thus facilitates the generation of precise segmentation results with accurate object boundaries, thus mitigating the ED challenge. Extensive experiments verify the superiority of our FEDER model in comparison with other SOTAs. + +Acknowledgements: This research is partly supported by the National Key R&D Program of China (Grant No. 2020AAA0108303), and Shenzhen Science and Technology Project (Grant No. JCYJ20200109143041798) & Shenzhen Stable Supporting Program (WDZC20200820200655001) & Shenzhen Key Laboratory of next generation interactive media innovative technology (Grant No. ZDSYS20210623092001004). + +# References + +[1] Bo Chang, Lili Meng, Eldad Haber, Lars Ruthotto, David Begert, and Elliot Holtham. Reversible architectures for arbitrarily deep residual neural networks. In AAAI, 2018. 3, 5 +[2] Xuelian Cheng, Huan Xiong, Deng-Ping Fan, Yiran Zhong, Mehrtash Harandi, Tom Drummond, and Zongyuan Ge. Implicit motion handling for video camouflaged object detection. In CVPR, pages 13864-13873, 2022. 6 +[3] Deng-Ping Fan, Ming-Ming Cheng, Yun Liu, Tao Li, and Ali Borji. Structure-measure: A new way to evaluate foreground maps. In ICCV, pages 4548-4557, 2017. 6 +[4] Deng-Ping Fan, Ge-Peng Ji, Ming-Ming Cheng, and Ling Shao. Concealed object detection. IEEE Trans. Pattern Anal. Mach. Intell., 2021. 3, 6, 7, 8 +[5] Deng-Ping Fan, Ge-Peng Ji, Xuebin Qin, and Ming-Ming Cheng. Cognitive vision inspired object segmentation metric and loss function. Sci. Sin. Inf., 6:6, 2021. 6 +[6] Deng-Ping Fan, Ge-Peng Ji, Guolei Sun, Ming-Ming Cheng, Jianbing Shen, and Ling Shao. Camouflaged object detection. In CVPR, pages 2777-2787, 2020. 2, 3, 6 +[7] Shang-Hua Gao, Ming-Ming Cheng, Kai Zhao, Xin-Yu Zhang, Ming-Hsuan Yang, and Philip Torr. Res2net: A new multi-scale backbone architecture. IEEE Trans. Pattern Anal. Mach. Intell., 43(2):652-662, 2019. 3 +[8] Wooseok Ha, Chandan Singh, Francois Lanusse, Srigokul Upadhyayula, and Bin Yu. Adaptive wavelet distillation from neural networks through interpretations. NIPS, 34:20669-20682, 2021. 3, 4 +[9] Eldad Haber and Lars Ruthotto. Stable architectures for deep neural networks. Inverse Probl, 34(1):014004, 2017. 5, 6 +[10] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, pages 770-778, 2016. 3 +[11] Jianqin Yin Yanbin Han Wendi Hou and Jinping Li. Detection of the mobile object with camouflage color under dynamic background based on optical flow. Procedia Eng., 15:2201-2205, 2011. 2 +[12] Sergey Ioffe and Christian Szegedy. Batch normalization: Accelerating deep network training by reducing internal covariate shift. In ICML, pages 448-456. PMLR, 2015. 4 +[13] Ge-Peng Ji, Lei Zhu, Mingchen Zhuge, and Keren Fu. Fast camouflaged object detection via edge-based reversible recalibration network. Pattern Recognit., 123:108414, 2022. 5 +[14] Qi Jia, Shuilian Yao, Yu Liu, Xin Fan, Risheng Liu, and Zhongxuan Luo. Segment, magnify and reiterate: Detecting camouflaged objects the hard way. In CVPR, pages 4713-4722, 2022. 1, 2, 6 +[15] Liming Jiang, Bo Dai, Wayne Wu, and Chen Change Loy. Focal frequency loss for image reconstruction and synthesis. In ICCV, pages 13919-13929, 2021. 2 +[16] Zhanghan Ke, Jiayu Sun, Kaican Li, Qiong Yan, and Rynson WH Lau. Modnet: Real-time trimap-free portrait matting via objective decomposition. In AAAI, pages 1140-1147, 2022. 3 +[17] Ajay Kumar. Computer-vision-based fabric defect detection: A survey. IEEE Trans. Ind. Electron., 55(1):348-363, 2008. 1 + +[18] Trung-Nghia Le, Tam V Nguyen, Zhongliang Nie, Minh-Triet Tran, and Akihiro Sugimoto. Anabranch network for camouflaged object segmentation. Comput. Vis. Image Underst., 184:45-56, 2019. 6 +[19] Bei Li, Quan Du, Tao Zhou, Yi Jing, Shuhan Zhou, Xin Zeng, Tong Xiao, Jingbo Zhu, Xuebo Liu, and Min Zhang. Ode transformer: An ordinary differential equation-inspired model for sequence generation. In ACL, pages 8335-8351, 2022. 3 +[20] Boyi Li, Felix Wu, Kilian Q Weinberger, and Serge Belongie. Positional normalization. NIPS, 32, 2019. 4 +[21] Juncheng Li, Faming Fang, Kangfu Mei, and Guixu Zhang. Multi-scale residual network for image super-resolution. In ECCV, pages 517-532, 2018. 4 +[22] Kai Li, Yulun Zhang, Kunpeng Li, and Yun Fu. Adversarial feature hallucination networks for few-shot learning. In CVPR, pages 13470-13479, 2020. 4 +[23] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In ICCV, pages 10012-10022, 2021. 4 +[24] Laetitia Loncan, Luis B De Almeida, José M Bioucas-Dias, Xavier Briottet, Jocelyn Chanussot, Nicolas Dobigeon, Sophie Fabre, Wenzhi Liao, Giorgio A Licciardi, Miguel Simoes, et al. Hyperspectral pansharpening: A review. IEEE Geosci. Remote Sens. Mag., 3(3):27-46, 2015. 4 +[25] Yunqiu Lv, Jing Zhang, Yuchao Dai, Aixuan Li, Bowen Liu, Nick Barnes, and Deng-Ping Fan. Simultaneously localize, segment and rank the camouflaged objects. In CVPR, pages 11591-11601, 2021. 6, 8 +[26] Yue Ma, Yali Wang, Yue Wu, Ziyu Lyu, Siran Chen, Xiu Li, and Yu Qiao. Visual knowledge graph for human action reasoning in videos. In ACM MM, pages 4132-4141, 2022. 4 +[27] Ran Margolin, Lihi Zelnik-Manor, and Ayellet Tal. How to evaluate foreground maps? In CVPR, pages 248-255, 2014. 6 +[28] Haiyang Mei, Ge-Peng Ji, Ziqi Wei, Xin Yang, Xiaopeng Wei, and Deng-Ping Fan. Camouflaged object segmentation with distraction mining. In CVPR, pages 8772-8781, 2021. 2, 6 +[29] Fausto Milletari, Nassir Navab, and Seyed-Ahmad Ahmadi. V-net: Fully convolutional neural networks for volumetric medical image segmentation. In 3DV, pages 565-571. IEEE, 2016. 6 +[30] Yuxin Pan, Yiwang Chen, Qiang Fu, Ping Zhang, and Xin Xu. Study on the camouflaged target detection method based on 3d convexity. Mod. Appl. Sci., 5(4):152, 2011. 2 +[31] Youwei Pang, Xiaoqi Zhao, Tian-Zhu Xiang, Lihe Zhang, and Huchuan Lu. Zoom in and out: A mixed-scale triplet network for camouflaged object detection. In CVPR, pages 2160-2170, 2022. 6 +[32] Jialun Pei, Tianyang Cheng, Deng-Ping Fan, He Tang, Chuanbo Chen, and Luc Van Gool. Osformer: One-stage camouflaged instance segmentation with transformers. In ECCV, pages 19-37. Springer, 2022. 6 + +[33] Natasha Price, Samuel Green, Jolyon Troscianko, Tom Tre-genza, and Martin Stevens. Background matching and disruptive coloration as habitat-specific strategies for camouflage. Sci. Rep., 9(1):1-10, 2019. 1 +[34] Xu Qin, Zhilin Wang, Yuanchao Bai, Xiaodong Xie, and Huizhu Jia. Ffa-net: Feature fusion attention network for single image dehazing. In AAAI, pages 11908-11915, 2020. 4 +[35] Dan Jeric Arcega Rustia, Chien Erh Lin, Jui-Yung Chung, Yi-Ji Zhuang, Ju-Chun Hsu, and Ta-Te Lin. Application of an image and environmental sensor network for automated greenhouse insect pest monitoring. J. Asia Pac. Entomol., 23(1):17-28, 2020. 1 +[36] Phil Sallee and Bruno Olshausen. Learning sparse multiscale image representations. NIPS, 15, 2002. 4 +[37] Jiawei Shen, Zhuoyan Li, Lei Yu, Gui-Song Xia, and Wen Yang. Implicit euler ode networks for single-image dehazing. In CVPRW, pages 218-219, 2020. 3 +[38] Chenyang Si, Weihao Yu, Pan Zhou, Yichen Zhou, Xinchao Wang, and Shuicheng Yan. Inception transformer. arXiv preprint arXiv:2205.12956, 2022. 5 +[39] Przemysław Skurowski, Hassan Abdulameer, J Błaszczyk, Tomasz Depta, Adam Kornacki, and P Koziel. Animal camouflage analysis: Chameleon database. Unpublished manuscript, 2(6):7, 2018. 6 +[40] Radomir S Stanković and Bogdan J Falkowski. The haar wavelet transform: its status and achievements. Comput. Electr. Eng., 29(1):25-44, 2003. 4, 8 +[41] Martin Stevens and Sami Merilaita. Animal camouflage: current issues and new perspectives. Philos. Trans. R. Soc. Lond., B, Biol. Sci., 364(1516):423-427, 2009. 1, 2, 3 +[42] Cheng Tai and E Weinan. Multiscale adaptive representation of signals: I. the basic framework. J. Mach. Learn. Res., 17(1):4875-4912, 2016. 4 + +[43] Christopher Torrence and Gilbert P Compo. A practical guide to wavelet analysis. Bull. Am. Meteorol. Soc., 79(1):61-78, 1998. 4 +[44] Dmitry Ulyanov, Andrea Vedaldi, and Victor Lempitsky. Instance normalization: The missing ingredient for fast stylization. arXiv preprint arXiv:1607.08022, 2016. 4 +[45] Qilong Wang, Banggu Wu, Pengfei Zhu, Peihua Li, Wangmeng Zuo, and Qinghua Hu. Eca-net: Efficient channel attention for deep convolutional neural networks. In CVPR, pages 13-19, 2020. 4 +[46] E Weinan. A proposal on machine learning via dynamical systems. Commun. Math. Stat., 1(5):1-11, 2017. 3, 5 +[47] Zhe Wu, Li Su, and Qingming Huang. Cascaded partial decoder for fast and accurate salient object detection. In CVPR, pages 3907-3916, 2019. 6 +[48] Fan Yang, Qiang Zhai, Xin Li, Rui Huang, Ao Luo, Hong Cheng, and Deng-Ping Fan. Uncertainty-guided transformer reasoning for camouflaged object detection. In ICCV, pages 4146-4155, 2021. 6 +[49] Jaejun Yoo, Youngjung Uh, Sanghyuk Chun, Byeongkyu Kang, and Jung-Woo Ha. Photorealistic style transfer via wavelet transforms. In ICCV, pages 9036-9045, 2019. 3 +[50] Qiang Zhai, Xin Li, Fan Yang, Chenglizhao Chen, Hong Cheng, and Deng-Ping Fan. Mutual graph learning for camouflaged object detection. In CVPR, pages 12997-13007, 2021. 5, 6 +[51] Yulun Zhang, Zhifei Zhang, Stephen DiVerdi, Zhaowen Wang, Jose Echevarria, and Yun Fu. Texture hallucination for large-factor painting super-resolution. In ECCV, pages 209-225. Springer, 2020. 8 +[52] Hongwei Zhu, Peng Li, Haoran Xie, Xuefeng Yan, Dong Liang, Dapeng Chen, Mingqiang Wei, and Jing Qin. I can find you! boundary-guided separated attention network for camouflaged object detection. In AAAI, 2022. 6 \ No newline at end of file diff --git a/2023/Camouflaged Object Detection With Feature Decomposition and Edge Reconstruction/images.zip b/2023/Camouflaged Object Detection With Feature Decomposition and Edge Reconstruction/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..115dc11203b2cbef62102770b5c451f9a586d159 --- /dev/null +++ b/2023/Camouflaged Object Detection With Feature Decomposition and Edge Reconstruction/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c87f259d94bf26e5265e580de4615d5123d822fd0f61817d48023aa3cb903d6a +size 916286 diff --git a/2023/Camouflaged Object Detection With Feature Decomposition and Edge Reconstruction/layout.json b/2023/Camouflaged Object Detection With Feature Decomposition and Edge Reconstruction/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..880f6ea0b47f3b4b1c6d4f9eecf119f8cff4ba04 --- /dev/null +++ b/2023/Camouflaged Object Detection With Feature Decomposition and Edge Reconstruction/layout.json @@ -0,0 +1,9648 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 100, + 103, + 493, + 138 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 100, + 103, + 493, + 138 + ], + "spans": [ + { + "bbox": [ + 100, + 103, + 493, + 138 + ], + "type": "text", + "content": "Camouflaged Object Detection with Feature Decomposition and Edge Reconstruction" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 160, + 545, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 160, + 545, + 203 + ], + "spans": [ + { + "bbox": [ + 53, + 160, + 545, + 203 + ], + "type": "text", + "content": "Chunming He" + }, + { + "bbox": [ + 53, + 160, + 545, + 203 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 53, + 160, + 545, + 203 + ], + "type": "text", + "content": ", Kai Li" + }, + { + "bbox": [ + 53, + 160, + 545, + 203 + ], + "type": "inline_equation", + "content": "^{2*}" + }, + { + "bbox": [ + 53, + 160, + 545, + 203 + ], + "type": "text", + "content": ", Yachao Zhang" + }, + { + "bbox": [ + 53, + 160, + 545, + 203 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 53, + 160, + 545, + 203 + ], + "type": "text", + "content": ", Longxiang Tang" + }, + { + "bbox": [ + 53, + 160, + 545, + 203 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 53, + 160, + 545, + 203 + ], + "type": "text", + "content": ", Yulun Zhang" + }, + { + "bbox": [ + 53, + 160, + 545, + 203 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 53, + 160, + 545, + 203 + ], + "type": "text", + "content": ", Zhenhua Guo" + }, + { + "bbox": [ + 53, + 160, + 545, + 203 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 53, + 160, + 545, + 203 + ], + "type": "text", + "content": ", Xiu Li" + }, + { + "bbox": [ + 53, + 160, + 545, + 203 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 53, + 160, + 545, + 203 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 53, + 160, + 545, + 203 + ], + "type": "text", + "content": "Shenzhen International Graduate School, Tsinghua University, " + }, + { + "bbox": [ + 53, + 160, + 545, + 203 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 53, + 160, + 545, + 203 + ], + "type": "text", + "content": "NEC Laboratories America, " + }, + { + "bbox": [ + 53, + 160, + 545, + 203 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 53, + 160, + 545, + 203 + ], + "type": "text", + "content": "ETH Zürich, " + }, + { + "bbox": [ + 53, + 160, + 545, + 203 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 53, + 160, + 545, + 203 + ], + "type": "text", + "content": "Tianyi Traffic Technology" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 119, + 205, + 476, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 119, + 205, + 476, + 217 + ], + "spans": [ + { + "bbox": [ + 119, + 205, + 476, + 217 + ], + "type": "text", + "content": "{chunminghe19990224,li.gml.kai,lloong.x,yulun100,cszguo}@gmail.com" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 153, + 220, + 436, + 231 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 153, + 220, + 436, + 231 + ], + "spans": [ + { + "bbox": [ + 153, + 220, + 436, + 231 + ], + "type": "text", + "content": "yachaozhang@stu.xmu.edu.cn, li.xiu@sz.tsinghua.edu.cn" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 143, + 258, + 191, + 271 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 258, + 191, + 271 + ], + "spans": [ + { + "bbox": [ + 143, + 258, + 191, + 271 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 284, + 290, + 631 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 284, + 290, + 631 + ], + "spans": [ + { + "bbox": [ + 47, + 284, + 290, + 631 + ], + "type": "text", + "content": "Camouflaged object detection (COD) aims to address the tough issue of identifying camouflaged objects visually blended into the surrounding backgrounds. COD is a challenging task due to the intrinsic similarity of camouflaged objects with the background, as well as their ambiguous boundaries. Existing approaches to this problem have developed various techniques to mimic the human visual system. Albeit effective in many cases, these methods still struggle when camouflaged objects are so deceptive to the vision system. In this paper, we propose the FEature Decomposition and Edge Reconstruction (FEDER) model for COD. The FEDER model addresses the intrinsic similarity of foreground and background by decomposing the features into different frequency bands using learnable wavelets. It then focuses on the most informative bands to mine subtle cues that differentiate foreground and background. To achieve this, a frequency attention module and a guidance-based feature aggregation module are developed. To combat the ambiguous boundary problem, we propose to learn an auxiliary edge reconstruction task alongside the COD task. We design an ordinary differential equation-inspired edge reconstruction module that generates exact edges. By learning the auxiliary task in conjunction with the COD task, the FEDER model can generate precise prediction maps with accurate object boundaries. Experiments show that our FEDER model significantly outperforms state-of-the-art methods with cheaper computational and memory costs. The code will be available at https://github.com/ChunmingHe/FEDER." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 652, + 127, + 664 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 652, + 127, + 664 + ], + "spans": [ + { + "bbox": [ + 48, + 652, + 127, + 664 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 672, + 287, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 672, + 287, + 696 + ], + "spans": [ + { + "bbox": [ + 47, + 672, + 287, + 696 + ], + "type": "text", + "content": "Camouflaged object detection (COD) aims to detect and segment objects \"seamlessly\" integrated into surrounding" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 310, + 259, + 542, + 452 + ], + "blocks": [ + { + "bbox": [ + 310, + 259, + 542, + 452 + ], + "lines": [ + { + "bbox": [ + 310, + 259, + 542, + 452 + ], + "spans": [ + { + "bbox": [ + 310, + 259, + 542, + 452 + ], + "type": "image", + "image_path": "2ceb23f8e240956ace12da1047172462b49289f379c43fd2be9387783d0522ab.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 336, + 452, + 369, + 461 + ], + "lines": [ + { + "bbox": [ + 336, + 452, + 369, + 461 + ], + "spans": [ + { + "bbox": [ + 336, + 452, + 369, + 461 + ], + "type": "text", + "content": "(a) Origin" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 395, + 452, + 417, + 460 + ], + "lines": [ + { + "bbox": [ + 395, + 452, + 417, + 460 + ], + "spans": [ + { + "bbox": [ + 395, + 452, + 417, + 460 + ], + "type": "text", + "content": "(b) GT" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 440, + 452, + 481, + 460 + ], + "lines": [ + { + "bbox": [ + 440, + 452, + 481, + 460 + ], + "spans": [ + { + "bbox": [ + 440, + 452, + 481, + 460 + ], + "type": "text", + "content": "(c) SegMaR" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 494, + 452, + 533, + 460 + ], + "lines": [ + { + "bbox": [ + 494, + 452, + 533, + 460 + ], + "spans": [ + { + "bbox": [ + 494, + 452, + 533, + 460 + ], + "type": "text", + "content": "(d) FEDER" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 306, + 463, + 545, + 497 + ], + "lines": [ + { + "bbox": [ + 306, + 463, + 545, + 497 + ], + "spans": [ + { + "bbox": [ + 306, + 463, + 545, + 497 + ], + "type": "text", + "content": "Figure 1. Results of SegMaR [14] and our method under the intrinsic similarity (IS) and edge disruption (ED) challenges. Our method better localizes the objects and produces clearer edges." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 510, + 545, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 510, + 545, + 617 + ], + "spans": [ + { + "bbox": [ + 304, + 510, + 545, + 617 + ], + "type": "text", + "content": "environments. COD is a challenging task as it needs to combat against excellent camouflage strategies, including background matching [41], disruptive coloration [33], etc., and distinguish the subtle differences between candidate objects and their backgrounds. Research in COD can simultaneously facilitate the development of visual perception for nuance discrimination and promote various valuable real-life applications, ranging from concealed defect detection [17] in industry to pest monitoring [35] in agriculture." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "content": "COD faces two main challenges. The first is the intrinsic similarity (IS) challenge, which occurs when camouflaged objects share similar colors and patterns with their backgrounds. This makes it difficult to even roughly localize those camouflaged objects. The second is the edge disruption (ED) challenge, which arises from extremely ambiguous object boundaries. Even if a rough localization is achieved, precise segmentation can barely be obtained." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 702, + 135, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 702, + 135, + 712 + ], + "spans": [ + { + "bbox": [ + 58, + 702, + 135, + 712 + ], + "type": "text", + "content": "*Corresponding author." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "22046" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 251 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 251 + ], + "type": "text", + "content": "To tackle these challenges, most existing works aim to develop models that mimic the human visual system [6,28]. However, since camouflage strategies are designed by prey to confuse the predator's visual system, and the intrinsic topological properties of candidate objects are not distinctive, such human perception-oriented attempts may struggle to identify subtle discriminative features and fail to effectively address the above challenges. For instance, as illustrated in Fig. 1, the state-of-the-art human perception-based COD method can only generate inaccurate prediction maps, such as the vague caddisfly and incomplete dog (Row 2 and 4), or even fail to detect camouflaged objects like the bird and snake (Row 1 and 3). Therefore, a better COD method should compensate for the \"flaw\" in human perception by emphasizing subtle discriminative features." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 252, + 289, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 252, + 289, + 361 + ], + "spans": [ + { + "bbox": [ + 46, + 252, + 289, + 361 + ], + "type": "text", + "content": "Based on the biological study [41], camouflaged objects often employ various camouflage strategies to conceal their discriminative differences, which mainly exist in texture details and global information distribution, within surrounding environments. Such a study inspires us to cope with the COD task by decomposing the camouflage scenario into different parts. This allows for the disentanglement of various intricate connections, enabling each part to be separately handled to fully excavate subtle discriminative cues." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 361, + 289, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 361, + 289, + 552 + ], + "spans": [ + { + "bbox": [ + 46, + 361, + 289, + 552 + ], + "type": "text", + "content": "With this inspiration, we propose the FEature Decomposition and Edge Reconstruction (FEDER) model for the COD task, which compensates for the deficiencies of human perception by emphasizing subtle discriminative features and effectively addresses the IS and ED challenges. Specifically, to combat the intractable localization problem caused by the IS challenge, we design the deep wavelet-like decomposition (DWD) strategy, which decomposes the extracted features into different frequency bands using learnable wavelet-like modules. Then, we focus on the most informative bands by filtering out noteworthy parts where discriminative cues are most likely to exist by a novel frequency attention (FA) module. Moreover, a guidance-based feature aggregation (GFA) module is proposed to aggregate the multi-scale decomposed features with attentional guidance to further emphasize discriminative information." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 552, + 289, + 672 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 552, + 289, + 672 + ], + "spans": [ + { + "bbox": [ + 46, + 552, + 289, + 672 + ], + "type": "text", + "content": "To address the ambiguous boundary problem of the ED challenge, we propose learning an auxiliary edge reconstruction task to encourage the network to excavate edge details. We design the ordinary differential equation (ODE)-inspired edge reconstruction (OER) module to reconstruct accurate and complete edge prediction maps using a high-order ODE solver, specifically, the second-order Runge-Kutta. Incorporating this auxiliary task with the COD task can facilitate the generation of precise segmentation results with accurate object boundaries." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 59, + 673, + 244, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 673, + 244, + 685 + ], + "spans": [ + { + "bbox": [ + 59, + 673, + 244, + 685 + ], + "type": "text", + "content": "Our contributions are summarized as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 57, + 689, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 689, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 57, + 689, + 288, + 714 + ], + "type": "text", + "content": "- We propose the FEature Decomposition and Edge Reconstruction (FEDER) model for the COD task. To" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 324, + 72, + 545, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 324, + 72, + 545, + 96 + ], + "spans": [ + { + "bbox": [ + 324, + 72, + 545, + 96 + ], + "type": "text", + "content": "the best of our knowledge, we are the first to approach COD from a decomposition perspective." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 316, + 98, + 545, + 258 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 316, + 98, + 545, + 158 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 98, + 545, + 158 + ], + "spans": [ + { + "bbox": [ + 316, + 98, + 545, + 158 + ], + "type": "text", + "content": "- To highlight the subtle discriminative features, we propose frequency attention modules to filter out the noteworthy parts of corresponding features and design the Guidance-based Feature Aggregation module to aggregate the multi-scale features with attentional guidance." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 316, + 161, + 545, + 220 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 161, + 545, + 220 + ], + "spans": [ + { + "bbox": [ + 316, + 161, + 545, + 220 + ], + "type": "text", + "content": "- We propose to learn an auxiliary edge reconstruction task along with the COD task to help generate precise segmentation maps with accurate object boundaries and design the ODE-inspired edge reconstruction module for complete edge prediction." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 316, + 222, + 545, + 258 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 222, + 545, + 258 + ], + "spans": [ + { + "bbox": [ + 316, + 222, + 545, + 258 + ], + "type": "text", + "content": "- The proposed FEDER significantly outperforms the state-of-the-art methods on four datasets by a large margin with cheaper computational and memory costs." + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 274, + 398, + 288 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 274, + 398, + 288 + ], + "spans": [ + { + "bbox": [ + 306, + 274, + 398, + 288 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 294, + 545, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 294, + 545, + 666 + ], + "spans": [ + { + "bbox": [ + 304, + 294, + 545, + 666 + ], + "type": "text", + "content": "Camouflaged object detection. Unlike existing object detection tasks, camouflaged object detection (COD) poses new challenges for mining subtle discriminative features under complex camouflage strategies [6, 11]. Early techniques utilized the hand-crafted operators for COD [11, 30], which were only applicable to camouflaged scenarios with simple backgrounds. Recent research has leveraged the huge capacity of deep learning to detect camouflaged objects in a learning manner [6, 14, 28]. Inspired by the hunting process of predators, SINet [6] designed a bio-inspired network to gradually search and locate the camouflaged object. PFNet [28] proposed the position module and focus module to imitate human identification with the distraction mining strategy. By simulating human behaviors in understanding complex scenarios, SegMaR [14] integrated segment, magnify and reiterate in a coarse-to-fine manner using the multistage strategy. However, these COD solutions mainly focus on mimicking biovision systems, which can be easily confused by complex camouflaged strategies and struggle to excavate the subtle discriminative features, thus failing to handle the IS and ED challenges (see Fig. 1). Unlike these human perception-oriented techniques, we first propose to address the COD task from a decomposition perspective by decomposing the extracted features into different frequency bands with learnable wavelets and filtering out the most informative bands to excavate those inconspicuous discriminative features, thus remedying the human visual deficiency and solving the IS challenge. To handle the ED challenge, we propose learning an auxiliary edge reconstruction task along with the COD task to facilitate the generation of precise segmentation results with clear object boundaries." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 666, + 545, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 666, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 666, + 545, + 715 + ], + "type": "text", + "content": "Deep wavelet decomposition. Deep wavelet decomposition is an effective tool to decompose image/feature into various frequency components and has gained immense popularity in many domains, such as image restoration [15]" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "22047" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 70, + 545, + 313 + ], + "blocks": [ + { + "bbox": [ + 50, + 70, + 545, + 313 + ], + "lines": [ + { + "bbox": [ + 50, + 70, + 545, + 313 + ], + "spans": [ + { + "bbox": [ + 50, + 70, + 545, + 313 + ], + "type": "image", + "image_path": "e1db8a771e890aa816d72c7592408874967dd9d3e4bfbfbba41e6d5092702720.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 316, + 546, + 351 + ], + "lines": [ + { + "bbox": [ + 46, + 316, + 546, + 351 + ], + "spans": [ + { + "bbox": [ + 46, + 316, + 546, + 351 + ], + "type": "text", + "content": "Figure 2. Overview of the proposed FEDER model. CFE and SED are short for camouflaged feature encoder and segmentation-oriented edge-assisted decoder. HFA/LFA and GFA denote high/low frequency attention modules and the guidance-based feature aggregation module. RRS and OER indicate the reversible re-calibration segmentation module and the ODE-inspired edge reconstruction module." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 365, + 287, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 365, + 287, + 426 + ], + "spans": [ + { + "bbox": [ + 46, + 365, + 287, + 426 + ], + "type": "text", + "content": "and style transfer [49]. To handle the IS challenge, we introduce deep wavelet decomposition into the COD task. Furthermore, to better accommodate the COD data, we employ the learnable wavelets for deep adaptive feature decomposition, whose coefficients are updated following AWD [8]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 426, + 288, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 426, + 288, + 582 + ], + "spans": [ + { + "bbox": [ + 46, + 426, + 288, + 582 + ], + "type": "text", + "content": "ODE-inspired network. Researchers have established a relationship between ODE and neural networks. [46] first analyzed ResNet from the perspective of discrete ODE and [1] further extended ResNet to an ODE-inspired network architecture with a more accurate transmission. Since then, ODE-inspired networks are widely utilized in many fields, such as image dehazing [37] and machine translation [19]. In this paper, to accommodate the fine-grained property of the edge, we propose an ODE-inspired edge reconstruction module with the second-order Runge-Kutta and a weighted gate mechanism, aiming to generate more accurate boundaries. Furthermore, we apply the Hamiltonian system to our OER module to ensure the stability of edge reconstruction." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 597, + 129, + 609 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 597, + 129, + 609 + ], + "spans": [ + { + "bbox": [ + 47, + 597, + 129, + 609 + ], + "type": "text", + "content": "3. Methodology" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 618, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 618, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 618, + 287, + 715 + ], + "type": "text", + "content": "Given a camouflaged image, we first extract a cascade of features using the camouflaged feature encoder (CFE). We then perform a wavelet-like decomposition (DWD) on the features to decompose them into different frequency bands. We select the most informative bands, such as the high-frequency and low-frequency components, for further analysis. These informative bands are processed by the frequency attention (FA) module and guidance-based feature" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 365, + 545, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 365, + 545, + 426 + ], + "spans": [ + { + "bbox": [ + 304, + 365, + 545, + 426 + ], + "type": "text", + "content": "aggregation (GFA) module to highlight the inconspicuous discriminative features. With the aggregated features, the segmentation-oriented edge-assisted decoder (SED) outputs both the segmentation map and the edge prediction map. Fig. 2 presents the framework of our FEDER model." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 434, + 504, + 447 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 434, + 504, + 447 + ], + "spans": [ + { + "bbox": [ + 305, + 434, + 504, + 447 + ], + "type": "text", + "content": "3.1. Camouflaged Feature Encoder (CFE)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 452, + 545, + 596 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 452, + 545, + 596 + ], + "spans": [ + { + "bbox": [ + 304, + 452, + 545, + 596 + ], + "type": "text", + "content": "Following SINet V2 [4], the basic encoder " + }, + { + "bbox": [ + 304, + 452, + 545, + 596 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 304, + 452, + 545, + 596 + ], + "type": "text", + "content": " adopts ResNet50 [10]/Res2Net50 [7] as its backbone. Given an image " + }, + { + "bbox": [ + 304, + 452, + 545, + 596 + ], + "type": "inline_equation", + "content": "I_{c}" + }, + { + "bbox": [ + 304, + 452, + 545, + 596 + ], + "type": "text", + "content": " of size " + }, + { + "bbox": [ + 304, + 452, + 545, + 596 + ], + "type": "inline_equation", + "content": "W\\times H" + }, + { + "bbox": [ + 304, + 452, + 545, + 596 + ], + "type": "text", + "content": ", the basic encoder " + }, + { + "bbox": [ + 304, + 452, + 545, + 596 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 304, + 452, + 545, + 596 + ], + "type": "text", + "content": " generates a set of feature maps " + }, + { + "bbox": [ + 304, + 452, + 545, + 596 + ], + "type": "inline_equation", + "content": "\\{f_k\\}_{k = 0}^4" + }, + { + "bbox": [ + 304, + 452, + 545, + 596 + ], + "type": "text", + "content": " with the resolution of " + }, + { + "bbox": [ + 304, + 452, + 545, + 596 + ], + "type": "inline_equation", + "content": "\\frac{H}{2^{k + 1}}\\times \\frac{W}{2^{k + 1}}" + }, + { + "bbox": [ + 304, + 452, + 545, + 596 + ], + "type": "text", + "content": ", R-Net [6] is cascaded to transform " + }, + { + "bbox": [ + 304, + 452, + 545, + 596 + ], + "type": "inline_equation", + "content": "\\{f_k\\}_{k = 1}^4" + }, + { + "bbox": [ + 304, + 452, + 545, + 596 + ], + "type": "text", + "content": " into a more informative and compact output, i.e., a series of 64-channel feature maps " + }, + { + "bbox": [ + 304, + 452, + 545, + 596 + ], + "type": "inline_equation", + "content": "\\{f_k^r\\}_{k = 1}^4" + }, + { + "bbox": [ + 304, + 452, + 545, + 596 + ], + "type": "text", + "content": ". Additionally, the last feature map " + }, + { + "bbox": [ + 304, + 452, + 545, + 596 + ], + "type": "inline_equation", + "content": "f_{4}" + }, + { + "bbox": [ + 304, + 452, + 545, + 596 + ], + "type": "text", + "content": " from the basic encoder " + }, + { + "bbox": [ + 304, + 452, + 545, + 596 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 304, + 452, + 545, + 596 + ], + "type": "text", + "content": " is further fed into an efficient atrous spatial pyramid pooling (e-ASPP) " + }, + { + "bbox": [ + 304, + 452, + 545, + 596 + ], + "type": "inline_equation", + "content": "A_{e}" + }, + { + "bbox": [ + 304, + 452, + 545, + 596 + ], + "type": "text", + "content": " [16] to enlarge the receptive field and fuse the multi-context information, resulting in " + }, + { + "bbox": [ + 304, + 452, + 545, + 596 + ], + "type": "inline_equation", + "content": "d_5^s = A_e(f_4)" + }, + { + "bbox": [ + 304, + 452, + 545, + 596 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 452, + 545, + 596 + ], + "type": "inline_equation", + "content": "d_5^s" + }, + { + "bbox": [ + 304, + 452, + 545, + 596 + ], + "type": "text", + "content": " is a coarse segmentation result with the same spatial resolution as " + }, + { + "bbox": [ + 304, + 452, + 545, + 596 + ], + "type": "inline_equation", + "content": "f_{4}" + }, + { + "bbox": [ + 304, + 452, + 545, + 596 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 604, + 485, + 617 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 604, + 485, + 617 + ], + "spans": [ + { + "bbox": [ + 306, + 604, + 485, + 617 + ], + "type": "text", + "content": "3.2. Deep Wavelet-like Decomposition" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 623, + 504, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 623, + 504, + 635 + ], + "spans": [ + { + "bbox": [ + 306, + 623, + 504, + 635 + ], + "type": "text", + "content": "3.2.1 Learnable Wavelet-like Decomposition" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 641, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 641, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 641, + 545, + 713 + ], + "type": "text", + "content": "Camouflaged objects share a high intrinsic similarity with the background, which poses challenges for common feature extractors to mine the inconspicuous discriminative features, ultimately resulting in the suppression of segmentation performance. Based on the biological study [41], the discriminative features of COD mainly exist in the high-" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "22048" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 72, + 220, + 134 + ], + "blocks": [ + { + "bbox": [ + 50, + 72, + 220, + 134 + ], + "lines": [ + { + "bbox": [ + 50, + 72, + 220, + 134 + ], + "spans": [ + { + "bbox": [ + 50, + 72, + 220, + 134 + ], + "type": "image", + "image_path": "6d23091521cfbf166c2637d633254fb290ef33cf79932c2078161c5135671d3e.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 196, + 137, + 395, + 148 + ], + "lines": [ + { + "bbox": [ + 196, + 137, + 395, + 148 + ], + "spans": [ + { + "bbox": [ + 196, + 137, + 395, + 148 + ], + "type": "text", + "content": "Figure 3. Details of the proposed GFA, HFA, and LFA" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 222, + 72, + 389, + 133 + ], + "blocks": [ + { + "bbox": [ + 222, + 72, + 389, + 133 + ], + "lines": [ + { + "bbox": [ + 222, + 72, + 389, + 133 + ], + "spans": [ + { + "bbox": [ + 222, + 72, + 389, + 133 + ], + "type": "image", + "image_path": "0197348a3e691a59e03b998b156935360fa977a868b609fd2185dcc4a5f125bd.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 391, + 72, + 544, + 134 + ], + "blocks": [ + { + "bbox": [ + 391, + 72, + 544, + 134 + ], + "lines": [ + { + "bbox": [ + 391, + 72, + 544, + 134 + ], + "spans": [ + { + "bbox": [ + 391, + 72, + 544, + 134 + ], + "type": "image", + "image_path": "7f6f84ff9fa0dff726a61ec5c2430ab3398798849268bc5678813f0ee5ea922c.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 161, + 287, + 233 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 161, + 287, + 233 + ], + "spans": [ + { + "bbox": [ + 46, + 161, + 287, + 233 + ], + "type": "text", + "content": "frequency (HF) component, e.g., texture and edge, and the low-frequency (LF) component, e.g., color and illumination. Inspired by this study, we propose to perform deep wavelet-like decomposition (DWD) on the extracted features " + }, + { + "bbox": [ + 46, + 161, + 287, + 233 + ], + "type": "inline_equation", + "content": "\\{f_k^r\\}_{k=1}^4" + }, + { + "bbox": [ + 46, + 161, + 287, + 233 + ], + "type": "text", + "content": " and select the most informative HF and LF components for further refinement. We decompose " + }, + { + "bbox": [ + 46, + 161, + 287, + 233 + ], + "type": "inline_equation", + "content": "f_k^r" + }, + { + "bbox": [ + 46, + 161, + 287, + 233 + ], + "type": "text", + "content": " as" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 73, + 235, + 287, + 248 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 73, + 235, + 287, + 248 + ], + "spans": [ + { + "bbox": [ + 73, + 235, + 287, + 248 + ], + "type": "interline_equation", + "content": "\\left(f _ {k} ^ {r}\\right) _ {H F} = W _ {H F} \\left(f _ {k} ^ {r}\\right), \\left(f _ {k} ^ {r}\\right) _ {L F} = W _ {L F} \\left(f _ {k} ^ {r}\\right), \\tag {1}", + "image_path": "dc6d7c6b2f35055175a5f9d50c35efe117e9534509b194f53f7945c2c55d74f8.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 249, + 287, + 333 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 249, + 287, + 333 + ], + "spans": [ + { + "bbox": [ + 46, + 249, + 287, + 333 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 249, + 287, + 333 + ], + "type": "inline_equation", + "content": "(f_k^r)_{HF}" + }, + { + "bbox": [ + 46, + 249, + 287, + 333 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 249, + 287, + 333 + ], + "type": "inline_equation", + "content": "(f_k^r)_{LF}" + }, + { + "bbox": [ + 46, + 249, + 287, + 333 + ], + "type": "text", + "content": " denote the HF and LF components of " + }, + { + "bbox": [ + 46, + 249, + 287, + 333 + ], + "type": "inline_equation", + "content": "f_k^r" + }, + { + "bbox": [ + 46, + 249, + 287, + 333 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 46, + 249, + 287, + 333 + ], + "type": "inline_equation", + "content": "W_{HF}" + }, + { + "bbox": [ + 46, + 249, + 287, + 333 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 249, + 287, + 333 + ], + "type": "inline_equation", + "content": "W_{LF}" + }, + { + "bbox": [ + 46, + 249, + 287, + 333 + ], + "type": "text", + "content": " represent the learnable HF and LF filters with the coefficients updated following AWD [8] and initialized by Haar wavelet [40]. The learned wavelet-like transformer is expected to better cater to COD data than manually-designed wavelets [36, 42], thus further facilitating the extraction of inconspicuous discriminative features." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 358, + 208, + 370 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 358, + 208, + 370 + ], + "spans": [ + { + "bbox": [ + 47, + 358, + 208, + 370 + ], + "type": "text", + "content": "3.2.2 Frequency Attention Modules" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 380, + 287, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 380, + 287, + 441 + ], + "spans": [ + { + "bbox": [ + 46, + 380, + 287, + 441 + ], + "type": "text", + "content": "To extract discriminative information from the decomposed features, we propose a high-frequency attention (HFA) module and a low-frequency attention (LFA) module, corresponding to HF and LF bands, respectively. The detailed structures of the two modules are illustrated in Fig. 3." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 443, + 287, + 563 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 443, + 287, + 563 + ], + "spans": [ + { + "bbox": [ + 46, + 443, + 287, + 563 + ], + "type": "text", + "content": "High-frequency attention module. We design the HFA module to accentuate those texture-rich regions for subtle discriminative feature extraction. Following [21, 26], we first apply a residual block for texture preservation, consisting of a " + }, + { + "bbox": [ + 46, + 443, + 287, + 563 + ], + "type": "inline_equation", + "content": "3 \\times 3" + }, + { + "bbox": [ + 46, + 443, + 287, + 563 + ], + "type": "text", + "content": " convolution layer, batch normalization (BN) [12], and ReLU. We then employ the joint attention module " + }, + { + "bbox": [ + 46, + 443, + 287, + 563 + ], + "type": "inline_equation", + "content": "J A(\\cdot)" + }, + { + "bbox": [ + 46, + 443, + 287, + 563 + ], + "type": "text", + "content": ", which includes spatial attention [34] and channel attention [45], to highlight noteworthy parts in both spatial and channel domains. Therefore, given HF features " + }, + { + "bbox": [ + 46, + 443, + 287, + 563 + ], + "type": "inline_equation", + "content": "(f_k^r)_{HF}" + }, + { + "bbox": [ + 46, + 443, + 287, + 563 + ], + "type": "text", + "content": ", the HF attention map " + }, + { + "bbox": [ + 46, + 443, + 287, + 563 + ], + "type": "inline_equation", + "content": "p_k^h" + }, + { + "bbox": [ + 46, + 443, + 287, + 563 + ], + "type": "text", + "content": " is formulated as follows:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 565, + 287, + 578 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 105, + 565, + 287, + 578 + ], + "spans": [ + { + "bbox": [ + 105, + 565, + 287, + 578 + ], + "type": "interline_equation", + "content": "p _ {k} ^ {h} = J A \\left(R e s B \\left(\\left(f _ {k} ^ {r}\\right) _ {H F}\\right)\\right), \\tag {2}", + "image_path": "eaa17ffa4d62f993c0f6e95b18d5cc638e5186a48909e5ab80cc7d8955fb3e51.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 579, + 259, + 591 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 579, + 259, + 591 + ], + "spans": [ + { + "bbox": [ + 46, + 579, + 259, + 591 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 579, + 259, + 591 + ], + "type": "inline_equation", + "content": "ResB(\\cdot)" + }, + { + "bbox": [ + 46, + 579, + 259, + 591 + ], + "type": "text", + "content": " denotes the residual block with BN." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 594, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 287, + 713 + ], + "type": "text", + "content": "Low-frequency attention module. Low-frequency components focus more on global information, such as color distribution and illumination, which inevitably leads to inevitably existing redundant components and slight perturbations [43]. To handle those problems, we design a comprehensive normalization strategy to suppress the undesired artifacts and provide cleaner global information for attention calculation both at the instance level and channel dimension, which can highlight those abnormal regions from a global perspective. Specifically, this module takes the de" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 161, + 528, + 174 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 161, + 528, + 174 + ], + "spans": [ + { + "bbox": [ + 305, + 161, + 528, + 174 + ], + "type": "text", + "content": "composed LF features " + }, + { + "bbox": [ + 305, + 161, + 528, + 174 + ], + "type": "inline_equation", + "content": "(f_k^r)_{LF}" + }, + { + "bbox": [ + 305, + 161, + 528, + 174 + ], + "type": "text", + "content": " as the input and outputs" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 349, + 175, + 545, + 190 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 349, + 175, + 545, + 190 + ], + "spans": [ + { + "bbox": [ + 349, + 175, + 545, + 190 + ], + "type": "interline_equation", + "content": "p _ {k} ^ {l} = J A \\left(P N \\left(R e s I N \\left(\\left(f _ {k} ^ {r}\\right) _ {L F}\\right)\\right)\\right), \\tag {3}", + "image_path": "d0ae01fd1263b66d027c4ebc4bf2058e9b1bad5f0a8bae0a808e367f76a16e3c.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 190, + 545, + 227 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 190, + 545, + 227 + ], + "spans": [ + { + "bbox": [ + 304, + 190, + 545, + 227 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 190, + 545, + 227 + ], + "type": "inline_equation", + "content": "ResIN(\\cdot), PN(\\cdot)" + }, + { + "bbox": [ + 304, + 190, + 545, + 227 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 190, + 545, + 227 + ], + "type": "inline_equation", + "content": "JA(\\cdot)" + }, + { + "bbox": [ + 304, + 190, + 545, + 227 + ], + "type": "text", + "content": " denote the instance normalization [44] constrained residual block, positional normalization [20], and joint attention, respectively." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 305, + 243, + 533, + 255 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 243, + 533, + 255 + ], + "spans": [ + { + "bbox": [ + 305, + 243, + 533, + 255 + ], + "type": "text", + "content": "3.2.3 Guidance-based Feature Aggregation Module" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 258, + 545, + 342 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 258, + 545, + 342 + ], + "spans": [ + { + "bbox": [ + 304, + 258, + 545, + 342 + ], + "type": "text", + "content": "As shown in Fig. 3, we propose a guidance-based feature aggregation (GFA) module to integrate the multi-scale decomposed features. Unlike existing heuristic-based feature aggregation strategies simply using concatenation [22, 24], GFA is specifically designed to address the key issue of COD, i.e., emphasizing the subtle discriminative features, by promoting inter-feature information interaction." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 343, + 546, + 464 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 343, + 546, + 464 + ], + "spans": [ + { + "bbox": [ + 304, + 343, + 546, + 464 + ], + "type": "text", + "content": "Taking HF bands as an example, GFA generates the aggregated feature " + }, + { + "bbox": [ + 304, + 343, + 546, + 464 + ], + "type": "inline_equation", + "content": "\\{f_{k - 1}^h\\}_{k = 2}^4" + }, + { + "bbox": [ + 304, + 343, + 546, + 464 + ], + "type": "text", + "content": " that combines deep semantic information of the low-resolution feature " + }, + { + "bbox": [ + 304, + 343, + 546, + 464 + ], + "type": "inline_equation", + "content": "(f_k^r)_{HF}" + }, + { + "bbox": [ + 304, + 343, + 546, + 464 + ], + "type": "text", + "content": " (at a higher level) and the abundant spatial details of the highresolution feature " + }, + { + "bbox": [ + 304, + 343, + 546, + 464 + ], + "type": "inline_equation", + "content": "(f_{k - 1}^{r})_{HF}" + }, + { + "bbox": [ + 304, + 343, + 546, + 464 + ], + "type": "text", + "content": " (at a lower level) with the guidance of the attention map " + }, + { + "bbox": [ + 304, + 343, + 546, + 464 + ], + "type": "inline_equation", + "content": "p_k^h" + }, + { + "bbox": [ + 304, + 343, + 546, + 464 + ], + "type": "text", + "content": ". Therefore, the aggregated feature " + }, + { + "bbox": [ + 304, + 343, + 546, + 464 + ], + "type": "inline_equation", + "content": "f_{k - 1}^h" + }, + { + "bbox": [ + 304, + 343, + 546, + 464 + ], + "type": "text", + "content": " can better highlight the subtle discriminative features. To extract the attention-guided semantic information, we first generate the down-sampled aggregated feature " + }, + { + "bbox": [ + 304, + 343, + 546, + 464 + ], + "type": "inline_equation", + "content": "f_{k - 1}^{dh}" + }, + { + "bbox": [ + 304, + 343, + 546, + 464 + ], + "type": "text", + "content": " with the window-based linear model [23]:" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 318, + 466, + 545, + 481 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 466, + 545, + 481 + ], + "spans": [ + { + "bbox": [ + 318, + 466, + 545, + 481 + ], + "type": "interline_equation", + "content": "\\left(f _ {k - 1} ^ {d h}\\right) _ {i} = \\sigma_ {w} d o w n \\left(\\left(f _ {k - 1} ^ {r}\\right) _ {H F}\\right) _ {i} + \\mu_ {w}, \\forall i \\in s _ {w}, \\tag {4}", + "image_path": "3264d5fc0ae23bf1aa3cc335ecb51e3b06bda6c31c013dac2c22d10f493190b8.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 483, + 545, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 483, + 545, + 529 + ], + "spans": [ + { + "bbox": [ + 304, + 483, + 545, + 529 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 483, + 545, + 529 + ], + "type": "inline_equation", + "content": "down(\\cdot), s_w" + }, + { + "bbox": [ + 304, + 483, + 545, + 529 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 483, + 545, + 529 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 483, + 545, + 529 + ], + "type": "text", + "content": " are the down-sampling operation, local window, and pixel point " + }, + { + "bbox": [ + 304, + 483, + 545, + 529 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 483, + 545, + 529 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 304, + 483, + 545, + 529 + ], + "type": "inline_equation", + "content": "\\{\\sigma_w, \\mu_w\\}" + }, + { + "bbox": [ + 304, + 483, + 545, + 529 + ], + "type": "text", + "content": " are linear aggregation coefficients for the pixels in window " + }, + { + "bbox": [ + 304, + 483, + 545, + 529 + ], + "type": "inline_equation", + "content": "s_w" + }, + { + "bbox": [ + 304, + 483, + 545, + 529 + ], + "type": "text", + "content": ", which can be acquired by optimizing the following objective function:" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 312, + 531, + 545, + 557 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 531, + 545, + 557 + ], + "spans": [ + { + "bbox": [ + 312, + 531, + 545, + 557 + ], + "type": "interline_equation", + "content": "\\min _ {\\sigma_ {w}, \\mu_ {w}} \\sum_ {i \\in s _ {w}} \\left[ \\left(p _ {k} ^ {h}\\right) _ {i} ^ {2} \\left(\\left(f _ {k - 1} ^ {d h}\\right) _ {i} - \\left(\\left(f _ {k} ^ {r}\\right) _ {H F}\\right) _ {i}\\right) ^ {2} + \\epsilon \\sigma_ {w} ^ {2} \\right], \\tag {5}", + "image_path": "7a2fc1a42def2fcc4f4aef16b09568a37ebb4c77ae43c94919b1471953012b47.jpg" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 304, + 559, + 545, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 559, + 545, + 582 + ], + "spans": [ + { + "bbox": [ + 304, + 559, + 545, + 582 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 559, + 545, + 582 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 304, + 559, + 545, + 582 + ], + "type": "text", + "content": " is a constraint value for " + }, + { + "bbox": [ + 304, + 559, + 545, + 582 + ], + "type": "inline_equation", + "content": "\\sigma_w" + }, + { + "bbox": [ + 304, + 559, + 545, + 582 + ], + "type": "text", + "content": ". See Supplementary Material (Supp) for derivations and solutions of " + }, + { + "bbox": [ + 304, + 559, + 545, + 582 + ], + "type": "inline_equation", + "content": "\\{\\sigma_w, \\mu_w\\}" + }, + { + "bbox": [ + 304, + 559, + 545, + 582 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 304, + 582, + 545, + 631 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 582, + 545, + 631 + ], + "spans": [ + { + "bbox": [ + 304, + 582, + 545, + 631 + ], + "type": "text", + "content": "Considering pixel " + }, + { + "bbox": [ + 304, + 582, + 545, + 631 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 582, + 545, + 631 + ], + "type": "text", + "content": " covered by multiple windows, we average those window-wise coefficients and get the specific aggregation coefficients " + }, + { + "bbox": [ + 304, + 582, + 545, + 631 + ], + "type": "inline_equation", + "content": "\\{\\sigma_i,\\mu_i\\}" + }, + { + "bbox": [ + 304, + 582, + 545, + 631 + ], + "type": "text", + "content": " for pixel " + }, + { + "bbox": [ + 304, + 582, + 545, + 631 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 582, + 545, + 631 + ], + "type": "text", + "content": ". By matrixing " + }, + { + "bbox": [ + 304, + 582, + 545, + 631 + ], + "type": "inline_equation", + "content": "\\{\\sigma_i,\\mu_i\\}" + }, + { + "bbox": [ + 304, + 582, + 545, + 631 + ], + "type": "text", + "content": " into " + }, + { + "bbox": [ + 304, + 582, + 545, + 631 + ], + "type": "inline_equation", + "content": "\\{\\sigma_i,\\mu_i\\}" + }, + { + "bbox": [ + 304, + 582, + 545, + 631 + ], + "type": "text", + "content": ", Eq. (4) can be rewritten as follows:" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 347, + 632, + 545, + 645 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 347, + 632, + 545, + 645 + ], + "spans": [ + { + "bbox": [ + 347, + 632, + 545, + 645 + ], + "type": "interline_equation", + "content": "f _ {k - 1} ^ {d h} = \\sigma_ {i} \\odot d o w n \\left(\\left(f _ {k - 1} ^ {r}\\right) _ {H F}\\right) + \\boldsymbol {\\mu} _ {i}, \\tag {6}", + "image_path": "4b9ba1ace37e4cc4777ff0f3d47c38dc3af3c33dda08a41580ac5b16094a25bb.jpg" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 304, + 647, + 545, + 683 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 647, + 545, + 683 + ], + "spans": [ + { + "bbox": [ + 304, + 647, + 545, + 683 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 647, + 545, + 683 + ], + "type": "inline_equation", + "content": "\\odot" + }, + { + "bbox": [ + 304, + 647, + 545, + 683 + ], + "type": "text", + "content": " is the Hadamard product. We then up-sample " + }, + { + "bbox": [ + 304, + 647, + 545, + 683 + ], + "type": "inline_equation", + "content": "\\{\\pmb {\\sigma}_i,\\pmb {\\mu}_i\\}" + }, + { + "bbox": [ + 304, + 647, + 545, + 683 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 304, + 647, + 545, + 683 + ], + "type": "inline_equation", + "content": "\\{\\pmb {\\sigma}_h,\\pmb {\\mu}_h\\}" + }, + { + "bbox": [ + 304, + 647, + 545, + 683 + ], + "type": "text", + "content": " and acquire the high-resolution ag-. ggregated feature " + }, + { + "bbox": [ + 304, + 647, + 545, + 683 + ], + "type": "inline_equation", + "content": "f_{k - 1}^{h}" + }, + { + "bbox": [ + 304, + 647, + 545, + 683 + ], + "type": "text", + "content": " for enriching spatial details:" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 339, + 685, + 545, + 716 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 339, + 685, + 545, + 716 + ], + "spans": [ + { + "bbox": [ + 339, + 685, + 545, + 716 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} f _ {k - 1} ^ {h} = G F A \\left(\\left(f _ {k} ^ {r}\\right) _ {H F}, \\left(f _ {k - 1} ^ {r}\\right) _ {H F}, p _ {k} ^ {h}\\right), \\tag {7} \\\\ = \\sigma_ {h} \\odot \\left(f _ {k - 1} ^ {r}\\right) _ {H F} + \\boldsymbol {\\mu} _ {h}. \\\\ \\end{array}", + "image_path": "7e05497af45c29e7366815ce49e65fffcc8afcc5157ac13431e621be46363c38.jpg" + } + ] + } + ], + "index": 26 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "22049" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 49, + 71, + 161, + 133 + ], + "blocks": [ + { + "bbox": [ + 49, + 71, + 161, + 133 + ], + "lines": [ + { + "bbox": [ + 49, + 71, + 161, + 133 + ], + "spans": [ + { + "bbox": [ + 49, + 71, + 161, + 133 + ], + "type": "image", + "image_path": "866a33e0554a38702327cb1a77fe6198f9cf0428ebcc76212aff636e29623f83.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 166, + 72, + 285, + 133 + ], + "blocks": [ + { + "bbox": [ + 166, + 72, + 285, + 133 + ], + "lines": [ + { + "bbox": [ + 166, + 72, + 285, + 133 + ], + "spans": [ + { + "bbox": [ + 166, + 72, + 285, + 133 + ], + "type": "image", + "image_path": "51c635ff8eb5e70df58596dfe41a541b3c8582289e6dedbb546269f96fdde41e.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 49, + 137, + 160, + 200 + ], + "blocks": [ + { + "bbox": [ + 49, + 137, + 160, + 200 + ], + "lines": [ + { + "bbox": [ + 49, + 137, + 160, + 200 + ], + "spans": [ + { + "bbox": [ + 49, + 137, + 160, + 200 + ], + "type": "image", + "image_path": "e906f30558141957c5ea9078f6ee271dbd79ad0f0545ea32cad7ca4d68cf224e.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 47, + 202, + 287, + 224 + ], + "lines": [ + { + "bbox": [ + 47, + 202, + 287, + 224 + ], + "spans": [ + { + "bbox": [ + 47, + 202, + 287, + 224 + ], + "type": "text", + "content": "Figure 4. Details of SED with RRS and OER modules. OER is an RK2-based module guaranteed with the Hamiltonian system." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 166, + 134, + 286, + 200 + ], + "blocks": [ + { + "bbox": [ + 166, + 134, + 286, + 200 + ], + "lines": [ + { + "bbox": [ + 166, + 134, + 286, + 200 + ], + "spans": [ + { + "bbox": [ + 166, + 134, + 286, + 200 + ], + "type": "image", + "image_path": "8364ece33d2de787d7b9c493ef51c7e15fa6b70747e11a81ae71218ffa5a40e1.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 239, + 286, + 267 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 239, + 286, + 267 + ], + "spans": [ + { + "bbox": [ + 46, + 239, + 286, + 267 + ], + "type": "text", + "content": "To iteratively acquire the aggregated features " + }, + { + "bbox": [ + 46, + 239, + 286, + 267 + ], + "type": "inline_equation", + "content": "\\left\\{f_{k - 1}^{h}\\right\\}_{k = 2}^{4}" + }, + { + "bbox": [ + 46, + 239, + 286, + 267 + ], + "type": "text", + "content": " we redefine GFA module by replacing " + }, + { + "bbox": [ + 46, + 239, + 286, + 267 + ], + "type": "inline_equation", + "content": "(f_k^r)_HF" + }, + { + "bbox": [ + 46, + 239, + 286, + 267 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 46, + 239, + 286, + 267 + ], + "type": "inline_equation", + "content": "f_{k}^{h}" + }, + { + "bbox": [ + 46, + 239, + 286, + 267 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 92, + 267, + 286, + 283 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 92, + 267, + 286, + 283 + ], + "spans": [ + { + "bbox": [ + 92, + 267, + 286, + 283 + ], + "type": "interline_equation", + "content": "f _ {k - 1} ^ {h} = G F A \\left(f _ {k} ^ {h}, \\left(f _ {k - 1} ^ {r}\\right) _ {H F}, p _ {k} ^ {h}\\right), \\tag {8}", + "image_path": "50cf99caca74b1b56261fde087256e6f3e1f48f4018f56dc6c7be86c9da13b95.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 283, + 287, + 369 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 283, + 287, + 369 + ], + "spans": [ + { + "bbox": [ + 46, + 283, + 287, + 369 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 283, + 287, + 369 + ], + "type": "inline_equation", + "content": "p_k^h = JA(ResB(f_k^h))" + }, + { + "bbox": [ + 46, + 283, + 287, + 369 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 283, + 287, + 369 + ], + "type": "inline_equation", + "content": "f_4^h = (f_4^r)_{HF}" + }, + { + "bbox": [ + 46, + 283, + 287, + 369 + ], + "type": "text", + "content": ". Guaranteed with frequency-specific attention, our aggregated features can emphasize more discriminative features than others by combining abundant spatial details and deep semantic information, thus better catering to the COD task. The calculation of the aggregated LF features " + }, + { + "bbox": [ + 46, + 283, + 287, + 369 + ], + "type": "inline_equation", + "content": "f_{k - 1}^{l}" + }, + { + "bbox": [ + 46, + 283, + 287, + 369 + ], + "type": "text", + "content": " are similar to " + }, + { + "bbox": [ + 46, + 283, + 287, + 369 + ], + "type": "inline_equation", + "content": "f_{k - 1}^{h}" + }, + { + "bbox": [ + 46, + 283, + 287, + 369 + ], + "type": "text", + "content": ", which can be seen in Supp." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 369, + 287, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 369, + 287, + 453 + ], + "spans": [ + { + "bbox": [ + 46, + 369, + 287, + 453 + ], + "type": "text", + "content": "Considering that bottom layers (at higher levels) focus more on HF details while top layers (at lower levels) care more about global information [38], we pass the aggregated HF/LF features into the bottom/top decoder layers along with the skip-connected encoded features " + }, + { + "bbox": [ + 46, + 369, + 287, + 453 + ], + "type": "inline_equation", + "content": "\\{f_k^r\\}_{k=1}^4" + }, + { + "bbox": [ + 46, + 369, + 287, + 453 + ], + "type": "text", + "content": ". To balance performance and efficiency, the integrated features " + }, + { + "bbox": [ + 46, + 369, + 287, + 453 + ], + "type": "inline_equation", + "content": "\\{f_k^o\\}_{k=1}^4" + }, + { + "bbox": [ + 46, + 369, + 287, + 453 + ], + "type": "text", + "content": " passed to the decoder are defined as:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 77, + 455, + 287, + 474 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 455, + 287, + 474 + ], + "spans": [ + { + "bbox": [ + 77, + 455, + 287, + 474 + ], + "type": "interline_equation", + "content": "f _ {1} ^ {o} = \\operatorname {c o n v 1} \\left(\\operatorname {c o n} \\left(f _ {1} ^ {r}, u p \\left(f _ {1} ^ {h}\\right)\\right)\\right), f _ {2} ^ {o} = f _ {2} ^ {r}, \\tag {9}", + "image_path": "26bd608f2bd230505fd1e4ae90832cf90376972ab9a8f2479758948eefeb9695.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 78, + 472, + 233, + 486 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 78, + 472, + 233, + 486 + ], + "spans": [ + { + "bbox": [ + 78, + 472, + 233, + 486 + ], + "type": "interline_equation", + "content": "f _ {3} ^ {o} = f _ {3} ^ {r}, f _ {4} ^ {o} = \\operatorname {c o n v 1} \\left(\\operatorname {c o n} \\left(f _ {4} ^ {r}, f _ {3} ^ {l}\\right)\\right),", + "image_path": "e0575fee593b59aae9b93b51b9fdfef7a2124c8be70f84630e3847f606f229a0.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 487, + 287, + 522 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 487, + 287, + 522 + ], + "spans": [ + { + "bbox": [ + 46, + 487, + 287, + 522 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 487, + 287, + 522 + ], + "type": "inline_equation", + "content": "up(\\cdot)" + }, + { + "bbox": [ + 46, + 487, + 287, + 522 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 487, + 287, + 522 + ], + "type": "inline_equation", + "content": "con(\\cdot)" + }, + { + "bbox": [ + 46, + 487, + 287, + 522 + ], + "type": "text", + "content": " denote the up-sampling operation and the concatenation operation. conv1 represents " + }, + { + "bbox": [ + 46, + 487, + 287, + 522 + ], + "type": "inline_equation", + "content": "1 \\times 1" + }, + { + "bbox": [ + 46, + 487, + 287, + 522 + ], + "type": "text", + "content": " convolution, which is used for channel-level integration." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 47, + 531, + 285, + 544 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 531, + 285, + 544 + ], + "spans": [ + { + "bbox": [ + 47, + 531, + 285, + 544 + ], + "type": "text", + "content": "3.3. Segmentation-oriented Edge-assisted Decoder" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 46, + 551, + 287, + 634 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 551, + 287, + 634 + ], + "spans": [ + { + "bbox": [ + 46, + 551, + 287, + 634 + ], + "type": "text", + "content": "In the segmentation-oriented edge-assisted decoder (SED) " + }, + { + "bbox": [ + 46, + 551, + 287, + 634 + ], + "type": "inline_equation", + "content": "\\{D_k\\}_{k=1}^4" + }, + { + "bbox": [ + 46, + 551, + 287, + 634 + ], + "type": "text", + "content": ", we propose to learn an auxiliary edge reconstruction task alongside the COD task to help generate precise segmentation maps with accurate boundaries. To be specific, as shown in Fig. 4, each decoder layer " + }, + { + "bbox": [ + 46, + 551, + 287, + 634 + ], + "type": "inline_equation", + "content": "D_k" + }, + { + "bbox": [ + 46, + 551, + 287, + 634 + ], + "type": "text", + "content": " consists of a reversible re-calibration segmentation (RRS) module and an ODE-inspired edge reconstruction (OER) module." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 47, + 649, + 285, + 661 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 649, + 285, + 661 + ], + "spans": [ + { + "bbox": [ + 47, + 649, + 285, + 661 + ], + "type": "text", + "content": "3.3.1 Reversible Re-calibration Segmentation Module" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 46, + 665, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 287, + 713 + ], + "type": "text", + "content": "Due to complex camouflage, prediction maps inevitably have some ambiguous regions with low confidence, we adopt a reverse strategy to excavate cues from these low-confidence regions by reversing the attention, which" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 305, + 72, + 545, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 545, + 168 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 545, + 168 + ], + "type": "text", + "content": "erases detected regions and amplifies the response in low-confidence regions, thus re-calibrating the misclassified regions. Specifically, we repeat the coarse segmentation map " + }, + { + "bbox": [ + 305, + 72, + 545, + 168 + ], + "type": "inline_equation", + "content": "\\{d_{k + 1}^s\\}_{k = 0}^3" + }, + { + "bbox": [ + 305, + 72, + 545, + 168 + ], + "type": "text", + "content": " as a 64-dimension tensor, normalize it to [0, 1] with Sigmoid " + }, + { + "bbox": [ + 305, + 72, + 545, + 168 + ], + "type": "inline_equation", + "content": "S(\\cdot)" + }, + { + "bbox": [ + 305, + 72, + 545, + 168 + ], + "type": "text", + "content": ", and reverse it by subtracting each element from 1. We then multiply the integrated feature " + }, + { + "bbox": [ + 305, + 72, + 545, + 168 + ], + "type": "inline_equation", + "content": "f_{k}^{o}" + }, + { + "bbox": [ + 305, + 72, + 545, + 168 + ], + "type": "text", + "content": " with the reversed map and concatenate it with the edge feature " + }, + { + "bbox": [ + 305, + 72, + 545, + 168 + ], + "type": "inline_equation", + "content": "d_{k}^{fe}" + }, + { + "bbox": [ + 305, + 72, + 545, + 168 + ], + "type": "text", + "content": " to obtain the segmentation result " + }, + { + "bbox": [ + 305, + 72, + 545, + 168 + ], + "type": "inline_equation", + "content": "d_{k}^{s}" + }, + { + "bbox": [ + 305, + 72, + 545, + 168 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 169, + 545, + 190 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 169, + 545, + 190 + ], + "spans": [ + { + "bbox": [ + 307, + 169, + 545, + 190 + ], + "type": "interline_equation", + "content": "d _ {k} ^ {s} = \\operatorname {c o n v 3} \\left(\\operatorname {c o n} \\left(d _ {k} ^ {f e}, f _ {k} ^ {o} \\odot r v (S (r p \\left(d _ {k + 1} ^ {s}\\right))\\right)\\right), \\tag {10}", + "image_path": "ce84e530a4f8e049b9ad83e8f6495076b2353124fb9b978cd42f8f4bef854d91.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 306, + 191, + 504, + 202 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 191, + 504, + 202 + ], + "spans": [ + { + "bbox": [ + 306, + 191, + 504, + 202 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 306, + 191, + 504, + 202 + ], + "type": "inline_equation", + "content": "rp(\\cdot)" + }, + { + "bbox": [ + 306, + 191, + 504, + 202 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 306, + 191, + 504, + 202 + ], + "type": "inline_equation", + "content": "rv(\\cdot)" + }, + { + "bbox": [ + 306, + 191, + 504, + 202 + ], + "type": "text", + "content": " denote repeat and reverse." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 305, + 220, + 525, + 232 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 220, + 525, + 232 + ], + "spans": [ + { + "bbox": [ + 305, + 220, + 525, + 232 + ], + "type": "text", + "content": "3.3.2 ODE-inspired Edge Reconstruction Module" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 237, + 545, + 321 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 237, + 545, + 321 + ], + "spans": [ + { + "bbox": [ + 304, + 237, + 545, + 321 + ], + "type": "text", + "content": "Existing methods tend to excavate edge information by incorporating certain priors within the residual network structure [13, 50]. However, either the intractable localization problem of the IS challenge or the ambiguous boundary problem of the ED challenge makes it difficult to design an appropriate edge prior for the COD task. In some cases, a biased prior can even reduce the segmentation performance." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 322, + 546, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 322, + 546, + 525 + ], + "spans": [ + { + "bbox": [ + 304, + 322, + 546, + 525 + ], + "type": "text", + "content": "Therefore, instead of exploiting prior knowledge, we focus on proposing an edge-friendly network architecture, i.e., the ODE-inspired edge reconstruction (OER) module. Compared with the traditional residual network structure that can be seen as the first-order Euler discretization approximation of ODE with nonnegligible truncation errors [46], the proposed OER module employs a higher-order ODE solver, specifically, a second-order Runge-Kutta (RK2), to provide more accurate numerical solutions in edge information processing. This better accommodates the fine-grained property of edges and facilitates the complete edge reconstruction, thus addressing the ED challenge. To ensure the flexibility of our OER module, we replace the fixed trade-off parameter in the RK2 solver with a weighted gate mechanism " + }, + { + "bbox": [ + 304, + 322, + 546, + 525 + ], + "type": "inline_equation", + "content": "g_{w}" + }, + { + "bbox": [ + 304, + 322, + 546, + 525 + ], + "type": "text", + "content": " with learnable coefficients. Given an input " + }, + { + "bbox": [ + 304, + 322, + 546, + 525 + ], + "type": "inline_equation", + "content": "e_t" + }, + { + "bbox": [ + 304, + 322, + 546, + 525 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 322, + 546, + 525 + ], + "type": "inline_equation", + "content": "e_t = conv1\\left(\\operatorname{con}\\left(f_k^o, rp\\left(d_{k+1}^s\\right)\\right)\\right)" + }, + { + "bbox": [ + 304, + 322, + 546, + 525 + ], + "type": "text", + "content": ", the proposed OER module can be formulated as follows:" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 344, + 527, + 485, + 540 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 344, + 527, + 485, + 540 + ], + "spans": [ + { + "bbox": [ + 344, + 527, + 485, + 540 + ], + "type": "interline_equation", + "content": "e _ {t + 1} = e _ {t} + g _ {w} F _ {1} + \\left(1 - g _ {w}\\right) F _ {2},", + "image_path": "5ca2a45114b249899ca8fb6b4608eccce061cdf13fc3c01b944c44b9e4d45b51.jpg" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 345, + 542, + 545, + 555 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 345, + 542, + 545, + 555 + ], + "spans": [ + { + "bbox": [ + 345, + 542, + 545, + 555 + ], + "type": "interline_equation", + "content": "g _ {w} = S \\left(\\sigma_ {g} \\operatorname {c o n} (F 1, F 2) + \\mu_ {g}\\right), \\tag {11}", + "image_path": "a8722b06fb797ce8ab8f389eaef88f019d6ded462c8378978c6d581859a56725.jpg" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 345, + 557, + 505, + 569 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 345, + 557, + 505, + 569 + ], + "spans": [ + { + "bbox": [ + 345, + 557, + 505, + 569 + ], + "type": "interline_equation", + "content": "F _ {1} = F \\left(e _ {t}, \\theta_ {t}\\right), F _ {2} = F \\left(e _ {t} + F _ {1}, \\theta_ {t}\\right),", + "image_path": "3650c9e43a9084bb0732ecfeadf0473c6ceb11c4176c8a906200c5c31c8c2137.jpg" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 304, + 572, + 545, + 656 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 572, + 545, + 656 + ], + "spans": [ + { + "bbox": [ + 304, + 572, + 545, + 656 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 572, + 545, + 656 + ], + "type": "inline_equation", + "content": "e_{t + 1} = d_k^{fe}" + }, + { + "bbox": [ + 304, + 572, + 545, + 656 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 572, + 545, + 656 + ], + "type": "inline_equation", + "content": "\\sigma_g" + }, + { + "bbox": [ + 304, + 572, + 545, + 656 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 572, + 545, + 656 + ], + "type": "inline_equation", + "content": "\\mu_g" + }, + { + "bbox": [ + 304, + 572, + 545, + 656 + ], + "type": "text", + "content": " are the learnable parameters in " + }, + { + "bbox": [ + 304, + 572, + 545, + 656 + ], + "type": "inline_equation", + "content": "g_w" + }, + { + "bbox": [ + 304, + 572, + 545, + 656 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 304, + 572, + 545, + 656 + ], + "type": "inline_equation", + "content": "\\{F_i\\}_{i = 1}^2" + }, + { + "bbox": [ + 304, + 572, + 545, + 656 + ], + "type": "text", + "content": " denotes the intermediate layers with shared parameters " + }, + { + "bbox": [ + 304, + 572, + 545, + 656 + ], + "type": "inline_equation", + "content": "\\theta_t" + }, + { + "bbox": [ + 304, + 572, + 545, + 656 + ], + "type": "text", + "content": " for efficiency. Following [1], we set " + }, + { + "bbox": [ + 304, + 572, + 545, + 656 + ], + "type": "inline_equation", + "content": "F_i" + }, + { + "bbox": [ + 304, + 572, + 545, + 656 + ], + "type": "text", + "content": " as a Conv-ReLU-Conv framework. To ensure the stability of OER, we apply the Hamiltonian system [9] to our OER module. By denoting Eq. (11) as " + }, + { + "bbox": [ + 304, + 572, + 545, + 656 + ], + "type": "inline_equation", + "content": "RK2(\\cdot)" + }, + { + "bbox": [ + 304, + 572, + 545, + 656 + ], + "type": "text", + "content": ", the Hamiltonian-theory-guaranteed OER module is defined as follows:" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 311, + 658, + 545, + 676 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 658, + 545, + 676 + ], + "spans": [ + { + "bbox": [ + 311, + 658, + 545, + 676 + ], + "type": "interline_equation", + "content": "e _ {t + 1} = \\operatorname {c o n} \\left(e _ {t + 1} ^ {1}, e _ {t + 1} ^ {2}\\right), e _ {t + 1} ^ {1} = e _ {t} ^ {1} + R K 2 _ {1} \\left(e _ {t} ^ {2}\\right), \\tag {12}", + "image_path": "e1fc8166bdc4af4df9c90f1c356aa43e08ba4fe4289ddf4db688aa4b61b0cb5d.jpg" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 313, + 673, + 476, + 688 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 673, + 476, + 688 + ], + "spans": [ + { + "bbox": [ + 313, + 673, + 476, + 688 + ], + "type": "interline_equation", + "content": "e _ {t + 1} ^ {2} = e _ {t} ^ {2} - R K 2 _ {2} \\left(e _ {t} ^ {1} + R K 2 _ {1} \\left(e _ {t} ^ {2}\\right)\\right),", + "image_path": "f91120ab092e737e7097a45067e38980ad798161ad2cae8dcc395bf6d10b5c3b.jpg" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 304, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 689, + 545, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 689, + 545, + 713 + ], + "type": "inline_equation", + "content": "e_t" + }, + { + "bbox": [ + 304, + 689, + 545, + 713 + ], + "type": "text", + "content": " is split to " + }, + { + "bbox": [ + 304, + 689, + 545, + 713 + ], + "type": "inline_equation", + "content": "e_t^1" + }, + { + "bbox": [ + 304, + 689, + 545, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 689, + 545, + 713 + ], + "type": "inline_equation", + "content": "e_t^2" + }, + { + "bbox": [ + 304, + 689, + 545, + 713 + ], + "type": "text", + "content": " in channel-wise. Note that the OER module in Eq. (12) is a reversible and sta" + } + ] + } + ], + "index": 28 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "22050" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 51, + 71, + 547, + 287 + ], + "blocks": [ + { + "bbox": [ + 51, + 71, + 547, + 287 + ], + "lines": [ + { + "bbox": [ + 51, + 71, + 547, + 287 + ], + "spans": [ + { + "bbox": [ + 51, + 71, + 547, + 287 + ], + "type": "table", + "html": "
MethodsPublicationsBackbonesCHAMELEON (76 images)CAMO (250 images)COD10K (2,026 images)NC4K (4,121 images)
M↓Fβ↑Eφ↑Sα↑M↓Fβ↑Eφ↑Sα↑M↓Fβ↑Eφ↑Sα↑
Common Setting: Single Input Scale and Single Stage
CPD [47]CVPR19ResNet500.0480.7750.8740.8570.1130.6750.7230.7160.0530.5780.7760.750
SINet [6]CVPR20ResNet500.0340.8230.9360.8720.0920.7120.8040.7450.0430.6670.8640.776
PFNet [28]CVPR21ResNet500.0330.8200.9310.8820.0850.7510.8410.7820.0400.6760.8770.800
MGL-R [50]CVPR21ResNet500.0310.8250.9170.8910.0880.7380.8120.7750.0350.6800.8510.814
LSR [25]CVPR21ResNet500.0300.8350.9350.8900.0800.7560.8380.7870.0370.6990.8800.804
UGTR [48]ICCV21ResNet500.0310.8050.9100.8880.0860.7470.8210.7840.0360.6700.8520.817
SLT-Net [2]CVPR22ResNet500.0300.8350.9400.8870.0820.7630.8480.7920.0360.6810.8750.804
SegMaR-1 [14]CVPR22ResNet500.0280.8280.9440.8920.0720.7720.8610.8050.0350.6990.8900.813
OSFormer [32]ECCV22ResNet500.0280.8360.9390.8910.0730.7670.8580.7990.0340.7010.8810.811
FEDER-R50—ResNet500.0280.8550.9470.8940.0690.7850.8730.8070.0320.7400.9000.823
SINet V2 [4]TPAMI22ResNet500.0300.8160.9420.8880.0700.7790.8820.8220.0370.6820.8870.815
BSA-Net [52]AAAI22ResNet500.0270.8510.9460.8950.0790.7680.8510.7960.0340.7230.8910.818
FEDER-R2N—ResNet500.0260.8560.9470.9030.0660.8070.8970.8360.0290.7480.9110.844
Other Setting: Multiple Input Scales (MIS)
ZoomNet [31]CVPR22ResNet500.0240.8580.9430.9020.0660.7920.8770.8200.0290.7400.8880.838
FEDER-MIS—ResNet500.0230.8690.9590.9060.0640.8010.8930.8270.0280.7560.9130.837
Other Setting: Multiple Stages (MS)
SegMaR-4 [14]CVPR22ResNet500.0250.8550.9550.9060.0710.7790.8650.8150.0330.7370.8960.833
FEDER-MS-4—ResNet500.0250.8740.9640.9070.0670.8090.8860.8220.0280.7520.9170.851
", + "image_path": "313bb79684d6aa130b56af1e289bed344d0d630ad4a7b5fc7a68e463be8f6739.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 288, + 546, + 321 + ], + "lines": [ + { + "bbox": [ + 46, + 288, + 546, + 321 + ], + "spans": [ + { + "bbox": [ + 46, + 288, + 546, + 321 + ], + "type": "text", + "content": "Table 1. Quantitative comparisons of the proposed FEDER and other state-of-the-art methods on four benchmarks. SegMaR-1 and SegMaR-4 denote SegMaR at one stage and four stages. R50 and R2N indicate ResNet50 and Res2Net50. The best results are marked in bold. For ResNet50 backbone in the common setting, the best two results are in red and blue fonts." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 47, + 327, + 287, + 364 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 327, + 287, + 364 + ], + "spans": [ + { + "bbox": [ + 47, + 327, + 287, + 364 + ], + "type": "text", + "content": "ble block [9], which further promotes the edge reconstruction performance. In this case, the final edge predictions " + }, + { + "bbox": [ + 47, + 327, + 287, + 364 + ], + "type": "inline_equation", + "content": "\\{d_k^e\\}_{k = 1}^4" + }, + { + "bbox": [ + 47, + 327, + 287, + 364 + ], + "type": "text", + "content": " can be acquired in the following manner:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 91, + 365, + 287, + 385 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 365, + 287, + 385 + ], + "spans": [ + { + "bbox": [ + 91, + 365, + 287, + 385 + ], + "type": "interline_equation", + "content": "d _ {k} ^ {e} = \\operatorname {c o n v 3} \\left(e _ {t + 1}\\right) = \\operatorname {c o n v 3} \\left(d _ {k} ^ {f e}\\right). \\tag {13}", + "image_path": "c10c1b9587dac31ec49f453e87e6c81f8f46b8ce8baac8af7cbf983fc50493e8.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 392, + 141, + 403 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 392, + 141, + 403 + ], + "spans": [ + { + "bbox": [ + 47, + 392, + 141, + 403 + ], + "type": "text", + "content": "3.4. Loss Functions" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 411, + 287, + 577 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 411, + 287, + 577 + ], + "spans": [ + { + "bbox": [ + 46, + 411, + 287, + 577 + ], + "type": "text", + "content": "The loss function of the proposed FEDER consists of two kinds of supervisions, namely the segmentation mask " + }, + { + "bbox": [ + 46, + 411, + 287, + 577 + ], + "type": "inline_equation", + "content": "GT_{s}" + }, + { + "bbox": [ + 46, + 411, + 287, + 577 + ], + "type": "text", + "content": " and edge " + }, + { + "bbox": [ + 46, + 411, + 287, + 577 + ], + "type": "inline_equation", + "content": "GT_{e}" + }, + { + "bbox": [ + 46, + 411, + 287, + 577 + ], + "type": "text", + "content": " of the camouflaged object, which correspond to the multi-scale segmentation maps " + }, + { + "bbox": [ + 46, + 411, + 287, + 577 + ], + "type": "inline_equation", + "content": "\\{d_k^s\\}_{k = 1}^5" + }, + { + "bbox": [ + 46, + 411, + 287, + 577 + ], + "type": "text", + "content": " and the multi-scale object edges " + }, + { + "bbox": [ + 46, + 411, + 287, + 577 + ], + "type": "inline_equation", + "content": "\\{d_k^e\\}_{k = 1}^4" + }, + { + "bbox": [ + 46, + 411, + 287, + 577 + ], + "type": "text", + "content": ". Following [4], we employ the weighted binary cross-entropy loss " + }, + { + "bbox": [ + 46, + 411, + 287, + 577 + ], + "type": "inline_equation", + "content": "L_{BCE}^{w}" + }, + { + "bbox": [ + 46, + 411, + 287, + 577 + ], + "type": "text", + "content": " and weighted intersection-over-union loss " + }, + { + "bbox": [ + 46, + 411, + 287, + 577 + ], + "type": "inline_equation", + "content": "L_{IoU}^{w}" + }, + { + "bbox": [ + 46, + 411, + 287, + 577 + ], + "type": "text", + "content": " for segmentation supervision, which focuses more on those hard pixels. For edge supervision, we use the dice loss " + }, + { + "bbox": [ + 46, + 411, + 287, + 577 + ], + "type": "inline_equation", + "content": "L_{dice}" + }, + { + "bbox": [ + 46, + 411, + 287, + 577 + ], + "type": "text", + "content": " [29] to overcome the extreme imbalance between the positive and negative samples. Furthermore, to handle the multi-scale outputs, we up-sampling all the outputs to match the size of their corresponding ground truths during training. Therefore, the total loss of our FEDER is formulated as follows:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 45, + 578, + 287, + 644 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 578, + 287, + 644 + ], + "spans": [ + { + "bbox": [ + 45, + 578, + 287, + 644 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} L _ {t} = \\sum_ {k = 1} ^ {5} \\frac {1}{2 ^ {k - 1}} \\left(L _ {B C E} ^ {w} \\left(d _ {k} ^ {s}, G T _ {s}\\right) + L _ {I o U} ^ {w} \\left(d _ {k} ^ {s}, G T _ {S}\\right)\\right) \\tag {14} \\\\ + \\sum_ {k = 1} ^ {4} \\frac {1}{2 ^ {k - 1}} L _ {d i c e} \\left(d _ {k} ^ {e}, G T _ {e}\\right). \\\\ \\end{array}", + "image_path": "47010d110712d699ce1969390d3804b2bab937030b17206692a18f0f79b7f19d.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 650, + 123, + 663 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 650, + 123, + 663 + ], + "spans": [ + { + "bbox": [ + 47, + 650, + 123, + 663 + ], + "type": "text", + "content": "4. Experiment" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 670, + 164, + 684 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 670, + 164, + 684 + ], + "spans": [ + { + "bbox": [ + 47, + 670, + 164, + 684 + ], + "type": "text", + "content": "4.1. Experimental Setup" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 689, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 689, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 689, + 287, + 714 + ], + "type": "text", + "content": "Datasets. We used four widely-used COD datasets for evaluation, including CHAMELEON [39], CAMO [18]," + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 327, + 545, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 327, + 545, + 434 + ], + "spans": [ + { + "bbox": [ + 304, + 327, + 545, + 434 + ], + "type": "text", + "content": "COD10K [4], and NC4K [25]. CHAMELEON contains 76 hand-annotated images. CAMO has 1,250 camouflaged images with 1,000 training images and 250 testing images. Currently, COD10K is the largest COD benchmark, with 3,040 training images and 2,026 testing images. NC4K is a large-scale COD testing dataset, comprising 4,121 images. Following [4], we form the training set with 3,040 images from COD10K and 1,000 images from CAMO, while the remaining camouflaged images are used for testing." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 435, + 545, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 435, + 545, + 495 + ], + "spans": [ + { + "bbox": [ + 304, + 435, + 545, + 495 + ], + "type": "text", + "content": "Evaluation metrics. Four commonly-used metrics are employed for COD task, including mean absolute error " + }, + { + "bbox": [ + 304, + 435, + 545, + 495 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 304, + 435, + 545, + 495 + ], + "type": "text", + "content": ", adaptive F-measure " + }, + { + "bbox": [ + 304, + 435, + 545, + 495 + ], + "type": "inline_equation", + "content": "F_{\\beta}" + }, + { + "bbox": [ + 304, + 435, + 545, + 495 + ], + "type": "text", + "content": " [27], mean E-measure " + }, + { + "bbox": [ + 304, + 435, + 545, + 495 + ], + "type": "inline_equation", + "content": "E_{\\phi}" + }, + { + "bbox": [ + 304, + 435, + 545, + 495 + ], + "type": "text", + "content": " [5], and structure measure " + }, + { + "bbox": [ + 304, + 435, + 545, + 495 + ], + "type": "inline_equation", + "content": "S_{\\alpha}" + }, + { + "bbox": [ + 304, + 435, + 545, + 495 + ], + "type": "text", + "content": " [3]. Larger " + }, + { + "bbox": [ + 304, + 435, + 545, + 495 + ], + "type": "inline_equation", + "content": "F_{\\beta}" + }, + { + "bbox": [ + 304, + 435, + 545, + 495 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 435, + 545, + 495 + ], + "type": "inline_equation", + "content": "E_{\\phi}" + }, + { + "bbox": [ + 304, + 435, + 545, + 495 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 435, + 545, + 495 + ], + "type": "inline_equation", + "content": "S_{\\alpha}" + }, + { + "bbox": [ + 304, + 435, + 545, + 495 + ], + "type": "text", + "content": ", and smaller " + }, + { + "bbox": [ + 304, + 435, + 545, + 495 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 304, + 435, + 545, + 495 + ], + "type": "text", + "content": " indicate better segmentation performance." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 495, + 545, + 602 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 495, + 545, + 602 + ], + "spans": [ + { + "bbox": [ + 304, + 495, + 545, + 602 + ], + "type": "text", + "content": "Implementation details. The proposed FEDER is implemented in PyTorch on two RTX3090TI GPUs and is optimized by the Adam with momentum terms (0.9, 0.999). Following the common setting [4, 6], our encoder (ResNet50 by default) is initialized with the model pre-trained on ImageNet [6]. In the training phase, the batch size is set to 36 and the learning rate is initialized to 0.0001, dividing by 10 every 80 epochs. For both training and inference phases, all images are resized as " + }, + { + "bbox": [ + 304, + 495, + 545, + 602 + ], + "type": "inline_equation", + "content": "384 \\times 384" + }, + { + "bbox": [ + 304, + 495, + 545, + 602 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 611, + 507, + 624 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 611, + 507, + 624 + ], + "spans": [ + { + "bbox": [ + 305, + 611, + 507, + 624 + ], + "type": "text", + "content": "4.2. Comparison with the State-of-the-arts" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "text", + "content": "Quantitative analysis. We compare the proposed FEDER with 12 cutting-edge techniques in three different settings, including the common setting (single input scale and single stage) and two other settings (multiple input scales (MIS) and multiple stages (MS)). In the MIS and MS settings, the proposed FEDER follows the practices of ZoomNet [31] and SegMaR [14], where FEDER-MS-4 means" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "type": "text", + "content": "22051" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 71, + 545, + 243 + ], + "blocks": [ + { + "bbox": [ + 50, + 71, + 545, + 243 + ], + "lines": [ + { + "bbox": [ + 50, + 71, + 545, + 243 + ], + "spans": [ + { + "bbox": [ + 50, + 71, + 545, + 243 + ], + "type": "image", + "image_path": "368ed6bce6fca44bf494230f7a13b13d55ef9c0209b9ae769f594f1212645419.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 245, + 546, + 266 + ], + "lines": [ + { + "bbox": [ + 46, + 245, + 546, + 266 + ], + "spans": [ + { + "bbox": [ + 46, + 245, + 546, + 266 + ], + "type": "text", + "content": "Figure 5. Visual comparisons of the proposed FEDER and other nine state-of-the-art methods. Our method can generate more accurate prediction maps with clearer boundaries than other methods. Please zoom in for more details." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 48, + 286, + 286, + 451 + ], + "blocks": [ + { + "bbox": [ + 48, + 286, + 286, + 451 + ], + "lines": [ + { + "bbox": [ + 48, + 286, + 286, + 451 + ], + "spans": [ + { + "bbox": [ + 48, + 286, + 286, + 451 + ], + "type": "image", + "image_path": "b039425d32da85e97fb5cfb7a542cad0ae8a695d7c926159c176cc6ea880885d.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 453, + 286, + 474 + ], + "lines": [ + { + "bbox": [ + 46, + 453, + 286, + 474 + ], + "spans": [ + { + "bbox": [ + 46, + 453, + 286, + 474 + ], + "type": "text", + "content": "Figure 6. Performance-Params-FLOPs comparisons of some state-of-the-art deep learning-based COD methods on COD10K [4]." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 495, + 286, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 495, + 286, + 685 + ], + "spans": [ + { + "bbox": [ + 46, + 495, + 286, + 685 + ], + "type": "text", + "content": "complete COD task with our FEDER at four stages following SegMaR-4. For a fair comparison, prediction maps of the above techniques are directly segmented by their provided models with no modifications. Besides, all prediction maps are evaluated with the same code. As shown in Tab. 1, our method achieves the best results in all settings and backbones, which comprehensively demonstrates the superiority of our FEDER. Notably, " + }, + { + "bbox": [ + 46, + 495, + 286, + 685 + ], + "type": "inline_equation", + "content": "COD10K" + }, + { + "bbox": [ + 46, + 495, + 286, + 685 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 495, + 286, + 685 + ], + "type": "inline_equation", + "content": "NC4K" + }, + { + "bbox": [ + 46, + 495, + 286, + 685 + ], + "type": "text", + "content": " are the two most challenging datasets in terms of the number of images and segmentation difficulty. In the common setting, the proposed FEDER surpasses the second best results " + }, + { + "bbox": [ + 46, + 495, + 286, + 685 + ], + "type": "inline_equation", + "content": "3.6\\%" + }, + { + "bbox": [ + 46, + 495, + 286, + 685 + ], + "type": "text", + "content": " on average over all metrics on " + }, + { + "bbox": [ + 46, + 495, + 286, + 685 + ], + "type": "inline_equation", + "content": "COD10K" + }, + { + "bbox": [ + 46, + 495, + 286, + 685 + ], + "type": "text", + "content": " dataset and " + }, + { + "bbox": [ + 46, + 495, + 286, + 685 + ], + "type": "inline_equation", + "content": "2.6\\%" + }, + { + "bbox": [ + 46, + 495, + 286, + 685 + ], + "type": "text", + "content": " on average on " + }, + { + "bbox": [ + 46, + 495, + 286, + 685 + ], + "type": "inline_equation", + "content": "NC4K" + }, + { + "bbox": [ + 46, + 495, + 286, + 685 + ], + "type": "text", + "content": " dataset. Such a big margin further confirms the effectiveness of the proposed method, both with the deep wavelet-like feature decomposition strategy and the ODE-inspired edge reconstruction module." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 689, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 689, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 689, + 286, + 713 + ], + "type": "text", + "content": "Qualitative analysis. Fig. 5 presents a visual comparison of our FEDER and other SOTAs. We select various typi" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 288, + 545, + 623 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 288, + 545, + 623 + ], + "spans": [ + { + "bbox": [ + 304, + 288, + 545, + 623 + ], + "type": "text", + "content": "cal and challenging camouflaged images and arrange them in order of the camouflaged object size, from smallest to largest. Note that most of these images suffer from the IS or ED challenge, which can confuse existing COD techniques, resulting in mislocalization, ambiguous boundaries, etc. In contrast to those methods, our FEDER can overcome such challenges and generate more competitive prediction maps in the following aspects: (a) more accurate localization of small objects. For those small objects under the IS challenge, precise localization is a significant problem due to subtle differences and can confuse most existing methods. Thanks to our HFA, LFA, and GFA modules, our FEDER can emphasize the inconspicuous discriminative features and thus ensure more accurate camouflaged object localization (in Rows 1 and 2). (b) clearer edges on large objects. For those large objects, our prediction maps can achieve much clearer boundaries than others (see Rows 7 and 8), which mainly attributes to our joint training strategy of edge and segmentation and our edge-friendly OER module. (c) stronger suppression of redundant information. In the IS challenge and degraded imaging scenarios, the detection performance can be inevitably influenced by redundant information, such as background noise. However, the proposed deep decomposition strategy can suppress the redundant information by filtering out those components with the most discriminative information, namely, the HF and LF components. Thus, as depicted in Rows 3 and 4 in Fig. 5, FEDER can generate more accurate prediction maps." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "text", + "content": "Efficiency analysis. We compare the performance, parameters, and FLOPs with other SOTAs on COD10K [4] in Fig. 6. As presented in Fig. 6, our proposed FEDER achieves the smallest FLOPs and parameters compared with other state-of-the-art deep learning-based COD techniques. Furthermore, our score in " + }, + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "inline_equation", + "content": "F_{\\beta}" + }, + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "text", + "content": " is much higher than other methods and surpasses the second best one in " + }, + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "inline_equation", + "content": "5.6\\%" + }, + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "22052" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 55, + 71, + 296, + 149 + ], + "blocks": [ + { + "bbox": [ + 55, + 71, + 296, + 149 + ], + "lines": [ + { + "bbox": [ + 55, + 71, + 296, + 149 + ], + "spans": [ + { + "bbox": [ + 55, + 71, + 296, + 149 + ], + "type": "table", + "html": "
MethodsCOD10K (2026 images)NC4K (4121 images)
M↓Fβ↑Eφ↑Sα↑M↓Fβ↑Eφ↑Sα↑
(a)w/o DWD0.0350.6970.8610.7940.0490.7760.8730.817
(b)FW->LW0.0320.7310.8950.8220.0460.8110.8990.844
(c)w/o HFA0.0330.7280.8950.8170.0470.8050.8910.840
(d)w/o LFA0.0330.7240.8870.8160.0470.8030.8920.841
(e)UFA->GFA0.0330.7200.8940.8100.0470.8040.8900.838
(f)FEDER0.0320.7400.9000.8230.0450.8170.9050.846
", + "image_path": "6b379310846587918299ab0bae1c54d6b23582215e138cf61e3db7612dc4f4fa.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 300, + 71, + 542, + 149 + ], + "blocks": [ + { + "bbox": [ + 108, + 152, + 240, + 161 + ], + "lines": [ + { + "bbox": [ + 108, + 152, + 240, + 161 + ], + "spans": [ + { + "bbox": [ + 108, + 152, + 240, + 161 + ], + "type": "text", + "content": "(a) Ablation study of DWD Component." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 300, + 71, + 542, + 149 + ], + "lines": [ + { + "bbox": [ + 300, + 71, + 542, + 149 + ], + "spans": [ + { + "bbox": [ + 300, + 71, + 542, + 149 + ], + "type": "table", + "html": "
MethodsCOD10K (2026 images)NC4K (4121 images)
M↓Fβ↑Eφ↑Sα↑M↓Fβ↑Eφ↑Sα↑
(a)w/o OER0.0340.7140.8710.8040.0480.7800.8890.834
(b)FC->WGM0.0320.7310.8990.8170.0460.8140.8950.842
(c)w/o HS0.0320.7230.8850.8210.0460.7980.8920.840
(d)RB->RK20.0330.7220.8810.8110.0470.8020.8950.843
(e)RK4->RK20.0310.7420.9050.8290.0450.8160.9060.848
(f)FEDER0.0320.7400.9000.8230.0450.8170.9050.846
", + "image_path": "78b13002d9327170e05a7c5090799ef11d76d04b338afdc7d2df98140cbe72b4.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 361, + 151, + 476, + 160 + ], + "lines": [ + { + "bbox": [ + 361, + 151, + 476, + 160 + ], + "spans": [ + { + "bbox": [ + 361, + 151, + 476, + 160 + ], + "type": "text", + "content": "(b) Ablation study of OER module." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 161, + 547, + 194 + ], + "lines": [ + { + "bbox": [ + 46, + 161, + 547, + 194 + ], + "spans": [ + { + "bbox": [ + 46, + 161, + 547, + 194 + ], + "type": "text", + "content": "Table 2. Ablation study on " + }, + { + "bbox": [ + 46, + 161, + 547, + 194 + ], + "type": "inline_equation", + "content": "COD10k" + }, + { + "bbox": [ + 46, + 161, + 547, + 194 + ], + "type": "text", + "content": " [4] and " + }, + { + "bbox": [ + 46, + 161, + 547, + 194 + ], + "type": "inline_equation", + "content": "NC4K" + }, + { + "bbox": [ + 46, + 161, + 547, + 194 + ], + "type": "text", + "content": " [25], where w/o means without. (a) FW, LW, and UFA are short for fixed wavelet [40], learnable wavelet, and upsampling-based feature aggregation [51]. (b) FC, WGM, HS, RB, and RK4 are short for fixed coefficient, weighted gate mechanism, Hamiltonian system, residual block, and forth-order Runge-Kutta. The best results are marked in bold." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 50, + 208, + 285, + 371 + ], + "blocks": [ + { + "bbox": [ + 50, + 208, + 285, + 371 + ], + "lines": [ + { + "bbox": [ + 50, + 208, + 285, + 371 + ], + "spans": [ + { + "bbox": [ + 50, + 208, + 285, + 371 + ], + "type": "image", + "image_path": "40688237d6af8db53c8358514051cbdb71c4b8ec2673fc9a246fcde553a3b2eb.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 372, + 287, + 405 + ], + "lines": [ + { + "bbox": [ + 46, + 372, + 287, + 405 + ], + "spans": [ + { + "bbox": [ + 46, + 372, + 287, + 405 + ], + "type": "text", + "content": "Figure 7. Visual verification of the effectiveness of the proposed components. B, D, and O denote baseline, DWD, and OER. Note that " + }, + { + "bbox": [ + 46, + 372, + 287, + 405 + ], + "type": "inline_equation", + "content": "\\mathrm{B + D + O}" + }, + { + "bbox": [ + 46, + 372, + 287, + 405 + ], + "type": "text", + "content": " is our proposed FEDER." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 420, + 141, + 433 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 420, + 141, + 433 + ], + "spans": [ + { + "bbox": [ + 47, + 420, + 141, + 433 + ], + "type": "text", + "content": "4.3. Ablation Study" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 438, + 287, + 462 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 438, + 287, + 462 + ], + "spans": [ + { + "bbox": [ + 46, + 438, + 287, + 462 + ], + "type": "text", + "content": "We conduct the ablation study on the two largest datasets, namely " + }, + { + "bbox": [ + 46, + 438, + 287, + 462 + ], + "type": "inline_equation", + "content": "COD10K" + }, + { + "bbox": [ + 46, + 438, + 287, + 462 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 438, + 287, + 462 + ], + "type": "inline_equation", + "content": "NC4K" + }, + { + "bbox": [ + 46, + 438, + 287, + 462 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 462, + 287, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 462, + 287, + 594 + ], + "spans": [ + { + "bbox": [ + 46, + 462, + 287, + 594 + ], + "type": "text", + "content": "Effect of DWD component. We demonstrate the effectiveness of our DWD component both in visual verification (see Fig. 7) and quantitative analysis (see Tab. 2a). In Fig. 7, the networks with DWD component (Rows 3 and 5) exhibit more accurate localizations and stronger redundant information suppression capacities. In addition, we present more detailed information in Tab. 2a to verify the validity of each part in the DWD component. As presented in Tab. 2a, we prove the superiority of the DWD component (in (a)), learnable wavelet-like decomposition strategy (in (b)), HFA/LFA module (in (c) and (d)), and GFA module (in (e))." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "text", + "content": "Effect of OER module. The efficacy of our OER module is demonstrated visually by Figs. 7 and 8. In Fig. 7, methods with the OER module can generate the prediction maps with clearer edges. Besides, Fig. 8 illustrates the advancement of our OER module in generating accurate and clear boundary information. We provide detailed information about the superiority of our OER module in Tab. 2b. Specifically, (b) and (c) verify the effectiveness of our weighted gate mechanism (learnable coefficient) and Hamiltonian system. We further compare the performance of the OER module with" + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 309, + 209, + 543, + 330 + ], + "blocks": [ + { + "bbox": [ + 309, + 209, + 543, + 330 + ], + "lines": [ + { + "bbox": [ + 309, + 209, + 543, + 330 + ], + "spans": [ + { + "bbox": [ + 309, + 209, + 543, + 330 + ], + "type": "image", + "image_path": "3f0873bdd35e953d7658e995808541f94246ae4772f04327e94e2abc3cdf0dfd.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 306, + 331, + 545, + 353 + ], + "lines": [ + { + "bbox": [ + 306, + 331, + 545, + 353 + ], + "spans": [ + { + "bbox": [ + 306, + 331, + 545, + 353 + ], + "type": "text", + "content": "Figure 8. Visualizations of the reconstructed edge from the OER module, where the edges are marked in red for well-read." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 368, + 546, + 429 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 368, + 546, + 429 + ], + "spans": [ + { + "bbox": [ + 305, + 368, + 546, + 429 + ], + "type": "text", + "content": "different Runge-Kutta methods, i.e., RB (RK1), RK2, and RK4. Notably, the COD results with RK2 significantly outperform those with RB and are slightly lower than that with RK4. Therefore, we integrate RK2 into our OER module for a balance of performance and efficiency." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 439, + 384, + 451 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 439, + 384, + 451 + ], + "spans": [ + { + "bbox": [ + 306, + 439, + 384, + 451 + ], + "type": "text", + "content": "5. Conclusions" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 459, + 546, + 614 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 459, + 546, + 614 + ], + "spans": [ + { + "bbox": [ + 304, + 459, + 546, + 614 + ], + "type": "text", + "content": "To address the IS and ED challenges, in this paper, we propose the FEDER model for COD. Specifically, we decompose the features into different frequency bands with learnable wavelets and filter out the most informative bands to excavate the subtle discriminative features with the HFA, LFA, and GFA modules, thereby solving the IS challenge. Besides, we propose to learn an auxiliary edge reconstruction task with our OER module to generate complete edges. Learning this auxiliary task along with the COD task thus facilitates the generation of precise segmentation results with accurate object boundaries, thus mitigating the ED challenge. Extensive experiments verify the superiority of our FEDER model in comparison with other SOTAs." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 615, + 547, + 710 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 615, + 547, + 710 + ], + "spans": [ + { + "bbox": [ + 304, + 615, + 547, + 710 + ], + "type": "text", + "content": "Acknowledgements: This research is partly supported by the National Key R&D Program of China (Grant No. 2020AAA0108303), and Shenzhen Science and Technology Project (Grant No. JCYJ20200109143041798) & Shenzhen Stable Supporting Program (WDZC20200820200655001) & Shenzhen Key Laboratory of next generation interactive media innovative technology (Grant No. ZDSYS20210623092001004)." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "type": "text", + "content": "22053" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 287, + 724 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 133 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 133 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 133 + ], + "type": "text", + "content": "[1] Bo Chang, Lili Meng, Eldad Haber, Lars Ruthotto, David Begert, and Elliot Holtham. Reversible architectures for arbitrarily deep residual neural networks. In AAAI, 2018. 3, 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 135, + 287, + 178 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 135, + 287, + 178 + ], + "spans": [ + { + "bbox": [ + 53, + 135, + 287, + 178 + ], + "type": "text", + "content": "[2] Xuelian Cheng, Huan Xiong, Deng-Ping Fan, Yiran Zhong, Mehrtash Harandi, Tom Drummond, and Zongyuan Ge. Implicit motion handling for video camouflaged object detection. In CVPR, pages 13864-13873, 2022. 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 178, + 287, + 212 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 178, + 287, + 212 + ], + "spans": [ + { + "bbox": [ + 53, + 178, + 287, + 212 + ], + "type": "text", + "content": "[3] Deng-Ping Fan, Ming-Ming Cheng, Yun Liu, Tao Li, and Ali Borji. Structure-measure: A new way to evaluate foreground maps. In ICCV, pages 4548-4557, 2017. 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 212, + 287, + 243 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 212, + 287, + 243 + ], + "spans": [ + { + "bbox": [ + 53, + 212, + 287, + 243 + ], + "type": "text", + "content": "[4] Deng-Ping Fan, Ge-Peng Ji, Ming-Ming Cheng, and Ling Shao. Concealed object detection. IEEE Trans. Pattern Anal. Mach. Intell., 2021. 3, 6, 7, 8" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 245, + 287, + 277 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 245, + 287, + 277 + ], + "spans": [ + { + "bbox": [ + 53, + 245, + 287, + 277 + ], + "type": "text", + "content": "[5] Deng-Ping Fan, Ge-Peng Ji, Xuebin Qin, and Ming-Ming Cheng. Cognitive vision inspired object segmentation metric and loss function. Sci. Sin. Inf., 6:6, 2021. 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 278, + 287, + 310 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 278, + 287, + 310 + ], + "spans": [ + { + "bbox": [ + 53, + 278, + 287, + 310 + ], + "type": "text", + "content": "[6] Deng-Ping Fan, Ge-Peng Ji, Guolei Sun, Ming-Ming Cheng, Jianbing Shen, and Ling Shao. Camouflaged object detection. In CVPR, pages 2777-2787, 2020. 2, 3, 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 311, + 287, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 311, + 287, + 354 + ], + "spans": [ + { + "bbox": [ + 53, + 311, + 287, + 354 + ], + "type": "text", + "content": "[7] Shang-Hua Gao, Ming-Ming Cheng, Kai Zhao, Xin-Yu Zhang, Ming-Hsuan Yang, and Philip Torr. Res2net: A new multi-scale backbone architecture. IEEE Trans. Pattern Anal. Mach. Intell., 43(2):652-662, 2019. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 354, + 287, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 354, + 287, + 397 + ], + "spans": [ + { + "bbox": [ + 53, + 354, + 287, + 397 + ], + "type": "text", + "content": "[8] Wooseok Ha, Chandan Singh, Francois Lanusse, Srigokul Upadhyayula, and Bin Yu. Adaptive wavelet distillation from neural networks through interpretations. NIPS, 34:20669-20682, 2021. 3, 4" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 399, + 287, + 419 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 399, + 287, + 419 + ], + "spans": [ + { + "bbox": [ + 53, + 399, + 287, + 419 + ], + "type": "text", + "content": "[9] Eldad Haber and Lars Ruthotto. Stable architectures for deep neural networks. Inverse Probl, 34(1):014004, 2017. 5, 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 420, + 287, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 420, + 287, + 453 + ], + "spans": [ + { + "bbox": [ + 48, + 420, + 287, + 453 + ], + "type": "text", + "content": "[10] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, pages 770-778, 2016. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 453, + 287, + 496 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 453, + 287, + 496 + ], + "spans": [ + { + "bbox": [ + 48, + 453, + 287, + 496 + ], + "type": "text", + "content": "[11] Jianqin Yin Yanbin Han Wendi Hou and Jinping Li. Detection of the mobile object with camouflage color under dynamic background based on optical flow. Procedia Eng., 15:2201-2205, 2011. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 497, + 287, + 529 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 497, + 287, + 529 + ], + "spans": [ + { + "bbox": [ + 48, + 497, + 287, + 529 + ], + "type": "text", + "content": "[12] Sergey Ioffe and Christian Szegedy. Batch normalization: Accelerating deep network training by reducing internal covariate shift. In ICML, pages 448-456. PMLR, 2015. 4" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 529, + 287, + 572 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 529, + 287, + 572 + ], + "spans": [ + { + "bbox": [ + 48, + 529, + 287, + 572 + ], + "type": "text", + "content": "[13] Ge-Peng Ji, Lei Zhu, Mingchen Zhuge, and Keren Fu. Fast camouflaged object detection via edge-based reversible recalibration network. Pattern Recognit., 123:108414, 2022. 5" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 574, + 287, + 616 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 574, + 287, + 616 + ], + "spans": [ + { + "bbox": [ + 48, + 574, + 287, + 616 + ], + "type": "text", + "content": "[14] Qi Jia, Shuilian Yao, Yu Liu, Xin Fan, Risheng Liu, and Zhongxuan Luo. Segment, magnify and reiterate: Detecting camouflaged objects the hard way. In CVPR, pages 4713-4722, 2022. 1, 2, 6" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 617, + 287, + 650 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 617, + 287, + 650 + ], + "spans": [ + { + "bbox": [ + 48, + 617, + 287, + 650 + ], + "type": "text", + "content": "[15] Liming Jiang, Bo Dai, Wayne Wu, and Chen Change Loy. Focal frequency loss for image reconstruction and synthesis. In ICCV, pages 13919-13929, 2021. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 48, + 651, + 287, + 693 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 651, + 287, + 693 + ], + "spans": [ + { + "bbox": [ + 48, + 651, + 287, + 693 + ], + "type": "text", + "content": "[16] Zhanghan Ke, Jiayu Sun, Kaican Li, Qiong Yan, and Rynson WH Lau. Modnet: Real-time trimap-free portrait matting via objective decomposition. In AAAI, pages 1140-1147, 2022. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 48, + 693, + 287, + 724 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 693, + 287, + 724 + ], + "spans": [ + { + "bbox": [ + 48, + 693, + 287, + 724 + ], + "type": "text", + "content": "[17] Ajay Kumar. Computer-vision-based fabric defect detection: A survey. IEEE Trans. Ind. Electron., 55(1):348-363, 2008. 1" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 546, + 701 + ], + "type": "list", + "angle": 0, + "index": 34, + "blocks": [ + { + "bbox": [ + 307, + 73, + 546, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 546, + 116 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 546, + 116 + ], + "type": "text", + "content": "[18] Trung-Nghia Le, Tam V Nguyen, Zhongliang Nie, Minh-Triet Tran, and Akihiro Sugimoto. Anabranch network for camouflaged object segmentation. Comput. Vis. Image Underst., 184:45-56, 2019. 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 118, + 546, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 546, + 172 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 546, + 172 + ], + "type": "text", + "content": "[19] Bei Li, Quan Du, Tao Zhou, Yi Jing, Shuhan Zhou, Xin Zeng, Tong Xiao, Jingbo Zhu, Xuebo Liu, and Min Zhang. Ode transformer: An ordinary differential equation-inspired model for sequence generation. In ACL, pages 8335-8351, 2022. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 174, + 546, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 174, + 546, + 196 + ], + "spans": [ + { + "bbox": [ + 307, + 174, + 546, + 196 + ], + "type": "text", + "content": "[20] Boyi Li, Felix Wu, Kilian Q Weinberger, and Serge Belongie. Positional normalization. NIPS, 32, 2019. 4" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 197, + 546, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 197, + 546, + 229 + ], + "spans": [ + { + "bbox": [ + 307, + 197, + 546, + 229 + ], + "type": "text", + "content": "[21] Juncheng Li, Faming Fang, Kangfu Mei, and Guixu Zhang. Multi-scale residual network for image super-resolution. In ECCV, pages 517-532, 2018. 4" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 231, + 546, + 263 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 231, + 546, + 263 + ], + "spans": [ + { + "bbox": [ + 307, + 231, + 546, + 263 + ], + "type": "text", + "content": "[22] Kai Li, Yulun Zhang, Kunpeng Li, and Yun Fu. Adversarial feature hallucination networks for few-shot learning. In CVPR, pages 13470-13479, 2020. 4" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 264, + 546, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 264, + 546, + 308 + ], + "spans": [ + { + "bbox": [ + 307, + 264, + 546, + 308 + ], + "type": "text", + "content": "[23] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In ICCV, pages 10012-10022, 2021. 4" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 309, + 546, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 309, + 546, + 364 + ], + "spans": [ + { + "bbox": [ + 307, + 309, + 546, + 364 + ], + "type": "text", + "content": "[24] Laetitia Loncan, Luis B De Almeida, José M Bioucas-Dias, Xavier Briottet, Jocelyn Chanussot, Nicolas Dobigeon, Sophie Fabre, Wenzhi Liao, Giorgio A Licciardi, Miguel Simoes, et al. Hyperspectral pansharpening: A review. IEEE Geosci. Remote Sens. Mag., 3(3):27-46, 2015. 4" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 365, + 546, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 365, + 546, + 408 + ], + "spans": [ + { + "bbox": [ + 307, + 365, + 546, + 408 + ], + "type": "text", + "content": "[25] Yunqiu Lv, Jing Zhang, Yuchao Dai, Aixuan Li, Bowen Liu, Nick Barnes, and Deng-Ping Fan. Simultaneously localize, segment and rank the camouflaged objects. In CVPR, pages 11591-11601, 2021. 6, 8" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 410, + 546, + 452 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 410, + 546, + 452 + ], + "spans": [ + { + "bbox": [ + 307, + 410, + 546, + 452 + ], + "type": "text", + "content": "[26] Yue Ma, Yali Wang, Yue Wu, Ziyu Lyu, Siran Chen, Xiu Li, and Yu Qiao. Visual knowledge graph for human action reasoning in videos. In ACM MM, pages 4132-4141, 2022. 4" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 455, + 546, + 486 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 455, + 546, + 486 + ], + "spans": [ + { + "bbox": [ + 307, + 455, + 546, + 486 + ], + "type": "text", + "content": "[27] Ran Margolin, Lihi Zelnik-Manor, and Ayellet Tal. How to evaluate foreground maps? In CVPR, pages 248-255, 2014. 6" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 488, + 546, + 532 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 488, + 546, + 532 + ], + "spans": [ + { + "bbox": [ + 307, + 488, + 546, + 532 + ], + "type": "text", + "content": "[28] Haiyang Mei, Ge-Peng Ji, Ziqi Wei, Xin Yang, Xiaopeng Wei, and Deng-Ping Fan. Camouflaged object segmentation with distraction mining. In CVPR, pages 8772-8781, 2021. 2, 6" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 307, + 533, + 546, + 576 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 533, + 546, + 576 + ], + "spans": [ + { + "bbox": [ + 307, + 533, + 546, + 576 + ], + "type": "text", + "content": "[29] Fausto Milletari, Nassir Navab, and Seyed-Ahmad Ahmadi. V-net: Fully convolutional neural networks for volumetric medical image segmentation. In 3DV, pages 565-571. IEEE, 2016. 6" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 307, + 578, + 546, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 578, + 546, + 611 + ], + "spans": [ + { + "bbox": [ + 307, + 578, + 546, + 611 + ], + "type": "text", + "content": "[30] Yuxin Pan, Yiwang Chen, Qiang Fu, Ping Zhang, and Xin Xu. Study on the camouflaged target detection method based on 3d convexity. Mod. Appl. Sci., 5(4):152, 2011. 2" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 307, + 613, + 546, + 655 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 613, + 546, + 655 + ], + "spans": [ + { + "bbox": [ + 307, + 613, + 546, + 655 + ], + "type": "text", + "content": "[31] Youwei Pang, Xiaoqi Zhao, Tian-Zhu Xiang, Lihe Zhang, and Huchuan Lu. Zoom in and out: A mixed-scale triplet network for camouflaged object detection. In CVPR, pages 2160-2170, 2022. 6" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 307, + 657, + 546, + 701 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 657, + 546, + 701 + ], + "spans": [ + { + "bbox": [ + 307, + 657, + 546, + 701 + ], + "type": "text", + "content": "[32] Jialun Pei, Tianyang Cheng, Deng-Ping Fan, He Tang, Chuanbo Chen, and Luc Van Gool. Osformer: One-stage camouflaged instance segmentation with transformers. In ECCV, pages 19-37. Springer, 2022. 6" + } + ] + } + ], + "index": 33 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "22054" + } + ] + } + ], + "index": 35 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 454 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "text", + "content": "[33] Natasha Price, Samuel Green, Jolyon Troscianko, Tom Tre-genza, and Martin Stevens. Background matching and disruptive coloration as habitat-specific strategies for camouflage. Sci. Rep., 9(1):1-10, 2019. 1" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 118, + 287, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 118, + 287, + 160 + ], + "spans": [ + { + "bbox": [ + 49, + 118, + 287, + 160 + ], + "type": "text", + "content": "[34] Xu Qin, Zhilin Wang, Yuanchao Bai, Xiaodong Xie, and Huizhu Jia. Ffa-net: Feature fusion attention network for single image dehazing. In AAAI, pages 11908-11915, 2020. 4" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 163, + 288, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 163, + 288, + 217 + ], + "spans": [ + { + "bbox": [ + 48, + 163, + 288, + 217 + ], + "type": "text", + "content": "[35] Dan Jeric Arcega Rustia, Chien Erh Lin, Jui-Yung Chung, Yi-Ji Zhuang, Ju-Chun Hsu, and Ta-Te Lin. Application of an image and environmental sensor network for automated greenhouse insect pest monitoring. J. Asia Pac. Entomol., 23(1):17-28, 2020. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 219, + 287, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 219, + 287, + 239 + ], + "spans": [ + { + "bbox": [ + 48, + 219, + 287, + 239 + ], + "type": "text", + "content": "[36] Phil Sallee and Bruno Olshausen. Learning sparse multiscale image representations. NIPS, 15, 2002. 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 241, + 287, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 241, + 287, + 274 + ], + "spans": [ + { + "bbox": [ + 48, + 241, + 287, + 274 + ], + "type": "text", + "content": "[37] Jiawei Shen, Zhuoyan Li, Lei Yu, Gui-Song Xia, and Wen Yang. Implicit euler ode networks for single-image dehazing. In CVPRW, pages 218-219, 2020. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 276, + 287, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 276, + 287, + 308 + ], + "spans": [ + { + "bbox": [ + 48, + 276, + 287, + 308 + ], + "type": "text", + "content": "[38] Chenyang Si, Weihao Yu, Pan Zhou, Yichen Zhou, Xinchao Wang, and Shuicheng Yan. Inception transformer. arXiv preprint arXiv:2205.12956, 2022. 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 309, + 287, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 309, + 287, + 353 + ], + "spans": [ + { + "bbox": [ + 48, + 309, + 287, + 353 + ], + "type": "text", + "content": "[39] Przemysław Skurowski, Hassan Abdulameer, J Błaszczyk, Tomasz Depta, Adam Kornacki, and P Koziel. Animal camouflage analysis: Chameleon database. Unpublished manuscript, 2(6):7, 2018. 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 354, + 287, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 354, + 287, + 387 + ], + "spans": [ + { + "bbox": [ + 48, + 354, + 287, + 387 + ], + "type": "text", + "content": "[40] Radomir S Stanković and Bogdan J Falkowski. The haar wavelet transform: its status and achievements. Comput. Electr. Eng., 29(1):25-44, 2003. 4, 8" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 388, + 287, + 421 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 388, + 287, + 421 + ], + "spans": [ + { + "bbox": [ + 48, + 388, + 287, + 421 + ], + "type": "text", + "content": "[41] Martin Stevens and Sami Merilaita. Animal camouflage: current issues and new perspectives. Philos. Trans. R. Soc. Lond., B, Biol. Sci., 364(1516):423-427, 2009. 1, 2, 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 422, + 287, + 454 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 422, + 287, + 454 + ], + "spans": [ + { + "bbox": [ + 48, + 422, + 287, + 454 + ], + "type": "text", + "content": "[42] Cheng Tai and E Weinan. Multiscale adaptive representation of signals: I. the basic framework. J. Mach. Learn. Res., 17(1):4875-4912, 2016. 4" + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 453 + ], + "type": "list", + "angle": 0, + "index": 21, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 105 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 105 + ], + "type": "text", + "content": "[43] Christopher Torrence and Gilbert P Compo. A practical guide to wavelet analysis. Bull. Am. Meteorol. Soc., 79(1):61-78, 1998. 4" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 307, + 107, + 545, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 107, + 545, + 140 + ], + "spans": [ + { + "bbox": [ + 307, + 107, + 545, + 140 + ], + "type": "text", + "content": "[44] Dmitry Ulyanov, Andrea Vedaldi, and Victor Lempitsky. Instance normalization: The missing ingredient for fast stylization. arXiv preprint arXiv:1607.08022, 2016. 4" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 307, + 141, + 545, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 141, + 545, + 184 + ], + "spans": [ + { + "bbox": [ + 307, + 141, + 545, + 184 + ], + "type": "text", + "content": "[45] Qilong Wang, Banggu Wu, Pengfei Zhu, Peihua Li, Wangmeng Zuo, and Qinghua Hu. Eca-net: Efficient channel attention for deep convolutional neural networks. In CVPR, pages 13-19, 2020. 4" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 307, + 186, + 545, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 186, + 545, + 207 + ], + "spans": [ + { + "bbox": [ + 307, + 186, + 545, + 207 + ], + "type": "text", + "content": "[46] E Weinan. A proposal on machine learning via dynamical systems. Commun. Math. Stat., 1(5):1-11, 2017. 3, 5" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 209, + 545, + 241 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 209, + 545, + 241 + ], + "spans": [ + { + "bbox": [ + 307, + 209, + 545, + 241 + ], + "type": "text", + "content": "[47] Zhe Wu, Li Su, and Qingming Huang. Cascaded partial decoder for fast and accurate salient object detection. In CVPR, pages 3907-3916, 2019. 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 243, + 545, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 243, + 545, + 285 + ], + "spans": [ + { + "bbox": [ + 307, + 243, + 545, + 285 + ], + "type": "text", + "content": "[48] Fan Yang, Qiang Zhai, Xin Li, Rui Huang, Ao Luo, Hong Cheng, and Deng-Ping Fan. Uncertainty-guided transformer reasoning for camouflaged object detection. In ICCV, pages 4146-4155, 2021. 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 287, + 545, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 287, + 545, + 319 + ], + "spans": [ + { + "bbox": [ + 307, + 287, + 545, + 319 + ], + "type": "text", + "content": "[49] Jaejun Yoo, Youngjung Uh, Sanghyuk Chun, Byeongkyu Kang, and Jung-Woo Ha. Photorealistic style transfer via wavelet transforms. In ICCV, pages 9036-9045, 2019. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 320, + 545, + 362 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 320, + 545, + 362 + ], + "spans": [ + { + "bbox": [ + 307, + 320, + 545, + 362 + ], + "type": "text", + "content": "[50] Qiang Zhai, Xin Li, Fan Yang, Chenglizhao Chen, Hong Cheng, and Deng-Ping Fan. Mutual graph learning for camouflaged object detection. In CVPR, pages 12997-13007, 2021. 5, 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 364, + 545, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 364, + 545, + 407 + ], + "spans": [ + { + "bbox": [ + 307, + 364, + 545, + 407 + ], + "type": "text", + "content": "[51] Yulun Zhang, Zhifei Zhang, Stephen DiVerdi, Zhaowen Wang, Jose Echevarria, and Yun Fu. Texture hallucination for large-factor painting super-resolution. In ECCV, pages 209-225. Springer, 2020. 8" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 409, + 545, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 409, + 545, + 453 + ], + "spans": [ + { + "bbox": [ + 307, + 409, + 545, + 453 + ], + "type": "text", + "content": "[52] Hongwei Zhu, Peng Li, Haoran Xie, Xuefeng Yan, Dong Liang, Dapeng Chen, Mingqiang Wei, and Jing Qin. I can find you! boundary-guided separated attention network for camouflaged object detection. In AAAI, 2022. 6" + } + ] + } + ], + "index": 20 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "22055" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Can't Steal_ Cont-Steal! Contrastive Stealing Attacks Against Image Encoders/d755a24e-203a-45cb-a84e-020f7fb471b0_content_list.json b/2023/Can't Steal_ Cont-Steal! Contrastive Stealing Attacks Against Image Encoders/d755a24e-203a-45cb-a84e-020f7fb471b0_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..ea7e403575c9ffd32fbcbcee1045684f04df499a --- /dev/null +++ b/2023/Can't Steal_ Cont-Steal! Contrastive Stealing Attacks Against Image Encoders/d755a24e-203a-45cb-a84e-020f7fb471b0_content_list.json @@ -0,0 +1,1771 @@ +[ + { + "type": "text", + "text": "Can't Steal? Cont-Steal! Contrastive Stealing Attacks Against Image Encoders", + "text_level": 1, + "bbox": [ + 212, + 130, + 756, + 175 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Zeyang Sha† Xinlei He† Ning Yu‡ Michael Backes† Yang Zhang†", + "bbox": [ + 200, + 202, + 754, + 222 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "†CISPA Helmholtz Center for Information Security ‡Salesforce Research {zeyang.sha, xinlei.he, director, zhang}@cispa.de, ning.yu@salesforce.com", + "bbox": [ + 158, + 238, + 805, + 273 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 308, + 313, + 325 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Self-supervised representation learning techniques have been developing rapidly to make full use of unlabeled images. They encode images into rich features that are oblivious to downstream tasks. Behind their revolutionary representation power, the requirements for dedicated model designs and a massive amount of computation resources expose image encoders to the risks of potential model stealing attacks - a cheap way to mimic the well-trained encoder performance while circumventing the demanding requirements. Yet conventional attacks only target supervised classifiers given their predicted labels and/or posteriors, which leaves the vulnerability of unsupervised encoders unexplored.", + "bbox": [ + 75, + 340, + 472, + 537 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this paper, we first instantiate the conventional stealing attacks against encoders and demonstrate their severer vulnerability compared with downstream classifiers. To better leverage the rich representation of encoders, we further propose Cont-Steal, a contrastive-learning-based attack, and validate its improved stealing effectiveness in various experiment settings. As a takeaway, we appeal to our community's attention to the intellectual property protection of representation learning techniques, especially to the defenses against encoder stealing attacks like ours.", + "bbox": [ + 75, + 537, + 470, + 690 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 717, + 209, + 734 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recent years have witnessed the great success of applying deep learning (DL) to computer vision tasks. Different from supervised DL models, self-supervised learning which transforms unlabeled data samples into rich representations, has gained more and more popularity.", + "bbox": [ + 75, + 743, + 468, + 819 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Behind its powerful representation, it is non-trivial to obtain a state-of-the-art image encoder. For instance, SimCLR [6] uses 128 TPU v3 cores to pre-train a ResNet-", + "bbox": [ + 75, + 819, + 468, + 864 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/e59fe95481358d05a780af7dd1b54c6d4f84ec72885d0eb1614d4d8fb718b935.jpg", + "image_caption": [ + "Figure 1. Model stealing attacks against classifiers (previous) v.s. model stealing attacks against encoders (ours). Previous works aim to steal a whole classifier using the predicted label or posteriors of a target model. In our work, we aim to steal the target encoder using its embeddings. The target encoder $(E_{t})$ is pre-trained and fixed, as shown in the solid frame. The surrogate encoder $(E_{s})$ is trainable by the adversary, as shown in the dashed frame." + ], + "image_footnote": [], + "bbox": [ + 501, + 311, + 870, + 506 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "50 encoder with a batch size of 4096. Therefore, many big companies provide cloud-based self-supervised learning encoder services for users. For instance, Cohere, $^{2}$ OpenAI, $^{3}$ and Clarifai $^{4}$ provide the embedding API of images and texts for commercial usage. There are many works [10, 26, 30] exploring security issues of encoder-based API. Therefore, it is a very important and urgent problem.", + "bbox": [ + 496, + 645, + 890, + 751 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "These kinds of service leave the possibility of model stealing attacks [5, 25, 28, 37, 44, 48, 50, 51]. In these attacks, the adversary aims to steal the parameters or functionalities of target models with only query access to them. A successful model stealing attack does not only threaten", + "bbox": [ + 496, + 752, + 893, + 828 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "See our code in https://github.com/zeyangsha/Cont-Steal.", + "bbox": [ + 76, + 875, + 467, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "$^{2}$ https://cohere.ai/ \n $^{3}$ https://beta.openai.com/docs/api-reference/embeddings/ \n $^{4}$ https://www clarifai.com/models/general-image-embedding/", + "bbox": [ + 500, + 837, + 890, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "16373", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "the intellectual property of the target model, but also serves as a stepping stone for further attacks such as adversarial examples [3, 4, 16, 38, 47, 52], backdoor attacks [7, 26, 41, 43], and membership inference attacks [22-24, 30, 33, 34, 40, 42, 45, 46]. So far, model stealing attacks concentrate on the supervised classifiers, i.e., the model responses are prediction posteriors or labels for a specific downstream task. The vulnerability of unsupervised image encoders is unfortunately unexplored.", + "bbox": [ + 75, + 90, + 468, + 224 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our Work. To fill this gap, we pioneer the systematic investigation of model stealing attacks against image encoders. In this work, the adversary's goal is to steal the functionalities of the target model. See Figure 1 for an overview and a comparison with previous works. More specifically, we focus on encoders trained by contrastive learning, which is one of the most cutting-edge unsupervised representation learning strategies that unleash the information of unlabeled data.", + "bbox": [ + 75, + 228, + 468, + 364 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We first instantiate the conventional stealing attacks against encoders and expose their vulnerability. Given an input image, the target encoder outputs its representation (referred to as embedding). Similar to model stealing attacks against classifiers, we consider the embedding as the \"ground truth\" label to guide the training procedure of a surrogate encoder on the adversary side. To measure the effectiveness of stealing attacks, we train an extra linear layer for the target and surrogate encoders towards the same downstream classification task. Preferably, the surrogate model should achieve both high classification accuracy and high agreement with the target predictions.", + "bbox": [ + 75, + 366, + 468, + 547 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We evaluate our attacks on five datasets against four contrastive learning encoders. Our results demonstrate that the conventional attacks are more effective against encoders than against downstream classifiers. For instance, when we steal the downstream classifier pre-trained by SimCLR on CIFAR10 (with posteriors as its responses) using STL10 as the surrogate dataset, the adversary can only achieve an accuracy of 0.359. The accuracy, however, increases to 0.500 instead when we steal its encoder (with the embedding as its response).", + "bbox": [ + 75, + 549, + 468, + 700 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Despite its encouraging performance, conventional attacks are not the most suitable ones against encoders. This is because they treat each image-embedding pair individually without interacting across pairs. Different embeddings are beneficial to each other as they can serve as anchors to better locate the position of the other embeddings in their space. Contrastive learning [6, 8, 17, 20, 27, 49, 53] is a straightforward idea to achieve this goal. It is formulated to enforce the embeddings of different augmentations of the same images closer and those of different images further.", + "bbox": [ + 75, + 702, + 468, + 852 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In a similar spirit, we propose Cont-Steal, a contrastive-learning-based model stealing attack against the encoder. The goal of Cont-Steal is to enforce the surrogate embed", + "bbox": [ + 76, + 854, + 468, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "ding of an image close to its target embedding (defined as a positive pair) and also push away embeddings of different images irrespective of being generated by the target or the surrogate encoders (defined as negative pairs).", + "bbox": [ + 496, + 90, + 890, + 151 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The comprehensive evaluation shows that Cont-Steal outperforms the conventional model stealing attacks to a large extent. For instance, when CIFAR10 is the target dataset, Cont-Steal achieves an accuracy of 0.714 on the SimCLR encoder pretrained on CIFAR10 with the surrogate dataset and downstream dataset being STL10, while the conventional attack only achieves 0.457 accuracy. Also, Cont-Steal is more query-efficient and dataset-independent (see Figure 9 for more details). This is because Cont-Steal leverages higher-order information across samples to mimic the functionality of the target encoder. To mitigate the attacks, we evaluate different defense mechanisms including noise, top- $k$ , rounding, and watermark. Our evaluations show that in most cases, these mechanisms cannot effectively defend against Cont-Steal. Among them, top- $k$ can reduce the attack performance to the largest extent. However, it also strongly limits the target model's utility.", + "bbox": [ + 496, + 152, + 892, + 407 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "As a takeaway, our attack further exposes the severe vulnerability of pre-trained encoders. We appeal to our community's attention to the intellectual property protection of representation learning techniques, especially to the defenses against encoder stealing attacks like ours.", + "bbox": [ + 496, + 407, + 890, + 484 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Threat Model", + "text_level": 1, + "bbox": [ + 500, + 497, + 640, + 513 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this work, for the encoder pre-trained with images, we consider image classification as the downstream task. We refer to the encoder as the target encoder. Then we treat both the encoder and the linear layer trained for the downstream task together as the target model. We first introduce the adversary's goal and then characterize different background knowledge that the adversary might have.", + "bbox": [ + 496, + 523, + 890, + 628 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Adversary's Goal. Following previous work [25, 28, 44], we taxonomize the adversary's goal into two dimensions, i.e., theft and utility. The theft adversary aims to build a surrogate encoder that has similar performance on the downstream tasks as the target encoder. Different from the thief adversary, the goal of the utility adversary is to construct a surrogate encoder that behaves normally on different downstream tasks. In this case, the surrogate encoder not only faithfully \"copies\" the behaviors of the target encoder, but also serves as a stepping stone to conduct other attacks.", + "bbox": [ + 496, + 628, + 890, + 779 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Adversary's Background Knowledge. We categorize the adversary's background knowledge into two dimensions, i.e., the knowledge of the target encoder and the distribution of the surrogate dataset.", + "bbox": [ + 496, + 779, + 890, + 839 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Regarding knowledge of the target encoder, we assume that the adversary only has black-box access to it, which means that they can only query the target encoder with an input image and obtain the corresponding output, i.e., the", + "bbox": [ + 496, + 839, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "16374", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "embedding of the input image.", + "bbox": [ + 76, + 90, + 281, + 106 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Regarding the surrogate dataset that is used to train the surrogate encoder, we consider two cases. First, we assume the adversary has the same training dataset as the target encoder. However, such an assumption may be hard to achieve as such datasets are usually private and protected by the model owner. In a more extreme case, we assume that the adversary has totally no information about the target encoder's training dataset, which means that they can only use a different distribution dataset to conduct the model stealing attacks. We later show that the adversary can still launch effective model stealing attacks against the target encoder given a surrogate dataset that is distributed differently compared to the target dataset.", + "bbox": [ + 75, + 106, + 468, + 301 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "For the model architecture that is used to train the surrogate encoder, we consider two cases. First, we assume the adversary is aware of the target encoder's architecture and can train the same architecture surrogate encoder. Then we relax our assumption that the adversary uses different architectures to train the surrogate encoder. Our evaluation shows that the choice of architecture does not have much impact on the attack performance (see Table 3), which makes the attack more realistic.", + "bbox": [ + 75, + 301, + 468, + 436 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Note that we also compare our attacks against the encoders to the traditional model stealing attacks that focus on the whole classifier (which has an encoder and a linear layer). If the attack targets a whole classifier, we assume the adversary may obtain the posteriors or the predicted label for an input image.", + "bbox": [ + 75, + 438, + 468, + 529 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Model Stealing Attacks", + "text_level": 1, + "bbox": [ + 76, + 541, + 295, + 558 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we first describe the conventional attacks against the encoders. Then, we propose a novel contrastive stealing framework, Cont-Steal, to steal the encoders more effectively.", + "bbox": [ + 75, + 566, + 468, + 628 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Conventional Attacks Against Encoders", + "text_level": 1, + "bbox": [ + 76, + 635, + 418, + 651 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The adversary takes two steps to conduct the model stealing attacks against the target encoder and one step for further evaluation.", + "bbox": [ + 75, + 659, + 468, + 702 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Obtain the Surrogate Dataset. To conduct model stealing attacks, the adversary first needs to obtain a surrogate dataset. Based on the knowledge of the target classifier's training dataset (target dataset), we consider two cases. If the adversary has full knowledge of the target dataset, they can directly leverage the target dataset itself as the surrogate dataset. Or the adversary has no knowledge of the target dataset, which means that they can only construct the surrogate dataset, which is distributed differently from the target dataset.", + "bbox": [ + 75, + 704, + 468, + 852 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Train the Surrogate Encoder. Slightly different from the classifier, the response of the encoder is an embedding, which is a feature vector. In this case, the adversary can still", + "bbox": [ + 75, + 854, + 468, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "leverage a similar loss function to optimize the surrogate encoder, which can be defined as follows:", + "bbox": [ + 498, + 90, + 890, + 119 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nL _ {M S} = \\sum_ {k = 1} ^ {N} l \\left(E _ {T} \\left(x _ {k}\\right), E _ {S} \\left(x _ {k}\\right)\\right) \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 578, + 133, + 890, + 175 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $E_{T}(\\cdot) / E_{S}(\\cdot)$ is the target/surrogate encoder, $N$ is the total number of samples on the surrogate dataset, and $l(\\cdot)$ is the MSE loss.", + "bbox": [ + 498, + 183, + 890, + 227 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Apply the Surrogate Encoder to Downstream Tasks. To evaluate the effectiveness of model stealing attacks against the encoder, the adversary can leverage the same downstream task to both the target and surrogate encoders. Concretely, the adversary trains an extra linear layer for the target and surrogate encoders, respectively. Note that we refer to the target/surrogate encoder and the extra linear layer as the target/surrogate classifiers. Then, the adversary quantifies the attack effectiveness by measuring the performance of the target/surrogate classifier on the downstream tasks.", + "bbox": [ + 496, + 229, + 890, + 380 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Cont-Steal Attacks Against Encoders", + "text_level": 1, + "bbox": [ + 498, + 391, + 820, + 407 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To better leverage the rich information from the embeddings, we propose Cont-Steal, a contrastive learning-based model stealing attacks against encoders, which leverages contrastive learning to enhance the stealing performance. Concretely, Cont-Steal aims to enforce the surrogate embedding of an image to get close to its target embedding (defined as a positive pair), and also push away embeddings of different images regardless of being generated by the target or the surrogate encoders (defined as negative pairs). There are three steps for the adversary to conduct contrastive stealing attacks against encoders and one step for further evaluation.", + "bbox": [ + 496, + 415, + 890, + 594 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Obtain the Surrogate Dataset. The adversary follows the same strategy as Section 3.1 to obtain the surrogate dataset. Data Augmentation. Our proposed Cont-Steal leverages data augmentation to transform an input image into its two augmented views. In this paper, we leverage RandAugment [11] as the augmentation method, which is made up of a group of advanced augmentation operations. Concretely, we set $n = 2$ and $m = 14$ following Cubuk et al. [11] where $n$ denotes the number of transformations to a given sample and $m$ represents the magnitude of global distortion.", + "bbox": [ + 496, + 597, + 890, + 747 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Train the Surrogate Encoder. Instead of querying the encoders with the original images, the adversary queries the encoders with the augmented views of them. Concretely, for an input image $x_{i}$ , we generate two augmented views of it, i.e., $\\widetilde{x}_{i,s}$ and $\\widetilde{x}_{i,t}$ , where $\\widetilde{x}_{i,s} / \\widetilde{x}_{i,t}$ is used to query the surrogate/target encoder. We consider $(\\widetilde{x}_{i,s},\\widetilde{x}_{j,t})$ as a positive pair if $i = j$ , and otherwise a negative pair.", + "bbox": [ + 496, + 750, + 890, + 854 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Given a mini-batch of $N$ samples, we generate $N$ augmented views as the input of the target encoder and another $N$ augmented views as the input of the surrogate encoders.", + "bbox": [ + 500, + 854, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "16375", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Concretely, the loss of Cont-Steal can be formulated as follows:", + "bbox": [ + 75, + 90, + 468, + 119 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nD _ {\\text {e n c o d e r}} ^ {+} (i) = \\exp \\left(\\operatorname {s i m} \\left(E _ {S} \\left(\\tilde {x} _ {i, s}\\right), E _ {T} \\left(\\tilde {x} _ {i, t}\\right)\\right) / \\tau\\right), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 91, + 152, + 468, + 169 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nD _ {\\text {e n c o d e r}} ^ {-} (i) = \\sum_ {k = 1} ^ {N} \\left(\\exp \\left(\\operatorname {s i m} \\left(E _ {S} \\left(\\widetilde {x} _ {i, s}\\right), E _ {T} \\left(\\widetilde {x} _ {k, t}\\right)\\right) / \\tau\\right)\\right), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 91, + 172, + 468, + 209 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nD _ {s e l f} ^ {-} (i) = \\sum_ {k = 1} ^ {N} \\mathbb {1} _ {[ k \\neq i ]} \\left(\\exp \\left(\\operatorname {s i m} \\left(E _ {S} \\left(\\widetilde {x} _ {i, s}\\right), E _ {S} \\left(\\widetilde {x} _ {k, s}\\right)\\right) / \\tau\\right)\\right), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 91, + 212, + 468, + 263 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nl (i) = - \\log \\frac {D _ {\\text {e n c o d e r}} ^ {+} (i)}{D _ {\\text {e n c o d e r}} ^ {-} (i) + D _ {\\text {s e l f}} ^ {-} (i)}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 91, + 267, + 468, + 301 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nL _ {\\text {C o n t - S t e a l}} = \\frac {\\sum_ {k = 1} ^ {N} l (k)}{N}, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 91, + 303, + 468, + 332 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $E_{S}(\\cdot)$ and $E_{T}(\\cdot)$ denotes the surrogate and target encoder, $sim(u,v) = u^{T}v / ||u||||v||$ represents the cosine similarity between $u$ and $v$ , and $\\tau$ is parameter to control the temperature.", + "bbox": [ + 75, + 349, + 468, + 409 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As illustrated in Figure 2, the conventional attack treats each embedding individually without interacting across pairs. However, different embeddings are beneficial to each other as they can serve as anchors to better locate the position of the other embeddings in their space. Cont-Steal maximizes the similarity of embeddings generated from the target and surrogate encoders for a positive pair $(\\widetilde{x}_{i,s},\\widetilde{x}_{i,t})$ (orange arrows in Figure 2). For the embedding generated from the target and surrogate encoders for any pair $(\\widetilde{x}_{i,s},\\widetilde{x}_{j,t})$ , contrastive stealing aims to make them more distant (green arrows in Figure 2). Besides, as pointed out by Chen et al. [6], contrastive learning benefits larger negative samples. To achieve this goal, we also consider the embeddings generated from the surrogate encoder for augmented views of different images, i.e., $(\\widetilde{x}_{i,s},\\widetilde{x}_{j,s})$ , as negative pairs minimize their similarity (blue arrows in Figure 2). We later show that such design can enhance the performance of contrastive stealing (see Table 5).", + "bbox": [ + 75, + 412, + 468, + 684 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Apply the Surrogate Encoder to Downstream Tasks. We follow Section 3.1 to evaluate the effectiveness of model stealing on downstream tasks.", + "bbox": [ + 75, + 686, + 468, + 731 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 76, + 752, + 207, + 768 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this section, we first describe the experimental setup in Section 4.1. Then we show the performance of the target encoders on the downstream tasks. Next, we summarize the performance of conventional attacks against classifiers and encoders in Section 4.2. Lastly, we evaluate the performance of Cont-Steal and conduct ablation studies to demonstrate its effectiveness under different settings in Section 4.3.", + "bbox": [ + 75, + 780, + 468, + 898 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/85b287c587173ea3c5b75df0fadd4f0b60bc30a80663ed68b901d2041a071eeb.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 509, + 95, + 869, + 256 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/29b7b7acc13ea667777bc9565837d2ce09ec9557351c5fd6bb3170d8f9e0b789.jpg", + "image_caption": [ + "Figure 2. Conventional attack (top) vs. Cont-Steal (bottom) against encoders. Conventional attack applies MSE loss to approximate target embeddings for each sample individually. Cont-Steal (bottom) introduces data augmentation and interacts across multiple samples: associating target/surrogate embeddings of the same images closer and repulsing those of different images farther away. The target encoder $(E_{t})$ is pre-trained and fixed, as shown in the solid frame. The surrogate encoder $(E_{s})$ is trainable by the adversary, as shown in the dashed frame." + ], + "image_footnote": [], + "bbox": [ + 519, + 265, + 877, + 454 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1. Experimental Setup", + "text_level": 1, + "bbox": [ + 500, + 621, + 689, + 638 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our encoders are pre-trained on CIFAR10 [1], and ImageNet [12]. We use four different kinds of contrastive methods: SimCLR [6], MoCo [20], BYOL [17] and SimSiam [8] to train a ResNet18 [21] as our target encoders. Our implementation is based on a PyTorch framework of contrastive learning. Then, these well pre-trained encoders will be applied to train downstream classifiers on CIFAR10 [1], STL10 [9], Fashion-MNIST [54], and SVHN [35]. In the experiments in the model stealing section, we use CIFAR10, STL10, Fashion-MNIST, and SVHN to conduct the attack.", + "bbox": [ + 496, + 645, + 890, + 810 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Agreement and accuracy are used as metrics to evaluate the model stealing attack's performance. The agreement will evaluate the similarity of surrogate encoders and target encoders in downstream tasks. The accuracy will evaluate", + "bbox": [ + 496, + 813, + 890, + 873 + ], + "page_idx": 3 + }, + { + "type": "page_footnote", + "text": "5https://github.com/vturrisi/solo-learn/", + "bbox": [ + 514, + 886, + 831, + 898 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "16376", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/4119feca9e7f1245c2f3a2a48efecc701d744f55cc9395ac73faa4bba0ec0e13.jpg", + "image_caption": [ + "(a) CIFAR10" + ], + "image_footnote": [], + "bbox": [ + 114, + 88, + 271, + 179 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/3afaced75a678288771aa36a8c03ae19f56260e927b5193ca45e454eac96b500.jpg", + "image_caption": [ + "(b) ImageNet100", + "Figure 3. The performance of target classifiers composed by target encoder and an extra linear layer. The encoders are pre-trained on CIFAR10 (a) and ImageNet100 (b). The x-axis represents different downstream datasets for the target encoder and classifier. The y-axis represents the target model's accuracy on downstream tasks." + ], + "image_footnote": [], + "bbox": [ + 272, + 88, + 429, + 180 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "the utility of surrogate models on downstream tasks. For each metric, a larger value is more desirable. During the stealing process, we set the batch size as 128 and the learning rate as 0.001. We show more results on the impact of hyperparameters in Supplementary Material in Section A.2", + "bbox": [ + 75, + 318, + 468, + 395 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Performance of Conventional Attacks", + "text_level": 1, + "bbox": [ + 76, + 407, + 401, + 422 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We first show the target encoder's performance in various downstream tasks. The results are summarized in Figure 3. We conduct our experiments to explore whether the encoders are more vulnerable to model stealing attacks. We show our results of target encoders and downstream classifiers both trained on CIFAR10 in Figure 4. In all cases, the adversary can get better attack performance by stealing encoders rather than classifiers. This gap becomes especially apparent when the adversary has absolutely no knowledge of the train data. This is because the rich information in embeddings can better facilitate the learning process of surrogate encoders. For instance, when the surrogate dataset is CIFAR10 (the same as the target downstream dataset), stealing SimCLR's embeddings can achieve 0.785 agreement, while stealing predicted labels can achieve 0.712 agreement. However, when the surrogate dataset is totally different from the downstream target dataset, e.g., SVHN, stealing embeddings from SimCLR can still achieve 0.507 agreement while the agreement of stealing predicted labels drops to 0.192. We show more results in Supplementary Material Section A.4 due to the page limitation.", + "bbox": [ + 75, + 431, + 468, + 748 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We also find that all model stealing attacks' accuracy and agreement are highly correlated. As shown in Figure 5, the agreement is highly correlated with the accuracy. This indicates that besides accuracy, the agreement can also be used as a metric to evaluate the performance of model stealing attacks. We show the result on Figure 5. It can be obviously seen that agreement is highly related to accuracy. We use the linear regression method to describe the relationship between agreement and accuracy and find that the relation function is $y = 0.940 * x$ .", + "bbox": [ + 75, + 750, + 470, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.3. Performance of Cont-Steal", + "text_level": 1, + "bbox": [ + 500, + 90, + 741, + 104 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As shown in Section 4.2, encoders are more vulnerable to model stealing attacks since the embedding usually contains richer information compared to the predicted label or posteriors. We then show that our proposed Cont-Steal can achieve better attack performance by making deeper use of embeddings' information.", + "bbox": [ + 496, + 114, + 893, + 204 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Figure 7 shows the attack performance when the target pre-training dataset is CIFAR10. Note that we also show the attack performance on other settings in the appendix. We discover that compared to conventional attacks against encoders, Cont-Steal can consistently achieve better performance. For instance, as shown in Figure 7d, when the target encoder is MoCo trained on CIFAR10, if the adversary uses STL10 to conduct model stealing attacks against encoders, the surrogate encoder can achieve 0.841 agreement in CIFAR10 downstream tasks with the Cont-Steal but only 0.479 with conventional attacks. Another finding is that compared to the same distribution surrogate dataset, our Cont-Steal can better enhance the performance when the surrogate dataset comes from a different distribution from the pre-trained dataset. For instance, when the target encoder is SimCLR trained on CIFAR10, Cont-Steal outperforms conventional attack by 0.055 agreement when the surrogate dataset is also CIFAR10, while the improvement increases to 0.207 and 0.214 when the surrogate dataset is STL10. We show more comparing results in Supplementary Material Section A.5. Note that our Cont-Steal also has great performance on other recent state-of-the-art visual models (ViT [13], MAE [19], and CLIP [39]), as we show in Section A.6, and can have better performance than other recent similar attacks [14, 31] shown in Section A.7.", + "bbox": [ + 496, + 205, + 892, + 582 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To better understand why Cont-Steal can always achieve better performance, we extract samples' embeddings generated by different encoders, i.e., the target encoder, surrogate encoder trained with the conventional attack, and surrogate encoder trained with the Cont-Steal, and project them into a 2-dimensional space using t-SNE. From the results summarized in Figure 6, we find that Cont-Steal can effectively mimic the pattern of the embeddings as the target encoder. However, the conventional attack fails to capture such patterns for a number of input samples, e.g., the outer circle in Figure 8c. This further demonstrates that Cont-Steal benefits from jointly considering different embeddings as they can serve as anchors to better locate the position of the other embeddings in their space. We also show some ablation study results on Supplementary Material Section A.2 to show that with the less surrogate dataset, less training epoch, and different model architecture, Cont-Steal can still achieve much better results than conventional steal. Also, we show further attacks based on the stole models on Supplementary Material Section A.3 to show that Cont-Steal can be used as a springboard for other attacks.", + "bbox": [ + 496, + 583, + 893, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "16377", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/7ca75274d1255e5a97d83fda6695ddbb645d584232f66126490b614d719f6ff4.jpg", + "image_caption": [ + "(a) CIFAR10" + ], + "image_footnote": [], + "bbox": [ + 135, + 88, + 305, + 186 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/3acc1371b453f6896ef3db12321196b9ba5addf89f3678e5148745bf8016db6d.jpg", + "image_caption": [ + "(b) STL10" + ], + "image_footnote": [], + "bbox": [ + 310, + 89, + 480, + 186 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/267e0b000c3a365c69ef675872af6ba0fea1e208686e7cfba7bb92933609cfa4.jpg", + "image_caption": [ + "(c) F-MNIST" + ], + "image_footnote": [], + "bbox": [ + 485, + 90, + 656, + 186 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/463bd51baf3228ac7f90595a7c4dd970088c74976ae63db34d790c86c07cfc9b.jpg", + "image_caption": [ + "(d) SVHN" + ], + "image_footnote": [], + "bbox": [ + 658, + 90, + 830, + 186 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/0df6cd4d210bf03d91d88af64c3b39213f8eea0c76feb7ad4ea94a44ebaf9ec0.jpg", + "image_caption": [ + "(e) CIFAR10" + ], + "image_footnote": [], + "bbox": [ + 135, + 205, + 305, + 301 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/c1cce1e7eaa711578e2ba8ee28d4231a3763ee7220dca3ede6f97bb3ba896016.jpg", + "image_caption": [ + "(f) STL10" + ], + "image_footnote": [], + "bbox": [ + 310, + 205, + 482, + 301 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/22dc3e19a6e284b471bffedb4704a1a7a3d41a0ac3c33a46a37f572e7acc5c40.jpg", + "image_caption": [ + "(g) F-MNIST" + ], + "image_footnote": [], + "bbox": [ + 486, + 205, + 656, + 301 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/5e8a4c3ece83db9a5eac59c9ab04268d4eb68b804a1d748c96459c978291c115.jpg", + "image_caption": [ + "(h) SVHN" + ], + "image_footnote": [], + "bbox": [ + 658, + 205, + 831, + 301 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/43b3f225b8da1930bb35212ee3efe3dd52b6137dc188db8b02133b94236864a9.jpg", + "image_caption": [ + "Figure 4. The performance of model stealing attack against target encoders and downstream classifiers both trained on CIFAR10. Target models can output predicted labels, posteriors, or embeddings. The adversary uses CIFAR10, STL10, Fashion-MNIST (F-MNIST), and SVHN to conduct model stealing attacks. The x-axis represents different kinds of target models. The first line's y-axis represents the agreement of the model stealing attack. The second line's y-axis represents the accuracy of the model stealing attack.", + "Figure 5. The relationship between accuracy and agreement. The x-axis is the agreement number, and the y-axis is the accuracy number." + ], + "image_footnote": [], + "bbox": [ + 109, + 416, + 437, + 604 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/fa618ec1342d2131ad747c287860fbedac41a52941a8f65e478920d3f2711474.jpg", + "image_caption": [ + "(a) Target encoder", + "Figure 6. The t-SNE projection of 1,000 randomly selected samples' embeddings from target encoder, surrogate encoder under Cont-Steal, and surrogate encoder under the conventional attack, respectively. Note that the target encoder is pre-trained by SimCLR on CIFAR10." + ], + "image_footnote": [], + "bbox": [ + 109, + 696, + 200, + 750 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/ce6d0ad655b75f8a0b188891b8ceb1d396043bf17d98bf60b9754f5a57f1499f.jpg", + "image_caption": [ + "(b) Cont-Steal" + ], + "image_footnote": [], + "bbox": [ + 230, + 696, + 318, + 750 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/09513a45e819a8fd6f954414b2106994d88fa140052013dcfc2aeb979f8ac24c.jpg", + "image_caption": [ + "(c) Conventional" + ], + "image_footnote": [], + "bbox": [ + 349, + 696, + 437, + 750 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/3790f71a65d7793137ad41010dd389be416a3747414c9ca3c337f2f828980146.jpg", + "table_caption": [ + "Table 1. The monetary and (training) time costs for normal training and Cont-Steal attack. Cont-Steal's monetary cost contains two parts: query cost and training cost. Note that we ignore the query time cost of Cont-Steal as it normally has a smaller value than the training time cost." + ], + "table_footnote": [], + "table_body": "
Monetary CostTime Cost
ModelNormal ($)Cont-Steal ($)Normal (h)Cont-Steal (h)
SimCLR58.6811.83 (1.83 + 10)20.010.62
MoCo54.8312.13 (2.13 + 10)18.690.73
BYOL61.4612.08 (2.08 + 10)20.960.71
SimSiam57.1412.00 (2.00 + 10)19.460.68
", + "bbox": [ + 509, + 492, + 880, + 588 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.4. Cost Analysis", + "text_level": 1, + "bbox": [ + 500, + 612, + 640, + 628 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "As we mentioned before, pre-train a state-of-the-art encoder is time-consuming and resource-demanding. We wonder if the model stealing attacks can steal the functionality of the encoder with much less cost. To this end, we evaluate the time and monetary cost of training an encoder from scratch or stealing a pre-trained encoder via ContSteal. The monetary cost of model stealing includes querying the target model and training the surrogate model. We refer to the query price as $1 for 1,000 queries based on AWS.6 Our experiment is conducted on 1 NVIDIA A100 whose price is$ 2.934 per hour based on google cloud.7", + "bbox": [ + 496, + 636, + 890, + 801 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The monetary and time cost is shown in Table 1. We observe that Cont-Steal can obtain a surrogate encoder with much less money and time cost than training the encoder from scratch. For instance, a ResNet18 trained by SimCLR", + "bbox": [ + 496, + 801, + 890, + 863 + ], + "page_idx": 5 + }, + { + "type": "page_footnote", + "text": "6https://aws.amazon.com/rekognition/pricing/", + "bbox": [ + 514, + 875, + 861, + 887 + ], + "page_idx": 5 + }, + { + "type": "page_footnote", + "text": "7https://cloud.google.com/compute/gpus-pricing/", + "bbox": [ + 519, + 888, + 887, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "16378", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/50bd10a636c3765a64687310ebe792f6a0595a0688a8683cb76ab2948a712c2c.jpg", + "image_caption": [ + "(a) CIFAR10" + ], + "image_footnote": [], + "bbox": [ + 135, + 88, + 305, + 186 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/7291b6de743fed6c4c3e879a0b1b9dd47616b89fb13aea5bb56545ce7bd0c926.jpg", + "image_caption": [ + "(b) STL10" + ], + "image_footnote": [], + "bbox": [ + 310, + 89, + 482, + 186 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/961afd3798f5ba4a144d9541282cff8b14ca0de8859b78790ad87c5977a8b80f.jpg", + "image_caption": [ + "(c) F-MNIST" + ], + "image_footnote": [], + "bbox": [ + 486, + 90, + 656, + 186 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/02426e26ce9d61527cae74aef17b4b1edca4a13464e0f4ed7e49c887ac2e037e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 658, + 90, + 830, + 186 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a7f645d73ca807d125679a62b078825367274826362cdafcdf9bfc9e8d61dc65.jpg", + "image_caption": [ + "(e) CIFAR10" + ], + "image_footnote": [], + "bbox": [ + 135, + 205, + 307, + 301 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/c796704c9354237234f9d12bdb37fc6b30d18e6feaaffc07e637417257d738bd.jpg", + "image_caption": [ + "(f) STL10" + ], + "image_footnote": [], + "bbox": [ + 310, + 205, + 482, + 301 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/c1af7f3f0ddd109c57a48bd35c17caaffd93df608c6633e0aa61bfe9e802aacd.jpg", + "image_caption": [ + "(g) F-MNIST" + ], + "image_footnote": [], + "bbox": [ + 486, + 205, + 656, + 301 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/fae52eea7c07e6a312dfd9279f7fd05195a2960cad8fb6e264c34d499d823721.jpg", + "image_caption": [ + "(d) SVHN", + "(h) SVHN" + ], + "image_footnote": [], + "bbox": [ + 658, + 205, + 830, + 301 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f470e76d0b9bd37680a0c9f53f2d7fa1faa050910bbe4869d6609abc624d6ae7.jpg", + "image_caption": [ + "Figure 7. The performance of Cont-Steal and conventional attack against target encoders trained on CIFAR10. The adversary uses CIFAR10, STL10, F-MNIST, and SVHN to conduct model stealing attacks. The adversary uses CIFAR10 as the downstream task to evaluate the attack performance. The x-axis represents different kinds of the target model. The first line's y-axis represents the agreement of the model stealing attack. The second line's y-axis represents the accuracy of the model stealing attack.", + "(a) Adding noise" + ], + "image_footnote": [], + "bbox": [ + 83, + 411, + 207, + 481 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/8b16877a46b1522e38dcade88684c045d8ef410010ca0bae3058eb925722b79c.jpg", + "image_caption": [ + "(b) Top- $k$", + "Figure 8. The performance of different defense methods. Target encoders are trained on CIFAR10. The downstream dataset and surrogate dataset are both STL10. The x-axis represents different defense levels. The y-axis represents the model's accuracy." + ], + "image_footnote": [], + "bbox": [ + 210, + 411, + 334, + 482 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/26ef07a82ee8c537eced95175b4f4d4475abb0193e362340d7032c969dd634fb.jpg", + "image_caption": [ + "(c) Rounding" + ], + "image_footnote": [], + "bbox": [ + 339, + 411, + 462, + 482 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "on CIFAR10 takes 20.01 hours and $58.68$ 's on 1 NVIDIA A100 GPU, while Cont-Steal only takes 0.62 hours and $11.83$ 's to steal an encoder that performs similarly on downstream tasks. The results demonstrate that Cont-Steal is able to construct surrogate encoders that perform similarly to the target encoders but with much less time and monetary cost.", + "bbox": [ + 75, + 594, + 468, + 686 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.5. Defenses", + "text_level": 1, + "bbox": [ + 76, + 695, + 181, + 709 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In this section, we will consider different defenses against model stealing attacks on encoders to evaluate the robustness of our proposed attack. We divided all defenses into three categories: perturbation-based defense [37] and watermark-based defense [2].", + "bbox": [ + 75, + 719, + 468, + 794 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Perturbation-based Defense. In this defense setting, the defender aims to perturb the output of the target model to limit the information the adversary can obtain. The common practice of this kind of defense includes adding noise [37], top- $k$ [37], and feature rounding [48].", + "bbox": [ + 75, + 794, + 468, + 869 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Adding noise means that the defender will introduce noise value to the original output of the model. In our case,", + "bbox": [ + 76, + 869, + 468, + 900 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "we consider adding Gaussian noise to the embeddings generated by the target encoder. We set the mean value to 0, and different noise levels represent different standard deviations of the Gaussian distribution. For Top- $k$ , the defender will only output the first $k$ largest number of each embedding (and set the rest as 0). In this way, the high-dimensional information of the image contained in embeddings can be appropriately reduced. Regarding feature rounding, the defender will truncate the values in the embedding to a specific digit. As a case study, we consider a ResNet18 encoder pre-trained on CIFAR10 with SimCLR and take STL10 to train its downstream classifier. The experimental results are summarized in Figure 8. We can observe that while adding noise and top- $k$ can reduce the model stealing attacks' performance, it may also degrade the target model performance to a large extent. For instance, when the noise increases from 0 to 10, the attack performance of ContSteal decreases from 0.729 to 0.410, while the target encoder's performance drops from 0.734 to 0.098. On the other hand, rounding only has a limited effect on both target model performance and attack performance. This indicates that perturbation-based defense cannot defend against the encoder's model stealing attack effectively since they cannot reach a good trade-off between attack performance and model utility.", + "bbox": [ + 496, + 414, + 892, + 792 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Watermark-based Defense. Watermark-based defense is also one of the most popular defense methods against model stealing attacks [2]. Watermark provides copyright protection by adding some specific identification to the target model. If the surrogate model is stolen from the watermarked target model, then ideally, it will contain the same watermark as well. Adi et al. [2] show that back", + "bbox": [ + 496, + 794, + 892, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "16379", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/5cc0eb05ceb5bd2b5e6307c739477dda2f74284d9912d0379f683db96a161765.jpg", + "table_caption": [ + "Table 2. Watermark defense. Pretrain dataset and surrogate dataset are both CIFAR10. Watermark leverages a watermark rate (wr) to verify the ownership of target models. A higher wr denotes a better verification performance." + ], + "table_footnote": [], + "table_body": "
DatasetTarget model (acc/wr)Cont-Steal (acc/wr)Baseline (acc/wr)
CIFAR100.864 / 0.9980.769 / 0.1300.871 / 0.095
STL100.721 / 0.9990.702 / 0.0340.733 / 0.111
SVHN0.501 / 0.9990.535 / 0.3030.492 / 0.103
F-MNIST0.857 / 0.9990.813 / 0.0610.850 / 0.099
", + "bbox": [ + 78, + 155, + 470, + 234 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "door technology can be used as the watermark to protect the model. In that sense, BadEncoder [26], a backdoor mechanism against the encoder, can be leveraged as a watermarking technology for our target encoder as well. The defenders first will train the watermarked (backdoored) encoder, where images with a certain trigger will cause misclassification. Then, if they find the surrogate model can also misclassify images with the same trigger, the defenders can claim ownership of the surrogate model.", + "bbox": [ + 75, + 258, + 468, + 393 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In our experiments, we leverage BadEncoder to watermark the encoder pre-trained on CIFAR10 by SimCLR, and leverage different downstream datasets to perform different tasks. We assume a strong adversary that has the same downstream dataset as the surrogate dataset. Also, we consider the baseline cases where the trigger samples are fed into the clean model to calculate the watermark rate (wr). As shown in Table 2), the watermark cannot be preserved as the surrogate models constructed by Cont-Steal have similar wr as the baseline model. For instance, when the downstream dataset is CIFAR10, Cont-Steal builds a surrogate model with 0.769 accuracy while only 0.130 wr, which is close to the baseline model. This indicates that Cont-Steal can bypass the watermarking technique as it can reach similar utility while reducing the wr to a large extent. Note that there is also another work to protect contrastive learning models from model stealing attacks using dataset inference [15]. We show in Supplementary Material in Section A.8 that this kind of defense can be easily bypassed by Cont-Steal.", + "bbox": [ + 75, + 395, + 470, + 696 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Related Work", + "text_level": 1, + "bbox": [ + 76, + 708, + 218, + 724 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Contrastive Learning. Contrastive learning is one of the most popular methods to train encoders. Current works [6, 8, 17, 20, 49, 53] propose different advanced contrastive learning algorithms. SimCLR, MoCo, BYOL, SimSiam are currently the mainstream frameworks of contrastive learning. Thus, we concentrate on them in this paper. There are many works on evaluating the security and privacy risks of contrastive learning. Previous works [23, 26, 30] propose membership inference attacks, attribution inference attacks, and backdoor attacks on contrastive learning. All proposed attacks show that contrastive-based models are vulnerable", + "bbox": [ + 75, + 734, + 470, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "to popular attacks. Therefore, the security issues of self-supervised learning deserve more attention.", + "bbox": [ + 496, + 90, + 890, + 121 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Model Stealing Attack. In model stealing, the adversary's goal is to steal part of the target model. Tramér et al. [48] proposed the first model stealing attack against black-box machine learning API to steal its parameters. Wang et al. [50] proposed the first hyperparameter stealing attacks against ML models. Oh et al. [36] also tried to steal machine learning model's architectures and hyperparameters. Orekondy et al. [37] proposed knockoff nets, which aim at stealing the functionality of black-box models. Krishna et al. [28] formulated a model stealing attack against BERT-based API. Besides, Wu et al. [51] and Shen et al. [44] perform model stealing attacks against Graph Neural Networks. These works often have relatively strong assumptions, such as the model family is known and the victim's data is partly available while we conduct model stealing attacks against encoders and relax the above assumption.", + "bbox": [ + 496, + 127, + 893, + 369 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 402, + 617, + 417 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we conduct the first model stealing risk assessment towards image encoders. Our evaluation shows that the encoder is more vulnerable to model stealing attacks compared to the classifier. This is because the embedding provided by the encoder contains richer information than the posteriors or predicted labels from whole classifiers.", + "bbox": [ + 496, + 434, + 890, + 525 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To better unleash the power from the embeddings, we propose Cont-Steal, a contrastive learning-based model stealing method against encoders. Concretely, Cont-Steal introduces different types of negative pairs as \"anchors\" to better navigate the surrogate encoder and learn the functionality of the target encoder. Extensive evaluations show that Cont-Steal consistently performs better than conventional attacks against encoders. And such an advantage is further amplified when the adversary has no information on the target dataset, a limited amount of data, and restricted query budgets. Our work points out that the threat of model stealing attacks against encoders is largely underestimated, which prompts the need for more effective intellectual property protection of representation learning techniques.", + "bbox": [ + 496, + 532, + 890, + 743 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments. We thank all anonymous reviewers for their constructive comments. This work is partially funded by the Helmholtz Association within the project \"Trustworthy Federated Data Analytics\" (TFDA) (funding number ZT-I-OO1 4) and by the European Health and Digital Executive Agency (HADEA) within the project \"Understanding the individual host response against Hepatitis D Virus to develop a personalized approach for the management of hepatitis D\" (D-Solve) (grant agreement number 101057917).", + "bbox": [ + 496, + 750, + 893, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "16380", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 174, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] https://www.cs.toronto.edu/~kriz/cifar.html.", + "[2] Yossi Adi, Carsten Baum, Moustapha Cisse, Benny Pinkas, and Joseph Keshet. Turning Your Weakness Into a Strength: Watermarking Deep Neural Networks by Backdoorsing. In USENIX Security Symposium (USENIX Security), pages 1615-1631. USENIX, 2018.", + "[3] Battista Biggio, Igino Corona, Davide Maiorca, Blaine Nelson, Nedim Srndic, Pavel Laskov, Giorgio Giacinto, and Fabio Roli. Evasion Attacks against Machine Learning at Test Time. In European Conference on Machine Learning and Principles and Practice of Knowledge Discovery in Databases (ECML/PKDD), pages 387-402. Springer, 2013.", + "[4] Nicholas Carlini and David Wagner. Towards Evaluating the Robustness of Neural Networks. In IEEE Symposium on Security and Privacy (S&P), pages 39-57. IEEE, 2017.", + "[5] Varun Chandrasekaran, Kamalika Chaudhuri, Irene Giacomelli, Somesh Jha, and Songbai Yan. Model Extraction and Active Learning. CoRR abs/1811.02054, 2018.", + "[6] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey E. Hinton. A Simple Framework for Contrastive Learning of Visual Representations. In International Conference on Machine Learning (ICML), pages 1597-1607. PMLR, 2020.", + "[7] Xiaoyi Chen, Ahmed Salem, Michael Backes, Shiqing Ma, Qingni Shen, Zhonghai Wu, and Yang Zhang. BadNL: Backdoor Attacks Against NLP Models with Semantic-preserving Improvements. In Annual Computer Security Applications Conference (ACSAC), pages 554-569. ACSAC, 2021.", + "[8] Xinlei Chen and Kaiming He. Exploring Simple Siamese Representation Learning. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 15750-15758. IEEE, 2021.", + "[9] Adam Coates, Andrew Y. Ng, and Honglak Lee. An Analysis of Single-Layer Networks in Unsupervised Feature Learning. In International Conference on Artificial Intelligence and Statistics (AISTATS), pages 215-223. JMLR, 2011.", + "[10] Tianshuo Cong, Xinlei He, and Yang Zhang. SSL-Guard: A Watermarking Scheme for Self-supervised" + ], + "bbox": [ + 78, + 114, + 467, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Learning Pre-trained Encoders. In ACM SIGSAC Conference on Computer and Communications Security (CCS), pages 579-593. ACM, 2022.", + "[11] Ekin Dogus Cubuk, Barret Zoph, Jon Shlens, and Quoc Le. RandAugment: Practical Automated Data Augmentation with a Reduced Search Space. In Annual Conference on Neural Information Processing Systems (NeurIPS), pages 18613-18624. NeurIPS, 2020.", + "[12] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. ImageNet: A large-scale hierarchical image database. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 248–255. IEEE, 2009.", + "[13] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale. In International Conference on Learning Representations (ICLR), 2021.", + "[14] Adam Dziedzic, Nikita Dhawan, Muhammad Ahmad Kaleem, Jonas Guan, and Nicolas Papernot. On the Difficulty of Defending Self-Supervised Learning against Model Extraction. In International Conference on Machine Learning (ICML). JMLR, 2022.", + "[15] Adam Dziedzic, Haonan Duan, Muhammad Ahmad Kaleem, Nikita Dhawan, Jonas Guan, Yannis Cattan, Franziska Boenisch, and Nicolas Papernot. Dataset Inference for Self-Supervised Models. In Annual Conference on Neural Information Processing Systems (NeurIPS). NeurIPS, 2022.", + "[16] Ian Goodfellow, Jonathon Shlens, and Christian Szegedy. Explaining and Harnessing Adversarial Examples. In International Conference on Learning Representations (ICLR), 2015.", + "[17] Jean-Bastien Grill, Florian Strub, Florent Altché, Corentin Tallec, Pierre H. Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Ávila Pires, Zhaohan Guo, Mohammad Gheshlaghi Azar, Bilal Piot, Koray Kavukcuoglu, Rémi Munos, and Michal Valko. Bootstrap Your Own Latent - A New Approach to Self-Supervised Learning. In Annual Conference on Neural Information Processing Systems (NeurIPS). NeurIPS, 2020.", + "[18] Yunhui Guo, Honghui Shi, Abhishek Kumar, Kristen Grauman, Tajana Rosing, and Rogério Schmidt" + ], + "bbox": [ + 501, + 90, + 890, + 901 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "16381", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Feris. SpotTune: Transfer Learning Through Adaptive Fine-Tuning. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 4805-4814. IEEE, 2019.", + "[19] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dólár, and Ross B. Girshick. Masked Autoencoders Are Scalable Vision Learners. CoRR abs/2111.06377, 2021.", + "[20] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross B. Girshick. Momentum Contrast for Unsupervised Visual Representation Learning. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 9726-9735. IEEE, 2020.", + "[21] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep Residual Learning for Image Recognition. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 770-778. IEEE, 2016.", + "[22] Xinlei He, Rui Wen, Yixin Wu, Michael Backes, Yun Shen, and Yang Zhang. Node-Level Membership Inference Attacks Against Graph Neural Networks. CoRR abs/2102.05429, 2021.", + "[23] Xinlei He and Yang Zhang. Quantifying and Mitigating Privacy Risks of Contrastive Learning. In ACM SIGSAC Conference on Computer and Communications Security (CCS), pages 845-863. ACM, 2021.", + "[24] Bo Hui, Yuchen Yang, Haolin Yuan, Philippe Burlina, Neil Zhenqiang Gong, and Yinzhi Cao. Practical Blind Membership Inference Attack via Differential Comparisons. In Network and Distributed System Security Symposium (NDSS). Internet Society, 2021.", + "[25] Matthew Jagielski, Nicholas Carlini, David Berthelot, Alex Kurakin, and Nicolas Papernot. High Accuracy and High Fidelity Extraction of Neural Networks. In USENIX Security Symposium (USENIX Security), pages 1345-1362. USENIX, 2020.", + "[26] Jinyuan Jia, Yupei Liu, and Neil Zhenqiang Gong. BadEncoder: Backdoor Attacks to Pre-trained Encoders in Self-Supervised Learning. In IEEE Symposium on Security and Privacy (S&P). IEEE, 2022.", + "[27] Prannay Khosla, Piotr Teterwak, Chen Wang, Aaron Sarna, Yonglong Tian, Phillip Isola, Aaron Maschinot, Ce Liu, and Dilip Krishnan. Supervised Contrastive Learning. In Annual Conference on Neural Information Processing Systems (NeurIPS). NeurIPS, 2020.", + "[28] Kalpesh Krishna, Gaurav Singh Tomar, Ankur P. Parikh, Nicolas Papernot, and Mohit Iyyer. Thieves on Sesame Street! Model Extraction of BERT-based" + ], + "bbox": [ + 78, + 90, + 468, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "APIs. In International Conference on Learning Representations (ICLR), 2020.", + "[29] Alexey Kurakin, Ian Goodfellow, and Samy Bengio. Adversarial Examples in the Physical World. CoRR abs/1607.02533, 2016.", + "[30] Hongbin Liu, Jinyuan Jia, Wenjie Qu, and Neil Zhenqiang Gong. EncoderMI: Membership Inference against Pre-trained Encoders in Contrastive Learning. In ACM SIGSAC Conference on Computer and Communications Security (CCS). ACM, 2021.", + "[31] Yupei Liu, Jinyuan Jia, Hongbin Liu, and Neil Zhenqiang Gong. StolenEncoder: Stealing Pre-trained Encoders in Self-supervised Learning. In ACM SIGSAC Conference on Computer and Communications Security (CCS), pages 2115-212. ACM, 2022.", + "[32] Aleksander Madry, Aleksandar Makelov, Ludwig Schmidt, Dimitris Tsipras, and Adrian Vladu. Towards Deep Learning Models Resistant to Adversarial Attacks. In International Conference on Learning Representations (ICLR), 2018.", + "[33] Milad Nasr, Reza Shokri, and Amir Houmansadr. Machine Learning with Membership Privacy using Adversarial Regularization. In ACM SIGSAC Conference on Computer and Communications Security (CCS), pages 634-646. ACM, 2018.", + "[34] Milad Nasr, Reza Shokri, and Amir Houmansadr. Comprehensive Privacy Analysis of Deep Learning: Passive and Active White-box Inference Attacks against Centralized and Federated Learning. In IEEE Symposium on Security and Privacy (S&P), pages 1021-1035. IEEE, 2019.", + "[35] Yuval Netzer, Tao Wang, Adam Coates, Alessandro Bissacco, Bo Wu, and Andrew Y. Ng. Reading Digits in Natural Images with Unsupervised Feature Learning. In Annual Conference on Neural Information Processing Systems (NIPS). NIPS, 2011.", + "[36] Seong Joon Oh, Max Augustin, Bernt Schiele, and Mario Fritz. Towards Reverse-Engineering Black-Box Neural Networks. In International Conference on Learning Representations (ICLR), 2018.", + "[37] Tribhuvanesh Orekondy, Bernt Schiele, and Mario Fritz. Knockoff Nets: Stealing Functionality of Black-Box Models. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 4954–4963. IEEE, 2019.", + "[38] Nicolas Papernot, Patrick D. McDaniel, Somesh Jha, Matt Fredrikson, Z. Berkay Celik, and Ananthram" + ], + "bbox": [ + 501, + 90, + 890, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "16382", + "bbox": [ + 480, + 945, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Swami. The Limitations of Deep Learning in Adversarial Settings. In IEEE European Symposium on Security and Privacy (Euro S&P), pages 372-387. IEEE, 2016.", + "[39] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning Transferable Visual Models From Natural Language Supervision. In International Conference on Machine Learning (ICML), pages 8748-8763. PMLR, 2021.", + "[40] Alexandre Sablayrolles, Matthijs Douze, Cordelia Schmid, Yann Ollivier, and Hervé Jégou. White-box vs Black-box: Bayes Optimal Strategies for Membership Inference. In International Conference on Machine Learning (ICML), pages 5558-5567. PMLR, 2019.", + "[41] Ahmed Salem, Michael Backes, and Yang Zhang. Don't Trigger Me! A Triggerless Backdoor Attack Against Deep Neural Networks. CoRR abs/2010.03282, 2020.", + "[42] Ahmed Salem, Yang Zhang, Mathias Humbert, Pascal Berrang, Mario Fritz, and Michael Backes. MLLeaks: Model and Data Independent Membership Inference Attacks and Defenses on Machine Learning Models. In Network and Distributed System Security Symposium (NDSS). Internet Society, 2019.", + "[43] Zeyang Sha, Xinlei He, Pascal Berrang, Mathias Humbert, and Yang Zhang. Fine-Tuning Is All You Need to Mitigate Backdoor Attacks. CoRR abs/2212.09067, 2022.", + "[44] Yun Shen, Xinlei He, Yufei Han, and Yang Zhang. Model Stealing Attacks Against Inductive Graph Neural Networks. In IEEE Symposium on Security and Privacy (S&P), pages 1175-1192. IEEE, 2022.", + "[45] Reza Shokri, Marco Stronati, Congzheng Song, and Vitaly Shmatikov. Membership Inference Attacks Against Machine Learning Models. In IEEE Symposium on Security and Privacy (S&P), pages 3-18. IEEE, 2017.", + "[46] Liwei Song and Prateek Mittal. Systematic Evaluation of Privacy Risks of Machine Learning Models. In USENIX Security Symposium (USENIX Security). USENIX, 2021.", + "[47] Florian Tramér, Alexey Kurakin, Nicolas Papernot, Ian Goodfellow, Dan Boneh, and Patrick McDaniel." + ], + "bbox": [ + 78, + 90, + 468, + 898 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Ensemble Adversarial Training: Attacks and Defenses. In International Conference on Learning Representations (ICLR), 2017.", + "[48] Florian Tramér, Fan Zhang, Ari Juels, Michael K. Reiter, and Thomas Ristenpart. Stealing Machine Learning Models via Prediction APIs. In USENIX Security Symposium (USENIX Security), pages 601-618. USENIX, 2016.", + "[49] Aäron van den Oord, Yazhe Li, and Oriol Vinyals. Representation Learning with Contrastive Predictive Coding. CoRR abs/1807.03748, 2018.", + "[50] Binghui Wang and Neil Zhenqiang Gong. Stealing Hyperparameters in Machine Learning. In IEEE Symposium on Security and Privacy (S&P), pages 36-52. IEEE, 2018.", + "[51] Bang Wu, Xiangwen Yang, Shirui Pan, and Xingliang Yuan. Model Extraction Attacks on Graph Neural Networks: Taxonomy and Realization. CoRR abs/2010.12751, 2020.", + "[52] Huijun Wu, Chen Wang, Yuriy Tyshetskiy, Andrew Docherty, Kai Lu, and Liming Zhu. Adversarial Examples for Graph Data: Deep Insights into Attack and Defense. In International Joint Conferences on Artificial Intelligence (IJCAI), pages 4816-4823. IJCAI, 2019.", + "[53] Zhirong Wu, Yuanjun Xiong, Stella X. Yu, and Dahua Lin. Unsupervised Feature Learning via Non-Parametric Instance Discrimination. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 3733–3742. IEEE, 2018.", + "[54] Han Xiao, Kashif Rasul, and Roland Vollgraf. Fashion-MNIST: a Novel Image Dataset for Benchmarking Machine Learning Algorithms. CoRR abs/1708.07747, 2017.", + "[55] Ziqing Yang, Zeyang Sha, Michael Backes, and Yang Zhang. From Visual Prompt Learning to Zero-Shot Transfer: Mapping Is All You Need. CoRR abs/2303.05266, 2023." + ], + "bbox": [ + 501, + 90, + 890, + 744 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "16383", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2023/Can't Steal_ Cont-Steal! Contrastive Stealing Attacks Against Image Encoders/d755a24e-203a-45cb-a84e-020f7fb471b0_model.json b/2023/Can't Steal_ Cont-Steal! Contrastive Stealing Attacks Against Image Encoders/d755a24e-203a-45cb-a84e-020f7fb471b0_model.json new file mode 100644 index 0000000000000000000000000000000000000000..b44d5d871971dc5d471df7676c9b189826a9243f --- /dev/null +++ b/2023/Can't Steal_ Cont-Steal! Contrastive Stealing Attacks Against Image Encoders/d755a24e-203a-45cb-a84e-020f7fb471b0_model.json @@ -0,0 +1,2620 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.214, + 0.131, + 0.758, + 0.176 + ], + "angle": 0, + "content": "Can't Steal? Cont-Steal! Contrastive Stealing Attacks Against Image Encoders" + }, + { + "type": "text", + "bbox": [ + 0.201, + 0.203, + 0.756, + 0.223 + ], + "angle": 0, + "content": "Zeyang Sha† Xinlei He† Ning Yu‡ Michael Backes† Yang Zhang†" + }, + { + "type": "text", + "bbox": [ + 0.16, + 0.239, + 0.806, + 0.275 + ], + "angle": 0, + "content": "†CISPA Helmholtz Center for Information Security ‡Salesforce Research {zeyang.sha, xinlei.he, director, zhang}@cispa.de, ning.yu@salesforce.com" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.309, + 0.314, + 0.326 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.341, + 0.473, + 0.538 + ], + "angle": 0, + "content": "Self-supervised representation learning techniques have been developing rapidly to make full use of unlabeled images. They encode images into rich features that are oblivious to downstream tasks. Behind their revolutionary representation power, the requirements for dedicated model designs and a massive amount of computation resources expose image encoders to the risks of potential model stealing attacks - a cheap way to mimic the well-trained encoder performance while circumventing the demanding requirements. Yet conventional attacks only target supervised classifiers given their predicted labels and/or posteriors, which leaves the vulnerability of unsupervised encoders unexplored." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.538, + 0.471, + 0.691 + ], + "angle": 0, + "content": "In this paper, we first instantiate the conventional stealing attacks against encoders and demonstrate their severer vulnerability compared with downstream classifiers. To better leverage the rich representation of encoders, we further propose Cont-Steal, a contrastive-learning-based attack, and validate its improved stealing effectiveness in various experiment settings. As a takeaway, we appeal to our community's attention to the intellectual property protection of representation learning techniques, especially to the defenses against encoder stealing attacks like ours." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.718, + 0.21, + 0.735 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.744, + 0.47, + 0.82 + ], + "angle": 0, + "content": "Recent years have witnessed the great success of applying deep learning (DL) to computer vision tasks. Different from supervised DL models, self-supervised learning which transforms unlabeled data samples into rich representations, has gained more and more popularity." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.82, + 0.47, + 0.866 + ], + "angle": 0, + "content": "Behind its powerful representation, it is non-trivial to obtain a state-of-the-art image encoder. For instance, SimCLR [6] uses 128 TPU v3 cores to pre-train a ResNet-" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.312, + 0.871, + 0.507 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.522, + 0.895, + 0.62 + ], + "angle": 0, + "content": "Figure 1. Model stealing attacks against classifiers (previous) v.s. model stealing attacks against encoders (ours). Previous works aim to steal a whole classifier using the predicted label or posteriors of a target model. In our work, we aim to steal the target encoder using its embeddings. The target encoder \\((E_{t})\\) is pre-trained and fixed, as shown in the solid frame. The surrogate encoder \\((E_{s})\\) is trainable by the adversary, as shown in the dashed frame." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.646, + 0.892, + 0.752 + ], + "angle": 0, + "content": "50 encoder with a batch size of 4096. Therefore, many big companies provide cloud-based self-supervised learning encoder services for users. For instance, Cohere,\\(^{2}\\) OpenAI,\\(^{3}\\) and Clarifai\\(^{4}\\) provide the embedding API of images and texts for commercial usage. There are many works [10, 26, 30] exploring security issues of encoder-based API. Therefore, it is a very important and urgent problem." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.753, + 0.894, + 0.829 + ], + "angle": 0, + "content": "These kinds of service leave the possibility of model stealing attacks [5, 25, 28, 37, 44, 48, 50, 51]. In these attacks, the adversary aims to steal the parameters or functionalities of target models with only query access to them. A successful model stealing attack does not only threaten" + }, + { + "type": "page_footnote", + "bbox": [ + 0.078, + 0.875, + 0.468, + 0.9 + ], + "angle": 0, + "content": "See our code in https://github.com/zeyangsha/Cont-Steal." + }, + { + "type": "page_footnote", + "bbox": [ + 0.5, + 0.838, + 0.892, + 0.901 + ], + "angle": 0, + "content": "\\(^{2}\\)https://cohere.ai/ \n\\(^{3}\\)https://beta.openai.com/docs/api-reference/embeddings/ \n\\(^{4}\\)https://www clarifai.com/models/general-image-embedding/" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16373" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.226 + ], + "angle": 0, + "content": "the intellectual property of the target model, but also serves as a stepping stone for further attacks such as adversarial examples [3, 4, 16, 38, 47, 52], backdoor attacks [7, 26, 41, 43], and membership inference attacks [22-24, 30, 33, 34, 40, 42, 45, 46]. So far, model stealing attacks concentrate on the supervised classifiers, i.e., the model responses are prediction posteriors or labels for a specific downstream task. The vulnerability of unsupervised image encoders is unfortunately unexplored." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.229, + 0.47, + 0.365 + ], + "angle": 0, + "content": "Our Work. To fill this gap, we pioneer the systematic investigation of model stealing attacks against image encoders. In this work, the adversary's goal is to steal the functionalities of the target model. See Figure 1 for an overview and a comparison with previous works. More specifically, we focus on encoders trained by contrastive learning, which is one of the most cutting-edge unsupervised representation learning strategies that unleash the information of unlabeled data." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.367, + 0.47, + 0.548 + ], + "angle": 0, + "content": "We first instantiate the conventional stealing attacks against encoders and expose their vulnerability. Given an input image, the target encoder outputs its representation (referred to as embedding). Similar to model stealing attacks against classifiers, we consider the embedding as the \"ground truth\" label to guide the training procedure of a surrogate encoder on the adversary side. To measure the effectiveness of stealing attacks, we train an extra linear layer for the target and surrogate encoders towards the same downstream classification task. Preferably, the surrogate model should achieve both high classification accuracy and high agreement with the target predictions." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.55, + 0.47, + 0.702 + ], + "angle": 0, + "content": "We evaluate our attacks on five datasets against four contrastive learning encoders. Our results demonstrate that the conventional attacks are more effective against encoders than against downstream classifiers. For instance, when we steal the downstream classifier pre-trained by SimCLR on CIFAR10 (with posteriors as its responses) using STL10 as the surrogate dataset, the adversary can only achieve an accuracy of 0.359. The accuracy, however, increases to 0.500 instead when we steal its encoder (with the embedding as its response)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.703, + 0.47, + 0.853 + ], + "angle": 0, + "content": "Despite its encouraging performance, conventional attacks are not the most suitable ones against encoders. This is because they treat each image-embedding pair individually without interacting across pairs. Different embeddings are beneficial to each other as they can serve as anchors to better locate the position of the other embeddings in their space. Contrastive learning [6, 8, 17, 20, 27, 49, 53] is a straightforward idea to achieve this goal. It is formulated to enforce the embeddings of different augmentations of the same images closer and those of different images further." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.856, + 0.47, + 0.901 + ], + "angle": 0, + "content": "In a similar spirit, we propose Cont-Steal, a contrastive-learning-based model stealing attack against the encoder. The goal of Cont-Steal is to enforce the surrogate embed" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.152 + ], + "angle": 0, + "content": "ding of an image close to its target embedding (defined as a positive pair) and also push away embeddings of different images irrespective of being generated by the target or the surrogate encoders (defined as negative pairs)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.153, + 0.893, + 0.409 + ], + "angle": 0, + "content": "The comprehensive evaluation shows that Cont-Steal outperforms the conventional model stealing attacks to a large extent. For instance, when CIFAR10 is the target dataset, Cont-Steal achieves an accuracy of 0.714 on the SimCLR encoder pretrained on CIFAR10 with the surrogate dataset and downstream dataset being STL10, while the conventional attack only achieves 0.457 accuracy. Also, Cont-Steal is more query-efficient and dataset-independent (see Figure 9 for more details). This is because Cont-Steal leverages higher-order information across samples to mimic the functionality of the target encoder. To mitigate the attacks, we evaluate different defense mechanisms including noise, top-\\(k\\), rounding, and watermark. Our evaluations show that in most cases, these mechanisms cannot effectively defend against Cont-Steal. Among them, top-\\(k\\) can reduce the attack performance to the largest extent. However, it also strongly limits the target model's utility." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.409, + 0.892, + 0.485 + ], + "angle": 0, + "content": "As a takeaway, our attack further exposes the severe vulnerability of pre-trained encoders. We appeal to our community's attention to the intellectual property protection of representation learning techniques, especially to the defenses against encoder stealing attacks like ours." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.498, + 0.642, + 0.514 + ], + "angle": 0, + "content": "2. Threat Model" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.524, + 0.892, + 0.629 + ], + "angle": 0, + "content": "In this work, for the encoder pre-trained with images, we consider image classification as the downstream task. We refer to the encoder as the target encoder. Then we treat both the encoder and the linear layer trained for the downstream task together as the target model. We first introduce the adversary's goal and then characterize different background knowledge that the adversary might have." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.63, + 0.892, + 0.78 + ], + "angle": 0, + "content": "Adversary's Goal. Following previous work [25, 28, 44], we taxonomize the adversary's goal into two dimensions, i.e., theft and utility. The theft adversary aims to build a surrogate encoder that has similar performance on the downstream tasks as the target encoder. Different from the thief adversary, the goal of the utility adversary is to construct a surrogate encoder that behaves normally on different downstream tasks. In this case, the surrogate encoder not only faithfully \"copies\" the behaviors of the target encoder, but also serves as a stepping stone to conduct other attacks." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.78, + 0.892, + 0.84 + ], + "angle": 0, + "content": "Adversary's Background Knowledge. We categorize the adversary's background knowledge into two dimensions, i.e., the knowledge of the target encoder and the distribution of the surrogate dataset." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Regarding knowledge of the target encoder, we assume that the adversary only has black-box access to it, which means that they can only query the target encoder with an input image and obtain the corresponding output, i.e., the" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "16374" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.282, + 0.107 + ], + "angle": 0, + "content": "embedding of the input image." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.107, + 0.47, + 0.302 + ], + "angle": 0, + "content": "Regarding the surrogate dataset that is used to train the surrogate encoder, we consider two cases. First, we assume the adversary has the same training dataset as the target encoder. However, such an assumption may be hard to achieve as such datasets are usually private and protected by the model owner. In a more extreme case, we assume that the adversary has totally no information about the target encoder's training dataset, which means that they can only use a different distribution dataset to conduct the model stealing attacks. We later show that the adversary can still launch effective model stealing attacks against the target encoder given a surrogate dataset that is distributed differently compared to the target dataset." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.303, + 0.469, + 0.437 + ], + "angle": 0, + "content": "For the model architecture that is used to train the surrogate encoder, we consider two cases. First, we assume the adversary is aware of the target encoder's architecture and can train the same architecture surrogate encoder. Then we relax our assumption that the adversary uses different architectures to train the surrogate encoder. Our evaluation shows that the choice of architecture does not have much impact on the attack performance (see Table 3), which makes the attack more realistic." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.439, + 0.469, + 0.53 + ], + "angle": 0, + "content": "Note that we also compare our attacks against the encoders to the traditional model stealing attacks that focus on the whole classifier (which has an encoder and a linear layer). If the attack targets a whole classifier, we assume the adversary may obtain the posteriors or the predicted label for an input image." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.542, + 0.296, + 0.559 + ], + "angle": 0, + "content": "3. Model Stealing Attacks" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.567, + 0.469, + 0.629 + ], + "angle": 0, + "content": "In this section, we first describe the conventional attacks against the encoders. Then, we propose a novel contrastive stealing framework, Cont-Steal, to steal the encoders more effectively." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.636, + 0.419, + 0.652 + ], + "angle": 0, + "content": "3.1. Conventional Attacks Against Encoders" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.66, + 0.469, + 0.703 + ], + "angle": 0, + "content": "The adversary takes two steps to conduct the model stealing attacks against the target encoder and one step for further evaluation." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.705, + 0.469, + 0.853 + ], + "angle": 0, + "content": "Obtain the Surrogate Dataset. To conduct model stealing attacks, the adversary first needs to obtain a surrogate dataset. Based on the knowledge of the target classifier's training dataset (target dataset), we consider two cases. If the adversary has full knowledge of the target dataset, they can directly leverage the target dataset itself as the surrogate dataset. Or the adversary has no knowledge of the target dataset, which means that they can only construct the surrogate dataset, which is distributed differently from the target dataset." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Train the Surrogate Encoder. Slightly different from the classifier, the response of the encoder is an embedding, which is a feature vector. In this case, the adversary can still" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.892, + 0.121 + ], + "angle": 0, + "content": "leverage a similar loss function to optimize the surrogate encoder, which can be defined as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.579, + 0.135, + 0.892, + 0.176 + ], + "angle": 0, + "content": "\\[\nL _ {M S} = \\sum_ {k = 1} ^ {N} l \\left(E _ {T} \\left(x _ {k}\\right), E _ {S} \\left(x _ {k}\\right)\\right) \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.184, + 0.892, + 0.228 + ], + "angle": 0, + "content": "where \\( E_{T}(\\cdot) / E_{S}(\\cdot) \\) is the target/surrogate encoder, \\( N \\) is the total number of samples on the surrogate dataset, and \\( l(\\cdot) \\) is the MSE loss." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.23, + 0.892, + 0.381 + ], + "angle": 0, + "content": "Apply the Surrogate Encoder to Downstream Tasks. To evaluate the effectiveness of model stealing attacks against the encoder, the adversary can leverage the same downstream task to both the target and surrogate encoders. Concretely, the adversary trains an extra linear layer for the target and surrogate encoders, respectively. Note that we refer to the target/surrogate encoder and the extra linear layer as the target/surrogate classifiers. Then, the adversary quantifies the attack effectiveness by measuring the performance of the target/surrogate classifier on the downstream tasks." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.392, + 0.821, + 0.408 + ], + "angle": 0, + "content": "3.2. Cont-Steal Attacks Against Encoders" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.416, + 0.892, + 0.595 + ], + "angle": 0, + "content": "To better leverage the rich information from the embeddings, we propose Cont-Steal, a contrastive learning-based model stealing attacks against encoders, which leverages contrastive learning to enhance the stealing performance. Concretely, Cont-Steal aims to enforce the surrogate embedding of an image to get close to its target embedding (defined as a positive pair), and also push away embeddings of different images regardless of being generated by the target or the surrogate encoders (defined as negative pairs). There are three steps for the adversary to conduct contrastive stealing attacks against encoders and one step for further evaluation." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.598, + 0.892, + 0.748 + ], + "angle": 0, + "content": "Obtain the Surrogate Dataset. The adversary follows the same strategy as Section 3.1 to obtain the surrogate dataset. Data Augmentation. Our proposed Cont-Steal leverages data augmentation to transform an input image into its two augmented views. In this paper, we leverage RandAugment [11] as the augmentation method, which is made up of a group of advanced augmentation operations. Concretely, we set \\( n = 2 \\) and \\( m = 14 \\) following Cubuk et al. [11] where \\( n \\) denotes the number of transformations to a given sample and \\( m \\) represents the magnitude of global distortion." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.892, + 0.855 + ], + "angle": 0, + "content": "Train the Surrogate Encoder. Instead of querying the encoders with the original images, the adversary queries the encoders with the augmented views of them. Concretely, for an input image \\( x_{i} \\), we generate two augmented views of it, i.e., \\( \\widetilde{x}_{i,s} \\) and \\( \\widetilde{x}_{i,t} \\), where \\( \\widetilde{x}_{i,s} / \\widetilde{x}_{i,t} \\) is used to query the surrogate/target encoder. We consider \\( (\\widetilde{x}_{i,s},\\widetilde{x}_{j,t}) \\) as a positive pair if \\( i = j \\), and otherwise a negative pair." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.856, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Given a mini-batch of \\(N\\) samples, we generate \\(N\\) augmented views as the input of the target encoder and another \\(N\\) augmented views as the input of the surrogate encoders." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16375" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.12 + ], + "angle": 0, + "content": "Concretely, the loss of Cont-Steal can be formulated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.093, + 0.153, + 0.47, + 0.17 + ], + "angle": 0, + "content": "\\[\nD _ {\\text {e n c o d e r}} ^ {+} (i) = \\exp \\left(\\operatorname {s i m} \\left(E _ {S} \\left(\\tilde {x} _ {i, s}\\right), E _ {T} \\left(\\tilde {x} _ {i, t}\\right)\\right) / \\tau\\right), \\tag {2}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.093, + 0.173, + 0.469, + 0.21 + ], + "angle": 0, + "content": "\\[\nD _ {\\text {e n c o d e r}} ^ {-} (i) = \\sum_ {k = 1} ^ {N} \\left(\\exp \\left(\\operatorname {s i m} \\left(E _ {S} \\left(\\widetilde {x} _ {i, s}\\right), E _ {T} \\left(\\widetilde {x} _ {k, t}\\right)\\right) / \\tau\\right)\\right), \\tag {3}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.093, + 0.213, + 0.469, + 0.264 + ], + "angle": 0, + "content": "\\[\nD _ {s e l f} ^ {-} (i) = \\sum_ {k = 1} ^ {N} \\mathbb {1} _ {[ k \\neq i ]} \\left(\\exp \\left(\\operatorname {s i m} \\left(E _ {S} \\left(\\widetilde {x} _ {i, s}\\right), E _ {S} \\left(\\widetilde {x} _ {k, s}\\right)\\right) / \\tau\\right)\\right), \\tag {4}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.093, + 0.268, + 0.469, + 0.302 + ], + "angle": 0, + "content": "\\[\nl (i) = - \\log \\frac {D _ {\\text {e n c o d e r}} ^ {+} (i)}{D _ {\\text {e n c o d e r}} ^ {-} (i) + D _ {\\text {s e l f}} ^ {-} (i)}, \\tag {5}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.093, + 0.304, + 0.469, + 0.333 + ], + "angle": 0, + "content": "\\[\nL _ {\\text {C o n t - S t e a l}} = \\frac {\\sum_ {k = 1} ^ {N} l (k)}{N}, \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.351, + 0.469, + 0.41 + ], + "angle": 0, + "content": "where \\( E_{S}(\\cdot) \\) and \\( E_{T}(\\cdot) \\) denotes the surrogate and target encoder, \\( sim(u,v) = u^{T}v / ||u||||v|| \\) represents the cosine similarity between \\( u \\) and \\( v \\), and \\( \\tau \\) is parameter to control the temperature." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.413, + 0.469, + 0.685 + ], + "angle": 0, + "content": "As illustrated in Figure 2, the conventional attack treats each embedding individually without interacting across pairs. However, different embeddings are beneficial to each other as they can serve as anchors to better locate the position of the other embeddings in their space. Cont-Steal maximizes the similarity of embeddings generated from the target and surrogate encoders for a positive pair \\((\\widetilde{x}_{i,s},\\widetilde{x}_{i,t})\\) (orange arrows in Figure 2). For the embedding generated from the target and surrogate encoders for any pair \\((\\widetilde{x}_{i,s},\\widetilde{x}_{j,t})\\), contrastive stealing aims to make them more distant (green arrows in Figure 2). Besides, as pointed out by Chen et al. [6], contrastive learning benefits larger negative samples. To achieve this goal, we also consider the embeddings generated from the surrogate encoder for augmented views of different images, i.e., \\((\\widetilde{x}_{i,s},\\widetilde{x}_{j,s})\\), as negative pairs minimize their similarity (blue arrows in Figure 2). We later show that such design can enhance the performance of contrastive stealing (see Table 5)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.687, + 0.469, + 0.732 + ], + "angle": 0, + "content": "Apply the Surrogate Encoder to Downstream Tasks. We follow Section 3.1 to evaluate the effectiveness of model stealing on downstream tasks." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.753, + 0.208, + 0.769 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.781, + 0.469, + 0.9 + ], + "angle": 0, + "content": "In this section, we first describe the experimental setup in Section 4.1. Then we show the performance of the target encoders on the downstream tasks. Next, we summarize the performance of conventional attacks against classifiers and encoders in Section 4.2. Lastly, we evaluate the performance of Cont-Steal and conduct ablation studies to demonstrate its effectiveness under different settings in Section 4.3." + }, + { + "type": "image", + "bbox": [ + 0.511, + 0.096, + 0.87, + 0.257 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.52, + 0.266, + 0.878, + 0.455 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.469, + 0.892, + 0.593 + ], + "angle": 0, + "content": "Figure 2. Conventional attack (top) vs. Cont-Steal (bottom) against encoders. Conventional attack applies MSE loss to approximate target embeddings for each sample individually. Cont-Steal (bottom) introduces data augmentation and interacts across multiple samples: associating target/surrogate embeddings of the same images closer and repulsing those of different images farther away. The target encoder \\((E_{t})\\) is pre-trained and fixed, as shown in the solid frame. The surrogate encoder \\((E_{s})\\) is trainable by the adversary, as shown in the dashed frame." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.622, + 0.69, + 0.639 + ], + "angle": 0, + "content": "4.1. Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.646, + 0.892, + 0.811 + ], + "angle": 0, + "content": "Our encoders are pre-trained on CIFAR10 [1], and ImageNet [12]. We use four different kinds of contrastive methods: SimCLR [6], MoCo [20], BYOL [17] and SimSiam [8] to train a ResNet18 [21] as our target encoders. Our implementation is based on a PyTorch framework of contrastive learning. Then, these well pre-trained encoders will be applied to train downstream classifiers on CIFAR10 [1], STL10 [9], Fashion-MNIST [54], and SVHN [35]. In the experiments in the model stealing section, we use CIFAR10, STL10, Fashion-MNIST, and SVHN to conduct the attack." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.814, + 0.892, + 0.874 + ], + "angle": 0, + "content": "Agreement and accuracy are used as metrics to evaluate the model stealing attack's performance. The agreement will evaluate the similarity of surrogate encoders and target encoders in downstream tasks. The accuracy will evaluate" + }, + { + "type": "page_footnote", + "bbox": [ + 0.516, + 0.887, + 0.833, + 0.9 + ], + "angle": 0, + "content": "5https://github.com/vturrisi/solo-learn/" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16376" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.116, + 0.089, + 0.272, + 0.18 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.159, + 0.183, + 0.231, + 0.195 + ], + "angle": 0, + "content": "(a) CIFAR10" + }, + { + "type": "image", + "bbox": [ + 0.274, + 0.089, + 0.43, + 0.181 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.307, + 0.183, + 0.4, + 0.196 + ], + "angle": 0, + "content": "(b) ImageNet100" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.207, + 0.471, + 0.29 + ], + "angle": 0, + "content": "Figure 3. The performance of target classifiers composed by target encoder and an extra linear layer. The encoders are pre-trained on CIFAR10 (a) and ImageNet100 (b). The x-axis represents different downstream datasets for the target encoder and classifier. The y-axis represents the target model's accuracy on downstream tasks." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.319, + 0.47, + 0.396 + ], + "angle": 0, + "content": "the utility of surrogate models on downstream tasks. For each metric, a larger value is more desirable. During the stealing process, we set the batch size as 128 and the learning rate as 0.001. We show more results on the impact of hyperparameters in Supplementary Material in Section A.2" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.408, + 0.402, + 0.423 + ], + "angle": 0, + "content": "4.2. Performance of Conventional Attacks" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.432, + 0.47, + 0.749 + ], + "angle": 0, + "content": "We first show the target encoder's performance in various downstream tasks. The results are summarized in Figure 3. We conduct our experiments to explore whether the encoders are more vulnerable to model stealing attacks. We show our results of target encoders and downstream classifiers both trained on CIFAR10 in Figure 4. In all cases, the adversary can get better attack performance by stealing encoders rather than classifiers. This gap becomes especially apparent when the adversary has absolutely no knowledge of the train data. This is because the rich information in embeddings can better facilitate the learning process of surrogate encoders. For instance, when the surrogate dataset is CIFAR10 (the same as the target downstream dataset), stealing SimCLR's embeddings can achieve 0.785 agreement, while stealing predicted labels can achieve 0.712 agreement. However, when the surrogate dataset is totally different from the downstream target dataset, e.g., SVHN, stealing embeddings from SimCLR can still achieve 0.507 agreement while the agreement of stealing predicted labels drops to 0.192. We show more results in Supplementary Material Section A.4 due to the page limitation." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.471, + 0.902 + ], + "angle": 0, + "content": "We also find that all model stealing attacks' accuracy and agreement are highly correlated. As shown in Figure 5, the agreement is highly correlated with the accuracy. This indicates that besides accuracy, the agreement can also be used as a metric to evaluate the performance of model stealing attacks. We show the result on Figure 5. It can be obviously seen that agreement is highly related to accuracy. We use the linear regression method to describe the relationship between agreement and accuracy and find that the relation function is \\( y = 0.940 * x \\)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.091, + 0.742, + 0.106 + ], + "angle": 0, + "content": "4.3. Performance of Cont-Steal" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.115, + 0.894, + 0.205 + ], + "angle": 0, + "content": "As shown in Section 4.2, encoders are more vulnerable to model stealing attacks since the embedding usually contains richer information compared to the predicted label or posteriors. We then show that our proposed Cont-Steal can achieve better attack performance by making deeper use of embeddings' information." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.206, + 0.893, + 0.583 + ], + "angle": 0, + "content": "Figure 7 shows the attack performance when the target pre-training dataset is CIFAR10. Note that we also show the attack performance on other settings in the appendix. We discover that compared to conventional attacks against encoders, Cont-Steal can consistently achieve better performance. For instance, as shown in Figure 7d, when the target encoder is MoCo trained on CIFAR10, if the adversary uses STL10 to conduct model stealing attacks against encoders, the surrogate encoder can achieve 0.841 agreement in CIFAR10 downstream tasks with the Cont-Steal but only 0.479 with conventional attacks. Another finding is that compared to the same distribution surrogate dataset, our Cont-Steal can better enhance the performance when the surrogate dataset comes from a different distribution from the pre-trained dataset. For instance, when the target encoder is SimCLR trained on CIFAR10, Cont-Steal outperforms conventional attack by 0.055 agreement when the surrogate dataset is also CIFAR10, while the improvement increases to 0.207 and 0.214 when the surrogate dataset is STL10. We show more comparing results in Supplementary Material Section A.5. Note that our Cont-Steal also has great performance on other recent state-of-the-art visual models (ViT [13], MAE [19], and CLIP [39]), as we show in Section A.6, and can have better performance than other recent similar attacks [14, 31] shown in Section A.7." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.584, + 0.894, + 0.901 + ], + "angle": 0, + "content": "To better understand why Cont-Steal can always achieve better performance, we extract samples' embeddings generated by different encoders, i.e., the target encoder, surrogate encoder trained with the conventional attack, and surrogate encoder trained with the Cont-Steal, and project them into a 2-dimensional space using t-SNE. From the results summarized in Figure 6, we find that Cont-Steal can effectively mimic the pattern of the embeddings as the target encoder. However, the conventional attack fails to capture such patterns for a number of input samples, e.g., the outer circle in Figure 8c. This further demonstrates that Cont-Steal benefits from jointly considering different embeddings as they can serve as anchors to better locate the position of the other embeddings in their space. We also show some ablation study results on Supplementary Material Section A.2 to show that with the less surrogate dataset, less training epoch, and different model architecture, Cont-Steal can still achieve much better results than conventional steal. Also, we show further attacks based on the stole models on Supplementary Material Section A.3 to show that Cont-Steal can be used as a springboard for other attacks." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16377" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.137, + 0.089, + 0.307, + 0.188 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.189, + 0.192, + 0.261, + 0.203 + ], + "angle": 0, + "content": "(a) CIFAR10" + }, + { + "type": "image", + "bbox": [ + 0.312, + 0.09, + 0.482, + 0.187 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.37, + 0.192, + 0.428, + 0.203 + ], + "angle": 0, + "content": "(b) STL10" + }, + { + "type": "image", + "bbox": [ + 0.486, + 0.091, + 0.657, + 0.187 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.536, + 0.192, + 0.609, + 0.203 + ], + "angle": 0, + "content": "(c) F-MNIST" + }, + { + "type": "image", + "bbox": [ + 0.66, + 0.091, + 0.831, + 0.187 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.719, + 0.192, + 0.774, + 0.203 + ], + "angle": 0, + "content": "(d) SVHN" + }, + { + "type": "image", + "bbox": [ + 0.137, + 0.206, + 0.307, + 0.302 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.189, + 0.307, + 0.261, + 0.319 + ], + "angle": 0, + "content": "(e) CIFAR10" + }, + { + "type": "image", + "bbox": [ + 0.312, + 0.206, + 0.483, + 0.302 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.371, + 0.307, + 0.427, + 0.319 + ], + "angle": 0, + "content": "(f) STL10" + }, + { + "type": "image", + "bbox": [ + 0.487, + 0.206, + 0.657, + 0.303 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.536, + 0.307, + 0.609, + 0.32 + ], + "angle": 0, + "content": "(g) F-MNIST" + }, + { + "type": "image", + "bbox": [ + 0.66, + 0.206, + 0.832, + 0.303 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.719, + 0.307, + 0.775, + 0.319 + ], + "angle": 0, + "content": "(h) SVHN" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.332, + 0.895, + 0.388 + ], + "angle": 0, + "content": "Figure 4. The performance of model stealing attack against target encoders and downstream classifiers both trained on CIFAR10. Target models can output predicted labels, posteriors, or embeddings. The adversary uses CIFAR10, STL10, Fashion-MNIST (F-MNIST), and SVHN to conduct model stealing attacks. The x-axis represents different kinds of target models. The first line's y-axis represents the agreement of the model stealing attack. The second line's y-axis represents the accuracy of the model stealing attack." + }, + { + "type": "image", + "bbox": [ + 0.11, + 0.417, + 0.439, + 0.605 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.615, + 0.471, + 0.656 + ], + "angle": 0, + "content": "Figure 5. The relationship between accuracy and agreement. The x-axis is the agreement number, and the y-axis is the accuracy number." + }, + { + "type": "image", + "bbox": [ + 0.11, + 0.697, + 0.201, + 0.75 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.105, + 0.76, + 0.204, + 0.772 + ], + "angle": 0, + "content": "(a) Target encoder" + }, + { + "type": "image", + "bbox": [ + 0.231, + 0.697, + 0.32, + 0.75 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.235, + 0.76, + 0.312, + 0.772 + ], + "angle": 0, + "content": "(b) Cont-Steal" + }, + { + "type": "image", + "bbox": [ + 0.351, + 0.697, + 0.439, + 0.75 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.349, + 0.76, + 0.439, + 0.772 + ], + "angle": 0, + "content": "(c) Conventional" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.785, + 0.471, + 0.853 + ], + "angle": 0, + "content": "Figure 6. The t-SNE projection of 1,000 randomly selected samples' embeddings from target encoder, surrogate encoder under Cont-Steal, and surrogate encoder under the conventional attack, respectively. Note that the target encoder is pre-trained by SimCLR on CIFAR10." + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.412, + 0.892, + 0.482 + ], + "angle": 0, + "content": "Table 1. The monetary and (training) time costs for normal training and Cont-Steal attack. Cont-Steal's monetary cost contains two parts: query cost and training cost. Note that we ignore the query time cost of Cont-Steal as it normally has a smaller value than the training time cost." + }, + { + "type": "table", + "bbox": [ + 0.511, + 0.493, + 0.882, + 0.589 + ], + "angle": 0, + "content": "
Monetary CostTime Cost
ModelNormal ($)Cont-Steal ($)Normal (h)Cont-Steal (h)
SimCLR58.6811.83 (1.83 + 10)20.010.62
MoCo54.8312.13 (2.13 + 10)18.690.73
BYOL61.4612.08 (2.08 + 10)20.960.71
SimSiam57.1412.00 (2.00 + 10)19.460.68
" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.613, + 0.641, + 0.629 + ], + "angle": 0, + "content": "4.4. Cost Analysis" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.637, + 0.892, + 0.803 + ], + "angle": 0, + "content": "As we mentioned before, pre-train a state-of-the-art encoder is time-consuming and resource-demanding. We wonder if the model stealing attacks can steal the functionality of the encoder with much less cost. To this end, we evaluate the time and monetary cost of training an encoder from scratch or stealing a pre-trained encoder via ContSteal. The monetary cost of model stealing includes querying the target model and training the surrogate model. We refer to the query price as \\(1 for 1,000 queries based on AWS.6 Our experiment is conducted on 1 NVIDIA A100 whose price is \\)2.934 per hour based on google cloud.7" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.803, + 0.892, + 0.864 + ], + "angle": 0, + "content": "The monetary and time cost is shown in Table 1. We observe that Cont-Steal can obtain a surrogate encoder with much less money and time cost than training the encoder from scratch. For instance, a ResNet18 trained by SimCLR" + }, + { + "type": "page_footnote", + "bbox": [ + 0.516, + 0.875, + 0.862, + 0.888 + ], + "angle": 0, + "content": "6https://aws.amazon.com/rekognition/pricing/" + }, + { + "type": "page_footnote", + "bbox": [ + 0.52, + 0.889, + 0.888, + 0.901 + ], + "angle": 0, + "content": "7https://cloud.google.com/compute/gpus-pricing/" + }, + { + "type": "list", + "bbox": [ + 0.516, + 0.875, + 0.888, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16378" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.137, + 0.089, + 0.307, + 0.188 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.189, + 0.192, + 0.261, + 0.203 + ], + "angle": 0, + "content": "(a) CIFAR10" + }, + { + "type": "image", + "bbox": [ + 0.312, + 0.09, + 0.483, + 0.187 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.37, + 0.192, + 0.428, + 0.203 + ], + "angle": 0, + "content": "(b) STL10" + }, + { + "type": "image", + "bbox": [ + 0.487, + 0.091, + 0.657, + 0.187 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.536, + 0.192, + 0.609, + 0.203 + ], + "angle": 0, + "content": "(c) F-MNIST" + }, + { + "type": "image", + "bbox": [ + 0.66, + 0.091, + 0.831, + 0.187 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.719, + 0.192, + 0.774, + 0.203 + ], + "angle": 0, + "content": "(d) SVHN" + }, + { + "type": "image", + "bbox": [ + 0.137, + 0.206, + 0.308, + 0.302 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.189, + 0.307, + 0.261, + 0.319 + ], + "angle": 0, + "content": "(e) CIFAR10" + }, + { + "type": "image", + "bbox": [ + 0.312, + 0.206, + 0.483, + 0.302 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.371, + 0.307, + 0.427, + 0.319 + ], + "angle": 0, + "content": "(f) STL10" + }, + { + "type": "image", + "bbox": [ + 0.487, + 0.206, + 0.657, + 0.302 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.536, + 0.307, + 0.609, + 0.32 + ], + "angle": 0, + "content": "(g) F-MNIST" + }, + { + "type": "image", + "bbox": [ + 0.66, + 0.206, + 0.831, + 0.302 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.719, + 0.307, + 0.774, + 0.319 + ], + "angle": 0, + "content": "(h) SVHN" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.332, + 0.895, + 0.388 + ], + "angle": 0, + "content": "Figure 7. The performance of Cont-Steal and conventional attack against target encoders trained on CIFAR10. The adversary uses CIFAR10, STL10, F-MNIST, and SVHN to conduct model stealing attacks. The adversary uses CIFAR10 as the downstream task to evaluate the attack performance. The x-axis represents different kinds of the target model. The first line's y-axis represents the agreement of the model stealing attack. The second line's y-axis represents the accuracy of the model stealing attack." + }, + { + "type": "image", + "bbox": [ + 0.084, + 0.412, + 0.209, + 0.482 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.102, + 0.488, + 0.192, + 0.5 + ], + "angle": 0, + "content": "(a) Adding noise" + }, + { + "type": "image", + "bbox": [ + 0.212, + 0.412, + 0.336, + 0.483 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.248, + 0.488, + 0.301, + 0.5 + ], + "angle": 0, + "content": "(b) Top- \\(k\\)" + }, + { + "type": "image", + "bbox": [ + 0.34, + 0.412, + 0.464, + 0.483 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.365, + 0.488, + 0.438, + 0.5 + ], + "angle": 0, + "content": "(c) Rounding" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.512, + 0.47, + 0.569 + ], + "angle": 0, + "content": "Figure 8. The performance of different defense methods. Target encoders are trained on CIFAR10. The downstream dataset and surrogate dataset are both STL10. The x-axis represents different defense levels. The y-axis represents the model's accuracy." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.595, + 0.47, + 0.687 + ], + "angle": 0, + "content": "on CIFAR10 takes 20.01 hours and \\(58.68\\)'s on 1 NVIDIA A100 GPU, while Cont-Steal only takes 0.62 hours and \\(11.83\\)'s to steal an encoder that performs similarly on downstream tasks. The results demonstrate that Cont-Steal is able to construct surrogate encoders that perform similarly to the target encoders but with much less time and monetary cost." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.696, + 0.182, + 0.71 + ], + "angle": 0, + "content": "4.5. Defenses" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.72, + 0.469, + 0.795 + ], + "angle": 0, + "content": "In this section, we will consider different defenses against model stealing attacks on encoders to evaluate the robustness of our proposed attack. We divided all defenses into three categories: perturbation-based defense [37] and watermark-based defense [2]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.795, + 0.469, + 0.87 + ], + "angle": 0, + "content": "Perturbation-based Defense. In this defense setting, the defender aims to perturb the output of the target model to limit the information the adversary can obtain. The common practice of this kind of defense includes adding noise [37], top-\\(k\\) [37], and feature rounding [48]." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Adding noise means that the defender will introduce noise value to the original output of the model. In our case," + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.415, + 0.893, + 0.793 + ], + "angle": 0, + "content": "we consider adding Gaussian noise to the embeddings generated by the target encoder. We set the mean value to 0, and different noise levels represent different standard deviations of the Gaussian distribution. For Top-\\(k\\), the defender will only output the first \\(k\\) largest number of each embedding (and set the rest as 0). In this way, the high-dimensional information of the image contained in embeddings can be appropriately reduced. Regarding feature rounding, the defender will truncate the values in the embedding to a specific digit. As a case study, we consider a ResNet18 encoder pre-trained on CIFAR10 with SimCLR and take STL10 to train its downstream classifier. The experimental results are summarized in Figure 8. We can observe that while adding noise and top-\\(k\\) can reduce the model stealing attacks' performance, it may also degrade the target model performance to a large extent. For instance, when the noise increases from 0 to 10, the attack performance of ContSteal decreases from 0.729 to 0.410, while the target encoder's performance drops from 0.734 to 0.098. On the other hand, rounding only has a limited effect on both target model performance and attack performance. This indicates that perturbation-based defense cannot defend against the encoder's model stealing attack effectively since they cannot reach a good trade-off between attack performance and model utility." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.795, + 0.893, + 0.901 + ], + "angle": 0, + "content": "Watermark-based Defense. Watermark-based defense is also one of the most popular defense methods against model stealing attacks [2]. Watermark provides copyright protection by adding some specific identification to the target model. If the surrogate model is stolen from the watermarked target model, then ideally, it will contain the same watermark as well. Adi et al. [2] show that back" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16379" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.089, + 0.472, + 0.146 + ], + "angle": 0, + "content": "Table 2. Watermark defense. Pretrain dataset and surrogate dataset are both CIFAR10. Watermark leverages a watermark rate (wr) to verify the ownership of target models. A higher wr denotes a better verification performance." + }, + { + "type": "table", + "bbox": [ + 0.08, + 0.156, + 0.472, + 0.236 + ], + "angle": 0, + "content": "
DatasetTarget model (acc/wr)Cont-Steal (acc/wr)Baseline (acc/wr)
CIFAR100.864 / 0.9980.769 / 0.1300.871 / 0.095
STL100.721 / 0.9990.702 / 0.0340.733 / 0.111
SVHN0.501 / 0.9990.535 / 0.3030.492 / 0.103
F-MNIST0.857 / 0.9990.813 / 0.0610.850 / 0.099
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.259, + 0.47, + 0.395 + ], + "angle": 0, + "content": "door technology can be used as the watermark to protect the model. In that sense, BadEncoder [26], a backdoor mechanism against the encoder, can be leveraged as a watermarking technology for our target encoder as well. The defenders first will train the watermarked (backdoored) encoder, where images with a certain trigger will cause misclassification. Then, if they find the surrogate model can also misclassify images with the same trigger, the defenders can claim ownership of the surrogate model." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.396, + 0.471, + 0.697 + ], + "angle": 0, + "content": "In our experiments, we leverage BadEncoder to watermark the encoder pre-trained on CIFAR10 by SimCLR, and leverage different downstream datasets to perform different tasks. We assume a strong adversary that has the same downstream dataset as the surrogate dataset. Also, we consider the baseline cases where the trigger samples are fed into the clean model to calculate the watermark rate (wr). As shown in Table 2), the watermark cannot be preserved as the surrogate models constructed by Cont-Steal have similar wr as the baseline model. For instance, when the downstream dataset is CIFAR10, Cont-Steal builds a surrogate model with 0.769 accuracy while only 0.130 wr, which is close to the baseline model. This indicates that Cont-Steal can bypass the watermarking technique as it can reach similar utility while reducing the wr to a large extent. Note that there is also another work to protect contrastive learning models from model stealing attacks using dataset inference [15]. We show in Supplementary Material in Section A.8 that this kind of defense can be easily bypassed by Cont-Steal." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.709, + 0.22, + 0.725 + ], + "angle": 0, + "content": "5. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.735, + 0.472, + 0.901 + ], + "angle": 0, + "content": "Contrastive Learning. Contrastive learning is one of the most popular methods to train encoders. Current works [6, 8, 17, 20, 49, 53] propose different advanced contrastive learning algorithms. SimCLR, MoCo, BYOL, SimSiam are currently the mainstream frameworks of contrastive learning. Thus, we concentrate on them in this paper. There are many works on evaluating the security and privacy risks of contrastive learning. Previous works [23, 26, 30] propose membership inference attacks, attribution inference attacks, and backdoor attacks on contrastive learning. All proposed attacks show that contrastive-based models are vulnerable" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.122 + ], + "angle": 0, + "content": "to popular attacks. Therefore, the security issues of self-supervised learning deserve more attention." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.128, + 0.895, + 0.371 + ], + "angle": 0, + "content": "Model Stealing Attack. In model stealing, the adversary's goal is to steal part of the target model. Tramér et al. [48] proposed the first model stealing attack against black-box machine learning API to steal its parameters. Wang et al. [50] proposed the first hyperparameter stealing attacks against ML models. Oh et al. [36] also tried to steal machine learning model's architectures and hyperparameters. Orekondy et al. [37] proposed knockoff nets, which aim at stealing the functionality of black-box models. Krishna et al. [28] formulated a model stealing attack against BERT-based API. Besides, Wu et al. [51] and Shen et al. [44] perform model stealing attacks against Graph Neural Networks. These works often have relatively strong assumptions, such as the model family is known and the victim's data is partly available while we conduct model stealing attacks against encoders and relax the above assumption." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.403, + 0.619, + 0.419 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.435, + 0.892, + 0.526 + ], + "angle": 0, + "content": "In this paper, we conduct the first model stealing risk assessment towards image encoders. Our evaluation shows that the encoder is more vulnerable to model stealing attacks compared to the classifier. This is because the embedding provided by the encoder contains richer information than the posteriors or predicted labels from whole classifiers." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.533, + 0.892, + 0.744 + ], + "angle": 0, + "content": "To better unleash the power from the embeddings, we propose Cont-Steal, a contrastive learning-based model stealing method against encoders. Concretely, Cont-Steal introduces different types of negative pairs as \"anchors\" to better navigate the surrogate encoder and learn the functionality of the target encoder. Extensive evaluations show that Cont-Steal consistently performs better than conventional attacks against encoders. And such an advantage is further amplified when the adversary has no information on the target dataset, a limited amount of data, and restricted query budgets. Our work points out that the threat of model stealing attacks against encoders is largely underestimated, which prompts the need for more effective intellectual property protection of representation learning techniques." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.895, + 0.901 + ], + "angle": 0, + "content": "Acknowledgments. We thank all anonymous reviewers for their constructive comments. This work is partially funded by the Helmholtz Association within the project \"Trustworthy Federated Data Analytics\" (TFDA) (funding number ZT-I-OO1 4) and by the European Health and Digital Executive Agency (HADEA) within the project \"Understanding the individual host response against Hepatitis D Virus to develop a personalized approach for the management of hepatitis D\" (D-Solve) (grant agreement number 101057917)." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "16380" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.176, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.116, + 0.468, + 0.145 + ], + "angle": 0, + "content": "[1] https://www.cs.toronto.edu/~kriz/cifar.html." + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.157, + 0.468, + 0.248 + ], + "angle": 0, + "content": "[2] Yossi Adi, Carsten Baum, Moustapha Cisse, Benny Pinkas, and Joseph Keshet. Turning Your Weakness Into a Strength: Watermarking Deep Neural Networks by Backdoorsing. In USENIX Security Symposium (USENIX Security), pages 1615-1631. USENIX, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.089, + 0.261, + 0.468, + 0.366 + ], + "angle": 0, + "content": "[3] Battista Biggio, Igino Corona, Davide Maiorca, Blaine Nelson, Nedim Srndic, Pavel Laskov, Giorgio Giacinto, and Fabio Roli. Evasion Attacks against Machine Learning at Test Time. In European Conference on Machine Learning and Principles and Practice of Knowledge Discovery in Databases (ECML/PKDD), pages 387-402. Springer, 2013." + }, + { + "type": "ref_text", + "bbox": [ + 0.089, + 0.378, + 0.468, + 0.437 + ], + "angle": 0, + "content": "[4] Nicholas Carlini and David Wagner. Towards Evaluating the Robustness of Neural Networks. In IEEE Symposium on Security and Privacy (S&P), pages 39-57. IEEE, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.089, + 0.45, + 0.468, + 0.509 + ], + "angle": 0, + "content": "[5] Varun Chandrasekaran, Kamalika Chaudhuri, Irene Giacomelli, Somesh Jha, and Songbai Yan. Model Extraction and Active Learning. CoRR abs/1811.02054, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.089, + 0.522, + 0.468, + 0.598 + ], + "angle": 0, + "content": "[6] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey E. Hinton. A Simple Framework for Contrastive Learning of Visual Representations. In International Conference on Machine Learning (ICML), pages 1597-1607. PMLR, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.089, + 0.61, + 0.468, + 0.7 + ], + "angle": 0, + "content": "[7] Xiaoyi Chen, Ahmed Salem, Michael Backes, Shiqing Ma, Qingni Shen, Zhonghai Wu, and Yang Zhang. BadNL: Backdoor Attacks Against NLP Models with Semantic-preserving Improvements. In Annual Computer Security Applications Conference (ACSAC), pages 554-569. ACSAC, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.089, + 0.712, + 0.468, + 0.772 + ], + "angle": 0, + "content": "[8] Xinlei Chen and Kaiming He. Exploring Simple Siamese Representation Learning. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 15750-15758. IEEE, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.089, + 0.784, + 0.468, + 0.858 + ], + "angle": 0, + "content": "[9] Adam Coates, Andrew Y. Ng, and Honglak Lee. An Analysis of Single-Layer Networks in Unsupervised Feature Learning. In International Conference on Artificial Intelligence and Statistics (AISTATS), pages 215-223. JMLR, 2011." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.871, + 0.468, + 0.902 + ], + "angle": 0, + "content": "[10] Tianshuo Cong, Xinlei He, and Yang Zhang. SSL-Guard: A Watermarking Scheme for Self-supervised" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.116, + 0.468, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.092, + 0.892, + 0.137 + ], + "angle": 0, + "content": "Learning Pre-trained Encoders. In ACM SIGSAC Conference on Computer and Communications Security (CCS), pages 579-593. ACM, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.148, + 0.892, + 0.236 + ], + "angle": 0, + "content": "[11] Ekin Dogus Cubuk, Barret Zoph, Jon Shlens, and Quoc Le. RandAugment: Practical Automated Data Augmentation with a Reduced Search Space. In Annual Conference on Neural Information Processing Systems (NeurIPS), pages 18613-18624. NeurIPS, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.249, + 0.892, + 0.323 + ], + "angle": 0, + "content": "[12] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. ImageNet: A large-scale hierarchical image database. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 248–255. IEEE, 2009." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.335, + 0.892, + 0.455 + ], + "angle": 0, + "content": "[13] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale. In International Conference on Learning Representations (ICLR), 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.467, + 0.892, + 0.541 + ], + "angle": 0, + "content": "[14] Adam Dziedzic, Nikita Dhawan, Muhammad Ahmad Kaleem, Jonas Guan, and Nicolas Papernot. On the Difficulty of Defending Self-Supervised Learning against Model Extraction. In International Conference on Machine Learning (ICML). JMLR, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.553, + 0.892, + 0.642 + ], + "angle": 0, + "content": "[15] Adam Dziedzic, Haonan Duan, Muhammad Ahmad Kaleem, Nikita Dhawan, Jonas Guan, Yannis Cattan, Franziska Boenisch, and Nicolas Papernot. Dataset Inference for Self-Supervised Models. In Annual Conference on Neural Information Processing Systems (NeurIPS). NeurIPS, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.654, + 0.892, + 0.713 + ], + "angle": 0, + "content": "[16] Ian Goodfellow, Jonathon Shlens, and Christian Szegedy. Explaining and Harnessing Adversarial Examples. In International Conference on Learning Representations (ICLR), 2015." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.725, + 0.892, + 0.858 + ], + "angle": 0, + "content": "[17] Jean-Bastien Grill, Florian Strub, Florent Altché, Corentin Tallec, Pierre H. Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Ávila Pires, Zhaohan Guo, Mohammad Gheshlaghi Azar, Bilal Piot, Koray Kavukcuoglu, Rémi Munos, and Michal Valko. Bootstrap Your Own Latent - A New Approach to Self-Supervised Learning. In Annual Conference on Neural Information Processing Systems (NeurIPS). NeurIPS, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.871, + 0.892, + 0.902 + ], + "angle": 0, + "content": "[18] Yunhui Guo, Honghui Shi, Abhishek Kumar, Kristen Grauman, Tajana Rosing, and Rogério Schmidt" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.092, + 0.892, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.518, + 0.957 + ], + "angle": 0, + "content": "16381" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.114, + 0.092, + 0.47, + 0.152 + ], + "angle": 0, + "content": "Feris. SpotTune: Transfer Learning Through Adaptive Fine-Tuning. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 4805-4814. IEEE, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.162, + 0.47, + 0.222 + ], + "angle": 0, + "content": "[19] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dólár, and Ross B. Girshick. Masked Autoencoders Are Scalable Vision Learners. CoRR abs/2111.06377, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.233, + 0.469, + 0.307 + ], + "angle": 0, + "content": "[20] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross B. Girshick. Momentum Contrast for Unsupervised Visual Representation Learning. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 9726-9735. IEEE, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.318, + 0.469, + 0.378 + ], + "angle": 0, + "content": "[21] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep Residual Learning for Image Recognition. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 770-778. IEEE, 2016." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.389, + 0.469, + 0.448 + ], + "angle": 0, + "content": "[22] Xinlei He, Rui Wen, Yixin Wu, Michael Backes, Yun Shen, and Yang Zhang. Node-Level Membership Inference Attacks Against Graph Neural Networks. CoRR abs/2102.05429, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.459, + 0.469, + 0.519 + ], + "angle": 0, + "content": "[23] Xinlei He and Yang Zhang. Quantifying and Mitigating Privacy Risks of Contrastive Learning. In ACM SIGSAC Conference on Computer and Communications Security (CCS), pages 845-863. ACM, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.529, + 0.469, + 0.605 + ], + "angle": 0, + "content": "[24] Bo Hui, Yuchen Yang, Haolin Yuan, Philippe Burlina, Neil Zhenqiang Gong, and Yinzhi Cao. Practical Blind Membership Inference Attack via Differential Comparisons. In Network and Distributed System Security Symposium (NDSS). Internet Society, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.615, + 0.469, + 0.69 + ], + "angle": 0, + "content": "[25] Matthew Jagielski, Nicholas Carlini, David Berthelot, Alex Kurakin, and Nicolas Papernot. High Accuracy and High Fidelity Extraction of Neural Networks. In USENIX Security Symposium (USENIX Security), pages 1345-1362. USENIX, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.7, + 0.469, + 0.76 + ], + "angle": 0, + "content": "[26] Jinyuan Jia, Yupei Liu, and Neil Zhenqiang Gong. BadEncoder: Backdoor Attacks to Pre-trained Encoders in Self-Supervised Learning. In IEEE Symposium on Security and Privacy (S&P). IEEE, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.77, + 0.469, + 0.845 + ], + "angle": 0, + "content": "[27] Prannay Khosla, Piotr Teterwak, Chen Wang, Aaron Sarna, Yonglong Tian, Phillip Isola, Aaron Maschinot, Ce Liu, and Dilip Krishnan. Supervised Contrastive Learning. In Annual Conference on Neural Information Processing Systems (NeurIPS). NeurIPS, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.856, + 0.469, + 0.9 + ], + "angle": 0, + "content": "[28] Kalpesh Krishna, Gaurav Singh Tomar, Ankur P. Parikh, Nicolas Papernot, and Mohit Iyyer. Thieves on Sesame Street! Model Extraction of BERT-based" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.536, + 0.092, + 0.89, + 0.121 + ], + "angle": 0, + "content": "APIs. In International Conference on Learning Representations (ICLR), 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.132, + 0.892, + 0.176 + ], + "angle": 0, + "content": "[29] Alexey Kurakin, Ian Goodfellow, and Samy Bengio. Adversarial Examples in the Physical World. CoRR abs/1607.02533, 2016." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.187, + 0.892, + 0.262 + ], + "angle": 0, + "content": "[30] Hongbin Liu, Jinyuan Jia, Wenjie Qu, and Neil Zhenqiang Gong. EncoderMI: Membership Inference against Pre-trained Encoders in Contrastive Learning. In ACM SIGSAC Conference on Computer and Communications Security (CCS). ACM, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.273, + 0.892, + 0.348 + ], + "angle": 0, + "content": "[31] Yupei Liu, Jinyuan Jia, Hongbin Liu, and Neil Zhenqiang Gong. StolenEncoder: Stealing Pre-trained Encoders in Self-supervised Learning. In ACM SIGSAC Conference on Computer and Communications Security (CCS), pages 2115-212. ACM, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.359, + 0.892, + 0.433 + ], + "angle": 0, + "content": "[32] Aleksander Madry, Aleksandar Makelov, Ludwig Schmidt, Dimitris Tsipras, and Adrian Vladu. Towards Deep Learning Models Resistant to Adversarial Attacks. In International Conference on Learning Representations (ICLR), 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.444, + 0.892, + 0.519 + ], + "angle": 0, + "content": "[33] Milad Nasr, Reza Shokri, and Amir Houmansadr. Machine Learning with Membership Privacy using Adversarial Regularization. In ACM SIGSAC Conference on Computer and Communications Security (CCS), pages 634-646. ACM, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.529, + 0.892, + 0.618 + ], + "angle": 0, + "content": "[34] Milad Nasr, Reza Shokri, and Amir Houmansadr. Comprehensive Privacy Analysis of Deep Learning: Passive and Active White-box Inference Attacks against Centralized and Federated Learning. In IEEE Symposium on Security and Privacy (S&P), pages 1021-1035. IEEE, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.63, + 0.892, + 0.705 + ], + "angle": 0, + "content": "[35] Yuval Netzer, Tao Wang, Adam Coates, Alessandro Bissacco, Bo Wu, and Andrew Y. Ng. Reading Digits in Natural Images with Unsupervised Feature Learning. In Annual Conference on Neural Information Processing Systems (NIPS). NIPS, 2011." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.715, + 0.892, + 0.775 + ], + "angle": 0, + "content": "[36] Seong Joon Oh, Max Augustin, Bernt Schiele, and Mario Fritz. Towards Reverse-Engineering Black-Box Neural Networks. In International Conference on Learning Representations (ICLR), 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.786, + 0.892, + 0.86 + ], + "angle": 0, + "content": "[37] Tribhuvanesh Orekondy, Bernt Schiele, and Mario Fritz. Knockoff Nets: Stealing Functionality of Black-Box Models. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 4954–4963. IEEE, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.871, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[38] Nicolas Papernot, Patrick D. McDaniel, Somesh Jha, Matt Fredrikson, Z. Berkay Celik, and Ananthram" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.092, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "16382" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.114, + 0.091, + 0.469, + 0.151 + ], + "angle": 0, + "content": "Swami. The Limitations of Deep Learning in Adversarial Settings. In IEEE European Symposium on Security and Privacy (Euro S&P), pages 372-387. IEEE, 2016." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.165, + 0.469, + 0.271 + ], + "angle": 0, + "content": "[39] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning Transferable Visual Models From Natural Language Supervision. In International Conference on Machine Learning (ICML), pages 8748-8763. PMLR, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.283, + 0.469, + 0.372 + ], + "angle": 0, + "content": "[40] Alexandre Sablayrolles, Matthijs Douze, Cordelia Schmid, Yann Ollivier, and Hervé Jégou. White-box vs Black-box: Bayes Optimal Strategies for Membership Inference. In International Conference on Machine Learning (ICML), pages 5558-5567. PMLR, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.387, + 0.469, + 0.446 + ], + "angle": 0, + "content": "[41] Ahmed Salem, Michael Backes, and Yang Zhang. Don't Trigger Me! A Triggerless Backdoor Attack Against Deep Neural Networks. CoRR abs/2010.03282, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.459, + 0.469, + 0.55 + ], + "angle": 0, + "content": "[42] Ahmed Salem, Yang Zhang, Mathias Humbert, Pascal Berrang, Mario Fritz, and Michael Backes. MLLeaks: Model and Data Independent Membership Inference Attacks and Defenses on Machine Learning Models. In Network and Distributed System Security Symposium (NDSS). Internet Society, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.563, + 0.469, + 0.622 + ], + "angle": 0, + "content": "[43] Zeyang Sha, Xinlei He, Pascal Berrang, Mathias Humbert, and Yang Zhang. Fine-Tuning Is All You Need to Mitigate Backdoor Attacks. CoRR abs/2212.09067, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.636, + 0.469, + 0.696 + ], + "angle": 0, + "content": "[44] Yun Shen, Xinlei He, Yufei Han, and Yang Zhang. Model Stealing Attacks Against Inductive Graph Neural Networks. In IEEE Symposium on Security and Privacy (S&P), pages 1175-1192. IEEE, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.709, + 0.469, + 0.784 + ], + "angle": 0, + "content": "[45] Reza Shokri, Marco Stronati, Congzheng Song, and Vitaly Shmatikov. Membership Inference Attacks Against Machine Learning Models. In IEEE Symposium on Security and Privacy (S&P), pages 3-18. IEEE, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.798, + 0.469, + 0.857 + ], + "angle": 0, + "content": "[46] Liwei Song and Prateek Mittal. Systematic Evaluation of Privacy Risks of Machine Learning Models. In USENIX Security Symposium (USENIX Security). USENIX, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.871, + 0.469, + 0.9 + ], + "angle": 0, + "content": "[47] Florian Tramér, Alexey Kurakin, Nicolas Papernot, Ian Goodfellow, Dan Boneh, and Patrick McDaniel." + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.091, + 0.469, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.536, + 0.092, + 0.892, + 0.136 + ], + "angle": 0, + "content": "Ensemble Adversarial Training: Attacks and Defenses. In International Conference on Learning Representations (ICLR), 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.147, + 0.892, + 0.221 + ], + "angle": 0, + "content": "[48] Florian Tramér, Fan Zhang, Ari Juels, Michael K. Reiter, and Thomas Ristenpart. Stealing Machine Learning Models via Prediction APIs. In USENIX Security Symposium (USENIX Security), pages 601-618. USENIX, 2016." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.233, + 0.892, + 0.277 + ], + "angle": 0, + "content": "[49] Aäron van den Oord, Yazhe Li, and Oriol Vinyals. Representation Learning with Contrastive Predictive Coding. CoRR abs/1807.03748, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.287, + 0.892, + 0.347 + ], + "angle": 0, + "content": "[50] Binghui Wang and Neil Zhenqiang Gong. Stealing Hyperparameters in Machine Learning. In IEEE Symposium on Security and Privacy (S&P), pages 36-52. IEEE, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.358, + 0.892, + 0.418 + ], + "angle": 0, + "content": "[51] Bang Wu, Xiangwen Yang, Shirui Pan, and Xingliang Yuan. Model Extraction Attacks on Graph Neural Networks: Taxonomy and Realization. CoRR abs/2010.12751, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.429, + 0.892, + 0.518 + ], + "angle": 0, + "content": "[52] Huijun Wu, Chen Wang, Yuriy Tyshetskiy, Andrew Docherty, Kai Lu, and Liming Zhu. Adversarial Examples for Graph Data: Deep Insights into Attack and Defense. In International Joint Conferences on Artificial Intelligence (IJCAI), pages 4816-4823. IJCAI, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.529, + 0.892, + 0.605 + ], + "angle": 0, + "content": "[53] Zhirong Wu, Yuanjun Xiong, Stella X. Yu, and Dahua Lin. Unsupervised Feature Learning via Non-Parametric Instance Discrimination. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 3733–3742. IEEE, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.615, + 0.892, + 0.674 + ], + "angle": 0, + "content": "[54] Han Xiao, Kashif Rasul, and Roland Vollgraf. Fashion-MNIST: a Novel Image Dataset for Benchmarking Machine Learning Algorithms. CoRR abs/1708.07747, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.685, + 0.892, + 0.745 + ], + "angle": 0, + "content": "[55] Ziqing Yang, Zeyang Sha, Michael Backes, and Yang Zhang. From Visual Prompt Learning to Zero-Shot Transfer: Mapping Is All You Need. CoRR abs/2303.05266, 2023." + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.092, + 0.892, + 0.745 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.956 + ], + "angle": 0, + "content": "16383" + } + ] +] \ No newline at end of file diff --git a/2023/Can't Steal_ Cont-Steal! Contrastive Stealing Attacks Against Image Encoders/d755a24e-203a-45cb-a84e-020f7fb471b0_origin.pdf b/2023/Can't Steal_ Cont-Steal! Contrastive Stealing Attacks Against Image Encoders/d755a24e-203a-45cb-a84e-020f7fb471b0_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..e8132c58747433bd2a7cb47aa816a25e32cc93c3 --- /dev/null +++ b/2023/Can't Steal_ Cont-Steal! Contrastive Stealing Attacks Against Image Encoders/d755a24e-203a-45cb-a84e-020f7fb471b0_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0641cde1d6c0dbdfd1fbc7bf73548bcbc47b5cd2d7f09f2fa25d1bfeb7b6b649 +size 2545408 diff --git a/2023/Can't Steal_ Cont-Steal! Contrastive Stealing Attacks Against Image Encoders/full.md b/2023/Can't Steal_ Cont-Steal! Contrastive Stealing Attacks Against Image Encoders/full.md new file mode 100644 index 0000000000000000000000000000000000000000..055724b7ab83a5ea98d285ba0fbc243a3af96871 --- /dev/null +++ b/2023/Can't Steal_ Cont-Steal! Contrastive Stealing Attacks Against Image Encoders/full.md @@ -0,0 +1,347 @@ +# Can't Steal? Cont-Steal! Contrastive Stealing Attacks Against Image Encoders + +Zeyang Sha† Xinlei He† Ning Yu‡ Michael Backes† Yang Zhang† + +†CISPA Helmholtz Center for Information Security ‡Salesforce Research {zeyang.sha, xinlei.he, director, zhang}@cispa.de, ning.yu@salesforce.com + +# Abstract + +Self-supervised representation learning techniques have been developing rapidly to make full use of unlabeled images. They encode images into rich features that are oblivious to downstream tasks. Behind their revolutionary representation power, the requirements for dedicated model designs and a massive amount of computation resources expose image encoders to the risks of potential model stealing attacks - a cheap way to mimic the well-trained encoder performance while circumventing the demanding requirements. Yet conventional attacks only target supervised classifiers given their predicted labels and/or posteriors, which leaves the vulnerability of unsupervised encoders unexplored. + +In this paper, we first instantiate the conventional stealing attacks against encoders and demonstrate their severer vulnerability compared with downstream classifiers. To better leverage the rich representation of encoders, we further propose Cont-Steal, a contrastive-learning-based attack, and validate its improved stealing effectiveness in various experiment settings. As a takeaway, we appeal to our community's attention to the intellectual property protection of representation learning techniques, especially to the defenses against encoder stealing attacks like ours. + +# 1. Introduction + +Recent years have witnessed the great success of applying deep learning (DL) to computer vision tasks. Different from supervised DL models, self-supervised learning which transforms unlabeled data samples into rich representations, has gained more and more popularity. + +Behind its powerful representation, it is non-trivial to obtain a state-of-the-art image encoder. For instance, SimCLR [6] uses 128 TPU v3 cores to pre-train a ResNet- + +![](images/e59fe95481358d05a780af7dd1b54c6d4f84ec72885d0eb1614d4d8fb718b935.jpg) +Figure 1. Model stealing attacks against classifiers (previous) v.s. model stealing attacks against encoders (ours). Previous works aim to steal a whole classifier using the predicted label or posteriors of a target model. In our work, we aim to steal the target encoder using its embeddings. The target encoder $(E_{t})$ is pre-trained and fixed, as shown in the solid frame. The surrogate encoder $(E_{s})$ is trainable by the adversary, as shown in the dashed frame. + +50 encoder with a batch size of 4096. Therefore, many big companies provide cloud-based self-supervised learning encoder services for users. For instance, Cohere, $^{2}$ OpenAI, $^{3}$ and Clarifai $^{4}$ provide the embedding API of images and texts for commercial usage. There are many works [10, 26, 30] exploring security issues of encoder-based API. Therefore, it is a very important and urgent problem. + +These kinds of service leave the possibility of model stealing attacks [5, 25, 28, 37, 44, 48, 50, 51]. In these attacks, the adversary aims to steal the parameters or functionalities of target models with only query access to them. A successful model stealing attack does not only threaten + +the intellectual property of the target model, but also serves as a stepping stone for further attacks such as adversarial examples [3, 4, 16, 38, 47, 52], backdoor attacks [7, 26, 41, 43], and membership inference attacks [22-24, 30, 33, 34, 40, 42, 45, 46]. So far, model stealing attacks concentrate on the supervised classifiers, i.e., the model responses are prediction posteriors or labels for a specific downstream task. The vulnerability of unsupervised image encoders is unfortunately unexplored. + +Our Work. To fill this gap, we pioneer the systematic investigation of model stealing attacks against image encoders. In this work, the adversary's goal is to steal the functionalities of the target model. See Figure 1 for an overview and a comparison with previous works. More specifically, we focus on encoders trained by contrastive learning, which is one of the most cutting-edge unsupervised representation learning strategies that unleash the information of unlabeled data. + +We first instantiate the conventional stealing attacks against encoders and expose their vulnerability. Given an input image, the target encoder outputs its representation (referred to as embedding). Similar to model stealing attacks against classifiers, we consider the embedding as the "ground truth" label to guide the training procedure of a surrogate encoder on the adversary side. To measure the effectiveness of stealing attacks, we train an extra linear layer for the target and surrogate encoders towards the same downstream classification task. Preferably, the surrogate model should achieve both high classification accuracy and high agreement with the target predictions. + +We evaluate our attacks on five datasets against four contrastive learning encoders. Our results demonstrate that the conventional attacks are more effective against encoders than against downstream classifiers. For instance, when we steal the downstream classifier pre-trained by SimCLR on CIFAR10 (with posteriors as its responses) using STL10 as the surrogate dataset, the adversary can only achieve an accuracy of 0.359. The accuracy, however, increases to 0.500 instead when we steal its encoder (with the embedding as its response). + +Despite its encouraging performance, conventional attacks are not the most suitable ones against encoders. This is because they treat each image-embedding pair individually without interacting across pairs. Different embeddings are beneficial to each other as they can serve as anchors to better locate the position of the other embeddings in their space. Contrastive learning [6, 8, 17, 20, 27, 49, 53] is a straightforward idea to achieve this goal. It is formulated to enforce the embeddings of different augmentations of the same images closer and those of different images further. + +In a similar spirit, we propose Cont-Steal, a contrastive-learning-based model stealing attack against the encoder. The goal of Cont-Steal is to enforce the surrogate embed + +ding of an image close to its target embedding (defined as a positive pair) and also push away embeddings of different images irrespective of being generated by the target or the surrogate encoders (defined as negative pairs). + +The comprehensive evaluation shows that Cont-Steal outperforms the conventional model stealing attacks to a large extent. For instance, when CIFAR10 is the target dataset, Cont-Steal achieves an accuracy of 0.714 on the SimCLR encoder pretrained on CIFAR10 with the surrogate dataset and downstream dataset being STL10, while the conventional attack only achieves 0.457 accuracy. Also, Cont-Steal is more query-efficient and dataset-independent (see Figure 9 for more details). This is because Cont-Steal leverages higher-order information across samples to mimic the functionality of the target encoder. To mitigate the attacks, we evaluate different defense mechanisms including noise, top- $k$ , rounding, and watermark. Our evaluations show that in most cases, these mechanisms cannot effectively defend against Cont-Steal. Among them, top- $k$ can reduce the attack performance to the largest extent. However, it also strongly limits the target model's utility. + +As a takeaway, our attack further exposes the severe vulnerability of pre-trained encoders. We appeal to our community's attention to the intellectual property protection of representation learning techniques, especially to the defenses against encoder stealing attacks like ours. + +# 2. Threat Model + +In this work, for the encoder pre-trained with images, we consider image classification as the downstream task. We refer to the encoder as the target encoder. Then we treat both the encoder and the linear layer trained for the downstream task together as the target model. We first introduce the adversary's goal and then characterize different background knowledge that the adversary might have. + +Adversary's Goal. Following previous work [25, 28, 44], we taxonomize the adversary's goal into two dimensions, i.e., theft and utility. The theft adversary aims to build a surrogate encoder that has similar performance on the downstream tasks as the target encoder. Different from the thief adversary, the goal of the utility adversary is to construct a surrogate encoder that behaves normally on different downstream tasks. In this case, the surrogate encoder not only faithfully "copies" the behaviors of the target encoder, but also serves as a stepping stone to conduct other attacks. + +Adversary's Background Knowledge. We categorize the adversary's background knowledge into two dimensions, i.e., the knowledge of the target encoder and the distribution of the surrogate dataset. + +Regarding knowledge of the target encoder, we assume that the adversary only has black-box access to it, which means that they can only query the target encoder with an input image and obtain the corresponding output, i.e., the + +embedding of the input image. + +Regarding the surrogate dataset that is used to train the surrogate encoder, we consider two cases. First, we assume the adversary has the same training dataset as the target encoder. However, such an assumption may be hard to achieve as such datasets are usually private and protected by the model owner. In a more extreme case, we assume that the adversary has totally no information about the target encoder's training dataset, which means that they can only use a different distribution dataset to conduct the model stealing attacks. We later show that the adversary can still launch effective model stealing attacks against the target encoder given a surrogate dataset that is distributed differently compared to the target dataset. + +For the model architecture that is used to train the surrogate encoder, we consider two cases. First, we assume the adversary is aware of the target encoder's architecture and can train the same architecture surrogate encoder. Then we relax our assumption that the adversary uses different architectures to train the surrogate encoder. Our evaluation shows that the choice of architecture does not have much impact on the attack performance (see Table 3), which makes the attack more realistic. + +Note that we also compare our attacks against the encoders to the traditional model stealing attacks that focus on the whole classifier (which has an encoder and a linear layer). If the attack targets a whole classifier, we assume the adversary may obtain the posteriors or the predicted label for an input image. + +# 3. Model Stealing Attacks + +In this section, we first describe the conventional attacks against the encoders. Then, we propose a novel contrastive stealing framework, Cont-Steal, to steal the encoders more effectively. + +# 3.1. Conventional Attacks Against Encoders + +The adversary takes two steps to conduct the model stealing attacks against the target encoder and one step for further evaluation. + +Obtain the Surrogate Dataset. To conduct model stealing attacks, the adversary first needs to obtain a surrogate dataset. Based on the knowledge of the target classifier's training dataset (target dataset), we consider two cases. If the adversary has full knowledge of the target dataset, they can directly leverage the target dataset itself as the surrogate dataset. Or the adversary has no knowledge of the target dataset, which means that they can only construct the surrogate dataset, which is distributed differently from the target dataset. + +Train the Surrogate Encoder. Slightly different from the classifier, the response of the encoder is an embedding, which is a feature vector. In this case, the adversary can still + +leverage a similar loss function to optimize the surrogate encoder, which can be defined as follows: + +$$ +L _ {M S} = \sum_ {k = 1} ^ {N} l \left(E _ {T} \left(x _ {k}\right), E _ {S} \left(x _ {k}\right)\right) \tag {1} +$$ + +where $E_{T}(\cdot) / E_{S}(\cdot)$ is the target/surrogate encoder, $N$ is the total number of samples on the surrogate dataset, and $l(\cdot)$ is the MSE loss. + +Apply the Surrogate Encoder to Downstream Tasks. To evaluate the effectiveness of model stealing attacks against the encoder, the adversary can leverage the same downstream task to both the target and surrogate encoders. Concretely, the adversary trains an extra linear layer for the target and surrogate encoders, respectively. Note that we refer to the target/surrogate encoder and the extra linear layer as the target/surrogate classifiers. Then, the adversary quantifies the attack effectiveness by measuring the performance of the target/surrogate classifier on the downstream tasks. + +# 3.2. Cont-Steal Attacks Against Encoders + +To better leverage the rich information from the embeddings, we propose Cont-Steal, a contrastive learning-based model stealing attacks against encoders, which leverages contrastive learning to enhance the stealing performance. Concretely, Cont-Steal aims to enforce the surrogate embedding of an image to get close to its target embedding (defined as a positive pair), and also push away embeddings of different images regardless of being generated by the target or the surrogate encoders (defined as negative pairs). There are three steps for the adversary to conduct contrastive stealing attacks against encoders and one step for further evaluation. + +Obtain the Surrogate Dataset. The adversary follows the same strategy as Section 3.1 to obtain the surrogate dataset. Data Augmentation. Our proposed Cont-Steal leverages data augmentation to transform an input image into its two augmented views. In this paper, we leverage RandAugment [11] as the augmentation method, which is made up of a group of advanced augmentation operations. Concretely, we set $n = 2$ and $m = 14$ following Cubuk et al. [11] where $n$ denotes the number of transformations to a given sample and $m$ represents the magnitude of global distortion. + +Train the Surrogate Encoder. Instead of querying the encoders with the original images, the adversary queries the encoders with the augmented views of them. Concretely, for an input image $x_{i}$ , we generate two augmented views of it, i.e., $\widetilde{x}_{i,s}$ and $\widetilde{x}_{i,t}$ , where $\widetilde{x}_{i,s} / \widetilde{x}_{i,t}$ is used to query the surrogate/target encoder. We consider $(\widetilde{x}_{i,s},\widetilde{x}_{j,t})$ as a positive pair if $i = j$ , and otherwise a negative pair. + +Given a mini-batch of $N$ samples, we generate $N$ augmented views as the input of the target encoder and another $N$ augmented views as the input of the surrogate encoders. + +Concretely, the loss of Cont-Steal can be formulated as follows: + +$$ +D _ {\text {e n c o d e r}} ^ {+} (i) = \exp \left(\operatorname {s i m} \left(E _ {S} \left(\tilde {x} _ {i, s}\right), E _ {T} \left(\tilde {x} _ {i, t}\right)\right) / \tau\right), \tag {2} +$$ + +$$ +D _ {\text {e n c o d e r}} ^ {-} (i) = \sum_ {k = 1} ^ {N} \left(\exp \left(\operatorname {s i m} \left(E _ {S} \left(\widetilde {x} _ {i, s}\right), E _ {T} \left(\widetilde {x} _ {k, t}\right)\right) / \tau\right)\right), \tag {3} +$$ + +$$ +D _ {s e l f} ^ {-} (i) = \sum_ {k = 1} ^ {N} \mathbb {1} _ {[ k \neq i ]} \left(\exp \left(\operatorname {s i m} \left(E _ {S} \left(\widetilde {x} _ {i, s}\right), E _ {S} \left(\widetilde {x} _ {k, s}\right)\right) / \tau\right)\right), \tag {4} +$$ + +$$ +l (i) = - \log \frac {D _ {\text {e n c o d e r}} ^ {+} (i)}{D _ {\text {e n c o d e r}} ^ {-} (i) + D _ {\text {s e l f}} ^ {-} (i)}, \tag {5} +$$ + +$$ +L _ {\text {C o n t - S t e a l}} = \frac {\sum_ {k = 1} ^ {N} l (k)}{N}, \tag {6} +$$ + +where $E_{S}(\cdot)$ and $E_{T}(\cdot)$ denotes the surrogate and target encoder, $sim(u,v) = u^{T}v / ||u||||v||$ represents the cosine similarity between $u$ and $v$ , and $\tau$ is parameter to control the temperature. + +As illustrated in Figure 2, the conventional attack treats each embedding individually without interacting across pairs. However, different embeddings are beneficial to each other as they can serve as anchors to better locate the position of the other embeddings in their space. Cont-Steal maximizes the similarity of embeddings generated from the target and surrogate encoders for a positive pair $(\widetilde{x}_{i,s},\widetilde{x}_{i,t})$ (orange arrows in Figure 2). For the embedding generated from the target and surrogate encoders for any pair $(\widetilde{x}_{i,s},\widetilde{x}_{j,t})$ , contrastive stealing aims to make them more distant (green arrows in Figure 2). Besides, as pointed out by Chen et al. [6], contrastive learning benefits larger negative samples. To achieve this goal, we also consider the embeddings generated from the surrogate encoder for augmented views of different images, i.e., $(\widetilde{x}_{i,s},\widetilde{x}_{j,s})$ , as negative pairs minimize their similarity (blue arrows in Figure 2). We later show that such design can enhance the performance of contrastive stealing (see Table 5). + +Apply the Surrogate Encoder to Downstream Tasks. We follow Section 3.1 to evaluate the effectiveness of model stealing on downstream tasks. + +# 4. Experiments + +In this section, we first describe the experimental setup in Section 4.1. Then we show the performance of the target encoders on the downstream tasks. Next, we summarize the performance of conventional attacks against classifiers and encoders in Section 4.2. Lastly, we evaluate the performance of Cont-Steal and conduct ablation studies to demonstrate its effectiveness under different settings in Section 4.3. + +![](images/85b287c587173ea3c5b75df0fadd4f0b60bc30a80663ed68b901d2041a071eeb.jpg) + +![](images/29b7b7acc13ea667777bc9565837d2ce09ec9557351c5fd6bb3170d8f9e0b789.jpg) +Figure 2. Conventional attack (top) vs. Cont-Steal (bottom) against encoders. Conventional attack applies MSE loss to approximate target embeddings for each sample individually. Cont-Steal (bottom) introduces data augmentation and interacts across multiple samples: associating target/surrogate embeddings of the same images closer and repulsing those of different images farther away. The target encoder $(E_{t})$ is pre-trained and fixed, as shown in the solid frame. The surrogate encoder $(E_{s})$ is trainable by the adversary, as shown in the dashed frame. + +# 4.1. Experimental Setup + +Our encoders are pre-trained on CIFAR10 [1], and ImageNet [12]. We use four different kinds of contrastive methods: SimCLR [6], MoCo [20], BYOL [17] and SimSiam [8] to train a ResNet18 [21] as our target encoders. Our implementation is based on a PyTorch framework of contrastive learning. Then, these well pre-trained encoders will be applied to train downstream classifiers on CIFAR10 [1], STL10 [9], Fashion-MNIST [54], and SVHN [35]. In the experiments in the model stealing section, we use CIFAR10, STL10, Fashion-MNIST, and SVHN to conduct the attack. + +Agreement and accuracy are used as metrics to evaluate the model stealing attack's performance. The agreement will evaluate the similarity of surrogate encoders and target encoders in downstream tasks. The accuracy will evaluate + +![](images/4119feca9e7f1245c2f3a2a48efecc701d744f55cc9395ac73faa4bba0ec0e13.jpg) +(a) CIFAR10 + +![](images/3afaced75a678288771aa36a8c03ae19f56260e927b5193ca45e454eac96b500.jpg) +(b) ImageNet100 +Figure 3. The performance of target classifiers composed by target encoder and an extra linear layer. The encoders are pre-trained on CIFAR10 (a) and ImageNet100 (b). The x-axis represents different downstream datasets for the target encoder and classifier. The y-axis represents the target model's accuracy on downstream tasks. + +the utility of surrogate models on downstream tasks. For each metric, a larger value is more desirable. During the stealing process, we set the batch size as 128 and the learning rate as 0.001. We show more results on the impact of hyperparameters in Supplementary Material in Section A.2 + +# 4.2. Performance of Conventional Attacks + +We first show the target encoder's performance in various downstream tasks. The results are summarized in Figure 3. We conduct our experiments to explore whether the encoders are more vulnerable to model stealing attacks. We show our results of target encoders and downstream classifiers both trained on CIFAR10 in Figure 4. In all cases, the adversary can get better attack performance by stealing encoders rather than classifiers. This gap becomes especially apparent when the adversary has absolutely no knowledge of the train data. This is because the rich information in embeddings can better facilitate the learning process of surrogate encoders. For instance, when the surrogate dataset is CIFAR10 (the same as the target downstream dataset), stealing SimCLR's embeddings can achieve 0.785 agreement, while stealing predicted labels can achieve 0.712 agreement. However, when the surrogate dataset is totally different from the downstream target dataset, e.g., SVHN, stealing embeddings from SimCLR can still achieve 0.507 agreement while the agreement of stealing predicted labels drops to 0.192. We show more results in Supplementary Material Section A.4 due to the page limitation. + +We also find that all model stealing attacks' accuracy and agreement are highly correlated. As shown in Figure 5, the agreement is highly correlated with the accuracy. This indicates that besides accuracy, the agreement can also be used as a metric to evaluate the performance of model stealing attacks. We show the result on Figure 5. It can be obviously seen that agreement is highly related to accuracy. We use the linear regression method to describe the relationship between agreement and accuracy and find that the relation function is $y = 0.940 * x$ . + +# 4.3. Performance of Cont-Steal + +As shown in Section 4.2, encoders are more vulnerable to model stealing attacks since the embedding usually contains richer information compared to the predicted label or posteriors. We then show that our proposed Cont-Steal can achieve better attack performance by making deeper use of embeddings' information. + +Figure 7 shows the attack performance when the target pre-training dataset is CIFAR10. Note that we also show the attack performance on other settings in the appendix. We discover that compared to conventional attacks against encoders, Cont-Steal can consistently achieve better performance. For instance, as shown in Figure 7d, when the target encoder is MoCo trained on CIFAR10, if the adversary uses STL10 to conduct model stealing attacks against encoders, the surrogate encoder can achieve 0.841 agreement in CIFAR10 downstream tasks with the Cont-Steal but only 0.479 with conventional attacks. Another finding is that compared to the same distribution surrogate dataset, our Cont-Steal can better enhance the performance when the surrogate dataset comes from a different distribution from the pre-trained dataset. For instance, when the target encoder is SimCLR trained on CIFAR10, Cont-Steal outperforms conventional attack by 0.055 agreement when the surrogate dataset is also CIFAR10, while the improvement increases to 0.207 and 0.214 when the surrogate dataset is STL10. We show more comparing results in Supplementary Material Section A.5. Note that our Cont-Steal also has great performance on other recent state-of-the-art visual models (ViT [13], MAE [19], and CLIP [39]), as we show in Section A.6, and can have better performance than other recent similar attacks [14, 31] shown in Section A.7. + +To better understand why Cont-Steal can always achieve better performance, we extract samples' embeddings generated by different encoders, i.e., the target encoder, surrogate encoder trained with the conventional attack, and surrogate encoder trained with the Cont-Steal, and project them into a 2-dimensional space using t-SNE. From the results summarized in Figure 6, we find that Cont-Steal can effectively mimic the pattern of the embeddings as the target encoder. However, the conventional attack fails to capture such patterns for a number of input samples, e.g., the outer circle in Figure 8c. This further demonstrates that Cont-Steal benefits from jointly considering different embeddings as they can serve as anchors to better locate the position of the other embeddings in their space. We also show some ablation study results on Supplementary Material Section A.2 to show that with the less surrogate dataset, less training epoch, and different model architecture, Cont-Steal can still achieve much better results than conventional steal. Also, we show further attacks based on the stole models on Supplementary Material Section A.3 to show that Cont-Steal can be used as a springboard for other attacks. + +![](images/7ca75274d1255e5a97d83fda6695ddbb645d584232f66126490b614d719f6ff4.jpg) +(a) CIFAR10 + +![](images/3acc1371b453f6896ef3db12321196b9ba5addf89f3678e5148745bf8016db6d.jpg) +(b) STL10 + +![](images/267e0b000c3a365c69ef675872af6ba0fea1e208686e7cfba7bb92933609cfa4.jpg) +(c) F-MNIST + +![](images/463bd51baf3228ac7f90595a7c4dd970088c74976ae63db34d790c86c07cfc9b.jpg) +(d) SVHN + +![](images/0df6cd4d210bf03d91d88af64c3b39213f8eea0c76feb7ad4ea94a44ebaf9ec0.jpg) +(e) CIFAR10 + +![](images/c1cce1e7eaa711578e2ba8ee28d4231a3763ee7220dca3ede6f97bb3ba896016.jpg) +(f) STL10 + +![](images/22dc3e19a6e284b471bffedb4704a1a7a3d41a0ac3c33a46a37f572e7acc5c40.jpg) +(g) F-MNIST + +![](images/5e8a4c3ece83db9a5eac59c9ab04268d4eb68b804a1d748c96459c978291c115.jpg) +(h) SVHN + +![](images/43b3f225b8da1930bb35212ee3efe3dd52b6137dc188db8b02133b94236864a9.jpg) +Figure 4. The performance of model stealing attack against target encoders and downstream classifiers both trained on CIFAR10. Target models can output predicted labels, posteriors, or embeddings. The adversary uses CIFAR10, STL10, Fashion-MNIST (F-MNIST), and SVHN to conduct model stealing attacks. The x-axis represents different kinds of target models. The first line's y-axis represents the agreement of the model stealing attack. The second line's y-axis represents the accuracy of the model stealing attack. +Figure 5. The relationship between accuracy and agreement. The x-axis is the agreement number, and the y-axis is the accuracy number. + +![](images/fa618ec1342d2131ad747c287860fbedac41a52941a8f65e478920d3f2711474.jpg) +(a) Target encoder +Figure 6. The t-SNE projection of 1,000 randomly selected samples' embeddings from target encoder, surrogate encoder under Cont-Steal, and surrogate encoder under the conventional attack, respectively. Note that the target encoder is pre-trained by SimCLR on CIFAR10. + +![](images/ce6d0ad655b75f8a0b188891b8ceb1d396043bf17d98bf60b9754f5a57f1499f.jpg) +(b) Cont-Steal + +![](images/09513a45e819a8fd6f954414b2106994d88fa140052013dcfc2aeb979f8ac24c.jpg) +(c) Conventional + +Table 1. The monetary and (training) time costs for normal training and Cont-Steal attack. Cont-Steal's monetary cost contains two parts: query cost and training cost. Note that we ignore the query time cost of Cont-Steal as it normally has a smaller value than the training time cost. + +
Monetary CostTime Cost
ModelNormal ($)Cont-Steal ($)Normal (h)Cont-Steal (h)
SimCLR58.6811.83 (1.83 + 10)20.010.62
MoCo54.8312.13 (2.13 + 10)18.690.73
BYOL61.4612.08 (2.08 + 10)20.960.71
SimSiam57.1412.00 (2.00 + 10)19.460.68
+ +# 4.4. Cost Analysis + +As we mentioned before, pre-train a state-of-the-art encoder is time-consuming and resource-demanding. We wonder if the model stealing attacks can steal the functionality of the encoder with much less cost. To this end, we evaluate the time and monetary cost of training an encoder from scratch or stealing a pre-trained encoder via ContSteal. The monetary cost of model stealing includes querying the target model and training the surrogate model. We refer to the query price as $1 for 1,000 queries based on AWS.6 Our experiment is conducted on 1 NVIDIA A100 whose price is$ 2.934 per hour based on google cloud.7 + +The monetary and time cost is shown in Table 1. We observe that Cont-Steal can obtain a surrogate encoder with much less money and time cost than training the encoder from scratch. For instance, a ResNet18 trained by SimCLR + +![](images/50bd10a636c3765a64687310ebe792f6a0595a0688a8683cb76ab2948a712c2c.jpg) +(a) CIFAR10 + +![](images/7291b6de743fed6c4c3e879a0b1b9dd47616b89fb13aea5bb56545ce7bd0c926.jpg) +(b) STL10 + +![](images/961afd3798f5ba4a144d9541282cff8b14ca0de8859b78790ad87c5977a8b80f.jpg) +(c) F-MNIST + +![](images/02426e26ce9d61527cae74aef17b4b1edca4a13464e0f4ed7e49c887ac2e037e.jpg) + +![](images/a7f645d73ca807d125679a62b078825367274826362cdafcdf9bfc9e8d61dc65.jpg) +(e) CIFAR10 + +![](images/c796704c9354237234f9d12bdb37fc6b30d18e6feaaffc07e637417257d738bd.jpg) +(f) STL10 + +![](images/c1af7f3f0ddd109c57a48bd35c17caaffd93df608c6633e0aa61bfe9e802aacd.jpg) +(g) F-MNIST + +![](images/fae52eea7c07e6a312dfd9279f7fd05195a2960cad8fb6e264c34d499d823721.jpg) +(d) SVHN +(h) SVHN + +![](images/f470e76d0b9bd37680a0c9f53f2d7fa1faa050910bbe4869d6609abc624d6ae7.jpg) +Figure 7. The performance of Cont-Steal and conventional attack against target encoders trained on CIFAR10. The adversary uses CIFAR10, STL10, F-MNIST, and SVHN to conduct model stealing attacks. The adversary uses CIFAR10 as the downstream task to evaluate the attack performance. The x-axis represents different kinds of the target model. The first line's y-axis represents the agreement of the model stealing attack. The second line's y-axis represents the accuracy of the model stealing attack. +(a) Adding noise + +![](images/8b16877a46b1522e38dcade88684c045d8ef410010ca0bae3058eb925722b79c.jpg) +(b) Top- $k$ +Figure 8. The performance of different defense methods. Target encoders are trained on CIFAR10. The downstream dataset and surrogate dataset are both STL10. The x-axis represents different defense levels. The y-axis represents the model's accuracy. + +![](images/26ef07a82ee8c537eced95175b4f4d4475abb0193e362340d7032c969dd634fb.jpg) +(c) Rounding + +on CIFAR10 takes 20.01 hours and $58.68$ 's on 1 NVIDIA A100 GPU, while Cont-Steal only takes 0.62 hours and $11.83$ 's to steal an encoder that performs similarly on downstream tasks. The results demonstrate that Cont-Steal is able to construct surrogate encoders that perform similarly to the target encoders but with much less time and monetary cost. + +# 4.5. Defenses + +In this section, we will consider different defenses against model stealing attacks on encoders to evaluate the robustness of our proposed attack. We divided all defenses into three categories: perturbation-based defense [37] and watermark-based defense [2]. + +Perturbation-based Defense. In this defense setting, the defender aims to perturb the output of the target model to limit the information the adversary can obtain. The common practice of this kind of defense includes adding noise [37], top- $k$ [37], and feature rounding [48]. + +Adding noise means that the defender will introduce noise value to the original output of the model. In our case, + +we consider adding Gaussian noise to the embeddings generated by the target encoder. We set the mean value to 0, and different noise levels represent different standard deviations of the Gaussian distribution. For Top- $k$ , the defender will only output the first $k$ largest number of each embedding (and set the rest as 0). In this way, the high-dimensional information of the image contained in embeddings can be appropriately reduced. Regarding feature rounding, the defender will truncate the values in the embedding to a specific digit. As a case study, we consider a ResNet18 encoder pre-trained on CIFAR10 with SimCLR and take STL10 to train its downstream classifier. The experimental results are summarized in Figure 8. We can observe that while adding noise and top- $k$ can reduce the model stealing attacks' performance, it may also degrade the target model performance to a large extent. For instance, when the noise increases from 0 to 10, the attack performance of ContSteal decreases from 0.729 to 0.410, while the target encoder's performance drops from 0.734 to 0.098. On the other hand, rounding only has a limited effect on both target model performance and attack performance. This indicates that perturbation-based defense cannot defend against the encoder's model stealing attack effectively since they cannot reach a good trade-off between attack performance and model utility. + +Watermark-based Defense. Watermark-based defense is also one of the most popular defense methods against model stealing attacks [2]. Watermark provides copyright protection by adding some specific identification to the target model. If the surrogate model is stolen from the watermarked target model, then ideally, it will contain the same watermark as well. Adi et al. [2] show that back + +Table 2. Watermark defense. Pretrain dataset and surrogate dataset are both CIFAR10. Watermark leverages a watermark rate (wr) to verify the ownership of target models. A higher wr denotes a better verification performance. + +
DatasetTarget model (acc/wr)Cont-Steal (acc/wr)Baseline (acc/wr)
CIFAR100.864 / 0.9980.769 / 0.1300.871 / 0.095
STL100.721 / 0.9990.702 / 0.0340.733 / 0.111
SVHN0.501 / 0.9990.535 / 0.3030.492 / 0.103
F-MNIST0.857 / 0.9990.813 / 0.0610.850 / 0.099
+ +door technology can be used as the watermark to protect the model. In that sense, BadEncoder [26], a backdoor mechanism against the encoder, can be leveraged as a watermarking technology for our target encoder as well. The defenders first will train the watermarked (backdoored) encoder, where images with a certain trigger will cause misclassification. Then, if they find the surrogate model can also misclassify images with the same trigger, the defenders can claim ownership of the surrogate model. + +In our experiments, we leverage BadEncoder to watermark the encoder pre-trained on CIFAR10 by SimCLR, and leverage different downstream datasets to perform different tasks. We assume a strong adversary that has the same downstream dataset as the surrogate dataset. Also, we consider the baseline cases where the trigger samples are fed into the clean model to calculate the watermark rate (wr). As shown in Table 2), the watermark cannot be preserved as the surrogate models constructed by Cont-Steal have similar wr as the baseline model. For instance, when the downstream dataset is CIFAR10, Cont-Steal builds a surrogate model with 0.769 accuracy while only 0.130 wr, which is close to the baseline model. This indicates that Cont-Steal can bypass the watermarking technique as it can reach similar utility while reducing the wr to a large extent. Note that there is also another work to protect contrastive learning models from model stealing attacks using dataset inference [15]. We show in Supplementary Material in Section A.8 that this kind of defense can be easily bypassed by Cont-Steal. + +# 5. Related Work + +Contrastive Learning. Contrastive learning is one of the most popular methods to train encoders. Current works [6, 8, 17, 20, 49, 53] propose different advanced contrastive learning algorithms. SimCLR, MoCo, BYOL, SimSiam are currently the mainstream frameworks of contrastive learning. Thus, we concentrate on them in this paper. There are many works on evaluating the security and privacy risks of contrastive learning. Previous works [23, 26, 30] propose membership inference attacks, attribution inference attacks, and backdoor attacks on contrastive learning. All proposed attacks show that contrastive-based models are vulnerable + +to popular attacks. Therefore, the security issues of self-supervised learning deserve more attention. + +Model Stealing Attack. In model stealing, the adversary's goal is to steal part of the target model. Tramér et al. [48] proposed the first model stealing attack against black-box machine learning API to steal its parameters. Wang et al. [50] proposed the first hyperparameter stealing attacks against ML models. Oh et al. [36] also tried to steal machine learning model's architectures and hyperparameters. Orekondy et al. [37] proposed knockoff nets, which aim at stealing the functionality of black-box models. Krishna et al. [28] formulated a model stealing attack against BERT-based API. Besides, Wu et al. [51] and Shen et al. [44] perform model stealing attacks against Graph Neural Networks. These works often have relatively strong assumptions, such as the model family is known and the victim's data is partly available while we conduct model stealing attacks against encoders and relax the above assumption. + +# 6. Conclusion + +In this paper, we conduct the first model stealing risk assessment towards image encoders. Our evaluation shows that the encoder is more vulnerable to model stealing attacks compared to the classifier. This is because the embedding provided by the encoder contains richer information than the posteriors or predicted labels from whole classifiers. + +To better unleash the power from the embeddings, we propose Cont-Steal, a contrastive learning-based model stealing method against encoders. Concretely, Cont-Steal introduces different types of negative pairs as "anchors" to better navigate the surrogate encoder and learn the functionality of the target encoder. Extensive evaluations show that Cont-Steal consistently performs better than conventional attacks against encoders. And such an advantage is further amplified when the adversary has no information on the target dataset, a limited amount of data, and restricted query budgets. Our work points out that the threat of model stealing attacks against encoders is largely underestimated, which prompts the need for more effective intellectual property protection of representation learning techniques. + +Acknowledgments. We thank all anonymous reviewers for their constructive comments. This work is partially funded by the Helmholtz Association within the project "Trustworthy Federated Data Analytics" (TFDA) (funding number ZT-I-OO1 4) and by the European Health and Digital Executive Agency (HADEA) within the project "Understanding the individual host response against Hepatitis D Virus to develop a personalized approach for the management of hepatitis D" (D-Solve) (grant agreement number 101057917). + +# References + +[1] https://www.cs.toronto.edu/~kriz/cifar.html. +[2] Yossi Adi, Carsten Baum, Moustapha Cisse, Benny Pinkas, and Joseph Keshet. Turning Your Weakness Into a Strength: Watermarking Deep Neural Networks by Backdoorsing. In USENIX Security Symposium (USENIX Security), pages 1615-1631. USENIX, 2018. +[3] Battista Biggio, Igino Corona, Davide Maiorca, Blaine Nelson, Nedim Srndic, Pavel Laskov, Giorgio Giacinto, and Fabio Roli. Evasion Attacks against Machine Learning at Test Time. In European Conference on Machine Learning and Principles and Practice of Knowledge Discovery in Databases (ECML/PKDD), pages 387-402. Springer, 2013. +[4] Nicholas Carlini and David Wagner. Towards Evaluating the Robustness of Neural Networks. In IEEE Symposium on Security and Privacy (S&P), pages 39-57. IEEE, 2017. +[5] Varun Chandrasekaran, Kamalika Chaudhuri, Irene Giacomelli, Somesh Jha, and Songbai Yan. Model Extraction and Active Learning. CoRR abs/1811.02054, 2018. +[6] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey E. Hinton. A Simple Framework for Contrastive Learning of Visual Representations. In International Conference on Machine Learning (ICML), pages 1597-1607. PMLR, 2020. +[7] Xiaoyi Chen, Ahmed Salem, Michael Backes, Shiqing Ma, Qingni Shen, Zhonghai Wu, and Yang Zhang. BadNL: Backdoor Attacks Against NLP Models with Semantic-preserving Improvements. In Annual Computer Security Applications Conference (ACSAC), pages 554-569. ACSAC, 2021. +[8] Xinlei Chen and Kaiming He. Exploring Simple Siamese Representation Learning. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 15750-15758. IEEE, 2021. +[9] Adam Coates, Andrew Y. Ng, and Honglak Lee. An Analysis of Single-Layer Networks in Unsupervised Feature Learning. In International Conference on Artificial Intelligence and Statistics (AISTATS), pages 215-223. JMLR, 2011. +[10] Tianshuo Cong, Xinlei He, and Yang Zhang. SSL-Guard: A Watermarking Scheme for Self-supervised + +Learning Pre-trained Encoders. In ACM SIGSAC Conference on Computer and Communications Security (CCS), pages 579-593. ACM, 2022. +[11] Ekin Dogus Cubuk, Barret Zoph, Jon Shlens, and Quoc Le. RandAugment: Practical Automated Data Augmentation with a Reduced Search Space. In Annual Conference on Neural Information Processing Systems (NeurIPS), pages 18613-18624. NeurIPS, 2020. +[12] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. ImageNet: A large-scale hierarchical image database. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 248–255. IEEE, 2009. +[13] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale. In International Conference on Learning Representations (ICLR), 2021. +[14] Adam Dziedzic, Nikita Dhawan, Muhammad Ahmad Kaleem, Jonas Guan, and Nicolas Papernot. On the Difficulty of Defending Self-Supervised Learning against Model Extraction. In International Conference on Machine Learning (ICML). JMLR, 2022. +[15] Adam Dziedzic, Haonan Duan, Muhammad Ahmad Kaleem, Nikita Dhawan, Jonas Guan, Yannis Cattan, Franziska Boenisch, and Nicolas Papernot. Dataset Inference for Self-Supervised Models. In Annual Conference on Neural Information Processing Systems (NeurIPS). NeurIPS, 2022. +[16] Ian Goodfellow, Jonathon Shlens, and Christian Szegedy. Explaining and Harnessing Adversarial Examples. In International Conference on Learning Representations (ICLR), 2015. +[17] Jean-Bastien Grill, Florian Strub, Florent Altché, Corentin Tallec, Pierre H. Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Ávila Pires, Zhaohan Guo, Mohammad Gheshlaghi Azar, Bilal Piot, Koray Kavukcuoglu, Rémi Munos, and Michal Valko. Bootstrap Your Own Latent - A New Approach to Self-Supervised Learning. In Annual Conference on Neural Information Processing Systems (NeurIPS). NeurIPS, 2020. +[18] Yunhui Guo, Honghui Shi, Abhishek Kumar, Kristen Grauman, Tajana Rosing, and Rogério Schmidt + +Feris. SpotTune: Transfer Learning Through Adaptive Fine-Tuning. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 4805-4814. IEEE, 2019. +[19] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dólár, and Ross B. Girshick. Masked Autoencoders Are Scalable Vision Learners. CoRR abs/2111.06377, 2021. +[20] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross B. Girshick. Momentum Contrast for Unsupervised Visual Representation Learning. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 9726-9735. IEEE, 2020. +[21] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep Residual Learning for Image Recognition. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 770-778. IEEE, 2016. +[22] Xinlei He, Rui Wen, Yixin Wu, Michael Backes, Yun Shen, and Yang Zhang. Node-Level Membership Inference Attacks Against Graph Neural Networks. CoRR abs/2102.05429, 2021. +[23] Xinlei He and Yang Zhang. Quantifying and Mitigating Privacy Risks of Contrastive Learning. In ACM SIGSAC Conference on Computer and Communications Security (CCS), pages 845-863. ACM, 2021. +[24] Bo Hui, Yuchen Yang, Haolin Yuan, Philippe Burlina, Neil Zhenqiang Gong, and Yinzhi Cao. Practical Blind Membership Inference Attack via Differential Comparisons. In Network and Distributed System Security Symposium (NDSS). Internet Society, 2021. +[25] Matthew Jagielski, Nicholas Carlini, David Berthelot, Alex Kurakin, and Nicolas Papernot. High Accuracy and High Fidelity Extraction of Neural Networks. In USENIX Security Symposium (USENIX Security), pages 1345-1362. USENIX, 2020. +[26] Jinyuan Jia, Yupei Liu, and Neil Zhenqiang Gong. BadEncoder: Backdoor Attacks to Pre-trained Encoders in Self-Supervised Learning. In IEEE Symposium on Security and Privacy (S&P). IEEE, 2022. +[27] Prannay Khosla, Piotr Teterwak, Chen Wang, Aaron Sarna, Yonglong Tian, Phillip Isola, Aaron Maschinot, Ce Liu, and Dilip Krishnan. Supervised Contrastive Learning. In Annual Conference on Neural Information Processing Systems (NeurIPS). NeurIPS, 2020. +[28] Kalpesh Krishna, Gaurav Singh Tomar, Ankur P. Parikh, Nicolas Papernot, and Mohit Iyyer. Thieves on Sesame Street! Model Extraction of BERT-based + +APIs. In International Conference on Learning Representations (ICLR), 2020. +[29] Alexey Kurakin, Ian Goodfellow, and Samy Bengio. Adversarial Examples in the Physical World. CoRR abs/1607.02533, 2016. +[30] Hongbin Liu, Jinyuan Jia, Wenjie Qu, and Neil Zhenqiang Gong. EncoderMI: Membership Inference against Pre-trained Encoders in Contrastive Learning. In ACM SIGSAC Conference on Computer and Communications Security (CCS). ACM, 2021. +[31] Yupei Liu, Jinyuan Jia, Hongbin Liu, and Neil Zhenqiang Gong. StolenEncoder: Stealing Pre-trained Encoders in Self-supervised Learning. In ACM SIGSAC Conference on Computer and Communications Security (CCS), pages 2115-212. ACM, 2022. +[32] Aleksander Madry, Aleksandar Makelov, Ludwig Schmidt, Dimitris Tsipras, and Adrian Vladu. Towards Deep Learning Models Resistant to Adversarial Attacks. In International Conference on Learning Representations (ICLR), 2018. +[33] Milad Nasr, Reza Shokri, and Amir Houmansadr. Machine Learning with Membership Privacy using Adversarial Regularization. In ACM SIGSAC Conference on Computer and Communications Security (CCS), pages 634-646. ACM, 2018. +[34] Milad Nasr, Reza Shokri, and Amir Houmansadr. Comprehensive Privacy Analysis of Deep Learning: Passive and Active White-box Inference Attacks against Centralized and Federated Learning. In IEEE Symposium on Security and Privacy (S&P), pages 1021-1035. IEEE, 2019. +[35] Yuval Netzer, Tao Wang, Adam Coates, Alessandro Bissacco, Bo Wu, and Andrew Y. Ng. Reading Digits in Natural Images with Unsupervised Feature Learning. In Annual Conference on Neural Information Processing Systems (NIPS). NIPS, 2011. +[36] Seong Joon Oh, Max Augustin, Bernt Schiele, and Mario Fritz. Towards Reverse-Engineering Black-Box Neural Networks. In International Conference on Learning Representations (ICLR), 2018. +[37] Tribhuvanesh Orekondy, Bernt Schiele, and Mario Fritz. Knockoff Nets: Stealing Functionality of Black-Box Models. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 4954–4963. IEEE, 2019. +[38] Nicolas Papernot, Patrick D. McDaniel, Somesh Jha, Matt Fredrikson, Z. Berkay Celik, and Ananthram + +Swami. The Limitations of Deep Learning in Adversarial Settings. In IEEE European Symposium on Security and Privacy (Euro S&P), pages 372-387. IEEE, 2016. +[39] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning Transferable Visual Models From Natural Language Supervision. In International Conference on Machine Learning (ICML), pages 8748-8763. PMLR, 2021. +[40] Alexandre Sablayrolles, Matthijs Douze, Cordelia Schmid, Yann Ollivier, and Hervé Jégou. White-box vs Black-box: Bayes Optimal Strategies for Membership Inference. In International Conference on Machine Learning (ICML), pages 5558-5567. PMLR, 2019. +[41] Ahmed Salem, Michael Backes, and Yang Zhang. Don't Trigger Me! A Triggerless Backdoor Attack Against Deep Neural Networks. CoRR abs/2010.03282, 2020. +[42] Ahmed Salem, Yang Zhang, Mathias Humbert, Pascal Berrang, Mario Fritz, and Michael Backes. MLLeaks: Model and Data Independent Membership Inference Attacks and Defenses on Machine Learning Models. In Network and Distributed System Security Symposium (NDSS). Internet Society, 2019. +[43] Zeyang Sha, Xinlei He, Pascal Berrang, Mathias Humbert, and Yang Zhang. Fine-Tuning Is All You Need to Mitigate Backdoor Attacks. CoRR abs/2212.09067, 2022. +[44] Yun Shen, Xinlei He, Yufei Han, and Yang Zhang. Model Stealing Attacks Against Inductive Graph Neural Networks. In IEEE Symposium on Security and Privacy (S&P), pages 1175-1192. IEEE, 2022. +[45] Reza Shokri, Marco Stronati, Congzheng Song, and Vitaly Shmatikov. Membership Inference Attacks Against Machine Learning Models. In IEEE Symposium on Security and Privacy (S&P), pages 3-18. IEEE, 2017. +[46] Liwei Song and Prateek Mittal. Systematic Evaluation of Privacy Risks of Machine Learning Models. In USENIX Security Symposium (USENIX Security). USENIX, 2021. +[47] Florian Tramér, Alexey Kurakin, Nicolas Papernot, Ian Goodfellow, Dan Boneh, and Patrick McDaniel. + +Ensemble Adversarial Training: Attacks and Defenses. In International Conference on Learning Representations (ICLR), 2017. +[48] Florian Tramér, Fan Zhang, Ari Juels, Michael K. Reiter, and Thomas Ristenpart. Stealing Machine Learning Models via Prediction APIs. In USENIX Security Symposium (USENIX Security), pages 601-618. USENIX, 2016. +[49] Aäron van den Oord, Yazhe Li, and Oriol Vinyals. Representation Learning with Contrastive Predictive Coding. CoRR abs/1807.03748, 2018. +[50] Binghui Wang and Neil Zhenqiang Gong. Stealing Hyperparameters in Machine Learning. In IEEE Symposium on Security and Privacy (S&P), pages 36-52. IEEE, 2018. +[51] Bang Wu, Xiangwen Yang, Shirui Pan, and Xingliang Yuan. Model Extraction Attacks on Graph Neural Networks: Taxonomy and Realization. CoRR abs/2010.12751, 2020. +[52] Huijun Wu, Chen Wang, Yuriy Tyshetskiy, Andrew Docherty, Kai Lu, and Liming Zhu. Adversarial Examples for Graph Data: Deep Insights into Attack and Defense. In International Joint Conferences on Artificial Intelligence (IJCAI), pages 4816-4823. IJCAI, 2019. +[53] Zhirong Wu, Yuanjun Xiong, Stella X. Yu, and Dahua Lin. Unsupervised Feature Learning via Non-Parametric Instance Discrimination. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 3733–3742. IEEE, 2018. +[54] Han Xiao, Kashif Rasul, and Roland Vollgraf. Fashion-MNIST: a Novel Image Dataset for Benchmarking Machine Learning Algorithms. CoRR abs/1708.07747, 2017. +[55] Ziqing Yang, Zeyang Sha, Michael Backes, and Yang Zhang. From Visual Prompt Learning to Zero-Shot Transfer: Mapping Is All You Need. CoRR abs/2303.05266, 2023. \ No newline at end of file diff --git a/2023/Can't Steal_ Cont-Steal! Contrastive Stealing Attacks Against Image Encoders/images.zip b/2023/Can't Steal_ Cont-Steal! Contrastive Stealing Attacks Against Image Encoders/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..b3d46c4aca346e2a4386bf8af7636c14d337f0ab --- /dev/null +++ b/2023/Can't Steal_ Cont-Steal! Contrastive Stealing Attacks Against Image Encoders/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8e999fbfa756ed370beaec2033f717be79e0588045e5b681cc670803ae0d30c8 +size 382208 diff --git a/2023/Can't Steal_ Cont-Steal! Contrastive Stealing Attacks Against Image Encoders/layout.json b/2023/Can't Steal_ Cont-Steal! Contrastive Stealing Attacks Against Image Encoders/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..78267d68cebda10bb13f0759bb12072eb19ea7b5 --- /dev/null +++ b/2023/Can't Steal_ Cont-Steal! Contrastive Stealing Attacks Against Image Encoders/layout.json @@ -0,0 +1,9036 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 130, + 103, + 463, + 139 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 103, + 463, + 139 + ], + "spans": [ + { + "bbox": [ + 130, + 103, + 463, + 139 + ], + "type": "text", + "content": "Can't Steal? Cont-Steal! Contrastive Stealing Attacks Against Image Encoders" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 123, + 160, + 462, + 176 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 123, + 160, + 462, + 176 + ], + "spans": [ + { + "bbox": [ + 123, + 160, + 462, + 176 + ], + "type": "text", + "content": "Zeyang Sha† Xinlei He† Ning Yu‡ Michael Backes† Yang Zhang†" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 97, + 189, + 493, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 97, + 189, + 493, + 217 + ], + "spans": [ + { + "bbox": [ + 97, + 189, + 493, + 217 + ], + "type": "text", + "content": "†CISPA Helmholtz Center for Information Security ‡Salesforce Research {zeyang.sha, xinlei.he, director, zhang}@cispa.de, ning.yu@salesforce.com" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 143, + 244, + 192, + 258 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 244, + 192, + 258 + ], + "spans": [ + { + "bbox": [ + 143, + 244, + 192, + 258 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 270, + 289, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 270, + 289, + 426 + ], + "spans": [ + { + "bbox": [ + 46, + 270, + 289, + 426 + ], + "type": "text", + "content": "Self-supervised representation learning techniques have been developing rapidly to make full use of unlabeled images. They encode images into rich features that are oblivious to downstream tasks. Behind their revolutionary representation power, the requirements for dedicated model designs and a massive amount of computation resources expose image encoders to the risks of potential model stealing attacks - a cheap way to mimic the well-trained encoder performance while circumventing the demanding requirements. Yet conventional attacks only target supervised classifiers given their predicted labels and/or posteriors, which leaves the vulnerability of unsupervised encoders unexplored." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 426, + 288, + 547 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 426, + 288, + 547 + ], + "spans": [ + { + "bbox": [ + 46, + 426, + 288, + 547 + ], + "type": "text", + "content": "In this paper, we first instantiate the conventional stealing attacks against encoders and demonstrate their severer vulnerability compared with downstream classifiers. To better leverage the rich representation of encoders, we further propose Cont-Steal, a contrastive-learning-based attack, and validate its improved stealing effectiveness in various experiment settings. As a takeaway, we appeal to our community's attention to the intellectual property protection of representation learning techniques, especially to the defenses against encoder stealing attacks like ours." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 568, + 128, + 582 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 568, + 128, + 582 + ], + "spans": [ + { + "bbox": [ + 47, + 568, + 128, + 582 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 589, + 287, + 649 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 589, + 287, + 649 + ], + "spans": [ + { + "bbox": [ + 46, + 589, + 287, + 649 + ], + "type": "text", + "content": "Recent years have witnessed the great success of applying deep learning (DL) to computer vision tasks. Different from supervised DL models, self-supervised learning which transforms unlabeled data samples into rich representations, has gained more and more popularity." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 649, + 287, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 649, + 287, + 685 + ], + "spans": [ + { + "bbox": [ + 46, + 649, + 287, + 685 + ], + "type": "text", + "content": "Behind its powerful representation, it is non-trivial to obtain a state-of-the-art image encoder. For instance, SimCLR [6] uses 128 TPU v3 cores to pre-train a ResNet-" + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 307, + 247, + 533, + 401 + ], + "blocks": [ + { + "bbox": [ + 307, + 247, + 533, + 401 + ], + "lines": [ + { + "bbox": [ + 307, + 247, + 533, + 401 + ], + "spans": [ + { + "bbox": [ + 307, + 247, + 533, + 401 + ], + "type": "image", + "image_path": "e59fe95481358d05a780af7dd1b54c6d4f84ec72885d0eb1614d4d8fb718b935.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 413, + 547, + 491 + ], + "lines": [ + { + "bbox": [ + 304, + 413, + 547, + 491 + ], + "spans": [ + { + "bbox": [ + 304, + 413, + 547, + 491 + ], + "type": "text", + "content": "Figure 1. Model stealing attacks against classifiers (previous) v.s. model stealing attacks against encoders (ours). Previous works aim to steal a whole classifier using the predicted label or posteriors of a target model. In our work, we aim to steal the target encoder using its embeddings. The target encoder " + }, + { + "bbox": [ + 304, + 413, + 547, + 491 + ], + "type": "inline_equation", + "content": "(E_{t})" + }, + { + "bbox": [ + 304, + 413, + 547, + 491 + ], + "type": "text", + "content": " is pre-trained and fixed, as shown in the solid frame. The surrogate encoder " + }, + { + "bbox": [ + 304, + 413, + 547, + 491 + ], + "type": "inline_equation", + "content": "(E_{s})" + }, + { + "bbox": [ + 304, + 413, + 547, + 491 + ], + "type": "text", + "content": " is trainable by the adversary, as shown in the dashed frame." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 511, + 545, + 595 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 511, + 545, + 595 + ], + "spans": [ + { + "bbox": [ + 304, + 511, + 545, + 595 + ], + "type": "text", + "content": "50 encoder with a batch size of 4096. Therefore, many big companies provide cloud-based self-supervised learning encoder services for users. For instance, Cohere," + }, + { + "bbox": [ + 304, + 511, + 545, + 595 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 304, + 511, + 545, + 595 + ], + "type": "text", + "content": " OpenAI," + }, + { + "bbox": [ + 304, + 511, + 545, + 595 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 304, + 511, + 545, + 595 + ], + "type": "text", + "content": " and Clarifai" + }, + { + "bbox": [ + 304, + 511, + 545, + 595 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 304, + 511, + 545, + 595 + ], + "type": "text", + "content": " provide the embedding API of images and texts for commercial usage. There are many works [10, 26, 30] exploring security issues of encoder-based API. Therefore, it is a very important and urgent problem." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 596, + 547, + 656 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 596, + 547, + 656 + ], + "spans": [ + { + "bbox": [ + 304, + 596, + 547, + 656 + ], + "type": "text", + "content": "These kinds of service leave the possibility of model stealing attacks [5, 25, 28, 37, 44, 48, 50, 51]. In these attacks, the adversary aims to steal the parameters or functionalities of target models with only query access to them. A successful model stealing attack does not only threaten" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 693, + 286, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 693, + 286, + 712 + ], + "spans": [ + { + "bbox": [ + 47, + 693, + 286, + 712 + ], + "type": "text", + "content": "See our code in https://github.com/zeyangsha/Cont-Steal." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 663, + 545, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 663, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 663, + 545, + 713 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 306, + 663, + 545, + 713 + ], + "type": "text", + "content": "https://cohere.ai/ \n" + }, + { + "bbox": [ + 306, + 663, + 545, + 713 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 306, + 663, + 545, + 713 + ], + "type": "text", + "content": "https://beta.openai.com/docs/api-reference/embeddings/ \n" + }, + { + "bbox": [ + 306, + 663, + 545, + 713 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 306, + 663, + 545, + 713 + ], + "type": "text", + "content": "https://www clarifai.com/models/general-image-embedding/" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16373" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 178 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 178 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 178 + ], + "type": "text", + "content": "the intellectual property of the target model, but also serves as a stepping stone for further attacks such as adversarial examples [3, 4, 16, 38, 47, 52], backdoor attacks [7, 26, 41, 43], and membership inference attacks [22-24, 30, 33, 34, 40, 42, 45, 46]. So far, model stealing attacks concentrate on the supervised classifiers, i.e., the model responses are prediction posteriors or labels for a specific downstream task. The vulnerability of unsupervised image encoders is unfortunately unexplored." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 181, + 287, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 181, + 287, + 289 + ], + "spans": [ + { + "bbox": [ + 46, + 181, + 287, + 289 + ], + "type": "text", + "content": "Our Work. To fill this gap, we pioneer the systematic investigation of model stealing attacks against image encoders. In this work, the adversary's goal is to steal the functionalities of the target model. See Figure 1 for an overview and a comparison with previous works. More specifically, we focus on encoders trained by contrastive learning, which is one of the most cutting-edge unsupervised representation learning strategies that unleash the information of unlabeled data." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 290, + 287, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 290, + 287, + 434 + ], + "spans": [ + { + "bbox": [ + 46, + 290, + 287, + 434 + ], + "type": "text", + "content": "We first instantiate the conventional stealing attacks against encoders and expose their vulnerability. Given an input image, the target encoder outputs its representation (referred to as embedding). Similar to model stealing attacks against classifiers, we consider the embedding as the \"ground truth\" label to guide the training procedure of a surrogate encoder on the adversary side. To measure the effectiveness of stealing attacks, we train an extra linear layer for the target and surrogate encoders towards the same downstream classification task. Preferably, the surrogate model should achieve both high classification accuracy and high agreement with the target predictions." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 435, + 287, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 435, + 287, + 555 + ], + "spans": [ + { + "bbox": [ + 46, + 435, + 287, + 555 + ], + "type": "text", + "content": "We evaluate our attacks on five datasets against four contrastive learning encoders. Our results demonstrate that the conventional attacks are more effective against encoders than against downstream classifiers. For instance, when we steal the downstream classifier pre-trained by SimCLR on CIFAR10 (with posteriors as its responses) using STL10 as the surrogate dataset, the adversary can only achieve an accuracy of 0.359. The accuracy, however, increases to 0.500 instead when we steal its encoder (with the embedding as its response)." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 556, + 287, + 675 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 556, + 287, + 675 + ], + "spans": [ + { + "bbox": [ + 46, + 556, + 287, + 675 + ], + "type": "text", + "content": "Despite its encouraging performance, conventional attacks are not the most suitable ones against encoders. This is because they treat each image-embedding pair individually without interacting across pairs. Different embeddings are beneficial to each other as they can serve as anchors to better locate the position of the other embeddings in their space. Contrastive learning [6, 8, 17, 20, 27, 49, 53] is a straightforward idea to achieve this goal. It is formulated to enforce the embeddings of different augmentations of the same images closer and those of different images further." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 677, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 677, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 677, + 287, + 713 + ], + "type": "text", + "content": "In a similar spirit, we propose Cont-Steal, a contrastive-learning-based model stealing attack against the encoder. The goal of Cont-Steal is to enforce the surrogate embed" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "text", + "content": "ding of an image close to its target embedding (defined as a positive pair) and also push away embeddings of different images irrespective of being generated by the target or the surrogate encoders (defined as negative pairs)." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 121, + 546, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 121, + 546, + 323 + ], + "spans": [ + { + "bbox": [ + 304, + 121, + 546, + 323 + ], + "type": "text", + "content": "The comprehensive evaluation shows that Cont-Steal outperforms the conventional model stealing attacks to a large extent. For instance, when CIFAR10 is the target dataset, Cont-Steal achieves an accuracy of 0.714 on the SimCLR encoder pretrained on CIFAR10 with the surrogate dataset and downstream dataset being STL10, while the conventional attack only achieves 0.457 accuracy. Also, Cont-Steal is more query-efficient and dataset-independent (see Figure 9 for more details). This is because Cont-Steal leverages higher-order information across samples to mimic the functionality of the target encoder. To mitigate the attacks, we evaluate different defense mechanisms including noise, top-" + }, + { + "bbox": [ + 304, + 121, + 546, + 323 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 121, + 546, + 323 + ], + "type": "text", + "content": ", rounding, and watermark. Our evaluations show that in most cases, these mechanisms cannot effectively defend against Cont-Steal. Among them, top-" + }, + { + "bbox": [ + 304, + 121, + 546, + 323 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 121, + 546, + 323 + ], + "type": "text", + "content": " can reduce the attack performance to the largest extent. However, it also strongly limits the target model's utility." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 323, + 545, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 323, + 545, + 384 + ], + "spans": [ + { + "bbox": [ + 304, + 323, + 545, + 384 + ], + "type": "text", + "content": "As a takeaway, our attack further exposes the severe vulnerability of pre-trained encoders. We appeal to our community's attention to the intellectual property protection of representation learning techniques, especially to the defenses against encoder stealing attacks like ours." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 394, + 392, + 407 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 394, + 392, + 407 + ], + "spans": [ + { + "bbox": [ + 306, + 394, + 392, + 407 + ], + "type": "text", + "content": "2. Threat Model" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 415, + 545, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 415, + 545, + 498 + ], + "spans": [ + { + "bbox": [ + 304, + 415, + 545, + 498 + ], + "type": "text", + "content": "In this work, for the encoder pre-trained with images, we consider image classification as the downstream task. We refer to the encoder as the target encoder. Then we treat both the encoder and the linear layer trained for the downstream task together as the target model. We first introduce the adversary's goal and then characterize different background knowledge that the adversary might have." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 498, + 545, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 498, + 545, + 617 + ], + "spans": [ + { + "bbox": [ + 304, + 498, + 545, + 617 + ], + "type": "text", + "content": "Adversary's Goal. Following previous work [25, 28, 44], we taxonomize the adversary's goal into two dimensions, i.e., theft and utility. The theft adversary aims to build a surrogate encoder that has similar performance on the downstream tasks as the target encoder. Different from the thief adversary, the goal of the utility adversary is to construct a surrogate encoder that behaves normally on different downstream tasks. In this case, the surrogate encoder not only faithfully \"copies\" the behaviors of the target encoder, but also serves as a stepping stone to conduct other attacks." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 617, + 545, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 617, + 545, + 665 + ], + "spans": [ + { + "bbox": [ + 304, + 617, + 545, + 665 + ], + "type": "text", + "content": "Adversary's Background Knowledge. We categorize the adversary's background knowledge into two dimensions, i.e., the knowledge of the target encoder and the distribution of the surrogate dataset." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "content": "Regarding knowledge of the target encoder, we assume that the adversary only has black-box access to it, which means that they can only query the target encoder with an input image and obtain the corresponding output, i.e., the" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "16374" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 172, + 84 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 172, + 84 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 172, + 84 + ], + "type": "text", + "content": "embedding of the input image." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 84, + 287, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 84, + 287, + 239 + ], + "spans": [ + { + "bbox": [ + 46, + 84, + 287, + 239 + ], + "type": "text", + "content": "Regarding the surrogate dataset that is used to train the surrogate encoder, we consider two cases. First, we assume the adversary has the same training dataset as the target encoder. However, such an assumption may be hard to achieve as such datasets are usually private and protected by the model owner. In a more extreme case, we assume that the adversary has totally no information about the target encoder's training dataset, which means that they can only use a different distribution dataset to conduct the model stealing attacks. We later show that the adversary can still launch effective model stealing attacks against the target encoder given a surrogate dataset that is distributed differently compared to the target dataset." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 239, + 287, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 239, + 287, + 346 + ], + "spans": [ + { + "bbox": [ + 46, + 239, + 287, + 346 + ], + "type": "text", + "content": "For the model architecture that is used to train the surrogate encoder, we consider two cases. First, we assume the adversary is aware of the target encoder's architecture and can train the same architecture surrogate encoder. Then we relax our assumption that the adversary uses different architectures to train the surrogate encoder. Our evaluation shows that the choice of architecture does not have much impact on the attack performance (see Table 3), which makes the attack more realistic." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 347, + 287, + 419 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 347, + 287, + 419 + ], + "spans": [ + { + "bbox": [ + 46, + 347, + 287, + 419 + ], + "type": "text", + "content": "Note that we also compare our attacks against the encoders to the traditional model stealing attacks that focus on the whole classifier (which has an encoder and a linear layer). If the attack targets a whole classifier, we assume the adversary may obtain the posteriors or the predicted label for an input image." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 429, + 181, + 442 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 429, + 181, + 442 + ], + "spans": [ + { + "bbox": [ + 47, + 429, + 181, + 442 + ], + "type": "text", + "content": "3. Model Stealing Attacks" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 449, + 287, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 449, + 287, + 498 + ], + "spans": [ + { + "bbox": [ + 46, + 449, + 287, + 498 + ], + "type": "text", + "content": "In this section, we first describe the conventional attacks against the encoders. Then, we propose a novel contrastive stealing framework, Cont-Steal, to steal the encoders more effectively." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 503, + 256, + 516 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 503, + 256, + 516 + ], + "spans": [ + { + "bbox": [ + 47, + 503, + 256, + 516 + ], + "type": "text", + "content": "3.1. Conventional Attacks Against Encoders" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 522, + 287, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 522, + 287, + 556 + ], + "spans": [ + { + "bbox": [ + 46, + 522, + 287, + 556 + ], + "type": "text", + "content": "The adversary takes two steps to conduct the model stealing attacks against the target encoder and one step for further evaluation." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 558, + 287, + 675 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 558, + 287, + 675 + ], + "spans": [ + { + "bbox": [ + 46, + 558, + 287, + 675 + ], + "type": "text", + "content": "Obtain the Surrogate Dataset. To conduct model stealing attacks, the adversary first needs to obtain a surrogate dataset. Based on the knowledge of the target classifier's training dataset (target dataset), we consider two cases. If the adversary has full knowledge of the target dataset, they can directly leverage the target dataset itself as the surrogate dataset. Or the adversary has no knowledge of the target dataset, which means that they can only construct the surrogate dataset, which is distributed differently from the target dataset." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "text", + "content": "Train the Surrogate Encoder. Slightly different from the classifier, the response of the encoder is an embedding, which is a feature vector. In this case, the adversary can still" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 72, + 545, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 545, + 95 + ], + "type": "text", + "content": "leverage a similar loss function to optimize the surrogate encoder, which can be defined as follows:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 354, + 106, + 545, + 139 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 354, + 106, + 545, + 139 + ], + "spans": [ + { + "bbox": [ + 354, + 106, + 545, + 139 + ], + "type": "interline_equation", + "content": "L _ {M S} = \\sum_ {k = 1} ^ {N} l \\left(E _ {T} \\left(x _ {k}\\right), E _ {S} \\left(x _ {k}\\right)\\right) \\tag {1}", + "image_path": "01c660ff5e6ded3ec9a01cd84587a897b65a7642464801b9a0d0fc910c23c6c5.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 145, + 545, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 145, + 545, + 180 + ], + "spans": [ + { + "bbox": [ + 305, + 145, + 545, + 180 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 145, + 545, + 180 + ], + "type": "inline_equation", + "content": "E_{T}(\\cdot) / E_{S}(\\cdot)" + }, + { + "bbox": [ + 305, + 145, + 545, + 180 + ], + "type": "text", + "content": " is the target/surrogate encoder, " + }, + { + "bbox": [ + 305, + 145, + 545, + 180 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 305, + 145, + 545, + 180 + ], + "type": "text", + "content": " is the total number of samples on the surrogate dataset, and " + }, + { + "bbox": [ + 305, + 145, + 545, + 180 + ], + "type": "inline_equation", + "content": "l(\\cdot)" + }, + { + "bbox": [ + 305, + 145, + 545, + 180 + ], + "type": "text", + "content": " is the MSE loss." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 182, + 545, + 301 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 182, + 545, + 301 + ], + "spans": [ + { + "bbox": [ + 304, + 182, + 545, + 301 + ], + "type": "text", + "content": "Apply the Surrogate Encoder to Downstream Tasks. To evaluate the effectiveness of model stealing attacks against the encoder, the adversary can leverage the same downstream task to both the target and surrogate encoders. Concretely, the adversary trains an extra linear layer for the target and surrogate encoders, respectively. Note that we refer to the target/surrogate encoder and the extra linear layer as the target/surrogate classifiers. Then, the adversary quantifies the attack effectiveness by measuring the performance of the target/surrogate classifier on the downstream tasks." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 310, + 502, + 323 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 310, + 502, + 323 + ], + "spans": [ + { + "bbox": [ + 305, + 310, + 502, + 323 + ], + "type": "text", + "content": "3.2. Cont-Steal Attacks Against Encoders" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 329, + 545, + 471 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 329, + 545, + 471 + ], + "spans": [ + { + "bbox": [ + 304, + 329, + 545, + 471 + ], + "type": "text", + "content": "To better leverage the rich information from the embeddings, we propose Cont-Steal, a contrastive learning-based model stealing attacks against encoders, which leverages contrastive learning to enhance the stealing performance. Concretely, Cont-Steal aims to enforce the surrogate embedding of an image to get close to its target embedding (defined as a positive pair), and also push away embeddings of different images regardless of being generated by the target or the surrogate encoders (defined as negative pairs). There are three steps for the adversary to conduct contrastive stealing attacks against encoders and one step for further evaluation." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 473, + 545, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 473, + 545, + 592 + ], + "spans": [ + { + "bbox": [ + 304, + 473, + 545, + 592 + ], + "type": "text", + "content": "Obtain the Surrogate Dataset. The adversary follows the same strategy as Section 3.1 to obtain the surrogate dataset. Data Augmentation. Our proposed Cont-Steal leverages data augmentation to transform an input image into its two augmented views. In this paper, we leverage RandAugment [11] as the augmentation method, which is made up of a group of advanced augmentation operations. Concretely, we set " + }, + { + "bbox": [ + 304, + 473, + 545, + 592 + ], + "type": "inline_equation", + "content": "n = 2" + }, + { + "bbox": [ + 304, + 473, + 545, + 592 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 473, + 545, + 592 + ], + "type": "inline_equation", + "content": "m = 14" + }, + { + "bbox": [ + 304, + 473, + 545, + 592 + ], + "type": "text", + "content": " following Cubuk et al. [11] where " + }, + { + "bbox": [ + 304, + 473, + 545, + 592 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 304, + 473, + 545, + 592 + ], + "type": "text", + "content": " denotes the number of transformations to a given sample and " + }, + { + "bbox": [ + 304, + 473, + 545, + 592 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 304, + 473, + 545, + 592 + ], + "type": "text", + "content": " represents the magnitude of global distortion." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 594, + 545, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 545, + 677 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 545, + 677 + ], + "type": "text", + "content": "Train the Surrogate Encoder. Instead of querying the encoders with the original images, the adversary queries the encoders with the augmented views of them. Concretely, for an input image " + }, + { + "bbox": [ + 304, + 594, + 545, + 677 + ], + "type": "inline_equation", + "content": "x_{i}" + }, + { + "bbox": [ + 304, + 594, + 545, + 677 + ], + "type": "text", + "content": ", we generate two augmented views of it, i.e., " + }, + { + "bbox": [ + 304, + 594, + 545, + 677 + ], + "type": "inline_equation", + "content": "\\widetilde{x}_{i,s}" + }, + { + "bbox": [ + 304, + 594, + 545, + 677 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 594, + 545, + 677 + ], + "type": "inline_equation", + "content": "\\widetilde{x}_{i,t}" + }, + { + "bbox": [ + 304, + 594, + 545, + 677 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 594, + 545, + 677 + ], + "type": "inline_equation", + "content": "\\widetilde{x}_{i,s} / \\widetilde{x}_{i,t}" + }, + { + "bbox": [ + 304, + 594, + 545, + 677 + ], + "type": "text", + "content": " is used to query the surrogate/target encoder. We consider " + }, + { + "bbox": [ + 304, + 594, + 545, + 677 + ], + "type": "inline_equation", + "content": "(\\widetilde{x}_{i,s},\\widetilde{x}_{j,t})" + }, + { + "bbox": [ + 304, + 594, + 545, + 677 + ], + "type": "text", + "content": " as a positive pair if " + }, + { + "bbox": [ + 304, + 594, + 545, + 677 + ], + "type": "inline_equation", + "content": "i = j" + }, + { + "bbox": [ + 304, + 594, + 545, + 677 + ], + "type": "text", + "content": ", and otherwise a negative pair." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 306, + 677, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 677, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 677, + 545, + 713 + ], + "type": "text", + "content": "Given a mini-batch of " + }, + { + "bbox": [ + 306, + 677, + 545, + 713 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 306, + 677, + 545, + 713 + ], + "type": "text", + "content": " samples, we generate " + }, + { + "bbox": [ + 306, + 677, + 545, + 713 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 306, + 677, + 545, + 713 + ], + "type": "text", + "content": " augmented views as the input of the target encoder and another " + }, + { + "bbox": [ + 306, + 677, + 545, + 713 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 306, + 677, + 545, + 713 + ], + "type": "text", + "content": " augmented views as the input of the surrogate encoders." + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16375" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 95 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 95 + ], + "type": "text", + "content": "Concretely, the loss of Cont-Steal can be formulated as follows:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 121, + 287, + 134 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 121, + 287, + 134 + ], + "spans": [ + { + "bbox": [ + 56, + 121, + 287, + 134 + ], + "type": "interline_equation", + "content": "D _ {\\text {e n c o d e r}} ^ {+} (i) = \\exp \\left(\\operatorname {s i m} \\left(E _ {S} \\left(\\tilde {x} _ {i, s}\\right), E _ {T} \\left(\\tilde {x} _ {i, t}\\right)\\right) / \\tau\\right), \\tag {2}", + "image_path": "9762116484446c106f60fc2733342892d33b3e7e732eec08b0fcb8762513a647.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 137, + 287, + 166 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 137, + 287, + 166 + ], + "spans": [ + { + "bbox": [ + 56, + 137, + 287, + 166 + ], + "type": "interline_equation", + "content": "D _ {\\text {e n c o d e r}} ^ {-} (i) = \\sum_ {k = 1} ^ {N} \\left(\\exp \\left(\\operatorname {s i m} \\left(E _ {S} \\left(\\widetilde {x} _ {i, s}\\right), E _ {T} \\left(\\widetilde {x} _ {k, t}\\right)\\right) / \\tau\\right)\\right), \\tag {3}", + "image_path": "7f5d4806ab84b005d56292fe90653d90cf5244e4e05e535499eacd10c858190d.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 168, + 287, + 209 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 168, + 287, + 209 + ], + "spans": [ + { + "bbox": [ + 56, + 168, + 287, + 209 + ], + "type": "interline_equation", + "content": "D _ {s e l f} ^ {-} (i) = \\sum_ {k = 1} ^ {N} \\mathbb {1} _ {[ k \\neq i ]} \\left(\\exp \\left(\\operatorname {s i m} \\left(E _ {S} \\left(\\widetilde {x} _ {i, s}\\right), E _ {S} \\left(\\widetilde {x} _ {k, s}\\right)\\right) / \\tau\\right)\\right), \\tag {4}", + "image_path": "40bf49908eda009a56857dd7b2747884df67992c960e32c0bee5113cfe85bb84.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 212, + 287, + 239 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 212, + 287, + 239 + ], + "spans": [ + { + "bbox": [ + 56, + 212, + 287, + 239 + ], + "type": "interline_equation", + "content": "l (i) = - \\log \\frac {D _ {\\text {e n c o d e r}} ^ {+} (i)}{D _ {\\text {e n c o d e r}} ^ {-} (i) + D _ {\\text {s e l f}} ^ {-} (i)}, \\tag {5}", + "image_path": "039205733c8da28060695d9af71b3cca728ae9ce932d2b88e1b93b0699d8de7d.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 240, + 287, + 263 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 240, + 287, + 263 + ], + "spans": [ + { + "bbox": [ + 56, + 240, + 287, + 263 + ], + "type": "interline_equation", + "content": "L _ {\\text {C o n t - S t e a l}} = \\frac {\\sum_ {k = 1} ^ {N} l (k)}{N}, \\tag {6}", + "image_path": "d6a7e2cdf12f1518f182a8ded789e8e1a7bd4a287a96cea1764319cbf60bb6b2.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 277, + 287, + 324 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 277, + 287, + 324 + ], + "spans": [ + { + "bbox": [ + 46, + 277, + 287, + 324 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 277, + 287, + 324 + ], + "type": "inline_equation", + "content": "E_{S}(\\cdot)" + }, + { + "bbox": [ + 46, + 277, + 287, + 324 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 277, + 287, + 324 + ], + "type": "inline_equation", + "content": "E_{T}(\\cdot)" + }, + { + "bbox": [ + 46, + 277, + 287, + 324 + ], + "type": "text", + "content": " denotes the surrogate and target encoder, " + }, + { + "bbox": [ + 46, + 277, + 287, + 324 + ], + "type": "inline_equation", + "content": "sim(u,v) = u^{T}v / ||u||||v||" + }, + { + "bbox": [ + 46, + 277, + 287, + 324 + ], + "type": "text", + "content": " represents the cosine similarity between " + }, + { + "bbox": [ + 46, + 277, + 287, + 324 + ], + "type": "inline_equation", + "content": "u" + }, + { + "bbox": [ + 46, + 277, + 287, + 324 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 277, + 287, + 324 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 46, + 277, + 287, + 324 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 277, + 287, + 324 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 46, + 277, + 287, + 324 + ], + "type": "text", + "content": " is parameter to control the temperature." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 327, + 287, + 542 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 327, + 287, + 542 + ], + "spans": [ + { + "bbox": [ + 46, + 327, + 287, + 542 + ], + "type": "text", + "content": "As illustrated in Figure 2, the conventional attack treats each embedding individually without interacting across pairs. However, different embeddings are beneficial to each other as they can serve as anchors to better locate the position of the other embeddings in their space. Cont-Steal maximizes the similarity of embeddings generated from the target and surrogate encoders for a positive pair " + }, + { + "bbox": [ + 46, + 327, + 287, + 542 + ], + "type": "inline_equation", + "content": "(\\widetilde{x}_{i,s},\\widetilde{x}_{i,t})" + }, + { + "bbox": [ + 46, + 327, + 287, + 542 + ], + "type": "text", + "content": " (orange arrows in Figure 2). For the embedding generated from the target and surrogate encoders for any pair " + }, + { + "bbox": [ + 46, + 327, + 287, + 542 + ], + "type": "inline_equation", + "content": "(\\widetilde{x}_{i,s},\\widetilde{x}_{j,t})" + }, + { + "bbox": [ + 46, + 327, + 287, + 542 + ], + "type": "text", + "content": ", contrastive stealing aims to make them more distant (green arrows in Figure 2). Besides, as pointed out by Chen et al. [6], contrastive learning benefits larger negative samples. To achieve this goal, we also consider the embeddings generated from the surrogate encoder for augmented views of different images, i.e., " + }, + { + "bbox": [ + 46, + 327, + 287, + 542 + ], + "type": "inline_equation", + "content": "(\\widetilde{x}_{i,s},\\widetilde{x}_{j,s})" + }, + { + "bbox": [ + 46, + 327, + 287, + 542 + ], + "type": "text", + "content": ", as negative pairs minimize their similarity (blue arrows in Figure 2). We later show that such design can enhance the performance of contrastive stealing (see Table 5)." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 544, + 287, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 544, + 287, + 579 + ], + "spans": [ + { + "bbox": [ + 46, + 544, + 287, + 579 + ], + "type": "text", + "content": "Apply the Surrogate Encoder to Downstream Tasks. We follow Section 3.1 to evaluate the effectiveness of model stealing on downstream tasks." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 596, + 127, + 609 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 596, + 127, + 609 + ], + "spans": [ + { + "bbox": [ + 47, + 596, + 127, + 609 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 618, + 287, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 618, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 46, + 618, + 287, + 712 + ], + "type": "text", + "content": "In this section, we first describe the experimental setup in Section 4.1. Then we show the performance of the target encoders on the downstream tasks. Next, we summarize the performance of conventional attacks against classifiers and encoders in Section 4.2. Lastly, we evaluate the performance of Cont-Steal and conduct ablation studies to demonstrate its effectiveness under different settings in Section 4.3." + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 312, + 76, + 532, + 203 + ], + "blocks": [ + { + "bbox": [ + 312, + 76, + 532, + 203 + ], + "lines": [ + { + "bbox": [ + 312, + 76, + 532, + 203 + ], + "spans": [ + { + "bbox": [ + 312, + 76, + 532, + 203 + ], + "type": "image", + "image_path": "85b287c587173ea3c5b75df0fadd4f0b60bc30a80663ed68b901d2041a071eeb.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 318, + 210, + 537, + 360 + ], + "blocks": [ + { + "bbox": [ + 318, + 210, + 537, + 360 + ], + "lines": [ + { + "bbox": [ + 318, + 210, + 537, + 360 + ], + "spans": [ + { + "bbox": [ + 318, + 210, + 537, + 360 + ], + "type": "image", + "image_path": "29b7b7acc13ea667777bc9565837d2ce09ec9557351c5fd6bb3170d8f9e0b789.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 371, + 545, + 469 + ], + "lines": [ + { + "bbox": [ + 304, + 371, + 545, + 469 + ], + "spans": [ + { + "bbox": [ + 304, + 371, + 545, + 469 + ], + "type": "text", + "content": "Figure 2. Conventional attack (top) vs. Cont-Steal (bottom) against encoders. Conventional attack applies MSE loss to approximate target embeddings for each sample individually. Cont-Steal (bottom) introduces data augmentation and interacts across multiple samples: associating target/surrogate embeddings of the same images closer and repulsing those of different images farther away. The target encoder " + }, + { + "bbox": [ + 304, + 371, + 545, + 469 + ], + "type": "inline_equation", + "content": "(E_{t})" + }, + { + "bbox": [ + 304, + 371, + 545, + 469 + ], + "type": "text", + "content": " is pre-trained and fixed, as shown in the solid frame. The surrogate encoder " + }, + { + "bbox": [ + 304, + 371, + 545, + 469 + ], + "type": "inline_equation", + "content": "(E_{s})" + }, + { + "bbox": [ + 304, + 371, + 545, + 469 + ], + "type": "text", + "content": " is trainable by the adversary, as shown in the dashed frame." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 492, + 422, + 506 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 492, + 422, + 506 + ], + "spans": [ + { + "bbox": [ + 306, + 492, + 422, + 506 + ], + "type": "text", + "content": "4.1. Experimental Setup" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 511, + 545, + 642 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 511, + 545, + 642 + ], + "spans": [ + { + "bbox": [ + 304, + 511, + 545, + 642 + ], + "type": "text", + "content": "Our encoders are pre-trained on CIFAR10 [1], and ImageNet [12]. We use four different kinds of contrastive methods: SimCLR [6], MoCo [20], BYOL [17] and SimSiam [8] to train a ResNet18 [21] as our target encoders. Our implementation is based on a PyTorch framework of contrastive learning. Then, these well pre-trained encoders will be applied to train downstream classifiers on CIFAR10 [1], STL10 [9], Fashion-MNIST [54], and SVHN [35]. In the experiments in the model stealing section, we use CIFAR10, STL10, Fashion-MNIST, and SVHN to conduct the attack." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 644, + 545, + 692 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 644, + 545, + 692 + ], + "spans": [ + { + "bbox": [ + 304, + 644, + 545, + 692 + ], + "type": "text", + "content": "Agreement and accuracy are used as metrics to evaluate the model stealing attack's performance. The agreement will evaluate the similarity of surrogate encoders and target encoders in downstream tasks. The accuracy will evaluate" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 315, + 702, + 509, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 702, + 509, + 712 + ], + "spans": [ + { + "bbox": [ + 315, + 702, + 509, + 712 + ], + "type": "text", + "content": "5https://github.com/vturrisi/solo-learn/" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16376" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 70, + 70, + 166, + 142 + ], + "blocks": [ + { + "bbox": [ + 70, + 70, + 166, + 142 + ], + "lines": [ + { + "bbox": [ + 70, + 70, + 166, + 142 + ], + "spans": [ + { + "bbox": [ + 70, + 70, + 166, + 142 + ], + "type": "image", + "image_path": "4119feca9e7f1245c2f3a2a48efecc701d744f55cc9395ac73faa4bba0ec0e13.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 97, + 144, + 141, + 154 + ], + "lines": [ + { + "bbox": [ + 97, + 144, + 141, + 154 + ], + "spans": [ + { + "bbox": [ + 97, + 144, + 141, + 154 + ], + "type": "text", + "content": "(a) CIFAR10" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 167, + 70, + 263, + 143 + ], + "blocks": [ + { + "bbox": [ + 167, + 70, + 263, + 143 + ], + "lines": [ + { + "bbox": [ + 167, + 70, + 263, + 143 + ], + "spans": [ + { + "bbox": [ + 167, + 70, + 263, + 143 + ], + "type": "image", + "image_path": "3afaced75a678288771aa36a8c03ae19f56260e927b5193ca45e454eac96b500.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 187, + 144, + 244, + 155 + ], + "lines": [ + { + "bbox": [ + 187, + 144, + 244, + 155 + ], + "spans": [ + { + "bbox": [ + 187, + 144, + 244, + 155 + ], + "type": "text", + "content": "(b) ImageNet100" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 163, + 288, + 229 + ], + "lines": [ + { + "bbox": [ + 46, + 163, + 288, + 229 + ], + "spans": [ + { + "bbox": [ + 46, + 163, + 288, + 229 + ], + "type": "text", + "content": "Figure 3. The performance of target classifiers composed by target encoder and an extra linear layer. The encoders are pre-trained on CIFAR10 (a) and ImageNet100 (b). The x-axis represents different downstream datasets for the target encoder and classifier. The y-axis represents the target model's accuracy on downstream tasks." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 252, + 287, + 313 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 252, + 287, + 313 + ], + "spans": [ + { + "bbox": [ + 46, + 252, + 287, + 313 + ], + "type": "text", + "content": "the utility of surrogate models on downstream tasks. For each metric, a larger value is more desirable. During the stealing process, we set the batch size as 128 and the learning rate as 0.001. We show more results on the impact of hyperparameters in Supplementary Material in Section A.2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 323, + 246, + 335 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 323, + 246, + 335 + ], + "spans": [ + { + "bbox": [ + 47, + 323, + 246, + 335 + ], + "type": "text", + "content": "4.2. Performance of Conventional Attacks" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 342, + 287, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 342, + 287, + 593 + ], + "spans": [ + { + "bbox": [ + 46, + 342, + 287, + 593 + ], + "type": "text", + "content": "We first show the target encoder's performance in various downstream tasks. The results are summarized in Figure 3. We conduct our experiments to explore whether the encoders are more vulnerable to model stealing attacks. We show our results of target encoders and downstream classifiers both trained on CIFAR10 in Figure 4. In all cases, the adversary can get better attack performance by stealing encoders rather than classifiers. This gap becomes especially apparent when the adversary has absolutely no knowledge of the train data. This is because the rich information in embeddings can better facilitate the learning process of surrogate encoders. For instance, when the surrogate dataset is CIFAR10 (the same as the target downstream dataset), stealing SimCLR's embeddings can achieve 0.785 agreement, while stealing predicted labels can achieve 0.712 agreement. However, when the surrogate dataset is totally different from the downstream target dataset, e.g., SVHN, stealing embeddings from SimCLR can still achieve 0.507 agreement while the agreement of stealing predicted labels drops to 0.192. We show more results in Supplementary Material Section A.4 due to the page limitation." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "text", + "content": "We also find that all model stealing attacks' accuracy and agreement are highly correlated. As shown in Figure 5, the agreement is highly correlated with the accuracy. This indicates that besides accuracy, the agreement can also be used as a metric to evaluate the performance of model stealing attacks. We show the result on Figure 5. It can be obviously seen that agreement is highly related to accuracy. We use the linear regression method to describe the relationship between agreement and accuracy and find that the relation function is " + }, + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "inline_equation", + "content": "y = 0.940 * x" + }, + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 72, + 454, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 454, + 83 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 454, + 83 + ], + "type": "text", + "content": "4.3. Performance of Cont-Steal" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 91, + 547, + 162 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 91, + 547, + 162 + ], + "spans": [ + { + "bbox": [ + 304, + 91, + 547, + 162 + ], + "type": "text", + "content": "As shown in Section 4.2, encoders are more vulnerable to model stealing attacks since the embedding usually contains richer information compared to the predicted label or posteriors. We then show that our proposed Cont-Steal can achieve better attack performance by making deeper use of embeddings' information." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 163, + 546, + 461 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 163, + 546, + 461 + ], + "spans": [ + { + "bbox": [ + 304, + 163, + 546, + 461 + ], + "type": "text", + "content": "Figure 7 shows the attack performance when the target pre-training dataset is CIFAR10. Note that we also show the attack performance on other settings in the appendix. We discover that compared to conventional attacks against encoders, Cont-Steal can consistently achieve better performance. For instance, as shown in Figure 7d, when the target encoder is MoCo trained on CIFAR10, if the adversary uses STL10 to conduct model stealing attacks against encoders, the surrogate encoder can achieve 0.841 agreement in CIFAR10 downstream tasks with the Cont-Steal but only 0.479 with conventional attacks. Another finding is that compared to the same distribution surrogate dataset, our Cont-Steal can better enhance the performance when the surrogate dataset comes from a different distribution from the pre-trained dataset. For instance, when the target encoder is SimCLR trained on CIFAR10, Cont-Steal outperforms conventional attack by 0.055 agreement when the surrogate dataset is also CIFAR10, while the improvement increases to 0.207 and 0.214 when the surrogate dataset is STL10. We show more comparing results in Supplementary Material Section A.5. Note that our Cont-Steal also has great performance on other recent state-of-the-art visual models (ViT [13], MAE [19], and CLIP [39]), as we show in Section A.6, and can have better performance than other recent similar attacks [14, 31] shown in Section A.7." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 462, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 462, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 462, + 547, + 713 + ], + "type": "text", + "content": "To better understand why Cont-Steal can always achieve better performance, we extract samples' embeddings generated by different encoders, i.e., the target encoder, surrogate encoder trained with the conventional attack, and surrogate encoder trained with the Cont-Steal, and project them into a 2-dimensional space using t-SNE. From the results summarized in Figure 6, we find that Cont-Steal can effectively mimic the pattern of the embeddings as the target encoder. However, the conventional attack fails to capture such patterns for a number of input samples, e.g., the outer circle in Figure 8c. This further demonstrates that Cont-Steal benefits from jointly considering different embeddings as they can serve as anchors to better locate the position of the other embeddings in their space. We also show some ablation study results on Supplementary Material Section A.2 to show that with the less surrogate dataset, less training epoch, and different model architecture, Cont-Steal can still achieve much better results than conventional steal. Also, we show further attacks based on the stole models on Supplementary Material Section A.3 to show that Cont-Steal can be used as a springboard for other attacks." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16377" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 83, + 70, + 187, + 148 + ], + "blocks": [ + { + "bbox": [ + 83, + 70, + 187, + 148 + ], + "lines": [ + { + "bbox": [ + 83, + 70, + 187, + 148 + ], + "spans": [ + { + "bbox": [ + 83, + 70, + 187, + 148 + ], + "type": "image", + "image_path": "7ca75274d1255e5a97d83fda6695ddbb645d584232f66126490b614d719f6ff4.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 115, + 152, + 159, + 160 + ], + "lines": [ + { + "bbox": [ + 115, + 152, + 159, + 160 + ], + "spans": [ + { + "bbox": [ + 115, + 152, + 159, + 160 + ], + "type": "text", + "content": "(a) CIFAR10" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 190, + 71, + 294, + 148 + ], + "blocks": [ + { + "bbox": [ + 190, + 71, + 294, + 148 + ], + "lines": [ + { + "bbox": [ + 190, + 71, + 294, + 148 + ], + "spans": [ + { + "bbox": [ + 190, + 71, + 294, + 148 + ], + "type": "image", + "image_path": "3acc1371b453f6896ef3db12321196b9ba5addf89f3678e5148745bf8016db6d.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 226, + 152, + 261, + 160 + ], + "lines": [ + { + "bbox": [ + 226, + 152, + 261, + 160 + ], + "spans": [ + { + "bbox": [ + 226, + 152, + 261, + 160 + ], + "type": "text", + "content": "(b) STL10" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 297, + 72, + 402, + 148 + ], + "blocks": [ + { + "bbox": [ + 297, + 72, + 402, + 148 + ], + "lines": [ + { + "bbox": [ + 297, + 72, + 402, + 148 + ], + "spans": [ + { + "bbox": [ + 297, + 72, + 402, + 148 + ], + "type": "image", + "image_path": "267e0b000c3a365c69ef675872af6ba0fea1e208686e7cfba7bb92933609cfa4.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 328, + 152, + 372, + 160 + ], + "lines": [ + { + "bbox": [ + 328, + 152, + 372, + 160 + ], + "spans": [ + { + "bbox": [ + 328, + 152, + 372, + 160 + ], + "type": "text", + "content": "(c) F-MNIST" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 403, + 72, + 508, + 148 + ], + "blocks": [ + { + "bbox": [ + 403, + 72, + 508, + 148 + ], + "lines": [ + { + "bbox": [ + 403, + 72, + 508, + 148 + ], + "spans": [ + { + "bbox": [ + 403, + 72, + 508, + 148 + ], + "type": "image", + "image_path": "463bd51baf3228ac7f90595a7c4dd970088c74976ae63db34d790c86c07cfc9b.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 440, + 152, + 473, + 160 + ], + "lines": [ + { + "bbox": [ + 440, + 152, + 473, + 160 + ], + "spans": [ + { + "bbox": [ + 440, + 152, + 473, + 160 + ], + "type": "text", + "content": "(d) SVHN" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 83, + 163, + 187, + 239 + ], + "blocks": [ + { + "bbox": [ + 83, + 163, + 187, + 239 + ], + "lines": [ + { + "bbox": [ + 83, + 163, + 187, + 239 + ], + "spans": [ + { + "bbox": [ + 83, + 163, + 187, + 239 + ], + "type": "image", + "image_path": "0df6cd4d210bf03d91d88af64c3b39213f8eea0c76feb7ad4ea94a44ebaf9ec0.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 115, + 243, + 159, + 252 + ], + "lines": [ + { + "bbox": [ + 115, + 243, + 159, + 252 + ], + "spans": [ + { + "bbox": [ + 115, + 243, + 159, + 252 + ], + "type": "text", + "content": "(e) CIFAR10" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 190, + 163, + 295, + 239 + ], + "blocks": [ + { + "bbox": [ + 190, + 163, + 295, + 239 + ], + "lines": [ + { + "bbox": [ + 190, + 163, + 295, + 239 + ], + "spans": [ + { + "bbox": [ + 190, + 163, + 295, + 239 + ], + "type": "image", + "image_path": "c1cce1e7eaa711578e2ba8ee28d4231a3763ee7220dca3ede6f97bb3ba896016.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 227, + 243, + 261, + 252 + ], + "lines": [ + { + "bbox": [ + 227, + 243, + 261, + 252 + ], + "spans": [ + { + "bbox": [ + 227, + 243, + 261, + 252 + ], + "type": "text", + "content": "(f) STL10" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 298, + 163, + 402, + 239 + ], + "blocks": [ + { + "bbox": [ + 298, + 163, + 402, + 239 + ], + "lines": [ + { + "bbox": [ + 298, + 163, + 402, + 239 + ], + "spans": [ + { + "bbox": [ + 298, + 163, + 402, + 239 + ], + "type": "image", + "image_path": "22dc3e19a6e284b471bffedb4704a1a7a3d41a0ac3c33a46a37f572e7acc5c40.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 328, + 243, + 372, + 253 + ], + "lines": [ + { + "bbox": [ + 328, + 243, + 372, + 253 + ], + "spans": [ + { + "bbox": [ + 328, + 243, + 372, + 253 + ], + "type": "text", + "content": "(g) F-MNIST" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 403, + 163, + 509, + 239 + ], + "blocks": [ + { + "bbox": [ + 403, + 163, + 509, + 239 + ], + "lines": [ + { + "bbox": [ + 403, + 163, + 509, + 239 + ], + "spans": [ + { + "bbox": [ + 403, + 163, + 509, + 239 + ], + "type": "image", + "image_path": "5e8a4c3ece83db9a5eac59c9ab04268d4eb68b804a1d748c96459c978291c115.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 440, + 243, + 474, + 252 + ], + "lines": [ + { + "bbox": [ + 440, + 243, + 474, + 252 + ], + "spans": [ + { + "bbox": [ + 440, + 243, + 474, + 252 + ], + "type": "text", + "content": "(h) SVHN" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 67, + 330, + 268, + 479 + ], + "blocks": [ + { + "bbox": [ + 46, + 262, + 547, + 307 + ], + "lines": [ + { + "bbox": [ + 46, + 262, + 547, + 307 + ], + "spans": [ + { + "bbox": [ + 46, + 262, + 547, + 307 + ], + "type": "text", + "content": "Figure 4. The performance of model stealing attack against target encoders and downstream classifiers both trained on CIFAR10. Target models can output predicted labels, posteriors, or embeddings. The adversary uses CIFAR10, STL10, Fashion-MNIST (F-MNIST), and SVHN to conduct model stealing attacks. The x-axis represents different kinds of target models. The first line's y-axis represents the agreement of the model stealing attack. The second line's y-axis represents the accuracy of the model stealing attack." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 67, + 330, + 268, + 479 + ], + "lines": [ + { + "bbox": [ + 67, + 330, + 268, + 479 + ], + "spans": [ + { + "bbox": [ + 67, + 330, + 268, + 479 + ], + "type": "image", + "image_path": "43b3f225b8da1930bb35212ee3efe3dd52b6137dc188db8b02133b94236864a9.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 487, + 288, + 519 + ], + "lines": [ + { + "bbox": [ + 46, + 487, + 288, + 519 + ], + "spans": [ + { + "bbox": [ + 46, + 487, + 288, + 519 + ], + "type": "text", + "content": "Figure 5. The relationship between accuracy and agreement. The x-axis is the agreement number, and the y-axis is the accuracy number." + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 67, + 552, + 123, + 594 + ], + "blocks": [ + { + "bbox": [ + 67, + 552, + 123, + 594 + ], + "lines": [ + { + "bbox": [ + 67, + 552, + 123, + 594 + ], + "spans": [ + { + "bbox": [ + 67, + 552, + 123, + 594 + ], + "type": "image", + "image_path": "fa618ec1342d2131ad747c287860fbedac41a52941a8f65e478920d3f2711474.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 64, + 601, + 124, + 611 + ], + "lines": [ + { + "bbox": [ + 64, + 601, + 124, + 611 + ], + "spans": [ + { + "bbox": [ + 64, + 601, + 124, + 611 + ], + "type": "text", + "content": "(a) Target encoder" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 621, + 288, + 675 + ], + "lines": [ + { + "bbox": [ + 46, + 621, + 288, + 675 + ], + "spans": [ + { + "bbox": [ + 46, + 621, + 288, + 675 + ], + "type": "text", + "content": "Figure 6. The t-SNE projection of 1,000 randomly selected samples' embeddings from target encoder, surrogate encoder under Cont-Steal, and surrogate encoder under the conventional attack, respectively. Note that the target encoder is pre-trained by SimCLR on CIFAR10." + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 141, + 552, + 195, + 594 + ], + "blocks": [ + { + "bbox": [ + 141, + 552, + 195, + 594 + ], + "lines": [ + { + "bbox": [ + 141, + 552, + 195, + 594 + ], + "spans": [ + { + "bbox": [ + 141, + 552, + 195, + 594 + ], + "type": "image", + "image_path": "ce6d0ad655b75f8a0b188891b8ceb1d396043bf17d98bf60b9754f5a57f1499f.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 143, + 601, + 190, + 611 + ], + "lines": [ + { + "bbox": [ + 143, + 601, + 190, + 611 + ], + "spans": [ + { + "bbox": [ + 143, + 601, + 190, + 611 + ], + "type": "text", + "content": "(b) Cont-Steal" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 214, + 552, + 268, + 594 + ], + "blocks": [ + { + "bbox": [ + 214, + 552, + 268, + 594 + ], + "lines": [ + { + "bbox": [ + 214, + 552, + 268, + 594 + ], + "spans": [ + { + "bbox": [ + 214, + 552, + 268, + 594 + ], + "type": "image", + "image_path": "09513a45e819a8fd6f954414b2106994d88fa140052013dcfc2aeb979f8ac24c.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 213, + 601, + 268, + 611 + ], + "lines": [ + { + "bbox": [ + 213, + 601, + 268, + 611 + ], + "spans": [ + { + "bbox": [ + 213, + 601, + 268, + 611 + ], + "type": "text", + "content": "(c) Conventional" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + } + ], + "index": 23 + }, + { + "type": "table", + "bbox": [ + 312, + 390, + 539, + 466 + ], + "blocks": [ + { + "bbox": [ + 305, + 326, + 545, + 381 + ], + "lines": [ + { + "bbox": [ + 305, + 326, + 545, + 381 + ], + "spans": [ + { + "bbox": [ + 305, + 326, + 545, + 381 + ], + "type": "text", + "content": "Table 1. The monetary and (training) time costs for normal training and Cont-Steal attack. Cont-Steal's monetary cost contains two parts: query cost and training cost. Note that we ignore the query time cost of Cont-Steal as it normally has a smaller value than the training time cost." + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 312, + 390, + 539, + 466 + ], + "lines": [ + { + "bbox": [ + 312, + 390, + 539, + 466 + ], + "spans": [ + { + "bbox": [ + 312, + 390, + 539, + 466 + ], + "type": "table", + "html": "
Monetary CostTime Cost
ModelNormal ($)Cont-Steal ($)Normal (h)Cont-Steal (h)
SimCLR58.6811.83 (1.83 + 10)20.010.62
MoCo54.8312.13 (2.13 + 10)18.690.73
BYOL61.4612.08 (2.08 + 10)20.960.71
SimSiam57.1412.00 (2.00 + 10)19.460.68
", + "image_path": "3790f71a65d7793137ad41010dd389be416a3747414c9ca3c337f2f828980146.jpg" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "table_body" + } + ], + "index": 27 + }, + { + "bbox": [ + 306, + 485, + 392, + 498 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 485, + 392, + 498 + ], + "spans": [ + { + "bbox": [ + 306, + 485, + 392, + 498 + ], + "type": "text", + "content": "4.4. Cost Analysis" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 304, + 504, + 545, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 504, + 545, + 635 + ], + "spans": [ + { + "bbox": [ + 304, + 504, + 545, + 635 + ], + "type": "text", + "content": "As we mentioned before, pre-train a state-of-the-art encoder is time-consuming and resource-demanding. We wonder if the model stealing attacks can steal the functionality of the encoder with much less cost. To this end, we evaluate the time and monetary cost of training an encoder from scratch or stealing a pre-trained encoder via ContSteal. The monetary cost of model stealing includes querying the target model and training the surrogate model. We refer to the query price as " + }, + { + "bbox": [ + 304, + 504, + 545, + 635 + ], + "type": "inline_equation", + "content": "1 for 1,000 queries based on AWS.6 Our experiment is conducted on 1 NVIDIA A100 whose price is" + }, + { + "bbox": [ + 304, + 504, + 545, + 635 + ], + "type": "text", + "content": "2.934 per hour based on google cloud.7" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 304, + 635, + 545, + 684 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 635, + 545, + 684 + ], + "spans": [ + { + "bbox": [ + 304, + 635, + 545, + 684 + ], + "type": "text", + "content": "The monetary and time cost is shown in Table 1. We observe that Cont-Steal can obtain a surrogate encoder with much less money and time cost than training the encoder from scratch. For instance, a ResNet18 trained by SimCLR" + } + ] + } + ], + "index": 30 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 315, + 693, + 527, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 693, + 527, + 703 + ], + "spans": [ + { + "bbox": [ + 315, + 693, + 527, + 703 + ], + "type": "text", + "content": "6https://aws.amazon.com/rekognition/pricing/" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 318, + 704, + 543, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 704, + 543, + 713 + ], + "spans": [ + { + "bbox": [ + 318, + 704, + 543, + 713 + ], + "type": "text", + "content": "7https://cloud.google.com/compute/gpus-pricing/" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16378" + } + ] + } + ], + "index": 34 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 83, + 70, + 187, + 148 + ], + "blocks": [ + { + "bbox": [ + 83, + 70, + 187, + 148 + ], + "lines": [ + { + "bbox": [ + 83, + 70, + 187, + 148 + ], + "spans": [ + { + "bbox": [ + 83, + 70, + 187, + 148 + ], + "type": "image", + "image_path": "50bd10a636c3765a64687310ebe792f6a0595a0688a8683cb76ab2948a712c2c.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 115, + 152, + 159, + 160 + ], + "lines": [ + { + "bbox": [ + 115, + 152, + 159, + 160 + ], + "spans": [ + { + "bbox": [ + 115, + 152, + 159, + 160 + ], + "type": "text", + "content": "(a) CIFAR10" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 190, + 71, + 295, + 148 + ], + "blocks": [ + { + "bbox": [ + 190, + 71, + 295, + 148 + ], + "lines": [ + { + "bbox": [ + 190, + 71, + 295, + 148 + ], + "spans": [ + { + "bbox": [ + 190, + 71, + 295, + 148 + ], + "type": "image", + "image_path": "7291b6de743fed6c4c3e879a0b1b9dd47616b89fb13aea5bb56545ce7bd0c926.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 226, + 152, + 261, + 160 + ], + "lines": [ + { + "bbox": [ + 226, + 152, + 261, + 160 + ], + "spans": [ + { + "bbox": [ + 226, + 152, + 261, + 160 + ], + "type": "text", + "content": "(b) STL10" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 298, + 72, + 402, + 148 + ], + "blocks": [ + { + "bbox": [ + 298, + 72, + 402, + 148 + ], + "lines": [ + { + "bbox": [ + 298, + 72, + 402, + 148 + ], + "spans": [ + { + "bbox": [ + 298, + 72, + 402, + 148 + ], + "type": "image", + "image_path": "961afd3798f5ba4a144d9541282cff8b14ca0de8859b78790ad87c5977a8b80f.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 328, + 152, + 372, + 160 + ], + "lines": [ + { + "bbox": [ + 328, + 152, + 372, + 160 + ], + "spans": [ + { + "bbox": [ + 328, + 152, + 372, + 160 + ], + "type": "text", + "content": "(c) F-MNIST" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 403, + 72, + 508, + 148 + ], + "blocks": [ + { + "bbox": [ + 403, + 72, + 508, + 148 + ], + "lines": [ + { + "bbox": [ + 403, + 72, + 508, + 148 + ], + "spans": [ + { + "bbox": [ + 403, + 72, + 508, + 148 + ], + "type": "image", + "image_path": "02426e26ce9d61527cae74aef17b4b1edca4a13464e0f4ed7e49c887ac2e037e.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 83, + 163, + 188, + 239 + ], + "blocks": [ + { + "bbox": [ + 83, + 163, + 188, + 239 + ], + "lines": [ + { + "bbox": [ + 83, + 163, + 188, + 239 + ], + "spans": [ + { + "bbox": [ + 83, + 163, + 188, + 239 + ], + "type": "image", + "image_path": "a7f645d73ca807d125679a62b078825367274826362cdafcdf9bfc9e8d61dc65.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 115, + 243, + 159, + 252 + ], + "lines": [ + { + "bbox": [ + 115, + 243, + 159, + 252 + ], + "spans": [ + { + "bbox": [ + 115, + 243, + 159, + 252 + ], + "type": "text", + "content": "(e) CIFAR10" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 190, + 163, + 295, + 239 + ], + "blocks": [ + { + "bbox": [ + 190, + 163, + 295, + 239 + ], + "lines": [ + { + "bbox": [ + 190, + 163, + 295, + 239 + ], + "spans": [ + { + "bbox": [ + 190, + 163, + 295, + 239 + ], + "type": "image", + "image_path": "c796704c9354237234f9d12bdb37fc6b30d18e6feaaffc07e637417257d738bd.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 227, + 243, + 261, + 252 + ], + "lines": [ + { + "bbox": [ + 227, + 243, + 261, + 252 + ], + "spans": [ + { + "bbox": [ + 227, + 243, + 261, + 252 + ], + "type": "text", + "content": "(f) STL10" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 298, + 163, + 402, + 239 + ], + "blocks": [ + { + "bbox": [ + 298, + 163, + 402, + 239 + ], + "lines": [ + { + "bbox": [ + 298, + 163, + 402, + 239 + ], + "spans": [ + { + "bbox": [ + 298, + 163, + 402, + 239 + ], + "type": "image", + "image_path": "c1af7f3f0ddd109c57a48bd35c17caaffd93df608c6633e0aa61bfe9e802aacd.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 328, + 243, + 372, + 253 + ], + "lines": [ + { + "bbox": [ + 328, + 243, + 372, + 253 + ], + "spans": [ + { + "bbox": [ + 328, + 243, + 372, + 253 + ], + "type": "text", + "content": "(g) F-MNIST" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 403, + 163, + 508, + 239 + ], + "blocks": [ + { + "bbox": [ + 440, + 152, + 473, + 160 + ], + "lines": [ + { + "bbox": [ + 440, + 152, + 473, + 160 + ], + "spans": [ + { + "bbox": [ + 440, + 152, + 473, + 160 + ], + "type": "text", + "content": "(d) SVHN" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 403, + 163, + 508, + 239 + ], + "lines": [ + { + "bbox": [ + 403, + 163, + 508, + 239 + ], + "spans": [ + { + "bbox": [ + 403, + 163, + 508, + 239 + ], + "type": "image", + "image_path": "fae52eea7c07e6a312dfd9279f7fd05195a2960cad8fb6e264c34d499d823721.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 440, + 243, + 473, + 252 + ], + "lines": [ + { + "bbox": [ + 440, + 243, + 473, + 252 + ], + "spans": [ + { + "bbox": [ + 440, + 243, + 473, + 252 + ], + "type": "text", + "content": "(h) SVHN" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 51, + 326, + 127, + 381 + ], + "blocks": [ + { + "bbox": [ + 46, + 262, + 547, + 307 + ], + "lines": [ + { + "bbox": [ + 46, + 262, + 547, + 307 + ], + "spans": [ + { + "bbox": [ + 46, + 262, + 547, + 307 + ], + "type": "text", + "content": "Figure 7. The performance of Cont-Steal and conventional attack against target encoders trained on CIFAR10. The adversary uses CIFAR10, STL10, F-MNIST, and SVHN to conduct model stealing attacks. The adversary uses CIFAR10 as the downstream task to evaluate the attack performance. The x-axis represents different kinds of the target model. The first line's y-axis represents the agreement of the model stealing attack. The second line's y-axis represents the accuracy of the model stealing attack." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 51, + 326, + 127, + 381 + ], + "lines": [ + { + "bbox": [ + 51, + 326, + 127, + 381 + ], + "spans": [ + { + "bbox": [ + 51, + 326, + 127, + 381 + ], + "type": "image", + "image_path": "f470e76d0b9bd37680a0c9f53f2d7fa1faa050910bbe4869d6609abc624d6ae7.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 62, + 386, + 117, + 396 + ], + "lines": [ + { + "bbox": [ + 62, + 386, + 117, + 396 + ], + "spans": [ + { + "bbox": [ + 62, + 386, + 117, + 396 + ], + "type": "text", + "content": "(a) Adding noise" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 129, + 326, + 205, + 382 + ], + "blocks": [ + { + "bbox": [ + 129, + 326, + 205, + 382 + ], + "lines": [ + { + "bbox": [ + 129, + 326, + 205, + 382 + ], + "spans": [ + { + "bbox": [ + 129, + 326, + 205, + 382 + ], + "type": "image", + "image_path": "8b16877a46b1522e38dcade88684c045d8ef410010ca0bae3058eb925722b79c.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 151, + 386, + 184, + 396 + ], + "lines": [ + { + "bbox": [ + 151, + 386, + 184, + 396 + ], + "spans": [ + { + "bbox": [ + 151, + 386, + 184, + 396 + ], + "type": "text", + "content": "(b) Top- " + }, + { + "bbox": [ + 151, + 386, + 184, + 396 + ], + "type": "inline_equation", + "content": "k" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 405, + 287, + 450 + ], + "lines": [ + { + "bbox": [ + 46, + 405, + 287, + 450 + ], + "spans": [ + { + "bbox": [ + 46, + 405, + 287, + 450 + ], + "type": "text", + "content": "Figure 8. The performance of different defense methods. Target encoders are trained on CIFAR10. The downstream dataset and surrogate dataset are both STL10. The x-axis represents different defense levels. The y-axis represents the model's accuracy." + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 208, + 326, + 283, + 382 + ], + "blocks": [ + { + "bbox": [ + 208, + 326, + 283, + 382 + ], + "lines": [ + { + "bbox": [ + 208, + 326, + 283, + 382 + ], + "spans": [ + { + "bbox": [ + 208, + 326, + 283, + 382 + ], + "type": "image", + "image_path": "26ef07a82ee8c537eced95175b4f4d4475abb0193e362340d7032c969dd634fb.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 223, + 386, + 268, + 396 + ], + "lines": [ + { + "bbox": [ + 223, + 386, + 268, + 396 + ], + "spans": [ + { + "bbox": [ + 223, + 386, + 268, + 396 + ], + "type": "text", + "content": "(c) Rounding" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + } + ], + "index": 21 + }, + { + "bbox": [ + 46, + 471, + 287, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 471, + 287, + 544 + ], + "spans": [ + { + "bbox": [ + 46, + 471, + 287, + 544 + ], + "type": "text", + "content": "on CIFAR10 takes 20.01 hours and " + }, + { + "bbox": [ + 46, + 471, + 287, + 544 + ], + "type": "inline_equation", + "content": "58.68" + }, + { + "bbox": [ + 46, + 471, + 287, + 544 + ], + "type": "text", + "content": "'s on 1 NVIDIA A100 GPU, while Cont-Steal only takes 0.62 hours and " + }, + { + "bbox": [ + 46, + 471, + 287, + 544 + ], + "type": "inline_equation", + "content": "11.83" + }, + { + "bbox": [ + 46, + 471, + 287, + 544 + ], + "type": "text", + "content": "'s to steal an encoder that performs similarly on downstream tasks. The results demonstrate that Cont-Steal is able to construct surrogate encoders that perform similarly to the target encoders but with much less time and monetary cost." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 47, + 551, + 111, + 562 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 551, + 111, + 562 + ], + "spans": [ + { + "bbox": [ + 47, + 551, + 111, + 562 + ], + "type": "text", + "content": "4.5. Defenses" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 46, + 570, + 287, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 570, + 287, + 629 + ], + "spans": [ + { + "bbox": [ + 46, + 570, + 287, + 629 + ], + "type": "text", + "content": "In this section, we will consider different defenses against model stealing attacks on encoders to evaluate the robustness of our proposed attack. We divided all defenses into three categories: perturbation-based defense [37] and watermark-based defense [2]." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 46, + 629, + 287, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 629, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 46, + 629, + 287, + 689 + ], + "type": "text", + "content": "Perturbation-based Defense. In this defense setting, the defender aims to perturb the output of the target model to limit the information the adversary can obtain. The common practice of this kind of defense includes adding noise [37], top-" + }, + { + "bbox": [ + 46, + 629, + 287, + 689 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 629, + 287, + 689 + ], + "type": "text", + "content": " [37], and feature rounding [48]." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "content": "Adding noise means that the defender will introduce noise value to the original output of the model. In our case," + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 304, + 328, + 546, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 328, + 546, + 628 + ], + "spans": [ + { + "bbox": [ + 304, + 328, + 546, + 628 + ], + "type": "text", + "content": "we consider adding Gaussian noise to the embeddings generated by the target encoder. We set the mean value to 0, and different noise levels represent different standard deviations of the Gaussian distribution. For Top-" + }, + { + "bbox": [ + 304, + 328, + 546, + 628 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 328, + 546, + 628 + ], + "type": "text", + "content": ", the defender will only output the first " + }, + { + "bbox": [ + 304, + 328, + 546, + 628 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 328, + 546, + 628 + ], + "type": "text", + "content": " largest number of each embedding (and set the rest as 0). In this way, the high-dimensional information of the image contained in embeddings can be appropriately reduced. Regarding feature rounding, the defender will truncate the values in the embedding to a specific digit. As a case study, we consider a ResNet18 encoder pre-trained on CIFAR10 with SimCLR and take STL10 to train its downstream classifier. The experimental results are summarized in Figure 8. We can observe that while adding noise and top-" + }, + { + "bbox": [ + 304, + 328, + 546, + 628 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 328, + 546, + 628 + ], + "type": "text", + "content": " can reduce the model stealing attacks' performance, it may also degrade the target model performance to a large extent. For instance, when the noise increases from 0 to 10, the attack performance of ContSteal decreases from 0.729 to 0.410, while the target encoder's performance drops from 0.734 to 0.098. On the other hand, rounding only has a limited effect on both target model performance and attack performance. This indicates that perturbation-based defense cannot defend against the encoder's model stealing attack effectively since they cannot reach a good trade-off between attack performance and model utility." + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 304, + 629, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 629, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 629, + 546, + 713 + ], + "type": "text", + "content": "Watermark-based Defense. Watermark-based defense is also one of the most popular defense methods against model stealing attacks [2]. Watermark provides copyright protection by adding some specific identification to the target model. If the surrogate model is stolen from the watermarked target model, then ideally, it will contain the same watermark as well. Adi et al. [2] show that back" + } + ] + } + ], + "index": 30 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16379" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 48, + 123, + 288, + 186 + ], + "blocks": [ + { + "bbox": [ + 47, + 70, + 288, + 115 + ], + "lines": [ + { + "bbox": [ + 47, + 70, + 288, + 115 + ], + "spans": [ + { + "bbox": [ + 47, + 70, + 288, + 115 + ], + "type": "text", + "content": "Table 2. Watermark defense. Pretrain dataset and surrogate dataset are both CIFAR10. Watermark leverages a watermark rate (wr) to verify the ownership of target models. A higher wr denotes a better verification performance." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 48, + 123, + 288, + 186 + ], + "lines": [ + { + "bbox": [ + 48, + 123, + 288, + 186 + ], + "spans": [ + { + "bbox": [ + 48, + 123, + 288, + 186 + ], + "type": "table", + "html": "
DatasetTarget model (acc/wr)Cont-Steal (acc/wr)Baseline (acc/wr)
CIFAR100.864 / 0.9980.769 / 0.1300.871 / 0.095
STL100.721 / 0.9990.702 / 0.0340.733 / 0.111
SVHN0.501 / 0.9990.535 / 0.3030.492 / 0.103
F-MNIST0.857 / 0.9990.813 / 0.0610.850 / 0.099
", + "image_path": "5cc0eb05ceb5bd2b5e6307c739477dda2f74284d9912d0379f683db96a161765.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 205, + 287, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 205, + 287, + 312 + ], + "spans": [ + { + "bbox": [ + 46, + 205, + 287, + 312 + ], + "type": "text", + "content": "door technology can be used as the watermark to protect the model. In that sense, BadEncoder [26], a backdoor mechanism against the encoder, can be leveraged as a watermarking technology for our target encoder as well. The defenders first will train the watermarked (backdoored) encoder, where images with a certain trigger will cause misclassification. Then, if they find the surrogate model can also misclassify images with the same trigger, the defenders can claim ownership of the surrogate model." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 313, + 288, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 313, + 288, + 552 + ], + "spans": [ + { + "bbox": [ + 46, + 313, + 288, + 552 + ], + "type": "text", + "content": "In our experiments, we leverage BadEncoder to watermark the encoder pre-trained on CIFAR10 by SimCLR, and leverage different downstream datasets to perform different tasks. We assume a strong adversary that has the same downstream dataset as the surrogate dataset. Also, we consider the baseline cases where the trigger samples are fed into the clean model to calculate the watermark rate (wr). As shown in Table 2), the watermark cannot be preserved as the surrogate models constructed by Cont-Steal have similar wr as the baseline model. For instance, when the downstream dataset is CIFAR10, Cont-Steal builds a surrogate model with 0.769 accuracy while only 0.130 wr, which is close to the baseline model. This indicates that Cont-Steal can bypass the watermarking technique as it can reach similar utility while reducing the wr to a large extent. Note that there is also another work to protect contrastive learning models from model stealing attacks using dataset inference [15]. We show in Supplementary Material in Section A.8 that this kind of defense can be easily bypassed by Cont-Steal." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 561, + 134, + 574 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 561, + 134, + 574 + ], + "spans": [ + { + "bbox": [ + 47, + 561, + 134, + 574 + ], + "type": "text", + "content": "5. Related Work" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 582, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 582, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 582, + 288, + 713 + ], + "type": "text", + "content": "Contrastive Learning. Contrastive learning is one of the most popular methods to train encoders. Current works [6, 8, 17, 20, 49, 53] propose different advanced contrastive learning algorithms. SimCLR, MoCo, BYOL, SimSiam are currently the mainstream frameworks of contrastive learning. Thus, we concentrate on them in this paper. There are many works on evaluating the security and privacy risks of contrastive learning. Previous works [23, 26, 30] propose membership inference attacks, attribution inference attacks, and backdoor attacks on contrastive learning. All proposed attacks show that contrastive-based models are vulnerable" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 72, + 545, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 96 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 96 + ], + "type": "text", + "content": "to popular attacks. Therefore, the security issues of self-supervised learning deserve more attention." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 101, + 547, + 293 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 101, + 547, + 293 + ], + "spans": [ + { + "bbox": [ + 304, + 101, + 547, + 293 + ], + "type": "text", + "content": "Model Stealing Attack. In model stealing, the adversary's goal is to steal part of the target model. Tramér et al. [48] proposed the first model stealing attack against black-box machine learning API to steal its parameters. Wang et al. [50] proposed the first hyperparameter stealing attacks against ML models. Oh et al. [36] also tried to steal machine learning model's architectures and hyperparameters. Orekondy et al. [37] proposed knockoff nets, which aim at stealing the functionality of black-box models. Krishna et al. [28] formulated a model stealing attack against BERT-based API. Besides, Wu et al. [51] and Shen et al. [44] perform model stealing attacks against Graph Neural Networks. These works often have relatively strong assumptions, such as the model family is known and the victim's data is partly available while we conduct model stealing attacks against encoders and relax the above assumption." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 319, + 378, + 331 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 319, + 378, + 331 + ], + "spans": [ + { + "bbox": [ + 306, + 319, + 378, + 331 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 344, + 545, + 416 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 344, + 545, + 416 + ], + "spans": [ + { + "bbox": [ + 304, + 344, + 545, + 416 + ], + "type": "text", + "content": "In this paper, we conduct the first model stealing risk assessment towards image encoders. Our evaluation shows that the encoder is more vulnerable to model stealing attacks compared to the classifier. This is because the embedding provided by the encoder contains richer information than the posteriors or predicted labels from whole classifiers." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 422, + 545, + 589 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 422, + 545, + 589 + ], + "spans": [ + { + "bbox": [ + 304, + 422, + 545, + 589 + ], + "type": "text", + "content": "To better unleash the power from the embeddings, we propose Cont-Steal, a contrastive learning-based model stealing method against encoders. Concretely, Cont-Steal introduces different types of negative pairs as \"anchors\" to better navigate the surrogate encoder and learn the functionality of the target encoder. Extensive evaluations show that Cont-Steal consistently performs better than conventional attacks against encoders. And such an advantage is further amplified when the adversary has no information on the target dataset, a limited amount of data, and restricted query budgets. Our work points out that the threat of model stealing attacks against encoders is largely underestimated, which prompts the need for more effective intellectual property protection of representation learning techniques." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 594, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 547, + 713 + ], + "type": "text", + "content": "Acknowledgments. We thank all anonymous reviewers for their constructive comments. This work is partially funded by the Helmholtz Association within the project \"Trustworthy Federated Data Analytics\" (TFDA) (funding number ZT-I-OO1 4) and by the European Health and Digital Executive Agency (HADEA) within the project \"Understanding the individual host response against Hepatitis D Virus to develop a personalized approach for the management of hepatitis D\" (D-Solve) (grant agreement number 101057917)." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "16380" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 286, + 714 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 53, + 91, + 286, + 114 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 286, + 114 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 286, + 114 + ], + "type": "text", + "content": "[1] https://www.cs.toronto.edu/~kriz/cifar.html." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 124, + 286, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 124, + 286, + 196 + ], + "spans": [ + { + "bbox": [ + 53, + 124, + 286, + 196 + ], + "type": "text", + "content": "[2] Yossi Adi, Carsten Baum, Moustapha Cisse, Benny Pinkas, and Joseph Keshet. Turning Your Weakness Into a Strength: Watermarking Deep Neural Networks by Backdoorsing. In USENIX Security Symposium (USENIX Security), pages 1615-1631. USENIX, 2018." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 206, + 286, + 289 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 206, + 286, + 289 + ], + "spans": [ + { + "bbox": [ + 54, + 206, + 286, + 289 + ], + "type": "text", + "content": "[3] Battista Biggio, Igino Corona, Davide Maiorca, Blaine Nelson, Nedim Srndic, Pavel Laskov, Giorgio Giacinto, and Fabio Roli. Evasion Attacks against Machine Learning at Test Time. In European Conference on Machine Learning and Principles and Practice of Knowledge Discovery in Databases (ECML/PKDD), pages 387-402. Springer, 2013." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 299, + 286, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 299, + 286, + 346 + ], + "spans": [ + { + "bbox": [ + 54, + 299, + 286, + 346 + ], + "type": "text", + "content": "[4] Nicholas Carlini and David Wagner. Towards Evaluating the Robustness of Neural Networks. In IEEE Symposium on Security and Privacy (S&P), pages 39-57. IEEE, 2017." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 356, + 286, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 356, + 286, + 403 + ], + "spans": [ + { + "bbox": [ + 54, + 356, + 286, + 403 + ], + "type": "text", + "content": "[5] Varun Chandrasekaran, Kamalika Chaudhuri, Irene Giacomelli, Somesh Jha, and Songbai Yan. Model Extraction and Active Learning. CoRR abs/1811.02054, 2018." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 413, + 286, + 473 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 413, + 286, + 473 + ], + "spans": [ + { + "bbox": [ + 54, + 413, + 286, + 473 + ], + "type": "text", + "content": "[6] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey E. Hinton. A Simple Framework for Contrastive Learning of Visual Representations. In International Conference on Machine Learning (ICML), pages 1597-1607. PMLR, 2020." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 54, + 483, + 286, + 554 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 483, + 286, + 554 + ], + "spans": [ + { + "bbox": [ + 54, + 483, + 286, + 554 + ], + "type": "text", + "content": "[7] Xiaoyi Chen, Ahmed Salem, Michael Backes, Shiqing Ma, Qingni Shen, Zhonghai Wu, and Yang Zhang. BadNL: Backdoor Attacks Against NLP Models with Semantic-preserving Improvements. In Annual Computer Security Applications Conference (ACSAC), pages 554-569. ACSAC, 2021." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 54, + 563, + 286, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 563, + 286, + 611 + ], + "spans": [ + { + "bbox": [ + 54, + 563, + 286, + 611 + ], + "type": "text", + "content": "[8] Xinlei Chen and Kaiming He. Exploring Simple Siamese Representation Learning. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 15750-15758. IEEE, 2021." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 54, + 620, + 286, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 620, + 286, + 679 + ], + "spans": [ + { + "bbox": [ + 54, + 620, + 286, + 679 + ], + "type": "text", + "content": "[9] Adam Coates, Andrew Y. Ng, and Honglak Lee. An Analysis of Single-Layer Networks in Unsupervised Feature Learning. In International Conference on Artificial Intelligence and Statistics (AISTATS), pages 215-223. JMLR, 2011." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 689, + 286, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 689, + 286, + 714 + ], + "spans": [ + { + "bbox": [ + 48, + 689, + 286, + 714 + ], + "type": "text", + "content": "[10] Tianshuo Cong, Xinlei He, and Yang Zhang. SSL-Guard: A Watermarking Scheme for Self-supervised" + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 72, + 545, + 714 + ], + "type": "list", + "angle": 0, + "index": 21, + "blocks": [ + { + "bbox": [ + 327, + 72, + 545, + 108 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 72, + 545, + 108 + ], + "spans": [ + { + "bbox": [ + 327, + 72, + 545, + 108 + ], + "type": "text", + "content": "Learning Pre-trained Encoders. In ACM SIGSAC Conference on Computer and Communications Security (CCS), pages 579-593. ACM, 2022." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 308, + 117, + 545, + 186 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 117, + 545, + 186 + ], + "spans": [ + { + "bbox": [ + 308, + 117, + 545, + 186 + ], + "type": "text", + "content": "[11] Ekin Dogus Cubuk, Barret Zoph, Jon Shlens, and Quoc Le. RandAugment: Practical Automated Data Augmentation with a Reduced Search Space. In Annual Conference on Neural Information Processing Systems (NeurIPS), pages 18613-18624. NeurIPS, 2020." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 307, + 197, + 545, + 255 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 197, + 545, + 255 + ], + "spans": [ + { + "bbox": [ + 307, + 197, + 545, + 255 + ], + "type": "text", + "content": "[12] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. ImageNet: A large-scale hierarchical image database. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 248–255. IEEE, 2009." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 308, + 265, + 545, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 265, + 545, + 360 + ], + "spans": [ + { + "bbox": [ + 308, + 265, + 545, + 360 + ], + "type": "text", + "content": "[13] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale. In International Conference on Learning Representations (ICLR), 2021." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 369, + 545, + 428 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 369, + 545, + 428 + ], + "spans": [ + { + "bbox": [ + 308, + 369, + 545, + 428 + ], + "type": "text", + "content": "[14] Adam Dziedzic, Nikita Dhawan, Muhammad Ahmad Kaleem, Jonas Guan, and Nicolas Papernot. On the Difficulty of Defending Self-Supervised Learning against Model Extraction. In International Conference on Machine Learning (ICML). JMLR, 2022." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 437, + 545, + 508 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 437, + 545, + 508 + ], + "spans": [ + { + "bbox": [ + 308, + 437, + 545, + 508 + ], + "type": "text", + "content": "[15] Adam Dziedzic, Haonan Duan, Muhammad Ahmad Kaleem, Nikita Dhawan, Jonas Guan, Yannis Cattan, Franziska Boenisch, and Nicolas Papernot. Dataset Inference for Self-Supervised Models. In Annual Conference on Neural Information Processing Systems (NeurIPS). NeurIPS, 2022." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 517, + 545, + 564 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 517, + 545, + 564 + ], + "spans": [ + { + "bbox": [ + 308, + 517, + 545, + 564 + ], + "type": "text", + "content": "[16] Ian Goodfellow, Jonathon Shlens, and Christian Szegedy. Explaining and Harnessing Adversarial Examples. In International Conference on Learning Representations (ICLR), 2015." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 574, + 545, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 574, + 545, + 679 + ], + "spans": [ + { + "bbox": [ + 308, + 574, + 545, + 679 + ], + "type": "text", + "content": "[17] Jean-Bastien Grill, Florian Strub, Florent Altché, Corentin Tallec, Pierre H. Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Ávila Pires, Zhaohan Guo, Mohammad Gheshlaghi Azar, Bilal Piot, Koray Kavukcuoglu, Rémi Munos, and Michal Valko. Bootstrap Your Own Latent - A New Approach to Self-Supervised Learning. In Annual Conference on Neural Information Processing Systems (NeurIPS). NeurIPS, 2020." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 689, + 545, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 689, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 308, + 689, + 545, + 714 + ], + "type": "text", + "content": "[18] Yunhui Guo, Honghui Shi, Abhishek Kumar, Kristen Grauman, Tajana Rosing, and Rogério Schmidt" + } + ] + } + ], + "index": 20 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "16381" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 69, + 72, + 287, + 120 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 72, + 287, + 120 + ], + "spans": [ + { + "bbox": [ + 69, + 72, + 287, + 120 + ], + "type": "text", + "content": "Feris. SpotTune: Transfer Learning Through Adaptive Fine-Tuning. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 4805-4814. IEEE, 2019." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 128, + 287, + 175 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 128, + 287, + 175 + ], + "spans": [ + { + "bbox": [ + 48, + 128, + 287, + 175 + ], + "type": "text", + "content": "[19] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dólár, and Ross B. Girshick. Masked Autoencoders Are Scalable Vision Learners. CoRR abs/2111.06377, 2021." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 184, + 287, + 243 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 184, + 287, + 243 + ], + "spans": [ + { + "bbox": [ + 48, + 184, + 287, + 243 + ], + "type": "text", + "content": "[20] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross B. Girshick. Momentum Contrast for Unsupervised Visual Representation Learning. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 9726-9735. IEEE, 2020." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 251, + 287, + 299 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 251, + 287, + 299 + ], + "spans": [ + { + "bbox": [ + 48, + 251, + 287, + 299 + ], + "type": "text", + "content": "[21] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep Residual Learning for Image Recognition. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 770-778. IEEE, 2016." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 308, + 287, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 308, + 287, + 354 + ], + "spans": [ + { + "bbox": [ + 48, + 308, + 287, + 354 + ], + "type": "text", + "content": "[22] Xinlei He, Rui Wen, Yixin Wu, Michael Backes, Yun Shen, and Yang Zhang. Node-Level Membership Inference Attacks Against Graph Neural Networks. CoRR abs/2102.05429, 2021." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 363, + 287, + 411 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 363, + 287, + 411 + ], + "spans": [ + { + "bbox": [ + 48, + 363, + 287, + 411 + ], + "type": "text", + "content": "[23] Xinlei He and Yang Zhang. Quantifying and Mitigating Privacy Risks of Contrastive Learning. In ACM SIGSAC Conference on Computer and Communications Security (CCS), pages 845-863. ACM, 2021." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 418, + 287, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 418, + 287, + 479 + ], + "spans": [ + { + "bbox": [ + 48, + 418, + 287, + 479 + ], + "type": "text", + "content": "[24] Bo Hui, Yuchen Yang, Haolin Yuan, Philippe Burlina, Neil Zhenqiang Gong, and Yinzhi Cao. Practical Blind Membership Inference Attack via Differential Comparisons. In Network and Distributed System Security Symposium (NDSS). Internet Society, 2021." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 487, + 287, + 546 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 487, + 287, + 546 + ], + "spans": [ + { + "bbox": [ + 48, + 487, + 287, + 546 + ], + "type": "text", + "content": "[25] Matthew Jagielski, Nicholas Carlini, David Berthelot, Alex Kurakin, and Nicolas Papernot. High Accuracy and High Fidelity Extraction of Neural Networks. In USENIX Security Symposium (USENIX Security), pages 1345-1362. USENIX, 2020." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 554, + 287, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 554, + 287, + 601 + ], + "spans": [ + { + "bbox": [ + 48, + 554, + 287, + 601 + ], + "type": "text", + "content": "[26] Jinyuan Jia, Yupei Liu, and Neil Zhenqiang Gong. BadEncoder: Backdoor Attacks to Pre-trained Encoders in Self-Supervised Learning. In IEEE Symposium on Security and Privacy (S&P). IEEE, 2022." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 609, + 287, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 609, + 287, + 669 + ], + "spans": [ + { + "bbox": [ + 48, + 609, + 287, + 669 + ], + "type": "text", + "content": "[27] Prannay Khosla, Piotr Teterwak, Chen Wang, Aaron Sarna, Yonglong Tian, Phillip Isola, Aaron Maschinot, Ce Liu, and Dilip Krishnan. Supervised Contrastive Learning. In Annual Conference on Neural Information Processing Systems (NeurIPS). NeurIPS, 2020." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 677, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 677, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 677, + 287, + 712 + ], + "type": "text", + "content": "[28] Kalpesh Krishna, Gaurav Singh Tomar, Ankur P. Parikh, Nicolas Papernot, and Mohit Iyyer. Thieves on Sesame Street! Model Extraction of BERT-based" + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 72, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 23, + "blocks": [ + { + "bbox": [ + 328, + 72, + 544, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 328, + 72, + 544, + 95 + ], + "spans": [ + { + "bbox": [ + 328, + 72, + 544, + 95 + ], + "type": "text", + "content": "APIs. In International Conference on Learning Representations (ICLR), 2020." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 307, + 104, + 545, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 104, + 545, + 139 + ], + "spans": [ + { + "bbox": [ + 307, + 104, + 545, + 139 + ], + "type": "text", + "content": "[29] Alexey Kurakin, Ian Goodfellow, and Samy Bengio. Adversarial Examples in the Physical World. CoRR abs/1607.02533, 2016." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 307, + 148, + 545, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 148, + 545, + 207 + ], + "spans": [ + { + "bbox": [ + 307, + 148, + 545, + 207 + ], + "type": "text", + "content": "[30] Hongbin Liu, Jinyuan Jia, Wenjie Qu, and Neil Zhenqiang Gong. EncoderMI: Membership Inference against Pre-trained Encoders in Contrastive Learning. In ACM SIGSAC Conference on Computer and Communications Security (CCS). ACM, 2021." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 216, + 545, + 275 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 216, + 545, + 275 + ], + "spans": [ + { + "bbox": [ + 307, + 216, + 545, + 275 + ], + "type": "text", + "content": "[31] Yupei Liu, Jinyuan Jia, Hongbin Liu, and Neil Zhenqiang Gong. StolenEncoder: Stealing Pre-trained Encoders in Self-supervised Learning. In ACM SIGSAC Conference on Computer and Communications Security (CCS), pages 2115-212. ACM, 2022." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 284, + 545, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 284, + 545, + 342 + ], + "spans": [ + { + "bbox": [ + 307, + 284, + 545, + 342 + ], + "type": "text", + "content": "[32] Aleksander Madry, Aleksandar Makelov, Ludwig Schmidt, Dimitris Tsipras, and Adrian Vladu. Towards Deep Learning Models Resistant to Adversarial Attacks. In International Conference on Learning Representations (ICLR), 2018." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 351, + 545, + 411 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 351, + 545, + 411 + ], + "spans": [ + { + "bbox": [ + 307, + 351, + 545, + 411 + ], + "type": "text", + "content": "[33] Milad Nasr, Reza Shokri, and Amir Houmansadr. Machine Learning with Membership Privacy using Adversarial Regularization. In ACM SIGSAC Conference on Computer and Communications Security (CCS), pages 634-646. ACM, 2018." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 418, + 545, + 489 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 418, + 545, + 489 + ], + "spans": [ + { + "bbox": [ + 307, + 418, + 545, + 489 + ], + "type": "text", + "content": "[34] Milad Nasr, Reza Shokri, and Amir Houmansadr. Comprehensive Privacy Analysis of Deep Learning: Passive and Active White-box Inference Attacks against Centralized and Federated Learning. In IEEE Symposium on Security and Privacy (S&P), pages 1021-1035. IEEE, 2019." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 498, + 545, + 558 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 498, + 545, + 558 + ], + "spans": [ + { + "bbox": [ + 307, + 498, + 545, + 558 + ], + "type": "text", + "content": "[35] Yuval Netzer, Tao Wang, Adam Coates, Alessandro Bissacco, Bo Wu, and Andrew Y. Ng. Reading Digits in Natural Images with Unsupervised Feature Learning. In Annual Conference on Neural Information Processing Systems (NIPS). NIPS, 2011." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 566, + 545, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 566, + 545, + 613 + ], + "spans": [ + { + "bbox": [ + 307, + 566, + 545, + 613 + ], + "type": "text", + "content": "[36] Seong Joon Oh, Max Augustin, Bernt Schiele, and Mario Fritz. Towards Reverse-Engineering Black-Box Neural Networks. In International Conference on Learning Representations (ICLR), 2018." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 622, + 545, + 681 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 622, + 545, + 681 + ], + "spans": [ + { + "bbox": [ + 307, + 622, + 545, + 681 + ], + "type": "text", + "content": "[37] Tribhuvanesh Orekondy, Bernt Schiele, and Mario Fritz. Knockoff Nets: Stealing Functionality of Black-Box Models. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 4954–4963. IEEE, 2019." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 689, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 689, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 307, + 689, + 545, + 712 + ], + "type": "text", + "content": "[38] Nicolas Papernot, Patrick D. McDaniel, Somesh Jha, Matt Fredrikson, Z. Berkay Celik, and Ananthram" + } + ] + } + ], + "index": 22 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "text", + "content": "16382" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 69, + 72, + 287, + 119 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 72, + 287, + 119 + ], + "spans": [ + { + "bbox": [ + 69, + 72, + 287, + 119 + ], + "type": "text", + "content": "Swami. The Limitations of Deep Learning in Adversarial Settings. In IEEE European Symposium on Security and Privacy (Euro S&P), pages 372-387. IEEE, 2016." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 130, + 287, + 214 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 130, + 287, + 214 + ], + "spans": [ + { + "bbox": [ + 48, + 130, + 287, + 214 + ], + "type": "text", + "content": "[39] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning Transferable Visual Models From Natural Language Supervision. In International Conference on Machine Learning (ICML), pages 8748-8763. PMLR, 2021." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 224, + 287, + 294 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 224, + 287, + 294 + ], + "spans": [ + { + "bbox": [ + 48, + 224, + 287, + 294 + ], + "type": "text", + "content": "[40] Alexandre Sablayrolles, Matthijs Douze, Cordelia Schmid, Yann Ollivier, and Hervé Jégou. White-box vs Black-box: Bayes Optimal Strategies for Membership Inference. In International Conference on Machine Learning (ICML), pages 5558-5567. PMLR, 2019." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 306, + 287, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 306, + 287, + 353 + ], + "spans": [ + { + "bbox": [ + 48, + 306, + 287, + 353 + ], + "type": "text", + "content": "[41] Ahmed Salem, Michael Backes, and Yang Zhang. Don't Trigger Me! A Triggerless Backdoor Attack Against Deep Neural Networks. CoRR abs/2010.03282, 2020." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 363, + 287, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 363, + 287, + 435 + ], + "spans": [ + { + "bbox": [ + 48, + 363, + 287, + 435 + ], + "type": "text", + "content": "[42] Ahmed Salem, Yang Zhang, Mathias Humbert, Pascal Berrang, Mario Fritz, and Michael Backes. MLLeaks: Model and Data Independent Membership Inference Attacks and Defenses on Machine Learning Models. In Network and Distributed System Security Symposium (NDSS). Internet Society, 2019." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 445, + 287, + 492 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 445, + 287, + 492 + ], + "spans": [ + { + "bbox": [ + 48, + 445, + 287, + 492 + ], + "type": "text", + "content": "[43] Zeyang Sha, Xinlei He, Pascal Berrang, Mathias Humbert, and Yang Zhang. Fine-Tuning Is All You Need to Mitigate Backdoor Attacks. CoRR abs/2212.09067, 2022." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 503, + 287, + 551 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 503, + 287, + 551 + ], + "spans": [ + { + "bbox": [ + 48, + 503, + 287, + 551 + ], + "type": "text", + "content": "[44] Yun Shen, Xinlei He, Yufei Han, and Yang Zhang. Model Stealing Attacks Against Inductive Graph Neural Networks. In IEEE Symposium on Security and Privacy (S&P), pages 1175-1192. IEEE, 2022." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 561, + 287, + 620 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 561, + 287, + 620 + ], + "spans": [ + { + "bbox": [ + 48, + 561, + 287, + 620 + ], + "type": "text", + "content": "[45] Reza Shokri, Marco Stronati, Congzheng Song, and Vitaly Shmatikov. Membership Inference Attacks Against Machine Learning Models. In IEEE Symposium on Security and Privacy (S&P), pages 3-18. IEEE, 2017." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 632, + 287, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 632, + 287, + 678 + ], + "spans": [ + { + "bbox": [ + 48, + 632, + 287, + 678 + ], + "type": "text", + "content": "[46] Liwei Song and Prateek Mittal. Systematic Evaluation of Privacy Risks of Machine Learning Models. In USENIX Security Symposium (USENIX Security). USENIX, 2021." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 689, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 689, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 689, + 287, + 712 + ], + "type": "text", + "content": "[47] Florian Tramér, Alexey Kurakin, Nicolas Papernot, Ian Goodfellow, Dan Boneh, and Patrick McDaniel." + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 72, + 545, + 590 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 328, + 72, + 545, + 107 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 328, + 72, + 545, + 107 + ], + "spans": [ + { + "bbox": [ + 328, + 72, + 545, + 107 + ], + "type": "text", + "content": "Ensemble Adversarial Training: Attacks and Defenses. In International Conference on Learning Representations (ICLR), 2017." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 307, + 116, + 545, + 175 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 116, + 545, + 175 + ], + "spans": [ + { + "bbox": [ + 307, + 116, + 545, + 175 + ], + "type": "text", + "content": "[48] Florian Tramér, Fan Zhang, Ari Juels, Michael K. Reiter, and Thomas Ristenpart. Stealing Machine Learning Models via Prediction APIs. In USENIX Security Symposium (USENIX Security), pages 601-618. USENIX, 2016." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 307, + 184, + 545, + 219 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 184, + 545, + 219 + ], + "spans": [ + { + "bbox": [ + 307, + 184, + 545, + 219 + ], + "type": "text", + "content": "[49] Aäron van den Oord, Yazhe Li, and Oriol Vinyals. Representation Learning with Contrastive Predictive Coding. CoRR abs/1807.03748, 2018." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 307, + 227, + 545, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 227, + 545, + 274 + ], + "spans": [ + { + "bbox": [ + 307, + 227, + 545, + 274 + ], + "type": "text", + "content": "[50] Binghui Wang and Neil Zhenqiang Gong. Stealing Hyperparameters in Machine Learning. In IEEE Symposium on Security and Privacy (S&P), pages 36-52. IEEE, 2018." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 283, + 545, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 283, + 545, + 331 + ], + "spans": [ + { + "bbox": [ + 307, + 283, + 545, + 331 + ], + "type": "text", + "content": "[51] Bang Wu, Xiangwen Yang, Shirui Pan, and Xingliang Yuan. Model Extraction Attacks on Graph Neural Networks: Taxonomy and Realization. CoRR abs/2010.12751, 2020." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 339, + 545, + 410 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 339, + 545, + 410 + ], + "spans": [ + { + "bbox": [ + 307, + 339, + 545, + 410 + ], + "type": "text", + "content": "[52] Huijun Wu, Chen Wang, Yuriy Tyshetskiy, Andrew Docherty, Kai Lu, and Liming Zhu. Adversarial Examples for Graph Data: Deep Insights into Attack and Defense. In International Joint Conferences on Artificial Intelligence (IJCAI), pages 4816-4823. IJCAI, 2019." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 418, + 545, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 418, + 545, + 479 + ], + "spans": [ + { + "bbox": [ + 307, + 418, + 545, + 479 + ], + "type": "text", + "content": "[53] Zhirong Wu, Yuanjun Xiong, Stella X. Yu, and Dahua Lin. Unsupervised Feature Learning via Non-Parametric Instance Discrimination. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 3733–3742. IEEE, 2018." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 487, + 545, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 487, + 545, + 533 + ], + "spans": [ + { + "bbox": [ + 307, + 487, + 545, + 533 + ], + "type": "text", + "content": "[54] Han Xiao, Kashif Rasul, and Roland Vollgraf. Fashion-MNIST: a Novel Image Dataset for Benchmarking Machine Learning Algorithms. CoRR abs/1708.07747, 2017." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 542, + 545, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 542, + 545, + 590 + ], + "spans": [ + { + "bbox": [ + 307, + 542, + 545, + 590 + ], + "type": "text", + "content": "[55] Ziqing Yang, Zeyang Sha, Michael Backes, and Yang Zhang. From Visual Prompt Learning to Zero-Shot Transfer: Mapping Is All You Need. CoRR abs/2303.05266, 2023." + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16383" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Canonical Fields_ Self-Supervised Learning of Pose-Canonicalized Neural Fields/54da6270-9672-4d8f-9893-3339952d04d6_content_list.json b/2023/Canonical Fields_ Self-Supervised Learning of Pose-Canonicalized Neural Fields/54da6270-9672-4d8f-9893-3339952d04d6_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..85886655eb864e27d8e50cb91d8fc8c0d978d54c --- /dev/null +++ b/2023/Canonical Fields_ Self-Supervised Learning of Pose-Canonicalized Neural Fields/54da6270-9672-4d8f-9893-3339952d04d6_content_list.json @@ -0,0 +1,1558 @@ +[ + { + "type": "text", + "text": "Canonical Fields: Self-Supervised Learning of Pose-Canonicalized Neural Fields", + "text_level": 1, + "bbox": [ + 80, + 130, + 888, + 151 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Rohith Agaram1", + "bbox": [ + 153, + 179, + 287, + 199 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Shaurya Dewan", + "bbox": [ + 326, + 180, + 460, + 198 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Madhava Krishna1", + "bbox": [ + 344, + 199, + 491, + 214 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Rahul Sajnani", + "bbox": [ + 501, + 180, + 624, + 198 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Adrien Poulenard3", + "bbox": [ + 663, + 180, + 813, + 196 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Srinath Sridhar", + "bbox": [ + 534, + 199, + 663, + 214 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ RRC, IIIT-Hyderabad", + "bbox": [ + 241, + 217, + 395, + 234 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ Brown University", + "bbox": [ + 428, + 217, + 555, + 234 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3Stanford University", + "bbox": [ + 588, + 217, + 727, + 234 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ivl.cs.brown.edu/projects/canonicalfields", + "bbox": [ + 320, + 237, + 643, + 253 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/3d97b753ffa37973fda8c606cdd8ac9d31052170813520d893589850268afb99.jpg", + "image_caption": [ + "Input NeRFs" + ], + "image_footnote": [], + "bbox": [ + 80, + 279, + 483, + 402 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/10d00d5ec06704c055fab2229ceebf64850b7ae311344eae598010bbdc85132c.jpg", + "image_caption": [ + "Canonical Fields", + "Figure 1. We present Canonical Field Network (CaFi-Net), a self-supervised method for 3D position and orientation (pose) canonicalization of objects represented as neural fields. We specifically focus on neural radiance fields (NeRFs) fitted to raw RGB images of arbitrarily posed objects (left), and obtain a canonical field (fixed novel view shown on right) where all objects in a category are consistently positioned and oriented." + ], + "image_footnote": [], + "bbox": [ + 483, + 279, + 890, + 402 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 465, + 312, + 481 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Coordinate-based implicit neural networks, or neural fields, have emerged as useful representations of shape and appearance in 3D computer vision. Despite advances, however, it remains challenging to build neural fields for categories of objects without datasets like ShapeNet that provide \"canonicalized\" object instances that are consistently aligned for their 3D position and orientation (pose). We present Canonical Field Network (CaFi-Net), a self-supervised method to canonicalize the 3D pose of instances from an object category represented as neural fields, specifically neural radiance fields (NeRFs). CaFi-Net directly learns from continuous and noisy radiance fields using a Siamese network architecture that is designed to extract equivariant field features for category-level canonicalization. During inference, our method takes pre-trained neural radiance fields of novel object instances at arbitrary 3D pose and estimates a canonical field with consistent 3D pose across the entire category. Extensive experiments on a new dataset of 1300 NeRF models across 13 object categories show that our method matches or exceeds the performance of 3D point cloud-based methods.", + "bbox": [ + 73, + 497, + 473, + 816 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 845, + 207, + 859 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Neural fields [59]—coordinate-based neural networks that implicitly parameterize signals—have recently gained", + "bbox": [ + 76, + 869, + 468, + 901 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "significant attention as representations of 3D shape [6,22, 28], view-dependent appearance [24, 42], and motion [25]. In particular, neural radiance fields (NeRFs) [24], have been successfully used in problems such as novel view synthesis [3,4,66], scene geometry extraction [55, 61], capturing dynamic scenes [19, 29, 30, 33, 52], 3D semantic segmentation [53, 68], and robotics [1, 13, 21].", + "bbox": [ + 496, + 467, + 890, + 573 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Despite the progress, it remains challenging to build neural fields that represent an entire category of objects. Previous methods sidestep the problem by overfitting on a single instance [24], or learning [22, 28, 63] on datasets like ShapeNet [5] that contain objects that are manually canonicalized – oriented consistently for 3D position and orientation (3D pose) across a category. This strong supervision makes it easier to learn over categories, but limits their application to data that contain these labels. Recent work has proposed methods for self-supervised learning of 3D pose canonicalization [40, 43, 47], however, these operate on 3D point clouds, meshes, or voxels – but not neural fields.", + "bbox": [ + 496, + 578, + 892, + 758 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this paper, we present Canonical Field Network (CaFi-Net), a self-supervised method for category-level canonicalization of the 3D position and orientation of objects represented as neural fields, specifically neural radiance fields. Canonicalizing neural fields is challenging because, unlike 3D point clouds or meshes, neural fields are continuous, noisy, and hard to manipulate since they are parameterized as the weights of a neural network [60]. To address these challenges, we first extend the notion of equiv", + "bbox": [ + 496, + 763, + 893, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "4500", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "arience to continuous vector fields and show how networks for processing 3D point clouds [50] can be extended to operate directly on neural radiance fields. We design CaFi-Net as a Siamese network that contains layers to extract equivariant features directly on vector fields. These field features are used to learn a canonical frame that is consistent across instances in the category. During inference, our method takes as input neural radiance fields of object instances from a category at arbitrary pose and estimates a canonical field that is consistent across the category. To handle noise in radiance fields from NeRF, our method incorporates density-based feature weighting and foreground-background clustering.", + "bbox": [ + 75, + 90, + 472, + 287 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our approach learns canonicalization without any supervision labels on a new dataset of 1300 pre-trained NeRF models of 13 common ShapeNet categories in arbitrary 3D pose (see Figure 1). We introduce several self-supervision loss functions that encourage the estimation of a consistent canonical pose. In addition, we present extensive quantitative comparisons with baselines and other methods on standardized canonicalization metrics [37] over 13 object categories. In particular, we show that our approach matches or exceeds the performance of 3D point cloud-based methods. This enables the new capability of directly operating on neural fields rather than converting them to point clouds for canonicalization. To sum up, we contribute:", + "bbox": [ + 75, + 287, + 472, + 484 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Canonical Field Network (CaFi-Net), the first method for self-supervised canonicalization of the 3D position and orientation (pose) of objects represented as neural radiance fields.", + "- A Siamese neural network architecture with equivariant feature extraction layers that are designed to directly operate on continuous and noisy radiance fields from NeRF.", + "- A public dataset of 1300 NeRF models from 13 ShapeNet categories including posed images, and weights for evaluating canonicalization performance." + ], + "bbox": [ + 96, + 494, + 467, + 662 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 76, + 678, + 218, + 694 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We focus our review of related work on neural fields, supervised canonicalization, equivariant neural network architectures, and self-supervised canonicalization.", + "bbox": [ + 75, + 704, + 467, + 747 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Neural Fields: Neural fields are emerging as useful representations for solving problems such as novel view synthesis [3, 4, 15, 24, 66], shape encoding and reconstruction [22, 28], dynamic reconstruction [19, 29, 30], appearance modeling [23, 36, 45], and human motion modeling [8, 18, 46, 51]. Generalization of neural fields to object categories remains difficult, but some methods have used pre-canonicalized datasets to circumvent this problem [63]. Neural fields are also used outside of computer vision, for instance in reconstructing proteins [69], physics [35], or au", + "bbox": [ + 75, + 750, + 467, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "dio [11]. Please see [49, 59] for more details. In this paper, our focus is on canonicalizing for the 3D pose of neural fields, specifically, neural radiance fields (NeRF) [24].", + "bbox": [ + 496, + 90, + 890, + 135 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Supervised Canonicalization: Datasets such as ShapeNet [5] and ModelNet40 [58] have 3D shapes that are manually pre-canonicalized. This inductive bias aids category-level generalization in problems such as 3D reconstruction [12, 22, 28, 48]. We can also use these datasets to formulate canonicalization as a supervised learning problem [54] enabling applications such as 6 degree-of-freedom object pose estimation, multi-view reconstruction [17, 44], and reconstruction of articulating [20, 65] and non-rigid objects [64]. However, our goal is to canonicalize without using manual pose labels.", + "bbox": [ + 496, + 136, + 890, + 301 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Equivariant Neural Networks: Pose-equivariant networks are equivariant to input pose [9, 50, 56, 57] by design. Some of these methods use Spherical Harmonic functions [10, 50, 56, 57], or vector neurons [9] to extract equivariant features. Equivariance is closely related to canonicalization since a canonical pose is also equivariant. Thus, previous methods have used pose-equivariance for canonicalization [37, 41, 43, 47]. However, these methods have thus far been limited to 3D point clouds, voxels, or meshes.", + "bbox": [ + 496, + 303, + 890, + 436 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Self-Supervised Canonicalization: Recent research has shown that self/weak supervision is sufficient for learning pose canonicalization on point clouds [37, 38, 43, 47], 2D key-points [26], and images [27]. None of these previous self-supervised methods can operate directly on neural fields – to the best of our knowledge, ours is the first.", + "bbox": [ + 496, + 438, + 890, + 527 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3. Background", + "text_level": 1, + "bbox": [ + 500, + 542, + 627, + 559 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We first provide background on the essential components and key ideas behind our method.", + "bbox": [ + 496, + 568, + 888, + 595 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Neural Radiance Fields (NeRF): Given posed RGB images, NeRF and its variants [23, 24] learn a neural network to synthesize novel views of complex scenes. At inference, for any 3D location and viewing direction, it estimates a direction-independent density value $\\sigma$ , and a direction-dependent color $c$ . Although designed for novel view synthesis, NeRF implicitly learns the 3D geometry and view-dependent appearance of scenes making it a useful representation of shape and appearance. While NeRFs excel at fitting a single object instance or scene, it struggles to generalize to object categories partly because objects can be in arbitrary 3D poses – thus, existing methods use precanonicalized datasets [63]. In this paper, we provide a way to canonicalize 3D pose without any supervision for objects represented as neural radiance fields. While our method is aimed at NeRFs, it can be extended to the 3D pose canonicalization of any neural field.", + "bbox": [ + 496, + 599, + 890, + 853 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Equivalence and Canonicalization: A function $\\Gamma$ over $x$ is said to be equivariant to a group operation $\\mathcal{O}$ if its output changes by a fixed mapping $M$ for any $\\mathcal{O}$ operating on", + "bbox": [ + 496, + 854, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "4501", + "bbox": [ + 480, + 944, + 513, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/4f7702db1fda005750345acb1f467aed69d3d49e34e8a2e86cc59ff2b522166f.jpg", + "image_caption": [ + "Figure 2. Rotation Equivariance in Vector Fields. Rotation equivariance in vector fields $\\mathcal{F}$ (here, a simulated 2D magnetic field) require two operations: (1) update the position of the green arrow (the vector), and (2) rotating the green arrow at its new position. A function $\\Gamma$ operating on $\\mathcal{F}$ is rotationally equivariant over the field if and only if $\\Gamma[(\\mathcal{R} \\cdot \\mathcal{F})(\\mathbf{x})] = M(\\mathcal{R})\\Gamma[\\mathcal{F}(\\mathcal{R}^{-1}\\mathbf{x})]$ , where $\\mathcal{R}$ is an $SO(2)$ rotation, and $M$ is a rotation-dependent mapping function." + ], + "image_footnote": [], + "bbox": [ + 81, + 87, + 472, + 244 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "the input $x$ , i.e., $\\Gamma(\\mathcal{O}x) = M(\\mathcal{O})\\Gamma(x)$ . In this paper, we are interested in equivariance to the $SE(3)$ group denoting 3D position and orientation (pose) – in particular we focus on the rotation group $SO(3)$ since 3D translation equivariance is readily achieved through mean centering [26]. Pose canonicalization is closely related to equivariance – our goal is to find an equivariant canonical 3D rigid transformation that maps the neural field of an object in an arbitrary pose to a canonical pose. We call the outputs of our method canonical fields – neural fields that are consistently oriented across a category of shapes.", + "bbox": [ + 75, + 397, + 468, + 565 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Rotation Equivalence in Vector Fields: Unlike 3D point clouds, rotation equivariance in fields is more involved. Consider a 2D vector field $\\mathcal{F}$ defined $\\forall \\mathbf{x} \\in \\mathbb{R}^2$ (see Figure 2). Rotating this field requires two operations: (1) updating the positions of vectors in the field to new rotated positions, and (2) rotating the directions of the vectors. A function $\\Gamma$ operating on $\\mathcal{F}$ is rotationally equivariant over the field if and only if $\\Gamma[(\\mathcal{R} \\cdot \\mathcal{F})(\\mathbf{x})] = M(\\mathcal{R})\\Gamma[\\mathcal{F}(\\mathcal{R}^{-1}\\mathbf{x})]$ , where $\\mathcal{R}$ is an $SO(2)$ rotation, and $M$ is a rotation-dependent mapping function. In this paper, we operate on NeRF's density field (a scalar field) but also extract equivariant features that can be vector-valued fields.", + "bbox": [ + 75, + 566, + 470, + 748 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Tensor Field Networks (TFNs): Many methods have been proposed for using equivariant features for rotation canonicalization of 3D data, for instance, spherical CNNs [7, 43], vector neurons [9], or capsule networks [47, 67]. However, these and other methods [38] are limited to 3D point cloud or voxel representations. In this paper, we extend Tensor Field Networks [34, 50] pose canonicalization directly on samples from NeRF's density field. TFNs operate by computing Type- $\\ell$ real-valued spherical harmonic functions", + "bbox": [ + 75, + 763, + 468, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "that are rotation equivariant: $D^{\\ell}(R)Y^{\\ell}(X) = Y^{\\ell}(Rx)$ . Where $D^{\\ell}:\\mathrm{SO}(3)\\to \\mathrm{SO}(2\\ell +1)$ is the Wigner matrix of type $\\ell$ (please see [34, 50] for details). For a 3D point set $P\\in \\mathbb{R}^{N\\times 3}$ , a rotation-equivariant TFN convolution layer (EQConv) at point $p\\in P$ is defined as: $\\mathrm{EQConv}^J (p) = Q^{(n,\\ell),J}\\left(\\sum_{y\\in^{2R}\\mathcal{N}}k_r^n (p - y)\\otimes s^\\ell [y]\\right)$ . Where $^{2R}\\mathcal{N}$ is a set containing neighbors of point $p$ at twice the resolution and $Q^{(n,\\ell),J}(\\cdot)$ is the Clebsch Gordon decomposition to combine type- $n$ kernel $k_r^n$ and type- $\\ell$ equivariant signal $s^\\ell$ . The EQConv $(\\cdot)$ layer aggregates type- $\\ell$ feature $s^\\ell$ at multiple receptive fields along with learnable weights to perform equivariant convolution. Our method uses the EQConv layer as the fundamental building block to canonicalize continuous fields.", + "bbox": [ + 496, + 89, + 893, + 309 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4. Method", + "text_level": 1, + "bbox": [ + 500, + 319, + 591, + 335 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our method assumes that we have pre-trained NeRF models of scenes containing single object instances at arbitrary poses from common shape categories. Our scenes contain primarily the object, but also some background - perfect foreground-background segmentation is not available. Our goal is to build a method that estimates a canonicalizing $SE(3)$ transformation that aligns these object NeRFs to a consistent category-level canonical field.", + "bbox": [ + 496, + 345, + 890, + 465 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Rather than operate on 3D point clouds, meshes, or voxels, our method (see Figure 3) operates directly on samples from the continuous neural radiance field. During training, we learn canonical fields from a pre-trained NeRF dataset in a fully self-supervised manner. We use a Siamese network architecture that extracts equivariant field features for canonicalization (see Figure 3). During inference, given a previously unseen NeRF model of a new object instance in an arbitrary pose, our method estimates the transformation that maps it to the canonical field.", + "bbox": [ + 496, + 467, + 892, + 617 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4.1. NeRF Sampling", + "text_level": 1, + "bbox": [ + 500, + 625, + 660, + 642 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "For a pre-trained neural radiance field that maps 3D position $(x, y, z)$ and direction $\\theta, \\phi$ to color and density $(\\mathbf{c}, \\sigma)$ , the input to our method consists of samples on the density field. We do not use color because it is direction-dependent in NeRF and also varies for different instances.", + "bbox": [ + 496, + 647, + 890, + 720 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Preprocessing: During both training and inference, we uniformly sample the pre-trained NeRF density field to find the object center and its bounding box (see supplementary document for details). We then sample a uniform grid in the density field of the object's bounding box. We empirically choose uniform grid sampling over fully random sampling since we found no difference. The uniformly sampled density grid is then normalized to obtain density values $\\sigma_{D}$ in the continuous range [0, 1], i.e.,", + "bbox": [ + 496, + 723, + 893, + 859 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {\\sigma} _ {D} := \\left\\{1 - e x p \\left(- d \\cdot \\Gamma_ {r} ^ {\\sigma} (x)\\right) \\mid x \\in \\boldsymbol {X} \\right\\}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 557, + 880, + 890, + 897 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "4502", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/d39ec9a4d2ad86ff29ef52690ae451f656d06e9adfc3938575f17f23d2437a23.jpg", + "image_caption": [ + "Figure 3. CaFi-Net samples a density field from NeRF and uses its density, position and density gradients as input signals (A) to canonicalize the field. We predict rotation equivariant features and weigh them by density (B) to guide our learning from the occupied regions of the scene. We then compute an invariant embedding by taking a dot product between equivariant features. This invariant embedding is used to canonicalize the field that enables rendering all the objects in the canonical frame (E). Our method also applies an inter-instance consistency loss (D) that aligns different instances of the same category in the canonical frame. We do not assume precanonicalized fields, and canonicalize in a self-supervised manner." + ], + "image_footnote": [], + "bbox": [ + 89, + 88, + 885, + 378 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $X$ is the regularly sampled grid within the object bounding box and $d$ is the depth step size used in the coarse NeRF sampling. Translation and scale canonicalization is simply achieved by moving the bounding box's center [26] and scaling it, so our next concern is on rotation canonicalization.", + "bbox": [ + 75, + 486, + 470, + 575 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.2. CaFi-Net", + "text_level": 1, + "bbox": [ + 76, + 588, + 187, + 603 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "CaFi-Net (Canonical Field Network) is our method for 3D rotation canonicalization using NeRF's density field (red box in Figure 3). We now describe its components.", + "bbox": [ + 75, + 612, + 468, + 657 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Signal Representation: At each of the sampled density field locations, we have several choices of signals to use as the input for CaFi-Net. For instance, we can (1) use the NeRF density value directly, (2) use the $xyz$ location of the sampled field, or (3) use the gradient of the density field. Densities alone do not capture the position and directional components of the field, so we use a combination of densities and gradients of densities at query locations of the grid as our input signal for CaFi-Net. Gradients of density capture the object surface and help in canonicalization (see ablation study in Section 5).", + "bbox": [ + 75, + 657, + 468, + 823 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "**Equivariant Convolution:** To canonicalize for rotation, we leverage equivariance properties of spherical harmonic functions to obtain rotation-equivariant learnable features on the field [34, 50]. Type- $\\ell$ spherical harmonic functions $(Y^{\\ell})$ are degree- $\\ell$ polynomials of points on the sphere that", + "bbox": [ + 75, + 825, + 470, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "are equivariant to $SO(2\\ell + 1)$ . The densities sampled from NeRF reconstructions are often noisy and contain outliers. To promote our method to learn only from occupied regions, we weigh each equivariant feature type by its density. This weighing has no effect on the equivariance of features (see supplementary document for proof). We have the option to weigh equivariant features with the local average of the density to smoothen the signal or to weigh directly by the density. We found weighing by density worked better than taking the local average density (see Section 5). We learn equivariant features of type $\\ell$ as:", + "bbox": [ + 496, + 486, + 893, + 654 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n{ } ^ { R } s ^ { \\ell } : = f _ { w } ( ^ { R } x ) \\left( W ^ { \\ell } ( \\operatorname { E Q C o n v } ^ { \\ell } ( \\nabla ( \\boldsymbol { \\sigma } _ { D } ( ^ { R } x ) ) ) ) + \\delta _ { 0 \\ell } b \\right) , \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 535, + 667, + 887, + 689 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nf _ {w} ^ {(R x)} := \\left\\{ \\begin{array}{l} \\operatorname {m e a n} _ {r} \\left(\\sigma_ {D} ^ {(R x)}\\right), \\quad o r \\\\ \\sigma_ {D} ^ {(R x)}, \\end{array} \\right. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 535, + 691, + 887, + 724 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where, $W^{\\ell}$ and $b$ are the weights and biases, ${}^{R}x$ is a 3D point queried at resolution $R$ and mean $r(\\sigma_D(Rx))$ is the average density at location $R_x$ from points sampled at radius $r$ . We hierarchically aggregate features at resolutions $\\frac{1}{2}, \\frac{1}{4}$ , and $\\frac{1}{8}$ to obtain global equivariant features. The EQConv convolution layer is the same as defined in Section 3. We employ non-linearities from [34] that preserve equivariance and capture better equivariant features. We also compute the max-pool of point-wise features of the last layer to obtain global type- $\\ell$ equivariant features $F^{\\ell}$ .", + "bbox": [ + 496, + 726, + 890, + 878 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Canonicalization: After obtaining the global type- $\\ell$", + "bbox": [ + 500, + 885, + 892, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4503", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "equivariant feature $F^{\\ell}$ , we compute its dot product with point-wise spherical harmonics $Z^{\\ell}$ scaled by their norms $Z^{\\ell}(\\mathbf{X}) = ||\\mathbf{X}||_{2}Y^{\\ell}(\\mathbf{X} / ||\\mathbf{X}||_{2})$ to obtain a pose-invariant embedding $H^{\\ell}$ [37]. As the two vectors $F^{\\ell}$ and $Z^{\\ell}$ are both equivariant to input rotation, their dot product is invariant. We use a linear layer on top of the pose-invariant embedding to estimate grid coordinates in the canonical frame $(P\\in \\mathbb{R}^{H\\times W\\times D\\times 3})$ for each point in $\\mathbf{X}$ and $M$ equivariant transformations $E\\in \\mathbb{R}^{M\\times 3\\times 3}$ [37] that orient the canonical coordinates to the input coordinates. We then choose the best canonicalizing transform $E_{b}$ and penalize it for canonicalization. The invariant embedding $H$ , grid coordinates $P$ and equivariant rotation $E$ are given as,", + "bbox": [ + 75, + 90, + 472, + 287 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nH ^ {\\ell} \\left(\\boldsymbol {\\sigma} _ {D}, \\boldsymbol {X}\\right) := \\langle F ^ {\\ell} \\left(\\boldsymbol {\\sigma} _ {D}\\right), Z ^ {\\ell} (\\boldsymbol {X}) \\rangle , \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 153, + 296, + 468, + 315 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nP := \\operatorname {M L P} _ {Y} \\left(H \\left(\\boldsymbol {\\sigma} _ {D}, \\boldsymbol {X}\\right)\\right), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 153, + 316, + 468, + 334 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nE := \\operatorname {M L P} _ {E} \\left(F \\left(\\boldsymbol {\\sigma} _ {D}, \\boldsymbol {X}\\right)\\right). \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 153, + 335, + 468, + 353 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Here, we have dropped the type- $\\ell$ notation for $H$ and $F$ as we concatenate all equivariant feature types.", + "bbox": [ + 75, + 363, + 468, + 393 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Siamese Network Architecture: Experimentally, canonicalizing noisy density fields is difficult without guidance on shape similarity within a category. We therefore use a Siamese training strategy [26] where two different object instances are forced to be consistently canonicalized (see the lower branch in Figure 3).", + "bbox": [ + 75, + 393, + 468, + 484 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Clustering: CaFi-Net predicts a canonical coordinate for every grid point, but unlike point clouds, fields have many unoccupied regions that do not provide any information. We penalize the network on foreground regions only by performing K-means clustering on the densities within the object bounding box with $K = 2$ and choose the cluster $C_f$ with a higher mean. Note that we still operate on continuous fields, but only cluster samples to guide our training.", + "bbox": [ + 75, + 484, + 470, + 606 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.3. Training", + "text_level": 1, + "bbox": [ + 76, + 614, + 181, + 630 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "CaFi-Net is trained on a large dataset of pre-trained NeRF models over 13 object categories (see Section 5). We use the following loss functions to train our model.", + "bbox": [ + 75, + 637, + 468, + 681 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Canonicalization Loss: To self-supervise our learning, we transform the predicted canonical grid coordinates to the input grid using the canonical rotation $E_{b}$ and compute a point-wise L2 distance between them. This loss forces the predicted invariant grid coordinates $P$ to reconstruct the shape and forces the predicted canonicalizing transform to be equivariant to the input transformation. The final loss is computed over the canonicalizing transform that minimizes this loss:", + "bbox": [ + 75, + 683, + 468, + 818 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nE _ {b} := \\min _ {j} \\left(\\operatorname {m e a n} _ {i \\in C _ {f}} \\left\\| X _ {i} - E _ {j} P _ {i} \\right\\| _ {2} ^ {2}\\right), \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 133, + 825, + 468, + 852 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {c a n o n}} := \\operatorname {m e a n} _ {i \\in C _ {f}} \\left| \\left| \\boldsymbol {X} _ {i} - E _ {b} P _ {i} \\right| \\right| _ {2} ^ {2}, \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 135, + 856, + 468, + 876 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $C_f$ is a set of points belonging to the foreground.", + "bbox": [ + 76, + 885, + 446, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "**Orthonormality Loss:** The predicted equivariant transformations $E$ should have orthonormal vectors and we use the following to force orthonormality using:", + "bbox": [ + 498, + 90, + 890, + 137 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nU _ {j}, D _ {j}, V _ {j} := \\operatorname {S V D} \\left(E _ {j}\\right), \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 573, + 148, + 890, + 165 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {o r t h o}} := \\frac {1}{M} \\sum_ {j} \\left| \\left| E _ {j} - U _ {j} V _ {j} ^ {T} \\right| \\right| _ {2}, \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 573, + 167, + 890, + 204 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\operatorname{SVD}(E_j)$ computes the singular value decomposition of the $j^{th}$ equivariant transformation.", + "bbox": [ + 498, + 215, + 888, + 244 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Siamese Shape Loss: We use a Siamese loss that penalizes for consistency between different shapes belonging to the same category. Given two un-canonicalized fields in different frames of reference $\\sigma_{D}^{1}$ and $\\sigma_{D}^{2}$ , we canonicalize both the fields to the canonical frame and compute chamfer distance between their predicted canonical coordinates for points in the foreground cluster $C_f$ , i.e.,", + "bbox": [ + 498, + 246, + 890, + 353 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {s i a m e s e}} := \\mathrm {C D} \\left(P \\left(\\boldsymbol {\\sigma} _ {D} ^ {1}\\right), P \\left(\\boldsymbol {\\sigma} _ {D} ^ {2}\\right)\\right). \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 560, + 363, + 890, + 382 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "This loss regularizes the training by ensuring that instances of the same category should be aligned in the canonical frame.", + "bbox": [ + 498, + 393, + 890, + 436 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Architecture and Hyper-parameters: CaFi-Net predicts a 128 dimensional invariant embedding for each point in space and performs convolution at resolutions $1/2$ , $1/4$ , and $1/8$ . Each equivariant convolution layer aggregates features from 512 neighboring points at twice the resolution. We then use equivariant non-linearities introduced in [34] by applying Batch Normalization [14] and ReLU activation [2] in the inverse spherical harmonic transform domain. We use a three-layer MLP with intermediate Batch Normalization and ReLU layers followed by a final linear layer to predict canonical coordinates. To train our network, we weigh the canonicalization loss $(\\mathcal{L}_{canon})$ the highest with weight 2.0 and 1.0 for all the other loss functions. All our models are written in PyTorch [31] and are trained for 300 epochs with a batch size of 2 on an Nvidia 1080-Ti GPU. We use the Adam Optimizer [16] with an initial learning rate of $6e - 4$ and L2 weight regularization of $1e - 5$ for all our experiments. Please see the supplementary document for more details.", + "bbox": [ + 498, + 438, + 892, + 724 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 500, + 739, + 632, + 757 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We now provide details on our extensive experiments to evaluate the quality of our pose canonicalization. To our knowledge, we are the first method for pose canonicalization of neural radiance fields which makes direct comparisons with other work challenging. Nonetheless, we have designed experiments to compare our method with three 3D point cloud-based canonicalization methods. Overall, we present three classes of experiments: (1) evaluation of the performance of our method on 13 different categories,", + "bbox": [ + 498, + 763, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "4504", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/1cfbfa9cb5d35d526940eea9e9ad2385c7c8fd304747111caaf33d4ee80b9fda.jpg", + "table_caption": [ + "Table 1. This table compares the canonicalization performance of our method (F - operating on fields) with other 3D point cloud-based methods (P) on three standard metrics (IC, CC and GEC) on our dataset of 13 categories. We compare with PCA, Canonical Capsules (CaCa) [47] and ConDor [37]. All metrics are multiplied by 100 for ease of reading. The top two performing methods are highlighted in magenta (best) and blue (second best). We are better than SOTA [37] on the Ground Truth Equivariance Consistency (GEC) and Category-Level Consistency (CC) with lower mean and median canonicalization error. However, we perform on par with ConDor [37] on the Instance-Level Consistency." + ], + "table_footnote": [], + "table_body": "
benchcabinetcarcellph.chaircouchfirearmlampmonitorplanespeakertablewater.avg.med.
Instance-Level Consistency (IC)↓
PCA (P)14.655.946.130.676.137.6315.0712.846.787.905.4010.3110.728.477.63
CaCa (P) [47]2.892.092.012.811.081.900.232.972.351.512.212.422.602.082.21
ConDor (P) [37]0.561.270.360.530.960.320.603.780.440.641.971.322.151.150.64
Ours (F)2.621.150.391.280.811.430.743.870.410.360.601.092.741.341.09
Category-Level Consistency (CC)↓
PCA (P)14.559.687.484.269.9013.4515.1311.8713.098.155.5110.4311.0710.3610.43
CaCa (P) [47]2.790.381.050.831.821.591.314.340.391.830.651.912.061.611.59
ConDor (P) [37]1.981.680.471.041.280.880.964.441.581.262.102.102.561.721.58
Ours (F)2.771.360.481.260.751.590.923.920.580.630.771.482.351.451.26
Ground Truth Equivalence Consistency (GEC)↓
PCA (P)14.405.866.321.646.328.1215.0112.496.828.165.3210.1711.278.618.12
CaCa (P) [47]3.652.282.402.741.902.221.284.692.482.092.223.012.792.592.40
ConDor (P) [37]2.121.790.501.171.340.931.044.581.591.312.222.242.671.811.59
Ours (F)3.261.510.541.430.941.901.014.370.630.660.811.633.021.671.43
", + "bbox": [ + 84, + 174, + 885, + 409 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "(2) comparison with other work, and (3) ablations to justify method design choices.", + "bbox": [ + 76, + 421, + 470, + 452 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Dataset: All our experiments use a new large synthetic NeRF dataset of shapes from 13 categories (that overlap with [12, 37, 47]) from the ShapeNet [5] dataset. For each category, we pick 100 instances from ShapeNet, rotate the shape randomly, and render 54 omnidirectional views sampled in a cube around the shape using Blender. Different from other datasets, we also simulate cluttered backgrounds. These posed views are used to train 100 NeRF models per category using the public PyTorch implementation of NeRF [62] for a total of 1300 NeRF models. To our knowledge, this is one of the largest (compared to [11, 53]) $360^{\\circ}$ synthetic NeRF datasets - we will publicly release the simulator, raw posed images, and trained NeRF weights. Each NeRF model is trained for 400 epochs with 1024 randomly selected rays per iteration, a coarse sampling resolution of 64 points, and a fine sampling resolution of 128 points along each ray. Rather than train NeRF to maximize PSNR on a novel view, we fix the number of epochs for all models to enable a fair comparison between instances. As a consequence, some NeRF models may have a lower PSNR, but our goal is not to increase NeRF quality but rather improve canonicalization quality - hence we fix the number of epochs. For each category, we set aside $20\\%$ of the models for testing. Metrics: Since there are no known metrics for evaluating canonical fields, we resort to 3 metrics used for 3D point cloud canonicalization [37]: (1) Instance-Level Consistency (IC), a metric that measures how consistently we can canonicalize the same instance in different poses, (2) Category-", + "bbox": [ + 75, + 462, + 472, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Level Consistency (CC) to measure how consistently we canonicalize across different instances in the same category, and (3) Ground Truth Equivariance Consistency (GEC), a variant of the Ground Truth Consistency (GC) metric proposed in [37] to measure canonicalization performance compared to manual labels. We use the GEC because we observed that the original GC metric is prone to degeneracy when the canonicalizing transform is identity. We fix this issue by taking three point clouds $P_{i}, P_{j}$ , and $P_{k}$ from a canonicalized dataset $\\mathcal{P}$ and rotating $P_{i}$ by $R_{1}, P_{j}$ by $R_{2}$ where both $R_{1}$ and $R_{2}$ are random rotations. Let $\\mathcal{C}(P)$ predict a canonicalizing rotation for a point cloud $P$ , we then compute the Ground Truth Equivariance consistency", + "bbox": [ + 496, + 421, + 893, + 619 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "as: $\\mathrm{GEC} \\coloneqq \\frac{1}{|\\mathcal{P}|^3} \\sum_{P_i, P_j, P_k \\in \\mathcal{P}} \\mathrm{CD}(\\mathcal{C}(R_1P_i)R_1P_k, \\mathcal{C}(R_2P_j)R_2P_k)$ , (12) where CD refers to the Chamfer distance. Here, we apply canonicalizing transforms of shape $P_i$ and $P_j$ to the shape $P_k$ . This modified metric will not be degenerate for identity canonicalizing transforms and is more rigorous in evaluating ground truth consistency. To compute our metrics, we used point clouds sampled from the Shapenet meshes [5] using poisson disk sampling.", + "bbox": [ + 496, + 627, + 898, + 763 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1. Evaluation & Comparisons", + "text_level": 1, + "bbox": [ + 498, + 770, + 746, + 787 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We first analyze the performance of our method on the three metrics across 13 different categories. In Table 1, we highlight our method that operates on a field as Ours (F). Our method performs well across all categories with low variance between different categories. We observe similar trends across metrics for a set of categories (e.g., firearms,lamps) indicating that category shape plays a", + "bbox": [ + 496, + 795, + 893, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "4505", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/ab570f5815e6488dc917d430a417903df7f3eb7066357e75f4fbb3419e7d0540.jpg", + "table_caption": [], + "table_footnote": [ + "(a) Choice of Signal Representation - Canonicalization metrics for using Gradients vs. $xyz$ locations as input signal. Gradients capture the object surface that help in canonicalization." + ], + "table_body": "
Categoryxyz vs. Gradient Signals
xyzGradient signal
Ground Truth Equivariance Consistency (GEC)↓
bench2.173.26
cellphone1.841.43
chair1.720.94
plane2.070.66
Average1.951.57
Instance-Level Consistency (IC)↓
bench2.042.62
cellphone1.61.28
chair1.40.81
plane1.880.36
Average1.731.26
Category-Level Consistency (CC)↓
bench1.992.77
cellphone1.491.26
chair1.520.75
plane1.90.63
Average1.721.35
", + "bbox": [ + 83, + 87, + 343, + 351 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/9bd5041d440aa74b96503a4414836d64a88eaa5a7f74d6311172097a47cd6687.jpg", + "table_caption": [], + "table_footnote": [ + "(b) Weighing Equivariant Signals by Local Average Density deteriorates the performance by smoothing out important details of the shape. We show canonicalization with (w) and without (w/o) weighing by the local averaged density." + ], + "table_body": "
CategoryLocal Average Density
w/ow
Ground Truth Equivariance Consistency (GEC)↓
bench3.263.38
cellphone1.431.63
chair0.941.3
plane0.660.64
Average1.571.73
Instance-Level Consistency (IC)↓
bench2.622.72
cellphone1.281.63
chair0.812.64
plane0.360.39
Average1.261.85
Category-Level Consistency (CC)↓
bench2.772.82
cellphone1.261.54
chair0.751.07
plane0.630.51
Average1.351.49
", + "bbox": [ + 354, + 87, + 614, + 351 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/38bb6101fee675aa122805e7ff3b28fd9dac0c43121a8ac47ccd4a35fc9abff3.jpg", + "table_caption": [], + "table_footnote": [ + "(c) Siamese Training improves performance on all canonicalization metrics on average. We show canonicalization performance with siamese and without (w/o) siamese training. The average of Ground Truth Equivariance Consistency GEC metric reduces to 1.57 from 1.86" + ], + "table_body": "
CategorySiamese Training
w/o siamesewith siamese
Ground Truth Equivariance Consistency (GEC)↓
bench3.523.26
cellphone1.611.43
chair1.020.94
plane1.310.66
Average1.861.57
Instance-Level Consistency (IC)↓
bench2.732.62
cellphone1.31.28
chair0.820.81
plane1.10.36
Average1.481.26
Category-Level Consistency (CC)↓
bench2.952.77
cellphone1.441.26
chair0.830.75
plane1.260.63
Average1.621.35
", + "bbox": [ + 625, + 87, + 885, + 351 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 2. Ablation studies to justify three key design choices: (a) signal representation, (b) density weighting, and (c) Siamese architecture.", + "bbox": [ + 76, + 431, + 890, + 446 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "role in canonicalization performance.", + "bbox": [ + 76, + 474, + 326, + 488 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Next, we compare our method against three benchmarks: Principal Component Analysis (PCA) [32], Canonical Capsules (CaCa) [47], and ConDor [37]. All prior methods canonicalize clean point clouds with all points on the surface. To have a fair comparison, we train the above methods on point clouds (P) sampled from NeRF scenes for points on and within the object. Most scenes have points in the range of 120-1200 when sampled uniformly on the $32^{3}$ grid. We resample the scene when points are less than 1024 and train both Canonical Capsules and ConDor on the resulting point clouds. CaFi-Net estimates a canonicalizing transform directly from the density field that we use to transform its corresponding ShapeNet point cloud for comparison against point canonicalizers. We follow the same protocol as [37] to measure and benchmark canonicalization performance for all the methods.", + "bbox": [ + 75, + 489, + 468, + 731 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Analysis: Table 1 compares the metrics for PCA (P), CaCa (P), ConDor (P), and CaFi-Net [Ours (F)]. We list point cloud canonicalizers with (P) and field canonicalizer as (F) in the table. On average, our method performs better than point cloud-based methods in the CC and GEC metrics and is on par with the state-of-the-art method (ConDor) in the IC metric. This is despite the fact that our method directly operates on the field. We observe that PCA underperforms in all categories suggesting that the noisy point clouds make it unreliable. Similarly we observe that CaCa fails to canonicalize in categories where high noise is observed. Given", + "bbox": [ + 75, + 734, + 470, + 902 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "that ConDor also uses TFNs in a different architecture, we conclude that the choice of TFNs makes canonicalization more robust to noise. Our method has the additional advantage of operating directly on the field. Please see Figure 4 for a qualitative comparison of the different methods.", + "bbox": [ + 496, + 474, + 890, + 550 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.2. Ablations", + "text_level": 1, + "bbox": [ + 500, + 559, + 609, + 574 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We conduct additional experiments to justify key design choices in our method. For all the ablation studies we use a smaller subset of our data consisting of 4 categories: bench, cellphone, chair, and airplane.", + "bbox": [ + 496, + 583, + 890, + 643 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Choice of Signal Representation: To select the appropriate signal representation to use (see Section 4), we conducted an ablation study using just the density field $xyz$ locations or its gradients. As seen in Table 2 (a), the average canonicalization error over 4 categories is lower when using gradient as compared to using density $xyz$ locations on all three metrics. We hypothesize that gradient of density weighs equivariant features closer to the surface of the object higher. This improves the canonicalization performance.", + "bbox": [ + 496, + 643, + 890, + 792 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Weighing Features by Local Average Density: Next, we justify why weighting the equivariant features using the density is helpful. Table 2 (b) shows results for weighting features by local average density compared to weighting features by density. Weighing features by local average smoothens out details that are necessary to canonicalize the volume. Our average ground truth canonicalization error", + "bbox": [ + 496, + 795, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "4506", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d570ced4afbb5858f75f9c7cf2ab418bfd5caed7eea2ab3266f24df6a4d21b34.jpg", + "image_caption": [ + "Input NeRFs", + "CaFi-Net (Ours)", + "PCA", + "Figure 4. This figure visually compares the canonicalization performance of our method with 3D point cloud-based methods. Input NeRFs shows one of the RGB views used to learn our NeRF input (notice how orientations of instances are random). CaFi-Net (Ours) shows the results of our canonical field rendered from a novel view unseen during NeRF training. PCA (blue), CaCa [47] (red), and ConDor [37]'s (green) results are also presented for the same instances. Our method matches or exceeds ConDor. We also show some failure cases for several instances (red box) which occur either due to thin structures or symmetry." + ], + "image_footnote": [], + "bbox": [ + 80, + 101, + 888, + 450 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "increases from 1.57 to 1.73 if we weigh features by local averaged densities. In future works, we will delve into better methods for reducing input signal noise that does not smoothen out the geometry details of objects to further improve our canonicalization.", + "bbox": [ + 75, + 527, + 468, + 602 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Siamese Architecture: Finally, we justify the need for a two-branch Siamese network architecture. Although our method can canonicalize even without this architecture, it makes it easier to learn over instances in a category as observed in [26]. Table 2 (c) compares our method on using a single branch architecture and a Siamese architecture. Clearly, the Siamese architecture helps improve results confirming our hypothesis. Enforcing different instances of the same category to align with each other reduces the Ground Truth Canonicalization error from 1.86 to 1.57. We also see a decrease in all the other metrics including Instance-Level Consistency.", + "bbox": [ + 75, + 604, + 470, + 785 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 76, + 799, + 194, + 814 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We presented Canonical Field Network (CaFi-Net), a method for self-supervised category-level canonicalization of the 3D pose of objects represented as neural radiance fields. CaFi-Net consists of a Siamese network architecture with rotation-equivariant convolution layers to extract", + "bbox": [ + 75, + 824, + 470, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "features for pose canonicalization. Our approach operates directly on NeRF's noisy but continuous density field. We train our method on a large dataset of 1300 NeRF models obtained for 13 common ShapeNet categories. During inference, our method is able to canonicalize arbitrarily oriented NeRFs. Experiments show that our method matches or outperforms 3D point cloud-based methods.", + "bbox": [ + 496, + 527, + 890, + 633 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Limitations & Future Work: Our approach has several limitations that future work should investigate. First, we require a $360^{\\circ}$ NeRF model of objects and cannot handle partial views from front-facing NeRFs. Second, canonicalizing densities with uncertainty is more difficult as the uncertainty increases for very thin/small structures that can be missed (see chair in Figure 4). Furthermore, we also inherit the issues that TFNs have with symmetry (see monitor case in Figure 4). In future work, we plan to extend our approach to create a large real dataset of common object categories without requiring manual pose canonicalization and combat the symmetry issue in CaFi-Net as in [39]", + "bbox": [ + 496, + 657, + 892, + 839 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments: This work was supported by NSF grant CNS-2038897, an AWS Cloud Credits award, NSF CloudBank, and a gift from Meta Reality Labs. We thank Chandradeep Pokhariya and Ishaan Shah.", + "bbox": [ + 496, + 839, + 893, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "4507", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Michal Adamkiewicz, Timothy Chen, Adam Caccavale, Rachel Gardner, Preston Culbertson, Jeannette Bohg, and Mac Schwager. Vision-only robot navigation in a neural radiance world. IEEE Robotics and Automation Letters, 7(2):4606-4613, 2022. 1", + "[2] Abien Fred Agarap. Deep learning using rectified linear units (relu). arXiv preprint arXiv:1803.08375, 2018. 5", + "[3] Jonathan T Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P Srinivasan. Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5855–5864, 2021. 1, 2", + "[4] Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Mip-nerf 360: Unbounded anti-aliased neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5470–5479, 2022. 1, 2", + "[5] Angel X. Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, Jianxiong Xiao, Li Yi, and Fisher Yu. ShapeNet: An Information-Rich 3D Model Repository. Technical Report arXiv:1512.03012 [cs.GR], Stanford University — Princeton University — Toyota Technological Institute at Chicago, 2015. 1, 2, 6", + "[6] Zhiqin Chen and Hao Zhang. Learning implicit fields for generative shape modeling. Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 1", + "[7] Taco S Cohen, Mario Geiger, Jonas Kohler, and Max Welling. Spherical cnns. arXiv preprint arXiv:1801.10130, 2018.3", + "[8] Enric Corona, Tomas Hodan, Minh Vo, Francesc Moreno-Noguer, Chris Sweeney, Richard Newcombe, and Lingni Ma. Lisa: Learning implicit shape and appearance of hands. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20533-20543, 2022. 2", + "[9] Congyue Deng, Or Litany, Yueqi Duan, Adrien Poulenard, Andrea Tagliasacchi, and Leonidas J. Guibas. Vector neurons: A general framework for so(3)-equivariant networks. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 12200-12209, October 2021. 2, 3", + "[10] Fabian B Fuchs, Daniel E Worrall, Volker Fischer, and Max Welling. Se (3)-transformers: 3d roto-translation equivariant attention networks. arXiv preprint arXiv:2006.10503, 2020. 2", + "[11] Ruohan Gao, Yen-Yu Chang, Shivani Mall, Li Fei-Fei, and Jiajun Wu. Objectfolder: A dataset of objects with implicit visual, auditory, and tactile representations. In Proceedings of the Conference on Robot Learning (CoRL), 2021. 2, 6", + "[12] Thibault Groueix, Matthew Fisher, Vladimir G Kim, Bryan C Russell, and Mathieu Aubry. A papier-mâché approach to learning 3d surface generation. In Proceedings of" + ], + "bbox": [ + 78, + 114, + 470, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "the IEEE conference on computer vision and pattern recognition, pages 216-224, 2018. 2, 6", + "[13] Jeffrey Ichnowski, Yahav Avigal, Justin Kerr, and Ken Goldberg. Dex-nerf: Using a neural radiance field to grasp transparent objects. Oct 2021. 1", + "[14] Sergey Ioffe and Christian Szegedy. Batch normalization: Accelerating deep network training by reducing internal covariate shift. In Francis Bach and David Blei, editors, Proceedings of the 32nd International Conference on Machine Learning, volume 37 of Proceedings of Machine Learning Research, pages 448-456, Lille, France, 07-09 Jul 2015. PMLR. 5", + "[15] Ajay Jain, Matthew Tancik, and Pieter Abbeel. Putting nerf on a diet: Semantically consistent few-shot view synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5885-5894, 2021. 2", + "[16] Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. CoRR, abs/1412.6980, 2015. 5", + "[17] Jiahui Lei, Srinath Sridhar, Paul Guerrero, Minhyuk Sung, Niloy Mitra, and Leonidas J Guibas. Pix2surf: Learning parametric 3d surface models of objects from images. In European Conference on Computer Vision, pages 121-138. Springer, 2020. 2", + "[18] Ruilong Li, Julian Tanke, Minh Vo, Michael Zollhofer, Jurgen Gall, Angjoo Kanazawa, and Christoph Lassner. Tava: Template-free animatable volumetric actors. arXiv preprint arXiv:2206.08929, 2022. 2", + "[19] Tianye Li, Mira Slavcheva, Michael Zollhoefer, Simon Green, Christoph Lassner, Changil Kim, Tanner Schmidt, Steven Lovegrove, Michael Goesele, Richard Newcombe, et al. Neural 3d video synthesis from multi-view video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5521-5531, 2022. 1, 2", + "[20] Xiaolong Li, He Wang, Li Yi, Leonidas J Guibas, A Lynn Abbott, and Shuran Song. Category-level articulated object pose estimation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3706-3715, 2020. 2", + "[21] Yunzhu Li, Shuang Li, Vincent Sitzmann, Pulkit Agrawal, and Antonio Torralba. 3d neural scene representations for visuomotor control. In Proceedings of Robotics: Science and Systems, 2021. 1", + "[22] Lars Mescheder, Michael Oechsle, Michael Niemeyer, Sebastian Nowozin, and Andreas Geiger. Occupancy networks: Learning 3d reconstruction in function space. In Proceedings IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2019. 1, 2", + "[23] Ben Mildenhall, Peter Hedman, Ricardo Martin-Brualla, Pratul P Srinivasan, and Jonathan T Barron. Nerf in the dark: High dynamic range view synthesis from noisy raw images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16190-16199, 2022. 2", + "[24] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view syn" + ], + "bbox": [ + 501, + 92, + 893, + 901 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "4508", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "thesis. In European conference on computer vision, pages 405-421. Springer, 2020. 1, 2", + "[25] Michael Niemeyer, Lars Mescheder, Michael Oechsle, and Andreas Geiger. Occupancy flow: 4d reconstruction by learning particle dynamics. In Proc. of the IEEE International Conf. on Computer Vision (ICCV), 2019. 1", + "[26] David Novotny, Nikhila Ravi, Benjamin Graham, Natalia Neverova, and Andrea Vedaldi. C3dpo: Canonical 3d pose networks for non-rigid structure from motion. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7688-7697, 2019. 2, 3, 4, 5, 8", + "[27] David Novotny, Roman Shapovalov, and Andrea Vedaldi. Canonical 3D Deformer Maps: Unifying parametric and non-parametric methods for dense weakly-supervised category reconstruction. In NeurIPS, 2020. 2", + "[28] Jeong Joon Park, Peter Florence, Julian Straub, Richard Newcombe, and Steven Lovegrove. Deepsdf: Learning continuous signed distance functions for shape representation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 165-174, 2019. 1, 2", + "[29] Keunhong Park, Utkarsh Sinha, Jonathan T. Barron, Sofien Bouaziz, Dan B Goldman, Steven M. Seitz, and Ricardo Martin-Brualla. Nerfies: Deformable neural radiance fields. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), 2021. 1, 2", + "[30] Keunhong Park, Utkarsh Sinha, Peter Hedman, Jonathan T. Barron, Sofien Bouaziz, Dan B Goldman, Ricardo MartinBrualla, and Steven M. Seitz. Hypernerf: A higher-dimensional representation for topologically varying neural radiance fields. arXiv preprint arXiv:2106.13228, 2021. 1, 2", + "[31] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, Alban Desmaison, Andreas Kopf, Edward Yang, Zachary DeVito, Martin Raison, Alykhan Tejani, Sasank Chilamkurthy, Benoit Steiner, Lu Fang, Junjie Bai, and Soumith Chintala. Pytorch: An imperative style, high-performance deep learning library. In H. Wallach, H. Larochelle, A. Beygelzimer, F. d'Alché-Buc, E. Fox, and R. Garnett, editors, Advances in Neural Information Processing Systems 32, pages 8024-8035. Curran Associates, Inc., 2019. 5", + "[32] Karl Pearson. Liii. on lines and planes of closest fit to systems of points in space. The London, Edinburgh, and Dublin philosophical magazine and journal of science, 2(11):559-572, 1901. 7", + "[33] Sida Peng, Junting Dong, Qianqian Wang, Shangzhan Zhang, Qing Shuai, Xiaowei Zhou, and Hujun Bao. Animatable neural radiance fields for modeling dynamic human bodies. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14314-14323, 2021. 1", + "[34] Adrien Poulenard and Leonidas J Guibas. A functional approach to rotation equivariant non-linearities for tensor field networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13174-13183, 2021. 3, 4, 5", + "[35] Maziar Raissi, Paris Perdikaris, and George E Karniadakis. Physics-informed neural networks: A deep learning framework for solving forward and inverse problems involving" + ], + "bbox": [ + 78, + 92, + 468, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "nonlinear partial differential equations. Journal of Computational Physics, 378:686-707, 2019. 2", + "[36] Viktor Rudnev, Mohamed Elgharib, William Smith, Lingjie Liu, Vladislav Golyanik, and Christian Theobalt. Nerf for outdoor scene relighting. In European Conference on Computer Vision (ECCV), 2022. 2", + "[37] Rahul Sajnani, Adrien Poulenard, Jivitesh Jain, Radhika Dua, Leonidas J Guibas, and Srinath Sridhar. Condor: Self-supervised canonicalization of 3d pose for partial shapes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16969-16979, 2022. 2, 5, 6, 7, 8", + "[38] Rahul Sajnani, AadilMehdi Sanchawala, Krishna Murthy Jatavallabhula, Srinath Sridhar, and K Madhava Krishna. Draco: Weakly supervised dense reconstruction and canonicalization of objects. In 2021 IEEE International Conference on Robotics and Automation (ICRA), pages 10302-10309. IEEE, 2021. 2, 3", + "[39] Ahyun Seo, Byungjin Kim, Suha Kwak, and Minsu Cho. Reflection and rotation symmetry detection via equivariant learning. In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 8", + "[40] Anthony Simeonov, Yilun Du, Andrea Tagliasacchi, Joshua B Tenenbaum, Alberto Rodriguez, Pulkit Agrawal, and Vincent Sitzmann. Neural descriptor fields: Se (3)-equivariant object representations for manipulation. In 2022 International Conference on Robotics and Automation (ICRA), pages 6394-6400. IEEE, 2022. 1", + "[41] Anthony Simeonov, Yilun Du, Andrea Tagliasacchi, Joshua B. Tenenbaum, Alberto Rodriguez, Pulkit Agrawal, and Vincent Sitzmann. Neural descriptor fields: Se(3)-equivariant object representations for manipulation. 2022. 2", + "[42] Vincent Sitzmann, Michael Zollhoefer, and Gordon Wetzstein. Scene representation networks: Continuous 3d-structure-aware neural scene representations. In H. Wallach, H. Larochelle, A. Beygelzimer, F. d'Alché-Buc, E. Fox, and R. Garnett, editors, Advances in Neural Information Processing Systems, volume 32. Curran Associates, Inc., 2019. 1", + "[43] Riccardo Spezialietti, Federico Stella, Marlon Marcon, Luciano Silva, Samuele Salti, and Luigi Di Stefano. Learning to orient surfaces by self-supervised spherical cnns. arXiv preprint arXiv:2011.03298, 2020. 1, 2, 3", + "[44] Srinath Sridhar, Davis Rempe, Julien Valentin, Sofien Bouaziz, and Leonidas J Guibas. Multiview aggregation for learning category-specific shape reconstruction. arXiv preprint arXiv:1907.01085, 2019. 2", + "[45] Pratul P Srinivasan, Boyang Deng, Xiuming Zhang, Matthew Tancik, Ben Mildenhall, and Jonathan T Barron. Nerv: Neural reflectance and visibility fields for relighting and view synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7495-7504, 2021. 2", + "[46] Shih-Yang Su, Frank Yu, Michael Zollhöfer, and Helge Rhodin. A-nerf: Articulated neural radiance fields for learning human shape, appearance, and pose. Advances in Neural Information Processing Systems, 34:12278-12291, 2021. 2" + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "4509", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[47] Weiwei Sun, Andrea Tagliasacchi, Boyang Deng, Sara Sabour, Soroosh Yazdani, Geoffrey Hinton, and Kwang Moo Yi. Canonical capsules: Unsupervised capsules in canonical pose. arXiv preprint arXiv:2012.04718, 2020. 1, 2, 3, 6, 7, 8", + "[48] Maxim Tatarchenko, Stephan R Richter, René Ranftl, Zhuwen Li, Vladlen Koltun, and Thomas Brox. What do single-view 3d reconstruction networks learn? In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3405-3414, 2019. 2", + "[49] Ayush Tewari, Justus Thies, Ben Mildenhall, Pratul Srinivasan, Edgar Tretschk, Yifan Wang, Christoph Lassner, Vincent Sitzmann, Ricardo Martin-Brualla, Stephen Lombardi, Tomas Simon, Christian Theobalt, Matthias Niessner, Jonathan T. Barron, Gordon Wetzstein, Michael Zollhoefer, and Vladislav Golyanik. Advances in neural rendering. Nov 2021. 2", + "[50] Nathaniel Thomas, Tess Smidt, Steven Kearnes, Lusann Yang, Li Li, Kai Kohlhoff, and Patrick Riley. Tensor field networks: Rotation-and translation-equivariant neural networks for 3d point clouds. arXiv preprint arXiv:1802.08219, 2018. 2, 3, 4", + "[51] Garvita Tiwari, Dimitrije Antic, Jan Eric Lenssen, Nikolaos Sarafianos, Tony Tung, and Gerard Pons-Moll. Pose-ndf: Modeling human pose manifolds with neural distance fields. In European Conference on Computer Vision (ECCV), October 2022. 2", + "[52] Edgar Tretschk, Ayush Tewari, Vladislav Golyanik, Michael Zollhöfer, Christoph Lassner, and Christian Theobalt. Non-rigid neural radiance fields: Reconstruction and novel view synthesis of a dynamic scene from monocular video. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12959–12970, 2021. 1", + "[53] Suhani Vora, Noha Radwan, Klaus Greff, Henning Meyer, Kyle Genova, Mehdi S. M. Sajjadi, Etienne Pot, Andrea Tagliasacchi, and Daniel Duckworth. Nesf: Neural semantic fields for generalizable semantic segmentation of 3d scenes, 2021. 1, 6", + "[54] He Wang, Srinath Sridhar, Jingwei Huang, Julien Valentin, Shuran Song, and Leonidas J Guibas. Normalized object coordinate space for category-level 6d object pose and size estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2642-2651, 2019. 2", + "[55] Peng Wang, Lingjie Liu, Yuan Liu, Christian Theobalt, Taku Komura, and Wenping Wang. Neus: Learning neural implicit surfaces by volume rendering for multi-view reconstruction. In Proceedings of the Thirtieth International Joint Conference on Artificial Intelligence (IJCAI). International Joint Conferences on Artificial Intelligence Organization, 2021. 1", + "[56] Maurice Weiler, Mario Geiger, Max Welling, Wouter Boomsma, and Taco Cohen. 3d steerable cnns: Learning rotationally equivariant features in volumetric data. In NIPS, pages 10381-10392, 2018. 2", + "[57] Maurice Weiler, Fred A Hamprecht, and Martin Storath. Learning steerable filters for rotation equivariant cnns. In Proc. CVPR, pages 849-858, 2018. 2", + "[58] Zhirong Wu, Shuran Song, Aditya Khosla, Fisher Yu, Linguang Zhang, Xiaou Tang, and Jianxiong Xiao. 3d" + ], + "bbox": [ + 78, + 90, + 470, + 901 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "shapenets: A deep representation for volumetric shapes. In Proc. CVPR, pages 1912-1920, 2015. 2", + "[59] Yiheng Xie, Towaki Takikawa, Shunsuke Saito, Or Litany, Shiqin Yan, Numair Khan, Federico Tombari, James Tompkin, Vincent Sitzmann, and Srinath Sridhar. Neural fields in visual computing and beyond. Computer Graphics Forum, 2022. 1, 2", + "[60] Guandao Yang, Serge Belongie, Bharath Hariharan, and Vladlen Koltun. Geometry processing with neural fields. Advances in Neural Information Processing Systems, 34:22483-22497, 2021. 1", + "[61] Lior Yariv, Jiatao Gu, Yoni Kasten, and Yaron Lipman. Volume rendering of neural implicit surfaces. In Advances in Neural Information Processing Systems (NeurIPS). Curran Associates, Inc., 2021. 1", + "[62] Lin Yen-Chen. Nerf-pytorch. https://github.com/yenchenlin/nerf-pytorch/, 2020.6", + "[63] Alex Yu, Vickie Ye, Matthew Tancik, and Angjoo Kanazawa. pixelnerf: Neural radiance fields from one or few images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4578-4587, 2021. 1, 2", + "[64] Kevin Zakka, Andy Zeng, Johnny Lee, and Shuran Song. Form2fit: Learning shape priors for generalizable assembly from disassembly. In 2020 IEEE International Conference on Robotics and Automation (ICRA), pages 9404-9410. IEEE, 2020. 2", + "[65] Ge Zhang, Or Litany, Srinath Sridhar, and Leonidas Guibas. Strobenet: Category-level multiview reconstruction of articulated objects. arXiv preprint arXiv:2105.08016, 2021. 2", + "[66] Kai Zhang, Gernot Riegler, Noah Snavely, and Vladlen Koltun. Nerf++: Analyzing and improving neural radiance fields. arXiv preprint arXiv:2010.07492, 2020. 1, 2", + "[67] Yongheng Zhao, Tolga Birdal, Haowen Deng, and Federico Tombari. 3d point capsule networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1009-1018, 2019. 3", + "[68] Shuaifeng Zhi, Tristan Laidlow, Stefan Leutenegger, and Andrew J. Davison. In-place scene labelling and understanding with implicit scene representation. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), 2021. 1", + "[69] Ellen D Zhong, Adam Lerer, Joseph H Davis, and Bonnie Berger. Cryodrgn2: Ab initio neural reconstruction of 3d protein structures from real cryo-em images. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4066-4075, 2021. 2" + ], + "bbox": [ + 501, + 92, + 893, + 756 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "4510", + "bbox": [ + 482, + 945, + 516, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2023/Canonical Fields_ Self-Supervised Learning of Pose-Canonicalized Neural Fields/54da6270-9672-4d8f-9893-3339952d04d6_model.json b/2023/Canonical Fields_ Self-Supervised Learning of Pose-Canonicalized Neural Fields/54da6270-9672-4d8f-9893-3339952d04d6_model.json new file mode 100644 index 0000000000000000000000000000000000000000..4bcd8a6b4e173b519fdae8e5039da033c0a7fb56 --- /dev/null +++ b/2023/Canonical Fields_ Self-Supervised Learning of Pose-Canonicalized Neural Fields/54da6270-9672-4d8f-9893-3339952d04d6_model.json @@ -0,0 +1,2400 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.081, + 0.131, + 0.89, + 0.152 + ], + "angle": 0, + "content": "Canonical Fields: Self-Supervised Learning of Pose-Canonicalized Neural Fields" + }, + { + "type": "text", + "bbox": [ + 0.154, + 0.18, + 0.288, + 0.2 + ], + "angle": 0, + "content": "Rohith Agaram1" + }, + { + "type": "text", + "bbox": [ + 0.328, + 0.181, + 0.462, + 0.199 + ], + "angle": 0, + "content": "Shaurya Dewan" + }, + { + "type": "text", + "bbox": [ + 0.345, + 0.2, + 0.493, + 0.215 + ], + "angle": 0, + "content": "Madhava Krishna1" + }, + { + "type": "text", + "bbox": [ + 0.503, + 0.181, + 0.625, + 0.199 + ], + "angle": 0, + "content": "Rahul Sajnani" + }, + { + "type": "text", + "bbox": [ + 0.665, + 0.181, + 0.814, + 0.198 + ], + "angle": 0, + "content": "Adrien Poulenard3" + }, + { + "type": "text", + "bbox": [ + 0.535, + 0.2, + 0.665, + 0.215 + ], + "angle": 0, + "content": "Srinath Sridhar" + }, + { + "type": "text", + "bbox": [ + 0.243, + 0.218, + 0.397, + 0.235 + ], + "angle": 0, + "content": "\\(^{1}\\)RRC, IIIT-Hyderabad" + }, + { + "type": "text", + "bbox": [ + 0.429, + 0.218, + 0.556, + 0.235 + ], + "angle": 0, + "content": "\\(^{2}\\)Brown University" + }, + { + "type": "text", + "bbox": [ + 0.589, + 0.218, + 0.728, + 0.235 + ], + "angle": 0, + "content": "3Stanford University" + }, + { + "type": "text", + "bbox": [ + 0.321, + 0.238, + 0.645, + 0.255 + ], + "angle": 0, + "content": "ivl.cs.brown.edu/projects/canonicalfields" + }, + { + "type": "image_caption", + "bbox": [ + 0.24, + 0.265, + 0.326, + 0.278 + ], + "angle": 0, + "content": "Input NeRFs" + }, + { + "type": "image_caption", + "bbox": [ + 0.632, + 0.265, + 0.745, + 0.277 + ], + "angle": 0, + "content": "Canonical Fields" + }, + { + "type": "image", + "bbox": [ + 0.081, + 0.28, + 0.485, + 0.404 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.485, + 0.28, + 0.891, + 0.404 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.407, + 0.893, + 0.461 + ], + "angle": 0, + "content": "Figure 1. We present Canonical Field Network (CaFi-Net), a self-supervised method for 3D position and orientation (pose) canonicalization of objects represented as neural fields. We specifically focus on neural radiance fields (NeRFs) fitted to raw RGB images of arbitrarily posed objects (left), and obtain a canonical field (fixed novel view shown on right) where all objects in a category are consistently positioned and oriented." + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.467, + 0.313, + 0.482 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.498, + 0.474, + 0.817 + ], + "angle": 0, + "content": "Coordinate-based implicit neural networks, or neural fields, have emerged as useful representations of shape and appearance in 3D computer vision. Despite advances, however, it remains challenging to build neural fields for categories of objects without datasets like ShapeNet that provide \"canonicalized\" object instances that are consistently aligned for their 3D position and orientation (pose). We present Canonical Field Network (CaFi-Net), a self-supervised method to canonicalize the 3D pose of instances from an object category represented as neural fields, specifically neural radiance fields (NeRFs). CaFi-Net directly learns from continuous and noisy radiance fields using a Siamese network architecture that is designed to extract equivariant field features for category-level canonicalization. During inference, our method takes pre-trained neural radiance fields of novel object instances at arbitrary 3D pose and estimates a canonical field with consistent 3D pose across the entire category. Extensive experiments on a new dataset of 1300 NeRF models across 13 object categories show that our method matches or exceeds the performance of 3D point cloud-based methods." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.846, + 0.208, + 0.86 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.47, + 0.902 + ], + "angle": 0, + "content": "Neural fields [59]—coordinate-based neural networks that implicitly parameterize signals—have recently gained" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.468, + 0.892, + 0.574 + ], + "angle": 0, + "content": "significant attention as representations of 3D shape [6,22, 28], view-dependent appearance [24, 42], and motion [25]. In particular, neural radiance fields (NeRFs) [24], have been successfully used in problems such as novel view synthesis [3,4,66], scene geometry extraction [55, 61], capturing dynamic scenes [19, 29, 30, 33, 52], 3D semantic segmentation [53, 68], and robotics [1, 13, 21]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.579, + 0.893, + 0.76 + ], + "angle": 0, + "content": "Despite the progress, it remains challenging to build neural fields that represent an entire category of objects. Previous methods sidestep the problem by overfitting on a single instance [24], or learning [22, 28, 63] on datasets like ShapeNet [5] that contain objects that are manually canonicalized – oriented consistently for 3D position and orientation (3D pose) across a category. This strong supervision makes it easier to learn over categories, but limits their application to data that contain these labels. Recent work has proposed methods for self-supervised learning of 3D pose canonicalization [40, 43, 47], however, these operate on 3D point clouds, meshes, or voxels – but not neural fields." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.894, + 0.902 + ], + "angle": 0, + "content": "In this paper, we present Canonical Field Network (CaFi-Net), a self-supervised method for category-level canonicalization of the 3D position and orientation of objects represented as neural fields, specifically neural radiance fields. Canonicalizing neural fields is challenging because, unlike 3D point clouds or meshes, neural fields are continuous, noisy, and hard to manipulate since they are parameterized as the weights of a neural network [60]. To address these challenges, we first extend the notion of equiv" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4500" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.288 + ], + "angle": 0, + "content": "arience to continuous vector fields and show how networks for processing 3D point clouds [50] can be extended to operate directly on neural radiance fields. We design CaFi-Net as a Siamese network that contains layers to extract equivariant features directly on vector fields. These field features are used to learn a canonical frame that is consistent across instances in the category. During inference, our method takes as input neural radiance fields of object instances from a category at arbitrary pose and estimates a canonical field that is consistent across the category. To handle noise in radiance fields from NeRF, our method incorporates density-based feature weighting and foreground-background clustering." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.288, + 0.473, + 0.486 + ], + "angle": 0, + "content": "Our approach learns canonicalization without any supervision labels on a new dataset of 1300 pre-trained NeRF models of 13 common ShapeNet categories in arbitrary 3D pose (see Figure 1). We introduce several self-supervision loss functions that encourage the estimation of a consistent canonical pose. In addition, we present extensive quantitative comparisons with baselines and other methods on standardized canonicalization metrics [37] over 13 object categories. In particular, we show that our approach matches or exceeds the performance of 3D point cloud-based methods. This enables the new capability of directly operating on neural fields rather than converting them to point clouds for canonicalization. To sum up, we contribute:" + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.495, + 0.468, + 0.554 + ], + "angle": 0, + "content": "- Canonical Field Network (CaFi-Net), the first method for self-supervised canonicalization of the 3D position and orientation (pose) of objects represented as neural radiance fields." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.558, + 0.468, + 0.614 + ], + "angle": 0, + "content": "- A Siamese neural network architecture with equivariant feature extraction layers that are designed to directly operate on continuous and noisy radiance fields from NeRF." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.619, + 0.468, + 0.663 + ], + "angle": 0, + "content": "- A public dataset of 1300 NeRF models from 13 ShapeNet categories including posed images, and weights for evaluating canonicalization performance." + }, + { + "type": "list", + "bbox": [ + 0.097, + 0.495, + 0.468, + 0.663 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.679, + 0.22, + 0.695 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.705, + 0.468, + 0.748 + ], + "angle": 0, + "content": "We focus our review of related work on neural fields, supervised canonicalization, equivariant neural network architectures, and self-supervised canonicalization." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.468, + 0.901 + ], + "angle": 0, + "content": "Neural Fields: Neural fields are emerging as useful representations for solving problems such as novel view synthesis [3, 4, 15, 24, 66], shape encoding and reconstruction [22, 28], dynamic reconstruction [19, 29, 30], appearance modeling [23, 36, 45], and human motion modeling [8, 18, 46, 51]. Generalization of neural fields to object categories remains difficult, but some methods have used pre-canonicalized datasets to circumvent this problem [63]. Neural fields are also used outside of computer vision, for instance in reconstructing proteins [69], physics [35], or au" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.136 + ], + "angle": 0, + "content": "dio [11]. Please see [49, 59] for more details. In this paper, our focus is on canonicalizing for the 3D pose of neural fields, specifically, neural radiance fields (NeRF) [24]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.137, + 0.892, + 0.303 + ], + "angle": 0, + "content": "Supervised Canonicalization: Datasets such as ShapeNet [5] and ModelNet40 [58] have 3D shapes that are manually pre-canonicalized. This inductive bias aids category-level generalization in problems such as 3D reconstruction [12, 22, 28, 48]. We can also use these datasets to formulate canonicalization as a supervised learning problem [54] enabling applications such as 6 degree-of-freedom object pose estimation, multi-view reconstruction [17, 44], and reconstruction of articulating [20, 65] and non-rigid objects [64]. However, our goal is to canonicalize without using manual pose labels." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.304, + 0.892, + 0.438 + ], + "angle": 0, + "content": "Equivariant Neural Networks: Pose-equivariant networks are equivariant to input pose [9, 50, 56, 57] by design. Some of these methods use Spherical Harmonic functions [10, 50, 56, 57], or vector neurons [9] to extract equivariant features. Equivariance is closely related to canonicalization since a canonical pose is also equivariant. Thus, previous methods have used pose-equivariance for canonicalization [37, 41, 43, 47]. However, these methods have thus far been limited to 3D point clouds, voxels, or meshes." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.439, + 0.892, + 0.529 + ], + "angle": 0, + "content": "Self-Supervised Canonicalization: Recent research has shown that self/weak supervision is sufficient for learning pose canonicalization on point clouds [37, 38, 43, 47], 2D key-points [26], and images [27]. None of these previous self-supervised methods can operate directly on neural fields – to the best of our knowledge, ours is the first." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.544, + 0.629, + 0.56 + ], + "angle": 0, + "content": "3. Background" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.569, + 0.89, + 0.597 + ], + "angle": 0, + "content": "We first provide background on the essential components and key ideas behind our method." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.6, + 0.892, + 0.854 + ], + "angle": 0, + "content": "Neural Radiance Fields (NeRF): Given posed RGB images, NeRF and its variants [23, 24] learn a neural network to synthesize novel views of complex scenes. At inference, for any 3D location and viewing direction, it estimates a direction-independent density value \\(\\sigma\\), and a direction-dependent color \\(c\\). Although designed for novel view synthesis, NeRF implicitly learns the 3D geometry and view-dependent appearance of scenes making it a useful representation of shape and appearance. While NeRFs excel at fitting a single object instance or scene, it struggles to generalize to object categories partly because objects can be in arbitrary 3D poses – thus, existing methods use precanonicalized datasets [63]. In this paper, we provide a way to canonicalize 3D pose without any supervision for objects represented as neural radiance fields. While our method is aimed at NeRFs, it can be extended to the 3D pose canonicalization of any neural field." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Equivalence and Canonicalization: A function \\(\\Gamma\\) over \\(x\\) is said to be equivariant to a group operation \\(\\mathcal{O}\\) if its output changes by a fixed mapping \\(M\\) for any \\(\\mathcal{O}\\) operating on" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "4501" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.088, + 0.473, + 0.246 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.257, + 0.473, + 0.37 + ], + "angle": 0, + "content": "Figure 2. Rotation Equivariance in Vector Fields. Rotation equivariance in vector fields \\(\\mathcal{F}\\) (here, a simulated 2D magnetic field) require two operations: (1) update the position of the green arrow (the vector), and (2) rotating the green arrow at its new position. A function \\(\\Gamma\\) operating on \\(\\mathcal{F}\\) is rotationally equivariant over the field if and only if \\(\\Gamma[(\\mathcal{R} \\cdot \\mathcal{F})(\\mathbf{x})] = M(\\mathcal{R})\\Gamma[\\mathcal{F}(\\mathcal{R}^{-1}\\mathbf{x})]\\), where \\(\\mathcal{R}\\) is an \\(SO(2)\\) rotation, and \\(M\\) is a rotation-dependent mapping function." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.398, + 0.47, + 0.566 + ], + "angle": 0, + "content": "the input \\( x \\), i.e., \\( \\Gamma(\\mathcal{O}x) = M(\\mathcal{O})\\Gamma(x) \\). In this paper, we are interested in equivariance to the \\( SE(3) \\) group denoting 3D position and orientation (pose) – in particular we focus on the rotation group \\( SO(3) \\) since 3D translation equivariance is readily achieved through mean centering [26]. Pose canonicalization is closely related to equivariance – our goal is to find an equivariant canonical 3D rigid transformation that maps the neural field of an object in an arbitrary pose to a canonical pose. We call the outputs of our method canonical fields – neural fields that are consistently oriented across a category of shapes." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.567, + 0.471, + 0.749 + ], + "angle": 0, + "content": "Rotation Equivalence in Vector Fields: Unlike 3D point clouds, rotation equivariance in fields is more involved. Consider a 2D vector field \\(\\mathcal{F}\\) defined \\(\\forall \\mathbf{x} \\in \\mathbb{R}^2\\) (see Figure 2). Rotating this field requires two operations: (1) updating the positions of vectors in the field to new rotated positions, and (2) rotating the directions of the vectors. A function \\(\\Gamma\\) operating on \\(\\mathcal{F}\\) is rotationally equivariant over the field if and only if \\(\\Gamma[(\\mathcal{R} \\cdot \\mathcal{F})(\\mathbf{x})] = M(\\mathcal{R})\\Gamma[\\mathcal{F}(\\mathcal{R}^{-1}\\mathbf{x})]\\), where \\(\\mathcal{R}\\) is an \\(SO(2)\\) rotation, and \\(M\\) is a rotation-dependent mapping function. In this paper, we operate on NeRF's density field (a scalar field) but also extract equivariant features that can be vector-valued fields." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.47, + 0.903 + ], + "angle": 0, + "content": "Tensor Field Networks (TFNs): Many methods have been proposed for using equivariant features for rotation canonicalization of 3D data, for instance, spherical CNNs [7, 43], vector neurons [9], or capsule networks [47, 67]. However, these and other methods [38] are limited to 3D point cloud or voxel representations. In this paper, we extend Tensor Field Networks [34, 50] pose canonicalization directly on samples from NeRF's density field. TFNs operate by computing Type- \\(\\ell\\) real-valued spherical harmonic functions" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.09, + 0.895, + 0.31 + ], + "angle": 0, + "content": "that are rotation equivariant: \\( D^{\\ell}(R)Y^{\\ell}(X) = Y^{\\ell}(Rx) \\). Where \\( D^{\\ell}:\\mathrm{SO}(3)\\to \\mathrm{SO}(2\\ell +1) \\) is the Wigner matrix of type \\( \\ell \\) (please see [34, 50] for details). For a 3D point set \\( P\\in \\mathbb{R}^{N\\times 3} \\), a rotation-equivariant TFN convolution layer (EQConv) at point \\( p\\in P \\) is defined as: \\( \\mathrm{EQConv}^J (p) = Q^{(n,\\ell),J}\\left(\\sum_{y\\in^{2R}\\mathcal{N}}k_r^n (p - y)\\otimes s^\\ell [y]\\right) \\). Where \\( ^{2R}\\mathcal{N} \\) is a set containing neighbors of point \\( p \\) at twice the resolution and \\( Q^{(n,\\ell),J}(\\cdot) \\) is the Clebsch Gordon decomposition to combine type- \\( n \\) kernel \\( k_r^n \\) and type- \\( \\ell \\) equivariant signal \\( s^\\ell \\). The EQConv \\( (\\cdot) \\) layer aggregates type- \\( \\ell \\) feature \\( s^\\ell \\) at multiple receptive fields along with learnable weights to perform equivariant convolution. Our method uses the EQConv layer as the fundamental building block to canonicalize continuous fields." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.32, + 0.593, + 0.337 + ], + "angle": 0, + "content": "4. Method" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.346, + 0.892, + 0.466 + ], + "angle": 0, + "content": "Our method assumes that we have pre-trained NeRF models of scenes containing single object instances at arbitrary poses from common shape categories. Our scenes contain primarily the object, but also some background - perfect foreground-background segmentation is not available. Our goal is to build a method that estimates a canonicalizing \\(SE(3)\\) transformation that aligns these object NeRFs to a consistent category-level canonical field." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.468, + 0.893, + 0.618 + ], + "angle": 0, + "content": "Rather than operate on 3D point clouds, meshes, or voxels, our method (see Figure 3) operates directly on samples from the continuous neural radiance field. During training, we learn canonical fields from a pre-trained NeRF dataset in a fully self-supervised manner. We use a Siamese network architecture that extracts equivariant field features for canonicalization (see Figure 3). During inference, given a previously unseen NeRF model of a new object instance in an arbitrary pose, our method estimates the transformation that maps it to the canonical field." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.625, + 0.661, + 0.643 + ], + "angle": 0, + "content": "4.1. NeRF Sampling" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.648, + 0.892, + 0.722 + ], + "angle": 0, + "content": "For a pre-trained neural radiance field that maps 3D position \\((x, y, z)\\) and direction \\(\\theta, \\phi\\) to color and density \\((\\mathbf{c}, \\sigma)\\), the input to our method consists of samples on the density field. We do not use color because it is direction-dependent in NeRF and also varies for different instances." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.724, + 0.895, + 0.86 + ], + "angle": 0, + "content": "Preprocessing: During both training and inference, we uniformly sample the pre-trained NeRF density field to find the object center and its bounding box (see supplementary document for details). We then sample a uniform grid in the density field of the object's bounding box. We empirically choose uniform grid sampling over fully random sampling since we found no difference. The uniformly sampled density grid is then normalized to obtain density values \\(\\sigma_{D}\\) in the continuous range [0, 1], i.e.," + }, + { + "type": "equation", + "bbox": [ + 0.558, + 0.881, + 0.892, + 0.898 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {\\sigma} _ {D} := \\left\\{1 - e x p \\left(- d \\cdot \\Gamma_ {r} ^ {\\sigma} (x)\\right) \\mid x \\in \\boldsymbol {X} \\right\\}, \\tag {1}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4502" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.091, + 0.089, + 0.887, + 0.379 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.388, + 0.896, + 0.472 + ], + "angle": 0, + "content": "Figure 3. CaFi-Net samples a density field from NeRF and uses its density, position and density gradients as input signals (A) to canonicalize the field. We predict rotation equivariant features and weigh them by density (B) to guide our learning from the occupied regions of the scene. We then compute an invariant embedding by taking a dot product between equivariant features. This invariant embedding is used to canonicalize the field that enables rendering all the objects in the canonical frame (E). Our method also applies an inter-instance consistency loss (D) that aligns different instances of the same category in the canonical frame. We do not assume precanonicalized fields, and canonicalize in a self-supervised manner." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.487, + 0.471, + 0.576 + ], + "angle": 0, + "content": "where \\( X \\) is the regularly sampled grid within the object bounding box and \\( d \\) is the depth step size used in the coarse NeRF sampling. Translation and scale canonicalization is simply achieved by moving the bounding box's center [26] and scaling it, so our next concern is on rotation canonicalization." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.589, + 0.188, + 0.604 + ], + "angle": 0, + "content": "4.2. CaFi-Net" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.613, + 0.47, + 0.658 + ], + "angle": 0, + "content": "CaFi-Net (Canonical Field Network) is our method for 3D rotation canonicalization using NeRF's density field (red box in Figure 3). We now describe its components." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.659, + 0.47, + 0.824 + ], + "angle": 0, + "content": "Signal Representation: At each of the sampled density field locations, we have several choices of signals to use as the input for CaFi-Net. For instance, we can (1) use the NeRF density value directly, (2) use the \\( xyz \\) location of the sampled field, or (3) use the gradient of the density field. Densities alone do not capture the position and directional components of the field, so we use a combination of densities and gradients of densities at query locations of the grid as our input signal for CaFi-Net. Gradients of density capture the object surface and help in canonicalization (see ablation study in Section 5)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.826, + 0.471, + 0.902 + ], + "angle": 0, + "content": "**Equivariant Convolution:** To canonicalize for rotation, we leverage equivariance properties of spherical harmonic functions to obtain rotation-equivariant learnable features on the field [34, 50]. Type-\\(\\ell\\) spherical harmonic functions \\((Y^{\\ell})\\) are degree-\\(\\ell\\) polynomials of points on the sphere that" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.487, + 0.895, + 0.655 + ], + "angle": 0, + "content": "are equivariant to \\( SO(2\\ell + 1) \\). The densities sampled from NeRF reconstructions are often noisy and contain outliers. To promote our method to learn only from occupied regions, we weigh each equivariant feature type by its density. This weighing has no effect on the equivariance of features (see supplementary document for proof). We have the option to weigh equivariant features with the local average of the density to smoothen the signal or to weigh directly by the density. We found weighing by density worked better than taking the local average density (see Section 5). We learn equivariant features of type \\( \\ell \\) as:" + }, + { + "type": "equation", + "bbox": [ + 0.537, + 0.669, + 0.888, + 0.69 + ], + "angle": 0, + "content": "\\[\n{ } ^ { R } s ^ { \\ell } : = f _ { w } ( ^ { R } x ) \\left( W ^ { \\ell } ( \\operatorname { E Q C o n v } ^ { \\ell } ( \\nabla ( \\boldsymbol { \\sigma } _ { D } ( ^ { R } x ) ) ) ) + \\delta _ { 0 \\ell } b \\right) , \\tag {2}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.537, + 0.693, + 0.888, + 0.726 + ], + "angle": 0, + "content": "\\[\nf _ {w} ^ {(R x)} := \\left\\{ \\begin{array}{l} \\operatorname {m e a n} _ {r} \\left(\\sigma_ {D} ^ {(R x)}\\right), \\quad o r \\\\ \\sigma_ {D} ^ {(R x)}, \\end{array} \\right. \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.727, + 0.892, + 0.88 + ], + "angle": 0, + "content": "where, \\( W^{\\ell} \\) and \\( b \\) are the weights and biases, \\( {}^{R}x \\) is a 3D point queried at resolution \\( R \\) and mean \\( r(\\sigma_D(Rx)) \\) is the average density at location \\( R_x \\) from points sampled at radius \\( r \\). We hierarchically aggregate features at resolutions \\( \\frac{1}{2}, \\frac{1}{4} \\), and \\( \\frac{1}{8} \\) to obtain global equivariant features. The EQConv convolution layer is the same as defined in Section 3. We employ non-linearities from [34] that preserve equivariance and capture better equivariant features. We also compute the max-pool of point-wise features of the last layer to obtain global type- \\( \\ell \\) equivariant features \\( F^{\\ell} \\)." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.886, + 0.893, + 0.902 + ], + "angle": 0, + "content": "Canonicalization: After obtaining the global type- \\(\\ell\\)" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4503" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.091, + 0.473, + 0.289 + ], + "angle": 0, + "content": "equivariant feature \\(F^{\\ell}\\), we compute its dot product with point-wise spherical harmonics \\(Z^{\\ell}\\) scaled by their norms \\(Z^{\\ell}(\\mathbf{X}) = ||\\mathbf{X}||_{2}Y^{\\ell}(\\mathbf{X} / ||\\mathbf{X}||_{2})\\) to obtain a pose-invariant embedding \\(H^{\\ell}\\) [37]. As the two vectors \\(F^{\\ell}\\) and \\(Z^{\\ell}\\) are both equivariant to input rotation, their dot product is invariant. We use a linear layer on top of the pose-invariant embedding to estimate grid coordinates in the canonical frame \\((P\\in \\mathbb{R}^{H\\times W\\times D\\times 3})\\) for each point in \\(\\mathbf{X}\\) and \\(M\\) equivariant transformations \\(E\\in \\mathbb{R}^{M\\times 3\\times 3}\\) [37] that orient the canonical coordinates to the input coordinates. We then choose the best canonicalizing transform \\(E_{b}\\) and penalize it for canonicalization. The invariant embedding \\(H\\), grid coordinates \\(P\\) and equivariant rotation \\(E\\) are given as," + }, + { + "type": "equation", + "bbox": [ + 0.154, + 0.297, + 0.47, + 0.316 + ], + "angle": 0, + "content": "\\[\nH ^ {\\ell} \\left(\\boldsymbol {\\sigma} _ {D}, \\boldsymbol {X}\\right) := \\langle F ^ {\\ell} \\left(\\boldsymbol {\\sigma} _ {D}\\right), Z ^ {\\ell} (\\boldsymbol {X}) \\rangle , \\tag {4}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.155, + 0.318, + 0.469, + 0.335 + ], + "angle": 0, + "content": "\\[\nP := \\operatorname {M L P} _ {Y} \\left(H \\left(\\boldsymbol {\\sigma} _ {D}, \\boldsymbol {X}\\right)\\right), \\tag {5}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.155, + 0.337, + 0.469, + 0.354 + ], + "angle": 0, + "content": "\\[\nE := \\operatorname {M L P} _ {E} \\left(F \\left(\\boldsymbol {\\sigma} _ {D}, \\boldsymbol {X}\\right)\\right). \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.364, + 0.47, + 0.395 + ], + "angle": 0, + "content": "Here, we have dropped the type-\\(\\ell\\) notation for \\(H\\) and \\(F\\) as we concatenate all equivariant feature types." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.395, + 0.47, + 0.485 + ], + "angle": 0, + "content": "Siamese Network Architecture: Experimentally, canonicalizing noisy density fields is difficult without guidance on shape similarity within a category. We therefore use a Siamese training strategy [26] where two different object instances are forced to be consistently canonicalized (see the lower branch in Figure 3)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.485, + 0.471, + 0.607 + ], + "angle": 0, + "content": "Clustering: CaFi-Net predicts a canonical coordinate for every grid point, but unlike point clouds, fields have many unoccupied regions that do not provide any information. We penalize the network on foreground regions only by performing K-means clustering on the densities within the object bounding box with \\(K = 2\\) and choose the cluster \\(C_f\\) with a higher mean. Note that we still operate on continuous fields, but only cluster samples to guide our training." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.615, + 0.182, + 0.631 + ], + "angle": 0, + "content": "4.3. Training" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.638, + 0.47, + 0.682 + ], + "angle": 0, + "content": "CaFi-Net is trained on a large dataset of pre-trained NeRF models over 13 object categories (see Section 5). We use the following loss functions to train our model." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.684, + 0.47, + 0.819 + ], + "angle": 0, + "content": "Canonicalization Loss: To self-supervise our learning, we transform the predicted canonical grid coordinates to the input grid using the canonical rotation \\( E_{b} \\) and compute a point-wise L2 distance between them. This loss forces the predicted invariant grid coordinates \\( P \\) to reconstruct the shape and forces the predicted canonicalizing transform to be equivariant to the input transformation. The final loss is computed over the canonicalizing transform that minimizes this loss:" + }, + { + "type": "equation", + "bbox": [ + 0.135, + 0.826, + 0.47, + 0.853 + ], + "angle": 0, + "content": "\\[\nE _ {b} := \\min _ {j} \\left(\\operatorname {m e a n} _ {i \\in C _ {f}} \\left\\| X _ {i} - E _ {j} P _ {i} \\right\\| _ {2} ^ {2}\\right), \\tag {7}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.136, + 0.857, + 0.469, + 0.877 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {c a n o n}} := \\operatorname {m e a n} _ {i \\in C _ {f}} \\left| \\left| \\boldsymbol {X} _ {i} - E _ {b} P _ {i} \\right| \\right| _ {2} ^ {2}, \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.886, + 0.447, + 0.902 + ], + "angle": 0, + "content": "where \\(C_f\\) is a set of points belonging to the foreground." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.892, + 0.138 + ], + "angle": 0, + "content": "**Orthonormality Loss:** The predicted equivariant transformations \\( E \\) should have orthonormal vectors and we use the following to force orthonormality using:" + }, + { + "type": "equation", + "bbox": [ + 0.574, + 0.149, + 0.892, + 0.166 + ], + "angle": 0, + "content": "\\[\nU _ {j}, D _ {j}, V _ {j} := \\operatorname {S V D} \\left(E _ {j}\\right), \\tag {9}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.575, + 0.169, + 0.892, + 0.205 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {o r t h o}} := \\frac {1}{M} \\sum_ {j} \\left| \\left| E _ {j} - U _ {j} V _ {j} ^ {T} \\right| \\right| _ {2}, \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.217, + 0.89, + 0.246 + ], + "angle": 0, + "content": "where \\(\\operatorname{SVD}(E_j)\\) computes the singular value decomposition of the \\(j^{th}\\) equivariant transformation." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.247, + 0.892, + 0.354 + ], + "angle": 0, + "content": "Siamese Shape Loss: We use a Siamese loss that penalizes for consistency between different shapes belonging to the same category. Given two un-canonicalized fields in different frames of reference \\(\\sigma_{D}^{1}\\) and \\(\\sigma_{D}^{2}\\), we canonicalize both the fields to the canonical frame and compute chamfer distance between their predicted canonical coordinates for points in the foreground cluster \\(C_f\\), i.e.," + }, + { + "type": "equation", + "bbox": [ + 0.562, + 0.364, + 0.892, + 0.383 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {s i a m e s e}} := \\mathrm {C D} \\left(P \\left(\\boldsymbol {\\sigma} _ {D} ^ {1}\\right), P \\left(\\boldsymbol {\\sigma} _ {D} ^ {2}\\right)\\right). \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.394, + 0.892, + 0.437 + ], + "angle": 0, + "content": "This loss regularizes the training by ensuring that instances of the same category should be aligned in the canonical frame." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.439, + 0.893, + 0.726 + ], + "angle": 0, + "content": "Architecture and Hyper-parameters: CaFi-Net predicts a 128 dimensional invariant embedding for each point in space and performs convolution at resolutions \\(1/2\\), \\(1/4\\), and \\(1/8\\). Each equivariant convolution layer aggregates features from 512 neighboring points at twice the resolution. We then use equivariant non-linearities introduced in [34] by applying Batch Normalization [14] and ReLU activation [2] in the inverse spherical harmonic transform domain. We use a three-layer MLP with intermediate Batch Normalization and ReLU layers followed by a final linear layer to predict canonical coordinates. To train our network, we weigh the canonicalization loss \\((\\mathcal{L}_{canon})\\) the highest with weight 2.0 and 1.0 for all the other loss functions. All our models are written in PyTorch [31] and are trained for 300 epochs with a batch size of 2 on an Nvidia 1080-Ti GPU. We use the Adam Optimizer [16] with an initial learning rate of \\(6e - 4\\) and L2 weight regularization of \\(1e - 5\\) for all our experiments. Please see the supplementary document for more details." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.74, + 0.633, + 0.758 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.765, + 0.892, + 0.901 + ], + "angle": 0, + "content": "We now provide details on our extensive experiments to evaluate the quality of our pose canonicalization. To our knowledge, we are the first method for pose canonicalization of neural radiance fields which makes direct comparisons with other work challenging. Nonetheless, we have designed experiments to compare our method with three 3D point cloud-based canonicalization methods. Overall, we present three classes of experiments: (1) evaluation of the performance of our method on 13 different categories," + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4504" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.089, + 0.895, + 0.174 + ], + "angle": 0, + "content": "Table 1. This table compares the canonicalization performance of our method (F - operating on fields) with other 3D point cloud-based methods (P) on three standard metrics (IC, CC and GEC) on our dataset of 13 categories. We compare with PCA, Canonical Capsules (CaCa) [47] and ConDor [37]. All metrics are multiplied by 100 for ease of reading. The top two performing methods are highlighted in magenta (best) and blue (second best). We are better than SOTA [37] on the Ground Truth Equivariance Consistency (GEC) and Category-Level Consistency (CC) with lower mean and median canonicalization error. However, we perform on par with ConDor [37] on the Instance-Level Consistency." + }, + { + "type": "table", + "bbox": [ + 0.085, + 0.175, + 0.887, + 0.41 + ], + "angle": 0, + "content": "
benchcabinetcarcellph.chaircouchfirearmlampmonitorplanespeakertablewater.avg.med.
Instance-Level Consistency (IC)↓
PCA (P)14.655.946.130.676.137.6315.0712.846.787.905.4010.3110.728.477.63
CaCa (P) [47]2.892.092.012.811.081.900.232.972.351.512.212.422.602.082.21
ConDor (P) [37]0.561.270.360.530.960.320.603.780.440.641.971.322.151.150.64
Ours (F)2.621.150.391.280.811.430.743.870.410.360.601.092.741.341.09
Category-Level Consistency (CC)↓
PCA (P)14.559.687.484.269.9013.4515.1311.8713.098.155.5110.4311.0710.3610.43
CaCa (P) [47]2.790.381.050.831.821.591.314.340.391.830.651.912.061.611.59
ConDor (P) [37]1.981.680.471.041.280.880.964.441.581.262.102.102.561.721.58
Ours (F)2.771.360.481.260.751.590.923.920.580.630.771.482.351.451.26
Ground Truth Equivalence Consistency (GEC)↓
PCA (P)14.405.866.321.646.328.1215.0112.496.828.165.3210.1711.278.618.12
CaCa (P) [47]3.652.282.402.741.902.221.284.692.482.092.223.012.792.592.40
ConDor (P) [37]2.121.790.501.171.340.931.044.581.591.312.222.242.671.811.59
Ours (F)3.261.510.541.430.941.901.014.370.630.660.811.633.021.671.43
" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.422, + 0.471, + 0.453 + ], + "angle": 0, + "content": "(2) comparison with other work, and (3) ablations to justify method design choices." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.463, + 0.473, + 0.904 + ], + "angle": 0, + "content": "Dataset: All our experiments use a new large synthetic NeRF dataset of shapes from 13 categories (that overlap with [12, 37, 47]) from the ShapeNet [5] dataset. For each category, we pick 100 instances from ShapeNet, rotate the shape randomly, and render 54 omnidirectional views sampled in a cube around the shape using Blender. Different from other datasets, we also simulate cluttered backgrounds. These posed views are used to train 100 NeRF models per category using the public PyTorch implementation of NeRF [62] for a total of 1300 NeRF models. To our knowledge, this is one of the largest (compared to [11, 53]) \\(360^{\\circ}\\) synthetic NeRF datasets - we will publicly release the simulator, raw posed images, and trained NeRF weights. Each NeRF model is trained for 400 epochs with 1024 randomly selected rays per iteration, a coarse sampling resolution of 64 points, and a fine sampling resolution of 128 points along each ray. Rather than train NeRF to maximize PSNR on a novel view, we fix the number of epochs for all models to enable a fair comparison between instances. As a consequence, some NeRF models may have a lower PSNR, but our goal is not to increase NeRF quality but rather improve canonicalization quality - hence we fix the number of epochs. For each category, we set aside \\(20\\%\\) of the models for testing. Metrics: Since there are no known metrics for evaluating canonical fields, we resort to 3 metrics used for 3D point cloud canonicalization [37]: (1) Instance-Level Consistency (IC), a metric that measures how consistently we can canonicalize the same instance in different poses, (2) Category-" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.422, + 0.895, + 0.621 + ], + "angle": 0, + "content": "Level Consistency (CC) to measure how consistently we canonicalize across different instances in the same category, and (3) Ground Truth Equivariance Consistency (GEC), a variant of the Ground Truth Consistency (GC) metric proposed in [37] to measure canonicalization performance compared to manual labels. We use the GEC because we observed that the original GC metric is prone to degeneracy when the canonicalizing transform is identity. We fix this issue by taking three point clouds \\( P_{i}, P_{j} \\), and \\( P_{k} \\) from a canonicalized dataset \\( \\mathcal{P} \\) and rotating \\( P_{i} \\) by \\( R_{1}, P_{j} \\) by \\( R_{2} \\) where both \\( R_{1} \\) and \\( R_{2} \\) are random rotations. Let \\( \\mathcal{C}(P) \\) predict a canonicalizing rotation for a point cloud \\( P \\), we then compute the Ground Truth Equivariance consistency" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.628, + 0.899, + 0.765 + ], + "angle": 0, + "content": "as: \\(\\mathrm{GEC} \\coloneqq \\frac{1}{|\\mathcal{P}|^3} \\sum_{P_i, P_j, P_k \\in \\mathcal{P}} \\mathrm{CD}(\\mathcal{C}(R_1P_i)R_1P_k, \\mathcal{C}(R_2P_j)R_2P_k)\\), (12) where CD refers to the Chamfer distance. Here, we apply canonicalizing transforms of shape \\(P_i\\) and \\(P_j\\) to the shape \\(P_k\\). This modified metric will not be degenerate for identity canonicalizing transforms and is more rigorous in evaluating ground truth consistency. To compute our metrics, we used point clouds sampled from the Shapenet meshes [5] using poisson disk sampling." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.771, + 0.748, + 0.788 + ], + "angle": 0, + "content": "5.1. Evaluation & Comparisons" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.796, + 0.895, + 0.903 + ], + "angle": 0, + "content": "We first analyze the performance of our method on the three metrics across 13 different categories. In Table 1, we highlight our method that operates on a field as Ours (F). Our method performs well across all categories with low variance between different categories. We observe similar trends across metrics for a set of categories (e.g., firearms,lamps) indicating that category shape plays a" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4505" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.084, + 0.088, + 0.344, + 0.352 + ], + "angle": 0, + "content": "
Categoryxyz vs. Gradient Signals
xyzGradient signal
Ground Truth Equivariance Consistency (GEC)↓
bench2.173.26
cellphone1.841.43
chair1.720.94
plane2.070.66
Average1.951.57
Instance-Level Consistency (IC)↓
bench2.042.62
cellphone1.61.28
chair1.40.81
plane1.880.36
Average1.731.26
Category-Level Consistency (CC)↓
bench1.992.77
cellphone1.491.26
chair1.520.75
plane1.90.63
Average1.721.35
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.077, + 0.356, + 0.346, + 0.406 + ], + "angle": 0, + "content": "(a) Choice of Signal Representation - Canonicalization metrics for using Gradients vs. \\( xyz \\) locations as input signal. Gradients capture the object surface that help in canonicalization." + }, + { + "type": "table", + "bbox": [ + 0.356, + 0.088, + 0.615, + 0.352 + ], + "angle": 0, + "content": "
CategoryLocal Average Density
w/ow
Ground Truth Equivariance Consistency (GEC)↓
bench3.263.38
cellphone1.431.63
chair0.941.3
plane0.660.64
Average1.571.73
Instance-Level Consistency (IC)↓
bench2.622.72
cellphone1.281.63
chair0.812.64
plane0.360.39
Average1.261.85
Category-Level Consistency (CC)↓
bench2.772.82
cellphone1.261.54
chair0.751.07
plane0.630.51
Average1.351.49
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.348, + 0.356, + 0.616, + 0.417 + ], + "angle": 0, + "content": "(b) Weighing Equivariant Signals by Local Average Density deteriorates the performance by smoothing out important details of the shape. We show canonicalization with (w) and without (w/o) weighing by the local averaged density." + }, + { + "type": "table", + "bbox": [ + 0.626, + 0.088, + 0.887, + 0.352 + ], + "angle": 0, + "content": "
CategorySiamese Training
w/o siamesewith siamese
Ground Truth Equivariance Consistency (GEC)↓
bench3.523.26
cellphone1.611.43
chair1.020.94
plane1.310.66
Average1.861.57
Instance-Level Consistency (IC)↓
bench2.732.62
cellphone1.31.28
chair0.820.81
plane1.10.36
Average1.481.26
Category-Level Consistency (CC)↓
bench2.952.77
cellphone1.441.26
chair0.830.75
plane1.260.63
Average1.621.35
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.619, + 0.356, + 0.89, + 0.428 + ], + "angle": 0, + "content": "(c) Siamese Training improves performance on all canonicalization metrics on average. We show canonicalization performance with siamese and without (w/o) siamese training. The average of Ground Truth Equivariance Consistency GEC metric reduces to 1.57 from 1.86" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.433, + 0.892, + 0.448 + ], + "angle": 0, + "content": "Table 2. Ablation studies to justify three key design choices: (a) signal representation, (b) density weighting, and (c) Siamese architecture." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.475, + 0.327, + 0.489 + ], + "angle": 0, + "content": "role in canonicalization performance." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.491, + 0.47, + 0.732 + ], + "angle": 0, + "content": "Next, we compare our method against three benchmarks: Principal Component Analysis (PCA) [32], Canonical Capsules (CaCa) [47], and ConDor [37]. All prior methods canonicalize clean point clouds with all points on the surface. To have a fair comparison, we train the above methods on point clouds (P) sampled from NeRF scenes for points on and within the object. Most scenes have points in the range of 120-1200 when sampled uniformly on the \\(32^{3}\\) grid. We resample the scene when points are less than 1024 and train both Canonical Capsules and ConDor on the resulting point clouds. CaFi-Net estimates a canonicalizing transform directly from the density field that we use to transform its corresponding ShapeNet point cloud for comparison against point canonicalizers. We follow the same protocol as [37] to measure and benchmark canonicalization performance for all the methods." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.735, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Analysis: Table 1 compares the metrics for PCA (P), CaCa (P), ConDor (P), and CaFi-Net [Ours (F)]. We list point cloud canonicalizers with (P) and field canonicalizer as (F) in the table. On average, our method performs better than point cloud-based methods in the CC and GEC metrics and is on par with the state-of-the-art method (ConDor) in the IC metric. This is despite the fact that our method directly operates on the field. We observe that PCA underperforms in all categories suggesting that the noisy point clouds make it unreliable. Similarly we observe that CaCa fails to canonicalize in categories where high noise is observed. Given" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.475, + 0.892, + 0.551 + ], + "angle": 0, + "content": "that ConDor also uses TFNs in a different architecture, we conclude that the choice of TFNs makes canonicalization more robust to noise. Our method has the additional advantage of operating directly on the field. Please see Figure 4 for a qualitative comparison of the different methods." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.56, + 0.611, + 0.575 + ], + "angle": 0, + "content": "5.2. Ablations" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.584, + 0.892, + 0.644 + ], + "angle": 0, + "content": "We conduct additional experiments to justify key design choices in our method. For all the ablation studies we use a smaller subset of our data consisting of 4 categories: bench, cellphone, chair, and airplane." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.644, + 0.892, + 0.794 + ], + "angle": 0, + "content": "Choice of Signal Representation: To select the appropriate signal representation to use (see Section 4), we conducted an ablation study using just the density field \\( xyz \\) locations or its gradients. As seen in Table 2 (a), the average canonicalization error over 4 categories is lower when using gradient as compared to using density \\( xyz \\) locations on all three metrics. We hypothesize that gradient of density weighs equivariant features closer to the surface of the object higher. This improves the canonicalization performance." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.796, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Weighing Features by Local Average Density: Next, we justify why weighting the equivariant features using the density is helpful. Table 2 (b) shows results for weighting features by local average density compared to weighting features by density. Weighing features by local average smoothens out details that are necessary to canonicalize the volume. Our average ground truth canonicalization error" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4506" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.185, + 0.088, + 0.262, + 0.1 + ], + "angle": 0, + "content": "Input NeRFs" + }, + { + "type": "image_caption", + "bbox": [ + 0.458, + 0.088, + 0.555, + 0.099 + ], + "angle": 0, + "content": "CaFi-Net (Ours)" + }, + { + "type": "image_caption", + "bbox": [ + 0.754, + 0.089, + 0.783, + 0.098 + ], + "angle": 0, + "content": "PCA" + }, + { + "type": "image", + "bbox": [ + 0.081, + 0.102, + 0.89, + 0.452 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.452, + 0.893, + 0.52 + ], + "angle": 0, + "content": "Figure 4. This figure visually compares the canonicalization performance of our method with 3D point cloud-based methods. Input NeRFs shows one of the RGB views used to learn our NeRF input (notice how orientations of instances are random). CaFi-Net (Ours) shows the results of our canonical field rendered from a novel view unseen during NeRF training. PCA (blue), CaCa [47] (red), and ConDor [37]'s (green) results are also presented for the same instances. Our method matches or exceeds ConDor. We also show some failure cases for several instances (red box) which occur either due to thin structures or symmetry." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.528, + 0.47, + 0.603 + ], + "angle": 0, + "content": "increases from 1.57 to 1.73 if we weigh features by local averaged densities. In future works, we will delve into better methods for reducing input signal noise that does not smoothen out the geometry details of objects to further improve our canonicalization." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.605, + 0.471, + 0.786 + ], + "angle": 0, + "content": "Siamese Architecture: Finally, we justify the need for a two-branch Siamese network architecture. Although our method can canonicalize even without this architecture, it makes it easier to learn over instances in a category as observed in [26]. Table 2 (c) compares our method on using a single branch architecture and a Siamese architecture. Clearly, the Siamese architecture helps improve results confirming our hypothesis. Enforcing different instances of the same category to align with each other reduces the Ground Truth Canonicalization error from 1.86 to 1.57. We also see a decrease in all the other metrics including Instance-Level Consistency." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.8, + 0.196, + 0.815 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.471, + 0.901 + ], + "angle": 0, + "content": "We presented Canonical Field Network (CaFi-Net), a method for self-supervised category-level canonicalization of the 3D pose of objects represented as neural radiance fields. CaFi-Net consists of a Siamese network architecture with rotation-equivariant convolution layers to extract" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.529, + 0.892, + 0.635 + ], + "angle": 0, + "content": "features for pose canonicalization. Our approach operates directly on NeRF's noisy but continuous density field. We train our method on a large dataset of 1300 NeRF models obtained for 13 common ShapeNet categories. During inference, our method is able to canonicalize arbitrarily oriented NeRFs. Experiments show that our method matches or outperforms 3D point cloud-based methods." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.659, + 0.893, + 0.84 + ], + "angle": 0, + "content": "Limitations & Future Work: Our approach has several limitations that future work should investigate. First, we require a \\(360^{\\circ}\\) NeRF model of objects and cannot handle partial views from front-facing NeRFs. Second, canonicalizing densities with uncertainty is more difficult as the uncertainty increases for very thin/small structures that can be missed (see chair in Figure 4). Furthermore, we also inherit the issues that TFNs have with symmetry (see monitor case in Figure 4). In future work, we plan to extend our approach to create a large real dataset of common object categories without requiring manual pose canonicalization and combat the symmetry issue in CaFi-Net as in [39]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.894, + 0.901 + ], + "angle": 0, + "content": "Acknowledgments: This work was supported by NSF grant CNS-2038897, an AWS Cloud Credits award, NSF CloudBank, and a gift from Meta Reality Labs. We thank Chandradeep Pokhariya and Ishaan Shah." + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4507" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.116, + 0.47, + 0.185 + ], + "angle": 0, + "content": "[1] Michal Adamkiewicz, Timothy Chen, Adam Caccavale, Rachel Gardner, Preston Culbertson, Jeannette Bohg, and Mac Schwager. Vision-only robot navigation in a neural radiance world. IEEE Robotics and Automation Letters, 7(2):4606-4613, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.187, + 0.47, + 0.215 + ], + "angle": 0, + "content": "[2] Abien Fred Agarap. Deep learning using rectified linear units (relu). arXiv preprint arXiv:1803.08375, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.217, + 0.471, + 0.299 + ], + "angle": 0, + "content": "[3] Jonathan T Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P Srinivasan. Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5855–5864, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.302, + 0.471, + 0.372 + ], + "angle": 0, + "content": "[4] Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Mip-nerf 360: Unbounded anti-aliased neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5470–5479, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.373, + 0.471, + 0.47 + ], + "angle": 0, + "content": "[5] Angel X. Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, Jianxiong Xiao, Li Yi, and Fisher Yu. ShapeNet: An Information-Rich 3D Model Repository. Technical Report arXiv:1512.03012 [cs.GR], Stanford University — Princeton University — Toyota Technological Institute at Chicago, 2015. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.472, + 0.471, + 0.527 + ], + "angle": 0, + "content": "[6] Zhiqin Chen and Hao Zhang. Learning implicit fields for generative shape modeling. Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.53, + 0.471, + 0.571 + ], + "angle": 0, + "content": "[7] Taco S Cohen, Mario Geiger, Jonas Kohler, and Max Welling. Spherical cnns. arXiv preprint arXiv:1801.10130, 2018.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.574, + 0.471, + 0.655 + ], + "angle": 0, + "content": "[8] Enric Corona, Tomas Hodan, Minh Vo, Francesc Moreno-Noguer, Chris Sweeney, Richard Newcombe, and Lingni Ma. Lisa: Learning implicit shape and appearance of hands. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20533-20543, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.659, + 0.471, + 0.741 + ], + "angle": 0, + "content": "[9] Congyue Deng, Or Litany, Yueqi Duan, Adrien Poulenard, Andrea Tagliasacchi, and Leonidas J. Guibas. Vector neurons: A general framework for so(3)-equivariant networks. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 12200-12209, October 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.744, + 0.47, + 0.799 + ], + "angle": 0, + "content": "[10] Fabian B Fuchs, Daniel E Worrall, Volker Fischer, and Max Welling. Se (3)-transformers: 3d roto-translation equivariant attention networks. arXiv preprint arXiv:2006.10503, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.802, + 0.471, + 0.857 + ], + "angle": 0, + "content": "[11] Ruohan Gao, Yen-Yu Chang, Shivani Mall, Li Fei-Fei, and Jiajun Wu. Objectfolder: A dataset of objects with implicit visual, auditory, and tactile representations. In Proceedings of the Conference on Robot Learning (CoRL), 2021. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.859, + 0.471, + 0.902 + ], + "angle": 0, + "content": "[12] Thibault Groueix, Matthew Fisher, Vladimir G Kim, Bryan C Russell, and Mathieu Aubry. A papier-mâché approach to learning 3d surface generation. In Proceedings of" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.116, + 0.471, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "the IEEE conference on computer vision and pattern recognition, pages 216-224, 2018. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.122, + 0.892, + 0.164 + ], + "angle": 0, + "content": "[13] Jeffrey Ichnowski, Yahav Avigal, Justin Kerr, and Ken Goldberg. Dex-nerf: Using a neural radiance field to grasp transparent objects. Oct 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.165, + 0.894, + 0.261 + ], + "angle": 0, + "content": "[14] Sergey Ioffe and Christian Szegedy. Batch normalization: Accelerating deep network training by reducing internal covariate shift. In Francis Bach and David Blei, editors, Proceedings of the 32nd International Conference on Machine Learning, volume 37 of Proceedings of Machine Learning Research, pages 448-456, Lille, France, 07-09 Jul 2015. PMLR. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.263, + 0.894, + 0.319 + ], + "angle": 0, + "content": "[15] Ajay Jain, Matthew Tancik, and Pieter Abbeel. Putting nerf on a diet: Semantically consistent few-shot view synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5885-5894, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.321, + 0.892, + 0.348 + ], + "angle": 0, + "content": "[16] Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. CoRR, abs/1412.6980, 2015. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.35, + 0.892, + 0.419 + ], + "angle": 0, + "content": "[17] Jiahui Lei, Srinath Sridhar, Paul Guerrero, Minhyuk Sung, Niloy Mitra, and Leonidas J Guibas. Pix2surf: Learning parametric 3d surface models of objects from images. In European Conference on Computer Vision, pages 121-138. Springer, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.421, + 0.892, + 0.475 + ], + "angle": 0, + "content": "[18] Ruilong Li, Julian Tanke, Minh Vo, Michael Zollhofer, Jurgen Gall, Angjoo Kanazawa, and Christoph Lassner. Tava: Template-free animatable volumetric actors. arXiv preprint arXiv:2206.08929, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.477, + 0.892, + 0.572 + ], + "angle": 0, + "content": "[19] Tianye Li, Mira Slavcheva, Michael Zollhoefer, Simon Green, Christoph Lassner, Changil Kim, Tanner Schmidt, Steven Lovegrove, Michael Goesele, Richard Newcombe, et al. Neural 3d video synthesis from multi-view video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5521-5531, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.576, + 0.892, + 0.644 + ], + "angle": 0, + "content": "[20] Xiaolong Li, He Wang, Li Yi, Leonidas J Guibas, A Lynn Abbott, and Shuran Song. Category-level articulated object pose estimation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3706-3715, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.647, + 0.892, + 0.702 + ], + "angle": 0, + "content": "[21] Yunzhu Li, Shuang Li, Vincent Sitzmann, Pulkit Agrawal, and Antonio Torralba. 3d neural scene representations for visuomotor control. In Proceedings of Robotics: Science and Systems, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.704, + 0.892, + 0.773 + ], + "angle": 0, + "content": "[22] Lars Mescheder, Michael Oechsle, Michael Niemeyer, Sebastian Nowozin, and Andreas Geiger. Occupancy networks: Learning 3d reconstruction in function space. In Proceedings IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2019. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.775, + 0.892, + 0.856 + ], + "angle": 0, + "content": "[23] Ben Mildenhall, Peter Hedman, Ricardo Martin-Brualla, Pratul P Srinivasan, and Jonathan T Barron. Nerf in the dark: High dynamic range view synthesis from noisy raw images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16190-16199, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.859, + 0.892, + 0.902 + ], + "angle": 0, + "content": "[24] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view syn" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4508" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.093, + 0.468, + 0.12 + ], + "angle": 0, + "content": "thesis. In European conference on computer vision, pages 405-421. Springer, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.122, + 0.469, + 0.176 + ], + "angle": 0, + "content": "[25] Michael Niemeyer, Lars Mescheder, Michael Oechsle, and Andreas Geiger. Occupancy flow: 4d reconstruction by learning particle dynamics. In Proc. of the IEEE International Conf. on Computer Vision (ICCV), 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.178, + 0.469, + 0.245 + ], + "angle": 0, + "content": "[26] David Novotny, Nikhila Ravi, Benjamin Graham, Natalia Neverova, and Andrea Vedaldi. C3dpo: Canonical 3d pose networks for non-rigid structure from motion. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7688-7697, 2019. 2, 3, 4, 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.247, + 0.469, + 0.302 + ], + "angle": 0, + "content": "[27] David Novotny, Roman Shapovalov, and Andrea Vedaldi. Canonical 3D Deformer Maps: Unifying parametric and non-parametric methods for dense weakly-supervised category reconstruction. In NeurIPS, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.304, + 0.469, + 0.371 + ], + "angle": 0, + "content": "[28] Jeong Joon Park, Peter Florence, Julian Straub, Richard Newcombe, and Steven Lovegrove. Deepsdf: Learning continuous signed distance functions for shape representation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 165-174, 2019. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.373, + 0.469, + 0.44 + ], + "angle": 0, + "content": "[29] Keunhong Park, Utkarsh Sinha, Jonathan T. Barron, Sofien Bouaziz, Dan B Goldman, Steven M. Seitz, and Ricardo Martin-Brualla. Nerfies: Deformable neural radiance fields. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.442, + 0.469, + 0.51 + ], + "angle": 0, + "content": "[30] Keunhong Park, Utkarsh Sinha, Peter Hedman, Jonathan T. Barron, Sofien Bouaziz, Dan B Goldman, Ricardo MartinBrualla, and Steven M. Seitz. Hypernerf: A higher-dimensional representation for topologically varying neural radiance fields. arXiv preprint arXiv:2106.13228, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.512, + 0.469, + 0.663 + ], + "angle": 0, + "content": "[31] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, Alban Desmaison, Andreas Kopf, Edward Yang, Zachary DeVito, Martin Raison, Alykhan Tejani, Sasank Chilamkurthy, Benoit Steiner, Lu Fang, Junjie Bai, and Soumith Chintala. Pytorch: An imperative style, high-performance deep learning library. In H. Wallach, H. Larochelle, A. Beygelzimer, F. d'Alché-Buc, E. Fox, and R. Garnett, editors, Advances in Neural Information Processing Systems 32, pages 8024-8035. Curran Associates, Inc., 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.665, + 0.469, + 0.718 + ], + "angle": 0, + "content": "[32] Karl Pearson. Liii. on lines and planes of closest fit to systems of points in space. The London, Edinburgh, and Dublin philosophical magazine and journal of science, 2(11):559-572, 1901. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.72, + 0.469, + 0.789 + ], + "angle": 0, + "content": "[33] Sida Peng, Junting Dong, Qianqian Wang, Shangzhan Zhang, Qing Shuai, Xiaowei Zhou, and Hujun Bao. Animatable neural radiance fields for modeling dynamic human bodies. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14314-14323, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.79, + 0.469, + 0.857 + ], + "angle": 0, + "content": "[34] Adrien Poulenard and Leonidas J Guibas. A functional approach to rotation equivariant non-linearities for tensor field networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13174-13183, 2021. 3, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.86, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[35] Maziar Raissi, Paris Perdikaris, and George E Karniadakis. Physics-informed neural networks: A deep learning framework for solving forward and inverse problems involving" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.093, + 0.469, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "nonlinear partial differential equations. Journal of Computational Physics, 378:686-707, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.122, + 0.892, + 0.176 + ], + "angle": 0, + "content": "[36] Viktor Rudnev, Mohamed Elgharib, William Smith, Lingjie Liu, Vladislav Golyanik, and Christian Theobalt. Nerf for outdoor scene relighting. In European Conference on Computer Vision (ECCV), 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.179, + 0.892, + 0.261 + ], + "angle": 0, + "content": "[37] Rahul Sajnani, Adrien Poulenard, Jivitesh Jain, Radhika Dua, Leonidas J Guibas, and Srinath Sridhar. Condor: Self-supervised canonicalization of 3d pose for partial shapes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16969-16979, 2022. 2, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.265, + 0.892, + 0.346 + ], + "angle": 0, + "content": "[38] Rahul Sajnani, AadilMehdi Sanchawala, Krishna Murthy Jatavallabhula, Srinath Sridhar, and K Madhava Krishna. Draco: Weakly supervised dense reconstruction and canonicalization of objects. In 2021 IEEE International Conference on Robotics and Automation (ICRA), pages 10302-10309. IEEE, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.349, + 0.892, + 0.404 + ], + "angle": 0, + "content": "[39] Ahyun Seo, Byungjin Kim, Suha Kwak, and Minsu Cho. Reflection and rotation symmetry detection via equivariant learning. In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.407, + 0.892, + 0.489 + ], + "angle": 0, + "content": "[40] Anthony Simeonov, Yilun Du, Andrea Tagliasacchi, Joshua B Tenenbaum, Alberto Rodriguez, Pulkit Agrawal, and Vincent Sitzmann. Neural descriptor fields: Se (3)-equivariant object representations for manipulation. In 2022 International Conference on Robotics and Automation (ICRA), pages 6394-6400. IEEE, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.491, + 0.892, + 0.558 + ], + "angle": 0, + "content": "[41] Anthony Simeonov, Yilun Du, Andrea Tagliasacchi, Joshua B. Tenenbaum, Alberto Rodriguez, Pulkit Agrawal, and Vincent Sitzmann. Neural descriptor fields: Se(3)-equivariant object representations for manipulation. 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.562, + 0.892, + 0.645 + ], + "angle": 0, + "content": "[42] Vincent Sitzmann, Michael Zollhoefer, and Gordon Wetzstein. Scene representation networks: Continuous 3d-structure-aware neural scene representations. In H. Wallach, H. Larochelle, A. Beygelzimer, F. d'Alché-Buc, E. Fox, and R. Garnett, editors, Advances in Neural Information Processing Systems, volume 32. Curran Associates, Inc., 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.647, + 0.892, + 0.702 + ], + "angle": 0, + "content": "[43] Riccardo Spezialietti, Federico Stella, Marlon Marcon, Luciano Silva, Samuele Salti, and Luigi Di Stefano. Learning to orient surfaces by self-supervised spherical cnns. arXiv preprint arXiv:2011.03298, 2020. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.704, + 0.892, + 0.759 + ], + "angle": 0, + "content": "[44] Srinath Sridhar, Davis Rempe, Julien Valentin, Sofien Bouaziz, and Leonidas J Guibas. Multiview aggregation for learning category-specific shape reconstruction. arXiv preprint arXiv:1907.01085, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.761, + 0.892, + 0.842 + ], + "angle": 0, + "content": "[45] Pratul P Srinivasan, Boyang Deng, Xiuming Zhang, Matthew Tancik, Ben Mildenhall, and Jonathan T Barron. Nerv: Neural reflectance and visibility fields for relighting and view synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7495-7504, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.846, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[46] Shih-Yang Su, Frank Yu, Michael Zollhöfer, and Helge Rhodin. A-nerf: Articulated neural radiance fields for learning human shape, appearance, and pose. Advances in Neural Information Processing Systems, 34:12278-12291, 2021. 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4509" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.148 + ], + "angle": 0, + "content": "[47] Weiwei Sun, Andrea Tagliasacchi, Boyang Deng, Sara Sabour, Soroosh Yazdani, Geoffrey Hinton, and Kwang Moo Yi. Canonical capsules: Unsupervised capsules in canonical pose. arXiv preprint arXiv:2012.04718, 2020. 1, 2, 3, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.149, + 0.471, + 0.218 + ], + "angle": 0, + "content": "[48] Maxim Tatarchenko, Stephan R Richter, René Ranftl, Zhuwen Li, Vladlen Koltun, and Thomas Brox. What do single-view 3d reconstruction networks learn? In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3405-3414, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.218, + 0.471, + 0.314 + ], + "angle": 0, + "content": "[49] Ayush Tewari, Justus Thies, Ben Mildenhall, Pratul Srinivasan, Edgar Tretschk, Yifan Wang, Christoph Lassner, Vincent Sitzmann, Ricardo Martin-Brualla, Stephen Lombardi, Tomas Simon, Christian Theobalt, Matthias Niessner, Jonathan T. Barron, Gordon Wetzstein, Michael Zollhoefer, and Vladislav Golyanik. Advances in neural rendering. Nov 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.316, + 0.471, + 0.383 + ], + "angle": 0, + "content": "[50] Nathaniel Thomas, Tess Smidt, Steven Kearnes, Lusann Yang, Li Li, Kai Kohlhoff, and Patrick Riley. Tensor field networks: Rotation-and translation-equivariant neural networks for 3d point clouds. arXiv preprint arXiv:1802.08219, 2018. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.385, + 0.471, + 0.453 + ], + "angle": 0, + "content": "[51] Garvita Tiwari, Dimitrije Antic, Jan Eric Lenssen, Nikolaos Sarafianos, Tony Tung, and Gerard Pons-Moll. Pose-ndf: Modeling human pose manifolds with neural distance fields. In European Conference on Computer Vision (ECCV), October 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.455, + 0.471, + 0.537 + ], + "angle": 0, + "content": "[52] Edgar Tretschk, Ayush Tewari, Vladislav Golyanik, Michael Zollhöfer, Christoph Lassner, and Christian Theobalt. Non-rigid neural radiance fields: Reconstruction and novel view synthesis of a dynamic scene from monocular video. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12959–12970, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.538, + 0.471, + 0.607 + ], + "angle": 0, + "content": "[53] Suhani Vora, Noha Radwan, Klaus Greff, Henning Meyer, Kyle Genova, Mehdi S. M. Sajjadi, Etienne Pot, Andrea Tagliasacchi, and Daniel Duckworth. Nesf: Neural semantic fields for generalizable semantic segmentation of 3d scenes, 2021. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.608, + 0.471, + 0.689 + ], + "angle": 0, + "content": "[54] He Wang, Srinath Sridhar, Jingwei Huang, Julien Valentin, Shuran Song, and Leonidas J Guibas. Normalized object coordinate space for category-level 6d object pose and size estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2642-2651, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.691, + 0.471, + 0.774 + ], + "angle": 0, + "content": "[55] Peng Wang, Lingjie Liu, Yuan Liu, Christian Theobalt, Taku Komura, and Wenping Wang. Neus: Learning neural implicit surfaces by volume rendering for multi-view reconstruction. In Proceedings of the Thirtieth International Joint Conference on Artificial Intelligence (IJCAI). International Joint Conferences on Artificial Intelligence Organization, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.775, + 0.471, + 0.83 + ], + "angle": 0, + "content": "[56] Maurice Weiler, Mario Geiger, Max Welling, Wouter Boomsma, and Taco Cohen. 3d steerable cnns: Learning rotationally equivariant features in volumetric data. In NIPS, pages 10381-10392, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.831, + 0.471, + 0.872 + ], + "angle": 0, + "content": "[57] Maurice Weiler, Fred A Hamprecht, and Martin Storath. Learning steerable filters for rotation equivariant cnns. In Proc. CVPR, pages 849-858, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.471, + 0.902 + ], + "angle": 0, + "content": "[58] Zhirong Wu, Shuran Song, Aditya Khosla, Fisher Yu, Linguang Zhang, Xiaou Tang, and Jianxiong Xiao. 3d" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.471, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "shapenets: A deep representation for volumetric shapes. In Proc. CVPR, pages 1912-1920, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.121, + 0.892, + 0.189 + ], + "angle": 0, + "content": "[59] Yiheng Xie, Towaki Takikawa, Shunsuke Saito, Or Litany, Shiqin Yan, Numair Khan, Federico Tombari, James Tompkin, Vincent Sitzmann, and Srinath Sridhar. Neural fields in visual computing and beyond. Computer Graphics Forum, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.191, + 0.892, + 0.247 + ], + "angle": 0, + "content": "[60] Guandao Yang, Serge Belongie, Bharath Hariharan, and Vladlen Koltun. Geometry processing with neural fields. Advances in Neural Information Processing Systems, 34:22483-22497, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.249, + 0.892, + 0.303 + ], + "angle": 0, + "content": "[61] Lior Yariv, Jiatao Gu, Yoni Kasten, and Yaron Lipman. Volume rendering of neural implicit surfaces. In Advances in Neural Information Processing Systems (NeurIPS). Curran Associates, Inc., 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.305, + 0.892, + 0.333 + ], + "angle": 0, + "content": "[62] Lin Yen-Chen. Nerf-pytorch. https://github.com/yenchenlin/nerf-pytorch/, 2020.6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.334, + 0.892, + 0.403 + ], + "angle": 0, + "content": "[63] Alex Yu, Vickie Ye, Matthew Tancik, and Angjoo Kanazawa. pixelnerf: Neural radiance fields from one or few images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4578-4587, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.404, + 0.892, + 0.473 + ], + "angle": 0, + "content": "[64] Kevin Zakka, Andy Zeng, Johnny Lee, and Shuran Song. Form2fit: Learning shape priors for generalizable assembly from disassembly. In 2020 IEEE International Conference on Robotics and Automation (ICRA), pages 9404-9410. IEEE, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.475, + 0.892, + 0.517 + ], + "angle": 0, + "content": "[65] Ge Zhang, Or Litany, Srinath Sridhar, and Leonidas Guibas. Strobenet: Category-level multiview reconstruction of articulated objects. arXiv preprint arXiv:2105.08016, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.518, + 0.892, + 0.559 + ], + "angle": 0, + "content": "[66] Kai Zhang, Gernot Riegler, Noah Snavely, and Vladlen Koltun. Nerf++: Analyzing and improving neural radiance fields. arXiv preprint arXiv:2010.07492, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.56, + 0.894, + 0.616 + ], + "angle": 0, + "content": "[67] Yongheng Zhao, Tolga Birdal, Haowen Deng, and Federico Tombari. 3d point capsule networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1009-1018, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.617, + 0.892, + 0.685 + ], + "angle": 0, + "content": "[68] Shuaifeng Zhi, Tristan Laidlow, Stefan Leutenegger, and Andrew J. Davison. In-place scene labelling and understanding with implicit scene representation. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.687, + 0.892, + 0.757 + ], + "angle": 0, + "content": "[69] Ellen D Zhong, Adam Lerer, Joseph H Davis, and Bonnie Berger. Cryodrgn2: Ab initio neural reconstruction of 3d protein structures from real cryo-em images. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4066-4075, 2021. 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.757 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.517, + 0.957 + ], + "angle": 0, + "content": "4510" + } + ] +] \ No newline at end of file diff --git a/2023/Canonical Fields_ Self-Supervised Learning of Pose-Canonicalized Neural Fields/54da6270-9672-4d8f-9893-3339952d04d6_origin.pdf b/2023/Canonical Fields_ Self-Supervised Learning of Pose-Canonicalized Neural Fields/54da6270-9672-4d8f-9893-3339952d04d6_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..58683a3cb8519d396354f39f71a10ca1ccae6f34 --- /dev/null +++ b/2023/Canonical Fields_ Self-Supervised Learning of Pose-Canonicalized Neural Fields/54da6270-9672-4d8f-9893-3339952d04d6_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:154ba249114c36a8c99e0256453f6e66c3439bbfdae7705055f04ab17d0c65bc +size 3167212 diff --git a/2023/Canonical Fields_ Self-Supervised Learning of Pose-Canonicalized Neural Fields/full.md b/2023/Canonical Fields_ Self-Supervised Learning of Pose-Canonicalized Neural Fields/full.md new file mode 100644 index 0000000000000000000000000000000000000000..8a9bb9122d013afc827cb5c68aa73562225279e7 --- /dev/null +++ b/2023/Canonical Fields_ Self-Supervised Learning of Pose-Canonicalized Neural Fields/full.md @@ -0,0 +1,335 @@ +# Canonical Fields: Self-Supervised Learning of Pose-Canonicalized Neural Fields + +Rohith Agaram1 + +Shaurya Dewan + +Madhava Krishna1 + +Rahul Sajnani + +Adrien Poulenard3 + +Srinath Sridhar + +$^{1}$ RRC, IIIT-Hyderabad + +$^{2}$ Brown University + +3Stanford University + +ivl.cs.brown.edu/projects/canonicalfields + +![](images/3d97b753ffa37973fda8c606cdd8ac9d31052170813520d893589850268afb99.jpg) +Input NeRFs + +![](images/10d00d5ec06704c055fab2229ceebf64850b7ae311344eae598010bbdc85132c.jpg) +Canonical Fields +Figure 1. We present Canonical Field Network (CaFi-Net), a self-supervised method for 3D position and orientation (pose) canonicalization of objects represented as neural fields. We specifically focus on neural radiance fields (NeRFs) fitted to raw RGB images of arbitrarily posed objects (left), and obtain a canonical field (fixed novel view shown on right) where all objects in a category are consistently positioned and oriented. + +# Abstract + +Coordinate-based implicit neural networks, or neural fields, have emerged as useful representations of shape and appearance in 3D computer vision. Despite advances, however, it remains challenging to build neural fields for categories of objects without datasets like ShapeNet that provide "canonicalized" object instances that are consistently aligned for their 3D position and orientation (pose). We present Canonical Field Network (CaFi-Net), a self-supervised method to canonicalize the 3D pose of instances from an object category represented as neural fields, specifically neural radiance fields (NeRFs). CaFi-Net directly learns from continuous and noisy radiance fields using a Siamese network architecture that is designed to extract equivariant field features for category-level canonicalization. During inference, our method takes pre-trained neural radiance fields of novel object instances at arbitrary 3D pose and estimates a canonical field with consistent 3D pose across the entire category. Extensive experiments on a new dataset of 1300 NeRF models across 13 object categories show that our method matches or exceeds the performance of 3D point cloud-based methods. + +# 1. Introduction + +Neural fields [59]—coordinate-based neural networks that implicitly parameterize signals—have recently gained + +significant attention as representations of 3D shape [6,22, 28], view-dependent appearance [24, 42], and motion [25]. In particular, neural radiance fields (NeRFs) [24], have been successfully used in problems such as novel view synthesis [3,4,66], scene geometry extraction [55, 61], capturing dynamic scenes [19, 29, 30, 33, 52], 3D semantic segmentation [53, 68], and robotics [1, 13, 21]. + +Despite the progress, it remains challenging to build neural fields that represent an entire category of objects. Previous methods sidestep the problem by overfitting on a single instance [24], or learning [22, 28, 63] on datasets like ShapeNet [5] that contain objects that are manually canonicalized – oriented consistently for 3D position and orientation (3D pose) across a category. This strong supervision makes it easier to learn over categories, but limits their application to data that contain these labels. Recent work has proposed methods for self-supervised learning of 3D pose canonicalization [40, 43, 47], however, these operate on 3D point clouds, meshes, or voxels – but not neural fields. + +In this paper, we present Canonical Field Network (CaFi-Net), a self-supervised method for category-level canonicalization of the 3D position and orientation of objects represented as neural fields, specifically neural radiance fields. Canonicalizing neural fields is challenging because, unlike 3D point clouds or meshes, neural fields are continuous, noisy, and hard to manipulate since they are parameterized as the weights of a neural network [60]. To address these challenges, we first extend the notion of equiv + +arience to continuous vector fields and show how networks for processing 3D point clouds [50] can be extended to operate directly on neural radiance fields. We design CaFi-Net as a Siamese network that contains layers to extract equivariant features directly on vector fields. These field features are used to learn a canonical frame that is consistent across instances in the category. During inference, our method takes as input neural radiance fields of object instances from a category at arbitrary pose and estimates a canonical field that is consistent across the category. To handle noise in radiance fields from NeRF, our method incorporates density-based feature weighting and foreground-background clustering. + +Our approach learns canonicalization without any supervision labels on a new dataset of 1300 pre-trained NeRF models of 13 common ShapeNet categories in arbitrary 3D pose (see Figure 1). We introduce several self-supervision loss functions that encourage the estimation of a consistent canonical pose. In addition, we present extensive quantitative comparisons with baselines and other methods on standardized canonicalization metrics [37] over 13 object categories. In particular, we show that our approach matches or exceeds the performance of 3D point cloud-based methods. This enables the new capability of directly operating on neural fields rather than converting them to point clouds for canonicalization. To sum up, we contribute: + +- Canonical Field Network (CaFi-Net), the first method for self-supervised canonicalization of the 3D position and orientation (pose) of objects represented as neural radiance fields. +- A Siamese neural network architecture with equivariant feature extraction layers that are designed to directly operate on continuous and noisy radiance fields from NeRF. +- A public dataset of 1300 NeRF models from 13 ShapeNet categories including posed images, and weights for evaluating canonicalization performance. + +# 2. Related Work + +We focus our review of related work on neural fields, supervised canonicalization, equivariant neural network architectures, and self-supervised canonicalization. + +Neural Fields: Neural fields are emerging as useful representations for solving problems such as novel view synthesis [3, 4, 15, 24, 66], shape encoding and reconstruction [22, 28], dynamic reconstruction [19, 29, 30], appearance modeling [23, 36, 45], and human motion modeling [8, 18, 46, 51]. Generalization of neural fields to object categories remains difficult, but some methods have used pre-canonicalized datasets to circumvent this problem [63]. Neural fields are also used outside of computer vision, for instance in reconstructing proteins [69], physics [35], or au + +dio [11]. Please see [49, 59] for more details. In this paper, our focus is on canonicalizing for the 3D pose of neural fields, specifically, neural radiance fields (NeRF) [24]. + +Supervised Canonicalization: Datasets such as ShapeNet [5] and ModelNet40 [58] have 3D shapes that are manually pre-canonicalized. This inductive bias aids category-level generalization in problems such as 3D reconstruction [12, 22, 28, 48]. We can also use these datasets to formulate canonicalization as a supervised learning problem [54] enabling applications such as 6 degree-of-freedom object pose estimation, multi-view reconstruction [17, 44], and reconstruction of articulating [20, 65] and non-rigid objects [64]. However, our goal is to canonicalize without using manual pose labels. + +Equivariant Neural Networks: Pose-equivariant networks are equivariant to input pose [9, 50, 56, 57] by design. Some of these methods use Spherical Harmonic functions [10, 50, 56, 57], or vector neurons [9] to extract equivariant features. Equivariance is closely related to canonicalization since a canonical pose is also equivariant. Thus, previous methods have used pose-equivariance for canonicalization [37, 41, 43, 47]. However, these methods have thus far been limited to 3D point clouds, voxels, or meshes. + +Self-Supervised Canonicalization: Recent research has shown that self/weak supervision is sufficient for learning pose canonicalization on point clouds [37, 38, 43, 47], 2D key-points [26], and images [27]. None of these previous self-supervised methods can operate directly on neural fields – to the best of our knowledge, ours is the first. + +# 3. Background + +We first provide background on the essential components and key ideas behind our method. + +Neural Radiance Fields (NeRF): Given posed RGB images, NeRF and its variants [23, 24] learn a neural network to synthesize novel views of complex scenes. At inference, for any 3D location and viewing direction, it estimates a direction-independent density value $\sigma$ , and a direction-dependent color $c$ . Although designed for novel view synthesis, NeRF implicitly learns the 3D geometry and view-dependent appearance of scenes making it a useful representation of shape and appearance. While NeRFs excel at fitting a single object instance or scene, it struggles to generalize to object categories partly because objects can be in arbitrary 3D poses – thus, existing methods use precanonicalized datasets [63]. In this paper, we provide a way to canonicalize 3D pose without any supervision for objects represented as neural radiance fields. While our method is aimed at NeRFs, it can be extended to the 3D pose canonicalization of any neural field. + +Equivalence and Canonicalization: A function $\Gamma$ over $x$ is said to be equivariant to a group operation $\mathcal{O}$ if its output changes by a fixed mapping $M$ for any $\mathcal{O}$ operating on + +![](images/4f7702db1fda005750345acb1f467aed69d3d49e34e8a2e86cc59ff2b522166f.jpg) +Figure 2. Rotation Equivariance in Vector Fields. Rotation equivariance in vector fields $\mathcal{F}$ (here, a simulated 2D magnetic field) require two operations: (1) update the position of the green arrow (the vector), and (2) rotating the green arrow at its new position. A function $\Gamma$ operating on $\mathcal{F}$ is rotationally equivariant over the field if and only if $\Gamma[(\mathcal{R} \cdot \mathcal{F})(\mathbf{x})] = M(\mathcal{R})\Gamma[\mathcal{F}(\mathcal{R}^{-1}\mathbf{x})]$ , where $\mathcal{R}$ is an $SO(2)$ rotation, and $M$ is a rotation-dependent mapping function. + +the input $x$ , i.e., $\Gamma(\mathcal{O}x) = M(\mathcal{O})\Gamma(x)$ . In this paper, we are interested in equivariance to the $SE(3)$ group denoting 3D position and orientation (pose) – in particular we focus on the rotation group $SO(3)$ since 3D translation equivariance is readily achieved through mean centering [26]. Pose canonicalization is closely related to equivariance – our goal is to find an equivariant canonical 3D rigid transformation that maps the neural field of an object in an arbitrary pose to a canonical pose. We call the outputs of our method canonical fields – neural fields that are consistently oriented across a category of shapes. + +Rotation Equivalence in Vector Fields: Unlike 3D point clouds, rotation equivariance in fields is more involved. Consider a 2D vector field $\mathcal{F}$ defined $\forall \mathbf{x} \in \mathbb{R}^2$ (see Figure 2). Rotating this field requires two operations: (1) updating the positions of vectors in the field to new rotated positions, and (2) rotating the directions of the vectors. A function $\Gamma$ operating on $\mathcal{F}$ is rotationally equivariant over the field if and only if $\Gamma[(\mathcal{R} \cdot \mathcal{F})(\mathbf{x})] = M(\mathcal{R})\Gamma[\mathcal{F}(\mathcal{R}^{-1}\mathbf{x})]$ , where $\mathcal{R}$ is an $SO(2)$ rotation, and $M$ is a rotation-dependent mapping function. In this paper, we operate on NeRF's density field (a scalar field) but also extract equivariant features that can be vector-valued fields. + +Tensor Field Networks (TFNs): Many methods have been proposed for using equivariant features for rotation canonicalization of 3D data, for instance, spherical CNNs [7, 43], vector neurons [9], or capsule networks [47, 67]. However, these and other methods [38] are limited to 3D point cloud or voxel representations. In this paper, we extend Tensor Field Networks [34, 50] pose canonicalization directly on samples from NeRF's density field. TFNs operate by computing Type- $\ell$ real-valued spherical harmonic functions + +that are rotation equivariant: $D^{\ell}(R)Y^{\ell}(X) = Y^{\ell}(Rx)$ . Where $D^{\ell}:\mathrm{SO}(3)\to \mathrm{SO}(2\ell +1)$ is the Wigner matrix of type $\ell$ (please see [34, 50] for details). For a 3D point set $P\in \mathbb{R}^{N\times 3}$ , a rotation-equivariant TFN convolution layer (EQConv) at point $p\in P$ is defined as: $\mathrm{EQConv}^J (p) = Q^{(n,\ell),J}\left(\sum_{y\in^{2R}\mathcal{N}}k_r^n (p - y)\otimes s^\ell [y]\right)$ . Where $^{2R}\mathcal{N}$ is a set containing neighbors of point $p$ at twice the resolution and $Q^{(n,\ell),J}(\cdot)$ is the Clebsch Gordon decomposition to combine type- $n$ kernel $k_r^n$ and type- $\ell$ equivariant signal $s^\ell$ . The EQConv $(\cdot)$ layer aggregates type- $\ell$ feature $s^\ell$ at multiple receptive fields along with learnable weights to perform equivariant convolution. Our method uses the EQConv layer as the fundamental building block to canonicalize continuous fields. + +# 4. Method + +Our method assumes that we have pre-trained NeRF models of scenes containing single object instances at arbitrary poses from common shape categories. Our scenes contain primarily the object, but also some background - perfect foreground-background segmentation is not available. Our goal is to build a method that estimates a canonicalizing $SE(3)$ transformation that aligns these object NeRFs to a consistent category-level canonical field. + +Rather than operate on 3D point clouds, meshes, or voxels, our method (see Figure 3) operates directly on samples from the continuous neural radiance field. During training, we learn canonical fields from a pre-trained NeRF dataset in a fully self-supervised manner. We use a Siamese network architecture that extracts equivariant field features for canonicalization (see Figure 3). During inference, given a previously unseen NeRF model of a new object instance in an arbitrary pose, our method estimates the transformation that maps it to the canonical field. + +# 4.1. NeRF Sampling + +For a pre-trained neural radiance field that maps 3D position $(x, y, z)$ and direction $\theta, \phi$ to color and density $(\mathbf{c}, \sigma)$ , the input to our method consists of samples on the density field. We do not use color because it is direction-dependent in NeRF and also varies for different instances. + +Preprocessing: During both training and inference, we uniformly sample the pre-trained NeRF density field to find the object center and its bounding box (see supplementary document for details). We then sample a uniform grid in the density field of the object's bounding box. We empirically choose uniform grid sampling over fully random sampling since we found no difference. The uniformly sampled density grid is then normalized to obtain density values $\sigma_{D}$ in the continuous range [0, 1], i.e., + +$$ +\boldsymbol {\sigma} _ {D} := \left\{1 - e x p \left(- d \cdot \Gamma_ {r} ^ {\sigma} (x)\right) \mid x \in \boldsymbol {X} \right\}, \tag {1} +$$ + +![](images/d39ec9a4d2ad86ff29ef52690ae451f656d06e9adfc3938575f17f23d2437a23.jpg) +Figure 3. CaFi-Net samples a density field from NeRF and uses its density, position and density gradients as input signals (A) to canonicalize the field. We predict rotation equivariant features and weigh them by density (B) to guide our learning from the occupied regions of the scene. We then compute an invariant embedding by taking a dot product between equivariant features. This invariant embedding is used to canonicalize the field that enables rendering all the objects in the canonical frame (E). Our method also applies an inter-instance consistency loss (D) that aligns different instances of the same category in the canonical frame. We do not assume precanonicalized fields, and canonicalize in a self-supervised manner. + +where $X$ is the regularly sampled grid within the object bounding box and $d$ is the depth step size used in the coarse NeRF sampling. Translation and scale canonicalization is simply achieved by moving the bounding box's center [26] and scaling it, so our next concern is on rotation canonicalization. + +# 4.2. CaFi-Net + +CaFi-Net (Canonical Field Network) is our method for 3D rotation canonicalization using NeRF's density field (red box in Figure 3). We now describe its components. + +Signal Representation: At each of the sampled density field locations, we have several choices of signals to use as the input for CaFi-Net. For instance, we can (1) use the NeRF density value directly, (2) use the $xyz$ location of the sampled field, or (3) use the gradient of the density field. Densities alone do not capture the position and directional components of the field, so we use a combination of densities and gradients of densities at query locations of the grid as our input signal for CaFi-Net. Gradients of density capture the object surface and help in canonicalization (see ablation study in Section 5). + +**Equivariant Convolution:** To canonicalize for rotation, we leverage equivariance properties of spherical harmonic functions to obtain rotation-equivariant learnable features on the field [34, 50]. Type- $\ell$ spherical harmonic functions $(Y^{\ell})$ are degree- $\ell$ polynomials of points on the sphere that + +are equivariant to $SO(2\ell + 1)$ . The densities sampled from NeRF reconstructions are often noisy and contain outliers. To promote our method to learn only from occupied regions, we weigh each equivariant feature type by its density. This weighing has no effect on the equivariance of features (see supplementary document for proof). We have the option to weigh equivariant features with the local average of the density to smoothen the signal or to weigh directly by the density. We found weighing by density worked better than taking the local average density (see Section 5). We learn equivariant features of type $\ell$ as: + +$$ +{ } ^ { R } s ^ { \ell } : = f _ { w } ( ^ { R } x ) \left( W ^ { \ell } ( \operatorname { E Q C o n v } ^ { \ell } ( \nabla ( \boldsymbol { \sigma } _ { D } ( ^ { R } x ) ) ) ) + \delta _ { 0 \ell } b \right) , \tag {2} +$$ + +$$ +f _ {w} ^ {(R x)} := \left\{ \begin{array}{l} \operatorname {m e a n} _ {r} \left(\sigma_ {D} ^ {(R x)}\right), \quad o r \\ \sigma_ {D} ^ {(R x)}, \end{array} \right. \tag {3} +$$ + +where, $W^{\ell}$ and $b$ are the weights and biases, ${}^{R}x$ is a 3D point queried at resolution $R$ and mean $r(\sigma_D(Rx))$ is the average density at location $R_x$ from points sampled at radius $r$ . We hierarchically aggregate features at resolutions $\frac{1}{2}, \frac{1}{4}$ , and $\frac{1}{8}$ to obtain global equivariant features. The EQConv convolution layer is the same as defined in Section 3. We employ non-linearities from [34] that preserve equivariance and capture better equivariant features. We also compute the max-pool of point-wise features of the last layer to obtain global type- $\ell$ equivariant features $F^{\ell}$ . + +Canonicalization: After obtaining the global type- $\ell$ + +equivariant feature $F^{\ell}$ , we compute its dot product with point-wise spherical harmonics $Z^{\ell}$ scaled by their norms $Z^{\ell}(\mathbf{X}) = ||\mathbf{X}||_{2}Y^{\ell}(\mathbf{X} / ||\mathbf{X}||_{2})$ to obtain a pose-invariant embedding $H^{\ell}$ [37]. As the two vectors $F^{\ell}$ and $Z^{\ell}$ are both equivariant to input rotation, their dot product is invariant. We use a linear layer on top of the pose-invariant embedding to estimate grid coordinates in the canonical frame $(P\in \mathbb{R}^{H\times W\times D\times 3})$ for each point in $\mathbf{X}$ and $M$ equivariant transformations $E\in \mathbb{R}^{M\times 3\times 3}$ [37] that orient the canonical coordinates to the input coordinates. We then choose the best canonicalizing transform $E_{b}$ and penalize it for canonicalization. The invariant embedding $H$ , grid coordinates $P$ and equivariant rotation $E$ are given as, + +$$ +H ^ {\ell} \left(\boldsymbol {\sigma} _ {D}, \boldsymbol {X}\right) := \langle F ^ {\ell} \left(\boldsymbol {\sigma} _ {D}\right), Z ^ {\ell} (\boldsymbol {X}) \rangle , \tag {4} +$$ + +$$ +P := \operatorname {M L P} _ {Y} \left(H \left(\boldsymbol {\sigma} _ {D}, \boldsymbol {X}\right)\right), \tag {5} +$$ + +$$ +E := \operatorname {M L P} _ {E} \left(F \left(\boldsymbol {\sigma} _ {D}, \boldsymbol {X}\right)\right). \tag {6} +$$ + +Here, we have dropped the type- $\ell$ notation for $H$ and $F$ as we concatenate all equivariant feature types. + +Siamese Network Architecture: Experimentally, canonicalizing noisy density fields is difficult without guidance on shape similarity within a category. We therefore use a Siamese training strategy [26] where two different object instances are forced to be consistently canonicalized (see the lower branch in Figure 3). + +Clustering: CaFi-Net predicts a canonical coordinate for every grid point, but unlike point clouds, fields have many unoccupied regions that do not provide any information. We penalize the network on foreground regions only by performing K-means clustering on the densities within the object bounding box with $K = 2$ and choose the cluster $C_f$ with a higher mean. Note that we still operate on continuous fields, but only cluster samples to guide our training. + +# 4.3. Training + +CaFi-Net is trained on a large dataset of pre-trained NeRF models over 13 object categories (see Section 5). We use the following loss functions to train our model. + +Canonicalization Loss: To self-supervise our learning, we transform the predicted canonical grid coordinates to the input grid using the canonical rotation $E_{b}$ and compute a point-wise L2 distance between them. This loss forces the predicted invariant grid coordinates $P$ to reconstruct the shape and forces the predicted canonicalizing transform to be equivariant to the input transformation. The final loss is computed over the canonicalizing transform that minimizes this loss: + +$$ +E _ {b} := \min _ {j} \left(\operatorname {m e a n} _ {i \in C _ {f}} \left\| X _ {i} - E _ {j} P _ {i} \right\| _ {2} ^ {2}\right), \tag {7} +$$ + +$$ +\mathcal {L} _ {\text {c a n o n}} := \operatorname {m e a n} _ {i \in C _ {f}} \left| \left| \boldsymbol {X} _ {i} - E _ {b} P _ {i} \right| \right| _ {2} ^ {2}, \tag {8} +$$ + +where $C_f$ is a set of points belonging to the foreground. + +**Orthonormality Loss:** The predicted equivariant transformations $E$ should have orthonormal vectors and we use the following to force orthonormality using: + +$$ +U _ {j}, D _ {j}, V _ {j} := \operatorname {S V D} \left(E _ {j}\right), \tag {9} +$$ + +$$ +\mathcal {L} _ {\text {o r t h o}} := \frac {1}{M} \sum_ {j} \left| \left| E _ {j} - U _ {j} V _ {j} ^ {T} \right| \right| _ {2}, \tag {10} +$$ + +where $\operatorname{SVD}(E_j)$ computes the singular value decomposition of the $j^{th}$ equivariant transformation. + +Siamese Shape Loss: We use a Siamese loss that penalizes for consistency between different shapes belonging to the same category. Given two un-canonicalized fields in different frames of reference $\sigma_{D}^{1}$ and $\sigma_{D}^{2}$ , we canonicalize both the fields to the canonical frame and compute chamfer distance between their predicted canonical coordinates for points in the foreground cluster $C_f$ , i.e., + +$$ +\mathcal {L} _ {\text {s i a m e s e}} := \mathrm {C D} \left(P \left(\boldsymbol {\sigma} _ {D} ^ {1}\right), P \left(\boldsymbol {\sigma} _ {D} ^ {2}\right)\right). \tag {11} +$$ + +This loss regularizes the training by ensuring that instances of the same category should be aligned in the canonical frame. + +Architecture and Hyper-parameters: CaFi-Net predicts a 128 dimensional invariant embedding for each point in space and performs convolution at resolutions $1/2$ , $1/4$ , and $1/8$ . Each equivariant convolution layer aggregates features from 512 neighboring points at twice the resolution. We then use equivariant non-linearities introduced in [34] by applying Batch Normalization [14] and ReLU activation [2] in the inverse spherical harmonic transform domain. We use a three-layer MLP with intermediate Batch Normalization and ReLU layers followed by a final linear layer to predict canonical coordinates. To train our network, we weigh the canonicalization loss $(\mathcal{L}_{canon})$ the highest with weight 2.0 and 1.0 for all the other loss functions. All our models are written in PyTorch [31] and are trained for 300 epochs with a batch size of 2 on an Nvidia 1080-Ti GPU. We use the Adam Optimizer [16] with an initial learning rate of $6e - 4$ and L2 weight regularization of $1e - 5$ for all our experiments. Please see the supplementary document for more details. + +# 5. Experiments + +We now provide details on our extensive experiments to evaluate the quality of our pose canonicalization. To our knowledge, we are the first method for pose canonicalization of neural radiance fields which makes direct comparisons with other work challenging. Nonetheless, we have designed experiments to compare our method with three 3D point cloud-based canonicalization methods. Overall, we present three classes of experiments: (1) evaluation of the performance of our method on 13 different categories, + +Table 1. This table compares the canonicalization performance of our method (F - operating on fields) with other 3D point cloud-based methods (P) on three standard metrics (IC, CC and GEC) on our dataset of 13 categories. We compare with PCA, Canonical Capsules (CaCa) [47] and ConDor [37]. All metrics are multiplied by 100 for ease of reading. The top two performing methods are highlighted in magenta (best) and blue (second best). We are better than SOTA [37] on the Ground Truth Equivariance Consistency (GEC) and Category-Level Consistency (CC) with lower mean and median canonicalization error. However, we perform on par with ConDor [37] on the Instance-Level Consistency. + +
benchcabinetcarcellph.chaircouchfirearmlampmonitorplanespeakertablewater.avg.med.
Instance-Level Consistency (IC)↓
PCA (P)14.655.946.130.676.137.6315.0712.846.787.905.4010.3110.728.477.63
CaCa (P) [47]2.892.092.012.811.081.900.232.972.351.512.212.422.602.082.21
ConDor (P) [37]0.561.270.360.530.960.320.603.780.440.641.971.322.151.150.64
Ours (F)2.621.150.391.280.811.430.743.870.410.360.601.092.741.341.09
Category-Level Consistency (CC)↓
PCA (P)14.559.687.484.269.9013.4515.1311.8713.098.155.5110.4311.0710.3610.43
CaCa (P) [47]2.790.381.050.831.821.591.314.340.391.830.651.912.061.611.59
ConDor (P) [37]1.981.680.471.041.280.880.964.441.581.262.102.102.561.721.58
Ours (F)2.771.360.481.260.751.590.923.920.580.630.771.482.351.451.26
Ground Truth Equivalence Consistency (GEC)↓
PCA (P)14.405.866.321.646.328.1215.0112.496.828.165.3210.1711.278.618.12
CaCa (P) [47]3.652.282.402.741.902.221.284.692.482.092.223.012.792.592.40
ConDor (P) [37]2.121.790.501.171.340.931.044.581.591.312.222.242.671.811.59
Ours (F)3.261.510.541.430.941.901.014.370.630.660.811.633.021.671.43
+ +(2) comparison with other work, and (3) ablations to justify method design choices. + +Dataset: All our experiments use a new large synthetic NeRF dataset of shapes from 13 categories (that overlap with [12, 37, 47]) from the ShapeNet [5] dataset. For each category, we pick 100 instances from ShapeNet, rotate the shape randomly, and render 54 omnidirectional views sampled in a cube around the shape using Blender. Different from other datasets, we also simulate cluttered backgrounds. These posed views are used to train 100 NeRF models per category using the public PyTorch implementation of NeRF [62] for a total of 1300 NeRF models. To our knowledge, this is one of the largest (compared to [11, 53]) $360^{\circ}$ synthetic NeRF datasets - we will publicly release the simulator, raw posed images, and trained NeRF weights. Each NeRF model is trained for 400 epochs with 1024 randomly selected rays per iteration, a coarse sampling resolution of 64 points, and a fine sampling resolution of 128 points along each ray. Rather than train NeRF to maximize PSNR on a novel view, we fix the number of epochs for all models to enable a fair comparison between instances. As a consequence, some NeRF models may have a lower PSNR, but our goal is not to increase NeRF quality but rather improve canonicalization quality - hence we fix the number of epochs. For each category, we set aside $20\%$ of the models for testing. Metrics: Since there are no known metrics for evaluating canonical fields, we resort to 3 metrics used for 3D point cloud canonicalization [37]: (1) Instance-Level Consistency (IC), a metric that measures how consistently we can canonicalize the same instance in different poses, (2) Category- + +Level Consistency (CC) to measure how consistently we canonicalize across different instances in the same category, and (3) Ground Truth Equivariance Consistency (GEC), a variant of the Ground Truth Consistency (GC) metric proposed in [37] to measure canonicalization performance compared to manual labels. We use the GEC because we observed that the original GC metric is prone to degeneracy when the canonicalizing transform is identity. We fix this issue by taking three point clouds $P_{i}, P_{j}$ , and $P_{k}$ from a canonicalized dataset $\mathcal{P}$ and rotating $P_{i}$ by $R_{1}, P_{j}$ by $R_{2}$ where both $R_{1}$ and $R_{2}$ are random rotations. Let $\mathcal{C}(P)$ predict a canonicalizing rotation for a point cloud $P$ , we then compute the Ground Truth Equivariance consistency + +as: $\mathrm{GEC} \coloneqq \frac{1}{|\mathcal{P}|^3} \sum_{P_i, P_j, P_k \in \mathcal{P}} \mathrm{CD}(\mathcal{C}(R_1P_i)R_1P_k, \mathcal{C}(R_2P_j)R_2P_k)$ , (12) where CD refers to the Chamfer distance. Here, we apply canonicalizing transforms of shape $P_i$ and $P_j$ to the shape $P_k$ . This modified metric will not be degenerate for identity canonicalizing transforms and is more rigorous in evaluating ground truth consistency. To compute our metrics, we used point clouds sampled from the Shapenet meshes [5] using poisson disk sampling. + +# 5.1. Evaluation & Comparisons + +We first analyze the performance of our method on the three metrics across 13 different categories. In Table 1, we highlight our method that operates on a field as Ours (F). Our method performs well across all categories with low variance between different categories. We observe similar trends across metrics for a set of categories (e.g., firearms,lamps) indicating that category shape plays a + +
Categoryxyz vs. Gradient Signals
xyzGradient signal
Ground Truth Equivariance Consistency (GEC)↓
bench2.173.26
cellphone1.841.43
chair1.720.94
plane2.070.66
Average1.951.57
Instance-Level Consistency (IC)↓
bench2.042.62
cellphone1.61.28
chair1.40.81
plane1.880.36
Average1.731.26
Category-Level Consistency (CC)↓
bench1.992.77
cellphone1.491.26
chair1.520.75
plane1.90.63
Average1.721.35
+ +(a) Choice of Signal Representation - Canonicalization metrics for using Gradients vs. $xyz$ locations as input signal. Gradients capture the object surface that help in canonicalization. + +
CategoryLocal Average Density
w/ow
Ground Truth Equivariance Consistency (GEC)↓
bench3.263.38
cellphone1.431.63
chair0.941.3
plane0.660.64
Average1.571.73
Instance-Level Consistency (IC)↓
bench2.622.72
cellphone1.281.63
chair0.812.64
plane0.360.39
Average1.261.85
Category-Level Consistency (CC)↓
bench2.772.82
cellphone1.261.54
chair0.751.07
plane0.630.51
Average1.351.49
+ +(b) Weighing Equivariant Signals by Local Average Density deteriorates the performance by smoothing out important details of the shape. We show canonicalization with (w) and without (w/o) weighing by the local averaged density. + +
CategorySiamese Training
w/o siamesewith siamese
Ground Truth Equivariance Consistency (GEC)↓
bench3.523.26
cellphone1.611.43
chair1.020.94
plane1.310.66
Average1.861.57
Instance-Level Consistency (IC)↓
bench2.732.62
cellphone1.31.28
chair0.820.81
plane1.10.36
Average1.481.26
Category-Level Consistency (CC)↓
bench2.952.77
cellphone1.441.26
chair0.830.75
plane1.260.63
Average1.621.35
+ +(c) Siamese Training improves performance on all canonicalization metrics on average. We show canonicalization performance with siamese and without (w/o) siamese training. The average of Ground Truth Equivariance Consistency GEC metric reduces to 1.57 from 1.86 + +Table 2. Ablation studies to justify three key design choices: (a) signal representation, (b) density weighting, and (c) Siamese architecture. + +role in canonicalization performance. + +Next, we compare our method against three benchmarks: Principal Component Analysis (PCA) [32], Canonical Capsules (CaCa) [47], and ConDor [37]. All prior methods canonicalize clean point clouds with all points on the surface. To have a fair comparison, we train the above methods on point clouds (P) sampled from NeRF scenes for points on and within the object. Most scenes have points in the range of 120-1200 when sampled uniformly on the $32^{3}$ grid. We resample the scene when points are less than 1024 and train both Canonical Capsules and ConDor on the resulting point clouds. CaFi-Net estimates a canonicalizing transform directly from the density field that we use to transform its corresponding ShapeNet point cloud for comparison against point canonicalizers. We follow the same protocol as [37] to measure and benchmark canonicalization performance for all the methods. + +Analysis: Table 1 compares the metrics for PCA (P), CaCa (P), ConDor (P), and CaFi-Net [Ours (F)]. We list point cloud canonicalizers with (P) and field canonicalizer as (F) in the table. On average, our method performs better than point cloud-based methods in the CC and GEC metrics and is on par with the state-of-the-art method (ConDor) in the IC metric. This is despite the fact that our method directly operates on the field. We observe that PCA underperforms in all categories suggesting that the noisy point clouds make it unreliable. Similarly we observe that CaCa fails to canonicalize in categories where high noise is observed. Given + +that ConDor also uses TFNs in a different architecture, we conclude that the choice of TFNs makes canonicalization more robust to noise. Our method has the additional advantage of operating directly on the field. Please see Figure 4 for a qualitative comparison of the different methods. + +# 5.2. Ablations + +We conduct additional experiments to justify key design choices in our method. For all the ablation studies we use a smaller subset of our data consisting of 4 categories: bench, cellphone, chair, and airplane. + +Choice of Signal Representation: To select the appropriate signal representation to use (see Section 4), we conducted an ablation study using just the density field $xyz$ locations or its gradients. As seen in Table 2 (a), the average canonicalization error over 4 categories is lower when using gradient as compared to using density $xyz$ locations on all three metrics. We hypothesize that gradient of density weighs equivariant features closer to the surface of the object higher. This improves the canonicalization performance. + +Weighing Features by Local Average Density: Next, we justify why weighting the equivariant features using the density is helpful. Table 2 (b) shows results for weighting features by local average density compared to weighting features by density. Weighing features by local average smoothens out details that are necessary to canonicalize the volume. Our average ground truth canonicalization error + +![](images/d570ced4afbb5858f75f9c7cf2ab418bfd5caed7eea2ab3266f24df6a4d21b34.jpg) +Input NeRFs +CaFi-Net (Ours) +PCA +Figure 4. This figure visually compares the canonicalization performance of our method with 3D point cloud-based methods. Input NeRFs shows one of the RGB views used to learn our NeRF input (notice how orientations of instances are random). CaFi-Net (Ours) shows the results of our canonical field rendered from a novel view unseen during NeRF training. PCA (blue), CaCa [47] (red), and ConDor [37]'s (green) results are also presented for the same instances. Our method matches or exceeds ConDor. We also show some failure cases for several instances (red box) which occur either due to thin structures or symmetry. + +increases from 1.57 to 1.73 if we weigh features by local averaged densities. In future works, we will delve into better methods for reducing input signal noise that does not smoothen out the geometry details of objects to further improve our canonicalization. + +Siamese Architecture: Finally, we justify the need for a two-branch Siamese network architecture. Although our method can canonicalize even without this architecture, it makes it easier to learn over instances in a category as observed in [26]. Table 2 (c) compares our method on using a single branch architecture and a Siamese architecture. Clearly, the Siamese architecture helps improve results confirming our hypothesis. Enforcing different instances of the same category to align with each other reduces the Ground Truth Canonicalization error from 1.86 to 1.57. We also see a decrease in all the other metrics including Instance-Level Consistency. + +# 6. Conclusion + +We presented Canonical Field Network (CaFi-Net), a method for self-supervised category-level canonicalization of the 3D pose of objects represented as neural radiance fields. CaFi-Net consists of a Siamese network architecture with rotation-equivariant convolution layers to extract + +features for pose canonicalization. Our approach operates directly on NeRF's noisy but continuous density field. We train our method on a large dataset of 1300 NeRF models obtained for 13 common ShapeNet categories. During inference, our method is able to canonicalize arbitrarily oriented NeRFs. Experiments show that our method matches or outperforms 3D point cloud-based methods. + +Limitations & Future Work: Our approach has several limitations that future work should investigate. First, we require a $360^{\circ}$ NeRF model of objects and cannot handle partial views from front-facing NeRFs. Second, canonicalizing densities with uncertainty is more difficult as the uncertainty increases for very thin/small structures that can be missed (see chair in Figure 4). Furthermore, we also inherit the issues that TFNs have with symmetry (see monitor case in Figure 4). In future work, we plan to extend our approach to create a large real dataset of common object categories without requiring manual pose canonicalization and combat the symmetry issue in CaFi-Net as in [39] + +Acknowledgments: This work was supported by NSF grant CNS-2038897, an AWS Cloud Credits award, NSF CloudBank, and a gift from Meta Reality Labs. We thank Chandradeep Pokhariya and Ishaan Shah. + +# References + +[1] Michal Adamkiewicz, Timothy Chen, Adam Caccavale, Rachel Gardner, Preston Culbertson, Jeannette Bohg, and Mac Schwager. Vision-only robot navigation in a neural radiance world. IEEE Robotics and Automation Letters, 7(2):4606-4613, 2022. 1 +[2] Abien Fred Agarap. Deep learning using rectified linear units (relu). arXiv preprint arXiv:1803.08375, 2018. 5 +[3] Jonathan T Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P Srinivasan. Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5855–5864, 2021. 1, 2 +[4] Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Mip-nerf 360: Unbounded anti-aliased neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5470–5479, 2022. 1, 2 +[5] Angel X. Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, Jianxiong Xiao, Li Yi, and Fisher Yu. ShapeNet: An Information-Rich 3D Model Repository. Technical Report arXiv:1512.03012 [cs.GR], Stanford University — Princeton University — Toyota Technological Institute at Chicago, 2015. 1, 2, 6 +[6] Zhiqin Chen and Hao Zhang. Learning implicit fields for generative shape modeling. Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 1 +[7] Taco S Cohen, Mario Geiger, Jonas Kohler, and Max Welling. Spherical cnns. arXiv preprint arXiv:1801.10130, 2018.3 +[8] Enric Corona, Tomas Hodan, Minh Vo, Francesc Moreno-Noguer, Chris Sweeney, Richard Newcombe, and Lingni Ma. Lisa: Learning implicit shape and appearance of hands. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20533-20543, 2022. 2 +[9] Congyue Deng, Or Litany, Yueqi Duan, Adrien Poulenard, Andrea Tagliasacchi, and Leonidas J. Guibas. Vector neurons: A general framework for so(3)-equivariant networks. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 12200-12209, October 2021. 2, 3 +[10] Fabian B Fuchs, Daniel E Worrall, Volker Fischer, and Max Welling. Se (3)-transformers: 3d roto-translation equivariant attention networks. arXiv preprint arXiv:2006.10503, 2020. 2 +[11] Ruohan Gao, Yen-Yu Chang, Shivani Mall, Li Fei-Fei, and Jiajun Wu. Objectfolder: A dataset of objects with implicit visual, auditory, and tactile representations. In Proceedings of the Conference on Robot Learning (CoRL), 2021. 2, 6 +[12] Thibault Groueix, Matthew Fisher, Vladimir G Kim, Bryan C Russell, and Mathieu Aubry. A papier-mâché approach to learning 3d surface generation. In Proceedings of + +the IEEE conference on computer vision and pattern recognition, pages 216-224, 2018. 2, 6 +[13] Jeffrey Ichnowski, Yahav Avigal, Justin Kerr, and Ken Goldberg. Dex-nerf: Using a neural radiance field to grasp transparent objects. Oct 2021. 1 +[14] Sergey Ioffe and Christian Szegedy. Batch normalization: Accelerating deep network training by reducing internal covariate shift. In Francis Bach and David Blei, editors, Proceedings of the 32nd International Conference on Machine Learning, volume 37 of Proceedings of Machine Learning Research, pages 448-456, Lille, France, 07-09 Jul 2015. PMLR. 5 +[15] Ajay Jain, Matthew Tancik, and Pieter Abbeel. Putting nerf on a diet: Semantically consistent few-shot view synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5885-5894, 2021. 2 +[16] Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. CoRR, abs/1412.6980, 2015. 5 +[17] Jiahui Lei, Srinath Sridhar, Paul Guerrero, Minhyuk Sung, Niloy Mitra, and Leonidas J Guibas. Pix2surf: Learning parametric 3d surface models of objects from images. In European Conference on Computer Vision, pages 121-138. Springer, 2020. 2 +[18] Ruilong Li, Julian Tanke, Minh Vo, Michael Zollhofer, Jurgen Gall, Angjoo Kanazawa, and Christoph Lassner. Tava: Template-free animatable volumetric actors. arXiv preprint arXiv:2206.08929, 2022. 2 +[19] Tianye Li, Mira Slavcheva, Michael Zollhoefer, Simon Green, Christoph Lassner, Changil Kim, Tanner Schmidt, Steven Lovegrove, Michael Goesele, Richard Newcombe, et al. Neural 3d video synthesis from multi-view video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5521-5531, 2022. 1, 2 +[20] Xiaolong Li, He Wang, Li Yi, Leonidas J Guibas, A Lynn Abbott, and Shuran Song. Category-level articulated object pose estimation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3706-3715, 2020. 2 +[21] Yunzhu Li, Shuang Li, Vincent Sitzmann, Pulkit Agrawal, and Antonio Torralba. 3d neural scene representations for visuomotor control. In Proceedings of Robotics: Science and Systems, 2021. 1 +[22] Lars Mescheder, Michael Oechsle, Michael Niemeyer, Sebastian Nowozin, and Andreas Geiger. Occupancy networks: Learning 3d reconstruction in function space. In Proceedings IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2019. 1, 2 +[23] Ben Mildenhall, Peter Hedman, Ricardo Martin-Brualla, Pratul P Srinivasan, and Jonathan T Barron. Nerf in the dark: High dynamic range view synthesis from noisy raw images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16190-16199, 2022. 2 +[24] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view syn + +thesis. In European conference on computer vision, pages 405-421. Springer, 2020. 1, 2 +[25] Michael Niemeyer, Lars Mescheder, Michael Oechsle, and Andreas Geiger. Occupancy flow: 4d reconstruction by learning particle dynamics. In Proc. of the IEEE International Conf. on Computer Vision (ICCV), 2019. 1 +[26] David Novotny, Nikhila Ravi, Benjamin Graham, Natalia Neverova, and Andrea Vedaldi. C3dpo: Canonical 3d pose networks for non-rigid structure from motion. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7688-7697, 2019. 2, 3, 4, 5, 8 +[27] David Novotny, Roman Shapovalov, and Andrea Vedaldi. Canonical 3D Deformer Maps: Unifying parametric and non-parametric methods for dense weakly-supervised category reconstruction. In NeurIPS, 2020. 2 +[28] Jeong Joon Park, Peter Florence, Julian Straub, Richard Newcombe, and Steven Lovegrove. Deepsdf: Learning continuous signed distance functions for shape representation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 165-174, 2019. 1, 2 +[29] Keunhong Park, Utkarsh Sinha, Jonathan T. Barron, Sofien Bouaziz, Dan B Goldman, Steven M. Seitz, and Ricardo Martin-Brualla. Nerfies: Deformable neural radiance fields. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), 2021. 1, 2 +[30] Keunhong Park, Utkarsh Sinha, Peter Hedman, Jonathan T. Barron, Sofien Bouaziz, Dan B Goldman, Ricardo MartinBrualla, and Steven M. Seitz. Hypernerf: A higher-dimensional representation for topologically varying neural radiance fields. arXiv preprint arXiv:2106.13228, 2021. 1, 2 +[31] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, Alban Desmaison, Andreas Kopf, Edward Yang, Zachary DeVito, Martin Raison, Alykhan Tejani, Sasank Chilamkurthy, Benoit Steiner, Lu Fang, Junjie Bai, and Soumith Chintala. Pytorch: An imperative style, high-performance deep learning library. In H. Wallach, H. Larochelle, A. Beygelzimer, F. d'Alché-Buc, E. Fox, and R. Garnett, editors, Advances in Neural Information Processing Systems 32, pages 8024-8035. Curran Associates, Inc., 2019. 5 +[32] Karl Pearson. Liii. on lines and planes of closest fit to systems of points in space. The London, Edinburgh, and Dublin philosophical magazine and journal of science, 2(11):559-572, 1901. 7 +[33] Sida Peng, Junting Dong, Qianqian Wang, Shangzhan Zhang, Qing Shuai, Xiaowei Zhou, and Hujun Bao. Animatable neural radiance fields for modeling dynamic human bodies. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14314-14323, 2021. 1 +[34] Adrien Poulenard and Leonidas J Guibas. A functional approach to rotation equivariant non-linearities for tensor field networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13174-13183, 2021. 3, 4, 5 +[35] Maziar Raissi, Paris Perdikaris, and George E Karniadakis. Physics-informed neural networks: A deep learning framework for solving forward and inverse problems involving + +nonlinear partial differential equations. Journal of Computational Physics, 378:686-707, 2019. 2 +[36] Viktor Rudnev, Mohamed Elgharib, William Smith, Lingjie Liu, Vladislav Golyanik, and Christian Theobalt. Nerf for outdoor scene relighting. In European Conference on Computer Vision (ECCV), 2022. 2 +[37] Rahul Sajnani, Adrien Poulenard, Jivitesh Jain, Radhika Dua, Leonidas J Guibas, and Srinath Sridhar. Condor: Self-supervised canonicalization of 3d pose for partial shapes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16969-16979, 2022. 2, 5, 6, 7, 8 +[38] Rahul Sajnani, AadilMehdi Sanchawala, Krishna Murthy Jatavallabhula, Srinath Sridhar, and K Madhava Krishna. Draco: Weakly supervised dense reconstruction and canonicalization of objects. In 2021 IEEE International Conference on Robotics and Automation (ICRA), pages 10302-10309. IEEE, 2021. 2, 3 +[39] Ahyun Seo, Byungjin Kim, Suha Kwak, and Minsu Cho. Reflection and rotation symmetry detection via equivariant learning. In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 8 +[40] Anthony Simeonov, Yilun Du, Andrea Tagliasacchi, Joshua B Tenenbaum, Alberto Rodriguez, Pulkit Agrawal, and Vincent Sitzmann. Neural descriptor fields: Se (3)-equivariant object representations for manipulation. In 2022 International Conference on Robotics and Automation (ICRA), pages 6394-6400. IEEE, 2022. 1 +[41] Anthony Simeonov, Yilun Du, Andrea Tagliasacchi, Joshua B. Tenenbaum, Alberto Rodriguez, Pulkit Agrawal, and Vincent Sitzmann. Neural descriptor fields: Se(3)-equivariant object representations for manipulation. 2022. 2 +[42] Vincent Sitzmann, Michael Zollhoefer, and Gordon Wetzstein. Scene representation networks: Continuous 3d-structure-aware neural scene representations. In H. Wallach, H. Larochelle, A. Beygelzimer, F. d'Alché-Buc, E. Fox, and R. Garnett, editors, Advances in Neural Information Processing Systems, volume 32. Curran Associates, Inc., 2019. 1 +[43] Riccardo Spezialietti, Federico Stella, Marlon Marcon, Luciano Silva, Samuele Salti, and Luigi Di Stefano. Learning to orient surfaces by self-supervised spherical cnns. arXiv preprint arXiv:2011.03298, 2020. 1, 2, 3 +[44] Srinath Sridhar, Davis Rempe, Julien Valentin, Sofien Bouaziz, and Leonidas J Guibas. Multiview aggregation for learning category-specific shape reconstruction. arXiv preprint arXiv:1907.01085, 2019. 2 +[45] Pratul P Srinivasan, Boyang Deng, Xiuming Zhang, Matthew Tancik, Ben Mildenhall, and Jonathan T Barron. Nerv: Neural reflectance and visibility fields for relighting and view synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7495-7504, 2021. 2 +[46] Shih-Yang Su, Frank Yu, Michael Zollhöfer, and Helge Rhodin. A-nerf: Articulated neural radiance fields for learning human shape, appearance, and pose. Advances in Neural Information Processing Systems, 34:12278-12291, 2021. 2 + +[47] Weiwei Sun, Andrea Tagliasacchi, Boyang Deng, Sara Sabour, Soroosh Yazdani, Geoffrey Hinton, and Kwang Moo Yi. Canonical capsules: Unsupervised capsules in canonical pose. arXiv preprint arXiv:2012.04718, 2020. 1, 2, 3, 6, 7, 8 +[48] Maxim Tatarchenko, Stephan R Richter, René Ranftl, Zhuwen Li, Vladlen Koltun, and Thomas Brox. What do single-view 3d reconstruction networks learn? In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3405-3414, 2019. 2 +[49] Ayush Tewari, Justus Thies, Ben Mildenhall, Pratul Srinivasan, Edgar Tretschk, Yifan Wang, Christoph Lassner, Vincent Sitzmann, Ricardo Martin-Brualla, Stephen Lombardi, Tomas Simon, Christian Theobalt, Matthias Niessner, Jonathan T. Barron, Gordon Wetzstein, Michael Zollhoefer, and Vladislav Golyanik. Advances in neural rendering. Nov 2021. 2 +[50] Nathaniel Thomas, Tess Smidt, Steven Kearnes, Lusann Yang, Li Li, Kai Kohlhoff, and Patrick Riley. Tensor field networks: Rotation-and translation-equivariant neural networks for 3d point clouds. arXiv preprint arXiv:1802.08219, 2018. 2, 3, 4 +[51] Garvita Tiwari, Dimitrije Antic, Jan Eric Lenssen, Nikolaos Sarafianos, Tony Tung, and Gerard Pons-Moll. Pose-ndf: Modeling human pose manifolds with neural distance fields. In European Conference on Computer Vision (ECCV), October 2022. 2 +[52] Edgar Tretschk, Ayush Tewari, Vladislav Golyanik, Michael Zollhöfer, Christoph Lassner, and Christian Theobalt. Non-rigid neural radiance fields: Reconstruction and novel view synthesis of a dynamic scene from monocular video. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12959–12970, 2021. 1 +[53] Suhani Vora, Noha Radwan, Klaus Greff, Henning Meyer, Kyle Genova, Mehdi S. M. Sajjadi, Etienne Pot, Andrea Tagliasacchi, and Daniel Duckworth. Nesf: Neural semantic fields for generalizable semantic segmentation of 3d scenes, 2021. 1, 6 +[54] He Wang, Srinath Sridhar, Jingwei Huang, Julien Valentin, Shuran Song, and Leonidas J Guibas. Normalized object coordinate space for category-level 6d object pose and size estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2642-2651, 2019. 2 +[55] Peng Wang, Lingjie Liu, Yuan Liu, Christian Theobalt, Taku Komura, and Wenping Wang. Neus: Learning neural implicit surfaces by volume rendering for multi-view reconstruction. In Proceedings of the Thirtieth International Joint Conference on Artificial Intelligence (IJCAI). International Joint Conferences on Artificial Intelligence Organization, 2021. 1 +[56] Maurice Weiler, Mario Geiger, Max Welling, Wouter Boomsma, and Taco Cohen. 3d steerable cnns: Learning rotationally equivariant features in volumetric data. In NIPS, pages 10381-10392, 2018. 2 +[57] Maurice Weiler, Fred A Hamprecht, and Martin Storath. Learning steerable filters for rotation equivariant cnns. In Proc. CVPR, pages 849-858, 2018. 2 +[58] Zhirong Wu, Shuran Song, Aditya Khosla, Fisher Yu, Linguang Zhang, Xiaou Tang, and Jianxiong Xiao. 3d + +shapenets: A deep representation for volumetric shapes. In Proc. CVPR, pages 1912-1920, 2015. 2 +[59] Yiheng Xie, Towaki Takikawa, Shunsuke Saito, Or Litany, Shiqin Yan, Numair Khan, Federico Tombari, James Tompkin, Vincent Sitzmann, and Srinath Sridhar. Neural fields in visual computing and beyond. Computer Graphics Forum, 2022. 1, 2 +[60] Guandao Yang, Serge Belongie, Bharath Hariharan, and Vladlen Koltun. Geometry processing with neural fields. Advances in Neural Information Processing Systems, 34:22483-22497, 2021. 1 +[61] Lior Yariv, Jiatao Gu, Yoni Kasten, and Yaron Lipman. Volume rendering of neural implicit surfaces. In Advances in Neural Information Processing Systems (NeurIPS). Curran Associates, Inc., 2021. 1 +[62] Lin Yen-Chen. Nerf-pytorch. https://github.com/yenchenlin/nerf-pytorch/, 2020.6 +[63] Alex Yu, Vickie Ye, Matthew Tancik, and Angjoo Kanazawa. pixelnerf: Neural radiance fields from one or few images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4578-4587, 2021. 1, 2 +[64] Kevin Zakka, Andy Zeng, Johnny Lee, and Shuran Song. Form2fit: Learning shape priors for generalizable assembly from disassembly. In 2020 IEEE International Conference on Robotics and Automation (ICRA), pages 9404-9410. IEEE, 2020. 2 +[65] Ge Zhang, Or Litany, Srinath Sridhar, and Leonidas Guibas. Strobenet: Category-level multiview reconstruction of articulated objects. arXiv preprint arXiv:2105.08016, 2021. 2 +[66] Kai Zhang, Gernot Riegler, Noah Snavely, and Vladlen Koltun. Nerf++: Analyzing and improving neural radiance fields. arXiv preprint arXiv:2010.07492, 2020. 1, 2 +[67] Yongheng Zhao, Tolga Birdal, Haowen Deng, and Federico Tombari. 3d point capsule networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1009-1018, 2019. 3 +[68] Shuaifeng Zhi, Tristan Laidlow, Stefan Leutenegger, and Andrew J. Davison. In-place scene labelling and understanding with implicit scene representation. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), 2021. 1 +[69] Ellen D Zhong, Adam Lerer, Joseph H Davis, and Bonnie Berger. Cryodrgn2: Ab initio neural reconstruction of 3d protein structures from real cryo-em images. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4066-4075, 2021. 2 \ No newline at end of file diff --git a/2023/Canonical Fields_ Self-Supervised Learning of Pose-Canonicalized Neural Fields/images.zip b/2023/Canonical Fields_ Self-Supervised Learning of Pose-Canonicalized Neural Fields/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..a08d9953a3d0e9e359927eb839877d9b738af32d --- /dev/null +++ b/2023/Canonical Fields_ Self-Supervised Learning of Pose-Canonicalized Neural Fields/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:34e05881d553031c3ba24b2abac38f41b99a674d11bba7d1119207385ec5de88 +size 517207 diff --git a/2023/Canonical Fields_ Self-Supervised Learning of Pose-Canonicalized Neural Fields/layout.json b/2023/Canonical Fields_ Self-Supervised Learning of Pose-Canonicalized Neural Fields/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..351a029563995a6a92ffe22073ff7e040a7b8c94 --- /dev/null +++ b/2023/Canonical Fields_ Self-Supervised Learning of Pose-Canonicalized Neural Fields/layout.json @@ -0,0 +1,9715 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 49, + 103, + 544, + 120 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 103, + 544, + 120 + ], + "spans": [ + { + "bbox": [ + 49, + 103, + 544, + 120 + ], + "type": "text", + "content": "Canonical Fields: Self-Supervised Learning of Pose-Canonicalized Neural Fields" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 94, + 142, + 176, + 158 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 94, + 142, + 176, + 158 + ], + "spans": [ + { + "bbox": [ + 94, + 142, + 176, + 158 + ], + "type": "text", + "content": "Rohith Agaram1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 200, + 143, + 282, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 200, + 143, + 282, + 157 + ], + "spans": [ + { + "bbox": [ + 200, + 143, + 282, + 157 + ], + "type": "text", + "content": "Shaurya Dewan" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 211, + 158, + 301, + 170 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 211, + 158, + 301, + 170 + ], + "spans": [ + { + "bbox": [ + 211, + 158, + 301, + 170 + ], + "type": "text", + "content": "Madhava Krishna1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 307, + 143, + 382, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 143, + 382, + 157 + ], + "spans": [ + { + "bbox": [ + 307, + 143, + 382, + 157 + ], + "type": "text", + "content": "Rahul Sajnani" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 406, + 143, + 498, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 406, + 143, + 498, + 156 + ], + "spans": [ + { + "bbox": [ + 406, + 143, + 498, + 156 + ], + "type": "text", + "content": "Adrien Poulenard3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 327, + 158, + 406, + 170 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 158, + 406, + 170 + ], + "spans": [ + { + "bbox": [ + 327, + 158, + 406, + 170 + ], + "type": "text", + "content": "Srinath Sridhar" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 148, + 172, + 242, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 148, + 172, + 242, + 186 + ], + "spans": [ + { + "bbox": [ + 148, + 172, + 242, + 186 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 148, + 172, + 242, + 186 + ], + "type": "text", + "content": "RRC, IIIT-Hyderabad" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 262, + 172, + 340, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 262, + 172, + 340, + 186 + ], + "spans": [ + { + "bbox": [ + 262, + 172, + 340, + 186 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 262, + 172, + 340, + 186 + ], + "type": "text", + "content": "Brown University" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 360, + 172, + 445, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 360, + 172, + 445, + 186 + ], + "spans": [ + { + "bbox": [ + 360, + 172, + 445, + 186 + ], + "type": "text", + "content": "3Stanford University" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 196, + 188, + 394, + 201 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 196, + 188, + 394, + 201 + ], + "spans": [ + { + "bbox": [ + 196, + 188, + 394, + 201 + ], + "type": "text", + "content": "ivl.cs.brown.edu/projects/canonicalfields" + } + ] + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 49, + 221, + 296, + 319 + ], + "blocks": [ + { + "bbox": [ + 146, + 209, + 199, + 220 + ], + "lines": [ + { + "bbox": [ + 146, + 209, + 199, + 220 + ], + "spans": [ + { + "bbox": [ + 146, + 209, + 199, + 220 + ], + "type": "text", + "content": "Input NeRFs" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 49, + 221, + 296, + 319 + ], + "lines": [ + { + "bbox": [ + 49, + 221, + 296, + 319 + ], + "spans": [ + { + "bbox": [ + 49, + 221, + 296, + 319 + ], + "type": "image", + "image_path": "3d97b753ffa37973fda8c606cdd8ac9d31052170813520d893589850268afb99.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 296, + 221, + 545, + 319 + ], + "blocks": [ + { + "bbox": [ + 386, + 209, + 455, + 219 + ], + "lines": [ + { + "bbox": [ + 386, + 209, + 455, + 219 + ], + "spans": [ + { + "bbox": [ + 386, + 209, + 455, + 219 + ], + "type": "text", + "content": "Canonical Fields" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 296, + 221, + 545, + 319 + ], + "lines": [ + { + "bbox": [ + 296, + 221, + 545, + 319 + ], + "spans": [ + { + "bbox": [ + 296, + 221, + 545, + 319 + ], + "type": "image", + "image_path": "10d00d5ec06704c055fab2229ceebf64850b7ae311344eae598010bbdc85132c.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 322, + 546, + 365 + ], + "lines": [ + { + "bbox": [ + 46, + 322, + 546, + 365 + ], + "spans": [ + { + "bbox": [ + 46, + 322, + 546, + 365 + ], + "type": "text", + "content": "Figure 1. We present Canonical Field Network (CaFi-Net), a self-supervised method for 3D position and orientation (pose) canonicalization of objects represented as neural fields. We specifically focus on neural radiance fields (NeRFs) fitted to raw RGB images of arbitrarily posed objects (left), and obtain a canonical field (fixed novel view shown on right) where all objects in a category are consistently positioned and oriented." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "bbox": [ + 143, + 369, + 191, + 381 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 369, + 191, + 381 + ], + "spans": [ + { + "bbox": [ + 143, + 369, + 191, + 381 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 45, + 394, + 290, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 394, + 290, + 647 + ], + "spans": [ + { + "bbox": [ + 45, + 394, + 290, + 647 + ], + "type": "text", + "content": "Coordinate-based implicit neural networks, or neural fields, have emerged as useful representations of shape and appearance in 3D computer vision. Despite advances, however, it remains challenging to build neural fields for categories of objects without datasets like ShapeNet that provide \"canonicalized\" object instances that are consistently aligned for their 3D position and orientation (pose). We present Canonical Field Network (CaFi-Net), a self-supervised method to canonicalize the 3D pose of instances from an object category represented as neural fields, specifically neural radiance fields (NeRFs). CaFi-Net directly learns from continuous and noisy radiance fields using a Siamese network architecture that is designed to extract equivariant field features for category-level canonicalization. During inference, our method takes pre-trained neural radiance fields of novel object instances at arbitrary 3D pose and estimates a canonical field with consistent 3D pose across the entire category. Extensive experiments on a new dataset of 1300 NeRF models across 13 object categories show that our method matches or exceeds the performance of 3D point cloud-based methods." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 47, + 670, + 127, + 681 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 670, + 127, + 681 + ], + "spans": [ + { + "bbox": [ + 47, + 670, + 127, + 681 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "type": "text", + "content": "Neural fields [59]—coordinate-based neural networks that implicitly parameterize signals—have recently gained" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 304, + 370, + 545, + 454 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 370, + 545, + 454 + ], + "spans": [ + { + "bbox": [ + 304, + 370, + 545, + 454 + ], + "type": "text", + "content": "significant attention as representations of 3D shape [6,22, 28], view-dependent appearance [24, 42], and motion [25]. In particular, neural radiance fields (NeRFs) [24], have been successfully used in problems such as novel view synthesis [3,4,66], scene geometry extraction [55, 61], capturing dynamic scenes [19, 29, 30, 33, 52], 3D semantic segmentation [53, 68], and robotics [1, 13, 21]." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 304, + 458, + 546, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 458, + 546, + 601 + ], + "spans": [ + { + "bbox": [ + 304, + 458, + 546, + 601 + ], + "type": "text", + "content": "Despite the progress, it remains challenging to build neural fields that represent an entire category of objects. Previous methods sidestep the problem by overfitting on a single instance [24], or learning [22, 28, 63] on datasets like ShapeNet [5] that contain objects that are manually canonicalized – oriented consistently for 3D position and orientation (3D pose) across a category. This strong supervision makes it easier to learn over categories, but limits their application to data that contain these labels. Recent work has proposed methods for self-supervised learning of 3D pose canonicalization [40, 43, 47], however, these operate on 3D point clouds, meshes, or voxels – but not neural fields." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 304, + 605, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 547, + 714 + ], + "type": "text", + "content": "In this paper, we present Canonical Field Network (CaFi-Net), a self-supervised method for category-level canonicalization of the 3D position and orientation of objects represented as neural fields, specifically neural radiance fields. Canonicalizing neural fields is challenging because, unlike 3D point clouds or meshes, neural fields are continuous, noisy, and hard to manipulate since they are parameterized as the weights of a neural network [60]. To address these challenges, we first extend the notion of equiv" + } + ] + } + ], + "index": 24 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4500" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 228 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 228 + ], + "type": "text", + "content": "arience to continuous vector fields and show how networks for processing 3D point clouds [50] can be extended to operate directly on neural radiance fields. We design CaFi-Net as a Siamese network that contains layers to extract equivariant features directly on vector fields. These field features are used to learn a canonical frame that is consistent across instances in the category. During inference, our method takes as input neural radiance fields of object instances from a category at arbitrary pose and estimates a canonical field that is consistent across the category. To handle noise in radiance fields from NeRF, our method incorporates density-based feature weighting and foreground-background clustering." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 228, + 289, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 228, + 289, + 384 + ], + "spans": [ + { + "bbox": [ + 46, + 228, + 289, + 384 + ], + "type": "text", + "content": "Our approach learns canonicalization without any supervision labels on a new dataset of 1300 pre-trained NeRF models of 13 common ShapeNet categories in arbitrary 3D pose (see Figure 1). We introduce several self-supervision loss functions that encourage the estimation of a consistent canonical pose. In addition, we present extensive quantitative comparisons with baselines and other methods on standardized canonicalization metrics [37] over 13 object categories. In particular, we show that our approach matches or exceeds the performance of 3D point cloud-based methods. This enables the new capability of directly operating on neural fields rather than converting them to point clouds for canonicalization. To sum up, we contribute:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 59, + 392, + 286, + 525 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 59, + 392, + 286, + 438 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 392, + 286, + 438 + ], + "spans": [ + { + "bbox": [ + 59, + 392, + 286, + 438 + ], + "type": "text", + "content": "- Canonical Field Network (CaFi-Net), the first method for self-supervised canonicalization of the 3D position and orientation (pose) of objects represented as neural radiance fields." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 59, + 441, + 286, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 441, + 286, + 486 + ], + "spans": [ + { + "bbox": [ + 59, + 441, + 286, + 486 + ], + "type": "text", + "content": "- A Siamese neural network architecture with equivariant feature extraction layers that are designed to directly operate on continuous and noisy radiance fields from NeRF." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 59, + 490, + 286, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 490, + 286, + 525 + ], + "spans": [ + { + "bbox": [ + 59, + 490, + 286, + 525 + ], + "type": "text", + "content": "- A public dataset of 1300 NeRF models from 13 ShapeNet categories including posed images, and weights for evaluating canonicalization performance." + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 47, + 537, + 134, + 550 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 537, + 134, + 550 + ], + "spans": [ + { + "bbox": [ + 47, + 537, + 134, + 550 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 558, + 286, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 558, + 286, + 592 + ], + "spans": [ + { + "bbox": [ + 46, + 558, + 286, + 592 + ], + "type": "text", + "content": "We focus our review of related work on neural fields, supervised canonicalization, equivariant neural network architectures, and self-supervised canonicalization." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 594, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 286, + 713 + ], + "type": "text", + "content": "Neural Fields: Neural fields are emerging as useful representations for solving problems such as novel view synthesis [3, 4, 15, 24, 66], shape encoding and reconstruction [22, 28], dynamic reconstruction [19, 29, 30], appearance modeling [23, 36, 45], and human motion modeling [8, 18, 46, 51]. Generalization of neural fields to object categories remains difficult, but some methods have used pre-canonicalized datasets to circumvent this problem [63]. Neural fields are also used outside of computer vision, for instance in reconstructing proteins [69], physics [35], or au" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 72, + 545, + 107 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 107 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 107 + ], + "type": "text", + "content": "dio [11]. Please see [49, 59] for more details. In this paper, our focus is on canonicalizing for the 3D pose of neural fields, specifically, neural radiance fields (NeRF) [24]." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 108, + 545, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 108, + 545, + 239 + ], + "spans": [ + { + "bbox": [ + 304, + 108, + 545, + 239 + ], + "type": "text", + "content": "Supervised Canonicalization: Datasets such as ShapeNet [5] and ModelNet40 [58] have 3D shapes that are manually pre-canonicalized. This inductive bias aids category-level generalization in problems such as 3D reconstruction [12, 22, 28, 48]. We can also use these datasets to formulate canonicalization as a supervised learning problem [54] enabling applications such as 6 degree-of-freedom object pose estimation, multi-view reconstruction [17, 44], and reconstruction of articulating [20, 65] and non-rigid objects [64]. However, our goal is to canonicalize without using manual pose labels." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 240, + 545, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 240, + 545, + 346 + ], + "spans": [ + { + "bbox": [ + 304, + 240, + 545, + 346 + ], + "type": "text", + "content": "Equivariant Neural Networks: Pose-equivariant networks are equivariant to input pose [9, 50, 56, 57] by design. Some of these methods use Spherical Harmonic functions [10, 50, 56, 57], or vector neurons [9] to extract equivariant features. Equivariance is closely related to canonicalization since a canonical pose is also equivariant. Thus, previous methods have used pose-equivariance for canonicalization [37, 41, 43, 47]. However, these methods have thus far been limited to 3D point clouds, voxels, or meshes." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 347, + 545, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 347, + 545, + 418 + ], + "spans": [ + { + "bbox": [ + 304, + 347, + 545, + 418 + ], + "type": "text", + "content": "Self-Supervised Canonicalization: Recent research has shown that self/weak supervision is sufficient for learning pose canonicalization on point clouds [37, 38, 43, 47], 2D key-points [26], and images [27]. None of these previous self-supervised methods can operate directly on neural fields – to the best of our knowledge, ours is the first." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 430, + 384, + 443 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 430, + 384, + 443 + ], + "spans": [ + { + "bbox": [ + 306, + 430, + 384, + 443 + ], + "type": "text", + "content": "3. Background" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 450, + 544, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 450, + 544, + 472 + ], + "spans": [ + { + "bbox": [ + 304, + 450, + 544, + 472 + ], + "type": "text", + "content": "We first provide background on the essential components and key ideas behind our method." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 475, + 545, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 475, + 545, + 676 + ], + "spans": [ + { + "bbox": [ + 304, + 475, + 545, + 676 + ], + "type": "text", + "content": "Neural Radiance Fields (NeRF): Given posed RGB images, NeRF and its variants [23, 24] learn a neural network to synthesize novel views of complex scenes. At inference, for any 3D location and viewing direction, it estimates a direction-independent density value " + }, + { + "bbox": [ + 304, + 475, + 545, + 676 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 304, + 475, + 545, + 676 + ], + "type": "text", + "content": ", and a direction-dependent color " + }, + { + "bbox": [ + 304, + 475, + 545, + 676 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 304, + 475, + 545, + 676 + ], + "type": "text", + "content": ". Although designed for novel view synthesis, NeRF implicitly learns the 3D geometry and view-dependent appearance of scenes making it a useful representation of shape and appearance. While NeRFs excel at fitting a single object instance or scene, it struggles to generalize to object categories partly because objects can be in arbitrary 3D poses – thus, existing methods use precanonicalized datasets [63]. In this paper, we provide a way to canonicalize 3D pose without any supervision for objects represented as neural radiance fields. While our method is aimed at NeRFs, it can be extended to the 3D pose canonicalization of any neural field." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "content": "Equivalence and Canonicalization: A function " + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\Gamma" + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "content": " over " + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "content": " is said to be equivariant to a group operation " + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{O}" + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "content": " if its output changes by a fixed mapping " + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "content": " for any " + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{O}" + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "content": " operating on" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 314, + 757 + ], + "type": "text", + "content": "4501" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 69, + 289, + 194 + ], + "blocks": [ + { + "bbox": [ + 50, + 69, + 289, + 194 + ], + "lines": [ + { + "bbox": [ + 50, + 69, + 289, + 194 + ], + "spans": [ + { + "bbox": [ + 50, + 69, + 289, + 194 + ], + "type": "image", + "image_path": "4f7702db1fda005750345acb1f467aed69d3d49e34e8a2e86cc59ff2b522166f.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 203, + 289, + 293 + ], + "lines": [ + { + "bbox": [ + 46, + 203, + 289, + 293 + ], + "spans": [ + { + "bbox": [ + 46, + 203, + 289, + 293 + ], + "type": "text", + "content": "Figure 2. Rotation Equivariance in Vector Fields. Rotation equivariance in vector fields " + }, + { + "bbox": [ + 46, + 203, + 289, + 293 + ], + "type": "inline_equation", + "content": "\\mathcal{F}" + }, + { + "bbox": [ + 46, + 203, + 289, + 293 + ], + "type": "text", + "content": " (here, a simulated 2D magnetic field) require two operations: (1) update the position of the green arrow (the vector), and (2) rotating the green arrow at its new position. A function " + }, + { + "bbox": [ + 46, + 203, + 289, + 293 + ], + "type": "inline_equation", + "content": "\\Gamma" + }, + { + "bbox": [ + 46, + 203, + 289, + 293 + ], + "type": "text", + "content": " operating on " + }, + { + "bbox": [ + 46, + 203, + 289, + 293 + ], + "type": "inline_equation", + "content": "\\mathcal{F}" + }, + { + "bbox": [ + 46, + 203, + 289, + 293 + ], + "type": "text", + "content": " is rotationally equivariant over the field if and only if " + }, + { + "bbox": [ + 46, + 203, + 289, + 293 + ], + "type": "inline_equation", + "content": "\\Gamma[(\\mathcal{R} \\cdot \\mathcal{F})(\\mathbf{x})] = M(\\mathcal{R})\\Gamma[\\mathcal{F}(\\mathcal{R}^{-1}\\mathbf{x})]" + }, + { + "bbox": [ + 46, + 203, + 289, + 293 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 203, + 289, + 293 + ], + "type": "inline_equation", + "content": "\\mathcal{R}" + }, + { + "bbox": [ + 46, + 203, + 289, + 293 + ], + "type": "text", + "content": " is an " + }, + { + "bbox": [ + 46, + 203, + 289, + 293 + ], + "type": "inline_equation", + "content": "SO(2)" + }, + { + "bbox": [ + 46, + 203, + 289, + 293 + ], + "type": "text", + "content": " rotation, and " + }, + { + "bbox": [ + 46, + 203, + 289, + 293 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 46, + 203, + 289, + 293 + ], + "type": "text", + "content": " is a rotation-dependent mapping function." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 315, + 287, + 448 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 315, + 287, + 448 + ], + "spans": [ + { + "bbox": [ + 46, + 315, + 287, + 448 + ], + "type": "text", + "content": "the input " + }, + { + "bbox": [ + 46, + 315, + 287, + 448 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 46, + 315, + 287, + 448 + ], + "type": "text", + "content": ", i.e., " + }, + { + "bbox": [ + 46, + 315, + 287, + 448 + ], + "type": "inline_equation", + "content": "\\Gamma(\\mathcal{O}x) = M(\\mathcal{O})\\Gamma(x)" + }, + { + "bbox": [ + 46, + 315, + 287, + 448 + ], + "type": "text", + "content": ". In this paper, we are interested in equivariance to the " + }, + { + "bbox": [ + 46, + 315, + 287, + 448 + ], + "type": "inline_equation", + "content": "SE(3)" + }, + { + "bbox": [ + 46, + 315, + 287, + 448 + ], + "type": "text", + "content": " group denoting 3D position and orientation (pose) – in particular we focus on the rotation group " + }, + { + "bbox": [ + 46, + 315, + 287, + 448 + ], + "type": "inline_equation", + "content": "SO(3)" + }, + { + "bbox": [ + 46, + 315, + 287, + 448 + ], + "type": "text", + "content": " since 3D translation equivariance is readily achieved through mean centering [26]. Pose canonicalization is closely related to equivariance – our goal is to find an equivariant canonical 3D rigid transformation that maps the neural field of an object in an arbitrary pose to a canonical pose. We call the outputs of our method canonical fields – neural fields that are consistently oriented across a category of shapes." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 449, + 288, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 449, + 288, + 593 + ], + "spans": [ + { + "bbox": [ + 46, + 449, + 288, + 593 + ], + "type": "text", + "content": "Rotation Equivalence in Vector Fields: Unlike 3D point clouds, rotation equivariance in fields is more involved. Consider a 2D vector field " + }, + { + "bbox": [ + 46, + 449, + 288, + 593 + ], + "type": "inline_equation", + "content": "\\mathcal{F}" + }, + { + "bbox": [ + 46, + 449, + 288, + 593 + ], + "type": "text", + "content": " defined " + }, + { + "bbox": [ + 46, + 449, + 288, + 593 + ], + "type": "inline_equation", + "content": "\\forall \\mathbf{x} \\in \\mathbb{R}^2" + }, + { + "bbox": [ + 46, + 449, + 288, + 593 + ], + "type": "text", + "content": " (see Figure 2). Rotating this field requires two operations: (1) updating the positions of vectors in the field to new rotated positions, and (2) rotating the directions of the vectors. A function " + }, + { + "bbox": [ + 46, + 449, + 288, + 593 + ], + "type": "inline_equation", + "content": "\\Gamma" + }, + { + "bbox": [ + 46, + 449, + 288, + 593 + ], + "type": "text", + "content": " operating on " + }, + { + "bbox": [ + 46, + 449, + 288, + 593 + ], + "type": "inline_equation", + "content": "\\mathcal{F}" + }, + { + "bbox": [ + 46, + 449, + 288, + 593 + ], + "type": "text", + "content": " is rotationally equivariant over the field if and only if " + }, + { + "bbox": [ + 46, + 449, + 288, + 593 + ], + "type": "inline_equation", + "content": "\\Gamma[(\\mathcal{R} \\cdot \\mathcal{F})(\\mathbf{x})] = M(\\mathcal{R})\\Gamma[\\mathcal{F}(\\mathcal{R}^{-1}\\mathbf{x})]" + }, + { + "bbox": [ + 46, + 449, + 288, + 593 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 449, + 288, + 593 + ], + "type": "inline_equation", + "content": "\\mathcal{R}" + }, + { + "bbox": [ + 46, + 449, + 288, + 593 + ], + "type": "text", + "content": " is an " + }, + { + "bbox": [ + 46, + 449, + 288, + 593 + ], + "type": "inline_equation", + "content": "SO(2)" + }, + { + "bbox": [ + 46, + 449, + 288, + 593 + ], + "type": "text", + "content": " rotation, and " + }, + { + "bbox": [ + 46, + 449, + 288, + 593 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 46, + 449, + 288, + 593 + ], + "type": "text", + "content": " is a rotation-dependent mapping function. In this paper, we operate on NeRF's density field (a scalar field) but also extract equivariant features that can be vector-valued fields." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 605, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 287, + 715 + ], + "type": "text", + "content": "Tensor Field Networks (TFNs): Many methods have been proposed for using equivariant features for rotation canonicalization of 3D data, for instance, spherical CNNs [7, 43], vector neurons [9], or capsule networks [47, 67]. However, these and other methods [38] are limited to 3D point cloud or voxel representations. In this paper, we extend Tensor Field Networks [34, 50] pose canonicalization directly on samples from NeRF's density field. TFNs operate by computing Type- " + }, + { + "bbox": [ + 46, + 605, + 287, + 715 + ], + "type": "inline_equation", + "content": "\\ell" + }, + { + "bbox": [ + 46, + 605, + 287, + 715 + ], + "type": "text", + "content": " real-valued spherical harmonic functions" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 71, + 547, + 245 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 71, + 547, + 245 + ], + "spans": [ + { + "bbox": [ + 304, + 71, + 547, + 245 + ], + "type": "text", + "content": "that are rotation equivariant: " + }, + { + "bbox": [ + 304, + 71, + 547, + 245 + ], + "type": "inline_equation", + "content": "D^{\\ell}(R)Y^{\\ell}(X) = Y^{\\ell}(Rx)" + }, + { + "bbox": [ + 304, + 71, + 547, + 245 + ], + "type": "text", + "content": ". Where " + }, + { + "bbox": [ + 304, + 71, + 547, + 245 + ], + "type": "inline_equation", + "content": "D^{\\ell}:\\mathrm{SO}(3)\\to \\mathrm{SO}(2\\ell +1)" + }, + { + "bbox": [ + 304, + 71, + 547, + 245 + ], + "type": "text", + "content": " is the Wigner matrix of type " + }, + { + "bbox": [ + 304, + 71, + 547, + 245 + ], + "type": "inline_equation", + "content": "\\ell" + }, + { + "bbox": [ + 304, + 71, + 547, + 245 + ], + "type": "text", + "content": " (please see [34, 50] for details). For a 3D point set " + }, + { + "bbox": [ + 304, + 71, + 547, + 245 + ], + "type": "inline_equation", + "content": "P\\in \\mathbb{R}^{N\\times 3}" + }, + { + "bbox": [ + 304, + 71, + 547, + 245 + ], + "type": "text", + "content": ", a rotation-equivariant TFN convolution layer (EQConv) at point " + }, + { + "bbox": [ + 304, + 71, + 547, + 245 + ], + "type": "inline_equation", + "content": "p\\in P" + }, + { + "bbox": [ + 304, + 71, + 547, + 245 + ], + "type": "text", + "content": " is defined as: " + }, + { + "bbox": [ + 304, + 71, + 547, + 245 + ], + "type": "inline_equation", + "content": "\\mathrm{EQConv}^J (p) = Q^{(n,\\ell),J}\\left(\\sum_{y\\in^{2R}\\mathcal{N}}k_r^n (p - y)\\otimes s^\\ell [y]\\right)" + }, + { + "bbox": [ + 304, + 71, + 547, + 245 + ], + "type": "text", + "content": ". Where " + }, + { + "bbox": [ + 304, + 71, + 547, + 245 + ], + "type": "inline_equation", + "content": "^{2R}\\mathcal{N}" + }, + { + "bbox": [ + 304, + 71, + 547, + 245 + ], + "type": "text", + "content": " is a set containing neighbors of point " + }, + { + "bbox": [ + 304, + 71, + 547, + 245 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 304, + 71, + 547, + 245 + ], + "type": "text", + "content": " at twice the resolution and " + }, + { + "bbox": [ + 304, + 71, + 547, + 245 + ], + "type": "inline_equation", + "content": "Q^{(n,\\ell),J}(\\cdot)" + }, + { + "bbox": [ + 304, + 71, + 547, + 245 + ], + "type": "text", + "content": " is the Clebsch Gordon decomposition to combine type- " + }, + { + "bbox": [ + 304, + 71, + 547, + 245 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 304, + 71, + 547, + 245 + ], + "type": "text", + "content": " kernel " + }, + { + "bbox": [ + 304, + 71, + 547, + 245 + ], + "type": "inline_equation", + "content": "k_r^n" + }, + { + "bbox": [ + 304, + 71, + 547, + 245 + ], + "type": "text", + "content": " and type- " + }, + { + "bbox": [ + 304, + 71, + 547, + 245 + ], + "type": "inline_equation", + "content": "\\ell" + }, + { + "bbox": [ + 304, + 71, + 547, + 245 + ], + "type": "text", + "content": " equivariant signal " + }, + { + "bbox": [ + 304, + 71, + 547, + 245 + ], + "type": "inline_equation", + "content": "s^\\ell" + }, + { + "bbox": [ + 304, + 71, + 547, + 245 + ], + "type": "text", + "content": ". The EQConv " + }, + { + "bbox": [ + 304, + 71, + 547, + 245 + ], + "type": "inline_equation", + "content": "(\\cdot)" + }, + { + "bbox": [ + 304, + 71, + 547, + 245 + ], + "type": "text", + "content": " layer aggregates type- " + }, + { + "bbox": [ + 304, + 71, + 547, + 245 + ], + "type": "inline_equation", + "content": "\\ell" + }, + { + "bbox": [ + 304, + 71, + 547, + 245 + ], + "type": "text", + "content": " feature " + }, + { + "bbox": [ + 304, + 71, + 547, + 245 + ], + "type": "inline_equation", + "content": "s^\\ell" + }, + { + "bbox": [ + 304, + 71, + 547, + 245 + ], + "type": "text", + "content": " at multiple receptive fields along with learnable weights to perform equivariant convolution. Our method uses the EQConv layer as the fundamental building block to canonicalize continuous fields." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 253, + 362, + 266 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 253, + 362, + 266 + ], + "spans": [ + { + "bbox": [ + 306, + 253, + 362, + 266 + ], + "type": "text", + "content": "4. Method" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 274, + 545, + 369 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 274, + 545, + 369 + ], + "spans": [ + { + "bbox": [ + 304, + 274, + 545, + 369 + ], + "type": "text", + "content": "Our method assumes that we have pre-trained NeRF models of scenes containing single object instances at arbitrary poses from common shape categories. Our scenes contain primarily the object, but also some background - perfect foreground-background segmentation is not available. Our goal is to build a method that estimates a canonicalizing " + }, + { + "bbox": [ + 304, + 274, + 545, + 369 + ], + "type": "inline_equation", + "content": "SE(3)" + }, + { + "bbox": [ + 304, + 274, + 545, + 369 + ], + "type": "text", + "content": " transformation that aligns these object NeRFs to a consistent category-level canonical field." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 370, + 546, + 489 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 370, + 546, + 489 + ], + "spans": [ + { + "bbox": [ + 304, + 370, + 546, + 489 + ], + "type": "text", + "content": "Rather than operate on 3D point clouds, meshes, or voxels, our method (see Figure 3) operates directly on samples from the continuous neural radiance field. During training, we learn canonical fields from a pre-trained NeRF dataset in a fully self-supervised manner. We use a Siamese network architecture that extracts equivariant field features for canonicalization (see Figure 3). During inference, given a previously unseen NeRF model of a new object instance in an arbitrary pose, our method estimates the transformation that maps it to the canonical field." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 495, + 404, + 509 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 495, + 404, + 509 + ], + "spans": [ + { + "bbox": [ + 306, + 495, + 404, + 509 + ], + "type": "text", + "content": "4.1. NeRF Sampling" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 513, + 545, + 571 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 513, + 545, + 571 + ], + "spans": [ + { + "bbox": [ + 304, + 513, + 545, + 571 + ], + "type": "text", + "content": "For a pre-trained neural radiance field that maps 3D position " + }, + { + "bbox": [ + 304, + 513, + 545, + 571 + ], + "type": "inline_equation", + "content": "(x, y, z)" + }, + { + "bbox": [ + 304, + 513, + 545, + 571 + ], + "type": "text", + "content": " and direction " + }, + { + "bbox": [ + 304, + 513, + 545, + 571 + ], + "type": "inline_equation", + "content": "\\theta, \\phi" + }, + { + "bbox": [ + 304, + 513, + 545, + 571 + ], + "type": "text", + "content": " to color and density " + }, + { + "bbox": [ + 304, + 513, + 545, + 571 + ], + "type": "inline_equation", + "content": "(\\mathbf{c}, \\sigma)" + }, + { + "bbox": [ + 304, + 513, + 545, + 571 + ], + "type": "text", + "content": ", the input to our method consists of samples on the density field. We do not use color because it is direction-dependent in NeRF and also varies for different instances." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 573, + 547, + 681 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 573, + 547, + 681 + ], + "spans": [ + { + "bbox": [ + 304, + 573, + 547, + 681 + ], + "type": "text", + "content": "Preprocessing: During both training and inference, we uniformly sample the pre-trained NeRF density field to find the object center and its bounding box (see supplementary document for details). We then sample a uniform grid in the density field of the object's bounding box. We empirically choose uniform grid sampling over fully random sampling since we found no difference. The uniformly sampled density grid is then normalized to obtain density values " + }, + { + "bbox": [ + 304, + 573, + 547, + 681 + ], + "type": "inline_equation", + "content": "\\sigma_{D}" + }, + { + "bbox": [ + 304, + 573, + 547, + 681 + ], + "type": "text", + "content": " in the continuous range [0, 1], i.e.," + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 341, + 697, + 545, + 711 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 341, + 697, + 545, + 711 + ], + "spans": [ + { + "bbox": [ + 341, + 697, + 545, + 711 + ], + "type": "interline_equation", + "content": "\\boldsymbol {\\sigma} _ {D} := \\left\\{1 - e x p \\left(- d \\cdot \\Gamma_ {r} ^ {\\sigma} (x)\\right) \\mid x \\in \\boldsymbol {X} \\right\\}, \\tag {1}", + "image_path": "4be3a0cf02156eb92727f0708618905cc3d939953a238013f36d011807c92cf5.jpg" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "4502" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 55, + 70, + 542, + 300 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 542, + 300 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 542, + 300 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 542, + 300 + ], + "type": "image", + "image_path": "d39ec9a4d2ad86ff29ef52690ae451f656d06e9adfc3938575f17f23d2437a23.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 307, + 548, + 373 + ], + "lines": [ + { + "bbox": [ + 46, + 307, + 548, + 373 + ], + "spans": [ + { + "bbox": [ + 46, + 307, + 548, + 373 + ], + "type": "text", + "content": "Figure 3. CaFi-Net samples a density field from NeRF and uses its density, position and density gradients as input signals (A) to canonicalize the field. We predict rotation equivariant features and weigh them by density (B) to guide our learning from the occupied regions of the scene. We then compute an invariant embedding by taking a dot product between equivariant features. This invariant embedding is used to canonicalize the field that enables rendering all the objects in the canonical frame (E). Our method also applies an inter-instance consistency loss (D) that aligns different instances of the same category in the canonical frame. We do not assume precanonicalized fields, and canonicalize in a self-supervised manner." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 385, + 288, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 385, + 288, + 456 + ], + "spans": [ + { + "bbox": [ + 46, + 385, + 288, + 456 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 385, + 288, + 456 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 46, + 385, + 288, + 456 + ], + "type": "text", + "content": " is the regularly sampled grid within the object bounding box and " + }, + { + "bbox": [ + 46, + 385, + 288, + 456 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 46, + 385, + 288, + 456 + ], + "type": "text", + "content": " is the depth step size used in the coarse NeRF sampling. Translation and scale canonicalization is simply achieved by moving the bounding box's center [26] and scaling it, so our next concern is on rotation canonicalization." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 466, + 115, + 478 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 466, + 115, + 478 + ], + "spans": [ + { + "bbox": [ + 47, + 466, + 115, + 478 + ], + "type": "text", + "content": "4.2. CaFi-Net" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 485, + 287, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 485, + 287, + 521 + ], + "spans": [ + { + "bbox": [ + 46, + 485, + 287, + 521 + ], + "type": "text", + "content": "CaFi-Net (Canonical Field Network) is our method for 3D rotation canonicalization using NeRF's density field (red box in Figure 3). We now describe its components." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 521, + 287, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 521, + 287, + 652 + ], + "spans": [ + { + "bbox": [ + 46, + 521, + 287, + 652 + ], + "type": "text", + "content": "Signal Representation: At each of the sampled density field locations, we have several choices of signals to use as the input for CaFi-Net. For instance, we can (1) use the NeRF density value directly, (2) use the " + }, + { + "bbox": [ + 46, + 521, + 287, + 652 + ], + "type": "inline_equation", + "content": "xyz" + }, + { + "bbox": [ + 46, + 521, + 287, + 652 + ], + "type": "text", + "content": " location of the sampled field, or (3) use the gradient of the density field. Densities alone do not capture the position and directional components of the field, so we use a combination of densities and gradients of densities at query locations of the grid as our input signal for CaFi-Net. Gradients of density capture the object surface and help in canonicalization (see ablation study in Section 5)." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 654, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 654, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 654, + 288, + 714 + ], + "type": "text", + "content": "**Equivariant Convolution:** To canonicalize for rotation, we leverage equivariance properties of spherical harmonic functions to obtain rotation-equivariant learnable features on the field [34, 50]. Type-" + }, + { + "bbox": [ + 46, + 654, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\ell" + }, + { + "bbox": [ + 46, + 654, + 288, + 714 + ], + "type": "text", + "content": " spherical harmonic functions " + }, + { + "bbox": [ + 46, + 654, + 288, + 714 + ], + "type": "inline_equation", + "content": "(Y^{\\ell})" + }, + { + "bbox": [ + 46, + 654, + 288, + 714 + ], + "type": "text", + "content": " are degree-" + }, + { + "bbox": [ + 46, + 654, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\ell" + }, + { + "bbox": [ + 46, + 654, + 288, + 714 + ], + "type": "text", + "content": " polynomials of points on the sphere that" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 385, + 547, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 385, + 547, + 518 + ], + "spans": [ + { + "bbox": [ + 304, + 385, + 547, + 518 + ], + "type": "text", + "content": "are equivariant to " + }, + { + "bbox": [ + 304, + 385, + 547, + 518 + ], + "type": "inline_equation", + "content": "SO(2\\ell + 1)" + }, + { + "bbox": [ + 304, + 385, + 547, + 518 + ], + "type": "text", + "content": ". The densities sampled from NeRF reconstructions are often noisy and contain outliers. To promote our method to learn only from occupied regions, we weigh each equivariant feature type by its density. This weighing has no effect on the equivariance of features (see supplementary document for proof). We have the option to weigh equivariant features with the local average of the density to smoothen the signal or to weigh directly by the density. We found weighing by density worked better than taking the local average density (see Section 5). We learn equivariant features of type " + }, + { + "bbox": [ + 304, + 385, + 547, + 518 + ], + "type": "inline_equation", + "content": "\\ell" + }, + { + "bbox": [ + 304, + 385, + 547, + 518 + ], + "type": "text", + "content": " as:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 328, + 529, + 543, + 546 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 328, + 529, + 543, + 546 + ], + "spans": [ + { + "bbox": [ + 328, + 529, + 543, + 546 + ], + "type": "interline_equation", + "content": "{ } ^ { R } s ^ { \\ell } : = f _ { w } ( ^ { R } x ) \\left( W ^ { \\ell } ( \\operatorname { E Q C o n v } ^ { \\ell } ( \\nabla ( \\boldsymbol { \\sigma } _ { D } ( ^ { R } x ) ) ) ) + \\delta _ { 0 \\ell } b \\right) , \\tag {2}", + "image_path": "8702f8e311764323079516496516242abcff8b0e5fb5e02249c9087864a60d54.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 328, + 548, + 543, + 574 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 328, + 548, + 543, + 574 + ], + "spans": [ + { + "bbox": [ + 328, + 548, + 543, + 574 + ], + "type": "interline_equation", + "content": "f _ {w} ^ {(R x)} := \\left\\{ \\begin{array}{l} \\operatorname {m e a n} _ {r} \\left(\\sigma_ {D} ^ {(R x)}\\right), \\quad o r \\\\ \\sigma_ {D} ^ {(R x)}, \\end{array} \\right. \\tag {3}", + "image_path": "39169401b352cfaba1f3a63b24ab29cbc5a805e12d93d7767970183ab739e5a3.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 575, + 545, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 575, + 545, + 696 + ], + "spans": [ + { + "bbox": [ + 304, + 575, + 545, + 696 + ], + "type": "text", + "content": "where, " + }, + { + "bbox": [ + 304, + 575, + 545, + 696 + ], + "type": "inline_equation", + "content": "W^{\\ell}" + }, + { + "bbox": [ + 304, + 575, + 545, + 696 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 575, + 545, + 696 + ], + "type": "inline_equation", + "content": "b" + }, + { + "bbox": [ + 304, + 575, + 545, + 696 + ], + "type": "text", + "content": " are the weights and biases, " + }, + { + "bbox": [ + 304, + 575, + 545, + 696 + ], + "type": "inline_equation", + "content": "{}^{R}x" + }, + { + "bbox": [ + 304, + 575, + 545, + 696 + ], + "type": "text", + "content": " is a 3D point queried at resolution " + }, + { + "bbox": [ + 304, + 575, + 545, + 696 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 304, + 575, + 545, + 696 + ], + "type": "text", + "content": " and mean " + }, + { + "bbox": [ + 304, + 575, + 545, + 696 + ], + "type": "inline_equation", + "content": "r(\\sigma_D(Rx))" + }, + { + "bbox": [ + 304, + 575, + 545, + 696 + ], + "type": "text", + "content": " is the average density at location " + }, + { + "bbox": [ + 304, + 575, + 545, + 696 + ], + "type": "inline_equation", + "content": "R_x" + }, + { + "bbox": [ + 304, + 575, + 545, + 696 + ], + "type": "text", + "content": " from points sampled at radius " + }, + { + "bbox": [ + 304, + 575, + 545, + 696 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 304, + 575, + 545, + 696 + ], + "type": "text", + "content": ". We hierarchically aggregate features at resolutions " + }, + { + "bbox": [ + 304, + 575, + 545, + 696 + ], + "type": "inline_equation", + "content": "\\frac{1}{2}, \\frac{1}{4}" + }, + { + "bbox": [ + 304, + 575, + 545, + 696 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 575, + 545, + 696 + ], + "type": "inline_equation", + "content": "\\frac{1}{8}" + }, + { + "bbox": [ + 304, + 575, + 545, + 696 + ], + "type": "text", + "content": " to obtain global equivariant features. The EQConv convolution layer is the same as defined in Section 3. We employ non-linearities from [34] that preserve equivariance and capture better equivariant features. We also compute the max-pool of point-wise features of the last layer to obtain global type- " + }, + { + "bbox": [ + 304, + 575, + 545, + 696 + ], + "type": "inline_equation", + "content": "\\ell" + }, + { + "bbox": [ + 304, + 575, + 545, + 696 + ], + "type": "text", + "content": " equivariant features " + }, + { + "bbox": [ + 304, + 575, + 545, + 696 + ], + "type": "inline_equation", + "content": "F^{\\ell}" + }, + { + "bbox": [ + 304, + 575, + 545, + 696 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 701, + 546, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 701, + 546, + 714 + ], + "spans": [ + { + "bbox": [ + 306, + 701, + 546, + 714 + ], + "type": "text", + "content": "Canonicalization: After obtaining the global type- " + }, + { + "bbox": [ + 306, + 701, + 546, + 714 + ], + "type": "inline_equation", + "content": "\\ell" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "4503" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 228 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 228 + ], + "type": "text", + "content": "equivariant feature " + }, + { + "bbox": [ + 46, + 72, + 289, + 228 + ], + "type": "inline_equation", + "content": "F^{\\ell}" + }, + { + "bbox": [ + 46, + 72, + 289, + 228 + ], + "type": "text", + "content": ", we compute its dot product with point-wise spherical harmonics " + }, + { + "bbox": [ + 46, + 72, + 289, + 228 + ], + "type": "inline_equation", + "content": "Z^{\\ell}" + }, + { + "bbox": [ + 46, + 72, + 289, + 228 + ], + "type": "text", + "content": " scaled by their norms " + }, + { + "bbox": [ + 46, + 72, + 289, + 228 + ], + "type": "inline_equation", + "content": "Z^{\\ell}(\\mathbf{X}) = ||\\mathbf{X}||_{2}Y^{\\ell}(\\mathbf{X} / ||\\mathbf{X}||_{2})" + }, + { + "bbox": [ + 46, + 72, + 289, + 228 + ], + "type": "text", + "content": " to obtain a pose-invariant embedding " + }, + { + "bbox": [ + 46, + 72, + 289, + 228 + ], + "type": "inline_equation", + "content": "H^{\\ell}" + }, + { + "bbox": [ + 46, + 72, + 289, + 228 + ], + "type": "text", + "content": " [37]. As the two vectors " + }, + { + "bbox": [ + 46, + 72, + 289, + 228 + ], + "type": "inline_equation", + "content": "F^{\\ell}" + }, + { + "bbox": [ + 46, + 72, + 289, + 228 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 72, + 289, + 228 + ], + "type": "inline_equation", + "content": "Z^{\\ell}" + }, + { + "bbox": [ + 46, + 72, + 289, + 228 + ], + "type": "text", + "content": " are both equivariant to input rotation, their dot product is invariant. We use a linear layer on top of the pose-invariant embedding to estimate grid coordinates in the canonical frame " + }, + { + "bbox": [ + 46, + 72, + 289, + 228 + ], + "type": "inline_equation", + "content": "(P\\in \\mathbb{R}^{H\\times W\\times D\\times 3})" + }, + { + "bbox": [ + 46, + 72, + 289, + 228 + ], + "type": "text", + "content": " for each point in " + }, + { + "bbox": [ + 46, + 72, + 289, + 228 + ], + "type": "inline_equation", + "content": "\\mathbf{X}" + }, + { + "bbox": [ + 46, + 72, + 289, + 228 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 72, + 289, + 228 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 46, + 72, + 289, + 228 + ], + "type": "text", + "content": " equivariant transformations " + }, + { + "bbox": [ + 46, + 72, + 289, + 228 + ], + "type": "inline_equation", + "content": "E\\in \\mathbb{R}^{M\\times 3\\times 3}" + }, + { + "bbox": [ + 46, + 72, + 289, + 228 + ], + "type": "text", + "content": " [37] that orient the canonical coordinates to the input coordinates. We then choose the best canonicalizing transform " + }, + { + "bbox": [ + 46, + 72, + 289, + 228 + ], + "type": "inline_equation", + "content": "E_{b}" + }, + { + "bbox": [ + 46, + 72, + 289, + 228 + ], + "type": "text", + "content": " and penalize it for canonicalization. The invariant embedding " + }, + { + "bbox": [ + 46, + 72, + 289, + 228 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 46, + 72, + 289, + 228 + ], + "type": "text", + "content": ", grid coordinates " + }, + { + "bbox": [ + 46, + 72, + 289, + 228 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 46, + 72, + 289, + 228 + ], + "type": "text", + "content": " and equivariant rotation " + }, + { + "bbox": [ + 46, + 72, + 289, + 228 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 46, + 72, + 289, + 228 + ], + "type": "text", + "content": " are given as," + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 94, + 235, + 287, + 250 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 94, + 235, + 287, + 250 + ], + "spans": [ + { + "bbox": [ + 94, + 235, + 287, + 250 + ], + "type": "interline_equation", + "content": "H ^ {\\ell} \\left(\\boldsymbol {\\sigma} _ {D}, \\boldsymbol {X}\\right) := \\langle F ^ {\\ell} \\left(\\boldsymbol {\\sigma} _ {D}\\right), Z ^ {\\ell} (\\boldsymbol {X}) \\rangle , \\tag {4}", + "image_path": "207bc95fb74733e25781291d3be9d0936c61f9e4168d0993ab21e0119c91bb3a.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 94, + 251, + 287, + 265 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 94, + 251, + 287, + 265 + ], + "spans": [ + { + "bbox": [ + 94, + 251, + 287, + 265 + ], + "type": "interline_equation", + "content": "P := \\operatorname {M L P} _ {Y} \\left(H \\left(\\boldsymbol {\\sigma} _ {D}, \\boldsymbol {X}\\right)\\right), \\tag {5}", + "image_path": "a6c4b9c4c6e1e942a8c09c584dd510830159dad7952bc611c6fea54833cf772a.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 94, + 266, + 287, + 280 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 94, + 266, + 287, + 280 + ], + "spans": [ + { + "bbox": [ + 94, + 266, + 287, + 280 + ], + "type": "interline_equation", + "content": "E := \\operatorname {M L P} _ {E} \\left(F \\left(\\boldsymbol {\\sigma} _ {D}, \\boldsymbol {X}\\right)\\right). \\tag {6}", + "image_path": "3ca4017492ee5291231f17e28e2bf1c947dc002202f50942763fcb973436efa7.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 288, + 287, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 288, + 287, + 312 + ], + "spans": [ + { + "bbox": [ + 46, + 288, + 287, + 312 + ], + "type": "text", + "content": "Here, we have dropped the type-" + }, + { + "bbox": [ + 46, + 288, + 287, + 312 + ], + "type": "inline_equation", + "content": "\\ell" + }, + { + "bbox": [ + 46, + 288, + 287, + 312 + ], + "type": "text", + "content": " notation for " + }, + { + "bbox": [ + 46, + 288, + 287, + 312 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 46, + 288, + 287, + 312 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 288, + 287, + 312 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 46, + 288, + 287, + 312 + ], + "type": "text", + "content": " as we concatenate all equivariant feature types." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 312, + 287, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 312, + 287, + 384 + ], + "spans": [ + { + "bbox": [ + 46, + 312, + 287, + 384 + ], + "type": "text", + "content": "Siamese Network Architecture: Experimentally, canonicalizing noisy density fields is difficult without guidance on shape similarity within a category. We therefore use a Siamese training strategy [26] where two different object instances are forced to be consistently canonicalized (see the lower branch in Figure 3)." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 384, + 288, + 480 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 384, + 288, + 480 + ], + "spans": [ + { + "bbox": [ + 46, + 384, + 288, + 480 + ], + "type": "text", + "content": "Clustering: CaFi-Net predicts a canonical coordinate for every grid point, but unlike point clouds, fields have many unoccupied regions that do not provide any information. We penalize the network on foreground regions only by performing K-means clustering on the densities within the object bounding box with " + }, + { + "bbox": [ + 46, + 384, + 288, + 480 + ], + "type": "inline_equation", + "content": "K = 2" + }, + { + "bbox": [ + 46, + 384, + 288, + 480 + ], + "type": "text", + "content": " and choose the cluster " + }, + { + "bbox": [ + 46, + 384, + 288, + 480 + ], + "type": "inline_equation", + "content": "C_f" + }, + { + "bbox": [ + 46, + 384, + 288, + 480 + ], + "type": "text", + "content": " with a higher mean. Note that we still operate on continuous fields, but only cluster samples to guide our training." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 487, + 111, + 499 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 487, + 111, + 499 + ], + "spans": [ + { + "bbox": [ + 47, + 487, + 111, + 499 + ], + "type": "text", + "content": "4.3. Training" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 505, + 287, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 505, + 287, + 540 + ], + "spans": [ + { + "bbox": [ + 46, + 505, + 287, + 540 + ], + "type": "text", + "content": "CaFi-Net is trained on a large dataset of pre-trained NeRF models over 13 object categories (see Section 5). We use the following loss functions to train our model." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 541, + 287, + 648 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 541, + 287, + 648 + ], + "spans": [ + { + "bbox": [ + 46, + 541, + 287, + 648 + ], + "type": "text", + "content": "Canonicalization Loss: To self-supervise our learning, we transform the predicted canonical grid coordinates to the input grid using the canonical rotation " + }, + { + "bbox": [ + 46, + 541, + 287, + 648 + ], + "type": "inline_equation", + "content": "E_{b}" + }, + { + "bbox": [ + 46, + 541, + 287, + 648 + ], + "type": "text", + "content": " and compute a point-wise L2 distance between them. This loss forces the predicted invariant grid coordinates " + }, + { + "bbox": [ + 46, + 541, + 287, + 648 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 46, + 541, + 287, + 648 + ], + "type": "text", + "content": " to reconstruct the shape and forces the predicted canonicalizing transform to be equivariant to the input transformation. The final loss is computed over the canonicalizing transform that minimizes this loss:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 82, + 654, + 287, + 675 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 82, + 654, + 287, + 675 + ], + "spans": [ + { + "bbox": [ + 82, + 654, + 287, + 675 + ], + "type": "interline_equation", + "content": "E _ {b} := \\min _ {j} \\left(\\operatorname {m e a n} _ {i \\in C _ {f}} \\left\\| X _ {i} - E _ {j} P _ {i} \\right\\| _ {2} ^ {2}\\right), \\tag {7}", + "image_path": "15ad8ec8c4e08e4c96d50ba0346b164dee8d21dfae8d2f8669eb58f4c801b52c.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 83, + 678, + 287, + 694 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 83, + 678, + 287, + 694 + ], + "spans": [ + { + "bbox": [ + 83, + 678, + 287, + 694 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {c a n o n}} := \\operatorname {m e a n} _ {i \\in C _ {f}} \\left| \\left| \\boldsymbol {X} _ {i} - E _ {b} P _ {i} \\right| \\right| _ {2} ^ {2}, \\tag {8}", + "image_path": "4bdbe8620c7fe4e36825778d110be2c4e299e27e3c8c2466082ea7be845ac7ed.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 47, + 701, + 273, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 701, + 273, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 701, + 273, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 701, + 273, + 714 + ], + "type": "inline_equation", + "content": "C_f" + }, + { + "bbox": [ + 47, + 701, + 273, + 714 + ], + "type": "text", + "content": " is a set of points belonging to the foreground." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 72, + 545, + 109 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 545, + 109 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 545, + 109 + ], + "type": "text", + "content": "**Orthonormality Loss:** The predicted equivariant transformations " + }, + { + "bbox": [ + 305, + 72, + 545, + 109 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 305, + 72, + 545, + 109 + ], + "type": "text", + "content": " should have orthonormal vectors and we use the following to force orthonormality using:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 351, + 118, + 545, + 131 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 351, + 118, + 545, + 131 + ], + "spans": [ + { + "bbox": [ + 351, + 118, + 545, + 131 + ], + "type": "interline_equation", + "content": "U _ {j}, D _ {j}, V _ {j} := \\operatorname {S V D} \\left(E _ {j}\\right), \\tag {9}", + "image_path": "7db796e551de4947e255b378fab6a0a839efd7bddfea636625e77c4a998aae48.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 351, + 133, + 545, + 162 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 351, + 133, + 545, + 162 + ], + "spans": [ + { + "bbox": [ + 351, + 133, + 545, + 162 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {o r t h o}} := \\frac {1}{M} \\sum_ {j} \\left| \\left| E _ {j} - U _ {j} V _ {j} ^ {T} \\right| \\right| _ {2}, \\tag {10}", + "image_path": "74a6e98747095c72dada9d54b64100b5f8d54d24d1f1727ee7ab176b425beb5e.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 305, + 171, + 544, + 194 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 171, + 544, + 194 + ], + "spans": [ + { + "bbox": [ + 305, + 171, + 544, + 194 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 171, + 544, + 194 + ], + "type": "inline_equation", + "content": "\\operatorname{SVD}(E_j)" + }, + { + "bbox": [ + 305, + 171, + 544, + 194 + ], + "type": "text", + "content": " computes the singular value decomposition of the " + }, + { + "bbox": [ + 305, + 171, + 544, + 194 + ], + "type": "inline_equation", + "content": "j^{th}" + }, + { + "bbox": [ + 305, + 171, + 544, + 194 + ], + "type": "text", + "content": " equivariant transformation." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 305, + 195, + 545, + 280 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 195, + 545, + 280 + ], + "spans": [ + { + "bbox": [ + 305, + 195, + 545, + 280 + ], + "type": "text", + "content": "Siamese Shape Loss: We use a Siamese loss that penalizes for consistency between different shapes belonging to the same category. Given two un-canonicalized fields in different frames of reference " + }, + { + "bbox": [ + 305, + 195, + 545, + 280 + ], + "type": "inline_equation", + "content": "\\sigma_{D}^{1}" + }, + { + "bbox": [ + 305, + 195, + 545, + 280 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 195, + 545, + 280 + ], + "type": "inline_equation", + "content": "\\sigma_{D}^{2}" + }, + { + "bbox": [ + 305, + 195, + 545, + 280 + ], + "type": "text", + "content": ", we canonicalize both the fields to the canonical frame and compute chamfer distance between their predicted canonical coordinates for points in the foreground cluster " + }, + { + "bbox": [ + 305, + 195, + 545, + 280 + ], + "type": "inline_equation", + "content": "C_f" + }, + { + "bbox": [ + 305, + 195, + 545, + 280 + ], + "type": "text", + "content": ", i.e.," + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 343, + 288, + 545, + 303 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 343, + 288, + 545, + 303 + ], + "spans": [ + { + "bbox": [ + 343, + 288, + 545, + 303 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {s i a m e s e}} := \\mathrm {C D} \\left(P \\left(\\boldsymbol {\\sigma} _ {D} ^ {1}\\right), P \\left(\\boldsymbol {\\sigma} _ {D} ^ {2}\\right)\\right). \\tag {11}", + "image_path": "fb2f10594ce594081b7650387982a12f748b08f7221ad37f04154696b61dc6b5.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 305, + 312, + 545, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 312, + 545, + 346 + ], + "spans": [ + { + "bbox": [ + 305, + 312, + 545, + 346 + ], + "type": "text", + "content": "This loss regularizes the training by ensuring that instances of the same category should be aligned in the canonical frame." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 305, + 347, + 546, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 347, + 546, + 574 + ], + "spans": [ + { + "bbox": [ + 305, + 347, + 546, + 574 + ], + "type": "text", + "content": "Architecture and Hyper-parameters: CaFi-Net predicts a 128 dimensional invariant embedding for each point in space and performs convolution at resolutions " + }, + { + "bbox": [ + 305, + 347, + 546, + 574 + ], + "type": "inline_equation", + "content": "1/2" + }, + { + "bbox": [ + 305, + 347, + 546, + 574 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 305, + 347, + 546, + 574 + ], + "type": "inline_equation", + "content": "1/4" + }, + { + "bbox": [ + 305, + 347, + 546, + 574 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 305, + 347, + 546, + 574 + ], + "type": "inline_equation", + "content": "1/8" + }, + { + "bbox": [ + 305, + 347, + 546, + 574 + ], + "type": "text", + "content": ". Each equivariant convolution layer aggregates features from 512 neighboring points at twice the resolution. We then use equivariant non-linearities introduced in [34] by applying Batch Normalization [14] and ReLU activation [2] in the inverse spherical harmonic transform domain. We use a three-layer MLP with intermediate Batch Normalization and ReLU layers followed by a final linear layer to predict canonical coordinates. To train our network, we weigh the canonicalization loss " + }, + { + "bbox": [ + 305, + 347, + 546, + 574 + ], + "type": "inline_equation", + "content": "(\\mathcal{L}_{canon})" + }, + { + "bbox": [ + 305, + 347, + 546, + 574 + ], + "type": "text", + "content": " the highest with weight 2.0 and 1.0 for all the other loss functions. All our models are written in PyTorch [31] and are trained for 300 epochs with a batch size of 2 on an Nvidia 1080-Ti GPU. We use the Adam Optimizer [16] with an initial learning rate of " + }, + { + "bbox": [ + 305, + 347, + 546, + 574 + ], + "type": "inline_equation", + "content": "6e - 4" + }, + { + "bbox": [ + 305, + 347, + 546, + 574 + ], + "type": "text", + "content": " and L2 weight regularization of " + }, + { + "bbox": [ + 305, + 347, + 546, + 574 + ], + "type": "inline_equation", + "content": "1e - 5" + }, + { + "bbox": [ + 305, + 347, + 546, + 574 + ], + "type": "text", + "content": " for all our experiments. Please see the supplementary document for more details." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 306, + 586, + 387, + 600 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 586, + 387, + 600 + ], + "spans": [ + { + "bbox": [ + 306, + 586, + 387, + 600 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 305, + 605, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 605, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 605, + 545, + 713 + ], + "type": "text", + "content": "We now provide details on our extensive experiments to evaluate the quality of our pose canonicalization. To our knowledge, we are the first method for pose canonicalization of neural radiance fields which makes direct comparisons with other work challenging. Nonetheless, we have designed experiments to compare our method with three 3D point cloud-based canonicalization methods. Overall, we present three classes of experiments: (1) evaluation of the performance of our method on 13 different categories," + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "4504" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 52, + 138, + 542, + 324 + ], + "blocks": [ + { + "bbox": [ + 46, + 70, + 547, + 137 + ], + "lines": [ + { + "bbox": [ + 46, + 70, + 547, + 137 + ], + "spans": [ + { + "bbox": [ + 46, + 70, + 547, + 137 + ], + "type": "text", + "content": "Table 1. This table compares the canonicalization performance of our method (F - operating on fields) with other 3D point cloud-based methods (P) on three standard metrics (IC, CC and GEC) on our dataset of 13 categories. We compare with PCA, Canonical Capsules (CaCa) [47] and ConDor [37]. All metrics are multiplied by 100 for ease of reading. The top two performing methods are highlighted in magenta (best) and blue (second best). We are better than SOTA [37] on the Ground Truth Equivariance Consistency (GEC) and Category-Level Consistency (CC) with lower mean and median canonicalization error. However, we perform on par with ConDor [37] on the Instance-Level Consistency." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 52, + 138, + 542, + 324 + ], + "lines": [ + { + "bbox": [ + 52, + 138, + 542, + 324 + ], + "spans": [ + { + "bbox": [ + 52, + 138, + 542, + 324 + ], + "type": "table", + "html": "
benchcabinetcarcellph.chaircouchfirearmlampmonitorplanespeakertablewater.avg.med.
Instance-Level Consistency (IC)↓
PCA (P)14.655.946.130.676.137.6315.0712.846.787.905.4010.3110.728.477.63
CaCa (P) [47]2.892.092.012.811.081.900.232.972.351.512.212.422.602.082.21
ConDor (P) [37]0.561.270.360.530.960.320.603.780.440.641.971.322.151.150.64
Ours (F)2.621.150.391.280.811.430.743.870.410.360.601.092.741.341.09
Category-Level Consistency (CC)↓
PCA (P)14.559.687.484.269.9013.4515.1311.8713.098.155.5110.4311.0710.3610.43
CaCa (P) [47]2.790.381.050.831.821.591.314.340.391.830.651.912.061.611.59
ConDor (P) [37]1.981.680.471.041.280.880.964.441.581.262.102.102.561.721.58
Ours (F)2.771.360.481.260.751.590.923.920.580.630.771.482.351.451.26
Ground Truth Equivalence Consistency (GEC)↓
PCA (P)14.405.866.321.646.328.1215.0112.496.828.165.3210.1711.278.618.12
CaCa (P) [47]3.652.282.402.741.902.221.284.692.482.092.223.012.792.592.40
ConDor (P) [37]2.121.790.501.171.340.931.044.581.591.312.222.242.671.811.59
Ours (F)3.261.510.541.430.941.901.014.370.630.660.811.633.021.671.43
", + "image_path": "1cfbfa9cb5d35d526940eea9e9ad2385c7c8fd304747111caaf33d4ee80b9fda.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 334, + 288, + 358 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 334, + 288, + 358 + ], + "spans": [ + { + "bbox": [ + 47, + 334, + 288, + 358 + ], + "type": "text", + "content": "(2) comparison with other work, and (3) ablations to justify method design choices." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 366, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 366, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 366, + 289, + 715 + ], + "type": "text", + "content": "Dataset: All our experiments use a new large synthetic NeRF dataset of shapes from 13 categories (that overlap with [12, 37, 47]) from the ShapeNet [5] dataset. For each category, we pick 100 instances from ShapeNet, rotate the shape randomly, and render 54 omnidirectional views sampled in a cube around the shape using Blender. Different from other datasets, we also simulate cluttered backgrounds. These posed views are used to train 100 NeRF models per category using the public PyTorch implementation of NeRF [62] for a total of 1300 NeRF models. To our knowledge, this is one of the largest (compared to [11, 53]) " + }, + { + "bbox": [ + 46, + 366, + 289, + 715 + ], + "type": "inline_equation", + "content": "360^{\\circ}" + }, + { + "bbox": [ + 46, + 366, + 289, + 715 + ], + "type": "text", + "content": " synthetic NeRF datasets - we will publicly release the simulator, raw posed images, and trained NeRF weights. Each NeRF model is trained for 400 epochs with 1024 randomly selected rays per iteration, a coarse sampling resolution of 64 points, and a fine sampling resolution of 128 points along each ray. Rather than train NeRF to maximize PSNR on a novel view, we fix the number of epochs for all models to enable a fair comparison between instances. As a consequence, some NeRF models may have a lower PSNR, but our goal is not to increase NeRF quality but rather improve canonicalization quality - hence we fix the number of epochs. For each category, we set aside " + }, + { + "bbox": [ + 46, + 366, + 289, + 715 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 46, + 366, + 289, + 715 + ], + "type": "text", + "content": " of the models for testing. Metrics: Since there are no known metrics for evaluating canonical fields, we resort to 3 metrics used for 3D point cloud canonicalization [37]: (1) Instance-Level Consistency (IC), a metric that measures how consistently we can canonicalize the same instance in different poses, (2) Category-" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 304, + 334, + 547, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 334, + 547, + 491 + ], + "spans": [ + { + "bbox": [ + 304, + 334, + 547, + 491 + ], + "type": "text", + "content": "Level Consistency (CC) to measure how consistently we canonicalize across different instances in the same category, and (3) Ground Truth Equivariance Consistency (GEC), a variant of the Ground Truth Consistency (GC) metric proposed in [37] to measure canonicalization performance compared to manual labels. We use the GEC because we observed that the original GC metric is prone to degeneracy when the canonicalizing transform is identity. We fix this issue by taking three point clouds " + }, + { + "bbox": [ + 304, + 334, + 547, + 491 + ], + "type": "inline_equation", + "content": "P_{i}, P_{j}" + }, + { + "bbox": [ + 304, + 334, + 547, + 491 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 334, + 547, + 491 + ], + "type": "inline_equation", + "content": "P_{k}" + }, + { + "bbox": [ + 304, + 334, + 547, + 491 + ], + "type": "text", + "content": " from a canonicalized dataset " + }, + { + "bbox": [ + 304, + 334, + 547, + 491 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 304, + 334, + 547, + 491 + ], + "type": "text", + "content": " and rotating " + }, + { + "bbox": [ + 304, + 334, + 547, + 491 + ], + "type": "inline_equation", + "content": "P_{i}" + }, + { + "bbox": [ + 304, + 334, + 547, + 491 + ], + "type": "text", + "content": " by " + }, + { + "bbox": [ + 304, + 334, + 547, + 491 + ], + "type": "inline_equation", + "content": "R_{1}, P_{j}" + }, + { + "bbox": [ + 304, + 334, + 547, + 491 + ], + "type": "text", + "content": " by " + }, + { + "bbox": [ + 304, + 334, + 547, + 491 + ], + "type": "inline_equation", + "content": "R_{2}" + }, + { + "bbox": [ + 304, + 334, + 547, + 491 + ], + "type": "text", + "content": " where both " + }, + { + "bbox": [ + 304, + 334, + 547, + 491 + ], + "type": "inline_equation", + "content": "R_{1}" + }, + { + "bbox": [ + 304, + 334, + 547, + 491 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 334, + 547, + 491 + ], + "type": "inline_equation", + "content": "R_{2}" + }, + { + "bbox": [ + 304, + 334, + 547, + 491 + ], + "type": "text", + "content": " are random rotations. Let " + }, + { + "bbox": [ + 304, + 334, + 547, + 491 + ], + "type": "inline_equation", + "content": "\\mathcal{C}(P)" + }, + { + "bbox": [ + 304, + 334, + 547, + 491 + ], + "type": "text", + "content": " predict a canonicalizing rotation for a point cloud " + }, + { + "bbox": [ + 304, + 334, + 547, + 491 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 304, + 334, + 547, + 491 + ], + "type": "text", + "content": ", we then compute the Ground Truth Equivariance consistency" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 497, + 550, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 497, + 550, + 605 + ], + "spans": [ + { + "bbox": [ + 304, + 497, + 550, + 605 + ], + "type": "text", + "content": "as: " + }, + { + "bbox": [ + 304, + 497, + 550, + 605 + ], + "type": "inline_equation", + "content": "\\mathrm{GEC} \\coloneqq \\frac{1}{|\\mathcal{P}|^3} \\sum_{P_i, P_j, P_k \\in \\mathcal{P}} \\mathrm{CD}(\\mathcal{C}(R_1P_i)R_1P_k, \\mathcal{C}(R_2P_j)R_2P_k)" + }, + { + "bbox": [ + 304, + 497, + 550, + 605 + ], + "type": "text", + "content": ", (12) where CD refers to the Chamfer distance. Here, we apply canonicalizing transforms of shape " + }, + { + "bbox": [ + 304, + 497, + 550, + 605 + ], + "type": "inline_equation", + "content": "P_i" + }, + { + "bbox": [ + 304, + 497, + 550, + 605 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 497, + 550, + 605 + ], + "type": "inline_equation", + "content": "P_j" + }, + { + "bbox": [ + 304, + 497, + 550, + 605 + ], + "type": "text", + "content": " to the shape " + }, + { + "bbox": [ + 304, + 497, + 550, + 605 + ], + "type": "inline_equation", + "content": "P_k" + }, + { + "bbox": [ + 304, + 497, + 550, + 605 + ], + "type": "text", + "content": ". This modified metric will not be degenerate for identity canonicalizing transforms and is more rigorous in evaluating ground truth consistency. To compute our metrics, we used point clouds sampled from the Shapenet meshes [5] using poisson disk sampling." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 305, + 610, + 457, + 624 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 610, + 457, + 624 + ], + "spans": [ + { + "bbox": [ + 305, + 610, + 457, + 624 + ], + "type": "text", + "content": "5.1. Evaluation & Comparisons" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 630, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 630, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 630, + 547, + 715 + ], + "type": "text", + "content": "We first analyze the performance of our method on the three metrics across 13 different categories. In Table 1, we highlight our method that operates on a field as Ours (F). Our method performs well across all categories with low variance between different categories. We observe similar trends across metrics for a set of categories (e.g., firearms,lamps) indicating that category shape plays a" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "4505" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 51, + 69, + 210, + 278 + ], + "blocks": [ + { + "bbox": [ + 51, + 69, + 210, + 278 + ], + "lines": [ + { + "bbox": [ + 51, + 69, + 210, + 278 + ], + "spans": [ + { + "bbox": [ + 51, + 69, + 210, + 278 + ], + "type": "table", + "html": "
Categoryxyz vs. Gradient Signals
xyzGradient signal
Ground Truth Equivariance Consistency (GEC)↓
bench2.173.26
cellphone1.841.43
chair1.720.94
plane2.070.66
Average1.951.57
Instance-Level Consistency (IC)↓
bench2.042.62
cellphone1.61.28
chair1.40.81
plane1.880.36
Average1.731.26
Category-Level Consistency (CC)↓
bench1.992.77
cellphone1.491.26
chair1.520.75
plane1.90.63
Average1.721.35
", + "image_path": "ab570f5815e6488dc917d430a417903df7f3eb7066357e75f4fbb3419e7d0540.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 47, + 281, + 211, + 321 + ], + "lines": [ + { + "bbox": [ + 47, + 281, + 211, + 321 + ], + "spans": [ + { + "bbox": [ + 47, + 281, + 211, + 321 + ], + "type": "text", + "content": "(a) Choice of Signal Representation - Canonicalization metrics for using Gradients vs. " + }, + { + "bbox": [ + 47, + 281, + 211, + 321 + ], + "type": "inline_equation", + "content": "xyz" + }, + { + "bbox": [ + 47, + 281, + 211, + 321 + ], + "type": "text", + "content": " locations as input signal. Gradients capture the object surface that help in canonicalization." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 217, + 69, + 376, + 278 + ], + "blocks": [ + { + "bbox": [ + 217, + 69, + 376, + 278 + ], + "lines": [ + { + "bbox": [ + 217, + 69, + 376, + 278 + ], + "spans": [ + { + "bbox": [ + 217, + 69, + 376, + 278 + ], + "type": "table", + "html": "
CategoryLocal Average Density
w/ow
Ground Truth Equivariance Consistency (GEC)↓
bench3.263.38
cellphone1.431.63
chair0.941.3
plane0.660.64
Average1.571.73
Instance-Level Consistency (IC)↓
bench2.622.72
cellphone1.281.63
chair0.812.64
plane0.360.39
Average1.261.85
Category-Level Consistency (CC)↓
bench2.772.82
cellphone1.261.54
chair0.751.07
plane0.630.51
Average1.351.49
", + "image_path": "9bd5041d440aa74b96503a4414836d64a88eaa5a7f74d6311172097a47cd6687.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 212, + 281, + 376, + 330 + ], + "lines": [ + { + "bbox": [ + 212, + 281, + 376, + 330 + ], + "spans": [ + { + "bbox": [ + 212, + 281, + 376, + 330 + ], + "type": "text", + "content": "(b) Weighing Equivariant Signals by Local Average Density deteriorates the performance by smoothing out important details of the shape. We show canonicalization with (w) and without (w/o) weighing by the local averaged density." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 383, + 69, + 542, + 278 + ], + "blocks": [ + { + "bbox": [ + 383, + 69, + 542, + 278 + ], + "lines": [ + { + "bbox": [ + 383, + 69, + 542, + 278 + ], + "spans": [ + { + "bbox": [ + 383, + 69, + 542, + 278 + ], + "type": "table", + "html": "
CategorySiamese Training
w/o siamesewith siamese
Ground Truth Equivariance Consistency (GEC)↓
bench3.523.26
cellphone1.611.43
chair1.020.94
plane1.310.66
Average1.861.57
Instance-Level Consistency (IC)↓
bench2.732.62
cellphone1.31.28
chair0.820.81
plane1.10.36
Average1.481.26
Category-Level Consistency (CC)↓
bench2.952.77
cellphone1.441.26
chair0.830.75
plane1.260.63
Average1.621.35
", + "image_path": "38bb6101fee675aa122805e7ff3b28fd9dac0c43121a8ac47ccd4a35fc9abff3.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 378, + 281, + 544, + 338 + ], + "lines": [ + { + "bbox": [ + 378, + 281, + 544, + 338 + ], + "spans": [ + { + "bbox": [ + 378, + 281, + 544, + 338 + ], + "type": "text", + "content": "(c) Siamese Training improves performance on all canonicalization metrics on average. We show canonicalization performance with siamese and without (w/o) siamese training. The average of Ground Truth Equivariance Consistency GEC metric reduces to 1.57 from 1.86" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 342, + 545, + 354 + ], + "lines": [ + { + "bbox": [ + 47, + 342, + 545, + 354 + ], + "spans": [ + { + "bbox": [ + 47, + 342, + 545, + 354 + ], + "type": "text", + "content": "Table 2. Ablation studies to justify three key design choices: (a) signal representation, (b) density weighting, and (c) Siamese architecture." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 47, + 376, + 200, + 387 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 376, + 200, + 387 + ], + "spans": [ + { + "bbox": [ + 47, + 376, + 200, + 387 + ], + "type": "text", + "content": "role in canonicalization performance." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 388, + 287, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 388, + 287, + 579 + ], + "spans": [ + { + "bbox": [ + 46, + 388, + 287, + 579 + ], + "type": "text", + "content": "Next, we compare our method against three benchmarks: Principal Component Analysis (PCA) [32], Canonical Capsules (CaCa) [47], and ConDor [37]. All prior methods canonicalize clean point clouds with all points on the surface. To have a fair comparison, we train the above methods on point clouds (P) sampled from NeRF scenes for points on and within the object. Most scenes have points in the range of 120-1200 when sampled uniformly on the " + }, + { + "bbox": [ + 46, + 388, + 287, + 579 + ], + "type": "inline_equation", + "content": "32^{3}" + }, + { + "bbox": [ + 46, + 388, + 287, + 579 + ], + "type": "text", + "content": " grid. We resample the scene when points are less than 1024 and train both Canonical Capsules and ConDor on the resulting point clouds. CaFi-Net estimates a canonicalizing transform directly from the density field that we use to transform its corresponding ShapeNet point cloud for comparison against point canonicalizers. We follow the same protocol as [37] to measure and benchmark canonicalization performance for all the methods." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 582, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 582, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 582, + 288, + 715 + ], + "type": "text", + "content": "Analysis: Table 1 compares the metrics for PCA (P), CaCa (P), ConDor (P), and CaFi-Net [Ours (F)]. We list point cloud canonicalizers with (P) and field canonicalizer as (F) in the table. On average, our method performs better than point cloud-based methods in the CC and GEC metrics and is on par with the state-of-the-art method (ConDor) in the IC metric. This is despite the fact that our method directly operates on the field. We observe that PCA underperforms in all categories suggesting that the noisy point clouds make it unreliable. Similarly we observe that CaCa fails to canonicalize in categories where high noise is observed. Given" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 376, + 545, + 436 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 376, + 545, + 436 + ], + "spans": [ + { + "bbox": [ + 304, + 376, + 545, + 436 + ], + "type": "text", + "content": "that ConDor also uses TFNs in a different architecture, we conclude that the choice of TFNs makes canonicalization more robust to noise. Our method has the additional advantage of operating directly on the field. Please see Figure 4 for a qualitative comparison of the different methods." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 443, + 373, + 455 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 443, + 373, + 455 + ], + "spans": [ + { + "bbox": [ + 306, + 443, + 373, + 455 + ], + "type": "text", + "content": "5.2. Ablations" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 462, + 545, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 462, + 545, + 510 + ], + "spans": [ + { + "bbox": [ + 304, + 462, + 545, + 510 + ], + "type": "text", + "content": "We conduct additional experiments to justify key design choices in our method. For all the ablation studies we use a smaller subset of our data consisting of 4 categories: bench, cellphone, chair, and airplane." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 510, + 545, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 510, + 545, + 628 + ], + "spans": [ + { + "bbox": [ + 304, + 510, + 545, + 628 + ], + "type": "text", + "content": "Choice of Signal Representation: To select the appropriate signal representation to use (see Section 4), we conducted an ablation study using just the density field " + }, + { + "bbox": [ + 304, + 510, + 545, + 628 + ], + "type": "inline_equation", + "content": "xyz" + }, + { + "bbox": [ + 304, + 510, + 545, + 628 + ], + "type": "text", + "content": " locations or its gradients. As seen in Table 2 (a), the average canonicalization error over 4 categories is lower when using gradient as compared to using density " + }, + { + "bbox": [ + 304, + 510, + 545, + 628 + ], + "type": "inline_equation", + "content": "xyz" + }, + { + "bbox": [ + 304, + 510, + 545, + 628 + ], + "type": "text", + "content": " locations on all three metrics. We hypothesize that gradient of density weighs equivariant features closer to the surface of the object higher. This improves the canonicalization performance." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "content": "Weighing Features by Local Average Density: Next, we justify why weighting the equivariant features using the density is helpful. Table 2 (b) shows results for weighting features by local average density compared to weighting features by density. Weighing features by local average smoothens out details that are necessary to canonicalize the volume. Our average ground truth canonicalization error" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "4506" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 49, + 80, + 544, + 357 + ], + "blocks": [ + { + "bbox": [ + 113, + 69, + 160, + 79 + ], + "lines": [ + { + "bbox": [ + 113, + 69, + 160, + 79 + ], + "spans": [ + { + "bbox": [ + 113, + 69, + 160, + 79 + ], + "type": "text", + "content": "Input NeRFs" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 280, + 69, + 339, + 78 + ], + "lines": [ + { + "bbox": [ + 280, + 69, + 339, + 78 + ], + "spans": [ + { + "bbox": [ + 280, + 69, + 339, + 78 + ], + "type": "text", + "content": "CaFi-Net (Ours)" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 461, + 70, + 479, + 77 + ], + "lines": [ + { + "bbox": [ + 461, + 70, + 479, + 77 + ], + "spans": [ + { + "bbox": [ + 461, + 70, + 479, + 77 + ], + "type": "text", + "content": "PCA" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 49, + 80, + 544, + 357 + ], + "lines": [ + { + "bbox": [ + 49, + 80, + 544, + 357 + ], + "spans": [ + { + "bbox": [ + 49, + 80, + 544, + 357 + ], + "type": "image", + "image_path": "d570ced4afbb5858f75f9c7cf2ab418bfd5caed7eea2ab3266f24df6a4d21b34.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 357, + 546, + 411 + ], + "lines": [ + { + "bbox": [ + 46, + 357, + 546, + 411 + ], + "spans": [ + { + "bbox": [ + 46, + 357, + 546, + 411 + ], + "type": "text", + "content": "Figure 4. This figure visually compares the canonicalization performance of our method with 3D point cloud-based methods. Input NeRFs shows one of the RGB views used to learn our NeRF input (notice how orientations of instances are random). CaFi-Net (Ours) shows the results of our canonical field rendered from a novel view unseen during NeRF training. PCA (blue), CaCa [47] (red), and ConDor [37]'s (green) results are also presented for the same instances. Our method matches or exceeds ConDor. We also show some failure cases for several instances (red box) which occur either due to thin structures or symmetry." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 418, + 287, + 477 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 418, + 287, + 477 + ], + "spans": [ + { + "bbox": [ + 46, + 418, + 287, + 477 + ], + "type": "text", + "content": "increases from 1.57 to 1.73 if we weigh features by local averaged densities. In future works, we will delve into better methods for reducing input signal noise that does not smoothen out the geometry details of objects to further improve our canonicalization." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 479, + 288, + 622 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 479, + 288, + 622 + ], + "spans": [ + { + "bbox": [ + 46, + 479, + 288, + 622 + ], + "type": "text", + "content": "Siamese Architecture: Finally, we justify the need for a two-branch Siamese network architecture. Although our method can canonicalize even without this architecture, it makes it easier to learn over instances in a category as observed in [26]. Table 2 (c) compares our method on using a single branch architecture and a Siamese architecture. Clearly, the Siamese architecture helps improve results confirming our hypothesis. Enforcing different instances of the same category to align with each other reduces the Ground Truth Canonicalization error from 1.86 to 1.57. We also see a decrease in all the other metrics including Instance-Level Consistency." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 633, + 119, + 645 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 633, + 119, + 645 + ], + "spans": [ + { + "bbox": [ + 47, + 633, + 119, + 645 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 653, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 288, + 713 + ], + "type": "text", + "content": "We presented Canonical Field Network (CaFi-Net), a method for self-supervised category-level canonicalization of the 3D pose of objects represented as neural radiance fields. CaFi-Net consists of a Siamese network architecture with rotation-equivariant convolution layers to extract" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 418, + 545, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 418, + 545, + 502 + ], + "spans": [ + { + "bbox": [ + 304, + 418, + 545, + 502 + ], + "type": "text", + "content": "features for pose canonicalization. Our approach operates directly on NeRF's noisy but continuous density field. We train our method on a large dataset of 1300 NeRF models obtained for 13 common ShapeNet categories. During inference, our method is able to canonicalize arbitrarily oriented NeRFs. Experiments show that our method matches or outperforms 3D point cloud-based methods." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 521, + 546, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 521, + 546, + 665 + ], + "spans": [ + { + "bbox": [ + 304, + 521, + 546, + 665 + ], + "type": "text", + "content": "Limitations & Future Work: Our approach has several limitations that future work should investigate. First, we require a " + }, + { + "bbox": [ + 304, + 521, + 546, + 665 + ], + "type": "inline_equation", + "content": "360^{\\circ}" + }, + { + "bbox": [ + 304, + 521, + 546, + 665 + ], + "type": "text", + "content": " NeRF model of objects and cannot handle partial views from front-facing NeRFs. Second, canonicalizing densities with uncertainty is more difficult as the uncertainty increases for very thin/small structures that can be missed (see chair in Figure 4). Furthermore, we also inherit the issues that TFNs have with symmetry (see monitor case in Figure 4). In future work, we plan to extend our approach to create a large real dataset of common object categories without requiring manual pose canonicalization and combat the symmetry issue in CaFi-Net as in [39]" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "type": "text", + "content": "Acknowledgments: This work was supported by NSF grant CNS-2038897, an AWS Cloud Credits award, NSF CloudBank, and a gift from Meta Reality Labs. We thank Chandradeep Pokhariya and Ishaan Shah." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "4507" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 714 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 146 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 146 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 146 + ], + "type": "text", + "content": "[1] Michal Adamkiewicz, Timothy Chen, Adam Caccavale, Rachel Gardner, Preston Culbertson, Jeannette Bohg, and Mac Schwager. Vision-only robot navigation in a neural radiance world. IEEE Robotics and Automation Letters, 7(2):4606-4613, 2022. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 148, + 287, + 170 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 148, + 287, + 170 + ], + "spans": [ + { + "bbox": [ + 53, + 148, + 287, + 170 + ], + "type": "text", + "content": "[2] Abien Fred Agarap. Deep learning using rectified linear units (relu). arXiv preprint arXiv:1803.08375, 2018. 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 171, + 288, + 236 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 171, + 288, + 236 + ], + "spans": [ + { + "bbox": [ + 53, + 171, + 288, + 236 + ], + "type": "text", + "content": "[3] Jonathan T Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P Srinivasan. Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5855–5864, 2021. 1, 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 239, + 288, + 294 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 239, + 288, + 294 + ], + "spans": [ + { + "bbox": [ + 53, + 239, + 288, + 294 + ], + "type": "text", + "content": "[4] Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Mip-nerf 360: Unbounded anti-aliased neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5470–5479, 2022. 1, 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 295, + 288, + 372 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 295, + 288, + 372 + ], + "spans": [ + { + "bbox": [ + 53, + 295, + 288, + 372 + ], + "type": "text", + "content": "[5] Angel X. Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, Jianxiong Xiao, Li Yi, and Fisher Yu. ShapeNet: An Information-Rich 3D Model Repository. Technical Report arXiv:1512.03012 [cs.GR], Stanford University — Princeton University — Toyota Technological Institute at Chicago, 2015. 1, 2, 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 373, + 288, + 417 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 373, + 288, + 417 + ], + "spans": [ + { + "bbox": [ + 53, + 373, + 288, + 417 + ], + "type": "text", + "content": "[6] Zhiqin Chen and Hao Zhang. Learning implicit fields for generative shape modeling. Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 419, + 288, + 452 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 419, + 288, + 452 + ], + "spans": [ + { + "bbox": [ + 53, + 419, + 288, + 452 + ], + "type": "text", + "content": "[7] Taco S Cohen, Mario Geiger, Jonas Kohler, and Max Welling. Spherical cnns. arXiv preprint arXiv:1801.10130, 2018.3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 454, + 288, + 518 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 454, + 288, + 518 + ], + "spans": [ + { + "bbox": [ + 53, + 454, + 288, + 518 + ], + "type": "text", + "content": "[8] Enric Corona, Tomas Hodan, Minh Vo, Francesc Moreno-Noguer, Chris Sweeney, Richard Newcombe, and Lingni Ma. Lisa: Learning implicit shape and appearance of hands. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20533-20543, 2022. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 521, + 288, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 521, + 288, + 586 + ], + "spans": [ + { + "bbox": [ + 53, + 521, + 288, + 586 + ], + "type": "text", + "content": "[9] Congyue Deng, Or Litany, Yueqi Duan, Adrien Poulenard, Andrea Tagliasacchi, and Leonidas J. Guibas. Vector neurons: A general framework for so(3)-equivariant networks. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 12200-12209, October 2021. 2, 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 589, + 287, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 589, + 287, + 632 + ], + "spans": [ + { + "bbox": [ + 48, + 589, + 287, + 632 + ], + "type": "text", + "content": "[10] Fabian B Fuchs, Daniel E Worrall, Volker Fischer, and Max Welling. Se (3)-transformers: 3d roto-translation equivariant attention networks. arXiv preprint arXiv:2006.10503, 2020. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 635, + 288, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 635, + 288, + 678 + ], + "spans": [ + { + "bbox": [ + 48, + 635, + 288, + 678 + ], + "type": "text", + "content": "[11] Ruohan Gao, Yen-Yu Chang, Shivani Mall, Li Fei-Fei, and Jiajun Wu. Objectfolder: A dataset of objects with implicit visual, auditory, and tactile representations. In Proceedings of the Conference on Robot Learning (CoRL), 2021. 2, 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 680, + 288, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 680, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 48, + 680, + 288, + 714 + ], + "type": "text", + "content": "[12] Thibault Groueix, Matthew Fisher, Vladimir G Kim, Bryan C Russell, and Mathieu Aubry. A papier-mâché approach to learning 3d surface generation. In Proceedings of" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 714 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "text", + "content": "the IEEE conference on computer vision and pattern recognition, pages 216-224, 2018. 2, 6" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 308, + 96, + 545, + 129 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 96, + 545, + 129 + ], + "spans": [ + { + "bbox": [ + 308, + 96, + 545, + 129 + ], + "type": "text", + "content": "[13] Jeffrey Ichnowski, Yahav Avigal, Justin Kerr, and Ken Goldberg. Dex-nerf: Using a neural radiance field to grasp transparent objects. Oct 2021. 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 130, + 547, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 130, + 547, + 206 + ], + "spans": [ + { + "bbox": [ + 307, + 130, + 547, + 206 + ], + "type": "text", + "content": "[14] Sergey Ioffe and Christian Szegedy. Batch normalization: Accelerating deep network training by reducing internal covariate shift. In Francis Bach and David Blei, editors, Proceedings of the 32nd International Conference on Machine Learning, volume 37 of Proceedings of Machine Learning Research, pages 448-456, Lille, France, 07-09 Jul 2015. PMLR. 5" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 208, + 547, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 208, + 547, + 252 + ], + "spans": [ + { + "bbox": [ + 308, + 208, + 547, + 252 + ], + "type": "text", + "content": "[15] Ajay Jain, Matthew Tancik, and Pieter Abbeel. Putting nerf on a diet: Semantically consistent few-shot view synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5885-5894, 2021. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 254, + 545, + 275 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 254, + 545, + 275 + ], + "spans": [ + { + "bbox": [ + 308, + 254, + 545, + 275 + ], + "type": "text", + "content": "[16] Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. CoRR, abs/1412.6980, 2015. 5" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 277, + 545, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 277, + 545, + 331 + ], + "spans": [ + { + "bbox": [ + 308, + 277, + 545, + 331 + ], + "type": "text", + "content": "[17] Jiahui Lei, Srinath Sridhar, Paul Guerrero, Minhyuk Sung, Niloy Mitra, and Leonidas J Guibas. Pix2surf: Learning parametric 3d surface models of objects from images. In European Conference on Computer Vision, pages 121-138. Springer, 2020. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 333, + 545, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 333, + 545, + 376 + ], + "spans": [ + { + "bbox": [ + 308, + 333, + 545, + 376 + ], + "type": "text", + "content": "[18] Ruilong Li, Julian Tanke, Minh Vo, Michael Zollhofer, Jurgen Gall, Angjoo Kanazawa, and Christoph Lassner. Tava: Template-free animatable volumetric actors. arXiv preprint arXiv:2206.08929, 2022. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 377, + 545, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 377, + 545, + 453 + ], + "spans": [ + { + "bbox": [ + 308, + 377, + 545, + 453 + ], + "type": "text", + "content": "[19] Tianye Li, Mira Slavcheva, Michael Zollhoefer, Simon Green, Christoph Lassner, Changil Kim, Tanner Schmidt, Steven Lovegrove, Michael Goesele, Richard Newcombe, et al. Neural 3d video synthesis from multi-view video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5521-5531, 2022. 1, 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 456, + 545, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 456, + 545, + 510 + ], + "spans": [ + { + "bbox": [ + 308, + 456, + 545, + 510 + ], + "type": "text", + "content": "[20] Xiaolong Li, He Wang, Li Yi, Leonidas J Guibas, A Lynn Abbott, and Shuran Song. Category-level articulated object pose estimation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3706-3715, 2020. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 512, + 545, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 512, + 545, + 555 + ], + "spans": [ + { + "bbox": [ + 308, + 512, + 545, + 555 + ], + "type": "text", + "content": "[21] Yunzhu Li, Shuang Li, Vincent Sitzmann, Pulkit Agrawal, and Antonio Torralba. 3d neural scene representations for visuomotor control. In Proceedings of Robotics: Science and Systems, 2021. 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 557, + 545, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 557, + 545, + 612 + ], + "spans": [ + { + "bbox": [ + 308, + 557, + 545, + 612 + ], + "type": "text", + "content": "[22] Lars Mescheder, Michael Oechsle, Michael Niemeyer, Sebastian Nowozin, and Andreas Geiger. Occupancy networks: Learning 3d reconstruction in function space. In Proceedings IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2019. 1, 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 613, + 545, + 677 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 613, + 545, + 677 + ], + "spans": [ + { + "bbox": [ + 308, + 613, + 545, + 677 + ], + "type": "text", + "content": "[23] Ben Mildenhall, Peter Hedman, Ricardo Martin-Brualla, Pratul P Srinivasan, and Jonathan T Barron. Nerf in the dark: High dynamic range view synthesis from noisy raw images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16190-16199, 2022. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 680, + 545, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 680, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 308, + 680, + 545, + 714 + ], + "type": "text", + "content": "[24] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view syn" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "4508" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 73, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 66, + 73, + 286, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 73, + 286, + 95 + ], + "spans": [ + { + "bbox": [ + 66, + 73, + 286, + 95 + ], + "type": "text", + "content": "thesis. In European conference on computer vision, pages 405-421. Springer, 2020. 1, 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 96, + 287, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 96, + 287, + 139 + ], + "spans": [ + { + "bbox": [ + 48, + 96, + 287, + 139 + ], + "type": "text", + "content": "[25] Michael Niemeyer, Lars Mescheder, Michael Oechsle, and Andreas Geiger. Occupancy flow: 4d reconstruction by learning particle dynamics. In Proc. of the IEEE International Conf. on Computer Vision (ICCV), 2019. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 140, + 287, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 140, + 287, + 194 + ], + "spans": [ + { + "bbox": [ + 48, + 140, + 287, + 194 + ], + "type": "text", + "content": "[26] David Novotny, Nikhila Ravi, Benjamin Graham, Natalia Neverova, and Andrea Vedaldi. C3dpo: Canonical 3d pose networks for non-rigid structure from motion. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7688-7697, 2019. 2, 3, 4, 5, 8" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 195, + 287, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 195, + 287, + 239 + ], + "spans": [ + { + "bbox": [ + 48, + 195, + 287, + 239 + ], + "type": "text", + "content": "[27] David Novotny, Roman Shapovalov, and Andrea Vedaldi. Canonical 3D Deformer Maps: Unifying parametric and non-parametric methods for dense weakly-supervised category reconstruction. In NeurIPS, 2020. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 240, + 287, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 240, + 287, + 293 + ], + "spans": [ + { + "bbox": [ + 48, + 240, + 287, + 293 + ], + "type": "text", + "content": "[28] Jeong Joon Park, Peter Florence, Julian Straub, Richard Newcombe, and Steven Lovegrove. Deepsdf: Learning continuous signed distance functions for shape representation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 165-174, 2019. 1, 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 295, + 287, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 295, + 287, + 348 + ], + "spans": [ + { + "bbox": [ + 48, + 295, + 287, + 348 + ], + "type": "text", + "content": "[29] Keunhong Park, Utkarsh Sinha, Jonathan T. Barron, Sofien Bouaziz, Dan B Goldman, Steven M. Seitz, and Ricardo Martin-Brualla. Nerfies: Deformable neural radiance fields. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), 2021. 1, 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 350, + 287, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 350, + 287, + 403 + ], + "spans": [ + { + "bbox": [ + 48, + 350, + 287, + 403 + ], + "type": "text", + "content": "[30] Keunhong Park, Utkarsh Sinha, Peter Hedman, Jonathan T. Barron, Sofien Bouaziz, Dan B Goldman, Ricardo MartinBrualla, and Steven M. Seitz. Hypernerf: A higher-dimensional representation for topologically varying neural radiance fields. arXiv preprint arXiv:2106.13228, 2021. 1, 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 405, + 287, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 405, + 287, + 525 + ], + "spans": [ + { + "bbox": [ + 48, + 405, + 287, + 525 + ], + "type": "text", + "content": "[31] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, Alban Desmaison, Andreas Kopf, Edward Yang, Zachary DeVito, Martin Raison, Alykhan Tejani, Sasank Chilamkurthy, Benoit Steiner, Lu Fang, Junjie Bai, and Soumith Chintala. Pytorch: An imperative style, high-performance deep learning library. In H. Wallach, H. Larochelle, A. Beygelzimer, F. d'Alché-Buc, E. Fox, and R. Garnett, editors, Advances in Neural Information Processing Systems 32, pages 8024-8035. Curran Associates, Inc., 2019. 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 526, + 287, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 526, + 287, + 568 + ], + "spans": [ + { + "bbox": [ + 48, + 526, + 287, + 568 + ], + "type": "text", + "content": "[32] Karl Pearson. Liii. on lines and planes of closest fit to systems of points in space. The London, Edinburgh, and Dublin philosophical magazine and journal of science, 2(11):559-572, 1901. 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 570, + 287, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 570, + 287, + 624 + ], + "spans": [ + { + "bbox": [ + 48, + 570, + 287, + 624 + ], + "type": "text", + "content": "[33] Sida Peng, Junting Dong, Qianqian Wang, Shangzhan Zhang, Qing Shuai, Xiaowei Zhou, and Hujun Bao. Animatable neural radiance fields for modeling dynamic human bodies. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14314-14323, 2021. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 625, + 287, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 625, + 287, + 678 + ], + "spans": [ + { + "bbox": [ + 48, + 625, + 287, + 678 + ], + "type": "text", + "content": "[34] Adrien Poulenard and Leonidas J Guibas. A functional approach to rotation equivariant non-linearities for tensor field networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13174-13183, 2021. 3, 4, 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 681, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 681, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 681, + 287, + 713 + ], + "type": "text", + "content": "[35] Maziar Raissi, Paris Perdikaris, and George E Karniadakis. Physics-informed neural networks: A deep learning framework for solving forward and inverse problems involving" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 25, + "blocks": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "type": "text", + "content": "nonlinear partial differential equations. Journal of Computational Physics, 378:686-707, 2019. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 307, + 96, + 545, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 96, + 545, + 139 + ], + "spans": [ + { + "bbox": [ + 307, + 96, + 545, + 139 + ], + "type": "text", + "content": "[36] Viktor Rudnev, Mohamed Elgharib, William Smith, Lingjie Liu, Vladislav Golyanik, and Christian Theobalt. Nerf for outdoor scene relighting. In European Conference on Computer Vision (ECCV), 2022. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 141, + 545, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 141, + 545, + 206 + ], + "spans": [ + { + "bbox": [ + 307, + 141, + 545, + 206 + ], + "type": "text", + "content": "[37] Rahul Sajnani, Adrien Poulenard, Jivitesh Jain, Radhika Dua, Leonidas J Guibas, and Srinath Sridhar. Condor: Self-supervised canonicalization of 3d pose for partial shapes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16969-16979, 2022. 2, 5, 6, 7, 8" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 209, + 545, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 209, + 545, + 274 + ], + "spans": [ + { + "bbox": [ + 307, + 209, + 545, + 274 + ], + "type": "text", + "content": "[38] Rahul Sajnani, AadilMehdi Sanchawala, Krishna Murthy Jatavallabhula, Srinath Sridhar, and K Madhava Krishna. Draco: Weakly supervised dense reconstruction and canonicalization of objects. In 2021 IEEE International Conference on Robotics and Automation (ICRA), pages 10302-10309. IEEE, 2021. 2, 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 276, + 545, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 276, + 545, + 319 + ], + "spans": [ + { + "bbox": [ + 307, + 276, + 545, + 319 + ], + "type": "text", + "content": "[39] Ahyun Seo, Byungjin Kim, Suha Kwak, and Minsu Cho. Reflection and rotation symmetry detection via equivariant learning. In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 8" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 322, + 545, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 322, + 545, + 387 + ], + "spans": [ + { + "bbox": [ + 307, + 322, + 545, + 387 + ], + "type": "text", + "content": "[40] Anthony Simeonov, Yilun Du, Andrea Tagliasacchi, Joshua B Tenenbaum, Alberto Rodriguez, Pulkit Agrawal, and Vincent Sitzmann. Neural descriptor fields: Se (3)-equivariant object representations for manipulation. In 2022 International Conference on Robotics and Automation (ICRA), pages 6394-6400. IEEE, 2022. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 388, + 545, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 388, + 545, + 441 + ], + "spans": [ + { + "bbox": [ + 307, + 388, + 545, + 441 + ], + "type": "text", + "content": "[41] Anthony Simeonov, Yilun Du, Andrea Tagliasacchi, Joshua B. Tenenbaum, Alberto Rodriguez, Pulkit Agrawal, and Vincent Sitzmann. Neural descriptor fields: Se(3)-equivariant object representations for manipulation. 2022. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 445, + 545, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 445, + 545, + 510 + ], + "spans": [ + { + "bbox": [ + 307, + 445, + 545, + 510 + ], + "type": "text", + "content": "[42] Vincent Sitzmann, Michael Zollhoefer, and Gordon Wetzstein. Scene representation networks: Continuous 3d-structure-aware neural scene representations. In H. Wallach, H. Larochelle, A. Beygelzimer, F. d'Alché-Buc, E. Fox, and R. Garnett, editors, Advances in Neural Information Processing Systems, volume 32. Curran Associates, Inc., 2019. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 512, + 545, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 512, + 545, + 555 + ], + "spans": [ + { + "bbox": [ + 307, + 512, + 545, + 555 + ], + "type": "text", + "content": "[43] Riccardo Spezialietti, Federico Stella, Marlon Marcon, Luciano Silva, Samuele Salti, and Luigi Di Stefano. Learning to orient surfaces by self-supervised spherical cnns. arXiv preprint arXiv:2011.03298, 2020. 1, 2, 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 557, + 545, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 557, + 545, + 601 + ], + "spans": [ + { + "bbox": [ + 307, + 557, + 545, + 601 + ], + "type": "text", + "content": "[44] Srinath Sridhar, Davis Rempe, Julien Valentin, Sofien Bouaziz, and Leonidas J Guibas. Multiview aggregation for learning category-specific shape reconstruction. arXiv preprint arXiv:1907.01085, 2019. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 602, + 545, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 602, + 545, + 666 + ], + "spans": [ + { + "bbox": [ + 307, + 602, + 545, + 666 + ], + "type": "text", + "content": "[45] Pratul P Srinivasan, Boyang Deng, Xiuming Zhang, Matthew Tancik, Ben Mildenhall, and Jonathan T Barron. Nerv: Neural reflectance and visibility fields for relighting and view synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7495-7504, 2021. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 670, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 670, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 670, + 545, + 713 + ], + "type": "text", + "content": "[46] Shih-Yang Su, Frank Yu, Michael Zollhöfer, and Helge Rhodin. A-nerf: Articulated neural radiance fields for learning human shape, appearance, and pose. Advances in Neural Information Processing Systems, 34:12278-12291, 2021. 2" + } + ] + } + ], + "index": 24 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "4509" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 714 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 117 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 117 + ], + "type": "text", + "content": "[47] Weiwei Sun, Andrea Tagliasacchi, Boyang Deng, Sara Sabour, Soroosh Yazdani, Geoffrey Hinton, and Kwang Moo Yi. Canonical capsules: Unsupervised capsules in canonical pose. arXiv preprint arXiv:2012.04718, 2020. 1, 2, 3, 6, 7, 8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 118, + 288, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 118, + 288, + 172 + ], + "spans": [ + { + "bbox": [ + 48, + 118, + 288, + 172 + ], + "type": "text", + "content": "[48] Maxim Tatarchenko, Stephan R Richter, René Ranftl, Zhuwen Li, Vladlen Koltun, and Thomas Brox. What do single-view 3d reconstruction networks learn? In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3405-3414, 2019. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 172, + 288, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 172, + 288, + 248 + ], + "spans": [ + { + "bbox": [ + 48, + 172, + 288, + 248 + ], + "type": "text", + "content": "[49] Ayush Tewari, Justus Thies, Ben Mildenhall, Pratul Srinivasan, Edgar Tretschk, Yifan Wang, Christoph Lassner, Vincent Sitzmann, Ricardo Martin-Brualla, Stephen Lombardi, Tomas Simon, Christian Theobalt, Matthias Niessner, Jonathan T. Barron, Gordon Wetzstein, Michael Zollhoefer, and Vladislav Golyanik. Advances in neural rendering. Nov 2021. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 250, + 288, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 250, + 288, + 303 + ], + "spans": [ + { + "bbox": [ + 48, + 250, + 288, + 303 + ], + "type": "text", + "content": "[50] Nathaniel Thomas, Tess Smidt, Steven Kearnes, Lusann Yang, Li Li, Kai Kohlhoff, and Patrick Riley. Tensor field networks: Rotation-and translation-equivariant neural networks for 3d point clouds. arXiv preprint arXiv:1802.08219, 2018. 2, 3, 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 304, + 288, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 304, + 288, + 358 + ], + "spans": [ + { + "bbox": [ + 48, + 304, + 288, + 358 + ], + "type": "text", + "content": "[51] Garvita Tiwari, Dimitrije Antic, Jan Eric Lenssen, Nikolaos Sarafianos, Tony Tung, and Gerard Pons-Moll. Pose-ndf: Modeling human pose manifolds with neural distance fields. In European Conference on Computer Vision (ECCV), October 2022. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 360, + 288, + 425 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 360, + 288, + 425 + ], + "spans": [ + { + "bbox": [ + 48, + 360, + 288, + 425 + ], + "type": "text", + "content": "[52] Edgar Tretschk, Ayush Tewari, Vladislav Golyanik, Michael Zollhöfer, Christoph Lassner, and Christian Theobalt. Non-rigid neural radiance fields: Reconstruction and novel view synthesis of a dynamic scene from monocular video. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12959–12970, 2021. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 426, + 288, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 426, + 288, + 480 + ], + "spans": [ + { + "bbox": [ + 48, + 426, + 288, + 480 + ], + "type": "text", + "content": "[53] Suhani Vora, Noha Radwan, Klaus Greff, Henning Meyer, Kyle Genova, Mehdi S. M. Sajjadi, Etienne Pot, Andrea Tagliasacchi, and Daniel Duckworth. Nesf: Neural semantic fields for generalizable semantic segmentation of 3d scenes, 2021. 1, 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 481, + 288, + 545 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 481, + 288, + 545 + ], + "spans": [ + { + "bbox": [ + 48, + 481, + 288, + 545 + ], + "type": "text", + "content": "[54] He Wang, Srinath Sridhar, Jingwei Huang, Julien Valentin, Shuran Song, and Leonidas J Guibas. Normalized object coordinate space for category-level 6d object pose and size estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2642-2651, 2019. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 547, + 288, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 547, + 288, + 613 + ], + "spans": [ + { + "bbox": [ + 48, + 547, + 288, + 613 + ], + "type": "text", + "content": "[55] Peng Wang, Lingjie Liu, Yuan Liu, Christian Theobalt, Taku Komura, and Wenping Wang. Neus: Learning neural implicit surfaces by volume rendering for multi-view reconstruction. In Proceedings of the Thirtieth International Joint Conference on Artificial Intelligence (IJCAI). International Joint Conferences on Artificial Intelligence Organization, 2021. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 613, + 288, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 613, + 288, + 657 + ], + "spans": [ + { + "bbox": [ + 48, + 613, + 288, + 657 + ], + "type": "text", + "content": "[56] Maurice Weiler, Mario Geiger, Max Welling, Wouter Boomsma, and Taco Cohen. 3d steerable cnns: Learning rotationally equivariant features in volumetric data. In NIPS, pages 10381-10392, 2018. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 658, + 288, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 288, + 690 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 288, + 690 + ], + "type": "text", + "content": "[57] Maurice Weiler, Fred A Hamprecht, and Martin Storath. Learning steerable filters for rotation equivariant cnns. In Proc. CVPR, pages 849-858, 2018. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 691, + 288, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 288, + 714 + ], + "type": "text", + "content": "[58] Zhirong Wu, Shuran Song, Aditya Khosla, Fisher Yu, Linguang Zhang, Xiaou Tang, and Jianxiong Xiao. 3d" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 599 + ], + "type": "list", + "angle": 0, + "index": 25, + "blocks": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "type": "text", + "content": "shapenets: A deep representation for volumetric shapes. In Proc. CVPR, pages 1912-1920, 2015. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 307, + 95, + 545, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 95, + 545, + 149 + ], + "spans": [ + { + "bbox": [ + 307, + 95, + 545, + 149 + ], + "type": "text", + "content": "[59] Yiheng Xie, Towaki Takikawa, Shunsuke Saito, Or Litany, Shiqin Yan, Numair Khan, Federico Tombari, James Tompkin, Vincent Sitzmann, and Srinath Sridhar. Neural fields in visual computing and beyond. Computer Graphics Forum, 2022. 1, 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 151, + 545, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 151, + 545, + 195 + ], + "spans": [ + { + "bbox": [ + 307, + 151, + 545, + 195 + ], + "type": "text", + "content": "[60] Guandao Yang, Serge Belongie, Bharath Hariharan, and Vladlen Koltun. Geometry processing with neural fields. Advances in Neural Information Processing Systems, 34:22483-22497, 2021. 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 197, + 545, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 197, + 545, + 239 + ], + "spans": [ + { + "bbox": [ + 307, + 197, + 545, + 239 + ], + "type": "text", + "content": "[61] Lior Yariv, Jiatao Gu, Yoni Kasten, and Yaron Lipman. Volume rendering of neural implicit surfaces. In Advances in Neural Information Processing Systems (NeurIPS). Curran Associates, Inc., 2021. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 241, + 545, + 263 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 241, + 545, + 263 + ], + "spans": [ + { + "bbox": [ + 307, + 241, + 545, + 263 + ], + "type": "text", + "content": "[62] Lin Yen-Chen. Nerf-pytorch. https://github.com/yenchenlin/nerf-pytorch/, 2020.6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 264, + 545, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 264, + 545, + 319 + ], + "spans": [ + { + "bbox": [ + 308, + 264, + 545, + 319 + ], + "type": "text", + "content": "[63] Alex Yu, Vickie Ye, Matthew Tancik, and Angjoo Kanazawa. pixelnerf: Neural radiance fields from one or few images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4578-4587, 2021. 1, 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 319, + 545, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 319, + 545, + 374 + ], + "spans": [ + { + "bbox": [ + 308, + 319, + 545, + 374 + ], + "type": "text", + "content": "[64] Kevin Zakka, Andy Zeng, Johnny Lee, and Shuran Song. Form2fit: Learning shape priors for generalizable assembly from disassembly. In 2020 IEEE International Conference on Robotics and Automation (ICRA), pages 9404-9410. IEEE, 2020. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 376, + 545, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 376, + 545, + 409 + ], + "spans": [ + { + "bbox": [ + 308, + 376, + 545, + 409 + ], + "type": "text", + "content": "[65] Ge Zhang, Or Litany, Srinath Sridhar, and Leonidas Guibas. Strobenet: Category-level multiview reconstruction of articulated objects. arXiv preprint arXiv:2105.08016, 2021. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 410, + 545, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 410, + 545, + 442 + ], + "spans": [ + { + "bbox": [ + 308, + 410, + 545, + 442 + ], + "type": "text", + "content": "[66] Kai Zhang, Gernot Riegler, Noah Snavely, and Vladlen Koltun. Nerf++: Analyzing and improving neural radiance fields. arXiv preprint arXiv:2010.07492, 2020. 1, 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 443, + 547, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 443, + 547, + 487 + ], + "spans": [ + { + "bbox": [ + 308, + 443, + 547, + 487 + ], + "type": "text", + "content": "[67] Yongheng Zhao, Tolga Birdal, Haowen Deng, and Federico Tombari. 3d point capsule networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1009-1018, 2019. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 488, + 545, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 488, + 545, + 542 + ], + "spans": [ + { + "bbox": [ + 308, + 488, + 545, + 542 + ], + "type": "text", + "content": "[68] Shuaifeng Zhi, Tristan Laidlow, Stefan Leutenegger, and Andrew J. Davison. In-place scene labelling and understanding with implicit scene representation. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), 2021. 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 544, + 545, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 544, + 545, + 599 + ], + "spans": [ + { + "bbox": [ + 308, + 544, + 545, + 599 + ], + "type": "text", + "content": "[69] Ellen D Zhong, Adam Lerer, Joseph H Davis, and Bonnie Berger. Cryodrgn2: Ab initio neural reconstruction of 3d protein structures from real cryo-em images. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4066-4075, 2021. 2" + } + ] + } + ], + "index": 24 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 316, + 757 + ], + "type": "text", + "content": "4510" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Cap4Video_ What Can Auxiliary Captions Do for Text-Video Retrieval_/926c868d-a104-48e5-bcb0-c1285d26ad5d_content_list.json b/2023/Cap4Video_ What Can Auxiliary Captions Do for Text-Video Retrieval_/926c868d-a104-48e5-bcb0-c1285d26ad5d_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..d6bb5fc2412e8efb7b6e179d3befcbf472766736 --- /dev/null +++ b/2023/Cap4Video_ What Can Auxiliary Captions Do for Text-Video Retrieval_/926c868d-a104-48e5-bcb0-c1285d26ad5d_content_list.json @@ -0,0 +1,1589 @@ +[ + { + "type": "text", + "text": "Cap4Video: What Can Auxiliary Captions Do for Text-Video Retrieval?", + "text_level": 1, + "bbox": [ + 122, + 130, + 846, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Wenhao Wu $^{1,2*}$ Haipeng Luo $^{3*}$ Bo Fang $^{3}$ Jingdong Wang $^{2}$ Wanli Ouyang $^{4,1}$ $^{1}$ The University of Sydney \n $^{2}$ Baidu Inc. \n $^{3}$ University of Chinese Academy of Sciences \n $^{4}$ Shanghai AI Laboratory \nwhwu.ucas@gmail.com", + "bbox": [ + 119, + 178, + 846, + 251 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 286, + 313, + 301 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Most existing text-video retrieval methods focus on cross-modal matching between the visual content of videos and textual query sentences. However, in real-world scenarios, online videos are often accompanied by relevant text information such as titles, tags, and even subtitles, which can be utilized to match textual queries. This insight has motivated us to propose a novel approach to text-video retrieval, where we directly generate associated captions from videos using zero-shot video captioning with knowledge from web-scale pre-trained models (e.g., CLIP and GPT-2). Given the generated captions, a natural question arises: what benefits do they bring to text-video retrieval? To answer this, we introduce Cap4Video, a new framework that leverages captions in three ways: i) Input data: video-caption pairs can augment the training data. ii) Intermediate feature interaction: we perform cross-modal feature interaction between the video and caption to produce enhanced video representations. iii) Output score: the Query-Caption matching branch can complement the original Query-Video matching branch for text-video retrieval. We conduct comprehensive ablation studies to demonstrate the effectiveness of our approach. Without any post-processing, Cap4Video achieves state-of-the-art performance on four standard text-video retrieval benchmarks: MSR-VTT $(51.4\\%)$ , VATEX $(66.6\\%)$ , MSVD $(51.8\\%)$ , and DiDeMo $(52.0\\%)$ . The code is available at https://github.com/whwu95/Cap4Video.", + "bbox": [ + 75, + 318, + 473, + 728 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 78, + 757, + 209, + 772 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Text-video retrieval is a fundamental task in videolanguage learning. With the rapid advancements in image-language pre-training [15, 30, 46, 47], researchers have focused on expanding pre-trained image-language models, especially CLIP [30], to tackle the text-video retrieval task. The research path has evolved from the most direct global", + "bbox": [ + 75, + 782, + 468, + 875 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/cd8bb2a400ae5fd45da3cd46c3d9b8da397d9bca24081d14ead1b6c00fa35874.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 506, + 285, + 877, + 330 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/893e334822d01329b3b958bcb5288b1fac058e0bb14c1c4e7ad1b735432f0bb7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 506, + 342, + 880, + 401 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/f08ac9db5530b7116bde109bf9af13292ae8bf4701bc6d49e256548d1794f9d7.jpg", + "image_caption": [ + "Figure 1. (a) An existing end-to-end learning paradigm for text-video retrieval. (b) Zero-shot video captioning achieved by guiding a large language model (LLM) such as GPT-2 [31] with CLIP [30]. (c) Our Cap4Video framework leverages the generated captions in three aspects: input data augmentation, intermediate feature interaction, and output score fusion." + ], + "image_footnote": [], + "bbox": [ + 506, + 411, + 887, + 564 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "matching (i.e., video-sentence alignment [11, 24]) to fine-grained matching (e.g., frame-word alignment [36], videoword alignment [13], multi-hierarchical alignment [9, 28], etc.). These studies have demonstrated remarkable performance and significantly outperformed previous models. Two key factors contribute to this improvement. Firstly, CLIP offers powerful visual and textual representations that are pre-aligned in the semantic embedding space, thereby reducing the challenge of cross-modal learning in videotext matching. Secondly, these methods can fine-tune the pre-trained vision and text encoders using sparsely sampled frames in an end-to-end manner. All of these methods aim to learn cross-modal alignment between the visual represen", + "bbox": [ + 496, + 703, + 893, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Equal contribution.", + "bbox": [ + 94, + 886, + 205, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "10704", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "tation of videos and the textual representation of the corresponding query, as depicted in Figure 1(a).", + "bbox": [ + 75, + 90, + 468, + 119 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "However, in real-life scenarios, online videos usually come with related content such as the video's title or tag on the video website. In addition to the visual signal in the video, the associated textual information can also be used to some extent to describe the video content and match the query (i.e., the common text-to-text retrieval). This raises a pertinent question: How can we generate associated text descriptions for videos? One possible solution is to crawl the video title from the video website. However, this method relies on annotations, and there is a risk that the video URL may have become invalid. Another automated solution is to generate captions using zero-shot video caption models. Therefore, we turn our attention to knowledge-rich pretrained models to handle such challenging open-set scenarios. We find that the recent study ZeroCap [34] provides a good practice to use frozen CLIP [30] and GPT-2 [31] for zero-shot image captioning. Thus, we leverage a video extension [33] of ZeroCap for generating captions in the video domain without any further training.", + "bbox": [ + 75, + 121, + 472, + 407 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "When provided with auxiliary captions, a natural question naturally arises: How can we leverage these captions to enhance the text-video retrieval task? In this paper, we propose the Cap4Video learning framework, as illustrated in Figure 1(c), which utilizes captions in three key ways: (i) Input Data: One simple approach is to augment the training data with the generated captions. Specifically, the given video and its generated caption can be treated as a matched pair, which serves as an additional positive sample pair for training beyond the query-video pairs. (ii) Intermediate Feature Interaction: Cross-modal interaction between the video and captions can be leveraged to improve the video representation. Specifically, we can exploit the complementary information between videos and captions to reduce redundant features from videos and learn more discriminative video representations. (iii) Output score: The generated caption can also represent the video's content, allowing us to employ query-caption matching to complement standard query-video matching for the text-video retrieval task. Moreover, a two-stream architecture can be utilized to reduce model bias and produce more robust results.", + "bbox": [ + 75, + 409, + 472, + 726 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We hope that our novel paradigm will encourage further investigation into the video-language learning. In summary, our contributions are as follows:", + "bbox": [ + 75, + 726, + 470, + 771 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We explore a novel problem: leveraging auxiliary captions to further enhance existing text-video retrieval. Besides labor-intensive manual crawling of video website titles, we investigate the potential of rich captions automatically generated by large language models (LLMs) to benefit text-video retrieval.", + "- We propose the Cap4Video learning framework," + ], + "bbox": [ + 94, + 782, + 468, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "which maximizes the utility of the auxiliary captions through three aspects: input data, feature interaction, and output score. Our framework improves the performance of existing query-video matching mechanisms, including global matching and fine-grained matching.", + "bbox": [ + 529, + 90, + 890, + 167 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "- Extensive experiments conducted on four video benchmarks demonstrate the effectiveness of our method. Our Cap4Video achieves state-of-the-art performance on MSR-VTT [44] (51.4%), VATEX [38] (66.6%), MSVD [43] (51.8%), and DiDeMo [1] (52.0%).", + "bbox": [ + 517, + 175, + 893, + 251 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Methodology", + "text_level": 1, + "bbox": [ + 500, + 262, + 635, + 280 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Background: Text-Video Matching", + "text_level": 1, + "bbox": [ + 500, + 287, + 805, + 304 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Text-video matching aims to evaluate the similarity between a given sentence $Q_{i}$ and a given video $V_{j}$ , typically using a similarity function $s(Q_{i},V_{j})$ . In text-to-video retrieval, the goal is to rank all videos based on their similarity scores to a given query sentence. To improve text-video retrieval, recent works [9, 11, 24] have applied CLIP [30] for initialization, leveraging pre-trained knowledge from image-text learning. Our baselines for text-video matching include two typical mechanisms: global matching and fine-grained matching, as illustrated in Figure 2(b).", + "bbox": [ + 496, + 310, + 893, + 460 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Global Matching is a commonly used technique in cross-modal contrastive learning [16, 24, 30]. In global matching, each modality is encoded independently to obtain global features, which are then used to calculate similarity. We train the visual encoder to output $F$ frame embeddings for a given video that samples $F$ frames. Similarly, the query encoder returns $W$ word embeddings and the [CLS] embedding as the global representation for a given query sentence that contains $W$ words. The frame embeddings are integrated using average pooling to obtain the global video embedding, which is then compared with the global query embedding to calculate similarity.", + "bbox": [ + 496, + 462, + 893, + 643 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Fine-grained Matching focuses on modeling the token-level alignment between two modalities, such as framework alignment. In order to achieve token-level patchword alignment for image-text learning, FILIP [45] and ColBERT [17] employ a Max-Mean pipeline. This pipeline finds the token-wise maximum similarity between patch and word tokens, and then averages the maximum similarity of tokens in the image or text to obtain the similarity between an image and a text or vice versa. Moreover, DRL [36] extends the token-wise alignment to text-video retrieval and introduces an attention mechanism to learn weighted pooling instead of mean pooling. We have adopted this mechanism as our enhanced baseline.", + "bbox": [ + 496, + 643, + 893, + 839 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Preprocessing: Caption Generation", + "text_level": 1, + "bbox": [ + 500, + 847, + 808, + 864 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To obtain auxiliary captions for a given video, we consider the following two approaches.", + "bbox": [ + 500, + 869, + 890, + 901 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "10705", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/5c9205f54a594b05cb3e4eb4d0fbf51423143a5bd711a36da850cfd413e19c99.jpg", + "image_caption": [ + "(a) Our Cap4Video", + "(b) Query-Video Matching: Two typical mechanisms", + "(c) Query-Caption Matching" + ], + "image_footnote": [], + "bbox": [ + 78, + 104, + 890, + 255 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/43d5630e5396a15b2a910fed787ac0af2fcc7fefd165e0035646efcd40e5ff13.jpg", + "image_caption": [ + "① Global embedding matching", + "Figure 2. An overview of our Cap4Video for text-video retrieval. We first generate captions using a zero-shot video captioner that combines CLIP [30] with GPT-2 [31], leveraging knowledge from both frozen web-scale models. We then utilize the pre-extracted caption information from three different perspectives: i) Input data: We use the video and captions to create new positive pairs for data augmentation during training. ii) Feature interaction: We perform feature interaction between video and caption to capture intra- and inter-modality context, yielding enhanced video representations. iii) Output score: The Query-Caption matching branch can complement the original Query-Video matching branch for text-video retrieval." + ], + "image_footnote": [], + "bbox": [ + 81, + 301, + 325, + 402 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/05872dffca63a7c7939be9d6844c32c5f148399b350605019d22356d859b3ce4.jpg", + "image_caption": [ + "② Fine-grained embedding matching" + ], + "image_footnote": [], + "bbox": [ + 330, + 301, + 630, + 401 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/29d4af8d46974b636e0b2b11ac5e56d999474addfb9339d9e2cf64bba9ad92a0.jpg", + "image_caption": [ + "Captions embeddings" + ], + "image_footnote": [], + "bbox": [ + 650, + 297, + 890, + 397 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Manual Crawling of Video Titles. We extract the video website title by crawling the original links (such as YouTube ID) of each video and utilize it as the caption. However, we skip this step for videos with expired links.", + "bbox": [ + 75, + 529, + 468, + 589 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Automatic Video Captioning. In contrast to the manual approach that relies on annotations, we leverage knowledge from the LLM to generate rich and diverse captions. Given the scalability of our framework, we aim to generate captions directly from downstream videos without any additional training, a process known as zero-shot video captioning. To achieve this, we follow [33, 34] and use GPT-2 [31] to predict the next word from an initial prompt, e.g., \"Video shows\". A calibrated CLIP [30] loss is then used to drive the model to generate sentences that describe the video, incorporating video-related knowledge into the auto-regressive process. See Supplementary for more details.", + "bbox": [ + 75, + 590, + 468, + 772 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.3. Data Augmentation with Auxiliary Captions", + "text_level": 1, + "bbox": [ + 76, + 785, + 450, + 801 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Auxiliary captions can be used to augment training data. For example, for a dataset consisting of $N$ videos and their corresponding query sentences, each video and its generated caption can be considered as a positive sample pair for training, in addition to the original query-video pairs. By selecting one caption per video, we can add at least $N$ pairs", + "bbox": [ + 75, + 810, + 468, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "as additional data augmentation during training.", + "bbox": [ + 500, + 530, + 815, + 544 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The automatic video captioner can generate multiple captions (e.g., 20) for each video. However, some of these captions may contain noise and may not be entirely relevant to the video content. To avoid negative effects on training, we use a filtering mechanism that evaluates the semantic similarity between each caption and the ground-truth query of the video using a pre-trained text encoder. The caption with the highest similarity is then chosen for data augmentation. Note that we only use the ground-truth query for caption filtering during the training phase.", + "bbox": [ + 496, + 546, + 890, + 698 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.4. Video-Caption Cross-Modal Interaction", + "text_level": 1, + "bbox": [ + 500, + 709, + 841, + 726 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We further consider taking advantage of the complementarity between videos and captions to reduce redundant features and learn more discriminative video representations. To preserve the pre-trained CLIP encoder architecture for efficient transfer learning, we limit the interaction to the final caption and frame embeddings. Specifically, we pass the frame embeddings $\\mathbf{e}_{\\mathbf{v}} = \\{\\mathbf{v}_1,\\mathbf{v}_2,\\dots ,\\mathbf{v}_F\\}$ and caption embeddings $\\mathbf{e_c} = \\{\\mathbf{c}_1,\\mathbf{c}_2,\\dots ,\\mathbf{c}_C\\}$ to the interaction module, where $F$ and $C$ represent the number of frames and captions, respectively. Figure 2(b) depicts several ways of interaction between the two modalities.", + "bbox": [ + 496, + 734, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "10706", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Sum. To obtain an enhanced frame embedding, an intuitive approach is to compute the sum of the global caption embedding $\\mathbf{c}_{\\mathbf{g}}$ and each frame embedding:", + "bbox": [ + 76, + 90, + 468, + 137 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {S u m} \\left(\\mathbf {v} _ {i}, \\mathbf {c} _ {\\mathbf {g}}\\right) = \\mathbf {v} _ {i} + \\mathbf {c} _ {\\mathbf {g}}, \\quad i = 1, \\dots , F, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 125, + 146, + 468, + 162 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathbf{v}_i\\in \\mathbb{R}^D$ is the $i$ -th frame embedding, and $\\mathbf{c_g}\\in \\mathbb{R}^D$ is computed by averaging the [CLS] embeddings of $C$ generated captions: $\\mathbf{c_g} = \\frac{1}{C}\\sum_{i = 1}^{C}\\mathbf{c}_i$ .", + "bbox": [ + 76, + 171, + 468, + 218 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "MLP. To model weighted combinations of each frame embedding and the global caption embedding $\\mathbf{c_g}$ , we concatenate them together and pass the result through a learnable Multi-layer Perceptron (MLP):", + "bbox": [ + 76, + 218, + 468, + 277 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {M L P} \\left(\\mathbf {v} _ {i}, \\mathbf {c} _ {\\mathbf {g}}\\right) = f _ {\\theta} \\left(\\left[ \\mathbf {v} _ {i}, \\mathbf {c} _ {\\mathbf {g}} \\right]\\right), \\quad \\text {f o r} i = 1, \\dots , F, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 98, + 287, + 468, + 305 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $[\\cdot ,\\cdot ]$ denotes the concatenation operation, $f_{\\theta}$ is the MLP with parameter $\\theta$", + "bbox": [ + 76, + 313, + 468, + 343 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Cross Transformer. We also investigate the use of selfattention [35] for interactions. The Cross Transformer operates on a sequence $\\{\\mathbf{e}_{\\mathbf{v}}, \\mathbf{e}_{\\mathbf{c}}\\} = \\{\\mathbf{v}_1, \\dots, \\mathbf{v}_F, \\mathbf{c}_1, \\dots, \\mathbf{c}_C\\}$ and processes them through $L$ encoder-style transformer blocks to generate final representations:", + "bbox": [ + 76, + 343, + 468, + 417 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {C r o s s} \\left(\\mathbf {e} _ {\\mathbf {v}}, \\mathbf {e} _ {\\mathbf {c}}\\right) = f _ {\\psi} \\left(\\left\\{\\mathbf {e} _ {\\mathbf {v}}, \\mathbf {e} _ {\\mathbf {c}} \\right\\}\\right), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 166, + 429, + 468, + 445 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\{\\}$ denotes that $\\mathbf{e_v}$ and $\\mathbf{e_c}$ form a sequence, and $f_{\\psi}$ represents the transformer encoders with parameter $\\psi$", + "bbox": [ + 76, + 454, + 467, + 484 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Co-attention Transformer. Co-attention [23] is another common method for exchanging information between modalities, allowing for mutual attention between video and caption. After this co-attentional transformer layer, we include $L$ transformer layers to model temporal information:", + "bbox": [ + 76, + 484, + 468, + 560 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {C o A t t n} \\left(\\mathbf {e} _ {\\mathbf {v}}, \\mathbf {e} _ {\\mathbf {c}}\\right) = f _ {\\phi_ {2}} \\left(f _ {\\phi_ {1}} \\left(\\left\\{\\mathbf {e} _ {\\mathbf {v}}, \\mathbf {e} _ {\\mathbf {c}} \\right\\}\\right)\\right), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 138, + 575, + 468, + 592 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $f_{\\phi_1}$ is the co-attentional transformer with parameter $\\phi_1$ and $f_{\\phi_2}$ is the transformer encoders with parameter $\\phi_2$ .", + "bbox": [ + 76, + 597, + 468, + 627 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The video-caption interaction module generates frame embeddings that can be further processed based on the type of matching needed. For global matching, the frame embeddings can be averaged to obtain a single video representation. Alternatively, for fine-grained matching, the individual frame embeddings can be retained.", + "bbox": [ + 75, + 627, + 468, + 717 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.5. Complementary Query-Caption Matching", + "text_level": 1, + "bbox": [ + 76, + 726, + 437, + 742 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Besides using the caption for data augmentation and video feature enhancement, it can also directly represent the video content, allowing for text-text retrieval. Specifically, each of the $C$ captions generated by the video is then passed through the caption encoder to obtain its [CLS] text embedding. These caption embeddings are then aggregated to form a global representation, as illustrated in Figure 2(c). The cosine similarity between this global caption embedding and the global query embedding is then calculated to complement the query-video matching.", + "bbox": [ + 75, + 750, + 468, + 900 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/5b5f3522b5e8a75ce629e18e42860268fcf62a74838d1f22f4eb6b0a07ae1473.jpg", + "image_caption": [ + "(a) Sum" + ], + "image_footnote": [], + "bbox": [ + 511, + 87, + 630, + 157 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/8ff8f77e0760782edf84a34a5fbe4b5bacb5620eee20a3123f8826c7987e564c.jpg", + "image_caption": [ + "(b) MLP" + ], + "image_footnote": [], + "bbox": [ + 683, + 87, + 893, + 157 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/f83a8122a918f22cf8f8b069a044a7a992d34651603c0f088a8751cbb80203aa.jpg", + "image_caption": [ + "(c) Cross Transformer", + "Figure 3. Illustration of four Video-Caption interaction strategies. The enhanced frame embeddings will be followed by a mean pooling for global matching or will remain for fine-grained matching." + ], + "image_footnote": [], + "bbox": [ + 511, + 183, + 648, + 270 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/407c3f1c9cbf64b35e7a8775624dd74fe29a5ff0d72c8ccac198a8deab4519a0.jpg", + "image_caption": [ + "(d) Co-attention Transformer", + "Caption Embedding", + "Frame Embedding" + ], + "image_footnote": [], + "bbox": [ + 684, + 180, + 893, + 268 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Notation. Let $\\{\\mathbf{e}_{\\mathbf{v}_i},\\mathbf{e}_{\\mathbf{t}_i},\\mathbf{e}_{\\mathbf{c}_i}\\}_{i = 1}^B$ be a batch of $B$ triples, where $\\mathbf{e}_{\\mathbf{v}_i}$ , $\\mathbf{e}_{\\mathbf{t}_i}$ , and $\\mathbf{e}_{\\mathbf{c}_i}$ denote the $i$ -th video, query, and caption embedding, respectively. Note that the term \"embedding\" used here is more general for convenience and can vary in meaning depending on the situation. For instance, in query-video global matching, $\\mathbf{e}_{\\mathbf{v}_i}$ and $\\mathbf{e}_{\\mathbf{t}_i}$ represent the averaged video feature and global [CLS] text feature, respectively. In query-video fine-grained matching, $\\mathbf{e}_{\\mathbf{v}_i}$ and $\\mathbf{e}_{\\mathbf{t}_i}$ represent a sequence of frame embeddings and a sequence of word embeddings, respectively. In query-caption matching, $\\mathbf{e}_{\\mathbf{c}_i}$ represents a sequence of caption embeddings, and $\\mathbf{e}_{\\mathbf{t}_i}$ represents a global [CLS] text feature.", + "bbox": [ + 498, + 398, + 890, + 582 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Learning Objectives. For the Query-Caption branch, we want the caption embedding $\\mathbf{e_c}$ and the query embedding $\\mathbf{e_t}$ to be close while they are related and far apart when they are not during training phase. We follow the common practice [24, 36] to consider the bidirectional learning objective. We employ symmetric cross-entropy loss to maximize the similarity between matched Query-Caption pairs and minimize the similarity for other pairs:", + "bbox": [ + 498, + 584, + 890, + 705 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {Q 2 C} = - \\frac {1}{B} \\sum_ {i} ^ {B} \\log \\frac {\\exp \\left(s _ {q c} \\left(\\mathbf {e} _ {\\mathbf {t} i} , \\mathbf {e} _ {\\mathbf {c} i}\\right) / \\tau\\right)}{\\sum_ {j} ^ {B} \\exp \\left(s _ {q c} \\left(\\mathbf {e} _ {\\mathbf {t} i} , \\mathbf {e} _ {\\mathbf {c} j}\\right) / \\tau\\right)},\n$$\n", + "text_format": "latex", + "bbox": [ + 521, + 722, + 846, + 763 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {C 2 Q} = - \\frac {1}{B} \\sum_ {i} ^ {B} \\log \\frac {\\exp \\left(s _ {q c} \\left(\\mathbf {e} _ {\\mathbf {t} i} , \\mathbf {e} _ {\\mathbf {c} i}\\right) / \\tau\\right)}{\\sum_ {j} ^ {B} \\exp \\left(s _ {q c} \\left(\\mathbf {e} _ {\\mathbf {t} j} , \\mathbf {e} _ {\\mathbf {c} i}\\right) / \\tau\\right)}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 522, + 766, + 890, + 808 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {Q C} = \\frac {1}{2} \\left(\\mathcal {L} _ {Q 2 C} + \\mathcal {L} _ {C 2 Q}\\right),\n$$\n", + "text_format": "latex", + "bbox": [ + 594, + 810, + 777, + 839 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $s_{qc}(\\cdot, \\cdot)$ represents the query-caption matching similarity function shown in Figure 2(c), and $\\tau$ refers to the temperature hyper-parameter for scaling. Similarly, the con", + "bbox": [ + 521, + 854, + 890, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "10707", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "trastive loss for Query-Video branch is formulated as:", + "bbox": [ + 76, + 90, + 434, + 107 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} _ {Q 2 V} = - \\frac {1}{B} \\sum_ {i} ^ {B} \\log \\frac {\\exp (s _ {q v} (\\mathbf {e} _ {\\mathbf {t} i} , \\mathbf {e} _ {\\mathbf {v} i}) / \\tau)}{\\sum_ {j} ^ {B} \\exp (s _ {q v} (\\mathbf {e} _ {\\mathbf {t} i} , \\mathbf {e} _ {\\mathbf {v} j}) / \\tau)}, \\\\ \\mathcal {L} _ {V 2 Q} = - \\frac {1}{B} \\sum_ {i} ^ {B} \\log \\frac {\\exp \\left(s _ {q v} \\left(\\mathbf {e} _ {\\mathbf {t} i} , \\mathbf {e} _ {\\mathbf {v} i}\\right) / \\tau\\right)}{\\sum_ {j} ^ {B} \\exp \\left(s _ {q v} \\left(\\mathbf {e} _ {\\mathbf {t} j} , \\mathbf {e} _ {\\mathbf {v} i}\\right) / \\tau\\right)}, \\tag {6} \\\\ \\mathcal {L} _ {Q V} = \\frac {1}{2} (\\mathcal {L} _ {Q 2 V} + \\mathcal {L} _ {V 2 Q}), \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 96, + 114, + 468, + 233 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $s_{qv}(\\cdot, \\cdot)$ represents the query-video matching (e.g., global matching, fine-grained matching) similarity function shown in Figure 2(b). The total loss $\\mathcal{L}$ is the sum of Query-Video loss $\\mathcal{L}_{QV}$ and Query-Caption loss $\\mathcal{L}_{QC}$ :", + "bbox": [ + 76, + 241, + 468, + 303 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\mathcal {L} _ {Q V} + \\mathcal {L} _ {Q C}. \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 207, + 311, + 468, + 330 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3. Experiments: Text-Video Retrieval", + "text_level": 1, + "bbox": [ + 76, + 339, + 395, + 356 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.1. Setups", + "text_level": 1, + "bbox": [ + 76, + 363, + 165, + 380 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Datasets. We conduct experiment on four popular benchmarks for video-to-text retrieval and text-to-video retrieval tasks. MSR-VTT [44] contains a total of 10K video clips, each having 20 captions. Following the data splits from [10,24,27], we train models with associated captions on the Training-9K set and report results on the test 1K-A set. DiDeMo [1] has 10K videos paired with 40K descriptions. Following previous works [2, 19, 24], we concatenate all descriptions of one video to a single query, acting as a video-paragraph retrieval task. VATEX [38] collects $\\sim 35\\mathrm{K}$ videos, each with multiple annotations. There are $\\sim 26\\mathrm{K}$ videos for training, 1,500 videos for validation and 1,500 videos for testing. MSVD [43] contains 1,970 videos with 80K captions, with $\\sim 40$ captions on average per video. There are 1,200, 100, and 670 videos in the train, validation, and test sets, respectively.", + "bbox": [ + 75, + 386, + 468, + 628 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Evaluation Metrics. For brevity, we abbreviate Recall at $K$ to $\\mathbb{R}@\\mathbb{K}$ ( $K = 1,5,10$ ) upon all datasets, which computes the percentage of correct videos among the top $K$ retrieved videos given textual queries (Text→Video, and vice versa). MdR, Median Rank, computes the median of the ground-truth in the retrieval ranking list. MnR, Mean Rank, computes the mean rank of the correct results in the retrieval ranking list. Note that for MdR and MnR, the lower score means the better (indicated as $\\downarrow$ ).", + "bbox": [ + 75, + 628, + 468, + 763 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Implementation Details. All experiments use the visual encoder in CLIP [30] as the video encoder, and the textual encoder in CLIP as both the caption encoder and query encoder. The caption encoder and query encoder share parameters. To reduce conflict between the two branches, the query-video branch is trained first, followed by the query-caption branch. The text length is fixed to 32, and the video length is fixed to 12 for all datasets except DiDeMo (64 max words and 64 frames). The initial learning rate is set to 1e-7", + "bbox": [ + 75, + 763, + 470, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "for the clip parameters and 1e-4 for the non-clip parameters. The model is trained with a batch size of 128 for 5 epochs, except for DiDeMo (15 epochs), using the Adam [18] optimizer. All learning rates follow the cosine schedule with a linear warmup [14] strategy. For the number of generated captions per video, we set $C$ to 30. The interaction module employs $L$ transformer layers, where $L$ is set to 4 for VA-TEX and MSR-VTT, and 1 for Didemo and MSVD. In the caption branch, the number of transformer layers is set to 2. For caption generation, we directly use the original pretrained CLIP and GPT-2, without any additional tuning.", + "bbox": [ + 496, + 90, + 890, + 258 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2. Comparison with State-of-the-Arts", + "text_level": 1, + "bbox": [ + 498, + 267, + 805, + 284 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this section, we compare our Cap4Video with recent state-of-the-art methods on four benchmarks: MSR-VTT [44], MSVD [43], VATEX [38], and DiDeMo [1].", + "bbox": [ + 496, + 291, + 890, + 337 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Table 1 shows the comparisons on DiDeMo, where our Cap4Video outperforms CLIP4Clip [24] by a significant margin of $9.2\\%$ in R@1 and exceeds DRL [36] by $3.0\\%$ , demonstrating the effectiveness of our method.", + "bbox": [ + 496, + 337, + 890, + 397 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Table 2 provides a comparison of our approach with recent state-of-the-art models on MSR-VTT. Our method achieves new state-of-the-art performance on text-to-video retrieval for both ViT-B/32 and ViT-B/16 backbones, significantly surpassing previous works. For instance, we achieve a $+4.8\\%$ higher R@1 than CLIP4Clip with the same ViT-B/32 on text-to-video retrieval. Additionally, our Cap4Video outperforms the recent TS2-Net [22] by $2.3\\%$ and $2.0\\%$ with ViT-B/32 and ViT-B/16, respectively.", + "bbox": [ + 496, + 398, + 890, + 532 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Table 3 and Table 4 show the results for the MSVD and VATEX datasets, respectively, where we use ViT-B/16 as our backbone. For MSVD, our Cap4Video achieves a remarkable performance of $51.8\\%$ R@1 and outperforms CLIP-based models CLIP4Clip [24] and X-Pool [25] by $6.6\\%$ and $4.6\\%$ on text-to-video retrieval, respectively. For VATEX, our approach also outperforms the recent state-of-the-art methods and achieves a $+7.5\\%$ R@1 improvement over TS2-Net [22] for text-to-video retrieval.", + "bbox": [ + 496, + 534, + 890, + 669 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In recent studies, several methods have been proposed", + "bbox": [ + 517, + 670, + 890, + 686 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/4bd930cfc7000be07d882d78e1de9a298570d6896ae8c377a7445757bff13a1a.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodR@1R@5R@10MdRMnR
CE [21]15.640.9-8.2-
ClipBERT [19]21.147.361.16.3-
Frozen [2]31.059.872.43.0-
TMVM [20]36.564.975.43.0-
CLIP4Clip [24]42.868.579.22.018.9
TS2-Net [22]41.871.682.02.014.8
HunYuan [28]45.075.683.42.012.0
DRL [36]49.076.584.52.0-
Cap4Video52.079.487.5110.5
", + "bbox": [ + 519, + 709, + 874, + 875 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Table 1. Results of text-to-video retrieval on the DiDeMo [1].", + "bbox": [ + 511, + 882, + 877, + 896 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "10708", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/da0da4daa866be38ac7bc836166b2349a80dbaef6de2b9f1fdea7b85ad6942fc.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodVenueText → VideoVideo → Text
R@1R@5R@10MdR↓MnR↓R@1R@5R@10MdR↓MnR↓
ClipBERT [19]CVPR'2022.046.859.96.0-----
MMT [10]ECCV'2026.657.169.64.0-27.057.569.73.721.3
T2VLAD [39]CVPR'2129.559.070.14.0-31.860.071.13.0
SupportSet [29]ICLR'2130.158.569.33.0-28.558.671.63.0-
Frozen [2]ICCV'2132.561.571.23.0------
BridgeFormer [12]CVPR'2237.664.875.1-------
TMVM [20]NeurIPS'2236.264.275.73.0-34.863.873.73.0-
CLIP-ViT-B/32
CLIP4Clip [24]arXiv'2144.571.481.62.015.342.770.980.62.011.6
CenterCLIP [48]SIGIR'2244.271.682.12.015.142.871.782.22.010.9
CAMoE [7]arXiv'2144.672.681.82.013.345.172.483.12.010.0
CLIP2Video [9]arXiv'2145.672.681.72.014.643.572.382.12.010.2
X-Pool [13]CVPR'2246.972.882.22.014.3-----
QB-Norm [4]CVPR'2247.273.083.02.0------
TS2-Net [22]ECCV'2247.074.583.82.013.045.374.183.72.09.2
DRL [36]arXiv'2247.474.683.82.0-45.373.983.32.0-
Cap4Video49.374.383.82.012.047.173.784.32.08.7
CLIP-ViT-B/16
CLIP2TV [11]arXiv'2148.374.682.82.014.946.575.484.92.010.2
CenterCLIP [48]SIGIR'2248.473.882.02.013.847.775.083.32.010.2
TS2-Net [22]ECCV'2249.475.685.32.013.546.675.984.92.08.9
DRL [36]arXiv'2250.276.584.71.0-48.976.385.42.0-
Cap4Video51.475.783.91.012.449.075.285.02.08
", + "bbox": [ + 83, + 88, + 885, + 474 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/691f1af6f839f6ab3d929e79bd1096144ce883588db3752e1a5dfcbe071bdab8.jpg", + "table_caption": [ + "Table 2. Retrieval results on the validation set of MSR-VTT 1K [44]. Here we report results without any post-processing operations (e.g., DSL [7] or QB-Norm [4]) during inference." + ], + "table_footnote": [], + "table_body": "
MethodR@1R@5R@10MdRMnR
CE [21]19.849.063.86.0-
SUPPORT [29]28.460.072.94.0-
CLIP [30]37.064.173.83.0-
Frozen [2]33.764.776.33.0-
TMVM [20]36.767.481.32.5-
CLIP4Clip [24]45.275.584.32.010.3
X-Pool [13]47.277.486.02.09.3
Cap4Video51.880.888.318.3
", + "bbox": [ + 96, + 532, + 450, + 686 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/b765fce0e6561e4b022bdd2e5d008c688fdd3c839dd6e5eacab29c777fd31df3.jpg", + "table_caption": [ + "Table 3. Results of text-to-video retrieval on the MSVD [43]." + ], + "table_footnote": [], + "table_body": "
MethodR@1R@5R@10MdRMnR
HGR [6]35.173.583.52.0-
CLIP [30]39.772.382.22.012.8
SUPPORT [29]44.982.189.71.0-
CLIP4Clip [24]55.989.295.01.03.9
Clip2Video [9]57.390.095.51.03.6
QB-Norm [4]58.888.393.81.0-
TS2-Net [22]59.190.095.21.03.5
Cap4Video66.693.197.012.7
", + "bbox": [ + 519, + 532, + 874, + 686 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 4. Results of text-to-video retrieval on the VATEX [38].", + "bbox": [ + 511, + 691, + 879, + 705 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "to improve text-video retrieval performance by adjusting similarity during inference using other query information. Notably, our results adheres to the standard retrieval logic, where the most relevant video is retrieved for each query from a set of videos, without any knowledge of the relationship between other queries and videos. Therefore, all results reported in our tables do not involve any post-processing procedures such as DSL [7] and QB-Norm [4]. Overall, the consistent state-of-the-art performance across four benchmarks demonstrates the effectiveness of our Cap4Video.", + "bbox": [ + 75, + 750, + 472, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.3. Ablation Study", + "text_level": 1, + "bbox": [ + 500, + 737, + 653, + 753 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this section, we provide detailed ablation studies to clarify the effects of each part of our design.", + "bbox": [ + 496, + 762, + 890, + 792 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Auxiliary Caption as Data Augmentation. We begin by investigating the impact of captions on data augmentation for training. In a real-world scenario, the original video title would naturally serve as an additional auxiliary caption. Therefore, we manually extracted the title from the video's original webpage and compared it to the caption generated by the GPT-2 model. Table 5 presents the re", + "bbox": [ + 496, + 795, + 890, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "10709", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/4b670fcfa9ce1435c2733532007b621c3174a1d92141fea3e67fd6a7aecb5541.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodGlobal MatchingFine-grained Matching
R@1R@5R@10MdR↓MnR↓R@1R@5R@10MdR↓MnR↓
Baseline42.870.479.0216.645.773.782.6213.1
+Different Sources of Caption as Data Augmentation
Video Title from Source URL43.871.180.9215.144.372.783.5213.1
Zero-shot Video Captioning44.270.781.5216.246.372.581.7212.9
+Different Number of Captions for Data Augmentation
Top-144.270.781.5216.246.372.581.7212.9
Top-343.371.781.6215.045.573.882.4212.7
Top-543.470.680.4216.245.672.782.7212.9
+Video-Caption Feature Interaction
Video Only44.270.781.5216.246.372.581.7212.9
Sum43.871.580.3216.147.273.382.8213.1
Concat-MLP37.566.178.4315.740.068.779.9212.7
Cross Transformer44.671.680.3214.647.975.483.0211.5
Co-attention Transformer45.371.280.9215.048.574.082.5212.7
+Query-Caption Matching Score
Query-Video Only45.371.280.9215.048.574.082.5212.7
Query-Caption Only30.355.267.5426.430.355.267.5426.4
Query-Video + Query-Caption45.671.781.2214.849.374.283.4212.1
", + "bbox": [ + 96, + 88, + 870, + 392 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "sults of using different sources of the caption, from which we observe that using captions generated by the web-scale model as data augmentation for training can lead to direct improvements in R@1 $(+1.4\\%, +0.6\\%)$ under both matching mechanisms. Using video titles can also bring a $1\\%$ improvement for global matching.", + "bbox": [ + 75, + 468, + 468, + 559 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We also explore the impact of the number of generated captions used for augmentation. We used the caption filtering mechanism mentioned in Sec. 2.3 to rank the relevance of captions and the ground-truth query, and selected different numbers of captions for training. The results demonstrated that using only one caption is sufficient.", + "bbox": [ + 75, + 561, + 468, + 652 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Benefit on Both Online and Offline Videos Scenarios. Our method is applicable for both online and offline videos. Offline videos are local videos with no title, while online videos have a title on the video website. Therefore, for online videos, the step of generating a caption with $\\mathrm{CLIP + GPT - 2}$ is skipped, and the website title is used directly as a caption. In Table 6, the results show that our method has significantly improved over the global matching baseline for both offline and online videos.", + "bbox": [ + 75, + 655, + 468, + 792 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Video-Caption Feature Interaction. As mentioned in Sec. 2.4, we have designed four approaches for Video-Caption feature interaction. Based on the results presented in Table 5, we can conclude the following: 1) The basic approach of Sum has been shown to enhance fine-grained matching by $0.9\\%$ R@1, but there is no noticeable improvement in global matching. 2) The MLP approach is difficult", + "bbox": [ + 75, + 794, + 468, + 902 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/1646cdf3574116cf8200a3a005a5bfbfaa8e9e5790f23bec7799911cfaa6806e.jpg", + "table_caption": [ + "Table 5. Component-wise evaluation of our framework on the MSR-VTT 1K validation set. With the ViT-B/32 backbone, we report the text-to-video retrieval results for two representative Query-Video matching mechanisms: global matching and fine-grained matching. The consistent improvement on two typical matching mechanisms demonstrates the generalization ability and effectiveness of our method." + ], + "table_footnote": [], + "table_body": "
Caption SourceR@1
BaselineN/A42.8
Cap4VideoOriginal Website Title45.8
Captioner (CLIP+GPT-2)45.6
", + "bbox": [ + 535, + 465, + 857, + 545 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 6. Exploring the effectiveness of captions from different sources on MSR-VTT 1k-A. Setting: ViT-B/32, global matching.", + "bbox": [ + 498, + 551, + 890, + 582 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "to optimize and performs poorly in both matching scenarios. We speculate that the MLP's operation in a black-box environment, despite creating a nonlinear metric space, may lead to degradation. 3) The Cross Transformer approach has demonstrated improvements of $+0.4\\%$ and $+1.6\\%$ in two matching settings, respectively. These enhancements may be attributed to the self-attention mechanism's ability to capture the inter-modal relationship between the video and caption. 4) Moreover, the Co-attention Transformer approach has significantly boosted performance, with gains of $+1.1\\%$ and $+2.2\\%$ for these two matching mechanisms. In summary, the results indicate that proper interaction between the video and generated caption can lead to better video representation and improved Query-Video matching.", + "bbox": [ + 496, + 611, + 890, + 823 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Query-Caption matching. We also investigate the Query-Caption matching branch for text-video retrieval. We aggregate caption embeddings using mean pooling to yield a global embedding. As shown in Table 5, the single Query-Caption matching branch achieves $30.3\\%$ R@1 on text-to", + "bbox": [ + 496, + 824, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "10710", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/bd0ea46b7b87631888f0581a4b24ae9bd771a918ab530b6e7aa7960641ff8350.jpg", + "table_caption": [ + "Query7765: a person is discussing a car." + ], + "table_footnote": [], + "table_body": "
VideoRank+ CaptionRank
6video of a car camera recording the driver's voice.1
4video showing the car in a parking spot.2
5video of SUV in the video below shows a salesman talking to an audience.3
", + "bbox": [ + 80, + 102, + 467, + 229 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/05d563cb803c0f28752c22ac4142287e6975054ea8ebe135b6ddb2fa8390fa14.jpg", + "table_caption": [ + "Query9616: person is recording the brown horse which is having fun." + ], + "table_footnote": [], + "table_body": "
VideoRank 2+ Caption video of the horse jumping over a fence at Ranch in Nevada was captured on camera.Rank 1
1video showing animation of a horse's simulation, which simulates the game.2
4video showing a horse simulation video game in which you could see your avatar being animated by the camera.3
", + "bbox": [ + 80, + 244, + 467, + 378 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Figure 4. The text-video results on the MSR-VTT 1K-A test set.", + "bbox": [ + 75, + 391, + 467, + 405 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Left: The ranking results of the query-video matching model.", + "bbox": [ + 76, + 406, + 467, + 419 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Right: The ranking results of Cap4Video, which incorporates generated captions to enhance retrieval. Please zoom in for best view.", + "bbox": [ + 76, + 420, + 467, + 446 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "video retrieval, outperforming previous query-video matching methods such as ClipBERT [19] (22.0%) and MMT [10] (26.6%). This suggests that the Query-Caption matching branch can complement the regular Query-Video matching branch for improved text-video retrieval. Combining the score of Query-Caption matching branch with Query-Video matching branch further improves performance (+0.8%).", + "bbox": [ + 75, + 474, + 467, + 580 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Overall, Cap4Video utilizes generated captions in three ways: input data augmentation, intermediate feature interaction, and output score fusion, leading to consistent improvements $(+2.8\\% / +3.6\\%)$ in both matching mechanisms.", + "bbox": [ + 75, + 582, + 467, + 642 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "3.4. Visualization", + "text_level": 1, + "bbox": [ + 76, + 654, + 212, + 667 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We provide two examples of videos retrieved by our Cap4Video method and a model without auxiliary captions. As illustrated in Figure 4, our approach successfully retrieves the ground-truth video with the assistance of the caption, while the video-only model returns multiple videos that are somewhat relevant to the query but not precise. See more qualitative results in Supplementary.", + "bbox": [ + 75, + 676, + 467, + 782 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4. Related Works", + "text_level": 1, + "bbox": [ + 76, + 799, + 225, + 814 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Zero-shot Image Captioning. In the field of natural language processing, transformer-based GPT models [5, 31] have been successful in generating text from prompts by training on large-scale text corpora. Similarly, CLIP [30], a vision-language alignment model trained on 400 million", + "bbox": [ + 75, + 824, + 467, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "image-text pairs, has demonstrated impressive zero-shot performance on vision tasks. However, research on transferring web-scale models to zero-shot video captioning remains limited. Recently, ZeroCap [34] proposed a method of using CLIP and the GPT-2 language model to generate textual descriptions of input images, leveraging knowledge from both models in a truly zero-shot manner without retraining or fine-tuning model parameters. MAGIC [32] has also used CLIP scores to align GPT-2 logits with corresponding images but requires fine-tuning on the MS-COCO caption text corpus. More recently, a study [33] extended the zero-shot capability of ZeroCap to the video domain. In this paper, we employ this video extension to generate auxiliary captions without any additional training.", + "bbox": [ + 496, + 90, + 890, + 301 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Text-Video Retrieval aims to retrieve relevant video content based on natural language descriptions. Early studies [6, 10, 21, 37, 39] focused on knowledge transfer from \"expert\" models and captured intra-modal and cross-modal interactions based on pre-extracted features. However, the performance of these methods is limited since they cannot perform end-to-end optimization. Recently, more methods have involved end-to-end training for text-video retrieval. One typical approach [2, 26, 27] is to first perform largescale text-video pre-training, then transfer the model to downstream text-video retrieval tasks. With the emergence of pre-trained Vision-Language Models (VLMs), there have been increased efforts to leverage them for improving video understanding [41, 42]. Thus, another training-efficient line is to directly expand the pre-trained VLM to the text-video retrieval task. CLIPBERT [19] enables affordable pioneering end-to-end training with a sparse sampling strategy. After that, recent works [3, 4, 8, 9, 11, 13, 22, 24, 40, 48] focus on transferring knowledge from CLIP models that have been pre-trained on 400M image-text pairs. The research path has evolved from the most direct global matching (i.e., video-sentence alignment [11, 24]) to fine-grained matching (e.g., frame-word alignment [36], video-word alignment [13], multi-hierarchical alignment [9, 28]). Unlike above CLIP-Based efforts on query-video matching, we propose to generate auxiliary captions from videos to improve text-video retrieval. Thus our method is compatible with both global and fine-grained matching.", + "bbox": [ + 496, + 303, + 890, + 724 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 739, + 617, + 753 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We introduce Cap4Video, a novel framework that leverages captions generated by web-scale language models to enhance text-video matching in three key ways: 1) Input data augmentation for training, 2) Intermediate videocaption feature interaction for compact video representations, and 3) Output score fusion for improved text-video retrieval. Our approach demonstrates consistent performance gains on four standard text-video retrieval benchmarks, outperforming state-of-the-art methods by a clear margin.", + "bbox": [ + 496, + 763, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "10711", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Lisa Anne Hendricks, Oliver Wang, Eli Shechtman, Josef Sivic, Trevor Darrell, and Bryan Russell. Localizing moments in video with natural language. In ICCV, pages 5803-5812, 2017. 2, 5", + "[2] Max Bain, Arsha Nagrani, Gül Varol, and Andrew Zisserman. Frozen in time: A joint video and image encoder for end-to-end retrieval. In ICCV, pages 1728-1738, 2021. 5, 6, 8", + "[3] Max Bain, Arsha Nagrani, Gül Varol, and Andrew Zisserman. A clip-hitchhiker's guide to long video retrieval. arXiv preprint arXiv:2205.08508, 2022. 8", + "[4] Simion-Vlad Bogolin, Ioana Croitoru, Hailin Jin, Yang Liu, and Samuel Albanie. Cross modal retrieval with querybank normalisation. In CVPR, pages 5194-5205, 2022. 6, 8", + "[5] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. NeurIPS, 33:1877-1901, 2020. 8", + "[6] Shizhe Chen, Yida Zhao, Qin Jin, and Qi Wu. Fine-grained video-text retrieval with hierarchical graph reasoning. In CVPR, pages 10638-10647, 2020. 6, 8", + "[7] Xing Cheng, Hezheng Lin, Xiangyu Wu, Fan Yang, and Dong Shen. Improving video-text retrieval by multi-stream corpus alignment and dual softmax loss. arXiv preprint arXiv:2109.04290, 2021. 6", + "[8] Bo Fang, Chang Liu, Yu Zhou, Min Yang, Yuxin Song, Fu Li, Weiping Wang, Xiangyang Ji, Wanli Ouyang, et al. Uatvr: Uncertainty-adaptive text-video retrieval. arXiv preprint arXiv:2301.06309, 2023. 8", + "[9] Han Fang, Pengfei Xiong, Luhui Xu, and Yu Chen. Clip2video: Mastering video-text retrieval via image clip. arXiv preprint arXiv:2106.11097, 2021. 1, 2, 6, 8", + "[10] Valentin Gabeur, Chen Sun, Karteek Alahari, and Cordelia Schmid. Multi-modal transformer for video retrieval. In ECCV, pages 214-229. Springer, 2020. 5, 6, 8", + "[11] Zijian Gao, Jingyu Liu, Sheng Chen, Dedan Chang, Hao Zhang, and Jinwei Yuan. Clip2tv: An empirical study on transformer-based methods for video-text retrieval. arXiv preprint arXiv:2111.05610, 2021. 1, 2, 6, 8", + "[12] Yuying Ge, Yixiao Ge, Xihui Liu, Dian Li, Ying Shan, Xiaohu Qie, and Ping Luo. Bridging video-text retrieval with multiple choice questions. In CVPR, pages 16167-16176, 2022. 6", + "[13] Satya Krishna Gorti, Noel Vouitsis, Junwei Ma, Keyvan Golestan, Maksims Volkovs, Animesh Garg, and Guangwei Yu. X-pool: Cross-modal language-video attention for text-video retrieval. In CVPR, pages 5006-5015, 2022. 1, 6, 8", + "[14] Priya Goyal, Piotr Dólar, Ross Girshick, Pieter Noordhuis, Lukasz Wesolowski, Aapo Kyrola, Andrew Tulloch, Yangqing Jia, and Kaiming He. Accurate, large minibatch sgd: Training imagenet in 1 hour. arXiv preprint arXiv:1706.02677, 2017. 5", + "[15] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom" + ], + "bbox": [ + 78, + 114, + 470, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In ICML, pages 4904-4916. PMLR, 2021. 1", + "[16] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In ICML, pages 4904-4916. PMLR, 2021. 2", + "[17] Omar Khattab and Matei Zaharia. Colbert: Efficient and effective passage search via contextualized late interaction over bert. In Proceedings of the 43rd International ACM SIGIR conference on research and development in Information Retrieval, pages 39-48, 2020. 2", + "[18] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In ICLR, 2015. 5", + "[19] Jie Lei, Linjie Li, Luowei Zhou, Zhe Gan, Tamara L Berg, Mohit Bansal, and Jingjing Liu. Less is more: Clipbert for video-and-language learning via sparse sampling. In CVPR, pages 7331-7341, 2021. 5, 6, 8", + "[20] Chengzhi Lin, Ancong Wu, Junwei Liang, Jun Zhang, Wenhang Ge, Wei-Shi Zheng, and Chunhua Shen. Text-adaptive multiple visual prototype matching for video-text retrieval. In NeurIPS, 2022. 5, 6", + "[21] Yang Liu, Samuel Albanie, Arsha Nagrani, and Andrew Zisserman. Use what you have: Video retrieval using representations from collaborative experts. arXiv preprint arXiv:1907.13487, 2019. 5, 6, 8", + "[22] Yuqi Liu, Pengfei Xiong, Luhui Xu, Shengming Cao, and Qin Jin. Ts2-net: Token shift and selection transformer for text-video retrieval. In ECCV, pages 319-335. Springer, 2022. 5, 6, 8", + "[23] Jiasen Lu, Dhruv Batra, Devi Parikh, and Stefan Lee. Vilbert: Pretraining task-agnostic visiolinguistic representations for vision-and-language tasks. NeurIPS, 32, 2019. 4", + "[24] Huaishao Luo, Lei Ji, Ming Zhong, Yang Chen, Wen Lei, Nan Duan, and Tianrui Li. Clip4clip: An empirical study of clip for end to end video clip retrieval and captioning. Neurocomputing, 508:293-304, 2022. 1, 2, 4, 5, 6, 8", + "[25] Yiwei Ma, Guohai Xu, Xiaoshuai Sun, Ming Yan, Ji Zhang, and Rongrong Ji. X-clip: End-to-end multi-grained contrastive learning for video-text retrieval. In ACM MM, pages 638–647, 2022. 5", + "[26] Antoine Miech, Jean-Baptiste Alayrac, Lucas Smaira, Ivan Laptev, Josef Sivic, and Andrew Zisserman. End-to-end learning of visual representations from uncurated instructional videos. In CVPR, pages 9879-9889, 2020. 8", + "[27] Antoine Miech, Dimitri Zhukov, Jean-Baptiste Alayrac, Makarand Tapaswi, Ivan Laptev, and Josef Sivic. Howto100m: Learning a text-video embedding by watching hundred million narrated video clips. In ICCV, pages 2630-2640, 2019. 5, 8", + "[28] Shaobo Min, Weijie Kong, Rong-Cheng Tu, Dihong Gong, Chengfei Cai, Wenzhe Zhao, Chenyang Liu, Sixiao Zheng, Hongfa Wang, Zhifeng Li, et al. Hunyuan_tvr for text-video retrivial. arXiv preprint arXiv:2204.03382, 2022. 1, 5, 8", + "[29] Mandela Patrick, Po-Yao Huang, Yuki Asano, Florian Metze, Alexander G Hauptmann, Joao F Henriques, and An" + ], + "bbox": [ + 503, + 92, + 890, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "10712", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "drea Vedaldi. Support-set bottlenecks for video-text representation learning. In ICLR, 2021. 6", + "[30] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, pages 8748-8763. PMLR, 2021. 1, 2, 3, 5, 6, 8", + "[31] Alec Radford, Jeff Wu, Rewon Child, David Luan, Dario Amodei, and Ilya Sutskever. Language models are unsupervised multitask learners. 2019. 1, 2, 3, 8", + "[32] Yixuan Su, Tian Lan, Yahui Liu, Fangyu Liu, Dani Yogatama, Yan Wang, Lingpeng Kong, and Nigel Collier. Language models can see: Plugging visual controls in text generation. arXiv preprint arXiv:2205.02655, 2022. 8", + "[33] Yoad Tewel, Yoav Shalev, Roy Nadler, Idan Schwartz, and Lior Wolf. Zero-shot video captioning with evolving pseudotokens. arXiv preprint arXiv:2207.11100, 2022. 2, 3, 8", + "[34] Yoad Tewel, Yoav Shalev, Idan Schwartz, and Lior Wolf. Zerocap: Zero-shot image-to-text generation for visual-semantic arithmetic. In CVPR, pages 17918-17928, 2022. 2, 3, 8", + "[35] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In NeurIPS, pages 5998-6008, 2017. 4", + "[36] Qiang Wang, Yanhao Zhang, Yun Zheng, Pan Pan, and Xian-Sheng Hua. Disentangled representation learning for text-video retrieval. arXiv preprint arXiv:2203.07111, 2022. 1, 2, 4, 5, 6, 8", + "[37] Wenzhe Wang, Mengdan Zhang, Runnan Chen, Guanyu Cai, Penghao Zhou, Pai Peng, Xiaowei Guo, Jian Wu, and Xing Sun. Dig into multi-modal cues for video retrieval with hierarchical alignment. In *IJCAI*, pages 1113–1121, 2021. 8", + "[38] Xin Wang, Jiawei Wu, Junkun Chen, Lei Li, Yuan-Fang Wang, and William Yang Wang. Vatex: A large-scale, high-quality multilingual dataset for video-and-language research. In ICCV, pages 4581-4591, 2019. 2, 5, 6", + "[39] Xiaohan Wang, Linchao Zhu, and Yi Yang. T2vlad: global-local sequence alignment for text-video retrieval. In CVPR, pages 5079-5088, 2021. 6, 8", + "[40] Xiaohan Wang, Linchao Zhu, Zhedong Zheng, Mingliang Xu, and Yi Yang. Align and tell: Boosting text-video retrieval with local alignment and fine-grained supervision. IEEE TMM, 2022. 8", + "[41] Wenhao Wu, Zhun Sun, and Wanli Ouyang. Revisiting classifier: Transferring vision-language models for video recognition. In AAAI, 2023. 8", + "[42] Wenhao Wu, Xiaohan Wang, Haipeng Luo, Jingdong Wang, Yi Yang, and Wanli Ouyang. Bidirectional cross-modal knowledge exploration for video recognition with pretrained vision-language models. In CVPR, 2023. 8", + "[43] Zuxuan Wu, Ting Yao, Yanwei Fu, and Yu-Gang Jiang. Deep learning for video classification and captioning. In Frontiers of multimedia research, pages 3-29. 2017. 2, 5, 6", + "[44] Jun Xu, Tao Mei, Ting Yao, and Yong Rui. Msr-vtt: A large video description dataset for bridging video and language. In CVPR, pages 5288-5296, 2016. 2, 5, 6" + ], + "bbox": [ + 78, + 90, + 468, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[45] Lewei Yao, Runhui Huang, Lu Hou, Guansong Lu, Minzhe Niu, Hang Xu, Xiaodan Liang, Zhenguo Li, Xin Jiang, and Chunjing Xu. Filip: Fine-grained interactive language-image pre-training. In ICLR, 2021. 2", + "[46] Jiahui Yu, Zirui Wang, Vijay Vasudevan, Legg Yeung, Mojtaba Seyhedhosseini, and Yonghui Wu. Coca: Contrastive captioners are image-text foundation models. arXiv preprint arXiv:2205.01917, 2022. 1", + "[47] Lu Yuan, Dongdong Chen, Yi-Ling Chen, Noel Codella, Xiyang Dai, Jianfeng Gao, Houdong Hu, Xuedong Huang, Boxin Li, Chunyuan Li, et al. Florence: A new foundation model for computer vision. arXiv preprint arXiv:2111.11432, 2021. 1", + "[48] Shuai Zhao, Linchao Zhu, Xiaohan Wang, and Yi Yang. Centerclip: Token clustering for efficient text-video retrieval. In Proceedings of the 45th International ACM SIGIR Conference on Research and Development in Information Retrieval, pages 970–981, 2022. 6, 8" + ], + "bbox": [ + 501, + 92, + 890, + 344 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10713", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2023/Cap4Video_ What Can Auxiliary Captions Do for Text-Video Retrieval_/926c868d-a104-48e5-bcb0-c1285d26ad5d_model.json b/2023/Cap4Video_ What Can Auxiliary Captions Do for Text-Video Retrieval_/926c868d-a104-48e5-bcb0-c1285d26ad5d_model.json new file mode 100644 index 0000000000000000000000000000000000000000..243223768d128e2f50646aff8ebaeee82b3ab086 --- /dev/null +++ b/2023/Cap4Video_ What Can Auxiliary Captions Do for Text-Video Retrieval_/926c868d-a104-48e5-bcb0-c1285d26ad5d_model.json @@ -0,0 +1,2233 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.123, + 0.131, + 0.848, + 0.153 + ], + "angle": 0, + "content": "Cap4Video: What Can Auxiliary Captions Do for Text-Video Retrieval?" + }, + { + "type": "text", + "bbox": [ + 0.12, + 0.179, + 0.848, + 0.252 + ], + "angle": 0, + "content": "Wenhao Wu\\(^{1,2*}\\) Haipeng Luo\\(^{3*}\\) Bo Fang\\(^{3}\\) Jingdong Wang\\(^{2}\\) Wanli Ouyang\\(^{4,1}\\) \n\\(^{1}\\)The University of Sydney \n\\(^{2}\\)Baidu Inc. \n\\(^{3}\\)University of Chinese Academy of Sciences \n\\(^{4}\\)Shanghai AI Laboratory \nwhwu.ucas@gmail.com" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.287, + 0.314, + 0.303 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.319, + 0.474, + 0.729 + ], + "angle": 0, + "content": "Most existing text-video retrieval methods focus on cross-modal matching between the visual content of videos and textual query sentences. However, in real-world scenarios, online videos are often accompanied by relevant text information such as titles, tags, and even subtitles, which can be utilized to match textual queries. This insight has motivated us to propose a novel approach to text-video retrieval, where we directly generate associated captions from videos using zero-shot video captioning with knowledge from web-scale pre-trained models (e.g., CLIP and GPT-2). Given the generated captions, a natural question arises: what benefits do they bring to text-video retrieval? To answer this, we introduce Cap4Video, a new framework that leverages captions in three ways: i) Input data: video-caption pairs can augment the training data. ii) Intermediate feature interaction: we perform cross-modal feature interaction between the video and caption to produce enhanced video representations. iii) Output score: the Query-Caption matching branch can complement the original Query-Video matching branch for text-video retrieval. We conduct comprehensive ablation studies to demonstrate the effectiveness of our approach. Without any post-processing, Cap4Video achieves state-of-the-art performance on four standard text-video retrieval benchmarks: MSR-VTT \\((51.4\\%)\\), VATEX \\((66.6\\%)\\), MSVD \\((51.8\\%)\\), and DiDeMo \\((52.0\\%)\\). The code is available at https://github.com/whwu95/Cap4Video." + }, + { + "type": "title", + "bbox": [ + 0.079, + 0.758, + 0.21, + 0.773 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.784, + 0.47, + 0.876 + ], + "angle": 0, + "content": "Text-video retrieval is a fundamental task in videolanguage learning. With the rapid advancements in image-language pre-training [15, 30, 46, 47], researchers have focused on expanding pre-trained image-language models, especially CLIP [30], to tackle the text-video retrieval task. The research path has evolved from the most direct global" + }, + { + "type": "image", + "bbox": [ + 0.507, + 0.286, + 0.878, + 0.332 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.508, + 0.343, + 0.882, + 0.402 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.508, + 0.412, + 0.888, + 0.565 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.586, + 0.894, + 0.671 + ], + "angle": 0, + "content": "Figure 1. (a) An existing end-to-end learning paradigm for text-video retrieval. (b) Zero-shot video captioning achieved by guiding a large language model (LLM) such as GPT-2 [31] with CLIP [30]. (c) Our Cap4Video framework leverages the generated captions in three aspects: input data augmentation, intermediate feature interaction, and output score fusion." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.704, + 0.895, + 0.903 + ], + "angle": 0, + "content": "matching (i.e., video-sentence alignment [11, 24]) to fine-grained matching (e.g., frame-word alignment [36], videoword alignment [13], multi-hierarchical alignment [9, 28], etc.). These studies have demonstrated remarkable performance and significantly outperformed previous models. Two key factors contribute to this improvement. Firstly, CLIP offers powerful visual and textual representations that are pre-aligned in the semantic embedding space, thereby reducing the challenge of cross-modal learning in videotext matching. Secondly, these methods can fine-tune the pre-trained vision and text encoders using sparsely sampled frames in an end-to-end manner. All of these methods aim to learn cross-modal alignment between the visual represen" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.887, + 0.206, + 0.9 + ], + "angle": 0, + "content": "*Equal contribution." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "10704" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.121 + ], + "angle": 0, + "content": "tation of videos and the textual representation of the corresponding query, as depicted in Figure 1(a)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.122, + 0.473, + 0.409 + ], + "angle": 0, + "content": "However, in real-life scenarios, online videos usually come with related content such as the video's title or tag on the video website. In addition to the visual signal in the video, the associated textual information can also be used to some extent to describe the video content and match the query (i.e., the common text-to-text retrieval). This raises a pertinent question: How can we generate associated text descriptions for videos? One possible solution is to crawl the video title from the video website. However, this method relies on annotations, and there is a risk that the video URL may have become invalid. Another automated solution is to generate captions using zero-shot video caption models. Therefore, we turn our attention to knowledge-rich pretrained models to handle such challenging open-set scenarios. We find that the recent study ZeroCap [34] provides a good practice to use frozen CLIP [30] and GPT-2 [31] for zero-shot image captioning. Thus, we leverage a video extension [33] of ZeroCap for generating captions in the video domain without any further training." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.41, + 0.473, + 0.727 + ], + "angle": 0, + "content": "When provided with auxiliary captions, a natural question naturally arises: How can we leverage these captions to enhance the text-video retrieval task? In this paper, we propose the Cap4Video learning framework, as illustrated in Figure 1(c), which utilizes captions in three key ways: (i) Input Data: One simple approach is to augment the training data with the generated captions. Specifically, the given video and its generated caption can be treated as a matched pair, which serves as an additional positive sample pair for training beyond the query-video pairs. (ii) Intermediate Feature Interaction: Cross-modal interaction between the video and captions can be leveraged to improve the video representation. Specifically, we can exploit the complementary information between videos and captions to reduce redundant features from videos and learn more discriminative video representations. (iii) Output score: The generated caption can also represent the video's content, allowing us to employ query-caption matching to complement standard query-video matching for the text-video retrieval task. Moreover, a two-stream architecture can be utilized to reduce model bias and produce more robust results." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.727, + 0.472, + 0.772 + ], + "angle": 0, + "content": "We hope that our novel paradigm will encourage further investigation into the video-language learning. In summary, our contributions are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.784, + 0.47, + 0.874 + ], + "angle": 0, + "content": "- We explore a novel problem: leveraging auxiliary captions to further enhance existing text-video retrieval. Besides labor-intensive manual crawling of video website titles, we investigate the potential of rich captions automatically generated by large language models (LLMs) to benefit text-video retrieval." + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.886, + 0.47, + 0.902 + ], + "angle": 0, + "content": "- We propose the Cap4Video learning framework," + }, + { + "type": "list", + "bbox": [ + 0.096, + 0.784, + 0.47, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.531, + 0.092, + 0.892, + 0.169 + ], + "angle": 0, + "content": "which maximizes the utility of the auxiliary captions through three aspects: input data, feature interaction, and output score. Our framework improves the performance of existing query-video matching mechanisms, including global matching and fine-grained matching." + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.176, + 0.894, + 0.252 + ], + "angle": 0, + "content": "- Extensive experiments conducted on four video benchmarks demonstrate the effectiveness of our method. Our Cap4Video achieves state-of-the-art performance on MSR-VTT [44] (51.4%), VATEX [38] (66.6%), MSVD [43] (51.8%), and DiDeMo [1] (52.0%)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.263, + 0.637, + 0.281 + ], + "angle": 0, + "content": "2. Methodology" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.288, + 0.807, + 0.305 + ], + "angle": 0, + "content": "2.1. Background: Text-Video Matching" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.311, + 0.895, + 0.462 + ], + "angle": 0, + "content": "Text-video matching aims to evaluate the similarity between a given sentence \\( Q_{i} \\) and a given video \\( V_{j} \\), typically using a similarity function \\( s(Q_{i},V_{j}) \\). In text-to-video retrieval, the goal is to rank all videos based on their similarity scores to a given query sentence. To improve text-video retrieval, recent works [9, 11, 24] have applied CLIP [30] for initialization, leveraging pre-trained knowledge from image-text learning. Our baselines for text-video matching include two typical mechanisms: global matching and fine-grained matching, as illustrated in Figure 2(b)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.463, + 0.895, + 0.644 + ], + "angle": 0, + "content": "Global Matching is a commonly used technique in cross-modal contrastive learning [16, 24, 30]. In global matching, each modality is encoded independently to obtain global features, which are then used to calculate similarity. We train the visual encoder to output \\( F \\) frame embeddings for a given video that samples \\( F \\) frames. Similarly, the query encoder returns \\( W \\) word embeddings and the [CLS] embedding as the global representation for a given query sentence that contains \\( W \\) words. The frame embeddings are integrated using average pooling to obtain the global video embedding, which is then compared with the global query embedding to calculate similarity." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.644, + 0.895, + 0.84 + ], + "angle": 0, + "content": "Fine-grained Matching focuses on modeling the token-level alignment between two modalities, such as framework alignment. In order to achieve token-level patchword alignment for image-text learning, FILIP [45] and ColBERT [17] employ a Max-Mean pipeline. This pipeline finds the token-wise maximum similarity between patch and word tokens, and then averages the maximum similarity of tokens in the image or text to obtain the similarity between an image and a text or vice versa. Moreover, DRL [36] extends the token-wise alignment to text-video retrieval and introduces an attention mechanism to learn weighted pooling instead of mean pooling. We have adopted this mechanism as our enhanced baseline." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.848, + 0.81, + 0.865 + ], + "angle": 0, + "content": "2.2. Preprocessing: Caption Generation" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.892, + 0.902 + ], + "angle": 0, + "content": "To obtain auxiliary captions for a given video, we consider the following two approaches." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "10705" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.081, + 0.09, + 0.225, + 0.105 + ], + "angle": 0, + "content": "(a) Our Cap4Video" + }, + { + "type": "image", + "bbox": [ + 0.08, + 0.106, + 0.891, + 0.256 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.083, + 0.263, + 0.468, + 0.278 + ], + "angle": 0, + "content": "(b) Query-Video Matching: Two typical mechanisms" + }, + { + "type": "image_caption", + "bbox": [ + 0.082, + 0.281, + 0.281, + 0.298 + ], + "angle": 0, + "content": "① Global embedding matching" + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.303, + 0.326, + 0.404 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.331, + 0.281, + 0.566, + 0.297 + ], + "angle": 0, + "content": "② Fine-grained embedding matching" + }, + { + "type": "image", + "bbox": [ + 0.331, + 0.303, + 0.631, + 0.402 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.639, + 0.263, + 0.851, + 0.279 + ], + "angle": 0, + "content": "(c) Query-Caption Matching" + }, + { + "type": "image_caption", + "bbox": [ + 0.603, + 0.282, + 0.738, + 0.296 + ], + "angle": 0, + "content": "Captions embeddings" + }, + { + "type": "image", + "bbox": [ + 0.651, + 0.298, + 0.891, + 0.398 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.42, + 0.893, + 0.503 + ], + "angle": 0, + "content": "Figure 2. An overview of our Cap4Video for text-video retrieval. We first generate captions using a zero-shot video captioner that combines CLIP [30] with GPT-2 [31], leveraging knowledge from both frozen web-scale models. We then utilize the pre-extracted caption information from three different perspectives: i) Input data: We use the video and captions to create new positive pairs for data augmentation during training. ii) Feature interaction: We perform feature interaction between video and caption to capture intra- and inter-modality context, yielding enhanced video representations. iii) Output score: The Query-Caption matching branch can complement the original Query-Video matching branch for text-video retrieval." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.53, + 0.47, + 0.59 + ], + "angle": 0, + "content": "Manual Crawling of Video Titles. We extract the video website title by crawling the original links (such as YouTube ID) of each video and utilize it as the caption. However, we skip this step for videos with expired links." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.592, + 0.47, + 0.773 + ], + "angle": 0, + "content": "Automatic Video Captioning. In contrast to the manual approach that relies on annotations, we leverage knowledge from the LLM to generate rich and diverse captions. Given the scalability of our framework, we aim to generate captions directly from downstream videos without any additional training, a process known as zero-shot video captioning. To achieve this, we follow [33, 34] and use GPT-2 [31] to predict the next word from an initial prompt, e.g., \"Video shows\". A calibrated CLIP [30] loss is then used to drive the model to generate sentences that describe the video, incorporating video-related knowledge into the auto-regressive process. See Supplementary for more details." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.786, + 0.452, + 0.803 + ], + "angle": 0, + "content": "2.3. Data Augmentation with Auxiliary Captions" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.811, + 0.47, + 0.903 + ], + "angle": 0, + "content": "Auxiliary captions can be used to augment training data. For example, for a dataset consisting of \\( N \\) videos and their corresponding query sentences, each video and its generated caption can be considered as a positive sample pair for training, in addition to the original query-video pairs. By selecting one caption per video, we can add at least \\( N \\) pairs" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.531, + 0.816, + 0.545 + ], + "angle": 0, + "content": "as additional data augmentation during training." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.547, + 0.892, + 0.699 + ], + "angle": 0, + "content": "The automatic video captioner can generate multiple captions (e.g., 20) for each video. However, some of these captions may contain noise and may not be entirely relevant to the video content. To avoid negative effects on training, we use a filtering mechanism that evaluates the semantic similarity between each caption and the ground-truth query of the video using a pre-trained text encoder. The caption with the highest similarity is then chosen for data augmentation. Note that we only use the ground-truth query for caption filtering during the training phase." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.71, + 0.843, + 0.727 + ], + "angle": 0, + "content": "2.4. Video-Caption Cross-Modal Interaction" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.735, + 0.892, + 0.901 + ], + "angle": 0, + "content": "We further consider taking advantage of the complementarity between videos and captions to reduce redundant features and learn more discriminative video representations. To preserve the pre-trained CLIP encoder architecture for efficient transfer learning, we limit the interaction to the final caption and frame embeddings. Specifically, we pass the frame embeddings \\(\\mathbf{e}_{\\mathbf{v}} = \\{\\mathbf{v}_1,\\mathbf{v}_2,\\dots ,\\mathbf{v}_F\\}\\) and caption embeddings \\(\\mathbf{e_c} = \\{\\mathbf{c}_1,\\mathbf{c}_2,\\dots ,\\mathbf{c}_C\\}\\) to the interaction module, where \\(F\\) and \\(C\\) represent the number of frames and captions, respectively. Figure 2(b) depicts several ways of interaction between the two modalities." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "10706" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.47, + 0.138 + ], + "angle": 0, + "content": "Sum. To obtain an enhanced frame embedding, an intuitive approach is to compute the sum of the global caption embedding \\(\\mathbf{c}_{\\mathbf{g}}\\) and each frame embedding:" + }, + { + "type": "equation", + "bbox": [ + 0.127, + 0.147, + 0.469, + 0.163 + ], + "angle": 0, + "content": "\\[\n\\operatorname {S u m} \\left(\\mathbf {v} _ {i}, \\mathbf {c} _ {\\mathbf {g}}\\right) = \\mathbf {v} _ {i} + \\mathbf {c} _ {\\mathbf {g}}, \\quad i = 1, \\dots , F, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.172, + 0.469, + 0.219 + ], + "angle": 0, + "content": "where \\(\\mathbf{v}_i\\in \\mathbb{R}^D\\) is the \\(i\\)-th frame embedding, and \\(\\mathbf{c_g}\\in \\mathbb{R}^D\\) is computed by averaging the [CLS] embeddings of \\(C\\) generated captions: \\(\\mathbf{c_g} = \\frac{1}{C}\\sum_{i = 1}^{C}\\mathbf{c}_i\\)." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.219, + 0.469, + 0.278 + ], + "angle": 0, + "content": "MLP. To model weighted combinations of each frame embedding and the global caption embedding \\(\\mathbf{c_g}\\), we concatenate them together and pass the result through a learnable Multi-layer Perceptron (MLP):" + }, + { + "type": "equation", + "bbox": [ + 0.099, + 0.288, + 0.469, + 0.306 + ], + "angle": 0, + "content": "\\[\n\\operatorname {M L P} \\left(\\mathbf {v} _ {i}, \\mathbf {c} _ {\\mathbf {g}}\\right) = f _ {\\theta} \\left(\\left[ \\mathbf {v} _ {i}, \\mathbf {c} _ {\\mathbf {g}} \\right]\\right), \\quad \\text {f o r} i = 1, \\dots , F, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.314, + 0.469, + 0.344 + ], + "angle": 0, + "content": "where \\([\\cdot ,\\cdot ]\\) denotes the concatenation operation, \\(f_{\\theta}\\) is the MLP with parameter \\(\\theta\\)" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.344, + 0.469, + 0.419 + ], + "angle": 0, + "content": "Cross Transformer. We also investigate the use of selfattention [35] for interactions. The Cross Transformer operates on a sequence \\(\\{\\mathbf{e}_{\\mathbf{v}}, \\mathbf{e}_{\\mathbf{c}}\\} = \\{\\mathbf{v}_1, \\dots, \\mathbf{v}_F, \\mathbf{c}_1, \\dots, \\mathbf{c}_C\\}\\) and processes them through \\(L\\) encoder-style transformer blocks to generate final representations:" + }, + { + "type": "equation", + "bbox": [ + 0.167, + 0.43, + 0.469, + 0.446 + ], + "angle": 0, + "content": "\\[\n\\operatorname {C r o s s} \\left(\\mathbf {e} _ {\\mathbf {v}}, \\mathbf {e} _ {\\mathbf {c}}\\right) = f _ {\\psi} \\left(\\left\\{\\mathbf {e} _ {\\mathbf {v}}, \\mathbf {e} _ {\\mathbf {c}} \\right\\}\\right), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.455, + 0.468, + 0.485 + ], + "angle": 0, + "content": "where \\(\\{\\}\\) denotes that \\(\\mathbf{e_v}\\) and \\(\\mathbf{e_c}\\) form a sequence, and \\(f_{\\psi}\\) represents the transformer encoders with parameter \\(\\psi\\)" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.486, + 0.469, + 0.561 + ], + "angle": 0, + "content": "Co-attention Transformer. Co-attention [23] is another common method for exchanging information between modalities, allowing for mutual attention between video and caption. After this co-attentional transformer layer, we include \\(L\\) transformer layers to model temporal information:" + }, + { + "type": "equation", + "bbox": [ + 0.14, + 0.576, + 0.469, + 0.593 + ], + "angle": 0, + "content": "\\[\n\\operatorname {C o A t t n} \\left(\\mathbf {e} _ {\\mathbf {v}}, \\mathbf {e} _ {\\mathbf {c}}\\right) = f _ {\\phi_ {2}} \\left(f _ {\\phi_ {1}} \\left(\\left\\{\\mathbf {e} _ {\\mathbf {v}}, \\mathbf {e} _ {\\mathbf {c}} \\right\\}\\right)\\right), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.598, + 0.469, + 0.628 + ], + "angle": 0, + "content": "where \\(f_{\\phi_1}\\) is the co-attentional transformer with parameter \\(\\phi_1\\) and \\(f_{\\phi_2}\\) is the transformer encoders with parameter \\(\\phi_2\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.628, + 0.469, + 0.718 + ], + "angle": 0, + "content": "The video-caption interaction module generates frame embeddings that can be further processed based on the type of matching needed. For global matching, the frame embeddings can be averaged to obtain a single video representation. Alternatively, for fine-grained matching, the individual frame embeddings can be retained." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.727, + 0.438, + 0.743 + ], + "angle": 0, + "content": "2.5. Complementary Query-Caption Matching" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.47, + 0.901 + ], + "angle": 0, + "content": "Besides using the caption for data augmentation and video feature enhancement, it can also directly represent the video content, allowing for text-text retrieval. Specifically, each of the \\(C\\) captions generated by the video is then passed through the caption encoder to obtain its [CLS] text embedding. These caption embeddings are then aggregated to form a global representation, as illustrated in Figure 2(c). The cosine similarity between this global caption embedding and the global query embedding is then calculated to complement the query-video matching." + }, + { + "type": "image", + "bbox": [ + 0.513, + 0.088, + 0.631, + 0.159 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.549, + 0.162, + 0.601, + 0.175 + ], + "angle": 0, + "content": "(a) Sum" + }, + { + "type": "image", + "bbox": [ + 0.684, + 0.088, + 0.894, + 0.158 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.761, + 0.161, + 0.816, + 0.174 + ], + "angle": 0, + "content": "(b) MLP" + }, + { + "type": "image", + "bbox": [ + 0.513, + 0.184, + 0.65, + 0.271 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.273, + 0.649, + 0.285 + ], + "angle": 0, + "content": "(c) Cross Transformer" + }, + { + "type": "image", + "bbox": [ + 0.685, + 0.181, + 0.894, + 0.269 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.7, + 0.273, + 0.879, + 0.285 + ], + "angle": 0, + "content": "(d) Co-attention Transformer" + }, + { + "type": "image_caption", + "bbox": [ + 0.54, + 0.295, + 0.691, + 0.31 + ], + "angle": 0, + "content": "Caption Embedding" + }, + { + "type": "image_caption", + "bbox": [ + 0.716, + 0.295, + 0.859, + 0.309 + ], + "angle": 0, + "content": "Frame Embedding" + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.327, + 0.892, + 0.369 + ], + "angle": 0, + "content": "Figure 3. Illustration of four Video-Caption interaction strategies. The enhanced frame embeddings will be followed by a mean pooling for global matching or will remain for fine-grained matching." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.4, + 0.892, + 0.583 + ], + "angle": 0, + "content": "Notation. Let \\(\\{\\mathbf{e}_{\\mathbf{v}_i},\\mathbf{e}_{\\mathbf{t}_i},\\mathbf{e}_{\\mathbf{c}_i}\\}_{i = 1}^B\\) be a batch of \\(B\\) triples, where \\(\\mathbf{e}_{\\mathbf{v}_i}\\), \\(\\mathbf{e}_{\\mathbf{t}_i}\\), and \\(\\mathbf{e}_{\\mathbf{c}_i}\\) denote the \\(i\\)-th video, query, and caption embedding, respectively. Note that the term \"embedding\" used here is more general for convenience and can vary in meaning depending on the situation. For instance, in query-video global matching, \\(\\mathbf{e}_{\\mathbf{v}_i}\\) and \\(\\mathbf{e}_{\\mathbf{t}_i}\\) represent the averaged video feature and global [CLS] text feature, respectively. In query-video fine-grained matching, \\(\\mathbf{e}_{\\mathbf{v}_i}\\) and \\(\\mathbf{e}_{\\mathbf{t}_i}\\) represent a sequence of frame embeddings and a sequence of word embeddings, respectively. In query-caption matching, \\(\\mathbf{e}_{\\mathbf{c}_i}\\) represents a sequence of caption embeddings, and \\(\\mathbf{e}_{\\mathbf{t}_i}\\) represents a global [CLS] text feature." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.585, + 0.892, + 0.706 + ], + "angle": 0, + "content": "Learning Objectives. For the Query-Caption branch, we want the caption embedding \\(\\mathbf{e_c}\\) and the query embedding \\(\\mathbf{e_t}\\) to be close while they are related and far apart when they are not during training phase. We follow the common practice [24, 36] to consider the bidirectional learning objective. We employ symmetric cross-entropy loss to maximize the similarity between matched Query-Caption pairs and minimize the similarity for other pairs:" + }, + { + "type": "equation", + "bbox": [ + 0.522, + 0.723, + 0.848, + 0.765 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {Q 2 C} = - \\frac {1}{B} \\sum_ {i} ^ {B} \\log \\frac {\\exp \\left(s _ {q c} \\left(\\mathbf {e} _ {\\mathbf {t} i} , \\mathbf {e} _ {\\mathbf {c} i}\\right) / \\tau\\right)}{\\sum_ {j} ^ {B} \\exp \\left(s _ {q c} \\left(\\mathbf {e} _ {\\mathbf {t} i} , \\mathbf {e} _ {\\mathbf {c} j}\\right) / \\tau\\right)},\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.524, + 0.767, + 0.891, + 0.809 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {C 2 Q} = - \\frac {1}{B} \\sum_ {i} ^ {B} \\log \\frac {\\exp \\left(s _ {q c} \\left(\\mathbf {e} _ {\\mathbf {t} i} , \\mathbf {e} _ {\\mathbf {c} i}\\right) / \\tau\\right)}{\\sum_ {j} ^ {B} \\exp \\left(s _ {q c} \\left(\\mathbf {e} _ {\\mathbf {t} j} , \\mathbf {e} _ {\\mathbf {c} i}\\right) / \\tau\\right)}, \\tag {5}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.595, + 0.811, + 0.779, + 0.84 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {Q C} = \\frac {1}{2} \\left(\\mathcal {L} _ {Q 2 C} + \\mathcal {L} _ {C 2 Q}\\right),\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.522, + 0.856, + 0.892, + 0.902 + ], + "angle": 0, + "content": "where \\( s_{qc}(\\cdot, \\cdot) \\) represents the query-caption matching similarity function shown in Figure 2(c), and \\( \\tau \\) refers to the temperature hyper-parameter for scaling. Similarly, the con" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "10707" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.435, + 0.108 + ], + "angle": 0, + "content": "trastive loss for Query-Video branch is formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.098, + 0.116, + 0.47, + 0.234 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} _ {Q 2 V} = - \\frac {1}{B} \\sum_ {i} ^ {B} \\log \\frac {\\exp (s _ {q v} (\\mathbf {e} _ {\\mathbf {t} i} , \\mathbf {e} _ {\\mathbf {v} i}) / \\tau)}{\\sum_ {j} ^ {B} \\exp (s _ {q v} (\\mathbf {e} _ {\\mathbf {t} i} , \\mathbf {e} _ {\\mathbf {v} j}) / \\tau)}, \\\\ \\mathcal {L} _ {V 2 Q} = - \\frac {1}{B} \\sum_ {i} ^ {B} \\log \\frac {\\exp \\left(s _ {q v} \\left(\\mathbf {e} _ {\\mathbf {t} i} , \\mathbf {e} _ {\\mathbf {v} i}\\right) / \\tau\\right)}{\\sum_ {j} ^ {B} \\exp \\left(s _ {q v} \\left(\\mathbf {e} _ {\\mathbf {t} j} , \\mathbf {e} _ {\\mathbf {v} i}\\right) / \\tau\\right)}, \\tag {6} \\\\ \\mathcal {L} _ {Q V} = \\frac {1}{2} (\\mathcal {L} _ {Q 2 V} + \\mathcal {L} _ {V 2 Q}), \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.242, + 0.47, + 0.304 + ], + "angle": 0, + "content": "where \\( s_{qv}(\\cdot, \\cdot) \\) represents the query-video matching (e.g., global matching, fine-grained matching) similarity function shown in Figure 2(b). The total loss \\( \\mathcal{L} \\) is the sum of Query-Video loss \\( \\mathcal{L}_{QV} \\) and Query-Caption loss \\( \\mathcal{L}_{QC} \\):" + }, + { + "type": "equation", + "bbox": [ + 0.209, + 0.313, + 0.47, + 0.331 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\mathcal {L} _ {Q V} + \\mathcal {L} _ {Q C}. \\tag {7}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.34, + 0.396, + 0.357 + ], + "angle": 0, + "content": "3. Experiments: Text-Video Retrieval" + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.364, + 0.166, + 0.381 + ], + "angle": 0, + "content": "3.1. Setups" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.387, + 0.47, + 0.629 + ], + "angle": 0, + "content": "Datasets. We conduct experiment on four popular benchmarks for video-to-text retrieval and text-to-video retrieval tasks. MSR-VTT [44] contains a total of 10K video clips, each having 20 captions. Following the data splits from [10,24,27], we train models with associated captions on the Training-9K set and report results on the test 1K-A set. DiDeMo [1] has 10K videos paired with 40K descriptions. Following previous works [2, 19, 24], we concatenate all descriptions of one video to a single query, acting as a video-paragraph retrieval task. VATEX [38] collects \\(\\sim 35\\mathrm{K}\\) videos, each with multiple annotations. There are \\(\\sim 26\\mathrm{K}\\) videos for training, 1,500 videos for validation and 1,500 videos for testing. MSVD [43] contains 1,970 videos with 80K captions, with \\(\\sim 40\\) captions on average per video. There are 1,200, 100, and 670 videos in the train, validation, and test sets, respectively." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.63, + 0.47, + 0.765 + ], + "angle": 0, + "content": "Evaluation Metrics. For brevity, we abbreviate Recall at \\( K \\) to \\( \\mathbb{R}@\\mathbb{K} \\) (\\( K = 1,5,10 \\)) upon all datasets, which computes the percentage of correct videos among the top \\( K \\) retrieved videos given textual queries (Text→Video, and vice versa). MdR, Median Rank, computes the median of the ground-truth in the retrieval ranking list. MnR, Mean Rank, computes the mean rank of the correct results in the retrieval ranking list. Note that for MdR and MnR, the lower score means the better (indicated as \\( \\downarrow \\))." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Implementation Details. All experiments use the visual encoder in CLIP [30] as the video encoder, and the textual encoder in CLIP as both the caption encoder and query encoder. The caption encoder and query encoder share parameters. To reduce conflict between the two branches, the query-video branch is trained first, followed by the query-caption branch. The text length is fixed to 32, and the video length is fixed to 12 for all datasets except DiDeMo (64 max words and 64 frames). The initial learning rate is set to 1e-7" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.259 + ], + "angle": 0, + "content": "for the clip parameters and 1e-4 for the non-clip parameters. The model is trained with a batch size of 128 for 5 epochs, except for DiDeMo (15 epochs), using the Adam [18] optimizer. All learning rates follow the cosine schedule with a linear warmup [14] strategy. For the number of generated captions per video, we set \\( C \\) to 30. The interaction module employs \\( L \\) transformer layers, where \\( L \\) is set to 4 for VA-TEX and MSR-VTT, and 1 for Didemo and MSVD. In the caption branch, the number of transformer layers is set to 2. For caption generation, we directly use the original pretrained CLIP and GPT-2, without any additional tuning." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.268, + 0.806, + 0.285 + ], + "angle": 0, + "content": "3.2. Comparison with State-of-the-Arts" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.292, + 0.892, + 0.338 + ], + "angle": 0, + "content": "In this section, we compare our Cap4Video with recent state-of-the-art methods on four benchmarks: MSR-VTT [44], MSVD [43], VATEX [38], and DiDeMo [1]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.338, + 0.892, + 0.398 + ], + "angle": 0, + "content": "Table 1 shows the comparisons on DiDeMo, where our Cap4Video outperforms CLIP4Clip [24] by a significant margin of \\(9.2\\%\\) in R@1 and exceeds DRL [36] by \\(3.0\\%\\), demonstrating the effectiveness of our method." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.399, + 0.892, + 0.534 + ], + "angle": 0, + "content": "Table 2 provides a comparison of our approach with recent state-of-the-art models on MSR-VTT. Our method achieves new state-of-the-art performance on text-to-video retrieval for both ViT-B/32 and ViT-B/16 backbones, significantly surpassing previous works. For instance, we achieve a \\(+4.8\\%\\) higher R@1 than CLIP4Clip with the same ViT-B/32 on text-to-video retrieval. Additionally, our Cap4Video outperforms the recent TS2-Net [22] by \\(2.3\\%\\) and \\(2.0\\%\\) with ViT-B/32 and ViT-B/16, respectively." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.535, + 0.892, + 0.67 + ], + "angle": 0, + "content": "Table 3 and Table 4 show the results for the MSVD and VATEX datasets, respectively, where we use ViT-B/16 as our backbone. For MSVD, our Cap4Video achieves a remarkable performance of \\(51.8\\%\\) R@1 and outperforms CLIP-based models CLIP4Clip [24] and X-Pool [25] by \\(6.6\\%\\) and \\(4.6\\%\\) on text-to-video retrieval, respectively. For VATEX, our approach also outperforms the recent state-of-the-art methods and achieves a \\(+7.5\\%\\) R@1 improvement over TS2-Net [22] for text-to-video retrieval." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.671, + 0.892, + 0.687 + ], + "angle": 0, + "content": "In recent studies, several methods have been proposed" + }, + { + "type": "table", + "bbox": [ + 0.52, + 0.71, + 0.875, + 0.876 + ], + "angle": 0, + "content": "
MethodR@1R@5R@10MdRMnR
CE [21]15.640.9-8.2-
ClipBERT [19]21.147.361.16.3-
Frozen [2]31.059.872.43.0-
TMVM [20]36.564.975.43.0-
CLIP4Clip [24]42.868.579.22.018.9
TS2-Net [22]41.871.682.02.014.8
HunYuan [28]45.075.683.42.012.0
DRL [36]49.076.584.52.0-
Cap4Video52.079.487.5110.5
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.883, + 0.879, + 0.897 + ], + "angle": 0, + "content": "Table 1. Results of text-to-video retrieval on the DiDeMo [1]." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "10708" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.084, + 0.089, + 0.887, + 0.475 + ], + "angle": 0, + "content": "
MethodVenueText → VideoVideo → Text
R@1R@5R@10MdR↓MnR↓R@1R@5R@10MdR↓MnR↓
ClipBERT [19]CVPR'2022.046.859.96.0-----
MMT [10]ECCV'2026.657.169.64.0-27.057.569.73.721.3
T2VLAD [39]CVPR'2129.559.070.14.0-31.860.071.13.0
SupportSet [29]ICLR'2130.158.569.33.0-28.558.671.63.0-
Frozen [2]ICCV'2132.561.571.23.0------
BridgeFormer [12]CVPR'2237.664.875.1-------
TMVM [20]NeurIPS'2236.264.275.73.0-34.863.873.73.0-
CLIP-ViT-B/32
CLIP4Clip [24]arXiv'2144.571.481.62.015.342.770.980.62.011.6
CenterCLIP [48]SIGIR'2244.271.682.12.015.142.871.782.22.010.9
CAMoE [7]arXiv'2144.672.681.82.013.345.172.483.12.010.0
CLIP2Video [9]arXiv'2145.672.681.72.014.643.572.382.12.010.2
X-Pool [13]CVPR'2246.972.882.22.014.3-----
QB-Norm [4]CVPR'2247.273.083.02.0------
TS2-Net [22]ECCV'2247.074.583.82.013.045.374.183.72.09.2
DRL [36]arXiv'2247.474.683.82.0-45.373.983.32.0-
Cap4Video49.374.383.82.012.047.173.784.32.08.7
CLIP-ViT-B/16
CLIP2TV [11]arXiv'2148.374.682.82.014.946.575.484.92.010.2
CenterCLIP [48]SIGIR'2248.473.882.02.013.847.775.083.32.010.2
TS2-Net [22]ECCV'2249.475.685.32.013.546.675.984.92.08.9
DRL [36]arXiv'2250.276.584.71.0-48.976.385.42.0-
Cap4Video51.475.783.91.012.449.075.285.02.08
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.482, + 0.892, + 0.511 + ], + "angle": 0, + "content": "Table 2. Retrieval results on the validation set of MSR-VTT 1K [44]. Here we report results without any post-processing operations (e.g., DSL [7] or QB-Norm [4]) during inference." + }, + { + "type": "table", + "bbox": [ + 0.097, + 0.534, + 0.451, + 0.687 + ], + "angle": 0, + "content": "
MethodR@1R@5R@10MdRMnR
CE [21]19.849.063.86.0-
SUPPORT [29]28.460.072.94.0-
CLIP [30]37.064.173.83.0-
Frozen [2]33.764.776.33.0-
TMVM [20]36.767.481.32.5-
CLIP4Clip [24]45.275.584.32.010.3
X-Pool [13]47.277.486.02.09.3
Cap4Video51.880.888.318.3
" + }, + { + "type": "table_caption", + "bbox": [ + 0.091, + 0.692, + 0.455, + 0.707 + ], + "angle": 0, + "content": "Table 3. Results of text-to-video retrieval on the MSVD [43]." + }, + { + "type": "table", + "bbox": [ + 0.521, + 0.534, + 0.875, + 0.687 + ], + "angle": 0, + "content": "
MethodR@1R@5R@10MdRMnR
HGR [6]35.173.583.52.0-
CLIP [30]39.772.382.22.012.8
SUPPORT [29]44.982.189.71.0-
CLIP4Clip [24]55.989.295.01.03.9
Clip2Video [9]57.390.095.51.03.6
QB-Norm [4]58.888.393.81.0-
TS2-Net [22]59.190.095.21.03.5
Cap4Video66.693.197.012.7
" + }, + { + "type": "table_caption", + "bbox": [ + 0.512, + 0.692, + 0.88, + 0.707 + ], + "angle": 0, + "content": "Table 4. Results of text-to-video retrieval on the VATEX [38]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.473, + 0.903 + ], + "angle": 0, + "content": "to improve text-video retrieval performance by adjusting similarity during inference using other query information. Notably, our results adheres to the standard retrieval logic, where the most relevant video is retrieved for each query from a set of videos, without any knowledge of the relationship between other queries and videos. Therefore, all results reported in our tables do not involve any post-processing procedures such as DSL [7] and QB-Norm [4]. Overall, the consistent state-of-the-art performance across four benchmarks demonstrates the effectiveness of our Cap4Video." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.738, + 0.655, + 0.754 + ], + "angle": 0, + "content": "3.3. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.763, + 0.892, + 0.793 + ], + "angle": 0, + "content": "In this section, we provide detailed ablation studies to clarify the effects of each part of our design." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.796, + 0.892, + 0.903 + ], + "angle": 0, + "content": "Auxiliary Caption as Data Augmentation. We begin by investigating the impact of captions on data augmentation for training. In a real-world scenario, the original video title would naturally serve as an additional auxiliary caption. Therefore, we manually extracted the title from the video's original webpage and compared it to the caption generated by the GPT-2 model. Table 5 presents the re" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "10709" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.098, + 0.089, + 0.872, + 0.393 + ], + "angle": 0, + "content": "
MethodGlobal MatchingFine-grained Matching
R@1R@5R@10MdR↓MnR↓R@1R@5R@10MdR↓MnR↓
Baseline42.870.479.0216.645.773.782.6213.1
+Different Sources of Caption as Data Augmentation
Video Title from Source URL43.871.180.9215.144.372.783.5213.1
Zero-shot Video Captioning44.270.781.5216.246.372.581.7212.9
+Different Number of Captions for Data Augmentation
Top-144.270.781.5216.246.372.581.7212.9
Top-343.371.781.6215.045.573.882.4212.7
Top-543.470.680.4216.245.672.782.7212.9
+Video-Caption Feature Interaction
Video Only44.270.781.5216.246.372.581.7212.9
Sum43.871.580.3216.147.273.382.8213.1
Concat-MLP37.566.178.4315.740.068.779.9212.7
Cross Transformer44.671.680.3214.647.975.483.0211.5
Co-attention Transformer45.371.280.9215.048.574.082.5212.7
+Query-Caption Matching Score
Query-Video Only45.371.280.9215.048.574.082.5212.7
Query-Caption Only30.355.267.5426.430.355.267.5426.4
Query-Video + Query-Caption45.671.781.2214.849.374.283.4212.1
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.4, + 0.895, + 0.444 + ], + "angle": 0, + "content": "Table 5. Component-wise evaluation of our framework on the MSR-VTT 1K validation set. With the ViT-B/32 backbone, we report the text-to-video retrieval results for two representative Query-Video matching mechanisms: global matching and fine-grained matching. The consistent improvement on two typical matching mechanisms demonstrates the generalization ability and effectiveness of our method." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.469, + 0.47, + 0.56 + ], + "angle": 0, + "content": "sults of using different sources of the caption, from which we observe that using captions generated by the web-scale model as data augmentation for training can lead to direct improvements in R@1 \\((+1.4\\%, +0.6\\%)\\) under both matching mechanisms. Using video titles can also bring a \\(1\\%\\) improvement for global matching." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.562, + 0.47, + 0.653 + ], + "angle": 0, + "content": "We also explore the impact of the number of generated captions used for augmentation. We used the caption filtering mechanism mentioned in Sec. 2.3 to rank the relevance of captions and the ground-truth query, and selected different numbers of captions for training. The results demonstrated that using only one caption is sufficient." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.656, + 0.47, + 0.793 + ], + "angle": 0, + "content": "Benefit on Both Online and Offline Videos Scenarios. Our method is applicable for both online and offline videos. Offline videos are local videos with no title, while online videos have a title on the video website. Therefore, for online videos, the step of generating a caption with \\(\\mathrm{CLIP + GPT - 2}\\) is skipped, and the website title is used directly as a caption. In Table 6, the results show that our method has significantly improved over the global matching baseline for both offline and online videos." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.795, + 0.47, + 0.903 + ], + "angle": 0, + "content": "Video-Caption Feature Interaction. As mentioned in Sec. 2.4, we have designed four approaches for Video-Caption feature interaction. Based on the results presented in Table 5, we can conclude the following: 1) The basic approach of Sum has been shown to enhance fine-grained matching by \\(0.9\\%\\) R@1, but there is no noticeable improvement in global matching. 2) The MLP approach is difficult" + }, + { + "type": "table", + "bbox": [ + 0.536, + 0.467, + 0.858, + 0.546 + ], + "angle": 0, + "content": "
Caption SourceR@1
BaselineN/A42.8
Cap4VideoOriginal Website Title45.8
Captioner (CLIP+GPT-2)45.6
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.553, + 0.892, + 0.583 + ], + "angle": 0, + "content": "Table 6. Exploring the effectiveness of captions from different sources on MSR-VTT 1k-A. Setting: ViT-B/32, global matching." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.612, + 0.892, + 0.824 + ], + "angle": 0, + "content": "to optimize and performs poorly in both matching scenarios. We speculate that the MLP's operation in a black-box environment, despite creating a nonlinear metric space, may lead to degradation. 3) The Cross Transformer approach has demonstrated improvements of \\(+0.4\\%\\) and \\(+1.6\\%\\) in two matching settings, respectively. These enhancements may be attributed to the self-attention mechanism's ability to capture the inter-modal relationship between the video and caption. 4) Moreover, the Co-attention Transformer approach has significantly boosted performance, with gains of \\(+1.1\\%\\) and \\(+2.2\\%\\) for these two matching mechanisms. In summary, the results indicate that proper interaction between the video and generated caption can lead to better video representation and improved Query-Video matching." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.825, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Query-Caption matching. We also investigate the Query-Caption matching branch for text-video retrieval. We aggregate caption embeddings using mean pooling to yield a global embedding. As shown in Table 5, the single Query-Caption matching branch achieves \\(30.3\\%\\) R@1 on text-to" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "10710" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.171, + 0.09, + 0.386, + 0.102 + ], + "angle": 0, + "content": "Query7765: a person is discussing a car." + }, + { + "type": "table", + "bbox": [ + 0.081, + 0.103, + 0.468, + 0.231 + ], + "angle": 0, + "content": "
VideoRank+ CaptionRank
6video of a car camera recording the driver's voice.1
4video showing the car in a parking spot.2
5video of SUV in the video below shows a salesman talking to an audience.3
" + }, + { + "type": "table_caption", + "bbox": [ + 0.094, + 0.232, + 0.456, + 0.244 + ], + "angle": 0, + "content": "Query9616: person is recording the brown horse which is having fun." + }, + { + "type": "table", + "bbox": [ + 0.081, + 0.245, + 0.468, + 0.379 + ], + "angle": 0, + "content": "
VideoRank 2+ Caption video of the horse jumping over a fence at Ranch in Nevada was captured on camera.Rank 1
1video showing animation of a horse's simulation, which simulates the game.2
4video showing a horse simulation video game in which you could see your avatar being animated by the camera.3
" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.392, + 0.468, + 0.406 + ], + "angle": 0, + "content": "Figure 4. The text-video results on the MSR-VTT 1K-A test set." + }, + { + "type": "image_caption", + "bbox": [ + 0.078, + 0.407, + 0.468, + 0.42 + ], + "angle": 0, + "content": "Left: The ranking results of the query-video matching model." + }, + { + "type": "image_caption", + "bbox": [ + 0.078, + 0.421, + 0.468, + 0.448 + ], + "angle": 0, + "content": "Right: The ranking results of Cap4Video, which incorporates generated captions to enhance retrieval. Please zoom in for best view." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.476, + 0.468, + 0.581 + ], + "angle": 0, + "content": "video retrieval, outperforming previous query-video matching methods such as ClipBERT [19] (22.0%) and MMT [10] (26.6%). This suggests that the Query-Caption matching branch can complement the regular Query-Video matching branch for improved text-video retrieval. Combining the score of Query-Caption matching branch with Query-Video matching branch further improves performance (+0.8%)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.583, + 0.468, + 0.643 + ], + "angle": 0, + "content": "Overall, Cap4Video utilizes generated captions in three ways: input data augmentation, intermediate feature interaction, and output score fusion, leading to consistent improvements \\((+2.8\\% / +3.6\\%)\\) in both matching mechanisms." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.655, + 0.214, + 0.669 + ], + "angle": 0, + "content": "3.4. Visualization" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.678, + 0.468, + 0.784 + ], + "angle": 0, + "content": "We provide two examples of videos retrieved by our Cap4Video method and a model without auxiliary captions. As illustrated in Figure 4, our approach successfully retrieves the ground-truth video with the assistance of the caption, while the video-only model returns multiple videos that are somewhat relevant to the query but not precise. See more qualitative results in Supplementary." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.8, + 0.226, + 0.815 + ], + "angle": 0, + "content": "4. Related Works" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.468, + 0.901 + ], + "angle": 0, + "content": "Zero-shot Image Captioning. In the field of natural language processing, transformer-based GPT models [5, 31] have been successful in generating text from prompts by training on large-scale text corpora. Similarly, CLIP [30], a vision-language alignment model trained on 400 million" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.303 + ], + "angle": 0, + "content": "image-text pairs, has demonstrated impressive zero-shot performance on vision tasks. However, research on transferring web-scale models to zero-shot video captioning remains limited. Recently, ZeroCap [34] proposed a method of using CLIP and the GPT-2 language model to generate textual descriptions of input images, leveraging knowledge from both models in a truly zero-shot manner without retraining or fine-tuning model parameters. MAGIC [32] has also used CLIP scores to align GPT-2 logits with corresponding images but requires fine-tuning on the MS-COCO caption text corpus. More recently, a study [33] extended the zero-shot capability of ZeroCap to the video domain. In this paper, we employ this video extension to generate auxiliary captions without any additional training." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.304, + 0.892, + 0.726 + ], + "angle": 0, + "content": "Text-Video Retrieval aims to retrieve relevant video content based on natural language descriptions. Early studies [6, 10, 21, 37, 39] focused on knowledge transfer from \"expert\" models and captured intra-modal and cross-modal interactions based on pre-extracted features. However, the performance of these methods is limited since they cannot perform end-to-end optimization. Recently, more methods have involved end-to-end training for text-video retrieval. One typical approach [2, 26, 27] is to first perform largescale text-video pre-training, then transfer the model to downstream text-video retrieval tasks. With the emergence of pre-trained Vision-Language Models (VLMs), there have been increased efforts to leverage them for improving video understanding [41, 42]. Thus, another training-efficient line is to directly expand the pre-trained VLM to the text-video retrieval task. CLIPBERT [19] enables affordable pioneering end-to-end training with a sparse sampling strategy. After that, recent works [3, 4, 8, 9, 11, 13, 22, 24, 40, 48] focus on transferring knowledge from CLIP models that have been pre-trained on 400M image-text pairs. The research path has evolved from the most direct global matching (i.e., video-sentence alignment [11, 24]) to fine-grained matching (e.g., frame-word alignment [36], video-word alignment [13], multi-hierarchical alignment [9, 28]). Unlike above CLIP-Based efforts on query-video matching, we propose to generate auxiliary captions from videos to improve text-video retrieval. Thus our method is compatible with both global and fine-grained matching." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.74, + 0.618, + 0.755 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.892, + 0.901 + ], + "angle": 0, + "content": "We introduce Cap4Video, a novel framework that leverages captions generated by web-scale language models to enhance text-video matching in three key ways: 1) Input data augmentation for training, 2) Intermediate videocaption feature interaction for compact video representations, and 3) Output score fusion for improved text-video retrieval. Our approach demonstrates consistent performance gains on four standard text-video retrieval benchmarks, outperforming state-of-the-art methods by a clear margin." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "10711" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.116, + 0.47, + 0.17 + ], + "angle": 0, + "content": "[1] Lisa Anne Hendricks, Oliver Wang, Eli Shechtman, Josef Sivic, Trevor Darrell, and Bryan Russell. Localizing moments in video with natural language. In ICCV, pages 5803-5812, 2017. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.173, + 0.471, + 0.227 + ], + "angle": 0, + "content": "[2] Max Bain, Arsha Nagrani, Gül Varol, and Andrew Zisserman. Frozen in time: A joint video and image encoder for end-to-end retrieval. In ICCV, pages 1728-1738, 2021. 5, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.23, + 0.471, + 0.272 + ], + "angle": 0, + "content": "[3] Max Bain, Arsha Nagrani, Gül Varol, and Andrew Zisserman. A clip-hitchhiker's guide to long video retrieval. arXiv preprint arXiv:2205.08508, 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.274, + 0.471, + 0.315 + ], + "angle": 0, + "content": "[4] Simion-Vlad Bogolin, Ioana Croitoru, Hailin Jin, Yang Liu, and Samuel Albanie. Cross modal retrieval with querybank normalisation. In CVPR, pages 5194-5205, 2022. 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.317, + 0.471, + 0.384 + ], + "angle": 0, + "content": "[5] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. NeurIPS, 33:1877-1901, 2020. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.387, + 0.471, + 0.429 + ], + "angle": 0, + "content": "[6] Shizhe Chen, Yida Zhao, Qin Jin, and Qi Wu. Fine-grained video-text retrieval with hierarchical graph reasoning. In CVPR, pages 10638-10647, 2020. 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.431, + 0.471, + 0.485 + ], + "angle": 0, + "content": "[7] Xing Cheng, Hezheng Lin, Xiangyu Wu, Fan Yang, and Dong Shen. Improving video-text retrieval by multi-stream corpus alignment and dual softmax loss. arXiv preprint arXiv:2109.04290, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.488, + 0.471, + 0.543 + ], + "angle": 0, + "content": "[8] Bo Fang, Chang Liu, Yu Zhou, Min Yang, Yuxin Song, Fu Li, Weiping Wang, Xiangyang Ji, Wanli Ouyang, et al. Uatvr: Uncertainty-adaptive text-video retrieval. arXiv preprint arXiv:2301.06309, 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.545, + 0.471, + 0.586 + ], + "angle": 0, + "content": "[9] Han Fang, Pengfei Xiong, Luhui Xu, and Yu Chen. Clip2video: Mastering video-text retrieval via image clip. arXiv preprint arXiv:2106.11097, 2021. 1, 2, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.588, + 0.471, + 0.629 + ], + "angle": 0, + "content": "[10] Valentin Gabeur, Chen Sun, Karteek Alahari, and Cordelia Schmid. Multi-modal transformer for video retrieval. In ECCV, pages 214-229. Springer, 2020. 5, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.631, + 0.471, + 0.687 + ], + "angle": 0, + "content": "[11] Zijian Gao, Jingyu Liu, Sheng Chen, Dedan Chang, Hao Zhang, and Jinwei Yuan. Clip2tv: An empirical study on transformer-based methods for video-text retrieval. arXiv preprint arXiv:2111.05610, 2021. 1, 2, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.689, + 0.471, + 0.742 + ], + "angle": 0, + "content": "[12] Yuying Ge, Yixiao Ge, Xihui Liu, Dian Li, Ying Shan, Xiaohu Qie, and Ping Luo. Bridging video-text retrieval with multiple choice questions. In CVPR, pages 16167-16176, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.746, + 0.471, + 0.8 + ], + "angle": 0, + "content": "[13] Satya Krishna Gorti, Noel Vouitsis, Junwei Ma, Keyvan Golestan, Maksims Volkovs, Animesh Garg, and Guangwei Yu. X-pool: Cross-modal language-video attention for text-video retrieval. In CVPR, pages 5006-5015, 2022. 1, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.803, + 0.471, + 0.87 + ], + "angle": 0, + "content": "[14] Priya Goyal, Piotr Dólar, Ross Girshick, Pieter Noordhuis, Lukasz Wesolowski, Aapo Kyrola, Andrew Tulloch, Yangqing Jia, and Kaiming He. Accurate, large minibatch sgd: Training imagenet in 1 hour. arXiv preprint arXiv:1706.02677, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.471, + 0.901 + ], + "angle": 0, + "content": "[15] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.116, + 0.471, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.134 + ], + "angle": 0, + "content": "Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In ICML, pages 4904-4916. PMLR, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.136, + 0.892, + 0.204 + ], + "angle": 0, + "content": "[16] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In ICML, pages 4904-4916. PMLR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.206, + 0.892, + 0.275 + ], + "angle": 0, + "content": "[17] Omar Khattab and Matei Zaharia. Colbert: Efficient and effective passage search via contextualized late interaction over bert. In Proceedings of the 43rd International ACM SIGIR conference on research and development in Information Retrieval, pages 39-48, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.277, + 0.892, + 0.304 + ], + "angle": 0, + "content": "[18] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In ICLR, 2015. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.306, + 0.892, + 0.36 + ], + "angle": 0, + "content": "[19] Jie Lei, Linjie Li, Luowei Zhou, Zhe Gan, Tamara L Berg, Mohit Bansal, and Jingjing Liu. Less is more: Clipbert for video-and-language learning via sparse sampling. In CVPR, pages 7331-7341, 2021. 5, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.362, + 0.892, + 0.417 + ], + "angle": 0, + "content": "[20] Chengzhi Lin, Ancong Wu, Junwei Liang, Jun Zhang, Wenhang Ge, Wei-Shi Zheng, and Chunhua Shen. Text-adaptive multiple visual prototype matching for video-text retrieval. In NeurIPS, 2022. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.42, + 0.892, + 0.473 + ], + "angle": 0, + "content": "[21] Yang Liu, Samuel Albanie, Arsha Nagrani, and Andrew Zisserman. Use what you have: Video retrieval using representations from collaborative experts. arXiv preprint arXiv:1907.13487, 2019. 5, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.476, + 0.892, + 0.53 + ], + "angle": 0, + "content": "[22] Yuqi Liu, Pengfei Xiong, Luhui Xu, Shengming Cao, and Qin Jin. Ts2-net: Token shift and selection transformer for text-video retrieval. In ECCV, pages 319-335. Springer, 2022. 5, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.533, + 0.892, + 0.574 + ], + "angle": 0, + "content": "[23] Jiasen Lu, Dhruv Batra, Devi Parikh, and Stefan Lee. Vilbert: Pretraining task-agnostic visiolinguistic representations for vision-and-language tasks. NeurIPS, 32, 2019. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.576, + 0.892, + 0.63 + ], + "angle": 0, + "content": "[24] Huaishao Luo, Lei Ji, Ming Zhong, Yang Chen, Wen Lei, Nan Duan, and Tianrui Li. Clip4clip: An empirical study of clip for end to end video clip retrieval and captioning. Neurocomputing, 508:293-304, 2022. 1, 2, 4, 5, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.633, + 0.892, + 0.687 + ], + "angle": 0, + "content": "[25] Yiwei Ma, Guohai Xu, Xiaoshuai Sun, Ming Yan, Ji Zhang, and Rongrong Ji. X-clip: End-to-end multi-grained contrastive learning for video-text retrieval. In ACM MM, pages 638–647, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.689, + 0.892, + 0.744 + ], + "angle": 0, + "content": "[26] Antoine Miech, Jean-Baptiste Alayrac, Lucas Smaira, Ivan Laptev, Josef Sivic, and Andrew Zisserman. End-to-end learning of visual representations from uncurated instructional videos. In CVPR, pages 9879-9889, 2020. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.746, + 0.892, + 0.814 + ], + "angle": 0, + "content": "[27] Antoine Miech, Dimitri Zhukov, Jean-Baptiste Alayrac, Makarand Tapaswi, Ivan Laptev, and Josef Sivic. Howto100m: Learning a text-video embedding by watching hundred million narrated video clips. In ICCV, pages 2630-2640, 2019. 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.817, + 0.892, + 0.871 + ], + "angle": 0, + "content": "[28] Shaobo Min, Weijie Kong, Rong-Cheng Tu, Dihong Gong, Chengfei Cai, Wenzhe Zhao, Chenyang Liu, Sixiao Zheng, Hongfa Wang, Zhifeng Li, et al. Hunyuan_tvr for text-video retrivial. arXiv preprint arXiv:2204.03382, 2022. 1, 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.873, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[29] Mandela Patrick, Po-Yao Huang, Yuki Asano, Florian Metze, Alexander G Hauptmann, Joao F Henriques, and An" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "10712" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.092, + 0.469, + 0.12 + ], + "angle": 0, + "content": "drea Vedaldi. Support-set bottlenecks for video-text representation learning. In ICLR, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.121, + 0.469, + 0.202 + ], + "angle": 0, + "content": "[30] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, pages 8748-8763. PMLR, 2021. 1, 2, 3, 5, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.204, + 0.469, + 0.245 + ], + "angle": 0, + "content": "[31] Alec Radford, Jeff Wu, Rewon Child, David Luan, Dario Amodei, and Ilya Sutskever. Language models are unsupervised multitask learners. 2019. 1, 2, 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.247, + 0.469, + 0.3 + ], + "angle": 0, + "content": "[32] Yixuan Su, Tian Lan, Yahui Liu, Fangyu Liu, Dani Yogatama, Yan Wang, Lingpeng Kong, and Nigel Collier. Language models can see: Plugging visual controls in text generation. arXiv preprint arXiv:2205.02655, 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.302, + 0.469, + 0.342 + ], + "angle": 0, + "content": "[33] Yoad Tewel, Yoav Shalev, Roy Nadler, Idan Schwartz, and Lior Wolf. Zero-shot video captioning with evolving pseudotokens. arXiv preprint arXiv:2207.11100, 2022. 2, 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.344, + 0.469, + 0.397 + ], + "angle": 0, + "content": "[34] Yoad Tewel, Yoav Shalev, Idan Schwartz, and Lior Wolf. Zerocap: Zero-shot image-to-text generation for visual-semantic arithmetic. In CVPR, pages 17918-17928, 2022. 2, 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.4, + 0.469, + 0.453 + ], + "angle": 0, + "content": "[35] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In NeurIPS, pages 5998-6008, 2017. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.456, + 0.469, + 0.509 + ], + "angle": 0, + "content": "[36] Qiang Wang, Yanhao Zhang, Yun Zheng, Pan Pan, and Xian-Sheng Hua. Disentangled representation learning for text-video retrieval. arXiv preprint arXiv:2203.07111, 2022. 1, 2, 4, 5, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.511, + 0.469, + 0.566 + ], + "angle": 0, + "content": "[37] Wenzhe Wang, Mengdan Zhang, Runnan Chen, Guanyu Cai, Penghao Zhou, Pai Peng, Xiaowei Guo, Jian Wu, and Xing Sun. Dig into multi-modal cues for video retrieval with hierarchical alignment. In *IJCAI*, pages 1113–1121, 2021. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.567, + 0.469, + 0.621 + ], + "angle": 0, + "content": "[38] Xin Wang, Jiawei Wu, Junkun Chen, Lei Li, Yuan-Fang Wang, and William Yang Wang. Vatex: A large-scale, high-quality multilingual dataset for video-and-language research. In ICCV, pages 4581-4591, 2019. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.623, + 0.469, + 0.663 + ], + "angle": 0, + "content": "[39] Xiaohan Wang, Linchao Zhu, and Yi Yang. T2vlad: global-local sequence alignment for text-video retrieval. In CVPR, pages 5079-5088, 2021. 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.665, + 0.469, + 0.718 + ], + "angle": 0, + "content": "[40] Xiaohan Wang, Linchao Zhu, Zhedong Zheng, Mingliang Xu, and Yi Yang. Align and tell: Boosting text-video retrieval with local alignment and fine-grained supervision. IEEE TMM, 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.72, + 0.469, + 0.76 + ], + "angle": 0, + "content": "[41] Wenhao Wu, Zhun Sun, and Wanli Ouyang. Revisiting classifier: Transferring vision-language models for video recognition. In AAAI, 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.762, + 0.469, + 0.816 + ], + "angle": 0, + "content": "[42] Wenhao Wu, Xiaohan Wang, Haipeng Luo, Jingdong Wang, Yi Yang, and Wanli Ouyang. Bidirectional cross-modal knowledge exploration for video recognition with pretrained vision-language models. In CVPR, 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.818, + 0.469, + 0.859 + ], + "angle": 0, + "content": "[43] Zuxuan Wu, Ting Yao, Yanwei Fu, and Yu-Gang Jiang. Deep learning for video classification and captioning. In Frontiers of multimedia research, pages 3-29. 2017. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.86, + 0.469, + 0.9 + ], + "angle": 0, + "content": "[44] Jun Xu, Tao Mei, Ting Yao, and Yong Rui. Msr-vtt: A large video description dataset for bridging video and language. In CVPR, pages 5288-5296, 2016. 2, 5, 6" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.469, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.147 + ], + "angle": 0, + "content": "[45] Lewei Yao, Runhui Huang, Lu Hou, Guansong Lu, Minzhe Niu, Hang Xu, Xiaodan Liang, Zhenguo Li, Xin Jiang, and Chunjing Xu. Filip: Fine-grained interactive language-image pre-training. In ICLR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.149, + 0.892, + 0.203 + ], + "angle": 0, + "content": "[46] Jiahui Yu, Zirui Wang, Vijay Vasudevan, Legg Yeung, Mojtaba Seyhedhosseini, and Yonghui Wu. Coca: Contrastive captioners are image-text foundation models. arXiv preprint arXiv:2205.01917, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.206, + 0.892, + 0.274 + ], + "angle": 0, + "content": "[47] Lu Yuan, Dongdong Chen, Yi-Ling Chen, Noel Codella, Xiyang Dai, Jianfeng Gao, Houdong Hu, Xuedong Huang, Boxin Li, Chunyuan Li, et al. Florence: A new foundation model for computer vision. arXiv preprint arXiv:2111.11432, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.277, + 0.892, + 0.345 + ], + "angle": 0, + "content": "[48] Shuai Zhao, Linchao Zhu, Xiaohan Wang, and Yi Yang. Centerclip: Token clustering for efficient text-video retrieval. In Proceedings of the 45th International ACM SIGIR Conference on Research and Development in Information Retrieval, pages 970–981, 2022. 6, 8" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.345 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.956 + ], + "angle": 0, + "content": "10713" + } + ] +] \ No newline at end of file diff --git a/2023/Cap4Video_ What Can Auxiliary Captions Do for Text-Video Retrieval_/926c868d-a104-48e5-bcb0-c1285d26ad5d_origin.pdf b/2023/Cap4Video_ What Can Auxiliary Captions Do for Text-Video Retrieval_/926c868d-a104-48e5-bcb0-c1285d26ad5d_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..7ea225c2a13b1a33093716bc6b49cc91362459c1 --- /dev/null +++ b/2023/Cap4Video_ What Can Auxiliary Captions Do for Text-Video Retrieval_/926c868d-a104-48e5-bcb0-c1285d26ad5d_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d876fda071a3e4bf63d92705a762f0080d6b14393e31baf5f2f0c025351ad23a +size 1056921 diff --git a/2023/Cap4Video_ What Can Auxiliary Captions Do for Text-Video Retrieval_/full.md b/2023/Cap4Video_ What Can Auxiliary Captions Do for Text-Video Retrieval_/full.md new file mode 100644 index 0000000000000000000000000000000000000000..b0e191d5897410645d90f0064e9aeeab1721c0ef --- /dev/null +++ b/2023/Cap4Video_ What Can Auxiliary Captions Do for Text-Video Retrieval_/full.md @@ -0,0 +1,325 @@ +# Cap4Video: What Can Auxiliary Captions Do for Text-Video Retrieval? + +Wenhao Wu $^{1,2*}$ Haipeng Luo $^{3*}$ Bo Fang $^{3}$ Jingdong Wang $^{2}$ Wanli Ouyang $^{4,1}$ $^{1}$ The University of Sydney + $^{2}$ Baidu Inc. + $^{3}$ University of Chinese Academy of Sciences + $^{4}$ Shanghai AI Laboratory +whwu.ucas@gmail.com + +# Abstract + +Most existing text-video retrieval methods focus on cross-modal matching between the visual content of videos and textual query sentences. However, in real-world scenarios, online videos are often accompanied by relevant text information such as titles, tags, and even subtitles, which can be utilized to match textual queries. This insight has motivated us to propose a novel approach to text-video retrieval, where we directly generate associated captions from videos using zero-shot video captioning with knowledge from web-scale pre-trained models (e.g., CLIP and GPT-2). Given the generated captions, a natural question arises: what benefits do they bring to text-video retrieval? To answer this, we introduce Cap4Video, a new framework that leverages captions in three ways: i) Input data: video-caption pairs can augment the training data. ii) Intermediate feature interaction: we perform cross-modal feature interaction between the video and caption to produce enhanced video representations. iii) Output score: the Query-Caption matching branch can complement the original Query-Video matching branch for text-video retrieval. We conduct comprehensive ablation studies to demonstrate the effectiveness of our approach. Without any post-processing, Cap4Video achieves state-of-the-art performance on four standard text-video retrieval benchmarks: MSR-VTT $(51.4\%)$ , VATEX $(66.6\%)$ , MSVD $(51.8\%)$ , and DiDeMo $(52.0\%)$ . The code is available at https://github.com/whwu95/Cap4Video. + +# 1. Introduction + +Text-video retrieval is a fundamental task in videolanguage learning. With the rapid advancements in image-language pre-training [15, 30, 46, 47], researchers have focused on expanding pre-trained image-language models, especially CLIP [30], to tackle the text-video retrieval task. The research path has evolved from the most direct global + +![](images/cd8bb2a400ae5fd45da3cd46c3d9b8da397d9bca24081d14ead1b6c00fa35874.jpg) + +![](images/893e334822d01329b3b958bcb5288b1fac058e0bb14c1c4e7ad1b735432f0bb7.jpg) + +![](images/f08ac9db5530b7116bde109bf9af13292ae8bf4701bc6d49e256548d1794f9d7.jpg) +Figure 1. (a) An existing end-to-end learning paradigm for text-video retrieval. (b) Zero-shot video captioning achieved by guiding a large language model (LLM) such as GPT-2 [31] with CLIP [30]. (c) Our Cap4Video framework leverages the generated captions in three aspects: input data augmentation, intermediate feature interaction, and output score fusion. + +matching (i.e., video-sentence alignment [11, 24]) to fine-grained matching (e.g., frame-word alignment [36], videoword alignment [13], multi-hierarchical alignment [9, 28], etc.). These studies have demonstrated remarkable performance and significantly outperformed previous models. Two key factors contribute to this improvement. Firstly, CLIP offers powerful visual and textual representations that are pre-aligned in the semantic embedding space, thereby reducing the challenge of cross-modal learning in videotext matching. Secondly, these methods can fine-tune the pre-trained vision and text encoders using sparsely sampled frames in an end-to-end manner. All of these methods aim to learn cross-modal alignment between the visual represen + +tation of videos and the textual representation of the corresponding query, as depicted in Figure 1(a). + +However, in real-life scenarios, online videos usually come with related content such as the video's title or tag on the video website. In addition to the visual signal in the video, the associated textual information can also be used to some extent to describe the video content and match the query (i.e., the common text-to-text retrieval). This raises a pertinent question: How can we generate associated text descriptions for videos? One possible solution is to crawl the video title from the video website. However, this method relies on annotations, and there is a risk that the video URL may have become invalid. Another automated solution is to generate captions using zero-shot video caption models. Therefore, we turn our attention to knowledge-rich pretrained models to handle such challenging open-set scenarios. We find that the recent study ZeroCap [34] provides a good practice to use frozen CLIP [30] and GPT-2 [31] for zero-shot image captioning. Thus, we leverage a video extension [33] of ZeroCap for generating captions in the video domain without any further training. + +When provided with auxiliary captions, a natural question naturally arises: How can we leverage these captions to enhance the text-video retrieval task? In this paper, we propose the Cap4Video learning framework, as illustrated in Figure 1(c), which utilizes captions in three key ways: (i) Input Data: One simple approach is to augment the training data with the generated captions. Specifically, the given video and its generated caption can be treated as a matched pair, which serves as an additional positive sample pair for training beyond the query-video pairs. (ii) Intermediate Feature Interaction: Cross-modal interaction between the video and captions can be leveraged to improve the video representation. Specifically, we can exploit the complementary information between videos and captions to reduce redundant features from videos and learn more discriminative video representations. (iii) Output score: The generated caption can also represent the video's content, allowing us to employ query-caption matching to complement standard query-video matching for the text-video retrieval task. Moreover, a two-stream architecture can be utilized to reduce model bias and produce more robust results. + +We hope that our novel paradigm will encourage further investigation into the video-language learning. In summary, our contributions are as follows: + +- We explore a novel problem: leveraging auxiliary captions to further enhance existing text-video retrieval. Besides labor-intensive manual crawling of video website titles, we investigate the potential of rich captions automatically generated by large language models (LLMs) to benefit text-video retrieval. +- We propose the Cap4Video learning framework, + +which maximizes the utility of the auxiliary captions through three aspects: input data, feature interaction, and output score. Our framework improves the performance of existing query-video matching mechanisms, including global matching and fine-grained matching. + +- Extensive experiments conducted on four video benchmarks demonstrate the effectiveness of our method. Our Cap4Video achieves state-of-the-art performance on MSR-VTT [44] (51.4%), VATEX [38] (66.6%), MSVD [43] (51.8%), and DiDeMo [1] (52.0%). + +# 2. Methodology + +# 2.1. Background: Text-Video Matching + +Text-video matching aims to evaluate the similarity between a given sentence $Q_{i}$ and a given video $V_{j}$ , typically using a similarity function $s(Q_{i},V_{j})$ . In text-to-video retrieval, the goal is to rank all videos based on their similarity scores to a given query sentence. To improve text-video retrieval, recent works [9, 11, 24] have applied CLIP [30] for initialization, leveraging pre-trained knowledge from image-text learning. Our baselines for text-video matching include two typical mechanisms: global matching and fine-grained matching, as illustrated in Figure 2(b). + +Global Matching is a commonly used technique in cross-modal contrastive learning [16, 24, 30]. In global matching, each modality is encoded independently to obtain global features, which are then used to calculate similarity. We train the visual encoder to output $F$ frame embeddings for a given video that samples $F$ frames. Similarly, the query encoder returns $W$ word embeddings and the [CLS] embedding as the global representation for a given query sentence that contains $W$ words. The frame embeddings are integrated using average pooling to obtain the global video embedding, which is then compared with the global query embedding to calculate similarity. + +Fine-grained Matching focuses on modeling the token-level alignment between two modalities, such as framework alignment. In order to achieve token-level patchword alignment for image-text learning, FILIP [45] and ColBERT [17] employ a Max-Mean pipeline. This pipeline finds the token-wise maximum similarity between patch and word tokens, and then averages the maximum similarity of tokens in the image or text to obtain the similarity between an image and a text or vice versa. Moreover, DRL [36] extends the token-wise alignment to text-video retrieval and introduces an attention mechanism to learn weighted pooling instead of mean pooling. We have adopted this mechanism as our enhanced baseline. + +# 2.2. Preprocessing: Caption Generation + +To obtain auxiliary captions for a given video, we consider the following two approaches. + +![](images/5c9205f54a594b05cb3e4eb4d0fbf51423143a5bd711a36da850cfd413e19c99.jpg) +(a) Our Cap4Video +(b) Query-Video Matching: Two typical mechanisms +(c) Query-Caption Matching + +![](images/43d5630e5396a15b2a910fed787ac0af2fcc7fefd165e0035646efcd40e5ff13.jpg) +① Global embedding matching +Figure 2. An overview of our Cap4Video for text-video retrieval. We first generate captions using a zero-shot video captioner that combines CLIP [30] with GPT-2 [31], leveraging knowledge from both frozen web-scale models. We then utilize the pre-extracted caption information from three different perspectives: i) Input data: We use the video and captions to create new positive pairs for data augmentation during training. ii) Feature interaction: We perform feature interaction between video and caption to capture intra- and inter-modality context, yielding enhanced video representations. iii) Output score: The Query-Caption matching branch can complement the original Query-Video matching branch for text-video retrieval. + +![](images/05872dffca63a7c7939be9d6844c32c5f148399b350605019d22356d859b3ce4.jpg) +② Fine-grained embedding matching + +![](images/29d4af8d46974b636e0b2b11ac5e56d999474addfb9339d9e2cf64bba9ad92a0.jpg) +Captions embeddings + +Manual Crawling of Video Titles. We extract the video website title by crawling the original links (such as YouTube ID) of each video and utilize it as the caption. However, we skip this step for videos with expired links. + +Automatic Video Captioning. In contrast to the manual approach that relies on annotations, we leverage knowledge from the LLM to generate rich and diverse captions. Given the scalability of our framework, we aim to generate captions directly from downstream videos without any additional training, a process known as zero-shot video captioning. To achieve this, we follow [33, 34] and use GPT-2 [31] to predict the next word from an initial prompt, e.g., "Video shows". A calibrated CLIP [30] loss is then used to drive the model to generate sentences that describe the video, incorporating video-related knowledge into the auto-regressive process. See Supplementary for more details. + +# 2.3. Data Augmentation with Auxiliary Captions + +Auxiliary captions can be used to augment training data. For example, for a dataset consisting of $N$ videos and their corresponding query sentences, each video and its generated caption can be considered as a positive sample pair for training, in addition to the original query-video pairs. By selecting one caption per video, we can add at least $N$ pairs + +as additional data augmentation during training. + +The automatic video captioner can generate multiple captions (e.g., 20) for each video. However, some of these captions may contain noise and may not be entirely relevant to the video content. To avoid negative effects on training, we use a filtering mechanism that evaluates the semantic similarity between each caption and the ground-truth query of the video using a pre-trained text encoder. The caption with the highest similarity is then chosen for data augmentation. Note that we only use the ground-truth query for caption filtering during the training phase. + +# 2.4. Video-Caption Cross-Modal Interaction + +We further consider taking advantage of the complementarity between videos and captions to reduce redundant features and learn more discriminative video representations. To preserve the pre-trained CLIP encoder architecture for efficient transfer learning, we limit the interaction to the final caption and frame embeddings. Specifically, we pass the frame embeddings $\mathbf{e}_{\mathbf{v}} = \{\mathbf{v}_1,\mathbf{v}_2,\dots ,\mathbf{v}_F\}$ and caption embeddings $\mathbf{e_c} = \{\mathbf{c}_1,\mathbf{c}_2,\dots ,\mathbf{c}_C\}$ to the interaction module, where $F$ and $C$ represent the number of frames and captions, respectively. Figure 2(b) depicts several ways of interaction between the two modalities. + +Sum. To obtain an enhanced frame embedding, an intuitive approach is to compute the sum of the global caption embedding $\mathbf{c}_{\mathbf{g}}$ and each frame embedding: + +$$ +\operatorname {S u m} \left(\mathbf {v} _ {i}, \mathbf {c} _ {\mathbf {g}}\right) = \mathbf {v} _ {i} + \mathbf {c} _ {\mathbf {g}}, \quad i = 1, \dots , F, \tag {1} +$$ + +where $\mathbf{v}_i\in \mathbb{R}^D$ is the $i$ -th frame embedding, and $\mathbf{c_g}\in \mathbb{R}^D$ is computed by averaging the [CLS] embeddings of $C$ generated captions: $\mathbf{c_g} = \frac{1}{C}\sum_{i = 1}^{C}\mathbf{c}_i$ . + +MLP. To model weighted combinations of each frame embedding and the global caption embedding $\mathbf{c_g}$ , we concatenate them together and pass the result through a learnable Multi-layer Perceptron (MLP): + +$$ +\operatorname {M L P} \left(\mathbf {v} _ {i}, \mathbf {c} _ {\mathbf {g}}\right) = f _ {\theta} \left(\left[ \mathbf {v} _ {i}, \mathbf {c} _ {\mathbf {g}} \right]\right), \quad \text {f o r} i = 1, \dots , F, \tag {2} +$$ + +where $[\cdot ,\cdot ]$ denotes the concatenation operation, $f_{\theta}$ is the MLP with parameter $\theta$ + +Cross Transformer. We also investigate the use of selfattention [35] for interactions. The Cross Transformer operates on a sequence $\{\mathbf{e}_{\mathbf{v}}, \mathbf{e}_{\mathbf{c}}\} = \{\mathbf{v}_1, \dots, \mathbf{v}_F, \mathbf{c}_1, \dots, \mathbf{c}_C\}$ and processes them through $L$ encoder-style transformer blocks to generate final representations: + +$$ +\operatorname {C r o s s} \left(\mathbf {e} _ {\mathbf {v}}, \mathbf {e} _ {\mathbf {c}}\right) = f _ {\psi} \left(\left\{\mathbf {e} _ {\mathbf {v}}, \mathbf {e} _ {\mathbf {c}} \right\}\right), \tag {3} +$$ + +where $\{\}$ denotes that $\mathbf{e_v}$ and $\mathbf{e_c}$ form a sequence, and $f_{\psi}$ represents the transformer encoders with parameter $\psi$ + +Co-attention Transformer. Co-attention [23] is another common method for exchanging information between modalities, allowing for mutual attention between video and caption. After this co-attentional transformer layer, we include $L$ transformer layers to model temporal information: + +$$ +\operatorname {C o A t t n} \left(\mathbf {e} _ {\mathbf {v}}, \mathbf {e} _ {\mathbf {c}}\right) = f _ {\phi_ {2}} \left(f _ {\phi_ {1}} \left(\left\{\mathbf {e} _ {\mathbf {v}}, \mathbf {e} _ {\mathbf {c}} \right\}\right)\right), \tag {4} +$$ + +where $f_{\phi_1}$ is the co-attentional transformer with parameter $\phi_1$ and $f_{\phi_2}$ is the transformer encoders with parameter $\phi_2$ . + +The video-caption interaction module generates frame embeddings that can be further processed based on the type of matching needed. For global matching, the frame embeddings can be averaged to obtain a single video representation. Alternatively, for fine-grained matching, the individual frame embeddings can be retained. + +# 2.5. Complementary Query-Caption Matching + +Besides using the caption for data augmentation and video feature enhancement, it can also directly represent the video content, allowing for text-text retrieval. Specifically, each of the $C$ captions generated by the video is then passed through the caption encoder to obtain its [CLS] text embedding. These caption embeddings are then aggregated to form a global representation, as illustrated in Figure 2(c). The cosine similarity between this global caption embedding and the global query embedding is then calculated to complement the query-video matching. + +![](images/5b5f3522b5e8a75ce629e18e42860268fcf62a74838d1f22f4eb6b0a07ae1473.jpg) +(a) Sum + +![](images/8ff8f77e0760782edf84a34a5fbe4b5bacb5620eee20a3123f8826c7987e564c.jpg) +(b) MLP + +![](images/f83a8122a918f22cf8f8b069a044a7a992d34651603c0f088a8751cbb80203aa.jpg) +(c) Cross Transformer +Figure 3. Illustration of four Video-Caption interaction strategies. The enhanced frame embeddings will be followed by a mean pooling for global matching or will remain for fine-grained matching. + +![](images/407c3f1c9cbf64b35e7a8775624dd74fe29a5ff0d72c8ccac198a8deab4519a0.jpg) +(d) Co-attention Transformer +Caption Embedding +Frame Embedding + +Notation. Let $\{\mathbf{e}_{\mathbf{v}_i},\mathbf{e}_{\mathbf{t}_i},\mathbf{e}_{\mathbf{c}_i}\}_{i = 1}^B$ be a batch of $B$ triples, where $\mathbf{e}_{\mathbf{v}_i}$ , $\mathbf{e}_{\mathbf{t}_i}$ , and $\mathbf{e}_{\mathbf{c}_i}$ denote the $i$ -th video, query, and caption embedding, respectively. Note that the term "embedding" used here is more general for convenience and can vary in meaning depending on the situation. For instance, in query-video global matching, $\mathbf{e}_{\mathbf{v}_i}$ and $\mathbf{e}_{\mathbf{t}_i}$ represent the averaged video feature and global [CLS] text feature, respectively. In query-video fine-grained matching, $\mathbf{e}_{\mathbf{v}_i}$ and $\mathbf{e}_{\mathbf{t}_i}$ represent a sequence of frame embeddings and a sequence of word embeddings, respectively. In query-caption matching, $\mathbf{e}_{\mathbf{c}_i}$ represents a sequence of caption embeddings, and $\mathbf{e}_{\mathbf{t}_i}$ represents a global [CLS] text feature. + +Learning Objectives. For the Query-Caption branch, we want the caption embedding $\mathbf{e_c}$ and the query embedding $\mathbf{e_t}$ to be close while they are related and far apart when they are not during training phase. We follow the common practice [24, 36] to consider the bidirectional learning objective. We employ symmetric cross-entropy loss to maximize the similarity between matched Query-Caption pairs and minimize the similarity for other pairs: + +$$ +\mathcal {L} _ {Q 2 C} = - \frac {1}{B} \sum_ {i} ^ {B} \log \frac {\exp \left(s _ {q c} \left(\mathbf {e} _ {\mathbf {t} i} , \mathbf {e} _ {\mathbf {c} i}\right) / \tau\right)}{\sum_ {j} ^ {B} \exp \left(s _ {q c} \left(\mathbf {e} _ {\mathbf {t} i} , \mathbf {e} _ {\mathbf {c} j}\right) / \tau\right)}, +$$ + +$$ +\mathcal {L} _ {C 2 Q} = - \frac {1}{B} \sum_ {i} ^ {B} \log \frac {\exp \left(s _ {q c} \left(\mathbf {e} _ {\mathbf {t} i} , \mathbf {e} _ {\mathbf {c} i}\right) / \tau\right)}{\sum_ {j} ^ {B} \exp \left(s _ {q c} \left(\mathbf {e} _ {\mathbf {t} j} , \mathbf {e} _ {\mathbf {c} i}\right) / \tau\right)}, \tag {5} +$$ + +$$ +\mathcal {L} _ {Q C} = \frac {1}{2} \left(\mathcal {L} _ {Q 2 C} + \mathcal {L} _ {C 2 Q}\right), +$$ + +where $s_{qc}(\cdot, \cdot)$ represents the query-caption matching similarity function shown in Figure 2(c), and $\tau$ refers to the temperature hyper-parameter for scaling. Similarly, the con + +trastive loss for Query-Video branch is formulated as: + +$$ +\begin{array}{l} \mathcal {L} _ {Q 2 V} = - \frac {1}{B} \sum_ {i} ^ {B} \log \frac {\exp (s _ {q v} (\mathbf {e} _ {\mathbf {t} i} , \mathbf {e} _ {\mathbf {v} i}) / \tau)}{\sum_ {j} ^ {B} \exp (s _ {q v} (\mathbf {e} _ {\mathbf {t} i} , \mathbf {e} _ {\mathbf {v} j}) / \tau)}, \\ \mathcal {L} _ {V 2 Q} = - \frac {1}{B} \sum_ {i} ^ {B} \log \frac {\exp \left(s _ {q v} \left(\mathbf {e} _ {\mathbf {t} i} , \mathbf {e} _ {\mathbf {v} i}\right) / \tau\right)}{\sum_ {j} ^ {B} \exp \left(s _ {q v} \left(\mathbf {e} _ {\mathbf {t} j} , \mathbf {e} _ {\mathbf {v} i}\right) / \tau\right)}, \tag {6} \\ \mathcal {L} _ {Q V} = \frac {1}{2} (\mathcal {L} _ {Q 2 V} + \mathcal {L} _ {V 2 Q}), \\ \end{array} +$$ + +where $s_{qv}(\cdot, \cdot)$ represents the query-video matching (e.g., global matching, fine-grained matching) similarity function shown in Figure 2(b). The total loss $\mathcal{L}$ is the sum of Query-Video loss $\mathcal{L}_{QV}$ and Query-Caption loss $\mathcal{L}_{QC}$ : + +$$ +\mathcal {L} = \mathcal {L} _ {Q V} + \mathcal {L} _ {Q C}. \tag {7} +$$ + +# 3. Experiments: Text-Video Retrieval + +# 3.1. Setups + +Datasets. We conduct experiment on four popular benchmarks for video-to-text retrieval and text-to-video retrieval tasks. MSR-VTT [44] contains a total of 10K video clips, each having 20 captions. Following the data splits from [10,24,27], we train models with associated captions on the Training-9K set and report results on the test 1K-A set. DiDeMo [1] has 10K videos paired with 40K descriptions. Following previous works [2, 19, 24], we concatenate all descriptions of one video to a single query, acting as a video-paragraph retrieval task. VATEX [38] collects $\sim 35\mathrm{K}$ videos, each with multiple annotations. There are $\sim 26\mathrm{K}$ videos for training, 1,500 videos for validation and 1,500 videos for testing. MSVD [43] contains 1,970 videos with 80K captions, with $\sim 40$ captions on average per video. There are 1,200, 100, and 670 videos in the train, validation, and test sets, respectively. + +Evaluation Metrics. For brevity, we abbreviate Recall at $K$ to $\mathbb{R}@\mathbb{K}$ ( $K = 1,5,10$ ) upon all datasets, which computes the percentage of correct videos among the top $K$ retrieved videos given textual queries (Text→Video, and vice versa). MdR, Median Rank, computes the median of the ground-truth in the retrieval ranking list. MnR, Mean Rank, computes the mean rank of the correct results in the retrieval ranking list. Note that for MdR and MnR, the lower score means the better (indicated as $\downarrow$ ). + +Implementation Details. All experiments use the visual encoder in CLIP [30] as the video encoder, and the textual encoder in CLIP as both the caption encoder and query encoder. The caption encoder and query encoder share parameters. To reduce conflict between the two branches, the query-video branch is trained first, followed by the query-caption branch. The text length is fixed to 32, and the video length is fixed to 12 for all datasets except DiDeMo (64 max words and 64 frames). The initial learning rate is set to 1e-7 + +for the clip parameters and 1e-4 for the non-clip parameters. The model is trained with a batch size of 128 for 5 epochs, except for DiDeMo (15 epochs), using the Adam [18] optimizer. All learning rates follow the cosine schedule with a linear warmup [14] strategy. For the number of generated captions per video, we set $C$ to 30. The interaction module employs $L$ transformer layers, where $L$ is set to 4 for VA-TEX and MSR-VTT, and 1 for Didemo and MSVD. In the caption branch, the number of transformer layers is set to 2. For caption generation, we directly use the original pretrained CLIP and GPT-2, without any additional tuning. + +# 3.2. Comparison with State-of-the-Arts + +In this section, we compare our Cap4Video with recent state-of-the-art methods on four benchmarks: MSR-VTT [44], MSVD [43], VATEX [38], and DiDeMo [1]. + +Table 1 shows the comparisons on DiDeMo, where our Cap4Video outperforms CLIP4Clip [24] by a significant margin of $9.2\%$ in R@1 and exceeds DRL [36] by $3.0\%$ , demonstrating the effectiveness of our method. + +Table 2 provides a comparison of our approach with recent state-of-the-art models on MSR-VTT. Our method achieves new state-of-the-art performance on text-to-video retrieval for both ViT-B/32 and ViT-B/16 backbones, significantly surpassing previous works. For instance, we achieve a $+4.8\%$ higher R@1 than CLIP4Clip with the same ViT-B/32 on text-to-video retrieval. Additionally, our Cap4Video outperforms the recent TS2-Net [22] by $2.3\%$ and $2.0\%$ with ViT-B/32 and ViT-B/16, respectively. + +Table 3 and Table 4 show the results for the MSVD and VATEX datasets, respectively, where we use ViT-B/16 as our backbone. For MSVD, our Cap4Video achieves a remarkable performance of $51.8\%$ R@1 and outperforms CLIP-based models CLIP4Clip [24] and X-Pool [25] by $6.6\%$ and $4.6\%$ on text-to-video retrieval, respectively. For VATEX, our approach also outperforms the recent state-of-the-art methods and achieves a $+7.5\%$ R@1 improvement over TS2-Net [22] for text-to-video retrieval. + +In recent studies, several methods have been proposed + +
MethodR@1R@5R@10MdRMnR
CE [21]15.640.9-8.2-
ClipBERT [19]21.147.361.16.3-
Frozen [2]31.059.872.43.0-
TMVM [20]36.564.975.43.0-
CLIP4Clip [24]42.868.579.22.018.9
TS2-Net [22]41.871.682.02.014.8
HunYuan [28]45.075.683.42.012.0
DRL [36]49.076.584.52.0-
Cap4Video52.079.487.5110.5
+ +Table 1. Results of text-to-video retrieval on the DiDeMo [1]. + +
MethodVenueText → VideoVideo → Text
R@1R@5R@10MdR↓MnR↓R@1R@5R@10MdR↓MnR↓
ClipBERT [19]CVPR'2022.046.859.96.0-----
MMT [10]ECCV'2026.657.169.64.0-27.057.569.73.721.3
T2VLAD [39]CVPR'2129.559.070.14.0-31.860.071.13.0
SupportSet [29]ICLR'2130.158.569.33.0-28.558.671.63.0-
Frozen [2]ICCV'2132.561.571.23.0------
BridgeFormer [12]CVPR'2237.664.875.1-------
TMVM [20]NeurIPS'2236.264.275.73.0-34.863.873.73.0-
CLIP-ViT-B/32
CLIP4Clip [24]arXiv'2144.571.481.62.015.342.770.980.62.011.6
CenterCLIP [48]SIGIR'2244.271.682.12.015.142.871.782.22.010.9
CAMoE [7]arXiv'2144.672.681.82.013.345.172.483.12.010.0
CLIP2Video [9]arXiv'2145.672.681.72.014.643.572.382.12.010.2
X-Pool [13]CVPR'2246.972.882.22.014.3-----
QB-Norm [4]CVPR'2247.273.083.02.0------
TS2-Net [22]ECCV'2247.074.583.82.013.045.374.183.72.09.2
DRL [36]arXiv'2247.474.683.82.0-45.373.983.32.0-
Cap4Video49.374.383.82.012.047.173.784.32.08.7
CLIP-ViT-B/16
CLIP2TV [11]arXiv'2148.374.682.82.014.946.575.484.92.010.2
CenterCLIP [48]SIGIR'2248.473.882.02.013.847.775.083.32.010.2
TS2-Net [22]ECCV'2249.475.685.32.013.546.675.984.92.08.9
DRL [36]arXiv'2250.276.584.71.0-48.976.385.42.0-
Cap4Video51.475.783.91.012.449.075.285.02.08
+ +Table 2. Retrieval results on the validation set of MSR-VTT 1K [44]. Here we report results without any post-processing operations (e.g., DSL [7] or QB-Norm [4]) during inference. + +
MethodR@1R@5R@10MdRMnR
CE [21]19.849.063.86.0-
SUPPORT [29]28.460.072.94.0-
CLIP [30]37.064.173.83.0-
Frozen [2]33.764.776.33.0-
TMVM [20]36.767.481.32.5-
CLIP4Clip [24]45.275.584.32.010.3
X-Pool [13]47.277.486.02.09.3
Cap4Video51.880.888.318.3
+ +Table 3. Results of text-to-video retrieval on the MSVD [43]. + +
MethodR@1R@5R@10MdRMnR
HGR [6]35.173.583.52.0-
CLIP [30]39.772.382.22.012.8
SUPPORT [29]44.982.189.71.0-
CLIP4Clip [24]55.989.295.01.03.9
Clip2Video [9]57.390.095.51.03.6
QB-Norm [4]58.888.393.81.0-
TS2-Net [22]59.190.095.21.03.5
Cap4Video66.693.197.012.7
+ +Table 4. Results of text-to-video retrieval on the VATEX [38]. + +to improve text-video retrieval performance by adjusting similarity during inference using other query information. Notably, our results adheres to the standard retrieval logic, where the most relevant video is retrieved for each query from a set of videos, without any knowledge of the relationship between other queries and videos. Therefore, all results reported in our tables do not involve any post-processing procedures such as DSL [7] and QB-Norm [4]. Overall, the consistent state-of-the-art performance across four benchmarks demonstrates the effectiveness of our Cap4Video. + +# 3.3. Ablation Study + +In this section, we provide detailed ablation studies to clarify the effects of each part of our design. + +Auxiliary Caption as Data Augmentation. We begin by investigating the impact of captions on data augmentation for training. In a real-world scenario, the original video title would naturally serve as an additional auxiliary caption. Therefore, we manually extracted the title from the video's original webpage and compared it to the caption generated by the GPT-2 model. Table 5 presents the re + +
MethodGlobal MatchingFine-grained Matching
R@1R@5R@10MdR↓MnR↓R@1R@5R@10MdR↓MnR↓
Baseline42.870.479.0216.645.773.782.6213.1
+Different Sources of Caption as Data Augmentation
Video Title from Source URL43.871.180.9215.144.372.783.5213.1
Zero-shot Video Captioning44.270.781.5216.246.372.581.7212.9
+Different Number of Captions for Data Augmentation
Top-144.270.781.5216.246.372.581.7212.9
Top-343.371.781.6215.045.573.882.4212.7
Top-543.470.680.4216.245.672.782.7212.9
+Video-Caption Feature Interaction
Video Only44.270.781.5216.246.372.581.7212.9
Sum43.871.580.3216.147.273.382.8213.1
Concat-MLP37.566.178.4315.740.068.779.9212.7
Cross Transformer44.671.680.3214.647.975.483.0211.5
Co-attention Transformer45.371.280.9215.048.574.082.5212.7
+Query-Caption Matching Score
Query-Video Only45.371.280.9215.048.574.082.5212.7
Query-Caption Only30.355.267.5426.430.355.267.5426.4
Query-Video + Query-Caption45.671.781.2214.849.374.283.4212.1
+ +sults of using different sources of the caption, from which we observe that using captions generated by the web-scale model as data augmentation for training can lead to direct improvements in R@1 $(+1.4\%, +0.6\%)$ under both matching mechanisms. Using video titles can also bring a $1\%$ improvement for global matching. + +We also explore the impact of the number of generated captions used for augmentation. We used the caption filtering mechanism mentioned in Sec. 2.3 to rank the relevance of captions and the ground-truth query, and selected different numbers of captions for training. The results demonstrated that using only one caption is sufficient. + +Benefit on Both Online and Offline Videos Scenarios. Our method is applicable for both online and offline videos. Offline videos are local videos with no title, while online videos have a title on the video website. Therefore, for online videos, the step of generating a caption with $\mathrm{CLIP + GPT - 2}$ is skipped, and the website title is used directly as a caption. In Table 6, the results show that our method has significantly improved over the global matching baseline for both offline and online videos. + +Video-Caption Feature Interaction. As mentioned in Sec. 2.4, we have designed four approaches for Video-Caption feature interaction. Based on the results presented in Table 5, we can conclude the following: 1) The basic approach of Sum has been shown to enhance fine-grained matching by $0.9\%$ R@1, but there is no noticeable improvement in global matching. 2) The MLP approach is difficult + +Table 5. Component-wise evaluation of our framework on the MSR-VTT 1K validation set. With the ViT-B/32 backbone, we report the text-to-video retrieval results for two representative Query-Video matching mechanisms: global matching and fine-grained matching. The consistent improvement on two typical matching mechanisms demonstrates the generalization ability and effectiveness of our method. + +
Caption SourceR@1
BaselineN/A42.8
Cap4VideoOriginal Website Title45.8
Captioner (CLIP+GPT-2)45.6
+ +Table 6. Exploring the effectiveness of captions from different sources on MSR-VTT 1k-A. Setting: ViT-B/32, global matching. + +to optimize and performs poorly in both matching scenarios. We speculate that the MLP's operation in a black-box environment, despite creating a nonlinear metric space, may lead to degradation. 3) The Cross Transformer approach has demonstrated improvements of $+0.4\%$ and $+1.6\%$ in two matching settings, respectively. These enhancements may be attributed to the self-attention mechanism's ability to capture the inter-modal relationship between the video and caption. 4) Moreover, the Co-attention Transformer approach has significantly boosted performance, with gains of $+1.1\%$ and $+2.2\%$ for these two matching mechanisms. In summary, the results indicate that proper interaction between the video and generated caption can lead to better video representation and improved Query-Video matching. + +Query-Caption matching. We also investigate the Query-Caption matching branch for text-video retrieval. We aggregate caption embeddings using mean pooling to yield a global embedding. As shown in Table 5, the single Query-Caption matching branch achieves $30.3\%$ R@1 on text-to + +Query7765: a person is discussing a car. + +
VideoRank+ CaptionRank
6video of a car camera recording the driver's voice.1
4video showing the car in a parking spot.2
5video of SUV in the video below shows a salesman talking to an audience.3
+ +Query9616: person is recording the brown horse which is having fun. + +
VideoRank 2+ Caption video of the horse jumping over a fence at Ranch in Nevada was captured on camera.Rank 1
1video showing animation of a horse's simulation, which simulates the game.2
4video showing a horse simulation video game in which you could see your avatar being animated by the camera.3
+ +Figure 4. The text-video results on the MSR-VTT 1K-A test set. + +Left: The ranking results of the query-video matching model. + +Right: The ranking results of Cap4Video, which incorporates generated captions to enhance retrieval. Please zoom in for best view. + +video retrieval, outperforming previous query-video matching methods such as ClipBERT [19] (22.0%) and MMT [10] (26.6%). This suggests that the Query-Caption matching branch can complement the regular Query-Video matching branch for improved text-video retrieval. Combining the score of Query-Caption matching branch with Query-Video matching branch further improves performance (+0.8%). + +Overall, Cap4Video utilizes generated captions in three ways: input data augmentation, intermediate feature interaction, and output score fusion, leading to consistent improvements $(+2.8\% / +3.6\%)$ in both matching mechanisms. + +# 3.4. Visualization + +We provide two examples of videos retrieved by our Cap4Video method and a model without auxiliary captions. As illustrated in Figure 4, our approach successfully retrieves the ground-truth video with the assistance of the caption, while the video-only model returns multiple videos that are somewhat relevant to the query but not precise. See more qualitative results in Supplementary. + +# 4. Related Works + +Zero-shot Image Captioning. In the field of natural language processing, transformer-based GPT models [5, 31] have been successful in generating text from prompts by training on large-scale text corpora. Similarly, CLIP [30], a vision-language alignment model trained on 400 million + +image-text pairs, has demonstrated impressive zero-shot performance on vision tasks. However, research on transferring web-scale models to zero-shot video captioning remains limited. Recently, ZeroCap [34] proposed a method of using CLIP and the GPT-2 language model to generate textual descriptions of input images, leveraging knowledge from both models in a truly zero-shot manner without retraining or fine-tuning model parameters. MAGIC [32] has also used CLIP scores to align GPT-2 logits with corresponding images but requires fine-tuning on the MS-COCO caption text corpus. More recently, a study [33] extended the zero-shot capability of ZeroCap to the video domain. In this paper, we employ this video extension to generate auxiliary captions without any additional training. + +Text-Video Retrieval aims to retrieve relevant video content based on natural language descriptions. Early studies [6, 10, 21, 37, 39] focused on knowledge transfer from "expert" models and captured intra-modal and cross-modal interactions based on pre-extracted features. However, the performance of these methods is limited since they cannot perform end-to-end optimization. Recently, more methods have involved end-to-end training for text-video retrieval. One typical approach [2, 26, 27] is to first perform largescale text-video pre-training, then transfer the model to downstream text-video retrieval tasks. With the emergence of pre-trained Vision-Language Models (VLMs), there have been increased efforts to leverage them for improving video understanding [41, 42]. Thus, another training-efficient line is to directly expand the pre-trained VLM to the text-video retrieval task. CLIPBERT [19] enables affordable pioneering end-to-end training with a sparse sampling strategy. After that, recent works [3, 4, 8, 9, 11, 13, 22, 24, 40, 48] focus on transferring knowledge from CLIP models that have been pre-trained on 400M image-text pairs. The research path has evolved from the most direct global matching (i.e., video-sentence alignment [11, 24]) to fine-grained matching (e.g., frame-word alignment [36], video-word alignment [13], multi-hierarchical alignment [9, 28]). Unlike above CLIP-Based efforts on query-video matching, we propose to generate auxiliary captions from videos to improve text-video retrieval. Thus our method is compatible with both global and fine-grained matching. + +# 5. Conclusion + +We introduce Cap4Video, a novel framework that leverages captions generated by web-scale language models to enhance text-video matching in three key ways: 1) Input data augmentation for training, 2) Intermediate videocaption feature interaction for compact video representations, and 3) Output score fusion for improved text-video retrieval. Our approach demonstrates consistent performance gains on four standard text-video retrieval benchmarks, outperforming state-of-the-art methods by a clear margin. + +# References + +[1] Lisa Anne Hendricks, Oliver Wang, Eli Shechtman, Josef Sivic, Trevor Darrell, and Bryan Russell. Localizing moments in video with natural language. In ICCV, pages 5803-5812, 2017. 2, 5 +[2] Max Bain, Arsha Nagrani, Gül Varol, and Andrew Zisserman. Frozen in time: A joint video and image encoder for end-to-end retrieval. In ICCV, pages 1728-1738, 2021. 5, 6, 8 +[3] Max Bain, Arsha Nagrani, Gül Varol, and Andrew Zisserman. A clip-hitchhiker's guide to long video retrieval. arXiv preprint arXiv:2205.08508, 2022. 8 +[4] Simion-Vlad Bogolin, Ioana Croitoru, Hailin Jin, Yang Liu, and Samuel Albanie. Cross modal retrieval with querybank normalisation. In CVPR, pages 5194-5205, 2022. 6, 8 +[5] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. NeurIPS, 33:1877-1901, 2020. 8 +[6] Shizhe Chen, Yida Zhao, Qin Jin, and Qi Wu. Fine-grained video-text retrieval with hierarchical graph reasoning. In CVPR, pages 10638-10647, 2020. 6, 8 +[7] Xing Cheng, Hezheng Lin, Xiangyu Wu, Fan Yang, and Dong Shen. Improving video-text retrieval by multi-stream corpus alignment and dual softmax loss. arXiv preprint arXiv:2109.04290, 2021. 6 +[8] Bo Fang, Chang Liu, Yu Zhou, Min Yang, Yuxin Song, Fu Li, Weiping Wang, Xiangyang Ji, Wanli Ouyang, et al. Uatvr: Uncertainty-adaptive text-video retrieval. arXiv preprint arXiv:2301.06309, 2023. 8 +[9] Han Fang, Pengfei Xiong, Luhui Xu, and Yu Chen. Clip2video: Mastering video-text retrieval via image clip. arXiv preprint arXiv:2106.11097, 2021. 1, 2, 6, 8 +[10] Valentin Gabeur, Chen Sun, Karteek Alahari, and Cordelia Schmid. Multi-modal transformer for video retrieval. In ECCV, pages 214-229. Springer, 2020. 5, 6, 8 +[11] Zijian Gao, Jingyu Liu, Sheng Chen, Dedan Chang, Hao Zhang, and Jinwei Yuan. Clip2tv: An empirical study on transformer-based methods for video-text retrieval. arXiv preprint arXiv:2111.05610, 2021. 1, 2, 6, 8 +[12] Yuying Ge, Yixiao Ge, Xihui Liu, Dian Li, Ying Shan, Xiaohu Qie, and Ping Luo. Bridging video-text retrieval with multiple choice questions. In CVPR, pages 16167-16176, 2022. 6 +[13] Satya Krishna Gorti, Noel Vouitsis, Junwei Ma, Keyvan Golestan, Maksims Volkovs, Animesh Garg, and Guangwei Yu. X-pool: Cross-modal language-video attention for text-video retrieval. In CVPR, pages 5006-5015, 2022. 1, 6, 8 +[14] Priya Goyal, Piotr Dólar, Ross Girshick, Pieter Noordhuis, Lukasz Wesolowski, Aapo Kyrola, Andrew Tulloch, Yangqing Jia, and Kaiming He. Accurate, large minibatch sgd: Training imagenet in 1 hour. arXiv preprint arXiv:1706.02677, 2017. 5 +[15] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom + +Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In ICML, pages 4904-4916. PMLR, 2021. 1 +[16] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In ICML, pages 4904-4916. PMLR, 2021. 2 +[17] Omar Khattab and Matei Zaharia. Colbert: Efficient and effective passage search via contextualized late interaction over bert. In Proceedings of the 43rd International ACM SIGIR conference on research and development in Information Retrieval, pages 39-48, 2020. 2 +[18] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In ICLR, 2015. 5 +[19] Jie Lei, Linjie Li, Luowei Zhou, Zhe Gan, Tamara L Berg, Mohit Bansal, and Jingjing Liu. Less is more: Clipbert for video-and-language learning via sparse sampling. In CVPR, pages 7331-7341, 2021. 5, 6, 8 +[20] Chengzhi Lin, Ancong Wu, Junwei Liang, Jun Zhang, Wenhang Ge, Wei-Shi Zheng, and Chunhua Shen. Text-adaptive multiple visual prototype matching for video-text retrieval. In NeurIPS, 2022. 5, 6 +[21] Yang Liu, Samuel Albanie, Arsha Nagrani, and Andrew Zisserman. Use what you have: Video retrieval using representations from collaborative experts. arXiv preprint arXiv:1907.13487, 2019. 5, 6, 8 +[22] Yuqi Liu, Pengfei Xiong, Luhui Xu, Shengming Cao, and Qin Jin. Ts2-net: Token shift and selection transformer for text-video retrieval. In ECCV, pages 319-335. Springer, 2022. 5, 6, 8 +[23] Jiasen Lu, Dhruv Batra, Devi Parikh, and Stefan Lee. Vilbert: Pretraining task-agnostic visiolinguistic representations for vision-and-language tasks. NeurIPS, 32, 2019. 4 +[24] Huaishao Luo, Lei Ji, Ming Zhong, Yang Chen, Wen Lei, Nan Duan, and Tianrui Li. Clip4clip: An empirical study of clip for end to end video clip retrieval and captioning. Neurocomputing, 508:293-304, 2022. 1, 2, 4, 5, 6, 8 +[25] Yiwei Ma, Guohai Xu, Xiaoshuai Sun, Ming Yan, Ji Zhang, and Rongrong Ji. X-clip: End-to-end multi-grained contrastive learning for video-text retrieval. In ACM MM, pages 638–647, 2022. 5 +[26] Antoine Miech, Jean-Baptiste Alayrac, Lucas Smaira, Ivan Laptev, Josef Sivic, and Andrew Zisserman. End-to-end learning of visual representations from uncurated instructional videos. In CVPR, pages 9879-9889, 2020. 8 +[27] Antoine Miech, Dimitri Zhukov, Jean-Baptiste Alayrac, Makarand Tapaswi, Ivan Laptev, and Josef Sivic. Howto100m: Learning a text-video embedding by watching hundred million narrated video clips. In ICCV, pages 2630-2640, 2019. 5, 8 +[28] Shaobo Min, Weijie Kong, Rong-Cheng Tu, Dihong Gong, Chengfei Cai, Wenzhe Zhao, Chenyang Liu, Sixiao Zheng, Hongfa Wang, Zhifeng Li, et al. Hunyuan_tvr for text-video retrivial. arXiv preprint arXiv:2204.03382, 2022. 1, 5, 8 +[29] Mandela Patrick, Po-Yao Huang, Yuki Asano, Florian Metze, Alexander G Hauptmann, Joao F Henriques, and An + +drea Vedaldi. Support-set bottlenecks for video-text representation learning. In ICLR, 2021. 6 +[30] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, pages 8748-8763. PMLR, 2021. 1, 2, 3, 5, 6, 8 +[31] Alec Radford, Jeff Wu, Rewon Child, David Luan, Dario Amodei, and Ilya Sutskever. Language models are unsupervised multitask learners. 2019. 1, 2, 3, 8 +[32] Yixuan Su, Tian Lan, Yahui Liu, Fangyu Liu, Dani Yogatama, Yan Wang, Lingpeng Kong, and Nigel Collier. Language models can see: Plugging visual controls in text generation. arXiv preprint arXiv:2205.02655, 2022. 8 +[33] Yoad Tewel, Yoav Shalev, Roy Nadler, Idan Schwartz, and Lior Wolf. Zero-shot video captioning with evolving pseudotokens. arXiv preprint arXiv:2207.11100, 2022. 2, 3, 8 +[34] Yoad Tewel, Yoav Shalev, Idan Schwartz, and Lior Wolf. Zerocap: Zero-shot image-to-text generation for visual-semantic arithmetic. In CVPR, pages 17918-17928, 2022. 2, 3, 8 +[35] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In NeurIPS, pages 5998-6008, 2017. 4 +[36] Qiang Wang, Yanhao Zhang, Yun Zheng, Pan Pan, and Xian-Sheng Hua. Disentangled representation learning for text-video retrieval. arXiv preprint arXiv:2203.07111, 2022. 1, 2, 4, 5, 6, 8 +[37] Wenzhe Wang, Mengdan Zhang, Runnan Chen, Guanyu Cai, Penghao Zhou, Pai Peng, Xiaowei Guo, Jian Wu, and Xing Sun. Dig into multi-modal cues for video retrieval with hierarchical alignment. In *IJCAI*, pages 1113–1121, 2021. 8 +[38] Xin Wang, Jiawei Wu, Junkun Chen, Lei Li, Yuan-Fang Wang, and William Yang Wang. Vatex: A large-scale, high-quality multilingual dataset for video-and-language research. In ICCV, pages 4581-4591, 2019. 2, 5, 6 +[39] Xiaohan Wang, Linchao Zhu, and Yi Yang. T2vlad: global-local sequence alignment for text-video retrieval. In CVPR, pages 5079-5088, 2021. 6, 8 +[40] Xiaohan Wang, Linchao Zhu, Zhedong Zheng, Mingliang Xu, and Yi Yang. Align and tell: Boosting text-video retrieval with local alignment and fine-grained supervision. IEEE TMM, 2022. 8 +[41] Wenhao Wu, Zhun Sun, and Wanli Ouyang. Revisiting classifier: Transferring vision-language models for video recognition. In AAAI, 2023. 8 +[42] Wenhao Wu, Xiaohan Wang, Haipeng Luo, Jingdong Wang, Yi Yang, and Wanli Ouyang. Bidirectional cross-modal knowledge exploration for video recognition with pretrained vision-language models. In CVPR, 2023. 8 +[43] Zuxuan Wu, Ting Yao, Yanwei Fu, and Yu-Gang Jiang. Deep learning for video classification and captioning. In Frontiers of multimedia research, pages 3-29. 2017. 2, 5, 6 +[44] Jun Xu, Tao Mei, Ting Yao, and Yong Rui. Msr-vtt: A large video description dataset for bridging video and language. In CVPR, pages 5288-5296, 2016. 2, 5, 6 + +[45] Lewei Yao, Runhui Huang, Lu Hou, Guansong Lu, Minzhe Niu, Hang Xu, Xiaodan Liang, Zhenguo Li, Xin Jiang, and Chunjing Xu. Filip: Fine-grained interactive language-image pre-training. In ICLR, 2021. 2 +[46] Jiahui Yu, Zirui Wang, Vijay Vasudevan, Legg Yeung, Mojtaba Seyhedhosseini, and Yonghui Wu. Coca: Contrastive captioners are image-text foundation models. arXiv preprint arXiv:2205.01917, 2022. 1 +[47] Lu Yuan, Dongdong Chen, Yi-Ling Chen, Noel Codella, Xiyang Dai, Jianfeng Gao, Houdong Hu, Xuedong Huang, Boxin Li, Chunyuan Li, et al. Florence: A new foundation model for computer vision. arXiv preprint arXiv:2111.11432, 2021. 1 +[48] Shuai Zhao, Linchao Zhu, Xiaohan Wang, and Yi Yang. Centerclip: Token clustering for efficient text-video retrieval. In Proceedings of the 45th International ACM SIGIR Conference on Research and Development in Information Retrieval, pages 970–981, 2022. 6, 8 \ No newline at end of file diff --git a/2023/Cap4Video_ What Can Auxiliary Captions Do for Text-Video Retrieval_/images.zip b/2023/Cap4Video_ What Can Auxiliary Captions Do for Text-Video Retrieval_/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..fc564d1a9d70eef180aa3dc3ea702daa2690ea26 --- /dev/null +++ b/2023/Cap4Video_ What Can Auxiliary Captions Do for Text-Video Retrieval_/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:32779e33c5faa6466ee18e3dae5b9fe2b57b0fccb511f7b0885d60c402045ddf +size 734256 diff --git a/2023/Cap4Video_ What Can Auxiliary Captions Do for Text-Video Retrieval_/layout.json b/2023/Cap4Video_ What Can Auxiliary Captions Do for Text-Video Retrieval_/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..407a163a9f3aeceaff861fbee1a52de1cafe1bd9 --- /dev/null +++ b/2023/Cap4Video_ What Can Auxiliary Captions Do for Text-Video Retrieval_/layout.json @@ -0,0 +1,8849 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 75, + 103, + 518, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 103, + 518, + 121 + ], + "spans": [ + { + "bbox": [ + 75, + 103, + 518, + 121 + ], + "type": "text", + "content": "Cap4Video: What Can Auxiliary Captions Do for Text-Video Retrieval?" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 73, + 141, + 518, + 199 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 73, + 141, + 518, + 199 + ], + "spans": [ + { + "bbox": [ + 73, + 141, + 518, + 199 + ], + "type": "text", + "content": "Wenhao Wu" + }, + { + "bbox": [ + 73, + 141, + 518, + 199 + ], + "type": "inline_equation", + "content": "^{1,2*}" + }, + { + "bbox": [ + 73, + 141, + 518, + 199 + ], + "type": "text", + "content": " Haipeng Luo" + }, + { + "bbox": [ + 73, + 141, + 518, + 199 + ], + "type": "inline_equation", + "content": "^{3*}" + }, + { + "bbox": [ + 73, + 141, + 518, + 199 + ], + "type": "text", + "content": " Bo Fang" + }, + { + "bbox": [ + 73, + 141, + 518, + 199 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 73, + 141, + 518, + 199 + ], + "type": "text", + "content": " Jingdong Wang" + }, + { + "bbox": [ + 73, + 141, + 518, + 199 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 73, + 141, + 518, + 199 + ], + "type": "text", + "content": " Wanli Ouyang" + }, + { + "bbox": [ + 73, + 141, + 518, + 199 + ], + "type": "inline_equation", + "content": "^{4,1}" + }, + { + "bbox": [ + 73, + 141, + 518, + 199 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 73, + 141, + 518, + 199 + ], + "type": "text", + "content": "The University of Sydney \n" + }, + { + "bbox": [ + 73, + 141, + 518, + 199 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 73, + 141, + 518, + 199 + ], + "type": "text", + "content": "Baidu Inc. \n" + }, + { + "bbox": [ + 73, + 141, + 518, + 199 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 73, + 141, + 518, + 199 + ], + "type": "text", + "content": "University of Chinese Academy of Sciences \n" + }, + { + "bbox": [ + 73, + 141, + 518, + 199 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 73, + 141, + 518, + 199 + ], + "type": "text", + "content": "Shanghai AI Laboratory \nwhwu.ucas@gmail.com" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 143, + 227, + 192, + 239 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 227, + 192, + 239 + ], + "spans": [ + { + "bbox": [ + 143, + 227, + 192, + 239 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 252, + 290, + 577 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 252, + 290, + 577 + ], + "spans": [ + { + "bbox": [ + 46, + 252, + 290, + 577 + ], + "type": "text", + "content": "Most existing text-video retrieval methods focus on cross-modal matching between the visual content of videos and textual query sentences. However, in real-world scenarios, online videos are often accompanied by relevant text information such as titles, tags, and even subtitles, which can be utilized to match textual queries. This insight has motivated us to propose a novel approach to text-video retrieval, where we directly generate associated captions from videos using zero-shot video captioning with knowledge from web-scale pre-trained models (e.g., CLIP and GPT-2). Given the generated captions, a natural question arises: what benefits do they bring to text-video retrieval? To answer this, we introduce Cap4Video, a new framework that leverages captions in three ways: i) Input data: video-caption pairs can augment the training data. ii) Intermediate feature interaction: we perform cross-modal feature interaction between the video and caption to produce enhanced video representations. iii) Output score: the Query-Caption matching branch can complement the original Query-Video matching branch for text-video retrieval. We conduct comprehensive ablation studies to demonstrate the effectiveness of our approach. Without any post-processing, Cap4Video achieves state-of-the-art performance on four standard text-video retrieval benchmarks: MSR-VTT " + }, + { + "bbox": [ + 46, + 252, + 290, + 577 + ], + "type": "inline_equation", + "content": "(51.4\\%)" + }, + { + "bbox": [ + 46, + 252, + 290, + 577 + ], + "type": "text", + "content": ", VATEX " + }, + { + "bbox": [ + 46, + 252, + 290, + 577 + ], + "type": "inline_equation", + "content": "(66.6\\%)" + }, + { + "bbox": [ + 46, + 252, + 290, + 577 + ], + "type": "text", + "content": ", MSVD " + }, + { + "bbox": [ + 46, + 252, + 290, + 577 + ], + "type": "inline_equation", + "content": "(51.8\\%)" + }, + { + "bbox": [ + 46, + 252, + 290, + 577 + ], + "type": "text", + "content": ", and DiDeMo " + }, + { + "bbox": [ + 46, + 252, + 290, + 577 + ], + "type": "inline_equation", + "content": "(52.0\\%)" + }, + { + "bbox": [ + 46, + 252, + 290, + 577 + ], + "type": "text", + "content": ". The code is available at https://github.com/whwu95/Cap4Video." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 600, + 128, + 612 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 600, + 128, + 612 + ], + "spans": [ + { + "bbox": [ + 48, + 600, + 128, + 612 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 620, + 287, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 620, + 287, + 693 + ], + "spans": [ + { + "bbox": [ + 46, + 620, + 287, + 693 + ], + "type": "text", + "content": "Text-video retrieval is a fundamental task in videolanguage learning. With the rapid advancements in image-language pre-training [15, 30, 46, 47], researchers have focused on expanding pre-trained image-language models, especially CLIP [30], to tackle the text-video retrieval task. The research path has evolved from the most direct global" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 310, + 226, + 537, + 262 + ], + "blocks": [ + { + "bbox": [ + 310, + 226, + 537, + 262 + ], + "lines": [ + { + "bbox": [ + 310, + 226, + 537, + 262 + ], + "spans": [ + { + "bbox": [ + 310, + 226, + 537, + 262 + ], + "type": "image", + "image_path": "cd8bb2a400ae5fd45da3cd46c3d9b8da397d9bca24081d14ead1b6c00fa35874.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 310, + 271, + 539, + 318 + ], + "blocks": [ + { + "bbox": [ + 310, + 271, + 539, + 318 + ], + "lines": [ + { + "bbox": [ + 310, + 271, + 539, + 318 + ], + "spans": [ + { + "bbox": [ + 310, + 271, + 539, + 318 + ], + "type": "image", + "image_path": "893e334822d01329b3b958bcb5288b1fac058e0bb14c1c4e7ad1b735432f0bb7.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 310, + 326, + 543, + 447 + ], + "blocks": [ + { + "bbox": [ + 310, + 326, + 543, + 447 + ], + "lines": [ + { + "bbox": [ + 310, + 326, + 543, + 447 + ], + "spans": [ + { + "bbox": [ + 310, + 326, + 543, + 447 + ], + "type": "image", + "image_path": "f08ac9db5530b7116bde109bf9af13292ae8bf4701bc6d49e256548d1794f9d7.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 464, + 547, + 531 + ], + "lines": [ + { + "bbox": [ + 304, + 464, + 547, + 531 + ], + "spans": [ + { + "bbox": [ + 304, + 464, + 547, + 531 + ], + "type": "text", + "content": "Figure 1. (a) An existing end-to-end learning paradigm for text-video retrieval. (b) Zero-shot video captioning achieved by guiding a large language model (LLM) such as GPT-2 [31] with CLIP [30]. (c) Our Cap4Video framework leverages the generated captions in three aspects: input data augmentation, intermediate feature interaction, and output score fusion." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 557, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 557, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 557, + 547, + 715 + ], + "type": "text", + "content": "matching (i.e., video-sentence alignment [11, 24]) to fine-grained matching (e.g., frame-word alignment [36], videoword alignment [13], multi-hierarchical alignment [9, 28], etc.). These studies have demonstrated remarkable performance and significantly outperformed previous models. Two key factors contribute to this improvement. Firstly, CLIP offers powerful visual and textual representations that are pre-aligned in the semantic embedding space, thereby reducing the challenge of cross-modal learning in videotext matching. Secondly, these methods can fine-tune the pre-trained vision and text encoders using sparsely sampled frames in an end-to-end manner. All of these methods aim to learn cross-modal alignment between the visual represen" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 702, + 126, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 702, + 126, + 712 + ], + "spans": [ + { + "bbox": [ + 58, + 702, + 126, + 712 + ], + "type": "text", + "content": "*Equal contribution." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "10704" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 95 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 95 + ], + "type": "text", + "content": "tation of videos and the textual representation of the corresponding query, as depicted in Figure 1(a)." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 96, + 289, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 96, + 289, + 323 + ], + "spans": [ + { + "bbox": [ + 46, + 96, + 289, + 323 + ], + "type": "text", + "content": "However, in real-life scenarios, online videos usually come with related content such as the video's title or tag on the video website. In addition to the visual signal in the video, the associated textual information can also be used to some extent to describe the video content and match the query (i.e., the common text-to-text retrieval). This raises a pertinent question: How can we generate associated text descriptions for videos? One possible solution is to crawl the video title from the video website. However, this method relies on annotations, and there is a risk that the video URL may have become invalid. Another automated solution is to generate captions using zero-shot video caption models. Therefore, we turn our attention to knowledge-rich pretrained models to handle such challenging open-set scenarios. We find that the recent study ZeroCap [34] provides a good practice to use frozen CLIP [30] and GPT-2 [31] for zero-shot image captioning. Thus, we leverage a video extension [33] of ZeroCap for generating captions in the video domain without any further training." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 324, + 289, + 575 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 324, + 289, + 575 + ], + "spans": [ + { + "bbox": [ + 46, + 324, + 289, + 575 + ], + "type": "text", + "content": "When provided with auxiliary captions, a natural question naturally arises: How can we leverage these captions to enhance the text-video retrieval task? In this paper, we propose the Cap4Video learning framework, as illustrated in Figure 1(c), which utilizes captions in three key ways: (i) Input Data: One simple approach is to augment the training data with the generated captions. Specifically, the given video and its generated caption can be treated as a matched pair, which serves as an additional positive sample pair for training beyond the query-video pairs. (ii) Intermediate Feature Interaction: Cross-modal interaction between the video and captions can be leveraged to improve the video representation. Specifically, we can exploit the complementary information between videos and captions to reduce redundant features from videos and learn more discriminative video representations. (iii) Output score: The generated caption can also represent the video's content, allowing us to employ query-caption matching to complement standard query-video matching for the text-video retrieval task. Moreover, a two-stream architecture can be utilized to reduce model bias and produce more robust results." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 575, + 288, + 611 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 575, + 288, + 611 + ], + "spans": [ + { + "bbox": [ + 46, + 575, + 288, + 611 + ], + "type": "text", + "content": "We hope that our novel paradigm will encourage further investigation into the video-language learning. In summary, our contributions are as follows:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 58, + 620, + 287, + 714 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 58, + 620, + 287, + 692 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 620, + 287, + 692 + ], + "spans": [ + { + "bbox": [ + 58, + 620, + 287, + 692 + ], + "type": "text", + "content": "- We explore a novel problem: leveraging auxiliary captions to further enhance existing text-video retrieval. Besides labor-intensive manual crawling of video website titles, we investigate the potential of rich captions automatically generated by large language models (LLMs) to benefit text-video retrieval." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 58, + 701, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 701, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 58, + 701, + 287, + 714 + ], + "type": "text", + "content": "- We propose the Cap4Video learning framework," + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 324, + 72, + 545, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 324, + 72, + 545, + 133 + ], + "spans": [ + { + "bbox": [ + 324, + 72, + 545, + 133 + ], + "type": "text", + "content": "which maximizes the utility of the auxiliary captions through three aspects: input data, feature interaction, and output score. Our framework improves the performance of existing query-video matching mechanisms, including global matching and fine-grained matching." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 317, + 139, + 547, + 199 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 139, + 547, + 199 + ], + "spans": [ + { + "bbox": [ + 317, + 139, + 547, + 199 + ], + "type": "text", + "content": "- Extensive experiments conducted on four video benchmarks demonstrate the effectiveness of our method. Our Cap4Video achieves state-of-the-art performance on MSR-VTT [44] (51.4%), VATEX [38] (66.6%), MSVD [43] (51.8%), and DiDeMo [1] (52.0%)." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 208, + 389, + 222 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 208, + 389, + 222 + ], + "spans": [ + { + "bbox": [ + 306, + 208, + 389, + 222 + ], + "type": "text", + "content": "2. Methodology" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 228, + 493, + 241 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 228, + 493, + 241 + ], + "spans": [ + { + "bbox": [ + 306, + 228, + 493, + 241 + ], + "type": "text", + "content": "2.1. Background: Text-Video Matching" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 246, + 547, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 246, + 547, + 365 + ], + "spans": [ + { + "bbox": [ + 304, + 246, + 547, + 365 + ], + "type": "text", + "content": "Text-video matching aims to evaluate the similarity between a given sentence " + }, + { + "bbox": [ + 304, + 246, + 547, + 365 + ], + "type": "inline_equation", + "content": "Q_{i}" + }, + { + "bbox": [ + 304, + 246, + 547, + 365 + ], + "type": "text", + "content": " and a given video " + }, + { + "bbox": [ + 304, + 246, + 547, + 365 + ], + "type": "inline_equation", + "content": "V_{j}" + }, + { + "bbox": [ + 304, + 246, + 547, + 365 + ], + "type": "text", + "content": ", typically using a similarity function " + }, + { + "bbox": [ + 304, + 246, + 547, + 365 + ], + "type": "inline_equation", + "content": "s(Q_{i},V_{j})" + }, + { + "bbox": [ + 304, + 246, + 547, + 365 + ], + "type": "text", + "content": ". In text-to-video retrieval, the goal is to rank all videos based on their similarity scores to a given query sentence. To improve text-video retrieval, recent works [9, 11, 24] have applied CLIP [30] for initialization, leveraging pre-trained knowledge from image-text learning. Our baselines for text-video matching include two typical mechanisms: global matching and fine-grained matching, as illustrated in Figure 2(b)." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 366, + 547, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 366, + 547, + 510 + ], + "spans": [ + { + "bbox": [ + 304, + 366, + 547, + 510 + ], + "type": "text", + "content": "Global Matching is a commonly used technique in cross-modal contrastive learning [16, 24, 30]. In global matching, each modality is encoded independently to obtain global features, which are then used to calculate similarity. We train the visual encoder to output " + }, + { + "bbox": [ + 304, + 366, + 547, + 510 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 304, + 366, + 547, + 510 + ], + "type": "text", + "content": " frame embeddings for a given video that samples " + }, + { + "bbox": [ + 304, + 366, + 547, + 510 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 304, + 366, + 547, + 510 + ], + "type": "text", + "content": " frames. Similarly, the query encoder returns " + }, + { + "bbox": [ + 304, + 366, + 547, + 510 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 304, + 366, + 547, + 510 + ], + "type": "text", + "content": " word embeddings and the [CLS] embedding as the global representation for a given query sentence that contains " + }, + { + "bbox": [ + 304, + 366, + 547, + 510 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 304, + 366, + 547, + 510 + ], + "type": "text", + "content": " words. The frame embeddings are integrated using average pooling to obtain the global video embedding, which is then compared with the global query embedding to calculate similarity." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 510, + 547, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 510, + 547, + 665 + ], + "spans": [ + { + "bbox": [ + 304, + 510, + 547, + 665 + ], + "type": "text", + "content": "Fine-grained Matching focuses on modeling the token-level alignment between two modalities, such as framework alignment. In order to achieve token-level patchword alignment for image-text learning, FILIP [45] and ColBERT [17] employ a Max-Mean pipeline. This pipeline finds the token-wise maximum similarity between patch and word tokens, and then averages the maximum similarity of tokens in the image or text to obtain the similarity between an image and a text or vice versa. Moreover, DRL [36] extends the token-wise alignment to text-video retrieval and introduces an attention mechanism to learn weighted pooling instead of mean pooling. We have adopted this mechanism as our enhanced baseline." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 671, + 495, + 685 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 671, + 495, + 685 + ], + "spans": [ + { + "bbox": [ + 306, + 671, + 495, + 685 + ], + "type": "text", + "content": "2.2. Preprocessing: Caption Generation" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 689, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 545, + 714 + ], + "type": "text", + "content": "To obtain auxiliary captions for a given video, we consider the following two approaches." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "10705" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 48, + 83, + 545, + 202 + ], + "blocks": [ + { + "bbox": [ + 49, + 71, + 137, + 83 + ], + "lines": [ + { + "bbox": [ + 49, + 71, + 137, + 83 + ], + "spans": [ + { + "bbox": [ + 49, + 71, + 137, + 83 + ], + "type": "text", + "content": "(a) Our Cap4Video" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 48, + 83, + 545, + 202 + ], + "lines": [ + { + "bbox": [ + 48, + 83, + 545, + 202 + ], + "spans": [ + { + "bbox": [ + 48, + 83, + 545, + 202 + ], + "type": "image", + "image_path": "5c9205f54a594b05cb3e4eb4d0fbf51423143a5bd711a36da850cfd413e19c99.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 50, + 208, + 286, + 220 + ], + "lines": [ + { + "bbox": [ + 50, + 208, + 286, + 220 + ], + "spans": [ + { + "bbox": [ + 50, + 208, + 286, + 220 + ], + "type": "text", + "content": "(b) Query-Video Matching: Two typical mechanisms" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 391, + 208, + 520, + 220 + ], + "lines": [ + { + "bbox": [ + 391, + 208, + 520, + 220 + ], + "spans": [ + { + "bbox": [ + 391, + 208, + 520, + 220 + ], + "type": "text", + "content": "(c) Query-Caption Matching" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 50, + 239, + 199, + 319 + ], + "blocks": [ + { + "bbox": [ + 50, + 222, + 171, + 236 + ], + "lines": [ + { + "bbox": [ + 50, + 222, + 171, + 236 + ], + "spans": [ + { + "bbox": [ + 50, + 222, + 171, + 236 + ], + "type": "text", + "content": "① Global embedding matching" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 50, + 239, + 199, + 319 + ], + "lines": [ + { + "bbox": [ + 50, + 239, + 199, + 319 + ], + "spans": [ + { + "bbox": [ + 50, + 239, + 199, + 319 + ], + "type": "image", + "image_path": "43d5630e5396a15b2a910fed787ac0af2fcc7fefd165e0035646efcd40e5ff13.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 332, + 546, + 398 + ], + "lines": [ + { + "bbox": [ + 46, + 332, + 546, + 398 + ], + "spans": [ + { + "bbox": [ + 46, + 332, + 546, + 398 + ], + "type": "text", + "content": "Figure 2. An overview of our Cap4Video for text-video retrieval. We first generate captions using a zero-shot video captioner that combines CLIP [30] with GPT-2 [31], leveraging knowledge from both frozen web-scale models. We then utilize the pre-extracted caption information from three different perspectives: i) Input data: We use the video and captions to create new positive pairs for data augmentation during training. ii) Feature interaction: We perform feature interaction between video and caption to capture intra- and inter-modality context, yielding enhanced video representations. iii) Output score: The Query-Caption matching branch can complement the original Query-Video matching branch for text-video retrieval." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 202, + 239, + 386, + 318 + ], + "blocks": [ + { + "bbox": [ + 202, + 222, + 346, + 235 + ], + "lines": [ + { + "bbox": [ + 202, + 222, + 346, + 235 + ], + "spans": [ + { + "bbox": [ + 202, + 222, + 346, + 235 + ], + "type": "text", + "content": "② Fine-grained embedding matching" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 202, + 239, + 386, + 318 + ], + "lines": [ + { + "bbox": [ + 202, + 239, + 386, + 318 + ], + "spans": [ + { + "bbox": [ + 202, + 239, + 386, + 318 + ], + "type": "image", + "image_path": "05872dffca63a7c7939be9d6844c32c5f148399b350605019d22356d859b3ce4.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 398, + 236, + 545, + 315 + ], + "blocks": [ + { + "bbox": [ + 369, + 223, + 451, + 234 + ], + "lines": [ + { + "bbox": [ + 369, + 223, + 451, + 234 + ], + "spans": [ + { + "bbox": [ + 369, + 223, + 451, + 234 + ], + "type": "text", + "content": "Captions embeddings" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 398, + 236, + 545, + 315 + ], + "lines": [ + { + "bbox": [ + 398, + 236, + 545, + 315 + ], + "spans": [ + { + "bbox": [ + 398, + 236, + 545, + 315 + ], + "type": "image", + "image_path": "29d4af8d46974b636e0b2b11ac5e56d999474addfb9339d9e2cf64bba9ad92a0.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 419, + 287, + 467 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 419, + 287, + 467 + ], + "spans": [ + { + "bbox": [ + 46, + 419, + 287, + 467 + ], + "type": "text", + "content": "Manual Crawling of Video Titles. We extract the video website title by crawling the original links (such as YouTube ID) of each video and utilize it as the caption. However, we skip this step for videos with expired links." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 468, + 287, + 612 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 468, + 287, + 612 + ], + "spans": [ + { + "bbox": [ + 46, + 468, + 287, + 612 + ], + "type": "text", + "content": "Automatic Video Captioning. In contrast to the manual approach that relies on annotations, we leverage knowledge from the LLM to generate rich and diverse captions. Given the scalability of our framework, we aim to generate captions directly from downstream videos without any additional training, a process known as zero-shot video captioning. To achieve this, we follow [33, 34] and use GPT-2 [31] to predict the next word from an initial prompt, e.g., \"Video shows\". A calibrated CLIP [30] loss is then used to drive the model to generate sentences that describe the video, incorporating video-related knowledge into the auto-regressive process. See Supplementary for more details." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 47, + 622, + 276, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 622, + 276, + 635 + ], + "spans": [ + { + "bbox": [ + 47, + 622, + 276, + 635 + ], + "type": "text", + "content": "2.3. Data Augmentation with Auxiliary Captions" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 46, + 642, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 642, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 642, + 287, + 715 + ], + "type": "text", + "content": "Auxiliary captions can be used to augment training data. For example, for a dataset consisting of " + }, + { + "bbox": [ + 46, + 642, + 287, + 715 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 642, + 287, + 715 + ], + "type": "text", + "content": " videos and their corresponding query sentences, each video and its generated caption can be considered as a positive sample pair for training, in addition to the original query-video pairs. By selecting one caption per video, we can add at least " + }, + { + "bbox": [ + 46, + 642, + 287, + 715 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 642, + 287, + 715 + ], + "type": "text", + "content": " pairs" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 420, + 499, + 431 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 420, + 499, + 431 + ], + "spans": [ + { + "bbox": [ + 306, + 420, + 499, + 431 + ], + "type": "text", + "content": "as additional data augmentation during training." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 433, + 545, + 553 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 433, + 545, + 553 + ], + "spans": [ + { + "bbox": [ + 304, + 433, + 545, + 553 + ], + "type": "text", + "content": "The automatic video captioner can generate multiple captions (e.g., 20) for each video. However, some of these captions may contain noise and may not be entirely relevant to the video content. To avoid negative effects on training, we use a filtering mechanism that evaluates the semantic similarity between each caption and the ground-truth query of the video using a pre-trained text encoder. The caption with the highest similarity is then chosen for data augmentation. Note that we only use the ground-truth query for caption filtering during the training phase." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 562, + 515, + 575 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 562, + 515, + 575 + ], + "spans": [ + { + "bbox": [ + 306, + 562, + 515, + 575 + ], + "type": "text", + "content": "2.4. Video-Caption Cross-Modal Interaction" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "content": "We further consider taking advantage of the complementarity between videos and captions to reduce redundant features and learn more discriminative video representations. To preserve the pre-trained CLIP encoder architecture for efficient transfer learning, we limit the interaction to the final caption and frame embeddings. Specifically, we pass the frame embeddings " + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{e}_{\\mathbf{v}} = \\{\\mathbf{v}_1,\\mathbf{v}_2,\\dots ,\\mathbf{v}_F\\}" + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "content": " and caption embeddings " + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{e_c} = \\{\\mathbf{c}_1,\\mathbf{c}_2,\\dots ,\\mathbf{c}_C\\}" + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "content": " to the interaction module, where " + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "content": " represent the number of frames and captions, respectively. Figure 2(b) depicts several ways of interaction between the two modalities." + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10706" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 287, + 109 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 287, + 109 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 287, + 109 + ], + "type": "text", + "content": "Sum. To obtain an enhanced frame embedding, an intuitive approach is to compute the sum of the global caption embedding " + }, + { + "bbox": [ + 47, + 72, + 287, + 109 + ], + "type": "inline_equation", + "content": "\\mathbf{c}_{\\mathbf{g}}" + }, + { + "bbox": [ + 47, + 72, + 287, + 109 + ], + "type": "text", + "content": " and each frame embedding:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 77, + 116, + 287, + 129 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 116, + 287, + 129 + ], + "spans": [ + { + "bbox": [ + 77, + 116, + 287, + 129 + ], + "type": "interline_equation", + "content": "\\operatorname {S u m} \\left(\\mathbf {v} _ {i}, \\mathbf {c} _ {\\mathbf {g}}\\right) = \\mathbf {v} _ {i} + \\mathbf {c} _ {\\mathbf {g}}, \\quad i = 1, \\dots , F, \\tag {1}", + "image_path": "87b756b808b51f8876aca2275335c7556b15b164e2a1d3f5c9c5f211b26a58c7.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 136, + 287, + 173 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 136, + 287, + 173 + ], + "spans": [ + { + "bbox": [ + 47, + 136, + 287, + 173 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 136, + 287, + 173 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_i\\in \\mathbb{R}^D" + }, + { + "bbox": [ + 47, + 136, + 287, + 173 + ], + "type": "text", + "content": " is the " + }, + { + "bbox": [ + 47, + 136, + 287, + 173 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 47, + 136, + 287, + 173 + ], + "type": "text", + "content": "-th frame embedding, and " + }, + { + "bbox": [ + 47, + 136, + 287, + 173 + ], + "type": "inline_equation", + "content": "\\mathbf{c_g}\\in \\mathbb{R}^D" + }, + { + "bbox": [ + 47, + 136, + 287, + 173 + ], + "type": "text", + "content": " is computed by averaging the [CLS] embeddings of " + }, + { + "bbox": [ + 47, + 136, + 287, + 173 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 47, + 136, + 287, + 173 + ], + "type": "text", + "content": " generated captions: " + }, + { + "bbox": [ + 47, + 136, + 287, + 173 + ], + "type": "inline_equation", + "content": "\\mathbf{c_g} = \\frac{1}{C}\\sum_{i = 1}^{C}\\mathbf{c}_i" + }, + { + "bbox": [ + 47, + 136, + 287, + 173 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 173, + 287, + 220 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 173, + 287, + 220 + ], + "spans": [ + { + "bbox": [ + 47, + 173, + 287, + 220 + ], + "type": "text", + "content": "MLP. To model weighted combinations of each frame embedding and the global caption embedding " + }, + { + "bbox": [ + 47, + 173, + 287, + 220 + ], + "type": "inline_equation", + "content": "\\mathbf{c_g}" + }, + { + "bbox": [ + 47, + 173, + 287, + 220 + ], + "type": "text", + "content": ", we concatenate them together and pass the result through a learnable Multi-layer Perceptron (MLP):" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 60, + 228, + 287, + 242 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 60, + 228, + 287, + 242 + ], + "spans": [ + { + "bbox": [ + 60, + 228, + 287, + 242 + ], + "type": "interline_equation", + "content": "\\operatorname {M L P} \\left(\\mathbf {v} _ {i}, \\mathbf {c} _ {\\mathbf {g}}\\right) = f _ {\\theta} \\left(\\left[ \\mathbf {v} _ {i}, \\mathbf {c} _ {\\mathbf {g}} \\right]\\right), \\quad \\text {f o r} i = 1, \\dots , F, \\tag {2}", + "image_path": "3c608d5e1b0e9f50c66f7a0fb9bed9c0449123201d5786c0a3fb4db52c5a66cf.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 248, + 287, + 272 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 248, + 287, + 272 + ], + "spans": [ + { + "bbox": [ + 47, + 248, + 287, + 272 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 248, + 287, + 272 + ], + "type": "inline_equation", + "content": "[\\cdot ,\\cdot ]" + }, + { + "bbox": [ + 47, + 248, + 287, + 272 + ], + "type": "text", + "content": " denotes the concatenation operation, " + }, + { + "bbox": [ + 47, + 248, + 287, + 272 + ], + "type": "inline_equation", + "content": "f_{\\theta}" + }, + { + "bbox": [ + 47, + 248, + 287, + 272 + ], + "type": "text", + "content": " is the MLP with parameter " + }, + { + "bbox": [ + 47, + 248, + 287, + 272 + ], + "type": "inline_equation", + "content": "\\theta" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 272, + 287, + 331 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 272, + 287, + 331 + ], + "spans": [ + { + "bbox": [ + 47, + 272, + 287, + 331 + ], + "type": "text", + "content": "Cross Transformer. We also investigate the use of selfattention [35] for interactions. The Cross Transformer operates on a sequence " + }, + { + "bbox": [ + 47, + 272, + 287, + 331 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{e}_{\\mathbf{v}}, \\mathbf{e}_{\\mathbf{c}}\\} = \\{\\mathbf{v}_1, \\dots, \\mathbf{v}_F, \\mathbf{c}_1, \\dots, \\mathbf{c}_C\\}" + }, + { + "bbox": [ + 47, + 272, + 287, + 331 + ], + "type": "text", + "content": " and processes them through " + }, + { + "bbox": [ + 47, + 272, + 287, + 331 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 47, + 272, + 287, + 331 + ], + "type": "text", + "content": " encoder-style transformer blocks to generate final representations:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 102, + 340, + 287, + 353 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 102, + 340, + 287, + 353 + ], + "spans": [ + { + "bbox": [ + 102, + 340, + 287, + 353 + ], + "type": "interline_equation", + "content": "\\operatorname {C r o s s} \\left(\\mathbf {e} _ {\\mathbf {v}}, \\mathbf {e} _ {\\mathbf {c}}\\right) = f _ {\\psi} \\left(\\left\\{\\mathbf {e} _ {\\mathbf {v}}, \\mathbf {e} _ {\\mathbf {c}} \\right\\}\\right), \\tag {3}", + "image_path": "a4aa60b173053cdd5ff90364ddabc0da7eecb6e9d4ba904c785f3b42128e1374.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 360, + 286, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 360, + 286, + 384 + ], + "spans": [ + { + "bbox": [ + 47, + 360, + 286, + 384 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 360, + 286, + 384 + ], + "type": "inline_equation", + "content": "\\{\\}" + }, + { + "bbox": [ + 47, + 360, + 286, + 384 + ], + "type": "text", + "content": " denotes that " + }, + { + "bbox": [ + 47, + 360, + 286, + 384 + ], + "type": "inline_equation", + "content": "\\mathbf{e_v}" + }, + { + "bbox": [ + 47, + 360, + 286, + 384 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 360, + 286, + 384 + ], + "type": "inline_equation", + "content": "\\mathbf{e_c}" + }, + { + "bbox": [ + 47, + 360, + 286, + 384 + ], + "type": "text", + "content": " form a sequence, and " + }, + { + "bbox": [ + 47, + 360, + 286, + 384 + ], + "type": "inline_equation", + "content": "f_{\\psi}" + }, + { + "bbox": [ + 47, + 360, + 286, + 384 + ], + "type": "text", + "content": " represents the transformer encoders with parameter " + }, + { + "bbox": [ + 47, + 360, + 286, + 384 + ], + "type": "inline_equation", + "content": "\\psi" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 384, + 287, + 444 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 384, + 287, + 444 + ], + "spans": [ + { + "bbox": [ + 47, + 384, + 287, + 444 + ], + "type": "text", + "content": "Co-attention Transformer. Co-attention [23] is another common method for exchanging information between modalities, allowing for mutual attention between video and caption. After this co-attentional transformer layer, we include " + }, + { + "bbox": [ + 47, + 384, + 287, + 444 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 47, + 384, + 287, + 444 + ], + "type": "text", + "content": " transformer layers to model temporal information:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 85, + 456, + 287, + 469 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 85, + 456, + 287, + 469 + ], + "spans": [ + { + "bbox": [ + 85, + 456, + 287, + 469 + ], + "type": "interline_equation", + "content": "\\operatorname {C o A t t n} \\left(\\mathbf {e} _ {\\mathbf {v}}, \\mathbf {e} _ {\\mathbf {c}}\\right) = f _ {\\phi_ {2}} \\left(f _ {\\phi_ {1}} \\left(\\left\\{\\mathbf {e} _ {\\mathbf {v}}, \\mathbf {e} _ {\\mathbf {c}} \\right\\}\\right)\\right), \\tag {4}", + "image_path": "e0c2ab07099c2f3a6fc464e8be81023c21fc35c0cb91a8fa564360baa4608f14.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 473, + 287, + 497 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 473, + 287, + 497 + ], + "spans": [ + { + "bbox": [ + 47, + 473, + 287, + 497 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 473, + 287, + 497 + ], + "type": "inline_equation", + "content": "f_{\\phi_1}" + }, + { + "bbox": [ + 47, + 473, + 287, + 497 + ], + "type": "text", + "content": " is the co-attentional transformer with parameter " + }, + { + "bbox": [ + 47, + 473, + 287, + 497 + ], + "type": "inline_equation", + "content": "\\phi_1" + }, + { + "bbox": [ + 47, + 473, + 287, + 497 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 473, + 287, + 497 + ], + "type": "inline_equation", + "content": "f_{\\phi_2}" + }, + { + "bbox": [ + 47, + 473, + 287, + 497 + ], + "type": "text", + "content": " is the transformer encoders with parameter " + }, + { + "bbox": [ + 47, + 473, + 287, + 497 + ], + "type": "inline_equation", + "content": "\\phi_2" + }, + { + "bbox": [ + 47, + 473, + 287, + 497 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 497, + 287, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 497, + 287, + 568 + ], + "spans": [ + { + "bbox": [ + 46, + 497, + 287, + 568 + ], + "type": "text", + "content": "The video-caption interaction module generates frame embeddings that can be further processed based on the type of matching needed. For global matching, the frame embeddings can be averaged to obtain a single video representation. Alternatively, for fine-grained matching, the individual frame embeddings can be retained." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 47, + 575, + 268, + 588 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 575, + 268, + 588 + ], + "spans": [ + { + "bbox": [ + 47, + 575, + 268, + 588 + ], + "type": "text", + "content": "2.5. Complementary Query-Caption Matching" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 46, + 594, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 287, + 713 + ], + "type": "text", + "content": "Besides using the caption for data augmentation and video feature enhancement, it can also directly represent the video content, allowing for text-text retrieval. Specifically, each of the " + }, + { + "bbox": [ + 46, + 594, + 287, + 713 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 46, + 594, + 287, + 713 + ], + "type": "text", + "content": " captions generated by the video is then passed through the caption encoder to obtain its [CLS] text embedding. These caption embeddings are then aggregated to form a global representation, as illustrated in Figure 2(c). The cosine similarity between this global caption embedding and the global query embedding is then calculated to complement the query-video matching." + } + ] + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 313, + 69, + 386, + 125 + ], + "blocks": [ + { + "bbox": [ + 313, + 69, + 386, + 125 + ], + "lines": [ + { + "bbox": [ + 313, + 69, + 386, + 125 + ], + "spans": [ + { + "bbox": [ + 313, + 69, + 386, + 125 + ], + "type": "image", + "image_path": "5b5f3522b5e8a75ce629e18e42860268fcf62a74838d1f22f4eb6b0a07ae1473.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 335, + 128, + 367, + 138 + ], + "lines": [ + { + "bbox": [ + 335, + 128, + 367, + 138 + ], + "spans": [ + { + "bbox": [ + 335, + 128, + 367, + 138 + ], + "type": "text", + "content": "(a) Sum" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 418, + 69, + 547, + 125 + ], + "blocks": [ + { + "bbox": [ + 418, + 69, + 547, + 125 + ], + "lines": [ + { + "bbox": [ + 418, + 69, + 547, + 125 + ], + "spans": [ + { + "bbox": [ + 418, + 69, + 547, + 125 + ], + "type": "image", + "image_path": "8ff8f77e0760782edf84a34a5fbe4b5bacb5620eee20a3123f8826c7987e564c.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 465, + 127, + 499, + 137 + ], + "lines": [ + { + "bbox": [ + 465, + 127, + 499, + 137 + ], + "spans": [ + { + "bbox": [ + 465, + 127, + 499, + 137 + ], + "type": "text", + "content": "(b) MLP" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 313, + 145, + 397, + 214 + ], + "blocks": [ + { + "bbox": [ + 313, + 145, + 397, + 214 + ], + "lines": [ + { + "bbox": [ + 313, + 145, + 397, + 214 + ], + "spans": [ + { + "bbox": [ + 313, + 145, + 397, + 214 + ], + "type": "image", + "image_path": "f83a8122a918f22cf8f8b069a044a7a992d34651603c0f088a8751cbb80203aa.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 216, + 397, + 225 + ], + "lines": [ + { + "bbox": [ + 313, + 216, + 397, + 225 + ], + "spans": [ + { + "bbox": [ + 313, + 216, + 397, + 225 + ], + "type": "text", + "content": "(c) Cross Transformer" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 305, + 258, + 545, + 292 + ], + "lines": [ + { + "bbox": [ + 305, + 258, + 545, + 292 + ], + "spans": [ + { + "bbox": [ + 305, + 258, + 545, + 292 + ], + "type": "text", + "content": "Figure 3. Illustration of four Video-Caption interaction strategies. The enhanced frame embeddings will be followed by a mean pooling for global matching or will remain for fine-grained matching." + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 419, + 143, + 547, + 213 + ], + "blocks": [ + { + "bbox": [ + 419, + 143, + 547, + 213 + ], + "lines": [ + { + "bbox": [ + 419, + 143, + 547, + 213 + ], + "spans": [ + { + "bbox": [ + 419, + 143, + 547, + 213 + ], + "type": "image", + "image_path": "407c3f1c9cbf64b35e7a8775624dd74fe29a5ff0d72c8ccac198a8deab4519a0.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 428, + 216, + 537, + 225 + ], + "lines": [ + { + "bbox": [ + 428, + 216, + 537, + 225 + ], + "spans": [ + { + "bbox": [ + 428, + 216, + 537, + 225 + ], + "type": "text", + "content": "(d) Co-attention Transformer" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 330, + 233, + 422, + 245 + ], + "lines": [ + { + "bbox": [ + 330, + 233, + 422, + 245 + ], + "spans": [ + { + "bbox": [ + 330, + 233, + 422, + 245 + ], + "type": "text", + "content": "Caption Embedding" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 438, + 233, + 525, + 244 + ], + "lines": [ + { + "bbox": [ + 438, + 233, + 525, + 244 + ], + "spans": [ + { + "bbox": [ + 438, + 233, + 525, + 244 + ], + "type": "text", + "content": "Frame Embedding" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + } + ], + "index": 21 + }, + { + "bbox": [ + 305, + 316, + 545, + 461 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 316, + 545, + 461 + ], + "spans": [ + { + "bbox": [ + 305, + 316, + 545, + 461 + ], + "type": "text", + "content": "Notation. Let " + }, + { + "bbox": [ + 305, + 316, + 545, + 461 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{e}_{\\mathbf{v}_i},\\mathbf{e}_{\\mathbf{t}_i},\\mathbf{e}_{\\mathbf{c}_i}\\}_{i = 1}^B" + }, + { + "bbox": [ + 305, + 316, + 545, + 461 + ], + "type": "text", + "content": " be a batch of " + }, + { + "bbox": [ + 305, + 316, + 545, + 461 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 305, + 316, + 545, + 461 + ], + "type": "text", + "content": " triples, where " + }, + { + "bbox": [ + 305, + 316, + 545, + 461 + ], + "type": "inline_equation", + "content": "\\mathbf{e}_{\\mathbf{v}_i}" + }, + { + "bbox": [ + 305, + 316, + 545, + 461 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 305, + 316, + 545, + 461 + ], + "type": "inline_equation", + "content": "\\mathbf{e}_{\\mathbf{t}_i}" + }, + { + "bbox": [ + 305, + 316, + 545, + 461 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 305, + 316, + 545, + 461 + ], + "type": "inline_equation", + "content": "\\mathbf{e}_{\\mathbf{c}_i}" + }, + { + "bbox": [ + 305, + 316, + 545, + 461 + ], + "type": "text", + "content": " denote the " + }, + { + "bbox": [ + 305, + 316, + 545, + 461 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 305, + 316, + 545, + 461 + ], + "type": "text", + "content": "-th video, query, and caption embedding, respectively. Note that the term \"embedding\" used here is more general for convenience and can vary in meaning depending on the situation. For instance, in query-video global matching, " + }, + { + "bbox": [ + 305, + 316, + 545, + 461 + ], + "type": "inline_equation", + "content": "\\mathbf{e}_{\\mathbf{v}_i}" + }, + { + "bbox": [ + 305, + 316, + 545, + 461 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 316, + 545, + 461 + ], + "type": "inline_equation", + "content": "\\mathbf{e}_{\\mathbf{t}_i}" + }, + { + "bbox": [ + 305, + 316, + 545, + 461 + ], + "type": "text", + "content": " represent the averaged video feature and global [CLS] text feature, respectively. In query-video fine-grained matching, " + }, + { + "bbox": [ + 305, + 316, + 545, + 461 + ], + "type": "inline_equation", + "content": "\\mathbf{e}_{\\mathbf{v}_i}" + }, + { + "bbox": [ + 305, + 316, + 545, + 461 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 316, + 545, + 461 + ], + "type": "inline_equation", + "content": "\\mathbf{e}_{\\mathbf{t}_i}" + }, + { + "bbox": [ + 305, + 316, + 545, + 461 + ], + "type": "text", + "content": " represent a sequence of frame embeddings and a sequence of word embeddings, respectively. In query-caption matching, " + }, + { + "bbox": [ + 305, + 316, + 545, + 461 + ], + "type": "inline_equation", + "content": "\\mathbf{e}_{\\mathbf{c}_i}" + }, + { + "bbox": [ + 305, + 316, + 545, + 461 + ], + "type": "text", + "content": " represents a sequence of caption embeddings, and " + }, + { + "bbox": [ + 305, + 316, + 545, + 461 + ], + "type": "inline_equation", + "content": "\\mathbf{e}_{\\mathbf{t}_i}" + }, + { + "bbox": [ + 305, + 316, + 545, + 461 + ], + "type": "text", + "content": " represents a global [CLS] text feature." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 305, + 463, + 545, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 463, + 545, + 559 + ], + "spans": [ + { + "bbox": [ + 305, + 463, + 545, + 559 + ], + "type": "text", + "content": "Learning Objectives. For the Query-Caption branch, we want the caption embedding " + }, + { + "bbox": [ + 305, + 463, + 545, + 559 + ], + "type": "inline_equation", + "content": "\\mathbf{e_c}" + }, + { + "bbox": [ + 305, + 463, + 545, + 559 + ], + "type": "text", + "content": " and the query embedding " + }, + { + "bbox": [ + 305, + 463, + 545, + 559 + ], + "type": "inline_equation", + "content": "\\mathbf{e_t}" + }, + { + "bbox": [ + 305, + 463, + 545, + 559 + ], + "type": "text", + "content": " to be close while they are related and far apart when they are not during training phase. We follow the common practice [24, 36] to consider the bidirectional learning objective. We employ symmetric cross-entropy loss to maximize the similarity between matched Query-Caption pairs and minimize the similarity for other pairs:" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 319, + 572, + 518, + 605 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 572, + 518, + 605 + ], + "spans": [ + { + "bbox": [ + 319, + 572, + 518, + 605 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {Q 2 C} = - \\frac {1}{B} \\sum_ {i} ^ {B} \\log \\frac {\\exp \\left(s _ {q c} \\left(\\mathbf {e} _ {\\mathbf {t} i} , \\mathbf {e} _ {\\mathbf {c} i}\\right) / \\tau\\right)}{\\sum_ {j} ^ {B} \\exp \\left(s _ {q c} \\left(\\mathbf {e} _ {\\mathbf {t} i} , \\mathbf {e} _ {\\mathbf {c} j}\\right) / \\tau\\right)},", + "image_path": "12ab5d4ea2e1239fb9780ec1f7dabd40780f4b06f063a6e30fd359c57d052d68.jpg" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 320, + 607, + 545, + 640 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 607, + 545, + 640 + ], + "spans": [ + { + "bbox": [ + 320, + 607, + 545, + 640 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {C 2 Q} = - \\frac {1}{B} \\sum_ {i} ^ {B} \\log \\frac {\\exp \\left(s _ {q c} \\left(\\mathbf {e} _ {\\mathbf {t} i} , \\mathbf {e} _ {\\mathbf {c} i}\\right) / \\tau\\right)}{\\sum_ {j} ^ {B} \\exp \\left(s _ {q c} \\left(\\mathbf {e} _ {\\mathbf {t} j} , \\mathbf {e} _ {\\mathbf {c} i}\\right) / \\tau\\right)}, \\tag {5}", + "image_path": "ac97f87cff729713404e83883e3cd3d1774032801985f2df66667868aeb49ab8.jpg" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 364, + 642, + 476, + 665 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 364, + 642, + 476, + 665 + ], + "spans": [ + { + "bbox": [ + 364, + 642, + 476, + 665 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {Q C} = \\frac {1}{2} \\left(\\mathcal {L} _ {Q 2 C} + \\mathcal {L} _ {C 2 Q}\\right),", + "image_path": "9dbc0af272acf14d13da25c08db96cd46e11c27cc7babfbc43499fa54db11624.jpg" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 319, + 677, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 677, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 319, + 677, + 545, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 319, + 677, + 545, + 714 + ], + "type": "inline_equation", + "content": "s_{qc}(\\cdot, \\cdot)" + }, + { + "bbox": [ + 319, + 677, + 545, + 714 + ], + "type": "text", + "content": " represents the query-caption matching similarity function shown in Figure 2(c), and " + }, + { + "bbox": [ + 319, + 677, + 545, + 714 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 319, + 677, + 545, + 714 + ], + "type": "text", + "content": " refers to the temperature hyper-parameter for scaling. Similarly, the con" + } + ] + } + ], + "index": 31 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10707" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 266, + 85 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 266, + 85 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 266, + 85 + ], + "type": "text", + "content": "trastive loss for Query-Video branch is formulated as:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 59, + 91, + 287, + 185 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 91, + 287, + 185 + ], + "spans": [ + { + "bbox": [ + 59, + 91, + 287, + 185 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} _ {Q 2 V} = - \\frac {1}{B} \\sum_ {i} ^ {B} \\log \\frac {\\exp (s _ {q v} (\\mathbf {e} _ {\\mathbf {t} i} , \\mathbf {e} _ {\\mathbf {v} i}) / \\tau)}{\\sum_ {j} ^ {B} \\exp (s _ {q v} (\\mathbf {e} _ {\\mathbf {t} i} , \\mathbf {e} _ {\\mathbf {v} j}) / \\tau)}, \\\\ \\mathcal {L} _ {V 2 Q} = - \\frac {1}{B} \\sum_ {i} ^ {B} \\log \\frac {\\exp \\left(s _ {q v} \\left(\\mathbf {e} _ {\\mathbf {t} i} , \\mathbf {e} _ {\\mathbf {v} i}\\right) / \\tau\\right)}{\\sum_ {j} ^ {B} \\exp \\left(s _ {q v} \\left(\\mathbf {e} _ {\\mathbf {t} j} , \\mathbf {e} _ {\\mathbf {v} i}\\right) / \\tau\\right)}, \\tag {6} \\\\ \\mathcal {L} _ {Q V} = \\frac {1}{2} (\\mathcal {L} _ {Q 2 V} + \\mathcal {L} _ {V 2 Q}), \\\\ \\end{array}", + "image_path": "5a43af8797a69170e753416d7a6ae9bca4d20ef175709d24288dc11800cc48be.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 191, + 287, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 191, + 287, + 240 + ], + "spans": [ + { + "bbox": [ + 47, + 191, + 287, + 240 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 191, + 287, + 240 + ], + "type": "inline_equation", + "content": "s_{qv}(\\cdot, \\cdot)" + }, + { + "bbox": [ + 47, + 191, + 287, + 240 + ], + "type": "text", + "content": " represents the query-video matching (e.g., global matching, fine-grained matching) similarity function shown in Figure 2(b). The total loss " + }, + { + "bbox": [ + 47, + 191, + 287, + 240 + ], + "type": "inline_equation", + "content": "\\mathcal{L}" + }, + { + "bbox": [ + 47, + 191, + 287, + 240 + ], + "type": "text", + "content": " is the sum of Query-Video loss " + }, + { + "bbox": [ + 47, + 191, + 287, + 240 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{QV}" + }, + { + "bbox": [ + 47, + 191, + 287, + 240 + ], + "type": "text", + "content": " and Query-Caption loss " + }, + { + "bbox": [ + 47, + 191, + 287, + 240 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{QC}" + }, + { + "bbox": [ + 47, + 191, + 287, + 240 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 127, + 247, + 287, + 262 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 127, + 247, + 287, + 262 + ], + "spans": [ + { + "bbox": [ + 127, + 247, + 287, + 262 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\mathcal {L} _ {Q V} + \\mathcal {L} _ {Q C}. \\tag {7}", + "image_path": "e31949e41bfb0c2e0a0847185afe3257584a6abde087d22265ffdbc402826d11.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 269, + 242, + 282 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 269, + 242, + 282 + ], + "spans": [ + { + "bbox": [ + 47, + 269, + 242, + 282 + ], + "type": "text", + "content": "3. Experiments: Text-Video Retrieval" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 288, + 101, + 301 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 288, + 101, + 301 + ], + "spans": [ + { + "bbox": [ + 47, + 288, + 101, + 301 + ], + "type": "text", + "content": "3.1. Setups" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 306, + 287, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 306, + 287, + 498 + ], + "spans": [ + { + "bbox": [ + 46, + 306, + 287, + 498 + ], + "type": "text", + "content": "Datasets. We conduct experiment on four popular benchmarks for video-to-text retrieval and text-to-video retrieval tasks. MSR-VTT [44] contains a total of 10K video clips, each having 20 captions. Following the data splits from [10,24,27], we train models with associated captions on the Training-9K set and report results on the test 1K-A set. DiDeMo [1] has 10K videos paired with 40K descriptions. Following previous works [2, 19, 24], we concatenate all descriptions of one video to a single query, acting as a video-paragraph retrieval task. VATEX [38] collects " + }, + { + "bbox": [ + 46, + 306, + 287, + 498 + ], + "type": "inline_equation", + "content": "\\sim 35\\mathrm{K}" + }, + { + "bbox": [ + 46, + 306, + 287, + 498 + ], + "type": "text", + "content": " videos, each with multiple annotations. There are " + }, + { + "bbox": [ + 46, + 306, + 287, + 498 + ], + "type": "inline_equation", + "content": "\\sim 26\\mathrm{K}" + }, + { + "bbox": [ + 46, + 306, + 287, + 498 + ], + "type": "text", + "content": " videos for training, 1,500 videos for validation and 1,500 videos for testing. MSVD [43] contains 1,970 videos with 80K captions, with " + }, + { + "bbox": [ + 46, + 306, + 287, + 498 + ], + "type": "inline_equation", + "content": "\\sim 40" + }, + { + "bbox": [ + 46, + 306, + 287, + 498 + ], + "type": "text", + "content": " captions on average per video. There are 1,200, 100, and 670 videos in the train, validation, and test sets, respectively." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 498, + 287, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 498, + 287, + 605 + ], + "spans": [ + { + "bbox": [ + 46, + 498, + 287, + 605 + ], + "type": "text", + "content": "Evaluation Metrics. For brevity, we abbreviate Recall at " + }, + { + "bbox": [ + 46, + 498, + 287, + 605 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 46, + 498, + 287, + 605 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 46, + 498, + 287, + 605 + ], + "type": "inline_equation", + "content": "\\mathbb{R}@\\mathbb{K}" + }, + { + "bbox": [ + 46, + 498, + 287, + 605 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 46, + 498, + 287, + 605 + ], + "type": "inline_equation", + "content": "K = 1,5,10" + }, + { + "bbox": [ + 46, + 498, + 287, + 605 + ], + "type": "text", + "content": ") upon all datasets, which computes the percentage of correct videos among the top " + }, + { + "bbox": [ + 46, + 498, + 287, + 605 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 46, + 498, + 287, + 605 + ], + "type": "text", + "content": " retrieved videos given textual queries (Text→Video, and vice versa). MdR, Median Rank, computes the median of the ground-truth in the retrieval ranking list. MnR, Mean Rank, computes the mean rank of the correct results in the retrieval ranking list. Note that for MdR and MnR, the lower score means the better (indicated as " + }, + { + "bbox": [ + 46, + 498, + 287, + 605 + ], + "type": "inline_equation", + "content": "\\downarrow" + }, + { + "bbox": [ + 46, + 498, + 287, + 605 + ], + "type": "text", + "content": ")." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 605, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 288, + 714 + ], + "type": "text", + "content": "Implementation Details. All experiments use the visual encoder in CLIP [30] as the video encoder, and the textual encoder in CLIP as both the caption encoder and query encoder. The caption encoder and query encoder share parameters. To reduce conflict between the two branches, the query-video branch is trained first, followed by the query-caption branch. The text length is fixed to 32, and the video length is fixed to 12 for all datasets except DiDeMo (64 max words and 64 frames). The initial learning rate is set to 1e-7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 72, + 545, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 205 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 205 + ], + "type": "text", + "content": "for the clip parameters and 1e-4 for the non-clip parameters. The model is trained with a batch size of 128 for 5 epochs, except for DiDeMo (15 epochs), using the Adam [18] optimizer. All learning rates follow the cosine schedule with a linear warmup [14] strategy. For the number of generated captions per video, we set " + }, + { + "bbox": [ + 304, + 72, + 545, + 205 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 304, + 72, + 545, + 205 + ], + "type": "text", + "content": " to 30. The interaction module employs " + }, + { + "bbox": [ + 304, + 72, + 545, + 205 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 304, + 72, + 545, + 205 + ], + "type": "text", + "content": " transformer layers, where " + }, + { + "bbox": [ + 304, + 72, + 545, + 205 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 304, + 72, + 545, + 205 + ], + "type": "text", + "content": " is set to 4 for VA-TEX and MSR-VTT, and 1 for Didemo and MSVD. In the caption branch, the number of transformer layers is set to 2. For caption generation, we directly use the original pretrained CLIP and GPT-2, without any additional tuning." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 212, + 493, + 225 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 212, + 493, + 225 + ], + "spans": [ + { + "bbox": [ + 305, + 212, + 493, + 225 + ], + "type": "text", + "content": "3.2. Comparison with State-of-the-Arts" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 231, + 545, + 267 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 231, + 545, + 267 + ], + "spans": [ + { + "bbox": [ + 304, + 231, + 545, + 267 + ], + "type": "text", + "content": "In this section, we compare our Cap4Video with recent state-of-the-art methods on four benchmarks: MSR-VTT [44], MSVD [43], VATEX [38], and DiDeMo [1]." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 267, + 545, + 315 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 267, + 545, + 315 + ], + "spans": [ + { + "bbox": [ + 304, + 267, + 545, + 315 + ], + "type": "text", + "content": "Table 1 shows the comparisons on DiDeMo, where our Cap4Video outperforms CLIP4Clip [24] by a significant margin of " + }, + { + "bbox": [ + 304, + 267, + 545, + 315 + ], + "type": "inline_equation", + "content": "9.2\\%" + }, + { + "bbox": [ + 304, + 267, + 545, + 315 + ], + "type": "text", + "content": " in R@1 and exceeds DRL [36] by " + }, + { + "bbox": [ + 304, + 267, + 545, + 315 + ], + "type": "inline_equation", + "content": "3.0\\%" + }, + { + "bbox": [ + 304, + 267, + 545, + 315 + ], + "type": "text", + "content": ", demonstrating the effectiveness of our method." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 316, + 545, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 316, + 545, + 422 + ], + "spans": [ + { + "bbox": [ + 304, + 316, + 545, + 422 + ], + "type": "text", + "content": "Table 2 provides a comparison of our approach with recent state-of-the-art models on MSR-VTT. Our method achieves new state-of-the-art performance on text-to-video retrieval for both ViT-B/32 and ViT-B/16 backbones, significantly surpassing previous works. For instance, we achieve a " + }, + { + "bbox": [ + 304, + 316, + 545, + 422 + ], + "type": "inline_equation", + "content": "+4.8\\%" + }, + { + "bbox": [ + 304, + 316, + 545, + 422 + ], + "type": "text", + "content": " higher R@1 than CLIP4Clip with the same ViT-B/32 on text-to-video retrieval. Additionally, our Cap4Video outperforms the recent TS2-Net [22] by " + }, + { + "bbox": [ + 304, + 316, + 545, + 422 + ], + "type": "inline_equation", + "content": "2.3\\%" + }, + { + "bbox": [ + 304, + 316, + 545, + 422 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 316, + 545, + 422 + ], + "type": "inline_equation", + "content": "2.0\\%" + }, + { + "bbox": [ + 304, + 316, + 545, + 422 + ], + "type": "text", + "content": " with ViT-B/32 and ViT-B/16, respectively." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 423, + 545, + 530 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 423, + 545, + 530 + ], + "spans": [ + { + "bbox": [ + 304, + 423, + 545, + 530 + ], + "type": "text", + "content": "Table 3 and Table 4 show the results for the MSVD and VATEX datasets, respectively, where we use ViT-B/16 as our backbone. For MSVD, our Cap4Video achieves a remarkable performance of " + }, + { + "bbox": [ + 304, + 423, + 545, + 530 + ], + "type": "inline_equation", + "content": "51.8\\%" + }, + { + "bbox": [ + 304, + 423, + 545, + 530 + ], + "type": "text", + "content": " R@1 and outperforms CLIP-based models CLIP4Clip [24] and X-Pool [25] by " + }, + { + "bbox": [ + 304, + 423, + 545, + 530 + ], + "type": "inline_equation", + "content": "6.6\\%" + }, + { + "bbox": [ + 304, + 423, + 545, + 530 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 423, + 545, + 530 + ], + "type": "inline_equation", + "content": "4.6\\%" + }, + { + "bbox": [ + 304, + 423, + 545, + 530 + ], + "type": "text", + "content": " on text-to-video retrieval, respectively. For VATEX, our approach also outperforms the recent state-of-the-art methods and achieves a " + }, + { + "bbox": [ + 304, + 423, + 545, + 530 + ], + "type": "inline_equation", + "content": "+7.5\\%" + }, + { + "bbox": [ + 304, + 423, + 545, + 530 + ], + "type": "text", + "content": " R@1 improvement over TS2-Net [22] for text-to-video retrieval." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 317, + 531, + 545, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 531, + 545, + 544 + ], + "spans": [ + { + "bbox": [ + 317, + 531, + 545, + 544 + ], + "type": "text", + "content": "In recent studies, several methods have been proposed" + } + ] + } + ], + "index": 15 + }, + { + "type": "table", + "bbox": [ + 318, + 562, + 535, + 693 + ], + "blocks": [ + { + "bbox": [ + 318, + 562, + 535, + 693 + ], + "lines": [ + { + "bbox": [ + 318, + 562, + 535, + 693 + ], + "spans": [ + { + "bbox": [ + 318, + 562, + 535, + 693 + ], + "type": "table", + "html": "
MethodR@1R@5R@10MdRMnR
CE [21]15.640.9-8.2-
ClipBERT [19]21.147.361.16.3-
Frozen [2]31.059.872.43.0-
TMVM [20]36.564.975.43.0-
CLIP4Clip [24]42.868.579.22.018.9
TS2-Net [22]41.871.682.02.014.8
HunYuan [28]45.075.683.42.012.0
DRL [36]49.076.584.52.0-
Cap4Video52.079.487.5110.5
", + "image_path": "4bd930cfc7000be07d882d78e1de9a298570d6896ae8c377a7445757bff13a1a.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "table_body" + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 699, + 537, + 710 + ], + "lines": [ + { + "bbox": [ + 313, + 699, + 537, + 710 + ], + "spans": [ + { + "bbox": [ + 313, + 699, + 537, + 710 + ], + "type": "text", + "content": "Table 1. Results of text-to-video retrieval on the DiDeMo [1]." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10708" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 51, + 70, + 542, + 376 + ], + "blocks": [ + { + "bbox": [ + 51, + 70, + 542, + 376 + ], + "lines": [ + { + "bbox": [ + 51, + 70, + 542, + 376 + ], + "spans": [ + { + "bbox": [ + 51, + 70, + 542, + 376 + ], + "type": "table", + "html": "
MethodVenueText → VideoVideo → Text
R@1R@5R@10MdR↓MnR↓R@1R@5R@10MdR↓MnR↓
ClipBERT [19]CVPR'2022.046.859.96.0-----
MMT [10]ECCV'2026.657.169.64.0-27.057.569.73.721.3
T2VLAD [39]CVPR'2129.559.070.14.0-31.860.071.13.0
SupportSet [29]ICLR'2130.158.569.33.0-28.558.671.63.0-
Frozen [2]ICCV'2132.561.571.23.0------
BridgeFormer [12]CVPR'2237.664.875.1-------
TMVM [20]NeurIPS'2236.264.275.73.0-34.863.873.73.0-
CLIP-ViT-B/32
CLIP4Clip [24]arXiv'2144.571.481.62.015.342.770.980.62.011.6
CenterCLIP [48]SIGIR'2244.271.682.12.015.142.871.782.22.010.9
CAMoE [7]arXiv'2144.672.681.82.013.345.172.483.12.010.0
CLIP2Video [9]arXiv'2145.672.681.72.014.643.572.382.12.010.2
X-Pool [13]CVPR'2246.972.882.22.014.3-----
QB-Norm [4]CVPR'2247.273.083.02.0------
TS2-Net [22]ECCV'2247.074.583.82.013.045.374.183.72.09.2
DRL [36]arXiv'2247.474.683.82.0-45.373.983.32.0-
Cap4Video49.374.383.82.012.047.173.784.32.08.7
CLIP-ViT-B/16
CLIP2TV [11]arXiv'2148.374.682.82.014.946.575.484.92.010.2
CenterCLIP [48]SIGIR'2248.473.882.02.013.847.775.083.32.010.2
TS2-Net [22]ECCV'2249.475.685.32.013.546.675.984.92.08.9
DRL [36]arXiv'2250.276.584.71.0-48.976.385.42.0-
Cap4Video51.475.783.91.012.449.075.285.02.08
", + "image_path": "da0da4daa866be38ac7bc836166b2349a80dbaef6de2b9f1fdea7b85ad6942fc.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 59, + 422, + 276, + 544 + ], + "blocks": [ + { + "bbox": [ + 47, + 381, + 545, + 404 + ], + "lines": [ + { + "bbox": [ + 47, + 381, + 545, + 404 + ], + "spans": [ + { + "bbox": [ + 47, + 381, + 545, + 404 + ], + "type": "text", + "content": "Table 2. Retrieval results on the validation set of MSR-VTT 1K [44]. Here we report results without any post-processing operations (e.g., DSL [7] or QB-Norm [4]) during inference." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 59, + 422, + 276, + 544 + ], + "lines": [ + { + "bbox": [ + 59, + 422, + 276, + 544 + ], + "spans": [ + { + "bbox": [ + 59, + 422, + 276, + 544 + ], + "type": "table", + "html": "
MethodR@1R@5R@10MdRMnR
CE [21]19.849.063.86.0-
SUPPORT [29]28.460.072.94.0-
CLIP [30]37.064.173.83.0-
Frozen [2]33.764.776.33.0-
TMVM [20]36.767.481.32.5-
CLIP4Clip [24]45.275.584.32.010.3
X-Pool [13]47.277.486.02.09.3
Cap4Video51.880.888.318.3
", + "image_path": "691f1af6f839f6ab3d929e79bd1096144ce883588db3752e1a5dfcbe071bdab8.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 318, + 422, + 535, + 544 + ], + "blocks": [ + { + "bbox": [ + 55, + 548, + 278, + 559 + ], + "lines": [ + { + "bbox": [ + 55, + 548, + 278, + 559 + ], + "spans": [ + { + "bbox": [ + 55, + 548, + 278, + 559 + ], + "type": "text", + "content": "Table 3. Results of text-to-video retrieval on the MSVD [43]." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 318, + 422, + 535, + 544 + ], + "lines": [ + { + "bbox": [ + 318, + 422, + 535, + 544 + ], + "spans": [ + { + "bbox": [ + 318, + 422, + 535, + 544 + ], + "type": "table", + "html": "
MethodR@1R@5R@10MdRMnR
HGR [6]35.173.583.52.0-
CLIP [30]39.772.382.22.012.8
SUPPORT [29]44.982.189.71.0-
CLIP4Clip [24]55.989.295.01.03.9
Clip2Video [9]57.390.095.51.03.6
QB-Norm [4]58.888.393.81.0-
TS2-Net [22]59.190.095.21.03.5
Cap4Video66.693.197.012.7
", + "image_path": "b765fce0e6561e4b022bdd2e5d008c688fdd3c839dd6e5eacab29c777fd31df3.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 548, + 538, + 559 + ], + "lines": [ + { + "bbox": [ + 313, + 548, + 538, + 559 + ], + "spans": [ + { + "bbox": [ + 313, + 548, + 538, + 559 + ], + "type": "text", + "content": "Table 4. Results of text-to-video retrieval on the VATEX [38]." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 594, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 289, + 715 + ], + "type": "text", + "content": "to improve text-video retrieval performance by adjusting similarity during inference using other query information. Notably, our results adheres to the standard retrieval logic, where the most relevant video is retrieved for each query from a set of videos, without any knowledge of the relationship between other queries and videos. Therefore, all results reported in our tables do not involve any post-processing procedures such as DSL [7] and QB-Norm [4]. Overall, the consistent state-of-the-art performance across four benchmarks demonstrates the effectiveness of our Cap4Video." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 584, + 400, + 597 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 584, + 400, + 597 + ], + "spans": [ + { + "bbox": [ + 306, + 584, + 400, + 597 + ], + "type": "text", + "content": "3.3. Ablation Study" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 604, + 545, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 604, + 545, + 628 + ], + "spans": [ + { + "bbox": [ + 304, + 604, + 545, + 628 + ], + "type": "text", + "content": "In this section, we provide detailed ablation studies to clarify the effects of each part of our design." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 630, + 545, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 630, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 630, + 545, + 715 + ], + "type": "text", + "content": "Auxiliary Caption as Data Augmentation. We begin by investigating the impact of captions on data augmentation for training. In a real-world scenario, the original video title would naturally serve as an additional auxiliary caption. Therefore, we manually extracted the title from the video's original webpage and compared it to the caption generated by the GPT-2 model. Table 5 presents the re" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "10709" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 59, + 70, + 533, + 311 + ], + "blocks": [ + { + "bbox": [ + 59, + 70, + 533, + 311 + ], + "lines": [ + { + "bbox": [ + 59, + 70, + 533, + 311 + ], + "spans": [ + { + "bbox": [ + 59, + 70, + 533, + 311 + ], + "type": "table", + "html": "
MethodGlobal MatchingFine-grained Matching
R@1R@5R@10MdR↓MnR↓R@1R@5R@10MdR↓MnR↓
Baseline42.870.479.0216.645.773.782.6213.1
+Different Sources of Caption as Data Augmentation
Video Title from Source URL43.871.180.9215.144.372.783.5213.1
Zero-shot Video Captioning44.270.781.5216.246.372.581.7212.9
+Different Number of Captions for Data Augmentation
Top-144.270.781.5216.246.372.581.7212.9
Top-343.371.781.6215.045.573.882.4212.7
Top-543.470.680.4216.245.672.782.7212.9
+Video-Caption Feature Interaction
Video Only44.270.781.5216.246.372.581.7212.9
Sum43.871.580.3216.147.273.382.8213.1
Concat-MLP37.566.178.4315.740.068.779.9212.7
Cross Transformer44.671.680.3214.647.975.483.0211.5
Co-attention Transformer45.371.280.9215.048.574.082.5212.7
+Query-Caption Matching Score
Query-Video Only45.371.280.9215.048.574.082.5212.7
Query-Caption Only30.355.267.5426.430.355.267.5426.4
Query-Video + Query-Caption45.671.781.2214.849.374.283.4212.1
", + "image_path": "4b670fcfa9ce1435c2733532007b621c3174a1d92141fea3e67fd6a7aecb5541.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 371, + 287, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 371, + 287, + 443 + ], + "spans": [ + { + "bbox": [ + 46, + 371, + 287, + 443 + ], + "type": "text", + "content": "sults of using different sources of the caption, from which we observe that using captions generated by the web-scale model as data augmentation for training can lead to direct improvements in R@1 " + }, + { + "bbox": [ + 46, + 371, + 287, + 443 + ], + "type": "inline_equation", + "content": "(+1.4\\%, +0.6\\%)" + }, + { + "bbox": [ + 46, + 371, + 287, + 443 + ], + "type": "text", + "content": " under both matching mechanisms. Using video titles can also bring a " + }, + { + "bbox": [ + 46, + 371, + 287, + 443 + ], + "type": "inline_equation", + "content": "1\\%" + }, + { + "bbox": [ + 46, + 371, + 287, + 443 + ], + "type": "text", + "content": " improvement for global matching." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 445, + 287, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 445, + 287, + 517 + ], + "spans": [ + { + "bbox": [ + 46, + 445, + 287, + 517 + ], + "type": "text", + "content": "We also explore the impact of the number of generated captions used for augmentation. We used the caption filtering mechanism mentioned in Sec. 2.3 to rank the relevance of captions and the ground-truth query, and selected different numbers of captions for training. The results demonstrated that using only one caption is sufficient." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 519, + 287, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 519, + 287, + 628 + ], + "spans": [ + { + "bbox": [ + 46, + 519, + 287, + 628 + ], + "type": "text", + "content": "Benefit on Both Online and Offline Videos Scenarios. Our method is applicable for both online and offline videos. Offline videos are local videos with no title, while online videos have a title on the video website. Therefore, for online videos, the step of generating a caption with " + }, + { + "bbox": [ + 46, + 519, + 287, + 628 + ], + "type": "inline_equation", + "content": "\\mathrm{CLIP + GPT - 2}" + }, + { + "bbox": [ + 46, + 519, + 287, + 628 + ], + "type": "text", + "content": " is skipped, and the website title is used directly as a caption. In Table 6, the results show that our method has significantly improved over the global matching baseline for both offline and online videos." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 629, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 629, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 629, + 287, + 715 + ], + "type": "text", + "content": "Video-Caption Feature Interaction. As mentioned in Sec. 2.4, we have designed four approaches for Video-Caption feature interaction. Based on the results presented in Table 5, we can conclude the following: 1) The basic approach of Sum has been shown to enhance fine-grained matching by " + }, + { + "bbox": [ + 46, + 629, + 287, + 715 + ], + "type": "inline_equation", + "content": "0.9\\%" + }, + { + "bbox": [ + 46, + 629, + 287, + 715 + ], + "type": "text", + "content": " R@1, but there is no noticeable improvement in global matching. 2) The MLP approach is difficult" + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 328, + 369, + 525, + 432 + ], + "blocks": [ + { + "bbox": [ + 46, + 316, + 547, + 351 + ], + "lines": [ + { + "bbox": [ + 46, + 316, + 547, + 351 + ], + "spans": [ + { + "bbox": [ + 46, + 316, + 547, + 351 + ], + "type": "text", + "content": "Table 5. Component-wise evaluation of our framework on the MSR-VTT 1K validation set. With the ViT-B/32 backbone, we report the text-to-video retrieval results for two representative Query-Video matching mechanisms: global matching and fine-grained matching. The consistent improvement on two typical matching mechanisms demonstrates the generalization ability and effectiveness of our method." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 328, + 369, + 525, + 432 + ], + "lines": [ + { + "bbox": [ + 328, + 369, + 525, + 432 + ], + "spans": [ + { + "bbox": [ + 328, + 369, + 525, + 432 + ], + "type": "table", + "html": "
Caption SourceR@1
BaselineN/A42.8
Cap4VideoOriginal Website Title45.8
Captioner (CLIP+GPT-2)45.6
", + "image_path": "1646cdf3574116cf8200a3a005a5bfbfaa8e9e5790f23bec7799911cfaa6806e.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 437, + 545, + 461 + ], + "lines": [ + { + "bbox": [ + 305, + 437, + 545, + 461 + ], + "spans": [ + { + "bbox": [ + 305, + 437, + 545, + 461 + ], + "type": "text", + "content": "Table 6. Exploring the effectiveness of captions from different sources on MSR-VTT 1k-A. Setting: ViT-B/32, global matching." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 484, + 545, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 484, + 545, + 652 + ], + "spans": [ + { + "bbox": [ + 304, + 484, + 545, + 652 + ], + "type": "text", + "content": "to optimize and performs poorly in both matching scenarios. We speculate that the MLP's operation in a black-box environment, despite creating a nonlinear metric space, may lead to degradation. 3) The Cross Transformer approach has demonstrated improvements of " + }, + { + "bbox": [ + 304, + 484, + 545, + 652 + ], + "type": "inline_equation", + "content": "+0.4\\%" + }, + { + "bbox": [ + 304, + 484, + 545, + 652 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 484, + 545, + 652 + ], + "type": "inline_equation", + "content": "+1.6\\%" + }, + { + "bbox": [ + 304, + 484, + 545, + 652 + ], + "type": "text", + "content": " in two matching settings, respectively. These enhancements may be attributed to the self-attention mechanism's ability to capture the inter-modal relationship between the video and caption. 4) Moreover, the Co-attention Transformer approach has significantly boosted performance, with gains of " + }, + { + "bbox": [ + 304, + 484, + 545, + 652 + ], + "type": "inline_equation", + "content": "+1.1\\%" + }, + { + "bbox": [ + 304, + 484, + 545, + 652 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 484, + 545, + 652 + ], + "type": "inline_equation", + "content": "+2.2\\%" + }, + { + "bbox": [ + 304, + 484, + 545, + 652 + ], + "type": "text", + "content": " for these two matching mechanisms. In summary, the results indicate that proper interaction between the video and generated caption can lead to better video representation and improved Query-Video matching." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "type": "text", + "content": "Query-Caption matching. We also investigate the Query-Caption matching branch for text-video retrieval. We aggregate caption embeddings using mean pooling to yield a global embedding. As shown in Table 5, the single Query-Caption matching branch achieves " + }, + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "type": "inline_equation", + "content": "30.3\\%" + }, + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "type": "text", + "content": " R@1 on text-to" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "10710" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 49, + 81, + 286, + 182 + ], + "blocks": [ + { + "bbox": [ + 104, + 71, + 236, + 80 + ], + "lines": [ + { + "bbox": [ + 104, + 71, + 236, + 80 + ], + "spans": [ + { + "bbox": [ + 104, + 71, + 236, + 80 + ], + "type": "text", + "content": "Query7765: a person is discussing a car." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 49, + 81, + 286, + 182 + ], + "lines": [ + { + "bbox": [ + 49, + 81, + 286, + 182 + ], + "spans": [ + { + "bbox": [ + 49, + 81, + 286, + 182 + ], + "type": "table", + "html": "
VideoRank+ CaptionRank
6video of a car camera recording the driver's voice.1
4video showing the car in a parking spot.2
5video of SUV in the video below shows a salesman talking to an audience.3
", + "image_path": "bd0ea46b7b87631888f0581a4b24ae9bd771a918ab530b6e7aa7960641ff8350.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 49, + 194, + 286, + 300 + ], + "blocks": [ + { + "bbox": [ + 57, + 183, + 279, + 193 + ], + "lines": [ + { + "bbox": [ + 57, + 183, + 279, + 193 + ], + "spans": [ + { + "bbox": [ + 57, + 183, + 279, + 193 + ], + "type": "text", + "content": "Query9616: person is recording the brown horse which is having fun." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 49, + 194, + 286, + 300 + ], + "lines": [ + { + "bbox": [ + 49, + 194, + 286, + 300 + ], + "spans": [ + { + "bbox": [ + 49, + 194, + 286, + 300 + ], + "type": "table", + "html": "
VideoRank 2+ Caption video of the horse jumping over a fence at Ranch in Nevada was captured on camera.Rank 1
1video showing animation of a horse's simulation, which simulates the game.2
4video showing a horse simulation video game in which you could see your avatar being animated by the camera.3
", + "image_path": "05d563cb803c0f28752c22ac4142287e6975054ea8ebe135b6ddb2fa8390fa14.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 310, + 286, + 321 + ], + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 310, + 286, + 321 + ], + "spans": [ + { + "bbox": [ + 46, + 310, + 286, + 321 + ], + "type": "text", + "content": "Figure 4. The text-video results on the MSR-VTT 1K-A test set." + } + ] + } + ], + "index": 4, + "type": "text" + }, + { + "bbox": [ + 47, + 322, + 286, + 332 + ], + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 322, + 286, + 332 + ], + "spans": [ + { + "bbox": [ + 47, + 322, + 286, + 332 + ], + "type": "text", + "content": "Left: The ranking results of the query-video matching model." + } + ] + } + ], + "index": 5, + "type": "text" + }, + { + "bbox": [ + 47, + 333, + 286, + 354 + ], + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 333, + 286, + 354 + ], + "spans": [ + { + "bbox": [ + 47, + 333, + 286, + 354 + ], + "type": "text", + "content": "Right: The ranking results of Cap4Video, which incorporates generated captions to enhance retrieval. Please zoom in for best view." + } + ] + } + ], + "index": 6, + "type": "text" + }, + { + "bbox": [ + 46, + 376, + 286, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 376, + 286, + 460 + ], + "spans": [ + { + "bbox": [ + 46, + 376, + 286, + 460 + ], + "type": "text", + "content": "video retrieval, outperforming previous query-video matching methods such as ClipBERT [19] (22.0%) and MMT [10] (26.6%). This suggests that the Query-Caption matching branch can complement the regular Query-Video matching branch for improved text-video retrieval. Combining the score of Query-Caption matching branch with Query-Video matching branch further improves performance (+0.8%)." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 461, + 286, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 461, + 286, + 509 + ], + "spans": [ + { + "bbox": [ + 46, + 461, + 286, + 509 + ], + "type": "text", + "content": "Overall, Cap4Video utilizes generated captions in three ways: input data augmentation, intermediate feature interaction, and output score fusion, leading to consistent improvements " + }, + { + "bbox": [ + 46, + 461, + 286, + 509 + ], + "type": "inline_equation", + "content": "(+2.8\\% / +3.6\\%)" + }, + { + "bbox": [ + 46, + 461, + 286, + 509 + ], + "type": "text", + "content": " in both matching mechanisms." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 518, + 130, + 529 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 518, + 130, + 529 + ], + "spans": [ + { + "bbox": [ + 47, + 518, + 130, + 529 + ], + "type": "text", + "content": "3.4. Visualization" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 536, + 286, + 620 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 536, + 286, + 620 + ], + "spans": [ + { + "bbox": [ + 46, + 536, + 286, + 620 + ], + "type": "text", + "content": "We provide two examples of videos retrieved by our Cap4Video method and a model without auxiliary captions. As illustrated in Figure 4, our approach successfully retrieves the ground-truth video with the assistance of the caption, while the video-only model returns multiple videos that are somewhat relevant to the query but not precise. See more qualitative results in Supplementary." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 633, + 138, + 645 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 633, + 138, + 645 + ], + "spans": [ + { + "bbox": [ + 47, + 633, + 138, + 645 + ], + "type": "text", + "content": "4. Related Works" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 653, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 286, + 713 + ], + "type": "text", + "content": "Zero-shot Image Captioning. In the field of natural language processing, transformer-based GPT models [5, 31] have been successful in generating text from prompts by training on large-scale text corpora. Similarly, CLIP [30], a vision-language alignment model trained on 400 million" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 72, + 545, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 239 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 239 + ], + "type": "text", + "content": "image-text pairs, has demonstrated impressive zero-shot performance on vision tasks. However, research on transferring web-scale models to zero-shot video captioning remains limited. Recently, ZeroCap [34] proposed a method of using CLIP and the GPT-2 language model to generate textual descriptions of input images, leveraging knowledge from both models in a truly zero-shot manner without retraining or fine-tuning model parameters. MAGIC [32] has also used CLIP scores to align GPT-2 logits with corresponding images but requires fine-tuning on the MS-COCO caption text corpus. More recently, a study [33] extended the zero-shot capability of ZeroCap to the video domain. In this paper, we employ this video extension to generate auxiliary captions without any additional training." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 240, + 545, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 240, + 545, + 574 + ], + "spans": [ + { + "bbox": [ + 304, + 240, + 545, + 574 + ], + "type": "text", + "content": "Text-Video Retrieval aims to retrieve relevant video content based on natural language descriptions. Early studies [6, 10, 21, 37, 39] focused on knowledge transfer from \"expert\" models and captured intra-modal and cross-modal interactions based on pre-extracted features. However, the performance of these methods is limited since they cannot perform end-to-end optimization. Recently, more methods have involved end-to-end training for text-video retrieval. One typical approach [2, 26, 27] is to first perform largescale text-video pre-training, then transfer the model to downstream text-video retrieval tasks. With the emergence of pre-trained Vision-Language Models (VLMs), there have been increased efforts to leverage them for improving video understanding [41, 42]. Thus, another training-efficient line is to directly expand the pre-trained VLM to the text-video retrieval task. CLIPBERT [19] enables affordable pioneering end-to-end training with a sparse sampling strategy. After that, recent works [3, 4, 8, 9, 11, 13, 22, 24, 40, 48] focus on transferring knowledge from CLIP models that have been pre-trained on 400M image-text pairs. The research path has evolved from the most direct global matching (i.e., video-sentence alignment [11, 24]) to fine-grained matching (e.g., frame-word alignment [36], video-word alignment [13], multi-hierarchical alignment [9, 28]). Unlike above CLIP-Based efforts on query-video matching, we propose to generate auxiliary captions from videos to improve text-video retrieval. Thus our method is compatible with both global and fine-grained matching." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 586, + 378, + 597 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 586, + 378, + 597 + ], + "spans": [ + { + "bbox": [ + 306, + 586, + 378, + 597 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": "We introduce Cap4Video, a novel framework that leverages captions generated by web-scale language models to enhance text-video matching in three key ways: 1) Input data augmentation for training, 2) Intermediate videocaption feature interaction for compact video representations, and 3) Output score fusion for improved text-video retrieval. Our approach demonstrates consistent performance gains on four standard text-video retrieval benchmarks, outperforming state-of-the-art methods by a clear margin." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10711" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 134 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 134 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 134 + ], + "type": "text", + "content": "[1] Lisa Anne Hendricks, Oliver Wang, Eli Shechtman, Josef Sivic, Trevor Darrell, and Bryan Russell. Localizing moments in video with natural language. In ICCV, pages 5803-5812, 2017. 2, 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 137, + 288, + 179 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 137, + 288, + 179 + ], + "spans": [ + { + "bbox": [ + 53, + 137, + 288, + 179 + ], + "type": "text", + "content": "[2] Max Bain, Arsha Nagrani, Gül Varol, and Andrew Zisserman. Frozen in time: A joint video and image encoder for end-to-end retrieval. In ICCV, pages 1728-1738, 2021. 5, 6, 8" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 182, + 288, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 182, + 288, + 215 + ], + "spans": [ + { + "bbox": [ + 53, + 182, + 288, + 215 + ], + "type": "text", + "content": "[3] Max Bain, Arsha Nagrani, Gül Varol, and Andrew Zisserman. A clip-hitchhiker's guide to long video retrieval. arXiv preprint arXiv:2205.08508, 2022. 8" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 217, + 288, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 217, + 288, + 249 + ], + "spans": [ + { + "bbox": [ + 53, + 217, + 288, + 249 + ], + "type": "text", + "content": "[4] Simion-Vlad Bogolin, Ioana Croitoru, Hailin Jin, Yang Liu, and Samuel Albanie. Cross modal retrieval with querybank normalisation. In CVPR, pages 5194-5205, 2022. 6, 8" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 251, + 288, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 251, + 288, + 304 + ], + "spans": [ + { + "bbox": [ + 53, + 251, + 288, + 304 + ], + "type": "text", + "content": "[5] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. NeurIPS, 33:1877-1901, 2020. 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 306, + 288, + 339 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 306, + 288, + 339 + ], + "spans": [ + { + "bbox": [ + 53, + 306, + 288, + 339 + ], + "type": "text", + "content": "[6] Shizhe Chen, Yida Zhao, Qin Jin, and Qi Wu. Fine-grained video-text retrieval with hierarchical graph reasoning. In CVPR, pages 10638-10647, 2020. 6, 8" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 341, + 288, + 384 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 341, + 288, + 384 + ], + "spans": [ + { + "bbox": [ + 53, + 341, + 288, + 384 + ], + "type": "text", + "content": "[7] Xing Cheng, Hezheng Lin, Xiangyu Wu, Fan Yang, and Dong Shen. Improving video-text retrieval by multi-stream corpus alignment and dual softmax loss. arXiv preprint arXiv:2109.04290, 2021. 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 386, + 288, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 386, + 288, + 430 + ], + "spans": [ + { + "bbox": [ + 53, + 386, + 288, + 430 + ], + "type": "text", + "content": "[8] Bo Fang, Chang Liu, Yu Zhou, Min Yang, Yuxin Song, Fu Li, Weiping Wang, Xiangyang Ji, Wanli Ouyang, et al. Uatvr: Uncertainty-adaptive text-video retrieval. arXiv preprint arXiv:2301.06309, 2023. 8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 431, + 288, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 431, + 288, + 464 + ], + "spans": [ + { + "bbox": [ + 53, + 431, + 288, + 464 + ], + "type": "text", + "content": "[9] Han Fang, Pengfei Xiong, Luhui Xu, and Yu Chen. Clip2video: Mastering video-text retrieval via image clip. arXiv preprint arXiv:2106.11097, 2021. 1, 2, 6, 8" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 465, + 288, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 465, + 288, + 498 + ], + "spans": [ + { + "bbox": [ + 48, + 465, + 288, + 498 + ], + "type": "text", + "content": "[10] Valentin Gabeur, Chen Sun, Karteek Alahari, and Cordelia Schmid. Multi-modal transformer for video retrieval. In ECCV, pages 214-229. Springer, 2020. 5, 6, 8" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 499, + 288, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 499, + 288, + 544 + ], + "spans": [ + { + "bbox": [ + 48, + 499, + 288, + 544 + ], + "type": "text", + "content": "[11] Zijian Gao, Jingyu Liu, Sheng Chen, Dedan Chang, Hao Zhang, and Jinwei Yuan. Clip2tv: An empirical study on transformer-based methods for video-text retrieval. arXiv preprint arXiv:2111.05610, 2021. 1, 2, 6, 8" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 545, + 288, + 587 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 545, + 288, + 587 + ], + "spans": [ + { + "bbox": [ + 48, + 545, + 288, + 587 + ], + "type": "text", + "content": "[12] Yuying Ge, Yixiao Ge, Xihui Liu, Dian Li, Ying Shan, Xiaohu Qie, and Ping Luo. Bridging video-text retrieval with multiple choice questions. In CVPR, pages 16167-16176, 2022. 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 590, + 288, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 590, + 288, + 633 + ], + "spans": [ + { + "bbox": [ + 48, + 590, + 288, + 633 + ], + "type": "text", + "content": "[13] Satya Krishna Gorti, Noel Vouitsis, Junwei Ma, Keyvan Golestan, Maksims Volkovs, Animesh Garg, and Guangwei Yu. X-pool: Cross-modal language-video attention for text-video retrieval. In CVPR, pages 5006-5015, 2022. 1, 6, 8" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 635, + 288, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 635, + 288, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 635, + 288, + 689 + ], + "type": "text", + "content": "[14] Priya Goyal, Piotr Dólar, Ross Girshick, Pieter Noordhuis, Lukasz Wesolowski, Aapo Kyrola, Andrew Tulloch, Yangqing Jia, and Kaiming He. Accurate, large minibatch sgd: Training imagenet in 1 hour. arXiv preprint arXiv:1706.02677, 2017. 5" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 691, + 288, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 288, + 713 + ], + "type": "text", + "content": "[15] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 32, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 106 + ], + "type": "text", + "content": "Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In ICML, pages 4904-4916. PMLR, 2021. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 107, + 545, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 107, + 545, + 161 + ], + "spans": [ + { + "bbox": [ + 308, + 107, + 545, + 161 + ], + "type": "text", + "content": "[16] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In ICML, pages 4904-4916. PMLR, 2021. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 163, + 545, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 163, + 545, + 217 + ], + "spans": [ + { + "bbox": [ + 308, + 163, + 545, + 217 + ], + "type": "text", + "content": "[17] Omar Khattab and Matei Zaharia. Colbert: Efficient and effective passage search via contextualized late interaction over bert. In Proceedings of the 43rd International ACM SIGIR conference on research and development in Information Retrieval, pages 39-48, 2020. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 219, + 545, + 240 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 219, + 545, + 240 + ], + "spans": [ + { + "bbox": [ + 308, + 219, + 545, + 240 + ], + "type": "text", + "content": "[18] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In ICLR, 2015. 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 242, + 545, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 242, + 545, + 285 + ], + "spans": [ + { + "bbox": [ + 308, + 242, + 545, + 285 + ], + "type": "text", + "content": "[19] Jie Lei, Linjie Li, Luowei Zhou, Zhe Gan, Tamara L Berg, Mohit Bansal, and Jingjing Liu. Less is more: Clipbert for video-and-language learning via sparse sampling. In CVPR, pages 7331-7341, 2021. 5, 6, 8" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 286, + 545, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 286, + 545, + 330 + ], + "spans": [ + { + "bbox": [ + 308, + 286, + 545, + 330 + ], + "type": "text", + "content": "[20] Chengzhi Lin, Ancong Wu, Junwei Liang, Jun Zhang, Wenhang Ge, Wei-Shi Zheng, and Chunhua Shen. Text-adaptive multiple visual prototype matching for video-text retrieval. In NeurIPS, 2022. 5, 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 332, + 545, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 332, + 545, + 374 + ], + "spans": [ + { + "bbox": [ + 308, + 332, + 545, + 374 + ], + "type": "text", + "content": "[21] Yang Liu, Samuel Albanie, Arsha Nagrani, and Andrew Zisserman. Use what you have: Video retrieval using representations from collaborative experts. arXiv preprint arXiv:1907.13487, 2019. 5, 6, 8" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 376, + 545, + 419 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 376, + 545, + 419 + ], + "spans": [ + { + "bbox": [ + 308, + 376, + 545, + 419 + ], + "type": "text", + "content": "[22] Yuqi Liu, Pengfei Xiong, Luhui Xu, Shengming Cao, and Qin Jin. Ts2-net: Token shift and selection transformer for text-video retrieval. In ECCV, pages 319-335. Springer, 2022. 5, 6, 8" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 422, + 545, + 454 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 422, + 545, + 454 + ], + "spans": [ + { + "bbox": [ + 308, + 422, + 545, + 454 + ], + "type": "text", + "content": "[23] Jiasen Lu, Dhruv Batra, Devi Parikh, and Stefan Lee. Vilbert: Pretraining task-agnostic visiolinguistic representations for vision-and-language tasks. NeurIPS, 32, 2019. 4" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 456, + 545, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 456, + 545, + 498 + ], + "spans": [ + { + "bbox": [ + 308, + 456, + 545, + 498 + ], + "type": "text", + "content": "[24] Huaishao Luo, Lei Ji, Ming Zhong, Yang Chen, Wen Lei, Nan Duan, and Tianrui Li. Clip4clip: An empirical study of clip for end to end video clip retrieval and captioning. Neurocomputing, 508:293-304, 2022. 1, 2, 4, 5, 6, 8" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 501, + 545, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 501, + 545, + 544 + ], + "spans": [ + { + "bbox": [ + 308, + 501, + 545, + 544 + ], + "type": "text", + "content": "[25] Yiwei Ma, Guohai Xu, Xiaoshuai Sun, Ming Yan, Ji Zhang, and Rongrong Ji. X-clip: End-to-end multi-grained contrastive learning for video-text retrieval. In ACM MM, pages 638–647, 2022. 5" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 545, + 545, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 545, + 545, + 589 + ], + "spans": [ + { + "bbox": [ + 308, + 545, + 545, + 589 + ], + "type": "text", + "content": "[26] Antoine Miech, Jean-Baptiste Alayrac, Lucas Smaira, Ivan Laptev, Josef Sivic, and Andrew Zisserman. End-to-end learning of visual representations from uncurated instructional videos. In CVPR, pages 9879-9889, 2020. 8" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 308, + 590, + 545, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 590, + 545, + 644 + ], + "spans": [ + { + "bbox": [ + 308, + 590, + 545, + 644 + ], + "type": "text", + "content": "[27] Antoine Miech, Dimitri Zhukov, Jean-Baptiste Alayrac, Makarand Tapaswi, Ivan Laptev, and Josef Sivic. Howto100m: Learning a text-video embedding by watching hundred million narrated video clips. In ICCV, pages 2630-2640, 2019. 5, 8" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 308, + 647, + 545, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 647, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 308, + 647, + 545, + 689 + ], + "type": "text", + "content": "[28] Shaobo Min, Weijie Kong, Rong-Cheng Tu, Dihong Gong, Chengfei Cai, Wenzhe Zhao, Chenyang Liu, Sixiao Zheng, Hongfa Wang, Zhifeng Li, et al. Hunyuan_tvr for text-video retrivial. arXiv preprint arXiv:2204.03382, 2022. 1, 5, 8" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 308, + 691, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 691, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 691, + 545, + 713 + ], + "type": "text", + "content": "[29] Mandela Patrick, Po-Yao Huang, Yuki Asano, Florian Metze, Alexander G Hauptmann, Joao F Henriques, and An" + } + ] + } + ], + "index": 31 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "10712" + } + ] + } + ], + "index": 33 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 66, + 72, + 287, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 72, + 287, + 95 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 287, + 95 + ], + "type": "text", + "content": "drea Vedaldi. Support-set bottlenecks for video-text representation learning. In ICLR, 2021. 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 95, + 287, + 159 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 95, + 287, + 159 + ], + "spans": [ + { + "bbox": [ + 48, + 95, + 287, + 159 + ], + "type": "text", + "content": "[30] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, pages 8748-8763. PMLR, 2021. 1, 2, 3, 5, 6, 8" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 161, + 287, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 161, + 287, + 194 + ], + "spans": [ + { + "bbox": [ + 48, + 161, + 287, + 194 + ], + "type": "text", + "content": "[31] Alec Radford, Jeff Wu, Rewon Child, David Luan, Dario Amodei, and Ilya Sutskever. Language models are unsupervised multitask learners. 2019. 1, 2, 3, 8" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 195, + 287, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 195, + 287, + 237 + ], + "spans": [ + { + "bbox": [ + 48, + 195, + 287, + 237 + ], + "type": "text", + "content": "[32] Yixuan Su, Tian Lan, Yahui Liu, Fangyu Liu, Dani Yogatama, Yan Wang, Lingpeng Kong, and Nigel Collier. Language models can see: Plugging visual controls in text generation. arXiv preprint arXiv:2205.02655, 2022. 8" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 239, + 287, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 239, + 287, + 270 + ], + "spans": [ + { + "bbox": [ + 48, + 239, + 287, + 270 + ], + "type": "text", + "content": "[33] Yoad Tewel, Yoav Shalev, Roy Nadler, Idan Schwartz, and Lior Wolf. Zero-shot video captioning with evolving pseudotokens. arXiv preprint arXiv:2207.11100, 2022. 2, 3, 8" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 272, + 287, + 314 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 272, + 287, + 314 + ], + "spans": [ + { + "bbox": [ + 48, + 272, + 287, + 314 + ], + "type": "text", + "content": "[34] Yoad Tewel, Yoav Shalev, Idan Schwartz, and Lior Wolf. Zerocap: Zero-shot image-to-text generation for visual-semantic arithmetic. In CVPR, pages 17918-17928, 2022. 2, 3, 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 316, + 287, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 316, + 287, + 358 + ], + "spans": [ + { + "bbox": [ + 48, + 316, + 287, + 358 + ], + "type": "text", + "content": "[35] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In NeurIPS, pages 5998-6008, 2017. 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 361, + 287, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 361, + 287, + 403 + ], + "spans": [ + { + "bbox": [ + 48, + 361, + 287, + 403 + ], + "type": "text", + "content": "[36] Qiang Wang, Yanhao Zhang, Yun Zheng, Pan Pan, and Xian-Sheng Hua. Disentangled representation learning for text-video retrieval. arXiv preprint arXiv:2203.07111, 2022. 1, 2, 4, 5, 6, 8" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 404, + 287, + 448 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 404, + 287, + 448 + ], + "spans": [ + { + "bbox": [ + 48, + 404, + 287, + 448 + ], + "type": "text", + "content": "[37] Wenzhe Wang, Mengdan Zhang, Runnan Chen, Guanyu Cai, Penghao Zhou, Pai Peng, Xiaowei Guo, Jian Wu, and Xing Sun. Dig into multi-modal cues for video retrieval with hierarchical alignment. In *IJCAI*, pages 1113–1121, 2021. 8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 449, + 287, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 449, + 287, + 491 + ], + "spans": [ + { + "bbox": [ + 48, + 449, + 287, + 491 + ], + "type": "text", + "content": "[38] Xin Wang, Jiawei Wu, Junkun Chen, Lei Li, Yuan-Fang Wang, and William Yang Wang. Vatex: A large-scale, high-quality multilingual dataset for video-and-language research. In ICCV, pages 4581-4591, 2019. 2, 5, 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 493, + 287, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 493, + 287, + 525 + ], + "spans": [ + { + "bbox": [ + 48, + 493, + 287, + 525 + ], + "type": "text", + "content": "[39] Xiaohan Wang, Linchao Zhu, and Yi Yang. T2vlad: global-local sequence alignment for text-video retrieval. In CVPR, pages 5079-5088, 2021. 6, 8" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 526, + 287, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 526, + 287, + 568 + ], + "spans": [ + { + "bbox": [ + 48, + 526, + 287, + 568 + ], + "type": "text", + "content": "[40] Xiaohan Wang, Linchao Zhu, Zhedong Zheng, Mingliang Xu, and Yi Yang. Align and tell: Boosting text-video retrieval with local alignment and fine-grained supervision. IEEE TMM, 2022. 8" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 570, + 287, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 570, + 287, + 601 + ], + "spans": [ + { + "bbox": [ + 48, + 570, + 287, + 601 + ], + "type": "text", + "content": "[41] Wenhao Wu, Zhun Sun, and Wanli Ouyang. Revisiting classifier: Transferring vision-language models for video recognition. In AAAI, 2023. 8" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 603, + 287, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 603, + 287, + 646 + ], + "spans": [ + { + "bbox": [ + 48, + 603, + 287, + 646 + ], + "type": "text", + "content": "[42] Wenhao Wu, Xiaohan Wang, Haipeng Luo, Jingdong Wang, Yi Yang, and Wanli Ouyang. Bidirectional cross-modal knowledge exploration for video recognition with pretrained vision-language models. In CVPR, 2023. 8" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 647, + 287, + 680 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 647, + 287, + 680 + ], + "spans": [ + { + "bbox": [ + 48, + 647, + 287, + 680 + ], + "type": "text", + "content": "[43] Zuxuan Wu, Ting Yao, Yanwei Fu, and Yu-Gang Jiang. Deep learning for video classification and captioning. In Frontiers of multimedia research, pages 3-29. 2017. 2, 5, 6" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 681, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 681, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 681, + 287, + 712 + ], + "type": "text", + "content": "[44] Jun Xu, Tao Mei, Ting Yao, and Yong Rui. Msr-vtt: A large video description dataset for bridging video and language. In CVPR, pages 5288-5296, 2016. 2, 5, 6" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 273 + ], + "type": "list", + "angle": 0, + "index": 21, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "text", + "content": "[45] Lewei Yao, Runhui Huang, Lu Hou, Guansong Lu, Minzhe Niu, Hang Xu, Xiaodan Liang, Zhenguo Li, Xin Jiang, and Chunjing Xu. Filip: Fine-grained interactive language-image pre-training. In ICLR, 2021. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 118, + 545, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 545, + 160 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 545, + 160 + ], + "type": "text", + "content": "[46] Jiahui Yu, Zirui Wang, Vijay Vasudevan, Legg Yeung, Mojtaba Seyhedhosseini, and Yonghui Wu. Coca: Contrastive captioners are image-text foundation models. arXiv preprint arXiv:2205.01917, 2022. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 163, + 545, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 163, + 545, + 217 + ], + "spans": [ + { + "bbox": [ + 307, + 163, + 545, + 217 + ], + "type": "text", + "content": "[47] Lu Yuan, Dongdong Chen, Yi-Ling Chen, Noel Codella, Xiyang Dai, Jianfeng Gao, Houdong Hu, Xuedong Huang, Boxin Li, Chunyuan Li, et al. Florence: A new foundation model for computer vision. arXiv preprint arXiv:2111.11432, 2021. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 219, + 545, + 273 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 219, + 545, + 273 + ], + "spans": [ + { + "bbox": [ + 307, + 219, + 545, + 273 + ], + "type": "text", + "content": "[48] Shuai Zhao, Linchao Zhu, Xiaohan Wang, and Yi Yang. Centerclip: Token clustering for efficient text-video retrieval. In Proceedings of the 45th International ACM SIGIR Conference on Research and Development in Information Retrieval, pages 970–981, 2022. 6, 8" + } + ] + } + ], + "index": 20 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10713" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/CapDet_ Unifying Dense Captioning and Open-World Detection Pretraining/c3eddfd4-ae1f-40f7-a7cf-96c54f1fea7b_content_list.json b/2023/CapDet_ Unifying Dense Captioning and Open-World Detection Pretraining/c3eddfd4-ae1f-40f7-a7cf-96c54f1fea7b_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..29b5340aee96194a975682a024ef992e41356917 --- /dev/null +++ b/2023/CapDet_ Unifying Dense Captioning and Open-World Detection Pretraining/c3eddfd4-ae1f-40f7-a7cf-96c54f1fea7b_content_list.json @@ -0,0 +1,1359 @@ +[ + { + "type": "text", + "text": "CapDet: Unifying Dense Captioning and Open-World Detection Pretraining", + "text_level": 1, + "bbox": [ + 101, + 128, + 867, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yanxin Long $^{1*}$ Youpeng Wen $^{1*}$ Jianhua Han $^{2}$ Hang Xu $^{2}$ Pengzhen Ren $^{1}$ Wei Zhang $^{2}$ Shen Zhao $^{1\\dagger}$ Xiaodan Liang $^{1,3\\dagger}$", + "bbox": [ + 168, + 179, + 799, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1Shenzhen Campus of Sun Yat-sen University 2Huawei Noah's Ark Lab 3MBZUAI longyx9@mail2.sysu.edu.cn, wenyoupeng0@outlook.com, hanjianhua4@huawei.com, chromexbjxh@gmail.com, renpzh@mail.sysu.edu.cn, wz.zhang@huawei.com, z-s-06@163.com, xdliang328@gmail.com", + "bbox": [ + 153, + 217, + 812, + 284 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 321, + 312, + 335 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Benefiting from large-scale vision-language pre-training on image-text pairs, open-world detection methods have shown superior generalization ability under the zero-shot or few-shot detection settings. However, a pre-defined category space is still required during the inference stage of existing methods and only the objects belonging to that space will be predicted. To introduce a \"real\" open-world detector, in this paper, we propose a novel method named CapDet to either predict under a given category list or directly generate the category of predicted bounding boxes. Specifically, we unify the open-world detection and dense caption tasks into a single yet effective framework by introducing an additional dense captioning head to generate the region-grounded captions. Besides, adding the captioning task will in turn benefit the generalization of detection performance since the captioning dataset covers more concepts. Experiment results show that by unifying the dense caption task, our CapDet has obtained significant performance improvements (e.g., $+2.1\\%$ mAP on LVIS rare classes) over the baseline method on LVIS (1203 classes). Besides, our CapDet also achieves state-of-the-art performance on dense captioning tasks, e.g., $15.44\\%$ mAP on VG V1.2 and $13.98\\%$ on the VG-COCO dataset.", + "bbox": [ + 75, + 347, + 473, + 694 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 702, + 207, + 717 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Most state-of-the-art object detection methods [33, 34, 50] benefit from a large number of densely annotated detection datasets (e.g., COCO [27], Object365 [36], LVIS [12]). However, this closed-world setting results in the model only being able to predict categories that appear in the training set. Considering the ubiquity of new concepts in real-world scenes, it is very challenging to locate and identify these new visual concepts. This predictive ability of new concepts in open-world scenarios has very important research value", + "bbox": [ + 75, + 728, + 468, + 863 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/366721316af1b64ee03ded007cf3162f6fdcbe9b26aaedea1fbbff23ca67b393.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 504, + 319, + 888, + 458 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/49d4c0e56a0c81fa4d585e5abfa1ce329a3cc45826d6c5e3b428f87956621d58.jpg", + "image_caption": [ + "Figure 1. Comparison of the different model predictions under OWD, OVD, and our setting. (a) OWD methods [14, 18, 48] are not able to describe the detailed category of the detected unknown objects and (b) the performance of OVD methods [8, 12, 41] usually depends on the pre-defined category list during the inference. (c) With the unification of two pipelines of dense captioning and open-world detection pre-training, our CapDet can either predict under a given category list or directly generate the description of predicted bounding boxes." + ], + "image_footnote": [], + "bbox": [ + 504, + 460, + 888, + 612 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "in real-world applications such as object search [29, 30], instance registration [45], and human-object interaction modeling [10].", + "bbox": [ + 496, + 744, + 890, + 790 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Currently, the open world scenario mainly includes two tasks: open world object detection [18] (OWD) and open-vocabulary object detection [44] (OVD). Although the paradigms of OWD and OVD tasks are closer to the real world, the former cannot describe the specific concept of the detected unknown objects and requires a pre-defined category list during the inference. Specifically, as shown", + "bbox": [ + 496, + 794, + 890, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 807, + 44 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Equal contribution. † Corresponding authors.", + "bbox": [ + 99, + 875, + 233, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "15233", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "in Figure 1, previous OWD methods [14,18,48] would recognize new concepts not in the predefined category space as \"unknown\". Further, another line of task OVD requires the model to learn a limited base class and generalize to novel classes. Compared to the zero-shot object detection (ZSD) proposed by [32], OVD allows the model to use external knowledge, e.g., knowledge distillation from a largescale vision-language pre-trained model [8, 12], image-caption pairs [44], image classification data [49], grounding data [25, 41, 46]. With the external knowledge, OVD methods show a superior generalization capacity to detect the novel classes within a given category space. However, as shown in Figure 1, when given an incomplete category list, OVD can only predict the concepts that appear in the given category list, otherwise, there will be recognition errors, (i.e., as illustrated in Figure 1 (b), the OVD methods prone to predict the \"wall socket\" as \"remote\", since the latter is in the category list but not the former).", + "bbox": [ + 75, + 90, + 472, + 363 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Thus, under the OVD setting, we mainly face the following two challenges: (i) it is difficult to define a complete list of categories; (ii) low response values on rare categories often lead to recognition errors. This is mainly because we cannot exhaustively enumerate new objects in the real world, and secondly, it is difficult to collect enough samples for rare classes. However, the fact that rare objects in the real world, even some new objects that are unknown to humans, such as UFOs, do not prevent people from using natural language to describe it as \"a flying vehicle that looks like a Frisbee\".", + "bbox": [ + 75, + 364, + 472, + 532 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Therefore, based on the above observations, in this paper, we consider a new setting that is closer to the open world and real scenes, i.e., we expect the model to both detect and recognize concepts in a given category list, and to generate corresponding natural language descriptions for new concepts or rare categories of objects. Early dense captioning methods [9, 17] can locate salient regions in images and generate the region-grounded captions with natural language. Inspired by this, to address the challenges faced in the OVD setting, we propose to unify the two pipelines of dense captioning and open-world detection pre-training into one training framework, called CapDet. It empowers the model with the ability to both accurately detect and recognize common object categories and generate dense captions for unknown and rare categories by unifying the two training tasks.", + "bbox": [ + 75, + 534, + 472, + 777 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Specifically, our CapDet constructs a unified data format for the dense captioning data and detection data. With the data unification, CapDet further adopts a unified pretraining paradigm including open-world object detection and dense captioning pre-training. For open-world detection pretraining, we treat the detection task as a semantic alignment task and adopt a dual encoder structure as [41] to locate and predict the given concepts list. The concepts", + "bbox": [ + 75, + 779, + 472, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "list contains category names in detection data and region-grounded captions in dense captioning data. For dense captioning pretraining, CapDet proposes a dense captioning head to take the predicted proposals as input to generate the region-grounded captions. Due to the rich visual concepts in the dense captioning data, the integration of dense captioning tasks will in turn benefit the generalization of detection performance.", + "bbox": [ + 496, + 90, + 893, + 210 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our experiments show that the integration of few dense captioning data brings in large improvement in the object detection datasets LVIS, e.g., $+2.7\\%$ mAP on LVIS. The unification of dense captioning and detection pre-training gains an additional $2.3\\%$ increment on LVIS and $2.1\\%$ increment on LVIS rare classes. Besides, our model also achieves state-of-the-art performance on dense captioning tasks. Note that our method is the first to unify dense captioning and open-world detection pretraining.", + "bbox": [ + 496, + 210, + 893, + 348 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To summarize, our contributions are three folds:", + "bbox": [ + 519, + 348, + 838, + 362 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose a novel open-vocabulary object detection framework CapDet, which cannot only detect and recognize concepts in a given category list but also generate corresponding natural language descriptions for new concept objects.", + "- We propose to unify the two pipelines of dense captioning and open-world detection pre-training into one training framework. Both two pre-training tasks are beneficial to each other.", + "- Experiments show that by unified dense captioning task and detection task, our CapDet gains significant performance improvements on the open-vocabulary object detection task (e.g., $+3.3\\%$ mAP on LVIS rare classes). Furthermore, our CapDet also achieves state-of-the-art performance on the dense captioning tasks, e.g., $15.44\\%$ mAP on Visual Genome (VG) V1.2 and $13.98\\%$ mAP on VG-COCO." + ], + "bbox": [ + 519, + 369, + 890, + 630 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 633, + 640, + 648 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Vision-Language Pre-training. Vision-Language Pretraining [6, 16, 31] as a scheme in the domains of natural language processing [1, 5] and computer vision [7] obtains continual attention currently. And it exhibits strong performance and generalization ability on various downstream vision and cross-modal tasks. Among them, CLIP [31] and ALIGN [16] as dual-stream methods utilize large-scale image-text pairs on the Internet by cross-modal contrastive learning to get excellent zero-shot classification ability. Single-stream methods [20, 24] unify visual and textual embeddings in a single transformer-based model, which can perform text generation tasks such as image caption and VQA. Some mixed architectures [23, 39] combine single-stream and dual-stream to explore a unified way of vision-language understanding and generation. However, these methods take low-resolution images as input and serve the", + "bbox": [ + 496, + 657, + 893, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "15234", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "task of classification and retrieval. Those vision-language pre-training approaches can not be applied to pure computer vision task directly, i.e., object detection task.", + "bbox": [ + 75, + 90, + 470, + 137 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Open World Object Detection / Open-Vocabulary Object Detection. Object detection is a core computer vision task, which aims at localizing objects using a bounding box and classifying them. The mature detection approaches which show great performance on supervised data include one-stage detectors (i.e.,YOLO [33],ATSS [47]) having a relatively high detection efficiency and two-stage detectors (i.e., Faster R-CNN [34],Mask R-CNN [15]) having good detection accuracy. However, how to generalize these methods to rare classes and novel concepts in the real world is a challenge. Currently, several object detection approaches for such open-world scenes have attracted extensive attention from academia and industry. These methods are divided into two tasks which are called open-world object detection and open-vocabulary object detection respectively depending on whether to detect the class of unknown classes.", + "bbox": [ + 73, + 143, + 468, + 400 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "For the OWD task, Zhao et al. [48] proposed a proposal advisor to assist in identifying unknown proposals without supervision and a class-specific expelling classifier to filter out confusing predictions. For the OVD task, GLIP [25] converts the detection data into grounding format and proposes a fusion module to learn semantic vision information in grounding data. K-Lite [38] reconstructs the input format of the data in GLIP from sequential to parallel and uses nouns hierarchy and definition to format text sequence. Det-CLIP [41] unifies detection, grounding, and image-text pair data in a paralleled formulation and constructs a concept dictionary to augment the text data, which strikes a balance between performance and efficiency. Differing from all these works, our CapDet can generate an open-set caption of each region proposal to cover situations where the semantics of new object instances are not in the given category list.", + "bbox": [ + 73, + 401, + 470, + 659 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Dense Captioning. Dense captioning aims at generating detailed descriptions for local regions, which usually needs to locate visual regions with semantic information and generate captions for these regions. J. Johnson et al. [17] utilized a fully convolutional localization network to locate regions of interest (RoIs) and then describe them. Afterward, many methods [26, 43] based on Faster-RCNN [34] and LSTM [11] are proposed to do dense captioning. X. Li et al. [26] arrange RoI features as a sequence and put them into LSTM with the guidance of the region features to form the complementary object context features. This method also needs ground truth bounding boxes auxiliary tests to achieve good results. But limited by the forget gate mechanism of LSTM, the inputted sequence cannot be too long. Then, the transformer-based method TDC [37] is proposed", + "bbox": [ + 75, + 674, + 468, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "to tackle the long sequence forgotten problem. Instead, our CapDet proposes a transformer-based caption head to generate a caption using a single-stage detector ATSS while simultaneously achieving open-world detection.", + "bbox": [ + 496, + 90, + 890, + 151 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 500, + 166, + 591, + 181 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The overview of our proposed CapDet is shown in Figure 2. To construct a detector to either predict under a given category list or directly generate the concepts of predicted bounding boxes, we incorporate detection data and dense caption data together. In this section, we will present a unified data format for the detection data and dense caption data in Section 3.1, the model architecture and pre-training objectives for open-world object detection pre-training in Section 3.2 and dense captioning in Section 3.3.", + "bbox": [ + 496, + 191, + 890, + 328 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Unified Formulation", + "text_level": 1, + "bbox": [ + 498, + 338, + 692, + 353 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We defined a unified triplet-wise data format $(x, \\{\\mathbf{b}_i\\}_{i=1}^N, y_{i=1}^N)$ for each sample from different sources. Specifically, $x \\in \\mathbb{R}^{3 \\times h \\times w}$ is the input image, $\\{\\mathbf{b}_i | \\mathbf{b}_i \\in \\mathbb{R}^4\\}_{i=1}^N$ denotes the bounding boxes coordinates for each region of the image, and the $y_{i=1}^N$ represents the concepts of the corresponding boxes. $N$ denotes the number of regions. A concept $y_i$ formatted as a sentence contains the category and textual description of the corresponding region. In detection data, a concept $y$ consists of the category name and the corresponding definition from the concept dictionary [41], while $y_i$ represents the region-grounded caption in dense caption data. For example, for an image $x$ in detection data, $y_i$ can be:", + "bbox": [ + 496, + 362, + 890, + 571 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\ny _ {i} = \\text {\" p e r s o n , a h u m a n b e i n g . \"}\n$$\n", + "text_format": "latex", + "bbox": [ + 580, + 574, + 807, + 590 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "For an image $x$ from dense captioning data, $y_{i}$ can be:", + "bbox": [ + 500, + 598, + 856, + 614 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\ny _ {i} = \\text {\" a n o u t l e t o n t h e w a l l .}\n$$\n", + "text_format": "latex", + "bbox": [ + 589, + 626, + 799, + 642 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "With the triplet, we can learn a unified image-text alignment objective on the detection data and the dense captioning data. The unified formulation also ensures the joint training of open-world object detection pre-training and dense captioning.", + "bbox": [ + 496, + 654, + 890, + 731 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Open-World Object Detection Pre-training", + "text_level": 1, + "bbox": [ + 500, + 739, + 864, + 758 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Based on the unified formulation of detection data and dense captioning data, we regard the captions of regions in dense captioning data as a kind of category and utilize two different sources of data for the open-world object detection pre-training. Compared with the limited class list of detection data, dense caption data contains richer concepts and more semantic information than class names of individual regions. On the other hand, localization and recognition are two essential tasks of object detection. Traditional object", + "bbox": [ + 496, + 763, + 890, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "15235", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/3edf93e562aaab7bfa54789f47818bdd10fd383c264bb50247aad7f92d884362.jpg", + "image_caption": [ + "Figure 2. The overall architecture of CapDet. The training paradigm of CapDet contains open-world object detection pre-training and dense captioning. In detection, CapDet contains a dual vision-language encoder. The image encoder generates region embeddings from detection and dense captioning data. The regression loss and centerness loss are introduced to regress the locations. The text encoder takes the category concepts as input to generate the embeddings from the [EOS] token. Then we treat the detection task as a matching task and adopt an alignment loss for the category embeddings and region embeddings. In dense captioning, an additional dense captioning head is proposed to take the region embeddings as input and generate the textual captions for corresponding regions with natural language." + ], + "image_footnote": [], + "bbox": [ + 138, + 90, + 831, + 431 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "detection always focuses on the salient objects in the image. While the dense captioning data contains lots of annotations which are just parts of an object, e.g., an ear of an elephant, it is not suitable to learn those annotations for the localization task. Therefore, we only calculate the localization loss on detection data.", + "bbox": [ + 75, + 537, + 470, + 626 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As shown in Figure 2, CapDet predicts the regions and treats the recognition task as a region-category matching task. For efficient learning on the matching task, we adopt the negative sampling proposed by [41] to provide negative concepts to enlarge the concept space in a batch. Specifically, for each iteration, we randomly sample a negative concept set and add to the positive concept set (N samples) in a batch to obtain the final concept set $y_{i = 1}^{M}$ , where M represents the sum of the number of positive and negative samples. Finally, we format the triplet to $(x,\\{\\mathbf{b}_i\\}_{i = 1}^N,y_{i = 1}^M)$ .", + "bbox": [ + 75, + 628, + 468, + 780 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "CapDet contains a dual vision-language encoder and takes the triplet $(x, \\{\\mathbf{b}_i\\}_{i=1}^N, y_{i=1}^M)$ as input. The image encoder $\\Phi_v$ is an object detector that can predict the bounding boxes of regions from the input image $x$ and output the region features $O \\in \\mathbb{R}^{K \\times D}$ . The text encoder $\\Phi_l$ takes the concept set $y_{i=1}^M$ as input and obtains the text embeddings $W \\in \\mathbb{R}^{M \\times D}$ from the special token [EOS] concatenated with the text input. $K, D$ denotes the number of predicted", + "bbox": [ + 75, + 780, + 470, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "regions and region feature dimensions. The alignment score matrix $S \\in \\mathbb{R}^{K \\times M}$ of regions and texts is calculated by:", + "bbox": [ + 498, + 537, + 890, + 568 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nO = \\Phi_ {v} (x), W = \\Phi_ {l} \\left(y _ {i = 1} ^ {M}\\right), S = O W ^ {T} \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 560, + 580, + 890, + 599 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $T$ denotes the transpose operation. A ground-truth alignment matrix $G \\in \\{0,1\\}^{K \\times M}$ is constructed to indicate the matching relation of regions and concepts. The alignment loss $\\mathcal{L}_{\\text{align}}$ is calculated by the predicted alignment scores of regions $S$ and the ground-truth alignment matrix $G$ . Following [25, 41], we adopt the ATSS [47] detector as an image encoder, and $\\mathcal{L}_{\\text{align}}$ is typically a sigmoid focal loss. As a one-stage detector, the localization loss contains centeredness loss $\\mathcal{L}_{\\text{cen}}$ and bounding box regression loss $\\mathcal{L}_{\\text{reg}}$ . The training objective of detection pre-training can be written as:", + "bbox": [ + 496, + 614, + 892, + 781 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\left\\{ \\begin{array}{l l} \\mathcal {L} _ {\\text {a l i g n}} + \\alpha \\mathcal {L} _ {\\text {r e g}} + \\beta \\mathcal {L} _ {\\text {c e n t e r}}, & \\text {f o r d e t e c t i o n} \\\\ \\mathcal {L} _ {\\text {a l i g n}}, & \\text {f o r d e n s e c a p t i o n i n g} \\end{array} \\right. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 503, + 794, + 890, + 845 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\alpha$ and $\\beta$ denote the weights for the centerness loss $\\mathcal{L}_{cen}$ and box regression loss $\\mathcal{L}_{reg}$ , respectively. The $\\mathcal{L}_{cen}$ is the sigmoid loss and the $\\mathcal{L}_{reg}$ is the GIOU loss [35].", + "bbox": [ + 496, + 848, + 890, + 893 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "15236", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/4508f6c0ede534f23083982b9625d33c5dfd9b80791abbebdd061d9f29e03eeb.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MODELBACKBONEPRE-TRAIN DATAIMAGES NUMBERAPLVIS APr / APc / APf
MASK-RCNN [15]SWIN-TLVIS0.1M34.119.1 / 34.0 / 37.0
ATSS [47]SWIN-TLVIS0.1M33.619.7 / 32.4 / 37.2
ATSS [47]SWIN-LLVIS0.1M43.930.6 / 43.7 / 46.3
MDETR [19]RN101GOLDG+0.77M24.220.9 / 24.3 / 24.2
GLIP-T(A) [25]SWIN-T+DH+FO3650.66M18.514.2 / 13.9 / 23.4
GLIP-T(C) [25]SWIN-T+DH+FO365,GOLDG1.43M24.917.7 / 19.5 / 31.0
GLIP-T [25]SWIN-T+DH+FO365,GOLDG,CAP4M5.43M26.020.8 / 21.4 / 31.0
K-LITE [38]SWIN-TO3650.66M21.314.8 / 18.6 / 24.8
K-LITE [38]SWIN-TO365,GOLDG1.43M26.117.2 / 24.6 / 29.0
GLIPv2-T [46]SWIN-T+DH+FO365,GOLDG,CAP4M5.43M29.0- / - / -
DETCLIP-T(A) [41]SWIN-TO3650.66M28.826.0 / 28.0 / 30.0
DETCLIP-T(B) [41]SWIN-TO365, GOLDG1.43M34.426.9 / 33.9 / 36.3
DETCLIP-T(C)* [41]SWIN-TO365, VG0.73M31.527.5 / 30.6 / 33.0
CAPDET (OURS)SWIN-TO365, VG0.73M33.829.6 / 32.8 / 35.5
", + "bbox": [ + 76, + 89, + 898, + 366 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Table 1. Zero-shot performance on LVIS [13] MiniVal5k datasets. $\\mathrm{AP}_r / \\mathrm{AP}_c / \\mathrm{AP}_f$ indicate the AP values for rare, common, and frequent categories, respectively. \"DH\" and \"F\" in GLIP [25] baselines stand for the dynamic head [3] and cross-modal fusion, respectively. Baselines with * are implemented with our code base. GoldG+ denotes the GoldG plus the COCO [27] caption dataset.", + "bbox": [ + 75, + 377, + 892, + 420 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Dense Captioning", + "text_level": 1, + "bbox": [ + 76, + 458, + 250, + 474 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The open-world object detection pre-training ensures CapDet gains the capacity to detect under given an arbitrary category list. However, when the given category list is not complete enough to cover the potential classes on a new domain data, the detector will perform worse on the categories which are not in the given list. Considering such limitation, we propose a dense captioning head $\\Phi_C$ to generate semantically rich concepts with natural language for the predicted proposals. In the dense captioning task, the model receives an image and produces a set of regions and the corresponding captions. The dense captioning head is a cross-modal decoder that takes the c predicted regions features $O$ generated by the image encoder as input. The captioning (i.e., language modeling) loss is calculated by:", + "bbox": [ + 75, + 481, + 468, + 691 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {c a p} = - \\log p \\left(y _ {i t} \\mid \\Phi_ {c} \\left(y _ {i (\\tau < t)}, O _ {i}\\right)\\right), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 148, + 702, + 468, + 720 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $y_{it}$ means the $t$ token in caption $y_{i}$ corresponding to region feature $O_{i}$ , and $y_{i(\\tau < t)}$ means tokens before $t$ in caption $y_{i}$ . The overall pre-training loss can be written as:", + "bbox": [ + 75, + 729, + 468, + 773 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = w _ {d} \\mathcal {L} _ {\\text {d e t}} + w _ {c} \\mathcal {L} _ {\\text {c a p}}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 191, + 784, + 468, + 801 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $w_{d}, w_{c}$ denote the weighting factor of $\\mathcal{L}_{det}$ and $\\mathcal{L}_{cap}$ .", + "bbox": [ + 76, + 810, + 467, + 825 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To minimize the gap in the type of bounding boxes between the detection data and dense captioning data, we propose a simple way to transform our detector as a class-agnostic detector and only select the top $k$ regions based on the centeredness scores to adapt to the dense captioning", + "bbox": [ + 75, + 825, + 468, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "task. We can fine-tune our CapDet on the dense captioning data to achieve better performance. Specifically, we propose \"object\" as the foreground concept and \"background\" as the background concept. The text encoder $\\Phi_l$ outputs the concept embeddings $W' \\in \\mathbb{R}^{2 \\times D}$ . Then the alignment scores $S' \\in \\mathbb{R}^{K \\times 2}$ is calculated by Eqn. 1. The captioning head takes the top $k$ most confident proposal embeddings based on centeredness scores as input to predict the region-grounded captions.", + "bbox": [ + 496, + 458, + 890, + 595 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiment", + "text_level": 1, + "bbox": [ + 500, + 614, + 624, + 633 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Implementation Details. For the image encoder, we adopt the Swin-T backbone proposed in SwinTransformer [28] which is pre-trained on ImageNet-1K [4]. We use 12 layers 8 heads transformer as our text encoder and load a base model checkpoint released by FILIP [42], in order to make a fair comparison with DetCLIP [41]. The structure of the dense captioning head is consistent with that in the text encoder but trained from scratch for a fair comparison. We employ AdamW [21] optimizer and set the batch size to 32. The learning rate is set to $1.4 \\times 10^{-4}$ for the parameters of the image encoder and detection head, and $1.4 \\times 10^{-5}$ for the text encoder and dense captioning head. When fine-tuning the VG dataset to do the dense captioning task, we set the learning rate to $1.4 \\times 10^{-4}$ . Without otherwise specified, all models are trained with 12 epochs and the learning rate is decayed with a factor of 0.1 at the 8-th and the 11-th epoch. The context token", + "bbox": [ + 496, + 643, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "15237", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/8f647bdc722a93a11d7ce02114dbac662df17023a0fec08ba576c8b005716469.jpg", + "image_caption": [ + "Figure 3. Qualitative visualizations between GLIP-T, DetCLIP-T(C) and CapDet. From top to down, the three rows of images show the LVIS zero-shot detection results of GLIP-T, DetCLIP-T(C), and CapDet respectively. All models are pre-trained on O365 and VG." + ], + "image_footnote": [], + "bbox": [ + 81, + 90, + 890, + 320 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "length for input text is set to 20. We set the number of input captions to 150, and the number of the region features N is determined by the feature map. The loss weight factor $w_{c}$ and $w_{d}$ are both set to 1.0. We build our model on MMDetection [2] code base.", + "bbox": [ + 75, + 398, + 470, + 474 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Dataset. Our CapDet is trained with two types of data, including detection data and caption data. Following DetCLIP [41], we use Object365 [36] (it will be abbreviated as O365 in the following paper) as detection data, and sample 0.66M data from O365 v2 for training. Following GLIP [25] and DetCLIP [41], LVIS [13] MiniVal5k (defined in [19]) which has 5000 images is used for detection evaluation. Moreover, we remove the training samples contained in the LVIS dataset for fair zero-transfer evaluation. For dense captioning data, we mainly conduct our experiments on VG [22] V1.2 and VG-COCO (defined in [37]). Following [37], we allocate 77398 images for training and 5000 images for validation and testing on VG. As demonstrated in [22], the ground-truth bounding boxes of VG are much denser than the other object detection datasets, i.e., the average number of per sample in MS COCO [27] is only 7.1 vs. 35.4 in VG. Then an intersection of VG V1.2 and MS COCO is proposed by [37] and is denoted as VG-COCO, which has 38080 images for training, 2489 for validation, and 2476 for testing.", + "bbox": [ + 75, + 498, + 468, + 801 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Benchmark Settings. We mainly evaluate our method on open-vocabulary object detection and dense captioning task. For open-vocabulary object detection, we evaluate the direct domain transfer on LVIS [13] which contains 1203 categories. Following [25, 41], we metric the zero-shot", + "bbox": [ + 75, + 824, + 468, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "detection performance by the Average Precision (AP) on a 5k subset. The annotations of LVIS data are split into three folds, i.e., rare, common, and frequency, based on the number of categories. Since there is almost no overlap between the rare classes and the classes of training dataset Objects365 [36], the AP of the rare classes shows a valuable zero-shot detection performance. For the dense captioning task, we follow the setting of [17] to evaluate the VG and VG-COCO. The evaluation metric we adopt is the mean Average Precision proposed by [17] which is calculated across a range of thresholds for both localization and language accuracy, i.e., the intersection over union (IOU) thresholds .3, .4, .5, .6, .7 are used for localization and the METEOR score' thresholds 0, .05, .1, .15, .2, .25 is adopted for evaluating the language generation.", + "bbox": [ + 496, + 398, + 893, + 627 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Open-world Detection Results", + "text_level": 1, + "bbox": [ + 498, + 633, + 767, + 648 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1 shows our zero-shot object detection performance on LVIS. We mainly train our CapDet with the backbone Swin-T [28] on the detection data Objects365 [36] and dense captioning data (VG [22]). Since DetCLIP does not report the performance on O365 and VG, we train DetCLIP on the two datasets under the same settings and denote it as DetCLIP-T(C) for a fair comparison. Comparing the 11th row and 12th row, our CapDet outperforms DetCLIP-T(C) on the same data scale and backbone with an extra simple caption head. Moreover, our model's zero-shot performance even surpasses the fully-supervised model with the same backbone by a large margin on rare classes, i.e., CapDet outperforms ATSS by $9.9\\%$ .", + "bbox": [ + 496, + 656, + 890, + 852 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Qualitative Visualizations Figure. 3 illustrates the detection results on LVIS [22] dataset from GLIP-T, DetCLIP-T(C), and CapDet. All three models are trained on O365", + "bbox": [ + 496, + 854, + 893, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "15238", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/9ed982972812d1c237f8b600b279b11086cb1ea5113a52fc4e174c9aae3b8d5b.jpg", + "image_caption": [ + "Figure 4. Qualitative visualizations between JIVC and CapDet. \"w/o ft\" means do caption without finetune, while \"w/ ft\" means with finetune." + ], + "image_footnote": [], + "bbox": [ + 86, + 90, + 883, + 458 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "and VG, and details are given in Section 4.3. Given a category list, the rare classes are detected more precisely by our CapDet, e.g., \"kitchen table\" in the first column, \"horse buggy\" in the third column, and \"fishbowl\" in the sixth column that our model CapDet detects correctly but the other two not.", + "bbox": [ + 75, + 505, + 470, + 595 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. Dense Captioning Results", + "text_level": 1, + "bbox": [ + 76, + 606, + 310, + 622 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Due to the target bounding boxes in dense captioning data containing lots of local structures of objects and being much denser than the bounding boxes in object detection data, we do not regress the bounding box in the pre-training stage. The previous works directly train on the dense captioning data and generate captions on the top $k$ proposals ranking by a confidence score. When fine-tuning our model on the VG dataset for the dense captioning tasks, we transform our CapDet into a class-agnostic detector. Specifically, we propose \"object\" as the foreground concept and \"background\" as the background concept for computing alignment scores. The scores are used as proposal confidences to predict the region-grounded captions.", + "bbox": [ + 75, + 628, + 468, + 824 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 2 and Table 3 show CapDet significantly outperforms the latest work TDC [37] by $2.5\\%$ on mAP on VG and TDC+ROCSU [37] by $2.08\\%$ , respectively. It is worth noticing that, even against given the ground-truth bounding boxes with the previous method COCG [26] denoted", + "bbox": [ + 75, + 825, + 470, + 902 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/8b07a72f56d866358c042469446500aeb6990624de40394e9abda968c0bfb63c.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodmAP(%)
FCLN [17]5.16
JIVC [40]9.96
ImgG [26]9.68
COCD [26]9.75
COCG [26]10.39
CAG-Net [43]10.51
TDC [37]11.90
CapDet (Ours)15.44
", + "bbox": [ + 596, + 503, + 797, + 662 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 2. Comparison of mAP (%) performance on dense captioning benchmark on the VG V1.2 dataset.", + "bbox": [ + 498, + 672, + 890, + 700 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "as COCOG>, our CapDet still gains a $43.80\\%$ mAP increase and achieves state-of-the-art. One important reason is that the excellent detection performance of our model assists the localization ability of dense captioning tasks.", + "bbox": [ + 496, + 717, + 890, + 777 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Qualitative Visualizations. Figure 4 shows a qualitative visualization comparison between JIVC [40] and our CapDet. The three image rows from top to bottom are the visualization of JIVC, CapDet without fine-tuning, and CapDet with finetuning. In the second row, CapDet can locate more objects than JIVC, owing to our model's superior localization performance. After finetuning, CapDet can further describe a region rather than a single object such as", + "bbox": [ + 496, + 779, + 892, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "15239", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/75c8e679af8f535aa8600950e38054ca478204c2200ad135ad6f1f9168fa17d8.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodmAP(%)
FCLN [17]4.23
JIVC [40]7.85
Max Pooling [26]7.86
ImgG [26]7.81
COCD [26]7.92
COCG [26]8.90
COCG-LocSiz [26]8.76
COCG&GT [26]9.79
TDC+ROCSU [37]11.9
CapDet (Ours)13.98
", + "bbox": [ + 158, + 88, + 390, + 276 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/176bc80291b39f541bbb7d8a1554cb82554100fb22f74f28abafb47763bc5c36.jpg", + "table_caption": [ + "Table 3. Comparison of mAP (%) performance on the dense captioning benchmark on the VG-COCO Dataset." + ], + "table_footnote": [], + "table_body": "
MODELDC HEADLVIS
APAPr / APc / APf
GLIP-T×30.422.5 / 29.0 / 33.0
GLIP-T✓33.127.0 / 32.1 / 35.0
DETCLIP-T×31.527.5 / 30.6 / 33.0
DETCLIP-T✓33.829.6 / 32.8 / 35.5
", + "bbox": [ + 81, + 335, + 465, + 434 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/bafbe26e5dd3eb83f5862dd315bad36dc4591d31c4c8b1ed25fef67a4a7e8800.jpg", + "table_caption": [ + "Table 4. Ablations on integrating our dense captioning head into different baselines." + ], + "table_footnote": [], + "table_body": "
Pre-training DataFine-tuneDCap mAP(%)Box mAP(%)
VG✘12.8627.65
O365,VG✘4.729.65
VG✓13.8328.58
O365,VG✓15.4430.61
", + "bbox": [ + 81, + 478, + 465, + 556 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 5. Ablations on incorporating data from different sources. \"DCap\" stands for the dense caption mAP.", + "bbox": [ + 76, + 566, + 467, + 595 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "\"two women in a kitchen\" in the 5-th column.", + "bbox": [ + 76, + 626, + 380, + 640 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.3. Ablation Studies", + "text_level": 1, + "bbox": [ + 76, + 652, + 240, + 667 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.3.1 Ablations for Unified Pre-training", + "text_level": 1, + "bbox": [ + 76, + 676, + 367, + 694 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effect on different baselines. Table 4 investigates the advantages of dense captioning heads on different baselines. We integrate our dense captioning head with GLIP-T or DetCLIP-T. The GLIP-T is implemented with parallel text encoding without external knowledge following the setting as ablations in [38] on our code base. All the results are pre-trained on Objects365 and VG. The results show that our dense captioning head is able to boost the generalization and model-agnostic.", + "bbox": [ + 75, + 702, + 468, + 838 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effect of dense captioning data. Table 1 shows the efficiency of incorporating dense captioning data. Specifically, only $0.07\\mathrm{M}$ data added, the DetCLIP-T(C) gains $+2.7\\%$ overall AP and $+1.5\\%$ $AP_{r}$ on LVIS compared to DetCLIP-", + "bbox": [ + 75, + 839, + 468, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "T(A). The performance of DetCLP-T(A) on rare categories also outperforms DetCLIP-T(C) train on Objects365 and GOLDG, while the data size is 1.43M vs. 0.73M.", + "bbox": [ + 498, + 90, + 890, + 135 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.3.2 Ablations for dense captioning", + "text_level": 1, + "bbox": [ + 500, + 143, + 767, + 159 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We investigate the impact of training policy and data from different sources on the dense captioning task. As shown in row1 in Table 5, our CapDet still achieves a significant performance which is directly trained on VG outperforms the previous task (i.e., TDC [37] in Table 2). Row2 is our CapDet and is pre-trained on Objects365 and VG, while only the bounding box in the Objects365 is regressed, and then transformed on a dense captioning task. Since the type of bounding boxes in dense captioning is different from the detection data, the result of the direct transforming to dense captioning is worse. However, we've proved that our model still keeps the dense captioning capacity on the salient objects in Figure 4. The results in row3 and row4 indicate that pre-training on the detection data Objects365 is also beneficial to the dense captioning task.", + "bbox": [ + 496, + 167, + 890, + 393 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Limitations", + "text_level": 1, + "bbox": [ + 500, + 397, + 622, + 412 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "These are a few issues that we need to improve in the future: (1) Although our unification training paradigm works well on open-vocabulary object detection and dense captioning task, the training of dense captioning generation costs lots of time. (2) In addition, existing dense captioning data is high-cost to collect. We will research how to collect large-scale dense captioning data by auto annotation and get better performance with the scaled-up data.", + "bbox": [ + 496, + 421, + 890, + 542 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 549, + 617, + 564 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we propose a novel open-world object detection method named CapDet. Our CapDet is more practical in the open world and real scenes. Specifically, CapDet introduces a unification training framework including open-world object detection pre-training and dense captioning. The unification enables our CapDet to localize and recognize concepts in an arbitrary given category list or directly generate textual captions for predicted new concept objects. Experiments show that the design of unification is both beneficial to open-world object detection tasks and dense captioning tasks. In the future, our CapDet can be easily injected into any open world and real scenes tasks. The unification framework can also be integrated into any other OWD/OVD methods to generate semantic-rich concepts for unknown/novel objects.", + "bbox": [ + 496, + 574, + 890, + 801 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements We gratefully acknowledge the support of MindSpore1, CANN (Compute Architecture for Neural Networks) and Ascend AI Processor used for this research.", + "bbox": [ + 496, + 811, + 890, + 872 + ], + "page_idx": 7 + }, + { + "type": "page_footnote", + "text": "$^{1}$ https://www.mindspore.cn/", + "bbox": [ + 524, + 886, + 676, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "15240", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 174, + 106 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Tom B. Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert-Voss, Gretchen Krueger, Tom Henighan, Rewon Child, Aditya Ramesh, Daniel M. Ziegler, Jeffrey Wu, Clemens Winter, Christopher Hesse, Mark Chen, Eric Sigler, Mateusz Litwin, Scott Gray, Benjamin Chess, Jack Clark, Christopher Berner, Sam McCandlish, Alec Radford, Ilya Sutskever, and Dario Amodei. Language models are few-shot learners. CoRR, abs/2005.14165, 2020. 2", + "[2] Kai Chen, Jiaqi Wang, Jiangmiao Pang, Yuhang Cao, Yu Xiong, Xiaoxiao Li, Shuyang Sun, Wansen Feng, Ziwei Liu, Jiarui Xu, Zheng Zhang, Dazhi Cheng, Chenchen Zhu, Tianheng Cheng, Qijie Zhao, Buyu Li, Xin Lu, Rui Zhu, Yue Wu, Jifeng Dai, Jingdong Wang, Jianping Shi, Wanli Ouyang, Chen Change Loy, and Dahua Lin. Mmdetection: Open mmlab detection toolbox and benchmark. ArXiv, abs/1906.07155, 2019. 6", + "[3] Xiyang Dai, Yinpeng Chen, Bin Xiao, Dongdong Chen, Mengchen Liu, Lu Yuan, and Lei Zhang. Dynamic head: Unifying object detection heads with attentions. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7373-7382, 2021. 5", + "[4] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, K. Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. 2009 IEEE Conference on Computer Vision and Pattern Recognition, pages 248-255, 2009. 5", + "[5] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. BERT: pre-training of deep bidirectional transformers for language understanding. CoRR, abs/1810.04805, 2018. 2", + "[6] Xiaoyi Dong, Yinglin Zheng, Jianmin Bao, Ting Zhang, Dongdong Chen, Hao Yang, Ming Zeng, Weiming Zhang, Lu Yuan, Dong Chen, et al. Maskclip: Masked self-distillation advances contrastive language-image pretraining. arXiv preprint arXiv:2208.12262, 2022. 2", + "[7] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. CoRR, abs/2010.11929, 2020. 2", + "[8] Yu Du, Fangyun Wei, Zihe Zhang, Miaojing Shi, Yue Gao, and Guoqi Li. Learning to prompt for open-vocabulary object detection with vision-language model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14084-14093, 2022. 1, 2", + "[9] Yiqi Gao, Xinglin Hou, Yuanmeng Zhang, Tiezheng Ge, Yunig Jiang, and Peng Wang. Caponimage: Context-driven dense-captioning on image. arXiv preprint arXiv:2204.12974, 2022. 2", + "[10] Georgia Gkioxari, Ross Girshick, Piotr Dólar, and Kaiming He. Detecting and recognizing human-object interactions. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 8359-8367, 2018. 1" + ], + "bbox": [ + 78, + 114, + 470, + 893 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[11] Alex Graves. Long short-term memory. Supervised sequence labelling with recurrent neural networks, pages 37-45, 2012. 3", + "[12] Xiuye Gu, Tsung-Yi Lin, Weicheng Kuo, and Yin Cui. Open-vocabulary object detection via vision and language knowledge distillation. arXiv preprint arXiv:2104.13921, 2021. 1, 2", + "[13] Agrim Gupta, Piotr Dollar, and Ross Girshick. Lvis: A dataset for large vocabulary instance segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5356-5364, 2019. 5, 6", + "[14] Akshita Gupta, Sanath Narayan, KJ Joseph, Salman Khan, Fahad Shahbaz Khan, and Mubarak Shah. Ow-detr: Open-world detection transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9235-9244, 2022. 1, 2", + "[15] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In Proceedings of the IEEE international conference on computer vision, pages 2961-2969, 2017. 3, 5", + "[16] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In International Conference on Machine Learning, pages 4904-4916. PMLR, 2021. 2", + "[17] Justin Johnson, Andrej Karpathy, and Li Fei-Fei. Densecap: Fully convolutional localization networks for dense captioning. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4565–4574, 2016. 2, 3, 6, 7, 8", + "[18] KJ Joseph, Salman Khan, Fahad Shahbaz Khan, and Vineeth N Balasubramanian. Towards open world object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5830-5840, 2021. 1, 2", + "[19] Aishwarya Kamath, Mannat Singh, Yann LeCun, Gabriel Synnaeve, Ishan Misra, and Nicolas Carion. Mdetr-modulated detection for end-to-end multi-modal understanding. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1780-1790, 2021. 5, 6", + "[20] Wonjae Kim, Bokyung Son, and Ildoo Kim. Vilt: Vision-and-language transformer without convolution or region supervision. In International Conference on Machine Learning, pages 5583-5594. PMLR, 2021. 2", + "[21] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014.5", + "[22] Ranjay Krishna, Yuke Zhu, Oliver Groth, Justin Johnson, Kenji Hata, Joshua Kravitz, Stephanie Chen, Yannis Kalantidis, Li-Jia Li, David A Shamma, et al. Visual genome: Connecting language and vision using crowdsourced dense image annotations. International journal of computer vision, 123(1):32-73, 2017. 6", + "[23] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. arXiv preprint arXiv:2201.12086, 2022. 2" + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "15241", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[24] Liunian Harold Li, Mark Yatskar, Da Yin, Cho-Jui Hsieh, and Kai-Wei Chang. Visualbert: A simple and perform-. mant baseline for vision and language. arXiv preprint arXiv:1908.03557, 2019. 2", + "[25] Liunian Harold Li, Pengchuan Zhang, Haotian Zhang, Jianwei Yang, Chunyuan Li, Yiwu Zhong, Lijuan Wang, Lu Yuan, Lei Zhang, Jenq-Neng Hwang, et al. Grounded language-image pre-training. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10965-10975, 2022. 2, 3, 4, 5, 6", + "[26] Xiangyang Li, Shuqiang Jiang, and Jungong Han. Learning object context for dense captioning. In Proceedings of the AAAI conference on artificial intelligence, volume 33, pages 8650-8657, 2019. 3, 7, 8", + "[27] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In European conference on computer vision, pages 740-755. Springer, 2014. 1, 5, 6", + "[28] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 9992-10002, 2021. 5, 6", + "[29] Jingjing Meng, Junsong Yuan, Jiong Yang, Gang Wang, and Yap-Peng Tan. Object instance search in videos via spatiotemporal trajectory discovery. IEEE Transactions on Multimedia, 18(1):116-127, 2015. 1", + "[30] James Philbin, Ondrej Chum, Michael Isard, Josef Sivic, and Andrew Zisserman. Object retrieval with large vocabularies and fast spatial matching. In 2007 IEEE conference on computer vision and pattern recognition, pages 1-8. IEEE, 2007. 1", + "[31] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 2", + "[32] Shafin Rahman, Salman Khan, and Fatih Porikli. Zero-shot object detection: Learning to simultaneously recognize and localize novel concepts. In Asian Conference on Computer Vision, pages 547-563. Springer, 2018. 2", + "[33] Joseph Redmon, Santosh Divvala, Ross Girshick, and Ali Farhadi. You only look once: Unified, real-time object detection. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 779-788, 2016. 1, 3", + "[34] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. Advances in neural information processing systems, 28, 2015. 1, 3", + "[35] Hamid Rezatofighi, Nathan Tsoi, JunYoung Gwak, Amir Sadeghian, Ian Reid, and Silvio Savarese. Generalized intersection over union: A metric and a loss for bounding box regression. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 658-666, 2019. 4" + ], + "bbox": [ + 78, + 90, + 470, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[36] Shuai Shao, Zeming Li, Tianyuan Zhang, Chao Peng, Gang Yu, Xiangyu Zhang, Jing Li, and Jian Sun. Objects365: A large-scale, high-quality dataset for object detection. In Proceedings of the IEEE/CVF international conference on computer vision, pages 8430-8439, 2019. 1, 6", + "[37] Zhuang Shao, Jungong Han, Demetris Marnerides, and Kurt Debattista. Region-object relation-aware dense captioning via transformer. IEEE Transactions on Neural Networks and Learning Systems, 2022. 3, 6, 7, 8", + "[38] Sheng Shen, Chunyuan Li, Xiaowei Hu, Yujia Xie, Jianwei Yang, Pengchuan Zhang, Anna Rohrbach, Zhe Gan, Lijuan Wang, Lu Yuan, et al. K-lite: Learning transferable visual models with external knowledge. arXiv preprint arXiv:2204.09222, 2022. 3, 5, 8", + "[39] Wenhui Wang, Hangbo Bao, Li Dong, and Furu Wei. Vlmo: Unified vision-language pre-training with mixture-of-modality-experts. arXiv preprint arXiv:2111.02358, 2021. 2", + "[40] Linjie Yang, Kevin Tang, Jianchao Yang, and Li-Jia Li. Dense captioning with joint inference and visual context. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2193-2202, 2017. 7, 8", + "[41] Lewei Yao, Jianhua Han, Youpeng Wen, Xiaodan Liang, Dan Xu, Wei Zhang, Zhenguo Li, Chunjing Xu, and Hang Xu. Detclip: Dictionary-enriched visual-concept paralleled pre-training for open-world detection. arXiv preprint arXiv:2209.09407, 2022. 1, 2, 3, 4, 5, 6", + "[42] Lewei Yao, Runhu Huang, Lu Hou, Guansong Lu, Minzhe Niu, Hang Xu, Xiaodan Liang, Zhenguo Li, Xin Jiang, and Chunjing Xu. Filip: Fine-grained interactive language-image pre-training. ArXiv, abs/2111.07783, 2022. 5", + "[43] Guojun Yin, Lu Sheng, Bin Liu, Nenghai Yu, Xiaogang Wang, and Jing Shao. Context and attribute grounded dense captioning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6241-6250, 2019. 3, 7", + "[44] Alireza Zareian, Kevin Dela Rosa, Derek Hao Hu, and Shih-Fu Chang. Open-vocabulary object detection using captions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14393-14402, 2021. 1, 2", + "[45] Haoyang Zhang and Xuming He. Deep free-form deformation network for object-mask registration. In Proceedings of the IEEE International Conference on Computer Vision, pages 4251-4259, 2017. 1", + "[46] Haotian Zhang, Pengchuan Zhang, Xiaowei Hu, Yen-Chun Chen, Liunian Harold Li, Xiyang Dai, Lijuan Wang, Lu Yuan, Jenq-Neng Hwang, and Jianfeng Gao. Glipv2: Unifying localization and vision-language understanding. arXiv preprint arXiv:2206.05836, 2022. 2, 5", + "[47] Shifeng Zhang, Cheng Chi, Yongqiang Yao, Zhen Lei, and Stan Z Li. Bridging the gap between anchor-based and anchor-free detection via adaptive training sample selection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9759-9768, 2020. 3, 4, 5", + "[48] Xiaowei Zhao, Xianglong Liu, Yifan Shen, Yuqing Ma, Yixuan Qiao, and Duorui Wang. Revisiting open world object detection. arXiv preprint arXiv:2201.00471, 2022. 1, 2, 3" + ], + "bbox": [ + 501, + 92, + 892, + 901 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "15242", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[49] Xingyi Zhou, Rohit Girdhar, Armand Joulin, Philipp Krahenbuhl, and Ishan Misra. Detecting twenty-thousand classes using image-level supervision. In European Conference on Computer Vision, pages 350-368. Springer, 2022. 2", + "[50] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable detr: Deformable transformers for end-to-end object detection. arXiv preprint arXiv:2010.04159, 2020. 1" + ], + "bbox": [ + 78, + 90, + 468, + 217 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "15243", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2023/CapDet_ Unifying Dense Captioning and Open-World Detection Pretraining/c3eddfd4-ae1f-40f7-a7cf-96c54f1fea7b_model.json b/2023/CapDet_ Unifying Dense Captioning and Open-World Detection Pretraining/c3eddfd4-ae1f-40f7-a7cf-96c54f1fea7b_model.json new file mode 100644 index 0000000000000000000000000000000000000000..5b1980684a3fdb5b8bd633d2790ea82663bb57cb --- /dev/null +++ b/2023/CapDet_ Unifying Dense Captioning and Open-World Detection Pretraining/c3eddfd4-ae1f-40f7-a7cf-96c54f1fea7b_model.json @@ -0,0 +1,1905 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.808, + 0.045 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.102, + 0.13, + 0.868, + 0.154 + ], + "angle": 0, + "content": "CapDet: Unifying Dense Captioning and Open-World Detection Pretraining" + }, + { + "type": "text", + "bbox": [ + 0.169, + 0.18, + 0.8, + 0.217 + ], + "angle": 0, + "content": "Yanxin Long\\(^{1*}\\) Youpeng Wen\\(^{1*}\\) Jianhua Han\\(^{2}\\) Hang Xu\\(^{2}\\) Pengzhen Ren\\(^{1}\\) Wei Zhang\\(^{2}\\) Shen Zhao\\(^{1\\dagger}\\) Xiaodan Liang\\(^{1,3\\dagger}\\)" + }, + { + "type": "text", + "bbox": [ + 0.154, + 0.218, + 0.813, + 0.285 + ], + "angle": 0, + "content": "1Shenzhen Campus of Sun Yat-sen University 2Huawei Noah's Ark Lab 3MBZUAI longyx9@mail2.sysu.edu.cn, wenyoupeng0@outlook.com, hanjianhua4@huawei.com, chromexbjxh@gmail.com, renpzh@mail.sysu.edu.cn, wz.zhang@huawei.com, z-s-06@163.com, xdliang328@gmail.com" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.322, + 0.313, + 0.337 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.348, + 0.474, + 0.695 + ], + "angle": 0, + "content": "Benefiting from large-scale vision-language pre-training on image-text pairs, open-world detection methods have shown superior generalization ability under the zero-shot or few-shot detection settings. However, a pre-defined category space is still required during the inference stage of existing methods and only the objects belonging to that space will be predicted. To introduce a \"real\" open-world detector, in this paper, we propose a novel method named CapDet to either predict under a given category list or directly generate the category of predicted bounding boxes. Specifically, we unify the open-world detection and dense caption tasks into a single yet effective framework by introducing an additional dense captioning head to generate the region-grounded captions. Besides, adding the captioning task will in turn benefit the generalization of detection performance since the captioning dataset covers more concepts. Experiment results show that by unifying the dense caption task, our CapDet has obtained significant performance improvements (e.g., \\(+2.1\\%\\) mAP on LVIS rare classes) over the baseline method on LVIS (1203 classes). Besides, our CapDet also achieves state-of-the-art performance on dense captioning tasks, e.g., \\(15.44\\%\\) mAP on VG V1.2 and \\(13.98\\%\\) on the VG-COCO dataset." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.703, + 0.208, + 0.718 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.729, + 0.47, + 0.864 + ], + "angle": 0, + "content": "Most state-of-the-art object detection methods [33, 34, 50] benefit from a large number of densely annotated detection datasets (e.g., COCO [27], Object365 [36], LVIS [12]). However, this closed-world setting results in the model only being able to predict categories that appear in the training set. Considering the ubiquity of new concepts in real-world scenes, it is very challenging to locate and identify these new visual concepts. This predictive ability of new concepts in open-world scenarios has very important research value" + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.32, + 0.889, + 0.459 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.506, + 0.462, + 0.889, + 0.613 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.614, + 0.892, + 0.739 + ], + "angle": 0, + "content": "Figure 1. Comparison of the different model predictions under OWD, OVD, and our setting. (a) OWD methods [14, 18, 48] are not able to describe the detailed category of the detected unknown objects and (b) the performance of OVD methods [8, 12, 41] usually depends on the pre-defined category list during the inference. (c) With the unification of two pipelines of dense captioning and open-world detection pre-training, our CapDet can either predict under a given category list or directly generate the description of predicted bounding boxes." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.745, + 0.891, + 0.791 + ], + "angle": 0, + "content": "in real-world applications such as object search [29, 30], instance registration [45], and human-object interaction modeling [10]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.795, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Currently, the open world scenario mainly includes two tasks: open world object detection [18] (OWD) and open-vocabulary object detection [44] (OVD). Although the paradigms of OWD and OVD tasks are closer to the real world, the former cannot describe the specific concept of the detected unknown objects and requires a pre-defined category list during the inference. Specifically, as shown" + }, + { + "type": "page_footnote", + "bbox": [ + 0.1, + 0.875, + 0.235, + 0.901 + ], + "angle": 0, + "content": "*Equal contribution. † Corresponding authors." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "15233" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.091, + 0.473, + 0.364 + ], + "angle": 0, + "content": "in Figure 1, previous OWD methods [14,18,48] would recognize new concepts not in the predefined category space as \"unknown\". Further, another line of task OVD requires the model to learn a limited base class and generalize to novel classes. Compared to the zero-shot object detection (ZSD) proposed by [32], OVD allows the model to use external knowledge, e.g., knowledge distillation from a largescale vision-language pre-trained model [8, 12], image-caption pairs [44], image classification data [49], grounding data [25, 41, 46]. With the external knowledge, OVD methods show a superior generalization capacity to detect the novel classes within a given category space. However, as shown in Figure 1, when given an incomplete category list, OVD can only predict the concepts that appear in the given category list, otherwise, there will be recognition errors, (i.e., as illustrated in Figure 1 (b), the OVD methods prone to predict the \"wall socket\" as \"remote\", since the latter is in the category list but not the former)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.366, + 0.473, + 0.533 + ], + "angle": 0, + "content": "Thus, under the OVD setting, we mainly face the following two challenges: (i) it is difficult to define a complete list of categories; (ii) low response values on rare categories often lead to recognition errors. This is mainly because we cannot exhaustively enumerate new objects in the real world, and secondly, it is difficult to collect enough samples for rare classes. However, the fact that rare objects in the real world, even some new objects that are unknown to humans, such as UFOs, do not prevent people from using natural language to describe it as \"a flying vehicle that looks like a Frisbee\"." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.535, + 0.473, + 0.779 + ], + "angle": 0, + "content": "Therefore, based on the above observations, in this paper, we consider a new setting that is closer to the open world and real scenes, i.e., we expect the model to both detect and recognize concepts in a given category list, and to generate corresponding natural language descriptions for new concepts or rare categories of objects. Early dense captioning methods [9, 17] can locate salient regions in images and generate the region-grounded captions with natural language. Inspired by this, to address the challenges faced in the OVD setting, we propose to unify the two pipelines of dense captioning and open-world detection pre-training into one training framework, called CapDet. It empowers the model with the ability to both accurately detect and recognize common object categories and generate dense captions for unknown and rare categories by unifying the two training tasks." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.78, + 0.473, + 0.904 + ], + "angle": 0, + "content": "Specifically, our CapDet constructs a unified data format for the dense captioning data and detection data. With the data unification, CapDet further adopts a unified pretraining paradigm including open-world object detection and dense captioning pre-training. For open-world detection pretraining, we treat the detection task as a semantic alignment task and adopt a dual encoder structure as [41] to locate and predict the given concepts list. The concepts" + }, + { + "type": "text", + "bbox": [ + 0.497, + 0.092, + 0.895, + 0.212 + ], + "angle": 0, + "content": "list contains category names in detection data and region-grounded captions in dense captioning data. For dense captioning pretraining, CapDet proposes a dense captioning head to take the predicted proposals as input to generate the region-grounded captions. Due to the rich visual concepts in the dense captioning data, the integration of dense captioning tasks will in turn benefit the generalization of detection performance." + }, + { + "type": "text", + "bbox": [ + 0.497, + 0.212, + 0.895, + 0.349 + ], + "angle": 0, + "content": "Our experiments show that the integration of few dense captioning data brings in large improvement in the object detection datasets LVIS, e.g., \\(+2.7\\%\\) mAP on LVIS. The unification of dense captioning and detection pre-training gains an additional \\(2.3\\%\\) increment on LVIS and \\(2.1\\%\\) increment on LVIS rare classes. Besides, our model also achieves state-of-the-art performance on dense captioning tasks. Note that our method is the first to unify dense captioning and open-world detection pretraining." + }, + { + "type": "text", + "bbox": [ + 0.52, + 0.349, + 0.839, + 0.363 + ], + "angle": 0, + "content": "To summarize, our contributions are three folds:" + }, + { + "type": "text", + "bbox": [ + 0.52, + 0.37, + 0.892, + 0.446 + ], + "angle": 0, + "content": "- We propose a novel open-vocabulary object detection framework CapDet, which cannot only detect and recognize concepts in a given category list but also generate corresponding natural language descriptions for new concept objects." + }, + { + "type": "text", + "bbox": [ + 0.52, + 0.449, + 0.892, + 0.507 + ], + "angle": 0, + "content": "- We propose to unify the two pipelines of dense captioning and open-world detection pre-training into one training framework. Both two pre-training tasks are beneficial to each other." + }, + { + "type": "text", + "bbox": [ + 0.52, + 0.512, + 0.892, + 0.631 + ], + "angle": 0, + "content": "- Experiments show that by unified dense captioning task and detection task, our CapDet gains significant performance improvements on the open-vocabulary object detection task (e.g., \\(+3.3\\%\\) mAP on LVIS rare classes). Furthermore, our CapDet also achieves state-of-the-art performance on the dense captioning tasks, e.g., \\(15.44\\%\\) mAP on Visual Genome (VG) V1.2 and \\(13.98\\%\\) mAP on VG-COCO." + }, + { + "type": "list", + "bbox": [ + 0.52, + 0.37, + 0.892, + 0.631 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.634, + 0.642, + 0.65 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.497, + 0.659, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Vision-Language Pre-training. Vision-Language Pretraining [6, 16, 31] as a scheme in the domains of natural language processing [1, 5] and computer vision [7] obtains continual attention currently. And it exhibits strong performance and generalization ability on various downstream vision and cross-modal tasks. Among them, CLIP [31] and ALIGN [16] as dual-stream methods utilize large-scale image-text pairs on the Internet by cross-modal contrastive learning to get excellent zero-shot classification ability. Single-stream methods [20, 24] unify visual and textual embeddings in a single transformer-based model, which can perform text generation tasks such as image caption and VQA. Some mixed architectures [23, 39] combine single-stream and dual-stream to explore a unified way of vision-language understanding and generation. However, these methods take low-resolution images as input and serve the" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "15234" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.471, + 0.138 + ], + "angle": 0, + "content": "task of classification and retrieval. Those vision-language pre-training approaches can not be applied to pure computer vision task directly, i.e., object detection task." + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.144, + 0.47, + 0.401 + ], + "angle": 0, + "content": "Open World Object Detection / Open-Vocabulary Object Detection. Object detection is a core computer vision task, which aims at localizing objects using a bounding box and classifying them. The mature detection approaches which show great performance on supervised data include one-stage detectors (i.e.,YOLO [33],ATSS [47]) having a relatively high detection efficiency and two-stage detectors (i.e., Faster R-CNN [34],Mask R-CNN [15]) having good detection accuracy. However, how to generalize these methods to rare classes and novel concepts in the real world is a challenge. Currently, several object detection approaches for such open-world scenes have attracted extensive attention from academia and industry. These methods are divided into two tasks which are called open-world object detection and open-vocabulary object detection respectively depending on whether to detect the class of unknown classes." + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.402, + 0.471, + 0.66 + ], + "angle": 0, + "content": "For the OWD task, Zhao et al. [48] proposed a proposal advisor to assist in identifying unknown proposals without supervision and a class-specific expelling classifier to filter out confusing predictions. For the OVD task, GLIP [25] converts the detection data into grounding format and proposes a fusion module to learn semantic vision information in grounding data. K-Lite [38] reconstructs the input format of the data in GLIP from sequential to parallel and uses nouns hierarchy and definition to format text sequence. Det-CLIP [41] unifies detection, grounding, and image-text pair data in a paralleled formulation and constructs a concept dictionary to augment the text data, which strikes a balance between performance and efficiency. Differing from all these works, our CapDet can generate an open-set caption of each region proposal to cover situations where the semantics of new object instances are not in the given category list." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.675, + 0.47, + 0.902 + ], + "angle": 0, + "content": "Dense Captioning. Dense captioning aims at generating detailed descriptions for local regions, which usually needs to locate visual regions with semantic information and generate captions for these regions. J. Johnson et al. [17] utilized a fully convolutional localization network to locate regions of interest (RoIs) and then describe them. Afterward, many methods [26, 43] based on Faster-RCNN [34] and LSTM [11] are proposed to do dense captioning. X. Li et al. [26] arrange RoI features as a sequence and put them into LSTM with the guidance of the region features to form the complementary object context features. This method also needs ground truth bounding boxes auxiliary tests to achieve good results. But limited by the forget gate mechanism of LSTM, the inputted sequence cannot be too long. Then, the transformer-based method TDC [37] is proposed" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.152 + ], + "angle": 0, + "content": "to tackle the long sequence forgotten problem. Instead, our CapDet proposes a transformer-based caption head to generate a caption using a single-stage detector ATSS while simultaneously achieving open-world detection." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.167, + 0.593, + 0.183 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.193, + 0.892, + 0.329 + ], + "angle": 0, + "content": "The overview of our proposed CapDet is shown in Figure 2. To construct a detector to either predict under a given category list or directly generate the concepts of predicted bounding boxes, we incorporate detection data and dense caption data together. In this section, we will present a unified data format for the detection data and dense caption data in Section 3.1, the model architecture and pre-training objectives for open-world object detection pre-training in Section 3.2 and dense captioning in Section 3.3." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.339, + 0.694, + 0.354 + ], + "angle": 0, + "content": "3.1. Unified Formulation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.363, + 0.892, + 0.572 + ], + "angle": 0, + "content": "We defined a unified triplet-wise data format \\((x, \\{\\mathbf{b}_i\\}_{i=1}^N, y_{i=1}^N)\\) for each sample from different sources. Specifically, \\(x \\in \\mathbb{R}^{3 \\times h \\times w}\\) is the input image, \\(\\{\\mathbf{b}_i | \\mathbf{b}_i \\in \\mathbb{R}^4\\}_{i=1}^N\\) denotes the bounding boxes coordinates for each region of the image, and the \\(y_{i=1}^N\\) represents the concepts of the corresponding boxes. \\(N\\) denotes the number of regions. A concept \\(y_i\\) formatted as a sentence contains the category and textual description of the corresponding region. In detection data, a concept \\(y\\) consists of the category name and the corresponding definition from the concept dictionary [41], while \\(y_i\\) represents the region-grounded caption in dense caption data. For example, for an image \\(x\\) in detection data, \\(y_i\\) can be:" + }, + { + "type": "equation", + "bbox": [ + 0.581, + 0.575, + 0.808, + 0.591 + ], + "angle": 0, + "content": "\\[\ny _ {i} = \\text {\" p e r s o n , a h u m a n b e i n g . \"}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.599, + 0.857, + 0.615 + ], + "angle": 0, + "content": "For an image \\( x \\) from dense captioning data, \\( y_{i} \\) can be:" + }, + { + "type": "equation", + "bbox": [ + 0.591, + 0.627, + 0.8, + 0.643 + ], + "angle": 0, + "content": "\\[\ny _ {i} = \\text {\" a n o u t l e t o n t h e w a l l .}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.655, + 0.892, + 0.732 + ], + "angle": 0, + "content": "With the triplet, we can learn a unified image-text alignment objective on the detection data and the dense captioning data. The unified formulation also ensures the joint training of open-world object detection pre-training and dense captioning." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.741, + 0.866, + 0.759 + ], + "angle": 0, + "content": "3.2. Open-World Object Detection Pre-training" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.892, + 0.903 + ], + "angle": 0, + "content": "Based on the unified formulation of detection data and dense captioning data, we regard the captions of regions in dense captioning data as a kind of category and utilize two different sources of data for the open-world object detection pre-training. Compared with the limited class list of detection data, dense caption data contains richer concepts and more semantic information than class names of individual regions. On the other hand, localization and recognition are two essential tasks of object detection. Traditional object" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15235" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.14, + 0.091, + 0.833, + 0.433 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.442, + 0.895, + 0.527 + ], + "angle": 0, + "content": "Figure 2. The overall architecture of CapDet. The training paradigm of CapDet contains open-world object detection pre-training and dense captioning. In detection, CapDet contains a dual vision-language encoder. The image encoder generates region embeddings from detection and dense captioning data. The regression loss and centerness loss are introduced to regress the locations. The text encoder takes the category concepts as input to generate the embeddings from the [EOS] token. Then we treat the detection task as a matching task and adopt an alignment loss for the category embeddings and region embeddings. In dense captioning, an additional dense captioning head is proposed to take the region embeddings as input and generate the textual captions for corresponding regions with natural language." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.538, + 0.471, + 0.627 + ], + "angle": 0, + "content": "detection always focuses on the salient objects in the image. While the dense captioning data contains lots of annotations which are just parts of an object, e.g., an ear of an elephant, it is not suitable to learn those annotations for the localization task. Therefore, we only calculate the localization loss on detection data." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.629, + 0.47, + 0.781 + ], + "angle": 0, + "content": "As shown in Figure 2, CapDet predicts the regions and treats the recognition task as a region-category matching task. For efficient learning on the matching task, we adopt the negative sampling proposed by [41] to provide negative concepts to enlarge the concept space in a batch. Specifically, for each iteration, we randomly sample a negative concept set and add to the positive concept set (N samples) in a batch to obtain the final concept set \\( y_{i = 1}^{M} \\), where M represents the sum of the number of positive and negative samples. Finally, we format the triplet to \\( (x,\\{\\mathbf{b}_i\\}_{i = 1}^N,y_{i = 1}^M) \\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.781, + 0.471, + 0.902 + ], + "angle": 0, + "content": "CapDet contains a dual vision-language encoder and takes the triplet \\((x, \\{\\mathbf{b}_i\\}_{i=1}^N, y_{i=1}^M)\\) as input. The image encoder \\(\\Phi_v\\) is an object detector that can predict the bounding boxes of regions from the input image \\(x\\) and output the region features \\(O \\in \\mathbb{R}^{K \\times D}\\). The text encoder \\(\\Phi_l\\) takes the concept set \\(y_{i=1}^M\\) as input and obtains the text embeddings \\(W \\in \\mathbb{R}^{M \\times D}\\) from the special token [EOS] concatenated with the text input. \\(K, D\\) denotes the number of predicted" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.538, + 0.892, + 0.569 + ], + "angle": 0, + "content": "regions and region feature dimensions. The alignment score matrix \\(S \\in \\mathbb{R}^{K \\times M}\\) of regions and texts is calculated by:" + }, + { + "type": "equation", + "bbox": [ + 0.561, + 0.582, + 0.892, + 0.601 + ], + "angle": 0, + "content": "\\[\nO = \\Phi_ {v} (x), W = \\Phi_ {l} \\left(y _ {i = 1} ^ {M}\\right), S = O W ^ {T} \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.616, + 0.893, + 0.782 + ], + "angle": 0, + "content": "where \\(T\\) denotes the transpose operation. A ground-truth alignment matrix \\(G \\in \\{0,1\\}^{K \\times M}\\) is constructed to indicate the matching relation of regions and concepts. The alignment loss \\(\\mathcal{L}_{\\text{align}}\\) is calculated by the predicted alignment scores of regions \\(S\\) and the ground-truth alignment matrix \\(G\\). Following [25, 41], we adopt the ATSS [47] detector as an image encoder, and \\(\\mathcal{L}_{\\text{align}}\\) is typically a sigmoid focal loss. As a one-stage detector, the localization loss contains centeredness loss \\(\\mathcal{L}_{\\text{cen}}\\) and bounding box regression loss \\(\\mathcal{L}_{\\text{reg}}\\). The training objective of detection pre-training can be written as:" + }, + { + "type": "equation", + "bbox": [ + 0.504, + 0.795, + 0.892, + 0.847 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\left\\{ \\begin{array}{l l} \\mathcal {L} _ {\\text {a l i g n}} + \\alpha \\mathcal {L} _ {\\text {r e g}} + \\beta \\mathcal {L} _ {\\text {c e n t e r}}, & \\text {f o r d e t e c t i o n} \\\\ \\mathcal {L} _ {\\text {a l i g n}}, & \\text {f o r d e n s e c a p t i o n i n g} \\end{array} \\right. \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.849, + 0.892, + 0.895 + ], + "angle": 0, + "content": "where \\(\\alpha\\) and \\(\\beta\\) denote the weights for the centerness loss \\(\\mathcal{L}_{cen}\\) and box regression loss \\(\\mathcal{L}_{reg}\\), respectively. The \\(\\mathcal{L}_{cen}\\) is the sigmoid loss and the \\(\\mathcal{L}_{reg}\\) is the GIOU loss [35]." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "15236" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.078, + 0.09, + 0.9, + 0.367 + ], + "angle": 0, + "content": "
MODELBACKBONEPRE-TRAIN DATAIMAGES NUMBERAPLVIS APr / APc / APf
MASK-RCNN [15]SWIN-TLVIS0.1M34.119.1 / 34.0 / 37.0
ATSS [47]SWIN-TLVIS0.1M33.619.7 / 32.4 / 37.2
ATSS [47]SWIN-LLVIS0.1M43.930.6 / 43.7 / 46.3
MDETR [19]RN101GOLDG+0.77M24.220.9 / 24.3 / 24.2
GLIP-T(A) [25]SWIN-T+DH+FO3650.66M18.514.2 / 13.9 / 23.4
GLIP-T(C) [25]SWIN-T+DH+FO365,GOLDG1.43M24.917.7 / 19.5 / 31.0
GLIP-T [25]SWIN-T+DH+FO365,GOLDG,CAP4M5.43M26.020.8 / 21.4 / 31.0
K-LITE [38]SWIN-TO3650.66M21.314.8 / 18.6 / 24.8
K-LITE [38]SWIN-TO365,GOLDG1.43M26.117.2 / 24.6 / 29.0
GLIPv2-T [46]SWIN-T+DH+FO365,GOLDG,CAP4M5.43M29.0- / - / -
DETCLIP-T(A) [41]SWIN-TO3650.66M28.826.0 / 28.0 / 30.0
DETCLIP-T(B) [41]SWIN-TO365, GOLDG1.43M34.426.9 / 33.9 / 36.3
DETCLIP-T(C)* [41]SWIN-TO365, VG0.73M31.527.5 / 30.6 / 33.0
CAPDET (OURS)SWIN-TO365, VG0.73M33.829.6 / 32.8 / 35.5
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.378, + 0.893, + 0.421 + ], + "angle": 0, + "content": "Table 1. Zero-shot performance on LVIS [13] MiniVal5k datasets. \\(\\mathrm{AP}_r / \\mathrm{AP}_c / \\mathrm{AP}_f\\) indicate the AP values for rare, common, and frequent categories, respectively. \"DH\" and \"F\" in GLIP [25] baselines stand for the dynamic head [3] and cross-modal fusion, respectively. Baselines with * are implemented with our code base. GoldG+ denotes the GoldG plus the COCO [27] caption dataset." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.459, + 0.251, + 0.475 + ], + "angle": 0, + "content": "3.3. Dense Captioning" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.482, + 0.47, + 0.693 + ], + "angle": 0, + "content": "The open-world object detection pre-training ensures CapDet gains the capacity to detect under given an arbitrary category list. However, when the given category list is not complete enough to cover the potential classes on a new domain data, the detector will perform worse on the categories which are not in the given list. Considering such limitation, we propose a dense captioning head \\(\\Phi_C\\) to generate semantically rich concepts with natural language for the predicted proposals. In the dense captioning task, the model receives an image and produces a set of regions and the corresponding captions. The dense captioning head is a cross-modal decoder that takes the c predicted regions features \\(O\\) generated by the image encoder as input. The captioning (i.e., language modeling) loss is calculated by:" + }, + { + "type": "equation", + "bbox": [ + 0.149, + 0.703, + 0.469, + 0.721 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {c a p} = - \\log p \\left(y _ {i t} \\mid \\Phi_ {c} \\left(y _ {i (\\tau < t)}, O _ {i}\\right)\\right), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.73, + 0.47, + 0.775 + ], + "angle": 0, + "content": "where \\( y_{it} \\) means the \\( t \\) token in caption \\( y_{i} \\) corresponding to region feature \\( O_{i} \\), and \\( y_{i(\\tau < t)} \\) means tokens before \\( t \\) in caption \\( y_{i} \\). The overall pre-training loss can be written as:" + }, + { + "type": "equation", + "bbox": [ + 0.192, + 0.785, + 0.469, + 0.802 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = w _ {d} \\mathcal {L} _ {\\text {d e t}} + w _ {c} \\mathcal {L} _ {\\text {c a p}}, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.811, + 0.468, + 0.826 + ], + "angle": 0, + "content": "where \\(w_{d}, w_{c}\\) denote the weighting factor of \\(\\mathcal{L}_{det}\\) and \\(\\mathcal{L}_{cap}\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.826, + 0.469, + 0.902 + ], + "angle": 0, + "content": "To minimize the gap in the type of bounding boxes between the detection data and dense captioning data, we propose a simple way to transform our detector as a class-agnostic detector and only select the top \\(k\\) regions based on the centeredness scores to adapt to the dense captioning" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.459, + 0.892, + 0.596 + ], + "angle": 0, + "content": "task. We can fine-tune our CapDet on the dense captioning data to achieve better performance. Specifically, we propose \"object\" as the foreground concept and \"background\" as the background concept. The text encoder \\(\\Phi_l\\) outputs the concept embeddings \\(W' \\in \\mathbb{R}^{2 \\times D}\\). Then the alignment scores \\(S' \\in \\mathbb{R}^{K \\times 2}\\) is calculated by Eqn. 1. The captioning head takes the top \\(k\\) most confident proposal embeddings based on centeredness scores as input to predict the region-grounded captions." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.616, + 0.625, + 0.634 + ], + "angle": 0, + "content": "4. Experiment" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.644, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Implementation Details. For the image encoder, we adopt the Swin-T backbone proposed in SwinTransformer [28] which is pre-trained on ImageNet-1K [4]. We use 12 layers 8 heads transformer as our text encoder and load a base model checkpoint released by FILIP [42], in order to make a fair comparison with DetCLIP [41]. The structure of the dense captioning head is consistent with that in the text encoder but trained from scratch for a fair comparison. We employ AdamW [21] optimizer and set the batch size to 32. The learning rate is set to \\(1.4 \\times 10^{-4}\\) for the parameters of the image encoder and detection head, and \\(1.4 \\times 10^{-5}\\) for the text encoder and dense captioning head. When fine-tuning the VG dataset to do the dense captioning task, we set the learning rate to \\(1.4 \\times 10^{-4}\\). Without otherwise specified, all models are trained with 12 epochs and the learning rate is decayed with a factor of 0.1 at the 8-th and the 11-th epoch. The context token" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15237" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.091, + 0.891, + 0.321 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.345, + 0.893, + 0.375 + ], + "angle": 0, + "content": "Figure 3. Qualitative visualizations between GLIP-T, DetCLIP-T(C) and CapDet. From top to down, the three rows of images show the LVIS zero-shot detection results of GLIP-T, DetCLIP-T(C), and CapDet respectively. All models are pre-trained on O365 and VG." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.4, + 0.471, + 0.476 + ], + "angle": 0, + "content": "length for input text is set to 20. We set the number of input captions to 150, and the number of the region features N is determined by the feature map. The loss weight factor \\( w_{c} \\) and \\( w_{d} \\) are both set to 1.0. We build our model on MMDetection [2] code base." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.499, + 0.47, + 0.802 + ], + "angle": 0, + "content": "Dataset. Our CapDet is trained with two types of data, including detection data and caption data. Following DetCLIP [41], we use Object365 [36] (it will be abbreviated as O365 in the following paper) as detection data, and sample 0.66M data from O365 v2 for training. Following GLIP [25] and DetCLIP [41], LVIS [13] MiniVal5k (defined in [19]) which has 5000 images is used for detection evaluation. Moreover, we remove the training samples contained in the LVIS dataset for fair zero-transfer evaluation. For dense captioning data, we mainly conduct our experiments on VG [22] V1.2 and VG-COCO (defined in [37]). Following [37], we allocate 77398 images for training and 5000 images for validation and testing on VG. As demonstrated in [22], the ground-truth bounding boxes of VG are much denser than the other object detection datasets, i.e., the average number of per sample in MS COCO [27] is only 7.1 vs. 35.4 in VG. Then an intersection of VG V1.2 and MS COCO is proposed by [37] and is denoted as VG-COCO, which has 38080 images for training, 2489 for validation, and 2476 for testing." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.47, + 0.903 + ], + "angle": 0, + "content": "Benchmark Settings. We mainly evaluate our method on open-vocabulary object detection and dense captioning task. For open-vocabulary object detection, we evaluate the direct domain transfer on LVIS [13] which contains 1203 categories. Following [25, 41], we metric the zero-shot" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.4, + 0.895, + 0.628 + ], + "angle": 0, + "content": "detection performance by the Average Precision (AP) on a 5k subset. The annotations of LVIS data are split into three folds, i.e., rare, common, and frequency, based on the number of categories. Since there is almost no overlap between the rare classes and the classes of training dataset Objects365 [36], the AP of the rare classes shows a valuable zero-shot detection performance. For the dense captioning task, we follow the setting of [17] to evaluate the VG and VG-COCO. The evaluation metric we adopt is the mean Average Precision proposed by [17] which is calculated across a range of thresholds for both localization and language accuracy, i.e., the intersection over union (IOU) thresholds .3, .4, .5, .6, .7 are used for localization and the METEOR score' thresholds 0, .05, .1, .15, .2, .25 is adopted for evaluating the language generation." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.634, + 0.768, + 0.649 + ], + "angle": 0, + "content": "4.1. Open-world Detection Results" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.657, + 0.892, + 0.853 + ], + "angle": 0, + "content": "Table 1 shows our zero-shot object detection performance on LVIS. We mainly train our CapDet with the backbone Swin-T [28] on the detection data Objects365 [36] and dense captioning data (VG [22]). Since DetCLIP does not report the performance on O365 and VG, we train DetCLIP on the two datasets under the same settings and denote it as DetCLIP-T(C) for a fair comparison. Comparing the 11th row and 12th row, our CapDet outperforms DetCLIP-T(C) on the same data scale and backbone with an extra simple caption head. Moreover, our model's zero-shot performance even surpasses the fully-supervised model with the same backbone by a large margin on rare classes, i.e., CapDet outperforms ATSS by \\(9.9\\%\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.895, + 0.901 + ], + "angle": 0, + "content": "Qualitative Visualizations Figure. 3 illustrates the detection results on LVIS [22] dataset from GLIP-T, DetCLIP-T(C), and CapDet. All three models are trained on O365" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15238" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.088, + 0.092, + 0.885, + 0.459 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.472, + 0.893, + 0.5 + ], + "angle": 0, + "content": "Figure 4. Qualitative visualizations between JIVC and CapDet. \"w/o ft\" means do caption without finetune, while \"w/ ft\" means with finetune." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.506, + 0.471, + 0.596 + ], + "angle": 0, + "content": "and VG, and details are given in Section 4.3. Given a category list, the rare classes are detected more precisely by our CapDet, e.g., \"kitchen table\" in the first column, \"horse buggy\" in the third column, and \"fishbowl\" in the sixth column that our model CapDet detects correctly but the other two not." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.607, + 0.312, + 0.623 + ], + "angle": 0, + "content": "4.2. Dense Captioning Results" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.629, + 0.47, + 0.825 + ], + "angle": 0, + "content": "Due to the target bounding boxes in dense captioning data containing lots of local structures of objects and being much denser than the bounding boxes in object detection data, we do not regress the bounding box in the pre-training stage. The previous works directly train on the dense captioning data and generate captions on the top \\(k\\) proposals ranking by a confidence score. When fine-tuning our model on the VG dataset for the dense captioning tasks, we transform our CapDet into a class-agnostic detector. Specifically, we propose \"object\" as the foreground concept and \"background\" as the background concept for computing alignment scores. The scores are used as proposal confidences to predict the region-grounded captions." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.826, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Table 2 and Table 3 show CapDet significantly outperforms the latest work TDC [37] by \\(2.5\\%\\) on mAP on VG and TDC+ROCSU [37] by \\(2.08\\%\\), respectively. It is worth noticing that, even against given the ground-truth bounding boxes with the previous method COCG [26] denoted" + }, + { + "type": "table", + "bbox": [ + 0.598, + 0.504, + 0.798, + 0.663 + ], + "angle": 0, + "content": "
MethodmAP(%)
FCLN [17]5.16
JIVC [40]9.96
ImgG [26]9.68
COCD [26]9.75
COCG [26]10.39
CAG-Net [43]10.51
TDC [37]11.90
CapDet (Ours)15.44
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.673, + 0.892, + 0.701 + ], + "angle": 0, + "content": "Table 2. Comparison of mAP (%) performance on dense captioning benchmark on the VG V1.2 dataset." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.718, + 0.892, + 0.779 + ], + "angle": 0, + "content": "as COCOG>, our CapDet still gains a \\(43.80\\%\\) mAP increase and achieves state-of-the-art. One important reason is that the excellent detection performance of our model assists the localization ability of dense captioning tasks." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.78, + 0.893, + 0.901 + ], + "angle": 0, + "content": "Qualitative Visualizations. Figure 4 shows a qualitative visualization comparison between JIVC [40] and our CapDet. The three image rows from top to bottom are the visualization of JIVC, CapDet without fine-tuning, and CapDet with finetuning. In the second row, CapDet can locate more objects than JIVC, owing to our model's superior localization performance. After finetuning, CapDet can further describe a region rather than a single object such as" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15239" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.159, + 0.089, + 0.391, + 0.277 + ], + "angle": 0, + "content": "
MethodmAP(%)
FCLN [17]4.23
JIVC [40]7.85
Max Pooling [26]7.86
ImgG [26]7.81
COCD [26]7.92
COCG [26]8.90
COCG-LocSiz [26]8.76
COCG&GT [26]9.79
TDC+ROCSU [37]11.9
CapDet (Ours)13.98
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.289, + 0.471, + 0.318 + ], + "angle": 0, + "content": "Table 3. Comparison of mAP (%) performance on the dense captioning benchmark on the VG-COCO Dataset." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.336, + 0.466, + 0.435 + ], + "angle": 0, + "content": "
MODELDC HEADLVIS
APAPr / APc / APf
GLIP-T×30.422.5 / 29.0 / 33.0
GLIP-T✓33.127.0 / 32.1 / 35.0
DETCLIP-T×31.527.5 / 30.6 / 33.0
DETCLIP-T✓33.829.6 / 32.8 / 35.5
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.446, + 0.469, + 0.473 + ], + "angle": 0, + "content": "Table 4. Ablations on integrating our dense captioning head into different baselines." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.479, + 0.466, + 0.557 + ], + "angle": 0, + "content": "
Pre-training DataFine-tuneDCap mAP(%)Box mAP(%)
VG✘12.8627.65
O365,VG✘4.729.65
VG✓13.8328.58
O365,VG✓15.4430.61
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.568, + 0.468, + 0.596 + ], + "angle": 0, + "content": "Table 5. Ablations on incorporating data from different sources. \"DCap\" stands for the dense caption mAP." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.627, + 0.381, + 0.641 + ], + "angle": 0, + "content": "\"two women in a kitchen\" in the 5-th column." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.654, + 0.241, + 0.669 + ], + "angle": 0, + "content": "4.3. Ablation Studies" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.678, + 0.368, + 0.695 + ], + "angle": 0, + "content": "4.3.1 Ablations for Unified Pre-training" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.703, + 0.469, + 0.839 + ], + "angle": 0, + "content": "Effect on different baselines. Table 4 investigates the advantages of dense captioning heads on different baselines. We integrate our dense captioning head with GLIP-T or DetCLIP-T. The GLIP-T is implemented with parallel text encoding without external knowledge following the setting as ablations in [38] on our code base. All the results are pre-trained on Objects365 and VG. The results show that our dense captioning head is able to boost the generalization and model-agnostic." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Effect of dense captioning data. Table 1 shows the efficiency of incorporating dense captioning data. Specifically, only \\(0.07\\mathrm{M}\\) data added, the DetCLIP-T(C) gains \\(+2.7\\%\\) overall AP and \\(+1.5\\%\\) \\(AP_{r}\\) on LVIS compared to DetCLIP-" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.892, + 0.136 + ], + "angle": 0, + "content": "T(A). The performance of DetCLP-T(A) on rare categories also outperforms DetCLIP-T(C) train on Objects365 and GOLDG, while the data size is 1.43M vs. 0.73M." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.145, + 0.768, + 0.16 + ], + "angle": 0, + "content": "4.3.2 Ablations for dense captioning" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.168, + 0.892, + 0.395 + ], + "angle": 0, + "content": "We investigate the impact of training policy and data from different sources on the dense captioning task. As shown in row1 in Table 5, our CapDet still achieves a significant performance which is directly trained on VG outperforms the previous task (i.e., TDC [37] in Table 2). Row2 is our CapDet and is pre-trained on Objects365 and VG, while only the bounding box in the Objects365 is regressed, and then transformed on a dense captioning task. Since the type of bounding boxes in dense captioning is different from the detection data, the result of the direct transforming to dense captioning is worse. However, we've proved that our model still keeps the dense captioning capacity on the salient objects in Figure 4. The results in row3 and row4 indicate that pre-training on the detection data Objects365 is also beneficial to the dense captioning task." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.398, + 0.623, + 0.413 + ], + "angle": 0, + "content": "5. Limitations" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.422, + 0.892, + 0.544 + ], + "angle": 0, + "content": "These are a few issues that we need to improve in the future: (1) Although our unification training paradigm works well on open-vocabulary object detection and dense captioning task, the training of dense captioning generation costs lots of time. (2) In addition, existing dense captioning data is high-cost to collect. We will research how to collect large-scale dense captioning data by auto annotation and get better performance with the scaled-up data." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.55, + 0.619, + 0.565 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.575, + 0.892, + 0.802 + ], + "angle": 0, + "content": "In this paper, we propose a novel open-world object detection method named CapDet. Our CapDet is more practical in the open world and real scenes. Specifically, CapDet introduces a unification training framework including open-world object detection pre-training and dense captioning. The unification enables our CapDet to localize and recognize concepts in an arbitrary given category list or directly generate textual captions for predicted new concept objects. Experiments show that the design of unification is both beneficial to open-world object detection tasks and dense captioning tasks. In the future, our CapDet can be easily injected into any open world and real scenes tasks. The unification framework can also be integrated into any other OWD/OVD methods to generate semantic-rich concepts for unknown/novel objects." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.813, + 0.892, + 0.873 + ], + "angle": 0, + "content": "Acknowledgements We gratefully acknowledge the support of MindSpore1, CANN (Compute Architecture for Neural Networks) and Ascend AI Processor used for this research." + }, + { + "type": "page_footnote", + "bbox": [ + 0.526, + 0.887, + 0.678, + 0.901 + ], + "angle": 0, + "content": "\\(^{1}\\)https://www.mindspore.cn/" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "15240" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.079, + 0.09, + 0.176, + 0.107 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.116, + 0.47, + 0.254 + ], + "angle": 0, + "content": "[1] Tom B. Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert-Voss, Gretchen Krueger, Tom Henighan, Rewon Child, Aditya Ramesh, Daniel M. Ziegler, Jeffrey Wu, Clemens Winter, Christopher Hesse, Mark Chen, Eric Sigler, Mateusz Litwin, Scott Gray, Benjamin Chess, Jack Clark, Christopher Berner, Sam McCandlish, Alec Radford, Ilya Sutskever, and Dario Amodei. Language models are few-shot learners. CoRR, abs/2005.14165, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.256, + 0.471, + 0.367 + ], + "angle": 0, + "content": "[2] Kai Chen, Jiaqi Wang, Jiangmiao Pang, Yuhang Cao, Yu Xiong, Xiaoxiao Li, Shuyang Sun, Wansen Feng, Ziwei Liu, Jiarui Xu, Zheng Zhang, Dazhi Cheng, Chenchen Zhu, Tianheng Cheng, Qijie Zhao, Buyu Li, Xin Lu, Rui Zhu, Yue Wu, Jifeng Dai, Jingdong Wang, Jianping Shi, Wanli Ouyang, Chen Change Loy, and Dahua Lin. Mmdetection: Open mmlab detection toolbox and benchmark. ArXiv, abs/1906.07155, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.368, + 0.471, + 0.438 + ], + "angle": 0, + "content": "[3] Xiyang Dai, Yinpeng Chen, Bin Xiao, Dongdong Chen, Mengchen Liu, Lu Yuan, and Lei Zhang. Dynamic head: Unifying object detection heads with attentions. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7373-7382, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.439, + 0.471, + 0.495 + ], + "angle": 0, + "content": "[4] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, K. Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. 2009 IEEE Conference on Computer Vision and Pattern Recognition, pages 248-255, 2009. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.497, + 0.471, + 0.552 + ], + "angle": 0, + "content": "[5] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. BERT: pre-training of deep bidirectional transformers for language understanding. CoRR, abs/1810.04805, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.554, + 0.47, + 0.623 + ], + "angle": 0, + "content": "[6] Xiaoyi Dong, Yinglin Zheng, Jianmin Bao, Ting Zhang, Dongdong Chen, Hao Yang, Ming Zeng, Weiming Zhang, Lu Yuan, Dong Chen, et al. Maskclip: Masked self-distillation advances contrastive language-image pretraining. arXiv preprint arXiv:2208.12262, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.625, + 0.47, + 0.707 + ], + "angle": 0, + "content": "[7] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. CoRR, abs/2010.11929, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.71, + 0.471, + 0.779 + ], + "angle": 0, + "content": "[8] Yu Du, Fangyun Wei, Zihe Zhang, Miaojing Shi, Yue Gao, and Guoqi Li. Learning to prompt for open-vocabulary object detection with vision-language model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14084-14093, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.781, + 0.471, + 0.835 + ], + "angle": 0, + "content": "[9] Yiqi Gao, Xinglin Hou, Yuanmeng Zhang, Tiezheng Ge, Yunig Jiang, and Peng Wang. Caponimage: Context-driven dense-captioning on image. arXiv preprint arXiv:2204.12974, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.837, + 0.471, + 0.894 + ], + "angle": 0, + "content": "[10] Georgia Gkioxari, Ross Girshick, Piotr Dólar, and Kaiming He. Detecting and recognizing human-object interactions. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 8359-8367, 2018. 1" + }, + { + "type": "list", + "bbox": [ + 0.079, + 0.116, + 0.471, + 0.894 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.133 + ], + "angle": 0, + "content": "[11] Alex Graves. Long short-term memory. Supervised sequence labelling with recurrent neural networks, pages 37-45, 2012. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.134, + 0.892, + 0.189 + ], + "angle": 0, + "content": "[12] Xiuye Gu, Tsung-Yi Lin, Weicheng Kuo, and Yin Cui. Open-vocabulary object detection via vision and language knowledge distillation. arXiv preprint arXiv:2104.13921, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.19, + 0.892, + 0.246 + ], + "angle": 0, + "content": "[13] Agrim Gupta, Piotr Dollar, and Ross Girshick. Lvis: A dataset for large vocabulary instance segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5356-5364, 2019. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.247, + 0.892, + 0.315 + ], + "angle": 0, + "content": "[14] Akshita Gupta, Sanath Narayan, KJ Joseph, Salman Khan, Fahad Shahbaz Khan, and Mubarak Shah. Ow-detr: Open-world detection transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9235-9244, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.316, + 0.892, + 0.369 + ], + "angle": 0, + "content": "[15] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In Proceedings of the IEEE international conference on computer vision, pages 2961-2969, 2017. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.371, + 0.892, + 0.453 + ], + "angle": 0, + "content": "[16] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In International Conference on Machine Learning, pages 4904-4916. PMLR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.455, + 0.892, + 0.524 + ], + "angle": 0, + "content": "[17] Justin Johnson, Andrej Karpathy, and Li Fei-Fei. Densecap: Fully convolutional localization networks for dense captioning. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4565–4574, 2016. 2, 3, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.525, + 0.892, + 0.593 + ], + "angle": 0, + "content": "[18] KJ Joseph, Salman Khan, Fahad Shahbaz Khan, and Vineeth N Balasubramanian. Towards open world object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5830-5840, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.595, + 0.892, + 0.663 + ], + "angle": 0, + "content": "[19] Aishwarya Kamath, Mannat Singh, Yann LeCun, Gabriel Synnaeve, Ishan Misra, and Nicolas Carion. Mdetr-modulated detection for end-to-end multi-modal understanding. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1780-1790, 2021. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.664, + 0.892, + 0.719 + ], + "angle": 0, + "content": "[20] Wonjae Kim, Bokyung Son, and Ildoo Kim. Vilt: Vision-and-language transformer without convolution or region supervision. In International Conference on Machine Learning, pages 5583-5594. PMLR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.72, + 0.892, + 0.761 + ], + "angle": 0, + "content": "[21] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014.5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.762, + 0.892, + 0.844 + ], + "angle": 0, + "content": "[22] Ranjay Krishna, Yuke Zhu, Oliver Groth, Justin Johnson, Kenji Hata, Joshua Kravitz, Stephanie Chen, Yannis Kalantidis, Li-Jia Li, David A Shamma, et al. Visual genome: Connecting language and vision using crowdsourced dense image annotations. International journal of computer vision, 123(1):32-73, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.845, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[23] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. arXiv preprint arXiv:2201.12086, 2022. 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "15241" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.147 + ], + "angle": 0, + "content": "[24] Liunian Harold Li, Mark Yatskar, Da Yin, Cho-Jui Hsieh, and Kai-Wei Chang. Visualbert: A simple and perform-. mant baseline for vision and language. arXiv preprint arXiv:1908.03557, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.149, + 0.471, + 0.234 + ], + "angle": 0, + "content": "[25] Liunian Harold Li, Pengchuan Zhang, Haotian Zhang, Jianwei Yang, Chunyuan Li, Yiwu Zhong, Lijuan Wang, Lu Yuan, Lei Zhang, Jenq-Neng Hwang, et al. Grounded language-image pre-training. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10965-10975, 2022. 2, 3, 4, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.235, + 0.471, + 0.289 + ], + "angle": 0, + "content": "[26] Xiangyang Li, Shuqiang Jiang, and Jungong Han. Learning object context for dense captioning. In Proceedings of the AAAI conference on artificial intelligence, volume 33, pages 8650-8657, 2019. 3, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.291, + 0.471, + 0.361 + ], + "angle": 0, + "content": "[27] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In European conference on computer vision, pages 740-755. Springer, 2014. 1, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.362, + 0.471, + 0.432 + ], + "angle": 0, + "content": "[28] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 9992-10002, 2021. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.434, + 0.471, + 0.489 + ], + "angle": 0, + "content": "[29] Jingjing Meng, Junsong Yuan, Jiong Yang, Gang Wang, and Yap-Peng Tan. Object instance search in videos via spatiotemporal trajectory discovery. IEEE Transactions on Multimedia, 18(1):116-127, 2015. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.49, + 0.471, + 0.558 + ], + "angle": 0, + "content": "[30] James Philbin, Ondrej Chum, Michael Isard, Josef Sivic, and Andrew Zisserman. Object retrieval with large vocabularies and fast spatial matching. In 2007 IEEE conference on computer vision and pattern recognition, pages 1-8. IEEE, 2007. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.561, + 0.471, + 0.645 + ], + "angle": 0, + "content": "[31] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.647, + 0.471, + 0.702 + ], + "angle": 0, + "content": "[32] Shafin Rahman, Salman Khan, and Fatih Porikli. Zero-shot object detection: Learning to simultaneously recognize and localize novel concepts. In Asian Conference on Computer Vision, pages 547-563. Springer, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.704, + 0.471, + 0.76 + ], + "angle": 0, + "content": "[33] Joseph Redmon, Santosh Divvala, Ross Girshick, and Ali Farhadi. You only look once: Unified, real-time object detection. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 779-788, 2016. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.761, + 0.471, + 0.816 + ], + "angle": 0, + "content": "[34] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. Advances in neural information processing systems, 28, 2015. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.818, + 0.471, + 0.9 + ], + "angle": 0, + "content": "[35] Hamid Rezatofighi, Nathan Tsoi, JunYoung Gwak, Amir Sadeghian, Ian Reid, and Silvio Savarese. Generalized intersection over union: A metric and a loss for bounding box regression. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 658-666, 2019. 4" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.471, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.893, + 0.162 + ], + "angle": 0, + "content": "[36] Shuai Shao, Zeming Li, Tianyuan Zhang, Chao Peng, Gang Yu, Xiangyu Zhang, Jing Li, and Jian Sun. Objects365: A large-scale, high-quality dataset for object detection. In Proceedings of the IEEE/CVF international conference on computer vision, pages 8430-8439, 2019. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.163, + 0.893, + 0.218 + ], + "angle": 0, + "content": "[37] Zhuang Shao, Jungong Han, Demetris Marnerides, and Kurt Debattista. Region-object relation-aware dense captioning via transformer. IEEE Transactions on Neural Networks and Learning Systems, 2022. 3, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.219, + 0.893, + 0.287 + ], + "angle": 0, + "content": "[38] Sheng Shen, Chunyuan Li, Xiaowei Hu, Yujia Xie, Jianwei Yang, Pengchuan Zhang, Anna Rohrbach, Zhe Gan, Lijuan Wang, Lu Yuan, et al. K-lite: Learning transferable visual models with external knowledge. arXiv preprint arXiv:2204.09222, 2022. 3, 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.288, + 0.893, + 0.33 + ], + "angle": 0, + "content": "[39] Wenhui Wang, Hangbo Bao, Li Dong, and Furu Wei. Vlmo: Unified vision-language pre-training with mixture-of-modality-experts. arXiv preprint arXiv:2111.02358, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.33, + 0.893, + 0.385 + ], + "angle": 0, + "content": "[40] Linjie Yang, Kevin Tang, Jianchao Yang, and Li-Jia Li. Dense captioning with joint inference and visual context. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2193-2202, 2017. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.386, + 0.893, + 0.455 + ], + "angle": 0, + "content": "[41] Lewei Yao, Jianhua Han, Youpeng Wen, Xiaodan Liang, Dan Xu, Wei Zhang, Zhenguo Li, Chunjing Xu, and Hang Xu. Detclip: Dictionary-enriched visual-concept paralleled pre-training for open-world detection. arXiv preprint arXiv:2209.09407, 2022. 1, 2, 3, 4, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.456, + 0.893, + 0.512 + ], + "angle": 0, + "content": "[42] Lewei Yao, Runhu Huang, Lu Hou, Guansong Lu, Minzhe Niu, Hang Xu, Xiaodan Liang, Zhenguo Li, Xin Jiang, and Chunjing Xu. Filip: Fine-grained interactive language-image pre-training. ArXiv, abs/2111.07783, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.512, + 0.893, + 0.58 + ], + "angle": 0, + "content": "[43] Guojun Yin, Lu Sheng, Bin Liu, Nenghai Yu, Xiaogang Wang, and Jing Shao. Context and attribute grounded dense captioning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6241-6250, 2019. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.58, + 0.893, + 0.649 + ], + "angle": 0, + "content": "[44] Alireza Zareian, Kevin Dela Rosa, Derek Hao Hu, and Shih-Fu Chang. Open-vocabulary object detection using captions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14393-14402, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.65, + 0.893, + 0.705 + ], + "angle": 0, + "content": "[45] Haoyang Zhang and Xuming He. Deep free-form deformation network for object-mask registration. In Proceedings of the IEEE International Conference on Computer Vision, pages 4251-4259, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.706, + 0.893, + 0.775 + ], + "angle": 0, + "content": "[46] Haotian Zhang, Pengchuan Zhang, Xiaowei Hu, Yen-Chun Chen, Liunian Harold Li, Xiyang Dai, Lijuan Wang, Lu Yuan, Jenq-Neng Hwang, and Jianfeng Gao. Glipv2: Unifying localization and vision-language understanding. arXiv preprint arXiv:2206.05836, 2022. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.776, + 0.893, + 0.858 + ], + "angle": 0, + "content": "[47] Shifeng Zhang, Cheng Chi, Yongqiang Yao, Zhen Lei, and Stan Z Li. Bridging the gap between anchor-based and anchor-free detection via adaptive training sample selection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9759-9768, 2020. 3, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.859, + 0.893, + 0.902 + ], + "angle": 0, + "content": "[48] Xiaowei Zhao, Xianglong Liu, Yifan Shen, Yuqing Ma, Yixuan Qiao, and Duorui Wang. Revisiting open world object detection. arXiv preprint arXiv:2201.00471, 2022. 1, 2, 3" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.893, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "15242" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.092, + 0.47, + 0.16 + ], + "angle": 0, + "content": "[49] Xingyi Zhou, Rohit Girdhar, Armand Joulin, Philipp Krahenbuhl, and Ishan Misra. Detecting twenty-thousand classes using image-level supervision. In European Conference on Computer Vision, pages 350-368. Springer, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.163, + 0.469, + 0.218 + ], + "angle": 0, + "content": "[50] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable detr: Deformable transformers for end-to-end object detection. arXiv preprint arXiv:2010.04159, 2020. 1" + }, + { + "type": "list", + "bbox": [ + 0.079, + 0.092, + 0.47, + 0.218 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.956 + ], + "angle": 0, + "content": "15243" + } + ] +] \ No newline at end of file diff --git a/2023/CapDet_ Unifying Dense Captioning and Open-World Detection Pretraining/c3eddfd4-ae1f-40f7-a7cf-96c54f1fea7b_origin.pdf b/2023/CapDet_ Unifying Dense Captioning and Open-World Detection Pretraining/c3eddfd4-ae1f-40f7-a7cf-96c54f1fea7b_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..26add4af171a2651dcf0400d5627d546e0be28fe --- /dev/null +++ b/2023/CapDet_ Unifying Dense Captioning and Open-World Detection Pretraining/c3eddfd4-ae1f-40f7-a7cf-96c54f1fea7b_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1a53cd40da13ae3896d09b3b695d7abd45abdf1f80024b6fb60132706b63f983 +size 2127027 diff --git a/2023/CapDet_ Unifying Dense Captioning and Open-World Detection Pretraining/full.md b/2023/CapDet_ Unifying Dense Captioning and Open-World Detection Pretraining/full.md new file mode 100644 index 0000000000000000000000000000000000000000..e3ca28b8704bda9b0b2b3da9e43e52f08139c869 --- /dev/null +++ b/2023/CapDet_ Unifying Dense Captioning and Open-World Detection Pretraining/full.md @@ -0,0 +1,260 @@ +# CapDet: Unifying Dense Captioning and Open-World Detection Pretraining + +Yanxin Long $^{1*}$ Youpeng Wen $^{1*}$ Jianhua Han $^{2}$ Hang Xu $^{2}$ Pengzhen Ren $^{1}$ Wei Zhang $^{2}$ Shen Zhao $^{1\dagger}$ Xiaodan Liang $^{1,3\dagger}$ + +1Shenzhen Campus of Sun Yat-sen University 2Huawei Noah's Ark Lab 3MBZUAI longyx9@mail2.sysu.edu.cn, wenyoupeng0@outlook.com, hanjianhua4@huawei.com, chromexbjxh@gmail.com, renpzh@mail.sysu.edu.cn, wz.zhang@huawei.com, z-s-06@163.com, xdliang328@gmail.com + +# Abstract + +Benefiting from large-scale vision-language pre-training on image-text pairs, open-world detection methods have shown superior generalization ability under the zero-shot or few-shot detection settings. However, a pre-defined category space is still required during the inference stage of existing methods and only the objects belonging to that space will be predicted. To introduce a "real" open-world detector, in this paper, we propose a novel method named CapDet to either predict under a given category list or directly generate the category of predicted bounding boxes. Specifically, we unify the open-world detection and dense caption tasks into a single yet effective framework by introducing an additional dense captioning head to generate the region-grounded captions. Besides, adding the captioning task will in turn benefit the generalization of detection performance since the captioning dataset covers more concepts. Experiment results show that by unifying the dense caption task, our CapDet has obtained significant performance improvements (e.g., $+2.1\%$ mAP on LVIS rare classes) over the baseline method on LVIS (1203 classes). Besides, our CapDet also achieves state-of-the-art performance on dense captioning tasks, e.g., $15.44\%$ mAP on VG V1.2 and $13.98\%$ on the VG-COCO dataset. + +# 1. Introduction + +Most state-of-the-art object detection methods [33, 34, 50] benefit from a large number of densely annotated detection datasets (e.g., COCO [27], Object365 [36], LVIS [12]). However, this closed-world setting results in the model only being able to predict categories that appear in the training set. Considering the ubiquity of new concepts in real-world scenes, it is very challenging to locate and identify these new visual concepts. This predictive ability of new concepts in open-world scenarios has very important research value + +![](images/366721316af1b64ee03ded007cf3162f6fdcbe9b26aaedea1fbbff23ca67b393.jpg) + +![](images/49d4c0e56a0c81fa4d585e5abfa1ce329a3cc45826d6c5e3b428f87956621d58.jpg) +Figure 1. Comparison of the different model predictions under OWD, OVD, and our setting. (a) OWD methods [14, 18, 48] are not able to describe the detailed category of the detected unknown objects and (b) the performance of OVD methods [8, 12, 41] usually depends on the pre-defined category list during the inference. (c) With the unification of two pipelines of dense captioning and open-world detection pre-training, our CapDet can either predict under a given category list or directly generate the description of predicted bounding boxes. + +in real-world applications such as object search [29, 30], instance registration [45], and human-object interaction modeling [10]. + +Currently, the open world scenario mainly includes two tasks: open world object detection [18] (OWD) and open-vocabulary object detection [44] (OVD). Although the paradigms of OWD and OVD tasks are closer to the real world, the former cannot describe the specific concept of the detected unknown objects and requires a pre-defined category list during the inference. Specifically, as shown + +in Figure 1, previous OWD methods [14,18,48] would recognize new concepts not in the predefined category space as "unknown". Further, another line of task OVD requires the model to learn a limited base class and generalize to novel classes. Compared to the zero-shot object detection (ZSD) proposed by [32], OVD allows the model to use external knowledge, e.g., knowledge distillation from a largescale vision-language pre-trained model [8, 12], image-caption pairs [44], image classification data [49], grounding data [25, 41, 46]. With the external knowledge, OVD methods show a superior generalization capacity to detect the novel classes within a given category space. However, as shown in Figure 1, when given an incomplete category list, OVD can only predict the concepts that appear in the given category list, otherwise, there will be recognition errors, (i.e., as illustrated in Figure 1 (b), the OVD methods prone to predict the "wall socket" as "remote", since the latter is in the category list but not the former). + +Thus, under the OVD setting, we mainly face the following two challenges: (i) it is difficult to define a complete list of categories; (ii) low response values on rare categories often lead to recognition errors. This is mainly because we cannot exhaustively enumerate new objects in the real world, and secondly, it is difficult to collect enough samples for rare classes. However, the fact that rare objects in the real world, even some new objects that are unknown to humans, such as UFOs, do not prevent people from using natural language to describe it as "a flying vehicle that looks like a Frisbee". + +Therefore, based on the above observations, in this paper, we consider a new setting that is closer to the open world and real scenes, i.e., we expect the model to both detect and recognize concepts in a given category list, and to generate corresponding natural language descriptions for new concepts or rare categories of objects. Early dense captioning methods [9, 17] can locate salient regions in images and generate the region-grounded captions with natural language. Inspired by this, to address the challenges faced in the OVD setting, we propose to unify the two pipelines of dense captioning and open-world detection pre-training into one training framework, called CapDet. It empowers the model with the ability to both accurately detect and recognize common object categories and generate dense captions for unknown and rare categories by unifying the two training tasks. + +Specifically, our CapDet constructs a unified data format for the dense captioning data and detection data. With the data unification, CapDet further adopts a unified pretraining paradigm including open-world object detection and dense captioning pre-training. For open-world detection pretraining, we treat the detection task as a semantic alignment task and adopt a dual encoder structure as [41] to locate and predict the given concepts list. The concepts + +list contains category names in detection data and region-grounded captions in dense captioning data. For dense captioning pretraining, CapDet proposes a dense captioning head to take the predicted proposals as input to generate the region-grounded captions. Due to the rich visual concepts in the dense captioning data, the integration of dense captioning tasks will in turn benefit the generalization of detection performance. + +Our experiments show that the integration of few dense captioning data brings in large improvement in the object detection datasets LVIS, e.g., $+2.7\%$ mAP on LVIS. The unification of dense captioning and detection pre-training gains an additional $2.3\%$ increment on LVIS and $2.1\%$ increment on LVIS rare classes. Besides, our model also achieves state-of-the-art performance on dense captioning tasks. Note that our method is the first to unify dense captioning and open-world detection pretraining. + +To summarize, our contributions are three folds: + +- We propose a novel open-vocabulary object detection framework CapDet, which cannot only detect and recognize concepts in a given category list but also generate corresponding natural language descriptions for new concept objects. +- We propose to unify the two pipelines of dense captioning and open-world detection pre-training into one training framework. Both two pre-training tasks are beneficial to each other. +- Experiments show that by unified dense captioning task and detection task, our CapDet gains significant performance improvements on the open-vocabulary object detection task (e.g., $+3.3\%$ mAP on LVIS rare classes). Furthermore, our CapDet also achieves state-of-the-art performance on the dense captioning tasks, e.g., $15.44\%$ mAP on Visual Genome (VG) V1.2 and $13.98\%$ mAP on VG-COCO. + +# 2. Related Work + +Vision-Language Pre-training. Vision-Language Pretraining [6, 16, 31] as a scheme in the domains of natural language processing [1, 5] and computer vision [7] obtains continual attention currently. And it exhibits strong performance and generalization ability on various downstream vision and cross-modal tasks. Among them, CLIP [31] and ALIGN [16] as dual-stream methods utilize large-scale image-text pairs on the Internet by cross-modal contrastive learning to get excellent zero-shot classification ability. Single-stream methods [20, 24] unify visual and textual embeddings in a single transformer-based model, which can perform text generation tasks such as image caption and VQA. Some mixed architectures [23, 39] combine single-stream and dual-stream to explore a unified way of vision-language understanding and generation. However, these methods take low-resolution images as input and serve the + +task of classification and retrieval. Those vision-language pre-training approaches can not be applied to pure computer vision task directly, i.e., object detection task. + +Open World Object Detection / Open-Vocabulary Object Detection. Object detection is a core computer vision task, which aims at localizing objects using a bounding box and classifying them. The mature detection approaches which show great performance on supervised data include one-stage detectors (i.e.,YOLO [33],ATSS [47]) having a relatively high detection efficiency and two-stage detectors (i.e., Faster R-CNN [34],Mask R-CNN [15]) having good detection accuracy. However, how to generalize these methods to rare classes and novel concepts in the real world is a challenge. Currently, several object detection approaches for such open-world scenes have attracted extensive attention from academia and industry. These methods are divided into two tasks which are called open-world object detection and open-vocabulary object detection respectively depending on whether to detect the class of unknown classes. + +For the OWD task, Zhao et al. [48] proposed a proposal advisor to assist in identifying unknown proposals without supervision and a class-specific expelling classifier to filter out confusing predictions. For the OVD task, GLIP [25] converts the detection data into grounding format and proposes a fusion module to learn semantic vision information in grounding data. K-Lite [38] reconstructs the input format of the data in GLIP from sequential to parallel and uses nouns hierarchy and definition to format text sequence. Det-CLIP [41] unifies detection, grounding, and image-text pair data in a paralleled formulation and constructs a concept dictionary to augment the text data, which strikes a balance between performance and efficiency. Differing from all these works, our CapDet can generate an open-set caption of each region proposal to cover situations where the semantics of new object instances are not in the given category list. + +Dense Captioning. Dense captioning aims at generating detailed descriptions for local regions, which usually needs to locate visual regions with semantic information and generate captions for these regions. J. Johnson et al. [17] utilized a fully convolutional localization network to locate regions of interest (RoIs) and then describe them. Afterward, many methods [26, 43] based on Faster-RCNN [34] and LSTM [11] are proposed to do dense captioning. X. Li et al. [26] arrange RoI features as a sequence and put them into LSTM with the guidance of the region features to form the complementary object context features. This method also needs ground truth bounding boxes auxiliary tests to achieve good results. But limited by the forget gate mechanism of LSTM, the inputted sequence cannot be too long. Then, the transformer-based method TDC [37] is proposed + +to tackle the long sequence forgotten problem. Instead, our CapDet proposes a transformer-based caption head to generate a caption using a single-stage detector ATSS while simultaneously achieving open-world detection. + +# 3. Method + +The overview of our proposed CapDet is shown in Figure 2. To construct a detector to either predict under a given category list or directly generate the concepts of predicted bounding boxes, we incorporate detection data and dense caption data together. In this section, we will present a unified data format for the detection data and dense caption data in Section 3.1, the model architecture and pre-training objectives for open-world object detection pre-training in Section 3.2 and dense captioning in Section 3.3. + +# 3.1. Unified Formulation + +We defined a unified triplet-wise data format $(x, \{\mathbf{b}_i\}_{i=1}^N, y_{i=1}^N)$ for each sample from different sources. Specifically, $x \in \mathbb{R}^{3 \times h \times w}$ is the input image, $\{\mathbf{b}_i | \mathbf{b}_i \in \mathbb{R}^4\}_{i=1}^N$ denotes the bounding boxes coordinates for each region of the image, and the $y_{i=1}^N$ represents the concepts of the corresponding boxes. $N$ denotes the number of regions. A concept $y_i$ formatted as a sentence contains the category and textual description of the corresponding region. In detection data, a concept $y$ consists of the category name and the corresponding definition from the concept dictionary [41], while $y_i$ represents the region-grounded caption in dense caption data. For example, for an image $x$ in detection data, $y_i$ can be: + +$$ +y _ {i} = \text {" p e r s o n , a h u m a n b e i n g . "} +$$ + +For an image $x$ from dense captioning data, $y_{i}$ can be: + +$$ +y _ {i} = \text {" a n o u t l e t o n t h e w a l l .} +$$ + +With the triplet, we can learn a unified image-text alignment objective on the detection data and the dense captioning data. The unified formulation also ensures the joint training of open-world object detection pre-training and dense captioning. + +# 3.2. Open-World Object Detection Pre-training + +Based on the unified formulation of detection data and dense captioning data, we regard the captions of regions in dense captioning data as a kind of category and utilize two different sources of data for the open-world object detection pre-training. Compared with the limited class list of detection data, dense caption data contains richer concepts and more semantic information than class names of individual regions. On the other hand, localization and recognition are two essential tasks of object detection. Traditional object + +![](images/3edf93e562aaab7bfa54789f47818bdd10fd383c264bb50247aad7f92d884362.jpg) +Figure 2. The overall architecture of CapDet. The training paradigm of CapDet contains open-world object detection pre-training and dense captioning. In detection, CapDet contains a dual vision-language encoder. The image encoder generates region embeddings from detection and dense captioning data. The regression loss and centerness loss are introduced to regress the locations. The text encoder takes the category concepts as input to generate the embeddings from the [EOS] token. Then we treat the detection task as a matching task and adopt an alignment loss for the category embeddings and region embeddings. In dense captioning, an additional dense captioning head is proposed to take the region embeddings as input and generate the textual captions for corresponding regions with natural language. + +detection always focuses on the salient objects in the image. While the dense captioning data contains lots of annotations which are just parts of an object, e.g., an ear of an elephant, it is not suitable to learn those annotations for the localization task. Therefore, we only calculate the localization loss on detection data. + +As shown in Figure 2, CapDet predicts the regions and treats the recognition task as a region-category matching task. For efficient learning on the matching task, we adopt the negative sampling proposed by [41] to provide negative concepts to enlarge the concept space in a batch. Specifically, for each iteration, we randomly sample a negative concept set and add to the positive concept set (N samples) in a batch to obtain the final concept set $y_{i = 1}^{M}$ , where M represents the sum of the number of positive and negative samples. Finally, we format the triplet to $(x,\{\mathbf{b}_i\}_{i = 1}^N,y_{i = 1}^M)$ . + +CapDet contains a dual vision-language encoder and takes the triplet $(x, \{\mathbf{b}_i\}_{i=1}^N, y_{i=1}^M)$ as input. The image encoder $\Phi_v$ is an object detector that can predict the bounding boxes of regions from the input image $x$ and output the region features $O \in \mathbb{R}^{K \times D}$ . The text encoder $\Phi_l$ takes the concept set $y_{i=1}^M$ as input and obtains the text embeddings $W \in \mathbb{R}^{M \times D}$ from the special token [EOS] concatenated with the text input. $K, D$ denotes the number of predicted + +regions and region feature dimensions. The alignment score matrix $S \in \mathbb{R}^{K \times M}$ of regions and texts is calculated by: + +$$ +O = \Phi_ {v} (x), W = \Phi_ {l} \left(y _ {i = 1} ^ {M}\right), S = O W ^ {T} \tag {1} +$$ + +where $T$ denotes the transpose operation. A ground-truth alignment matrix $G \in \{0,1\}^{K \times M}$ is constructed to indicate the matching relation of regions and concepts. The alignment loss $\mathcal{L}_{\text{align}}$ is calculated by the predicted alignment scores of regions $S$ and the ground-truth alignment matrix $G$ . Following [25, 41], we adopt the ATSS [47] detector as an image encoder, and $\mathcal{L}_{\text{align}}$ is typically a sigmoid focal loss. As a one-stage detector, the localization loss contains centeredness loss $\mathcal{L}_{\text{cen}}$ and bounding box regression loss $\mathcal{L}_{\text{reg}}$ . The training objective of detection pre-training can be written as: + +$$ +\mathcal {L} = \left\{ \begin{array}{l l} \mathcal {L} _ {\text {a l i g n}} + \alpha \mathcal {L} _ {\text {r e g}} + \beta \mathcal {L} _ {\text {c e n t e r}}, & \text {f o r d e t e c t i o n} \\ \mathcal {L} _ {\text {a l i g n}}, & \text {f o r d e n s e c a p t i o n i n g} \end{array} \right. \tag {2} +$$ + +where $\alpha$ and $\beta$ denote the weights for the centerness loss $\mathcal{L}_{cen}$ and box regression loss $\mathcal{L}_{reg}$ , respectively. The $\mathcal{L}_{cen}$ is the sigmoid loss and the $\mathcal{L}_{reg}$ is the GIOU loss [35]. + +
MODELBACKBONEPRE-TRAIN DATAIMAGES NUMBERAPLVIS APr / APc / APf
MASK-RCNN [15]SWIN-TLVIS0.1M34.119.1 / 34.0 / 37.0
ATSS [47]SWIN-TLVIS0.1M33.619.7 / 32.4 / 37.2
ATSS [47]SWIN-LLVIS0.1M43.930.6 / 43.7 / 46.3
MDETR [19]RN101GOLDG+0.77M24.220.9 / 24.3 / 24.2
GLIP-T(A) [25]SWIN-T+DH+FO3650.66M18.514.2 / 13.9 / 23.4
GLIP-T(C) [25]SWIN-T+DH+FO365,GOLDG1.43M24.917.7 / 19.5 / 31.0
GLIP-T [25]SWIN-T+DH+FO365,GOLDG,CAP4M5.43M26.020.8 / 21.4 / 31.0
K-LITE [38]SWIN-TO3650.66M21.314.8 / 18.6 / 24.8
K-LITE [38]SWIN-TO365,GOLDG1.43M26.117.2 / 24.6 / 29.0
GLIPv2-T [46]SWIN-T+DH+FO365,GOLDG,CAP4M5.43M29.0- / - / -
DETCLIP-T(A) [41]SWIN-TO3650.66M28.826.0 / 28.0 / 30.0
DETCLIP-T(B) [41]SWIN-TO365, GOLDG1.43M34.426.9 / 33.9 / 36.3
DETCLIP-T(C)* [41]SWIN-TO365, VG0.73M31.527.5 / 30.6 / 33.0
CAPDET (OURS)SWIN-TO365, VG0.73M33.829.6 / 32.8 / 35.5
+ +Table 1. Zero-shot performance on LVIS [13] MiniVal5k datasets. $\mathrm{AP}_r / \mathrm{AP}_c / \mathrm{AP}_f$ indicate the AP values for rare, common, and frequent categories, respectively. "DH" and "F" in GLIP [25] baselines stand for the dynamic head [3] and cross-modal fusion, respectively. Baselines with * are implemented with our code base. GoldG+ denotes the GoldG plus the COCO [27] caption dataset. + +# 3.3. Dense Captioning + +The open-world object detection pre-training ensures CapDet gains the capacity to detect under given an arbitrary category list. However, when the given category list is not complete enough to cover the potential classes on a new domain data, the detector will perform worse on the categories which are not in the given list. Considering such limitation, we propose a dense captioning head $\Phi_C$ to generate semantically rich concepts with natural language for the predicted proposals. In the dense captioning task, the model receives an image and produces a set of regions and the corresponding captions. The dense captioning head is a cross-modal decoder that takes the c predicted regions features $O$ generated by the image encoder as input. The captioning (i.e., language modeling) loss is calculated by: + +$$ +\mathcal {L} _ {c a p} = - \log p \left(y _ {i t} \mid \Phi_ {c} \left(y _ {i (\tau < t)}, O _ {i}\right)\right), \tag {3} +$$ + +where $y_{it}$ means the $t$ token in caption $y_{i}$ corresponding to region feature $O_{i}$ , and $y_{i(\tau < t)}$ means tokens before $t$ in caption $y_{i}$ . The overall pre-training loss can be written as: + +$$ +\mathcal {L} = w _ {d} \mathcal {L} _ {\text {d e t}} + w _ {c} \mathcal {L} _ {\text {c a p}}, \tag {4} +$$ + +where $w_{d}, w_{c}$ denote the weighting factor of $\mathcal{L}_{det}$ and $\mathcal{L}_{cap}$ . + +To minimize the gap in the type of bounding boxes between the detection data and dense captioning data, we propose a simple way to transform our detector as a class-agnostic detector and only select the top $k$ regions based on the centeredness scores to adapt to the dense captioning + +task. We can fine-tune our CapDet on the dense captioning data to achieve better performance. Specifically, we propose "object" as the foreground concept and "background" as the background concept. The text encoder $\Phi_l$ outputs the concept embeddings $W' \in \mathbb{R}^{2 \times D}$ . Then the alignment scores $S' \in \mathbb{R}^{K \times 2}$ is calculated by Eqn. 1. The captioning head takes the top $k$ most confident proposal embeddings based on centeredness scores as input to predict the region-grounded captions. + +# 4. Experiment + +Implementation Details. For the image encoder, we adopt the Swin-T backbone proposed in SwinTransformer [28] which is pre-trained on ImageNet-1K [4]. We use 12 layers 8 heads transformer as our text encoder and load a base model checkpoint released by FILIP [42], in order to make a fair comparison with DetCLIP [41]. The structure of the dense captioning head is consistent with that in the text encoder but trained from scratch for a fair comparison. We employ AdamW [21] optimizer and set the batch size to 32. The learning rate is set to $1.4 \times 10^{-4}$ for the parameters of the image encoder and detection head, and $1.4 \times 10^{-5}$ for the text encoder and dense captioning head. When fine-tuning the VG dataset to do the dense captioning task, we set the learning rate to $1.4 \times 10^{-4}$ . Without otherwise specified, all models are trained with 12 epochs and the learning rate is decayed with a factor of 0.1 at the 8-th and the 11-th epoch. The context token + +![](images/8f647bdc722a93a11d7ce02114dbac662df17023a0fec08ba576c8b005716469.jpg) +Figure 3. Qualitative visualizations between GLIP-T, DetCLIP-T(C) and CapDet. From top to down, the three rows of images show the LVIS zero-shot detection results of GLIP-T, DetCLIP-T(C), and CapDet respectively. All models are pre-trained on O365 and VG. + +length for input text is set to 20. We set the number of input captions to 150, and the number of the region features N is determined by the feature map. The loss weight factor $w_{c}$ and $w_{d}$ are both set to 1.0. We build our model on MMDetection [2] code base. + +Dataset. Our CapDet is trained with two types of data, including detection data and caption data. Following DetCLIP [41], we use Object365 [36] (it will be abbreviated as O365 in the following paper) as detection data, and sample 0.66M data from O365 v2 for training. Following GLIP [25] and DetCLIP [41], LVIS [13] MiniVal5k (defined in [19]) which has 5000 images is used for detection evaluation. Moreover, we remove the training samples contained in the LVIS dataset for fair zero-transfer evaluation. For dense captioning data, we mainly conduct our experiments on VG [22] V1.2 and VG-COCO (defined in [37]). Following [37], we allocate 77398 images for training and 5000 images for validation and testing on VG. As demonstrated in [22], the ground-truth bounding boxes of VG are much denser than the other object detection datasets, i.e., the average number of per sample in MS COCO [27] is only 7.1 vs. 35.4 in VG. Then an intersection of VG V1.2 and MS COCO is proposed by [37] and is denoted as VG-COCO, which has 38080 images for training, 2489 for validation, and 2476 for testing. + +Benchmark Settings. We mainly evaluate our method on open-vocabulary object detection and dense captioning task. For open-vocabulary object detection, we evaluate the direct domain transfer on LVIS [13] which contains 1203 categories. Following [25, 41], we metric the zero-shot + +detection performance by the Average Precision (AP) on a 5k subset. The annotations of LVIS data are split into three folds, i.e., rare, common, and frequency, based on the number of categories. Since there is almost no overlap between the rare classes and the classes of training dataset Objects365 [36], the AP of the rare classes shows a valuable zero-shot detection performance. For the dense captioning task, we follow the setting of [17] to evaluate the VG and VG-COCO. The evaluation metric we adopt is the mean Average Precision proposed by [17] which is calculated across a range of thresholds for both localization and language accuracy, i.e., the intersection over union (IOU) thresholds .3, .4, .5, .6, .7 are used for localization and the METEOR score' thresholds 0, .05, .1, .15, .2, .25 is adopted for evaluating the language generation. + +# 4.1. Open-world Detection Results + +Table 1 shows our zero-shot object detection performance on LVIS. We mainly train our CapDet with the backbone Swin-T [28] on the detection data Objects365 [36] and dense captioning data (VG [22]). Since DetCLIP does not report the performance on O365 and VG, we train DetCLIP on the two datasets under the same settings and denote it as DetCLIP-T(C) for a fair comparison. Comparing the 11th row and 12th row, our CapDet outperforms DetCLIP-T(C) on the same data scale and backbone with an extra simple caption head. Moreover, our model's zero-shot performance even surpasses the fully-supervised model with the same backbone by a large margin on rare classes, i.e., CapDet outperforms ATSS by $9.9\%$ . + +Qualitative Visualizations Figure. 3 illustrates the detection results on LVIS [22] dataset from GLIP-T, DetCLIP-T(C), and CapDet. All three models are trained on O365 + +![](images/9ed982972812d1c237f8b600b279b11086cb1ea5113a52fc4e174c9aae3b8d5b.jpg) +Figure 4. Qualitative visualizations between JIVC and CapDet. "w/o ft" means do caption without finetune, while "w/ ft" means with finetune. + +and VG, and details are given in Section 4.3. Given a category list, the rare classes are detected more precisely by our CapDet, e.g., "kitchen table" in the first column, "horse buggy" in the third column, and "fishbowl" in the sixth column that our model CapDet detects correctly but the other two not. + +# 4.2. Dense Captioning Results + +Due to the target bounding boxes in dense captioning data containing lots of local structures of objects and being much denser than the bounding boxes in object detection data, we do not regress the bounding box in the pre-training stage. The previous works directly train on the dense captioning data and generate captions on the top $k$ proposals ranking by a confidence score. When fine-tuning our model on the VG dataset for the dense captioning tasks, we transform our CapDet into a class-agnostic detector. Specifically, we propose "object" as the foreground concept and "background" as the background concept for computing alignment scores. The scores are used as proposal confidences to predict the region-grounded captions. + +Table 2 and Table 3 show CapDet significantly outperforms the latest work TDC [37] by $2.5\%$ on mAP on VG and TDC+ROCSU [37] by $2.08\%$ , respectively. It is worth noticing that, even against given the ground-truth bounding boxes with the previous method COCG [26] denoted + +
MethodmAP(%)
FCLN [17]5.16
JIVC [40]9.96
ImgG [26]9.68
COCD [26]9.75
COCG [26]10.39
CAG-Net [43]10.51
TDC [37]11.90
CapDet (Ours)15.44
+ +Table 2. Comparison of mAP (%) performance on dense captioning benchmark on the VG V1.2 dataset. + +as COCOG>, our CapDet still gains a $43.80\%$ mAP increase and achieves state-of-the-art. One important reason is that the excellent detection performance of our model assists the localization ability of dense captioning tasks. + +Qualitative Visualizations. Figure 4 shows a qualitative visualization comparison between JIVC [40] and our CapDet. The three image rows from top to bottom are the visualization of JIVC, CapDet without fine-tuning, and CapDet with finetuning. In the second row, CapDet can locate more objects than JIVC, owing to our model's superior localization performance. After finetuning, CapDet can further describe a region rather than a single object such as + +
MethodmAP(%)
FCLN [17]4.23
JIVC [40]7.85
Max Pooling [26]7.86
ImgG [26]7.81
COCD [26]7.92
COCG [26]8.90
COCG-LocSiz [26]8.76
COCG&GT [26]9.79
TDC+ROCSU [37]11.9
CapDet (Ours)13.98
+ +Table 3. Comparison of mAP (%) performance on the dense captioning benchmark on the VG-COCO Dataset. + +
MODELDC HEADLVIS
APAPr / APc / APf
GLIP-T×30.422.5 / 29.0 / 33.0
GLIP-T✓33.127.0 / 32.1 / 35.0
DETCLIP-T×31.527.5 / 30.6 / 33.0
DETCLIP-T✓33.829.6 / 32.8 / 35.5
+ +Table 4. Ablations on integrating our dense captioning head into different baselines. + +
Pre-training DataFine-tuneDCap mAP(%)Box mAP(%)
VG✘12.8627.65
O365,VG✘4.729.65
VG✓13.8328.58
O365,VG✓15.4430.61
+ +Table 5. Ablations on incorporating data from different sources. "DCap" stands for the dense caption mAP. + +"two women in a kitchen" in the 5-th column. + +# 4.3. Ablation Studies + +# 4.3.1 Ablations for Unified Pre-training + +Effect on different baselines. Table 4 investigates the advantages of dense captioning heads on different baselines. We integrate our dense captioning head with GLIP-T or DetCLIP-T. The GLIP-T is implemented with parallel text encoding without external knowledge following the setting as ablations in [38] on our code base. All the results are pre-trained on Objects365 and VG. The results show that our dense captioning head is able to boost the generalization and model-agnostic. + +Effect of dense captioning data. Table 1 shows the efficiency of incorporating dense captioning data. Specifically, only $0.07\mathrm{M}$ data added, the DetCLIP-T(C) gains $+2.7\%$ overall AP and $+1.5\%$ $AP_{r}$ on LVIS compared to DetCLIP- + +T(A). The performance of DetCLP-T(A) on rare categories also outperforms DetCLIP-T(C) train on Objects365 and GOLDG, while the data size is 1.43M vs. 0.73M. + +# 4.3.2 Ablations for dense captioning + +We investigate the impact of training policy and data from different sources on the dense captioning task. As shown in row1 in Table 5, our CapDet still achieves a significant performance which is directly trained on VG outperforms the previous task (i.e., TDC [37] in Table 2). Row2 is our CapDet and is pre-trained on Objects365 and VG, while only the bounding box in the Objects365 is regressed, and then transformed on a dense captioning task. Since the type of bounding boxes in dense captioning is different from the detection data, the result of the direct transforming to dense captioning is worse. However, we've proved that our model still keeps the dense captioning capacity on the salient objects in Figure 4. The results in row3 and row4 indicate that pre-training on the detection data Objects365 is also beneficial to the dense captioning task. + +# 5. Limitations + +These are a few issues that we need to improve in the future: (1) Although our unification training paradigm works well on open-vocabulary object detection and dense captioning task, the training of dense captioning generation costs lots of time. (2) In addition, existing dense captioning data is high-cost to collect. We will research how to collect large-scale dense captioning data by auto annotation and get better performance with the scaled-up data. + +# 6. Conclusion + +In this paper, we propose a novel open-world object detection method named CapDet. Our CapDet is more practical in the open world and real scenes. Specifically, CapDet introduces a unification training framework including open-world object detection pre-training and dense captioning. The unification enables our CapDet to localize and recognize concepts in an arbitrary given category list or directly generate textual captions for predicted new concept objects. Experiments show that the design of unification is both beneficial to open-world object detection tasks and dense captioning tasks. In the future, our CapDet can be easily injected into any open world and real scenes tasks. The unification framework can also be integrated into any other OWD/OVD methods to generate semantic-rich concepts for unknown/novel objects. + +Acknowledgements We gratefully acknowledge the support of MindSpore1, CANN (Compute Architecture for Neural Networks) and Ascend AI Processor used for this research. + +# References + +[1] Tom B. Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert-Voss, Gretchen Krueger, Tom Henighan, Rewon Child, Aditya Ramesh, Daniel M. Ziegler, Jeffrey Wu, Clemens Winter, Christopher Hesse, Mark Chen, Eric Sigler, Mateusz Litwin, Scott Gray, Benjamin Chess, Jack Clark, Christopher Berner, Sam McCandlish, Alec Radford, Ilya Sutskever, and Dario Amodei. Language models are few-shot learners. CoRR, abs/2005.14165, 2020. 2 +[2] Kai Chen, Jiaqi Wang, Jiangmiao Pang, Yuhang Cao, Yu Xiong, Xiaoxiao Li, Shuyang Sun, Wansen Feng, Ziwei Liu, Jiarui Xu, Zheng Zhang, Dazhi Cheng, Chenchen Zhu, Tianheng Cheng, Qijie Zhao, Buyu Li, Xin Lu, Rui Zhu, Yue Wu, Jifeng Dai, Jingdong Wang, Jianping Shi, Wanli Ouyang, Chen Change Loy, and Dahua Lin. Mmdetection: Open mmlab detection toolbox and benchmark. ArXiv, abs/1906.07155, 2019. 6 +[3] Xiyang Dai, Yinpeng Chen, Bin Xiao, Dongdong Chen, Mengchen Liu, Lu Yuan, and Lei Zhang. Dynamic head: Unifying object detection heads with attentions. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7373-7382, 2021. 5 +[4] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, K. Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. 2009 IEEE Conference on Computer Vision and Pattern Recognition, pages 248-255, 2009. 5 +[5] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. BERT: pre-training of deep bidirectional transformers for language understanding. CoRR, abs/1810.04805, 2018. 2 +[6] Xiaoyi Dong, Yinglin Zheng, Jianmin Bao, Ting Zhang, Dongdong Chen, Hao Yang, Ming Zeng, Weiming Zhang, Lu Yuan, Dong Chen, et al. Maskclip: Masked self-distillation advances contrastive language-image pretraining. arXiv preprint arXiv:2208.12262, 2022. 2 +[7] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. CoRR, abs/2010.11929, 2020. 2 +[8] Yu Du, Fangyun Wei, Zihe Zhang, Miaojing Shi, Yue Gao, and Guoqi Li. Learning to prompt for open-vocabulary object detection with vision-language model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14084-14093, 2022. 1, 2 +[9] Yiqi Gao, Xinglin Hou, Yuanmeng Zhang, Tiezheng Ge, Yunig Jiang, and Peng Wang. Caponimage: Context-driven dense-captioning on image. arXiv preprint arXiv:2204.12974, 2022. 2 +[10] Georgia Gkioxari, Ross Girshick, Piotr Dólar, and Kaiming He. Detecting and recognizing human-object interactions. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 8359-8367, 2018. 1 + +[11] Alex Graves. Long short-term memory. Supervised sequence labelling with recurrent neural networks, pages 37-45, 2012. 3 +[12] Xiuye Gu, Tsung-Yi Lin, Weicheng Kuo, and Yin Cui. Open-vocabulary object detection via vision and language knowledge distillation. arXiv preprint arXiv:2104.13921, 2021. 1, 2 +[13] Agrim Gupta, Piotr Dollar, and Ross Girshick. Lvis: A dataset for large vocabulary instance segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5356-5364, 2019. 5, 6 +[14] Akshita Gupta, Sanath Narayan, KJ Joseph, Salman Khan, Fahad Shahbaz Khan, and Mubarak Shah. Ow-detr: Open-world detection transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9235-9244, 2022. 1, 2 +[15] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In Proceedings of the IEEE international conference on computer vision, pages 2961-2969, 2017. 3, 5 +[16] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In International Conference on Machine Learning, pages 4904-4916. PMLR, 2021. 2 +[17] Justin Johnson, Andrej Karpathy, and Li Fei-Fei. Densecap: Fully convolutional localization networks for dense captioning. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4565–4574, 2016. 2, 3, 6, 7, 8 +[18] KJ Joseph, Salman Khan, Fahad Shahbaz Khan, and Vineeth N Balasubramanian. Towards open world object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5830-5840, 2021. 1, 2 +[19] Aishwarya Kamath, Mannat Singh, Yann LeCun, Gabriel Synnaeve, Ishan Misra, and Nicolas Carion. Mdetr-modulated detection for end-to-end multi-modal understanding. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1780-1790, 2021. 5, 6 +[20] Wonjae Kim, Bokyung Son, and Ildoo Kim. Vilt: Vision-and-language transformer without convolution or region supervision. In International Conference on Machine Learning, pages 5583-5594. PMLR, 2021. 2 +[21] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014.5 +[22] Ranjay Krishna, Yuke Zhu, Oliver Groth, Justin Johnson, Kenji Hata, Joshua Kravitz, Stephanie Chen, Yannis Kalantidis, Li-Jia Li, David A Shamma, et al. Visual genome: Connecting language and vision using crowdsourced dense image annotations. International journal of computer vision, 123(1):32-73, 2017. 6 +[23] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. arXiv preprint arXiv:2201.12086, 2022. 2 + +[24] Liunian Harold Li, Mark Yatskar, Da Yin, Cho-Jui Hsieh, and Kai-Wei Chang. Visualbert: A simple and perform-. mant baseline for vision and language. arXiv preprint arXiv:1908.03557, 2019. 2 +[25] Liunian Harold Li, Pengchuan Zhang, Haotian Zhang, Jianwei Yang, Chunyuan Li, Yiwu Zhong, Lijuan Wang, Lu Yuan, Lei Zhang, Jenq-Neng Hwang, et al. Grounded language-image pre-training. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10965-10975, 2022. 2, 3, 4, 5, 6 +[26] Xiangyang Li, Shuqiang Jiang, and Jungong Han. Learning object context for dense captioning. In Proceedings of the AAAI conference on artificial intelligence, volume 33, pages 8650-8657, 2019. 3, 7, 8 +[27] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In European conference on computer vision, pages 740-755. Springer, 2014. 1, 5, 6 +[28] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 9992-10002, 2021. 5, 6 +[29] Jingjing Meng, Junsong Yuan, Jiong Yang, Gang Wang, and Yap-Peng Tan. Object instance search in videos via spatiotemporal trajectory discovery. IEEE Transactions on Multimedia, 18(1):116-127, 2015. 1 +[30] James Philbin, Ondrej Chum, Michael Isard, Josef Sivic, and Andrew Zisserman. Object retrieval with large vocabularies and fast spatial matching. In 2007 IEEE conference on computer vision and pattern recognition, pages 1-8. IEEE, 2007. 1 +[31] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 2 +[32] Shafin Rahman, Salman Khan, and Fatih Porikli. Zero-shot object detection: Learning to simultaneously recognize and localize novel concepts. In Asian Conference on Computer Vision, pages 547-563. Springer, 2018. 2 +[33] Joseph Redmon, Santosh Divvala, Ross Girshick, and Ali Farhadi. You only look once: Unified, real-time object detection. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 779-788, 2016. 1, 3 +[34] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. Advances in neural information processing systems, 28, 2015. 1, 3 +[35] Hamid Rezatofighi, Nathan Tsoi, JunYoung Gwak, Amir Sadeghian, Ian Reid, and Silvio Savarese. Generalized intersection over union: A metric and a loss for bounding box regression. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 658-666, 2019. 4 + +[36] Shuai Shao, Zeming Li, Tianyuan Zhang, Chao Peng, Gang Yu, Xiangyu Zhang, Jing Li, and Jian Sun. Objects365: A large-scale, high-quality dataset for object detection. In Proceedings of the IEEE/CVF international conference on computer vision, pages 8430-8439, 2019. 1, 6 +[37] Zhuang Shao, Jungong Han, Demetris Marnerides, and Kurt Debattista. Region-object relation-aware dense captioning via transformer. IEEE Transactions on Neural Networks and Learning Systems, 2022. 3, 6, 7, 8 +[38] Sheng Shen, Chunyuan Li, Xiaowei Hu, Yujia Xie, Jianwei Yang, Pengchuan Zhang, Anna Rohrbach, Zhe Gan, Lijuan Wang, Lu Yuan, et al. K-lite: Learning transferable visual models with external knowledge. arXiv preprint arXiv:2204.09222, 2022. 3, 5, 8 +[39] Wenhui Wang, Hangbo Bao, Li Dong, and Furu Wei. Vlmo: Unified vision-language pre-training with mixture-of-modality-experts. arXiv preprint arXiv:2111.02358, 2021. 2 +[40] Linjie Yang, Kevin Tang, Jianchao Yang, and Li-Jia Li. Dense captioning with joint inference and visual context. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2193-2202, 2017. 7, 8 +[41] Lewei Yao, Jianhua Han, Youpeng Wen, Xiaodan Liang, Dan Xu, Wei Zhang, Zhenguo Li, Chunjing Xu, and Hang Xu. Detclip: Dictionary-enriched visual-concept paralleled pre-training for open-world detection. arXiv preprint arXiv:2209.09407, 2022. 1, 2, 3, 4, 5, 6 +[42] Lewei Yao, Runhu Huang, Lu Hou, Guansong Lu, Minzhe Niu, Hang Xu, Xiaodan Liang, Zhenguo Li, Xin Jiang, and Chunjing Xu. Filip: Fine-grained interactive language-image pre-training. ArXiv, abs/2111.07783, 2022. 5 +[43] Guojun Yin, Lu Sheng, Bin Liu, Nenghai Yu, Xiaogang Wang, and Jing Shao. Context and attribute grounded dense captioning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6241-6250, 2019. 3, 7 +[44] Alireza Zareian, Kevin Dela Rosa, Derek Hao Hu, and Shih-Fu Chang. Open-vocabulary object detection using captions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14393-14402, 2021. 1, 2 +[45] Haoyang Zhang and Xuming He. Deep free-form deformation network for object-mask registration. In Proceedings of the IEEE International Conference on Computer Vision, pages 4251-4259, 2017. 1 +[46] Haotian Zhang, Pengchuan Zhang, Xiaowei Hu, Yen-Chun Chen, Liunian Harold Li, Xiyang Dai, Lijuan Wang, Lu Yuan, Jenq-Neng Hwang, and Jianfeng Gao. Glipv2: Unifying localization and vision-language understanding. arXiv preprint arXiv:2206.05836, 2022. 2, 5 +[47] Shifeng Zhang, Cheng Chi, Yongqiang Yao, Zhen Lei, and Stan Z Li. Bridging the gap between anchor-based and anchor-free detection via adaptive training sample selection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9759-9768, 2020. 3, 4, 5 +[48] Xiaowei Zhao, Xianglong Liu, Yifan Shen, Yuqing Ma, Yixuan Qiao, and Duorui Wang. Revisiting open world object detection. arXiv preprint arXiv:2201.00471, 2022. 1, 2, 3 + +[49] Xingyi Zhou, Rohit Girdhar, Armand Joulin, Philipp Krahenbuhl, and Ishan Misra. Detecting twenty-thousand classes using image-level supervision. In European Conference on Computer Vision, pages 350-368. Springer, 2022. 2 +[50] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable detr: Deformable transformers for end-to-end object detection. arXiv preprint arXiv:2010.04159, 2020. 1 \ No newline at end of file diff --git a/2023/CapDet_ Unifying Dense Captioning and Open-World Detection Pretraining/images.zip b/2023/CapDet_ Unifying Dense Captioning and Open-World Detection Pretraining/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..e1a051698b36fb45c2cc6b7eb06b7b72de29334f --- /dev/null +++ b/2023/CapDet_ Unifying Dense Captioning and Open-World Detection Pretraining/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a196208bc72fba5331dd9254788c7cc31e7cdfe49ab32be80f86526da6839c52 +size 841401 diff --git a/2023/CapDet_ Unifying Dense Captioning and Open-World Detection Pretraining/layout.json b/2023/CapDet_ Unifying Dense Captioning and Open-World Detection Pretraining/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..da39fd6ec60c882f8b070b15a06f3c5cf322fb3b --- /dev/null +++ b/2023/CapDet_ Unifying Dense Captioning and Open-World Detection Pretraining/layout.json @@ -0,0 +1,7497 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 62, + 102, + 531, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 102, + 531, + 121 + ], + "spans": [ + { + "bbox": [ + 62, + 102, + 531, + 121 + ], + "type": "text", + "content": "CapDet: Unifying Dense Captioning and Open-World Detection Pretraining" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 103, + 142, + 489, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 103, + 142, + 489, + 171 + ], + "spans": [ + { + "bbox": [ + 103, + 142, + 489, + 171 + ], + "type": "text", + "content": "Yanxin Long" + }, + { + "bbox": [ + 103, + 142, + 489, + 171 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 103, + 142, + 489, + 171 + ], + "type": "text", + "content": " Youpeng Wen" + }, + { + "bbox": [ + 103, + 142, + 489, + 171 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 103, + 142, + 489, + 171 + ], + "type": "text", + "content": " Jianhua Han" + }, + { + "bbox": [ + 103, + 142, + 489, + 171 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 103, + 142, + 489, + 171 + ], + "type": "text", + "content": " Hang Xu" + }, + { + "bbox": [ + 103, + 142, + 489, + 171 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 103, + 142, + 489, + 171 + ], + "type": "text", + "content": " Pengzhen Ren" + }, + { + "bbox": [ + 103, + 142, + 489, + 171 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 103, + 142, + 489, + 171 + ], + "type": "text", + "content": " Wei Zhang" + }, + { + "bbox": [ + 103, + 142, + 489, + 171 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 103, + 142, + 489, + 171 + ], + "type": "text", + "content": " Shen Zhao" + }, + { + "bbox": [ + 103, + 142, + 489, + 171 + ], + "type": "inline_equation", + "content": "^{1\\dagger}" + }, + { + "bbox": [ + 103, + 142, + 489, + 171 + ], + "type": "text", + "content": " Xiaodan Liang" + }, + { + "bbox": [ + 103, + 142, + 489, + 171 + ], + "type": "inline_equation", + "content": "^{1,3\\dagger}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 94, + 172, + 497, + 225 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 94, + 172, + 497, + 225 + ], + "spans": [ + { + "bbox": [ + 94, + 172, + 497, + 225 + ], + "type": "text", + "content": "1Shenzhen Campus of Sun Yat-sen University 2Huawei Noah's Ark Lab 3MBZUAI longyx9@mail2.sysu.edu.cn, wenyoupeng0@outlook.com, hanjianhua4@huawei.com, chromexbjxh@gmail.com, renpzh@mail.sysu.edu.cn, wz.zhang@huawei.com, z-s-06@163.com, xdliang328@gmail.com" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 143, + 255, + 191, + 266 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 255, + 191, + 266 + ], + "spans": [ + { + "bbox": [ + 143, + 255, + 191, + 266 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 275, + 290, + 550 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 275, + 290, + 550 + ], + "spans": [ + { + "bbox": [ + 46, + 275, + 290, + 550 + ], + "type": "text", + "content": "Benefiting from large-scale vision-language pre-training on image-text pairs, open-world detection methods have shown superior generalization ability under the zero-shot or few-shot detection settings. However, a pre-defined category space is still required during the inference stage of existing methods and only the objects belonging to that space will be predicted. To introduce a \"real\" open-world detector, in this paper, we propose a novel method named CapDet to either predict under a given category list or directly generate the category of predicted bounding boxes. Specifically, we unify the open-world detection and dense caption tasks into a single yet effective framework by introducing an additional dense captioning head to generate the region-grounded captions. Besides, adding the captioning task will in turn benefit the generalization of detection performance since the captioning dataset covers more concepts. Experiment results show that by unifying the dense caption task, our CapDet has obtained significant performance improvements (e.g., " + }, + { + "bbox": [ + 46, + 275, + 290, + 550 + ], + "type": "inline_equation", + "content": "+2.1\\%" + }, + { + "bbox": [ + 46, + 275, + 290, + 550 + ], + "type": "text", + "content": " mAP on LVIS rare classes) over the baseline method on LVIS (1203 classes). Besides, our CapDet also achieves state-of-the-art performance on dense captioning tasks, e.g., " + }, + { + "bbox": [ + 46, + 275, + 290, + 550 + ], + "type": "inline_equation", + "content": "15.44\\%" + }, + { + "bbox": [ + 46, + 275, + 290, + 550 + ], + "type": "text", + "content": " mAP on VG V1.2 and " + }, + { + "bbox": [ + 46, + 275, + 290, + 550 + ], + "type": "inline_equation", + "content": "13.98\\%" + }, + { + "bbox": [ + 46, + 275, + 290, + 550 + ], + "type": "text", + "content": " on the VG-COCO dataset." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 556, + 127, + 568 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 556, + 127, + 568 + ], + "spans": [ + { + "bbox": [ + 47, + 556, + 127, + 568 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 577, + 287, + 684 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 577, + 287, + 684 + ], + "spans": [ + { + "bbox": [ + 46, + 577, + 287, + 684 + ], + "type": "text", + "content": "Most state-of-the-art object detection methods [33, 34, 50] benefit from a large number of densely annotated detection datasets (e.g., COCO [27], Object365 [36], LVIS [12]). However, this closed-world setting results in the model only being able to predict categories that appear in the training set. Considering the ubiquity of new concepts in real-world scenes, it is very challenging to locate and identify these new visual concepts. This predictive ability of new concepts in open-world scenarios has very important research value" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 309, + 253, + 544, + 363 + ], + "blocks": [ + { + "bbox": [ + 309, + 253, + 544, + 363 + ], + "lines": [ + { + "bbox": [ + 309, + 253, + 544, + 363 + ], + "spans": [ + { + "bbox": [ + 309, + 253, + 544, + 363 + ], + "type": "image", + "image_path": "366721316af1b64ee03ded007cf3162f6fdcbe9b26aaedea1fbbff23ca67b393.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 309, + 365, + 544, + 485 + ], + "blocks": [ + { + "bbox": [ + 309, + 365, + 544, + 485 + ], + "lines": [ + { + "bbox": [ + 309, + 365, + 544, + 485 + ], + "spans": [ + { + "bbox": [ + 309, + 365, + 544, + 485 + ], + "type": "image", + "image_path": "49d4c0e56a0c81fa4d585e5abfa1ce329a3cc45826d6c5e3b428f87956621d58.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 486, + 545, + 585 + ], + "lines": [ + { + "bbox": [ + 304, + 486, + 545, + 585 + ], + "spans": [ + { + "bbox": [ + 304, + 486, + 545, + 585 + ], + "type": "text", + "content": "Figure 1. Comparison of the different model predictions under OWD, OVD, and our setting. (a) OWD methods [14, 18, 48] are not able to describe the detailed category of the detected unknown objects and (b) the performance of OVD methods [8, 12, 41] usually depends on the pre-defined category list during the inference. (c) With the unification of two pipelines of dense captioning and open-world detection pre-training, our CapDet can either predict under a given category list or directly generate the description of predicted bounding boxes." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 590, + 545, + 626 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 590, + 545, + 626 + ], + "spans": [ + { + "bbox": [ + 304, + 590, + 545, + 626 + ], + "type": "text", + "content": "in real-world applications such as object search [29, 30], instance registration [45], and human-object interaction modeling [10]." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "text", + "content": "Currently, the open world scenario mainly includes two tasks: open world object detection [18] (OWD) and open-vocabulary object detection [44] (OVD). Although the paradigms of OWD and OVD tasks are closer to the real world, the former cannot describe the specific concept of the detected unknown objects and requires a pre-defined category list during the inference. Specifically, as shown" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 494, + 35 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 494, + 35 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 494, + 35 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 693, + 143, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 693, + 143, + 713 + ], + "spans": [ + { + "bbox": [ + 61, + 693, + 143, + 713 + ], + "type": "text", + "content": "*Equal contribution. † Corresponding authors." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15233" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 288 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 288 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 288 + ], + "type": "text", + "content": "in Figure 1, previous OWD methods [14,18,48] would recognize new concepts not in the predefined category space as \"unknown\". Further, another line of task OVD requires the model to learn a limited base class and generalize to novel classes. Compared to the zero-shot object detection (ZSD) proposed by [32], OVD allows the model to use external knowledge, e.g., knowledge distillation from a largescale vision-language pre-trained model [8, 12], image-caption pairs [44], image classification data [49], grounding data [25, 41, 46]. With the external knowledge, OVD methods show a superior generalization capacity to detect the novel classes within a given category space. However, as shown in Figure 1, when given an incomplete category list, OVD can only predict the concepts that appear in the given category list, otherwise, there will be recognition errors, (i.e., as illustrated in Figure 1 (b), the OVD methods prone to predict the \"wall socket\" as \"remote\", since the latter is in the category list but not the former)." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 289, + 289, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 289, + 289, + 422 + ], + "spans": [ + { + "bbox": [ + 46, + 289, + 289, + 422 + ], + "type": "text", + "content": "Thus, under the OVD setting, we mainly face the following two challenges: (i) it is difficult to define a complete list of categories; (ii) low response values on rare categories often lead to recognition errors. This is mainly because we cannot exhaustively enumerate new objects in the real world, and secondly, it is difficult to collect enough samples for rare classes. However, the fact that rare objects in the real world, even some new objects that are unknown to humans, such as UFOs, do not prevent people from using natural language to describe it as \"a flying vehicle that looks like a Frisbee\"." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 423, + 289, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 423, + 289, + 616 + ], + "spans": [ + { + "bbox": [ + 46, + 423, + 289, + 616 + ], + "type": "text", + "content": "Therefore, based on the above observations, in this paper, we consider a new setting that is closer to the open world and real scenes, i.e., we expect the model to both detect and recognize concepts in a given category list, and to generate corresponding natural language descriptions for new concepts or rare categories of objects. Early dense captioning methods [9, 17] can locate salient regions in images and generate the region-grounded captions with natural language. Inspired by this, to address the challenges faced in the OVD setting, we propose to unify the two pipelines of dense captioning and open-world detection pre-training into one training framework, called CapDet. It empowers the model with the ability to both accurately detect and recognize common object categories and generate dense captions for unknown and rare categories by unifying the two training tasks." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 617, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 617, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 617, + 289, + 715 + ], + "type": "text", + "content": "Specifically, our CapDet constructs a unified data format for the dense captioning data and detection data. With the data unification, CapDet further adopts a unified pretraining paradigm including open-world object detection and dense captioning pre-training. For open-world detection pretraining, we treat the detection task as a semantic alignment task and adopt a dual encoder structure as [41] to locate and predict the given concepts list. The concepts" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 304, + 72, + 547, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 167 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 167 + ], + "type": "text", + "content": "list contains category names in detection data and region-grounded captions in dense captioning data. For dense captioning pretraining, CapDet proposes a dense captioning head to take the predicted proposals as input to generate the region-grounded captions. Due to the rich visual concepts in the dense captioning data, the integration of dense captioning tasks will in turn benefit the generalization of detection performance." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 167, + 547, + 276 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 167, + 547, + 276 + ], + "spans": [ + { + "bbox": [ + 304, + 167, + 547, + 276 + ], + "type": "text", + "content": "Our experiments show that the integration of few dense captioning data brings in large improvement in the object detection datasets LVIS, e.g., " + }, + { + "bbox": [ + 304, + 167, + 547, + 276 + ], + "type": "inline_equation", + "content": "+2.7\\%" + }, + { + "bbox": [ + 304, + 167, + 547, + 276 + ], + "type": "text", + "content": " mAP on LVIS. The unification of dense captioning and detection pre-training gains an additional " + }, + { + "bbox": [ + 304, + 167, + 547, + 276 + ], + "type": "inline_equation", + "content": "2.3\\%" + }, + { + "bbox": [ + 304, + 167, + 547, + 276 + ], + "type": "text", + "content": " increment on LVIS and " + }, + { + "bbox": [ + 304, + 167, + 547, + 276 + ], + "type": "inline_equation", + "content": "2.1\\%" + }, + { + "bbox": [ + 304, + 167, + 547, + 276 + ], + "type": "text", + "content": " increment on LVIS rare classes. Besides, our model also achieves state-of-the-art performance on dense captioning tasks. Note that our method is the first to unify dense captioning and open-world detection pretraining." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 318, + 276, + 513, + 287 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 276, + 513, + 287 + ], + "spans": [ + { + "bbox": [ + 318, + 276, + 513, + 287 + ], + "type": "text", + "content": "To summarize, our contributions are three folds:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 318, + 293, + 545, + 499 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 318, + 293, + 545, + 353 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 293, + 545, + 353 + ], + "spans": [ + { + "bbox": [ + 318, + 293, + 545, + 353 + ], + "type": "text", + "content": "- We propose a novel open-vocabulary object detection framework CapDet, which cannot only detect and recognize concepts in a given category list but also generate corresponding natural language descriptions for new concept objects." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 318, + 355, + 545, + 401 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 355, + 545, + 401 + ], + "spans": [ + { + "bbox": [ + 318, + 355, + 545, + 401 + ], + "type": "text", + "content": "- We propose to unify the two pipelines of dense captioning and open-world detection pre-training into one training framework. Both two pre-training tasks are beneficial to each other." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 318, + 405, + 545, + 499 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 405, + 545, + 499 + ], + "spans": [ + { + "bbox": [ + 318, + 405, + 545, + 499 + ], + "type": "text", + "content": "- Experiments show that by unified dense captioning task and detection task, our CapDet gains significant performance improvements on the open-vocabulary object detection task (e.g., " + }, + { + "bbox": [ + 318, + 405, + 545, + 499 + ], + "type": "inline_equation", + "content": "+3.3\\%" + }, + { + "bbox": [ + 318, + 405, + 545, + 499 + ], + "type": "text", + "content": " mAP on LVIS rare classes). Furthermore, our CapDet also achieves state-of-the-art performance on the dense captioning tasks, e.g., " + }, + { + "bbox": [ + 318, + 405, + 545, + 499 + ], + "type": "inline_equation", + "content": "15.44\\%" + }, + { + "bbox": [ + 318, + 405, + 545, + 499 + ], + "type": "text", + "content": " mAP on Visual Genome (VG) V1.2 and " + }, + { + "bbox": [ + 318, + 405, + 545, + 499 + ], + "type": "inline_equation", + "content": "13.98\\%" + }, + { + "bbox": [ + 318, + 405, + 545, + 499 + ], + "type": "text", + "content": " mAP on VG-COCO." + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 502, + 392, + 514 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 502, + 392, + 514 + ], + "spans": [ + { + "bbox": [ + 306, + 502, + 392, + 514 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 521, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 521, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 521, + 547, + 715 + ], + "type": "text", + "content": "Vision-Language Pre-training. Vision-Language Pretraining [6, 16, 31] as a scheme in the domains of natural language processing [1, 5] and computer vision [7] obtains continual attention currently. And it exhibits strong performance and generalization ability on various downstream vision and cross-modal tasks. Among them, CLIP [31] and ALIGN [16] as dual-stream methods utilize large-scale image-text pairs on the Internet by cross-modal contrastive learning to get excellent zero-shot classification ability. Single-stream methods [20, 24] unify visual and textual embeddings in a single transformer-based model, which can perform text generation tasks such as image caption and VQA. Some mixed architectures [23, 39] combine single-stream and dual-stream to explore a unified way of vision-language understanding and generation. However, these methods take low-resolution images as input and serve the" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "15234" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 288, + 109 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 288, + 109 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 288, + 109 + ], + "type": "text", + "content": "task of classification and retrieval. Those vision-language pre-training approaches can not be applied to pure computer vision task directly, i.e., object detection task." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 45, + 114, + 287, + 317 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 114, + 287, + 317 + ], + "spans": [ + { + "bbox": [ + 45, + 114, + 287, + 317 + ], + "type": "text", + "content": "Open World Object Detection / Open-Vocabulary Object Detection. Object detection is a core computer vision task, which aims at localizing objects using a bounding box and classifying them. The mature detection approaches which show great performance on supervised data include one-stage detectors (i.e.,YOLO [33],ATSS [47]) having a relatively high detection efficiency and two-stage detectors (i.e., Faster R-CNN [34],Mask R-CNN [15]) having good detection accuracy. However, how to generalize these methods to rare classes and novel concepts in the real world is a challenge. Currently, several object detection approaches for such open-world scenes have attracted extensive attention from academia and industry. These methods are divided into two tasks which are called open-world object detection and open-vocabulary object detection respectively depending on whether to detect the class of unknown classes." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 45, + 318, + 288, + 522 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 318, + 288, + 522 + ], + "spans": [ + { + "bbox": [ + 45, + 318, + 288, + 522 + ], + "type": "text", + "content": "For the OWD task, Zhao et al. [48] proposed a proposal advisor to assist in identifying unknown proposals without supervision and a class-specific expelling classifier to filter out confusing predictions. For the OVD task, GLIP [25] converts the detection data into grounding format and proposes a fusion module to learn semantic vision information in grounding data. K-Lite [38] reconstructs the input format of the data in GLIP from sequential to parallel and uses nouns hierarchy and definition to format text sequence. Det-CLIP [41] unifies detection, grounding, and image-text pair data in a paralleled formulation and constructs a concept dictionary to augment the text data, which strikes a balance between performance and efficiency. Differing from all these works, our CapDet can generate an open-set caption of each region proposal to cover situations where the semantics of new object instances are not in the given category list." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 534, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 534, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 534, + 287, + 714 + ], + "type": "text", + "content": "Dense Captioning. Dense captioning aims at generating detailed descriptions for local regions, which usually needs to locate visual regions with semantic information and generate captions for these regions. J. Johnson et al. [17] utilized a fully convolutional localization network to locate regions of interest (RoIs) and then describe them. Afterward, many methods [26, 43] based on Faster-RCNN [34] and LSTM [11] are proposed to do dense captioning. X. Li et al. [26] arrange RoI features as a sequence and put them into LSTM with the guidance of the region features to form the complementary object context features. This method also needs ground truth bounding boxes auxiliary tests to achieve good results. But limited by the forget gate mechanism of LSTM, the inputted sequence cannot be too long. Then, the transformer-based method TDC [37] is proposed" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "text", + "content": "to tackle the long sequence forgotten problem. Instead, our CapDet proposes a transformer-based caption head to generate a caption using a single-stage detector ATSS while simultaneously achieving open-world detection." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 306, + 132, + 362, + 144 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 132, + 362, + 144 + ], + "spans": [ + { + "bbox": [ + 306, + 132, + 362, + 144 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 152, + 545, + 260 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 152, + 545, + 260 + ], + "spans": [ + { + "bbox": [ + 304, + 152, + 545, + 260 + ], + "type": "text", + "content": "The overview of our proposed CapDet is shown in Figure 2. To construct a detector to either predict under a given category list or directly generate the concepts of predicted bounding boxes, we incorporate detection data and dense caption data together. In this section, we will present a unified data format for the detection data and dense caption data in Section 3.1, the model architecture and pre-training objectives for open-world object detection pre-training in Section 3.2 and dense captioning in Section 3.3." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 268, + 424, + 280 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 268, + 424, + 280 + ], + "spans": [ + { + "bbox": [ + 305, + 268, + 424, + 280 + ], + "type": "text", + "content": "3.1. Unified Formulation" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 287, + 545, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 287, + 545, + 453 + ], + "spans": [ + { + "bbox": [ + 304, + 287, + 545, + 453 + ], + "type": "text", + "content": "We defined a unified triplet-wise data format " + }, + { + "bbox": [ + 304, + 287, + 545, + 453 + ], + "type": "inline_equation", + "content": "(x, \\{\\mathbf{b}_i\\}_{i=1}^N, y_{i=1}^N)" + }, + { + "bbox": [ + 304, + 287, + 545, + 453 + ], + "type": "text", + "content": " for each sample from different sources. Specifically, " + }, + { + "bbox": [ + 304, + 287, + 545, + 453 + ], + "type": "inline_equation", + "content": "x \\in \\mathbb{R}^{3 \\times h \\times w}" + }, + { + "bbox": [ + 304, + 287, + 545, + 453 + ], + "type": "text", + "content": " is the input image, " + }, + { + "bbox": [ + 304, + 287, + 545, + 453 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{b}_i | \\mathbf{b}_i \\in \\mathbb{R}^4\\}_{i=1}^N" + }, + { + "bbox": [ + 304, + 287, + 545, + 453 + ], + "type": "text", + "content": " denotes the bounding boxes coordinates for each region of the image, and the " + }, + { + "bbox": [ + 304, + 287, + 545, + 453 + ], + "type": "inline_equation", + "content": "y_{i=1}^N" + }, + { + "bbox": [ + 304, + 287, + 545, + 453 + ], + "type": "text", + "content": " represents the concepts of the corresponding boxes. " + }, + { + "bbox": [ + 304, + 287, + 545, + 453 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 287, + 545, + 453 + ], + "type": "text", + "content": " denotes the number of regions. A concept " + }, + { + "bbox": [ + 304, + 287, + 545, + 453 + ], + "type": "inline_equation", + "content": "y_i" + }, + { + "bbox": [ + 304, + 287, + 545, + 453 + ], + "type": "text", + "content": " formatted as a sentence contains the category and textual description of the corresponding region. In detection data, a concept " + }, + { + "bbox": [ + 304, + 287, + 545, + 453 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 304, + 287, + 545, + 453 + ], + "type": "text", + "content": " consists of the category name and the corresponding definition from the concept dictionary [41], while " + }, + { + "bbox": [ + 304, + 287, + 545, + 453 + ], + "type": "inline_equation", + "content": "y_i" + }, + { + "bbox": [ + 304, + 287, + 545, + 453 + ], + "type": "text", + "content": " represents the region-grounded caption in dense caption data. For example, for an image " + }, + { + "bbox": [ + 304, + 287, + 545, + 453 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 304, + 287, + 545, + 453 + ], + "type": "text", + "content": " in detection data, " + }, + { + "bbox": [ + 304, + 287, + 545, + 453 + ], + "type": "inline_equation", + "content": "y_i" + }, + { + "bbox": [ + 304, + 287, + 545, + 453 + ], + "type": "text", + "content": " can be:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 355, + 455, + 494, + 468 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 355, + 455, + 494, + 468 + ], + "spans": [ + { + "bbox": [ + 355, + 455, + 494, + 468 + ], + "type": "interline_equation", + "content": "y _ {i} = \\text {\" p e r s o n , a h u m a n b e i n g . \"}", + "image_path": "920a3ded9642f4be6db68ca593943c970831702f36942f45a708ba5b57f4597e.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 474, + 524, + 487 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 474, + 524, + 487 + ], + "spans": [ + { + "bbox": [ + 306, + 474, + 524, + 487 + ], + "type": "text", + "content": "For an image " + }, + { + "bbox": [ + 306, + 474, + 524, + 487 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 306, + 474, + 524, + 487 + ], + "type": "text", + "content": " from dense captioning data, " + }, + { + "bbox": [ + 306, + 474, + 524, + 487 + ], + "type": "inline_equation", + "content": "y_{i}" + }, + { + "bbox": [ + 306, + 474, + 524, + 487 + ], + "type": "text", + "content": " can be:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 361, + 496, + 489, + 509 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 361, + 496, + 489, + 509 + ], + "spans": [ + { + "bbox": [ + 361, + 496, + 489, + 509 + ], + "type": "interline_equation", + "content": "y _ {i} = \\text {\" a n o u t l e t o n t h e w a l l .}", + "image_path": "cb5d4554c719de831ae061cfafbb80898de75463f6e809b0244285e659193ce2.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 518, + 545, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 518, + 545, + 579 + ], + "spans": [ + { + "bbox": [ + 304, + 518, + 545, + 579 + ], + "type": "text", + "content": "With the triplet, we can learn a unified image-text alignment objective on the detection data and the dense captioning data. The unified formulation also ensures the joint training of open-world object detection pre-training and dense captioning." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 586, + 529, + 601 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 586, + 529, + 601 + ], + "spans": [ + { + "bbox": [ + 306, + 586, + 529, + 601 + ], + "type": "text", + "content": "3.2. Open-World Object Detection Pre-training" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 605, + 545, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 545, + 715 + ], + "type": "text", + "content": "Based on the unified formulation of detection data and dense captioning data, we regard the captions of regions in dense captioning data as a kind of category and utilize two different sources of data for the open-world object detection pre-training. Compared with the limited class list of detection data, dense caption data contains richer concepts and more semantic information than class names of individual regions. On the other hand, localization and recognition are two essential tasks of object detection. Traditional object" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15235" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 85, + 72, + 509, + 342 + ], + "blocks": [ + { + "bbox": [ + 85, + 72, + 509, + 342 + ], + "lines": [ + { + "bbox": [ + 85, + 72, + 509, + 342 + ], + "spans": [ + { + "bbox": [ + 85, + 72, + 509, + 342 + ], + "type": "image", + "image_path": "3edf93e562aaab7bfa54789f47818bdd10fd383c264bb50247aad7f92d884362.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 350, + 547, + 417 + ], + "lines": [ + { + "bbox": [ + 46, + 350, + 547, + 417 + ], + "spans": [ + { + "bbox": [ + 46, + 350, + 547, + 417 + ], + "type": "text", + "content": "Figure 2. The overall architecture of CapDet. The training paradigm of CapDet contains open-world object detection pre-training and dense captioning. In detection, CapDet contains a dual vision-language encoder. The image encoder generates region embeddings from detection and dense captioning data. The regression loss and centerness loss are introduced to regress the locations. The text encoder takes the category concepts as input to generate the embeddings from the [EOS] token. Then we treat the detection task as a matching task and adopt an alignment loss for the category embeddings and region embeddings. In dense captioning, an additional dense captioning head is proposed to take the region embeddings as input and generate the textual captions for corresponding regions with natural language." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 426, + 288, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 426, + 288, + 496 + ], + "spans": [ + { + "bbox": [ + 46, + 426, + 288, + 496 + ], + "type": "text", + "content": "detection always focuses on the salient objects in the image. While the dense captioning data contains lots of annotations which are just parts of an object, e.g., an ear of an elephant, it is not suitable to learn those annotations for the localization task. Therefore, we only calculate the localization loss on detection data." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 498, + 287, + 618 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 498, + 287, + 618 + ], + "spans": [ + { + "bbox": [ + 46, + 498, + 287, + 618 + ], + "type": "text", + "content": "As shown in Figure 2, CapDet predicts the regions and treats the recognition task as a region-category matching task. For efficient learning on the matching task, we adopt the negative sampling proposed by [41] to provide negative concepts to enlarge the concept space in a batch. Specifically, for each iteration, we randomly sample a negative concept set and add to the positive concept set (N samples) in a batch to obtain the final concept set " + }, + { + "bbox": [ + 46, + 498, + 287, + 618 + ], + "type": "inline_equation", + "content": "y_{i = 1}^{M}" + }, + { + "bbox": [ + 46, + 498, + 287, + 618 + ], + "type": "text", + "content": ", where M represents the sum of the number of positive and negative samples. Finally, we format the triplet to " + }, + { + "bbox": [ + 46, + 498, + 287, + 618 + ], + "type": "inline_equation", + "content": "(x,\\{\\mathbf{b}_i\\}_{i = 1}^N,y_{i = 1}^M)" + }, + { + "bbox": [ + 46, + 498, + 287, + 618 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 618, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 618, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 618, + 288, + 714 + ], + "type": "text", + "content": "CapDet contains a dual vision-language encoder and takes the triplet " + }, + { + "bbox": [ + 46, + 618, + 288, + 714 + ], + "type": "inline_equation", + "content": "(x, \\{\\mathbf{b}_i\\}_{i=1}^N, y_{i=1}^M)" + }, + { + "bbox": [ + 46, + 618, + 288, + 714 + ], + "type": "text", + "content": " as input. The image encoder " + }, + { + "bbox": [ + 46, + 618, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\Phi_v" + }, + { + "bbox": [ + 46, + 618, + 288, + 714 + ], + "type": "text", + "content": " is an object detector that can predict the bounding boxes of regions from the input image " + }, + { + "bbox": [ + 46, + 618, + 288, + 714 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 46, + 618, + 288, + 714 + ], + "type": "text", + "content": " and output the region features " + }, + { + "bbox": [ + 46, + 618, + 288, + 714 + ], + "type": "inline_equation", + "content": "O \\in \\mathbb{R}^{K \\times D}" + }, + { + "bbox": [ + 46, + 618, + 288, + 714 + ], + "type": "text", + "content": ". The text encoder " + }, + { + "bbox": [ + 46, + 618, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\Phi_l" + }, + { + "bbox": [ + 46, + 618, + 288, + 714 + ], + "type": "text", + "content": " takes the concept set " + }, + { + "bbox": [ + 46, + 618, + 288, + 714 + ], + "type": "inline_equation", + "content": "y_{i=1}^M" + }, + { + "bbox": [ + 46, + 618, + 288, + 714 + ], + "type": "text", + "content": " as input and obtains the text embeddings " + }, + { + "bbox": [ + 46, + 618, + 288, + 714 + ], + "type": "inline_equation", + "content": "W \\in \\mathbb{R}^{M \\times D}" + }, + { + "bbox": [ + 46, + 618, + 288, + 714 + ], + "type": "text", + "content": " from the special token [EOS] concatenated with the text input. " + }, + { + "bbox": [ + 46, + 618, + 288, + 714 + ], + "type": "inline_equation", + "content": "K, D" + }, + { + "bbox": [ + 46, + 618, + 288, + 714 + ], + "type": "text", + "content": " denotes the number of predicted" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 305, + 426, + 545, + 450 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 426, + 545, + 450 + ], + "spans": [ + { + "bbox": [ + 305, + 426, + 545, + 450 + ], + "type": "text", + "content": "regions and region feature dimensions. The alignment score matrix " + }, + { + "bbox": [ + 305, + 426, + 545, + 450 + ], + "type": "inline_equation", + "content": "S \\in \\mathbb{R}^{K \\times M}" + }, + { + "bbox": [ + 305, + 426, + 545, + 450 + ], + "type": "text", + "content": " of regions and texts is calculated by:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 343, + 460, + 545, + 475 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 343, + 460, + 545, + 475 + ], + "spans": [ + { + "bbox": [ + 343, + 460, + 545, + 475 + ], + "type": "interline_equation", + "content": "O = \\Phi_ {v} (x), W = \\Phi_ {l} \\left(y _ {i = 1} ^ {M}\\right), S = O W ^ {T} \\tag {1}", + "image_path": "5ea342af178f8a3a2e5702b62bde1e9b5be6e04d5819f3dcb7f300772337a5d3.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 487, + 546, + 619 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 487, + 546, + 619 + ], + "spans": [ + { + "bbox": [ + 304, + 487, + 546, + 619 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 487, + 546, + 619 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 304, + 487, + 546, + 619 + ], + "type": "text", + "content": " denotes the transpose operation. A ground-truth alignment matrix " + }, + { + "bbox": [ + 304, + 487, + 546, + 619 + ], + "type": "inline_equation", + "content": "G \\in \\{0,1\\}^{K \\times M}" + }, + { + "bbox": [ + 304, + 487, + 546, + 619 + ], + "type": "text", + "content": " is constructed to indicate the matching relation of regions and concepts. The alignment loss " + }, + { + "bbox": [ + 304, + 487, + 546, + 619 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\text{align}}" + }, + { + "bbox": [ + 304, + 487, + 546, + 619 + ], + "type": "text", + "content": " is calculated by the predicted alignment scores of regions " + }, + { + "bbox": [ + 304, + 487, + 546, + 619 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 304, + 487, + 546, + 619 + ], + "type": "text", + "content": " and the ground-truth alignment matrix " + }, + { + "bbox": [ + 304, + 487, + 546, + 619 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 304, + 487, + 546, + 619 + ], + "type": "text", + "content": ". Following [25, 41], we adopt the ATSS [47] detector as an image encoder, and " + }, + { + "bbox": [ + 304, + 487, + 546, + 619 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\text{align}}" + }, + { + "bbox": [ + 304, + 487, + 546, + 619 + ], + "type": "text", + "content": " is typically a sigmoid focal loss. As a one-stage detector, the localization loss contains centeredness loss " + }, + { + "bbox": [ + 304, + 487, + 546, + 619 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\text{cen}}" + }, + { + "bbox": [ + 304, + 487, + 546, + 619 + ], + "type": "text", + "content": " and bounding box regression loss " + }, + { + "bbox": [ + 304, + 487, + 546, + 619 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\text{reg}}" + }, + { + "bbox": [ + 304, + 487, + 546, + 619 + ], + "type": "text", + "content": ". The training objective of detection pre-training can be written as:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 308, + 629, + 545, + 670 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 629, + 545, + 670 + ], + "spans": [ + { + "bbox": [ + 308, + 629, + 545, + 670 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\left\\{ \\begin{array}{l l} \\mathcal {L} _ {\\text {a l i g n}} + \\alpha \\mathcal {L} _ {\\text {r e g}} + \\beta \\mathcal {L} _ {\\text {c e n t e r}}, & \\text {f o r d e t e c t i o n} \\\\ \\mathcal {L} _ {\\text {a l i g n}}, & \\text {f o r d e n s e c a p t i o n i n g} \\end{array} \\right. \\tag {2}", + "image_path": "38e8a832ab0529bc01ccabf96bceb8535a600bdb25436bbe9e5793d45ec438cb.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 672, + 545, + 708 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 672, + 545, + 708 + ], + "spans": [ + { + "bbox": [ + 304, + 672, + 545, + 708 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 672, + 545, + 708 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 304, + 672, + 545, + 708 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 672, + 545, + 708 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 304, + 672, + 545, + 708 + ], + "type": "text", + "content": " denote the weights for the centerness loss " + }, + { + "bbox": [ + 304, + 672, + 545, + 708 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{cen}" + }, + { + "bbox": [ + 304, + 672, + 545, + 708 + ], + "type": "text", + "content": " and box regression loss " + }, + { + "bbox": [ + 304, + 672, + 545, + 708 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{reg}" + }, + { + "bbox": [ + 304, + 672, + 545, + 708 + ], + "type": "text", + "content": ", respectively. The " + }, + { + "bbox": [ + 304, + 672, + 545, + 708 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{cen}" + }, + { + "bbox": [ + 304, + 672, + 545, + 708 + ], + "type": "text", + "content": " is the sigmoid loss and the " + }, + { + "bbox": [ + 304, + 672, + 545, + 708 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{reg}" + }, + { + "bbox": [ + 304, + 672, + 545, + 708 + ], + "type": "text", + "content": " is the GIOU loss [35]." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "15236" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 47, + 71, + 550, + 290 + ], + "blocks": [ + { + "bbox": [ + 47, + 71, + 550, + 290 + ], + "lines": [ + { + "bbox": [ + 47, + 71, + 550, + 290 + ], + "spans": [ + { + "bbox": [ + 47, + 71, + 550, + 290 + ], + "type": "table", + "html": "
MODELBACKBONEPRE-TRAIN DATAIMAGES NUMBERAPLVIS APr / APc / APf
MASK-RCNN [15]SWIN-TLVIS0.1M34.119.1 / 34.0 / 37.0
ATSS [47]SWIN-TLVIS0.1M33.619.7 / 32.4 / 37.2
ATSS [47]SWIN-LLVIS0.1M43.930.6 / 43.7 / 46.3
MDETR [19]RN101GOLDG+0.77M24.220.9 / 24.3 / 24.2
GLIP-T(A) [25]SWIN-T+DH+FO3650.66M18.514.2 / 13.9 / 23.4
GLIP-T(C) [25]SWIN-T+DH+FO365,GOLDG1.43M24.917.7 / 19.5 / 31.0
GLIP-T [25]SWIN-T+DH+FO365,GOLDG,CAP4M5.43M26.020.8 / 21.4 / 31.0
K-LITE [38]SWIN-TO3650.66M21.314.8 / 18.6 / 24.8
K-LITE [38]SWIN-TO365,GOLDG1.43M26.117.2 / 24.6 / 29.0
GLIPv2-T [46]SWIN-T+DH+FO365,GOLDG,CAP4M5.43M29.0- / - / -
DETCLIP-T(A) [41]SWIN-TO3650.66M28.826.0 / 28.0 / 30.0
DETCLIP-T(B) [41]SWIN-TO365, GOLDG1.43M34.426.9 / 33.9 / 36.3
DETCLIP-T(C)* [41]SWIN-TO365, VG0.73M31.527.5 / 30.6 / 33.0
CAPDET (OURS)SWIN-TO365, VG0.73M33.829.6 / 32.8 / 35.5
", + "image_path": "4508f6c0ede534f23083982b9625d33c5dfd9b80791abbebdd061d9f29e03eeb.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 299, + 546, + 333 + ], + "lines": [ + { + "bbox": [ + 46, + 299, + 546, + 333 + ], + "spans": [ + { + "bbox": [ + 46, + 299, + 546, + 333 + ], + "type": "text", + "content": "Table 1. Zero-shot performance on LVIS [13] MiniVal5k datasets. " + }, + { + "bbox": [ + 46, + 299, + 546, + 333 + ], + "type": "inline_equation", + "content": "\\mathrm{AP}_r / \\mathrm{AP}_c / \\mathrm{AP}_f" + }, + { + "bbox": [ + 46, + 299, + 546, + 333 + ], + "type": "text", + "content": " indicate the AP values for rare, common, and frequent categories, respectively. \"DH\" and \"F\" in GLIP [25] baselines stand for the dynamic head [3] and cross-modal fusion, respectively. Baselines with * are implemented with our code base. GoldG+ denotes the GoldG plus the COCO [27] caption dataset." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 47, + 363, + 153, + 376 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 363, + 153, + 376 + ], + "spans": [ + { + "bbox": [ + 47, + 363, + 153, + 376 + ], + "type": "text", + "content": "3.3. Dense Captioning" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 381, + 287, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 381, + 287, + 548 + ], + "spans": [ + { + "bbox": [ + 46, + 381, + 287, + 548 + ], + "type": "text", + "content": "The open-world object detection pre-training ensures CapDet gains the capacity to detect under given an arbitrary category list. However, when the given category list is not complete enough to cover the potential classes on a new domain data, the detector will perform worse on the categories which are not in the given list. Considering such limitation, we propose a dense captioning head " + }, + { + "bbox": [ + 46, + 381, + 287, + 548 + ], + "type": "inline_equation", + "content": "\\Phi_C" + }, + { + "bbox": [ + 46, + 381, + 287, + 548 + ], + "type": "text", + "content": " to generate semantically rich concepts with natural language for the predicted proposals. In the dense captioning task, the model receives an image and produces a set of regions and the corresponding captions. The dense captioning head is a cross-modal decoder that takes the c predicted regions features " + }, + { + "bbox": [ + 46, + 381, + 287, + 548 + ], + "type": "inline_equation", + "content": "O" + }, + { + "bbox": [ + 46, + 381, + 287, + 548 + ], + "type": "text", + "content": " generated by the image encoder as input. The captioning (i.e., language modeling) loss is calculated by:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 91, + 556, + 287, + 571 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 556, + 287, + 571 + ], + "spans": [ + { + "bbox": [ + 91, + 556, + 287, + 571 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {c a p} = - \\log p \\left(y _ {i t} \\mid \\Phi_ {c} \\left(y _ {i (\\tau < t)}, O _ {i}\\right)\\right), \\tag {3}", + "image_path": "3ed9ff202e7c979b73a418ce4f7923b14fac46ec9728c39054ed0981dc93f89e.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 578, + 287, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 578, + 287, + 613 + ], + "spans": [ + { + "bbox": [ + 46, + 578, + 287, + 613 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 578, + 287, + 613 + ], + "type": "inline_equation", + "content": "y_{it}" + }, + { + "bbox": [ + 46, + 578, + 287, + 613 + ], + "type": "text", + "content": " means the " + }, + { + "bbox": [ + 46, + 578, + 287, + 613 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 578, + 287, + 613 + ], + "type": "text", + "content": " token in caption " + }, + { + "bbox": [ + 46, + 578, + 287, + 613 + ], + "type": "inline_equation", + "content": "y_{i}" + }, + { + "bbox": [ + 46, + 578, + 287, + 613 + ], + "type": "text", + "content": " corresponding to region feature " + }, + { + "bbox": [ + 46, + 578, + 287, + 613 + ], + "type": "inline_equation", + "content": "O_{i}" + }, + { + "bbox": [ + 46, + 578, + 287, + 613 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 578, + 287, + 613 + ], + "type": "inline_equation", + "content": "y_{i(\\tau < t)}" + }, + { + "bbox": [ + 46, + 578, + 287, + 613 + ], + "type": "text", + "content": " means tokens before " + }, + { + "bbox": [ + 46, + 578, + 287, + 613 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 578, + 287, + 613 + ], + "type": "text", + "content": " in caption " + }, + { + "bbox": [ + 46, + 578, + 287, + 613 + ], + "type": "inline_equation", + "content": "y_{i}" + }, + { + "bbox": [ + 46, + 578, + 287, + 613 + ], + "type": "text", + "content": ". The overall pre-training loss can be written as:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 117, + 621, + 287, + 635 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 117, + 621, + 287, + 635 + ], + "spans": [ + { + "bbox": [ + 117, + 621, + 287, + 635 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = w _ {d} \\mathcal {L} _ {\\text {d e t}} + w _ {c} \\mathcal {L} _ {\\text {c a p}}, \\tag {4}", + "image_path": "9ecfee54a5e516ab4ddb5d7a20b2dbf6374bc845149756ec32c9b1050b438fc6.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 642, + 286, + 654 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 642, + 286, + 654 + ], + "spans": [ + { + "bbox": [ + 47, + 642, + 286, + 654 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 642, + 286, + 654 + ], + "type": "inline_equation", + "content": "w_{d}, w_{c}" + }, + { + "bbox": [ + 47, + 642, + 286, + 654 + ], + "type": "text", + "content": " denote the weighting factor of " + }, + { + "bbox": [ + 47, + 642, + 286, + 654 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{det}" + }, + { + "bbox": [ + 47, + 642, + 286, + 654 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 642, + 286, + 654 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{cap}" + }, + { + "bbox": [ + 47, + 642, + 286, + 654 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 654, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 654, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 654, + 287, + 714 + ], + "type": "text", + "content": "To minimize the gap in the type of bounding boxes between the detection data and dense captioning data, we propose a simple way to transform our detector as a class-agnostic detector and only select the top " + }, + { + "bbox": [ + 46, + 654, + 287, + 714 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 654, + 287, + 714 + ], + "type": "text", + "content": " regions based on the centeredness scores to adapt to the dense captioning" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 363, + 545, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 363, + 545, + 472 + ], + "spans": [ + { + "bbox": [ + 304, + 363, + 545, + 472 + ], + "type": "text", + "content": "task. We can fine-tune our CapDet on the dense captioning data to achieve better performance. Specifically, we propose \"object\" as the foreground concept and \"background\" as the background concept. The text encoder " + }, + { + "bbox": [ + 304, + 363, + 545, + 472 + ], + "type": "inline_equation", + "content": "\\Phi_l" + }, + { + "bbox": [ + 304, + 363, + 545, + 472 + ], + "type": "text", + "content": " outputs the concept embeddings " + }, + { + "bbox": [ + 304, + 363, + 545, + 472 + ], + "type": "inline_equation", + "content": "W' \\in \\mathbb{R}^{2 \\times D}" + }, + { + "bbox": [ + 304, + 363, + 545, + 472 + ], + "type": "text", + "content": ". Then the alignment scores " + }, + { + "bbox": [ + 304, + 363, + 545, + 472 + ], + "type": "inline_equation", + "content": "S' \\in \\mathbb{R}^{K \\times 2}" + }, + { + "bbox": [ + 304, + 363, + 545, + 472 + ], + "type": "text", + "content": " is calculated by Eqn. 1. The captioning head takes the top " + }, + { + "bbox": [ + 304, + 363, + 545, + 472 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 363, + 545, + 472 + ], + "type": "text", + "content": " most confident proposal embeddings based on centeredness scores as input to predict the region-grounded captions." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 487, + 382, + 502 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 487, + 382, + 502 + ], + "spans": [ + { + "bbox": [ + 306, + 487, + 382, + 502 + ], + "type": "text", + "content": "4. Experiment" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "type": "text", + "content": "Implementation Details. For the image encoder, we adopt the Swin-T backbone proposed in SwinTransformer [28] which is pre-trained on ImageNet-1K [4]. We use 12 layers 8 heads transformer as our text encoder and load a base model checkpoint released by FILIP [42], in order to make a fair comparison with DetCLIP [41]. The structure of the dense captioning head is consistent with that in the text encoder but trained from scratch for a fair comparison. We employ AdamW [21] optimizer and set the batch size to 32. The learning rate is set to " + }, + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "type": "inline_equation", + "content": "1.4 \\times 10^{-4}" + }, + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "type": "text", + "content": " for the parameters of the image encoder and detection head, and " + }, + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "type": "inline_equation", + "content": "1.4 \\times 10^{-5}" + }, + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "type": "text", + "content": " for the text encoder and dense captioning head. When fine-tuning the VG dataset to do the dense captioning task, we set the learning rate to " + }, + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "type": "inline_equation", + "content": "1.4 \\times 10^{-4}" + }, + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "type": "text", + "content": ". Without otherwise specified, all models are trained with 12 epochs and the learning rate is decayed with a factor of 0.1 at the 8-th and the 11-th epoch. The context token" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15237" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 72, + 545, + 254 + ], + "blocks": [ + { + "bbox": [ + 50, + 72, + 545, + 254 + ], + "lines": [ + { + "bbox": [ + 50, + 72, + 545, + 254 + ], + "spans": [ + { + "bbox": [ + 50, + 72, + 545, + 254 + ], + "type": "image", + "image_path": "8f647bdc722a93a11d7ce02114dbac662df17023a0fec08ba576c8b005716469.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 273, + 546, + 297 + ], + "lines": [ + { + "bbox": [ + 46, + 273, + 546, + 297 + ], + "spans": [ + { + "bbox": [ + 46, + 273, + 546, + 297 + ], + "type": "text", + "content": "Figure 3. Qualitative visualizations between GLIP-T, DetCLIP-T(C) and CapDet. From top to down, the three rows of images show the LVIS zero-shot detection results of GLIP-T, DetCLIP-T(C), and CapDet respectively. All models are pre-trained on O365 and VG." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 316, + 288, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 316, + 288, + 376 + ], + "spans": [ + { + "bbox": [ + 46, + 316, + 288, + 376 + ], + "type": "text", + "content": "length for input text is set to 20. We set the number of input captions to 150, and the number of the region features N is determined by the feature map. The loss weight factor " + }, + { + "bbox": [ + 46, + 316, + 288, + 376 + ], + "type": "inline_equation", + "content": "w_{c}" + }, + { + "bbox": [ + 46, + 316, + 288, + 376 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 316, + 288, + 376 + ], + "type": "inline_equation", + "content": "w_{d}" + }, + { + "bbox": [ + 46, + 316, + 288, + 376 + ], + "type": "text", + "content": " are both set to 1.0. We build our model on MMDetection [2] code base." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 395, + 287, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 395, + 287, + 635 + ], + "spans": [ + { + "bbox": [ + 46, + 395, + 287, + 635 + ], + "type": "text", + "content": "Dataset. Our CapDet is trained with two types of data, including detection data and caption data. Following DetCLIP [41], we use Object365 [36] (it will be abbreviated as O365 in the following paper) as detection data, and sample 0.66M data from O365 v2 for training. Following GLIP [25] and DetCLIP [41], LVIS [13] MiniVal5k (defined in [19]) which has 5000 images is used for detection evaluation. Moreover, we remove the training samples contained in the LVIS dataset for fair zero-transfer evaluation. For dense captioning data, we mainly conduct our experiments on VG [22] V1.2 and VG-COCO (defined in [37]). Following [37], we allocate 77398 images for training and 5000 images for validation and testing on VG. As demonstrated in [22], the ground-truth bounding boxes of VG are much denser than the other object detection datasets, i.e., the average number of per sample in MS COCO [27] is only 7.1 vs. 35.4 in VG. Then an intersection of VG V1.2 and MS COCO is proposed by [37] and is denoted as VG-COCO, which has 38080 images for training, 2489 for validation, and 2476 for testing." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 653, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 287, + 715 + ], + "type": "text", + "content": "Benchmark Settings. We mainly evaluate our method on open-vocabulary object detection and dense captioning task. For open-vocabulary object detection, we evaluate the direct domain transfer on LVIS [13] which contains 1203 categories. Following [25, 41], we metric the zero-shot" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 316, + 547, + 497 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 316, + 547, + 497 + ], + "spans": [ + { + "bbox": [ + 304, + 316, + 547, + 497 + ], + "type": "text", + "content": "detection performance by the Average Precision (AP) on a 5k subset. The annotations of LVIS data are split into three folds, i.e., rare, common, and frequency, based on the number of categories. Since there is almost no overlap between the rare classes and the classes of training dataset Objects365 [36], the AP of the rare classes shows a valuable zero-shot detection performance. For the dense captioning task, we follow the setting of [17] to evaluate the VG and VG-COCO. The evaluation metric we adopt is the mean Average Precision proposed by [17] which is calculated across a range of thresholds for both localization and language accuracy, i.e., the intersection over union (IOU) thresholds .3, .4, .5, .6, .7 are used for localization and the METEOR score' thresholds 0, .05, .1, .15, .2, .25 is adopted for evaluating the language generation." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 305, + 502, + 470, + 514 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 502, + 470, + 514 + ], + "spans": [ + { + "bbox": [ + 305, + 502, + 470, + 514 + ], + "type": "text", + "content": "4.1. Open-world Detection Results" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 520, + 545, + 675 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 520, + 545, + 675 + ], + "spans": [ + { + "bbox": [ + 304, + 520, + 545, + 675 + ], + "type": "text", + "content": "Table 1 shows our zero-shot object detection performance on LVIS. We mainly train our CapDet with the backbone Swin-T [28] on the detection data Objects365 [36] and dense captioning data (VG [22]). Since DetCLIP does not report the performance on O365 and VG, we train DetCLIP on the two datasets under the same settings and denote it as DetCLIP-T(C) for a fair comparison. Comparing the 11th row and 12th row, our CapDet outperforms DetCLIP-T(C) on the same data scale and backbone with an extra simple caption head. Moreover, our model's zero-shot performance even surpasses the fully-supervised model with the same backbone by a large margin on rare classes, i.e., CapDet outperforms ATSS by " + }, + { + "bbox": [ + 304, + 520, + 545, + 675 + ], + "type": "inline_equation", + "content": "9.9\\%" + }, + { + "bbox": [ + 304, + 520, + 545, + 675 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 677, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 547, + 713 + ], + "type": "text", + "content": "Qualitative Visualizations Figure. 3 illustrates the detection results on LVIS [22] dataset from GLIP-T, DetCLIP-T(C), and CapDet. All three models are trained on O365" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15238" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 53, + 72, + 541, + 363 + ], + "blocks": [ + { + "bbox": [ + 53, + 72, + 541, + 363 + ], + "lines": [ + { + "bbox": [ + 53, + 72, + 541, + 363 + ], + "spans": [ + { + "bbox": [ + 53, + 72, + 541, + 363 + ], + "type": "image", + "image_path": "9ed982972812d1c237f8b600b279b11086cb1ea5113a52fc4e174c9aae3b8d5b.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 373, + 546, + 396 + ], + "lines": [ + { + "bbox": [ + 46, + 373, + 546, + 396 + ], + "spans": [ + { + "bbox": [ + 46, + 373, + 546, + 396 + ], + "type": "text", + "content": "Figure 4. Qualitative visualizations between JIVC and CapDet. \"w/o ft\" means do caption without finetune, while \"w/ ft\" means with finetune." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 400, + 288, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 400, + 288, + 472 + ], + "spans": [ + { + "bbox": [ + 46, + 400, + 288, + 472 + ], + "type": "text", + "content": "and VG, and details are given in Section 4.3. Given a category list, the rare classes are detected more precisely by our CapDet, e.g., \"kitchen table\" in the first column, \"horse buggy\" in the third column, and \"fishbowl\" in the sixth column that our model CapDet detects correctly but the other two not." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 480, + 190, + 493 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 480, + 190, + 493 + ], + "spans": [ + { + "bbox": [ + 47, + 480, + 190, + 493 + ], + "type": "text", + "content": "4.2. Dense Captioning Results" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 498, + 287, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 498, + 287, + 653 + ], + "spans": [ + { + "bbox": [ + 46, + 498, + 287, + 653 + ], + "type": "text", + "content": "Due to the target bounding boxes in dense captioning data containing lots of local structures of objects and being much denser than the bounding boxes in object detection data, we do not regress the bounding box in the pre-training stage. The previous works directly train on the dense captioning data and generate captions on the top " + }, + { + "bbox": [ + 46, + 498, + 287, + 653 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 498, + 287, + 653 + ], + "type": "text", + "content": " proposals ranking by a confidence score. When fine-tuning our model on the VG dataset for the dense captioning tasks, we transform our CapDet into a class-agnostic detector. Specifically, we propose \"object\" as the foreground concept and \"background\" as the background concept for computing alignment scores. The scores are used as proposal confidences to predict the region-grounded captions." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 654, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 654, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 654, + 288, + 715 + ], + "type": "text", + "content": "Table 2 and Table 3 show CapDet significantly outperforms the latest work TDC [37] by " + }, + { + "bbox": [ + 46, + 654, + 288, + 715 + ], + "type": "inline_equation", + "content": "2.5\\%" + }, + { + "bbox": [ + 46, + 654, + 288, + 715 + ], + "type": "text", + "content": " on mAP on VG and TDC+ROCSU [37] by " + }, + { + "bbox": [ + 46, + 654, + 288, + 715 + ], + "type": "inline_equation", + "content": "2.08\\%" + }, + { + "bbox": [ + 46, + 654, + 288, + 715 + ], + "type": "text", + "content": ", respectively. It is worth noticing that, even against given the ground-truth bounding boxes with the previous method COCG [26] denoted" + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 365, + 399, + 488, + 525 + ], + "blocks": [ + { + "bbox": [ + 365, + 399, + 488, + 525 + ], + "lines": [ + { + "bbox": [ + 365, + 399, + 488, + 525 + ], + "spans": [ + { + "bbox": [ + 365, + 399, + 488, + 525 + ], + "type": "table", + "html": "
MethodmAP(%)
FCLN [17]5.16
JIVC [40]9.96
ImgG [26]9.68
COCD [26]9.75
COCG [26]10.39
CAG-Net [43]10.51
TDC [37]11.90
CapDet (Ours)15.44
", + "image_path": "8b07a72f56d866358c042469446500aeb6990624de40394e9abda968c0bfb63c.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 533, + 545, + 555 + ], + "lines": [ + { + "bbox": [ + 305, + 533, + 545, + 555 + ], + "spans": [ + { + "bbox": [ + 305, + 533, + 545, + 555 + ], + "type": "text", + "content": "Table 2. Comparison of mAP (%) performance on dense captioning benchmark on the VG V1.2 dataset." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 568, + 545, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 568, + 545, + 616 + ], + "spans": [ + { + "bbox": [ + 304, + 568, + 545, + 616 + ], + "type": "text", + "content": "as COCOG>, our CapDet still gains a " + }, + { + "bbox": [ + 304, + 568, + 545, + 616 + ], + "type": "inline_equation", + "content": "43.80\\%" + }, + { + "bbox": [ + 304, + 568, + 545, + 616 + ], + "type": "text", + "content": " mAP increase and achieves state-of-the-art. One important reason is that the excellent detection performance of our model assists the localization ability of dense captioning tasks." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 617, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 617, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 617, + 546, + 713 + ], + "type": "text", + "content": "Qualitative Visualizations. Figure 4 shows a qualitative visualization comparison between JIVC [40] and our CapDet. The three image rows from top to bottom are the visualization of JIVC, CapDet without fine-tuning, and CapDet with finetuning. In the second row, CapDet can locate more objects than JIVC, owing to our model's superior localization performance. After finetuning, CapDet can further describe a region rather than a single object such as" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15239" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 97, + 70, + 239, + 219 + ], + "blocks": [ + { + "bbox": [ + 97, + 70, + 239, + 219 + ], + "lines": [ + { + "bbox": [ + 97, + 70, + 239, + 219 + ], + "spans": [ + { + "bbox": [ + 97, + 70, + 239, + 219 + ], + "type": "table", + "html": "
MethodmAP(%)
FCLN [17]4.23
JIVC [40]7.85
Max Pooling [26]7.86
ImgG [26]7.81
COCD [26]7.92
COCG [26]8.90
COCG-LocSiz [26]8.76
COCG&GT [26]9.79
TDC+ROCSU [37]11.9
CapDet (Ours)13.98
", + "image_path": "75c8e679af8f535aa8600950e38054ca478204c2200ad135ad6f1f9168fa17d8.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 50, + 266, + 285, + 344 + ], + "blocks": [ + { + "bbox": [ + 47, + 228, + 288, + 251 + ], + "lines": [ + { + "bbox": [ + 47, + 228, + 288, + 251 + ], + "spans": [ + { + "bbox": [ + 47, + 228, + 288, + 251 + ], + "type": "text", + "content": "Table 3. Comparison of mAP (%) performance on the dense captioning benchmark on the VG-COCO Dataset." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 50, + 266, + 285, + 344 + ], + "lines": [ + { + "bbox": [ + 50, + 266, + 285, + 344 + ], + "spans": [ + { + "bbox": [ + 50, + 266, + 285, + 344 + ], + "type": "table", + "html": "
MODELDC HEADLVIS
APAPr / APc / APf
GLIP-T×30.422.5 / 29.0 / 33.0
GLIP-T✓33.127.0 / 32.1 / 35.0
DETCLIP-T×31.527.5 / 30.6 / 33.0
DETCLIP-T✓33.829.6 / 32.8 / 35.5
", + "image_path": "176bc80291b39f541bbb7d8a1554cb82554100fb22f74f28abafb47763bc5c36.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 50, + 379, + 285, + 441 + ], + "blocks": [ + { + "bbox": [ + 47, + 353, + 287, + 374 + ], + "lines": [ + { + "bbox": [ + 47, + 353, + 287, + 374 + ], + "spans": [ + { + "bbox": [ + 47, + 353, + 287, + 374 + ], + "type": "text", + "content": "Table 4. Ablations on integrating our dense captioning head into different baselines." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 50, + 379, + 285, + 441 + ], + "lines": [ + { + "bbox": [ + 50, + 379, + 285, + 441 + ], + "spans": [ + { + "bbox": [ + 50, + 379, + 285, + 441 + ], + "type": "table", + "html": "
Pre-training DataFine-tuneDCap mAP(%)Box mAP(%)
VG✘12.8627.65
O365,VG✘4.729.65
VG✓13.8328.58
O365,VG✓15.4430.61
", + "image_path": "bafbe26e5dd3eb83f5862dd315bad36dc4591d31c4c8b1ed25fef67a4a7e8800.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 449, + 286, + 472 + ], + "lines": [ + { + "bbox": [ + 47, + 449, + 286, + 472 + ], + "spans": [ + { + "bbox": [ + 47, + 449, + 286, + 472 + ], + "type": "text", + "content": "Table 5. Ablations on incorporating data from different sources. \"DCap\" stands for the dense caption mAP." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 47, + 496, + 233, + 507 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 496, + 233, + 507 + ], + "spans": [ + { + "bbox": [ + 47, + 496, + 233, + 507 + ], + "type": "text", + "content": "\"two women in a kitchen\" in the 5-th column." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 517, + 147, + 529 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 517, + 147, + 529 + ], + "spans": [ + { + "bbox": [ + 47, + 517, + 147, + 529 + ], + "type": "text", + "content": "4.3. Ablation Studies" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 536, + 225, + 550 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 536, + 225, + 550 + ], + "spans": [ + { + "bbox": [ + 47, + 536, + 225, + 550 + ], + "type": "text", + "content": "4.3.1 Ablations for Unified Pre-training" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 556, + 287, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 556, + 287, + 664 + ], + "spans": [ + { + "bbox": [ + 46, + 556, + 287, + 664 + ], + "type": "text", + "content": "Effect on different baselines. Table 4 investigates the advantages of dense captioning heads on different baselines. We integrate our dense captioning head with GLIP-T or DetCLIP-T. The GLIP-T is implemented with parallel text encoding without external knowledge following the setting as ablations in [38] on our code base. All the results are pre-trained on Objects365 and VG. The results show that our dense captioning head is able to boost the generalization and model-agnostic." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 665, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 287, + 713 + ], + "type": "text", + "content": "Effect of dense captioning data. Table 1 shows the efficiency of incorporating dense captioning data. Specifically, only " + }, + { + "bbox": [ + 46, + 665, + 287, + 713 + ], + "type": "inline_equation", + "content": "0.07\\mathrm{M}" + }, + { + "bbox": [ + 46, + 665, + 287, + 713 + ], + "type": "text", + "content": " data added, the DetCLIP-T(C) gains " + }, + { + "bbox": [ + 46, + 665, + 287, + 713 + ], + "type": "inline_equation", + "content": "+2.7\\%" + }, + { + "bbox": [ + 46, + 665, + 287, + 713 + ], + "type": "text", + "content": " overall AP and " + }, + { + "bbox": [ + 46, + 665, + 287, + 713 + ], + "type": "inline_equation", + "content": "+1.5\\%" + }, + { + "bbox": [ + 46, + 665, + 287, + 713 + ], + "type": "inline_equation", + "content": "AP_{r}" + }, + { + "bbox": [ + 46, + 665, + 287, + 713 + ], + "type": "text", + "content": " on LVIS compared to DetCLIP-" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 72, + 545, + 107 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 545, + 107 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 545, + 107 + ], + "type": "text", + "content": "T(A). The performance of DetCLP-T(A) on rare categories also outperforms DetCLIP-T(C) train on Objects365 and GOLDG, while the data size is 1.43M vs. 0.73M." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 114, + 470, + 126 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 114, + 470, + 126 + ], + "spans": [ + { + "bbox": [ + 306, + 114, + 470, + 126 + ], + "type": "text", + "content": "4.3.2 Ablations for dense captioning" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 133, + 545, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 133, + 545, + 312 + ], + "spans": [ + { + "bbox": [ + 304, + 133, + 545, + 312 + ], + "type": "text", + "content": "We investigate the impact of training policy and data from different sources on the dense captioning task. As shown in row1 in Table 5, our CapDet still achieves a significant performance which is directly trained on VG outperforms the previous task (i.e., TDC [37] in Table 2). Row2 is our CapDet and is pre-trained on Objects365 and VG, while only the bounding box in the Objects365 is regressed, and then transformed on a dense captioning task. Since the type of bounding boxes in dense captioning is different from the detection data, the result of the direct transforming to dense captioning is worse. However, we've proved that our model still keeps the dense captioning capacity on the salient objects in Figure 4. The results in row3 and row4 indicate that pre-training on the detection data Objects365 is also beneficial to the dense captioning task." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 315, + 381, + 327 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 315, + 381, + 327 + ], + "spans": [ + { + "bbox": [ + 306, + 315, + 381, + 327 + ], + "type": "text", + "content": "5. Limitations" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 334, + 545, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 334, + 545, + 430 + ], + "spans": [ + { + "bbox": [ + 304, + 334, + 545, + 430 + ], + "type": "text", + "content": "These are a few issues that we need to improve in the future: (1) Although our unification training paradigm works well on open-vocabulary object detection and dense captioning task, the training of dense captioning generation costs lots of time. (2) In addition, existing dense captioning data is high-cost to collect. We will research how to collect large-scale dense captioning data by auto annotation and get better performance with the scaled-up data." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 435, + 378, + 447 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 435, + 378, + 447 + ], + "spans": [ + { + "bbox": [ + 306, + 435, + 378, + 447 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 455, + 545, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 455, + 545, + 635 + ], + "spans": [ + { + "bbox": [ + 304, + 455, + 545, + 635 + ], + "type": "text", + "content": "In this paper, we propose a novel open-world object detection method named CapDet. Our CapDet is more practical in the open world and real scenes. Specifically, CapDet introduces a unification training framework including open-world object detection pre-training and dense captioning. The unification enables our CapDet to localize and recognize concepts in an arbitrary given category list or directly generate textual captions for predicted new concept objects. Experiments show that the design of unification is both beneficial to open-world object detection tasks and dense captioning tasks. In the future, our CapDet can be easily injected into any open world and real scenes tasks. The unification framework can also be integrated into any other OWD/OVD methods to generate semantic-rich concepts for unknown/novel objects." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 643, + 545, + 691 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 643, + 545, + 691 + ], + "spans": [ + { + "bbox": [ + 304, + 643, + 545, + 691 + ], + "type": "text", + "content": "Acknowledgements We gratefully acknowledge the support of MindSpore1, CANN (Compute Architecture for Neural Networks) and Ascend AI Processor used for this research." + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 321, + 702, + 414, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 702, + 414, + 713 + ], + "spans": [ + { + "bbox": [ + 321, + 702, + 414, + 713 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 321, + 702, + 414, + 713 + ], + "type": "text", + "content": "https://www.mindspore.cn/" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "15240" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 708 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 201 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 201 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 201 + ], + "type": "text", + "content": "[1] Tom B. Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert-Voss, Gretchen Krueger, Tom Henighan, Rewon Child, Aditya Ramesh, Daniel M. Ziegler, Jeffrey Wu, Clemens Winter, Christopher Hesse, Mark Chen, Eric Sigler, Mateusz Litwin, Scott Gray, Benjamin Chess, Jack Clark, Christopher Berner, Sam McCandlish, Alec Radford, Ilya Sutskever, and Dario Amodei. Language models are few-shot learners. CoRR, abs/2005.14165, 2020. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 202, + 288, + 290 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 202, + 288, + 290 + ], + "spans": [ + { + "bbox": [ + 53, + 202, + 288, + 290 + ], + "type": "text", + "content": "[2] Kai Chen, Jiaqi Wang, Jiangmiao Pang, Yuhang Cao, Yu Xiong, Xiaoxiao Li, Shuyang Sun, Wansen Feng, Ziwei Liu, Jiarui Xu, Zheng Zhang, Dazhi Cheng, Chenchen Zhu, Tianheng Cheng, Qijie Zhao, Buyu Li, Xin Lu, Rui Zhu, Yue Wu, Jifeng Dai, Jingdong Wang, Jianping Shi, Wanli Ouyang, Chen Change Loy, and Dahua Lin. Mmdetection: Open mmlab detection toolbox and benchmark. ArXiv, abs/1906.07155, 2019. 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 291, + 288, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 291, + 288, + 346 + ], + "spans": [ + { + "bbox": [ + 53, + 291, + 288, + 346 + ], + "type": "text", + "content": "[3] Xiyang Dai, Yinpeng Chen, Bin Xiao, Dongdong Chen, Mengchen Liu, Lu Yuan, and Lei Zhang. Dynamic head: Unifying object detection heads with attentions. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7373-7382, 2021. 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 347, + 288, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 347, + 288, + 392 + ], + "spans": [ + { + "bbox": [ + 53, + 347, + 288, + 392 + ], + "type": "text", + "content": "[4] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, K. Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. 2009 IEEE Conference on Computer Vision and Pattern Recognition, pages 248-255, 2009. 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 393, + 288, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 393, + 288, + 437 + ], + "spans": [ + { + "bbox": [ + 53, + 393, + 288, + 437 + ], + "type": "text", + "content": "[5] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. BERT: pre-training of deep bidirectional transformers for language understanding. CoRR, abs/1810.04805, 2018. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 438, + 287, + 493 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 438, + 287, + 493 + ], + "spans": [ + { + "bbox": [ + 53, + 438, + 287, + 493 + ], + "type": "text", + "content": "[6] Xiaoyi Dong, Yinglin Zheng, Jianmin Bao, Ting Zhang, Dongdong Chen, Hao Yang, Ming Zeng, Weiming Zhang, Lu Yuan, Dong Chen, et al. Maskclip: Masked self-distillation advances contrastive language-image pretraining. arXiv preprint arXiv:2208.12262, 2022. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 495, + 287, + 559 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 495, + 287, + 559 + ], + "spans": [ + { + "bbox": [ + 53, + 495, + 287, + 559 + ], + "type": "text", + "content": "[7] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. CoRR, abs/2010.11929, 2020. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 562, + 288, + 616 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 562, + 288, + 616 + ], + "spans": [ + { + "bbox": [ + 53, + 562, + 288, + 616 + ], + "type": "text", + "content": "[8] Yu Du, Fangyun Wei, Zihe Zhang, Miaojing Shi, Yue Gao, and Guoqi Li. Learning to prompt for open-vocabulary object detection with vision-language model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14084-14093, 2022. 1, 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 618, + 288, + 661 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 618, + 288, + 661 + ], + "spans": [ + { + "bbox": [ + 53, + 618, + 288, + 661 + ], + "type": "text", + "content": "[9] Yiqi Gao, Xinglin Hou, Yuanmeng Zhang, Tiezheng Ge, Yunig Jiang, and Peng Wang. Caponimage: Context-driven dense-captioning on image. arXiv preprint arXiv:2204.12974, 2022. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 662, + 288, + 708 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 662, + 288, + 708 + ], + "spans": [ + { + "bbox": [ + 48, + 662, + 288, + 708 + ], + "type": "text", + "content": "[10] Georgia Gkioxari, Ross Girshick, Piotr Dólar, and Kaiming He. Detecting and recognizing human-object interactions. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 8359-8367, 2018. 1" + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 25, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 105 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 105 + ], + "type": "text", + "content": "[11] Alex Graves. Long short-term memory. Supervised sequence labelling with recurrent neural networks, pages 37-45, 2012. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 307, + 106, + 545, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 106, + 545, + 149 + ], + "spans": [ + { + "bbox": [ + 307, + 106, + 545, + 149 + ], + "type": "text", + "content": "[12] Xiuye Gu, Tsung-Yi Lin, Weicheng Kuo, and Yin Cui. Open-vocabulary object detection via vision and language knowledge distillation. arXiv preprint arXiv:2104.13921, 2021. 1, 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 307, + 150, + 545, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 150, + 545, + 194 + ], + "spans": [ + { + "bbox": [ + 307, + 150, + 545, + 194 + ], + "type": "text", + "content": "[13] Agrim Gupta, Piotr Dollar, and Ross Girshick. Lvis: A dataset for large vocabulary instance segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5356-5364, 2019. 5, 6" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 195, + 545, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 195, + 545, + 249 + ], + "spans": [ + { + "bbox": [ + 307, + 195, + 545, + 249 + ], + "type": "text", + "content": "[14] Akshita Gupta, Sanath Narayan, KJ Joseph, Salman Khan, Fahad Shahbaz Khan, and Mubarak Shah. Ow-detr: Open-world detection transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9235-9244, 2022. 1, 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 250, + 545, + 292 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 250, + 545, + 292 + ], + "spans": [ + { + "bbox": [ + 307, + 250, + 545, + 292 + ], + "type": "text", + "content": "[15] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In Proceedings of the IEEE international conference on computer vision, pages 2961-2969, 2017. 3, 5" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 293, + 545, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 293, + 545, + 358 + ], + "spans": [ + { + "bbox": [ + 307, + 293, + 545, + 358 + ], + "type": "text", + "content": "[16] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In International Conference on Machine Learning, pages 4904-4916. PMLR, 2021. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 360, + 545, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 360, + 545, + 415 + ], + "spans": [ + { + "bbox": [ + 307, + 360, + 545, + 415 + ], + "type": "text", + "content": "[17] Justin Johnson, Andrej Karpathy, and Li Fei-Fei. Densecap: Fully convolutional localization networks for dense captioning. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4565–4574, 2016. 2, 3, 6, 7, 8" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 415, + 545, + 469 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 415, + 545, + 469 + ], + "spans": [ + { + "bbox": [ + 307, + 415, + 545, + 469 + ], + "type": "text", + "content": "[18] KJ Joseph, Salman Khan, Fahad Shahbaz Khan, and Vineeth N Balasubramanian. Towards open world object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5830-5840, 2021. 1, 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 471, + 545, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 471, + 545, + 525 + ], + "spans": [ + { + "bbox": [ + 307, + 471, + 545, + 525 + ], + "type": "text", + "content": "[19] Aishwarya Kamath, Mannat Singh, Yann LeCun, Gabriel Synnaeve, Ishan Misra, and Nicolas Carion. Mdetr-modulated detection for end-to-end multi-modal understanding. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1780-1790, 2021. 5, 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 525, + 545, + 569 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 525, + 545, + 569 + ], + "spans": [ + { + "bbox": [ + 307, + 525, + 545, + 569 + ], + "type": "text", + "content": "[20] Wonjae Kim, Bokyung Son, and Ildoo Kim. Vilt: Vision-and-language transformer without convolution or region supervision. In International Conference on Machine Learning, pages 5583-5594. PMLR, 2021. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 570, + 545, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 570, + 545, + 602 + ], + "spans": [ + { + "bbox": [ + 307, + 570, + 545, + 602 + ], + "type": "text", + "content": "[21] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014.5" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 603, + 545, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 603, + 545, + 668 + ], + "spans": [ + { + "bbox": [ + 307, + 603, + 545, + 668 + ], + "type": "text", + "content": "[22] Ranjay Krishna, Yuke Zhu, Oliver Groth, Justin Johnson, Kenji Hata, Joshua Kravitz, Stephanie Chen, Yannis Kalantidis, Li-Jia Li, David A Shamma, et al. Visual genome: Connecting language and vision using crowdsourced dense image annotations. International journal of computer vision, 123(1):32-73, 2017. 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 669, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 669, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 669, + 545, + 713 + ], + "type": "text", + "content": "[23] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. arXiv preprint arXiv:2201.12086, 2022. 2" + } + ] + } + ], + "index": 24 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15241" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "text", + "content": "[24] Liunian Harold Li, Mark Yatskar, Da Yin, Cho-Jui Hsieh, and Kai-Wei Chang. Visualbert: A simple and perform-. mant baseline for vision and language. arXiv preprint arXiv:1908.03557, 2019. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 118, + 288, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 118, + 288, + 185 + ], + "spans": [ + { + "bbox": [ + 48, + 118, + 288, + 185 + ], + "type": "text", + "content": "[25] Liunian Harold Li, Pengchuan Zhang, Haotian Zhang, Jianwei Yang, Chunyuan Li, Yiwu Zhong, Lijuan Wang, Lu Yuan, Lei Zhang, Jenq-Neng Hwang, et al. Grounded language-image pre-training. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10965-10975, 2022. 2, 3, 4, 5, 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 186, + 288, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 186, + 288, + 228 + ], + "spans": [ + { + "bbox": [ + 48, + 186, + 288, + 228 + ], + "type": "text", + "content": "[26] Xiangyang Li, Shuqiang Jiang, and Jungong Han. Learning object context for dense captioning. In Proceedings of the AAAI conference on artificial intelligence, volume 33, pages 8650-8657, 2019. 3, 7, 8" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 230, + 288, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 230, + 288, + 285 + ], + "spans": [ + { + "bbox": [ + 48, + 230, + 288, + 285 + ], + "type": "text", + "content": "[27] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In European conference on computer vision, pages 740-755. Springer, 2014. 1, 5, 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 286, + 288, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 286, + 288, + 342 + ], + "spans": [ + { + "bbox": [ + 48, + 286, + 288, + 342 + ], + "type": "text", + "content": "[28] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 9992-10002, 2021. 5, 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 343, + 288, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 343, + 288, + 387 + ], + "spans": [ + { + "bbox": [ + 48, + 343, + 288, + 387 + ], + "type": "text", + "content": "[29] Jingjing Meng, Junsong Yuan, Jiong Yang, Gang Wang, and Yap-Peng Tan. Object instance search in videos via spatiotemporal trajectory discovery. IEEE Transactions on Multimedia, 18(1):116-127, 2015. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 388, + 288, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 388, + 288, + 441 + ], + "spans": [ + { + "bbox": [ + 48, + 388, + 288, + 441 + ], + "type": "text", + "content": "[30] James Philbin, Ondrej Chum, Michael Isard, Josef Sivic, and Andrew Zisserman. Object retrieval with large vocabularies and fast spatial matching. In 2007 IEEE conference on computer vision and pattern recognition, pages 1-8. IEEE, 2007. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 444, + 288, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 444, + 288, + 510 + ], + "spans": [ + { + "bbox": [ + 48, + 444, + 288, + 510 + ], + "type": "text", + "content": "[31] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 512, + 288, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 512, + 288, + 555 + ], + "spans": [ + { + "bbox": [ + 48, + 512, + 288, + 555 + ], + "type": "text", + "content": "[32] Shafin Rahman, Salman Khan, and Fatih Porikli. Zero-shot object detection: Learning to simultaneously recognize and localize novel concepts. In Asian Conference on Computer Vision, pages 547-563. Springer, 2018. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 557, + 288, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 557, + 288, + 601 + ], + "spans": [ + { + "bbox": [ + 48, + 557, + 288, + 601 + ], + "type": "text", + "content": "[33] Joseph Redmon, Santosh Divvala, Ross Girshick, and Ali Farhadi. You only look once: Unified, real-time object detection. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 779-788, 2016. 1, 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 602, + 288, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 602, + 288, + 646 + ], + "spans": [ + { + "bbox": [ + 48, + 602, + 288, + 646 + ], + "type": "text", + "content": "[34] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. Advances in neural information processing systems, 28, 2015. 1, 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 647, + 288, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 647, + 288, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 647, + 288, + 712 + ], + "type": "text", + "content": "[35] Hamid Rezatofighi, Nathan Tsoi, JunYoung Gwak, Amir Sadeghian, Ian Reid, and Silvio Savarese. Generalized intersection over union: A metric and a loss for bounding box regression. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 658-666, 2019. 4" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 546, + 714 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 307, + 73, + 546, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 546, + 128 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 546, + 128 + ], + "type": "text", + "content": "[36] Shuai Shao, Zeming Li, Tianyuan Zhang, Chao Peng, Gang Yu, Xiangyu Zhang, Jing Li, and Jian Sun. Objects365: A large-scale, high-quality dataset for object detection. In Proceedings of the IEEE/CVF international conference on computer vision, pages 8430-8439, 2019. 1, 6" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 307, + 129, + 546, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 129, + 546, + 172 + ], + "spans": [ + { + "bbox": [ + 307, + 129, + 546, + 172 + ], + "type": "text", + "content": "[37] Zhuang Shao, Jungong Han, Demetris Marnerides, and Kurt Debattista. Region-object relation-aware dense captioning via transformer. IEEE Transactions on Neural Networks and Learning Systems, 2022. 3, 6, 7, 8" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 173, + 546, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 173, + 546, + 227 + ], + "spans": [ + { + "bbox": [ + 307, + 173, + 546, + 227 + ], + "type": "text", + "content": "[38] Sheng Shen, Chunyuan Li, Xiaowei Hu, Yujia Xie, Jianwei Yang, Pengchuan Zhang, Anna Rohrbach, Zhe Gan, Lijuan Wang, Lu Yuan, et al. K-lite: Learning transferable visual models with external knowledge. arXiv preprint arXiv:2204.09222, 2022. 3, 5, 8" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 228, + 546, + 261 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 228, + 546, + 261 + ], + "spans": [ + { + "bbox": [ + 307, + 228, + 546, + 261 + ], + "type": "text", + "content": "[39] Wenhui Wang, Hangbo Bao, Li Dong, and Furu Wei. Vlmo: Unified vision-language pre-training with mixture-of-modality-experts. arXiv preprint arXiv:2111.02358, 2021. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 261, + 546, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 261, + 546, + 304 + ], + "spans": [ + { + "bbox": [ + 307, + 261, + 546, + 304 + ], + "type": "text", + "content": "[40] Linjie Yang, Kevin Tang, Jianchao Yang, and Li-Jia Li. Dense captioning with joint inference and visual context. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2193-2202, 2017. 7, 8" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 305, + 546, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 305, + 546, + 360 + ], + "spans": [ + { + "bbox": [ + 307, + 305, + 546, + 360 + ], + "type": "text", + "content": "[41] Lewei Yao, Jianhua Han, Youpeng Wen, Xiaodan Liang, Dan Xu, Wei Zhang, Zhenguo Li, Chunjing Xu, and Hang Xu. Detclip: Dictionary-enriched visual-concept paralleled pre-training for open-world detection. arXiv preprint arXiv:2209.09407, 2022. 1, 2, 3, 4, 5, 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 361, + 546, + 405 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 361, + 546, + 405 + ], + "spans": [ + { + "bbox": [ + 307, + 361, + 546, + 405 + ], + "type": "text", + "content": "[42] Lewei Yao, Runhu Huang, Lu Hou, Guansong Lu, Minzhe Niu, Hang Xu, Xiaodan Liang, Zhenguo Li, Xin Jiang, and Chunjing Xu. Filip: Fine-grained interactive language-image pre-training. ArXiv, abs/2111.07783, 2022. 5" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 405, + 546, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 405, + 546, + 459 + ], + "spans": [ + { + "bbox": [ + 307, + 405, + 546, + 459 + ], + "type": "text", + "content": "[43] Guojun Yin, Lu Sheng, Bin Liu, Nenghai Yu, Xiaogang Wang, and Jing Shao. Context and attribute grounded dense captioning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6241-6250, 2019. 3, 7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 459, + 546, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 459, + 546, + 514 + ], + "spans": [ + { + "bbox": [ + 307, + 459, + 546, + 514 + ], + "type": "text", + "content": "[44] Alireza Zareian, Kevin Dela Rosa, Derek Hao Hu, and Shih-Fu Chang. Open-vocabulary object detection using captions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14393-14402, 2021. 1, 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 514, + 546, + 558 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 514, + 546, + 558 + ], + "spans": [ + { + "bbox": [ + 307, + 514, + 546, + 558 + ], + "type": "text", + "content": "[45] Haoyang Zhang and Xuming He. Deep free-form deformation network for object-mask registration. In Proceedings of the IEEE International Conference on Computer Vision, pages 4251-4259, 2017. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 559, + 546, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 559, + 546, + 613 + ], + "spans": [ + { + "bbox": [ + 307, + 559, + 546, + 613 + ], + "type": "text", + "content": "[46] Haotian Zhang, Pengchuan Zhang, Xiaowei Hu, Yen-Chun Chen, Liunian Harold Li, Xiyang Dai, Lijuan Wang, Lu Yuan, Jenq-Neng Hwang, and Jianfeng Gao. Glipv2: Unifying localization and vision-language understanding. arXiv preprint arXiv:2206.05836, 2022. 2, 5" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 614, + 546, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 614, + 546, + 679 + ], + "spans": [ + { + "bbox": [ + 307, + 614, + 546, + 679 + ], + "type": "text", + "content": "[47] Shifeng Zhang, Cheng Chi, Yongqiang Yao, Zhen Lei, and Stan Z Li. Bridging the gap between anchor-based and anchor-free detection via adaptive training sample selection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9759-9768, 2020. 3, 4, 5" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 680, + 546, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 680, + 546, + 714 + ], + "spans": [ + { + "bbox": [ + 307, + 680, + 546, + 714 + ], + "type": "text", + "content": "[48] Xiaowei Zhao, Xianglong Liu, Yifan Shen, Yuqing Ma, Yixuan Qiao, and Duorui Wang. Revisiting open world object detection. arXiv preprint arXiv:2201.00471, 2022. 1, 2, 3" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "15242" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 172 + ], + "type": "list", + "angle": 0, + "index": 2, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 126 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 126 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 126 + ], + "type": "text", + "content": "[49] Xingyi Zhou, Rohit Girdhar, Armand Joulin, Philipp Krahenbuhl, and Ishan Misra. Detecting twenty-thousand classes using image-level supervision. In European Conference on Computer Vision, pages 350-368. Springer, 2022. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 129, + 287, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 129, + 287, + 172 + ], + "spans": [ + { + "bbox": [ + 48, + 129, + 287, + 172 + ], + "type": "text", + "content": "[50] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable detr: Deformable transformers for end-to-end object detection. arXiv preprint arXiv:2010.04159, 2020. 1" + } + ] + } + ], + "index": 1 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "15243" + } + ] + } + ], + "index": 3 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Cascade Evidential Learning for Open-World Weakly-Supervised Temporal Action Localization/3c2843a4-d81f-4df7-8b2c-5c29eaeafad1_content_list.json b/2023/Cascade Evidential Learning for Open-World Weakly-Supervised Temporal Action Localization/3c2843a4-d81f-4df7-8b2c-5c29eaeafad1_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..b66cb94ce019adae1357e7e2cc613187940a5c0b --- /dev/null +++ b/2023/Cascade Evidential Learning for Open-World Weakly-Supervised Temporal Action Localization/3c2843a4-d81f-4df7-8b2c-5c29eaeafad1_content_list.json @@ -0,0 +1,1480 @@ +[ + { + "type": "text", + "text": "Cascade Evidential Learning for Open-world Weakly-supervised Temporal Action Localization", + "text_level": 1, + "bbox": [ + 107, + 160, + 863, + 204 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Mengyuan Chen $^{1,2}$ , Junyu Gao $^{1,2}$ , and Changsheng Xu $^{1,2,3}$", + "bbox": [ + 252, + 226, + 715, + 245 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 State Key Laboratory of Multimodal Artificial Intelligence Systems (MAIS), Institute of Automation, Chinese Academy of Sciences (CASIA)", + "bbox": [ + 176, + 245, + 793, + 276 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ School of Artificial Intelligence, University of Chinese Academy of Sciences (UCAS)", + "bbox": [ + 139, + 276, + 828, + 292 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{3}$ Peng Cheng Laboratory, ShenZhen, China", + "bbox": [ + 310, + 292, + 660, + 308 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "chenmengyuan2021@ia.ac.cn; { junyu.gao, csxu}@nlpr.ia.ac.cn", + "bbox": [ + 225, + 311, + 737, + 325 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 357, + 312, + 372 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Targeting at recognizing and localizing action instances with only video-level labels during training, Weakly-supervised Temporal Action Localization (WTAL) has achieved significant progress in recent years. However, living in the dynamically changing open world where unknown actions constantly spring up, the closed-set assumption of existing WTAL methods is invalid. Compared with traditional open-set recognition tasks, Open-world WTAL (OW-TAL) is challenging since not only are the annotations of unknown samples unavailable, but also the fine-grained annotations of known action instances can only be inferred ambiguously from the video category labels. To address this problem, we propose a Cascade Evidential Learning framework at an evidence level, which targets at OWTAL for the first time. Our method jointly leverages multi-scale temporal contexts and knowledge-guided prototype information to progressively collect cascade and enhanced evidence for known action, unknown action, and background separation. Extensive experiments conducted on THUMOS-14 and ActivityNet-v1.3 verify the effectiveness of our method. Besides the classification metrics adopted by previous open-set recognition methods, we also evaluate our method on localization metrics which are more reasonable for OWTAL.", + "bbox": [ + 75, + 388, + 473, + 706 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 77, + 747, + 208, + 762 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Targeting at recognizing and localizing action instances with only video-level labels during training, Weakly-supervised Temporal Action Localization (WTAL) has attracted increasing attention from both academia and industry [9, 11, 18, 19, 37, 43]. Unlike fully-supervised TAL, WTAL only requires video-level action labels during training. However, the closed-set assumption of existing WTAL", + "bbox": [ + 75, + 769, + 470, + 866 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/daa5b0178a8ff34dde03014efcda7b9b40d5693769d885a40a2c4ce4507a0d42.jpg", + "image_caption": [ + "Training Phase" + ], + "image_footnote": [], + "bbox": [ + 502, + 368, + 692, + 458 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/b1b53169730a0881bee51cdd58636c5d7c6bb46c1f52ac76484b9c323c349076.jpg", + "image_caption": [ + "Testing Phase", + "Figure 1. Illustration of the training and testing phases of OWTAL. With only video-level labels for training, OWTAL aims to localize both known and unknown action instances in testing videos." + ], + "image_footnote": [], + "bbox": [ + 695, + 368, + 889, + 459 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "methods is invalid in the dynamically changing real world, since with the development of society never-before-seen human action categories are constantly emerging. Therefore, to address this problem, we consider a different WTAL setting in this work, termed as Open-world WTAL (OWTAL).", + "bbox": [ + 497, + 517, + 892, + 587 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Different from the traditional WTAL task, as shown in Figure 1, OWTAL allows testing videos to contain action instances of unknown categories, which have never appeared during training. Therefore, temporal boundaries of both known and unknown action instances are expected to be predicted. Compared with its fully-supervised counterpart Open Set TAL [3], OWTAL is challenging in two aspects: (1) Ambiguity of annotations of closed-set (known) action instances. Previous works indicate that the closed-set and open-set performance are highly correlated [42]. However, under the OWTAL setting, not only are the annotations of unknown action instances unavailable, but also the fine-grained annotations of known ones can only be inferred ambiguously from the video category labels. During training, the known action instances that the model needs to focus on are prone to be disturbed by the background snippets, which hinders the learning of the closed-set actions, thus making it extremely difficult to differentiate the unknown actions, the known actions, and the background. (2) Lack of reasonable metrics. The traditional Open Set Recognition (OSR) aims", + "bbox": [ + 497, + 590, + 894, + 866 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 129, + 49, + 200, + 85 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 252, + 48, + 791, + 87 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "14741", + "bbox": [ + 465, + 885, + 504, + 897 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "for classification while the goal of OWTAL is to perform localization instead, thus the classification metrics commonly adopted by OSR are not sufficient for OWTAL.", + "bbox": [ + 75, + 125, + 468, + 167 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In order to alleviate the negative impact caused by the weak annotations of known action instances, we propose a Cascade Evidential Learning method for owtaL (CELL), which progressively collects cascaded evidence by considering both temporal contexts in multi-scale ranges and inter-video correlations under the guidance of prior knowledge. Since the goal of OWTAL is to locate the consecutive known/unknown action segments of various temporal lengths in open-world scenarios, perceiving temporal contexts in diverse ranges is essential. We argue that it is meaningful to endow individual snippet features with the ability of sensing multi-scale neighborhood video segments, and thus a Multi-scale Extended-range Perception module (Section 3.2) is designed to obtain more discriminative video features for initial evidence collection, taking advantage of the temporal contexts. Due to the large intra-action variation in visual patterns and the lack of prior knowledge guidance, the known action instances which visually deviate from the common ones are likely to be misidentified with the initial evidence collected from individual videos. Therefore, we design a Knowledge-guided Bipolar Prototype Learning strategy (Section 3.3), where a semantic relation graph is constructed to provide prior knowledge guidance for the bipolar prototype learning among videos, thus perceiving the open-world more comprehensively. We use this strategy to generate a series of evidence calibration factors for further cascade evidence enhancement. Finally, a Cascade Evidence Enhancement module (Section 3.4) is designed for enhancing the initial evidence with the calibration factors, and the uncertainty estimated from the cascaded evidence is used for the known/unknown judgment. Extensive experiments conducted on THUMOS-14 and ActivityNet verify the effectiveness. Besides the various classification metrics adopted by previous works, we also evaluate our method on localization metrics which are more in line with the needs of real-world applications.", + "bbox": [ + 75, + 169, + 472, + 666 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To summarize, our contribution is threefold:", + "bbox": [ + 95, + 667, + 389, + 681 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- To tackle the unique challenges of OWTAL, we propose a cascade evidential learning framework, which progressively collects comprehensive evidence for known action, unknown action, and background separation. Localization metrics which meet the needs of the real-world more closely are adopted for evaluation.", + "- To achieve OWTAL without fine-grained annotations, the proposed CELL jointly leverages multi-scale temporal contexts and knowledge-guided prototype information during the evidence cascade learning process.", + "- We conduct comprehensive experiments on two popular WTAL benchmarks, THUMOS-14 and ActivityNetv1.3, and achieve significant performance improvement" + ], + "bbox": [ + 84, + 684, + 468, + 866 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "over various baselines. Experiments show that CELL enables existing methods to well adapt to the more practical open-world settings.", + "bbox": [ + 521, + 125, + 890, + 167 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 499, + 180, + 640, + 194 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Weakly-supervised Temporal Action Localization. Originated from UntrimmedNet [43], the pioneer work to utilize video-level action category annotations as the weak supervision for TAL, existing WTAL methods commonly adopt a multiple instance learning strategy and can be roughly categorized into erasing-based, attention-based and uncertainty-based methods. Erasing-based methods [38, 49, 50] erase the most discriminative segments to mitigate the single snippet cheating issue [48]. Attention-based methods [10, 19, 21, 29, 33, 36] employ the attention mechanism to select snippets most likely to be the foreground with activation scores. The newly-arising uncertainty-based methods [9, 24, 46] utilize uncertainty to address the inevitable action-background ambiguity caused by the weakly-supervised setting. However, all existing WTAL methods are rooted in the closed-set assumption, which impedes their application to real-world scenarios. In contrast, we address the WTAL task under the open-world setting by collecting reliable cascaded evidence, which has not been explored in prior work.", + "bbox": [ + 497, + 201, + 890, + 463 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Open Set Recognition. Open Set Recognition (OSR) aims to recognize known classes and reject the unknown. [34] first formalize the OSR problem and provide a basic framework. [4] propose the first DNN-based OSR method Openmax, which rejects unknown classes by modeling the distance of activation vectors with Extreme Value Theory (EVT). GAN-based methods are also developed towards this task. [30] generate samples similar to the training data but not belonging to the known classes, and then utilize the generated unknown class samples to train an open-set classifier. Other approaches include prototype-based methods [7, 8], which reject unknown classes by calculating the maximum distance between the input sample and the learned closed-set prototypes, and reconstruction-based methods [31, 40, 47] utilize the reconstruction error in the test phase as an open-set indicator. Some works [28, 51] attempt to seek assistance from external knowledge, but few of them have explicitly modeled unknown class information in an end-to-end framework.", + "bbox": [ + 497, + 465, + 890, + 726 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Most existing approaches focus on open-set recognition tasks while few works attend to temporal localization-related vision field. [3] formalize the open-set temporal action localization problem and propose an OpenTAL framework. However, its fully-supervised setting requires large amounts of fine-grained labels whose annotation is time-consuming, error-prone, and costly. Besides, the evaluation metrics in [3] are mainly classification ones. In contrast, in this paper, we address the challenging OWTAL problem which also abandons the closed-set assumption but only re", + "bbox": [ + 497, + 728, + 890, + 866 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "14742", + "bbox": [ + 465, + 885, + 505, + 897 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "quires easily available video-level class labels, and employ comprehensive localization-related metrics for evaluation.", + "bbox": [ + 75, + 125, + 467, + 153 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Evidential Deep Learning. Based on Dempster-Shafer Theory (DST) [45] and Subjective Logic theory (SL) [22], Evidential Deep Learning (EDL) allows uncertainty estimation in a single forward pass [41] by collecting scalar evidence for each category and parameterizing a Dirichlet distribution, which models the distribution of classification probabilities, over the collected evidence. As a newly arising trustworthy method, EDL has achieved remarkable progress in various computer vision tasks, including action localization [9], image classification [35], regression [1], multi-view classification [17, 27], out-of-distribution detection [20], long-tail learning [25], and open-set action recognition and localization [2, 3]. However, current EDL algorithms only focus on collecting evidence from individual samples, while they neglect the complementary role that the potential correlations among samples can play in evidence collection. To address this issue, in this work we propose a cascade evidence calibration paradigm which effectively enhances the collected evidence.", + "bbox": [ + 75, + 154, + 468, + 416 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Our Approach", + "text_level": 1, + "bbox": [ + 75, + 428, + 225, + 444 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Notations and Preliminaries", + "text_level": 1, + "bbox": [ + 75, + 445, + 327, + 459 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Problem formulation of OWTAL. OWTAL targets at detecting known/unknown action instances in untrimmed videos. Formally, we are given a training set $\\{\\mathcal{V},\\mathcal{Y}\\}$ with $N$ training videos, where $\\mathcal{V}$ denotes an untrimmed video, and $\\pmb {y}\\in \\mathbb{R}^{C}$ is its multi-hot label indicating the action categories that the action instances in this video belong to, where $C$ is the number of known action categories. During testing, the goal is to use the learned model to output a set of quadruplets $\\{c_r,t_r^s,t_r^e,\\gamma_r\\}_{r = 1}^R$ ,where $R$ is the number of action instances in $\\mathcal{V}$ $c_{r},t_{r}^{s},t_{r}^{e}$ and $\\gamma_{r}$ denote the action category, the start and end timestamps, and the confidence score, respectively. $c_{r}\\in \\{1,\\dots,C + 1\\}$ , where $C + 1$ denotes the unknown action category.", + "bbox": [ + 75, + 465, + 468, + 643 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Feature extraction. Following previous works [10,46], we split the untrimmed video $\\mathcal{V}$ into $T$ non-overlapping 16-frame snippets and then use pre-trained networks, e.g., the I3D model [23], to extract RGB and optical flow features. Then, the above features are concatenated and then fed into a fusion module, e.g., convolutional layers [19, 33], to obtain the snippet-wise feature $\\pmb{X} = [\\pmb{x}_1,\\dots,\\pmb{x}_T]\\in \\mathbb{R}^{T\\times D}$ , where $D$ is the feature dimension.", + "bbox": [ + 75, + 644, + 468, + 755 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Multi-scale Extended-range Perception for Initial Evidence Collection", + "text_level": 1, + "bbox": [ + 75, + 764, + 468, + 792 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "A video may contain several known/unknown action instances with various temporal lengths in the open-world. Since the goal of OWTAL is to locate these consecutive video segments, perceiving temporal contexts in diverse ranges is essential. We argue that it is meaningful to endow", + "bbox": [ + 75, + 796, + 468, + 866 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "individual snippet features with the ability of sensing multiscale neighborhood video segments. As shown in Figure 2, we design a Multi-scale Extended-range Perception module (MEP) to obtain more discriminative video features, taking advantage of the temporal contexts. Note that some anchor-based temporal modeling methods [6, 12, 14, 44] also leverage multi-scale temporal information. However, their goal is to obtain the fused temporal feature for proposal generation, while the MEP aims to enhance individual snippet features for initial evidence collection.", + "bbox": [ + 497, + 125, + 890, + 262 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In MEP, each snippet $t$ is assigned with a series of $t$ -centered multi-scale video segments, whose context-aware features can contribute to the extension of snippet-specific perception ranges. Here, the temporal boundaries of the $t$ -centered video segments can be denoted as set $\\Omega_t = \\{(s_t^m, e_t^m) | m \\in \\mathcal{M}\\}$ , where $\\mathcal{M}$ is a pre-defined length set, $s_t^m = \\max(0, t - m)$ , and $e_t^m = \\min(T, t + m)$ . For brevity, we denote the temporal average of the feature sequence of the extended video segment $(s_t^m, e_t^m)$ as $f_t^m \\in \\mathbb{R}^D$ , which can be given by:", + "bbox": [ + 497, + 263, + 892, + 402 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {f} _ {t} ^ {m} = \\frac {1}{e _ {t} ^ {m} - s _ {t} ^ {m} + 1} \\sum_ {s _ {t} ^ {m} \\leq i \\leq e _ {t} ^ {m}} \\boldsymbol {x} _ {i}. \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 581, + 406, + 890, + 441 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To extend the perception range of snippet $t$ , we perform feature enhancement by fusing $\\boldsymbol{x}_t$ with $\\boldsymbol{f}_t^m$ , during which the weight of $\\boldsymbol{f}_t^m$ is positively correlated with the similarity between them. Such operation is motivated by the simple expectation that a snippet-level feature would absorb meaningful contextual information mainly from extended video segments that are highly relevant to it. Therefore, the snippet-level feature with extended-range perception, denoted as $\\tilde{\\boldsymbol{F}} = [\\tilde{\\boldsymbol{f}}_1,\\dots ,\\tilde{\\boldsymbol{f}}_T]$ , can be formulated as:", + "bbox": [ + 497, + 447, + 890, + 572 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {\\boldsymbol {f}} _ {t} = \\left(1 - \\alpha_ {t}\\right) \\varphi_ {1} (\\boldsymbol {x} _ {t}) + \\alpha_ {t} \\sum_ {m \\in \\mathcal {M}} \\delta \\left(\\omega_ {t} ^ {m}\\right) \\varphi_ {2} \\left(\\boldsymbol {f} _ {t} ^ {m}\\right), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 522, + 578, + 890, + 606 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\delta (\\cdot)$ denotes the Softmax operation with a temperature factor, $\\varphi_{1,2}$ are fully-connected layers for feature embedding, $\\alpha_{t} \\in [0,1]$ is the fusion weight representing the scaled average of the cosine similarities $\\omega_{t}^{m}$ between $\\pmb{x}_{t}$ and $\\pmb{f}_t^m$ , and $\\alpha_{t}$ and $\\omega_t^m$ can be calculated by:", + "bbox": [ + 497, + 612, + 890, + 682 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\alpha_ {t} = \\frac {1}{2 | \\mathcal {M} |} \\sum_ {m \\in \\mathcal {M}} \\left(\\omega_ {t} ^ {m} + 1\\right), \\omega_ {t} ^ {m} = \\cos \\left(\\frac {\\boldsymbol {x} _ {t}}{\\left| \\boldsymbol {x} _ {t} \\right|}, \\frac {\\boldsymbol {f} _ {t} ^ {m}}{\\left| \\boldsymbol {f} _ {t} ^ {m} \\right|}\\right). \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 507, + 688, + 890, + 733 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Thereafter, similar to previous methods [10, 19], we utilize an attention module to predict an attention score sequence $\\mathbf{A} = [A_{1},\\dots,A_{T}]\\in \\mathbb{R}^{T}$ for the input video $\\nu$ representing the probabilities of snippets belonging to the foreground, then select the top-ranked snippets. Specifically, for the input video we obtain the set $\\Theta$ of snippets with the top- $k$ attention scores by the following equation:", + "bbox": [ + 497, + 735, + 890, + 831 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\Theta = \\arg \\max _ {\\Theta} \\sum_ {t \\in \\Theta , | \\Theta | = k} A _ {t}, \\quad \\Theta \\subset \\{1, 2, \\dots , T \\}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 522, + 837, + 890, + 868 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "14743", + "bbox": [ + 465, + 885, + 504, + 897 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/395baac5f70b6598677f6663cca078ae6d31e0f1e20fd363883ca19a848f68ff.jpg", + "image_caption": [ + "Figure 2. Overall framework of the proposed CELL for OWTAL. Firstly, a Multi-scale Extended-range Perception (MEP) module is designed to collect the initial evidence by perceiving temporal contexts in diverse ranges. Then we propose a Knowledge-guided Bipolar Prototype Learning (KBPL) module to obtain evidence calibration factors from inter-video correlations and prior knowledge. Finally, a Cascade Evidence Enhancement (CEE) module is designed for enhancing the initial evidence with the calibration factors, and the uncertainty derived from the cascaded evidence is used for the known/unknown judgment." + ], + "image_footnote": [], + "bbox": [ + 159, + 123, + 806, + 372 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $k = \\lceil T / r \\rceil$ , $r$ is a scaling factor. According to the snippet index in $\\Theta$ , we select them from $\\tilde{\\pmb{F}}$ and calculate the average as the video-level feature, denoted as $\\tilde{F}_{vid}$ .", + "bbox": [ + 75, + 448, + 468, + 491 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "With the multi-scale-context-aware video-level feature, we feed $\\tilde{\\pmb{F}}_{vid}$ into an evidence collector $f$ , a DNN parameterized by $\\theta$ , to collect initial evidence $e_{init} \\in \\mathbb{R}^C$ :", + "bbox": [ + 75, + 491, + 468, + 533 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {e} _ {\\text {i n i t}} = g \\left(f \\left(\\tilde {\\boldsymbol {F}} _ {v i d}; \\boldsymbol {\\theta}\\right)\\right), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 194, + 543, + 468, + 560 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $g$ denotes a scale function, e.g., ReLU, Softplus or Exp, to ensure the collected evidence non-negative.", + "bbox": [ + 75, + 571, + 468, + 599 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Based on Subjective Logic theory [22], Evidential Deep Learning (EDL) allows uncertainty estimation in a single forward pass [41] by collecting evidence of each category and parameterizing a Dirichlet distribution, which models the distribution of class probabilities, over the collected evidence. Our collected initial evidence $e_{init}$ can be optimized by the following loss function:", + "bbox": [ + 75, + 599, + 468, + 695 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {i n i t}} = \\sum_ {i = 1} ^ {C} y _ {i} \\left(\\log S _ {\\text {i n i t}} - \\log \\alpha_ {\\text {i n i t}, i}\\right), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 142, + 707, + 468, + 745 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $S_{init} = \\sum_{i}^{C}\\alpha_{init,i},\\alpha_{init,i} = e_{init,i} + 1$ , and Eq. (6) is actually the deformation of the negative logarithm of the marginal likelihood in the EDL paradigm [35]. Besides, the video-level classification uncertainty $u_{init}$ can be derived by the EDL theory as $u_{init} = C / S_{init}$ . Note that uncertainty $u_{init}$ is inversely proportional to the total evidence of all closed-set categories, thus it reflects the probability that the video contains unknown actions.", + "bbox": [ + 75, + 757, + 470, + 869 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Knowledge-guided Bipolar Prototype Learning for Evidence Calibration Factors", + "text_level": 1, + "bbox": [ + 499, + 448, + 890, + 477 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Due to the differences in action components, scenes, shooting angles and other aspects, videos tend to keep a large intra-action variation in visual patterns, and the known action instances which visually deviate from the majority are likely to be mistakenly identified as unknown ones. We argue that the initial evidence collected by a single video is insufficient for handling this issue. Therefore, we resort to the bipolar prototype learning strategy, in which both positive and negative prototypes are set for each category, to explore inter-video correlations for mining the intrinsic information of actions and eliminating the interference caused by background snippets. However, even if the intra-action visual variance is mitigated to some extent, it is still challenging to effectively detect the unknown action instances in the open-world environment for the lack of prior knowledge. To tackle the problem, a semantic relation graph is constructed to simulate the open world, so as to provide prior knowledge guidance for the bipolar prototype learning among videos, finally obtaining more accurate evidence calibration factors for further cascade evidence enhancement.", + "bbox": [ + 497, + 486, + 890, + 763 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Based on the above observations, as shown in Figure 2, we propose a Knowledge-guided Bipolar Prototype Learning module (KBPL) to obtain evidence calibration factors from inter-video correlations and prior knowledge. Firstly, we construct a semantic relation graph with concepts related to known categories as vertie and their relations as edges. Specifically, the vertie set $\\mathcal{N}$ includes concepts of known", + "bbox": [ + 497, + 769, + 892, + 866 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "14744", + "bbox": [ + 465, + 885, + 505, + 897 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "classes, denoted as $\\mathcal{N}_{kno}$ , and the neighbor action concepts in ConceptNet [39] with the top- $p$ relation strength, denoted as $\\mathcal{N}_{nei}$ , thus $\\mathcal{N} = \\mathcal{N}_{kno} \\bigcup \\mathcal{N}_{nei}$ . Following a standard graph construction operation [13, 15], edges are connected in our graph if the relation strength between two concepts (vertex) is larger than a threshold. The features of vertex in $\\mathcal{N}$ are initialized by the Glove-300 [32] embedding vectors, denoted as $\\{n_j\\}_{j=1}^{|\\mathcal{N}|}$ , and then updated by:", + "bbox": [ + 75, + 125, + 472, + 239 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {\\boldsymbol {n}} _ {j} = h _ {2} \\left(\\underset {z \\in \\{1, 2, 3 \\}} {\\operatorname {C o n c a t}} \\left(\\sum_ {l \\in \\mathcal {S} _ {j, z}} \\delta \\left(\\eta_ {j, l}\\right) h _ {1} \\left(\\boldsymbol {n} _ {l}\\right)\\right)\\right) + h _ {3} \\left(\\boldsymbol {n} _ {j}\\right), \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 94, + 250, + 468, + 281 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $h_{1,2,3}$ denote fully-connected layers, $\\delta(\\cdot)$ denotes the Softmax operation, $S_{j,z}$ denotes the index set of the $z$ -hop neighbor vertex of node $j$ , and $\\eta_{j,l}$ represents the cosine similarity of $\\mathbf{n}_l$ and $\\mathbf{n}_j$ . Although many other complicated and advanced graph network methods can be adopted, we select the direct update manner here for simplicity.", + "bbox": [ + 75, + 290, + 468, + 372 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "With the concept relation graph, for category $i$ , we use the updated representation of its corresponding vertex to generate the positive prototype, i.e. $Q_{i}^{pos} = \\phi_{1}(n_{i})$ , and the average updated representations of its neighbor vertex are adopted for producing the negative one, i.e. $Q_{i}^{neg} = \\phi_{2}(\\sum_{j\\in S_{i}}n_{j} / |S_{i}|)$ , where $\\phi_{1,2}$ are fully-connected layers for prototype embedding. Then, for video $\\nu_{b}$ in the minibatch, where $b = 1,\\dots,B$ , and $B$ is the mini-batch size, we define the calibration factor $\\xi_{b,i}$ by simultaneously considering the distances between the video-level feature $\\tilde{F}_{vid,b}$ and the bipolar prototypes $Q_{i}^{pos}$ and $Q_{i}^{neg}$ as:", + "bbox": [ + 75, + 372, + 468, + 526 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\xi_ {b, i} = \\hat {D} (\\tilde {\\boldsymbol {F}} _ {v i d, b}, \\boldsymbol {Q} _ {i} ^ {n e g}) - \\hat {D} (\\tilde {\\boldsymbol {F}} _ {v i d, b}, \\boldsymbol {Q} _ {i} ^ {p o s}), \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 129, + 536, + 468, + 553 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\hat{D}$ is a distance calculation method, which can be Euclidean distance or negative cosine similarity.", + "bbox": [ + 75, + 562, + 468, + 591 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To guide the learning of prototypes by inter-video correlations, for each category, we pull close the videos belonging to the category to the corresponding positive prototype, and push away the videos of other classes in the same minibatch, while opposite operations are performed for the negative prototype. Specifically, the loss function $\\mathcal{L}_{kbpl}$ guiding the pos-neg prototype learning is designed as:", + "bbox": [ + 75, + 591, + 468, + 688 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {k b p l} = - \\frac {1}{C} \\sum_ {i = 1} ^ {C} \\mathbb {I} \\left(\\mathcal {B} _ {i} \\neq \\emptyset\\right) \\log \\left(\\frac {\\sum_ {b \\in \\mathcal {B} _ {i}} \\exp \\left(\\xi_ {b , i}\\right)}{\\sum_ {b = 1} ^ {B} \\exp \\left(\\xi_ {b , i}\\right)}\\right), \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 95, + 698, + 468, + 745 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathbb{I}(\\cdot)$ is an indicator function, $\\mathcal{B}_i = \\{b|y_{b,i} = 1\\}$ , and $\\pmb{y}_b$ is the ground-truth label of the video $\\mathcal{V}_b$ .", + "bbox": [ + 75, + 745, + 468, + 774 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Cascade Evidence Enhancement", + "text_level": 1, + "bbox": [ + 75, + 779, + 363, + 792 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As shown in Figure 2, we then perform cascade evidence enhancement, which scales the initial evidence collected from the MEP module by the learned calibration factors. As we stated above, the uncertainty $u_{init}$ represents the probability that a video contains unknown actions. For example,", + "bbox": [ + 75, + 796, + 468, + 866 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "given a test video, when $u_{init}$ is close to the extreme values (close to zero for known actions and close to the maximum for unknown ones), we can recognize known or unknown classes with high confidence [35], otherwise the model has relatively low confidence in the known/unknown judgment. Specifically, we design an uncertainty-aware adaptor $\\zeta$ as a Gaussian-like estimation function for evidence cascade:", + "bbox": [ + 497, + 125, + 890, + 220 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\zeta \\left(u _ {i n i t}; \\mu , \\sigma\\right) = \\exp \\left(- \\frac {\\left(u _ {i n i t} - \\mu\\right) ^ {2}}{2 \\sigma^ {2}}\\right), \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 563, + 232, + 890, + 262 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mu$ is set to the value beyond the $u_{init}$ of $95\\%$ training videos, and $\\sigma$ is a learnable scalar parameter which controls the shape of the Gaussian-like function.", + "bbox": [ + 497, + 273, + 890, + 314 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "If the above uncertainty-aware adaptor $\\zeta$ of an input video is high, it means it is difficult for the model to perform known/unknown judgment by only considering the initial evidence collected inside the video, thus it is necessary to employ the information provided by the inter-video correlations to carry out evidence calibration. Specifically, the cascade evidence calibration process can be formulated as:", + "bbox": [ + 497, + 315, + 890, + 412 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\ne _ {c a s c, i} = (1 + \\zeta (u _ {i n i t}) \\tanh (\\xi_ {i})) e _ {i n i t, i}, \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 563, + 427, + 890, + 440 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where the $\\tanh (\\cdot)$ function ensures the scaling range of $e_{init,i}$ to be (0,2) for a stable learning, and $e_{casc} = [e_{casc,1},\\dots,e_{casc,C}] \\in \\mathbb{R}^C$ is the enhanced cascaded evidence. The subscript $b$ (video index) is omitted in this equation for simplicity. $e_{casc}$ is then optimized by an EDL loss $\\mathcal{L}_{casc}$ , whose form is identical with Eq. (6), and the cascaded video-level uncertainty $u_{casc}$ for known/unknown judgment can be obtained in the same way with $u_{init}$ .", + "bbox": [ + 497, + 453, + 890, + 564 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.5. Training and Inference", + "text_level": 1, + "bbox": [ + 499, + 572, + 712, + 586 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Training. By combining all the optimization objectives introduced above, we obtain the final loss function as:", + "bbox": [ + 499, + 591, + 889, + 617 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\mathcal {L} _ {c l s} + \\mathcal {L} _ {i n i t} + \\mathcal {L} _ {k b p l} + \\mathcal {L} _ {c a s c}, \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 579, + 633, + 890, + 646 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathcal{L}_{cls}$ is the standard video-level classification loss, e.g., MIL loss [10, 19], for optimizing the backbone.", + "bbox": [ + 499, + 659, + 889, + 686 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Inference. For each video, we predict its action categories set $\\Psi$ (including the Unknown class if exist) by Algorithm 1, where the closed-set video-level classification score $P = [P_{1},\\dots,P_{C}]$ is obtained by averaging the classification activation sequence (CAS) of snippets in $\\Theta$ . Then we use a threshold strategy to obtain action snippet candidates following the standard process [18, 19]. Finally, for each action class in $\\Psi$ we group continuous snippets into action proposals, whose confidence scores $\\gamma$ are estimated according to the CAS scores of the corresponding class (for Known classes) or attention scores (for the Unknown class), and perform non-maximum-suppression (NMS) to remove duplicated proposals.", + "bbox": [ + 499, + 687, + 890, + 866 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "14745", + "bbox": [ + 465, + 885, + 504, + 897 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/14954c6395ddb2b6309aaa03faf3929d6b45319701eacd5559d5039e8d3b1488.jpg", + "table_caption": [ + "Table 1. OWTAL localization results on THUMOS-14 evaluated by both top- $K$ mAP@Avg and traditional mAP@Avg. The averages of the mAP of Known classes and the AP of the Unknown class are reported. The values are averaged on t-IoU [0.1:0.1:0.5] and [0.1:0.1:0.7]." + ], + "table_footnote": [], + "table_body": "
MethodsTop-K mAP@Avg(%)mAP@Avg(%)
Top-5Top-10Top-20Top-50Top-100
0.1-0.50.1-0.70.1-0.50.1-0.70.1-0.50.1-0.70.1-0.50.1-0.70.1-0.50.1-0.70.1-0.50.1-0.7
ASM-Loc + Trivial4.33.57.86.313.310.721.116.926.120.929.523.6
ASM-Loc + SoftMax10.98.916.913.523.418.728.823.029.823.830.124.0
ASM-Loc + OpenMax10.18.115.312.121.216.726.120.627.321.527.621.7
ASM-Loc + ARPL10.08.216.413.423.619.330.424.832.026.032.426.4
ASM-Loc + EDL11.39.217.414.024.219.430.524.331.925.432.225.7
ASM-Loc + CELL(Ours)11.29.318.014.625.520.732.426.434.127.834.728.1
CO2-Net + Trivial5.54.49.57.716.713.525.920.930.925.034.427.9
CO2-Net + SoftMax11.39.117.814.325.120.232.226.033.727.334.227.8
CO2-Net + OpenMax10.38.416.313.223.018.629.123.530.424.730.825.0
CO2-Net + ARPL11.69.518.314.925.720.933.327.135.128.735.729.2
CO2-Net + EDL11.29.117.614.324.820.032.226.034.027.534.628.1
CO2-Net + CELL(Ours)12.610.320.116.428.123.036.930.338.931.839.532.3
", + "bbox": [ + 80, + 150, + 902, + 382 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/507031e04b7331bfda72caedcab9cf2607faac1bea16683ce55a54ba3fb8fd3d.jpg", + "table_caption": [ + "Table 2. OWTAL localization results on THUMOS-14 evaluated by both top- $K$ mAP@Avg and traditional mAP@Avg. The mAP of Known classes (K) and the AP of the Unknown class (U) averaged on t-IoU thresholds [0.1:0.1:0.7] are reported respectively. The video-level known-unknown classification accuracy (v-Acc(%) ) is also presented for reference." + ], + "table_footnote": [], + "table_body": "
MethodsTop-K mAP@Avg(%)mAP@Avg(%)v-Acc(%)
Top-5Top-10Top-20Top-50Top-100
KUKUKUKUKUKU
CO2-Net + Trivial4.44.47.87.615.111.923.818.030.519.436.219.628.1
CO2-Net + SoftMax12.65.619.39.326.413.932.719.234.520.035.620.071.2
CO2-Net + OpenMax11.94.918.48.025.112.031.016.132.716.733.616.570.5
CO2-Net + ARPL12.36.719.410.426.615.232.621.734.622.735.822.673.0
CO2-Net + EDL12.45.719.19.426.313.832.719.434.820.235.920.370.9
CO2-Net + CELL(Ours)13.37.320.512.227.618.334.026.635.927.636.827.874.9
", + "bbox": [ + 90, + 434, + 878, + 590 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experimental Results", + "text_level": 1, + "bbox": [ + 75, + 599, + 280, + 615 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Experimental Setup", + "text_level": 1, + "bbox": [ + 75, + 621, + 265, + 636 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Datasets. We conduct experiments on two popular benchmarks for OWTAL, THUMOS-14 and ActivityNet-v1.3. THUMOS-14 contains 200 validation videos and 213 test videos annotated from 20 action categories, and each video contains 15.4 action instances on average. ActivityNet-v1.3 contains 10,024 training videos and 4,926 validation videos from 200 action categories, and each video contains 1.6 action instances on average. Following the pioneer work [3], we randomly remove 1/4 action categories of THUMOS-14 training set and keep the entire THUMOS-14 testing set to enable open-set evaluation, and such random removal is repeated three times. To further increase the openness and verify the applicability of our proposed method for open-world scenarios, we train our model on the known split of THUMOS-14 and test on the testing set of ActivityNet-v1.3. Note that since ActivityNet-v1.3 covers", + "bbox": [ + 75, + 644, + 470, + 866 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "most categories in THUMOS-14, we manually remove 14 action categories which are semantically overlapping with the THUMOS-14 as previous work did.", + "bbox": [ + 497, + 600, + 890, + 642 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation Metrics. We adopt three sets of metrics to measure the performance of the OWTAL task. (1) Vanilla mAP. To evaluate the localization performance, following previous approaches [10, 19, 33], we employ the mean Average Precision (mAP) under different temporal Intersection over Union (t-IoU) thresholds as metrics. The t-IoU thresholds for THUMOS-14 is [0.1:0.1:0.7] and for ActivityNetv1.3 is [0.5:0.05:0.95]. (2) Top- $K$ mAP. In our experiments, we observe that it only brings slight performance loss on the traditional mAP for models to generate a large number of redundant proposals. However, in real-world applications our goal is to provide users with accurate action proposals of a limited number, rather than an extreme largescale proposal set consisting of numerous redundant proposals, although the set may contain even more accurate predictions. Inspired by the above observation, we design a", + "bbox": [ + 497, + 644, + 892, + 866 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "14746", + "bbox": [ + 465, + 885, + 505, + 897 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Algorithm 1 Action Categories Inference Procedure", + "text_level": 1, + "bbox": [ + 78, + 125, + 423, + 138 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Input: Untrimmed testing video $\\nu$", + "bbox": [ + 78, + 141, + 312, + 154 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Require: Trained CELL model.", + "bbox": [ + 78, + 155, + 290, + 168 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Require: Threshold $\\tau$ obtained from training data (Please refer to Implementation Details).", + "bbox": [ + 78, + 169, + 468, + 195 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Output: Set $\\Psi$ of action categories in $\\mathcal{V}$ .", + "bbox": [ + 78, + 195, + 346, + 209 + ], + "page_idx": 6 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1: Predict the closed-set classification score $P$ and cascaded video-level uncertainty $u_{\\text{casc}}$ by CELL.", + "2: if $u_{\\text{casc}} < \\tau$ then", + "3: $\\Psi = \\{i|P_i > 0.2\\}$ ▷OnlyKnownClasses", + "4: if $\\Psi = \\emptyset$ then", + "5: $\\Psi = \\arg \\max_{i} P_{i}$ ▷OnlyKnownClasses", + "6: end if", + "7: else if $\\arg \\max_{i} P_{i} > 0.5$ then", + "8: $\\Psi = \\{\\arg \\max_{i} P_{i}, C + 1\\}$", + "9: Both Known and Unknown Classes", + "10: else", + "11: $\\Psi = \\{C + 1\\}$ ▷OnlyUnknownClasses", + "12: end if", + "13: return $\\Psi$" + ], + "bbox": [ + 80, + 212, + 468, + 403 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "variant metric of the vanilla mAP to further meet the needs of real-world applications, termed as top- $K$ mAP, which restricts the maximum number of generated proposals by selecting proposals with top- $K$ confidence scores. When $K$ is not restricted, the top- $K$ mAP degrades to the traditional mAP metric. In our experiments $K$ is set to 5, 10, 20, 50, 100 for THUMOS-14 and 1, 3, 5 for ActivityNet-v1.3, respectively. (3) Classification metrics. For known/unknown classification, following [3], four classification metrics including the False Alarm Rate at True Positive Rate of $95\\%$ (FAR@95) (smaller value indicates better performance), the Area Under the Receiver Operating Characteristic (AUROC) curve, the Area Under the Precision-Recall (AUPR), and the Open Set Detection Rate (OSDR) are also employed. To validate the effectiveness of Algorithm 1, we additionally adopt the video-level known-unknown classification accuracy (v-Acc(\\%)) metric for evaluation. Note that a testing video may belong to three types which have: (i) only known actions, (ii) only unknown actions, or (iii) both known and unknown actions.", + "bbox": [ + 75, + 435, + 468, + 711 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Implementation Details. We adopt CO2-Net [19] and ASM-Loc [18] as our backbone networks. Following existing methods, we use I3D [5] model pretrained on the Kinetics [23] dataset to extract both the RGB and optical flow features. The feature dimension $D$ is 2048, the evidence scale function $g$ in Eq. (6) is Exp, and the distance calculation method $\\hat{D}$ in Eq. (8) is negative cosine similarity. The number of the sampled snippets $T$ for THUMOS-14 and ActivityNet-v1.3 is 500 and 160, and the scaling factor $r$ is 7. $\\xi$ in Eq. (10) is updated for every 25 epochs. The batch size is set to 10 and 16, and the learning rate is 5e-5 and 1e", + "bbox": [ + 75, + 714, + 468, + 866 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4 in CO2-Net and ASM-Loc, respectively. Following previous works [2, 40], we determine the threshold $\\tau$ of uncertainty $u_{\\text{casc}}$ in Algorithm 1 by ensuring $95\\%$ training videos to be recognized as known. Our model is implemented with Python 3.7 and PyTorch 1.11.0. All experiments are conducted on a single RTX3090 GPU.", + "bbox": [ + 497, + 125, + 890, + 207 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. Comparison with State-of-the-art Methods", + "text_level": 1, + "bbox": [ + 497, + 214, + 868, + 228 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Our proposed CELL method is compared with the following baselines based on the ASM-Loc [18] and the CO2-Net [19]: (1) Trivial: assuming that all testing videos contain both known and unknown actions without any known/unknown judgment. (2) SoftMax: utilizing the maximum softmax probability to identify the unknown. (3) OpenMax: appending the unknown score to the softmax scores by OpenMax [4] in testing. (4) ARPL: using the Adversarial Reciprocal Points Learning (ARPL) [7] to identify the unknown. (5) EDL: vanilla EDL is used to replace the softmax classification head for uncertainty estimation.", + "bbox": [ + 497, + 232, + 890, + 383 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Comparison results on THUMOS-14. Table 1 and Table 2 represent the OWTAL localization results on the THUMOS-14 dataset. The results show that our proposed method consistently outperforms the baselines by large margins on all metrics. We notice that totally without any known/unknown judgment, the Trivial method shows comparable performance with the SoftMax baseline on the traditional mAP metric, for the embarrassing property that it only brings slight performance loss on the traditional mAP to generate numerous error redundant proposals, while on our proposed top- $k$ mAP metrics the performance of the Trivial version shows a reasonable gap. The results also show that Openmax does not work well on the OWTAL task, which may lie in that Openmax often fails to recognize visually indistinguishable samples [16], especially in weakly-supervised video analysis field. Note that the other two SOTA approaches ARPL and EDL perform well but still far behind the proposed CELL.", + "bbox": [ + 497, + 384, + 890, + 631 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 3 displays the known/unknown classification results evaluated on the metrics adopted by [3], showing that our method outperforms the EDL baseline and even achieves comparable results with fully-supervised methods on FAR@95, AUROC, and AUPR. We infer that the large gap between weakly-supervised methods and fully-supervised ones on OSDR is due to the difference between backbone networks. OSDR is defined as the area under the curve of Correct Detection Rate (CDR) and False Positive Rate (FPR), and the CDR indicates the fraction of known actions which are positively localized and correctly classified, while the FPR denotes the fraction of unknown actions that are positively localized but falsely classified. CO2-Net adopts the localization-by-classification strategy commonly used in WTAL method, thus the positively localized actions are usually accompanied by correct class predictions, while AFSD [26] adopts an anchor-free method whose localize", + "bbox": [ + 497, + 631, + 890, + 866 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "14747", + "bbox": [ + 465, + 885, + 504, + 897 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "tion performance does not entirely rely on the classification results, thus resulting in the large gap on OSDR. Note that compared with our employed mAP-related metrics, OSDR is still a classification metric thus cannot fully exhibit the superiority of different models.", + "bbox": [ + 75, + 125, + 470, + 195 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/51d7830257195f0f41c4cf81500fe33bf728ffaaaeddebf2f814b5d556010148.jpg", + "table_caption": [ + "Table 3. OWTAL classification results on THUMOS-14 evaluated by FAR@95, AUROC, AUPR and OSDR. In this experiment we use CO2-Net as our backbone." + ], + "table_footnote": [], + "table_body": "
SupervisionMethodsFAR@95(↓)AUROCAUPROSDR
FullySoftMax85.5854.7031.8523.40
OpenMax90.3453.2633.1713.66
EDL81.4264.0540.0536.26
OpenTAL70.9678.3358.6242.91
WeaklyCO2+EDL87.3767.1144.4763.56
CELL(Ours)68.8674.6254.2569.81
", + "bbox": [ + 80, + 256, + 468, + 357 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/45f17487ecd280aaa3f781ba27854155493d9e5004b7be397d896944bedb01c4.jpg", + "table_caption": [ + "Table 4. OWTAL localization results on ActivityNet-v1.3. The values are averaged on t-IoU thresholds [0.5:0.05:0.95]. The metric v-Acc(%) is also presented for reference." + ], + "table_footnote": [], + "table_body": "
MethodsTop-K mAP@Avg(%)mAP@Avg(%)v-Acc(%)
Top1Top3Top5
ASM-Loc + SoftMax12.713.213.513.586.37
ASM-Loc + OpenMax12.012.612.812.981.74
ASM-Loc + ARPL17.017.617.818.096.81
ASM-Loc + EDL16.016.716.917.093.38
ASM-Loc + CELL(Ours)17.918.418.618.997.83
", + "bbox": [ + 80, + 427, + 468, + 524 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/e9e3f02adf708b68d66b952abdc19bfb0f54c61dc54b68c2ead31246fe2ccc7e.jpg", + "table_caption": [ + "Table 5. Ablation study of our proposed CELL. MEP denotes the multi-scale extended-range perception module, KBPL represents the knowledge-guided bipolar prototype learning module, and UA denotes the uncertainty-aware adaptor $\\zeta$ described in Eq. (10)." + ], + "table_footnote": [], + "table_body": "
ExpMEPKBPLUATop-k mAP@Avg(%)mAP@Avg(%)
Top-10Top-20Top-50
1XXX14.320.226.027.8
2✓XX15.321.226.929.0
3X✓X15.221.126.929.2
4✓✓X16.022.429.431.6
5✓✓✓16.423.030.332.3
", + "bbox": [ + 87, + 606, + 457, + 705 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Comparison results on ActivityNet-v1.3. To further increase the openness and verify the applicability of our proposed method for open-world scenarios, we further train our model on the known splits of THUMOS-14 and test on ActivityNet-v1.3. Although the testing set only consists of unknown action categories, in order to ensure a fair comparison, the model is still tested according to Algorithm 1, without using any additional prior knowledge. As shown in Table 4, our proposed CELL obtains consistently favorable performance. Compared with EDL, CELL shows superior performance with an absolute gain of $1.9\\%$ in mAP@Avg.", + "bbox": [ + 75, + 714, + 468, + 866 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.3. Ablation Study", + "text_level": 1, + "bbox": [ + 499, + 125, + 652, + 140 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This section shows the effectiveness of our modules. In Table 5, the impact of progressively adding each component is presented, proving their contributions clearly. According to the cascade steps of our method, we perform ablation study on three components, Multi-scale Extended-range Perception module (MEP), Knowledge-guided Bipolar Prototype Learning module (KBPL), and the Uncertainty-aware Adaptor (UA) in CEE module. Note that without MEP we replace the fused feature by the average of snippet features in $\\Theta$ , and without UA we simply set $\\zeta$ in Eq. (11) to 1. As shown in Table 5, every step of our method brings effective performance improvement. Moreover, it is noteworthy that MEP perceiving temporal contexts in multiscale ranges and KBPL leveraging inter-video correlations and prior knowledge can enhance and complement each other, thus significantly improving the performance.", + "bbox": [ + 497, + 144, + 890, + 366 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusions", + "text_level": 1, + "bbox": [ + 499, + 374, + 626, + 388 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Targeting at open-world weakly-supervised temporal action localization (OWTAL), we propose a cascade evidential learning framework, which entails three main components: (1) Multi-scale extended-range perception module perceiving temporal contexts in diverse ranges for collecting initial evidence; (2) Knowledge-guided bipolar prototype learning strategy exploring inter-video relations under the guidance of prior knowledge for seeking supplementary evidence support; (3) Cascade evidence enhancement for final evidence calibration. In our extensive experiments, CELL achieves state-of-the-art performance on various metrics. Several limitations of this work are noteworthy. Firstly, never-before-seen human action categories are constantly emerging, thus lifelong learning paradigm which can keep embracing new action categories is even more suitable for OWTAL. Secondly, the class annotation granularity of the training set will interfere with the unknown action identification in open-world scenarios. Specifically, if some fine-grained actions are not labeled during training, e.g., the run-up movement in HighJump videos in THUMOS14, the model may misjudge such unknown actions as the background in testing. We assume the problem can be alleviated by training on large-scale datasets with fine-grained annotations. These directions are left as future work.", + "bbox": [ + 497, + 395, + 890, + 726 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 499, + 736, + 665, + 751 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This work was supported by the National Key Research & Development Plan of China under Grant 2020AAA0106200, in part by the National Natural Science Foundation of China under Grants 62036012, U21B2044, 62236008, 62102415, 61721004, 62072286, 62072455, 62002355, in part by Beijing Natural Science Foundation (L201001), and in part by Open Research Projects of Zhejiang Lab (NO.2022RC0AB02).", + "bbox": [ + 497, + 755, + 890, + 866 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "14748", + "bbox": [ + 465, + 885, + 505, + 897 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 124, + 171, + 138 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Alexander Amini, Wilko Schwarting, Ava Soleimany, and Daniela Rus. Deep evidential regression. In NeurIPS, 2020. 3", + "[2] Wentao Bao, Qi Yu, and Yu Kong. Evidential deep learning for open set action recognition. In ICCV, 2021. 3, 7", + "[3] Wentao Bao, Qi Yu, and Yu Kong. Opental: Towards open set temporal action localization. In CVPR, 2022. 1, 2, 3, 6, 7", + "[4] Abhijit Bendale and Terrance E Boult. Towards open set deep networks. In CVPR, 2016. 2, 7", + "[5] Joao Carreira and Andrew Zisserman. Quo vadis, action recognition? a new model and the kinetics dataset. In CVPR, 2017. 7", + "[6] Yu-Wei Chao, Sudheendra Vijayanarasimhan, Bryan Seybold, David A Ross, Jia Deng, and Rahul Sukthankar. Rethinking the faster r-cnn architecture for temporal action localization. In CVPR, 2018. 3", + "[7] Guangyao Chen, Peixi Peng, Xiangqian Wang, and Yonghong Tian. Adversarial reciprocal points learning for open set recognition. arXiv preprint arXiv:2103.00953, 2021. 2, 7", + "[8] Guangyao Chen, Limeng Qiao, Yemin Shi, Peixi Peng, Jia Li, Tiejun Huang, Shiliang Pu, and Yonghong Tian. Learning open set network with discriminative reciprocal points. In ECCV. Springer, 2020. 2", + "[9] Mengyuan Chen, Junyu Gao, Shicai Yang, and Changsheng Xu. Dual-evidential learning for weakly-supervised temporal action localization. In ECCV. Springer, 2022. 1, 2, 3", + "[10] Junyu Gao, Mengyuan Chen, and Changsheng Xu. Fine-grained temporal contrastive learning for weakly-supervised temporal action localization. In CVPR, 2022. 2, 3, 5, 6", + "[11] Junyu Gao and Changsheng Xu. Fast video moment retrieval. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1523-1532, 2021. 1", + "[12] Junyu Gao, Tianzhu Zhang, and Changsheng Xu. Graph convolutional tracking. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 4649-4659, 2019. 3", + "[13] Junyu Gao, Tianzhu Zhang, and Changsheng Xu. I know the relationships: Zero-shot action recognition via two-stream graph convolutional networks and knowledge graphs. In AAAI, 2019. 5", + "[14] Junyu Gao, Tianzhu Zhang, and Changsheng Xu. Smart: Joint sampling and regression for visual tracking. IEEE Transactions on Image Processing, 28(8):3923-3935, 2019. 3", + "[15] Junyu Gao, Tianzhu Zhang, and Changsheng Xu. Learning to model relationships for zero-shot video classification. IEEE Transactions on Pattern Analysis and Machine Intelligence, 43(10):3476-3491, 2021. 5", + "[16] Chuanxing Geng, Sheng-jun Huang, and Songcan Chen. Recent advances in open set recognition: A survey. IEEE transactions on pattern analysis and machine intelligence, 43(10):3614-3631, 2020. 7", + "[17] Zongbo Han, Changqing Zhang, Huazhu Fu, and Joey Tianyi Zhou. Trusted multi-view classification. In ICLR, 2020. 3" + ], + "bbox": [ + 78, + 147, + 468, + 855 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[18] Bo He, Xitong Yang, Le Kang, Zhiyu Cheng, Xin Zhou, and Abhinav Shrivastava. Asm-loc: Action-aware segment modeling for weakly-supervised temporal action localization. In CVPR, 2022. 1, 5, 7", + "[19] Fa-Ting Hong, Jia-Chang Feng, Dan Xu, Ying Shan, and Wei-Shi Zheng. Cross-modal consensus network for weakly supervised temporal action localization. In ACM MM, 2021. 1, 2, 3, 5, 6, 7", + "[20] Yibo Hu, Yuzhe Ou, Xujiang Zhao, Jin-Hee Cho, and Feng Chen. Multidimensional uncertainty-aware evidential neural networks. In AAAI, 2021. 3", + "[21] Linjiang Huang, Liang Wang, and Hongsheng Li. Weakly supervised temporal action localization via representative snippet knowledge propagation. In CVPR, 2022. 2", + "[22] Audun Jsang. Subjective logic: A formalism for reasoning under uncertainty. Springer Verlag, 2016. 3, 4", + "[23] Will Kay, Joao Carreira, Karen Simonyan, Brian Zhang, Chloe Hillier, Sudheendra Vijayanarasimhan, Fabio Viola, Tim Green, Trevor Back, Paul Natsev, et al. The kinetics human action video dataset. arXiv:1705.06950, 2017. 3, 7", + "[24] Pilhyeon Lee, Jinglu Wang, Yan Lu, and Hyeran Byun. Weakly-supervised temporal action localization by uncertainty modeling. In AAAI, 2021. 2", + "[25] Bolian Li, Zongbo Han, Haining Li, Huazhu Fu, and Changqing Zhang. Trustworthy long-tailed classification. arXiv preprint arXiv:2111.09030, 2021. 3", + "[26] Chuming Lin, Chengming Xu, Donghao Luo, Yabiao Wang, Ying Tai, Chengjie Wang, Jilin Li, Feiyue Huang, and Yanwei Fu. Learning salient boundary feature for anchor-free temporal action localization. In CVPR, 2021. 7", + "[27] Wei Liu, Xiaodong Yue, Yufei Chen, and Thierry Denoeux. Trusted multi-view deep learning with opinion aggregation. In AAAI, 2022. 3", + "[28] Vincent Lonij, Ambrish Rawat, and Maria-Irina Nicolae. Open-world visual recognition using knowledge graphs. arXiv preprint arXiv:1708.08310, 2017. 2", + "[29] Wang Luo, Tianzhu Zhang, Wenfei Yang, Jingen Liu, Tao Mei, Feng Wu, and Yongdong Zhang. Action unit memory network for weakly supervised temporal action localization. In CVPR, 2021. 2", + "[30] Lawrence Neal, Matthew Olson, Xiaoli Fern, Weng-Keen Wong, and Fuxin Li. Open set learning with counterfactual images. In ECCV, 2018. 2", + "[31] Poojan Oza and Vishal M Patel. C2ae: Class conditioned auto-encoder for open-set recognition. In CVPR, 2019. 2", + "[32] Jeffrey Pennington, Richard Socher, and Christopher D Manning. Glove: Global vectors for word representation. In EMNLP, 2014. 5", + "[33] Sanqing Qu, Guang Chen, Zhijun Li, Lijun Zhang, Fan Lu, and Alois Knoll. Acm-net: Action context modeling network for weakly-supervised temporal action localization. arXiv preprint arXiv:2104.02967, 2021. 2, 3, 6", + "[34] Walter J Scheirer, Anderson de Rezende Rocha, Archana Sapkota, and Terrance E Boult. Toward open set recognition. IEEE transactions on pattern analysis and machine intelligence, 35(7):1757-1772, 2012. 2" + ], + "bbox": [ + 502, + 125, + 890, + 852 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "14749", + "bbox": [ + 465, + 887, + 504, + 897 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[35] Murat Sensoy, Lance Kaplan, and Melih Kandemir. Evidential deep learning to quantify classification uncertainty. In NeurIPS, 2018. 3, 4, 5", + "[36] Baifeng Shi, Qi Dai, Yadong Mu, and Jingdong Wang. Weakly-supervised action localization by generative attention modeling. In CVPR, 2020. 2", + "[37] Zheng Shou, Hang Gao, Lei Zhang, Kazuyuki Miyazawa, and Shih-Fu Chang. Autoloc: Weakly-supervised temporal action localization in untrimmed videos. In ECCV, 2018. 1", + "[38] Krishna Kumar Singh and Yong Jae Lee. Hide-and-seek: Forcing a network to be meticulous for weakly-supervised object and action localization. In ICCV, 2017. 2", + "[39] Robyn Speer, Joshua Chin, and Catherine Havasi. Concept-net 5.5: An open multilingual graph of general knowledge. In AAAI, 2017. 5", + "[40] Xin Sun, Zhenning Yang, Chi Zhang, Keck-Voon Ling, and Guohao Peng. Conditional gaussian distribution learning for open set recognition. In CVPR, 2020. 2, 7", + "[41] Dennis Ulmer. A survey on evidential deep learning for single-pass uncertainty estimation. arXiv preprint arXiv:2110.03051, 2021. 3, 4", + "[42] Sagar Vaze, Kai Han, Andrea Vedaldi, and Andrew Zisserman. Open-set recognition: A good closed-set classifier is all you need. In ICLR, 2022. 1", + "[43] Limin Wang, Yuanjun Xiong, Dahua Lin, and Luc Van Gool. Untrimmednets for weakly supervised action recognition and detection. In CVPR, 2017. 1, 2", + "[44] Huijuan Xu, Abir Das, and Kate Saenko. R-c3d: Region convolutional 3d network for temporal activity detection. In ICCV, 2017. 3" + ], + "bbox": [ + 78, + 125, + 468, + 505 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[45] Ronald R Yager and Liping Liu. Classic works of the Dempster-Shafer theory of belief functions, volume 219. Springer, 2008. 3", + "[46] Wenfei Yang, Tianzhu Zhang, Xiaoyuan Yu, Tian Qi, Yongdong Zhang, and Feng Wu. Uncertainty guided collaborative training for weakly supervised temporal action detection. In CVPR, 2021. 2, 3", + "[47] Ryota Yoshihashi, Wen Shao, Rei Kawakami, Shaodi You, Makoto Iida, and Takeshi Naemura. Classification-reconstruction learning for open-set recognition. In CVPR, 2019. 2", + "[48] Can Zhang, Meng Cao, Dongming Yang, Jie Chen, and Yuexian Zou. Cola: Weakly-supervised temporal action localization with snippet contrastive learning. In CVPR, 2021. 2", + "[49] Chengwei Zhang, Yunlu Xu, Zhanzhan Cheng, Yi Niu, Shiliang Pu, Fei Wu, and Futai Zou. Adversarial seeded sequence growing for weakly-supervised temporal action localization. In ACM MM, 2019. 2", + "[50] Jia-Xing Zhong, Nannan Li, Weijie Kong, Tao Zhang, Thomas H Li, and Ge Li. Step-by-step erasion, one-by-one collection: a weakly supervised temporal action detector. In ACM MM, 2018. 2", + "[51] Zhun Zhong, Linchao Zhu, Zhiming Luo, Shaozi Li, Yi Yang, and Nicu Sebe. Openmix: Reviving known knowledge for discovering novel visual categories in an open world. In CVPR, 2021. 2" + ], + "bbox": [ + 502, + 127, + 890, + 473 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "14750", + "bbox": [ + 465, + 885, + 504, + 897 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2023/Cascade Evidential Learning for Open-World Weakly-Supervised Temporal Action Localization/3c2843a4-d81f-4df7-8b2c-5c29eaeafad1_model.json b/2023/Cascade Evidential Learning for Open-World Weakly-Supervised Temporal Action Localization/3c2843a4-d81f-4df7-8b2c-5c29eaeafad1_model.json new file mode 100644 index 0000000000000000000000000000000000000000..d926c94119af8a8255ffa2f6ace83b7057c7bd02 --- /dev/null +++ b/2023/Cascade Evidential Learning for Open-World Weakly-Supervised Temporal Action Localization/3c2843a4-d81f-4df7-8b2c-5c29eaeafad1_model.json @@ -0,0 +1,2189 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.13, + 0.051, + 0.201, + 0.086 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.253, + 0.049, + 0.792, + 0.088 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.108, + 0.161, + 0.865, + 0.205 + ], + "angle": 0, + "content": "Cascade Evidential Learning for Open-world Weakly-supervised Temporal Action Localization" + }, + { + "type": "text", + "bbox": [ + 0.253, + 0.228, + 0.717, + 0.246 + ], + "angle": 0, + "content": "Mengyuan Chen\\(^{1,2}\\), Junyu Gao\\(^{1,2}\\), and Changsheng Xu\\(^{1,2,3}\\)" + }, + { + "type": "text", + "bbox": [ + 0.178, + 0.246, + 0.794, + 0.277 + ], + "angle": 0, + "content": "1 State Key Laboratory of Multimodal Artificial Intelligence Systems (MAIS), Institute of Automation, Chinese Academy of Sciences (CASIA)" + }, + { + "type": "text", + "bbox": [ + 0.141, + 0.277, + 0.83, + 0.293 + ], + "angle": 0, + "content": "\\(^{2}\\) School of Artificial Intelligence, University of Chinese Academy of Sciences (UCAS)" + }, + { + "type": "text", + "bbox": [ + 0.312, + 0.293, + 0.662, + 0.309 + ], + "angle": 0, + "content": "\\(^{3}\\) Peng Cheng Laboratory, ShenZhen, China" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.312, + 0.739, + 0.326 + ], + "angle": 0, + "content": "chenmengyuan2021@ia.ac.cn; { junyu.gao, csxu}@nlpr.ia.ac.cn" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.358, + 0.314, + 0.373 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.389, + 0.474, + 0.707 + ], + "angle": 0, + "content": "Targeting at recognizing and localizing action instances with only video-level labels during training, Weakly-supervised Temporal Action Localization (WTAL) has achieved significant progress in recent years. However, living in the dynamically changing open world where unknown actions constantly spring up, the closed-set assumption of existing WTAL methods is invalid. Compared with traditional open-set recognition tasks, Open-world WTAL (OW-TAL) is challenging since not only are the annotations of unknown samples unavailable, but also the fine-grained annotations of known action instances can only be inferred ambiguously from the video category labels. To address this problem, we propose a Cascade Evidential Learning framework at an evidence level, which targets at OWTAL for the first time. Our method jointly leverages multi-scale temporal contexts and knowledge-guided prototype information to progressively collect cascade and enhanced evidence for known action, unknown action, and background separation. Extensive experiments conducted on THUMOS-14 and ActivityNet-v1.3 verify the effectiveness of our method. Besides the classification metrics adopted by previous open-set recognition methods, we also evaluate our method on localization metrics which are more reasonable for OWTAL." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.748, + 0.21, + 0.763 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.77, + 0.471, + 0.868 + ], + "angle": 0, + "content": "Targeting at recognizing and localizing action instances with only video-level labels during training, Weakly-supervised Temporal Action Localization (WTAL) has attracted increasing attention from both academia and industry [9, 11, 18, 19, 37, 43]. Unlike fully-supervised TAL, WTAL only requires video-level action labels during training. However, the closed-set assumption of existing WTAL" + }, + { + "type": "image_caption", + "bbox": [ + 0.554, + 0.357, + 0.645, + 0.369 + ], + "angle": 0, + "content": "Training Phase" + }, + { + "type": "image", + "bbox": [ + 0.504, + 0.369, + 0.693, + 0.459 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.756, + 0.357, + 0.842, + 0.368 + ], + "angle": 0, + "content": "Testing Phase" + }, + { + "type": "image", + "bbox": [ + 0.697, + 0.369, + 0.89, + 0.46 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.468, + 0.893, + 0.507 + ], + "angle": 0, + "content": "Figure 1. Illustration of the training and testing phases of OWTAL. With only video-level labels for training, OWTAL aims to localize both known and unknown action instances in testing videos." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.518, + 0.893, + 0.588 + ], + "angle": 0, + "content": "methods is invalid in the dynamically changing real world, since with the development of society never-before-seen human action categories are constantly emerging. Therefore, to address this problem, we consider a different WTAL setting in this work, termed as Open-world WTAL (OWTAL)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.591, + 0.895, + 0.868 + ], + "angle": 0, + "content": "Different from the traditional WTAL task, as shown in Figure 1, OWTAL allows testing videos to contain action instances of unknown categories, which have never appeared during training. Therefore, temporal boundaries of both known and unknown action instances are expected to be predicted. Compared with its fully-supervised counterpart Open Set TAL [3], OWTAL is challenging in two aspects: (1) Ambiguity of annotations of closed-set (known) action instances. Previous works indicate that the closed-set and open-set performance are highly correlated [42]. However, under the OWTAL setting, not only are the annotations of unknown action instances unavailable, but also the fine-grained annotations of known ones can only be inferred ambiguously from the video category labels. During training, the known action instances that the model needs to focus on are prone to be disturbed by the background snippets, which hinders the learning of the closed-set actions, thus making it extremely difficult to differentiate the unknown actions, the known actions, and the background. (2) Lack of reasonable metrics. The traditional Open Set Recognition (OSR) aims" + }, + { + "type": "page_number", + "bbox": [ + 0.466, + 0.887, + 0.505, + 0.898 + ], + "angle": 0, + "content": "14741" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.127, + 0.47, + 0.168 + ], + "angle": 0, + "content": "for classification while the goal of OWTAL is to perform localization instead, thus the classification metrics commonly adopted by OSR are not sufficient for OWTAL." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.17, + 0.473, + 0.667 + ], + "angle": 0, + "content": "In order to alleviate the negative impact caused by the weak annotations of known action instances, we propose a Cascade Evidential Learning method for owtaL (CELL), which progressively collects cascaded evidence by considering both temporal contexts in multi-scale ranges and inter-video correlations under the guidance of prior knowledge. Since the goal of OWTAL is to locate the consecutive known/unknown action segments of various temporal lengths in open-world scenarios, perceiving temporal contexts in diverse ranges is essential. We argue that it is meaningful to endow individual snippet features with the ability of sensing multi-scale neighborhood video segments, and thus a Multi-scale Extended-range Perception module (Section 3.2) is designed to obtain more discriminative video features for initial evidence collection, taking advantage of the temporal contexts. Due to the large intra-action variation in visual patterns and the lack of prior knowledge guidance, the known action instances which visually deviate from the common ones are likely to be misidentified with the initial evidence collected from individual videos. Therefore, we design a Knowledge-guided Bipolar Prototype Learning strategy (Section 3.3), where a semantic relation graph is constructed to provide prior knowledge guidance for the bipolar prototype learning among videos, thus perceiving the open-world more comprehensively. We use this strategy to generate a series of evidence calibration factors for further cascade evidence enhancement. Finally, a Cascade Evidence Enhancement module (Section 3.4) is designed for enhancing the initial evidence with the calibration factors, and the uncertainty estimated from the cascaded evidence is used for the known/unknown judgment. Extensive experiments conducted on THUMOS-14 and ActivityNet verify the effectiveness. Besides the various classification metrics adopted by previous works, we also evaluate our method on localization metrics which are more in line with the needs of real-world applications." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.668, + 0.391, + 0.682 + ], + "angle": 0, + "content": "To summarize, our contribution is threefold:" + }, + { + "type": "text", + "bbox": [ + 0.085, + 0.685, + 0.47, + 0.767 + ], + "angle": 0, + "content": "- To tackle the unique challenges of OWTAL, we propose a cascade evidential learning framework, which progressively collects comprehensive evidence for known action, unknown action, and background separation. Localization metrics which meet the needs of the real-world more closely are adopted for evaluation." + }, + { + "type": "text", + "bbox": [ + 0.085, + 0.769, + 0.469, + 0.825 + ], + "angle": 0, + "content": "- To achieve OWTAL without fine-grained annotations, the proposed CELL jointly leverages multi-scale temporal contexts and knowledge-guided prototype information during the evidence cascade learning process." + }, + { + "type": "text", + "bbox": [ + 0.085, + 0.826, + 0.469, + 0.867 + ], + "angle": 0, + "content": "- We conduct comprehensive experiments on two popular WTAL benchmarks, THUMOS-14 and ActivityNetv1.3, and achieve significant performance improvement" + }, + { + "type": "list", + "bbox": [ + 0.085, + 0.685, + 0.47, + 0.867 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.522, + 0.127, + 0.892, + 0.168 + ], + "angle": 0, + "content": "over various baselines. Experiments show that CELL enables existing methods to well adapt to the more practical open-world settings." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.181, + 0.642, + 0.195 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.202, + 0.892, + 0.464 + ], + "angle": 0, + "content": "Weakly-supervised Temporal Action Localization. Originated from UntrimmedNet [43], the pioneer work to utilize video-level action category annotations as the weak supervision for TAL, existing WTAL methods commonly adopt a multiple instance learning strategy and can be roughly categorized into erasing-based, attention-based and uncertainty-based methods. Erasing-based methods [38, 49, 50] erase the most discriminative segments to mitigate the single snippet cheating issue [48]. Attention-based methods [10, 19, 21, 29, 33, 36] employ the attention mechanism to select snippets most likely to be the foreground with activation scores. The newly-arising uncertainty-based methods [9, 24, 46] utilize uncertainty to address the inevitable action-background ambiguity caused by the weakly-supervised setting. However, all existing WTAL methods are rooted in the closed-set assumption, which impedes their application to real-world scenarios. In contrast, we address the WTAL task under the open-world setting by collecting reliable cascaded evidence, which has not been explored in prior work." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.466, + 0.892, + 0.727 + ], + "angle": 0, + "content": "Open Set Recognition. Open Set Recognition (OSR) aims to recognize known classes and reject the unknown. [34] first formalize the OSR problem and provide a basic framework. [4] propose the first DNN-based OSR method Openmax, which rejects unknown classes by modeling the distance of activation vectors with Extreme Value Theory (EVT). GAN-based methods are also developed towards this task. [30] generate samples similar to the training data but not belonging to the known classes, and then utilize the generated unknown class samples to train an open-set classifier. Other approaches include prototype-based methods [7, 8], which reject unknown classes by calculating the maximum distance between the input sample and the learned closed-set prototypes, and reconstruction-based methods [31, 40, 47] utilize the reconstruction error in the test phase as an open-set indicator. Some works [28, 51] attempt to seek assistance from external knowledge, but few of them have explicitly modeled unknown class information in an end-to-end framework." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.729, + 0.892, + 0.867 + ], + "angle": 0, + "content": "Most existing approaches focus on open-set recognition tasks while few works attend to temporal localization-related vision field. [3] formalize the open-set temporal action localization problem and propose an OpenTAL framework. However, its fully-supervised setting requires large amounts of fine-grained labels whose annotation is time-consuming, error-prone, and costly. Besides, the evaluation metrics in [3] are mainly classification ones. In contrast, in this paper, we address the challenging OWTAL problem which also abandons the closed-set assumption but only re" + }, + { + "type": "page_number", + "bbox": [ + 0.466, + 0.887, + 0.506, + 0.898 + ], + "angle": 0, + "content": "14742" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.126, + 0.468, + 0.154 + ], + "angle": 0, + "content": "quires easily available video-level class labels, and employ comprehensive localization-related metrics for evaluation." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.155, + 0.47, + 0.417 + ], + "angle": 0, + "content": "Evidential Deep Learning. Based on Dempster-Shafer Theory (DST) [45] and Subjective Logic theory (SL) [22], Evidential Deep Learning (EDL) allows uncertainty estimation in a single forward pass [41] by collecting scalar evidence for each category and parameterizing a Dirichlet distribution, which models the distribution of classification probabilities, over the collected evidence. As a newly arising trustworthy method, EDL has achieved remarkable progress in various computer vision tasks, including action localization [9], image classification [35], regression [1], multi-view classification [17, 27], out-of-distribution detection [20], long-tail learning [25], and open-set action recognition and localization [2, 3]. However, current EDL algorithms only focus on collecting evidence from individual samples, while they neglect the complementary role that the potential correlations among samples can play in evidence collection. To address this issue, in this work we propose a cascade evidence calibration paradigm which effectively enhances the collected evidence." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.429, + 0.226, + 0.445 + ], + "angle": 0, + "content": "3. Our Approach" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.446, + 0.329, + 0.46 + ], + "angle": 0, + "content": "3.1. Notations and Preliminaries" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.466, + 0.469, + 0.644 + ], + "angle": 0, + "content": "Problem formulation of OWTAL. OWTAL targets at detecting known/unknown action instances in untrimmed videos. Formally, we are given a training set \\(\\{\\mathcal{V},\\mathcal{Y}\\}\\) with \\(N\\) training videos, where \\(\\mathcal{V}\\) denotes an untrimmed video, and \\(\\pmb {y}\\in \\mathbb{R}^{C}\\) is its multi-hot label indicating the action categories that the action instances in this video belong to, where \\(C\\) is the number of known action categories. During testing, the goal is to use the learned model to output a set of quadruplets \\(\\{c_r,t_r^s,t_r^e,\\gamma_r\\}_{r = 1}^R\\) ,where \\(R\\) is the number of action instances in \\(\\mathcal{V}\\) \\(c_{r},t_{r}^{s},t_{r}^{e}\\) and \\(\\gamma_{r}\\) denote the action category, the start and end timestamps, and the confidence score, respectively. \\(c_{r}\\in \\{1,\\dots,C + 1\\}\\) , where \\(C + 1\\) denotes the unknown action category." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.646, + 0.469, + 0.756 + ], + "angle": 0, + "content": "Feature extraction. Following previous works [10,46], we split the untrimmed video \\(\\mathcal{V}\\) into \\(T\\) non-overlapping 16-frame snippets and then use pre-trained networks, e.g., the I3D model [23], to extract RGB and optical flow features. Then, the above features are concatenated and then fed into a fusion module, e.g., convolutional layers [19, 33], to obtain the snippet-wise feature \\(\\pmb{X} = [\\pmb{x}_1,\\dots,\\pmb{x}_T]\\in \\mathbb{R}^{T\\times D}\\), where \\(D\\) is the feature dimension." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.765, + 0.469, + 0.793 + ], + "angle": 0, + "content": "3.2. Multi-scale Extended-range Perception for Initial Evidence Collection" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.798, + 0.469, + 0.867 + ], + "angle": 0, + "content": "A video may contain several known/unknown action instances with various temporal lengths in the open-world. Since the goal of OWTAL is to locate these consecutive video segments, perceiving temporal contexts in diverse ranges is essential. We argue that it is meaningful to endow" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.126, + 0.892, + 0.263 + ], + "angle": 0, + "content": "individual snippet features with the ability of sensing multiscale neighborhood video segments. As shown in Figure 2, we design a Multi-scale Extended-range Perception module (MEP) to obtain more discriminative video features, taking advantage of the temporal contexts. Note that some anchor-based temporal modeling methods [6, 12, 14, 44] also leverage multi-scale temporal information. However, their goal is to obtain the fused temporal feature for proposal generation, while the MEP aims to enhance individual snippet features for initial evidence collection." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.264, + 0.893, + 0.403 + ], + "angle": 0, + "content": "In MEP, each snippet \\( t \\) is assigned with a series of \\( t \\)-centered multi-scale video segments, whose context-aware features can contribute to the extension of snippet-specific perception ranges. Here, the temporal boundaries of the \\( t \\)-centered video segments can be denoted as set \\( \\Omega_t = \\{(s_t^m, e_t^m) | m \\in \\mathcal{M}\\} \\), where \\( \\mathcal{M} \\) is a pre-defined length set, \\( s_t^m = \\max(0, t - m) \\), and \\( e_t^m = \\min(T, t + m) \\). For brevity, we denote the temporal average of the feature sequence of the extended video segment \\( (s_t^m, e_t^m) \\) as \\( f_t^m \\in \\mathbb{R}^D \\), which can be given by:" + }, + { + "type": "equation", + "bbox": [ + 0.582, + 0.407, + 0.892, + 0.442 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {f} _ {t} ^ {m} = \\frac {1}{e _ {t} ^ {m} - s _ {t} ^ {m} + 1} \\sum_ {s _ {t} ^ {m} \\leq i \\leq e _ {t} ^ {m}} \\boldsymbol {x} _ {i}. \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.448, + 0.892, + 0.573 + ], + "angle": 0, + "content": "To extend the perception range of snippet \\(t\\), we perform feature enhancement by fusing \\(\\boldsymbol{x}_t\\) with \\(\\boldsymbol{f}_t^m\\), during which the weight of \\(\\boldsymbol{f}_t^m\\) is positively correlated with the similarity between them. Such operation is motivated by the simple expectation that a snippet-level feature would absorb meaningful contextual information mainly from extended video segments that are highly relevant to it. Therefore, the snippet-level feature with extended-range perception, denoted as \\(\\tilde{\\boldsymbol{F}} = [\\tilde{\\boldsymbol{f}}_1,\\dots ,\\tilde{\\boldsymbol{f}}_T]\\), can be formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.523, + 0.579, + 0.891, + 0.608 + ], + "angle": 0, + "content": "\\[\n\\tilde {\\boldsymbol {f}} _ {t} = \\left(1 - \\alpha_ {t}\\right) \\varphi_ {1} (\\boldsymbol {x} _ {t}) + \\alpha_ {t} \\sum_ {m \\in \\mathcal {M}} \\delta \\left(\\omega_ {t} ^ {m}\\right) \\varphi_ {2} \\left(\\boldsymbol {f} _ {t} ^ {m}\\right), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.614, + 0.891, + 0.684 + ], + "angle": 0, + "content": "where \\(\\delta (\\cdot)\\) denotes the Softmax operation with a temperature factor, \\(\\varphi_{1,2}\\) are fully-connected layers for feature embedding, \\(\\alpha_{t} \\in [0,1]\\) is the fusion weight representing the scaled average of the cosine similarities \\(\\omega_{t}^{m}\\) between \\(\\pmb{x}_{t}\\) and \\(\\pmb{f}_t^m\\), and \\(\\alpha_{t}\\) and \\(\\omega_t^m\\) can be calculated by:" + }, + { + "type": "equation", + "bbox": [ + 0.509, + 0.689, + 0.891, + 0.735 + ], + "angle": 0, + "content": "\\[\n\\alpha_ {t} = \\frac {1}{2 | \\mathcal {M} |} \\sum_ {m \\in \\mathcal {M}} \\left(\\omega_ {t} ^ {m} + 1\\right), \\omega_ {t} ^ {m} = \\cos \\left(\\frac {\\boldsymbol {x} _ {t}}{\\left| \\boldsymbol {x} _ {t} \\right|}, \\frac {\\boldsymbol {f} _ {t} ^ {m}}{\\left| \\boldsymbol {f} _ {t} ^ {m} \\right|}\\right). \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.736, + 0.892, + 0.832 + ], + "angle": 0, + "content": "Thereafter, similar to previous methods [10, 19], we utilize an attention module to predict an attention score sequence \\(\\mathbf{A} = [A_{1},\\dots,A_{T}]\\in \\mathbb{R}^{T}\\) for the input video \\(\\nu\\) representing the probabilities of snippets belonging to the foreground, then select the top-ranked snippets. Specifically, for the input video we obtain the set \\(\\Theta\\) of snippets with the top-\\(k\\) attention scores by the following equation:" + }, + { + "type": "equation", + "bbox": [ + 0.523, + 0.838, + 0.891, + 0.869 + ], + "angle": 0, + "content": "\\[\n\\Theta = \\arg \\max _ {\\Theta} \\sum_ {t \\in \\Theta , | \\Theta | = k} A _ {t}, \\quad \\Theta \\subset \\{1, 2, \\dots , T \\}, \\tag {4}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.466, + 0.887, + 0.505, + 0.898 + ], + "angle": 0, + "content": "14743" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.161, + 0.124, + 0.808, + 0.374 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.378, + 0.893, + 0.442 + ], + "angle": 0, + "content": "Figure 2. Overall framework of the proposed CELL for OWTAL. Firstly, a Multi-scale Extended-range Perception (MEP) module is designed to collect the initial evidence by perceiving temporal contexts in diverse ranges. Then we propose a Knowledge-guided Bipolar Prototype Learning (KBPL) module to obtain evidence calibration factors from inter-video correlations and prior knowledge. Finally, a Cascade Evidence Enhancement (CEE) module is designed for enhancing the initial evidence with the calibration factors, and the uncertainty derived from the cascaded evidence is used for the known/unknown judgment." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.45, + 0.47, + 0.492 + ], + "angle": 0, + "content": "where \\( k = \\lceil T / r \\rceil \\), \\( r \\) is a scaling factor. According to the snippet index in \\( \\Theta \\), we select them from \\( \\tilde{\\pmb{F}} \\) and calculate the average as the video-level feature, denoted as \\( \\tilde{F}_{vid} \\)." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.492, + 0.47, + 0.534 + ], + "angle": 0, + "content": "With the multi-scale-context-aware video-level feature, we feed \\(\\tilde{\\pmb{F}}_{vid}\\) into an evidence collector \\(f\\), a DNN parameterized by \\(\\theta\\), to collect initial evidence \\(e_{init} \\in \\mathbb{R}^C\\):" + }, + { + "type": "equation", + "bbox": [ + 0.196, + 0.545, + 0.47, + 0.561 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {e} _ {\\text {i n i t}} = g \\left(f \\left(\\tilde {\\boldsymbol {F}} _ {v i d}; \\boldsymbol {\\theta}\\right)\\right), \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.572, + 0.469, + 0.6 + ], + "angle": 0, + "content": "where \\(g\\) denotes a scale function, e.g., ReLU, Softplus or Exp, to ensure the collected evidence non-negative." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.6, + 0.47, + 0.697 + ], + "angle": 0, + "content": "Based on Subjective Logic theory [22], Evidential Deep Learning (EDL) allows uncertainty estimation in a single forward pass [41] by collecting evidence of each category and parameterizing a Dirichlet distribution, which models the distribution of class probabilities, over the collected evidence. Our collected initial evidence \\( e_{init} \\) can be optimized by the following loss function:" + }, + { + "type": "equation", + "bbox": [ + 0.143, + 0.708, + 0.47, + 0.746 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {i n i t}} = \\sum_ {i = 1} ^ {C} y _ {i} \\left(\\log S _ {\\text {i n i t}} - \\log \\alpha_ {\\text {i n i t}, i}\\right), \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.758, + 0.471, + 0.87 + ], + "angle": 0, + "content": "where \\( S_{init} = \\sum_{i}^{C}\\alpha_{init,i},\\alpha_{init,i} = e_{init,i} + 1 \\), and Eq. (6) is actually the deformation of the negative logarithm of the marginal likelihood in the EDL paradigm [35]. Besides, the video-level classification uncertainty \\( u_{init} \\) can be derived by the EDL theory as \\( u_{init} = C / S_{init} \\). Note that uncertainty \\( u_{init} \\) is inversely proportional to the total evidence of all closed-set categories, thus it reflects the probability that the video contains unknown actions." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.449, + 0.892, + 0.478 + ], + "angle": 0, + "content": "3.3. Knowledge-guided Bipolar Prototype Learning for Evidence Calibration Factors" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.488, + 0.892, + 0.764 + ], + "angle": 0, + "content": "Due to the differences in action components, scenes, shooting angles and other aspects, videos tend to keep a large intra-action variation in visual patterns, and the known action instances which visually deviate from the majority are likely to be mistakenly identified as unknown ones. We argue that the initial evidence collected by a single video is insufficient for handling this issue. Therefore, we resort to the bipolar prototype learning strategy, in which both positive and negative prototypes are set for each category, to explore inter-video correlations for mining the intrinsic information of actions and eliminating the interference caused by background snippets. However, even if the intra-action visual variance is mitigated to some extent, it is still challenging to effectively detect the unknown action instances in the open-world environment for the lack of prior knowledge. To tackle the problem, a semantic relation graph is constructed to simulate the open world, so as to provide prior knowledge guidance for the bipolar prototype learning among videos, finally obtaining more accurate evidence calibration factors for further cascade evidence enhancement." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.77, + 0.893, + 0.868 + ], + "angle": 0, + "content": "Based on the above observations, as shown in Figure 2, we propose a Knowledge-guided Bipolar Prototype Learning module (KBPL) to obtain evidence calibration factors from inter-video correlations and prior knowledge. Firstly, we construct a semantic relation graph with concepts related to known categories as vertie and their relations as edges. Specifically, the vertie set \\(\\mathcal{N}\\) includes concepts of known" + }, + { + "type": "page_number", + "bbox": [ + 0.466, + 0.887, + 0.506, + 0.898 + ], + "angle": 0, + "content": "14744" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.126, + 0.473, + 0.241 + ], + "angle": 0, + "content": "classes, denoted as \\(\\mathcal{N}_{kno}\\), and the neighbor action concepts in ConceptNet [39] with the top-\\(p\\) relation strength, denoted as \\(\\mathcal{N}_{nei}\\), thus \\(\\mathcal{N} = \\mathcal{N}_{kno} \\bigcup \\mathcal{N}_{nei}\\). Following a standard graph construction operation [13, 15], edges are connected in our graph if the relation strength between two concepts (vertex) is larger than a threshold. The features of vertex in \\(\\mathcal{N}\\) are initialized by the Glove-300 [32] embedding vectors, denoted as \\(\\{n_j\\}_{j=1}^{|\\mathcal{N}|}\\), and then updated by:" + }, + { + "type": "equation", + "bbox": [ + 0.095, + 0.251, + 0.47, + 0.282 + ], + "angle": 0, + "content": "\\[\n\\tilde {\\boldsymbol {n}} _ {j} = h _ {2} \\left(\\underset {z \\in \\{1, 2, 3 \\}} {\\operatorname {C o n c a t}} \\left(\\sum_ {l \\in \\mathcal {S} _ {j, z}} \\delta \\left(\\eta_ {j, l}\\right) h _ {1} \\left(\\boldsymbol {n} _ {l}\\right)\\right)\\right) + h _ {3} \\left(\\boldsymbol {n} _ {j}\\right), \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.291, + 0.47, + 0.374 + ], + "angle": 0, + "content": "where \\( h_{1,2,3} \\) denote fully-connected layers, \\( \\delta(\\cdot) \\) denotes the Softmax operation, \\( S_{j,z} \\) denotes the index set of the \\( z \\)-hop neighbor vertex of node \\( j \\), and \\( \\eta_{j,l} \\) represents the cosine similarity of \\( \\mathbf{n}_l \\) and \\( \\mathbf{n}_j \\). Although many other complicated and advanced graph network methods can be adopted, we select the direct update manner here for simplicity." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.374, + 0.47, + 0.527 + ], + "angle": 0, + "content": "With the concept relation graph, for category \\(i\\), we use the updated representation of its corresponding vertex to generate the positive prototype, i.e. \\(Q_{i}^{pos} = \\phi_{1}(n_{i})\\), and the average updated representations of its neighbor vertex are adopted for producing the negative one, i.e. \\(Q_{i}^{neg} = \\phi_{2}(\\sum_{j\\in S_{i}}n_{j} / |S_{i}|)\\), where \\(\\phi_{1,2}\\) are fully-connected layers for prototype embedding. Then, for video \\(\\nu_{b}\\) in the minibatch, where \\(b = 1,\\dots,B\\), and \\(B\\) is the mini-batch size, we define the calibration factor \\(\\xi_{b,i}\\) by simultaneously considering the distances between the video-level feature \\(\\tilde{F}_{vid,b}\\) and the bipolar prototypes \\(Q_{i}^{pos}\\) and \\(Q_{i}^{neg}\\) as:" + }, + { + "type": "equation", + "bbox": [ + 0.13, + 0.537, + 0.469, + 0.554 + ], + "angle": 0, + "content": "\\[\n\\xi_ {b, i} = \\hat {D} (\\tilde {\\boldsymbol {F}} _ {v i d, b}, \\boldsymbol {Q} _ {i} ^ {n e g}) - \\hat {D} (\\tilde {\\boldsymbol {F}} _ {v i d, b}, \\boldsymbol {Q} _ {i} ^ {p o s}), \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.564, + 0.469, + 0.592 + ], + "angle": 0, + "content": "where \\(\\hat{D}\\) is a distance calculation method, which can be Euclidean distance or negative cosine similarity." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.592, + 0.469, + 0.689 + ], + "angle": 0, + "content": "To guide the learning of prototypes by inter-video correlations, for each category, we pull close the videos belonging to the category to the corresponding positive prototype, and push away the videos of other classes in the same minibatch, while opposite operations are performed for the negative prototype. Specifically, the loss function \\(\\mathcal{L}_{kbpl}\\) guiding the pos-neg prototype learning is designed as:" + }, + { + "type": "equation", + "bbox": [ + 0.097, + 0.699, + 0.469, + 0.747 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {k b p l} = - \\frac {1}{C} \\sum_ {i = 1} ^ {C} \\mathbb {I} \\left(\\mathcal {B} _ {i} \\neq \\emptyset\\right) \\log \\left(\\frac {\\sum_ {b \\in \\mathcal {B} _ {i}} \\exp \\left(\\xi_ {b , i}\\right)}{\\sum_ {b = 1} ^ {B} \\exp \\left(\\xi_ {b , i}\\right)}\\right), \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.747, + 0.469, + 0.775 + ], + "angle": 0, + "content": "where \\(\\mathbb{I}(\\cdot)\\) is an indicator function, \\(\\mathcal{B}_i = \\{b|y_{b,i} = 1\\}\\), and \\(\\pmb{y}_b\\) is the ground-truth label of the video \\(\\mathcal{V}_b\\)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.78, + 0.364, + 0.793 + ], + "angle": 0, + "content": "3.4. Cascade Evidence Enhancement" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.798, + 0.469, + 0.867 + ], + "angle": 0, + "content": "As shown in Figure 2, we then perform cascade evidence enhancement, which scales the initial evidence collected from the MEP module by the learned calibration factors. As we stated above, the uncertainty \\( u_{init} \\) represents the probability that a video contains unknown actions. For example," + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.127, + 0.892, + 0.222 + ], + "angle": 0, + "content": "given a test video, when \\( u_{init} \\) is close to the extreme values (close to zero for known actions and close to the maximum for unknown ones), we can recognize known or unknown classes with high confidence [35], otherwise the model has relatively low confidence in the known/unknown judgment. Specifically, we design an uncertainty-aware adaptor \\( \\zeta \\) as a Gaussian-like estimation function for evidence cascade:" + }, + { + "type": "equation", + "bbox": [ + 0.564, + 0.233, + 0.892, + 0.263 + ], + "angle": 0, + "content": "\\[\n\\zeta \\left(u _ {i n i t}; \\mu , \\sigma\\right) = \\exp \\left(- \\frac {\\left(u _ {i n i t} - \\mu\\right) ^ {2}}{2 \\sigma^ {2}}\\right), \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.274, + 0.891, + 0.315 + ], + "angle": 0, + "content": "where \\(\\mu\\) is set to the value beyond the \\(u_{init}\\) of \\(95\\%\\) training videos, and \\(\\sigma\\) is a learnable scalar parameter which controls the shape of the Gaussian-like function." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.317, + 0.892, + 0.413 + ], + "angle": 0, + "content": "If the above uncertainty-aware adaptor \\(\\zeta\\) of an input video is high, it means it is difficult for the model to perform known/unknown judgment by only considering the initial evidence collected inside the video, thus it is necessary to employ the information provided by the inter-video correlations to carry out evidence calibration. Specifically, the cascade evidence calibration process can be formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.564, + 0.428, + 0.891, + 0.441 + ], + "angle": 0, + "content": "\\[\ne _ {c a s c, i} = (1 + \\zeta (u _ {i n i t}) \\tanh (\\xi_ {i})) e _ {i n i t, i}, \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.454, + 0.892, + 0.565 + ], + "angle": 0, + "content": "where the \\(\\tanh (\\cdot)\\) function ensures the scaling range of \\(e_{init,i}\\) to be (0,2) for a stable learning, and \\(e_{casc} = [e_{casc,1},\\dots,e_{casc,C}] \\in \\mathbb{R}^C\\) is the enhanced cascaded evidence. The subscript \\(b\\) (video index) is omitted in this equation for simplicity. \\(e_{casc}\\) is then optimized by an EDL loss \\(\\mathcal{L}_{casc}\\), whose form is identical with Eq. (6), and the cascaded video-level uncertainty \\(u_{casc}\\) for known/unknown judgment can be obtained in the same way with \\(u_{init}\\)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.573, + 0.714, + 0.587 + ], + "angle": 0, + "content": "3.5. Training and Inference" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.592, + 0.89, + 0.618 + ], + "angle": 0, + "content": "Training. By combining all the optimization objectives introduced above, we obtain the final loss function as:" + }, + { + "type": "equation", + "bbox": [ + 0.58, + 0.634, + 0.891, + 0.647 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\mathcal {L} _ {c l s} + \\mathcal {L} _ {i n i t} + \\mathcal {L} _ {k b p l} + \\mathcal {L} _ {c a s c}, \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.66, + 0.89, + 0.687 + ], + "angle": 0, + "content": "where \\(\\mathcal{L}_{cls}\\) is the standard video-level classification loss, e.g., MIL loss [10, 19], for optimizing the backbone." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.688, + 0.892, + 0.868 + ], + "angle": 0, + "content": "Inference. For each video, we predict its action categories set \\(\\Psi\\) (including the Unknown class if exist) by Algorithm 1, where the closed-set video-level classification score \\(P = [P_{1},\\dots,P_{C}]\\) is obtained by averaging the classification activation sequence (CAS) of snippets in \\(\\Theta\\). Then we use a threshold strategy to obtain action snippet candidates following the standard process [18, 19]. Finally, for each action class in \\(\\Psi\\) we group continuous snippets into action proposals, whose confidence scores \\(\\gamma\\) are estimated according to the CAS scores of the corresponding class (for Known classes) or attention scores (for the Unknown class), and perform non-maximum-suppression (NMS) to remove duplicated proposals." + }, + { + "type": "page_number", + "bbox": [ + 0.466, + 0.887, + 0.505, + 0.898 + ], + "angle": 0, + "content": "14745" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.078, + 0.124, + 0.892, + 0.15 + ], + "angle": 0, + "content": "Table 1. OWTAL localization results on THUMOS-14 evaluated by both top- \\( K \\) mAP@Avg and traditional mAP@Avg. The averages of the mAP of Known classes and the AP of the Unknown class are reported. The values are averaged on t-IoU [0.1:0.1:0.5] and [0.1:0.1:0.7]." + }, + { + "type": "table", + "bbox": [ + 0.081, + 0.152, + 0.904, + 0.383 + ], + "angle": 0, + "content": "
MethodsTop-K mAP@Avg(%)mAP@Avg(%)
Top-5Top-10Top-20Top-50Top-100
0.1-0.50.1-0.70.1-0.50.1-0.70.1-0.50.1-0.70.1-0.50.1-0.70.1-0.50.1-0.70.1-0.50.1-0.7
ASM-Loc + Trivial4.33.57.86.313.310.721.116.926.120.929.523.6
ASM-Loc + SoftMax10.98.916.913.523.418.728.823.029.823.830.124.0
ASM-Loc + OpenMax10.18.115.312.121.216.726.120.627.321.527.621.7
ASM-Loc + ARPL10.08.216.413.423.619.330.424.832.026.032.426.4
ASM-Loc + EDL11.39.217.414.024.219.430.524.331.925.432.225.7
ASM-Loc + CELL(Ours)11.29.318.014.625.520.732.426.434.127.834.728.1
CO2-Net + Trivial5.54.49.57.716.713.525.920.930.925.034.427.9
CO2-Net + SoftMax11.39.117.814.325.120.232.226.033.727.334.227.8
CO2-Net + OpenMax10.38.416.313.223.018.629.123.530.424.730.825.0
CO2-Net + ARPL11.69.518.314.925.720.933.327.135.128.735.729.2
CO2-Net + EDL11.29.117.614.324.820.032.226.034.027.534.628.1
CO2-Net + CELL(Ours)12.610.320.116.428.123.036.930.338.931.839.532.3
" + }, + { + "type": "table_caption", + "bbox": [ + 0.078, + 0.395, + 0.893, + 0.433 + ], + "angle": 0, + "content": "Table 2. OWTAL localization results on THUMOS-14 evaluated by both top- \\( K \\) mAP@Avg and traditional mAP@Avg. The mAP of Known classes (K) and the AP of the Unknown class (U) averaged on t-IoU thresholds [0.1:0.1:0.7] are reported respectively. The video-level known-unknown classification accuracy (v-Acc(%) ) is also presented for reference." + }, + { + "type": "table", + "bbox": [ + 0.092, + 0.435, + 0.88, + 0.591 + ], + "angle": 0, + "content": "
MethodsTop-K mAP@Avg(%)mAP@Avg(%)v-Acc(%)
Top-5Top-10Top-20Top-50Top-100
KUKUKUKUKUKU
CO2-Net + Trivial4.44.47.87.615.111.923.818.030.519.436.219.628.1
CO2-Net + SoftMax12.65.619.39.326.413.932.719.234.520.035.620.071.2
CO2-Net + OpenMax11.94.918.48.025.112.031.016.132.716.733.616.570.5
CO2-Net + ARPL12.36.719.410.426.615.232.621.734.622.735.822.673.0
CO2-Net + EDL12.45.719.19.426.313.832.719.434.820.235.920.370.9
CO2-Net + CELL(Ours)13.37.320.512.227.618.334.026.635.927.636.827.874.9
" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.6, + 0.282, + 0.616 + ], + "angle": 0, + "content": "4. Experimental Results" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.622, + 0.267, + 0.637 + ], + "angle": 0, + "content": "4.1. Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.646, + 0.471, + 0.867 + ], + "angle": 0, + "content": "Datasets. We conduct experiments on two popular benchmarks for OWTAL, THUMOS-14 and ActivityNet-v1.3. THUMOS-14 contains 200 validation videos and 213 test videos annotated from 20 action categories, and each video contains 15.4 action instances on average. ActivityNet-v1.3 contains 10,024 training videos and 4,926 validation videos from 200 action categories, and each video contains 1.6 action instances on average. Following the pioneer work [3], we randomly remove 1/4 action categories of THUMOS-14 training set and keep the entire THUMOS-14 testing set to enable open-set evaluation, and such random removal is repeated three times. To further increase the openness and verify the applicability of our proposed method for open-world scenarios, we train our model on the known split of THUMOS-14 and test on the testing set of ActivityNet-v1.3. Note that since ActivityNet-v1.3 covers" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.601, + 0.892, + 0.643 + ], + "angle": 0, + "content": "most categories in THUMOS-14, we manually remove 14 action categories which are semantically overlapping with the THUMOS-14 as previous work did." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.646, + 0.893, + 0.868 + ], + "angle": 0, + "content": "Evaluation Metrics. We adopt three sets of metrics to measure the performance of the OWTAL task. (1) Vanilla mAP. To evaluate the localization performance, following previous approaches [10, 19, 33], we employ the mean Average Precision (mAP) under different temporal Intersection over Union (t-IoU) thresholds as metrics. The t-IoU thresholds for THUMOS-14 is [0.1:0.1:0.7] and for ActivityNetv1.3 is [0.5:0.05:0.95]. (2) Top-\\(K\\) mAP. In our experiments, we observe that it only brings slight performance loss on the traditional mAP for models to generate a large number of redundant proposals. However, in real-world applications our goal is to provide users with accurate action proposals of a limited number, rather than an extreme largescale proposal set consisting of numerous redundant proposals, although the set may contain even more accurate predictions. Inspired by the above observation, we design a" + }, + { + "type": "page_number", + "bbox": [ + 0.466, + 0.887, + 0.506, + 0.898 + ], + "angle": 0, + "content": "14746" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.126, + 0.425, + 0.14 + ], + "angle": 0, + "content": "Algorithm 1 Action Categories Inference Procedure" + }, + { + "type": "text", + "bbox": [ + 0.08, + 0.142, + 0.313, + 0.155 + ], + "angle": 0, + "content": "Input: Untrimmed testing video \\(\\nu\\)" + }, + { + "type": "text", + "bbox": [ + 0.08, + 0.156, + 0.291, + 0.169 + ], + "angle": 0, + "content": "Require: Trained CELL model." + }, + { + "type": "text", + "bbox": [ + 0.08, + 0.17, + 0.469, + 0.196 + ], + "angle": 0, + "content": "Require: Threshold \\(\\tau\\) obtained from training data (Please refer to Implementation Details)." + }, + { + "type": "text", + "bbox": [ + 0.08, + 0.197, + 0.347, + 0.21 + ], + "angle": 0, + "content": "Output: Set \\(\\Psi\\) of action categories in \\(\\mathcal{V}\\)." + }, + { + "type": "text", + "bbox": [ + 0.088, + 0.213, + 0.469, + 0.24 + ], + "angle": 0, + "content": "1: Predict the closed-set classification score \\( P \\) and cascaded video-level uncertainty \\( u_{\\text{casc}} \\) by CELL." + }, + { + "type": "text", + "bbox": [ + 0.088, + 0.242, + 0.225, + 0.254 + ], + "angle": 0, + "content": "2: if \\( u_{\\text{casc}} < \\tau \\) then" + }, + { + "type": "text", + "bbox": [ + 0.088, + 0.255, + 0.468, + 0.269 + ], + "angle": 0, + "content": "3: \\(\\Psi = \\{i|P_i > 0.2\\}\\) ▷OnlyKnownClasses" + }, + { + "type": "text", + "bbox": [ + 0.088, + 0.27, + 0.231, + 0.281 + ], + "angle": 0, + "content": "4: if \\(\\Psi = \\emptyset\\) then" + }, + { + "type": "text", + "bbox": [ + 0.088, + 0.283, + 0.468, + 0.295 + ], + "angle": 0, + "content": "5: \\(\\Psi = \\arg \\max_{i} P_{i}\\) ▷OnlyKnownClasses" + }, + { + "type": "text", + "bbox": [ + 0.088, + 0.297, + 0.175, + 0.307 + ], + "angle": 0, + "content": "6: end if" + }, + { + "type": "text", + "bbox": [ + 0.088, + 0.31, + 0.311, + 0.322 + ], + "angle": 0, + "content": "7: else if \\(\\arg \\max_{i} P_{i} > 0.5\\) then" + }, + { + "type": "text", + "bbox": [ + 0.088, + 0.324, + 0.312, + 0.337 + ], + "angle": 0, + "content": "8: \\(\\Psi = \\{\\arg \\max_{i} P_{i}, C + 1\\}\\)" + }, + { + "type": "text", + "bbox": [ + 0.088, + 0.338, + 0.468, + 0.35 + ], + "angle": 0, + "content": "9: Both Known and Unknown Classes" + }, + { + "type": "text", + "bbox": [ + 0.082, + 0.352, + 0.136, + 0.362 + ], + "angle": 0, + "content": "10: else" + }, + { + "type": "text", + "bbox": [ + 0.082, + 0.365, + 0.468, + 0.379 + ], + "angle": 0, + "content": "11: \\(\\Psi = \\{C + 1\\}\\) ▷OnlyUnknownClasses" + }, + { + "type": "text", + "bbox": [ + 0.082, + 0.38, + 0.151, + 0.391 + ], + "angle": 0, + "content": "12: end if" + }, + { + "type": "text", + "bbox": [ + 0.082, + 0.393, + 0.171, + 0.404 + ], + "angle": 0, + "content": "13: return \\(\\Psi\\)" + }, + { + "type": "list", + "bbox": [ + 0.082, + 0.213, + 0.469, + 0.404 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.437, + 0.469, + 0.712 + ], + "angle": 0, + "content": "variant metric of the vanilla mAP to further meet the needs of real-world applications, termed as top-\\(K\\) mAP, which restricts the maximum number of generated proposals by selecting proposals with top-\\(K\\) confidence scores. When \\(K\\) is not restricted, the top-\\(K\\) mAP degrades to the traditional mAP metric. In our experiments \\(K\\) is set to 5, 10, 20, 50, 100 for THUMOS-14 and 1, 3, 5 for ActivityNet-v1.3, respectively. (3) Classification metrics. For known/unknown classification, following [3], four classification metrics including the False Alarm Rate at True Positive Rate of \\(95\\%\\) (FAR@95) (smaller value indicates better performance), the Area Under the Receiver Operating Characteristic (AUROC) curve, the Area Under the Precision-Recall (AUPR), and the Open Set Detection Rate (OSDR) are also employed. To validate the effectiveness of Algorithm 1, we additionally adopt the video-level known-unknown classification accuracy (v-Acc(\\%)) metric for evaluation. Note that a testing video may belong to three types which have: (i) only known actions, (ii) only unknown actions, or (iii) both known and unknown actions." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.715, + 0.469, + 0.867 + ], + "angle": 0, + "content": "Implementation Details. We adopt CO2-Net [19] and ASM-Loc [18] as our backbone networks. Following existing methods, we use I3D [5] model pretrained on the Kinetics [23] dataset to extract both the RGB and optical flow features. The feature dimension \\( D \\) is 2048, the evidence scale function \\( g \\) in Eq. (6) is Exp, and the distance calculation method \\( \\hat{D} \\) in Eq. (8) is negative cosine similarity. The number of the sampled snippets \\( T \\) for THUMOS-14 and ActivityNet-v1.3 is 500 and 160, and the scaling factor \\( r \\) is 7. \\( \\xi \\) in Eq. (10) is updated for every 25 epochs. The batch size is set to 10 and 16, and the learning rate is 5e-5 and 1e" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.126, + 0.892, + 0.209 + ], + "angle": 0, + "content": "4 in CO2-Net and ASM-Loc, respectively. Following previous works [2, 40], we determine the threshold \\(\\tau\\) of uncertainty \\(u_{\\text{casc}}\\) in Algorithm 1 by ensuring \\(95\\%\\) training videos to be recognized as known. Our model is implemented with Python 3.7 and PyTorch 1.11.0. All experiments are conducted on a single RTX3090 GPU." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.215, + 0.869, + 0.229 + ], + "angle": 0, + "content": "4.2. Comparison with State-of-the-art Methods" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.233, + 0.892, + 0.384 + ], + "angle": 0, + "content": "Our proposed CELL method is compared with the following baselines based on the ASM-Loc [18] and the CO2-Net [19]: (1) Trivial: assuming that all testing videos contain both known and unknown actions without any known/unknown judgment. (2) SoftMax: utilizing the maximum softmax probability to identify the unknown. (3) OpenMax: appending the unknown score to the softmax scores by OpenMax [4] in testing. (4) ARPL: using the Adversarial Reciprocal Points Learning (ARPL) [7] to identify the unknown. (5) EDL: vanilla EDL is used to replace the softmax classification head for uncertainty estimation." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.385, + 0.892, + 0.632 + ], + "angle": 0, + "content": "Comparison results on THUMOS-14. Table 1 and Table 2 represent the OWTAL localization results on the THUMOS-14 dataset. The results show that our proposed method consistently outperforms the baselines by large margins on all metrics. We notice that totally without any known/unknown judgment, the Trivial method shows comparable performance with the SoftMax baseline on the traditional mAP metric, for the embarrassing property that it only brings slight performance loss on the traditional mAP to generate numerous error redundant proposals, while on our proposed top-\\(k\\) mAP metrics the performance of the Trivial version shows a reasonable gap. The results also show that Openmax does not work well on the OWTAL task, which may lie in that Openmax often fails to recognize visually indistinguishable samples [16], especially in weakly-supervised video analysis field. Note that the other two SOTA approaches ARPL and EDL perform well but still far behind the proposed CELL." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.633, + 0.892, + 0.867 + ], + "angle": 0, + "content": "Table 3 displays the known/unknown classification results evaluated on the metrics adopted by [3], showing that our method outperforms the EDL baseline and even achieves comparable results with fully-supervised methods on FAR@95, AUROC, and AUPR. We infer that the large gap between weakly-supervised methods and fully-supervised ones on OSDR is due to the difference between backbone networks. OSDR is defined as the area under the curve of Correct Detection Rate (CDR) and False Positive Rate (FPR), and the CDR indicates the fraction of known actions which are positively localized and correctly classified, while the FPR denotes the fraction of unknown actions that are positively localized but falsely classified. CO2-Net adopts the localization-by-classification strategy commonly used in WTAL method, thus the positively localized actions are usually accompanied by correct class predictions, while AFSD [26] adopts an anchor-free method whose localize" + }, + { + "type": "page_number", + "bbox": [ + 0.466, + 0.887, + 0.505, + 0.898 + ], + "angle": 0, + "content": "14747" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.126, + 0.471, + 0.196 + ], + "angle": 0, + "content": "tion performance does not entirely rely on the classification results, thus resulting in the large gap on OSDR. Note that compared with our employed mAP-related metrics, OSDR is still a classification metric thus cannot fully exhibit the superiority of different models." + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.213, + 0.47, + 0.252 + ], + "angle": 0, + "content": "Table 3. OWTAL classification results on THUMOS-14 evaluated by FAR@95, AUROC, AUPR and OSDR. In this experiment we use CO2-Net as our backbone." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.257, + 0.469, + 0.358 + ], + "angle": 0, + "content": "
SupervisionMethodsFAR@95(↓)AUROCAUPROSDR
FullySoftMax85.5854.7031.8523.40
OpenMax90.3453.2633.1713.66
EDL81.4264.0540.0536.26
OpenTAL70.9678.3358.6242.91
WeaklyCO2+EDL87.3767.1144.4763.56
CELL(Ours)68.8674.6254.2569.81
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.388, + 0.47, + 0.426 + ], + "angle": 0, + "content": "Table 4. OWTAL localization results on ActivityNet-v1.3. The values are averaged on t-IoU thresholds [0.5:0.05:0.95]. The metric v-Acc(%) is also presented for reference." + }, + { + "type": "table", + "bbox": [ + 0.081, + 0.428, + 0.47, + 0.525 + ], + "angle": 0, + "content": "
MethodsTop-K mAP@Avg(%)mAP@Avg(%)v-Acc(%)
Top1Top3Top5
ASM-Loc + SoftMax12.713.213.513.586.37
ASM-Loc + OpenMax12.012.612.812.981.74
ASM-Loc + ARPL17.017.617.818.096.81
ASM-Loc + EDL16.016.716.917.093.38
ASM-Loc + CELL(Ours)17.918.418.618.997.83
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.554, + 0.469, + 0.605 + ], + "angle": 0, + "content": "Table 5. Ablation study of our proposed CELL. MEP denotes the multi-scale extended-range perception module, KBPL represents the knowledge-guided bipolar prototype learning module, and UA denotes the uncertainty-aware adaptor \\(\\zeta\\) described in Eq. (10)." + }, + { + "type": "table", + "bbox": [ + 0.088, + 0.607, + 0.458, + 0.706 + ], + "angle": 0, + "content": "
ExpMEPKBPLUATop-k mAP@Avg(%)mAP@Avg(%)
Top-10Top-20Top-50
1XXX14.320.226.027.8
2✓XX15.321.226.929.0
3X✓X15.221.126.929.2
4✓✓X16.022.429.431.6
5✓✓✓16.423.030.332.3
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.716, + 0.47, + 0.868 + ], + "angle": 0, + "content": "Comparison results on ActivityNet-v1.3. To further increase the openness and verify the applicability of our proposed method for open-world scenarios, we further train our model on the known splits of THUMOS-14 and test on ActivityNet-v1.3. Although the testing set only consists of unknown action categories, in order to ensure a fair comparison, the model is still tested according to Algorithm 1, without using any additional prior knowledge. As shown in Table 4, our proposed CELL obtains consistently favorable performance. Compared with EDL, CELL shows superior performance with an absolute gain of \\(1.9\\%\\) in mAP@Avg." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.126, + 0.653, + 0.141 + ], + "angle": 0, + "content": "4.3. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.145, + 0.892, + 0.367 + ], + "angle": 0, + "content": "This section shows the effectiveness of our modules. In Table 5, the impact of progressively adding each component is presented, proving their contributions clearly. According to the cascade steps of our method, we perform ablation study on three components, Multi-scale Extended-range Perception module (MEP), Knowledge-guided Bipolar Prototype Learning module (KBPL), and the Uncertainty-aware Adaptor (UA) in CEE module. Note that without MEP we replace the fused feature by the average of snippet features in \\(\\Theta\\), and without UA we simply set \\(\\zeta\\) in Eq. (11) to 1. As shown in Table 5, every step of our method brings effective performance improvement. Moreover, it is noteworthy that MEP perceiving temporal contexts in multiscale ranges and KBPL leveraging inter-video correlations and prior knowledge can enhance and complement each other, thus significantly improving the performance." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.375, + 0.627, + 0.389 + ], + "angle": 0, + "content": "5. Conclusions" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.396, + 0.892, + 0.727 + ], + "angle": 0, + "content": "Targeting at open-world weakly-supervised temporal action localization (OWTAL), we propose a cascade evidential learning framework, which entails three main components: (1) Multi-scale extended-range perception module perceiving temporal contexts in diverse ranges for collecting initial evidence; (2) Knowledge-guided bipolar prototype learning strategy exploring inter-video relations under the guidance of prior knowledge for seeking supplementary evidence support; (3) Cascade evidence enhancement for final evidence calibration. In our extensive experiments, CELL achieves state-of-the-art performance on various metrics. Several limitations of this work are noteworthy. Firstly, never-before-seen human action categories are constantly emerging, thus lifelong learning paradigm which can keep embracing new action categories is even more suitable for OWTAL. Secondly, the class annotation granularity of the training set will interfere with the unknown action identification in open-world scenarios. Specifically, if some fine-grained actions are not labeled during training, e.g., the run-up movement in HighJump videos in THUMOS14, the model may misjudge such unknown actions as the background in testing. We assume the problem can be alleviated by training on large-scale datasets with fine-grained annotations. These directions are left as future work." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.737, + 0.667, + 0.752 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.756, + 0.892, + 0.868 + ], + "angle": 0, + "content": "This work was supported by the National Key Research & Development Plan of China under Grant 2020AAA0106200, in part by the National Natural Science Foundation of China under Grants 62036012, U21B2044, 62236008, 62102415, 61721004, 62072286, 62072455, 62002355, in part by Beijing Natural Science Foundation (L201001), and in part by Open Research Projects of Zhejiang Lab (NO.2022RC0AB02)." + }, + { + "type": "page_number", + "bbox": [ + 0.466, + 0.887, + 0.506, + 0.898 + ], + "angle": 0, + "content": "14748" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.125, + 0.173, + 0.139 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.148, + 0.469, + 0.185 + ], + "angle": 0, + "content": "[1] Alexander Amini, Wilko Schwarting, Ava Soleimany, and Daniela Rus. Deep evidential regression. In NeurIPS, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.187, + 0.469, + 0.212 + ], + "angle": 0, + "content": "[2] Wentao Bao, Qi Yu, and Yu Kong. Evidential deep learning for open set action recognition. In ICCV, 2021. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.213, + 0.469, + 0.239 + ], + "angle": 0, + "content": "[3] Wentao Bao, Qi Yu, and Yu Kong. Opental: Towards open set temporal action localization. In CVPR, 2022. 1, 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.24, + 0.469, + 0.265 + ], + "angle": 0, + "content": "[4] Abhijit Bendale and Terrance E Boult. Towards open set deep networks. In CVPR, 2016. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.266, + 0.469, + 0.302 + ], + "angle": 0, + "content": "[5] Joao Carreira and Andrew Zisserman. Quo vadis, action recognition? a new model and the kinetics dataset. In CVPR, 2017. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.304, + 0.469, + 0.354 + ], + "angle": 0, + "content": "[6] Yu-Wei Chao, Sudheendra Vijayanarasimhan, Bryan Seybold, David A Ross, Jia Deng, and Rahul Sukthankar. Rethinking the faster r-cnn architecture for temporal action localization. In CVPR, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.355, + 0.469, + 0.405 + ], + "angle": 0, + "content": "[7] Guangyao Chen, Peixi Peng, Xiangqian Wang, and Yonghong Tian. Adversarial reciprocal points learning for open set recognition. arXiv preprint arXiv:2103.00953, 2021. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.407, + 0.469, + 0.457 + ], + "angle": 0, + "content": "[8] Guangyao Chen, Limeng Qiao, Yemin Shi, Peixi Peng, Jia Li, Tiejun Huang, Shiliang Pu, and Yonghong Tian. Learning open set network with discriminative reciprocal points. In ECCV. Springer, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.459, + 0.469, + 0.496 + ], + "angle": 0, + "content": "[9] Mengyuan Chen, Junyu Gao, Shicai Yang, and Changsheng Xu. Dual-evidential learning for weakly-supervised temporal action localization. In ECCV. Springer, 2022. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.498, + 0.469, + 0.535 + ], + "angle": 0, + "content": "[10] Junyu Gao, Mengyuan Chen, and Changsheng Xu. Fine-grained temporal contrastive learning for weakly-supervised temporal action localization. In CVPR, 2022. 2, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.536, + 0.469, + 0.574 + ], + "angle": 0, + "content": "[11] Junyu Gao and Changsheng Xu. Fast video moment retrieval. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1523-1532, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.576, + 0.469, + 0.624 + ], + "angle": 0, + "content": "[12] Junyu Gao, Tianzhu Zhang, and Changsheng Xu. Graph convolutional tracking. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 4649-4659, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.626, + 0.469, + 0.675 + ], + "angle": 0, + "content": "[13] Junyu Gao, Tianzhu Zhang, and Changsheng Xu. I know the relationships: Zero-shot action recognition via two-stream graph convolutional networks and knowledge graphs. In AAAI, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.678, + 0.469, + 0.726 + ], + "angle": 0, + "content": "[14] Junyu Gao, Tianzhu Zhang, and Changsheng Xu. Smart: Joint sampling and regression for visual tracking. IEEE Transactions on Image Processing, 28(8):3923-3935, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.729, + 0.469, + 0.78 + ], + "angle": 0, + "content": "[15] Junyu Gao, Tianzhu Zhang, and Changsheng Xu. Learning to model relationships for zero-shot video classification. IEEE Transactions on Pattern Analysis and Machine Intelligence, 43(10):3476-3491, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.781, + 0.469, + 0.83 + ], + "angle": 0, + "content": "[16] Chuanxing Geng, Sheng-jun Huang, and Songcan Chen. Recent advances in open set recognition: A survey. IEEE transactions on pattern analysis and machine intelligence, 43(10):3614-3631, 2020. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.832, + 0.469, + 0.856 + ], + "angle": 0, + "content": "[17] Zongbo Han, Changqing Zhang, Huazhu Fu, and Joey Tianyi Zhou. Trusted multi-view classification. In ICLR, 2020. 3" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.148, + 0.469, + 0.856 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.127, + 0.892, + 0.177 + ], + "angle": 0, + "content": "[18] Bo He, Xitong Yang, Le Kang, Zhiyu Cheng, Xin Zhou, and Abhinav Shrivastava. Asm-loc: Action-aware segment modeling for weakly-supervised temporal action localization. In CVPR, 2022. 1, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.179, + 0.892, + 0.229 + ], + "angle": 0, + "content": "[19] Fa-Ting Hong, Jia-Chang Feng, Dan Xu, Ying Shan, and Wei-Shi Zheng. Cross-modal consensus network for weakly supervised temporal action localization. In ACM MM, 2021. 1, 2, 3, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.231, + 0.892, + 0.268 + ], + "angle": 0, + "content": "[20] Yibo Hu, Yuzhe Ou, Xujiang Zhao, Jin-Hee Cho, and Feng Chen. Multidimensional uncertainty-aware evidential neural networks. In AAAI, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.27, + 0.892, + 0.307 + ], + "angle": 0, + "content": "[21] Linjiang Huang, Liang Wang, and Hongsheng Li. Weakly supervised temporal action localization via representative snippet knowledge propagation. In CVPR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.309, + 0.892, + 0.334 + ], + "angle": 0, + "content": "[22] Audun Jsang. Subjective logic: A formalism for reasoning under uncertainty. Springer Verlag, 2016. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.336, + 0.892, + 0.385 + ], + "angle": 0, + "content": "[23] Will Kay, Joao Carreira, Karen Simonyan, Brian Zhang, Chloe Hillier, Sudheendra Vijayanarasimhan, Fabio Viola, Tim Green, Trevor Back, Paul Natsev, et al. The kinetics human action video dataset. arXiv:1705.06950, 2017. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.387, + 0.892, + 0.424 + ], + "angle": 0, + "content": "[24] Pilhyeon Lee, Jinglu Wang, Yan Lu, and Hyeran Byun. Weakly-supervised temporal action localization by uncertainty modeling. In AAAI, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.426, + 0.892, + 0.464 + ], + "angle": 0, + "content": "[25] Bolian Li, Zongbo Han, Haining Li, Huazhu Fu, and Changqing Zhang. Trustworthy long-tailed classification. arXiv preprint arXiv:2111.09030, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.466, + 0.892, + 0.515 + ], + "angle": 0, + "content": "[26] Chuming Lin, Chengming Xu, Donghao Luo, Yabiao Wang, Ying Tai, Chengjie Wang, Jilin Li, Feiyue Huang, and Yanwei Fu. Learning salient boundary feature for anchor-free temporal action localization. In CVPR, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.517, + 0.892, + 0.554 + ], + "angle": 0, + "content": "[27] Wei Liu, Xiaodong Yue, Yufei Chen, and Thierry Denoeux. Trusted multi-view deep learning with opinion aggregation. In AAAI, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.556, + 0.892, + 0.593 + ], + "angle": 0, + "content": "[28] Vincent Lonij, Ambrish Rawat, and Maria-Irina Nicolae. Open-world visual recognition using knowledge graphs. arXiv preprint arXiv:1708.08310, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.595, + 0.892, + 0.644 + ], + "angle": 0, + "content": "[29] Wang Luo, Tianzhu Zhang, Wenfei Yang, Jingen Liu, Tao Mei, Feng Wu, and Yongdong Zhang. Action unit memory network for weakly supervised temporal action localization. In CVPR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.646, + 0.892, + 0.683 + ], + "angle": 0, + "content": "[30] Lawrence Neal, Matthew Olson, Xiaoli Fern, Weng-Keen Wong, and Fuxin Li. Open set learning with counterfactual images. In ECCV, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.686, + 0.892, + 0.711 + ], + "angle": 0, + "content": "[31] Poojan Oza and Vishal M Patel. C2ae: Class conditioned auto-encoder for open-set recognition. In CVPR, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.712, + 0.892, + 0.748 + ], + "angle": 0, + "content": "[32] Jeffrey Pennington, Richard Socher, and Christopher D Manning. Glove: Global vectors for word representation. In EMNLP, 2014. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.75, + 0.892, + 0.801 + ], + "angle": 0, + "content": "[33] Sanqing Qu, Guang Chen, Zhijun Li, Lijun Zhang, Fan Lu, and Alois Knoll. Acm-net: Action context modeling network for weakly-supervised temporal action localization. arXiv preprint arXiv:2104.02967, 2021. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.803, + 0.892, + 0.853 + ], + "angle": 0, + "content": "[34] Walter J Scheirer, Anderson de Rezende Rocha, Archana Sapkota, and Terrance E Boult. Toward open set recognition. IEEE transactions on pattern analysis and machine intelligence, 35(7):1757-1772, 2012. 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.127, + 0.892, + 0.853 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.467, + 0.888, + 0.505, + 0.898 + ], + "angle": 0, + "content": "14749" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.127, + 0.47, + 0.164 + ], + "angle": 0, + "content": "[35] Murat Sensoy, Lance Kaplan, and Melih Kandemir. Evidential deep learning to quantify classification uncertainty. In NeurIPS, 2018. 3, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.166, + 0.47, + 0.202 + ], + "angle": 0, + "content": "[36] Baifeng Shi, Qi Dai, Yadong Mu, and Jingdong Wang. Weakly-supervised action localization by generative attention modeling. In CVPR, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.204, + 0.469, + 0.24 + ], + "angle": 0, + "content": "[37] Zheng Shou, Hang Gao, Lei Zhang, Kazuyuki Miyazawa, and Shih-Fu Chang. Autoloc: Weakly-supervised temporal action localization in untrimmed videos. In ECCV, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.242, + 0.469, + 0.278 + ], + "angle": 0, + "content": "[38] Krishna Kumar Singh and Yong Jae Lee. Hide-and-seek: Forcing a network to be meticulous for weakly-supervised object and action localization. In ICCV, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.28, + 0.469, + 0.316 + ], + "angle": 0, + "content": "[39] Robyn Speer, Joshua Chin, and Catherine Havasi. Concept-net 5.5: An open multilingual graph of general knowledge. In AAAI, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.317, + 0.469, + 0.354 + ], + "angle": 0, + "content": "[40] Xin Sun, Zhenning Yang, Chi Zhang, Keck-Voon Ling, and Guohao Peng. Conditional gaussian distribution learning for open set recognition. In CVPR, 2020. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.355, + 0.469, + 0.392 + ], + "angle": 0, + "content": "[41] Dennis Ulmer. A survey on evidential deep learning for single-pass uncertainty estimation. arXiv preprint arXiv:2110.03051, 2021. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.393, + 0.469, + 0.43 + ], + "angle": 0, + "content": "[42] Sagar Vaze, Kai Han, Andrea Vedaldi, and Andrew Zisserman. Open-set recognition: A good closed-set classifier is all you need. In ICLR, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.431, + 0.469, + 0.467 + ], + "angle": 0, + "content": "[43] Limin Wang, Yuanjun Xiong, Dahua Lin, and Luc Van Gool. Untrimmednets for weakly supervised action recognition and detection. In CVPR, 2017. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.469, + 0.469, + 0.506 + ], + "angle": 0, + "content": "[44] Huijuan Xu, Abir Das, and Kate Saenko. R-c3d: Region convolutional 3d network for temporal activity detection. In ICCV, 2017. 3" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.127, + 0.47, + 0.506 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.128, + 0.892, + 0.165 + ], + "angle": 0, + "content": "[45] Ronald R Yager and Liping Liu. Classic works of the Dempster-Shafer theory of belief functions, volume 219. Springer, 2008. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.167, + 0.892, + 0.216 + ], + "angle": 0, + "content": "[46] Wenfei Yang, Tianzhu Zhang, Xiaoyuan Yu, Tian Qi, Yongdong Zhang, and Feng Wu. Uncertainty guided collaborative training for weakly supervised temporal action detection. In CVPR, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.218, + 0.892, + 0.268 + ], + "angle": 0, + "content": "[47] Ryota Yoshihashi, Wen Shao, Rei Kawakami, Shaodi You, Makoto Iida, and Takeshi Naemura. Classification-reconstruction learning for open-set recognition. In CVPR, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.27, + 0.892, + 0.318 + ], + "angle": 0, + "content": "[48] Can Zhang, Meng Cao, Dongming Yang, Jie Chen, and Yuexian Zou. Cola: Weakly-supervised temporal action localization with snippet contrastive learning. In CVPR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.321, + 0.892, + 0.371 + ], + "angle": 0, + "content": "[49] Chengwei Zhang, Yunlu Xu, Zhanzhan Cheng, Yi Niu, Shiliang Pu, Fei Wu, and Futai Zou. Adversarial seeded sequence growing for weakly-supervised temporal action localization. In ACM MM, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.373, + 0.892, + 0.422 + ], + "angle": 0, + "content": "[50] Jia-Xing Zhong, Nannan Li, Weijie Kong, Tao Zhang, Thomas H Li, and Ge Li. Step-by-step erasion, one-by-one collection: a weakly supervised temporal action detector. In ACM MM, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.424, + 0.892, + 0.474 + ], + "angle": 0, + "content": "[51] Zhun Zhong, Linchao Zhu, Zhiming Luo, Shaozi Li, Yi Yang, and Nicu Sebe. Openmix: Reviving known knowledge for discovering novel visual categories in an open world. In CVPR, 2021. 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.128, + 0.892, + 0.474 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.467, + 0.887, + 0.505, + 0.898 + ], + "angle": 0, + "content": "14750" + } + ] +] \ No newline at end of file diff --git a/2023/Cascade Evidential Learning for Open-World Weakly-Supervised Temporal Action Localization/3c2843a4-d81f-4df7-8b2c-5c29eaeafad1_origin.pdf b/2023/Cascade Evidential Learning for Open-World Weakly-Supervised Temporal Action Localization/3c2843a4-d81f-4df7-8b2c-5c29eaeafad1_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..f80665bcb48d74627205618cd38281ce60699a33 --- /dev/null +++ b/2023/Cascade Evidential Learning for Open-World Weakly-Supervised Temporal Action Localization/3c2843a4-d81f-4df7-8b2c-5c29eaeafad1_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3e8408885617e74856fdd9e7379e543c454351ae0b25b05e5d6a725194360599 +size 607500 diff --git a/2023/Cascade Evidential Learning for Open-World Weakly-Supervised Temporal Action Localization/full.md b/2023/Cascade Evidential Learning for Open-World Weakly-Supervised Temporal Action Localization/full.md new file mode 100644 index 0000000000000000000000000000000000000000..921dc8d4f7c96600d418762a7f907b51ed6dbf7c --- /dev/null +++ b/2023/Cascade Evidential Learning for Open-World Weakly-Supervised Temporal Action Localization/full.md @@ -0,0 +1,316 @@ +# Cascade Evidential Learning for Open-world Weakly-supervised Temporal Action Localization + +Mengyuan Chen $^{1,2}$ , Junyu Gao $^{1,2}$ , and Changsheng Xu $^{1,2,3}$ + +1 State Key Laboratory of Multimodal Artificial Intelligence Systems (MAIS), Institute of Automation, Chinese Academy of Sciences (CASIA) + +$^{2}$ School of Artificial Intelligence, University of Chinese Academy of Sciences (UCAS) + +$^{3}$ Peng Cheng Laboratory, ShenZhen, China + +chenmengyuan2021@ia.ac.cn; { junyu.gao, csxu}@nlpr.ia.ac.cn + +# Abstract + +Targeting at recognizing and localizing action instances with only video-level labels during training, Weakly-supervised Temporal Action Localization (WTAL) has achieved significant progress in recent years. However, living in the dynamically changing open world where unknown actions constantly spring up, the closed-set assumption of existing WTAL methods is invalid. Compared with traditional open-set recognition tasks, Open-world WTAL (OW-TAL) is challenging since not only are the annotations of unknown samples unavailable, but also the fine-grained annotations of known action instances can only be inferred ambiguously from the video category labels. To address this problem, we propose a Cascade Evidential Learning framework at an evidence level, which targets at OWTAL for the first time. Our method jointly leverages multi-scale temporal contexts and knowledge-guided prototype information to progressively collect cascade and enhanced evidence for known action, unknown action, and background separation. Extensive experiments conducted on THUMOS-14 and ActivityNet-v1.3 verify the effectiveness of our method. Besides the classification metrics adopted by previous open-set recognition methods, we also evaluate our method on localization metrics which are more reasonable for OWTAL. + +# 1. Introduction + +Targeting at recognizing and localizing action instances with only video-level labels during training, Weakly-supervised Temporal Action Localization (WTAL) has attracted increasing attention from both academia and industry [9, 11, 18, 19, 37, 43]. Unlike fully-supervised TAL, WTAL only requires video-level action labels during training. However, the closed-set assumption of existing WTAL + +![](images/daa5b0178a8ff34dde03014efcda7b9b40d5693769d885a40a2c4ce4507a0d42.jpg) +Training Phase + +![](images/b1b53169730a0881bee51cdd58636c5d7c6bb46c1f52ac76484b9c323c349076.jpg) +Testing Phase +Figure 1. Illustration of the training and testing phases of OWTAL. With only video-level labels for training, OWTAL aims to localize both known and unknown action instances in testing videos. + +methods is invalid in the dynamically changing real world, since with the development of society never-before-seen human action categories are constantly emerging. Therefore, to address this problem, we consider a different WTAL setting in this work, termed as Open-world WTAL (OWTAL). + +Different from the traditional WTAL task, as shown in Figure 1, OWTAL allows testing videos to contain action instances of unknown categories, which have never appeared during training. Therefore, temporal boundaries of both known and unknown action instances are expected to be predicted. Compared with its fully-supervised counterpart Open Set TAL [3], OWTAL is challenging in two aspects: (1) Ambiguity of annotations of closed-set (known) action instances. Previous works indicate that the closed-set and open-set performance are highly correlated [42]. However, under the OWTAL setting, not only are the annotations of unknown action instances unavailable, but also the fine-grained annotations of known ones can only be inferred ambiguously from the video category labels. During training, the known action instances that the model needs to focus on are prone to be disturbed by the background snippets, which hinders the learning of the closed-set actions, thus making it extremely difficult to differentiate the unknown actions, the known actions, and the background. (2) Lack of reasonable metrics. The traditional Open Set Recognition (OSR) aims + +for classification while the goal of OWTAL is to perform localization instead, thus the classification metrics commonly adopted by OSR are not sufficient for OWTAL. + +In order to alleviate the negative impact caused by the weak annotations of known action instances, we propose a Cascade Evidential Learning method for owtaL (CELL), which progressively collects cascaded evidence by considering both temporal contexts in multi-scale ranges and inter-video correlations under the guidance of prior knowledge. Since the goal of OWTAL is to locate the consecutive known/unknown action segments of various temporal lengths in open-world scenarios, perceiving temporal contexts in diverse ranges is essential. We argue that it is meaningful to endow individual snippet features with the ability of sensing multi-scale neighborhood video segments, and thus a Multi-scale Extended-range Perception module (Section 3.2) is designed to obtain more discriminative video features for initial evidence collection, taking advantage of the temporal contexts. Due to the large intra-action variation in visual patterns and the lack of prior knowledge guidance, the known action instances which visually deviate from the common ones are likely to be misidentified with the initial evidence collected from individual videos. Therefore, we design a Knowledge-guided Bipolar Prototype Learning strategy (Section 3.3), where a semantic relation graph is constructed to provide prior knowledge guidance for the bipolar prototype learning among videos, thus perceiving the open-world more comprehensively. We use this strategy to generate a series of evidence calibration factors for further cascade evidence enhancement. Finally, a Cascade Evidence Enhancement module (Section 3.4) is designed for enhancing the initial evidence with the calibration factors, and the uncertainty estimated from the cascaded evidence is used for the known/unknown judgment. Extensive experiments conducted on THUMOS-14 and ActivityNet verify the effectiveness. Besides the various classification metrics adopted by previous works, we also evaluate our method on localization metrics which are more in line with the needs of real-world applications. + +To summarize, our contribution is threefold: + +- To tackle the unique challenges of OWTAL, we propose a cascade evidential learning framework, which progressively collects comprehensive evidence for known action, unknown action, and background separation. Localization metrics which meet the needs of the real-world more closely are adopted for evaluation. +- To achieve OWTAL without fine-grained annotations, the proposed CELL jointly leverages multi-scale temporal contexts and knowledge-guided prototype information during the evidence cascade learning process. +- We conduct comprehensive experiments on two popular WTAL benchmarks, THUMOS-14 and ActivityNetv1.3, and achieve significant performance improvement + +over various baselines. Experiments show that CELL enables existing methods to well adapt to the more practical open-world settings. + +# 2. Related Work + +Weakly-supervised Temporal Action Localization. Originated from UntrimmedNet [43], the pioneer work to utilize video-level action category annotations as the weak supervision for TAL, existing WTAL methods commonly adopt a multiple instance learning strategy and can be roughly categorized into erasing-based, attention-based and uncertainty-based methods. Erasing-based methods [38, 49, 50] erase the most discriminative segments to mitigate the single snippet cheating issue [48]. Attention-based methods [10, 19, 21, 29, 33, 36] employ the attention mechanism to select snippets most likely to be the foreground with activation scores. The newly-arising uncertainty-based methods [9, 24, 46] utilize uncertainty to address the inevitable action-background ambiguity caused by the weakly-supervised setting. However, all existing WTAL methods are rooted in the closed-set assumption, which impedes their application to real-world scenarios. In contrast, we address the WTAL task under the open-world setting by collecting reliable cascaded evidence, which has not been explored in prior work. + +Open Set Recognition. Open Set Recognition (OSR) aims to recognize known classes and reject the unknown. [34] first formalize the OSR problem and provide a basic framework. [4] propose the first DNN-based OSR method Openmax, which rejects unknown classes by modeling the distance of activation vectors with Extreme Value Theory (EVT). GAN-based methods are also developed towards this task. [30] generate samples similar to the training data but not belonging to the known classes, and then utilize the generated unknown class samples to train an open-set classifier. Other approaches include prototype-based methods [7, 8], which reject unknown classes by calculating the maximum distance between the input sample and the learned closed-set prototypes, and reconstruction-based methods [31, 40, 47] utilize the reconstruction error in the test phase as an open-set indicator. Some works [28, 51] attempt to seek assistance from external knowledge, but few of them have explicitly modeled unknown class information in an end-to-end framework. + +Most existing approaches focus on open-set recognition tasks while few works attend to temporal localization-related vision field. [3] formalize the open-set temporal action localization problem and propose an OpenTAL framework. However, its fully-supervised setting requires large amounts of fine-grained labels whose annotation is time-consuming, error-prone, and costly. Besides, the evaluation metrics in [3] are mainly classification ones. In contrast, in this paper, we address the challenging OWTAL problem which also abandons the closed-set assumption but only re + +quires easily available video-level class labels, and employ comprehensive localization-related metrics for evaluation. + +Evidential Deep Learning. Based on Dempster-Shafer Theory (DST) [45] and Subjective Logic theory (SL) [22], Evidential Deep Learning (EDL) allows uncertainty estimation in a single forward pass [41] by collecting scalar evidence for each category and parameterizing a Dirichlet distribution, which models the distribution of classification probabilities, over the collected evidence. As a newly arising trustworthy method, EDL has achieved remarkable progress in various computer vision tasks, including action localization [9], image classification [35], regression [1], multi-view classification [17, 27], out-of-distribution detection [20], long-tail learning [25], and open-set action recognition and localization [2, 3]. However, current EDL algorithms only focus on collecting evidence from individual samples, while they neglect the complementary role that the potential correlations among samples can play in evidence collection. To address this issue, in this work we propose a cascade evidence calibration paradigm which effectively enhances the collected evidence. + +# 3. Our Approach + +# 3.1. Notations and Preliminaries + +Problem formulation of OWTAL. OWTAL targets at detecting known/unknown action instances in untrimmed videos. Formally, we are given a training set $\{\mathcal{V},\mathcal{Y}\}$ with $N$ training videos, where $\mathcal{V}$ denotes an untrimmed video, and $\pmb {y}\in \mathbb{R}^{C}$ is its multi-hot label indicating the action categories that the action instances in this video belong to, where $C$ is the number of known action categories. During testing, the goal is to use the learned model to output a set of quadruplets $\{c_r,t_r^s,t_r^e,\gamma_r\}_{r = 1}^R$ ,where $R$ is the number of action instances in $\mathcal{V}$ $c_{r},t_{r}^{s},t_{r}^{e}$ and $\gamma_{r}$ denote the action category, the start and end timestamps, and the confidence score, respectively. $c_{r}\in \{1,\dots,C + 1\}$ , where $C + 1$ denotes the unknown action category. + +Feature extraction. Following previous works [10,46], we split the untrimmed video $\mathcal{V}$ into $T$ non-overlapping 16-frame snippets and then use pre-trained networks, e.g., the I3D model [23], to extract RGB and optical flow features. Then, the above features are concatenated and then fed into a fusion module, e.g., convolutional layers [19, 33], to obtain the snippet-wise feature $\pmb{X} = [\pmb{x}_1,\dots,\pmb{x}_T]\in \mathbb{R}^{T\times D}$ , where $D$ is the feature dimension. + +# 3.2. Multi-scale Extended-range Perception for Initial Evidence Collection + +A video may contain several known/unknown action instances with various temporal lengths in the open-world. Since the goal of OWTAL is to locate these consecutive video segments, perceiving temporal contexts in diverse ranges is essential. We argue that it is meaningful to endow + +individual snippet features with the ability of sensing multiscale neighborhood video segments. As shown in Figure 2, we design a Multi-scale Extended-range Perception module (MEP) to obtain more discriminative video features, taking advantage of the temporal contexts. Note that some anchor-based temporal modeling methods [6, 12, 14, 44] also leverage multi-scale temporal information. However, their goal is to obtain the fused temporal feature for proposal generation, while the MEP aims to enhance individual snippet features for initial evidence collection. + +In MEP, each snippet $t$ is assigned with a series of $t$ -centered multi-scale video segments, whose context-aware features can contribute to the extension of snippet-specific perception ranges. Here, the temporal boundaries of the $t$ -centered video segments can be denoted as set $\Omega_t = \{(s_t^m, e_t^m) | m \in \mathcal{M}\}$ , where $\mathcal{M}$ is a pre-defined length set, $s_t^m = \max(0, t - m)$ , and $e_t^m = \min(T, t + m)$ . For brevity, we denote the temporal average of the feature sequence of the extended video segment $(s_t^m, e_t^m)$ as $f_t^m \in \mathbb{R}^D$ , which can be given by: + +$$ +\boldsymbol {f} _ {t} ^ {m} = \frac {1}{e _ {t} ^ {m} - s _ {t} ^ {m} + 1} \sum_ {s _ {t} ^ {m} \leq i \leq e _ {t} ^ {m}} \boldsymbol {x} _ {i}. \tag {1} +$$ + +To extend the perception range of snippet $t$ , we perform feature enhancement by fusing $\boldsymbol{x}_t$ with $\boldsymbol{f}_t^m$ , during which the weight of $\boldsymbol{f}_t^m$ is positively correlated with the similarity between them. Such operation is motivated by the simple expectation that a snippet-level feature would absorb meaningful contextual information mainly from extended video segments that are highly relevant to it. Therefore, the snippet-level feature with extended-range perception, denoted as $\tilde{\boldsymbol{F}} = [\tilde{\boldsymbol{f}}_1,\dots ,\tilde{\boldsymbol{f}}_T]$ , can be formulated as: + +$$ +\tilde {\boldsymbol {f}} _ {t} = \left(1 - \alpha_ {t}\right) \varphi_ {1} (\boldsymbol {x} _ {t}) + \alpha_ {t} \sum_ {m \in \mathcal {M}} \delta \left(\omega_ {t} ^ {m}\right) \varphi_ {2} \left(\boldsymbol {f} _ {t} ^ {m}\right), \tag {2} +$$ + +where $\delta (\cdot)$ denotes the Softmax operation with a temperature factor, $\varphi_{1,2}$ are fully-connected layers for feature embedding, $\alpha_{t} \in [0,1]$ is the fusion weight representing the scaled average of the cosine similarities $\omega_{t}^{m}$ between $\pmb{x}_{t}$ and $\pmb{f}_t^m$ , and $\alpha_{t}$ and $\omega_t^m$ can be calculated by: + +$$ +\alpha_ {t} = \frac {1}{2 | \mathcal {M} |} \sum_ {m \in \mathcal {M}} \left(\omega_ {t} ^ {m} + 1\right), \omega_ {t} ^ {m} = \cos \left(\frac {\boldsymbol {x} _ {t}}{\left| \boldsymbol {x} _ {t} \right|}, \frac {\boldsymbol {f} _ {t} ^ {m}}{\left| \boldsymbol {f} _ {t} ^ {m} \right|}\right). \tag {3} +$$ + +Thereafter, similar to previous methods [10, 19], we utilize an attention module to predict an attention score sequence $\mathbf{A} = [A_{1},\dots,A_{T}]\in \mathbb{R}^{T}$ for the input video $\nu$ representing the probabilities of snippets belonging to the foreground, then select the top-ranked snippets. Specifically, for the input video we obtain the set $\Theta$ of snippets with the top- $k$ attention scores by the following equation: + +$$ +\Theta = \arg \max _ {\Theta} \sum_ {t \in \Theta , | \Theta | = k} A _ {t}, \quad \Theta \subset \{1, 2, \dots , T \}, \tag {4} +$$ + +![](images/395baac5f70b6598677f6663cca078ae6d31e0f1e20fd363883ca19a848f68ff.jpg) +Figure 2. Overall framework of the proposed CELL for OWTAL. Firstly, a Multi-scale Extended-range Perception (MEP) module is designed to collect the initial evidence by perceiving temporal contexts in diverse ranges. Then we propose a Knowledge-guided Bipolar Prototype Learning (KBPL) module to obtain evidence calibration factors from inter-video correlations and prior knowledge. Finally, a Cascade Evidence Enhancement (CEE) module is designed for enhancing the initial evidence with the calibration factors, and the uncertainty derived from the cascaded evidence is used for the known/unknown judgment. + +where $k = \lceil T / r \rceil$ , $r$ is a scaling factor. According to the snippet index in $\Theta$ , we select them from $\tilde{\pmb{F}}$ and calculate the average as the video-level feature, denoted as $\tilde{F}_{vid}$ . + +With the multi-scale-context-aware video-level feature, we feed $\tilde{\pmb{F}}_{vid}$ into an evidence collector $f$ , a DNN parameterized by $\theta$ , to collect initial evidence $e_{init} \in \mathbb{R}^C$ : + +$$ +\boldsymbol {e} _ {\text {i n i t}} = g \left(f \left(\tilde {\boldsymbol {F}} _ {v i d}; \boldsymbol {\theta}\right)\right), \tag {5} +$$ + +where $g$ denotes a scale function, e.g., ReLU, Softplus or Exp, to ensure the collected evidence non-negative. + +Based on Subjective Logic theory [22], Evidential Deep Learning (EDL) allows uncertainty estimation in a single forward pass [41] by collecting evidence of each category and parameterizing a Dirichlet distribution, which models the distribution of class probabilities, over the collected evidence. Our collected initial evidence $e_{init}$ can be optimized by the following loss function: + +$$ +\mathcal {L} _ {\text {i n i t}} = \sum_ {i = 1} ^ {C} y _ {i} \left(\log S _ {\text {i n i t}} - \log \alpha_ {\text {i n i t}, i}\right), \tag {6} +$$ + +where $S_{init} = \sum_{i}^{C}\alpha_{init,i},\alpha_{init,i} = e_{init,i} + 1$ , and Eq. (6) is actually the deformation of the negative logarithm of the marginal likelihood in the EDL paradigm [35]. Besides, the video-level classification uncertainty $u_{init}$ can be derived by the EDL theory as $u_{init} = C / S_{init}$ . Note that uncertainty $u_{init}$ is inversely proportional to the total evidence of all closed-set categories, thus it reflects the probability that the video contains unknown actions. + +# 3.3. Knowledge-guided Bipolar Prototype Learning for Evidence Calibration Factors + +Due to the differences in action components, scenes, shooting angles and other aspects, videos tend to keep a large intra-action variation in visual patterns, and the known action instances which visually deviate from the majority are likely to be mistakenly identified as unknown ones. We argue that the initial evidence collected by a single video is insufficient for handling this issue. Therefore, we resort to the bipolar prototype learning strategy, in which both positive and negative prototypes are set for each category, to explore inter-video correlations for mining the intrinsic information of actions and eliminating the interference caused by background snippets. However, even if the intra-action visual variance is mitigated to some extent, it is still challenging to effectively detect the unknown action instances in the open-world environment for the lack of prior knowledge. To tackle the problem, a semantic relation graph is constructed to simulate the open world, so as to provide prior knowledge guidance for the bipolar prototype learning among videos, finally obtaining more accurate evidence calibration factors for further cascade evidence enhancement. + +Based on the above observations, as shown in Figure 2, we propose a Knowledge-guided Bipolar Prototype Learning module (KBPL) to obtain evidence calibration factors from inter-video correlations and prior knowledge. Firstly, we construct a semantic relation graph with concepts related to known categories as vertie and their relations as edges. Specifically, the vertie set $\mathcal{N}$ includes concepts of known + +classes, denoted as $\mathcal{N}_{kno}$ , and the neighbor action concepts in ConceptNet [39] with the top- $p$ relation strength, denoted as $\mathcal{N}_{nei}$ , thus $\mathcal{N} = \mathcal{N}_{kno} \bigcup \mathcal{N}_{nei}$ . Following a standard graph construction operation [13, 15], edges are connected in our graph if the relation strength between two concepts (vertex) is larger than a threshold. The features of vertex in $\mathcal{N}$ are initialized by the Glove-300 [32] embedding vectors, denoted as $\{n_j\}_{j=1}^{|\mathcal{N}|}$ , and then updated by: + +$$ +\tilde {\boldsymbol {n}} _ {j} = h _ {2} \left(\underset {z \in \{1, 2, 3 \}} {\operatorname {C o n c a t}} \left(\sum_ {l \in \mathcal {S} _ {j, z}} \delta \left(\eta_ {j, l}\right) h _ {1} \left(\boldsymbol {n} _ {l}\right)\right)\right) + h _ {3} \left(\boldsymbol {n} _ {j}\right), \tag {7} +$$ + +where $h_{1,2,3}$ denote fully-connected layers, $\delta(\cdot)$ denotes the Softmax operation, $S_{j,z}$ denotes the index set of the $z$ -hop neighbor vertex of node $j$ , and $\eta_{j,l}$ represents the cosine similarity of $\mathbf{n}_l$ and $\mathbf{n}_j$ . Although many other complicated and advanced graph network methods can be adopted, we select the direct update manner here for simplicity. + +With the concept relation graph, for category $i$ , we use the updated representation of its corresponding vertex to generate the positive prototype, i.e. $Q_{i}^{pos} = \phi_{1}(n_{i})$ , and the average updated representations of its neighbor vertex are adopted for producing the negative one, i.e. $Q_{i}^{neg} = \phi_{2}(\sum_{j\in S_{i}}n_{j} / |S_{i}|)$ , where $\phi_{1,2}$ are fully-connected layers for prototype embedding. Then, for video $\nu_{b}$ in the minibatch, where $b = 1,\dots,B$ , and $B$ is the mini-batch size, we define the calibration factor $\xi_{b,i}$ by simultaneously considering the distances between the video-level feature $\tilde{F}_{vid,b}$ and the bipolar prototypes $Q_{i}^{pos}$ and $Q_{i}^{neg}$ as: + +$$ +\xi_ {b, i} = \hat {D} (\tilde {\boldsymbol {F}} _ {v i d, b}, \boldsymbol {Q} _ {i} ^ {n e g}) - \hat {D} (\tilde {\boldsymbol {F}} _ {v i d, b}, \boldsymbol {Q} _ {i} ^ {p o s}), \tag {8} +$$ + +where $\hat{D}$ is a distance calculation method, which can be Euclidean distance or negative cosine similarity. + +To guide the learning of prototypes by inter-video correlations, for each category, we pull close the videos belonging to the category to the corresponding positive prototype, and push away the videos of other classes in the same minibatch, while opposite operations are performed for the negative prototype. Specifically, the loss function $\mathcal{L}_{kbpl}$ guiding the pos-neg prototype learning is designed as: + +$$ +\mathcal {L} _ {k b p l} = - \frac {1}{C} \sum_ {i = 1} ^ {C} \mathbb {I} \left(\mathcal {B} _ {i} \neq \emptyset\right) \log \left(\frac {\sum_ {b \in \mathcal {B} _ {i}} \exp \left(\xi_ {b , i}\right)}{\sum_ {b = 1} ^ {B} \exp \left(\xi_ {b , i}\right)}\right), \tag {9} +$$ + +where $\mathbb{I}(\cdot)$ is an indicator function, $\mathcal{B}_i = \{b|y_{b,i} = 1\}$ , and $\pmb{y}_b$ is the ground-truth label of the video $\mathcal{V}_b$ . + +# 3.4. Cascade Evidence Enhancement + +As shown in Figure 2, we then perform cascade evidence enhancement, which scales the initial evidence collected from the MEP module by the learned calibration factors. As we stated above, the uncertainty $u_{init}$ represents the probability that a video contains unknown actions. For example, + +given a test video, when $u_{init}$ is close to the extreme values (close to zero for known actions and close to the maximum for unknown ones), we can recognize known or unknown classes with high confidence [35], otherwise the model has relatively low confidence in the known/unknown judgment. Specifically, we design an uncertainty-aware adaptor $\zeta$ as a Gaussian-like estimation function for evidence cascade: + +$$ +\zeta \left(u _ {i n i t}; \mu , \sigma\right) = \exp \left(- \frac {\left(u _ {i n i t} - \mu\right) ^ {2}}{2 \sigma^ {2}}\right), \tag {10} +$$ + +where $\mu$ is set to the value beyond the $u_{init}$ of $95\%$ training videos, and $\sigma$ is a learnable scalar parameter which controls the shape of the Gaussian-like function. + +If the above uncertainty-aware adaptor $\zeta$ of an input video is high, it means it is difficult for the model to perform known/unknown judgment by only considering the initial evidence collected inside the video, thus it is necessary to employ the information provided by the inter-video correlations to carry out evidence calibration. Specifically, the cascade evidence calibration process can be formulated as: + +$$ +e _ {c a s c, i} = (1 + \zeta (u _ {i n i t}) \tanh (\xi_ {i})) e _ {i n i t, i}, \tag {11} +$$ + +where the $\tanh (\cdot)$ function ensures the scaling range of $e_{init,i}$ to be (0,2) for a stable learning, and $e_{casc} = [e_{casc,1},\dots,e_{casc,C}] \in \mathbb{R}^C$ is the enhanced cascaded evidence. The subscript $b$ (video index) is omitted in this equation for simplicity. $e_{casc}$ is then optimized by an EDL loss $\mathcal{L}_{casc}$ , whose form is identical with Eq. (6), and the cascaded video-level uncertainty $u_{casc}$ for known/unknown judgment can be obtained in the same way with $u_{init}$ . + +# 3.5. Training and Inference + +Training. By combining all the optimization objectives introduced above, we obtain the final loss function as: + +$$ +\mathcal {L} = \mathcal {L} _ {c l s} + \mathcal {L} _ {i n i t} + \mathcal {L} _ {k b p l} + \mathcal {L} _ {c a s c}, \tag {12} +$$ + +where $\mathcal{L}_{cls}$ is the standard video-level classification loss, e.g., MIL loss [10, 19], for optimizing the backbone. + +Inference. For each video, we predict its action categories set $\Psi$ (including the Unknown class if exist) by Algorithm 1, where the closed-set video-level classification score $P = [P_{1},\dots,P_{C}]$ is obtained by averaging the classification activation sequence (CAS) of snippets in $\Theta$ . Then we use a threshold strategy to obtain action snippet candidates following the standard process [18, 19]. Finally, for each action class in $\Psi$ we group continuous snippets into action proposals, whose confidence scores $\gamma$ are estimated according to the CAS scores of the corresponding class (for Known classes) or attention scores (for the Unknown class), and perform non-maximum-suppression (NMS) to remove duplicated proposals. + +Table 1. OWTAL localization results on THUMOS-14 evaluated by both top- $K$ mAP@Avg and traditional mAP@Avg. The averages of the mAP of Known classes and the AP of the Unknown class are reported. The values are averaged on t-IoU [0.1:0.1:0.5] and [0.1:0.1:0.7]. + +
MethodsTop-K mAP@Avg(%)mAP@Avg(%)
Top-5Top-10Top-20Top-50Top-100
0.1-0.50.1-0.70.1-0.50.1-0.70.1-0.50.1-0.70.1-0.50.1-0.70.1-0.50.1-0.70.1-0.50.1-0.7
ASM-Loc + Trivial4.33.57.86.313.310.721.116.926.120.929.523.6
ASM-Loc + SoftMax10.98.916.913.523.418.728.823.029.823.830.124.0
ASM-Loc + OpenMax10.18.115.312.121.216.726.120.627.321.527.621.7
ASM-Loc + ARPL10.08.216.413.423.619.330.424.832.026.032.426.4
ASM-Loc + EDL11.39.217.414.024.219.430.524.331.925.432.225.7
ASM-Loc + CELL(Ours)11.29.318.014.625.520.732.426.434.127.834.728.1
CO2-Net + Trivial5.54.49.57.716.713.525.920.930.925.034.427.9
CO2-Net + SoftMax11.39.117.814.325.120.232.226.033.727.334.227.8
CO2-Net + OpenMax10.38.416.313.223.018.629.123.530.424.730.825.0
CO2-Net + ARPL11.69.518.314.925.720.933.327.135.128.735.729.2
CO2-Net + EDL11.29.117.614.324.820.032.226.034.027.534.628.1
CO2-Net + CELL(Ours)12.610.320.116.428.123.036.930.338.931.839.532.3
+ +Table 2. OWTAL localization results on THUMOS-14 evaluated by both top- $K$ mAP@Avg and traditional mAP@Avg. The mAP of Known classes (K) and the AP of the Unknown class (U) averaged on t-IoU thresholds [0.1:0.1:0.7] are reported respectively. The video-level known-unknown classification accuracy (v-Acc(%) ) is also presented for reference. + +
MethodsTop-K mAP@Avg(%)mAP@Avg(%)v-Acc(%)
Top-5Top-10Top-20Top-50Top-100
KUKUKUKUKUKU
CO2-Net + Trivial4.44.47.87.615.111.923.818.030.519.436.219.628.1
CO2-Net + SoftMax12.65.619.39.326.413.932.719.234.520.035.620.071.2
CO2-Net + OpenMax11.94.918.48.025.112.031.016.132.716.733.616.570.5
CO2-Net + ARPL12.36.719.410.426.615.232.621.734.622.735.822.673.0
CO2-Net + EDL12.45.719.19.426.313.832.719.434.820.235.920.370.9
CO2-Net + CELL(Ours)13.37.320.512.227.618.334.026.635.927.636.827.874.9
+ +# 4. Experimental Results + +# 4.1. Experimental Setup + +Datasets. We conduct experiments on two popular benchmarks for OWTAL, THUMOS-14 and ActivityNet-v1.3. THUMOS-14 contains 200 validation videos and 213 test videos annotated from 20 action categories, and each video contains 15.4 action instances on average. ActivityNet-v1.3 contains 10,024 training videos and 4,926 validation videos from 200 action categories, and each video contains 1.6 action instances on average. Following the pioneer work [3], we randomly remove 1/4 action categories of THUMOS-14 training set and keep the entire THUMOS-14 testing set to enable open-set evaluation, and such random removal is repeated three times. To further increase the openness and verify the applicability of our proposed method for open-world scenarios, we train our model on the known split of THUMOS-14 and test on the testing set of ActivityNet-v1.3. Note that since ActivityNet-v1.3 covers + +most categories in THUMOS-14, we manually remove 14 action categories which are semantically overlapping with the THUMOS-14 as previous work did. + +Evaluation Metrics. We adopt three sets of metrics to measure the performance of the OWTAL task. (1) Vanilla mAP. To evaluate the localization performance, following previous approaches [10, 19, 33], we employ the mean Average Precision (mAP) under different temporal Intersection over Union (t-IoU) thresholds as metrics. The t-IoU thresholds for THUMOS-14 is [0.1:0.1:0.7] and for ActivityNetv1.3 is [0.5:0.05:0.95]. (2) Top- $K$ mAP. In our experiments, we observe that it only brings slight performance loss on the traditional mAP for models to generate a large number of redundant proposals. However, in real-world applications our goal is to provide users with accurate action proposals of a limited number, rather than an extreme largescale proposal set consisting of numerous redundant proposals, although the set may contain even more accurate predictions. Inspired by the above observation, we design a + +# Algorithm 1 Action Categories Inference Procedure + +Input: Untrimmed testing video $\nu$ + +Require: Trained CELL model. + +Require: Threshold $\tau$ obtained from training data (Please refer to Implementation Details). + +Output: Set $\Psi$ of action categories in $\mathcal{V}$ . + +1: Predict the closed-set classification score $P$ and cascaded video-level uncertainty $u_{\text{casc}}$ by CELL. +2: if $u_{\text{casc}} < \tau$ then +3: $\Psi = \{i|P_i > 0.2\}$ ▷OnlyKnownClasses +4: if $\Psi = \emptyset$ then +5: $\Psi = \arg \max_{i} P_{i}$ ▷OnlyKnownClasses +6: end if +7: else if $\arg \max_{i} P_{i} > 0.5$ then +8: $\Psi = \{\arg \max_{i} P_{i}, C + 1\}$ +9: Both Known and Unknown Classes +10: else +11: $\Psi = \{C + 1\}$ ▷OnlyUnknownClasses +12: end if +13: return $\Psi$ + +variant metric of the vanilla mAP to further meet the needs of real-world applications, termed as top- $K$ mAP, which restricts the maximum number of generated proposals by selecting proposals with top- $K$ confidence scores. When $K$ is not restricted, the top- $K$ mAP degrades to the traditional mAP metric. In our experiments $K$ is set to 5, 10, 20, 50, 100 for THUMOS-14 and 1, 3, 5 for ActivityNet-v1.3, respectively. (3) Classification metrics. For known/unknown classification, following [3], four classification metrics including the False Alarm Rate at True Positive Rate of $95\%$ (FAR@95) (smaller value indicates better performance), the Area Under the Receiver Operating Characteristic (AUROC) curve, the Area Under the Precision-Recall (AUPR), and the Open Set Detection Rate (OSDR) are also employed. To validate the effectiveness of Algorithm 1, we additionally adopt the video-level known-unknown classification accuracy (v-Acc(\%)) metric for evaluation. Note that a testing video may belong to three types which have: (i) only known actions, (ii) only unknown actions, or (iii) both known and unknown actions. + +Implementation Details. We adopt CO2-Net [19] and ASM-Loc [18] as our backbone networks. Following existing methods, we use I3D [5] model pretrained on the Kinetics [23] dataset to extract both the RGB and optical flow features. The feature dimension $D$ is 2048, the evidence scale function $g$ in Eq. (6) is Exp, and the distance calculation method $\hat{D}$ in Eq. (8) is negative cosine similarity. The number of the sampled snippets $T$ for THUMOS-14 and ActivityNet-v1.3 is 500 and 160, and the scaling factor $r$ is 7. $\xi$ in Eq. (10) is updated for every 25 epochs. The batch size is set to 10 and 16, and the learning rate is 5e-5 and 1e + +4 in CO2-Net and ASM-Loc, respectively. Following previous works [2, 40], we determine the threshold $\tau$ of uncertainty $u_{\text{casc}}$ in Algorithm 1 by ensuring $95\%$ training videos to be recognized as known. Our model is implemented with Python 3.7 and PyTorch 1.11.0. All experiments are conducted on a single RTX3090 GPU. + +# 4.2. Comparison with State-of-the-art Methods + +Our proposed CELL method is compared with the following baselines based on the ASM-Loc [18] and the CO2-Net [19]: (1) Trivial: assuming that all testing videos contain both known and unknown actions without any known/unknown judgment. (2) SoftMax: utilizing the maximum softmax probability to identify the unknown. (3) OpenMax: appending the unknown score to the softmax scores by OpenMax [4] in testing. (4) ARPL: using the Adversarial Reciprocal Points Learning (ARPL) [7] to identify the unknown. (5) EDL: vanilla EDL is used to replace the softmax classification head for uncertainty estimation. + +Comparison results on THUMOS-14. Table 1 and Table 2 represent the OWTAL localization results on the THUMOS-14 dataset. The results show that our proposed method consistently outperforms the baselines by large margins on all metrics. We notice that totally without any known/unknown judgment, the Trivial method shows comparable performance with the SoftMax baseline on the traditional mAP metric, for the embarrassing property that it only brings slight performance loss on the traditional mAP to generate numerous error redundant proposals, while on our proposed top- $k$ mAP metrics the performance of the Trivial version shows a reasonable gap. The results also show that Openmax does not work well on the OWTAL task, which may lie in that Openmax often fails to recognize visually indistinguishable samples [16], especially in weakly-supervised video analysis field. Note that the other two SOTA approaches ARPL and EDL perform well but still far behind the proposed CELL. + +Table 3 displays the known/unknown classification results evaluated on the metrics adopted by [3], showing that our method outperforms the EDL baseline and even achieves comparable results with fully-supervised methods on FAR@95, AUROC, and AUPR. We infer that the large gap between weakly-supervised methods and fully-supervised ones on OSDR is due to the difference between backbone networks. OSDR is defined as the area under the curve of Correct Detection Rate (CDR) and False Positive Rate (FPR), and the CDR indicates the fraction of known actions which are positively localized and correctly classified, while the FPR denotes the fraction of unknown actions that are positively localized but falsely classified. CO2-Net adopts the localization-by-classification strategy commonly used in WTAL method, thus the positively localized actions are usually accompanied by correct class predictions, while AFSD [26] adopts an anchor-free method whose localize + +tion performance does not entirely rely on the classification results, thus resulting in the large gap on OSDR. Note that compared with our employed mAP-related metrics, OSDR is still a classification metric thus cannot fully exhibit the superiority of different models. + +Table 3. OWTAL classification results on THUMOS-14 evaluated by FAR@95, AUROC, AUPR and OSDR. In this experiment we use CO2-Net as our backbone. + +
SupervisionMethodsFAR@95(↓)AUROCAUPROSDR
FullySoftMax85.5854.7031.8523.40
OpenMax90.3453.2633.1713.66
EDL81.4264.0540.0536.26
OpenTAL70.9678.3358.6242.91
WeaklyCO2+EDL87.3767.1144.4763.56
CELL(Ours)68.8674.6254.2569.81
+ +Table 4. OWTAL localization results on ActivityNet-v1.3. The values are averaged on t-IoU thresholds [0.5:0.05:0.95]. The metric v-Acc(%) is also presented for reference. + +
MethodsTop-K mAP@Avg(%)mAP@Avg(%)v-Acc(%)
Top1Top3Top5
ASM-Loc + SoftMax12.713.213.513.586.37
ASM-Loc + OpenMax12.012.612.812.981.74
ASM-Loc + ARPL17.017.617.818.096.81
ASM-Loc + EDL16.016.716.917.093.38
ASM-Loc + CELL(Ours)17.918.418.618.997.83
+ +Table 5. Ablation study of our proposed CELL. MEP denotes the multi-scale extended-range perception module, KBPL represents the knowledge-guided bipolar prototype learning module, and UA denotes the uncertainty-aware adaptor $\zeta$ described in Eq. (10). + +
ExpMEPKBPLUATop-k mAP@Avg(%)mAP@Avg(%)
Top-10Top-20Top-50
1XXX14.320.226.027.8
2✓XX15.321.226.929.0
3X✓X15.221.126.929.2
4✓✓X16.022.429.431.6
5✓✓✓16.423.030.332.3
+ +Comparison results on ActivityNet-v1.3. To further increase the openness and verify the applicability of our proposed method for open-world scenarios, we further train our model on the known splits of THUMOS-14 and test on ActivityNet-v1.3. Although the testing set only consists of unknown action categories, in order to ensure a fair comparison, the model is still tested according to Algorithm 1, without using any additional prior knowledge. As shown in Table 4, our proposed CELL obtains consistently favorable performance. Compared with EDL, CELL shows superior performance with an absolute gain of $1.9\%$ in mAP@Avg. + +# 4.3. Ablation Study + +This section shows the effectiveness of our modules. In Table 5, the impact of progressively adding each component is presented, proving their contributions clearly. According to the cascade steps of our method, we perform ablation study on three components, Multi-scale Extended-range Perception module (MEP), Knowledge-guided Bipolar Prototype Learning module (KBPL), and the Uncertainty-aware Adaptor (UA) in CEE module. Note that without MEP we replace the fused feature by the average of snippet features in $\Theta$ , and without UA we simply set $\zeta$ in Eq. (11) to 1. As shown in Table 5, every step of our method brings effective performance improvement. Moreover, it is noteworthy that MEP perceiving temporal contexts in multiscale ranges and KBPL leveraging inter-video correlations and prior knowledge can enhance and complement each other, thus significantly improving the performance. + +# 5. Conclusions + +Targeting at open-world weakly-supervised temporal action localization (OWTAL), we propose a cascade evidential learning framework, which entails three main components: (1) Multi-scale extended-range perception module perceiving temporal contexts in diverse ranges for collecting initial evidence; (2) Knowledge-guided bipolar prototype learning strategy exploring inter-video relations under the guidance of prior knowledge for seeking supplementary evidence support; (3) Cascade evidence enhancement for final evidence calibration. In our extensive experiments, CELL achieves state-of-the-art performance on various metrics. Several limitations of this work are noteworthy. Firstly, never-before-seen human action categories are constantly emerging, thus lifelong learning paradigm which can keep embracing new action categories is even more suitable for OWTAL. Secondly, the class annotation granularity of the training set will interfere with the unknown action identification in open-world scenarios. Specifically, if some fine-grained actions are not labeled during training, e.g., the run-up movement in HighJump videos in THUMOS14, the model may misjudge such unknown actions as the background in testing. We assume the problem can be alleviated by training on large-scale datasets with fine-grained annotations. These directions are left as future work. + +# Acknowledgements + +This work was supported by the National Key Research & Development Plan of China under Grant 2020AAA0106200, in part by the National Natural Science Foundation of China under Grants 62036012, U21B2044, 62236008, 62102415, 61721004, 62072286, 62072455, 62002355, in part by Beijing Natural Science Foundation (L201001), and in part by Open Research Projects of Zhejiang Lab (NO.2022RC0AB02). + +# References + +[1] Alexander Amini, Wilko Schwarting, Ava Soleimany, and Daniela Rus. Deep evidential regression. In NeurIPS, 2020. 3 +[2] Wentao Bao, Qi Yu, and Yu Kong. Evidential deep learning for open set action recognition. In ICCV, 2021. 3, 7 +[3] Wentao Bao, Qi Yu, and Yu Kong. Opental: Towards open set temporal action localization. In CVPR, 2022. 1, 2, 3, 6, 7 +[4] Abhijit Bendale and Terrance E Boult. Towards open set deep networks. In CVPR, 2016. 2, 7 +[5] Joao Carreira and Andrew Zisserman. Quo vadis, action recognition? a new model and the kinetics dataset. In CVPR, 2017. 7 +[6] Yu-Wei Chao, Sudheendra Vijayanarasimhan, Bryan Seybold, David A Ross, Jia Deng, and Rahul Sukthankar. Rethinking the faster r-cnn architecture for temporal action localization. In CVPR, 2018. 3 +[7] Guangyao Chen, Peixi Peng, Xiangqian Wang, and Yonghong Tian. Adversarial reciprocal points learning for open set recognition. arXiv preprint arXiv:2103.00953, 2021. 2, 7 +[8] Guangyao Chen, Limeng Qiao, Yemin Shi, Peixi Peng, Jia Li, Tiejun Huang, Shiliang Pu, and Yonghong Tian. Learning open set network with discriminative reciprocal points. In ECCV. Springer, 2020. 2 +[9] Mengyuan Chen, Junyu Gao, Shicai Yang, and Changsheng Xu. Dual-evidential learning for weakly-supervised temporal action localization. In ECCV. Springer, 2022. 1, 2, 3 +[10] Junyu Gao, Mengyuan Chen, and Changsheng Xu. Fine-grained temporal contrastive learning for weakly-supervised temporal action localization. In CVPR, 2022. 2, 3, 5, 6 +[11] Junyu Gao and Changsheng Xu. Fast video moment retrieval. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1523-1532, 2021. 1 +[12] Junyu Gao, Tianzhu Zhang, and Changsheng Xu. Graph convolutional tracking. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 4649-4659, 2019. 3 +[13] Junyu Gao, Tianzhu Zhang, and Changsheng Xu. I know the relationships: Zero-shot action recognition via two-stream graph convolutional networks and knowledge graphs. In AAAI, 2019. 5 +[14] Junyu Gao, Tianzhu Zhang, and Changsheng Xu. Smart: Joint sampling and regression for visual tracking. IEEE Transactions on Image Processing, 28(8):3923-3935, 2019. 3 +[15] Junyu Gao, Tianzhu Zhang, and Changsheng Xu. Learning to model relationships for zero-shot video classification. IEEE Transactions on Pattern Analysis and Machine Intelligence, 43(10):3476-3491, 2021. 5 +[16] Chuanxing Geng, Sheng-jun Huang, and Songcan Chen. Recent advances in open set recognition: A survey. IEEE transactions on pattern analysis and machine intelligence, 43(10):3614-3631, 2020. 7 +[17] Zongbo Han, Changqing Zhang, Huazhu Fu, and Joey Tianyi Zhou. Trusted multi-view classification. In ICLR, 2020. 3 + +[18] Bo He, Xitong Yang, Le Kang, Zhiyu Cheng, Xin Zhou, and Abhinav Shrivastava. Asm-loc: Action-aware segment modeling for weakly-supervised temporal action localization. In CVPR, 2022. 1, 5, 7 +[19] Fa-Ting Hong, Jia-Chang Feng, Dan Xu, Ying Shan, and Wei-Shi Zheng. Cross-modal consensus network for weakly supervised temporal action localization. In ACM MM, 2021. 1, 2, 3, 5, 6, 7 +[20] Yibo Hu, Yuzhe Ou, Xujiang Zhao, Jin-Hee Cho, and Feng Chen. Multidimensional uncertainty-aware evidential neural networks. In AAAI, 2021. 3 +[21] Linjiang Huang, Liang Wang, and Hongsheng Li. Weakly supervised temporal action localization via representative snippet knowledge propagation. In CVPR, 2022. 2 +[22] Audun Jsang. Subjective logic: A formalism for reasoning under uncertainty. Springer Verlag, 2016. 3, 4 +[23] Will Kay, Joao Carreira, Karen Simonyan, Brian Zhang, Chloe Hillier, Sudheendra Vijayanarasimhan, Fabio Viola, Tim Green, Trevor Back, Paul Natsev, et al. The kinetics human action video dataset. arXiv:1705.06950, 2017. 3, 7 +[24] Pilhyeon Lee, Jinglu Wang, Yan Lu, and Hyeran Byun. Weakly-supervised temporal action localization by uncertainty modeling. In AAAI, 2021. 2 +[25] Bolian Li, Zongbo Han, Haining Li, Huazhu Fu, and Changqing Zhang. Trustworthy long-tailed classification. arXiv preprint arXiv:2111.09030, 2021. 3 +[26] Chuming Lin, Chengming Xu, Donghao Luo, Yabiao Wang, Ying Tai, Chengjie Wang, Jilin Li, Feiyue Huang, and Yanwei Fu. Learning salient boundary feature for anchor-free temporal action localization. In CVPR, 2021. 7 +[27] Wei Liu, Xiaodong Yue, Yufei Chen, and Thierry Denoeux. Trusted multi-view deep learning with opinion aggregation. In AAAI, 2022. 3 +[28] Vincent Lonij, Ambrish Rawat, and Maria-Irina Nicolae. Open-world visual recognition using knowledge graphs. arXiv preprint arXiv:1708.08310, 2017. 2 +[29] Wang Luo, Tianzhu Zhang, Wenfei Yang, Jingen Liu, Tao Mei, Feng Wu, and Yongdong Zhang. Action unit memory network for weakly supervised temporal action localization. In CVPR, 2021. 2 +[30] Lawrence Neal, Matthew Olson, Xiaoli Fern, Weng-Keen Wong, and Fuxin Li. Open set learning with counterfactual images. In ECCV, 2018. 2 +[31] Poojan Oza and Vishal M Patel. C2ae: Class conditioned auto-encoder for open-set recognition. In CVPR, 2019. 2 +[32] Jeffrey Pennington, Richard Socher, and Christopher D Manning. Glove: Global vectors for word representation. In EMNLP, 2014. 5 +[33] Sanqing Qu, Guang Chen, Zhijun Li, Lijun Zhang, Fan Lu, and Alois Knoll. Acm-net: Action context modeling network for weakly-supervised temporal action localization. arXiv preprint arXiv:2104.02967, 2021. 2, 3, 6 +[34] Walter J Scheirer, Anderson de Rezende Rocha, Archana Sapkota, and Terrance E Boult. Toward open set recognition. IEEE transactions on pattern analysis and machine intelligence, 35(7):1757-1772, 2012. 2 + +[35] Murat Sensoy, Lance Kaplan, and Melih Kandemir. Evidential deep learning to quantify classification uncertainty. In NeurIPS, 2018. 3, 4, 5 +[36] Baifeng Shi, Qi Dai, Yadong Mu, and Jingdong Wang. Weakly-supervised action localization by generative attention modeling. In CVPR, 2020. 2 +[37] Zheng Shou, Hang Gao, Lei Zhang, Kazuyuki Miyazawa, and Shih-Fu Chang. Autoloc: Weakly-supervised temporal action localization in untrimmed videos. In ECCV, 2018. 1 +[38] Krishna Kumar Singh and Yong Jae Lee. Hide-and-seek: Forcing a network to be meticulous for weakly-supervised object and action localization. In ICCV, 2017. 2 +[39] Robyn Speer, Joshua Chin, and Catherine Havasi. Concept-net 5.5: An open multilingual graph of general knowledge. In AAAI, 2017. 5 +[40] Xin Sun, Zhenning Yang, Chi Zhang, Keck-Voon Ling, and Guohao Peng. Conditional gaussian distribution learning for open set recognition. In CVPR, 2020. 2, 7 +[41] Dennis Ulmer. A survey on evidential deep learning for single-pass uncertainty estimation. arXiv preprint arXiv:2110.03051, 2021. 3, 4 +[42] Sagar Vaze, Kai Han, Andrea Vedaldi, and Andrew Zisserman. Open-set recognition: A good closed-set classifier is all you need. In ICLR, 2022. 1 +[43] Limin Wang, Yuanjun Xiong, Dahua Lin, and Luc Van Gool. Untrimmednets for weakly supervised action recognition and detection. In CVPR, 2017. 1, 2 +[44] Huijuan Xu, Abir Das, and Kate Saenko. R-c3d: Region convolutional 3d network for temporal activity detection. In ICCV, 2017. 3 + +[45] Ronald R Yager and Liping Liu. Classic works of the Dempster-Shafer theory of belief functions, volume 219. Springer, 2008. 3 +[46] Wenfei Yang, Tianzhu Zhang, Xiaoyuan Yu, Tian Qi, Yongdong Zhang, and Feng Wu. Uncertainty guided collaborative training for weakly supervised temporal action detection. In CVPR, 2021. 2, 3 +[47] Ryota Yoshihashi, Wen Shao, Rei Kawakami, Shaodi You, Makoto Iida, and Takeshi Naemura. Classification-reconstruction learning for open-set recognition. In CVPR, 2019. 2 +[48] Can Zhang, Meng Cao, Dongming Yang, Jie Chen, and Yuexian Zou. Cola: Weakly-supervised temporal action localization with snippet contrastive learning. In CVPR, 2021. 2 +[49] Chengwei Zhang, Yunlu Xu, Zhanzhan Cheng, Yi Niu, Shiliang Pu, Fei Wu, and Futai Zou. Adversarial seeded sequence growing for weakly-supervised temporal action localization. In ACM MM, 2019. 2 +[50] Jia-Xing Zhong, Nannan Li, Weijie Kong, Tao Zhang, Thomas H Li, and Ge Li. Step-by-step erasion, one-by-one collection: a weakly supervised temporal action detector. In ACM MM, 2018. 2 +[51] Zhun Zhong, Linchao Zhu, Zhiming Luo, Shaozi Li, Yi Yang, and Nicu Sebe. Openmix: Reviving known knowledge for discovering novel visual categories in an open world. In CVPR, 2021. 2 \ No newline at end of file diff --git a/2023/Cascade Evidential Learning for Open-World Weakly-Supervised Temporal Action Localization/images.zip b/2023/Cascade Evidential Learning for Open-World Weakly-Supervised Temporal Action Localization/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..f7da03b8066a19e7f186f3767f0edbaaf7e2d772 --- /dev/null +++ b/2023/Cascade Evidential Learning for Open-World Weakly-Supervised Temporal Action Localization/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b564078a691d2c32d9ee50f0faa436b746075fb59feea42ab34db355cb61c6b6 +size 418593 diff --git a/2023/Cascade Evidential Learning for Open-World Weakly-Supervised Temporal Action Localization/layout.json b/2023/Cascade Evidential Learning for Open-World Weakly-Supervised Temporal Action Localization/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..f765abe7a4d285ad4da83209807e0bb3365216f6 --- /dev/null +++ b/2023/Cascade Evidential Learning for Open-World Weakly-Supervised Temporal Action Localization/layout.json @@ -0,0 +1,9606 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 64, + 135, + 514, + 172 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 135, + 514, + 172 + ], + "spans": [ + { + "bbox": [ + 64, + 135, + 514, + 172 + ], + "type": "text", + "content": "Cascade Evidential Learning for Open-world Weakly-supervised Temporal Action Localization" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 150, + 191, + 426, + 207 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 150, + 191, + 426, + 207 + ], + "spans": [ + { + "bbox": [ + 150, + 191, + 426, + 207 + ], + "type": "text", + "content": "Mengyuan Chen" + }, + { + "bbox": [ + 150, + 191, + 426, + 207 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 150, + 191, + 426, + 207 + ], + "type": "text", + "content": ", Junyu Gao" + }, + { + "bbox": [ + 150, + 191, + 426, + 207 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 150, + 191, + 426, + 207 + ], + "type": "text", + "content": ", and Changsheng Xu" + }, + { + "bbox": [ + 150, + 191, + 426, + 207 + ], + "type": "inline_equation", + "content": "^{1,2,3}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 207, + 472, + 233 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 105, + 207, + 472, + 233 + ], + "spans": [ + { + "bbox": [ + 105, + 207, + 472, + 233 + ], + "type": "text", + "content": "1 State Key Laboratory of Multimodal Artificial Intelligence Systems (MAIS), Institute of Automation, Chinese Academy of Sciences (CASIA)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 83, + 233, + 493, + 246 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 83, + 233, + 493, + 246 + ], + "spans": [ + { + "bbox": [ + 83, + 233, + 493, + 246 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 83, + 233, + 493, + 246 + ], + "type": "text", + "content": " School of Artificial Intelligence, University of Chinese Academy of Sciences (UCAS)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 185, + 246, + 393, + 260 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 185, + 246, + 393, + 260 + ], + "spans": [ + { + "bbox": [ + 185, + 246, + 393, + 260 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 185, + 246, + 393, + 260 + ], + "type": "text", + "content": " Peng Cheng Laboratory, ShenZhen, China" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 134, + 262, + 439, + 274 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 262, + 439, + 274 + ], + "spans": [ + { + "bbox": [ + 134, + 262, + 439, + 274 + ], + "type": "text", + "content": "chenmengyuan2021@ia.ac.cn; { junyu.gao, csxu}@nlpr.ia.ac.cn" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 139, + 301, + 186, + 314 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 139, + 301, + 186, + 314 + ], + "spans": [ + { + "bbox": [ + 139, + 301, + 186, + 314 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 45, + 327, + 282, + 595 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 327, + 282, + 595 + ], + "spans": [ + { + "bbox": [ + 45, + 327, + 282, + 595 + ], + "type": "text", + "content": "Targeting at recognizing and localizing action instances with only video-level labels during training, Weakly-supervised Temporal Action Localization (WTAL) has achieved significant progress in recent years. However, living in the dynamically changing open world where unknown actions constantly spring up, the closed-set assumption of existing WTAL methods is invalid. Compared with traditional open-set recognition tasks, Open-world WTAL (OW-TAL) is challenging since not only are the annotations of unknown samples unavailable, but also the fine-grained annotations of known action instances can only be inferred ambiguously from the video category labels. To address this problem, we propose a Cascade Evidential Learning framework at an evidence level, which targets at OWTAL for the first time. Our method jointly leverages multi-scale temporal contexts and knowledge-guided prototype information to progressively collect cascade and enhanced evidence for known action, unknown action, and background separation. Extensive experiments conducted on THUMOS-14 and ActivityNet-v1.3 verify the effectiveness of our method. Besides the classification metrics adopted by previous open-set recognition methods, we also evaluate our method on localization metrics which are more reasonable for OWTAL." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 629, + 124, + 642 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 629, + 124, + 642 + ], + "spans": [ + { + "bbox": [ + 46, + 629, + 124, + 642 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 45, + 648, + 280, + 730 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 648, + 280, + 730 + ], + "spans": [ + { + "bbox": [ + 45, + 648, + 280, + 730 + ], + "type": "text", + "content": "Targeting at recognizing and localizing action instances with only video-level labels during training, Weakly-supervised Temporal Action Localization (WTAL) has attracted increasing attention from both academia and industry [9, 11, 18, 19, 37, 43]. Unlike fully-supervised TAL, WTAL only requires video-level action labels during training. However, the closed-set assumption of existing WTAL" + } + ] + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 299, + 310, + 412, + 386 + ], + "blocks": [ + { + "bbox": [ + 329, + 300, + 383, + 310 + ], + "lines": [ + { + "bbox": [ + 329, + 300, + 383, + 310 + ], + "spans": [ + { + "bbox": [ + 329, + 300, + 383, + 310 + ], + "type": "text", + "content": "Training Phase" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 299, + 310, + 412, + 386 + ], + "lines": [ + { + "bbox": [ + 299, + 310, + 412, + 386 + ], + "spans": [ + { + "bbox": [ + 299, + 310, + 412, + 386 + ], + "type": "image", + "image_path": "daa5b0178a8ff34dde03014efcda7b9b40d5693769d885a40a2c4ce4507a0d42.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 414, + 310, + 529, + 387 + ], + "blocks": [ + { + "bbox": [ + 449, + 300, + 500, + 309 + ], + "lines": [ + { + "bbox": [ + 449, + 300, + 500, + 309 + ], + "spans": [ + { + "bbox": [ + 449, + 300, + 500, + 309 + ], + "type": "text", + "content": "Testing Phase" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 414, + 310, + 529, + 387 + ], + "lines": [ + { + "bbox": [ + 414, + 310, + 529, + 387 + ], + "spans": [ + { + "bbox": [ + 414, + 310, + 529, + 387 + ], + "type": "image", + "image_path": "b1b53169730a0881bee51cdd58636c5d7c6bb46c1f52ac76484b9c323c349076.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 296, + 394, + 531, + 426 + ], + "lines": [ + { + "bbox": [ + 296, + 394, + 531, + 426 + ], + "spans": [ + { + "bbox": [ + 296, + 394, + 531, + 426 + ], + "type": "text", + "content": "Figure 1. Illustration of the training and testing phases of OWTAL. With only video-level labels for training, OWTAL aims to localize both known and unknown action instances in testing videos." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "bbox": [ + 296, + 436, + 531, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 436, + 531, + 495 + ], + "spans": [ + { + "bbox": [ + 296, + 436, + 531, + 495 + ], + "type": "text", + "content": "methods is invalid in the dynamically changing real world, since with the development of society never-before-seen human action categories are constantly emerging. Therefore, to address this problem, we consider a different WTAL setting in this work, termed as Open-world WTAL (OWTAL)." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 296, + 497, + 532, + 730 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 497, + 532, + 730 + ], + "spans": [ + { + "bbox": [ + 296, + 497, + 532, + 730 + ], + "type": "text", + "content": "Different from the traditional WTAL task, as shown in Figure 1, OWTAL allows testing videos to contain action instances of unknown categories, which have never appeared during training. Therefore, temporal boundaries of both known and unknown action instances are expected to be predicted. Compared with its fully-supervised counterpart Open Set TAL [3], OWTAL is challenging in two aspects: (1) Ambiguity of annotations of closed-set (known) action instances. Previous works indicate that the closed-set and open-set performance are highly correlated [42]. However, under the OWTAL setting, not only are the annotations of unknown action instances unavailable, but also the fine-grained annotations of known ones can only be inferred ambiguously from the video category labels. During training, the known action instances that the model needs to focus on are prone to be disturbed by the background snippets, which hinders the learning of the closed-set actions, thus making it extremely difficult to differentiate the unknown actions, the known actions, and the background. (2) Lack of reasonable metrics. The traditional Open Set Recognition (OSR) aims" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 77, + 42, + 119, + 72 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 42, + 119, + 72 + ], + "spans": [ + { + "bbox": [ + 77, + 42, + 119, + 72 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 150, + 41, + 471, + 74 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 150, + 41, + 471, + 74 + ], + "spans": [ + { + "bbox": [ + 150, + 41, + 471, + 74 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 277, + 746, + 300, + 756 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 277, + 746, + 300, + 756 + ], + "spans": [ + { + "bbox": [ + 277, + 746, + 300, + 756 + ], + "type": "text", + "content": "14741" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 595, + 842 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 45, + 106, + 279, + 141 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 106, + 279, + 141 + ], + "spans": [ + { + "bbox": [ + 45, + 106, + 279, + 141 + ], + "type": "text", + "content": "for classification while the goal of OWTAL is to perform localization instead, thus the classification metrics commonly adopted by OSR are not sufficient for OWTAL." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 45, + 143, + 281, + 561 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 143, + 281, + 561 + ], + "spans": [ + { + "bbox": [ + 45, + 143, + 281, + 561 + ], + "type": "text", + "content": "In order to alleviate the negative impact caused by the weak annotations of known action instances, we propose a Cascade Evidential Learning method for owtaL (CELL), which progressively collects cascaded evidence by considering both temporal contexts in multi-scale ranges and inter-video correlations under the guidance of prior knowledge. Since the goal of OWTAL is to locate the consecutive known/unknown action segments of various temporal lengths in open-world scenarios, perceiving temporal contexts in diverse ranges is essential. We argue that it is meaningful to endow individual snippet features with the ability of sensing multi-scale neighborhood video segments, and thus a Multi-scale Extended-range Perception module (Section 3.2) is designed to obtain more discriminative video features for initial evidence collection, taking advantage of the temporal contexts. Due to the large intra-action variation in visual patterns and the lack of prior knowledge guidance, the known action instances which visually deviate from the common ones are likely to be misidentified with the initial evidence collected from individual videos. Therefore, we design a Knowledge-guided Bipolar Prototype Learning strategy (Section 3.3), where a semantic relation graph is constructed to provide prior knowledge guidance for the bipolar prototype learning among videos, thus perceiving the open-world more comprehensively. We use this strategy to generate a series of evidence calibration factors for further cascade evidence enhancement. Finally, a Cascade Evidence Enhancement module (Section 3.4) is designed for enhancing the initial evidence with the calibration factors, and the uncertainty estimated from the cascaded evidence is used for the known/unknown judgment. Extensive experiments conducted on THUMOS-14 and ActivityNet verify the effectiveness. Besides the various classification metrics adopted by previous works, we also evaluate our method on localization metrics which are more in line with the needs of real-world applications." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 57, + 562, + 232, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 562, + 232, + 574 + ], + "spans": [ + { + "bbox": [ + 57, + 562, + 232, + 574 + ], + "type": "text", + "content": "To summarize, our contribution is threefold:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 50, + 576, + 279, + 730 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 50, + 576, + 279, + 645 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 50, + 576, + 279, + 645 + ], + "spans": [ + { + "bbox": [ + 50, + 576, + 279, + 645 + ], + "type": "text", + "content": "- To tackle the unique challenges of OWTAL, we propose a cascade evidential learning framework, which progressively collects comprehensive evidence for known action, unknown action, and background separation. Localization metrics which meet the needs of the real-world more closely are adopted for evaluation." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 50, + 647, + 279, + 694 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 50, + 647, + 279, + 694 + ], + "spans": [ + { + "bbox": [ + 50, + 647, + 279, + 694 + ], + "type": "text", + "content": "- To achieve OWTAL without fine-grained annotations, the proposed CELL jointly leverages multi-scale temporal contexts and knowledge-guided prototype information during the evidence cascade learning process." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 50, + 695, + 279, + 730 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 50, + 695, + 279, + 730 + ], + "spans": [ + { + "bbox": [ + 50, + 695, + 279, + 730 + ], + "type": "text", + "content": "- We conduct comprehensive experiments on two popular WTAL benchmarks, THUMOS-14 and ActivityNetv1.3, and achieve significant performance improvement" + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 310, + 106, + 530, + 141 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 310, + 106, + 530, + 141 + ], + "spans": [ + { + "bbox": [ + 310, + 106, + 530, + 141 + ], + "type": "text", + "content": "over various baselines. Experiments show that CELL enables existing methods to well adapt to the more practical open-world settings." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 297, + 152, + 381, + 164 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 152, + 381, + 164 + ], + "spans": [ + { + "bbox": [ + 297, + 152, + 381, + 164 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 296, + 170, + 530, + 390 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 170, + 530, + 390 + ], + "spans": [ + { + "bbox": [ + 296, + 170, + 530, + 390 + ], + "type": "text", + "content": "Weakly-supervised Temporal Action Localization. Originated from UntrimmedNet [43], the pioneer work to utilize video-level action category annotations as the weak supervision for TAL, existing WTAL methods commonly adopt a multiple instance learning strategy and can be roughly categorized into erasing-based, attention-based and uncertainty-based methods. Erasing-based methods [38, 49, 50] erase the most discriminative segments to mitigate the single snippet cheating issue [48]. Attention-based methods [10, 19, 21, 29, 33, 36] employ the attention mechanism to select snippets most likely to be the foreground with activation scores. The newly-arising uncertainty-based methods [9, 24, 46] utilize uncertainty to address the inevitable action-background ambiguity caused by the weakly-supervised setting. However, all existing WTAL methods are rooted in the closed-set assumption, which impedes their application to real-world scenarios. In contrast, we address the WTAL task under the open-world setting by collecting reliable cascaded evidence, which has not been explored in prior work." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 296, + 392, + 530, + 612 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 392, + 530, + 612 + ], + "spans": [ + { + "bbox": [ + 296, + 392, + 530, + 612 + ], + "type": "text", + "content": "Open Set Recognition. Open Set Recognition (OSR) aims to recognize known classes and reject the unknown. [34] first formalize the OSR problem and provide a basic framework. [4] propose the first DNN-based OSR method Openmax, which rejects unknown classes by modeling the distance of activation vectors with Extreme Value Theory (EVT). GAN-based methods are also developed towards this task. [30] generate samples similar to the training data but not belonging to the known classes, and then utilize the generated unknown class samples to train an open-set classifier. Other approaches include prototype-based methods [7, 8], which reject unknown classes by calculating the maximum distance between the input sample and the learned closed-set prototypes, and reconstruction-based methods [31, 40, 47] utilize the reconstruction error in the test phase as an open-set indicator. Some works [28, 51] attempt to seek assistance from external knowledge, but few of them have explicitly modeled unknown class information in an end-to-end framework." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 296, + 613, + 530, + 730 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 613, + 530, + 730 + ], + "spans": [ + { + "bbox": [ + 296, + 613, + 530, + 730 + ], + "type": "text", + "content": "Most existing approaches focus on open-set recognition tasks while few works attend to temporal localization-related vision field. [3] formalize the open-set temporal action localization problem and propose an OpenTAL framework. However, its fully-supervised setting requires large amounts of fine-grained labels whose annotation is time-consuming, error-prone, and costly. Besides, the evaluation metrics in [3] are mainly classification ones. In contrast, in this paper, we address the challenging OWTAL problem which also abandons the closed-set assumption but only re" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 277, + 746, + 301, + 756 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 277, + 746, + 301, + 756 + ], + "spans": [ + { + "bbox": [ + 277, + 746, + 301, + 756 + ], + "type": "text", + "content": "14742" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 595, + 842 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 45, + 106, + 278, + 129 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 106, + 278, + 129 + ], + "spans": [ + { + "bbox": [ + 45, + 106, + 278, + 129 + ], + "type": "text", + "content": "quires easily available video-level class labels, and employ comprehensive localization-related metrics for evaluation." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 45, + 130, + 279, + 351 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 130, + 279, + 351 + ], + "spans": [ + { + "bbox": [ + 45, + 130, + 279, + 351 + ], + "type": "text", + "content": "Evidential Deep Learning. Based on Dempster-Shafer Theory (DST) [45] and Subjective Logic theory (SL) [22], Evidential Deep Learning (EDL) allows uncertainty estimation in a single forward pass [41] by collecting scalar evidence for each category and parameterizing a Dirichlet distribution, which models the distribution of classification probabilities, over the collected evidence. As a newly arising trustworthy method, EDL has achieved remarkable progress in various computer vision tasks, including action localization [9], image classification [35], regression [1], multi-view classification [17, 27], out-of-distribution detection [20], long-tail learning [25], and open-set action recognition and localization [2, 3]. However, current EDL algorithms only focus on collecting evidence from individual samples, while they neglect the complementary role that the potential correlations among samples can play in evidence collection. To address this issue, in this work we propose a cascade evidence calibration paradigm which effectively enhances the collected evidence." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 45, + 361, + 134, + 374 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 361, + 134, + 374 + ], + "spans": [ + { + "bbox": [ + 45, + 361, + 134, + 374 + ], + "type": "text", + "content": "3. Our Approach" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 45, + 375, + 195, + 387 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 375, + 195, + 387 + ], + "spans": [ + { + "bbox": [ + 45, + 375, + 195, + 387 + ], + "type": "text", + "content": "3.1. Notations and Preliminaries" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 45, + 392, + 279, + 542 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 392, + 279, + 542 + ], + "spans": [ + { + "bbox": [ + 45, + 392, + 279, + 542 + ], + "type": "text", + "content": "Problem formulation of OWTAL. OWTAL targets at detecting known/unknown action instances in untrimmed videos. Formally, we are given a training set " + }, + { + "bbox": [ + 45, + 392, + 279, + 542 + ], + "type": "inline_equation", + "content": "\\{\\mathcal{V},\\mathcal{Y}\\}" + }, + { + "bbox": [ + 45, + 392, + 279, + 542 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 45, + 392, + 279, + 542 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 45, + 392, + 279, + 542 + ], + "type": "text", + "content": " training videos, where " + }, + { + "bbox": [ + 45, + 392, + 279, + 542 + ], + "type": "inline_equation", + "content": "\\mathcal{V}" + }, + { + "bbox": [ + 45, + 392, + 279, + 542 + ], + "type": "text", + "content": " denotes an untrimmed video, and " + }, + { + "bbox": [ + 45, + 392, + 279, + 542 + ], + "type": "inline_equation", + "content": "\\pmb {y}\\in \\mathbb{R}^{C}" + }, + { + "bbox": [ + 45, + 392, + 279, + 542 + ], + "type": "text", + "content": " is its multi-hot label indicating the action categories that the action instances in this video belong to, where " + }, + { + "bbox": [ + 45, + 392, + 279, + 542 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 45, + 392, + 279, + 542 + ], + "type": "text", + "content": " is the number of known action categories. During testing, the goal is to use the learned model to output a set of quadruplets " + }, + { + "bbox": [ + 45, + 392, + 279, + 542 + ], + "type": "inline_equation", + "content": "\\{c_r,t_r^s,t_r^e,\\gamma_r\\}_{r = 1}^R" + }, + { + "bbox": [ + 45, + 392, + 279, + 542 + ], + "type": "text", + "content": " ,where " + }, + { + "bbox": [ + 45, + 392, + 279, + 542 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 45, + 392, + 279, + 542 + ], + "type": "text", + "content": " is the number of action instances in " + }, + { + "bbox": [ + 45, + 392, + 279, + 542 + ], + "type": "inline_equation", + "content": "\\mathcal{V}" + }, + { + "bbox": [ + 45, + 392, + 279, + 542 + ], + "type": "inline_equation", + "content": "c_{r},t_{r}^{s},t_{r}^{e}" + }, + { + "bbox": [ + 45, + 392, + 279, + 542 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 45, + 392, + 279, + 542 + ], + "type": "inline_equation", + "content": "\\gamma_{r}" + }, + { + "bbox": [ + 45, + 392, + 279, + 542 + ], + "type": "text", + "content": " denote the action category, the start and end timestamps, and the confidence score, respectively. " + }, + { + "bbox": [ + 45, + 392, + 279, + 542 + ], + "type": "inline_equation", + "content": "c_{r}\\in \\{1,\\dots,C + 1\\}" + }, + { + "bbox": [ + 45, + 392, + 279, + 542 + ], + "type": "text", + "content": " , where " + }, + { + "bbox": [ + 45, + 392, + 279, + 542 + ], + "type": "inline_equation", + "content": "C + 1" + }, + { + "bbox": [ + 45, + 392, + 279, + 542 + ], + "type": "text", + "content": " denotes the unknown action category." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 45, + 543, + 279, + 636 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 543, + 279, + 636 + ], + "spans": [ + { + "bbox": [ + 45, + 543, + 279, + 636 + ], + "type": "text", + "content": "Feature extraction. Following previous works [10,46], we split the untrimmed video " + }, + { + "bbox": [ + 45, + 543, + 279, + 636 + ], + "type": "inline_equation", + "content": "\\mathcal{V}" + }, + { + "bbox": [ + 45, + 543, + 279, + 636 + ], + "type": "text", + "content": " into " + }, + { + "bbox": [ + 45, + 543, + 279, + 636 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 45, + 543, + 279, + 636 + ], + "type": "text", + "content": " non-overlapping 16-frame snippets and then use pre-trained networks, e.g., the I3D model [23], to extract RGB and optical flow features. Then, the above features are concatenated and then fed into a fusion module, e.g., convolutional layers [19, 33], to obtain the snippet-wise feature " + }, + { + "bbox": [ + 45, + 543, + 279, + 636 + ], + "type": "inline_equation", + "content": "\\pmb{X} = [\\pmb{x}_1,\\dots,\\pmb{x}_T]\\in \\mathbb{R}^{T\\times D}" + }, + { + "bbox": [ + 45, + 543, + 279, + 636 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 45, + 543, + 279, + 636 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 45, + 543, + 279, + 636 + ], + "type": "text", + "content": " is the feature dimension." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 45, + 644, + 279, + 667 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 644, + 279, + 667 + ], + "spans": [ + { + "bbox": [ + 45, + 644, + 279, + 667 + ], + "type": "text", + "content": "3.2. Multi-scale Extended-range Perception for Initial Evidence Collection" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 45, + 671, + 279, + 730 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 671, + 279, + 730 + ], + "spans": [ + { + "bbox": [ + 45, + 671, + 279, + 730 + ], + "type": "text", + "content": "A video may contain several known/unknown action instances with various temporal lengths in the open-world. Since the goal of OWTAL is to locate these consecutive video segments, perceiving temporal contexts in diverse ranges is essential. We argue that it is meaningful to endow" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 296, + 106, + 530, + 221 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 106, + 530, + 221 + ], + "spans": [ + { + "bbox": [ + 296, + 106, + 530, + 221 + ], + "type": "text", + "content": "individual snippet features with the ability of sensing multiscale neighborhood video segments. As shown in Figure 2, we design a Multi-scale Extended-range Perception module (MEP) to obtain more discriminative video features, taking advantage of the temporal contexts. Note that some anchor-based temporal modeling methods [6, 12, 14, 44] also leverage multi-scale temporal information. However, their goal is to obtain the fused temporal feature for proposal generation, while the MEP aims to enhance individual snippet features for initial evidence collection." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 296, + 222, + 531, + 339 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 222, + 531, + 339 + ], + "spans": [ + { + "bbox": [ + 296, + 222, + 531, + 339 + ], + "type": "text", + "content": "In MEP, each snippet " + }, + { + "bbox": [ + 296, + 222, + 531, + 339 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 296, + 222, + 531, + 339 + ], + "type": "text", + "content": " is assigned with a series of " + }, + { + "bbox": [ + 296, + 222, + 531, + 339 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 296, + 222, + 531, + 339 + ], + "type": "text", + "content": "-centered multi-scale video segments, whose context-aware features can contribute to the extension of snippet-specific perception ranges. Here, the temporal boundaries of the " + }, + { + "bbox": [ + 296, + 222, + 531, + 339 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 296, + 222, + 531, + 339 + ], + "type": "text", + "content": "-centered video segments can be denoted as set " + }, + { + "bbox": [ + 296, + 222, + 531, + 339 + ], + "type": "inline_equation", + "content": "\\Omega_t = \\{(s_t^m, e_t^m) | m \\in \\mathcal{M}\\}" + }, + { + "bbox": [ + 296, + 222, + 531, + 339 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 296, + 222, + 531, + 339 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 296, + 222, + 531, + 339 + ], + "type": "text", + "content": " is a pre-defined length set, " + }, + { + "bbox": [ + 296, + 222, + 531, + 339 + ], + "type": "inline_equation", + "content": "s_t^m = \\max(0, t - m)" + }, + { + "bbox": [ + 296, + 222, + 531, + 339 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 296, + 222, + 531, + 339 + ], + "type": "inline_equation", + "content": "e_t^m = \\min(T, t + m)" + }, + { + "bbox": [ + 296, + 222, + 531, + 339 + ], + "type": "text", + "content": ". For brevity, we denote the temporal average of the feature sequence of the extended video segment " + }, + { + "bbox": [ + 296, + 222, + 531, + 339 + ], + "type": "inline_equation", + "content": "(s_t^m, e_t^m)" + }, + { + "bbox": [ + 296, + 222, + 531, + 339 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 296, + 222, + 531, + 339 + ], + "type": "inline_equation", + "content": "f_t^m \\in \\mathbb{R}^D" + }, + { + "bbox": [ + 296, + 222, + 531, + 339 + ], + "type": "text", + "content": ", which can be given by:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 346, + 342, + 530, + 372 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 346, + 342, + 530, + 372 + ], + "spans": [ + { + "bbox": [ + 346, + 342, + 530, + 372 + ], + "type": "interline_equation", + "content": "\\boldsymbol {f} _ {t} ^ {m} = \\frac {1}{e _ {t} ^ {m} - s _ {t} ^ {m} + 1} \\sum_ {s _ {t} ^ {m} \\leq i \\leq e _ {t} ^ {m}} \\boldsymbol {x} _ {i}. \\tag {1}", + "image_path": "b242d6bd6dc1e0f199e4d59c81e65eb5d75b1c6e0733038d514bd032f0f69959.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 296, + 377, + 530, + 482 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 377, + 530, + 482 + ], + "spans": [ + { + "bbox": [ + 296, + 377, + 530, + 482 + ], + "type": "text", + "content": "To extend the perception range of snippet " + }, + { + "bbox": [ + 296, + 377, + 530, + 482 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 296, + 377, + 530, + 482 + ], + "type": "text", + "content": ", we perform feature enhancement by fusing " + }, + { + "bbox": [ + 296, + 377, + 530, + 482 + ], + "type": "inline_equation", + "content": "\\boldsymbol{x}_t" + }, + { + "bbox": [ + 296, + 377, + 530, + 482 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 296, + 377, + 530, + 482 + ], + "type": "inline_equation", + "content": "\\boldsymbol{f}_t^m" + }, + { + "bbox": [ + 296, + 377, + 530, + 482 + ], + "type": "text", + "content": ", during which the weight of " + }, + { + "bbox": [ + 296, + 377, + 530, + 482 + ], + "type": "inline_equation", + "content": "\\boldsymbol{f}_t^m" + }, + { + "bbox": [ + 296, + 377, + 530, + 482 + ], + "type": "text", + "content": " is positively correlated with the similarity between them. Such operation is motivated by the simple expectation that a snippet-level feature would absorb meaningful contextual information mainly from extended video segments that are highly relevant to it. Therefore, the snippet-level feature with extended-range perception, denoted as " + }, + { + "bbox": [ + 296, + 377, + 530, + 482 + ], + "type": "inline_equation", + "content": "\\tilde{\\boldsymbol{F}} = [\\tilde{\\boldsymbol{f}}_1,\\dots ,\\tilde{\\boldsymbol{f}}_T]" + }, + { + "bbox": [ + 296, + 377, + 530, + 482 + ], + "type": "text", + "content": ", can be formulated as:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 311, + 487, + 530, + 511 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 487, + 530, + 511 + ], + "spans": [ + { + "bbox": [ + 311, + 487, + 530, + 511 + ], + "type": "interline_equation", + "content": "\\tilde {\\boldsymbol {f}} _ {t} = \\left(1 - \\alpha_ {t}\\right) \\varphi_ {1} (\\boldsymbol {x} _ {t}) + \\alpha_ {t} \\sum_ {m \\in \\mathcal {M}} \\delta \\left(\\omega_ {t} ^ {m}\\right) \\varphi_ {2} \\left(\\boldsymbol {f} _ {t} ^ {m}\\right), \\tag {2}", + "image_path": "9ab8adeb1f6d2d5eb435fc98937e55010fa5fd85c9cf1ef2444212a92f6185ca.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 296, + 516, + 530, + 575 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 516, + 530, + 575 + ], + "spans": [ + { + "bbox": [ + 296, + 516, + 530, + 575 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 296, + 516, + 530, + 575 + ], + "type": "inline_equation", + "content": "\\delta (\\cdot)" + }, + { + "bbox": [ + 296, + 516, + 530, + 575 + ], + "type": "text", + "content": " denotes the Softmax operation with a temperature factor, " + }, + { + "bbox": [ + 296, + 516, + 530, + 575 + ], + "type": "inline_equation", + "content": "\\varphi_{1,2}" + }, + { + "bbox": [ + 296, + 516, + 530, + 575 + ], + "type": "text", + "content": " are fully-connected layers for feature embedding, " + }, + { + "bbox": [ + 296, + 516, + 530, + 575 + ], + "type": "inline_equation", + "content": "\\alpha_{t} \\in [0,1]" + }, + { + "bbox": [ + 296, + 516, + 530, + 575 + ], + "type": "text", + "content": " is the fusion weight representing the scaled average of the cosine similarities " + }, + { + "bbox": [ + 296, + 516, + 530, + 575 + ], + "type": "inline_equation", + "content": "\\omega_{t}^{m}" + }, + { + "bbox": [ + 296, + 516, + 530, + 575 + ], + "type": "text", + "content": " between " + }, + { + "bbox": [ + 296, + 516, + 530, + 575 + ], + "type": "inline_equation", + "content": "\\pmb{x}_{t}" + }, + { + "bbox": [ + 296, + 516, + 530, + 575 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 296, + 516, + 530, + 575 + ], + "type": "inline_equation", + "content": "\\pmb{f}_t^m" + }, + { + "bbox": [ + 296, + 516, + 530, + 575 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 296, + 516, + 530, + 575 + ], + "type": "inline_equation", + "content": "\\alpha_{t}" + }, + { + "bbox": [ + 296, + 516, + 530, + 575 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 296, + 516, + 530, + 575 + ], + "type": "inline_equation", + "content": "\\omega_t^m" + }, + { + "bbox": [ + 296, + 516, + 530, + 575 + ], + "type": "text", + "content": " can be calculated by:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 302, + 580, + 530, + 618 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 580, + 530, + 618 + ], + "spans": [ + { + "bbox": [ + 302, + 580, + 530, + 618 + ], + "type": "interline_equation", + "content": "\\alpha_ {t} = \\frac {1}{2 | \\mathcal {M} |} \\sum_ {m \\in \\mathcal {M}} \\left(\\omega_ {t} ^ {m} + 1\\right), \\omega_ {t} ^ {m} = \\cos \\left(\\frac {\\boldsymbol {x} _ {t}}{\\left| \\boldsymbol {x} _ {t} \\right|}, \\frac {\\boldsymbol {f} _ {t} ^ {m}}{\\left| \\boldsymbol {f} _ {t} ^ {m} \\right|}\\right). \\tag {3}", + "image_path": "3303b3113f22b0f38714f459ca6c704c755f85bf4ea32ba1f1947826f4210910.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 296, + 619, + 530, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 619, + 530, + 700 + ], + "spans": [ + { + "bbox": [ + 296, + 619, + 530, + 700 + ], + "type": "text", + "content": "Thereafter, similar to previous methods [10, 19], we utilize an attention module to predict an attention score sequence " + }, + { + "bbox": [ + 296, + 619, + 530, + 700 + ], + "type": "inline_equation", + "content": "\\mathbf{A} = [A_{1},\\dots,A_{T}]\\in \\mathbb{R}^{T}" + }, + { + "bbox": [ + 296, + 619, + 530, + 700 + ], + "type": "text", + "content": " for the input video " + }, + { + "bbox": [ + 296, + 619, + 530, + 700 + ], + "type": "inline_equation", + "content": "\\nu" + }, + { + "bbox": [ + 296, + 619, + 530, + 700 + ], + "type": "text", + "content": " representing the probabilities of snippets belonging to the foreground, then select the top-ranked snippets. Specifically, for the input video we obtain the set " + }, + { + "bbox": [ + 296, + 619, + 530, + 700 + ], + "type": "inline_equation", + "content": "\\Theta" + }, + { + "bbox": [ + 296, + 619, + 530, + 700 + ], + "type": "text", + "content": " of snippets with the top-" + }, + { + "bbox": [ + 296, + 619, + 530, + 700 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 296, + 619, + 530, + 700 + ], + "type": "text", + "content": " attention scores by the following equation:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 311, + 705, + 530, + 731 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 705, + 530, + 731 + ], + "spans": [ + { + "bbox": [ + 311, + 705, + 530, + 731 + ], + "type": "interline_equation", + "content": "\\Theta = \\arg \\max _ {\\Theta} \\sum_ {t \\in \\Theta , | \\Theta | = k} A _ {t}, \\quad \\Theta \\subset \\{1, 2, \\dots , T \\}, \\tag {4}", + "image_path": "1efb7ac01bbf05e796707b02bb5a03c63852d38925dd8f7838a1f49170bc1d31.jpg" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 277, + 746, + 300, + 756 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 277, + 746, + 300, + 756 + ], + "spans": [ + { + "bbox": [ + 277, + 746, + 300, + 756 + ], + "type": "text", + "content": "14743" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 595, + 842 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 95, + 104, + 480, + 314 + ], + "blocks": [ + { + "bbox": [ + 95, + 104, + 480, + 314 + ], + "lines": [ + { + "bbox": [ + 95, + 104, + 480, + 314 + ], + "spans": [ + { + "bbox": [ + 95, + 104, + 480, + 314 + ], + "type": "image", + "image_path": "395baac5f70b6598677f6663cca078ae6d31e0f1e20fd363883ca19a848f68ff.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 45, + 318, + 531, + 372 + ], + "lines": [ + { + "bbox": [ + 45, + 318, + 531, + 372 + ], + "spans": [ + { + "bbox": [ + 45, + 318, + 531, + 372 + ], + "type": "text", + "content": "Figure 2. Overall framework of the proposed CELL for OWTAL. Firstly, a Multi-scale Extended-range Perception (MEP) module is designed to collect the initial evidence by perceiving temporal contexts in diverse ranges. Then we propose a Knowledge-guided Bipolar Prototype Learning (KBPL) module to obtain evidence calibration factors from inter-video correlations and prior knowledge. Finally, a Cascade Evidence Enhancement (CEE) module is designed for enhancing the initial evidence with the calibration factors, and the uncertainty derived from the cascaded evidence is used for the known/unknown judgment." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 45, + 378, + 279, + 414 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 378, + 279, + 414 + ], + "spans": [ + { + "bbox": [ + 45, + 378, + 279, + 414 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 45, + 378, + 279, + 414 + ], + "type": "inline_equation", + "content": "k = \\lceil T / r \\rceil" + }, + { + "bbox": [ + 45, + 378, + 279, + 414 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 45, + 378, + 279, + 414 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 45, + 378, + 279, + 414 + ], + "type": "text", + "content": " is a scaling factor. According to the snippet index in " + }, + { + "bbox": [ + 45, + 378, + 279, + 414 + ], + "type": "inline_equation", + "content": "\\Theta" + }, + { + "bbox": [ + 45, + 378, + 279, + 414 + ], + "type": "text", + "content": ", we select them from " + }, + { + "bbox": [ + 45, + 378, + 279, + 414 + ], + "type": "inline_equation", + "content": "\\tilde{\\pmb{F}}" + }, + { + "bbox": [ + 45, + 378, + 279, + 414 + ], + "type": "text", + "content": " and calculate the average as the video-level feature, denoted as " + }, + { + "bbox": [ + 45, + 378, + 279, + 414 + ], + "type": "inline_equation", + "content": "\\tilde{F}_{vid}" + }, + { + "bbox": [ + 45, + 378, + 279, + 414 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 45, + 414, + 279, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 414, + 279, + 449 + ], + "spans": [ + { + "bbox": [ + 45, + 414, + 279, + 449 + ], + "type": "text", + "content": "With the multi-scale-context-aware video-level feature, we feed " + }, + { + "bbox": [ + 45, + 414, + 279, + 449 + ], + "type": "inline_equation", + "content": "\\tilde{\\pmb{F}}_{vid}" + }, + { + "bbox": [ + 45, + 414, + 279, + 449 + ], + "type": "text", + "content": " into an evidence collector " + }, + { + "bbox": [ + 45, + 414, + 279, + 449 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 45, + 414, + 279, + 449 + ], + "type": "text", + "content": ", a DNN parameterized by " + }, + { + "bbox": [ + 45, + 414, + 279, + 449 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 45, + 414, + 279, + 449 + ], + "type": "text", + "content": ", to collect initial evidence " + }, + { + "bbox": [ + 45, + 414, + 279, + 449 + ], + "type": "inline_equation", + "content": "e_{init} \\in \\mathbb{R}^C" + }, + { + "bbox": [ + 45, + 414, + 279, + 449 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 116, + 458, + 279, + 472 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 116, + 458, + 279, + 472 + ], + "spans": [ + { + "bbox": [ + 116, + 458, + 279, + 472 + ], + "type": "interline_equation", + "content": "\\boldsymbol {e} _ {\\text {i n i t}} = g \\left(f \\left(\\tilde {\\boldsymbol {F}} _ {v i d}; \\boldsymbol {\\theta}\\right)\\right), \\tag {5}", + "image_path": "3d4083828ed1f788c92960b21891e7eb9052f7834a0dc8331a6fa11c1aa7bc34.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 45, + 481, + 279, + 505 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 481, + 279, + 505 + ], + "spans": [ + { + "bbox": [ + 45, + 481, + 279, + 505 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 45, + 481, + 279, + 505 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 45, + 481, + 279, + 505 + ], + "type": "text", + "content": " denotes a scale function, e.g., ReLU, Softplus or Exp, to ensure the collected evidence non-negative." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 45, + 505, + 279, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 505, + 279, + 586 + ], + "spans": [ + { + "bbox": [ + 45, + 505, + 279, + 586 + ], + "type": "text", + "content": "Based on Subjective Logic theory [22], Evidential Deep Learning (EDL) allows uncertainty estimation in a single forward pass [41] by collecting evidence of each category and parameterizing a Dirichlet distribution, which models the distribution of class probabilities, over the collected evidence. Our collected initial evidence " + }, + { + "bbox": [ + 45, + 505, + 279, + 586 + ], + "type": "inline_equation", + "content": "e_{init}" + }, + { + "bbox": [ + 45, + 505, + 279, + 586 + ], + "type": "text", + "content": " can be optimized by the following loss function:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 85, + 596, + 279, + 628 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 85, + 596, + 279, + 628 + ], + "spans": [ + { + "bbox": [ + 85, + 596, + 279, + 628 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {i n i t}} = \\sum_ {i = 1} ^ {C} y _ {i} \\left(\\log S _ {\\text {i n i t}} - \\log \\alpha_ {\\text {i n i t}, i}\\right), \\tag {6}", + "image_path": "ff16ae0541e0cabda2b60a2be55c2560b4048ad4ca0d178d8dd81e98329667b0.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 45, + 638, + 280, + 732 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 638, + 280, + 732 + ], + "spans": [ + { + "bbox": [ + 45, + 638, + 280, + 732 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 45, + 638, + 280, + 732 + ], + "type": "inline_equation", + "content": "S_{init} = \\sum_{i}^{C}\\alpha_{init,i},\\alpha_{init,i} = e_{init,i} + 1" + }, + { + "bbox": [ + 45, + 638, + 280, + 732 + ], + "type": "text", + "content": ", and Eq. (6) is actually the deformation of the negative logarithm of the marginal likelihood in the EDL paradigm [35]. Besides, the video-level classification uncertainty " + }, + { + "bbox": [ + 45, + 638, + 280, + 732 + ], + "type": "inline_equation", + "content": "u_{init}" + }, + { + "bbox": [ + 45, + 638, + 280, + 732 + ], + "type": "text", + "content": " can be derived by the EDL theory as " + }, + { + "bbox": [ + 45, + 638, + 280, + 732 + ], + "type": "inline_equation", + "content": "u_{init} = C / S_{init}" + }, + { + "bbox": [ + 45, + 638, + 280, + 732 + ], + "type": "text", + "content": ". Note that uncertainty " + }, + { + "bbox": [ + 45, + 638, + 280, + 732 + ], + "type": "inline_equation", + "content": "u_{init}" + }, + { + "bbox": [ + 45, + 638, + 280, + 732 + ], + "type": "text", + "content": " is inversely proportional to the total evidence of all closed-set categories, thus it reflects the probability that the video contains unknown actions." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 297, + 378, + 530, + 402 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 378, + 530, + 402 + ], + "spans": [ + { + "bbox": [ + 297, + 378, + 530, + 402 + ], + "type": "text", + "content": "3.3. Knowledge-guided Bipolar Prototype Learning for Evidence Calibration Factors" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 296, + 410, + 530, + 643 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 410, + 530, + 643 + ], + "spans": [ + { + "bbox": [ + 296, + 410, + 530, + 643 + ], + "type": "text", + "content": "Due to the differences in action components, scenes, shooting angles and other aspects, videos tend to keep a large intra-action variation in visual patterns, and the known action instances which visually deviate from the majority are likely to be mistakenly identified as unknown ones. We argue that the initial evidence collected by a single video is insufficient for handling this issue. Therefore, we resort to the bipolar prototype learning strategy, in which both positive and negative prototypes are set for each category, to explore inter-video correlations for mining the intrinsic information of actions and eliminating the interference caused by background snippets. However, even if the intra-action visual variance is mitigated to some extent, it is still challenging to effectively detect the unknown action instances in the open-world environment for the lack of prior knowledge. To tackle the problem, a semantic relation graph is constructed to simulate the open world, so as to provide prior knowledge guidance for the bipolar prototype learning among videos, finally obtaining more accurate evidence calibration factors for further cascade evidence enhancement." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 296, + 648, + 531, + 730 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 648, + 531, + 730 + ], + "spans": [ + { + "bbox": [ + 296, + 648, + 531, + 730 + ], + "type": "text", + "content": "Based on the above observations, as shown in Figure 2, we propose a Knowledge-guided Bipolar Prototype Learning module (KBPL) to obtain evidence calibration factors from inter-video correlations and prior knowledge. Firstly, we construct a semantic relation graph with concepts related to known categories as vertie and their relations as edges. Specifically, the vertie set " + }, + { + "bbox": [ + 296, + 648, + 531, + 730 + ], + "type": "inline_equation", + "content": "\\mathcal{N}" + }, + { + "bbox": [ + 296, + 648, + 531, + 730 + ], + "type": "text", + "content": " includes concepts of known" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 277, + 746, + 301, + 756 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 277, + 746, + 301, + 756 + ], + "spans": [ + { + "bbox": [ + 277, + 746, + 301, + 756 + ], + "type": "text", + "content": "14744" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 595, + 842 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 45, + 106, + 281, + 202 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 106, + 281, + 202 + ], + "spans": [ + { + "bbox": [ + 45, + 106, + 281, + 202 + ], + "type": "text", + "content": "classes, denoted as " + }, + { + "bbox": [ + 45, + 106, + 281, + 202 + ], + "type": "inline_equation", + "content": "\\mathcal{N}_{kno}" + }, + { + "bbox": [ + 45, + 106, + 281, + 202 + ], + "type": "text", + "content": ", and the neighbor action concepts in ConceptNet [39] with the top-" + }, + { + "bbox": [ + 45, + 106, + 281, + 202 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 45, + 106, + 281, + 202 + ], + "type": "text", + "content": " relation strength, denoted as " + }, + { + "bbox": [ + 45, + 106, + 281, + 202 + ], + "type": "inline_equation", + "content": "\\mathcal{N}_{nei}" + }, + { + "bbox": [ + 45, + 106, + 281, + 202 + ], + "type": "text", + "content": ", thus " + }, + { + "bbox": [ + 45, + 106, + 281, + 202 + ], + "type": "inline_equation", + "content": "\\mathcal{N} = \\mathcal{N}_{kno} \\bigcup \\mathcal{N}_{nei}" + }, + { + "bbox": [ + 45, + 106, + 281, + 202 + ], + "type": "text", + "content": ". Following a standard graph construction operation [13, 15], edges are connected in our graph if the relation strength between two concepts (vertex) is larger than a threshold. The features of vertex in " + }, + { + "bbox": [ + 45, + 106, + 281, + 202 + ], + "type": "inline_equation", + "content": "\\mathcal{N}" + }, + { + "bbox": [ + 45, + 106, + 281, + 202 + ], + "type": "text", + "content": " are initialized by the Glove-300 [32] embedding vectors, denoted as " + }, + { + "bbox": [ + 45, + 106, + 281, + 202 + ], + "type": "inline_equation", + "content": "\\{n_j\\}_{j=1}^{|\\mathcal{N}|}" + }, + { + "bbox": [ + 45, + 106, + 281, + 202 + ], + "type": "text", + "content": ", and then updated by:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 211, + 279, + 237 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 211, + 279, + 237 + ], + "spans": [ + { + "bbox": [ + 56, + 211, + 279, + 237 + ], + "type": "interline_equation", + "content": "\\tilde {\\boldsymbol {n}} _ {j} = h _ {2} \\left(\\underset {z \\in \\{1, 2, 3 \\}} {\\operatorname {C o n c a t}} \\left(\\sum_ {l \\in \\mathcal {S} _ {j, z}} \\delta \\left(\\eta_ {j, l}\\right) h _ {1} \\left(\\boldsymbol {n} _ {l}\\right)\\right)\\right) + h _ {3} \\left(\\boldsymbol {n} _ {j}\\right), \\tag {7}", + "image_path": "6f54ad84bc4c3e6e220f603a24c1f90a60489198df00920f668c6aa01bc0af3f.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 45, + 245, + 279, + 314 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 245, + 279, + 314 + ], + "spans": [ + { + "bbox": [ + 45, + 245, + 279, + 314 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 45, + 245, + 279, + 314 + ], + "type": "inline_equation", + "content": "h_{1,2,3}" + }, + { + "bbox": [ + 45, + 245, + 279, + 314 + ], + "type": "text", + "content": " denote fully-connected layers, " + }, + { + "bbox": [ + 45, + 245, + 279, + 314 + ], + "type": "inline_equation", + "content": "\\delta(\\cdot)" + }, + { + "bbox": [ + 45, + 245, + 279, + 314 + ], + "type": "text", + "content": " denotes the Softmax operation, " + }, + { + "bbox": [ + 45, + 245, + 279, + 314 + ], + "type": "inline_equation", + "content": "S_{j,z}" + }, + { + "bbox": [ + 45, + 245, + 279, + 314 + ], + "type": "text", + "content": " denotes the index set of the " + }, + { + "bbox": [ + 45, + 245, + 279, + 314 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 45, + 245, + 279, + 314 + ], + "type": "text", + "content": "-hop neighbor vertex of node " + }, + { + "bbox": [ + 45, + 245, + 279, + 314 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 45, + 245, + 279, + 314 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 45, + 245, + 279, + 314 + ], + "type": "inline_equation", + "content": "\\eta_{j,l}" + }, + { + "bbox": [ + 45, + 245, + 279, + 314 + ], + "type": "text", + "content": " represents the cosine similarity of " + }, + { + "bbox": [ + 45, + 245, + 279, + 314 + ], + "type": "inline_equation", + "content": "\\mathbf{n}_l" + }, + { + "bbox": [ + 45, + 245, + 279, + 314 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 45, + 245, + 279, + 314 + ], + "type": "inline_equation", + "content": "\\mathbf{n}_j" + }, + { + "bbox": [ + 45, + 245, + 279, + 314 + ], + "type": "text", + "content": ". Although many other complicated and advanced graph network methods can be adopted, we select the direct update manner here for simplicity." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 45, + 314, + 279, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 314, + 279, + 443 + ], + "spans": [ + { + "bbox": [ + 45, + 314, + 279, + 443 + ], + "type": "text", + "content": "With the concept relation graph, for category " + }, + { + "bbox": [ + 45, + 314, + 279, + 443 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 45, + 314, + 279, + 443 + ], + "type": "text", + "content": ", we use the updated representation of its corresponding vertex to generate the positive prototype, i.e. " + }, + { + "bbox": [ + 45, + 314, + 279, + 443 + ], + "type": "inline_equation", + "content": "Q_{i}^{pos} = \\phi_{1}(n_{i})" + }, + { + "bbox": [ + 45, + 314, + 279, + 443 + ], + "type": "text", + "content": ", and the average updated representations of its neighbor vertex are adopted for producing the negative one, i.e. " + }, + { + "bbox": [ + 45, + 314, + 279, + 443 + ], + "type": "inline_equation", + "content": "Q_{i}^{neg} = \\phi_{2}(\\sum_{j\\in S_{i}}n_{j} / |S_{i}|)" + }, + { + "bbox": [ + 45, + 314, + 279, + 443 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 45, + 314, + 279, + 443 + ], + "type": "inline_equation", + "content": "\\phi_{1,2}" + }, + { + "bbox": [ + 45, + 314, + 279, + 443 + ], + "type": "text", + "content": " are fully-connected layers for prototype embedding. Then, for video " + }, + { + "bbox": [ + 45, + 314, + 279, + 443 + ], + "type": "inline_equation", + "content": "\\nu_{b}" + }, + { + "bbox": [ + 45, + 314, + 279, + 443 + ], + "type": "text", + "content": " in the minibatch, where " + }, + { + "bbox": [ + 45, + 314, + 279, + 443 + ], + "type": "inline_equation", + "content": "b = 1,\\dots,B" + }, + { + "bbox": [ + 45, + 314, + 279, + 443 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 45, + 314, + 279, + 443 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 45, + 314, + 279, + 443 + ], + "type": "text", + "content": " is the mini-batch size, we define the calibration factor " + }, + { + "bbox": [ + 45, + 314, + 279, + 443 + ], + "type": "inline_equation", + "content": "\\xi_{b,i}" + }, + { + "bbox": [ + 45, + 314, + 279, + 443 + ], + "type": "text", + "content": " by simultaneously considering the distances between the video-level feature " + }, + { + "bbox": [ + 45, + 314, + 279, + 443 + ], + "type": "inline_equation", + "content": "\\tilde{F}_{vid,b}" + }, + { + "bbox": [ + 45, + 314, + 279, + 443 + ], + "type": "text", + "content": " and the bipolar prototypes " + }, + { + "bbox": [ + 45, + 314, + 279, + 443 + ], + "type": "inline_equation", + "content": "Q_{i}^{pos}" + }, + { + "bbox": [ + 45, + 314, + 279, + 443 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 45, + 314, + 279, + 443 + ], + "type": "inline_equation", + "content": "Q_{i}^{neg}" + }, + { + "bbox": [ + 45, + 314, + 279, + 443 + ], + "type": "text", + "content": " as:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 77, + 452, + 279, + 466 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 452, + 279, + 466 + ], + "spans": [ + { + "bbox": [ + 77, + 452, + 279, + 466 + ], + "type": "interline_equation", + "content": "\\xi_ {b, i} = \\hat {D} (\\tilde {\\boldsymbol {F}} _ {v i d, b}, \\boldsymbol {Q} _ {i} ^ {n e g}) - \\hat {D} (\\tilde {\\boldsymbol {F}} _ {v i d, b}, \\boldsymbol {Q} _ {i} ^ {p o s}), \\tag {8}", + "image_path": "7e83f0bdd6c06692da59adba2165dc30417d1542085affe93e947af0742a30f6.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 45, + 474, + 279, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 474, + 279, + 498 + ], + "spans": [ + { + "bbox": [ + 45, + 474, + 279, + 498 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 45, + 474, + 279, + 498 + ], + "type": "inline_equation", + "content": "\\hat{D}" + }, + { + "bbox": [ + 45, + 474, + 279, + 498 + ], + "type": "text", + "content": " is a distance calculation method, which can be Euclidean distance or negative cosine similarity." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 45, + 498, + 279, + 580 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 498, + 279, + 580 + ], + "spans": [ + { + "bbox": [ + 45, + 498, + 279, + 580 + ], + "type": "text", + "content": "To guide the learning of prototypes by inter-video correlations, for each category, we pull close the videos belonging to the category to the corresponding positive prototype, and push away the videos of other classes in the same minibatch, while opposite operations are performed for the negative prototype. Specifically, the loss function " + }, + { + "bbox": [ + 45, + 498, + 279, + 580 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{kbpl}" + }, + { + "bbox": [ + 45, + 498, + 279, + 580 + ], + "type": "text", + "content": " guiding the pos-neg prototype learning is designed as:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 57, + 588, + 279, + 628 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 588, + 279, + 628 + ], + "spans": [ + { + "bbox": [ + 57, + 588, + 279, + 628 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {k b p l} = - \\frac {1}{C} \\sum_ {i = 1} ^ {C} \\mathbb {I} \\left(\\mathcal {B} _ {i} \\neq \\emptyset\\right) \\log \\left(\\frac {\\sum_ {b \\in \\mathcal {B} _ {i}} \\exp \\left(\\xi_ {b , i}\\right)}{\\sum_ {b = 1} ^ {B} \\exp \\left(\\xi_ {b , i}\\right)}\\right), \\tag {9}", + "image_path": "787424210558a5f5756b8e9479c407e1b0cc182adb353442e8462cabc13dfb21.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 45, + 628, + 279, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 628, + 279, + 652 + ], + "spans": [ + { + "bbox": [ + 45, + 628, + 279, + 652 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 45, + 628, + 279, + 652 + ], + "type": "inline_equation", + "content": "\\mathbb{I}(\\cdot)" + }, + { + "bbox": [ + 45, + 628, + 279, + 652 + ], + "type": "text", + "content": " is an indicator function, " + }, + { + "bbox": [ + 45, + 628, + 279, + 652 + ], + "type": "inline_equation", + "content": "\\mathcal{B}_i = \\{b|y_{b,i} = 1\\}" + }, + { + "bbox": [ + 45, + 628, + 279, + 652 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 45, + 628, + 279, + 652 + ], + "type": "inline_equation", + "content": "\\pmb{y}_b" + }, + { + "bbox": [ + 45, + 628, + 279, + 652 + ], + "type": "text", + "content": " is the ground-truth label of the video " + }, + { + "bbox": [ + 45, + 628, + 279, + 652 + ], + "type": "inline_equation", + "content": "\\mathcal{V}_b" + }, + { + "bbox": [ + 45, + 628, + 279, + 652 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 45, + 656, + 216, + 667 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 656, + 216, + 667 + ], + "spans": [ + { + "bbox": [ + 45, + 656, + 216, + 667 + ], + "type": "text", + "content": "3.4. Cascade Evidence Enhancement" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 45, + 671, + 279, + 730 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 671, + 279, + 730 + ], + "spans": [ + { + "bbox": [ + 45, + 671, + 279, + 730 + ], + "type": "text", + "content": "As shown in Figure 2, we then perform cascade evidence enhancement, which scales the initial evidence collected from the MEP module by the learned calibration factors. As we stated above, the uncertainty " + }, + { + "bbox": [ + 45, + 671, + 279, + 730 + ], + "type": "inline_equation", + "content": "u_{init}" + }, + { + "bbox": [ + 45, + 671, + 279, + 730 + ], + "type": "text", + "content": " represents the probability that a video contains unknown actions. For example," + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 296, + 106, + 530, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 106, + 530, + 186 + ], + "spans": [ + { + "bbox": [ + 296, + 106, + 530, + 186 + ], + "type": "text", + "content": "given a test video, when " + }, + { + "bbox": [ + 296, + 106, + 530, + 186 + ], + "type": "inline_equation", + "content": "u_{init}" + }, + { + "bbox": [ + 296, + 106, + 530, + 186 + ], + "type": "text", + "content": " is close to the extreme values (close to zero for known actions and close to the maximum for unknown ones), we can recognize known or unknown classes with high confidence [35], otherwise the model has relatively low confidence in the known/unknown judgment. Specifically, we design an uncertainty-aware adaptor " + }, + { + "bbox": [ + 296, + 106, + 530, + 186 + ], + "type": "inline_equation", + "content": "\\zeta" + }, + { + "bbox": [ + 296, + 106, + 530, + 186 + ], + "type": "text", + "content": " as a Gaussian-like estimation function for evidence cascade:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 335, + 196, + 530, + 221 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 196, + 530, + 221 + ], + "spans": [ + { + "bbox": [ + 335, + 196, + 530, + 221 + ], + "type": "interline_equation", + "content": "\\zeta \\left(u _ {i n i t}; \\mu , \\sigma\\right) = \\exp \\left(- \\frac {\\left(u _ {i n i t} - \\mu\\right) ^ {2}}{2 \\sigma^ {2}}\\right), \\tag {10}", + "image_path": "af1f3c205db1c4fd1b4ab72782ccad01c90e2be80662ad4deb506b1d5268e594.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 296, + 230, + 530, + 265 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 230, + 530, + 265 + ], + "spans": [ + { + "bbox": [ + 296, + 230, + 530, + 265 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 296, + 230, + 530, + 265 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 296, + 230, + 530, + 265 + ], + "type": "text", + "content": " is set to the value beyond the " + }, + { + "bbox": [ + 296, + 230, + 530, + 265 + ], + "type": "inline_equation", + "content": "u_{init}" + }, + { + "bbox": [ + 296, + 230, + 530, + 265 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 296, + 230, + 530, + 265 + ], + "type": "inline_equation", + "content": "95\\%" + }, + { + "bbox": [ + 296, + 230, + 530, + 265 + ], + "type": "text", + "content": " training videos, and " + }, + { + "bbox": [ + 296, + 230, + 530, + 265 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 296, + 230, + 530, + 265 + ], + "type": "text", + "content": " is a learnable scalar parameter which controls the shape of the Gaussian-like function." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 296, + 266, + 530, + 347 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 266, + 530, + 347 + ], + "spans": [ + { + "bbox": [ + 296, + 266, + 530, + 347 + ], + "type": "text", + "content": "If the above uncertainty-aware adaptor " + }, + { + "bbox": [ + 296, + 266, + 530, + 347 + ], + "type": "inline_equation", + "content": "\\zeta" + }, + { + "bbox": [ + 296, + 266, + 530, + 347 + ], + "type": "text", + "content": " of an input video is high, it means it is difficult for the model to perform known/unknown judgment by only considering the initial evidence collected inside the video, thus it is necessary to employ the information provided by the inter-video correlations to carry out evidence calibration. Specifically, the cascade evidence calibration process can be formulated as:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 335, + 360, + 530, + 371 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 360, + 530, + 371 + ], + "spans": [ + { + "bbox": [ + 335, + 360, + 530, + 371 + ], + "type": "interline_equation", + "content": "e _ {c a s c, i} = (1 + \\zeta (u _ {i n i t}) \\tanh (\\xi_ {i})) e _ {i n i t, i}, \\tag {11}", + "image_path": "82ff9ff56064642ede826672d2f55b60d0fa2f6b47497d0648b9bd7847536e62.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 296, + 382, + 530, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 382, + 530, + 475 + ], + "spans": [ + { + "bbox": [ + 296, + 382, + 530, + 475 + ], + "type": "text", + "content": "where the " + }, + { + "bbox": [ + 296, + 382, + 530, + 475 + ], + "type": "inline_equation", + "content": "\\tanh (\\cdot)" + }, + { + "bbox": [ + 296, + 382, + 530, + 475 + ], + "type": "text", + "content": " function ensures the scaling range of " + }, + { + "bbox": [ + 296, + 382, + 530, + 475 + ], + "type": "inline_equation", + "content": "e_{init,i}" + }, + { + "bbox": [ + 296, + 382, + 530, + 475 + ], + "type": "text", + "content": " to be (0,2) for a stable learning, and " + }, + { + "bbox": [ + 296, + 382, + 530, + 475 + ], + "type": "inline_equation", + "content": "e_{casc} = [e_{casc,1},\\dots,e_{casc,C}] \\in \\mathbb{R}^C" + }, + { + "bbox": [ + 296, + 382, + 530, + 475 + ], + "type": "text", + "content": " is the enhanced cascaded evidence. The subscript " + }, + { + "bbox": [ + 296, + 382, + 530, + 475 + ], + "type": "inline_equation", + "content": "b" + }, + { + "bbox": [ + 296, + 382, + 530, + 475 + ], + "type": "text", + "content": " (video index) is omitted in this equation for simplicity. " + }, + { + "bbox": [ + 296, + 382, + 530, + 475 + ], + "type": "inline_equation", + "content": "e_{casc}" + }, + { + "bbox": [ + 296, + 382, + 530, + 475 + ], + "type": "text", + "content": " is then optimized by an EDL loss " + }, + { + "bbox": [ + 296, + 382, + 530, + 475 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{casc}" + }, + { + "bbox": [ + 296, + 382, + 530, + 475 + ], + "type": "text", + "content": ", whose form is identical with Eq. (6), and the cascaded video-level uncertainty " + }, + { + "bbox": [ + 296, + 382, + 530, + 475 + ], + "type": "inline_equation", + "content": "u_{casc}" + }, + { + "bbox": [ + 296, + 382, + 530, + 475 + ], + "type": "text", + "content": " for known/unknown judgment can be obtained in the same way with " + }, + { + "bbox": [ + 296, + 382, + 530, + 475 + ], + "type": "inline_equation", + "content": "u_{init}" + }, + { + "bbox": [ + 296, + 382, + 530, + 475 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 297, + 482, + 424, + 494 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 482, + 424, + 494 + ], + "spans": [ + { + "bbox": [ + 297, + 482, + 424, + 494 + ], + "type": "text", + "content": "3.5. Training and Inference" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 297, + 498, + 529, + 520 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 498, + 529, + 520 + ], + "spans": [ + { + "bbox": [ + 297, + 498, + 529, + 520 + ], + "type": "text", + "content": "Training. By combining all the optimization objectives introduced above, we obtain the final loss function as:" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 345, + 533, + 530, + 544 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 345, + 533, + 530, + 544 + ], + "spans": [ + { + "bbox": [ + 345, + 533, + 530, + 544 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\mathcal {L} _ {c l s} + \\mathcal {L} _ {i n i t} + \\mathcal {L} _ {k b p l} + \\mathcal {L} _ {c a s c}, \\tag {12}", + "image_path": "40e9e48ac09178041ca3f477b4fe4e1f1de4424f10dc6e96f2c7b17df40edc74.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 297, + 555, + 529, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 555, + 529, + 578 + ], + "spans": [ + { + "bbox": [ + 297, + 555, + 529, + 578 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 297, + 555, + 529, + 578 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{cls}" + }, + { + "bbox": [ + 297, + 555, + 529, + 578 + ], + "type": "text", + "content": " is the standard video-level classification loss, e.g., MIL loss [10, 19], for optimizing the backbone." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 297, + 579, + 530, + 730 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 579, + 530, + 730 + ], + "spans": [ + { + "bbox": [ + 297, + 579, + 530, + 730 + ], + "type": "text", + "content": "Inference. For each video, we predict its action categories set " + }, + { + "bbox": [ + 297, + 579, + 530, + 730 + ], + "type": "inline_equation", + "content": "\\Psi" + }, + { + "bbox": [ + 297, + 579, + 530, + 730 + ], + "type": "text", + "content": " (including the Unknown class if exist) by Algorithm 1, where the closed-set video-level classification score " + }, + { + "bbox": [ + 297, + 579, + 530, + 730 + ], + "type": "inline_equation", + "content": "P = [P_{1},\\dots,P_{C}]" + }, + { + "bbox": [ + 297, + 579, + 530, + 730 + ], + "type": "text", + "content": " is obtained by averaging the classification activation sequence (CAS) of snippets in " + }, + { + "bbox": [ + 297, + 579, + 530, + 730 + ], + "type": "inline_equation", + "content": "\\Theta" + }, + { + "bbox": [ + 297, + 579, + 530, + 730 + ], + "type": "text", + "content": ". Then we use a threshold strategy to obtain action snippet candidates following the standard process [18, 19]. Finally, for each action class in " + }, + { + "bbox": [ + 297, + 579, + 530, + 730 + ], + "type": "inline_equation", + "content": "\\Psi" + }, + { + "bbox": [ + 297, + 579, + 530, + 730 + ], + "type": "text", + "content": " we group continuous snippets into action proposals, whose confidence scores " + }, + { + "bbox": [ + 297, + 579, + 530, + 730 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 297, + 579, + 530, + 730 + ], + "type": "text", + "content": " are estimated according to the CAS scores of the corresponding class (for Known classes) or attention scores (for the Unknown class), and perform non-maximum-suppression (NMS) to remove duplicated proposals." + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 277, + 746, + 300, + 756 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 277, + 746, + 300, + 756 + ], + "spans": [ + { + "bbox": [ + 277, + 746, + 300, + 756 + ], + "type": "text", + "content": "14745" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 595, + 842 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 48, + 127, + 537, + 322 + ], + "blocks": [ + { + "bbox": [ + 46, + 104, + 530, + 126 + ], + "lines": [ + { + "bbox": [ + 46, + 104, + 530, + 126 + ], + "spans": [ + { + "bbox": [ + 46, + 104, + 530, + 126 + ], + "type": "text", + "content": "Table 1. OWTAL localization results on THUMOS-14 evaluated by both top- " + }, + { + "bbox": [ + 46, + 104, + 530, + 126 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 46, + 104, + 530, + 126 + ], + "type": "text", + "content": " mAP@Avg and traditional mAP@Avg. The averages of the mAP of Known classes and the AP of the Unknown class are reported. The values are averaged on t-IoU [0.1:0.1:0.5] and [0.1:0.1:0.7]." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 48, + 127, + 537, + 322 + ], + "lines": [ + { + "bbox": [ + 48, + 127, + 537, + 322 + ], + "spans": [ + { + "bbox": [ + 48, + 127, + 537, + 322 + ], + "type": "table", + "html": "
MethodsTop-K mAP@Avg(%)mAP@Avg(%)
Top-5Top-10Top-20Top-50Top-100
0.1-0.50.1-0.70.1-0.50.1-0.70.1-0.50.1-0.70.1-0.50.1-0.70.1-0.50.1-0.70.1-0.50.1-0.7
ASM-Loc + Trivial4.33.57.86.313.310.721.116.926.120.929.523.6
ASM-Loc + SoftMax10.98.916.913.523.418.728.823.029.823.830.124.0
ASM-Loc + OpenMax10.18.115.312.121.216.726.120.627.321.527.621.7
ASM-Loc + ARPL10.08.216.413.423.619.330.424.832.026.032.426.4
ASM-Loc + EDL11.39.217.414.024.219.430.524.331.925.432.225.7
ASM-Loc + CELL(Ours)11.29.318.014.625.520.732.426.434.127.834.728.1
CO2-Net + Trivial5.54.49.57.716.713.525.920.930.925.034.427.9
CO2-Net + SoftMax11.39.117.814.325.120.232.226.033.727.334.227.8
CO2-Net + OpenMax10.38.416.313.223.018.629.123.530.424.730.825.0
CO2-Net + ARPL11.69.518.314.925.720.933.327.135.128.735.729.2
CO2-Net + EDL11.29.117.614.324.820.032.226.034.027.534.628.1
CO2-Net + CELL(Ours)12.610.320.116.428.123.036.930.338.931.839.532.3
", + "image_path": "14954c6395ddb2b6309aaa03faf3929d6b45319701eacd5559d5039e8d3b1488.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 54, + 366, + 523, + 497 + ], + "blocks": [ + { + "bbox": [ + 46, + 332, + 531, + 364 + ], + "lines": [ + { + "bbox": [ + 46, + 332, + 531, + 364 + ], + "spans": [ + { + "bbox": [ + 46, + 332, + 531, + 364 + ], + "type": "text", + "content": "Table 2. OWTAL localization results on THUMOS-14 evaluated by both top- " + }, + { + "bbox": [ + 46, + 332, + 531, + 364 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 46, + 332, + 531, + 364 + ], + "type": "text", + "content": " mAP@Avg and traditional mAP@Avg. The mAP of Known classes (K) and the AP of the Unknown class (U) averaged on t-IoU thresholds [0.1:0.1:0.7] are reported respectively. The video-level known-unknown classification accuracy (v-Acc(%) ) is also presented for reference." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 54, + 366, + 523, + 497 + ], + "lines": [ + { + "bbox": [ + 54, + 366, + 523, + 497 + ], + "spans": [ + { + "bbox": [ + 54, + 366, + 523, + 497 + ], + "type": "table", + "html": "
MethodsTop-K mAP@Avg(%)mAP@Avg(%)v-Acc(%)
Top-5Top-10Top-20Top-50Top-100
KUKUKUKUKUKU
CO2-Net + Trivial4.44.47.87.615.111.923.818.030.519.436.219.628.1
CO2-Net + SoftMax12.65.619.39.326.413.932.719.234.520.035.620.071.2
CO2-Net + OpenMax11.94.918.48.025.112.031.016.132.716.733.616.570.5
CO2-Net + ARPL12.36.719.410.426.615.232.621.734.622.735.822.673.0
CO2-Net + EDL12.45.719.19.426.313.832.719.434.820.235.920.370.9
CO2-Net + CELL(Ours)13.37.320.512.227.618.334.026.635.927.636.827.874.9
", + "image_path": "507031e04b7331bfda72caedcab9cf2607faac1bea16683ce55a54ba3fb8fd3d.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 45, + 505, + 167, + 518 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 505, + 167, + 518 + ], + "spans": [ + { + "bbox": [ + 45, + 505, + 167, + 518 + ], + "type": "text", + "content": "4. Experimental Results" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 45, + 523, + 158, + 536 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 523, + 158, + 536 + ], + "spans": [ + { + "bbox": [ + 45, + 523, + 158, + 536 + ], + "type": "text", + "content": "4.1. Experimental Setup" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 45, + 543, + 280, + 730 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 543, + 280, + 730 + ], + "spans": [ + { + "bbox": [ + 45, + 543, + 280, + 730 + ], + "type": "text", + "content": "Datasets. We conduct experiments on two popular benchmarks for OWTAL, THUMOS-14 and ActivityNet-v1.3. THUMOS-14 contains 200 validation videos and 213 test videos annotated from 20 action categories, and each video contains 15.4 action instances on average. ActivityNet-v1.3 contains 10,024 training videos and 4,926 validation videos from 200 action categories, and each video contains 1.6 action instances on average. Following the pioneer work [3], we randomly remove 1/4 action categories of THUMOS-14 training set and keep the entire THUMOS-14 testing set to enable open-set evaluation, and such random removal is repeated three times. To further increase the openness and verify the applicability of our proposed method for open-world scenarios, we train our model on the known split of THUMOS-14 and test on the testing set of ActivityNet-v1.3. Note that since ActivityNet-v1.3 covers" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 296, + 506, + 530, + 541 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 506, + 530, + 541 + ], + "spans": [ + { + "bbox": [ + 296, + 506, + 530, + 541 + ], + "type": "text", + "content": "most categories in THUMOS-14, we manually remove 14 action categories which are semantically overlapping with the THUMOS-14 as previous work did." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 296, + 543, + 531, + 730 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 543, + 531, + 730 + ], + "spans": [ + { + "bbox": [ + 296, + 543, + 531, + 730 + ], + "type": "text", + "content": "Evaluation Metrics. We adopt three sets of metrics to measure the performance of the OWTAL task. (1) Vanilla mAP. To evaluate the localization performance, following previous approaches [10, 19, 33], we employ the mean Average Precision (mAP) under different temporal Intersection over Union (t-IoU) thresholds as metrics. The t-IoU thresholds for THUMOS-14 is [0.1:0.1:0.7] and for ActivityNetv1.3 is [0.5:0.05:0.95]. (2) Top-" + }, + { + "bbox": [ + 296, + 543, + 531, + 730 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 296, + 543, + 531, + 730 + ], + "type": "text", + "content": " mAP. In our experiments, we observe that it only brings slight performance loss on the traditional mAP for models to generate a large number of redundant proposals. However, in real-world applications our goal is to provide users with accurate action proposals of a limited number, rather than an extreme largescale proposal set consisting of numerous redundant proposals, although the set may contain even more accurate predictions. Inspired by the above observation, we design a" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 277, + 746, + 301, + 756 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 277, + 746, + 301, + 756 + ], + "spans": [ + { + "bbox": [ + 277, + 746, + 301, + 756 + ], + "type": "text", + "content": "14746" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 595, + 842 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 106, + 252, + 117 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 106, + 252, + 117 + ], + "spans": [ + { + "bbox": [ + 47, + 106, + 252, + 117 + ], + "type": "text", + "content": "Algorithm 1 Action Categories Inference Procedure" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 119, + 186, + 130 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 119, + 186, + 130 + ], + "spans": [ + { + "bbox": [ + 47, + 119, + 186, + 130 + ], + "type": "text", + "content": "Input: Untrimmed testing video " + }, + { + "bbox": [ + 47, + 119, + 186, + 130 + ], + "type": "inline_equation", + "content": "\\nu" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 131, + 173, + 142 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 131, + 173, + 142 + ], + "spans": [ + { + "bbox": [ + 47, + 131, + 173, + 142 + ], + "type": "text", + "content": "Require: Trained CELL model." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 143, + 279, + 165 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 143, + 279, + 165 + ], + "spans": [ + { + "bbox": [ + 47, + 143, + 279, + 165 + ], + "type": "text", + "content": "Require: Threshold " + }, + { + "bbox": [ + 47, + 143, + 279, + 165 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 47, + 143, + 279, + 165 + ], + "type": "text", + "content": " obtained from training data (Please refer to Implementation Details)." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 165, + 206, + 176 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 165, + 206, + 176 + ], + "spans": [ + { + "bbox": [ + 47, + 165, + 206, + 176 + ], + "type": "text", + "content": "Output: Set " + }, + { + "bbox": [ + 47, + 165, + 206, + 176 + ], + "type": "inline_equation", + "content": "\\Psi" + }, + { + "bbox": [ + 47, + 165, + 206, + 176 + ], + "type": "text", + "content": " of action categories in " + }, + { + "bbox": [ + 47, + 165, + 206, + 176 + ], + "type": "inline_equation", + "content": "\\mathcal{V}" + }, + { + "bbox": [ + 47, + 165, + 206, + 176 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 179, + 279, + 340 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 52, + 179, + 279, + 202 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 52, + 179, + 279, + 202 + ], + "spans": [ + { + "bbox": [ + 52, + 179, + 279, + 202 + ], + "type": "text", + "content": "1: Predict the closed-set classification score " + }, + { + "bbox": [ + 52, + 179, + 279, + 202 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 52, + 179, + 279, + 202 + ], + "type": "text", + "content": " and cascaded video-level uncertainty " + }, + { + "bbox": [ + 52, + 179, + 279, + 202 + ], + "type": "inline_equation", + "content": "u_{\\text{casc}}" + }, + { + "bbox": [ + 52, + 179, + 279, + 202 + ], + "type": "text", + "content": " by CELL." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 52, + 203, + 133, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 52, + 203, + 133, + 213 + ], + "spans": [ + { + "bbox": [ + 52, + 203, + 133, + 213 + ], + "type": "text", + "content": "2: if " + }, + { + "bbox": [ + 52, + 203, + 133, + 213 + ], + "type": "inline_equation", + "content": "u_{\\text{casc}} < \\tau" + }, + { + "bbox": [ + 52, + 203, + 133, + 213 + ], + "type": "text", + "content": " then" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 52, + 214, + 278, + 226 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 52, + 214, + 278, + 226 + ], + "spans": [ + { + "bbox": [ + 52, + 214, + 278, + 226 + ], + "type": "text", + "content": "3: " + }, + { + "bbox": [ + 52, + 214, + 278, + 226 + ], + "type": "inline_equation", + "content": "\\Psi = \\{i|P_i > 0.2\\}" + }, + { + "bbox": [ + 52, + 214, + 278, + 226 + ], + "type": "text", + "content": " ▷OnlyKnownClasses" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 52, + 227, + 137, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 52, + 227, + 137, + 236 + ], + "spans": [ + { + "bbox": [ + 52, + 227, + 137, + 236 + ], + "type": "text", + "content": "4: if " + }, + { + "bbox": [ + 52, + 227, + 137, + 236 + ], + "type": "inline_equation", + "content": "\\Psi = \\emptyset" + }, + { + "bbox": [ + 52, + 227, + 137, + 236 + ], + "type": "text", + "content": " then" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 52, + 238, + 278, + 248 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 52, + 238, + 278, + 248 + ], + "spans": [ + { + "bbox": [ + 52, + 238, + 278, + 248 + ], + "type": "text", + "content": "5: " + }, + { + "bbox": [ + 52, + 238, + 278, + 248 + ], + "type": "inline_equation", + "content": "\\Psi = \\arg \\max_{i} P_{i}" + }, + { + "bbox": [ + 52, + 238, + 278, + 248 + ], + "type": "text", + "content": " ▷OnlyKnownClasses" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 52, + 250, + 104, + 258 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 52, + 250, + 104, + 258 + ], + "spans": [ + { + "bbox": [ + 52, + 250, + 104, + 258 + ], + "type": "text", + "content": "6: end if" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 52, + 261, + 185, + 271 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 52, + 261, + 185, + 271 + ], + "spans": [ + { + "bbox": [ + 52, + 261, + 185, + 271 + ], + "type": "text", + "content": "7: else if " + }, + { + "bbox": [ + 52, + 261, + 185, + 271 + ], + "type": "inline_equation", + "content": "\\arg \\max_{i} P_{i} > 0.5" + }, + { + "bbox": [ + 52, + 261, + 185, + 271 + ], + "type": "text", + "content": " then" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 52, + 272, + 185, + 283 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 52, + 272, + 185, + 283 + ], + "spans": [ + { + "bbox": [ + 52, + 272, + 185, + 283 + ], + "type": "text", + "content": "8: " + }, + { + "bbox": [ + 52, + 272, + 185, + 283 + ], + "type": "inline_equation", + "content": "\\Psi = \\{\\arg \\max_{i} P_{i}, C + 1\\}" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 52, + 284, + 278, + 294 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 52, + 284, + 278, + 294 + ], + "spans": [ + { + "bbox": [ + 52, + 284, + 278, + 294 + ], + "type": "text", + "content": "9: Both Known and Unknown Classes" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 296, + 80, + 304 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 296, + 80, + 304 + ], + "spans": [ + { + "bbox": [ + 48, + 296, + 80, + 304 + ], + "type": "text", + "content": "10: else" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 307, + 278, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 307, + 278, + 319 + ], + "spans": [ + { + "bbox": [ + 48, + 307, + 278, + 319 + ], + "type": "text", + "content": "11: " + }, + { + "bbox": [ + 48, + 307, + 278, + 319 + ], + "type": "inline_equation", + "content": "\\Psi = \\{C + 1\\}" + }, + { + "bbox": [ + 48, + 307, + 278, + 319 + ], + "type": "text", + "content": " ▷OnlyUnknownClasses" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 48, + 319, + 89, + 329 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 319, + 89, + 329 + ], + "spans": [ + { + "bbox": [ + 48, + 319, + 89, + 329 + ], + "type": "text", + "content": "12: end if" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 48, + 330, + 101, + 340 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 330, + 101, + 340 + ], + "spans": [ + { + "bbox": [ + 48, + 330, + 101, + 340 + ], + "type": "text", + "content": "13: return " + }, + { + "bbox": [ + 48, + 330, + 101, + 340 + ], + "type": "inline_equation", + "content": "\\Psi" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 45, + 367, + 279, + 599 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 367, + 279, + 599 + ], + "spans": [ + { + "bbox": [ + 45, + 367, + 279, + 599 + ], + "type": "text", + "content": "variant metric of the vanilla mAP to further meet the needs of real-world applications, termed as top-" + }, + { + "bbox": [ + 45, + 367, + 279, + 599 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 45, + 367, + 279, + 599 + ], + "type": "text", + "content": " mAP, which restricts the maximum number of generated proposals by selecting proposals with top-" + }, + { + "bbox": [ + 45, + 367, + 279, + 599 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 45, + 367, + 279, + 599 + ], + "type": "text", + "content": " confidence scores. When " + }, + { + "bbox": [ + 45, + 367, + 279, + 599 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 45, + 367, + 279, + 599 + ], + "type": "text", + "content": " is not restricted, the top-" + }, + { + "bbox": [ + 45, + 367, + 279, + 599 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 45, + 367, + 279, + 599 + ], + "type": "text", + "content": " mAP degrades to the traditional mAP metric. In our experiments " + }, + { + "bbox": [ + 45, + 367, + 279, + 599 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 45, + 367, + 279, + 599 + ], + "type": "text", + "content": " is set to 5, 10, 20, 50, 100 for THUMOS-14 and 1, 3, 5 for ActivityNet-v1.3, respectively. (3) Classification metrics. For known/unknown classification, following [3], four classification metrics including the False Alarm Rate at True Positive Rate of " + }, + { + "bbox": [ + 45, + 367, + 279, + 599 + ], + "type": "inline_equation", + "content": "95\\%" + }, + { + "bbox": [ + 45, + 367, + 279, + 599 + ], + "type": "text", + "content": " (FAR@95) (smaller value indicates better performance), the Area Under the Receiver Operating Characteristic (AUROC) curve, the Area Under the Precision-Recall (AUPR), and the Open Set Detection Rate (OSDR) are also employed. To validate the effectiveness of Algorithm 1, we additionally adopt the video-level known-unknown classification accuracy (v-Acc(\\%)) metric for evaluation. Note that a testing video may belong to three types which have: (i) only known actions, (ii) only unknown actions, or (iii) both known and unknown actions." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 45, + 602, + 279, + 730 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 602, + 279, + 730 + ], + "spans": [ + { + "bbox": [ + 45, + 602, + 279, + 730 + ], + "type": "text", + "content": "Implementation Details. We adopt CO2-Net [19] and ASM-Loc [18] as our backbone networks. Following existing methods, we use I3D [5] model pretrained on the Kinetics [23] dataset to extract both the RGB and optical flow features. The feature dimension " + }, + { + "bbox": [ + 45, + 602, + 279, + 730 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 45, + 602, + 279, + 730 + ], + "type": "text", + "content": " is 2048, the evidence scale function " + }, + { + "bbox": [ + 45, + 602, + 279, + 730 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 45, + 602, + 279, + 730 + ], + "type": "text", + "content": " in Eq. (6) is Exp, and the distance calculation method " + }, + { + "bbox": [ + 45, + 602, + 279, + 730 + ], + "type": "inline_equation", + "content": "\\hat{D}" + }, + { + "bbox": [ + 45, + 602, + 279, + 730 + ], + "type": "text", + "content": " in Eq. (8) is negative cosine similarity. The number of the sampled snippets " + }, + { + "bbox": [ + 45, + 602, + 279, + 730 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 45, + 602, + 279, + 730 + ], + "type": "text", + "content": " for THUMOS-14 and ActivityNet-v1.3 is 500 and 160, and the scaling factor " + }, + { + "bbox": [ + 45, + 602, + 279, + 730 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 45, + 602, + 279, + 730 + ], + "type": "text", + "content": " is 7. " + }, + { + "bbox": [ + 45, + 602, + 279, + 730 + ], + "type": "inline_equation", + "content": "\\xi" + }, + { + "bbox": [ + 45, + 602, + 279, + 730 + ], + "type": "text", + "content": " in Eq. (10) is updated for every 25 epochs. The batch size is set to 10 and 16, and the learning rate is 5e-5 and 1e" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 296, + 106, + 530, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 106, + 530, + 175 + ], + "spans": [ + { + "bbox": [ + 296, + 106, + 530, + 175 + ], + "type": "text", + "content": "4 in CO2-Net and ASM-Loc, respectively. Following previous works [2, 40], we determine the threshold " + }, + { + "bbox": [ + 296, + 106, + 530, + 175 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 296, + 106, + 530, + 175 + ], + "type": "text", + "content": " of uncertainty " + }, + { + "bbox": [ + 296, + 106, + 530, + 175 + ], + "type": "inline_equation", + "content": "u_{\\text{casc}}" + }, + { + "bbox": [ + 296, + 106, + 530, + 175 + ], + "type": "text", + "content": " in Algorithm 1 by ensuring " + }, + { + "bbox": [ + 296, + 106, + 530, + 175 + ], + "type": "inline_equation", + "content": "95\\%" + }, + { + "bbox": [ + 296, + 106, + 530, + 175 + ], + "type": "text", + "content": " training videos to be recognized as known. Our model is implemented with Python 3.7 and PyTorch 1.11.0. All experiments are conducted on a single RTX3090 GPU." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 296, + 181, + 517, + 192 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 181, + 517, + 192 + ], + "spans": [ + { + "bbox": [ + 296, + 181, + 517, + 192 + ], + "type": "text", + "content": "4.2. Comparison with State-of-the-art Methods" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 296, + 196, + 530, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 196, + 530, + 323 + ], + "spans": [ + { + "bbox": [ + 296, + 196, + 530, + 323 + ], + "type": "text", + "content": "Our proposed CELL method is compared with the following baselines based on the ASM-Loc [18] and the CO2-Net [19]: (1) Trivial: assuming that all testing videos contain both known and unknown actions without any known/unknown judgment. (2) SoftMax: utilizing the maximum softmax probability to identify the unknown. (3) OpenMax: appending the unknown score to the softmax scores by OpenMax [4] in testing. (4) ARPL: using the Adversarial Reciprocal Points Learning (ARPL) [7] to identify the unknown. (5) EDL: vanilla EDL is used to replace the softmax classification head for uncertainty estimation." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 296, + 324, + 530, + 532 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 324, + 530, + 532 + ], + "spans": [ + { + "bbox": [ + 296, + 324, + 530, + 532 + ], + "type": "text", + "content": "Comparison results on THUMOS-14. Table 1 and Table 2 represent the OWTAL localization results on the THUMOS-14 dataset. The results show that our proposed method consistently outperforms the baselines by large margins on all metrics. We notice that totally without any known/unknown judgment, the Trivial method shows comparable performance with the SoftMax baseline on the traditional mAP metric, for the embarrassing property that it only brings slight performance loss on the traditional mAP to generate numerous error redundant proposals, while on our proposed top-" + }, + { + "bbox": [ + 296, + 324, + 530, + 532 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 296, + 324, + 530, + 532 + ], + "type": "text", + "content": " mAP metrics the performance of the Trivial version shows a reasonable gap. The results also show that Openmax does not work well on the OWTAL task, which may lie in that Openmax often fails to recognize visually indistinguishable samples [16], especially in weakly-supervised video analysis field. Note that the other two SOTA approaches ARPL and EDL perform well but still far behind the proposed CELL." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 296, + 532, + 530, + 730 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 532, + 530, + 730 + ], + "spans": [ + { + "bbox": [ + 296, + 532, + 530, + 730 + ], + "type": "text", + "content": "Table 3 displays the known/unknown classification results evaluated on the metrics adopted by [3], showing that our method outperforms the EDL baseline and even achieves comparable results with fully-supervised methods on FAR@95, AUROC, and AUPR. We infer that the large gap between weakly-supervised methods and fully-supervised ones on OSDR is due to the difference between backbone networks. OSDR is defined as the area under the curve of Correct Detection Rate (CDR) and False Positive Rate (FPR), and the CDR indicates the fraction of known actions which are positively localized and correctly classified, while the FPR denotes the fraction of unknown actions that are positively localized but falsely classified. CO2-Net adopts the localization-by-classification strategy commonly used in WTAL method, thus the positively localized actions are usually accompanied by correct class predictions, while AFSD [26] adopts an anchor-free method whose localize" + } + ] + } + ], + "index": 25 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 277, + 746, + 300, + 756 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 277, + 746, + 300, + 756 + ], + "spans": [ + { + "bbox": [ + 277, + 746, + 300, + 756 + ], + "type": "text", + "content": "14747" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 595, + 842 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 45, + 106, + 280, + 165 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 106, + 280, + 165 + ], + "spans": [ + { + "bbox": [ + 45, + 106, + 280, + 165 + ], + "type": "text", + "content": "tion performance does not entirely rely on the classification results, thus resulting in the large gap on OSDR. Note that compared with our employed mAP-related metrics, OSDR is still a classification metric thus cannot fully exhibit the superiority of different models." + } + ] + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 48, + 216, + 279, + 301 + ], + "blocks": [ + { + "bbox": [ + 45, + 179, + 279, + 212 + ], + "lines": [ + { + "bbox": [ + 45, + 179, + 279, + 212 + ], + "spans": [ + { + "bbox": [ + 45, + 179, + 279, + 212 + ], + "type": "text", + "content": "Table 3. OWTAL classification results on THUMOS-14 evaluated by FAR@95, AUROC, AUPR and OSDR. In this experiment we use CO2-Net as our backbone." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 48, + 216, + 279, + 301 + ], + "lines": [ + { + "bbox": [ + 48, + 216, + 279, + 301 + ], + "spans": [ + { + "bbox": [ + 48, + 216, + 279, + 301 + ], + "type": "table", + "html": "
SupervisionMethodsFAR@95(↓)AUROCAUPROSDR
FullySoftMax85.5854.7031.8523.40
OpenMax90.3453.2633.1713.66
EDL81.4264.0540.0536.26
OpenTAL70.9678.3358.6242.91
WeaklyCO2+EDL87.3767.1144.4763.56
CELL(Ours)68.8674.6254.2569.81
", + "image_path": "51d7830257195f0f41c4cf81500fe33bf728ffaaaeddebf2f814b5d556010148.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 48, + 360, + 279, + 442 + ], + "blocks": [ + { + "bbox": [ + 45, + 326, + 279, + 358 + ], + "lines": [ + { + "bbox": [ + 45, + 326, + 279, + 358 + ], + "spans": [ + { + "bbox": [ + 45, + 326, + 279, + 358 + ], + "type": "text", + "content": "Table 4. OWTAL localization results on ActivityNet-v1.3. The values are averaged on t-IoU thresholds [0.5:0.05:0.95]. The metric v-Acc(%) is also presented for reference." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 48, + 360, + 279, + 442 + ], + "lines": [ + { + "bbox": [ + 48, + 360, + 279, + 442 + ], + "spans": [ + { + "bbox": [ + 48, + 360, + 279, + 442 + ], + "type": "table", + "html": "
MethodsTop-K mAP@Avg(%)mAP@Avg(%)v-Acc(%)
Top1Top3Top5
ASM-Loc + SoftMax12.713.213.513.586.37
ASM-Loc + OpenMax12.012.612.812.981.74
ASM-Loc + ARPL17.017.617.818.096.81
ASM-Loc + EDL16.016.716.917.093.38
ASM-Loc + CELL(Ours)17.918.418.618.997.83
", + "image_path": "45f17487ecd280aaa3f781ba27854155493d9e5004b7be397d896944bedb01c4.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 52, + 511, + 272, + 594 + ], + "blocks": [ + { + "bbox": [ + 45, + 466, + 279, + 509 + ], + "lines": [ + { + "bbox": [ + 45, + 466, + 279, + 509 + ], + "spans": [ + { + "bbox": [ + 45, + 466, + 279, + 509 + ], + "type": "text", + "content": "Table 5. Ablation study of our proposed CELL. MEP denotes the multi-scale extended-range perception module, KBPL represents the knowledge-guided bipolar prototype learning module, and UA denotes the uncertainty-aware adaptor " + }, + { + "bbox": [ + 45, + 466, + 279, + 509 + ], + "type": "inline_equation", + "content": "\\zeta" + }, + { + "bbox": [ + 45, + 466, + 279, + 509 + ], + "type": "text", + "content": " described in Eq. (10)." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 52, + 511, + 272, + 594 + ], + "lines": [ + { + "bbox": [ + 52, + 511, + 272, + 594 + ], + "spans": [ + { + "bbox": [ + 52, + 511, + 272, + 594 + ], + "type": "table", + "html": "
ExpMEPKBPLUATop-k mAP@Avg(%)mAP@Avg(%)
Top-10Top-20Top-50
1XXX14.320.226.027.8
2✓XX15.321.226.929.0
3X✓X15.221.126.929.2
4✓✓X16.022.429.431.6
5✓✓✓16.423.030.332.3
", + "image_path": "e9e3f02adf708b68d66b952abdc19bfb0f54c61dc54b68c2ead31246fe2ccc7e.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 45, + 602, + 279, + 730 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 602, + 279, + 730 + ], + "spans": [ + { + "bbox": [ + 45, + 602, + 279, + 730 + ], + "type": "text", + "content": "Comparison results on ActivityNet-v1.3. To further increase the openness and verify the applicability of our proposed method for open-world scenarios, we further train our model on the known splits of THUMOS-14 and test on ActivityNet-v1.3. Although the testing set only consists of unknown action categories, in order to ensure a fair comparison, the model is still tested according to Algorithm 1, without using any additional prior knowledge. As shown in Table 4, our proposed CELL obtains consistently favorable performance. Compared with EDL, CELL shows superior performance with an absolute gain of " + }, + { + "bbox": [ + 45, + 602, + 279, + 730 + ], + "type": "inline_equation", + "content": "1.9\\%" + }, + { + "bbox": [ + 45, + 602, + 279, + 730 + ], + "type": "text", + "content": " in mAP@Avg." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 297, + 106, + 388, + 118 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 106, + 388, + 118 + ], + "spans": [ + { + "bbox": [ + 297, + 106, + 388, + 118 + ], + "type": "text", + "content": "4.3. Ablation Study" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 296, + 122, + 530, + 309 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 122, + 530, + 309 + ], + "spans": [ + { + "bbox": [ + 296, + 122, + 530, + 309 + ], + "type": "text", + "content": "This section shows the effectiveness of our modules. In Table 5, the impact of progressively adding each component is presented, proving their contributions clearly. According to the cascade steps of our method, we perform ablation study on three components, Multi-scale Extended-range Perception module (MEP), Knowledge-guided Bipolar Prototype Learning module (KBPL), and the Uncertainty-aware Adaptor (UA) in CEE module. Note that without MEP we replace the fused feature by the average of snippet features in " + }, + { + "bbox": [ + 296, + 122, + 530, + 309 + ], + "type": "inline_equation", + "content": "\\Theta" + }, + { + "bbox": [ + 296, + 122, + 530, + 309 + ], + "type": "text", + "content": ", and without UA we simply set " + }, + { + "bbox": [ + 296, + 122, + 530, + 309 + ], + "type": "inline_equation", + "content": "\\zeta" + }, + { + "bbox": [ + 296, + 122, + 530, + 309 + ], + "type": "text", + "content": " in Eq. (11) to 1. As shown in Table 5, every step of our method brings effective performance improvement. Moreover, it is noteworthy that MEP perceiving temporal contexts in multiscale ranges and KBPL leveraging inter-video correlations and prior knowledge can enhance and complement each other, thus significantly improving the performance." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 297, + 315, + 373, + 327 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 315, + 373, + 327 + ], + "spans": [ + { + "bbox": [ + 297, + 315, + 373, + 327 + ], + "type": "text", + "content": "5. Conclusions" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 296, + 333, + 530, + 612 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 333, + 530, + 612 + ], + "spans": [ + { + "bbox": [ + 296, + 333, + 530, + 612 + ], + "type": "text", + "content": "Targeting at open-world weakly-supervised temporal action localization (OWTAL), we propose a cascade evidential learning framework, which entails three main components: (1) Multi-scale extended-range perception module perceiving temporal contexts in diverse ranges for collecting initial evidence; (2) Knowledge-guided bipolar prototype learning strategy exploring inter-video relations under the guidance of prior knowledge for seeking supplementary evidence support; (3) Cascade evidence enhancement for final evidence calibration. In our extensive experiments, CELL achieves state-of-the-art performance on various metrics. Several limitations of this work are noteworthy. Firstly, never-before-seen human action categories are constantly emerging, thus lifelong learning paradigm which can keep embracing new action categories is even more suitable for OWTAL. Secondly, the class annotation granularity of the training set will interfere with the unknown action identification in open-world scenarios. Specifically, if some fine-grained actions are not labeled during training, e.g., the run-up movement in HighJump videos in THUMOS14, the model may misjudge such unknown actions as the background in testing. We assume the problem can be alleviated by training on large-scale datasets with fine-grained annotations. These directions are left as future work." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 297, + 620, + 396, + 633 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 620, + 396, + 633 + ], + "spans": [ + { + "bbox": [ + 297, + 620, + 396, + 633 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 296, + 636, + 530, + 730 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 636, + 530, + 730 + ], + "spans": [ + { + "bbox": [ + 296, + 636, + 530, + 730 + ], + "type": "text", + "content": "This work was supported by the National Key Research & Development Plan of China under Grant 2020AAA0106200, in part by the National Natural Science Foundation of China under Grants 62036012, U21B2044, 62236008, 62102415, 61721004, 62072286, 62072455, 62002355, in part by Beijing Natural Science Foundation (L201001), and in part by Open Research Projects of Zhejiang Lab (NO.2022RC0AB02)." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 277, + 746, + 301, + 756 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 277, + 746, + 301, + 756 + ], + "spans": [ + { + "bbox": [ + 277, + 746, + 301, + 756 + ], + "type": "text", + "content": "14748" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 595, + 842 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 105, + 102, + 117 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 105, + 102, + 117 + ], + "spans": [ + { + "bbox": [ + 47, + 105, + 102, + 117 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 124, + 279, + 720 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 51, + 124, + 279, + 155 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 51, + 124, + 279, + 155 + ], + "spans": [ + { + "bbox": [ + 51, + 124, + 279, + 155 + ], + "type": "text", + "content": "[1] Alexander Amini, Wilko Schwarting, Ava Soleimany, and Daniela Rus. Deep evidential regression. In NeurIPS, 2020. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 51, + 157, + 279, + 178 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 51, + 157, + 279, + 178 + ], + "spans": [ + { + "bbox": [ + 51, + 157, + 279, + 178 + ], + "type": "text", + "content": "[2] Wentao Bao, Qi Yu, and Yu Kong. Evidential deep learning for open set action recognition. In ICCV, 2021. 3, 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 51, + 179, + 279, + 201 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 51, + 179, + 279, + 201 + ], + "spans": [ + { + "bbox": [ + 51, + 179, + 279, + 201 + ], + "type": "text", + "content": "[3] Wentao Bao, Qi Yu, and Yu Kong. Opental: Towards open set temporal action localization. In CVPR, 2022. 1, 2, 3, 6, 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 51, + 202, + 279, + 223 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 51, + 202, + 279, + 223 + ], + "spans": [ + { + "bbox": [ + 51, + 202, + 279, + 223 + ], + "type": "text", + "content": "[4] Abhijit Bendale and Terrance E Boult. Towards open set deep networks. In CVPR, 2016. 2, 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 51, + 223, + 279, + 254 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 51, + 223, + 279, + 254 + ], + "spans": [ + { + "bbox": [ + 51, + 223, + 279, + 254 + ], + "type": "text", + "content": "[5] Joao Carreira and Andrew Zisserman. Quo vadis, action recognition? a new model and the kinetics dataset. In CVPR, 2017. 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 51, + 255, + 279, + 298 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 51, + 255, + 279, + 298 + ], + "spans": [ + { + "bbox": [ + 51, + 255, + 279, + 298 + ], + "type": "text", + "content": "[6] Yu-Wei Chao, Sudheendra Vijayanarasimhan, Bryan Seybold, David A Ross, Jia Deng, and Rahul Sukthankar. Rethinking the faster r-cnn architecture for temporal action localization. In CVPR, 2018. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 51, + 298, + 279, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 51, + 298, + 279, + 341 + ], + "spans": [ + { + "bbox": [ + 51, + 298, + 279, + 341 + ], + "type": "text", + "content": "[7] Guangyao Chen, Peixi Peng, Xiangqian Wang, and Yonghong Tian. Adversarial reciprocal points learning for open set recognition. arXiv preprint arXiv:2103.00953, 2021. 2, 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 51, + 342, + 279, + 384 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 51, + 342, + 279, + 384 + ], + "spans": [ + { + "bbox": [ + 51, + 342, + 279, + 384 + ], + "type": "text", + "content": "[8] Guangyao Chen, Limeng Qiao, Yemin Shi, Peixi Peng, Jia Li, Tiejun Huang, Shiliang Pu, and Yonghong Tian. Learning open set network with discriminative reciprocal points. In ECCV. Springer, 2020. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 51, + 386, + 279, + 417 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 51, + 386, + 279, + 417 + ], + "spans": [ + { + "bbox": [ + 51, + 386, + 279, + 417 + ], + "type": "text", + "content": "[9] Mengyuan Chen, Junyu Gao, Shicai Yang, and Changsheng Xu. Dual-evidential learning for weakly-supervised temporal action localization. In ECCV. Springer, 2022. 1, 2, 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 419, + 279, + 450 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 419, + 279, + 450 + ], + "spans": [ + { + "bbox": [ + 47, + 419, + 279, + 450 + ], + "type": "text", + "content": "[10] Junyu Gao, Mengyuan Chen, and Changsheng Xu. Fine-grained temporal contrastive learning for weakly-supervised temporal action localization. In CVPR, 2022. 2, 3, 5, 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 451, + 279, + 483 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 451, + 279, + 483 + ], + "spans": [ + { + "bbox": [ + 47, + 451, + 279, + 483 + ], + "type": "text", + "content": "[11] Junyu Gao and Changsheng Xu. Fast video moment retrieval. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1523-1532, 2021. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 47, + 484, + 279, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 484, + 279, + 525 + ], + "spans": [ + { + "bbox": [ + 47, + 484, + 279, + 525 + ], + "type": "text", + "content": "[12] Junyu Gao, Tianzhu Zhang, and Changsheng Xu. Graph convolutional tracking. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 4649-4659, 2019. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 47, + 527, + 279, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 527, + 279, + 568 + ], + "spans": [ + { + "bbox": [ + 47, + 527, + 279, + 568 + ], + "type": "text", + "content": "[13] Junyu Gao, Tianzhu Zhang, and Changsheng Xu. I know the relationships: Zero-shot action recognition via two-stream graph convolutional networks and knowledge graphs. In AAAI, 2019. 5" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 47, + 570, + 279, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 570, + 279, + 611 + ], + "spans": [ + { + "bbox": [ + 47, + 570, + 279, + 611 + ], + "type": "text", + "content": "[14] Junyu Gao, Tianzhu Zhang, and Changsheng Xu. Smart: Joint sampling and regression for visual tracking. IEEE Transactions on Image Processing, 28(8):3923-3935, 2019. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 47, + 613, + 279, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 613, + 279, + 656 + ], + "spans": [ + { + "bbox": [ + 47, + 613, + 279, + 656 + ], + "type": "text", + "content": "[15] Junyu Gao, Tianzhu Zhang, and Changsheng Xu. Learning to model relationships for zero-shot video classification. IEEE Transactions on Pattern Analysis and Machine Intelligence, 43(10):3476-3491, 2021. 5" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 47, + 657, + 279, + 698 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 657, + 279, + 698 + ], + "spans": [ + { + "bbox": [ + 47, + 657, + 279, + 698 + ], + "type": "text", + "content": "[16] Chuanxing Geng, Sheng-jun Huang, and Songcan Chen. Recent advances in open set recognition: A survey. IEEE transactions on pattern analysis and machine intelligence, 43(10):3614-3631, 2020. 7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 47, + 700, + 279, + 720 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 700, + 279, + 720 + ], + "spans": [ + { + "bbox": [ + 47, + 700, + 279, + 720 + ], + "type": "text", + "content": "[17] Zongbo Han, Changqing Zhang, Huazhu Fu, and Joey Tianyi Zhou. Trusted multi-view classification. In ICLR, 2020. 3" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 299, + 106, + 530, + 718 + ], + "type": "list", + "angle": 0, + "index": 36, + "blocks": [ + { + "bbox": [ + 299, + 106, + 530, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 106, + 530, + 149 + ], + "spans": [ + { + "bbox": [ + 299, + 106, + 530, + 149 + ], + "type": "text", + "content": "[18] Bo He, Xitong Yang, Le Kang, Zhiyu Cheng, Xin Zhou, and Abhinav Shrivastava. Asm-loc: Action-aware segment modeling for weakly-supervised temporal action localization. In CVPR, 2022. 1, 5, 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 299, + 150, + 530, + 192 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 150, + 530, + 192 + ], + "spans": [ + { + "bbox": [ + 299, + 150, + 530, + 192 + ], + "type": "text", + "content": "[19] Fa-Ting Hong, Jia-Chang Feng, Dan Xu, Ying Shan, and Wei-Shi Zheng. Cross-modal consensus network for weakly supervised temporal action localization. In ACM MM, 2021. 1, 2, 3, 5, 6, 7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 299, + 194, + 530, + 225 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 194, + 530, + 225 + ], + "spans": [ + { + "bbox": [ + 299, + 194, + 530, + 225 + ], + "type": "text", + "content": "[20] Yibo Hu, Yuzhe Ou, Xujiang Zhao, Jin-Hee Cho, and Feng Chen. Multidimensional uncertainty-aware evidential neural networks. In AAAI, 2021. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 299, + 227, + 530, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 227, + 530, + 258 + ], + "spans": [ + { + "bbox": [ + 299, + 227, + 530, + 258 + ], + "type": "text", + "content": "[21] Linjiang Huang, Liang Wang, and Hongsheng Li. Weakly supervised temporal action localization via representative snippet knowledge propagation. In CVPR, 2022. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 299, + 260, + 530, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 260, + 530, + 281 + ], + "spans": [ + { + "bbox": [ + 299, + 260, + 530, + 281 + ], + "type": "text", + "content": "[22] Audun Jsang. Subjective logic: A formalism for reasoning under uncertainty. Springer Verlag, 2016. 3, 4" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 299, + 282, + 530, + 324 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 282, + 530, + 324 + ], + "spans": [ + { + "bbox": [ + 299, + 282, + 530, + 324 + ], + "type": "text", + "content": "[23] Will Kay, Joao Carreira, Karen Simonyan, Brian Zhang, Chloe Hillier, Sudheendra Vijayanarasimhan, Fabio Viola, Tim Green, Trevor Back, Paul Natsev, et al. The kinetics human action video dataset. arXiv:1705.06950, 2017. 3, 7" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 299, + 325, + 530, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 325, + 530, + 357 + ], + "spans": [ + { + "bbox": [ + 299, + 325, + 530, + 357 + ], + "type": "text", + "content": "[24] Pilhyeon Lee, Jinglu Wang, Yan Lu, and Hyeran Byun. Weakly-supervised temporal action localization by uncertainty modeling. In AAAI, 2021. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 299, + 358, + 530, + 390 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 358, + 530, + 390 + ], + "spans": [ + { + "bbox": [ + 299, + 358, + 530, + 390 + ], + "type": "text", + "content": "[25] Bolian Li, Zongbo Han, Haining Li, Huazhu Fu, and Changqing Zhang. Trustworthy long-tailed classification. arXiv preprint arXiv:2111.09030, 2021. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 299, + 392, + 530, + 433 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 392, + 530, + 433 + ], + "spans": [ + { + "bbox": [ + 299, + 392, + 530, + 433 + ], + "type": "text", + "content": "[26] Chuming Lin, Chengming Xu, Donghao Luo, Yabiao Wang, Ying Tai, Chengjie Wang, Jilin Li, Feiyue Huang, and Yanwei Fu. Learning salient boundary feature for anchor-free temporal action localization. In CVPR, 2021. 7" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 299, + 435, + 530, + 466 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 435, + 530, + 466 + ], + "spans": [ + { + "bbox": [ + 299, + 435, + 530, + 466 + ], + "type": "text", + "content": "[27] Wei Liu, Xiaodong Yue, Yufei Chen, and Thierry Denoeux. Trusted multi-view deep learning with opinion aggregation. In AAAI, 2022. 3" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 299, + 468, + 530, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 468, + 530, + 499 + ], + "spans": [ + { + "bbox": [ + 299, + 468, + 530, + 499 + ], + "type": "text", + "content": "[28] Vincent Lonij, Ambrish Rawat, and Maria-Irina Nicolae. Open-world visual recognition using knowledge graphs. arXiv preprint arXiv:1708.08310, 2017. 2" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 299, + 500, + 530, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 500, + 530, + 542 + ], + "spans": [ + { + "bbox": [ + 299, + 500, + 530, + 542 + ], + "type": "text", + "content": "[29] Wang Luo, Tianzhu Zhang, Wenfei Yang, Jingen Liu, Tao Mei, Feng Wu, and Yongdong Zhang. Action unit memory network for weakly supervised temporal action localization. In CVPR, 2021. 2" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 299, + 543, + 530, + 575 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 543, + 530, + 575 + ], + "spans": [ + { + "bbox": [ + 299, + 543, + 530, + 575 + ], + "type": "text", + "content": "[30] Lawrence Neal, Matthew Olson, Xiaoli Fern, Weng-Keen Wong, and Fuxin Li. Open set learning with counterfactual images. In ECCV, 2018. 2" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 299, + 577, + 530, + 598 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 577, + 530, + 598 + ], + "spans": [ + { + "bbox": [ + 299, + 577, + 530, + 598 + ], + "type": "text", + "content": "[31] Poojan Oza and Vishal M Patel. C2ae: Class conditioned auto-encoder for open-set recognition. In CVPR, 2019. 2" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 299, + 599, + 530, + 629 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 599, + 530, + 629 + ], + "spans": [ + { + "bbox": [ + 299, + 599, + 530, + 629 + ], + "type": "text", + "content": "[32] Jeffrey Pennington, Richard Socher, and Christopher D Manning. Glove: Global vectors for word representation. In EMNLP, 2014. 5" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 299, + 631, + 530, + 674 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 631, + 530, + 674 + ], + "spans": [ + { + "bbox": [ + 299, + 631, + 530, + 674 + ], + "type": "text", + "content": "[33] Sanqing Qu, Guang Chen, Zhijun Li, Lijun Zhang, Fan Lu, and Alois Knoll. Acm-net: Action context modeling network for weakly-supervised temporal action localization. arXiv preprint arXiv:2104.02967, 2021. 2, 3, 6" + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 299, + 676, + 530, + 718 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 676, + 530, + 718 + ], + "spans": [ + { + "bbox": [ + 299, + 676, + 530, + 718 + ], + "type": "text", + "content": "[34] Walter J Scheirer, Anderson de Rezende Rocha, Archana Sapkota, and Terrance E Boult. Toward open set recognition. IEEE transactions on pattern analysis and machine intelligence, 35(7):1757-1772, 2012. 2" + } + ] + } + ], + "index": 35 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 277, + 747, + 300, + 756 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 277, + 747, + 300, + 756 + ], + "spans": [ + { + "bbox": [ + 277, + 747, + 300, + 756 + ], + "type": "text", + "content": "14749" + } + ] + } + ], + "index": 37 + } + ], + "page_size": [ + 595, + 842 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 106, + 279, + 426 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 47, + 106, + 279, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 106, + 279, + 138 + ], + "spans": [ + { + "bbox": [ + 47, + 106, + 279, + 138 + ], + "type": "text", + "content": "[35] Murat Sensoy, Lance Kaplan, and Melih Kandemir. Evidential deep learning to quantify classification uncertainty. In NeurIPS, 2018. 3, 4, 5" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 139, + 279, + 170 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 139, + 279, + 170 + ], + "spans": [ + { + "bbox": [ + 47, + 139, + 279, + 170 + ], + "type": "text", + "content": "[36] Baifeng Shi, Qi Dai, Yadong Mu, and Jingdong Wang. Weakly-supervised action localization by generative attention modeling. In CVPR, 2020. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 171, + 279, + 202 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 171, + 279, + 202 + ], + "spans": [ + { + "bbox": [ + 47, + 171, + 279, + 202 + ], + "type": "text", + "content": "[37] Zheng Shou, Hang Gao, Lei Zhang, Kazuyuki Miyazawa, and Shih-Fu Chang. Autoloc: Weakly-supervised temporal action localization in untrimmed videos. In ECCV, 2018. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 203, + 279, + 234 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 203, + 279, + 234 + ], + "spans": [ + { + "bbox": [ + 47, + 203, + 279, + 234 + ], + "type": "text", + "content": "[38] Krishna Kumar Singh and Yong Jae Lee. Hide-and-seek: Forcing a network to be meticulous for weakly-supervised object and action localization. In ICCV, 2017. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 235, + 279, + 266 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 235, + 279, + 266 + ], + "spans": [ + { + "bbox": [ + 47, + 235, + 279, + 266 + ], + "type": "text", + "content": "[39] Robyn Speer, Joshua Chin, and Catherine Havasi. Concept-net 5.5: An open multilingual graph of general knowledge. In AAAI, 2017. 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 266, + 279, + 298 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 266, + 279, + 298 + ], + "spans": [ + { + "bbox": [ + 47, + 266, + 279, + 298 + ], + "type": "text", + "content": "[40] Xin Sun, Zhenning Yang, Chi Zhang, Keck-Voon Ling, and Guohao Peng. Conditional gaussian distribution learning for open set recognition. In CVPR, 2020. 2, 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 298, + 279, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 298, + 279, + 330 + ], + "spans": [ + { + "bbox": [ + 47, + 298, + 279, + 330 + ], + "type": "text", + "content": "[41] Dennis Ulmer. A survey on evidential deep learning for single-pass uncertainty estimation. arXiv preprint arXiv:2110.03051, 2021. 3, 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 330, + 279, + 362 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 330, + 279, + 362 + ], + "spans": [ + { + "bbox": [ + 47, + 330, + 279, + 362 + ], + "type": "text", + "content": "[42] Sagar Vaze, Kai Han, Andrea Vedaldi, and Andrew Zisserman. Open-set recognition: A good closed-set classifier is all you need. In ICLR, 2022. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 362, + 279, + 393 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 362, + 279, + 393 + ], + "spans": [ + { + "bbox": [ + 47, + 362, + 279, + 393 + ], + "type": "text", + "content": "[43] Limin Wang, Yuanjun Xiong, Dahua Lin, and Luc Van Gool. Untrimmednets for weakly supervised action recognition and detection. In CVPR, 2017. 1, 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 394, + 279, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 394, + 279, + 426 + ], + "spans": [ + { + "bbox": [ + 47, + 394, + 279, + 426 + ], + "type": "text", + "content": "[44] Huijuan Xu, Abir Das, and Kate Saenko. R-c3d: Region convolutional 3d network for temporal activity detection. In ICCV, 2017. 3" + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 299, + 107, + 530, + 399 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 299, + 107, + 530, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 107, + 530, + 138 + ], + "spans": [ + { + "bbox": [ + 299, + 107, + 530, + 138 + ], + "type": "text", + "content": "[45] Ronald R Yager and Liping Liu. Classic works of the Dempster-Shafer theory of belief functions, volume 219. Springer, 2008. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 299, + 140, + 530, + 181 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 140, + 530, + 181 + ], + "spans": [ + { + "bbox": [ + 299, + 140, + 530, + 181 + ], + "type": "text", + "content": "[46] Wenfei Yang, Tianzhu Zhang, Xiaoyuan Yu, Tian Qi, Yongdong Zhang, and Feng Wu. Uncertainty guided collaborative training for weakly supervised temporal action detection. In CVPR, 2021. 2, 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 299, + 183, + 530, + 225 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 183, + 530, + 225 + ], + "spans": [ + { + "bbox": [ + 299, + 183, + 530, + 225 + ], + "type": "text", + "content": "[47] Ryota Yoshihashi, Wen Shao, Rei Kawakami, Shaodi You, Makoto Iida, and Takeshi Naemura. Classification-reconstruction learning for open-set recognition. In CVPR, 2019. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 299, + 227, + 530, + 267 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 227, + 530, + 267 + ], + "spans": [ + { + "bbox": [ + 299, + 227, + 530, + 267 + ], + "type": "text", + "content": "[48] Can Zhang, Meng Cao, Dongming Yang, Jie Chen, and Yuexian Zou. Cola: Weakly-supervised temporal action localization with snippet contrastive learning. In CVPR, 2021. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 299, + 270, + 530, + 312 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 270, + 530, + 312 + ], + "spans": [ + { + "bbox": [ + 299, + 270, + 530, + 312 + ], + "type": "text", + "content": "[49] Chengwei Zhang, Yunlu Xu, Zhanzhan Cheng, Yi Niu, Shiliang Pu, Fei Wu, and Futai Zou. Adversarial seeded sequence growing for weakly-supervised temporal action localization. In ACM MM, 2019. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 299, + 314, + 530, + 355 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 314, + 530, + 355 + ], + "spans": [ + { + "bbox": [ + 299, + 314, + 530, + 355 + ], + "type": "text", + "content": "[50] Jia-Xing Zhong, Nannan Li, Weijie Kong, Tao Zhang, Thomas H Li, and Ge Li. Step-by-step erasion, one-by-one collection: a weakly supervised temporal action detector. In ACM MM, 2018. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 299, + 357, + 530, + 399 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 357, + 530, + 399 + ], + "spans": [ + { + "bbox": [ + 299, + 357, + 530, + 399 + ], + "type": "text", + "content": "[51] Zhun Zhong, Linchao Zhu, Zhiming Luo, Shaozi Li, Yi Yang, and Nicu Sebe. Openmix: Reviving known knowledge for discovering novel visual categories in an open world. In CVPR, 2021. 2" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 277, + 746, + 300, + 756 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 277, + 746, + 300, + 756 + ], + "spans": [ + { + "bbox": [ + 277, + 746, + 300, + 756 + ], + "type": "text", + "content": "14750" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 595, + 842 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Cascaded Local Implicit Transformer for Arbitrary-Scale Super-Resolution/ae047e51-87a0-4eec-807f-144baf2a48b8_content_list.json b/2023/Cascaded Local Implicit Transformer for Arbitrary-Scale Super-Resolution/ae047e51-87a0-4eec-807f-144baf2a48b8_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..512771cbb395d7ae7d3eb6f9c93f96ed86a50821 --- /dev/null +++ b/2023/Cascaded Local Implicit Transformer for Arbitrary-Scale Super-Resolution/ae047e51-87a0-4eec-807f-144baf2a48b8_content_list.json @@ -0,0 +1,1476 @@ +[ + { + "type": "text", + "text": " Cascaded Local Implicit Transformer for Arbitrary-Scale Super-Resolution", + "text_level": 1, + "bbox": [ + 102, + 63, + 867, + 85 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Hao-Wei Chen\\*,1,2 Yu-Syuan Xu\\*,2 Min-Fong Hong\\* Yi-Min Tsai\\* Hsien-Kai Kuo\\* Chun-Yi Lee\\* Elsa Lab, National Tsing Hua University \\*MediaTek Inc.", + "bbox": [ + 86, + 94, + 877, + 135 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{jaroslaw1007, cylee}@gapp.nthu.edu.tw", + "bbox": [ + 316, + 138, + 655, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{Yu-Syuan.Xu, romulus.hong, Yi-Min.Tsai, Hsienkai.Kuo}@mediatek.com", + "bbox": [ + 183, + 155, + 784, + 170 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 220, + 313, + 236 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Implicit neural representation has recently shown a promising ability in representing images with arbitrary resolutions. In this paper, we present a Local Implicit Transformer (LIT), which integrates the attention mechanism and frequency encoding technique into a local implicit image function. We design a cross-scale local attention block to effectively aggregate local features and a local frequency encoding block to combine positional encoding with Fourier domain information for constructing high-resolution images. To further improve representative power, we propose a Cascaded LIT (CLIT) that exploits multi-scale features, along with a cumulative training strategy that gradually increases the upsampling scales during training. We have conducted extensive experiments to validate the effectiveness of these components and analyze various training strategies. The qualitative and quantitative results demonstrate that LIT and CLIT achieve favorable results and outperform the prior works in arbitrary super-resolution tasks.", + "bbox": [ + 73, + 243, + 473, + 517 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 527, + 209, + 542 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Single Image Super-Resolution (SISR) is the process of reconstructing high-resolution (HR) images from their corresponding low-resolution (LR) counterparts. SISR has long been recognized as a challenging task in the low-level vision domain due to its ill-posed nature, and has attracted a number of researchers dedicated to this field of study over the past decade [1-21]. A line of SISR research referred to as 'fixed-scale SR' [1-15] focuses on extracting feature embeddings from LR images and leveraging these embeddings to upsample images with a predefined factor through learnable deconvolutions [3] or sub-pixel convolutions [4]. Despite their success, many of the proposed approaches necessitate a distinct deep neural network model for each upsampling scale, which is usually restricted to a limited selection of integers (e.g., $2 \\times$ , $3 \\times$ , $4 \\times$ ). Such a limitation constrains the potential applications and deployment options of SISR models. To overcome this limitation, approaches for upsampling LR images in a continuous manner via a single model emerge and attracted considerable attention recently.", + "bbox": [ + 75, + 546, + 468, + 833 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Over the past few years, arbitrary-scale SR has emerged and attracted considerable attention from researchers [16-21]. Apart from the pioneering work Meta-SR [16], recent", + "bbox": [ + 76, + 833, + 468, + 878 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/7dd09a46aa8a3ae0b389404446b4cb11bc731515f8fca5e1d6da6786a3498803.jpg", + "image_caption": [ + "LR", + "HR" + ], + "image_footnote": [], + "bbox": [ + 522, + 203, + 691, + 262 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/d4d3d990051fde0f40d410629c121267ff31938a40770d96a1caec70438d2d47.jpg", + "image_caption": [ + "LR" + ], + "image_footnote": [], + "bbox": [ + 714, + 203, + 792, + 262 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/bcf88dce662e00f036c471a554bf325f4d362e8c2be90085943be045957dd27e.jpg", + "image_caption": [ + "HR" + ], + "image_footnote": [], + "bbox": [ + 800, + 204, + 879, + 263 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/35530f43854e42ca6df6273a074902f118a82adb1809fbb418332d116d1950f5.jpg", + "image_caption": [ + "(a)", + "LIIF", + "Loc" + ], + "image_footnote": [], + "bbox": [ + 524, + 273, + 691, + 335 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/254c026769d360b5c4d72475f0f5d2476cfbb5e39efffd59a99453f68d2be3e2.jpg", + "image_caption": [ + "LIIF" + ], + "image_footnote": [], + "bbox": [ + 714, + 273, + 879, + 335 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/b71c8b963ebff36469aa32a422c08a5118dd1fb4dc208e9d66a6ecbf64f4925b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 524, + 340, + 697, + 405 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/d8f8c1a76b4eaec2c548a4266eed42cb65b7eb618ee9845cad898259dbbac04e.jpg", + "image_caption": [ + "Ours" + ], + "image_footnote": [], + "bbox": [ + 714, + 340, + 792, + 405 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/113b3566563e8c0ffdcc6c13bc5b1e00884b78b508154fe15859ca0082a9038f.jpg", + "image_caption": [ + "Figure 1. An illustration and comparison of different approaches that take into account nearby pixels for continuous upsampling: (a) the local ensemble method used in [17], and (b) our proposed local attention mechanism." + ], + "image_footnote": [], + "bbox": [ + 803, + 340, + 888, + 407 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "endevavors [17-21] have achieved arbitrary-scale SR by replacing the upsampling layers commonly adopted by previous approaches with local implicit image functions, and have demonstrated favorable performance. These local implicit functions employ multi-layer perceptrons (MLPs) to map 2D coordinates and corresponding latent representations to RGB values. Fig. 1 illustrates how different approaches sample latent representations based on the queried coordinates (depicted as the red dots). Fig. 1 (a) illustrates the local ensemble technique adopted by contemporary mainstream methods [17-21]. It calculates the RGB value of the queried coordinate by taking the weighted average of those of the surrounding four pixels based on their relative distances to the queried coordinate. This approach, however, does not consider contextual information and relies solely on distance. For instance, in Fig. 1, the queried coordinates are intentionally designed to lie on edges. However, merely calculating the weighted average of pixels fails to reflect the contextual information about the image content, thereby preventing the accurate capture of the necessary features for performing SR. As a result, although pixel distance plays a vital role in SR tasks, it is essential to concentrate more on the contextual information in an image.", + "bbox": [ + 496, + 458, + 892, + 806 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In light of the above observations, we propose a Local Implicit Transformer (LIT), which expands the numbers of referenced latent vectors and accounts for the feature correlation in the context by exploiting the attention mechanism [22]. LIT comprises a Cross-Scale Local Attention Block (CSLAB), a Local Frequency Encoding Block", + "bbox": [ + 496, + 809, + 893, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation.", + "bbox": [ + 236, + 0, + 807, + 18 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "Except for this watermark, it is identical to the accepted version;", + "bbox": [ + 323, + 16, + 723, + 30 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 292, + 31, + 753, + 45 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Equal contribution", + "bbox": [ + 101, + 887, + 209, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "18257", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "(LFEB), and a decoder. CSLAB generates attention maps based on the bilinearly interpolated latent vectors at the queried coordinates and the key latent vectors sampled from a grid of coordinates with relative positional bias [23, 24]. The first and second columns of Fig. 1 (b) visualize the attention maps generated by LIT, where the attention areas align closely with the edges. By applying attention maps to feature embeddings, the RGB values of the queried coordinates can be contextually predicted. Moreover, inspired by [21, 25, 26], we introduce LFEB, which projects relative coordinates into latent space to address the spectral bias problem [27] of an implicit neural function. Specifically, relative coordinates are encoded into relative positional encoding, which are multiplied with the frequency encoding extracted from the feature embedding in the Fourier domain to generate the frequency embeddings. This design enables the frequency embedding to integrate the relative positional encoding with texture information, thereby augmenting the expressivity of relative coordinates. Finally, a decoder is adopted to produce RGB values by taking advantage of the attention feature embedding and the frequency embedding.", + "bbox": [ + 76, + 90, + 472, + 409 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In order to address the issue of diverse scaling factors and achieve arbitrary-scale super-resolution, it is crucial to consider the role of upsampling factors in constructing high-resolution images within the local implicit image function. However, simultaneously training a local implicit image function with a wide range of upsampling factors (e.g., $1 \\times \\sim 30 \\times$ ) poses significant challenges. As a result, we propose a cumulative training strategy to incrementally enhance the fuction's its representative power. The strategy initially trains the local implicit image function with small upsampling factors and then finetunes it with alternatively sampled small and large ones. Furthermore, we present Cascaded LIT (CLIT) to harness the advantages of multiscale feature embeddings, complementing missing details and information during one-step upsampling. The combination of the cumulative training strategy and CLIT enables efficient and effective handling of arbitrary-scale SR tasks.", + "bbox": [ + 76, + 415, + 472, + 674 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The main contributions of our work are summarized as follows: (1) We introduce the LIT architecture, which incorporates the local attention mechanism into arbitrary-scale SR (2) We further develop a cumulative training strategy and the cascaded framework CLIT to effectively handle large-scale upsampling. (3) We carry out comprehensive analyses of the performance impacts for LIT and CLIT. The extensive experimental findings demonstrate that the proposed LIT and CLIT are able to yield remarkable or comparable results across a wide range of benchmark datasets.", + "bbox": [ + 76, + 680, + 472, + 833 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The paper is organized as follows. Section 2 reviews the related work. Section 3 walks through the proposed LIT and CLIT frameworks and the implementation details. Section 4 presents the experimental results. Section 5 concludes.", + "bbox": [ + 76, + 839, + 472, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 89, + 640, + 106 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Implicit neural representation. Implicit neural representation is a technique for representing continuous-domain signals via coordinate-based multi-layer perceptrons (MLPs). Its concept has been adopted in various 3D tasks, e.g., 3D object shape modeling [28-32], 3D scene reconstruction [33-36], and 3D structure rendering [25, 37-39]. For example, NeRF [25] employs implicit neural representation to perform novel view synthesis, which maps coordinates to RGB colors for a specific scene. In the past few years, 2D applications of implicit neural representation have been attempted as well, such as image representation [40, 41] and super-resolution [17-21]. Our work is related to a technique called 'local implicit neural representation' [17, 21], which encodes LR images to feature embeddings such that similar information could be shared within local regions. Such local implicit neural representations are exploited to upscale LR images to HR ones.", + "bbox": [ + 496, + 114, + 893, + 375 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Single image super-resolution. In the past several years, various deep neural network (DNN) based architectures [1-15] have been proposed for SISR. Among these works, SR-CNN [1] pioneered the use of convolutional neural networks (CNNs) to achieve SISR in an end-to-end manner. It is later followed by several subsequent works that incorporated more complicated model architectures, such as residual blocks [6, 7], dense connections [8, 9], attention based mechanisms [10-12], or cascaded frameworks [5, 42, 43], to extract more effective feature representations for SISR. Recently, transformer-based methods [13-15] were introduced to SISR and achieved promising performance.", + "bbox": [ + 496, + 387, + 893, + 570 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Arbitrary-scale super-resolution. As discussed in Section 1, most of the contemporary SISR works limit their upsampling scales to specific integer values, and are required to train a distinct model for each upsampling scale. To overcome such a limitation, several approaches [16-21] were proposed to train a unified model for arbitrary upsampling scales. Meta-SR [16] proposed a meta-upscale module for predicting the weights of their convolutional filters from coordinates and scales. The predicted weights are then utilized to perform convolutions to generate HR images. In contrast to Meta-SR, LIIF [17] employs an MLP as a local implicit function, which takes a queried coordinate in an HR image, its nearby feature representations extracted from the corresponding LR image, as well as a cell size to predict an RGB value for that coordinate. UltraSR [18] and IPE [19] extended LIIF by replacing coordinates with the embedded ones to deal with the spectral bias issue [25, 27, 41, 44, 45] inherent in MLPs. LTE [21] further introduced a local texture estimator that transforms coordinates into Fourier domain information to enrich the representational capability of its local implicit function. Different from the above ap", + "bbox": [ + 496, + 583, + 893, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "18258", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/f4d95a8a3964cbfddd8daf78d2de9aac78c5aa7fa8793fc059c5fe65e9b8d4c6.jpg", + "image_caption": [ + "Figure 2. The proposed LIT framework. The local sampling operation samples input embeddings based on a grid of coordinates." + ], + "image_footnote": [], + "bbox": [ + 83, + 89, + 890, + 325 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "proaches, our proposed methodology exploits a novel local attention mechanism and a cascaded framework to deal with the arbitrary-scale SR. In order to fairly compare with the above approaches, we similarly adopt EDSR [7], RDN [9] and SwinIR [14] as the encoders for our LIT and CLIT.", + "bbox": [ + 75, + 357, + 468, + 434 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Methodology", + "text_level": 1, + "bbox": [ + 76, + 446, + 210, + 463 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we first provide an overview of the proposed LIT framework, followed by the implementation details of it and its main modules. We then discuss our cumulative training strategy, as well as the framework of CLIT.", + "bbox": [ + 75, + 470, + 468, + 532 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Overview of the LIT Framework", + "text_level": 1, + "bbox": [ + 76, + 541, + 366, + 556 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "LIT is a framework that employs a novel cross-scaled local attention mechanism and a local frequency encoding technique to perform arbitrary-scale SR tasks. Fig. 2 (a) provides an overview of the proposed framework, which aims at producing an HR image $I^{HR} \\in \\mathbb{R}^{r_hH \\times r_wW \\times 3}$ at 2D HR coordinates $\\mathbf{x}^{HR} \\in \\mathcal{X}$ from a given LR image $I^{LR} \\in \\mathbb{R}^{H \\times W \\times 3}$ at 2D LR coordinates $\\mathbf{x}^{LR} \\in \\mathcal{X}$ based on an arbitrary upsampling scale $\\mathbf{r} = \\{r_h, r_w\\}$ , where $\\mathcal{X}$ is the 2D coordinate space that is used to represent an image in the continuous domain. An encoder $E_{\\theta}$ first extracts a feature embedding $\\mathcal{Z} \\in \\mathbb{R}^{H \\times W \\times C}$ from $I^{LR}$ . The extracted $\\mathcal{Z}$ is then forwarded into LIT along with the 2D coordinates of $I^{HR}$ and a cell $= (2 / s_h, 2 / s_w)$ to generate the RGB values of a residual image $I_r^{HR} \\in \\mathbb{R}^{r_hH \\times r_wW \\times 3}$ in a pixel-wise fashion. Lastly, the residual image $I_r^{HR}$ is combined with a bilinearly upsampled image $I_{\\uparrow}^{HR} \\in \\mathbb{R}^{r_hH \\times r_wW \\times 3}$ via element-wise addition to derive the final HR image $I^{HR}$ .", + "bbox": [ + 75, + 564, + 472, + 824 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Local Implicit Transformer", + "text_level": 1, + "bbox": [ + 76, + 832, + 323, + 848 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "LIT is developed for mapping any 2D coordinate in the continuous image domain to an RGB color. As highlighted in Fig. 2 (b), it is composed of a cross-scale local attention", + "bbox": [ + 75, + 854, + 468, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "block (CSLAB), a local frequency encoding block (LFB), and a decoder $D_{\\phi}$ paraterizerized by $\\phi$ . The former two blocks are responsible for estimating a local latent embedding $\\tilde{z} \\in \\mathbb{R}^{G_h G_w \\times C}$ and a local frequency embedding $\\tilde{f} \\in \\mathbb{R}^{G_h G_w \\times C}$ , respectively, where $G_h$ and $G_w$ denote the height and width of local grids employed for performing local coordinate sampling, as depicted in Fig. 3. On the other hand, $D_{\\phi}$ utilizes these embeddings along with the provided cell to generate $I_r^{HR}$ . More specifically, LIT first projects $\\mathcal{Z}$ using four separate convolutional layers to obtain four latent embeddings, corresponding to query $q$ , key $k$ , value $v$ , and frequency $f$ . Based on a queried HR coordinate $x_q \\in \\mathbf{x}^{HR}$ , CSLAB and LFB estimate $\\tilde{z}$ and $\\tilde{f}$ as follows:", + "bbox": [ + 496, + 357, + 890, + 555 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {z} = C S L A B (\\delta x, q, k, v), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 606, + 566, + 890, + 582 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {f} = L F E B (\\delta x, f), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 625, + 590, + 890, + 609 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\delta \\mathbf {x} = \\left\\{x _ {q} - x ^ {(i, j)} \\right\\} _ {i \\in \\{1, 2, \\dots , G _ {h} \\}, j \\in \\{1, 2, \\dots , G _ {w} \\}}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 537, + 614, + 890, + 643 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $x^{(i,j)}\\in \\mathbf{x}^{LR}$ denotes an LR coordinate in the local grid indexed by $(i,j)$ , and $\\delta \\mathbf{x}$ represents the set of local relative coordinates defined by Eq. (3). The local grid is sampled in a manner that positions its center $x^{(\\lfloor G_h / 2\\rfloor +1,\\lfloor G_w / 2\\rfloor +1)}$ at the LR coordinate closest to $x_{q}$ . The query latent vector $q\\in \\mathbb{R}^{1\\times C}$ at the HR coordinate $x_{q}$ is computed by bilinear interpolation, while the remainder of the local latent embeddings $k\\in \\mathbb{R}^{G_hG_w\\times C}$ , $v\\in \\mathbb{R}^{G_hG_w\\times C}$ , and $f\\in \\mathbb{R}^{G_hG_w\\times C}$ are sampled at the local LR coordinates $\\mathbf{x} = \\{x^{i,j}\\}_{i\\in \\{1,2,\\dots,G_h\\},j\\in \\{1,2,\\dots,G_w\\}}$ . With the local latent embedding $\\tilde{z}$ and the local frequency embedding $\\tilde{f}$ , the function of $D_{\\phi}$ is formulated as follows:", + "bbox": [ + 496, + 648, + 890, + 830 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nI ^ {r} \\left(x _ {q}\\right) = D _ {\\phi} \\left(\\tilde {z}, \\tilde {f}, c\\right), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 620, + 840, + 890, + 859 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $I^{r}(x_{q})$ is the predicted RGB value at the queried coordinate $x_{q}$ , and $c = \\{HR_{\\Delta h}, HR_{\\Delta w}\\}$ denotes the cell that", + "bbox": [ + 496, + 869, + 890, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "18259", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/4b721537abf77883775292eb0d479c54213e81a052d46033b70500a91fbec677.jpg", + "image_caption": [ + "Figure 3. An illustration of the proposed local coordinate sampling scheme. The red rectangular region, outlined by a dashed line, represents the local grid with dimensions $G_{h} \\times G_{w}$ . The red dots indicate the sampled local LR coordinates, while the gray dots correspond to the HR coordinates. The grid dimensions $LR_{\\Delta h,w}$ and $HR_{\\Delta h,w}$ represent the unit sizes of pixels in LR and HR images, respectively." + ], + "image_footnote": [], + "bbox": [ + 78, + 85, + 472, + 306 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "represents the height and width of a pixel in an HR image, as illustrated in Fig. 3. $D_{\\phi}$ is implemented as a five-layer MLP utilizing Gaussian Error Linear Unit (GELU) activation [46], and is employed consistently across all images.", + "bbox": [ + 75, + 398, + 468, + 460 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Cross-scale local attention block. LIT exploits CSLAB to perform a local attention mechanism over a local grid to generate a local latent embedding $\\tilde{z}$ for each HR coordinate, as illustrated in Fig. 2 (c). CSLAB first calculates the inner product of $q$ and $k$ , adds the relative positional bias $B$ to the result, and obtains an attention matrix. This attention matrix is subsequently normalized by a Softmax operation to produce a local attention map. Finally, CSLAB performs element-wise multiplication of $v$ and the local attention map to derive $\\tilde{z}$ . The overall procedure is formulated as follows:", + "bbox": [ + 75, + 478, + 468, + 630 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\tilde {z} = \\operatorname {s o f t m a x} \\left(\\frac {q k ^ {\\top}}{\\sqrt {C}} + B\\right) \\times v, (5) \\\\ B = F C (\\gamma (\\delta \\mathbf {x})), (6) \\\\ \\gamma (\\delta \\mathbf {x}) = \\left[ \\sin \\left(2 ^ {0} \\delta \\mathbf {x}\\right), \\cos \\left(2 ^ {0} \\delta \\mathbf {x}\\right), \\dots \\right. (7) \\\\ , \\sin \\left(2 ^ {L - 1} \\delta \\mathbf {x}\\right), \\cos \\left(2 ^ {L - 1} \\delta \\mathbf {x}\\right) ], \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 148, + 638, + 468, + 744 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $C$ denotes the channel dimension of the local key latent embedding $k$ , $FC$ represents a fully-connected layer, $\\gamma$ is the positional encoding function, and $L$ is a hyperparameter. In this work, $L$ is set to 10, and the multi-head attention mechanism adopted is formulated in Eq. (8) as follows:", + "bbox": [ + 75, + 751, + 468, + 827 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {z} = \\operatorname {c o n c a t} \\left(\\operatorname {s o f t m a x} \\left(\\frac {q _ {i} k _ {i} ^ {\\top}}{\\sqrt {C / H}} + B _ {i}\\right) \\times v _ {i}\\right), \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 120, + 837, + 468, + 875 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $H$ is the number of attention heads and $i\\in [1,\\dots ,H]$", + "bbox": [ + 76, + 885, + 468, + 901 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/a9ca32d0bd97b5f4a24bbb9fda0903384bfb89b5b21fc8834c22b9195181d570.jpg", + "image_caption": [ + "Figure 4. An overview of the proposed CLIT framework." + ], + "image_footnote": [], + "bbox": [ + 504, + 90, + 890, + 244 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Local frequency encoding block. Fig. 2 (d) illustrates the local frequency encoding block, which incorporates dominant-frequency components to generate the local frequency embedding $\\tilde{f}$ . In order to account for dominant-frequency information, LFBB first encodes local frequency latent embedding $f$ using Fast Fourier Transform (FFT), which is then multiplied by the relative positional encoding to generate $\\tilde{f}$ . The resultant local frequency embedding $\\tilde{f}$ is then concatenated with the flattened local latent embedding $\\tilde{z}$ to form a mixing embedding, which can be used by the decoder consisting of MLPs to predict RGB values [47].", + "bbox": [ + 498, + 275, + 890, + 441 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Cumulative Training Strategy", + "text_level": 1, + "bbox": [ + 500, + 453, + 766, + 469 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this section, we discuss our proposed cumulative training strategy, which is developed for enhancing the performance of arbitrary-scale SR. The cumulative training strategy focuses on the schedule of the cell sizes selected during the training phase, as \"cell decoding\" has been recognized as an essential input to a local implicit image function [17]. Recent studies [17, 19] have observed that the effect of cell decoding on the performance of arbitrary-scale SR is prominent for in-distribution upsampling, but degrades significantly for out-of-distribution large-scale upsampling. Such trends are demonstrated in Table 3 of [17] and Table 4 of [19]. The authors in [21] also constrained their cell sizes to in-distribution ranges to mitigate the negative impact during evaluation. To overcome the degradation issue for out-of-distribution cell sizes, incorporating large cell sizes during training appears to be a promising solution. However, simply training the local implicit image function with a diverse range of cell sizes at once leads to a performance drop. According to our experimental observations presented in Section 4.4, we find that training the local implicit image function by alternatively switching between large and small cells offers positive impacts on the performance. Based on the above observations, we propose a cumulative training strategy which first trains the local implicit image function with large cell sizes, and finetunes it with the alternative training strategy to improve the performance on different upsampling scales. More quantitative results of our training strategies can be found in Section 4.4.", + "bbox": [ + 496, + 478, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "18260", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/1c1a785986547ec8de67eccb11a849f76cff301b4641887444182acfa3e80442.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Method×2×3×4×6×12×18×24×30
Bicubic [7]31.0128.2226.6624.8222.2721.0020.1919.59
EDSR-baseline [7]34.5530.9028.94-----
EDSR-baseline-Meta-SR [16, 17]34.6430.9328.9226.6123.5522.0321.0620.37
EDSR-baseline-LIIF [17]34.6730.9629.0026.7523.7122.1721.1820.48
EDSR-baseline-UltraSR [18]34.6931.0229.0526.8123.7522.2121.2120.51
EDSR-baseline-IPE [19]34.7231.0129.0426.7923.7522.2121.2220.51
EDSR-baseline-LTE [21]34.7231.0229.0426.8123.7822.2321.2420.53
EDSR-baseline-CLIT (Ours)34.8131.1229.1526.9223.8322.2921.2620.53
RDN-Meta-SR [16, 17]35.0031.2729.2526.8823.7322.1821.1720.47
RDN-LIIF [17]34.9931.2629.2726.9923.8922.3421.3120.59
RDN-UltraSR [18]35.0031.3029.3227.0323.7322.3621.3320.61
RDN-IPE [19]35.0431.3229.3227.0423.9322.3821.3420.63
RDN-LTE [21]35.0431.3229.3327.0423.9522.4021.3620.64
RDN-CLIT (Ours)35.1031.3929.3927.1224.0122.4521.3820.64
SwinIR-MetaSR [16, 21]35.1531.4029.3326.9423.8022.2621.2620.54
SwinIR-LIIF [17, 21]35.1731.4629.4627.1524.0222.4321.4020.67
SwinIR-LTE [21]35.2431.5029.5127.2024.0922.5021.4720.73
SwinIR-CLIT (Ours)35.2931.5529.5527.2624.1122.5121.4520.70
", + "bbox": [ + 102, + 88, + 872, + 391 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Table 1. The average PSNR (dB) on the DIV2K [48] validation set. The results are obtained from the original manuscripts [17-19, 21]. The best and second-best performing results are highlighted by the red and blue colors, respectively.", + "bbox": [ + 75, + 393, + 890, + 419 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Cascaded Local Implicit Transformer", + "text_level": 1, + "bbox": [ + 76, + 422, + 403, + 439 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Previous works [17, 21] typically address arbitrary-scale SR through a single step of upsampling. However, one-step upsampling struggles to reconstruct an HR image when the upsampling scale is large [5, 42, 43]. In light of this, we propose a cascaded upsampling strategy, called \"Cascaded LIT(CLIT)\" to predict residual images from multi-scale feature embeddings. In an $N$ -branched CLIT, the multi-scale feature embeddings $\\mathcal{Z}^1, \\mathcal{Z}^2, \\dots, \\mathcal{Z}^N$ are derived as follows:", + "bbox": [ + 75, + 449, + 470, + 569 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {Z} ^ {N} = \\mathcal {Z} \\uparrow_ {s ^ {1} \\times s ^ {2} \\times \\dots \\times s ^ {N - 1}}, \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 181, + 585, + 468, + 612 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $s^1 = 1$ and $s\\in \\mathbf{s}$", + "bbox": [ + 183, + 607, + 346, + 622 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\uparrow$ is a bilinear upsampling function and $s$ is a set of scaling factors, which are configurable hypermeters. For a branch $i,i\\in [1,\\dots,N]$ , $\\mathrm{LIT}^i$ estimates the residual image $I_r^i$ from the feature embedding $\\mathcal{Z}^i$ with the coordinate and the corresponding cells. Lastly, the final HR image $I^{HR}\\in \\mathbb{R}^{r_hH\\times r_wW\\times 3}$ can be estimated as the following equation:", + "bbox": [ + 76, + 638, + 468, + 729 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nI ^ {H R} = \\lambda^ {N - 1} I _ {r} ^ {1} + \\lambda^ {N - 2} I _ {r} ^ {2} + \\dots + \\lambda^ {0} I _ {r} ^ {N} + I _ {\\uparrow} ^ {H R}. \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 91, + 744, + 468, + 763 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\lambda$ is a discount factor, with a default value of 0.75. During the training phase, CLIT is trained using the proposed cumulative training strategy. Initially, $\\mathrm{LIT}^1$ is trained with the strategy employed by [17]. Subsequently, $\\mathrm{LIT}^1$ is finetuned and $\\mathrm{LIT}^2$ is initialized by applying the alternative training strategy. By incrementally incorporating LITs into CLIT, the performance is progressively enhanced. The details of the training strategies are specified in Section 4.", + "bbox": [ + 75, + 779, + 468, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experimental Results", + "text_level": 1, + "bbox": [ + 500, + 422, + 705, + 439 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this section, we present the experimental results and discuss their implications. We begin with a brief introduction to our experimental setup in Section 4.1. Following that, we evaluate our CLIT with different datasets in Section 4.2. The, Section 4.3 showcases the learned attention maps. Finally, ablation studies for various configurations of the proposed CLIT and LIT are compared in Section 4.4.", + "bbox": [ + 496, + 449, + 890, + 555 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Experimental Setup", + "text_level": 1, + "bbox": [ + 498, + 566, + 689, + 583 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Dataset. We use the DIV2K dataset [48] for network training. It consists of 1,000 images in 2K resolutions and provides low-resolution counterparts with down-sampling scales, $\\times 2$ , $\\times 3$ , $\\times 4$ , which are generated by the bicubic interpolation method. On the other hand, we also evaluate the performance on the validation set of DIV2K [48], Set5 [49], Set14 [50], B100 [51] and Urban100 [52] in terms of peak signal-to-noise (PSNR) values.", + "bbox": [ + 496, + 590, + 890, + 710 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Training details During training, we feed batches of size $48 \\times 48$ low-resolution images into the framework, following the prior works [7]. For each batch, a single upsampling scale is sampled from a uniform distribution $r \\sim \\mathcal{U}(1,4)$ . With single upsampling scales $s$ , a batch of HR images is cropped into patches of size $48r \\times 48r$ , while the corresponding LR images are cropped into patches of $48 \\times 48$ . The patches are augmented by randomly horizontal flipping, vertical flipping, and $90^{\\circ}$ rotating. Then, we sample $48^2$ pixels (coordinate-RGB pairs) on each HR patch as the ground-truths. We set the batch size to 32 and use", + "bbox": [ + 496, + 734, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "18261", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/c8f790d45fa284914bdd715039abf777e30d550a11dc804d42703976814f10aa.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodSet5Set14B100Urban100
×2×3×4×6×8×2×3×4×6×8×2×3×4×6×8×2×3×4×6×8
RDN [9]38.2434.7132.47--34.0130.5728.81--32.3429.2627.72--32.8928.8026.61--
RDN-Meta-SR [16, 17]38.2234.6332.3829.0426.9633.9830.5428.7826.5124.9732.3329.2627.7125.9024.8332.9228.8226.5523.9922.59
RDN-LLIF [17]38.1734.6832.5029.1527.1433.9730.5328.8026.6425.1532.3229.2627.7425.9824.9132.8728.8226.6824.2022.79
RDN-UltraSR [18]38.2134.6732.4929.3327.2433.9730.5928.8626.6925.2532.3529.2927.7726.0124.9632.9728.9226.7824.3022.87
RDN-IPE [19]38.1134.6832.5129.2527.2233.9430.4728.7526.5825.0932.3129.2827.7626.0024.9332.9728.8226.7624.2622.87
RDN-LTE [21]38.2334.7232.6129.3227.2634.0930.5828.8826.7125.1632.3629.3027.7726.0124.9533.0428.9726.8124.2822.88
RDN-CLIT (Ours)38.2634.7932.6929.5427.3434.0930.6928.9326.8325.3632.3929.3327.8026.0725.0033.1429.0526.9324.4423.04
SwinIR [14]38.3534.8932.72--34.1430.7728.94--32.4429.3727.83--33.4029.2927.07--
SwinIR-MetaSR [16, 21]38.2634.7732.4729.0927.0234.1430.6628.8526.5825.0932.3929.3127.7525.9424.8633.2929.1226.7624.1622.75
SwinIR-LLIF [17]38.2834.8732.7329.4627.3634.1430.7528.9826.8225.3432.3929.3427.8426.0725.0133.3629.3327.1524.5923.14
SwinIR-LTE [21]38.3334.8932.8129.5027.3534.2530.8029.0626.8625.4232.4429.3927.8626.0925.0333.5029.4127.2424.6223.17
SwinIR-CLIT (Ours)38.4134.9732.8629.6927.6234.2730.8529.0826.9425.5532.4629.4227.9126.1525.0933.5629.4327.2524.7723.33
", + "bbox": [ + 80, + 88, + 893, + 213 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 2. The average PSNR (dB) on Set5 [49], Set14 [50], B100 [51], and Urban100 [52]. The results are obtained from the original manuscripts [17-19, 21]. The best and second-best performing results are highlighted by the red and blue colors, respectively.", + "bbox": [ + 76, + 215, + 890, + 239 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/6e0bd67caa71965e7c5090c88a45e21384e9a979558071c4cab1b76adb97b29a.jpg", + "image_caption": [ + "Figure 5. The qualitative results of LIIF [17], LTE [21], and our CLIT with using RDN [9] as the encoder." + ], + "image_footnote": [], + "bbox": [ + 130, + 244, + 838, + 683 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "the Adam optimizer [53] together with L1 loss for training. We train LIT for 1,000 epochs, and the learning rate is initialized at $1e^{-4}$ and decayed by a factor 0.5 at [200, 400, 600, 800] epochs. For cumulative training of CLIT, as discussed in Section 3.4, we sample $N$ scale factors $\\{s^1, s^2, \\dots, s^N\\}$ from the distribution $\\mathcal{U}(1,4)$ according the number of LITs $N$ in the train step. The total upsampling scale $r = s^1 \\times s^2 \\dots \\times s^N$ is the product of all scale factors. If $48r \\times 48r$ is greater than the whole HR image, we clip the scale factor of stage 1. To train $N$ LITs, we fine-tune the model for $500 \\times N$ epochs, and the learning rate is initialized at $1e^{-4}$ and decayed by a factor 0.5 at $[100 \\times N, 200 \\times N, 300 \\times N, 400 \\times N]$ epochs.", + "bbox": [ + 75, + 703, + 470, + 898 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Validation of CLIT", + "text_level": 1, + "bbox": [ + 500, + 702, + 683, + 715 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Quantitative results. We first compare our proposed CLIT to existing local implicit neural representation methods for arbitrary-scale SR, including LIIF [17], UltraSR [18], IPE [19], and LTE [21]. Table 1 summarizes the quantitative results in terms of PSNR(dB) on DIV2k [48]. As shown in Table 1, CLIT achieves the best performance when EDSR-Baseline [7], RDN [9] are used as the encoders. These results demonstrate the advantage of CLIT.", + "bbox": [ + 496, + 729, + 890, + 851 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 2 compares CLIT to prior works [17-19, 21] on widely-used datasets, including Set5 [49], Set14 [50], B100 [51], and Urban100 [52], with RDN and SwinIR.", + "bbox": [ + 498, + 854, + 892, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "18262", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/ebe25e5c577baf8ff4da43660cc0c4fc9ae38ee8c4d5fe45500e97fbf15af4e7.jpg", + "image_caption": [ + "Figure 6. The qualitative results of LIIF [17], LTE [21], and our CLIT using RDN [9] as the encoder and non-integer upsampling scales." + ], + "image_footnote": [], + "bbox": [ + 117, + 88, + 854, + 431 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Note that RDN [9] is trained and evaluated for specific upsampling scales. CLIT outperforms those existing methods in most cases across all datasets and scales, even achieving a 0.26dB PSNR improvement on Set5 for the $\\times 8$ scale.", + "bbox": [ + 75, + 450, + 470, + 513 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Qualitative results. Fig. 5 compares the qualitative results of CLIT with the baseline methods LIIF [17] and LTE [21], on various dataset DIV2k [48], Set14 [50], B100 [51], and Urban100 [52] with variant upsampling scales. The official codes provided by them were used to produce the results. In the first row, the SR visualization results with a $\\times 12$ upsampling scale are depicted. It can be observed that both LIIF and LTE struggle to reconstruct the letters continuously, whereas the CLIT result demonstrates continuity for these alphabets, particularly for 'M' and 'S'. In the second row, the stripes on the zebra appear blurry in the LIIF and LTE results, whereas they are more distinct in the CLIT result. In the third row, the cross pattern on the tablecloth is not clearly rendered by LIIF and LTE. In contrast, CLIT is able to successfully generate clean crosses and sharp lines. In the fourth row, despite the blurriness of the original LR image, CLIT captures the texture and produces straight lines on the flag, showcasing its effectiveness.", + "bbox": [ + 75, + 522, + 468, + 792 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Fig. 6 presents the results of text image enhancement employing progressively increasing non-integer upsampling factors, alongside a comparison with earlier methods, LIIF [17] and LTE [21]. The input image first undergoes a downscaling process by a factor of 2.2 to generate an LR image. Subsequently, the LR image is then upscaled using predetermined non-integer factors $\\times 1.6$ , $\\times 2.5$ , $\\times 3.4$ , $\\times 4.3$ .", + "bbox": [ + 75, + 795, + 470, + 902 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b8439088d65aa6981d1b276c707e5fe6fdc5623441f9c4b85d08480d415e21a2.jpg", + "image_caption": [ + "Figure 7. A visualization of the local attention maps of the coordinates highlight as red dots." + ], + "image_footnote": [], + "bbox": [ + 501, + 449, + 890, + 684 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "It can be observed that our proposed model effectively captures the patterns of text within the image, and accurately estimates words and numbers with enhanced sharpness and clarity. The enhancement becomes particularly evident in the word 'infinity' in the first row of contexts, 'SUNSHINE' in the fifth row, and the number '0' in the final row, as each of these elements demonstrates considerable improvements.", + "bbox": [ + 496, + 717, + 890, + 824 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Visualization Local Attention Maps", + "text_level": 1, + "bbox": [ + 500, + 832, + 808, + 849 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In Fig. 7, we provide visualizations of local attention maps. The LR images are generated by applying bicubic downsampling to HR images with scale factors", + "bbox": [ + 496, + 854, + 890, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "18263", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/e39d5472bde49fccf2d1db39d07238c5bc0af4080b256a344ffd3c8eba63838d.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
In-distributionOut-of-distribution
×2×3×4×6×8×12
LIT34.7931.0729.1026.8425.4723.77
LIT (-a)34.7131.0029.0226.7625.4023.70
LIT (-f)34.7831.0729.0926.8325.4623.77
LIT (-c)34.6531.0029.0426.8025.4523.77
LIT (+e)34.7931.0829.1126.8525.4823.78
", + "bbox": [ + 81, + 88, + 468, + 183 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "$\\{ \\times 4.5, \\times 6, \\times 6 \\}$ . Subsequently, the proposed CLIT with RDN [9] encoder is utilized to produce HR predictions. It can be observed that the attention maps closely align with the edges, indicating that the cross-scale local attention block effectively captures similar latent codes within a local area. This substantiates that our proposed design enables the generation of SR images from LR counterparts, ultimately producing clean and sharp edges in the output.", + "bbox": [ + 75, + 239, + 470, + 361 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4. Ablation Studies", + "text_level": 1, + "bbox": [ + 76, + 369, + 240, + 383 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this section, we present a series of ablation analyses to substantiate the design decisions proposed in this paper. All of the ablation experiments are conducted on the DIV2K [48] validation set, utilizing the EDSR-baseline [7] as the encoder and employing the PSNR metric for evaluation.", + "bbox": [ + 75, + 392, + 468, + 468 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Validation of the design choices. Table 3 presents a summary of the quantitative contributions associated with each LIT component. A significant improvement is observed when adopting the cross-scale local attention block, while a relatively minor gain with the frequency encoding block by comparing LTE with LTE (-a) and LTE (-f), respectively. In addition, LTE (-c) reveals that removing cell decoding leads to more severe performance degradation for in-distribution upsampling scales compared to out-of-distribution counterparts, highlighting the importance of cell decoding for in-distribution upsampling scales. Finally, incorporating the local ensemble technique in LIT (+e) results in a further, albeit modest, enhancement of the overall performance.", + "bbox": [ + 75, + 470, + 468, + 667 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The effectiveness of the local grid. Table 4 provides a performance comparison of various local grid sizes employed within the local coordinate sampling scheme, as discussed in Section 3. The results presented in Table 4 reveal that increasing the local grid size contributes to performance enhancements, albeit at the cost of extended training times. As a result, this study adopts a $7 \\times 7$ local grid size to strike an ideal balance between effectiveness and efficiency.", + "bbox": [ + 75, + 670, + 468, + 792 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Analysis on the training strategy. Table 5 presents a quantitative comparison of the proposed cumulative training strategy with the other training strategies for training an LIT. The baseline strategy trains LIT using upsampling scales sampled from a uniform distribution, $r \\sim \\mathcal{U}(1,4)$ . Expanding the sampling scale distribution to $r \\sim \\mathcal{U}(1,12)$ improves the performance of large-scale upsampling while", + "bbox": [ + 75, + 795, + 468, + 902 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/2066e509767a63192a25f6193caaf3b3832a55f5940af776f7cac64b5e97f7cc.jpg", + "table_caption": [ + "Table 3. The average PSNR (dB) for different design choices. The notations -a/f/c denote the removal of the cross-scale local attention block, frequency encoding block, and cell size, respectively, and +e denotes the addition of local ensemble. The best performing one is indicated in bolded." + ], + "table_footnote": [], + "table_body": "
Local gird\nGh × GwIn-distributionOut-of-distributionTraining time
×2×3×4×6×8×12
1 × 134.7131.0029.0226.7625.4023.7020.9
3 × 334.7631.0429.0726.8125.4623.7623.2
5 × 534.7831.0729.0826.8325.4723.7828.4
7 × 734.7931.0729.1026.8425.4723.7733.6
9 × 934.7931.0629.1026.8425.4723.7753.0
15 × 1534.8331.1029.1126.8625.4923.7876.1
25 × 2534.8631.1329.1526.8825.5023.80190.9
", + "bbox": [ + 504, + 87, + 893, + 181 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/a7d9beeca4edb2adb023d68b58b181c25f4d46f7d7eaff77c4e839c450e64ca3.jpg", + "table_caption": [ + "Table 4. The average PSNR (dB) and the training time (hour) of multiple local grid sizes. The bolded numbers correspond to the best performance." + ], + "table_footnote": [], + "table_body": "
Training strategy×2×3×4×6×8×12
Training with r ~ U(1, 4)34.7931.0729.1026.8425.4723.77
Training with r ~ U(1, 12)34.6931.0629.1126.8725.5223.82
Alternative training strategy34.7531.0929.1226.8825.5223.81
Cumulative training strategy34.7831.1029.1426.8925.5423.83
", + "bbox": [ + 503, + 212, + 890, + 266 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 5. The average PSNR (dB) of different training strategies. The best performing results are indicated in **bold**.", + "bbox": [ + 498, + 268, + 890, + 292 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "compromising the performance of small-scale upsampling. To achieve high-quality results across all upsampling scale ranges, the alternative training strategy trains LIT alternatively by switching the sampling scales between $r \\sim \\mathcal{U}(1,4)$ and $r \\sim \\mathcal{U}(4,12)$ . On the other hand, the proposed cumulative training strategy first trains LIT with $r \\sim \\mathcal{U}(1,4)$ , followed by fine-tuning using the alternative training strategy. This progressive approach improves the performance of LIT by gradually broadening the distribution of the upsampling scales. Comparing the results of the cumulatively trained LIT with the results of EDSR-baseline-CLIT in Table 1 reveals that the proposed cumulative training strategy works for both LIT and CLIT. Moreover, the results from these tables suggest that the cascaded framework is more effective in addressing arbitrary-scale SR tasks.", + "bbox": [ + 496, + 296, + 890, + 523 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 526, + 617, + 541 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we introduced the attention mechanism and frequency encoding technique to address arbitrary-scale SR tasks. To achieve this objective, we proposed the LIT framework, which consists of a cross-scale local attention block and a local frequency encoding block. The former is designed to find the local latent embedding to reconstruct the corresponding RGB value, while the latter focuses on encoding coordinates with the frequency information derived from the feature embeddings. To enhance the capability of capturing fine details, we proposed CLIT and the corresponding cumulative training strategy that trains the model with progressively increasing upsampling scales. Based on the experimental results, both quantitative and qualitative assessments showcase the superior performance of LIT and CLIT in comparison to existing methods. Furthermore, our comprehensive analyses validated the effectiveness of the training strategies and the components employed by them.", + "bbox": [ + 496, + 545, + 890, + 801 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 500, + 806, + 666, + 821 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The authors gratefully acknowledge the support from the National Science and Technology Council (NSTC) in Taiwan under grant numbers MOST 111-2223-E-007-004-MY3 and MOST 111-2628-E-007-010, as well as the financial and technical support from MediaTek Inc., Taiwan.", + "bbox": [ + 496, + 825, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "18264", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] C. Dong, C. C. Loy, K. He, and X. Tang. Image superresolution using deep convolutional networks. IEEE Trans. Pattern Analysis and Machine Intelligence (TPAMI), 38(2): 295-307, 2016. 1, 2", + "[2] J. Kim, J. K. Lee, and K. M. Lee. Accurate image superresolution using very deep convolutional networks. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), pages 1646-1654, 2016.", + "[3] C. Dong, C. C. Loy, and X. Tang. Accelerating the superresolution convolutional neural network. In Proc. European Conf. on Computer Vision (ECCV), pages 391-407, 2016. 1", + "[4] W. Shi, J. Caballero, F. Huszar, J. Totz, A. P. Aitken, et al. Real-time single image and video super-resolution using an efficient sub-pixel convolutional neural network. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), pages 1874-1883, 2016. 1", + "[5] W. Lai, J. Huang, N. Ahuja, and M. Yang. Fast and accurate image super-resolution with deep laplacian pyramid networks. IEEE Trans. Pattern Analysis and Machine Intelligence (TPAMI), 41(11):2599-2613, 2019. 2, 5", + "[6] C. Ledig, L. Theis, F. Huszar, J. Caballero, A. Cunningham, et al. Photo-realistic single image super-resolution using a generative adversarial network. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), pages 105-114, 2017. 2", + "[7] B. Lim, S. Son, H. Kim, S. Nah, and K. M. Lee. Enhanced deep residual networks for single image super-resolution. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition Workshop (CVPRW), pages 1132-1140, 2017. 2, 3, 5, 6, 8", + "[8] X. Wang, K. Yu, S. Wu, J. Gu, Y. Liu, et al. Esrgan: Enhanced super-resolution generative adversarial networks. In Proc. European Conf. on Computer Vision Workshop (ECCVW), pages 63-79, 2018. 2", + "[9] Y. Zhang, Y. Tian, Y. Kong, B. Zhong, and Y. Fu. Residual dense network for image super-resolution. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), pages 2472-2481, 2018. 2, 3, 6, 7, 8", + "[10] Y. Zhang, K. Li, K. Li, L. Wang, B. Zhong, et al. Image super-resolution using very deep residual channel attention networks. In Proc. European Conf. on Computer Vision (ECCV), pages 294-310, 2018. 2", + "[11] T. Dai, J. Cai, Y. Zhang, S. Xia, and L. Zhang. Second-order attention network for single image super-resolution. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), pages 11065-11074, 2019.", + "[12] Y. Mei, Y. Fan, Y. Zhou, L. Huang, T. S. Huang, et al. Image super-resolution with cross-scale non-local attention and exhaustive self-exemplars mining. In Proc. IEEE Conf." + ], + "bbox": [ + 78, + 114, + 470, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "on Computer Vision and Pattern Recognition (CVPR), pages 5689-5698, 2020. 2", + "[13] H. Chen, Y. Wang, T. Guo, C. Xu, Y. Deng, et al. Pre-trained image processing transformer. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), pages 12294-12305, 2021. 2", + "[14] J. Liang, J. Cao, G. Sun, K. Zhang, L. Van Gool, et al. Swinir: Image restoration using swim transformer. In Proc. IEEE Int. Conf. on Computer Vision Workshop (ICCVW), pages 1833-1844, 2021. 3, 6", + "[15] X. Chen, X. Wang, J. Zhou, and C. Dong. Activating more pixels in image super-resolution transformer. CoRR, abs/2205.04437, 2022. 1, 2", + "[16] X. Hu, H. Mu, X. Zhang, Z. Wang, T. Tan, et al. Meta-sr: A magnification-arbitrary network for super-resolution. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), pages 1575-1584, 2019. 1, 2, 5, 6", + "[17] Y. Chen, S. Liu, and X. Wang. Learning continuous image representation with local implicit image function. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), pages 8628-8638, 2021. 1, 2, 4, 5, 6, 7", + "[18] X. Xu, Z. Wang, and H. Shi. Ultrasr: Spatial encoding is a missing key for implicit image function-based arbitrary-scale super-resolution. CoRR, abs/2103.12716, 2021. 2, 5, 6", + "[19] Y. Liu, Y. Guo, and S. Zhang. Enhancing multi-scale implicit learning in image super-resolution with integrated positional encoding. CoRR, abs/2112.05756, 2021. 2, 4, 5, 6", + "[20] J. Yang, S. Shen, H. Yue, and K. Li. Implicit transformer network for screen content image continuous super-resolution. In Proc. Conf. on Neural Information Processing Systems (NeurIPS), pages 13304-13315, 2021.", + "[21] J. Lee and K. H. Jin. Local texture estimator for implicit representation function. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), pages 1929-1938, 2022. 1, 2, 4, 5, 6, 7", + "[22] A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, et al. Attention is all you need. In Proc. Conf. on Neural Information Processing Systems (NeurIPS), pages 5998-6008, 2017. 1", + "[23] Z. Liu, Y. Lin, Y. Cao, H. Hu, Y. Wei, et al. Swin transformer: Hierarchical vision transformer using shifted windows. In Proc. IEEE Int. Conf. on Computer Vision (ICCV), 2021. 2", + "[24] W. Wang, L. Yao, L. Chen, B. Lin, D. Cai, et al. Crossformer: A versatile vision transformer hinging on cross-scale attention. In Proc. Int. Conf. on Learning Representations (ICLR), 2022. 2", + "[25] B. Mildenhall, P. P. Srinivasan, M. Tancik, J. T. Barron, R. Ramamoorthi, et al. Nerf: Representing scenes as neural radiance fields for view synthesis. In Proc. European Conf. on Computer Vision (ECCV), pages 405-421, 2020. 2" + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "18265", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[26] M. Tancik, P. P. Srinivasan, B. Mildenhall, S. Fridovich-Keil, N. Raghavan, et al. Fourier features let networks learn high frequency functions in low dimensional domains. In Proc. Conf. on Neural Information Processing Systems (NeurIPS), pages 7537–7547, 2020. 2", + "[27] N. Rahaman, A. Baratin, D. Arpit, F. Draxler, M. Lin, et al. On the spectral bias of neural networks. In Proc. Int. Conf. on Machine Learning (ICML), pages 5301-5310, 2019. 2", + "[28] Z. Chen and H. Zhang. Learning implicit fields for generative shape modeling. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), pages 5939-5948, 2019. 2", + "[29] J. J. Park, P. Florence, J. Straub, R. A. Newcombe, and S. Lovegrove. Deepsdf: Learning continuous signed distance functions for shape representation. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), pages 165-174, 2019.", + "[30] L. M. Mescheder, M. Oechsle, M. Niemeyer, S. Nowozin, and A. Geiger. Occupancy networks: Learning 3d reconstruction in function space. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), pages 4460-4470, 2019.", + "[31] K. Genova, F. Cole, A. Sud, A. Sarna, and T. A. Funkhouser. Local deep implicit functions for 3d shape. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), pages 4856-4865, 2020.", + "[32] M. Atzmon and Y. Lipman. SAL: sign agnostic learning of shapes from raw data. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), pages 2562-2571, 2020. 2", + "[33] V. Sitzmann, M. Zollhöfer, and G. Wetzstein. Scene representation networks: Continuous 3d-structure-aware neural scene representations. In Proc. Conf. on Neural Information Processing Systems (NeurIPS), pages 1119–1130, 2019. 2", + "[34] C. M. Jiang, A. Sud, A. Makadia, J. Huang, M. Nießner, et al. Local implicit grid representations for 3d scenes. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), pages 6000-6009, 2020.", + "[35] R. Chabra, J. E. Lenssen, E. Ilg, T. Schmidt, J. Straub, et al. Deep local shapes: Learning local SDF priors for detailed 3d reconstruction. In Proc. European Conf. on Computer Vision (ECCV), pages 608-625, 2020.", + "[36] S. Peng, M. Niemeyer, L. M. Mescheder, M. Pollefeys, and A. Geiger. Convolutional occupancy networks. In Proc. European Conf. on Computer Vision (ECCV), pages 523-540, 2020. 2", + "[37] M. Niemeyer, L. Mescheder, M. Oechsle, and A. Geiger. Differentiable volumetric rendering: Learning implicit 3d representations without 3d supervision. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), pages 3501-3512, 2020. 2" + ], + "bbox": [ + 78, + 90, + 468, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[38] L. Liu, J. Gu, K. Z. Lin, T. Chua, and C. Theobalt. Neural sparse voxel fields. In Proc. Conf. on Neural Information Processing Systems (NeurIPS), 2020.", + "[39] J. T. Barron, B. Mildenhall, M. Tancik, P. Hedman, R. Martin-Brualla, et al. Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In Proc. IEEE Int. Conf. on Computer Vision (ICCV), pages 5835–5844, 2021. 2", + "[40] S. Klocek, L. Maziarka, M. Wolczyk, J. Tabor, J. Nowak, et al. Hypernetwork functional image representation. In Artificial Neural Networks and Machine Learning, pages 496-510, 2019. 2", + "[41] V. Sitzmann, J. N. P. Martel, A. W. Bergman, D. B. Lindell, and G. Wetzstein. Implicit neural representations with periodic activation functions. In Proc. Conf. on Neural Information Processing Systems (NeurIPS), pages 7462-7473, 2020. 2", + "[42] K. Grm, W. J. Scheirer, and V. Struc. Face hallucination using cascaded super-resolution and identity priors. IEEE Trans. Image Processing, pages 2150-2165, 2020. 2, 5", + "[43] Z. Fu, Y. Kong, Y. Zheng, H. Ye, W. Hu, et al. Cascaded detail-preserving networks for super-resolution of document images. In Int. Conf. on Doc. Analysis Recognition, ICDAR, pages 240-245, 2019. 2, 5", + "[44] R. Basri, M. Galun, A. Geifman, D. W. Jacobs, Y. Kasten, et al. Frequency bias in neural networks for input of nonuniform density. In Proc. Int. Conf. on Machine Learning (ICML), pages 685-694, 2020. 2", + "[45] M. Tancik, B. Mildenhall, T. Wang, D. Schmidt, P. P. Srinivasan, et al. Learned initializations for optimizing coordinate-based neural representations. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), pages 2846-2855, 2021. 2", + "[46] D. Hendrycks and K. Gimpel. Bridging nonlinearities and stochastic regularizers with gaussian error linear units. CoRR, abs/1606.08415, 2016. 4", + "[47] J. Lee-Thorp, J. Ainslie, I. Eckstein, and S. Ontanon. FNet: Mixing tokens with fourier transforms. In Proc. Conf. of North American Chapter of the Association for Computational Linguistics: Human Language Technologies (NAACL), pages 4296-4313, 2022. 4", + "[48] E. Agustsson and R. Timofte. NTIRE 2017 challenge on single image super-resolution: Dataset and study. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition Workshop (CVPRW), pages 1122-1131, 2017. 5, 6, 7, 8", + "[49] M. Bevilacqua, A. Roumy, C. Guillemot, and M. Alberi-Morel. Low-complexity single-image super-resolution based on nonnegative neighbor embedding. In Proc. British Machine Vision Conf. (BMVC), pages 1-10, 2012. 5, 6" + ], + "bbox": [ + 501, + 90, + 890, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "18266", + "bbox": [ + 480, + 945, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[50] R. Zeyde, M. Elad, and M. Protter. On single image scale-up using sparse-representations. In *Curves and Surfaces*, volume 6920 of Lecture Notes in Computer Science, pages 711-730, 2010. 5, 6, 7", + "[51] D. R. Martin, C. C. Fowlkes, D. Tal, and J. Malik. A database of human segmented natural images and its application to evaluating segmentation algorithms and measuring ecological statistics. In Proc. IEEE Int. Conf. on Computer Vision (ICCV), pages 416-425, 2001. 5, 6, 7", + "[52] J. Huang, A. Singh, and N. Ahuja. Single image superresolution from transformed self-exemplars. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), pages 5197-5206, 2015. 5, 6, 7", + "[53] D. P. Kingma and J. Ba. Adam: A method for stochastic optimization. In Proc. Int. Conf. on Learning Representations (ICLR), 2015. 6" + ], + "bbox": [ + 78, + 90, + 468, + 343 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "18267", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2023/Cascaded Local Implicit Transformer for Arbitrary-Scale Super-Resolution/ae047e51-87a0-4eec-807f-144baf2a48b8_model.json b/2023/Cascaded Local Implicit Transformer for Arbitrary-Scale Super-Resolution/ae047e51-87a0-4eec-807f-144baf2a48b8_model.json new file mode 100644 index 0000000000000000000000000000000000000000..42a0ec1d03657ef420bb6ad211c9af1ab03be8c7 --- /dev/null +++ b/2023/Cascaded Local Implicit Transformer for Arbitrary-Scale Super-Resolution/ae047e51-87a0-4eec-807f-144baf2a48b8_model.json @@ -0,0 +1,2125 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.808, + 0.02 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation." + }, + { + "type": "header", + "bbox": [ + 0.325, + 0.017, + 0.725, + 0.031 + ], + "angle": 0, + "content": "Except for this watermark, it is identical to the accepted version;" + }, + { + "type": "header", + "bbox": [ + 0.294, + 0.032, + 0.754, + 0.046 + ], + "angle": 0, + "content": "the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.104, + 0.064, + 0.868, + 0.086 + ], + "angle": 0, + "content": " Cascaded Local Implicit Transformer for Arbitrary-Scale Super-Resolution" + }, + { + "type": "text", + "bbox": [ + 0.087, + 0.095, + 0.879, + 0.136 + ], + "angle": 0, + "content": "Hao-Wei Chen\\*,1,2 Yu-Syuan Xu\\*,2 Min-Fong Hong\\* Yi-Min Tsai\\* Hsien-Kai Kuo\\* Chun-Yi Lee\\* Elsa Lab, National Tsing Hua University \\*MediaTek Inc." + }, + { + "type": "text", + "bbox": [ + 0.318, + 0.139, + 0.656, + 0.154 + ], + "angle": 0, + "content": "{jaroslaw1007, cylee}@gapp.nthu.edu.tw" + }, + { + "type": "text", + "bbox": [ + 0.184, + 0.156, + 0.785, + 0.171 + ], + "angle": 0, + "content": "{Yu-Syuan.Xu, romulus.hong, Yi-Min.Tsai, Hsienkai.Kuo}@mediatek.com" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.221, + 0.314, + 0.237 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.074, + 0.244, + 0.474, + 0.518 + ], + "angle": 0, + "content": "Implicit neural representation has recently shown a promising ability in representing images with arbitrary resolutions. In this paper, we present a Local Implicit Transformer (LIT), which integrates the attention mechanism and frequency encoding technique into a local implicit image function. We design a cross-scale local attention block to effectively aggregate local features and a local frequency encoding block to combine positional encoding with Fourier domain information for constructing high-resolution images. To further improve representative power, we propose a Cascaded LIT (CLIT) that exploits multi-scale features, along with a cumulative training strategy that gradually increases the upsampling scales during training. We have conducted extensive experiments to validate the effectiveness of these components and analyze various training strategies. The qualitative and quantitative results demonstrate that LIT and CLIT achieve favorable results and outperform the prior works in arbitrary super-resolution tasks." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.529, + 0.21, + 0.544 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.547, + 0.47, + 0.834 + ], + "angle": 0, + "content": "Single Image Super-Resolution (SISR) is the process of reconstructing high-resolution (HR) images from their corresponding low-resolution (LR) counterparts. SISR has long been recognized as a challenging task in the low-level vision domain due to its ill-posed nature, and has attracted a number of researchers dedicated to this field of study over the past decade [1-21]. A line of SISR research referred to as 'fixed-scale SR' [1-15] focuses on extracting feature embeddings from LR images and leveraging these embeddings to upsample images with a predefined factor through learnable deconvolutions [3] or sub-pixel convolutions [4]. Despite their success, many of the proposed approaches necessitate a distinct deep neural network model for each upsampling scale, which is usually restricted to a limited selection of integers (e.g., \\( 2 \\times \\), \\( 3 \\times \\), \\( 4 \\times \\)). Such a limitation constrains the potential applications and deployment options of SISR models. To overcome this limitation, approaches for upsampling LR images in a continuous manner via a single model emerge and attracted considerable attention recently." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.834, + 0.47, + 0.88 + ], + "angle": 0, + "content": "Over the past few years, arbitrary-scale SR has emerged and attracted considerable attention from researchers [16-21]. Apart from the pioneering work Meta-SR [16], recent" + }, + { + "type": "image_caption", + "bbox": [ + 0.501, + 0.301, + 0.518, + 0.312 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.523, + 0.204, + 0.692, + 0.263 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.556, + 0.263, + 0.569, + 0.271 + ], + "angle": 0, + "content": "LR" + }, + { + "type": "image_caption", + "bbox": [ + 0.642, + 0.263, + 0.659, + 0.271 + ], + "angle": 0, + "content": "HR" + }, + { + "type": "image", + "bbox": [ + 0.715, + 0.204, + 0.793, + 0.263 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.746, + 0.263, + 0.761, + 0.271 + ], + "angle": 0, + "content": "LR" + }, + { + "type": "image", + "bbox": [ + 0.802, + 0.205, + 0.88, + 0.264 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.834, + 0.264, + 0.849, + 0.271 + ], + "angle": 0, + "content": "HR" + }, + { + "type": "image", + "bbox": [ + 0.526, + 0.275, + 0.692, + 0.336 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.553, + 0.336, + 0.571, + 0.343 + ], + "angle": 0, + "content": "LIIF" + }, + { + "type": "image", + "bbox": [ + 0.715, + 0.275, + 0.88, + 0.336 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.746, + 0.336, + 0.761, + 0.343 + ], + "angle": 0, + "content": "LIIF" + }, + { + "type": "image", + "bbox": [ + 0.526, + 0.342, + 0.699, + 0.406 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.617, + 0.336, + 0.634, + 0.343 + ], + "angle": 0, + "content": "Loc" + }, + { + "type": "image", + "bbox": [ + 0.715, + 0.342, + 0.793, + 0.406 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.744, + 0.402, + 0.766, + 0.41 + ], + "angle": 0, + "content": "Ours" + }, + { + "type": "image", + "bbox": [ + 0.805, + 0.342, + 0.89, + 0.408 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.412, + 0.892, + 0.45 + ], + "angle": 0, + "content": "Figure 1. An illustration and comparison of different approaches that take into account nearby pixels for continuous upsampling: (a) the local ensemble method used in [17], and (b) our proposed local attention mechanism." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.459, + 0.893, + 0.807 + ], + "angle": 0, + "content": "endevavors [17-21] have achieved arbitrary-scale SR by replacing the upsampling layers commonly adopted by previous approaches with local implicit image functions, and have demonstrated favorable performance. These local implicit functions employ multi-layer perceptrons (MLPs) to map 2D coordinates and corresponding latent representations to RGB values. Fig. 1 illustrates how different approaches sample latent representations based on the queried coordinates (depicted as the red dots). Fig. 1 (a) illustrates the local ensemble technique adopted by contemporary mainstream methods [17-21]. It calculates the RGB value of the queried coordinate by taking the weighted average of those of the surrounding four pixels based on their relative distances to the queried coordinate. This approach, however, does not consider contextual information and relies solely on distance. For instance, in Fig. 1, the queried coordinates are intentionally designed to lie on edges. However, merely calculating the weighted average of pixels fails to reflect the contextual information about the image content, thereby preventing the accurate capture of the necessary features for performing SR. As a result, although pixel distance plays a vital role in SR tasks, it is essential to concentrate more on the contextual information in an image." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.81, + 0.894, + 0.902 + ], + "angle": 0, + "content": "In light of the above observations, we propose a Local Implicit Transformer (LIT), which expands the numbers of referenced latent vectors and accounts for the feature correlation in the context by exploiting the attention mechanism [22]. LIT comprises a Cross-Scale Local Attention Block (CSLAB), a Local Frequency Encoding Block" + }, + { + "type": "page_footnote", + "bbox": [ + 0.102, + 0.888, + 0.21, + 0.9 + ], + "angle": 0, + "content": "*Equal contribution" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18257" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.091, + 0.473, + 0.41 + ], + "angle": 0, + "content": "(LFEB), and a decoder. CSLAB generates attention maps based on the bilinearly interpolated latent vectors at the queried coordinates and the key latent vectors sampled from a grid of coordinates with relative positional bias [23, 24]. The first and second columns of Fig. 1 (b) visualize the attention maps generated by LIT, where the attention areas align closely with the edges. By applying attention maps to feature embeddings, the RGB values of the queried coordinates can be contextually predicted. Moreover, inspired by [21, 25, 26], we introduce LFEB, which projects relative coordinates into latent space to address the spectral bias problem [27] of an implicit neural function. Specifically, relative coordinates are encoded into relative positional encoding, which are multiplied with the frequency encoding extracted from the feature embedding in the Fourier domain to generate the frequency embeddings. This design enables the frequency embedding to integrate the relative positional encoding with texture information, thereby augmenting the expressivity of relative coordinates. Finally, a decoder is adopted to produce RGB values by taking advantage of the attention feature embedding and the frequency embedding." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.416, + 0.473, + 0.675 + ], + "angle": 0, + "content": "In order to address the issue of diverse scaling factors and achieve arbitrary-scale super-resolution, it is crucial to consider the role of upsampling factors in constructing high-resolution images within the local implicit image function. However, simultaneously training a local implicit image function with a wide range of upsampling factors (e.g., \\(1 \\times \\sim 30 \\times\\)) poses significant challenges. As a result, we propose a cumulative training strategy to incrementally enhance the fuction's its representative power. The strategy initially trains the local implicit image function with small upsampling factors and then finetunes it with alternatively sampled small and large ones. Furthermore, we present Cascaded LIT (CLIT) to harness the advantages of multiscale feature embeddings, complementing missing details and information during one-step upsampling. The combination of the cumulative training strategy and CLIT enables efficient and effective handling of arbitrary-scale SR tasks." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.681, + 0.473, + 0.834 + ], + "angle": 0, + "content": "The main contributions of our work are summarized as follows: (1) We introduce the LIT architecture, which incorporates the local attention mechanism into arbitrary-scale SR (2) We further develop a cumulative training strategy and the cascaded framework CLIT to effectively handle large-scale upsampling. (3) We carry out comprehensive analyses of the performance impacts for LIT and CLIT. The extensive experimental findings demonstrate that the proposed LIT and CLIT are able to yield remarkable or comparable results across a wide range of benchmark datasets." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.84, + 0.473, + 0.903 + ], + "angle": 0, + "content": "The paper is organized as follows. Section 2 reviews the related work. Section 3 walks through the proposed LIT and CLIT frameworks and the implementation details. Section 4 presents the experimental results. Section 5 concludes." + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.09, + 0.642, + 0.107 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.115, + 0.895, + 0.375 + ], + "angle": 0, + "content": "Implicit neural representation. Implicit neural representation is a technique for representing continuous-domain signals via coordinate-based multi-layer perceptrons (MLPs). Its concept has been adopted in various 3D tasks, e.g., 3D object shape modeling [28-32], 3D scene reconstruction [33-36], and 3D structure rendering [25, 37-39]. For example, NeRF [25] employs implicit neural representation to perform novel view synthesis, which maps coordinates to RGB colors for a specific scene. In the past few years, 2D applications of implicit neural representation have been attempted as well, such as image representation [40, 41] and super-resolution [17-21]. Our work is related to a technique called 'local implicit neural representation' [17, 21], which encodes LR images to feature embeddings such that similar information could be shared within local regions. Such local implicit neural representations are exploited to upscale LR images to HR ones." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.388, + 0.895, + 0.571 + ], + "angle": 0, + "content": "Single image super-resolution. In the past several years, various deep neural network (DNN) based architectures [1-15] have been proposed for SISR. Among these works, SR-CNN [1] pioneered the use of convolutional neural networks (CNNs) to achieve SISR in an end-to-end manner. It is later followed by several subsequent works that incorporated more complicated model architectures, such as residual blocks [6, 7], dense connections [8, 9], attention based mechanisms [10-12], or cascaded frameworks [5, 42, 43], to extract more effective feature representations for SISR. Recently, transformer-based methods [13-15] were introduced to SISR and achieved promising performance." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.584, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Arbitrary-scale super-resolution. As discussed in Section 1, most of the contemporary SISR works limit their upsampling scales to specific integer values, and are required to train a distinct model for each upsampling scale. To overcome such a limitation, several approaches [16-21] were proposed to train a unified model for arbitrary upsampling scales. Meta-SR [16] proposed a meta-upscale module for predicting the weights of their convolutional filters from coordinates and scales. The predicted weights are then utilized to perform convolutions to generate HR images. In contrast to Meta-SR, LIIF [17] employs an MLP as a local implicit function, which takes a queried coordinate in an HR image, its nearby feature representations extracted from the corresponding LR image, as well as a cell size to predict an RGB value for that coordinate. UltraSR [18] and IPE [19] extended LIIF by replacing coordinates with the embedded ones to deal with the spectral bias issue [25, 27, 41, 44, 45] inherent in MLPs. LTE [21] further introduced a local texture estimator that transforms coordinates into Fourier domain information to enrich the representational capability of its local implicit function. Different from the above ap" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "18258" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.084, + 0.09, + 0.891, + 0.326 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.145, + 0.327, + 0.822, + 0.34 + ], + "angle": 0, + "content": "Figure 2. The proposed LIT framework. The local sampling operation samples input embeddings based on a grid of coordinates." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.358, + 0.47, + 0.435 + ], + "angle": 0, + "content": "proaches, our proposed methodology exploits a novel local attention mechanism and a cascaded framework to deal with the arbitrary-scale SR. In order to fairly compare with the above approaches, we similarly adopt EDSR [7], RDN [9] and SwinIR [14] as the encoders for our LIT and CLIT." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.447, + 0.212, + 0.464 + ], + "angle": 0, + "content": "3. Methodology" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.472, + 0.469, + 0.533 + ], + "angle": 0, + "content": "In this section, we first provide an overview of the proposed LIT framework, followed by the implementation details of it and its main modules. We then discuss our cumulative training strategy, as well as the framework of CLIT." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.542, + 0.367, + 0.557 + ], + "angle": 0, + "content": "3.1. Overview of the LIT Framework" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.565, + 0.473, + 0.825 + ], + "angle": 0, + "content": "LIT is a framework that employs a novel cross-scaled local attention mechanism and a local frequency encoding technique to perform arbitrary-scale SR tasks. Fig. 2 (a) provides an overview of the proposed framework, which aims at producing an HR image \\( I^{HR} \\in \\mathbb{R}^{r_hH \\times r_wW \\times 3} \\) at 2D HR coordinates \\( \\mathbf{x}^{HR} \\in \\mathcal{X} \\) from a given LR image \\( I^{LR} \\in \\mathbb{R}^{H \\times W \\times 3} \\) at 2D LR coordinates \\( \\mathbf{x}^{LR} \\in \\mathcal{X} \\) based on an arbitrary upsampling scale \\( \\mathbf{r} = \\{r_h, r_w\\} \\), where \\( \\mathcal{X} \\) is the 2D coordinate space that is used to represent an image in the continuous domain. An encoder \\( E_{\\theta} \\) first extracts a feature embedding \\( \\mathcal{Z} \\in \\mathbb{R}^{H \\times W \\times C} \\) from \\( I^{LR} \\). The extracted \\( \\mathcal{Z} \\) is then forwarded into LIT along with the 2D coordinates of \\( I^{HR} \\) and a cell \\( = (2 / s_h, 2 / s_w) \\) to generate the RGB values of a residual image \\( I_r^{HR} \\in \\mathbb{R}^{r_hH \\times r_wW \\times 3} \\) in a pixel-wise fashion. Lastly, the residual image \\( I_r^{HR} \\) is combined with a bilinearly upsampled image \\( I_{\\uparrow}^{HR} \\in \\mathbb{R}^{r_hH \\times r_wW \\times 3} \\) via element-wise addition to derive the final HR image \\( I^{HR} \\)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.833, + 0.325, + 0.849 + ], + "angle": 0, + "content": "3.2. Local Implicit Transformer" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.47, + 0.901 + ], + "angle": 0, + "content": "LIT is developed for mapping any 2D coordinate in the continuous image domain to an RGB color. As highlighted in Fig. 2 (b), it is composed of a cross-scale local attention" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.358, + 0.892, + 0.556 + ], + "angle": 0, + "content": "block (CSLAB), a local frequency encoding block (LFB), and a decoder \\(D_{\\phi}\\) paraterizerized by \\(\\phi\\). The former two blocks are responsible for estimating a local latent embedding \\(\\tilde{z} \\in \\mathbb{R}^{G_h G_w \\times C}\\) and a local frequency embedding \\(\\tilde{f} \\in \\mathbb{R}^{G_h G_w \\times C}\\), respectively, where \\(G_h\\) and \\(G_w\\) denote the height and width of local grids employed for performing local coordinate sampling, as depicted in Fig. 3. On the other hand, \\(D_{\\phi}\\) utilizes these embeddings along with the provided cell to generate \\(I_r^{HR}\\). More specifically, LIT first projects \\(\\mathcal{Z}\\) using four separate convolutional layers to obtain four latent embeddings, corresponding to query \\(q\\), key \\(k\\), value \\(v\\), and frequency \\(f\\). Based on a queried HR coordinate \\(x_q \\in \\mathbf{x}^{HR}\\), CSLAB and LFB estimate \\(\\tilde{z}\\) and \\(\\tilde{f}\\) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.607, + 0.567, + 0.891, + 0.583 + ], + "angle": 0, + "content": "\\[\n\\tilde {z} = C S L A B (\\delta x, q, k, v), \\tag {1}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.627, + 0.592, + 0.891, + 0.61 + ], + "angle": 0, + "content": "\\[\n\\tilde {f} = L F E B (\\delta x, f), \\tag {2}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.538, + 0.615, + 0.891, + 0.644 + ], + "angle": 0, + "content": "\\[\n\\delta \\mathbf {x} = \\left\\{x _ {q} - x ^ {(i, j)} \\right\\} _ {i \\in \\{1, 2, \\dots , G _ {h} \\}, j \\in \\{1, 2, \\dots , G _ {w} \\}}, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.649, + 0.892, + 0.832 + ], + "angle": 0, + "content": "where \\(x^{(i,j)}\\in \\mathbf{x}^{LR}\\) denotes an LR coordinate in the local grid indexed by \\((i,j)\\), and \\(\\delta \\mathbf{x}\\) represents the set of local relative coordinates defined by Eq. (3). The local grid is sampled in a manner that positions its center \\(x^{(\\lfloor G_h / 2\\rfloor +1,\\lfloor G_w / 2\\rfloor +1)}\\) at the LR coordinate closest to \\(x_{q}\\). The query latent vector \\(q\\in \\mathbb{R}^{1\\times C}\\) at the HR coordinate \\(x_{q}\\) is computed by bilinear interpolation, while the remainder of the local latent embeddings \\(k\\in \\mathbb{R}^{G_hG_w\\times C}\\), \\(v\\in \\mathbb{R}^{G_hG_w\\times C}\\), and \\(f\\in \\mathbb{R}^{G_hG_w\\times C}\\) are sampled at the local LR coordinates \\(\\mathbf{x} = \\{x^{i,j}\\}_{i\\in \\{1,2,\\dots,G_h\\},j\\in \\{1,2,\\dots,G_w\\}}\\). With the local latent embedding \\(\\tilde{z}\\) and the local frequency embedding \\(\\tilde{f}\\), the function of \\(D_{\\phi}\\) is formulated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.621, + 0.842, + 0.891, + 0.861 + ], + "angle": 0, + "content": "\\[\nI ^ {r} \\left(x _ {q}\\right) = D _ {\\phi} \\left(\\tilde {z}, \\tilde {f}, c\\right), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.871, + 0.892, + 0.903 + ], + "angle": 0, + "content": "where \\(I^{r}(x_{q})\\) is the predicted RGB value at the queried coordinate \\(x_{q}\\), and \\(c = \\{HR_{\\Delta h}, HR_{\\Delta w}\\}\\) denotes the cell that" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18259" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.08, + 0.087, + 0.473, + 0.307 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.31, + 0.471, + 0.382 + ], + "angle": 0, + "content": "Figure 3. An illustration of the proposed local coordinate sampling scheme. The red rectangular region, outlined by a dashed line, represents the local grid with dimensions \\( G_{h} \\times G_{w} \\). The red dots indicate the sampled local LR coordinates, while the gray dots correspond to the HR coordinates. The grid dimensions \\( LR_{\\Delta h,w} \\) and \\( HR_{\\Delta h,w} \\) represent the unit sizes of pixels in LR and HR images, respectively." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.4, + 0.47, + 0.462 + ], + "angle": 0, + "content": "represents the height and width of a pixel in an HR image, as illustrated in Fig. 3. \\(D_{\\phi}\\) is implemented as a five-layer MLP utilizing Gaussian Error Linear Unit (GELU) activation [46], and is employed consistently across all images." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.479, + 0.47, + 0.631 + ], + "angle": 0, + "content": "Cross-scale local attention block. LIT exploits CSLAB to perform a local attention mechanism over a local grid to generate a local latent embedding \\(\\tilde{z}\\) for each HR coordinate, as illustrated in Fig. 2 (c). CSLAB first calculates the inner product of \\(q\\) and \\(k\\), adds the relative positional bias \\(B\\) to the result, and obtains an attention matrix. This attention matrix is subsequently normalized by a Softmax operation to produce a local attention map. Finally, CSLAB performs element-wise multiplication of \\(v\\) and the local attention map to derive \\(\\tilde{z}\\). The overall procedure is formulated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.15, + 0.64, + 0.47, + 0.746 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\tilde {z} = \\operatorname {s o f t m a x} \\left(\\frac {q k ^ {\\top}}{\\sqrt {C}} + B\\right) \\times v, (5) \\\\ B = F C (\\gamma (\\delta \\mathbf {x})), (6) \\\\ \\gamma (\\delta \\mathbf {x}) = \\left[ \\sin \\left(2 ^ {0} \\delta \\mathbf {x}\\right), \\cos \\left(2 ^ {0} \\delta \\mathbf {x}\\right), \\dots \\right. (7) \\\\ , \\sin \\left(2 ^ {L - 1} \\delta \\mathbf {x}\\right), \\cos \\left(2 ^ {L - 1} \\delta \\mathbf {x}\\right) ], \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.752, + 0.469, + 0.828 + ], + "angle": 0, + "content": "where \\(C\\) denotes the channel dimension of the local key latent embedding \\(k\\), \\(FC\\) represents a fully-connected layer, \\(\\gamma\\) is the positional encoding function, and \\(L\\) is a hyperparameter. In this work, \\(L\\) is set to 10, and the multi-head attention mechanism adopted is formulated in Eq. (8) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.122, + 0.838, + 0.469, + 0.875 + ], + "angle": 0, + "content": "\\[\n\\tilde {z} = \\operatorname {c o n c a t} \\left(\\operatorname {s o f t m a x} \\left(\\frac {q _ {i} k _ {i} ^ {\\top}}{\\sqrt {C / H}} + B _ {i}\\right) \\times v _ {i}\\right), \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.886, + 0.469, + 0.902 + ], + "angle": 0, + "content": "where \\(H\\) is the number of attention heads and \\(i\\in [1,\\dots ,H]\\)" + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.091, + 0.891, + 0.246 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.544, + 0.25, + 0.847, + 0.263 + ], + "angle": 0, + "content": "Figure 4. An overview of the proposed CLIT framework." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.276, + 0.892, + 0.442 + ], + "angle": 0, + "content": "Local frequency encoding block. Fig. 2 (d) illustrates the local frequency encoding block, which incorporates dominant-frequency components to generate the local frequency embedding \\(\\tilde{f}\\). In order to account for dominant-frequency information, LFBB first encodes local frequency latent embedding \\(f\\) using Fast Fourier Transform (FFT), which is then multiplied by the relative positional encoding to generate \\(\\tilde{f}\\). The resultant local frequency embedding \\(\\tilde{f}\\) is then concatenated with the flattened local latent embedding \\(\\tilde{z}\\) to form a mixing embedding, which can be used by the decoder consisting of MLPs to predict RGB values [47]." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.454, + 0.767, + 0.47 + ], + "angle": 0, + "content": "3.3. Cumulative Training Strategy" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.479, + 0.892, + 0.901 + ], + "angle": 0, + "content": "In this section, we discuss our proposed cumulative training strategy, which is developed for enhancing the performance of arbitrary-scale SR. The cumulative training strategy focuses on the schedule of the cell sizes selected during the training phase, as \"cell decoding\" has been recognized as an essential input to a local implicit image function [17]. Recent studies [17, 19] have observed that the effect of cell decoding on the performance of arbitrary-scale SR is prominent for in-distribution upsampling, but degrades significantly for out-of-distribution large-scale upsampling. Such trends are demonstrated in Table 3 of [17] and Table 4 of [19]. The authors in [21] also constrained their cell sizes to in-distribution ranges to mitigate the negative impact during evaluation. To overcome the degradation issue for out-of-distribution cell sizes, incorporating large cell sizes during training appears to be a promising solution. However, simply training the local implicit image function with a diverse range of cell sizes at once leads to a performance drop. According to our experimental observations presented in Section 4.4, we find that training the local implicit image function by alternatively switching between large and small cells offers positive impacts on the performance. Based on the above observations, we propose a cumulative training strategy which first trains the local implicit image function with large cell sizes, and finetunes it with the alternative training strategy to improve the performance on different upsampling scales. More quantitative results of our training strategies can be found in Section 4.4." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18260" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.104, + 0.089, + 0.874, + 0.392 + ], + "angle": 0, + "content": "
Method×2×3×4×6×12×18×24×30
Bicubic [7]31.0128.2226.6624.8222.2721.0020.1919.59
EDSR-baseline [7]34.5530.9028.94-----
EDSR-baseline-Meta-SR [16, 17]34.6430.9328.9226.6123.5522.0321.0620.37
EDSR-baseline-LIIF [17]34.6730.9629.0026.7523.7122.1721.1820.48
EDSR-baseline-UltraSR [18]34.6931.0229.0526.8123.7522.2121.2120.51
EDSR-baseline-IPE [19]34.7231.0129.0426.7923.7522.2121.2220.51
EDSR-baseline-LTE [21]34.7231.0229.0426.8123.7822.2321.2420.53
EDSR-baseline-CLIT (Ours)34.8131.1229.1526.9223.8322.2921.2620.53
RDN-Meta-SR [16, 17]35.0031.2729.2526.8823.7322.1821.1720.47
RDN-LIIF [17]34.9931.2629.2726.9923.8922.3421.3120.59
RDN-UltraSR [18]35.0031.3029.3227.0323.7322.3621.3320.61
RDN-IPE [19]35.0431.3229.3227.0423.9322.3821.3420.63
RDN-LTE [21]35.0431.3229.3327.0423.9522.4021.3620.64
RDN-CLIT (Ours)35.1031.3929.3927.1224.0122.4521.3820.64
SwinIR-MetaSR [16, 21]35.1531.4029.3326.9423.8022.2621.2620.54
SwinIR-LIIF [17, 21]35.1731.4629.4627.1524.0222.4321.4020.67
SwinIR-LTE [21]35.2431.5029.5127.2024.0922.5021.4720.73
SwinIR-CLIT (Ours)35.2931.5529.5527.2624.1122.5121.4520.70
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.395, + 0.892, + 0.42 + ], + "angle": 0, + "content": "Table 1. The average PSNR (dB) on the DIV2K [48] validation set. The results are obtained from the original manuscripts [17-19, 21]. The best and second-best performing results are highlighted by the red and blue colors, respectively." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.424, + 0.404, + 0.44 + ], + "angle": 0, + "content": "3.4. Cascaded Local Implicit Transformer" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.45, + 0.471, + 0.57 + ], + "angle": 0, + "content": "Previous works [17, 21] typically address arbitrary-scale SR through a single step of upsampling. However, one-step upsampling struggles to reconstruct an HR image when the upsampling scale is large [5, 42, 43]. In light of this, we propose a cascaded upsampling strategy, called \"Cascaded LIT(CLIT)\" to predict residual images from multi-scale feature embeddings. In an \\(N\\)-branched CLIT, the multi-scale feature embeddings \\(\\mathcal{Z}^1, \\mathcal{Z}^2, \\dots, \\mathcal{Z}^N\\) are derived as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.182, + 0.586, + 0.469, + 0.613 + ], + "angle": 0, + "content": "\\[\n\\mathcal {Z} ^ {N} = \\mathcal {Z} \\uparrow_ {s ^ {1} \\times s ^ {2} \\times \\dots \\times s ^ {N - 1}}, \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.184, + 0.608, + 0.348, + 0.623 + ], + "angle": 0, + "content": "where \\(s^1 = 1\\) and \\(s\\in \\mathbf{s}\\)" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.64, + 0.47, + 0.731 + ], + "angle": 0, + "content": "where \\(\\uparrow\\) is a bilinear upsampling function and \\(s\\) is a set of scaling factors, which are configurable hypermeters. For a branch \\(i,i\\in [1,\\dots,N]\\), \\(\\mathrm{LIT}^i\\) estimates the residual image \\(I_r^i\\) from the feature embedding \\(\\mathcal{Z}^i\\) with the coordinate and the corresponding cells. Lastly, the final HR image \\(I^{HR}\\in \\mathbb{R}^{r_hH\\times r_wW\\times 3}\\) can be estimated as the following equation:" + }, + { + "type": "equation", + "bbox": [ + 0.093, + 0.745, + 0.47, + 0.765 + ], + "angle": 0, + "content": "\\[\nI ^ {H R} = \\lambda^ {N - 1} I _ {r} ^ {1} + \\lambda^ {N - 2} I _ {r} ^ {2} + \\dots + \\lambda^ {0} I _ {r} ^ {N} + I _ {\\uparrow} ^ {H R}. \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.78, + 0.469, + 0.901 + ], + "angle": 0, + "content": "where \\(\\lambda\\) is a discount factor, with a default value of 0.75. During the training phase, CLIT is trained using the proposed cumulative training strategy. Initially, \\(\\mathrm{LIT}^1\\) is trained with the strategy employed by [17]. Subsequently, \\(\\mathrm{LIT}^1\\) is finetuned and \\(\\mathrm{LIT}^2\\) is initialized by applying the alternative training strategy. By incrementally incorporating LITs into CLIT, the performance is progressively enhanced. The details of the training strategies are specified in Section 4." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.424, + 0.706, + 0.44 + ], + "angle": 0, + "content": "4. Experimental Results" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.45, + 0.892, + 0.556 + ], + "angle": 0, + "content": "In this section, we present the experimental results and discuss their implications. We begin with a brief introduction to our experimental setup in Section 4.1. Following that, we evaluate our CLIT with different datasets in Section 4.2. The, Section 4.3 showcases the learned attention maps. Finally, ablation studies for various configurations of the proposed CLIT and LIT are compared in Section 4.4." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.567, + 0.691, + 0.584 + ], + "angle": 0, + "content": "4.1. Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.592, + 0.892, + 0.712 + ], + "angle": 0, + "content": "Dataset. We use the DIV2K dataset [48] for network training. It consists of 1,000 images in 2K resolutions and provides low-resolution counterparts with down-sampling scales, \\(\\times 2\\), \\(\\times 3\\), \\(\\times 4\\), which are generated by the bicubic interpolation method. On the other hand, we also evaluate the performance on the validation set of DIV2K [48], Set5 [49], Set14 [50], B100 [51] and Urban100 [52] in terms of peak signal-to-noise (PSNR) values." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.735, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Training details During training, we feed batches of size \\(48 \\times 48\\) low-resolution images into the framework, following the prior works [7]. For each batch, a single upsampling scale is sampled from a uniform distribution \\(r \\sim \\mathcal{U}(1,4)\\). With single upsampling scales \\(s\\), a batch of HR images is cropped into patches of size \\(48r \\times 48r\\), while the corresponding LR images are cropped into patches of \\(48 \\times 48\\). The patches are augmented by randomly horizontal flipping, vertical flipping, and \\(90^{\\circ}\\) rotating. Then, we sample \\(48^2\\) pixels (coordinate-RGB pairs) on each HR patch as the ground-truths. We set the batch size to 32 and use" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "18261" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.081, + 0.089, + 0.894, + 0.214 + ], + "angle": 0, + "content": "
MethodSet5Set14B100Urban100
×2×3×4×6×8×2×3×4×6×8×2×3×4×6×8×2×3×4×6×8
RDN [9]38.2434.7132.47--34.0130.5728.81--32.3429.2627.72--32.8928.8026.61--
RDN-Meta-SR [16, 17]38.2234.6332.3829.0426.9633.9830.5428.7826.5124.9732.3329.2627.7125.9024.8332.9228.8226.5523.9922.59
RDN-LLIF [17]38.1734.6832.5029.1527.1433.9730.5328.8026.6425.1532.3229.2627.7425.9824.9132.8728.8226.6824.2022.79
RDN-UltraSR [18]38.2134.6732.4929.3327.2433.9730.5928.8626.6925.2532.3529.2927.7726.0124.9632.9728.9226.7824.3022.87
RDN-IPE [19]38.1134.6832.5129.2527.2233.9430.4728.7526.5825.0932.3129.2827.7626.0024.9332.9728.8226.7624.2622.87
RDN-LTE [21]38.2334.7232.6129.3227.2634.0930.5828.8826.7125.1632.3629.3027.7726.0124.9533.0428.9726.8124.2822.88
RDN-CLIT (Ours)38.2634.7932.6929.5427.3434.0930.6928.9326.8325.3632.3929.3327.8026.0725.0033.1429.0526.9324.4423.04
SwinIR [14]38.3534.8932.72--34.1430.7728.94--32.4429.3727.83--33.4029.2927.07--
SwinIR-MetaSR [16, 21]38.2634.7732.4729.0927.0234.1430.6628.8526.5825.0932.3929.3127.7525.9424.8633.2929.1226.7624.1622.75
SwinIR-LLIF [17]38.2834.8732.7329.4627.3634.1430.7528.9826.8225.3432.3929.3427.8426.0725.0133.3629.3327.1524.5923.14
SwinIR-LTE [21]38.3334.8932.8129.5027.3534.2530.8029.0626.8625.4232.4429.3927.8626.0925.0333.5029.4127.2424.6223.17
SwinIR-CLIT (Ours)38.4134.9732.8629.6927.6234.2730.8529.0826.9425.5532.4629.4227.9126.1525.0933.5629.4327.2524.7723.33
" + }, + { + "type": "table_caption", + "bbox": [ + 0.078, + 0.216, + 0.891, + 0.24 + ], + "angle": 0, + "content": "Table 2. The average PSNR (dB) on Set5 [49], Set14 [50], B100 [51], and Urban100 [52]. The results are obtained from the original manuscripts [17-19, 21]. The best and second-best performing results are highlighted by the red and blue colors, respectively." + }, + { + "type": "image", + "bbox": [ + 0.131, + 0.246, + 0.839, + 0.684 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.205, + 0.685, + 0.763, + 0.699 + ], + "angle": 0, + "content": "Figure 5. The qualitative results of LIIF [17], LTE [21], and our CLIT with using RDN [9] as the encoder." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.704, + 0.471, + 0.9 + ], + "angle": 0, + "content": "the Adam optimizer [53] together with L1 loss for training. We train LIT for 1,000 epochs, and the learning rate is initialized at \\( 1e^{-4} \\) and decayed by a factor 0.5 at [200, 400, 600, 800] epochs. For cumulative training of CLIT, as discussed in Section 3.4, we sample \\( N \\) scale factors \\( \\{s^1, s^2, \\dots, s^N\\} \\) from the distribution \\( \\mathcal{U}(1,4) \\) according the number of LITs \\( N \\) in the train step. The total upsampling scale \\( r = s^1 \\times s^2 \\dots \\times s^N \\) is the product of all scale factors. If \\( 48r \\times 48r \\) is greater than the whole HR image, we clip the scale factor of stage 1. To train \\( N \\) LITs, we fine-tune the model for \\( 500 \\times N \\) epochs, and the learning rate is initialized at \\( 1e^{-4} \\) and decayed by a factor 0.5 at \\( [100 \\times N, 200 \\times N, 300 \\times N, 400 \\times N] \\) epochs." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.703, + 0.684, + 0.717 + ], + "angle": 0, + "content": "4.2. Validation of CLIT" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.73, + 0.892, + 0.852 + ], + "angle": 0, + "content": "Quantitative results. We first compare our proposed CLIT to existing local implicit neural representation methods for arbitrary-scale SR, including LIIF [17], UltraSR [18], IPE [19], and LTE [21]. Table 1 summarizes the quantitative results in terms of PSNR(dB) on DIV2k [48]. As shown in Table 1, CLIT achieves the best performance when EDSR-Baseline [7], RDN [9] are used as the encoders. These results demonstrate the advantage of CLIT." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.856, + 0.893, + 0.901 + ], + "angle": 0, + "content": "Table 2 compares CLIT to prior works [17-19, 21] on widely-used datasets, including Set5 [49], Set14 [50], B100 [51], and Urban100 [52], with RDN and SwinIR." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18262" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.118, + 0.089, + 0.855, + 0.432 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.127, + 0.433, + 0.841, + 0.447 + ], + "angle": 0, + "content": "Figure 6. The qualitative results of LIIF [17], LTE [21], and our CLIT using RDN [9] as the encoder and non-integer upsampling scales." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.451, + 0.471, + 0.514 + ], + "angle": 0, + "content": "Note that RDN [9] is trained and evaluated for specific upsampling scales. CLIT outperforms those existing methods in most cases across all datasets and scales, even achieving a 0.26dB PSNR improvement on Set5 for the \\(\\times 8\\) scale." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.523, + 0.47, + 0.794 + ], + "angle": 0, + "content": "Qualitative results. Fig. 5 compares the qualitative results of CLIT with the baseline methods LIIF [17] and LTE [21], on various dataset DIV2k [48], Set14 [50], B100 [51], and Urban100 [52] with variant upsampling scales. The official codes provided by them were used to produce the results. In the first row, the SR visualization results with a \\(\\times 12\\) upsampling scale are depicted. It can be observed that both LIIF and LTE struggle to reconstruct the letters continuously, whereas the CLIT result demonstrates continuity for these alphabets, particularly for 'M' and 'S'. In the second row, the stripes on the zebra appear blurry in the LIIF and LTE results, whereas they are more distinct in the CLIT result. In the third row, the cross pattern on the tablecloth is not clearly rendered by LIIF and LTE. In contrast, CLIT is able to successfully generate clean crosses and sharp lines. In the fourth row, despite the blurriness of the original LR image, CLIT captures the texture and produces straight lines on the flag, showcasing its effectiveness." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.796, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Fig. 6 presents the results of text image enhancement employing progressively increasing non-integer upsampling factors, alongside a comparison with earlier methods, LIIF [17] and LTE [21]. The input image first undergoes a downscaling process by a factor of 2.2 to generate an LR image. Subsequently, the LR image is then upscaled using predetermined non-integer factors \\(\\times 1.6\\), \\(\\times 2.5\\), \\(\\times 3.4\\), \\(\\times 4.3\\)." + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.45, + 0.892, + 0.685 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.688, + 0.892, + 0.713 + ], + "angle": 0, + "content": "Figure 7. A visualization of the local attention maps of the coordinates highlight as red dots." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.718, + 0.892, + 0.825 + ], + "angle": 0, + "content": "It can be observed that our proposed model effectively captures the patterns of text within the image, and accurately estimates words and numbers with enhanced sharpness and clarity. The enhancement becomes particularly evident in the word 'infinity' in the first row of contexts, 'SUNSHINE' in the fifth row, and the number '0' in the final row, as each of these elements demonstrates considerable improvements." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.833, + 0.809, + 0.85 + ], + "angle": 0, + "content": "4.3. Visualization Local Attention Maps" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.892, + 0.903 + ], + "angle": 0, + "content": "In Fig. 7, we provide visualizations of local attention maps. The LR images are generated by applying bicubic downsampling to HR images with scale factors" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18263" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.082, + 0.089, + 0.47, + 0.184 + ], + "angle": 0, + "content": "
In-distributionOut-of-distribution
×2×3×4×6×8×12
LIT34.7931.0729.1026.8425.4723.77
LIT (-a)34.7131.0029.0226.7625.4023.70
LIT (-f)34.7831.0729.0926.8325.4623.77
LIT (-c)34.6531.0029.0426.8025.4523.77
LIT (+e)34.7931.0829.1126.8525.4823.78
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.186, + 0.47, + 0.236 + ], + "angle": 0, + "content": "Table 3. The average PSNR (dB) for different design choices. The notations -a/f/c denote the removal of the cross-scale local attention block, frequency encoding block, and cell size, respectively, and +e denotes the addition of local ensemble. The best performing one is indicated in bolded." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.24, + 0.471, + 0.362 + ], + "angle": 0, + "content": "\\(\\{ \\times 4.5, \\times 6, \\times 6 \\}\\). Subsequently, the proposed CLIT with RDN [9] encoder is utilized to produce HR predictions. It can be observed that the attention maps closely align with the edges, indicating that the cross-scale local attention block effectively captures similar latent codes within a local area. This substantiates that our proposed design enables the generation of SR images from LR counterparts, ultimately producing clean and sharp edges in the output." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.37, + 0.241, + 0.385 + ], + "angle": 0, + "content": "4.4. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.393, + 0.47, + 0.469 + ], + "angle": 0, + "content": "In this section, we present a series of ablation analyses to substantiate the design decisions proposed in this paper. All of the ablation experiments are conducted on the DIV2K [48] validation set, utilizing the EDSR-baseline [7] as the encoder and employing the PSNR metric for evaluation." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.472, + 0.47, + 0.669 + ], + "angle": 0, + "content": "Validation of the design choices. Table 3 presents a summary of the quantitative contributions associated with each LIT component. A significant improvement is observed when adopting the cross-scale local attention block, while a relatively minor gain with the frequency encoding block by comparing LTE with LTE (-a) and LTE (-f), respectively. In addition, LTE (-c) reveals that removing cell decoding leads to more severe performance degradation for in-distribution upsampling scales compared to out-of-distribution counterparts, highlighting the importance of cell decoding for in-distribution upsampling scales. Finally, incorporating the local ensemble technique in LIT (+e) results in a further, albeit modest, enhancement of the overall performance." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.671, + 0.47, + 0.793 + ], + "angle": 0, + "content": "The effectiveness of the local grid. Table 4 provides a performance comparison of various local grid sizes employed within the local coordinate sampling scheme, as discussed in Section 3. The results presented in Table 4 reveal that increasing the local grid size contributes to performance enhancements, albeit at the cost of extended training times. As a result, this study adopts a \\(7 \\times 7\\) local grid size to strike an ideal balance between effectiveness and efficiency." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.796, + 0.47, + 0.903 + ], + "angle": 0, + "content": "Analysis on the training strategy. Table 5 presents a quantitative comparison of the proposed cumulative training strategy with the other training strategies for training an LIT. The baseline strategy trains LIT using upsampling scales sampled from a uniform distribution, \\( r \\sim \\mathcal{U}(1,4) \\). Expanding the sampling scale distribution to \\( r \\sim \\mathcal{U}(1,12) \\) improves the performance of large-scale upsampling while" + }, + { + "type": "table", + "bbox": [ + 0.505, + 0.088, + 0.894, + 0.183 + ], + "angle": 0, + "content": "
Local gird\nGh × GwIn-distributionOut-of-distributionTraining time
×2×3×4×6×8×12
1 × 134.7131.0029.0226.7625.4023.7020.9
3 × 334.7631.0429.0726.8125.4623.7623.2
5 × 534.7831.0729.0826.8325.4723.7828.4
7 × 734.7931.0729.1026.8425.4723.7733.6
9 × 934.7931.0629.1026.8425.4723.7753.0
15 × 1534.8331.1029.1126.8625.4923.7876.1
25 × 2534.8631.1329.1526.8825.5023.80190.9
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.185, + 0.892, + 0.21 + ], + "angle": 0, + "content": "Table 4. The average PSNR (dB) and the training time (hour) of multiple local grid sizes. The bolded numbers correspond to the best performance." + }, + { + "type": "table", + "bbox": [ + 0.504, + 0.213, + 0.892, + 0.267 + ], + "angle": 0, + "content": "
Training strategy×2×3×4×6×8×12
Training with r ~ U(1, 4)34.7931.0729.1026.8425.4723.77
Training with r ~ U(1, 12)34.6931.0629.1126.8725.5223.82
Alternative training strategy34.7531.0929.1226.8825.5223.81
Cumulative training strategy34.7831.1029.1426.8925.5423.83
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.269, + 0.892, + 0.293 + ], + "angle": 0, + "content": "Table 5. The average PSNR (dB) of different training strategies. The best performing results are indicated in **bold**." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.297, + 0.892, + 0.524 + ], + "angle": 0, + "content": "compromising the performance of small-scale upsampling. To achieve high-quality results across all upsampling scale ranges, the alternative training strategy trains LIT alternatively by switching the sampling scales between \\( r \\sim \\mathcal{U}(1,4) \\) and \\( r \\sim \\mathcal{U}(4,12) \\). On the other hand, the proposed cumulative training strategy first trains LIT with \\( r \\sim \\mathcal{U}(1,4) \\), followed by fine-tuning using the alternative training strategy. This progressive approach improves the performance of LIT by gradually broadening the distribution of the upsampling scales. Comparing the results of the cumulatively trained LIT with the results of EDSR-baseline-CLIT in Table 1 reveals that the proposed cumulative training strategy works for both LIT and CLIT. Moreover, the results from these tables suggest that the cascaded framework is more effective in addressing arbitrary-scale SR tasks." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.527, + 0.619, + 0.542 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.546, + 0.892, + 0.803 + ], + "angle": 0, + "content": "In this paper, we introduced the attention mechanism and frequency encoding technique to address arbitrary-scale SR tasks. To achieve this objective, we proposed the LIT framework, which consists of a cross-scale local attention block and a local frequency encoding block. The former is designed to find the local latent embedding to reconstruct the corresponding RGB value, while the latter focuses on encoding coordinates with the frequency information derived from the feature embeddings. To enhance the capability of capturing fine details, we proposed CLIT and the corresponding cumulative training strategy that trains the model with progressively increasing upsampling scales. Based on the experimental results, both quantitative and qualitative assessments showcase the superior performance of LIT and CLIT in comparison to existing methods. Furthermore, our comprehensive analyses validated the effectiveness of the training strategies and the components employed by them." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.807, + 0.667, + 0.823 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.826, + 0.892, + 0.901 + ], + "angle": 0, + "content": "The authors gratefully acknowledge the support from the National Science and Technology Council (NSTC) in Taiwan under grant numbers MOST 111-2223-E-007-004-MY3 and MOST 111-2628-E-007-010, as well as the financial and technical support from MediaTek Inc., Taiwan." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18264" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.17 + ], + "angle": 0, + "content": "[1] C. Dong, C. C. Loy, K. He, and X. Tang. Image superresolution using deep convolutional networks. IEEE Trans. Pattern Analysis and Machine Intelligence (TPAMI), 38(2): 295-307, 2016. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.181, + 0.471, + 0.236 + ], + "angle": 0, + "content": "[2] J. Kim, J. K. Lee, and K. M. Lee. Accurate image superresolution using very deep convolutional networks. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), pages 1646-1654, 2016." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.245, + 0.47, + 0.288 + ], + "angle": 0, + "content": "[3] C. Dong, C. C. Loy, and X. Tang. Accelerating the superresolution convolutional neural network. In Proc. European Conf. on Computer Vision (ECCV), pages 391-407, 2016. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.297, + 0.468, + 0.366 + ], + "angle": 0, + "content": "[4] W. Shi, J. Caballero, F. Huszar, J. Totz, A. P. Aitken, et al. Real-time single image and video super-resolution using an efficient sub-pixel convolutional neural network. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), pages 1874-1883, 2016. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.375, + 0.468, + 0.431 + ], + "angle": 0, + "content": "[5] W. Lai, J. Huang, N. Ahuja, and M. Yang. Fast and accurate image super-resolution with deep laplacian pyramid networks. IEEE Trans. Pattern Analysis and Machine Intelligence (TPAMI), 41(11):2599-2613, 2019. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.441, + 0.468, + 0.509 + ], + "angle": 0, + "content": "[6] C. Ledig, L. Theis, F. Huszar, J. Caballero, A. Cunningham, et al. Photo-realistic single image super-resolution using a generative adversarial network. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), pages 105-114, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.519, + 0.468, + 0.588 + ], + "angle": 0, + "content": "[7] B. Lim, S. Son, H. Kim, S. Nah, and K. M. Lee. Enhanced deep residual networks for single image super-resolution. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition Workshop (CVPRW), pages 1132-1140, 2017. 2, 3, 5, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.599, + 0.468, + 0.655 + ], + "angle": 0, + "content": "[8] X. Wang, K. Yu, S. Wu, J. Gu, Y. Liu, et al. Esrgan: Enhanced super-resolution generative adversarial networks. In Proc. European Conf. on Computer Vision Workshop (ECCVW), pages 63-79, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.664, + 0.468, + 0.72 + ], + "angle": 0, + "content": "[9] Y. Zhang, Y. Tian, Y. Kong, B. Zhong, and Y. Fu. Residual dense network for image super-resolution. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), pages 2472-2481, 2018. 2, 3, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.728, + 0.468, + 0.784 + ], + "angle": 0, + "content": "[10] Y. Zhang, K. Li, K. Li, L. Wang, B. Zhong, et al. Image super-resolution using very deep residual channel attention networks. In Proc. European Conf. on Computer Vision (ECCV), pages 294-310, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.794, + 0.468, + 0.849 + ], + "angle": 0, + "content": "[11] T. Dai, J. Cai, Y. Zhang, S. Xia, and L. Zhang. Second-order attention network for single image super-resolution. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), pages 11065-11074, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.859, + 0.468, + 0.902 + ], + "angle": 0, + "content": "[12] Y. Mei, Y. Fan, Y. Zhou, L. Huang, T. S. Huang, et al. Image super-resolution with cross-scale non-local attention and exhaustive self-exemplars mining. In Proc. IEEE Conf." + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.471, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "on Computer Vision and Pattern Recognition (CVPR), pages 5689-5698, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.129, + 0.892, + 0.184 + ], + "angle": 0, + "content": "[13] H. Chen, Y. Wang, T. Guo, C. Xu, Y. Deng, et al. Pre-trained image processing transformer. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), pages 12294-12305, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.193, + 0.892, + 0.249 + ], + "angle": 0, + "content": "[14] J. Liang, J. Cao, G. Sun, K. Zhang, L. Van Gool, et al. Swinir: Image restoration using swim transformer. In Proc. IEEE Int. Conf. on Computer Vision Workshop (ICCVW), pages 1833-1844, 2021. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.258, + 0.892, + 0.299 + ], + "angle": 0, + "content": "[15] X. Chen, X. Wang, J. Zhou, and C. Dong. Activating more pixels in image super-resolution transformer. CoRR, abs/2205.04437, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.308, + 0.892, + 0.363 + ], + "angle": 0, + "content": "[16] X. Hu, H. Mu, X. Zhang, Z. Wang, T. Tan, et al. Meta-sr: A magnification-arbitrary network for super-resolution. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), pages 1575-1584, 2019. 1, 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.373, + 0.892, + 0.429 + ], + "angle": 0, + "content": "[17] Y. Chen, S. Liu, and X. Wang. Learning continuous image representation with local implicit image function. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), pages 8628-8638, 2021. 1, 2, 4, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.437, + 0.892, + 0.479 + ], + "angle": 0, + "content": "[18] X. Xu, Z. Wang, and H. Shi. Ultrasr: Spatial encoding is a missing key for implicit image function-based arbitrary-scale super-resolution. CoRR, abs/2103.12716, 2021. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.487, + 0.892, + 0.529 + ], + "angle": 0, + "content": "[19] Y. Liu, Y. Guo, and S. Zhang. Enhancing multi-scale implicit learning in image super-resolution with integrated positional encoding. CoRR, abs/2112.05756, 2021. 2, 4, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.538, + 0.892, + 0.593 + ], + "angle": 0, + "content": "[20] J. Yang, S. Shen, H. Yue, and K. Li. Implicit transformer network for screen content image continuous super-resolution. In Proc. Conf. on Neural Information Processing Systems (NeurIPS), pages 13304-13315, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.602, + 0.892, + 0.657 + ], + "angle": 0, + "content": "[21] J. Lee and K. H. Jin. Local texture estimator for implicit representation function. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), pages 1929-1938, 2022. 1, 2, 4, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.666, + 0.892, + 0.72 + ], + "angle": 0, + "content": "[22] A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, et al. Attention is all you need. In Proc. Conf. on Neural Information Processing Systems (NeurIPS), pages 5998-6008, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.73, + 0.892, + 0.772 + ], + "angle": 0, + "content": "[23] Z. Liu, Y. Lin, Y. Cao, H. Hu, Y. Wei, et al. Swin transformer: Hierarchical vision transformer using shifted windows. In Proc. IEEE Int. Conf. on Computer Vision (ICCV), 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.781, + 0.892, + 0.836 + ], + "angle": 0, + "content": "[24] W. Wang, L. Yao, L. Chen, B. Lin, D. Cai, et al. Crossformer: A versatile vision transformer hinging on cross-scale attention. In Proc. Int. Conf. on Learning Representations (ICLR), 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.845, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[25] B. Mildenhall, P. P. Srinivasan, M. Tancik, J. T. Barron, R. Ramamoorthi, et al. Nerf: Representing scenes as neural radiance fields for view synthesis. In Proc. European Conf. on Computer Vision (ECCV), pages 405-421, 2020. 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18265" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.162 + ], + "angle": 0, + "content": "[26] M. Tancik, P. P. Srinivasan, B. Mildenhall, S. Fridovich-Keil, N. Raghavan, et al. Fourier features let networks learn high frequency functions in low dimensional domains. In Proc. Conf. on Neural Information Processing Systems (NeurIPS), pages 7537–7547, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.172, + 0.469, + 0.214 + ], + "angle": 0, + "content": "[27] N. Rahaman, A. Baratin, D. Arpit, F. Draxler, M. Lin, et al. On the spectral bias of neural networks. In Proc. Int. Conf. on Machine Learning (ICML), pages 5301-5310, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.224, + 0.469, + 0.267 + ], + "angle": 0, + "content": "[28] Z. Chen and H. Zhang. Learning implicit fields for generative shape modeling. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), pages 5939-5948, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.276, + 0.469, + 0.345 + ], + "angle": 0, + "content": "[29] J. J. Park, P. Florence, J. Straub, R. A. Newcombe, and S. Lovegrove. Deepsdf: Learning continuous signed distance functions for shape representation. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), pages 165-174, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.356, + 0.469, + 0.424 + ], + "angle": 0, + "content": "[30] L. M. Mescheder, M. Oechsle, M. Niemeyer, S. Nowozin, and A. Geiger. Occupancy networks: Learning 3d reconstruction in function space. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), pages 4460-4470, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.435, + 0.469, + 0.491 + ], + "angle": 0, + "content": "[31] K. Genova, F. Cole, A. Sud, A. Sarna, and T. A. Funkhouser. Local deep implicit functions for 3d shape. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), pages 4856-4865, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.501, + 0.469, + 0.556 + ], + "angle": 0, + "content": "[32] M. Atzmon and Y. Lipman. SAL: sign agnostic learning of shapes from raw data. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), pages 2562-2571, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.567, + 0.469, + 0.624 + ], + "angle": 0, + "content": "[33] V. Sitzmann, M. Zollhöfer, and G. Wetzstein. Scene representation networks: Continuous 3d-structure-aware neural scene representations. In Proc. Conf. on Neural Information Processing Systems (NeurIPS), pages 1119–1130, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.633, + 0.469, + 0.689 + ], + "angle": 0, + "content": "[34] C. M. Jiang, A. Sud, A. Makadia, J. Huang, M. Nießner, et al. Local implicit grid representations for 3d scenes. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), pages 6000-6009, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.699, + 0.469, + 0.755 + ], + "angle": 0, + "content": "[35] R. Chabra, J. E. Lenssen, E. Ilg, T. Schmidt, J. Straub, et al. Deep local shapes: Learning local SDF priors for detailed 3d reconstruction. In Proc. European Conf. on Computer Vision (ECCV), pages 608-625, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.765, + 0.469, + 0.82 + ], + "angle": 0, + "content": "[36] S. Peng, M. Niemeyer, L. M. Mescheder, M. Pollefeys, and A. Geiger. Convolutional occupancy networks. In Proc. European Conf. on Computer Vision (ECCV), pages 523-540, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.831, + 0.469, + 0.9 + ], + "angle": 0, + "content": "[37] M. Niemeyer, L. Mescheder, M. Oechsle, and A. Geiger. Differentiable volumetric rendering: Learning implicit 3d representations without 3d supervision. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), pages 3501-3512, 2020. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.092, + 0.892, + 0.135 + ], + "angle": 0, + "content": "[38] L. Liu, J. Gu, K. Z. Lin, T. Chua, and C. Theobalt. Neural sparse voxel fields. In Proc. Conf. on Neural Information Processing Systems (NeurIPS), 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.145, + 0.892, + 0.214 + ], + "angle": 0, + "content": "[39] J. T. Barron, B. Mildenhall, M. Tancik, P. Hedman, R. Martin-Brualla, et al. Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In Proc. IEEE Int. Conf. on Computer Vision (ICCV), pages 5835–5844, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.226, + 0.892, + 0.281 + ], + "angle": 0, + "content": "[40] S. Klocek, L. Maziarka, M. Wolczyk, J. Tabor, J. Nowak, et al. Hypernetwork functional image representation. In Artificial Neural Networks and Machine Learning, pages 496-510, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.293, + 0.892, + 0.362 + ], + "angle": 0, + "content": "[41] V. Sitzmann, J. N. P. Martel, A. W. Bergman, D. B. Lindell, and G. Wetzstein. Implicit neural representations with periodic activation functions. In Proc. Conf. on Neural Information Processing Systems (NeurIPS), pages 7462-7473, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.374, + 0.892, + 0.417 + ], + "angle": 0, + "content": "[42] K. Grm, W. J. Scheirer, and V. Struc. Face hallucination using cascaded super-resolution and identity priors. IEEE Trans. Image Processing, pages 2150-2165, 2020. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.428, + 0.892, + 0.484 + ], + "angle": 0, + "content": "[43] Z. Fu, Y. Kong, Y. Zheng, H. Ye, W. Hu, et al. Cascaded detail-preserving networks for super-resolution of document images. In Int. Conf. on Doc. Analysis Recognition, ICDAR, pages 240-245, 2019. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.495, + 0.892, + 0.551 + ], + "angle": 0, + "content": "[44] R. Basri, M. Galun, A. Geifman, D. W. Jacobs, Y. Kasten, et al. Frequency bias in neural networks for input of nonuniform density. In Proc. Int. Conf. on Machine Learning (ICML), pages 685-694, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.562, + 0.892, + 0.632 + ], + "angle": 0, + "content": "[45] M. Tancik, B. Mildenhall, T. Wang, D. Schmidt, P. P. Srinivasan, et al. Learned initializations for optimizing coordinate-based neural representations. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), pages 2846-2855, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.643, + 0.892, + 0.685 + ], + "angle": 0, + "content": "[46] D. Hendrycks and K. Gimpel. Bridging nonlinearities and stochastic regularizers with gaussian error linear units. CoRR, abs/1606.08415, 2016. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.697, + 0.892, + 0.766 + ], + "angle": 0, + "content": "[47] J. Lee-Thorp, J. Ainslie, I. Eckstein, and S. Ontanon. FNet: Mixing tokens with fourier transforms. In Proc. Conf. of North American Chapter of the Association for Computational Linguistics: Human Language Technologies (NAACL), pages 4296-4313, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.778, + 0.892, + 0.834 + ], + "angle": 0, + "content": "[48] E. Agustsson and R. Timofte. NTIRE 2017 challenge on single image super-resolution: Dataset and study. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition Workshop (CVPRW), pages 1122-1131, 2017. 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.845, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[49] M. Bevilacqua, A. Roumy, C. Guillemot, and M. Alberi-Morel. Low-complexity single-image super-resolution based on nonnegative neighbor embedding. In Proc. British Machine Vision Conf. (BMVC), pages 1-10, 2012. 5, 6" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.092, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "18266" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.147 + ], + "angle": 0, + "content": "[50] R. Zeyde, M. Elad, and M. Protter. On single image scale-up using sparse-representations. In *Curves and Surfaces*, volume 6920 of Lecture Notes in Computer Science, pages 711-730, 2010. 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.158, + 0.47, + 0.227 + ], + "angle": 0, + "content": "[51] D. R. Martin, C. C. Fowlkes, D. Tal, and J. Malik. A database of human segmented natural images and its application to evaluating segmentation algorithms and measuring ecological statistics. In Proc. IEEE Int. Conf. on Computer Vision (ICCV), pages 416-425, 2001. 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.238, + 0.47, + 0.293 + ], + "angle": 0, + "content": "[52] J. Huang, A. Singh, and N. Ahuja. Single image superresolution from transformed self-exemplars. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), pages 5197-5206, 2015. 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.303, + 0.47, + 0.344 + ], + "angle": 0, + "content": "[53] D. P. Kingma and J. Ba. Adam: A method for stochastic optimization. In Proc. Int. Conf. on Learning Representations (ICLR), 2015. 6" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.344 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.956 + ], + "angle": 0, + "content": "18267" + } + ] +] \ No newline at end of file diff --git a/2023/Cascaded Local Implicit Transformer for Arbitrary-Scale Super-Resolution/ae047e51-87a0-4eec-807f-144baf2a48b8_origin.pdf b/2023/Cascaded Local Implicit Transformer for Arbitrary-Scale Super-Resolution/ae047e51-87a0-4eec-807f-144baf2a48b8_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..9e3552df5cba9b27669f257e679e5eb7760295fb --- /dev/null +++ b/2023/Cascaded Local Implicit Transformer for Arbitrary-Scale Super-Resolution/ae047e51-87a0-4eec-807f-144baf2a48b8_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3b20305bb76090e96c03a7cae49ad3540e5a8bc6465e5f5762a0744de0fb6d4e +size 1195637 diff --git a/2023/Cascaded Local Implicit Transformer for Arbitrary-Scale Super-Resolution/full.md b/2023/Cascaded Local Implicit Transformer for Arbitrary-Scale Super-Resolution/full.md new file mode 100644 index 0000000000000000000000000000000000000000..f2b347b6779ce4fe5083b181fd62bda9816126ca --- /dev/null +++ b/2023/Cascaded Local Implicit Transformer for Arbitrary-Scale Super-Resolution/full.md @@ -0,0 +1,291 @@ +# Cascaded Local Implicit Transformer for Arbitrary-Scale Super-Resolution + +Hao-Wei Chen\*,1,2 Yu-Syuan Xu\*,2 Min-Fong Hong\* Yi-Min Tsai\* Hsien-Kai Kuo\* Chun-Yi Lee\* Elsa Lab, National Tsing Hua University \*MediaTek Inc. + +{jaroslaw1007, cylee}@gapp.nthu.edu.tw + +{Yu-Syuan.Xu, romulus.hong, Yi-Min.Tsai, Hsienkai.Kuo}@mediatek.com + +# Abstract + +Implicit neural representation has recently shown a promising ability in representing images with arbitrary resolutions. In this paper, we present a Local Implicit Transformer (LIT), which integrates the attention mechanism and frequency encoding technique into a local implicit image function. We design a cross-scale local attention block to effectively aggregate local features and a local frequency encoding block to combine positional encoding with Fourier domain information for constructing high-resolution images. To further improve representative power, we propose a Cascaded LIT (CLIT) that exploits multi-scale features, along with a cumulative training strategy that gradually increases the upsampling scales during training. We have conducted extensive experiments to validate the effectiveness of these components and analyze various training strategies. The qualitative and quantitative results demonstrate that LIT and CLIT achieve favorable results and outperform the prior works in arbitrary super-resolution tasks. + +# 1. Introduction + +Single Image Super-Resolution (SISR) is the process of reconstructing high-resolution (HR) images from their corresponding low-resolution (LR) counterparts. SISR has long been recognized as a challenging task in the low-level vision domain due to its ill-posed nature, and has attracted a number of researchers dedicated to this field of study over the past decade [1-21]. A line of SISR research referred to as 'fixed-scale SR' [1-15] focuses on extracting feature embeddings from LR images and leveraging these embeddings to upsample images with a predefined factor through learnable deconvolutions [3] or sub-pixel convolutions [4]. Despite their success, many of the proposed approaches necessitate a distinct deep neural network model for each upsampling scale, which is usually restricted to a limited selection of integers (e.g., $2 \times$ , $3 \times$ , $4 \times$ ). Such a limitation constrains the potential applications and deployment options of SISR models. To overcome this limitation, approaches for upsampling LR images in a continuous manner via a single model emerge and attracted considerable attention recently. + +Over the past few years, arbitrary-scale SR has emerged and attracted considerable attention from researchers [16-21]. Apart from the pioneering work Meta-SR [16], recent + +![](images/7dd09a46aa8a3ae0b389404446b4cb11bc731515f8fca5e1d6da6786a3498803.jpg) +LR +HR + +![](images/d4d3d990051fde0f40d410629c121267ff31938a40770d96a1caec70438d2d47.jpg) +LR + +![](images/bcf88dce662e00f036c471a554bf325f4d362e8c2be90085943be045957dd27e.jpg) +HR + +![](images/35530f43854e42ca6df6273a074902f118a82adb1809fbb418332d116d1950f5.jpg) +(a) +LIIF +Loc + +![](images/254c026769d360b5c4d72475f0f5d2476cfbb5e39efffd59a99453f68d2be3e2.jpg) +LIIF + +![](images/b71c8b963ebff36469aa32a422c08a5118dd1fb4dc208e9d66a6ecbf64f4925b.jpg) + +![](images/d8f8c1a76b4eaec2c548a4266eed42cb65b7eb618ee9845cad898259dbbac04e.jpg) +Ours + +![](images/113b3566563e8c0ffdcc6c13bc5b1e00884b78b508154fe15859ca0082a9038f.jpg) +Figure 1. An illustration and comparison of different approaches that take into account nearby pixels for continuous upsampling: (a) the local ensemble method used in [17], and (b) our proposed local attention mechanism. + +endevavors [17-21] have achieved arbitrary-scale SR by replacing the upsampling layers commonly adopted by previous approaches with local implicit image functions, and have demonstrated favorable performance. These local implicit functions employ multi-layer perceptrons (MLPs) to map 2D coordinates and corresponding latent representations to RGB values. Fig. 1 illustrates how different approaches sample latent representations based on the queried coordinates (depicted as the red dots). Fig. 1 (a) illustrates the local ensemble technique adopted by contemporary mainstream methods [17-21]. It calculates the RGB value of the queried coordinate by taking the weighted average of those of the surrounding four pixels based on their relative distances to the queried coordinate. This approach, however, does not consider contextual information and relies solely on distance. For instance, in Fig. 1, the queried coordinates are intentionally designed to lie on edges. However, merely calculating the weighted average of pixels fails to reflect the contextual information about the image content, thereby preventing the accurate capture of the necessary features for performing SR. As a result, although pixel distance plays a vital role in SR tasks, it is essential to concentrate more on the contextual information in an image. + +In light of the above observations, we propose a Local Implicit Transformer (LIT), which expands the numbers of referenced latent vectors and accounts for the feature correlation in the context by exploiting the attention mechanism [22]. LIT comprises a Cross-Scale Local Attention Block (CSLAB), a Local Frequency Encoding Block + +(LFEB), and a decoder. CSLAB generates attention maps based on the bilinearly interpolated latent vectors at the queried coordinates and the key latent vectors sampled from a grid of coordinates with relative positional bias [23, 24]. The first and second columns of Fig. 1 (b) visualize the attention maps generated by LIT, where the attention areas align closely with the edges. By applying attention maps to feature embeddings, the RGB values of the queried coordinates can be contextually predicted. Moreover, inspired by [21, 25, 26], we introduce LFEB, which projects relative coordinates into latent space to address the spectral bias problem [27] of an implicit neural function. Specifically, relative coordinates are encoded into relative positional encoding, which are multiplied with the frequency encoding extracted from the feature embedding in the Fourier domain to generate the frequency embeddings. This design enables the frequency embedding to integrate the relative positional encoding with texture information, thereby augmenting the expressivity of relative coordinates. Finally, a decoder is adopted to produce RGB values by taking advantage of the attention feature embedding and the frequency embedding. + +In order to address the issue of diverse scaling factors and achieve arbitrary-scale super-resolution, it is crucial to consider the role of upsampling factors in constructing high-resolution images within the local implicit image function. However, simultaneously training a local implicit image function with a wide range of upsampling factors (e.g., $1 \times \sim 30 \times$ ) poses significant challenges. As a result, we propose a cumulative training strategy to incrementally enhance the fuction's its representative power. The strategy initially trains the local implicit image function with small upsampling factors and then finetunes it with alternatively sampled small and large ones. Furthermore, we present Cascaded LIT (CLIT) to harness the advantages of multiscale feature embeddings, complementing missing details and information during one-step upsampling. The combination of the cumulative training strategy and CLIT enables efficient and effective handling of arbitrary-scale SR tasks. + +The main contributions of our work are summarized as follows: (1) We introduce the LIT architecture, which incorporates the local attention mechanism into arbitrary-scale SR (2) We further develop a cumulative training strategy and the cascaded framework CLIT to effectively handle large-scale upsampling. (3) We carry out comprehensive analyses of the performance impacts for LIT and CLIT. The extensive experimental findings demonstrate that the proposed LIT and CLIT are able to yield remarkable or comparable results across a wide range of benchmark datasets. + +The paper is organized as follows. Section 2 reviews the related work. Section 3 walks through the proposed LIT and CLIT frameworks and the implementation details. Section 4 presents the experimental results. Section 5 concludes. + +# 2. Related Work + +Implicit neural representation. Implicit neural representation is a technique for representing continuous-domain signals via coordinate-based multi-layer perceptrons (MLPs). Its concept has been adopted in various 3D tasks, e.g., 3D object shape modeling [28-32], 3D scene reconstruction [33-36], and 3D structure rendering [25, 37-39]. For example, NeRF [25] employs implicit neural representation to perform novel view synthesis, which maps coordinates to RGB colors for a specific scene. In the past few years, 2D applications of implicit neural representation have been attempted as well, such as image representation [40, 41] and super-resolution [17-21]. Our work is related to a technique called 'local implicit neural representation' [17, 21], which encodes LR images to feature embeddings such that similar information could be shared within local regions. Such local implicit neural representations are exploited to upscale LR images to HR ones. + +Single image super-resolution. In the past several years, various deep neural network (DNN) based architectures [1-15] have been proposed for SISR. Among these works, SR-CNN [1] pioneered the use of convolutional neural networks (CNNs) to achieve SISR in an end-to-end manner. It is later followed by several subsequent works that incorporated more complicated model architectures, such as residual blocks [6, 7], dense connections [8, 9], attention based mechanisms [10-12], or cascaded frameworks [5, 42, 43], to extract more effective feature representations for SISR. Recently, transformer-based methods [13-15] were introduced to SISR and achieved promising performance. + +Arbitrary-scale super-resolution. As discussed in Section 1, most of the contemporary SISR works limit their upsampling scales to specific integer values, and are required to train a distinct model for each upsampling scale. To overcome such a limitation, several approaches [16-21] were proposed to train a unified model for arbitrary upsampling scales. Meta-SR [16] proposed a meta-upscale module for predicting the weights of their convolutional filters from coordinates and scales. The predicted weights are then utilized to perform convolutions to generate HR images. In contrast to Meta-SR, LIIF [17] employs an MLP as a local implicit function, which takes a queried coordinate in an HR image, its nearby feature representations extracted from the corresponding LR image, as well as a cell size to predict an RGB value for that coordinate. UltraSR [18] and IPE [19] extended LIIF by replacing coordinates with the embedded ones to deal with the spectral bias issue [25, 27, 41, 44, 45] inherent in MLPs. LTE [21] further introduced a local texture estimator that transforms coordinates into Fourier domain information to enrich the representational capability of its local implicit function. Different from the above ap + +![](images/f4d95a8a3964cbfddd8daf78d2de9aac78c5aa7fa8793fc059c5fe65e9b8d4c6.jpg) +Figure 2. The proposed LIT framework. The local sampling operation samples input embeddings based on a grid of coordinates. + +proaches, our proposed methodology exploits a novel local attention mechanism and a cascaded framework to deal with the arbitrary-scale SR. In order to fairly compare with the above approaches, we similarly adopt EDSR [7], RDN [9] and SwinIR [14] as the encoders for our LIT and CLIT. + +# 3. Methodology + +In this section, we first provide an overview of the proposed LIT framework, followed by the implementation details of it and its main modules. We then discuss our cumulative training strategy, as well as the framework of CLIT. + +# 3.1. Overview of the LIT Framework + +LIT is a framework that employs a novel cross-scaled local attention mechanism and a local frequency encoding technique to perform arbitrary-scale SR tasks. Fig. 2 (a) provides an overview of the proposed framework, which aims at producing an HR image $I^{HR} \in \mathbb{R}^{r_hH \times r_wW \times 3}$ at 2D HR coordinates $\mathbf{x}^{HR} \in \mathcal{X}$ from a given LR image $I^{LR} \in \mathbb{R}^{H \times W \times 3}$ at 2D LR coordinates $\mathbf{x}^{LR} \in \mathcal{X}$ based on an arbitrary upsampling scale $\mathbf{r} = \{r_h, r_w\}$ , where $\mathcal{X}$ is the 2D coordinate space that is used to represent an image in the continuous domain. An encoder $E_{\theta}$ first extracts a feature embedding $\mathcal{Z} \in \mathbb{R}^{H \times W \times C}$ from $I^{LR}$ . The extracted $\mathcal{Z}$ is then forwarded into LIT along with the 2D coordinates of $I^{HR}$ and a cell $= (2 / s_h, 2 / s_w)$ to generate the RGB values of a residual image $I_r^{HR} \in \mathbb{R}^{r_hH \times r_wW \times 3}$ in a pixel-wise fashion. Lastly, the residual image $I_r^{HR}$ is combined with a bilinearly upsampled image $I_{\uparrow}^{HR} \in \mathbb{R}^{r_hH \times r_wW \times 3}$ via element-wise addition to derive the final HR image $I^{HR}$ . + +# 3.2. Local Implicit Transformer + +LIT is developed for mapping any 2D coordinate in the continuous image domain to an RGB color. As highlighted in Fig. 2 (b), it is composed of a cross-scale local attention + +block (CSLAB), a local frequency encoding block (LFB), and a decoder $D_{\phi}$ paraterizerized by $\phi$ . The former two blocks are responsible for estimating a local latent embedding $\tilde{z} \in \mathbb{R}^{G_h G_w \times C}$ and a local frequency embedding $\tilde{f} \in \mathbb{R}^{G_h G_w \times C}$ , respectively, where $G_h$ and $G_w$ denote the height and width of local grids employed for performing local coordinate sampling, as depicted in Fig. 3. On the other hand, $D_{\phi}$ utilizes these embeddings along with the provided cell to generate $I_r^{HR}$ . More specifically, LIT first projects $\mathcal{Z}$ using four separate convolutional layers to obtain four latent embeddings, corresponding to query $q$ , key $k$ , value $v$ , and frequency $f$ . Based on a queried HR coordinate $x_q \in \mathbf{x}^{HR}$ , CSLAB and LFB estimate $\tilde{z}$ and $\tilde{f}$ as follows: + +$$ +\tilde {z} = C S L A B (\delta x, q, k, v), \tag {1} +$$ + +$$ +\tilde {f} = L F E B (\delta x, f), \tag {2} +$$ + +$$ +\delta \mathbf {x} = \left\{x _ {q} - x ^ {(i, j)} \right\} _ {i \in \{1, 2, \dots , G _ {h} \}, j \in \{1, 2, \dots , G _ {w} \}}, \tag {3} +$$ + +where $x^{(i,j)}\in \mathbf{x}^{LR}$ denotes an LR coordinate in the local grid indexed by $(i,j)$ , and $\delta \mathbf{x}$ represents the set of local relative coordinates defined by Eq. (3). The local grid is sampled in a manner that positions its center $x^{(\lfloor G_h / 2\rfloor +1,\lfloor G_w / 2\rfloor +1)}$ at the LR coordinate closest to $x_{q}$ . The query latent vector $q\in \mathbb{R}^{1\times C}$ at the HR coordinate $x_{q}$ is computed by bilinear interpolation, while the remainder of the local latent embeddings $k\in \mathbb{R}^{G_hG_w\times C}$ , $v\in \mathbb{R}^{G_hG_w\times C}$ , and $f\in \mathbb{R}^{G_hG_w\times C}$ are sampled at the local LR coordinates $\mathbf{x} = \{x^{i,j}\}_{i\in \{1,2,\dots,G_h\},j\in \{1,2,\dots,G_w\}}$ . With the local latent embedding $\tilde{z}$ and the local frequency embedding $\tilde{f}$ , the function of $D_{\phi}$ is formulated as follows: + +$$ +I ^ {r} \left(x _ {q}\right) = D _ {\phi} \left(\tilde {z}, \tilde {f}, c\right), \tag {4} +$$ + +where $I^{r}(x_{q})$ is the predicted RGB value at the queried coordinate $x_{q}$ , and $c = \{HR_{\Delta h}, HR_{\Delta w}\}$ denotes the cell that + +![](images/4b721537abf77883775292eb0d479c54213e81a052d46033b70500a91fbec677.jpg) +Figure 3. An illustration of the proposed local coordinate sampling scheme. The red rectangular region, outlined by a dashed line, represents the local grid with dimensions $G_{h} \times G_{w}$ . The red dots indicate the sampled local LR coordinates, while the gray dots correspond to the HR coordinates. The grid dimensions $LR_{\Delta h,w}$ and $HR_{\Delta h,w}$ represent the unit sizes of pixels in LR and HR images, respectively. + +represents the height and width of a pixel in an HR image, as illustrated in Fig. 3. $D_{\phi}$ is implemented as a five-layer MLP utilizing Gaussian Error Linear Unit (GELU) activation [46], and is employed consistently across all images. + +Cross-scale local attention block. LIT exploits CSLAB to perform a local attention mechanism over a local grid to generate a local latent embedding $\tilde{z}$ for each HR coordinate, as illustrated in Fig. 2 (c). CSLAB first calculates the inner product of $q$ and $k$ , adds the relative positional bias $B$ to the result, and obtains an attention matrix. This attention matrix is subsequently normalized by a Softmax operation to produce a local attention map. Finally, CSLAB performs element-wise multiplication of $v$ and the local attention map to derive $\tilde{z}$ . The overall procedure is formulated as follows: + +$$ +\begin{array}{l} \tilde {z} = \operatorname {s o f t m a x} \left(\frac {q k ^ {\top}}{\sqrt {C}} + B\right) \times v, (5) \\ B = F C (\gamma (\delta \mathbf {x})), (6) \\ \gamma (\delta \mathbf {x}) = \left[ \sin \left(2 ^ {0} \delta \mathbf {x}\right), \cos \left(2 ^ {0} \delta \mathbf {x}\right), \dots \right. (7) \\ , \sin \left(2 ^ {L - 1} \delta \mathbf {x}\right), \cos \left(2 ^ {L - 1} \delta \mathbf {x}\right) ], \\ \end{array} +$$ + +where $C$ denotes the channel dimension of the local key latent embedding $k$ , $FC$ represents a fully-connected layer, $\gamma$ is the positional encoding function, and $L$ is a hyperparameter. In this work, $L$ is set to 10, and the multi-head attention mechanism adopted is formulated in Eq. (8) as follows: + +$$ +\tilde {z} = \operatorname {c o n c a t} \left(\operatorname {s o f t m a x} \left(\frac {q _ {i} k _ {i} ^ {\top}}{\sqrt {C / H}} + B _ {i}\right) \times v _ {i}\right), \tag {8} +$$ + +where $H$ is the number of attention heads and $i\in [1,\dots ,H]$ + +![](images/a9ca32d0bd97b5f4a24bbb9fda0903384bfb89b5b21fc8834c22b9195181d570.jpg) +Figure 4. An overview of the proposed CLIT framework. + +Local frequency encoding block. Fig. 2 (d) illustrates the local frequency encoding block, which incorporates dominant-frequency components to generate the local frequency embedding $\tilde{f}$ . In order to account for dominant-frequency information, LFBB first encodes local frequency latent embedding $f$ using Fast Fourier Transform (FFT), which is then multiplied by the relative positional encoding to generate $\tilde{f}$ . The resultant local frequency embedding $\tilde{f}$ is then concatenated with the flattened local latent embedding $\tilde{z}$ to form a mixing embedding, which can be used by the decoder consisting of MLPs to predict RGB values [47]. + +# 3.3. Cumulative Training Strategy + +In this section, we discuss our proposed cumulative training strategy, which is developed for enhancing the performance of arbitrary-scale SR. The cumulative training strategy focuses on the schedule of the cell sizes selected during the training phase, as "cell decoding" has been recognized as an essential input to a local implicit image function [17]. Recent studies [17, 19] have observed that the effect of cell decoding on the performance of arbitrary-scale SR is prominent for in-distribution upsampling, but degrades significantly for out-of-distribution large-scale upsampling. Such trends are demonstrated in Table 3 of [17] and Table 4 of [19]. The authors in [21] also constrained their cell sizes to in-distribution ranges to mitigate the negative impact during evaluation. To overcome the degradation issue for out-of-distribution cell sizes, incorporating large cell sizes during training appears to be a promising solution. However, simply training the local implicit image function with a diverse range of cell sizes at once leads to a performance drop. According to our experimental observations presented in Section 4.4, we find that training the local implicit image function by alternatively switching between large and small cells offers positive impacts on the performance. Based on the above observations, we propose a cumulative training strategy which first trains the local implicit image function with large cell sizes, and finetunes it with the alternative training strategy to improve the performance on different upsampling scales. More quantitative results of our training strategies can be found in Section 4.4. + +
Method×2×3×4×6×12×18×24×30
Bicubic [7]31.0128.2226.6624.8222.2721.0020.1919.59
EDSR-baseline [7]34.5530.9028.94-----
EDSR-baseline-Meta-SR [16, 17]34.6430.9328.9226.6123.5522.0321.0620.37
EDSR-baseline-LIIF [17]34.6730.9629.0026.7523.7122.1721.1820.48
EDSR-baseline-UltraSR [18]34.6931.0229.0526.8123.7522.2121.2120.51
EDSR-baseline-IPE [19]34.7231.0129.0426.7923.7522.2121.2220.51
EDSR-baseline-LTE [21]34.7231.0229.0426.8123.7822.2321.2420.53
EDSR-baseline-CLIT (Ours)34.8131.1229.1526.9223.8322.2921.2620.53
RDN-Meta-SR [16, 17]35.0031.2729.2526.8823.7322.1821.1720.47
RDN-LIIF [17]34.9931.2629.2726.9923.8922.3421.3120.59
RDN-UltraSR [18]35.0031.3029.3227.0323.7322.3621.3320.61
RDN-IPE [19]35.0431.3229.3227.0423.9322.3821.3420.63
RDN-LTE [21]35.0431.3229.3327.0423.9522.4021.3620.64
RDN-CLIT (Ours)35.1031.3929.3927.1224.0122.4521.3820.64
SwinIR-MetaSR [16, 21]35.1531.4029.3326.9423.8022.2621.2620.54
SwinIR-LIIF [17, 21]35.1731.4629.4627.1524.0222.4321.4020.67
SwinIR-LTE [21]35.2431.5029.5127.2024.0922.5021.4720.73
SwinIR-CLIT (Ours)35.2931.5529.5527.2624.1122.5121.4520.70
+ +Table 1. The average PSNR (dB) on the DIV2K [48] validation set. The results are obtained from the original manuscripts [17-19, 21]. The best and second-best performing results are highlighted by the red and blue colors, respectively. + +# 3.4. Cascaded Local Implicit Transformer + +Previous works [17, 21] typically address arbitrary-scale SR through a single step of upsampling. However, one-step upsampling struggles to reconstruct an HR image when the upsampling scale is large [5, 42, 43]. In light of this, we propose a cascaded upsampling strategy, called "Cascaded LIT(CLIT)" to predict residual images from multi-scale feature embeddings. In an $N$ -branched CLIT, the multi-scale feature embeddings $\mathcal{Z}^1, \mathcal{Z}^2, \dots, \mathcal{Z}^N$ are derived as follows: + +$$ +\mathcal {Z} ^ {N} = \mathcal {Z} \uparrow_ {s ^ {1} \times s ^ {2} \times \dots \times s ^ {N - 1}}, \tag {9} +$$ + +where $s^1 = 1$ and $s\in \mathbf{s}$ + +where $\uparrow$ is a bilinear upsampling function and $s$ is a set of scaling factors, which are configurable hypermeters. For a branch $i,i\in [1,\dots,N]$ , $\mathrm{LIT}^i$ estimates the residual image $I_r^i$ from the feature embedding $\mathcal{Z}^i$ with the coordinate and the corresponding cells. Lastly, the final HR image $I^{HR}\in \mathbb{R}^{r_hH\times r_wW\times 3}$ can be estimated as the following equation: + +$$ +I ^ {H R} = \lambda^ {N - 1} I _ {r} ^ {1} + \lambda^ {N - 2} I _ {r} ^ {2} + \dots + \lambda^ {0} I _ {r} ^ {N} + I _ {\uparrow} ^ {H R}. \tag {10} +$$ + +where $\lambda$ is a discount factor, with a default value of 0.75. During the training phase, CLIT is trained using the proposed cumulative training strategy. Initially, $\mathrm{LIT}^1$ is trained with the strategy employed by [17]. Subsequently, $\mathrm{LIT}^1$ is finetuned and $\mathrm{LIT}^2$ is initialized by applying the alternative training strategy. By incrementally incorporating LITs into CLIT, the performance is progressively enhanced. The details of the training strategies are specified in Section 4. + +# 4. Experimental Results + +In this section, we present the experimental results and discuss their implications. We begin with a brief introduction to our experimental setup in Section 4.1. Following that, we evaluate our CLIT with different datasets in Section 4.2. The, Section 4.3 showcases the learned attention maps. Finally, ablation studies for various configurations of the proposed CLIT and LIT are compared in Section 4.4. + +# 4.1. Experimental Setup + +Dataset. We use the DIV2K dataset [48] for network training. It consists of 1,000 images in 2K resolutions and provides low-resolution counterparts with down-sampling scales, $\times 2$ , $\times 3$ , $\times 4$ , which are generated by the bicubic interpolation method. On the other hand, we also evaluate the performance on the validation set of DIV2K [48], Set5 [49], Set14 [50], B100 [51] and Urban100 [52] in terms of peak signal-to-noise (PSNR) values. + +Training details During training, we feed batches of size $48 \times 48$ low-resolution images into the framework, following the prior works [7]. For each batch, a single upsampling scale is sampled from a uniform distribution $r \sim \mathcal{U}(1,4)$ . With single upsampling scales $s$ , a batch of HR images is cropped into patches of size $48r \times 48r$ , while the corresponding LR images are cropped into patches of $48 \times 48$ . The patches are augmented by randomly horizontal flipping, vertical flipping, and $90^{\circ}$ rotating. Then, we sample $48^2$ pixels (coordinate-RGB pairs) on each HR patch as the ground-truths. We set the batch size to 32 and use + +
MethodSet5Set14B100Urban100
×2×3×4×6×8×2×3×4×6×8×2×3×4×6×8×2×3×4×6×8
RDN [9]38.2434.7132.47--34.0130.5728.81--32.3429.2627.72--32.8928.8026.61--
RDN-Meta-SR [16, 17]38.2234.6332.3829.0426.9633.9830.5428.7826.5124.9732.3329.2627.7125.9024.8332.9228.8226.5523.9922.59
RDN-LLIF [17]38.1734.6832.5029.1527.1433.9730.5328.8026.6425.1532.3229.2627.7425.9824.9132.8728.8226.6824.2022.79
RDN-UltraSR [18]38.2134.6732.4929.3327.2433.9730.5928.8626.6925.2532.3529.2927.7726.0124.9632.9728.9226.7824.3022.87
RDN-IPE [19]38.1134.6832.5129.2527.2233.9430.4728.7526.5825.0932.3129.2827.7626.0024.9332.9728.8226.7624.2622.87
RDN-LTE [21]38.2334.7232.6129.3227.2634.0930.5828.8826.7125.1632.3629.3027.7726.0124.9533.0428.9726.8124.2822.88
RDN-CLIT (Ours)38.2634.7932.6929.5427.3434.0930.6928.9326.8325.3632.3929.3327.8026.0725.0033.1429.0526.9324.4423.04
SwinIR [14]38.3534.8932.72--34.1430.7728.94--32.4429.3727.83--33.4029.2927.07--
SwinIR-MetaSR [16, 21]38.2634.7732.4729.0927.0234.1430.6628.8526.5825.0932.3929.3127.7525.9424.8633.2929.1226.7624.1622.75
SwinIR-LLIF [17]38.2834.8732.7329.4627.3634.1430.7528.9826.8225.3432.3929.3427.8426.0725.0133.3629.3327.1524.5923.14
SwinIR-LTE [21]38.3334.8932.8129.5027.3534.2530.8029.0626.8625.4232.4429.3927.8626.0925.0333.5029.4127.2424.6223.17
SwinIR-CLIT (Ours)38.4134.9732.8629.6927.6234.2730.8529.0826.9425.5532.4629.4227.9126.1525.0933.5629.4327.2524.7723.33
+ +Table 2. The average PSNR (dB) on Set5 [49], Set14 [50], B100 [51], and Urban100 [52]. The results are obtained from the original manuscripts [17-19, 21]. The best and second-best performing results are highlighted by the red and blue colors, respectively. + +![](images/6e0bd67caa71965e7c5090c88a45e21384e9a979558071c4cab1b76adb97b29a.jpg) +Figure 5. The qualitative results of LIIF [17], LTE [21], and our CLIT with using RDN [9] as the encoder. + +the Adam optimizer [53] together with L1 loss for training. We train LIT for 1,000 epochs, and the learning rate is initialized at $1e^{-4}$ and decayed by a factor 0.5 at [200, 400, 600, 800] epochs. For cumulative training of CLIT, as discussed in Section 3.4, we sample $N$ scale factors $\{s^1, s^2, \dots, s^N\}$ from the distribution $\mathcal{U}(1,4)$ according the number of LITs $N$ in the train step. The total upsampling scale $r = s^1 \times s^2 \dots \times s^N$ is the product of all scale factors. If $48r \times 48r$ is greater than the whole HR image, we clip the scale factor of stage 1. To train $N$ LITs, we fine-tune the model for $500 \times N$ epochs, and the learning rate is initialized at $1e^{-4}$ and decayed by a factor 0.5 at $[100 \times N, 200 \times N, 300 \times N, 400 \times N]$ epochs. + +# 4.2. Validation of CLIT + +Quantitative results. We first compare our proposed CLIT to existing local implicit neural representation methods for arbitrary-scale SR, including LIIF [17], UltraSR [18], IPE [19], and LTE [21]. Table 1 summarizes the quantitative results in terms of PSNR(dB) on DIV2k [48]. As shown in Table 1, CLIT achieves the best performance when EDSR-Baseline [7], RDN [9] are used as the encoders. These results demonstrate the advantage of CLIT. + +Table 2 compares CLIT to prior works [17-19, 21] on widely-used datasets, including Set5 [49], Set14 [50], B100 [51], and Urban100 [52], with RDN and SwinIR. + +![](images/ebe25e5c577baf8ff4da43660cc0c4fc9ae38ee8c4d5fe45500e97fbf15af4e7.jpg) +Figure 6. The qualitative results of LIIF [17], LTE [21], and our CLIT using RDN [9] as the encoder and non-integer upsampling scales. + +Note that RDN [9] is trained and evaluated for specific upsampling scales. CLIT outperforms those existing methods in most cases across all datasets and scales, even achieving a 0.26dB PSNR improvement on Set5 for the $\times 8$ scale. + +Qualitative results. Fig. 5 compares the qualitative results of CLIT with the baseline methods LIIF [17] and LTE [21], on various dataset DIV2k [48], Set14 [50], B100 [51], and Urban100 [52] with variant upsampling scales. The official codes provided by them were used to produce the results. In the first row, the SR visualization results with a $\times 12$ upsampling scale are depicted. It can be observed that both LIIF and LTE struggle to reconstruct the letters continuously, whereas the CLIT result demonstrates continuity for these alphabets, particularly for 'M' and 'S'. In the second row, the stripes on the zebra appear blurry in the LIIF and LTE results, whereas they are more distinct in the CLIT result. In the third row, the cross pattern on the tablecloth is not clearly rendered by LIIF and LTE. In contrast, CLIT is able to successfully generate clean crosses and sharp lines. In the fourth row, despite the blurriness of the original LR image, CLIT captures the texture and produces straight lines on the flag, showcasing its effectiveness. + +Fig. 6 presents the results of text image enhancement employing progressively increasing non-integer upsampling factors, alongside a comparison with earlier methods, LIIF [17] and LTE [21]. The input image first undergoes a downscaling process by a factor of 2.2 to generate an LR image. Subsequently, the LR image is then upscaled using predetermined non-integer factors $\times 1.6$ , $\times 2.5$ , $\times 3.4$ , $\times 4.3$ . + +![](images/b8439088d65aa6981d1b276c707e5fe6fdc5623441f9c4b85d08480d415e21a2.jpg) +Figure 7. A visualization of the local attention maps of the coordinates highlight as red dots. + +It can be observed that our proposed model effectively captures the patterns of text within the image, and accurately estimates words and numbers with enhanced sharpness and clarity. The enhancement becomes particularly evident in the word 'infinity' in the first row of contexts, 'SUNSHINE' in the fifth row, and the number '0' in the final row, as each of these elements demonstrates considerable improvements. + +# 4.3. Visualization Local Attention Maps + +In Fig. 7, we provide visualizations of local attention maps. The LR images are generated by applying bicubic downsampling to HR images with scale factors + +
In-distributionOut-of-distribution
×2×3×4×6×8×12
LIT34.7931.0729.1026.8425.4723.77
LIT (-a)34.7131.0029.0226.7625.4023.70
LIT (-f)34.7831.0729.0926.8325.4623.77
LIT (-c)34.6531.0029.0426.8025.4523.77
LIT (+e)34.7931.0829.1126.8525.4823.78
+ +$\{ \times 4.5, \times 6, \times 6 \}$ . Subsequently, the proposed CLIT with RDN [9] encoder is utilized to produce HR predictions. It can be observed that the attention maps closely align with the edges, indicating that the cross-scale local attention block effectively captures similar latent codes within a local area. This substantiates that our proposed design enables the generation of SR images from LR counterparts, ultimately producing clean and sharp edges in the output. + +# 4.4. Ablation Studies + +In this section, we present a series of ablation analyses to substantiate the design decisions proposed in this paper. All of the ablation experiments are conducted on the DIV2K [48] validation set, utilizing the EDSR-baseline [7] as the encoder and employing the PSNR metric for evaluation. + +Validation of the design choices. Table 3 presents a summary of the quantitative contributions associated with each LIT component. A significant improvement is observed when adopting the cross-scale local attention block, while a relatively minor gain with the frequency encoding block by comparing LTE with LTE (-a) and LTE (-f), respectively. In addition, LTE (-c) reveals that removing cell decoding leads to more severe performance degradation for in-distribution upsampling scales compared to out-of-distribution counterparts, highlighting the importance of cell decoding for in-distribution upsampling scales. Finally, incorporating the local ensemble technique in LIT (+e) results in a further, albeit modest, enhancement of the overall performance. + +The effectiveness of the local grid. Table 4 provides a performance comparison of various local grid sizes employed within the local coordinate sampling scheme, as discussed in Section 3. The results presented in Table 4 reveal that increasing the local grid size contributes to performance enhancements, albeit at the cost of extended training times. As a result, this study adopts a $7 \times 7$ local grid size to strike an ideal balance between effectiveness and efficiency. + +Analysis on the training strategy. Table 5 presents a quantitative comparison of the proposed cumulative training strategy with the other training strategies for training an LIT. The baseline strategy trains LIT using upsampling scales sampled from a uniform distribution, $r \sim \mathcal{U}(1,4)$ . Expanding the sampling scale distribution to $r \sim \mathcal{U}(1,12)$ improves the performance of large-scale upsampling while + +Table 3. The average PSNR (dB) for different design choices. The notations -a/f/c denote the removal of the cross-scale local attention block, frequency encoding block, and cell size, respectively, and +e denotes the addition of local ensemble. The best performing one is indicated in bolded. + +
Local gird +Gh × GwIn-distributionOut-of-distributionTraining time
×2×3×4×6×8×12
1 × 134.7131.0029.0226.7625.4023.7020.9
3 × 334.7631.0429.0726.8125.4623.7623.2
5 × 534.7831.0729.0826.8325.4723.7828.4
7 × 734.7931.0729.1026.8425.4723.7733.6
9 × 934.7931.0629.1026.8425.4723.7753.0
15 × 1534.8331.1029.1126.8625.4923.7876.1
25 × 2534.8631.1329.1526.8825.5023.80190.9
+ +Table 4. The average PSNR (dB) and the training time (hour) of multiple local grid sizes. The bolded numbers correspond to the best performance. + +
Training strategy×2×3×4×6×8×12
Training with r ~ U(1, 4)34.7931.0729.1026.8425.4723.77
Training with r ~ U(1, 12)34.6931.0629.1126.8725.5223.82
Alternative training strategy34.7531.0929.1226.8825.5223.81
Cumulative training strategy34.7831.1029.1426.8925.5423.83
+ +Table 5. The average PSNR (dB) of different training strategies. The best performing results are indicated in **bold**. + +compromising the performance of small-scale upsampling. To achieve high-quality results across all upsampling scale ranges, the alternative training strategy trains LIT alternatively by switching the sampling scales between $r \sim \mathcal{U}(1,4)$ and $r \sim \mathcal{U}(4,12)$ . On the other hand, the proposed cumulative training strategy first trains LIT with $r \sim \mathcal{U}(1,4)$ , followed by fine-tuning using the alternative training strategy. This progressive approach improves the performance of LIT by gradually broadening the distribution of the upsampling scales. Comparing the results of the cumulatively trained LIT with the results of EDSR-baseline-CLIT in Table 1 reveals that the proposed cumulative training strategy works for both LIT and CLIT. Moreover, the results from these tables suggest that the cascaded framework is more effective in addressing arbitrary-scale SR tasks. + +# 5. Conclusion + +In this paper, we introduced the attention mechanism and frequency encoding technique to address arbitrary-scale SR tasks. To achieve this objective, we proposed the LIT framework, which consists of a cross-scale local attention block and a local frequency encoding block. The former is designed to find the local latent embedding to reconstruct the corresponding RGB value, while the latter focuses on encoding coordinates with the frequency information derived from the feature embeddings. To enhance the capability of capturing fine details, we proposed CLIT and the corresponding cumulative training strategy that trains the model with progressively increasing upsampling scales. Based on the experimental results, both quantitative and qualitative assessments showcase the superior performance of LIT and CLIT in comparison to existing methods. Furthermore, our comprehensive analyses validated the effectiveness of the training strategies and the components employed by them. + +# Acknowledgements + +The authors gratefully acknowledge the support from the National Science and Technology Council (NSTC) in Taiwan under grant numbers MOST 111-2223-E-007-004-MY3 and MOST 111-2628-E-007-010, as well as the financial and technical support from MediaTek Inc., Taiwan. + +# References + +[1] C. Dong, C. C. Loy, K. He, and X. Tang. Image superresolution using deep convolutional networks. IEEE Trans. Pattern Analysis and Machine Intelligence (TPAMI), 38(2): 295-307, 2016. 1, 2 +[2] J. Kim, J. K. Lee, and K. M. Lee. Accurate image superresolution using very deep convolutional networks. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), pages 1646-1654, 2016. +[3] C. Dong, C. C. Loy, and X. Tang. Accelerating the superresolution convolutional neural network. In Proc. European Conf. on Computer Vision (ECCV), pages 391-407, 2016. 1 +[4] W. Shi, J. Caballero, F. Huszar, J. Totz, A. P. Aitken, et al. Real-time single image and video super-resolution using an efficient sub-pixel convolutional neural network. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), pages 1874-1883, 2016. 1 +[5] W. Lai, J. Huang, N. Ahuja, and M. Yang. Fast and accurate image super-resolution with deep laplacian pyramid networks. IEEE Trans. Pattern Analysis and Machine Intelligence (TPAMI), 41(11):2599-2613, 2019. 2, 5 +[6] C. Ledig, L. Theis, F. Huszar, J. Caballero, A. Cunningham, et al. Photo-realistic single image super-resolution using a generative adversarial network. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), pages 105-114, 2017. 2 +[7] B. Lim, S. Son, H. Kim, S. Nah, and K. M. Lee. Enhanced deep residual networks for single image super-resolution. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition Workshop (CVPRW), pages 1132-1140, 2017. 2, 3, 5, 6, 8 +[8] X. Wang, K. Yu, S. Wu, J. Gu, Y. Liu, et al. Esrgan: Enhanced super-resolution generative adversarial networks. In Proc. European Conf. on Computer Vision Workshop (ECCVW), pages 63-79, 2018. 2 +[9] Y. Zhang, Y. Tian, Y. Kong, B. Zhong, and Y. Fu. Residual dense network for image super-resolution. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), pages 2472-2481, 2018. 2, 3, 6, 7, 8 +[10] Y. Zhang, K. Li, K. Li, L. Wang, B. Zhong, et al. Image super-resolution using very deep residual channel attention networks. In Proc. European Conf. on Computer Vision (ECCV), pages 294-310, 2018. 2 +[11] T. Dai, J. Cai, Y. Zhang, S. Xia, and L. Zhang. Second-order attention network for single image super-resolution. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), pages 11065-11074, 2019. +[12] Y. Mei, Y. Fan, Y. Zhou, L. Huang, T. S. Huang, et al. Image super-resolution with cross-scale non-local attention and exhaustive self-exemplars mining. In Proc. IEEE Conf. + +on Computer Vision and Pattern Recognition (CVPR), pages 5689-5698, 2020. 2 +[13] H. Chen, Y. Wang, T. Guo, C. Xu, Y. Deng, et al. Pre-trained image processing transformer. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), pages 12294-12305, 2021. 2 +[14] J. Liang, J. Cao, G. Sun, K. Zhang, L. Van Gool, et al. Swinir: Image restoration using swim transformer. In Proc. IEEE Int. Conf. on Computer Vision Workshop (ICCVW), pages 1833-1844, 2021. 3, 6 +[15] X. Chen, X. Wang, J. Zhou, and C. Dong. Activating more pixels in image super-resolution transformer. CoRR, abs/2205.04437, 2022. 1, 2 +[16] X. Hu, H. Mu, X. Zhang, Z. Wang, T. Tan, et al. Meta-sr: A magnification-arbitrary network for super-resolution. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), pages 1575-1584, 2019. 1, 2, 5, 6 +[17] Y. Chen, S. Liu, and X. Wang. Learning continuous image representation with local implicit image function. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), pages 8628-8638, 2021. 1, 2, 4, 5, 6, 7 +[18] X. Xu, Z. Wang, and H. Shi. Ultrasr: Spatial encoding is a missing key for implicit image function-based arbitrary-scale super-resolution. CoRR, abs/2103.12716, 2021. 2, 5, 6 +[19] Y. Liu, Y. Guo, and S. Zhang. Enhancing multi-scale implicit learning in image super-resolution with integrated positional encoding. CoRR, abs/2112.05756, 2021. 2, 4, 5, 6 +[20] J. Yang, S. Shen, H. Yue, and K. Li. Implicit transformer network for screen content image continuous super-resolution. In Proc. Conf. on Neural Information Processing Systems (NeurIPS), pages 13304-13315, 2021. +[21] J. Lee and K. H. Jin. Local texture estimator for implicit representation function. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), pages 1929-1938, 2022. 1, 2, 4, 5, 6, 7 +[22] A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, et al. Attention is all you need. In Proc. Conf. on Neural Information Processing Systems (NeurIPS), pages 5998-6008, 2017. 1 +[23] Z. Liu, Y. Lin, Y. Cao, H. Hu, Y. Wei, et al. Swin transformer: Hierarchical vision transformer using shifted windows. In Proc. IEEE Int. Conf. on Computer Vision (ICCV), 2021. 2 +[24] W. Wang, L. Yao, L. Chen, B. Lin, D. Cai, et al. Crossformer: A versatile vision transformer hinging on cross-scale attention. In Proc. Int. Conf. on Learning Representations (ICLR), 2022. 2 +[25] B. Mildenhall, P. P. Srinivasan, M. Tancik, J. T. Barron, R. Ramamoorthi, et al. Nerf: Representing scenes as neural radiance fields for view synthesis. In Proc. European Conf. on Computer Vision (ECCV), pages 405-421, 2020. 2 + +[26] M. Tancik, P. P. Srinivasan, B. Mildenhall, S. Fridovich-Keil, N. Raghavan, et al. Fourier features let networks learn high frequency functions in low dimensional domains. In Proc. Conf. on Neural Information Processing Systems (NeurIPS), pages 7537–7547, 2020. 2 +[27] N. Rahaman, A. Baratin, D. Arpit, F. Draxler, M. Lin, et al. On the spectral bias of neural networks. In Proc. Int. Conf. on Machine Learning (ICML), pages 5301-5310, 2019. 2 +[28] Z. Chen and H. Zhang. Learning implicit fields for generative shape modeling. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), pages 5939-5948, 2019. 2 +[29] J. J. Park, P. Florence, J. Straub, R. A. Newcombe, and S. Lovegrove. Deepsdf: Learning continuous signed distance functions for shape representation. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), pages 165-174, 2019. +[30] L. M. Mescheder, M. Oechsle, M. Niemeyer, S. Nowozin, and A. Geiger. Occupancy networks: Learning 3d reconstruction in function space. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), pages 4460-4470, 2019. +[31] K. Genova, F. Cole, A. Sud, A. Sarna, and T. A. Funkhouser. Local deep implicit functions for 3d shape. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), pages 4856-4865, 2020. +[32] M. Atzmon and Y. Lipman. SAL: sign agnostic learning of shapes from raw data. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), pages 2562-2571, 2020. 2 +[33] V. Sitzmann, M. Zollhöfer, and G. Wetzstein. Scene representation networks: Continuous 3d-structure-aware neural scene representations. In Proc. Conf. on Neural Information Processing Systems (NeurIPS), pages 1119–1130, 2019. 2 +[34] C. M. Jiang, A. Sud, A. Makadia, J. Huang, M. Nießner, et al. Local implicit grid representations for 3d scenes. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), pages 6000-6009, 2020. +[35] R. Chabra, J. E. Lenssen, E. Ilg, T. Schmidt, J. Straub, et al. Deep local shapes: Learning local SDF priors for detailed 3d reconstruction. In Proc. European Conf. on Computer Vision (ECCV), pages 608-625, 2020. +[36] S. Peng, M. Niemeyer, L. M. Mescheder, M. Pollefeys, and A. Geiger. Convolutional occupancy networks. In Proc. European Conf. on Computer Vision (ECCV), pages 523-540, 2020. 2 +[37] M. Niemeyer, L. Mescheder, M. Oechsle, and A. Geiger. Differentiable volumetric rendering: Learning implicit 3d representations without 3d supervision. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), pages 3501-3512, 2020. 2 + +[38] L. Liu, J. Gu, K. Z. Lin, T. Chua, and C. Theobalt. Neural sparse voxel fields. In Proc. Conf. on Neural Information Processing Systems (NeurIPS), 2020. +[39] J. T. Barron, B. Mildenhall, M. Tancik, P. Hedman, R. Martin-Brualla, et al. Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In Proc. IEEE Int. Conf. on Computer Vision (ICCV), pages 5835–5844, 2021. 2 +[40] S. Klocek, L. Maziarka, M. Wolczyk, J. Tabor, J. Nowak, et al. Hypernetwork functional image representation. In Artificial Neural Networks and Machine Learning, pages 496-510, 2019. 2 +[41] V. Sitzmann, J. N. P. Martel, A. W. Bergman, D. B. Lindell, and G. Wetzstein. Implicit neural representations with periodic activation functions. In Proc. Conf. on Neural Information Processing Systems (NeurIPS), pages 7462-7473, 2020. 2 +[42] K. Grm, W. J. Scheirer, and V. Struc. Face hallucination using cascaded super-resolution and identity priors. IEEE Trans. Image Processing, pages 2150-2165, 2020. 2, 5 +[43] Z. Fu, Y. Kong, Y. Zheng, H. Ye, W. Hu, et al. Cascaded detail-preserving networks for super-resolution of document images. In Int. Conf. on Doc. Analysis Recognition, ICDAR, pages 240-245, 2019. 2, 5 +[44] R. Basri, M. Galun, A. Geifman, D. W. Jacobs, Y. Kasten, et al. Frequency bias in neural networks for input of nonuniform density. In Proc. Int. Conf. on Machine Learning (ICML), pages 685-694, 2020. 2 +[45] M. Tancik, B. Mildenhall, T. Wang, D. Schmidt, P. P. Srinivasan, et al. Learned initializations for optimizing coordinate-based neural representations. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), pages 2846-2855, 2021. 2 +[46] D. Hendrycks and K. Gimpel. Bridging nonlinearities and stochastic regularizers with gaussian error linear units. CoRR, abs/1606.08415, 2016. 4 +[47] J. Lee-Thorp, J. Ainslie, I. Eckstein, and S. Ontanon. FNet: Mixing tokens with fourier transforms. In Proc. Conf. of North American Chapter of the Association for Computational Linguistics: Human Language Technologies (NAACL), pages 4296-4313, 2022. 4 +[48] E. Agustsson and R. Timofte. NTIRE 2017 challenge on single image super-resolution: Dataset and study. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition Workshop (CVPRW), pages 1122-1131, 2017. 5, 6, 7, 8 +[49] M. Bevilacqua, A. Roumy, C. Guillemot, and M. Alberi-Morel. Low-complexity single-image super-resolution based on nonnegative neighbor embedding. In Proc. British Machine Vision Conf. (BMVC), pages 1-10, 2012. 5, 6 + +[50] R. Zeyde, M. Elad, and M. Protter. On single image scale-up using sparse-representations. In *Curves and Surfaces*, volume 6920 of Lecture Notes in Computer Science, pages 711-730, 2010. 5, 6, 7 +[51] D. R. Martin, C. C. Fowlkes, D. Tal, and J. Malik. A database of human segmented natural images and its application to evaluating segmentation algorithms and measuring ecological statistics. In Proc. IEEE Int. Conf. on Computer Vision (ICCV), pages 416-425, 2001. 5, 6, 7 +[52] J. Huang, A. Singh, and N. Ahuja. Single image superresolution from transformed self-exemplars. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), pages 5197-5206, 2015. 5, 6, 7 +[53] D. P. Kingma and J. Ba. Adam: A method for stochastic optimization. In Proc. Int. Conf. on Learning Representations (ICLR), 2015. 6 \ No newline at end of file diff --git a/2023/Cascaded Local Implicit Transformer for Arbitrary-Scale Super-Resolution/images.zip b/2023/Cascaded Local Implicit Transformer for Arbitrary-Scale Super-Resolution/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..cadb7e23e2d673feb926e951299612a40975f387 --- /dev/null +++ b/2023/Cascaded Local Implicit Transformer for Arbitrary-Scale Super-Resolution/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a69128ffa99beb3d10632f6f96414e7e37959251f8467e96106795da26ee4584 +size 1003531 diff --git a/2023/Cascaded Local Implicit Transformer for Arbitrary-Scale Super-Resolution/layout.json b/2023/Cascaded Local Implicit Transformer for Arbitrary-Scale Super-Resolution/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..77c8a6e1d90bc1aa69cf686a1e531ba51b954c58 --- /dev/null +++ b/2023/Cascaded Local Implicit Transformer for Arbitrary-Scale Super-Resolution/layout.json @@ -0,0 +1,9046 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 63, + 50, + 531, + 68 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 50, + 531, + 68 + ], + "spans": [ + { + "bbox": [ + 63, + 50, + 531, + 68 + ], + "type": "text", + "content": " Cascaded Local Implicit Transformer for Arbitrary-Scale Super-Resolution" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 75, + 537, + 107 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 75, + 537, + 107 + ], + "spans": [ + { + "bbox": [ + 53, + 75, + 537, + 107 + ], + "type": "text", + "content": "Hao-Wei Chen\\*,1,2 Yu-Syuan Xu\\*,2 Min-Fong Hong\\* Yi-Min Tsai\\* Hsien-Kai Kuo\\* Chun-Yi Lee\\* Elsa Lab, National Tsing Hua University \\*MediaTek Inc." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 194, + 110, + 401, + 121 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 194, + 110, + 401, + 121 + ], + "spans": [ + { + "bbox": [ + 194, + 110, + 401, + 121 + ], + "type": "text", + "content": "{jaroslaw1007, cylee}@gapp.nthu.edu.tw" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 112, + 123, + 480, + 135 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 112, + 123, + 480, + 135 + ], + "spans": [ + { + "bbox": [ + 112, + 123, + 480, + 135 + ], + "type": "text", + "content": "{Yu-Syuan.Xu, romulus.hong, Yi-Min.Tsai, Hsienkai.Kuo}@mediatek.com" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 143, + 175, + 192, + 187 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 175, + 192, + 187 + ], + "spans": [ + { + "bbox": [ + 143, + 175, + 192, + 187 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 45, + 193, + 290, + 410 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 193, + 290, + 410 + ], + "spans": [ + { + "bbox": [ + 45, + 193, + 290, + 410 + ], + "type": "text", + "content": "Implicit neural representation has recently shown a promising ability in representing images with arbitrary resolutions. In this paper, we present a Local Implicit Transformer (LIT), which integrates the attention mechanism and frequency encoding technique into a local implicit image function. We design a cross-scale local attention block to effectively aggregate local features and a local frequency encoding block to combine positional encoding with Fourier domain information for constructing high-resolution images. To further improve representative power, we propose a Cascaded LIT (CLIT) that exploits multi-scale features, along with a cumulative training strategy that gradually increases the upsampling scales during training. We have conducted extensive experiments to validate the effectiveness of these components and analyze various training strategies. The qualitative and quantitative results demonstrate that LIT and CLIT achieve favorable results and outperform the prior works in arbitrary super-resolution tasks." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 418, + 128, + 430 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 418, + 128, + 430 + ], + "spans": [ + { + "bbox": [ + 47, + 418, + 128, + 430 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 433, + 287, + 660 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 433, + 287, + 660 + ], + "spans": [ + { + "bbox": [ + 46, + 433, + 287, + 660 + ], + "type": "text", + "content": "Single Image Super-Resolution (SISR) is the process of reconstructing high-resolution (HR) images from their corresponding low-resolution (LR) counterparts. SISR has long been recognized as a challenging task in the low-level vision domain due to its ill-posed nature, and has attracted a number of researchers dedicated to this field of study over the past decade [1-21]. A line of SISR research referred to as 'fixed-scale SR' [1-15] focuses on extracting feature embeddings from LR images and leveraging these embeddings to upsample images with a predefined factor through learnable deconvolutions [3] or sub-pixel convolutions [4]. Despite their success, many of the proposed approaches necessitate a distinct deep neural network model for each upsampling scale, which is usually restricted to a limited selection of integers (e.g., " + }, + { + "bbox": [ + 46, + 433, + 287, + 660 + ], + "type": "inline_equation", + "content": "2 \\times" + }, + { + "bbox": [ + 46, + 433, + 287, + 660 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 433, + 287, + 660 + ], + "type": "inline_equation", + "content": "3 \\times" + }, + { + "bbox": [ + 46, + 433, + 287, + 660 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 433, + 287, + 660 + ], + "type": "inline_equation", + "content": "4 \\times" + }, + { + "bbox": [ + 46, + 433, + 287, + 660 + ], + "type": "text", + "content": "). Such a limitation constrains the potential applications and deployment options of SISR models. To overcome this limitation, approaches for upsampling LR images in a continuous manner via a single model emerge and attracted considerable attention recently." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 47, + 660, + 287, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 660, + 287, + 696 + ], + "spans": [ + { + "bbox": [ + 47, + 660, + 287, + 696 + ], + "type": "text", + "content": "Over the past few years, arbitrary-scale SR has emerged and attracted considerable attention from researchers [16-21]. Apart from the pioneering work Meta-SR [16], recent" + } + ] + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 320, + 161, + 423, + 208 + ], + "blocks": [ + { + "bbox": [ + 320, + 161, + 423, + 208 + ], + "lines": [ + { + "bbox": [ + 320, + 161, + 423, + 208 + ], + "spans": [ + { + "bbox": [ + 320, + 161, + 423, + 208 + ], + "type": "image", + "image_path": "7dd09a46aa8a3ae0b389404446b4cb11bc731515f8fca5e1d6da6786a3498803.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 340, + 208, + 348, + 214 + ], + "lines": [ + { + "bbox": [ + 340, + 208, + 348, + 214 + ], + "spans": [ + { + "bbox": [ + 340, + 208, + 348, + 214 + ], + "type": "text", + "content": "LR" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 392, + 208, + 403, + 214 + ], + "lines": [ + { + "bbox": [ + 392, + 208, + 403, + 214 + ], + "spans": [ + { + "bbox": [ + 392, + 208, + 403, + 214 + ], + "type": "text", + "content": "HR" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 437, + 161, + 485, + 208 + ], + "blocks": [ + { + "bbox": [ + 437, + 161, + 485, + 208 + ], + "lines": [ + { + "bbox": [ + 437, + 161, + 485, + 208 + ], + "spans": [ + { + "bbox": [ + 437, + 161, + 485, + 208 + ], + "type": "image", + "image_path": "d4d3d990051fde0f40d410629c121267ff31938a40770d96a1caec70438d2d47.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 456, + 208, + 465, + 214 + ], + "lines": [ + { + "bbox": [ + 456, + 208, + 465, + 214 + ], + "spans": [ + { + "bbox": [ + 456, + 208, + 465, + 214 + ], + "type": "text", + "content": "LR" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 490, + 162, + 538, + 209 + ], + "blocks": [ + { + "bbox": [ + 490, + 162, + 538, + 209 + ], + "lines": [ + { + "bbox": [ + 490, + 162, + 538, + 209 + ], + "spans": [ + { + "bbox": [ + 490, + 162, + 538, + 209 + ], + "type": "image", + "image_path": "bcf88dce662e00f036c471a554bf325f4d362e8c2be90085943be045957dd27e.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 510, + 209, + 519, + 214 + ], + "lines": [ + { + "bbox": [ + 510, + 209, + 519, + 214 + ], + "spans": [ + { + "bbox": [ + 510, + 209, + 519, + 214 + ], + "type": "text", + "content": "HR" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 321, + 217, + 423, + 266 + ], + "blocks": [ + { + "bbox": [ + 306, + 238, + 317, + 247 + ], + "lines": [ + { + "bbox": [ + 306, + 238, + 317, + 247 + ], + "spans": [ + { + "bbox": [ + 306, + 238, + 317, + 247 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 321, + 217, + 423, + 266 + ], + "lines": [ + { + "bbox": [ + 321, + 217, + 423, + 266 + ], + "spans": [ + { + "bbox": [ + 321, + 217, + 423, + 266 + ], + "type": "image", + "image_path": "35530f43854e42ca6df6273a074902f118a82adb1809fbb418332d116d1950f5.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 338, + 266, + 349, + 271 + ], + "lines": [ + { + "bbox": [ + 338, + 266, + 349, + 271 + ], + "spans": [ + { + "bbox": [ + 338, + 266, + 349, + 271 + ], + "type": "text", + "content": "LIIF" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 377, + 266, + 388, + 271 + ], + "lines": [ + { + "bbox": [ + 377, + 266, + 388, + 271 + ], + "spans": [ + { + "bbox": [ + 377, + 266, + 388, + 271 + ], + "type": "text", + "content": "Loc" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_caption" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 437, + 217, + 538, + 266 + ], + "blocks": [ + { + "bbox": [ + 437, + 217, + 538, + 266 + ], + "lines": [ + { + "bbox": [ + 437, + 217, + 538, + 266 + ], + "spans": [ + { + "bbox": [ + 437, + 217, + 538, + 266 + ], + "type": "image", + "image_path": "254c026769d360b5c4d72475f0f5d2476cfbb5e39efffd59a99453f68d2be3e2.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 456, + 266, + 465, + 271 + ], + "lines": [ + { + "bbox": [ + 456, + 266, + 465, + 271 + ], + "spans": [ + { + "bbox": [ + 456, + 266, + 465, + 271 + ], + "type": "text", + "content": "LIIF" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 321, + 270, + 427, + 321 + ], + "blocks": [ + { + "bbox": [ + 321, + 270, + 427, + 321 + ], + "lines": [ + { + "bbox": [ + 321, + 270, + 427, + 321 + ], + "spans": [ + { + "bbox": [ + 321, + 270, + 427, + 321 + ], + "type": "image", + "image_path": "b71c8b963ebff36469aa32a422c08a5118dd1fb4dc208e9d66a6ecbf64f4925b.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 437, + 270, + 485, + 321 + ], + "blocks": [ + { + "bbox": [ + 437, + 270, + 485, + 321 + ], + "lines": [ + { + "bbox": [ + 437, + 270, + 485, + 321 + ], + "spans": [ + { + "bbox": [ + 437, + 270, + 485, + 321 + ], + "type": "image", + "image_path": "d8f8c1a76b4eaec2c548a4266eed42cb65b7eb618ee9845cad898259dbbac04e.jpg" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 455, + 318, + 468, + 324 + ], + "lines": [ + { + "bbox": [ + 455, + 318, + 468, + 324 + ], + "spans": [ + { + "bbox": [ + 455, + 318, + 468, + 324 + ], + "type": "text", + "content": "Ours" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_caption" + } + ], + "index": 27 + }, + { + "type": "image", + "bbox": [ + 492, + 270, + 544, + 323 + ], + "blocks": [ + { + "bbox": [ + 492, + 270, + 544, + 323 + ], + "lines": [ + { + "bbox": [ + 492, + 270, + 544, + 323 + ], + "spans": [ + { + "bbox": [ + 492, + 270, + 544, + 323 + ], + "type": "image", + "image_path": "113b3566563e8c0ffdcc6c13bc5b1e00884b78b508154fe15859ca0082a9038f.jpg" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 326, + 545, + 356 + ], + "lines": [ + { + "bbox": [ + 305, + 326, + 545, + 356 + ], + "spans": [ + { + "bbox": [ + 305, + 326, + 545, + 356 + ], + "type": "text", + "content": "Figure 1. An illustration and comparison of different approaches that take into account nearby pixels for continuous upsampling: (a) the local ensemble method used in [17], and (b) our proposed local attention mechanism." + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_caption" + } + ], + "index": 29 + }, + { + "bbox": [ + 304, + 363, + 546, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 363, + 546, + 639 + ], + "spans": [ + { + "bbox": [ + 304, + 363, + 546, + 639 + ], + "type": "text", + "content": "endevavors [17-21] have achieved arbitrary-scale SR by replacing the upsampling layers commonly adopted by previous approaches with local implicit image functions, and have demonstrated favorable performance. These local implicit functions employ multi-layer perceptrons (MLPs) to map 2D coordinates and corresponding latent representations to RGB values. Fig. 1 illustrates how different approaches sample latent representations based on the queried coordinates (depicted as the red dots). Fig. 1 (a) illustrates the local ensemble technique adopted by contemporary mainstream methods [17-21]. It calculates the RGB value of the queried coordinate by taking the weighted average of those of the surrounding four pixels based on their relative distances to the queried coordinate. This approach, however, does not consider contextual information and relies solely on distance. For instance, in Fig. 1, the queried coordinates are intentionally designed to lie on edges. However, merely calculating the weighted average of pixels fails to reflect the contextual information about the image content, thereby preventing the accurate capture of the necessary features for performing SR. As a result, although pixel distance plays a vital role in SR tasks, it is essential to concentrate more on the contextual information in an image." + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 304, + 641, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 641, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 641, + 547, + 714 + ], + "type": "text", + "content": "In light of the above observations, we propose a Local Implicit Transformer (LIT), which expands the numbers of referenced latent vectors and accounts for the feature correlation in the context by exploiting the attention mechanism [22]. LIT comprises a Cross-Scale Local Attention Block (CSLAB), a Local Frequency Encoding Block" + } + ] + } + ], + "index": 32 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 494, + 15 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 494, + 15 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 494, + 15 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 198, + 13, + 443, + 24 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 198, + 13, + 443, + 24 + ], + "spans": [ + { + "bbox": [ + 198, + 13, + 443, + 24 + ], + "type": "text", + "content": "Except for this watermark, it is identical to the accepted version;" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 179, + 25, + 461, + 36 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 179, + 25, + 461, + 36 + ], + "spans": [ + { + "bbox": [ + 179, + 25, + 461, + 36 + ], + "type": "text", + "content": "the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 703, + 128, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 703, + 128, + 712 + ], + "spans": [ + { + "bbox": [ + 62, + 703, + 128, + 712 + ], + "type": "text", + "content": "*Equal contribution" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18257" + } + ] + } + ], + "index": 34 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 289, + 324 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 289, + 324 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 289, + 324 + ], + "type": "text", + "content": "(LFEB), and a decoder. CSLAB generates attention maps based on the bilinearly interpolated latent vectors at the queried coordinates and the key latent vectors sampled from a grid of coordinates with relative positional bias [23, 24]. The first and second columns of Fig. 1 (b) visualize the attention maps generated by LIT, where the attention areas align closely with the edges. By applying attention maps to feature embeddings, the RGB values of the queried coordinates can be contextually predicted. Moreover, inspired by [21, 25, 26], we introduce LFEB, which projects relative coordinates into latent space to address the spectral bias problem [27] of an implicit neural function. Specifically, relative coordinates are encoded into relative positional encoding, which are multiplied with the frequency encoding extracted from the feature embedding in the Fourier domain to generate the frequency embeddings. This design enables the frequency embedding to integrate the relative positional encoding with texture information, thereby augmenting the expressivity of relative coordinates. Finally, a decoder is adopted to produce RGB values by taking advantage of the attention feature embedding and the frequency embedding." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 329, + 289, + 534 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 329, + 289, + 534 + ], + "spans": [ + { + "bbox": [ + 47, + 329, + 289, + 534 + ], + "type": "text", + "content": "In order to address the issue of diverse scaling factors and achieve arbitrary-scale super-resolution, it is crucial to consider the role of upsampling factors in constructing high-resolution images within the local implicit image function. However, simultaneously training a local implicit image function with a wide range of upsampling factors (e.g., " + }, + { + "bbox": [ + 47, + 329, + 289, + 534 + ], + "type": "inline_equation", + "content": "1 \\times \\sim 30 \\times" + }, + { + "bbox": [ + 47, + 329, + 289, + 534 + ], + "type": "text", + "content": ") poses significant challenges. As a result, we propose a cumulative training strategy to incrementally enhance the fuction's its representative power. The strategy initially trains the local implicit image function with small upsampling factors and then finetunes it with alternatively sampled small and large ones. Furthermore, we present Cascaded LIT (CLIT) to harness the advantages of multiscale feature embeddings, complementing missing details and information during one-step upsampling. The combination of the cumulative training strategy and CLIT enables efficient and effective handling of arbitrary-scale SR tasks." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 539, + 289, + 660 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 539, + 289, + 660 + ], + "spans": [ + { + "bbox": [ + 47, + 539, + 289, + 660 + ], + "type": "text", + "content": "The main contributions of our work are summarized as follows: (1) We introduce the LIT architecture, which incorporates the local attention mechanism into arbitrary-scale SR (2) We further develop a cumulative training strategy and the cascaded framework CLIT to effectively handle large-scale upsampling. (3) We carry out comprehensive analyses of the performance impacts for LIT and CLIT. The extensive experimental findings demonstrate that the proposed LIT and CLIT are able to yield remarkable or comparable results across a wide range of benchmark datasets." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 665, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 665, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 47, + 665, + 289, + 715 + ], + "type": "text", + "content": "The paper is organized as follows. Section 2 reviews the related work. Section 3 walks through the proposed LIT and CLIT frameworks and the implementation details. Section 4 presents the experimental results. Section 5 concludes." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 306, + 71, + 392, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 71, + 392, + 84 + ], + "spans": [ + { + "bbox": [ + 306, + 71, + 392, + 84 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 91, + 547, + 297 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 91, + 547, + 297 + ], + "spans": [ + { + "bbox": [ + 304, + 91, + 547, + 297 + ], + "type": "text", + "content": "Implicit neural representation. Implicit neural representation is a technique for representing continuous-domain signals via coordinate-based multi-layer perceptrons (MLPs). Its concept has been adopted in various 3D tasks, e.g., 3D object shape modeling [28-32], 3D scene reconstruction [33-36], and 3D structure rendering [25, 37-39]. For example, NeRF [25] employs implicit neural representation to perform novel view synthesis, which maps coordinates to RGB colors for a specific scene. In the past few years, 2D applications of implicit neural representation have been attempted as well, such as image representation [40, 41] and super-resolution [17-21]. Our work is related to a technique called 'local implicit neural representation' [17, 21], which encodes LR images to feature embeddings such that similar information could be shared within local regions. Such local implicit neural representations are exploited to upscale LR images to HR ones." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 307, + 547, + 452 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 307, + 547, + 452 + ], + "spans": [ + { + "bbox": [ + 304, + 307, + 547, + 452 + ], + "type": "text", + "content": "Single image super-resolution. In the past several years, various deep neural network (DNN) based architectures [1-15] have been proposed for SISR. Among these works, SR-CNN [1] pioneered the use of convolutional neural networks (CNNs) to achieve SISR in an end-to-end manner. It is later followed by several subsequent works that incorporated more complicated model architectures, such as residual blocks [6, 7], dense connections [8, 9], attention based mechanisms [10-12], or cascaded frameworks [5, 42, 43], to extract more effective feature representations for SISR. Recently, transformer-based methods [13-15] were introduced to SISR and achieved promising performance." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 462, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 462, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 462, + 547, + 715 + ], + "type": "text", + "content": "Arbitrary-scale super-resolution. As discussed in Section 1, most of the contemporary SISR works limit their upsampling scales to specific integer values, and are required to train a distinct model for each upsampling scale. To overcome such a limitation, several approaches [16-21] were proposed to train a unified model for arbitrary upsampling scales. Meta-SR [16] proposed a meta-upscale module for predicting the weights of their convolutional filters from coordinates and scales. The predicted weights are then utilized to perform convolutions to generate HR images. In contrast to Meta-SR, LIIF [17] employs an MLP as a local implicit function, which takes a queried coordinate in an HR image, its nearby feature representations extracted from the corresponding LR image, as well as a cell size to predict an RGB value for that coordinate. UltraSR [18] and IPE [19] extended LIIF by replacing coordinates with the embedded ones to deal with the spectral bias issue [25, 27, 41, 44, 45] inherent in MLPs. LTE [21] further introduced a local texture estimator that transforms coordinates into Fourier domain information to enrich the representational capability of its local implicit function. Different from the above ap" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "18258" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 51, + 71, + 545, + 258 + ], + "blocks": [ + { + "bbox": [ + 51, + 71, + 545, + 258 + ], + "lines": [ + { + "bbox": [ + 51, + 71, + 545, + 258 + ], + "spans": [ + { + "bbox": [ + 51, + 71, + 545, + 258 + ], + "type": "image", + "image_path": "f4d95a8a3964cbfddd8daf78d2de9aac78c5aa7fa8793fc059c5fe65e9b8d4c6.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 88, + 258, + 503, + 269 + ], + "lines": [ + { + "bbox": [ + 88, + 258, + 503, + 269 + ], + "spans": [ + { + "bbox": [ + 88, + 258, + 503, + 269 + ], + "type": "text", + "content": "Figure 2. The proposed LIT framework. The local sampling operation samples input embeddings based on a grid of coordinates." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 283, + 287, + 344 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 283, + 287, + 344 + ], + "spans": [ + { + "bbox": [ + 46, + 283, + 287, + 344 + ], + "type": "text", + "content": "proaches, our proposed methodology exploits a novel local attention mechanism and a cascaded framework to deal with the arbitrary-scale SR. In order to fairly compare with the above approaches, we similarly adopt EDSR [7], RDN [9] and SwinIR [14] as the encoders for our LIT and CLIT." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 354, + 129, + 367 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 354, + 129, + 367 + ], + "spans": [ + { + "bbox": [ + 47, + 354, + 129, + 367 + ], + "type": "text", + "content": "3. Methodology" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 373, + 287, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 373, + 287, + 422 + ], + "spans": [ + { + "bbox": [ + 46, + 373, + 287, + 422 + ], + "type": "text", + "content": "In this section, we first provide an overview of the proposed LIT framework, followed by the implementation details of it and its main modules. We then discuss our cumulative training strategy, as well as the framework of CLIT." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 429, + 224, + 441 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 429, + 224, + 441 + ], + "spans": [ + { + "bbox": [ + 47, + 429, + 224, + 441 + ], + "type": "text", + "content": "3.1. Overview of the LIT Framework" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 447, + 289, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 447, + 289, + 653 + ], + "spans": [ + { + "bbox": [ + 46, + 447, + 289, + 653 + ], + "type": "text", + "content": "LIT is a framework that employs a novel cross-scaled local attention mechanism and a local frequency encoding technique to perform arbitrary-scale SR tasks. Fig. 2 (a) provides an overview of the proposed framework, which aims at producing an HR image " + }, + { + "bbox": [ + 46, + 447, + 289, + 653 + ], + "type": "inline_equation", + "content": "I^{HR} \\in \\mathbb{R}^{r_hH \\times r_wW \\times 3}" + }, + { + "bbox": [ + 46, + 447, + 289, + 653 + ], + "type": "text", + "content": " at 2D HR coordinates " + }, + { + "bbox": [ + 46, + 447, + 289, + 653 + ], + "type": "inline_equation", + "content": "\\mathbf{x}^{HR} \\in \\mathcal{X}" + }, + { + "bbox": [ + 46, + 447, + 289, + 653 + ], + "type": "text", + "content": " from a given LR image " + }, + { + "bbox": [ + 46, + 447, + 289, + 653 + ], + "type": "inline_equation", + "content": "I^{LR} \\in \\mathbb{R}^{H \\times W \\times 3}" + }, + { + "bbox": [ + 46, + 447, + 289, + 653 + ], + "type": "text", + "content": " at 2D LR coordinates " + }, + { + "bbox": [ + 46, + 447, + 289, + 653 + ], + "type": "inline_equation", + "content": "\\mathbf{x}^{LR} \\in \\mathcal{X}" + }, + { + "bbox": [ + 46, + 447, + 289, + 653 + ], + "type": "text", + "content": " based on an arbitrary upsampling scale " + }, + { + "bbox": [ + 46, + 447, + 289, + 653 + ], + "type": "inline_equation", + "content": "\\mathbf{r} = \\{r_h, r_w\\}" + }, + { + "bbox": [ + 46, + 447, + 289, + 653 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 447, + 289, + 653 + ], + "type": "inline_equation", + "content": "\\mathcal{X}" + }, + { + "bbox": [ + 46, + 447, + 289, + 653 + ], + "type": "text", + "content": " is the 2D coordinate space that is used to represent an image in the continuous domain. An encoder " + }, + { + "bbox": [ + 46, + 447, + 289, + 653 + ], + "type": "inline_equation", + "content": "E_{\\theta}" + }, + { + "bbox": [ + 46, + 447, + 289, + 653 + ], + "type": "text", + "content": " first extracts a feature embedding " + }, + { + "bbox": [ + 46, + 447, + 289, + 653 + ], + "type": "inline_equation", + "content": "\\mathcal{Z} \\in \\mathbb{R}^{H \\times W \\times C}" + }, + { + "bbox": [ + 46, + 447, + 289, + 653 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 46, + 447, + 289, + 653 + ], + "type": "inline_equation", + "content": "I^{LR}" + }, + { + "bbox": [ + 46, + 447, + 289, + 653 + ], + "type": "text", + "content": ". The extracted " + }, + { + "bbox": [ + 46, + 447, + 289, + 653 + ], + "type": "inline_equation", + "content": "\\mathcal{Z}" + }, + { + "bbox": [ + 46, + 447, + 289, + 653 + ], + "type": "text", + "content": " is then forwarded into LIT along with the 2D coordinates of " + }, + { + "bbox": [ + 46, + 447, + 289, + 653 + ], + "type": "inline_equation", + "content": "I^{HR}" + }, + { + "bbox": [ + 46, + 447, + 289, + 653 + ], + "type": "text", + "content": " and a cell " + }, + { + "bbox": [ + 46, + 447, + 289, + 653 + ], + "type": "inline_equation", + "content": "= (2 / s_h, 2 / s_w)" + }, + { + "bbox": [ + 46, + 447, + 289, + 653 + ], + "type": "text", + "content": " to generate the RGB values of a residual image " + }, + { + "bbox": [ + 46, + 447, + 289, + 653 + ], + "type": "inline_equation", + "content": "I_r^{HR} \\in \\mathbb{R}^{r_hH \\times r_wW \\times 3}" + }, + { + "bbox": [ + 46, + 447, + 289, + 653 + ], + "type": "text", + "content": " in a pixel-wise fashion. Lastly, the residual image " + }, + { + "bbox": [ + 46, + 447, + 289, + 653 + ], + "type": "inline_equation", + "content": "I_r^{HR}" + }, + { + "bbox": [ + 46, + 447, + 289, + 653 + ], + "type": "text", + "content": " is combined with a bilinearly upsampled image " + }, + { + "bbox": [ + 46, + 447, + 289, + 653 + ], + "type": "inline_equation", + "content": "I_{\\uparrow}^{HR} \\in \\mathbb{R}^{r_hH \\times r_wW \\times 3}" + }, + { + "bbox": [ + 46, + 447, + 289, + 653 + ], + "type": "text", + "content": " via element-wise addition to derive the final HR image " + }, + { + "bbox": [ + 46, + 447, + 289, + 653 + ], + "type": "inline_equation", + "content": "I^{HR}" + }, + { + "bbox": [ + 46, + 447, + 289, + 653 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 659, + 198, + 672 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 659, + 198, + 672 + ], + "spans": [ + { + "bbox": [ + 47, + 659, + 198, + 672 + ], + "type": "text", + "content": "3.2. Local Implicit Transformer" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "text", + "content": "LIT is developed for mapping any 2D coordinate in the continuous image domain to an RGB color. As highlighted in Fig. 2 (b), it is composed of a cross-scale local attention" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 283, + 545, + 440 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 283, + 545, + 440 + ], + "spans": [ + { + "bbox": [ + 304, + 283, + 545, + 440 + ], + "type": "text", + "content": "block (CSLAB), a local frequency encoding block (LFB), and a decoder " + }, + { + "bbox": [ + 304, + 283, + 545, + 440 + ], + "type": "inline_equation", + "content": "D_{\\phi}" + }, + { + "bbox": [ + 304, + 283, + 545, + 440 + ], + "type": "text", + "content": " paraterizerized by " + }, + { + "bbox": [ + 304, + 283, + 545, + 440 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 304, + 283, + 545, + 440 + ], + "type": "text", + "content": ". The former two blocks are responsible for estimating a local latent embedding " + }, + { + "bbox": [ + 304, + 283, + 545, + 440 + ], + "type": "inline_equation", + "content": "\\tilde{z} \\in \\mathbb{R}^{G_h G_w \\times C}" + }, + { + "bbox": [ + 304, + 283, + 545, + 440 + ], + "type": "text", + "content": " and a local frequency embedding " + }, + { + "bbox": [ + 304, + 283, + 545, + 440 + ], + "type": "inline_equation", + "content": "\\tilde{f} \\in \\mathbb{R}^{G_h G_w \\times C}" + }, + { + "bbox": [ + 304, + 283, + 545, + 440 + ], + "type": "text", + "content": ", respectively, where " + }, + { + "bbox": [ + 304, + 283, + 545, + 440 + ], + "type": "inline_equation", + "content": "G_h" + }, + { + "bbox": [ + 304, + 283, + 545, + 440 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 283, + 545, + 440 + ], + "type": "inline_equation", + "content": "G_w" + }, + { + "bbox": [ + 304, + 283, + 545, + 440 + ], + "type": "text", + "content": " denote the height and width of local grids employed for performing local coordinate sampling, as depicted in Fig. 3. On the other hand, " + }, + { + "bbox": [ + 304, + 283, + 545, + 440 + ], + "type": "inline_equation", + "content": "D_{\\phi}" + }, + { + "bbox": [ + 304, + 283, + 545, + 440 + ], + "type": "text", + "content": " utilizes these embeddings along with the provided cell to generate " + }, + { + "bbox": [ + 304, + 283, + 545, + 440 + ], + "type": "inline_equation", + "content": "I_r^{HR}" + }, + { + "bbox": [ + 304, + 283, + 545, + 440 + ], + "type": "text", + "content": ". More specifically, LIT first projects " + }, + { + "bbox": [ + 304, + 283, + 545, + 440 + ], + "type": "inline_equation", + "content": "\\mathcal{Z}" + }, + { + "bbox": [ + 304, + 283, + 545, + 440 + ], + "type": "text", + "content": " using four separate convolutional layers to obtain four latent embeddings, corresponding to query " + }, + { + "bbox": [ + 304, + 283, + 545, + 440 + ], + "type": "inline_equation", + "content": "q" + }, + { + "bbox": [ + 304, + 283, + 545, + 440 + ], + "type": "text", + "content": ", key " + }, + { + "bbox": [ + 304, + 283, + 545, + 440 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 283, + 545, + 440 + ], + "type": "text", + "content": ", value " + }, + { + "bbox": [ + 304, + 283, + 545, + 440 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 304, + 283, + 545, + 440 + ], + "type": "text", + "content": ", and frequency " + }, + { + "bbox": [ + 304, + 283, + 545, + 440 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 304, + 283, + 545, + 440 + ], + "type": "text", + "content": ". Based on a queried HR coordinate " + }, + { + "bbox": [ + 304, + 283, + 545, + 440 + ], + "type": "inline_equation", + "content": "x_q \\in \\mathbf{x}^{HR}" + }, + { + "bbox": [ + 304, + 283, + 545, + 440 + ], + "type": "text", + "content": ", CSLAB and LFB estimate " + }, + { + "bbox": [ + 304, + 283, + 545, + 440 + ], + "type": "inline_equation", + "content": "\\tilde{z}" + }, + { + "bbox": [ + 304, + 283, + 545, + 440 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 283, + 545, + 440 + ], + "type": "inline_equation", + "content": "\\tilde{f}" + }, + { + "bbox": [ + 304, + 283, + 545, + 440 + ], + "type": "text", + "content": " as follows:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 371, + 449, + 545, + 461 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 371, + 449, + 545, + 461 + ], + "spans": [ + { + "bbox": [ + 371, + 449, + 545, + 461 + ], + "type": "interline_equation", + "content": "\\tilde {z} = C S L A B (\\delta x, q, k, v), \\tag {1}", + "image_path": "2075c477ab81548f6432bacd869505685d6bbaa2d4829ac9b2b9edb46a3f8608.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 383, + 468, + 545, + 483 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 383, + 468, + 545, + 483 + ], + "spans": [ + { + "bbox": [ + 383, + 468, + 545, + 483 + ], + "type": "interline_equation", + "content": "\\tilde {f} = L F E B (\\delta x, f), \\tag {2}", + "image_path": "d82f3a77612cf9449547bc25d96e4970c578bbeb0efd2e78a2b4af777679f1d1.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 329, + 487, + 545, + 510 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 329, + 487, + 545, + 510 + ], + "spans": [ + { + "bbox": [ + 329, + 487, + 545, + 510 + ], + "type": "interline_equation", + "content": "\\delta \\mathbf {x} = \\left\\{x _ {q} - x ^ {(i, j)} \\right\\} _ {i \\in \\{1, 2, \\dots , G _ {h} \\}, j \\in \\{1, 2, \\dots , G _ {w} \\}}, \\tag {3}", + "image_path": "9d9e69c0793f3a022296d2e88a265ce736dac7055e2bc430be7320fcfff809f6.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 514, + 545, + 658 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 514, + 545, + 658 + ], + "spans": [ + { + "bbox": [ + 304, + 514, + 545, + 658 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 514, + 545, + 658 + ], + "type": "inline_equation", + "content": "x^{(i,j)}\\in \\mathbf{x}^{LR}" + }, + { + "bbox": [ + 304, + 514, + 545, + 658 + ], + "type": "text", + "content": " denotes an LR coordinate in the local grid indexed by " + }, + { + "bbox": [ + 304, + 514, + 545, + 658 + ], + "type": "inline_equation", + "content": "(i,j)" + }, + { + "bbox": [ + 304, + 514, + 545, + 658 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 514, + 545, + 658 + ], + "type": "inline_equation", + "content": "\\delta \\mathbf{x}" + }, + { + "bbox": [ + 304, + 514, + 545, + 658 + ], + "type": "text", + "content": " represents the set of local relative coordinates defined by Eq. (3). The local grid is sampled in a manner that positions its center " + }, + { + "bbox": [ + 304, + 514, + 545, + 658 + ], + "type": "inline_equation", + "content": "x^{(\\lfloor G_h / 2\\rfloor +1,\\lfloor G_w / 2\\rfloor +1)}" + }, + { + "bbox": [ + 304, + 514, + 545, + 658 + ], + "type": "text", + "content": " at the LR coordinate closest to " + }, + { + "bbox": [ + 304, + 514, + 545, + 658 + ], + "type": "inline_equation", + "content": "x_{q}" + }, + { + "bbox": [ + 304, + 514, + 545, + 658 + ], + "type": "text", + "content": ". The query latent vector " + }, + { + "bbox": [ + 304, + 514, + 545, + 658 + ], + "type": "inline_equation", + "content": "q\\in \\mathbb{R}^{1\\times C}" + }, + { + "bbox": [ + 304, + 514, + 545, + 658 + ], + "type": "text", + "content": " at the HR coordinate " + }, + { + "bbox": [ + 304, + 514, + 545, + 658 + ], + "type": "inline_equation", + "content": "x_{q}" + }, + { + "bbox": [ + 304, + 514, + 545, + 658 + ], + "type": "text", + "content": " is computed by bilinear interpolation, while the remainder of the local latent embeddings " + }, + { + "bbox": [ + 304, + 514, + 545, + 658 + ], + "type": "inline_equation", + "content": "k\\in \\mathbb{R}^{G_hG_w\\times C}" + }, + { + "bbox": [ + 304, + 514, + 545, + 658 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 514, + 545, + 658 + ], + "type": "inline_equation", + "content": "v\\in \\mathbb{R}^{G_hG_w\\times C}" + }, + { + "bbox": [ + 304, + 514, + 545, + 658 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 514, + 545, + 658 + ], + "type": "inline_equation", + "content": "f\\in \\mathbb{R}^{G_hG_w\\times C}" + }, + { + "bbox": [ + 304, + 514, + 545, + 658 + ], + "type": "text", + "content": " are sampled at the local LR coordinates " + }, + { + "bbox": [ + 304, + 514, + 545, + 658 + ], + "type": "inline_equation", + "content": "\\mathbf{x} = \\{x^{i,j}\\}_{i\\in \\{1,2,\\dots,G_h\\},j\\in \\{1,2,\\dots,G_w\\}}" + }, + { + "bbox": [ + 304, + 514, + 545, + 658 + ], + "type": "text", + "content": ". With the local latent embedding " + }, + { + "bbox": [ + 304, + 514, + 545, + 658 + ], + "type": "inline_equation", + "content": "\\tilde{z}" + }, + { + "bbox": [ + 304, + 514, + 545, + 658 + ], + "type": "text", + "content": " and the local frequency embedding " + }, + { + "bbox": [ + 304, + 514, + 545, + 658 + ], + "type": "inline_equation", + "content": "\\tilde{f}" + }, + { + "bbox": [ + 304, + 514, + 545, + 658 + ], + "type": "text", + "content": ", the function of " + }, + { + "bbox": [ + 304, + 514, + 545, + 658 + ], + "type": "inline_equation", + "content": "D_{\\phi}" + }, + { + "bbox": [ + 304, + 514, + 545, + 658 + ], + "type": "text", + "content": " is formulated as follows:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 380, + 666, + 545, + 681 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 380, + 666, + 545, + 681 + ], + "spans": [ + { + "bbox": [ + 380, + 666, + 545, + 681 + ], + "type": "interline_equation", + "content": "I ^ {r} \\left(x _ {q}\\right) = D _ {\\phi} \\left(\\tilde {z}, \\tilde {f}, c\\right), \\tag {4}", + "image_path": "6641b0cfe4e637a5e941133b0c51b38c7d4f5a73f1e141664c07630639a02dfe.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 689, + 545, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 689, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 689, + 545, + 715 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 689, + 545, + 715 + ], + "type": "inline_equation", + "content": "I^{r}(x_{q})" + }, + { + "bbox": [ + 304, + 689, + 545, + 715 + ], + "type": "text", + "content": " is the predicted RGB value at the queried coordinate " + }, + { + "bbox": [ + 304, + 689, + 545, + 715 + ], + "type": "inline_equation", + "content": "x_{q}" + }, + { + "bbox": [ + 304, + 689, + 545, + 715 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 689, + 545, + 715 + ], + "type": "inline_equation", + "content": "c = \\{HR_{\\Delta h}, HR_{\\Delta w}\\}" + }, + { + "bbox": [ + 304, + 689, + 545, + 715 + ], + "type": "text", + "content": " denotes the cell that" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18259" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 48, + 68, + 289, + 243 + ], + "blocks": [ + { + "bbox": [ + 48, + 68, + 289, + 243 + ], + "lines": [ + { + "bbox": [ + 48, + 68, + 289, + 243 + ], + "spans": [ + { + "bbox": [ + 48, + 68, + 289, + 243 + ], + "type": "image", + "image_path": "4b721537abf77883775292eb0d479c54213e81a052d46033b70500a91fbec677.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 245, + 288, + 302 + ], + "lines": [ + { + "bbox": [ + 46, + 245, + 288, + 302 + ], + "spans": [ + { + "bbox": [ + 46, + 245, + 288, + 302 + ], + "type": "text", + "content": "Figure 3. An illustration of the proposed local coordinate sampling scheme. The red rectangular region, outlined by a dashed line, represents the local grid with dimensions " + }, + { + "bbox": [ + 46, + 245, + 288, + 302 + ], + "type": "inline_equation", + "content": "G_{h} \\times G_{w}" + }, + { + "bbox": [ + 46, + 245, + 288, + 302 + ], + "type": "text", + "content": ". The red dots indicate the sampled local LR coordinates, while the gray dots correspond to the HR coordinates. The grid dimensions " + }, + { + "bbox": [ + 46, + 245, + 288, + 302 + ], + "type": "inline_equation", + "content": "LR_{\\Delta h,w}" + }, + { + "bbox": [ + 46, + 245, + 288, + 302 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 245, + 288, + 302 + ], + "type": "inline_equation", + "content": "HR_{\\Delta h,w}" + }, + { + "bbox": [ + 46, + 245, + 288, + 302 + ], + "type": "text", + "content": " represent the unit sizes of pixels in LR and HR images, respectively." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 316, + 287, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 316, + 287, + 365 + ], + "spans": [ + { + "bbox": [ + 46, + 316, + 287, + 365 + ], + "type": "text", + "content": "represents the height and width of a pixel in an HR image, as illustrated in Fig. 3. " + }, + { + "bbox": [ + 46, + 316, + 287, + 365 + ], + "type": "inline_equation", + "content": "D_{\\phi}" + }, + { + "bbox": [ + 46, + 316, + 287, + 365 + ], + "type": "text", + "content": " is implemented as a five-layer MLP utilizing Gaussian Error Linear Unit (GELU) activation [46], and is employed consistently across all images." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 379, + 287, + 499 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 379, + 287, + 499 + ], + "spans": [ + { + "bbox": [ + 46, + 379, + 287, + 499 + ], + "type": "text", + "content": "Cross-scale local attention block. LIT exploits CSLAB to perform a local attention mechanism over a local grid to generate a local latent embedding " + }, + { + "bbox": [ + 46, + 379, + 287, + 499 + ], + "type": "inline_equation", + "content": "\\tilde{z}" + }, + { + "bbox": [ + 46, + 379, + 287, + 499 + ], + "type": "text", + "content": " for each HR coordinate, as illustrated in Fig. 2 (c). CSLAB first calculates the inner product of " + }, + { + "bbox": [ + 46, + 379, + 287, + 499 + ], + "type": "inline_equation", + "content": "q" + }, + { + "bbox": [ + 46, + 379, + 287, + 499 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 379, + 287, + 499 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 379, + 287, + 499 + ], + "type": "text", + "content": ", adds the relative positional bias " + }, + { + "bbox": [ + 46, + 379, + 287, + 499 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 46, + 379, + 287, + 499 + ], + "type": "text", + "content": " to the result, and obtains an attention matrix. This attention matrix is subsequently normalized by a Softmax operation to produce a local attention map. Finally, CSLAB performs element-wise multiplication of " + }, + { + "bbox": [ + 46, + 379, + 287, + 499 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 46, + 379, + 287, + 499 + ], + "type": "text", + "content": " and the local attention map to derive " + }, + { + "bbox": [ + 46, + 379, + 287, + 499 + ], + "type": "inline_equation", + "content": "\\tilde{z}" + }, + { + "bbox": [ + 46, + 379, + 287, + 499 + ], + "type": "text", + "content": ". The overall procedure is formulated as follows:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 91, + 506, + 287, + 590 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 506, + 287, + 590 + ], + "spans": [ + { + "bbox": [ + 91, + 506, + 287, + 590 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\tilde {z} = \\operatorname {s o f t m a x} \\left(\\frac {q k ^ {\\top}}{\\sqrt {C}} + B\\right) \\times v, (5) \\\\ B = F C (\\gamma (\\delta \\mathbf {x})), (6) \\\\ \\gamma (\\delta \\mathbf {x}) = \\left[ \\sin \\left(2 ^ {0} \\delta \\mathbf {x}\\right), \\cos \\left(2 ^ {0} \\delta \\mathbf {x}\\right), \\dots \\right. (7) \\\\ , \\sin \\left(2 ^ {L - 1} \\delta \\mathbf {x}\\right), \\cos \\left(2 ^ {L - 1} \\delta \\mathbf {x}\\right) ], \\\\ \\end{array}", + "image_path": "44029d5a1e9d097a19b81754c1711034274bb50885b623843061fb0dea378e79.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 595, + 287, + 655 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 595, + 287, + 655 + ], + "spans": [ + { + "bbox": [ + 46, + 595, + 287, + 655 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 595, + 287, + 655 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 46, + 595, + 287, + 655 + ], + "type": "text", + "content": " denotes the channel dimension of the local key latent embedding " + }, + { + "bbox": [ + 46, + 595, + 287, + 655 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 595, + 287, + 655 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 595, + 287, + 655 + ], + "type": "inline_equation", + "content": "FC" + }, + { + "bbox": [ + 46, + 595, + 287, + 655 + ], + "type": "text", + "content": " represents a fully-connected layer, " + }, + { + "bbox": [ + 46, + 595, + 287, + 655 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 46, + 595, + 287, + 655 + ], + "type": "text", + "content": " is the positional encoding function, and " + }, + { + "bbox": [ + 46, + 595, + 287, + 655 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 46, + 595, + 287, + 655 + ], + "type": "text", + "content": " is a hyperparameter. In this work, " + }, + { + "bbox": [ + 46, + 595, + 287, + 655 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 46, + 595, + 287, + 655 + ], + "type": "text", + "content": " is set to 10, and the multi-head attention mechanism adopted is formulated in Eq. (8) as follows:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 74, + 663, + 287, + 693 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 74, + 663, + 287, + 693 + ], + "spans": [ + { + "bbox": [ + 74, + 663, + 287, + 693 + ], + "type": "interline_equation", + "content": "\\tilde {z} = \\operatorname {c o n c a t} \\left(\\operatorname {s o f t m a x} \\left(\\frac {q _ {i} k _ {i} ^ {\\top}}{\\sqrt {C / H}} + B _ {i}\\right) \\times v _ {i}\\right), \\tag {8}", + "image_path": "e3c9e0d38bf461aaba041daf2fedcf6b8f6dcd306e3c710610de5ef1186a4f09.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 701, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 701, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 701, + 287, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 701, + 287, + 714 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 47, + 701, + 287, + 714 + ], + "type": "text", + "content": " is the number of attention heads and " + }, + { + "bbox": [ + 47, + 701, + 287, + 714 + ], + "type": "inline_equation", + "content": "i\\in [1,\\dots ,H]" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 309, + 72, + 545, + 194 + ], + "blocks": [ + { + "bbox": [ + 309, + 72, + 545, + 194 + ], + "lines": [ + { + "bbox": [ + 309, + 72, + 545, + 194 + ], + "spans": [ + { + "bbox": [ + 309, + 72, + 545, + 194 + ], + "type": "image", + "image_path": "a9ca32d0bd97b5f4a24bbb9fda0903384bfb89b5b21fc8834c22b9195181d570.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 332, + 198, + 518, + 208 + ], + "lines": [ + { + "bbox": [ + 332, + 198, + 518, + 208 + ], + "spans": [ + { + "bbox": [ + 332, + 198, + 518, + 208 + ], + "type": "text", + "content": "Figure 4. An overview of the proposed CLIT framework." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 218, + 545, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 218, + 545, + 350 + ], + "spans": [ + { + "bbox": [ + 305, + 218, + 545, + 350 + ], + "type": "text", + "content": "Local frequency encoding block. Fig. 2 (d) illustrates the local frequency encoding block, which incorporates dominant-frequency components to generate the local frequency embedding " + }, + { + "bbox": [ + 305, + 218, + 545, + 350 + ], + "type": "inline_equation", + "content": "\\tilde{f}" + }, + { + "bbox": [ + 305, + 218, + 545, + 350 + ], + "type": "text", + "content": ". In order to account for dominant-frequency information, LFBB first encodes local frequency latent embedding " + }, + { + "bbox": [ + 305, + 218, + 545, + 350 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 305, + 218, + 545, + 350 + ], + "type": "text", + "content": " using Fast Fourier Transform (FFT), which is then multiplied by the relative positional encoding to generate " + }, + { + "bbox": [ + 305, + 218, + 545, + 350 + ], + "type": "inline_equation", + "content": "\\tilde{f}" + }, + { + "bbox": [ + 305, + 218, + 545, + 350 + ], + "type": "text", + "content": ". The resultant local frequency embedding " + }, + { + "bbox": [ + 305, + 218, + 545, + 350 + ], + "type": "inline_equation", + "content": "\\tilde{f}" + }, + { + "bbox": [ + 305, + 218, + 545, + 350 + ], + "type": "text", + "content": " is then concatenated with the flattened local latent embedding " + }, + { + "bbox": [ + 305, + 218, + 545, + 350 + ], + "type": "inline_equation", + "content": "\\tilde{z}" + }, + { + "bbox": [ + 305, + 218, + 545, + 350 + ], + "type": "text", + "content": " to form a mixing embedding, which can be used by the decoder consisting of MLPs to predict RGB values [47]." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 359, + 469, + 372 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 359, + 469, + 372 + ], + "spans": [ + { + "bbox": [ + 306, + 359, + 469, + 372 + ], + "type": "text", + "content": "3.3. Cumulative Training Strategy" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 379, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 379, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 379, + 545, + 713 + ], + "type": "text", + "content": "In this section, we discuss our proposed cumulative training strategy, which is developed for enhancing the performance of arbitrary-scale SR. The cumulative training strategy focuses on the schedule of the cell sizes selected during the training phase, as \"cell decoding\" has been recognized as an essential input to a local implicit image function [17]. Recent studies [17, 19] have observed that the effect of cell decoding on the performance of arbitrary-scale SR is prominent for in-distribution upsampling, but degrades significantly for out-of-distribution large-scale upsampling. Such trends are demonstrated in Table 3 of [17] and Table 4 of [19]. The authors in [21] also constrained their cell sizes to in-distribution ranges to mitigate the negative impact during evaluation. To overcome the degradation issue for out-of-distribution cell sizes, incorporating large cell sizes during training appears to be a promising solution. However, simply training the local implicit image function with a diverse range of cell sizes at once leads to a performance drop. According to our experimental observations presented in Section 4.4, we find that training the local implicit image function by alternatively switching between large and small cells offers positive impacts on the performance. Based on the above observations, we propose a cumulative training strategy which first trains the local implicit image function with large cell sizes, and finetunes it with the alternative training strategy to improve the performance on different upsampling scales. More quantitative results of our training strategies can be found in Section 4.4." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18260" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 63, + 70, + 534, + 310 + ], + "blocks": [ + { + "bbox": [ + 63, + 70, + 534, + 310 + ], + "lines": [ + { + "bbox": [ + 63, + 70, + 534, + 310 + ], + "spans": [ + { + "bbox": [ + 63, + 70, + 534, + 310 + ], + "type": "table", + "html": "
Method×2×3×4×6×12×18×24×30
Bicubic [7]31.0128.2226.6624.8222.2721.0020.1919.59
EDSR-baseline [7]34.5530.9028.94-----
EDSR-baseline-Meta-SR [16, 17]34.6430.9328.9226.6123.5522.0321.0620.37
EDSR-baseline-LIIF [17]34.6730.9629.0026.7523.7122.1721.1820.48
EDSR-baseline-UltraSR [18]34.6931.0229.0526.8123.7522.2121.2120.51
EDSR-baseline-IPE [19]34.7231.0129.0426.7923.7522.2121.2220.51
EDSR-baseline-LTE [21]34.7231.0229.0426.8123.7822.2321.2420.53
EDSR-baseline-CLIT (Ours)34.8131.1229.1526.9223.8322.2921.2620.53
RDN-Meta-SR [16, 17]35.0031.2729.2526.8823.7322.1821.1720.47
RDN-LIIF [17]34.9931.2629.2726.9923.8922.3421.3120.59
RDN-UltraSR [18]35.0031.3029.3227.0323.7322.3621.3320.61
RDN-IPE [19]35.0431.3229.3227.0423.9322.3821.3420.63
RDN-LTE [21]35.0431.3229.3327.0423.9522.4021.3620.64
RDN-CLIT (Ours)35.1031.3929.3927.1224.0122.4521.3820.64
SwinIR-MetaSR [16, 21]35.1531.4029.3326.9423.8022.2621.2620.54
SwinIR-LIIF [17, 21]35.1731.4629.4627.1524.0222.4321.4020.67
SwinIR-LTE [21]35.2431.5029.5127.2024.0922.5021.4720.73
SwinIR-CLIT (Ours)35.2931.5529.5527.2624.1122.5121.4520.70
", + "image_path": "1c1a785986547ec8de67eccb11a849f76cff301b4641887444182acfa3e80442.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 312, + 545, + 332 + ], + "lines": [ + { + "bbox": [ + 46, + 312, + 545, + 332 + ], + "spans": [ + { + "bbox": [ + 46, + 312, + 545, + 332 + ], + "type": "text", + "content": "Table 1. The average PSNR (dB) on the DIV2K [48] validation set. The results are obtained from the original manuscripts [17-19, 21]. The best and second-best performing results are highlighted by the red and blue colors, respectively." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 47, + 335, + 247, + 348 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 335, + 247, + 348 + ], + "spans": [ + { + "bbox": [ + 47, + 335, + 247, + 348 + ], + "type": "text", + "content": "3.4. Cascaded Local Implicit Transformer" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 356, + 288, + 451 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 356, + 288, + 451 + ], + "spans": [ + { + "bbox": [ + 46, + 356, + 288, + 451 + ], + "type": "text", + "content": "Previous works [17, 21] typically address arbitrary-scale SR through a single step of upsampling. However, one-step upsampling struggles to reconstruct an HR image when the upsampling scale is large [5, 42, 43]. In light of this, we propose a cascaded upsampling strategy, called \"Cascaded LIT(CLIT)\" to predict residual images from multi-scale feature embeddings. In an " + }, + { + "bbox": [ + 46, + 356, + 288, + 451 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 356, + 288, + 451 + ], + "type": "text", + "content": "-branched CLIT, the multi-scale feature embeddings " + }, + { + "bbox": [ + 46, + 356, + 288, + 451 + ], + "type": "inline_equation", + "content": "\\mathcal{Z}^1, \\mathcal{Z}^2, \\dots, \\mathcal{Z}^N" + }, + { + "bbox": [ + 46, + 356, + 288, + 451 + ], + "type": "text", + "content": " are derived as follows:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 111, + 464, + 287, + 485 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 111, + 464, + 287, + 485 + ], + "spans": [ + { + "bbox": [ + 111, + 464, + 287, + 485 + ], + "type": "interline_equation", + "content": "\\mathcal {Z} ^ {N} = \\mathcal {Z} \\uparrow_ {s ^ {1} \\times s ^ {2} \\times \\dots \\times s ^ {N - 1}}, \\tag {9}", + "image_path": "63c9147ddfbc25a5aa0f67563432558f226bc2137c6f35bfee4f02d831ce8b32.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 112, + 481, + 212, + 493 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 112, + 481, + 212, + 493 + ], + "spans": [ + { + "bbox": [ + 112, + 481, + 212, + 493 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 112, + 481, + 212, + 493 + ], + "type": "inline_equation", + "content": "s^1 = 1" + }, + { + "bbox": [ + 112, + 481, + 212, + 493 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 112, + 481, + 212, + 493 + ], + "type": "inline_equation", + "content": "s\\in \\mathbf{s}" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 506, + 287, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 506, + 287, + 578 + ], + "spans": [ + { + "bbox": [ + 47, + 506, + 287, + 578 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 506, + 287, + 578 + ], + "type": "inline_equation", + "content": "\\uparrow" + }, + { + "bbox": [ + 47, + 506, + 287, + 578 + ], + "type": "text", + "content": " is a bilinear upsampling function and " + }, + { + "bbox": [ + 47, + 506, + 287, + 578 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 47, + 506, + 287, + 578 + ], + "type": "text", + "content": " is a set of scaling factors, which are configurable hypermeters. For a branch " + }, + { + "bbox": [ + 47, + 506, + 287, + 578 + ], + "type": "inline_equation", + "content": "i,i\\in [1,\\dots,N]" + }, + { + "bbox": [ + 47, + 506, + 287, + 578 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 47, + 506, + 287, + 578 + ], + "type": "inline_equation", + "content": "\\mathrm{LIT}^i" + }, + { + "bbox": [ + 47, + 506, + 287, + 578 + ], + "type": "text", + "content": " estimates the residual image " + }, + { + "bbox": [ + 47, + 506, + 287, + 578 + ], + "type": "inline_equation", + "content": "I_r^i" + }, + { + "bbox": [ + 47, + 506, + 287, + 578 + ], + "type": "text", + "content": " from the feature embedding " + }, + { + "bbox": [ + 47, + 506, + 287, + 578 + ], + "type": "inline_equation", + "content": "\\mathcal{Z}^i" + }, + { + "bbox": [ + 47, + 506, + 287, + 578 + ], + "type": "text", + "content": " with the coordinate and the corresponding cells. Lastly, the final HR image " + }, + { + "bbox": [ + 47, + 506, + 287, + 578 + ], + "type": "inline_equation", + "content": "I^{HR}\\in \\mathbb{R}^{r_hH\\times r_wW\\times 3}" + }, + { + "bbox": [ + 47, + 506, + 287, + 578 + ], + "type": "text", + "content": " can be estimated as the following equation:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 590, + 287, + 605 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 590, + 287, + 605 + ], + "spans": [ + { + "bbox": [ + 56, + 590, + 287, + 605 + ], + "type": "interline_equation", + "content": "I ^ {H R} = \\lambda^ {N - 1} I _ {r} ^ {1} + \\lambda^ {N - 2} I _ {r} ^ {2} + \\dots + \\lambda^ {0} I _ {r} ^ {N} + I _ {\\uparrow} ^ {H R}. \\tag {10}", + "image_path": "09095e9ba8910c8e67957eaf165246ea8c221d95db4ae8157a2cd586957fbd91.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 617, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 617, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 617, + 287, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 617, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 46, + 617, + 287, + 713 + ], + "type": "text", + "content": " is a discount factor, with a default value of 0.75. During the training phase, CLIT is trained using the proposed cumulative training strategy. Initially, " + }, + { + "bbox": [ + 46, + 617, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{LIT}^1" + }, + { + "bbox": [ + 46, + 617, + 287, + 713 + ], + "type": "text", + "content": " is trained with the strategy employed by [17]. Subsequently, " + }, + { + "bbox": [ + 46, + 617, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{LIT}^1" + }, + { + "bbox": [ + 46, + 617, + 287, + 713 + ], + "type": "text", + "content": " is finetuned and " + }, + { + "bbox": [ + 46, + 617, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{LIT}^2" + }, + { + "bbox": [ + 46, + 617, + 287, + 713 + ], + "type": "text", + "content": " is initialized by applying the alternative training strategy. By incrementally incorporating LITs into CLIT, the performance is progressively enhanced. The details of the training strategies are specified in Section 4." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 335, + 432, + 348 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 335, + 432, + 348 + ], + "spans": [ + { + "bbox": [ + 306, + 335, + 432, + 348 + ], + "type": "text", + "content": "4. Experimental Results" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 356, + 545, + 440 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 356, + 545, + 440 + ], + "spans": [ + { + "bbox": [ + 304, + 356, + 545, + 440 + ], + "type": "text", + "content": "In this section, we present the experimental results and discuss their implications. We begin with a brief introduction to our experimental setup in Section 4.1. Following that, we evaluate our CLIT with different datasets in Section 4.2. The, Section 4.3 showcases the learned attention maps. Finally, ablation studies for various configurations of the proposed CLIT and LIT are compared in Section 4.4." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 449, + 422, + 462 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 449, + 422, + 462 + ], + "spans": [ + { + "bbox": [ + 305, + 449, + 422, + 462 + ], + "type": "text", + "content": "4.1. Experimental Setup" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 468, + 545, + 563 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 468, + 545, + 563 + ], + "spans": [ + { + "bbox": [ + 304, + 468, + 545, + 563 + ], + "type": "text", + "content": "Dataset. We use the DIV2K dataset [48] for network training. It consists of 1,000 images in 2K resolutions and provides low-resolution counterparts with down-sampling scales, " + }, + { + "bbox": [ + 304, + 468, + 545, + 563 + ], + "type": "inline_equation", + "content": "\\times 2" + }, + { + "bbox": [ + 304, + 468, + 545, + 563 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 468, + 545, + 563 + ], + "type": "inline_equation", + "content": "\\times 3" + }, + { + "bbox": [ + 304, + 468, + 545, + 563 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 468, + 545, + 563 + ], + "type": "inline_equation", + "content": "\\times 4" + }, + { + "bbox": [ + 304, + 468, + 545, + 563 + ], + "type": "text", + "content": ", which are generated by the bicubic interpolation method. On the other hand, we also evaluate the performance on the validation set of DIV2K [48], Set5 [49], Set14 [50], B100 [51] and Urban100 [52] in terms of peak signal-to-noise (PSNR) values." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "content": "Training details During training, we feed batches of size " + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "inline_equation", + "content": "48 \\times 48" + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "content": " low-resolution images into the framework, following the prior works [7]. For each batch, a single upsampling scale is sampled from a uniform distribution " + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "inline_equation", + "content": "r \\sim \\mathcal{U}(1,4)" + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "content": ". With single upsampling scales " + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "content": ", a batch of HR images is cropped into patches of size " + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "inline_equation", + "content": "48r \\times 48r" + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "content": ", while the corresponding LR images are cropped into patches of " + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "inline_equation", + "content": "48 \\times 48" + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "content": ". The patches are augmented by randomly horizontal flipping, vertical flipping, and " + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "inline_equation", + "content": "90^{\\circ}" + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "content": " rotating. Then, we sample " + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "inline_equation", + "content": "48^2" + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "content": " pixels (coordinate-RGB pairs) on each HR patch as the ground-truths. We set the batch size to 32 and use" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18261" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 49, + 70, + 547, + 169 + ], + "blocks": [ + { + "bbox": [ + 49, + 70, + 547, + 169 + ], + "lines": [ + { + "bbox": [ + 49, + 70, + 547, + 169 + ], + "spans": [ + { + "bbox": [ + 49, + 70, + 547, + 169 + ], + "type": "table", + "html": "
MethodSet5Set14B100Urban100
×2×3×4×6×8×2×3×4×6×8×2×3×4×6×8×2×3×4×6×8
RDN [9]38.2434.7132.47--34.0130.5728.81--32.3429.2627.72--32.8928.8026.61--
RDN-Meta-SR [16, 17]38.2234.6332.3829.0426.9633.9830.5428.7826.5124.9732.3329.2627.7125.9024.8332.9228.8226.5523.9922.59
RDN-LLIF [17]38.1734.6832.5029.1527.1433.9730.5328.8026.6425.1532.3229.2627.7425.9824.9132.8728.8226.6824.2022.79
RDN-UltraSR [18]38.2134.6732.4929.3327.2433.9730.5928.8626.6925.2532.3529.2927.7726.0124.9632.9728.9226.7824.3022.87
RDN-IPE [19]38.1134.6832.5129.2527.2233.9430.4728.7526.5825.0932.3129.2827.7626.0024.9332.9728.8226.7624.2622.87
RDN-LTE [21]38.2334.7232.6129.3227.2634.0930.5828.8826.7125.1632.3629.3027.7726.0124.9533.0428.9726.8124.2822.88
RDN-CLIT (Ours)38.2634.7932.6929.5427.3434.0930.6928.9326.8325.3632.3929.3327.8026.0725.0033.1429.0526.9324.4423.04
SwinIR [14]38.3534.8932.72--34.1430.7728.94--32.4429.3727.83--33.4029.2927.07--
SwinIR-MetaSR [16, 21]38.2634.7732.4729.0927.0234.1430.6628.8526.5825.0932.3929.3127.7525.9424.8633.2929.1226.7624.1622.75
SwinIR-LLIF [17]38.2834.8732.7329.4627.3634.1430.7528.9826.8225.3432.3929.3427.8426.0725.0133.3629.3327.1524.5923.14
SwinIR-LTE [21]38.3334.8932.8129.5027.3534.2530.8029.0626.8625.4232.4429.3927.8626.0925.0333.5029.4127.2424.6223.17
SwinIR-CLIT (Ours)38.4134.9732.8629.6927.6234.2730.8529.0826.9425.5532.4629.4227.9126.1525.0933.5629.4327.2524.7723.33
", + "image_path": "c8f790d45fa284914bdd715039abf777e30d550a11dc804d42703976814f10aa.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 171, + 545, + 190 + ], + "lines": [ + { + "bbox": [ + 47, + 171, + 545, + 190 + ], + "spans": [ + { + "bbox": [ + 47, + 171, + 545, + 190 + ], + "type": "text", + "content": "Table 2. The average PSNR (dB) on Set5 [49], Set14 [50], B100 [51], and Urban100 [52]. The results are obtained from the original manuscripts [17-19, 21]. The best and second-best performing results are highlighted by the red and blue colors, respectively." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 80, + 194, + 513, + 541 + ], + "blocks": [ + { + "bbox": [ + 80, + 194, + 513, + 541 + ], + "lines": [ + { + "bbox": [ + 80, + 194, + 513, + 541 + ], + "spans": [ + { + "bbox": [ + 80, + 194, + 513, + 541 + ], + "type": "image", + "image_path": "6e0bd67caa71965e7c5090c88a45e21384e9a979558071c4cab1b76adb97b29a.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 125, + 542, + 466, + 553 + ], + "lines": [ + { + "bbox": [ + 125, + 542, + 466, + 553 + ], + "spans": [ + { + "bbox": [ + 125, + 542, + 466, + 553 + ], + "type": "text", + "content": "Figure 5. The qualitative results of LIIF [17], LTE [21], and our CLIT with using RDN [9] as the encoder." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 557, + 288, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 557, + 288, + 712 + ], + "spans": [ + { + "bbox": [ + 46, + 557, + 288, + 712 + ], + "type": "text", + "content": "the Adam optimizer [53] together with L1 loss for training. We train LIT for 1,000 epochs, and the learning rate is initialized at " + }, + { + "bbox": [ + 46, + 557, + 288, + 712 + ], + "type": "inline_equation", + "content": "1e^{-4}" + }, + { + "bbox": [ + 46, + 557, + 288, + 712 + ], + "type": "text", + "content": " and decayed by a factor 0.5 at [200, 400, 600, 800] epochs. For cumulative training of CLIT, as discussed in Section 3.4, we sample " + }, + { + "bbox": [ + 46, + 557, + 288, + 712 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 557, + 288, + 712 + ], + "type": "text", + "content": " scale factors " + }, + { + "bbox": [ + 46, + 557, + 288, + 712 + ], + "type": "inline_equation", + "content": "\\{s^1, s^2, \\dots, s^N\\}" + }, + { + "bbox": [ + 46, + 557, + 288, + 712 + ], + "type": "text", + "content": " from the distribution " + }, + { + "bbox": [ + 46, + 557, + 288, + 712 + ], + "type": "inline_equation", + "content": "\\mathcal{U}(1,4)" + }, + { + "bbox": [ + 46, + 557, + 288, + 712 + ], + "type": "text", + "content": " according the number of LITs " + }, + { + "bbox": [ + 46, + 557, + 288, + 712 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 557, + 288, + 712 + ], + "type": "text", + "content": " in the train step. The total upsampling scale " + }, + { + "bbox": [ + 46, + 557, + 288, + 712 + ], + "type": "inline_equation", + "content": "r = s^1 \\times s^2 \\dots \\times s^N" + }, + { + "bbox": [ + 46, + 557, + 288, + 712 + ], + "type": "text", + "content": " is the product of all scale factors. If " + }, + { + "bbox": [ + 46, + 557, + 288, + 712 + ], + "type": "inline_equation", + "content": "48r \\times 48r" + }, + { + "bbox": [ + 46, + 557, + 288, + 712 + ], + "type": "text", + "content": " is greater than the whole HR image, we clip the scale factor of stage 1. To train " + }, + { + "bbox": [ + 46, + 557, + 288, + 712 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 557, + 288, + 712 + ], + "type": "text", + "content": " LITs, we fine-tune the model for " + }, + { + "bbox": [ + 46, + 557, + 288, + 712 + ], + "type": "inline_equation", + "content": "500 \\times N" + }, + { + "bbox": [ + 46, + 557, + 288, + 712 + ], + "type": "text", + "content": " epochs, and the learning rate is initialized at " + }, + { + "bbox": [ + 46, + 557, + 288, + 712 + ], + "type": "inline_equation", + "content": "1e^{-4}" + }, + { + "bbox": [ + 46, + 557, + 288, + 712 + ], + "type": "text", + "content": " and decayed by a factor 0.5 at " + }, + { + "bbox": [ + 46, + 557, + 288, + 712 + ], + "type": "inline_equation", + "content": "[100 \\times N, 200 \\times N, 300 \\times N, 400 \\times N]" + }, + { + "bbox": [ + 46, + 557, + 288, + 712 + ], + "type": "text", + "content": " epochs." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 306, + 556, + 418, + 567 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 556, + 418, + 567 + ], + "spans": [ + { + "bbox": [ + 306, + 556, + 418, + 567 + ], + "type": "text", + "content": "4.2. Validation of CLIT" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 578, + 545, + 674 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 578, + 545, + 674 + ], + "spans": [ + { + "bbox": [ + 304, + 578, + 545, + 674 + ], + "type": "text", + "content": "Quantitative results. We first compare our proposed CLIT to existing local implicit neural representation methods for arbitrary-scale SR, including LIIF [17], UltraSR [18], IPE [19], and LTE [21]. Table 1 summarizes the quantitative results in terms of PSNR(dB) on DIV2k [48]. As shown in Table 1, CLIT achieves the best performance when EDSR-Baseline [7], RDN [9] are used as the encoders. These results demonstrate the advantage of CLIT." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 677, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 677, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 677, + 546, + 713 + ], + "type": "text", + "content": "Table 2 compares CLIT to prior works [17-19, 21] on widely-used datasets, including Set5 [49], Set14 [50], B100 [51], and Urban100 [52], with RDN and SwinIR." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18262" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 72, + 70, + 523, + 342 + ], + "blocks": [ + { + "bbox": [ + 72, + 70, + 523, + 342 + ], + "lines": [ + { + "bbox": [ + 72, + 70, + 523, + 342 + ], + "spans": [ + { + "bbox": [ + 72, + 70, + 523, + 342 + ], + "type": "image", + "image_path": "ebe25e5c577baf8ff4da43660cc0c4fc9ae38ee8c4d5fe45500e97fbf15af4e7.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 77, + 342, + 514, + 354 + ], + "lines": [ + { + "bbox": [ + 77, + 342, + 514, + 354 + ], + "spans": [ + { + "bbox": [ + 77, + 342, + 514, + 354 + ], + "type": "text", + "content": "Figure 6. The qualitative results of LIIF [17], LTE [21], and our CLIT using RDN [9] as the encoder and non-integer upsampling scales." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 357, + 288, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 357, + 288, + 407 + ], + "spans": [ + { + "bbox": [ + 46, + 357, + 288, + 407 + ], + "type": "text", + "content": "Note that RDN [9] is trained and evaluated for specific upsampling scales. CLIT outperforms those existing methods in most cases across all datasets and scales, even achieving a 0.26dB PSNR improvement on Set5 for the " + }, + { + "bbox": [ + 46, + 357, + 288, + 407 + ], + "type": "inline_equation", + "content": "\\times 8" + }, + { + "bbox": [ + 46, + 357, + 288, + 407 + ], + "type": "text", + "content": " scale." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 414, + 287, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 414, + 287, + 628 + ], + "spans": [ + { + "bbox": [ + 46, + 414, + 287, + 628 + ], + "type": "text", + "content": "Qualitative results. Fig. 5 compares the qualitative results of CLIT with the baseline methods LIIF [17] and LTE [21], on various dataset DIV2k [48], Set14 [50], B100 [51], and Urban100 [52] with variant upsampling scales. The official codes provided by them were used to produce the results. In the first row, the SR visualization results with a " + }, + { + "bbox": [ + 46, + 414, + 287, + 628 + ], + "type": "inline_equation", + "content": "\\times 12" + }, + { + "bbox": [ + 46, + 414, + 287, + 628 + ], + "type": "text", + "content": " upsampling scale are depicted. It can be observed that both LIIF and LTE struggle to reconstruct the letters continuously, whereas the CLIT result demonstrates continuity for these alphabets, particularly for 'M' and 'S'. In the second row, the stripes on the zebra appear blurry in the LIIF and LTE results, whereas they are more distinct in the CLIT result. In the third row, the cross pattern on the tablecloth is not clearly rendered by LIIF and LTE. In contrast, CLIT is able to successfully generate clean crosses and sharp lines. In the fourth row, despite the blurriness of the original LR image, CLIT captures the texture and produces straight lines on the flag, showcasing its effectiveness." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 630, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 630, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 630, + 288, + 715 + ], + "type": "text", + "content": "Fig. 6 presents the results of text image enhancement employing progressively increasing non-integer upsampling factors, alongside a comparison with earlier methods, LIIF [17] and LTE [21]. The input image first undergoes a downscaling process by a factor of 2.2 to generate an LR image. Subsequently, the LR image is then upscaled using predetermined non-integer factors " + }, + { + "bbox": [ + 46, + 630, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\times 1.6" + }, + { + "bbox": [ + 46, + 630, + 288, + 715 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 630, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\times 2.5" + }, + { + "bbox": [ + 46, + 630, + 288, + 715 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 630, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\times 3.4" + }, + { + "bbox": [ + 46, + 630, + 288, + 715 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 630, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\times 4.3" + }, + { + "bbox": [ + 46, + 630, + 288, + 715 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 307, + 356, + 545, + 542 + ], + "blocks": [ + { + "bbox": [ + 307, + 356, + 545, + 542 + ], + "lines": [ + { + "bbox": [ + 307, + 356, + 545, + 542 + ], + "spans": [ + { + "bbox": [ + 307, + 356, + 545, + 542 + ], + "type": "image", + "image_path": "b8439088d65aa6981d1b276c707e5fe6fdc5623441f9c4b85d08480d415e21a2.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 544, + 545, + 564 + ], + "lines": [ + { + "bbox": [ + 305, + 544, + 545, + 564 + ], + "spans": [ + { + "bbox": [ + 305, + 544, + 545, + 564 + ], + "type": "text", + "content": "Figure 7. A visualization of the local attention maps of the coordinates highlight as red dots." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 568, + 545, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 568, + 545, + 653 + ], + "spans": [ + { + "bbox": [ + 304, + 568, + 545, + 653 + ], + "type": "text", + "content": "It can be observed that our proposed model effectively captures the patterns of text within the image, and accurately estimates words and numbers with enhanced sharpness and clarity. The enhancement becomes particularly evident in the word 'infinity' in the first row of contexts, 'SUNSHINE' in the fifth row, and the number '0' in the final row, as each of these elements demonstrates considerable improvements." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 659, + 495, + 673 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 659, + 495, + 673 + ], + "spans": [ + { + "bbox": [ + 306, + 659, + 495, + 673 + ], + "type": "text", + "content": "4.3. Visualization Local Attention Maps" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 677, + 545, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 545, + 715 + ], + "type": "text", + "content": "In Fig. 7, we provide visualizations of local attention maps. The LR images are generated by applying bicubic downsampling to HR images with scale factors" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18263" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 50, + 70, + 287, + 145 + ], + "blocks": [ + { + "bbox": [ + 50, + 70, + 287, + 145 + ], + "lines": [ + { + "bbox": [ + 50, + 70, + 287, + 145 + ], + "spans": [ + { + "bbox": [ + 50, + 70, + 287, + 145 + ], + "type": "table", + "html": "
In-distributionOut-of-distribution
×2×3×4×6×8×12
LIT34.7931.0729.1026.8425.4723.77
LIT (-a)34.7131.0029.0226.7625.4023.70
LIT (-f)34.7831.0729.0926.8325.4623.77
LIT (-c)34.6531.0029.0426.8025.4523.77
LIT (+e)34.7931.0829.1126.8525.4823.78
", + "image_path": "e39d5472bde49fccf2d1db39d07238c5bc0af4080b256a344ffd3c8eba63838d.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 190, + 288, + 286 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 190, + 288, + 286 + ], + "spans": [ + { + "bbox": [ + 46, + 190, + 288, + 286 + ], + "type": "inline_equation", + "content": "\\{ \\times 4.5, \\times 6, \\times 6 \\}" + }, + { + "bbox": [ + 46, + 190, + 288, + 286 + ], + "type": "text", + "content": ". Subsequently, the proposed CLIT with RDN [9] encoder is utilized to produce HR predictions. It can be observed that the attention maps closely align with the edges, indicating that the cross-scale local attention block effectively captures similar latent codes within a local area. This substantiates that our proposed design enables the generation of SR images from LR counterparts, ultimately producing clean and sharp edges in the output." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 293, + 147, + 304 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 293, + 147, + 304 + ], + "spans": [ + { + "bbox": [ + 47, + 293, + 147, + 304 + ], + "type": "text", + "content": "4.4. Ablation Studies" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 311, + 287, + 371 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 311, + 287, + 371 + ], + "spans": [ + { + "bbox": [ + 46, + 311, + 287, + 371 + ], + "type": "text", + "content": "In this section, we present a series of ablation analyses to substantiate the design decisions proposed in this paper. All of the ablation experiments are conducted on the DIV2K [48] validation set, utilizing the EDSR-baseline [7] as the encoder and employing the PSNR metric for evaluation." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 373, + 287, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 373, + 287, + 529 + ], + "spans": [ + { + "bbox": [ + 46, + 373, + 287, + 529 + ], + "type": "text", + "content": "Validation of the design choices. Table 3 presents a summary of the quantitative contributions associated with each LIT component. A significant improvement is observed when adopting the cross-scale local attention block, while a relatively minor gain with the frequency encoding block by comparing LTE with LTE (-a) and LTE (-f), respectively. In addition, LTE (-c) reveals that removing cell decoding leads to more severe performance degradation for in-distribution upsampling scales compared to out-of-distribution counterparts, highlighting the importance of cell decoding for in-distribution upsampling scales. Finally, incorporating the local ensemble technique in LIT (+e) results in a further, albeit modest, enhancement of the overall performance." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 531, + 287, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 531, + 287, + 628 + ], + "spans": [ + { + "bbox": [ + 46, + 531, + 287, + 628 + ], + "type": "text", + "content": "The effectiveness of the local grid. Table 4 provides a performance comparison of various local grid sizes employed within the local coordinate sampling scheme, as discussed in Section 3. The results presented in Table 4 reveal that increasing the local grid size contributes to performance enhancements, albeit at the cost of extended training times. As a result, this study adopts a " + }, + { + "bbox": [ + 46, + 531, + 287, + 628 + ], + "type": "inline_equation", + "content": "7 \\times 7" + }, + { + "bbox": [ + 46, + 531, + 287, + 628 + ], + "type": "text", + "content": " local grid size to strike an ideal balance between effectiveness and efficiency." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 630, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 630, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 630, + 287, + 715 + ], + "type": "text", + "content": "Analysis on the training strategy. Table 5 presents a quantitative comparison of the proposed cumulative training strategy with the other training strategies for training an LIT. The baseline strategy trains LIT using upsampling scales sampled from a uniform distribution, " + }, + { + "bbox": [ + 46, + 630, + 287, + 715 + ], + "type": "inline_equation", + "content": "r \\sim \\mathcal{U}(1,4)" + }, + { + "bbox": [ + 46, + 630, + 287, + 715 + ], + "type": "text", + "content": ". Expanding the sampling scale distribution to " + }, + { + "bbox": [ + 46, + 630, + 287, + 715 + ], + "type": "inline_equation", + "content": "r \\sim \\mathcal{U}(1,12)" + }, + { + "bbox": [ + 46, + 630, + 287, + 715 + ], + "type": "text", + "content": " improves the performance of large-scale upsampling while" + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 309, + 69, + 547, + 144 + ], + "blocks": [ + { + "bbox": [ + 46, + 147, + 287, + 186 + ], + "lines": [ + { + "bbox": [ + 46, + 147, + 287, + 186 + ], + "spans": [ + { + "bbox": [ + 46, + 147, + 287, + 186 + ], + "type": "text", + "content": "Table 3. The average PSNR (dB) for different design choices. The notations -a/f/c denote the removal of the cross-scale local attention block, frequency encoding block, and cell size, respectively, and +e denotes the addition of local ensemble. The best performing one is indicated in bolded." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 309, + 69, + 547, + 144 + ], + "lines": [ + { + "bbox": [ + 309, + 69, + 547, + 144 + ], + "spans": [ + { + "bbox": [ + 309, + 69, + 547, + 144 + ], + "type": "table", + "html": "
Local gird\nGh × GwIn-distributionOut-of-distributionTraining time
×2×3×4×6×8×12
1 × 134.7131.0029.0226.7625.4023.7020.9
3 × 334.7631.0429.0726.8125.4623.7623.2
5 × 534.7831.0729.0826.8325.4723.7828.4
7 × 734.7931.0729.1026.8425.4723.7733.6
9 × 934.7931.0629.1026.8425.4723.7753.0
15 × 1534.8331.1029.1126.8625.4923.7876.1
25 × 2534.8631.1329.1526.8825.5023.80190.9
", + "image_path": "2066e509767a63192a25f6193caaf3b3832a55f5940af776f7cac64b5e97f7cc.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 308, + 168, + 545, + 211 + ], + "blocks": [ + { + "bbox": [ + 305, + 146, + 545, + 166 + ], + "lines": [ + { + "bbox": [ + 305, + 146, + 545, + 166 + ], + "spans": [ + { + "bbox": [ + 305, + 146, + 545, + 166 + ], + "type": "text", + "content": "Table 4. The average PSNR (dB) and the training time (hour) of multiple local grid sizes. The bolded numbers correspond to the best performance." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 308, + 168, + 545, + 211 + ], + "lines": [ + { + "bbox": [ + 308, + 168, + 545, + 211 + ], + "spans": [ + { + "bbox": [ + 308, + 168, + 545, + 211 + ], + "type": "table", + "html": "
Training strategy×2×3×4×6×8×12
Training with r ~ U(1, 4)34.7931.0729.1026.8425.4723.77
Training with r ~ U(1, 12)34.6931.0629.1126.8725.5223.82
Alternative training strategy34.7531.0929.1226.8825.5223.81
Cumulative training strategy34.7831.1029.1426.8925.5423.83
", + "image_path": "a7d9beeca4edb2adb023d68b58b181c25f4d46f7d7eaff77c4e839c450e64ca3.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 213, + 545, + 232 + ], + "lines": [ + { + "bbox": [ + 305, + 213, + 545, + 232 + ], + "spans": [ + { + "bbox": [ + 305, + 213, + 545, + 232 + ], + "type": "text", + "content": "Table 5. The average PSNR (dB) of different training strategies. The best performing results are indicated in **bold**." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 235, + 545, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 235, + 545, + 415 + ], + "spans": [ + { + "bbox": [ + 304, + 235, + 545, + 415 + ], + "type": "text", + "content": "compromising the performance of small-scale upsampling. To achieve high-quality results across all upsampling scale ranges, the alternative training strategy trains LIT alternatively by switching the sampling scales between " + }, + { + "bbox": [ + 304, + 235, + 545, + 415 + ], + "type": "inline_equation", + "content": "r \\sim \\mathcal{U}(1,4)" + }, + { + "bbox": [ + 304, + 235, + 545, + 415 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 235, + 545, + 415 + ], + "type": "inline_equation", + "content": "r \\sim \\mathcal{U}(4,12)" + }, + { + "bbox": [ + 304, + 235, + 545, + 415 + ], + "type": "text", + "content": ". On the other hand, the proposed cumulative training strategy first trains LIT with " + }, + { + "bbox": [ + 304, + 235, + 545, + 415 + ], + "type": "inline_equation", + "content": "r \\sim \\mathcal{U}(1,4)" + }, + { + "bbox": [ + 304, + 235, + 545, + 415 + ], + "type": "text", + "content": ", followed by fine-tuning using the alternative training strategy. This progressive approach improves the performance of LIT by gradually broadening the distribution of the upsampling scales. Comparing the results of the cumulatively trained LIT with the results of EDSR-baseline-CLIT in Table 1 reveals that the proposed cumulative training strategy works for both LIT and CLIT. Moreover, the results from these tables suggest that the cascaded framework is more effective in addressing arbitrary-scale SR tasks." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 417, + 378, + 429 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 417, + 378, + 429 + ], + "spans": [ + { + "bbox": [ + 306, + 417, + 378, + 429 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 432, + 545, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 432, + 545, + 635 + ], + "spans": [ + { + "bbox": [ + 304, + 432, + 545, + 635 + ], + "type": "text", + "content": "In this paper, we introduced the attention mechanism and frequency encoding technique to address arbitrary-scale SR tasks. To achieve this objective, we proposed the LIT framework, which consists of a cross-scale local attention block and a local frequency encoding block. The former is designed to find the local latent embedding to reconstruct the corresponding RGB value, while the latter focuses on encoding coordinates with the frequency information derived from the feature embeddings. To enhance the capability of capturing fine details, we proposed CLIT and the corresponding cumulative training strategy that trains the model with progressively increasing upsampling scales. Based on the experimental results, both quantitative and qualitative assessments showcase the superior performance of LIT and CLIT in comparison to existing methods. Furthermore, our comprehensive analyses validated the effectiveness of the training strategies and the components employed by them." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 639, + 408, + 651 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 639, + 408, + 651 + ], + "spans": [ + { + "bbox": [ + 306, + 639, + 408, + 651 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "content": "The authors gratefully acknowledge the support from the National Science and Technology Council (NSTC) in Taiwan under grant numbers MOST 111-2223-E-007-004-MY3 and MOST 111-2628-E-007-010, as well as the financial and technical support from MediaTek Inc., Taiwan." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18264" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 714 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 134 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 134 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 134 + ], + "type": "text", + "content": "[1] C. Dong, C. C. Loy, K. He, and X. Tang. Image superresolution using deep convolutional networks. IEEE Trans. Pattern Analysis and Machine Intelligence (TPAMI), 38(2): 295-307, 2016. 1, 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 143, + 288, + 186 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 143, + 288, + 186 + ], + "spans": [ + { + "bbox": [ + 53, + 143, + 288, + 186 + ], + "type": "text", + "content": "[2] J. Kim, J. K. Lee, and K. M. Lee. Accurate image superresolution using very deep convolutional networks. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), pages 1646-1654, 2016." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 194, + 287, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 194, + 287, + 228 + ], + "spans": [ + { + "bbox": [ + 53, + 194, + 287, + 228 + ], + "type": "text", + "content": "[3] C. Dong, C. C. Loy, and X. Tang. Accelerating the superresolution convolutional neural network. In Proc. European Conf. on Computer Vision (ECCV), pages 391-407, 2016. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 235, + 286, + 289 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 235, + 286, + 289 + ], + "spans": [ + { + "bbox": [ + 53, + 235, + 286, + 289 + ], + "type": "text", + "content": "[4] W. Shi, J. Caballero, F. Huszar, J. Totz, A. P. Aitken, et al. Real-time single image and video super-resolution using an efficient sub-pixel convolutional neural network. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), pages 1874-1883, 2016. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 297, + 286, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 297, + 286, + 341 + ], + "spans": [ + { + "bbox": [ + 53, + 297, + 286, + 341 + ], + "type": "text", + "content": "[5] W. Lai, J. Huang, N. Ahuja, and M. Yang. Fast and accurate image super-resolution with deep laplacian pyramid networks. IEEE Trans. Pattern Analysis and Machine Intelligence (TPAMI), 41(11):2599-2613, 2019. 2, 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 349, + 286, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 349, + 286, + 403 + ], + "spans": [ + { + "bbox": [ + 53, + 349, + 286, + 403 + ], + "type": "text", + "content": "[6] C. Ledig, L. Theis, F. Huszar, J. Caballero, A. Cunningham, et al. Photo-realistic single image super-resolution using a generative adversarial network. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), pages 105-114, 2017. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 411, + 286, + 465 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 411, + 286, + 465 + ], + "spans": [ + { + "bbox": [ + 53, + 411, + 286, + 465 + ], + "type": "text", + "content": "[7] B. Lim, S. Son, H. Kim, S. Nah, and K. M. Lee. Enhanced deep residual networks for single image super-resolution. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition Workshop (CVPRW), pages 1132-1140, 2017. 2, 3, 5, 6, 8" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 474, + 286, + 518 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 474, + 286, + 518 + ], + "spans": [ + { + "bbox": [ + 53, + 474, + 286, + 518 + ], + "type": "text", + "content": "[8] X. Wang, K. Yu, S. Wu, J. Gu, Y. Liu, et al. Esrgan: Enhanced super-resolution generative adversarial networks. In Proc. European Conf. on Computer Vision Workshop (ECCVW), pages 63-79, 2018. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 525, + 286, + 570 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 525, + 286, + 570 + ], + "spans": [ + { + "bbox": [ + 53, + 525, + 286, + 570 + ], + "type": "text", + "content": "[9] Y. Zhang, Y. Tian, Y. Kong, B. Zhong, and Y. Fu. Residual dense network for image super-resolution. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), pages 2472-2481, 2018. 2, 3, 6, 7, 8" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 576, + 286, + 620 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 576, + 286, + 620 + ], + "spans": [ + { + "bbox": [ + 48, + 576, + 286, + 620 + ], + "type": "text", + "content": "[10] Y. Zhang, K. Li, K. Li, L. Wang, B. Zhong, et al. Image super-resolution using very deep residual channel attention networks. In Proc. European Conf. on Computer Vision (ECCV), pages 294-310, 2018. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 628, + 286, + 672 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 628, + 286, + 672 + ], + "spans": [ + { + "bbox": [ + 48, + 628, + 286, + 672 + ], + "type": "text", + "content": "[11] T. Dai, J. Cai, Y. Zhang, S. Xia, and L. Zhang. Second-order attention network for single image super-resolution. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), pages 11065-11074, 2019." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 680, + 286, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 680, + 286, + 714 + ], + "spans": [ + { + "bbox": [ + 48, + 680, + 286, + 714 + ], + "type": "text", + "content": "[12] Y. Mei, Y. Fan, Y. Zhou, L. Huang, T. S. Huang, et al. Image super-resolution with cross-scale non-local attention and exhaustive self-exemplars mining. In Proc. IEEE Conf." + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "type": "text", + "content": "on Computer Vision and Pattern Recognition (CVPR), pages 5689-5698, 2020. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 308, + 102, + 545, + 145 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 102, + 545, + 145 + ], + "spans": [ + { + "bbox": [ + 308, + 102, + 545, + 145 + ], + "type": "text", + "content": "[13] H. Chen, Y. Wang, T. Guo, C. Xu, Y. Deng, et al. Pre-trained image processing transformer. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), pages 12294-12305, 2021. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 152, + 545, + 197 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 152, + 545, + 197 + ], + "spans": [ + { + "bbox": [ + 307, + 152, + 545, + 197 + ], + "type": "text", + "content": "[14] J. Liang, J. Cao, G. Sun, K. Zhang, L. Van Gool, et al. Swinir: Image restoration using swim transformer. In Proc. IEEE Int. Conf. on Computer Vision Workshop (ICCVW), pages 1833-1844, 2021. 3, 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 204, + 545, + 236 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 204, + 545, + 236 + ], + "spans": [ + { + "bbox": [ + 308, + 204, + 545, + 236 + ], + "type": "text", + "content": "[15] X. Chen, X. Wang, J. Zhou, and C. Dong. Activating more pixels in image super-resolution transformer. CoRR, abs/2205.04437, 2022. 1, 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 243, + 545, + 287 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 243, + 545, + 287 + ], + "spans": [ + { + "bbox": [ + 308, + 243, + 545, + 287 + ], + "type": "text", + "content": "[16] X. Hu, H. Mu, X. Zhang, Z. Wang, T. Tan, et al. Meta-sr: A magnification-arbitrary network for super-resolution. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), pages 1575-1584, 2019. 1, 2, 5, 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 295, + 545, + 339 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 295, + 545, + 339 + ], + "spans": [ + { + "bbox": [ + 308, + 295, + 545, + 339 + ], + "type": "text", + "content": "[17] Y. Chen, S. Liu, and X. Wang. Learning continuous image representation with local implicit image function. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), pages 8628-8638, 2021. 1, 2, 4, 5, 6, 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 346, + 545, + 379 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 346, + 545, + 379 + ], + "spans": [ + { + "bbox": [ + 308, + 346, + 545, + 379 + ], + "type": "text", + "content": "[18] X. Xu, Z. Wang, and H. Shi. Ultrasr: Spatial encoding is a missing key for implicit image function-based arbitrary-scale super-resolution. CoRR, abs/2103.12716, 2021. 2, 5, 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 385, + 545, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 385, + 545, + 418 + ], + "spans": [ + { + "bbox": [ + 308, + 385, + 545, + 418 + ], + "type": "text", + "content": "[19] Y. Liu, Y. Guo, and S. Zhang. Enhancing multi-scale implicit learning in image super-resolution with integrated positional encoding. CoRR, abs/2112.05756, 2021. 2, 4, 5, 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 426, + 545, + 469 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 426, + 545, + 469 + ], + "spans": [ + { + "bbox": [ + 308, + 426, + 545, + 469 + ], + "type": "text", + "content": "[20] J. Yang, S. Shen, H. Yue, and K. Li. Implicit transformer network for screen content image continuous super-resolution. In Proc. Conf. on Neural Information Processing Systems (NeurIPS), pages 13304-13315, 2021." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 476, + 545, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 476, + 545, + 520 + ], + "spans": [ + { + "bbox": [ + 308, + 476, + 545, + 520 + ], + "type": "text", + "content": "[21] J. Lee and K. H. Jin. Local texture estimator for implicit representation function. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), pages 1929-1938, 2022. 1, 2, 4, 5, 6, 7" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 527, + 545, + 570 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 527, + 545, + 570 + ], + "spans": [ + { + "bbox": [ + 308, + 527, + 545, + 570 + ], + "type": "text", + "content": "[22] A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, et al. Attention is all you need. In Proc. Conf. on Neural Information Processing Systems (NeurIPS), pages 5998-6008, 2017. 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 578, + 545, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 578, + 545, + 611 + ], + "spans": [ + { + "bbox": [ + 308, + 578, + 545, + 611 + ], + "type": "text", + "content": "[23] Z. Liu, Y. Lin, Y. Cao, H. Hu, Y. Wei, et al. Swin transformer: Hierarchical vision transformer using shifted windows. In Proc. IEEE Int. Conf. on Computer Vision (ICCV), 2021. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 618, + 545, + 662 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 618, + 545, + 662 + ], + "spans": [ + { + "bbox": [ + 308, + 618, + 545, + 662 + ], + "type": "text", + "content": "[24] W. Wang, L. Yao, L. Chen, B. Lin, D. Cai, et al. Crossformer: A versatile vision transformer hinging on cross-scale attention. In Proc. Int. Conf. on Learning Representations (ICLR), 2022. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 669, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 669, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 669, + 545, + 713 + ], + "type": "text", + "content": "[25] B. Mildenhall, P. P. Srinivasan, M. Tancik, J. T. Barron, R. Ramamoorthi, et al. Nerf: Representing scenes as neural radiance fields for view synthesis. In Proc. European Conf. on Computer Vision (ECCV), pages 405-421, 2020. 2" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18265" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 128 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 128 + ], + "type": "text", + "content": "[26] M. Tancik, P. P. Srinivasan, B. Mildenhall, S. Fridovich-Keil, N. Raghavan, et al. Fourier features let networks learn high frequency functions in low dimensional domains. In Proc. Conf. on Neural Information Processing Systems (NeurIPS), pages 7537–7547, 2020. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 136, + 287, + 169 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 136, + 287, + 169 + ], + "spans": [ + { + "bbox": [ + 48, + 136, + 287, + 169 + ], + "type": "text", + "content": "[27] N. Rahaman, A. Baratin, D. Arpit, F. Draxler, M. Lin, et al. On the spectral bias of neural networks. In Proc. Int. Conf. on Machine Learning (ICML), pages 5301-5310, 2019. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 177, + 287, + 211 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 177, + 287, + 211 + ], + "spans": [ + { + "bbox": [ + 48, + 177, + 287, + 211 + ], + "type": "text", + "content": "[28] Z. Chen and H. Zhang. Learning implicit fields for generative shape modeling. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), pages 5939-5948, 2019. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 218, + 287, + 273 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 218, + 287, + 273 + ], + "spans": [ + { + "bbox": [ + 48, + 218, + 287, + 273 + ], + "type": "text", + "content": "[29] J. J. Park, P. Florence, J. Straub, R. A. Newcombe, and S. Lovegrove. Deepsdf: Learning continuous signed distance functions for shape representation. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), pages 165-174, 2019." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 281, + 287, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 281, + 287, + 335 + ], + "spans": [ + { + "bbox": [ + 48, + 281, + 287, + 335 + ], + "type": "text", + "content": "[30] L. M. Mescheder, M. Oechsle, M. Niemeyer, S. Nowozin, and A. Geiger. Occupancy networks: Learning 3d reconstruction in function space. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), pages 4460-4470, 2019." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 344, + 287, + 388 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 344, + 287, + 388 + ], + "spans": [ + { + "bbox": [ + 48, + 344, + 287, + 388 + ], + "type": "text", + "content": "[31] K. Genova, F. Cole, A. Sud, A. Sarna, and T. A. Funkhouser. Local deep implicit functions for 3d shape. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), pages 4856-4865, 2020." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 396, + 287, + 440 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 396, + 287, + 440 + ], + "spans": [ + { + "bbox": [ + 48, + 396, + 287, + 440 + ], + "type": "text", + "content": "[32] M. Atzmon and Y. Lipman. SAL: sign agnostic learning of shapes from raw data. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), pages 2562-2571, 2020. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 449, + 287, + 494 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 449, + 287, + 494 + ], + "spans": [ + { + "bbox": [ + 48, + 449, + 287, + 494 + ], + "type": "text", + "content": "[33] V. Sitzmann, M. Zollhöfer, and G. Wetzstein. Scene representation networks: Continuous 3d-structure-aware neural scene representations. In Proc. Conf. on Neural Information Processing Systems (NeurIPS), pages 1119–1130, 2019. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 501, + 287, + 545 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 501, + 287, + 545 + ], + "spans": [ + { + "bbox": [ + 48, + 501, + 287, + 545 + ], + "type": "text", + "content": "[34] C. M. Jiang, A. Sud, A. Makadia, J. Huang, M. Nießner, et al. Local implicit grid representations for 3d scenes. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), pages 6000-6009, 2020." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 553, + 287, + 597 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 553, + 287, + 597 + ], + "spans": [ + { + "bbox": [ + 48, + 553, + 287, + 597 + ], + "type": "text", + "content": "[35] R. Chabra, J. E. Lenssen, E. Ilg, T. Schmidt, J. Straub, et al. Deep local shapes: Learning local SDF priors for detailed 3d reconstruction. In Proc. European Conf. on Computer Vision (ECCV), pages 608-625, 2020." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 605, + 287, + 649 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 605, + 287, + 649 + ], + "spans": [ + { + "bbox": [ + 48, + 605, + 287, + 649 + ], + "type": "text", + "content": "[36] S. Peng, M. Niemeyer, L. M. Mescheder, M. Pollefeys, and A. Geiger. Convolutional occupancy networks. In Proc. European Conf. on Computer Vision (ECCV), pages 523-540, 2020. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 658, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 287, + 712 + ], + "type": "text", + "content": "[37] M. Niemeyer, L. Mescheder, M. Oechsle, and A. Geiger. Differentiable volumetric rendering: Learning implicit 3d representations without 3d supervision. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), pages 3501-3512, 2020. 2" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 72, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 25, + "blocks": [ + { + "bbox": [ + 307, + 72, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 72, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 307, + 72, + 545, + 106 + ], + "type": "text", + "content": "[38] L. Liu, J. Gu, K. Z. Lin, T. Chua, and C. Theobalt. Neural sparse voxel fields. In Proc. Conf. on Neural Information Processing Systems (NeurIPS), 2020." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 307, + 114, + 545, + 169 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 114, + 545, + 169 + ], + "spans": [ + { + "bbox": [ + 307, + 114, + 545, + 169 + ], + "type": "text", + "content": "[39] J. T. Barron, B. Mildenhall, M. Tancik, P. Hedman, R. Martin-Brualla, et al. Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In Proc. IEEE Int. Conf. on Computer Vision (ICCV), pages 5835–5844, 2021. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 178, + 545, + 222 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 178, + 545, + 222 + ], + "spans": [ + { + "bbox": [ + 307, + 178, + 545, + 222 + ], + "type": "text", + "content": "[40] S. Klocek, L. Maziarka, M. Wolczyk, J. Tabor, J. Nowak, et al. Hypernetwork functional image representation. In Artificial Neural Networks and Machine Learning, pages 496-510, 2019. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 232, + 545, + 286 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 232, + 545, + 286 + ], + "spans": [ + { + "bbox": [ + 307, + 232, + 545, + 286 + ], + "type": "text", + "content": "[41] V. Sitzmann, J. N. P. Martel, A. W. Bergman, D. B. Lindell, and G. Wetzstein. Implicit neural representations with periodic activation functions. In Proc. Conf. on Neural Information Processing Systems (NeurIPS), pages 7462-7473, 2020. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 296, + 545, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 296, + 545, + 330 + ], + "spans": [ + { + "bbox": [ + 307, + 296, + 545, + 330 + ], + "type": "text", + "content": "[42] K. Grm, W. J. Scheirer, and V. Struc. Face hallucination using cascaded super-resolution and identity priors. IEEE Trans. Image Processing, pages 2150-2165, 2020. 2, 5" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 338, + 545, + 383 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 338, + 545, + 383 + ], + "spans": [ + { + "bbox": [ + 307, + 338, + 545, + 383 + ], + "type": "text", + "content": "[43] Z. Fu, Y. Kong, Y. Zheng, H. Ye, W. Hu, et al. Cascaded detail-preserving networks for super-resolution of document images. In Int. Conf. on Doc. Analysis Recognition, ICDAR, pages 240-245, 2019. 2, 5" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 392, + 545, + 436 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 392, + 545, + 436 + ], + "spans": [ + { + "bbox": [ + 307, + 392, + 545, + 436 + ], + "type": "text", + "content": "[44] R. Basri, M. Galun, A. Geifman, D. W. Jacobs, Y. Kasten, et al. Frequency bias in neural networks for input of nonuniform density. In Proc. Int. Conf. on Machine Learning (ICML), pages 685-694, 2020. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 445, + 545, + 500 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 445, + 545, + 500 + ], + "spans": [ + { + "bbox": [ + 307, + 445, + 545, + 500 + ], + "type": "text", + "content": "[45] M. Tancik, B. Mildenhall, T. Wang, D. Schmidt, P. P. Srinivasan, et al. Learned initializations for optimizing coordinate-based neural representations. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), pages 2846-2855, 2021. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 509, + 545, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 509, + 545, + 542 + ], + "spans": [ + { + "bbox": [ + 307, + 509, + 545, + 542 + ], + "type": "text", + "content": "[46] D. Hendrycks and K. Gimpel. Bridging nonlinearities and stochastic regularizers with gaussian error linear units. CoRR, abs/1606.08415, 2016. 4" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 552, + 545, + 606 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 552, + 545, + 606 + ], + "spans": [ + { + "bbox": [ + 307, + 552, + 545, + 606 + ], + "type": "text", + "content": "[47] J. Lee-Thorp, J. Ainslie, I. Eckstein, and S. Ontanon. FNet: Mixing tokens with fourier transforms. In Proc. Conf. of North American Chapter of the Association for Computational Linguistics: Human Language Technologies (NAACL), pages 4296-4313, 2022. 4" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 616, + 545, + 660 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 616, + 545, + 660 + ], + "spans": [ + { + "bbox": [ + 307, + 616, + 545, + 660 + ], + "type": "text", + "content": "[48] E. Agustsson and R. Timofte. NTIRE 2017 challenge on single image super-resolution: Dataset and study. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition Workshop (CVPRW), pages 1122-1131, 2017. 5, 6, 7, 8" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 669, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 669, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 307, + 669, + 545, + 712 + ], + "type": "text", + "content": "[49] M. Bevilacqua, A. Roumy, C. Guillemot, and M. Alberi-Morel. Low-complexity single-image super-resolution based on nonnegative neighbor embedding. In Proc. British Machine Vision Conf. (BMVC), pages 1-10, 2012. 5, 6" + } + ] + } + ], + "index": 24 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "text", + "content": "18266" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 272 + ], + "type": "list", + "angle": 0, + "index": 4, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "text", + "content": "[50] R. Zeyde, M. Elad, and M. Protter. On single image scale-up using sparse-representations. In *Curves and Surfaces*, volume 6920 of Lecture Notes in Computer Science, pages 711-730, 2010. 5, 6, 7" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 125, + 287, + 179 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 125, + 287, + 179 + ], + "spans": [ + { + "bbox": [ + 48, + 125, + 287, + 179 + ], + "type": "text", + "content": "[51] D. R. Martin, C. C. Fowlkes, D. Tal, and J. Malik. A database of human segmented natural images and its application to evaluating segmentation algorithms and measuring ecological statistics. In Proc. IEEE Int. Conf. on Computer Vision (ICCV), pages 416-425, 2001. 5, 6, 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 188, + 287, + 232 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 188, + 287, + 232 + ], + "spans": [ + { + "bbox": [ + 48, + 188, + 287, + 232 + ], + "type": "text", + "content": "[52] J. Huang, A. Singh, and N. Ahuja. Single image superresolution from transformed self-exemplars. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), pages 5197-5206, 2015. 5, 6, 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 239, + 287, + 272 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 239, + 287, + 272 + ], + "spans": [ + { + "bbox": [ + 48, + 239, + 287, + 272 + ], + "type": "text", + "content": "[53] D. P. Kingma and J. Ba. Adam: A method for stochastic optimization. In Proc. Int. Conf. on Learning Representations (ICLR), 2015. 6" + } + ] + } + ], + "index": 3 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "18267" + } + ] + } + ], + "index": 5 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Castling-ViT_ Compressing Self-Attention via Switching Towards Linear-Angular Attention at Vision Transformer Inference/d74c49bb-effe-4caf-8858-0066e120e08d_content_list.json b/2023/Castling-ViT_ Compressing Self-Attention via Switching Towards Linear-Angular Attention at Vision Transformer Inference/d74c49bb-effe-4caf-8858-0066e120e08d_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..ef71e26832890eb3c563cb6de1afbc0b4c31ba26 --- /dev/null +++ b/2023/Castling-ViT_ Compressing Self-Attention via Switching Towards Linear-Angular Attention at Vision Transformer Inference/d74c49bb-effe-4caf-8858-0066e120e08d_content_list.json @@ -0,0 +1,1558 @@ +[ + { + "type": "text", + "text": "Castling-ViT: Compressing Self-Attention via Switching Towards Linear-Angular Attention at Vision Transformer Inference", + "text_level": 1, + "bbox": [ + 137, + 130, + 825, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Haoran You $^{1,2,\\dagger,*}$ , Yunyang Xiong $^{2,*}$ , Xiaoliang Dai $^{2}$ , Bichen Wu $^{2}$ , Peizhao Zhang $^{2}$ , Haoqi Fan $^{2}$ , Peter Vajda $^{2}$ , Yingyan (Celine) Lin $^{1}$ $^{1}$ Georgia Institute of Technology, $^{2}$ Meta Research \n{haoran.you, celine.lin}@gatech.edu \n{yunyang, xiaoliang dai, wbc, stzpz, haoqifan, vajdap}@meta.com", + "bbox": [ + 156, + 193, + 812, + 282 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 306, + 313, + 321 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Vision Transformers (ViTs) have shown impressive performance but still require a high computation cost as compared to convolutional neural networks (CNNs), one reason is that ViTs' attention measures global similarities and thus has a quadratic complexity with the number of input tokens. Existing efficient ViTs adopt local attention or linear attention, which sacrifice ViTs' capabilities of capturing either global or local context. In this work, we ask an important research question: Can ViTs learn both global and local context while being more efficient during inference? To this end, we propose a framework called Castling-ViT, which trains ViTs using both linear-angular attention and masked softmax-based quadratic attention, but then switches to having only linear-angular attention during inference. Our Castling-ViT leverages angular kernels to measure the similarities between queries and keys via spectral angles. And we further simplify it with two techniques: (1) a novel linear-angular attention mechanism: we decompose the angular kernels into linear terms and high-order residuals, and only keep the linear terms; and (2) we adopt two parameterized modules to approximate high-order residuals: a depthwise convolution and an auxiliary masked softmax attention to help learn global and local information, where the masks for softmax attention are regularized to gradually become zeros and thus incur no overhead during inference. Extensive experiments validate the effectiveness of our Castling-ViT, e.g., achieving up to a $1.8\\%$ higher accuracy or $40\\%$ MACs reduction on classification and 1.2 higher mAP on detection under comparable FLOPs, as compared to ViTs with vanilla softmax-based attentions. Project page is available at here.", + "bbox": [ + 76, + 338, + 472, + 806 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 823, + 209, + 839 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Vision Transformers (ViTs) have made significant progress in image classification, object detection, and many", + "bbox": [ + 76, + 845, + 468, + 876 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/07c00801ece4ca556f08c973c3b493a5ed8898b8012feb40e2288e0cba1fd932.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 506, + 306, + 888, + 428 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/d17453d5505c6838894302c6cb13f95985a07f7942fc59096ef8ab3348323632.jpg", + "image_caption": [ + "Figure 1. Castling-ViT over SOTA baselines on (1) ImageNet [18] image classification and (2) COCO [36] object detection." + ], + "image_footnote": [], + "bbox": [ + 504, + 431, + 888, + 551 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "other applications. It is well recognized that the superior performance achieved by ViTs is largely attributed to their self-attention modules that can better capture global context [20, 57, 64]. Nevertheless, ViTs' powerful self-attention module comes at the cost of quadratic complexity with the number of input tokens, causing a major efficiency bottleneck to ViTs' achievable runtime (i.e., inference latency) [3, 8, 32, 58, 66, 69, 76]. To mitigate this issue, linear attention designs have been developed to alleviate the vanilla ViT attention's quadratic complexity. In particular, existing efforts can be categorized into two clusters: (1) ViTs with local attention by restricting the attention window size [38, 53], sharing the attention queries [2], or representing the attention queries/keys with low rank matrices [58]; and (2) ViTs with kernel-based linear attention, which approximate the non-linearity softmax function by decomposing it into separate kernel embeddings. This enables a change in the matrix computation order for a reduced computational complexity [5, 6, 14, 29, 37, 39, 43, 66].", + "bbox": [ + 496, + 593, + 890, + 880 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Despite their promise in alleviating ViTs' complexity", + "bbox": [ + 517, + 883, + 890, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 44 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Equal contribution. † Work done while interning at Meta Research.", + "bbox": [ + 94, + 886, + 455, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "14431", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "and thus inference runtime, both the local and linear attention compromise ViTs' performance due to the lack of capabilities to capture global or local context. To marry the best of both worlds, we advocate training ViTs with both (1) efficient but less powerful linear attention, i.e., without the high-order residuals in angular kernel expansion, and (2) powerful yet costly softmax-based masked attention. The latter helps approximate high-order residuals at the early training stage while being dropped during inference, based on an assumption that the remaining networks can gradually learn the high-order components at the later training stage [67]. This concept resembles the \"castling\" move in chess when two pieces are moved at once. While it sounds promising, there are still two challenges to achieve this. First, existing linear attention modules still underperform their vanilla softmax-based counterparts. Therefore, a better linear attention is crucial for the final performance. We find that angular kernels perform equally as softmax-based attentions in terms of similarity measurements. While they still suffer from a quadratic complexity, they can be divided into linear terms and high-order residuals. The challenge is how to construct ViTs with only the linear terms. Second, doing so would require that the trained ViTs merely rely on the linear terms towards the end of training, which would call for an approximation of the above high-order residuals. The challenge is that how we can resort to costly but powerful modules to approximate high-order residuals during training but does not incur extra inference cost.", + "bbox": [ + 76, + 90, + 472, + 513 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this work, we develop techniques to tackle those challenges, and make the following contributions:", + "bbox": [ + 76, + 513, + 470, + 546 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose a framework called Castling-ViT, which trains ViTs using both linear-angular attention and masked softmax-based quadratic attention, but then switches to having only linear-angular attentions during ViT inference to save computational costs.", + "- We develop a new linear-angular attention leveraging angular kernels to close the accuracy gap between linear attention and softmax-based attention. It expands angular kernels where linear terms are kept while complex high-order residuals are approximated.", + "- We use two parameterized modules to approximate the high-order residuals above: a depthwise convolution and an auxiliary masked softmax-based attention, where the latter's attention masks are regularized to gradually become zeros to avoid inference overhead.", + "- We conduct extensive experiments to validate the effectiveness of the proposed Castling-ViT. Results on classification, detection, and segmentation tasks consistently demonstrate its superior performance ( $\\uparrow 1.8\\%$ top-1 accuracy or $\\uparrow 1.2$ mAP) or efficiency (40% MACs savings) over state-of-the-art (SOTA) CNNs and ViTs." + ], + "bbox": [ + 96, + 556, + 468, + 888 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 500, + 89, + 650, + 106 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Vision Transformers (ViTs). ViT [20] beats CNNs with a simple encoder-only transformer architecture taking the splitted non-overlapped image patches as sequential inputs, but relies on costly pretraining on a huge JFT-300M dataset [50]. Later, DeiT [54] and T2T-ViT [72] leverage an improved ViT training recipe or enhanced tokenization mechanism to achieve a comparable accuracy without the necessity of costly pretraining. To further improve ViTs' achievable accuracy-efficiency trade-offs, CrossViT [9], PiT [26], PVT [59], MViT [21] and Swin-Transformer [38] propose a pyramid-like architecture, which is commonly used in CNNs [16, 28, 65]; DynamicViT [46], A-ViT [68], ToME [4], and MIA-Former [71] propose to adaptively identify and remove unnecessary input tokens for saving computational costs. With the goal of deploying ViTs in resource-constrained devices, various efficient ViT architectures have been proposed [6, 24, 35, 40]. For example, LeViT [24], CvT [63], and MobileViT [40] adopt more efficient self-attention implementation or incorporate convolutional feature extraction blocks into their early layers; EfficientFormer [35] further enables pure ViTs to run as fast as MobileNets. In contrast, our Castling-ViT explores whether ViTs can learn both global and local features while still being efficient at runtime. Also, we target a generic linear-angular attention that can serve as a drop-in replacement for all kinds of ViT architectures and thus is orthogonal to new ViT architecture designs.", + "bbox": [ + 496, + 112, + 893, + 521 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Efficient ViT Variants. As commonly recognized, ViTs rely heavily on their self-attention module which is however costly due to its quadratic computational complexity with the total number of input tokens [29, 76]. To make the self-attention module more efficient, a surge of linear attention works have been proposed and can be roughly categorized into two groups: local attention [2, 38, 56, 58] or kernel-based linear attention [1, 5, 6, 14, 29, 37, 39, 43, 66]. For kernel-based linear attention, common designs approximate the softmax function [5, 14, 29] or the full self-attention matrix [39, 66] with orthogonal features or kernel embeddings, then the computation order can be changed from (QK)V to Q(KV). For example, [29] and [6] decompose the exponential terms in softmax-based attention into kernel functions and exchange the computation order. Despite their decent performance, currently kernel-based linear attention in general underperform the softmax-based attention. Recent works [8, 17] also unify low rank approximated and sparse attention (can also be dropped at inference) to improve ViTs' accuracy-efficiency tradeoffs. Different from the above works, we explore from a new perspective by taking spectral angles into consideration when measuring the similarities among tokens, resulting in linear-angular attention that can achieve comparable or even better performance than softmax-based attention. More efficient ViT variants", + "bbox": [ + 496, + 523, + 893, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "14432", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "are supplied to the Appendix.", + "bbox": [ + 76, + 90, + 274, + 104 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "ViTs have also been used as the backbones for downstream tasks, e.g., detection [34, 60, 77] and segmentation [11,42]. We supply more literature review to Appendix.", + "bbox": [ + 75, + 106, + 468, + 151 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. The Proposed Methods", + "text_level": 1, + "bbox": [ + 76, + 165, + 295, + 181 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Preliminary of Self-Attention", + "text_level": 1, + "bbox": [ + 76, + 190, + 336, + 205 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Self-Attention. Self-attention module is a core component of the Transformer [20, 57], and usually consists of multiple heads. Each head captures global-context information by measuring pairwise correlations among all $N$ tokens ( $N$ denotes the total number of tokens) as defined below:", + "bbox": [ + 75, + 209, + 470, + 285 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {H} _ {t} ^ {k} = \\sum_ {i = 1} ^ {N} \\frac {\\exp \\left(\\mathbf {Q} _ {t} \\cdot \\mathbf {K} _ {i} ^ {T} / \\sqrt {d}\\right)}{\\sum_ {j = 1} ^ {N} \\exp \\left(\\mathbf {Q} _ {t} \\cdot \\mathbf {K} _ {j} ^ {T} / \\sqrt {d}\\right)} \\cdot \\mathbf {V} _ {i}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 130, + 291, + 468, + 334 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $t\\in \\{1,\\dots ,N\\} ,k\\in \\{1,\\dots ,M\\}$ $\\mathbf{H}_t^k$ refers to the $t$ th row of the $m$ -th head's attention matrix H. $\\mathbf{Q}_t,\\mathbf{K}_t,\\mathbf{V}_t\\in$ $\\mathbf{R}^d$ , are the query, key, and value vectors obtained by linearly projecting the input $\\mathbf{X_t}\\in \\mathbb{R}^D$ with three learnable weight matrices $W^{Q},W^{K},W^{V}\\in \\mathbb{R}^{D\\times d}$ . The attention head first computes the inner product between the query-key pairs, then scales the product results to stabilize the training and uses Softmax to normalize the resulting attention scores, and finally computes a weighted sum over all value vectors. Outputs from all attention heads are then concatenated together before a final linear projection with learnable weights. Note that $\\exp (\\cdot)$ denotes an exponential function. Computing Eq. (1) has a quadratic complexity of $\\mathcal{O}(N^2)$", + "bbox": [ + 75, + 340, + 468, + 537 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Kernel-based Linear Attention. The core idea of linear attention [29, 49, 58] is to decompose the similarity measurement function into separate kernel embeddings, i.e., $\\mathrm{Sim}(\\mathbf{Q},\\mathbf{K})\\approx \\phi (\\mathbf{Q})\\phi (\\mathbf{K})^T$ so that we can change the computation order to $\\phi (\\mathbf{Q})(\\phi (\\mathbf{K})^T\\mathbf{V})$ based on the associative property of matrix multiplication. In this way, the attention complexity is quadratic to the feature dimension $d$ instead of the token length $N$ . One straightforward implementation of linear attention is to use Gaussian RBF kernels to measure the similarity, which can serve as an unbiased estimation of $\\exp (\\langle \\cdot ,\\cdot \\rangle)$ in Eq. (1):", + "bbox": [ + 75, + 537, + 470, + 705 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\exp \\left(\\frac {\\langle \\mathbf {x} , \\mathbf {y} \\rangle}{\\sigma^ {2}}\\right) = \\underbrace {\\exp \\left(\\frac {- \\| \\mathbf {x} - \\mathbf {y} \\| ^ {2}}{2 \\sigma^ {2}}\\right)} _ {\\text {G a u s s i a n R B F K e r n e l}} \\exp \\left(\\frac {\\| \\mathbf {x} \\| ^ {2} + \\| \\mathbf {y} \\| ^ {2}}{2 \\sigma^ {2}}\\right) \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 84, + 710, + 468, + 760 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\langle \\cdot, \\cdot \\rangle$ denotes the inner product operator. According to [45], we can induce a function $\\phi(\\cdot)$ to approximate the Gaussian RBF kernel, mapping the input space to the feature space. Assuming that both $\\mathbf{Q}$ and $\\mathbf{K}$ are normalized as unit row vectors along the feature dimension, then the attention formula can be approximated by:", + "bbox": [ + 75, + 766, + 468, + 858 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {H} _ {t} ^ {k} = \\frac {\\phi (\\mathbf {Q} _ {t}) \\cdot \\sum_ {i = 1} ^ {N} \\left(\\phi \\left(\\mathbf {K} _ {i}\\right) ^ {T} \\cdot \\mathbf {V} _ {i}\\right)}{\\phi (\\mathbf {Q} _ {t}) \\cdot \\sum_ {j = 1} ^ {N} \\phi \\left(\\mathbf {K} _ {j} ^ {T}\\right)} \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 151, + 863, + 468, + 905 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/27ef88097b699bd0084f362693ea21f7f8ebf624aff2d6eee1f482d95dfa630f.jpg", + "image_caption": [ + "Figure 2. Illustration of self-attention used in LeViT and MViTv2." + ], + "image_footnote": [], + "bbox": [ + 501, + 88, + 893, + 309 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "While doing so can alleviate the attention complexity to become linear w.r.t. the token length $N$ , it often causes a nontrivial accuracy drop as compared to the corresponding ViTs with vanilla softmax-based attention [6, 13, 72].", + "bbox": [ + 496, + 335, + 890, + 396 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Revisit Attention Designs. Recent ViTs have achieved a low complexity for classification tasks (e.g., LeViT [24]) but still be costly for downstream tasks due to high input resolutions, i.e., the number of tokens (e.g., MViTv2 [34]). Therefore, one dilemma is the tradeoff between the model efficiency and generalizability. As illustrated in Fig. 2, LeViT shrinks the feature dimension to be more efficient on low-resolution tasks but cannot be well generalized to downstream tasks, due to the resulting (1) feature bottlenecks, i.e., insufficient feature dimensions, and (2) fixed attention biases. On the other hand, MViTv2 performs token pooling for better fitting high-resolution tasks but is less effective for low-resolution due to the caused token bottlenecks, i.e., insufficient number of tokens. Our ablation studies in Table 1 show that (1) for ImageNet classification, both token and feature pooling can lead to accuracy drops in ViTs; and (2) for downsampling layers, post-Q pooling (i.e., pooling after linear projection) together with residual connections in MViTv2 performs better than pre-Q pooling (i.e., pooling during linear projection) in LeViT. Therefore, we develop Castling-LeViTs on top of attention with merely post-Q pooling and residual connections.", + "bbox": [ + 496, + 398, + 892, + 731 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. The Proposed Castling-ViT Framework", + "text_level": 1, + "bbox": [ + 500, + 742, + 841, + 758 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Castling-ViT Overview. Fig. 4 illustrates an overview of the proposed Castling-ViT, which makes linear attention more powerful than previous designs while still being efficient during inference. In particular, we propose (1) a novel kernel-based linear-angular attention from the spectral angle perspective to close the accuracy gap between linear attention and softmax-based attention; and (2) a training augmentation method that leverages softmax-based attention as an auxiliary branch to assist the linear-angular atten", + "bbox": [ + 496, + 763, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "14433", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "table", + "img_path": "images/bbb9916f1a67e1f02942854ed649fd31c72638e0d41d4465e5a5d08b4af6b2bf.jpg", + "table_caption": [ + "Table 1. Analyze the attention design on ImageNet classification." + ], + "table_footnote": [], + "table_body": "
PoolingResidual QMACs (M)Top-1 Accuracy (%)
TokenFeat.Pre-QPost-Q
✓✓68578.00
✓✓66178.19
✓73679.11
✓✓✓77178.73
✓✓✓76377.23
✓✓83880.05
", + "bbox": [ + 80, + 104, + 468, + 200 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "tion only during ViT training. Note that a mask is applied to the auxiliary branch to manifest linear-angular attention and drop the auxiliary branch.", + "bbox": [ + 76, + 210, + 468, + 257 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2.1 Linear-Angular Attention", + "text_level": 1, + "bbox": [ + 76, + 266, + 312, + 281 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Angular Kernel. In addition to the previously adopted polynomial, exponential, or RBF kernel that focuses on spatial similarity measurements [29, 37], we propose to consider measuring spectral similarity via angular kernel as an alternative to existing softmax-based attentions, leading to similar or better performance since it additionally takes into consideration the nature of spectral characteristics, e.g., the spectral angle as a distance measurement function [27, 30]. Such a spectral angle between two vectors is defined as:", + "bbox": [ + 75, + 286, + 468, + 422 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\theta \\left(\\mathbf {x} _ {i}, \\mathbf {x} _ {j}\\right) = \\arccos \\left(\\frac {\\left\\langle \\mathbf {x} _ {i} , \\mathbf {x} _ {j} \\right\\rangle}{\\left\\| \\mathbf {x} _ {i} \\right\\| \\cdot \\left\\| \\mathbf {x} _ {j} \\right\\|}\\right), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 150, + 428, + 468, + 462 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\|\\cdot\\|$ is the Euclidean distance and $\\langle \\cdot, \\cdot \\rangle$ is the inner product. The output range of $\\theta$ is $[0, \\pi]$ . Such an angle can be used as a distance. In our design, we define the angular kernel as a similarity measurement function between the queries $\\mathbf{Q}$ and keys $\\mathbf{K}$ as:", + "bbox": [ + 75, + 467, + 468, + 544 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {S i m} \\left(\\mathbf {Q} _ {i}, \\mathbf {K} _ {j}\\right) = 1 - \\frac {1}{\\pi} \\cdot \\theta \\left(\\mathbf {Q} _ {i}, \\mathbf {K} _ {j}\\right), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 151, + 547, + 468, + 578 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "and the output range is thus [0, 1]. With more aligned $\\mathbf{Q}_i$ and $\\mathbf{K}_j$ , $\\theta$ is closer to 0 and thus the similarity is closer to 1; In contrast, if $\\mathbf{Q}_i$ and $\\mathbf{K}_j$ have opposite features, $\\theta$ is closer to $\\pi$ and thus the similarity is closer to 0.", + "bbox": [ + 76, + 583, + 468, + 643 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Properties of Angular Kernel and Its Feature Space. One property of our angular kernel is that replacing the similarity measurement in self-attention with such a kernel provides an efficient way to implicitly map the input data to a high (even infinite after expansion) dimensional feature space [27], where the distances/angles are calculated based on a rich feature structure. Let $\\phi(\\cdot)$ denotes the implicit map induced by this kernel, the norm of mapped input data is:", + "bbox": [ + 75, + 643, + 468, + 765 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\left\\| \\phi \\left(\\mathbf {x} _ {i}\\right) \\right\\| ^ {2} = \\operatorname {S i m} \\left(\\mathbf {x} _ {i}, \\mathbf {x} _ {i}\\right) = 1 - \\frac {1}{\\pi} \\cdot 0 = 1, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 130, + 770, + 468, + 799 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "which means that all data in the input space are mapped onto the sphere of radius 1 in the feature space. Also, the distance between two input features is given by:", + "bbox": [ + 76, + 804, + 468, + 849 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\left\\| \\phi (\\mathbf {x} _ {i}) - \\phi (\\mathbf {x} _ {j}) \\right\\| ^ {2} = 1 + 1 - 2 \\cdot \\phi (\\mathbf {x} _ {i}) \\phi (\\mathbf {x} _ {j}) ^ {T} (7) \\\\ = 2 \\cdot \\left(1 - \\operatorname {S i m} \\left(\\mathbf {x} _ {i}, \\mathbf {x} _ {j}\\right)\\right) = \\frac {2}{\\pi} \\cdot \\theta \\left(\\mathbf {x} _ {i}, \\mathbf {x} _ {j}\\right). (7) \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 101, + 854, + 468, + 905 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/fae8d3265027027362ce0f59815be2c815433d2e751bb7501aaa5a537dd1f64b.jpg", + "image_caption": [ + "(a) Input space" + ], + "image_footnote": [], + "bbox": [ + 537, + 95, + 630, + 157 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/8dde8b37267758b9431238ed8673c7d607d39af714e6d08bb0c4afbaf7b1e2ca.jpg", + "image_caption": [ + "(b) Feature space", + "Figure 3. (a) A simple 2D input space and (b) the corresponding 3D feature space after applying the angular kernel to the 2D input data. Note that real input data/features are of higher dimensions." + ], + "image_footnote": [], + "bbox": [ + 663, + 87, + 851, + 160 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "That is, the square (Euclid) distance and the spectral angle is positively correlated and the distance range is $[0, 2]$ . As shown in Fig. 3, the angles in the input space are correlated to the feature distance after applying our angular kernel.", + "bbox": [ + 498, + 222, + 890, + 282 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Expansion of Angular Kernel. A natural following question is how to incorporate the above angular kernel for designing linear attention given its quadratic complexity w.r.t. the input token length. Recalling from trigonometric identities and the expansion of the arccos function into an infinite series, we reformulate the similarity function as:", + "bbox": [ + 498, + 282, + 890, + 373 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\operatorname {S i m} \\left(\\mathbf {Q} _ {i}, \\mathbf {K} _ {j}\\right) = 1 - \\frac {1}{\\pi} \\cdot \\left(\\frac {\\pi}{2} - \\arcsin \\left(\\mathbf {Q} _ {i} \\cdot \\mathbf {K} _ {j} ^ {T}\\right)\\right) \\\\ = \\underbrace {\\frac {1}{2} + \\frac {1}{\\pi} \\cdot \\left(\\mathbf {Q} _ {i} \\cdot \\mathbf {K} _ {j} ^ {T}\\right)} _ {\\text {L i n e a r - a n g u l a r t e r m s}} \\tag {8} \\\\ + \\underbrace {\\frac {1}{\\pi} \\cdot \\sum_ {k = 1} ^ {\\infty} \\frac {(2 k) !}{2 ^ {2 k} (k !) ^ {2} (2 k + 1)} \\left(\\mathbf {Q} _ {i} \\cdot \\mathbf {K} _ {j} ^ {T}\\right) ^ {2 k + 1}} _ {\\text {H i g h - o r d e r r e s i d u a l s w h e n k \\geq 1 (n o n - l i n e a r)}}, \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 514, + 380, + 890, + 525 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathbf{Q}_i\\cdot \\mathbf{K}_j^T$ denotes the normalized linear kernel function $\\langle \\mathbf{Q}_i,\\mathbf{K}_j\\rangle /\\| \\mathbf{Q}_i\\| \\cdot \\| \\mathbf{K}_j\\|$ , which is equivalent to the inner product if $\\mathbf{Q}_i$ and $\\mathbf{K}_j$ are unit vectors. We see that the first linear-angular terms can be directly used as the similarity measurement in linear attention, while the remaining higher-order terms of an infinite series introduce a much higher complexity. As such, we propose to adopt a relaxation to approximate it.", + "bbox": [ + 498, + 534, + 890, + 654 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Linear-Angular Attention. To construction our linear- angular attention, we leverage parametrized DNN modules to approximate the expectation of the high-order residuals, i.e., $\\mathbf{M}_{ij} = \\mathbb{E}_{\\mathbf{Q}_i\\sim p_{Q_i},\\mathbf{K}_j\\sim p_{K_j}}\\left(\\sum_{k = 1}^{\\infty}\\alpha_k(\\mathbf{Q}_i\\cdot \\mathbf{K}_j^T)^{2k + 1}\\right)$ where $\\alpha_{k}$ is the coefficient for the $k$ -th order term in Eq. (8). The key question is how to design such DNN modules. Kim et al. [31] conducted an analysis on average attention weights of ViT models, which show a strong inductive bias to attend to neighboring tokens. To capture this, we introduce a learnable depthwise convolution (DWConv) module which is applied on all the value tokens, so $\\mathbf{M}_{ij}$ by-design attends to nearby tokens, as illustrated in Fig. 4. Second, DWConv is limited by its receptive field [41] while the averaged attention map also has a small number of offdiagonal scores to connect to nonadjacent tokens [31]. To capture that, we adopt a sparse softmax-based attention as", + "bbox": [ + 496, + 655, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "14434", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/ec02e87fdc198761e8fc24ad61e5ff865315daa98184d7b07602f7e5ef3e52ea.jpg", + "image_caption": [ + "Figure 4. The visualization of our Castling-ViT with both linear-angular attention and auxiliary softmax-based self-attention." + ], + "image_footnote": [], + "bbox": [ + 78, + 88, + 468, + 349 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "described in Section 3.2.2. In practice, we found that such a sparse softmax attention tends to converge to all zeros after adding a simple threshold regularization. We provide more explanation and visualization in Sec. 3.2.2 and Sec. 4.5.", + "bbox": [ + 75, + 390, + 470, + 450 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this way, our attention module can be formulated as:", + "bbox": [ + 96, + 450, + 460, + 465 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathbf {H} = \\operatorname {S i m} (\\mathbf {Q}, \\mathbf {K}) \\cdot \\mathbf {V} \\approx \\frac {1}{2} \\cdot \\mathbf {V} + \\frac {1}{\\pi} \\cdot \\mathbf {Q} \\cdot \\left(\\mathbf {K} ^ {T} \\cdot \\mathbf {V}\\right) \\tag {9} \\\\ + \\mathbf {M} _ {D W} \\cdot \\mathbf {V} + \\mathbf {M} _ {S p a r s e A t t n}, \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 94, + 472, + 468, + 520 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $1 / \\pi \\cdot Q\\cdot (K^T\\cdot V)$ is the linear term with $\\mathcal{O}(N)$ complexity, $\\mathbf{M}_{DW}\\cdot V$ is the matrix form of DWConv, and $\\mathbf{M}_{SparseAttn}$ is the normalized sparse softmax attention. The overall complexity to compute Eq. (9) is linear to the input token length, where the MACs of DWConv is also negligible (e.g., $< 1\\%$ of the total MACs). Also, normalizations $(\\mathrm{x} / \\| \\mathrm{x}\\|_{2})$ are inserted to the $\\mathbf{Q} / \\mathbf{K}$ and sparse attention branches to help the similarity measurement following [5].", + "bbox": [ + 75, + 527, + 468, + 650 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2.2 Switch Towards Linear-Angular Attention", + "text_level": 1, + "bbox": [ + 76, + 654, + 426, + 667 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Recall that we add a sparse softmax-based attention as an auxiliary branch in the linear-angular attention to help approximate the high-order residuals. Such a costly attention can be potentially dropped without hurting the inference accuracy, drawing inspiration from [67] that the remaining network can gradually learn the high-order/frequency components at the later training stage. Next, we explain how it is constructed and dropped.", + "bbox": [ + 75, + 674, + 468, + 794 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Sparse Training Augmentation. As illustrated in Fig. 4, we adopt a masked softmax-based attention as an auxiliary branch to augment ViT training. In particular, we first use a comparator with a predefined threshold $\\epsilon$ to generate a binary mask, where attentions greater than $\\epsilon$ are set to 1 and 0 otherwise. These masks are then applied to the attention maps to generate masked attention maps that will be", + "bbox": [ + 75, + 795, + 468, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "summed up with our linear-angular attention together as the final attention. The resulting sparse attention is given by:", + "bbox": [ + 498, + 90, + 890, + 121 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {M} _ {\\text {S p a r s e A t t n}} (\\mathbf {Q}, \\mathbf {K}) = \\operatorname {M a s k} _ {\\varepsilon} \\left(\\operatorname {S o f t m a x} \\left(\\mathbf {Q} \\cdot \\mathbf {K} ^ {\\mathrm {T}}\\right)\\right). \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 511, + 128, + 890, + 147 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathsf{Mask}_{\\epsilon}(\\mathbf{x}) = \\mathbf{x}$ if $\\mathbf{x} > \\epsilon$ else 0, acting as an elementwise threshold function. As such, the sparse attention captures the higher attention scores (i.e., strong local features), and can potentially complement our linear-angular attention by supplementing the missing higher-order terms. Such an assumption aligns well with recent findings that (1) low-rank and sparse approximations complement each other [8, 17, 31]; and (2) linear attention lacks local feature extraction capabilities over its softmax-based counterpart [6].", + "bbox": [ + 496, + 157, + 890, + 290 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Castling During ViT Inference. As we are targeting efficient ViT inference, it is desired to reduce or completely remove the costly softmax-based attentions while only keeping our linear-angular attention at runtime, i.e., performing castling. In our experiments, we found that under the sparsity regularization above, the softmax attention naturally converges to all zeros as the training progresses. We consider both fixed and dynamic schedules for $\\epsilon$ and find that our Castling-ViT is not sensitive to neither the threshold value nor the threshold schedule for a given task. Given a fixed mask threshold (e.g., $\\epsilon = 0.02$ in image classification experiments), the masks become all zeros at latter training stages and thus the auxiliary branch can be removed without hurting the model accuracy. We supply the visualization of the mask-evolving trajectory and our conjecture for understanding such a phenomenon in Sec. 4.5.", + "bbox": [ + 496, + 291, + 892, + 532 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 498, + 541, + 633, + 558 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Experiment Settings", + "text_level": 1, + "bbox": [ + 500, + 561, + 692, + 579 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Tasks, Datasets, and Models. Tasks and Datasets. We consider three benchmark datasets and three representative vision tasks to demonstrate the superiority of the proposed Castling-ViT, including image classification on ImageNet dataset [18] with 1.2 million training and 50K validation images; Object detection on COCO dataset [36] with 118K training and 5K validation images; Semantic segmentation on ADE20K dataset [75] with 20K/2K/3K images for training, validation, and testing, respectively. Models. We apply our proposed Castling-ViT idea on top of various models. For the classification task, we consider LeViT [24], MViTv2 [34], and DeiT [54]; For the detection task, we consider models with efficient ViT backbones (e.g., PicoDet [70] with modified ESNet and LCNet backbones [15] with transformer blocks); For the segmentation task, we consider Mask2former [11] with ViT-Base backbone.", + "bbox": [ + 496, + 580, + 890, + 823 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Training Settings. For the classification task, we use a SGD optimizer with 0.9 momentum and $2 \\times 10^{-5}$ weight decay to train ViTs for 1000 epochs using 64 V100 GPUs, with each card having 64 (LeViT) or 32 (MViTv2/DeiT) batch sizes. The learning rate is 2.0", + "bbox": [ + 496, + 823, + 892, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "14435", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/6cf2c60dafa4275a4fb2ccd33d5b0b932d477d6df0693ca0c09fcc6307ac5e9b.jpg", + "table_caption": [ + "Table 2. Castling-ViT over SOTA baselines on ImageNet." + ], + "table_footnote": [], + "table_body": "
MACs RangesModelsParams (M)MACs (G)Top-1 Acc.(%)Top-5 Acc.(%)
<1GMobileNetV2-1.4 [47]6.90.5874.7-
EfficientNet-B0 [51]5.30.3977.193.3
EfficientNet-B1 [51]7.80.7079.194.4
MobileViT-XS [40]2.30.7074.892.3
LeViT-128 [24]9.20.4178.6-
LeViT-192 [24]10.90.6680.0-
Castling-LeViT-12810.50.4979.694.6
Castling-LeViT-19212.70.8281.395.5
1~3GEfficientNet-B3 [51]12.01.8081.695.7
HRFormer-T [73]8.01.8078.5-
DeiT-T [54]5.61.2574.591.9
LeViT-256 [24]18.91.1281.6-
LeViT-384 [24]39.12.3582.6-
Caslting-DeiT-T5.641.1876.092.5
Castling-LeViT-25622.01.4082.696.1
Castling-LeViT-38445.82.9083.796.7
3~10GRegNetY-4G [44]21.04.0080.0-
RegNetY-8G [44]39.08.0081.7-
EfficientNet-B4 [51]19.04.2082.996.4
EfficientNet-B5 [51]30.09.9083.696.7
Swin-T [38]29.04.5081.395.5
Swin-S [38]50.08.7083.296.2
CSWin-T [19]23.04.3082.7-
CSWin-S [19]35.06.9083.6-
DeiT-S [54]21.94.6081.295.4
AutoFormer-S [10]22.95.1081.795.7
PVTv2-V2 [61]25.04.0082.0-
MViTv2-T [34]24.04.6882.3-
MViTv2-S [34]34.76.9583.6-
Castling-MViTv2-T24.14.5084.196.8
Castling-MViTv2-S34.76.9584.697.0
>10GCaiT-S36 [55]68.013.9083.3-
Swin-B [38]88.015.4083.596.5
CSWin-B [19]78.015.0084.2-
AutoFormer-B [10]54.011.0082.495.7
MViTv2-B [34]51.210.0784.4-
DeiT-B [54]86.317.5683.496.5
Castling-DeiT-B87.2217.2884.2-
Castling-MViTv2-B51.99.8285.097.2
", + "bbox": [ + 81, + 104, + 465, + 555 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "with first 11 epochs warm-up starting from 0.01 and decays by a factor of 0.9875 per epoch [62]. Also, we use the distillation [54] based on a teacher model with a $85.5\\%$ accuracy. For the detection task, we adopt SGD optimizer with momentum 0.9 and weight decay 4e-5 to train models on COCO. All models are trained on 8 V100 GPUs with each card having 80 batch sizes following PicoDet's training recipe [70]. Also, we follow LeViT's training recipe [24] to pretrain backbones on ImageNet; For the segmentation task, we follow Mask2former's training recipe [11] to train models on ADE20K, where ViT backbones are pretrained following MAE [25] if specified.", + "bbox": [ + 75, + 566, + 467, + 747 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Baselines and Evaluation Metrics. Baselines. For the classification task, we compare the proposed Castling-ViT with LeViT [24], MviTv2 [34], DeiT [54], Swin [38], CSWin [19], PVT [59], etc. For the detection task, we compare with FBNetV5 [62], YOLOX [23], YOLOv5, MobileDet [65], and EfficientDet [52]. For the segmentation task, we compare with Mask2former [11] with ViT backbones. Evaluation Metrics. We evaluate the Castling-ViT and all baselines in terms of accuracy-efficiency tradeoffs. Specifically, the accuracy metrics refer to top-1/5 accuracy", + "bbox": [ + 75, + 750, + 467, + 901 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/f80cbe153f3a2fa53c3b8868014ff6ed91dafbe483535704e0366c50d03a13db.jpg", + "image_caption": [ + "Figure 5. Castling vs. baseline ViT on ImageNet." + ], + "image_footnote": [], + "bbox": [ + 506, + 92, + 694, + 188 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/f151dba8746532bcfb4368a3d9945e3c144e9e15828109e90f06fe66e7b983be.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 702, + 93, + 885, + 189 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "for the classification task; AP, $\\mathrm{AP}^{50}$ , $\\mathrm{AP}^{75}$ for the detection task (AP: average precision); mIoU, mAcc, and pAcc for the segmentation task (mIoU: mean intersection over union). For efficiency metrics, we compare the number of model parameters or inference FLOPs (or MACs).", + "bbox": [ + 496, + 219, + 890, + 295 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Castling-ViT over SOTA Baslines", + "text_level": 1, + "bbox": [ + 500, + 305, + 792, + 320 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Image Classification. To evaluate the effectiveness of our proposed techniques on the image classification task, we apply the proposed Castling-ViT idea to three typical or SOTA ViT architectures: DeiT [54] as typical ViTs, LeViT [24] as representative efficient ViTs, and MViTv2 [34] as representative hierarchical ViTs for downstream applications of high input resolutions, and compare their performance over baselines on ImageNet. As shown in Tab. 2, the comparison across a large MACs (or FLOPs) range starting from 0.4G to 17G. We categorize ViT models into four regimes: $<1\\mathrm{G}$ , $1 \\sim 3\\mathrm{G}$ , $3 \\sim 10\\mathrm{G}$ , and $>10\\mathrm{G}$ , and select baselines in each regime to benchmark separately for the clarity purpose. Castling-ViTs consistently perform better than all baselines across various MACs ranges in terms of the accuracy-efficiency tradeoff. For example, Castling-LeViT achieves $82.6\\%$ top-1 accuracy with only 1.40G MACs while LeViT requires 2.35G FLOPs instead, i.e., $\\downarrow 40\\%$ MACs; On the other hand, under comparable MACs, Castling-MViTv2 achieves $84.1\\%$ accuracy vs. MViTv2 with $82.3\\%$ accuracy instead, i.e., $\\uparrow 1.8\\%$ top-1 accuracy. Overall, Castling-ViT achieves an improved accuracy of $\\uparrow 0.5\\% \\sim \\uparrow 6.6\\%$ , $\\uparrow 1.0\\% \\sim \\uparrow 8.1\\%$ , $\\uparrow 1.0\\% \\sim \\uparrow 4.1\\%$ , and $\\uparrow 0.6\\% \\sim \\uparrow 2.6\\%$ over baselines under $<1\\mathrm{G}$ MACs, $1 \\sim 3\\mathrm{G}$ MACs, $3 \\sim 10\\mathrm{G}$ MACs, and $>10\\mathrm{G}$ MACs, respectively. Note that we calculate improvements under comparable MACs. In addition to the overall comparison, we also visualize the apple-to/apple benchmark, e.g., Castling-LeViT vs. LeViT, to validate the effectiveness of proposed techniques. As shown in Fig. 5, Castling-LeViT/MViTv2 achieves $\\downarrow 25.7\\% \\sim \\downarrow 55.3\\%$ MACs reductions under comparable accuracies or offers a comparable or better accuracy ( $\\uparrow 0.6\\% \\sim \\uparrow 1.8\\%$ ) under comparable MACs over corresponding LeViT/MViTv2 baselines.", + "bbox": [ + 496, + 325, + 890, + 823 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Object Detection. We also extend the Castling-ViT to the downstream object detection task and compare it with previous efficient detectors on COCO dataset to evaluate its efficacy. Specifically, we construct the detector with modified ESNet [70] or LCNet [15] (replace the last one or two", + "bbox": [ + 496, + 824, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "14436", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/4fdcabe63d76940ca8a0011891ca230c105fec9dd352254758919fed4f7dbc4b.jpg", + "table_caption": [ + "Table 3. Castling-ViT over SOTA baselines on COCO detection, where Castling-ViT† means that the detection model adopts modified ESNet (use transformer blocks in the last stage; referred to ESNet-ViT) as backbone while Castling-ViT‡ uses modified LCNet (use transformer blocks in the last two stage; referred to LCNet-ViT) as backbone. Also, “-S/M/L” and “-320/416” represent small/medium/large model variants and 320x320/416x416 input resolutions, respectively." + ], + "table_footnote": [], + "table_body": "
ModelsParams (M)MACs (G)mAPAP50AP75
MobileDet-DSP [65]4.851.4327.3--
MobileDet-DSP-Fused [65]9.153.2229.1--
EfficientDet-D0 [52]3.902.5434.653.037.1
YOLOv5-N1.902.3028.045.7-
YOLOX-Nano [23]0.910.5425.8--
YOLOX-Tiny [23]5.063.2332.8--
FBNetV5-AC-224x320 [62]-0.7125.0--
FBNetV5-AR-224x320 [62]-0.9126.4--
FBNetV5-A-224x320 [62]-1.3527.2--
FBNetV5-AC-320x640 [62]-1.3728.9--
FBNetV5-AR-320x640 [62]-1.8030.4--
Castling-ViT-S-320†3.250.6228.142.329.2
Castling-ViT-S-416†3.251.0631.346.732.5
Castling-ViT-M-416†6.012.0034.049.535.9
Castling-ViT-L-416†9.323.0335.050.737.2
Castling-ViT-L-416*13.105.3137.353.439.6
", + "bbox": [ + 80, + 202, + 468, + 412 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "stages with transformer blocks) as backbones and follow PicoDet [70]'s detection head design as well as their training recipe. Fig. 1 and Tab. 3 show the overall comparison between the proposed Castling-ViT and other baselines. We can see that our Castling-ViT consistently achieves better accuracy-efficiency tradeoffs, leading to $\\uparrow 6.0$ , $\\uparrow 2.2 \\sim \\uparrow 2.3$ , $\\uparrow 4.0 \\sim \\uparrow 5.9$ , $\\uparrow 3.1 \\sim \\uparrow 4.0$ mAP improvements as compared to YOLOv5, YOLOX [23], MobileDet [65], and FBNetv5 [62], respectively, under comparable or even less MACs. As for EfficientDet [52], our method achieves comparable accuracy-efficiency trade-offs. Apart from the overall comparison with baselines, we also provide the apple-to-Apple comparison with detectors with softmax-based attention, we supply the comparison results and Castling-ViT's breakdown analysis to Sec. 4.4. This set of experiments validate the effectiveness of the proposed Castling-ViT for servering as efficient detector backbones (e.g., $\\leq 3G$ MACs) in the object detection task.", + "bbox": [ + 75, + 428, + 468, + 699 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Semantic Segmentation. We further extend Castling-ViT to the semantic segmentation task to evaluate its effectiveness. Specifically, we use ViT-Base as the backbone in Mask2former [11] framework to serve as our baseline and testbed. Then we build the Castling-ViT-Base as the backbone to benchmark on the ADE20K dataset. As shown in Tab. 4, Mask2former with Castling-ViT backbone achieves $15\\%$ total MACs reductions and $19\\%$ backbone MACs reduction under comparable or slightly better mIoU, i.e., $\\uparrow 0.13\\%$ and $\\uparrow 0.52\\%$ without or with MAE pretraining on ImageNet [25]. This set of experiments validate that our proposed techniques could be well generalized to various downstream tasks that require large input resolutions.", + "bbox": [ + 75, + 704, + 468, + 902 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/1124711e09ba1fa5f4dc6e83156c8fad29a685452d8968e146223caf8429f68c.jpg", + "table_caption": [ + "Table 4. Castling-ViT over baselines on the ADE20K segmentation dataset, serving as backbones in the Mask2former framework. Note that we report both total MACs and backbone MACs, i.e., $(\\cdot)$ ." + ], + "table_footnote": [], + "table_body": "
Mask2former w/MAE \nPretrainParams \n(M)MACs \n(G)mIoUmAccpAcc
ViT-BaseN118229 (182)34.5446.3675.84
Castling-ViT-BaseN118195 (147)34.6746.4776.20
ViT-BaseY118229 (182)47.9261.0083.02
Castling-ViT-BaseY118195 (147)48.4461.8283.29
", + "bbox": [ + 503, + 132, + 890, + 200 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/50e242aafcb40b2bf76c639f90f87a71c0874514f720cd2f46615d5f6ed89a10.jpg", + "table_caption": [ + "Table 5. Ablation studies of Castling-ViT with various kernel-based linear attentions, where MC refers to mean-centering kernel function applied to $\\mathbf{Q} / \\mathbf{K}$ ; Softmax also means the kernel function; ReLU-S and ReLU-E denote applying ReLU as kernels for decomposing softmax or exponential terms, respectively. All models are trained on COCO from scratch without pretraining." + ], + "table_footnote": [], + "table_body": "
BackboneLinear AttnS-320S-416M-416
mAPAP50mAPAP50mAPAP50
LCNetN/A25.539.029.744.234.349.9
LCNet-ViTN/A25.939.629.844.434.950.4
LCNet-ViTMC [17]22.834.926.339.631.546.1
Softmax [49]26.039.629.243.434.449.8
Cosine [5]24.437.427.241.032.948.0
ReLU-S23.936.7--30.945.3
ReLU-E [6]26.540.230.244.734.349.7
Lin.-Angular27.040.630.445.535.551.1
", + "bbox": [ + 504, + 294, + 890, + 407 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Linear-Angular Attention over SOTA Baselines", + "text_level": 1, + "bbox": [ + 498, + 424, + 890, + 441 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We also conduct ablation studies among various kinds of kernels used in linear attention to evaluate the superiority of our proposed linear-angular kernel. Also shown in Tab. 5, we compare the angular kernel with five other commonly adopted kernels, results on three model and resolution settings consistently demonstrate that our proposed angular kernel helps achieve better mAP, e.g., $\\uparrow 4.0\\% \\sim \\uparrow 4.2\\%$ over MC [17], $\\uparrow 1.0\\% \\sim \\uparrow 1.2\\%$ over Softmax [49], $\\uparrow 2.6\\% \\sim \\uparrow 3.2\\%$ over Cosine [5], $\\uparrow 3.1\\% \\sim \\uparrow 4.6\\%$ and $\\uparrow 0.2\\% \\sim \\uparrow 1.2\\%$ over ReLU-S and ReLU-E [6], respectively. This set of experiments validate the superiority of the proposed linear-angular kernels.", + "bbox": [ + 496, + 446, + 890, + 627 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Ablation Studies of Castling-ViT", + "text_level": 1, + "bbox": [ + 498, + 636, + 785, + 652 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We conduct ablation studies on Castling-ViT's linear- angular attention, added DWConv, as well as auxiliary masked softmax-based attention, as shown in Tab. 6 and Tab. 7, where the experiments are performed on CastlingViT-S-320\\*/M-416\\*, respectively, without pretraining on ImageNet. Note that here \\* means that we use LCNet-ViT, i.e., replacing all layers in last two stages with transformer blocks, as backbones. Results in three tables show detailed performance breakdown and consistently demonstrate that all components in our proposed Castling-ViT contribute to the final performance. Specifically, linear-angular attention itself already achieves $\\downarrow 0.1\\% \\sim \\uparrow 0.5\\%$ mAP improvements while leading to $10.1\\% \\sim 15.3\\%$ MACs reductions simultaneously. Adding DWConv in the middle of MLP layers further leads to $\\downarrow 0.1\\% \\sim \\uparrow 0.6\\%$ mAP improvements while only incur negligible MACs overhead. Adding aux", + "bbox": [ + 496, + 657, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "14437", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/0e8692f7a7ea0d9f713cceddc428971960983d06b90946db75fc03db6dcc821c.jpg", + "table_caption": [ + "Table 6. Ablation studies of Castling-ViT-S-320 w/ LCNet backbone, or LCNet-ViT backbone + linear-angular attention (denoted as Lin.) + DWConv (denoted as DW.) + auxiliary softmax-based attention (denoted as SparseAttn.)." + ], + "table_footnote": [], + "table_body": "
Castling-ViT-S-320 w/Params (M)MACs (G)mAPAP50AP75
LCNet1.100.4825.539.026.5
LCNet-ViT2.140.6925.939.626.9
+ Lin.2.140.6226.239.727.3
+ Lin.+DW.2.140.6226.840.728.0
+ Lin.+DW.+SparseAttn.2.140.6227.040.628.0
", + "bbox": [ + 80, + 146, + 468, + 234 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/c591186ebac976780e92fa48d5640ec668c8028435f16f989f746ad267fd87cc.jpg", + "table_caption": [ + "Table 7. Ablation studies of Castling-ViT-M-416 w/ LCNet backbone, or LCNet-ViT backbone + Lin. + DW. + SparseAttn." + ], + "table_footnote": [], + "table_body": "
Castling-ViT-M-416 w/Params (M)MACs (G)mAPAP50AP75
LCNet3.322.1534.349.936.5
LCNet-ViT7.533.5134.950.437.0
+ Lin.7.533.1534.850.336.7
+ Lin.+DW.7.533.1535.150.936.9
+ Lin.+DW.+SparseAttn.7.533.1535.351.137.3
", + "bbox": [ + 80, + 268, + 468, + 357 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "illary masked softmax-based attention further increases the mAP by $0.2\\%$ without incurring any overhead since it only assists the training while being removed during inference. Note that although LCNet-ViT backbone leads to better mAP, it also introduces more MACs as compared to efficient convolutions. As such, LCNet-ViT still slightly underperforms LCNet backbones in terms of accuracy-efficiency trade-offs, since we directly replace the CNN backbones with transformer blocks, whose architecture could not be optimal for ViTs. It remains to an open problem to build ViT based models that achieves higher accuracy-efficiency trade-off than pure ConvNet models.", + "bbox": [ + 75, + 372, + 468, + 553 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.5. Discussion on the Auxiliary Branch", + "text_level": 1, + "bbox": [ + 76, + 561, + 383, + 578 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Trajectory of Nonzeros in Masks. To further understand the effect of the auxiliary branch, we visualize the trajectories of nonzeros in masks of auxiliary softmax-based attention. As shown in Fig. 6, we count the nonzeros in the first attention layer's masks throughout the training of both Castling-ViT-S-416", + "bbox": [ + 75, + 584, + 254, + 777 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "and Castling-ViT-M-416 with the LCNet-ViT backbone on COCO. we observe that the introduced auxiliary attention will only assist the training in the early or middle training stages and will gradually vanish towards all zeros in the later training stage, which is well aligned with the assumption in our Castling-ViT. This set of experiments validate the idea of performing \"castling\" in ViTs, i.e., drop the auxiliary branch without sacrificing performance.", + "bbox": [ + 75, + 779, + 468, + 901 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/5279c2953fa8e8015eb33abc95d73b9575265d7ccf39c40c1efc3d7cca79218c.jpg", + "image_caption": [ + "Figure 6. Visualizing the trajectories of nonzeros in auxiliary masks during training." + ], + "image_footnote": [], + "bbox": [ + 295, + 588, + 459, + 724 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/d75300cba95932f023a0bd0c5f4439fede7dee84c03a47788d5601f602833679.jpg", + "image_caption": [ + "Figure 7. Left: Visualizing the taget similarity function, linear-angular terms, and DNNs' fitting curve; Middle: Visualizing the loss trajectory during DNN training; Right: Visualizing the learning trajectory of both low and high frequency components." + ], + "image_footnote": [], + "bbox": [ + 504, + 89, + 629, + 167 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/80a8e7ce0330b01dacacd833646e1813cd7573a0f392c088e9016f4ee2192c35.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 629, + 90, + 751, + 169 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/06ba51f9e7a7ddd23a070f84e831f634c43e5996188046b58fa7e971ca0d3996.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 758, + 88, + 890, + 169 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Conjecture of Castling. We conduct a synthetic experiment to give an analogy for explaining the castling phenomenon. In Fig. 7 (Left), we visualize the curve of (1) angular similarity function (denoted in Red), (2) linear-angular terms (denoted in Green), and (3) two-layer DNN's approximation. We see that only keeping linear-angular terms leads to distortion, while DNN is capable of learning the missing high-frequency parts, whose loss trajectory is shown in Fig. 7 (Middle). Also, the learning trajectory of frequency components is visualized in Fig. 7 (Right), following F-Principle [67], where x-axis is training steps, y-axis categorizes both low and high frequency components, blue/red colors refer to large/small difference between the learned frequency components and target frequency components, respectively. We observe that DNNs fit target functions from low to high frequencies during training. It indicates that DWConv itself is not sufficient for approximating the high-order residual at early training stages. Costly attention that contains high-order residuals is then desired to help training while being dropped at inference since the remaining networks can gradually learn the high-frequency components at later training stages [67].", + "bbox": [ + 496, + 234, + 890, + 566 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 575, + 617, + 590 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We present Castling-ViT that trains ViTs with both linear-angular and softmax-based quadratic attention but switches to only having the former during inference. Castling-ViT leverages angular kernels to measure the similarities between the queries and keys via spectral angles and highlights two enablers: (1) a new linear-angular attention mechanism: we decompose angular kernels to linear terms and high-order residuals, and keep only the former for inference; and (2) we approximate the high-order residuals using a depthwise convolution and an auxiliary masked softmax attention whose masks gradually become zeros during training without incurring inference overhead. Extensive experiments consistently validate Castling-ViT's advantages.", + "bbox": [ + 496, + 597, + 890, + 792 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgment", + "text_level": 1, + "bbox": [ + 500, + 801, + 651, + 819 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The work is supported in part by the National Science Foundation (NSF) RTML program (Award number: 1937592) and the CoCoSys, one of the seven centers in JUMP 2.0, a Semiconductor Research Corporation (SRC) program sponsored by DARPA.", + "bbox": [ + 496, + 825, + 890, + 901 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "14438", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Alaaeldin Ali, Hugo Touvron, Mathilde Caron, Piotr Bojanowski, Matthijs Douze, Armand Joulin, Ivan Laptev, Natalia Neverova, Gabriel Synnaeve, Jakob Verbeek, et al. Xcit: Cross-covariance image transformers. Advances in neural information processing systems, 34:20014-20027, 2021. 2, 13", + "[2] Moab Arar, Ariel Shamir, and Amit H Bermano. Learned queries for efficient local attention. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10841-10852, 2022. 1, 2, 13", + "[3] Iz Beltagy, Matthew E Peters, and Arman Cohan. Longformer: The long-document transformer. arXiv preprint arXiv:2004.05150, 2020. 1, 13", + "[4] Daniel Bolya, Cheng-Yang Fu, Xiaoliang Dai, Peizhao Zhang, Christoph Feichtenhofer, and Judy Hoffman. Token merging: Your vit but faster. arXiv preprint arXiv:2210.09461, 2022. 2", + "[5] Daniel Bolya, Cheng-Yang Fu, Xiaoliang Dai, Peizhao Zhang, and Judy Hoffman. Hydra attention: Efficient attention with many heads. arXiv preprint arXiv:2209.07484, 2022. 1, 2, 5, 7, 13", + "[6] Han Cai, Chuang Gan, and Song Han. Efficientvit: Enhanced linear attention for high-resolution low-computation visual recognition. arXiv preprint arXiv:2205.14756, 2022. 1, 2, 3, 5, 7, 13, 14", + "[7] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In European conference on computer vision, pages 213-229. Springer, 2020. 13", + "[8] Beidi Chen, Tri Dao, Eric Winsor, Zhao Song, Atri Rudra, and Christopher Ré. Scatterbrain: Unifying sparse and low-rank attention. Advances in Neural Information Processing Systems, 34:17413-17426, 2021. 1, 2, 5, 13", + "[9] Chun-Fu Chen, Quanfu Fan, and Rameswar Panda. Crossvit: Cross-attention multi-scale vision transformer for image classification. arXiv preprint arXiv:2103.14899, 2021. 2", + "[10] Minghao Chen, Houwen Peng, Jianlong Fu, and Haibin Ling. Autoformer: Searching transformers for visual recognition. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12270-12280, 2021. 6", + "[11] Bowen Cheng, Ishan Misra, Alexander G Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1290–1299, 2022. 3, 5, 6, 7, 13", + "[12] Bowen Cheng, Alex Schwing, and Alexander Kirillov. Per-pixel classification is not all you need for semantic segmentation. Advances in Neural Information Processing Systems, 34:17864-17875, 2021. 13", + "[13] Krzysztof Choromanski, Valerii Likhosherstov, David Dohan, Xingyou Song, Andreea Gane, Tamas Sarlos, Peter Hawkins, Jared Davis, Afroz Mohiuddin, Lukasz Kaiser," + ], + "bbox": [ + 78, + 114, + 470, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "et al. Rethinking attention with performers. arXiv preprint arXiv:2009.14794, 2020.3", + "[14] Krzysztof Marcin Choromanski, Valerii Likhosherstov, David Dohan, Xingyou Song, Andreea Gane, Tamas Sarlos, Peter Hawkins, Jared Quincy Davis, Afroz Mohiuddin, Lukasz Kaiser, David Benjamin Belanger, Lucy J Colwell, and Adrian Weller. Rethinking attention with performers. In International Conference on Learning Representations, 2021. 1, 2, 13", + "[15] Cheng Cui, Tingquan Gao, Shengyu Wei, Yuning Du, Ruoyu Guo, Shuilong Dong, Bin Lu, Ying Zhou, Xueying Lv, Qiwen Liu, et al. Pp-lcnet: A lightweight cpu convolutional neural network. arXiv preprint arXiv:2109.15099, 2021. 5, 6", + "[16] Xiaoliang Dai, Alvin Wan, P. Zhang, B. Wu, Zijian He, Zhen Wei, K. Chen, Yuandong Tian, Matthew E. Yu, Péter Vajda, and J. Gonzalez. Fbnetv3: Joint architecture-recipe search using neural acquisition function. ArXiv, abs/2006.02049, 2020.2", + "[17] Jyotikrishna Dass, Shang Wu, Huihong Shi, Chaojian Li, Zhifan Ye, Zhongfeng Wang, and Yingyan Lin. Vitality: Unifying low-rank and sparse approximation for vision transformer acceleration with a linear taylor attention. In The 29th IEEE International Symposium on High-Performance Computer Architecture (HPCA 2023), 2023. 2, 5, 7, 14", + "[18] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 1, 5", + "[19] Xiaoyi Dong, Jianmin Bao, Dongdong Chen, Weiming Zhang, Nenghai Yu, Lu Yuan, Dong Chen, and Baining Guo. Cswin transformer: A general vision transformer backbone with cross-shaped windows. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12124-12134, 2022. 6", + "[20] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations, 2021. 1, 2, 3", + "[21] Haoqi Fan, Bo Xiong, Karttikeya Mangalam, Yanghao Li, Zhicheng Yan, Jitendra Malik, and Christoph Feichtenhofer. Multiscale vision transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6824-6835, 2021. 2", + "[22] Peng Gao, Teli Ma, Hongsheng Li, Jifeng Dai, and Yu Qiao. Convmae: Masked convolution meets masked autoencoders. arXiv preprint arXiv:2205.03892, 2022. 13", + "[23] Zheng Ge, Songtao Liu, Feng Wang, Zeming Li, and Jian Sun. Yolox: Exceeding yolo series in 2021. arXiv preprint arXiv:2107.08430, 2021. 6, 7, 14", + "[24] Ben Graham, Alaaeldin El-Nouby, Hugo Touvron, Pierre Stock, Armand Joulin, Hervé Jégou, and Matthijs Douze. Levit: a vision transformer in convnet's clothing for faster inference. arXiv preprint arXiv:2104.01136, 2021. 2, 3, 5, 6, 13, 14" + ], + "bbox": [ + 503, + 92, + 890, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "14439", + "bbox": [ + 480, + 945, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[25] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollár, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16000-16009, 2022. 6, 7, 13", + "[26] Byeongho Heo, Sangdoo Yun, Dongyoon Han, Sanghyuk Chun, Junsuk Choe, and Seong Joon Oh. Rethinking spatial dimensions of vision transformers. arXiv preprint arXiv:2103.16302, 2021. 2", + "[27] Paul Honeine and Cedric Richard. The angular kernel in machine learning for hyperspectral data classification. In 2010 2nd Workshop on Hyperspectral Image and Signal Processing: Evolution in Remote Sensing, pages 1-4. IEEE, 2010. 4", + "[28] Andrew Howard, Mark Sandler, Grace Chu, Liang-Chieh Chen, Bo Chen, Mingxing Tan, Weijun Wang, Yukun Zhu, Ruoming Pang, Vijay Vasudevan, et al. Searching for mobilenetv3. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1314-1324, 2019. 2", + "[29] Angelos Katharopoulos, Apoorv Vyas, Nikolaos Pappas, and François Fleuret. Transformers are rnns: Fast autoregressive transformers with linear attention. In International Conference on Machine Learning, pages 5156-5165. PMLR, 2020. 1, 2, 3, 4, 13", + "[30] Nirmal Keshava. Distance metrics and band selection in hyperspectral processing with applications to material identification and spectral libraries. IEEE Transactions on Geoscience and remote sensing, 42(7):1552-1565, 2004. 4", + "[31] Kyungmin Kim, Bichen Wu, Xiaoliang Dai, Peizhao Zhang, Zhicheng Yan, Peter Vajda, and Seon Joo Kim. Rethinking the self-attention in vision transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3071–3075, 2021. 4, 5", + "[32] Nikita Kitaev, Lukasz Kaiser, and Anselm Levskaya. Reformer: The efficient transformer. arXiv preprint arXiv:2001.04451, 2020. 1, 13", + "[33] Yanghao Li, Hanzi Mao, Ross Girshick, and Kaiming He. Exploring plain vision transformer backbones for object detection. arXiv preprint arXiv:2203.16527, 2022. 13", + "[34] Yanghao Li, Chao-Yuan Wu, Haoqi Fan, Karttikeya Mangalam, Bo Xiong, Jitendra Malik, and Christoph Feichtenhofer. Mviv2: Improved multiscale vision transformers for classification and detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4804-4814, 2022. 3, 5, 6, 13, 14", + "[35] Yanyu Li, Geng Yuan, Yang Wen, Eric Hu, Georgios Evangelidis, Sergey Tulyakov, Yanzhi Wang, and Jian Ren. Efficientformer: Vision transformers at mobilenet speed. arXiv preprint arXiv:2206.01191, 2022. 2", + "[36] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In European conference on computer vision, pages 740-755. Springer, 2014. 1, 5", + "[37] Jing Liu, Zizheng Pan, Haoyu He, Jianfei Cai, and Bohan Zhuang. Ecoformer: Energy-saving attention with linear complexity. In NeurIPS, 2022. 1, 2, 4, 13" + ], + "bbox": [ + 78, + 90, + 470, + 901 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[38] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10012-10022, 2021. 1, 2, 6, 13, 14", + "[39] Jiachen Lu, Jinghan Yao, Junge Zhang, Xiatian Zhu, Hang Xu, Weiguo Gao, Chunjing Xu, Tao Xiang, and Li Zhang. Soft: softmax-free transformer with linear complexity. Advances in Neural Information Processing Systems, 34:21297-21309, 2021. 1, 2, 13", + "[40] Sachin Mehta and Mohammad Rastegari. Mobilevit: lightweight, general-purpose, and mobile-friendly vision transformer. arXiv preprint arXiv:2110.02178, 2021. 2, 6", + "[41] John Moody and Christian Darken. _Learning with localized receptive fields_. Yale Univ., Department of Computer Science, 1988. 4", + "[42]Zipeng Qin, Jianbo Liu, Xiaolin Zhang, Maoqing Tian, Aojun Zhou, Shuai Yi, and Hongsheng Li. Pyramid fusion transformer for semantic segmentation. arXiv preprint arXiv:2201.04019, 2022.3, 13", + "[43] Zhen Qin, Weixuan Sun, Hui Deng, Dongxu Li, Yunshen Wei, Baohong Lv, Junjie Yan, Lingpeng Kong, and Yiran Zhong. cosformer: Rethinking softmax in attention. In International Conference on Learning Representations, 2022. 1, 2, 13", + "[44] Ilija Radosavovic, Raj Prateek Kosaraju, Ross Girshick, Kaiming He, and Piotr Dólár. Designing network design spaces. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10428-10436, 2020. 6", + "[45] Ali Rahimi and Benjamin Recht. Random features for large-scale kernel machines. Advances in neural information processing systems, 20, 2007. 3", + "[46] Yongming Rao, Wenliang Zhao, Benlin Liu, Jiwen Lu, Jie Zhou, and Cho-Jui Hsieh. Dynamicvit: Efficient vision transformers with dynamic token sparsification. Advances in neural information processing systems, 34:13937-13949, 2021. 2", + "[47] Mark Sandler, Andrew Howard, Menglong Zhu, Andrey Zh-moginov, and Liang-Chieh Chen. Mobilenetv2: Inverted residuals and linear bottlenecks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4510-4520, 2018. 6", + "[48] Robert R. Schaller. Moore's law: past, present and future. IEEE spectrum, 1997. 14", + "[49] Zhuoran Shen, Mingyuan Zhang, Haiyu Zhao, Shuai Yi, and Hongsheng Li. Efficient attention: Attention with linear complexities. In Proceedings of the IEEE/CVF winter conference on applications of computer vision, pages 3531-3539, 2021. 3, 7", + "[50] Chen Sun, Abhinav Shrivastava, Saurabh Singh, and Abhinav Gupta. Revisiting unreasonable effectiveness of data in deep learning era. In ICCV, 2017. 2", + "[51] Mingxing Tan and Quoc Le. Efficientnet: Rethinking model scaling for convolutional neural networks. In International conference on machine learning, pages 6105-6114. PMLR, 2019. 6" + ], + "bbox": [ + 501, + 92, + 890, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "14440", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[52] Mingxing Tan, Ruoming Pang, and Quoc V Le. Efficientdet: Scalable and efficient object detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10781-10790, 2020. 6, 7", + "[53] Shitao Tang, Jiahui Zhang, Siyu Zhu, and Ping Tan. Quadtree attention for vision transformers. In International Conference on Learning Representations, 2022. 1", + "[54] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Herve Jegou. Training data-efficient image transformers & distillation through attention. In Proceedings of the 38th International Conference on Machine Learning, volume 139 of Proceedings of Machine Learning Research, pages 10347-10357. PMLR, 18-24 Jul 2021. 2, 5, 6, 14", + "[55] Hugo Touvron, Matthieu Cord, Alexandre Sablayrolles, Gabriel Synnaeve, and Hervé Jégou. Going deeper with image transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 32-42, 2021. 6", + "[56] Zhengzhong Tu, Hossein Talebi, Han Zhang, Feng Yang, Peyman Milanfar, Alan Bovik, and Yinxiao Li. Maxvit: Multi-axis vision transformer. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXIV, pages 459-479. Springer, 2022. 2, 13", + "[57] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Ilia Polosukhin. Attention is all you need. In Advances in Neural Information Processing Systems, volume 30. Curran Associates, Inc., 2017. 1, 3", + "[58] Sinong Wang, Belinda Z. Li, Madian Khabsa, Han Fang, and Hao Ma. Linformer: Self-attention with linear complexity, 2020. 1, 2, 3, 13", + "[59] Wenhai Wang, Enze Xie, Xiang Li, Deng-Ping Fan, Kaitao Song, Ding Liang, Tong Lu, Ping Luo, and Ling Shao. Pyramid vision transformer: A versatile backbone for dense prediction without convolutions. arXiv preprint arXiv:2102.12122, 2021. 2, 6, 14", + "[60] Wenhai Wang, Enze Xie, Xiang Li, Deng-Ping Fan, Kaitao Song, Ding Liang, Tong Lu, Ping Luo, and Ling Shao. Pyramid vision transformer: A versatile backbone for dense prediction without convolutions. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 568-578, October 2021. 3, 13", + "[61] Wenhai Wang, Enze Xie, Xiang Li, Deng-Ping Fan, Kaitao Song, Ding Liang, Tong Lu, Ping Luo, and Ling Shao. Pvt v2: Improved baselines with pyramid vision transformer. Computational Visual Media, 8(3):415-424, 2022. 6", + "[62] Bichen Wu, Chaojian Li, Hang Zhang, Xiaoliang Dai, Peizhao Zhang, Matthew Yu, Jialiang Wang, Yingyan Lin, and Peter Vajda. Fbnetv5: Neural architecture search for multiple tasks in one run. arXiv preprint arXiv:2111.10007, 2021.6,7", + "[63] Haiping Wu, Bin Xiao, Noel Codella, Mengchen Liu, Xiyang Dai, Lu Yuan, and Lei Zhang. Cvt: Introducing convolutions to vision transformers. arXiv preprint arXiv:2103.15808, 2021. 2" + ], + "bbox": [ + 78, + 90, + 470, + 900 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[64] Zhanghao Wu, Zhijian Liu, Ji Lin, Yujun Lin, and Song Han. Lite transformer with long-short range attention. In International Conference on Learning Representations, 2020. 1", + "[65] Yunyang Xiong, Hanxiao Liu, Suyog Gupta, Berkin Akin, Gabriel Bender, Pieter-Jan Kindermans, Mingxing Tan, Vikas Singh, and Bo Chen. Mobiledets: Searching for object detection architectures for mobile accelerators. arXiv preprint arXiv:2004.14525, 2020. 2, 6, 7", + "[66] Yunyang Xiong, Zhanpeng Zeng, Rudrasis Chakraborty, Mingxing Tan, Glenn Fung, Yin Li, and Vikas Singh. Nystromformer: A nystrom-based algorithm for approximating self-attention. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 35, pages 14138–14148, 2021. 1, 2, 13", + "[67] Zhi-Qin John Xu, Yaoyu Zhang, Tao Luo, Yanyang Xiao, and Zheng Ma. Frequency principle: Fourier analysis sheds light on deep neural networks. arXiv preprint arXiv:1901.06523, 2019. 2, 5, 8", + "[68] Hongxu Yin, Arash Vahdat, Jose M Alvarez, Arun Mallya, Jan Kautz, and Pavlo Molchanov. A-vit: Adaptive tokens for efficient vision transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10809-10818, 2022. 2", + "[69] Haoran You, Zhanyi Sun, Huihong Shi, Zhongzhi Yu, Yang Zhao, Yongan Zhang, Chaojian Li, Baopu Li, and Yingyan Lin. Vitcod: Vision transformer acceleration via dedicated algorithm and accelerator co-design. In The 29th IEEE International Symposium on High-Performance Computer Architecture (HPCA 2023), 2023. 1", + "[70] Guanghua Yu, Qinyao Chang, Wenyu Lv, Chang Xu, Cheng Cui, Wei Ji, Qingqing Dang, Kaipeng Deng, Guanzhong Wang, Yuning Du, et al. Pp-picodet: A better real-time object detector on mobile devices. arXiv preprint arXiv:2111.00902, 2021. 5, 6, 7", + "[71] Zhongzhi Yu, Yonggan Fu, Sicheng Li, Chaojian Li, and Yingyan Lin. Mia-former: Efficient and robust vision transformers via multi-grained input-adaptation. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 36, pages 8962-8970, 2022. 2", + "[72] Li Yuan, Yunpeng Chen, Tao Wang, Weihao Yu, Yujun Shi, Zi-Hang Jiang, Francis EH Tay, Jiashi Feng, and Shuicheng Yan. Tokens-to-token vit: Training vision transformers from scratch onImagenet. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 558-567, 2021. 2, 3", + "[73] Yuhui Yuan, Rao Fu, Lang Huang, Weihong Lin, Chao Zhang, Xilin Chen, and Jingdong Wang. Hrformer: High-resolution vision transformer for dense predict. Advances in Neural Information Processing Systems, 34:7281-7293, 2021.6", + "[74] Zixiao Zhang, Xiaogiang Lu, Guojin Cao, Yuting Yang, Licheng Jiao, and Fang Liu. Vit-yolo: Transformer-based yolo for object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2799-2808, 2021. 13", + "[75] Bolei Zhou, Hang Zhao, Xavier Puig, Sanja Fidler, Adela Barriuso, and Antonio Torralba. Scene parsing through" + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "14441", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "ade20k dataset. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 633-641, 2017. 5", + "[76] Chen Zhu, Wei Ping, Chaowei Xiao, Mohammad Shoeybi, Tom Goldstein, Anima Anandkumar, and Bryan Catanzaro. Long-short transformer: Efficient transformers for language and vision. Advances in Neural Information Processing Systems, 34, 2021. 1, 2", + "[77] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable detr: Deformable transformers for end-to-end object detection. arXiv preprint arXiv:2010.04159, 2020. 3, 13" + ], + "bbox": [ + 78, + 90, + 467, + 260 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "14442", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 11 + } +] \ No newline at end of file diff --git a/2023/Castling-ViT_ Compressing Self-Attention via Switching Towards Linear-Angular Attention at Vision Transformer Inference/d74c49bb-effe-4caf-8858-0066e120e08d_model.json b/2023/Castling-ViT_ Compressing Self-Attention via Switching Towards Linear-Angular Attention at Vision Transformer Inference/d74c49bb-effe-4caf-8858-0066e120e08d_model.json new file mode 100644 index 0000000000000000000000000000000000000000..be0e341d7a6b6aca652faef18df99e7bd3bf45cf --- /dev/null +++ b/2023/Castling-ViT_ Compressing Self-Attention via Switching Towards Linear-Angular Attention at Vision Transformer Inference/d74c49bb-effe-4caf-8858-0066e120e08d_model.json @@ -0,0 +1,2468 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.0, + 0.812, + 0.045 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.138, + 0.131, + 0.826, + 0.178 + ], + "angle": 0, + "content": "Castling-ViT: Compressing Self-Attention via Switching Towards Linear-Angular Attention at Vision Transformer Inference" + }, + { + "type": "text", + "bbox": [ + 0.157, + 0.194, + 0.813, + 0.284 + ], + "angle": 0, + "content": "Haoran You\\(^{1,2,\\dagger,*}\\), Yunyang Xiong\\(^{2,*}\\), Xiaoliang Dai\\(^{2}\\), Bichen Wu\\(^{2}\\), Peizhao Zhang\\(^{2}\\), Haoqi Fan\\(^{2}\\), Peter Vajda\\(^{2}\\), Yingyan (Celine) Lin\\(^{1}\\) \n\\(^{1}\\)Georgia Institute of Technology, \\(^{2}\\)Meta Research \n{haoran.you, celine.lin}@gatech.edu \n{yunyang, xiaoliang dai, wbc, stzpz, haoqifan, vajdap}@meta.com" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.307, + 0.314, + 0.323 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.339, + 0.473, + 0.807 + ], + "angle": 0, + "content": "Vision Transformers (ViTs) have shown impressive performance but still require a high computation cost as compared to convolutional neural networks (CNNs), one reason is that ViTs' attention measures global similarities and thus has a quadratic complexity with the number of input tokens. Existing efficient ViTs adopt local attention or linear attention, which sacrifice ViTs' capabilities of capturing either global or local context. In this work, we ask an important research question: Can ViTs learn both global and local context while being more efficient during inference? To this end, we propose a framework called Castling-ViT, which trains ViTs using both linear-angular attention and masked softmax-based quadratic attention, but then switches to having only linear-angular attention during inference. Our Castling-ViT leverages angular kernels to measure the similarities between queries and keys via spectral angles. And we further simplify it with two techniques: (1) a novel linear-angular attention mechanism: we decompose the angular kernels into linear terms and high-order residuals, and only keep the linear terms; and (2) we adopt two parameterized modules to approximate high-order residuals: a depthwise convolution and an auxiliary masked softmax attention to help learn global and local information, where the masks for softmax attention are regularized to gradually become zeros and thus incur no overhead during inference. Extensive experiments validate the effectiveness of our Castling-ViT, e.g., achieving up to a \\(1.8\\%\\) higher accuracy or \\(40\\%\\) MACs reduction on classification and 1.2 higher mAP on detection under comparable FLOPs, as compared to ViTs with vanilla softmax-based attentions. Project page is available at here." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.824, + 0.21, + 0.84 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.846, + 0.47, + 0.877 + ], + "angle": 0, + "content": "Vision Transformers (ViTs) have made significant progress in image classification, object detection, and many" + }, + { + "type": "image", + "bbox": [ + 0.507, + 0.307, + 0.889, + 0.429 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.506, + 0.432, + 0.889, + 0.552 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.5, + 0.553, + 0.892, + 0.582 + ], + "angle": 0, + "content": "Figure 1. Castling-ViT over SOTA baselines on (1) ImageNet [18] image classification and (2) COCO [36] object detection." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.594, + 0.892, + 0.881 + ], + "angle": 0, + "content": "other applications. It is well recognized that the superior performance achieved by ViTs is largely attributed to their self-attention modules that can better capture global context [20, 57, 64]. Nevertheless, ViTs' powerful self-attention module comes at the cost of quadratic complexity with the number of input tokens, causing a major efficiency bottleneck to ViTs' achievable runtime (i.e., inference latency) [3, 8, 32, 58, 66, 69, 76]. To mitigate this issue, linear attention designs have been developed to alleviate the vanilla ViT attention's quadratic complexity. In particular, existing efforts can be categorized into two clusters: (1) ViTs with local attention by restricting the attention window size [38, 53], sharing the attention queries [2], or representing the attention queries/keys with low rank matrices [58]; and (2) ViTs with kernel-based linear attention, which approximate the non-linearity softmax function by decomposing it into separate kernel embeddings. This enables a change in the matrix computation order for a reduced computational complexity [5, 6, 14, 29, 37, 39, 43, 66]." + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.885, + 0.892, + 0.902 + ], + "angle": 0, + "content": "Despite their promise in alleviating ViTs' complexity" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.887, + 0.456, + 0.9 + ], + "angle": 0, + "content": "*Equal contribution. † Work done while interning at Meta Research." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "14431" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.473, + 0.514 + ], + "angle": 0, + "content": "and thus inference runtime, both the local and linear attention compromise ViTs' performance due to the lack of capabilities to capture global or local context. To marry the best of both worlds, we advocate training ViTs with both (1) efficient but less powerful linear attention, i.e., without the high-order residuals in angular kernel expansion, and (2) powerful yet costly softmax-based masked attention. The latter helps approximate high-order residuals at the early training stage while being dropped during inference, based on an assumption that the remaining networks can gradually learn the high-order components at the later training stage [67]. This concept resembles the \"castling\" move in chess when two pieces are moved at once. While it sounds promising, there are still two challenges to achieve this. First, existing linear attention modules still underperform their vanilla softmax-based counterparts. Therefore, a better linear attention is crucial for the final performance. We find that angular kernels perform equally as softmax-based attentions in terms of similarity measurements. While they still suffer from a quadratic complexity, they can be divided into linear terms and high-order residuals. The challenge is how to construct ViTs with only the linear terms. Second, doing so would require that the trained ViTs merely rely on the linear terms towards the end of training, which would call for an approximation of the above high-order residuals. The challenge is that how we can resort to costly but powerful modules to approximate high-order residuals during training but does not incur extra inference cost." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.515, + 0.472, + 0.547 + ], + "angle": 0, + "content": "In this work, we develop techniques to tackle those challenges, and make the following contributions:" + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.557, + 0.47, + 0.633 + ], + "angle": 0, + "content": "- We propose a framework called Castling-ViT, which trains ViTs using both linear-angular attention and masked softmax-based quadratic attention, but then switches to having only linear-angular attentions during ViT inference to save computational costs." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.638, + 0.47, + 0.715 + ], + "angle": 0, + "content": "- We develop a new linear-angular attention leveraging angular kernels to close the accuracy gap between linear attention and softmax-based attention. It expands angular kernels where linear terms are kept while complex high-order residuals are approximated." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.719, + 0.47, + 0.795 + ], + "angle": 0, + "content": "- We use two parameterized modules to approximate the high-order residuals above: a depthwise convolution and an auxiliary masked softmax-based attention, where the latter's attention masks are regularized to gradually become zeros to avoid inference overhead." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.799, + 0.47, + 0.89 + ], + "angle": 0, + "content": "- We conduct extensive experiments to validate the effectiveness of the proposed Castling-ViT. Results on classification, detection, and segmentation tasks consistently demonstrate its superior performance (\\( \\uparrow 1.8\\% \\) top-1 accuracy or \\( \\uparrow 1.2 \\) mAP) or efficiency (40% MACs savings) over state-of-the-art (SOTA) CNNs and ViTs." + }, + { + "type": "list", + "bbox": [ + 0.097, + 0.557, + 0.47, + 0.89 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.09, + 0.651, + 0.107 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.113, + 0.895, + 0.522 + ], + "angle": 0, + "content": "Vision Transformers (ViTs). ViT [20] beats CNNs with a simple encoder-only transformer architecture taking the splitted non-overlapped image patches as sequential inputs, but relies on costly pretraining on a huge JFT-300M dataset [50]. Later, DeiT [54] and T2T-ViT [72] leverage an improved ViT training recipe or enhanced tokenization mechanism to achieve a comparable accuracy without the necessity of costly pretraining. To further improve ViTs' achievable accuracy-efficiency trade-offs, CrossViT [9], PiT [26], PVT [59], MViT [21] and Swin-Transformer [38] propose a pyramid-like architecture, which is commonly used in CNNs [16, 28, 65]; DynamicViT [46], A-ViT [68], ToME [4], and MIA-Former [71] propose to adaptively identify and remove unnecessary input tokens for saving computational costs. With the goal of deploying ViTs in resource-constrained devices, various efficient ViT architectures have been proposed [6, 24, 35, 40]. For example, LeViT [24], CvT [63], and MobileViT [40] adopt more efficient self-attention implementation or incorporate convolutional feature extraction blocks into their early layers; EfficientFormer [35] further enables pure ViTs to run as fast as MobileNets. In contrast, our Castling-ViT explores whether ViTs can learn both global and local features while still being efficient at runtime. Also, we target a generic linear-angular attention that can serve as a drop-in replacement for all kinds of ViT architectures and thus is orthogonal to new ViT architecture designs." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.524, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Efficient ViT Variants. As commonly recognized, ViTs rely heavily on their self-attention module which is however costly due to its quadratic computational complexity with the total number of input tokens [29, 76]. To make the self-attention module more efficient, a surge of linear attention works have been proposed and can be roughly categorized into two groups: local attention [2, 38, 56, 58] or kernel-based linear attention [1, 5, 6, 14, 29, 37, 39, 43, 66]. For kernel-based linear attention, common designs approximate the softmax function [5, 14, 29] or the full self-attention matrix [39, 66] with orthogonal features or kernel embeddings, then the computation order can be changed from (QK)V to Q(KV). For example, [29] and [6] decompose the exponential terms in softmax-based attention into kernel functions and exchange the computation order. Despite their decent performance, currently kernel-based linear attention in general underperform the softmax-based attention. Recent works [8, 17] also unify low rank approximated and sparse attention (can also be dropped at inference) to improve ViTs' accuracy-efficiency tradeoffs. Different from the above works, we explore from a new perspective by taking spectral angles into consideration when measuring the similarities among tokens, resulting in linear-angular attention that can achieve comparable or even better performance than softmax-based attention. More efficient ViT variants" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "14432" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.275, + 0.106 + ], + "angle": 0, + "content": "are supplied to the Appendix." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.107, + 0.47, + 0.152 + ], + "angle": 0, + "content": "ViTs have also been used as the backbones for downstream tasks, e.g., detection [34, 60, 77] and segmentation [11,42]. We supply more literature review to Appendix." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.166, + 0.296, + 0.183 + ], + "angle": 0, + "content": "3. The Proposed Methods" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.191, + 0.338, + 0.207 + ], + "angle": 0, + "content": "3.1. Preliminary of Self-Attention" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.21, + 0.471, + 0.286 + ], + "angle": 0, + "content": "Self-Attention. Self-attention module is a core component of the Transformer [20, 57], and usually consists of multiple heads. Each head captures global-context information by measuring pairwise correlations among all \\(N\\) tokens (\\(N\\) denotes the total number of tokens) as defined below:" + }, + { + "type": "equation", + "bbox": [ + 0.132, + 0.292, + 0.47, + 0.335 + ], + "angle": 0, + "content": "\\[\n\\mathbf {H} _ {t} ^ {k} = \\sum_ {i = 1} ^ {N} \\frac {\\exp \\left(\\mathbf {Q} _ {t} \\cdot \\mathbf {K} _ {i} ^ {T} / \\sqrt {d}\\right)}{\\sum_ {j = 1} ^ {N} \\exp \\left(\\mathbf {Q} _ {t} \\cdot \\mathbf {K} _ {j} ^ {T} / \\sqrt {d}\\right)} \\cdot \\mathbf {V} _ {i}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.341, + 0.47, + 0.539 + ], + "angle": 0, + "content": "where \\(t\\in \\{1,\\dots ,N\\} ,k\\in \\{1,\\dots ,M\\}\\) \\(\\mathbf{H}_t^k\\) refers to the \\(t\\) th row of the \\(m\\) -th head's attention matrix H. \\(\\mathbf{Q}_t,\\mathbf{K}_t,\\mathbf{V}_t\\in\\) \\(\\mathbf{R}^d\\) , are the query, key, and value vectors obtained by linearly projecting the input \\(\\mathbf{X_t}\\in \\mathbb{R}^D\\) with three learnable weight matrices \\(W^{Q},W^{K},W^{V}\\in \\mathbb{R}^{D\\times d}\\) . The attention head first computes the inner product between the query-key pairs, then scales the product results to stabilize the training and uses Softmax to normalize the resulting attention scores, and finally computes a weighted sum over all value vectors. Outputs from all attention heads are then concatenated together before a final linear projection with learnable weights. Note that \\(\\exp (\\cdot)\\) denotes an exponential function. Computing Eq. (1) has a quadratic complexity of \\(\\mathcal{O}(N^2)\\)" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.539, + 0.471, + 0.706 + ], + "angle": 0, + "content": "Kernel-based Linear Attention. The core idea of linear attention [29, 49, 58] is to decompose the similarity measurement function into separate kernel embeddings, i.e., \\(\\mathrm{Sim}(\\mathbf{Q},\\mathbf{K})\\approx \\phi (\\mathbf{Q})\\phi (\\mathbf{K})^T\\) so that we can change the computation order to \\(\\phi (\\mathbf{Q})(\\phi (\\mathbf{K})^T\\mathbf{V})\\) based on the associative property of matrix multiplication. In this way, the attention complexity is quadratic to the feature dimension \\(d\\) instead of the token length \\(N\\). One straightforward implementation of linear attention is to use Gaussian RBF kernels to measure the similarity, which can serve as an unbiased estimation of \\(\\exp (\\langle \\cdot ,\\cdot \\rangle)\\) in Eq. (1):" + }, + { + "type": "equation", + "bbox": [ + 0.086, + 0.711, + 0.469, + 0.761 + ], + "angle": 0, + "content": "\\[\n\\exp \\left(\\frac {\\langle \\mathbf {x} , \\mathbf {y} \\rangle}{\\sigma^ {2}}\\right) = \\underbrace {\\exp \\left(\\frac {- \\| \\mathbf {x} - \\mathbf {y} \\| ^ {2}}{2 \\sigma^ {2}}\\right)} _ {\\text {G a u s s i a n R B F K e r n e l}} \\exp \\left(\\frac {\\| \\mathbf {x} \\| ^ {2} + \\| \\mathbf {y} \\| ^ {2}}{2 \\sigma^ {2}}\\right) \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.767, + 0.47, + 0.859 + ], + "angle": 0, + "content": "where \\(\\langle \\cdot, \\cdot \\rangle\\) denotes the inner product operator. According to [45], we can induce a function \\(\\phi(\\cdot)\\) to approximate the Gaussian RBF kernel, mapping the input space to the feature space. Assuming that both \\(\\mathbf{Q}\\) and \\(\\mathbf{K}\\) are normalized as unit row vectors along the feature dimension, then the attention formula can be approximated by:" + }, + { + "type": "equation", + "bbox": [ + 0.152, + 0.864, + 0.469, + 0.906 + ], + "angle": 0, + "content": "\\[\n\\mathbf {H} _ {t} ^ {k} = \\frac {\\phi (\\mathbf {Q} _ {t}) \\cdot \\sum_ {i = 1} ^ {N} \\left(\\phi \\left(\\mathbf {K} _ {i}\\right) ^ {T} \\cdot \\mathbf {V} _ {i}\\right)}{\\phi (\\mathbf {Q} _ {t}) \\cdot \\sum_ {j = 1} ^ {N} \\phi \\left(\\mathbf {K} _ {j} ^ {T}\\right)} \\tag {3}\n\\]" + }, + { + "type": "image", + "bbox": [ + 0.502, + 0.089, + 0.895, + 0.31 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.5, + 0.313, + 0.892, + 0.328 + ], + "angle": 0, + "content": "Figure 2. Illustration of self-attention used in LeViT and MViTv2." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.337, + 0.892, + 0.397 + ], + "angle": 0, + "content": "While doing so can alleviate the attention complexity to become linear w.r.t. the token length \\(N\\), it often causes a nontrivial accuracy drop as compared to the corresponding ViTs with vanilla softmax-based attention [6, 13, 72]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.399, + 0.893, + 0.732 + ], + "angle": 0, + "content": "Revisit Attention Designs. Recent ViTs have achieved a low complexity for classification tasks (e.g., LeViT [24]) but still be costly for downstream tasks due to high input resolutions, i.e., the number of tokens (e.g., MViTv2 [34]). Therefore, one dilemma is the tradeoff between the model efficiency and generalizability. As illustrated in Fig. 2, LeViT shrinks the feature dimension to be more efficient on low-resolution tasks but cannot be well generalized to downstream tasks, due to the resulting (1) feature bottlenecks, i.e., insufficient feature dimensions, and (2) fixed attention biases. On the other hand, MViTv2 performs token pooling for better fitting high-resolution tasks but is less effective for low-resolution due to the caused token bottlenecks, i.e., insufficient number of tokens. Our ablation studies in Table 1 show that (1) for ImageNet classification, both token and feature pooling can lead to accuracy drops in ViTs; and (2) for downsampling layers, post-Q pooling (i.e., pooling after linear projection) together with residual connections in MViTv2 performs better than pre-Q pooling (i.e., pooling during linear projection) in LeViT. Therefore, we develop Castling-LeViTs on top of attention with merely post-Q pooling and residual connections." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.743, + 0.842, + 0.759 + ], + "angle": 0, + "content": "3.2. The Proposed Castling-ViT Framework" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Castling-ViT Overview. Fig. 4 illustrates an overview of the proposed Castling-ViT, which makes linear attention more powerful than previous designs while still being efficient during inference. In particular, we propose (1) a novel kernel-based linear-angular attention from the spectral angle perspective to close the accuracy gap between linear attention and softmax-based attention; and (2) a training augmentation method that leverages softmax-based attention as an auxiliary branch to assist the linear-angular atten" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "14433" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.08, + 0.09, + 0.467, + 0.103 + ], + "angle": 0, + "content": "Table 1. Analyze the attention design on ImageNet classification." + }, + { + "type": "table", + "bbox": [ + 0.081, + 0.105, + 0.47, + 0.202 + ], + "angle": 0, + "content": "
PoolingResidual QMACs (M)Top-1 Accuracy (%)
TokenFeat.Pre-QPost-Q
✓✓68578.00
✓✓66178.19
✓73679.11
✓✓✓77178.73
✓✓✓76377.23
✓✓83880.05
" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.212, + 0.47, + 0.258 + ], + "angle": 0, + "content": "tion only during ViT training. Note that a mask is applied to the auxiliary branch to manifest linear-angular attention and drop the auxiliary branch." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.267, + 0.313, + 0.282 + ], + "angle": 0, + "content": "3.2.1 Linear-Angular Attention" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.287, + 0.47, + 0.424 + ], + "angle": 0, + "content": "Angular Kernel. In addition to the previously adopted polynomial, exponential, or RBF kernel that focuses on spatial similarity measurements [29, 37], we propose to consider measuring spectral similarity via angular kernel as an alternative to existing softmax-based attentions, leading to similar or better performance since it additionally takes into consideration the nature of spectral characteristics, e.g., the spectral angle as a distance measurement function [27, 30]. Such a spectral angle between two vectors is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.151, + 0.429, + 0.47, + 0.463 + ], + "angle": 0, + "content": "\\[\n\\theta \\left(\\mathbf {x} _ {i}, \\mathbf {x} _ {j}\\right) = \\arccos \\left(\\frac {\\left\\langle \\mathbf {x} _ {i} , \\mathbf {x} _ {j} \\right\\rangle}{\\left\\| \\mathbf {x} _ {i} \\right\\| \\cdot \\left\\| \\mathbf {x} _ {j} \\right\\|}\\right), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.468, + 0.47, + 0.545 + ], + "angle": 0, + "content": "where \\( \\|\\cdot\\| \\) is the Euclidean distance and \\( \\langle \\cdot, \\cdot \\rangle \\) is the inner product. The output range of \\( \\theta \\) is \\( [0, \\pi] \\). Such an angle can be used as a distance. In our design, we define the angular kernel as a similarity measurement function between the queries \\( \\mathbf{Q} \\) and keys \\( \\mathbf{K} \\) as:" + }, + { + "type": "equation", + "bbox": [ + 0.152, + 0.549, + 0.469, + 0.579 + ], + "angle": 0, + "content": "\\[\n\\operatorname {S i m} \\left(\\mathbf {Q} _ {i}, \\mathbf {K} _ {j}\\right) = 1 - \\frac {1}{\\pi} \\cdot \\theta \\left(\\mathbf {Q} _ {i}, \\mathbf {K} _ {j}\\right), \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.584, + 0.47, + 0.645 + ], + "angle": 0, + "content": "and the output range is thus [0, 1]. With more aligned \\(\\mathbf{Q}_i\\) and \\(\\mathbf{K}_j\\), \\(\\theta\\) is closer to 0 and thus the similarity is closer to 1; In contrast, if \\(\\mathbf{Q}_i\\) and \\(\\mathbf{K}_j\\) have opposite features, \\(\\theta\\) is closer to \\(\\pi\\) and thus the similarity is closer to 0." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.645, + 0.47, + 0.766 + ], + "angle": 0, + "content": "Properties of Angular Kernel and Its Feature Space. One property of our angular kernel is that replacing the similarity measurement in self-attention with such a kernel provides an efficient way to implicitly map the input data to a high (even infinite after expansion) dimensional feature space [27], where the distances/angles are calculated based on a rich feature structure. Let \\(\\phi(\\cdot)\\) denotes the implicit map induced by this kernel, the norm of mapped input data is:" + }, + { + "type": "equation", + "bbox": [ + 0.132, + 0.771, + 0.469, + 0.8 + ], + "angle": 0, + "content": "\\[\n\\left\\| \\phi \\left(\\mathbf {x} _ {i}\\right) \\right\\| ^ {2} = \\operatorname {S i m} \\left(\\mathbf {x} _ {i}, \\mathbf {x} _ {i}\\right) = 1 - \\frac {1}{\\pi} \\cdot 0 = 1, \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.805, + 0.47, + 0.85 + ], + "angle": 0, + "content": "which means that all data in the input space are mapped onto the sphere of radius 1 in the feature space. Also, the distance between two input features is given by:" + }, + { + "type": "equation", + "bbox": [ + 0.102, + 0.855, + 0.47, + 0.906 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\left\\| \\phi (\\mathbf {x} _ {i}) - \\phi (\\mathbf {x} _ {j}) \\right\\| ^ {2} = 1 + 1 - 2 \\cdot \\phi (\\mathbf {x} _ {i}) \\phi (\\mathbf {x} _ {j}) ^ {T} (7) \\\\ = 2 \\cdot \\left(1 - \\operatorname {S i m} \\left(\\mathbf {x} _ {i}, \\mathbf {x} _ {j}\\right)\\right) = \\frac {2}{\\pi} \\cdot \\theta \\left(\\mathbf {x} _ {i}, \\mathbf {x} _ {j}\\right). (7) \\\\ \\end{array}\n\\]" + }, + { + "type": "image", + "bbox": [ + 0.539, + 0.096, + 0.632, + 0.159 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.547, + 0.163, + 0.619, + 0.174 + ], + "angle": 0, + "content": "(a) Input space" + }, + { + "type": "image", + "bbox": [ + 0.665, + 0.088, + 0.852, + 0.161 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.704, + 0.163, + 0.788, + 0.174 + ], + "angle": 0, + "content": "(b) Feature space" + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.175, + 0.892, + 0.217 + ], + "angle": 0, + "content": "Figure 3. (a) A simple 2D input space and (b) the corresponding 3D feature space after applying the angular kernel to the 2D input data. Note that real input data/features are of higher dimensions." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.223, + 0.892, + 0.283 + ], + "angle": 0, + "content": "That is, the square (Euclid) distance and the spectral angle is positively correlated and the distance range is \\([0, 2]\\). As shown in Fig. 3, the angles in the input space are correlated to the feature distance after applying our angular kernel." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.283, + 0.892, + 0.374 + ], + "angle": 0, + "content": "Expansion of Angular Kernel. A natural following question is how to incorporate the above angular kernel for designing linear attention given its quadratic complexity w.r.t. the input token length. Recalling from trigonometric identities and the expansion of the arccos function into an infinite series, we reformulate the similarity function as:" + }, + { + "type": "equation", + "bbox": [ + 0.516, + 0.381, + 0.891, + 0.526 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\operatorname {S i m} \\left(\\mathbf {Q} _ {i}, \\mathbf {K} _ {j}\\right) = 1 - \\frac {1}{\\pi} \\cdot \\left(\\frac {\\pi}{2} - \\arcsin \\left(\\mathbf {Q} _ {i} \\cdot \\mathbf {K} _ {j} ^ {T}\\right)\\right) \\\\ = \\underbrace {\\frac {1}{2} + \\frac {1}{\\pi} \\cdot \\left(\\mathbf {Q} _ {i} \\cdot \\mathbf {K} _ {j} ^ {T}\\right)} _ {\\text {L i n e a r - a n g u l a r t e r m s}} \\tag {8} \\\\ + \\underbrace {\\frac {1}{\\pi} \\cdot \\sum_ {k = 1} ^ {\\infty} \\frac {(2 k) !}{2 ^ {2 k} (k !) ^ {2} (2 k + 1)} \\left(\\mathbf {Q} _ {i} \\cdot \\mathbf {K} _ {j} ^ {T}\\right) ^ {2 k + 1}} _ {\\text {H i g h - o r d e r r e s i d u a l s w h e n k \\geq 1 (n o n - l i n e a r)}}, \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.535, + 0.892, + 0.655 + ], + "angle": 0, + "content": "where \\(\\mathbf{Q}_i\\cdot \\mathbf{K}_j^T\\) denotes the normalized linear kernel function \\(\\langle \\mathbf{Q}_i,\\mathbf{K}_j\\rangle /\\| \\mathbf{Q}_i\\| \\cdot \\| \\mathbf{K}_j\\|\\), which is equivalent to the inner product if \\(\\mathbf{Q}_i\\) and \\(\\mathbf{K}_j\\) are unit vectors. We see that the first linear-angular terms can be directly used as the similarity measurement in linear attention, while the remaining higher-order terms of an infinite series introduce a much higher complexity. As such, we propose to adopt a relaxation to approximate it." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.656, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Linear-Angular Attention. To construction our linear- angular attention, we leverage parametrized DNN modules to approximate the expectation of the high-order residuals, i.e., \\(\\mathbf{M}_{ij} = \\mathbb{E}_{\\mathbf{Q}_i\\sim p_{Q_i},\\mathbf{K}_j\\sim p_{K_j}}\\left(\\sum_{k = 1}^{\\infty}\\alpha_k(\\mathbf{Q}_i\\cdot \\mathbf{K}_j^T)^{2k + 1}\\right)\\) where \\(\\alpha_{k}\\) is the coefficient for the \\(k\\) -th order term in Eq. (8). The key question is how to design such DNN modules. Kim et al. [31] conducted an analysis on average attention weights of ViT models, which show a strong inductive bias to attend to neighboring tokens. To capture this, we introduce a learnable depthwise convolution (DWConv) module which is applied on all the value tokens, so \\(\\mathbf{M}_{ij}\\) by-design attends to nearby tokens, as illustrated in Fig. 4. Second, DWConv is limited by its receptive field [41] while the averaged attention map also has a small number of offdiagonal scores to connect to nonadjacent tokens [31]. To capture that, we adopt a sparse softmax-based attention as" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "14434" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.08, + 0.089, + 0.47, + 0.35 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.352, + 0.47, + 0.381 + ], + "angle": 0, + "content": "Figure 4. The visualization of our Castling-ViT with both linear-angular attention and auxiliary softmax-based self-attention." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.391, + 0.471, + 0.451 + ], + "angle": 0, + "content": "described in Section 3.2.2. In practice, we found that such a sparse softmax attention tends to converge to all zeros after adding a simple threshold regularization. We provide more explanation and visualization in Sec. 3.2.2 and Sec. 4.5." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.452, + 0.462, + 0.466 + ], + "angle": 0, + "content": "In this way, our attention module can be formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.095, + 0.473, + 0.469, + 0.521 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathbf {H} = \\operatorname {S i m} (\\mathbf {Q}, \\mathbf {K}) \\cdot \\mathbf {V} \\approx \\frac {1}{2} \\cdot \\mathbf {V} + \\frac {1}{\\pi} \\cdot \\mathbf {Q} \\cdot \\left(\\mathbf {K} ^ {T} \\cdot \\mathbf {V}\\right) \\tag {9} \\\\ + \\mathbf {M} _ {D W} \\cdot \\mathbf {V} + \\mathbf {M} _ {S p a r s e A t t n}, \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.528, + 0.47, + 0.651 + ], + "angle": 0, + "content": "where \\(1 / \\pi \\cdot Q\\cdot (K^T\\cdot V)\\) is the linear term with \\(\\mathcal{O}(N)\\) complexity, \\(\\mathbf{M}_{DW}\\cdot V\\) is the matrix form of DWConv, and \\(\\mathbf{M}_{SparseAttn}\\) is the normalized sparse softmax attention. The overall complexity to compute Eq. (9) is linear to the input token length, where the MACs of DWConv is also negligible (e.g., \\(< 1\\%\\) of the total MACs). Also, normalizations \\((\\mathrm{x} / \\| \\mathrm{x}\\|_{2})\\) are inserted to the \\(\\mathbf{Q} / \\mathbf{K}\\) and sparse attention branches to help the similarity measurement following [5]." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.655, + 0.427, + 0.669 + ], + "angle": 0, + "content": "3.2.2 Switch Towards Linear-Angular Attention" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.675, + 0.469, + 0.795 + ], + "angle": 0, + "content": "Recall that we add a sparse softmax-based attention as an auxiliary branch in the linear-angular attention to help approximate the high-order residuals. Such a costly attention can be potentially dropped without hurting the inference accuracy, drawing inspiration from [67] that the remaining network can gradually learn the high-order/frequency components at the later training stage. Next, we explain how it is constructed and dropped." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.796, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Sparse Training Augmentation. As illustrated in Fig. 4, we adopt a masked softmax-based attention as an auxiliary branch to augment ViT training. In particular, we first use a comparator with a predefined threshold \\(\\epsilon\\) to generate a binary mask, where attentions greater than \\(\\epsilon\\) are set to 1 and 0 otherwise. These masks are then applied to the attention maps to generate masked attention maps that will be" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.892, + 0.122 + ], + "angle": 0, + "content": "summed up with our linear-angular attention together as the final attention. The resulting sparse attention is given by:" + }, + { + "type": "equation", + "bbox": [ + 0.512, + 0.13, + 0.892, + 0.148 + ], + "angle": 0, + "content": "\\[\n\\mathbf {M} _ {\\text {S p a r s e A t t n}} (\\mathbf {Q}, \\mathbf {K}) = \\operatorname {M a s k} _ {\\varepsilon} \\left(\\operatorname {S o f t m a x} \\left(\\mathbf {Q} \\cdot \\mathbf {K} ^ {\\mathrm {T}}\\right)\\right). \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.158, + 0.892, + 0.291 + ], + "angle": 0, + "content": "where \\(\\mathsf{Mask}_{\\epsilon}(\\mathbf{x}) = \\mathbf{x}\\) if \\(\\mathbf{x} > \\epsilon\\) else 0, acting as an elementwise threshold function. As such, the sparse attention captures the higher attention scores (i.e., strong local features), and can potentially complement our linear-angular attention by supplementing the missing higher-order terms. Such an assumption aligns well with recent findings that (1) low-rank and sparse approximations complement each other [8, 17, 31]; and (2) linear attention lacks local feature extraction capabilities over its softmax-based counterpart [6]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.292, + 0.893, + 0.534 + ], + "angle": 0, + "content": "Castling During ViT Inference. As we are targeting efficient ViT inference, it is desired to reduce or completely remove the costly softmax-based attentions while only keeping our linear-angular attention at runtime, i.e., performing castling. In our experiments, we found that under the sparsity regularization above, the softmax attention naturally converges to all zeros as the training progresses. We consider both fixed and dynamic schedules for \\(\\epsilon\\) and find that our Castling-ViT is not sensitive to neither the threshold value nor the threshold schedule for a given task. Given a fixed mask threshold (e.g., \\(\\epsilon = 0.02\\) in image classification experiments), the masks become all zeros at latter training stages and thus the auxiliary branch can be removed without hurting the model accuracy. We supply the visualization of the mask-evolving trajectory and our conjecture for understanding such a phenomenon in Sec. 4.5." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.542, + 0.634, + 0.559 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.563, + 0.694, + 0.58 + ], + "angle": 0, + "content": "4.1. Experiment Settings" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.582, + 0.892, + 0.824 + ], + "angle": 0, + "content": "Tasks, Datasets, and Models. Tasks and Datasets. We consider three benchmark datasets and three representative vision tasks to demonstrate the superiority of the proposed Castling-ViT, including image classification on ImageNet dataset [18] with 1.2 million training and 50K validation images; Object detection on COCO dataset [36] with 118K training and 5K validation images; Semantic segmentation on ADE20K dataset [75] with 20K/2K/3K images for training, validation, and testing, respectively. Models. We apply our proposed Castling-ViT idea on top of various models. For the classification task, we consider LeViT [24], MViTv2 [34], and DeiT [54]; For the detection task, we consider models with efficient ViT backbones (e.g., PicoDet [70] with modified ESNet and LCNet backbones [15] with transformer blocks); For the segmentation task, we consider Mask2former [11] with ViT-Base backbone." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.824, + 0.893, + 0.901 + ], + "angle": 0, + "content": "Training Settings. For the classification task, we use a SGD optimizer with 0.9 momentum and \\(2 \\times 10^{-5}\\) weight decay to train ViTs for 1000 epochs using 64 V100 GPUs, with each card having 64 (LeViT) or 32 (MViTv2/DeiT) batch sizes. The learning rate is 2.0" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "14435" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.102, + 0.09, + 0.444, + 0.103 + ], + "angle": 0, + "content": "Table 2. Castling-ViT over SOTA baselines on ImageNet." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.105, + 0.467, + 0.556 + ], + "angle": 0, + "content": "
MACs RangesModelsParams (M)MACs (G)Top-1 Acc.(%)Top-5 Acc.(%)
<1GMobileNetV2-1.4 [47]6.90.5874.7-
EfficientNet-B0 [51]5.30.3977.193.3
EfficientNet-B1 [51]7.80.7079.194.4
MobileViT-XS [40]2.30.7074.892.3
LeViT-128 [24]9.20.4178.6-
LeViT-192 [24]10.90.6680.0-
Castling-LeViT-12810.50.4979.694.6
Castling-LeViT-19212.70.8281.395.5
1~3GEfficientNet-B3 [51]12.01.8081.695.7
HRFormer-T [73]8.01.8078.5-
DeiT-T [54]5.61.2574.591.9
LeViT-256 [24]18.91.1281.6-
LeViT-384 [24]39.12.3582.6-
Caslting-DeiT-T5.641.1876.092.5
Castling-LeViT-25622.01.4082.696.1
Castling-LeViT-38445.82.9083.796.7
3~10GRegNetY-4G [44]21.04.0080.0-
RegNetY-8G [44]39.08.0081.7-
EfficientNet-B4 [51]19.04.2082.996.4
EfficientNet-B5 [51]30.09.9083.696.7
Swin-T [38]29.04.5081.395.5
Swin-S [38]50.08.7083.296.2
CSWin-T [19]23.04.3082.7-
CSWin-S [19]35.06.9083.6-
DeiT-S [54]21.94.6081.295.4
AutoFormer-S [10]22.95.1081.795.7
PVTv2-V2 [61]25.04.0082.0-
MViTv2-T [34]24.04.6882.3-
MViTv2-S [34]34.76.9583.6-
Castling-MViTv2-T24.14.5084.196.8
Castling-MViTv2-S34.76.9584.697.0
>10GCaiT-S36 [55]68.013.9083.3-
Swin-B [38]88.015.4083.596.5
CSWin-B [19]78.015.0084.2-
AutoFormer-B [10]54.011.0082.495.7
MViTv2-B [34]51.210.0784.4-
DeiT-B [54]86.317.5683.496.5
Castling-DeiT-B87.2217.2884.2-
Castling-MViTv2-B51.99.8285.097.2
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.567, + 0.468, + 0.748 + ], + "angle": 0, + "content": "with first 11 epochs warm-up starting from 0.01 and decays by a factor of 0.9875 per epoch [62]. Also, we use the distillation [54] based on a teacher model with a \\(85.5\\%\\) accuracy. For the detection task, we adopt SGD optimizer with momentum 0.9 and weight decay 4e-5 to train models on COCO. All models are trained on 8 V100 GPUs with each card having 80 batch sizes following PicoDet's training recipe [70]. Also, we follow LeViT's training recipe [24] to pretrain backbones on ImageNet; For the segmentation task, we follow Mask2former's training recipe [11] to train models on ADE20K, where ViT backbones are pretrained following MAE [25] if specified." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.468, + 0.902 + ], + "angle": 0, + "content": "Baselines and Evaluation Metrics. Baselines. For the classification task, we compare the proposed Castling-ViT with LeViT [24], MviTv2 [34], DeiT [54], Swin [38], CSWin [19], PVT [59], etc. For the detection task, we compare with FBNetV5 [62], YOLOX [23], YOLOv5, MobileDet [65], and EfficientDet [52]. For the segmentation task, we compare with Mask2former [11] with ViT backbones. Evaluation Metrics. We evaluate the Castling-ViT and all baselines in terms of accuracy-efficiency tradeoffs. Specifically, the accuracy metrics refer to top-1/5 accuracy" + }, + { + "type": "image", + "bbox": [ + 0.508, + 0.093, + 0.695, + 0.189 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.703, + 0.094, + 0.886, + 0.19 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.548, + 0.197, + 0.842, + 0.211 + ], + "angle": 0, + "content": "Figure 5. Castling vs. baseline ViT on ImageNet." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.22, + 0.892, + 0.296 + ], + "angle": 0, + "content": "for the classification task; AP, \\(\\mathrm{AP}^{50}\\), \\(\\mathrm{AP}^{75}\\) for the detection task (AP: average precision); mIoU, mAcc, and pAcc for the segmentation task (mIoU: mean intersection over union). For efficiency metrics, we compare the number of model parameters or inference FLOPs (or MACs)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.306, + 0.793, + 0.321 + ], + "angle": 0, + "content": "4.2. Castling-ViT over SOTA Baslines" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.326, + 0.892, + 0.824 + ], + "angle": 0, + "content": "Image Classification. To evaluate the effectiveness of our proposed techniques on the image classification task, we apply the proposed Castling-ViT idea to three typical or SOTA ViT architectures: DeiT [54] as typical ViTs, LeViT [24] as representative efficient ViTs, and MViTv2 [34] as representative hierarchical ViTs for downstream applications of high input resolutions, and compare their performance over baselines on ImageNet. As shown in Tab. 2, the comparison across a large MACs (or FLOPs) range starting from 0.4G to 17G. We categorize ViT models into four regimes: \\(<1\\mathrm{G}\\), \\(1 \\sim 3\\mathrm{G}\\), \\(3 \\sim 10\\mathrm{G}\\), and \\(>10\\mathrm{G}\\), and select baselines in each regime to benchmark separately for the clarity purpose. Castling-ViTs consistently perform better than all baselines across various MACs ranges in terms of the accuracy-efficiency tradeoff. For example, Castling-LeViT achieves \\(82.6\\%\\) top-1 accuracy with only 1.40G MACs while LeViT requires 2.35G FLOPs instead, i.e., \\(\\downarrow 40\\%\\) MACs; On the other hand, under comparable MACs, Castling-MViTv2 achieves \\(84.1\\%\\) accuracy vs. MViTv2 with \\(82.3\\%\\) accuracy instead, i.e., \\(\\uparrow 1.8\\%\\) top-1 accuracy. Overall, Castling-ViT achieves an improved accuracy of \\(\\uparrow 0.5\\% \\sim \\uparrow 6.6\\%\\), \\(\\uparrow 1.0\\% \\sim \\uparrow 8.1\\%\\), \\(\\uparrow 1.0\\% \\sim \\uparrow 4.1\\%\\), and \\(\\uparrow 0.6\\% \\sim \\uparrow 2.6\\%\\) over baselines under \\(<1\\mathrm{G}\\) MACs, \\(1 \\sim 3\\mathrm{G}\\) MACs, \\(3 \\sim 10\\mathrm{G}\\) MACs, and \\(>10\\mathrm{G}\\) MACs, respectively. Note that we calculate improvements under comparable MACs. In addition to the overall comparison, we also visualize the apple-to/apple benchmark, e.g., Castling-LeViT vs. LeViT, to validate the effectiveness of proposed techniques. As shown in Fig. 5, Castling-LeViT/MViTv2 achieves \\(\\downarrow 25.7\\% \\sim \\downarrow 55.3\\%\\) MACs reductions under comparable accuracies or offers a comparable or better accuracy (\\(\\uparrow 0.6\\% \\sim \\uparrow 1.8\\%\\)) under comparable MACs over corresponding LeViT/MViTv2 baselines." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.825, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Object Detection. We also extend the Castling-ViT to the downstream object detection task and compare it with previous efficient detectors on COCO dataset to evaluate its efficacy. Specifically, we construct the detector with modified ESNet [70] or LCNet [15] (replace the last one or two" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "14436" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.089, + 0.472, + 0.201 + ], + "angle": 0, + "content": "Table 3. Castling-ViT over SOTA baselines on COCO detection, where Castling-ViT† means that the detection model adopts modified ESNet (use transformer blocks in the last stage; referred to ESNet-ViT) as backbone while Castling-ViT‡ uses modified LCNet (use transformer blocks in the last two stage; referred to LCNet-ViT) as backbone. Also, “-S/M/L” and “-320/416” represent small/medium/large model variants and 320x320/416x416 input resolutions, respectively." + }, + { + "type": "table", + "bbox": [ + 0.081, + 0.203, + 0.47, + 0.413 + ], + "angle": 0, + "content": "
ModelsParams (M)MACs (G)mAPAP50AP75
MobileDet-DSP [65]4.851.4327.3--
MobileDet-DSP-Fused [65]9.153.2229.1--
EfficientDet-D0 [52]3.902.5434.653.037.1
YOLOv5-N1.902.3028.045.7-
YOLOX-Nano [23]0.910.5425.8--
YOLOX-Tiny [23]5.063.2332.8--
FBNetV5-AC-224x320 [62]-0.7125.0--
FBNetV5-AR-224x320 [62]-0.9126.4--
FBNetV5-A-224x320 [62]-1.3527.2--
FBNetV5-AC-320x640 [62]-1.3728.9--
FBNetV5-AR-320x640 [62]-1.8030.4--
Castling-ViT-S-320†3.250.6228.142.329.2
Castling-ViT-S-416†3.251.0631.346.732.5
Castling-ViT-M-416†6.012.0034.049.535.9
Castling-ViT-L-416†9.323.0335.050.737.2
Castling-ViT-L-416*13.105.3137.353.439.6
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.429, + 0.47, + 0.7 + ], + "angle": 0, + "content": "stages with transformer blocks) as backbones and follow PicoDet [70]'s detection head design as well as their training recipe. Fig. 1 and Tab. 3 show the overall comparison between the proposed Castling-ViT and other baselines. We can see that our Castling-ViT consistently achieves better accuracy-efficiency tradeoffs, leading to \\(\\uparrow 6.0\\), \\(\\uparrow 2.2 \\sim \\uparrow 2.3\\), \\(\\uparrow 4.0 \\sim \\uparrow 5.9\\), \\(\\uparrow 3.1 \\sim \\uparrow 4.0\\) mAP improvements as compared to YOLOv5, YOLOX [23], MobileDet [65], and FBNetv5 [62], respectively, under comparable or even less MACs. As for EfficientDet [52], our method achieves comparable accuracy-efficiency trade-offs. Apart from the overall comparison with baselines, we also provide the apple-to-Apple comparison with detectors with softmax-based attention, we supply the comparison results and Castling-ViT's breakdown analysis to Sec. 4.4. This set of experiments validate the effectiveness of the proposed Castling-ViT for servering as efficient detector backbones (e.g., \\(\\leq 3G\\) MACs) in the object detection task." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.705, + 0.47, + 0.903 + ], + "angle": 0, + "content": "Semantic Segmentation. We further extend Castling-ViT to the semantic segmentation task to evaluate its effectiveness. Specifically, we use ViT-Base as the backbone in Mask2former [11] framework to serve as our baseline and testbed. Then we build the Castling-ViT-Base as the backbone to benchmark on the ADE20K dataset. As shown in Tab. 4, Mask2former with Castling-ViT backbone achieves \\(15\\%\\) total MACs reductions and \\(19\\%\\) backbone MACs reduction under comparable or slightly better mIoU, i.e., \\(\\uparrow 0.13\\%\\) and \\(\\uparrow 0.52\\%\\) without or with MAE pretraining on ImageNet [25]. This set of experiments validate that our proposed techniques could be well generalized to various downstream tasks that require large input resolutions." + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.089, + 0.892, + 0.132 + ], + "angle": 0, + "content": "Table 4. Castling-ViT over baselines on the ADE20K segmentation dataset, serving as backbones in the Mask2former framework. Note that we report both total MACs and backbone MACs, i.e., \\((\\cdot)\\)." + }, + { + "type": "table", + "bbox": [ + 0.504, + 0.133, + 0.892, + 0.201 + ], + "angle": 0, + "content": "
Mask2former w/MAE \nPretrainParams \n(M)MACs \n(G)mIoUmAccpAcc
ViT-BaseN118229 (182)34.5446.3675.84
Castling-ViT-BaseN118195 (147)34.6746.4776.20
ViT-BaseY118229 (182)47.9261.0083.02
Castling-ViT-BaseY118195 (147)48.4461.8283.29
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.21, + 0.892, + 0.294 + ], + "angle": 0, + "content": "Table 5. Ablation studies of Castling-ViT with various kernel-based linear attentions, where MC refers to mean-centering kernel function applied to \\(\\mathbf{Q} / \\mathbf{K}\\); Softmax also means the kernel function; ReLU-S and ReLU-E denote applying ReLU as kernels for decomposing softmax or exponential terms, respectively. All models are trained on COCO from scratch without pretraining." + }, + { + "type": "table", + "bbox": [ + 0.505, + 0.295, + 0.892, + 0.408 + ], + "angle": 0, + "content": "
BackboneLinear AttnS-320S-416M-416
mAPAP50mAPAP50mAPAP50
LCNetN/A25.539.029.744.234.349.9
LCNet-ViTN/A25.939.629.844.434.950.4
LCNet-ViTMC [17]22.834.926.339.631.546.1
Softmax [49]26.039.629.243.434.449.8
Cosine [5]24.437.427.241.032.948.0
ReLU-S23.936.7--30.945.3
ReLU-E [6]26.540.230.244.734.349.7
Lin.-Angular27.040.630.445.535.551.1
" + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.425, + 0.892, + 0.442 + ], + "angle": 0, + "content": "4.3. Linear-Angular Attention over SOTA Baselines" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.447, + 0.892, + 0.628 + ], + "angle": 0, + "content": "We also conduct ablation studies among various kinds of kernels used in linear attention to evaluate the superiority of our proposed linear-angular kernel. Also shown in Tab. 5, we compare the angular kernel with five other commonly adopted kernels, results on three model and resolution settings consistently demonstrate that our proposed angular kernel helps achieve better mAP, e.g., \\( \\uparrow 4.0\\% \\sim \\uparrow 4.2\\% \\) over MC [17], \\( \\uparrow 1.0\\% \\sim \\uparrow 1.2\\% \\) over Softmax [49], \\( \\uparrow 2.6\\% \\sim \\uparrow 3.2\\% \\) over Cosine [5], \\( \\uparrow 3.1\\% \\sim \\uparrow 4.6\\% \\) and \\( \\uparrow 0.2\\% \\sim \\uparrow 1.2\\% \\) over ReLU-S and ReLU-E [6], respectively. This set of experiments validate the superiority of the proposed linear-angular kernels." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.637, + 0.787, + 0.654 + ], + "angle": 0, + "content": "4.4. Ablation Studies of Castling-ViT" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.659, + 0.892, + 0.901 + ], + "angle": 0, + "content": "We conduct ablation studies on Castling-ViT's linear- angular attention, added DWConv, as well as auxiliary masked softmax-based attention, as shown in Tab. 6 and Tab. 7, where the experiments are performed on CastlingViT-S-320\\*/M-416\\*, respectively, without pretraining on ImageNet. Note that here \\* means that we use LCNet-ViT, i.e., replacing all layers in last two stages with transformer blocks, as backbones. Results in three tables show detailed performance breakdown and consistently demonstrate that all components in our proposed Castling-ViT contribute to the final performance. Specifically, linear-angular attention itself already achieves \\(\\downarrow 0.1\\% \\sim \\uparrow 0.5\\%\\) mAP improvements while leading to \\(10.1\\% \\sim 15.3\\%\\) MACs reductions simultaneously. Adding DWConv in the middle of MLP layers further leads to \\(\\downarrow 0.1\\% \\sim \\uparrow 0.6\\%\\) mAP improvements while only incur negligible MACs overhead. Adding aux" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "14437" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.089, + 0.472, + 0.146 + ], + "angle": 0, + "content": "Table 6. Ablation studies of Castling-ViT-S-320 w/ LCNet backbone, or LCNet-ViT backbone + linear-angular attention (denoted as Lin.) + DWConv (denoted as DW.) + auxiliary softmax-based attention (denoted as SparseAttn.)." + }, + { + "type": "table", + "bbox": [ + 0.081, + 0.147, + 0.47, + 0.235 + ], + "angle": 0, + "content": "
Castling-ViT-S-320 w/Params (M)MACs (G)mAPAP50AP75
LCNet1.100.4825.539.026.5
LCNet-ViT2.140.6925.939.626.9
+ Lin.2.140.6226.239.727.3
+ Lin.+DW.2.140.6226.840.728.0
+ Lin.+DW.+SparseAttn.2.140.6227.040.628.0
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.24, + 0.469, + 0.268 + ], + "angle": 0, + "content": "Table 7. Ablation studies of Castling-ViT-M-416 w/ LCNet backbone, or LCNet-ViT backbone + Lin. + DW. + SparseAttn." + }, + { + "type": "table", + "bbox": [ + 0.081, + 0.27, + 0.469, + 0.358 + ], + "angle": 0, + "content": "
Castling-ViT-M-416 w/Params (M)MACs (G)mAPAP50AP75
LCNet3.322.1534.349.936.5
LCNet-ViT7.533.5134.950.437.0
+ Lin.7.533.1534.850.336.7
+ Lin.+DW.7.533.1535.150.936.9
+ Lin.+DW.+SparseAttn.7.533.1535.351.137.3
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.373, + 0.47, + 0.554 + ], + "angle": 0, + "content": "illary masked softmax-based attention further increases the mAP by \\(0.2\\%\\) without incurring any overhead since it only assists the training while being removed during inference. Note that although LCNet-ViT backbone leads to better mAP, it also introduces more MACs as compared to efficient convolutions. As such, LCNet-ViT still slightly underperforms LCNet backbones in terms of accuracy-efficiency trade-offs, since we directly replace the CNN backbones with transformer blocks, whose architecture could not be optimal for ViTs. It remains to an open problem to build ViT based models that achieves higher accuracy-efficiency trade-off than pure ConvNet models." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.563, + 0.384, + 0.579 + ], + "angle": 0, + "content": "4.5. Discussion on the Auxiliary Branch" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.585, + 0.255, + 0.779 + ], + "angle": 0, + "content": "Trajectory of Nonzeros in Masks. To further understand the effect of the auxiliary branch, we visualize the trajectories of nonzeros in masks of auxiliary softmax-based attention. As shown in Fig. 6, we count the nonzeros in the first attention layer's masks throughout the training of both Castling-ViT-S-416" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.78, + 0.47, + 0.902 + ], + "angle": 0, + "content": "and Castling-ViT-M-416 with the LCNet-ViT backbone on COCO. we observe that the introduced auxiliary attention will only assist the training in the early or middle training stages and will gradually vanish towards all zeros in the later training stage, which is well aligned with the assumption in our Castling-ViT. This set of experiments validate the idea of performing \"castling\" in ViTs, i.e., drop the auxiliary branch without sacrificing performance." + }, + { + "type": "image", + "bbox": [ + 0.297, + 0.589, + 0.46, + 0.725 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.284, + 0.728, + 0.47, + 0.771 + ], + "angle": 0, + "content": "Figure 6. Visualizing the trajectories of nonzeros in auxiliary masks during training." + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.09, + 0.63, + 0.169 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.63, + 0.091, + 0.753, + 0.17 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.759, + 0.089, + 0.892, + 0.17 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.173, + 0.892, + 0.229 + ], + "angle": 0, + "content": "Figure 7. Left: Visualizing the taget similarity function, linear-angular terms, and DNNs' fitting curve; Middle: Visualizing the loss trajectory during DNN training; Right: Visualizing the learning trajectory of both low and high frequency components." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.236, + 0.892, + 0.568 + ], + "angle": 0, + "content": "Conjecture of Castling. We conduct a synthetic experiment to give an analogy for explaining the castling phenomenon. In Fig. 7 (Left), we visualize the curve of (1) angular similarity function (denoted in Red), (2) linear-angular terms (denoted in Green), and (3) two-layer DNN's approximation. We see that only keeping linear-angular terms leads to distortion, while DNN is capable of learning the missing high-frequency parts, whose loss trajectory is shown in Fig. 7 (Middle). Also, the learning trajectory of frequency components is visualized in Fig. 7 (Right), following F-Principle [67], where x-axis is training steps, y-axis categorizes both low and high frequency components, blue/red colors refer to large/small difference between the learned frequency components and target frequency components, respectively. We observe that DNNs fit target functions from low to high frequencies during training. It indicates that DWConv itself is not sufficient for approximating the high-order residual at early training stages. Costly attention that contains high-order residuals is then desired to help training while being dropped at inference since the remaining networks can gradually learn the high-frequency components at later training stages [67]." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.576, + 0.619, + 0.592 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.598, + 0.892, + 0.794 + ], + "angle": 0, + "content": "We present Castling-ViT that trains ViTs with both linear-angular and softmax-based quadratic attention but switches to only having the former during inference. Castling-ViT leverages angular kernels to measure the similarities between the queries and keys via spectral angles and highlights two enablers: (1) a new linear-angular attention mechanism: we decompose angular kernels to linear terms and high-order residuals, and keep only the former for inference; and (2) we approximate the high-order residuals using a depthwise convolution and an auxiliary masked softmax attention whose masks gradually become zeros during training without incurring inference overhead. Extensive experiments consistently validate Castling-ViT's advantages." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.803, + 0.652, + 0.82 + ], + "angle": 0, + "content": "Acknowledgment" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.826, + 0.892, + 0.902 + ], + "angle": 0, + "content": "The work is supported in part by the National Science Foundation (NSF) RTML program (Award number: 1937592) and the CoCoSys, one of the seven centers in JUMP 2.0, a Semiconductor Research Corporation (SRC) program sponsored by DARPA." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "14438" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.105 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.116, + 0.47, + 0.197 + ], + "angle": 0, + "content": "[1] Alaaeldin Ali, Hugo Touvron, Mathilde Caron, Piotr Bojanowski, Matthijs Douze, Armand Joulin, Ivan Laptev, Natalia Neverova, Gabriel Synnaeve, Jakob Verbeek, et al. Xcit: Cross-covariance image transformers. Advances in neural information processing systems, 34:20014-20027, 2021. 2, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.201, + 0.472, + 0.257 + ], + "angle": 0, + "content": "[2] Moab Arar, Ariel Shamir, and Amit H Bermano. Learned queries for efficient local attention. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10841-10852, 2022. 1, 2, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.259, + 0.471, + 0.3 + ], + "angle": 0, + "content": "[3] Iz Beltagy, Matthew E Peters, and Arman Cohan. Longformer: The long-document transformer. arXiv preprint arXiv:2004.05150, 2020. 1, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.302, + 0.471, + 0.357 + ], + "angle": 0, + "content": "[4] Daniel Bolya, Cheng-Yang Fu, Xiaoliang Dai, Peizhao Zhang, Christoph Feichtenhofer, and Judy Hoffman. Token merging: Your vit but faster. arXiv preprint arXiv:2210.09461, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.36, + 0.47, + 0.413 + ], + "angle": 0, + "content": "[5] Daniel Bolya, Cheng-Yang Fu, Xiaoliang Dai, Peizhao Zhang, and Judy Hoffman. Hydra attention: Efficient attention with many heads. arXiv preprint arXiv:2209.07484, 2022. 1, 2, 5, 7, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.417, + 0.47, + 0.471 + ], + "angle": 0, + "content": "[6] Han Cai, Chuang Gan, and Song Han. Efficientvit: Enhanced linear attention for high-resolution low-computation visual recognition. arXiv preprint arXiv:2205.14756, 2022. 1, 2, 3, 5, 7, 13, 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.474, + 0.47, + 0.543 + ], + "angle": 0, + "content": "[7] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In European conference on computer vision, pages 213-229. Springer, 2020. 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.546, + 0.47, + 0.6 + ], + "angle": 0, + "content": "[8] Beidi Chen, Tri Dao, Eric Winsor, Zhao Song, Atri Rudra, and Christopher Ré. Scatterbrain: Unifying sparse and low-rank attention. Advances in Neural Information Processing Systems, 34:17413-17426, 2021. 1, 2, 5, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.603, + 0.47, + 0.644 + ], + "angle": 0, + "content": "[9] Chun-Fu Chen, Quanfu Fan, and Rameswar Panda. Crossvit: Cross-attention multi-scale vision transformer for image classification. arXiv preprint arXiv:2103.14899, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.647, + 0.47, + 0.714 + ], + "angle": 0, + "content": "[10] Minghao Chen, Houwen Peng, Jianlong Fu, and Haibin Ling. Autoformer: Searching transformers for visual recognition. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12270-12280, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.717, + 0.47, + 0.799 + ], + "angle": 0, + "content": "[11] Bowen Cheng, Ishan Misra, Alexander G Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1290–1299, 2022. 3, 5, 6, 7, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.803, + 0.47, + 0.856 + ], + "angle": 0, + "content": "[12] Bowen Cheng, Alex Schwing, and Alexander Kirillov. Per-pixel classification is not all you need for semantic segmentation. Advances in Neural Information Processing Systems, 34:17864-17875, 2021. 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.859, + 0.47, + 0.901 + ], + "angle": 0, + "content": "[13] Krzysztof Choromanski, Valerii Likhosherstov, David Dohan, Xingyou Song, Andreea Gane, Tamas Sarlos, Peter Hawkins, Jared Davis, Afroz Mohiuddin, Lukasz Kaiser," + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.116, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "et al. Rethinking attention with performers. arXiv preprint arXiv:2009.14794, 2020.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.122, + 0.892, + 0.217 + ], + "angle": 0, + "content": "[14] Krzysztof Marcin Choromanski, Valerii Likhosherstov, David Dohan, Xingyou Song, Andreea Gane, Tamas Sarlos, Peter Hawkins, Jared Quincy Davis, Afroz Mohiuddin, Lukasz Kaiser, David Benjamin Belanger, Lucy J Colwell, and Adrian Weller. Rethinking attention with performers. In International Conference on Learning Representations, 2021. 1, 2, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.219, + 0.892, + 0.286 + ], + "angle": 0, + "content": "[15] Cheng Cui, Tingquan Gao, Shengyu Wei, Yuning Du, Ruoyu Guo, Shuilong Dong, Bin Lu, Ying Zhou, Xueying Lv, Qiwen Liu, et al. Pp-lcnet: A lightweight cpu convolutional neural network. arXiv preprint arXiv:2109.15099, 2021. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.288, + 0.892, + 0.356 + ], + "angle": 0, + "content": "[16] Xiaoliang Dai, Alvin Wan, P. Zhang, B. Wu, Zijian He, Zhen Wei, K. Chen, Yuandong Tian, Matthew E. Yu, Péter Vajda, and J. Gonzalez. Fbnetv3: Joint architecture-recipe search using neural acquisition function. ArXiv, abs/2006.02049, 2020.2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.358, + 0.892, + 0.44 + ], + "angle": 0, + "content": "[17] Jyotikrishna Dass, Shang Wu, Huihong Shi, Chaojian Li, Zhifan Ye, Zhongfeng Wang, and Yingyan Lin. Vitality: Unifying low-rank and sparse approximation for vision transformer acceleration with a linear taylor attention. In The 29th IEEE International Symposium on High-Performance Computer Architecture (HPCA 2023), 2023. 2, 5, 7, 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.442, + 0.892, + 0.496 + ], + "angle": 0, + "content": "[18] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 1, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.498, + 0.892, + 0.579 + ], + "angle": 0, + "content": "[19] Xiaoyi Dong, Jianmin Bao, Dongdong Chen, Weiming Zhang, Nenghai Yu, Lu Yuan, Dong Chen, and Baining Guo. Cswin transformer: A general vision transformer backbone with cross-shaped windows. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12124-12134, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.581, + 0.892, + 0.677 + ], + "angle": 0, + "content": "[20] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations, 2021. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.679, + 0.892, + 0.747 + ], + "angle": 0, + "content": "[21] Haoqi Fan, Bo Xiong, Karttikeya Mangalam, Yanghao Li, Zhicheng Yan, Jitendra Malik, and Christoph Feichtenhofer. Multiscale vision transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6824-6835, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.749, + 0.892, + 0.789 + ], + "angle": 0, + "content": "[22] Peng Gao, Teli Ma, Hongsheng Li, Jifeng Dai, and Yu Qiao. Convmae: Masked convolution meets masked autoencoders. arXiv preprint arXiv:2205.03892, 2022. 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.791, + 0.892, + 0.83 + ], + "angle": 0, + "content": "[23] Zheng Ge, Songtao Liu, Feng Wang, Zeming Li, and Jian Sun. Yolox: Exceeding yolo series in 2021. arXiv preprint arXiv:2107.08430, 2021. 6, 7, 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.832, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[24] Ben Graham, Alaaeldin El-Nouby, Hugo Touvron, Pierre Stock, Armand Joulin, Hervé Jégou, and Matthijs Douze. Levit: a vision transformer in convnet's clothing for faster inference. arXiv preprint arXiv:2104.01136, 2021. 2, 3, 5, 6, 13, 14" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "14439" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.161 + ], + "angle": 0, + "content": "[25] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollár, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16000-16009, 2022. 6, 7, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.162, + 0.471, + 0.218 + ], + "angle": 0, + "content": "[26] Byeongho Heo, Sangdoo Yun, Dongyoon Han, Sanghyuk Chun, Junsuk Choe, and Seong Joon Oh. Rethinking spatial dimensions of vision transformers. arXiv preprint arXiv:2103.16302, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.219, + 0.471, + 0.286 + ], + "angle": 0, + "content": "[27] Paul Honeine and Cedric Richard. The angular kernel in machine learning for hyperspectral data classification. In 2010 2nd Workshop on Hyperspectral Image and Signal Processing: Evolution in Remote Sensing, pages 1-4. IEEE, 2010. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.288, + 0.471, + 0.369 + ], + "angle": 0, + "content": "[28] Andrew Howard, Mark Sandler, Grace Chu, Liang-Chieh Chen, Bo Chen, Mingxing Tan, Weijun Wang, Yukun Zhu, Ruoming Pang, Vijay Vasudevan, et al. Searching for mobilenetv3. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1314-1324, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.371, + 0.471, + 0.44 + ], + "angle": 0, + "content": "[29] Angelos Katharopoulos, Apoorv Vyas, Nikolaos Pappas, and François Fleuret. Transformers are rnns: Fast autoregressive transformers with linear attention. In International Conference on Machine Learning, pages 5156-5165. PMLR, 2020. 1, 2, 3, 4, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.441, + 0.471, + 0.496 + ], + "angle": 0, + "content": "[30] Nirmal Keshava. Distance metrics and band selection in hyperspectral processing with applications to material identification and spectral libraries. IEEE Transactions on Geoscience and remote sensing, 42(7):1552-1565, 2004. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.497, + 0.471, + 0.566 + ], + "angle": 0, + "content": "[31] Kyungmin Kim, Bichen Wu, Xiaoliang Dai, Peizhao Zhang, Zhicheng Yan, Peter Vajda, and Seon Joo Kim. Rethinking the self-attention in vision transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3071–3075, 2021. 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.567, + 0.471, + 0.607 + ], + "angle": 0, + "content": "[32] Nikita Kitaev, Lukasz Kaiser, and Anselm Levskaya. Reformer: The efficient transformer. arXiv preprint arXiv:2001.04451, 2020. 1, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.608, + 0.471, + 0.65 + ], + "angle": 0, + "content": "[33] Yanghao Li, Hanzi Mao, Ross Girshick, and Kaiming He. Exploring plain vision transformer backbones for object detection. arXiv preprint arXiv:2203.16527, 2022. 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.651, + 0.471, + 0.734 + ], + "angle": 0, + "content": "[34] Yanghao Li, Chao-Yuan Wu, Haoqi Fan, Karttikeya Mangalam, Bo Xiong, Jitendra Malik, and Christoph Feichtenhofer. Mviv2: Improved multiscale vision transformers for classification and detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4804-4814, 2022. 3, 5, 6, 13, 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.735, + 0.471, + 0.789 + ], + "angle": 0, + "content": "[35] Yanyu Li, Geng Yuan, Yang Wen, Eric Hu, Georgios Evangelidis, Sergey Tulyakov, Yanzhi Wang, and Jian Ren. Efficientformer: Vision transformers at mobilenet speed. arXiv preprint arXiv:2206.01191, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.79, + 0.471, + 0.858 + ], + "angle": 0, + "content": "[36] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In European conference on computer vision, pages 740-755. Springer, 2014. 1, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.859, + 0.471, + 0.902 + ], + "angle": 0, + "content": "[37] Jing Liu, Zizheng Pan, Haoyu He, Jianfei Cai, and Bohan Zhuang. Ecoformer: Energy-saving attention with linear complexity. In NeurIPS, 2022. 1, 2, 4, 13" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.471, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.162 + ], + "angle": 0, + "content": "[38] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10012-10022, 2021. 1, 2, 6, 13, 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.163, + 0.892, + 0.232 + ], + "angle": 0, + "content": "[39] Jiachen Lu, Jinghan Yao, Junge Zhang, Xiatian Zhu, Hang Xu, Weiguo Gao, Chunjing Xu, Tao Xiang, and Li Zhang. Soft: softmax-free transformer with linear complexity. Advances in Neural Information Processing Systems, 34:21297-21309, 2021. 1, 2, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.234, + 0.892, + 0.276 + ], + "angle": 0, + "content": "[40] Sachin Mehta and Mohammad Rastegari. Mobilevit: lightweight, general-purpose, and mobile-friendly vision transformer. arXiv preprint arXiv:2110.02178, 2021. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.277, + 0.892, + 0.318 + ], + "angle": 0, + "content": "[41] John Moody and Christian Darken. _Learning with localized receptive fields_. Yale Univ., Department of Computer Science, 1988. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.32, + 0.892, + 0.375 + ], + "angle": 0, + "content": "[42]Zipeng Qin, Jianbo Liu, Xiaolin Zhang, Maoqing Tian, Aojun Zhou, Shuai Yi, and Hongsheng Li. Pyramid fusion transformer for semantic segmentation. arXiv preprint arXiv:2201.04019, 2022.3, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.377, + 0.892, + 0.445 + ], + "angle": 0, + "content": "[43] Zhen Qin, Weixuan Sun, Hui Deng, Dongxu Li, Yunshen Wei, Baohong Lv, Junjie Yan, Lingpeng Kong, and Yiran Zhong. cosformer: Rethinking softmax in attention. In International Conference on Learning Representations, 2022. 1, 2, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.447, + 0.892, + 0.516 + ], + "angle": 0, + "content": "[44] Ilija Radosavovic, Raj Prateek Kosaraju, Ross Girshick, Kaiming He, and Piotr Dólár. Designing network design spaces. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10428-10436, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.518, + 0.892, + 0.56 + ], + "angle": 0, + "content": "[45] Ali Rahimi and Benjamin Recht. Random features for large-scale kernel machines. Advances in neural information processing systems, 20, 2007. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.561, + 0.892, + 0.629 + ], + "angle": 0, + "content": "[46] Yongming Rao, Wenliang Zhao, Benlin Liu, Jiwen Lu, Jie Zhou, and Cho-Jui Hsieh. Dynamicvit: Efficient vision transformers with dynamic token sparsification. Advances in neural information processing systems, 34:13937-13949, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.632, + 0.892, + 0.702 + ], + "angle": 0, + "content": "[47] Mark Sandler, Andrew Howard, Menglong Zhu, Andrey Zh-moginov, and Liang-Chieh Chen. Mobilenetv2: Inverted residuals and linear bottlenecks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4510-4520, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.703, + 0.892, + 0.731 + ], + "angle": 0, + "content": "[48] Robert R. Schaller. Moore's law: past, present and future. IEEE spectrum, 1997. 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.732, + 0.892, + 0.8 + ], + "angle": 0, + "content": "[49] Zhuoran Shen, Mingyuan Zhang, Haiyu Zhao, Shuai Yi, and Hongsheng Li. Efficient attention: Attention with linear complexities. In Proceedings of the IEEE/CVF winter conference on applications of computer vision, pages 3531-3539, 2021. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.802, + 0.892, + 0.844 + ], + "angle": 0, + "content": "[50] Chen Sun, Abhinav Shrivastava, Saurabh Singh, and Abhinav Gupta. Revisiting unreasonable effectiveness of data in deep learning era. In ICCV, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.845, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[51] Mingxing Tan and Quoc Le. Efficientnet: Rethinking model scaling for convolutional neural networks. In International conference on machine learning, pages 6105-6114. PMLR, 2019. 6" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "14440" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.148 + ], + "angle": 0, + "content": "[52] Mingxing Tan, Ruoming Pang, and Quoc V Le. Efficientdet: Scalable and efficient object detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10781-10790, 2020. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.149, + 0.471, + 0.192 + ], + "angle": 0, + "content": "[53] Shitao Tang, Jiahui Zhang, Siyu Zhu, and Ping Tan. Quadtree attention for vision transformers. In International Conference on Learning Representations, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.194, + 0.471, + 0.289 + ], + "angle": 0, + "content": "[54] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Herve Jegou. Training data-efficient image transformers & distillation through attention. In Proceedings of the 38th International Conference on Machine Learning, volume 139 of Proceedings of Machine Learning Research, pages 10347-10357. PMLR, 18-24 Jul 2021. 2, 5, 6, 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.291, + 0.471, + 0.359 + ], + "angle": 0, + "content": "[55] Hugo Touvron, Matthieu Cord, Alexandre Sablayrolles, Gabriel Synnaeve, and Hervé Jégou. Going deeper with image transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 32-42, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.362, + 0.47, + 0.446 + ], + "angle": 0, + "content": "[56] Zhengzhong Tu, Hossein Talebi, Han Zhang, Feng Yang, Peyman Milanfar, Alan Bovik, and Yinxiao Li. Maxvit: Multi-axis vision transformer. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXIV, pages 459-479. Springer, 2022. 2, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.447, + 0.47, + 0.516 + ], + "angle": 0, + "content": "[57] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Ilia Polosukhin. Attention is all you need. In Advances in Neural Information Processing Systems, volume 30. Curran Associates, Inc., 2017. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.518, + 0.47, + 0.559 + ], + "angle": 0, + "content": "[58] Sinong Wang, Belinda Z. Li, Madian Khabsa, Han Fang, and Hao Ma. Linformer: Self-attention with linear complexity, 2020. 1, 2, 3, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.562, + 0.47, + 0.631 + ], + "angle": 0, + "content": "[59] Wenhai Wang, Enze Xie, Xiang Li, Deng-Ping Fan, Kaitao Song, Ding Liang, Tong Lu, Ping Luo, and Ling Shao. Pyramid vision transformer: A versatile backbone for dense prediction without convolutions. arXiv preprint arXiv:2102.12122, 2021. 2, 6, 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.633, + 0.47, + 0.716 + ], + "angle": 0, + "content": "[60] Wenhai Wang, Enze Xie, Xiang Li, Deng-Ping Fan, Kaitao Song, Ding Liang, Tong Lu, Ping Luo, and Ling Shao. Pyramid vision transformer: A versatile backbone for dense prediction without convolutions. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 568-578, October 2021. 3, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.718, + 0.47, + 0.773 + ], + "angle": 0, + "content": "[61] Wenhai Wang, Enze Xie, Xiang Li, Deng-Ping Fan, Kaitao Song, Ding Liang, Tong Lu, Ping Luo, and Ling Shao. Pvt v2: Improved baselines with pyramid vision transformer. Computational Visual Media, 8(3):415-424, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.775, + 0.47, + 0.843 + ], + "angle": 0, + "content": "[62] Bichen Wu, Chaojian Li, Hang Zhang, Xiaoliang Dai, Peizhao Zhang, Matthew Yu, Jialiang Wang, Yingyan Lin, and Peter Vajda. Fbnetv5: Neural architecture search for multiple tasks in one run. arXiv preprint arXiv:2111.10007, 2021.6,7" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.845, + 0.47, + 0.901 + ], + "angle": 0, + "content": "[63] Haiping Wu, Bin Xiao, Noel Codella, Mengchen Liu, Xiyang Dai, Lu Yuan, and Lei Zhang. Cvt: Introducing convolutions to vision transformers. arXiv preprint arXiv:2103.15808, 2021. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.471, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.146 + ], + "angle": 0, + "content": "[64] Zhanghao Wu, Zhijian Liu, Ji Lin, Yujun Lin, and Song Han. Lite transformer with long-short range attention. In International Conference on Learning Representations, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.149, + 0.892, + 0.218 + ], + "angle": 0, + "content": "[65] Yunyang Xiong, Hanxiao Liu, Suyog Gupta, Berkin Akin, Gabriel Bender, Pieter-Jan Kindermans, Mingxing Tan, Vikas Singh, and Bo Chen. Mobiledets: Searching for object detection architectures for mobile accelerators. arXiv preprint arXiv:2004.14525, 2020. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.219, + 0.892, + 0.3 + ], + "angle": 0, + "content": "[66] Yunyang Xiong, Zhanpeng Zeng, Rudrasis Chakraborty, Mingxing Tan, Glenn Fung, Yin Li, and Vikas Singh. Nystromformer: A nystrom-based algorithm for approximating self-attention. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 35, pages 14138–14148, 2021. 1, 2, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.302, + 0.892, + 0.357 + ], + "angle": 0, + "content": "[67] Zhi-Qin John Xu, Yaoyu Zhang, Tao Luo, Yanyang Xiao, and Zheng Ma. Frequency principle: Fourier analysis sheds light on deep neural networks. arXiv preprint arXiv:1901.06523, 2019. 2, 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.358, + 0.892, + 0.426 + ], + "angle": 0, + "content": "[68] Hongxu Yin, Arash Vahdat, Jose M Alvarez, Arun Mallya, Jan Kautz, and Pavlo Molchanov. A-vit: Adaptive tokens for efficient vision transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10809-10818, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.428, + 0.892, + 0.51 + ], + "angle": 0, + "content": "[69] Haoran You, Zhanyi Sun, Huihong Shi, Zhongzhi Yu, Yang Zhao, Yongan Zhang, Chaojian Li, Baopu Li, and Yingyan Lin. Vitcod: Vision transformer acceleration via dedicated algorithm and accelerator co-design. In The 29th IEEE International Symposium on High-Performance Computer Architecture (HPCA 2023), 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.511, + 0.892, + 0.58 + ], + "angle": 0, + "content": "[70] Guanghua Yu, Qinyao Chang, Wenyu Lv, Chang Xu, Cheng Cui, Wei Ji, Qingqing Dang, Kaipeng Deng, Guanzhong Wang, Yuning Du, et al. Pp-picodet: A better real-time object detector on mobile devices. arXiv preprint arXiv:2111.00902, 2021. 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.581, + 0.892, + 0.65 + ], + "angle": 0, + "content": "[71] Zhongzhi Yu, Yonggan Fu, Sicheng Li, Chaojian Li, and Yingyan Lin. Mia-former: Efficient and robust vision transformers via multi-grained input-adaptation. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 36, pages 8962-8970, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.651, + 0.892, + 0.732 + ], + "angle": 0, + "content": "[72] Li Yuan, Yunpeng Chen, Tao Wang, Weihao Yu, Yujun Shi, Zi-Hang Jiang, Francis EH Tay, Jiashi Feng, and Shuicheng Yan. Tokens-to-token vit: Training vision transformers from scratch onImagenet. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 558-567, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.734, + 0.892, + 0.802 + ], + "angle": 0, + "content": "[73] Yuhui Yuan, Rao Fu, Lang Huang, Weihong Lin, Chao Zhang, Xilin Chen, and Jingdong Wang. Hrformer: High-resolution vision transformer for dense predict. Advances in Neural Information Processing Systems, 34:7281-7293, 2021.6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.804, + 0.892, + 0.872 + ], + "angle": 0, + "content": "[74] Zixiao Zhang, Xiaogiang Lu, Guojin Cao, Yuting Yang, Licheng Jiao, and Fang Liu. Vit-yolo: Transformer-based yolo for object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2799-2808, 2021. 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.873, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[75] Bolei Zhou, Hang Zhao, Xavier Puig, Sanja Fidler, Adela Barriuso, and Antonio Torralba. Scene parsing through" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.518, + 0.957 + ], + "angle": 0, + "content": "14441" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.092, + 0.468, + 0.133 + ], + "angle": 0, + "content": "ade20k dataset. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 633-641, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.136, + 0.468, + 0.205 + ], + "angle": 0, + "content": "[76] Chen Zhu, Wei Ping, Chaowei Xiao, Mohammad Shoeybi, Tom Goldstein, Anima Anandkumar, and Bryan Catanzaro. Long-short transformer: Efficient transformers for language and vision. Advances in Neural Information Processing Systems, 34, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.207, + 0.468, + 0.261 + ], + "angle": 0, + "content": "[77] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable detr: Deformable transformers for end-to-end object detection. arXiv preprint arXiv:2010.04159, 2020. 3, 13" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.468, + 0.261 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.956 + ], + "angle": 0, + "content": "14442" + } + ] +] \ No newline at end of file diff --git a/2023/Castling-ViT_ Compressing Self-Attention via Switching Towards Linear-Angular Attention at Vision Transformer Inference/d74c49bb-effe-4caf-8858-0066e120e08d_origin.pdf b/2023/Castling-ViT_ Compressing Self-Attention via Switching Towards Linear-Angular Attention at Vision Transformer Inference/d74c49bb-effe-4caf-8858-0066e120e08d_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..e47fa6cb9fb5f067915749dfdbf8ea27cd09d67b --- /dev/null +++ b/2023/Castling-ViT_ Compressing Self-Attention via Switching Towards Linear-Angular Attention at Vision Transformer Inference/d74c49bb-effe-4caf-8858-0066e120e08d_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4a097d38ff5bd13f441fa3c3aea5bd77dc6c634b2493eab23b6c3203445e3f19 +size 1560811 diff --git a/2023/Castling-ViT_ Compressing Self-Attention via Switching Towards Linear-Angular Attention at Vision Transformer Inference/full.md b/2023/Castling-ViT_ Compressing Self-Attention via Switching Towards Linear-Angular Attention at Vision Transformer Inference/full.md new file mode 100644 index 0000000000000000000000000000000000000000..ce0bf7142afad2bc52e33c4f24825530f7f3deef --- /dev/null +++ b/2023/Castling-ViT_ Compressing Self-Attention via Switching Towards Linear-Angular Attention at Vision Transformer Inference/full.md @@ -0,0 +1,338 @@ +# Castling-ViT: Compressing Self-Attention via Switching Towards Linear-Angular Attention at Vision Transformer Inference + +Haoran You $^{1,2,\dagger,*}$ , Yunyang Xiong $^{2,*}$ , Xiaoliang Dai $^{2}$ , Bichen Wu $^{2}$ , Peizhao Zhang $^{2}$ , Haoqi Fan $^{2}$ , Peter Vajda $^{2}$ , Yingyan (Celine) Lin $^{1}$ $^{1}$ Georgia Institute of Technology, $^{2}$ Meta Research +{haoran.you, celine.lin}@gatech.edu +{yunyang, xiaoliang dai, wbc, stzpz, haoqifan, vajdap}@meta.com + +# Abstract + +Vision Transformers (ViTs) have shown impressive performance but still require a high computation cost as compared to convolutional neural networks (CNNs), one reason is that ViTs' attention measures global similarities and thus has a quadratic complexity with the number of input tokens. Existing efficient ViTs adopt local attention or linear attention, which sacrifice ViTs' capabilities of capturing either global or local context. In this work, we ask an important research question: Can ViTs learn both global and local context while being more efficient during inference? To this end, we propose a framework called Castling-ViT, which trains ViTs using both linear-angular attention and masked softmax-based quadratic attention, but then switches to having only linear-angular attention during inference. Our Castling-ViT leverages angular kernels to measure the similarities between queries and keys via spectral angles. And we further simplify it with two techniques: (1) a novel linear-angular attention mechanism: we decompose the angular kernels into linear terms and high-order residuals, and only keep the linear terms; and (2) we adopt two parameterized modules to approximate high-order residuals: a depthwise convolution and an auxiliary masked softmax attention to help learn global and local information, where the masks for softmax attention are regularized to gradually become zeros and thus incur no overhead during inference. Extensive experiments validate the effectiveness of our Castling-ViT, e.g., achieving up to a $1.8\%$ higher accuracy or $40\%$ MACs reduction on classification and 1.2 higher mAP on detection under comparable FLOPs, as compared to ViTs with vanilla softmax-based attentions. Project page is available at here. + +# 1. Introduction + +Vision Transformers (ViTs) have made significant progress in image classification, object detection, and many + +![](images/07c00801ece4ca556f08c973c3b493a5ed8898b8012feb40e2288e0cba1fd932.jpg) + +![](images/d17453d5505c6838894302c6cb13f95985a07f7942fc59096ef8ab3348323632.jpg) +Figure 1. Castling-ViT over SOTA baselines on (1) ImageNet [18] image classification and (2) COCO [36] object detection. + +other applications. It is well recognized that the superior performance achieved by ViTs is largely attributed to their self-attention modules that can better capture global context [20, 57, 64]. Nevertheless, ViTs' powerful self-attention module comes at the cost of quadratic complexity with the number of input tokens, causing a major efficiency bottleneck to ViTs' achievable runtime (i.e., inference latency) [3, 8, 32, 58, 66, 69, 76]. To mitigate this issue, linear attention designs have been developed to alleviate the vanilla ViT attention's quadratic complexity. In particular, existing efforts can be categorized into two clusters: (1) ViTs with local attention by restricting the attention window size [38, 53], sharing the attention queries [2], or representing the attention queries/keys with low rank matrices [58]; and (2) ViTs with kernel-based linear attention, which approximate the non-linearity softmax function by decomposing it into separate kernel embeddings. This enables a change in the matrix computation order for a reduced computational complexity [5, 6, 14, 29, 37, 39, 43, 66]. + +Despite their promise in alleviating ViTs' complexity + +and thus inference runtime, both the local and linear attention compromise ViTs' performance due to the lack of capabilities to capture global or local context. To marry the best of both worlds, we advocate training ViTs with both (1) efficient but less powerful linear attention, i.e., without the high-order residuals in angular kernel expansion, and (2) powerful yet costly softmax-based masked attention. The latter helps approximate high-order residuals at the early training stage while being dropped during inference, based on an assumption that the remaining networks can gradually learn the high-order components at the later training stage [67]. This concept resembles the "castling" move in chess when two pieces are moved at once. While it sounds promising, there are still two challenges to achieve this. First, existing linear attention modules still underperform their vanilla softmax-based counterparts. Therefore, a better linear attention is crucial for the final performance. We find that angular kernels perform equally as softmax-based attentions in terms of similarity measurements. While they still suffer from a quadratic complexity, they can be divided into linear terms and high-order residuals. The challenge is how to construct ViTs with only the linear terms. Second, doing so would require that the trained ViTs merely rely on the linear terms towards the end of training, which would call for an approximation of the above high-order residuals. The challenge is that how we can resort to costly but powerful modules to approximate high-order residuals during training but does not incur extra inference cost. + +In this work, we develop techniques to tackle those challenges, and make the following contributions: + +- We propose a framework called Castling-ViT, which trains ViTs using both linear-angular attention and masked softmax-based quadratic attention, but then switches to having only linear-angular attentions during ViT inference to save computational costs. +- We develop a new linear-angular attention leveraging angular kernels to close the accuracy gap between linear attention and softmax-based attention. It expands angular kernels where linear terms are kept while complex high-order residuals are approximated. +- We use two parameterized modules to approximate the high-order residuals above: a depthwise convolution and an auxiliary masked softmax-based attention, where the latter's attention masks are regularized to gradually become zeros to avoid inference overhead. +- We conduct extensive experiments to validate the effectiveness of the proposed Castling-ViT. Results on classification, detection, and segmentation tasks consistently demonstrate its superior performance ( $\uparrow 1.8\%$ top-1 accuracy or $\uparrow 1.2$ mAP) or efficiency (40% MACs savings) over state-of-the-art (SOTA) CNNs and ViTs. + +# 2. Related Works + +Vision Transformers (ViTs). ViT [20] beats CNNs with a simple encoder-only transformer architecture taking the splitted non-overlapped image patches as sequential inputs, but relies on costly pretraining on a huge JFT-300M dataset [50]. Later, DeiT [54] and T2T-ViT [72] leverage an improved ViT training recipe or enhanced tokenization mechanism to achieve a comparable accuracy without the necessity of costly pretraining. To further improve ViTs' achievable accuracy-efficiency trade-offs, CrossViT [9], PiT [26], PVT [59], MViT [21] and Swin-Transformer [38] propose a pyramid-like architecture, which is commonly used in CNNs [16, 28, 65]; DynamicViT [46], A-ViT [68], ToME [4], and MIA-Former [71] propose to adaptively identify and remove unnecessary input tokens for saving computational costs. With the goal of deploying ViTs in resource-constrained devices, various efficient ViT architectures have been proposed [6, 24, 35, 40]. For example, LeViT [24], CvT [63], and MobileViT [40] adopt more efficient self-attention implementation or incorporate convolutional feature extraction blocks into their early layers; EfficientFormer [35] further enables pure ViTs to run as fast as MobileNets. In contrast, our Castling-ViT explores whether ViTs can learn both global and local features while still being efficient at runtime. Also, we target a generic linear-angular attention that can serve as a drop-in replacement for all kinds of ViT architectures and thus is orthogonal to new ViT architecture designs. + +Efficient ViT Variants. As commonly recognized, ViTs rely heavily on their self-attention module which is however costly due to its quadratic computational complexity with the total number of input tokens [29, 76]. To make the self-attention module more efficient, a surge of linear attention works have been proposed and can be roughly categorized into two groups: local attention [2, 38, 56, 58] or kernel-based linear attention [1, 5, 6, 14, 29, 37, 39, 43, 66]. For kernel-based linear attention, common designs approximate the softmax function [5, 14, 29] or the full self-attention matrix [39, 66] with orthogonal features or kernel embeddings, then the computation order can be changed from (QK)V to Q(KV). For example, [29] and [6] decompose the exponential terms in softmax-based attention into kernel functions and exchange the computation order. Despite their decent performance, currently kernel-based linear attention in general underperform the softmax-based attention. Recent works [8, 17] also unify low rank approximated and sparse attention (can also be dropped at inference) to improve ViTs' accuracy-efficiency tradeoffs. Different from the above works, we explore from a new perspective by taking spectral angles into consideration when measuring the similarities among tokens, resulting in linear-angular attention that can achieve comparable or even better performance than softmax-based attention. More efficient ViT variants + +are supplied to the Appendix. + +ViTs have also been used as the backbones for downstream tasks, e.g., detection [34, 60, 77] and segmentation [11,42]. We supply more literature review to Appendix. + +# 3. The Proposed Methods + +# 3.1. Preliminary of Self-Attention + +Self-Attention. Self-attention module is a core component of the Transformer [20, 57], and usually consists of multiple heads. Each head captures global-context information by measuring pairwise correlations among all $N$ tokens ( $N$ denotes the total number of tokens) as defined below: + +$$ +\mathbf {H} _ {t} ^ {k} = \sum_ {i = 1} ^ {N} \frac {\exp \left(\mathbf {Q} _ {t} \cdot \mathbf {K} _ {i} ^ {T} / \sqrt {d}\right)}{\sum_ {j = 1} ^ {N} \exp \left(\mathbf {Q} _ {t} \cdot \mathbf {K} _ {j} ^ {T} / \sqrt {d}\right)} \cdot \mathbf {V} _ {i}, \tag {1} +$$ + +where $t\in \{1,\dots ,N\} ,k\in \{1,\dots ,M\}$ $\mathbf{H}_t^k$ refers to the $t$ th row of the $m$ -th head's attention matrix H. $\mathbf{Q}_t,\mathbf{K}_t,\mathbf{V}_t\in$ $\mathbf{R}^d$ , are the query, key, and value vectors obtained by linearly projecting the input $\mathbf{X_t}\in \mathbb{R}^D$ with three learnable weight matrices $W^{Q},W^{K},W^{V}\in \mathbb{R}^{D\times d}$ . The attention head first computes the inner product between the query-key pairs, then scales the product results to stabilize the training and uses Softmax to normalize the resulting attention scores, and finally computes a weighted sum over all value vectors. Outputs from all attention heads are then concatenated together before a final linear projection with learnable weights. Note that $\exp (\cdot)$ denotes an exponential function. Computing Eq. (1) has a quadratic complexity of $\mathcal{O}(N^2)$ + +Kernel-based Linear Attention. The core idea of linear attention [29, 49, 58] is to decompose the similarity measurement function into separate kernel embeddings, i.e., $\mathrm{Sim}(\mathbf{Q},\mathbf{K})\approx \phi (\mathbf{Q})\phi (\mathbf{K})^T$ so that we can change the computation order to $\phi (\mathbf{Q})(\phi (\mathbf{K})^T\mathbf{V})$ based on the associative property of matrix multiplication. In this way, the attention complexity is quadratic to the feature dimension $d$ instead of the token length $N$ . One straightforward implementation of linear attention is to use Gaussian RBF kernels to measure the similarity, which can serve as an unbiased estimation of $\exp (\langle \cdot ,\cdot \rangle)$ in Eq. (1): + +$$ +\exp \left(\frac {\langle \mathbf {x} , \mathbf {y} \rangle}{\sigma^ {2}}\right) = \underbrace {\exp \left(\frac {- \| \mathbf {x} - \mathbf {y} \| ^ {2}}{2 \sigma^ {2}}\right)} _ {\text {G a u s s i a n R B F K e r n e l}} \exp \left(\frac {\| \mathbf {x} \| ^ {2} + \| \mathbf {y} \| ^ {2}}{2 \sigma^ {2}}\right) \tag {2} +$$ + +where $\langle \cdot, \cdot \rangle$ denotes the inner product operator. According to [45], we can induce a function $\phi(\cdot)$ to approximate the Gaussian RBF kernel, mapping the input space to the feature space. Assuming that both $\mathbf{Q}$ and $\mathbf{K}$ are normalized as unit row vectors along the feature dimension, then the attention formula can be approximated by: + +$$ +\mathbf {H} _ {t} ^ {k} = \frac {\phi (\mathbf {Q} _ {t}) \cdot \sum_ {i = 1} ^ {N} \left(\phi \left(\mathbf {K} _ {i}\right) ^ {T} \cdot \mathbf {V} _ {i}\right)}{\phi (\mathbf {Q} _ {t}) \cdot \sum_ {j = 1} ^ {N} \phi \left(\mathbf {K} _ {j} ^ {T}\right)} \tag {3} +$$ + +![](images/27ef88097b699bd0084f362693ea21f7f8ebf624aff2d6eee1f482d95dfa630f.jpg) +Figure 2. Illustration of self-attention used in LeViT and MViTv2. + +While doing so can alleviate the attention complexity to become linear w.r.t. the token length $N$ , it often causes a nontrivial accuracy drop as compared to the corresponding ViTs with vanilla softmax-based attention [6, 13, 72]. + +Revisit Attention Designs. Recent ViTs have achieved a low complexity for classification tasks (e.g., LeViT [24]) but still be costly for downstream tasks due to high input resolutions, i.e., the number of tokens (e.g., MViTv2 [34]). Therefore, one dilemma is the tradeoff between the model efficiency and generalizability. As illustrated in Fig. 2, LeViT shrinks the feature dimension to be more efficient on low-resolution tasks but cannot be well generalized to downstream tasks, due to the resulting (1) feature bottlenecks, i.e., insufficient feature dimensions, and (2) fixed attention biases. On the other hand, MViTv2 performs token pooling for better fitting high-resolution tasks but is less effective for low-resolution due to the caused token bottlenecks, i.e., insufficient number of tokens. Our ablation studies in Table 1 show that (1) for ImageNet classification, both token and feature pooling can lead to accuracy drops in ViTs; and (2) for downsampling layers, post-Q pooling (i.e., pooling after linear projection) together with residual connections in MViTv2 performs better than pre-Q pooling (i.e., pooling during linear projection) in LeViT. Therefore, we develop Castling-LeViTs on top of attention with merely post-Q pooling and residual connections. + +# 3.2. The Proposed Castling-ViT Framework + +Castling-ViT Overview. Fig. 4 illustrates an overview of the proposed Castling-ViT, which makes linear attention more powerful than previous designs while still being efficient during inference. In particular, we propose (1) a novel kernel-based linear-angular attention from the spectral angle perspective to close the accuracy gap between linear attention and softmax-based attention; and (2) a training augmentation method that leverages softmax-based attention as an auxiliary branch to assist the linear-angular atten + +Table 1. Analyze the attention design on ImageNet classification. + +
PoolingResidual QMACs (M)Top-1 Accuracy (%)
TokenFeat.Pre-QPost-Q
✓✓68578.00
✓✓66178.19
✓73679.11
✓✓✓77178.73
✓✓✓76377.23
✓✓83880.05
+ +tion only during ViT training. Note that a mask is applied to the auxiliary branch to manifest linear-angular attention and drop the auxiliary branch. + +# 3.2.1 Linear-Angular Attention + +Angular Kernel. In addition to the previously adopted polynomial, exponential, or RBF kernel that focuses on spatial similarity measurements [29, 37], we propose to consider measuring spectral similarity via angular kernel as an alternative to existing softmax-based attentions, leading to similar or better performance since it additionally takes into consideration the nature of spectral characteristics, e.g., the spectral angle as a distance measurement function [27, 30]. Such a spectral angle between two vectors is defined as: + +$$ +\theta \left(\mathbf {x} _ {i}, \mathbf {x} _ {j}\right) = \arccos \left(\frac {\left\langle \mathbf {x} _ {i} , \mathbf {x} _ {j} \right\rangle}{\left\| \mathbf {x} _ {i} \right\| \cdot \left\| \mathbf {x} _ {j} \right\|}\right), \tag {4} +$$ + +where $\|\cdot\|$ is the Euclidean distance and $\langle \cdot, \cdot \rangle$ is the inner product. The output range of $\theta$ is $[0, \pi]$ . Such an angle can be used as a distance. In our design, we define the angular kernel as a similarity measurement function between the queries $\mathbf{Q}$ and keys $\mathbf{K}$ as: + +$$ +\operatorname {S i m} \left(\mathbf {Q} _ {i}, \mathbf {K} _ {j}\right) = 1 - \frac {1}{\pi} \cdot \theta \left(\mathbf {Q} _ {i}, \mathbf {K} _ {j}\right), \tag {5} +$$ + +and the output range is thus [0, 1]. With more aligned $\mathbf{Q}_i$ and $\mathbf{K}_j$ , $\theta$ is closer to 0 and thus the similarity is closer to 1; In contrast, if $\mathbf{Q}_i$ and $\mathbf{K}_j$ have opposite features, $\theta$ is closer to $\pi$ and thus the similarity is closer to 0. + +Properties of Angular Kernel and Its Feature Space. One property of our angular kernel is that replacing the similarity measurement in self-attention with such a kernel provides an efficient way to implicitly map the input data to a high (even infinite after expansion) dimensional feature space [27], where the distances/angles are calculated based on a rich feature structure. Let $\phi(\cdot)$ denotes the implicit map induced by this kernel, the norm of mapped input data is: + +$$ +\left\| \phi \left(\mathbf {x} _ {i}\right) \right\| ^ {2} = \operatorname {S i m} \left(\mathbf {x} _ {i}, \mathbf {x} _ {i}\right) = 1 - \frac {1}{\pi} \cdot 0 = 1, \tag {6} +$$ + +which means that all data in the input space are mapped onto the sphere of radius 1 in the feature space. Also, the distance between two input features is given by: + +$$ +\begin{array}{l} \left\| \phi (\mathbf {x} _ {i}) - \phi (\mathbf {x} _ {j}) \right\| ^ {2} = 1 + 1 - 2 \cdot \phi (\mathbf {x} _ {i}) \phi (\mathbf {x} _ {j}) ^ {T} (7) \\ = 2 \cdot \left(1 - \operatorname {S i m} \left(\mathbf {x} _ {i}, \mathbf {x} _ {j}\right)\right) = \frac {2}{\pi} \cdot \theta \left(\mathbf {x} _ {i}, \mathbf {x} _ {j}\right). (7) \\ \end{array} +$$ + +![](images/fae8d3265027027362ce0f59815be2c815433d2e751bb7501aaa5a537dd1f64b.jpg) +(a) Input space + +![](images/8dde8b37267758b9431238ed8673c7d607d39af714e6d08bb0c4afbaf7b1e2ca.jpg) +(b) Feature space +Figure 3. (a) A simple 2D input space and (b) the corresponding 3D feature space after applying the angular kernel to the 2D input data. Note that real input data/features are of higher dimensions. + +That is, the square (Euclid) distance and the spectral angle is positively correlated and the distance range is $[0, 2]$ . As shown in Fig. 3, the angles in the input space are correlated to the feature distance after applying our angular kernel. + +Expansion of Angular Kernel. A natural following question is how to incorporate the above angular kernel for designing linear attention given its quadratic complexity w.r.t. the input token length. Recalling from trigonometric identities and the expansion of the arccos function into an infinite series, we reformulate the similarity function as: + +$$ +\begin{array}{l} \operatorname {S i m} \left(\mathbf {Q} _ {i}, \mathbf {K} _ {j}\right) = 1 - \frac {1}{\pi} \cdot \left(\frac {\pi}{2} - \arcsin \left(\mathbf {Q} _ {i} \cdot \mathbf {K} _ {j} ^ {T}\right)\right) \\ = \underbrace {\frac {1}{2} + \frac {1}{\pi} \cdot \left(\mathbf {Q} _ {i} \cdot \mathbf {K} _ {j} ^ {T}\right)} _ {\text {L i n e a r - a n g u l a r t e r m s}} \tag {8} \\ + \underbrace {\frac {1}{\pi} \cdot \sum_ {k = 1} ^ {\infty} \frac {(2 k) !}{2 ^ {2 k} (k !) ^ {2} (2 k + 1)} \left(\mathbf {Q} _ {i} \cdot \mathbf {K} _ {j} ^ {T}\right) ^ {2 k + 1}} _ {\text {H i g h - o r d e r r e s i d u a l s w h e n k \geq 1 (n o n - l i n e a r)}}, \\ \end{array} +$$ + +where $\mathbf{Q}_i\cdot \mathbf{K}_j^T$ denotes the normalized linear kernel function $\langle \mathbf{Q}_i,\mathbf{K}_j\rangle /\| \mathbf{Q}_i\| \cdot \| \mathbf{K}_j\|$ , which is equivalent to the inner product if $\mathbf{Q}_i$ and $\mathbf{K}_j$ are unit vectors. We see that the first linear-angular terms can be directly used as the similarity measurement in linear attention, while the remaining higher-order terms of an infinite series introduce a much higher complexity. As such, we propose to adopt a relaxation to approximate it. + +Linear-Angular Attention. To construction our linear- angular attention, we leverage parametrized DNN modules to approximate the expectation of the high-order residuals, i.e., $\mathbf{M}_{ij} = \mathbb{E}_{\mathbf{Q}_i\sim p_{Q_i},\mathbf{K}_j\sim p_{K_j}}\left(\sum_{k = 1}^{\infty}\alpha_k(\mathbf{Q}_i\cdot \mathbf{K}_j^T)^{2k + 1}\right)$ where $\alpha_{k}$ is the coefficient for the $k$ -th order term in Eq. (8). The key question is how to design such DNN modules. Kim et al. [31] conducted an analysis on average attention weights of ViT models, which show a strong inductive bias to attend to neighboring tokens. To capture this, we introduce a learnable depthwise convolution (DWConv) module which is applied on all the value tokens, so $\mathbf{M}_{ij}$ by-design attends to nearby tokens, as illustrated in Fig. 4. Second, DWConv is limited by its receptive field [41] while the averaged attention map also has a small number of offdiagonal scores to connect to nonadjacent tokens [31]. To capture that, we adopt a sparse softmax-based attention as + +![](images/ec02e87fdc198761e8fc24ad61e5ff865315daa98184d7b07602f7e5ef3e52ea.jpg) +Figure 4. The visualization of our Castling-ViT with both linear-angular attention and auxiliary softmax-based self-attention. + +described in Section 3.2.2. In practice, we found that such a sparse softmax attention tends to converge to all zeros after adding a simple threshold regularization. We provide more explanation and visualization in Sec. 3.2.2 and Sec. 4.5. + +In this way, our attention module can be formulated as: + +$$ +\begin{array}{l} \mathbf {H} = \operatorname {S i m} (\mathbf {Q}, \mathbf {K}) \cdot \mathbf {V} \approx \frac {1}{2} \cdot \mathbf {V} + \frac {1}{\pi} \cdot \mathbf {Q} \cdot \left(\mathbf {K} ^ {T} \cdot \mathbf {V}\right) \tag {9} \\ + \mathbf {M} _ {D W} \cdot \mathbf {V} + \mathbf {M} _ {S p a r s e A t t n}, \\ \end{array} +$$ + +where $1 / \pi \cdot Q\cdot (K^T\cdot V)$ is the linear term with $\mathcal{O}(N)$ complexity, $\mathbf{M}_{DW}\cdot V$ is the matrix form of DWConv, and $\mathbf{M}_{SparseAttn}$ is the normalized sparse softmax attention. The overall complexity to compute Eq. (9) is linear to the input token length, where the MACs of DWConv is also negligible (e.g., $< 1\%$ of the total MACs). Also, normalizations $(\mathrm{x} / \| \mathrm{x}\|_{2})$ are inserted to the $\mathbf{Q} / \mathbf{K}$ and sparse attention branches to help the similarity measurement following [5]. + +# 3.2.2 Switch Towards Linear-Angular Attention + +Recall that we add a sparse softmax-based attention as an auxiliary branch in the linear-angular attention to help approximate the high-order residuals. Such a costly attention can be potentially dropped without hurting the inference accuracy, drawing inspiration from [67] that the remaining network can gradually learn the high-order/frequency components at the later training stage. Next, we explain how it is constructed and dropped. + +Sparse Training Augmentation. As illustrated in Fig. 4, we adopt a masked softmax-based attention as an auxiliary branch to augment ViT training. In particular, we first use a comparator with a predefined threshold $\epsilon$ to generate a binary mask, where attentions greater than $\epsilon$ are set to 1 and 0 otherwise. These masks are then applied to the attention maps to generate masked attention maps that will be + +summed up with our linear-angular attention together as the final attention. The resulting sparse attention is given by: + +$$ +\mathbf {M} _ {\text {S p a r s e A t t n}} (\mathbf {Q}, \mathbf {K}) = \operatorname {M a s k} _ {\varepsilon} \left(\operatorname {S o f t m a x} \left(\mathbf {Q} \cdot \mathbf {K} ^ {\mathrm {T}}\right)\right). \tag {10} +$$ + +where $\mathsf{Mask}_{\epsilon}(\mathbf{x}) = \mathbf{x}$ if $\mathbf{x} > \epsilon$ else 0, acting as an elementwise threshold function. As such, the sparse attention captures the higher attention scores (i.e., strong local features), and can potentially complement our linear-angular attention by supplementing the missing higher-order terms. Such an assumption aligns well with recent findings that (1) low-rank and sparse approximations complement each other [8, 17, 31]; and (2) linear attention lacks local feature extraction capabilities over its softmax-based counterpart [6]. + +Castling During ViT Inference. As we are targeting efficient ViT inference, it is desired to reduce or completely remove the costly softmax-based attentions while only keeping our linear-angular attention at runtime, i.e., performing castling. In our experiments, we found that under the sparsity regularization above, the softmax attention naturally converges to all zeros as the training progresses. We consider both fixed and dynamic schedules for $\epsilon$ and find that our Castling-ViT is not sensitive to neither the threshold value nor the threshold schedule for a given task. Given a fixed mask threshold (e.g., $\epsilon = 0.02$ in image classification experiments), the masks become all zeros at latter training stages and thus the auxiliary branch can be removed without hurting the model accuracy. We supply the visualization of the mask-evolving trajectory and our conjecture for understanding such a phenomenon in Sec. 4.5. + +# 4. Experiments + +# 4.1. Experiment Settings + +Tasks, Datasets, and Models. Tasks and Datasets. We consider three benchmark datasets and three representative vision tasks to demonstrate the superiority of the proposed Castling-ViT, including image classification on ImageNet dataset [18] with 1.2 million training and 50K validation images; Object detection on COCO dataset [36] with 118K training and 5K validation images; Semantic segmentation on ADE20K dataset [75] with 20K/2K/3K images for training, validation, and testing, respectively. Models. We apply our proposed Castling-ViT idea on top of various models. For the classification task, we consider LeViT [24], MViTv2 [34], and DeiT [54]; For the detection task, we consider models with efficient ViT backbones (e.g., PicoDet [70] with modified ESNet and LCNet backbones [15] with transformer blocks); For the segmentation task, we consider Mask2former [11] with ViT-Base backbone. + +Training Settings. For the classification task, we use a SGD optimizer with 0.9 momentum and $2 \times 10^{-5}$ weight decay to train ViTs for 1000 epochs using 64 V100 GPUs, with each card having 64 (LeViT) or 32 (MViTv2/DeiT) batch sizes. The learning rate is 2.0 + +Table 2. Castling-ViT over SOTA baselines on ImageNet. + +
MACs RangesModelsParams (M)MACs (G)Top-1 Acc.(%)Top-5 Acc.(%)
<1GMobileNetV2-1.4 [47]6.90.5874.7-
EfficientNet-B0 [51]5.30.3977.193.3
EfficientNet-B1 [51]7.80.7079.194.4
MobileViT-XS [40]2.30.7074.892.3
LeViT-128 [24]9.20.4178.6-
LeViT-192 [24]10.90.6680.0-
Castling-LeViT-12810.50.4979.694.6
Castling-LeViT-19212.70.8281.395.5
1~3GEfficientNet-B3 [51]12.01.8081.695.7
HRFormer-T [73]8.01.8078.5-
DeiT-T [54]5.61.2574.591.9
LeViT-256 [24]18.91.1281.6-
LeViT-384 [24]39.12.3582.6-
Caslting-DeiT-T5.641.1876.092.5
Castling-LeViT-25622.01.4082.696.1
Castling-LeViT-38445.82.9083.796.7
3~10GRegNetY-4G [44]21.04.0080.0-
RegNetY-8G [44]39.08.0081.7-
EfficientNet-B4 [51]19.04.2082.996.4
EfficientNet-B5 [51]30.09.9083.696.7
Swin-T [38]29.04.5081.395.5
Swin-S [38]50.08.7083.296.2
CSWin-T [19]23.04.3082.7-
CSWin-S [19]35.06.9083.6-
DeiT-S [54]21.94.6081.295.4
AutoFormer-S [10]22.95.1081.795.7
PVTv2-V2 [61]25.04.0082.0-
MViTv2-T [34]24.04.6882.3-
MViTv2-S [34]34.76.9583.6-
Castling-MViTv2-T24.14.5084.196.8
Castling-MViTv2-S34.76.9584.697.0
>10GCaiT-S36 [55]68.013.9083.3-
Swin-B [38]88.015.4083.596.5
CSWin-B [19]78.015.0084.2-
AutoFormer-B [10]54.011.0082.495.7
MViTv2-B [34]51.210.0784.4-
DeiT-B [54]86.317.5683.496.5
Castling-DeiT-B87.2217.2884.2-
Castling-MViTv2-B51.99.8285.097.2
+ +with first 11 epochs warm-up starting from 0.01 and decays by a factor of 0.9875 per epoch [62]. Also, we use the distillation [54] based on a teacher model with a $85.5\%$ accuracy. For the detection task, we adopt SGD optimizer with momentum 0.9 and weight decay 4e-5 to train models on COCO. All models are trained on 8 V100 GPUs with each card having 80 batch sizes following PicoDet's training recipe [70]. Also, we follow LeViT's training recipe [24] to pretrain backbones on ImageNet; For the segmentation task, we follow Mask2former's training recipe [11] to train models on ADE20K, where ViT backbones are pretrained following MAE [25] if specified. + +Baselines and Evaluation Metrics. Baselines. For the classification task, we compare the proposed Castling-ViT with LeViT [24], MviTv2 [34], DeiT [54], Swin [38], CSWin [19], PVT [59], etc. For the detection task, we compare with FBNetV5 [62], YOLOX [23], YOLOv5, MobileDet [65], and EfficientDet [52]. For the segmentation task, we compare with Mask2former [11] with ViT backbones. Evaluation Metrics. We evaluate the Castling-ViT and all baselines in terms of accuracy-efficiency tradeoffs. Specifically, the accuracy metrics refer to top-1/5 accuracy + +![](images/f80cbe153f3a2fa53c3b8868014ff6ed91dafbe483535704e0366c50d03a13db.jpg) +Figure 5. Castling vs. baseline ViT on ImageNet. + +![](images/f151dba8746532bcfb4368a3d9945e3c144e9e15828109e90f06fe66e7b983be.jpg) + +for the classification task; AP, $\mathrm{AP}^{50}$ , $\mathrm{AP}^{75}$ for the detection task (AP: average precision); mIoU, mAcc, and pAcc for the segmentation task (mIoU: mean intersection over union). For efficiency metrics, we compare the number of model parameters or inference FLOPs (or MACs). + +# 4.2. Castling-ViT over SOTA Baslines + +Image Classification. To evaluate the effectiveness of our proposed techniques on the image classification task, we apply the proposed Castling-ViT idea to three typical or SOTA ViT architectures: DeiT [54] as typical ViTs, LeViT [24] as representative efficient ViTs, and MViTv2 [34] as representative hierarchical ViTs for downstream applications of high input resolutions, and compare their performance over baselines on ImageNet. As shown in Tab. 2, the comparison across a large MACs (or FLOPs) range starting from 0.4G to 17G. We categorize ViT models into four regimes: $<1\mathrm{G}$ , $1 \sim 3\mathrm{G}$ , $3 \sim 10\mathrm{G}$ , and $>10\mathrm{G}$ , and select baselines in each regime to benchmark separately for the clarity purpose. Castling-ViTs consistently perform better than all baselines across various MACs ranges in terms of the accuracy-efficiency tradeoff. For example, Castling-LeViT achieves $82.6\%$ top-1 accuracy with only 1.40G MACs while LeViT requires 2.35G FLOPs instead, i.e., $\downarrow 40\%$ MACs; On the other hand, under comparable MACs, Castling-MViTv2 achieves $84.1\%$ accuracy vs. MViTv2 with $82.3\%$ accuracy instead, i.e., $\uparrow 1.8\%$ top-1 accuracy. Overall, Castling-ViT achieves an improved accuracy of $\uparrow 0.5\% \sim \uparrow 6.6\%$ , $\uparrow 1.0\% \sim \uparrow 8.1\%$ , $\uparrow 1.0\% \sim \uparrow 4.1\%$ , and $\uparrow 0.6\% \sim \uparrow 2.6\%$ over baselines under $<1\mathrm{G}$ MACs, $1 \sim 3\mathrm{G}$ MACs, $3 \sim 10\mathrm{G}$ MACs, and $>10\mathrm{G}$ MACs, respectively. Note that we calculate improvements under comparable MACs. In addition to the overall comparison, we also visualize the apple-to/apple benchmark, e.g., Castling-LeViT vs. LeViT, to validate the effectiveness of proposed techniques. As shown in Fig. 5, Castling-LeViT/MViTv2 achieves $\downarrow 25.7\% \sim \downarrow 55.3\%$ MACs reductions under comparable accuracies or offers a comparable or better accuracy ( $\uparrow 0.6\% \sim \uparrow 1.8\%$ ) under comparable MACs over corresponding LeViT/MViTv2 baselines. + +Object Detection. We also extend the Castling-ViT to the downstream object detection task and compare it with previous efficient detectors on COCO dataset to evaluate its efficacy. Specifically, we construct the detector with modified ESNet [70] or LCNet [15] (replace the last one or two + +Table 3. Castling-ViT over SOTA baselines on COCO detection, where Castling-ViT† means that the detection model adopts modified ESNet (use transformer blocks in the last stage; referred to ESNet-ViT) as backbone while Castling-ViT‡ uses modified LCNet (use transformer blocks in the last two stage; referred to LCNet-ViT) as backbone. Also, “-S/M/L” and “-320/416” represent small/medium/large model variants and 320x320/416x416 input resolutions, respectively. + +
ModelsParams (M)MACs (G)mAPAP50AP75
MobileDet-DSP [65]4.851.4327.3--
MobileDet-DSP-Fused [65]9.153.2229.1--
EfficientDet-D0 [52]3.902.5434.653.037.1
YOLOv5-N1.902.3028.045.7-
YOLOX-Nano [23]0.910.5425.8--
YOLOX-Tiny [23]5.063.2332.8--
FBNetV5-AC-224x320 [62]-0.7125.0--
FBNetV5-AR-224x320 [62]-0.9126.4--
FBNetV5-A-224x320 [62]-1.3527.2--
FBNetV5-AC-320x640 [62]-1.3728.9--
FBNetV5-AR-320x640 [62]-1.8030.4--
Castling-ViT-S-320†3.250.6228.142.329.2
Castling-ViT-S-416†3.251.0631.346.732.5
Castling-ViT-M-416†6.012.0034.049.535.9
Castling-ViT-L-416†9.323.0335.050.737.2
Castling-ViT-L-416*13.105.3137.353.439.6
+ +stages with transformer blocks) as backbones and follow PicoDet [70]'s detection head design as well as their training recipe. Fig. 1 and Tab. 3 show the overall comparison between the proposed Castling-ViT and other baselines. We can see that our Castling-ViT consistently achieves better accuracy-efficiency tradeoffs, leading to $\uparrow 6.0$ , $\uparrow 2.2 \sim \uparrow 2.3$ , $\uparrow 4.0 \sim \uparrow 5.9$ , $\uparrow 3.1 \sim \uparrow 4.0$ mAP improvements as compared to YOLOv5, YOLOX [23], MobileDet [65], and FBNetv5 [62], respectively, under comparable or even less MACs. As for EfficientDet [52], our method achieves comparable accuracy-efficiency trade-offs. Apart from the overall comparison with baselines, we also provide the apple-to-Apple comparison with detectors with softmax-based attention, we supply the comparison results and Castling-ViT's breakdown analysis to Sec. 4.4. This set of experiments validate the effectiveness of the proposed Castling-ViT for servering as efficient detector backbones (e.g., $\leq 3G$ MACs) in the object detection task. + +Semantic Segmentation. We further extend Castling-ViT to the semantic segmentation task to evaluate its effectiveness. Specifically, we use ViT-Base as the backbone in Mask2former [11] framework to serve as our baseline and testbed. Then we build the Castling-ViT-Base as the backbone to benchmark on the ADE20K dataset. As shown in Tab. 4, Mask2former with Castling-ViT backbone achieves $15\%$ total MACs reductions and $19\%$ backbone MACs reduction under comparable or slightly better mIoU, i.e., $\uparrow 0.13\%$ and $\uparrow 0.52\%$ without or with MAE pretraining on ImageNet [25]. This set of experiments validate that our proposed techniques could be well generalized to various downstream tasks that require large input resolutions. + +Table 4. Castling-ViT over baselines on the ADE20K segmentation dataset, serving as backbones in the Mask2former framework. Note that we report both total MACs and backbone MACs, i.e., $(\cdot)$ . + +
Mask2former w/MAE +PretrainParams +(M)MACs +(G)mIoUmAccpAcc
ViT-BaseN118229 (182)34.5446.3675.84
Castling-ViT-BaseN118195 (147)34.6746.4776.20
ViT-BaseY118229 (182)47.9261.0083.02
Castling-ViT-BaseY118195 (147)48.4461.8283.29
+ +Table 5. Ablation studies of Castling-ViT with various kernel-based linear attentions, where MC refers to mean-centering kernel function applied to $\mathbf{Q} / \mathbf{K}$ ; Softmax also means the kernel function; ReLU-S and ReLU-E denote applying ReLU as kernels for decomposing softmax or exponential terms, respectively. All models are trained on COCO from scratch without pretraining. + +
BackboneLinear AttnS-320S-416M-416
mAPAP50mAPAP50mAPAP50
LCNetN/A25.539.029.744.234.349.9
LCNet-ViTN/A25.939.629.844.434.950.4
LCNet-ViTMC [17]22.834.926.339.631.546.1
Softmax [49]26.039.629.243.434.449.8
Cosine [5]24.437.427.241.032.948.0
ReLU-S23.936.7--30.945.3
ReLU-E [6]26.540.230.244.734.349.7
Lin.-Angular27.040.630.445.535.551.1
+ +# 4.3. Linear-Angular Attention over SOTA Baselines + +We also conduct ablation studies among various kinds of kernels used in linear attention to evaluate the superiority of our proposed linear-angular kernel. Also shown in Tab. 5, we compare the angular kernel with five other commonly adopted kernels, results on three model and resolution settings consistently demonstrate that our proposed angular kernel helps achieve better mAP, e.g., $\uparrow 4.0\% \sim \uparrow 4.2\%$ over MC [17], $\uparrow 1.0\% \sim \uparrow 1.2\%$ over Softmax [49], $\uparrow 2.6\% \sim \uparrow 3.2\%$ over Cosine [5], $\uparrow 3.1\% \sim \uparrow 4.6\%$ and $\uparrow 0.2\% \sim \uparrow 1.2\%$ over ReLU-S and ReLU-E [6], respectively. This set of experiments validate the superiority of the proposed linear-angular kernels. + +# 4.4. Ablation Studies of Castling-ViT + +We conduct ablation studies on Castling-ViT's linear- angular attention, added DWConv, as well as auxiliary masked softmax-based attention, as shown in Tab. 6 and Tab. 7, where the experiments are performed on CastlingViT-S-320\*/M-416\*, respectively, without pretraining on ImageNet. Note that here \* means that we use LCNet-ViT, i.e., replacing all layers in last two stages with transformer blocks, as backbones. Results in three tables show detailed performance breakdown and consistently demonstrate that all components in our proposed Castling-ViT contribute to the final performance. Specifically, linear-angular attention itself already achieves $\downarrow 0.1\% \sim \uparrow 0.5\%$ mAP improvements while leading to $10.1\% \sim 15.3\%$ MACs reductions simultaneously. Adding DWConv in the middle of MLP layers further leads to $\downarrow 0.1\% \sim \uparrow 0.6\%$ mAP improvements while only incur negligible MACs overhead. Adding aux + +Table 6. Ablation studies of Castling-ViT-S-320 w/ LCNet backbone, or LCNet-ViT backbone + linear-angular attention (denoted as Lin.) + DWConv (denoted as DW.) + auxiliary softmax-based attention (denoted as SparseAttn.). + +
Castling-ViT-S-320 w/Params (M)MACs (G)mAPAP50AP75
LCNet1.100.4825.539.026.5
LCNet-ViT2.140.6925.939.626.9
+ Lin.2.140.6226.239.727.3
+ Lin.+DW.2.140.6226.840.728.0
+ Lin.+DW.+SparseAttn.2.140.6227.040.628.0
+ +Table 7. Ablation studies of Castling-ViT-M-416 w/ LCNet backbone, or LCNet-ViT backbone + Lin. + DW. + SparseAttn. + +
Castling-ViT-M-416 w/Params (M)MACs (G)mAPAP50AP75
LCNet3.322.1534.349.936.5
LCNet-ViT7.533.5134.950.437.0
+ Lin.7.533.1534.850.336.7
+ Lin.+DW.7.533.1535.150.936.9
+ Lin.+DW.+SparseAttn.7.533.1535.351.137.3
+ +illary masked softmax-based attention further increases the mAP by $0.2\%$ without incurring any overhead since it only assists the training while being removed during inference. Note that although LCNet-ViT backbone leads to better mAP, it also introduces more MACs as compared to efficient convolutions. As such, LCNet-ViT still slightly underperforms LCNet backbones in terms of accuracy-efficiency trade-offs, since we directly replace the CNN backbones with transformer blocks, whose architecture could not be optimal for ViTs. It remains to an open problem to build ViT based models that achieves higher accuracy-efficiency trade-off than pure ConvNet models. + +# 4.5. Discussion on the Auxiliary Branch + +Trajectory of Nonzeros in Masks. To further understand the effect of the auxiliary branch, we visualize the trajectories of nonzeros in masks of auxiliary softmax-based attention. As shown in Fig. 6, we count the nonzeros in the first attention layer's masks throughout the training of both Castling-ViT-S-416 + +and Castling-ViT-M-416 with the LCNet-ViT backbone on COCO. we observe that the introduced auxiliary attention will only assist the training in the early or middle training stages and will gradually vanish towards all zeros in the later training stage, which is well aligned with the assumption in our Castling-ViT. This set of experiments validate the idea of performing "castling" in ViTs, i.e., drop the auxiliary branch without sacrificing performance. + +![](images/5279c2953fa8e8015eb33abc95d73b9575265d7ccf39c40c1efc3d7cca79218c.jpg) +Figure 6. Visualizing the trajectories of nonzeros in auxiliary masks during training. + +![](images/d75300cba95932f023a0bd0c5f4439fede7dee84c03a47788d5601f602833679.jpg) +Figure 7. Left: Visualizing the taget similarity function, linear-angular terms, and DNNs' fitting curve; Middle: Visualizing the loss trajectory during DNN training; Right: Visualizing the learning trajectory of both low and high frequency components. + +![](images/80a8e7ce0330b01dacacd833646e1813cd7573a0f392c088e9016f4ee2192c35.jpg) + +![](images/06ba51f9e7a7ddd23a070f84e831f634c43e5996188046b58fa7e971ca0d3996.jpg) + +Conjecture of Castling. We conduct a synthetic experiment to give an analogy for explaining the castling phenomenon. In Fig. 7 (Left), we visualize the curve of (1) angular similarity function (denoted in Red), (2) linear-angular terms (denoted in Green), and (3) two-layer DNN's approximation. We see that only keeping linear-angular terms leads to distortion, while DNN is capable of learning the missing high-frequency parts, whose loss trajectory is shown in Fig. 7 (Middle). Also, the learning trajectory of frequency components is visualized in Fig. 7 (Right), following F-Principle [67], where x-axis is training steps, y-axis categorizes both low and high frequency components, blue/red colors refer to large/small difference between the learned frequency components and target frequency components, respectively. We observe that DNNs fit target functions from low to high frequencies during training. It indicates that DWConv itself is not sufficient for approximating the high-order residual at early training stages. Costly attention that contains high-order residuals is then desired to help training while being dropped at inference since the remaining networks can gradually learn the high-frequency components at later training stages [67]. + +# 5. Conclusion + +We present Castling-ViT that trains ViTs with both linear-angular and softmax-based quadratic attention but switches to only having the former during inference. Castling-ViT leverages angular kernels to measure the similarities between the queries and keys via spectral angles and highlights two enablers: (1) a new linear-angular attention mechanism: we decompose angular kernels to linear terms and high-order residuals, and keep only the former for inference; and (2) we approximate the high-order residuals using a depthwise convolution and an auxiliary masked softmax attention whose masks gradually become zeros during training without incurring inference overhead. Extensive experiments consistently validate Castling-ViT's advantages. + +# Acknowledgment + +The work is supported in part by the National Science Foundation (NSF) RTML program (Award number: 1937592) and the CoCoSys, one of the seven centers in JUMP 2.0, a Semiconductor Research Corporation (SRC) program sponsored by DARPA. + +# References + +[1] Alaaeldin Ali, Hugo Touvron, Mathilde Caron, Piotr Bojanowski, Matthijs Douze, Armand Joulin, Ivan Laptev, Natalia Neverova, Gabriel Synnaeve, Jakob Verbeek, et al. Xcit: Cross-covariance image transformers. Advances in neural information processing systems, 34:20014-20027, 2021. 2, 13 +[2] Moab Arar, Ariel Shamir, and Amit H Bermano. Learned queries for efficient local attention. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10841-10852, 2022. 1, 2, 13 +[3] Iz Beltagy, Matthew E Peters, and Arman Cohan. Longformer: The long-document transformer. arXiv preprint arXiv:2004.05150, 2020. 1, 13 +[4] Daniel Bolya, Cheng-Yang Fu, Xiaoliang Dai, Peizhao Zhang, Christoph Feichtenhofer, and Judy Hoffman. Token merging: Your vit but faster. arXiv preprint arXiv:2210.09461, 2022. 2 +[5] Daniel Bolya, Cheng-Yang Fu, Xiaoliang Dai, Peizhao Zhang, and Judy Hoffman. Hydra attention: Efficient attention with many heads. arXiv preprint arXiv:2209.07484, 2022. 1, 2, 5, 7, 13 +[6] Han Cai, Chuang Gan, and Song Han. Efficientvit: Enhanced linear attention for high-resolution low-computation visual recognition. arXiv preprint arXiv:2205.14756, 2022. 1, 2, 3, 5, 7, 13, 14 +[7] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In European conference on computer vision, pages 213-229. Springer, 2020. 13 +[8] Beidi Chen, Tri Dao, Eric Winsor, Zhao Song, Atri Rudra, and Christopher Ré. Scatterbrain: Unifying sparse and low-rank attention. Advances in Neural Information Processing Systems, 34:17413-17426, 2021. 1, 2, 5, 13 +[9] Chun-Fu Chen, Quanfu Fan, and Rameswar Panda. Crossvit: Cross-attention multi-scale vision transformer for image classification. arXiv preprint arXiv:2103.14899, 2021. 2 +[10] Minghao Chen, Houwen Peng, Jianlong Fu, and Haibin Ling. Autoformer: Searching transformers for visual recognition. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12270-12280, 2021. 6 +[11] Bowen Cheng, Ishan Misra, Alexander G Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1290–1299, 2022. 3, 5, 6, 7, 13 +[12] Bowen Cheng, Alex Schwing, and Alexander Kirillov. Per-pixel classification is not all you need for semantic segmentation. Advances in Neural Information Processing Systems, 34:17864-17875, 2021. 13 +[13] Krzysztof Choromanski, Valerii Likhosherstov, David Dohan, Xingyou Song, Andreea Gane, Tamas Sarlos, Peter Hawkins, Jared Davis, Afroz Mohiuddin, Lukasz Kaiser, + +et al. Rethinking attention with performers. arXiv preprint arXiv:2009.14794, 2020.3 +[14] Krzysztof Marcin Choromanski, Valerii Likhosherstov, David Dohan, Xingyou Song, Andreea Gane, Tamas Sarlos, Peter Hawkins, Jared Quincy Davis, Afroz Mohiuddin, Lukasz Kaiser, David Benjamin Belanger, Lucy J Colwell, and Adrian Weller. Rethinking attention with performers. In International Conference on Learning Representations, 2021. 1, 2, 13 +[15] Cheng Cui, Tingquan Gao, Shengyu Wei, Yuning Du, Ruoyu Guo, Shuilong Dong, Bin Lu, Ying Zhou, Xueying Lv, Qiwen Liu, et al. Pp-lcnet: A lightweight cpu convolutional neural network. arXiv preprint arXiv:2109.15099, 2021. 5, 6 +[16] Xiaoliang Dai, Alvin Wan, P. Zhang, B. Wu, Zijian He, Zhen Wei, K. Chen, Yuandong Tian, Matthew E. Yu, Péter Vajda, and J. Gonzalez. Fbnetv3: Joint architecture-recipe search using neural acquisition function. ArXiv, abs/2006.02049, 2020.2 +[17] Jyotikrishna Dass, Shang Wu, Huihong Shi, Chaojian Li, Zhifan Ye, Zhongfeng Wang, and Yingyan Lin. Vitality: Unifying low-rank and sparse approximation for vision transformer acceleration with a linear taylor attention. In The 29th IEEE International Symposium on High-Performance Computer Architecture (HPCA 2023), 2023. 2, 5, 7, 14 +[18] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 1, 5 +[19] Xiaoyi Dong, Jianmin Bao, Dongdong Chen, Weiming Zhang, Nenghai Yu, Lu Yuan, Dong Chen, and Baining Guo. Cswin transformer: A general vision transformer backbone with cross-shaped windows. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12124-12134, 2022. 6 +[20] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations, 2021. 1, 2, 3 +[21] Haoqi Fan, Bo Xiong, Karttikeya Mangalam, Yanghao Li, Zhicheng Yan, Jitendra Malik, and Christoph Feichtenhofer. Multiscale vision transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6824-6835, 2021. 2 +[22] Peng Gao, Teli Ma, Hongsheng Li, Jifeng Dai, and Yu Qiao. Convmae: Masked convolution meets masked autoencoders. arXiv preprint arXiv:2205.03892, 2022. 13 +[23] Zheng Ge, Songtao Liu, Feng Wang, Zeming Li, and Jian Sun. Yolox: Exceeding yolo series in 2021. arXiv preprint arXiv:2107.08430, 2021. 6, 7, 14 +[24] Ben Graham, Alaaeldin El-Nouby, Hugo Touvron, Pierre Stock, Armand Joulin, Hervé Jégou, and Matthijs Douze. Levit: a vision transformer in convnet's clothing for faster inference. arXiv preprint arXiv:2104.01136, 2021. 2, 3, 5, 6, 13, 14 + +[25] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollár, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16000-16009, 2022. 6, 7, 13 +[26] Byeongho Heo, Sangdoo Yun, Dongyoon Han, Sanghyuk Chun, Junsuk Choe, and Seong Joon Oh. Rethinking spatial dimensions of vision transformers. arXiv preprint arXiv:2103.16302, 2021. 2 +[27] Paul Honeine and Cedric Richard. The angular kernel in machine learning for hyperspectral data classification. In 2010 2nd Workshop on Hyperspectral Image and Signal Processing: Evolution in Remote Sensing, pages 1-4. IEEE, 2010. 4 +[28] Andrew Howard, Mark Sandler, Grace Chu, Liang-Chieh Chen, Bo Chen, Mingxing Tan, Weijun Wang, Yukun Zhu, Ruoming Pang, Vijay Vasudevan, et al. Searching for mobilenetv3. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1314-1324, 2019. 2 +[29] Angelos Katharopoulos, Apoorv Vyas, Nikolaos Pappas, and François Fleuret. Transformers are rnns: Fast autoregressive transformers with linear attention. In International Conference on Machine Learning, pages 5156-5165. PMLR, 2020. 1, 2, 3, 4, 13 +[30] Nirmal Keshava. Distance metrics and band selection in hyperspectral processing with applications to material identification and spectral libraries. IEEE Transactions on Geoscience and remote sensing, 42(7):1552-1565, 2004. 4 +[31] Kyungmin Kim, Bichen Wu, Xiaoliang Dai, Peizhao Zhang, Zhicheng Yan, Peter Vajda, and Seon Joo Kim. Rethinking the self-attention in vision transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3071–3075, 2021. 4, 5 +[32] Nikita Kitaev, Lukasz Kaiser, and Anselm Levskaya. Reformer: The efficient transformer. arXiv preprint arXiv:2001.04451, 2020. 1, 13 +[33] Yanghao Li, Hanzi Mao, Ross Girshick, and Kaiming He. Exploring plain vision transformer backbones for object detection. arXiv preprint arXiv:2203.16527, 2022. 13 +[34] Yanghao Li, Chao-Yuan Wu, Haoqi Fan, Karttikeya Mangalam, Bo Xiong, Jitendra Malik, and Christoph Feichtenhofer. Mviv2: Improved multiscale vision transformers for classification and detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4804-4814, 2022. 3, 5, 6, 13, 14 +[35] Yanyu Li, Geng Yuan, Yang Wen, Eric Hu, Georgios Evangelidis, Sergey Tulyakov, Yanzhi Wang, and Jian Ren. Efficientformer: Vision transformers at mobilenet speed. arXiv preprint arXiv:2206.01191, 2022. 2 +[36] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In European conference on computer vision, pages 740-755. Springer, 2014. 1, 5 +[37] Jing Liu, Zizheng Pan, Haoyu He, Jianfei Cai, and Bohan Zhuang. Ecoformer: Energy-saving attention with linear complexity. In NeurIPS, 2022. 1, 2, 4, 13 + +[38] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10012-10022, 2021. 1, 2, 6, 13, 14 +[39] Jiachen Lu, Jinghan Yao, Junge Zhang, Xiatian Zhu, Hang Xu, Weiguo Gao, Chunjing Xu, Tao Xiang, and Li Zhang. Soft: softmax-free transformer with linear complexity. Advances in Neural Information Processing Systems, 34:21297-21309, 2021. 1, 2, 13 +[40] Sachin Mehta and Mohammad Rastegari. Mobilevit: lightweight, general-purpose, and mobile-friendly vision transformer. arXiv preprint arXiv:2110.02178, 2021. 2, 6 +[41] John Moody and Christian Darken. _Learning with localized receptive fields_. Yale Univ., Department of Computer Science, 1988. 4 +[42]Zipeng Qin, Jianbo Liu, Xiaolin Zhang, Maoqing Tian, Aojun Zhou, Shuai Yi, and Hongsheng Li. Pyramid fusion transformer for semantic segmentation. arXiv preprint arXiv:2201.04019, 2022.3, 13 +[43] Zhen Qin, Weixuan Sun, Hui Deng, Dongxu Li, Yunshen Wei, Baohong Lv, Junjie Yan, Lingpeng Kong, and Yiran Zhong. cosformer: Rethinking softmax in attention. In International Conference on Learning Representations, 2022. 1, 2, 13 +[44] Ilija Radosavovic, Raj Prateek Kosaraju, Ross Girshick, Kaiming He, and Piotr Dólár. Designing network design spaces. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10428-10436, 2020. 6 +[45] Ali Rahimi and Benjamin Recht. Random features for large-scale kernel machines. Advances in neural information processing systems, 20, 2007. 3 +[46] Yongming Rao, Wenliang Zhao, Benlin Liu, Jiwen Lu, Jie Zhou, and Cho-Jui Hsieh. Dynamicvit: Efficient vision transformers with dynamic token sparsification. Advances in neural information processing systems, 34:13937-13949, 2021. 2 +[47] Mark Sandler, Andrew Howard, Menglong Zhu, Andrey Zh-moginov, and Liang-Chieh Chen. Mobilenetv2: Inverted residuals and linear bottlenecks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4510-4520, 2018. 6 +[48] Robert R. Schaller. Moore's law: past, present and future. IEEE spectrum, 1997. 14 +[49] Zhuoran Shen, Mingyuan Zhang, Haiyu Zhao, Shuai Yi, and Hongsheng Li. Efficient attention: Attention with linear complexities. In Proceedings of the IEEE/CVF winter conference on applications of computer vision, pages 3531-3539, 2021. 3, 7 +[50] Chen Sun, Abhinav Shrivastava, Saurabh Singh, and Abhinav Gupta. Revisiting unreasonable effectiveness of data in deep learning era. In ICCV, 2017. 2 +[51] Mingxing Tan and Quoc Le. Efficientnet: Rethinking model scaling for convolutional neural networks. In International conference on machine learning, pages 6105-6114. PMLR, 2019. 6 + +[52] Mingxing Tan, Ruoming Pang, and Quoc V Le. Efficientdet: Scalable and efficient object detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10781-10790, 2020. 6, 7 +[53] Shitao Tang, Jiahui Zhang, Siyu Zhu, and Ping Tan. Quadtree attention for vision transformers. In International Conference on Learning Representations, 2022. 1 +[54] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Herve Jegou. Training data-efficient image transformers & distillation through attention. In Proceedings of the 38th International Conference on Machine Learning, volume 139 of Proceedings of Machine Learning Research, pages 10347-10357. PMLR, 18-24 Jul 2021. 2, 5, 6, 14 +[55] Hugo Touvron, Matthieu Cord, Alexandre Sablayrolles, Gabriel Synnaeve, and Hervé Jégou. Going deeper with image transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 32-42, 2021. 6 +[56] Zhengzhong Tu, Hossein Talebi, Han Zhang, Feng Yang, Peyman Milanfar, Alan Bovik, and Yinxiao Li. Maxvit: Multi-axis vision transformer. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXIV, pages 459-479. Springer, 2022. 2, 13 +[57] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Ilia Polosukhin. Attention is all you need. In Advances in Neural Information Processing Systems, volume 30. Curran Associates, Inc., 2017. 1, 3 +[58] Sinong Wang, Belinda Z. Li, Madian Khabsa, Han Fang, and Hao Ma. Linformer: Self-attention with linear complexity, 2020. 1, 2, 3, 13 +[59] Wenhai Wang, Enze Xie, Xiang Li, Deng-Ping Fan, Kaitao Song, Ding Liang, Tong Lu, Ping Luo, and Ling Shao. Pyramid vision transformer: A versatile backbone for dense prediction without convolutions. arXiv preprint arXiv:2102.12122, 2021. 2, 6, 14 +[60] Wenhai Wang, Enze Xie, Xiang Li, Deng-Ping Fan, Kaitao Song, Ding Liang, Tong Lu, Ping Luo, and Ling Shao. Pyramid vision transformer: A versatile backbone for dense prediction without convolutions. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 568-578, October 2021. 3, 13 +[61] Wenhai Wang, Enze Xie, Xiang Li, Deng-Ping Fan, Kaitao Song, Ding Liang, Tong Lu, Ping Luo, and Ling Shao. Pvt v2: Improved baselines with pyramid vision transformer. Computational Visual Media, 8(3):415-424, 2022. 6 +[62] Bichen Wu, Chaojian Li, Hang Zhang, Xiaoliang Dai, Peizhao Zhang, Matthew Yu, Jialiang Wang, Yingyan Lin, and Peter Vajda. Fbnetv5: Neural architecture search for multiple tasks in one run. arXiv preprint arXiv:2111.10007, 2021.6,7 +[63] Haiping Wu, Bin Xiao, Noel Codella, Mengchen Liu, Xiyang Dai, Lu Yuan, and Lei Zhang. Cvt: Introducing convolutions to vision transformers. arXiv preprint arXiv:2103.15808, 2021. 2 + +[64] Zhanghao Wu, Zhijian Liu, Ji Lin, Yujun Lin, and Song Han. Lite transformer with long-short range attention. In International Conference on Learning Representations, 2020. 1 +[65] Yunyang Xiong, Hanxiao Liu, Suyog Gupta, Berkin Akin, Gabriel Bender, Pieter-Jan Kindermans, Mingxing Tan, Vikas Singh, and Bo Chen. Mobiledets: Searching for object detection architectures for mobile accelerators. arXiv preprint arXiv:2004.14525, 2020. 2, 6, 7 +[66] Yunyang Xiong, Zhanpeng Zeng, Rudrasis Chakraborty, Mingxing Tan, Glenn Fung, Yin Li, and Vikas Singh. Nystromformer: A nystrom-based algorithm for approximating self-attention. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 35, pages 14138–14148, 2021. 1, 2, 13 +[67] Zhi-Qin John Xu, Yaoyu Zhang, Tao Luo, Yanyang Xiao, and Zheng Ma. Frequency principle: Fourier analysis sheds light on deep neural networks. arXiv preprint arXiv:1901.06523, 2019. 2, 5, 8 +[68] Hongxu Yin, Arash Vahdat, Jose M Alvarez, Arun Mallya, Jan Kautz, and Pavlo Molchanov. A-vit: Adaptive tokens for efficient vision transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10809-10818, 2022. 2 +[69] Haoran You, Zhanyi Sun, Huihong Shi, Zhongzhi Yu, Yang Zhao, Yongan Zhang, Chaojian Li, Baopu Li, and Yingyan Lin. Vitcod: Vision transformer acceleration via dedicated algorithm and accelerator co-design. In The 29th IEEE International Symposium on High-Performance Computer Architecture (HPCA 2023), 2023. 1 +[70] Guanghua Yu, Qinyao Chang, Wenyu Lv, Chang Xu, Cheng Cui, Wei Ji, Qingqing Dang, Kaipeng Deng, Guanzhong Wang, Yuning Du, et al. Pp-picodet: A better real-time object detector on mobile devices. arXiv preprint arXiv:2111.00902, 2021. 5, 6, 7 +[71] Zhongzhi Yu, Yonggan Fu, Sicheng Li, Chaojian Li, and Yingyan Lin. Mia-former: Efficient and robust vision transformers via multi-grained input-adaptation. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 36, pages 8962-8970, 2022. 2 +[72] Li Yuan, Yunpeng Chen, Tao Wang, Weihao Yu, Yujun Shi, Zi-Hang Jiang, Francis EH Tay, Jiashi Feng, and Shuicheng Yan. Tokens-to-token vit: Training vision transformers from scratch onImagenet. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 558-567, 2021. 2, 3 +[73] Yuhui Yuan, Rao Fu, Lang Huang, Weihong Lin, Chao Zhang, Xilin Chen, and Jingdong Wang. Hrformer: High-resolution vision transformer for dense predict. Advances in Neural Information Processing Systems, 34:7281-7293, 2021.6 +[74] Zixiao Zhang, Xiaogiang Lu, Guojin Cao, Yuting Yang, Licheng Jiao, and Fang Liu. Vit-yolo: Transformer-based yolo for object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2799-2808, 2021. 13 +[75] Bolei Zhou, Hang Zhao, Xavier Puig, Sanja Fidler, Adela Barriuso, and Antonio Torralba. Scene parsing through + +ade20k dataset. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 633-641, 2017. 5 +[76] Chen Zhu, Wei Ping, Chaowei Xiao, Mohammad Shoeybi, Tom Goldstein, Anima Anandkumar, and Bryan Catanzaro. Long-short transformer: Efficient transformers for language and vision. Advances in Neural Information Processing Systems, 34, 2021. 1, 2 +[77] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable detr: Deformable transformers for end-to-end object detection. arXiv preprint arXiv:2010.04159, 2020. 3, 13 \ No newline at end of file diff --git a/2023/Castling-ViT_ Compressing Self-Attention via Switching Towards Linear-Angular Attention at Vision Transformer Inference/images.zip b/2023/Castling-ViT_ Compressing Self-Attention via Switching Towards Linear-Angular Attention at Vision Transformer Inference/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..e33dd5be7788c9ffc8afbf1d1544142ebed60bc5 --- /dev/null +++ b/2023/Castling-ViT_ Compressing Self-Attention via Switching Towards Linear-Angular Attention at Vision Transformer Inference/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f5352999260859169cbcf2366606d2f61d4bddb31adfcd5af21826b4d14b73e6 +size 618109 diff --git a/2023/Castling-ViT_ Compressing Self-Attention via Switching Towards Linear-Angular Attention at Vision Transformer Inference/layout.json b/2023/Castling-ViT_ Compressing Self-Attention via Switching Towards Linear-Angular Attention at Vision Transformer Inference/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..c1e44402f384b70beefeab474a78e467ce02a15d --- /dev/null +++ b/2023/Castling-ViT_ Compressing Self-Attention via Switching Towards Linear-Angular Attention at Vision Transformer Inference/layout.json @@ -0,0 +1,9817 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 84, + 103, + 505, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 84, + 103, + 505, + 140 + ], + "spans": [ + { + "bbox": [ + 84, + 103, + 505, + 140 + ], + "type": "text", + "content": "Castling-ViT: Compressing Self-Attention via Switching Towards Linear-Angular Attention at Vision Transformer Inference" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 96, + 153, + 497, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 96, + 153, + 497, + 224 + ], + "spans": [ + { + "bbox": [ + 96, + 153, + 497, + 224 + ], + "type": "text", + "content": "Haoran You" + }, + { + "bbox": [ + 96, + 153, + 497, + 224 + ], + "type": "inline_equation", + "content": "^{1,2,\\dagger,*}" + }, + { + "bbox": [ + 96, + 153, + 497, + 224 + ], + "type": "text", + "content": ", Yunyang Xiong" + }, + { + "bbox": [ + 96, + 153, + 497, + 224 + ], + "type": "inline_equation", + "content": "^{2,*}" + }, + { + "bbox": [ + 96, + 153, + 497, + 224 + ], + "type": "text", + "content": ", Xiaoliang Dai" + }, + { + "bbox": [ + 96, + 153, + 497, + 224 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 96, + 153, + 497, + 224 + ], + "type": "text", + "content": ", Bichen Wu" + }, + { + "bbox": [ + 96, + 153, + 497, + 224 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 96, + 153, + 497, + 224 + ], + "type": "text", + "content": ", Peizhao Zhang" + }, + { + "bbox": [ + 96, + 153, + 497, + 224 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 96, + 153, + 497, + 224 + ], + "type": "text", + "content": ", Haoqi Fan" + }, + { + "bbox": [ + 96, + 153, + 497, + 224 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 96, + 153, + 497, + 224 + ], + "type": "text", + "content": ", Peter Vajda" + }, + { + "bbox": [ + 96, + 153, + 497, + 224 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 96, + 153, + 497, + 224 + ], + "type": "text", + "content": ", Yingyan (Celine) Lin" + }, + { + "bbox": [ + 96, + 153, + 497, + 224 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 96, + 153, + 497, + 224 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 96, + 153, + 497, + 224 + ], + "type": "text", + "content": "Georgia Institute of Technology, " + }, + { + "bbox": [ + 96, + 153, + 497, + 224 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 96, + 153, + 497, + 224 + ], + "type": "text", + "content": "Meta Research \n{haoran.you, celine.lin}@gatech.edu \n{yunyang, xiaoliang dai, wbc, stzpz, haoqifan, vajdap}@meta.com" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 143, + 243, + 192, + 255 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 243, + 192, + 255 + ], + "spans": [ + { + "bbox": [ + 143, + 243, + 192, + 255 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 268, + 289, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 268, + 289, + 639 + ], + "spans": [ + { + "bbox": [ + 47, + 268, + 289, + 639 + ], + "type": "text", + "content": "Vision Transformers (ViTs) have shown impressive performance but still require a high computation cost as compared to convolutional neural networks (CNNs), one reason is that ViTs' attention measures global similarities and thus has a quadratic complexity with the number of input tokens. Existing efficient ViTs adopt local attention or linear attention, which sacrifice ViTs' capabilities of capturing either global or local context. In this work, we ask an important research question: Can ViTs learn both global and local context while being more efficient during inference? To this end, we propose a framework called Castling-ViT, which trains ViTs using both linear-angular attention and masked softmax-based quadratic attention, but then switches to having only linear-angular attention during inference. Our Castling-ViT leverages angular kernels to measure the similarities between queries and keys via spectral angles. And we further simplify it with two techniques: (1) a novel linear-angular attention mechanism: we decompose the angular kernels into linear terms and high-order residuals, and only keep the linear terms; and (2) we adopt two parameterized modules to approximate high-order residuals: a depthwise convolution and an auxiliary masked softmax attention to help learn global and local information, where the masks for softmax attention are regularized to gradually become zeros and thus incur no overhead during inference. Extensive experiments validate the effectiveness of our Castling-ViT, e.g., achieving up to a " + }, + { + "bbox": [ + 47, + 268, + 289, + 639 + ], + "type": "inline_equation", + "content": "1.8\\%" + }, + { + "bbox": [ + 47, + 268, + 289, + 639 + ], + "type": "text", + "content": " higher accuracy or " + }, + { + "bbox": [ + 47, + 268, + 289, + 639 + ], + "type": "inline_equation", + "content": "40\\%" + }, + { + "bbox": [ + 47, + 268, + 289, + 639 + ], + "type": "text", + "content": " MACs reduction on classification and 1.2 higher mAP on detection under comparable FLOPs, as compared to ViTs with vanilla softmax-based attentions. Project page is available at here." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 652, + 128, + 665 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 652, + 128, + 665 + ], + "spans": [ + { + "bbox": [ + 47, + 652, + 128, + 665 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 670, + 287, + 694 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 670, + 287, + 694 + ], + "spans": [ + { + "bbox": [ + 47, + 670, + 287, + 694 + ], + "type": "text", + "content": "Vision Transformers (ViTs) have made significant progress in image classification, object detection, and many" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 310, + 243, + 544, + 339 + ], + "blocks": [ + { + "bbox": [ + 310, + 243, + 544, + 339 + ], + "lines": [ + { + "bbox": [ + 310, + 243, + 544, + 339 + ], + "spans": [ + { + "bbox": [ + 310, + 243, + 544, + 339 + ], + "type": "image", + "image_path": "07c00801ece4ca556f08c973c3b493a5ed8898b8012feb40e2288e0cba1fd932.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 309, + 342, + 544, + 437 + ], + "blocks": [ + { + "bbox": [ + 309, + 342, + 544, + 437 + ], + "lines": [ + { + "bbox": [ + 309, + 342, + 544, + 437 + ], + "spans": [ + { + "bbox": [ + 309, + 342, + 544, + 437 + ], + "type": "image", + "image_path": "d17453d5505c6838894302c6cb13f95985a07f7942fc59096ef8ab3348323632.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 306, + 437, + 545, + 460 + ], + "lines": [ + { + "bbox": [ + 306, + 437, + 545, + 460 + ], + "spans": [ + { + "bbox": [ + 306, + 437, + 545, + 460 + ], + "type": "text", + "content": "Figure 1. Castling-ViT over SOTA baselines on (1) ImageNet [18] image classification and (2) COCO [36] object detection." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 470, + 545, + 697 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 470, + 545, + 697 + ], + "spans": [ + { + "bbox": [ + 304, + 470, + 545, + 697 + ], + "type": "text", + "content": "other applications. It is well recognized that the superior performance achieved by ViTs is largely attributed to their self-attention modules that can better capture global context [20, 57, 64]. Nevertheless, ViTs' powerful self-attention module comes at the cost of quadratic complexity with the number of input tokens, causing a major efficiency bottleneck to ViTs' achievable runtime (i.e., inference latency) [3, 8, 32, 58, 66, 69, 76]. To mitigate this issue, linear attention designs have been developed to alleviate the vanilla ViT attention's quadratic complexity. In particular, existing efforts can be categorized into two clusters: (1) ViTs with local attention by restricting the attention window size [38, 53], sharing the attention queries [2], or representing the attention queries/keys with low rank matrices [58]; and (2) ViTs with kernel-based linear attention, which approximate the non-linearity softmax function by decomposing it into separate kernel embeddings. This enables a change in the matrix computation order for a reduced computational complexity [5, 6, 14, 29, 37, 39, 43, 66]." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 317, + 700, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 700, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 317, + 700, + 545, + 714 + ], + "type": "text", + "content": "Despite their promise in alleviating ViTs' complexity" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 702, + 279, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 702, + 279, + 712 + ], + "spans": [ + { + "bbox": [ + 58, + 702, + 279, + 712 + ], + "type": "text", + "content": "*Equal contribution. † Work done while interning at Meta Research." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14431" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 289, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 289, + 407 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 289, + 407 + ], + "type": "text", + "content": "and thus inference runtime, both the local and linear attention compromise ViTs' performance due to the lack of capabilities to capture global or local context. To marry the best of both worlds, we advocate training ViTs with both (1) efficient but less powerful linear attention, i.e., without the high-order residuals in angular kernel expansion, and (2) powerful yet costly softmax-based masked attention. The latter helps approximate high-order residuals at the early training stage while being dropped during inference, based on an assumption that the remaining networks can gradually learn the high-order components at the later training stage [67]. This concept resembles the \"castling\" move in chess when two pieces are moved at once. While it sounds promising, there are still two challenges to achieve this. First, existing linear attention modules still underperform their vanilla softmax-based counterparts. Therefore, a better linear attention is crucial for the final performance. We find that angular kernels perform equally as softmax-based attentions in terms of similarity measurements. While they still suffer from a quadratic complexity, they can be divided into linear terms and high-order residuals. The challenge is how to construct ViTs with only the linear terms. Second, doing so would require that the trained ViTs merely rely on the linear terms towards the end of training, which would call for an approximation of the above high-order residuals. The challenge is that how we can resort to costly but powerful modules to approximate high-order residuals during training but does not incur extra inference cost." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 407, + 288, + 433 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 407, + 288, + 433 + ], + "spans": [ + { + "bbox": [ + 47, + 407, + 288, + 433 + ], + "type": "text", + "content": "In this work, we develop techniques to tackle those challenges, and make the following contributions:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 59, + 441, + 287, + 704 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 59, + 441, + 287, + 501 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 441, + 287, + 501 + ], + "spans": [ + { + "bbox": [ + 59, + 441, + 287, + 501 + ], + "type": "text", + "content": "- We propose a framework called Castling-ViT, which trains ViTs using both linear-angular attention and masked softmax-based quadratic attention, but then switches to having only linear-angular attentions during ViT inference to save computational costs." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 59, + 505, + 287, + 566 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 505, + 287, + 566 + ], + "spans": [ + { + "bbox": [ + 59, + 505, + 287, + 566 + ], + "type": "text", + "content": "- We develop a new linear-angular attention leveraging angular kernels to close the accuracy gap between linear attention and softmax-based attention. It expands angular kernels where linear terms are kept while complex high-order residuals are approximated." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 59, + 569, + 287, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 569, + 287, + 629 + ], + "spans": [ + { + "bbox": [ + 59, + 569, + 287, + 629 + ], + "type": "text", + "content": "- We use two parameterized modules to approximate the high-order residuals above: a depthwise convolution and an auxiliary masked softmax-based attention, where the latter's attention masks are regularized to gradually become zeros to avoid inference overhead." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 59, + 632, + 287, + 704 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 632, + 287, + 704 + ], + "spans": [ + { + "bbox": [ + 59, + 632, + 287, + 704 + ], + "type": "text", + "content": "- We conduct extensive experiments to validate the effectiveness of the proposed Castling-ViT. Results on classification, detection, and segmentation tasks consistently demonstrate its superior performance (" + }, + { + "bbox": [ + 59, + 632, + 287, + 704 + ], + "type": "inline_equation", + "content": "\\uparrow 1.8\\%" + }, + { + "bbox": [ + 59, + 632, + 287, + 704 + ], + "type": "text", + "content": " top-1 accuracy or " + }, + { + "bbox": [ + 59, + 632, + 287, + 704 + ], + "type": "inline_equation", + "content": "\\uparrow 1.2" + }, + { + "bbox": [ + 59, + 632, + 287, + 704 + ], + "type": "text", + "content": " mAP) or efficiency (40% MACs savings) over state-of-the-art (SOTA) CNNs and ViTs." + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 71, + 398, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 71, + 398, + 84 + ], + "spans": [ + { + "bbox": [ + 306, + 71, + 398, + 84 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 89, + 547, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 89, + 547, + 413 + ], + "spans": [ + { + "bbox": [ + 304, + 89, + 547, + 413 + ], + "type": "text", + "content": "Vision Transformers (ViTs). ViT [20] beats CNNs with a simple encoder-only transformer architecture taking the splitted non-overlapped image patches as sequential inputs, but relies on costly pretraining on a huge JFT-300M dataset [50]. Later, DeiT [54] and T2T-ViT [72] leverage an improved ViT training recipe or enhanced tokenization mechanism to achieve a comparable accuracy without the necessity of costly pretraining. To further improve ViTs' achievable accuracy-efficiency trade-offs, CrossViT [9], PiT [26], PVT [59], MViT [21] and Swin-Transformer [38] propose a pyramid-like architecture, which is commonly used in CNNs [16, 28, 65]; DynamicViT [46], A-ViT [68], ToME [4], and MIA-Former [71] propose to adaptively identify and remove unnecessary input tokens for saving computational costs. With the goal of deploying ViTs in resource-constrained devices, various efficient ViT architectures have been proposed [6, 24, 35, 40]. For example, LeViT [24], CvT [63], and MobileViT [40] adopt more efficient self-attention implementation or incorporate convolutional feature extraction blocks into their early layers; EfficientFormer [35] further enables pure ViTs to run as fast as MobileNets. In contrast, our Castling-ViT explores whether ViTs can learn both global and local features while still being efficient at runtime. Also, we target a generic linear-angular attention that can serve as a drop-in replacement for all kinds of ViT architectures and thus is orthogonal to new ViT architecture designs." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 415, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 415, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 415, + 547, + 715 + ], + "type": "text", + "content": "Efficient ViT Variants. As commonly recognized, ViTs rely heavily on their self-attention module which is however costly due to its quadratic computational complexity with the total number of input tokens [29, 76]. To make the self-attention module more efficient, a surge of linear attention works have been proposed and can be roughly categorized into two groups: local attention [2, 38, 56, 58] or kernel-based linear attention [1, 5, 6, 14, 29, 37, 39, 43, 66]. For kernel-based linear attention, common designs approximate the softmax function [5, 14, 29] or the full self-attention matrix [39, 66] with orthogonal features or kernel embeddings, then the computation order can be changed from (QK)V to Q(KV). For example, [29] and [6] decompose the exponential terms in softmax-based attention into kernel functions and exchange the computation order. Despite their decent performance, currently kernel-based linear attention in general underperform the softmax-based attention. Recent works [8, 17] also unify low rank approximated and sparse attention (can also be dropped at inference) to improve ViTs' accuracy-efficiency tradeoffs. Different from the above works, we explore from a new perspective by taking spectral angles into consideration when measuring the similarities among tokens, resulting in linear-angular attention that can achieve comparable or even better performance than softmax-based attention. More efficient ViT variants" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "14432" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 168, + 83 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 168, + 83 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 168, + 83 + ], + "type": "text", + "content": "are supplied to the Appendix." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 84, + 287, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 84, + 287, + 120 + ], + "spans": [ + { + "bbox": [ + 46, + 84, + 287, + 120 + ], + "type": "text", + "content": "ViTs have also been used as the backbones for downstream tasks, e.g., detection [34, 60, 77] and segmentation [11,42]. We supply more literature review to Appendix." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 131, + 181, + 144 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 131, + 181, + 144 + ], + "spans": [ + { + "bbox": [ + 47, + 131, + 181, + 144 + ], + "type": "text", + "content": "3. The Proposed Methods" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 151, + 206, + 163 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 151, + 206, + 163 + ], + "spans": [ + { + "bbox": [ + 47, + 151, + 206, + 163 + ], + "type": "text", + "content": "3.1. Preliminary of Self-Attention" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 166, + 288, + 226 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 166, + 288, + 226 + ], + "spans": [ + { + "bbox": [ + 46, + 166, + 288, + 226 + ], + "type": "text", + "content": "Self-Attention. Self-attention module is a core component of the Transformer [20, 57], and usually consists of multiple heads. Each head captures global-context information by measuring pairwise correlations among all " + }, + { + "bbox": [ + 46, + 166, + 288, + 226 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 166, + 288, + 226 + ], + "type": "text", + "content": " tokens (" + }, + { + "bbox": [ + 46, + 166, + 288, + 226 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 166, + 288, + 226 + ], + "type": "text", + "content": " denotes the total number of tokens) as defined below:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 80, + 231, + 287, + 265 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 80, + 231, + 287, + 265 + ], + "spans": [ + { + "bbox": [ + 80, + 231, + 287, + 265 + ], + "type": "interline_equation", + "content": "\\mathbf {H} _ {t} ^ {k} = \\sum_ {i = 1} ^ {N} \\frac {\\exp \\left(\\mathbf {Q} _ {t} \\cdot \\mathbf {K} _ {i} ^ {T} / \\sqrt {d}\\right)}{\\sum_ {j = 1} ^ {N} \\exp \\left(\\mathbf {Q} _ {t} \\cdot \\mathbf {K} _ {j} ^ {T} / \\sqrt {d}\\right)} \\cdot \\mathbf {V} _ {i}, \\tag {1}", + "image_path": "47f3b89551c68473d9db30da5009d9edc7e6173bbbcf8321984ae11bc6df104f.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 270, + 287, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 270, + 287, + 426 + ], + "spans": [ + { + "bbox": [ + 46, + 270, + 287, + 426 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 270, + 287, + 426 + ], + "type": "inline_equation", + "content": "t\\in \\{1,\\dots ,N\\} ,k\\in \\{1,\\dots ,M\\}" + }, + { + "bbox": [ + 46, + 270, + 287, + 426 + ], + "type": "inline_equation", + "content": "\\mathbf{H}_t^k" + }, + { + "bbox": [ + 46, + 270, + 287, + 426 + ], + "type": "text", + "content": " refers to the " + }, + { + "bbox": [ + 46, + 270, + 287, + 426 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 270, + 287, + 426 + ], + "type": "text", + "content": " th row of the " + }, + { + "bbox": [ + 46, + 270, + 287, + 426 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 46, + 270, + 287, + 426 + ], + "type": "text", + "content": " -th head's attention matrix H. " + }, + { + "bbox": [ + 46, + 270, + 287, + 426 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}_t,\\mathbf{K}_t,\\mathbf{V}_t\\in" + }, + { + "bbox": [ + 46, + 270, + 287, + 426 + ], + "type": "inline_equation", + "content": "\\mathbf{R}^d" + }, + { + "bbox": [ + 46, + 270, + 287, + 426 + ], + "type": "text", + "content": " , are the query, key, and value vectors obtained by linearly projecting the input " + }, + { + "bbox": [ + 46, + 270, + 287, + 426 + ], + "type": "inline_equation", + "content": "\\mathbf{X_t}\\in \\mathbb{R}^D" + }, + { + "bbox": [ + 46, + 270, + 287, + 426 + ], + "type": "text", + "content": " with three learnable weight matrices " + }, + { + "bbox": [ + 46, + 270, + 287, + 426 + ], + "type": "inline_equation", + "content": "W^{Q},W^{K},W^{V}\\in \\mathbb{R}^{D\\times d}" + }, + { + "bbox": [ + 46, + 270, + 287, + 426 + ], + "type": "text", + "content": " . The attention head first computes the inner product between the query-key pairs, then scales the product results to stabilize the training and uses Softmax to normalize the resulting attention scores, and finally computes a weighted sum over all value vectors. Outputs from all attention heads are then concatenated together before a final linear projection with learnable weights. Note that " + }, + { + "bbox": [ + 46, + 270, + 287, + 426 + ], + "type": "inline_equation", + "content": "\\exp (\\cdot)" + }, + { + "bbox": [ + 46, + 270, + 287, + 426 + ], + "type": "text", + "content": " denotes an exponential function. Computing Eq. (1) has a quadratic complexity of " + }, + { + "bbox": [ + 46, + 270, + 287, + 426 + ], + "type": "inline_equation", + "content": "\\mathcal{O}(N^2)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 426, + 288, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 426, + 288, + 559 + ], + "spans": [ + { + "bbox": [ + 46, + 426, + 288, + 559 + ], + "type": "text", + "content": "Kernel-based Linear Attention. The core idea of linear attention [29, 49, 58] is to decompose the similarity measurement function into separate kernel embeddings, i.e., " + }, + { + "bbox": [ + 46, + 426, + 288, + 559 + ], + "type": "inline_equation", + "content": "\\mathrm{Sim}(\\mathbf{Q},\\mathbf{K})\\approx \\phi (\\mathbf{Q})\\phi (\\mathbf{K})^T" + }, + { + "bbox": [ + 46, + 426, + 288, + 559 + ], + "type": "text", + "content": " so that we can change the computation order to " + }, + { + "bbox": [ + 46, + 426, + 288, + 559 + ], + "type": "inline_equation", + "content": "\\phi (\\mathbf{Q})(\\phi (\\mathbf{K})^T\\mathbf{V})" + }, + { + "bbox": [ + 46, + 426, + 288, + 559 + ], + "type": "text", + "content": " based on the associative property of matrix multiplication. In this way, the attention complexity is quadratic to the feature dimension " + }, + { + "bbox": [ + 46, + 426, + 288, + 559 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 46, + 426, + 288, + 559 + ], + "type": "text", + "content": " instead of the token length " + }, + { + "bbox": [ + 46, + 426, + 288, + 559 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 426, + 288, + 559 + ], + "type": "text", + "content": ". One straightforward implementation of linear attention is to use Gaussian RBF kernels to measure the similarity, which can serve as an unbiased estimation of " + }, + { + "bbox": [ + 46, + 426, + 288, + 559 + ], + "type": "inline_equation", + "content": "\\exp (\\langle \\cdot ,\\cdot \\rangle)" + }, + { + "bbox": [ + 46, + 426, + 288, + 559 + ], + "type": "text", + "content": " in Eq. (1):" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 52, + 563, + 287, + 602 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 52, + 563, + 287, + 602 + ], + "spans": [ + { + "bbox": [ + 52, + 563, + 287, + 602 + ], + "type": "interline_equation", + "content": "\\exp \\left(\\frac {\\langle \\mathbf {x} , \\mathbf {y} \\rangle}{\\sigma^ {2}}\\right) = \\underbrace {\\exp \\left(\\frac {- \\| \\mathbf {x} - \\mathbf {y} \\| ^ {2}}{2 \\sigma^ {2}}\\right)} _ {\\text {G a u s s i a n R B F K e r n e l}} \\exp \\left(\\frac {\\| \\mathbf {x} \\| ^ {2} + \\| \\mathbf {y} \\| ^ {2}}{2 \\sigma^ {2}}\\right) \\tag {2}", + "image_path": "daff120b598875046cdb97cc233ce5e3cb0ad57bfce72d979638f0664d139609.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 607, + 287, + 680 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 607, + 287, + 680 + ], + "spans": [ + { + "bbox": [ + 46, + 607, + 287, + 680 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 607, + 287, + 680 + ], + "type": "inline_equation", + "content": "\\langle \\cdot, \\cdot \\rangle" + }, + { + "bbox": [ + 46, + 607, + 287, + 680 + ], + "type": "text", + "content": " denotes the inner product operator. According to [45], we can induce a function " + }, + { + "bbox": [ + 46, + 607, + 287, + 680 + ], + "type": "inline_equation", + "content": "\\phi(\\cdot)" + }, + { + "bbox": [ + 46, + 607, + 287, + 680 + ], + "type": "text", + "content": " to approximate the Gaussian RBF kernel, mapping the input space to the feature space. Assuming that both " + }, + { + "bbox": [ + 46, + 607, + 287, + 680 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}" + }, + { + "bbox": [ + 46, + 607, + 287, + 680 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 607, + 287, + 680 + ], + "type": "inline_equation", + "content": "\\mathbf{K}" + }, + { + "bbox": [ + 46, + 607, + 287, + 680 + ], + "type": "text", + "content": " are normalized as unit row vectors along the feature dimension, then the attention formula can be approximated by:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 93, + 684, + 287, + 717 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 93, + 684, + 287, + 717 + ], + "spans": [ + { + "bbox": [ + 93, + 684, + 287, + 717 + ], + "type": "interline_equation", + "content": "\\mathbf {H} _ {t} ^ {k} = \\frac {\\phi (\\mathbf {Q} _ {t}) \\cdot \\sum_ {i = 1} ^ {N} \\left(\\phi \\left(\\mathbf {K} _ {i}\\right) ^ {T} \\cdot \\mathbf {V} _ {i}\\right)}{\\phi (\\mathbf {Q} _ {t}) \\cdot \\sum_ {j = 1} ^ {N} \\phi \\left(\\mathbf {K} _ {j} ^ {T}\\right)} \\tag {3}", + "image_path": "f9c0d131b073be1bf1a36e0cd3c52332be7906b946787242b014b4d76973b2f1.jpg" + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 307, + 70, + 547, + 245 + ], + "blocks": [ + { + "bbox": [ + 307, + 70, + 547, + 245 + ], + "lines": [ + { + "bbox": [ + 307, + 70, + 547, + 245 + ], + "spans": [ + { + "bbox": [ + 307, + 70, + 547, + 245 + ], + "type": "image", + "image_path": "27ef88097b699bd0084f362693ea21f7f8ebf624aff2d6eee1f482d95dfa630f.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 306, + 247, + 545, + 259 + ], + "lines": [ + { + "bbox": [ + 306, + 247, + 545, + 259 + ], + "spans": [ + { + "bbox": [ + 306, + 247, + 545, + 259 + ], + "type": "text", + "content": "Figure 2. Illustration of self-attention used in LeViT and MViTv2." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 266, + 545, + 314 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 266, + 545, + 314 + ], + "spans": [ + { + "bbox": [ + 304, + 266, + 545, + 314 + ], + "type": "text", + "content": "While doing so can alleviate the attention complexity to become linear w.r.t. the token length " + }, + { + "bbox": [ + 304, + 266, + 545, + 314 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 266, + 545, + 314 + ], + "type": "text", + "content": ", it often causes a nontrivial accuracy drop as compared to the corresponding ViTs with vanilla softmax-based attention [6, 13, 72]." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 316, + 546, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 316, + 546, + 579 + ], + "spans": [ + { + "bbox": [ + 304, + 316, + 546, + 579 + ], + "type": "text", + "content": "Revisit Attention Designs. Recent ViTs have achieved a low complexity for classification tasks (e.g., LeViT [24]) but still be costly for downstream tasks due to high input resolutions, i.e., the number of tokens (e.g., MViTv2 [34]). Therefore, one dilemma is the tradeoff between the model efficiency and generalizability. As illustrated in Fig. 2, LeViT shrinks the feature dimension to be more efficient on low-resolution tasks but cannot be well generalized to downstream tasks, due to the resulting (1) feature bottlenecks, i.e., insufficient feature dimensions, and (2) fixed attention biases. On the other hand, MViTv2 performs token pooling for better fitting high-resolution tasks but is less effective for low-resolution due to the caused token bottlenecks, i.e., insufficient number of tokens. Our ablation studies in Table 1 show that (1) for ImageNet classification, both token and feature pooling can lead to accuracy drops in ViTs; and (2) for downsampling layers, post-Q pooling (i.e., pooling after linear projection) together with residual connections in MViTv2 performs better than pre-Q pooling (i.e., pooling during linear projection) in LeViT. Therefore, we develop Castling-LeViTs on top of attention with merely post-Q pooling and residual connections." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 588, + 515, + 601 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 588, + 515, + 601 + ], + "spans": [ + { + "bbox": [ + 306, + 588, + 515, + 601 + ], + "type": "text", + "content": "3.2. The Proposed Castling-ViT Framework" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": "Castling-ViT Overview. Fig. 4 illustrates an overview of the proposed Castling-ViT, which makes linear attention more powerful than previous designs while still being efficient during inference. In particular, we propose (1) a novel kernel-based linear-angular attention from the spectral angle perspective to close the accuracy gap between linear attention and softmax-based attention; and (2) a training augmentation method that leverages softmax-based attention as an auxiliary branch to assist the linear-angular atten" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14433" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 49, + 83, + 287, + 159 + ], + "blocks": [ + { + "bbox": [ + 48, + 71, + 285, + 81 + ], + "lines": [ + { + "bbox": [ + 48, + 71, + 285, + 81 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 285, + 81 + ], + "type": "text", + "content": "Table 1. Analyze the attention design on ImageNet classification." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 49, + 83, + 287, + 159 + ], + "lines": [ + { + "bbox": [ + 49, + 83, + 287, + 159 + ], + "spans": [ + { + "bbox": [ + 49, + 83, + 287, + 159 + ], + "type": "table", + "html": "
PoolingResidual QMACs (M)Top-1 Accuracy (%)
TokenFeat.Pre-QPost-Q
✓✓68578.00
✓✓66178.19
✓73679.11
✓✓✓77178.73
✓✓✓76377.23
✓✓83880.05
", + "image_path": "bbb9916f1a67e1f02942854ed649fd31c72638e0d41d4465e5a5d08b4af6b2bf.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 167, + 287, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 167, + 287, + 204 + ], + "spans": [ + { + "bbox": [ + 47, + 167, + 287, + 204 + ], + "type": "text", + "content": "tion only during ViT training. Note that a mask is applied to the auxiliary branch to manifest linear-angular attention and drop the auxiliary branch." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 211, + 191, + 223 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 211, + 191, + 223 + ], + "spans": [ + { + "bbox": [ + 47, + 211, + 191, + 223 + ], + "type": "text", + "content": "3.2.1 Linear-Angular Attention" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 227, + 287, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 227, + 287, + 335 + ], + "spans": [ + { + "bbox": [ + 46, + 227, + 287, + 335 + ], + "type": "text", + "content": "Angular Kernel. In addition to the previously adopted polynomial, exponential, or RBF kernel that focuses on spatial similarity measurements [29, 37], we propose to consider measuring spectral similarity via angular kernel as an alternative to existing softmax-based attentions, leading to similar or better performance since it additionally takes into consideration the nature of spectral characteristics, e.g., the spectral angle as a distance measurement function [27, 30]. Such a spectral angle between two vectors is defined as:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 92, + 339, + 287, + 366 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 92, + 339, + 287, + 366 + ], + "spans": [ + { + "bbox": [ + 92, + 339, + 287, + 366 + ], + "type": "interline_equation", + "content": "\\theta \\left(\\mathbf {x} _ {i}, \\mathbf {x} _ {j}\\right) = \\arccos \\left(\\frac {\\left\\langle \\mathbf {x} _ {i} , \\mathbf {x} _ {j} \\right\\rangle}{\\left\\| \\mathbf {x} _ {i} \\right\\| \\cdot \\left\\| \\mathbf {x} _ {j} \\right\\|}\\right), \\tag {4}", + "image_path": "64f1368e6a6ce7f450c70c62687d90c30f0b747d0478ec71e06792cd44f640d2.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 370, + 287, + 431 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 370, + 287, + 431 + ], + "spans": [ + { + "bbox": [ + 46, + 370, + 287, + 431 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 370, + 287, + 431 + ], + "type": "inline_equation", + "content": "\\|\\cdot\\|" + }, + { + "bbox": [ + 46, + 370, + 287, + 431 + ], + "type": "text", + "content": " is the Euclidean distance and " + }, + { + "bbox": [ + 46, + 370, + 287, + 431 + ], + "type": "inline_equation", + "content": "\\langle \\cdot, \\cdot \\rangle" + }, + { + "bbox": [ + 46, + 370, + 287, + 431 + ], + "type": "text", + "content": " is the inner product. The output range of " + }, + { + "bbox": [ + 46, + 370, + 287, + 431 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 46, + 370, + 287, + 431 + ], + "type": "text", + "content": " is " + }, + { + "bbox": [ + 46, + 370, + 287, + 431 + ], + "type": "inline_equation", + "content": "[0, \\pi]" + }, + { + "bbox": [ + 46, + 370, + 287, + 431 + ], + "type": "text", + "content": ". Such an angle can be used as a distance. In our design, we define the angular kernel as a similarity measurement function between the queries " + }, + { + "bbox": [ + 46, + 370, + 287, + 431 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}" + }, + { + "bbox": [ + 46, + 370, + 287, + 431 + ], + "type": "text", + "content": " and keys " + }, + { + "bbox": [ + 46, + 370, + 287, + 431 + ], + "type": "inline_equation", + "content": "\\mathbf{K}" + }, + { + "bbox": [ + 46, + 370, + 287, + 431 + ], + "type": "text", + "content": " as:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 93, + 434, + 287, + 458 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 93, + 434, + 287, + 458 + ], + "spans": [ + { + "bbox": [ + 93, + 434, + 287, + 458 + ], + "type": "interline_equation", + "content": "\\operatorname {S i m} \\left(\\mathbf {Q} _ {i}, \\mathbf {K} _ {j}\\right) = 1 - \\frac {1}{\\pi} \\cdot \\theta \\left(\\mathbf {Q} _ {i}, \\mathbf {K} _ {j}\\right), \\tag {5}", + "image_path": "2a7dd36c4f1fb17226d9f7356417378334e93706dcca60ed1439c008b3ab8ee0.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 462, + 287, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 462, + 287, + 510 + ], + "spans": [ + { + "bbox": [ + 47, + 462, + 287, + 510 + ], + "type": "text", + "content": "and the output range is thus [0, 1]. With more aligned " + }, + { + "bbox": [ + 47, + 462, + 287, + 510 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}_i" + }, + { + "bbox": [ + 47, + 462, + 287, + 510 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 462, + 287, + 510 + ], + "type": "inline_equation", + "content": "\\mathbf{K}_j" + }, + { + "bbox": [ + 47, + 462, + 287, + 510 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 47, + 462, + 287, + 510 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 47, + 462, + 287, + 510 + ], + "type": "text", + "content": " is closer to 0 and thus the similarity is closer to 1; In contrast, if " + }, + { + "bbox": [ + 47, + 462, + 287, + 510 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}_i" + }, + { + "bbox": [ + 47, + 462, + 287, + 510 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 462, + 287, + 510 + ], + "type": "inline_equation", + "content": "\\mathbf{K}_j" + }, + { + "bbox": [ + 47, + 462, + 287, + 510 + ], + "type": "text", + "content": " have opposite features, " + }, + { + "bbox": [ + 47, + 462, + 287, + 510 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 47, + 462, + 287, + 510 + ], + "type": "text", + "content": " is closer to " + }, + { + "bbox": [ + 47, + 462, + 287, + 510 + ], + "type": "inline_equation", + "content": "\\pi" + }, + { + "bbox": [ + 47, + 462, + 287, + 510 + ], + "type": "text", + "content": " and thus the similarity is closer to 0." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 510, + 287, + 606 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 510, + 287, + 606 + ], + "spans": [ + { + "bbox": [ + 46, + 510, + 287, + 606 + ], + "type": "text", + "content": "Properties of Angular Kernel and Its Feature Space. One property of our angular kernel is that replacing the similarity measurement in self-attention with such a kernel provides an efficient way to implicitly map the input data to a high (even infinite after expansion) dimensional feature space [27], where the distances/angles are calculated based on a rich feature structure. Let " + }, + { + "bbox": [ + 46, + 510, + 287, + 606 + ], + "type": "inline_equation", + "content": "\\phi(\\cdot)" + }, + { + "bbox": [ + 46, + 510, + 287, + 606 + ], + "type": "text", + "content": " denotes the implicit map induced by this kernel, the norm of mapped input data is:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 80, + 610, + 287, + 633 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 80, + 610, + 287, + 633 + ], + "spans": [ + { + "bbox": [ + 80, + 610, + 287, + 633 + ], + "type": "interline_equation", + "content": "\\left\\| \\phi \\left(\\mathbf {x} _ {i}\\right) \\right\\| ^ {2} = \\operatorname {S i m} \\left(\\mathbf {x} _ {i}, \\mathbf {x} _ {i}\\right) = 1 - \\frac {1}{\\pi} \\cdot 0 = 1, \\tag {6}", + "image_path": "b18c281786c88b37d6cfe5e38f0fefd8533cdf9c447f38bbb8d15f88c039fe2c.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 637, + 287, + 673 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 637, + 287, + 673 + ], + "spans": [ + { + "bbox": [ + 47, + 637, + 287, + 673 + ], + "type": "text", + "content": "which means that all data in the input space are mapped onto the sphere of radius 1 in the feature space. Also, the distance between two input features is given by:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 62, + 677, + 287, + 717 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 677, + 287, + 717 + ], + "spans": [ + { + "bbox": [ + 62, + 677, + 287, + 717 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\left\\| \\phi (\\mathbf {x} _ {i}) - \\phi (\\mathbf {x} _ {j}) \\right\\| ^ {2} = 1 + 1 - 2 \\cdot \\phi (\\mathbf {x} _ {i}) \\phi (\\mathbf {x} _ {j}) ^ {T} (7) \\\\ = 2 \\cdot \\left(1 - \\operatorname {S i m} \\left(\\mathbf {x} _ {i}, \\mathbf {x} _ {j}\\right)\\right) = \\frac {2}{\\pi} \\cdot \\theta \\left(\\mathbf {x} _ {i}, \\mathbf {x} _ {j}\\right). (7) \\\\ \\end{array}", + "image_path": "2ff405178e20427c9926a8030d2e472d12c3ba5d927ebe903d69b7bbcc1324ff.jpg" + } + ] + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 329, + 76, + 386, + 125 + ], + "blocks": [ + { + "bbox": [ + 329, + 76, + 386, + 125 + ], + "lines": [ + { + "bbox": [ + 329, + 76, + 386, + 125 + ], + "spans": [ + { + "bbox": [ + 329, + 76, + 386, + 125 + ], + "type": "image", + "image_path": "fae8d3265027027362ce0f59815be2c815433d2e751bb7501aaa5a537dd1f64b.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 334, + 129, + 378, + 137 + ], + "lines": [ + { + "bbox": [ + 334, + 129, + 378, + 137 + ], + "spans": [ + { + "bbox": [ + 334, + 129, + 378, + 137 + ], + "type": "text", + "content": "(a) Input space" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 406, + 69, + 521, + 127 + ], + "blocks": [ + { + "bbox": [ + 406, + 69, + 521, + 127 + ], + "lines": [ + { + "bbox": [ + 406, + 69, + 521, + 127 + ], + "spans": [ + { + "bbox": [ + 406, + 69, + 521, + 127 + ], + "type": "image", + "image_path": "8dde8b37267758b9431238ed8673c7d607d39af714e6d08bb0c4afbaf7b1e2ca.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 430, + 129, + 482, + 137 + ], + "lines": [ + { + "bbox": [ + 430, + 129, + 482, + 137 + ], + "spans": [ + { + "bbox": [ + 430, + 129, + 482, + 137 + ], + "type": "text", + "content": "(b) Feature space" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 305, + 138, + 545, + 171 + ], + "lines": [ + { + "bbox": [ + 305, + 138, + 545, + 171 + ], + "spans": [ + { + "bbox": [ + 305, + 138, + 545, + 171 + ], + "type": "text", + "content": "Figure 3. (a) A simple 2D input space and (b) the corresponding 3D feature space after applying the angular kernel to the 2D input data. Note that real input data/features are of higher dimensions." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "bbox": [ + 305, + 176, + 545, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 176, + 545, + 224 + ], + "spans": [ + { + "bbox": [ + 305, + 176, + 545, + 224 + ], + "type": "text", + "content": "That is, the square (Euclid) distance and the spectral angle is positively correlated and the distance range is " + }, + { + "bbox": [ + 305, + 176, + 545, + 224 + ], + "type": "inline_equation", + "content": "[0, 2]" + }, + { + "bbox": [ + 305, + 176, + 545, + 224 + ], + "type": "text", + "content": ". As shown in Fig. 3, the angles in the input space are correlated to the feature distance after applying our angular kernel." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 305, + 224, + 545, + 296 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 224, + 545, + 296 + ], + "spans": [ + { + "bbox": [ + 305, + 224, + 545, + 296 + ], + "type": "text", + "content": "Expansion of Angular Kernel. A natural following question is how to incorporate the above angular kernel for designing linear attention given its quadratic complexity w.r.t. the input token length. Recalling from trigonometric identities and the expansion of the arccos function into an infinite series, we reformulate the similarity function as:" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 315, + 301, + 545, + 416 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 301, + 545, + 416 + ], + "spans": [ + { + "bbox": [ + 315, + 301, + 545, + 416 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\operatorname {S i m} \\left(\\mathbf {Q} _ {i}, \\mathbf {K} _ {j}\\right) = 1 - \\frac {1}{\\pi} \\cdot \\left(\\frac {\\pi}{2} - \\arcsin \\left(\\mathbf {Q} _ {i} \\cdot \\mathbf {K} _ {j} ^ {T}\\right)\\right) \\\\ = \\underbrace {\\frac {1}{2} + \\frac {1}{\\pi} \\cdot \\left(\\mathbf {Q} _ {i} \\cdot \\mathbf {K} _ {j} ^ {T}\\right)} _ {\\text {L i n e a r - a n g u l a r t e r m s}} \\tag {8} \\\\ + \\underbrace {\\frac {1}{\\pi} \\cdot \\sum_ {k = 1} ^ {\\infty} \\frac {(2 k) !}{2 ^ {2 k} (k !) ^ {2} (2 k + 1)} \\left(\\mathbf {Q} _ {i} \\cdot \\mathbf {K} _ {j} ^ {T}\\right) ^ {2 k + 1}} _ {\\text {H i g h - o r d e r r e s i d u a l s w h e n k \\geq 1 (n o n - l i n e a r)}}, \\\\ \\end{array}", + "image_path": "c1fa4242ca7e149271050c94e3a6e0c0ae5455132401fed598914f84e06848b4.jpg" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 305, + 423, + 545, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 423, + 545, + 518 + ], + "spans": [ + { + "bbox": [ + 305, + 423, + 545, + 518 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 423, + 545, + 518 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}_i\\cdot \\mathbf{K}_j^T" + }, + { + "bbox": [ + 305, + 423, + 545, + 518 + ], + "type": "text", + "content": " denotes the normalized linear kernel function " + }, + { + "bbox": [ + 305, + 423, + 545, + 518 + ], + "type": "inline_equation", + "content": "\\langle \\mathbf{Q}_i,\\mathbf{K}_j\\rangle /\\| \\mathbf{Q}_i\\| \\cdot \\| \\mathbf{K}_j\\|" + }, + { + "bbox": [ + 305, + 423, + 545, + 518 + ], + "type": "text", + "content": ", which is equivalent to the inner product if " + }, + { + "bbox": [ + 305, + 423, + 545, + 518 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}_i" + }, + { + "bbox": [ + 305, + 423, + 545, + 518 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 423, + 545, + 518 + ], + "type": "inline_equation", + "content": "\\mathbf{K}_j" + }, + { + "bbox": [ + 305, + 423, + 545, + 518 + ], + "type": "text", + "content": " are unit vectors. We see that the first linear-angular terms can be directly used as the similarity measurement in linear attention, while the remaining higher-order terms of an infinite series introduce a much higher complexity. As such, we propose to adopt a relaxation to approximate it." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 304, + 519, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 519, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 519, + 545, + 713 + ], + "type": "text", + "content": "Linear-Angular Attention. To construction our linear- angular attention, we leverage parametrized DNN modules to approximate the expectation of the high-order residuals, i.e., " + }, + { + "bbox": [ + 304, + 519, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_{ij} = \\mathbb{E}_{\\mathbf{Q}_i\\sim p_{Q_i},\\mathbf{K}_j\\sim p_{K_j}}\\left(\\sum_{k = 1}^{\\infty}\\alpha_k(\\mathbf{Q}_i\\cdot \\mathbf{K}_j^T)^{2k + 1}\\right)" + }, + { + "bbox": [ + 304, + 519, + 545, + 713 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 304, + 519, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\alpha_{k}" + }, + { + "bbox": [ + 304, + 519, + 545, + 713 + ], + "type": "text", + "content": " is the coefficient for the " + }, + { + "bbox": [ + 304, + 519, + 545, + 713 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 519, + 545, + 713 + ], + "type": "text", + "content": " -th order term in Eq. (8). The key question is how to design such DNN modules. Kim et al. [31] conducted an analysis on average attention weights of ViT models, which show a strong inductive bias to attend to neighboring tokens. To capture this, we introduce a learnable depthwise convolution (DWConv) module which is applied on all the value tokens, so " + }, + { + "bbox": [ + 304, + 519, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_{ij}" + }, + { + "bbox": [ + 304, + 519, + 545, + 713 + ], + "type": "text", + "content": " by-design attends to nearby tokens, as illustrated in Fig. 4. Second, DWConv is limited by its receptive field [41] while the averaged attention map also has a small number of offdiagonal scores to connect to nonadjacent tokens [31]. To capture that, we adopt a sparse softmax-based attention as" + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14434" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 48, + 70, + 287, + 277 + ], + "blocks": [ + { + "bbox": [ + 48, + 70, + 287, + 277 + ], + "lines": [ + { + "bbox": [ + 48, + 70, + 287, + 277 + ], + "spans": [ + { + "bbox": [ + 48, + 70, + 287, + 277 + ], + "type": "image", + "image_path": "ec02e87fdc198761e8fc24ad61e5ff865315daa98184d7b07602f7e5ef3e52ea.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 47, + 278, + 287, + 301 + ], + "lines": [ + { + "bbox": [ + 47, + 278, + 287, + 301 + ], + "spans": [ + { + "bbox": [ + 47, + 278, + 287, + 301 + ], + "type": "text", + "content": "Figure 4. The visualization of our Castling-ViT with both linear-angular attention and auxiliary softmax-based self-attention." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 309, + 288, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 309, + 288, + 357 + ], + "spans": [ + { + "bbox": [ + 46, + 309, + 288, + 357 + ], + "type": "text", + "content": "described in Section 3.2.2. In practice, we found that such a sparse softmax attention tends to converge to all zeros after adding a simple threshold regularization. We provide more explanation and visualization in Sec. 3.2.2 and Sec. 4.5." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 59, + 357, + 282, + 369 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 357, + 282, + 369 + ], + "spans": [ + { + "bbox": [ + 59, + 357, + 282, + 369 + ], + "type": "text", + "content": "In this way, our attention module can be formulated as:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 58, + 374, + 287, + 412 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 374, + 287, + 412 + ], + "spans": [ + { + "bbox": [ + 58, + 374, + 287, + 412 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathbf {H} = \\operatorname {S i m} (\\mathbf {Q}, \\mathbf {K}) \\cdot \\mathbf {V} \\approx \\frac {1}{2} \\cdot \\mathbf {V} + \\frac {1}{\\pi} \\cdot \\mathbf {Q} \\cdot \\left(\\mathbf {K} ^ {T} \\cdot \\mathbf {V}\\right) \\tag {9} \\\\ + \\mathbf {M} _ {D W} \\cdot \\mathbf {V} + \\mathbf {M} _ {S p a r s e A t t n}, \\\\ \\end{array}", + "image_path": "4505c7be847931373d875c57a0d2d44f30e884ff980a15daf5860e0f94d3ecc1.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 418, + 287, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 418, + 287, + 515 + ], + "spans": [ + { + "bbox": [ + 46, + 418, + 287, + 515 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 418, + 287, + 515 + ], + "type": "inline_equation", + "content": "1 / \\pi \\cdot Q\\cdot (K^T\\cdot V)" + }, + { + "bbox": [ + 46, + 418, + 287, + 515 + ], + "type": "text", + "content": " is the linear term with " + }, + { + "bbox": [ + 46, + 418, + 287, + 515 + ], + "type": "inline_equation", + "content": "\\mathcal{O}(N)" + }, + { + "bbox": [ + 46, + 418, + 287, + 515 + ], + "type": "text", + "content": " complexity, " + }, + { + "bbox": [ + 46, + 418, + 287, + 515 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_{DW}\\cdot V" + }, + { + "bbox": [ + 46, + 418, + 287, + 515 + ], + "type": "text", + "content": " is the matrix form of DWConv, and " + }, + { + "bbox": [ + 46, + 418, + 287, + 515 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_{SparseAttn}" + }, + { + "bbox": [ + 46, + 418, + 287, + 515 + ], + "type": "text", + "content": " is the normalized sparse softmax attention. The overall complexity to compute Eq. (9) is linear to the input token length, where the MACs of DWConv is also negligible (e.g., " + }, + { + "bbox": [ + 46, + 418, + 287, + 515 + ], + "type": "inline_equation", + "content": "< 1\\%" + }, + { + "bbox": [ + 46, + 418, + 287, + 515 + ], + "type": "text", + "content": " of the total MACs). Also, normalizations " + }, + { + "bbox": [ + 46, + 418, + 287, + 515 + ], + "type": "inline_equation", + "content": "(\\mathrm{x} / \\| \\mathrm{x}\\|_{2})" + }, + { + "bbox": [ + 46, + 418, + 287, + 515 + ], + "type": "text", + "content": " are inserted to the " + }, + { + "bbox": [ + 46, + 418, + 287, + 515 + ], + "type": "inline_equation", + "content": "\\mathbf{Q} / \\mathbf{K}" + }, + { + "bbox": [ + 46, + 418, + 287, + 515 + ], + "type": "text", + "content": " and sparse attention branches to help the similarity measurement following [5]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 518, + 261, + 529 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 518, + 261, + 529 + ], + "spans": [ + { + "bbox": [ + 47, + 518, + 261, + 529 + ], + "type": "text", + "content": "3.2.2 Switch Towards Linear-Angular Attention" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 534, + 287, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 534, + 287, + 629 + ], + "spans": [ + { + "bbox": [ + 46, + 534, + 287, + 629 + ], + "type": "text", + "content": "Recall that we add a sparse softmax-based attention as an auxiliary branch in the linear-angular attention to help approximate the high-order residuals. Such a costly attention can be potentially dropped without hurting the inference accuracy, drawing inspiration from [67] that the remaining network can gradually learn the high-order/frequency components at the later training stage. Next, we explain how it is constructed and dropped." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 630, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 630, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 630, + 287, + 713 + ], + "type": "text", + "content": "Sparse Training Augmentation. As illustrated in Fig. 4, we adopt a masked softmax-based attention as an auxiliary branch to augment ViT training. In particular, we first use a comparator with a predefined threshold " + }, + { + "bbox": [ + 46, + 630, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 46, + 630, + 287, + 713 + ], + "type": "text", + "content": " to generate a binary mask, where attentions greater than " + }, + { + "bbox": [ + 46, + 630, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 46, + 630, + 287, + 713 + ], + "type": "text", + "content": " are set to 1 and 0 otherwise. These masks are then applied to the attention maps to generate masked attention maps that will be" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 72, + 545, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 545, + 96 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 545, + 96 + ], + "type": "text", + "content": "summed up with our linear-angular attention together as the final attention. The resulting sparse attention is given by:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 102, + 545, + 117 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 102, + 545, + 117 + ], + "spans": [ + { + "bbox": [ + 313, + 102, + 545, + 117 + ], + "type": "interline_equation", + "content": "\\mathbf {M} _ {\\text {S p a r s e A t t n}} (\\mathbf {Q}, \\mathbf {K}) = \\operatorname {M a s k} _ {\\varepsilon} \\left(\\operatorname {S o f t m a x} \\left(\\mathbf {Q} \\cdot \\mathbf {K} ^ {\\mathrm {T}}\\right)\\right). \\tag {10}", + "image_path": "4b1ae8c95754ab70eab5111a2c960e0bb7fa5619347803c83137adf220e9c75e.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 125, + 545, + 230 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 125, + 545, + 230 + ], + "spans": [ + { + "bbox": [ + 304, + 125, + 545, + 230 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 125, + 545, + 230 + ], + "type": "inline_equation", + "content": "\\mathsf{Mask}_{\\epsilon}(\\mathbf{x}) = \\mathbf{x}" + }, + { + "bbox": [ + 304, + 125, + 545, + 230 + ], + "type": "text", + "content": " if " + }, + { + "bbox": [ + 304, + 125, + 545, + 230 + ], + "type": "inline_equation", + "content": "\\mathbf{x} > \\epsilon" + }, + { + "bbox": [ + 304, + 125, + 545, + 230 + ], + "type": "text", + "content": " else 0, acting as an elementwise threshold function. As such, the sparse attention captures the higher attention scores (i.e., strong local features), and can potentially complement our linear-angular attention by supplementing the missing higher-order terms. Such an assumption aligns well with recent findings that (1) low-rank and sparse approximations complement each other [8, 17, 31]; and (2) linear attention lacks local feature extraction capabilities over its softmax-based counterpart [6]." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 231, + 546, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 231, + 546, + 422 + ], + "spans": [ + { + "bbox": [ + 304, + 231, + 546, + 422 + ], + "type": "text", + "content": "Castling During ViT Inference. As we are targeting efficient ViT inference, it is desired to reduce or completely remove the costly softmax-based attentions while only keeping our linear-angular attention at runtime, i.e., performing castling. In our experiments, we found that under the sparsity regularization above, the softmax attention naturally converges to all zeros as the training progresses. We consider both fixed and dynamic schedules for " + }, + { + "bbox": [ + 304, + 231, + 546, + 422 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 304, + 231, + 546, + 422 + ], + "type": "text", + "content": " and find that our Castling-ViT is not sensitive to neither the threshold value nor the threshold schedule for a given task. Given a fixed mask threshold (e.g., " + }, + { + "bbox": [ + 304, + 231, + 546, + 422 + ], + "type": "inline_equation", + "content": "\\epsilon = 0.02" + }, + { + "bbox": [ + 304, + 231, + 546, + 422 + ], + "type": "text", + "content": " in image classification experiments), the masks become all zeros at latter training stages and thus the auxiliary branch can be removed without hurting the model accuracy. We supply the visualization of the mask-evolving trajectory and our conjecture for understanding such a phenomenon in Sec. 4.5." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 429, + 388, + 442 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 429, + 388, + 442 + ], + "spans": [ + { + "bbox": [ + 305, + 429, + 388, + 442 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 445, + 424, + 459 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 445, + 424, + 459 + ], + "spans": [ + { + "bbox": [ + 306, + 445, + 424, + 459 + ], + "type": "text", + "content": "4.1. Experiment Settings" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 460, + 545, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 460, + 545, + 652 + ], + "spans": [ + { + "bbox": [ + 304, + 460, + 545, + 652 + ], + "type": "text", + "content": "Tasks, Datasets, and Models. Tasks and Datasets. We consider three benchmark datasets and three representative vision tasks to demonstrate the superiority of the proposed Castling-ViT, including image classification on ImageNet dataset [18] with 1.2 million training and 50K validation images; Object detection on COCO dataset [36] with 118K training and 5K validation images; Semantic segmentation on ADE20K dataset [75] with 20K/2K/3K images for training, validation, and testing, respectively. Models. We apply our proposed Castling-ViT idea on top of various models. For the classification task, we consider LeViT [24], MViTv2 [34], and DeiT [54]; For the detection task, we consider models with efficient ViT backbones (e.g., PicoDet [70] with modified ESNet and LCNet backbones [15] with transformer blocks); For the segmentation task, we consider Mask2former [11] with ViT-Base backbone." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 652, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 652, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 652, + 546, + 713 + ], + "type": "text", + "content": "Training Settings. For the classification task, we use a SGD optimizer with 0.9 momentum and " + }, + { + "bbox": [ + 304, + 652, + 546, + 713 + ], + "type": "inline_equation", + "content": "2 \\times 10^{-5}" + }, + { + "bbox": [ + 304, + 652, + 546, + 713 + ], + "type": "text", + "content": " weight decay to train ViTs for 1000 epochs using 64 V100 GPUs, with each card having 64 (LeViT) or 32 (MViTv2/DeiT) batch sizes. The learning rate is 2.0" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14435" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 50, + 83, + 285, + 440 + ], + "blocks": [ + { + "bbox": [ + 62, + 71, + 271, + 81 + ], + "lines": [ + { + "bbox": [ + 62, + 71, + 271, + 81 + ], + "spans": [ + { + "bbox": [ + 62, + 71, + 271, + 81 + ], + "type": "text", + "content": "Table 2. Castling-ViT over SOTA baselines on ImageNet." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 50, + 83, + 285, + 440 + ], + "lines": [ + { + "bbox": [ + 50, + 83, + 285, + 440 + ], + "spans": [ + { + "bbox": [ + 50, + 83, + 285, + 440 + ], + "type": "table", + "html": "
MACs RangesModelsParams (M)MACs (G)Top-1 Acc.(%)Top-5 Acc.(%)
<1GMobileNetV2-1.4 [47]6.90.5874.7-
EfficientNet-B0 [51]5.30.3977.193.3
EfficientNet-B1 [51]7.80.7079.194.4
MobileViT-XS [40]2.30.7074.892.3
LeViT-128 [24]9.20.4178.6-
LeViT-192 [24]10.90.6680.0-
Castling-LeViT-12810.50.4979.694.6
Castling-LeViT-19212.70.8281.395.5
1~3GEfficientNet-B3 [51]12.01.8081.695.7
HRFormer-T [73]8.01.8078.5-
DeiT-T [54]5.61.2574.591.9
LeViT-256 [24]18.91.1281.6-
LeViT-384 [24]39.12.3582.6-
Caslting-DeiT-T5.641.1876.092.5
Castling-LeViT-25622.01.4082.696.1
Castling-LeViT-38445.82.9083.796.7
3~10GRegNetY-4G [44]21.04.0080.0-
RegNetY-8G [44]39.08.0081.7-
EfficientNet-B4 [51]19.04.2082.996.4
EfficientNet-B5 [51]30.09.9083.696.7
Swin-T [38]29.04.5081.395.5
Swin-S [38]50.08.7083.296.2
CSWin-T [19]23.04.3082.7-
CSWin-S [19]35.06.9083.6-
DeiT-S [54]21.94.6081.295.4
AutoFormer-S [10]22.95.1081.795.7
PVTv2-V2 [61]25.04.0082.0-
MViTv2-T [34]24.04.6882.3-
MViTv2-S [34]34.76.9583.6-
Castling-MViTv2-T24.14.5084.196.8
Castling-MViTv2-S34.76.9584.697.0
>10GCaiT-S36 [55]68.013.9083.3-
Swin-B [38]88.015.4083.596.5
CSWin-B [19]78.015.0084.2-
AutoFormer-B [10]54.011.0082.495.7
MViTv2-B [34]51.210.0784.4-
DeiT-B [54]86.317.5683.496.5
Castling-DeiT-B87.2217.2884.2-
Castling-MViTv2-B51.99.8285.097.2
", + "image_path": "6cf2c60dafa4275a4fb2ccd33d5b0b932d477d6df0693ca0c09fcc6307ac5e9b.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 449, + 286, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 449, + 286, + 592 + ], + "spans": [ + { + "bbox": [ + 46, + 449, + 286, + 592 + ], + "type": "text", + "content": "with first 11 epochs warm-up starting from 0.01 and decays by a factor of 0.9875 per epoch [62]. Also, we use the distillation [54] based on a teacher model with a " + }, + { + "bbox": [ + 46, + 449, + 286, + 592 + ], + "type": "inline_equation", + "content": "85.5\\%" + }, + { + "bbox": [ + 46, + 449, + 286, + 592 + ], + "type": "text", + "content": " accuracy. For the detection task, we adopt SGD optimizer with momentum 0.9 and weight decay 4e-5 to train models on COCO. All models are trained on 8 V100 GPUs with each card having 80 batch sizes following PicoDet's training recipe [70]. Also, we follow LeViT's training recipe [24] to pretrain backbones on ImageNet; For the segmentation task, we follow Mask2former's training recipe [11] to train models on ADE20K, where ViT backbones are pretrained following MAE [25] if specified." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 594, + 286, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 286, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 286, + 714 + ], + "type": "text", + "content": "Baselines and Evaluation Metrics. Baselines. For the classification task, we compare the proposed Castling-ViT with LeViT [24], MviTv2 [34], DeiT [54], Swin [38], CSWin [19], PVT [59], etc. For the detection task, we compare with FBNetV5 [62], YOLOX [23], YOLOv5, MobileDet [65], and EfficientDet [52]. For the segmentation task, we compare with Mask2former [11] with ViT backbones. Evaluation Metrics. We evaluate the Castling-ViT and all baselines in terms of accuracy-efficiency tradeoffs. Specifically, the accuracy metrics refer to top-1/5 accuracy" + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 310, + 73, + 425, + 149 + ], + "blocks": [ + { + "bbox": [ + 310, + 73, + 425, + 149 + ], + "lines": [ + { + "bbox": [ + 310, + 73, + 425, + 149 + ], + "spans": [ + { + "bbox": [ + 310, + 73, + 425, + 149 + ], + "type": "image", + "image_path": "f80cbe153f3a2fa53c3b8868014ff6ed91dafbe483535704e0366c50d03a13db.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 335, + 156, + 515, + 167 + ], + "lines": [ + { + "bbox": [ + 335, + 156, + 515, + 167 + ], + "spans": [ + { + "bbox": [ + 335, + 156, + 515, + 167 + ], + "type": "text", + "content": "Figure 5. Castling vs. baseline ViT on ImageNet." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 430, + 74, + 542, + 150 + ], + "blocks": [ + { + "bbox": [ + 430, + 74, + 542, + 150 + ], + "lines": [ + { + "bbox": [ + 430, + 74, + 542, + 150 + ], + "spans": [ + { + "bbox": [ + 430, + 74, + 542, + 150 + ], + "type": "image", + "image_path": "f151dba8746532bcfb4368a3d9945e3c144e9e15828109e90f06fe66e7b983be.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 174, + 545, + 234 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 174, + 545, + 234 + ], + "spans": [ + { + "bbox": [ + 304, + 174, + 545, + 234 + ], + "type": "text", + "content": "for the classification task; AP, " + }, + { + "bbox": [ + 304, + 174, + 545, + 234 + ], + "type": "inline_equation", + "content": "\\mathrm{AP}^{50}" + }, + { + "bbox": [ + 304, + 174, + 545, + 234 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 174, + 545, + 234 + ], + "type": "inline_equation", + "content": "\\mathrm{AP}^{75}" + }, + { + "bbox": [ + 304, + 174, + 545, + 234 + ], + "type": "text", + "content": " for the detection task (AP: average precision); mIoU, mAcc, and pAcc for the segmentation task (mIoU: mean intersection over union). For efficiency metrics, we compare the number of model parameters or inference FLOPs (or MACs)." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 242, + 485, + 254 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 242, + 485, + 254 + ], + "spans": [ + { + "bbox": [ + 306, + 242, + 485, + 254 + ], + "type": "text", + "content": "4.2. Castling-ViT over SOTA Baslines" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 258, + 545, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 258, + 545, + 652 + ], + "spans": [ + { + "bbox": [ + 304, + 258, + 545, + 652 + ], + "type": "text", + "content": "Image Classification. To evaluate the effectiveness of our proposed techniques on the image classification task, we apply the proposed Castling-ViT idea to three typical or SOTA ViT architectures: DeiT [54] as typical ViTs, LeViT [24] as representative efficient ViTs, and MViTv2 [34] as representative hierarchical ViTs for downstream applications of high input resolutions, and compare their performance over baselines on ImageNet. As shown in Tab. 2, the comparison across a large MACs (or FLOPs) range starting from 0.4G to 17G. We categorize ViT models into four regimes: " + }, + { + "bbox": [ + 304, + 258, + 545, + 652 + ], + "type": "inline_equation", + "content": "<1\\mathrm{G}" + }, + { + "bbox": [ + 304, + 258, + 545, + 652 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 258, + 545, + 652 + ], + "type": "inline_equation", + "content": "1 \\sim 3\\mathrm{G}" + }, + { + "bbox": [ + 304, + 258, + 545, + 652 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 258, + 545, + 652 + ], + "type": "inline_equation", + "content": "3 \\sim 10\\mathrm{G}" + }, + { + "bbox": [ + 304, + 258, + 545, + 652 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 258, + 545, + 652 + ], + "type": "inline_equation", + "content": ">10\\mathrm{G}" + }, + { + "bbox": [ + 304, + 258, + 545, + 652 + ], + "type": "text", + "content": ", and select baselines in each regime to benchmark separately for the clarity purpose. Castling-ViTs consistently perform better than all baselines across various MACs ranges in terms of the accuracy-efficiency tradeoff. For example, Castling-LeViT achieves " + }, + { + "bbox": [ + 304, + 258, + 545, + 652 + ], + "type": "inline_equation", + "content": "82.6\\%" + }, + { + "bbox": [ + 304, + 258, + 545, + 652 + ], + "type": "text", + "content": " top-1 accuracy with only 1.40G MACs while LeViT requires 2.35G FLOPs instead, i.e., " + }, + { + "bbox": [ + 304, + 258, + 545, + 652 + ], + "type": "inline_equation", + "content": "\\downarrow 40\\%" + }, + { + "bbox": [ + 304, + 258, + 545, + 652 + ], + "type": "text", + "content": " MACs; On the other hand, under comparable MACs, Castling-MViTv2 achieves " + }, + { + "bbox": [ + 304, + 258, + 545, + 652 + ], + "type": "inline_equation", + "content": "84.1\\%" + }, + { + "bbox": [ + 304, + 258, + 545, + 652 + ], + "type": "text", + "content": " accuracy vs. MViTv2 with " + }, + { + "bbox": [ + 304, + 258, + 545, + 652 + ], + "type": "inline_equation", + "content": "82.3\\%" + }, + { + "bbox": [ + 304, + 258, + 545, + 652 + ], + "type": "text", + "content": " accuracy instead, i.e., " + }, + { + "bbox": [ + 304, + 258, + 545, + 652 + ], + "type": "inline_equation", + "content": "\\uparrow 1.8\\%" + }, + { + "bbox": [ + 304, + 258, + 545, + 652 + ], + "type": "text", + "content": " top-1 accuracy. Overall, Castling-ViT achieves an improved accuracy of " + }, + { + "bbox": [ + 304, + 258, + 545, + 652 + ], + "type": "inline_equation", + "content": "\\uparrow 0.5\\% \\sim \\uparrow 6.6\\%" + }, + { + "bbox": [ + 304, + 258, + 545, + 652 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 258, + 545, + 652 + ], + "type": "inline_equation", + "content": "\\uparrow 1.0\\% \\sim \\uparrow 8.1\\%" + }, + { + "bbox": [ + 304, + 258, + 545, + 652 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 258, + 545, + 652 + ], + "type": "inline_equation", + "content": "\\uparrow 1.0\\% \\sim \\uparrow 4.1\\%" + }, + { + "bbox": [ + 304, + 258, + 545, + 652 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 258, + 545, + 652 + ], + "type": "inline_equation", + "content": "\\uparrow 0.6\\% \\sim \\uparrow 2.6\\%" + }, + { + "bbox": [ + 304, + 258, + 545, + 652 + ], + "type": "text", + "content": " over baselines under " + }, + { + "bbox": [ + 304, + 258, + 545, + 652 + ], + "type": "inline_equation", + "content": "<1\\mathrm{G}" + }, + { + "bbox": [ + 304, + 258, + 545, + 652 + ], + "type": "text", + "content": " MACs, " + }, + { + "bbox": [ + 304, + 258, + 545, + 652 + ], + "type": "inline_equation", + "content": "1 \\sim 3\\mathrm{G}" + }, + { + "bbox": [ + 304, + 258, + 545, + 652 + ], + "type": "text", + "content": " MACs, " + }, + { + "bbox": [ + 304, + 258, + 545, + 652 + ], + "type": "inline_equation", + "content": "3 \\sim 10\\mathrm{G}" + }, + { + "bbox": [ + 304, + 258, + 545, + 652 + ], + "type": "text", + "content": " MACs, and " + }, + { + "bbox": [ + 304, + 258, + 545, + 652 + ], + "type": "inline_equation", + "content": ">10\\mathrm{G}" + }, + { + "bbox": [ + 304, + 258, + 545, + 652 + ], + "type": "text", + "content": " MACs, respectively. Note that we calculate improvements under comparable MACs. In addition to the overall comparison, we also visualize the apple-to/apple benchmark, e.g., Castling-LeViT vs. LeViT, to validate the effectiveness of proposed techniques. As shown in Fig. 5, Castling-LeViT/MViTv2 achieves " + }, + { + "bbox": [ + 304, + 258, + 545, + 652 + ], + "type": "inline_equation", + "content": "\\downarrow 25.7\\% \\sim \\downarrow 55.3\\%" + }, + { + "bbox": [ + 304, + 258, + 545, + 652 + ], + "type": "text", + "content": " MACs reductions under comparable accuracies or offers a comparable or better accuracy (" + }, + { + "bbox": [ + 304, + 258, + 545, + 652 + ], + "type": "inline_equation", + "content": "\\uparrow 0.6\\% \\sim \\uparrow 1.8\\%" + }, + { + "bbox": [ + 304, + 258, + 545, + 652 + ], + "type": "text", + "content": ") under comparable MACs over corresponding LeViT/MViTv2 baselines." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "type": "text", + "content": "Object Detection. We also extend the Castling-ViT to the downstream object detection task and compare it with previous efficient detectors on COCO dataset to evaluate its efficacy. Specifically, we construct the detector with modified ESNet [70] or LCNet [15] (replace the last one or two" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "14436" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 49, + 160, + 287, + 327 + ], + "blocks": [ + { + "bbox": [ + 46, + 70, + 288, + 159 + ], + "lines": [ + { + "bbox": [ + 46, + 70, + 288, + 159 + ], + "spans": [ + { + "bbox": [ + 46, + 70, + 288, + 159 + ], + "type": "text", + "content": "Table 3. Castling-ViT over SOTA baselines on COCO detection, where Castling-ViT† means that the detection model adopts modified ESNet (use transformer blocks in the last stage; referred to ESNet-ViT) as backbone while Castling-ViT‡ uses modified LCNet (use transformer blocks in the last two stage; referred to LCNet-ViT) as backbone. Also, “-S/M/L” and “-320/416” represent small/medium/large model variants and 320x320/416x416 input resolutions, respectively." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 49, + 160, + 287, + 327 + ], + "lines": [ + { + "bbox": [ + 49, + 160, + 287, + 327 + ], + "spans": [ + { + "bbox": [ + 49, + 160, + 287, + 327 + ], + "type": "table", + "html": "
ModelsParams (M)MACs (G)mAPAP50AP75
MobileDet-DSP [65]4.851.4327.3--
MobileDet-DSP-Fused [65]9.153.2229.1--
EfficientDet-D0 [52]3.902.5434.653.037.1
YOLOv5-N1.902.3028.045.7-
YOLOX-Nano [23]0.910.5425.8--
YOLOX-Tiny [23]5.063.2332.8--
FBNetV5-AC-224x320 [62]-0.7125.0--
FBNetV5-AR-224x320 [62]-0.9126.4--
FBNetV5-A-224x320 [62]-1.3527.2--
FBNetV5-AC-320x640 [62]-1.3728.9--
FBNetV5-AR-320x640 [62]-1.8030.4--
Castling-ViT-S-320†3.250.6228.142.329.2
Castling-ViT-S-416†3.251.0631.346.732.5
Castling-ViT-M-416†6.012.0034.049.535.9
Castling-ViT-L-416†9.323.0335.050.737.2
Castling-ViT-L-416*13.105.3137.353.439.6
", + "image_path": "4fdcabe63d76940ca8a0011891ca230c105fec9dd352254758919fed4f7dbc4b.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 339, + 287, + 554 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 339, + 287, + 554 + ], + "spans": [ + { + "bbox": [ + 46, + 339, + 287, + 554 + ], + "type": "text", + "content": "stages with transformer blocks) as backbones and follow PicoDet [70]'s detection head design as well as their training recipe. Fig. 1 and Tab. 3 show the overall comparison between the proposed Castling-ViT and other baselines. We can see that our Castling-ViT consistently achieves better accuracy-efficiency tradeoffs, leading to " + }, + { + "bbox": [ + 46, + 339, + 287, + 554 + ], + "type": "inline_equation", + "content": "\\uparrow 6.0" + }, + { + "bbox": [ + 46, + 339, + 287, + 554 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 339, + 287, + 554 + ], + "type": "inline_equation", + "content": "\\uparrow 2.2 \\sim \\uparrow 2.3" + }, + { + "bbox": [ + 46, + 339, + 287, + 554 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 339, + 287, + 554 + ], + "type": "inline_equation", + "content": "\\uparrow 4.0 \\sim \\uparrow 5.9" + }, + { + "bbox": [ + 46, + 339, + 287, + 554 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 339, + 287, + 554 + ], + "type": "inline_equation", + "content": "\\uparrow 3.1 \\sim \\uparrow 4.0" + }, + { + "bbox": [ + 46, + 339, + 287, + 554 + ], + "type": "text", + "content": " mAP improvements as compared to YOLOv5, YOLOX [23], MobileDet [65], and FBNetv5 [62], respectively, under comparable or even less MACs. As for EfficientDet [52], our method achieves comparable accuracy-efficiency trade-offs. Apart from the overall comparison with baselines, we also provide the apple-to-Apple comparison with detectors with softmax-based attention, we supply the comparison results and Castling-ViT's breakdown analysis to Sec. 4.4. This set of experiments validate the effectiveness of the proposed Castling-ViT for servering as efficient detector backbones (e.g., " + }, + { + "bbox": [ + 46, + 339, + 287, + 554 + ], + "type": "inline_equation", + "content": "\\leq 3G" + }, + { + "bbox": [ + 46, + 339, + 287, + 554 + ], + "type": "text", + "content": " MACs) in the object detection task." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 558, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 558, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 558, + 287, + 715 + ], + "type": "text", + "content": "Semantic Segmentation. We further extend Castling-ViT to the semantic segmentation task to evaluate its effectiveness. Specifically, we use ViT-Base as the backbone in Mask2former [11] framework to serve as our baseline and testbed. Then we build the Castling-ViT-Base as the backbone to benchmark on the ADE20K dataset. As shown in Tab. 4, Mask2former with Castling-ViT backbone achieves " + }, + { + "bbox": [ + 46, + 558, + 287, + 715 + ], + "type": "inline_equation", + "content": "15\\%" + }, + { + "bbox": [ + 46, + 558, + 287, + 715 + ], + "type": "text", + "content": " total MACs reductions and " + }, + { + "bbox": [ + 46, + 558, + 287, + 715 + ], + "type": "inline_equation", + "content": "19\\%" + }, + { + "bbox": [ + 46, + 558, + 287, + 715 + ], + "type": "text", + "content": " backbone MACs reduction under comparable or slightly better mIoU, i.e., " + }, + { + "bbox": [ + 46, + 558, + 287, + 715 + ], + "type": "inline_equation", + "content": "\\uparrow 0.13\\%" + }, + { + "bbox": [ + 46, + 558, + 287, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 558, + 287, + 715 + ], + "type": "inline_equation", + "content": "\\uparrow 0.52\\%" + }, + { + "bbox": [ + 46, + 558, + 287, + 715 + ], + "type": "text", + "content": " without or with MAE pretraining on ImageNet [25]. This set of experiments validate that our proposed techniques could be well generalized to various downstream tasks that require large input resolutions." + } + ] + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 308, + 105, + 545, + 159 + ], + "blocks": [ + { + "bbox": [ + 305, + 70, + 545, + 104 + ], + "lines": [ + { + "bbox": [ + 305, + 70, + 545, + 104 + ], + "spans": [ + { + "bbox": [ + 305, + 70, + 545, + 104 + ], + "type": "text", + "content": "Table 4. Castling-ViT over baselines on the ADE20K segmentation dataset, serving as backbones in the Mask2former framework. Note that we report both total MACs and backbone MACs, i.e., " + }, + { + "bbox": [ + 305, + 70, + 545, + 104 + ], + "type": "inline_equation", + "content": "(\\cdot)" + }, + { + "bbox": [ + 305, + 70, + 545, + 104 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 308, + 105, + 545, + 159 + ], + "lines": [ + { + "bbox": [ + 308, + 105, + 545, + 159 + ], + "spans": [ + { + "bbox": [ + 308, + 105, + 545, + 159 + ], + "type": "table", + "html": "
Mask2former w/MAE \nPretrainParams \n(M)MACs \n(G)mIoUmAccpAcc
ViT-BaseN118229 (182)34.5446.3675.84
Castling-ViT-BaseN118195 (147)34.6746.4776.20
ViT-BaseY118229 (182)47.9261.0083.02
Castling-ViT-BaseY118195 (147)48.4461.8283.29
", + "image_path": "1124711e09ba1fa5f4dc6e83156c8fad29a685452d8968e146223caf8429f68c.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 309, + 233, + 545, + 323 + ], + "blocks": [ + { + "bbox": [ + 305, + 166, + 545, + 232 + ], + "lines": [ + { + "bbox": [ + 305, + 166, + 545, + 232 + ], + "spans": [ + { + "bbox": [ + 305, + 166, + 545, + 232 + ], + "type": "text", + "content": "Table 5. Ablation studies of Castling-ViT with various kernel-based linear attentions, where MC refers to mean-centering kernel function applied to " + }, + { + "bbox": [ + 305, + 166, + 545, + 232 + ], + "type": "inline_equation", + "content": "\\mathbf{Q} / \\mathbf{K}" + }, + { + "bbox": [ + 305, + 166, + 545, + 232 + ], + "type": "text", + "content": "; Softmax also means the kernel function; ReLU-S and ReLU-E denote applying ReLU as kernels for decomposing softmax or exponential terms, respectively. All models are trained on COCO from scratch without pretraining." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 309, + 233, + 545, + 323 + ], + "lines": [ + { + "bbox": [ + 309, + 233, + 545, + 323 + ], + "spans": [ + { + "bbox": [ + 309, + 233, + 545, + 323 + ], + "type": "table", + "html": "
BackboneLinear AttnS-320S-416M-416
mAPAP50mAPAP50mAPAP50
LCNetN/A25.539.029.744.234.349.9
LCNet-ViTN/A25.939.629.844.434.950.4
LCNet-ViTMC [17]22.834.926.339.631.546.1
Softmax [49]26.039.629.243.434.449.8
Cosine [5]24.437.427.241.032.948.0
ReLU-S23.936.7--30.945.3
ReLU-E [6]26.540.230.244.734.349.7
Lin.-Angular27.040.630.445.535.551.1
", + "image_path": "50e242aafcb40b2bf76c639f90f87a71c0874514f720cd2f46615d5f6ed89a10.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 336, + 545, + 350 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 336, + 545, + 350 + ], + "spans": [ + { + "bbox": [ + 305, + 336, + 545, + 350 + ], + "type": "text", + "content": "4.3. Linear-Angular Attention over SOTA Baselines" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 354, + 545, + 497 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 354, + 545, + 497 + ], + "spans": [ + { + "bbox": [ + 304, + 354, + 545, + 497 + ], + "type": "text", + "content": "We also conduct ablation studies among various kinds of kernels used in linear attention to evaluate the superiority of our proposed linear-angular kernel. Also shown in Tab. 5, we compare the angular kernel with five other commonly adopted kernels, results on three model and resolution settings consistently demonstrate that our proposed angular kernel helps achieve better mAP, e.g., " + }, + { + "bbox": [ + 304, + 354, + 545, + 497 + ], + "type": "inline_equation", + "content": "\\uparrow 4.0\\% \\sim \\uparrow 4.2\\%" + }, + { + "bbox": [ + 304, + 354, + 545, + 497 + ], + "type": "text", + "content": " over MC [17], " + }, + { + "bbox": [ + 304, + 354, + 545, + 497 + ], + "type": "inline_equation", + "content": "\\uparrow 1.0\\% \\sim \\uparrow 1.2\\%" + }, + { + "bbox": [ + 304, + 354, + 545, + 497 + ], + "type": "text", + "content": " over Softmax [49], " + }, + { + "bbox": [ + 304, + 354, + 545, + 497 + ], + "type": "inline_equation", + "content": "\\uparrow 2.6\\% \\sim \\uparrow 3.2\\%" + }, + { + "bbox": [ + 304, + 354, + 545, + 497 + ], + "type": "text", + "content": " over Cosine [5], " + }, + { + "bbox": [ + 304, + 354, + 545, + 497 + ], + "type": "inline_equation", + "content": "\\uparrow 3.1\\% \\sim \\uparrow 4.6\\%" + }, + { + "bbox": [ + 304, + 354, + 545, + 497 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 354, + 545, + 497 + ], + "type": "inline_equation", + "content": "\\uparrow 0.2\\% \\sim \\uparrow 1.2\\%" + }, + { + "bbox": [ + 304, + 354, + 545, + 497 + ], + "type": "text", + "content": " over ReLU-S and ReLU-E [6], respectively. This set of experiments validate the superiority of the proposed linear-angular kernels." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 504, + 481, + 517 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 504, + 481, + 517 + ], + "spans": [ + { + "bbox": [ + 305, + 504, + 481, + 517 + ], + "type": "text", + "content": "4.4. Ablation Studies of Castling-ViT" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 521, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 521, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 521, + 545, + 713 + ], + "type": "text", + "content": "We conduct ablation studies on Castling-ViT's linear- angular attention, added DWConv, as well as auxiliary masked softmax-based attention, as shown in Tab. 6 and Tab. 7, where the experiments are performed on CastlingViT-S-320\\*/M-416\\*, respectively, without pretraining on ImageNet. Note that here \\* means that we use LCNet-ViT, i.e., replacing all layers in last two stages with transformer blocks, as backbones. Results in three tables show detailed performance breakdown and consistently demonstrate that all components in our proposed Castling-ViT contribute to the final performance. Specifically, linear-angular attention itself already achieves " + }, + { + "bbox": [ + 304, + 521, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\downarrow 0.1\\% \\sim \\uparrow 0.5\\%" + }, + { + "bbox": [ + 304, + 521, + 545, + 713 + ], + "type": "text", + "content": " mAP improvements while leading to " + }, + { + "bbox": [ + 304, + 521, + 545, + 713 + ], + "type": "inline_equation", + "content": "10.1\\% \\sim 15.3\\%" + }, + { + "bbox": [ + 304, + 521, + 545, + 713 + ], + "type": "text", + "content": " MACs reductions simultaneously. Adding DWConv in the middle of MLP layers further leads to " + }, + { + "bbox": [ + 304, + 521, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\downarrow 0.1\\% \\sim \\uparrow 0.6\\%" + }, + { + "bbox": [ + 304, + 521, + 545, + 713 + ], + "type": "text", + "content": " mAP improvements while only incur negligible MACs overhead. Adding aux" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14437" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 49, + 116, + 287, + 186 + ], + "blocks": [ + { + "bbox": [ + 47, + 70, + 288, + 115 + ], + "lines": [ + { + "bbox": [ + 47, + 70, + 288, + 115 + ], + "spans": [ + { + "bbox": [ + 47, + 70, + 288, + 115 + ], + "type": "text", + "content": "Table 6. Ablation studies of Castling-ViT-S-320 w/ LCNet backbone, or LCNet-ViT backbone + linear-angular attention (denoted as Lin.) + DWConv (denoted as DW.) + auxiliary softmax-based attention (denoted as SparseAttn.)." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 49, + 116, + 287, + 186 + ], + "lines": [ + { + "bbox": [ + 49, + 116, + 287, + 186 + ], + "spans": [ + { + "bbox": [ + 49, + 116, + 287, + 186 + ], + "type": "table", + "html": "
Castling-ViT-S-320 w/Params (M)MACs (G)mAPAP50AP75
LCNet1.100.4825.539.026.5
LCNet-ViT2.140.6925.939.626.9
+ Lin.2.140.6226.239.727.3
+ Lin.+DW.2.140.6226.840.728.0
+ Lin.+DW.+SparseAttn.2.140.6227.040.628.0
", + "image_path": "0e8692f7a7ea0d9f713cceddc428971960983d06b90946db75fc03db6dcc821c.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 49, + 213, + 287, + 283 + ], + "blocks": [ + { + "bbox": [ + 47, + 190, + 287, + 212 + ], + "lines": [ + { + "bbox": [ + 47, + 190, + 287, + 212 + ], + "spans": [ + { + "bbox": [ + 47, + 190, + 287, + 212 + ], + "type": "text", + "content": "Table 7. Ablation studies of Castling-ViT-M-416 w/ LCNet backbone, or LCNet-ViT backbone + Lin. + DW. + SparseAttn." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 49, + 213, + 287, + 283 + ], + "lines": [ + { + "bbox": [ + 49, + 213, + 287, + 283 + ], + "spans": [ + { + "bbox": [ + 49, + 213, + 287, + 283 + ], + "type": "table", + "html": "
Castling-ViT-M-416 w/Params (M)MACs (G)mAPAP50AP75
LCNet3.322.1534.349.936.5
LCNet-ViT7.533.5134.950.437.0
+ Lin.7.533.1534.850.336.7
+ Lin.+DW.7.533.1535.150.936.9
+ Lin.+DW.+SparseAttn.7.533.1535.351.137.3
", + "image_path": "c591186ebac976780e92fa48d5640ec668c8028435f16f989f746ad267fd87cc.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 295, + 287, + 438 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 295, + 287, + 438 + ], + "spans": [ + { + "bbox": [ + 46, + 295, + 287, + 438 + ], + "type": "text", + "content": "illary masked softmax-based attention further increases the mAP by " + }, + { + "bbox": [ + 46, + 295, + 287, + 438 + ], + "type": "inline_equation", + "content": "0.2\\%" + }, + { + "bbox": [ + 46, + 295, + 287, + 438 + ], + "type": "text", + "content": " without incurring any overhead since it only assists the training while being removed during inference. Note that although LCNet-ViT backbone leads to better mAP, it also introduces more MACs as compared to efficient convolutions. As such, LCNet-ViT still slightly underperforms LCNet backbones in terms of accuracy-efficiency trade-offs, since we directly replace the CNN backbones with transformer blocks, whose architecture could not be optimal for ViTs. It remains to an open problem to build ViT based models that achieves higher accuracy-efficiency trade-off than pure ConvNet models." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 445, + 235, + 458 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 445, + 235, + 458 + ], + "spans": [ + { + "bbox": [ + 47, + 445, + 235, + 458 + ], + "type": "text", + "content": "4.5. Discussion on the Auxiliary Branch" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 463, + 156, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 463, + 156, + 616 + ], + "spans": [ + { + "bbox": [ + 46, + 463, + 156, + 616 + ], + "type": "text", + "content": "Trajectory of Nonzeros in Masks. To further understand the effect of the auxiliary branch, we visualize the trajectories of nonzeros in masks of auxiliary softmax-based attention. As shown in Fig. 6, we count the nonzeros in the first attention layer's masks throughout the training of both Castling-ViT-S-416" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 617, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 617, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 617, + 287, + 714 + ], + "type": "text", + "content": "and Castling-ViT-M-416 with the LCNet-ViT backbone on COCO. we observe that the introduced auxiliary attention will only assist the training in the early or middle training stages and will gradually vanish towards all zeros in the later training stage, which is well aligned with the assumption in our Castling-ViT. This set of experiments validate the idea of performing \"castling\" in ViTs, i.e., drop the auxiliary branch without sacrificing performance." + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 181, + 466, + 281, + 574 + ], + "blocks": [ + { + "bbox": [ + 181, + 466, + 281, + 574 + ], + "lines": [ + { + "bbox": [ + 181, + 466, + 281, + 574 + ], + "spans": [ + { + "bbox": [ + 181, + 466, + 281, + 574 + ], + "type": "image", + "image_path": "5279c2953fa8e8015eb33abc95d73b9575265d7ccf39c40c1efc3d7cca79218c.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 173, + 576, + 287, + 610 + ], + "lines": [ + { + "bbox": [ + 173, + 576, + 287, + 610 + ], + "spans": [ + { + "bbox": [ + 173, + 576, + 287, + 610 + ], + "type": "text", + "content": "Figure 6. Visualizing the trajectories of nonzeros in auxiliary masks during training." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 309, + 71, + 385, + 133 + ], + "blocks": [ + { + "bbox": [ + 309, + 71, + 385, + 133 + ], + "lines": [ + { + "bbox": [ + 309, + 71, + 385, + 133 + ], + "spans": [ + { + "bbox": [ + 309, + 71, + 385, + 133 + ], + "type": "image", + "image_path": "d75300cba95932f023a0bd0c5f4439fede7dee84c03a47788d5601f602833679.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 137, + 545, + 181 + ], + "lines": [ + { + "bbox": [ + 304, + 137, + 545, + 181 + ], + "spans": [ + { + "bbox": [ + 304, + 137, + 545, + 181 + ], + "type": "text", + "content": "Figure 7. Left: Visualizing the taget similarity function, linear-angular terms, and DNNs' fitting curve; Middle: Visualizing the loss trajectory during DNN training; Right: Visualizing the learning trajectory of both low and high frequency components." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 385, + 72, + 460, + 134 + ], + "blocks": [ + { + "bbox": [ + 385, + 72, + 460, + 134 + ], + "lines": [ + { + "bbox": [ + 385, + 72, + 460, + 134 + ], + "spans": [ + { + "bbox": [ + 385, + 72, + 460, + 134 + ], + "type": "image", + "image_path": "80a8e7ce0330b01dacacd833646e1813cd7573a0f392c088e9016f4ee2192c35.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 464, + 70, + 545, + 134 + ], + "blocks": [ + { + "bbox": [ + 464, + 70, + 545, + 134 + ], + "lines": [ + { + "bbox": [ + 464, + 70, + 545, + 134 + ], + "spans": [ + { + "bbox": [ + 464, + 70, + 545, + 134 + ], + "type": "image", + "image_path": "06ba51f9e7a7ddd23a070f84e831f634c43e5996188046b58fa7e971ca0d3996.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 186, + 545, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 186, + 545, + 449 + ], + "spans": [ + { + "bbox": [ + 304, + 186, + 545, + 449 + ], + "type": "text", + "content": "Conjecture of Castling. We conduct a synthetic experiment to give an analogy for explaining the castling phenomenon. In Fig. 7 (Left), we visualize the curve of (1) angular similarity function (denoted in Red), (2) linear-angular terms (denoted in Green), and (3) two-layer DNN's approximation. We see that only keeping linear-angular terms leads to distortion, while DNN is capable of learning the missing high-frequency parts, whose loss trajectory is shown in Fig. 7 (Middle). Also, the learning trajectory of frequency components is visualized in Fig. 7 (Right), following F-Principle [67], where x-axis is training steps, y-axis categorizes both low and high frequency components, blue/red colors refer to large/small difference between the learned frequency components and target frequency components, respectively. We observe that DNNs fit target functions from low to high frequencies during training. It indicates that DWConv itself is not sufficient for approximating the high-order residual at early training stages. Costly attention that contains high-order residuals is then desired to help training while being dropped at inference since the remaining networks can gradually learn the high-frequency components at later training stages [67]." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 456, + 378, + 468 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 456, + 378, + 468 + ], + "spans": [ + { + "bbox": [ + 306, + 456, + 378, + 468 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 473, + 545, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 473, + 545, + 628 + ], + "spans": [ + { + "bbox": [ + 304, + 473, + 545, + 628 + ], + "type": "text", + "content": "We present Castling-ViT that trains ViTs with both linear-angular and softmax-based quadratic attention but switches to only having the former during inference. Castling-ViT leverages angular kernels to measure the similarities between the queries and keys via spectral angles and highlights two enablers: (1) a new linear-angular attention mechanism: we decompose angular kernels to linear terms and high-order residuals, and keep only the former for inference; and (2) we approximate the high-order residuals using a depthwise convolution and an auxiliary masked softmax attention whose masks gradually become zeros during training without incurring inference overhead. Extensive experiments consistently validate Castling-ViT's advantages." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 635, + 399, + 649 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 635, + 399, + 649 + ], + "spans": [ + { + "bbox": [ + 306, + 635, + 399, + 649 + ], + "type": "text", + "content": "Acknowledgment" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 654, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 545, + 714 + ], + "type": "text", + "content": "The work is supported in part by the National Science Foundation (NSF) RTML program (Award number: 1937592) and the CoCoSys, one of the seven centers in JUMP 2.0, a Semiconductor Research Corporation (SRC) program sponsored by DARPA." + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14438" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 156 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 156 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 156 + ], + "type": "text", + "content": "[1] Alaaeldin Ali, Hugo Touvron, Mathilde Caron, Piotr Bojanowski, Matthijs Douze, Armand Joulin, Ivan Laptev, Natalia Neverova, Gabriel Synnaeve, Jakob Verbeek, et al. Xcit: Cross-covariance image transformers. Advances in neural information processing systems, 34:20014-20027, 2021. 2, 13" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 159, + 288, + 203 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 159, + 288, + 203 + ], + "spans": [ + { + "bbox": [ + 53, + 159, + 288, + 203 + ], + "type": "text", + "content": "[2] Moab Arar, Ariel Shamir, and Amit H Bermano. Learned queries for efficient local attention. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10841-10852, 2022. 1, 2, 13" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 205, + 288, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 205, + 288, + 237 + ], + "spans": [ + { + "bbox": [ + 53, + 205, + 288, + 237 + ], + "type": "text", + "content": "[3] Iz Beltagy, Matthew E Peters, and Arman Cohan. Longformer: The long-document transformer. arXiv preprint arXiv:2004.05150, 2020. 1, 13" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 239, + 288, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 239, + 288, + 282 + ], + "spans": [ + { + "bbox": [ + 53, + 239, + 288, + 282 + ], + "type": "text", + "content": "[4] Daniel Bolya, Cheng-Yang Fu, Xiaoliang Dai, Peizhao Zhang, Christoph Feichtenhofer, and Judy Hoffman. Token merging: Your vit but faster. arXiv preprint arXiv:2210.09461, 2022. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 285, + 287, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 285, + 287, + 327 + ], + "spans": [ + { + "bbox": [ + 53, + 285, + 287, + 327 + ], + "type": "text", + "content": "[5] Daniel Bolya, Cheng-Yang Fu, Xiaoliang Dai, Peizhao Zhang, and Judy Hoffman. Hydra attention: Efficient attention with many heads. arXiv preprint arXiv:2209.07484, 2022. 1, 2, 5, 7, 13" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 330, + 287, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 330, + 287, + 373 + ], + "spans": [ + { + "bbox": [ + 53, + 330, + 287, + 373 + ], + "type": "text", + "content": "[6] Han Cai, Chuang Gan, and Song Han. Efficientvit: Enhanced linear attention for high-resolution low-computation visual recognition. arXiv preprint arXiv:2205.14756, 2022. 1, 2, 3, 5, 7, 13, 14" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 375, + 287, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 375, + 287, + 430 + ], + "spans": [ + { + "bbox": [ + 53, + 375, + 287, + 430 + ], + "type": "text", + "content": "[7] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In European conference on computer vision, pages 213-229. Springer, 2020. 13" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 432, + 287, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 432, + 287, + 475 + ], + "spans": [ + { + "bbox": [ + 53, + 432, + 287, + 475 + ], + "type": "text", + "content": "[8] Beidi Chen, Tri Dao, Eric Winsor, Zhao Song, Atri Rudra, and Christopher Ré. Scatterbrain: Unifying sparse and low-rank attention. Advances in Neural Information Processing Systems, 34:17413-17426, 2021. 1, 2, 5, 13" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 477, + 287, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 477, + 287, + 510 + ], + "spans": [ + { + "bbox": [ + 53, + 477, + 287, + 510 + ], + "type": "text", + "content": "[9] Chun-Fu Chen, Quanfu Fan, and Rameswar Panda. Crossvit: Cross-attention multi-scale vision transformer for image classification. arXiv preprint arXiv:2103.14899, 2021. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 512, + 287, + 565 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 512, + 287, + 565 + ], + "spans": [ + { + "bbox": [ + 48, + 512, + 287, + 565 + ], + "type": "text", + "content": "[10] Minghao Chen, Houwen Peng, Jianlong Fu, and Haibin Ling. Autoformer: Searching transformers for visual recognition. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12270-12280, 2021. 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 567, + 287, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 567, + 287, + 632 + ], + "spans": [ + { + "bbox": [ + 48, + 567, + 287, + 632 + ], + "type": "text", + "content": "[11] Bowen Cheng, Ishan Misra, Alexander G Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1290–1299, 2022. 3, 5, 6, 7, 13" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 635, + 287, + 677 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 635, + 287, + 677 + ], + "spans": [ + { + "bbox": [ + 48, + 635, + 287, + 677 + ], + "type": "text", + "content": "[12] Bowen Cheng, Alex Schwing, and Alexander Kirillov. Per-pixel classification is not all you need for semantic segmentation. Advances in Neural Information Processing Systems, 34:17864-17875, 2021. 13" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "type": "text", + "content": "[13] Krzysztof Choromanski, Valerii Likhosherstov, David Dohan, Xingyou Song, Andreea Gane, Tamas Sarlos, Peter Hawkins, Jared Davis, Afroz Mohiuddin, Lukasz Kaiser," + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "text", + "content": "et al. Rethinking attention with performers. arXiv preprint arXiv:2009.14794, 2020.3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 96, + 545, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 96, + 545, + 171 + ], + "spans": [ + { + "bbox": [ + 308, + 96, + 545, + 171 + ], + "type": "text", + "content": "[14] Krzysztof Marcin Choromanski, Valerii Likhosherstov, David Dohan, Xingyou Song, Andreea Gane, Tamas Sarlos, Peter Hawkins, Jared Quincy Davis, Afroz Mohiuddin, Lukasz Kaiser, David Benjamin Belanger, Lucy J Colwell, and Adrian Weller. Rethinking attention with performers. In International Conference on Learning Representations, 2021. 1, 2, 13" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 173, + 545, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 173, + 545, + 226 + ], + "spans": [ + { + "bbox": [ + 308, + 173, + 545, + 226 + ], + "type": "text", + "content": "[15] Cheng Cui, Tingquan Gao, Shengyu Wei, Yuning Du, Ruoyu Guo, Shuilong Dong, Bin Lu, Ying Zhou, Xueying Lv, Qiwen Liu, et al. Pp-lcnet: A lightweight cpu convolutional neural network. arXiv preprint arXiv:2109.15099, 2021. 5, 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 228, + 545, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 228, + 545, + 281 + ], + "spans": [ + { + "bbox": [ + 308, + 228, + 545, + 281 + ], + "type": "text", + "content": "[16] Xiaoliang Dai, Alvin Wan, P. Zhang, B. Wu, Zijian He, Zhen Wei, K. Chen, Yuandong Tian, Matthew E. Yu, Péter Vajda, and J. Gonzalez. Fbnetv3: Joint architecture-recipe search using neural acquisition function. ArXiv, abs/2006.02049, 2020.2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 283, + 545, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 283, + 545, + 348 + ], + "spans": [ + { + "bbox": [ + 308, + 283, + 545, + 348 + ], + "type": "text", + "content": "[17] Jyotikrishna Dass, Shang Wu, Huihong Shi, Chaojian Li, Zhifan Ye, Zhongfeng Wang, and Yingyan Lin. Vitality: Unifying low-rank and sparse approximation for vision transformer acceleration with a linear taylor attention. In The 29th IEEE International Symposium on High-Performance Computer Architecture (HPCA 2023), 2023. 2, 5, 7, 14" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 350, + 545, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 350, + 545, + 392 + ], + "spans": [ + { + "bbox": [ + 308, + 350, + 545, + 392 + ], + "type": "text", + "content": "[18] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 1, 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 394, + 545, + 458 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 394, + 545, + 458 + ], + "spans": [ + { + "bbox": [ + 308, + 394, + 545, + 458 + ], + "type": "text", + "content": "[19] Xiaoyi Dong, Jianmin Bao, Dongdong Chen, Weiming Zhang, Nenghai Yu, Lu Yuan, Dong Chen, and Baining Guo. Cswin transformer: A general vision transformer backbone with cross-shaped windows. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12124-12134, 2022. 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 460, + 545, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 460, + 545, + 536 + ], + "spans": [ + { + "bbox": [ + 308, + 460, + 545, + 536 + ], + "type": "text", + "content": "[20] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations, 2021. 1, 2, 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 537, + 545, + 591 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 537, + 545, + 591 + ], + "spans": [ + { + "bbox": [ + 308, + 537, + 545, + 591 + ], + "type": "text", + "content": "[21] Haoqi Fan, Bo Xiong, Karttikeya Mangalam, Yanghao Li, Zhicheng Yan, Jitendra Malik, and Christoph Feichtenhofer. Multiscale vision transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6824-6835, 2021. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 593, + 545, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 593, + 545, + 624 + ], + "spans": [ + { + "bbox": [ + 308, + 593, + 545, + 624 + ], + "type": "text", + "content": "[22] Peng Gao, Teli Ma, Hongsheng Li, Jifeng Dai, and Yu Qiao. Convmae: Masked convolution meets masked autoencoders. arXiv preprint arXiv:2205.03892, 2022. 13" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 626, + 545, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 626, + 545, + 657 + ], + "spans": [ + { + "bbox": [ + 308, + 626, + 545, + 657 + ], + "type": "text", + "content": "[23] Zheng Ge, Songtao Liu, Feng Wang, Zeming Li, and Jian Sun. Yolox: Exceeding yolo series in 2021. arXiv preprint arXiv:2107.08430, 2021. 6, 7, 14" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 658, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 658, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 308, + 658, + 545, + 712 + ], + "type": "text", + "content": "[24] Ben Graham, Alaaeldin El-Nouby, Hugo Touvron, Pierre Stock, Armand Joulin, Hervé Jégou, and Matthijs Douze. Levit: a vision transformer in convnet's clothing for faster inference. arXiv preprint arXiv:2104.01136, 2021. 2, 3, 5, 6, 13, 14" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "text", + "content": "14439" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 714 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "text", + "content": "[25] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollár, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16000-16009, 2022. 6, 7, 13" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 128, + 288, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 128, + 288, + 172 + ], + "spans": [ + { + "bbox": [ + 48, + 128, + 288, + 172 + ], + "type": "text", + "content": "[26] Byeongho Heo, Sangdoo Yun, Dongyoon Han, Sanghyuk Chun, Junsuk Choe, and Seong Joon Oh. Rethinking spatial dimensions of vision transformers. arXiv preprint arXiv:2103.16302, 2021. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 173, + 288, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 173, + 288, + 226 + ], + "spans": [ + { + "bbox": [ + 48, + 173, + 288, + 226 + ], + "type": "text", + "content": "[27] Paul Honeine and Cedric Richard. The angular kernel in machine learning for hyperspectral data classification. In 2010 2nd Workshop on Hyperspectral Image and Signal Processing: Evolution in Remote Sensing, pages 1-4. IEEE, 2010. 4" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 228, + 288, + 292 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 228, + 288, + 292 + ], + "spans": [ + { + "bbox": [ + 48, + 228, + 288, + 292 + ], + "type": "text", + "content": "[28] Andrew Howard, Mark Sandler, Grace Chu, Liang-Chieh Chen, Bo Chen, Mingxing Tan, Weijun Wang, Yukun Zhu, Ruoming Pang, Vijay Vasudevan, et al. Searching for mobilenetv3. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1314-1324, 2019. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 293, + 288, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 293, + 288, + 348 + ], + "spans": [ + { + "bbox": [ + 48, + 293, + 288, + 348 + ], + "type": "text", + "content": "[29] Angelos Katharopoulos, Apoorv Vyas, Nikolaos Pappas, and François Fleuret. Transformers are rnns: Fast autoregressive transformers with linear attention. In International Conference on Machine Learning, pages 5156-5165. PMLR, 2020. 1, 2, 3, 4, 13" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 349, + 288, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 349, + 288, + 392 + ], + "spans": [ + { + "bbox": [ + 48, + 349, + 288, + 392 + ], + "type": "text", + "content": "[30] Nirmal Keshava. Distance metrics and band selection in hyperspectral processing with applications to material identification and spectral libraries. IEEE Transactions on Geoscience and remote sensing, 42(7):1552-1565, 2004. 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 393, + 288, + 448 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 393, + 288, + 448 + ], + "spans": [ + { + "bbox": [ + 48, + 393, + 288, + 448 + ], + "type": "text", + "content": "[31] Kyungmin Kim, Bichen Wu, Xiaoliang Dai, Peizhao Zhang, Zhicheng Yan, Peter Vajda, and Seon Joo Kim. Rethinking the self-attention in vision transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3071–3075, 2021. 4, 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 449, + 288, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 449, + 288, + 480 + ], + "spans": [ + { + "bbox": [ + 48, + 449, + 288, + 480 + ], + "type": "text", + "content": "[32] Nikita Kitaev, Lukasz Kaiser, and Anselm Levskaya. Reformer: The efficient transformer. arXiv preprint arXiv:2001.04451, 2020. 1, 13" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 481, + 288, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 481, + 288, + 514 + ], + "spans": [ + { + "bbox": [ + 48, + 481, + 288, + 514 + ], + "type": "text", + "content": "[33] Yanghao Li, Hanzi Mao, Ross Girshick, and Kaiming He. Exploring plain vision transformer backbones for object detection. arXiv preprint arXiv:2203.16527, 2022. 13" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 515, + 288, + 581 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 515, + 288, + 581 + ], + "spans": [ + { + "bbox": [ + 48, + 515, + 288, + 581 + ], + "type": "text", + "content": "[34] Yanghao Li, Chao-Yuan Wu, Haoqi Fan, Karttikeya Mangalam, Bo Xiong, Jitendra Malik, and Christoph Feichtenhofer. Mviv2: Improved multiscale vision transformers for classification and detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4804-4814, 2022. 3, 5, 6, 13, 14" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 582, + 288, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 582, + 288, + 624 + ], + "spans": [ + { + "bbox": [ + 48, + 582, + 288, + 624 + ], + "type": "text", + "content": "[35] Yanyu Li, Geng Yuan, Yang Wen, Eric Hu, Georgios Evangelidis, Sergey Tulyakov, Yanzhi Wang, and Jian Ren. Efficientformer: Vision transformers at mobilenet speed. arXiv preprint arXiv:2206.01191, 2022. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 625, + 288, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 625, + 288, + 679 + ], + "spans": [ + { + "bbox": [ + 48, + 625, + 288, + 679 + ], + "type": "text", + "content": "[36] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In European conference on computer vision, pages 740-755. Springer, 2014. 1, 5" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 680, + 288, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 680, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 48, + 680, + 288, + 714 + ], + "type": "text", + "content": "[37] Jing Liu, Zizheng Pan, Haoyu He, Jianfei Cai, and Bohan Zhuang. Ecoformer: Energy-saving attention with linear complexity. In NeurIPS, 2022. 1, 2, 4, 13" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 128 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 128 + ], + "type": "text", + "content": "[38] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10012-10022, 2021. 1, 2, 6, 13, 14" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 129, + 545, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 129, + 545, + 183 + ], + "spans": [ + { + "bbox": [ + 307, + 129, + 545, + 183 + ], + "type": "text", + "content": "[39] Jiachen Lu, Jinghan Yao, Junge Zhang, Xiatian Zhu, Hang Xu, Weiguo Gao, Chunjing Xu, Tao Xiang, and Li Zhang. Soft: softmax-free transformer with linear complexity. Advances in Neural Information Processing Systems, 34:21297-21309, 2021. 1, 2, 13" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 185, + 545, + 218 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 185, + 545, + 218 + ], + "spans": [ + { + "bbox": [ + 307, + 185, + 545, + 218 + ], + "type": "text", + "content": "[40] Sachin Mehta and Mohammad Rastegari. Mobilevit: lightweight, general-purpose, and mobile-friendly vision transformer. arXiv preprint arXiv:2110.02178, 2021. 2, 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 219, + 545, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 219, + 545, + 251 + ], + "spans": [ + { + "bbox": [ + 307, + 219, + 545, + 251 + ], + "type": "text", + "content": "[41] John Moody and Christian Darken. _Learning with localized receptive fields_. Yale Univ., Department of Computer Science, 1988. 4" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 253, + 545, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 253, + 545, + 297 + ], + "spans": [ + { + "bbox": [ + 307, + 253, + 545, + 297 + ], + "type": "text", + "content": "[42]Zipeng Qin, Jianbo Liu, Xiaolin Zhang, Maoqing Tian, Aojun Zhou, Shuai Yi, and Hongsheng Li. Pyramid fusion transformer for semantic segmentation. arXiv preprint arXiv:2201.04019, 2022.3, 13" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 298, + 545, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 298, + 545, + 352 + ], + "spans": [ + { + "bbox": [ + 307, + 298, + 545, + 352 + ], + "type": "text", + "content": "[43] Zhen Qin, Weixuan Sun, Hui Deng, Dongxu Li, Yunshen Wei, Baohong Lv, Junjie Yan, Lingpeng Kong, and Yiran Zhong. cosformer: Rethinking softmax in attention. In International Conference on Learning Representations, 2022. 1, 2, 13" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 354, + 545, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 354, + 545, + 408 + ], + "spans": [ + { + "bbox": [ + 307, + 354, + 545, + 408 + ], + "type": "text", + "content": "[44] Ilija Radosavovic, Raj Prateek Kosaraju, Ross Girshick, Kaiming He, and Piotr Dólár. Designing network design spaces. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10428-10436, 2020. 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 410, + 545, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 410, + 545, + 443 + ], + "spans": [ + { + "bbox": [ + 307, + 410, + 545, + 443 + ], + "type": "text", + "content": "[45] Ali Rahimi and Benjamin Recht. Random features for large-scale kernel machines. Advances in neural information processing systems, 20, 2007. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 444, + 545, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 444, + 545, + 498 + ], + "spans": [ + { + "bbox": [ + 307, + 444, + 545, + 498 + ], + "type": "text", + "content": "[46] Yongming Rao, Wenliang Zhao, Benlin Liu, Jiwen Lu, Jie Zhou, and Cho-Jui Hsieh. Dynamicvit: Efficient vision transformers with dynamic token sparsification. Advances in neural information processing systems, 34:13937-13949, 2021. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 500, + 545, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 500, + 545, + 555 + ], + "spans": [ + { + "bbox": [ + 307, + 500, + 545, + 555 + ], + "type": "text", + "content": "[47] Mark Sandler, Andrew Howard, Menglong Zhu, Andrey Zh-moginov, and Liang-Chieh Chen. Mobilenetv2: Inverted residuals and linear bottlenecks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4510-4520, 2018. 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 556, + 545, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 556, + 545, + 578 + ], + "spans": [ + { + "bbox": [ + 307, + 556, + 545, + 578 + ], + "type": "text", + "content": "[48] Robert R. Schaller. Moore's law: past, present and future. IEEE spectrum, 1997. 14" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 579, + 545, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 579, + 545, + 633 + ], + "spans": [ + { + "bbox": [ + 307, + 579, + 545, + 633 + ], + "type": "text", + "content": "[49] Zhuoran Shen, Mingyuan Zhang, Haiyu Zhao, Shuai Yi, and Hongsheng Li. Efficient attention: Attention with linear complexities. In Proceedings of the IEEE/CVF winter conference on applications of computer vision, pages 3531-3539, 2021. 3, 7" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 635, + 545, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 635, + 545, + 668 + ], + "spans": [ + { + "bbox": [ + 307, + 635, + 545, + 668 + ], + "type": "text", + "content": "[50] Chen Sun, Abhinav Shrivastava, Saurabh Singh, and Abhinav Gupta. Revisiting unreasonable effectiveness of data in deep learning era. In ICCV, 2017. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 669, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 669, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 307, + 669, + 545, + 712 + ], + "type": "text", + "content": "[51] Mingxing Tan and Quoc Le. Efficientnet: Rethinking model scaling for convolutional neural networks. In International conference on machine learning, pages 6105-6114. PMLR, 2019. 6" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "14440" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 117 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 117 + ], + "type": "text", + "content": "[52] Mingxing Tan, Ruoming Pang, and Quoc V Le. Efficientdet: Scalable and efficient object detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10781-10790, 2020. 6, 7" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 118, + 288, + 152 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 118, + 288, + 152 + ], + "spans": [ + { + "bbox": [ + 49, + 118, + 288, + 152 + ], + "type": "text", + "content": "[53] Shitao Tang, Jiahui Zhang, Siyu Zhu, and Ping Tan. Quadtree attention for vision transformers. In International Conference on Learning Representations, 2022. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 153, + 288, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 153, + 288, + 228 + ], + "spans": [ + { + "bbox": [ + 49, + 153, + 288, + 228 + ], + "type": "text", + "content": "[54] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Herve Jegou. Training data-efficient image transformers & distillation through attention. In Proceedings of the 38th International Conference on Machine Learning, volume 139 of Proceedings of Machine Learning Research, pages 10347-10357. PMLR, 18-24 Jul 2021. 2, 5, 6, 14" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 230, + 288, + 284 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 230, + 288, + 284 + ], + "spans": [ + { + "bbox": [ + 49, + 230, + 288, + 284 + ], + "type": "text", + "content": "[55] Hugo Touvron, Matthieu Cord, Alexandre Sablayrolles, Gabriel Synnaeve, and Hervé Jégou. Going deeper with image transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 32-42, 2021. 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 49, + 286, + 287, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 286, + 287, + 353 + ], + "spans": [ + { + "bbox": [ + 49, + 286, + 287, + 353 + ], + "type": "text", + "content": "[56] Zhengzhong Tu, Hossein Talebi, Han Zhang, Feng Yang, Peyman Milanfar, Alan Bovik, and Yinxiao Li. Maxvit: Multi-axis vision transformer. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXIV, pages 459-479. Springer, 2022. 2, 13" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 49, + 354, + 287, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 354, + 287, + 408 + ], + "spans": [ + { + "bbox": [ + 49, + 354, + 287, + 408 + ], + "type": "text", + "content": "[57] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Ilia Polosukhin. Attention is all you need. In Advances in Neural Information Processing Systems, volume 30. Curran Associates, Inc., 2017. 1, 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 49, + 410, + 287, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 410, + 287, + 442 + ], + "spans": [ + { + "bbox": [ + 49, + 410, + 287, + 442 + ], + "type": "text", + "content": "[58] Sinong Wang, Belinda Z. Li, Madian Khabsa, Han Fang, and Hao Ma. Linformer: Self-attention with linear complexity, 2020. 1, 2, 3, 13" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 49, + 445, + 287, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 445, + 287, + 499 + ], + "spans": [ + { + "bbox": [ + 49, + 445, + 287, + 499 + ], + "type": "text", + "content": "[59] Wenhai Wang, Enze Xie, Xiang Li, Deng-Ping Fan, Kaitao Song, Ding Liang, Tong Lu, Ping Luo, and Ling Shao. Pyramid vision transformer: A versatile backbone for dense prediction without convolutions. arXiv preprint arXiv:2102.12122, 2021. 2, 6, 14" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 49, + 501, + 287, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 501, + 287, + 567 + ], + "spans": [ + { + "bbox": [ + 49, + 501, + 287, + 567 + ], + "type": "text", + "content": "[60] Wenhai Wang, Enze Xie, Xiang Li, Deng-Ping Fan, Kaitao Song, Ding Liang, Tong Lu, Ping Luo, and Ling Shao. Pyramid vision transformer: A versatile backbone for dense prediction without convolutions. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 568-578, October 2021. 3, 13" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 49, + 568, + 287, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 568, + 287, + 612 + ], + "spans": [ + { + "bbox": [ + 49, + 568, + 287, + 612 + ], + "type": "text", + "content": "[61] Wenhai Wang, Enze Xie, Xiang Li, Deng-Ping Fan, Kaitao Song, Ding Liang, Tong Lu, Ping Luo, and Ling Shao. Pvt v2: Improved baselines with pyramid vision transformer. Computational Visual Media, 8(3):415-424, 2022. 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 49, + 613, + 287, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 613, + 287, + 667 + ], + "spans": [ + { + "bbox": [ + 49, + 613, + 287, + 667 + ], + "type": "text", + "content": "[62] Bichen Wu, Chaojian Li, Hang Zhang, Xiaoliang Dai, Peizhao Zhang, Matthew Yu, Jialiang Wang, Yingyan Lin, and Peter Vajda. Fbnetv5: Neural architecture search for multiple tasks in one run. arXiv preprint arXiv:2111.10007, 2021.6,7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 49, + 669, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 669, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 49, + 669, + 287, + 713 + ], + "type": "text", + "content": "[63] Haiping Wu, Bin Xiao, Noel Codella, Mengchen Liu, Xiyang Dai, Lu Yuan, and Lei Zhang. Cvt: Introducing convolutions to vision transformers. arXiv preprint arXiv:2103.15808, 2021. 2" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 25, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 115 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 115 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 115 + ], + "type": "text", + "content": "[64] Zhanghao Wu, Zhijian Liu, Ji Lin, Yujun Lin, and Song Han. Lite transformer with long-short range attention. In International Conference on Learning Representations, 2020. 1" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 308, + 118, + 545, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 118, + 545, + 172 + ], + "spans": [ + { + "bbox": [ + 308, + 118, + 545, + 172 + ], + "type": "text", + "content": "[65] Yunyang Xiong, Hanxiao Liu, Suyog Gupta, Berkin Akin, Gabriel Bender, Pieter-Jan Kindermans, Mingxing Tan, Vikas Singh, and Bo Chen. Mobiledets: Searching for object detection architectures for mobile accelerators. arXiv preprint arXiv:2004.14525, 2020. 2, 6, 7" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 308, + 173, + 545, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 173, + 545, + 237 + ], + "spans": [ + { + "bbox": [ + 308, + 173, + 545, + 237 + ], + "type": "text", + "content": "[66] Yunyang Xiong, Zhanpeng Zeng, Rudrasis Chakraborty, Mingxing Tan, Glenn Fung, Yin Li, and Vikas Singh. Nystromformer: A nystrom-based algorithm for approximating self-attention. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 35, pages 14138–14148, 2021. 1, 2, 13" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 239, + 545, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 239, + 545, + 282 + ], + "spans": [ + { + "bbox": [ + 308, + 239, + 545, + 282 + ], + "type": "text", + "content": "[67] Zhi-Qin John Xu, Yaoyu Zhang, Tao Luo, Yanyang Xiao, and Zheng Ma. Frequency principle: Fourier analysis sheds light on deep neural networks. arXiv preprint arXiv:1901.06523, 2019. 2, 5, 8" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 283, + 545, + 337 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 283, + 545, + 337 + ], + "spans": [ + { + "bbox": [ + 308, + 283, + 545, + 337 + ], + "type": "text", + "content": "[68] Hongxu Yin, Arash Vahdat, Jose M Alvarez, Arun Mallya, Jan Kautz, and Pavlo Molchanov. A-vit: Adaptive tokens for efficient vision transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10809-10818, 2022. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 338, + 545, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 338, + 545, + 403 + ], + "spans": [ + { + "bbox": [ + 308, + 338, + 545, + 403 + ], + "type": "text", + "content": "[69] Haoran You, Zhanyi Sun, Huihong Shi, Zhongzhi Yu, Yang Zhao, Yongan Zhang, Chaojian Li, Baopu Li, and Yingyan Lin. Vitcod: Vision transformer acceleration via dedicated algorithm and accelerator co-design. In The 29th IEEE International Symposium on High-Performance Computer Architecture (HPCA 2023), 2023. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 404, + 545, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 404, + 545, + 459 + ], + "spans": [ + { + "bbox": [ + 308, + 404, + 545, + 459 + ], + "type": "text", + "content": "[70] Guanghua Yu, Qinyao Chang, Wenyu Lv, Chang Xu, Cheng Cui, Wei Ji, Qingqing Dang, Kaipeng Deng, Guanzhong Wang, Yuning Du, et al. Pp-picodet: A better real-time object detector on mobile devices. arXiv preprint arXiv:2111.00902, 2021. 5, 6, 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 460, + 545, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 460, + 545, + 514 + ], + "spans": [ + { + "bbox": [ + 308, + 460, + 545, + 514 + ], + "type": "text", + "content": "[71] Zhongzhi Yu, Yonggan Fu, Sicheng Li, Chaojian Li, and Yingyan Lin. Mia-former: Efficient and robust vision transformers via multi-grained input-adaptation. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 36, pages 8962-8970, 2022. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 515, + 545, + 579 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 515, + 545, + 579 + ], + "spans": [ + { + "bbox": [ + 308, + 515, + 545, + 579 + ], + "type": "text", + "content": "[72] Li Yuan, Yunpeng Chen, Tao Wang, Weihao Yu, Yujun Shi, Zi-Hang Jiang, Francis EH Tay, Jiashi Feng, and Shuicheng Yan. Tokens-to-token vit: Training vision transformers from scratch onImagenet. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 558-567, 2021. 2, 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 581, + 545, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 581, + 545, + 635 + ], + "spans": [ + { + "bbox": [ + 308, + 581, + 545, + 635 + ], + "type": "text", + "content": "[73] Yuhui Yuan, Rao Fu, Lang Huang, Weihong Lin, Chao Zhang, Xilin Chen, and Jingdong Wang. Hrformer: High-resolution vision transformer for dense predict. Advances in Neural Information Processing Systems, 34:7281-7293, 2021.6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 636, + 545, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 636, + 545, + 690 + ], + "spans": [ + { + "bbox": [ + 308, + 636, + 545, + 690 + ], + "type": "text", + "content": "[74] Zixiao Zhang, Xiaogiang Lu, Guojin Cao, Yuting Yang, Licheng Jiao, and Fang Liu. Vit-yolo: Transformer-based yolo for object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2799-2808, 2021. 13" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 691, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 691, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 691, + 545, + 713 + ], + "type": "text", + "content": "[75] Bolei Zhou, Hang Zhao, Xavier Puig, Sanja Fidler, Adela Barriuso, and Antonio Torralba. Scene parsing through" + } + ] + } + ], + "index": 24 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "14441" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 286, + 206 + ], + "type": "list", + "angle": 0, + "index": 3, + "blocks": [ + { + "bbox": [ + 66, + 72, + 286, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 72, + 286, + 105 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 286, + 105 + ], + "type": "text", + "content": "ade20k dataset. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 633-641, 2017. 5" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 107, + 286, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 107, + 286, + 162 + ], + "spans": [ + { + "bbox": [ + 48, + 107, + 286, + 162 + ], + "type": "text", + "content": "[76] Chen Zhu, Wei Ping, Chaowei Xiao, Mohammad Shoeybi, Tom Goldstein, Anima Anandkumar, and Bryan Catanzaro. Long-short transformer: Efficient transformers for language and vision. Advances in Neural Information Processing Systems, 34, 2021. 1, 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 163, + 286, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 163, + 286, + 206 + ], + "spans": [ + { + "bbox": [ + 49, + 163, + 286, + 206 + ], + "type": "text", + "content": "[77] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable detr: Deformable transformers for end-to-end object detection. arXiv preprint arXiv:2010.04159, 2020. 3, 13" + } + ] + } + ], + "index": 2 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "14442" + } + ] + } + ], + "index": 4 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Catch Missing Details_ Image Reconstruction With Frequency Augmented Variational Autoencoder/0a253bb0-c0db-46ba-9d61-8be0d603b311_content_list.json b/2023/Catch Missing Details_ Image Reconstruction With Frequency Augmented Variational Autoencoder/0a253bb0-c0db-46ba-9d61-8be0d603b311_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..cac81fcefbb49c9558bbe1455102bd6fd59ca0bf --- /dev/null +++ b/2023/Catch Missing Details_ Image Reconstruction With Frequency Augmented Variational Autoencoder/0a253bb0-c0db-46ba-9d61-8be0d603b311_content_list.json @@ -0,0 +1,1780 @@ +[ + { + "type": "text", + "text": "Catch Missing Details: Image Reconstruction with Frequency Augmented Variational Autoencoder", + "text_level": 1, + "bbox": [ + 114, + 128, + 854, + 174 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Xinmiao Lin", + "bbox": [ + 220, + 204, + 328, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Rochester Institute of Technology", + "bbox": [ + 138, + 222, + 408, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "x13439@rit.edu", + "bbox": [ + 210, + 242, + 339, + 255 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Chiuman Ho", + "bbox": [ + 294, + 263, + 398, + 280 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "OPPO US Research", + "bbox": [ + 267, + 282, + 426, + 297 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "chiuman@oppo.com", + "bbox": [ + 274, + 303, + 419, + 316 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yikang Li", + "bbox": [ + 496, + 203, + 581, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "OPPO US Research", + "bbox": [ + 459, + 222, + 619, + 237 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "yikang.li1@oppo.com", + "bbox": [ + 452, + 242, + 624, + 256 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Jenhao Hsiao", + "bbox": [ + 694, + 204, + 803, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "OPPO US Research", + "bbox": [ + 669, + 222, + 828, + 237 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "mark@oppo.com", + "bbox": [ + 689, + 243, + 807, + 256 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 351, + 313, + 367 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The popular VQ-VAE models reconstruct images through learning a discrete codebook but suffer from a significant issue in the rapid quality degradation of image reconstruction as the compression rate rises. One major reason is that a higher compression rate induces more loss of visual signals on the higher frequency spectrum which reflect the details on pixel space. In this paper, a Frequency Complement Module (FCM) architecture is proposed to capture the missing frequency information for enhancing reconstruction quality. The FCM can be easily incorporated into the VQ-VAE structure, and we refer to the new model as Frequency Augmented VAE (FA-VAE). In addition, a Dynamic Spectrum Loss (DSL) is introduced to guide the FCMs to balance between various frequencies dynamically for optimal reconstruction. FA-VAE is further extended to the text-to-image synthesis task, and a Cross-attention Autoregressive Transformer (CAT) is proposed to obtain more precise semantic attributes in texts. Extensive reconstruction experiments with different compression rates are conducted on several benchmark datasets, and the results demonstrate that the proposed FA-VAE is able to restore more faithfully the details compared to SOTA methods. CAT also shows improved generation quality with better image-text semantic alignment.", + "bbox": [ + 75, + 385, + 473, + 750 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 782, + 209, + 797 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "VQ-VAE models [6,11,25,29,39,46] reconstruct images through learning a discrete codebook of latent embeddings. They gained wide popularity due to the scalable and versatile codebook, which can be broadly applied to many visual tasks such as image synthesis [11,49] and inpainting [4,31]. A higher compression rate is typically preferable in VQ-", + "bbox": [ + 75, + 809, + 468, + 902 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yu Kong \nMichigan State University", + "text_level": 1, + "bbox": [ + 491, + 263, + 702, + 300 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "yukong@msu.edu", + "bbox": [ + 531, + 303, + 658, + 316 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/47894a809973432e07174ef8fa48bbe6243ccdecc6c21b8ad23d4aa1df795a9e.jpg", + "image_caption": [ + "Figure 1. Images and their frequency maps. Row 1: original and reconstructed images. Row 2: the frequency maps of images, frequency increases in any direction away from the center. $f$ is the compression rate. With a greater compression rate, more details are lost during reconstruction, i.e. eyes and mouth shape, and hair texture (pointed with red arrows) which align with the loss of high-frequency features. All frequency figures in this paper use the same colormap. rFID [14] and lpips [16] are lower the better, and frequency values increase from red to green, zoom in for better visualization." + ], + "image_footnote": [], + "bbox": [ + 501, + 352, + 890, + 503 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "VAE models since it provides memory efficiency and better learning of coherent semantics structures [11, 40].", + "bbox": [ + 496, + 652, + 892, + 683 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "One main challenge quickly arises for a higher compression rate, which severely compromises reconstruction accuracy. Figure 1 row 1 shows that although the reconstructed images at higher compression rates appear consistent with the original image, details inconsistencies such as the color and contour of the lips become apparent upon closer scrutiny. Figure 1 row 2 reveals that similar degradation also manifests on the frequency domains where features towards the middle and higher frequency spectrum are the least recoverable with greater compression rate.", + "bbox": [ + 496, + 686, + 892, + 837 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Several causes stand behind this gap between pixel and frequency space. The convolutional nature of autoencoders is prone to spectral bias, which favors learning low-frequency features [22, 36]. This challenge is further ag-", + "bbox": [ + 496, + 839, + 893, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "1736", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "gravated when current methods exclusively design losses or improve model architecture for better semantics resemblance [11, 16, 25] but often neglect the alignment on the frequency domain [12, 15]. On top of that, it is intuitively more challenging for a decoder to reconstruct an image patch from a single codebook embedding (high compression) than multiple embeddings (less compression). The reason is that the former mixes up features of incomplete and diverse frequencies, while the latter could preserve more fine-grained and complete features at various frequencies.", + "bbox": [ + 75, + 90, + 472, + 242 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Inspired by these insights, the Frequency Augmented VAE (FA-VAE) model is proposed, which aims to improve reconstruction quality by achieving better alignment on the frequency spectrums between the original and reconstructed images. More specifically, new modules named Frequency Complement Modules (FCM) are crafted and embedded at multiple layers of FA-VAE's decoder to learn to complement the decoder's features with missing frequencies.", + "bbox": [ + 76, + 244, + 472, + 364 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We observe that valuable middle and high frequencies are mingled with the encoder's feature maps during the compression via an encoder, shown in Figure 3 row 4. Therefore, a new loss termed Spectrum Loss (SL) is proposed to guide FCMs to generate missing features that align with the same level's encoder features on the frequency domain. Since most image semantics reside on the low-frequency spectrum [48], SL prioritizes learning lower-frequency features with diminishing weights as frequencies increase.", + "bbox": [ + 75, + 367, + 472, + 518 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Interestingly, we discover that checkerboard patterns appear in the complemented decoder's features with SL, although better reconstruction performance is achieved (Figure 3 column 4). We speculate that because SL sets a deterministic range for the low-frequency spectrum when applying weights on the frequencies without considering that the importance of a frequency can vary from layer to layer. Thus, an improved loss function Dynamic Spectrum Loss (DSL) is crafted on top of SL with a learnable component to adjust the range of low-frequency spectrum dynamically for optimal reconstruction. DSL can improve reconstruction quality even further than SL without the unnatural checkerboard artifacts in the features (Figure 3 column 5).", + "bbox": [ + 75, + 520, + 472, + 717 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We further extend FA-VAE to the text-to-image generation task and propose the Cross-attention Autoregressive Transformer (CAT) model. We first observe that only using one or a few token embeddings is a coarse representation of lengthy texts [8, 27, 33]. Thus CAT uses all token embeddings as a condition for more precise guidance. Moreover, existing works typically use self-attention, and the text condition is embedded merely at the beginning of the generation [11, 49]. This mechanism becomes problematic in the autoregressive generation because one image token is generated at a time, thus the text condition gradually loosens its connection with the generated tokens. To circumvent this", + "bbox": [ + 75, + 719, + 472, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "issue, CAT embeds a cross-attention mechanism that allows the text condition to guide each step generation.", + "bbox": [ + 498, + 90, + 890, + 121 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To summarize, our work includes the following contributions:", + "bbox": [ + 498, + 121, + 890, + 151 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose a new type of architecture called Frequency Augmented VAE (FA-VAE) for improving image reconstruction through achieving more accurate details reconstruction.", + "- We propose a new loss called Spectrum Loss (SL) and its enhanced version Dynamic Spectrum Loss (D SL), which guides the Frequency Complement Modules (FCM) in FA-VAE to adaptively learn different low/high frequency mixtures for optimal reconstruction.", + "- We propose a new Cross-attention Autoregressive Transformer (CAT) for text-to-image generation using more fine-grained textual embeddings as a condition with a cross-attention mechanism for better image-text semantic alignment." + ], + "bbox": [ + 517, + 151, + 890, + 377 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work & Background", + "text_level": 1, + "bbox": [ + 500, + 391, + 769, + 409 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Image Reconstruction Vector Quantized-Variational AutoEncoder (VQ-VAE) [46] extends the Variational AutoEncoder structure [23] and proposes to encode images into discrete latent codes using vector quantization (VQ). Then a generative model, such as an autoregressive transformer [13], can be trained and paired with the decoder in VQ-VAE to synthesize new images. Later works [11,25,39] further improve the reconstruction and generation quality by improving the generative model architecture or the quantization efficiency. Since VQ-VAE models operate on discrete latent spaces, they cannot be directly and fairly compared to other VAE-based models that employ continuous latent space [1,6,23,29,45,51]. In contrast to current VQ-VAE models, our proposed FA-VAE technique improves reconstruction through the frequency angle and is more generalized since the FCMs can be readily extended to other neural networks that share the same VQ-VAE structure.", + "bbox": [ + 496, + 417, + 890, + 672 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Few other works do image reconstruction via the frequency perspective. FFL [15] proposed a loss function to penalize differences on the hard frequencies. A new module, which is designed in the continuous latent space for image compression, is proposed to be incorporated into the encoder and decoder in [12]. To our best knowledge, FA-VAE is the first work that aims to improve image reconstruction on discrete latent space through the frequency perspective.", + "bbox": [ + 496, + 674, + 890, + 795 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Image Generation Image generation can be achieved via GAN-based models [3, 18, 42, 44, 52, 55], which synthesize images from noise vectors unconditionally or conditioned on different inputs such as texts, masks, etc. It becomes cumbersome to train one model for one application. Thus, StyleGAN [21] encodes semantic attributes into a continuous latent space, and subsequent works [19, 20,", + "bbox": [ + 496, + 795, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "1737", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/54efa40dc53ecdaf9a16925bc2ca99f28be9dba84c340692f3429da4f3c8bce4.jpg", + "image_caption": [ + "Figure 2. Frequency Augmented VAE (FA-VAE). (a) The encoder $\\mathcal{E}$ encodes the images $X$ onto discrete latent codebook space $\\hat{z}_q$ which is used by the decoder $\\mathcal{G}$ to reconstruct images $\\hat{X}$ . DSL (b) guides the FCMs (c) to learn to complement the reconstructed features with missing features of important frequencies in order to improve reconstruction quality. A more detailed figure is in the supplement." + ], + "image_footnote": [], + "bbox": [ + 76, + 88, + 426, + 268 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/b8d3e2c0f9326268fe96599a321282dea36c00873d5856951720531629ea190d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 426, + 88, + 890, + 268 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "24, 32, 43, 50, 54] leverage this space for generating images conditioned on attributes or textual descriptions. However, StyleGAN-based models cannot scale to large datasets when the number of attributes becomes substantially large because this demands to increase model's size. In contrast, the codebook in VQ-VAE models [46] is scalable to large datasets without additional model complexity. Diffusion-based models [5, 7, 10, 27, 28, 37, 40, 41] generate images from Gaussian noise through a reverse diffusion process can often require substantial training and large datasets. Most text-to-image generation models commonly use one of a few embeddings for a condition [8, 27, 38, 49] from pretrained models such as CLIP [33] or T5-XXL [35]. In contrast, our proposed autoregressive transformer CAT uses the embeddings from all the text tokens for more fine-grained guidance during image generation.", + "bbox": [ + 75, + 321, + 472, + 563 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "VQ-VAE We now describe the VQ-VAE model [46] as it is the backbone used in the FA-VAE model, which is presented in Figure 2. VQ-VAE [46] and related models [11, 25, 38, 39] consist of an encoder that encodes the images into a codebook of embeddings. Then, a decoder is trained to reconstruct the images from a discrete set of codebook embeddings. In this paper, reconstruction loss in VQ-GAN is utilized which is:", + "bbox": [ + 75, + 564, + 470, + 684 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {r e c} = \\left\\| X - \\hat {X} \\right\\| _ {1} + \\mathcal {L} _ {p i p s} (X - \\hat {X}) \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 148, + 694, + 468, + 714 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $X$ and $\\hat{X}$ are the original and reconstructed images. We also use the adversarial setting of VQ-GAN model which introduced a discriminator compared to VQ-VAE models [39, 46], more details could be referred to [11].", + "bbox": [ + 76, + 724, + 468, + 786 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Methodology", + "text_level": 1, + "bbox": [ + 76, + 800, + 210, + 816 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The proposed Frequency Augmented VAE (FA-VAE) is presented in Figure 2. In this section, we describe how FA-VAE ameliorates reconstruction quality by bridging the spectrum domain gap between reconstructed and original images. By explicitly embedding Frequency Complement", + "bbox": [ + 76, + 824, + 470, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Modules (FCM) into the decoder, Dynamic Spectrum Loss (DSL) leverages important features frequency-wise from the encoder and guides the FCMs to complement the features of the decoder for better frequency restoration at different reconstruction stages.", + "bbox": [ + 496, + 321, + 893, + 397 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Frequency Augmented VAE (FA-VAE)", + "text_level": 1, + "bbox": [ + 498, + 410, + 830, + 426 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Let the images be $X \\in \\mathbb{R}^{H \\times W \\times 3}$ . The codebook $\\mathfrak{C}$ is a set of $|\\mathfrak{C}|$ embeddings, such as $\\mathfrak{C} = \\{\\mathfrak{c}_i | i = 1, \\dots, |\\mathfrak{c}|\\} \\in \\mathbb{R}^{n_z}$ and $n_z$ is the length of one codebook embedding. FA-VAE consists of an encoder $\\mathcal{E}$ that encodes the images $X$ into latent space representations such as $z = \\mathcal{E}(x)$ , for $z \\in \\mathbb{R}^{h \\times w \\times n_z}$ and $x \\in X$ , $(h \\times w)$ is the resolution of the encoded representation. Let $f = H / h = W / w$ be the downsampling factor or compression rate. Each feature of $z$ is approximated by the vector quantization block $\\mathcal{Q}$ using the nearest codebook entries, and it can be presented as:", + "bbox": [ + 496, + 433, + 893, + 585 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {Q} (z) = \\hat {z} _ {q} = \\underset {c _ {k} \\in \\mathcal {C}} {\\arg \\min } \\| z _ {i j} - \\mathfrak {c} _ {k} \\| \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 584, + 599, + 890, + 625 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\hat{z}_q$ is the quantized latent embedding and subsequently used by the decoder $\\mathcal{G}$ to produce the reconstructed image $\\hat{x} = \\mathcal{G}(\\hat{z}_q)$ .", + "bbox": [ + 496, + 638, + 893, + 685 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1.1 Frequency Complement Modules (FCM)", + "text_level": 1, + "bbox": [ + 500, + 708, + 836, + 723 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Motivation Figure 1 shows that a higher compression rate leads to more significant reconstruction disparities on the higher-frequency spectrum. Existing models [11, 25, 39] guide the reconstructed images to be more aligned on the pixel and feature spaces with the original images (Eq. 1) but neglect frequency spectrum alignment and leave the encoder and decoder without further guidance. Figure 3 column 2 shows that the encoder activations $\\mathcal{A}_1$ of baseline VQ-GAN [11] contain rich high-frequency features (row 3 & 4), but the decoder's activations could mainly restore low-frequency features (row 5 & 6).", + "bbox": [ + 496, + 734, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "1738", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Therefore we propose Frequency Complement Modules (FCM), illustrated in Figure 2 (c), which aims to complement the decoder's feature maps $\\mathcal{B}_i$ with features of missing frequencies using the encoder activations $\\mathcal{A}_i$ . The FCMs $\\mathcal{F}_i$ consist of sequences of convolution layers and activations. The decoder $\\mathcal{G}$ with FCMs embedded can be represented as:", + "bbox": [ + 76, + 90, + 472, + 183 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {B} _ {M - 1} = \\mathcal {G} _ {M} \\left(\\mathcal {C} _ {M} + \\hat {z} _ {q}\\right) = \\mathcal {G} _ {M} \\left(\\mathcal {F} _ {M} (\\hat {z} _ {q}) + \\hat {z} _ {q}\\right)\n$$\n", + "text_format": "latex", + "bbox": [ + 104, + 207, + 421, + 224 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\dots \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 130, + 227, + 468, + 241 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {B} _ {1} = \\mathcal {G} _ {2} \\left(\\mathcal {C} _ {2} + \\mathcal {B} _ {2}\\right) = \\mathcal {G} _ {2} \\left(\\mathcal {F} _ {2} \\left(\\mathcal {B} _ {2}\\right) + \\mathcal {B} _ {2}\\right)\n$$\n", + "text_format": "latex", + "bbox": [ + 130, + 244, + 395, + 261 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Similarly, the encoder $\\mathcal{E}$ can be abstracted to:", + "bbox": [ + 76, + 273, + 379, + 287 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nz = \\mathcal {E} (x) = \\mathcal {E} _ {M} \\left(\\dots \\left(\\mathcal {E} _ {2} (\\mathcal {E} _ {1} (x)))\\right)\\right) \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 156, + 301, + 468, + 319 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathcal{G}_i$ is the $i$ -th layer of the decoder, and $\\mathcal{A}_i = \\mathcal{E}_i(x)$ is the outputs or feature maps of the corresponding encoder block. The outputs of previous blocks $\\mathcal{B}_i$ are complemented by outputs $\\mathcal{C}_i = \\mathcal{F}_i(\\mathcal{B}_{i + 1})$ , which contain the frequency-rich features learned from the encoder activations $\\mathcal{A}_i$ . The following section describes how Dynamic Frequency Loss (DSL) guides the learning of FCMs to specifically learn to complement $\\mathcal{B}_i$ with the features of missing frequencies.", + "bbox": [ + 75, + 330, + 468, + 450 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In the paper, the frequency feature compensation is implemented by addition (in Eq. 3). However, the FCM is flexible in adopting any architecture and merging techniques, as illustrated by the three examples in Figure 5. Note that no architecture limitation is imposed on the encoder and decoder blocks as long as $\\mathcal{A}_i$ and $\\mathcal{B}_i$ share the same resolution. VQ-GAN [11] and related models [25,46] already have similar architecture, and VQ-GAN is chosen as the backbone for FA-VAE.", + "bbox": [ + 75, + 453, + 468, + 587 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1.2 Dynamic Spectrum Loss (DSL)", + "text_level": 1, + "bbox": [ + 76, + 609, + 348, + 626 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Motivation Spectrum Loss (SL) is proposed to guide explicitly the outputs of FCMs $\\mathcal{C}_i$ to be more aligned with the encoder's activations $\\mathcal{A}_i$ on the frequency spectrum because the latter contains rich features on the higher frequency spectrum (Figure 3). Moreover, to account for the varying importance of frequencies across different decoder stages, Dynamic Spectrum Loss (DSL) is proposed, a more generalized variant of SL, and has the ability to adjust weights put on higher-frequency spectrum adaptively. Thus, each decoder block's outputs can be enriched with features of the most critical frequencies for an accurate reconstruction.", + "bbox": [ + 75, + 635, + 468, + 801 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Background The outputs $\\mathcal{A}_i$ of encoder block $\\mathcal{E}_i$ and $\\mathcal{C}_i$ of FCM block $\\mathcal{F}_i$ are first transformed to the frequency domain using Discrete Fourier Transform (DFT) as follows:", + "bbox": [ + 76, + 801, + 470, + 848 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nF (u, v) = \\sum_ {x = 0} ^ {M - 1} \\sum_ {y = 0} ^ {N - 1} f (x, y) \\cdot e ^ {- i 2 \\pi \\left(\\frac {u x}{M} + \\frac {v y}{N}\\right)} \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 125, + 859, + 468, + 902 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/b90b2d4f5584033841c23e4b3bf314d6aad03e39cb0d50246f860c60030de423.jpg", + "image_caption": [ + "Figure 3. Image, activations, and their frequency maps. Ours* is FA-VAE model with FFL [15], Ours‡ is FA-VAE with SL and Ours is FA-VAE model with DSL. DSL shows a more harmonious balance between low- and high-frequencies (row 5) and more accurate reconstruction in mouth and hair textures. The frequency maps are plotted using the average of all channels and the contrast is adjusted to emphasize the higher frequency spectrums." + ], + "image_footnote": [], + "bbox": [ + 517, + 90, + 882, + 425 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $e$ and $i$ are Euler's number and the imaginary unit. $M \\times N$ is the spatial resolution of feature maps and the Fourier Transform (Eq. 5) is applied to each of them. $f(x,y)$ is the value at $(x,y)$ of each feature map. $F(u,v)$ is the corresponding value at $(u,v)$ coordinates on the frequency spectrum. The Focal Frequency Loss (FFL) [15] can be presented as:", + "bbox": [ + 498, + 530, + 890, + 636 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {F F L} \\left(\\mathcal {A} _ {i}, \\mathcal {C} _ {i}\\right) = \\frac {1}{M N \\left| \\mathcal {C} _ {i} \\right|} \\sum_ {c = 0} ^ {\\left| \\mathcal {C} _ {i} \\right| - 1} \\sum_ {u = 0} ^ {M - 1} \\sum_ {v = 0} ^ {N - 1} w (u, v) J (u, v), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 506, + 648, + 890, + 703 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $w(u,v) = |F_{\\mathcal{A}_i}(u,v) - F_{\\mathcal{C}_i}(u,v)|$ are the weights put on each frequency. $J(u,v) = |F_{\\mathcal{A}_i}(u,v) - F_{\\mathcal{C}_i}(u,v)|^2$ is the main error function based on the frequency difference. The $|\\mathcal{C}_i|$ is the number of feature maps in $\\mathcal{A}_i$ and $\\mathcal{C}_i$ . Both real and imaginary parts of the frequency domain are considered, more details are in [15].", + "bbox": [ + 496, + 704, + 890, + 794 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Limitations of FFL Eq. 6 demonstrates that FFL puts modulating term $w(u, v)$ to focus learning on the hardest frequencies for reconstruction, which are the higher frequencies following our observation (Figures 1 & 3). However, this could not be ideal because features on the lower frequencies define the image content, and overemphasizing the higher frequencies could over-constrain the learning", + "bbox": [ + 496, + 795, + 890, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "1739", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/8a2e58a0498ae831d9b106743f2ee53ecdd60c1343d9197660e65a05a86863b1.jpg", + "image_caption": [ + "Figure 4. Cross-attention Autoregressive Transformer (CAT)." + ], + "image_footnote": [], + "bbox": [ + 76, + 87, + 467, + 199 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "and lead to sub-optimal reconstruction, see Figure 7 row 1 column 3 and Table 2 row 3. Moreover, Figure 3 column 3 shows FCMs guided by a simple FFL can improve reconstruction performance (Ours\\*). However the frequency maps of the decoder activations $\\mathcal{C}_1$ contain excessive noise due to overemphasis on the higher frequency spectrum, and lower frequencies are neglected (row 6).", + "bbox": [ + 75, + 219, + 468, + 325 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Spectrum Loss (SL) Thus, we propose to apply a low-pass filter on the weights $w(u,v)$ in Eq. 6 to penalize more mismatch in the lower-frequency domain and gradually diminish the penalizing weights towards higher frequency spectrum. Therefore, let the Gaussian kernels be $K_{i}(\\mu ,\\sigma)$ with weights initialized using mean and standard deviation $(\\mu ,\\sigma_{i})$ , and applied over the feature maps as:", + "bbox": [ + 76, + 325, + 470, + 433 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n(\\hat {\\mathcal {A}} _ {i}, \\hat {\\mathcal {C}} _ {i}) = (K _ {i} (\\mu , \\sigma_ {i}) \\star \\mathcal {A} _ {i}, K _ {i} (\\mu , \\sigma_ {i}) \\star \\mathcal {C} _ {i}). \\qquad (7)\n$$\n", + "text_format": "latex", + "bbox": [ + 125, + 440, + 468, + 460 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where the $\\star$ is the convolution operation. Then, the Spectrum Loss (SL) is defined as:", + "bbox": [ + 76, + 469, + 468, + 500 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {S L} \\left(\\mathcal {A} _ {i}, \\mathcal {C} _ {i}\\right) = \\operatorname {F F L} \\left(\\hat {\\mathcal {A}} _ {i}, \\hat {\\mathcal {C}} _ {i}\\right). \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 179, + 508, + 468, + 529 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Limitations of SL Up until now, a fixed Gaussian filter $K_{i}(\\mu ,\\sigma_{i})$ is applied over all $\\mathcal{A}_i$ and $\\mathcal{C}_i$ of all encoder and decoder blocks. Figure 3 column 4 demonstrates that SL $(\\mathrm{Ours}^{\\ddagger})$ improves reconstruction on the lower frequency spectrum, leading to better reconstruction than the baseline model (Table 2 row 5 vs row 1). However, checkerboard artifacts are also present (row 5) on $\\mathcal{C}_1$ . One reason is that deterministic variance $\\sigma_{i}$ assumes that decoder activations across different levels require the same amount of higher frequency features for accurate reconstruction, which could be an over-rigid constraint on the learning. Another reason could be that the same $\\sigma_{i}$ magnifies the checkerboard effects produced by upsampling or striding operations in CNNs which are normally circumvented by subsequent convolutional layers [2, 30]. In the observation of the experiments, we find that the checkerboard effects can be intensified with the same $\\sigma_{i}$ in different layers due to the bounding effect between low frequency and high frequency (shown in Figure 3 column 4 row 6).", + "bbox": [ + 75, + 537, + 470, + 824 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Dynamic Spectrum Loss (DSL) Therefore, we propose to optimize the variances $\\sigma_{i}$ instead of setting them as static hyperparameters to suit the different amounts of frequencies needed for each block's $\\mathcal{B}_i$ . The new Dynamic Spectrum Loss (DSL) is a more generalized form of Spectrum Loss", + "bbox": [ + 75, + 825, + 472, + 902 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/884396341330d2834aad988835914292b738c9ff0e91466b21cc7ee2afc6bf40.jpg", + "image_caption": [ + "(a) FCM" + ], + "image_footnote": [], + "bbox": [ + 514, + 92, + 617, + 253 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/a185b0a415dcdd350010fbac884698b49521abaacf2dc9398fea1b8bd44aff9e.jpg", + "image_caption": [ + "(b) FCM w/ residual" + ], + "image_footnote": [], + "bbox": [ + 630, + 89, + 738, + 255 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/2c8618e727623ee74960bb7ffaa264f574e20e07e09360b454a36d66134545a6.jpg", + "image_caption": [ + "(c) FCM w/ attention", + "Figure 5. FCM with different architectures and connections." + ], + "image_footnote": [], + "bbox": [ + 759, + 88, + 859, + 255 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "(SL) with learnable $\\sigma_{i}$ . Note that DSL also includes FFL as a special form when used on the original and reconstructed images without the Gaussian filters.", + "bbox": [ + 498, + 289, + 892, + 334 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Then the total reconstruction loss for FA-VAE can be described as:", + "bbox": [ + 498, + 334, + 890, + 363 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} _ {r e c} = \\alpha \\operatorname {F F L} (X, \\hat {X}) + \\beta \\sum_ {i = 0} ^ {M - 1} \\operatorname {D S L} \\left(\\mathcal {A} _ {i}, \\mathcal {C} _ {i}\\right) \\tag {9} \\\\ + \\left\\| X - \\hat {X} \\right\\| _ {1} + \\mathcal {L} _ {p i p s} (X - \\hat {X}). \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 547, + 367, + 890, + 431 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathrm{FFL}(X - \\hat{X})$ is the Focal Frequency Loss applied on the original reconstructed images. The second term in Eq. 9 is the DSL loss applied over the outputs of the encoder and FCM blocks. The first two losses aim to minimize the frequency spectrum differences on the images and internal feature maps. The third and fourth losses act on the pixel and feature maps space [11]. $\\alpha$ and $\\beta$ are hyperparameters. $\\sigma_{i}$ are model parameters and optimized as:", + "bbox": [ + 498, + 438, + 892, + 561 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\sigma_ {i} ^ {*}, \\mathcal {E} ^ {*}, \\mathcal {G} ^ {*}, \\mathcal {C} ^ {*} = \\underset {\\sigma_ {i}, \\mathcal {E}, \\mathcal {G}, \\mathcal {C}} {\\arg \\min } \\left(\\mathcal {L} _ {\\text {r e c}} + \\mathcal {L} _ {Q}\\right) \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 566, + 566, + 890, + 592 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "$\\mathcal{L}_Q$ is the quantization loss which minimizes the difference between the codebook embeddings and the embeddings given by the encoder $\\mathcal{E}$ , more details in [11]. We also use the $L_{2}$ regularization on the codebook embeddings during quantization as in [53] and exponential moving average (EMA) for updating the codebook [46] as they provide more stable training for the quantization block $\\mathcal{Q}$ .", + "bbox": [ + 498, + 598, + 890, + 703 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Benefits of DSL The advantages of learnable $\\sigma_{i}$ in DSL are shown in Figure 3 column 5 where the reconstructed activations show no checkerboard artifacts as in column 4. Moreover, the reconstructed activations $C_1$ are more similar to the encoder activations $A_{1}$ on the frequency spectrum. Compared to the baseline model (column 2), our model exhibits a more harmonious balance between low- and highfrequencies, leading to more accurate reconstructed images.", + "bbox": [ + 496, + 704, + 892, + 825 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2. Cross-attention Autoregressive Transformer (CAT)", + "text_level": 1, + "bbox": [ + 500, + 832, + 890, + 862 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We further extend FA-VAE to the text-to-image generation task and introduce a new Cross-attention Autoregres", + "bbox": [ + 498, + 869, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "1740", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/0ba7b2117c5b13a6e7cbcc7d10418207d814a840d3dd6f2e4fc0de38d9078575.jpg", + "image_caption": [ + "Figure 6. Reconstruction on ImageNet [9], label: Japanese terrier." + ], + "image_footnote": [], + "bbox": [ + 125, + 89, + 851, + 294 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "sive Transformer (CAT), presented in Figure 4. CAT uses all the token embeddings of a textual description given by the pretrained text encoder of CLIP model [33], while existing works mostly use one or partial textual embeddings [8, 38, 49]. Furthermore, CAT uses a cross-attention mechanism with the text tokens to guide the image generation at each step. This more fine-grained text condition allows the generation to capture more precisely the relationships of semantic attributes between text and image.", + "bbox": [ + 75, + 319, + 472, + 455 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\np (s \\mid c) = \\prod_ {i} p \\left(s _ {i} \\mid s _ {< i}, c\\right) \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 187, + 469, + 468, + 501 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Then the predicted indices $z'$ can be decoded to an image using FA-VAE's decoder $G$ . The loss is to maximize the log-likelihood of the data representations,", + "bbox": [ + 75, + 508, + 468, + 555 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {C A T}} = \\mathbb {E} _ {x \\sim p (x)} [ - \\log p (s) ] \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 179, + 566, + 468, + 584 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this paper, the GPT2 model [34], which is the default setting in [11, 49], is utilized as the backbone for CAT, and cross-attention mechanism inspired from [47] is applied with the GPT2 structure.", + "bbox": [ + 75, + 595, + 468, + 656 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 76, + 670, + 209, + 686 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Experimental Details", + "text_level": 1, + "bbox": [ + 76, + 695, + 276, + 710 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The datasets used in this paper are: (1) MultiModal CelebA-HQ dataset [50] with 30,000 high-resolution celebrity face images and each image comes with ten captions; (2) Flickr-Faces-HQ (FFHQ) dataset [21] of 70,000 high-resolution face images; (3) ImageNet [9] that contains around 1 million images from 1000 categories. Experiments are performed on V-100 GPUs and results are reported on the validation sets unless specified otherwise. For a fair comparison with existing models, the resolutions used for training are $256 \\times 256$ on all datasets. The details of training settings and hyperparameters are provided in the supplementary materials.", + "bbox": [ + 75, + 719, + 470, + 901 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/01040bda16eed070d927281fb1db8a90d01da9569144389ad52d856728b67117.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelDatasetCodebook Size(h × w)rFID ↓
RQ-VAE [25]FFHQ2048(8 × 8)5.33
FA-VAE (Ours)FFHQ2048(16 × 16)4.98
VQ-VAE-2 [39]ImageNet512(64 × 64) & (32 × 32)~ 10 (train)
VQ-GAN [40]ImageNet8192(64 × 64)1.06
FA-VAE (Ours)ImageNet8192(64 × 64)0.40
DALL-E [38]ImageNet8192(32 × 32)32.01
VQ-GAN [11]ImageNet16384(16 × 16)5.15
VQ-GAN [11]ImageNet1024(16 × 16)7.94
VQ-GAN [25]ImageNet16384(8 × 8)17.95
RQ-VAE† [46]ImageNet16384(8 × 8)10.77
RQ-VAE* [25]ImageNet16384(8 × 8)4.73
FA-VAE (Ours)ImageNet16384(16 × 16)4.60
", + "bbox": [ + 506, + 316, + 885, + 513 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1. Reconstruction results on the validation data of FFHQ [21] and ImageNet [9] respectively, more results in the supplement. $\\dagger$ means depth 2, and $*$ means depth 4 from RQ-VAE [25].", + "bbox": [ + 498, + 516, + 890, + 559 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Image Reconstruction", + "text_level": 1, + "bbox": [ + 500, + 571, + 705, + 585 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Reconstruction on FFHQ and Imagenet The experiment results of reconstruction on FFHQ [21] and ImageNet [9] are presented in Table 1. The results of the baselines are from the original paper. Note that smaller $h$ and $w$ means that the downsampling factor or the compression rate $f$ is larger. FA-VAE model shows improved reconstruction quality over baseline models across different compression rates in both datasets. The main reason is that FA-VAE can successfully reconstruct the important middle and high-frequencies which are neglected in baseline models.", + "bbox": [ + 496, + 595, + 890, + 747 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Figure 6 further supports our previous claim and demonstrates clear diverging reconstruction qualities between baseline models VQ-GAN [11], DALL-E [38], and our FA-VAE model. With a larger compression rate, Figure 6 row 2 shows that more middle and high frequencies are compressed in VQ-GAN and DALL-E. In comparison, FA-VAE can reconstruct middle and high frequencies more accurately, which translates to good representations with improved semantics and local details in Figure 6 row 1. For instance, at a compression rate of 16, the dog's tongue of", + "bbox": [ + 496, + 750, + 892, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "1741", + "bbox": [ + 483, + 944, + 513, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/d2f6ddcb3d6c15a1d5d328fbfe6eb42d7ae0dc3a6b8f88ab33955627e99b92d3.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ablation onFFL(X, X̂)SL(Ai, Ci)FCM Fkernel K_i(μ, σ_i)DSL(Ai, Ci)kernel size μinitial value σ_ipair-wise σ_iL1 ↓Lpips ↓rFID ↓
1VQ-GAN [11]XXXXXXXX0.1210.3010.12
2VQ-GAN + Style [53]XXXXXXXX0.0850.2311.90
3VQ-GAN [15]✓XXXXXXX0.1140.3530.65
4SL w/o kernel✓XCONVXXXXX0.0820.227.04
5✓✓CONVXXXXX0.0820.227.02
6SL w/ kernel✓✓CONV✓X93X0.0800.227.39
7FCM architecture✓✓CONV✓✓93X0.0810.215.90
8✓✓RES✓✓93X0.0780.216.44
9✓✓ATTN✓✓93X0.0890.237.49
10DSL kernel size μ✓✓RES✓✓33✓0.0810.216.66
11✓✓RES✓✓53✓0.0810.217.04
12✓✓RES✓✓93✓0.0820.226.53
13✓✓RES✓✓113✓0.0830.227.48
14✓✓RES✓✓153✓0.0830.226.36
", + "bbox": [ + 76, + 88, + 895, + 323 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 2. Ablation studies on the CelebA-HQ validation dataset [17], visualizations are in Figure 7. Words put in bold to highlight the changing component.", + "bbox": [ + 75, + 323, + 893, + 354 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/baaff5b0e1bf3eb64c21e14321a4c13220115cfb0846f9f227cf826f024b83db.jpg", + "image_caption": [ + "Figure 7. Reconstruction comparisons for ablation studies on CelebA-HQ. The figure number represents the setting in the corresponding row in Table 2. $\\mathrm{SL}^*$ is SL without Gaussian kernel, $\\mathrm{DSL}^{**}$ is DSL with non pairwise $\\sigma$ . FCM is convolution architecture by default." + ], + "image_footnote": [], + "bbox": [ + 71, + 357, + 478, + 616 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "FA-VAE's reconstruction is more similar to the original image than VQ-GAN in terms of color and shape; more qualitative results are in the supplement.", + "bbox": [ + 75, + 698, + 468, + 744 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Ablation Studies on FCM and DSL In Table 2, we perform ablation studies on different architectures of FCM in combination with different settings of the Spectrum Loss (SL) and Dynamic Spectrum Loss (DSL). Figure 7 gives the accompanying visualizations of the quantitative results.", + "bbox": [ + 75, + 746, + 468, + 823 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "First, as motivated in the Section 3, Table 2 row 3 shows that considering all frequencies as equally important leads the VQ-GAN model to poor performance in terms of rFID and lpips (see Figure 7 image 3). In comparison, results in Table 2 rows 4 demonstrate that FCMs alone can help", + "bbox": [ + 75, + 824, + 470, + 902 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/932a0db2c62244545bb9d924242c6cdc083e564b88fe6a69abd09bc669dbf422.jpg", + "image_caption": [ + "Figure 8. Zero-shot reconstruction on CelebA-HQ [17], FFHQ [21] and photos from the wild, taken from [11]. The models are trained on ImageNet on different compression rates." + ], + "image_footnote": [], + "bbox": [ + 493, + 359, + 901, + 698 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "FA-VAE model achieve a better reconstruction because the residual connections in FCMs preserve better the information flow. The reconstruction is further improved when combined with a simple SL on images level (row 5) or blocks level (row 6). Although SL (row 6) shows slightly inferior reconstruction due to neglection of frequency importance variance across decoder blocks, Figure 7 image 5 shows that the details, such as earrings and hairs, can still be more enhanced than the baseline model shown in image 1.", + "bbox": [ + 496, + 750, + 893, + 898 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "1742", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "the woman has high cheekbones, wavy hair, big nose, mouth open, big lips, pointy nose, and", + "bbox": [ + 89, + 92, + 470, + 99 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "arched eyebrows. she is young and wears heavy makeup, and lipstick.", + "bbox": [ + 137, + 99, + 423, + 109 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/cfe241304298cd6e8d8b818b986c9dc29b5a870d0f74404f1028cf72449795fb.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 83, + 108, + 210, + 205 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/5d53cae184c22bb6f3549e040602db9ec6ee33bf23c152baa7e129404438f150.jpg", + "image_caption": [ + "Figure 9. Text-to-Image generation on CelebA-HQ-MM [17]" + ], + "image_footnote": [], + "bbox": [ + 218, + 109, + 348, + 205 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/f4002f6c85c47a9bfad67f5021422678f8801addf8b46a7ef97ec0227f6de5c7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 354, + 109, + 483, + 205 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "He has goatee and wears necktie.", + "bbox": [ + 619, + 95, + 761, + 104 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/b18c3107873dd5e4fef8daa512f186606422465d46391a7cd90fcafc290df8fc.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 491, + 108, + 620, + 205 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/19510cdb08f01382673ee2673a3e2b70c5c9e73c5d46748decc291b48cf89fa8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 625, + 108, + 754, + 205 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/e56a840f29eeaf83d60f908e87ab6a197b0f13d34f9bd022bc3950abe5f5cfdd.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 763, + 108, + 890, + 205 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Then, Table 2 rows 7-9 compare the performance when the architecture of FCM varies as illustrated in Figure 5; note that instead of using same $\\sigma_{i}$ for $(\\mathcal{A}_i,\\mathcal{C}_i)$ , we use two learnable $\\sigma$ for $\\mathcal{A}_i$ and $\\mathcal{C}_i$ respectively. Quantitatively, when everything else is held equal, the convolution architecture shows better reconstruction than residual and attention architectures. Similarly, Figure 7 image 7 resembles the original image more than images 8 and 9 in terms of face radiance, smoothness, lip color, and shape. The reason is that residual connection enriches the outputs of FCM, which defies the purpose of enriching features of higher frequencies of the decoder features motivated in section 3.1.1. The attention mechanism is at a disadvantage here because the decoder and the encoder exclusively use convolutions.", + "bbox": [ + 75, + 239, + 468, + 450 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Finally, varying the size of kernels in DSL (Table 2 row 10 - 14) show quite similar quantitative reconstruction performance, while qualitatively, Figure 7 shows that a larger kernel size tends to produce smoother reconstructions. One reason could be that a larger kernel size tends to smooth more the feature maps because more surrounding values are used during convolution. Thus, in other experiments, we choose a kernel size of 3, and the effects of kernel sizes are open to future works.", + "bbox": [ + 75, + 453, + 468, + 588 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Zero-shot Reconstruction To further demonstrate the reconstruction capabilities of FA-VAE models, Table 3 gives the zero-shot reconstruction performance evaluated on CelebA-HQ [17] and FFHQ [21] using models trained on ImageNet [9]. The accompanying qualitative results are in Figure 8. Overall, FA-VAE displays impressive transferability capability with more faithful reconstruction in details, for instance, the light contrast in column 1 and book details in column 3. Note also that the common metric used in image compression PSNR is not perfectly correlated with rFID, but we put the metric here for reference.", + "bbox": [ + 75, + 590, + 468, + 756 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.3. Image Synthesis", + "text_level": 1, + "bbox": [ + 76, + 770, + 238, + 786 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 4 shows the text-to-image generation performance on CelebA-HQ-MM [17]. Our proposed autoregressive transformer CAT yields a better generation quality than other GAN-based models, including AttnGAN [52], ControlGAN [26], which are solely designed for image generation. Figure 9 shows that CAT can generate satisfactory images conditioned on text inputs of varying lengths. At-", + "bbox": [ + 75, + 794, + 468, + 900 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/cecdb9180f22dc7c0437c53c13d84134ffdf033419b35e4fde51e1903fcea4c4.jpg", + "table_caption": [], + "table_footnote": [ + "Table 3. Zero-shot reconstruction results on the validation data of FFHQ [21] and CelebA-HQ [17] using models trained on ImageNet [9]. $f$ is the downsampling factor, the codebook sizes for $f = \\{ {16},8,4\\}$ are $\\{ {16384},{16384},{8192}\\}$ respectively." + ], + "table_body": "
CelebAFFHQ
Pretrained ModelfrFID ↓PSNR ↑rFID ↓PSNR ↑
VQ-GAN [40]168.6223.406.8322.68
VQ-GAN [11]44.1131.203.0630.82
FA-VAE (Ours)166.5222.596.1921.95
FA-VAE (Ours)42.2531.391.4630.85
", + "bbox": [ + 506, + 237, + 887, + 338 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/7fbaac8d8421c9e877aee0aabaa65043276a85b3a1756d769d87188fe91626d6.jpg", + "table_caption": [], + "table_footnote": [ + "Table 4. Text-to-image generation on CelebA-HQ MM [50]. tributes such as \"mouth open\" and \"arched eyebrows\" are captured during generation because all the tokens embeddings are used as a condition which gives more precise guidance. More quantitative and qualitative results on different datasets are in the supplement." + ], + "table_body": "
ModelFID ↓
AttnGAN [52]125.98
ControlGAN [26]116.32
DM-GAN [55]131.05
DF-GAN [44]137.60
TediGAN [50]106.37
LAFITE [54]12.54
CAT (Ours)10.23
", + "bbox": [ + 591, + 401, + 802, + 518 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 625, + 619, + 641 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we introduce the Frequency Augmented VAE (FA-VAE) model, which aims to improve reconstruction quality by bridging the gaps in the frequency domains between original and reconstructed images. New modules named Frequency Complement Modules (FCM) are crafted and guided under the new (Dynamic) Spectrum Loss ((D)SL) to learn to complement the reconstructed features of missing frequencies. A new Cross-attention Autoregressive Transformer (CAT) is proposed for achieving more precise textual-image alignment in the text-to-image generation task. FA-VAE shows improved reconstruction on various datasets compared to SOTA methods.", + "bbox": [ + 496, + 650, + 890, + 830 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgement", + "text_level": 1, + "bbox": [ + 500, + 845, + 660, + 862 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This research is supported in part by the National Science Foundation under award No. 2040209.", + "bbox": [ + 500, + 869, + 890, + 898 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "1743", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Jyoti Aneja, Alexander G. Schwing, Jan Kautz, and Arash Vahdat. Ncp-vae: Variational autoencoders with noise contrastive priors. *ArXiv*, abs/2010.02917, 2020. 2", + "[2] Valerio Biscione and Jeffrey S. Bowers. Convolutional neural networks are not invariant to translation, but they can learn to be. JMLR, 2022. 5", + "[3] Andrew Brock, Jeff Donahue, and Karen Simonyan. Large scale GAN training for high fidelity natural image synthesis. In ICLR, 2019. 2", + "[4] Huiwen Chang, Han Zhang, Lu Jiang, Ce Liu, and William T. Freeman. Maskgit: Masked generative image transformer. In CVPR, 2022. 1", + "[5] Ting Chen, Ruixiang Zhang, and Geoffrey Hinton. Analog bits: Generating discrete data using diffusion models with self-conditioning, 2022. 3", + "[6] Rewon Child. Very deep $\\{\\mathrm{vae}\\}$ s generalize autoregressive models and can outperform them on images. In ICLR, 2021. 1, 2", + "[7] Jooyoung Choi, Jungbeom Lee, Chaehun Shin, Sungwon Kim, Hyunwoo Kim, and Sungroh Yoon. Perception prioritized training of diffusion models. In CVPR, 2022. 3", + "[8] Katherine Crowson, Stella Rose Biderman, Daniel Kornis, Dashiell Stander, Eric Hallahan, Louis Castricato, and Edward Raff. Vqgan-clip: Open domain image generation and editing with natural language guidance. In ECCV, 2022. 2, 3, 6", + "[9] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In CVPR, 2009. 6, 8", + "[10] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. In NeurIPS, 2021. 3", + "[11] Patrick Esser, Robin Rombach, and Björn Ommer. Taming transformers for high-resolution image synthesis. In CVPR, 2021. 1, 2, 3, 4, 5, 6, 7, 8", + "[12] Ge Gao, Pei You, Rong Pan, Shunyuan Han, Yuanyuan Zhang, Yuchao Dai, and Hojae Lee. Neural image compression via attentional multi-scale back projection and frequency decomposition. In ICCV, 2021. 2", + "[13] Karol Gregor, Ivo Danihelka, Andriy Mnih, Charles Blundell, and Daan Wierstra. Deep autoregressive networks. In ICML, 2014. 2", + "[14] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. In NeurIPS, 2017. 1", + "[15] Liming Jiang, Bo Dai, Wayne Wu, and Chen Change Loy. Focal frequency loss for image reconstruction and synthesis. In ICCV, 2021. 2, 4, 7", + "[16] Justin Johnson, Alexandre Alahi, and Li Fei-Fei. Perceptual losses for real-time style transfer and super-resolution. In ECCV, 2016. 1, 2", + "[17] Tero Karras, Timo Aila, Samuli Laine, and Jaakko Lehtinen. Progressive growing of gans for improved quality, stability, and variation. CoRR, abs/1710.10196, 2017. 7, 8" + ], + "bbox": [ + 78, + 116, + 467, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[18] Tero Karras, Timo Aila, Samuli Laine, and Jaakko Lehtinen. Progressive growing of GANs for improved quality, stability, and variation. In ICLR, 2018. 2", + "[19] Tero Karras, Miika Aittala, Janne Hellsten, Samuli Laine, Jaakko Lehtinen, and Timo Aila. Training generative adversarial networks with limited data. In NeurIPS, 2020. 2", + "[20] Tero Karras, Miika Aittala, Samuli Laine, Erik Härkönen, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Alias-free generative adversarial networks. In NeurIPS, 2021. 2", + "[21] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In CVPR, 2019. 2, 6, 7, 8", + "[22] Soo Ye Kim, Kfir Aberman, Nori Kanazawa, Rahul Garg, Neal Wadhwa, Huiwen Chang, Nikhil Karnad, Munchurl Kim, and Orly Liba. Zoom-to-inpaint: Image inpainting with high frequency details. In CVPRW, 2022. 1", + "[23] Diederik P Kingma and Max Welling. Auto-encoding variational bayes, 2013. 2", + "[24] Umut Kocasari, Alara Dirik, Mert Tiftikci, and Pinar Yanardag. Stylemc: Multi-channel based fast text-guided image generation and manipulation. 2022 IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 3441-3450, 2022. 2", + "[25] Doyup Lee, Chiheon Kim, Saehoon Kim, Minsu Cho, and Wook-Shin Han. Autoregressive image generation using residual quantization. In CVPR, 2022. 1, 2, 3, 4, 6", + "[26] Bowen Li, Xiaojuan Qi, Thomas Lukasiewicz, and Philip H. S. Torr. Controllable text-to-image generation. In NeurIPS, 2019. 8", + "[27] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. GLIDE: towards photorealistic image generation and editing with text-guided diffusion models. In ICML, 2022. 2, 3", + "[28] Alexander Quinn Nichol and Prafulla Dhariwal. Improved denoising diffusion probabilistic models. In ICML, 2021. 3", + "[29] Gaurav Parmar, Dacheng Li, Kwonjoon Lee, and Zhuowen Tu. Dual contradistinctive generative autoencoder. In CVPR, 2021. 1, 2", + "[30] Yash Patel, Srikar Appalaraju, and R. Manmatha. Deep perceptual compression, 2019. 5", + "[31] Jialun Peng, Dong Liu, Songcen Xu, and Houqiang Li. Generating diverse structure for image inpainting with hierarchical vq-vae. In CVPR, 2021. 1", + "[32] Stanislav Pidhorskyi, Donald A. Adjeroh, and Gianfranco Doretto. Adversarial latent autoencoders. In CVPR, 2020. 2", + "[33] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In ICML, 2021. 2, 3, 6", + "[34] Alec Radford, Jeff Wu, Rewon Child, David Luan, Dario Amodei, and Ilya Sutskever. Language models are unsupervised multitask learners. 2019. 6" + ], + "bbox": [ + 501, + 92, + 890, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "1744", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[35] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J. Liu. Exploring the limits of transfer learning with a unified text-to-text transformer. JMLR, 21(140):1-67, 2020. 3", + "[36] Nasim Rahaman, Aristide Baratin, Devansh Arpit, Felix Dräxler, Min Lin, Fred A. Hamprecht, Yoshua Bengio, and Aaron C. Courville. On the spectral bias of neural networks. In ICML, 2019. 1", + "[37] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. ArXiv, abs/2204.06125, 2022. 3", + "[38] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In ICML, 2021. 3, 6", + "[39] Ali Razavi, Aaron van den Oord, and Oriol Vinyals. Generating diverse high-fidelity images with vq-vae-2. In NeurIPS, 2019. 1, 2, 3, 6", + "[40] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 1, 3, 6, 8", + "[41] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L. Denton, Seyed Kamyar Seyed Ghasemipour, Burcu Karagol Ayan, Seyedeh Sara Mahdavi, Raphael Gontijo Lopes, Tim Salimans, Jonathan Ho, David J. Fleet, and Mohammad Norouzi. Photorealistic text-to-image diffusion models with deep language understanding. ArXiv, abs/2205.11487, 2022. 3", + "[42] Edgar Schonfeld, Bernt Schiele, and Anna Khoreva. A u-net based discriminator for generative adversarial networks. In CVPR, 2020. 2", + "[43] Jianxin Sun, Qiyao Deng, Qi Li, Muyi Sun, Min Ren, and Zhenan Sun. Anyface: Free-style text-to-face synthesis and manipulation. In CVPR, 2022. 2", + "[44] Ming Tao, Hao Tang, Fei Wu, Xiao-Yuan Jing, Bing-Kun Bao, and Changsheng Xu. Df-gan: A simple and effective baseline for text-to-image synthesis. In CVPR, 2022. 2, 8", + "[45] Arash Vahdat and Jan Kautz. NVAE: A deep hierarchical variational autoencoder. In NeurIPS, 2020. 2", + "[46] Aaron van den Oord, Oriol Vinyals, and koray kavukcuoglu. Neural discrete representation learning. In NeurIPS, 2017. 1, 2, 3, 4, 5, 6", + "[47] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In NeurIPS, 2017. 6", + "[48] Haohan Wang, Xindi Wu, Zeyi Huang, and Eric P. Xing. High-frequency component helps explain the generalization of convolutional neural networks. In CVPR, 2020. 2", + "[49] Zihao Wang, Wei Liu, Qian He, Xinglong Wu, and Zili Yi. Clip-gen: Language-free training of a text-to-image generator with clip. arXiv preprint arXiv:2203.00386, 2022. 1, 2, 3, 6", + "[50] Weihao Xia, Yujiu Yang, Jing-Hao Xue, and Baoyuan Wu. Tedigan: Text-guided diverse face image generation and manipulation. In CVPR, 2021. 2, 6, 8" + ], + "bbox": [ + 78, + 90, + 468, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[51] Zhisheng Xiao, Karsten Kreis, Jan Kautz, and Arash Vahdat. Vaebm: A symbiosis between variational autoencoders and energy-based models. In ICLR, 2021. 2", + "[52] Tao Xu, Pengchuan Zhang, Qiuyuan Huang, Han Zhang, Zhe Gan, Xiaolei Huang, and Xiaodong He. Attngan: Fine-grained text to image generation with attentional generative adversarial networks. In CVPR, 2018. 2, 8", + "[53] Jiahui Yu, Xin Li, Jing Yu Koh, Han Zhang, Ruoming Pang, James Qin, Alexander Ku, Yuanzhong Xu, Jason Baldridge, and Yonghui Wu. Vector-quantized image modeling with improved VQGAN. In ICLR, 2022. 5, 7", + "[54] Yufan Zhou, Ruiyi Zhang, Changyou Chen, Chunyuan Li, Chris Tensmeyer, Tong Yu, Jiuming Gu, Jinhui Xu, and Tong Sun. Lafite: Towards language-free training for text-to-image generation. CVPR, 2022. 2, 8", + "[55] Minfeng Zhu, Pingbo Pan, Wei Chen, and Yi Yang. Dm-gan: Dynamic memory generative adversarial networks for text-to-image synthesis. In CVPR, 2019. 2, 8" + ], + "bbox": [ + 501, + 92, + 890, + 345 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "1745", + "bbox": [ + 483, + 945, + 514, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2023/Catch Missing Details_ Image Reconstruction With Frequency Augmented Variational Autoencoder/0a253bb0-c0db-46ba-9d61-8be0d603b311_model.json b/2023/Catch Missing Details_ Image Reconstruction With Frequency Augmented Variational Autoencoder/0a253bb0-c0db-46ba-9d61-8be0d603b311_model.json new file mode 100644 index 0000000000000000000000000000000000000000..d3d296c70171399beea045145e38ba9ed33a3357 --- /dev/null +++ b/2023/Catch Missing Details_ Image Reconstruction With Frequency Augmented Variational Autoencoder/0a253bb0-c0db-46ba-9d61-8be0d603b311_model.json @@ -0,0 +1,2420 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.115, + 0.13, + 0.856, + 0.175 + ], + "angle": 0, + "content": "Catch Missing Details: Image Reconstruction with Frequency Augmented Variational Autoencoder" + }, + { + "type": "text", + "bbox": [ + 0.222, + 0.205, + 0.329, + 0.221 + ], + "angle": 0, + "content": "Xinmiao Lin" + }, + { + "type": "text", + "bbox": [ + 0.14, + 0.223, + 0.41, + 0.24 + ], + "angle": 0, + "content": "Rochester Institute of Technology" + }, + { + "type": "text", + "bbox": [ + 0.212, + 0.243, + 0.34, + 0.256 + ], + "angle": 0, + "content": "x13439@rit.edu" + }, + { + "type": "text", + "bbox": [ + 0.295, + 0.265, + 0.4, + 0.281 + ], + "angle": 0, + "content": "Chiuman Ho" + }, + { + "type": "text", + "bbox": [ + 0.269, + 0.283, + 0.428, + 0.298 + ], + "angle": 0, + "content": "OPPO US Research" + }, + { + "type": "text", + "bbox": [ + 0.276, + 0.304, + 0.421, + 0.317 + ], + "angle": 0, + "content": "chiuman@oppo.com" + }, + { + "type": "text", + "bbox": [ + 0.497, + 0.204, + 0.582, + 0.222 + ], + "angle": 0, + "content": "Yikang Li" + }, + { + "type": "text", + "bbox": [ + 0.46, + 0.223, + 0.62, + 0.238 + ], + "angle": 0, + "content": "OPPO US Research" + }, + { + "type": "text", + "bbox": [ + 0.454, + 0.243, + 0.625, + 0.257 + ], + "angle": 0, + "content": "yikang.li1@oppo.com" + }, + { + "type": "text", + "bbox": [ + 0.695, + 0.205, + 0.804, + 0.221 + ], + "angle": 0, + "content": "Jenhao Hsiao" + }, + { + "type": "text", + "bbox": [ + 0.67, + 0.223, + 0.829, + 0.238 + ], + "angle": 0, + "content": "OPPO US Research" + }, + { + "type": "text", + "bbox": [ + 0.691, + 0.244, + 0.808, + 0.257 + ], + "angle": 0, + "content": "mark@oppo.com" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.352, + 0.314, + 0.368 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.386, + 0.474, + 0.751 + ], + "angle": 0, + "content": "The popular VQ-VAE models reconstruct images through learning a discrete codebook but suffer from a significant issue in the rapid quality degradation of image reconstruction as the compression rate rises. One major reason is that a higher compression rate induces more loss of visual signals on the higher frequency spectrum which reflect the details on pixel space. In this paper, a Frequency Complement Module (FCM) architecture is proposed to capture the missing frequency information for enhancing reconstruction quality. The FCM can be easily incorporated into the VQ-VAE structure, and we refer to the new model as Frequency Augmented VAE (FA-VAE). In addition, a Dynamic Spectrum Loss (DSL) is introduced to guide the FCMs to balance between various frequencies dynamically for optimal reconstruction. FA-VAE is further extended to the text-to-image synthesis task, and a Cross-attention Autoregressive Transformer (CAT) is proposed to obtain more precise semantic attributes in texts. Extensive reconstruction experiments with different compression rates are conducted on several benchmark datasets, and the results demonstrate that the proposed FA-VAE is able to restore more faithfully the details compared to SOTA methods. CAT also shows improved generation quality with better image-text semantic alignment." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.783, + 0.21, + 0.799 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.81, + 0.47, + 0.903 + ], + "angle": 0, + "content": "VQ-VAE models [6,11,25,29,39,46] reconstruct images through learning a discrete codebook of latent embeddings. They gained wide popularity due to the scalable and versatile codebook, which can be broadly applied to many visual tasks such as image synthesis [11,49] and inpainting [4,31]. A higher compression rate is typically preferable in VQ-" + }, + { + "type": "title", + "bbox": [ + 0.492, + 0.265, + 0.704, + 0.301 + ], + "angle": 0, + "content": "Yu Kong \nMichigan State University" + }, + { + "type": "text", + "bbox": [ + 0.532, + 0.304, + 0.659, + 0.317 + ], + "angle": 0, + "content": "yukong@msu.edu" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.353, + 0.891, + 0.504 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.507, + 0.893, + 0.645 + ], + "angle": 0, + "content": "Figure 1. Images and their frequency maps. Row 1: original and reconstructed images. Row 2: the frequency maps of images, frequency increases in any direction away from the center. \\( f \\) is the compression rate. With a greater compression rate, more details are lost during reconstruction, i.e. eyes and mouth shape, and hair texture (pointed with red arrows) which align with the loss of high-frequency features. All frequency figures in this paper use the same colormap. rFID [14] and lpips [16] are lower the better, and frequency values increase from red to green, zoom in for better visualization." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.653, + 0.893, + 0.684 + ], + "angle": 0, + "content": "VAE models since it provides memory efficiency and better learning of coherent semantics structures [11, 40]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.687, + 0.893, + 0.838 + ], + "angle": 0, + "content": "One main challenge quickly arises for a higher compression rate, which severely compromises reconstruction accuracy. Figure 1 row 1 shows that although the reconstructed images at higher compression rates appear consistent with the original image, details inconsistencies such as the color and contour of the lips become apparent upon closer scrutiny. Figure 1 row 2 reveals that similar degradation also manifests on the frequency domains where features towards the middle and higher frequency spectrum are the least recoverable with greater compression rate." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.894, + 0.903 + ], + "angle": 0, + "content": "Several causes stand behind this gap between pixel and frequency space. The convolutional nature of autoencoders is prone to spectral bias, which favors learning low-frequency features [22, 36]. This challenge is further ag-" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1736" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.243 + ], + "angle": 0, + "content": "gravated when current methods exclusively design losses or improve model architecture for better semantics resemblance [11, 16, 25] but often neglect the alignment on the frequency domain [12, 15]. On top of that, it is intuitively more challenging for a decoder to reconstruct an image patch from a single codebook embedding (high compression) than multiple embeddings (less compression). The reason is that the former mixes up features of incomplete and diverse frequencies, while the latter could preserve more fine-grained and complete features at various frequencies." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.245, + 0.473, + 0.366 + ], + "angle": 0, + "content": "Inspired by these insights, the Frequency Augmented VAE (FA-VAE) model is proposed, which aims to improve reconstruction quality by achieving better alignment on the frequency spectrums between the original and reconstructed images. More specifically, new modules named Frequency Complement Modules (FCM) are crafted and embedded at multiple layers of FA-VAE's decoder to learn to complement the decoder's features with missing frequencies." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.368, + 0.473, + 0.519 + ], + "angle": 0, + "content": "We observe that valuable middle and high frequencies are mingled with the encoder's feature maps during the compression via an encoder, shown in Figure 3 row 4. Therefore, a new loss termed Spectrum Loss (SL) is proposed to guide FCMs to generate missing features that align with the same level's encoder features on the frequency domain. Since most image semantics reside on the low-frequency spectrum [48], SL prioritizes learning lower-frequency features with diminishing weights as frequencies increase." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.521, + 0.473, + 0.718 + ], + "angle": 0, + "content": "Interestingly, we discover that checkerboard patterns appear in the complemented decoder's features with SL, although better reconstruction performance is achieved (Figure 3 column 4). We speculate that because SL sets a deterministic range for the low-frequency spectrum when applying weights on the frequencies without considering that the importance of a frequency can vary from layer to layer. Thus, an improved loss function Dynamic Spectrum Loss (DSL) is crafted on top of SL with a learnable component to adjust the range of low-frequency spectrum dynamically for optimal reconstruction. DSL can improve reconstruction quality even further than SL without the unnatural checkerboard artifacts in the features (Figure 3 column 5)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.72, + 0.473, + 0.902 + ], + "angle": 0, + "content": "We further extend FA-VAE to the text-to-image generation task and propose the Cross-attention Autoregressive Transformer (CAT) model. We first observe that only using one or a few token embeddings is a coarse representation of lengthy texts [8, 27, 33]. Thus CAT uses all token embeddings as a condition for more precise guidance. Moreover, existing works typically use self-attention, and the text condition is embedded merely at the beginning of the generation [11, 49]. This mechanism becomes problematic in the autoregressive generation because one image token is generated at a time, thus the text condition gradually loosens its connection with the generated tokens. To circumvent this" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.892, + 0.122 + ], + "angle": 0, + "content": "issue, CAT embeds a cross-attention mechanism that allows the text condition to guide each step generation." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.122, + 0.892, + 0.152 + ], + "angle": 0, + "content": "To summarize, our work includes the following contributions:" + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.152, + 0.892, + 0.211 + ], + "angle": 0, + "content": "- We propose a new type of architecture called Frequency Augmented VAE (FA-VAE) for improving image reconstruction through achieving more accurate details reconstruction." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.213, + 0.892, + 0.301 + ], + "angle": 0, + "content": "- We propose a new loss called Spectrum Loss (SL) and its enhanced version Dynamic Spectrum Loss (D SL), which guides the Frequency Complement Modules (FCM) in FA-VAE to adaptively learn different low/high frequency mixtures for optimal reconstruction." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.303, + 0.892, + 0.378 + ], + "angle": 0, + "content": "- We propose a new Cross-attention Autoregressive Transformer (CAT) for text-to-image generation using more fine-grained textual embeddings as a condition with a cross-attention mechanism for better image-text semantic alignment." + }, + { + "type": "list", + "bbox": [ + 0.518, + 0.152, + 0.892, + 0.378 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.392, + 0.771, + 0.41 + ], + "angle": 0, + "content": "2. Related Work & Background" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.418, + 0.892, + 0.674 + ], + "angle": 0, + "content": "Image Reconstruction Vector Quantized-Variational AutoEncoder (VQ-VAE) [46] extends the Variational AutoEncoder structure [23] and proposes to encode images into discrete latent codes using vector quantization (VQ). Then a generative model, such as an autoregressive transformer [13], can be trained and paired with the decoder in VQ-VAE to synthesize new images. Later works [11,25,39] further improve the reconstruction and generation quality by improving the generative model architecture or the quantization efficiency. Since VQ-VAE models operate on discrete latent spaces, they cannot be directly and fairly compared to other VAE-based models that employ continuous latent space [1,6,23,29,45,51]. In contrast to current VQ-VAE models, our proposed FA-VAE technique improves reconstruction through the frequency angle and is more generalized since the FCMs can be readily extended to other neural networks that share the same VQ-VAE structure." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.675, + 0.892, + 0.796 + ], + "angle": 0, + "content": "Few other works do image reconstruction via the frequency perspective. FFL [15] proposed a loss function to penalize differences on the hard frequencies. A new module, which is designed in the continuous latent space for image compression, is proposed to be incorporated into the encoder and decoder in [12]. To our best knowledge, FA-VAE is the first work that aims to improve image reconstruction on discrete latent space through the frequency perspective." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.796, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Image Generation Image generation can be achieved via GAN-based models [3, 18, 42, 44, 52, 55], which synthesize images from noise vectors unconditionally or conditioned on different inputs such as texts, masks, etc. It becomes cumbersome to train one model for one application. Thus, StyleGAN [21] encodes semantic attributes into a continuous latent space, and subsequent works [19, 20," + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1737" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.078, + 0.089, + 0.427, + 0.27 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.427, + 0.089, + 0.891, + 0.27 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.075, + 0.276, + 0.893, + 0.318 + ], + "angle": 0, + "content": "Figure 2. Frequency Augmented VAE (FA-VAE). (a) The encoder \\(\\mathcal{E}\\) encodes the images \\(X\\) onto discrete latent codebook space \\(\\hat{z}_q\\) which is used by the decoder \\(\\mathcal{G}\\) to reconstruct images \\(\\hat{X}\\). DSL (b) guides the FCMs (c) to learn to complement the reconstructed features with missing features of important frequencies in order to improve reconstruction quality. A more detailed figure is in the supplement." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.322, + 0.473, + 0.564 + ], + "angle": 0, + "content": "24, 32, 43, 50, 54] leverage this space for generating images conditioned on attributes or textual descriptions. However, StyleGAN-based models cannot scale to large datasets when the number of attributes becomes substantially large because this demands to increase model's size. In contrast, the codebook in VQ-VAE models [46] is scalable to large datasets without additional model complexity. Diffusion-based models [5, 7, 10, 27, 28, 37, 40, 41] generate images from Gaussian noise through a reverse diffusion process can often require substantial training and large datasets. Most text-to-image generation models commonly use one of a few embeddings for a condition [8, 27, 38, 49] from pretrained models such as CLIP [33] or T5-XXL [35]. In contrast, our proposed autoregressive transformer CAT uses the embeddings from all the text tokens for more fine-grained guidance during image generation." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.565, + 0.471, + 0.685 + ], + "angle": 0, + "content": "VQ-VAE We now describe the VQ-VAE model [46] as it is the backbone used in the FA-VAE model, which is presented in Figure 2. VQ-VAE [46] and related models [11, 25, 38, 39] consist of an encoder that encodes the images into a codebook of embeddings. Then, a decoder is trained to reconstruct the images from a discrete set of codebook embeddings. In this paper, reconstruction loss in VQ-GAN is utilized which is:" + }, + { + "type": "equation", + "bbox": [ + 0.149, + 0.695, + 0.47, + 0.715 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {r e c} = \\left\\| X - \\hat {X} \\right\\| _ {1} + \\mathcal {L} _ {p i p s} (X - \\hat {X}) \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.726, + 0.47, + 0.787 + ], + "angle": 0, + "content": "where \\( X \\) and \\( \\hat{X} \\) are the original and reconstructed images. We also use the adversarial setting of VQ-GAN model which introduced a discriminator compared to VQ-VAE models [39, 46], more details could be referred to [11]." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.801, + 0.212, + 0.818 + ], + "angle": 0, + "content": "3. Methodology" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.825, + 0.471, + 0.902 + ], + "angle": 0, + "content": "The proposed Frequency Augmented VAE (FA-VAE) is presented in Figure 2. In this section, we describe how FA-VAE ameliorates reconstruction quality by bridging the spectrum domain gap between reconstructed and original images. By explicitly embedding Frequency Complement" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.322, + 0.894, + 0.398 + ], + "angle": 0, + "content": "Modules (FCM) into the decoder, Dynamic Spectrum Loss (DSL) leverages important features frequency-wise from the encoder and guides the FCMs to complement the features of the decoder for better frequency restoration at different reconstruction stages." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.411, + 0.831, + 0.427 + ], + "angle": 0, + "content": "3.1. Frequency Augmented VAE (FA-VAE)" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.434, + 0.895, + 0.587 + ], + "angle": 0, + "content": "Let the images be \\(X \\in \\mathbb{R}^{H \\times W \\times 3}\\). The codebook \\(\\mathfrak{C}\\) is a set of \\(|\\mathfrak{C}|\\) embeddings, such as \\(\\mathfrak{C} = \\{\\mathfrak{c}_i | i = 1, \\dots, |\\mathfrak{c}|\\} \\in \\mathbb{R}^{n_z}\\) and \\(n_z\\) is the length of one codebook embedding. FA-VAE consists of an encoder \\(\\mathcal{E}\\) that encodes the images \\(X\\) into latent space representations such as \\(z = \\mathcal{E}(x)\\), for \\(z \\in \\mathbb{R}^{h \\times w \\times n_z}\\) and \\(x \\in X\\), \\((h \\times w)\\) is the resolution of the encoded representation. Let \\(f = H / h = W / w\\) be the downsampling factor or compression rate. Each feature of \\(z\\) is approximated by the vector quantization block \\(\\mathcal{Q}\\) using the nearest codebook entries, and it can be presented as:" + }, + { + "type": "equation", + "bbox": [ + 0.586, + 0.6, + 0.892, + 0.626 + ], + "angle": 0, + "content": "\\[\n\\mathcal {Q} (z) = \\hat {z} _ {q} = \\underset {c _ {k} \\in \\mathcal {C}} {\\arg \\min } \\| z _ {i j} - \\mathfrak {c} _ {k} \\| \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.639, + 0.894, + 0.686 + ], + "angle": 0, + "content": "where \\(\\hat{z}_q\\) is the quantized latent embedding and subsequently used by the decoder \\(\\mathcal{G}\\) to produce the reconstructed image \\(\\hat{x} = \\mathcal{G}(\\hat{z}_q)\\)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.709, + 0.838, + 0.724 + ], + "angle": 0, + "content": "3.1.1 Frequency Complement Modules (FCM)" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.735, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Motivation Figure 1 shows that a higher compression rate leads to more significant reconstruction disparities on the higher-frequency spectrum. Existing models [11, 25, 39] guide the reconstructed images to be more aligned on the pixel and feature spaces with the original images (Eq. 1) but neglect frequency spectrum alignment and leave the encoder and decoder without further guidance. Figure 3 column 2 shows that the encoder activations \\(\\mathcal{A}_1\\) of baseline VQ-GAN [11] contain rich high-frequency features (row 3 & 4), but the decoder's activations could mainly restore low-frequency features (row 5 & 6)." + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1738" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.473, + 0.184 + ], + "angle": 0, + "content": "Therefore we propose Frequency Complement Modules (FCM), illustrated in Figure 2 (c), which aims to complement the decoder's feature maps \\(\\mathcal{B}_i\\) with features of missing frequencies using the encoder activations \\(\\mathcal{A}_i\\). The FCMs \\(\\mathcal{F}_i\\) consist of sequences of convolution layers and activations. The decoder \\(\\mathcal{G}\\) with FCMs embedded can be represented as:" + }, + { + "type": "equation", + "bbox": [ + 0.105, + 0.208, + 0.422, + 0.225 + ], + "angle": 0, + "content": "\\[\n\\mathcal {B} _ {M - 1} = \\mathcal {G} _ {M} \\left(\\mathcal {C} _ {M} + \\hat {z} _ {q}\\right) = \\mathcal {G} _ {M} \\left(\\mathcal {F} _ {M} (\\hat {z} _ {q}) + \\hat {z} _ {q}\\right)\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.132, + 0.228, + 0.469, + 0.242 + ], + "angle": 0, + "content": "\\[\n\\dots \\tag {3}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.131, + 0.246, + 0.397, + 0.262 + ], + "angle": 0, + "content": "\\[\n\\mathcal {B} _ {1} = \\mathcal {G} _ {2} \\left(\\mathcal {C} _ {2} + \\mathcal {B} _ {2}\\right) = \\mathcal {G} _ {2} \\left(\\mathcal {F} _ {2} \\left(\\mathcal {B} _ {2}\\right) + \\mathcal {B} _ {2}\\right)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.274, + 0.38, + 0.289 + ], + "angle": 0, + "content": "Similarly, the encoder \\(\\mathcal{E}\\) can be abstracted to:" + }, + { + "type": "equation", + "bbox": [ + 0.158, + 0.303, + 0.47, + 0.32 + ], + "angle": 0, + "content": "\\[\nz = \\mathcal {E} (x) = \\mathcal {E} _ {M} \\left(\\dots \\left(\\mathcal {E} _ {2} (\\mathcal {E} _ {1} (x)))\\right)\\right) \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.332, + 0.469, + 0.452 + ], + "angle": 0, + "content": "where \\(\\mathcal{G}_i\\) is the \\(i\\)-th layer of the decoder, and \\(\\mathcal{A}_i = \\mathcal{E}_i(x)\\) is the outputs or feature maps of the corresponding encoder block. The outputs of previous blocks \\(\\mathcal{B}_i\\) are complemented by outputs \\(\\mathcal{C}_i = \\mathcal{F}_i(\\mathcal{B}_{i + 1})\\), which contain the frequency-rich features learned from the encoder activations \\(\\mathcal{A}_i\\). The following section describes how Dynamic Frequency Loss (DSL) guides the learning of FCMs to specifically learn to complement \\(\\mathcal{B}_i\\) with the features of missing frequencies." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.454, + 0.47, + 0.588 + ], + "angle": 0, + "content": "In the paper, the frequency feature compensation is implemented by addition (in Eq. 3). However, the FCM is flexible in adopting any architecture and merging techniques, as illustrated by the three examples in Figure 5. Note that no architecture limitation is imposed on the encoder and decoder blocks as long as \\(\\mathcal{A}_i\\) and \\(\\mathcal{B}_i\\) share the same resolution. VQ-GAN [11] and related models [25,46] already have similar architecture, and VQ-GAN is chosen as the backbone for FA-VAE." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.611, + 0.349, + 0.627 + ], + "angle": 0, + "content": "3.1.2 Dynamic Spectrum Loss (DSL)" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.636, + 0.47, + 0.802 + ], + "angle": 0, + "content": "Motivation Spectrum Loss (SL) is proposed to guide explicitly the outputs of FCMs \\(\\mathcal{C}_i\\) to be more aligned with the encoder's activations \\(\\mathcal{A}_i\\) on the frequency spectrum because the latter contains rich features on the higher frequency spectrum (Figure 3). Moreover, to account for the varying importance of frequencies across different decoder stages, Dynamic Spectrum Loss (DSL) is proposed, a more generalized variant of SL, and has the ability to adjust weights put on higher-frequency spectrum adaptively. Thus, each decoder block's outputs can be enriched with features of the most critical frequencies for an accurate reconstruction." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.803, + 0.471, + 0.849 + ], + "angle": 0, + "content": "Background The outputs \\(\\mathcal{A}_i\\) of encoder block \\(\\mathcal{E}_i\\) and \\(\\mathcal{C}_i\\) of FCM block \\(\\mathcal{F}_i\\) are first transformed to the frequency domain using Discrete Fourier Transform (DFT) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.126, + 0.861, + 0.47, + 0.904 + ], + "angle": 0, + "content": "\\[\nF (u, v) = \\sum_ {x = 0} ^ {M - 1} \\sum_ {y = 0} ^ {N - 1} f (x, y) \\cdot e ^ {- i 2 \\pi \\left(\\frac {u x}{M} + \\frac {v y}{N}\\right)} \\tag {5}\n\\]" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.091, + 0.883, + 0.426 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.428, + 0.892, + 0.526 + ], + "angle": 0, + "content": "Figure 3. Image, activations, and their frequency maps. Ours* is FA-VAE model with FFL [15], Ours‡ is FA-VAE with SL and Ours is FA-VAE model with DSL. DSL shows a more harmonious balance between low- and high-frequencies (row 5) and more accurate reconstruction in mouth and hair textures. The frequency maps are plotted using the average of all channels and the contrast is adjusted to emphasize the higher frequency spectrums." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.531, + 0.892, + 0.637 + ], + "angle": 0, + "content": "where \\( e \\) and \\( i \\) are Euler's number and the imaginary unit. \\( M \\times N \\) is the spatial resolution of feature maps and the Fourier Transform (Eq. 5) is applied to each of them. \\( f(x,y) \\) is the value at \\( (x,y) \\) of each feature map. \\( F(u,v) \\) is the corresponding value at \\( (u,v) \\) coordinates on the frequency spectrum. The Focal Frequency Loss (FFL) [15] can be presented as:" + }, + { + "type": "equation", + "bbox": [ + 0.507, + 0.649, + 0.892, + 0.704 + ], + "angle": 0, + "content": "\\[\n\\operatorname {F F L} \\left(\\mathcal {A} _ {i}, \\mathcal {C} _ {i}\\right) = \\frac {1}{M N \\left| \\mathcal {C} _ {i} \\right|} \\sum_ {c = 0} ^ {\\left| \\mathcal {C} _ {i} \\right| - 1} \\sum_ {u = 0} ^ {M - 1} \\sum_ {v = 0} ^ {N - 1} w (u, v) J (u, v), \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.705, + 0.892, + 0.795 + ], + "angle": 0, + "content": "where \\( w(u,v) = |F_{\\mathcal{A}_i}(u,v) - F_{\\mathcal{C}_i}(u,v)| \\) are the weights put on each frequency. \\( J(u,v) = |F_{\\mathcal{A}_i}(u,v) - F_{\\mathcal{C}_i}(u,v)|^2 \\) is the main error function based on the frequency difference. The \\( |\\mathcal{C}_i| \\) is the number of feature maps in \\( \\mathcal{A}_i \\) and \\( \\mathcal{C}_i \\). Both real and imaginary parts of the frequency domain are considered, more details are in [15]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.796, + 0.892, + 0.902 + ], + "angle": 0, + "content": "Limitations of FFL Eq. 6 demonstrates that FFL puts modulating term \\( w(u, v) \\) to focus learning on the hardest frequencies for reconstruction, which are the higher frequencies following our observation (Figures 1 & 3). However, this could not be ideal because features on the lower frequencies define the image content, and overemphasizing the higher frequencies could over-constrain the learning" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1739" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.078, + 0.088, + 0.468, + 0.2 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.202, + 0.457, + 0.216 + ], + "angle": 0, + "content": "Figure 4. Cross-attention Autoregressive Transformer (CAT)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.22, + 0.47, + 0.326 + ], + "angle": 0, + "content": "and lead to sub-optimal reconstruction, see Figure 7 row 1 column 3 and Table 2 row 3. Moreover, Figure 3 column 3 shows FCMs guided by a simple FFL can improve reconstruction performance (Ours\\*). However the frequency maps of the decoder activations \\(\\mathcal{C}_1\\) contain excessive noise due to overemphasis on the higher frequency spectrum, and lower frequencies are neglected (row 6)." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.327, + 0.471, + 0.434 + ], + "angle": 0, + "content": "Spectrum Loss (SL) Thus, we propose to apply a low-pass filter on the weights \\( w(u,v) \\) in Eq. 6 to penalize more mismatch in the lower-frequency domain and gradually diminish the penalizing weights towards higher frequency spectrum. Therefore, let the Gaussian kernels be \\( K_{i}(\\mu ,\\sigma) \\) with weights initialized using mean and standard deviation \\( (\\mu ,\\sigma_{i}) \\), and applied over the feature maps as:" + }, + { + "type": "equation", + "bbox": [ + 0.127, + 0.441, + 0.47, + 0.462 + ], + "angle": 0, + "content": "\\[\n(\\hat {\\mathcal {A}} _ {i}, \\hat {\\mathcal {C}} _ {i}) = (K _ {i} (\\mu , \\sigma_ {i}) \\star \\mathcal {A} _ {i}, K _ {i} (\\mu , \\sigma_ {i}) \\star \\mathcal {C} _ {i}). \\qquad (7)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.47, + 0.47, + 0.501 + ], + "angle": 0, + "content": "where the \\(\\star\\) is the convolution operation. Then, the Spectrum Loss (SL) is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.181, + 0.51, + 0.47, + 0.53 + ], + "angle": 0, + "content": "\\[\n\\operatorname {S L} \\left(\\mathcal {A} _ {i}, \\mathcal {C} _ {i}\\right) = \\operatorname {F F L} \\left(\\hat {\\mathcal {A}} _ {i}, \\hat {\\mathcal {C}} _ {i}\\right). \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.538, + 0.471, + 0.825 + ], + "angle": 0, + "content": "Limitations of SL Up until now, a fixed Gaussian filter \\( K_{i}(\\mu ,\\sigma_{i}) \\) is applied over all \\( \\mathcal{A}_i \\) and \\( \\mathcal{C}_i \\) of all encoder and decoder blocks. Figure 3 column 4 demonstrates that SL \\( (\\mathrm{Ours}^{\\ddagger}) \\) improves reconstruction on the lower frequency spectrum, leading to better reconstruction than the baseline model (Table 2 row 5 vs row 1). However, checkerboard artifacts are also present (row 5) on \\( \\mathcal{C}_1 \\). One reason is that deterministic variance \\( \\sigma_{i} \\) assumes that decoder activations across different levels require the same amount of higher frequency features for accurate reconstruction, which could be an over-rigid constraint on the learning. Another reason could be that the same \\( \\sigma_{i} \\) magnifies the checkerboard effects produced by upsampling or striding operations in CNNs which are normally circumvented by subsequent convolutional layers [2, 30]. In the observation of the experiments, we find that the checkerboard effects can be intensified with the same \\( \\sigma_{i} \\) in different layers due to the bounding effect between low frequency and high frequency (shown in Figure 3 column 4 row 6)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.826, + 0.473, + 0.903 + ], + "angle": 0, + "content": "Dynamic Spectrum Loss (DSL) Therefore, we propose to optimize the variances \\(\\sigma_{i}\\) instead of setting them as static hyperparameters to suit the different amounts of frequencies needed for each block's \\(\\mathcal{B}_i\\). The new Dynamic Spectrum Loss (DSL) is a more generalized form of Spectrum Loss" + }, + { + "type": "image", + "bbox": [ + 0.515, + 0.093, + 0.619, + 0.255 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.541, + 0.257, + 0.59, + 0.269 + ], + "angle": 0, + "content": "(a) FCM" + }, + { + "type": "image", + "bbox": [ + 0.631, + 0.09, + 0.74, + 0.256 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.631, + 0.257, + 0.741, + 0.269 + ], + "angle": 0, + "content": "(b) FCM w/ residual" + }, + { + "type": "image", + "bbox": [ + 0.761, + 0.089, + 0.861, + 0.256 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.754, + 0.257, + 0.868, + 0.269 + ], + "angle": 0, + "content": "(c) FCM w/ attention" + }, + { + "type": "image_caption", + "bbox": [ + 0.516, + 0.273, + 0.875, + 0.287 + ], + "angle": 0, + "content": "Figure 5. FCM with different architectures and connections." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.29, + 0.893, + 0.335 + ], + "angle": 0, + "content": "(SL) with learnable \\(\\sigma_{i}\\). Note that DSL also includes FFL as a special form when used on the original and reconstructed images without the Gaussian filters." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.335, + 0.892, + 0.364 + ], + "angle": 0, + "content": "Then the total reconstruction loss for FA-VAE can be described as:" + }, + { + "type": "equation", + "bbox": [ + 0.549, + 0.368, + 0.892, + 0.432 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} _ {r e c} = \\alpha \\operatorname {F F L} (X, \\hat {X}) + \\beta \\sum_ {i = 0} ^ {M - 1} \\operatorname {D S L} \\left(\\mathcal {A} _ {i}, \\mathcal {C} _ {i}\\right) \\tag {9} \\\\ + \\left\\| X - \\hat {X} \\right\\| _ {1} + \\mathcal {L} _ {p i p s} (X - \\hat {X}). \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.439, + 0.893, + 0.562 + ], + "angle": 0, + "content": "where \\(\\mathrm{FFL}(X - \\hat{X})\\) is the Focal Frequency Loss applied on the original reconstructed images. The second term in Eq. 9 is the DSL loss applied over the outputs of the encoder and FCM blocks. The first two losses aim to minimize the frequency spectrum differences on the images and internal feature maps. The third and fourth losses act on the pixel and feature maps space [11]. \\(\\alpha\\) and \\(\\beta\\) are hyperparameters. \\(\\sigma_{i}\\) are model parameters and optimized as:" + }, + { + "type": "equation", + "bbox": [ + 0.567, + 0.567, + 0.892, + 0.593 + ], + "angle": 0, + "content": "\\[\n\\sigma_ {i} ^ {*}, \\mathcal {E} ^ {*}, \\mathcal {G} ^ {*}, \\mathcal {C} ^ {*} = \\underset {\\sigma_ {i}, \\mathcal {E}, \\mathcal {G}, \\mathcal {C}} {\\arg \\min } \\left(\\mathcal {L} _ {\\text {r e c}} + \\mathcal {L} _ {Q}\\right) \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.599, + 0.892, + 0.704 + ], + "angle": 0, + "content": "\\(\\mathcal{L}_Q\\) is the quantization loss which minimizes the difference between the codebook embeddings and the embeddings given by the encoder \\(\\mathcal{E}\\), more details in [11]. We also use the \\(L_{2}\\) regularization on the codebook embeddings during quantization as in [53] and exponential moving average (EMA) for updating the codebook [46] as they provide more stable training for the quantization block \\(\\mathcal{Q}\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.705, + 0.893, + 0.826 + ], + "angle": 0, + "content": "Benefits of DSL The advantages of learnable \\(\\sigma_{i}\\) in DSL are shown in Figure 3 column 5 where the reconstructed activations show no checkerboard artifacts as in column 4. Moreover, the reconstructed activations \\(C_1\\) are more similar to the encoder activations \\(A_{1}\\) on the frequency spectrum. Compared to the baseline model (column 2), our model exhibits a more harmonious balance between low- and highfrequencies, leading to more accurate reconstructed images." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.833, + 0.892, + 0.863 + ], + "angle": 0, + "content": "3.2. Cross-attention Autoregressive Transformer (CAT)" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.871, + 0.892, + 0.901 + ], + "angle": 0, + "content": "We further extend FA-VAE to the text-to-image generation task and introduce a new Cross-attention Autoregres" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1740" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.126, + 0.09, + 0.852, + 0.295 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.289, + 0.298, + 0.68, + 0.312 + ], + "angle": 0, + "content": "Figure 6. Reconstruction on ImageNet [9], label: Japanese terrier." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.32, + 0.473, + 0.457 + ], + "angle": 0, + "content": "sive Transformer (CAT), presented in Figure 4. CAT uses all the token embeddings of a textual description given by the pretrained text encoder of CLIP model [33], while existing works mostly use one or partial textual embeddings [8, 38, 49]. Furthermore, CAT uses a cross-attention mechanism with the text tokens to guide the image generation at each step. This more fine-grained text condition allows the generation to capture more precisely the relationships of semantic attributes between text and image." + }, + { + "type": "equation", + "bbox": [ + 0.188, + 0.47, + 0.469, + 0.502 + ], + "angle": 0, + "content": "\\[\np (s \\mid c) = \\prod_ {i} p \\left(s _ {i} \\mid s _ {< i}, c\\right) \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.509, + 0.47, + 0.556 + ], + "angle": 0, + "content": "Then the predicted indices \\( z' \\) can be decoded to an image using FA-VAE's decoder \\( G \\). The loss is to maximize the log-likelihood of the data representations," + }, + { + "type": "equation", + "bbox": [ + 0.18, + 0.567, + 0.469, + 0.585 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {C A T}} = \\mathbb {E} _ {x \\sim p (x)} [ - \\log p (s) ] \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.596, + 0.47, + 0.657 + ], + "angle": 0, + "content": "In this paper, the GPT2 model [34], which is the default setting in [11, 49], is utilized as the backbone for CAT, and cross-attention mechanism inspired from [47] is applied with the GPT2 structure." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.671, + 0.21, + 0.688 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.696, + 0.277, + 0.712 + ], + "angle": 0, + "content": "4.1. Experimental Details" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.72, + 0.471, + 0.902 + ], + "angle": 0, + "content": "The datasets used in this paper are: (1) MultiModal CelebA-HQ dataset [50] with 30,000 high-resolution celebrity face images and each image comes with ten captions; (2) Flickr-Faces-HQ (FFHQ) dataset [21] of 70,000 high-resolution face images; (3) ImageNet [9] that contains around 1 million images from 1000 categories. Experiments are performed on V-100 GPUs and results are reported on the validation sets unless specified otherwise. For a fair comparison with existing models, the resolutions used for training are \\(256 \\times 256\\) on all datasets. The details of training settings and hyperparameters are provided in the supplementary materials." + }, + { + "type": "table", + "bbox": [ + 0.508, + 0.318, + 0.887, + 0.515 + ], + "angle": 0, + "content": "
ModelDatasetCodebook Size(h × w)rFID ↓
RQ-VAE [25]FFHQ2048(8 × 8)5.33
FA-VAE (Ours)FFHQ2048(16 × 16)4.98
VQ-VAE-2 [39]ImageNet512(64 × 64) & (32 × 32)~ 10 (train)
VQ-GAN [40]ImageNet8192(64 × 64)1.06
FA-VAE (Ours)ImageNet8192(64 × 64)0.40
DALL-E [38]ImageNet8192(32 × 32)32.01
VQ-GAN [11]ImageNet16384(16 × 16)5.15
VQ-GAN [11]ImageNet1024(16 × 16)7.94
VQ-GAN [25]ImageNet16384(8 × 8)17.95
RQ-VAE† [46]ImageNet16384(8 × 8)10.77
RQ-VAE* [25]ImageNet16384(8 × 8)4.73
FA-VAE (Ours)ImageNet16384(16 × 16)4.60
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.517, + 0.892, + 0.56 + ], + "angle": 0, + "content": "Table 1. Reconstruction results on the validation data of FFHQ [21] and ImageNet [9] respectively, more results in the supplement. \\(\\dagger\\) means depth 2, and \\(*\\) means depth 4 from RQ-VAE [25]." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.572, + 0.707, + 0.587 + ], + "angle": 0, + "content": "4.2. Image Reconstruction" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.597, + 0.892, + 0.748 + ], + "angle": 0, + "content": "Reconstruction on FFHQ and Imagenet The experiment results of reconstruction on FFHQ [21] and ImageNet [9] are presented in Table 1. The results of the baselines are from the original paper. Note that smaller \\( h \\) and \\( w \\) means that the downsampling factor or the compression rate \\( f \\) is larger. FA-VAE model shows improved reconstruction quality over baseline models across different compression rates in both datasets. The main reason is that FA-VAE can successfully reconstruct the important middle and high-frequencies which are neglected in baseline models." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.893, + 0.902 + ], + "angle": 0, + "content": "Figure 6 further supports our previous claim and demonstrates clear diverging reconstruction qualities between baseline models VQ-GAN [11], DALL-E [38], and our FA-VAE model. With a larger compression rate, Figure 6 row 2 shows that more middle and high frequencies are compressed in VQ-GAN and DALL-E. In comparison, FA-VAE can reconstruct middle and high frequencies more accurately, which translates to good representations with improved semantics and local details in Figure 6 row 1. For instance, at a compression rate of 16, the dog's tongue of" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "1741" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.078, + 0.089, + 0.896, + 0.324 + ], + "angle": 0, + "content": "
ablation onFFL(X, X̂)SL(Ai, Ci)FCM Fkernel K_i(μ, σ_i)DSL(Ai, Ci)kernel size μinitial value σ_ipair-wise σ_iL1 ↓Lpips ↓rFID ↓
1VQ-GAN [11]XXXXXXXX0.1210.3010.12
2VQ-GAN + Style [53]XXXXXXXX0.0850.2311.90
3VQ-GAN [15]✓XXXXXXX0.1140.3530.65
4SL w/o kernel✓XCONVXXXXX0.0820.227.04
5✓✓CONVXXXXX0.0820.227.02
6SL w/ kernel✓✓CONV✓X93X0.0800.227.39
7FCM architecture✓✓CONV✓✓93X0.0810.215.90
8✓✓RES✓✓93X0.0780.216.44
9✓✓ATTN✓✓93X0.0890.237.49
10DSL kernel size μ✓✓RES✓✓33✓0.0810.216.66
11✓✓RES✓✓53✓0.0810.217.04
12✓✓RES✓✓93✓0.0820.226.53
13✓✓RES✓✓113✓0.0830.227.48
14✓✓RES✓✓153✓0.0830.226.36
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.324, + 0.895, + 0.355 + ], + "angle": 0, + "content": "Table 2. Ablation studies on the CelebA-HQ validation dataset [17], visualizations are in Figure 7. Words put in bold to highlight the changing component." + }, + { + "type": "image", + "bbox": [ + 0.073, + 0.358, + 0.479, + 0.617 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.618, + 0.472, + 0.687 + ], + "angle": 0, + "content": "Figure 7. Reconstruction comparisons for ablation studies on CelebA-HQ. The figure number represents the setting in the corresponding row in Table 2. \\(\\mathrm{SL}^*\\) is SL without Gaussian kernel, \\(\\mathrm{DSL}^{**}\\) is DSL with non pairwise \\(\\sigma\\). FCM is convolution architecture by default." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.699, + 0.47, + 0.745 + ], + "angle": 0, + "content": "FA-VAE's reconstruction is more similar to the original image than VQ-GAN in terms of color and shape; more qualitative results are in the supplement." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.747, + 0.47, + 0.824 + ], + "angle": 0, + "content": "Ablation Studies on FCM and DSL In Table 2, we perform ablation studies on different architectures of FCM in combination with different settings of the Spectrum Loss (SL) and Dynamic Spectrum Loss (DSL). Figure 7 gives the accompanying visualizations of the quantitative results." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.471, + 0.903 + ], + "angle": 0, + "content": "First, as motivated in the Section 3, Table 2 row 3 shows that considering all frequencies as equally important leads the VQ-GAN model to poor performance in terms of rFID and lpips (see Figure 7 image 3). In comparison, results in Table 2 rows 4 demonstrate that FCMs alone can help" + }, + { + "type": "image", + "bbox": [ + 0.494, + 0.361, + 0.902, + 0.699 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.703, + 0.895, + 0.746 + ], + "angle": 0, + "content": "Figure 8. Zero-shot reconstruction on CelebA-HQ [17], FFHQ [21] and photos from the wild, taken from [11]. The models are trained on ImageNet on different compression rates." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.895, + 0.9 + ], + "angle": 0, + "content": "FA-VAE model achieve a better reconstruction because the residual connections in FCMs preserve better the information flow. The reconstruction is further improved when combined with a simple SL on images level (row 5) or blocks level (row 6). Although SL (row 6) shows slightly inferior reconstruction due to neglection of frequency importance variance across decoder blocks, Figure 7 image 5 shows that the details, such as earrings and hairs, can still be more enhanced than the baseline model shown in image 1." + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1742" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.091, + 0.093, + 0.472, + 0.101 + ], + "angle": 0, + "content": "the woman has high cheekbones, wavy hair, big nose, mouth open, big lips, pointy nose, and" + }, + { + "type": "text", + "bbox": [ + 0.138, + 0.101, + 0.424, + 0.11 + ], + "angle": 0, + "content": "arched eyebrows. she is young and wears heavy makeup, and lipstick." + }, + { + "type": "image", + "bbox": [ + 0.084, + 0.109, + 0.212, + 0.207 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.219, + 0.111, + 0.349, + 0.207 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.356, + 0.111, + 0.484, + 0.207 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.62, + 0.097, + 0.762, + 0.106 + ], + "angle": 0, + "content": "He has goatee and wears necktie." + }, + { + "type": "image", + "bbox": [ + 0.492, + 0.109, + 0.621, + 0.207 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.627, + 0.109, + 0.756, + 0.207 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.764, + 0.109, + 0.892, + 0.207 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.302, + 0.218, + 0.668, + 0.233 + ], + "angle": 0, + "content": "Figure 9. Text-to-Image generation on CelebA-HQ-MM [17]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.241, + 0.47, + 0.451 + ], + "angle": 0, + "content": "Then, Table 2 rows 7-9 compare the performance when the architecture of FCM varies as illustrated in Figure 5; note that instead of using same \\(\\sigma_{i}\\) for \\((\\mathcal{A}_i,\\mathcal{C}_i)\\), we use two learnable \\(\\sigma\\) for \\(\\mathcal{A}_i\\) and \\(\\mathcal{C}_i\\) respectively. Quantitatively, when everything else is held equal, the convolution architecture shows better reconstruction than residual and attention architectures. Similarly, Figure 7 image 7 resembles the original image more than images 8 and 9 in terms of face radiance, smoothness, lip color, and shape. The reason is that residual connection enriches the outputs of FCM, which defies the purpose of enriching features of higher frequencies of the decoder features motivated in section 3.1.1. The attention mechanism is at a disadvantage here because the decoder and the encoder exclusively use convolutions." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.454, + 0.47, + 0.589 + ], + "angle": 0, + "content": "Finally, varying the size of kernels in DSL (Table 2 row 10 - 14) show quite similar quantitative reconstruction performance, while qualitatively, Figure 7 shows that a larger kernel size tends to produce smoother reconstructions. One reason could be that a larger kernel size tends to smooth more the feature maps because more surrounding values are used during convolution. Thus, in other experiments, we choose a kernel size of 3, and the effects of kernel sizes are open to future works." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.592, + 0.47, + 0.757 + ], + "angle": 0, + "content": "Zero-shot Reconstruction To further demonstrate the reconstruction capabilities of FA-VAE models, Table 3 gives the zero-shot reconstruction performance evaluated on CelebA-HQ [17] and FFHQ [21] using models trained on ImageNet [9]. The accompanying qualitative results are in Figure 8. Overall, FA-VAE displays impressive transferability capability with more faithful reconstruction in details, for instance, the light contrast in column 1 and book details in column 3. Note also that the common metric used in image compression PSNR is not perfectly correlated with rFID, but we put the metric here for reference." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.771, + 0.239, + 0.787 + ], + "angle": 0, + "content": "4.3. Image Synthesis" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.795, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Table 4 shows the text-to-image generation performance on CelebA-HQ-MM [17]. Our proposed autoregressive transformer CAT yields a better generation quality than other GAN-based models, including AttnGAN [52], ControlGAN [26], which are solely designed for image generation. Figure 9 shows that CAT can generate satisfactory images conditioned on text inputs of varying lengths. At-" + }, + { + "type": "table", + "bbox": [ + 0.507, + 0.238, + 0.888, + 0.339 + ], + "angle": 0, + "content": "
CelebAFFHQ
Pretrained ModelfrFID ↓PSNR ↑rFID ↓PSNR ↑
VQ-GAN [40]168.6223.406.8322.68
VQ-GAN [11]44.1131.203.0630.82
FA-VAE (Ours)166.5222.596.1921.95
FA-VAE (Ours)42.2531.391.4630.85
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.499, + 0.341, + 0.894, + 0.398 + ], + "angle": 0, + "content": "Table 3. Zero-shot reconstruction results on the validation data of FFHQ [21] and CelebA-HQ [17] using models trained on ImageNet [9]. \\( f \\) is the downsampling factor, the codebook sizes for \\( f = \\{ {16},8,4\\} \\) are \\( \\{ {16384},{16384},{8192}\\} \\) respectively." + }, + { + "type": "table", + "bbox": [ + 0.593, + 0.402, + 0.803, + 0.519 + ], + "angle": 0, + "content": "
ModelFID ↓
AttnGAN [52]125.98
ControlGAN [26]116.32
DM-GAN [55]131.05
DF-GAN [44]137.60
TediGAN [50]106.37
LAFITE [54]12.54
CAT (Ours)10.23
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.498, + 0.522, + 0.892, + 0.613 + ], + "angle": 0, + "content": "Table 4. Text-to-image generation on CelebA-HQ MM [50]. tributes such as \"mouth open\" and \"arched eyebrows\" are captured during generation because all the tokens embeddings are used as a condition which gives more precise guidance. More quantitative and qualitative results on different datasets are in the supplement." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.626, + 0.62, + 0.642 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.651, + 0.892, + 0.832 + ], + "angle": 0, + "content": "In this paper, we introduce the Frequency Augmented VAE (FA-VAE) model, which aims to improve reconstruction quality by bridging the gaps in the frequency domains between original and reconstructed images. New modules named Frequency Complement Modules (FCM) are crafted and guided under the new (Dynamic) Spectrum Loss ((D)SL) to learn to complement the reconstructed features of missing frequencies. A new Cross-attention Autoregressive Transformer (CAT) is proposed for achieving more precise textual-image alignment in the text-to-image generation task. FA-VAE shows improved reconstruction on various datasets compared to SOTA methods." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.846, + 0.661, + 0.863 + ], + "angle": 0, + "content": "Acknowledgement" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.891, + 0.9 + ], + "angle": 0, + "content": "This research is supported in part by the National Science Foundation under award No. 2040209." + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1743" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.117, + 0.468, + 0.158 + ], + "angle": 0, + "content": "[1] Jyoti Aneja, Alexander G. Schwing, Jan Kautz, and Arash Vahdat. Ncp-vae: Variational autoencoders with noise contrastive priors. *ArXiv*, abs/2010.02917, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.16, + 0.468, + 0.201 + ], + "angle": 0, + "content": "[2] Valerio Biscione and Jeffrey S. Bowers. Convolutional neural networks are not invariant to translation, but they can learn to be. JMLR, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.204, + 0.468, + 0.245 + ], + "angle": 0, + "content": "[3] Andrew Brock, Jeff Donahue, and Karen Simonyan. Large scale GAN training for high fidelity natural image synthesis. In ICLR, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.248, + 0.468, + 0.288 + ], + "angle": 0, + "content": "[4] Huiwen Chang, Han Zhang, Lu Jiang, Ce Liu, and William T. Freeman. Maskgit: Masked generative image transformer. In CVPR, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.292, + 0.468, + 0.332 + ], + "angle": 0, + "content": "[5] Ting Chen, Ruixiang Zhang, and Geoffrey Hinton. Analog bits: Generating discrete data using diffusion models with self-conditioning, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.336, + 0.468, + 0.376 + ], + "angle": 0, + "content": "[6] Rewon Child. Very deep \\(\\{\\mathrm{vae}\\}\\)s generalize autoregressive models and can outperform them on images. In ICLR, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.379, + 0.468, + 0.42 + ], + "angle": 0, + "content": "[7] Jooyoung Choi, Jungbeom Lee, Chaehun Shin, Sungwon Kim, Hyunwoo Kim, and Sungroh Yoon. Perception prioritized training of diffusion models. In CVPR, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.423, + 0.468, + 0.491 + ], + "angle": 0, + "content": "[8] Katherine Crowson, Stella Rose Biderman, Daniel Kornis, Dashiell Stander, Eric Hallahan, Louis Castricato, and Edward Raff. Vqgan-clip: Open domain image generation and editing with natural language guidance. In ECCV, 2022. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.495, + 0.468, + 0.535 + ], + "angle": 0, + "content": "[9] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In CVPR, 2009. 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.539, + 0.468, + 0.566 + ], + "angle": 0, + "content": "[10] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. In NeurIPS, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.569, + 0.468, + 0.609 + ], + "angle": 0, + "content": "[11] Patrick Esser, Robin Rombach, and Björn Ommer. Taming transformers for high-resolution image synthesis. In CVPR, 2021. 1, 2, 3, 4, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.613, + 0.468, + 0.667 + ], + "angle": 0, + "content": "[12] Ge Gao, Pei You, Rong Pan, Shunyuan Han, Yuanyuan Zhang, Yuchao Dai, and Hojae Lee. Neural image compression via attentional multi-scale back projection and frequency decomposition. In ICCV, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.671, + 0.468, + 0.711 + ], + "angle": 0, + "content": "[13] Karol Gregor, Ivo Danihelka, Andriy Mnih, Charles Blundell, and Daan Wierstra. Deep autoregressive networks. In ICML, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.714, + 0.468, + 0.768 + ], + "angle": 0, + "content": "[14] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. In NeurIPS, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.772, + 0.468, + 0.812 + ], + "angle": 0, + "content": "[15] Liming Jiang, Bo Dai, Wayne Wu, and Chen Change Loy. Focal frequency loss for image reconstruction and synthesis. In ICCV, 2021. 2, 4, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.816, + 0.468, + 0.856 + ], + "angle": 0, + "content": "[16] Justin Johnson, Alexandre Alahi, and Li Fei-Fei. Perceptual losses for real-time style transfer and super-resolution. In ECCV, 2016. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.859, + 0.468, + 0.9 + ], + "angle": 0, + "content": "[17] Tero Karras, Timo Aila, Samuli Laine, and Jaakko Lehtinen. Progressive growing of gans for improved quality, stability, and variation. CoRR, abs/1710.10196, 2017. 7, 8" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.117, + 0.468, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.134 + ], + "angle": 0, + "content": "[18] Tero Karras, Timo Aila, Samuli Laine, and Jaakko Lehtinen. Progressive growing of GANs for improved quality, stability, and variation. In ICLR, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.137, + 0.892, + 0.177 + ], + "angle": 0, + "content": "[19] Tero Karras, Miika Aittala, Janne Hellsten, Samuli Laine, Jaakko Lehtinen, and Timo Aila. Training generative adversarial networks with limited data. In NeurIPS, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.18, + 0.892, + 0.221 + ], + "angle": 0, + "content": "[20] Tero Karras, Miika Aittala, Samuli Laine, Erik Härkönen, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Alias-free generative adversarial networks. In NeurIPS, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.224, + 0.892, + 0.263 + ], + "angle": 0, + "content": "[21] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In CVPR, 2019. 2, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.267, + 0.892, + 0.322 + ], + "angle": 0, + "content": "[22] Soo Ye Kim, Kfir Aberman, Nori Kanazawa, Rahul Garg, Neal Wadhwa, Huiwen Chang, Nikhil Karnad, Munchurl Kim, and Orly Liba. Zoom-to-inpaint: Image inpainting with high frequency details. In CVPRW, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.325, + 0.892, + 0.352 + ], + "angle": 0, + "content": "[23] Diederik P Kingma and Max Welling. Auto-encoding variational bayes, 2013. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.354, + 0.892, + 0.422 + ], + "angle": 0, + "content": "[24] Umut Kocasari, Alara Dirik, Mert Tiftikci, and Pinar Yanardag. Stylemc: Multi-channel based fast text-guided image generation and manipulation. 2022 IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 3441-3450, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.426, + 0.892, + 0.466 + ], + "angle": 0, + "content": "[25] Doyup Lee, Chiheon Kim, Saehoon Kim, Minsu Cho, and Wook-Shin Han. Autoregressive image generation using residual quantization. In CVPR, 2022. 1, 2, 3, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.469, + 0.892, + 0.509 + ], + "angle": 0, + "content": "[26] Bowen Li, Xiaojuan Qi, Thomas Lukasiewicz, and Philip H. S. Torr. Controllable text-to-image generation. In NeurIPS, 2019. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.513, + 0.892, + 0.58 + ], + "angle": 0, + "content": "[27] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. GLIDE: towards photorealistic image generation and editing with text-guided diffusion models. In ICML, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.584, + 0.892, + 0.611 + ], + "angle": 0, + "content": "[28] Alexander Quinn Nichol and Prafulla Dhariwal. Improved denoising diffusion probabilistic models. In ICML, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.614, + 0.892, + 0.654 + ], + "angle": 0, + "content": "[29] Gaurav Parmar, Dacheng Li, Kwonjoon Lee, and Zhuowen Tu. Dual contradistinctive generative autoencoder. In CVPR, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.657, + 0.892, + 0.685 + ], + "angle": 0, + "content": "[30] Yash Patel, Srikar Appalaraju, and R. Manmatha. Deep perceptual compression, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.687, + 0.892, + 0.728 + ], + "angle": 0, + "content": "[31] Jialun Peng, Dong Liu, Songcen Xu, and Houqiang Li. Generating diverse structure for image inpainting with hierarchical vq-vae. In CVPR, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.731, + 0.892, + 0.77 + ], + "angle": 0, + "content": "[32] Stanislav Pidhorskyi, Donald A. Adjeroh, and Gianfranco Doretto. Adversarial latent autoencoders. In CVPR, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.774, + 0.892, + 0.856 + ], + "angle": 0, + "content": "[33] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In ICML, 2021. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.859, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[34] Alec Radford, Jeff Wu, Rewon Child, David Luan, Dario Amodei, and Ilya Sutskever. Language models are unsupervised multitask learners. 2019. 6" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1744" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.16 + ], + "angle": 0, + "content": "[35] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J. Liu. Exploring the limits of transfer learning with a unified text-to-text transformer. JMLR, 21(140):1-67, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.165, + 0.47, + 0.219 + ], + "angle": 0, + "content": "[36] Nasim Rahaman, Aristide Baratin, Devansh Arpit, Felix Dräxler, Min Lin, Fred A. Hamprecht, Yoshua Bengio, and Aaron C. Courville. On the spectral bias of neural networks. In ICML, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.222, + 0.469, + 0.262 + ], + "angle": 0, + "content": "[37] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. ArXiv, abs/2204.06125, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.266, + 0.469, + 0.307 + ], + "angle": 0, + "content": "[38] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In ICML, 2021. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.31, + 0.469, + 0.35 + ], + "angle": 0, + "content": "[39] Ali Razavi, Aaron van den Oord, and Oriol Vinyals. Generating diverse high-fidelity images with vq-vae-2. In NeurIPS, 2019. 1, 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.354, + 0.469, + 0.407 + ], + "angle": 0, + "content": "[40] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 1, 3, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.411, + 0.469, + 0.507 + ], + "angle": 0, + "content": "[41] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L. Denton, Seyed Kamyar Seyed Ghasemipour, Burcu Karagol Ayan, Seyedeh Sara Mahdavi, Raphael Gontijo Lopes, Tim Salimans, Jonathan Ho, David J. Fleet, and Mohammad Norouzi. Photorealistic text-to-image diffusion models with deep language understanding. ArXiv, abs/2205.11487, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.511, + 0.469, + 0.55 + ], + "angle": 0, + "content": "[42] Edgar Schonfeld, Bernt Schiele, and Anna Khoreva. A u-net based discriminator for generative adversarial networks. In CVPR, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.554, + 0.469, + 0.594 + ], + "angle": 0, + "content": "[43] Jianxin Sun, Qiyao Deng, Qi Li, Muyi Sun, Min Ren, and Zhenan Sun. Anyface: Free-style text-to-face synthesis and manipulation. In CVPR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.598, + 0.469, + 0.639 + ], + "angle": 0, + "content": "[44] Ming Tao, Hao Tang, Fei Wu, Xiao-Yuan Jing, Bing-Kun Bao, and Changsheng Xu. Df-gan: A simple and effective baseline for text-to-image synthesis. In CVPR, 2022. 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.641, + 0.469, + 0.668 + ], + "angle": 0, + "content": "[45] Arash Vahdat and Jan Kautz. NVAE: A deep hierarchical variational autoencoder. In NeurIPS, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.672, + 0.469, + 0.711 + ], + "angle": 0, + "content": "[46] Aaron van den Oord, Oriol Vinyals, and koray kavukcuoglu. Neural discrete representation learning. In NeurIPS, 2017. 1, 2, 3, 4, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.715, + 0.469, + 0.755 + ], + "angle": 0, + "content": "[47] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In NeurIPS, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.758, + 0.469, + 0.799 + ], + "angle": 0, + "content": "[48] Haohan Wang, Xindi Wu, Zeyi Huang, and Eric P. Xing. High-frequency component helps explain the generalization of convolutional neural networks. In CVPR, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.802, + 0.469, + 0.856 + ], + "angle": 0, + "content": "[49] Zihao Wang, Wei Liu, Qian He, Xinglong Wu, and Zili Yi. Clip-gen: Language-free training of a text-to-image generator with clip. arXiv preprint arXiv:2203.00386, 2022. 1, 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.86, + 0.469, + 0.9 + ], + "angle": 0, + "content": "[50] Weihao Xia, Yujiu Yang, Jing-Hao Xue, and Baoyuan Wu. Tedigan: Text-guided diverse face image generation and manipulation. In CVPR, 2021. 2, 6, 8" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.134 + ], + "angle": 0, + "content": "[51] Zhisheng Xiao, Karsten Kreis, Jan Kautz, and Arash Vahdat. Vaebm: A symbiosis between variational autoencoders and energy-based models. In ICLR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.137, + 0.892, + 0.19 + ], + "angle": 0, + "content": "[52] Tao Xu, Pengchuan Zhang, Qiuyuan Huang, Han Zhang, Zhe Gan, Xiaolei Huang, and Xiaodong He. Attngan: Fine-grained text to image generation with attentional generative adversarial networks. In CVPR, 2018. 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.193, + 0.892, + 0.247 + ], + "angle": 0, + "content": "[53] Jiahui Yu, Xin Li, Jing Yu Koh, Han Zhang, Ruoming Pang, James Qin, Alexander Ku, Yuanzhong Xu, Jason Baldridge, and Yonghui Wu. Vector-quantized image modeling with improved VQGAN. In ICLR, 2022. 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.25, + 0.892, + 0.303 + ], + "angle": 0, + "content": "[54] Yufan Zhou, Ruiyi Zhang, Changyou Chen, Chunyuan Li, Chris Tensmeyer, Tong Yu, Jiuming Gu, Jinhui Xu, and Tong Sun. Lafite: Towards language-free training for text-to-image generation. CVPR, 2022. 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.306, + 0.892, + 0.347 + ], + "angle": 0, + "content": "[55] Minfeng Zhu, Pingbo Pan, Wei Chen, and Yi Yang. Dm-gan: Dynamic memory generative adversarial networks for text-to-image synthesis. In CVPR, 2019. 2, 8" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.347 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1745" + } + ] +] \ No newline at end of file diff --git a/2023/Catch Missing Details_ Image Reconstruction With Frequency Augmented Variational Autoencoder/0a253bb0-c0db-46ba-9d61-8be0d603b311_origin.pdf b/2023/Catch Missing Details_ Image Reconstruction With Frequency Augmented Variational Autoencoder/0a253bb0-c0db-46ba-9d61-8be0d603b311_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..23b50d3cbbf4feda6afd8d47d314f693ccdc0114 --- /dev/null +++ b/2023/Catch Missing Details_ Image Reconstruction With Frequency Augmented Variational Autoencoder/0a253bb0-c0db-46ba-9d61-8be0d603b311_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d611948b32a9399ed3751c538af379682e08f8f4500d231beef01a6cf056db25 +size 6220898 diff --git a/2023/Catch Missing Details_ Image Reconstruction With Frequency Augmented Variational Autoencoder/full.md b/2023/Catch Missing Details_ Image Reconstruction With Frequency Augmented Variational Autoencoder/full.md new file mode 100644 index 0000000000000000000000000000000000000000..873aa18d21fe66518d71fdbbfc451322344c3e6f --- /dev/null +++ b/2023/Catch Missing Details_ Image Reconstruction With Frequency Augmented Variational Autoencoder/full.md @@ -0,0 +1,365 @@ +# Catch Missing Details: Image Reconstruction with Frequency Augmented Variational Autoencoder + +Xinmiao Lin + +Rochester Institute of Technology + +x13439@rit.edu + +Chiuman Ho + +OPPO US Research + +chiuman@oppo.com + +Yikang Li + +OPPO US Research + +yikang.li1@oppo.com + +Jenhao Hsiao + +OPPO US Research + +mark@oppo.com + +# Abstract + +The popular VQ-VAE models reconstruct images through learning a discrete codebook but suffer from a significant issue in the rapid quality degradation of image reconstruction as the compression rate rises. One major reason is that a higher compression rate induces more loss of visual signals on the higher frequency spectrum which reflect the details on pixel space. In this paper, a Frequency Complement Module (FCM) architecture is proposed to capture the missing frequency information for enhancing reconstruction quality. The FCM can be easily incorporated into the VQ-VAE structure, and we refer to the new model as Frequency Augmented VAE (FA-VAE). In addition, a Dynamic Spectrum Loss (DSL) is introduced to guide the FCMs to balance between various frequencies dynamically for optimal reconstruction. FA-VAE is further extended to the text-to-image synthesis task, and a Cross-attention Autoregressive Transformer (CAT) is proposed to obtain more precise semantic attributes in texts. Extensive reconstruction experiments with different compression rates are conducted on several benchmark datasets, and the results demonstrate that the proposed FA-VAE is able to restore more faithfully the details compared to SOTA methods. CAT also shows improved generation quality with better image-text semantic alignment. + +# 1. Introduction + +VQ-VAE models [6,11,25,29,39,46] reconstruct images through learning a discrete codebook of latent embeddings. They gained wide popularity due to the scalable and versatile codebook, which can be broadly applied to many visual tasks such as image synthesis [11,49] and inpainting [4,31]. A higher compression rate is typically preferable in VQ- + +# Yu Kong +Michigan State University + +yukong@msu.edu + +![](images/47894a809973432e07174ef8fa48bbe6243ccdecc6c21b8ad23d4aa1df795a9e.jpg) +Figure 1. Images and their frequency maps. Row 1: original and reconstructed images. Row 2: the frequency maps of images, frequency increases in any direction away from the center. $f$ is the compression rate. With a greater compression rate, more details are lost during reconstruction, i.e. eyes and mouth shape, and hair texture (pointed with red arrows) which align with the loss of high-frequency features. All frequency figures in this paper use the same colormap. rFID [14] and lpips [16] are lower the better, and frequency values increase from red to green, zoom in for better visualization. + +VAE models since it provides memory efficiency and better learning of coherent semantics structures [11, 40]. + +One main challenge quickly arises for a higher compression rate, which severely compromises reconstruction accuracy. Figure 1 row 1 shows that although the reconstructed images at higher compression rates appear consistent with the original image, details inconsistencies such as the color and contour of the lips become apparent upon closer scrutiny. Figure 1 row 2 reveals that similar degradation also manifests on the frequency domains where features towards the middle and higher frequency spectrum are the least recoverable with greater compression rate. + +Several causes stand behind this gap between pixel and frequency space. The convolutional nature of autoencoders is prone to spectral bias, which favors learning low-frequency features [22, 36]. This challenge is further ag- + +gravated when current methods exclusively design losses or improve model architecture for better semantics resemblance [11, 16, 25] but often neglect the alignment on the frequency domain [12, 15]. On top of that, it is intuitively more challenging for a decoder to reconstruct an image patch from a single codebook embedding (high compression) than multiple embeddings (less compression). The reason is that the former mixes up features of incomplete and diverse frequencies, while the latter could preserve more fine-grained and complete features at various frequencies. + +Inspired by these insights, the Frequency Augmented VAE (FA-VAE) model is proposed, which aims to improve reconstruction quality by achieving better alignment on the frequency spectrums between the original and reconstructed images. More specifically, new modules named Frequency Complement Modules (FCM) are crafted and embedded at multiple layers of FA-VAE's decoder to learn to complement the decoder's features with missing frequencies. + +We observe that valuable middle and high frequencies are mingled with the encoder's feature maps during the compression via an encoder, shown in Figure 3 row 4. Therefore, a new loss termed Spectrum Loss (SL) is proposed to guide FCMs to generate missing features that align with the same level's encoder features on the frequency domain. Since most image semantics reside on the low-frequency spectrum [48], SL prioritizes learning lower-frequency features with diminishing weights as frequencies increase. + +Interestingly, we discover that checkerboard patterns appear in the complemented decoder's features with SL, although better reconstruction performance is achieved (Figure 3 column 4). We speculate that because SL sets a deterministic range for the low-frequency spectrum when applying weights on the frequencies without considering that the importance of a frequency can vary from layer to layer. Thus, an improved loss function Dynamic Spectrum Loss (DSL) is crafted on top of SL with a learnable component to adjust the range of low-frequency spectrum dynamically for optimal reconstruction. DSL can improve reconstruction quality even further than SL without the unnatural checkerboard artifacts in the features (Figure 3 column 5). + +We further extend FA-VAE to the text-to-image generation task and propose the Cross-attention Autoregressive Transformer (CAT) model. We first observe that only using one or a few token embeddings is a coarse representation of lengthy texts [8, 27, 33]. Thus CAT uses all token embeddings as a condition for more precise guidance. Moreover, existing works typically use self-attention, and the text condition is embedded merely at the beginning of the generation [11, 49]. This mechanism becomes problematic in the autoregressive generation because one image token is generated at a time, thus the text condition gradually loosens its connection with the generated tokens. To circumvent this + +issue, CAT embeds a cross-attention mechanism that allows the text condition to guide each step generation. + +To summarize, our work includes the following contributions: + +- We propose a new type of architecture called Frequency Augmented VAE (FA-VAE) for improving image reconstruction through achieving more accurate details reconstruction. +- We propose a new loss called Spectrum Loss (SL) and its enhanced version Dynamic Spectrum Loss (D SL), which guides the Frequency Complement Modules (FCM) in FA-VAE to adaptively learn different low/high frequency mixtures for optimal reconstruction. +- We propose a new Cross-attention Autoregressive Transformer (CAT) for text-to-image generation using more fine-grained textual embeddings as a condition with a cross-attention mechanism for better image-text semantic alignment. + +# 2. Related Work & Background + +Image Reconstruction Vector Quantized-Variational AutoEncoder (VQ-VAE) [46] extends the Variational AutoEncoder structure [23] and proposes to encode images into discrete latent codes using vector quantization (VQ). Then a generative model, such as an autoregressive transformer [13], can be trained and paired with the decoder in VQ-VAE to synthesize new images. Later works [11,25,39] further improve the reconstruction and generation quality by improving the generative model architecture or the quantization efficiency. Since VQ-VAE models operate on discrete latent spaces, they cannot be directly and fairly compared to other VAE-based models that employ continuous latent space [1,6,23,29,45,51]. In contrast to current VQ-VAE models, our proposed FA-VAE technique improves reconstruction through the frequency angle and is more generalized since the FCMs can be readily extended to other neural networks that share the same VQ-VAE structure. + +Few other works do image reconstruction via the frequency perspective. FFL [15] proposed a loss function to penalize differences on the hard frequencies. A new module, which is designed in the continuous latent space for image compression, is proposed to be incorporated into the encoder and decoder in [12]. To our best knowledge, FA-VAE is the first work that aims to improve image reconstruction on discrete latent space through the frequency perspective. + +Image Generation Image generation can be achieved via GAN-based models [3, 18, 42, 44, 52, 55], which synthesize images from noise vectors unconditionally or conditioned on different inputs such as texts, masks, etc. It becomes cumbersome to train one model for one application. Thus, StyleGAN [21] encodes semantic attributes into a continuous latent space, and subsequent works [19, 20, + +![](images/54efa40dc53ecdaf9a16925bc2ca99f28be9dba84c340692f3429da4f3c8bce4.jpg) +Figure 2. Frequency Augmented VAE (FA-VAE). (a) The encoder $\mathcal{E}$ encodes the images $X$ onto discrete latent codebook space $\hat{z}_q$ which is used by the decoder $\mathcal{G}$ to reconstruct images $\hat{X}$ . DSL (b) guides the FCMs (c) to learn to complement the reconstructed features with missing features of important frequencies in order to improve reconstruction quality. A more detailed figure is in the supplement. + +![](images/b8d3e2c0f9326268fe96599a321282dea36c00873d5856951720531629ea190d.jpg) + +24, 32, 43, 50, 54] leverage this space for generating images conditioned on attributes or textual descriptions. However, StyleGAN-based models cannot scale to large datasets when the number of attributes becomes substantially large because this demands to increase model's size. In contrast, the codebook in VQ-VAE models [46] is scalable to large datasets without additional model complexity. Diffusion-based models [5, 7, 10, 27, 28, 37, 40, 41] generate images from Gaussian noise through a reverse diffusion process can often require substantial training and large datasets. Most text-to-image generation models commonly use one of a few embeddings for a condition [8, 27, 38, 49] from pretrained models such as CLIP [33] or T5-XXL [35]. In contrast, our proposed autoregressive transformer CAT uses the embeddings from all the text tokens for more fine-grained guidance during image generation. + +VQ-VAE We now describe the VQ-VAE model [46] as it is the backbone used in the FA-VAE model, which is presented in Figure 2. VQ-VAE [46] and related models [11, 25, 38, 39] consist of an encoder that encodes the images into a codebook of embeddings. Then, a decoder is trained to reconstruct the images from a discrete set of codebook embeddings. In this paper, reconstruction loss in VQ-GAN is utilized which is: + +$$ +\mathcal {L} _ {r e c} = \left\| X - \hat {X} \right\| _ {1} + \mathcal {L} _ {p i p s} (X - \hat {X}) \tag {1} +$$ + +where $X$ and $\hat{X}$ are the original and reconstructed images. We also use the adversarial setting of VQ-GAN model which introduced a discriminator compared to VQ-VAE models [39, 46], more details could be referred to [11]. + +# 3. Methodology + +The proposed Frequency Augmented VAE (FA-VAE) is presented in Figure 2. In this section, we describe how FA-VAE ameliorates reconstruction quality by bridging the spectrum domain gap between reconstructed and original images. By explicitly embedding Frequency Complement + +Modules (FCM) into the decoder, Dynamic Spectrum Loss (DSL) leverages important features frequency-wise from the encoder and guides the FCMs to complement the features of the decoder for better frequency restoration at different reconstruction stages. + +# 3.1. Frequency Augmented VAE (FA-VAE) + +Let the images be $X \in \mathbb{R}^{H \times W \times 3}$ . The codebook $\mathfrak{C}$ is a set of $|\mathfrak{C}|$ embeddings, such as $\mathfrak{C} = \{\mathfrak{c}_i | i = 1, \dots, |\mathfrak{c}|\} \in \mathbb{R}^{n_z}$ and $n_z$ is the length of one codebook embedding. FA-VAE consists of an encoder $\mathcal{E}$ that encodes the images $X$ into latent space representations such as $z = \mathcal{E}(x)$ , for $z \in \mathbb{R}^{h \times w \times n_z}$ and $x \in X$ , $(h \times w)$ is the resolution of the encoded representation. Let $f = H / h = W / w$ be the downsampling factor or compression rate. Each feature of $z$ is approximated by the vector quantization block $\mathcal{Q}$ using the nearest codebook entries, and it can be presented as: + +$$ +\mathcal {Q} (z) = \hat {z} _ {q} = \underset {c _ {k} \in \mathcal {C}} {\arg \min } \| z _ {i j} - \mathfrak {c} _ {k} \| \tag {2} +$$ + +where $\hat{z}_q$ is the quantized latent embedding and subsequently used by the decoder $\mathcal{G}$ to produce the reconstructed image $\hat{x} = \mathcal{G}(\hat{z}_q)$ . + +# 3.1.1 Frequency Complement Modules (FCM) + +Motivation Figure 1 shows that a higher compression rate leads to more significant reconstruction disparities on the higher-frequency spectrum. Existing models [11, 25, 39] guide the reconstructed images to be more aligned on the pixel and feature spaces with the original images (Eq. 1) but neglect frequency spectrum alignment and leave the encoder and decoder without further guidance. Figure 3 column 2 shows that the encoder activations $\mathcal{A}_1$ of baseline VQ-GAN [11] contain rich high-frequency features (row 3 & 4), but the decoder's activations could mainly restore low-frequency features (row 5 & 6). + +Therefore we propose Frequency Complement Modules (FCM), illustrated in Figure 2 (c), which aims to complement the decoder's feature maps $\mathcal{B}_i$ with features of missing frequencies using the encoder activations $\mathcal{A}_i$ . The FCMs $\mathcal{F}_i$ consist of sequences of convolution layers and activations. The decoder $\mathcal{G}$ with FCMs embedded can be represented as: + +$$ +\mathcal {B} _ {M - 1} = \mathcal {G} _ {M} \left(\mathcal {C} _ {M} + \hat {z} _ {q}\right) = \mathcal {G} _ {M} \left(\mathcal {F} _ {M} (\hat {z} _ {q}) + \hat {z} _ {q}\right) +$$ + +$$ +\dots \tag {3} +$$ + +$$ +\mathcal {B} _ {1} = \mathcal {G} _ {2} \left(\mathcal {C} _ {2} + \mathcal {B} _ {2}\right) = \mathcal {G} _ {2} \left(\mathcal {F} _ {2} \left(\mathcal {B} _ {2}\right) + \mathcal {B} _ {2}\right) +$$ + +Similarly, the encoder $\mathcal{E}$ can be abstracted to: + +$$ +z = \mathcal {E} (x) = \mathcal {E} _ {M} \left(\dots \left(\mathcal {E} _ {2} (\mathcal {E} _ {1} (x)))\right)\right) \tag {4} +$$ + +where $\mathcal{G}_i$ is the $i$ -th layer of the decoder, and $\mathcal{A}_i = \mathcal{E}_i(x)$ is the outputs or feature maps of the corresponding encoder block. The outputs of previous blocks $\mathcal{B}_i$ are complemented by outputs $\mathcal{C}_i = \mathcal{F}_i(\mathcal{B}_{i + 1})$ , which contain the frequency-rich features learned from the encoder activations $\mathcal{A}_i$ . The following section describes how Dynamic Frequency Loss (DSL) guides the learning of FCMs to specifically learn to complement $\mathcal{B}_i$ with the features of missing frequencies. + +In the paper, the frequency feature compensation is implemented by addition (in Eq. 3). However, the FCM is flexible in adopting any architecture and merging techniques, as illustrated by the three examples in Figure 5. Note that no architecture limitation is imposed on the encoder and decoder blocks as long as $\mathcal{A}_i$ and $\mathcal{B}_i$ share the same resolution. VQ-GAN [11] and related models [25,46] already have similar architecture, and VQ-GAN is chosen as the backbone for FA-VAE. + +# 3.1.2 Dynamic Spectrum Loss (DSL) + +Motivation Spectrum Loss (SL) is proposed to guide explicitly the outputs of FCMs $\mathcal{C}_i$ to be more aligned with the encoder's activations $\mathcal{A}_i$ on the frequency spectrum because the latter contains rich features on the higher frequency spectrum (Figure 3). Moreover, to account for the varying importance of frequencies across different decoder stages, Dynamic Spectrum Loss (DSL) is proposed, a more generalized variant of SL, and has the ability to adjust weights put on higher-frequency spectrum adaptively. Thus, each decoder block's outputs can be enriched with features of the most critical frequencies for an accurate reconstruction. + +Background The outputs $\mathcal{A}_i$ of encoder block $\mathcal{E}_i$ and $\mathcal{C}_i$ of FCM block $\mathcal{F}_i$ are first transformed to the frequency domain using Discrete Fourier Transform (DFT) as follows: + +$$ +F (u, v) = \sum_ {x = 0} ^ {M - 1} \sum_ {y = 0} ^ {N - 1} f (x, y) \cdot e ^ {- i 2 \pi \left(\frac {u x}{M} + \frac {v y}{N}\right)} \tag {5} +$$ + +![](images/b90b2d4f5584033841c23e4b3bf314d6aad03e39cb0d50246f860c60030de423.jpg) +Figure 3. Image, activations, and their frequency maps. Ours* is FA-VAE model with FFL [15], Ours‡ is FA-VAE with SL and Ours is FA-VAE model with DSL. DSL shows a more harmonious balance between low- and high-frequencies (row 5) and more accurate reconstruction in mouth and hair textures. The frequency maps are plotted using the average of all channels and the contrast is adjusted to emphasize the higher frequency spectrums. + +where $e$ and $i$ are Euler's number and the imaginary unit. $M \times N$ is the spatial resolution of feature maps and the Fourier Transform (Eq. 5) is applied to each of them. $f(x,y)$ is the value at $(x,y)$ of each feature map. $F(u,v)$ is the corresponding value at $(u,v)$ coordinates on the frequency spectrum. The Focal Frequency Loss (FFL) [15] can be presented as: + +$$ +\operatorname {F F L} \left(\mathcal {A} _ {i}, \mathcal {C} _ {i}\right) = \frac {1}{M N \left| \mathcal {C} _ {i} \right|} \sum_ {c = 0} ^ {\left| \mathcal {C} _ {i} \right| - 1} \sum_ {u = 0} ^ {M - 1} \sum_ {v = 0} ^ {N - 1} w (u, v) J (u, v), \tag {6} +$$ + +where $w(u,v) = |F_{\mathcal{A}_i}(u,v) - F_{\mathcal{C}_i}(u,v)|$ are the weights put on each frequency. $J(u,v) = |F_{\mathcal{A}_i}(u,v) - F_{\mathcal{C}_i}(u,v)|^2$ is the main error function based on the frequency difference. The $|\mathcal{C}_i|$ is the number of feature maps in $\mathcal{A}_i$ and $\mathcal{C}_i$ . Both real and imaginary parts of the frequency domain are considered, more details are in [15]. + +Limitations of FFL Eq. 6 demonstrates that FFL puts modulating term $w(u, v)$ to focus learning on the hardest frequencies for reconstruction, which are the higher frequencies following our observation (Figures 1 & 3). However, this could not be ideal because features on the lower frequencies define the image content, and overemphasizing the higher frequencies could over-constrain the learning + +![](images/8a2e58a0498ae831d9b106743f2ee53ecdd60c1343d9197660e65a05a86863b1.jpg) +Figure 4. Cross-attention Autoregressive Transformer (CAT). + +and lead to sub-optimal reconstruction, see Figure 7 row 1 column 3 and Table 2 row 3. Moreover, Figure 3 column 3 shows FCMs guided by a simple FFL can improve reconstruction performance (Ours\*). However the frequency maps of the decoder activations $\mathcal{C}_1$ contain excessive noise due to overemphasis on the higher frequency spectrum, and lower frequencies are neglected (row 6). + +Spectrum Loss (SL) Thus, we propose to apply a low-pass filter on the weights $w(u,v)$ in Eq. 6 to penalize more mismatch in the lower-frequency domain and gradually diminish the penalizing weights towards higher frequency spectrum. Therefore, let the Gaussian kernels be $K_{i}(\mu ,\sigma)$ with weights initialized using mean and standard deviation $(\mu ,\sigma_{i})$ , and applied over the feature maps as: + +$$ +(\hat {\mathcal {A}} _ {i}, \hat {\mathcal {C}} _ {i}) = (K _ {i} (\mu , \sigma_ {i}) \star \mathcal {A} _ {i}, K _ {i} (\mu , \sigma_ {i}) \star \mathcal {C} _ {i}). \qquad (7) +$$ + +where the $\star$ is the convolution operation. Then, the Spectrum Loss (SL) is defined as: + +$$ +\operatorname {S L} \left(\mathcal {A} _ {i}, \mathcal {C} _ {i}\right) = \operatorname {F F L} \left(\hat {\mathcal {A}} _ {i}, \hat {\mathcal {C}} _ {i}\right). \tag {8} +$$ + +Limitations of SL Up until now, a fixed Gaussian filter $K_{i}(\mu ,\sigma_{i})$ is applied over all $\mathcal{A}_i$ and $\mathcal{C}_i$ of all encoder and decoder blocks. Figure 3 column 4 demonstrates that SL $(\mathrm{Ours}^{\ddagger})$ improves reconstruction on the lower frequency spectrum, leading to better reconstruction than the baseline model (Table 2 row 5 vs row 1). However, checkerboard artifacts are also present (row 5) on $\mathcal{C}_1$ . One reason is that deterministic variance $\sigma_{i}$ assumes that decoder activations across different levels require the same amount of higher frequency features for accurate reconstruction, which could be an over-rigid constraint on the learning. Another reason could be that the same $\sigma_{i}$ magnifies the checkerboard effects produced by upsampling or striding operations in CNNs which are normally circumvented by subsequent convolutional layers [2, 30]. In the observation of the experiments, we find that the checkerboard effects can be intensified with the same $\sigma_{i}$ in different layers due to the bounding effect between low frequency and high frequency (shown in Figure 3 column 4 row 6). + +Dynamic Spectrum Loss (DSL) Therefore, we propose to optimize the variances $\sigma_{i}$ instead of setting them as static hyperparameters to suit the different amounts of frequencies needed for each block's $\mathcal{B}_i$ . The new Dynamic Spectrum Loss (DSL) is a more generalized form of Spectrum Loss + +![](images/884396341330d2834aad988835914292b738c9ff0e91466b21cc7ee2afc6bf40.jpg) +(a) FCM + +![](images/a185b0a415dcdd350010fbac884698b49521abaacf2dc9398fea1b8bd44aff9e.jpg) +(b) FCM w/ residual + +![](images/2c8618e727623ee74960bb7ffaa264f574e20e07e09360b454a36d66134545a6.jpg) +(c) FCM w/ attention +Figure 5. FCM with different architectures and connections. + +(SL) with learnable $\sigma_{i}$ . Note that DSL also includes FFL as a special form when used on the original and reconstructed images without the Gaussian filters. + +Then the total reconstruction loss for FA-VAE can be described as: + +$$ +\begin{array}{l} \mathcal {L} _ {r e c} = \alpha \operatorname {F F L} (X, \hat {X}) + \beta \sum_ {i = 0} ^ {M - 1} \operatorname {D S L} \left(\mathcal {A} _ {i}, \mathcal {C} _ {i}\right) \tag {9} \\ + \left\| X - \hat {X} \right\| _ {1} + \mathcal {L} _ {p i p s} (X - \hat {X}). \\ \end{array} +$$ + +where $\mathrm{FFL}(X - \hat{X})$ is the Focal Frequency Loss applied on the original reconstructed images. The second term in Eq. 9 is the DSL loss applied over the outputs of the encoder and FCM blocks. The first two losses aim to minimize the frequency spectrum differences on the images and internal feature maps. The third and fourth losses act on the pixel and feature maps space [11]. $\alpha$ and $\beta$ are hyperparameters. $\sigma_{i}$ are model parameters and optimized as: + +$$ +\sigma_ {i} ^ {*}, \mathcal {E} ^ {*}, \mathcal {G} ^ {*}, \mathcal {C} ^ {*} = \underset {\sigma_ {i}, \mathcal {E}, \mathcal {G}, \mathcal {C}} {\arg \min } \left(\mathcal {L} _ {\text {r e c}} + \mathcal {L} _ {Q}\right) \tag {10} +$$ + +$\mathcal{L}_Q$ is the quantization loss which minimizes the difference between the codebook embeddings and the embeddings given by the encoder $\mathcal{E}$ , more details in [11]. We also use the $L_{2}$ regularization on the codebook embeddings during quantization as in [53] and exponential moving average (EMA) for updating the codebook [46] as they provide more stable training for the quantization block $\mathcal{Q}$ . + +Benefits of DSL The advantages of learnable $\sigma_{i}$ in DSL are shown in Figure 3 column 5 where the reconstructed activations show no checkerboard artifacts as in column 4. Moreover, the reconstructed activations $C_1$ are more similar to the encoder activations $A_{1}$ on the frequency spectrum. Compared to the baseline model (column 2), our model exhibits a more harmonious balance between low- and highfrequencies, leading to more accurate reconstructed images. + +# 3.2. Cross-attention Autoregressive Transformer (CAT) + +We further extend FA-VAE to the text-to-image generation task and introduce a new Cross-attention Autoregres + +![](images/0ba7b2117c5b13a6e7cbcc7d10418207d814a840d3dd6f2e4fc0de38d9078575.jpg) +Figure 6. Reconstruction on ImageNet [9], label: Japanese terrier. + +sive Transformer (CAT), presented in Figure 4. CAT uses all the token embeddings of a textual description given by the pretrained text encoder of CLIP model [33], while existing works mostly use one or partial textual embeddings [8, 38, 49]. Furthermore, CAT uses a cross-attention mechanism with the text tokens to guide the image generation at each step. This more fine-grained text condition allows the generation to capture more precisely the relationships of semantic attributes between text and image. + +$$ +p (s \mid c) = \prod_ {i} p \left(s _ {i} \mid s _ {< i}, c\right) \tag {11} +$$ + +Then the predicted indices $z'$ can be decoded to an image using FA-VAE's decoder $G$ . The loss is to maximize the log-likelihood of the data representations, + +$$ +\mathcal {L} _ {\mathrm {C A T}} = \mathbb {E} _ {x \sim p (x)} [ - \log p (s) ] \tag {12} +$$ + +In this paper, the GPT2 model [34], which is the default setting in [11, 49], is utilized as the backbone for CAT, and cross-attention mechanism inspired from [47] is applied with the GPT2 structure. + +# 4. Experiments + +# 4.1. Experimental Details + +The datasets used in this paper are: (1) MultiModal CelebA-HQ dataset [50] with 30,000 high-resolution celebrity face images and each image comes with ten captions; (2) Flickr-Faces-HQ (FFHQ) dataset [21] of 70,000 high-resolution face images; (3) ImageNet [9] that contains around 1 million images from 1000 categories. Experiments are performed on V-100 GPUs and results are reported on the validation sets unless specified otherwise. For a fair comparison with existing models, the resolutions used for training are $256 \times 256$ on all datasets. The details of training settings and hyperparameters are provided in the supplementary materials. + +
ModelDatasetCodebook Size(h × w)rFID ↓
RQ-VAE [25]FFHQ2048(8 × 8)5.33
FA-VAE (Ours)FFHQ2048(16 × 16)4.98
VQ-VAE-2 [39]ImageNet512(64 × 64) & (32 × 32)~ 10 (train)
VQ-GAN [40]ImageNet8192(64 × 64)1.06
FA-VAE (Ours)ImageNet8192(64 × 64)0.40
DALL-E [38]ImageNet8192(32 × 32)32.01
VQ-GAN [11]ImageNet16384(16 × 16)5.15
VQ-GAN [11]ImageNet1024(16 × 16)7.94
VQ-GAN [25]ImageNet16384(8 × 8)17.95
RQ-VAE† [46]ImageNet16384(8 × 8)10.77
RQ-VAE* [25]ImageNet16384(8 × 8)4.73
FA-VAE (Ours)ImageNet16384(16 × 16)4.60
+ +Table 1. Reconstruction results on the validation data of FFHQ [21] and ImageNet [9] respectively, more results in the supplement. $\dagger$ means depth 2, and $*$ means depth 4 from RQ-VAE [25]. + +# 4.2. Image Reconstruction + +Reconstruction on FFHQ and Imagenet The experiment results of reconstruction on FFHQ [21] and ImageNet [9] are presented in Table 1. The results of the baselines are from the original paper. Note that smaller $h$ and $w$ means that the downsampling factor or the compression rate $f$ is larger. FA-VAE model shows improved reconstruction quality over baseline models across different compression rates in both datasets. The main reason is that FA-VAE can successfully reconstruct the important middle and high-frequencies which are neglected in baseline models. + +Figure 6 further supports our previous claim and demonstrates clear diverging reconstruction qualities between baseline models VQ-GAN [11], DALL-E [38], and our FA-VAE model. With a larger compression rate, Figure 6 row 2 shows that more middle and high frequencies are compressed in VQ-GAN and DALL-E. In comparison, FA-VAE can reconstruct middle and high frequencies more accurately, which translates to good representations with improved semantics and local details in Figure 6 row 1. For instance, at a compression rate of 16, the dog's tongue of + +
ablation onFFL(X, X̂)SL(Ai, Ci)FCM Fkernel K_i(μ, σ_i)DSL(Ai, Ci)kernel size μinitial value σ_ipair-wise σ_iL1 ↓Lpips ↓rFID ↓
1VQ-GAN [11]XXXXXXXX0.1210.3010.12
2VQ-GAN + Style [53]XXXXXXXX0.0850.2311.90
3VQ-GAN [15]✓XXXXXXX0.1140.3530.65
4SL w/o kernel✓XCONVXXXXX0.0820.227.04
5✓✓CONVXXXXX0.0820.227.02
6SL w/ kernel✓✓CONV✓X93X0.0800.227.39
7FCM architecture✓✓CONV✓✓93X0.0810.215.90
8✓✓RES✓✓93X0.0780.216.44
9✓✓ATTN✓✓93X0.0890.237.49
10DSL kernel size μ✓✓RES✓✓33✓0.0810.216.66
11✓✓RES✓✓53✓0.0810.217.04
12✓✓RES✓✓93✓0.0820.226.53
13✓✓RES✓✓113✓0.0830.227.48
14✓✓RES✓✓153✓0.0830.226.36
+ +Table 2. Ablation studies on the CelebA-HQ validation dataset [17], visualizations are in Figure 7. Words put in bold to highlight the changing component. + +![](images/baaff5b0e1bf3eb64c21e14321a4c13220115cfb0846f9f227cf826f024b83db.jpg) +Figure 7. Reconstruction comparisons for ablation studies on CelebA-HQ. The figure number represents the setting in the corresponding row in Table 2. $\mathrm{SL}^*$ is SL without Gaussian kernel, $\mathrm{DSL}^{**}$ is DSL with non pairwise $\sigma$ . FCM is convolution architecture by default. + +FA-VAE's reconstruction is more similar to the original image than VQ-GAN in terms of color and shape; more qualitative results are in the supplement. + +Ablation Studies on FCM and DSL In Table 2, we perform ablation studies on different architectures of FCM in combination with different settings of the Spectrum Loss (SL) and Dynamic Spectrum Loss (DSL). Figure 7 gives the accompanying visualizations of the quantitative results. + +First, as motivated in the Section 3, Table 2 row 3 shows that considering all frequencies as equally important leads the VQ-GAN model to poor performance in terms of rFID and lpips (see Figure 7 image 3). In comparison, results in Table 2 rows 4 demonstrate that FCMs alone can help + +![](images/932a0db2c62244545bb9d924242c6cdc083e564b88fe6a69abd09bc669dbf422.jpg) +Figure 8. Zero-shot reconstruction on CelebA-HQ [17], FFHQ [21] and photos from the wild, taken from [11]. The models are trained on ImageNet on different compression rates. + +FA-VAE model achieve a better reconstruction because the residual connections in FCMs preserve better the information flow. The reconstruction is further improved when combined with a simple SL on images level (row 5) or blocks level (row 6). Although SL (row 6) shows slightly inferior reconstruction due to neglection of frequency importance variance across decoder blocks, Figure 7 image 5 shows that the details, such as earrings and hairs, can still be more enhanced than the baseline model shown in image 1. + +the woman has high cheekbones, wavy hair, big nose, mouth open, big lips, pointy nose, and + +arched eyebrows. she is young and wears heavy makeup, and lipstick. + +![](images/cfe241304298cd6e8d8b818b986c9dc29b5a870d0f74404f1028cf72449795fb.jpg) + +![](images/5d53cae184c22bb6f3549e040602db9ec6ee33bf23c152baa7e129404438f150.jpg) +Figure 9. Text-to-Image generation on CelebA-HQ-MM [17] + +![](images/f4002f6c85c47a9bfad67f5021422678f8801addf8b46a7ef97ec0227f6de5c7.jpg) + +He has goatee and wears necktie. + +![](images/b18c3107873dd5e4fef8daa512f186606422465d46391a7cd90fcafc290df8fc.jpg) + +![](images/19510cdb08f01382673ee2673a3e2b70c5c9e73c5d46748decc291b48cf89fa8.jpg) + +![](images/e56a840f29eeaf83d60f908e87ab6a197b0f13d34f9bd022bc3950abe5f5cfdd.jpg) + +Then, Table 2 rows 7-9 compare the performance when the architecture of FCM varies as illustrated in Figure 5; note that instead of using same $\sigma_{i}$ for $(\mathcal{A}_i,\mathcal{C}_i)$ , we use two learnable $\sigma$ for $\mathcal{A}_i$ and $\mathcal{C}_i$ respectively. Quantitatively, when everything else is held equal, the convolution architecture shows better reconstruction than residual and attention architectures. Similarly, Figure 7 image 7 resembles the original image more than images 8 and 9 in terms of face radiance, smoothness, lip color, and shape. The reason is that residual connection enriches the outputs of FCM, which defies the purpose of enriching features of higher frequencies of the decoder features motivated in section 3.1.1. The attention mechanism is at a disadvantage here because the decoder and the encoder exclusively use convolutions. + +Finally, varying the size of kernels in DSL (Table 2 row 10 - 14) show quite similar quantitative reconstruction performance, while qualitatively, Figure 7 shows that a larger kernel size tends to produce smoother reconstructions. One reason could be that a larger kernel size tends to smooth more the feature maps because more surrounding values are used during convolution. Thus, in other experiments, we choose a kernel size of 3, and the effects of kernel sizes are open to future works. + +Zero-shot Reconstruction To further demonstrate the reconstruction capabilities of FA-VAE models, Table 3 gives the zero-shot reconstruction performance evaluated on CelebA-HQ [17] and FFHQ [21] using models trained on ImageNet [9]. The accompanying qualitative results are in Figure 8. Overall, FA-VAE displays impressive transferability capability with more faithful reconstruction in details, for instance, the light contrast in column 1 and book details in column 3. Note also that the common metric used in image compression PSNR is not perfectly correlated with rFID, but we put the metric here for reference. + +# 4.3. Image Synthesis + +Table 4 shows the text-to-image generation performance on CelebA-HQ-MM [17]. Our proposed autoregressive transformer CAT yields a better generation quality than other GAN-based models, including AttnGAN [52], ControlGAN [26], which are solely designed for image generation. Figure 9 shows that CAT can generate satisfactory images conditioned on text inputs of varying lengths. At- + +
CelebAFFHQ
Pretrained ModelfrFID ↓PSNR ↑rFID ↓PSNR ↑
VQ-GAN [40]168.6223.406.8322.68
VQ-GAN [11]44.1131.203.0630.82
FA-VAE (Ours)166.5222.596.1921.95
FA-VAE (Ours)42.2531.391.4630.85
+ +Table 3. Zero-shot reconstruction results on the validation data of FFHQ [21] and CelebA-HQ [17] using models trained on ImageNet [9]. $f$ is the downsampling factor, the codebook sizes for $f = \{ {16},8,4\}$ are $\{ {16384},{16384},{8192}\}$ respectively. + +
ModelFID ↓
AttnGAN [52]125.98
ControlGAN [26]116.32
DM-GAN [55]131.05
DF-GAN [44]137.60
TediGAN [50]106.37
LAFITE [54]12.54
CAT (Ours)10.23
+ +Table 4. Text-to-image generation on CelebA-HQ MM [50]. tributes such as "mouth open" and "arched eyebrows" are captured during generation because all the tokens embeddings are used as a condition which gives more precise guidance. More quantitative and qualitative results on different datasets are in the supplement. + +# 5. Conclusion + +In this paper, we introduce the Frequency Augmented VAE (FA-VAE) model, which aims to improve reconstruction quality by bridging the gaps in the frequency domains between original and reconstructed images. New modules named Frequency Complement Modules (FCM) are crafted and guided under the new (Dynamic) Spectrum Loss ((D)SL) to learn to complement the reconstructed features of missing frequencies. A new Cross-attention Autoregressive Transformer (CAT) is proposed for achieving more precise textual-image alignment in the text-to-image generation task. FA-VAE shows improved reconstruction on various datasets compared to SOTA methods. + +# Acknowledgement + +This research is supported in part by the National Science Foundation under award No. 2040209. + +# References + +[1] Jyoti Aneja, Alexander G. Schwing, Jan Kautz, and Arash Vahdat. Ncp-vae: Variational autoencoders with noise contrastive priors. *ArXiv*, abs/2010.02917, 2020. 2 +[2] Valerio Biscione and Jeffrey S. Bowers. Convolutional neural networks are not invariant to translation, but they can learn to be. JMLR, 2022. 5 +[3] Andrew Brock, Jeff Donahue, and Karen Simonyan. Large scale GAN training for high fidelity natural image synthesis. In ICLR, 2019. 2 +[4] Huiwen Chang, Han Zhang, Lu Jiang, Ce Liu, and William T. Freeman. Maskgit: Masked generative image transformer. In CVPR, 2022. 1 +[5] Ting Chen, Ruixiang Zhang, and Geoffrey Hinton. Analog bits: Generating discrete data using diffusion models with self-conditioning, 2022. 3 +[6] Rewon Child. Very deep $\{\mathrm{vae}\}$ s generalize autoregressive models and can outperform them on images. In ICLR, 2021. 1, 2 +[7] Jooyoung Choi, Jungbeom Lee, Chaehun Shin, Sungwon Kim, Hyunwoo Kim, and Sungroh Yoon. Perception prioritized training of diffusion models. In CVPR, 2022. 3 +[8] Katherine Crowson, Stella Rose Biderman, Daniel Kornis, Dashiell Stander, Eric Hallahan, Louis Castricato, and Edward Raff. Vqgan-clip: Open domain image generation and editing with natural language guidance. In ECCV, 2022. 2, 3, 6 +[9] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In CVPR, 2009. 6, 8 +[10] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. In NeurIPS, 2021. 3 +[11] Patrick Esser, Robin Rombach, and Björn Ommer. Taming transformers for high-resolution image synthesis. In CVPR, 2021. 1, 2, 3, 4, 5, 6, 7, 8 +[12] Ge Gao, Pei You, Rong Pan, Shunyuan Han, Yuanyuan Zhang, Yuchao Dai, and Hojae Lee. Neural image compression via attentional multi-scale back projection and frequency decomposition. In ICCV, 2021. 2 +[13] Karol Gregor, Ivo Danihelka, Andriy Mnih, Charles Blundell, and Daan Wierstra. Deep autoregressive networks. In ICML, 2014. 2 +[14] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. In NeurIPS, 2017. 1 +[15] Liming Jiang, Bo Dai, Wayne Wu, and Chen Change Loy. Focal frequency loss for image reconstruction and synthesis. In ICCV, 2021. 2, 4, 7 +[16] Justin Johnson, Alexandre Alahi, and Li Fei-Fei. Perceptual losses for real-time style transfer and super-resolution. In ECCV, 2016. 1, 2 +[17] Tero Karras, Timo Aila, Samuli Laine, and Jaakko Lehtinen. Progressive growing of gans for improved quality, stability, and variation. CoRR, abs/1710.10196, 2017. 7, 8 + +[18] Tero Karras, Timo Aila, Samuli Laine, and Jaakko Lehtinen. Progressive growing of GANs for improved quality, stability, and variation. In ICLR, 2018. 2 +[19] Tero Karras, Miika Aittala, Janne Hellsten, Samuli Laine, Jaakko Lehtinen, and Timo Aila. Training generative adversarial networks with limited data. In NeurIPS, 2020. 2 +[20] Tero Karras, Miika Aittala, Samuli Laine, Erik Härkönen, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Alias-free generative adversarial networks. In NeurIPS, 2021. 2 +[21] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In CVPR, 2019. 2, 6, 7, 8 +[22] Soo Ye Kim, Kfir Aberman, Nori Kanazawa, Rahul Garg, Neal Wadhwa, Huiwen Chang, Nikhil Karnad, Munchurl Kim, and Orly Liba. Zoom-to-inpaint: Image inpainting with high frequency details. In CVPRW, 2022. 1 +[23] Diederik P Kingma and Max Welling. Auto-encoding variational bayes, 2013. 2 +[24] Umut Kocasari, Alara Dirik, Mert Tiftikci, and Pinar Yanardag. Stylemc: Multi-channel based fast text-guided image generation and manipulation. 2022 IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 3441-3450, 2022. 2 +[25] Doyup Lee, Chiheon Kim, Saehoon Kim, Minsu Cho, and Wook-Shin Han. Autoregressive image generation using residual quantization. In CVPR, 2022. 1, 2, 3, 4, 6 +[26] Bowen Li, Xiaojuan Qi, Thomas Lukasiewicz, and Philip H. S. Torr. Controllable text-to-image generation. In NeurIPS, 2019. 8 +[27] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. GLIDE: towards photorealistic image generation and editing with text-guided diffusion models. In ICML, 2022. 2, 3 +[28] Alexander Quinn Nichol and Prafulla Dhariwal. Improved denoising diffusion probabilistic models. In ICML, 2021. 3 +[29] Gaurav Parmar, Dacheng Li, Kwonjoon Lee, and Zhuowen Tu. Dual contradistinctive generative autoencoder. In CVPR, 2021. 1, 2 +[30] Yash Patel, Srikar Appalaraju, and R. Manmatha. Deep perceptual compression, 2019. 5 +[31] Jialun Peng, Dong Liu, Songcen Xu, and Houqiang Li. Generating diverse structure for image inpainting with hierarchical vq-vae. In CVPR, 2021. 1 +[32] Stanislav Pidhorskyi, Donald A. Adjeroh, and Gianfranco Doretto. Adversarial latent autoencoders. In CVPR, 2020. 2 +[33] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In ICML, 2021. 2, 3, 6 +[34] Alec Radford, Jeff Wu, Rewon Child, David Luan, Dario Amodei, and Ilya Sutskever. Language models are unsupervised multitask learners. 2019. 6 + +[35] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J. Liu. Exploring the limits of transfer learning with a unified text-to-text transformer. JMLR, 21(140):1-67, 2020. 3 +[36] Nasim Rahaman, Aristide Baratin, Devansh Arpit, Felix Dräxler, Min Lin, Fred A. Hamprecht, Yoshua Bengio, and Aaron C. Courville. On the spectral bias of neural networks. In ICML, 2019. 1 +[37] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. ArXiv, abs/2204.06125, 2022. 3 +[38] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In ICML, 2021. 3, 6 +[39] Ali Razavi, Aaron van den Oord, and Oriol Vinyals. Generating diverse high-fidelity images with vq-vae-2. In NeurIPS, 2019. 1, 2, 3, 6 +[40] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 1, 3, 6, 8 +[41] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L. Denton, Seyed Kamyar Seyed Ghasemipour, Burcu Karagol Ayan, Seyedeh Sara Mahdavi, Raphael Gontijo Lopes, Tim Salimans, Jonathan Ho, David J. Fleet, and Mohammad Norouzi. Photorealistic text-to-image diffusion models with deep language understanding. ArXiv, abs/2205.11487, 2022. 3 +[42] Edgar Schonfeld, Bernt Schiele, and Anna Khoreva. A u-net based discriminator for generative adversarial networks. In CVPR, 2020. 2 +[43] Jianxin Sun, Qiyao Deng, Qi Li, Muyi Sun, Min Ren, and Zhenan Sun. Anyface: Free-style text-to-face synthesis and manipulation. In CVPR, 2022. 2 +[44] Ming Tao, Hao Tang, Fei Wu, Xiao-Yuan Jing, Bing-Kun Bao, and Changsheng Xu. Df-gan: A simple and effective baseline for text-to-image synthesis. In CVPR, 2022. 2, 8 +[45] Arash Vahdat and Jan Kautz. NVAE: A deep hierarchical variational autoencoder. In NeurIPS, 2020. 2 +[46] Aaron van den Oord, Oriol Vinyals, and koray kavukcuoglu. Neural discrete representation learning. In NeurIPS, 2017. 1, 2, 3, 4, 5, 6 +[47] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In NeurIPS, 2017. 6 +[48] Haohan Wang, Xindi Wu, Zeyi Huang, and Eric P. Xing. High-frequency component helps explain the generalization of convolutional neural networks. In CVPR, 2020. 2 +[49] Zihao Wang, Wei Liu, Qian He, Xinglong Wu, and Zili Yi. Clip-gen: Language-free training of a text-to-image generator with clip. arXiv preprint arXiv:2203.00386, 2022. 1, 2, 3, 6 +[50] Weihao Xia, Yujiu Yang, Jing-Hao Xue, and Baoyuan Wu. Tedigan: Text-guided diverse face image generation and manipulation. In CVPR, 2021. 2, 6, 8 + +[51] Zhisheng Xiao, Karsten Kreis, Jan Kautz, and Arash Vahdat. Vaebm: A symbiosis between variational autoencoders and energy-based models. In ICLR, 2021. 2 +[52] Tao Xu, Pengchuan Zhang, Qiuyuan Huang, Han Zhang, Zhe Gan, Xiaolei Huang, and Xiaodong He. Attngan: Fine-grained text to image generation with attentional generative adversarial networks. In CVPR, 2018. 2, 8 +[53] Jiahui Yu, Xin Li, Jing Yu Koh, Han Zhang, Ruoming Pang, James Qin, Alexander Ku, Yuanzhong Xu, Jason Baldridge, and Yonghui Wu. Vector-quantized image modeling with improved VQGAN. In ICLR, 2022. 5, 7 +[54] Yufan Zhou, Ruiyi Zhang, Changyou Chen, Chunyuan Li, Chris Tensmeyer, Tong Yu, Jiuming Gu, Jinhui Xu, and Tong Sun. Lafite: Towards language-free training for text-to-image generation. CVPR, 2022. 2, 8 +[55] Minfeng Zhu, Pingbo Pan, Wei Chen, and Yi Yang. Dm-gan: Dynamic memory generative adversarial networks for text-to-image synthesis. In CVPR, 2019. 2, 8 \ No newline at end of file diff --git a/2023/Catch Missing Details_ Image Reconstruction With Frequency Augmented Variational Autoencoder/images.zip b/2023/Catch Missing Details_ Image Reconstruction With Frequency Augmented Variational Autoencoder/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..0d43dba37497eb8793a36caa60ef088a7c1a2907 --- /dev/null +++ b/2023/Catch Missing Details_ Image Reconstruction With Frequency Augmented Variational Autoencoder/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e6696213c58ab7cbb04f53389ec1de001392f3dc424b1a14984135892d27a44b +size 810638 diff --git a/2023/Catch Missing Details_ Image Reconstruction With Frequency Augmented Variational Autoencoder/layout.json b/2023/Catch Missing Details_ Image Reconstruction With Frequency Augmented Variational Autoencoder/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..72d898c2eea8d691a889a01ea11ff5d8a3130f1c --- /dev/null +++ b/2023/Catch Missing Details_ Image Reconstruction With Frequency Augmented Variational Autoencoder/layout.json @@ -0,0 +1,9595 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 70, + 102, + 523, + 138 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 102, + 523, + 138 + ], + "spans": [ + { + "bbox": [ + 70, + 102, + 523, + 138 + ], + "type": "text", + "content": "Catch Missing Details: Image Reconstruction with Frequency Augmented Variational Autoencoder" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 135, + 162, + 201, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 162, + 201, + 175 + ], + "spans": [ + { + "bbox": [ + 135, + 162, + 201, + 175 + ], + "type": "text", + "content": "Xinmiao Lin" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 85, + 176, + 250, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 85, + 176, + 250, + 190 + ], + "spans": [ + { + "bbox": [ + 85, + 176, + 250, + 190 + ], + "type": "text", + "content": "Rochester Institute of Technology" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 129, + 192, + 208, + 202 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 129, + 192, + 208, + 202 + ], + "spans": [ + { + "bbox": [ + 129, + 192, + 208, + 202 + ], + "type": "text", + "content": "x13439@rit.edu" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 180, + 209, + 244, + 222 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 180, + 209, + 244, + 222 + ], + "spans": [ + { + "bbox": [ + 180, + 209, + 244, + 222 + ], + "type": "text", + "content": "Chiuman Ho" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 164, + 224, + 261, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 164, + 224, + 261, + 236 + ], + "spans": [ + { + "bbox": [ + 164, + 224, + 261, + 236 + ], + "type": "text", + "content": "OPPO US Research" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 168, + 240, + 257, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 168, + 240, + 257, + 251 + ], + "spans": [ + { + "bbox": [ + 168, + 240, + 257, + 251 + ], + "type": "text", + "content": "chiuman@oppo.com" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 161, + 356, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 161, + 356, + 175 + ], + "spans": [ + { + "bbox": [ + 304, + 161, + 356, + 175 + ], + "type": "text", + "content": "Yikang Li" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 281, + 176, + 379, + 188 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 281, + 176, + 379, + 188 + ], + "spans": [ + { + "bbox": [ + 281, + 176, + 379, + 188 + ], + "type": "text", + "content": "OPPO US Research" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 277, + 192, + 382, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 277, + 192, + 382, + 203 + ], + "spans": [ + { + "bbox": [ + 277, + 192, + 382, + 203 + ], + "type": "text", + "content": "yikang.li1@oppo.com" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 425, + 162, + 492, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 425, + 162, + 492, + 175 + ], + "spans": [ + { + "bbox": [ + 425, + 162, + 492, + 175 + ], + "type": "text", + "content": "Jenhao Hsiao" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 410, + 176, + 507, + 188 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 410, + 176, + 507, + 188 + ], + "spans": [ + { + "bbox": [ + 410, + 176, + 507, + 188 + ], + "type": "text", + "content": "OPPO US Research" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 422, + 193, + 494, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 422, + 193, + 494, + 203 + ], + "spans": [ + { + "bbox": [ + 422, + 193, + 494, + 203 + ], + "type": "text", + "content": "mark@oppo.com" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 143, + 278, + 192, + 291 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 278, + 192, + 291 + ], + "spans": [ + { + "bbox": [ + 143, + 278, + 192, + 291 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 46, + 305, + 290, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 305, + 290, + 594 + ], + "spans": [ + { + "bbox": [ + 46, + 305, + 290, + 594 + ], + "type": "text", + "content": "The popular VQ-VAE models reconstruct images through learning a discrete codebook but suffer from a significant issue in the rapid quality degradation of image reconstruction as the compression rate rises. One major reason is that a higher compression rate induces more loss of visual signals on the higher frequency spectrum which reflect the details on pixel space. In this paper, a Frequency Complement Module (FCM) architecture is proposed to capture the missing frequency information for enhancing reconstruction quality. The FCM can be easily incorporated into the VQ-VAE structure, and we refer to the new model as Frequency Augmented VAE (FA-VAE). In addition, a Dynamic Spectrum Loss (DSL) is introduced to guide the FCMs to balance between various frequencies dynamically for optimal reconstruction. FA-VAE is further extended to the text-to-image synthesis task, and a Cross-attention Autoregressive Transformer (CAT) is proposed to obtain more precise semantic attributes in texts. Extensive reconstruction experiments with different compression rates are conducted on several benchmark datasets, and the results demonstrate that the proposed FA-VAE is able to restore more faithfully the details compared to SOTA methods. CAT also shows improved generation quality with better image-text semantic alignment." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 47, + 620, + 128, + 632 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 620, + 128, + 632 + ], + "spans": [ + { + "bbox": [ + 47, + 620, + 128, + 632 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 46, + 641, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 641, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 641, + 287, + 715 + ], + "type": "text", + "content": "VQ-VAE models [6,11,25,29,39,46] reconstruct images through learning a discrete codebook of latent embeddings. They gained wide popularity due to the scalable and versatile codebook, which can be broadly applied to many visual tasks such as image synthesis [11,49] and inpainting [4,31]. A higher compression rate is typically preferable in VQ-" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 301, + 209, + 430, + 238 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 301, + 209, + 430, + 238 + ], + "spans": [ + { + "bbox": [ + 301, + 209, + 430, + 238 + ], + "type": "text", + "content": "Yu Kong \nMichigan State University" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 325, + 240, + 403, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 240, + 403, + 251 + ], + "spans": [ + { + "bbox": [ + 325, + 240, + 403, + 251 + ], + "type": "text", + "content": "yukong@msu.edu" + } + ] + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 307, + 279, + 545, + 399 + ], + "blocks": [ + { + "bbox": [ + 307, + 279, + 545, + 399 + ], + "lines": [ + { + "bbox": [ + 307, + 279, + 545, + 399 + ], + "spans": [ + { + "bbox": [ + 307, + 279, + 545, + 399 + ], + "type": "image", + "image_path": "47894a809973432e07174ef8fa48bbe6243ccdecc6c21b8ad23d4aa1df795a9e.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 401, + 546, + 510 + ], + "lines": [ + { + "bbox": [ + 304, + 401, + 546, + 510 + ], + "spans": [ + { + "bbox": [ + 304, + 401, + 546, + 510 + ], + "type": "text", + "content": "Figure 1. Images and their frequency maps. Row 1: original and reconstructed images. Row 2: the frequency maps of images, frequency increases in any direction away from the center. " + }, + { + "bbox": [ + 304, + 401, + 546, + 510 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 304, + 401, + 546, + 510 + ], + "type": "text", + "content": " is the compression rate. With a greater compression rate, more details are lost during reconstruction, i.e. eyes and mouth shape, and hair texture (pointed with red arrows) which align with the loss of high-frequency features. All frequency figures in this paper use the same colormap. rFID [14] and lpips [16] are lower the better, and frequency values increase from red to green, zoom in for better visualization." + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + } + ], + "index": 21 + }, + { + "bbox": [ + 304, + 517, + 546, + 541 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 517, + 546, + 541 + ], + "spans": [ + { + "bbox": [ + 304, + 517, + 546, + 541 + ], + "type": "text", + "content": "VAE models since it provides memory efficiency and better learning of coherent semantics structures [11, 40]." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 304, + 544, + 546, + 663 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 544, + 546, + 663 + ], + "spans": [ + { + "bbox": [ + 304, + 544, + 546, + 663 + ], + "type": "text", + "content": "One main challenge quickly arises for a higher compression rate, which severely compromises reconstruction accuracy. Figure 1 row 1 shows that although the reconstructed images at higher compression rates appear consistent with the original image, details inconsistencies such as the color and contour of the lips become apparent upon closer scrutiny. Figure 1 row 2 reveals that similar degradation also manifests on the frequency domains where features towards the middle and higher frequency spectrum are the least recoverable with greater compression rate." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 304, + 665, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 547, + 715 + ], + "type": "text", + "content": "Several causes stand behind this gap between pixel and frequency space. The convolutional nature of autoencoders is prone to spectral bias, which favors learning low-frequency features [22, 36]. This challenge is further ag-" + } + ] + } + ], + "index": 25 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1736" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 192 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 192 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 192 + ], + "type": "text", + "content": "gravated when current methods exclusively design losses or improve model architecture for better semantics resemblance [11, 16, 25] but often neglect the alignment on the frequency domain [12, 15]. On top of that, it is intuitively more challenging for a decoder to reconstruct an image patch from a single codebook embedding (high compression) than multiple embeddings (less compression). The reason is that the former mixes up features of incomplete and diverse frequencies, while the latter could preserve more fine-grained and complete features at various frequencies." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 194, + 289, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 194, + 289, + 289 + ], + "spans": [ + { + "bbox": [ + 47, + 194, + 289, + 289 + ], + "type": "text", + "content": "Inspired by these insights, the Frequency Augmented VAE (FA-VAE) model is proposed, which aims to improve reconstruction quality by achieving better alignment on the frequency spectrums between the original and reconstructed images. More specifically, new modules named Frequency Complement Modules (FCM) are crafted and embedded at multiple layers of FA-VAE's decoder to learn to complement the decoder's features with missing frequencies." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 291, + 289, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 291, + 289, + 411 + ], + "spans": [ + { + "bbox": [ + 46, + 291, + 289, + 411 + ], + "type": "text", + "content": "We observe that valuable middle and high frequencies are mingled with the encoder's feature maps during the compression via an encoder, shown in Figure 3 row 4. Therefore, a new loss termed Spectrum Loss (SL) is proposed to guide FCMs to generate missing features that align with the same level's encoder features on the frequency domain. Since most image semantics reside on the low-frequency spectrum [48], SL prioritizes learning lower-frequency features with diminishing weights as frequencies increase." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 412, + 289, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 412, + 289, + 568 + ], + "spans": [ + { + "bbox": [ + 46, + 412, + 289, + 568 + ], + "type": "text", + "content": "Interestingly, we discover that checkerboard patterns appear in the complemented decoder's features with SL, although better reconstruction performance is achieved (Figure 3 column 4). We speculate that because SL sets a deterministic range for the low-frequency spectrum when applying weights on the frequencies without considering that the importance of a frequency can vary from layer to layer. Thus, an improved loss function Dynamic Spectrum Loss (DSL) is crafted on top of SL with a learnable component to adjust the range of low-frequency spectrum dynamically for optimal reconstruction. DSL can improve reconstruction quality even further than SL without the unnatural checkerboard artifacts in the features (Figure 3 column 5)." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 570, + 289, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 570, + 289, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 570, + 289, + 714 + ], + "type": "text", + "content": "We further extend FA-VAE to the text-to-image generation task and propose the Cross-attention Autoregressive Transformer (CAT) model. We first observe that only using one or a few token embeddings is a coarse representation of lengthy texts [8, 27, 33]. Thus CAT uses all token embeddings as a condition for more precise guidance. Moreover, existing works typically use self-attention, and the text condition is embedded merely at the beginning of the generation [11, 49]. This mechanism becomes problematic in the autoregressive generation because one image token is generated at a time, thus the text condition gradually loosens its connection with the generated tokens. To circumvent this" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 305, + 72, + 545, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 545, + 96 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 545, + 96 + ], + "type": "text", + "content": "issue, CAT embeds a cross-attention mechanism that allows the text condition to guide each step generation." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 305, + 96, + 545, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 96, + 545, + 120 + ], + "spans": [ + { + "bbox": [ + 305, + 96, + 545, + 120 + ], + "type": "text", + "content": "To summarize, our work includes the following contributions:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 317, + 120, + 545, + 299 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 317, + 120, + 545, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 120, + 545, + 167 + ], + "spans": [ + { + "bbox": [ + 317, + 120, + 545, + 167 + ], + "type": "text", + "content": "- We propose a new type of architecture called Frequency Augmented VAE (FA-VAE) for improving image reconstruction through achieving more accurate details reconstruction." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 317, + 168, + 545, + 238 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 168, + 545, + 238 + ], + "spans": [ + { + "bbox": [ + 317, + 168, + 545, + 238 + ], + "type": "text", + "content": "- We propose a new loss called Spectrum Loss (SL) and its enhanced version Dynamic Spectrum Loss (D SL), which guides the Frequency Complement Modules (FCM) in FA-VAE to adaptively learn different low/high frequency mixtures for optimal reconstruction." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 317, + 239, + 545, + 299 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 239, + 545, + 299 + ], + "spans": [ + { + "bbox": [ + 317, + 239, + 545, + 299 + ], + "type": "text", + "content": "- We propose a new Cross-attention Autoregressive Transformer (CAT) for text-to-image generation using more fine-grained textual embeddings as a condition with a cross-attention mechanism for better image-text semantic alignment." + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 310, + 471, + 324 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 310, + 471, + 324 + ], + "spans": [ + { + "bbox": [ + 306, + 310, + 471, + 324 + ], + "type": "text", + "content": "2. Related Work & Background" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 331, + 545, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 331, + 545, + 533 + ], + "spans": [ + { + "bbox": [ + 304, + 331, + 545, + 533 + ], + "type": "text", + "content": "Image Reconstruction Vector Quantized-Variational AutoEncoder (VQ-VAE) [46] extends the Variational AutoEncoder structure [23] and proposes to encode images into discrete latent codes using vector quantization (VQ). Then a generative model, such as an autoregressive transformer [13], can be trained and paired with the decoder in VQ-VAE to synthesize new images. Later works [11,25,39] further improve the reconstruction and generation quality by improving the generative model architecture or the quantization efficiency. Since VQ-VAE models operate on discrete latent spaces, they cannot be directly and fairly compared to other VAE-based models that employ continuous latent space [1,6,23,29,45,51]. In contrast to current VQ-VAE models, our proposed FA-VAE technique improves reconstruction through the frequency angle and is more generalized since the FCMs can be readily extended to other neural networks that share the same VQ-VAE structure." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 534, + 545, + 630 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 534, + 545, + 630 + ], + "spans": [ + { + "bbox": [ + 304, + 534, + 545, + 630 + ], + "type": "text", + "content": "Few other works do image reconstruction via the frequency perspective. FFL [15] proposed a loss function to penalize differences on the hard frequencies. A new module, which is designed in the continuous latent space for image compression, is proposed to be incorporated into the encoder and decoder in [12]. To our best knowledge, FA-VAE is the first work that aims to improve image reconstruction on discrete latent space through the frequency perspective." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "content": "Image Generation Image generation can be achieved via GAN-based models [3, 18, 42, 44, 52, 55], which synthesize images from noise vectors unconditionally or conditioned on different inputs such as texts, masks, etc. It becomes cumbersome to train one model for one application. Thus, StyleGAN [21] encodes semantic attributes into a continuous latent space, and subsequent works [19, 20," + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1737" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 47, + 70, + 261, + 213 + ], + "blocks": [ + { + "bbox": [ + 47, + 70, + 261, + 213 + ], + "lines": [ + { + "bbox": [ + 47, + 70, + 261, + 213 + ], + "spans": [ + { + "bbox": [ + 47, + 70, + 261, + 213 + ], + "type": "image", + "image_path": "54efa40dc53ecdaf9a16925bc2ca99f28be9dba84c340692f3429da4f3c8bce4.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 45, + 218, + 546, + 251 + ], + "lines": [ + { + "bbox": [ + 45, + 218, + 546, + 251 + ], + "spans": [ + { + "bbox": [ + 45, + 218, + 546, + 251 + ], + "type": "text", + "content": "Figure 2. Frequency Augmented VAE (FA-VAE). (a) The encoder " + }, + { + "bbox": [ + 45, + 218, + 546, + 251 + ], + "type": "inline_equation", + "content": "\\mathcal{E}" + }, + { + "bbox": [ + 45, + 218, + 546, + 251 + ], + "type": "text", + "content": " encodes the images " + }, + { + "bbox": [ + 45, + 218, + 546, + 251 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 45, + 218, + 546, + 251 + ], + "type": "text", + "content": " onto discrete latent codebook space " + }, + { + "bbox": [ + 45, + 218, + 546, + 251 + ], + "type": "inline_equation", + "content": "\\hat{z}_q" + }, + { + "bbox": [ + 45, + 218, + 546, + 251 + ], + "type": "text", + "content": " which is used by the decoder " + }, + { + "bbox": [ + 45, + 218, + 546, + 251 + ], + "type": "inline_equation", + "content": "\\mathcal{G}" + }, + { + "bbox": [ + 45, + 218, + 546, + 251 + ], + "type": "text", + "content": " to reconstruct images " + }, + { + "bbox": [ + 45, + 218, + 546, + 251 + ], + "type": "inline_equation", + "content": "\\hat{X}" + }, + { + "bbox": [ + 45, + 218, + 546, + 251 + ], + "type": "text", + "content": ". DSL (b) guides the FCMs (c) to learn to complement the reconstructed features with missing features of important frequencies in order to improve reconstruction quality. A more detailed figure is in the supplement." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 261, + 70, + 545, + 213 + ], + "blocks": [ + { + "bbox": [ + 261, + 70, + 545, + 213 + ], + "lines": [ + { + "bbox": [ + 261, + 70, + 545, + 213 + ], + "spans": [ + { + "bbox": [ + 261, + 70, + 545, + 213 + ], + "type": "image", + "image_path": "b8d3e2c0f9326268fe96599a321282dea36c00873d5856951720531629ea190d.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 255, + 289, + 446 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 255, + 289, + 446 + ], + "spans": [ + { + "bbox": [ + 46, + 255, + 289, + 446 + ], + "type": "text", + "content": "24, 32, 43, 50, 54] leverage this space for generating images conditioned on attributes or textual descriptions. However, StyleGAN-based models cannot scale to large datasets when the number of attributes becomes substantially large because this demands to increase model's size. In contrast, the codebook in VQ-VAE models [46] is scalable to large datasets without additional model complexity. Diffusion-based models [5, 7, 10, 27, 28, 37, 40, 41] generate images from Gaussian noise through a reverse diffusion process can often require substantial training and large datasets. Most text-to-image generation models commonly use one of a few embeddings for a condition [8, 27, 38, 49] from pretrained models such as CLIP [33] or T5-XXL [35]. In contrast, our proposed autoregressive transformer CAT uses the embeddings from all the text tokens for more fine-grained guidance during image generation." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 447, + 288, + 542 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 447, + 288, + 542 + ], + "spans": [ + { + "bbox": [ + 46, + 447, + 288, + 542 + ], + "type": "text", + "content": "VQ-VAE We now describe the VQ-VAE model [46] as it is the backbone used in the FA-VAE model, which is presented in Figure 2. VQ-VAE [46] and related models [11, 25, 38, 39] consist of an encoder that encodes the images into a codebook of embeddings. Then, a decoder is trained to reconstruct the images from a discrete set of codebook embeddings. In this paper, reconstruction loss in VQ-GAN is utilized which is:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 91, + 550, + 287, + 566 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 550, + 287, + 566 + ], + "spans": [ + { + "bbox": [ + 91, + 550, + 287, + 566 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {r e c} = \\left\\| X - \\hat {X} \\right\\| _ {1} + \\mathcal {L} _ {p i p s} (X - \\hat {X}) \\tag {1}", + "image_path": "2b731650fd5f6fb086a137fcc91fcec7153ab5abce2ed315f9f83cf5cdd75321.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 574, + 287, + 623 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 574, + 287, + 623 + ], + "spans": [ + { + "bbox": [ + 47, + 574, + 287, + 623 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 574, + 287, + 623 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 47, + 574, + 287, + 623 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 574, + 287, + 623 + ], + "type": "inline_equation", + "content": "\\hat{X}" + }, + { + "bbox": [ + 47, + 574, + 287, + 623 + ], + "type": "text", + "content": " are the original and reconstructed images. We also use the adversarial setting of VQ-GAN model which introduced a discriminator compared to VQ-VAE models [39, 46], more details could be referred to [11]." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 634, + 129, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 634, + 129, + 647 + ], + "spans": [ + { + "bbox": [ + 47, + 634, + 129, + 647 + ], + "type": "text", + "content": "3. Methodology" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 653, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 653, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 653, + 288, + 714 + ], + "type": "text", + "content": "The proposed Frequency Augmented VAE (FA-VAE) is presented in Figure 2. In this section, we describe how FA-VAE ameliorates reconstruction quality by bridging the spectrum domain gap between reconstructed and original images. By explicitly embedding Frequency Complement" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 255, + 547, + 315 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 255, + 547, + 315 + ], + "spans": [ + { + "bbox": [ + 304, + 255, + 547, + 315 + ], + "type": "text", + "content": "Modules (FCM) into the decoder, Dynamic Spectrum Loss (DSL) leverages important features frequency-wise from the encoder and guides the FCMs to complement the features of the decoder for better frequency restoration at different reconstruction stages." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 325, + 508, + 338 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 325, + 508, + 338 + ], + "spans": [ + { + "bbox": [ + 305, + 325, + 508, + 338 + ], + "type": "text", + "content": "3.1. Frequency Augmented VAE (FA-VAE)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 343, + 547, + 464 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 343, + 547, + 464 + ], + "spans": [ + { + "bbox": [ + 304, + 343, + 547, + 464 + ], + "type": "text", + "content": "Let the images be " + }, + { + "bbox": [ + 304, + 343, + 547, + 464 + ], + "type": "inline_equation", + "content": "X \\in \\mathbb{R}^{H \\times W \\times 3}" + }, + { + "bbox": [ + 304, + 343, + 547, + 464 + ], + "type": "text", + "content": ". The codebook " + }, + { + "bbox": [ + 304, + 343, + 547, + 464 + ], + "type": "inline_equation", + "content": "\\mathfrak{C}" + }, + { + "bbox": [ + 304, + 343, + 547, + 464 + ], + "type": "text", + "content": " is a set of " + }, + { + "bbox": [ + 304, + 343, + 547, + 464 + ], + "type": "inline_equation", + "content": "|\\mathfrak{C}|" + }, + { + "bbox": [ + 304, + 343, + 547, + 464 + ], + "type": "text", + "content": " embeddings, such as " + }, + { + "bbox": [ + 304, + 343, + 547, + 464 + ], + "type": "inline_equation", + "content": "\\mathfrak{C} = \\{\\mathfrak{c}_i | i = 1, \\dots, |\\mathfrak{c}|\\} \\in \\mathbb{R}^{n_z}" + }, + { + "bbox": [ + 304, + 343, + 547, + 464 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 343, + 547, + 464 + ], + "type": "inline_equation", + "content": "n_z" + }, + { + "bbox": [ + 304, + 343, + 547, + 464 + ], + "type": "text", + "content": " is the length of one codebook embedding. FA-VAE consists of an encoder " + }, + { + "bbox": [ + 304, + 343, + 547, + 464 + ], + "type": "inline_equation", + "content": "\\mathcal{E}" + }, + { + "bbox": [ + 304, + 343, + 547, + 464 + ], + "type": "text", + "content": " that encodes the images " + }, + { + "bbox": [ + 304, + 343, + 547, + 464 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 304, + 343, + 547, + 464 + ], + "type": "text", + "content": " into latent space representations such as " + }, + { + "bbox": [ + 304, + 343, + 547, + 464 + ], + "type": "inline_equation", + "content": "z = \\mathcal{E}(x)" + }, + { + "bbox": [ + 304, + 343, + 547, + 464 + ], + "type": "text", + "content": ", for " + }, + { + "bbox": [ + 304, + 343, + 547, + 464 + ], + "type": "inline_equation", + "content": "z \\in \\mathbb{R}^{h \\times w \\times n_z}" + }, + { + "bbox": [ + 304, + 343, + 547, + 464 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 343, + 547, + 464 + ], + "type": "inline_equation", + "content": "x \\in X" + }, + { + "bbox": [ + 304, + 343, + 547, + 464 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 343, + 547, + 464 + ], + "type": "inline_equation", + "content": "(h \\times w)" + }, + { + "bbox": [ + 304, + 343, + 547, + 464 + ], + "type": "text", + "content": " is the resolution of the encoded representation. Let " + }, + { + "bbox": [ + 304, + 343, + 547, + 464 + ], + "type": "inline_equation", + "content": "f = H / h = W / w" + }, + { + "bbox": [ + 304, + 343, + 547, + 464 + ], + "type": "text", + "content": " be the downsampling factor or compression rate. Each feature of " + }, + { + "bbox": [ + 304, + 343, + 547, + 464 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 304, + 343, + 547, + 464 + ], + "type": "text", + "content": " is approximated by the vector quantization block " + }, + { + "bbox": [ + 304, + 343, + 547, + 464 + ], + "type": "inline_equation", + "content": "\\mathcal{Q}" + }, + { + "bbox": [ + 304, + 343, + 547, + 464 + ], + "type": "text", + "content": " using the nearest codebook entries, and it can be presented as:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 358, + 475, + 545, + 495 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 358, + 475, + 545, + 495 + ], + "spans": [ + { + "bbox": [ + 358, + 475, + 545, + 495 + ], + "type": "interline_equation", + "content": "\\mathcal {Q} (z) = \\hat {z} _ {q} = \\underset {c _ {k} \\in \\mathcal {C}} {\\arg \\min } \\| z _ {i j} - \\mathfrak {c} _ {k} \\| \\tag {2}", + "image_path": "c8accb8d6048eb4dac0e18776def3197922218c4c4f3f098f07f71dc316c57bf.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 506, + 547, + 543 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 506, + 547, + 543 + ], + "spans": [ + { + "bbox": [ + 304, + 506, + 547, + 543 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 506, + 547, + 543 + ], + "type": "inline_equation", + "content": "\\hat{z}_q" + }, + { + "bbox": [ + 304, + 506, + 547, + 543 + ], + "type": "text", + "content": " is the quantized latent embedding and subsequently used by the decoder " + }, + { + "bbox": [ + 304, + 506, + 547, + 543 + ], + "type": "inline_equation", + "content": "\\mathcal{G}" + }, + { + "bbox": [ + 304, + 506, + 547, + 543 + ], + "type": "text", + "content": " to produce the reconstructed image " + }, + { + "bbox": [ + 304, + 506, + 547, + 543 + ], + "type": "inline_equation", + "content": "\\hat{x} = \\mathcal{G}(\\hat{z}_q)" + }, + { + "bbox": [ + 304, + 506, + 547, + 543 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 561, + 512, + 573 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 561, + 512, + 573 + ], + "spans": [ + { + "bbox": [ + 306, + 561, + 512, + 573 + ], + "type": "text", + "content": "3.1.1 Frequency Complement Modules (FCM)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "content": "Motivation Figure 1 shows that a higher compression rate leads to more significant reconstruction disparities on the higher-frequency spectrum. Existing models [11, 25, 39] guide the reconstructed images to be more aligned on the pixel and feature spaces with the original images (Eq. 1) but neglect frequency spectrum alignment and leave the encoder and decoder without further guidance. Figure 3 column 2 shows that the encoder activations " + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{A}_1" + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "content": " of baseline VQ-GAN [11] contain rich high-frequency features (row 3 & 4), but the decoder's activations could mainly restore low-frequency features (row 5 & 6)." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1738" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 289, + 145 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 289, + 145 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 289, + 145 + ], + "type": "text", + "content": "Therefore we propose Frequency Complement Modules (FCM), illustrated in Figure 2 (c), which aims to complement the decoder's feature maps " + }, + { + "bbox": [ + 47, + 72, + 289, + 145 + ], + "type": "inline_equation", + "content": "\\mathcal{B}_i" + }, + { + "bbox": [ + 47, + 72, + 289, + 145 + ], + "type": "text", + "content": " with features of missing frequencies using the encoder activations " + }, + { + "bbox": [ + 47, + 72, + 289, + 145 + ], + "type": "inline_equation", + "content": "\\mathcal{A}_i" + }, + { + "bbox": [ + 47, + 72, + 289, + 145 + ], + "type": "text", + "content": ". The FCMs " + }, + { + "bbox": [ + 47, + 72, + 289, + 145 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_i" + }, + { + "bbox": [ + 47, + 72, + 289, + 145 + ], + "type": "text", + "content": " consist of sequences of convolution layers and activations. The decoder " + }, + { + "bbox": [ + 47, + 72, + 289, + 145 + ], + "type": "inline_equation", + "content": "\\mathcal{G}" + }, + { + "bbox": [ + 47, + 72, + 289, + 145 + ], + "type": "text", + "content": " with FCMs embedded can be represented as:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 64, + 164, + 258, + 178 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 164, + 258, + 178 + ], + "spans": [ + { + "bbox": [ + 64, + 164, + 258, + 178 + ], + "type": "interline_equation", + "content": "\\mathcal {B} _ {M - 1} = \\mathcal {G} _ {M} \\left(\\mathcal {C} _ {M} + \\hat {z} _ {q}\\right) = \\mathcal {G} _ {M} \\left(\\mathcal {F} _ {M} (\\hat {z} _ {q}) + \\hat {z} _ {q}\\right)", + "image_path": "816e6670fede55e86ac4f6f0b046056e8a420baac5dfdd4451f6b1bb60801764.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 80, + 180, + 287, + 191 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 80, + 180, + 287, + 191 + ], + "spans": [ + { + "bbox": [ + 80, + 180, + 287, + 191 + ], + "type": "interline_equation", + "content": "\\dots \\tag {3}", + "image_path": "fa5372ad1f89c7f0dde87979602515d92bcac9906f3d1336fb21762f7632365e.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 80, + 194, + 242, + 207 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 80, + 194, + 242, + 207 + ], + "spans": [ + { + "bbox": [ + 80, + 194, + 242, + 207 + ], + "type": "interline_equation", + "content": "\\mathcal {B} _ {1} = \\mathcal {G} _ {2} \\left(\\mathcal {C} _ {2} + \\mathcal {B} _ {2}\\right) = \\mathcal {G} _ {2} \\left(\\mathcal {F} _ {2} \\left(\\mathcal {B} _ {2}\\right) + \\mathcal {B} _ {2}\\right)", + "image_path": "bd19f482e6af321f681db52957f4f5cde7266dac1f07c13f6cc2ac7884abdc95.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 217, + 232, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 217, + 232, + 228 + ], + "spans": [ + { + "bbox": [ + 47, + 217, + 232, + 228 + ], + "type": "text", + "content": "Similarly, the encoder " + }, + { + "bbox": [ + 47, + 217, + 232, + 228 + ], + "type": "inline_equation", + "content": "\\mathcal{E}" + }, + { + "bbox": [ + 47, + 217, + 232, + 228 + ], + "type": "text", + "content": " can be abstracted to:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 96, + 239, + 287, + 253 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 96, + 239, + 287, + 253 + ], + "spans": [ + { + "bbox": [ + 96, + 239, + 287, + 253 + ], + "type": "interline_equation", + "content": "z = \\mathcal {E} (x) = \\mathcal {E} _ {M} \\left(\\dots \\left(\\mathcal {E} _ {2} (\\mathcal {E} _ {1} (x)))\\right)\\right) \\tag {4}", + "image_path": "2a8dcdd4fbd1eafe0a3147e3c279cc3fc9d44ad3ecc778840d641030f1142a46.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 262, + 287, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 262, + 287, + 357 + ], + "spans": [ + { + "bbox": [ + 46, + 262, + 287, + 357 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 262, + 287, + 357 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_i" + }, + { + "bbox": [ + 46, + 262, + 287, + 357 + ], + "type": "text", + "content": " is the " + }, + { + "bbox": [ + 46, + 262, + 287, + 357 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 262, + 287, + 357 + ], + "type": "text", + "content": "-th layer of the decoder, and " + }, + { + "bbox": [ + 46, + 262, + 287, + 357 + ], + "type": "inline_equation", + "content": "\\mathcal{A}_i = \\mathcal{E}_i(x)" + }, + { + "bbox": [ + 46, + 262, + 287, + 357 + ], + "type": "text", + "content": " is the outputs or feature maps of the corresponding encoder block. The outputs of previous blocks " + }, + { + "bbox": [ + 46, + 262, + 287, + 357 + ], + "type": "inline_equation", + "content": "\\mathcal{B}_i" + }, + { + "bbox": [ + 46, + 262, + 287, + 357 + ], + "type": "text", + "content": " are complemented by outputs " + }, + { + "bbox": [ + 46, + 262, + 287, + 357 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_i = \\mathcal{F}_i(\\mathcal{B}_{i + 1})" + }, + { + "bbox": [ + 46, + 262, + 287, + 357 + ], + "type": "text", + "content": ", which contain the frequency-rich features learned from the encoder activations " + }, + { + "bbox": [ + 46, + 262, + 287, + 357 + ], + "type": "inline_equation", + "content": "\\mathcal{A}_i" + }, + { + "bbox": [ + 46, + 262, + 287, + 357 + ], + "type": "text", + "content": ". The following section describes how Dynamic Frequency Loss (DSL) guides the learning of FCMs to specifically learn to complement " + }, + { + "bbox": [ + 46, + 262, + 287, + 357 + ], + "type": "inline_equation", + "content": "\\mathcal{B}_i" + }, + { + "bbox": [ + 46, + 262, + 287, + 357 + ], + "type": "text", + "content": " with the features of missing frequencies." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 359, + 287, + 465 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 359, + 287, + 465 + ], + "spans": [ + { + "bbox": [ + 46, + 359, + 287, + 465 + ], + "type": "text", + "content": "In the paper, the frequency feature compensation is implemented by addition (in Eq. 3). However, the FCM is flexible in adopting any architecture and merging techniques, as illustrated by the three examples in Figure 5. Note that no architecture limitation is imposed on the encoder and decoder blocks as long as " + }, + { + "bbox": [ + 46, + 359, + 287, + 465 + ], + "type": "inline_equation", + "content": "\\mathcal{A}_i" + }, + { + "bbox": [ + 46, + 359, + 287, + 465 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 359, + 287, + 465 + ], + "type": "inline_equation", + "content": "\\mathcal{B}_i" + }, + { + "bbox": [ + 46, + 359, + 287, + 465 + ], + "type": "text", + "content": " share the same resolution. VQ-GAN [11] and related models [25,46] already have similar architecture, and VQ-GAN is chosen as the backbone for FA-VAE." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 483, + 213, + 496 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 483, + 213, + 496 + ], + "spans": [ + { + "bbox": [ + 47, + 483, + 213, + 496 + ], + "type": "text", + "content": "3.1.2 Dynamic Spectrum Loss (DSL)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 503, + 287, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 503, + 287, + 635 + ], + "spans": [ + { + "bbox": [ + 46, + 503, + 287, + 635 + ], + "type": "text", + "content": "Motivation Spectrum Loss (SL) is proposed to guide explicitly the outputs of FCMs " + }, + { + "bbox": [ + 46, + 503, + 287, + 635 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_i" + }, + { + "bbox": [ + 46, + 503, + 287, + 635 + ], + "type": "text", + "content": " to be more aligned with the encoder's activations " + }, + { + "bbox": [ + 46, + 503, + 287, + 635 + ], + "type": "inline_equation", + "content": "\\mathcal{A}_i" + }, + { + "bbox": [ + 46, + 503, + 287, + 635 + ], + "type": "text", + "content": " on the frequency spectrum because the latter contains rich features on the higher frequency spectrum (Figure 3). Moreover, to account for the varying importance of frequencies across different decoder stages, Dynamic Spectrum Loss (DSL) is proposed, a more generalized variant of SL, and has the ability to adjust weights put on higher-frequency spectrum adaptively. Thus, each decoder block's outputs can be enriched with features of the most critical frequencies for an accurate reconstruction." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 635, + 288, + 672 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 635, + 288, + 672 + ], + "spans": [ + { + "bbox": [ + 47, + 635, + 288, + 672 + ], + "type": "text", + "content": "Background The outputs " + }, + { + "bbox": [ + 47, + 635, + 288, + 672 + ], + "type": "inline_equation", + "content": "\\mathcal{A}_i" + }, + { + "bbox": [ + 47, + 635, + 288, + 672 + ], + "type": "text", + "content": " of encoder block " + }, + { + "bbox": [ + 47, + 635, + 288, + 672 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_i" + }, + { + "bbox": [ + 47, + 635, + 288, + 672 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 635, + 288, + 672 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_i" + }, + { + "bbox": [ + 47, + 635, + 288, + 672 + ], + "type": "text", + "content": " of FCM block " + }, + { + "bbox": [ + 47, + 635, + 288, + 672 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_i" + }, + { + "bbox": [ + 47, + 635, + 288, + 672 + ], + "type": "text", + "content": " are first transformed to the frequency domain using Discrete Fourier Transform (DFT) as follows:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 77, + 681, + 287, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 681, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 77, + 681, + 287, + 715 + ], + "type": "interline_equation", + "content": "F (u, v) = \\sum_ {x = 0} ^ {M - 1} \\sum_ {y = 0} ^ {N - 1} f (x, y) \\cdot e ^ {- i 2 \\pi \\left(\\frac {u x}{M} + \\frac {v y}{N}\\right)} \\tag {5}", + "image_path": "fc0b7b8bc4e30078059a3c433427431f2228c1dcbfe48a02b588ec5644a23ec6.jpg" + } + ] + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 317, + 72, + 540, + 337 + ], + "blocks": [ + { + "bbox": [ + 317, + 72, + 540, + 337 + ], + "lines": [ + { + "bbox": [ + 317, + 72, + 540, + 337 + ], + "spans": [ + { + "bbox": [ + 317, + 72, + 540, + 337 + ], + "type": "image", + "image_path": "b90b2d4f5584033841c23e4b3bf314d6aad03e39cb0d50246f860c60030de423.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 338, + 545, + 416 + ], + "lines": [ + { + "bbox": [ + 305, + 338, + 545, + 416 + ], + "spans": [ + { + "bbox": [ + 305, + 338, + 545, + 416 + ], + "type": "text", + "content": "Figure 3. Image, activations, and their frequency maps. Ours* is FA-VAE model with FFL [15], Ours‡ is FA-VAE with SL and Ours is FA-VAE model with DSL. DSL shows a more harmonious balance between low- and high-frequencies (row 5) and more accurate reconstruction in mouth and hair textures. The frequency maps are plotted using the average of all channels and the contrast is adjusted to emphasize the higher frequency spectrums." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 420, + 545, + 504 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 420, + 545, + 504 + ], + "spans": [ + { + "bbox": [ + 305, + 420, + 545, + 504 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 420, + 545, + 504 + ], + "type": "inline_equation", + "content": "e" + }, + { + "bbox": [ + 305, + 420, + 545, + 504 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 420, + 545, + 504 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 305, + 420, + 545, + 504 + ], + "type": "text", + "content": " are Euler's number and the imaginary unit. " + }, + { + "bbox": [ + 305, + 420, + 545, + 504 + ], + "type": "inline_equation", + "content": "M \\times N" + }, + { + "bbox": [ + 305, + 420, + 545, + 504 + ], + "type": "text", + "content": " is the spatial resolution of feature maps and the Fourier Transform (Eq. 5) is applied to each of them. " + }, + { + "bbox": [ + 305, + 420, + 545, + 504 + ], + "type": "inline_equation", + "content": "f(x,y)" + }, + { + "bbox": [ + 305, + 420, + 545, + 504 + ], + "type": "text", + "content": " is the value at " + }, + { + "bbox": [ + 305, + 420, + 545, + 504 + ], + "type": "inline_equation", + "content": "(x,y)" + }, + { + "bbox": [ + 305, + 420, + 545, + 504 + ], + "type": "text", + "content": " of each feature map. " + }, + { + "bbox": [ + 305, + 420, + 545, + 504 + ], + "type": "inline_equation", + "content": "F(u,v)" + }, + { + "bbox": [ + 305, + 420, + 545, + 504 + ], + "type": "text", + "content": " is the corresponding value at " + }, + { + "bbox": [ + 305, + 420, + 545, + 504 + ], + "type": "inline_equation", + "content": "(u,v)" + }, + { + "bbox": [ + 305, + 420, + 545, + 504 + ], + "type": "text", + "content": " coordinates on the frequency spectrum. The Focal Frequency Loss (FFL) [15] can be presented as:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 310, + 514, + 545, + 557 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 310, + 514, + 545, + 557 + ], + "spans": [ + { + "bbox": [ + 310, + 514, + 545, + 557 + ], + "type": "interline_equation", + "content": "\\operatorname {F F L} \\left(\\mathcal {A} _ {i}, \\mathcal {C} _ {i}\\right) = \\frac {1}{M N \\left| \\mathcal {C} _ {i} \\right|} \\sum_ {c = 0} ^ {\\left| \\mathcal {C} _ {i} \\right| - 1} \\sum_ {u = 0} ^ {M - 1} \\sum_ {v = 0} ^ {N - 1} w (u, v) J (u, v), \\tag {6}", + "image_path": "82df8de322766e440c72dad40c3587ba7d7e4030da68d972df08c5e81937b2dc.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 558, + 545, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 558, + 545, + 629 + ], + "spans": [ + { + "bbox": [ + 304, + 558, + 545, + 629 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 558, + 545, + 629 + ], + "type": "inline_equation", + "content": "w(u,v) = |F_{\\mathcal{A}_i}(u,v) - F_{\\mathcal{C}_i}(u,v)|" + }, + { + "bbox": [ + 304, + 558, + 545, + 629 + ], + "type": "text", + "content": " are the weights put on each frequency. " + }, + { + "bbox": [ + 304, + 558, + 545, + 629 + ], + "type": "inline_equation", + "content": "J(u,v) = |F_{\\mathcal{A}_i}(u,v) - F_{\\mathcal{C}_i}(u,v)|^2" + }, + { + "bbox": [ + 304, + 558, + 545, + 629 + ], + "type": "text", + "content": " is the main error function based on the frequency difference. The " + }, + { + "bbox": [ + 304, + 558, + 545, + 629 + ], + "type": "inline_equation", + "content": "|\\mathcal{C}_i|" + }, + { + "bbox": [ + 304, + 558, + 545, + 629 + ], + "type": "text", + "content": " is the number of feature maps in " + }, + { + "bbox": [ + 304, + 558, + 545, + 629 + ], + "type": "inline_equation", + "content": "\\mathcal{A}_i" + }, + { + "bbox": [ + 304, + 558, + 545, + 629 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 558, + 545, + 629 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_i" + }, + { + "bbox": [ + 304, + 558, + 545, + 629 + ], + "type": "text", + "content": ". Both real and imaginary parts of the frequency domain are considered, more details are in [15]." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 630, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 630, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 630, + 545, + 714 + ], + "type": "text", + "content": "Limitations of FFL Eq. 6 demonstrates that FFL puts modulating term " + }, + { + "bbox": [ + 304, + 630, + 545, + 714 + ], + "type": "inline_equation", + "content": "w(u, v)" + }, + { + "bbox": [ + 304, + 630, + 545, + 714 + ], + "type": "text", + "content": " to focus learning on the hardest frequencies for reconstruction, which are the higher frequencies following our observation (Figures 1 & 3). However, this could not be ideal because features on the lower frequencies define the image content, and overemphasizing the higher frequencies could over-constrain the learning" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1739" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 47, + 69, + 286, + 158 + ], + "blocks": [ + { + "bbox": [ + 47, + 69, + 286, + 158 + ], + "lines": [ + { + "bbox": [ + 47, + 69, + 286, + 158 + ], + "spans": [ + { + "bbox": [ + 47, + 69, + 286, + 158 + ], + "type": "image", + "image_path": "8a2e58a0498ae831d9b106743f2ee53ecdd60c1343d9197660e65a05a86863b1.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 159, + 279, + 171 + ], + "lines": [ + { + "bbox": [ + 55, + 159, + 279, + 171 + ], + "spans": [ + { + "bbox": [ + 55, + 159, + 279, + 171 + ], + "type": "text", + "content": "Figure 4. Cross-attention Autoregressive Transformer (CAT)." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 174, + 287, + 258 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 174, + 287, + 258 + ], + "spans": [ + { + "bbox": [ + 46, + 174, + 287, + 258 + ], + "type": "text", + "content": "and lead to sub-optimal reconstruction, see Figure 7 row 1 column 3 and Table 2 row 3. Moreover, Figure 3 column 3 shows FCMs guided by a simple FFL can improve reconstruction performance (Ours\\*). However the frequency maps of the decoder activations " + }, + { + "bbox": [ + 46, + 174, + 287, + 258 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_1" + }, + { + "bbox": [ + 46, + 174, + 287, + 258 + ], + "type": "text", + "content": " contain excessive noise due to overemphasis on the higher frequency spectrum, and lower frequencies are neglected (row 6)." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 258, + 288, + 343 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 258, + 288, + 343 + ], + "spans": [ + { + "bbox": [ + 47, + 258, + 288, + 343 + ], + "type": "text", + "content": "Spectrum Loss (SL) Thus, we propose to apply a low-pass filter on the weights " + }, + { + "bbox": [ + 47, + 258, + 288, + 343 + ], + "type": "inline_equation", + "content": "w(u,v)" + }, + { + "bbox": [ + 47, + 258, + 288, + 343 + ], + "type": "text", + "content": " in Eq. 6 to penalize more mismatch in the lower-frequency domain and gradually diminish the penalizing weights towards higher frequency spectrum. Therefore, let the Gaussian kernels be " + }, + { + "bbox": [ + 47, + 258, + 288, + 343 + ], + "type": "inline_equation", + "content": "K_{i}(\\mu ,\\sigma)" + }, + { + "bbox": [ + 47, + 258, + 288, + 343 + ], + "type": "text", + "content": " with weights initialized using mean and standard deviation " + }, + { + "bbox": [ + 47, + 258, + 288, + 343 + ], + "type": "inline_equation", + "content": "(\\mu ,\\sigma_{i})" + }, + { + "bbox": [ + 47, + 258, + 288, + 343 + ], + "type": "text", + "content": ", and applied over the feature maps as:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 77, + 349, + 287, + 365 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 349, + 287, + 365 + ], + "spans": [ + { + "bbox": [ + 77, + 349, + 287, + 365 + ], + "type": "interline_equation", + "content": "(\\hat {\\mathcal {A}} _ {i}, \\hat {\\mathcal {C}} _ {i}) = (K _ {i} (\\mu , \\sigma_ {i}) \\star \\mathcal {A} _ {i}, K _ {i} (\\mu , \\sigma_ {i}) \\star \\mathcal {C} _ {i}). \\qquad (7)", + "image_path": "34c728cd7361094e8fac03c03a60efb62d758e88f6311d7f1ea02ef7d71bdc9c.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 372, + 287, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 372, + 287, + 396 + ], + "spans": [ + { + "bbox": [ + 47, + 372, + 287, + 396 + ], + "type": "text", + "content": "where the " + }, + { + "bbox": [ + 47, + 372, + 287, + 396 + ], + "type": "inline_equation", + "content": "\\star" + }, + { + "bbox": [ + 47, + 372, + 287, + 396 + ], + "type": "text", + "content": " is the convolution operation. Then, the Spectrum Loss (SL) is defined as:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 110, + 403, + 287, + 419 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 110, + 403, + 287, + 419 + ], + "spans": [ + { + "bbox": [ + 110, + 403, + 287, + 419 + ], + "type": "interline_equation", + "content": "\\operatorname {S L} \\left(\\mathcal {A} _ {i}, \\mathcal {C} _ {i}\\right) = \\operatorname {F F L} \\left(\\hat {\\mathcal {A}} _ {i}, \\hat {\\mathcal {C}} _ {i}\\right). \\tag {8}", + "image_path": "3f7475a467cb4e6150b40a066c61ebdc82e3a502b5d157c5562a3c85f56e508c.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 426, + 288, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 426, + 288, + 653 + ], + "spans": [ + { + "bbox": [ + 46, + 426, + 288, + 653 + ], + "type": "text", + "content": "Limitations of SL Up until now, a fixed Gaussian filter " + }, + { + "bbox": [ + 46, + 426, + 288, + 653 + ], + "type": "inline_equation", + "content": "K_{i}(\\mu ,\\sigma_{i})" + }, + { + "bbox": [ + 46, + 426, + 288, + 653 + ], + "type": "text", + "content": " is applied over all " + }, + { + "bbox": [ + 46, + 426, + 288, + 653 + ], + "type": "inline_equation", + "content": "\\mathcal{A}_i" + }, + { + "bbox": [ + 46, + 426, + 288, + 653 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 426, + 288, + 653 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_i" + }, + { + "bbox": [ + 46, + 426, + 288, + 653 + ], + "type": "text", + "content": " of all encoder and decoder blocks. Figure 3 column 4 demonstrates that SL " + }, + { + "bbox": [ + 46, + 426, + 288, + 653 + ], + "type": "inline_equation", + "content": "(\\mathrm{Ours}^{\\ddagger})" + }, + { + "bbox": [ + 46, + 426, + 288, + 653 + ], + "type": "text", + "content": " improves reconstruction on the lower frequency spectrum, leading to better reconstruction than the baseline model (Table 2 row 5 vs row 1). However, checkerboard artifacts are also present (row 5) on " + }, + { + "bbox": [ + 46, + 426, + 288, + 653 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_1" + }, + { + "bbox": [ + 46, + 426, + 288, + 653 + ], + "type": "text", + "content": ". One reason is that deterministic variance " + }, + { + "bbox": [ + 46, + 426, + 288, + 653 + ], + "type": "inline_equation", + "content": "\\sigma_{i}" + }, + { + "bbox": [ + 46, + 426, + 288, + 653 + ], + "type": "text", + "content": " assumes that decoder activations across different levels require the same amount of higher frequency features for accurate reconstruction, which could be an over-rigid constraint on the learning. Another reason could be that the same " + }, + { + "bbox": [ + 46, + 426, + 288, + 653 + ], + "type": "inline_equation", + "content": "\\sigma_{i}" + }, + { + "bbox": [ + 46, + 426, + 288, + 653 + ], + "type": "text", + "content": " magnifies the checkerboard effects produced by upsampling or striding operations in CNNs which are normally circumvented by subsequent convolutional layers [2, 30]. In the observation of the experiments, we find that the checkerboard effects can be intensified with the same " + }, + { + "bbox": [ + 46, + 426, + 288, + 653 + ], + "type": "inline_equation", + "content": "\\sigma_{i}" + }, + { + "bbox": [ + 46, + 426, + 288, + 653 + ], + "type": "text", + "content": " in different layers due to the bounding effect between low frequency and high frequency (shown in Figure 3 column 4 row 6)." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 654, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 654, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 654, + 289, + 715 + ], + "type": "text", + "content": "Dynamic Spectrum Loss (DSL) Therefore, we propose to optimize the variances " + }, + { + "bbox": [ + 46, + 654, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\sigma_{i}" + }, + { + "bbox": [ + 46, + 654, + 289, + 715 + ], + "type": "text", + "content": " instead of setting them as static hyperparameters to suit the different amounts of frequencies needed for each block's " + }, + { + "bbox": [ + 46, + 654, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{B}_i" + }, + { + "bbox": [ + 46, + 654, + 289, + 715 + ], + "type": "text", + "content": ". The new Dynamic Spectrum Loss (DSL) is a more generalized form of Spectrum Loss" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 315, + 73, + 378, + 201 + ], + "blocks": [ + { + "bbox": [ + 315, + 73, + 378, + 201 + ], + "lines": [ + { + "bbox": [ + 315, + 73, + 378, + 201 + ], + "spans": [ + { + "bbox": [ + 315, + 73, + 378, + 201 + ], + "type": "image", + "image_path": "884396341330d2834aad988835914292b738c9ff0e91466b21cc7ee2afc6bf40.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 331, + 203, + 361, + 213 + ], + "lines": [ + { + "bbox": [ + 331, + 203, + 361, + 213 + ], + "spans": [ + { + "bbox": [ + 331, + 203, + 361, + 213 + ], + "type": "text", + "content": "(a) FCM" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 386, + 71, + 452, + 202 + ], + "blocks": [ + { + "bbox": [ + 386, + 71, + 452, + 202 + ], + "lines": [ + { + "bbox": [ + 386, + 71, + 452, + 202 + ], + "spans": [ + { + "bbox": [ + 386, + 71, + 452, + 202 + ], + "type": "image", + "image_path": "a185b0a415dcdd350010fbac884698b49521abaacf2dc9398fea1b8bd44aff9e.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 386, + 203, + 453, + 213 + ], + "lines": [ + { + "bbox": [ + 386, + 203, + 453, + 213 + ], + "spans": [ + { + "bbox": [ + 386, + 203, + 453, + 213 + ], + "type": "text", + "content": "(b) FCM w/ residual" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 465, + 70, + 526, + 202 + ], + "blocks": [ + { + "bbox": [ + 465, + 70, + 526, + 202 + ], + "lines": [ + { + "bbox": [ + 465, + 70, + 526, + 202 + ], + "spans": [ + { + "bbox": [ + 465, + 70, + 526, + 202 + ], + "type": "image", + "image_path": "2c8618e727623ee74960bb7ffaa264f574e20e07e09360b454a36d66134545a6.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 461, + 203, + 531, + 213 + ], + "lines": [ + { + "bbox": [ + 461, + 203, + 531, + 213 + ], + "spans": [ + { + "bbox": [ + 461, + 203, + 531, + 213 + ], + "type": "text", + "content": "(c) FCM w/ attention" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 315, + 216, + 535, + 227 + ], + "lines": [ + { + "bbox": [ + 315, + 216, + 535, + 227 + ], + "spans": [ + { + "bbox": [ + 315, + 216, + 535, + 227 + ], + "type": "text", + "content": "Figure 5. FCM with different architectures and connections." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 229, + 546, + 265 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 229, + 546, + 265 + ], + "spans": [ + { + "bbox": [ + 305, + 229, + 546, + 265 + ], + "type": "text", + "content": "(SL) with learnable " + }, + { + "bbox": [ + 305, + 229, + 546, + 265 + ], + "type": "inline_equation", + "content": "\\sigma_{i}" + }, + { + "bbox": [ + 305, + 229, + 546, + 265 + ], + "type": "text", + "content": ". Note that DSL also includes FFL as a special form when used on the original and reconstructed images without the Gaussian filters." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 305, + 265, + 545, + 288 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 265, + 545, + 288 + ], + "spans": [ + { + "bbox": [ + 305, + 265, + 545, + 288 + ], + "type": "text", + "content": "Then the total reconstruction loss for FA-VAE can be described as:" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 335, + 291, + 545, + 342 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 291, + 545, + 342 + ], + "spans": [ + { + "bbox": [ + 335, + 291, + 545, + 342 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} _ {r e c} = \\alpha \\operatorname {F F L} (X, \\hat {X}) + \\beta \\sum_ {i = 0} ^ {M - 1} \\operatorname {D S L} \\left(\\mathcal {A} _ {i}, \\mathcal {C} _ {i}\\right) \\tag {9} \\\\ + \\left\\| X - \\hat {X} \\right\\| _ {1} + \\mathcal {L} _ {p i p s} (X - \\hat {X}). \\\\ \\end{array}", + "image_path": "3d5c202074d7f5dd76ae1eac27238f37f3548a80f0e471ddbe6e453d30fa2989.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 305, + 347, + 546, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 347, + 546, + 445 + ], + "spans": [ + { + "bbox": [ + 305, + 347, + 546, + 445 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 347, + 546, + 445 + ], + "type": "inline_equation", + "content": "\\mathrm{FFL}(X - \\hat{X})" + }, + { + "bbox": [ + 305, + 347, + 546, + 445 + ], + "type": "text", + "content": " is the Focal Frequency Loss applied on the original reconstructed images. The second term in Eq. 9 is the DSL loss applied over the outputs of the encoder and FCM blocks. The first two losses aim to minimize the frequency spectrum differences on the images and internal feature maps. The third and fourth losses act on the pixel and feature maps space [11]. " + }, + { + "bbox": [ + 305, + 347, + 546, + 445 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 305, + 347, + 546, + 445 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 347, + 546, + 445 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 305, + 347, + 546, + 445 + ], + "type": "text", + "content": " are hyperparameters. " + }, + { + "bbox": [ + 305, + 347, + 546, + 445 + ], + "type": "inline_equation", + "content": "\\sigma_{i}" + }, + { + "bbox": [ + 305, + 347, + 546, + 445 + ], + "type": "text", + "content": " are model parameters and optimized as:" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 347, + 449, + 545, + 469 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 347, + 449, + 545, + 469 + ], + "spans": [ + { + "bbox": [ + 347, + 449, + 545, + 469 + ], + "type": "interline_equation", + "content": "\\sigma_ {i} ^ {*}, \\mathcal {E} ^ {*}, \\mathcal {G} ^ {*}, \\mathcal {C} ^ {*} = \\underset {\\sigma_ {i}, \\mathcal {E}, \\mathcal {G}, \\mathcal {C}} {\\arg \\min } \\left(\\mathcal {L} _ {\\text {r e c}} + \\mathcal {L} _ {Q}\\right) \\tag {10}", + "image_path": "20cff33bb0d7dc4acb7aeda358cf6b984df4e0830656d1752a7bcf0f5e19ac2f.jpg" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 305, + 474, + 545, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 474, + 545, + 557 + ], + "spans": [ + { + "bbox": [ + 305, + 474, + 545, + 557 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_Q" + }, + { + "bbox": [ + 305, + 474, + 545, + 557 + ], + "type": "text", + "content": " is the quantization loss which minimizes the difference between the codebook embeddings and the embeddings given by the encoder " + }, + { + "bbox": [ + 305, + 474, + 545, + 557 + ], + "type": "inline_equation", + "content": "\\mathcal{E}" + }, + { + "bbox": [ + 305, + 474, + 545, + 557 + ], + "type": "text", + "content": ", more details in [11]. We also use the " + }, + { + "bbox": [ + 305, + 474, + 545, + 557 + ], + "type": "inline_equation", + "content": "L_{2}" + }, + { + "bbox": [ + 305, + 474, + 545, + 557 + ], + "type": "text", + "content": " regularization on the codebook embeddings during quantization as in [53] and exponential moving average (EMA) for updating the codebook [46] as they provide more stable training for the quantization block " + }, + { + "bbox": [ + 305, + 474, + 545, + 557 + ], + "type": "inline_equation", + "content": "\\mathcal{Q}" + }, + { + "bbox": [ + 305, + 474, + 545, + 557 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 304, + 558, + 546, + 654 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 558, + 546, + 654 + ], + "spans": [ + { + "bbox": [ + 304, + 558, + 546, + 654 + ], + "type": "text", + "content": "Benefits of DSL The advantages of learnable " + }, + { + "bbox": [ + 304, + 558, + 546, + 654 + ], + "type": "inline_equation", + "content": "\\sigma_{i}" + }, + { + "bbox": [ + 304, + 558, + 546, + 654 + ], + "type": "text", + "content": " in DSL are shown in Figure 3 column 5 where the reconstructed activations show no checkerboard artifacts as in column 4. Moreover, the reconstructed activations " + }, + { + "bbox": [ + 304, + 558, + 546, + 654 + ], + "type": "inline_equation", + "content": "C_1" + }, + { + "bbox": [ + 304, + 558, + 546, + 654 + ], + "type": "text", + "content": " are more similar to the encoder activations " + }, + { + "bbox": [ + 304, + 558, + 546, + 654 + ], + "type": "inline_equation", + "content": "A_{1}" + }, + { + "bbox": [ + 304, + 558, + 546, + 654 + ], + "type": "text", + "content": " on the frequency spectrum. Compared to the baseline model (column 2), our model exhibits a more harmonious balance between low- and highfrequencies, leading to more accurate reconstructed images." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 306, + 659, + 545, + 683 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 659, + 545, + 683 + ], + "spans": [ + { + "bbox": [ + 306, + 659, + 545, + 683 + ], + "type": "text", + "content": "3.2. Cross-attention Autoregressive Transformer (CAT)" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "text", + "content": "We further extend FA-VAE to the text-to-image generation task and introduce a new Cross-attention Autoregres" + } + ] + } + ], + "index": 24 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1740" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 77, + 71, + 521, + 233 + ], + "blocks": [ + { + "bbox": [ + 77, + 71, + 521, + 233 + ], + "lines": [ + { + "bbox": [ + 77, + 71, + 521, + 233 + ], + "spans": [ + { + "bbox": [ + 77, + 71, + 521, + 233 + ], + "type": "image", + "image_path": "0ba7b2117c5b13a6e7cbcc7d10418207d814a840d3dd6f2e4fc0de38d9078575.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 176, + 236, + 416, + 247 + ], + "lines": [ + { + "bbox": [ + 176, + 236, + 416, + 247 + ], + "spans": [ + { + "bbox": [ + 176, + 236, + 416, + 247 + ], + "type": "text", + "content": "Figure 6. Reconstruction on ImageNet [9], label: Japanese terrier." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 253, + 289, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 253, + 289, + 361 + ], + "spans": [ + { + "bbox": [ + 46, + 253, + 289, + 361 + ], + "type": "text", + "content": "sive Transformer (CAT), presented in Figure 4. CAT uses all the token embeddings of a textual description given by the pretrained text encoder of CLIP model [33], while existing works mostly use one or partial textual embeddings [8, 38, 49]. Furthermore, CAT uses a cross-attention mechanism with the text tokens to guide the image generation at each step. This more fine-grained text condition allows the generation to capture more precisely the relationships of semantic attributes between text and image." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 115, + 372, + 287, + 397 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 115, + 372, + 287, + 397 + ], + "spans": [ + { + "bbox": [ + 115, + 372, + 287, + 397 + ], + "type": "interline_equation", + "content": "p (s \\mid c) = \\prod_ {i} p \\left(s _ {i} \\mid s _ {< i}, c\\right) \\tag {11}", + "image_path": "53e153dc3f570fe55ec886d780c2bfb42e853b3c2312be23bc562e3264a05b2d.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 403, + 287, + 440 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 403, + 287, + 440 + ], + "spans": [ + { + "bbox": [ + 46, + 403, + 287, + 440 + ], + "type": "text", + "content": "Then the predicted indices " + }, + { + "bbox": [ + 46, + 403, + 287, + 440 + ], + "type": "inline_equation", + "content": "z'" + }, + { + "bbox": [ + 46, + 403, + 287, + 440 + ], + "type": "text", + "content": " can be decoded to an image using FA-VAE's decoder " + }, + { + "bbox": [ + 46, + 403, + 287, + 440 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 46, + 403, + 287, + 440 + ], + "type": "text", + "content": ". The loss is to maximize the log-likelihood of the data representations," + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 110, + 449, + 287, + 463 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 110, + 449, + 287, + 463 + ], + "spans": [ + { + "bbox": [ + 110, + 449, + 287, + 463 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {C A T}} = \\mathbb {E} _ {x \\sim p (x)} [ - \\log p (s) ] \\tag {12}", + "image_path": "aa3a278ac30c22479acb8bd9e85a6496ce6605b2b13ded6a5f6e7ddd502ca43c.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 472, + 287, + 520 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 472, + 287, + 520 + ], + "spans": [ + { + "bbox": [ + 46, + 472, + 287, + 520 + ], + "type": "text", + "content": "In this paper, the GPT2 model [34], which is the default setting in [11, 49], is utilized as the backbone for CAT, and cross-attention mechanism inspired from [47] is applied with the GPT2 structure." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 531, + 128, + 544 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 531, + 128, + 544 + ], + "spans": [ + { + "bbox": [ + 47, + 531, + 128, + 544 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 551, + 169, + 563 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 551, + 169, + 563 + ], + "spans": [ + { + "bbox": [ + 47, + 551, + 169, + 563 + ], + "type": "text", + "content": "4.1. Experimental Details" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "type": "text", + "content": "The datasets used in this paper are: (1) MultiModal CelebA-HQ dataset [50] with 30,000 high-resolution celebrity face images and each image comes with ten captions; (2) Flickr-Faces-HQ (FFHQ) dataset [21] of 70,000 high-resolution face images; (3) ImageNet [9] that contains around 1 million images from 1000 categories. Experiments are performed on V-100 GPUs and results are reported on the validation sets unless specified otherwise. For a fair comparison with existing models, the resolutions used for training are " + }, + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "type": "text", + "content": " on all datasets. The details of training settings and hyperparameters are provided in the supplementary materials." + } + ] + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 310, + 251, + 542, + 407 + ], + "blocks": [ + { + "bbox": [ + 310, + 251, + 542, + 407 + ], + "lines": [ + { + "bbox": [ + 310, + 251, + 542, + 407 + ], + "spans": [ + { + "bbox": [ + 310, + 251, + 542, + 407 + ], + "type": "table", + "html": "
ModelDatasetCodebook Size(h × w)rFID ↓
RQ-VAE [25]FFHQ2048(8 × 8)5.33
FA-VAE (Ours)FFHQ2048(16 × 16)4.98
VQ-VAE-2 [39]ImageNet512(64 × 64) & (32 × 32)~ 10 (train)
VQ-GAN [40]ImageNet8192(64 × 64)1.06
FA-VAE (Ours)ImageNet8192(64 × 64)0.40
DALL-E [38]ImageNet8192(32 × 32)32.01
VQ-GAN [11]ImageNet16384(16 × 16)5.15
VQ-GAN [11]ImageNet1024(16 × 16)7.94
VQ-GAN [25]ImageNet16384(8 × 8)17.95
RQ-VAE† [46]ImageNet16384(8 × 8)10.77
RQ-VAE* [25]ImageNet16384(8 × 8)4.73
FA-VAE (Ours)ImageNet16384(16 × 16)4.60
", + "image_path": "01040bda16eed070d927281fb1db8a90d01da9569144389ad52d856728b67117.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 409, + 545, + 443 + ], + "lines": [ + { + "bbox": [ + 305, + 409, + 545, + 443 + ], + "spans": [ + { + "bbox": [ + 305, + 409, + 545, + 443 + ], + "type": "text", + "content": "Table 1. Reconstruction results on the validation data of FFHQ [21] and ImageNet [9] respectively, more results in the supplement. " + }, + { + "bbox": [ + 305, + 409, + 545, + 443 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 305, + 409, + 545, + 443 + ], + "type": "text", + "content": " means depth 2, and " + }, + { + "bbox": [ + 305, + 409, + 545, + 443 + ], + "type": "inline_equation", + "content": "*" + }, + { + "bbox": [ + 305, + 409, + 545, + 443 + ], + "type": "text", + "content": " means depth 4 from RQ-VAE [25]." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 306, + 453, + 432, + 464 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 453, + 432, + 464 + ], + "spans": [ + { + "bbox": [ + 306, + 453, + 432, + 464 + ], + "type": "text", + "content": "4.2. Image Reconstruction" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 472, + 545, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 472, + 545, + 592 + ], + "spans": [ + { + "bbox": [ + 304, + 472, + 545, + 592 + ], + "type": "text", + "content": "Reconstruction on FFHQ and Imagenet The experiment results of reconstruction on FFHQ [21] and ImageNet [9] are presented in Table 1. The results of the baselines are from the original paper. Note that smaller " + }, + { + "bbox": [ + 304, + 472, + 545, + 592 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 304, + 472, + 545, + 592 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 472, + 545, + 592 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 304, + 472, + 545, + 592 + ], + "type": "text", + "content": " means that the downsampling factor or the compression rate " + }, + { + "bbox": [ + 304, + 472, + 545, + 592 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 304, + 472, + 545, + 592 + ], + "type": "text", + "content": " is larger. FA-VAE model shows improved reconstruction quality over baseline models across different compression rates in both datasets. The main reason is that FA-VAE can successfully reconstruct the important middle and high-frequencies which are neglected in baseline models." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 594, + 546, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 546, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 546, + 714 + ], + "type": "text", + "content": "Figure 6 further supports our previous claim and demonstrates clear diverging reconstruction qualities between baseline models VQ-GAN [11], DALL-E [38], and our FA-VAE model. With a larger compression rate, Figure 6 row 2 shows that more middle and high frequencies are compressed in VQ-GAN and DALL-E. In comparison, FA-VAE can reconstruct middle and high frequencies more accurately, which translates to good representations with improved semantics and local details in Figure 6 row 1. For instance, at a compression rate of 16, the dog's tongue of" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 314, + 757 + ], + "type": "text", + "content": "1741" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 47, + 70, + 548, + 256 + ], + "blocks": [ + { + "bbox": [ + 47, + 70, + 548, + 256 + ], + "lines": [ + { + "bbox": [ + 47, + 70, + 548, + 256 + ], + "spans": [ + { + "bbox": [ + 47, + 70, + 548, + 256 + ], + "type": "table", + "html": "
ablation onFFL(X, X̂)SL(Ai, Ci)FCM Fkernel K_i(μ, σ_i)DSL(Ai, Ci)kernel size μinitial value σ_ipair-wise σ_iL1 ↓Lpips ↓rFID ↓
1VQ-GAN [11]XXXXXXXX0.1210.3010.12
2VQ-GAN + Style [53]XXXXXXXX0.0850.2311.90
3VQ-GAN [15]✓XXXXXXX0.1140.3530.65
4SL w/o kernel✓XCONVXXXXX0.0820.227.04
5✓✓CONVXXXXX0.0820.227.02
6SL w/ kernel✓✓CONV✓X93X0.0800.227.39
7FCM architecture✓✓CONV✓✓93X0.0810.215.90
8✓✓RES✓✓93X0.0780.216.44
9✓✓ATTN✓✓93X0.0890.237.49
10DSL kernel size μ✓✓RES✓✓33✓0.0810.216.66
11✓✓RES✓✓53✓0.0810.217.04
12✓✓RES✓✓93✓0.0820.226.53
13✓✓RES✓✓113✓0.0830.227.48
14✓✓RES✓✓153✓0.0830.226.36
", + "image_path": "d2f6ddcb3d6c15a1d5d328fbfe6eb42d7ae0dc3a6b8f88ab33955627e99b92d3.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 256, + 547, + 281 + ], + "lines": [ + { + "bbox": [ + 46, + 256, + 547, + 281 + ], + "spans": [ + { + "bbox": [ + 46, + 256, + 547, + 281 + ], + "type": "text", + "content": "Table 2. Ablation studies on the CelebA-HQ validation dataset [17], visualizations are in Figure 7. Words put in bold to highlight the changing component." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 44, + 283, + 293, + 488 + ], + "blocks": [ + { + "bbox": [ + 44, + 283, + 293, + 488 + ], + "lines": [ + { + "bbox": [ + 44, + 283, + 293, + 488 + ], + "spans": [ + { + "bbox": [ + 44, + 283, + 293, + 488 + ], + "type": "image", + "image_path": "baaff5b0e1bf3eb64c21e14321a4c13220115cfb0846f9f227cf826f024b83db.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 489, + 288, + 544 + ], + "lines": [ + { + "bbox": [ + 46, + 489, + 288, + 544 + ], + "spans": [ + { + "bbox": [ + 46, + 489, + 288, + 544 + ], + "type": "text", + "content": "Figure 7. Reconstruction comparisons for ablation studies on CelebA-HQ. The figure number represents the setting in the corresponding row in Table 2. " + }, + { + "bbox": [ + 46, + 489, + 288, + 544 + ], + "type": "inline_equation", + "content": "\\mathrm{SL}^*" + }, + { + "bbox": [ + 46, + 489, + 288, + 544 + ], + "type": "text", + "content": " is SL without Gaussian kernel, " + }, + { + "bbox": [ + 46, + 489, + 288, + 544 + ], + "type": "inline_equation", + "content": "\\mathrm{DSL}^{**}" + }, + { + "bbox": [ + 46, + 489, + 288, + 544 + ], + "type": "text", + "content": " is DSL with non pairwise " + }, + { + "bbox": [ + 46, + 489, + 288, + 544 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 46, + 489, + 288, + 544 + ], + "type": "text", + "content": ". FCM is convolution architecture by default." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 553, + 287, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 553, + 287, + 590 + ], + "spans": [ + { + "bbox": [ + 46, + 553, + 287, + 590 + ], + "type": "text", + "content": "FA-VAE's reconstruction is more similar to the original image than VQ-GAN in terms of color and shape; more qualitative results are in the supplement." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 591, + 287, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 591, + 287, + 652 + ], + "spans": [ + { + "bbox": [ + 46, + 591, + 287, + 652 + ], + "type": "text", + "content": "Ablation Studies on FCM and DSL In Table 2, we perform ablation studies on different architectures of FCM in combination with different settings of the Spectrum Loss (SL) and Dynamic Spectrum Loss (DSL). Figure 7 gives the accompanying visualizations of the quantitative results." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "type": "text", + "content": "First, as motivated in the Section 3, Table 2 row 3 shows that considering all frequencies as equally important leads the VQ-GAN model to poor performance in terms of rFID and lpips (see Figure 7 image 3). In comparison, results in Table 2 rows 4 demonstrate that FCMs alone can help" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 302, + 285, + 552, + 553 + ], + "blocks": [ + { + "bbox": [ + 302, + 285, + 552, + 553 + ], + "lines": [ + { + "bbox": [ + 302, + 285, + 552, + 553 + ], + "spans": [ + { + "bbox": [ + 302, + 285, + 552, + 553 + ], + "type": "image", + "image_path": "932a0db2c62244545bb9d924242c6cdc083e564b88fe6a69abd09bc669dbf422.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 556, + 547, + 590 + ], + "lines": [ + { + "bbox": [ + 304, + 556, + 547, + 590 + ], + "spans": [ + { + "bbox": [ + 304, + 556, + 547, + 590 + ], + "type": "text", + "content": "Figure 8. Zero-shot reconstruction on CelebA-HQ [17], FFHQ [21] and photos from the wild, taken from [11]. The models are trained on ImageNet on different compression rates." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 594, + 547, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 547, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 547, + 712 + ], + "type": "text", + "content": "FA-VAE model achieve a better reconstruction because the residual connections in FCMs preserve better the information flow. The reconstruction is further improved when combined with a simple SL on images level (row 5) or blocks level (row 6). Although SL (row 6) shows slightly inferior reconstruction due to neglection of frequency importance variance across decoder blocks, Figure 7 image 5 shows that the details, such as earrings and hairs, can still be more enhanced than the baseline model shown in image 1." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1742" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 73, + 288, + 79 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 73, + 288, + 79 + ], + "spans": [ + { + "bbox": [ + 55, + 73, + 288, + 79 + ], + "type": "text", + "content": "the woman has high cheekbones, wavy hair, big nose, mouth open, big lips, pointy nose, and" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 84, + 79, + 259, + 87 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 84, + 79, + 259, + 87 + ], + "spans": [ + { + "bbox": [ + 84, + 79, + 259, + 87 + ], + "type": "text", + "content": "arched eyebrows. she is young and wears heavy makeup, and lipstick." + } + ] + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 51, + 86, + 129, + 163 + ], + "blocks": [ + { + "bbox": [ + 51, + 86, + 129, + 163 + ], + "lines": [ + { + "bbox": [ + 51, + 86, + 129, + 163 + ], + "spans": [ + { + "bbox": [ + 51, + 86, + 129, + 163 + ], + "type": "image", + "image_path": "cfe241304298cd6e8d8b818b986c9dc29b5a870d0f74404f1028cf72449795fb.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 134, + 87, + 213, + 163 + ], + "blocks": [ + { + "bbox": [ + 134, + 87, + 213, + 163 + ], + "lines": [ + { + "bbox": [ + 134, + 87, + 213, + 163 + ], + "spans": [ + { + "bbox": [ + 134, + 87, + 213, + 163 + ], + "type": "image", + "image_path": "5d53cae184c22bb6f3549e040602db9ec6ee33bf23c152baa7e129404438f150.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 184, + 172, + 408, + 184 + ], + "lines": [ + { + "bbox": [ + 184, + 172, + 408, + 184 + ], + "spans": [ + { + "bbox": [ + 184, + 172, + 408, + 184 + ], + "type": "text", + "content": "Figure 9. Text-to-Image generation on CelebA-HQ-MM [17]" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 217, + 87, + 296, + 163 + ], + "blocks": [ + { + "bbox": [ + 217, + 87, + 296, + 163 + ], + "lines": [ + { + "bbox": [ + 217, + 87, + 296, + 163 + ], + "spans": [ + { + "bbox": [ + 217, + 87, + 296, + 163 + ], + "type": "image", + "image_path": "f4002f6c85c47a9bfad67f5021422678f8801addf8b46a7ef97ec0227f6de5c7.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 379, + 76, + 466, + 83 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 379, + 76, + 466, + 83 + ], + "spans": [ + { + "bbox": [ + 379, + 76, + 466, + 83 + ], + "type": "text", + "content": "He has goatee and wears necktie." + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 301, + 86, + 380, + 163 + ], + "blocks": [ + { + "bbox": [ + 301, + 86, + 380, + 163 + ], + "lines": [ + { + "bbox": [ + 301, + 86, + 380, + 163 + ], + "spans": [ + { + "bbox": [ + 301, + 86, + 380, + 163 + ], + "type": "image", + "image_path": "b18c3107873dd5e4fef8daa512f186606422465d46391a7cd90fcafc290df8fc.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 383, + 86, + 462, + 163 + ], + "blocks": [ + { + "bbox": [ + 383, + 86, + 462, + 163 + ], + "lines": [ + { + "bbox": [ + 383, + 86, + 462, + 163 + ], + "spans": [ + { + "bbox": [ + 383, + 86, + 462, + 163 + ], + "type": "image", + "image_path": "19510cdb08f01382673ee2673a3e2b70c5c9e73c5d46748decc291b48cf89fa8.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 467, + 86, + 545, + 163 + ], + "blocks": [ + { + "bbox": [ + 467, + 86, + 545, + 163 + ], + "lines": [ + { + "bbox": [ + 467, + 86, + 545, + 163 + ], + "spans": [ + { + "bbox": [ + 467, + 86, + 545, + 163 + ], + "type": "image", + "image_path": "e56a840f29eeaf83d60f908e87ab6a197b0f13d34f9bd022bc3950abe5f5cfdd.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 190, + 287, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 190, + 287, + 357 + ], + "spans": [ + { + "bbox": [ + 46, + 190, + 287, + 357 + ], + "type": "text", + "content": "Then, Table 2 rows 7-9 compare the performance when the architecture of FCM varies as illustrated in Figure 5; note that instead of using same " + }, + { + "bbox": [ + 46, + 190, + 287, + 357 + ], + "type": "inline_equation", + "content": "\\sigma_{i}" + }, + { + "bbox": [ + 46, + 190, + 287, + 357 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 46, + 190, + 287, + 357 + ], + "type": "inline_equation", + "content": "(\\mathcal{A}_i,\\mathcal{C}_i)" + }, + { + "bbox": [ + 46, + 190, + 287, + 357 + ], + "type": "text", + "content": ", we use two learnable " + }, + { + "bbox": [ + 46, + 190, + 287, + 357 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 46, + 190, + 287, + 357 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 46, + 190, + 287, + 357 + ], + "type": "inline_equation", + "content": "\\mathcal{A}_i" + }, + { + "bbox": [ + 46, + 190, + 287, + 357 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 190, + 287, + 357 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_i" + }, + { + "bbox": [ + 46, + 190, + 287, + 357 + ], + "type": "text", + "content": " respectively. Quantitatively, when everything else is held equal, the convolution architecture shows better reconstruction than residual and attention architectures. Similarly, Figure 7 image 7 resembles the original image more than images 8 and 9 in terms of face radiance, smoothness, lip color, and shape. The reason is that residual connection enriches the outputs of FCM, which defies the purpose of enriching features of higher frequencies of the decoder features motivated in section 3.1.1. The attention mechanism is at a disadvantage here because the decoder and the encoder exclusively use convolutions." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 359, + 287, + 466 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 359, + 287, + 466 + ], + "spans": [ + { + "bbox": [ + 46, + 359, + 287, + 466 + ], + "type": "text", + "content": "Finally, varying the size of kernels in DSL (Table 2 row 10 - 14) show quite similar quantitative reconstruction performance, while qualitatively, Figure 7 shows that a larger kernel size tends to produce smoother reconstructions. One reason could be that a larger kernel size tends to smooth more the feature maps because more surrounding values are used during convolution. Thus, in other experiments, we choose a kernel size of 3, and the effects of kernel sizes are open to future works." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 468, + 287, + 599 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 468, + 287, + 599 + ], + "spans": [ + { + "bbox": [ + 46, + 468, + 287, + 599 + ], + "type": "text", + "content": "Zero-shot Reconstruction To further demonstrate the reconstruction capabilities of FA-VAE models, Table 3 gives the zero-shot reconstruction performance evaluated on CelebA-HQ [17] and FFHQ [21] using models trained on ImageNet [9]. The accompanying qualitative results are in Figure 8. Overall, FA-VAE displays impressive transferability capability with more faithful reconstruction in details, for instance, the light contrast in column 1 and book details in column 3. Note also that the common metric used in image compression PSNR is not perfectly correlated with rFID, but we put the metric here for reference." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 47, + 610, + 146, + 623 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 610, + 146, + 623 + ], + "spans": [ + { + "bbox": [ + 47, + 610, + 146, + 623 + ], + "type": "text", + "content": "4.3. Image Synthesis" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 46, + 629, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 629, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 629, + 287, + 713 + ], + "type": "text", + "content": "Table 4 shows the text-to-image generation performance on CelebA-HQ-MM [17]. Our proposed autoregressive transformer CAT yields a better generation quality than other GAN-based models, including AttnGAN [52], ControlGAN [26], which are solely designed for image generation. Figure 9 shows that CAT can generate satisfactory images conditioned on text inputs of varying lengths. At-" + } + ] + } + ], + "index": 14 + }, + { + "type": "table", + "bbox": [ + 310, + 188, + 543, + 268 + ], + "blocks": [ + { + "bbox": [ + 310, + 188, + 543, + 268 + ], + "lines": [ + { + "bbox": [ + 310, + 188, + 543, + 268 + ], + "spans": [ + { + "bbox": [ + 310, + 188, + 543, + 268 + ], + "type": "table", + "html": "
CelebAFFHQ
Pretrained ModelfrFID ↓PSNR ↑rFID ↓PSNR ↑
VQ-GAN [40]168.6223.406.8322.68
VQ-GAN [11]44.1131.203.0630.82
FA-VAE (Ours)166.5222.596.1921.95
FA-VAE (Ours)42.2531.391.4630.85
", + "image_path": "cecdb9180f22dc7c0437c53c13d84134ffdf033419b35e4fde51e1903fcea4c4.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 305, + 270, + 547, + 315 + ], + "lines": [ + { + "bbox": [ + 305, + 270, + 547, + 315 + ], + "spans": [ + { + "bbox": [ + 305, + 270, + 547, + 315 + ], + "type": "text", + "content": "Table 3. Zero-shot reconstruction results on the validation data of FFHQ [21] and CelebA-HQ [17] using models trained on ImageNet [9]. " + }, + { + "bbox": [ + 305, + 270, + 547, + 315 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 305, + 270, + 547, + 315 + ], + "type": "text", + "content": " is the downsampling factor, the codebook sizes for " + }, + { + "bbox": [ + 305, + 270, + 547, + 315 + ], + "type": "inline_equation", + "content": "f = \\{ {16},8,4\\}" + }, + { + "bbox": [ + 305, + 270, + 547, + 315 + ], + "type": "text", + "content": " are " + }, + { + "bbox": [ + 305, + 270, + 547, + 315 + ], + "type": "inline_equation", + "content": "\\{ {16384},{16384},{8192}\\}" + }, + { + "bbox": [ + 305, + 270, + 547, + 315 + ], + "type": "text", + "content": " respectively." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 15 + }, + { + "type": "table", + "bbox": [ + 362, + 318, + 491, + 411 + ], + "blocks": [ + { + "bbox": [ + 362, + 318, + 491, + 411 + ], + "lines": [ + { + "bbox": [ + 362, + 318, + 491, + 411 + ], + "spans": [ + { + "bbox": [ + 362, + 318, + 491, + 411 + ], + "type": "table", + "html": "
ModelFID ↓
AttnGAN [52]125.98
ControlGAN [26]116.32
DM-GAN [55]131.05
DF-GAN [44]137.60
TediGAN [50]106.37
LAFITE [54]12.54
CAT (Ours)10.23
", + "image_path": "7fbaac8d8421c9e877aee0aabaa65043276a85b3a1756d769d87188fe91626d6.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 304, + 413, + 545, + 485 + ], + "lines": [ + { + "bbox": [ + 304, + 413, + 545, + 485 + ], + "spans": [ + { + "bbox": [ + 304, + 413, + 545, + 485 + ], + "type": "text", + "content": "Table 4. Text-to-image generation on CelebA-HQ MM [50]. tributes such as \"mouth open\" and \"arched eyebrows\" are captured during generation because all the tokens embeddings are used as a condition which gives more precise guidance. More quantitative and qualitative results on different datasets are in the supplement." + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 17 + }, + { + "bbox": [ + 306, + 495, + 379, + 508 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 495, + 379, + 508 + ], + "spans": [ + { + "bbox": [ + 306, + 495, + 379, + 508 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 515, + 545, + 658 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 515, + 545, + 658 + ], + "spans": [ + { + "bbox": [ + 304, + 515, + 545, + 658 + ], + "type": "text", + "content": "In this paper, we introduce the Frequency Augmented VAE (FA-VAE) model, which aims to improve reconstruction quality by bridging the gaps in the frequency domains between original and reconstructed images. New modules named Frequency Complement Modules (FCM) are crafted and guided under the new (Dynamic) Spectrum Loss ((D)SL) to learn to complement the reconstructed features of missing frequencies. A new Cross-attention Autoregressive Transformer (CAT) is proposed for achieving more precise textual-image alignment in the text-to-image generation task. FA-VAE shows improved reconstruction on various datasets compared to SOTA methods." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 306, + 670, + 404, + 683 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 670, + 404, + 683 + ], + "spans": [ + { + "bbox": [ + 306, + 670, + 404, + 683 + ], + "type": "text", + "content": "Acknowledgement" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 306, + 689, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 545, + 712 + ], + "type": "text", + "content": "This research is supported in part by the National Science Foundation under award No. 2040209." + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1743" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 92, + 286, + 712 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 53, + 92, + 286, + 125 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 92, + 286, + 125 + ], + "spans": [ + { + "bbox": [ + 53, + 92, + 286, + 125 + ], + "type": "text", + "content": "[1] Jyoti Aneja, Alexander G. Schwing, Jan Kautz, and Arash Vahdat. Ncp-vae: Variational autoencoders with noise contrastive priors. *ArXiv*, abs/2010.02917, 2020. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 126, + 286, + 159 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 126, + 286, + 159 + ], + "spans": [ + { + "bbox": [ + 53, + 126, + 286, + 159 + ], + "type": "text", + "content": "[2] Valerio Biscione and Jeffrey S. Bowers. Convolutional neural networks are not invariant to translation, but they can learn to be. JMLR, 2022. 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 161, + 286, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 161, + 286, + 194 + ], + "spans": [ + { + "bbox": [ + 53, + 161, + 286, + 194 + ], + "type": "text", + "content": "[3] Andrew Brock, Jeff Donahue, and Karen Simonyan. Large scale GAN training for high fidelity natural image synthesis. In ICLR, 2019. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 196, + 286, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 196, + 286, + 228 + ], + "spans": [ + { + "bbox": [ + 53, + 196, + 286, + 228 + ], + "type": "text", + "content": "[4] Huiwen Chang, Han Zhang, Lu Jiang, Ce Liu, and William T. Freeman. Maskgit: Masked generative image transformer. In CVPR, 2022. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 231, + 286, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 231, + 286, + 262 + ], + "spans": [ + { + "bbox": [ + 53, + 231, + 286, + 262 + ], + "type": "text", + "content": "[5] Ting Chen, Ruixiang Zhang, and Geoffrey Hinton. Analog bits: Generating discrete data using diffusion models with self-conditioning, 2022. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 266, + 286, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 266, + 286, + 297 + ], + "spans": [ + { + "bbox": [ + 53, + 266, + 286, + 297 + ], + "type": "text", + "content": "[6] Rewon Child. Very deep " + }, + { + "bbox": [ + 53, + 266, + 286, + 297 + ], + "type": "inline_equation", + "content": "\\{\\mathrm{vae}\\}" + }, + { + "bbox": [ + 53, + 266, + 286, + 297 + ], + "type": "text", + "content": "s generalize autoregressive models and can outperform them on images. In ICLR, 2021. 1, 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 300, + 286, + 332 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 300, + 286, + 332 + ], + "spans": [ + { + "bbox": [ + 53, + 300, + 286, + 332 + ], + "type": "text", + "content": "[7] Jooyoung Choi, Jungbeom Lee, Chaehun Shin, Sungwon Kim, Hyunwoo Kim, and Sungroh Yoon. Perception prioritized training of diffusion models. In CVPR, 2022. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 335, + 286, + 388 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 335, + 286, + 388 + ], + "spans": [ + { + "bbox": [ + 53, + 335, + 286, + 388 + ], + "type": "text", + "content": "[8] Katherine Crowson, Stella Rose Biderman, Daniel Kornis, Dashiell Stander, Eric Hallahan, Louis Castricato, and Edward Raff. Vqgan-clip: Open domain image generation and editing with natural language guidance. In ECCV, 2022. 2, 3, 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 392, + 286, + 423 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 392, + 286, + 423 + ], + "spans": [ + { + "bbox": [ + 53, + 392, + 286, + 423 + ], + "type": "text", + "content": "[9] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In CVPR, 2009. 6, 8" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 426, + 286, + 448 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 426, + 286, + 448 + ], + "spans": [ + { + "bbox": [ + 48, + 426, + 286, + 448 + ], + "type": "text", + "content": "[10] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. In NeurIPS, 2021. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 450, + 286, + 482 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 450, + 286, + 482 + ], + "spans": [ + { + "bbox": [ + 48, + 450, + 286, + 482 + ], + "type": "text", + "content": "[11] Patrick Esser, Robin Rombach, and Björn Ommer. Taming transformers for high-resolution image synthesis. In CVPR, 2021. 1, 2, 3, 4, 5, 6, 7, 8" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 485, + 286, + 528 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 485, + 286, + 528 + ], + "spans": [ + { + "bbox": [ + 48, + 485, + 286, + 528 + ], + "type": "text", + "content": "[12] Ge Gao, Pei You, Rong Pan, Shunyuan Han, Yuanyuan Zhang, Yuchao Dai, and Hojae Lee. Neural image compression via attentional multi-scale back projection and frequency decomposition. In ICCV, 2021. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 531, + 286, + 563 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 531, + 286, + 563 + ], + "spans": [ + { + "bbox": [ + 48, + 531, + 286, + 563 + ], + "type": "text", + "content": "[13] Karol Gregor, Ivo Danihelka, Andriy Mnih, Charles Blundell, and Daan Wierstra. Deep autoregressive networks. In ICML, 2014. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 565, + 286, + 608 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 565, + 286, + 608 + ], + "spans": [ + { + "bbox": [ + 48, + 565, + 286, + 608 + ], + "type": "text", + "content": "[14] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. In NeurIPS, 2017. 1" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 611, + 286, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 611, + 286, + 643 + ], + "spans": [ + { + "bbox": [ + 48, + 611, + 286, + 643 + ], + "type": "text", + "content": "[15] Liming Jiang, Bo Dai, Wayne Wu, and Chen Change Loy. Focal frequency loss for image reconstruction and synthesis. In ICCV, 2021. 2, 4, 7" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 48, + 646, + 286, + 677 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 646, + 286, + 677 + ], + "spans": [ + { + "bbox": [ + 48, + 646, + 286, + 677 + ], + "type": "text", + "content": "[16] Justin Johnson, Alexandre Alahi, and Li Fei-Fei. Perceptual losses for real-time style transfer and super-resolution. In ECCV, 2016. 1, 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 48, + 680, + 286, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 680, + 286, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 680, + 286, + 712 + ], + "type": "text", + "content": "[17] Tero Karras, Timo Aila, Samuli Laine, and Jaakko Lehtinen. Progressive growing of gans for improved quality, stability, and variation. CoRR, abs/1710.10196, 2017. 7, 8" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 36, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "text", + "content": "[18] Tero Karras, Timo Aila, Samuli Laine, and Jaakko Lehtinen. Progressive growing of GANs for improved quality, stability, and variation. In ICLR, 2018. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 108, + 545, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 108, + 545, + 140 + ], + "spans": [ + { + "bbox": [ + 307, + 108, + 545, + 140 + ], + "type": "text", + "content": "[19] Tero Karras, Miika Aittala, Janne Hellsten, Samuli Laine, Jaakko Lehtinen, and Timo Aila. Training generative adversarial networks with limited data. In NeurIPS, 2020. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 142, + 545, + 175 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 142, + 545, + 175 + ], + "spans": [ + { + "bbox": [ + 307, + 142, + 545, + 175 + ], + "type": "text", + "content": "[20] Tero Karras, Miika Aittala, Samuli Laine, Erik Härkönen, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Alias-free generative adversarial networks. In NeurIPS, 2021. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 177, + 545, + 208 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 177, + 545, + 208 + ], + "spans": [ + { + "bbox": [ + 307, + 177, + 545, + 208 + ], + "type": "text", + "content": "[21] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In CVPR, 2019. 2, 6, 7, 8" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 211, + 545, + 255 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 211, + 545, + 255 + ], + "spans": [ + { + "bbox": [ + 307, + 211, + 545, + 255 + ], + "type": "text", + "content": "[22] Soo Ye Kim, Kfir Aberman, Nori Kanazawa, Rahul Garg, Neal Wadhwa, Huiwen Chang, Nikhil Karnad, Munchurl Kim, and Orly Liba. Zoom-to-inpaint: Image inpainting with high frequency details. In CVPRW, 2022. 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 257, + 545, + 278 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 257, + 545, + 278 + ], + "spans": [ + { + "bbox": [ + 307, + 257, + 545, + 278 + ], + "type": "text", + "content": "[23] Diederik P Kingma and Max Welling. Auto-encoding variational bayes, 2013. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 280, + 545, + 334 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 280, + 545, + 334 + ], + "spans": [ + { + "bbox": [ + 307, + 280, + 545, + 334 + ], + "type": "text", + "content": "[24] Umut Kocasari, Alara Dirik, Mert Tiftikci, and Pinar Yanardag. Stylemc: Multi-channel based fast text-guided image generation and manipulation. 2022 IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 3441-3450, 2022. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 337, + 545, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 337, + 545, + 369 + ], + "spans": [ + { + "bbox": [ + 307, + 337, + 545, + 369 + ], + "type": "text", + "content": "[25] Doyup Lee, Chiheon Kim, Saehoon Kim, Minsu Cho, and Wook-Shin Han. Autoregressive image generation using residual quantization. In CVPR, 2022. 1, 2, 3, 4, 6" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 371, + 545, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 371, + 545, + 403 + ], + "spans": [ + { + "bbox": [ + 307, + 371, + 545, + 403 + ], + "type": "text", + "content": "[26] Bowen Li, Xiaojuan Qi, Thomas Lukasiewicz, and Philip H. S. Torr. Controllable text-to-image generation. In NeurIPS, 2019. 8" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 406, + 545, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 406, + 545, + 459 + ], + "spans": [ + { + "bbox": [ + 307, + 406, + 545, + 459 + ], + "type": "text", + "content": "[27] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. GLIDE: towards photorealistic image generation and editing with text-guided diffusion models. In ICML, 2022. 2, 3" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 462, + 545, + 483 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 462, + 545, + 483 + ], + "spans": [ + { + "bbox": [ + 307, + 462, + 545, + 483 + ], + "type": "text", + "content": "[28] Alexander Quinn Nichol and Prafulla Dhariwal. Improved denoising diffusion probabilistic models. In ICML, 2021. 3" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 307, + 486, + 545, + 517 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 486, + 545, + 517 + ], + "spans": [ + { + "bbox": [ + 307, + 486, + 545, + 517 + ], + "type": "text", + "content": "[29] Gaurav Parmar, Dacheng Li, Kwonjoon Lee, and Zhuowen Tu. Dual contradistinctive generative autoencoder. In CVPR, 2021. 1, 2" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 307, + 520, + 545, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 520, + 545, + 542 + ], + "spans": [ + { + "bbox": [ + 307, + 520, + 545, + 542 + ], + "type": "text", + "content": "[30] Yash Patel, Srikar Appalaraju, and R. Manmatha. Deep perceptual compression, 2019. 5" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 307, + 544, + 545, + 576 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 544, + 545, + 576 + ], + "spans": [ + { + "bbox": [ + 307, + 544, + 545, + 576 + ], + "type": "text", + "content": "[31] Jialun Peng, Dong Liu, Songcen Xu, and Houqiang Li. Generating diverse structure for image inpainting with hierarchical vq-vae. In CVPR, 2021. 1" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 307, + 578, + 545, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 578, + 545, + 609 + ], + "spans": [ + { + "bbox": [ + 307, + 578, + 545, + 609 + ], + "type": "text", + "content": "[32] Stanislav Pidhorskyi, Donald A. Adjeroh, and Gianfranco Doretto. Adversarial latent autoencoders. In CVPR, 2020. 2" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 307, + 613, + 545, + 677 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 613, + 545, + 677 + ], + "spans": [ + { + "bbox": [ + 307, + 613, + 545, + 677 + ], + "type": "text", + "content": "[33] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In ICML, 2021. 2, 3, 6" + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 307, + 680, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 680, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 307, + 680, + 545, + 712 + ], + "type": "text", + "content": "[34] Alec Radford, Jeff Wu, Rewon Child, David Luan, Dario Amodei, and Ilya Sutskever. Language models are unsupervised multitask learners. 2019. 6" + } + ] + } + ], + "index": 35 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1744" + } + ] + } + ], + "index": 37 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 126 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 126 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 126 + ], + "type": "text", + "content": "[35] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J. Liu. Exploring the limits of transfer learning with a unified text-to-text transformer. JMLR, 21(140):1-67, 2020. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 130, + 287, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 130, + 287, + 173 + ], + "spans": [ + { + "bbox": [ + 48, + 130, + 287, + 173 + ], + "type": "text", + "content": "[36] Nasim Rahaman, Aristide Baratin, Devansh Arpit, Felix Dräxler, Min Lin, Fred A. Hamprecht, Yoshua Bengio, and Aaron C. Courville. On the spectral bias of neural networks. In ICML, 2019. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 175, + 287, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 175, + 287, + 207 + ], + "spans": [ + { + "bbox": [ + 48, + 175, + 287, + 207 + ], + "type": "text", + "content": "[37] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. ArXiv, abs/2204.06125, 2022. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 210, + 287, + 243 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 210, + 287, + 243 + ], + "spans": [ + { + "bbox": [ + 48, + 210, + 287, + 243 + ], + "type": "text", + "content": "[38] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In ICML, 2021. 3, 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 245, + 287, + 277 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 245, + 287, + 277 + ], + "spans": [ + { + "bbox": [ + 48, + 245, + 287, + 277 + ], + "type": "text", + "content": "[39] Ali Razavi, Aaron van den Oord, and Oriol Vinyals. Generating diverse high-fidelity images with vq-vae-2. In NeurIPS, 2019. 1, 2, 3, 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 280, + 287, + 322 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 280, + 287, + 322 + ], + "spans": [ + { + "bbox": [ + 48, + 280, + 287, + 322 + ], + "type": "text", + "content": "[40] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 1, 3, 6, 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 325, + 287, + 401 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 325, + 287, + 401 + ], + "spans": [ + { + "bbox": [ + 48, + 325, + 287, + 401 + ], + "type": "text", + "content": "[41] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L. Denton, Seyed Kamyar Seyed Ghasemipour, Burcu Karagol Ayan, Seyedeh Sara Mahdavi, Raphael Gontijo Lopes, Tim Salimans, Jonathan Ho, David J. Fleet, and Mohammad Norouzi. Photorealistic text-to-image diffusion models with deep language understanding. ArXiv, abs/2205.11487, 2022. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 404, + 287, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 404, + 287, + 435 + ], + "spans": [ + { + "bbox": [ + 48, + 404, + 287, + 435 + ], + "type": "text", + "content": "[42] Edgar Schonfeld, Bernt Schiele, and Anna Khoreva. A u-net based discriminator for generative adversarial networks. In CVPR, 2020. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 438, + 287, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 438, + 287, + 470 + ], + "spans": [ + { + "bbox": [ + 48, + 438, + 287, + 470 + ], + "type": "text", + "content": "[43] Jianxin Sun, Qiyao Deng, Qi Li, Muyi Sun, Min Ren, and Zhenan Sun. Anyface: Free-style text-to-face synthesis and manipulation. In CVPR, 2022. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 473, + 287, + 506 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 473, + 287, + 506 + ], + "spans": [ + { + "bbox": [ + 48, + 473, + 287, + 506 + ], + "type": "text", + "content": "[44] Ming Tao, Hao Tang, Fei Wu, Xiao-Yuan Jing, Bing-Kun Bao, and Changsheng Xu. Df-gan: A simple and effective baseline for text-to-image synthesis. In CVPR, 2022. 2, 8" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 507, + 287, + 529 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 507, + 287, + 529 + ], + "spans": [ + { + "bbox": [ + 48, + 507, + 287, + 529 + ], + "type": "text", + "content": "[45] Arash Vahdat and Jan Kautz. NVAE: A deep hierarchical variational autoencoder. In NeurIPS, 2020. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 532, + 287, + 563 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 532, + 287, + 563 + ], + "spans": [ + { + "bbox": [ + 48, + 532, + 287, + 563 + ], + "type": "text", + "content": "[46] Aaron van den Oord, Oriol Vinyals, and koray kavukcuoglu. Neural discrete representation learning. In NeurIPS, 2017. 1, 2, 3, 4, 5, 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 566, + 287, + 597 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 566, + 287, + 597 + ], + "spans": [ + { + "bbox": [ + 48, + 566, + 287, + 597 + ], + "type": "text", + "content": "[47] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In NeurIPS, 2017. 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 600, + 287, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 600, + 287, + 632 + ], + "spans": [ + { + "bbox": [ + 48, + 600, + 287, + 632 + ], + "type": "text", + "content": "[48] Haohan Wang, Xindi Wu, Zeyi Huang, and Eric P. Xing. High-frequency component helps explain the generalization of convolutional neural networks. In CVPR, 2020. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 635, + 287, + 677 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 635, + 287, + 677 + ], + "spans": [ + { + "bbox": [ + 48, + 635, + 287, + 677 + ], + "type": "text", + "content": "[49] Zihao Wang, Wei Liu, Qian He, Xinglong Wu, and Zili Yi. Clip-gen: Language-free training of a text-to-image generator with clip. arXiv preprint arXiv:2203.00386, 2022. 1, 2, 3, 6" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 681, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 681, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 681, + 287, + 712 + ], + "type": "text", + "content": "[50] Weihao Xia, Yujiu Yang, Jing-Hao Xue, and Baoyuan Wu. Tedigan: Text-guided diverse face image generation and manipulation. In CVPR, 2021. 2, 6, 8" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 274 + ], + "type": "list", + "angle": 0, + "index": 22, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "text", + "content": "[51] Zhisheng Xiao, Karsten Kreis, Jan Kautz, and Arash Vahdat. Vaebm: A symbiosis between variational autoencoders and energy-based models. In ICLR, 2021. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 108, + 545, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 108, + 545, + 150 + ], + "spans": [ + { + "bbox": [ + 307, + 108, + 545, + 150 + ], + "type": "text", + "content": "[52] Tao Xu, Pengchuan Zhang, Qiuyuan Huang, Han Zhang, Zhe Gan, Xiaolei Huang, and Xiaodong He. Attngan: Fine-grained text to image generation with attentional generative adversarial networks. In CVPR, 2018. 2, 8" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 152, + 545, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 152, + 545, + 195 + ], + "spans": [ + { + "bbox": [ + 307, + 152, + 545, + 195 + ], + "type": "text", + "content": "[53] Jiahui Yu, Xin Li, Jing Yu Koh, Han Zhang, Ruoming Pang, James Qin, Alexander Ku, Yuanzhong Xu, Jason Baldridge, and Yonghui Wu. Vector-quantized image modeling with improved VQGAN. In ICLR, 2022. 5, 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 198, + 545, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 198, + 545, + 239 + ], + "spans": [ + { + "bbox": [ + 307, + 198, + 545, + 239 + ], + "type": "text", + "content": "[54] Yufan Zhou, Ruiyi Zhang, Changyou Chen, Chunyuan Li, Chris Tensmeyer, Tong Yu, Jiuming Gu, Jinhui Xu, and Tong Sun. Lafite: Towards language-free training for text-to-image generation. CVPR, 2022. 2, 8" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 242, + 545, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 242, + 545, + 274 + ], + "spans": [ + { + "bbox": [ + 307, + 242, + 545, + 274 + ], + "type": "text", + "content": "[55] Minfeng Zhu, Pingbo Pan, Wei Chen, and Yi Yang. Dm-gan: Dynamic memory generative adversarial networks for text-to-image synthesis. In CVPR, 2019. 2, 8" + } + ] + } + ], + "index": 21 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "type": "text", + "content": "1745" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Category Query Learning for Human-Object Interaction Classification/c48c8f3f-b487-4c24-815b-6385a57d02e8_content_list.json b/2023/Category Query Learning for Human-Object Interaction Classification/c48c8f3f-b487-4c24-815b-6385a57d02e8_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..59b055db5019635662f16e41711a6d5b38efd5e0 --- /dev/null +++ b/2023/Category Query Learning for Human-Object Interaction Classification/c48c8f3f-b487-4c24-815b-6385a57d02e8_content_list.json @@ -0,0 +1,1862 @@ +[ + { + "type": "text", + "text": "Category Query Learning for Human-Object Interaction Classification", + "text_level": 1, + "bbox": [ + 127, + 130, + 841, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Chi Xie $^{1\\dagger}$ Fangao Zeng $^{2‡}$ Yue Hu $^{3‡}$ Shuang Liang $^{1*}$ Yichen Wei $^{2‡}$ $^{1}$ Tongji University $^{2}$ MEGVII Technology $^{3}$ Shanghai Jiao Tong University $^{1}$ {chixie, shuangliang}@tongji.edu.cn", + "bbox": [ + 173, + 179, + 795, + 228 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "2zfg472988436@163.com,wei_yi_chen@hotmail.com 318671129361@sjtu.edu.cn", + "bbox": [ + 174, + 233, + 810, + 250 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 286, + 312, + 301 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Unlike most previous HOI methods that focus on learning better human-object features, we propose a novel and complementary approach called category query learning. Such queries are explicitly associated to interaction categories, converted to image specific category representation via a transformer decoder, and learnt via an auxiliary image-level classification task. This idea is motivated by an earlier multi-label image classification method, but is for the first time applied for the challenging human-object interaction classification task. Our method is simple, general and effective. It is validated on three representative HOI baselines and achieves new state-of-the-art results on two benchmarks. Code will be available at https://github.com/charles-xie/CQL.", + "bbox": [ + 75, + 319, + 473, + 530 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 563, + 209, + 579 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Human-Object Interaction (HOI) detection has attracted a lot of interests in recent years [3, 8-10, 21, 33]. The task consists of two sub-tasks. The first is human and object detection. It is usually performed by common object detection methods. The second is interaction classification of each human-object (HO) pair. This sub-task is very challenging due to the complex appearance variations in the interaction categories. See Fig. 1 for examples. It is the focus of most previous HOI methods, as well as this work.", + "bbox": [ + 75, + 590, + 468, + 724 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Most previous HOI methods focus on learning better human-object features, including modeling relation and context via GNN [7, 29, 34, 37] or attention mechanism [8, 34, 46], decoupling localization and classification [22, 41, 48], leveraging vision-language knowledge [6, 22] and introducing multi-scale feature to transformer [16]. However, for interaction classification they all adopt the simple linear classifier that performs the dot product of the", + "bbox": [ + 75, + 727, + 468, + 849 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/3075a4e4a0d7b0b2428e126ff8afd3b3dc0a7619be00fad2022df5f19c9f8ac1.jpg", + "image_caption": [ + "(a) person fly kite" + ], + "image_footnote": [], + "bbox": [ + 501, + 284, + 694, + 385 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/34ce8e97a2ba6758f9ba26c1eaf19d64dcf450e62bd5da932f9f922319a890f2.jpg", + "image_caption": [ + "person fly airplane" + ], + "image_footnote": [], + "bbox": [ + 697, + 285, + 890, + 385 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/02d5d4e4abf81c7fe8dd3fb7833919f4d4f4db83d6e10ced0f40417a6658b050.jpg", + "image_caption": [ + "(b) person hold fork" + ], + "image_footnote": [], + "bbox": [ + 501, + 397, + 694, + 497 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/afe9e0d7b15e9a43e4021404cf5e3e203cf80b73db034c463fefc552f98f45e0.jpg", + "image_caption": [ + "person hold elephant", + "Figure 1. Interaction classification is inherently challenging. In (a), \"fly\" is semantically polysemic, resulting in different objects, poses and relative positions. In (b), when \"hold\" is associated with different objects, the appearance, scene background, and human poses are largely different." + ], + "image_footnote": [], + "bbox": [ + 697, + 397, + 890, + 497 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "human-object feature and a static weight vector, which represents an interaction category.", + "bbox": [ + 498, + 613, + 890, + 643 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this work, we propose a new approach that enhances the above paradigm and complements most previous HOI methods. It is motivated by the recent work Query2label [25], a transformer-based classification network. It proposes a new concept we call category-specific query. Unlike the queries in other transformer methods, each query is associated to a specific and fixed image category during training and inference. This one-to-one binding makes the query learn to model each category more effectively. The queries are converted to image specific category representations via a transformer decoder. This method achieves excellent performance on multi-label image classification task.", + "bbox": [ + 496, + 643, + 892, + 824 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We extend this approach for human-object interaction classification. Essentially, our approach replaces traditional category representation as a static weight vector in previous HOI methods with category queries learnt as described above. The same linear classifier is adopted. Such cate", + "bbox": [ + 496, + 825, + 893, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author.", + "bbox": [ + 93, + 862, + 222, + 875 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "$\\dagger$ Work done during internship at MEGVII technology.", + "bbox": [ + 96, + 875, + 380, + 887 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "$\\ddagger$ Work done while worked at MEGVII.", + "bbox": [ + 96, + 887, + 303, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "15275", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "gory queries are more effective, and adaptive for different images, giving rise to better modeling of the complex variations in each interaction category. This is the crucial difference between this work and a simple adaption of [25] to HOI. Notably, this work is the first to address the category weight representation problem in the HOI community.", + "bbox": [ + 75, + 90, + 467, + 181 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Note that our proposed category specific query is different and not related to those queries in other transformer-based HOI methods [4, 15, 33, 49]. Specifically, category queries extract image-level features as the category representation. The queries in other methods are human-object instance-level features and category-agnostic.", + "bbox": [ + 75, + 181, + 467, + 273 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our method is simple, lightweight and general. The overview is in Fig. 2. It is complementary to any off-the-shelf HOI method that provides human-object features. The modification of both inference and training is small. The incurred additional cost is marginal.", + "bbox": [ + 75, + 273, + 467, + 349 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In experiments, our approach is validated on three representative and strong HOI baseline methods, two transformer-based methods [22, 33] and a traditional two-stage method [42]. They are all significantly improved by our approach. New state-of-the-art results are obtained on two benchmarks. Specifically, we obtain $36.03\\mathrm{mAP}$ on HICO-DET. Comprehensive ablation studies and in-depth discussion are also provided to verify the effectiveness of implementation details in our approach. It turns out that our method is more effective on challenging images that contain more human-object instances, a property that is rarely discussed by previous HOI methods.", + "bbox": [ + 75, + 351, + 467, + 531 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 76, + 547, + 218, + 564 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Instance Query Learning in HOI Detection", + "text_level": 1, + "bbox": [ + 76, + 574, + 441, + 589 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "DETR [2] firstly proposes the concept of object instance query for object detection task. Such queries essentially learn the priors of both object appearance and spatial location. DETR leverages those queries to probe image features through a transformer [35] and localize unique objects in the image. Motivated by its great success, many works [15, 22, 33, 41, 48, 49] adapt such detection transformer framework to HOI detection by simply treating the HOI triplet [33, 49] or H-O pair [4, 15] as an object. A few of them [6, 30, 45] pay attention to adapting the plain query to this task. DOQ [30] proposes a knowledge distillation model using oracle queries to facilitate the representation learning of a transformer-based detector; HQM [45] explicitly constructs hard positive queries from ground truth to train the model to be less vulnerable to spacial variations; CATN [6] utilizes the object category prior generated from external object detector and language model for query initialization. In summary, these transformer-based methods use each query to aggregate context information not restricted to one interaction category, in order to predict", + "bbox": [ + 75, + 598, + 467, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "a potential HOI instance at a specific location.", + "bbox": [ + 500, + 90, + 803, + 104 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In DETR [2] and its variants [4,15,22,33,41,49] in HOI, as the queries are category-agnostic, their association to object categories are dynamic and unstable during training. This could be problematic. For example, it is well known that the convergence of DETR training is slow. In contrast, our proposed query is category-specific. The learning is guided by image-level classification task and stable. Such queries learn category-specific priors and are good representation for interaction categories.", + "bbox": [ + 496, + 106, + 890, + 243 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Feature Learning in HOI Detection", + "text_level": 1, + "bbox": [ + 500, + 255, + 805, + 271 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Early methods. Based on two-stage detection framework, early works make many efforts to help feature learning, including employing architectures effective in modeling relation and context like GNN [7, 29, 34, 37] and attention module [8, 34, 46], leveraging fine-grained visual features [11, 17, 19, 20, 36] like human pose and introducing language prior [1, 7, 17, 27, 46].", + "bbox": [ + 496, + 279, + 890, + 385 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Transformer-based methods. Motivated by DETR [49], many methods [4, 15, 33, 49] leverage transformer architecture [35] and extend the object query in DETR to HOI query. With the help of HOI query and transformer's built-in attention mechanism, those methods learn effective feature representation for HOI triplet or H-O pair.", + "bbox": [ + 496, + 386, + 890, + 477 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Based on those pioneer transformer-based methods, recently, many methods are proposed to further help feature learning, by decoupling H-O pair localization and interaction classification [22, 41, 48] or exploiting multi-scale feature in transformer architecture [16]. Some works [6, 22] leverage vision-language knowledge in CLIP [31] or design a pretrained model [40] specifically for HOI; others utilize information like human poses [39] or spatial configurations [13] that has been used in early HOI detectors.", + "bbox": [ + 496, + 478, + 890, + 613 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Relation to the proposed method. Previous methods learn the H-O feature while ours learns the category query as the category representation feature. Thus, they are complementary. The interaction classification is simply by measuring the similarity between the two types of features. The integration of our method to previous HOI methods is simple.", + "bbox": [ + 496, + 614, + 890, + 705 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3. Our Method", + "text_level": 1, + "bbox": [ + 500, + 722, + 629, + 737 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The overview of our method is in Fig. 2. It consists of two components, the image-level category query learning (top block) and human-object interaction classification (bottom right block).", + "bbox": [ + 496, + 748, + 890, + 808 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The first component is detailed in Sec. 3.1. It is briefly summarized here. A number of queries (embedding vectors) are associated to human-object interaction category, in a one-to-one manner. Such queries interact with image features (provided by a baseline HOI method) through a transformer decoder [35] and become image-specific queries.", + "bbox": [ + 496, + 809, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "15276", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/a2e1ee732a1a14a4cdac05c06211e634cc8a176b620996b5cf18b6d23f76a740.jpg", + "image_caption": [ + "Figure 2. Overview of our method. It consists of two components (top and bottom right). It can be integrated with any baseline HOI method (bottom left) that provides image feature $I$ and human-object instance features $F_{i}$ . See Sec. 3 for details." + ], + "image_footnote": [], + "bbox": [ + 122, + 89, + 849, + 330 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Learning of both the queries and decoder weights is supervised by an auxiliary image-level classification task. In this way, the queries are learnt to capture category-specific feature and become good feature representation for these categories. Besides some minor details, this step is the same as the previous work Query2label [25], which is for multi-label image classification task.", + "bbox": [ + 75, + 388, + 468, + 492 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The second component is detailed in Sec. 3.2. For the first time, we adopt the category query learning method for human-object interaction classification tasks. The cosine similarity between the category query and human-object feature is used for interaction classification. Thus, it works with any HOI method that provides human-object features. Besides, the image-level classification results turn out moderately helpful in an score integration step, which is an extra technique that benefits the performance.", + "bbox": [ + 75, + 493, + 470, + 628 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Overall, the proposed method is simple, effective, lightweight and general. It can be combined with most previous HOI detection methods (bottom left block in Fig. 2), with small modification, as elaborated in Sec. 4.", + "bbox": [ + 75, + 628, + 468, + 689 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Image-level Category Query Learning", + "text_level": 1, + "bbox": [ + 76, + 700, + 406, + 717 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Similar as in Query2Label [25], for $K$ human-object interaction categories, we define their one-to-one corresponding category queries, which are learnable embedding vectors, $\\{Q_k\\} \\in \\mathbb{R}^{K\\times D}$ , where $D$ is the vector dimension.", + "bbox": [ + 75, + 723, + 468, + 782 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Each query $Q_{k}$ aggregates image features $\\mathbf{I} \\in \\mathbb{R}^{H \\times W \\times D}$ through a transformer decoder and is updated to image specific query $Q_{k}'$ ,", + "bbox": [ + 75, + 782, + 468, + 830 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\left\\{Q _ {k} ^ {\\prime} \\right\\} = \\operatorname {d e c o d e r} \\left(\\left\\{Q _ {k} \\right\\}, I\\right). \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 178, + 840, + 468, + 859 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Note that the decoder structure has several variants, which are studied in [5]. Our experiments show that the", + "bbox": [ + 76, + 869, + 470, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "structure is of minor importance, as discussed in Sec. 5.4. Specifically, our decoder consists of two layers, each of which consisting of a cross-attention layer, a self-attention layer and a FFN layer, in order.", + "bbox": [ + 496, + 388, + 892, + 446 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Then, each image-level classification probabilities $p_k$ is computed by applying a category-specific fully-connected layer and a sigmoid activation on the updated query $Q_{k}^{\\prime}$ .", + "bbox": [ + 496, + 448, + 893, + 494 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\np _ {k} = \\operatorname {s i g m o i d} \\left(\\mathrm {F C} \\left(Q _ {k} ^ {\\prime}\\right)\\right) \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 611, + 510, + 890, + 527 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Learning of the category query $\\{Q_k\\}$ and the decoder weights is supervised by common image classification losses. To deal with the label imbalance problem, focal loss [23] and asymmetric loss (ASL) [32] are used. Asymmetric loss is a variant of focal loss. It is more robust for high label imbalance and noises. Our experiments (see Sec. 5.4) show that it is slightly better.", + "bbox": [ + 496, + 535, + 892, + 641 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Specifically, with the classification probability $p_k$ and the shifted probability $p_k' = \\max(p_k - m, 0)$ , the asymmetric loss is", + "bbox": [ + 496, + 641, + 893, + 685 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {i m g} = \\frac {1}{K} \\sum_ {k = 1} ^ {K} \\left\\{ \\begin{array}{l l} (1 - p _ {k}) ^ {\\gamma +} \\log \\left(p _ {k}\\right), & y _ {k} = 1, \\\\ \\left(p _ {k} ^ {\\prime}\\right) ^ {\\gamma -} \\log \\left(1 - p _ {k} ^ {\\prime}\\right), & y _ {k} = 0, \\end{array} \\right. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 511, + 696, + 890, + 738 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where the binary label $y_{k}$ indicates the existence of category $k$ in the image, and $\\gamma +, \\gamma -$ as well as $m$ are hyperparameters. We use the default values in ASL [32], $\\gamma + = 0$ , $\\gamma - = 4$ and $m = 0.05$ .", + "bbox": [ + 496, + 750, + 890, + 808 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this way, the category queries are learnt to encode the category priors. Figure 3 is the visualization of the heatmaps in the cross-attention layer of the decoder. Each category query learns to locate the human body parts related to discriminative feature of its corresponding interaction category, e.g., in Fig. 3c, the query of \"hold\" highlights", + "bbox": [ + 496, + 810, + 892, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "15277", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "the hand region, while in the same image the query of \"ride\" highlights the foot region. It qualitatively demonstrates that the query learning is effective in encoding category-specific information.", + "bbox": [ + 75, + 90, + 468, + 151 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The updated queries adaptively extract category-related features for each image, with the help of the transformer decoder's built-in multi-head cross-attention layer.", + "bbox": [ + 76, + 154, + 470, + 200 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Interaction Classification with Category Query", + "text_level": 1, + "bbox": [ + 76, + 217, + 467, + 233 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this step, we apply the updated category queries $\\{Q_k'\\}$ as the weights for interaction classification. Given the $i$ -th human-object instance, its classification probability score for $k$ -th interaction category is simply the cosine similarity between its feature $F_i$ and the category query feature $Q_k'$ .", + "bbox": [ + 75, + 243, + 467, + 320 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\ns _ {i, k} = \\operatorname {s i g m o i d} \\left(\\frac {Q _ {k} ^ {\\prime} \\cdot F _ {i}}{\\left| Q _ {k} ^ {\\prime} \\right| \\times \\left| F _ {i} \\right|}\\right). \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 171, + 340, + 468, + 375 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "There are no restrictions or assumptions on the human-object feature $F_{i}$ . Most previous HOI methods should be applicable here.", + "bbox": [ + 75, + 391, + 468, + 436 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this step, the traditional classification weight from static parameters is replaced with the category query adaptive on each image. This behavior is essentially different from [25], which uses category queries as image feature.", + "bbox": [ + 75, + 439, + 468, + 500 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "By \"adaptive\", we mean that the queries are updated dynamically according to the image contents. As exemplified in Fig. 3a and Fig. 3c, the queries of \"hold\" learn to highlight different interactive areas in different images and update themselves with features from these areas via attention. This shows the query learning is adaptive to image.", + "bbox": [ + 75, + 502, + 468, + 593 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As discussed in Sec. 5.4, the classification step in Eq. (4) is crucial to make the category query learning effective on human-object interaction classification. Without this step, the image-level category query learning using only image-level classification from Query2label [25] is of little use.", + "bbox": [ + 75, + 595, + 468, + 672 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Score integration step. The segmentation method in [12] discards certain categories during pixel classification that have low image classification scores. Motivated by this method, we take a similar score integration step. The image-level classification score $\\{p_i\\}$ is used to enhance the human-object instance classification. The idea is that, the instance score should be higher if the image-level probability is higher. Our implementation is similar as in [12]. During both training and testing, for each image, the top- $\\kappa$ categories ( $\\kappa = 70$ in this work) with higher image classification scores $\\{p_i\\}$ are selected. The instance score $s_{i,k}$ is slightly modulated such that it becomes higher if the rank of category $k$ is higher. This strategy gives rise to moderate improvement, as verified in Tab. 4. We left the implementation details in the supplementary materials.", + "bbox": [ + 75, + 674, + 468, + 900 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/99f8282d3ac2f3524a9f012e38cb046233dabf2bf8ff19f79576baca3c9a90b0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 501, + 88, + 629, + 155 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/fe379372a13e190926267d60ca26365d5a1c8c3fdec1a768296f4aba55d26d7c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 632, + 88, + 759, + 155 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/2b0bc16ea2c372fd882e3a9223f14266858b79022b8fc3d45678b1c1a80631f0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 761, + 88, + 890, + 155 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/0c455d5cab7e46117e8c591ebceb56b8f57bafb24bc6662232973cbab7dab19f.jpg", + "image_caption": [ + "(a) Image with \"person-break-baseball bat\" and \"hold-baseball bat\"." + ], + "image_footnote": [], + "bbox": [ + 501, + 174, + 629, + 241 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/51e2755600f28b9ba7c8639160c56def36c898efa95c3d24d8d9c08303494f85.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 632, + 174, + 759, + 241 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/9647b202997c06d7eaa3c35abeecd263747d3dd6ee99e228e39f754859e401d7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 761, + 174, + 890, + 241 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/8f7492dcb75316b0373578dff4aa2d091da4a403ff996384eacc1f96bf4dd570.jpg", + "image_caption": [ + "(b) Image with \"person-direct-airplane\" and \"fly-airplane\".", + "(c) Image with \"person-hold-horse\" and \"ride-horse\".", + "Figure 3. From left to right: image, attention maps of the cross-attention layer in the decoder for different interaction categories." + ], + "image_footnote": [], + "bbox": [ + 501, + 260, + 629, + 327 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/9ceaf0fc4813a46f45b113af842239ef578ebfa73b0cc7aa99f6cd2a4f59e989.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 632, + 260, + 759, + 327 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/6aa65ad1b0f1578a4bcc497a7396b04fb3d878abc1f22548143f7c6548f7f973.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 761, + 260, + 890, + 327 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Integration to Off-the-shelf HOI Detectors", + "text_level": 1, + "bbox": [ + 500, + 402, + 879, + 420 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As shown in Fig. 2, our method is ready to integrate with any baseline HOI method that provides image feature $\\mathbf{I}$ and human-object instance feature $\\{F_i\\}$ . The integration is simple. During inference, the human-object instance interaction classification part is replaced by our method in Sec. 3.2, the top and bottom right block in Fig. 2.", + "bbox": [ + 496, + 429, + 890, + 520 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "During training, the original loss in the baseline HOI method $\\mathcal{L}_{base}$ is added to our image classification loss in Eq. (3). The final loss $\\mathcal{L}$ for training is", + "bbox": [ + 496, + 520, + 890, + 566 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\mathcal {L} _ {\\text {b a s e}} + \\lambda * \\mathcal {L} _ {\\text {i m g}}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 614, + 584, + 890, + 601 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where the weight $\\lambda$ is 1.0 by default. All other hyper parameters and details during training remain the same as in the baseline HOI method.", + "bbox": [ + 496, + 609, + 890, + 654 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Thus, our method is general and applicable to most existing HOI methods. In this work, we select three representative yet different baseline methods to verify the effectiveness of our approach, as described below.", + "bbox": [ + 496, + 656, + 890, + 715 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "QPIC [33] is the first to introduce transformer method into HOI task. It is also the baseline for many recent works [13, 28, 41, 45, 48]. Its performance is much better than early one-stage [14, 21, 38] and two-stage [7, 11, 17] methods while keeping a simple and end-to-end architecture. It consists of a CNN backbone as well as a transformer encoder and decoder.", + "bbox": [ + 496, + 717, + 890, + 821 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "During our integration, the feature map in its transformer encoder is used as the image feature $\\mathbf{I}$ . The human-object feature $\\{F_i\\}$ is the query feature in its decoder.", + "bbox": [ + 496, + 824, + 890, + 869 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "SCG [42] is a traditional two-stage method and the best in this category. It is also one of the best methods that", + "bbox": [ + 496, + 869, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "15278", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/4df4208d2accdb28e88a97e882663661a848ff76439b76ecb20c1cedb5bfd21b.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodPipelineE2EHICO-DETV-COCOEfficiency
FullRareNon-RareS1S2#ParamsFPS
QPIC [33]transformer✓28.9321.6231.1261.3963.6541M19.5
+ Ourstransformer✓31.08(+2.15)23.9033.2263.67(+2.28)65.4946M(+5M)18.3(-6.2%)
SCG [42]two-stageX31.2824.1633.4056.9362.5157M4.5
+ Ourstwo-stageX32.74(+1.46)26.2534.6859.14(+2.21)65.6164M(+7M)4.1(-8.9%)
GEN-VLKT [22]transformerX33.6929.9434.8164.8966.7442M21.7
+ OurstransformerX35.36(+1.67)32.9736.0766.40(+1.51)69.1747M(+5M)20.6(-5.1%)
", + "bbox": [ + 78, + 88, + 893, + 205 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Table 1. The performance numbers of three different baseline HOI methods with and without integration of our method, on two datasets. \"E2E\" denotes whether a HOI detector is end-to-end. All models are tested on Tesla V100.", + "bbox": [ + 75, + 214, + 890, + 244 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "does not use transformer. It uses a multi-stream graph neural network(GNN) for interaction classification. In our experiment, the detection boxes are from a fine-tuned detector provided by DRG [7] for HICO-DET and a fine-tuned DETR for V-COCO.", + "bbox": [ + 75, + 258, + 468, + 333 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "During our integration, the CNN feature map in the backbone of SCG is used as image feature $\\mathbf{I}$ . The human-object feature $\\{F_i\\}$ is generated through RoI pooling with detected human and object boxes and fused with the GNN.", + "bbox": [ + 75, + 335, + 468, + 396 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "GEN-VLKT [22] is also transformer-based, but not end-to-end as pairwise NMS [41] is used for post-processing. It is the current state-of-the-art method. It uses two parallel decoders for object detection and interaction classification, namely instance decoder and interaction decoder.", + "bbox": [ + 75, + 397, + 468, + 472 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "During our integration, the feature map in its transformer encoder is used as image feature $\\mathbf{I}$ . The query feature in the interaction decoder is used as the human-object feature $\\{F_i\\}$ . Note that, unlike the majority of HOI detection methods, the original GEN-VLKT uses HOI categories rather than interaction categories during interaction classification. Our experiments still use interaction categories, in order to be consistent with most other methods.", + "bbox": [ + 75, + 474, + 468, + 594 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 76, + 611, + 209, + 628 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this section, we verify the applicability and effectiveness of the proposed method through experiments. In Sec. 5.1, we introduce the experimental settings. Then we demonstrate the effectiveness of the proposed method over 3 baselines in Sec. 5.2, and show it achieves SOTA results on major benchmarks in Sec. 5.3. Next, in Sec. 5.4 we conduct comprehensive ablation studies on the key components as well as detailed technical designs. Lastly, we provide some analysis and visualization in Sec. 5.5.", + "bbox": [ + 75, + 637, + 468, + 773 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.1. Datasets", + "text_level": 1, + "bbox": [ + 76, + 785, + 179, + 800 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "HICO-DET [3] and V-COCO [10] are two widely-used HOI benchmarks. HICO-DET contains 47,776 images, with 38,118 for training and 9,658 for testing. There are 600 HOI categories in HICO-DET, consisting of 117 interaction classes and 80 object classes. Each HOI category is composed of an interaction and an object. V-COCO is", + "bbox": [ + 75, + 809, + 468, + 902 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "a subset of MS-COCO [24] with HOI annotations, including 10,346 images (2,533 for training, 2,867 for validation and 4,946 for testing). It has 80 object categories same with HICO-DET and 29 interaction categories.", + "bbox": [ + 496, + 258, + 890, + 319 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Evaluation metrics. For HICO-DET, we adopt the commonly used mAP metric [3]. Each prediction is a $\\langle$ human, interaction, object $\\rangle$ triplet. A prediction is a true positive only when the human and object bounding boxes both have IoU $>0.5$ w.r.t. ground truth and the interaction classification result is correct. We evaluate the performance in two different settings following [3]. In the known object setting, for each HOI category, we evaluate the prediction only on the images containing the target object category. In default setting, the detection result of each category is evaluated on the full test set. In each setting, we report the mAP over (1) all 600 HOI categories (Full), (2) 138 categories with less than 10 training samples (Rare), and (3) the remaining 462 categories (Non-rare). For V-COCO, we use the role mAP following [10], under both scenario #1 (including objects) and #2 (ignoring objects). The performance is evaluated using its official evaluation toolkit.", + "bbox": [ + 496, + 320, + 892, + 577 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.2. Improvement on Three Different Baselines", + "text_level": 1, + "bbox": [ + 498, + 588, + 862, + 604 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Tab. 1 summarizes the performance of the three baseline HOI methods before and after integration of our method. Backbone is ResNet50. All these methods are significantly improved. Specifically, QPIC [33] is improved by $2.15\\mathrm{mAP}$ , making it competitive with those more recent works [16, 41, 43]. SCG [42] is improved by 1.46 mAP, demonstrating that our method is not limited to transformer-based baselines. The current SOTA method GEN-VLKT [22] is improved by $1.67\\mathrm{mAP}$ , producing the new SOTA result (also refer to Tab. 2).", + "bbox": [ + 496, + 612, + 890, + 762 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "On the V-COCO dataset [10], the performance improvement is similar, which is 2.28, 2.21 and $1.51\\mathrm{mAP}$ on QPIC, SCG and GEN-VLKT, respectively.", + "bbox": [ + 496, + 763, + 890, + 808 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Notably, for SCG [42], as the object detector is fixed during the training of its interaction classification network, the improvement by our method is purely due to better interaction classification, not a better fine-tuned CNN backbone or a better object detector. This further consolidates that the category query learning is effective.", + "bbox": [ + 496, + 810, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "15279", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/d7771f26a4592a7996c1641309e6c92b449727c80d684243974858a5bc968310.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodDetectorBackboneDefaultKnown Object
FullRareNon-rareFullRareNon-rare
DRG [7]HICO-DETResNet50-FPN24.5319.4726.0427.9823.1129.43
GG-Net [47]HICO-DETHourglass10423.4716.4825.6027.3620.2329.48
IDN [18]HICO-DETResNet5026.2922.6127.3928.2424.4729.37
QPIC [33]HICO-DETResNet5029.0721.8531.2331.6824.1433.93
SCG [42]HICO-DETResNet50-FPN31.3324.7233.3134.3727.1836.52
CDN [41]HICO-DETResNet5031.7827.5533.0534.5329.7335.96
DT [48]HICO-DETResNet5031.7527.4533.0334.5030.1335.81
STIP [44]HICO-DETResNet5031.6027.7532.7534.4130.1235.69
HQM [45]HICO-DETResNet5032.4728.1533.76---
MSTR [16]HICO-DETResNet5031.1725.3132.9234.0228.8335.57
RLIP [40]COCO+VGResNet5032.8426.8534.63---
IF [26]HICO-DETResNet5033.5130.3034.4636.2833.1637.21
GEN-VLKT-B [22]HICO-DETResNet5033.7529.2535.1036.7832.7537.99
GEN-VLKT-M [22]HICO-DETResNet10134.7831.5035.7738.0734.9439.01
GEN-VLKT-L [22]HICO-DETResNet10134.9531.1836.0838.2234.3639.37
BodyPartMap [39]HICO-DETResNet5035.1533.7135.5837.5635.8738.06
GEN-VLKT-B + OursHICO-DETResNet5035.3632.9736.0738.4334.8539.50
GEN-VLKT-M + OursHICO-DETResNet10135.8332.9136.7038.7935.2839.84
GEN-VLKT-L + OursHICO-DETResNet10136.0333.1636.8938.8235.5139.81
", + "bbox": [ + 158, + 88, + 812, + 375 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/d8a4328f669a88eec8d47905c521006b6c621c1b6c9ebfb69415047c9bbf3942.jpg", + "table_caption": [ + "Table 2. The proposed method achieves state-of-the-art on HICO-DET [3]. The best results are marked in bold." + ], + "table_footnote": [], + "table_body": "
MethodBackboneScenario #1Scenario #2
DRG [7]R50FPN51.0-
SCG [42]R5054.260.9
GG-Net [47]HG10454.7-
QPIC [33]R5058.861.0
HQM [45]R5063.6-
CDN [41]R5061.763.8
GEN-VLKT-B [22]R5062.464.5
GEN-VLKT-M [22]R10163.365.6
GEN-VLKT-L [22]R10163.665.9
MSTR [16]R5062.065.2
BodyPartMap [39]R5063.065.1
IF [26]R5063.065.2
DT [48]R5066.268.5
STIP [44]R5065.169.7
GEN-VLKT-B + OursR5066.469.2
GEN-VLKT-M + OursR10166.869.8
GEN-VLKT-L + OursR10166.569.9
", + "bbox": [ + 78, + 412, + 468, + 667 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To verify that the performance improvement is not due to a larger model, we also compare the model size and running speed. Our method increases the model parameters by a few millions, which is small compared to the original model size. The running speed, measured by FPS, is only decreased by a few percent. The marginal additional cost shows that our method is quite lightweight.", + "bbox": [ + 75, + 724, + 470, + 833 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.3. Comparison with State-of-the-art", + "text_level": 1, + "bbox": [ + 76, + 845, + 370, + 862 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Tab. 2 and Tab. 3 compare our method with many previous methods for HICO-DET and V-COCO datasets, respec", + "bbox": [ + 75, + 869, + 470, + 902 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/9aa52c15fd14b12f7e1c2ea21a9b26c1a9c2531c6e57ce051540a806c3a5b0ad.jpg", + "table_caption": [ + "Table 3. Comparison with state-of-the-art methods on V-COCO [10] dataset. The best results are marked in bold." + ], + "table_footnote": [], + "table_body": "
C1C2C3FullRareNon-Rare
a---33.6929.9434.81
b✓--33.86 (+0.17)31.1234.68
c✓✓-34.98 (+1.29)31.7335.95
d✓✓✓35.36 (+1.67)32.9736.07
", + "bbox": [ + 500, + 412, + 890, + 491 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 4. Ablation study of several variants of our method, starting from the baseline (a) to our approach (d). C1, C2 and C3 are described in Sec. 5.4. The best results are marked in bold.", + "bbox": [ + 498, + 502, + 890, + 544 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "tively. GEN-VLKT [22] is used as our baseline.", + "bbox": [ + 498, + 561, + 816, + 575 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "On HICO-DET, our result with ResNet50 backbone already outperforms all previous methods under both default and known object settings. With the stronger ResNet101 backbone, our method achieves the new state-of-the-art 36.03 full mAP under default settings and 38.82 under known object settings.", + "bbox": [ + 496, + 578, + 890, + 667 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "On V-COCO dataset, our method achieves the new state-of-the-art performance on Scenario 1, with an AP of 66.4 for ResNet50, surpassing [48]. For scenario 2, it is comparable with the state-of-the-art [44].", + "bbox": [ + 496, + 669, + 890, + 729 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.4. Ablation Experiments", + "text_level": 1, + "bbox": [ + 500, + 739, + 705, + 757 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We perform various ablation experiments to validate the effectiveness of different components in our method. HICO-DET dataset and GEN-VLKT [22] baseline are used.", + "bbox": [ + 496, + 763, + 890, + 809 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The proposed method can be divided into 3 components: C1 means applying Query2Label [25] to the baseline detector as a multi-task learning (with feature extractor shared). In detail, it adds the queries, the decoder and image classification loss (Eq. (1), Eq. (2) and Eq. (5)). C2 means using the learned query in C1 as adaptive interaction classification", + "bbox": [ + 496, + 810, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "15280", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/7fd5c2aae2cdb3d7c3b031283727ab588c36474d5764b26dc06f27430412bfc8.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
loss typeλFullRareNon-Rare
-034.2130.1535.42
focal loss [23]0.534.4331.0635.44
1.034.5131.0835.53
1.534.3531.5435.19
2.034.2931.1835.22
ASL [32]0.534.5731.9135.36
1.034.9831.7335.95
1.534.7732.0835.57
2.034.4131.9235.15
", + "bbox": [ + 96, + 88, + 450, + 244 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/32a48e79528c09fcf51ce5510070690cb016cb204a800d90beea99164b411da3.jpg", + "table_caption": [ + "Table 5. Ablation on the type and weight $\\lambda$ of the image classification loss. The best results are marked in bold." + ], + "table_footnote": [], + "table_body": "
Layer structureFullRareNon-Rare
S→C→F34.7332.0935.52
C→S→F34.9831.7335.95
C→F34.6331.3435.61
", + "bbox": [ + 112, + 287, + 433, + 353 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/b2518088af689198302fcf2598dafda5211319cb939bf809da95ffd2080271bb.jpg", + "table_caption": [ + "Table 6. Ablation on structure of each layer in the category decoder. Here \"S\", \"C\" and \"F\" stands for the self-attention, cross-attention and FFN in a standard transformer [35]." + ], + "table_footnote": [], + "table_body": "
LFullRareNon-Rare
134.6631.4735.61
234.9831.7335.95
334.8631.5635.85
", + "bbox": [ + 158, + 411, + 388, + 474 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 7. Ablation on the number of layers in the category decoder.", + "bbox": [ + 76, + 486, + 467, + 500 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "weight (Eq. (4)). This is the key component of the proposed method, which makes a distinction between the proposed method and a simple adaption of Query2label [25] to HOI task. C3 denotes the score integration step in Sec. 3.2.", + "bbox": [ + 76, + 518, + 468, + 579 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To validate our approach in Fig. 2, several variants with these components in our approach are experimented and summarized in Tab. 4. First, (b) is a simple combination of Query2label [25] and the baseline detector (a) in a multitask setting. (b) is only slightly better than (a), showing that simply applying Query2Label to HOI is barely helpful.", + "bbox": [ + 76, + 580, + 468, + 672 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Second, variant (c) significantly boosts (b), indicating using the queries as adaptive classification weights is the key to the performance improvement. This shows the effectiveness of our most crucial technical design in C2: applying the learned category queries as adaptive interaction classification weights.", + "bbox": [ + 76, + 672, + 468, + 762 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Finally, the complete approach is (d). It further adds the integration step of image classification and instance classification score on (c). This technique produces moderate improvement over (c), i.e., $1.67\\mathrm{mAP}$ vs. $1.29\\mathrm{mAP}$ .", + "bbox": [ + 76, + 763, + 468, + 823 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Image classification loss. Tab. 5 compares different loss functions and weights. First, when $\\lambda = 0$ , which means no image-level supervision is applied, the improvement over the baseline (33.69 mAP) drops to only 0.52 mAP. This demonstrates the image classification supervision is essen", + "bbox": [ + 76, + 824, + 468, + 898 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/827f372f985eaa1b832f1f34b3aaa43a0ee9b577522936d14a3ed751e170a219.jpg", + "image_caption": [ + "Figure 4. Some qualitative comparison between the baseline and the proposed method on HICO-DET. From left to right, column 1: true positive (TP) detection results, whose interaction score is increased by our method; column 2: false positive (FP) detection results, whose interaction score is decreased by our method; column 3: corresponding image-level GT and predictions by our method. Scores on the left and right of an image are the interaction classification scores of the visualized instance from the baseline and our method. Best viewed in color. More in supplementary materials." + ], + "image_footnote": [], + "bbox": [ + 506, + 92, + 888, + 356 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "tial to make the category query learning effective, with either focal loss or ASL. Additionally, ASL is slightly better than focal loss. By default, $\\lambda = 1.0$ is adopted.", + "bbox": [ + 498, + 513, + 890, + 559 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Is asymmetric loss the key? In the ablation above, we can see that ASL does help our image-level query learning. However, it is not the major reason for the performance improvement. To figure out this, we replace focal loss in the plain baseline GEN-VLKT with ASL and the result is only slightly better by $0.08\\mathrm{mAP}$ .", + "bbox": [ + 498, + 561, + 890, + 651 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Decoder structure. Tab. 6 compares several structures of the decoder. Compared to the standard decoder [2, 35] $(\\mathrm{S} \\rightarrow \\mathrm{C} \\rightarrow \\mathrm{F}$ in the table), putting cross-attention first $(\\mathrm{C} \\rightarrow \\mathrm{S} \\rightarrow \\mathrm{F})$ is slightly better (by $0.25 \\mathrm{mAP}$ ) without extra computation. If we remove self-attention $(\\mathrm{C} \\rightarrow \\mathrm{F})$ , the performance drops by $0.35 \\mathrm{mAP}$ compared with the $\\mathrm{S} \\rightarrow \\mathrm{C} \\rightarrow \\mathrm{F}$ setting. This is probably because self-attention helps to learn the dependencies between different category queries.", + "bbox": [ + 498, + 652, + 890, + 772 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Tab. 7 compares different numbers of decoder layers, denoted as $L$ . We find that $L = 2$ is sufficient. More layers do not help the performance.", + "bbox": [ + 498, + 773, + 890, + 819 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.5. Discussions and Analysis", + "text_level": 1, + "bbox": [ + 500, + 830, + 725, + 847 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To understand why the category query learning is effective for human-object interaction classification, we provide some analysis and qualitative results.", + "bbox": [ + 498, + 854, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "15281", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/199b250f04f984c175003beebe08d073e57ff36e116e51d60d06aaf1e7eeec19.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 102, + 93, + 442, + 210 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/2e06852b319c9a0790037fa9831788f097d23b9a7f7b6c7bb0d8c4d1db475004.jpg", + "image_caption": [ + "Figure 5. Performance evaluated on image partitions with different $n = 1,2,3,4,5, > 5$ instances for an interaction category. Top: mAP numbers for three baselines (dashed) as well as our integration (solid). Bottom: relative mAP ratios for improvement produced by our method (solid) and between two arbitrary previous methods (dashed)." + ], + "image_footnote": [], + "bbox": [ + 101, + 217, + 439, + 333 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The attention maps in the cross-attention layer of the decoder is visualized in Fig. 3. For different category queries, the corresponding attention maps show they learn to capture the semantics of the category, while being adaptive to different images. For example, in Fig. 3a, the broken part of the bat in the air is highlighted for \"break\" while the left part in the hand is highlighted for \"hold\". This is similar for Fig. 3b. In Fig. 3c, the attention map highlights many instances with corresponding action \"hold\" and \"ride\".", + "bbox": [ + 75, + 441, + 467, + 578 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To qualitatively demonstrate how our method helps, we visualize some cases of the baseline and the proposed method in Fig. 4. In the first case, the baseline predicts a TP of \"person sit on bus\" with a low score 0.15, and a FP of \"person board bus\" with a high score 0.29. Our model also predicts the same TP and FP, but lowers the FP score to 0.07 and lifts the TP score to 0.38. Besides, our model correctly predicts image-level scores: the wrong category \"board\" is given a low score of 0.06 while the four correct categories are given high scores. In the second case, our model successfully predicts the TP \"person dribble\" missed by the baseline and suppress the FP \"person hold sports\" from 0.6 to 0.11 with the help of correct image-level classification result. This is similar for the third case.", + "bbox": [ + 75, + 580, + 467, + 791 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Last, as our category query learning is performed on image level, we conjecture that it is more helpful for images with dense human-object interactions. In such images, an human-object instance is relatively small and hard to learn good feature on its own. However, it may benefit more from the global image level category query feature, which aggregates more information from other similar instances in this", + "bbox": [ + 75, + 795, + 467, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "image. To validate this conjecture, we partition the images according to their \"interaction density\" and check whether our method produces larger improvement on images that are \"denser\". Specifically, for each interaction category, its mAP is evaluated separately on six different image partition subsets, where each image contains different numbers $(n = 1, 2, 3, 4, 5$ and $>5$ ) of human-object instances of this category. The mAP results of the three baseline methods and their integrated versions (as in Tab. 1) are shown in Fig. 5 (top, dashed vs. solid lines). It shows that: 1) the mAP is lower for larger n, indicating the \"denser\" images are more challenging; 2) our method improves the baselines consistently on all different partitions.", + "bbox": [ + 496, + 90, + 890, + 286 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To check whether the proposed method is more effective on \"denser\" images, we use the relative mAP improvement, which is a ratio, $\\frac{mAP_{ours} - mAP_{baseline}}{mAP_{baseline}}$ , for analysis. The ratio curves of the three baselines are shown in Fig. 5 (bottom, solid lines). It is clear that the relative improvement becomes larger for larger $n$ . This indicates that the image-level category query learning is more effective on these challenging dense images.", + "bbox": [ + 496, + 287, + 890, + 407 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To further verify that this behavior is not commonly true, we also compute the relative ratio of three more comparisons, \"gen-vlkt vs. qpic\", \"gen-vlkt vs. scg\" and \"scg vs. qpic\", in which the former outperforms the latter. These curves are also shown in Fig. 5 (bottom, dashed lines). There is no clear pattern in these curves, indicating that the performance gap between two arbitrary HOI methods are in general not related to the image \"density\".", + "bbox": [ + 496, + 407, + 890, + 529 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 542, + 617, + 558 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This work proposes a novel approach for the human-object interaction classification sub-task in HOI detection. We study the problem of interaction category modeling, in contrast to most previous methods focusing on human-object feature learning. We adopt the concept of category query in a previous method [25] for HOI, for the first time, and show that it is simple, general and highly effective.", + "bbox": [ + 496, + 568, + 890, + 672 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Clearly, this idea of category query modeling is not limited to multi-label image classification and HOI detection. We hope it is useful for other vision tasks.", + "bbox": [ + 496, + 674, + 890, + 717 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgment. This work was supported in part by the National Natural Science Foundation of China under Grant 62076183, 61936014 and 61976159, in part by the Natural Science Foundation of Shanghai under Grant 20ZR1473500, in part by the Shanghai Science and Technology Innovation Action Project of under Grant 20511100700 and 22511105300, in part by the Shanghai Municipal Science and Technology Major Project under Grant 2021SHZDZX0100, and in part by the Fundamental Research Funds for the Central Universities. The authors would also like to thank the anonymous reviewers for their careful work and valuable suggestions.", + "bbox": [ + 496, + 719, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "15282", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] A Bansal, S. S Rambhatla, A Shrivastava, and R Chellappa. Detecting human-object interactions via functional generalization. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 34, pages 10460-10469, 2020. 2", + "[2] N Carion, F Massa, G Synnaeve, N Usunier, A Kirillov, and S Zagoruyko. End-to-end object detection with transformers. In European Conference on Computer Vision, pages 213-229. Springer, 2020. 2, 7", + "[3] Y.-W Chao, Y Liu, X Liu, H Zeng, and J Deng. Learning to detect human-object interactions. In 2018 iee winter conference on applications of computer vision (wacv), pages 381-389. IEEE, 2018. 1, 5, 6", + "[4] M Chen, Y Liao, S Liu, Z Chen, F Wang, and C Qian. Reformulating hoi detection as adaptive set prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9004–9013, 2021. 2", + "[5] B Cheng, I Misra, A. G Schwing, A Kirillov, and R Girdhar. Masked-attention mask transformer for universal image segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1290–1299, 2022. 3", + "[6] L Dong, Z Li, K Xu, Z Zhang, L Yan, S Zhong, and X Zou. Category-aware transformer network for better human-object interaction detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19538-19547, 2022. 1, 2", + "[7] C Gao, J Xu, Y Zou, and J.-B Huang. Drg: Dual relation graph for human-object interaction detection. In European Conference on Computer Vision, pages 696-712. Springer, 2020. 1, 2, 4, 5, 6", + "[8] C Gao, Y Zou, and J.-B Huang. ican: Instance-centric attention network for human-object interaction detection. arXiv preprint arXiv:1808.10437, 2018. 1, 2", + "[9] G Gkioxari, R Girshick, P Dólár, and K He. Detecting and recognizing human-object interactions. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 8359-8367, 2018. 1", + "[10] S Gupta and J Malik. Visual semantic role labeling. arXiv preprint arXiv:1505.04474, 2015. 1, 5, 6", + "[11] T Gupta, A Schwing, and D Hoiem. No-frills human-object interaction detection: Factorization, layout encodings, and training techniques. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9677-9685, 2019. 2, 4", + "[12] H He, Y Yuan, X Yue, and H Hu. Rankseg: Adaptive pixel classification with image category ranking for segmentation. In Proceedings of the European Conference on Computer Vision (ECCV), pages 682-700. Springer Nature Switzerland Cham, 2022. 4", + "[13] A Iftekhar, H Chen, K Kundu, X Li, J Tighe, and D Modolo. What to look at and where: Semantic and spatial refined transformer for detecting human-object interactions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5353-5363, 2022. 2, 4", + "[14] B Kim, T Choi, J Kang, and H. J Kim. Uniondet: Union-level detector towards real-time human-object interaction de" + ], + "bbox": [ + 78, + 114, + 470, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "tlection. In European Conference on Computer Vision, pages 498-514. Springer, 2020. 4", + "[15] B Kim, J Lee, J Kang, E.-S Kim, and H. J Kim. Hotr: End-to-end human-object interaction detection with transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 74-83, 2021. 2", + "[16] B Kim, J Mun, K.-W On, M Shin, J Lee, and E.-S Kim. Mstr: Multi-scale transformer for end-to-end human-object interaction detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19578-19587, 2022. 1, 2, 5, 6", + "[17] D.-J Kim, X Sun, J Choi, S Lin, and I. S Kweon. Detecting human-object interactions with action co-occurrence priors. In European Conference on Computer Vision, pages 718–736. Springer, 2020. 2, 4", + "[18] Y.-L Li, X Liu, X Wu, Y Li, and C Lu. Hoi analysis: Integrating and decomposing human-object interaction. Advances in Neural Information Processing Systems, 33:5011-5022, 2020. 6", + "[19] Y.-L Li, L Xu, X Liu, X Huang, Y Xu, S Wang, H.-S Fang, Z Ma, M Chen, and C Lu. Pastanet: Toward human activity knowledge engine. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 382-391, 2020. 2", + "[20] Y.-L Li, S Zhou, X Huang, L Xu, Z Ma, H.-S Fang, Y Wang, and C Lu. Transferable interactiveness knowledge for human-object interaction detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3585-3594, 2019. 2", + "[21] Y Liao, S Liu, F Wang, Y Chen, C Qian, and J Feng. Ppdm: Parallel point detection and matching for real-time human-object interaction detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 482-490, 2020. 1, 4", + "[22] Y Liao, A Zhang, M Lu, Y Wang, X Li, and S Liu. Gen-vlkt: Simplify association and enhance interaction understanding for hoi detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20123-20132, 2022. 1, 2, 5, 6", + "[23] T.-Y Lin, P Goyal, R Girshick, K He, and P Dollar. Focal loss for dense object detection. In Proceedings of the IEEE international conference on computer vision, pages 2980-2988, 2017. 3, 7", + "[24] T.-Y Lin, M Maire, S Belongie, J Hays, P Perona, D Ramanan, P Dollár, and C. L Zitnick. Microsoft coco: Common objects in context. In European conference on computer vision, pages 740-755. Springer, 2014. 5", + "[25] S Liu, L Zhang, X Yang, H Su, and J Zhu. Query2label: A simple transformer way to multi-label classification. arXiv preprint arXiv:2107.10834, 2021. 1, 2, 3, 4, 6, 7, 8", + "[26] X Liu, Y.-L Li, X Wu, Y.-W Tai, C Lu, and C.-K Tang. Interactiveness field in human-object interactions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20113-20122, 2022. 6", + "[27] Y Liu, Q Chen, and A Zisserman. Amplifying key cues for human-object-interaction detection. In European Conference on Computer Vision, pages 248-265. Springer, 2020. 2" + ], + "bbox": [ + 503, + 92, + 890, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "15283", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[28] J Park, S Lee, H Heo, H. K Choi, and H. J Kim. Consistency learning via decoding path augmentation for transformers in human object interaction detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1019-1028, 2022. 4", + "[29] S Qi, W Wang, B Jia, J Shen, and S.-C Zhu. Learning human-object interactions by graph parsing neural networks. In Proceedings of the European Conference on Computer Vision (ECCV), pages 401-417, 2018. 1, 2", + "[30] X Qu, C Ding, X Li, X Zhong, and D Tao. Distillation using oracle queries for transformer-based human-object interaction detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19558-19567, 2022. 2", + "[31] A Radford, J. W Kim, C Hallacy, A Ramesh, G Goh, S Agarwal, G Sastry, A Askell, P Mishkin, J Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 2", + "[32] T Ridnik, E Ben-Baruch, N Zamir, A Noy, I Friedman, M Protter, and L Zelnik-Manor. Asymmetric loss for multi-label classification. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 82-91, 2021. 3, 7", + "[33] M Tamura, H Ohashi, and T Yoshinaga. Qpic: Query-based pairwise human-object interaction detection with image-wide contextual information. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10410-10419, 2021. 1, 2, 4, 5, 6", + "[34] O Ulutan, A Iftekhar, and B. S Manjunath. Vsgnet: Spatial attention network for detecting human object interactions using graph convolutions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13617-13626, 2020. 1, 2", + "[35] A Vaswani, N Shazeer, N Parmar, J Uszkoreit, L Jones, A. N Gomez, L Kaiser, and I Polosukhin. Attention is all you need. In Advances in neural information processing systems, pages 5998-6008, 2017. 2, 7", + "[36] B Wan, D Zhou, Y Liu, R Li, and X He. Pose-aware multilevel feature network for human object interaction detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9469-9478, 2019. 2", + "[37] H Wang, W.-s Zheng, and L Yingbiao. Contextual heterogeneous graph network for human-object interaction detection. In European Conference on Computer Vision, pages 248-264. Springer, 2020. 1, 2", + "[38] T Wang, T Yang, M Danelljan, F. S Khan, X Zhang, and J Sun. Learning human-object interaction detection using interaction points. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4116-4125, 2020. 4", + "[39] X Wu, Y.-L Li, X Liu, J Zhang, Y Wu, and C Lu. Mining cross-person cues for body-part interactiveness learning in hoi detection. In Proceedings of the European Conference on Computer Vision (ECCV), pages 121-136. Springer, 2022. 2, 6", + "[40] H Yuan, J Jiang, S Albanie, T Feng, Z Huang, D Ni, and M Tang. Rlip: Relational language-image pre-training for" + ], + "bbox": [ + 78, + 90, + 470, + 901 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "human-object interaction detection. In Advances in Neural Information Processing Systems, 2022. 2, 6", + "[41] A Zhang, Y Liao, S Liu, M Lu, Y Wang, C Gao, and X Li. Mining the benefits of two-stage and one-stage hoi detection. In Thirty-Fifth Conference on Neural Information Processing Systems, 2021. 1, 2, 4, 5, 6", + "[42] F. Z Zhang, D Campbell, and S Gould. Spatially conditioned graphs for detecting human-object interactions. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13319-13327, 2021. 2, 4, 5, 6", + "[43] F. Z Zhang, D Campbell, and S Gould. Efficient two-stage detection of human-object interactions with a novel unary-pairwise transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20104-20112, 2022. 5", + "[44] Y Zhang, Y Pan, T Yao, R Huang, T Mei, and C.-W Chen. Exploring structure-aware transformer over interaction proposals for human-object interaction detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19548-19557, 2022. 6", + "[45] X Zhong, C Ding, Z Li, and S Huang. Towards hard-positive query mining for detr-based human-object interaction detection. In Proceedings of the European Conference on Computer Vision (ECCV), pages 444-460. Springer, 2022. 2, 4, 6", + "[46] X Zhong, C Ding, X Qu, and D Tao. Polysemy deciphering network for robust human-object interaction detection. International Journal of Computer Vision, 129(6):1910-1929, 2021. 1, 2", + "[47] X Zhong, X Qu, C Ding, and D Tao. Glance and gaze: Inferring action-aware points for one-stage human-object interaction detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13234-13243, 2021. 6", + "[48] D Zhou, Z Liu, J Wang, L Wang, T Hu, E Ding, and J Wang. Human-object interaction detection via disentangled transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19568-19577, 2022. 1, 2, 4, 6", + "[49] C Zou, B Wang, Y Hu, J Liu, Q Wu, Y Zhao, B Li, C Zhang, C Zhang, Y Wei, et al. End-to-end human object interaction detection with hoi transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11825-11834, 2021. 2" + ], + "bbox": [ + 501, + 92, + 890, + 712 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "15284", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2023/Category Query Learning for Human-Object Interaction Classification/c48c8f3f-b487-4c24-815b-6385a57d02e8_model.json b/2023/Category Query Learning for Human-Object Interaction Classification/c48c8f3f-b487-4c24-815b-6385a57d02e8_model.json new file mode 100644 index 0000000000000000000000000000000000000000..055d1b62d537b84ddf5779168e25755cef33fd47 --- /dev/null +++ b/2023/Category Query Learning for Human-Object Interaction Classification/c48c8f3f-b487-4c24-815b-6385a57d02e8_model.json @@ -0,0 +1,2464 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.128, + 0.131, + 0.843, + 0.154 + ], + "angle": 0, + "content": "Category Query Learning for Human-Object Interaction Classification" + }, + { + "type": "text", + "bbox": [ + 0.174, + 0.18, + 0.796, + 0.229 + ], + "angle": 0, + "content": "Chi Xie\\(^{1\\dagger}\\) Fangao Zeng\\(^{2‡}\\) Yue Hu\\(^{3‡}\\) Shuang Liang\\(^{1*}\\) Yichen Wei\\(^{2‡}\\) \n\\(^{1}\\)Tongji University \\(^{2}\\)MEGVII Technology \\(^{3}\\)Shanghai Jiao Tong University \\(^{1}\\) {chixie, shuangliang}@tongji.edu.cn" + }, + { + "type": "text", + "bbox": [ + 0.175, + 0.234, + 0.812, + 0.251 + ], + "angle": 0, + "content": "2zfg472988436@163.com,wei_yi_chen@hotmail.com 318671129361@sjtu.edu.cn" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.287, + 0.313, + 0.303 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.32, + 0.474, + 0.531 + ], + "angle": 0, + "content": "Unlike most previous HOI methods that focus on learning better human-object features, we propose a novel and complementary approach called category query learning. Such queries are explicitly associated to interaction categories, converted to image specific category representation via a transformer decoder, and learnt via an auxiliary image-level classification task. This idea is motivated by an earlier multi-label image classification method, but is for the first time applied for the challenging human-object interaction classification task. Our method is simple, general and effective. It is validated on three representative HOI baselines and achieves new state-of-the-art results on two benchmarks. Code will be available at https://github.com/charles-xie/CQL." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.564, + 0.21, + 0.58 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.591, + 0.47, + 0.726 + ], + "angle": 0, + "content": "Human-Object Interaction (HOI) detection has attracted a lot of interests in recent years [3, 8-10, 21, 33]. The task consists of two sub-tasks. The first is human and object detection. It is usually performed by common object detection methods. The second is interaction classification of each human-object (HO) pair. This sub-task is very challenging due to the complex appearance variations in the interaction categories. See Fig. 1 for examples. It is the focus of most previous HOI methods, as well as this work." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.728, + 0.47, + 0.85 + ], + "angle": 0, + "content": "Most previous HOI methods focus on learning better human-object features, including modeling relation and context via GNN [7, 29, 34, 37] or attention mechanism [8, 34, 46], decoupling localization and classification [22, 41, 48], leveraging vision-language knowledge [6, 22] and introducing multi-scale feature to transformer [16]. However, for interaction classification they all adopt the simple linear classifier that performs the dot product of the" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.285, + 0.695, + 0.386 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.554, + 0.387, + 0.645, + 0.397 + ], + "angle": 0, + "content": "(a) person fly kite" + }, + { + "type": "image", + "bbox": [ + 0.698, + 0.286, + 0.892, + 0.386 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.747, + 0.387, + 0.845, + 0.396 + ], + "angle": 0, + "content": "person fly airplane" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.398, + 0.696, + 0.498 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.545, + 0.499, + 0.65, + 0.51 + ], + "angle": 0, + "content": "(b) person hold fork" + }, + { + "type": "image", + "bbox": [ + 0.698, + 0.398, + 0.892, + 0.498 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.739, + 0.5, + 0.852, + 0.51 + ], + "angle": 0, + "content": "person hold elephant" + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.527, + 0.894, + 0.597 + ], + "angle": 0, + "content": "Figure 1. Interaction classification is inherently challenging. In (a), \"fly\" is semantically polysemic, resulting in different objects, poses and relative positions. In (b), when \"hold\" is associated with different objects, the appearance, scene background, and human poses are largely different." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.614, + 0.892, + 0.644 + ], + "angle": 0, + "content": "human-object feature and a static weight vector, which represents an interaction category." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.645, + 0.893, + 0.825 + ], + "angle": 0, + "content": "In this work, we propose a new approach that enhances the above paradigm and complements most previous HOI methods. It is motivated by the recent work Query2label [25], a transformer-based classification network. It proposes a new concept we call category-specific query. Unlike the queries in other transformer methods, each query is associated to a specific and fixed image category during training and inference. This one-to-one binding makes the query learn to model each category more effectively. The queries are converted to image specific category representations via a transformer decoder. This method achieves excellent performance on multi-label image classification task." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.826, + 0.894, + 0.901 + ], + "angle": 0, + "content": "We extend this approach for human-object interaction classification. Essentially, our approach replaces traditional category representation as a static weight vector in previous HOI methods with category queries learnt as described above. The same linear classifier is adopted. Such cate" + }, + { + "type": "page_footnote", + "bbox": [ + 0.094, + 0.863, + 0.223, + 0.876 + ], + "angle": 0, + "content": "*Corresponding author." + }, + { + "type": "page_footnote", + "bbox": [ + 0.097, + 0.876, + 0.382, + 0.888 + ], + "angle": 0, + "content": "\\(\\dagger\\) Work done during internship at MEGVII technology." + }, + { + "type": "page_footnote", + "bbox": [ + 0.097, + 0.888, + 0.304, + 0.899 + ], + "angle": 0, + "content": "\\(\\ddagger\\)Work done while worked at MEGVII." + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.863, + 0.382, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15275" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.468, + 0.182 + ], + "angle": 0, + "content": "gory queries are more effective, and adaptive for different images, giving rise to better modeling of the complex variations in each interaction category. This is the crucial difference between this work and a simple adaption of [25] to HOI. Notably, this work is the first to address the category weight representation problem in the HOI community." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.183, + 0.468, + 0.274 + ], + "angle": 0, + "content": "Note that our proposed category specific query is different and not related to those queries in other transformer-based HOI methods [4, 15, 33, 49]. Specifically, category queries extract image-level features as the category representation. The queries in other methods are human-object instance-level features and category-agnostic." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.275, + 0.468, + 0.35 + ], + "angle": 0, + "content": "Our method is simple, lightweight and general. The overview is in Fig. 2. It is complementary to any off-the-shelf HOI method that provides human-object features. The modification of both inference and training is small. The incurred additional cost is marginal." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.352, + 0.468, + 0.532 + ], + "angle": 0, + "content": "In experiments, our approach is validated on three representative and strong HOI baseline methods, two transformer-based methods [22, 33] and a traditional two-stage method [42]. They are all significantly improved by our approach. New state-of-the-art results are obtained on two benchmarks. Specifically, we obtain \\(36.03\\mathrm{mAP}\\) on HICO-DET. Comprehensive ablation studies and in-depth discussion are also provided to verify the effectiveness of implementation details in our approach. It turns out that our method is more effective on challenging images that contain more human-object instances, a property that is rarely discussed by previous HOI methods." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.549, + 0.219, + 0.565 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.575, + 0.442, + 0.59 + ], + "angle": 0, + "content": "2.1. Instance Query Learning in HOI Detection" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.599, + 0.468, + 0.901 + ], + "angle": 0, + "content": "DETR [2] firstly proposes the concept of object instance query for object detection task. Such queries essentially learn the priors of both object appearance and spatial location. DETR leverages those queries to probe image features through a transformer [35] and localize unique objects in the image. Motivated by its great success, many works [15, 22, 33, 41, 48, 49] adapt such detection transformer framework to HOI detection by simply treating the HOI triplet [33, 49] or H-O pair [4, 15] as an object. A few of them [6, 30, 45] pay attention to adapting the plain query to this task. DOQ [30] proposes a knowledge distillation model using oracle queries to facilitate the representation learning of a transformer-based detector; HQM [45] explicitly constructs hard positive queries from ground truth to train the model to be less vulnerable to spacial variations; CATN [6] utilizes the object category prior generated from external object detector and language model for query initialization. In summary, these transformer-based methods use each query to aggregate context information not restricted to one interaction category, in order to predict" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.092, + 0.805, + 0.106 + ], + "angle": 0, + "content": "a potential HOI instance at a specific location." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.107, + 0.892, + 0.244 + ], + "angle": 0, + "content": "In DETR [2] and its variants [4,15,22,33,41,49] in HOI, as the queries are category-agnostic, their association to object categories are dynamic and unstable during training. This could be problematic. For example, it is well known that the convergence of DETR training is slow. In contrast, our proposed query is category-specific. The learning is guided by image-level classification task and stable. Such queries learn category-specific priors and are good representation for interaction categories." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.256, + 0.807, + 0.272 + ], + "angle": 0, + "content": "2.2. Feature Learning in HOI Detection" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.28, + 0.892, + 0.386 + ], + "angle": 0, + "content": "Early methods. Based on two-stage detection framework, early works make many efforts to help feature learning, including employing architectures effective in modeling relation and context like GNN [7, 29, 34, 37] and attention module [8, 34, 46], leveraging fine-grained visual features [11, 17, 19, 20, 36] like human pose and introducing language prior [1, 7, 17, 27, 46]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.387, + 0.892, + 0.478 + ], + "angle": 0, + "content": "Transformer-based methods. Motivated by DETR [49], many methods [4, 15, 33, 49] leverage transformer architecture [35] and extend the object query in DETR to HOI query. With the help of HOI query and transformer's built-in attention mechanism, those methods learn effective feature representation for HOI triplet or H-O pair." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.479, + 0.892, + 0.614 + ], + "angle": 0, + "content": "Based on those pioneer transformer-based methods, recently, many methods are proposed to further help feature learning, by decoupling H-O pair localization and interaction classification [22, 41, 48] or exploiting multi-scale feature in transformer architecture [16]. Some works [6, 22] leverage vision-language knowledge in CLIP [31] or design a pretrained model [40] specifically for HOI; others utilize information like human poses [39] or spatial configurations [13] that has been used in early HOI detectors." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.615, + 0.892, + 0.706 + ], + "angle": 0, + "content": "Relation to the proposed method. Previous methods learn the H-O feature while ours learns the category query as the category representation feature. Thus, they are complementary. The interaction classification is simply by measuring the similarity between the two types of features. The integration of our method to previous HOI methods is simple." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.723, + 0.63, + 0.738 + ], + "angle": 0, + "content": "3. Our Method" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.749, + 0.892, + 0.809 + ], + "angle": 0, + "content": "The overview of our method is in Fig. 2. It consists of two components, the image-level category query learning (top block) and human-object interaction classification (bottom right block)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.81, + 0.892, + 0.901 + ], + "angle": 0, + "content": "The first component is detailed in Sec. 3.1. It is briefly summarized here. A number of queries (embedding vectors) are associated to human-object interaction category, in a one-to-one manner. Such queries interact with image features (provided by a baseline HOI method) through a transformer decoder [35] and become image-specific queries." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15276" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.123, + 0.09, + 0.85, + 0.332 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.343, + 0.895, + 0.374 + ], + "angle": 0, + "content": "Figure 2. Overview of our method. It consists of two components (top and bottom right). It can be integrated with any baseline HOI method (bottom left) that provides image feature \\( I \\) and human-object instance features \\( F_{i} \\). See Sec. 3 for details." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.389, + 0.47, + 0.493 + ], + "angle": 0, + "content": "Learning of both the queries and decoder weights is supervised by an auxiliary image-level classification task. In this way, the queries are learnt to capture category-specific feature and become good feature representation for these categories. Besides some minor details, this step is the same as the previous work Query2label [25], which is for multi-label image classification task." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.494, + 0.471, + 0.63 + ], + "angle": 0, + "content": "The second component is detailed in Sec. 3.2. For the first time, we adopt the category query learning method for human-object interaction classification tasks. The cosine similarity between the category query and human-object feature is used for interaction classification. Thus, it works with any HOI method that provides human-object features. Besides, the image-level classification results turn out moderately helpful in an score integration step, which is an extra technique that benefits the performance." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.63, + 0.47, + 0.69 + ], + "angle": 0, + "content": "Overall, the proposed method is simple, effective, lightweight and general. It can be combined with most previous HOI detection methods (bottom left block in Fig. 2), with small modification, as elaborated in Sec. 4." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.701, + 0.407, + 0.718 + ], + "angle": 0, + "content": "3.1. Image-level Category Query Learning" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.724, + 0.47, + 0.784 + ], + "angle": 0, + "content": "Similar as in Query2Label [25], for \\(K\\) human-object interaction categories, we define their one-to-one corresponding category queries, which are learnable embedding vectors, \\(\\{Q_k\\} \\in \\mathbb{R}^{K\\times D}\\), where \\(D\\) is the vector dimension." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.784, + 0.469, + 0.831 + ], + "angle": 0, + "content": "Each query \\( Q_{k} \\) aggregates image features \\( \\mathbf{I} \\in \\mathbb{R}^{H \\times W \\times D} \\) through a transformer decoder and is updated to image specific query \\( Q_{k}' \\)," + }, + { + "type": "equation", + "bbox": [ + 0.179, + 0.842, + 0.47, + 0.86 + ], + "angle": 0, + "content": "\\[\n\\left\\{Q _ {k} ^ {\\prime} \\right\\} = \\operatorname {d e c o d e r} \\left(\\left\\{Q _ {k} \\right\\}, I\\right). \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Note that the decoder structure has several variants, which are studied in [5]. Our experiments show that the" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.389, + 0.893, + 0.448 + ], + "angle": 0, + "content": "structure is of minor importance, as discussed in Sec. 5.4. Specifically, our decoder consists of two layers, each of which consisting of a cross-attention layer, a self-attention layer and a FFN layer, in order." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.449, + 0.894, + 0.496 + ], + "angle": 0, + "content": "Then, each image-level classification probabilities \\( p_k \\) is computed by applying a category-specific fully-connected layer and a sigmoid activation on the updated query \\( Q_{k}^{\\prime} \\)." + }, + { + "type": "equation", + "bbox": [ + 0.612, + 0.511, + 0.892, + 0.528 + ], + "angle": 0, + "content": "\\[\np _ {k} = \\operatorname {s i g m o i d} \\left(\\mathrm {F C} \\left(Q _ {k} ^ {\\prime}\\right)\\right) \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.536, + 0.893, + 0.642 + ], + "angle": 0, + "content": "Learning of the category query \\(\\{Q_k\\}\\) and the decoder weights is supervised by common image classification losses. To deal with the label imbalance problem, focal loss [23] and asymmetric loss (ASL) [32] are used. Asymmetric loss is a variant of focal loss. It is more robust for high label imbalance and noises. Our experiments (see Sec. 5.4) show that it is slightly better." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.642, + 0.894, + 0.686 + ], + "angle": 0, + "content": "Specifically, with the classification probability \\( p_k \\) and the shifted probability \\( p_k' = \\max(p_k - m, 0) \\), the asymmetric loss is" + }, + { + "type": "equation", + "bbox": [ + 0.513, + 0.697, + 0.892, + 0.739 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {i m g} = \\frac {1}{K} \\sum_ {k = 1} ^ {K} \\left\\{ \\begin{array}{l l} (1 - p _ {k}) ^ {\\gamma +} \\log \\left(p _ {k}\\right), & y _ {k} = 1, \\\\ \\left(p _ {k} ^ {\\prime}\\right) ^ {\\gamma -} \\log \\left(1 - p _ {k} ^ {\\prime}\\right), & y _ {k} = 0, \\end{array} \\right. \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.892, + 0.809 + ], + "angle": 0, + "content": "where the binary label \\( y_{k} \\) indicates the existence of category \\( k \\) in the image, and \\( \\gamma +, \\gamma - \\) as well as \\( m \\) are hyperparameters. We use the default values in ASL [32], \\( \\gamma + = 0 \\), \\( \\gamma - = 4 \\) and \\( m = 0.05 \\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.811, + 0.893, + 0.902 + ], + "angle": 0, + "content": "In this way, the category queries are learnt to encode the category priors. Figure 3 is the visualization of the heatmaps in the cross-attention layer of the decoder. Each category query learns to locate the human body parts related to discriminative feature of its corresponding interaction category, e.g., in Fig. 3c, the query of \"hold\" highlights" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15277" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.152 + ], + "angle": 0, + "content": "the hand region, while in the same image the query of \"ride\" highlights the foot region. It qualitatively demonstrates that the query learning is effective in encoding category-specific information." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.155, + 0.471, + 0.201 + ], + "angle": 0, + "content": "The updated queries adaptively extract category-related features for each image, with the help of the transformer decoder's built-in multi-head cross-attention layer." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.218, + 0.468, + 0.234 + ], + "angle": 0, + "content": "3.2. Interaction Classification with Category Query" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.244, + 0.468, + 0.321 + ], + "angle": 0, + "content": "In this step, we apply the updated category queries \\(\\{Q_k'\\}\\) as the weights for interaction classification. Given the \\(i\\)-th human-object instance, its classification probability score for \\(k\\)-th interaction category is simply the cosine similarity between its feature \\(F_i\\) and the category query feature \\(Q_k'\\)." + }, + { + "type": "equation", + "bbox": [ + 0.172, + 0.342, + 0.469, + 0.375 + ], + "angle": 0, + "content": "\\[\ns _ {i, k} = \\operatorname {s i g m o i d} \\left(\\frac {Q _ {k} ^ {\\prime} \\cdot F _ {i}}{\\left| Q _ {k} ^ {\\prime} \\right| \\times \\left| F _ {i} \\right|}\\right). \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.392, + 0.469, + 0.437 + ], + "angle": 0, + "content": "There are no restrictions or assumptions on the human-object feature \\( F_{i} \\). Most previous HOI methods should be applicable here." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.44, + 0.469, + 0.501 + ], + "angle": 0, + "content": "In this step, the traditional classification weight from static parameters is replaced with the category query adaptive on each image. This behavior is essentially different from [25], which uses category queries as image feature." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.503, + 0.469, + 0.594 + ], + "angle": 0, + "content": "By \"adaptive\", we mean that the queries are updated dynamically according to the image contents. As exemplified in Fig. 3a and Fig. 3c, the queries of \"hold\" learn to highlight different interactive areas in different images and update themselves with features from these areas via attention. This shows the query learning is adaptive to image." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.597, + 0.469, + 0.673 + ], + "angle": 0, + "content": "As discussed in Sec. 5.4, the classification step in Eq. (4) is crucial to make the category query learning effective on human-object interaction classification. Without this step, the image-level category query learning using only image-level classification from Query2label [25] is of little use." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.675, + 0.47, + 0.901 + ], + "angle": 0, + "content": "Score integration step. The segmentation method in [12] discards certain categories during pixel classification that have low image classification scores. Motivated by this method, we take a similar score integration step. The image-level classification score \\(\\{p_i\\}\\) is used to enhance the human-object instance classification. The idea is that, the instance score should be higher if the image-level probability is higher. Our implementation is similar as in [12]. During both training and testing, for each image, the top-\\(\\kappa\\) categories (\\(\\kappa = 70\\) in this work) with higher image classification scores \\(\\{p_i\\}\\) are selected. The instance score \\(s_{i,k}\\) is slightly modulated such that it becomes higher if the rank of category \\(k\\) is higher. This strategy gives rise to moderate improvement, as verified in Tab. 4. We left the implementation details in the supplementary materials." + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.089, + 0.63, + 0.156 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.633, + 0.089, + 0.76, + 0.156 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.763, + 0.089, + 0.892, + 0.156 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.518, + 0.162, + 0.874, + 0.174 + ], + "angle": 0, + "content": "(a) Image with \"person-break-baseball bat\" and \"hold-baseball bat\"." + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.175, + 0.63, + 0.242 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.633, + 0.175, + 0.76, + 0.242 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.763, + 0.175, + 0.892, + 0.242 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.541, + 0.248, + 0.85, + 0.261 + ], + "angle": 0, + "content": "(b) Image with \"person-direct-airplane\" and \"fly-airplane\"." + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.261, + 0.63, + 0.328 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.633, + 0.261, + 0.76, + 0.328 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.763, + 0.261, + 0.892, + 0.328 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.555, + 0.333, + 0.838, + 0.346 + ], + "angle": 0, + "content": "(c) Image with \"person-hold-horse\" and \"ride-horse\"." + }, + { + "type": "image_caption", + "bbox": [ + 0.5, + 0.358, + 0.892, + 0.387 + ], + "angle": 0, + "content": "Figure 3. From left to right: image, attention maps of the cross-attention layer in the decoder for different interaction categories." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.404, + 0.88, + 0.421 + ], + "angle": 0, + "content": "4. Integration to Off-the-shelf HOI Detectors" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.43, + 0.892, + 0.521 + ], + "angle": 0, + "content": "As shown in Fig. 2, our method is ready to integrate with any baseline HOI method that provides image feature \\(\\mathbf{I}\\) and human-object instance feature \\(\\{F_i\\}\\). The integration is simple. During inference, the human-object instance interaction classification part is replaced by our method in Sec. 3.2, the top and bottom right block in Fig. 2." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.521, + 0.892, + 0.567 + ], + "angle": 0, + "content": "During training, the original loss in the baseline HOI method \\(\\mathcal{L}_{base}\\) is added to our image classification loss in Eq. (3). The final loss \\(\\mathcal{L}\\) for training is" + }, + { + "type": "equation", + "bbox": [ + 0.616, + 0.585, + 0.891, + 0.602 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\mathcal {L} _ {\\text {b a s e}} + \\lambda * \\mathcal {L} _ {\\text {i m g}}, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.61, + 0.892, + 0.655 + ], + "angle": 0, + "content": "where the weight \\(\\lambda\\) is 1.0 by default. All other hyper parameters and details during training remain the same as in the baseline HOI method." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.657, + 0.892, + 0.717 + ], + "angle": 0, + "content": "Thus, our method is general and applicable to most existing HOI methods. In this work, we select three representative yet different baseline methods to verify the effectiveness of our approach, as described below." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.718, + 0.892, + 0.822 + ], + "angle": 0, + "content": "QPIC [33] is the first to introduce transformer method into HOI task. It is also the baseline for many recent works [13, 28, 41, 45, 48]. Its performance is much better than early one-stage [14, 21, 38] and two-stage [7, 11, 17] methods while keeping a simple and end-to-end architecture. It consists of a CNN backbone as well as a transformer encoder and decoder." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.825, + 0.892, + 0.87 + ], + "angle": 0, + "content": "During our integration, the feature map in its transformer encoder is used as the image feature \\(\\mathbf{I}\\). The human-object feature \\(\\{F_i\\}\\) is the query feature in its decoder." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.871, + 0.892, + 0.901 + ], + "angle": 0, + "content": "SCG [42] is a traditional two-stage method and the best in this category. It is also one of the best methods that" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15278" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.08, + 0.089, + 0.895, + 0.206 + ], + "angle": 0, + "content": "
MethodPipelineE2EHICO-DETV-COCOEfficiency
FullRareNon-RareS1S2#ParamsFPS
QPIC [33]transformer✓28.9321.6231.1261.3963.6541M19.5
+ Ourstransformer✓31.08(+2.15)23.9033.2263.67(+2.28)65.4946M(+5M)18.3(-6.2%)
SCG [42]two-stageX31.2824.1633.4056.9362.5157M4.5
+ Ourstwo-stageX32.74(+1.46)26.2534.6859.14(+2.21)65.6164M(+7M)4.1(-8.9%)
GEN-VLKT [22]transformerX33.6929.9434.8164.8966.7442M21.7
+ OurstransformerX35.36(+1.67)32.9736.0766.40(+1.51)69.1747M(+5M)20.6(-5.1%)
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.215, + 0.891, + 0.245 + ], + "angle": 0, + "content": "Table 1. The performance numbers of three different baseline HOI methods with and without integration of our method, on two datasets. \"E2E\" denotes whether a HOI detector is end-to-end. All models are tested on Tesla V100." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.26, + 0.469, + 0.334 + ], + "angle": 0, + "content": "does not use transformer. It uses a multi-stream graph neural network(GNN) for interaction classification. In our experiment, the detection boxes are from a fine-tuned detector provided by DRG [7] for HICO-DET and a fine-tuned DETR for V-COCO." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.336, + 0.469, + 0.397 + ], + "angle": 0, + "content": "During our integration, the CNN feature map in the backbone of SCG is used as image feature \\(\\mathbf{I}\\). The human-object feature \\(\\{F_i\\}\\) is generated through RoI pooling with detected human and object boxes and fused with the GNN." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.398, + 0.469, + 0.473 + ], + "angle": 0, + "content": "GEN-VLKT [22] is also transformer-based, but not end-to-end as pairwise NMS [41] is used for post-processing. It is the current state-of-the-art method. It uses two parallel decoders for object detection and interaction classification, namely instance decoder and interaction decoder." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.475, + 0.47, + 0.595 + ], + "angle": 0, + "content": "During our integration, the feature map in its transformer encoder is used as image feature \\(\\mathbf{I}\\). The query feature in the interaction decoder is used as the human-object feature \\(\\{F_i\\}\\). Note that, unlike the majority of HOI detection methods, the original GEN-VLKT uses HOI categories rather than interaction categories during interaction classification. Our experiments still use interaction categories, in order to be consistent with most other methods." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.612, + 0.21, + 0.629 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.638, + 0.47, + 0.775 + ], + "angle": 0, + "content": "In this section, we verify the applicability and effectiveness of the proposed method through experiments. In Sec. 5.1, we introduce the experimental settings. Then we demonstrate the effectiveness of the proposed method over 3 baselines in Sec. 5.2, and show it achieves SOTA results on major benchmarks in Sec. 5.3. Next, in Sec. 5.4 we conduct comprehensive ablation studies on the key components as well as detailed technical designs. Lastly, we provide some analysis and visualization in Sec. 5.5." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.786, + 0.18, + 0.801 + ], + "angle": 0, + "content": "5.1. Datasets" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.81, + 0.47, + 0.903 + ], + "angle": 0, + "content": "HICO-DET [3] and V-COCO [10] are two widely-used HOI benchmarks. HICO-DET contains 47,776 images, with 38,118 for training and 9,658 for testing. There are 600 HOI categories in HICO-DET, consisting of 117 interaction classes and 80 object classes. Each HOI category is composed of an interaction and an object. V-COCO is" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.26, + 0.892, + 0.32 + ], + "angle": 0, + "content": "a subset of MS-COCO [24] with HOI annotations, including 10,346 images (2,533 for training, 2,867 for validation and 4,946 for testing). It has 80 object categories same with HICO-DET and 29 interaction categories." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.321, + 0.893, + 0.578 + ], + "angle": 0, + "content": "Evaluation metrics. For HICO-DET, we adopt the commonly used mAP metric [3]. Each prediction is a \\(\\langle\\) human, interaction, object \\(\\rangle\\) triplet. A prediction is a true positive only when the human and object bounding boxes both have IoU \\(>0.5\\) w.r.t. ground truth and the interaction classification result is correct. We evaluate the performance in two different settings following [3]. In the known object setting, for each HOI category, we evaluate the prediction only on the images containing the target object category. In default setting, the detection result of each category is evaluated on the full test set. In each setting, we report the mAP over (1) all 600 HOI categories (Full), (2) 138 categories with less than 10 training samples (Rare), and (3) the remaining 462 categories (Non-rare). For V-COCO, we use the role mAP following [10], under both scenario #1 (including objects) and #2 (ignoring objects). The performance is evaluated using its official evaluation toolkit." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.589, + 0.864, + 0.605 + ], + "angle": 0, + "content": "5.2. Improvement on Three Different Baselines" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.613, + 0.892, + 0.763 + ], + "angle": 0, + "content": "Tab. 1 summarizes the performance of the three baseline HOI methods before and after integration of our method. Backbone is ResNet50. All these methods are significantly improved. Specifically, QPIC [33] is improved by \\(2.15\\mathrm{mAP}\\), making it competitive with those more recent works [16, 41, 43]. SCG [42] is improved by 1.46 mAP, demonstrating that our method is not limited to transformer-based baselines. The current SOTA method GEN-VLKT [22] is improved by \\(1.67\\mathrm{mAP}\\), producing the new SOTA result (also refer to Tab. 2)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.892, + 0.809 + ], + "angle": 0, + "content": "On the V-COCO dataset [10], the performance improvement is similar, which is 2.28, 2.21 and \\(1.51\\mathrm{mAP}\\) on QPIC, SCG and GEN-VLKT, respectively." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.811, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Notably, for SCG [42], as the object detector is fixed during the training of its interaction classification network, the improvement by our method is purely due to better interaction classification, not a better fine-tuned CNN backbone or a better object detector. This further consolidates that the category query learning is effective." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15279" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.16, + 0.089, + 0.813, + 0.375 + ], + "angle": 0, + "content": "
MethodDetectorBackboneDefaultKnown Object
FullRareNon-rareFullRareNon-rare
DRG [7]HICO-DETResNet50-FPN24.5319.4726.0427.9823.1129.43
GG-Net [47]HICO-DETHourglass10423.4716.4825.6027.3620.2329.48
IDN [18]HICO-DETResNet5026.2922.6127.3928.2424.4729.37
QPIC [33]HICO-DETResNet5029.0721.8531.2331.6824.1433.93
SCG [42]HICO-DETResNet50-FPN31.3324.7233.3134.3727.1836.52
CDN [41]HICO-DETResNet5031.7827.5533.0534.5329.7335.96
DT [48]HICO-DETResNet5031.7527.4533.0334.5030.1335.81
STIP [44]HICO-DETResNet5031.6027.7532.7534.4130.1235.69
HQM [45]HICO-DETResNet5032.4728.1533.76---
MSTR [16]HICO-DETResNet5031.1725.3132.9234.0228.8335.57
RLIP [40]COCO+VGResNet5032.8426.8534.63---
IF [26]HICO-DETResNet5033.5130.3034.4636.2833.1637.21
GEN-VLKT-B [22]HICO-DETResNet5033.7529.2535.1036.7832.7537.99
GEN-VLKT-M [22]HICO-DETResNet10134.7831.5035.7738.0734.9439.01
GEN-VLKT-L [22]HICO-DETResNet10134.9531.1836.0838.2234.3639.37
BodyPartMap [39]HICO-DETResNet5035.1533.7135.5837.5635.8738.06
GEN-VLKT-B + OursHICO-DETResNet5035.3632.9736.0738.4334.8539.50
GEN-VLKT-M + OursHICO-DETResNet10135.8332.9136.7038.7935.2839.84
GEN-VLKT-L + OursHICO-DETResNet10136.0333.1636.8938.8235.5139.81
" + }, + { + "type": "table_caption", + "bbox": [ + 0.155, + 0.386, + 0.816, + 0.4 + ], + "angle": 0, + "content": "Table 2. The proposed method achieves state-of-the-art on HICO-DET [3]. The best results are marked in bold." + }, + { + "type": "table", + "bbox": [ + 0.08, + 0.414, + 0.47, + 0.668 + ], + "angle": 0, + "content": "
MethodBackboneScenario #1Scenario #2
DRG [7]R50FPN51.0-
SCG [42]R5054.260.9
GG-Net [47]HG10454.7-
QPIC [33]R5058.861.0
HQM [45]R5063.6-
CDN [41]R5061.763.8
GEN-VLKT-B [22]R5062.464.5
GEN-VLKT-M [22]R10163.365.6
GEN-VLKT-L [22]R10163.665.9
MSTR [16]R5062.065.2
BodyPartMap [39]R5063.065.1
IF [26]R5063.065.2
DT [48]R5066.268.5
STIP [44]R5065.169.7
GEN-VLKT-B + OursR5066.469.2
GEN-VLKT-M + OursR10166.869.8
GEN-VLKT-L + OursR10166.569.9
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.678, + 0.47, + 0.707 + ], + "angle": 0, + "content": "Table 3. Comparison with state-of-the-art methods on V-COCO [10] dataset. The best results are marked in bold." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.726, + 0.471, + 0.834 + ], + "angle": 0, + "content": "To verify that the performance improvement is not due to a larger model, we also compare the model size and running speed. Our method increases the model parameters by a few millions, which is small compared to the original model size. The running speed, measured by FPS, is only decreased by a few percent. The marginal additional cost shows that our method is quite lightweight." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.846, + 0.371, + 0.863 + ], + "angle": 0, + "content": "5.3. Comparison with State-of-the-art" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.871, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Tab. 2 and Tab. 3 compare our method with many previous methods for HICO-DET and V-COCO datasets, respec" + }, + { + "type": "table", + "bbox": [ + 0.501, + 0.414, + 0.892, + 0.492 + ], + "angle": 0, + "content": "
C1C2C3FullRareNon-Rare
a---33.6929.9434.81
b✓--33.86 (+0.17)31.1234.68
c✓✓-34.98 (+1.29)31.7335.95
d✓✓✓35.36 (+1.67)32.9736.07
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.503, + 0.892, + 0.545 + ], + "angle": 0, + "content": "Table 4. Ablation study of several variants of our method, starting from the baseline (a) to our approach (d). C1, C2 and C3 are described in Sec. 5.4. The best results are marked in bold." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.563, + 0.817, + 0.577 + ], + "angle": 0, + "content": "tively. GEN-VLKT [22] is used as our baseline." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.579, + 0.892, + 0.669 + ], + "angle": 0, + "content": "On HICO-DET, our result with ResNet50 backbone already outperforms all previous methods under both default and known object settings. With the stronger ResNet101 backbone, our method achieves the new state-of-the-art 36.03 full mAP under default settings and 38.82 under known object settings." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.67, + 0.892, + 0.73 + ], + "angle": 0, + "content": "On V-COCO dataset, our method achieves the new state-of-the-art performance on Scenario 1, with an AP of 66.4 for ResNet50, surpassing [48]. For scenario 2, it is comparable with the state-of-the-art [44]." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.741, + 0.707, + 0.758 + ], + "angle": 0, + "content": "5.4. Ablation Experiments" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.892, + 0.81 + ], + "angle": 0, + "content": "We perform various ablation experiments to validate the effectiveness of different components in our method. HICO-DET dataset and GEN-VLKT [22] baseline are used." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.811, + 0.892, + 0.901 + ], + "angle": 0, + "content": "The proposed method can be divided into 3 components: C1 means applying Query2Label [25] to the baseline detector as a multi-task learning (with feature extractor shared). In detail, it adds the queries, the decoder and image classification loss (Eq. (1), Eq. (2) and Eq. (5)). C2 means using the learned query in C1 as adaptive interaction classification" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "15280" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.098, + 0.089, + 0.451, + 0.245 + ], + "angle": 0, + "content": "
loss typeλFullRareNon-Rare
-034.2130.1535.42
focal loss [23]0.534.4331.0635.44
1.034.5131.0835.53
1.534.3531.5435.19
2.034.2931.1835.22
ASL [32]0.534.5731.9135.36
1.034.9831.7335.95
1.534.7732.0835.57
2.034.4131.9235.15
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.256, + 0.47, + 0.283 + ], + "angle": 0, + "content": "Table 5. Ablation on the type and weight \\( \\lambda \\) of the image classification loss. The best results are marked in bold." + }, + { + "type": "table", + "bbox": [ + 0.114, + 0.289, + 0.434, + 0.354 + ], + "angle": 0, + "content": "
Layer structureFullRareNon-Rare
S→C→F34.7332.0935.52
C→S→F34.9831.7335.95
C→F34.6331.3435.61
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.365, + 0.47, + 0.406 + ], + "angle": 0, + "content": "Table 6. Ablation on structure of each layer in the category decoder. Here \"S\", \"C\" and \"F\" stands for the self-attention, cross-attention and FFN in a standard transformer [35]." + }, + { + "type": "table", + "bbox": [ + 0.159, + 0.412, + 0.39, + 0.476 + ], + "angle": 0, + "content": "
LFullRareNon-Rare
134.6631.4735.61
234.9831.7335.95
334.8631.5635.85
" + }, + { + "type": "table_caption", + "bbox": [ + 0.078, + 0.487, + 0.468, + 0.501 + ], + "angle": 0, + "content": "Table 7. Ablation on the number of layers in the category decoder." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.52, + 0.469, + 0.58 + ], + "angle": 0, + "content": "weight (Eq. (4)). This is the key component of the proposed method, which makes a distinction between the proposed method and a simple adaption of Query2label [25] to HOI task. C3 denotes the score integration step in Sec. 3.2." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.581, + 0.469, + 0.673 + ], + "angle": 0, + "content": "To validate our approach in Fig. 2, several variants with these components in our approach are experimented and summarized in Tab. 4. First, (b) is a simple combination of Query2label [25] and the baseline detector (a) in a multitask setting. (b) is only slightly better than (a), showing that simply applying Query2Label to HOI is barely helpful." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.673, + 0.469, + 0.763 + ], + "angle": 0, + "content": "Second, variant (c) significantly boosts (b), indicating using the queries as adaptive classification weights is the key to the performance improvement. This shows the effectiveness of our most crucial technical design in C2: applying the learned category queries as adaptive interaction classification weights." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.764, + 0.469, + 0.824 + ], + "angle": 0, + "content": "Finally, the complete approach is (d). It further adds the integration step of image classification and instance classification score on (c). This technique produces moderate improvement over (c), i.e., \\(1.67\\mathrm{mAP}\\) vs. \\(1.29\\mathrm{mAP}\\)." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.825, + 0.469, + 0.9 + ], + "angle": 0, + "content": "Image classification loss. Tab. 5 compares different loss functions and weights. First, when \\(\\lambda = 0\\), which means no image-level supervision is applied, the improvement over the baseline (33.69 mAP) drops to only 0.52 mAP. This demonstrates the image classification supervision is essen" + }, + { + "type": "image", + "bbox": [ + 0.508, + 0.093, + 0.889, + 0.357 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.372, + 0.892, + 0.496 + ], + "angle": 0, + "content": "Figure 4. Some qualitative comparison between the baseline and the proposed method on HICO-DET. From left to right, column 1: true positive (TP) detection results, whose interaction score is increased by our method; column 2: false positive (FP) detection results, whose interaction score is decreased by our method; column 3: corresponding image-level GT and predictions by our method. Scores on the left and right of an image are the interaction classification scores of the visualized instance from the baseline and our method. Best viewed in color. More in supplementary materials." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.515, + 0.891, + 0.56 + ], + "angle": 0, + "content": "tial to make the category query learning effective, with either focal loss or ASL. Additionally, ASL is slightly better than focal loss. By default, \\(\\lambda = 1.0\\) is adopted." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.562, + 0.892, + 0.652 + ], + "angle": 0, + "content": "Is asymmetric loss the key? In the ablation above, we can see that ASL does help our image-level query learning. However, it is not the major reason for the performance improvement. To figure out this, we replace focal loss in the plain baseline GEN-VLKT with ASL and the result is only slightly better by \\(0.08\\mathrm{mAP}\\)." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.653, + 0.892, + 0.773 + ], + "angle": 0, + "content": "Decoder structure. Tab. 6 compares several structures of the decoder. Compared to the standard decoder [2, 35] \\((\\mathrm{S} \\rightarrow \\mathrm{C} \\rightarrow \\mathrm{F}\\) in the table), putting cross-attention first \\((\\mathrm{C} \\rightarrow \\mathrm{S} \\rightarrow \\mathrm{F})\\) is slightly better (by \\(0.25 \\mathrm{mAP}\\)) without extra computation. If we remove self-attention \\((\\mathrm{C} \\rightarrow \\mathrm{F})\\), the performance drops by \\(0.35 \\mathrm{mAP}\\) compared with the \\(\\mathrm{S} \\rightarrow \\mathrm{C} \\rightarrow \\mathrm{F}\\) setting. This is probably because self-attention helps to learn the dependencies between different category queries." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.774, + 0.892, + 0.82 + ], + "angle": 0, + "content": "Tab. 7 compares different numbers of decoder layers, denoted as \\( L \\). We find that \\( L = 2 \\) is sufficient. More layers do not help the performance." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.832, + 0.727, + 0.848 + ], + "angle": 0, + "content": "5.5. Discussions and Analysis" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.856, + 0.892, + 0.901 + ], + "angle": 0, + "content": "To understand why the category query learning is effective for human-object interaction classification, we provide some analysis and qualitative results." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "15281" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.104, + 0.094, + 0.444, + 0.212 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.102, + 0.218, + 0.441, + 0.334 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.338, + 0.47, + 0.42 + ], + "angle": 0, + "content": "Figure 5. Performance evaluated on image partitions with different \\( n = 1,2,3,4,5, > 5 \\) instances for an interaction category. Top: mAP numbers for three baselines (dashed) as well as our integration (solid). Bottom: relative mAP ratios for improvement produced by our method (solid) and between two arbitrary previous methods (dashed)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.443, + 0.468, + 0.579 + ], + "angle": 0, + "content": "The attention maps in the cross-attention layer of the decoder is visualized in Fig. 3. For different category queries, the corresponding attention maps show they learn to capture the semantics of the category, while being adaptive to different images. For example, in Fig. 3a, the broken part of the bat in the air is highlighted for \"break\" while the left part in the hand is highlighted for \"hold\". This is similar for Fig. 3b. In Fig. 3c, the attention map highlights many instances with corresponding action \"hold\" and \"ride\"." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.581, + 0.468, + 0.792 + ], + "angle": 0, + "content": "To qualitatively demonstrate how our method helps, we visualize some cases of the baseline and the proposed method in Fig. 4. In the first case, the baseline predicts a TP of \"person sit on bus\" with a low score 0.15, and a FP of \"person board bus\" with a high score 0.29. Our model also predicts the same TP and FP, but lowers the FP score to 0.07 and lifts the TP score to 0.38. Besides, our model correctly predicts image-level scores: the wrong category \"board\" is given a low score of 0.06 while the four correct categories are given high scores. In the second case, our model successfully predicts the TP \"person dribble\" missed by the baseline and suppress the FP \"person hold sports\" from 0.6 to 0.11 with the help of correct image-level classification result. This is similar for the third case." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.796, + 0.468, + 0.901 + ], + "angle": 0, + "content": "Last, as our category query learning is performed on image level, we conjecture that it is more helpful for images with dense human-object interactions. In such images, an human-object instance is relatively small and hard to learn good feature on its own. However, it may benefit more from the global image level category query feature, which aggregates more information from other similar instances in this" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.287 + ], + "angle": 0, + "content": "image. To validate this conjecture, we partition the images according to their \"interaction density\" and check whether our method produces larger improvement on images that are \"denser\". Specifically, for each interaction category, its mAP is evaluated separately on six different image partition subsets, where each image contains different numbers \\((n = 1, 2, 3, 4, 5\\) and \\(>5\\)) of human-object instances of this category. The mAP results of the three baseline methods and their integrated versions (as in Tab. 1) are shown in Fig. 5 (top, dashed vs. solid lines). It shows that: 1) the mAP is lower for larger n, indicating the \"denser\" images are more challenging; 2) our method improves the baselines consistently on all different partitions." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.288, + 0.892, + 0.409 + ], + "angle": 0, + "content": "To check whether the proposed method is more effective on \"denser\" images, we use the relative mAP improvement, which is a ratio, \\(\\frac{mAP_{ours} - mAP_{baseline}}{mAP_{baseline}}\\), for analysis. The ratio curves of the three baselines are shown in Fig. 5 (bottom, solid lines). It is clear that the relative improvement becomes larger for larger \\(n\\). This indicates that the image-level category query learning is more effective on these challenging dense images." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.409, + 0.892, + 0.53 + ], + "angle": 0, + "content": "To further verify that this behavior is not commonly true, we also compute the relative ratio of three more comparisons, \"gen-vlkt vs. qpic\", \"gen-vlkt vs. scg\" and \"scg vs. qpic\", in which the former outperforms the latter. These curves are also shown in Fig. 5 (bottom, dashed lines). There is no clear pattern in these curves, indicating that the performance gap between two arbitrary HOI methods are in general not related to the image \"density\"." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.544, + 0.619, + 0.559 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.569, + 0.892, + 0.674 + ], + "angle": 0, + "content": "This work proposes a novel approach for the human-object interaction classification sub-task in HOI detection. We study the problem of interaction category modeling, in contrast to most previous methods focusing on human-object feature learning. We adopt the concept of category query in a previous method [25] for HOI, for the first time, and show that it is simple, general and highly effective." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.675, + 0.892, + 0.718 + ], + "angle": 0, + "content": "Clearly, this idea of category query modeling is not limited to multi-label image classification and HOI detection. We hope it is useful for other vision tasks." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Acknowledgment. This work was supported in part by the National Natural Science Foundation of China under Grant 62076183, 61936014 and 61976159, in part by the Natural Science Foundation of Shanghai under Grant 20ZR1473500, in part by the Shanghai Science and Technology Innovation Action Project of under Grant 20511100700 and 22511105300, in part by the Shanghai Municipal Science and Technology Major Project under Grant 2021SHZDZX0100, and in part by the Fundamental Research Funds for the Central Universities. The authors would also like to thank the anonymous reviewers for their careful work and valuable suggestions." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.956 + ], + "angle": 0, + "content": "15282" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.105 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.17 + ], + "angle": 0, + "content": "[1] A Bansal, S. S Rambhatla, A Shrivastava, and R Chellappa. Detecting human-object interactions via functional generalization. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 34, pages 10460-10469, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.171, + 0.472, + 0.226 + ], + "angle": 0, + "content": "[2] N Carion, F Massa, G Synnaeve, N Usunier, A Kirillov, and S Zagoruyko. End-to-end object detection with transformers. In European Conference on Computer Vision, pages 213-229. Springer, 2020. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.228, + 0.471, + 0.282 + ], + "angle": 0, + "content": "[3] Y.-W Chao, Y Liu, X Liu, H Zeng, and J Deng. Learning to detect human-object interactions. In 2018 iee winter conference on applications of computer vision (wacv), pages 381-389. IEEE, 2018. 1, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.284, + 0.471, + 0.338 + ], + "angle": 0, + "content": "[4] M Chen, Y Liao, S Liu, Z Chen, F Wang, and C Qian. Reformulating hoi detection as adaptive set prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9004–9013, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.34, + 0.471, + 0.407 + ], + "angle": 0, + "content": "[5] B Cheng, I Misra, A. G Schwing, A Kirillov, and R Girdhar. Masked-attention mask transformer for universal image segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1290–1299, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.41, + 0.471, + 0.478 + ], + "angle": 0, + "content": "[6] L Dong, Z Li, K Xu, Z Zhang, L Yan, S Zhong, and X Zou. Category-aware transformer network for better human-object interaction detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19538-19547, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.48, + 0.471, + 0.535 + ], + "angle": 0, + "content": "[7] C Gao, J Xu, Y Zou, and J.-B Huang. Drg: Dual relation graph for human-object interaction detection. In European Conference on Computer Vision, pages 696-712. Springer, 2020. 1, 2, 4, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.536, + 0.471, + 0.577 + ], + "angle": 0, + "content": "[8] C Gao, Y Zou, and J.-B Huang. ican: Instance-centric attention network for human-object interaction detection. arXiv preprint arXiv:1808.10437, 2018. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.579, + 0.471, + 0.633 + ], + "angle": 0, + "content": "[9] G Gkioxari, R Girshick, P Dólár, and K He. Detecting and recognizing human-object interactions. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 8359-8367, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.635, + 0.47, + 0.662 + ], + "angle": 0, + "content": "[10] S Gupta and J Malik. Visual semantic role labeling. arXiv preprint arXiv:1505.04474, 2015. 1, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.664, + 0.47, + 0.731 + ], + "angle": 0, + "content": "[11] T Gupta, A Schwing, and D Hoiem. No-frills human-object interaction detection: Factorization, layout encodings, and training techniques. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9677-9685, 2019. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.733, + 0.47, + 0.801 + ], + "angle": 0, + "content": "[12] H He, Y Yuan, X Yue, and H Hu. Rankseg: Adaptive pixel classification with image category ranking for segmentation. In Proceedings of the European Conference on Computer Vision (ECCV), pages 682-700. Springer Nature Switzerland Cham, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.803, + 0.47, + 0.871 + ], + "angle": 0, + "content": "[13] A Iftekhar, H Chen, K Kundu, X Li, J Tighe, and D Modolo. What to look at and where: Semantic and spatial refined transformer for detecting human-object interactions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5353-5363, 2022. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.47, + 0.901 + ], + "angle": 0, + "content": "[14] B Kim, T Choi, J Kang, and H. J Kim. Uniondet: Union-level detector towards real-time human-object interaction de" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "tlection. In European Conference on Computer Vision, pages 498-514. Springer, 2020. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.121, + 0.892, + 0.176 + ], + "angle": 0, + "content": "[15] B Kim, J Lee, J Kang, E.-S Kim, and H. J Kim. Hotr: End-to-end human-object interaction detection with transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 74-83, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.177, + 0.892, + 0.245 + ], + "angle": 0, + "content": "[16] B Kim, J Mun, K.-W On, M Shin, J Lee, and E.-S Kim. Mstr: Multi-scale transformer for end-to-end human-object interaction detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19578-19587, 2022. 1, 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.247, + 0.892, + 0.302 + ], + "angle": 0, + "content": "[17] D.-J Kim, X Sun, J Choi, S Lin, and I. S Kweon. Detecting human-object interactions with action co-occurrence priors. In European Conference on Computer Vision, pages 718–736. Springer, 2020. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.303, + 0.892, + 0.356 + ], + "angle": 0, + "content": "[18] Y.-L Li, X Liu, X Wu, Y Li, and C Lu. Hoi analysis: Integrating and decomposing human-object interaction. Advances in Neural Information Processing Systems, 33:5011-5022, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.358, + 0.892, + 0.426 + ], + "angle": 0, + "content": "[19] Y.-L Li, L Xu, X Liu, X Huang, Y Xu, S Wang, H.-S Fang, Z Ma, M Chen, and C Lu. Pastanet: Toward human activity knowledge engine. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 382-391, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.428, + 0.892, + 0.496 + ], + "angle": 0, + "content": "[20] Y.-L Li, S Zhou, X Huang, L Xu, Z Ma, H.-S Fang, Y Wang, and C Lu. Transferable interactiveness knowledge for human-object interaction detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3585-3594, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.498, + 0.892, + 0.566 + ], + "angle": 0, + "content": "[21] Y Liao, S Liu, F Wang, Y Chen, C Qian, and J Feng. Ppdm: Parallel point detection and matching for real-time human-object interaction detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 482-490, 2020. 1, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.567, + 0.892, + 0.635 + ], + "angle": 0, + "content": "[22] Y Liao, A Zhang, M Lu, Y Wang, X Li, and S Liu. Gen-vlkt: Simplify association and enhance interaction understanding for hoi detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20123-20132, 2022. 1, 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.637, + 0.892, + 0.691 + ], + "angle": 0, + "content": "[23] T.-Y Lin, P Goyal, R Girshick, K He, and P Dollar. Focal loss for dense object detection. In Proceedings of the IEEE international conference on computer vision, pages 2980-2988, 2017. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.692, + 0.892, + 0.748 + ], + "angle": 0, + "content": "[24] T.-Y Lin, M Maire, S Belongie, J Hays, P Perona, D Ramanan, P Dollár, and C. L Zitnick. Microsoft coco: Common objects in context. In European conference on computer vision, pages 740-755. Springer, 2014. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.749, + 0.892, + 0.789 + ], + "angle": 0, + "content": "[25] S Liu, L Zhang, X Yang, H Su, and J Zhu. Query2label: A simple transformer way to multi-label classification. arXiv preprint arXiv:2107.10834, 2021. 1, 2, 3, 4, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.79, + 0.892, + 0.845 + ], + "angle": 0, + "content": "[26] X Liu, Y.-L Li, X Wu, Y.-W Tai, C Lu, and C.-K Tang. Interactiveness field in human-object interactions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20113-20122, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.846, + 0.892, + 0.899 + ], + "angle": 0, + "content": "[27] Y Liu, Q Chen, and A Zisserman. Amplifying key cues for human-object-interaction detection. In European Conference on Computer Vision, pages 248-265. Springer, 2020. 2" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.892, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15283" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.161 + ], + "angle": 0, + "content": "[28] J Park, S Lee, H Heo, H. K Choi, and H. J Kim. Consistency learning via decoding path augmentation for transformers in human object interaction detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1019-1028, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.162, + 0.47, + 0.218 + ], + "angle": 0, + "content": "[29] S Qi, W Wang, B Jia, J Shen, and S.-C Zhu. Learning human-object interactions by graph parsing neural networks. In Proceedings of the European Conference on Computer Vision (ECCV), pages 401-417, 2018. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.218, + 0.471, + 0.286 + ], + "angle": 0, + "content": "[30] X Qu, C Ding, X Li, X Zhong, and D Tao. Distillation using oracle queries for transformer-based human-object interaction detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19558-19567, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.288, + 0.471, + 0.357 + ], + "angle": 0, + "content": "[31] A Radford, J. W Kim, C Hallacy, A Ramesh, G Goh, S Agarwal, G Sastry, A Askell, P Mishkin, J Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.358, + 0.47, + 0.425 + ], + "angle": 0, + "content": "[32] T Ridnik, E Ben-Baruch, N Zamir, A Noy, I Friedman, M Protter, and L Zelnik-Manor. Asymmetric loss for multi-label classification. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 82-91, 2021. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.427, + 0.47, + 0.496 + ], + "angle": 0, + "content": "[33] M Tamura, H Ohashi, and T Yoshinaga. Qpic: Query-based pairwise human-object interaction detection with image-wide contextual information. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10410-10419, 2021. 1, 2, 4, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.497, + 0.47, + 0.566 + ], + "angle": 0, + "content": "[34] O Ulutan, A Iftekhar, and B. S Manjunath. Vsgnet: Spatial attention network for detecting human object interactions using graph convolutions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13617-13626, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.567, + 0.469, + 0.621 + ], + "angle": 0, + "content": "[35] A Vaswani, N Shazeer, N Parmar, J Uszkoreit, L Jones, A. N Gomez, L Kaiser, and I Polosukhin. Attention is all you need. In Advances in neural information processing systems, pages 5998-6008, 2017. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.622, + 0.469, + 0.678 + ], + "angle": 0, + "content": "[36] B Wan, D Zhou, Y Liu, R Li, and X He. Pose-aware multilevel feature network for human object interaction detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9469-9478, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.679, + 0.469, + 0.734 + ], + "angle": 0, + "content": "[37] H Wang, W.-s Zheng, and L Yingbiao. Contextual heterogeneous graph network for human-object interaction detection. In European Conference on Computer Vision, pages 248-264. Springer, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.735, + 0.469, + 0.802 + ], + "angle": 0, + "content": "[38] T Wang, T Yang, M Danelljan, F. S Khan, X Zhang, and J Sun. Learning human-object interaction detection using interaction points. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4116-4125, 2020. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.804, + 0.469, + 0.871 + ], + "angle": 0, + "content": "[39] X Wu, Y.-L Li, X Liu, J Zhang, Y Wu, and C Lu. Mining cross-person cues for body-part interactiveness learning in hoi detection. In Proceedings of the European Conference on Computer Vision (ECCV), pages 121-136. Springer, 2022. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.469, + 0.902 + ], + "angle": 0, + "content": "[40] H Yuan, J Jiang, S Albanie, T Feng, Z Huang, D Ni, and M Tang. Rlip: Relational language-image pre-training for" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.471, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "human-object interaction detection. In Advances in Neural Information Processing Systems, 2022. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.122, + 0.892, + 0.177 + ], + "angle": 0, + "content": "[41] A Zhang, Y Liao, S Liu, M Lu, Y Wang, C Gao, and X Li. Mining the benefits of two-stage and one-stage hoi detection. In Thirty-Fifth Conference on Neural Information Processing Systems, 2021. 1, 2, 4, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.179, + 0.892, + 0.234 + ], + "angle": 0, + "content": "[42] F. Z Zhang, D Campbell, and S Gould. Spatially conditioned graphs for detecting human-object interactions. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13319-13327, 2021. 2, 4, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.235, + 0.892, + 0.303 + ], + "angle": 0, + "content": "[43] F. Z Zhang, D Campbell, and S Gould. Efficient two-stage detection of human-object interactions with a novel unary-pairwise transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20104-20112, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.305, + 0.892, + 0.375 + ], + "angle": 0, + "content": "[44] Y Zhang, Y Pan, T Yao, R Huang, T Mei, and C.-W Chen. Exploring structure-aware transformer over interaction proposals for human-object interaction detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19548-19557, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.376, + 0.892, + 0.444 + ], + "angle": 0, + "content": "[45] X Zhong, C Ding, Z Li, and S Huang. Towards hard-positive query mining for detr-based human-object interaction detection. In Proceedings of the European Conference on Computer Vision (ECCV), pages 444-460. Springer, 2022. 2, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.446, + 0.892, + 0.5 + ], + "angle": 0, + "content": "[46] X Zhong, C Ding, X Qu, and D Tao. Polysemy deciphering network for robust human-object interaction detection. International Journal of Computer Vision, 129(6):1910-1929, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.502, + 0.892, + 0.571 + ], + "angle": 0, + "content": "[47] X Zhong, X Qu, C Ding, and D Tao. Glance and gaze: Inferring action-aware points for one-stage human-object interaction detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13234-13243, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.573, + 0.892, + 0.642 + ], + "angle": 0, + "content": "[48] D Zhou, Z Liu, J Wang, L Wang, T Hu, E Ding, and J Wang. Human-object interaction detection via disentangled transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19568-19577, 2022. 1, 2, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.643, + 0.892, + 0.713 + ], + "angle": 0, + "content": "[49] C Zou, B Wang, Y Hu, J Liu, Q Wu, Y Zhao, B Li, C Zhang, C Zhang, Y Wei, et al. End-to-end human object interaction detection with hoi transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11825-11834, 2021. 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.713 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "15284" + } + ] +] \ No newline at end of file diff --git a/2023/Category Query Learning for Human-Object Interaction Classification/c48c8f3f-b487-4c24-815b-6385a57d02e8_origin.pdf b/2023/Category Query Learning for Human-Object Interaction Classification/c48c8f3f-b487-4c24-815b-6385a57d02e8_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..30f06748bd7e650509ce8b6338dd5596a1e826c9 --- /dev/null +++ b/2023/Category Query Learning for Human-Object Interaction Classification/c48c8f3f-b487-4c24-815b-6385a57d02e8_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:eee6d4cef4525e454e66e376d50aaecbd2f184a79f82c7d9dc31c55cd18bdf0a +size 2418599 diff --git a/2023/Category Query Learning for Human-Object Interaction Classification/full.md b/2023/Category Query Learning for Human-Object Interaction Classification/full.md new file mode 100644 index 0000000000000000000000000000000000000000..84b303c7b886cb911212f1733c524c6a28fcacfd --- /dev/null +++ b/2023/Category Query Learning for Human-Object Interaction Classification/full.md @@ -0,0 +1,354 @@ +# Category Query Learning for Human-Object Interaction Classification + +Chi Xie $^{1\dagger}$ Fangao Zeng $^{2‡}$ Yue Hu $^{3‡}$ Shuang Liang $^{1*}$ Yichen Wei $^{2‡}$ $^{1}$ Tongji University $^{2}$ MEGVII Technology $^{3}$ Shanghai Jiao Tong University $^{1}$ {chixie, shuangliang}@tongji.edu.cn + +2zfg472988436@163.com,wei_yi_chen@hotmail.com 318671129361@sjtu.edu.cn + +# Abstract + +Unlike most previous HOI methods that focus on learning better human-object features, we propose a novel and complementary approach called category query learning. Such queries are explicitly associated to interaction categories, converted to image specific category representation via a transformer decoder, and learnt via an auxiliary image-level classification task. This idea is motivated by an earlier multi-label image classification method, but is for the first time applied for the challenging human-object interaction classification task. Our method is simple, general and effective. It is validated on three representative HOI baselines and achieves new state-of-the-art results on two benchmarks. Code will be available at https://github.com/charles-xie/CQL. + +# 1. Introduction + +Human-Object Interaction (HOI) detection has attracted a lot of interests in recent years [3, 8-10, 21, 33]. The task consists of two sub-tasks. The first is human and object detection. It is usually performed by common object detection methods. The second is interaction classification of each human-object (HO) pair. This sub-task is very challenging due to the complex appearance variations in the interaction categories. See Fig. 1 for examples. It is the focus of most previous HOI methods, as well as this work. + +Most previous HOI methods focus on learning better human-object features, including modeling relation and context via GNN [7, 29, 34, 37] or attention mechanism [8, 34, 46], decoupling localization and classification [22, 41, 48], leveraging vision-language knowledge [6, 22] and introducing multi-scale feature to transformer [16]. However, for interaction classification they all adopt the simple linear classifier that performs the dot product of the + +![](images/3075a4e4a0d7b0b2428e126ff8afd3b3dc0a7619be00fad2022df5f19c9f8ac1.jpg) +(a) person fly kite + +![](images/34ce8e97a2ba6758f9ba26c1eaf19d64dcf450e62bd5da932f9f922319a890f2.jpg) +person fly airplane + +![](images/02d5d4e4abf81c7fe8dd3fb7833919f4d4f4db83d6e10ced0f40417a6658b050.jpg) +(b) person hold fork + +![](images/afe9e0d7b15e9a43e4021404cf5e3e203cf80b73db034c463fefc552f98f45e0.jpg) +person hold elephant +Figure 1. Interaction classification is inherently challenging. In (a), "fly" is semantically polysemic, resulting in different objects, poses and relative positions. In (b), when "hold" is associated with different objects, the appearance, scene background, and human poses are largely different. + +human-object feature and a static weight vector, which represents an interaction category. + +In this work, we propose a new approach that enhances the above paradigm and complements most previous HOI methods. It is motivated by the recent work Query2label [25], a transformer-based classification network. It proposes a new concept we call category-specific query. Unlike the queries in other transformer methods, each query is associated to a specific and fixed image category during training and inference. This one-to-one binding makes the query learn to model each category more effectively. The queries are converted to image specific category representations via a transformer decoder. This method achieves excellent performance on multi-label image classification task. + +We extend this approach for human-object interaction classification. Essentially, our approach replaces traditional category representation as a static weight vector in previous HOI methods with category queries learnt as described above. The same linear classifier is adopted. Such cate + +gory queries are more effective, and adaptive for different images, giving rise to better modeling of the complex variations in each interaction category. This is the crucial difference between this work and a simple adaption of [25] to HOI. Notably, this work is the first to address the category weight representation problem in the HOI community. + +Note that our proposed category specific query is different and not related to those queries in other transformer-based HOI methods [4, 15, 33, 49]. Specifically, category queries extract image-level features as the category representation. The queries in other methods are human-object instance-level features and category-agnostic. + +Our method is simple, lightweight and general. The overview is in Fig. 2. It is complementary to any off-the-shelf HOI method that provides human-object features. The modification of both inference and training is small. The incurred additional cost is marginal. + +In experiments, our approach is validated on three representative and strong HOI baseline methods, two transformer-based methods [22, 33] and a traditional two-stage method [42]. They are all significantly improved by our approach. New state-of-the-art results are obtained on two benchmarks. Specifically, we obtain $36.03\mathrm{mAP}$ on HICO-DET. Comprehensive ablation studies and in-depth discussion are also provided to verify the effectiveness of implementation details in our approach. It turns out that our method is more effective on challenging images that contain more human-object instances, a property that is rarely discussed by previous HOI methods. + +# 2. Related Work + +# 2.1. Instance Query Learning in HOI Detection + +DETR [2] firstly proposes the concept of object instance query for object detection task. Such queries essentially learn the priors of both object appearance and spatial location. DETR leverages those queries to probe image features through a transformer [35] and localize unique objects in the image. Motivated by its great success, many works [15, 22, 33, 41, 48, 49] adapt such detection transformer framework to HOI detection by simply treating the HOI triplet [33, 49] or H-O pair [4, 15] as an object. A few of them [6, 30, 45] pay attention to adapting the plain query to this task. DOQ [30] proposes a knowledge distillation model using oracle queries to facilitate the representation learning of a transformer-based detector; HQM [45] explicitly constructs hard positive queries from ground truth to train the model to be less vulnerable to spacial variations; CATN [6] utilizes the object category prior generated from external object detector and language model for query initialization. In summary, these transformer-based methods use each query to aggregate context information not restricted to one interaction category, in order to predict + +a potential HOI instance at a specific location. + +In DETR [2] and its variants [4,15,22,33,41,49] in HOI, as the queries are category-agnostic, their association to object categories are dynamic and unstable during training. This could be problematic. For example, it is well known that the convergence of DETR training is slow. In contrast, our proposed query is category-specific. The learning is guided by image-level classification task and stable. Such queries learn category-specific priors and are good representation for interaction categories. + +# 2.2. Feature Learning in HOI Detection + +Early methods. Based on two-stage detection framework, early works make many efforts to help feature learning, including employing architectures effective in modeling relation and context like GNN [7, 29, 34, 37] and attention module [8, 34, 46], leveraging fine-grained visual features [11, 17, 19, 20, 36] like human pose and introducing language prior [1, 7, 17, 27, 46]. + +Transformer-based methods. Motivated by DETR [49], many methods [4, 15, 33, 49] leverage transformer architecture [35] and extend the object query in DETR to HOI query. With the help of HOI query and transformer's built-in attention mechanism, those methods learn effective feature representation for HOI triplet or H-O pair. + +Based on those pioneer transformer-based methods, recently, many methods are proposed to further help feature learning, by decoupling H-O pair localization and interaction classification [22, 41, 48] or exploiting multi-scale feature in transformer architecture [16]. Some works [6, 22] leverage vision-language knowledge in CLIP [31] or design a pretrained model [40] specifically for HOI; others utilize information like human poses [39] or spatial configurations [13] that has been used in early HOI detectors. + +Relation to the proposed method. Previous methods learn the H-O feature while ours learns the category query as the category representation feature. Thus, they are complementary. The interaction classification is simply by measuring the similarity between the two types of features. The integration of our method to previous HOI methods is simple. + +# 3. Our Method + +The overview of our method is in Fig. 2. It consists of two components, the image-level category query learning (top block) and human-object interaction classification (bottom right block). + +The first component is detailed in Sec. 3.1. It is briefly summarized here. A number of queries (embedding vectors) are associated to human-object interaction category, in a one-to-one manner. Such queries interact with image features (provided by a baseline HOI method) through a transformer decoder [35] and become image-specific queries. + +![](images/a2e1ee732a1a14a4cdac05c06211e634cc8a176b620996b5cf18b6d23f76a740.jpg) +Figure 2. Overview of our method. It consists of two components (top and bottom right). It can be integrated with any baseline HOI method (bottom left) that provides image feature $I$ and human-object instance features $F_{i}$ . See Sec. 3 for details. + +Learning of both the queries and decoder weights is supervised by an auxiliary image-level classification task. In this way, the queries are learnt to capture category-specific feature and become good feature representation for these categories. Besides some minor details, this step is the same as the previous work Query2label [25], which is for multi-label image classification task. + +The second component is detailed in Sec. 3.2. For the first time, we adopt the category query learning method for human-object interaction classification tasks. The cosine similarity between the category query and human-object feature is used for interaction classification. Thus, it works with any HOI method that provides human-object features. Besides, the image-level classification results turn out moderately helpful in an score integration step, which is an extra technique that benefits the performance. + +Overall, the proposed method is simple, effective, lightweight and general. It can be combined with most previous HOI detection methods (bottom left block in Fig. 2), with small modification, as elaborated in Sec. 4. + +# 3.1. Image-level Category Query Learning + +Similar as in Query2Label [25], for $K$ human-object interaction categories, we define their one-to-one corresponding category queries, which are learnable embedding vectors, $\{Q_k\} \in \mathbb{R}^{K\times D}$ , where $D$ is the vector dimension. + +Each query $Q_{k}$ aggregates image features $\mathbf{I} \in \mathbb{R}^{H \times W \times D}$ through a transformer decoder and is updated to image specific query $Q_{k}'$ , + +$$ +\left\{Q _ {k} ^ {\prime} \right\} = \operatorname {d e c o d e r} \left(\left\{Q _ {k} \right\}, I\right). \tag {1} +$$ + +Note that the decoder structure has several variants, which are studied in [5]. Our experiments show that the + +structure is of minor importance, as discussed in Sec. 5.4. Specifically, our decoder consists of two layers, each of which consisting of a cross-attention layer, a self-attention layer and a FFN layer, in order. + +Then, each image-level classification probabilities $p_k$ is computed by applying a category-specific fully-connected layer and a sigmoid activation on the updated query $Q_{k}^{\prime}$ . + +$$ +p _ {k} = \operatorname {s i g m o i d} \left(\mathrm {F C} \left(Q _ {k} ^ {\prime}\right)\right) \tag {2} +$$ + +Learning of the category query $\{Q_k\}$ and the decoder weights is supervised by common image classification losses. To deal with the label imbalance problem, focal loss [23] and asymmetric loss (ASL) [32] are used. Asymmetric loss is a variant of focal loss. It is more robust for high label imbalance and noises. Our experiments (see Sec. 5.4) show that it is slightly better. + +Specifically, with the classification probability $p_k$ and the shifted probability $p_k' = \max(p_k - m, 0)$ , the asymmetric loss is + +$$ +\mathcal {L} _ {i m g} = \frac {1}{K} \sum_ {k = 1} ^ {K} \left\{ \begin{array}{l l} (1 - p _ {k}) ^ {\gamma +} \log \left(p _ {k}\right), & y _ {k} = 1, \\ \left(p _ {k} ^ {\prime}\right) ^ {\gamma -} \log \left(1 - p _ {k} ^ {\prime}\right), & y _ {k} = 0, \end{array} \right. \tag {3} +$$ + +where the binary label $y_{k}$ indicates the existence of category $k$ in the image, and $\gamma +, \gamma -$ as well as $m$ are hyperparameters. We use the default values in ASL [32], $\gamma + = 0$ , $\gamma - = 4$ and $m = 0.05$ . + +In this way, the category queries are learnt to encode the category priors. Figure 3 is the visualization of the heatmaps in the cross-attention layer of the decoder. Each category query learns to locate the human body parts related to discriminative feature of its corresponding interaction category, e.g., in Fig. 3c, the query of "hold" highlights + +the hand region, while in the same image the query of "ride" highlights the foot region. It qualitatively demonstrates that the query learning is effective in encoding category-specific information. + +The updated queries adaptively extract category-related features for each image, with the help of the transformer decoder's built-in multi-head cross-attention layer. + +# 3.2. Interaction Classification with Category Query + +In this step, we apply the updated category queries $\{Q_k'\}$ as the weights for interaction classification. Given the $i$ -th human-object instance, its classification probability score for $k$ -th interaction category is simply the cosine similarity between its feature $F_i$ and the category query feature $Q_k'$ . + +$$ +s _ {i, k} = \operatorname {s i g m o i d} \left(\frac {Q _ {k} ^ {\prime} \cdot F _ {i}}{\left| Q _ {k} ^ {\prime} \right| \times \left| F _ {i} \right|}\right). \tag {4} +$$ + +There are no restrictions or assumptions on the human-object feature $F_{i}$ . Most previous HOI methods should be applicable here. + +In this step, the traditional classification weight from static parameters is replaced with the category query adaptive on each image. This behavior is essentially different from [25], which uses category queries as image feature. + +By "adaptive", we mean that the queries are updated dynamically according to the image contents. As exemplified in Fig. 3a and Fig. 3c, the queries of "hold" learn to highlight different interactive areas in different images and update themselves with features from these areas via attention. This shows the query learning is adaptive to image. + +As discussed in Sec. 5.4, the classification step in Eq. (4) is crucial to make the category query learning effective on human-object interaction classification. Without this step, the image-level category query learning using only image-level classification from Query2label [25] is of little use. + +Score integration step. The segmentation method in [12] discards certain categories during pixel classification that have low image classification scores. Motivated by this method, we take a similar score integration step. The image-level classification score $\{p_i\}$ is used to enhance the human-object instance classification. The idea is that, the instance score should be higher if the image-level probability is higher. Our implementation is similar as in [12]. During both training and testing, for each image, the top- $\kappa$ categories ( $\kappa = 70$ in this work) with higher image classification scores $\{p_i\}$ are selected. The instance score $s_{i,k}$ is slightly modulated such that it becomes higher if the rank of category $k$ is higher. This strategy gives rise to moderate improvement, as verified in Tab. 4. We left the implementation details in the supplementary materials. + +![](images/99f8282d3ac2f3524a9f012e38cb046233dabf2bf8ff19f79576baca3c9a90b0.jpg) + +![](images/fe379372a13e190926267d60ca26365d5a1c8c3fdec1a768296f4aba55d26d7c.jpg) + +![](images/2b0bc16ea2c372fd882e3a9223f14266858b79022b8fc3d45678b1c1a80631f0.jpg) + +![](images/0c455d5cab7e46117e8c591ebceb56b8f57bafb24bc6662232973cbab7dab19f.jpg) +(a) Image with "person-break-baseball bat" and "hold-baseball bat". + +![](images/51e2755600f28b9ba7c8639160c56def36c898efa95c3d24d8d9c08303494f85.jpg) + +![](images/9647b202997c06d7eaa3c35abeecd263747d3dd6ee99e228e39f754859e401d7.jpg) + +![](images/8f7492dcb75316b0373578dff4aa2d091da4a403ff996384eacc1f96bf4dd570.jpg) +(b) Image with "person-direct-airplane" and "fly-airplane". +(c) Image with "person-hold-horse" and "ride-horse". +Figure 3. From left to right: image, attention maps of the cross-attention layer in the decoder for different interaction categories. + +![](images/9ceaf0fc4813a46f45b113af842239ef578ebfa73b0cc7aa99f6cd2a4f59e989.jpg) + +![](images/6aa65ad1b0f1578a4bcc497a7396b04fb3d878abc1f22548143f7c6548f7f973.jpg) + +# 4. Integration to Off-the-shelf HOI Detectors + +As shown in Fig. 2, our method is ready to integrate with any baseline HOI method that provides image feature $\mathbf{I}$ and human-object instance feature $\{F_i\}$ . The integration is simple. During inference, the human-object instance interaction classification part is replaced by our method in Sec. 3.2, the top and bottom right block in Fig. 2. + +During training, the original loss in the baseline HOI method $\mathcal{L}_{base}$ is added to our image classification loss in Eq. (3). The final loss $\mathcal{L}$ for training is + +$$ +\mathcal {L} = \mathcal {L} _ {\text {b a s e}} + \lambda * \mathcal {L} _ {\text {i m g}}, \tag {5} +$$ + +where the weight $\lambda$ is 1.0 by default. All other hyper parameters and details during training remain the same as in the baseline HOI method. + +Thus, our method is general and applicable to most existing HOI methods. In this work, we select three representative yet different baseline methods to verify the effectiveness of our approach, as described below. + +QPIC [33] is the first to introduce transformer method into HOI task. It is also the baseline for many recent works [13, 28, 41, 45, 48]. Its performance is much better than early one-stage [14, 21, 38] and two-stage [7, 11, 17] methods while keeping a simple and end-to-end architecture. It consists of a CNN backbone as well as a transformer encoder and decoder. + +During our integration, the feature map in its transformer encoder is used as the image feature $\mathbf{I}$ . The human-object feature $\{F_i\}$ is the query feature in its decoder. + +SCG [42] is a traditional two-stage method and the best in this category. It is also one of the best methods that + +
MethodPipelineE2EHICO-DETV-COCOEfficiency
FullRareNon-RareS1S2#ParamsFPS
QPIC [33]transformer✓28.9321.6231.1261.3963.6541M19.5
+ Ourstransformer✓31.08(+2.15)23.9033.2263.67(+2.28)65.4946M(+5M)18.3(-6.2%)
SCG [42]two-stageX31.2824.1633.4056.9362.5157M4.5
+ Ourstwo-stageX32.74(+1.46)26.2534.6859.14(+2.21)65.6164M(+7M)4.1(-8.9%)
GEN-VLKT [22]transformerX33.6929.9434.8164.8966.7442M21.7
+ OurstransformerX35.36(+1.67)32.9736.0766.40(+1.51)69.1747M(+5M)20.6(-5.1%)
+ +Table 1. The performance numbers of three different baseline HOI methods with and without integration of our method, on two datasets. "E2E" denotes whether a HOI detector is end-to-end. All models are tested on Tesla V100. + +does not use transformer. It uses a multi-stream graph neural network(GNN) for interaction classification. In our experiment, the detection boxes are from a fine-tuned detector provided by DRG [7] for HICO-DET and a fine-tuned DETR for V-COCO. + +During our integration, the CNN feature map in the backbone of SCG is used as image feature $\mathbf{I}$ . The human-object feature $\{F_i\}$ is generated through RoI pooling with detected human and object boxes and fused with the GNN. + +GEN-VLKT [22] is also transformer-based, but not end-to-end as pairwise NMS [41] is used for post-processing. It is the current state-of-the-art method. It uses two parallel decoders for object detection and interaction classification, namely instance decoder and interaction decoder. + +During our integration, the feature map in its transformer encoder is used as image feature $\mathbf{I}$ . The query feature in the interaction decoder is used as the human-object feature $\{F_i\}$ . Note that, unlike the majority of HOI detection methods, the original GEN-VLKT uses HOI categories rather than interaction categories during interaction classification. Our experiments still use interaction categories, in order to be consistent with most other methods. + +# 5. Experiments + +In this section, we verify the applicability and effectiveness of the proposed method through experiments. In Sec. 5.1, we introduce the experimental settings. Then we demonstrate the effectiveness of the proposed method over 3 baselines in Sec. 5.2, and show it achieves SOTA results on major benchmarks in Sec. 5.3. Next, in Sec. 5.4 we conduct comprehensive ablation studies on the key components as well as detailed technical designs. Lastly, we provide some analysis and visualization in Sec. 5.5. + +# 5.1. Datasets + +HICO-DET [3] and V-COCO [10] are two widely-used HOI benchmarks. HICO-DET contains 47,776 images, with 38,118 for training and 9,658 for testing. There are 600 HOI categories in HICO-DET, consisting of 117 interaction classes and 80 object classes. Each HOI category is composed of an interaction and an object. V-COCO is + +a subset of MS-COCO [24] with HOI annotations, including 10,346 images (2,533 for training, 2,867 for validation and 4,946 for testing). It has 80 object categories same with HICO-DET and 29 interaction categories. + +Evaluation metrics. For HICO-DET, we adopt the commonly used mAP metric [3]. Each prediction is a $\langle$ human, interaction, object $\rangle$ triplet. A prediction is a true positive only when the human and object bounding boxes both have IoU $>0.5$ w.r.t. ground truth and the interaction classification result is correct. We evaluate the performance in two different settings following [3]. In the known object setting, for each HOI category, we evaluate the prediction only on the images containing the target object category. In default setting, the detection result of each category is evaluated on the full test set. In each setting, we report the mAP over (1) all 600 HOI categories (Full), (2) 138 categories with less than 10 training samples (Rare), and (3) the remaining 462 categories (Non-rare). For V-COCO, we use the role mAP following [10], under both scenario #1 (including objects) and #2 (ignoring objects). The performance is evaluated using its official evaluation toolkit. + +# 5.2. Improvement on Three Different Baselines + +Tab. 1 summarizes the performance of the three baseline HOI methods before and after integration of our method. Backbone is ResNet50. All these methods are significantly improved. Specifically, QPIC [33] is improved by $2.15\mathrm{mAP}$ , making it competitive with those more recent works [16, 41, 43]. SCG [42] is improved by 1.46 mAP, demonstrating that our method is not limited to transformer-based baselines. The current SOTA method GEN-VLKT [22] is improved by $1.67\mathrm{mAP}$ , producing the new SOTA result (also refer to Tab. 2). + +On the V-COCO dataset [10], the performance improvement is similar, which is 2.28, 2.21 and $1.51\mathrm{mAP}$ on QPIC, SCG and GEN-VLKT, respectively. + +Notably, for SCG [42], as the object detector is fixed during the training of its interaction classification network, the improvement by our method is purely due to better interaction classification, not a better fine-tuned CNN backbone or a better object detector. This further consolidates that the category query learning is effective. + +
MethodDetectorBackboneDefaultKnown Object
FullRareNon-rareFullRareNon-rare
DRG [7]HICO-DETResNet50-FPN24.5319.4726.0427.9823.1129.43
GG-Net [47]HICO-DETHourglass10423.4716.4825.6027.3620.2329.48
IDN [18]HICO-DETResNet5026.2922.6127.3928.2424.4729.37
QPIC [33]HICO-DETResNet5029.0721.8531.2331.6824.1433.93
SCG [42]HICO-DETResNet50-FPN31.3324.7233.3134.3727.1836.52
CDN [41]HICO-DETResNet5031.7827.5533.0534.5329.7335.96
DT [48]HICO-DETResNet5031.7527.4533.0334.5030.1335.81
STIP [44]HICO-DETResNet5031.6027.7532.7534.4130.1235.69
HQM [45]HICO-DETResNet5032.4728.1533.76---
MSTR [16]HICO-DETResNet5031.1725.3132.9234.0228.8335.57
RLIP [40]COCO+VGResNet5032.8426.8534.63---
IF [26]HICO-DETResNet5033.5130.3034.4636.2833.1637.21
GEN-VLKT-B [22]HICO-DETResNet5033.7529.2535.1036.7832.7537.99
GEN-VLKT-M [22]HICO-DETResNet10134.7831.5035.7738.0734.9439.01
GEN-VLKT-L [22]HICO-DETResNet10134.9531.1836.0838.2234.3639.37
BodyPartMap [39]HICO-DETResNet5035.1533.7135.5837.5635.8738.06
GEN-VLKT-B + OursHICO-DETResNet5035.3632.9736.0738.4334.8539.50
GEN-VLKT-M + OursHICO-DETResNet10135.8332.9136.7038.7935.2839.84
GEN-VLKT-L + OursHICO-DETResNet10136.0333.1636.8938.8235.5139.81
+ +Table 2. The proposed method achieves state-of-the-art on HICO-DET [3]. The best results are marked in bold. + +
MethodBackboneScenario #1Scenario #2
DRG [7]R50FPN51.0-
SCG [42]R5054.260.9
GG-Net [47]HG10454.7-
QPIC [33]R5058.861.0
HQM [45]R5063.6-
CDN [41]R5061.763.8
GEN-VLKT-B [22]R5062.464.5
GEN-VLKT-M [22]R10163.365.6
GEN-VLKT-L [22]R10163.665.9
MSTR [16]R5062.065.2
BodyPartMap [39]R5063.065.1
IF [26]R5063.065.2
DT [48]R5066.268.5
STIP [44]R5065.169.7
GEN-VLKT-B + OursR5066.469.2
GEN-VLKT-M + OursR10166.869.8
GEN-VLKT-L + OursR10166.569.9
+ +To verify that the performance improvement is not due to a larger model, we also compare the model size and running speed. Our method increases the model parameters by a few millions, which is small compared to the original model size. The running speed, measured by FPS, is only decreased by a few percent. The marginal additional cost shows that our method is quite lightweight. + +# 5.3. Comparison with State-of-the-art + +Tab. 2 and Tab. 3 compare our method with many previous methods for HICO-DET and V-COCO datasets, respec + +Table 3. Comparison with state-of-the-art methods on V-COCO [10] dataset. The best results are marked in bold. + +
C1C2C3FullRareNon-Rare
a---33.6929.9434.81
b✓--33.86 (+0.17)31.1234.68
c✓✓-34.98 (+1.29)31.7335.95
d✓✓✓35.36 (+1.67)32.9736.07
+ +Table 4. Ablation study of several variants of our method, starting from the baseline (a) to our approach (d). C1, C2 and C3 are described in Sec. 5.4. The best results are marked in bold. + +tively. GEN-VLKT [22] is used as our baseline. + +On HICO-DET, our result with ResNet50 backbone already outperforms all previous methods under both default and known object settings. With the stronger ResNet101 backbone, our method achieves the new state-of-the-art 36.03 full mAP under default settings and 38.82 under known object settings. + +On V-COCO dataset, our method achieves the new state-of-the-art performance on Scenario 1, with an AP of 66.4 for ResNet50, surpassing [48]. For scenario 2, it is comparable with the state-of-the-art [44]. + +# 5.4. Ablation Experiments + +We perform various ablation experiments to validate the effectiveness of different components in our method. HICO-DET dataset and GEN-VLKT [22] baseline are used. + +The proposed method can be divided into 3 components: C1 means applying Query2Label [25] to the baseline detector as a multi-task learning (with feature extractor shared). In detail, it adds the queries, the decoder and image classification loss (Eq. (1), Eq. (2) and Eq. (5)). C2 means using the learned query in C1 as adaptive interaction classification + +
loss typeλFullRareNon-Rare
-034.2130.1535.42
focal loss [23]0.534.4331.0635.44
1.034.5131.0835.53
1.534.3531.5435.19
2.034.2931.1835.22
ASL [32]0.534.5731.9135.36
1.034.9831.7335.95
1.534.7732.0835.57
2.034.4131.9235.15
+ +Table 5. Ablation on the type and weight $\lambda$ of the image classification loss. The best results are marked in bold. + +
Layer structureFullRareNon-Rare
S→C→F34.7332.0935.52
C→S→F34.9831.7335.95
C→F34.6331.3435.61
+ +Table 6. Ablation on structure of each layer in the category decoder. Here "S", "C" and "F" stands for the self-attention, cross-attention and FFN in a standard transformer [35]. + +
LFullRareNon-Rare
134.6631.4735.61
234.9831.7335.95
334.8631.5635.85
+ +Table 7. Ablation on the number of layers in the category decoder. + +weight (Eq. (4)). This is the key component of the proposed method, which makes a distinction between the proposed method and a simple adaption of Query2label [25] to HOI task. C3 denotes the score integration step in Sec. 3.2. + +To validate our approach in Fig. 2, several variants with these components in our approach are experimented and summarized in Tab. 4. First, (b) is a simple combination of Query2label [25] and the baseline detector (a) in a multitask setting. (b) is only slightly better than (a), showing that simply applying Query2Label to HOI is barely helpful. + +Second, variant (c) significantly boosts (b), indicating using the queries as adaptive classification weights is the key to the performance improvement. This shows the effectiveness of our most crucial technical design in C2: applying the learned category queries as adaptive interaction classification weights. + +Finally, the complete approach is (d). It further adds the integration step of image classification and instance classification score on (c). This technique produces moderate improvement over (c), i.e., $1.67\mathrm{mAP}$ vs. $1.29\mathrm{mAP}$ . + +Image classification loss. Tab. 5 compares different loss functions and weights. First, when $\lambda = 0$ , which means no image-level supervision is applied, the improvement over the baseline (33.69 mAP) drops to only 0.52 mAP. This demonstrates the image classification supervision is essen + +![](images/827f372f985eaa1b832f1f34b3aaa43a0ee9b577522936d14a3ed751e170a219.jpg) +Figure 4. Some qualitative comparison between the baseline and the proposed method on HICO-DET. From left to right, column 1: true positive (TP) detection results, whose interaction score is increased by our method; column 2: false positive (FP) detection results, whose interaction score is decreased by our method; column 3: corresponding image-level GT and predictions by our method. Scores on the left and right of an image are the interaction classification scores of the visualized instance from the baseline and our method. Best viewed in color. More in supplementary materials. + +tial to make the category query learning effective, with either focal loss or ASL. Additionally, ASL is slightly better than focal loss. By default, $\lambda = 1.0$ is adopted. + +Is asymmetric loss the key? In the ablation above, we can see that ASL does help our image-level query learning. However, it is not the major reason for the performance improvement. To figure out this, we replace focal loss in the plain baseline GEN-VLKT with ASL and the result is only slightly better by $0.08\mathrm{mAP}$ . + +Decoder structure. Tab. 6 compares several structures of the decoder. Compared to the standard decoder [2, 35] $(\mathrm{S} \rightarrow \mathrm{C} \rightarrow \mathrm{F}$ in the table), putting cross-attention first $(\mathrm{C} \rightarrow \mathrm{S} \rightarrow \mathrm{F})$ is slightly better (by $0.25 \mathrm{mAP}$ ) without extra computation. If we remove self-attention $(\mathrm{C} \rightarrow \mathrm{F})$ , the performance drops by $0.35 \mathrm{mAP}$ compared with the $\mathrm{S} \rightarrow \mathrm{C} \rightarrow \mathrm{F}$ setting. This is probably because self-attention helps to learn the dependencies between different category queries. + +Tab. 7 compares different numbers of decoder layers, denoted as $L$ . We find that $L = 2$ is sufficient. More layers do not help the performance. + +# 5.5. Discussions and Analysis + +To understand why the category query learning is effective for human-object interaction classification, we provide some analysis and qualitative results. + +![](images/199b250f04f984c175003beebe08d073e57ff36e116e51d60d06aaf1e7eeec19.jpg) + +![](images/2e06852b319c9a0790037fa9831788f097d23b9a7f7b6c7bb0d8c4d1db475004.jpg) +Figure 5. Performance evaluated on image partitions with different $n = 1,2,3,4,5, > 5$ instances for an interaction category. Top: mAP numbers for three baselines (dashed) as well as our integration (solid). Bottom: relative mAP ratios for improvement produced by our method (solid) and between two arbitrary previous methods (dashed). + +The attention maps in the cross-attention layer of the decoder is visualized in Fig. 3. For different category queries, the corresponding attention maps show they learn to capture the semantics of the category, while being adaptive to different images. For example, in Fig. 3a, the broken part of the bat in the air is highlighted for "break" while the left part in the hand is highlighted for "hold". This is similar for Fig. 3b. In Fig. 3c, the attention map highlights many instances with corresponding action "hold" and "ride". + +To qualitatively demonstrate how our method helps, we visualize some cases of the baseline and the proposed method in Fig. 4. In the first case, the baseline predicts a TP of "person sit on bus" with a low score 0.15, and a FP of "person board bus" with a high score 0.29. Our model also predicts the same TP and FP, but lowers the FP score to 0.07 and lifts the TP score to 0.38. Besides, our model correctly predicts image-level scores: the wrong category "board" is given a low score of 0.06 while the four correct categories are given high scores. In the second case, our model successfully predicts the TP "person dribble" missed by the baseline and suppress the FP "person hold sports" from 0.6 to 0.11 with the help of correct image-level classification result. This is similar for the third case. + +Last, as our category query learning is performed on image level, we conjecture that it is more helpful for images with dense human-object interactions. In such images, an human-object instance is relatively small and hard to learn good feature on its own. However, it may benefit more from the global image level category query feature, which aggregates more information from other similar instances in this + +image. To validate this conjecture, we partition the images according to their "interaction density" and check whether our method produces larger improvement on images that are "denser". Specifically, for each interaction category, its mAP is evaluated separately on six different image partition subsets, where each image contains different numbers $(n = 1, 2, 3, 4, 5$ and $>5$ ) of human-object instances of this category. The mAP results of the three baseline methods and their integrated versions (as in Tab. 1) are shown in Fig. 5 (top, dashed vs. solid lines). It shows that: 1) the mAP is lower for larger n, indicating the "denser" images are more challenging; 2) our method improves the baselines consistently on all different partitions. + +To check whether the proposed method is more effective on "denser" images, we use the relative mAP improvement, which is a ratio, $\frac{mAP_{ours} - mAP_{baseline}}{mAP_{baseline}}$ , for analysis. The ratio curves of the three baselines are shown in Fig. 5 (bottom, solid lines). It is clear that the relative improvement becomes larger for larger $n$ . This indicates that the image-level category query learning is more effective on these challenging dense images. + +To further verify that this behavior is not commonly true, we also compute the relative ratio of three more comparisons, "gen-vlkt vs. qpic", "gen-vlkt vs. scg" and "scg vs. qpic", in which the former outperforms the latter. These curves are also shown in Fig. 5 (bottom, dashed lines). There is no clear pattern in these curves, indicating that the performance gap between two arbitrary HOI methods are in general not related to the image "density". + +# 6. Conclusion + +This work proposes a novel approach for the human-object interaction classification sub-task in HOI detection. We study the problem of interaction category modeling, in contrast to most previous methods focusing on human-object feature learning. We adopt the concept of category query in a previous method [25] for HOI, for the first time, and show that it is simple, general and highly effective. + +Clearly, this idea of category query modeling is not limited to multi-label image classification and HOI detection. We hope it is useful for other vision tasks. + +Acknowledgment. This work was supported in part by the National Natural Science Foundation of China under Grant 62076183, 61936014 and 61976159, in part by the Natural Science Foundation of Shanghai under Grant 20ZR1473500, in part by the Shanghai Science and Technology Innovation Action Project of under Grant 20511100700 and 22511105300, in part by the Shanghai Municipal Science and Technology Major Project under Grant 2021SHZDZX0100, and in part by the Fundamental Research Funds for the Central Universities. The authors would also like to thank the anonymous reviewers for their careful work and valuable suggestions. + +# References + +[1] A Bansal, S. S Rambhatla, A Shrivastava, and R Chellappa. Detecting human-object interactions via functional generalization. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 34, pages 10460-10469, 2020. 2 +[2] N Carion, F Massa, G Synnaeve, N Usunier, A Kirillov, and S Zagoruyko. End-to-end object detection with transformers. In European Conference on Computer Vision, pages 213-229. Springer, 2020. 2, 7 +[3] Y.-W Chao, Y Liu, X Liu, H Zeng, and J Deng. Learning to detect human-object interactions. In 2018 iee winter conference on applications of computer vision (wacv), pages 381-389. IEEE, 2018. 1, 5, 6 +[4] M Chen, Y Liao, S Liu, Z Chen, F Wang, and C Qian. Reformulating hoi detection as adaptive set prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9004–9013, 2021. 2 +[5] B Cheng, I Misra, A. G Schwing, A Kirillov, and R Girdhar. Masked-attention mask transformer for universal image segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1290–1299, 2022. 3 +[6] L Dong, Z Li, K Xu, Z Zhang, L Yan, S Zhong, and X Zou. Category-aware transformer network for better human-object interaction detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19538-19547, 2022. 1, 2 +[7] C Gao, J Xu, Y Zou, and J.-B Huang. Drg: Dual relation graph for human-object interaction detection. In European Conference on Computer Vision, pages 696-712. Springer, 2020. 1, 2, 4, 5, 6 +[8] C Gao, Y Zou, and J.-B Huang. ican: Instance-centric attention network for human-object interaction detection. arXiv preprint arXiv:1808.10437, 2018. 1, 2 +[9] G Gkioxari, R Girshick, P Dólár, and K He. Detecting and recognizing human-object interactions. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 8359-8367, 2018. 1 +[10] S Gupta and J Malik. Visual semantic role labeling. arXiv preprint arXiv:1505.04474, 2015. 1, 5, 6 +[11] T Gupta, A Schwing, and D Hoiem. No-frills human-object interaction detection: Factorization, layout encodings, and training techniques. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9677-9685, 2019. 2, 4 +[12] H He, Y Yuan, X Yue, and H Hu. Rankseg: Adaptive pixel classification with image category ranking for segmentation. In Proceedings of the European Conference on Computer Vision (ECCV), pages 682-700. Springer Nature Switzerland Cham, 2022. 4 +[13] A Iftekhar, H Chen, K Kundu, X Li, J Tighe, and D Modolo. What to look at and where: Semantic and spatial refined transformer for detecting human-object interactions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5353-5363, 2022. 2, 4 +[14] B Kim, T Choi, J Kang, and H. J Kim. Uniondet: Union-level detector towards real-time human-object interaction de + +tlection. In European Conference on Computer Vision, pages 498-514. Springer, 2020. 4 +[15] B Kim, J Lee, J Kang, E.-S Kim, and H. J Kim. Hotr: End-to-end human-object interaction detection with transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 74-83, 2021. 2 +[16] B Kim, J Mun, K.-W On, M Shin, J Lee, and E.-S Kim. Mstr: Multi-scale transformer for end-to-end human-object interaction detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19578-19587, 2022. 1, 2, 5, 6 +[17] D.-J Kim, X Sun, J Choi, S Lin, and I. S Kweon. Detecting human-object interactions with action co-occurrence priors. In European Conference on Computer Vision, pages 718–736. Springer, 2020. 2, 4 +[18] Y.-L Li, X Liu, X Wu, Y Li, and C Lu. Hoi analysis: Integrating and decomposing human-object interaction. Advances in Neural Information Processing Systems, 33:5011-5022, 2020. 6 +[19] Y.-L Li, L Xu, X Liu, X Huang, Y Xu, S Wang, H.-S Fang, Z Ma, M Chen, and C Lu. Pastanet: Toward human activity knowledge engine. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 382-391, 2020. 2 +[20] Y.-L Li, S Zhou, X Huang, L Xu, Z Ma, H.-S Fang, Y Wang, and C Lu. Transferable interactiveness knowledge for human-object interaction detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3585-3594, 2019. 2 +[21] Y Liao, S Liu, F Wang, Y Chen, C Qian, and J Feng. Ppdm: Parallel point detection and matching for real-time human-object interaction detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 482-490, 2020. 1, 4 +[22] Y Liao, A Zhang, M Lu, Y Wang, X Li, and S Liu. Gen-vlkt: Simplify association and enhance interaction understanding for hoi detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20123-20132, 2022. 1, 2, 5, 6 +[23] T.-Y Lin, P Goyal, R Girshick, K He, and P Dollar. Focal loss for dense object detection. In Proceedings of the IEEE international conference on computer vision, pages 2980-2988, 2017. 3, 7 +[24] T.-Y Lin, M Maire, S Belongie, J Hays, P Perona, D Ramanan, P Dollár, and C. L Zitnick. Microsoft coco: Common objects in context. In European conference on computer vision, pages 740-755. Springer, 2014. 5 +[25] S Liu, L Zhang, X Yang, H Su, and J Zhu. Query2label: A simple transformer way to multi-label classification. arXiv preprint arXiv:2107.10834, 2021. 1, 2, 3, 4, 6, 7, 8 +[26] X Liu, Y.-L Li, X Wu, Y.-W Tai, C Lu, and C.-K Tang. Interactiveness field in human-object interactions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20113-20122, 2022. 6 +[27] Y Liu, Q Chen, and A Zisserman. Amplifying key cues for human-object-interaction detection. In European Conference on Computer Vision, pages 248-265. Springer, 2020. 2 + +[28] J Park, S Lee, H Heo, H. K Choi, and H. J Kim. Consistency learning via decoding path augmentation for transformers in human object interaction detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1019-1028, 2022. 4 +[29] S Qi, W Wang, B Jia, J Shen, and S.-C Zhu. Learning human-object interactions by graph parsing neural networks. In Proceedings of the European Conference on Computer Vision (ECCV), pages 401-417, 2018. 1, 2 +[30] X Qu, C Ding, X Li, X Zhong, and D Tao. Distillation using oracle queries for transformer-based human-object interaction detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19558-19567, 2022. 2 +[31] A Radford, J. W Kim, C Hallacy, A Ramesh, G Goh, S Agarwal, G Sastry, A Askell, P Mishkin, J Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 2 +[32] T Ridnik, E Ben-Baruch, N Zamir, A Noy, I Friedman, M Protter, and L Zelnik-Manor. Asymmetric loss for multi-label classification. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 82-91, 2021. 3, 7 +[33] M Tamura, H Ohashi, and T Yoshinaga. Qpic: Query-based pairwise human-object interaction detection with image-wide contextual information. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10410-10419, 2021. 1, 2, 4, 5, 6 +[34] O Ulutan, A Iftekhar, and B. S Manjunath. Vsgnet: Spatial attention network for detecting human object interactions using graph convolutions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13617-13626, 2020. 1, 2 +[35] A Vaswani, N Shazeer, N Parmar, J Uszkoreit, L Jones, A. N Gomez, L Kaiser, and I Polosukhin. Attention is all you need. In Advances in neural information processing systems, pages 5998-6008, 2017. 2, 7 +[36] B Wan, D Zhou, Y Liu, R Li, and X He. Pose-aware multilevel feature network for human object interaction detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9469-9478, 2019. 2 +[37] H Wang, W.-s Zheng, and L Yingbiao. Contextual heterogeneous graph network for human-object interaction detection. In European Conference on Computer Vision, pages 248-264. Springer, 2020. 1, 2 +[38] T Wang, T Yang, M Danelljan, F. S Khan, X Zhang, and J Sun. Learning human-object interaction detection using interaction points. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4116-4125, 2020. 4 +[39] X Wu, Y.-L Li, X Liu, J Zhang, Y Wu, and C Lu. Mining cross-person cues for body-part interactiveness learning in hoi detection. In Proceedings of the European Conference on Computer Vision (ECCV), pages 121-136. Springer, 2022. 2, 6 +[40] H Yuan, J Jiang, S Albanie, T Feng, Z Huang, D Ni, and M Tang. Rlip: Relational language-image pre-training for + +human-object interaction detection. In Advances in Neural Information Processing Systems, 2022. 2, 6 +[41] A Zhang, Y Liao, S Liu, M Lu, Y Wang, C Gao, and X Li. Mining the benefits of two-stage and one-stage hoi detection. In Thirty-Fifth Conference on Neural Information Processing Systems, 2021. 1, 2, 4, 5, 6 +[42] F. Z Zhang, D Campbell, and S Gould. Spatially conditioned graphs for detecting human-object interactions. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13319-13327, 2021. 2, 4, 5, 6 +[43] F. Z Zhang, D Campbell, and S Gould. Efficient two-stage detection of human-object interactions with a novel unary-pairwise transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20104-20112, 2022. 5 +[44] Y Zhang, Y Pan, T Yao, R Huang, T Mei, and C.-W Chen. Exploring structure-aware transformer over interaction proposals for human-object interaction detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19548-19557, 2022. 6 +[45] X Zhong, C Ding, Z Li, and S Huang. Towards hard-positive query mining for detr-based human-object interaction detection. In Proceedings of the European Conference on Computer Vision (ECCV), pages 444-460. Springer, 2022. 2, 4, 6 +[46] X Zhong, C Ding, X Qu, and D Tao. Polysemy deciphering network for robust human-object interaction detection. International Journal of Computer Vision, 129(6):1910-1929, 2021. 1, 2 +[47] X Zhong, X Qu, C Ding, and D Tao. Glance and gaze: Inferring action-aware points for one-stage human-object interaction detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13234-13243, 2021. 6 +[48] D Zhou, Z Liu, J Wang, L Wang, T Hu, E Ding, and J Wang. Human-object interaction detection via disentangled transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19568-19577, 2022. 1, 2, 4, 6 +[49] C Zou, B Wang, Y Hu, J Liu, Q Wu, Y Zhao, B Li, C Zhang, C Zhang, Y Wei, et al. End-to-end human object interaction detection with hoi transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11825-11834, 2021. 2 \ No newline at end of file diff --git a/2023/Category Query Learning for Human-Object Interaction Classification/images.zip b/2023/Category Query Learning for Human-Object Interaction Classification/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..ab60a6ef869468682de2c07fbfad4a9d25d97d29 --- /dev/null +++ b/2023/Category Query Learning for Human-Object Interaction Classification/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f43888a0beec998b66d272d8b8d4a6966bda976f75037654318adc41e0ebf599 +size 660973 diff --git a/2023/Category Query Learning for Human-Object Interaction Classification/layout.json b/2023/Category Query Learning for Human-Object Interaction Classification/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..96cbe0ff72cce7e4ce935bbcda7b3a5d03e0de49 --- /dev/null +++ b/2023/Category Query Learning for Human-Object Interaction Classification/layout.json @@ -0,0 +1,9203 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 78, + 103, + 515, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 78, + 103, + 515, + 121 + ], + "spans": [ + { + "bbox": [ + 78, + 103, + 515, + 121 + ], + "type": "text", + "content": "Category Query Learning for Human-Object Interaction Classification" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 106, + 142, + 487, + 181 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 106, + 142, + 487, + 181 + ], + "spans": [ + { + "bbox": [ + 106, + 142, + 487, + 181 + ], + "type": "text", + "content": "Chi Xie" + }, + { + "bbox": [ + 106, + 142, + 487, + 181 + ], + "type": "inline_equation", + "content": "^{1\\dagger}" + }, + { + "bbox": [ + 106, + 142, + 487, + 181 + ], + "type": "text", + "content": " Fangao Zeng" + }, + { + "bbox": [ + 106, + 142, + 487, + 181 + ], + "type": "inline_equation", + "content": "^{2‡}" + }, + { + "bbox": [ + 106, + 142, + 487, + 181 + ], + "type": "text", + "content": " Yue Hu" + }, + { + "bbox": [ + 106, + 142, + 487, + 181 + ], + "type": "inline_equation", + "content": "^{3‡}" + }, + { + "bbox": [ + 106, + 142, + 487, + 181 + ], + "type": "text", + "content": " Shuang Liang" + }, + { + "bbox": [ + 106, + 142, + 487, + 181 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 106, + 142, + 487, + 181 + ], + "type": "text", + "content": " Yichen Wei" + }, + { + "bbox": [ + 106, + 142, + 487, + 181 + ], + "type": "inline_equation", + "content": "^{2‡}" + }, + { + "bbox": [ + 106, + 142, + 487, + 181 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 106, + 142, + 487, + 181 + ], + "type": "text", + "content": "Tongji University " + }, + { + "bbox": [ + 106, + 142, + 487, + 181 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 106, + 142, + 487, + 181 + ], + "type": "text", + "content": "MEGVII Technology " + }, + { + "bbox": [ + 106, + 142, + 487, + 181 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 106, + 142, + 487, + 181 + ], + "type": "text", + "content": "Shanghai Jiao Tong University " + }, + { + "bbox": [ + 106, + 142, + 487, + 181 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 106, + 142, + 487, + 181 + ], + "type": "text", + "content": " {chixie, shuangliang}@tongji.edu.cn" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 107, + 185, + 496, + 198 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 107, + 185, + 496, + 198 + ], + "spans": [ + { + "bbox": [ + 107, + 185, + 496, + 198 + ], + "type": "text", + "content": "2zfg472988436@163.com,wei_yi_chen@hotmail.com 318671129361@sjtu.edu.cn" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 143, + 227, + 191, + 239 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 227, + 191, + 239 + ], + "spans": [ + { + "bbox": [ + 143, + 227, + 191, + 239 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 253, + 290, + 420 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 253, + 290, + 420 + ], + "spans": [ + { + "bbox": [ + 46, + 253, + 290, + 420 + ], + "type": "text", + "content": "Unlike most previous HOI methods that focus on learning better human-object features, we propose a novel and complementary approach called category query learning. Such queries are explicitly associated to interaction categories, converted to image specific category representation via a transformer decoder, and learnt via an auxiliary image-level classification task. This idea is motivated by an earlier multi-label image classification method, but is for the first time applied for the challenging human-object interaction classification task. Our method is simple, general and effective. It is validated on three representative HOI baselines and achieves new state-of-the-art results on two benchmarks. Code will be available at https://github.com/charles-xie/CQL." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 446, + 128, + 459 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 446, + 128, + 459 + ], + "spans": [ + { + "bbox": [ + 47, + 446, + 128, + 459 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 468, + 287, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 468, + 287, + 574 + ], + "spans": [ + { + "bbox": [ + 46, + 468, + 287, + 574 + ], + "type": "text", + "content": "Human-Object Interaction (HOI) detection has attracted a lot of interests in recent years [3, 8-10, 21, 33]. The task consists of two sub-tasks. The first is human and object detection. It is usually performed by common object detection methods. The second is interaction classification of each human-object (HO) pair. This sub-task is very challenging due to the complex appearance variations in the interaction categories. See Fig. 1 for examples. It is the focus of most previous HOI methods, as well as this work." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 576, + 287, + 673 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 576, + 287, + 673 + ], + "spans": [ + { + "bbox": [ + 46, + 576, + 287, + 673 + ], + "type": "text", + "content": "Most previous HOI methods focus on learning better human-object features, including modeling relation and context via GNN [7, 29, 34, 37] or attention mechanism [8, 34, 46], decoupling localization and classification [22, 41, 48], leveraging vision-language knowledge [6, 22] and introducing multi-scale feature to transformer [16]. However, for interaction classification they all adopt the simple linear classifier that performs the dot product of the" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 307, + 225, + 425, + 305 + ], + "blocks": [ + { + "bbox": [ + 307, + 225, + 425, + 305 + ], + "lines": [ + { + "bbox": [ + 307, + 225, + 425, + 305 + ], + "spans": [ + { + "bbox": [ + 307, + 225, + 425, + 305 + ], + "type": "image", + "image_path": "3075a4e4a0d7b0b2428e126ff8afd3b3dc0a7619be00fad2022df5f19c9f8ac1.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 339, + 306, + 394, + 314 + ], + "lines": [ + { + "bbox": [ + 339, + 306, + 394, + 314 + ], + "spans": [ + { + "bbox": [ + 339, + 306, + 394, + 314 + ], + "type": "text", + "content": "(a) person fly kite" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 427, + 226, + 545, + 305 + ], + "blocks": [ + { + "bbox": [ + 427, + 226, + 545, + 305 + ], + "lines": [ + { + "bbox": [ + 427, + 226, + 545, + 305 + ], + "spans": [ + { + "bbox": [ + 427, + 226, + 545, + 305 + ], + "type": "image", + "image_path": "34ce8e97a2ba6758f9ba26c1eaf19d64dcf450e62bd5da932f9f922319a890f2.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 457, + 306, + 517, + 313 + ], + "lines": [ + { + "bbox": [ + 457, + 306, + 517, + 313 + ], + "spans": [ + { + "bbox": [ + 457, + 306, + 517, + 313 + ], + "type": "text", + "content": "person fly airplane" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 307, + 315, + 425, + 394 + ], + "blocks": [ + { + "bbox": [ + 307, + 315, + 425, + 394 + ], + "lines": [ + { + "bbox": [ + 307, + 315, + 425, + 394 + ], + "spans": [ + { + "bbox": [ + 307, + 315, + 425, + 394 + ], + "type": "image", + "image_path": "02d5d4e4abf81c7fe8dd3fb7833919f4d4f4db83d6e10ced0f40417a6658b050.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 333, + 395, + 397, + 403 + ], + "lines": [ + { + "bbox": [ + 333, + 395, + 397, + 403 + ], + "spans": [ + { + "bbox": [ + 333, + 395, + 397, + 403 + ], + "type": "text", + "content": "(b) person hold fork" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 427, + 315, + 545, + 394 + ], + "blocks": [ + { + "bbox": [ + 427, + 315, + 545, + 394 + ], + "lines": [ + { + "bbox": [ + 427, + 315, + 545, + 394 + ], + "spans": [ + { + "bbox": [ + 427, + 315, + 545, + 394 + ], + "type": "image", + "image_path": "afe9e0d7b15e9a43e4021404cf5e3e203cf80b73db034c463fefc552f98f45e0.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 452, + 396, + 521, + 403 + ], + "lines": [ + { + "bbox": [ + 452, + 396, + 521, + 403 + ], + "spans": [ + { + "bbox": [ + 452, + 396, + 521, + 403 + ], + "type": "text", + "content": "person hold elephant" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 305, + 417, + 547, + 472 + ], + "lines": [ + { + "bbox": [ + 305, + 417, + 547, + 472 + ], + "spans": [ + { + "bbox": [ + 305, + 417, + 547, + 472 + ], + "type": "text", + "content": "Figure 1. Interaction classification is inherently challenging. In (a), \"fly\" is semantically polysemic, resulting in different objects, poses and relative positions. In (b), when \"hold\" is associated with different objects, the appearance, scene background, and human poses are largely different." + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "bbox": [ + 305, + 486, + 545, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 486, + 545, + 510 + ], + "spans": [ + { + "bbox": [ + 305, + 486, + 545, + 510 + ], + "type": "text", + "content": "human-object feature and a static weight vector, which represents an interaction category." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 510, + 546, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 510, + 546, + 653 + ], + "spans": [ + { + "bbox": [ + 304, + 510, + 546, + 653 + ], + "type": "text", + "content": "In this work, we propose a new approach that enhances the above paradigm and complements most previous HOI methods. It is motivated by the recent work Query2label [25], a transformer-based classification network. It proposes a new concept we call category-specific query. Unlike the queries in other transformer methods, each query is associated to a specific and fixed image category during training and inference. This one-to-one binding makes the query learn to model each category more effectively. The queries are converted to image specific category representations via a transformer decoder. This method achieves excellent performance on multi-label image classification task." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 654, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 547, + 713 + ], + "type": "text", + "content": "We extend this approach for human-object interaction classification. Essentially, our approach replaces traditional category representation as a static weight vector in previous HOI methods with category queries learnt as described above. The same linear classifier is adopted. Such cate" + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 57, + 683, + 136, + 693 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 683, + 136, + 693 + ], + "spans": [ + { + "bbox": [ + 57, + 683, + 136, + 693 + ], + "type": "text", + "content": "*Corresponding author." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 59, + 693, + 233, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 693, + 233, + 703 + ], + "spans": [ + { + "bbox": [ + 59, + 693, + 233, + 703 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 59, + 693, + 233, + 703 + ], + "type": "text", + "content": " Work done during internship at MEGVII technology." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 59, + 703, + 186, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 703, + 186, + 712 + ], + "spans": [ + { + "bbox": [ + 59, + 703, + 186, + 712 + ], + "type": "inline_equation", + "content": "\\ddagger" + }, + { + "bbox": [ + 59, + 703, + 186, + 712 + ], + "type": "text", + "content": "Work done while worked at MEGVII." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15275" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 286, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 286, + 144 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 286, + 144 + ], + "type": "text", + "content": "gory queries are more effective, and adaptive for different images, giving rise to better modeling of the complex variations in each interaction category. This is the crucial difference between this work and a simple adaption of [25] to HOI. Notably, this work is the first to address the category weight representation problem in the HOI community." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 144, + 286, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 144, + 286, + 217 + ], + "spans": [ + { + "bbox": [ + 46, + 144, + 286, + 217 + ], + "type": "text", + "content": "Note that our proposed category specific query is different and not related to those queries in other transformer-based HOI methods [4, 15, 33, 49]. Specifically, category queries extract image-level features as the category representation. The queries in other methods are human-object instance-level features and category-agnostic." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 217, + 286, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 217, + 286, + 277 + ], + "spans": [ + { + "bbox": [ + 46, + 217, + 286, + 277 + ], + "type": "text", + "content": "Our method is simple, lightweight and general. The overview is in Fig. 2. It is complementary to any off-the-shelf HOI method that provides human-object features. The modification of both inference and training is small. The incurred additional cost is marginal." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 278, + 286, + 421 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 278, + 286, + 421 + ], + "spans": [ + { + "bbox": [ + 46, + 278, + 286, + 421 + ], + "type": "text", + "content": "In experiments, our approach is validated on three representative and strong HOI baseline methods, two transformer-based methods [22, 33] and a traditional two-stage method [42]. They are all significantly improved by our approach. New state-of-the-art results are obtained on two benchmarks. Specifically, we obtain " + }, + { + "bbox": [ + 46, + 278, + 286, + 421 + ], + "type": "inline_equation", + "content": "36.03\\mathrm{mAP}" + }, + { + "bbox": [ + 46, + 278, + 286, + 421 + ], + "type": "text", + "content": " on HICO-DET. Comprehensive ablation studies and in-depth discussion are also provided to verify the effectiveness of implementation details in our approach. It turns out that our method is more effective on challenging images that contain more human-object instances, a property that is rarely discussed by previous HOI methods." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 434, + 134, + 447 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 434, + 134, + 447 + ], + "spans": [ + { + "bbox": [ + 47, + 434, + 134, + 447 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 455, + 270, + 467 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 455, + 270, + 467 + ], + "spans": [ + { + "bbox": [ + 47, + 455, + 270, + 467 + ], + "type": "text", + "content": "2.1. Instance Query Learning in HOI Detection" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 474, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 474, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 474, + 286, + 713 + ], + "type": "text", + "content": "DETR [2] firstly proposes the concept of object instance query for object detection task. Such queries essentially learn the priors of both object appearance and spatial location. DETR leverages those queries to probe image features through a transformer [35] and localize unique objects in the image. Motivated by its great success, many works [15, 22, 33, 41, 48, 49] adapt such detection transformer framework to HOI detection by simply treating the HOI triplet [33, 49] or H-O pair [4, 15] as an object. A few of them [6, 30, 45] pay attention to adapting the plain query to this task. DOQ [30] proposes a knowledge distillation model using oracle queries to facilitate the representation learning of a transformer-based detector; HQM [45] explicitly constructs hard positive queries from ground truth to train the model to be less vulnerable to spacial variations; CATN [6] utilizes the object category prior generated from external object detector and language model for query initialization. In summary, these transformer-based methods use each query to aggregate context information not restricted to one interaction category, in order to predict" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 72, + 492, + 83 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 492, + 83 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 492, + 83 + ], + "type": "text", + "content": "a potential HOI instance at a specific location." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 84, + 545, + 193 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 84, + 545, + 193 + ], + "spans": [ + { + "bbox": [ + 304, + 84, + 545, + 193 + ], + "type": "text", + "content": "In DETR [2] and its variants [4,15,22,33,41,49] in HOI, as the queries are category-agnostic, their association to object categories are dynamic and unstable during training. This could be problematic. For example, it is well known that the convergence of DETR training is slow. In contrast, our proposed query is category-specific. The learning is guided by image-level classification task and stable. Such queries learn category-specific priors and are good representation for interaction categories." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 202, + 493, + 215 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 202, + 493, + 215 + ], + "spans": [ + { + "bbox": [ + 306, + 202, + 493, + 215 + ], + "type": "text", + "content": "2.2. Feature Learning in HOI Detection" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 221, + 545, + 305 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 221, + 545, + 305 + ], + "spans": [ + { + "bbox": [ + 304, + 221, + 545, + 305 + ], + "type": "text", + "content": "Early methods. Based on two-stage detection framework, early works make many efforts to help feature learning, including employing architectures effective in modeling relation and context like GNN [7, 29, 34, 37] and attention module [8, 34, 46], leveraging fine-grained visual features [11, 17, 19, 20, 36] like human pose and introducing language prior [1, 7, 17, 27, 46]." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 306, + 545, + 378 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 306, + 545, + 378 + ], + "spans": [ + { + "bbox": [ + 304, + 306, + 545, + 378 + ], + "type": "text", + "content": "Transformer-based methods. Motivated by DETR [49], many methods [4, 15, 33, 49] leverage transformer architecture [35] and extend the object query in DETR to HOI query. With the help of HOI query and transformer's built-in attention mechanism, those methods learn effective feature representation for HOI triplet or H-O pair." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 379, + 545, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 379, + 545, + 486 + ], + "spans": [ + { + "bbox": [ + 304, + 379, + 545, + 486 + ], + "type": "text", + "content": "Based on those pioneer transformer-based methods, recently, many methods are proposed to further help feature learning, by decoupling H-O pair localization and interaction classification [22, 41, 48] or exploiting multi-scale feature in transformer architecture [16]. Some works [6, 22] leverage vision-language knowledge in CLIP [31] or design a pretrained model [40] specifically for HOI; others utilize information like human poses [39] or spatial configurations [13] that has been used in early HOI detectors." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 487, + 545, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 487, + 545, + 559 + ], + "spans": [ + { + "bbox": [ + 304, + 487, + 545, + 559 + ], + "type": "text", + "content": "Relation to the proposed method. Previous methods learn the H-O feature while ours learns the category query as the category representation feature. Thus, they are complementary. The interaction classification is simply by measuring the similarity between the two types of features. The integration of our method to previous HOI methods is simple." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 572, + 385, + 584 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 572, + 385, + 584 + ], + "spans": [ + { + "bbox": [ + 306, + 572, + 385, + 584 + ], + "type": "text", + "content": "3. Our Method" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 593, + 545, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 593, + 545, + 640 + ], + "spans": [ + { + "bbox": [ + 304, + 593, + 545, + 640 + ], + "type": "text", + "content": "The overview of our method is in Fig. 2. It consists of two components, the image-level category query learning (top block) and human-object interaction classification (bottom right block)." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 641, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 641, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 641, + 545, + 713 + ], + "type": "text", + "content": "The first component is detailed in Sec. 3.1. It is briefly summarized here. A number of queries (embedding vectors) are associated to human-object interaction category, in a one-to-one manner. Such queries interact with image features (provided by a baseline HOI method) through a transformer decoder [35] and become image-specific queries." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15276" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 75, + 71, + 520, + 262 + ], + "blocks": [ + { + "bbox": [ + 75, + 71, + 520, + 262 + ], + "lines": [ + { + "bbox": [ + 75, + 71, + 520, + 262 + ], + "spans": [ + { + "bbox": [ + 75, + 71, + 520, + 262 + ], + "type": "image", + "image_path": "a2e1ee732a1a14a4cdac05c06211e634cc8a176b620996b5cf18b6d23f76a740.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 271, + 547, + 296 + ], + "lines": [ + { + "bbox": [ + 46, + 271, + 547, + 296 + ], + "spans": [ + { + "bbox": [ + 46, + 271, + 547, + 296 + ], + "type": "text", + "content": "Figure 2. Overview of our method. It consists of two components (top and bottom right). It can be integrated with any baseline HOI method (bottom left) that provides image feature " + }, + { + "bbox": [ + 46, + 271, + 547, + 296 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 46, + 271, + 547, + 296 + ], + "type": "text", + "content": " and human-object instance features " + }, + { + "bbox": [ + 46, + 271, + 547, + 296 + ], + "type": "inline_equation", + "content": "F_{i}" + }, + { + "bbox": [ + 46, + 271, + 547, + 296 + ], + "type": "text", + "content": ". See Sec. 3 for details." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 308, + 287, + 390 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 308, + 287, + 390 + ], + "spans": [ + { + "bbox": [ + 46, + 308, + 287, + 390 + ], + "type": "text", + "content": "Learning of both the queries and decoder weights is supervised by an auxiliary image-level classification task. In this way, the queries are learnt to capture category-specific feature and become good feature representation for these categories. Besides some minor details, this step is the same as the previous work Query2label [25], which is for multi-label image classification task." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 391, + 288, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 391, + 288, + 498 + ], + "spans": [ + { + "bbox": [ + 46, + 391, + 288, + 498 + ], + "type": "text", + "content": "The second component is detailed in Sec. 3.2. For the first time, we adopt the category query learning method for human-object interaction classification tasks. The cosine similarity between the category query and human-object feature is used for interaction classification. Thus, it works with any HOI method that provides human-object features. Besides, the image-level classification results turn out moderately helpful in an score integration step, which is an extra technique that benefits the performance." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 498, + 287, + 546 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 498, + 287, + 546 + ], + "spans": [ + { + "bbox": [ + 46, + 498, + 287, + 546 + ], + "type": "text", + "content": "Overall, the proposed method is simple, effective, lightweight and general. It can be combined with most previous HOI detection methods (bottom left block in Fig. 2), with small modification, as elaborated in Sec. 4." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 555, + 249, + 568 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 555, + 249, + 568 + ], + "spans": [ + { + "bbox": [ + 47, + 555, + 249, + 568 + ], + "type": "text", + "content": "3.1. Image-level Category Query Learning" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 573, + 287, + 620 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 573, + 287, + 620 + ], + "spans": [ + { + "bbox": [ + 46, + 573, + 287, + 620 + ], + "type": "text", + "content": "Similar as in Query2Label [25], for " + }, + { + "bbox": [ + 46, + 573, + 287, + 620 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 46, + 573, + 287, + 620 + ], + "type": "text", + "content": " human-object interaction categories, we define their one-to-one corresponding category queries, which are learnable embedding vectors, " + }, + { + "bbox": [ + 46, + 573, + 287, + 620 + ], + "type": "inline_equation", + "content": "\\{Q_k\\} \\in \\mathbb{R}^{K\\times D}" + }, + { + "bbox": [ + 46, + 573, + 287, + 620 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 573, + 287, + 620 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 46, + 573, + 287, + 620 + ], + "type": "text", + "content": " is the vector dimension." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 620, + 287, + 658 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 620, + 287, + 658 + ], + "spans": [ + { + "bbox": [ + 46, + 620, + 287, + 658 + ], + "type": "text", + "content": "Each query " + }, + { + "bbox": [ + 46, + 620, + 287, + 658 + ], + "type": "inline_equation", + "content": "Q_{k}" + }, + { + "bbox": [ + 46, + 620, + 287, + 658 + ], + "type": "text", + "content": " aggregates image features " + }, + { + "bbox": [ + 46, + 620, + 287, + 658 + ], + "type": "inline_equation", + "content": "\\mathbf{I} \\in \\mathbb{R}^{H \\times W \\times D}" + }, + { + "bbox": [ + 46, + 620, + 287, + 658 + ], + "type": "text", + "content": " through a transformer decoder and is updated to image specific query " + }, + { + "bbox": [ + 46, + 620, + 287, + 658 + ], + "type": "inline_equation", + "content": "Q_{k}'" + }, + { + "bbox": [ + 46, + 620, + 287, + 658 + ], + "type": "text", + "content": "," + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 109, + 666, + 287, + 681 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 109, + 666, + 287, + 681 + ], + "spans": [ + { + "bbox": [ + 109, + 666, + 287, + 681 + ], + "type": "interline_equation", + "content": "\\left\\{Q _ {k} ^ {\\prime} \\right\\} = \\operatorname {d e c o d e r} \\left(\\left\\{Q _ {k} \\right\\}, I\\right). \\tag {1}", + "image_path": "87ed69ef24e2184080d608cc764f6fccfdc4db4d79c00a705e28edce8fa28f73.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 689, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 288, + 715 + ], + "type": "text", + "content": "Note that the decoder structure has several variants, which are studied in [5]. Our experiments show that the" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 308, + 546, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 308, + 546, + 354 + ], + "spans": [ + { + "bbox": [ + 304, + 308, + 546, + 354 + ], + "type": "text", + "content": "structure is of minor importance, as discussed in Sec. 5.4. Specifically, our decoder consists of two layers, each of which consisting of a cross-attention layer, a self-attention layer and a FFN layer, in order." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 355, + 547, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 355, + 547, + 392 + ], + "spans": [ + { + "bbox": [ + 304, + 355, + 547, + 392 + ], + "type": "text", + "content": "Then, each image-level classification probabilities " + }, + { + "bbox": [ + 304, + 355, + 547, + 392 + ], + "type": "inline_equation", + "content": "p_k" + }, + { + "bbox": [ + 304, + 355, + 547, + 392 + ], + "type": "text", + "content": " is computed by applying a category-specific fully-connected layer and a sigmoid activation on the updated query " + }, + { + "bbox": [ + 304, + 355, + 547, + 392 + ], + "type": "inline_equation", + "content": "Q_{k}^{\\prime}" + }, + { + "bbox": [ + 304, + 355, + 547, + 392 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 374, + 404, + 545, + 418 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 374, + 404, + 545, + 418 + ], + "spans": [ + { + "bbox": [ + 374, + 404, + 545, + 418 + ], + "type": "interline_equation", + "content": "p _ {k} = \\operatorname {s i g m o i d} \\left(\\mathrm {F C} \\left(Q _ {k} ^ {\\prime}\\right)\\right) \\tag {2}", + "image_path": "115bde69856e7a3b70a05e28b3e62abdfec8c51789bf780dc1379dd69b245e93.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 424, + 546, + 508 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 424, + 546, + 508 + ], + "spans": [ + { + "bbox": [ + 304, + 424, + 546, + 508 + ], + "type": "text", + "content": "Learning of the category query " + }, + { + "bbox": [ + 304, + 424, + 546, + 508 + ], + "type": "inline_equation", + "content": "\\{Q_k\\}" + }, + { + "bbox": [ + 304, + 424, + 546, + 508 + ], + "type": "text", + "content": " and the decoder weights is supervised by common image classification losses. To deal with the label imbalance problem, focal loss [23] and asymmetric loss (ASL) [32] are used. Asymmetric loss is a variant of focal loss. It is more robust for high label imbalance and noises. Our experiments (see Sec. 5.4) show that it is slightly better." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 508, + 547, + 543 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 508, + 547, + 543 + ], + "spans": [ + { + "bbox": [ + 304, + 508, + 547, + 543 + ], + "type": "text", + "content": "Specifically, with the classification probability " + }, + { + "bbox": [ + 304, + 508, + 547, + 543 + ], + "type": "inline_equation", + "content": "p_k" + }, + { + "bbox": [ + 304, + 508, + 547, + 543 + ], + "type": "text", + "content": " and the shifted probability " + }, + { + "bbox": [ + 304, + 508, + 547, + 543 + ], + "type": "inline_equation", + "content": "p_k' = \\max(p_k - m, 0)" + }, + { + "bbox": [ + 304, + 508, + 547, + 543 + ], + "type": "text", + "content": ", the asymmetric loss is" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 552, + 545, + 585 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 552, + 545, + 585 + ], + "spans": [ + { + "bbox": [ + 313, + 552, + 545, + 585 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {i m g} = \\frac {1}{K} \\sum_ {k = 1} ^ {K} \\left\\{ \\begin{array}{l l} (1 - p _ {k}) ^ {\\gamma +} \\log \\left(p _ {k}\\right), & y _ {k} = 1, \\\\ \\left(p _ {k} ^ {\\prime}\\right) ^ {\\gamma -} \\log \\left(1 - p _ {k} ^ {\\prime}\\right), & y _ {k} = 0, \\end{array} \\right. \\tag {3}", + "image_path": "d2406313cdb456899142e940ca6b46bedb48a3c1413d406e6187516ea0516cfb.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 594, + 545, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 545, + 640 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 545, + 640 + ], + "type": "text", + "content": "where the binary label " + }, + { + "bbox": [ + 304, + 594, + 545, + 640 + ], + "type": "inline_equation", + "content": "y_{k}" + }, + { + "bbox": [ + 304, + 594, + 545, + 640 + ], + "type": "text", + "content": " indicates the existence of category " + }, + { + "bbox": [ + 304, + 594, + 545, + 640 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 594, + 545, + 640 + ], + "type": "text", + "content": " in the image, and " + }, + { + "bbox": [ + 304, + 594, + 545, + 640 + ], + "type": "inline_equation", + "content": "\\gamma +, \\gamma -" + }, + { + "bbox": [ + 304, + 594, + 545, + 640 + ], + "type": "text", + "content": " as well as " + }, + { + "bbox": [ + 304, + 594, + 545, + 640 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 304, + 594, + 545, + 640 + ], + "type": "text", + "content": " are hyperparameters. We use the default values in ASL [32], " + }, + { + "bbox": [ + 304, + 594, + 545, + 640 + ], + "type": "inline_equation", + "content": "\\gamma + = 0" + }, + { + "bbox": [ + 304, + 594, + 545, + 640 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 594, + 545, + 640 + ], + "type": "inline_equation", + "content": "\\gamma - = 4" + }, + { + "bbox": [ + 304, + 594, + 545, + 640 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 594, + 545, + 640 + ], + "type": "inline_equation", + "content": "m = 0.05" + }, + { + "bbox": [ + 304, + 594, + 545, + 640 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 642, + 546, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 642, + 546, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 642, + 546, + 714 + ], + "type": "text", + "content": "In this way, the category queries are learnt to encode the category priors. Figure 3 is the visualization of the heatmaps in the cross-attention layer of the decoder. Each category query learns to locate the human body parts related to discriminative feature of its corresponding interaction category, e.g., in Fig. 3c, the query of \"hold\" highlights" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15277" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "text", + "content": "the hand region, while in the same image the query of \"ride\" highlights the foot region. It qualitatively demonstrates that the query learning is effective in encoding category-specific information." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 122, + 288, + 159 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 122, + 288, + 159 + ], + "spans": [ + { + "bbox": [ + 47, + 122, + 288, + 159 + ], + "type": "text", + "content": "The updated queries adaptively extract category-related features for each image, with the help of the transformer decoder's built-in multi-head cross-attention layer." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 172, + 286, + 185 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 172, + 286, + 185 + ], + "spans": [ + { + "bbox": [ + 47, + 172, + 286, + 185 + ], + "type": "text", + "content": "3.2. Interaction Classification with Category Query" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 193, + 286, + 254 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 193, + 286, + 254 + ], + "spans": [ + { + "bbox": [ + 46, + 193, + 286, + 254 + ], + "type": "text", + "content": "In this step, we apply the updated category queries " + }, + { + "bbox": [ + 46, + 193, + 286, + 254 + ], + "type": "inline_equation", + "content": "\\{Q_k'\\}" + }, + { + "bbox": [ + 46, + 193, + 286, + 254 + ], + "type": "text", + "content": " as the weights for interaction classification. Given the " + }, + { + "bbox": [ + 46, + 193, + 286, + 254 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 193, + 286, + 254 + ], + "type": "text", + "content": "-th human-object instance, its classification probability score for " + }, + { + "bbox": [ + 46, + 193, + 286, + 254 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 193, + 286, + 254 + ], + "type": "text", + "content": "-th interaction category is simply the cosine similarity between its feature " + }, + { + "bbox": [ + 46, + 193, + 286, + 254 + ], + "type": "inline_equation", + "content": "F_i" + }, + { + "bbox": [ + 46, + 193, + 286, + 254 + ], + "type": "text", + "content": " and the category query feature " + }, + { + "bbox": [ + 46, + 193, + 286, + 254 + ], + "type": "inline_equation", + "content": "Q_k'" + }, + { + "bbox": [ + 46, + 193, + 286, + 254 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 270, + 287, + 297 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 105, + 270, + 287, + 297 + ], + "spans": [ + { + "bbox": [ + 105, + 270, + 287, + 297 + ], + "type": "interline_equation", + "content": "s _ {i, k} = \\operatorname {s i g m o i d} \\left(\\frac {Q _ {k} ^ {\\prime} \\cdot F _ {i}}{\\left| Q _ {k} ^ {\\prime} \\right| \\times \\left| F _ {i} \\right|}\\right). \\tag {4}", + "image_path": "fcc0ba0a73f010fc549c6c8422bda620c3df7b654b4a26c73c8adde2b7a1a3a7.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 310, + 287, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 310, + 287, + 346 + ], + "spans": [ + { + "bbox": [ + 46, + 310, + 287, + 346 + ], + "type": "text", + "content": "There are no restrictions or assumptions on the human-object feature " + }, + { + "bbox": [ + 46, + 310, + 287, + 346 + ], + "type": "inline_equation", + "content": "F_{i}" + }, + { + "bbox": [ + 46, + 310, + 287, + 346 + ], + "type": "text", + "content": ". Most previous HOI methods should be applicable here." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 348, + 287, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 348, + 287, + 396 + ], + "spans": [ + { + "bbox": [ + 46, + 348, + 287, + 396 + ], + "type": "text", + "content": "In this step, the traditional classification weight from static parameters is replaced with the category query adaptive on each image. This behavior is essentially different from [25], which uses category queries as image feature." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 398, + 287, + 470 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 398, + 287, + 470 + ], + "spans": [ + { + "bbox": [ + 46, + 398, + 287, + 470 + ], + "type": "text", + "content": "By \"adaptive\", we mean that the queries are updated dynamically according to the image contents. As exemplified in Fig. 3a and Fig. 3c, the queries of \"hold\" learn to highlight different interactive areas in different images and update themselves with features from these areas via attention. This shows the query learning is adaptive to image." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 472, + 287, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 472, + 287, + 533 + ], + "spans": [ + { + "bbox": [ + 46, + 472, + 287, + 533 + ], + "type": "text", + "content": "As discussed in Sec. 5.4, the classification step in Eq. (4) is crucial to make the category query learning effective on human-object interaction classification. Without this step, the image-level category query learning using only image-level classification from Query2label [25] is of little use." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 534, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 534, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 534, + 287, + 713 + ], + "type": "text", + "content": "Score integration step. The segmentation method in [12] discards certain categories during pixel classification that have low image classification scores. Motivated by this method, we take a similar score integration step. The image-level classification score " + }, + { + "bbox": [ + 46, + 534, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\{p_i\\}" + }, + { + "bbox": [ + 46, + 534, + 287, + 713 + ], + "type": "text", + "content": " is used to enhance the human-object instance classification. The idea is that, the instance score should be higher if the image-level probability is higher. Our implementation is similar as in [12]. During both training and testing, for each image, the top-" + }, + { + "bbox": [ + 46, + 534, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\kappa" + }, + { + "bbox": [ + 46, + 534, + 287, + 713 + ], + "type": "text", + "content": " categories (" + }, + { + "bbox": [ + 46, + 534, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\kappa = 70" + }, + { + "bbox": [ + 46, + 534, + 287, + 713 + ], + "type": "text", + "content": " in this work) with higher image classification scores " + }, + { + "bbox": [ + 46, + 534, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\{p_i\\}" + }, + { + "bbox": [ + 46, + 534, + 287, + 713 + ], + "type": "text", + "content": " are selected. The instance score " + }, + { + "bbox": [ + 46, + 534, + 287, + 713 + ], + "type": "inline_equation", + "content": "s_{i,k}" + }, + { + "bbox": [ + 46, + 534, + 287, + 713 + ], + "type": "text", + "content": " is slightly modulated such that it becomes higher if the rank of category " + }, + { + "bbox": [ + 46, + 534, + 287, + 713 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 534, + 287, + 713 + ], + "type": "text", + "content": " is higher. This strategy gives rise to moderate improvement, as verified in Tab. 4. We left the implementation details in the supplementary materials." + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 307, + 70, + 385, + 123 + ], + "blocks": [ + { + "bbox": [ + 307, + 70, + 385, + 123 + ], + "lines": [ + { + "bbox": [ + 307, + 70, + 385, + 123 + ], + "spans": [ + { + "bbox": [ + 307, + 70, + 385, + 123 + ], + "type": "image", + "image_path": "99f8282d3ac2f3524a9f012e38cb046233dabf2bf8ff19f79576baca3c9a90b0.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 387, + 70, + 465, + 123 + ], + "blocks": [ + { + "bbox": [ + 387, + 70, + 465, + 123 + ], + "lines": [ + { + "bbox": [ + 387, + 70, + 465, + 123 + ], + "spans": [ + { + "bbox": [ + 387, + 70, + 465, + 123 + ], + "type": "image", + "image_path": "fe379372a13e190926267d60ca26365d5a1c8c3fdec1a768296f4aba55d26d7c.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 466, + 70, + 545, + 123 + ], + "blocks": [ + { + "bbox": [ + 466, + 70, + 545, + 123 + ], + "lines": [ + { + "bbox": [ + 466, + 70, + 545, + 123 + ], + "spans": [ + { + "bbox": [ + 466, + 70, + 545, + 123 + ], + "type": "image", + "image_path": "2b0bc16ea2c372fd882e3a9223f14266858b79022b8fc3d45678b1c1a80631f0.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 307, + 138, + 385, + 191 + ], + "blocks": [ + { + "bbox": [ + 317, + 128, + 534, + 137 + ], + "lines": [ + { + "bbox": [ + 317, + 128, + 534, + 137 + ], + "spans": [ + { + "bbox": [ + 317, + 128, + 534, + 137 + ], + "type": "text", + "content": "(a) Image with \"person-break-baseball bat\" and \"hold-baseball bat\"." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 307, + 138, + 385, + 191 + ], + "lines": [ + { + "bbox": [ + 307, + 138, + 385, + 191 + ], + "spans": [ + { + "bbox": [ + 307, + 138, + 385, + 191 + ], + "type": "image", + "image_path": "0c455d5cab7e46117e8c591ebceb56b8f57bafb24bc6662232973cbab7dab19f.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 387, + 138, + 465, + 191 + ], + "blocks": [ + { + "bbox": [ + 387, + 138, + 465, + 191 + ], + "lines": [ + { + "bbox": [ + 387, + 138, + 465, + 191 + ], + "spans": [ + { + "bbox": [ + 387, + 138, + 465, + 191 + ], + "type": "image", + "image_path": "51e2755600f28b9ba7c8639160c56def36c898efa95c3d24d8d9c08303494f85.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 466, + 138, + 545, + 191 + ], + "blocks": [ + { + "bbox": [ + 466, + 138, + 545, + 191 + ], + "lines": [ + { + "bbox": [ + 466, + 138, + 545, + 191 + ], + "spans": [ + { + "bbox": [ + 466, + 138, + 545, + 191 + ], + "type": "image", + "image_path": "9647b202997c06d7eaa3c35abeecd263747d3dd6ee99e228e39f754859e401d7.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 307, + 206, + 385, + 259 + ], + "blocks": [ + { + "bbox": [ + 331, + 196, + 520, + 206 + ], + "lines": [ + { + "bbox": [ + 331, + 196, + 520, + 206 + ], + "spans": [ + { + "bbox": [ + 331, + 196, + 520, + 206 + ], + "type": "text", + "content": "(b) Image with \"person-direct-airplane\" and \"fly-airplane\"." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 307, + 206, + 385, + 259 + ], + "lines": [ + { + "bbox": [ + 307, + 206, + 385, + 259 + ], + "spans": [ + { + "bbox": [ + 307, + 206, + 385, + 259 + ], + "type": "image", + "image_path": "8f7492dcb75316b0373578dff4aa2d091da4a403ff996384eacc1f96bf4dd570.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 339, + 263, + 512, + 274 + ], + "lines": [ + { + "bbox": [ + 339, + 263, + 512, + 274 + ], + "spans": [ + { + "bbox": [ + 339, + 263, + 512, + 274 + ], + "type": "text", + "content": "(c) Image with \"person-hold-horse\" and \"ride-horse\"." + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 306, + 283, + 545, + 306 + ], + "lines": [ + { + "bbox": [ + 306, + 283, + 545, + 306 + ], + "spans": [ + { + "bbox": [ + 306, + 283, + 545, + 306 + ], + "type": "text", + "content": "Figure 3. From left to right: image, attention maps of the cross-attention layer in the decoder for different interaction categories." + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 387, + 206, + 465, + 259 + ], + "blocks": [ + { + "bbox": [ + 387, + 206, + 465, + 259 + ], + "lines": [ + { + "bbox": [ + 387, + 206, + 465, + 259 + ], + "spans": [ + { + "bbox": [ + 387, + 206, + 465, + 259 + ], + "type": "image", + "image_path": "9ceaf0fc4813a46f45b113af842239ef578ebfa73b0cc7aa99f6cd2a4f59e989.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 466, + 206, + 545, + 259 + ], + "blocks": [ + { + "bbox": [ + 466, + 206, + 545, + 259 + ], + "lines": [ + { + "bbox": [ + 466, + 206, + 545, + 259 + ], + "spans": [ + { + "bbox": [ + 466, + 206, + 545, + 259 + ], + "type": "image", + "image_path": "6aa65ad1b0f1578a4bcc497a7396b04fb3d878abc1f22548143f7c6548f7f973.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + } + ], + "index": 20 + }, + { + "bbox": [ + 306, + 319, + 538, + 333 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 319, + 538, + 333 + ], + "spans": [ + { + "bbox": [ + 306, + 319, + 538, + 333 + ], + "type": "text", + "content": "4. Integration to Off-the-shelf HOI Detectors" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 304, + 340, + 545, + 412 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 340, + 545, + 412 + ], + "spans": [ + { + "bbox": [ + 304, + 340, + 545, + 412 + ], + "type": "text", + "content": "As shown in Fig. 2, our method is ready to integrate with any baseline HOI method that provides image feature " + }, + { + "bbox": [ + 304, + 340, + 545, + 412 + ], + "type": "inline_equation", + "content": "\\mathbf{I}" + }, + { + "bbox": [ + 304, + 340, + 545, + 412 + ], + "type": "text", + "content": " and human-object instance feature " + }, + { + "bbox": [ + 304, + 340, + 545, + 412 + ], + "type": "inline_equation", + "content": "\\{F_i\\}" + }, + { + "bbox": [ + 304, + 340, + 545, + 412 + ], + "type": "text", + "content": ". The integration is simple. During inference, the human-object instance interaction classification part is replaced by our method in Sec. 3.2, the top and bottom right block in Fig. 2." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 304, + 412, + 545, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 412, + 545, + 449 + ], + "spans": [ + { + "bbox": [ + 304, + 412, + 545, + 449 + ], + "type": "text", + "content": "During training, the original loss in the baseline HOI method " + }, + { + "bbox": [ + 304, + 412, + 545, + 449 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{base}" + }, + { + "bbox": [ + 304, + 412, + 545, + 449 + ], + "type": "text", + "content": " is added to our image classification loss in Eq. (3). The final loss " + }, + { + "bbox": [ + 304, + 412, + 545, + 449 + ], + "type": "inline_equation", + "content": "\\mathcal{L}" + }, + { + "bbox": [ + 304, + 412, + 545, + 449 + ], + "type": "text", + "content": " for training is" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 376, + 463, + 545, + 476 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 376, + 463, + 545, + 476 + ], + "spans": [ + { + "bbox": [ + 376, + 463, + 545, + 476 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\mathcal {L} _ {\\text {b a s e}} + \\lambda * \\mathcal {L} _ {\\text {i m g}}, \\tag {5}", + "image_path": "8179a82a868cc477f02cb6f1e791c94dd7b25a4cfbe053bc7f03f2d9e0c4dee1.jpg" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 304, + 483, + 545, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 483, + 545, + 518 + ], + "spans": [ + { + "bbox": [ + 304, + 483, + 545, + 518 + ], + "type": "text", + "content": "where the weight " + }, + { + "bbox": [ + 304, + 483, + 545, + 518 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 304, + 483, + 545, + 518 + ], + "type": "text", + "content": " is 1.0 by default. All other hyper parameters and details during training remain the same as in the baseline HOI method." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 304, + 520, + 545, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 520, + 545, + 567 + ], + "spans": [ + { + "bbox": [ + 304, + 520, + 545, + 567 + ], + "type": "text", + "content": "Thus, our method is general and applicable to most existing HOI methods. In this work, we select three representative yet different baseline methods to verify the effectiveness of our approach, as described below." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 304, + 568, + 545, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 568, + 545, + 651 + ], + "spans": [ + { + "bbox": [ + 304, + 568, + 545, + 651 + ], + "type": "text", + "content": "QPIC [33] is the first to introduce transformer method into HOI task. It is also the baseline for many recent works [13, 28, 41, 45, 48]. Its performance is much better than early one-stage [14, 21, 38] and two-stage [7, 11, 17] methods while keeping a simple and end-to-end architecture. It consists of a CNN backbone as well as a transformer encoder and decoder." + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 304, + 653, + 545, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 653, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 304, + 653, + 545, + 689 + ], + "type": "text", + "content": "During our integration, the feature map in its transformer encoder is used as the image feature " + }, + { + "bbox": [ + 304, + 653, + 545, + 689 + ], + "type": "inline_equation", + "content": "\\mathbf{I}" + }, + { + "bbox": [ + 304, + 653, + 545, + 689 + ], + "type": "text", + "content": ". The human-object feature " + }, + { + "bbox": [ + 304, + 653, + 545, + 689 + ], + "type": "inline_equation", + "content": "\\{F_i\\}" + }, + { + "bbox": [ + 304, + 653, + 545, + 689 + ], + "type": "text", + "content": " is the query feature in its decoder." + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 304, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 689, + 545, + 713 + ], + "type": "text", + "content": "SCG [42] is a traditional two-stage method and the best in this category. It is also one of the best methods that" + } + ] + } + ], + "index": 31 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15278" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 48, + 70, + 547, + 163 + ], + "blocks": [ + { + "bbox": [ + 48, + 70, + 547, + 163 + ], + "lines": [ + { + "bbox": [ + 48, + 70, + 547, + 163 + ], + "spans": [ + { + "bbox": [ + 48, + 70, + 547, + 163 + ], + "type": "table", + "html": "
MethodPipelineE2EHICO-DETV-COCOEfficiency
FullRareNon-RareS1S2#ParamsFPS
QPIC [33]transformer✓28.9321.6231.1261.3963.6541M19.5
+ Ourstransformer✓31.08(+2.15)23.9033.2263.67(+2.28)65.4946M(+5M)18.3(-6.2%)
SCG [42]two-stageX31.2824.1633.4056.9362.5157M4.5
+ Ourstwo-stageX32.74(+1.46)26.2534.6859.14(+2.21)65.6164M(+7M)4.1(-8.9%)
GEN-VLKT [22]transformerX33.6929.9434.8164.8966.7442M21.7
+ OurstransformerX35.36(+1.67)32.9736.0766.40(+1.51)69.1747M(+5M)20.6(-5.1%)
", + "image_path": "4df4208d2accdb28e88a97e882663661a848ff76439b76ecb20c1cedb5bfd21b.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 170, + 545, + 194 + ], + "lines": [ + { + "bbox": [ + 46, + 170, + 545, + 194 + ], + "spans": [ + { + "bbox": [ + 46, + 170, + 545, + 194 + ], + "type": "text", + "content": "Table 1. The performance numbers of three different baseline HOI methods with and without integration of our method, on two datasets. \"E2E\" denotes whether a HOI detector is end-to-end. All models are tested on Tesla V100." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 205, + 287, + 264 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 205, + 287, + 264 + ], + "spans": [ + { + "bbox": [ + 46, + 205, + 287, + 264 + ], + "type": "text", + "content": "does not use transformer. It uses a multi-stream graph neural network(GNN) for interaction classification. In our experiment, the detection boxes are from a fine-tuned detector provided by DRG [7] for HICO-DET and a fine-tuned DETR for V-COCO." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 266, + 287, + 314 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 266, + 287, + 314 + ], + "spans": [ + { + "bbox": [ + 46, + 266, + 287, + 314 + ], + "type": "text", + "content": "During our integration, the CNN feature map in the backbone of SCG is used as image feature " + }, + { + "bbox": [ + 46, + 266, + 287, + 314 + ], + "type": "inline_equation", + "content": "\\mathbf{I}" + }, + { + "bbox": [ + 46, + 266, + 287, + 314 + ], + "type": "text", + "content": ". The human-object feature " + }, + { + "bbox": [ + 46, + 266, + 287, + 314 + ], + "type": "inline_equation", + "content": "\\{F_i\\}" + }, + { + "bbox": [ + 46, + 266, + 287, + 314 + ], + "type": "text", + "content": " is generated through RoI pooling with detected human and object boxes and fused with the GNN." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 315, + 287, + 374 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 315, + 287, + 374 + ], + "spans": [ + { + "bbox": [ + 46, + 315, + 287, + 374 + ], + "type": "text", + "content": "GEN-VLKT [22] is also transformer-based, but not end-to-end as pairwise NMS [41] is used for post-processing. It is the current state-of-the-art method. It uses two parallel decoders for object detection and interaction classification, namely instance decoder and interaction decoder." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 376, + 287, + 471 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 376, + 287, + 471 + ], + "spans": [ + { + "bbox": [ + 46, + 376, + 287, + 471 + ], + "type": "text", + "content": "During our integration, the feature map in its transformer encoder is used as image feature " + }, + { + "bbox": [ + 46, + 376, + 287, + 471 + ], + "type": "inline_equation", + "content": "\\mathbf{I}" + }, + { + "bbox": [ + 46, + 376, + 287, + 471 + ], + "type": "text", + "content": ". The query feature in the interaction decoder is used as the human-object feature " + }, + { + "bbox": [ + 46, + 376, + 287, + 471 + ], + "type": "inline_equation", + "content": "\\{F_i\\}" + }, + { + "bbox": [ + 46, + 376, + 287, + 471 + ], + "type": "text", + "content": ". Note that, unlike the majority of HOI detection methods, the original GEN-VLKT uses HOI categories rather than interaction categories during interaction classification. Our experiments still use interaction categories, in order to be consistent with most other methods." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 484, + 128, + 498 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 484, + 128, + 498 + ], + "spans": [ + { + "bbox": [ + 47, + 484, + 128, + 498 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 505, + 287, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 505, + 287, + 613 + ], + "spans": [ + { + "bbox": [ + 46, + 505, + 287, + 613 + ], + "type": "text", + "content": "In this section, we verify the applicability and effectiveness of the proposed method through experiments. In Sec. 5.1, we introduce the experimental settings. Then we demonstrate the effectiveness of the proposed method over 3 baselines in Sec. 5.2, and show it achieves SOTA results on major benchmarks in Sec. 5.3. Next, in Sec. 5.4 we conduct comprehensive ablation studies on the key components as well as detailed technical designs. Lastly, we provide some analysis and visualization in Sec. 5.5." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 622, + 110, + 634 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 622, + 110, + 634 + ], + "spans": [ + { + "bbox": [ + 47, + 622, + 110, + 634 + ], + "type": "text", + "content": "5.1. Datasets" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 641, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 641, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 641, + 287, + 715 + ], + "type": "text", + "content": "HICO-DET [3] and V-COCO [10] are two widely-used HOI benchmarks. HICO-DET contains 47,776 images, with 38,118 for training and 9,658 for testing. There are 600 HOI categories in HICO-DET, consisting of 117 interaction classes and 80 object classes. Each HOI category is composed of an interaction and an object. V-COCO is" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 205, + 545, + 253 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 205, + 545, + 253 + ], + "spans": [ + { + "bbox": [ + 304, + 205, + 545, + 253 + ], + "type": "text", + "content": "a subset of MS-COCO [24] with HOI annotations, including 10,346 images (2,533 for training, 2,867 for validation and 4,946 for testing). It has 80 object categories same with HICO-DET and 29 interaction categories." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 254, + 546, + 457 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 254, + 546, + 457 + ], + "spans": [ + { + "bbox": [ + 304, + 254, + 546, + 457 + ], + "type": "text", + "content": "Evaluation metrics. For HICO-DET, we adopt the commonly used mAP metric [3]. Each prediction is a " + }, + { + "bbox": [ + 304, + 254, + 546, + 457 + ], + "type": "inline_equation", + "content": "\\langle" + }, + { + "bbox": [ + 304, + 254, + 546, + 457 + ], + "type": "text", + "content": " human, interaction, object " + }, + { + "bbox": [ + 304, + 254, + 546, + 457 + ], + "type": "inline_equation", + "content": "\\rangle" + }, + { + "bbox": [ + 304, + 254, + 546, + 457 + ], + "type": "text", + "content": " triplet. A prediction is a true positive only when the human and object bounding boxes both have IoU " + }, + { + "bbox": [ + 304, + 254, + 546, + 457 + ], + "type": "inline_equation", + "content": ">0.5" + }, + { + "bbox": [ + 304, + 254, + 546, + 457 + ], + "type": "text", + "content": " w.r.t. ground truth and the interaction classification result is correct. We evaluate the performance in two different settings following [3]. In the known object setting, for each HOI category, we evaluate the prediction only on the images containing the target object category. In default setting, the detection result of each category is evaluated on the full test set. In each setting, we report the mAP over (1) all 600 HOI categories (Full), (2) 138 categories with less than 10 training samples (Rare), and (3) the remaining 462 categories (Non-rare). For V-COCO, we use the role mAP following [10], under both scenario #1 (including objects) and #2 (ignoring objects). The performance is evaluated using its official evaluation toolkit." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 466, + 528, + 479 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 466, + 528, + 479 + ], + "spans": [ + { + "bbox": [ + 305, + 466, + 528, + 479 + ], + "type": "text", + "content": "5.2. Improvement on Three Different Baselines" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 485, + 545, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 485, + 545, + 604 + ], + "spans": [ + { + "bbox": [ + 304, + 485, + 545, + 604 + ], + "type": "text", + "content": "Tab. 1 summarizes the performance of the three baseline HOI methods before and after integration of our method. Backbone is ResNet50. All these methods are significantly improved. Specifically, QPIC [33] is improved by " + }, + { + "bbox": [ + 304, + 485, + 545, + 604 + ], + "type": "inline_equation", + "content": "2.15\\mathrm{mAP}" + }, + { + "bbox": [ + 304, + 485, + 545, + 604 + ], + "type": "text", + "content": ", making it competitive with those more recent works [16, 41, 43]. SCG [42] is improved by 1.46 mAP, demonstrating that our method is not limited to transformer-based baselines. The current SOTA method GEN-VLKT [22] is improved by " + }, + { + "bbox": [ + 304, + 485, + 545, + 604 + ], + "type": "inline_equation", + "content": "1.67\\mathrm{mAP}" + }, + { + "bbox": [ + 304, + 485, + 545, + 604 + ], + "type": "text", + "content": ", producing the new SOTA result (also refer to Tab. 2)." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 605, + 545, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 545, + 640 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 545, + 640 + ], + "type": "text", + "content": "On the V-COCO dataset [10], the performance improvement is similar, which is 2.28, 2.21 and " + }, + { + "bbox": [ + 304, + 605, + 545, + 640 + ], + "type": "inline_equation", + "content": "1.51\\mathrm{mAP}" + }, + { + "bbox": [ + 304, + 605, + 545, + 640 + ], + "type": "text", + "content": " on QPIC, SCG and GEN-VLKT, respectively." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "type": "text", + "content": "Notably, for SCG [42], as the object detector is fixed during the training of its interaction classification network, the improvement by our method is purely due to better interaction classification, not a better fine-tuned CNN backbone or a better object detector. This further consolidates that the category query learning is effective." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15279" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 97, + 70, + 497, + 297 + ], + "blocks": [ + { + "bbox": [ + 97, + 70, + 497, + 297 + ], + "lines": [ + { + "bbox": [ + 97, + 70, + 497, + 297 + ], + "spans": [ + { + "bbox": [ + 97, + 70, + 497, + 297 + ], + "type": "table", + "html": "
MethodDetectorBackboneDefaultKnown Object
FullRareNon-rareFullRareNon-rare
DRG [7]HICO-DETResNet50-FPN24.5319.4726.0427.9823.1129.43
GG-Net [47]HICO-DETHourglass10423.4716.4825.6027.3620.2329.48
IDN [18]HICO-DETResNet5026.2922.6127.3928.2424.4729.37
QPIC [33]HICO-DETResNet5029.0721.8531.2331.6824.1433.93
SCG [42]HICO-DETResNet50-FPN31.3324.7233.3134.3727.1836.52
CDN [41]HICO-DETResNet5031.7827.5533.0534.5329.7335.96
DT [48]HICO-DETResNet5031.7527.4533.0334.5030.1335.81
STIP [44]HICO-DETResNet5031.6027.7532.7534.4130.1235.69
HQM [45]HICO-DETResNet5032.4728.1533.76---
MSTR [16]HICO-DETResNet5031.1725.3132.9234.0228.8335.57
RLIP [40]COCO+VGResNet5032.8426.8534.63---
IF [26]HICO-DETResNet5033.5130.3034.4636.2833.1637.21
GEN-VLKT-B [22]HICO-DETResNet5033.7529.2535.1036.7832.7537.99
GEN-VLKT-M [22]HICO-DETResNet10134.7831.5035.7738.0734.9439.01
GEN-VLKT-L [22]HICO-DETResNet10134.9531.1836.0838.2234.3639.37
BodyPartMap [39]HICO-DETResNet5035.1533.7135.5837.5635.8738.06
GEN-VLKT-B + OursHICO-DETResNet5035.3632.9736.0738.4334.8539.50
GEN-VLKT-M + OursHICO-DETResNet10135.8332.9136.7038.7935.2839.84
GEN-VLKT-L + OursHICO-DETResNet10136.0333.1636.8938.8235.5139.81
", + "image_path": "d7771f26a4592a7996c1641309e6c92b449727c80d684243974858a5bc968310.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 48, + 327, + 287, + 529 + ], + "blocks": [ + { + "bbox": [ + 94, + 305, + 499, + 316 + ], + "lines": [ + { + "bbox": [ + 94, + 305, + 499, + 316 + ], + "spans": [ + { + "bbox": [ + 94, + 305, + 499, + 316 + ], + "type": "text", + "content": "Table 2. The proposed method achieves state-of-the-art on HICO-DET [3]. The best results are marked in bold." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 48, + 327, + 287, + 529 + ], + "lines": [ + { + "bbox": [ + 48, + 327, + 287, + 529 + ], + "spans": [ + { + "bbox": [ + 48, + 327, + 287, + 529 + ], + "type": "table", + "html": "
MethodBackboneScenario #1Scenario #2
DRG [7]R50FPN51.0-
SCG [42]R5054.260.9
GG-Net [47]HG10454.7-
QPIC [33]R5058.861.0
HQM [45]R5063.6-
CDN [41]R5061.763.8
GEN-VLKT-B [22]R5062.464.5
GEN-VLKT-M [22]R10163.365.6
GEN-VLKT-L [22]R10163.665.9
MSTR [16]R5062.065.2
BodyPartMap [39]R5063.065.1
IF [26]R5063.065.2
DT [48]R5066.268.5
STIP [44]R5065.169.7
GEN-VLKT-B + OursR5066.469.2
GEN-VLKT-M + OursR10166.869.8
GEN-VLKT-L + OursR10166.569.9
", + "image_path": "d8a4328f669a88eec8d47905c521006b6c621c1b6c9ebfb69415047c9bbf3942.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 574, + 288, + 660 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 574, + 288, + 660 + ], + "spans": [ + { + "bbox": [ + 46, + 574, + 288, + 660 + ], + "type": "text", + "content": "To verify that the performance improvement is not due to a larger model, we also compare the model size and running speed. Our method increases the model parameters by a few millions, which is small compared to the original model size. The running speed, measured by FPS, is only decreased by a few percent. The marginal additional cost shows that our method is quite lightweight." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 670, + 227, + 683 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 670, + 227, + 683 + ], + "spans": [ + { + "bbox": [ + 47, + 670, + 227, + 683 + ], + "type": "text", + "content": "5.3. Comparison with State-of-the-art" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 689, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 689, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 689, + 288, + 715 + ], + "type": "text", + "content": "Tab. 2 and Tab. 3 compare our method with many previous methods for HICO-DET and V-COCO datasets, respec" + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 306, + 327, + 545, + 389 + ], + "blocks": [ + { + "bbox": [ + 46, + 536, + 287, + 559 + ], + "lines": [ + { + "bbox": [ + 46, + 536, + 287, + 559 + ], + "spans": [ + { + "bbox": [ + 46, + 536, + 287, + 559 + ], + "type": "text", + "content": "Table 3. Comparison with state-of-the-art methods on V-COCO [10] dataset. The best results are marked in bold." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 306, + 327, + 545, + 389 + ], + "lines": [ + { + "bbox": [ + 306, + 327, + 545, + 389 + ], + "spans": [ + { + "bbox": [ + 306, + 327, + 545, + 389 + ], + "type": "table", + "html": "
C1C2C3FullRareNon-Rare
a---33.6929.9434.81
b✓--33.86 (+0.17)31.1234.68
c✓✓-34.98 (+1.29)31.7335.95
d✓✓✓35.36 (+1.67)32.9736.07
", + "image_path": "9aa52c15fd14b12f7e1c2ea21a9b26c1a9c2531c6e57ce051540a806c3a5b0ad.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 398, + 545, + 431 + ], + "lines": [ + { + "bbox": [ + 305, + 398, + 545, + 431 + ], + "spans": [ + { + "bbox": [ + 305, + 398, + 545, + 431 + ], + "type": "text", + "content": "Table 4. Ablation study of several variants of our method, starting from the baseline (a) to our approach (d). C1, C2 and C3 are described in Sec. 5.4. The best results are marked in bold." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 305, + 445, + 500, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 445, + 500, + 456 + ], + "spans": [ + { + "bbox": [ + 305, + 445, + 500, + 456 + ], + "type": "text", + "content": "tively. GEN-VLKT [22] is used as our baseline." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 458, + 545, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 458, + 545, + 529 + ], + "spans": [ + { + "bbox": [ + 304, + 458, + 545, + 529 + ], + "type": "text", + "content": "On HICO-DET, our result with ResNet50 backbone already outperforms all previous methods under both default and known object settings. With the stronger ResNet101 backbone, our method achieves the new state-of-the-art 36.03 full mAP under default settings and 38.82 under known object settings." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 530, + 545, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 530, + 545, + 578 + ], + "spans": [ + { + "bbox": [ + 304, + 530, + 545, + 578 + ], + "type": "text", + "content": "On V-COCO dataset, our method achieves the new state-of-the-art performance on Scenario 1, with an AP of 66.4 for ResNet50, surpassing [48]. For scenario 2, it is comparable with the state-of-the-art [44]." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 586, + 432, + 600 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 586, + 432, + 600 + ], + "spans": [ + { + "bbox": [ + 306, + 586, + 432, + 600 + ], + "type": "text", + "content": "5.4. Ablation Experiments" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 605, + 545, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 545, + 641 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 545, + 641 + ], + "type": "text", + "content": "We perform various ablation experiments to validate the effectiveness of different components in our method. HICO-DET dataset and GEN-VLKT [22] baseline are used." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "type": "text", + "content": "The proposed method can be divided into 3 components: C1 means applying Query2Label [25] to the baseline detector as a multi-task learning (with feature extractor shared). In detail, it adds the queries, the decoder and image classification loss (Eq. (1), Eq. (2) and Eq. (5)). C2 means using the learned query in C1 as adaptive interaction classification" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "15280" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 59, + 70, + 276, + 194 + ], + "blocks": [ + { + "bbox": [ + 59, + 70, + 276, + 194 + ], + "lines": [ + { + "bbox": [ + 59, + 70, + 276, + 194 + ], + "spans": [ + { + "bbox": [ + 59, + 70, + 276, + 194 + ], + "type": "table", + "html": "
loss typeλFullRareNon-Rare
-034.2130.1535.42
focal loss [23]0.534.4331.0635.44
1.034.5131.0835.53
1.534.3531.5435.19
2.034.2931.1835.22
ASL [32]0.534.5731.9135.36
1.034.9831.7335.95
1.534.7732.0835.57
2.034.4131.9235.15
", + "image_path": "7fd5c2aae2cdb3d7c3b031283727ab588c36474d5764b26dc06f27430412bfc8.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 69, + 228, + 265, + 280 + ], + "blocks": [ + { + "bbox": [ + 47, + 202, + 287, + 224 + ], + "lines": [ + { + "bbox": [ + 47, + 202, + 287, + 224 + ], + "spans": [ + { + "bbox": [ + 47, + 202, + 287, + 224 + ], + "type": "text", + "content": "Table 5. Ablation on the type and weight " + }, + { + "bbox": [ + 47, + 202, + 287, + 224 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 47, + 202, + 287, + 224 + ], + "type": "text", + "content": " of the image classification loss. The best results are marked in bold." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 69, + 228, + 265, + 280 + ], + "lines": [ + { + "bbox": [ + 69, + 228, + 265, + 280 + ], + "spans": [ + { + "bbox": [ + 69, + 228, + 265, + 280 + ], + "type": "table", + "html": "
Layer structureFullRareNon-Rare
S→C→F34.7332.0935.52
C→S→F34.9831.7335.95
C→F34.6331.3435.61
", + "image_path": "32a48e79528c09fcf51ce5510070690cb016cb204a800d90beea99164b411da3.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 97, + 326, + 238, + 376 + ], + "blocks": [ + { + "bbox": [ + 47, + 289, + 287, + 321 + ], + "lines": [ + { + "bbox": [ + 47, + 289, + 287, + 321 + ], + "spans": [ + { + "bbox": [ + 47, + 289, + 287, + 321 + ], + "type": "text", + "content": "Table 6. Ablation on structure of each layer in the category decoder. Here \"S\", \"C\" and \"F\" stands for the self-attention, cross-attention and FFN in a standard transformer [35]." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 97, + 326, + 238, + 376 + ], + "lines": [ + { + "bbox": [ + 97, + 326, + 238, + 376 + ], + "spans": [ + { + "bbox": [ + 97, + 326, + 238, + 376 + ], + "type": "table", + "html": "
LFullRareNon-Rare
134.6631.4735.61
234.9831.7335.95
334.8631.5635.85
", + "image_path": "b2518088af689198302fcf2598dafda5211319cb939bf809da95ffd2080271bb.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 385, + 286, + 396 + ], + "lines": [ + { + "bbox": [ + 47, + 385, + 286, + 396 + ], + "spans": [ + { + "bbox": [ + 47, + 385, + 286, + 396 + ], + "type": "text", + "content": "Table 7. Ablation on the number of layers in the category decoder." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 47, + 411, + 287, + 459 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 411, + 287, + 459 + ], + "spans": [ + { + "bbox": [ + 47, + 411, + 287, + 459 + ], + "type": "text", + "content": "weight (Eq. (4)). This is the key component of the proposed method, which makes a distinction between the proposed method and a simple adaption of Query2label [25] to HOI task. C3 denotes the score integration step in Sec. 3.2." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 460, + 287, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 460, + 287, + 533 + ], + "spans": [ + { + "bbox": [ + 47, + 460, + 287, + 533 + ], + "type": "text", + "content": "To validate our approach in Fig. 2, several variants with these components in our approach are experimented and summarized in Tab. 4. First, (b) is a simple combination of Query2label [25] and the baseline detector (a) in a multitask setting. (b) is only slightly better than (a), showing that simply applying Query2Label to HOI is barely helpful." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 533, + 287, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 533, + 287, + 604 + ], + "spans": [ + { + "bbox": [ + 47, + 533, + 287, + 604 + ], + "type": "text", + "content": "Second, variant (c) significantly boosts (b), indicating using the queries as adaptive classification weights is the key to the performance improvement. This shows the effectiveness of our most crucial technical design in C2: applying the learned category queries as adaptive interaction classification weights." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 605, + 287, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 605, + 287, + 652 + ], + "spans": [ + { + "bbox": [ + 47, + 605, + 287, + 652 + ], + "type": "text", + "content": "Finally, the complete approach is (d). It further adds the integration step of image classification and instance classification score on (c). This technique produces moderate improvement over (c), i.e., " + }, + { + "bbox": [ + 47, + 605, + 287, + 652 + ], + "type": "inline_equation", + "content": "1.67\\mathrm{mAP}" + }, + { + "bbox": [ + 47, + 605, + 287, + 652 + ], + "type": "text", + "content": " vs. " + }, + { + "bbox": [ + 47, + 605, + 287, + 652 + ], + "type": "inline_equation", + "content": "1.29\\mathrm{mAP}" + }, + { + "bbox": [ + 47, + 605, + 287, + 652 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 653, + 287, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 653, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 47, + 653, + 287, + 712 + ], + "type": "text", + "content": "Image classification loss. Tab. 5 compares different loss functions and weights. First, when " + }, + { + "bbox": [ + 47, + 653, + 287, + 712 + ], + "type": "inline_equation", + "content": "\\lambda = 0" + }, + { + "bbox": [ + 47, + 653, + 287, + 712 + ], + "type": "text", + "content": ", which means no image-level supervision is applied, the improvement over the baseline (33.69 mAP) drops to only 0.52 mAP. This demonstrates the image classification supervision is essen" + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 310, + 73, + 544, + 282 + ], + "blocks": [ + { + "bbox": [ + 310, + 73, + 544, + 282 + ], + "lines": [ + { + "bbox": [ + 310, + 73, + 544, + 282 + ], + "spans": [ + { + "bbox": [ + 310, + 73, + 544, + 282 + ], + "type": "image", + "image_path": "827f372f985eaa1b832f1f34b3aaa43a0ee9b577522936d14a3ed751e170a219.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 294, + 545, + 392 + ], + "lines": [ + { + "bbox": [ + 305, + 294, + 545, + 392 + ], + "spans": [ + { + "bbox": [ + 305, + 294, + 545, + 392 + ], + "type": "text", + "content": "Figure 4. Some qualitative comparison between the baseline and the proposed method on HICO-DET. From left to right, column 1: true positive (TP) detection results, whose interaction score is increased by our method; column 2: false positive (FP) detection results, whose interaction score is decreased by our method; column 3: corresponding image-level GT and predictions by our method. Scores on the left and right of an image are the interaction classification scores of the visualized instance from the baseline and our method. Best viewed in color. More in supplementary materials." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 407, + 545, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 407, + 545, + 443 + ], + "spans": [ + { + "bbox": [ + 305, + 407, + 545, + 443 + ], + "type": "text", + "content": "tial to make the category query learning effective, with either focal loss or ASL. Additionally, ASL is slightly better than focal loss. By default, " + }, + { + "bbox": [ + 305, + 407, + 545, + 443 + ], + "type": "inline_equation", + "content": "\\lambda = 1.0" + }, + { + "bbox": [ + 305, + 407, + 545, + 443 + ], + "type": "text", + "content": " is adopted." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 445, + 545, + 516 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 445, + 545, + 516 + ], + "spans": [ + { + "bbox": [ + 305, + 445, + 545, + 516 + ], + "type": "text", + "content": "Is asymmetric loss the key? In the ablation above, we can see that ASL does help our image-level query learning. However, it is not the major reason for the performance improvement. To figure out this, we replace focal loss in the plain baseline GEN-VLKT with ASL and the result is only slightly better by " + }, + { + "bbox": [ + 305, + 445, + 545, + 516 + ], + "type": "inline_equation", + "content": "0.08\\mathrm{mAP}" + }, + { + "bbox": [ + 305, + 445, + 545, + 516 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 517, + 545, + 612 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 517, + 545, + 612 + ], + "spans": [ + { + "bbox": [ + 305, + 517, + 545, + 612 + ], + "type": "text", + "content": "Decoder structure. Tab. 6 compares several structures of the decoder. Compared to the standard decoder [2, 35] " + }, + { + "bbox": [ + 305, + 517, + 545, + 612 + ], + "type": "inline_equation", + "content": "(\\mathrm{S} \\rightarrow \\mathrm{C} \\rightarrow \\mathrm{F}" + }, + { + "bbox": [ + 305, + 517, + 545, + 612 + ], + "type": "text", + "content": " in the table), putting cross-attention first " + }, + { + "bbox": [ + 305, + 517, + 545, + 612 + ], + "type": "inline_equation", + "content": "(\\mathrm{C} \\rightarrow \\mathrm{S} \\rightarrow \\mathrm{F})" + }, + { + "bbox": [ + 305, + 517, + 545, + 612 + ], + "type": "text", + "content": " is slightly better (by " + }, + { + "bbox": [ + 305, + 517, + 545, + 612 + ], + "type": "inline_equation", + "content": "0.25 \\mathrm{mAP}" + }, + { + "bbox": [ + 305, + 517, + 545, + 612 + ], + "type": "text", + "content": ") without extra computation. If we remove self-attention " + }, + { + "bbox": [ + 305, + 517, + 545, + 612 + ], + "type": "inline_equation", + "content": "(\\mathrm{C} \\rightarrow \\mathrm{F})" + }, + { + "bbox": [ + 305, + 517, + 545, + 612 + ], + "type": "text", + "content": ", the performance drops by " + }, + { + "bbox": [ + 305, + 517, + 545, + 612 + ], + "type": "inline_equation", + "content": "0.35 \\mathrm{mAP}" + }, + { + "bbox": [ + 305, + 517, + 545, + 612 + ], + "type": "text", + "content": " compared with the " + }, + { + "bbox": [ + 305, + 517, + 545, + 612 + ], + "type": "inline_equation", + "content": "\\mathrm{S} \\rightarrow \\mathrm{C} \\rightarrow \\mathrm{F}" + }, + { + "bbox": [ + 305, + 517, + 545, + 612 + ], + "type": "text", + "content": " setting. This is probably because self-attention helps to learn the dependencies between different category queries." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 305, + 613, + 545, + 649 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 613, + 545, + 649 + ], + "spans": [ + { + "bbox": [ + 305, + 613, + 545, + 649 + ], + "type": "text", + "content": "Tab. 7 compares different numbers of decoder layers, denoted as " + }, + { + "bbox": [ + 305, + 613, + 545, + 649 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 305, + 613, + 545, + 649 + ], + "type": "text", + "content": ". We find that " + }, + { + "bbox": [ + 305, + 613, + 545, + 649 + ], + "type": "inline_equation", + "content": "L = 2" + }, + { + "bbox": [ + 305, + 613, + 545, + 649 + ], + "type": "text", + "content": " is sufficient. More layers do not help the performance." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 658, + 444, + 671 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 658, + 444, + 671 + ], + "spans": [ + { + "bbox": [ + 306, + 658, + 444, + 671 + ], + "type": "text", + "content": "5.5. Discussions and Analysis" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 305, + 677, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 677, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 677, + 545, + 713 + ], + "type": "text", + "content": "To understand why the category query learning is effective for human-object interaction classification, we provide some analysis and qualitative results." + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15281" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 63, + 74, + 271, + 167 + ], + "blocks": [ + { + "bbox": [ + 63, + 74, + 271, + 167 + ], + "lines": [ + { + "bbox": [ + 63, + 74, + 271, + 167 + ], + "spans": [ + { + "bbox": [ + 63, + 74, + 271, + 167 + ], + "type": "image", + "image_path": "199b250f04f984c175003beebe08d073e57ff36e116e51d60d06aaf1e7eeec19.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 62, + 172, + 269, + 264 + ], + "blocks": [ + { + "bbox": [ + 62, + 172, + 269, + 264 + ], + "lines": [ + { + "bbox": [ + 62, + 172, + 269, + 264 + ], + "spans": [ + { + "bbox": [ + 62, + 172, + 269, + 264 + ], + "type": "image", + "image_path": "2e06852b319c9a0790037fa9831788f097d23b9a7f7b6c7bb0d8c4d1db475004.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 267, + 287, + 332 + ], + "lines": [ + { + "bbox": [ + 46, + 267, + 287, + 332 + ], + "spans": [ + { + "bbox": [ + 46, + 267, + 287, + 332 + ], + "type": "text", + "content": "Figure 5. Performance evaluated on image partitions with different " + }, + { + "bbox": [ + 46, + 267, + 287, + 332 + ], + "type": "inline_equation", + "content": "n = 1,2,3,4,5, > 5" + }, + { + "bbox": [ + 46, + 267, + 287, + 332 + ], + "type": "text", + "content": " instances for an interaction category. Top: mAP numbers for three baselines (dashed) as well as our integration (solid). Bottom: relative mAP ratios for improvement produced by our method (solid) and between two arbitrary previous methods (dashed)." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 350, + 286, + 458 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 350, + 286, + 458 + ], + "spans": [ + { + "bbox": [ + 46, + 350, + 286, + 458 + ], + "type": "text", + "content": "The attention maps in the cross-attention layer of the decoder is visualized in Fig. 3. For different category queries, the corresponding attention maps show they learn to capture the semantics of the category, while being adaptive to different images. For example, in Fig. 3a, the broken part of the bat in the air is highlighted for \"break\" while the left part in the hand is highlighted for \"hold\". This is similar for Fig. 3b. In Fig. 3c, the attention map highlights many instances with corresponding action \"hold\" and \"ride\"." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 460, + 286, + 627 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 460, + 286, + 627 + ], + "spans": [ + { + "bbox": [ + 46, + 460, + 286, + 627 + ], + "type": "text", + "content": "To qualitatively demonstrate how our method helps, we visualize some cases of the baseline and the proposed method in Fig. 4. In the first case, the baseline predicts a TP of \"person sit on bus\" with a low score 0.15, and a FP of \"person board bus\" with a high score 0.29. Our model also predicts the same TP and FP, but lowers the FP score to 0.07 and lifts the TP score to 0.38. Besides, our model correctly predicts image-level scores: the wrong category \"board\" is given a low score of 0.06 while the four correct categories are given high scores. In the second case, our model successfully predicts the TP \"person dribble\" missed by the baseline and suppress the FP \"person hold sports\" from 0.6 to 0.11 with the help of correct image-level classification result. This is similar for the third case." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 630, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 630, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 630, + 286, + 713 + ], + "type": "text", + "content": "Last, as our category query learning is performed on image level, we conjecture that it is more helpful for images with dense human-object interactions. In such images, an human-object instance is relatively small and hard to learn good feature on its own. However, it may benefit more from the global image level category query feature, which aggregates more information from other similar instances in this" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 72, + 545, + 227 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 227 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 227 + ], + "type": "text", + "content": "image. To validate this conjecture, we partition the images according to their \"interaction density\" and check whether our method produces larger improvement on images that are \"denser\". Specifically, for each interaction category, its mAP is evaluated separately on six different image partition subsets, where each image contains different numbers " + }, + { + "bbox": [ + 304, + 72, + 545, + 227 + ], + "type": "inline_equation", + "content": "(n = 1, 2, 3, 4, 5" + }, + { + "bbox": [ + 304, + 72, + 545, + 227 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 72, + 545, + 227 + ], + "type": "inline_equation", + "content": ">5" + }, + { + "bbox": [ + 304, + 72, + 545, + 227 + ], + "type": "text", + "content": ") of human-object instances of this category. The mAP results of the three baseline methods and their integrated versions (as in Tab. 1) are shown in Fig. 5 (top, dashed vs. solid lines). It shows that: 1) the mAP is lower for larger n, indicating the \"denser\" images are more challenging; 2) our method improves the baselines consistently on all different partitions." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 228, + 545, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 228, + 545, + 323 + ], + "spans": [ + { + "bbox": [ + 304, + 228, + 545, + 323 + ], + "type": "text", + "content": "To check whether the proposed method is more effective on \"denser\" images, we use the relative mAP improvement, which is a ratio, " + }, + { + "bbox": [ + 304, + 228, + 545, + 323 + ], + "type": "inline_equation", + "content": "\\frac{mAP_{ours} - mAP_{baseline}}{mAP_{baseline}}" + }, + { + "bbox": [ + 304, + 228, + 545, + 323 + ], + "type": "text", + "content": ", for analysis. The ratio curves of the three baselines are shown in Fig. 5 (bottom, solid lines). It is clear that the relative improvement becomes larger for larger " + }, + { + "bbox": [ + 304, + 228, + 545, + 323 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 304, + 228, + 545, + 323 + ], + "type": "text", + "content": ". This indicates that the image-level category query learning is more effective on these challenging dense images." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 323, + 545, + 419 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 323, + 545, + 419 + ], + "spans": [ + { + "bbox": [ + 304, + 323, + 545, + 419 + ], + "type": "text", + "content": "To further verify that this behavior is not commonly true, we also compute the relative ratio of three more comparisons, \"gen-vlkt vs. qpic\", \"gen-vlkt vs. scg\" and \"scg vs. qpic\", in which the former outperforms the latter. These curves are also shown in Fig. 5 (bottom, dashed lines). There is no clear pattern in these curves, indicating that the performance gap between two arbitrary HOI methods are in general not related to the image \"density\"." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 430, + 378, + 442 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 430, + 378, + 442 + ], + "spans": [ + { + "bbox": [ + 306, + 430, + 378, + 442 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 450, + 545, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 450, + 545, + 533 + ], + "spans": [ + { + "bbox": [ + 304, + 450, + 545, + 533 + ], + "type": "text", + "content": "This work proposes a novel approach for the human-object interaction classification sub-task in HOI detection. We study the problem of interaction category modeling, in contrast to most previous methods focusing on human-object feature learning. We adopt the concept of category query in a previous method [25] for HOI, for the first time, and show that it is simple, general and highly effective." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 534, + 545, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 534, + 545, + 568 + ], + "spans": [ + { + "bbox": [ + 304, + 534, + 545, + 568 + ], + "type": "text", + "content": "Clearly, this idea of category query modeling is not limited to multi-label image classification and HOI detection. We hope it is useful for other vision tasks." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "text", + "content": "Acknowledgment. This work was supported in part by the National Natural Science Foundation of China under Grant 62076183, 61936014 and 61976159, in part by the Natural Science Foundation of Shanghai under Grant 20ZR1473500, in part by the Shanghai Science and Technology Innovation Action Project of under Grant 20511100700 and 22511105300, in part by the Shanghai Municipal Science and Technology Major Project under Grant 2021SHZDZX0100, and in part by the Fundamental Research Funds for the Central Universities. The authors would also like to thank the anonymous reviewers for their careful work and valuable suggestions." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15282" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 134 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 134 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 134 + ], + "type": "text", + "content": "[1] A Bansal, S. S Rambhatla, A Shrivastava, and R Chellappa. Detecting human-object interactions via functional generalization. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 34, pages 10460-10469, 2020. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 135, + 288, + 178 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 135, + 288, + 178 + ], + "spans": [ + { + "bbox": [ + 53, + 135, + 288, + 178 + ], + "type": "text", + "content": "[2] N Carion, F Massa, G Synnaeve, N Usunier, A Kirillov, and S Zagoruyko. End-to-end object detection with transformers. In European Conference on Computer Vision, pages 213-229. Springer, 2020. 2, 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 180, + 288, + 223 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 180, + 288, + 223 + ], + "spans": [ + { + "bbox": [ + 53, + 180, + 288, + 223 + ], + "type": "text", + "content": "[3] Y.-W Chao, Y Liu, X Liu, H Zeng, and J Deng. Learning to detect human-object interactions. In 2018 iee winter conference on applications of computer vision (wacv), pages 381-389. IEEE, 2018. 1, 5, 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 224, + 288, + 267 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 224, + 288, + 267 + ], + "spans": [ + { + "bbox": [ + 53, + 224, + 288, + 267 + ], + "type": "text", + "content": "[4] M Chen, Y Liao, S Liu, Z Chen, F Wang, and C Qian. Reformulating hoi detection as adaptive set prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9004–9013, 2021. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 269, + 288, + 322 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 269, + 288, + 322 + ], + "spans": [ + { + "bbox": [ + 53, + 269, + 288, + 322 + ], + "type": "text", + "content": "[5] B Cheng, I Misra, A. G Schwing, A Kirillov, and R Girdhar. Masked-attention mask transformer for universal image segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1290–1299, 2022. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 324, + 288, + 378 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 324, + 288, + 378 + ], + "spans": [ + { + "bbox": [ + 53, + 324, + 288, + 378 + ], + "type": "text", + "content": "[6] L Dong, Z Li, K Xu, Z Zhang, L Yan, S Zhong, and X Zou. Category-aware transformer network for better human-object interaction detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19538-19547, 2022. 1, 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 380, + 288, + 423 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 380, + 288, + 423 + ], + "spans": [ + { + "bbox": [ + 53, + 380, + 288, + 423 + ], + "type": "text", + "content": "[7] C Gao, J Xu, Y Zou, and J.-B Huang. Drg: Dual relation graph for human-object interaction detection. In European Conference on Computer Vision, pages 696-712. Springer, 2020. 1, 2, 4, 5, 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 424, + 288, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 424, + 288, + 456 + ], + "spans": [ + { + "bbox": [ + 53, + 424, + 288, + 456 + ], + "type": "text", + "content": "[8] C Gao, Y Zou, and J.-B Huang. ican: Instance-centric attention network for human-object interaction detection. arXiv preprint arXiv:1808.10437, 2018. 1, 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 458, + 288, + 501 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 458, + 288, + 501 + ], + "spans": [ + { + "bbox": [ + 53, + 458, + 288, + 501 + ], + "type": "text", + "content": "[9] G Gkioxari, R Girshick, P Dólár, and K He. Detecting and recognizing human-object interactions. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 8359-8367, 2018. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 502, + 287, + 524 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 502, + 287, + 524 + ], + "spans": [ + { + "bbox": [ + 48, + 502, + 287, + 524 + ], + "type": "text", + "content": "[10] S Gupta and J Malik. Visual semantic role labeling. arXiv preprint arXiv:1505.04474, 2015. 1, 5, 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 525, + 287, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 525, + 287, + 578 + ], + "spans": [ + { + "bbox": [ + 48, + 525, + 287, + 578 + ], + "type": "text", + "content": "[11] T Gupta, A Schwing, and D Hoiem. No-frills human-object interaction detection: Factorization, layout encodings, and training techniques. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9677-9685, 2019. 2, 4" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 580, + 287, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 580, + 287, + 634 + ], + "spans": [ + { + "bbox": [ + 48, + 580, + 287, + 634 + ], + "type": "text", + "content": "[12] H He, Y Yuan, X Yue, and H Hu. Rankseg: Adaptive pixel classification with image category ranking for segmentation. In Proceedings of the European Conference on Computer Vision (ECCV), pages 682-700. Springer Nature Switzerland Cham, 2022. 4" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 635, + 287, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 635, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 635, + 287, + 689 + ], + "type": "text", + "content": "[13] A Iftekhar, H Chen, K Kundu, X Li, J Tighe, and D Modolo. What to look at and where: Semantic and spatial refined transformer for detecting human-object interactions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5353-5363, 2022. 2, 4" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "type": "text", + "content": "[14] B Kim, T Choi, J Kang, and H. J Kim. Uniondet: Union-level detector towards real-time human-object interaction de" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "text", + "content": "tlection. In European Conference on Computer Vision, pages 498-514. Springer, 2020. 4" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 95, + 545, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 95, + 545, + 139 + ], + "spans": [ + { + "bbox": [ + 308, + 95, + 545, + 139 + ], + "type": "text", + "content": "[15] B Kim, J Lee, J Kang, E.-S Kim, and H. J Kim. Hotr: End-to-end human-object interaction detection with transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 74-83, 2021. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 140, + 545, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 140, + 545, + 194 + ], + "spans": [ + { + "bbox": [ + 308, + 140, + 545, + 194 + ], + "type": "text", + "content": "[16] B Kim, J Mun, K.-W On, M Shin, J Lee, and E.-S Kim. Mstr: Multi-scale transformer for end-to-end human-object interaction detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19578-19587, 2022. 1, 2, 5, 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 195, + 545, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 195, + 545, + 239 + ], + "spans": [ + { + "bbox": [ + 308, + 195, + 545, + 239 + ], + "type": "text", + "content": "[17] D.-J Kim, X Sun, J Choi, S Lin, and I. S Kweon. Detecting human-object interactions with action co-occurrence priors. In European Conference on Computer Vision, pages 718–736. Springer, 2020. 2, 4" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 239, + 545, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 239, + 545, + 281 + ], + "spans": [ + { + "bbox": [ + 308, + 239, + 545, + 281 + ], + "type": "text", + "content": "[18] Y.-L Li, X Liu, X Wu, Y Li, and C Lu. Hoi analysis: Integrating and decomposing human-object interaction. Advances in Neural Information Processing Systems, 33:5011-5022, 2020. 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 283, + 545, + 337 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 283, + 545, + 337 + ], + "spans": [ + { + "bbox": [ + 308, + 283, + 545, + 337 + ], + "type": "text", + "content": "[19] Y.-L Li, L Xu, X Liu, X Huang, Y Xu, S Wang, H.-S Fang, Z Ma, M Chen, and C Lu. Pastanet: Toward human activity knowledge engine. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 382-391, 2020. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 338, + 545, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 338, + 545, + 392 + ], + "spans": [ + { + "bbox": [ + 308, + 338, + 545, + 392 + ], + "type": "text", + "content": "[20] Y.-L Li, S Zhou, X Huang, L Xu, Z Ma, H.-S Fang, Y Wang, and C Lu. Transferable interactiveness knowledge for human-object interaction detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3585-3594, 2019. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 394, + 545, + 448 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 394, + 545, + 448 + ], + "spans": [ + { + "bbox": [ + 308, + 394, + 545, + 448 + ], + "type": "text", + "content": "[21] Y Liao, S Liu, F Wang, Y Chen, C Qian, and J Feng. Ppdm: Parallel point detection and matching for real-time human-object interaction detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 482-490, 2020. 1, 4" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 449, + 545, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 449, + 545, + 502 + ], + "spans": [ + { + "bbox": [ + 308, + 449, + 545, + 502 + ], + "type": "text", + "content": "[22] Y Liao, A Zhang, M Lu, Y Wang, X Li, and S Liu. Gen-vlkt: Simplify association and enhance interaction understanding for hoi detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20123-20132, 2022. 1, 2, 5, 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 504, + 545, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 504, + 545, + 547 + ], + "spans": [ + { + "bbox": [ + 308, + 504, + 545, + 547 + ], + "type": "text", + "content": "[23] T.-Y Lin, P Goyal, R Girshick, K He, and P Dollar. Focal loss for dense object detection. In Proceedings of the IEEE international conference on computer vision, pages 2980-2988, 2017. 3, 7" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 548, + 545, + 592 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 548, + 545, + 592 + ], + "spans": [ + { + "bbox": [ + 308, + 548, + 545, + 592 + ], + "type": "text", + "content": "[24] T.-Y Lin, M Maire, S Belongie, J Hays, P Perona, D Ramanan, P Dollár, and C. L Zitnick. Microsoft coco: Common objects in context. In European conference on computer vision, pages 740-755. Springer, 2014. 5" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 593, + 545, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 593, + 545, + 624 + ], + "spans": [ + { + "bbox": [ + 308, + 593, + 545, + 624 + ], + "type": "text", + "content": "[25] S Liu, L Zhang, X Yang, H Su, and J Zhu. Query2label: A simple transformer way to multi-label classification. arXiv preprint arXiv:2107.10834, 2021. 1, 2, 3, 4, 6, 7, 8" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 625, + 545, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 625, + 545, + 669 + ], + "spans": [ + { + "bbox": [ + 308, + 625, + 545, + 669 + ], + "type": "text", + "content": "[26] X Liu, Y.-L Li, X Wu, Y.-W Tai, C Lu, and C.-K Tang. Interactiveness field in human-object interactions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20113-20122, 2022. 6" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 308, + 670, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 670, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 308, + 670, + 545, + 712 + ], + "type": "text", + "content": "[27] Y Liu, Q Chen, and A Zisserman. Amplifying key cues for human-object-interaction detection. In European Conference on Computer Vision, pages 248-265. Springer, 2020. 2" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15283" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 714 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "text", + "content": "[28] J Park, S Lee, H Heo, H. K Choi, and H. J Kim. Consistency learning via decoding path augmentation for transformers in human object interaction detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1019-1028, 2022. 4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 128, + 287, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 128, + 287, + 172 + ], + "spans": [ + { + "bbox": [ + 48, + 128, + 287, + 172 + ], + "type": "text", + "content": "[29] S Qi, W Wang, B Jia, J Shen, and S.-C Zhu. Learning human-object interactions by graph parsing neural networks. In Proceedings of the European Conference on Computer Vision (ECCV), pages 401-417, 2018. 1, 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 172, + 288, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 172, + 288, + 226 + ], + "spans": [ + { + "bbox": [ + 48, + 172, + 288, + 226 + ], + "type": "text", + "content": "[30] X Qu, C Ding, X Li, X Zhong, and D Tao. Distillation using oracle queries for transformer-based human-object interaction detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19558-19567, 2022. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 228, + 288, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 228, + 288, + 282 + ], + "spans": [ + { + "bbox": [ + 48, + 228, + 288, + 282 + ], + "type": "text", + "content": "[31] A Radford, J. W Kim, C Hallacy, A Ramesh, G Goh, S Agarwal, G Sastry, A Askell, P Mishkin, J Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 283, + 287, + 336 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 283, + 287, + 336 + ], + "spans": [ + { + "bbox": [ + 48, + 283, + 287, + 336 + ], + "type": "text", + "content": "[32] T Ridnik, E Ben-Baruch, N Zamir, A Noy, I Friedman, M Protter, and L Zelnik-Manor. Asymmetric loss for multi-label classification. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 82-91, 2021. 3, 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 338, + 287, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 338, + 287, + 392 + ], + "spans": [ + { + "bbox": [ + 48, + 338, + 287, + 392 + ], + "type": "text", + "content": "[33] M Tamura, H Ohashi, and T Yoshinaga. Qpic: Query-based pairwise human-object interaction detection with image-wide contextual information. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10410-10419, 2021. 1, 2, 4, 5, 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 393, + 287, + 448 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 393, + 287, + 448 + ], + "spans": [ + { + "bbox": [ + 48, + 393, + 287, + 448 + ], + "type": "text", + "content": "[34] O Ulutan, A Iftekhar, and B. S Manjunath. Vsgnet: Spatial attention network for detecting human object interactions using graph convolutions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13617-13626, 2020. 1, 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 449, + 287, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 449, + 287, + 491 + ], + "spans": [ + { + "bbox": [ + 48, + 449, + 287, + 491 + ], + "type": "text", + "content": "[35] A Vaswani, N Shazeer, N Parmar, J Uszkoreit, L Jones, A. N Gomez, L Kaiser, and I Polosukhin. Attention is all you need. In Advances in neural information processing systems, pages 5998-6008, 2017. 2, 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 492, + 287, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 492, + 287, + 536 + ], + "spans": [ + { + "bbox": [ + 48, + 492, + 287, + 536 + ], + "type": "text", + "content": "[36] B Wan, D Zhou, Y Liu, R Li, and X He. Pose-aware multilevel feature network for human object interaction detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9469-9478, 2019. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 537, + 287, + 581 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 537, + 287, + 581 + ], + "spans": [ + { + "bbox": [ + 48, + 537, + 287, + 581 + ], + "type": "text", + "content": "[37] H Wang, W.-s Zheng, and L Yingbiao. Contextual heterogeneous graph network for human-object interaction detection. In European Conference on Computer Vision, pages 248-264. Springer, 2020. 1, 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 582, + 287, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 582, + 287, + 635 + ], + "spans": [ + { + "bbox": [ + 48, + 582, + 287, + 635 + ], + "type": "text", + "content": "[38] T Wang, T Yang, M Danelljan, F. S Khan, X Zhang, and J Sun. Learning human-object interaction detection using interaction points. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4116-4125, 2020. 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 636, + 287, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 636, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 636, + 287, + 689 + ], + "type": "text", + "content": "[39] X Wu, Y.-L Li, X Liu, J Zhang, Y Wu, and C Lu. Mining cross-person cues for body-part interactiveness learning in hoi detection. In Proceedings of the European Conference on Computer Vision (ECCV), pages 121-136. Springer, 2022. 2, 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 691, + 287, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 287, + 714 + ], + "type": "text", + "content": "[40] H Yuan, J Jiang, S Albanie, T Feng, Z Huang, D Ni, and M Tang. Rlip: Relational language-image pre-training for" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 564 + ], + "type": "list", + "angle": 0, + "index": 24, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "text", + "content": "human-object interaction detection. In Advances in Neural Information Processing Systems, 2022. 2, 6" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 96, + 545, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 96, + 545, + 140 + ], + "spans": [ + { + "bbox": [ + 307, + 96, + 545, + 140 + ], + "type": "text", + "content": "[41] A Zhang, Y Liao, S Liu, M Lu, Y Wang, C Gao, and X Li. Mining the benefits of two-stage and one-stage hoi detection. In Thirty-Fifth Conference on Neural Information Processing Systems, 2021. 1, 2, 4, 5, 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 141, + 545, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 141, + 545, + 185 + ], + "spans": [ + { + "bbox": [ + 307, + 141, + 545, + 185 + ], + "type": "text", + "content": "[42] F. Z Zhang, D Campbell, and S Gould. Spatially conditioned graphs for detecting human-object interactions. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13319-13327, 2021. 2, 4, 5, 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 186, + 545, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 186, + 545, + 239 + ], + "spans": [ + { + "bbox": [ + 307, + 186, + 545, + 239 + ], + "type": "text", + "content": "[43] F. Z Zhang, D Campbell, and S Gould. Efficient two-stage detection of human-object interactions with a novel unary-pairwise transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20104-20112, 2022. 5" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 241, + 545, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 241, + 545, + 297 + ], + "spans": [ + { + "bbox": [ + 307, + 241, + 545, + 297 + ], + "type": "text", + "content": "[44] Y Zhang, Y Pan, T Yao, R Huang, T Mei, and C.-W Chen. Exploring structure-aware transformer over interaction proposals for human-object interaction detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19548-19557, 2022. 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 297, + 545, + 351 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 297, + 545, + 351 + ], + "spans": [ + { + "bbox": [ + 307, + 297, + 545, + 351 + ], + "type": "text", + "content": "[45] X Zhong, C Ding, Z Li, and S Huang. Towards hard-positive query mining for detr-based human-object interaction detection. In Proceedings of the European Conference on Computer Vision (ECCV), pages 444-460. Springer, 2022. 2, 4, 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 353, + 545, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 353, + 545, + 396 + ], + "spans": [ + { + "bbox": [ + 307, + 353, + 545, + 396 + ], + "type": "text", + "content": "[46] X Zhong, C Ding, X Qu, and D Tao. Polysemy deciphering network for robust human-object interaction detection. International Journal of Computer Vision, 129(6):1910-1929, 2021. 1, 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 397, + 545, + 452 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 397, + 545, + 452 + ], + "spans": [ + { + "bbox": [ + 307, + 397, + 545, + 452 + ], + "type": "text", + "content": "[47] X Zhong, X Qu, C Ding, and D Tao. Glance and gaze: Inferring action-aware points for one-stage human-object interaction detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13234-13243, 2021. 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 453, + 545, + 508 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 453, + 545, + 508 + ], + "spans": [ + { + "bbox": [ + 307, + 453, + 545, + 508 + ], + "type": "text", + "content": "[48] D Zhou, Z Liu, J Wang, L Wang, T Hu, E Ding, and J Wang. Human-object interaction detection via disentangled transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19568-19577, 2022. 1, 2, 4, 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 509, + 545, + 564 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 509, + 545, + 564 + ], + "spans": [ + { + "bbox": [ + 307, + 509, + 545, + 564 + ], + "type": "text", + "content": "[49] C Zou, B Wang, Y Hu, J Liu, Q Wu, Y Zhao, B Li, C Zhang, C Zhang, Y Wei, et al. End-to-end human object interaction detection with hoi transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11825-11834, 2021. 2" + } + ] + } + ], + "index": 23 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "15284" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Causally-Aware Intraoperative Imputation for Overall Survival Time Prediction/58789d3b-4b75-4c2b-ad82-075b8268b238_content_list.json b/2023/Causally-Aware Intraoperative Imputation for Overall Survival Time Prediction/58789d3b-4b75-4c2b-ad82-075b8268b238_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..e73a7a0a8b03af73a1c0623751c56525be61da0f --- /dev/null +++ b/2023/Causally-Aware Intraoperative Imputation for Overall Survival Time Prediction/58789d3b-4b75-4c2b-ad82-075b8268b238_content_list.json @@ -0,0 +1,1417 @@ +[ + { + "type": "text", + "text": "Causally-Aware Intraoperative Imputation for Overall Survival Time Prediction", + "text_level": 1, + "bbox": [ + 81, + 130, + 888, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Xiang Li $^{1,*}$ , Xuelin Qian $^{1,*}$ , Litian Liang $^{1,*}$ , Lingjie Kong $^{1}$ , Qiaole Dong $^{1}$ , Jiejun Chen $^{2}$ , Dingxia Liu $^{2}$ , Xiuzhong Yao $^{2,\\dagger}$ , Yanwei Fu $^{1,\\dagger}$", + "bbox": [ + 125, + 179, + 843, + 217 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ School of Data Science, Fudan University", + "bbox": [ + 315, + 233, + 653, + 251 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ Department of Radiology, Zhongshan Hospital, Fudan University", + "bbox": [ + 222, + 252, + 745, + 268 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{li_x20, xlqian, ltliang19, qldong18, jjchen20, yanweifu}@fudan.edu.cn", + "bbox": [ + 197, + 271, + 769, + 287 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{jkong22,dxliu21}@m.fudan.edu.cn yao.xiuzhong@zs-hospital.sh.cn", + "bbox": [ + 187, + 290, + 776, + 305 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/0738ecee42815d685eaa17534391c0dc9b1ca91398d3944bca76e79b91c8f7a7.jpg", + "image_caption": [ + "Figure 1. Our idea illustration. Patient A and B have different OS time (109 months and 16 months, respectively) but similar pre-operative image-based patterns and indexes; thus, preoperative based model fails to distinguish A and B (their predicted OS time are both larger than 9 years). However, their intra-operative attributes, which can describe the severity of disease in a more informative way, are very different from each other (e.g., G-score, S-score, Clinicopathologic hepatocirrhosis). Therefore, it can help the model discriminate A and B. In the right image, we show that our CAWIM – that leverages intra-operative indexes in the training stage – can largely improve the prediction power than other methods. Our method managed to correctly classify these two cases, and the overall performance of our CAWIM surpasses other baseline models by approximately 10 points on 4-category classification task." + ], + "image_footnote": [], + "bbox": [ + 89, + 335, + 883, + 470 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 578, + 313, + 594 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Previous efforts in vision community are mostly made on learning good representations from visual patterns. Beyond this, this paper emphasizes the high-level ability of causal reasoning. We thus present a case study of solving the challenging task of Overall Survival (OS) time in primary liver cancers. Critically, the prediction of OS time at the early stage remains challenging, due to the unobvious image patterns of reflecting the OS. To this end, we propose a causal inference system by leveraging the intraoperative attributes and the correlation among them, as an intermediate supervision to bridge the gap between the images and the final OS. Particularly, we build a causal graph, and train the images to estimate the intraoperative attributes for final OS prediction. We present a novel Causally-aware Intraoperative Imputation Model (CAWIM) that can sequentially predict each attribute using its parent nodes in the estimated causal graph. To determine the causal directions,", + "bbox": [ + 75, + 609, + 470, + 864 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "we propose a splitting-voting mechanism, which votes for the direction for each pair of adjacent nodes among multiple predictions obtained via causal discovery from heterogeneity. The practicability and effectiveness of our method are demonstrated by the promising results on liver cancer dataset of 361 patients with long-term observations.", + "bbox": [ + 500, + 580, + 892, + 671 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 500, + 705, + 630, + 722 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The success of recent deep learning model is largely attributed to learning the good representations for visual patterns. Such representations essentially facilitate various vision task, such as recognition and synthesis [15, 25, 33]. Nevertheless, one important goal for the vision community is to model and summarize the relationships of observed variables of a system, in order to enable well predictions on similar data. Essentially, it is desirable to understand how the system is changed if one modifies these relationships under certain conditions, e.g., the effects of a treatment in healthcare. Thus this demands the high-", + "bbox": [ + 496, + 734, + 892, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Equal contribution", + "bbox": [ + 94, + 875, + 202, + 887 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "† Corresponding author", + "bbox": [ + 96, + 887, + 218, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "15681", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "level ability of causal reasoning beyond the previous efforts of only learning good representations for visual patterns [1, 5, 16, 21, 35, 37]. This naturally leads into our task of causal inference.", + "bbox": [ + 75, + 90, + 468, + 151 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "This paper presents a case study of solving the challenging task of Overall Survival (OS) time estimation in Primary Liver Cancers (PLC). Generally, the liver cancer remains one of the most common malignancies worldwide in the 21st century, as there are about one million new cases every year [36]. The five-year survival rate for advanced PLC is only about $5\\%$ [7]. Therefore, early and accurate prediction of OS time estimation can provide informative guidance for individualized treatment planning and reducing burden of medical resources [6, 27, 32, 34]. On the other hand, one shall easily notice that with the renaissance of deep learning, great achievements have been made on medical imaging analysis, such as diagnosis, segmentation, and classification [16, 21, 35, 37]. Unfortunately, it still remains challenging for experienced clinicians to predict OS time at early stage, even with advanced modern diagnostic tools such as Magnetic Resonance Imaging (MRI) or tumor marker tests [3, 11], and the deep learning tools [1, 5].", + "bbox": [ + 75, + 154, + 472, + 426 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Some studies propose to leverage deep neural networks for OS time prediction. They focus on fusion learning of multi-modal image features and some basic information (i.e., age and gender) [5, 8, 14, 22, 24, 30, 39]. Nevertheless, the accurate prediction based on only preoperative information (such as image and tumor marker indexes) is still challenging, possibly due to missing information from early diagnosis stage to the final stage. This missing information includes the texture and pathological attributes of the liver, health level of the patient, and post-operative treatment [9, 17, 20]. For example, as in Fig. 1, patient A and B with different OS time have almost identical preoperative indicators cause, history of disease, etc., making the preoperative-based model hard for discrimination.", + "bbox": [ + 75, + 429, + 472, + 641 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To amend this problem, we present a causal inference system that can well utilize the intraoperative information, which is pretty easy to be accessed in training data. According to medical priorities, such information records pathological attributes, which can be more reflective about the patient's health level and the postoperative recovery. Inspired by this, we propose to leverage this auxiliary information to help build our causal inference system. Again, we take the example in Fig. 1. Although preoperative information cannot differentiate patient A from B, their intraoperative index shows great difference in distribution, which can thus be employed for the discrimination. Additionally, there are many indicators from medical experts that the intraoperative indexes are related to each other. For instance, the clinicopathologic hepatocirrhosis is dependent on the hepatocirrhosis; the sum of tumor diameter is affected by the number of tumors; the fibrosis (reflected on S-Score) can prob", + "bbox": [ + 75, + 643, + 472, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "ably deteriorate to cirrhosis [4], etc. By leveraging these relationships, we can better understand the causal inference concrete from both observed data modelling and medical expert-level knowledge of these variables.", + "bbox": [ + 496, + 90, + 890, + 151 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To this end, we encapsulate these priors and the inspired proposals into a new method, dubbed as Causally-Aware Intraoperative Imputation Model (CAWIM). It incorporates causal discovery module to sequentially estimate intraoperative indexes as an intermediate stage towards final OS time prediction. Specifically, our model is composed of two key steps: i) estimating the intraoperative indexes using preoperative information, i.e., image and indexes; ii) followed by OS time prediction using estimated intraoperative indexes and preoperative information. To achieve more accurate prediction of intraoperative indexes that is determinant to the prediction power of the whole method, we propose a Causally-aware Directed Acyclic Graph (CaDAG) module. It learns the causal structure represented as a DAG over intraoperative features. To identify the causal relations beyond the traditional PC algorithm [26], we propose a splitting-voting mechanism, which is inspired by the recent work [18] that learn the causal structure with the assistance of an auxiliary domain index variable. Our proposed mechanism can not only identify the causal relations even when this domain index variable is not available, but also can be theoretically guaranteed that the learned graph is not acyclic. During test stage, we sequentially estimate each intraoperative index with preoperative information and additionally, its parent set among other intraoperative indexes. The utility of our method can be demonstrated by a significant improvement of OS time prediction on an in-house liver cancer dataset, as well as better interpretability of learned causal structure, more accurate estimation of intraoperative indexes and more interpretable visualization results.", + "bbox": [ + 496, + 155, + 892, + 625 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In a nutshell, we for the first time present a case study of building a causally-aware intraoperative imputation system for the challenging task of overall survival time prediction. The proposed method of building the casual inference system, can be naturally extended to other similar medical tasks. Our key contributions are listed as follows. (1) New Paradigm for OS time Prediction. We propose to leverage intraoperative indexes as an intermediate stage during training. The leverage of this information can significantly alleviate the \"missing information\" issue. To the best of our knowledge, we are the first to leverage auxiliary information (in addition to preoperative features) for OS time prediction. (2) Causal Structure Learning. We propose a novel splitting-voting mechanism that can identify the causal structure even when the domain index variable is missing. (3) Better Prediction Power. Our method can significantly improve the prediction power for liver cancer over the competitors. The methods are evaluated on the medical", + "bbox": [ + 496, + 628, + 893, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "15682", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "dataset, which, to the best of our knowledge, is the largest primary live cancer dataset.", + "bbox": [ + 76, + 90, + 468, + 121 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2. Related Work and Preliminaries", + "text_level": 1, + "bbox": [ + 76, + 135, + 372, + 150 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Previous efforts [23,39] directly take the OS time prediction as the vanilla classification task by employing various deep models, whilst we reformulate it by causal discovery method for the first time. We will review the most related literature here.", + "bbox": [ + 76, + 160, + 467, + 234 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "MRI based deep models. OS time is defined as the duration from patient's first scan to cancer-related death [6]. Since Magnetic Resonance Image (MRI) plays an important role in tumor-related studies, various MRI based deep learning model have been proposed for feature extraction and representing [28], multi-modal feature fusion [10, 30, 38, 39], and improvement on model structure [5]. However, since missing information during disease progression can mediate the correlation from the preoperative stage to the final stage, the results of these image based methods are typically unsatisfied. In contrast, our method can amend this problem by leveraging intraoperative indexes as a bridge.", + "bbox": [ + 75, + 236, + 467, + 416 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Causal discovery. Causal discovery is to learn the causal graph over endogenous variables. Typical methods include PC algorithm [19], a constraint-based method that implement iterative conditional independence test to identify the skeleton and directions via v-structure. Under Markovian and faithfulness assumptions [26, 31], this method is provable to identify an equivalent class of the causal graph. To further identify more directions, recent work [18] leveraged a domain index variable that splits the dataset into multiple domains. With these heterogeneous data, they can identify mutable variables that change across the domain index variable. On the basis of this, they further identified edges among these mutable variable set by implementing Hilbert Schmidt Independence Criterion (HSIC) norm [13]. However, such domain index may not be available in many scenarios, making the identification of these directions difficult. Besides, this method may not ensure that the learned graph is acyclic. To address these issues, we in this paper propose a splitting-voting mechanism to mine the causal relations among intraoperative indexes so as for leveraging each other to better estimate these indexes, which is provable to return a directed acyclic graph.", + "bbox": [ + 75, + 417, + 467, + 748 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Preliminaries. (1) Causal Discovery. By faithfulness and Markovian assumptions [26], we can use conditional independent tests for causal discovery. That is learning the causal graph over $\\mathbf{V}$ . (2) Structural causal model (SCM). Before fully developing our contributions in methodology, we give as preliminary the definition of SCM, which is defined as $\\langle G, \\mathcal{F}, P\\langle \\epsilon \\rangle \\rangle$ : i) directed acycle graph $G = (\\mathbf{V}, \\mathbf{E})$ is as causal structure, with node set $\\mathbf{V}$ and edge set $\\mathbf{E}$ ; ii) the autonomous structural functions $\\mathcal{F} = \\{f_k\\}_{V_k \\in V}$ , where disturbance on $V_k$ has no effect on others; iii) probability", + "bbox": [ + 75, + 750, + 467, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "measure $P(\\epsilon)$ for exogenous variables $\\{\\epsilon_k\\}_k$ . Given the assumption that $\\{\\epsilon_k\\}_k$ are independent, each $\\mathrm{P}$ can be obtained by Causal Markov Condition with $G$ as $\\mathrm{P}(\\{V_K = v_k\\}_{V_k \\in \\mathbf{V}}) = \\prod_k \\mathrm{P}(V_k = v_k | Pa(k) = pa(k))$ . (3) Causal discovery from heterogeneity. In [18], the authors proposed to leverage a domain index variable $C$ to identify some causal directions. This method is built upon the PC and faithfulness assumption at a distributional level. It first learns the skeleton of the graph and identifies a variable set that is affected by $E$ , followed by identifying the directions between $E$ and its neighbors. Specifically, if $C \\to V_i, C \\to V_j$ , and $V_i \\bot V_j | X, C$ for some deconfounding set $X$ between $V_i$ and $V_j$ , $\\{P(V_i | X, C = c)\\}$ is independent of $\\{P(V_j | V_i, C = c, X)\\}$ , while $\\{P(V_j | X, C = c)\\}$ and $\\{P(V_j | V_i, C = c, X)\\}$ are dependent. We can test these Independence using Hilbert Schmidt Independence Criterion (HSIC) [12] norms $\\delta_{v_i \\to v_j | X}$ and $\\delta_{v_j \\to v_i | X}$ . One can obtain that $V_i \\to V_j$ if $\\delta_{v_i \\to V_j} < \\alpha$ with pre-set significance level $\\alpha$ . However, in many applications such domain index variable is missing, i.e., all data are pooled together. In this paper, we propose a splitting-voting mechanism (in the CaDAG module) to identify causal relations. In the following, we first introduce some basic assumptions that our method is built upon.", + "bbox": [ + 496, + 90, + 890, + 454 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Methodology", + "text_level": 1, + "bbox": [ + 500, + 465, + 633, + 482 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Problem setting. For the OS time classification task, suppose we have $\\{X_{i},\\mathbf{B}_{i},\\mathbf{A}_{i},y_{i}\\} \\sim_{i.i.d}P(X,\\mathbf{B},\\mathbf{A},y)$ where $X$ denotes the image acquired from structural Magnetic Resonance Image (sMRI); $\\mathbf{B}$ , $\\mathbf{A}$ denote the preoperative and intraoperative attributes, respectively. We have the final label $y$ . Our goal is to predict $y$ from the image $X$ and preoperative attributes $\\mathbf{B}$ that are recorded before surgery. Our method is built upon the following three assumptions, commonly utilized in the causal inference works [18, 26].", + "bbox": [ + 496, + 489, + 890, + 626 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Definition 1 (Causal Graph) We assume the causal graph over $\\mathbf{A}$ is a directed acyclic graph (DAG) and denote the corresponding SCM as $M\\coloneqq \\langle G\\coloneqq (\\mathbf{A},\\mathbf{E}),\\mathcal{F},P(\\varepsilon)\\rangle$", + "bbox": [ + 496, + 633, + 890, + 681 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Definition 2 (Markovian and Faithfulness) For", + "text_level": 1, + "bbox": [ + 498, + 688, + 830, + 702 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "triplets of disjoint sets $\\mathbf{V}_i, \\mathbf{V}_j, \\mathbf{V}_k$ , it holds that $\\mathbf{V}_i \\perp_d \\mathbf{V}_j | \\mathbf{V}_k \\leftrightarrow \\mathbf{V}_i \\perp \\mathbf{V}_j | \\mathbf{V}_k$ , where $\\perp_d$ and $\\perp$ respectively mean $d$ -separation and probability independence. This is the common property in DAG.", + "bbox": [ + 496, + 703, + 890, + 763 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Definition 3 (Distributional Faithfulness) If $X_{i} \\to X_{j}$ and at least $E \\to V_{i}$ or $E \\to V_{j}$ holds, $\\{P^{e}(V_{i}|V_{j},\\mathbf{Z})\\}$ is dependent to $\\{P^{e}(V_{j}|\\mathbf{Z})\\}$ , where $\\mathbf{Z}$ denotes the minimal deconfounding set. Particularly, $\\mathbf{Z}$ is a deconfounding set between $V_{i}$ and $V_{j}$ if we have $V_{i} \\perp V_{j}|\\mathbf{Z}$ and $\\mathbf{Z} \\cap (\\mathrm{De}(V_i) \\cup \\mathrm{De}(V_j)) = \\emptyset$ .", + "bbox": [ + 496, + 771, + 890, + 862 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In term of this assumption, we further have the following result for edge orientation [18].", + "bbox": [ + 496, + 869, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "15683", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Theorem 4 (Theorem 2 in [18]) Denote $E$ as domain index variable. Under assumptions 1, 2, 3, for each adjacent pair $(V_i, V_j)$ such that $E \\to V_i$ or $E \\to V_j$ holds, we have $V_i \\to V_j$ if $\\{P^e(V_i|\\mathbf{Z})\\}$ and $\\{P^e(V_j|V_i,\\mathbf{Z})\\}$ are independent while $\\{P^e(V_j|\\mathbf{Z})\\}$ and $\\{P^e(V_i|V_j,\\mathbf{Z})\\}$ are dependent.", + "bbox": [ + 75, + 90, + 472, + 169 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1. Causally-aware Directed Acyclic Graph", + "text_level": 1, + "bbox": [ + 76, + 181, + 419, + 199 + ], + "page_idx": 3 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "The Causally-aware DAG. We propose a novel Causally-Aware Intraoperative Imputation Model (CAWIM). The pipeline of our method is shown in Fig. 2. It is composed of Causally-aware Directed Acycled Graph (CaDAG), and a classification model with MRI imaging encoder. The CaDAG enables the causally-aware intraoperative reasoning, by managing to better represent and encode the correlation of intraoperative information. Specifically, we first use preoperative information, i.e., MRI images and indexes to estimate the intraoperative features. The estimated intraoperative features $\\tilde{\\mathbf{A}}$ , together with the MRI $X$ and preoperative features $\\mathbf{B}$ are then utilized for OS classification. To accurately estimate intraoperative indexes $\\mathbf{A}$ , we propose CaDAG to learn causal structure over $\\mathbf{A}$ , and then leverage the parent nodes of each index for sequentially imputation during test stage. We give the details about the CaDAG and the intraoperative imputation model for classification, while some theoretical proofs, and the specification of our networks are in the appendix.", + "We aim at learning the causal structure over $\\mathbf{A}$ . However, without the domain index, we actually have the difficulty of dividing the whole dataset into multiple domains. To this end, we propose a splitting-voting mechanism for this problem. Particularly, this mechanism first generates the domain-index variable $E$ by randomly splitting the whole dataset into multiple domains for $m > 0$ times. Then for each time we identify the mutable variable set and generate a prediction for each identifiable causal directions, with the assistance of domain index variable $E$ . Finally, we vote for a final direction among $m$ predictions for each pair of variables. As shown in Fig. 3, the whole procedure is a sequence of the four steps. Step i), ii) and iii) provide preliminary causal relationship between each pair of nodes. Step iv) is our proposed splitting-voting, aiming to finally determine all the causal directions and output a causal DAG.", + "i) Learning skeleton of the causal graph. We first implement PC algorithm to learn the skeleton [26]. That is an undirected graph over all attributes. This can be achieved by iterative conditional independence test. Then we determine $A_{i}$ and $A_{j}$ to be adjacent if they are not independent conditional on any subsets. Finally, we utilize the v-structure to learn the equivalence class of DAG.", + "ii) Random splitting and identification of mutable variable sets. To determine more directions, we randomly split the dataset into multiple domains for $m$ times. For each time we have a domain index variable $E_{i}$ and multiple do" + ], + "bbox": [ + 75, + 205, + 472, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "mains $\\{\\mathcal{D}_e^i\\}$ [18]; and we identify the mutable variable set $\\mathbf{M}_i$ such that each $X \\in \\mathbf{M}_i$ has $X \\not\\perp E_i|\\mathbf{A} - X$ is conditionally dependent to $E_i$ . We thus can implement conditional independent tests to identify $\\mathbf{M}_i$ . Steps i) and ii) are summarized in Alg. 1.", + "bbox": [ + 496, + 90, + 893, + 167 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Algorithm 1 Identify the skeleton and mutable set.", + "text_level": 1, + "bbox": [ + 500, + 180, + 836, + 196 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "INPUT: $\\{\\mathcal{D}_e^i\\coloneqq \\{\\mathbf{A}_i^e\\}_{i = 1}^{n_e}|e\\}$ , domain index variable $E_{i}$ .", + "bbox": [ + 500, + 210, + 875, + 227 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "OUTPUT: Equivalent class of DAG and $\\mathbf{M}_i$ .", + "bbox": [ + 500, + 227, + 802, + 243 + ], + "page_idx": 3 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1: Implement the PC algorithm to learn the equivalence class of DAG via v-structure.", + "2: For each $A \\in \\mathbf{A}$ , add the edge $E_i \\to A$ in the graph iff $E_i \\not\\perp A$ given any subsets." + ], + "bbox": [ + 508, + 255, + 890, + 315 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "iii) Identifying directions via changed causal models. For each adjacent pair $A_{i}$ and $A_{j}$ such that at least one of them is adjacent to $E$ , we identify the causal directions among $\\mathbf{M}_i$ via HSIC Norm [12] by testing the independence between $\\{P^{e}(A_{i}|A_{j},\\mathbf{Z})\\}$ and $\\{P^{e}(A_{j}|\\mathbf{Z})\\}$ if $A_{i}\\gets A_{j}$ , and that between $\\{P^{e}(A_{j}|A_{i},C)\\}$ and $\\{P^{e}(A_{i}|\\mathbf{Z})\\}$ for determining the direction between $A_{i}$ and $A_{j}$ . Here, $\\mathbf{Z}$ denotes the deconfounding set. In this regard, the direction between $A_{i}$ and $A_{j}$ is then determined according to the following rule: if $\\hat{\\Delta}_{A_i\\to A_j|\\mathbf{Z},E} < \\hat{\\Delta}_{A_j\\to A_i|C,E}$ , output $A_{i}\\rightarrow A_{j}$ ; if $\\hat{\\Delta}_{A_i\\to A_j|\\mathbf{Z},E} > \\hat{\\Delta}_{A_j\\to A_i|\\mathbf{Z},E}$ , output $A_{i}\\leftarrow A_{j}$ ; if $\\hat{\\Delta}_{A_i\\to A_j|\\mathbf{Z},E} > \\alpha$ and $\\hat{\\Delta}_{A_j\\to A_i|\\mathbf{Z},E} > \\alpha$ , where $\\alpha$ is a threshold, making the direction undetermined.", + "bbox": [ + 496, + 335, + 893, + 540 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Algorithm 2 Identify causal directions among M.", + "text_level": 1, + "bbox": [ + 500, + 553, + 833, + 569 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "INPUT: $E$ ; skeleton and $M$ via Alg.1.", + "bbox": [ + 500, + 582, + 756, + 597 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "OUTPUT: Directed graph among M.", + "bbox": [ + 500, + 598, + 751, + 614 + ], + "page_idx": 3 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1: For each adjacent $(A_{i},A_{j})$ such that one of $V_{i}\\in \\mathbf{M}$", + "2: Detect deconfounding set $\\mathbf{Z}$", + "3: Calculate $\\hat{\\Delta}_{A_i\\to A_j|\\mathbf{Z},E} > \\alpha$ and $\\hat{\\Delta}_{A_j\\to A_i|\\mathbf{Z},E} > \\alpha$", + "4: Determine $A_{i} \\to A_{j}$ or $A_{j} \\to A_{i}$ ." + ], + "bbox": [ + 508, + 625, + 883, + 686 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "iv) Voting for the direction of each edge. For each time $i$ we implement Alg. 2 to predict an direction of the edge $(A_{i}, A_{j})$ . Then we vote for $A_{i} \\rightarrow A_{j}$ if the frequency $f_{i \\rightarrow j}$ of $A_{i} \\rightarrow A_{j}$ surpasses a pre-set threshold $\\alpha$ ( $0 < \\alpha < 1$ ). Finally, if there exists a cycle of $N$ variables: $A_{i_1} \\rightarrow A_{i_2} \\rightarrow \\ldots \\rightarrow A_{i_1}$ , we orient $(A_{i_k}, A_{i_{k+1}})$ with the minimum frequency among $\\{f_{i_k, i_{k+1}}, f_{i_N, i_1}\\}$ .", + "bbox": [ + 496, + 707, + 890, + 815 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We show that this procedure can generate a DAG over $\\mathbf{A}$ , with the proof in appendix.", + "bbox": [ + 500, + 828, + 890, + 859 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Theorem 5 Under definitions and assumptions 1, 2, 3, the learned graph via our CaDAG is a directed acyclic graph.", + "bbox": [ + 496, + 869, + 890, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "15684", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/084fd24618bb2d6424aa8db962c2987f2924f680a8e6a27dc2f13c2f73f95380.jpg", + "image_caption": [ + "Figure 2. Overview of our CAWIM model. We concatenate the MRI features encoded by the MRI Imaging Encoder, preoperative indicators, and the estimated intraoperative indicators obtained from CaDAG to perform the final prediction on OS time." + ], + "image_footnote": [], + "bbox": [ + 81, + 85, + 883, + 220 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/c6333c92c34b3ea843d0f41e1738c5c833cd42cf5b4bbda28713c5d270735f4e.jpg", + "image_caption": [ + "Figure 3. Schematic diagram of a four-variable case $A_{1}$ , $A_{2}$ , $A_{3}$ , and $A_{4}$ , with the real DAG shown in the top left corner. Through step i) and ii) in the upper part, a Markov Equivalence Class of 3 causal graphs is observed with the direction between $A_{1}$ and $A_{2}$ undetermined. For step iii), we split the dataset into $(n = 3)$ domains and consider the 4 circumstances shown fig(a) to (d) in the lower left part. For step iv), we vote for the direction between $A_{1}$ and $A_{2}$ as $A_{1} \\rightarrow A_{2}$ according to the number of splitting $(n = 3)$ , frequency $f_{1\\rightarrow 2} = 2 / 3$ and threshold $\\alpha = 2 / 3$ , as shown in the lower right part. Finally, we obtain an inferred DAG with each direction determined." + ], + "image_footnote": [], + "bbox": [ + 117, + 255, + 861, + 468 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2. Intraoperative Imputation for Classification", + "text_level": 1, + "bbox": [ + 76, + 549, + 449, + 565 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Given MRI image $X \\in \\mathbb{R}^{H \\times W}$ of a patient that contains multiple slides, we first resample these slides into a fixed $K$ slides for each patient. As each slice is a gray image, we repeat the images for 3 times along the channel dimension; and then we employ a neural network to extract features from each slice. The extracted features are denoted as $F_{1}, \\ldots, F_{K}$ . In order to aggregate the information of all slices, we divide $F_{1}, \\ldots, F_{K}$ into $g$ groups according to the acquired position. Each group contains consecutive $\\lfloor K / g \\rfloor$ slices. Then, we implement a Multi-head Cross-Attention (MCA) mechanism inside each group to aggregate features:", + "bbox": [ + 75, + 571, + 468, + 739 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {h e a d} _ {j} = \\operatorname {s o f t m a x} \\left(Q ^ {j} \\left(F \\mathbf {W} _ {K} ^ {j}\\right) ^ {T}\\right) \\left(F \\mathbf {W} _ {V} ^ {j}\\right) \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 137, + 747, + 468, + 766 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {M C A} = \\left[ \\text {h e a d} _ {1}; \\dots ; \\text {h e a d} _ {h} \\right] \\mathbf {W} _ {O} \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 163, + 773, + 468, + 789 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $F$ is the slice's feature within a group, the query vector $Q^{j}$ and linear projection layers $\\mathbf{W}_K^j$ , $\\mathbf{W}_V^j$ , $\\mathbf{W}_O$ , $1 \\leq j \\leq h$ are all learnable parameters. Finally, we concatenate $\\{F_k\\}$ from $g$ groups as our final image feature $I \\in \\mathbb{R}^{g \\times d}$ with $d$ denoting the dimension of each $F_k$ with $k = 1, \\ldots, K$ .", + "bbox": [ + 75, + 794, + 468, + 869 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Afterwards, we sequentially estimate each $A_{j}$ by preoperative information, as our learned causal graph over $\\mathbf{A}$ . So", + "bbox": [ + 75, + 869, + 468, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "for each $A_{j}$ , we predict", + "bbox": [ + 500, + 549, + 658, + 565 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {A} _ {j} = f _ {\\text {i n t r o}} ^ {j} (\\mathbf {B}, I, \\operatorname {P a} (A _ {j})) \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 609, + 578, + 890, + 597 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathrm{Pa}(A_j)$ denotes the parent node set of $A_{j}$ . We use cross entropy loss to optimize $f_{\\mathrm{intra}}^j$ if $A_{j}$ is categorical; and optimize $f_{\\mathrm{intra}}^j$ if $A_{j}$ is continuous. After obtaining all predicted $\\{\\hat{A}_i\\}_{i = 1}^d$ , we use $(I,\\mathbf{B},\\hat{\\mathbf{A}})$ to predict $Y$ via $f_{\\theta}(I,\\mathbf{B},\\hat{\\mathbf{A}})$ and obtain $\\theta$ by optimizing the cross-entropy $\\mathcal{L}(y,f_{\\theta}(I,\\mathbf{B},\\hat{\\mathbf{A}}))$ . During test stage, given a new sample $(x,\\mathbf{b})$ , we first extract $I$ and then sequentially estimate $\\mathbf{a}_i$ and finally predict $y$ using $f_{\\theta}(x,\\mathbf{b},\\hat{\\mathbf{a}})$ .", + "bbox": [ + 496, + 609, + 890, + 739 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 500, + 753, + 632, + 770 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Dataset overview. Ethics Committee of Zhongshan Hospital, Fudan University approved the protocol of this study and waived the requirement for patient-informed consent (B2021-325R). Because open source data sets generally lack sufficient intraoperative information, it takes us around ten years to collect and build our own medical dataset. Particularly, we conduct a search through the medical records in the hospital information system, and build a dataset with", + "bbox": [ + 496, + 779, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "15685", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "439 patients infected by primary liver cancer. Long-term medical data of every patient has been recorded from the first scan to the cancer-related death. Till the submission of this manuscript, around $90\\%$ patients have unfortunately passed away. This gives us the ground-truth OS labels. Information like MRI scanning before the surgery, blood test and operation reports, follow-up records are also collected. After removing samples with missing information, 361 patients were enrolled, with 306 $(84.8\\%)$ men, and the average age was 53.7 $(\\pm 11.7)$ years. Specifically, each patient has MRI scanning with four modalities, including $\\mathrm{T}_{1}$ weighted images $(\\mathrm{T}_{1}$ -WI), $\\mathrm{T}_{2}$ weighted images $(\\mathrm{T}_{2}$ -WI), $\\mathrm{T}_{1}$ in-phase $(\\mathrm{T}_{1}$ -IP) and contrast-enhanced $\\mathrm{T}_{1}$ at the arterial phase $(\\mathrm{T}_{1}$ ce-AP). Following the guidance of medical experts, we select 11 preoperative and 20 intraoperative indicators that are related to OS time for our study. Please refer to the appendix for details.", + "bbox": [ + 75, + 90, + 472, + 347 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Data preprocessing. In our dataset, the survival time ranges from 0 to 130 months, which are cauterized into four classes: short-term survival ( $\\leq 36$ months), middle-short-term survival (between 36 and 72 months), middle-long-term survival (between 72 and 108 months), and long-term survival ( $\\geq 108$ months). In our paper, we consider the classification task: given a new sample, our goal is to classify which class of OS time this sample belongs to. To extract $I$ from MRI image, we resample $K = 20$ slices and employ Resnet-34 for feature extraction. We set $g = 4$ and the dimension $F_{k}$ as 512. For $f_{\\mathrm{intro}}^{j}$ and $f_{\\theta}$ , we employ a multi-layer perceptron (MLP) for parameterization.", + "bbox": [ + 75, + 349, + 472, + 532 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We encode categorical covariates into dummy variables and implement zero-mean normalization for continuous variables. For MRI images, we resize them into $256 \\times 256$ as input. We split the whole dataset into 5 folds. To remove the effect of imbalance across classes and randomness, we adopt average precision, recall, and $\\mathrm{F_1}$ -Score measurements over five folds for classification evaluation.", + "bbox": [ + 75, + 532, + 472, + 638 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation details. We use the ImageNet pretrained Resnet-34 to initialize our CNN encoder. We train 150 epochs for all methods with $\\mathrm{T}_{1}$ ce-AP modality, using SGD as the optimizer with learning rate $3e - 3$ and $3e - 2$ for CNN encoder and later MLP, respectively. We set weight decay to $5e - 4$ and nesterov momentum factor to 0.9. Besides, we decrease the learning rate every 60 epochs by a factor of 1/10. We set the Batch size to 8. In our voting mechanism of CaDAG, we in total implement Alg. 2 for seven times and set the voting threshold as 4/7. Our model is trained with Pytorch on NVIDIA GeForce RTX 3090 GPUs.", + "bbox": [ + 75, + 641, + 472, + 808 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Compared Methods. We compare our method with four baseline methods: i) Random: random guess among four classes as a trivial baseline, of which the expected $\\mathrm{F_1}$ score is $25\\%$ ; ii) MRI: train the end-to-end method with $\\mathrm{T_1}$ ce-AP only as input for classification; iii) MRI+pre: additionally take preoperative features as input, on the basis of the", + "bbox": [ + 75, + 809, + 472, + 902 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/38acba7d80a01d78bdbf124f3e4f7946bb4a5a2d4d3c0fe63ac7a1b5d87dad43.jpg", + "table_caption": [ + "Table 1. Comparisons of Our CAWIM with Other Baselines." + ], + "table_footnote": [], + "table_body": "
ModelPrecisionRecallF1-Score
(a)Random26.04±2.1024.54±0.8023.15±0.64
(b)MRI40.79±7.9733.45±2.6631.88±3.92
(c)MRI+Pre36.99±5.3035.62±3.5033.83±3.64
(d)gt Intra41.26±2.7441.62±1.9539.77±2.47
(e)CAWIM45.58±5.8643.70±5.8942.21±4.92
", + "bbox": [ + 501, + 104, + 887, + 196 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "method MRI; iv) ground-truth (GT) Intra: only take the ground-truth intraoperative features as input. It is worth to mention that although the intraoperative features are not allowed to use during test stage, the comparison with method iv (i.e., GT Intra) can provide information that to which extent, the intraoperative features can help OS time prediction.", + "bbox": [ + 496, + 209, + 890, + 303 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Results and Analysis", + "text_level": 1, + "bbox": [ + 498, + 311, + 696, + 328 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1 records comparisons of our proposed CAWIM (Tab. 1(d)) with other baseline models. (Tab. 1(a), Tab. 1(b), and Tab. 1(c)). Obviously, our CAWIM constantly outperform others in all metrics. Specifically, CAWIM improved $\\text{MRI} + \\text{Pre}$ by nearly $10\\%$ in terms of $\\text{F}_1$ -Score, which can be contributed to the prediction of intraoperative features equipped with causally aware prediction mechanism. Without this mechanism, using MRI and preoperative features only can outperform random guess by only $8.33\\%$ , which demonstrates the limited information of the image for OS time. In contrast, it is also interesting to note from Tab. 1(d) that using intraoperative features can be more informative than MRI features, in terms of OS prediction, which is natural as the attributes that described the functions and textures of the liver could affect the postoperative recovery and thus the OS time. Finally, our CAWIM (Tab. 1(e)) that combines the information of intraoperative features and the information from images, can achieve further improvement.", + "bbox": [ + 496, + 335, + 890, + 609 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Ablation Study", + "text_level": 1, + "bbox": [ + 498, + 619, + 653, + 635 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To test the effectiveness of preoperative features, the particularly intraoperative features, and our causally-aware module for OS time prediction, we implement ablation studies and summarize the results in Tab. 2. Specifically, Tab. 2(a), i.e., w/o CaDAG replaces our causally-aware module with an end-to-end estimation network from (MRI, preoperative features) to intraoperative features; Tab. 2(b), i.e., w/o Intra is the same to Tab. 1(c), i.e., MRI+pre in Tab. 1 that directly predict OS time using MRI and preoperative features without estimated intraoperative features; Tab. 2(c), i.e., w/o Pre is the same to our CAWIM except that it does not utilize preoperative features for intraoperative estimation and OS prediction.", + "bbox": [ + 496, + 642, + 890, + 839 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The results in Tab. 2 show that the deletion of each module can lead to a descent of all metrics. Specifically, Tab. 2(a) shows a significant performance drop by $5.6\\%$ in $\\mathrm{F_1}$ -Score, which indicates the effectiveness of our causally", + "bbox": [ + 496, + 839, + 893, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "15686", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "aware module in estimating intraoperative features in OS time prediction. This phenomena can be explained by more accurate intraoperative estimation equipped with our causally-aware module, as shown in Fig. 4 of $\\mathrm{F_1}$ -score, accuracy and $R$ -squared score. We will leave detailed discussions in Section of \"Further Discussion\". Tab. 2(b), as the same to MRI+pre in Tab. 1, shows a $8.3\\%$ performance drop, which indicates the additional information provided by the MRI image and preoperative features. Finally, the $3.78\\%$ degradation of Tab. 2(c) compared to ours validates the effectiveness of preoperative features in OS prediction, which can also be explained by the estimation results of intraoperative features in Fig. 4. In a word, we can observe a significant drop of $F_{1}$ -score if we remove causally-aware module and intraoperative features, which can demonstrate the effectiveness of our methods.", + "bbox": [ + 75, + 90, + 472, + 330 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Besides, it is also interesting to compare Tab. 2(b) and Tab. 2(c) with we observe that the performance drop for $w/o$ Intra is more significant than that for $w/o$ Pre, with respectively an average of $8.4\\%$ and $3.8\\%$ drop in $F_1$ -Score. This again, indicates more information of intraoperative indexes provided than preoperative indexes in predicting OS time. Indeed, we obtained Precision $27.68\\%$ , Recall $26.65\\%$ and F1-Score $23.85\\%$ if we only predicted OS time using preoperative features, which is comparable to the result from random guess. On the other hand, we can observe from Fig. 4 that preoperative indexes are beneficial to the estimation of intraoperative features, which indicating the necessity of using intraoperative features as a bridge between preoperative information and final OS time prediction.", + "bbox": [ + 75, + 333, + 472, + 545 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Medical Interpretation of Causal Graph", + "text_level": 1, + "bbox": [ + 76, + 558, + 421, + 574 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To further explain the effectiveness of our causally-aware module, we in this subsection present our learned causal graph over intraoperative features in Fig. 5 and the corresponding medical interpretation of learned causal relations. In general, most of the parent-child relationships in our CaDAG are in line with common sense or can be explained by prior medical knowledge, e.g., sum of tumor diameter is determined by number of tumors, the edge from $G$ -score to $S$ -score also correspond to the strong correlation between fibrosis and hepatitis (please refer to appendix for detailed descriptions of intraoperative indexes).", + "bbox": [ + 75, + 582, + 468, + 748 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Further, those indicators with non-visual features that access severity of pathology are often determined by other attributes of the liver; therefore, these features often come up in deeper layers in Fig. 5. For instance, $G$ -score and $S$ -score reflect the level of inflammation and fibrosis that are artificially defined according to Metavir scoring system [2]. Those indicators tend to rely on other intraoperative features besides merely MRI and preoperative information, e.g. clinicopathologic hepatocirrhosis is reasoned from hepatocirrhosis and other covariants. For this reason, the addi", + "bbox": [ + 75, + 750, + 470, + 900 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/fc83786945b1467c438a851b6f2288ead83388d9ffb12c0eff6fb633166358c2.jpg", + "table_caption": [ + "Table 2. Results of Ablation Studies. (a) estimates each intraoperative features naively from (MRI, preoperative indexes), without the causally-aware module. (b) is the same to $M R I + P r e$ in Tab. 1. (c) is the CAWIM without preoperative features for intraoperative indexes estimation and OS time prediction." + ], + "table_footnote": [], + "table_body": "
ModelPrecisionRecallF1-Score
(a)w/o CaDAG36.98±2.3837.78±1.9136.62±1.87
(b)w/o Intra36.99±5.3035.62±3.5033.83±3.64
(c)w/o Pre38.92±7.7141.57±4.5338.43±5.84
(d)CAWIM45.58±5.8643.70±5.8942.21±4.92
", + "bbox": [ + 501, + 160, + 903, + 238 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "tional leverage of hepatocirrhosis lead to more accurate estimation of clinicopathologic hepatocirrhosis, as shown in Fig. 4 (a). On the other hand, we also observe some associations that cannot be explained well, e.g. the causal relationship between sum of tumor diameter and cell type, which may due to the existence of unobserved confounders.", + "bbox": [ + 496, + 248, + 893, + 339 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Visualization", + "text_level": 1, + "bbox": [ + 500, + 347, + 637, + 361 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In this section, we visualize the high-response area of our CAWIM Fig. 6(c) and the version without CaDAG Fig. 6(b) using Grad-CAM [29]. As shown in Fig. 6, the detected regions of our method can be more concentrated on the liver. This result can be explained with Fig.4 and Fig. 5 in a more complementary way. Specifically, as the learned causal graph is medical explainable and with better estimation of intraoperative features, the model is driven to locate on liver-related regions. More visualization results are left in appendix due to space limit.", + "bbox": [ + 496, + 369, + 893, + 521 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.5. Further Discussion", + "text_level": 1, + "bbox": [ + 500, + 530, + 683, + 544 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Intraoperative prediction. In our method, the estimation of intraoperative indicators are important for final prediction. Fig. 4 shows the metrics of some typical intraoperative prediction by i) $w/o \\, CaDAG$ and ii) $w/o \\, Pre$ with $F_1$ -Score and accuracy for discrete variants (Fig. 4 (a, b)) and $R$ -square for continuous ones (Fig. 4 (c)). Fig. 4 (a, b) shows that our CaDAG can generally enhance the estimation of categorical variables hepatocirrhosis, cell type, differentiation, etc. Nevertheless, we also find that such improvements will decrease if corresponding indicators are in deep layers of the graph in Fig. 5. For example, the accuracy of $G$ -Score and $S$ -score is only comparable and even worse than that of $w/o \\, CaDAG$ . A possible reason is the accumulated error along the directed path for estimation. On the other hand, as shown by the results of $R$ -square (closer to 1, the better the prediction) does not bring about notable improvement in regression tasks, (e.g., ascites and tumor diameter).", + "bbox": [ + 496, + 553, + 893, + 809 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We also observe an evident performance degradation of $w / o$ Pre for most of those discrete or continuous variants. This implies that although operative alone has limited contribution to OS time prediction, it still matters in our CAWIM, because of the less missing information due to shorter time interval from the preoperative stage to the", + "bbox": [ + 496, + 810, + 893, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "15687", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/01d26054c451b1c255ec9bdb018d00e30cb4ff489eaf6f6bc84a3c82a4a2f274.jpg", + "image_caption": [ + "(a)" + ], + "image_footnote": [], + "bbox": [ + 60, + 98, + 330, + 236 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/f3d9ee8a3d02e3efb04474ac8234a698f83a470aadc7468a30acd9ec12c36a4c.jpg", + "image_caption": [ + "(b)", + "Figure 4. Quantity Results of Intraoperative Prediction. Intraoperative prediction are categorized into classification ((a) and (b)) and regression problem ((c)) according to the data form. (a) and (b) shows the prediction $\\mathrm{F}_1$ -Score and accuracy rate of the discrete variants respectively. The numbers after the variant name in (a) and (b) are the number of classes, i.e. hepatocirrhosis is a 6-category classification task. (c) shows the R-squared score for continuous variants." + ], + "image_footnote": [], + "bbox": [ + 343, + 99, + 620, + 236 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/fd6aa08290bb102b2db018cf78102e9766530d0ae8a7d21ec1a482407f4086db.jpg", + "image_caption": [ + "(c)" + ], + "image_footnote": [], + "bbox": [ + 635, + 99, + 880, + 236 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/a4777bdbafa993ada2bb5a5b5a1f855219cff6f6c8b8d872dc8b1be1d363964e.jpg", + "image_caption": [ + "Figure 5. Learned DAG over Intraoperative Indexes A." + ], + "image_footnote": [], + "bbox": [ + 107, + 330, + 433, + 465 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/7d7d315d9ce95c101dc90f817294af7392c4bb7fad917562d9e90d5d6976aaa8.jpg", + "image_caption": [ + "Figure 6. Heat Maps via Grad-CAM [29] on 5 Patients. (a) shows original images and the bounding box of liver. (b) and (c) shows the area of interest $w/o$ and with CaDAG, respectively." + ], + "image_footnote": [], + "bbox": [ + 86, + 492, + 433, + 655 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "surgery stage than to the final stage. This observation validates the effect of using intraoperative features.", + "bbox": [ + 75, + 717, + 468, + 747 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "General Framework for OS time prediction. In general, predicting OS time from preoperative images and indexes is medically important but challenging, as there exists a lot of missing information due to a long time interval between the preoperative stage to the final OS stage. To amend this problem, it is informative to leverage some intermediate information. In this paper, we show that with this leverage and the modeling of this information via causal discovery, our CAWIM model enjoys convincing medical interpretability, more concentrated location of liver region, more accurate", + "bbox": [ + 75, + 750, + 472, + 901 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "estimation of intraoperative features, and finally better classification results on OS time. We thus believe that our framework can be beneficial to OS time prediction in other scenarios, e.g., other diseases or intermediate information in addition to intraoperative indexes.", + "bbox": [ + 496, + 325, + 892, + 400 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Limitations. We find in Fig. 5 that although most of learned causal relations are consistent with medical priors, there exist relations that cannot be explained well: it is hard to determine the causal order for some pairs of variables: e.g., the causal relationship between the sum of tumor diameter and cell type. This may due to unobserved confounders between these pairs of variables, which may be alleviated by learning hidden representations that can explain the associations among these pairs of variables. We will leave this exploration in our future work.", + "bbox": [ + 496, + 402, + 892, + 555 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 580, + 617, + 595 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We propose a novel OS time prediction paradigm, which is the first to leverage intraoperative attributes by causal structure learning. Our method significantly outperforms baselines by a large margin. We demonstrate that this improvement is contributed to highly interpretable learned causal structure, accurate estimation of intraoperative indexes, and identification of disease-related regions. We believe our method, especially the leverage of intraoperative information equipped with causal discovery can potentially benefit other scenarios. For limitation, we shall relax the SCM assumption by allowing the existence of unobserved confounders. We leave it as the future work.", + "bbox": [ + 496, + 609, + 892, + 790 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgment. Ethics Committee of Zhongshan Hospital, Fudan University approved the protocol of this study and waived the requirement for patient-informed consent (B2021-325R). The authors are also with Shanghai Key Lab of Intelligent Information Processing, and Fudan ISTBI-ZJNU Algorithm Centre for Brain-inspired Intelligence, Zhejiang Normal University, Jinhua, China.", + "bbox": [ + 496, + 794, + 893, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "15688", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Shekoofeh Azizi, Basil Mustafa, Fiona Ryan, Zachary Beaver, Jan Freyberg, Jonathan Deaton, Aaron Loh, Alan Karthikesalingam, Simon Kornblith, Ting Chen, et al. Big self-supervised models advance medical image classification. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3478-3488, 2021. 2", + "[2] Pierre Bedossa and Thierry Poynard. An algorithm for the grading of activity in chronic hepatitis c. Hepatology, 24(2):289-293, 1996. 7", + "[3] Nicholas A Christakis, Julia L Smith, Colin Murray Parkes, and Elizabeth B Lamont. Extent and determinants of error in doctors' prognoses in terminally ill patients: prospective cohort study commentary: Why do doctors overestimate? commentary: Prognoses should be based on proved indices not intuition. *Bmj*, 320(7233):469–473, 2000. 2", + "[4] Agostino Colli, Mirella Fraquelli, Marco Andreola, Barbara Marino, Enrico Zuccoli, and Dario Conte. Severe liver fibrosis or cirrhosis: accuracy of us for detection—analysis of 300 cases. Radiology, 227(1):89-94, 2003. 2", + "[5] Yin Dai, Yifan Gao, and Fayu Liu. Transmed: Transformers advance multi-modal medical image classification. Diagnostics, 11(8):1384, 2021. 2, 3", + "[6] James J Driscoll and Oliver Rixe. Overall survival: still the gold standard: why overall survival remains the definitive end point in cancer clinical trials. The Cancer Journal, 15(5):401-405, 2009. 2, 3", + "[7] J Faivre, D Forman, J Esteve, M Obradovic, M Sant, EURO-CARE Working Group, et al. Survival of patients with primary liver cancer, pancreatic cancer and biliary tract cancer in europe. European Journal of Cancer, 34(14):2184-2190, 1998. 2", + "[8] Ming Fan, Zuhui Liu, Sudan Xie, Maosheng Xu, Shiwei Wang, Xin Gao, and Lihua Li. Integration of dynamic contrast-enhanced magnetic resonance imaging and t2-weighted imaging radiomic features by a canonical correlation analysis-based feature fusion method to predict histological grade in ductal breast carcinoma. Physics in Medicine & Biology, 64(21):215001, 2019. 2", + "[9] Sheung Tat Fan, Chung Mau Lo, Ronnie TP Poon, Chun Yeung, Chi Leung Liu, Wai Key Yuen, Chi Ming Lam, Kelvin KC Ng, and See Ching Chan. Continuous improvement of survival outcomes of resection of hepatocellular carcinoma: a 20-year experience. Annals of surgery, 253(4):745-758, 2011. 2", + "[10] Xue Feng, Nicholas J Tustison, Sohil H Patel, and Craig H Meyer. Brain tumor segmentation using an ensemble of 3d u-nets and overall survival prediction using radiomic features. Frontiers in computational neuroscience, 14:25, 2020. 3", + "[11] Paul Glare, Kiran Virik, Mark Jones, Malcolm Hudson, Steffen Eychmuller, John Simes, and Nicholas Christakis. A systematic review of physicians' survival predictions in terminally ill cancer patients. *Bmj*, 327(7408):195, 2003. 2", + "[12] Arthur Gretton, Olivier Bousquet, Alex Smola, and Bernhard Schölkopf. Measuring statistical dependence with hilbert-schmidt norms. In International conference on algorithmic learning theory, pages 63-77. Springer, 2005. 3, 4" + ], + "bbox": [ + 78, + 114, + 470, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[13] Arthur Gretton, Kenji Fukumizu, Choon Teo, Le Song, Bernhard Scholkopf, and Alex Smola. A kernel statistical test of independence. Advances in neural information processing systems, 20, 2007. 3", + "[14] Lu Guo, Ping Wang, Ranran Sun, Chengwen Yang, Ning Zhang, Yu Guo, and Yuanming Feng. A fuzzy feature fusion method for auto-segmentation of gliomas with multimodality diffusion and perfusion magnetic resonance images in radiotherapy. Scientific reports, 8(1):1-11, 2018. 2", + "[15] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 1", + "[16] Mohammad Hesam Hesamian, Wenjing Jia, Xiangjian He, and Paul Kennedy. Deep learning techniques for medical image segmentation: achievements and challenges. Journal of digital imaging, 32(4):582-596, 2019. 2", + "[17] Karin W Houben and John L McCall. Liver transplantation for hepatocellular carcinoma in patients without underlying liver disease: a systematic review. *Liver Transplantation and Surgery*, 5(2):91–95, 1999. 2", + "[18] Biwei Huang, Kun Zhang, Jiji Zhang, Joseph D Ramsey, Ruben Sanchez-Romero, Clark Glymour, and Bernhard Scholkopf. Causal discovery from heterogeneous/nonstationary data. J. Mach. Learn. Res., 21(89):1-53, 2020. 2, 3, 4", + "[19] Markus Kalisch and Peter Buhlman. Estimating high-dimensional directed acyclic graphs with the pc-algorithm. Journal of Machine Learning Research, 8(3), 2007. 3", + "[20] Christophe Laurent, Jean Frédéric Blanc, Steele Nobili, Antonio Sa Cunha, Brigitte le Bail, Paulette Bioulac-Sage, Charles Balabaud, Maylis Capdepont, and Jean Saric. Prognostic factors and longterm survival after hepatic resection for hepatocellular carcinoma originating from noncirrhotic liver. Journal of the American College of Surgeons, 201(5):656-662, 2005. 2", + "[21] Qing Li, Weidong Cai, Xiaogang Wang, Yun Zhou, David Dagan Feng, and Mei Chen. Medical image classification with convolutional neural network. In 2014 13th international conference on control automation robotics & vision (ICARCV), pages 844-848. IEEE, 2014. 2", + "[22] Tao Li, Wu Li, Yehui Yang, and Wensheng Zhang. Classification of brain disease in magnetic resonance images using two-stage local feature fusion. PloS one, 12(2):e0171749, 2017. 2", + "[23] Dong Nie, Han Zhang, Ehsan Adeli, Luyan Liu, and Dinggang Shen. 3d deep learning for multi-modal imaging-guided survival time prediction of brain tumor patients. In International conference on medical image computing and computer-assisted intervention, pages 212-220. Springer, 2016. 3", + "[24] Byeonggwan Noh, Young Mok Park, Yujin Kwon, Chang In Choi, Byung Kwan Choi, Yo-Han Park, Kwangho Yang, Sunju Lee, Taeyoung Ha, YunKyong Hyon, et al. Machine learning-based survival rate prediction of korean hepatocellular carcinoma patients using multi-center data. BMC gastroenterology, 22(1):1-9, 2022. 2" + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "15689", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[25] Augustus Odena, Christopher Olah, and Jonathon Shlens. Conditional image synthesis with auxiliary classifier gans. In International conference on machine learning, pages 2642-2651. PMLR, 2017. 1", + "[26] Judea Pearl. Causality. Cambridge university press, 2009. 2, 3, 4", + "[27] Markus Peck-Radosavljevic. Drug therapy for advanced-stage liver cancer. Liver cancer, 3(2):125-131, 2014. 2", + "[28] Whitney B Pope, James Sayre, Alla Perlina, J Pablo Villablanca, Paul S Mischel, and Timothy F Cloughesy. Mr imaging correlates of survival in patients with high-grade gliomas. American Journal of Neuroradiology, 26(10):2466-2474, 2005. 3", + "[29] Ramprasaath R. Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra. Grad-cam: Visual explanations from deep networks via gradient-based localization. In 2017 IEEE International Conference on Computer Vision (ICCV), pages 618-626, 2017. 7, 8", + "[30] Wei Shao, Tongxin Wang, Liang Sun, Tianhan Dong, Zhi Han, Zhi Huang, Jie Zhang, Daoqiang Zhang, and Kun Huang. Multi-task multi-modal learning for joint diagnosis and prognosis of human cancers. Medical Image Analysis, 65:101795, 2020. 2, 3", + "[31] Pater Spirtes, Clark Glymour, Richard Scheines, Stuart Kauffman, Valerio Aimale, and Frank Wimberly. Constructing bayesian network models of gene expression networks from microarray data. 2000. 3", + "[32] Wen Tang, Haoyue Zhang, Pengxin Yu, Han Kang, and Rongguo Zhang. Mmmna-net for overall survival time prediction of brain tumor patients. arXiv preprint arXiv:2206.06267, 2022. 2", + "[33] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 1", + "[34] Jing-Houng Wang, Chi-Sin Changchien, Tsung-Hui Hu, Chuan-Mo Lee, Kwong-Ming Kee, Chih-Yun Lin, Chao-Long Chen, Tai-Yi Chen, Yu-Jie Huang, and Sheng-Nan Lu. The efficacy of treatment schedules according to Barcelona clinic liver cancer staging for hepatocellular carcinoma-survival analysis of 3892 patients. European journal of cancer, 44(7):1000-1006, 2008. 2", + "[35] Samir S Yadav and Shivajirao M Jadhav. Deep convolutional neural network based medical image classification for disease diagnosis. Journal of Big Data, 6(1):1-18, 2019. 2", + "[36] Naoki Yamanaka, Eizo Okamoto, Tsuyosi Oriyama, Jiro Fujimoto, Kazutaka Furukawa, Eisuke Kawamura, Tsuneo Tanaka, and Fumito Tomoda. A prediction scoring system to select the surgical treatment of liver cancer. further refinement based on 10 years of use. Annals of surgery, 219(4):342, 1994. 2", + "[37] Jianpeng Zhang, Yutong Xie, Qi Wu, and Yong Xia. Medical image classification using synergic deep learning. Medical image analysis, 54:10-19, 2019. 2", + "[38] Fan Zhou, Tengfei Li, Heng Li, and Hongtu Zhu. Tpcnn: two-phase patch-based convolutional neural network for automatic brain tumor segmentation and survival prediction." + ], + "bbox": [ + 78, + 90, + 468, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "In International MICCAI Brainlesion Workshop, pages 274-286. Springer, 2017. 3", + "[39] Tao Zhou, H. Fu, Yu Zhang, Changqing Zhang, Xiankai Lu, Jianbing Shen, and Ling Shao. M2net: Multi-modal multichannel network for overall survival time prediction of brain tumor patients. ArXiv, abs/2006.10135, 2020. 2, 3" + ], + "bbox": [ + 503, + 92, + 890, + 176 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "15690", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2023/Causally-Aware Intraoperative Imputation for Overall Survival Time Prediction/58789d3b-4b75-4c2b-ad82-075b8268b238_model.json b/2023/Causally-Aware Intraoperative Imputation for Overall Survival Time Prediction/58789d3b-4b75-4c2b-ad82-075b8268b238_model.json new file mode 100644 index 0000000000000000000000000000000000000000..f47e784582dd1548d99f30a88c54a5a9bbdc7883 --- /dev/null +++ b/2023/Causally-Aware Intraoperative Imputation for Overall Survival Time Prediction/58789d3b-4b75-4c2b-ad82-075b8268b238_model.json @@ -0,0 +1,1991 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.082, + 0.131, + 0.89, + 0.153 + ], + "angle": 0, + "content": "Causally-Aware Intraoperative Imputation for Overall Survival Time Prediction" + }, + { + "type": "text", + "bbox": [ + 0.126, + 0.18, + 0.844, + 0.218 + ], + "angle": 0, + "content": "Xiang Li\\(^{1,*}\\), Xuelin Qian\\(^{1,*}\\), Litian Liang\\(^{1,*}\\), Lingjie Kong\\(^{1}\\), Qiaole Dong\\(^{1}\\), Jiejun Chen\\(^{2}\\), Dingxia Liu\\(^{2}\\), Xiuzhong Yao\\(^{2,\\dagger}\\), Yanwei Fu\\(^{1,\\dagger}\\)" + }, + { + "type": "text", + "bbox": [ + 0.316, + 0.234, + 0.654, + 0.252 + ], + "angle": 0, + "content": "\\(^{1}\\)School of Data Science, Fudan University" + }, + { + "type": "text", + "bbox": [ + 0.223, + 0.253, + 0.746, + 0.27 + ], + "angle": 0, + "content": "\\(^{2}\\)Department of Radiology, Zhongshan Hospital, Fudan University" + }, + { + "type": "text", + "bbox": [ + 0.199, + 0.272, + 0.771, + 0.288 + ], + "angle": 0, + "content": "{li_x20, xlqian, ltliang19, qldong18, jjchen20, yanweifu}@fudan.edu.cn" + }, + { + "type": "text", + "bbox": [ + 0.189, + 0.291, + 0.777, + 0.306 + ], + "angle": 0, + "content": "{jkong22,dxliu21}@m.fudan.edu.cn yao.xiuzhong@zs-hospital.sh.cn" + }, + { + "type": "image", + "bbox": [ + 0.09, + 0.336, + 0.885, + 0.472 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.473, + 0.893, + 0.57 + ], + "angle": 0, + "content": "Figure 1. Our idea illustration. Patient A and B have different OS time (109 months and 16 months, respectively) but similar pre-operative image-based patterns and indexes; thus, preoperative based model fails to distinguish A and B (their predicted OS time are both larger than 9 years). However, their intra-operative attributes, which can describe the severity of disease in a more informative way, are very different from each other (e.g., G-score, S-score, Clinicopathologic hepatocirrhosis). Therefore, it can help the model discriminate A and B. In the right image, we show that our CAWIM – that leverages intra-operative indexes in the training stage – can largely improve the prediction power than other methods. Our method managed to correctly classify these two cases, and the overall performance of our CAWIM surpasses other baseline models by approximately 10 points on 4-category classification task." + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.579, + 0.314, + 0.595 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.611, + 0.471, + 0.866 + ], + "angle": 0, + "content": "Previous efforts in vision community are mostly made on learning good representations from visual patterns. Beyond this, this paper emphasizes the high-level ability of causal reasoning. We thus present a case study of solving the challenging task of Overall Survival (OS) time in primary liver cancers. Critically, the prediction of OS time at the early stage remains challenging, due to the unobvious image patterns of reflecting the OS. To this end, we propose a causal inference system by leveraging the intraoperative attributes and the correlation among them, as an intermediate supervision to bridge the gap between the images and the final OS. Particularly, we build a causal graph, and train the images to estimate the intraoperative attributes for final OS prediction. We present a novel Causally-aware Intraoperative Imputation Model (CAWIM) that can sequentially predict each attribute using its parent nodes in the estimated causal graph. To determine the causal directions," + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.581, + 0.893, + 0.672 + ], + "angle": 0, + "content": "we propose a splitting-voting mechanism, which votes for the direction for each pair of adjacent nodes among multiple predictions obtained via causal discovery from heterogeneity. The practicability and effectiveness of our method are demonstrated by the promising results on liver cancer dataset of 361 patients with long-term observations." + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.707, + 0.631, + 0.723 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.735, + 0.893, + 0.901 + ], + "angle": 0, + "content": "The success of recent deep learning model is largely attributed to learning the good representations for visual patterns. Such representations essentially facilitate various vision task, such as recognition and synthesis [15, 25, 33]. Nevertheless, one important goal for the vision community is to model and summarize the relationships of observed variables of a system, in order to enable well predictions on similar data. Essentially, it is desirable to understand how the system is changed if one modifies these relationships under certain conditions, e.g., the effects of a treatment in healthcare. Thus this demands the high-" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.875, + 0.204, + 0.888 + ], + "angle": 0, + "content": "*Equal contribution" + }, + { + "type": "page_footnote", + "bbox": [ + 0.098, + 0.888, + 0.22, + 0.9 + ], + "angle": 0, + "content": "† Corresponding author" + }, + { + "type": "list", + "bbox": [ + 0.096, + 0.875, + 0.22, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "15681" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.152 + ], + "angle": 0, + "content": "level ability of causal reasoning beyond the previous efforts of only learning good representations for visual patterns [1, 5, 16, 21, 35, 37]. This naturally leads into our task of causal inference." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.155, + 0.473, + 0.428 + ], + "angle": 0, + "content": "This paper presents a case study of solving the challenging task of Overall Survival (OS) time estimation in Primary Liver Cancers (PLC). Generally, the liver cancer remains one of the most common malignancies worldwide in the 21st century, as there are about one million new cases every year [36]. The five-year survival rate for advanced PLC is only about \\(5\\%\\) [7]. Therefore, early and accurate prediction of OS time estimation can provide informative guidance for individualized treatment planning and reducing burden of medical resources [6, 27, 32, 34]. On the other hand, one shall easily notice that with the renaissance of deep learning, great achievements have been made on medical imaging analysis, such as diagnosis, segmentation, and classification [16, 21, 35, 37]. Unfortunately, it still remains challenging for experienced clinicians to predict OS time at early stage, even with advanced modern diagnostic tools such as Magnetic Resonance Imaging (MRI) or tumor marker tests [3, 11], and the deep learning tools [1, 5]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.43, + 0.473, + 0.642 + ], + "angle": 0, + "content": "Some studies propose to leverage deep neural networks for OS time prediction. They focus on fusion learning of multi-modal image features and some basic information (i.e., age and gender) [5, 8, 14, 22, 24, 30, 39]. Nevertheless, the accurate prediction based on only preoperative information (such as image and tumor marker indexes) is still challenging, possibly due to missing information from early diagnosis stage to the final stage. This missing information includes the texture and pathological attributes of the liver, health level of the patient, and post-operative treatment [9, 17, 20]. For example, as in Fig. 1, patient A and B with different OS time have almost identical preoperative indicators cause, history of disease, etc., making the preoperative-based model hard for discrimination." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.644, + 0.473, + 0.903 + ], + "angle": 0, + "content": "To amend this problem, we present a causal inference system that can well utilize the intraoperative information, which is pretty easy to be accessed in training data. According to medical priorities, such information records pathological attributes, which can be more reflective about the patient's health level and the postoperative recovery. Inspired by this, we propose to leverage this auxiliary information to help build our causal inference system. Again, we take the example in Fig. 1. Although preoperative information cannot differentiate patient A from B, their intraoperative index shows great difference in distribution, which can thus be employed for the discrimination. Additionally, there are many indicators from medical experts that the intraoperative indexes are related to each other. For instance, the clinicopathologic hepatocirrhosis is dependent on the hepatocirrhosis; the sum of tumor diameter is affected by the number of tumors; the fibrosis (reflected on S-Score) can prob" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.152 + ], + "angle": 0, + "content": "ably deteriorate to cirrhosis [4], etc. By leveraging these relationships, we can better understand the causal inference concrete from both observed data modelling and medical expert-level knowledge of these variables." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.156, + 0.893, + 0.625 + ], + "angle": 0, + "content": "To this end, we encapsulate these priors and the inspired proposals into a new method, dubbed as Causally-Aware Intraoperative Imputation Model (CAWIM). It incorporates causal discovery module to sequentially estimate intraoperative indexes as an intermediate stage towards final OS time prediction. Specifically, our model is composed of two key steps: i) estimating the intraoperative indexes using preoperative information, i.e., image and indexes; ii) followed by OS time prediction using estimated intraoperative indexes and preoperative information. To achieve more accurate prediction of intraoperative indexes that is determinant to the prediction power of the whole method, we propose a Causally-aware Directed Acyclic Graph (CaDAG) module. It learns the causal structure represented as a DAG over intraoperative features. To identify the causal relations beyond the traditional PC algorithm [26], we propose a splitting-voting mechanism, which is inspired by the recent work [18] that learn the causal structure with the assistance of an auxiliary domain index variable. Our proposed mechanism can not only identify the causal relations even when this domain index variable is not available, but also can be theoretically guaranteed that the learned graph is not acyclic. During test stage, we sequentially estimate each intraoperative index with preoperative information and additionally, its parent set among other intraoperative indexes. The utility of our method can be demonstrated by a significant improvement of OS time prediction on an in-house liver cancer dataset, as well as better interpretability of learned causal structure, more accurate estimation of intraoperative indexes and more interpretable visualization results." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.629, + 0.895, + 0.903 + ], + "angle": 0, + "content": "In a nutshell, we for the first time present a case study of building a causally-aware intraoperative imputation system for the challenging task of overall survival time prediction. The proposed method of building the casual inference system, can be naturally extended to other similar medical tasks. Our key contributions are listed as follows. (1) New Paradigm for OS time Prediction. We propose to leverage intraoperative indexes as an intermediate stage during training. The leverage of this information can significantly alleviate the \"missing information\" issue. To the best of our knowledge, we are the first to leverage auxiliary information (in addition to preoperative features) for OS time prediction. (2) Causal Structure Learning. We propose a novel splitting-voting mechanism that can identify the causal structure even when the domain index variable is missing. (3) Better Prediction Power. Our method can significantly improve the prediction power for liver cancer over the competitors. The methods are evaluated on the medical" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "15682" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.47, + 0.122 + ], + "angle": 0, + "content": "dataset, which, to the best of our knowledge, is the largest primary live cancer dataset." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.136, + 0.373, + 0.151 + ], + "angle": 0, + "content": "2. Related Work and Preliminaries" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.161, + 0.468, + 0.235 + ], + "angle": 0, + "content": "Previous efforts [23,39] directly take the OS time prediction as the vanilla classification task by employing various deep models, whilst we reformulate it by causal discovery method for the first time. We will review the most related literature here." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.237, + 0.468, + 0.417 + ], + "angle": 0, + "content": "MRI based deep models. OS time is defined as the duration from patient's first scan to cancer-related death [6]. Since Magnetic Resonance Image (MRI) plays an important role in tumor-related studies, various MRI based deep learning model have been proposed for feature extraction and representing [28], multi-modal feature fusion [10, 30, 38, 39], and improvement on model structure [5]. However, since missing information during disease progression can mediate the correlation from the preoperative stage to the final stage, the results of these image based methods are typically unsatisfied. In contrast, our method can amend this problem by leveraging intraoperative indexes as a bridge." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.418, + 0.468, + 0.749 + ], + "angle": 0, + "content": "Causal discovery. Causal discovery is to learn the causal graph over endogenous variables. Typical methods include PC algorithm [19], a constraint-based method that implement iterative conditional independence test to identify the skeleton and directions via v-structure. Under Markovian and faithfulness assumptions [26, 31], this method is provable to identify an equivalent class of the causal graph. To further identify more directions, recent work [18] leveraged a domain index variable that splits the dataset into multiple domains. With these heterogeneous data, they can identify mutable variables that change across the domain index variable. On the basis of this, they further identified edges among these mutable variable set by implementing Hilbert Schmidt Independence Criterion (HSIC) norm [13]. However, such domain index may not be available in many scenarios, making the identification of these directions difficult. Besides, this method may not ensure that the learned graph is acyclic. To address these issues, we in this paper propose a splitting-voting mechanism to mine the causal relations among intraoperative indexes so as for leveraging each other to better estimate these indexes, which is provable to return a directed acyclic graph." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.468, + 0.901 + ], + "angle": 0, + "content": "Preliminaries. (1) Causal Discovery. By faithfulness and Markovian assumptions [26], we can use conditional independent tests for causal discovery. That is learning the causal graph over \\(\\mathbf{V}\\). (2) Structural causal model (SCM). Before fully developing our contributions in methodology, we give as preliminary the definition of SCM, which is defined as \\(\\langle G, \\mathcal{F}, P\\langle \\epsilon \\rangle \\rangle\\): i) directed acycle graph \\(G = (\\mathbf{V}, \\mathbf{E})\\) is as causal structure, with node set \\(\\mathbf{V}\\) and edge set \\(\\mathbf{E}\\); ii) the autonomous structural functions \\(\\mathcal{F} = \\{f_k\\}_{V_k \\in V}\\), where disturbance on \\(V_k\\) has no effect on others; iii) probability" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.455 + ], + "angle": 0, + "content": "measure \\( P(\\epsilon) \\) for exogenous variables \\( \\{\\epsilon_k\\}_k \\). Given the assumption that \\( \\{\\epsilon_k\\}_k \\) are independent, each \\( \\mathrm{P} \\) can be obtained by Causal Markov Condition with \\( G \\) as \\( \\mathrm{P}(\\{V_K = v_k\\}_{V_k \\in \\mathbf{V}}) = \\prod_k \\mathrm{P}(V_k = v_k | Pa(k) = pa(k)) \\). (3) Causal discovery from heterogeneity. In [18], the authors proposed to leverage a domain index variable \\( C \\) to identify some causal directions. This method is built upon the PC and faithfulness assumption at a distributional level. It first learns the skeleton of the graph and identifies a variable set that is affected by \\( E \\), followed by identifying the directions between \\( E \\) and its neighbors. Specifically, if \\( C \\to V_i, C \\to V_j \\), and \\( V_i \\bot V_j | X, C \\) for some deconfounding set \\( X \\) between \\( V_i \\) and \\( V_j \\), \\( \\{P(V_i | X, C = c)\\} \\) is independent of \\( \\{P(V_j | V_i, C = c, X)\\} \\), while \\( \\{P(V_j | X, C = c)\\} \\) and \\( \\{P(V_j | V_i, C = c, X)\\} \\) are dependent. We can test these Independence using Hilbert Schmidt Independence Criterion (HSIC) [12] norms \\( \\delta_{v_i \\to v_j | X} \\) and \\( \\delta_{v_j \\to v_i | X} \\). One can obtain that \\( V_i \\to V_j \\) if \\( \\delta_{v_i \\to V_j} < \\alpha \\) with pre-set significance level \\( \\alpha \\). However, in many applications such domain index variable is missing, i.e., all data are pooled together. In this paper, we propose a splitting-voting mechanism (in the CaDAG module) to identify causal relations. In the following, we first introduce some basic assumptions that our method is built upon." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.466, + 0.634, + 0.483 + ], + "angle": 0, + "content": "3. Methodology" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.491, + 0.892, + 0.627 + ], + "angle": 0, + "content": "Problem setting. For the OS time classification task, suppose we have \\(\\{X_{i},\\mathbf{B}_{i},\\mathbf{A}_{i},y_{i}\\} \\sim_{i.i.d}P(X,\\mathbf{B},\\mathbf{A},y)\\) where \\(X\\) denotes the image acquired from structural Magnetic Resonance Image (sMRI); \\(\\mathbf{B}\\), \\(\\mathbf{A}\\) denote the preoperative and intraoperative attributes, respectively. We have the final label \\(y\\). Our goal is to predict \\(y\\) from the image \\(X\\) and preoperative attributes \\(\\mathbf{B}\\) that are recorded before surgery. Our method is built upon the following three assumptions, commonly utilized in the causal inference works [18, 26]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.635, + 0.892, + 0.682 + ], + "angle": 0, + "content": "Definition 1 (Causal Graph) We assume the causal graph over \\(\\mathbf{A}\\) is a directed acyclic graph (DAG) and denote the corresponding SCM as \\(M\\coloneqq \\langle G\\coloneqq (\\mathbf{A},\\mathbf{E}),\\mathcal{F},P(\\varepsilon)\\rangle\\)" + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.689, + 0.831, + 0.703 + ], + "angle": 0, + "content": "Definition 2 (Markovian and Faithfulness) For" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.704, + 0.892, + 0.764 + ], + "angle": 0, + "content": "triplets of disjoint sets \\(\\mathbf{V}_i, \\mathbf{V}_j, \\mathbf{V}_k\\), it holds that \\(\\mathbf{V}_i \\perp_d \\mathbf{V}_j | \\mathbf{V}_k \\leftrightarrow \\mathbf{V}_i \\perp \\mathbf{V}_j | \\mathbf{V}_k\\), where \\(\\perp_d\\) and \\(\\perp\\) respectively mean \\(d\\)-separation and probability independence. This is the common property in DAG." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.772, + 0.892, + 0.863 + ], + "angle": 0, + "content": "Definition 3 (Distributional Faithfulness) If \\( X_{i} \\to X_{j} \\) and at least \\( E \\to V_{i} \\) or \\( E \\to V_{j} \\) holds, \\( \\{P^{e}(V_{i}|V_{j},\\mathbf{Z})\\} \\) is dependent to \\( \\{P^{e}(V_{j}|\\mathbf{Z})\\} \\), where \\( \\mathbf{Z} \\) denotes the minimal deconfounding set. Particularly, \\( \\mathbf{Z} \\) is a deconfounding set between \\( V_{i} \\) and \\( V_{j} \\) if we have \\( V_{i} \\perp V_{j}|\\mathbf{Z} \\) and \\( \\mathbf{Z} \\cap (\\mathrm{De}(V_i) \\cup \\mathrm{De}(V_j)) = \\emptyset \\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.871, + 0.892, + 0.901 + ], + "angle": 0, + "content": "In term of this assumption, we further have the following result for edge orientation [18]." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15683" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.091, + 0.473, + 0.17 + ], + "angle": 0, + "content": "Theorem 4 (Theorem 2 in [18]) Denote \\( E \\) as domain index variable. Under assumptions 1, 2, 3, for each adjacent pair \\( (V_i, V_j) \\) such that \\( E \\to V_i \\) or \\( E \\to V_j \\) holds, we have \\( V_i \\to V_j \\) if \\( \\{P^e(V_i|\\mathbf{Z})\\} \\) and \\( \\{P^e(V_j|V_i,\\mathbf{Z})\\} \\) are independent while \\( \\{P^e(V_j|\\mathbf{Z})\\} \\) and \\( \\{P^e(V_i|V_j,\\mathbf{Z})\\} \\) are dependent." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.183, + 0.421, + 0.2 + ], + "angle": 0, + "content": "3.1. Causally-aware Directed Acyclic Graph" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.207, + 0.473, + 0.493 + ], + "angle": 0, + "content": "The Causally-aware DAG. We propose a novel Causally-Aware Intraoperative Imputation Model (CAWIM). The pipeline of our method is shown in Fig. 2. It is composed of Causally-aware Directed Acycled Graph (CaDAG), and a classification model with MRI imaging encoder. The CaDAG enables the causally-aware intraoperative reasoning, by managing to better represent and encode the correlation of intraoperative information. Specifically, we first use preoperative information, i.e., MRI images and indexes to estimate the intraoperative features. The estimated intraoperative features \\(\\tilde{\\mathbf{A}}\\), together with the MRI \\(X\\) and preoperative features \\(\\mathbf{B}\\) are then utilized for OS classification. To accurately estimate intraoperative indexes \\(\\mathbf{A}\\), we propose CaDAG to learn causal structure over \\(\\mathbf{A}\\), and then leverage the parent nodes of each index for sequentially imputation during test stage. We give the details about the CaDAG and the intraoperative imputation model for classification, while some theoretical proofs, and the specification of our networks are in the appendix." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.494, + 0.47, + 0.734 + ], + "angle": 0, + "content": "We aim at learning the causal structure over \\(\\mathbf{A}\\). However, without the domain index, we actually have the difficulty of dividing the whole dataset into multiple domains. To this end, we propose a splitting-voting mechanism for this problem. Particularly, this mechanism first generates the domain-index variable \\(E\\) by randomly splitting the whole dataset into multiple domains for \\(m > 0\\) times. Then for each time we identify the mutable variable set and generate a prediction for each identifiable causal directions, with the assistance of domain index variable \\(E\\). Finally, we vote for a final direction among \\(m\\) predictions for each pair of variables. As shown in Fig. 3, the whole procedure is a sequence of the four steps. Step i), ii) and iii) provide preliminary causal relationship between each pair of nodes. Step iv) is our proposed splitting-voting, aiming to finally determine all the causal directions and output a causal DAG." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.735, + 0.47, + 0.84 + ], + "angle": 0, + "content": "i) Learning skeleton of the causal graph. We first implement PC algorithm to learn the skeleton [26]. That is an undirected graph over all attributes. This can be achieved by iterative conditional independence test. Then we determine \\( A_{i} \\) and \\( A_{j} \\) to be adjacent if they are not independent conditional on any subsets. Finally, we utilize the v-structure to learn the equivalence class of DAG." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.471, + 0.902 + ], + "angle": 0, + "content": "ii) Random splitting and identification of mutable variable sets. To determine more directions, we randomly split the dataset into multiple domains for \\( m \\) times. For each time we have a domain index variable \\( E_{i} \\) and multiple do" + }, + { + "type": "list", + "bbox": [ + 0.076, + 0.207, + 0.473, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.091, + 0.895, + 0.168 + ], + "angle": 0, + "content": "mains \\(\\{\\mathcal{D}_e^i\\}\\) [18]; and we identify the mutable variable set \\(\\mathbf{M}_i\\) such that each \\(X \\in \\mathbf{M}_i\\) has \\(X \\not\\perp E_i|\\mathbf{A} - X\\) is conditionally dependent to \\(E_i\\). We thus can implement conditional independent tests to identify \\(\\mathbf{M}_i\\). Steps i) and ii) are summarized in Alg. 1." + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.181, + 0.838, + 0.197 + ], + "angle": 0, + "content": "Algorithm 1 Identify the skeleton and mutable set." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.212, + 0.877, + 0.228 + ], + "angle": 0, + "content": "INPUT: \\(\\{\\mathcal{D}_e^i\\coloneqq \\{\\mathbf{A}_i^e\\}_{i = 1}^{n_e}|e\\}\\), domain index variable \\(E_{i}\\)." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.228, + 0.803, + 0.244 + ], + "angle": 0, + "content": "OUTPUT: Equivalent class of DAG and \\(\\mathbf{M}_i\\)." + }, + { + "type": "text", + "bbox": [ + 0.509, + 0.256, + 0.892, + 0.285 + ], + "angle": 0, + "content": "1: Implement the PC algorithm to learn the equivalence class of DAG via v-structure." + }, + { + "type": "text", + "bbox": [ + 0.509, + 0.286, + 0.892, + 0.316 + ], + "angle": 0, + "content": "2: For each \\(A \\in \\mathbf{A}\\), add the edge \\(E_i \\to A\\) in the graph iff \\(E_i \\not\\perp A\\) given any subsets." + }, + { + "type": "list", + "bbox": [ + 0.509, + 0.256, + 0.892, + 0.316 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.336, + 0.895, + 0.541 + ], + "angle": 0, + "content": "iii) Identifying directions via changed causal models. For each adjacent pair \\( A_{i} \\) and \\( A_{j} \\) such that at least one of them is adjacent to \\( E \\), we identify the causal directions among \\( \\mathbf{M}_i \\) via HSIC Norm [12] by testing the independence between \\( \\{P^{e}(A_{i}|A_{j},\\mathbf{Z})\\} \\) and \\( \\{P^{e}(A_{j}|\\mathbf{Z})\\} \\) if \\( A_{i}\\gets A_{j} \\), and that between \\( \\{P^{e}(A_{j}|A_{i},C)\\} \\) and \\( \\{P^{e}(A_{i}|\\mathbf{Z})\\} \\) for determining the direction between \\( A_{i} \\) and \\( A_{j} \\). Here, \\( \\mathbf{Z} \\) denotes the deconfounding set. In this regard, the direction between \\( A_{i} \\) and \\( A_{j} \\) is then determined according to the following rule: if \\( \\hat{\\Delta}_{A_i\\to A_j|\\mathbf{Z},E} < \\hat{\\Delta}_{A_j\\to A_i|C,E} \\), output \\( A_{i}\\rightarrow A_{j} \\); if \\( \\hat{\\Delta}_{A_i\\to A_j|\\mathbf{Z},E} > \\hat{\\Delta}_{A_j\\to A_i|\\mathbf{Z},E} \\), output \\( A_{i}\\leftarrow A_{j} \\); if \\( \\hat{\\Delta}_{A_i\\to A_j|\\mathbf{Z},E} > \\alpha \\) and \\( \\hat{\\Delta}_{A_j\\to A_i|\\mathbf{Z},E} > \\alpha \\), where \\( \\alpha \\) is a threshold, making the direction undetermined." + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.554, + 0.834, + 0.57 + ], + "angle": 0, + "content": "Algorithm 2 Identify causal directions among M." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.583, + 0.758, + 0.598 + ], + "angle": 0, + "content": "INPUT: \\(E\\); skeleton and \\(M\\) via Alg.1." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.599, + 0.753, + 0.615 + ], + "angle": 0, + "content": "OUTPUT: Directed graph among M." + }, + { + "type": "text", + "bbox": [ + 0.509, + 0.626, + 0.877, + 0.642 + ], + "angle": 0, + "content": "1: For each adjacent \\((A_{i},A_{j})\\) such that one of \\(V_{i}\\in \\mathbf{M}\\)" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.643, + 0.736, + 0.656 + ], + "angle": 0, + "content": "2: Detect deconfounding set \\(\\mathbf{Z}\\)" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.656, + 0.885, + 0.673 + ], + "angle": 0, + "content": "3: Calculate \\(\\hat{\\Delta}_{A_i\\to A_j|\\mathbf{Z},E} > \\alpha\\) and \\(\\hat{\\Delta}_{A_j\\to A_i|\\mathbf{Z},E} > \\alpha\\)" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.673, + 0.771, + 0.688 + ], + "angle": 0, + "content": "4: Determine \\( A_{i} \\to A_{j} \\) or \\( A_{j} \\to A_{i} \\)." + }, + { + "type": "list", + "bbox": [ + 0.509, + 0.626, + 0.885, + 0.688 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.708, + 0.892, + 0.816 + ], + "angle": 0, + "content": "iv) Voting for the direction of each edge. For each time \\(i\\) we implement Alg. 2 to predict an direction of the edge \\((A_{i}, A_{j})\\). Then we vote for \\(A_{i} \\rightarrow A_{j}\\) if the frequency \\(f_{i \\rightarrow j}\\) of \\(A_{i} \\rightarrow A_{j}\\) surpasses a pre-set threshold \\(\\alpha\\) (\\(0 < \\alpha < 1\\)). Finally, if there exists a cycle of \\(N\\) variables: \\(A_{i_1} \\rightarrow A_{i_2} \\rightarrow \\ldots \\rightarrow A_{i_1}\\), we orient \\((A_{i_k}, A_{i_{k+1}})\\) with the minimum frequency among \\(\\{f_{i_k, i_{k+1}}, f_{i_N, i_1}\\}\\)." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.829, + 0.892, + 0.86 + ], + "angle": 0, + "content": "We show that this procedure can generate a DAG over \\(\\mathbf{A}\\), with the proof in appendix." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.871, + 0.892, + 0.902 + ], + "angle": 0, + "content": "Theorem 5 Under definitions and assumptions 1, 2, 3, the learned graph via our CaDAG is a directed acyclic graph." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "15684" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.086, + 0.885, + 0.221 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.225, + 0.893, + 0.253 + ], + "angle": 0, + "content": "Figure 2. Overview of our CAWIM model. We concatenate the MRI features encoded by the MRI Imaging Encoder, preoperative indicators, and the estimated intraoperative indicators obtained from CaDAG to perform the final prediction on OS time." + }, + { + "type": "image", + "bbox": [ + 0.119, + 0.256, + 0.862, + 0.469 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.472, + 0.893, + 0.542 + ], + "angle": 0, + "content": "Figure 3. Schematic diagram of a four-variable case \\( A_{1} \\), \\( A_{2} \\), \\( A_{3} \\), and \\( A_{4} \\), with the real DAG shown in the top left corner. Through step i) and ii) in the upper part, a Markov Equivalence Class of 3 causal graphs is observed with the direction between \\( A_{1} \\) and \\( A_{2} \\) undetermined. For step iii), we split the dataset into \\( (n = 3) \\) domains and consider the 4 circumstances shown fig(a) to (d) in the lower left part. For step iv), we vote for the direction between \\( A_{1} \\) and \\( A_{2} \\) as \\( A_{1} \\rightarrow A_{2} \\) according to the number of splitting \\( (n = 3) \\), frequency \\( f_{1\\rightarrow 2} = 2 / 3 \\) and threshold \\( \\alpha = 2 / 3 \\), as shown in the lower right part. Finally, we obtain an inferred DAG with each direction determined." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.55, + 0.45, + 0.566 + ], + "angle": 0, + "content": "3.2. Intraoperative Imputation for Classification" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.572, + 0.47, + 0.74 + ], + "angle": 0, + "content": "Given MRI image \\( X \\in \\mathbb{R}^{H \\times W} \\) of a patient that contains multiple slides, we first resample these slides into a fixed \\( K \\) slides for each patient. As each slice is a gray image, we repeat the images for 3 times along the channel dimension; and then we employ a neural network to extract features from each slice. The extracted features are denoted as \\( F_{1}, \\ldots, F_{K} \\). In order to aggregate the information of all slices, we divide \\( F_{1}, \\ldots, F_{K} \\) into \\( g \\) groups according to the acquired position. Each group contains consecutive \\( \\lfloor K / g \\rfloor \\) slices. Then, we implement a Multi-head Cross-Attention (MCA) mechanism inside each group to aggregate features:" + }, + { + "type": "equation", + "bbox": [ + 0.138, + 0.748, + 0.469, + 0.767 + ], + "angle": 0, + "content": "\\[\n\\operatorname {h e a d} _ {j} = \\operatorname {s o f t m a x} \\left(Q ^ {j} \\left(F \\mathbf {W} _ {K} ^ {j}\\right) ^ {T}\\right) \\left(F \\mathbf {W} _ {V} ^ {j}\\right) \\tag {1}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.165, + 0.774, + 0.469, + 0.79 + ], + "angle": 0, + "content": "\\[\n\\mathbf {M C A} = \\left[ \\text {h e a d} _ {1}; \\dots ; \\text {h e a d} _ {h} \\right] \\mathbf {W} _ {O} \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.795, + 0.469, + 0.87 + ], + "angle": 0, + "content": "where \\(F\\) is the slice's feature within a group, the query vector \\(Q^{j}\\) and linear projection layers \\(\\mathbf{W}_K^j\\), \\(\\mathbf{W}_V^j\\), \\(\\mathbf{W}_O\\), \\(1 \\leq j \\leq h\\) are all learnable parameters. Finally, we concatenate \\(\\{F_k\\}\\) from \\(g\\) groups as our final image feature \\(I \\in \\mathbb{R}^{g \\times d}\\) with \\(d\\) denoting the dimension of each \\(F_k\\) with \\(k = 1, \\ldots, K\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.871, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Afterwards, we sequentially estimate each \\(A_{j}\\) by preoperative information, as our learned causal graph over \\(\\mathbf{A}\\). So" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.55, + 0.659, + 0.566 + ], + "angle": 0, + "content": "for each \\(A_{j}\\), we predict" + }, + { + "type": "equation", + "bbox": [ + 0.61, + 0.579, + 0.892, + 0.598 + ], + "angle": 0, + "content": "\\[\n\\hat {A} _ {j} = f _ {\\text {i n t r o}} ^ {j} (\\mathbf {B}, I, \\operatorname {P a} (A _ {j})) \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.61, + 0.892, + 0.74 + ], + "angle": 0, + "content": "where \\(\\mathrm{Pa}(A_j)\\) denotes the parent node set of \\(A_{j}\\). We use cross entropy loss to optimize \\(f_{\\mathrm{intra}}^j\\) if \\(A_{j}\\) is categorical; and optimize \\(f_{\\mathrm{intra}}^j\\) if \\(A_{j}\\) is continuous. After obtaining all predicted \\(\\{\\hat{A}_i\\}_{i = 1}^d\\), we use \\((I,\\mathbf{B},\\hat{\\mathbf{A}})\\) to predict \\(Y\\) via \\(f_{\\theta}(I,\\mathbf{B},\\hat{\\mathbf{A}})\\) and obtain \\(\\theta\\) by optimizing the cross-entropy \\(\\mathcal{L}(y,f_{\\theta}(I,\\mathbf{B},\\hat{\\mathbf{A}}))\\). During test stage, given a new sample \\((x,\\mathbf{b})\\), we first extract \\(I\\) and then sequentially estimate \\(\\mathbf{a}_i\\) and finally predict \\(y\\) using \\(f_{\\theta}(x,\\mathbf{b},\\hat{\\mathbf{a}})\\)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.755, + 0.633, + 0.771 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.78, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Dataset overview. Ethics Committee of Zhongshan Hospital, Fudan University approved the protocol of this study and waived the requirement for patient-informed consent (B2021-325R). Because open source data sets generally lack sufficient intraoperative information, it takes us around ten years to collect and build our own medical dataset. Particularly, we conduct a search through the medical records in the hospital information system, and build a dataset with" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "15685" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.091, + 0.473, + 0.348 + ], + "angle": 0, + "content": "439 patients infected by primary liver cancer. Long-term medical data of every patient has been recorded from the first scan to the cancer-related death. Till the submission of this manuscript, around \\(90\\%\\) patients have unfortunately passed away. This gives us the ground-truth OS labels. Information like MRI scanning before the surgery, blood test and operation reports, follow-up records are also collected. After removing samples with missing information, 361 patients were enrolled, with 306 \\((84.8\\%)\\) men, and the average age was 53.7 \\((\\pm 11.7)\\) years. Specifically, each patient has MRI scanning with four modalities, including \\(\\mathrm{T}_{1}\\) weighted images \\((\\mathrm{T}_{1}\\)-WI), \\(\\mathrm{T}_{2}\\) weighted images \\((\\mathrm{T}_{2}\\)-WI), \\(\\mathrm{T}_{1}\\) in-phase \\((\\mathrm{T}_{1}\\)-IP) and contrast-enhanced \\(\\mathrm{T}_{1}\\) at the arterial phase \\((\\mathrm{T}_{1}\\)ce-AP). Following the guidance of medical experts, we select 11 preoperative and 20 intraoperative indicators that are related to OS time for our study. Please refer to the appendix for details." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.35, + 0.473, + 0.533 + ], + "angle": 0, + "content": "Data preprocessing. In our dataset, the survival time ranges from 0 to 130 months, which are cauterized into four classes: short-term survival (\\(\\leq 36\\) months), middle-short-term survival (between 36 and 72 months), middle-long-term survival (between 72 and 108 months), and long-term survival (\\(\\geq 108\\) months). In our paper, we consider the classification task: given a new sample, our goal is to classify which class of OS time this sample belongs to. To extract \\(I\\) from MRI image, we resample \\(K = 20\\) slices and employ Resnet-34 for feature extraction. We set \\(g = 4\\) and the dimension \\(F_{k}\\) as 512. For \\(f_{\\mathrm{intro}}^{j}\\) and \\(f_{\\theta}\\), we employ a multi-layer perceptron (MLP) for parameterization." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.534, + 0.473, + 0.64 + ], + "angle": 0, + "content": "We encode categorical covariates into dummy variables and implement zero-mean normalization for continuous variables. For MRI images, we resize them into \\(256 \\times 256\\) as input. We split the whole dataset into 5 folds. To remove the effect of imbalance across classes and randomness, we adopt average precision, recall, and \\(\\mathrm{F_1}\\)-Score measurements over five folds for classification evaluation." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.642, + 0.473, + 0.809 + ], + "angle": 0, + "content": "Implementation details. We use the ImageNet pretrained Resnet-34 to initialize our CNN encoder. We train 150 epochs for all methods with \\(\\mathrm{T}_{1}\\)ce-AP modality, using SGD as the optimizer with learning rate \\(3e - 3\\) and \\(3e - 2\\) for CNN encoder and later MLP, respectively. We set weight decay to \\(5e - 4\\) and nesterov momentum factor to 0.9. Besides, we decrease the learning rate every 60 epochs by a factor of 1/10. We set the Batch size to 8. In our voting mechanism of CaDAG, we in total implement Alg. 2 for seven times and set the voting threshold as 4/7. Our model is trained with Pytorch on NVIDIA GeForce RTX 3090 GPUs." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.81, + 0.473, + 0.903 + ], + "angle": 0, + "content": "Compared Methods. We compare our method with four baseline methods: i) Random: random guess among four classes as a trivial baseline, of which the expected \\(\\mathrm{F_1}\\) score is \\(25\\%\\); ii) MRI: train the end-to-end method with \\(\\mathrm{T_1}\\)ce-AP only as input for classification; iii) MRI+pre: additionally take preoperative features as input, on the basis of the" + }, + { + "type": "table_caption", + "bbox": [ + 0.515, + 0.089, + 0.877, + 0.103 + ], + "angle": 0, + "content": "Table 1. Comparisons of Our CAWIM with Other Baselines." + }, + { + "type": "table", + "bbox": [ + 0.502, + 0.106, + 0.888, + 0.197 + ], + "angle": 0, + "content": "
ModelPrecisionRecallF1-Score
(a)Random26.04±2.1024.54±0.8023.15±0.64
(b)MRI40.79±7.9733.45±2.6631.88±3.92
(c)MRI+Pre36.99±5.3035.62±3.5033.83±3.64
(d)gt Intra41.26±2.7441.62±1.9539.77±2.47
(e)CAWIM45.58±5.8643.70±5.8942.21±4.92
" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.21, + 0.892, + 0.304 + ], + "angle": 0, + "content": "method MRI; iv) ground-truth (GT) Intra: only take the ground-truth intraoperative features as input. It is worth to mention that although the intraoperative features are not allowed to use during test stage, the comparison with method iv (i.e., GT Intra) can provide information that to which extent, the intraoperative features can help OS time prediction." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.312, + 0.697, + 0.329 + ], + "angle": 0, + "content": "4.1. Results and Analysis" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.336, + 0.892, + 0.61 + ], + "angle": 0, + "content": "Table 1 records comparisons of our proposed CAWIM (Tab. 1(d)) with other baseline models. (Tab. 1(a), Tab. 1(b), and Tab. 1(c)). Obviously, our CAWIM constantly outperform others in all metrics. Specifically, CAWIM improved \\( \\text{MRI} + \\text{Pre} \\) by nearly \\( 10\\% \\) in terms of \\( \\text{F}_1 \\)-Score, which can be contributed to the prediction of intraoperative features equipped with causally aware prediction mechanism. Without this mechanism, using MRI and preoperative features only can outperform random guess by only \\( 8.33\\% \\), which demonstrates the limited information of the image for OS time. In contrast, it is also interesting to note from Tab. 1(d) that using intraoperative features can be more informative than MRI features, in terms of OS prediction, which is natural as the attributes that described the functions and textures of the liver could affect the postoperative recovery and thus the OS time. Finally, our CAWIM (Tab. 1(e)) that combines the information of intraoperative features and the information from images, can achieve further improvement." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.62, + 0.655, + 0.636 + ], + "angle": 0, + "content": "4.2. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.643, + 0.892, + 0.84 + ], + "angle": 0, + "content": "To test the effectiveness of preoperative features, the particularly intraoperative features, and our causally-aware module for OS time prediction, we implement ablation studies and summarize the results in Tab. 2. Specifically, Tab. 2(a), i.e., w/o CaDAG replaces our causally-aware module with an end-to-end estimation network from (MRI, preoperative features) to intraoperative features; Tab. 2(b), i.e., w/o Intra is the same to Tab. 1(c), i.e., MRI+pre in Tab. 1 that directly predict OS time using MRI and preoperative features without estimated intraoperative features; Tab. 2(c), i.e., w/o Pre is the same to our CAWIM except that it does not utilize preoperative features for intraoperative estimation and OS prediction." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.894, + 0.903 + ], + "angle": 0, + "content": "The results in Tab. 2 show that the deletion of each module can lead to a descent of all metrics. Specifically, Tab. 2(a) shows a significant performance drop by \\(5.6\\%\\) in \\(\\mathrm{F_1}\\)-Score, which indicates the effectiveness of our causally" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "15686" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.332 + ], + "angle": 0, + "content": "aware module in estimating intraoperative features in OS time prediction. This phenomena can be explained by more accurate intraoperative estimation equipped with our causally-aware module, as shown in Fig. 4 of \\(\\mathrm{F_1}\\)-score, accuracy and \\(R\\)-squared score. We will leave detailed discussions in Section of \"Further Discussion\". Tab. 2(b), as the same to MRI+pre in Tab. 1, shows a \\(8.3\\%\\) performance drop, which indicates the additional information provided by the MRI image and preoperative features. Finally, the \\(3.78\\%\\) degradation of Tab. 2(c) compared to ours validates the effectiveness of preoperative features in OS prediction, which can also be explained by the estimation results of intraoperative features in Fig. 4. In a word, we can observe a significant drop of \\(F_{1}\\)-score if we remove causally-aware module and intraoperative features, which can demonstrate the effectiveness of our methods." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.334, + 0.473, + 0.546 + ], + "angle": 0, + "content": "Besides, it is also interesting to compare Tab. 2(b) and Tab. 2(c) with we observe that the performance drop for \\( w/o \\) Intra is more significant than that for \\( w/o \\) Pre, with respectively an average of \\( 8.4\\% \\) and \\( 3.8\\% \\) drop in \\( F_1 \\)-Score. This again, indicates more information of intraoperative indexes provided than preoperative indexes in predicting OS time. Indeed, we obtained Precision \\( 27.68\\% \\), Recall \\( 26.65\\% \\) and F1-Score \\( 23.85\\% \\) if we only predicted OS time using preoperative features, which is comparable to the result from random guess. On the other hand, we can observe from Fig. 4 that preoperative indexes are beneficial to the estimation of intraoperative features, which indicating the necessity of using intraoperative features as a bridge between preoperative information and final OS time prediction." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.559, + 0.422, + 0.575 + ], + "angle": 0, + "content": "4.3. Medical Interpretation of Causal Graph" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.583, + 0.47, + 0.749 + ], + "angle": 0, + "content": "To further explain the effectiveness of our causally-aware module, we in this subsection present our learned causal graph over intraoperative features in Fig. 5 and the corresponding medical interpretation of learned causal relations. In general, most of the parent-child relationships in our CaDAG are in line with common sense or can be explained by prior medical knowledge, e.g., sum of tumor diameter is determined by number of tumors, the edge from \\( G \\)-score to \\( S \\)-score also correspond to the strong correlation between fibrosis and hepatitis (please refer to appendix for detailed descriptions of intraoperative indexes)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.471, + 0.901 + ], + "angle": 0, + "content": "Further, those indicators with non-visual features that access severity of pathology are often determined by other attributes of the liver; therefore, these features often come up in deeper layers in Fig. 5. For instance, \\( G \\)-score and \\( S \\)-score reflect the level of inflammation and fibrosis that are artificially defined according to Metavir scoring system [2]. Those indicators tend to rely on other intraoperative features besides merely MRI and preoperative information, e.g. clinicopathologic hepatocirrhosis is reasoned from hepatocirrhosis and other covariants. For this reason, the addi" + }, + { + "type": "table_caption", + "bbox": [ + 0.498, + 0.089, + 0.895, + 0.159 + ], + "angle": 0, + "content": "Table 2. Results of Ablation Studies. (a) estimates each intraoperative features naively from (MRI, preoperative indexes), without the causally-aware module. (b) is the same to \\( M R I + P r e \\) in Tab. 1. (c) is the CAWIM without preoperative features for intraoperative indexes estimation and OS time prediction." + }, + { + "type": "table", + "bbox": [ + 0.502, + 0.161, + 0.905, + 0.239 + ], + "angle": 0, + "content": "
ModelPrecisionRecallF1-Score
(a)w/o CaDAG36.98±2.3837.78±1.9136.62±1.87
(b)w/o Intra36.99±5.3035.62±3.5033.83±3.64
(c)w/o Pre38.92±7.7141.57±4.5338.43±5.84
(d)CAWIM45.58±5.8643.70±5.8942.21±4.92
" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.249, + 0.895, + 0.34 + ], + "angle": 0, + "content": "tional leverage of hepatocirrhosis lead to more accurate estimation of clinicopathologic hepatocirrhosis, as shown in Fig. 4 (a). On the other hand, we also observe some associations that cannot be explained well, e.g. the causal relationship between sum of tumor diameter and cell type, which may due to the existence of unobserved confounders." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.348, + 0.638, + 0.362 + ], + "angle": 0, + "content": "4.4. Visualization" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.371, + 0.895, + 0.522 + ], + "angle": 0, + "content": "In this section, we visualize the high-response area of our CAWIM Fig. 6(c) and the version without CaDAG Fig. 6(b) using Grad-CAM [29]. As shown in Fig. 6, the detected regions of our method can be more concentrated on the liver. This result can be explained with Fig.4 and Fig. 5 in a more complementary way. Specifically, as the learned causal graph is medical explainable and with better estimation of intraoperative features, the model is driven to locate on liver-related regions. More visualization results are left in appendix due to space limit." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.531, + 0.684, + 0.545 + ], + "angle": 0, + "content": "4.5. Further Discussion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.554, + 0.895, + 0.81 + ], + "angle": 0, + "content": "Intraoperative prediction. In our method, the estimation of intraoperative indicators are important for final prediction. Fig. 4 shows the metrics of some typical intraoperative prediction by i) \\( w/o \\, CaDAG \\) and ii) \\( w/o \\, Pre \\) with \\( F_1 \\)-Score and accuracy for discrete variants (Fig. 4 (a, b)) and \\( R \\)-square for continuous ones (Fig. 4 (c)). Fig. 4 (a, b) shows that our CaDAG can generally enhance the estimation of categorical variables hepatocirrhosis, cell type, differentiation, etc. Nevertheless, we also find that such improvements will decrease if corresponding indicators are in deep layers of the graph in Fig. 5. For example, the accuracy of \\( G \\)-Score and \\( S \\)-score is only comparable and even worse than that of \\( w/o \\, CaDAG \\). A possible reason is the accumulated error along the directed path for estimation. On the other hand, as shown by the results of \\( R \\)-square (closer to 1, the better the prediction) does not bring about notable improvement in regression tasks, (e.g., ascites and tumor diameter)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.811, + 0.894, + 0.901 + ], + "angle": 0, + "content": "We also observe an evident performance degradation of \\( w / o \\) Pre for most of those discrete or continuous variants. This implies that although operative alone has limited contribution to OS time prediction, it still matters in our CAWIM, because of the less missing information due to shorter time interval from the preoperative stage to the" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15687" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.061, + 0.099, + 0.331, + 0.237 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.168, + 0.241, + 0.188, + 0.255 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.344, + 0.101, + 0.621, + 0.237 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.477, + 0.24, + 0.499, + 0.255 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image", + "bbox": [ + 0.636, + 0.101, + 0.882, + 0.237 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.75, + 0.241, + 0.769, + 0.255 + ], + "angle": 0, + "content": "(c)" + }, + { + "type": "image_caption", + "bbox": [ + 0.075, + 0.256, + 0.894, + 0.312 + ], + "angle": 0, + "content": "Figure 4. Quantity Results of Intraoperative Prediction. Intraoperative prediction are categorized into classification ((a) and (b)) and regression problem ((c)) according to the data form. (a) and (b) shows the prediction \\( \\mathrm{F}_1 \\)-Score and accuracy rate of the discrete variants respectively. The numbers after the variant name in (a) and (b) are the number of classes, i.e. hepatocirrhosis is a 6-category classification task. (c) shows the R-squared score for continuous variants." + }, + { + "type": "image", + "bbox": [ + 0.108, + 0.331, + 0.434, + 0.466 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.106, + 0.47, + 0.435, + 0.484 + ], + "angle": 0, + "content": "Figure 5. Learned DAG over Intraoperative Indexes A." + }, + { + "type": "image", + "bbox": [ + 0.088, + 0.493, + 0.434, + 0.656 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.659, + 0.47, + 0.701 + ], + "angle": 0, + "content": "Figure 6. Heat Maps via Grad-CAM [29] on 5 Patients. (a) shows original images and the bounding box of liver. (b) and (c) shows the area of interest \\( w/o \\) and with CaDAG, respectively." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.718, + 0.469, + 0.748 + ], + "angle": 0, + "content": "surgery stage than to the final stage. This observation validates the effect of using intraoperative features." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.473, + 0.902 + ], + "angle": 0, + "content": "General Framework for OS time prediction. In general, predicting OS time from preoperative images and indexes is medically important but challenging, as there exists a lot of missing information due to a long time interval between the preoperative stage to the final OS stage. To amend this problem, it is informative to leverage some intermediate information. In this paper, we show that with this leverage and the modeling of this information via causal discovery, our CAWIM model enjoys convincing medical interpretability, more concentrated location of liver region, more accurate" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.326, + 0.893, + 0.401 + ], + "angle": 0, + "content": "estimation of intraoperative features, and finally better classification results on OS time. We thus believe that our framework can be beneficial to OS time prediction in other scenarios, e.g., other diseases or intermediate information in addition to intraoperative indexes." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.404, + 0.893, + 0.556 + ], + "angle": 0, + "content": "Limitations. We find in Fig. 5 that although most of learned causal relations are consistent with medical priors, there exist relations that cannot be explained well: it is hard to determine the causal order for some pairs of variables: e.g., the causal relationship between the sum of tumor diameter and cell type. This may due to unobserved confounders between these pairs of variables, which may be alleviated by learning hidden representations that can explain the associations among these pairs of variables. We will leave this exploration in our future work." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.581, + 0.619, + 0.597 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.61, + 0.893, + 0.791 + ], + "angle": 0, + "content": "We propose a novel OS time prediction paradigm, which is the first to leverage intraoperative attributes by causal structure learning. Our method significantly outperforms baselines by a large margin. We demonstrate that this improvement is contributed to highly interpretable learned causal structure, accurate estimation of intraoperative indexes, and identification of disease-related regions. We believe our method, especially the leverage of intraoperative information equipped with causal discovery can potentially benefit other scenarios. For limitation, we shall relax the SCM assumption by allowing the existence of unobserved confounders. We leave it as the future work." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.795, + 0.894, + 0.903 + ], + "angle": 0, + "content": "Acknowledgment. Ethics Committee of Zhongshan Hospital, Fudan University approved the protocol of this study and waived the requirement for patient-informed consent (B2021-325R). The authors are also with Shanghai Key Lab of Intelligent Information Processing, and Fudan ISTBI-ZJNU Algorithm Centre for Brain-inspired Intelligence, Zhejiang Normal University, Jinhua, China." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15688" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.105 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.198 + ], + "angle": 0, + "content": "[1] Shekoofeh Azizi, Basil Mustafa, Fiona Ryan, Zachary Beaver, Jan Freyberg, Jonathan Deaton, Aaron Loh, Alan Karthikesalingam, Simon Kornblith, Ting Chen, et al. Big self-supervised models advance medical image classification. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3478-3488, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.2, + 0.472, + 0.24 + ], + "angle": 0, + "content": "[2] Pierre Bedossa and Thierry Poynard. An algorithm for the grading of activity in chronic hepatitis c. Hepatology, 24(2):289-293, 1996. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.242, + 0.47, + 0.324 + ], + "angle": 0, + "content": "[3] Nicholas A Christakis, Julia L Smith, Colin Murray Parkes, and Elizabeth B Lamont. Extent and determinants of error in doctors' prognoses in terminally ill patients: prospective cohort study commentary: Why do doctors overestimate? commentary: Prognoses should be based on proved indices not intuition. *Bmj*, 320(7233):469–473, 2000. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.326, + 0.47, + 0.381 + ], + "angle": 0, + "content": "[4] Agostino Colli, Mirella Fraquelli, Marco Andreola, Barbara Marino, Enrico Zuccoli, and Dario Conte. Severe liver fibrosis or cirrhosis: accuracy of us for detection—analysis of 300 cases. Radiology, 227(1):89-94, 2003. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.382, + 0.47, + 0.422 + ], + "angle": 0, + "content": "[5] Yin Dai, Yifan Gao, and Fayu Liu. Transmed: Transformers advance multi-modal medical image classification. Diagnostics, 11(8):1384, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.424, + 0.47, + 0.479 + ], + "angle": 0, + "content": "[6] James J Driscoll and Oliver Rixe. Overall survival: still the gold standard: why overall survival remains the definitive end point in cancer clinical trials. The Cancer Journal, 15(5):401-405, 2009. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.481, + 0.47, + 0.548 + ], + "angle": 0, + "content": "[7] J Faivre, D Forman, J Esteve, M Obradovic, M Sant, EURO-CARE Working Group, et al. Survival of patients with primary liver cancer, pancreatic cancer and biliary tract cancer in europe. European Journal of Cancer, 34(14):2184-2190, 1998. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.551, + 0.47, + 0.648 + ], + "angle": 0, + "content": "[8] Ming Fan, Zuhui Liu, Sudan Xie, Maosheng Xu, Shiwei Wang, Xin Gao, and Lihua Li. Integration of dynamic contrast-enhanced magnetic resonance imaging and t2-weighted imaging radiomic features by a canonical correlation analysis-based feature fusion method to predict histological grade in ductal breast carcinoma. Physics in Medicine & Biology, 64(21):215001, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.649, + 0.47, + 0.731 + ], + "angle": 0, + "content": "[9] Sheung Tat Fan, Chung Mau Lo, Ronnie TP Poon, Chun Yeung, Chi Leung Liu, Wai Key Yuen, Chi Ming Lam, Kelvin KC Ng, and See Ching Chan. Continuous improvement of survival outcomes of resection of hepatocellular carcinoma: a 20-year experience. Annals of surgery, 253(4):745-758, 2011. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.733, + 0.47, + 0.788 + ], + "angle": 0, + "content": "[10] Xue Feng, Nicholas J Tustison, Sohil H Patel, and Craig H Meyer. Brain tumor segmentation using an ensemble of 3d u-nets and overall survival prediction using radiomic features. Frontiers in computational neuroscience, 14:25, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.789, + 0.47, + 0.844 + ], + "angle": 0, + "content": "[11] Paul Glare, Kiran Virik, Mark Jones, Malcolm Hudson, Steffen Eychmuller, John Simes, and Nicholas Christakis. A systematic review of physicians' survival predictions in terminally ill cancer patients. *Bmj*, 327(7408):195, 2003. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.845, + 0.47, + 0.901 + ], + "angle": 0, + "content": "[12] Arthur Gretton, Olivier Bousquet, Alex Smola, and Bernhard Schölkopf. Measuring statistical dependence with hilbert-schmidt norms. In International conference on algorithmic learning theory, pages 63-77. Springer, 2005. 3, 4" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.147 + ], + "angle": 0, + "content": "[13] Arthur Gretton, Kenji Fukumizu, Choon Teo, Le Song, Bernhard Scholkopf, and Alex Smola. A kernel statistical test of independence. Advances in neural information processing systems, 20, 2007. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.15, + 0.892, + 0.219 + ], + "angle": 0, + "content": "[14] Lu Guo, Ping Wang, Ranran Sun, Chengwen Yang, Ning Zhang, Yu Guo, and Yuanming Feng. A fuzzy feature fusion method for auto-segmentation of gliomas with multimodality diffusion and perfusion magnetic resonance images in radiotherapy. Scientific reports, 8(1):1-11, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.221, + 0.892, + 0.276 + ], + "angle": 0, + "content": "[15] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.278, + 0.892, + 0.332 + ], + "angle": 0, + "content": "[16] Mohammad Hesam Hesamian, Wenjing Jia, Xiangjian He, and Paul Kennedy. Deep learning techniques for medical image segmentation: achievements and challenges. Journal of digital imaging, 32(4):582-596, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.334, + 0.892, + 0.389 + ], + "angle": 0, + "content": "[17] Karin W Houben and John L McCall. Liver transplantation for hepatocellular carcinoma in patients without underlying liver disease: a systematic review. *Liver Transplantation and Surgery*, 5(2):91–95, 1999. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.392, + 0.892, + 0.46 + ], + "angle": 0, + "content": "[18] Biwei Huang, Kun Zhang, Jiji Zhang, Joseph D Ramsey, Ruben Sanchez-Romero, Clark Glymour, and Bernhard Scholkopf. Causal discovery from heterogeneous/nonstationary data. J. Mach. Learn. Res., 21(89):1-53, 2020. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.463, + 0.892, + 0.504 + ], + "angle": 0, + "content": "[19] Markus Kalisch and Peter Buhlman. Estimating high-dimensional directed acyclic graphs with the pc-algorithm. Journal of Machine Learning Research, 8(3), 2007. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.506, + 0.892, + 0.602 + ], + "angle": 0, + "content": "[20] Christophe Laurent, Jean Frédéric Blanc, Steele Nobili, Antonio Sa Cunha, Brigitte le Bail, Paulette Bioulac-Sage, Charles Balabaud, Maylis Capdepont, and Jean Saric. Prognostic factors and longterm survival after hepatic resection for hepatocellular carcinoma originating from noncirrhotic liver. Journal of the American College of Surgeons, 201(5):656-662, 2005. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.605, + 0.892, + 0.673 + ], + "angle": 0, + "content": "[21] Qing Li, Weidong Cai, Xiaogang Wang, Yun Zhou, David Dagan Feng, and Mei Chen. Medical image classification with convolutional neural network. In 2014 13th international conference on control automation robotics & vision (ICARCV), pages 844-848. IEEE, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.676, + 0.892, + 0.73 + ], + "angle": 0, + "content": "[22] Tao Li, Wu Li, Yehui Yang, and Wensheng Zhang. Classification of brain disease in magnetic resonance images using two-stage local feature fusion. PloS one, 12(2):e0171749, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.733, + 0.892, + 0.813 + ], + "angle": 0, + "content": "[23] Dong Nie, Han Zhang, Ehsan Adeli, Luyan Liu, and Dinggang Shen. 3d deep learning for multi-modal imaging-guided survival time prediction of brain tumor patients. In International conference on medical image computing and computer-assisted intervention, pages 212-220. Springer, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.817, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[24] Byeonggwan Noh, Young Mok Park, Yujin Kwon, Chang In Choi, Byung Kwan Choi, Yo-Han Park, Kwangho Yang, Sunju Lee, Taeyoung Ha, YunKyong Hyon, et al. Machine learning-based survival rate prediction of korean hepatocellular carcinoma patients using multi-center data. BMC gastroenterology, 22(1):1-9, 2022. 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "15689" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.147 + ], + "angle": 0, + "content": "[25] Augustus Odena, Christopher Olah, and Jonathon Shlens. Conditional image synthesis with auxiliary classifier gans. In International conference on machine learning, pages 2642-2651. PMLR, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.149, + 0.468, + 0.175 + ], + "angle": 0, + "content": "[26] Judea Pearl. Causality. Cambridge university press, 2009. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.178, + 0.468, + 0.204 + ], + "angle": 0, + "content": "[27] Markus Peck-Radosavljevic. Drug therapy for advanced-stage liver cancer. Liver cancer, 3(2):125-131, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.206, + 0.468, + 0.272 + ], + "angle": 0, + "content": "[28] Whitney B Pope, James Sayre, Alla Perlina, J Pablo Villablanca, Paul S Mischel, and Timothy F Cloughesy. Mr imaging correlates of survival in patients with high-grade gliomas. American Journal of Neuroradiology, 26(10):2466-2474, 2005. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.275, + 0.468, + 0.355 + ], + "angle": 0, + "content": "[29] Ramprasaath R. Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra. Grad-cam: Visual explanations from deep networks via gradient-based localization. In 2017 IEEE International Conference on Computer Vision (ICCV), pages 618-626, 2017. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.359, + 0.468, + 0.425 + ], + "angle": 0, + "content": "[30] Wei Shao, Tongxin Wang, Liang Sun, Tianhan Dong, Zhi Han, Zhi Huang, Jie Zhang, Daoqiang Zhang, and Kun Huang. Multi-task multi-modal learning for joint diagnosis and prognosis of human cancers. Medical Image Analysis, 65:101795, 2020. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.428, + 0.468, + 0.482 + ], + "angle": 0, + "content": "[31] Pater Spirtes, Clark Glymour, Richard Scheines, Stuart Kauffman, Valerio Aimale, and Frank Wimberly. Constructing bayesian network models of gene expression networks from microarray data. 2000. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.485, + 0.468, + 0.537 + ], + "angle": 0, + "content": "[32] Wen Tang, Haoyue Zhang, Pengxin Yu, Han Kang, and Rongguo Zhang. Mmmna-net for overall survival time prediction of brain tumor patients. arXiv preprint arXiv:2206.06267, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.54, + 0.468, + 0.594 + ], + "angle": 0, + "content": "[33] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.596, + 0.468, + 0.69 + ], + "angle": 0, + "content": "[34] Jing-Houng Wang, Chi-Sin Changchien, Tsung-Hui Hu, Chuan-Mo Lee, Kwong-Ming Kee, Chih-Yun Lin, Chao-Long Chen, Tai-Yi Chen, Yu-Jie Huang, and Sheng-Nan Lu. The efficacy of treatment schedules according to Barcelona clinic liver cancer staging for hepatocellular carcinoma-survival analysis of 3892 patients. European journal of cancer, 44(7):1000-1006, 2008. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.693, + 0.468, + 0.734 + ], + "angle": 0, + "content": "[35] Samir S Yadav and Shivajirao M Jadhav. Deep convolutional neural network based medical image classification for disease diagnosis. Journal of Big Data, 6(1):1-18, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.735, + 0.468, + 0.815 + ], + "angle": 0, + "content": "[36] Naoki Yamanaka, Eizo Okamoto, Tsuyosi Oriyama, Jiro Fujimoto, Kazutaka Furukawa, Eisuke Kawamura, Tsuneo Tanaka, and Fumito Tomoda. A prediction scoring system to select the surgical treatment of liver cancer. further refinement based on 10 years of use. Annals of surgery, 219(4):342, 1994. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.818, + 0.468, + 0.858 + ], + "angle": 0, + "content": "[37] Jianpeng Zhang, Yutong Xie, Qi Wu, and Yong Xia. Medical image classification using synergic deep learning. Medical image analysis, 54:10-19, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.861, + 0.468, + 0.9 + ], + "angle": 0, + "content": "[38] Fan Zhou, Tengfei Li, Heng Li, and Hongtu Zhu. Tpcnn: two-phase patch-based convolutional neural network for automatic brain tumor segmentation and survival prediction." + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "In International MICCAI Brainlesion Workshop, pages 274-286. Springer, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.122, + 0.892, + 0.177 + ], + "angle": 0, + "content": "[39] Tao Zhou, H. Fu, Yu Zhang, Changqing Zhang, Xiankai Lu, Jianbing Shen, and Ling Shao. M2net: Multi-modal multichannel network for overall survival time prediction of brain tumor patients. ArXiv, abs/2006.10135, 2020. 2, 3" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.892, + 0.177 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15690" + } + ] +] \ No newline at end of file diff --git a/2023/Causally-Aware Intraoperative Imputation for Overall Survival Time Prediction/58789d3b-4b75-4c2b-ad82-075b8268b238_origin.pdf b/2023/Causally-Aware Intraoperative Imputation for Overall Survival Time Prediction/58789d3b-4b75-4c2b-ad82-075b8268b238_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..2e2806365b872bcb495587f4af4514d015b5a2de --- /dev/null +++ b/2023/Causally-Aware Intraoperative Imputation for Overall Survival Time Prediction/58789d3b-4b75-4c2b-ad82-075b8268b238_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:16f7e75493ae5f72da3038f392f3950c1a486031e1458cd032fee45084e50bdf +size 1048868 diff --git a/2023/Causally-Aware Intraoperative Imputation for Overall Survival Time Prediction/full.md b/2023/Causally-Aware Intraoperative Imputation for Overall Survival Time Prediction/full.md new file mode 100644 index 0000000000000000000000000000000000000000..4c8584609d0c83f89e180c3d514f3446570d95e7 --- /dev/null +++ b/2023/Causally-Aware Intraoperative Imputation for Overall Survival Time Prediction/full.md @@ -0,0 +1,267 @@ +# Causally-Aware Intraoperative Imputation for Overall Survival Time Prediction + +Xiang Li $^{1,*}$ , Xuelin Qian $^{1,*}$ , Litian Liang $^{1,*}$ , Lingjie Kong $^{1}$ , Qiaole Dong $^{1}$ , Jiejun Chen $^{2}$ , Dingxia Liu $^{2}$ , Xiuzhong Yao $^{2,\dagger}$ , Yanwei Fu $^{1,\dagger}$ + +$^{1}$ School of Data Science, Fudan University + +$^{2}$ Department of Radiology, Zhongshan Hospital, Fudan University + +{li_x20, xlqian, ltliang19, qldong18, jjchen20, yanweifu}@fudan.edu.cn + +{jkong22,dxliu21}@m.fudan.edu.cn yao.xiuzhong@zs-hospital.sh.cn + +![](images/0738ecee42815d685eaa17534391c0dc9b1ca91398d3944bca76e79b91c8f7a7.jpg) +Figure 1. Our idea illustration. Patient A and B have different OS time (109 months and 16 months, respectively) but similar pre-operative image-based patterns and indexes; thus, preoperative based model fails to distinguish A and B (their predicted OS time are both larger than 9 years). However, their intra-operative attributes, which can describe the severity of disease in a more informative way, are very different from each other (e.g., G-score, S-score, Clinicopathologic hepatocirrhosis). Therefore, it can help the model discriminate A and B. In the right image, we show that our CAWIM – that leverages intra-operative indexes in the training stage – can largely improve the prediction power than other methods. Our method managed to correctly classify these two cases, and the overall performance of our CAWIM surpasses other baseline models by approximately 10 points on 4-category classification task. + +# Abstract + +Previous efforts in vision community are mostly made on learning good representations from visual patterns. Beyond this, this paper emphasizes the high-level ability of causal reasoning. We thus present a case study of solving the challenging task of Overall Survival (OS) time in primary liver cancers. Critically, the prediction of OS time at the early stage remains challenging, due to the unobvious image patterns of reflecting the OS. To this end, we propose a causal inference system by leveraging the intraoperative attributes and the correlation among them, as an intermediate supervision to bridge the gap between the images and the final OS. Particularly, we build a causal graph, and train the images to estimate the intraoperative attributes for final OS prediction. We present a novel Causally-aware Intraoperative Imputation Model (CAWIM) that can sequentially predict each attribute using its parent nodes in the estimated causal graph. To determine the causal directions, + +we propose a splitting-voting mechanism, which votes for the direction for each pair of adjacent nodes among multiple predictions obtained via causal discovery from heterogeneity. The practicability and effectiveness of our method are demonstrated by the promising results on liver cancer dataset of 361 patients with long-term observations. + +# 1. Introduction + +The success of recent deep learning model is largely attributed to learning the good representations for visual patterns. Such representations essentially facilitate various vision task, such as recognition and synthesis [15, 25, 33]. Nevertheless, one important goal for the vision community is to model and summarize the relationships of observed variables of a system, in order to enable well predictions on similar data. Essentially, it is desirable to understand how the system is changed if one modifies these relationships under certain conditions, e.g., the effects of a treatment in healthcare. Thus this demands the high- + +level ability of causal reasoning beyond the previous efforts of only learning good representations for visual patterns [1, 5, 16, 21, 35, 37]. This naturally leads into our task of causal inference. + +This paper presents a case study of solving the challenging task of Overall Survival (OS) time estimation in Primary Liver Cancers (PLC). Generally, the liver cancer remains one of the most common malignancies worldwide in the 21st century, as there are about one million new cases every year [36]. The five-year survival rate for advanced PLC is only about $5\%$ [7]. Therefore, early and accurate prediction of OS time estimation can provide informative guidance for individualized treatment planning and reducing burden of medical resources [6, 27, 32, 34]. On the other hand, one shall easily notice that with the renaissance of deep learning, great achievements have been made on medical imaging analysis, such as diagnosis, segmentation, and classification [16, 21, 35, 37]. Unfortunately, it still remains challenging for experienced clinicians to predict OS time at early stage, even with advanced modern diagnostic tools such as Magnetic Resonance Imaging (MRI) or tumor marker tests [3, 11], and the deep learning tools [1, 5]. + +Some studies propose to leverage deep neural networks for OS time prediction. They focus on fusion learning of multi-modal image features and some basic information (i.e., age and gender) [5, 8, 14, 22, 24, 30, 39]. Nevertheless, the accurate prediction based on only preoperative information (such as image and tumor marker indexes) is still challenging, possibly due to missing information from early diagnosis stage to the final stage. This missing information includes the texture and pathological attributes of the liver, health level of the patient, and post-operative treatment [9, 17, 20]. For example, as in Fig. 1, patient A and B with different OS time have almost identical preoperative indicators cause, history of disease, etc., making the preoperative-based model hard for discrimination. + +To amend this problem, we present a causal inference system that can well utilize the intraoperative information, which is pretty easy to be accessed in training data. According to medical priorities, such information records pathological attributes, which can be more reflective about the patient's health level and the postoperative recovery. Inspired by this, we propose to leverage this auxiliary information to help build our causal inference system. Again, we take the example in Fig. 1. Although preoperative information cannot differentiate patient A from B, their intraoperative index shows great difference in distribution, which can thus be employed for the discrimination. Additionally, there are many indicators from medical experts that the intraoperative indexes are related to each other. For instance, the clinicopathologic hepatocirrhosis is dependent on the hepatocirrhosis; the sum of tumor diameter is affected by the number of tumors; the fibrosis (reflected on S-Score) can prob + +ably deteriorate to cirrhosis [4], etc. By leveraging these relationships, we can better understand the causal inference concrete from both observed data modelling and medical expert-level knowledge of these variables. + +To this end, we encapsulate these priors and the inspired proposals into a new method, dubbed as Causally-Aware Intraoperative Imputation Model (CAWIM). It incorporates causal discovery module to sequentially estimate intraoperative indexes as an intermediate stage towards final OS time prediction. Specifically, our model is composed of two key steps: i) estimating the intraoperative indexes using preoperative information, i.e., image and indexes; ii) followed by OS time prediction using estimated intraoperative indexes and preoperative information. To achieve more accurate prediction of intraoperative indexes that is determinant to the prediction power of the whole method, we propose a Causally-aware Directed Acyclic Graph (CaDAG) module. It learns the causal structure represented as a DAG over intraoperative features. To identify the causal relations beyond the traditional PC algorithm [26], we propose a splitting-voting mechanism, which is inspired by the recent work [18] that learn the causal structure with the assistance of an auxiliary domain index variable. Our proposed mechanism can not only identify the causal relations even when this domain index variable is not available, but also can be theoretically guaranteed that the learned graph is not acyclic. During test stage, we sequentially estimate each intraoperative index with preoperative information and additionally, its parent set among other intraoperative indexes. The utility of our method can be demonstrated by a significant improvement of OS time prediction on an in-house liver cancer dataset, as well as better interpretability of learned causal structure, more accurate estimation of intraoperative indexes and more interpretable visualization results. + +In a nutshell, we for the first time present a case study of building a causally-aware intraoperative imputation system for the challenging task of overall survival time prediction. The proposed method of building the casual inference system, can be naturally extended to other similar medical tasks. Our key contributions are listed as follows. (1) New Paradigm for OS time Prediction. We propose to leverage intraoperative indexes as an intermediate stage during training. The leverage of this information can significantly alleviate the "missing information" issue. To the best of our knowledge, we are the first to leverage auxiliary information (in addition to preoperative features) for OS time prediction. (2) Causal Structure Learning. We propose a novel splitting-voting mechanism that can identify the causal structure even when the domain index variable is missing. (3) Better Prediction Power. Our method can significantly improve the prediction power for liver cancer over the competitors. The methods are evaluated on the medical + +dataset, which, to the best of our knowledge, is the largest primary live cancer dataset. + +# 2. Related Work and Preliminaries + +Previous efforts [23,39] directly take the OS time prediction as the vanilla classification task by employing various deep models, whilst we reformulate it by causal discovery method for the first time. We will review the most related literature here. + +MRI based deep models. OS time is defined as the duration from patient's first scan to cancer-related death [6]. Since Magnetic Resonance Image (MRI) plays an important role in tumor-related studies, various MRI based deep learning model have been proposed for feature extraction and representing [28], multi-modal feature fusion [10, 30, 38, 39], and improvement on model structure [5]. However, since missing information during disease progression can mediate the correlation from the preoperative stage to the final stage, the results of these image based methods are typically unsatisfied. In contrast, our method can amend this problem by leveraging intraoperative indexes as a bridge. + +Causal discovery. Causal discovery is to learn the causal graph over endogenous variables. Typical methods include PC algorithm [19], a constraint-based method that implement iterative conditional independence test to identify the skeleton and directions via v-structure. Under Markovian and faithfulness assumptions [26, 31], this method is provable to identify an equivalent class of the causal graph. To further identify more directions, recent work [18] leveraged a domain index variable that splits the dataset into multiple domains. With these heterogeneous data, they can identify mutable variables that change across the domain index variable. On the basis of this, they further identified edges among these mutable variable set by implementing Hilbert Schmidt Independence Criterion (HSIC) norm [13]. However, such domain index may not be available in many scenarios, making the identification of these directions difficult. Besides, this method may not ensure that the learned graph is acyclic. To address these issues, we in this paper propose a splitting-voting mechanism to mine the causal relations among intraoperative indexes so as for leveraging each other to better estimate these indexes, which is provable to return a directed acyclic graph. + +Preliminaries. (1) Causal Discovery. By faithfulness and Markovian assumptions [26], we can use conditional independent tests for causal discovery. That is learning the causal graph over $\mathbf{V}$ . (2) Structural causal model (SCM). Before fully developing our contributions in methodology, we give as preliminary the definition of SCM, which is defined as $\langle G, \mathcal{F}, P\langle \epsilon \rangle \rangle$ : i) directed acycle graph $G = (\mathbf{V}, \mathbf{E})$ is as causal structure, with node set $\mathbf{V}$ and edge set $\mathbf{E}$ ; ii) the autonomous structural functions $\mathcal{F} = \{f_k\}_{V_k \in V}$ , where disturbance on $V_k$ has no effect on others; iii) probability + +measure $P(\epsilon)$ for exogenous variables $\{\epsilon_k\}_k$ . Given the assumption that $\{\epsilon_k\}_k$ are independent, each $\mathrm{P}$ can be obtained by Causal Markov Condition with $G$ as $\mathrm{P}(\{V_K = v_k\}_{V_k \in \mathbf{V}}) = \prod_k \mathrm{P}(V_k = v_k | Pa(k) = pa(k))$ . (3) Causal discovery from heterogeneity. In [18], the authors proposed to leverage a domain index variable $C$ to identify some causal directions. This method is built upon the PC and faithfulness assumption at a distributional level. It first learns the skeleton of the graph and identifies a variable set that is affected by $E$ , followed by identifying the directions between $E$ and its neighbors. Specifically, if $C \to V_i, C \to V_j$ , and $V_i \bot V_j | X, C$ for some deconfounding set $X$ between $V_i$ and $V_j$ , $\{P(V_i | X, C = c)\}$ is independent of $\{P(V_j | V_i, C = c, X)\}$ , while $\{P(V_j | X, C = c)\}$ and $\{P(V_j | V_i, C = c, X)\}$ are dependent. We can test these Independence using Hilbert Schmidt Independence Criterion (HSIC) [12] norms $\delta_{v_i \to v_j | X}$ and $\delta_{v_j \to v_i | X}$ . One can obtain that $V_i \to V_j$ if $\delta_{v_i \to V_j} < \alpha$ with pre-set significance level $\alpha$ . However, in many applications such domain index variable is missing, i.e., all data are pooled together. In this paper, we propose a splitting-voting mechanism (in the CaDAG module) to identify causal relations. In the following, we first introduce some basic assumptions that our method is built upon. + +# 3. Methodology + +Problem setting. For the OS time classification task, suppose we have $\{X_{i},\mathbf{B}_{i},\mathbf{A}_{i},y_{i}\} \sim_{i.i.d}P(X,\mathbf{B},\mathbf{A},y)$ where $X$ denotes the image acquired from structural Magnetic Resonance Image (sMRI); $\mathbf{B}$ , $\mathbf{A}$ denote the preoperative and intraoperative attributes, respectively. We have the final label $y$ . Our goal is to predict $y$ from the image $X$ and preoperative attributes $\mathbf{B}$ that are recorded before surgery. Our method is built upon the following three assumptions, commonly utilized in the causal inference works [18, 26]. + +Definition 1 (Causal Graph) We assume the causal graph over $\mathbf{A}$ is a directed acyclic graph (DAG) and denote the corresponding SCM as $M\coloneqq \langle G\coloneqq (\mathbf{A},\mathbf{E}),\mathcal{F},P(\varepsilon)\rangle$ + +# Definition 2 (Markovian and Faithfulness) For + +triplets of disjoint sets $\mathbf{V}_i, \mathbf{V}_j, \mathbf{V}_k$ , it holds that $\mathbf{V}_i \perp_d \mathbf{V}_j | \mathbf{V}_k \leftrightarrow \mathbf{V}_i \perp \mathbf{V}_j | \mathbf{V}_k$ , where $\perp_d$ and $\perp$ respectively mean $d$ -separation and probability independence. This is the common property in DAG. + +Definition 3 (Distributional Faithfulness) If $X_{i} \to X_{j}$ and at least $E \to V_{i}$ or $E \to V_{j}$ holds, $\{P^{e}(V_{i}|V_{j},\mathbf{Z})\}$ is dependent to $\{P^{e}(V_{j}|\mathbf{Z})\}$ , where $\mathbf{Z}$ denotes the minimal deconfounding set. Particularly, $\mathbf{Z}$ is a deconfounding set between $V_{i}$ and $V_{j}$ if we have $V_{i} \perp V_{j}|\mathbf{Z}$ and $\mathbf{Z} \cap (\mathrm{De}(V_i) \cup \mathrm{De}(V_j)) = \emptyset$ . + +In term of this assumption, we further have the following result for edge orientation [18]. + +Theorem 4 (Theorem 2 in [18]) Denote $E$ as domain index variable. Under assumptions 1, 2, 3, for each adjacent pair $(V_i, V_j)$ such that $E \to V_i$ or $E \to V_j$ holds, we have $V_i \to V_j$ if $\{P^e(V_i|\mathbf{Z})\}$ and $\{P^e(V_j|V_i,\mathbf{Z})\}$ are independent while $\{P^e(V_j|\mathbf{Z})\}$ and $\{P^e(V_i|V_j,\mathbf{Z})\}$ are dependent. + +# 3.1. Causally-aware Directed Acyclic Graph + +The Causally-aware DAG. We propose a novel Causally-Aware Intraoperative Imputation Model (CAWIM). The pipeline of our method is shown in Fig. 2. It is composed of Causally-aware Directed Acycled Graph (CaDAG), and a classification model with MRI imaging encoder. The CaDAG enables the causally-aware intraoperative reasoning, by managing to better represent and encode the correlation of intraoperative information. Specifically, we first use preoperative information, i.e., MRI images and indexes to estimate the intraoperative features. The estimated intraoperative features $\tilde{\mathbf{A}}$ , together with the MRI $X$ and preoperative features $\mathbf{B}$ are then utilized for OS classification. To accurately estimate intraoperative indexes $\mathbf{A}$ , we propose CaDAG to learn causal structure over $\mathbf{A}$ , and then leverage the parent nodes of each index for sequentially imputation during test stage. We give the details about the CaDAG and the intraoperative imputation model for classification, while some theoretical proofs, and the specification of our networks are in the appendix. +We aim at learning the causal structure over $\mathbf{A}$ . However, without the domain index, we actually have the difficulty of dividing the whole dataset into multiple domains. To this end, we propose a splitting-voting mechanism for this problem. Particularly, this mechanism first generates the domain-index variable $E$ by randomly splitting the whole dataset into multiple domains for $m > 0$ times. Then for each time we identify the mutable variable set and generate a prediction for each identifiable causal directions, with the assistance of domain index variable $E$ . Finally, we vote for a final direction among $m$ predictions for each pair of variables. As shown in Fig. 3, the whole procedure is a sequence of the four steps. Step i), ii) and iii) provide preliminary causal relationship between each pair of nodes. Step iv) is our proposed splitting-voting, aiming to finally determine all the causal directions and output a causal DAG. +i) Learning skeleton of the causal graph. We first implement PC algorithm to learn the skeleton [26]. That is an undirected graph over all attributes. This can be achieved by iterative conditional independence test. Then we determine $A_{i}$ and $A_{j}$ to be adjacent if they are not independent conditional on any subsets. Finally, we utilize the v-structure to learn the equivalence class of DAG. +ii) Random splitting and identification of mutable variable sets. To determine more directions, we randomly split the dataset into multiple domains for $m$ times. For each time we have a domain index variable $E_{i}$ and multiple do + +mains $\{\mathcal{D}_e^i\}$ [18]; and we identify the mutable variable set $\mathbf{M}_i$ such that each $X \in \mathbf{M}_i$ has $X \not\perp E_i|\mathbf{A} - X$ is conditionally dependent to $E_i$ . We thus can implement conditional independent tests to identify $\mathbf{M}_i$ . Steps i) and ii) are summarized in Alg. 1. + +# Algorithm 1 Identify the skeleton and mutable set. + +INPUT: $\{\mathcal{D}_e^i\coloneqq \{\mathbf{A}_i^e\}_{i = 1}^{n_e}|e\}$ , domain index variable $E_{i}$ . + +OUTPUT: Equivalent class of DAG and $\mathbf{M}_i$ . + +1: Implement the PC algorithm to learn the equivalence class of DAG via v-structure. +2: For each $A \in \mathbf{A}$ , add the edge $E_i \to A$ in the graph iff $E_i \not\perp A$ given any subsets. + +iii) Identifying directions via changed causal models. For each adjacent pair $A_{i}$ and $A_{j}$ such that at least one of them is adjacent to $E$ , we identify the causal directions among $\mathbf{M}_i$ via HSIC Norm [12] by testing the independence between $\{P^{e}(A_{i}|A_{j},\mathbf{Z})\}$ and $\{P^{e}(A_{j}|\mathbf{Z})\}$ if $A_{i}\gets A_{j}$ , and that between $\{P^{e}(A_{j}|A_{i},C)\}$ and $\{P^{e}(A_{i}|\mathbf{Z})\}$ for determining the direction between $A_{i}$ and $A_{j}$ . Here, $\mathbf{Z}$ denotes the deconfounding set. In this regard, the direction between $A_{i}$ and $A_{j}$ is then determined according to the following rule: if $\hat{\Delta}_{A_i\to A_j|\mathbf{Z},E} < \hat{\Delta}_{A_j\to A_i|C,E}$ , output $A_{i}\rightarrow A_{j}$ ; if $\hat{\Delta}_{A_i\to A_j|\mathbf{Z},E} > \hat{\Delta}_{A_j\to A_i|\mathbf{Z},E}$ , output $A_{i}\leftarrow A_{j}$ ; if $\hat{\Delta}_{A_i\to A_j|\mathbf{Z},E} > \alpha$ and $\hat{\Delta}_{A_j\to A_i|\mathbf{Z},E} > \alpha$ , where $\alpha$ is a threshold, making the direction undetermined. + +# Algorithm 2 Identify causal directions among M. + +INPUT: $E$ ; skeleton and $M$ via Alg.1. + +OUTPUT: Directed graph among M. + +1: For each adjacent $(A_{i},A_{j})$ such that one of $V_{i}\in \mathbf{M}$ +2: Detect deconfounding set $\mathbf{Z}$ +3: Calculate $\hat{\Delta}_{A_i\to A_j|\mathbf{Z},E} > \alpha$ and $\hat{\Delta}_{A_j\to A_i|\mathbf{Z},E} > \alpha$ +4: Determine $A_{i} \to A_{j}$ or $A_{j} \to A_{i}$ . + +iv) Voting for the direction of each edge. For each time $i$ we implement Alg. 2 to predict an direction of the edge $(A_{i}, A_{j})$ . Then we vote for $A_{i} \rightarrow A_{j}$ if the frequency $f_{i \rightarrow j}$ of $A_{i} \rightarrow A_{j}$ surpasses a pre-set threshold $\alpha$ ( $0 < \alpha < 1$ ). Finally, if there exists a cycle of $N$ variables: $A_{i_1} \rightarrow A_{i_2} \rightarrow \ldots \rightarrow A_{i_1}$ , we orient $(A_{i_k}, A_{i_{k+1}})$ with the minimum frequency among $\{f_{i_k, i_{k+1}}, f_{i_N, i_1}\}$ . + +We show that this procedure can generate a DAG over $\mathbf{A}$ , with the proof in appendix. + +Theorem 5 Under definitions and assumptions 1, 2, 3, the learned graph via our CaDAG is a directed acyclic graph. + +![](images/084fd24618bb2d6424aa8db962c2987f2924f680a8e6a27dc2f13c2f73f95380.jpg) +Figure 2. Overview of our CAWIM model. We concatenate the MRI features encoded by the MRI Imaging Encoder, preoperative indicators, and the estimated intraoperative indicators obtained from CaDAG to perform the final prediction on OS time. + +![](images/c6333c92c34b3ea843d0f41e1738c5c833cd42cf5b4bbda28713c5d270735f4e.jpg) +Figure 3. Schematic diagram of a four-variable case $A_{1}$ , $A_{2}$ , $A_{3}$ , and $A_{4}$ , with the real DAG shown in the top left corner. Through step i) and ii) in the upper part, a Markov Equivalence Class of 3 causal graphs is observed with the direction between $A_{1}$ and $A_{2}$ undetermined. For step iii), we split the dataset into $(n = 3)$ domains and consider the 4 circumstances shown fig(a) to (d) in the lower left part. For step iv), we vote for the direction between $A_{1}$ and $A_{2}$ as $A_{1} \rightarrow A_{2}$ according to the number of splitting $(n = 3)$ , frequency $f_{1\rightarrow 2} = 2 / 3$ and threshold $\alpha = 2 / 3$ , as shown in the lower right part. Finally, we obtain an inferred DAG with each direction determined. + +# 3.2. Intraoperative Imputation for Classification + +Given MRI image $X \in \mathbb{R}^{H \times W}$ of a patient that contains multiple slides, we first resample these slides into a fixed $K$ slides for each patient. As each slice is a gray image, we repeat the images for 3 times along the channel dimension; and then we employ a neural network to extract features from each slice. The extracted features are denoted as $F_{1}, \ldots, F_{K}$ . In order to aggregate the information of all slices, we divide $F_{1}, \ldots, F_{K}$ into $g$ groups according to the acquired position. Each group contains consecutive $\lfloor K / g \rfloor$ slices. Then, we implement a Multi-head Cross-Attention (MCA) mechanism inside each group to aggregate features: + +$$ +\operatorname {h e a d} _ {j} = \operatorname {s o f t m a x} \left(Q ^ {j} \left(F \mathbf {W} _ {K} ^ {j}\right) ^ {T}\right) \left(F \mathbf {W} _ {V} ^ {j}\right) \tag {1} +$$ + +$$ +\mathbf {M C A} = \left[ \text {h e a d} _ {1}; \dots ; \text {h e a d} _ {h} \right] \mathbf {W} _ {O} \tag {2} +$$ + +where $F$ is the slice's feature within a group, the query vector $Q^{j}$ and linear projection layers $\mathbf{W}_K^j$ , $\mathbf{W}_V^j$ , $\mathbf{W}_O$ , $1 \leq j \leq h$ are all learnable parameters. Finally, we concatenate $\{F_k\}$ from $g$ groups as our final image feature $I \in \mathbb{R}^{g \times d}$ with $d$ denoting the dimension of each $F_k$ with $k = 1, \ldots, K$ . + +Afterwards, we sequentially estimate each $A_{j}$ by preoperative information, as our learned causal graph over $\mathbf{A}$ . So + +for each $A_{j}$ , we predict + +$$ +\hat {A} _ {j} = f _ {\text {i n t r o}} ^ {j} (\mathbf {B}, I, \operatorname {P a} (A _ {j})) \tag {3} +$$ + +where $\mathrm{Pa}(A_j)$ denotes the parent node set of $A_{j}$ . We use cross entropy loss to optimize $f_{\mathrm{intra}}^j$ if $A_{j}$ is categorical; and optimize $f_{\mathrm{intra}}^j$ if $A_{j}$ is continuous. After obtaining all predicted $\{\hat{A}_i\}_{i = 1}^d$ , we use $(I,\mathbf{B},\hat{\mathbf{A}})$ to predict $Y$ via $f_{\theta}(I,\mathbf{B},\hat{\mathbf{A}})$ and obtain $\theta$ by optimizing the cross-entropy $\mathcal{L}(y,f_{\theta}(I,\mathbf{B},\hat{\mathbf{A}}))$ . During test stage, given a new sample $(x,\mathbf{b})$ , we first extract $I$ and then sequentially estimate $\mathbf{a}_i$ and finally predict $y$ using $f_{\theta}(x,\mathbf{b},\hat{\mathbf{a}})$ . + +# 4. Experiments + +Dataset overview. Ethics Committee of Zhongshan Hospital, Fudan University approved the protocol of this study and waived the requirement for patient-informed consent (B2021-325R). Because open source data sets generally lack sufficient intraoperative information, it takes us around ten years to collect and build our own medical dataset. Particularly, we conduct a search through the medical records in the hospital information system, and build a dataset with + +439 patients infected by primary liver cancer. Long-term medical data of every patient has been recorded from the first scan to the cancer-related death. Till the submission of this manuscript, around $90\%$ patients have unfortunately passed away. This gives us the ground-truth OS labels. Information like MRI scanning before the surgery, blood test and operation reports, follow-up records are also collected. After removing samples with missing information, 361 patients were enrolled, with 306 $(84.8\%)$ men, and the average age was 53.7 $(\pm 11.7)$ years. Specifically, each patient has MRI scanning with four modalities, including $\mathrm{T}_{1}$ weighted images $(\mathrm{T}_{1}$ -WI), $\mathrm{T}_{2}$ weighted images $(\mathrm{T}_{2}$ -WI), $\mathrm{T}_{1}$ in-phase $(\mathrm{T}_{1}$ -IP) and contrast-enhanced $\mathrm{T}_{1}$ at the arterial phase $(\mathrm{T}_{1}$ ce-AP). Following the guidance of medical experts, we select 11 preoperative and 20 intraoperative indicators that are related to OS time for our study. Please refer to the appendix for details. + +Data preprocessing. In our dataset, the survival time ranges from 0 to 130 months, which are cauterized into four classes: short-term survival ( $\leq 36$ months), middle-short-term survival (between 36 and 72 months), middle-long-term survival (between 72 and 108 months), and long-term survival ( $\geq 108$ months). In our paper, we consider the classification task: given a new sample, our goal is to classify which class of OS time this sample belongs to. To extract $I$ from MRI image, we resample $K = 20$ slices and employ Resnet-34 for feature extraction. We set $g = 4$ and the dimension $F_{k}$ as 512. For $f_{\mathrm{intro}}^{j}$ and $f_{\theta}$ , we employ a multi-layer perceptron (MLP) for parameterization. + +We encode categorical covariates into dummy variables and implement zero-mean normalization for continuous variables. For MRI images, we resize them into $256 \times 256$ as input. We split the whole dataset into 5 folds. To remove the effect of imbalance across classes and randomness, we adopt average precision, recall, and $\mathrm{F_1}$ -Score measurements over five folds for classification evaluation. + +Implementation details. We use the ImageNet pretrained Resnet-34 to initialize our CNN encoder. We train 150 epochs for all methods with $\mathrm{T}_{1}$ ce-AP modality, using SGD as the optimizer with learning rate $3e - 3$ and $3e - 2$ for CNN encoder and later MLP, respectively. We set weight decay to $5e - 4$ and nesterov momentum factor to 0.9. Besides, we decrease the learning rate every 60 epochs by a factor of 1/10. We set the Batch size to 8. In our voting mechanism of CaDAG, we in total implement Alg. 2 for seven times and set the voting threshold as 4/7. Our model is trained with Pytorch on NVIDIA GeForce RTX 3090 GPUs. + +Compared Methods. We compare our method with four baseline methods: i) Random: random guess among four classes as a trivial baseline, of which the expected $\mathrm{F_1}$ score is $25\%$ ; ii) MRI: train the end-to-end method with $\mathrm{T_1}$ ce-AP only as input for classification; iii) MRI+pre: additionally take preoperative features as input, on the basis of the + +Table 1. Comparisons of Our CAWIM with Other Baselines. + +
ModelPrecisionRecallF1-Score
(a)Random26.04±2.1024.54±0.8023.15±0.64
(b)MRI40.79±7.9733.45±2.6631.88±3.92
(c)MRI+Pre36.99±5.3035.62±3.5033.83±3.64
(d)gt Intra41.26±2.7441.62±1.9539.77±2.47
(e)CAWIM45.58±5.8643.70±5.8942.21±4.92
+ +method MRI; iv) ground-truth (GT) Intra: only take the ground-truth intraoperative features as input. It is worth to mention that although the intraoperative features are not allowed to use during test stage, the comparison with method iv (i.e., GT Intra) can provide information that to which extent, the intraoperative features can help OS time prediction. + +# 4.1. Results and Analysis + +Table 1 records comparisons of our proposed CAWIM (Tab. 1(d)) with other baseline models. (Tab. 1(a), Tab. 1(b), and Tab. 1(c)). Obviously, our CAWIM constantly outperform others in all metrics. Specifically, CAWIM improved $\text{MRI} + \text{Pre}$ by nearly $10\%$ in terms of $\text{F}_1$ -Score, which can be contributed to the prediction of intraoperative features equipped with causally aware prediction mechanism. Without this mechanism, using MRI and preoperative features only can outperform random guess by only $8.33\%$ , which demonstrates the limited information of the image for OS time. In contrast, it is also interesting to note from Tab. 1(d) that using intraoperative features can be more informative than MRI features, in terms of OS prediction, which is natural as the attributes that described the functions and textures of the liver could affect the postoperative recovery and thus the OS time. Finally, our CAWIM (Tab. 1(e)) that combines the information of intraoperative features and the information from images, can achieve further improvement. + +# 4.2. Ablation Study + +To test the effectiveness of preoperative features, the particularly intraoperative features, and our causally-aware module for OS time prediction, we implement ablation studies and summarize the results in Tab. 2. Specifically, Tab. 2(a), i.e., w/o CaDAG replaces our causally-aware module with an end-to-end estimation network from (MRI, preoperative features) to intraoperative features; Tab. 2(b), i.e., w/o Intra is the same to Tab. 1(c), i.e., MRI+pre in Tab. 1 that directly predict OS time using MRI and preoperative features without estimated intraoperative features; Tab. 2(c), i.e., w/o Pre is the same to our CAWIM except that it does not utilize preoperative features for intraoperative estimation and OS prediction. + +The results in Tab. 2 show that the deletion of each module can lead to a descent of all metrics. Specifically, Tab. 2(a) shows a significant performance drop by $5.6\%$ in $\mathrm{F_1}$ -Score, which indicates the effectiveness of our causally + +aware module in estimating intraoperative features in OS time prediction. This phenomena can be explained by more accurate intraoperative estimation equipped with our causally-aware module, as shown in Fig. 4 of $\mathrm{F_1}$ -score, accuracy and $R$ -squared score. We will leave detailed discussions in Section of "Further Discussion". Tab. 2(b), as the same to MRI+pre in Tab. 1, shows a $8.3\%$ performance drop, which indicates the additional information provided by the MRI image and preoperative features. Finally, the $3.78\%$ degradation of Tab. 2(c) compared to ours validates the effectiveness of preoperative features in OS prediction, which can also be explained by the estimation results of intraoperative features in Fig. 4. In a word, we can observe a significant drop of $F_{1}$ -score if we remove causally-aware module and intraoperative features, which can demonstrate the effectiveness of our methods. + +Besides, it is also interesting to compare Tab. 2(b) and Tab. 2(c) with we observe that the performance drop for $w/o$ Intra is more significant than that for $w/o$ Pre, with respectively an average of $8.4\%$ and $3.8\%$ drop in $F_1$ -Score. This again, indicates more information of intraoperative indexes provided than preoperative indexes in predicting OS time. Indeed, we obtained Precision $27.68\%$ , Recall $26.65\%$ and F1-Score $23.85\%$ if we only predicted OS time using preoperative features, which is comparable to the result from random guess. On the other hand, we can observe from Fig. 4 that preoperative indexes are beneficial to the estimation of intraoperative features, which indicating the necessity of using intraoperative features as a bridge between preoperative information and final OS time prediction. + +# 4.3. Medical Interpretation of Causal Graph + +To further explain the effectiveness of our causally-aware module, we in this subsection present our learned causal graph over intraoperative features in Fig. 5 and the corresponding medical interpretation of learned causal relations. In general, most of the parent-child relationships in our CaDAG are in line with common sense or can be explained by prior medical knowledge, e.g., sum of tumor diameter is determined by number of tumors, the edge from $G$ -score to $S$ -score also correspond to the strong correlation between fibrosis and hepatitis (please refer to appendix for detailed descriptions of intraoperative indexes). + +Further, those indicators with non-visual features that access severity of pathology are often determined by other attributes of the liver; therefore, these features often come up in deeper layers in Fig. 5. For instance, $G$ -score and $S$ -score reflect the level of inflammation and fibrosis that are artificially defined according to Metavir scoring system [2]. Those indicators tend to rely on other intraoperative features besides merely MRI and preoperative information, e.g. clinicopathologic hepatocirrhosis is reasoned from hepatocirrhosis and other covariants. For this reason, the addi + +Table 2. Results of Ablation Studies. (a) estimates each intraoperative features naively from (MRI, preoperative indexes), without the causally-aware module. (b) is the same to $M R I + P r e$ in Tab. 1. (c) is the CAWIM without preoperative features for intraoperative indexes estimation and OS time prediction. + +
ModelPrecisionRecallF1-Score
(a)w/o CaDAG36.98±2.3837.78±1.9136.62±1.87
(b)w/o Intra36.99±5.3035.62±3.5033.83±3.64
(c)w/o Pre38.92±7.7141.57±4.5338.43±5.84
(d)CAWIM45.58±5.8643.70±5.8942.21±4.92
+ +tional leverage of hepatocirrhosis lead to more accurate estimation of clinicopathologic hepatocirrhosis, as shown in Fig. 4 (a). On the other hand, we also observe some associations that cannot be explained well, e.g. the causal relationship between sum of tumor diameter and cell type, which may due to the existence of unobserved confounders. + +# 4.4. Visualization + +In this section, we visualize the high-response area of our CAWIM Fig. 6(c) and the version without CaDAG Fig. 6(b) using Grad-CAM [29]. As shown in Fig. 6, the detected regions of our method can be more concentrated on the liver. This result can be explained with Fig.4 and Fig. 5 in a more complementary way. Specifically, as the learned causal graph is medical explainable and with better estimation of intraoperative features, the model is driven to locate on liver-related regions. More visualization results are left in appendix due to space limit. + +# 4.5. Further Discussion + +Intraoperative prediction. In our method, the estimation of intraoperative indicators are important for final prediction. Fig. 4 shows the metrics of some typical intraoperative prediction by i) $w/o \, CaDAG$ and ii) $w/o \, Pre$ with $F_1$ -Score and accuracy for discrete variants (Fig. 4 (a, b)) and $R$ -square for continuous ones (Fig. 4 (c)). Fig. 4 (a, b) shows that our CaDAG can generally enhance the estimation of categorical variables hepatocirrhosis, cell type, differentiation, etc. Nevertheless, we also find that such improvements will decrease if corresponding indicators are in deep layers of the graph in Fig. 5. For example, the accuracy of $G$ -Score and $S$ -score is only comparable and even worse than that of $w/o \, CaDAG$ . A possible reason is the accumulated error along the directed path for estimation. On the other hand, as shown by the results of $R$ -square (closer to 1, the better the prediction) does not bring about notable improvement in regression tasks, (e.g., ascites and tumor diameter). + +We also observe an evident performance degradation of $w / o$ Pre for most of those discrete or continuous variants. This implies that although operative alone has limited contribution to OS time prediction, it still matters in our CAWIM, because of the less missing information due to shorter time interval from the preoperative stage to the + +![](images/01d26054c451b1c255ec9bdb018d00e30cb4ff489eaf6f6bc84a3c82a4a2f274.jpg) +(a) + +![](images/f3d9ee8a3d02e3efb04474ac8234a698f83a470aadc7468a30acd9ec12c36a4c.jpg) +(b) +Figure 4. Quantity Results of Intraoperative Prediction. Intraoperative prediction are categorized into classification ((a) and (b)) and regression problem ((c)) according to the data form. (a) and (b) shows the prediction $\mathrm{F}_1$ -Score and accuracy rate of the discrete variants respectively. The numbers after the variant name in (a) and (b) are the number of classes, i.e. hepatocirrhosis is a 6-category classification task. (c) shows the R-squared score for continuous variants. + +![](images/fd6aa08290bb102b2db018cf78102e9766530d0ae8a7d21ec1a482407f4086db.jpg) +(c) + +![](images/a4777bdbafa993ada2bb5a5b5a1f855219cff6f6c8b8d872dc8b1be1d363964e.jpg) +Figure 5. Learned DAG over Intraoperative Indexes A. + +![](images/7d7d315d9ce95c101dc90f817294af7392c4bb7fad917562d9e90d5d6976aaa8.jpg) +Figure 6. Heat Maps via Grad-CAM [29] on 5 Patients. (a) shows original images and the bounding box of liver. (b) and (c) shows the area of interest $w/o$ and with CaDAG, respectively. + +surgery stage than to the final stage. This observation validates the effect of using intraoperative features. + +General Framework for OS time prediction. In general, predicting OS time from preoperative images and indexes is medically important but challenging, as there exists a lot of missing information due to a long time interval between the preoperative stage to the final OS stage. To amend this problem, it is informative to leverage some intermediate information. In this paper, we show that with this leverage and the modeling of this information via causal discovery, our CAWIM model enjoys convincing medical interpretability, more concentrated location of liver region, more accurate + +estimation of intraoperative features, and finally better classification results on OS time. We thus believe that our framework can be beneficial to OS time prediction in other scenarios, e.g., other diseases or intermediate information in addition to intraoperative indexes. + +Limitations. We find in Fig. 5 that although most of learned causal relations are consistent with medical priors, there exist relations that cannot be explained well: it is hard to determine the causal order for some pairs of variables: e.g., the causal relationship between the sum of tumor diameter and cell type. This may due to unobserved confounders between these pairs of variables, which may be alleviated by learning hidden representations that can explain the associations among these pairs of variables. We will leave this exploration in our future work. + +# 5. Conclusion + +We propose a novel OS time prediction paradigm, which is the first to leverage intraoperative attributes by causal structure learning. Our method significantly outperforms baselines by a large margin. We demonstrate that this improvement is contributed to highly interpretable learned causal structure, accurate estimation of intraoperative indexes, and identification of disease-related regions. We believe our method, especially the leverage of intraoperative information equipped with causal discovery can potentially benefit other scenarios. For limitation, we shall relax the SCM assumption by allowing the existence of unobserved confounders. We leave it as the future work. + +Acknowledgment. Ethics Committee of Zhongshan Hospital, Fudan University approved the protocol of this study and waived the requirement for patient-informed consent (B2021-325R). The authors are also with Shanghai Key Lab of Intelligent Information Processing, and Fudan ISTBI-ZJNU Algorithm Centre for Brain-inspired Intelligence, Zhejiang Normal University, Jinhua, China. + +# References + +[1] Shekoofeh Azizi, Basil Mustafa, Fiona Ryan, Zachary Beaver, Jan Freyberg, Jonathan Deaton, Aaron Loh, Alan Karthikesalingam, Simon Kornblith, Ting Chen, et al. Big self-supervised models advance medical image classification. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3478-3488, 2021. 2 +[2] Pierre Bedossa and Thierry Poynard. An algorithm for the grading of activity in chronic hepatitis c. Hepatology, 24(2):289-293, 1996. 7 +[3] Nicholas A Christakis, Julia L Smith, Colin Murray Parkes, and Elizabeth B Lamont. Extent and determinants of error in doctors' prognoses in terminally ill patients: prospective cohort study commentary: Why do doctors overestimate? commentary: Prognoses should be based on proved indices not intuition. *Bmj*, 320(7233):469–473, 2000. 2 +[4] Agostino Colli, Mirella Fraquelli, Marco Andreola, Barbara Marino, Enrico Zuccoli, and Dario Conte. Severe liver fibrosis or cirrhosis: accuracy of us for detection—analysis of 300 cases. Radiology, 227(1):89-94, 2003. 2 +[5] Yin Dai, Yifan Gao, and Fayu Liu. Transmed: Transformers advance multi-modal medical image classification. Diagnostics, 11(8):1384, 2021. 2, 3 +[6] James J Driscoll and Oliver Rixe. Overall survival: still the gold standard: why overall survival remains the definitive end point in cancer clinical trials. The Cancer Journal, 15(5):401-405, 2009. 2, 3 +[7] J Faivre, D Forman, J Esteve, M Obradovic, M Sant, EURO-CARE Working Group, et al. Survival of patients with primary liver cancer, pancreatic cancer and biliary tract cancer in europe. European Journal of Cancer, 34(14):2184-2190, 1998. 2 +[8] Ming Fan, Zuhui Liu, Sudan Xie, Maosheng Xu, Shiwei Wang, Xin Gao, and Lihua Li. Integration of dynamic contrast-enhanced magnetic resonance imaging and t2-weighted imaging radiomic features by a canonical correlation analysis-based feature fusion method to predict histological grade in ductal breast carcinoma. Physics in Medicine & Biology, 64(21):215001, 2019. 2 +[9] Sheung Tat Fan, Chung Mau Lo, Ronnie TP Poon, Chun Yeung, Chi Leung Liu, Wai Key Yuen, Chi Ming Lam, Kelvin KC Ng, and See Ching Chan. Continuous improvement of survival outcomes of resection of hepatocellular carcinoma: a 20-year experience. Annals of surgery, 253(4):745-758, 2011. 2 +[10] Xue Feng, Nicholas J Tustison, Sohil H Patel, and Craig H Meyer. Brain tumor segmentation using an ensemble of 3d u-nets and overall survival prediction using radiomic features. Frontiers in computational neuroscience, 14:25, 2020. 3 +[11] Paul Glare, Kiran Virik, Mark Jones, Malcolm Hudson, Steffen Eychmuller, John Simes, and Nicholas Christakis. A systematic review of physicians' survival predictions in terminally ill cancer patients. *Bmj*, 327(7408):195, 2003. 2 +[12] Arthur Gretton, Olivier Bousquet, Alex Smola, and Bernhard Schölkopf. Measuring statistical dependence with hilbert-schmidt norms. In International conference on algorithmic learning theory, pages 63-77. Springer, 2005. 3, 4 + +[13] Arthur Gretton, Kenji Fukumizu, Choon Teo, Le Song, Bernhard Scholkopf, and Alex Smola. A kernel statistical test of independence. Advances in neural information processing systems, 20, 2007. 3 +[14] Lu Guo, Ping Wang, Ranran Sun, Chengwen Yang, Ning Zhang, Yu Guo, and Yuanming Feng. A fuzzy feature fusion method for auto-segmentation of gliomas with multimodality diffusion and perfusion magnetic resonance images in radiotherapy. Scientific reports, 8(1):1-11, 2018. 2 +[15] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 1 +[16] Mohammad Hesam Hesamian, Wenjing Jia, Xiangjian He, and Paul Kennedy. Deep learning techniques for medical image segmentation: achievements and challenges. Journal of digital imaging, 32(4):582-596, 2019. 2 +[17] Karin W Houben and John L McCall. Liver transplantation for hepatocellular carcinoma in patients without underlying liver disease: a systematic review. *Liver Transplantation and Surgery*, 5(2):91–95, 1999. 2 +[18] Biwei Huang, Kun Zhang, Jiji Zhang, Joseph D Ramsey, Ruben Sanchez-Romero, Clark Glymour, and Bernhard Scholkopf. Causal discovery from heterogeneous/nonstationary data. J. Mach. Learn. Res., 21(89):1-53, 2020. 2, 3, 4 +[19] Markus Kalisch and Peter Buhlman. Estimating high-dimensional directed acyclic graphs with the pc-algorithm. Journal of Machine Learning Research, 8(3), 2007. 3 +[20] Christophe Laurent, Jean Frédéric Blanc, Steele Nobili, Antonio Sa Cunha, Brigitte le Bail, Paulette Bioulac-Sage, Charles Balabaud, Maylis Capdepont, and Jean Saric. Prognostic factors and longterm survival after hepatic resection for hepatocellular carcinoma originating from noncirrhotic liver. Journal of the American College of Surgeons, 201(5):656-662, 2005. 2 +[21] Qing Li, Weidong Cai, Xiaogang Wang, Yun Zhou, David Dagan Feng, and Mei Chen. Medical image classification with convolutional neural network. In 2014 13th international conference on control automation robotics & vision (ICARCV), pages 844-848. IEEE, 2014. 2 +[22] Tao Li, Wu Li, Yehui Yang, and Wensheng Zhang. Classification of brain disease in magnetic resonance images using two-stage local feature fusion. PloS one, 12(2):e0171749, 2017. 2 +[23] Dong Nie, Han Zhang, Ehsan Adeli, Luyan Liu, and Dinggang Shen. 3d deep learning for multi-modal imaging-guided survival time prediction of brain tumor patients. In International conference on medical image computing and computer-assisted intervention, pages 212-220. Springer, 2016. 3 +[24] Byeonggwan Noh, Young Mok Park, Yujin Kwon, Chang In Choi, Byung Kwan Choi, Yo-Han Park, Kwangho Yang, Sunju Lee, Taeyoung Ha, YunKyong Hyon, et al. Machine learning-based survival rate prediction of korean hepatocellular carcinoma patients using multi-center data. BMC gastroenterology, 22(1):1-9, 2022. 2 + +[25] Augustus Odena, Christopher Olah, and Jonathon Shlens. Conditional image synthesis with auxiliary classifier gans. In International conference on machine learning, pages 2642-2651. PMLR, 2017. 1 +[26] Judea Pearl. Causality. Cambridge university press, 2009. 2, 3, 4 +[27] Markus Peck-Radosavljevic. Drug therapy for advanced-stage liver cancer. Liver cancer, 3(2):125-131, 2014. 2 +[28] Whitney B Pope, James Sayre, Alla Perlina, J Pablo Villablanca, Paul S Mischel, and Timothy F Cloughesy. Mr imaging correlates of survival in patients with high-grade gliomas. American Journal of Neuroradiology, 26(10):2466-2474, 2005. 3 +[29] Ramprasaath R. Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra. Grad-cam: Visual explanations from deep networks via gradient-based localization. In 2017 IEEE International Conference on Computer Vision (ICCV), pages 618-626, 2017. 7, 8 +[30] Wei Shao, Tongxin Wang, Liang Sun, Tianhan Dong, Zhi Han, Zhi Huang, Jie Zhang, Daoqiang Zhang, and Kun Huang. Multi-task multi-modal learning for joint diagnosis and prognosis of human cancers. Medical Image Analysis, 65:101795, 2020. 2, 3 +[31] Pater Spirtes, Clark Glymour, Richard Scheines, Stuart Kauffman, Valerio Aimale, and Frank Wimberly. Constructing bayesian network models of gene expression networks from microarray data. 2000. 3 +[32] Wen Tang, Haoyue Zhang, Pengxin Yu, Han Kang, and Rongguo Zhang. Mmmna-net for overall survival time prediction of brain tumor patients. arXiv preprint arXiv:2206.06267, 2022. 2 +[33] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 1 +[34] Jing-Houng Wang, Chi-Sin Changchien, Tsung-Hui Hu, Chuan-Mo Lee, Kwong-Ming Kee, Chih-Yun Lin, Chao-Long Chen, Tai-Yi Chen, Yu-Jie Huang, and Sheng-Nan Lu. The efficacy of treatment schedules according to Barcelona clinic liver cancer staging for hepatocellular carcinoma-survival analysis of 3892 patients. European journal of cancer, 44(7):1000-1006, 2008. 2 +[35] Samir S Yadav and Shivajirao M Jadhav. Deep convolutional neural network based medical image classification for disease diagnosis. Journal of Big Data, 6(1):1-18, 2019. 2 +[36] Naoki Yamanaka, Eizo Okamoto, Tsuyosi Oriyama, Jiro Fujimoto, Kazutaka Furukawa, Eisuke Kawamura, Tsuneo Tanaka, and Fumito Tomoda. A prediction scoring system to select the surgical treatment of liver cancer. further refinement based on 10 years of use. Annals of surgery, 219(4):342, 1994. 2 +[37] Jianpeng Zhang, Yutong Xie, Qi Wu, and Yong Xia. Medical image classification using synergic deep learning. Medical image analysis, 54:10-19, 2019. 2 +[38] Fan Zhou, Tengfei Li, Heng Li, and Hongtu Zhu. Tpcnn: two-phase patch-based convolutional neural network for automatic brain tumor segmentation and survival prediction. + +In International MICCAI Brainlesion Workshop, pages 274-286. Springer, 2017. 3 +[39] Tao Zhou, H. Fu, Yu Zhang, Changqing Zhang, Xiankai Lu, Jianbing Shen, and Ling Shao. M2net: Multi-modal multichannel network for overall survival time prediction of brain tumor patients. ArXiv, abs/2006.10135, 2020. 2, 3 \ No newline at end of file diff --git a/2023/Causally-Aware Intraoperative Imputation for Overall Survival Time Prediction/images.zip b/2023/Causally-Aware Intraoperative Imputation for Overall Survival Time Prediction/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..90adcdc4f860625fb7e21ff2288765a500f3f20a --- /dev/null +++ b/2023/Causally-Aware Intraoperative Imputation for Overall Survival Time Prediction/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1ca6c9a2a9956eba1d335ff0382d927c6a9f16e69280bfeb0fa110b5ba3927dc +size 399133 diff --git a/2023/Causally-Aware Intraoperative Imputation for Overall Survival Time Prediction/layout.json b/2023/Causally-Aware Intraoperative Imputation for Overall Survival Time Prediction/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..8c1f8d9f47df103ef572121918e427a297228d97 --- /dev/null +++ b/2023/Causally-Aware Intraoperative Imputation for Overall Survival Time Prediction/layout.json @@ -0,0 +1,11092 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 50, + 103, + 544, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 50, + 103, + 544, + 121 + ], + "spans": [ + { + "bbox": [ + 50, + 103, + 544, + 121 + ], + "type": "text", + "content": "Causally-Aware Intraoperative Imputation for Overall Survival Time Prediction" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 77, + 142, + 516, + 172 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 142, + 516, + 172 + ], + "spans": [ + { + "bbox": [ + 77, + 142, + 516, + 172 + ], + "type": "text", + "content": "Xiang Li" + }, + { + "bbox": [ + 77, + 142, + 516, + 172 + ], + "type": "inline_equation", + "content": "^{1,*}" + }, + { + "bbox": [ + 77, + 142, + 516, + 172 + ], + "type": "text", + "content": ", Xuelin Qian" + }, + { + "bbox": [ + 77, + 142, + 516, + 172 + ], + "type": "inline_equation", + "content": "^{1,*}" + }, + { + "bbox": [ + 77, + 142, + 516, + 172 + ], + "type": "text", + "content": ", Litian Liang" + }, + { + "bbox": [ + 77, + 142, + 516, + 172 + ], + "type": "inline_equation", + "content": "^{1,*}" + }, + { + "bbox": [ + 77, + 142, + 516, + 172 + ], + "type": "text", + "content": ", Lingjie Kong" + }, + { + "bbox": [ + 77, + 142, + 516, + 172 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 77, + 142, + 516, + 172 + ], + "type": "text", + "content": ", Qiaole Dong" + }, + { + "bbox": [ + 77, + 142, + 516, + 172 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 77, + 142, + 516, + 172 + ], + "type": "text", + "content": ", Jiejun Chen" + }, + { + "bbox": [ + 77, + 142, + 516, + 172 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 77, + 142, + 516, + 172 + ], + "type": "text", + "content": ", Dingxia Liu" + }, + { + "bbox": [ + 77, + 142, + 516, + 172 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 77, + 142, + 516, + 172 + ], + "type": "text", + "content": ", Xiuzhong Yao" + }, + { + "bbox": [ + 77, + 142, + 516, + 172 + ], + "type": "inline_equation", + "content": "^{2,\\dagger}" + }, + { + "bbox": [ + 77, + 142, + 516, + 172 + ], + "type": "text", + "content": ", Yanwei Fu" + }, + { + "bbox": [ + 77, + 142, + 516, + 172 + ], + "type": "inline_equation", + "content": "^{1,\\dagger}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 193, + 185, + 400, + 199 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 193, + 185, + 400, + 199 + ], + "spans": [ + { + "bbox": [ + 193, + 185, + 400, + 199 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 193, + 185, + 400, + 199 + ], + "type": "text", + "content": "School of Data Science, Fudan University" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 136, + 200, + 456, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 200, + 456, + 213 + ], + "spans": [ + { + "bbox": [ + 136, + 200, + 456, + 213 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 136, + 200, + 456, + 213 + ], + "type": "text", + "content": "Department of Radiology, Zhongshan Hospital, Fudan University" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 121, + 215, + 471, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 121, + 215, + 471, + 228 + ], + "spans": [ + { + "bbox": [ + 121, + 215, + 471, + 228 + ], + "type": "text", + "content": "{li_x20, xlqian, ltliang19, qldong18, jjchen20, yanweifu}@fudan.edu.cn" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 115, + 230, + 475, + 242 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 115, + 230, + 475, + 242 + ], + "spans": [ + { + "bbox": [ + 115, + 230, + 475, + 242 + ], + "type": "text", + "content": "{jkong22,dxliu21}@m.fudan.edu.cn yao.xiuzhong@zs-hospital.sh.cn" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 55, + 266, + 541, + 373 + ], + "blocks": [ + { + "bbox": [ + 55, + 266, + 541, + 373 + ], + "lines": [ + { + "bbox": [ + 55, + 266, + 541, + 373 + ], + "spans": [ + { + "bbox": [ + 55, + 266, + 541, + 373 + ], + "type": "image", + "image_path": "0738ecee42815d685eaa17534391c0dc9b1ca91398d3944bca76e79b91c8f7a7.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 374, + 546, + 451 + ], + "lines": [ + { + "bbox": [ + 46, + 374, + 546, + 451 + ], + "spans": [ + { + "bbox": [ + 46, + 374, + 546, + 451 + ], + "type": "text", + "content": "Figure 1. Our idea illustration. Patient A and B have different OS time (109 months and 16 months, respectively) but similar pre-operative image-based patterns and indexes; thus, preoperative based model fails to distinguish A and B (their predicted OS time are both larger than 9 years). However, their intra-operative attributes, which can describe the severity of disease in a more informative way, are very different from each other (e.g., G-score, S-score, Clinicopathologic hepatocirrhosis). Therefore, it can help the model discriminate A and B. In the right image, we show that our CAWIM – that leverages intra-operative indexes in the training stage – can largely improve the prediction power than other methods. Our method managed to correctly classify these two cases, and the overall performance of our CAWIM surpasses other baseline models by approximately 10 points on 4-category classification task." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 143, + 458, + 192, + 471 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 458, + 192, + 471 + ], + "spans": [ + { + "bbox": [ + 143, + 458, + 192, + 471 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 483, + 288, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 483, + 288, + 685 + ], + "spans": [ + { + "bbox": [ + 46, + 483, + 288, + 685 + ], + "type": "text", + "content": "Previous efforts in vision community are mostly made on learning good representations from visual patterns. Beyond this, this paper emphasizes the high-level ability of causal reasoning. We thus present a case study of solving the challenging task of Overall Survival (OS) time in primary liver cancers. Critically, the prediction of OS time at the early stage remains challenging, due to the unobvious image patterns of reflecting the OS. To this end, we propose a causal inference system by leveraging the intraoperative attributes and the correlation among them, as an intermediate supervision to bridge the gap between the images and the final OS. Particularly, we build a causal graph, and train the images to estimate the intraoperative attributes for final OS prediction. We present a novel Causally-aware Intraoperative Imputation Model (CAWIM) that can sequentially predict each attribute using its parent nodes in the estimated causal graph. To determine the causal directions," + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 460, + 546, + 532 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 460, + 546, + 532 + ], + "spans": [ + { + "bbox": [ + 306, + 460, + 546, + 532 + ], + "type": "text", + "content": "we propose a splitting-voting mechanism, which votes for the direction for each pair of adjacent nodes among multiple predictions obtained via causal discovery from heterogeneity. The practicability and effectiveness of our method are demonstrated by the promising results on liver cancer dataset of 361 patients with long-term observations." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 559, + 386, + 572 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 559, + 386, + 572 + ], + "spans": [ + { + "bbox": [ + 306, + 559, + 386, + 572 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 582, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 582, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 582, + 546, + 713 + ], + "type": "text", + "content": "The success of recent deep learning model is largely attributed to learning the good representations for visual patterns. Such representations essentially facilitate various vision task, such as recognition and synthesis [15, 25, 33]. Nevertheless, one important goal for the vision community is to model and summarize the relationships of observed variables of a system, in order to enable well predictions on similar data. Essentially, it is desirable to understand how the system is changed if one modifies these relationships under certain conditions, e.g., the effects of a treatment in healthcare. Thus this demands the high-" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 693, + 124, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 693, + 124, + 703 + ], + "spans": [ + { + "bbox": [ + 58, + 693, + 124, + 703 + ], + "type": "text", + "content": "*Equal contribution" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 59, + 703, + 134, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 703, + 134, + 712 + ], + "spans": [ + { + "bbox": [ + 59, + 703, + 134, + 712 + ], + "type": "text", + "content": "† Corresponding author" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15681" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "text", + "content": "level ability of causal reasoning beyond the previous efforts of only learning good representations for visual patterns [1, 5, 16, 21, 35, 37]. This naturally leads into our task of causal inference." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 122, + 289, + 338 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 122, + 289, + 338 + ], + "spans": [ + { + "bbox": [ + 46, + 122, + 289, + 338 + ], + "type": "text", + "content": "This paper presents a case study of solving the challenging task of Overall Survival (OS) time estimation in Primary Liver Cancers (PLC). Generally, the liver cancer remains one of the most common malignancies worldwide in the 21st century, as there are about one million new cases every year [36]. The five-year survival rate for advanced PLC is only about " + }, + { + "bbox": [ + 46, + 122, + 289, + 338 + ], + "type": "inline_equation", + "content": "5\\%" + }, + { + "bbox": [ + 46, + 122, + 289, + 338 + ], + "type": "text", + "content": " [7]. Therefore, early and accurate prediction of OS time estimation can provide informative guidance for individualized treatment planning and reducing burden of medical resources [6, 27, 32, 34]. On the other hand, one shall easily notice that with the renaissance of deep learning, great achievements have been made on medical imaging analysis, such as diagnosis, segmentation, and classification [16, 21, 35, 37]. Unfortunately, it still remains challenging for experienced clinicians to predict OS time at early stage, even with advanced modern diagnostic tools such as Magnetic Resonance Imaging (MRI) or tumor marker tests [3, 11], and the deep learning tools [1, 5]." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 340, + 289, + 508 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 340, + 289, + 508 + ], + "spans": [ + { + "bbox": [ + 46, + 340, + 289, + 508 + ], + "type": "text", + "content": "Some studies propose to leverage deep neural networks for OS time prediction. They focus on fusion learning of multi-modal image features and some basic information (i.e., age and gender) [5, 8, 14, 22, 24, 30, 39]. Nevertheless, the accurate prediction based on only preoperative information (such as image and tumor marker indexes) is still challenging, possibly due to missing information from early diagnosis stage to the final stage. This missing information includes the texture and pathological attributes of the liver, health level of the patient, and post-operative treatment [9, 17, 20]. For example, as in Fig. 1, patient A and B with different OS time have almost identical preoperative indicators cause, history of disease, etc., making the preoperative-based model hard for discrimination." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 510, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 510, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 510, + 289, + 715 + ], + "type": "text", + "content": "To amend this problem, we present a causal inference system that can well utilize the intraoperative information, which is pretty easy to be accessed in training data. According to medical priorities, such information records pathological attributes, which can be more reflective about the patient's health level and the postoperative recovery. Inspired by this, we propose to leverage this auxiliary information to help build our causal inference system. Again, we take the example in Fig. 1. Although preoperative information cannot differentiate patient A from B, their intraoperative index shows great difference in distribution, which can thus be employed for the discrimination. Additionally, there are many indicators from medical experts that the intraoperative indexes are related to each other. For instance, the clinicopathologic hepatocirrhosis is dependent on the hepatocirrhosis; the sum of tumor diameter is affected by the number of tumors; the fibrosis (reflected on S-Score) can prob" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "text", + "content": "ably deteriorate to cirrhosis [4], etc. By leveraging these relationships, we can better understand the causal inference concrete from both observed data modelling and medical expert-level knowledge of these variables." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 123, + 546, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 123, + 546, + 495 + ], + "spans": [ + { + "bbox": [ + 304, + 123, + 546, + 495 + ], + "type": "text", + "content": "To this end, we encapsulate these priors and the inspired proposals into a new method, dubbed as Causally-Aware Intraoperative Imputation Model (CAWIM). It incorporates causal discovery module to sequentially estimate intraoperative indexes as an intermediate stage towards final OS time prediction. Specifically, our model is composed of two key steps: i) estimating the intraoperative indexes using preoperative information, i.e., image and indexes; ii) followed by OS time prediction using estimated intraoperative indexes and preoperative information. To achieve more accurate prediction of intraoperative indexes that is determinant to the prediction power of the whole method, we propose a Causally-aware Directed Acyclic Graph (CaDAG) module. It learns the causal structure represented as a DAG over intraoperative features. To identify the causal relations beyond the traditional PC algorithm [26], we propose a splitting-voting mechanism, which is inspired by the recent work [18] that learn the causal structure with the assistance of an auxiliary domain index variable. Our proposed mechanism can not only identify the causal relations even when this domain index variable is not available, but also can be theoretically guaranteed that the learned graph is not acyclic. During test stage, we sequentially estimate each intraoperative index with preoperative information and additionally, its parent set among other intraoperative indexes. The utility of our method can be demonstrated by a significant improvement of OS time prediction on an in-house liver cancer dataset, as well as better interpretability of learned causal structure, more accurate estimation of intraoperative indexes and more interpretable visualization results." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 498, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 498, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 498, + 547, + 715 + ], + "type": "text", + "content": "In a nutshell, we for the first time present a case study of building a causally-aware intraoperative imputation system for the challenging task of overall survival time prediction. The proposed method of building the casual inference system, can be naturally extended to other similar medical tasks. Our key contributions are listed as follows. (1) New Paradigm for OS time Prediction. We propose to leverage intraoperative indexes as an intermediate stage during training. The leverage of this information can significantly alleviate the \"missing information\" issue. To the best of our knowledge, we are the first to leverage auxiliary information (in addition to preoperative features) for OS time prediction. (2) Causal Structure Learning. We propose a novel splitting-voting mechanism that can identify the causal structure even when the domain index variable is missing. (3) Better Prediction Power. Our method can significantly improve the prediction power for liver cancer over the competitors. The methods are evaluated on the medical" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "15682" + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 287, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 287, + 96 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 287, + 96 + ], + "type": "text", + "content": "dataset, which, to the best of our knowledge, is the largest primary live cancer dataset." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 107, + 228, + 119 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 107, + 228, + 119 + ], + "spans": [ + { + "bbox": [ + 47, + 107, + 228, + 119 + ], + "type": "text", + "content": "2. Related Work and Preliminaries" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 127, + 286, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 127, + 286, + 186 + ], + "spans": [ + { + "bbox": [ + 47, + 127, + 286, + 186 + ], + "type": "text", + "content": "Previous efforts [23,39] directly take the OS time prediction as the vanilla classification task by employing various deep models, whilst we reformulate it by causal discovery method for the first time. We will review the most related literature here." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 187, + 286, + 330 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 187, + 286, + 330 + ], + "spans": [ + { + "bbox": [ + 46, + 187, + 286, + 330 + ], + "type": "text", + "content": "MRI based deep models. OS time is defined as the duration from patient's first scan to cancer-related death [6]. Since Magnetic Resonance Image (MRI) plays an important role in tumor-related studies, various MRI based deep learning model have been proposed for feature extraction and representing [28], multi-modal feature fusion [10, 30, 38, 39], and improvement on model structure [5]. However, since missing information during disease progression can mediate the correlation from the preoperative stage to the final stage, the results of these image based methods are typically unsatisfied. In contrast, our method can amend this problem by leveraging intraoperative indexes as a bridge." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 331, + 286, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 331, + 286, + 593 + ], + "spans": [ + { + "bbox": [ + 46, + 331, + 286, + 593 + ], + "type": "text", + "content": "Causal discovery. Causal discovery is to learn the causal graph over endogenous variables. Typical methods include PC algorithm [19], a constraint-based method that implement iterative conditional independence test to identify the skeleton and directions via v-structure. Under Markovian and faithfulness assumptions [26, 31], this method is provable to identify an equivalent class of the causal graph. To further identify more directions, recent work [18] leveraged a domain index variable that splits the dataset into multiple domains. With these heterogeneous data, they can identify mutable variables that change across the domain index variable. On the basis of this, they further identified edges among these mutable variable set by implementing Hilbert Schmidt Independence Criterion (HSIC) norm [13]. However, such domain index may not be available in many scenarios, making the identification of these directions difficult. Besides, this method may not ensure that the learned graph is acyclic. To address these issues, we in this paper propose a splitting-voting mechanism to mine the causal relations among intraoperative indexes so as for leveraging each other to better estimate these indexes, which is provable to return a directed acyclic graph." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 594, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 286, + 713 + ], + "type": "text", + "content": "Preliminaries. (1) Causal Discovery. By faithfulness and Markovian assumptions [26], we can use conditional independent tests for causal discovery. That is learning the causal graph over " + }, + { + "bbox": [ + 46, + 594, + 286, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{V}" + }, + { + "bbox": [ + 46, + 594, + 286, + 713 + ], + "type": "text", + "content": ". (2) Structural causal model (SCM). Before fully developing our contributions in methodology, we give as preliminary the definition of SCM, which is defined as " + }, + { + "bbox": [ + 46, + 594, + 286, + 713 + ], + "type": "inline_equation", + "content": "\\langle G, \\mathcal{F}, P\\langle \\epsilon \\rangle \\rangle" + }, + { + "bbox": [ + 46, + 594, + 286, + 713 + ], + "type": "text", + "content": ": i) directed acycle graph " + }, + { + "bbox": [ + 46, + 594, + 286, + 713 + ], + "type": "inline_equation", + "content": "G = (\\mathbf{V}, \\mathbf{E})" + }, + { + "bbox": [ + 46, + 594, + 286, + 713 + ], + "type": "text", + "content": " is as causal structure, with node set " + }, + { + "bbox": [ + 46, + 594, + 286, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{V}" + }, + { + "bbox": [ + 46, + 594, + 286, + 713 + ], + "type": "text", + "content": " and edge set " + }, + { + "bbox": [ + 46, + 594, + 286, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{E}" + }, + { + "bbox": [ + 46, + 594, + 286, + 713 + ], + "type": "text", + "content": "; ii) the autonomous structural functions " + }, + { + "bbox": [ + 46, + 594, + 286, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{F} = \\{f_k\\}_{V_k \\in V}" + }, + { + "bbox": [ + 46, + 594, + 286, + 713 + ], + "type": "text", + "content": ", where disturbance on " + }, + { + "bbox": [ + 46, + 594, + 286, + 713 + ], + "type": "inline_equation", + "content": "V_k" + }, + { + "bbox": [ + 46, + 594, + 286, + 713 + ], + "type": "text", + "content": " has no effect on others; iii) probability" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 72, + 545, + 360 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 360 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 360 + ], + "type": "text", + "content": "measure " + }, + { + "bbox": [ + 304, + 72, + 545, + 360 + ], + "type": "inline_equation", + "content": "P(\\epsilon)" + }, + { + "bbox": [ + 304, + 72, + 545, + 360 + ], + "type": "text", + "content": " for exogenous variables " + }, + { + "bbox": [ + 304, + 72, + 545, + 360 + ], + "type": "inline_equation", + "content": "\\{\\epsilon_k\\}_k" + }, + { + "bbox": [ + 304, + 72, + 545, + 360 + ], + "type": "text", + "content": ". Given the assumption that " + }, + { + "bbox": [ + 304, + 72, + 545, + 360 + ], + "type": "inline_equation", + "content": "\\{\\epsilon_k\\}_k" + }, + { + "bbox": [ + 304, + 72, + 545, + 360 + ], + "type": "text", + "content": " are independent, each " + }, + { + "bbox": [ + 304, + 72, + 545, + 360 + ], + "type": "inline_equation", + "content": "\\mathrm{P}" + }, + { + "bbox": [ + 304, + 72, + 545, + 360 + ], + "type": "text", + "content": " can be obtained by Causal Markov Condition with " + }, + { + "bbox": [ + 304, + 72, + 545, + 360 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 304, + 72, + 545, + 360 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 304, + 72, + 545, + 360 + ], + "type": "inline_equation", + "content": "\\mathrm{P}(\\{V_K = v_k\\}_{V_k \\in \\mathbf{V}}) = \\prod_k \\mathrm{P}(V_k = v_k | Pa(k) = pa(k))" + }, + { + "bbox": [ + 304, + 72, + 545, + 360 + ], + "type": "text", + "content": ". (3) Causal discovery from heterogeneity. In [18], the authors proposed to leverage a domain index variable " + }, + { + "bbox": [ + 304, + 72, + 545, + 360 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 304, + 72, + 545, + 360 + ], + "type": "text", + "content": " to identify some causal directions. This method is built upon the PC and faithfulness assumption at a distributional level. It first learns the skeleton of the graph and identifies a variable set that is affected by " + }, + { + "bbox": [ + 304, + 72, + 545, + 360 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 304, + 72, + 545, + 360 + ], + "type": "text", + "content": ", followed by identifying the directions between " + }, + { + "bbox": [ + 304, + 72, + 545, + 360 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 304, + 72, + 545, + 360 + ], + "type": "text", + "content": " and its neighbors. Specifically, if " + }, + { + "bbox": [ + 304, + 72, + 545, + 360 + ], + "type": "inline_equation", + "content": "C \\to V_i, C \\to V_j" + }, + { + "bbox": [ + 304, + 72, + 545, + 360 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 72, + 545, + 360 + ], + "type": "inline_equation", + "content": "V_i \\bot V_j | X, C" + }, + { + "bbox": [ + 304, + 72, + 545, + 360 + ], + "type": "text", + "content": " for some deconfounding set " + }, + { + "bbox": [ + 304, + 72, + 545, + 360 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 304, + 72, + 545, + 360 + ], + "type": "text", + "content": " between " + }, + { + "bbox": [ + 304, + 72, + 545, + 360 + ], + "type": "inline_equation", + "content": "V_i" + }, + { + "bbox": [ + 304, + 72, + 545, + 360 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 72, + 545, + 360 + ], + "type": "inline_equation", + "content": "V_j" + }, + { + "bbox": [ + 304, + 72, + 545, + 360 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 72, + 545, + 360 + ], + "type": "inline_equation", + "content": "\\{P(V_i | X, C = c)\\}" + }, + { + "bbox": [ + 304, + 72, + 545, + 360 + ], + "type": "text", + "content": " is independent of " + }, + { + "bbox": [ + 304, + 72, + 545, + 360 + ], + "type": "inline_equation", + "content": "\\{P(V_j | V_i, C = c, X)\\}" + }, + { + "bbox": [ + 304, + 72, + 545, + 360 + ], + "type": "text", + "content": ", while " + }, + { + "bbox": [ + 304, + 72, + 545, + 360 + ], + "type": "inline_equation", + "content": "\\{P(V_j | X, C = c)\\}" + }, + { + "bbox": [ + 304, + 72, + 545, + 360 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 72, + 545, + 360 + ], + "type": "inline_equation", + "content": "\\{P(V_j | V_i, C = c, X)\\}" + }, + { + "bbox": [ + 304, + 72, + 545, + 360 + ], + "type": "text", + "content": " are dependent. We can test these Independence using Hilbert Schmidt Independence Criterion (HSIC) [12] norms " + }, + { + "bbox": [ + 304, + 72, + 545, + 360 + ], + "type": "inline_equation", + "content": "\\delta_{v_i \\to v_j | X}" + }, + { + "bbox": [ + 304, + 72, + 545, + 360 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 72, + 545, + 360 + ], + "type": "inline_equation", + "content": "\\delta_{v_j \\to v_i | X}" + }, + { + "bbox": [ + 304, + 72, + 545, + 360 + ], + "type": "text", + "content": ". One can obtain that " + }, + { + "bbox": [ + 304, + 72, + 545, + 360 + ], + "type": "inline_equation", + "content": "V_i \\to V_j" + }, + { + "bbox": [ + 304, + 72, + 545, + 360 + ], + "type": "text", + "content": " if " + }, + { + "bbox": [ + 304, + 72, + 545, + 360 + ], + "type": "inline_equation", + "content": "\\delta_{v_i \\to V_j} < \\alpha" + }, + { + "bbox": [ + 304, + 72, + 545, + 360 + ], + "type": "text", + "content": " with pre-set significance level " + }, + { + "bbox": [ + 304, + 72, + 545, + 360 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 304, + 72, + 545, + 360 + ], + "type": "text", + "content": ". However, in many applications such domain index variable is missing, i.e., all data are pooled together. In this paper, we propose a splitting-voting mechanism (in the CaDAG module) to identify causal relations. In the following, we first introduce some basic assumptions that our method is built upon." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 369, + 388, + 382 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 369, + 388, + 382 + ], + "spans": [ + { + "bbox": [ + 306, + 369, + 388, + 382 + ], + "type": "text", + "content": "3. Methodology" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 388, + 545, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 388, + 545, + 496 + ], + "spans": [ + { + "bbox": [ + 304, + 388, + 545, + 496 + ], + "type": "text", + "content": "Problem setting. For the OS time classification task, suppose we have " + }, + { + "bbox": [ + 304, + 388, + 545, + 496 + ], + "type": "inline_equation", + "content": "\\{X_{i},\\mathbf{B}_{i},\\mathbf{A}_{i},y_{i}\\} \\sim_{i.i.d}P(X,\\mathbf{B},\\mathbf{A},y)" + }, + { + "bbox": [ + 304, + 388, + 545, + 496 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 304, + 388, + 545, + 496 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 304, + 388, + 545, + 496 + ], + "type": "text", + "content": " denotes the image acquired from structural Magnetic Resonance Image (sMRI); " + }, + { + "bbox": [ + 304, + 388, + 545, + 496 + ], + "type": "inline_equation", + "content": "\\mathbf{B}" + }, + { + "bbox": [ + 304, + 388, + 545, + 496 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 388, + 545, + 496 + ], + "type": "inline_equation", + "content": "\\mathbf{A}" + }, + { + "bbox": [ + 304, + 388, + 545, + 496 + ], + "type": "text", + "content": " denote the preoperative and intraoperative attributes, respectively. We have the final label " + }, + { + "bbox": [ + 304, + 388, + 545, + 496 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 304, + 388, + 545, + 496 + ], + "type": "text", + "content": ". Our goal is to predict " + }, + { + "bbox": [ + 304, + 388, + 545, + 496 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 304, + 388, + 545, + 496 + ], + "type": "text", + "content": " from the image " + }, + { + "bbox": [ + 304, + 388, + 545, + 496 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 304, + 388, + 545, + 496 + ], + "type": "text", + "content": " and preoperative attributes " + }, + { + "bbox": [ + 304, + 388, + 545, + 496 + ], + "type": "inline_equation", + "content": "\\mathbf{B}" + }, + { + "bbox": [ + 304, + 388, + 545, + 496 + ], + "type": "text", + "content": " that are recorded before surgery. Our method is built upon the following three assumptions, commonly utilized in the causal inference works [18, 26]." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 502, + 545, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 502, + 545, + 540 + ], + "spans": [ + { + "bbox": [ + 304, + 502, + 545, + 540 + ], + "type": "text", + "content": "Definition 1 (Causal Graph) We assume the causal graph over " + }, + { + "bbox": [ + 304, + 502, + 545, + 540 + ], + "type": "inline_equation", + "content": "\\mathbf{A}" + }, + { + "bbox": [ + 304, + 502, + 545, + 540 + ], + "type": "text", + "content": " is a directed acyclic graph (DAG) and denote the corresponding SCM as " + }, + { + "bbox": [ + 304, + 502, + 545, + 540 + ], + "type": "inline_equation", + "content": "M\\coloneqq \\langle G\\coloneqq (\\mathbf{A},\\mathbf{E}),\\mathcal{F},P(\\varepsilon)\\rangle" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 545, + 508, + 556 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 545, + 508, + 556 + ], + "spans": [ + { + "bbox": [ + 305, + 545, + 508, + 556 + ], + "type": "text", + "content": "Definition 2 (Markovian and Faithfulness) For" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 557, + 545, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 557, + 545, + 605 + ], + "spans": [ + { + "bbox": [ + 304, + 557, + 545, + 605 + ], + "type": "text", + "content": "triplets of disjoint sets " + }, + { + "bbox": [ + 304, + 557, + 545, + 605 + ], + "type": "inline_equation", + "content": "\\mathbf{V}_i, \\mathbf{V}_j, \\mathbf{V}_k" + }, + { + "bbox": [ + 304, + 557, + 545, + 605 + ], + "type": "text", + "content": ", it holds that " + }, + { + "bbox": [ + 304, + 557, + 545, + 605 + ], + "type": "inline_equation", + "content": "\\mathbf{V}_i \\perp_d \\mathbf{V}_j | \\mathbf{V}_k \\leftrightarrow \\mathbf{V}_i \\perp \\mathbf{V}_j | \\mathbf{V}_k" + }, + { + "bbox": [ + 304, + 557, + 545, + 605 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 557, + 545, + 605 + ], + "type": "inline_equation", + "content": "\\perp_d" + }, + { + "bbox": [ + 304, + 557, + 545, + 605 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 557, + 545, + 605 + ], + "type": "inline_equation", + "content": "\\perp" + }, + { + "bbox": [ + 304, + 557, + 545, + 605 + ], + "type": "text", + "content": " respectively mean " + }, + { + "bbox": [ + 304, + 557, + 545, + 605 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 304, + 557, + 545, + 605 + ], + "type": "text", + "content": "-separation and probability independence. This is the common property in DAG." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 611, + 545, + 683 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 611, + 545, + 683 + ], + "spans": [ + { + "bbox": [ + 304, + 611, + 545, + 683 + ], + "type": "text", + "content": "Definition 3 (Distributional Faithfulness) If " + }, + { + "bbox": [ + 304, + 611, + 545, + 683 + ], + "type": "inline_equation", + "content": "X_{i} \\to X_{j}" + }, + { + "bbox": [ + 304, + 611, + 545, + 683 + ], + "type": "text", + "content": " and at least " + }, + { + "bbox": [ + 304, + 611, + 545, + 683 + ], + "type": "inline_equation", + "content": "E \\to V_{i}" + }, + { + "bbox": [ + 304, + 611, + 545, + 683 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 304, + 611, + 545, + 683 + ], + "type": "inline_equation", + "content": "E \\to V_{j}" + }, + { + "bbox": [ + 304, + 611, + 545, + 683 + ], + "type": "text", + "content": " holds, " + }, + { + "bbox": [ + 304, + 611, + 545, + 683 + ], + "type": "inline_equation", + "content": "\\{P^{e}(V_{i}|V_{j},\\mathbf{Z})\\}" + }, + { + "bbox": [ + 304, + 611, + 545, + 683 + ], + "type": "text", + "content": " is dependent to " + }, + { + "bbox": [ + 304, + 611, + 545, + 683 + ], + "type": "inline_equation", + "content": "\\{P^{e}(V_{j}|\\mathbf{Z})\\}" + }, + { + "bbox": [ + 304, + 611, + 545, + 683 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 611, + 545, + 683 + ], + "type": "inline_equation", + "content": "\\mathbf{Z}" + }, + { + "bbox": [ + 304, + 611, + 545, + 683 + ], + "type": "text", + "content": " denotes the minimal deconfounding set. Particularly, " + }, + { + "bbox": [ + 304, + 611, + 545, + 683 + ], + "type": "inline_equation", + "content": "\\mathbf{Z}" + }, + { + "bbox": [ + 304, + 611, + 545, + 683 + ], + "type": "text", + "content": " is a deconfounding set between " + }, + { + "bbox": [ + 304, + 611, + 545, + 683 + ], + "type": "inline_equation", + "content": "V_{i}" + }, + { + "bbox": [ + 304, + 611, + 545, + 683 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 611, + 545, + 683 + ], + "type": "inline_equation", + "content": "V_{j}" + }, + { + "bbox": [ + 304, + 611, + 545, + 683 + ], + "type": "text", + "content": " if we have " + }, + { + "bbox": [ + 304, + 611, + 545, + 683 + ], + "type": "inline_equation", + "content": "V_{i} \\perp V_{j}|\\mathbf{Z}" + }, + { + "bbox": [ + 304, + 611, + 545, + 683 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 611, + 545, + 683 + ], + "type": "inline_equation", + "content": "\\mathbf{Z} \\cap (\\mathrm{De}(V_i) \\cup \\mathrm{De}(V_j)) = \\emptyset" + }, + { + "bbox": [ + 304, + 611, + 545, + 683 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 689, + 545, + 713 + ], + "type": "text", + "content": "In term of this assumption, we further have the following result for edge orientation [18]." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15683" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 134 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 134 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 134 + ], + "type": "text", + "content": "Theorem 4 (Theorem 2 in [18]) Denote " + }, + { + "bbox": [ + 46, + 72, + 289, + 134 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 46, + 72, + 289, + 134 + ], + "type": "text", + "content": " as domain index variable. Under assumptions 1, 2, 3, for each adjacent pair " + }, + { + "bbox": [ + 46, + 72, + 289, + 134 + ], + "type": "inline_equation", + "content": "(V_i, V_j)" + }, + { + "bbox": [ + 46, + 72, + 289, + 134 + ], + "type": "text", + "content": " such that " + }, + { + "bbox": [ + 46, + 72, + 289, + 134 + ], + "type": "inline_equation", + "content": "E \\to V_i" + }, + { + "bbox": [ + 46, + 72, + 289, + 134 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 46, + 72, + 289, + 134 + ], + "type": "inline_equation", + "content": "E \\to V_j" + }, + { + "bbox": [ + 46, + 72, + 289, + 134 + ], + "type": "text", + "content": " holds, we have " + }, + { + "bbox": [ + 46, + 72, + 289, + 134 + ], + "type": "inline_equation", + "content": "V_i \\to V_j" + }, + { + "bbox": [ + 46, + 72, + 289, + 134 + ], + "type": "text", + "content": " if " + }, + { + "bbox": [ + 46, + 72, + 289, + 134 + ], + "type": "inline_equation", + "content": "\\{P^e(V_i|\\mathbf{Z})\\}" + }, + { + "bbox": [ + 46, + 72, + 289, + 134 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 72, + 289, + 134 + ], + "type": "inline_equation", + "content": "\\{P^e(V_j|V_i,\\mathbf{Z})\\}" + }, + { + "bbox": [ + 46, + 72, + 289, + 134 + ], + "type": "text", + "content": " are independent while " + }, + { + "bbox": [ + 46, + 72, + 289, + 134 + ], + "type": "inline_equation", + "content": "\\{P^e(V_j|\\mathbf{Z})\\}" + }, + { + "bbox": [ + 46, + 72, + 289, + 134 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 72, + 289, + 134 + ], + "type": "inline_equation", + "content": "\\{P^e(V_i|V_j,\\mathbf{Z})\\}" + }, + { + "bbox": [ + 46, + 72, + 289, + 134 + ], + "type": "text", + "content": " are dependent." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 144, + 257, + 158 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 144, + 257, + 158 + ], + "spans": [ + { + "bbox": [ + 47, + 144, + 257, + 158 + ], + "type": "text", + "content": "3.1. Causally-aware Directed Acyclic Graph" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 163, + 289, + 714 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 46, + 163, + 289, + 390 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 163, + 289, + 390 + ], + "spans": [ + { + "bbox": [ + 46, + 163, + 289, + 390 + ], + "type": "text", + "content": "The Causally-aware DAG. We propose a novel Causally-Aware Intraoperative Imputation Model (CAWIM). The pipeline of our method is shown in Fig. 2. It is composed of Causally-aware Directed Acycled Graph (CaDAG), and a classification model with MRI imaging encoder. The CaDAG enables the causally-aware intraoperative reasoning, by managing to better represent and encode the correlation of intraoperative information. Specifically, we first use preoperative information, i.e., MRI images and indexes to estimate the intraoperative features. The estimated intraoperative features " + }, + { + "bbox": [ + 46, + 163, + 289, + 390 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{A}}" + }, + { + "bbox": [ + 46, + 163, + 289, + 390 + ], + "type": "text", + "content": ", together with the MRI " + }, + { + "bbox": [ + 46, + 163, + 289, + 390 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 46, + 163, + 289, + 390 + ], + "type": "text", + "content": " and preoperative features " + }, + { + "bbox": [ + 46, + 163, + 289, + 390 + ], + "type": "inline_equation", + "content": "\\mathbf{B}" + }, + { + "bbox": [ + 46, + 163, + 289, + 390 + ], + "type": "text", + "content": " are then utilized for OS classification. To accurately estimate intraoperative indexes " + }, + { + "bbox": [ + 46, + 163, + 289, + 390 + ], + "type": "inline_equation", + "content": "\\mathbf{A}" + }, + { + "bbox": [ + 46, + 163, + 289, + 390 + ], + "type": "text", + "content": ", we propose CaDAG to learn causal structure over " + }, + { + "bbox": [ + 46, + 163, + 289, + 390 + ], + "type": "inline_equation", + "content": "\\mathbf{A}" + }, + { + "bbox": [ + 46, + 163, + 289, + 390 + ], + "type": "text", + "content": ", and then leverage the parent nodes of each index for sequentially imputation during test stage. We give the details about the CaDAG and the intraoperative imputation model for classification, while some theoretical proofs, and the specification of our networks are in the appendix." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 391, + 287, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 391, + 287, + 581 + ], + "spans": [ + { + "bbox": [ + 46, + 391, + 287, + 581 + ], + "type": "text", + "content": "We aim at learning the causal structure over " + }, + { + "bbox": [ + 46, + 391, + 287, + 581 + ], + "type": "inline_equation", + "content": "\\mathbf{A}" + }, + { + "bbox": [ + 46, + 391, + 287, + 581 + ], + "type": "text", + "content": ". However, without the domain index, we actually have the difficulty of dividing the whole dataset into multiple domains. To this end, we propose a splitting-voting mechanism for this problem. Particularly, this mechanism first generates the domain-index variable " + }, + { + "bbox": [ + 46, + 391, + 287, + 581 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 46, + 391, + 287, + 581 + ], + "type": "text", + "content": " by randomly splitting the whole dataset into multiple domains for " + }, + { + "bbox": [ + 46, + 391, + 287, + 581 + ], + "type": "inline_equation", + "content": "m > 0" + }, + { + "bbox": [ + 46, + 391, + 287, + 581 + ], + "type": "text", + "content": " times. Then for each time we identify the mutable variable set and generate a prediction for each identifiable causal directions, with the assistance of domain index variable " + }, + { + "bbox": [ + 46, + 391, + 287, + 581 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 46, + 391, + 287, + 581 + ], + "type": "text", + "content": ". Finally, we vote for a final direction among " + }, + { + "bbox": [ + 46, + 391, + 287, + 581 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 46, + 391, + 287, + 581 + ], + "type": "text", + "content": " predictions for each pair of variables. As shown in Fig. 3, the whole procedure is a sequence of the four steps. Step i), ii) and iii) provide preliminary causal relationship between each pair of nodes. Step iv) is our proposed splitting-voting, aiming to finally determine all the causal directions and output a causal DAG." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 582, + 287, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 582, + 287, + 665 + ], + "spans": [ + { + "bbox": [ + 46, + 582, + 287, + 665 + ], + "type": "text", + "content": "i) Learning skeleton of the causal graph. We first implement PC algorithm to learn the skeleton [26]. That is an undirected graph over all attributes. This can be achieved by iterative conditional independence test. Then we determine " + }, + { + "bbox": [ + 46, + 582, + 287, + 665 + ], + "type": "inline_equation", + "content": "A_{i}" + }, + { + "bbox": [ + 46, + 582, + 287, + 665 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 582, + 287, + 665 + ], + "type": "inline_equation", + "content": "A_{j}" + }, + { + "bbox": [ + 46, + 582, + 287, + 665 + ], + "type": "text", + "content": " to be adjacent if they are not independent conditional on any subsets. Finally, we utilize the v-structure to learn the equivalence class of DAG." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 665, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 288, + 714 + ], + "type": "text", + "content": "ii) Random splitting and identification of mutable variable sets. To determine more directions, we randomly split the dataset into multiple domains for " + }, + { + "bbox": [ + 46, + 665, + 288, + 714 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 46, + 665, + 288, + 714 + ], + "type": "text", + "content": " times. For each time we have a domain index variable " + }, + { + "bbox": [ + 46, + 665, + 288, + 714 + ], + "type": "inline_equation", + "content": "E_{i}" + }, + { + "bbox": [ + 46, + 665, + 288, + 714 + ], + "type": "text", + "content": " and multiple do" + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 304, + 72, + 547, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 133 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 133 + ], + "type": "text", + "content": "mains " + }, + { + "bbox": [ + 304, + 72, + 547, + 133 + ], + "type": "inline_equation", + "content": "\\{\\mathcal{D}_e^i\\}" + }, + { + "bbox": [ + 304, + 72, + 547, + 133 + ], + "type": "text", + "content": " [18]; and we identify the mutable variable set " + }, + { + "bbox": [ + 304, + 72, + 547, + 133 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_i" + }, + { + "bbox": [ + 304, + 72, + 547, + 133 + ], + "type": "text", + "content": " such that each " + }, + { + "bbox": [ + 304, + 72, + 547, + 133 + ], + "type": "inline_equation", + "content": "X \\in \\mathbf{M}_i" + }, + { + "bbox": [ + 304, + 72, + 547, + 133 + ], + "type": "text", + "content": " has " + }, + { + "bbox": [ + 304, + 72, + 547, + 133 + ], + "type": "inline_equation", + "content": "X \\not\\perp E_i|\\mathbf{A} - X" + }, + { + "bbox": [ + 304, + 72, + 547, + 133 + ], + "type": "text", + "content": " is conditionally dependent to " + }, + { + "bbox": [ + 304, + 72, + 547, + 133 + ], + "type": "inline_equation", + "content": "E_i" + }, + { + "bbox": [ + 304, + 72, + 547, + 133 + ], + "type": "text", + "content": ". We thus can implement conditional independent tests to identify " + }, + { + "bbox": [ + 304, + 72, + 547, + 133 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_i" + }, + { + "bbox": [ + 304, + 72, + 547, + 133 + ], + "type": "text", + "content": ". Steps i) and ii) are summarized in Alg. 1." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 143, + 512, + 156 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 143, + 512, + 156 + ], + "spans": [ + { + "bbox": [ + 306, + 143, + 512, + 156 + ], + "type": "text", + "content": "Algorithm 1 Identify the skeleton and mutable set." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 167, + 536, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 167, + 536, + 180 + ], + "spans": [ + { + "bbox": [ + 306, + 167, + 536, + 180 + ], + "type": "text", + "content": "INPUT: " + }, + { + "bbox": [ + 306, + 167, + 536, + 180 + ], + "type": "inline_equation", + "content": "\\{\\mathcal{D}_e^i\\coloneqq \\{\\mathbf{A}_i^e\\}_{i = 1}^{n_e}|e\\}" + }, + { + "bbox": [ + 306, + 167, + 536, + 180 + ], + "type": "text", + "content": ", domain index variable " + }, + { + "bbox": [ + 306, + 167, + 536, + 180 + ], + "type": "inline_equation", + "content": "E_{i}" + }, + { + "bbox": [ + 306, + 167, + 536, + 180 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 180, + 491, + 193 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 180, + 491, + 193 + ], + "spans": [ + { + "bbox": [ + 306, + 180, + 491, + 193 + ], + "type": "text", + "content": "OUTPUT: Equivalent class of DAG and " + }, + { + "bbox": [ + 306, + 180, + 491, + 193 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_i" + }, + { + "bbox": [ + 306, + 180, + 491, + 193 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 311, + 202, + 545, + 250 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 311, + 202, + 545, + 225 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 202, + 545, + 225 + ], + "spans": [ + { + "bbox": [ + 311, + 202, + 545, + 225 + ], + "type": "text", + "content": "1: Implement the PC algorithm to learn the equivalence class of DAG via v-structure." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 311, + 226, + 545, + 250 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 226, + 545, + 250 + ], + "spans": [ + { + "bbox": [ + 311, + 226, + 545, + 250 + ], + "type": "text", + "content": "2: For each " + }, + { + "bbox": [ + 311, + 226, + 545, + 250 + ], + "type": "inline_equation", + "content": "A \\in \\mathbf{A}" + }, + { + "bbox": [ + 311, + 226, + 545, + 250 + ], + "type": "text", + "content": ", add the edge " + }, + { + "bbox": [ + 311, + 226, + 545, + 250 + ], + "type": "inline_equation", + "content": "E_i \\to A" + }, + { + "bbox": [ + 311, + 226, + 545, + 250 + ], + "type": "text", + "content": " in the graph iff " + }, + { + "bbox": [ + 311, + 226, + 545, + 250 + ], + "type": "inline_equation", + "content": "E_i \\not\\perp A" + }, + { + "bbox": [ + 311, + 226, + 545, + 250 + ], + "type": "text", + "content": " given any subsets." + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 304, + 266, + 547, + 428 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 266, + 547, + 428 + ], + "spans": [ + { + "bbox": [ + 304, + 266, + 547, + 428 + ], + "type": "text", + "content": "iii) Identifying directions via changed causal models. For each adjacent pair " + }, + { + "bbox": [ + 304, + 266, + 547, + 428 + ], + "type": "inline_equation", + "content": "A_{i}" + }, + { + "bbox": [ + 304, + 266, + 547, + 428 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 266, + 547, + 428 + ], + "type": "inline_equation", + "content": "A_{j}" + }, + { + "bbox": [ + 304, + 266, + 547, + 428 + ], + "type": "text", + "content": " such that at least one of them is adjacent to " + }, + { + "bbox": [ + 304, + 266, + 547, + 428 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 304, + 266, + 547, + 428 + ], + "type": "text", + "content": ", we identify the causal directions among " + }, + { + "bbox": [ + 304, + 266, + 547, + 428 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_i" + }, + { + "bbox": [ + 304, + 266, + 547, + 428 + ], + "type": "text", + "content": " via HSIC Norm [12] by testing the independence between " + }, + { + "bbox": [ + 304, + 266, + 547, + 428 + ], + "type": "inline_equation", + "content": "\\{P^{e}(A_{i}|A_{j},\\mathbf{Z})\\}" + }, + { + "bbox": [ + 304, + 266, + 547, + 428 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 266, + 547, + 428 + ], + "type": "inline_equation", + "content": "\\{P^{e}(A_{j}|\\mathbf{Z})\\}" + }, + { + "bbox": [ + 304, + 266, + 547, + 428 + ], + "type": "text", + "content": " if " + }, + { + "bbox": [ + 304, + 266, + 547, + 428 + ], + "type": "inline_equation", + "content": "A_{i}\\gets A_{j}" + }, + { + "bbox": [ + 304, + 266, + 547, + 428 + ], + "type": "text", + "content": ", and that between " + }, + { + "bbox": [ + 304, + 266, + 547, + 428 + ], + "type": "inline_equation", + "content": "\\{P^{e}(A_{j}|A_{i},C)\\}" + }, + { + "bbox": [ + 304, + 266, + 547, + 428 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 266, + 547, + 428 + ], + "type": "inline_equation", + "content": "\\{P^{e}(A_{i}|\\mathbf{Z})\\}" + }, + { + "bbox": [ + 304, + 266, + 547, + 428 + ], + "type": "text", + "content": " for determining the direction between " + }, + { + "bbox": [ + 304, + 266, + 547, + 428 + ], + "type": "inline_equation", + "content": "A_{i}" + }, + { + "bbox": [ + 304, + 266, + 547, + 428 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 266, + 547, + 428 + ], + "type": "inline_equation", + "content": "A_{j}" + }, + { + "bbox": [ + 304, + 266, + 547, + 428 + ], + "type": "text", + "content": ". Here, " + }, + { + "bbox": [ + 304, + 266, + 547, + 428 + ], + "type": "inline_equation", + "content": "\\mathbf{Z}" + }, + { + "bbox": [ + 304, + 266, + 547, + 428 + ], + "type": "text", + "content": " denotes the deconfounding set. In this regard, the direction between " + }, + { + "bbox": [ + 304, + 266, + 547, + 428 + ], + "type": "inline_equation", + "content": "A_{i}" + }, + { + "bbox": [ + 304, + 266, + 547, + 428 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 266, + 547, + 428 + ], + "type": "inline_equation", + "content": "A_{j}" + }, + { + "bbox": [ + 304, + 266, + 547, + 428 + ], + "type": "text", + "content": " is then determined according to the following rule: if " + }, + { + "bbox": [ + 304, + 266, + 547, + 428 + ], + "type": "inline_equation", + "content": "\\hat{\\Delta}_{A_i\\to A_j|\\mathbf{Z},E} < \\hat{\\Delta}_{A_j\\to A_i|C,E}" + }, + { + "bbox": [ + 304, + 266, + 547, + 428 + ], + "type": "text", + "content": ", output " + }, + { + "bbox": [ + 304, + 266, + 547, + 428 + ], + "type": "inline_equation", + "content": "A_{i}\\rightarrow A_{j}" + }, + { + "bbox": [ + 304, + 266, + 547, + 428 + ], + "type": "text", + "content": "; if " + }, + { + "bbox": [ + 304, + 266, + 547, + 428 + ], + "type": "inline_equation", + "content": "\\hat{\\Delta}_{A_i\\to A_j|\\mathbf{Z},E} > \\hat{\\Delta}_{A_j\\to A_i|\\mathbf{Z},E}" + }, + { + "bbox": [ + 304, + 266, + 547, + 428 + ], + "type": "text", + "content": ", output " + }, + { + "bbox": [ + 304, + 266, + 547, + 428 + ], + "type": "inline_equation", + "content": "A_{i}\\leftarrow A_{j}" + }, + { + "bbox": [ + 304, + 266, + 547, + 428 + ], + "type": "text", + "content": "; if " + }, + { + "bbox": [ + 304, + 266, + 547, + 428 + ], + "type": "inline_equation", + "content": "\\hat{\\Delta}_{A_i\\to A_j|\\mathbf{Z},E} > \\alpha" + }, + { + "bbox": [ + 304, + 266, + 547, + 428 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 266, + 547, + 428 + ], + "type": "inline_equation", + "content": "\\hat{\\Delta}_{A_j\\to A_i|\\mathbf{Z},E} > \\alpha" + }, + { + "bbox": [ + 304, + 266, + 547, + 428 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 266, + 547, + 428 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 304, + 266, + 547, + 428 + ], + "type": "text", + "content": " is a threshold, making the direction undetermined." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 438, + 510, + 451 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 438, + 510, + 451 + ], + "spans": [ + { + "bbox": [ + 306, + 438, + 510, + 451 + ], + "type": "text", + "content": "Algorithm 2 Identify causal directions among M." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 461, + 463, + 473 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 461, + 463, + 473 + ], + "spans": [ + { + "bbox": [ + 306, + 461, + 463, + 473 + ], + "type": "text", + "content": "INPUT: " + }, + { + "bbox": [ + 306, + 461, + 463, + 473 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 306, + 461, + 463, + 473 + ], + "type": "text", + "content": "; skeleton and " + }, + { + "bbox": [ + 306, + 461, + 463, + 473 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 306, + 461, + 463, + 473 + ], + "type": "text", + "content": " via Alg.1." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 474, + 460, + 487 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 474, + 460, + 487 + ], + "spans": [ + { + "bbox": [ + 306, + 474, + 460, + 487 + ], + "type": "text", + "content": "OUTPUT: Directed graph among M." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 311, + 495, + 541, + 544 + ], + "type": "list", + "angle": 0, + "index": 22, + "blocks": [ + { + "bbox": [ + 311, + 495, + 536, + 508 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 495, + 536, + 508 + ], + "spans": [ + { + "bbox": [ + 311, + 495, + 536, + 508 + ], + "type": "text", + "content": "1: For each adjacent " + }, + { + "bbox": [ + 311, + 495, + 536, + 508 + ], + "type": "inline_equation", + "content": "(A_{i},A_{j})" + }, + { + "bbox": [ + 311, + 495, + 536, + 508 + ], + "type": "text", + "content": " such that one of " + }, + { + "bbox": [ + 311, + 495, + 536, + 508 + ], + "type": "inline_equation", + "content": "V_{i}\\in \\mathbf{M}" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 312, + 509, + 450, + 519 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 509, + 450, + 519 + ], + "spans": [ + { + "bbox": [ + 312, + 509, + 450, + 519 + ], + "type": "text", + "content": "2: Detect deconfounding set " + }, + { + "bbox": [ + 312, + 509, + 450, + 519 + ], + "type": "inline_equation", + "content": "\\mathbf{Z}" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 312, + 519, + 541, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 519, + 541, + 533 + ], + "spans": [ + { + "bbox": [ + 312, + 519, + 541, + 533 + ], + "type": "text", + "content": "3: Calculate " + }, + { + "bbox": [ + 312, + 519, + 541, + 533 + ], + "type": "inline_equation", + "content": "\\hat{\\Delta}_{A_i\\to A_j|\\mathbf{Z},E} > \\alpha" + }, + { + "bbox": [ + 312, + 519, + 541, + 533 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 312, + 519, + 541, + 533 + ], + "type": "inline_equation", + "content": "\\hat{\\Delta}_{A_j\\to A_i|\\mathbf{Z},E} > \\alpha" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 312, + 533, + 471, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 533, + 471, + 544 + ], + "spans": [ + { + "bbox": [ + 312, + 533, + 471, + 544 + ], + "type": "text", + "content": "4: Determine " + }, + { + "bbox": [ + 312, + 533, + 471, + 544 + ], + "type": "inline_equation", + "content": "A_{i} \\to A_{j}" + }, + { + "bbox": [ + 312, + 533, + 471, + 544 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 312, + 533, + 471, + 544 + ], + "type": "inline_equation", + "content": "A_{j} \\to A_{i}" + }, + { + "bbox": [ + 312, + 533, + 471, + 544 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 21 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 304, + 560, + 545, + 646 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 560, + 545, + 646 + ], + "spans": [ + { + "bbox": [ + 304, + 560, + 545, + 646 + ], + "type": "text", + "content": "iv) Voting for the direction of each edge. For each time " + }, + { + "bbox": [ + 304, + 560, + 545, + 646 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 560, + 545, + 646 + ], + "type": "text", + "content": " we implement Alg. 2 to predict an direction of the edge " + }, + { + "bbox": [ + 304, + 560, + 545, + 646 + ], + "type": "inline_equation", + "content": "(A_{i}, A_{j})" + }, + { + "bbox": [ + 304, + 560, + 545, + 646 + ], + "type": "text", + "content": ". Then we vote for " + }, + { + "bbox": [ + 304, + 560, + 545, + 646 + ], + "type": "inline_equation", + "content": "A_{i} \\rightarrow A_{j}" + }, + { + "bbox": [ + 304, + 560, + 545, + 646 + ], + "type": "text", + "content": " if the frequency " + }, + { + "bbox": [ + 304, + 560, + 545, + 646 + ], + "type": "inline_equation", + "content": "f_{i \\rightarrow j}" + }, + { + "bbox": [ + 304, + 560, + 545, + 646 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 304, + 560, + 545, + 646 + ], + "type": "inline_equation", + "content": "A_{i} \\rightarrow A_{j}" + }, + { + "bbox": [ + 304, + 560, + 545, + 646 + ], + "type": "text", + "content": " surpasses a pre-set threshold " + }, + { + "bbox": [ + 304, + 560, + 545, + 646 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 304, + 560, + 545, + 646 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 304, + 560, + 545, + 646 + ], + "type": "inline_equation", + "content": "0 < \\alpha < 1" + }, + { + "bbox": [ + 304, + 560, + 545, + 646 + ], + "type": "text", + "content": "). Finally, if there exists a cycle of " + }, + { + "bbox": [ + 304, + 560, + 545, + 646 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 560, + 545, + 646 + ], + "type": "text", + "content": " variables: " + }, + { + "bbox": [ + 304, + 560, + 545, + 646 + ], + "type": "inline_equation", + "content": "A_{i_1} \\rightarrow A_{i_2} \\rightarrow \\ldots \\rightarrow A_{i_1}" + }, + { + "bbox": [ + 304, + 560, + 545, + 646 + ], + "type": "text", + "content": ", we orient " + }, + { + "bbox": [ + 304, + 560, + 545, + 646 + ], + "type": "inline_equation", + "content": "(A_{i_k}, A_{i_{k+1}})" + }, + { + "bbox": [ + 304, + 560, + 545, + 646 + ], + "type": "text", + "content": " with the minimum frequency among " + }, + { + "bbox": [ + 304, + 560, + 545, + 646 + ], + "type": "inline_equation", + "content": "\\{f_{i_k, i_{k+1}}, f_{i_N, i_1}\\}" + }, + { + "bbox": [ + 304, + 560, + 545, + 646 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 306, + 656, + 545, + 681 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 656, + 545, + 681 + ], + "spans": [ + { + "bbox": [ + 306, + 656, + 545, + 681 + ], + "type": "text", + "content": "We show that this procedure can generate a DAG over " + }, + { + "bbox": [ + 306, + 656, + 545, + 681 + ], + "type": "inline_equation", + "content": "\\mathbf{A}" + }, + { + "bbox": [ + 306, + 656, + 545, + 681 + ], + "type": "text", + "content": ", with the proof in appendix." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 304, + 689, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 689, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 689, + 545, + 714 + ], + "type": "text", + "content": "Theorem 5 Under definitions and assumptions 1, 2, 3, the learned graph via our CaDAG is a directed acyclic graph." + } + ] + } + ], + "index": 25 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "15684" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 68, + 541, + 175 + ], + "blocks": [ + { + "bbox": [ + 50, + 68, + 541, + 175 + ], + "lines": [ + { + "bbox": [ + 50, + 68, + 541, + 175 + ], + "spans": [ + { + "bbox": [ + 50, + 68, + 541, + 175 + ], + "type": "image", + "image_path": "084fd24618bb2d6424aa8db962c2987f2924f680a8e6a27dc2f13c2f73f95380.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 178, + 546, + 200 + ], + "lines": [ + { + "bbox": [ + 46, + 178, + 546, + 200 + ], + "spans": [ + { + "bbox": [ + 46, + 178, + 546, + 200 + ], + "type": "text", + "content": "Figure 2. Overview of our CAWIM model. We concatenate the MRI features encoded by the MRI Imaging Encoder, preoperative indicators, and the estimated intraoperative indicators obtained from CaDAG to perform the final prediction on OS time." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 72, + 202, + 527, + 371 + ], + "blocks": [ + { + "bbox": [ + 72, + 202, + 527, + 371 + ], + "lines": [ + { + "bbox": [ + 72, + 202, + 527, + 371 + ], + "spans": [ + { + "bbox": [ + 72, + 202, + 527, + 371 + ], + "type": "image", + "image_path": "c6333c92c34b3ea843d0f41e1738c5c833cd42cf5b4bbda28713c5d270735f4e.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 373, + 546, + 429 + ], + "lines": [ + { + "bbox": [ + 46, + 373, + 546, + 429 + ], + "spans": [ + { + "bbox": [ + 46, + 373, + 546, + 429 + ], + "type": "text", + "content": "Figure 3. Schematic diagram of a four-variable case " + }, + { + "bbox": [ + 46, + 373, + 546, + 429 + ], + "type": "inline_equation", + "content": "A_{1}" + }, + { + "bbox": [ + 46, + 373, + 546, + 429 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 373, + 546, + 429 + ], + "type": "inline_equation", + "content": "A_{2}" + }, + { + "bbox": [ + 46, + 373, + 546, + 429 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 373, + 546, + 429 + ], + "type": "inline_equation", + "content": "A_{3}" + }, + { + "bbox": [ + 46, + 373, + 546, + 429 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 373, + 546, + 429 + ], + "type": "inline_equation", + "content": "A_{4}" + }, + { + "bbox": [ + 46, + 373, + 546, + 429 + ], + "type": "text", + "content": ", with the real DAG shown in the top left corner. Through step i) and ii) in the upper part, a Markov Equivalence Class of 3 causal graphs is observed with the direction between " + }, + { + "bbox": [ + 46, + 373, + 546, + 429 + ], + "type": "inline_equation", + "content": "A_{1}" + }, + { + "bbox": [ + 46, + 373, + 546, + 429 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 373, + 546, + 429 + ], + "type": "inline_equation", + "content": "A_{2}" + }, + { + "bbox": [ + 46, + 373, + 546, + 429 + ], + "type": "text", + "content": " undetermined. For step iii), we split the dataset into " + }, + { + "bbox": [ + 46, + 373, + 546, + 429 + ], + "type": "inline_equation", + "content": "(n = 3)" + }, + { + "bbox": [ + 46, + 373, + 546, + 429 + ], + "type": "text", + "content": " domains and consider the 4 circumstances shown fig(a) to (d) in the lower left part. For step iv), we vote for the direction between " + }, + { + "bbox": [ + 46, + 373, + 546, + 429 + ], + "type": "inline_equation", + "content": "A_{1}" + }, + { + "bbox": [ + 46, + 373, + 546, + 429 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 373, + 546, + 429 + ], + "type": "inline_equation", + "content": "A_{2}" + }, + { + "bbox": [ + 46, + 373, + 546, + 429 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 46, + 373, + 546, + 429 + ], + "type": "inline_equation", + "content": "A_{1} \\rightarrow A_{2}" + }, + { + "bbox": [ + 46, + 373, + 546, + 429 + ], + "type": "text", + "content": " according to the number of splitting " + }, + { + "bbox": [ + 46, + 373, + 546, + 429 + ], + "type": "inline_equation", + "content": "(n = 3)" + }, + { + "bbox": [ + 46, + 373, + 546, + 429 + ], + "type": "text", + "content": ", frequency " + }, + { + "bbox": [ + 46, + 373, + 546, + 429 + ], + "type": "inline_equation", + "content": "f_{1\\rightarrow 2} = 2 / 3" + }, + { + "bbox": [ + 46, + 373, + 546, + 429 + ], + "type": "text", + "content": " and threshold " + }, + { + "bbox": [ + 46, + 373, + 546, + 429 + ], + "type": "inline_equation", + "content": "\\alpha = 2 / 3" + }, + { + "bbox": [ + 46, + 373, + 546, + 429 + ], + "type": "text", + "content": ", as shown in the lower right part. Finally, we obtain an inferred DAG with each direction determined." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 435, + 275, + 448 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 435, + 275, + 448 + ], + "spans": [ + { + "bbox": [ + 47, + 435, + 275, + 448 + ], + "type": "text", + "content": "3.2. Intraoperative Imputation for Classification" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 453, + 287, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 453, + 287, + 586 + ], + "spans": [ + { + "bbox": [ + 46, + 453, + 287, + 586 + ], + "type": "text", + "content": "Given MRI image " + }, + { + "bbox": [ + 46, + 453, + 287, + 586 + ], + "type": "inline_equation", + "content": "X \\in \\mathbb{R}^{H \\times W}" + }, + { + "bbox": [ + 46, + 453, + 287, + 586 + ], + "type": "text", + "content": " of a patient that contains multiple slides, we first resample these slides into a fixed " + }, + { + "bbox": [ + 46, + 453, + 287, + 586 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 46, + 453, + 287, + 586 + ], + "type": "text", + "content": " slides for each patient. As each slice is a gray image, we repeat the images for 3 times along the channel dimension; and then we employ a neural network to extract features from each slice. The extracted features are denoted as " + }, + { + "bbox": [ + 46, + 453, + 287, + 586 + ], + "type": "inline_equation", + "content": "F_{1}, \\ldots, F_{K}" + }, + { + "bbox": [ + 46, + 453, + 287, + 586 + ], + "type": "text", + "content": ". In order to aggregate the information of all slices, we divide " + }, + { + "bbox": [ + 46, + 453, + 287, + 586 + ], + "type": "inline_equation", + "content": "F_{1}, \\ldots, F_{K}" + }, + { + "bbox": [ + 46, + 453, + 287, + 586 + ], + "type": "text", + "content": " into " + }, + { + "bbox": [ + 46, + 453, + 287, + 586 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 46, + 453, + 287, + 586 + ], + "type": "text", + "content": " groups according to the acquired position. Each group contains consecutive " + }, + { + "bbox": [ + 46, + 453, + 287, + 586 + ], + "type": "inline_equation", + "content": "\\lfloor K / g \\rfloor" + }, + { + "bbox": [ + 46, + 453, + 287, + 586 + ], + "type": "text", + "content": " slices. Then, we implement a Multi-head Cross-Attention (MCA) mechanism inside each group to aggregate features:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 84, + 592, + 287, + 607 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 84, + 592, + 287, + 607 + ], + "spans": [ + { + "bbox": [ + 84, + 592, + 287, + 607 + ], + "type": "interline_equation", + "content": "\\operatorname {h e a d} _ {j} = \\operatorname {s o f t m a x} \\left(Q ^ {j} \\left(F \\mathbf {W} _ {K} ^ {j}\\right) ^ {T}\\right) \\left(F \\mathbf {W} _ {V} ^ {j}\\right) \\tag {1}", + "image_path": "844858f15127a4060f9eb8bdae238bb66016cdb04028ab015489d795941bf4cb.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 100, + 613, + 287, + 625 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 100, + 613, + 287, + 625 + ], + "spans": [ + { + "bbox": [ + 100, + 613, + 287, + 625 + ], + "type": "interline_equation", + "content": "\\mathbf {M C A} = \\left[ \\text {h e a d} _ {1}; \\dots ; \\text {h e a d} _ {h} \\right] \\mathbf {W} _ {O} \\tag {2}", + "image_path": "5abad3394dd25a58584bef4097b9dd8a48520a87adc5fa977e0854f4ccc6276f.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 629, + 287, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 629, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 46, + 629, + 287, + 689 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 629, + 287, + 689 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 46, + 629, + 287, + 689 + ], + "type": "text", + "content": " is the slice's feature within a group, the query vector " + }, + { + "bbox": [ + 46, + 629, + 287, + 689 + ], + "type": "inline_equation", + "content": "Q^{j}" + }, + { + "bbox": [ + 46, + 629, + 287, + 689 + ], + "type": "text", + "content": " and linear projection layers " + }, + { + "bbox": [ + 46, + 629, + 287, + 689 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_K^j" + }, + { + "bbox": [ + 46, + 629, + 287, + 689 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 629, + 287, + 689 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_V^j" + }, + { + "bbox": [ + 46, + 629, + 287, + 689 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 629, + 287, + 689 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_O" + }, + { + "bbox": [ + 46, + 629, + 287, + 689 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 629, + 287, + 689 + ], + "type": "inline_equation", + "content": "1 \\leq j \\leq h" + }, + { + "bbox": [ + 46, + 629, + 287, + 689 + ], + "type": "text", + "content": " are all learnable parameters. Finally, we concatenate " + }, + { + "bbox": [ + 46, + 629, + 287, + 689 + ], + "type": "inline_equation", + "content": "\\{F_k\\}" + }, + { + "bbox": [ + 46, + 629, + 287, + 689 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 46, + 629, + 287, + 689 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 46, + 629, + 287, + 689 + ], + "type": "text", + "content": " groups as our final image feature " + }, + { + "bbox": [ + 46, + 629, + 287, + 689 + ], + "type": "inline_equation", + "content": "I \\in \\mathbb{R}^{g \\times d}" + }, + { + "bbox": [ + 46, + 629, + 287, + 689 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 46, + 629, + 287, + 689 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 46, + 629, + 287, + 689 + ], + "type": "text", + "content": " denoting the dimension of each " + }, + { + "bbox": [ + 46, + 629, + 287, + 689 + ], + "type": "inline_equation", + "content": "F_k" + }, + { + "bbox": [ + 46, + 629, + 287, + 689 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 46, + 629, + 287, + 689 + ], + "type": "inline_equation", + "content": "k = 1, \\ldots, K" + }, + { + "bbox": [ + 46, + 629, + 287, + 689 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 689, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 689, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 689, + 287, + 713 + ], + "type": "text", + "content": "Afterwards, we sequentially estimate each " + }, + { + "bbox": [ + 46, + 689, + 287, + 713 + ], + "type": "inline_equation", + "content": "A_{j}" + }, + { + "bbox": [ + 46, + 689, + 287, + 713 + ], + "type": "text", + "content": " by preoperative information, as our learned causal graph over " + }, + { + "bbox": [ + 46, + 689, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{A}" + }, + { + "bbox": [ + 46, + 689, + 287, + 713 + ], + "type": "text", + "content": ". So" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 435, + 403, + 448 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 435, + 403, + 448 + ], + "spans": [ + { + "bbox": [ + 306, + 435, + 403, + 448 + ], + "type": "text", + "content": "for each " + }, + { + "bbox": [ + 306, + 435, + 403, + 448 + ], + "type": "inline_equation", + "content": "A_{j}" + }, + { + "bbox": [ + 306, + 435, + 403, + 448 + ], + "type": "text", + "content": ", we predict" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 373, + 458, + 545, + 473 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 373, + 458, + 545, + 473 + ], + "spans": [ + { + "bbox": [ + 373, + 458, + 545, + 473 + ], + "type": "interline_equation", + "content": "\\hat {A} _ {j} = f _ {\\text {i n t r o}} ^ {j} (\\mathbf {B}, I, \\operatorname {P a} (A _ {j})) \\tag {3}", + "image_path": "5f44195c40b53ddf315d3ef026bbd1cbc222b144e95cd59d64d119e9ad8e98cd.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 483, + 545, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 483, + 545, + 586 + ], + "spans": [ + { + "bbox": [ + 304, + 483, + 545, + 586 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 483, + 545, + 586 + ], + "type": "inline_equation", + "content": "\\mathrm{Pa}(A_j)" + }, + { + "bbox": [ + 304, + 483, + 545, + 586 + ], + "type": "text", + "content": " denotes the parent node set of " + }, + { + "bbox": [ + 304, + 483, + 545, + 586 + ], + "type": "inline_equation", + "content": "A_{j}" + }, + { + "bbox": [ + 304, + 483, + 545, + 586 + ], + "type": "text", + "content": ". We use cross entropy loss to optimize " + }, + { + "bbox": [ + 304, + 483, + 545, + 586 + ], + "type": "inline_equation", + "content": "f_{\\mathrm{intra}}^j" + }, + { + "bbox": [ + 304, + 483, + 545, + 586 + ], + "type": "text", + "content": " if " + }, + { + "bbox": [ + 304, + 483, + 545, + 586 + ], + "type": "inline_equation", + "content": "A_{j}" + }, + { + "bbox": [ + 304, + 483, + 545, + 586 + ], + "type": "text", + "content": " is categorical; and optimize " + }, + { + "bbox": [ + 304, + 483, + 545, + 586 + ], + "type": "inline_equation", + "content": "f_{\\mathrm{intra}}^j" + }, + { + "bbox": [ + 304, + 483, + 545, + 586 + ], + "type": "text", + "content": " if " + }, + { + "bbox": [ + 304, + 483, + 545, + 586 + ], + "type": "inline_equation", + "content": "A_{j}" + }, + { + "bbox": [ + 304, + 483, + 545, + 586 + ], + "type": "text", + "content": " is continuous. After obtaining all predicted " + }, + { + "bbox": [ + 304, + 483, + 545, + 586 + ], + "type": "inline_equation", + "content": "\\{\\hat{A}_i\\}_{i = 1}^d" + }, + { + "bbox": [ + 304, + 483, + 545, + 586 + ], + "type": "text", + "content": ", we use " + }, + { + "bbox": [ + 304, + 483, + 545, + 586 + ], + "type": "inline_equation", + "content": "(I,\\mathbf{B},\\hat{\\mathbf{A}})" + }, + { + "bbox": [ + 304, + 483, + 545, + 586 + ], + "type": "text", + "content": " to predict " + }, + { + "bbox": [ + 304, + 483, + 545, + 586 + ], + "type": "inline_equation", + "content": "Y" + }, + { + "bbox": [ + 304, + 483, + 545, + 586 + ], + "type": "text", + "content": " via " + }, + { + "bbox": [ + 304, + 483, + 545, + 586 + ], + "type": "inline_equation", + "content": "f_{\\theta}(I,\\mathbf{B},\\hat{\\mathbf{A}})" + }, + { + "bbox": [ + 304, + 483, + 545, + 586 + ], + "type": "text", + "content": " and obtain " + }, + { + "bbox": [ + 304, + 483, + 545, + 586 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 304, + 483, + 545, + 586 + ], + "type": "text", + "content": " by optimizing the cross-entropy " + }, + { + "bbox": [ + 304, + 483, + 545, + 586 + ], + "type": "inline_equation", + "content": "\\mathcal{L}(y,f_{\\theta}(I,\\mathbf{B},\\hat{\\mathbf{A}}))" + }, + { + "bbox": [ + 304, + 483, + 545, + 586 + ], + "type": "text", + "content": ". During test stage, given a new sample " + }, + { + "bbox": [ + 304, + 483, + 545, + 586 + ], + "type": "inline_equation", + "content": "(x,\\mathbf{b})" + }, + { + "bbox": [ + 304, + 483, + 545, + 586 + ], + "type": "text", + "content": ", we first extract " + }, + { + "bbox": [ + 304, + 483, + 545, + 586 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 304, + 483, + 545, + 586 + ], + "type": "text", + "content": " and then sequentially estimate " + }, + { + "bbox": [ + 304, + 483, + 545, + 586 + ], + "type": "inline_equation", + "content": "\\mathbf{a}_i" + }, + { + "bbox": [ + 304, + 483, + 545, + 586 + ], + "type": "text", + "content": " and finally predict " + }, + { + "bbox": [ + 304, + 483, + 545, + 586 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 304, + 483, + 545, + 586 + ], + "type": "text", + "content": " using " + }, + { + "bbox": [ + 304, + 483, + 545, + 586 + ], + "type": "inline_equation", + "content": "f_{\\theta}(x,\\mathbf{b},\\hat{\\mathbf{a}})" + }, + { + "bbox": [ + 304, + 483, + 545, + 586 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 597, + 387, + 610 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 597, + 387, + 610 + ], + "spans": [ + { + "bbox": [ + 306, + 597, + 387, + 610 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "content": "Dataset overview. Ethics Committee of Zhongshan Hospital, Fudan University approved the protocol of this study and waived the requirement for patient-informed consent (B2021-325R). Because open source data sets generally lack sufficient intraoperative information, it takes us around ten years to collect and build our own medical dataset. Particularly, we conduct a search through the medical records in the hospital information system, and build a dataset with" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15685" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 275 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 275 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 275 + ], + "type": "text", + "content": "439 patients infected by primary liver cancer. Long-term medical data of every patient has been recorded from the first scan to the cancer-related death. Till the submission of this manuscript, around " + }, + { + "bbox": [ + 46, + 72, + 289, + 275 + ], + "type": "inline_equation", + "content": "90\\%" + }, + { + "bbox": [ + 46, + 72, + 289, + 275 + ], + "type": "text", + "content": " patients have unfortunately passed away. This gives us the ground-truth OS labels. Information like MRI scanning before the surgery, blood test and operation reports, follow-up records are also collected. After removing samples with missing information, 361 patients were enrolled, with 306 " + }, + { + "bbox": [ + 46, + 72, + 289, + 275 + ], + "type": "inline_equation", + "content": "(84.8\\%)" + }, + { + "bbox": [ + 46, + 72, + 289, + 275 + ], + "type": "text", + "content": " men, and the average age was 53.7 " + }, + { + "bbox": [ + 46, + 72, + 289, + 275 + ], + "type": "inline_equation", + "content": "(\\pm 11.7)" + }, + { + "bbox": [ + 46, + 72, + 289, + 275 + ], + "type": "text", + "content": " years. Specifically, each patient has MRI scanning with four modalities, including " + }, + { + "bbox": [ + 46, + 72, + 289, + 275 + ], + "type": "inline_equation", + "content": "\\mathrm{T}_{1}" + }, + { + "bbox": [ + 46, + 72, + 289, + 275 + ], + "type": "text", + "content": " weighted images " + }, + { + "bbox": [ + 46, + 72, + 289, + 275 + ], + "type": "inline_equation", + "content": "(\\mathrm{T}_{1}" + }, + { + "bbox": [ + 46, + 72, + 289, + 275 + ], + "type": "text", + "content": "-WI), " + }, + { + "bbox": [ + 46, + 72, + 289, + 275 + ], + "type": "inline_equation", + "content": "\\mathrm{T}_{2}" + }, + { + "bbox": [ + 46, + 72, + 289, + 275 + ], + "type": "text", + "content": " weighted images " + }, + { + "bbox": [ + 46, + 72, + 289, + 275 + ], + "type": "inline_equation", + "content": "(\\mathrm{T}_{2}" + }, + { + "bbox": [ + 46, + 72, + 289, + 275 + ], + "type": "text", + "content": "-WI), " + }, + { + "bbox": [ + 46, + 72, + 289, + 275 + ], + "type": "inline_equation", + "content": "\\mathrm{T}_{1}" + }, + { + "bbox": [ + 46, + 72, + 289, + 275 + ], + "type": "text", + "content": " in-phase " + }, + { + "bbox": [ + 46, + 72, + 289, + 275 + ], + "type": "inline_equation", + "content": "(\\mathrm{T}_{1}" + }, + { + "bbox": [ + 46, + 72, + 289, + 275 + ], + "type": "text", + "content": "-IP) and contrast-enhanced " + }, + { + "bbox": [ + 46, + 72, + 289, + 275 + ], + "type": "inline_equation", + "content": "\\mathrm{T}_{1}" + }, + { + "bbox": [ + 46, + 72, + 289, + 275 + ], + "type": "text", + "content": " at the arterial phase " + }, + { + "bbox": [ + 46, + 72, + 289, + 275 + ], + "type": "inline_equation", + "content": "(\\mathrm{T}_{1}" + }, + { + "bbox": [ + 46, + 72, + 289, + 275 + ], + "type": "text", + "content": "ce-AP). Following the guidance of medical experts, we select 11 preoperative and 20 intraoperative indicators that are related to OS time for our study. Please refer to the appendix for details." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 277, + 289, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 277, + 289, + 422 + ], + "spans": [ + { + "bbox": [ + 46, + 277, + 289, + 422 + ], + "type": "text", + "content": "Data preprocessing. In our dataset, the survival time ranges from 0 to 130 months, which are cauterized into four classes: short-term survival (" + }, + { + "bbox": [ + 46, + 277, + 289, + 422 + ], + "type": "inline_equation", + "content": "\\leq 36" + }, + { + "bbox": [ + 46, + 277, + 289, + 422 + ], + "type": "text", + "content": " months), middle-short-term survival (between 36 and 72 months), middle-long-term survival (between 72 and 108 months), and long-term survival (" + }, + { + "bbox": [ + 46, + 277, + 289, + 422 + ], + "type": "inline_equation", + "content": "\\geq 108" + }, + { + "bbox": [ + 46, + 277, + 289, + 422 + ], + "type": "text", + "content": " months). In our paper, we consider the classification task: given a new sample, our goal is to classify which class of OS time this sample belongs to. To extract " + }, + { + "bbox": [ + 46, + 277, + 289, + 422 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 46, + 277, + 289, + 422 + ], + "type": "text", + "content": " from MRI image, we resample " + }, + { + "bbox": [ + 46, + 277, + 289, + 422 + ], + "type": "inline_equation", + "content": "K = 20" + }, + { + "bbox": [ + 46, + 277, + 289, + 422 + ], + "type": "text", + "content": " slices and employ Resnet-34 for feature extraction. We set " + }, + { + "bbox": [ + 46, + 277, + 289, + 422 + ], + "type": "inline_equation", + "content": "g = 4" + }, + { + "bbox": [ + 46, + 277, + 289, + 422 + ], + "type": "text", + "content": " and the dimension " + }, + { + "bbox": [ + 46, + 277, + 289, + 422 + ], + "type": "inline_equation", + "content": "F_{k}" + }, + { + "bbox": [ + 46, + 277, + 289, + 422 + ], + "type": "text", + "content": " as 512. For " + }, + { + "bbox": [ + 46, + 277, + 289, + 422 + ], + "type": "inline_equation", + "content": "f_{\\mathrm{intro}}^{j}" + }, + { + "bbox": [ + 46, + 277, + 289, + 422 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 277, + 289, + 422 + ], + "type": "inline_equation", + "content": "f_{\\theta}" + }, + { + "bbox": [ + 46, + 277, + 289, + 422 + ], + "type": "text", + "content": ", we employ a multi-layer perceptron (MLP) for parameterization." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 422, + 289, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 422, + 289, + 506 + ], + "spans": [ + { + "bbox": [ + 46, + 422, + 289, + 506 + ], + "type": "text", + "content": "We encode categorical covariates into dummy variables and implement zero-mean normalization for continuous variables. For MRI images, we resize them into " + }, + { + "bbox": [ + 46, + 422, + 289, + 506 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 46, + 422, + 289, + 506 + ], + "type": "text", + "content": " as input. We split the whole dataset into 5 folds. To remove the effect of imbalance across classes and randomness, we adopt average precision, recall, and " + }, + { + "bbox": [ + 46, + 422, + 289, + 506 + ], + "type": "inline_equation", + "content": "\\mathrm{F_1}" + }, + { + "bbox": [ + 46, + 422, + 289, + 506 + ], + "type": "text", + "content": "-Score measurements over five folds for classification evaluation." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 508, + 289, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 508, + 289, + 640 + ], + "spans": [ + { + "bbox": [ + 46, + 508, + 289, + 640 + ], + "type": "text", + "content": "Implementation details. We use the ImageNet pretrained Resnet-34 to initialize our CNN encoder. We train 150 epochs for all methods with " + }, + { + "bbox": [ + 46, + 508, + 289, + 640 + ], + "type": "inline_equation", + "content": "\\mathrm{T}_{1}" + }, + { + "bbox": [ + 46, + 508, + 289, + 640 + ], + "type": "text", + "content": "ce-AP modality, using SGD as the optimizer with learning rate " + }, + { + "bbox": [ + 46, + 508, + 289, + 640 + ], + "type": "inline_equation", + "content": "3e - 3" + }, + { + "bbox": [ + 46, + 508, + 289, + 640 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 508, + 289, + 640 + ], + "type": "inline_equation", + "content": "3e - 2" + }, + { + "bbox": [ + 46, + 508, + 289, + 640 + ], + "type": "text", + "content": " for CNN encoder and later MLP, respectively. We set weight decay to " + }, + { + "bbox": [ + 46, + 508, + 289, + 640 + ], + "type": "inline_equation", + "content": "5e - 4" + }, + { + "bbox": [ + 46, + 508, + 289, + 640 + ], + "type": "text", + "content": " and nesterov momentum factor to 0.9. Besides, we decrease the learning rate every 60 epochs by a factor of 1/10. We set the Batch size to 8. In our voting mechanism of CaDAG, we in total implement Alg. 2 for seven times and set the voting threshold as 4/7. Our model is trained with Pytorch on NVIDIA GeForce RTX 3090 GPUs." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 641, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 641, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 641, + 289, + 715 + ], + "type": "text", + "content": "Compared Methods. We compare our method with four baseline methods: i) Random: random guess among four classes as a trivial baseline, of which the expected " + }, + { + "bbox": [ + 46, + 641, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\mathrm{F_1}" + }, + { + "bbox": [ + 46, + 641, + 289, + 715 + ], + "type": "text", + "content": " score is " + }, + { + "bbox": [ + 46, + 641, + 289, + 715 + ], + "type": "inline_equation", + "content": "25\\%" + }, + { + "bbox": [ + 46, + 641, + 289, + 715 + ], + "type": "text", + "content": "; ii) MRI: train the end-to-end method with " + }, + { + "bbox": [ + 46, + 641, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\mathrm{T_1}" + }, + { + "bbox": [ + 46, + 641, + 289, + 715 + ], + "type": "text", + "content": "ce-AP only as input for classification; iii) MRI+pre: additionally take preoperative features as input, on the basis of the" + } + ] + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 307, + 83, + 543, + 156 + ], + "blocks": [ + { + "bbox": [ + 315, + 70, + 536, + 81 + ], + "lines": [ + { + "bbox": [ + 315, + 70, + 536, + 81 + ], + "spans": [ + { + "bbox": [ + 315, + 70, + 536, + 81 + ], + "type": "text", + "content": "Table 1. Comparisons of Our CAWIM with Other Baselines." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 307, + 83, + 543, + 156 + ], + "lines": [ + { + "bbox": [ + 307, + 83, + 543, + 156 + ], + "spans": [ + { + "bbox": [ + 307, + 83, + 543, + 156 + ], + "type": "table", + "html": "
ModelPrecisionRecallF1-Score
(a)Random26.04±2.1024.54±0.8023.15±0.64
(b)MRI40.79±7.9733.45±2.6631.88±3.92
(c)MRI+Pre36.99±5.3035.62±3.5033.83±3.64
(d)gt Intra41.26±2.7441.62±1.9539.77±2.47
(e)CAWIM45.58±5.8643.70±5.8942.21±4.92
", + "image_path": "38acba7d80a01d78bdbf124f3e4f7946bb4a5a2d4d3c0fe63ac7a1b5d87dad43.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 166, + 545, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 166, + 545, + 240 + ], + "spans": [ + { + "bbox": [ + 304, + 166, + 545, + 240 + ], + "type": "text", + "content": "method MRI; iv) ground-truth (GT) Intra: only take the ground-truth intraoperative features as input. It is worth to mention that although the intraoperative features are not allowed to use during test stage, the comparison with method iv (i.e., GT Intra) can provide information that to which extent, the intraoperative features can help OS time prediction." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 247, + 426, + 260 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 247, + 426, + 260 + ], + "spans": [ + { + "bbox": [ + 305, + 247, + 426, + 260 + ], + "type": "text", + "content": "4.1. Results and Analysis" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 266, + 545, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 266, + 545, + 483 + ], + "spans": [ + { + "bbox": [ + 304, + 266, + 545, + 483 + ], + "type": "text", + "content": "Table 1 records comparisons of our proposed CAWIM (Tab. 1(d)) with other baseline models. (Tab. 1(a), Tab. 1(b), and Tab. 1(c)). Obviously, our CAWIM constantly outperform others in all metrics. Specifically, CAWIM improved " + }, + { + "bbox": [ + 304, + 266, + 545, + 483 + ], + "type": "inline_equation", + "content": "\\text{MRI} + \\text{Pre}" + }, + { + "bbox": [ + 304, + 266, + 545, + 483 + ], + "type": "text", + "content": " by nearly " + }, + { + "bbox": [ + 304, + 266, + 545, + 483 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 304, + 266, + 545, + 483 + ], + "type": "text", + "content": " in terms of " + }, + { + "bbox": [ + 304, + 266, + 545, + 483 + ], + "type": "inline_equation", + "content": "\\text{F}_1" + }, + { + "bbox": [ + 304, + 266, + 545, + 483 + ], + "type": "text", + "content": "-Score, which can be contributed to the prediction of intraoperative features equipped with causally aware prediction mechanism. Without this mechanism, using MRI and preoperative features only can outperform random guess by only " + }, + { + "bbox": [ + 304, + 266, + 545, + 483 + ], + "type": "inline_equation", + "content": "8.33\\%" + }, + { + "bbox": [ + 304, + 266, + 545, + 483 + ], + "type": "text", + "content": ", which demonstrates the limited information of the image for OS time. In contrast, it is also interesting to note from Tab. 1(d) that using intraoperative features can be more informative than MRI features, in terms of OS prediction, which is natural as the attributes that described the functions and textures of the liver could affect the postoperative recovery and thus the OS time. Finally, our CAWIM (Tab. 1(e)) that combines the information of intraoperative features and the information from images, can achieve further improvement." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 491, + 400, + 503 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 491, + 400, + 503 + ], + "spans": [ + { + "bbox": [ + 305, + 491, + 400, + 503 + ], + "type": "text", + "content": "4.2. Ablation Study" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 509, + 545, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 509, + 545, + 665 + ], + "spans": [ + { + "bbox": [ + 304, + 509, + 545, + 665 + ], + "type": "text", + "content": "To test the effectiveness of preoperative features, the particularly intraoperative features, and our causally-aware module for OS time prediction, we implement ablation studies and summarize the results in Tab. 2. Specifically, Tab. 2(a), i.e., w/o CaDAG replaces our causally-aware module with an end-to-end estimation network from (MRI, preoperative features) to intraoperative features; Tab. 2(b), i.e., w/o Intra is the same to Tab. 1(c), i.e., MRI+pre in Tab. 1 that directly predict OS time using MRI and preoperative features without estimated intraoperative features; Tab. 2(c), i.e., w/o Pre is the same to our CAWIM except that it does not utilize preoperative features for intraoperative estimation and OS prediction." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 665, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 547, + 715 + ], + "type": "text", + "content": "The results in Tab. 2 show that the deletion of each module can lead to a descent of all metrics. Specifically, Tab. 2(a) shows a significant performance drop by " + }, + { + "bbox": [ + 304, + 665, + 547, + 715 + ], + "type": "inline_equation", + "content": "5.6\\%" + }, + { + "bbox": [ + 304, + 665, + 547, + 715 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 304, + 665, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\mathrm{F_1}" + }, + { + "bbox": [ + 304, + 665, + 547, + 715 + ], + "type": "text", + "content": "-Score, which indicates the effectiveness of our causally" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "15686" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 262 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 262 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 262 + ], + "type": "text", + "content": "aware module in estimating intraoperative features in OS time prediction. This phenomena can be explained by more accurate intraoperative estimation equipped with our causally-aware module, as shown in Fig. 4 of " + }, + { + "bbox": [ + 46, + 72, + 289, + 262 + ], + "type": "inline_equation", + "content": "\\mathrm{F_1}" + }, + { + "bbox": [ + 46, + 72, + 289, + 262 + ], + "type": "text", + "content": "-score, accuracy and " + }, + { + "bbox": [ + 46, + 72, + 289, + 262 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 46, + 72, + 289, + 262 + ], + "type": "text", + "content": "-squared score. We will leave detailed discussions in Section of \"Further Discussion\". Tab. 2(b), as the same to MRI+pre in Tab. 1, shows a " + }, + { + "bbox": [ + 46, + 72, + 289, + 262 + ], + "type": "inline_equation", + "content": "8.3\\%" + }, + { + "bbox": [ + 46, + 72, + 289, + 262 + ], + "type": "text", + "content": " performance drop, which indicates the additional information provided by the MRI image and preoperative features. Finally, the " + }, + { + "bbox": [ + 46, + 72, + 289, + 262 + ], + "type": "inline_equation", + "content": "3.78\\%" + }, + { + "bbox": [ + 46, + 72, + 289, + 262 + ], + "type": "text", + "content": " degradation of Tab. 2(c) compared to ours validates the effectiveness of preoperative features in OS prediction, which can also be explained by the estimation results of intraoperative features in Fig. 4. In a word, we can observe a significant drop of " + }, + { + "bbox": [ + 46, + 72, + 289, + 262 + ], + "type": "inline_equation", + "content": "F_{1}" + }, + { + "bbox": [ + 46, + 72, + 289, + 262 + ], + "type": "text", + "content": "-score if we remove causally-aware module and intraoperative features, which can demonstrate the effectiveness of our methods." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 264, + 289, + 432 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 264, + 289, + 432 + ], + "spans": [ + { + "bbox": [ + 46, + 264, + 289, + 432 + ], + "type": "text", + "content": "Besides, it is also interesting to compare Tab. 2(b) and Tab. 2(c) with we observe that the performance drop for " + }, + { + "bbox": [ + 46, + 264, + 289, + 432 + ], + "type": "inline_equation", + "content": "w/o" + }, + { + "bbox": [ + 46, + 264, + 289, + 432 + ], + "type": "text", + "content": " Intra is more significant than that for " + }, + { + "bbox": [ + 46, + 264, + 289, + 432 + ], + "type": "inline_equation", + "content": "w/o" + }, + { + "bbox": [ + 46, + 264, + 289, + 432 + ], + "type": "text", + "content": " Pre, with respectively an average of " + }, + { + "bbox": [ + 46, + 264, + 289, + 432 + ], + "type": "inline_equation", + "content": "8.4\\%" + }, + { + "bbox": [ + 46, + 264, + 289, + 432 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 264, + 289, + 432 + ], + "type": "inline_equation", + "content": "3.8\\%" + }, + { + "bbox": [ + 46, + 264, + 289, + 432 + ], + "type": "text", + "content": " drop in " + }, + { + "bbox": [ + 46, + 264, + 289, + 432 + ], + "type": "inline_equation", + "content": "F_1" + }, + { + "bbox": [ + 46, + 264, + 289, + 432 + ], + "type": "text", + "content": "-Score. This again, indicates more information of intraoperative indexes provided than preoperative indexes in predicting OS time. Indeed, we obtained Precision " + }, + { + "bbox": [ + 46, + 264, + 289, + 432 + ], + "type": "inline_equation", + "content": "27.68\\%" + }, + { + "bbox": [ + 46, + 264, + 289, + 432 + ], + "type": "text", + "content": ", Recall " + }, + { + "bbox": [ + 46, + 264, + 289, + 432 + ], + "type": "inline_equation", + "content": "26.65\\%" + }, + { + "bbox": [ + 46, + 264, + 289, + 432 + ], + "type": "text", + "content": " and F1-Score " + }, + { + "bbox": [ + 46, + 264, + 289, + 432 + ], + "type": "inline_equation", + "content": "23.85\\%" + }, + { + "bbox": [ + 46, + 264, + 289, + 432 + ], + "type": "text", + "content": " if we only predicted OS time using preoperative features, which is comparable to the result from random guess. On the other hand, we can observe from Fig. 4 that preoperative indexes are beneficial to the estimation of intraoperative features, which indicating the necessity of using intraoperative features as a bridge between preoperative information and final OS time prediction." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 442, + 258, + 455 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 442, + 258, + 455 + ], + "spans": [ + { + "bbox": [ + 47, + 442, + 258, + 455 + ], + "type": "text", + "content": "4.3. Medical Interpretation of Causal Graph" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 461, + 287, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 461, + 287, + 593 + ], + "spans": [ + { + "bbox": [ + 46, + 461, + 287, + 593 + ], + "type": "text", + "content": "To further explain the effectiveness of our causally-aware module, we in this subsection present our learned causal graph over intraoperative features in Fig. 5 and the corresponding medical interpretation of learned causal relations. In general, most of the parent-child relationships in our CaDAG are in line with common sense or can be explained by prior medical knowledge, e.g., sum of tumor diameter is determined by number of tumors, the edge from " + }, + { + "bbox": [ + 46, + 461, + 287, + 593 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 46, + 461, + 287, + 593 + ], + "type": "text", + "content": "-score to " + }, + { + "bbox": [ + 46, + 461, + 287, + 593 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 46, + 461, + 287, + 593 + ], + "type": "text", + "content": "-score also correspond to the strong correlation between fibrosis and hepatitis (please refer to appendix for detailed descriptions of intraoperative indexes)." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 594, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 288, + 713 + ], + "type": "text", + "content": "Further, those indicators with non-visual features that access severity of pathology are often determined by other attributes of the liver; therefore, these features often come up in deeper layers in Fig. 5. For instance, " + }, + { + "bbox": [ + 46, + 594, + 288, + 713 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 46, + 594, + 288, + 713 + ], + "type": "text", + "content": "-score and " + }, + { + "bbox": [ + 46, + 594, + 288, + 713 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 46, + 594, + 288, + 713 + ], + "type": "text", + "content": "-score reflect the level of inflammation and fibrosis that are artificially defined according to Metavir scoring system [2]. Those indicators tend to rely on other intraoperative features besides merely MRI and preoperative information, e.g. clinicopathologic hepatocirrhosis is reasoned from hepatocirrhosis and other covariants. For this reason, the addi" + } + ] + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 307, + 127, + 553, + 189 + ], + "blocks": [ + { + "bbox": [ + 304, + 70, + 547, + 125 + ], + "lines": [ + { + "bbox": [ + 304, + 70, + 547, + 125 + ], + "spans": [ + { + "bbox": [ + 304, + 70, + 547, + 125 + ], + "type": "text", + "content": "Table 2. Results of Ablation Studies. (a) estimates each intraoperative features naively from (MRI, preoperative indexes), without the causally-aware module. (b) is the same to " + }, + { + "bbox": [ + 304, + 70, + 547, + 125 + ], + "type": "inline_equation", + "content": "M R I + P r e" + }, + { + "bbox": [ + 304, + 70, + 547, + 125 + ], + "type": "text", + "content": " in Tab. 1. (c) is the CAWIM without preoperative features for intraoperative indexes estimation and OS time prediction." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 307, + 127, + 553, + 189 + ], + "lines": [ + { + "bbox": [ + 307, + 127, + 553, + 189 + ], + "spans": [ + { + "bbox": [ + 307, + 127, + 553, + 189 + ], + "type": "table", + "html": "
ModelPrecisionRecallF1-Score
(a)w/o CaDAG36.98±2.3837.78±1.9136.62±1.87
(b)w/o Intra36.99±5.3035.62±3.5033.83±3.64
(c)w/o Pre38.92±7.7141.57±4.5338.43±5.84
(d)CAWIM45.58±5.8643.70±5.8942.21±4.92
", + "image_path": "fc83786945b1467c438a851b6f2288ead83388d9ffb12c0eff6fb633166358c2.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 197, + 547, + 269 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 197, + 547, + 269 + ], + "spans": [ + { + "bbox": [ + 304, + 197, + 547, + 269 + ], + "type": "text", + "content": "tional leverage of hepatocirrhosis lead to more accurate estimation of clinicopathologic hepatocirrhosis, as shown in Fig. 4 (a). On the other hand, we also observe some associations that cannot be explained well, e.g. the causal relationship between sum of tumor diameter and cell type, which may due to the existence of unobserved confounders." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 275, + 390, + 286 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 275, + 390, + 286 + ], + "spans": [ + { + "bbox": [ + 306, + 275, + 390, + 286 + ], + "type": "text", + "content": "4.4. Visualization" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 293, + 547, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 293, + 547, + 413 + ], + "spans": [ + { + "bbox": [ + 304, + 293, + 547, + 413 + ], + "type": "text", + "content": "In this section, we visualize the high-response area of our CAWIM Fig. 6(c) and the version without CaDAG Fig. 6(b) using Grad-CAM [29]. As shown in Fig. 6, the detected regions of our method can be more concentrated on the liver. This result can be explained with Fig.4 and Fig. 5 in a more complementary way. Specifically, as the learned causal graph is medical explainable and with better estimation of intraoperative features, the model is driven to locate on liver-related regions. More visualization results are left in appendix due to space limit." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 420, + 418, + 431 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 420, + 418, + 431 + ], + "spans": [ + { + "bbox": [ + 306, + 420, + 418, + 431 + ], + "type": "text", + "content": "4.5. Further Discussion" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 438, + 547, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 438, + 547, + 641 + ], + "spans": [ + { + "bbox": [ + 304, + 438, + 547, + 641 + ], + "type": "text", + "content": "Intraoperative prediction. In our method, the estimation of intraoperative indicators are important for final prediction. Fig. 4 shows the metrics of some typical intraoperative prediction by i) " + }, + { + "bbox": [ + 304, + 438, + 547, + 641 + ], + "type": "inline_equation", + "content": "w/o \\, CaDAG" + }, + { + "bbox": [ + 304, + 438, + 547, + 641 + ], + "type": "text", + "content": " and ii) " + }, + { + "bbox": [ + 304, + 438, + 547, + 641 + ], + "type": "inline_equation", + "content": "w/o \\, Pre" + }, + { + "bbox": [ + 304, + 438, + 547, + 641 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 304, + 438, + 547, + 641 + ], + "type": "inline_equation", + "content": "F_1" + }, + { + "bbox": [ + 304, + 438, + 547, + 641 + ], + "type": "text", + "content": "-Score and accuracy for discrete variants (Fig. 4 (a, b)) and " + }, + { + "bbox": [ + 304, + 438, + 547, + 641 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 304, + 438, + 547, + 641 + ], + "type": "text", + "content": "-square for continuous ones (Fig. 4 (c)). Fig. 4 (a, b) shows that our CaDAG can generally enhance the estimation of categorical variables hepatocirrhosis, cell type, differentiation, etc. Nevertheless, we also find that such improvements will decrease if corresponding indicators are in deep layers of the graph in Fig. 5. For example, the accuracy of " + }, + { + "bbox": [ + 304, + 438, + 547, + 641 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 304, + 438, + 547, + 641 + ], + "type": "text", + "content": "-Score and " + }, + { + "bbox": [ + 304, + 438, + 547, + 641 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 304, + 438, + 547, + 641 + ], + "type": "text", + "content": "-score is only comparable and even worse than that of " + }, + { + "bbox": [ + 304, + 438, + 547, + 641 + ], + "type": "inline_equation", + "content": "w/o \\, CaDAG" + }, + { + "bbox": [ + 304, + 438, + 547, + 641 + ], + "type": "text", + "content": ". A possible reason is the accumulated error along the directed path for estimation. On the other hand, as shown by the results of " + }, + { + "bbox": [ + 304, + 438, + 547, + 641 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 304, + 438, + 547, + 641 + ], + "type": "text", + "content": "-square (closer to 1, the better the prediction) does not bring about notable improvement in regression tasks, (e.g., ascites and tumor diameter)." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 642, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 642, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 642, + 547, + 713 + ], + "type": "text", + "content": "We also observe an evident performance degradation of " + }, + { + "bbox": [ + 304, + 642, + 547, + 713 + ], + "type": "inline_equation", + "content": "w / o" + }, + { + "bbox": [ + 304, + 642, + 547, + 713 + ], + "type": "text", + "content": " Pre for most of those discrete or continuous variants. This implies that although operative alone has limited contribution to OS time prediction, it still matters in our CAWIM, because of the less missing information due to shorter time interval from the preoperative stage to the" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15687" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 37, + 78, + 202, + 187 + ], + "blocks": [ + { + "bbox": [ + 37, + 78, + 202, + 187 + ], + "lines": [ + { + "bbox": [ + 37, + 78, + 202, + 187 + ], + "spans": [ + { + "bbox": [ + 37, + 78, + 202, + 187 + ], + "type": "image", + "image_path": "01d26054c451b1c255ec9bdb018d00e30cb4ff489eaf6f6bc84a3c82a4a2f274.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 102, + 190, + 115, + 201 + ], + "lines": [ + { + "bbox": [ + 102, + 190, + 115, + 201 + ], + "spans": [ + { + "bbox": [ + 102, + 190, + 115, + 201 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 210, + 79, + 380, + 187 + ], + "blocks": [ + { + "bbox": [ + 210, + 79, + 380, + 187 + ], + "lines": [ + { + "bbox": [ + 210, + 79, + 380, + 187 + ], + "spans": [ + { + "bbox": [ + 210, + 79, + 380, + 187 + ], + "type": "image", + "image_path": "f3d9ee8a3d02e3efb04474ac8234a698f83a470aadc7468a30acd9ec12c36a4c.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 291, + 190, + 305, + 201 + ], + "lines": [ + { + "bbox": [ + 291, + 190, + 305, + 201 + ], + "spans": [ + { + "bbox": [ + 291, + 190, + 305, + 201 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 45, + 202, + 547, + 247 + ], + "lines": [ + { + "bbox": [ + 45, + 202, + 547, + 247 + ], + "spans": [ + { + "bbox": [ + 45, + 202, + 547, + 247 + ], + "type": "text", + "content": "Figure 4. Quantity Results of Intraoperative Prediction. Intraoperative prediction are categorized into classification ((a) and (b)) and regression problem ((c)) according to the data form. (a) and (b) shows the prediction " + }, + { + "bbox": [ + 45, + 202, + 547, + 247 + ], + "type": "inline_equation", + "content": "\\mathrm{F}_1" + }, + { + "bbox": [ + 45, + 202, + 547, + 247 + ], + "type": "text", + "content": "-Score and accuracy rate of the discrete variants respectively. The numbers after the variant name in (a) and (b) are the number of classes, i.e. hepatocirrhosis is a 6-category classification task. (c) shows the R-squared score for continuous variants." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 389, + 79, + 539, + 187 + ], + "blocks": [ + { + "bbox": [ + 389, + 79, + 539, + 187 + ], + "lines": [ + { + "bbox": [ + 389, + 79, + 539, + 187 + ], + "spans": [ + { + "bbox": [ + 389, + 79, + 539, + 187 + ], + "type": "image", + "image_path": "fd6aa08290bb102b2db018cf78102e9766530d0ae8a7d21ec1a482407f4086db.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 459, + 190, + 470, + 201 + ], + "lines": [ + { + "bbox": [ + 459, + 190, + 470, + 201 + ], + "spans": [ + { + "bbox": [ + 459, + 190, + 470, + 201 + ], + "type": "text", + "content": "(c)" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 66, + 262, + 265, + 369 + ], + "blocks": [ + { + "bbox": [ + 66, + 262, + 265, + 369 + ], + "lines": [ + { + "bbox": [ + 66, + 262, + 265, + 369 + ], + "spans": [ + { + "bbox": [ + 66, + 262, + 265, + 369 + ], + "type": "image", + "image_path": "a4777bdbafa993ada2bb5a5b5a1f855219cff6f6c8b8d872dc8b1be1d363964e.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 64, + 372, + 266, + 383 + ], + "lines": [ + { + "bbox": [ + 64, + 372, + 266, + 383 + ], + "spans": [ + { + "bbox": [ + 64, + 372, + 266, + 383 + ], + "type": "text", + "content": "Figure 5. Learned DAG over Intraoperative Indexes A." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 53, + 390, + 265, + 519 + ], + "blocks": [ + { + "bbox": [ + 53, + 390, + 265, + 519 + ], + "lines": [ + { + "bbox": [ + 53, + 390, + 265, + 519 + ], + "spans": [ + { + "bbox": [ + 53, + 390, + 265, + 519 + ], + "type": "image", + "image_path": "7d7d315d9ce95c101dc90f817294af7392c4bb7fad917562d9e90d5d6976aaa8.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 521, + 287, + 555 + ], + "lines": [ + { + "bbox": [ + 46, + 521, + 287, + 555 + ], + "spans": [ + { + "bbox": [ + 46, + 521, + 287, + 555 + ], + "type": "text", + "content": "Figure 6. Heat Maps via Grad-CAM [29] on 5 Patients. (a) shows original images and the bounding box of liver. (b) and (c) shows the area of interest " + }, + { + "bbox": [ + 46, + 521, + 287, + 555 + ], + "type": "inline_equation", + "content": "w/o" + }, + { + "bbox": [ + 46, + 521, + 287, + 555 + ], + "type": "text", + "content": " and with CaDAG, respectively." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 568, + 287, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 568, + 287, + 592 + ], + "spans": [ + { + "bbox": [ + 46, + 568, + 287, + 592 + ], + "type": "text", + "content": "surgery stage than to the final stage. This observation validates the effect of using intraoperative features." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 594, + 289, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 289, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 289, + 714 + ], + "type": "text", + "content": "General Framework for OS time prediction. In general, predicting OS time from preoperative images and indexes is medically important but challenging, as there exists a lot of missing information due to a long time interval between the preoperative stage to the final OS stage. To amend this problem, it is informative to leverage some intermediate information. In this paper, we show that with this leverage and the modeling of this information via causal discovery, our CAWIM model enjoys convincing medical interpretability, more concentrated location of liver region, more accurate" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 258, + 546, + 317 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 258, + 546, + 317 + ], + "spans": [ + { + "bbox": [ + 304, + 258, + 546, + 317 + ], + "type": "text", + "content": "estimation of intraoperative features, and finally better classification results on OS time. We thus believe that our framework can be beneficial to OS time prediction in other scenarios, e.g., other diseases or intermediate information in addition to intraoperative indexes." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 319, + 546, + 440 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 319, + 546, + 440 + ], + "spans": [ + { + "bbox": [ + 304, + 319, + 546, + 440 + ], + "type": "text", + "content": "Limitations. We find in Fig. 5 that although most of learned causal relations are consistent with medical priors, there exist relations that cannot be explained well: it is hard to determine the causal order for some pairs of variables: e.g., the causal relationship between the sum of tumor diameter and cell type. This may due to unobserved confounders between these pairs of variables, which may be alleviated by learning hidden representations that can explain the associations among these pairs of variables. We will leave this exploration in our future work." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 460, + 378, + 472 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 460, + 378, + 472 + ], + "spans": [ + { + "bbox": [ + 306, + 460, + 378, + 472 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 483, + 546, + 626 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 483, + 546, + 626 + ], + "spans": [ + { + "bbox": [ + 304, + 483, + 546, + 626 + ], + "type": "text", + "content": "We propose a novel OS time prediction paradigm, which is the first to leverage intraoperative attributes by causal structure learning. Our method significantly outperforms baselines by a large margin. We demonstrate that this improvement is contributed to highly interpretable learned causal structure, accurate estimation of intraoperative indexes, and identification of disease-related regions. We believe our method, especially the leverage of intraoperative information equipped with causal discovery can potentially benefit other scenarios. For limitation, we shall relax the SCM assumption by allowing the existence of unobserved confounders. We leave it as the future work." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 629, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 629, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 629, + 547, + 715 + ], + "type": "text", + "content": "Acknowledgment. Ethics Committee of Zhongshan Hospital, Fudan University approved the protocol of this study and waived the requirement for patient-informed consent (B2021-325R). The authors are also with Shanghai Key Lab of Intelligent Information Processing, and Fudan ISTBI-ZJNU Algorithm Centre for Brain-inspired Intelligence, Zhejiang Normal University, Jinhua, China." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15688" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 156 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 156 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 156 + ], + "type": "text", + "content": "[1] Shekoofeh Azizi, Basil Mustafa, Fiona Ryan, Zachary Beaver, Jan Freyberg, Jonathan Deaton, Aaron Loh, Alan Karthikesalingam, Simon Kornblith, Ting Chen, et al. Big self-supervised models advance medical image classification. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3478-3488, 2021. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 158, + 288, + 190 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 158, + 288, + 190 + ], + "spans": [ + { + "bbox": [ + 53, + 158, + 288, + 190 + ], + "type": "text", + "content": "[2] Pierre Bedossa and Thierry Poynard. An algorithm for the grading of activity in chronic hepatitis c. Hepatology, 24(2):289-293, 1996. 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 191, + 287, + 256 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 191, + 287, + 256 + ], + "spans": [ + { + "bbox": [ + 53, + 191, + 287, + 256 + ], + "type": "text", + "content": "[3] Nicholas A Christakis, Julia L Smith, Colin Murray Parkes, and Elizabeth B Lamont. Extent and determinants of error in doctors' prognoses in terminally ill patients: prospective cohort study commentary: Why do doctors overestimate? commentary: Prognoses should be based on proved indices not intuition. *Bmj*, 320(7233):469–473, 2000. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 258, + 287, + 301 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 258, + 287, + 301 + ], + "spans": [ + { + "bbox": [ + 53, + 258, + 287, + 301 + ], + "type": "text", + "content": "[4] Agostino Colli, Mirella Fraquelli, Marco Andreola, Barbara Marino, Enrico Zuccoli, and Dario Conte. Severe liver fibrosis or cirrhosis: accuracy of us for detection—analysis of 300 cases. Radiology, 227(1):89-94, 2003. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 302, + 287, + 334 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 302, + 287, + 334 + ], + "spans": [ + { + "bbox": [ + 53, + 302, + 287, + 334 + ], + "type": "text", + "content": "[5] Yin Dai, Yifan Gao, and Fayu Liu. Transmed: Transformers advance multi-modal medical image classification. Diagnostics, 11(8):1384, 2021. 2, 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 335, + 287, + 379 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 335, + 287, + 379 + ], + "spans": [ + { + "bbox": [ + 53, + 335, + 287, + 379 + ], + "type": "text", + "content": "[6] James J Driscoll and Oliver Rixe. Overall survival: still the gold standard: why overall survival remains the definitive end point in cancer clinical trials. The Cancer Journal, 15(5):401-405, 2009. 2, 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 380, + 287, + 434 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 380, + 287, + 434 + ], + "spans": [ + { + "bbox": [ + 53, + 380, + 287, + 434 + ], + "type": "text", + "content": "[7] J Faivre, D Forman, J Esteve, M Obradovic, M Sant, EURO-CARE Working Group, et al. Survival of patients with primary liver cancer, pancreatic cancer and biliary tract cancer in europe. European Journal of Cancer, 34(14):2184-2190, 1998. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 436, + 287, + 513 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 436, + 287, + 513 + ], + "spans": [ + { + "bbox": [ + 53, + 436, + 287, + 513 + ], + "type": "text", + "content": "[8] Ming Fan, Zuhui Liu, Sudan Xie, Maosheng Xu, Shiwei Wang, Xin Gao, and Lihua Li. Integration of dynamic contrast-enhanced magnetic resonance imaging and t2-weighted imaging radiomic features by a canonical correlation analysis-based feature fusion method to predict histological grade in ductal breast carcinoma. Physics in Medicine & Biology, 64(21):215001, 2019. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 514, + 287, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 514, + 287, + 578 + ], + "spans": [ + { + "bbox": [ + 53, + 514, + 287, + 578 + ], + "type": "text", + "content": "[9] Sheung Tat Fan, Chung Mau Lo, Ronnie TP Poon, Chun Yeung, Chi Leung Liu, Wai Key Yuen, Chi Ming Lam, Kelvin KC Ng, and See Ching Chan. Continuous improvement of survival outcomes of resection of hepatocellular carcinoma: a 20-year experience. Annals of surgery, 253(4):745-758, 2011. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 580, + 287, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 580, + 287, + 624 + ], + "spans": [ + { + "bbox": [ + 48, + 580, + 287, + 624 + ], + "type": "text", + "content": "[10] Xue Feng, Nicholas J Tustison, Sohil H Patel, and Craig H Meyer. Brain tumor segmentation using an ensemble of 3d u-nets and overall survival prediction using radiomic features. Frontiers in computational neuroscience, 14:25, 2020. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 624, + 287, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 624, + 287, + 668 + ], + "spans": [ + { + "bbox": [ + 48, + 624, + 287, + 668 + ], + "type": "text", + "content": "[11] Paul Glare, Kiran Virik, Mark Jones, Malcolm Hudson, Steffen Eychmuller, John Simes, and Nicholas Christakis. A systematic review of physicians' survival predictions in terminally ill cancer patients. *Bmj*, 327(7408):195, 2003. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 669, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 669, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 669, + 287, + 713 + ], + "type": "text", + "content": "[12] Arthur Gretton, Olivier Bousquet, Alex Smola, and Bernhard Schölkopf. Measuring statistical dependence with hilbert-schmidt norms. In International conference on algorithmic learning theory, pages 63-77. Springer, 2005. 3, 4" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "text", + "content": "[13] Arthur Gretton, Kenji Fukumizu, Choon Teo, Le Song, Bernhard Scholkopf, and Alex Smola. A kernel statistical test of independence. Advances in neural information processing systems, 20, 2007. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 118, + 545, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 545, + 173 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 545, + 173 + ], + "type": "text", + "content": "[14] Lu Guo, Ping Wang, Ranran Sun, Chengwen Yang, Ning Zhang, Yu Guo, and Yuanming Feng. A fuzzy feature fusion method for auto-segmentation of gliomas with multimodality diffusion and perfusion magnetic resonance images in radiotherapy. Scientific reports, 8(1):1-11, 2018. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 175, + 545, + 218 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 175, + 545, + 218 + ], + "spans": [ + { + "bbox": [ + 307, + 175, + 545, + 218 + ], + "type": "text", + "content": "[15] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 220, + 545, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 220, + 545, + 262 + ], + "spans": [ + { + "bbox": [ + 307, + 220, + 545, + 262 + ], + "type": "text", + "content": "[16] Mohammad Hesam Hesamian, Wenjing Jia, Xiangjian He, and Paul Kennedy. Deep learning techniques for medical image segmentation: achievements and challenges. Journal of digital imaging, 32(4):582-596, 2019. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 264, + 545, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 264, + 545, + 308 + ], + "spans": [ + { + "bbox": [ + 307, + 264, + 545, + 308 + ], + "type": "text", + "content": "[17] Karin W Houben and John L McCall. Liver transplantation for hepatocellular carcinoma in patients without underlying liver disease: a systematic review. *Liver Transplantation and Surgery*, 5(2):91–95, 1999. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 310, + 545, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 310, + 545, + 364 + ], + "spans": [ + { + "bbox": [ + 307, + 310, + 545, + 364 + ], + "type": "text", + "content": "[18] Biwei Huang, Kun Zhang, Jiji Zhang, Joseph D Ramsey, Ruben Sanchez-Romero, Clark Glymour, and Bernhard Scholkopf. Causal discovery from heterogeneous/nonstationary data. J. Mach. Learn. Res., 21(89):1-53, 2020. 2, 3, 4" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 366, + 545, + 399 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 366, + 545, + 399 + ], + "spans": [ + { + "bbox": [ + 307, + 366, + 545, + 399 + ], + "type": "text", + "content": "[19] Markus Kalisch and Peter Buhlman. Estimating high-dimensional directed acyclic graphs with the pc-algorithm. Journal of Machine Learning Research, 8(3), 2007. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 400, + 545, + 476 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 400, + 545, + 476 + ], + "spans": [ + { + "bbox": [ + 307, + 400, + 545, + 476 + ], + "type": "text", + "content": "[20] Christophe Laurent, Jean Frédéric Blanc, Steele Nobili, Antonio Sa Cunha, Brigitte le Bail, Paulette Bioulac-Sage, Charles Balabaud, Maylis Capdepont, and Jean Saric. Prognostic factors and longterm survival after hepatic resection for hepatocellular carcinoma originating from noncirrhotic liver. Journal of the American College of Surgeons, 201(5):656-662, 2005. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 479, + 545, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 479, + 545, + 533 + ], + "spans": [ + { + "bbox": [ + 307, + 479, + 545, + 533 + ], + "type": "text", + "content": "[21] Qing Li, Weidong Cai, Xiaogang Wang, Yun Zhou, David Dagan Feng, and Mei Chen. Medical image classification with convolutional neural network. In 2014 13th international conference on control automation robotics & vision (ICARCV), pages 844-848. IEEE, 2014. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 535, + 545, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 535, + 545, + 578 + ], + "spans": [ + { + "bbox": [ + 307, + 535, + 545, + 578 + ], + "type": "text", + "content": "[22] Tao Li, Wu Li, Yehui Yang, and Wensheng Zhang. Classification of brain disease in magnetic resonance images using two-stage local feature fusion. PloS one, 12(2):e0171749, 2017. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 580, + 545, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 580, + 545, + 643 + ], + "spans": [ + { + "bbox": [ + 307, + 580, + 545, + 643 + ], + "type": "text", + "content": "[23] Dong Nie, Han Zhang, Ehsan Adeli, Luyan Liu, and Dinggang Shen. 3d deep learning for multi-modal imaging-guided survival time prediction of brain tumor patients. In International conference on medical image computing and computer-assisted intervention, pages 212-220. Springer, 2016. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 647, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 647, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 647, + 545, + 713 + ], + "type": "text", + "content": "[24] Byeonggwan Noh, Young Mok Park, Yujin Kwon, Chang In Choi, Byung Kwan Choi, Yo-Han Park, Kwangho Yang, Sunju Lee, Taeyoung Ha, YunKyong Hyon, et al. Machine learning-based survival rate prediction of korean hepatocellular carcinoma patients using multi-center data. BMC gastroenterology, 22(1):1-9, 2022. 2" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "15689" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "text", + "content": "[25] Augustus Odena, Christopher Olah, and Jonathon Shlens. Conditional image synthesis with auxiliary classifier gans. In International conference on machine learning, pages 2642-2651. PMLR, 2017. 1" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 118, + 286, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 118, + 286, + 138 + ], + "spans": [ + { + "bbox": [ + 49, + 118, + 286, + 138 + ], + "type": "text", + "content": "[26] Judea Pearl. Causality. Cambridge university press, 2009. 2, 3, 4" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 140, + 286, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 140, + 286, + 161 + ], + "spans": [ + { + "bbox": [ + 49, + 140, + 286, + 161 + ], + "type": "text", + "content": "[27] Markus Peck-Radosavljevic. Drug therapy for advanced-stage liver cancer. Liver cancer, 3(2):125-131, 2014. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 163, + 286, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 163, + 286, + 215 + ], + "spans": [ + { + "bbox": [ + 49, + 163, + 286, + 215 + ], + "type": "text", + "content": "[28] Whitney B Pope, James Sayre, Alla Perlina, J Pablo Villablanca, Paul S Mischel, and Timothy F Cloughesy. Mr imaging correlates of survival in patients with high-grade gliomas. American Journal of Neuroradiology, 26(10):2466-2474, 2005. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 49, + 217, + 286, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 217, + 286, + 281 + ], + "spans": [ + { + "bbox": [ + 49, + 217, + 286, + 281 + ], + "type": "text", + "content": "[29] Ramprasaath R. Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra. Grad-cam: Visual explanations from deep networks via gradient-based localization. In 2017 IEEE International Conference on Computer Vision (ICCV), pages 618-626, 2017. 7, 8" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 49, + 284, + 286, + 336 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 284, + 286, + 336 + ], + "spans": [ + { + "bbox": [ + 49, + 284, + 286, + 336 + ], + "type": "text", + "content": "[30] Wei Shao, Tongxin Wang, Liang Sun, Tianhan Dong, Zhi Han, Zhi Huang, Jie Zhang, Daoqiang Zhang, and Kun Huang. Multi-task multi-modal learning for joint diagnosis and prognosis of human cancers. Medical Image Analysis, 65:101795, 2020. 2, 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 49, + 338, + 286, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 338, + 286, + 381 + ], + "spans": [ + { + "bbox": [ + 49, + 338, + 286, + 381 + ], + "type": "text", + "content": "[31] Pater Spirtes, Clark Glymour, Richard Scheines, Stuart Kauffman, Valerio Aimale, and Frank Wimberly. Constructing bayesian network models of gene expression networks from microarray data. 2000. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 49, + 384, + 286, + 425 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 384, + 286, + 425 + ], + "spans": [ + { + "bbox": [ + 49, + 384, + 286, + 425 + ], + "type": "text", + "content": "[32] Wen Tang, Haoyue Zhang, Pengxin Yu, Han Kang, and Rongguo Zhang. Mmmna-net for overall survival time prediction of brain tumor patients. arXiv preprint arXiv:2206.06267, 2022. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 49, + 427, + 286, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 427, + 286, + 470 + ], + "spans": [ + { + "bbox": [ + 49, + 427, + 286, + 470 + ], + "type": "text", + "content": "[33] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 49, + 472, + 286, + 546 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 472, + 286, + 546 + ], + "spans": [ + { + "bbox": [ + 49, + 472, + 286, + 546 + ], + "type": "text", + "content": "[34] Jing-Houng Wang, Chi-Sin Changchien, Tsung-Hui Hu, Chuan-Mo Lee, Kwong-Ming Kee, Chih-Yun Lin, Chao-Long Chen, Tai-Yi Chen, Yu-Jie Huang, and Sheng-Nan Lu. The efficacy of treatment schedules according to Barcelona clinic liver cancer staging for hepatocellular carcinoma-survival analysis of 3892 patients. European journal of cancer, 44(7):1000-1006, 2008. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 49, + 548, + 286, + 581 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 548, + 286, + 581 + ], + "spans": [ + { + "bbox": [ + 49, + 548, + 286, + 581 + ], + "type": "text", + "content": "[35] Samir S Yadav and Shivajirao M Jadhav. Deep convolutional neural network based medical image classification for disease diagnosis. Journal of Big Data, 6(1):1-18, 2019. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 49, + 582, + 286, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 582, + 286, + 645 + ], + "spans": [ + { + "bbox": [ + 49, + 582, + 286, + 645 + ], + "type": "text", + "content": "[36] Naoki Yamanaka, Eizo Okamoto, Tsuyosi Oriyama, Jiro Fujimoto, Kazutaka Furukawa, Eisuke Kawamura, Tsuneo Tanaka, and Fumito Tomoda. A prediction scoring system to select the surgical treatment of liver cancer. further refinement based on 10 years of use. Annals of surgery, 219(4):342, 1994. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 49, + 647, + 286, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 647, + 286, + 679 + ], + "spans": [ + { + "bbox": [ + 49, + 647, + 286, + 679 + ], + "type": "text", + "content": "[37] Jianpeng Zhang, Yutong Xie, Qi Wu, and Yong Xia. Medical image classification using synergic deep learning. Medical image analysis, 54:10-19, 2019. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 49, + 681, + 286, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 681, + 286, + 712 + ], + "spans": [ + { + "bbox": [ + 49, + 681, + 286, + 712 + ], + "type": "text", + "content": "[38] Fan Zhou, Tengfei Li, Heng Li, and Hongtu Zhu. Tpcnn: two-phase patch-based convolutional neural network for automatic brain tumor segmentation and survival prediction." + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 545, + 140 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "type": "text", + "content": "In International MICCAI Brainlesion Workshop, pages 274-286. Springer, 2017. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 96, + 545, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 96, + 545, + 140 + ], + "spans": [ + { + "bbox": [ + 308, + 96, + 545, + 140 + ], + "type": "text", + "content": "[39] Tao Zhou, H. Fu, Yu Zhang, Changqing Zhang, Xiankai Lu, Jianbing Shen, and Ling Shao. M2net: Multi-modal multichannel network for overall survival time prediction of brain tumor patients. ArXiv, abs/2006.10135, 2020. 2, 3" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "15690" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/CelebV-Text_ A Large-Scale Facial Text-Video Dataset/b3bc8119-1bc3-419d-89b1-046a6b2d92d9_content_list.json b/2023/CelebV-Text_ A Large-Scale Facial Text-Video Dataset/b3bc8119-1bc3-419d-89b1-046a6b2d92d9_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..6a2997ae6be426209b12da99e3e9215cbf35ebe3 --- /dev/null +++ b/2023/CelebV-Text_ A Large-Scale Facial Text-Video Dataset/b3bc8119-1bc3-419d-89b1-046a6b2d92d9_content_list.json @@ -0,0 +1,1566 @@ +[ + { + "type": "text", + "text": "CelebV-Text: A Large-Scale Facial Text-Video Dataset", + "text_level": 1, + "bbox": [ + 209, + 114, + 759, + 135 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Jianhui $\\mathrm{Yu}^{1*}$ Hao Zhu $^{2*}$ Liming Jiang $^{3}$ Chen Change Loy $^{3}$ Weidong Cai $^{1}$ Wayne Wu $^{4}$", + "bbox": [ + 96, + 148, + 869, + 167 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ University of Sydney $^{2}$ SenseTime Research $^{3}$ S-Lab, Nanyang Technological University $^{4}$ Shanghai AI Laboratory", + "bbox": [ + 124, + 167, + 844, + 184 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "jianhui.yu@sydney.edu.au haozhu96@gmail.com {liming002,ccloy}@ntu.edu.sg", + "bbox": [ + 148, + 185, + 812, + 199 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "tom.cai@sydney.edu.au wuwenyan0503@gmail.com", + "bbox": [ + 279, + 200, + 684, + 215 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/f0109e28e042bfb7ecc4716cf4e58dae3e14efdf34943c697065303fe469de7e.jpg", + "image_caption": [ + "(a) Video Samples (left: general appearance, middle: detailed appearance, right: light conditions)", + "Figure 1. Overview of CelebV-Text. CelebV-Text contains (a) 70,000 video samples and (b) 1,400,000 text descriptions. Each video sample is annotated with general appearance, detailed appearance, light conditions, action, emotion, and light directions." + ], + "image_footnote": [], + "bbox": [ + 80, + 241, + 638, + 494 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/fcf8b3338799a55aa4546a38e787d33d2e42b686dbba211adb062ffd7cec9900.jpg", + "image_caption": [ + "(b) Text Descriptions" + ], + "image_footnote": [], + "bbox": [ + 643, + 239, + 885, + 494 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 549, + 312, + 564 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Text-driven generation models are flourishing in video generation and editing. However, face-centric text-to-video generation remains a challenge due to the lack of a suitable dataset containing high-quality videos and highly relevant texts. This paper presents CelebV-Text, a large-scale, diverse, and high-quality dataset of facial text-video pairs, to facilitate research on facial text-to-video generation tasks. CelebV-Text comprises 70,000 in-the-wild face video clips with diverse visual content, each paired with 20 texts generated using the proposed semi-automatic text generation strategy. The provided texts are of high quality, describing both static and dynamic attributes precisely. The superiority of CelebV-Text over other datasets is demonstrated via comprehensive statistical analysis of the videos, texts, and text-video relevance. The effectiveness and potential of CelebV-Text are further shown through extensive self-evaluation. A benchmark is constructed with representative methods to standardize the evaluation of the facial text-to-video generation task. All data and models are publicly available1.", + "bbox": [ + 75, + 575, + 470, + 853 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 501, + 549, + 630, + 564 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Text-driven video generation has recently garnered significant attention in the fields of computer vision and computer graphics. By using text as input, video content can be generated and controlled, inspiring numerous applications in both academia and industry [5,34,43,47]. However, text-to-video generation still faces many challenges, particularly in the face-centric scenario where generated video frames often lack quality [18, 34, 37] or have weak relevance to input texts [2,4,39,67]. We believe that one of the main issues is the absence of a well-suited facial text-video dataset containing high-quality video samples and text descriptions of various attributes highly relevant to videos.", + "bbox": [ + 496, + 575, + 890, + 743 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Constructing a high-quality facial text-video dataset poses several challenges, mainly in three aspects. 1) Data collection. The quality and quantity of video samples largely determine the quality of generated videos [11, 45, 48, 60]. However, obtaining such a large-scale dataset with high-quality samples while maintaining a natural distribution and smooth video motion is challenging. 2) Data annotation. The relevance of text-video pairs needs to be ensured. This requires a comprehensive coverage of text for describing the content and motion appearing in the video, such as light conditions and head movements. 3) Text gen", + "bbox": [ + 496, + 746, + 892, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Equal contribution.", + "bbox": [ + 94, + 875, + 205, + 887 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "1Project page: https://celebv-text.github.io", + "bbox": [ + 96, + 887, + 405, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "14805", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "eration. Producing diverse and natural texts are non-trivial. Manual text generation is expensive and not scalable. While auto-text generation is easily extensible, it is limited in naturalness.", + "bbox": [ + 76, + 90, + 467, + 145 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To overcome the challenges mentioned above, we carefully design a comprehensive data construction pipeline that includes data collection and processing, data annotation, and semi-auto text generation. First, to obtain raw videos, we follow the data collection steps of CelebV-HQ, which has proven to be effective in [66]. We introduce a minor modification to the video processing step to improve the video's smoothness further. Next, to ensure highly relevant text-video pairs, we analyze videos from both temporal dynamics and static content and establish a set of attributes that may or may not change over time. Finally, we propose a semi-auto template-based method to generate texts that are diverse and natural. Our approach leverages the advantages of both auto- and manual-text methods. Specifically, we design a rich variety of grammar templates as [10,52] to parse annotation and manual texts, which are flexibly combined and modified to achieve high diversity, complexity, and naturalness.", + "bbox": [ + 75, + 147, + 467, + 396 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "With the proposed pipeline, we create CelebV-Text, a Large-Scale Facial Text-Video Dataset, which includes 70,000 in-the-wild video clips with a resolution of at least $512 \\times 512$ and 1,400,000 text descriptions with 20 for each clip. As depicted in Figure 1, CelebV-Text consists of high-quality video samples and text descriptions for realistic face video generation. Each video is annotated with three types of static attributes (40 general appearances, 5 detailed appearances, and 6 light conditions) and three types of dynamic attributes (37 actions, 8 emotions, and 6 light directions). All dynamic attributes are densely annotated with start and end timestamps, while manual-texts are provided for labels that cannot be discretized. Furthermore, we have designed three templates for each attribute type, resulting in a total of 18 templates that can be flexibly combined. All attributes and manual-texts are naturally described in our generated texts.", + "bbox": [ + 75, + 398, + 467, + 633 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "CelebV-Text surpasses existing face video datasets [11] in terms of resolution (over 2 times higher), number of samples, and more diverse distribution. In addition, the texts in CelebV-Text exhibit higher diversity, complexity, and naturalness than those in text-video datasets [19, 66]. CelebV-Text also shows high relevance of text-video pairs, validated by our text-video retrieval experiments [17]. To further examine the effectiveness and potential of CelebV-Text, we evaluate it on a representative baseline [19] for facial text-to-video generation. Our results show better relevance between generated face videos and texts when compared to a state-of-the-art large-scale pretrained model [26]. Furthermore, we show that a simple modification of [19] with text interpolation can significantly improve temporal coherence. Finally, we present a new benchmark for text-to-video generation to standardize the facial text-to-video generation task, which includes representative models [5, 19] on three text-video datasets.", + "bbox": [ + 75, + 635, + 467, + 883 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The main contributions of this work are summarized as", + "bbox": [ + 96, + 885, + 467, + 898 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "follows: 1) We propose CelebV-Text, the first large-scale facial text-video dataset with high-quality videos, as well as rich and highly-relevant texts, to facilitate research in facial text-to-video generation. 2) Comprehensive statistical analyses are conducted to examine video/text quality and diversity, as well as text-video relevance, demonstrating the superiority of CelebV-Text. 3) A series of self-evaluations are performed to demonstrate the effectiveness and potential of CelebV-Text. 4) A new benchmark for text-to-video generation is constructed to promote the standardization of the facial text-to-video generation task.", + "bbox": [ + 496, + 90, + 890, + 244 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 260, + 640, + 276 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Text-to-Video Generation. Text-driven video generation, which involves generating videos from text descriptions, has recently gained significant interest as a challenging task. Mittal et al. [43] first introduced this task to generate semantically consistent videos conditioned on encoded captions. Other studies, such as [5, 15, 46], attempt to generate video samples conditioned on encoded text inputs. However, due to the low richness of text descriptions and the small number of data samples, the generated video samples are often at low resolution or lack relevance with the input texts. More recently, several works [19, 26, 27, 55, 57–59] have employed discrete latent codes [16, 54] for more realistic video generation. Some of these works treat videos as a sequence of independent images [19, 26, 58, 59], while Phenaki [55] considers temporal relations between each frame for a more robust video decoding process. Another branch of studies leverage diffusion models for text-to-video generation [20, 24, 25, 51], which require millions or billions of samples to achieve high-quality generation. While text-to-video generation methods are rapidly evolving, they are generally designed for generating general videos. Among these methods, only MMVID [19] has conducted specific experiments with face-centric descriptions. One possible reason for this is that facial text-to-video generation requires more accurate and detailed text descriptions than general tasks. However, there is currently no suitable dataset available that provides such properties for face-centric text-to-video generation.", + "bbox": [ + 496, + 286, + 890, + 675 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Multimodal Datasets. Existing multimodal datasets can be categorized into two classes: open-world and closed-world. Open-world datasets [3,9,13,32,35,36,40,43,49,50,62,65] are widely used for text-to-image/video generation tasks. Some of them have manual annotations [13, 32, 35, 50, 62] and part of them are directly collected from the Internet, such as subtitles [40, 49]. Closed-world datasets are mostly composed of images or videos collected in constrained environment with corresponding information such as text. CLEVR [28] is a synthetic text-image dataset produced by arranging 3D objects with different shapes under a controlled background. While MUGEN [21] is a video-audiotext dataset that was collected using CoinRun [12] by introducing audio and new interactions. The corresponding text is produced by human annotators and grammar templates.", + "bbox": [ + 496, + 676, + 890, + 883 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Multimodal face datasets also exist. Modified MUG [1]", + "bbox": [ + 517, + 885, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "14806", + "bbox": [ + 480, + 945, + 519, + 955 + ], + "page_idx": 1 + }, + { + "type": "table", + "img_path": "images/97120df4bda02a1a091b792dc9856012e28a2c3c55b51fe4e5b3b1aa21522161.jpg", + "table_caption": [ + "Table 1. In-the-wild face video dataset comparison. The symbol “#” indicates the number. The abbreviations “Res.”, “Dura.”, “App.”, “Cond.”, “Act.”, “Emo.”, and “Dir.” stand for Resolution, Duration, Appearance, Condition, Action, Emotion, and Direction, respectively. The “half checkmark” denotes that CelebV-HQ consists of action attributes with no timestamp." + ], + "table_footnote": [], + "table_body": "
DatasetsMeta InformationAttribute LabelsText
StaticDynamic
#SamplesRes.Dura.General App.Detail App.Light Cond.Act.Emo.Light Dir.Auto Manual
CelebV [60]5256×2562hrs
VoxCeleb2 [11]150,480224×2242442hrs
CelebV-HQ [66]35,666512×51268hrs✓XX✓✓XX
MM-Vox [19]19,522224×224323hrs✓XXXXX✓
CelebV-Text70,000512×512+279hrs✓✓✓✓✓✓✓
", + "bbox": [ + 99, + 121, + 867, + 262 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "is a closed-world text-video dataset that contains 1,039 videos with subjects showing different emotions, where the text descriptions are generated from facial emotions using a fixed template [30]. MM-Vox [19] contains 19,522 face videos from VoxCeleb [45], with 36 facial attributes manually labeled following CelebA [38] and text descriptions generated via Probabilistic Context-Free Grammar (PCFG) [61]. However, both datasets only contain language descriptions related to static facial attributes without considering the temporal state change (i.e., emotion or action) presented in the original face videos. Moreover, the limited label annotations restrict the diversity of the text descriptions, making them sub-optimal for studying the text-to-video generation task on the face domain. CelebV-HQ [66] is the latest high-quality face video dataset that covers facial annotations, including appearance, movement, and emotion. However, it only provides discrete labels and timestamps, with no text descriptions.", + "bbox": [ + 75, + 279, + 472, + 531 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Dataset Construction", + "text_level": 1, + "bbox": [ + 76, + 546, + 279, + 561 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this work, we aim to build a facial text-video dataset, which requires not only large-scale video samples of high quality, but also natural and diverse text descriptions that are highly relevant to videos. To achieve so, we propose an efficient pipeline, as shown in Figure 2, to construct CelebV-Text, including Data Collection & Processing, Data Annotation, and Semi-auto Text Generation.", + "bbox": [ + 75, + 571, + 468, + 670 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Data Collection & Processing", + "text_level": 1, + "bbox": [ + 76, + 680, + 339, + 698 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Collection. We follow the same strategy as CelebV-HQ [66] due to its effectiveness in large-scale high-quality data collection. Specifically, we firstly generate a large number of queries, including human names, movie titles, vlogs and so on, to retrieve videos that contain human faces with temporally dynamic state changes and abundant facial attributes. Our data are collected from open world with videos downloaded from online resources. Videos with low resolution $(< 512^{2})$ , low time duration $(< 5s)$ , and having appeared in CelebV-HQ are filtered out.", + "bbox": [ + 75, + 704, + 468, + 843 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Processing. To sample high-quality and diverse video clips from our raw collections, similar steps are followed as CelebV-HQ [66] with modifications. We first filter out video clips with bounding box regions less than $512^{2}$ rather", + "bbox": [ + 75, + 844, + 470, + 902 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/eedc5f3f0b2e60ba783324e68f63307e42cfcdf071714f39f4be781746fcf053.jpg", + "image_caption": [ + "Figure 2. Pipeline of our dataset construction process. The pipeline includes data collection & processing, data annotation, and semi-auto text generation." + ], + "image_footnote": [], + "bbox": [ + 506, + 277, + 890, + 468 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "than resize them to the same resolution. In this way, clips are not upsampled or downsampled hence the video quality would not be affected, which leads to various resolutions of collected videos: $56.4\\%$ with $512^2 \\sim 1024^2$ , and $43.6\\%$ for $1024^2+$ . To reduce the face area noise when the background changes, we further change the video splitting strategy. In addition to our focus on the same human motion [6] and identity [14] present in adjacent frames, we split the video into different clips when the background changes by a toolkit1.", + "bbox": [ + 498, + 531, + 892, + 672 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Data Annotation", + "text_level": 1, + "bbox": [ + 500, + 681, + 665, + 698 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The annotation process is a core part in CelebV-Text construction, which would greatly affect the relevance of text-video pairs, as our designed text templates heavily depend on the annotation results. Here, we first describe how we design attributes, and then give details about the annotation strategy for face videos.", + "bbox": [ + 498, + 705, + 890, + 789 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Attributes Design. Temporal dynamic is the key difference between images and videos. However, as shown in Table 1, most face video datasets focus on static attributes where attribute information does not change over time, such as appearance. Dynamic attributes that change over time, such as emotion and face actions, are often neglected. In the fol", + "bbox": [ + 498, + 789, + 890, + 875 + ], + "page_idx": 2 + }, + { + "type": "page_footnote", + "text": "1 https://github.com/Breakthrough/PySceneDetect", + "bbox": [ + 517, + 886, + 782, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "14807", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/a913896e83a99c179f17204cbfab4c4efac3155f5fef60f086907c9c4feecf3c.jpg", + "image_caption": [ + "Figure 3. Dataset distribution comparison. The distributions of appearance attributes, action attributes, and light directions." + ], + "image_footnote": [], + "bbox": [ + 122, + 74, + 380, + 244 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/28c5acb5920de6e0ddd8f704943dcbbe09d28a7eced1a6c54706017cf64a6290.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 379, + 74, + 635, + 244 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/744e07655ba3d485ce2a50ff5b2975ac45985578192d20caa4c48711963ed744.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 635, + 74, + 834, + 244 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "lowing, we decouple face videos into static and dynamic categories and details are given as follows.", + "bbox": [ + 75, + 270, + 468, + 297 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "1) Static. The current dataset [19] only considers static information such as the appearance attribute, which includes 40 classes as CelebA [38]. In contrast, we define static information to include three types of attributes: general appearance, detailed appearance, and light conditions. General appearance attributes follow the same definition as CelebA [38]. Detailed appearance attributes including five classes are proposed for realistic face generation, i.e., scar, mole, freckle, dimple, and one-eyed. We define light conditions in a restricted manner to include light color temperature [22] and brightness [7], with a total of 6 classes.", + "bbox": [ + 75, + 297, + 468, + 450 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2) Dynamic. Here, we design three dynamic attributes, i.e., action, emotion, and light directions. For action attributes, we follow CelebV-HQ [66] and expand their action list by two classes, i.e., squint and blink. For emotion attributes, we select the 8 emotion setting in Affectnet [44], including neutral, anger, contempt, disgust, fear, happiness, sadness, and surprise. For light direction attributes, we derive and modify classes from [29] and give 6 light direction classes. Complete lists are given in the Appendix. Moreover, as shown in Table 1, CelebV-HQ [66] is the only dataset giving timestamps of dynamic attributes. Following their idea, we densely annotate all dynamic attributes of CelebV-Text with the start and end time.", + "bbox": [ + 75, + 452, + 468, + 632 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Automatic and Manual Annotation. Based on our attributes design, we find that some attributes can be annotated automatically (e.g., appearance) while some need manual annotations (e.g., timestamps of dynamic attributes). Considering the dataset quality and cost of expense, our annotation strategy includes both automatic and manual annotations.", + "bbox": [ + 75, + 633, + 468, + 729 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For automatic annotation, we first investigate algorithms and select designed attributes that can be automatically annotated. We then test different algorithms on our dataset and keep those giving annotation accuracy of $85\\%$ or higher. This process yields all light condition labels, all appearance labels, and all emotion labels suitable for automatic annotation. Algorithms for different labels we finally chose are reported in the Appendix. Automatic annotation results can be further revised by human workers to improve accuracy in a less costly way.", + "bbox": [ + 75, + 732, + 468, + 871 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For manual annotation, we hire and train human workers following [66] to annotate attributes that are filtered out by", + "bbox": [ + 76, + 872, + 468, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "an automatic annotation process. In this case, we manually annotate dynamic attributes, i.e., action and light directions, to give both class labels and exact timestamps. In addition, it is hard to represent detailed appearance attributes by the discrete label, e.g., the characteristics of scars or moles. We therefore ask annotators to give a natural description for each attribute, describing exact positions relative to face parts. These designs greatly enhance the relevance between the final text and the video.", + "bbox": [ + 496, + 270, + 890, + 395 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Semi-auto Text Generation", + "text_level": 1, + "bbox": [ + 498, + 404, + 741, + 419 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Multimodal text-video datasets collect texts via three common methods: subtitles [4, 40, 67], manual-text generation [3, 9, 32, 56, 62], and auto-text generation [5, 21, 27]. However, it is difficult for the individual method to generate texts with high relevance to videos, natural expression, and high diversity. Specifically, although subtitles are easy to obtain, they can pose weakly relevant text-video pairs and introduce noise, making the dataset quality hard to control. Moreover, manual-text generation method is time and cost consuming, as natural language descriptions are required for each video. In this case, increasing the data scale is quite hard as more workers are needed to describe new videos, which does not meet the efficiency and scalability of annotation. Finally, auto-text generation is flexible and scalable, as abundant texts can be simultaneously generated given annotation results of collected videos. However, the diversity, complexity, and naturalness of generated texts can be impacted by the designed grammar templates.", + "bbox": [ + 496, + 426, + 890, + 678 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To this end, we propose a semi-auto template-based text generation strategy that combines both manual-text and auto-text generation methods. Specifically, as mentioned in Section 3.2, manual-texts are required to describe detailed appearance attributes. Annotated attribute information is fed into our designed template for auto-text generation.", + "bbox": [ + 496, + 678, + 890, + 761 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To make our template as natural as possible, we first ask each annotator to describe 10 different face videos for each attribute. We then analyze the grammar structure (i.e., parse tree banks) along with online corpora following [10, 31], and find the most three common grammar structures for each attribute. Finally, we utilize probabilistic context-free grammar [52, 61] and modify the grammar structures to design our own templates. Texts are generated based on templates with synonym replacement using NLTK [8] to increase our generation diversity. Details of our template de", + "bbox": [ + 496, + 761, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "14808", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/c5c729fb58a120a7c4b9b09e97194b28978b2f7c66e2e7beb3395396282b25e8.jpg", + "image_caption": [ + "(a) Image quality distribution" + ], + "image_footnote": [], + "bbox": [ + 106, + 82, + 411, + 172 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/8ae208a61353f51058e45f49014aa986ea1c2b064485a4f62d0a05ee160c2b33.jpg", + "image_caption": [ + "(b) Video quality distribution", + "Figure 4. Dataset quality distribution. The metrics used are BRISQUE [42] and VSFA [33] respectively." + ], + "image_footnote": [], + "bbox": [ + 106, + 195, + 405, + 277 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "signs are in the Appendix.", + "bbox": [ + 76, + 328, + 251, + 342 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Statistical Analysis of CelebV-Text", + "text_level": 1, + "bbox": [ + 76, + 354, + 388, + 372 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this section, we compare CelebV-Text with the two most relevant and representative face video datasets [19,66]. We perform a comprehensive analysis of CelebV-Text in terms of video, text, and text-video relevance. To verify the effectiveness of our designed grammar templates, we generate text descriptions for CelebV-HQ based on its attributes for comparison. For simplicity, we use \"CelebV-HQ\" to denote this variant in the following.", + "bbox": [ + 75, + 380, + 468, + 491 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Video Comparisons", + "text_level": 1, + "bbox": [ + 76, + 501, + 264, + 517 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We briefly compare the overall statistics of existing face video datasets [11, 19, 60, 66] in Table 1. As reported, CelebV-Text contains 70,000 video clips with a total duration of around 279 hours. Each video is accompanied by 20 sentences describing all 6 designed attributes. Compared to CelebV [60], CelebV-Text has a larger scale and higher resolution. Although VoxCeleb2 [11] has more samples than CelebV-Text, its video distribution is limited as most videos are mainly talking faces. Moreover, video samples of both CelebV-HQ [66] and CelebV-Text are collected in open-world with diverse queries so that they are rich in distribution, while CelebV-Text has about 2 times video data, more video attributes, and highly relevant text descriptions. Finally, compared to the only existing facial text-video dataset MM-Vox [19], CelebV-Text overpasses MM-Vox in terms of scale and quality.", + "bbox": [ + 75, + 523, + 468, + 746 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Attributes Distribution. In order to better present the distribution of different attributes in CelebV-Text, we pick and divide general appearance, action, and light direction attributes into groups. More distributions and division designs are provided in the Appendix. Specifically, all 40 general appearance classes are divided into 5 groups shown in Figure 3 (a). Facial features (e.g., double chin, big nose, and oval face) account for the most portion around $45\\%$ . The elementary group is twice large than the beard type, accounting for around $25\\%$ and $12\\%$ , respectively. Fewer samples are located to the hairstyle and accessories groups,", + "bbox": [ + 75, + 746, + 468, + 900 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/8f1c27f9666ba9bebd95f0a8a9d2ec53224c929a45807271dd4d9d8d189191b1.jpg", + "image_caption": [ + "(a) Number of word distribution" + ], + "image_footnote": [], + "bbox": [ + 516, + 84, + 692, + 180 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/0fcc4b2effa71e2482d61d55b7b5b1005caf599f0c78b81d51fb0e376a4f946f.jpg", + "image_caption": [ + "(b) 4-Grams distribution", + "MM-Vox" + ], + "image_footnote": [], + "bbox": [ + 697, + 85, + 875, + 180 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/a84628f0117593e033e8be4469d122b0db503a029981c50b4f5c21a92a69dffa.jpg", + "image_caption": [], + "image_footnote": [ + "This man with grey hair has a big nose and bags under eyes. He is wearing eyeglasses and earrings, with freckles above the eye areas. The man is firstly gazing for a short time and then he wags his head for a short time, and finally he keeps gazing for the rest of the time. He firstly remains neutral for a long time and then he turns angry for the rest of the time. The video is slightly dark in warm light. The light direction is front lighting for the whole time. CelebV-Text" + ], + "bbox": [ + 526, + 194, + 586, + 241 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/c9689ecab2ebe6ca19b83c6809ae2af7c720137dc660b41a353b2d6903cfadb8.jpg", + "image_caption": [ + "(c) Video and text examples", + "Figure 5. Text distribution. CelebV-Text achieves better performance in both 4-gram and number words distribution." + ], + "image_footnote": [ + "a man has bags under eyes , receding hairline and straight hairhe is wearing goatee . he is chubby . he has beard , archedeyebrows and black hair ." + ], + "bbox": [ + 527, + 247, + 589, + 292 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "taking around $10\\%$ and $8\\%$ , respectively. Besides, action attributes are divided into 5 groups in Figure 3 (b), where it is clear that head-related actions account for the largest portion of around $60\\%$ , followed by eyes-related actions of around $20\\%$ . The interaction group (e.g., eat), feeling group (e.g., smile), and daily group (e.g., sleep) account for around $9\\%$ , $7\\%$ , and $4\\%$ , respectively. Finally, for light directions (Figure 3 (c)), most samples contain the front lighting and the remaining ones are evenly distributed.", + "bbox": [ + 496, + 353, + 890, + 479 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Video Quality Distribution. We follow [66] to analyze the quality of our collected videos. To demonstrate the superiority of CelebV-Text, we compare with MM-Vox [19] and CelebV-HQ [66], where mean BRISQUE [42] and VSFA [33] are used to evaluate the image and video quality, respectively. Image quality of all datasets is shown in Figure 4 (a), where CelebV-Text and CelebV-HQ achieve comparable quality, higher than MM-Vox by a large margin. Video quality of all datasets is shown in Figure 4 (b), where CelebV-Text has the best quality, which is due to the effect of the video split method mentioned in Section 3.1, alleviating the discontinuity during background transitions.", + "bbox": [ + 496, + 479, + 890, + 647 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Text Comparisons", + "text_level": 1, + "bbox": [ + 500, + 655, + 674, + 671 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In addition to a large number of video samples, text descriptions of CelebV-Text are longer and more detailed than those in MM-Vox [19] and CelebV-HQ [66] (see Figure 5 (a)), where the average text length of MM-Vox, CelebV-HQ, and CelebV-Text are 28.39, 31.06, and 67.15. Distributions of Celeb-HQ and MM-Vox are close, but there are more words in CelebV-Text to describe a video due to the comprehensive annotation.", + "bbox": [ + 496, + 676, + 890, + 787 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To validate the linguistic diversity of the generated texts, comparisons are conducted among the three datasets following [56]. Specifically, we report the unique part-of-speech (POS) tags (i.e., verb, noun, adjective, and adverb) of the three datasets in Table 3. Obviously, due to our comprehensively designed attribute list and the number of templates, CelebV-Text presents a wider variety of text styles, covering a broader range of face attributes that are static or", + "bbox": [ + 496, + 787, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "14809", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/e053a57d299a3b0c6dfaa09f338e98fa83362fcd2c5eb7070d247701db3515de.jpg", + "table_caption": [ + "Table 2. Multimodal retrieval results. Clip2Video [17] is leveraged to measure the text-video relevance via retrieval experiments. Bold values indicate the best results, underlined ones indicate the second best." + ], + "table_footnote": [], + "table_body": "
Text ⇒ VideoVideo ⇒ Text
DescriptionDatasetR@1(↑)R@5(↑)R@10(↑)MdR(↓)MnR(↓)R@1(↑)R@5(↑)R@10(↑)MdR(↓)MnR(↓)
(a) App.MM-Vox [19]1.59.015.752.068.82.09.214.643.057.8
CelebV-HQ [66]5.919.229.727.052.27.220.732.427.046.9
CelebV-Text6.121.335.526.349.17.420.729.926.648.3
(b) App.+Emo.CelebV-HQ [66]6.520.130.825.048.07.925.538.817.037.0
CelebV-Text6.623.437.126.047.68.127.234.718.238.3
(c) App.+Emo.+Act.CelebV-Text6.924.139.225.846.78.027.637.116.736.1
", + "bbox": [ + 80, + 119, + 893, + 236 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/39f3b3b671181fbce26fb6f2f7293d5e439754ca18dad09257ba3b3c52aa18bd.jpg", + "table_caption": [ + "Table 3. Number of unique POS tags. The numbers of unique POS tags for MM-Vox, CelebV-HQ, and CelebV-Text." + ], + "table_footnote": [], + "table_body": "
Dataset#Verb#Adj.#Noun#Adv.
MM-Vox [19]520380
CelebV-HQ [66]1024506
CelebV-Text967817424
", + "bbox": [ + 112, + 277, + 437, + 342 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "dynamic in the temporal domain.", + "bbox": [ + 76, + 358, + 299, + 373 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In addition, we further examine the naturalness and complexity of our texts compared to MM-Vox, where we modify [63] to calculate the type-token vocabulary curve for all captions. As shown in Figure 5 (b) where unique 4-grams are selected as the types [56], it is evident that due to our grammar structures and synonym replacement, the linguistic naturalness (vocabulary use) and complexity (vocabulary size) of our CelebV-Text are much better. Please refer to Appendix for more $n$ -grams results.", + "bbox": [ + 75, + 378, + 468, + 505 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Text-Video Relevance", + "text_level": 1, + "bbox": [ + 76, + 531, + 279, + 545 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To quantitatively validate our text-video relevance, we conduct text-video retrieval tasks on three datasets: MMVox [19], CelebV-HQ [66], and CelebV-Text. Rather than use conventional frame-wise clip score as most works [24, 51, 55], we follow [17] to compute feature similarities between texts and videos with the consideration of temporal dynamics, which reflects accurate multimodal interactions across the two modalities. Recall at rank K (R@K), median rank (MdR), and mean rank (MnR) [17, 41, 64] are used as evaluation metrics, where the higher R@K, the lower median rank and mean rank indicate better performance.", + "bbox": [ + 75, + 559, + 468, + 715 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We first examine the performance given texts with descriptions of general appearance in Table 2 (a). Results of CelebV-HQ and CelebV-Text are both better than MM-Vox for two retrieval tasks, which indicates our designed templates can produce texts more relevant to videos than MM-Vox. We further add descriptions about dynamic emotion changes to CelebV-HQ and CelebV-Text in Table 2 (b). Similar results are achieved in both datasets, which reflects that our annotation accuracy on static appearance attributes is as good as CelebV-HQ. Finally, we append action descriptions to CelebV-Text in Table 2 (c), which achieves the best performance on most metrics, verifying the relevance between our generated texts and video samples.", + "bbox": [ + 75, + 719, + 470, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Experiment", + "text_level": 1, + "bbox": [ + 498, + 250, + 625, + 267 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this section, we first conduct facial text-to-video generation to validate the effectiveness of our CelebV-Text dataset. We then benchmark representative approaches on facial text-to-video generation task.", + "bbox": [ + 496, + 273, + 893, + 330 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1. High-relevance Text-to-Video Generation", + "text_level": 1, + "bbox": [ + 498, + 335, + 852, + 353 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To show the benefits brought by our text descriptions which depict both static and dynamic attributes, we conduct experiments to show the effectiveness of CelebV-Text. Experiments are mainly based on a recent open-sourced state-of-the-art method, MMVID [19], and compared with $\\mathrm{CogVideo}^2$ [26], which is a large-scale pretrained text-to-video model, trained on millions of text-image/video pairs. Static Face Video Generation. To validate the effectiveness of our facial text-video dataset in static attributes, we use the models stated above to generate videos conditioned on general appearance, face details, and light conditions descriptions, respectively. Specifically, we first train MMVID [19] from scratch solely on CelebV-Text. We then generate 3 input texts including individual descriptions of each of the static attributes. Generated texts are fed into both MMVID [19] and $\\mathrm{CogVideo}$ [26] and corresponding video outputs are examined.", + "bbox": [ + 496, + 359, + 890, + 595 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Visualization results of general appearance are shown in Figure 6 (a), which prove the effectiveness of our dataset. We observe that although CogVideo can output the face video given a text description, the text-video pair is not quite relevant, such as \"bags under eyes\" and \"wavy hair\". However, MMVID [19] produces videos with high relevance to input texts, containing all attributes described in the text. More results are shown in the Appendix.", + "bbox": [ + 496, + 595, + 890, + 707 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Dynamic Face Video Generation. We follow the above experimental setting to validate the effectiveness of our dataset with dynamic attribute changes (i.e., emotion, action and light direction). Due to the difficulty in modelling state change [27, 51], we follow [5] to apply test-time interpolation to MMVID [19], named MMVID-interp, to improve the text encoding and better understand the dynamics. Details of our modification are shown in the Appendix.", + "bbox": [ + 496, + 707, + 890, + 818 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In Figure 6 (b), we observe that CogVideo fails to reflect the temporal change described in the input text, i.e., smile", + "bbox": [ + 498, + 818, + 893, + 847 + ], + "page_idx": 5 + }, + { + "type": "page_footnote", + "text": "2We choose CogVideo [26] as the representative large-scale model for comparison, since the inference code and pretrained models of other large-scale methods (e.g., CogVideo [26], Phenaki [55],Imagen Video [24], and Make-A-Video [51]) are not public.", + "bbox": [ + 498, + 853, + 893, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "14810", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/efae64ba66f72d93c5ff2220733fd9f97a67492b842e56d1eb24ba07da6c4fbf.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 135, + 80, + 178, + 109 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The man has black hair and bags under eyes. He has a beard, goatee and arched eyebrows.", + "bbox": [ + 179, + 88, + 727, + 102 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ba52a7552e850638ece4b6b6087f7e53b959ae5de42549044c9215ee4b796997.jpg", + "image_caption": [ + "(a) Static - General Appearance" + ], + "image_footnote": [], + "bbox": [ + 102, + 112, + 866, + 266 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/74b2b38e3ce9c9cf64c6b6bbcf306e44b89fed6a6b604f4806da868eab81ce21.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 133, + 280, + 174, + 308 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The young woman is wearing lipstick and earrings. She firstly smiles and then turns her head.", + "bbox": [ + 176, + 287, + 748, + 301 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/3932e1044d7d61834cb037e35b9f0c6004423953232f3ed597b93f856772e966.jpg", + "image_caption": [ + "(b) Dynamic - Action", + "Figure 6. Qualitative results of facial text-to-video generation. The generated samples are given texts describing (a) the static attribute and (b) dynamic attribute." + ], + "image_footnote": [], + "bbox": [ + 101, + 309, + 864, + 534 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/2a5b0c7d0e2c1351f7e2b85915032c0ba9b614092494afbc88176411b50665db.jpg", + "table_caption": [ + "Table 4. Benchmark of text-to-video generation on different datasets. $\\downarrow$ means a lower value is better and $\\uparrow$ means the opposite.", + "(a) Quantitative results on general appearance descriptions." + ], + "table_footnote": [], + "table_body": "
DatasetMethodFVD(↓)FID(↓)CLIPSIM(↑)
MM-Vox [19]TFGAN [5]502.28 ± 1.66760.24 ± 16.010.165 ± 0.022
MMVID [19]65.79 ± 1.8138.81 ± 3.660.170 ± 0.020
CelebV-HQ [66]TFGAN [5]428.04 ± 1.76616.24 ± 17.450.168 ± 0.021
MMVID [19]73.65 ± 1.4363.86 ± 3.660.172 ± 0.019
CelebV-TextTFGAN [5]403.04 ± 1.34589.24 ± 16.460.177 ± 0.012
MMVID [19]66.69 ± 1.3558.70 ± 4.670.198 ± 0.014
", + "bbox": [ + 80, + 611, + 460, + 689 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "$\\rightarrow$ turn. However, both MMVID [19] and MMVID-interp trained on CelebV-Text can successfully model the dynamic attribute changes, which demonstrates the effectiveness of our dataset. In addition, we find that MMVID [19] cannot preserve some attributes well (e.g., earrings), while MMVID-interp can stabilize the sampling process, validating the effectiveness of our modification. More results are shown in the Appendix.", + "bbox": [ + 75, + 699, + 467, + 811 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Note that CogVideo [26] has a much larger model size ( $\\sim 100$ times larger than MMVID [19]) and is trained on much large text-video data ( $\\sim 75$ times larger than CelebV-Text). However, video samples produced by CogVideo [26] shown in Figure 6 are of a lower quality than the ones by MMVID [19] trained solely on CelebV-Text, where gener", + "bbox": [ + 75, + 815, + 467, + 900 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/2839e4c4536ba0d9f794d000acc62263c4f09bdaa3205a831d9b62686d925b6b.jpg", + "table_caption": [ + "(b) Quantitative results on dynamic descriptions of CelebV-Text." + ], + "table_footnote": [], + "table_body": "
DatasetMethodFVD(↓)FID(↓)CLIPSIM(↑)
CelebV-TextTFGAN [5]442.30 ± 2.56623.17 ± 18.880.158 ± 0.024
App.+Emo.MMVID [19]82.78 ± 1.4761.58 ± 3.990.176 ± 0.008
MMVID-interp72.87 ± 1.2341.57 ± 3.560.182 ± 0.010
CelebV-TextTFGAN [5]571.34 ± 4.54784.93 ± 20.130.154 ± 0.028
App.+Act.MMVID [19]109.25 ± 2.1182.55 ± 4.370.174 ± 0.019
MMVID-interp80.81 ± 2.5570.88 ± 4.770.176 ± 0.020
", + "bbox": [ + 506, + 611, + 883, + 689 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "ated faces are not in a high relevance to input texts, demonstrating the effectiveness of our facial text-video dataset.", + "bbox": [ + 498, + 700, + 888, + 728 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.2. Benchmark on Facial Text-to-Video Generation", + "text_level": 1, + "bbox": [ + 500, + 734, + 890, + 750 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "As the domain of text-to-video generation is currently thriving, there exists only one benchmark in the face domain, MM-Vox [19]. We expand [19] and construct a benchmark of facial text-to-video generation tasks on three datasets: MM-Vox [19], CelebV-HQ [66] with texts generated by our templates, and CelebV-Text. We choose two representative methods3, TFGAN [5] and MMVID [19], to evaluate their performances on all datasets.", + "bbox": [ + 496, + 756, + 890, + 868 + ], + "page_idx": 6 + }, + { + "type": "page_footnote", + "text": "3Other methods, e.g., CogVideo [26], Phenaki [55], Imagen Video [24], and Make-A-Video [51] are not included since their training codes are not", + "bbox": [ + 500, + 875, + 890, + 898 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "14811", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/49b1fc43a2f5568f612a67090fb86cc277c7a252e1dda53770aee87931ba751f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 102, + 84, + 142, + 112 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "He has bushy eyebrows, beard and wavy hair. He has got 5 o'clock shadow and brown hair. He has bags under eyes and sideburns with mustache.", + "bbox": [ + 143, + 84, + 833, + 112 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/abe76c6b731b63028efed25c276ff8711a484ae88577c50204f564b470ab8381.jpg", + "image_caption": [ + "Figure 7. Qualitative results on three facial text-video datasets. Red and yellow regions indicate the missing of \"bags under eyes\" and the existence of \"wavy hair\" and \"bags under eyes\"." + ], + "image_footnote": [], + "bbox": [ + 84, + 114, + 851, + 339 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Quantitative Results. For thorough benchmark construction, we evaluate baseline methods given variant texts including static and dynamic attributes. We use FVD [53] (temporal consistency), FID [23] (individual frame quality), and CLIPSIM [57] (text-video relevance) as evaluation metrics following [19] and report detailed results for appearance, action, and emotion in Table 4. Evaluation steps are repeated over ten runs with mean values and standard errors reported as well. All other values are shown in the Appendix. It can be seen from Table 4 that MMVID [19] obtains good FVD/FID/CLIPSIM metrics over TFGAN [5] which fails to generate reasonable video outputs. In addition, when input texts contain descriptions about a dynamic state change in the temporal domain, the generated video quality by MMVID [19] decreases, which encourages future methods to focus more on cross-modal understanding and consistent video generation. Moreover, the performance of MMVID-interp is better than MMVID [19] on all metrics, validating the effectiveness of our modification mentioned in Section 5.1. Due to challenges posed by our dataset and text-to-video generation task, there is still considerable room to improve.", + "bbox": [ + 75, + 378, + 468, + 685 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Qualitative Results. Video samples generated from MMVID [19] trained on different datasets are shown in Figure 7, where all video frames are of $128^2$ . We can see that video samples generated by MMVID [19] trained on different datasets are of high quality with temporal consistency. However, MMVID [19] trained on MM-Vox [19] can sometimes fail to generate attributes mentioned in the input texts. More generated video samples with dynamic attribute changes are shown in the Appendix.", + "bbox": [ + 75, + 685, + 468, + 811 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Discussion", + "text_level": 1, + "bbox": [ + 76, + 824, + 191, + 839 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We have proposed CelebV-Text, a large-scale, high-quality, and diverse facial text-video dataset with static and", + "bbox": [ + 76, + 848, + 468, + 878 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "dynamic attributes. CelebV-Text contains 70,000 video clips, each of which is accompanied by 20 individual sentences describing both static and dynamic factors. Through extensive statistical analysis and experiments, we have demonstrated the superiority and effectiveness of CelebV-Text. In the future, we plan to further enlarge CelebV-Text in both scale and diversity. We may further explore several new tasks based on CelebV-Text, such as fine-grained control of video face, adaptation of general pretrained models to the face domain, and text-driven 3D-aware facial video generation.", + "bbox": [ + 500, + 378, + 890, + 532 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Ethical Consideration. CelebV-Text is intended for research purposes only. While the raw videos will not be released, the data annotations, links to raw videos, and data processing tools will be made available after undergoing a rigorous legality check procedure at our institution. It is worth noting that our data annotation does not include any personal biometric information such as identity. Only generic attribute information such as gender, hair color, and motion is annotated. Additionally, synthetic videos generated in this work do not exhibit bias or certain biometric information (e.g., big lips or big nose), alleviating ethical concerns. CelebV-Text may be used for deepfakes, but it can also be used for forgery detection tasks to prevent such issues. We will try our best to control the application and acquisition procedure of CelebV-Text to avoid potential misuse and abuse. In the future, we plan to use synthetic face generation frameworks to generate synthetic face videos to address the ethical shortcomings of existing real-world face video datasets.", + "bbox": [ + 496, + 541, + 892, + 806 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgement. CelebV-Text is developed under OpenXDLab - an open platform for X-Dimension high-quality data. This study is supported by the RIE2020 Industry Alignment Fund Industry Collaboration Projects (IAF-ICP) Funding Initiative, in-kind contribution from the industry partner(s), and the MOE AcRF Tier 1 (RG16/21).", + "bbox": [ + 496, + 815, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_footnote", + "text": "public so far.", + "bbox": [ + 76, + 887, + 148, + 898 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "14812", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Niki Aifanti, Christos Papachristou, and Anastasios Delopoulos. The mug facial expression database. In WIAMIS, 2010. 2", + "[2] Jean-Baptiste Alayrac, Piotr Bojanowski, Nishant Agrawal, Josef Sivic, Ivan Laptev, and Simon Lacoste-Julien. Unsupervised learning from narrated instruction videos. In CVPR, 2016. 1", + "[3] Lisa Anne Hendricks, Oliver Wang, Eli Shechtman, Josef Sivic, Trevor Darrell, and Bryan Russell. Localizing moments in video with natural language. In ICCV, 2017. 2, 4", + "[4] Max Bain, Arsha Nagrani, Gül Varol, and Andrew Zisserman. Frozen in time: A joint video and image encoder for end-to-end retrieval. In ICCV, 2021. 1, 4", + "[5] Yogesh Balaji, Martin Renqiang Min, Bing Bai, Rama Chellappa, and Hans Peter Graf. Conditional gan with discriminative filter generation for text-to-video synthesis. In IJCAI, 2019. 1, 2, 4, 6, 7, 8", + "[6] Alex Bewley, Zongyuan Ge, Lionel Ott, Fabio Ramos, and Ben Upcroft. Simple online and realtime tracking. In ICIP. IEEE, 2016. 3", + "[7] Sergey Bezryadin, Pavel Bourov, and Dmitry Ilinih. Brightness calculation in digital image processing. In TDPF, 2007. 4", + "[8] Steven Bird, Ewan Klein, and Edward Loper. Natural language processing with Python: analyzing text with the natural language toolkit. \"O'Reilly Media, Inc.\", 2009. 4", + "[9] David Chen and William B Dolan. Collecting highly parallel data for paraphrase evaluation. In ACL, 2011. 2, 4", + "[10] Danqi Chen and Christopher D Manning. A fast and accurate dependency parser using neural networks. In EMNLP, 2014. 2, 4", + "[11] Joon Son Chung, Arsha Nagrani, and Andrew Zisserman. Voxceleb2: Deep speaker recognition. In INTERSPEECH, 2018. 1, 2, 3, 5", + "[12] Karl Cobbe, Oleg Klimov, Chris Hesse, Taehoon Kim, and John Schulman. Quantifying generalization in reinforcement learning. In ICML, 2019. 2", + "[13] Dima Damen, Hazel Doughty, Giovanni Maria Farinella, Antonino Furnari, Evangelos Kazakos, Jian Ma, Davide Moltisanti, Jonathan Munro, Toby Perrett, Will Price, et al. Rescaling egocentric vision: collection, pipeline and challenges for epic-kitchens-100. In IJCV, 2022. 2", + "[14] Jiankang Deng, Jia Guo, Niannan Xue, and Stefanos Zafeiriou. Arcface: Additive angular margin loss for deep face recognition. In CVPR, 2019. 3", + "[15] Kangle Deng, Tianyi Fei, Xin Huang, and Yuxin Peng. IRCAN: Introspective recurrent convolutional gan for text-to-video generation. In IJCAI, pages 2216-2222, 2019. 2", + "[16] Patrick Esser, Robin Rombach, and Bjorn Ommer. Taming transformers for high-resolution image synthesis. In CVPR, 2021. 2", + "[17] Han Fang, Pengfei Xiong, Luhui Xu, and Yu Chen. Clip2video: Mastering video-text retrieval via image clip. arXiv preprint arXiv:2106.11097, 2021. 2, 6", + "[18] Tanmay Gupta, Dustin Schwenk, Ali Farhadi, Derek Hoiem, and Aniruddha Kembhavi. Imagine this! scripts to compositions to videos. In ECCV, 2018. 1", + "[19] Ligong Han, Jian Ren, Hsin-Ying Lee, Francesco Barbieri, Kyle Olszewski, Shervin Minaee, Dimitris Metaxas, and" + ], + "bbox": [ + 78, + 114, + 467, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Sergey Tulyakov. Show me what and tell me how: Video synthesis via multimodal conditioning. In CVPR, 2022. 2, 3, 4, 5, 6, 7, 8", + "[20] William Harvey, Saeid Naderiparizi, Vaden Masrani, Christian Weilbach, and Frank Wood. Flexible diffusion modeling of long videos. arXiv preprint arXiv:2205.11495, 2022. 2", + "[21] Thomas Hayes, Songyang Zhang, Xi Yin, Guan Pang, Sasha Sheng, Harry Yang, Songwei Ge, Isabelle Hu, and Devi Parikh. Mugen: A playground for video-audio-text multimodal understanding and generation. arXiv preprint arXiv:2204.08058, 2022. 2, 4", + "[22] Javier Hernandez-Andres, Raymond L Lee, and Javier Romero. Calculating correlated color temperatures across the entire gamut of daylight and skylight chromaticities. In Applied optics, 1999. 4", + "[23] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. NeurIPS, 2017. 8", + "[24] Jonathan Ho, William Chan, Chitwan Sahara, Jay Whang, Ruiqi Gao, Alexey Gritsenko, Diederik P Kingma, Ben Poole, Mohammad Norouzi, David J Fleet, et al. Imagen video: High definition video generation with diffusion models. arXiv preprint arXiv:2210.02303, 2022. 2, 6, 7", + "[25] Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan, Mohammad Norouzi, and David J Fleet. Video diffusion models. arXiv preprint arXiv:2204.03458, 2022. 2", + "[26] Wenyi Hong, Ming Ding, Wendi Zheng, Xinghan Liu, and Jie Tang. Cogvideo: Large-scale pretraining for text-to-video generation via transformers. arXiv preprint arXiv:2205.15868, 2022. 2, 6, 7", + "[27] Yaosi Hu, Chong Luo, and Zhenzhong Chen. Make it move: controllable image-to-video generation with text descriptions. In CVPR, 2022. 2, 4, 6", + "[28] Justin Johnson, Bharath Hariharan, Laurens Van Der Maaten, Li Fei-Fei, C Lawrence Zitnick, and Ross Girshick. Clevr: A diagnostic dataset for compositional language and elementary visual reasoning. In CVPR, 2017. 2", + "[29] Peter Kán and Hannes Kafumann. Deep light: light source estimation for augmented reality using deep learning. The Visual Computer, 2019. 4", + "[30] Doyeon Kim, Donggyu Joo, and Junmo Kim. Tivgan: Text to image to video generation with step-by-step evolutionary generator. In IEEE Access, 2020. 3", + "[31] Dan Klein and Christopher D Manning. Accurate unlexicalized parsing. In ACL, 2003. 4", + "[32] Ranjay Krishna, Kenji Hata, Frederic Ren, Li Fei-Fei, and Juan Carlos Niebles. Dense-captioning events in videos. In ICCV, 2017. 2, 4", + "[33] Dingquan Li, Tingting Jiang, and Ming Jiang. Quality assessment of in-the-wild videos. In ACM MM, pages 2351-2359, 2019. 5", + "[34] Yitong Li, Martin Min, Dinghan Shen, David Carlson, and Lawrence Carin. Video generation from text. In Proceedings of the AAAI conference on artificial intelligence, 2018. 1", + "[35] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dálár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In ECCV, 2014. 2", + "[36] Wen Liu, Zhixin Piao, Jie Min, Wenhan Luo, Lin Ma, and Shenghua Gao. Liquid warping gan: A unified framework" + ], + "bbox": [ + 503, + 92, + 890, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "14813", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "for human motion imitation, appearance transfer and novel view synthesis. In ICCV, 2019. 2", + "[37] Yue Liu, Xin Wang, Yitian Yuan, and Wenwu Zhu. Cross-modal dual learning for sentence-to-video generation. In ACM MM, 2019. 1", + "[38] Ziwei Liu, Ping Luo, Xiaogang Wang, and Xiaou Tang. Deep learning face attributes in the wild. In ICCV, 2015. 3, 4", + "[39] Jonathan Malmaud, Jonathan Huang, Vivek Rathod, Nick Johnston, Andrew Rabinovich, and Kevin Murphy. What's cookin'? interpreting cooking videos using text, speech and vision. In ACL, 2015. 1", + "[40] Antoine Miech, Dimitri Zhukov, Jean-Baptiste Alayrac, Makarand Tapaswi, Ivan Laptev, and Josef Sivic. Howto100m: Learning a text-video embedding by watching hundred million narrated video clips. In CVPR, 2019. 2, 4", + "[41] Niluthpol Chowdhury Mithun, Juncheng Li, Florian Metze, and Amit K Roy-Chowdhury. Learning joint embedding with multimodal cues for cross-modal video-text retrieval. In ICMR, 2018. 6", + "[42] Anish Mittal, Anush Krishna Moorthy, and Alan Conrad Bovik. No-reference image quality assessment in the spatial domain. In TIP, 2012. 5", + "[43] Gaurav Mittal, Tanya Marwah, and Vineeth N Balasubramanian. Sync画: Automatic video generation using deep recurrent attentive architectures. In ACM MM, 2017. 1, 2", + "[44] Ali Mollahosseini, Behzad Hasani, and Mohammad H Ma-hoor. Affectnet: A database for facial expression, valence, and arousal computing in the wild. In TAC, 2017. 4", + "[45] A. Nagrani, J. S. Chung, and A. Zisserman. Voxceleb: a large-scale speaker identification dataset. In INTER-SPEECH, 2017. 1, 3", + "[46] Yingwei Pan, Zhaofan Qiu, Ting Yao, Houqiang Li, and Tao Mei. To create what you tell: Generating videos from captions. In ACM MM, 2017. 2", + "[47] MarcAurelio Ranzato, Arthur Szlam, Joan Bruna, Michael Mathieu, Ronan Collobert, and Sumit Chopra. Video (language) modeling: a baseline for generative models of natural videos. arXiv preprint arXiv:1412.6604, 2014. 1", + "[48] Andreas Rössler, Davide Cozzolino, Luisa Verdoliva, Christian Riess, Justus Thies, and Matthias Nießner. Faceforensics: A large-scale video dataset for forgery detection in human faces. arXiv preprint arXiv:1803.09179, 2018. 1", + "[49] Christoph Schuhmann, Richard Vencu, Romain Beaumont, Robert Kaczmarczyk, Clayton Mullis, Aarush Katta, Theo Coombes, Jenia Jitsev, and Aran Komatsuzaki. Laion-400m: Open dataset of clip-filtered 400 million image-text pairs. arXiv preprint arXiv:2111.02114, 2021. 2", + "[50] Piyush Sharma, Nan Ding, Sebastian Goodman, and Radu Soricut. Conceptual captions: A cleaned, hypernymed, image alt-text dataset for automatic image captioning. In ACL, 2018. 2", + "[51] Uriel Singer, Adam Polyak, Thomas Hayes, Xi Yin, Jie An, Songyang Zhang, Qiyuan Hu, Harry Yang, Oron Ashual, Oran Gafni, et al. Make-a-video: Text-to-video generation without text-video data. arXiv preprint arXiv:2209.14792, 2022. 2, 6, 7", + "[52] David Stap, Maurits Bleeker, Sarah Ibrahimi, and Maartje ter Hoeve. Conditional image generation and manipulation for user-specified content. arXiv preprint arXiv:2005.04909, 2020. 2, 4" + ], + "bbox": [ + 78, + 92, + 468, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[53] Thomas Unterthiner, Sjoerd van Steenkiste, Karol Kurach, Raphael Marinier, Marcin Michalski, and Sylvain Gelly. Towards accurate generative models of video: A new metric & challenges. arXiv preprint arXiv:1812.01717, 2018. 8", + "[54] Aaron Van Den Oord, Oriol Vinyals, et al. Neural discrete representation learning. NeurIPS, 2017. 2", + "[55] Ruben Villegas, Mohammad Babaeizadeh, Pieter-Jan Kindermans, Hernan Moraldo, Han Zhang, Mohammad Taghi Saffar, Santiago Castro, Julius Kunze, and Dumitru Erhan. Phenaki: Variable length video generation from open domain textual description. arXiv preprint arXiv:2210.02399, 2022. 2, 6, 7", + "[56] Xin Wang, Jiawei Wu, Junkun Chen, Lei Li, Yuan-Fang Wang, and William Yang Wang. Vatex: A large-scale, high-quality multilingual dataset for video-and-language research. In ICCV, 2019. 4, 5, 6", + "[57] Chenfei Wu, Lun Huang, Qianxi Zhang, Binyang Li, Lei Ji, Fan Yang, Guillermo Sapiro, and Nan Duan. Godiva: Generating open-domain videos from natural descriptions. arXiv preprint arXiv:2104.14806, 2021. 2, 8", + "[58] Chenfei Wu, Jian Liang, Xiaowei Hu, Zhe Gan, Jianfeng Wang, Lijuan Wang, Zicheng Liu, Yuejian Fang, and Nan Duan. Nuwa-infinity: Autoregressive over autoregressive generation for infinite visual synthesis. In NeurIPS, 2022. 2", + "[59] Chenfei Wu, Jian Liang, Lei Ji, Fan Yang, Yuejian Fang, Daxin Jiang, and Nan Duan. N\\`uwa: Visual synthesis pretraining for neural visual world creation. In ECCV, 2022. 2", + "[60] Wayne Wu, Yunxuan Zhang, Cheng Li, Chen Qian, and Chen Change Loy. Reenactgan: Learning to reenact faces via boundary transfer. In ECCV, 2018. 1, 3, 5", + "[61] Weihao Xia, Yujiu Yang, Jing-Hao Xue, and Baoyuan Wu. Tedigan: Text-guided diverse face image generation and manipulation. In CVPR, 2021. 3, 4", + "[62] Jun Xu, Tao Mei, Ting Yao, and Yong Rui. Msr-vtt: A large video description dataset for bridging video and language. In CVPR, 2016. 2, 4", + "[63] Gilbert Youmans. Measuring lexical style and competence: The type-token vocabulary curve. Style, 1990. 6", + "[64] Bowen Zhang, Hexiang Hu, and Fei Sha. Cross-modal and hierarchical modeling of video and text. In ECCV, 2018. 6", + "[65] Luowei Zhou, Chenliang Xu, and Jason J Corso. Towards automatic learning of procedures from web instructional videos. In AAAI, 2018. 2", + "[66] Hao Zhu, Wayne Wu, Wentao Zhu, Liming Jiang, Siwei Tang, Li Zhang, Ziwei Liu, and Chen Change Loy. CelebV-HQ: A large-scale video facial attributes dataset. In ECCV, 2022. 2, 3, 4, 5, 6, 7", + "[67] Dimitri Zhukov, Jean-Baptiste Alayrac, Ramadan Gokberk Cinbis, David Fouhey, Ivan Laptev, and Josef Sivic. Cross-task weakly supervised learning from instructional videos. In CVPR, 2019. 1, 4" + ], + "bbox": [ + 501, + 92, + 890, + 785 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "14814", + "bbox": [ + 480, + 945, + 519, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2023/CelebV-Text_ A Large-Scale Facial Text-Video Dataset/b3bc8119-1bc3-419d-89b1-046a6b2d92d9_model.json b/2023/CelebV-Text_ A Large-Scale Facial Text-Video Dataset/b3bc8119-1bc3-419d-89b1-046a6b2d92d9_model.json new file mode 100644 index 0000000000000000000000000000000000000000..9879859ad3ed8ada2ec33957b43613e1b58acc72 --- /dev/null +++ b/2023/CelebV-Text_ A Large-Scale Facial Text-Video Dataset/b3bc8119-1bc3-419d-89b1-046a6b2d92d9_model.json @@ -0,0 +1,2442 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.21, + 0.115, + 0.761, + 0.136 + ], + "angle": 0, + "content": "CelebV-Text: A Large-Scale Facial Text-Video Dataset" + }, + { + "type": "text", + "bbox": [ + 0.098, + 0.149, + 0.87, + 0.168 + ], + "angle": 0, + "content": "Jianhui \\(\\mathrm{Yu}^{1*}\\) Hao Zhu\\(^{2*}\\) Liming Jiang\\(^{3}\\) Chen Change Loy\\(^{3}\\) Weidong Cai\\(^{1}\\) Wayne Wu\\(^{4}\\)" + }, + { + "type": "text", + "bbox": [ + 0.125, + 0.169, + 0.845, + 0.185 + ], + "angle": 0, + "content": "\\(^{1}\\)University of Sydney \\(^{2}\\)SenseTime Research \\(^{3}\\)S-Lab, Nanyang Technological University \\(^{4}\\)Shanghai AI Laboratory" + }, + { + "type": "text", + "bbox": [ + 0.15, + 0.186, + 0.813, + 0.2 + ], + "angle": 0, + "content": "jianhui.yu@sydney.edu.au haozhu96@gmail.com {liming002,ccloy}@ntu.edu.sg" + }, + { + "type": "text", + "bbox": [ + 0.28, + 0.202, + 0.686, + 0.216 + ], + "angle": 0, + "content": "tom.cai@sydney.edu.au wuwenyan0503@gmail.com" + }, + { + "type": "image", + "bbox": [ + 0.081, + 0.242, + 0.639, + 0.495 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.154, + 0.495, + 0.566, + 0.506 + ], + "angle": 0, + "content": "(a) Video Samples (left: general appearance, middle: detailed appearance, right: light conditions)" + }, + { + "type": "image", + "bbox": [ + 0.645, + 0.241, + 0.886, + 0.495 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.719, + 0.495, + 0.812, + 0.506 + ], + "angle": 0, + "content": "(b) Text Descriptions" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.51, + 0.893, + 0.537 + ], + "angle": 0, + "content": "Figure 1. Overview of CelebV-Text. CelebV-Text contains (a) 70,000 video samples and (b) 1,400,000 text descriptions. Each video sample is annotated with general appearance, detailed appearance, light conditions, action, emotion, and light directions." + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.55, + 0.313, + 0.565 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.577, + 0.471, + 0.854 + ], + "angle": 0, + "content": "Text-driven generation models are flourishing in video generation and editing. However, face-centric text-to-video generation remains a challenge due to the lack of a suitable dataset containing high-quality videos and highly relevant texts. This paper presents CelebV-Text, a large-scale, diverse, and high-quality dataset of facial text-video pairs, to facilitate research on facial text-to-video generation tasks. CelebV-Text comprises 70,000 in-the-wild face video clips with diverse visual content, each paired with 20 texts generated using the proposed semi-automatic text generation strategy. The provided texts are of high quality, describing both static and dynamic attributes precisely. The superiority of CelebV-Text over other datasets is demonstrated via comprehensive statistical analysis of the videos, texts, and text-video relevance. The effectiveness and potential of CelebV-Text are further shown through extensive self-evaluation. A benchmark is constructed with representative methods to standardize the evaluation of the facial text-to-video generation task. All data and models are publicly available1." + }, + { + "type": "title", + "bbox": [ + 0.502, + 0.55, + 0.631, + 0.565 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.577, + 0.892, + 0.744 + ], + "angle": 0, + "content": "Text-driven video generation has recently garnered significant attention in the fields of computer vision and computer graphics. By using text as input, video content can be generated and controlled, inspiring numerous applications in both academia and industry [5,34,43,47]. However, text-to-video generation still faces many challenges, particularly in the face-centric scenario where generated video frames often lack quality [18, 34, 37] or have weak relevance to input texts [2,4,39,67]. We believe that one of the main issues is the absence of a well-suited facial text-video dataset containing high-quality video samples and text descriptions of various attributes highly relevant to videos." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.747, + 0.893, + 0.901 + ], + "angle": 0, + "content": "Constructing a high-quality facial text-video dataset poses several challenges, mainly in three aspects. 1) Data collection. The quality and quantity of video samples largely determine the quality of generated videos [11, 45, 48, 60]. However, obtaining such a large-scale dataset with high-quality samples while maintaining a natural distribution and smooth video motion is challenging. 2) Data annotation. The relevance of text-video pairs needs to be ensured. This requires a comprehensive coverage of text for describing the content and motion appearing in the video, such as light conditions and head movements. 3) Text gen" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.875, + 0.207, + 0.888 + ], + "angle": 0, + "content": "*Equal contribution." + }, + { + "type": "page_footnote", + "bbox": [ + 0.097, + 0.888, + 0.406, + 0.901 + ], + "angle": 0, + "content": "1Project page: https://celebv-text.github.io" + }, + { + "type": "list", + "bbox": [ + 0.096, + 0.875, + 0.406, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "14805" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.468, + 0.146 + ], + "angle": 0, + "content": "eration. Producing diverse and natural texts are non-trivial. Manual text generation is expensive and not scalable. While auto-text generation is easily extensible, it is limited in naturalness." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.148, + 0.468, + 0.397 + ], + "angle": 0, + "content": "To overcome the challenges mentioned above, we carefully design a comprehensive data construction pipeline that includes data collection and processing, data annotation, and semi-auto text generation. First, to obtain raw videos, we follow the data collection steps of CelebV-HQ, which has proven to be effective in [66]. We introduce a minor modification to the video processing step to improve the video's smoothness further. Next, to ensure highly relevant text-video pairs, we analyze videos from both temporal dynamics and static content and establish a set of attributes that may or may not change over time. Finally, we propose a semi-auto template-based method to generate texts that are diverse and natural. Our approach leverages the advantages of both auto- and manual-text methods. Specifically, we design a rich variety of grammar templates as [10,52] to parse annotation and manual texts, which are flexibly combined and modified to achieve high diversity, complexity, and naturalness." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.399, + 0.468, + 0.635 + ], + "angle": 0, + "content": "With the proposed pipeline, we create CelebV-Text, a Large-Scale Facial Text-Video Dataset, which includes 70,000 in-the-wild video clips with a resolution of at least \\(512 \\times 512\\) and 1,400,000 text descriptions with 20 for each clip. As depicted in Figure 1, CelebV-Text consists of high-quality video samples and text descriptions for realistic face video generation. Each video is annotated with three types of static attributes (40 general appearances, 5 detailed appearances, and 6 light conditions) and three types of dynamic attributes (37 actions, 8 emotions, and 6 light directions). All dynamic attributes are densely annotated with start and end timestamps, while manual-texts are provided for labels that cannot be discretized. Furthermore, we have designed three templates for each attribute type, resulting in a total of 18 templates that can be flexibly combined. All attributes and manual-texts are naturally described in our generated texts." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.636, + 0.468, + 0.884 + ], + "angle": 0, + "content": "CelebV-Text surpasses existing face video datasets [11] in terms of resolution (over 2 times higher), number of samples, and more diverse distribution. In addition, the texts in CelebV-Text exhibit higher diversity, complexity, and naturalness than those in text-video datasets [19, 66]. CelebV-Text also shows high relevance of text-video pairs, validated by our text-video retrieval experiments [17]. To further examine the effectiveness and potential of CelebV-Text, we evaluate it on a representative baseline [19] for facial text-to-video generation. Our results show better relevance between generated face videos and texts when compared to a state-of-the-art large-scale pretrained model [26]. Furthermore, we show that a simple modification of [19] with text interpolation can significantly improve temporal coherence. Finally, we present a new benchmark for text-to-video generation to standardize the facial text-to-video generation task, which includes representative models [5, 19] on three text-video datasets." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.886, + 0.468, + 0.899 + ], + "angle": 0, + "content": "The main contributions of this work are summarized as" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.245 + ], + "angle": 0, + "content": "follows: 1) We propose CelebV-Text, the first large-scale facial text-video dataset with high-quality videos, as well as rich and highly-relevant texts, to facilitate research in facial text-to-video generation. 2) Comprehensive statistical analyses are conducted to examine video/text quality and diversity, as well as text-video relevance, demonstrating the superiority of CelebV-Text. 3) A series of self-evaluations are performed to demonstrate the effectiveness and potential of CelebV-Text. 4) A new benchmark for text-to-video generation is constructed to promote the standardization of the facial text-to-video generation task." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.261, + 0.642, + 0.277 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.287, + 0.892, + 0.676 + ], + "angle": 0, + "content": "Text-to-Video Generation. Text-driven video generation, which involves generating videos from text descriptions, has recently gained significant interest as a challenging task. Mittal et al. [43] first introduced this task to generate semantically consistent videos conditioned on encoded captions. Other studies, such as [5, 15, 46], attempt to generate video samples conditioned on encoded text inputs. However, due to the low richness of text descriptions and the small number of data samples, the generated video samples are often at low resolution or lack relevance with the input texts. More recently, several works [19, 26, 27, 55, 57–59] have employed discrete latent codes [16, 54] for more realistic video generation. Some of these works treat videos as a sequence of independent images [19, 26, 58, 59], while Phenaki [55] considers temporal relations between each frame for a more robust video decoding process. Another branch of studies leverage diffusion models for text-to-video generation [20, 24, 25, 51], which require millions or billions of samples to achieve high-quality generation. While text-to-video generation methods are rapidly evolving, they are generally designed for generating general videos. Among these methods, only MMVID [19] has conducted specific experiments with face-centric descriptions. One possible reason for this is that facial text-to-video generation requires more accurate and detailed text descriptions than general tasks. However, there is currently no suitable dataset available that provides such properties for face-centric text-to-video generation." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.677, + 0.892, + 0.884 + ], + "angle": 0, + "content": "Multimodal Datasets. Existing multimodal datasets can be categorized into two classes: open-world and closed-world. Open-world datasets [3,9,13,32,35,36,40,43,49,50,62,65] are widely used for text-to-image/video generation tasks. Some of them have manual annotations [13, 32, 35, 50, 62] and part of them are directly collected from the Internet, such as subtitles [40, 49]. Closed-world datasets are mostly composed of images or videos collected in constrained environment with corresponding information such as text. CLEVR [28] is a synthetic text-image dataset produced by arranging 3D objects with different shapes under a controlled background. While MUGEN [21] is a video-audiotext dataset that was collected using CoinRun [12] by introducing audio and new interactions. The corresponding text is produced by human annotators and grammar templates." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.886, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Multimodal face datasets also exist. Modified MUG [1]" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "14806" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.078, + 0.892, + 0.118 + ], + "angle": 0, + "content": "Table 1. In-the-wild face video dataset comparison. The symbol “#” indicates the number. The abbreviations “Res.”, “Dura.”, “App.”, “Cond.”, “Act.”, “Emo.”, and “Dir.” stand for Resolution, Duration, Appearance, Condition, Action, Emotion, and Direction, respectively. The “half checkmark” denotes that CelebV-HQ consists of action attributes with no timestamp." + }, + { + "type": "table", + "bbox": [ + 0.1, + 0.122, + 0.868, + 0.263 + ], + "angle": 0, + "content": "
DatasetsMeta InformationAttribute LabelsText
StaticDynamic
#SamplesRes.Dura.General App.Detail App.Light Cond.Act.Emo.Light Dir.Auto Manual
CelebV [60]5256×2562hrs
VoxCeleb2 [11]150,480224×2242442hrs
CelebV-HQ [66]35,666512×51268hrs✓XX✓✓XX
MM-Vox [19]19,522224×224323hrs✓XXXXX✓
CelebV-Text70,000512×512+279hrs✓✓✓✓✓✓✓
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.28, + 0.473, + 0.532 + ], + "angle": 0, + "content": "is a closed-world text-video dataset that contains 1,039 videos with subjects showing different emotions, where the text descriptions are generated from facial emotions using a fixed template [30]. MM-Vox [19] contains 19,522 face videos from VoxCeleb [45], with 36 facial attributes manually labeled following CelebA [38] and text descriptions generated via Probabilistic Context-Free Grammar (PCFG) [61]. However, both datasets only contain language descriptions related to static facial attributes without considering the temporal state change (i.e., emotion or action) presented in the original face videos. Moreover, the limited label annotations restrict the diversity of the text descriptions, making them sub-optimal for studying the text-to-video generation task on the face domain. CelebV-HQ [66] is the latest high-quality face video dataset that covers facial annotations, including appearance, movement, and emotion. However, it only provides discrete labels and timestamps, with no text descriptions." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.547, + 0.281, + 0.563 + ], + "angle": 0, + "content": "3. Dataset Construction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.572, + 0.47, + 0.671 + ], + "angle": 0, + "content": "In this work, we aim to build a facial text-video dataset, which requires not only large-scale video samples of high quality, but also natural and diverse text descriptions that are highly relevant to videos. To achieve so, we propose an efficient pipeline, as shown in Figure 2, to construct CelebV-Text, including Data Collection & Processing, Data Annotation, and Semi-auto Text Generation." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.681, + 0.34, + 0.699 + ], + "angle": 0, + "content": "3.1. Data Collection & Processing" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.705, + 0.47, + 0.844 + ], + "angle": 0, + "content": "Collection. We follow the same strategy as CelebV-HQ [66] due to its effectiveness in large-scale high-quality data collection. Specifically, we firstly generate a large number of queries, including human names, movie titles, vlogs and so on, to retrieve videos that contain human faces with temporally dynamic state changes and abundant facial attributes. Our data are collected from open world with videos downloaded from online resources. Videos with low resolution \\((< 512^{2})\\), low time duration \\((< 5s)\\), and having appeared in CelebV-HQ are filtered out." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.845, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Processing. To sample high-quality and diverse video clips from our raw collections, similar steps are followed as CelebV-HQ [66] with modifications. We first filter out video clips with bounding box regions less than \\(512^{2}\\) rather" + }, + { + "type": "image", + "bbox": [ + 0.507, + 0.279, + 0.891, + 0.469 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.482, + 0.892, + 0.523 + ], + "angle": 0, + "content": "Figure 2. Pipeline of our dataset construction process. The pipeline includes data collection & processing, data annotation, and semi-auto text generation." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.532, + 0.893, + 0.673 + ], + "angle": 0, + "content": "than resize them to the same resolution. In this way, clips are not upsampled or downsampled hence the video quality would not be affected, which leads to various resolutions of collected videos: \\(56.4\\%\\) with \\(512^2 \\sim 1024^2\\), and \\(43.6\\%\\) for \\(1024^2+\\). To reduce the face area noise when the background changes, we further change the video splitting strategy. In addition to our focus on the same human motion [6] and identity [14] present in adjacent frames, we split the video into different clips when the background changes by a toolkit1." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.683, + 0.666, + 0.699 + ], + "angle": 0, + "content": "3.2. Data Annotation" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.706, + 0.892, + 0.79 + ], + "angle": 0, + "content": "The annotation process is a core part in CelebV-Text construction, which would greatly affect the relevance of text-video pairs, as our designed text templates heavily depend on the annotation results. Here, we first describe how we design attributes, and then give details about the annotation strategy for face videos." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.79, + 0.892, + 0.875 + ], + "angle": 0, + "content": "Attributes Design. Temporal dynamic is the key difference between images and videos. However, as shown in Table 1, most face video datasets focus on static attributes where attribute information does not change over time, such as appearance. Dynamic attributes that change over time, such as emotion and face actions, are often neglected. In the fol" + }, + { + "type": "page_footnote", + "bbox": [ + 0.518, + 0.887, + 0.783, + 0.901 + ], + "angle": 0, + "content": "1 https://github.com/Breakthrough/PySceneDetect" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "14807" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.123, + 0.075, + 0.382, + 0.245 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.38, + 0.075, + 0.637, + 0.246 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.637, + 0.075, + 0.835, + 0.246 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.113, + 0.248, + 0.855, + 0.262 + ], + "angle": 0, + "content": "Figure 3. Dataset distribution comparison. The distributions of appearance attributes, action attributes, and light directions." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.271, + 0.469, + 0.299 + ], + "angle": 0, + "content": "lowing, we decouple face videos into static and dynamic categories and details are given as follows." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.299, + 0.469, + 0.452 + ], + "angle": 0, + "content": "1) Static. The current dataset [19] only considers static information such as the appearance attribute, which includes 40 classes as CelebA [38]. In contrast, we define static information to include three types of attributes: general appearance, detailed appearance, and light conditions. General appearance attributes follow the same definition as CelebA [38]. Detailed appearance attributes including five classes are proposed for realistic face generation, i.e., scar, mole, freckle, dimple, and one-eyed. We define light conditions in a restricted manner to include light color temperature [22] and brightness [7], with a total of 6 classes." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.453, + 0.469, + 0.633 + ], + "angle": 0, + "content": "2) Dynamic. Here, we design three dynamic attributes, i.e., action, emotion, and light directions. For action attributes, we follow CelebV-HQ [66] and expand their action list by two classes, i.e., squint and blink. For emotion attributes, we select the 8 emotion setting in Affectnet [44], including neutral, anger, contempt, disgust, fear, happiness, sadness, and surprise. For light direction attributes, we derive and modify classes from [29] and give 6 light direction classes. Complete lists are given in the Appendix. Moreover, as shown in Table 1, CelebV-HQ [66] is the only dataset giving timestamps of dynamic attributes. Following their idea, we densely annotate all dynamic attributes of CelebV-Text with the start and end time." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.634, + 0.469, + 0.731 + ], + "angle": 0, + "content": "Automatic and Manual Annotation. Based on our attributes design, we find that some attributes can be annotated automatically (e.g., appearance) while some need manual annotations (e.g., timestamps of dynamic attributes). Considering the dataset quality and cost of expense, our annotation strategy includes both automatic and manual annotations." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.733, + 0.469, + 0.872 + ], + "angle": 0, + "content": "For automatic annotation, we first investigate algorithms and select designed attributes that can be automatically annotated. We then test different algorithms on our dataset and keep those giving annotation accuracy of \\(85\\%\\) or higher. This process yields all light condition labels, all appearance labels, and all emotion labels suitable for automatic annotation. Algorithms for different labels we finally chose are reported in the Appendix. Automatic annotation results can be further revised by human workers to improve accuracy in a less costly way." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.873, + 0.469, + 0.901 + ], + "angle": 0, + "content": "For manual annotation, we hire and train human workers following [66] to annotate attributes that are filtered out by" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.271, + 0.892, + 0.396 + ], + "angle": 0, + "content": "an automatic annotation process. In this case, we manually annotate dynamic attributes, i.e., action and light directions, to give both class labels and exact timestamps. In addition, it is hard to represent detailed appearance attributes by the discrete label, e.g., the characteristics of scars or moles. We therefore ask annotators to give a natural description for each attribute, describing exact positions relative to face parts. These designs greatly enhance the relevance between the final text and the video." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.405, + 0.743, + 0.42 + ], + "angle": 0, + "content": "3.3. Semi-auto Text Generation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.427, + 0.892, + 0.679 + ], + "angle": 0, + "content": "Multimodal text-video datasets collect texts via three common methods: subtitles [4, 40, 67], manual-text generation [3, 9, 32, 56, 62], and auto-text generation [5, 21, 27]. However, it is difficult for the individual method to generate texts with high relevance to videos, natural expression, and high diversity. Specifically, although subtitles are easy to obtain, they can pose weakly relevant text-video pairs and introduce noise, making the dataset quality hard to control. Moreover, manual-text generation method is time and cost consuming, as natural language descriptions are required for each video. In this case, increasing the data scale is quite hard as more workers are needed to describe new videos, which does not meet the efficiency and scalability of annotation. Finally, auto-text generation is flexible and scalable, as abundant texts can be simultaneously generated given annotation results of collected videos. However, the diversity, complexity, and naturalness of generated texts can be impacted by the designed grammar templates." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.679, + 0.892, + 0.762 + ], + "angle": 0, + "content": "To this end, we propose a semi-auto template-based text generation strategy that combines both manual-text and auto-text generation methods. Specifically, as mentioned in Section 3.2, manual-texts are required to describe detailed appearance attributes. Annotated attribute information is fed into our designed template for auto-text generation." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.762, + 0.892, + 0.901 + ], + "angle": 0, + "content": "To make our template as natural as possible, we first ask each annotator to describe 10 different face videos for each attribute. We then analyze the grammar structure (i.e., parse tree banks) along with online corpora following [10, 31], and find the most three common grammar structures for each attribute. Finally, we utilize probabilistic context-free grammar [52, 61] and modify the grammar structures to design our own templates. Texts are generated based on templates with synonym replacement using NLTK [8] to increase our generation diversity. Details of our template de" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.956 + ], + "angle": 0, + "content": "14808" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.107, + 0.083, + 0.413, + 0.173 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.189, + 0.175, + 0.322, + 0.186 + ], + "angle": 0, + "content": "(a) Image quality distribution" + }, + { + "type": "image", + "bbox": [ + 0.107, + 0.196, + 0.406, + 0.279 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.197, + 0.283, + 0.329, + 0.293 + ], + "angle": 0, + "content": "(b) Video quality distribution" + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.293, + 0.468, + 0.318 + ], + "angle": 0, + "content": "Figure 4. Dataset quality distribution. The metrics used are BRISQUE [42] and VSFA [33] respectively." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.329, + 0.252, + 0.343 + ], + "angle": 0, + "content": "signs are in the Appendix." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.356, + 0.389, + 0.373 + ], + "angle": 0, + "content": "4. Statistical Analysis of CelebV-Text" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.381, + 0.469, + 0.492 + ], + "angle": 0, + "content": "In this section, we compare CelebV-Text with the two most relevant and representative face video datasets [19,66]. We perform a comprehensive analysis of CelebV-Text in terms of video, text, and text-video relevance. To verify the effectiveness of our designed grammar templates, we generate text descriptions for CelebV-HQ based on its attributes for comparison. For simplicity, we use \"CelebV-HQ\" to denote this variant in the following." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.502, + 0.265, + 0.518 + ], + "angle": 0, + "content": "4.1. Video Comparisons" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.524, + 0.469, + 0.747 + ], + "angle": 0, + "content": "We briefly compare the overall statistics of existing face video datasets [11, 19, 60, 66] in Table 1. As reported, CelebV-Text contains 70,000 video clips with a total duration of around 279 hours. Each video is accompanied by 20 sentences describing all 6 designed attributes. Compared to CelebV [60], CelebV-Text has a larger scale and higher resolution. Although VoxCeleb2 [11] has more samples than CelebV-Text, its video distribution is limited as most videos are mainly talking faces. Moreover, video samples of both CelebV-HQ [66] and CelebV-Text are collected in open-world with diverse queries so that they are rich in distribution, while CelebV-Text has about 2 times video data, more video attributes, and highly relevant text descriptions. Finally, compared to the only existing facial text-video dataset MM-Vox [19], CelebV-Text overpasses MM-Vox in terms of scale and quality." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.747, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Attributes Distribution. In order to better present the distribution of different attributes in CelebV-Text, we pick and divide general appearance, action, and light direction attributes into groups. More distributions and division designs are provided in the Appendix. Specifically, all 40 general appearance classes are divided into 5 groups shown in Figure 3 (a). Facial features (e.g., double chin, big nose, and oval face) account for the most portion around \\(45\\%\\). The elementary group is twice large than the beard type, accounting for around \\(25\\%\\) and \\(12\\%\\), respectively. Fewer samples are located to the hairstyle and accessories groups," + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.085, + 0.693, + 0.181 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.545, + 0.183, + 0.682, + 0.191 + ], + "angle": 0, + "content": "(a) Number of word distribution" + }, + { + "type": "image", + "bbox": [ + 0.698, + 0.086, + 0.877, + 0.181 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.751, + 0.183, + 0.854, + 0.192 + ], + "angle": 0, + "content": "(b) 4-Grams distribution" + }, + { + "type": "image", + "bbox": [ + 0.527, + 0.195, + 0.588, + 0.242 + ], + "angle": 0, + "content": null + }, + { + "type": "image_footnote", + "bbox": [ + 0.59, + 0.197, + 0.865, + 0.243 + ], + "angle": 0, + "content": "This man with grey hair has a big nose and bags under eyes. He is wearing eyeglasses and earrings, with freckles above the eye areas. The man is firstly gazing for a short time and then he wags his head for a short time, and finally he keeps gazing for the rest of the time. He firstly remains neutral for a long time and then he turns angry for the rest of the time. The video is slightly dark in warm light. The light direction is front lighting for the whole time. CelebV-Text" + }, + { + "type": "image", + "bbox": [ + 0.528, + 0.248, + 0.59, + 0.294 + ], + "angle": 0, + "content": null + }, + { + "type": "image_footnote", + "bbox": [ + 0.601, + 0.258, + 0.835, + 0.283 + ], + "angle": 0, + "content": "a man has bags under eyes , receding hairline and straight hairhe is wearing goatee . he is chubby . he has beard , archedeyebrows and black hair ." + }, + { + "type": "image_caption", + "bbox": [ + 0.81, + 0.281, + 0.848, + 0.289 + ], + "angle": 0, + "content": "MM-Vox" + }, + { + "type": "image_caption", + "bbox": [ + 0.651, + 0.298, + 0.773, + 0.307 + ], + "angle": 0, + "content": "(c) Video and text examples" + }, + { + "type": "image_caption", + "bbox": [ + 0.5, + 0.309, + 0.89, + 0.335 + ], + "angle": 0, + "content": "Figure 5. Text distribution. CelebV-Text achieves better performance in both 4-gram and number words distribution." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.354, + 0.892, + 0.48 + ], + "angle": 0, + "content": "taking around \\(10\\%\\) and \\(8\\%\\), respectively. Besides, action attributes are divided into 5 groups in Figure 3 (b), where it is clear that head-related actions account for the largest portion of around \\(60\\%\\), followed by eyes-related actions of around \\(20\\%\\). The interaction group (e.g., eat), feeling group (e.g., smile), and daily group (e.g., sleep) account for around \\(9\\%\\), \\(7\\%\\), and \\(4\\%\\), respectively. Finally, for light directions (Figure 3 (c)), most samples contain the front lighting and the remaining ones are evenly distributed." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.48, + 0.892, + 0.648 + ], + "angle": 0, + "content": "Video Quality Distribution. We follow [66] to analyze the quality of our collected videos. To demonstrate the superiority of CelebV-Text, we compare with MM-Vox [19] and CelebV-HQ [66], where mean BRISQUE [42] and VSFA [33] are used to evaluate the image and video quality, respectively. Image quality of all datasets is shown in Figure 4 (a), where CelebV-Text and CelebV-HQ achieve comparable quality, higher than MM-Vox by a large margin. Video quality of all datasets is shown in Figure 4 (b), where CelebV-Text has the best quality, which is due to the effect of the video split method mentioned in Section 3.1, alleviating the discontinuity during background transitions." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.656, + 0.676, + 0.672 + ], + "angle": 0, + "content": "4.2. Text Comparisons" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.678, + 0.891, + 0.789 + ], + "angle": 0, + "content": "In addition to a large number of video samples, text descriptions of CelebV-Text are longer and more detailed than those in MM-Vox [19] and CelebV-HQ [66] (see Figure 5 (a)), where the average text length of MM-Vox, CelebV-HQ, and CelebV-Text are 28.39, 31.06, and 67.15. Distributions of Celeb-HQ and MM-Vox are close, but there are more words in CelebV-Text to describe a video due to the comprehensive annotation." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.789, + 0.892, + 0.901 + ], + "angle": 0, + "content": "To validate the linguistic diversity of the generated texts, comparisons are conducted among the three datasets following [56]. Specifically, we report the unique part-of-speech (POS) tags (i.e., verb, noun, adjective, and adverb) of the three datasets in Table 3. Obviously, due to our comprehensively designed attribute list and the number of templates, CelebV-Text presents a wider variety of text styles, covering a broader range of face attributes that are static or" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "14809" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.089, + 0.895, + 0.116 + ], + "angle": 0, + "content": "Table 2. Multimodal retrieval results. Clip2Video [17] is leveraged to measure the text-video relevance via retrieval experiments. Bold values indicate the best results, underlined ones indicate the second best." + }, + { + "type": "table", + "bbox": [ + 0.081, + 0.12, + 0.894, + 0.237 + ], + "angle": 0, + "content": "
Text ⇒ VideoVideo ⇒ Text
DescriptionDatasetR@1(↑)R@5(↑)R@10(↑)MdR(↓)MnR(↓)R@1(↑)R@5(↑)R@10(↑)MdR(↓)MnR(↓)
(a) App.MM-Vox [19]1.59.015.752.068.82.09.214.643.057.8
CelebV-HQ [66]5.919.229.727.052.27.220.732.427.046.9
CelebV-Text6.121.335.526.349.17.420.729.926.648.3
(b) App.+Emo.CelebV-HQ [66]6.520.130.825.048.07.925.538.817.037.0
CelebV-Text6.623.437.126.047.68.127.234.718.238.3
(c) App.+Emo.+Act.CelebV-Text6.924.139.225.846.78.027.637.116.736.1
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.249, + 0.47, + 0.276 + ], + "angle": 0, + "content": "Table 3. Number of unique POS tags. The numbers of unique POS tags for MM-Vox, CelebV-HQ, and CelebV-Text." + }, + { + "type": "table", + "bbox": [ + 0.113, + 0.279, + 0.438, + 0.343 + ], + "angle": 0, + "content": "
Dataset#Verb#Adj.#Noun#Adv.
MM-Vox [19]520380
CelebV-HQ [66]1024506
CelebV-Text967817424
" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.359, + 0.3, + 0.374 + ], + "angle": 0, + "content": "dynamic in the temporal domain." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.379, + 0.47, + 0.506 + ], + "angle": 0, + "content": "In addition, we further examine the naturalness and complexity of our texts compared to MM-Vox, where we modify [63] to calculate the type-token vocabulary curve for all captions. As shown in Figure 5 (b) where unique 4-grams are selected as the types [56], it is evident that due to our grammar structures and synonym replacement, the linguistic naturalness (vocabulary use) and complexity (vocabulary size) of our CelebV-Text are much better. Please refer to Appendix for more \\( n \\)-grams results." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.532, + 0.281, + 0.546 + ], + "angle": 0, + "content": "4.3. Text-Video Relevance" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.56, + 0.47, + 0.716 + ], + "angle": 0, + "content": "To quantitatively validate our text-video relevance, we conduct text-video retrieval tasks on three datasets: MMVox [19], CelebV-HQ [66], and CelebV-Text. Rather than use conventional frame-wise clip score as most works [24, 51, 55], we follow [17] to compute feature similarities between texts and videos with the consideration of temporal dynamics, which reflects accurate multimodal interactions across the two modalities. Recall at rank K (R@K), median rank (MdR), and mean rank (MnR) [17, 41, 64] are used as evaluation metrics, where the higher R@K, the lower median rank and mean rank indicate better performance." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.72, + 0.471, + 0.902 + ], + "angle": 0, + "content": "We first examine the performance given texts with descriptions of general appearance in Table 2 (a). Results of CelebV-HQ and CelebV-Text are both better than MM-Vox for two retrieval tasks, which indicates our designed templates can produce texts more relevant to videos than MM-Vox. We further add descriptions about dynamic emotion changes to CelebV-HQ and CelebV-Text in Table 2 (b). Similar results are achieved in both datasets, which reflects that our annotation accuracy on static appearance attributes is as good as CelebV-HQ. Finally, we append action descriptions to CelebV-Text in Table 2 (c), which achieves the best performance on most metrics, verifying the relevance between our generated texts and video samples." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.25, + 0.627, + 0.268 + ], + "angle": 0, + "content": "5. Experiment" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.275, + 0.894, + 0.332 + ], + "angle": 0, + "content": "In this section, we first conduct facial text-to-video generation to validate the effectiveness of our CelebV-Text dataset. We then benchmark representative approaches on facial text-to-video generation task." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.337, + 0.854, + 0.354 + ], + "angle": 0, + "content": "5.1. High-relevance Text-to-Video Generation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.36, + 0.892, + 0.596 + ], + "angle": 0, + "content": "To show the benefits brought by our text descriptions which depict both static and dynamic attributes, we conduct experiments to show the effectiveness of CelebV-Text. Experiments are mainly based on a recent open-sourced state-of-the-art method, MMVID [19], and compared with \\(\\mathrm{CogVideo}^2\\) [26], which is a large-scale pretrained text-to-video model, trained on millions of text-image/video pairs. Static Face Video Generation. To validate the effectiveness of our facial text-video dataset in static attributes, we use the models stated above to generate videos conditioned on general appearance, face details, and light conditions descriptions, respectively. Specifically, we first train MMVID [19] from scratch solely on CelebV-Text. We then generate 3 input texts including individual descriptions of each of the static attributes. Generated texts are fed into both MMVID [19] and \\(\\mathrm{CogVideo}\\) [26] and corresponding video outputs are examined." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.597, + 0.892, + 0.708 + ], + "angle": 0, + "content": "Visualization results of general appearance are shown in Figure 6 (a), which prove the effectiveness of our dataset. We observe that although CogVideo can output the face video given a text description, the text-video pair is not quite relevant, such as \"bags under eyes\" and \"wavy hair\". However, MMVID [19] produces videos with high relevance to input texts, containing all attributes described in the text. More results are shown in the Appendix." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.708, + 0.892, + 0.819 + ], + "angle": 0, + "content": "Dynamic Face Video Generation. We follow the above experimental setting to validate the effectiveness of our dataset with dynamic attribute changes (i.e., emotion, action and light direction). Due to the difficulty in modelling state change [27, 51], we follow [5] to apply test-time interpolation to MMVID [19], named MMVID-interp, to improve the text encoding and better understand the dynamics. Details of our modification are shown in the Appendix." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.819, + 0.894, + 0.848 + ], + "angle": 0, + "content": "In Figure 6 (b), we observe that CogVideo fails to reflect the temporal change described in the input text, i.e., smile" + }, + { + "type": "page_footnote", + "bbox": [ + 0.499, + 0.854, + 0.894, + 0.901 + ], + "angle": 0, + "content": "2We choose CogVideo [26] as the representative large-scale model for comparison, since the inference code and pretrained models of other large-scale methods (e.g., CogVideo [26], Phenaki [55],Imagen Video [24], and Make-A-Video [51]) are not public." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "14810" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.137, + 0.081, + 0.179, + 0.111 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.18, + 0.089, + 0.728, + 0.103 + ], + "angle": 0, + "content": "The man has black hair and bags under eyes. He has a beard, goatee and arched eyebrows." + }, + { + "type": "image", + "bbox": [ + 0.104, + 0.113, + 0.867, + 0.267 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.404, + 0.269, + 0.597, + 0.282 + ], + "angle": 0, + "content": "(a) Static - General Appearance" + }, + { + "type": "image", + "bbox": [ + 0.135, + 0.281, + 0.176, + 0.309 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.178, + 0.288, + 0.749, + 0.302 + ], + "angle": 0, + "content": "The young woman is wearing lipstick and earrings. She firstly smiles and then turns her head." + }, + { + "type": "image", + "bbox": [ + 0.102, + 0.31, + 0.865, + 0.535 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.435, + 0.537, + 0.566, + 0.549 + ], + "angle": 0, + "content": "(b) Dynamic - Action" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.55, + 0.89, + 0.575 + ], + "angle": 0, + "content": "Figure 6. Qualitative results of facial text-to-video generation. The generated samples are given texts describing (a) the static attribute and (b) dynamic attribute." + }, + { + "type": "table_caption", + "bbox": [ + 0.097, + 0.583, + 0.87, + 0.597 + ], + "angle": 0, + "content": "Table 4. Benchmark of text-to-video generation on different datasets. \\( \\downarrow \\) means a lower value is better and \\( \\uparrow \\) means the opposite." + }, + { + "type": "table_caption", + "bbox": [ + 0.114, + 0.598, + 0.425, + 0.609 + ], + "angle": 0, + "content": "(a) Quantitative results on general appearance descriptions." + }, + { + "type": "table", + "bbox": [ + 0.081, + 0.612, + 0.462, + 0.69 + ], + "angle": 0, + "content": "
DatasetMethodFVD(↓)FID(↓)CLIPSIM(↑)
MM-Vox [19]TFGAN [5]502.28 ± 1.66760.24 ± 16.010.165 ± 0.022
MMVID [19]65.79 ± 1.8138.81 ± 3.660.170 ± 0.020
CelebV-HQ [66]TFGAN [5]428.04 ± 1.76616.24 ± 17.450.168 ± 0.021
MMVID [19]73.65 ± 1.4363.86 ± 3.660.172 ± 0.019
CelebV-TextTFGAN [5]403.04 ± 1.34589.24 ± 16.460.177 ± 0.012
MMVID [19]66.69 ± 1.3558.70 ± 4.670.198 ± 0.014
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.7, + 0.468, + 0.813 + ], + "angle": 0, + "content": "\\(\\rightarrow\\) turn. However, both MMVID [19] and MMVID-interp trained on CelebV-Text can successfully model the dynamic attribute changes, which demonstrates the effectiveness of our dataset. In addition, we find that MMVID [19] cannot preserve some attributes well (e.g., earrings), while MMVID-interp can stabilize the sampling process, validating the effectiveness of our modification. More results are shown in the Appendix." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.816, + 0.468, + 0.901 + ], + "angle": 0, + "content": "Note that CogVideo [26] has a much larger model size (\\(\\sim 100\\) times larger than MMVID [19]) and is trained on much large text-video data (\\(\\sim 75\\) times larger than CelebV-Text). However, video samples produced by CogVideo [26] shown in Figure 6 are of a lower quality than the ones by MMVID [19] trained solely on CelebV-Text, where gener" + }, + { + "type": "table_caption", + "bbox": [ + 0.525, + 0.598, + 0.865, + 0.609 + ], + "angle": 0, + "content": "(b) Quantitative results on dynamic descriptions of CelebV-Text." + }, + { + "type": "table", + "bbox": [ + 0.507, + 0.612, + 0.885, + 0.69 + ], + "angle": 0, + "content": "
DatasetMethodFVD(↓)FID(↓)CLIPSIM(↑)
CelebV-TextTFGAN [5]442.30 ± 2.56623.17 ± 18.880.158 ± 0.024
App.+Emo.MMVID [19]82.78 ± 1.4761.58 ± 3.990.176 ± 0.008
MMVID-interp72.87 ± 1.2341.57 ± 3.560.182 ± 0.010
CelebV-TextTFGAN [5]571.34 ± 4.54784.93 ± 20.130.154 ± 0.028
App.+Act.MMVID [19]109.25 ± 2.1182.55 ± 4.370.174 ± 0.019
MMVID-interp80.81 ± 2.5570.88 ± 4.770.176 ± 0.020
" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.701, + 0.89, + 0.729 + ], + "angle": 0, + "content": "ated faces are not in a high relevance to input texts, demonstrating the effectiveness of our facial text-video dataset." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.735, + 0.891, + 0.75 + ], + "angle": 0, + "content": "5.2. Benchmark on Facial Text-to-Video Generation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.757, + 0.892, + 0.869 + ], + "angle": 0, + "content": "As the domain of text-to-video generation is currently thriving, there exists only one benchmark in the face domain, MM-Vox [19]. We expand [19] and construct a benchmark of facial text-to-video generation tasks on three datasets: MM-Vox [19], CelebV-HQ [66] with texts generated by our templates, and CelebV-Text. We choose two representative methods3, TFGAN [5] and MMVID [19], to evaluate their performances on all datasets." + }, + { + "type": "page_footnote", + "bbox": [ + 0.5, + 0.876, + 0.892, + 0.9 + ], + "angle": 0, + "content": "3Other methods, e.g., CogVideo [26], Phenaki [55], Imagen Video [24], and Make-A-Video [51] are not included since their training codes are not" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "14811" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.104, + 0.085, + 0.143, + 0.113 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.145, + 0.085, + 0.834, + 0.113 + ], + "angle": 0, + "content": "He has bushy eyebrows, beard and wavy hair. He has got 5 o'clock shadow and brown hair. He has bags under eyes and sideburns with mustache." + }, + { + "type": "image", + "bbox": [ + 0.086, + 0.115, + 0.852, + 0.34 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.342, + 0.892, + 0.369 + ], + "angle": 0, + "content": "Figure 7. Qualitative results on three facial text-video datasets. Red and yellow regions indicate the missing of \"bags under eyes\" and the existence of \"wavy hair\" and \"bags under eyes\"." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.38, + 0.47, + 0.686 + ], + "angle": 0, + "content": "Quantitative Results. For thorough benchmark construction, we evaluate baseline methods given variant texts including static and dynamic attributes. We use FVD [53] (temporal consistency), FID [23] (individual frame quality), and CLIPSIM [57] (text-video relevance) as evaluation metrics following [19] and report detailed results for appearance, action, and emotion in Table 4. Evaluation steps are repeated over ten runs with mean values and standard errors reported as well. All other values are shown in the Appendix. It can be seen from Table 4 that MMVID [19] obtains good FVD/FID/CLIPSIM metrics over TFGAN [5] which fails to generate reasonable video outputs. In addition, when input texts contain descriptions about a dynamic state change in the temporal domain, the generated video quality by MMVID [19] decreases, which encourages future methods to focus more on cross-modal understanding and consistent video generation. Moreover, the performance of MMVID-interp is better than MMVID [19] on all metrics, validating the effectiveness of our modification mentioned in Section 5.1. Due to challenges posed by our dataset and text-to-video generation task, there is still considerable room to improve." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.686, + 0.469, + 0.812 + ], + "angle": 0, + "content": "Qualitative Results. Video samples generated from MMVID [19] trained on different datasets are shown in Figure 7, where all video frames are of \\(128^2\\). We can see that video samples generated by MMVID [19] trained on different datasets are of high quality with temporal consistency. However, MMVID [19] trained on MM-Vox [19] can sometimes fail to generate attributes mentioned in the input texts. More generated video samples with dynamic attribute changes are shown in the Appendix." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.825, + 0.192, + 0.84 + ], + "angle": 0, + "content": "6. Discussion" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.849, + 0.47, + 0.879 + ], + "angle": 0, + "content": "We have proposed CelebV-Text, a large-scale, high-quality, and diverse facial text-video dataset with static and" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.38, + 0.892, + 0.534 + ], + "angle": 0, + "content": "dynamic attributes. CelebV-Text contains 70,000 video clips, each of which is accompanied by 20 individual sentences describing both static and dynamic factors. Through extensive statistical analysis and experiments, we have demonstrated the superiority and effectiveness of CelebV-Text. In the future, we plan to further enlarge CelebV-Text in both scale and diversity. We may further explore several new tasks based on CelebV-Text, such as fine-grained control of video face, adaptation of general pretrained models to the face domain, and text-driven 3D-aware facial video generation." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.542, + 0.893, + 0.807 + ], + "angle": 0, + "content": "Ethical Consideration. CelebV-Text is intended for research purposes only. While the raw videos will not be released, the data annotations, links to raw videos, and data processing tools will be made available after undergoing a rigorous legality check procedure at our institution. It is worth noting that our data annotation does not include any personal biometric information such as identity. Only generic attribute information such as gender, hair color, and motion is annotated. Additionally, synthetic videos generated in this work do not exhibit bias or certain biometric information (e.g., big lips or big nose), alleviating ethical concerns. CelebV-Text may be used for deepfakes, but it can also be used for forgery detection tasks to prevent such issues. We will try our best to control the application and acquisition procedure of CelebV-Text to avoid potential misuse and abuse. In the future, we plan to use synthetic face generation frameworks to generate synthetic face videos to address the ethical shortcomings of existing real-world face video datasets." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.816, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Acknowledgement. CelebV-Text is developed under OpenXDLab - an open platform for X-Dimension high-quality data. This study is supported by the RIE2020 Industry Alignment Fund Industry Collaboration Projects (IAF-ICP) Funding Initiative, in-kind contribution from the industry partner(s), and the MOE AcRF Tier 1 (RG16/21)." + }, + { + "type": "page_footnote", + "bbox": [ + 0.078, + 0.888, + 0.15, + 0.9 + ], + "angle": 0, + "content": "public so far." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "14812" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.105 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.468, + 0.152 + ], + "angle": 0, + "content": "[1] Niki Aifanti, Christos Papachristou, and Anastasios Delopoulos. The mug facial expression database. In WIAMIS, 2010. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.154, + 0.468, + 0.204 + ], + "angle": 0, + "content": "[2] Jean-Baptiste Alayrac, Piotr Bojanowski, Nishant Agrawal, Josef Sivic, Ivan Laptev, and Simon Lacoste-Julien. Unsupervised learning from narrated instruction videos. In CVPR, 2016. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.206, + 0.468, + 0.255 + ], + "angle": 0, + "content": "[3] Lisa Anne Hendricks, Oliver Wang, Eli Shechtman, Josef Sivic, Trevor Darrell, and Bryan Russell. Localizing moments in video with natural language. In ICCV, 2017. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.258, + 0.468, + 0.296 + ], + "angle": 0, + "content": "[4] Max Bain, Arsha Nagrani, Gül Varol, and Andrew Zisserman. Frozen in time: A joint video and image encoder for end-to-end retrieval. In ICCV, 2021. 1, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.298, + 0.468, + 0.348 + ], + "angle": 0, + "content": "[5] Yogesh Balaji, Martin Renqiang Min, Bing Bai, Rama Chellappa, and Hans Peter Graf. Conditional gan with discriminative filter generation for text-to-video synthesis. In IJCAI, 2019. 1, 2, 4, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.35, + 0.468, + 0.387 + ], + "angle": 0, + "content": "[6] Alex Bewley, Zongyuan Ge, Lionel Ott, Fabio Ramos, and Ben Upcroft. Simple online and realtime tracking. In ICIP. IEEE, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.389, + 0.468, + 0.426 + ], + "angle": 0, + "content": "[7] Sergey Bezryadin, Pavel Bourov, and Dmitry Ilinih. Brightness calculation in digital image processing. In TDPF, 2007. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.429, + 0.468, + 0.467 + ], + "angle": 0, + "content": "[8] Steven Bird, Ewan Klein, and Edward Loper. Natural language processing with Python: analyzing text with the natural language toolkit. \"O'Reilly Media, Inc.\", 2009. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.468, + 0.468, + 0.494 + ], + "angle": 0, + "content": "[9] David Chen and William B Dolan. Collecting highly parallel data for paraphrase evaluation. In ACL, 2011. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.495, + 0.468, + 0.532 + ], + "angle": 0, + "content": "[10] Danqi Chen and Christopher D Manning. A fast and accurate dependency parser using neural networks. In EMNLP, 2014. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.534, + 0.468, + 0.572 + ], + "angle": 0, + "content": "[11] Joon Son Chung, Arsha Nagrani, and Andrew Zisserman. Voxceleb2: Deep speaker recognition. In INTERSPEECH, 2018. 1, 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.573, + 0.468, + 0.611 + ], + "angle": 0, + "content": "[12] Karl Cobbe, Oleg Klimov, Chris Hesse, Taehoon Kim, and John Schulman. Quantifying generalization in reinforcement learning. In ICML, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.613, + 0.468, + 0.676 + ], + "angle": 0, + "content": "[13] Dima Damen, Hazel Doughty, Giovanni Maria Farinella, Antonino Furnari, Evangelos Kazakos, Jian Ma, Davide Moltisanti, Jonathan Munro, Toby Perrett, Will Price, et al. Rescaling egocentric vision: collection, pipeline and challenges for epic-kitchens-100. In IJCV, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.678, + 0.468, + 0.716 + ], + "angle": 0, + "content": "[14] Jiankang Deng, Jia Guo, Niannan Xue, and Stefanos Zafeiriou. Arcface: Additive angular margin loss for deep face recognition. In CVPR, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.717, + 0.468, + 0.755 + ], + "angle": 0, + "content": "[15] Kangle Deng, Tianyi Fei, Xin Huang, and Yuxin Peng. IRCAN: Introspective recurrent convolutional gan for text-to-video generation. In IJCAI, pages 2216-2222, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.757, + 0.468, + 0.793 + ], + "angle": 0, + "content": "[16] Patrick Esser, Robin Rombach, and Bjorn Ommer. Taming transformers for high-resolution image synthesis. In CVPR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.796, + 0.468, + 0.834 + ], + "angle": 0, + "content": "[17] Han Fang, Pengfei Xiong, Luhui Xu, and Yu Chen. Clip2video: Mastering video-text retrieval via image clip. arXiv preprint arXiv:2106.11097, 2021. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.836, + 0.468, + 0.873 + ], + "angle": 0, + "content": "[18] Tanmay Gupta, Dustin Schwenk, Ali Farhadi, Derek Hoiem, and Aniruddha Kembhavi. Imagine this! scripts to compositions to videos. In ECCV, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.874, + 0.468, + 0.9 + ], + "angle": 0, + "content": "[19] Ligong Han, Jian Ren, Hsin-Ying Lee, Francesco Barbieri, Kyle Olszewski, Shervin Minaee, Dimitris Metaxas, and" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.468, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.892, + 0.131 + ], + "angle": 0, + "content": "Sergey Tulyakov. Show me what and tell me how: Video synthesis via multimodal conditioning. In CVPR, 2022. 2, 3, 4, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.133, + 0.892, + 0.17 + ], + "angle": 0, + "content": "[20] William Harvey, Saeid Naderiparizi, Vaden Masrani, Christian Weilbach, and Frank Wood. Flexible diffusion modeling of long videos. arXiv preprint arXiv:2205.11495, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.172, + 0.892, + 0.234 + ], + "angle": 0, + "content": "[21] Thomas Hayes, Songyang Zhang, Xi Yin, Guan Pang, Sasha Sheng, Harry Yang, Songwei Ge, Isabelle Hu, and Devi Parikh. Mugen: A playground for video-audio-text multimodal understanding and generation. arXiv preprint arXiv:2204.08058, 2022. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.236, + 0.892, + 0.287 + ], + "angle": 0, + "content": "[22] Javier Hernandez-Andres, Raymond L Lee, and Javier Romero. Calculating correlated color temperatures across the entire gamut of daylight and skylight chromaticities. In Applied optics, 1999. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.288, + 0.892, + 0.338 + ], + "angle": 0, + "content": "[23] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. NeurIPS, 2017. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.34, + 0.892, + 0.403 + ], + "angle": 0, + "content": "[24] Jonathan Ho, William Chan, Chitwan Sahara, Jay Whang, Ruiqi Gao, Alexey Gritsenko, Diederik P Kingma, Ben Poole, Mohammad Norouzi, David J Fleet, et al. Imagen video: High definition video generation with diffusion models. arXiv preprint arXiv:2210.02303, 2022. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.405, + 0.892, + 0.443 + ], + "angle": 0, + "content": "[25] Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan, Mohammad Norouzi, and David J Fleet. Video diffusion models. arXiv preprint arXiv:2204.03458, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.445, + 0.892, + 0.494 + ], + "angle": 0, + "content": "[26] Wenyi Hong, Ming Ding, Wendi Zheng, Xinghan Liu, and Jie Tang. Cogvideo: Large-scale pretraining for text-to-video generation via transformers. arXiv preprint arXiv:2205.15868, 2022. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.496, + 0.892, + 0.534 + ], + "angle": 0, + "content": "[27] Yaosi Hu, Chong Luo, and Zhenzhong Chen. Make it move: controllable image-to-video generation with text descriptions. In CVPR, 2022. 2, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.536, + 0.892, + 0.598 + ], + "angle": 0, + "content": "[28] Justin Johnson, Bharath Hariharan, Laurens Van Der Maaten, Li Fei-Fei, C Lawrence Zitnick, and Ross Girshick. Clevr: A diagnostic dataset for compositional language and elementary visual reasoning. In CVPR, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.6, + 0.892, + 0.638 + ], + "angle": 0, + "content": "[29] Peter Kán and Hannes Kafumann. Deep light: light source estimation for augmented reality using deep learning. The Visual Computer, 2019. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.639, + 0.892, + 0.677 + ], + "angle": 0, + "content": "[30] Doyeon Kim, Donggyu Joo, and Junmo Kim. Tivgan: Text to image to video generation with step-by-step evolutionary generator. In IEEE Access, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.679, + 0.892, + 0.704 + ], + "angle": 0, + "content": "[31] Dan Klein and Christopher D Manning. Accurate unlexicalized parsing. In ACL, 2003. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.705, + 0.892, + 0.743 + ], + "angle": 0, + "content": "[32] Ranjay Krishna, Kenji Hata, Frederic Ren, Li Fei-Fei, and Juan Carlos Niebles. Dense-captioning events in videos. In ICCV, 2017. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.745, + 0.892, + 0.782 + ], + "angle": 0, + "content": "[33] Dingquan Li, Tingting Jiang, and Ming Jiang. Quality assessment of in-the-wild videos. In ACM MM, pages 2351-2359, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.784, + 0.892, + 0.822 + ], + "angle": 0, + "content": "[34] Yitong Li, Martin Min, Dinghan Shen, David Carlson, and Lawrence Carin. Video generation from text. In Proceedings of the AAAI conference on artificial intelligence, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.823, + 0.892, + 0.873 + ], + "angle": 0, + "content": "[35] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dálár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In ECCV, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.874, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[36] Wen Liu, Zhixin Piao, Jie Min, Wenhan Luo, Lin Ma, and Shenghua Gao. Liquid warping gan: A unified framework" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.956 + ], + "angle": 0, + "content": "14813" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.093, + 0.469, + 0.119 + ], + "angle": 0, + "content": "for human motion imitation, appearance transfer and novel view synthesis. In ICCV, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.121, + 0.469, + 0.159 + ], + "angle": 0, + "content": "[37] Yue Liu, Xin Wang, Yitian Yuan, and Wenwu Zhu. Cross-modal dual learning for sentence-to-video generation. In ACM MM, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.161, + 0.468, + 0.198 + ], + "angle": 0, + "content": "[38] Ziwei Liu, Ping Luo, Xiaogang Wang, and Xiaou Tang. Deep learning face attributes in the wild. In ICCV, 2015. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.201, + 0.469, + 0.251 + ], + "angle": 0, + "content": "[39] Jonathan Malmaud, Jonathan Huang, Vivek Rathod, Nick Johnston, Andrew Rabinovich, and Kevin Murphy. What's cookin'? interpreting cooking videos using text, speech and vision. In ACL, 2015. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.254, + 0.469, + 0.305 + ], + "angle": 0, + "content": "[40] Antoine Miech, Dimitri Zhukov, Jean-Baptiste Alayrac, Makarand Tapaswi, Ivan Laptev, and Josef Sivic. Howto100m: Learning a text-video embedding by watching hundred million narrated video clips. In CVPR, 2019. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.307, + 0.469, + 0.357 + ], + "angle": 0, + "content": "[41] Niluthpol Chowdhury Mithun, Juncheng Li, Florian Metze, and Amit K Roy-Chowdhury. Learning joint embedding with multimodal cues for cross-modal video-text retrieval. In ICMR, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.36, + 0.469, + 0.397 + ], + "angle": 0, + "content": "[42] Anish Mittal, Anush Krishna Moorthy, and Alan Conrad Bovik. No-reference image quality assessment in the spatial domain. In TIP, 2012. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.4, + 0.469, + 0.438 + ], + "angle": 0, + "content": "[43] Gaurav Mittal, Tanya Marwah, and Vineeth N Balasubramanian. Sync画: Automatic video generation using deep recurrent attentive architectures. In ACM MM, 2017. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.44, + 0.469, + 0.478 + ], + "angle": 0, + "content": "[44] Ali Mollahosseini, Behzad Hasani, and Mohammad H Ma-hoor. Affectnet: A database for facial expression, valence, and arousal computing in the wild. In TAC, 2017. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.48, + 0.469, + 0.518 + ], + "angle": 0, + "content": "[45] A. Nagrani, J. S. Chung, and A. Zisserman. Voxceleb: a large-scale speaker identification dataset. In INTER-SPEECH, 2017. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.521, + 0.469, + 0.558 + ], + "angle": 0, + "content": "[46] Yingwei Pan, Zhaofan Qiu, Ting Yao, Houqiang Li, and Tao Mei. To create what you tell: Generating videos from captions. In ACM MM, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.561, + 0.469, + 0.611 + ], + "angle": 0, + "content": "[47] MarcAurelio Ranzato, Arthur Szlam, Joan Bruna, Michael Mathieu, Ronan Collobert, and Sumit Chopra. Video (language) modeling: a baseline for generative models of natural videos. arXiv preprint arXiv:1412.6604, 2014. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.613, + 0.469, + 0.664 + ], + "angle": 0, + "content": "[48] Andreas Rössler, Davide Cozzolino, Luisa Verdoliva, Christian Riess, Justus Thies, and Matthias Nießner. Faceforensics: A large-scale video dataset for forgery detection in human faces. arXiv preprint arXiv:1803.09179, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.666, + 0.469, + 0.729 + ], + "angle": 0, + "content": "[49] Christoph Schuhmann, Richard Vencu, Romain Beaumont, Robert Kaczmarczyk, Clayton Mullis, Aarush Katta, Theo Coombes, Jenia Jitsev, and Aran Komatsuzaki. Laion-400m: Open dataset of clip-filtered 400 million image-text pairs. arXiv preprint arXiv:2111.02114, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.732, + 0.469, + 0.781 + ], + "angle": 0, + "content": "[50] Piyush Sharma, Nan Ding, Sebastian Goodman, and Radu Soricut. Conceptual captions: A cleaned, hypernymed, image alt-text dataset for automatic image captioning. In ACL, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.784, + 0.469, + 0.847 + ], + "angle": 0, + "content": "[51] Uriel Singer, Adam Polyak, Thomas Hayes, Xi Yin, Jie An, Songyang Zhang, Qiyuan Hu, Harry Yang, Oron Ashual, Oran Gafni, et al. Make-a-video: Text-to-video generation without text-video data. arXiv preprint arXiv:2209.14792, 2022. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.85, + 0.469, + 0.899 + ], + "angle": 0, + "content": "[52] David Stap, Maurits Bleeker, Sarah Ibrahimi, and Maartje ter Hoeve. Conditional image generation and manipulation for user-specified content. arXiv preprint arXiv:2005.04909, 2020. 2, 4" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.093, + 0.469, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.145 + ], + "angle": 0, + "content": "[53] Thomas Unterthiner, Sjoerd van Steenkiste, Karol Kurach, Raphael Marinier, Marcin Michalski, and Sylvain Gelly. Towards accurate generative models of video: A new metric & challenges. arXiv preprint arXiv:1812.01717, 2018. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.146, + 0.892, + 0.171 + ], + "angle": 0, + "content": "[54] Aaron Van Den Oord, Oriol Vinyals, et al. Neural discrete representation learning. NeurIPS, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.173, + 0.892, + 0.248 + ], + "angle": 0, + "content": "[55] Ruben Villegas, Mohammad Babaeizadeh, Pieter-Jan Kindermans, Hernan Moraldo, Han Zhang, Mohammad Taghi Saffar, Santiago Castro, Julius Kunze, and Dumitru Erhan. Phenaki: Variable length video generation from open domain textual description. arXiv preprint arXiv:2210.02399, 2022. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.251, + 0.892, + 0.3 + ], + "angle": 0, + "content": "[56] Xin Wang, Jiawei Wu, Junkun Chen, Lei Li, Yuan-Fang Wang, and William Yang Wang. Vatex: A large-scale, high-quality multilingual dataset for video-and-language research. In ICCV, 2019. 4, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.302, + 0.892, + 0.353 + ], + "angle": 0, + "content": "[57] Chenfei Wu, Lun Huang, Qianxi Zhang, Binyang Li, Lei Ji, Fan Yang, Guillermo Sapiro, and Nan Duan. Godiva: Generating open-domain videos from natural descriptions. arXiv preprint arXiv:2104.14806, 2021. 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.355, + 0.892, + 0.417 + ], + "angle": 0, + "content": "[58] Chenfei Wu, Jian Liang, Xiaowei Hu, Zhe Gan, Jianfeng Wang, Lijuan Wang, Zicheng Liu, Yuejian Fang, and Nan Duan. Nuwa-infinity: Autoregressive over autoregressive generation for infinite visual synthesis. In NeurIPS, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.42, + 0.892, + 0.468 + ], + "angle": 0, + "content": "[59] Chenfei Wu, Jian Liang, Lei Ji, Fan Yang, Yuejian Fang, Daxin Jiang, and Nan Duan. N\\`uwa: Visual synthesis pretraining for neural visual world creation. In ECCV, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.472, + 0.892, + 0.51 + ], + "angle": 0, + "content": "[60] Wayne Wu, Yunxuan Zhang, Cheng Li, Chen Qian, and Chen Change Loy. Reenactgan: Learning to reenact faces via boundary transfer. In ECCV, 2018. 1, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.511, + 0.892, + 0.55 + ], + "angle": 0, + "content": "[61] Weihao Xia, Yujiu Yang, Jing-Hao Xue, and Baoyuan Wu. Tedigan: Text-guided diverse face image generation and manipulation. In CVPR, 2021. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.551, + 0.892, + 0.588 + ], + "angle": 0, + "content": "[62] Jun Xu, Tao Mei, Ting Yao, and Yong Rui. Msr-vtt: A large video description dataset for bridging video and language. In CVPR, 2016. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.59, + 0.892, + 0.616 + ], + "angle": 0, + "content": "[63] Gilbert Youmans. Measuring lexical style and competence: The type-token vocabulary curve. Style, 1990. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.617, + 0.892, + 0.642 + ], + "angle": 0, + "content": "[64] Bowen Zhang, Hexiang Hu, and Fei Sha. Cross-modal and hierarchical modeling of video and text. In ECCV, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.643, + 0.892, + 0.681 + ], + "angle": 0, + "content": "[65] Luowei Zhou, Chenliang Xu, and Jason J Corso. Towards automatic learning of procedures from web instructional videos. In AAAI, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.683, + 0.892, + 0.734 + ], + "angle": 0, + "content": "[66] Hao Zhu, Wayne Wu, Wentao Zhu, Liming Jiang, Siwei Tang, Li Zhang, Ziwei Liu, and Chen Change Loy. CelebV-HQ: A large-scale video facial attributes dataset. In ECCV, 2022. 2, 3, 4, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.736, + 0.892, + 0.786 + ], + "angle": 0, + "content": "[67] Dimitri Zhukov, Jean-Baptiste Alayrac, Ramadan Gokberk Cinbis, David Fouhey, Ivan Laptev, and Josef Sivic. Cross-task weakly supervised learning from instructional videos. In CVPR, 2019. 1, 4" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.786 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "14814" + } + ] +] \ No newline at end of file diff --git a/2023/CelebV-Text_ A Large-Scale Facial Text-Video Dataset/b3bc8119-1bc3-419d-89b1-046a6b2d92d9_origin.pdf b/2023/CelebV-Text_ A Large-Scale Facial Text-Video Dataset/b3bc8119-1bc3-419d-89b1-046a6b2d92d9_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..07cd090c1a89955a0346f047514cd5392cc0b624 --- /dev/null +++ b/2023/CelebV-Text_ A Large-Scale Facial Text-Video Dataset/b3bc8119-1bc3-419d-89b1-046a6b2d92d9_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5da1f6e2c375afc9561ea9ec317a6207ef28e7b857e484b47e973948c82c12d6 +size 3143544 diff --git a/2023/CelebV-Text_ A Large-Scale Facial Text-Video Dataset/full.md b/2023/CelebV-Text_ A Large-Scale Facial Text-Video Dataset/full.md new file mode 100644 index 0000000000000000000000000000000000000000..9ba47131c5e65f4638ed313a110393767f92f53f --- /dev/null +++ b/2023/CelebV-Text_ A Large-Scale Facial Text-Video Dataset/full.md @@ -0,0 +1,308 @@ +# CelebV-Text: A Large-Scale Facial Text-Video Dataset + +Jianhui $\mathrm{Yu}^{1*}$ Hao Zhu $^{2*}$ Liming Jiang $^{3}$ Chen Change Loy $^{3}$ Weidong Cai $^{1}$ Wayne Wu $^{4}$ + +$^{1}$ University of Sydney $^{2}$ SenseTime Research $^{3}$ S-Lab, Nanyang Technological University $^{4}$ Shanghai AI Laboratory + +jianhui.yu@sydney.edu.au haozhu96@gmail.com {liming002,ccloy}@ntu.edu.sg + +tom.cai@sydney.edu.au wuwenyan0503@gmail.com + +![](images/f0109e28e042bfb7ecc4716cf4e58dae3e14efdf34943c697065303fe469de7e.jpg) +(a) Video Samples (left: general appearance, middle: detailed appearance, right: light conditions) +Figure 1. Overview of CelebV-Text. CelebV-Text contains (a) 70,000 video samples and (b) 1,400,000 text descriptions. Each video sample is annotated with general appearance, detailed appearance, light conditions, action, emotion, and light directions. + +![](images/fcf8b3338799a55aa4546a38e787d33d2e42b686dbba211adb062ffd7cec9900.jpg) +(b) Text Descriptions + +# Abstract + +Text-driven generation models are flourishing in video generation and editing. However, face-centric text-to-video generation remains a challenge due to the lack of a suitable dataset containing high-quality videos and highly relevant texts. This paper presents CelebV-Text, a large-scale, diverse, and high-quality dataset of facial text-video pairs, to facilitate research on facial text-to-video generation tasks. CelebV-Text comprises 70,000 in-the-wild face video clips with diverse visual content, each paired with 20 texts generated using the proposed semi-automatic text generation strategy. The provided texts are of high quality, describing both static and dynamic attributes precisely. The superiority of CelebV-Text over other datasets is demonstrated via comprehensive statistical analysis of the videos, texts, and text-video relevance. The effectiveness and potential of CelebV-Text are further shown through extensive self-evaluation. A benchmark is constructed with representative methods to standardize the evaluation of the facial text-to-video generation task. All data and models are publicly available1. + +# 1. Introduction + +Text-driven video generation has recently garnered significant attention in the fields of computer vision and computer graphics. By using text as input, video content can be generated and controlled, inspiring numerous applications in both academia and industry [5,34,43,47]. However, text-to-video generation still faces many challenges, particularly in the face-centric scenario where generated video frames often lack quality [18, 34, 37] or have weak relevance to input texts [2,4,39,67]. We believe that one of the main issues is the absence of a well-suited facial text-video dataset containing high-quality video samples and text descriptions of various attributes highly relevant to videos. + +Constructing a high-quality facial text-video dataset poses several challenges, mainly in three aspects. 1) Data collection. The quality and quantity of video samples largely determine the quality of generated videos [11, 45, 48, 60]. However, obtaining such a large-scale dataset with high-quality samples while maintaining a natural distribution and smooth video motion is challenging. 2) Data annotation. The relevance of text-video pairs needs to be ensured. This requires a comprehensive coverage of text for describing the content and motion appearing in the video, such as light conditions and head movements. 3) Text gen + +eration. Producing diverse and natural texts are non-trivial. Manual text generation is expensive and not scalable. While auto-text generation is easily extensible, it is limited in naturalness. + +To overcome the challenges mentioned above, we carefully design a comprehensive data construction pipeline that includes data collection and processing, data annotation, and semi-auto text generation. First, to obtain raw videos, we follow the data collection steps of CelebV-HQ, which has proven to be effective in [66]. We introduce a minor modification to the video processing step to improve the video's smoothness further. Next, to ensure highly relevant text-video pairs, we analyze videos from both temporal dynamics and static content and establish a set of attributes that may or may not change over time. Finally, we propose a semi-auto template-based method to generate texts that are diverse and natural. Our approach leverages the advantages of both auto- and manual-text methods. Specifically, we design a rich variety of grammar templates as [10,52] to parse annotation and manual texts, which are flexibly combined and modified to achieve high diversity, complexity, and naturalness. + +With the proposed pipeline, we create CelebV-Text, a Large-Scale Facial Text-Video Dataset, which includes 70,000 in-the-wild video clips with a resolution of at least $512 \times 512$ and 1,400,000 text descriptions with 20 for each clip. As depicted in Figure 1, CelebV-Text consists of high-quality video samples and text descriptions for realistic face video generation. Each video is annotated with three types of static attributes (40 general appearances, 5 detailed appearances, and 6 light conditions) and three types of dynamic attributes (37 actions, 8 emotions, and 6 light directions). All dynamic attributes are densely annotated with start and end timestamps, while manual-texts are provided for labels that cannot be discretized. Furthermore, we have designed three templates for each attribute type, resulting in a total of 18 templates that can be flexibly combined. All attributes and manual-texts are naturally described in our generated texts. + +CelebV-Text surpasses existing face video datasets [11] in terms of resolution (over 2 times higher), number of samples, and more diverse distribution. In addition, the texts in CelebV-Text exhibit higher diversity, complexity, and naturalness than those in text-video datasets [19, 66]. CelebV-Text also shows high relevance of text-video pairs, validated by our text-video retrieval experiments [17]. To further examine the effectiveness and potential of CelebV-Text, we evaluate it on a representative baseline [19] for facial text-to-video generation. Our results show better relevance between generated face videos and texts when compared to a state-of-the-art large-scale pretrained model [26]. Furthermore, we show that a simple modification of [19] with text interpolation can significantly improve temporal coherence. Finally, we present a new benchmark for text-to-video generation to standardize the facial text-to-video generation task, which includes representative models [5, 19] on three text-video datasets. + +The main contributions of this work are summarized as + +follows: 1) We propose CelebV-Text, the first large-scale facial text-video dataset with high-quality videos, as well as rich and highly-relevant texts, to facilitate research in facial text-to-video generation. 2) Comprehensive statistical analyses are conducted to examine video/text quality and diversity, as well as text-video relevance, demonstrating the superiority of CelebV-Text. 3) A series of self-evaluations are performed to demonstrate the effectiveness and potential of CelebV-Text. 4) A new benchmark for text-to-video generation is constructed to promote the standardization of the facial text-to-video generation task. + +# 2. Related Work + +Text-to-Video Generation. Text-driven video generation, which involves generating videos from text descriptions, has recently gained significant interest as a challenging task. Mittal et al. [43] first introduced this task to generate semantically consistent videos conditioned on encoded captions. Other studies, such as [5, 15, 46], attempt to generate video samples conditioned on encoded text inputs. However, due to the low richness of text descriptions and the small number of data samples, the generated video samples are often at low resolution or lack relevance with the input texts. More recently, several works [19, 26, 27, 55, 57–59] have employed discrete latent codes [16, 54] for more realistic video generation. Some of these works treat videos as a sequence of independent images [19, 26, 58, 59], while Phenaki [55] considers temporal relations between each frame for a more robust video decoding process. Another branch of studies leverage diffusion models for text-to-video generation [20, 24, 25, 51], which require millions or billions of samples to achieve high-quality generation. While text-to-video generation methods are rapidly evolving, they are generally designed for generating general videos. Among these methods, only MMVID [19] has conducted specific experiments with face-centric descriptions. One possible reason for this is that facial text-to-video generation requires more accurate and detailed text descriptions than general tasks. However, there is currently no suitable dataset available that provides such properties for face-centric text-to-video generation. + +Multimodal Datasets. Existing multimodal datasets can be categorized into two classes: open-world and closed-world. Open-world datasets [3,9,13,32,35,36,40,43,49,50,62,65] are widely used for text-to-image/video generation tasks. Some of them have manual annotations [13, 32, 35, 50, 62] and part of them are directly collected from the Internet, such as subtitles [40, 49]. Closed-world datasets are mostly composed of images or videos collected in constrained environment with corresponding information such as text. CLEVR [28] is a synthetic text-image dataset produced by arranging 3D objects with different shapes under a controlled background. While MUGEN [21] is a video-audiotext dataset that was collected using CoinRun [12] by introducing audio and new interactions. The corresponding text is produced by human annotators and grammar templates. + +Multimodal face datasets also exist. Modified MUG [1] + +Table 1. In-the-wild face video dataset comparison. The symbol “#” indicates the number. The abbreviations “Res.”, “Dura.”, “App.”, “Cond.”, “Act.”, “Emo.”, and “Dir.” stand for Resolution, Duration, Appearance, Condition, Action, Emotion, and Direction, respectively. The “half checkmark” denotes that CelebV-HQ consists of action attributes with no timestamp. + +
DatasetsMeta InformationAttribute LabelsText
StaticDynamic
#SamplesRes.Dura.General App.Detail App.Light Cond.Act.Emo.Light Dir.Auto Manual
CelebV [60]5256×2562hrs
VoxCeleb2 [11]150,480224×2242442hrs
CelebV-HQ [66]35,666512×51268hrs✓XX✓✓XX
MM-Vox [19]19,522224×224323hrs✓XXXXX✓
CelebV-Text70,000512×512+279hrs✓✓✓✓✓✓✓
+ +is a closed-world text-video dataset that contains 1,039 videos with subjects showing different emotions, where the text descriptions are generated from facial emotions using a fixed template [30]. MM-Vox [19] contains 19,522 face videos from VoxCeleb [45], with 36 facial attributes manually labeled following CelebA [38] and text descriptions generated via Probabilistic Context-Free Grammar (PCFG) [61]. However, both datasets only contain language descriptions related to static facial attributes without considering the temporal state change (i.e., emotion or action) presented in the original face videos. Moreover, the limited label annotations restrict the diversity of the text descriptions, making them sub-optimal for studying the text-to-video generation task on the face domain. CelebV-HQ [66] is the latest high-quality face video dataset that covers facial annotations, including appearance, movement, and emotion. However, it only provides discrete labels and timestamps, with no text descriptions. + +# 3. Dataset Construction + +In this work, we aim to build a facial text-video dataset, which requires not only large-scale video samples of high quality, but also natural and diverse text descriptions that are highly relevant to videos. To achieve so, we propose an efficient pipeline, as shown in Figure 2, to construct CelebV-Text, including Data Collection & Processing, Data Annotation, and Semi-auto Text Generation. + +# 3.1. Data Collection & Processing + +Collection. We follow the same strategy as CelebV-HQ [66] due to its effectiveness in large-scale high-quality data collection. Specifically, we firstly generate a large number of queries, including human names, movie titles, vlogs and so on, to retrieve videos that contain human faces with temporally dynamic state changes and abundant facial attributes. Our data are collected from open world with videos downloaded from online resources. Videos with low resolution $(< 512^{2})$ , low time duration $(< 5s)$ , and having appeared in CelebV-HQ are filtered out. + +Processing. To sample high-quality and diverse video clips from our raw collections, similar steps are followed as CelebV-HQ [66] with modifications. We first filter out video clips with bounding box regions less than $512^{2}$ rather + +![](images/eedc5f3f0b2e60ba783324e68f63307e42cfcdf071714f39f4be781746fcf053.jpg) +Figure 2. Pipeline of our dataset construction process. The pipeline includes data collection & processing, data annotation, and semi-auto text generation. + +than resize them to the same resolution. In this way, clips are not upsampled or downsampled hence the video quality would not be affected, which leads to various resolutions of collected videos: $56.4\%$ with $512^2 \sim 1024^2$ , and $43.6\%$ for $1024^2+$ . To reduce the face area noise when the background changes, we further change the video splitting strategy. In addition to our focus on the same human motion [6] and identity [14] present in adjacent frames, we split the video into different clips when the background changes by a toolkit1. + +# 3.2. Data Annotation + +The annotation process is a core part in CelebV-Text construction, which would greatly affect the relevance of text-video pairs, as our designed text templates heavily depend on the annotation results. Here, we first describe how we design attributes, and then give details about the annotation strategy for face videos. + +Attributes Design. Temporal dynamic is the key difference between images and videos. However, as shown in Table 1, most face video datasets focus on static attributes where attribute information does not change over time, such as appearance. Dynamic attributes that change over time, such as emotion and face actions, are often neglected. In the fol + +![](images/a913896e83a99c179f17204cbfab4c4efac3155f5fef60f086907c9c4feecf3c.jpg) +Figure 3. Dataset distribution comparison. The distributions of appearance attributes, action attributes, and light directions. + +![](images/28c5acb5920de6e0ddd8f704943dcbbe09d28a7eced1a6c54706017cf64a6290.jpg) + +![](images/744e07655ba3d485ce2a50ff5b2975ac45985578192d20caa4c48711963ed744.jpg) + +lowing, we decouple face videos into static and dynamic categories and details are given as follows. + +1) Static. The current dataset [19] only considers static information such as the appearance attribute, which includes 40 classes as CelebA [38]. In contrast, we define static information to include three types of attributes: general appearance, detailed appearance, and light conditions. General appearance attributes follow the same definition as CelebA [38]. Detailed appearance attributes including five classes are proposed for realistic face generation, i.e., scar, mole, freckle, dimple, and one-eyed. We define light conditions in a restricted manner to include light color temperature [22] and brightness [7], with a total of 6 classes. + +2) Dynamic. Here, we design three dynamic attributes, i.e., action, emotion, and light directions. For action attributes, we follow CelebV-HQ [66] and expand their action list by two classes, i.e., squint and blink. For emotion attributes, we select the 8 emotion setting in Affectnet [44], including neutral, anger, contempt, disgust, fear, happiness, sadness, and surprise. For light direction attributes, we derive and modify classes from [29] and give 6 light direction classes. Complete lists are given in the Appendix. Moreover, as shown in Table 1, CelebV-HQ [66] is the only dataset giving timestamps of dynamic attributes. Following their idea, we densely annotate all dynamic attributes of CelebV-Text with the start and end time. + +Automatic and Manual Annotation. Based on our attributes design, we find that some attributes can be annotated automatically (e.g., appearance) while some need manual annotations (e.g., timestamps of dynamic attributes). Considering the dataset quality and cost of expense, our annotation strategy includes both automatic and manual annotations. + +For automatic annotation, we first investigate algorithms and select designed attributes that can be automatically annotated. We then test different algorithms on our dataset and keep those giving annotation accuracy of $85\%$ or higher. This process yields all light condition labels, all appearance labels, and all emotion labels suitable for automatic annotation. Algorithms for different labels we finally chose are reported in the Appendix. Automatic annotation results can be further revised by human workers to improve accuracy in a less costly way. + +For manual annotation, we hire and train human workers following [66] to annotate attributes that are filtered out by + +an automatic annotation process. In this case, we manually annotate dynamic attributes, i.e., action and light directions, to give both class labels and exact timestamps. In addition, it is hard to represent detailed appearance attributes by the discrete label, e.g., the characteristics of scars or moles. We therefore ask annotators to give a natural description for each attribute, describing exact positions relative to face parts. These designs greatly enhance the relevance between the final text and the video. + +# 3.3. Semi-auto Text Generation + +Multimodal text-video datasets collect texts via three common methods: subtitles [4, 40, 67], manual-text generation [3, 9, 32, 56, 62], and auto-text generation [5, 21, 27]. However, it is difficult for the individual method to generate texts with high relevance to videos, natural expression, and high diversity. Specifically, although subtitles are easy to obtain, they can pose weakly relevant text-video pairs and introduce noise, making the dataset quality hard to control. Moreover, manual-text generation method is time and cost consuming, as natural language descriptions are required for each video. In this case, increasing the data scale is quite hard as more workers are needed to describe new videos, which does not meet the efficiency and scalability of annotation. Finally, auto-text generation is flexible and scalable, as abundant texts can be simultaneously generated given annotation results of collected videos. However, the diversity, complexity, and naturalness of generated texts can be impacted by the designed grammar templates. + +To this end, we propose a semi-auto template-based text generation strategy that combines both manual-text and auto-text generation methods. Specifically, as mentioned in Section 3.2, manual-texts are required to describe detailed appearance attributes. Annotated attribute information is fed into our designed template for auto-text generation. + +To make our template as natural as possible, we first ask each annotator to describe 10 different face videos for each attribute. We then analyze the grammar structure (i.e., parse tree banks) along with online corpora following [10, 31], and find the most three common grammar structures for each attribute. Finally, we utilize probabilistic context-free grammar [52, 61] and modify the grammar structures to design our own templates. Texts are generated based on templates with synonym replacement using NLTK [8] to increase our generation diversity. Details of our template de + +![](images/c5c729fb58a120a7c4b9b09e97194b28978b2f7c66e2e7beb3395396282b25e8.jpg) +(a) Image quality distribution + +![](images/8ae208a61353f51058e45f49014aa986ea1c2b064485a4f62d0a05ee160c2b33.jpg) +(b) Video quality distribution +Figure 4. Dataset quality distribution. The metrics used are BRISQUE [42] and VSFA [33] respectively. + +signs are in the Appendix. + +# 4. Statistical Analysis of CelebV-Text + +In this section, we compare CelebV-Text with the two most relevant and representative face video datasets [19,66]. We perform a comprehensive analysis of CelebV-Text in terms of video, text, and text-video relevance. To verify the effectiveness of our designed grammar templates, we generate text descriptions for CelebV-HQ based on its attributes for comparison. For simplicity, we use "CelebV-HQ" to denote this variant in the following. + +# 4.1. Video Comparisons + +We briefly compare the overall statistics of existing face video datasets [11, 19, 60, 66] in Table 1. As reported, CelebV-Text contains 70,000 video clips with a total duration of around 279 hours. Each video is accompanied by 20 sentences describing all 6 designed attributes. Compared to CelebV [60], CelebV-Text has a larger scale and higher resolution. Although VoxCeleb2 [11] has more samples than CelebV-Text, its video distribution is limited as most videos are mainly talking faces. Moreover, video samples of both CelebV-HQ [66] and CelebV-Text are collected in open-world with diverse queries so that they are rich in distribution, while CelebV-Text has about 2 times video data, more video attributes, and highly relevant text descriptions. Finally, compared to the only existing facial text-video dataset MM-Vox [19], CelebV-Text overpasses MM-Vox in terms of scale and quality. + +Attributes Distribution. In order to better present the distribution of different attributes in CelebV-Text, we pick and divide general appearance, action, and light direction attributes into groups. More distributions and division designs are provided in the Appendix. Specifically, all 40 general appearance classes are divided into 5 groups shown in Figure 3 (a). Facial features (e.g., double chin, big nose, and oval face) account for the most portion around $45\%$ . The elementary group is twice large than the beard type, accounting for around $25\%$ and $12\%$ , respectively. Fewer samples are located to the hairstyle and accessories groups, + +![](images/8f1c27f9666ba9bebd95f0a8a9d2ec53224c929a45807271dd4d9d8d189191b1.jpg) +(a) Number of word distribution + +![](images/0fcc4b2effa71e2482d61d55b7b5b1005caf599f0c78b81d51fb0e376a4f946f.jpg) +(b) 4-Grams distribution +MM-Vox + +![](images/a84628f0117593e033e8be4469d122b0db503a029981c50b4f5c21a92a69dffa.jpg) +This man with grey hair has a big nose and bags under eyes. He is wearing eyeglasses and earrings, with freckles above the eye areas. The man is firstly gazing for a short time and then he wags his head for a short time, and finally he keeps gazing for the rest of the time. He firstly remains neutral for a long time and then he turns angry for the rest of the time. The video is slightly dark in warm light. The light direction is front lighting for the whole time. CelebV-Text + +(c) Video and text examples +Figure 5. Text distribution. CelebV-Text achieves better performance in both 4-gram and number words distribution. +![](images/c9689ecab2ebe6ca19b83c6809ae2af7c720137dc660b41a353b2d6903cfadb8.jpg) +a man has bags under eyes , receding hairline and straight hairhe is wearing goatee . he is chubby . he has beard , archedeyebrows and black hair . + +taking around $10\%$ and $8\%$ , respectively. Besides, action attributes are divided into 5 groups in Figure 3 (b), where it is clear that head-related actions account for the largest portion of around $60\%$ , followed by eyes-related actions of around $20\%$ . The interaction group (e.g., eat), feeling group (e.g., smile), and daily group (e.g., sleep) account for around $9\%$ , $7\%$ , and $4\%$ , respectively. Finally, for light directions (Figure 3 (c)), most samples contain the front lighting and the remaining ones are evenly distributed. + +Video Quality Distribution. We follow [66] to analyze the quality of our collected videos. To demonstrate the superiority of CelebV-Text, we compare with MM-Vox [19] and CelebV-HQ [66], where mean BRISQUE [42] and VSFA [33] are used to evaluate the image and video quality, respectively. Image quality of all datasets is shown in Figure 4 (a), where CelebV-Text and CelebV-HQ achieve comparable quality, higher than MM-Vox by a large margin. Video quality of all datasets is shown in Figure 4 (b), where CelebV-Text has the best quality, which is due to the effect of the video split method mentioned in Section 3.1, alleviating the discontinuity during background transitions. + +# 4.2. Text Comparisons + +In addition to a large number of video samples, text descriptions of CelebV-Text are longer and more detailed than those in MM-Vox [19] and CelebV-HQ [66] (see Figure 5 (a)), where the average text length of MM-Vox, CelebV-HQ, and CelebV-Text are 28.39, 31.06, and 67.15. Distributions of Celeb-HQ and MM-Vox are close, but there are more words in CelebV-Text to describe a video due to the comprehensive annotation. + +To validate the linguistic diversity of the generated texts, comparisons are conducted among the three datasets following [56]. Specifically, we report the unique part-of-speech (POS) tags (i.e., verb, noun, adjective, and adverb) of the three datasets in Table 3. Obviously, due to our comprehensively designed attribute list and the number of templates, CelebV-Text presents a wider variety of text styles, covering a broader range of face attributes that are static or + +Table 2. Multimodal retrieval results. Clip2Video [17] is leveraged to measure the text-video relevance via retrieval experiments. Bold values indicate the best results, underlined ones indicate the second best. + +
Text ⇒ VideoVideo ⇒ Text
DescriptionDatasetR@1(↑)R@5(↑)R@10(↑)MdR(↓)MnR(↓)R@1(↑)R@5(↑)R@10(↑)MdR(↓)MnR(↓)
(a) App.MM-Vox [19]1.59.015.752.068.82.09.214.643.057.8
CelebV-HQ [66]5.919.229.727.052.27.220.732.427.046.9
CelebV-Text6.121.335.526.349.17.420.729.926.648.3
(b) App.+Emo.CelebV-HQ [66]6.520.130.825.048.07.925.538.817.037.0
CelebV-Text6.623.437.126.047.68.127.234.718.238.3
(c) App.+Emo.+Act.CelebV-Text6.924.139.225.846.78.027.637.116.736.1
+ +Table 3. Number of unique POS tags. The numbers of unique POS tags for MM-Vox, CelebV-HQ, and CelebV-Text. + +
Dataset#Verb#Adj.#Noun#Adv.
MM-Vox [19]520380
CelebV-HQ [66]1024506
CelebV-Text967817424
+ +dynamic in the temporal domain. + +In addition, we further examine the naturalness and complexity of our texts compared to MM-Vox, where we modify [63] to calculate the type-token vocabulary curve for all captions. As shown in Figure 5 (b) where unique 4-grams are selected as the types [56], it is evident that due to our grammar structures and synonym replacement, the linguistic naturalness (vocabulary use) and complexity (vocabulary size) of our CelebV-Text are much better. Please refer to Appendix for more $n$ -grams results. + +# 4.3. Text-Video Relevance + +To quantitatively validate our text-video relevance, we conduct text-video retrieval tasks on three datasets: MMVox [19], CelebV-HQ [66], and CelebV-Text. Rather than use conventional frame-wise clip score as most works [24, 51, 55], we follow [17] to compute feature similarities between texts and videos with the consideration of temporal dynamics, which reflects accurate multimodal interactions across the two modalities. Recall at rank K (R@K), median rank (MdR), and mean rank (MnR) [17, 41, 64] are used as evaluation metrics, where the higher R@K, the lower median rank and mean rank indicate better performance. + +We first examine the performance given texts with descriptions of general appearance in Table 2 (a). Results of CelebV-HQ and CelebV-Text are both better than MM-Vox for two retrieval tasks, which indicates our designed templates can produce texts more relevant to videos than MM-Vox. We further add descriptions about dynamic emotion changes to CelebV-HQ and CelebV-Text in Table 2 (b). Similar results are achieved in both datasets, which reflects that our annotation accuracy on static appearance attributes is as good as CelebV-HQ. Finally, we append action descriptions to CelebV-Text in Table 2 (c), which achieves the best performance on most metrics, verifying the relevance between our generated texts and video samples. + +# 5. Experiment + +In this section, we first conduct facial text-to-video generation to validate the effectiveness of our CelebV-Text dataset. We then benchmark representative approaches on facial text-to-video generation task. + +# 5.1. High-relevance Text-to-Video Generation + +To show the benefits brought by our text descriptions which depict both static and dynamic attributes, we conduct experiments to show the effectiveness of CelebV-Text. Experiments are mainly based on a recent open-sourced state-of-the-art method, MMVID [19], and compared with $\mathrm{CogVideo}^2$ [26], which is a large-scale pretrained text-to-video model, trained on millions of text-image/video pairs. Static Face Video Generation. To validate the effectiveness of our facial text-video dataset in static attributes, we use the models stated above to generate videos conditioned on general appearance, face details, and light conditions descriptions, respectively. Specifically, we first train MMVID [19] from scratch solely on CelebV-Text. We then generate 3 input texts including individual descriptions of each of the static attributes. Generated texts are fed into both MMVID [19] and $\mathrm{CogVideo}$ [26] and corresponding video outputs are examined. + +Visualization results of general appearance are shown in Figure 6 (a), which prove the effectiveness of our dataset. We observe that although CogVideo can output the face video given a text description, the text-video pair is not quite relevant, such as "bags under eyes" and "wavy hair". However, MMVID [19] produces videos with high relevance to input texts, containing all attributes described in the text. More results are shown in the Appendix. + +Dynamic Face Video Generation. We follow the above experimental setting to validate the effectiveness of our dataset with dynamic attribute changes (i.e., emotion, action and light direction). Due to the difficulty in modelling state change [27, 51], we follow [5] to apply test-time interpolation to MMVID [19], named MMVID-interp, to improve the text encoding and better understand the dynamics. Details of our modification are shown in the Appendix. + +In Figure 6 (b), we observe that CogVideo fails to reflect the temporal change described in the input text, i.e., smile + +![](images/efae64ba66f72d93c5ff2220733fd9f97a67492b842e56d1eb24ba07da6c4fbf.jpg) + +The man has black hair and bags under eyes. He has a beard, goatee and arched eyebrows. + +![](images/ba52a7552e850638ece4b6b6087f7e53b959ae5de42549044c9215ee4b796997.jpg) +(a) Static - General Appearance + +![](images/74b2b38e3ce9c9cf64c6b6bbcf306e44b89fed6a6b604f4806da868eab81ce21.jpg) + +The young woman is wearing lipstick and earrings. She firstly smiles and then turns her head. + +![](images/3932e1044d7d61834cb037e35b9f0c6004423953232f3ed597b93f856772e966.jpg) +(b) Dynamic - Action +Figure 6. Qualitative results of facial text-to-video generation. The generated samples are given texts describing (a) the static attribute and (b) dynamic attribute. + +Table 4. Benchmark of text-to-video generation on different datasets. $\downarrow$ means a lower value is better and $\uparrow$ means the opposite. +(a) Quantitative results on general appearance descriptions. + +
DatasetMethodFVD(↓)FID(↓)CLIPSIM(↑)
MM-Vox [19]TFGAN [5]502.28 ± 1.66760.24 ± 16.010.165 ± 0.022
MMVID [19]65.79 ± 1.8138.81 ± 3.660.170 ± 0.020
CelebV-HQ [66]TFGAN [5]428.04 ± 1.76616.24 ± 17.450.168 ± 0.021
MMVID [19]73.65 ± 1.4363.86 ± 3.660.172 ± 0.019
CelebV-TextTFGAN [5]403.04 ± 1.34589.24 ± 16.460.177 ± 0.012
MMVID [19]66.69 ± 1.3558.70 ± 4.670.198 ± 0.014
+ +$\rightarrow$ turn. However, both MMVID [19] and MMVID-interp trained on CelebV-Text can successfully model the dynamic attribute changes, which demonstrates the effectiveness of our dataset. In addition, we find that MMVID [19] cannot preserve some attributes well (e.g., earrings), while MMVID-interp can stabilize the sampling process, validating the effectiveness of our modification. More results are shown in the Appendix. + +Note that CogVideo [26] has a much larger model size ( $\sim 100$ times larger than MMVID [19]) and is trained on much large text-video data ( $\sim 75$ times larger than CelebV-Text). However, video samples produced by CogVideo [26] shown in Figure 6 are of a lower quality than the ones by MMVID [19] trained solely on CelebV-Text, where gener + +(b) Quantitative results on dynamic descriptions of CelebV-Text. + +
DatasetMethodFVD(↓)FID(↓)CLIPSIM(↑)
CelebV-TextTFGAN [5]442.30 ± 2.56623.17 ± 18.880.158 ± 0.024
App.+Emo.MMVID [19]82.78 ± 1.4761.58 ± 3.990.176 ± 0.008
MMVID-interp72.87 ± 1.2341.57 ± 3.560.182 ± 0.010
CelebV-TextTFGAN [5]571.34 ± 4.54784.93 ± 20.130.154 ± 0.028
App.+Act.MMVID [19]109.25 ± 2.1182.55 ± 4.370.174 ± 0.019
MMVID-interp80.81 ± 2.5570.88 ± 4.770.176 ± 0.020
+ +ated faces are not in a high relevance to input texts, demonstrating the effectiveness of our facial text-video dataset. + +# 5.2. Benchmark on Facial Text-to-Video Generation + +As the domain of text-to-video generation is currently thriving, there exists only one benchmark in the face domain, MM-Vox [19]. We expand [19] and construct a benchmark of facial text-to-video generation tasks on three datasets: MM-Vox [19], CelebV-HQ [66] with texts generated by our templates, and CelebV-Text. We choose two representative methods3, TFGAN [5] and MMVID [19], to evaluate their performances on all datasets. + +![](images/49b1fc43a2f5568f612a67090fb86cc277c7a252e1dda53770aee87931ba751f.jpg) + +He has bushy eyebrows, beard and wavy hair. He has got 5 o'clock shadow and brown hair. He has bags under eyes and sideburns with mustache. + +![](images/abe76c6b731b63028efed25c276ff8711a484ae88577c50204f564b470ab8381.jpg) +Figure 7. Qualitative results on three facial text-video datasets. Red and yellow regions indicate the missing of "bags under eyes" and the existence of "wavy hair" and "bags under eyes". + +Quantitative Results. For thorough benchmark construction, we evaluate baseline methods given variant texts including static and dynamic attributes. We use FVD [53] (temporal consistency), FID [23] (individual frame quality), and CLIPSIM [57] (text-video relevance) as evaluation metrics following [19] and report detailed results for appearance, action, and emotion in Table 4. Evaluation steps are repeated over ten runs with mean values and standard errors reported as well. All other values are shown in the Appendix. It can be seen from Table 4 that MMVID [19] obtains good FVD/FID/CLIPSIM metrics over TFGAN [5] which fails to generate reasonable video outputs. In addition, when input texts contain descriptions about a dynamic state change in the temporal domain, the generated video quality by MMVID [19] decreases, which encourages future methods to focus more on cross-modal understanding and consistent video generation. Moreover, the performance of MMVID-interp is better than MMVID [19] on all metrics, validating the effectiveness of our modification mentioned in Section 5.1. Due to challenges posed by our dataset and text-to-video generation task, there is still considerable room to improve. + +Qualitative Results. Video samples generated from MMVID [19] trained on different datasets are shown in Figure 7, where all video frames are of $128^2$ . We can see that video samples generated by MMVID [19] trained on different datasets are of high quality with temporal consistency. However, MMVID [19] trained on MM-Vox [19] can sometimes fail to generate attributes mentioned in the input texts. More generated video samples with dynamic attribute changes are shown in the Appendix. + +# 6. Discussion + +We have proposed CelebV-Text, a large-scale, high-quality, and diverse facial text-video dataset with static and + +dynamic attributes. CelebV-Text contains 70,000 video clips, each of which is accompanied by 20 individual sentences describing both static and dynamic factors. Through extensive statistical analysis and experiments, we have demonstrated the superiority and effectiveness of CelebV-Text. In the future, we plan to further enlarge CelebV-Text in both scale and diversity. We may further explore several new tasks based on CelebV-Text, such as fine-grained control of video face, adaptation of general pretrained models to the face domain, and text-driven 3D-aware facial video generation. + +Ethical Consideration. CelebV-Text is intended for research purposes only. While the raw videos will not be released, the data annotations, links to raw videos, and data processing tools will be made available after undergoing a rigorous legality check procedure at our institution. It is worth noting that our data annotation does not include any personal biometric information such as identity. Only generic attribute information such as gender, hair color, and motion is annotated. Additionally, synthetic videos generated in this work do not exhibit bias or certain biometric information (e.g., big lips or big nose), alleviating ethical concerns. CelebV-Text may be used for deepfakes, but it can also be used for forgery detection tasks to prevent such issues. We will try our best to control the application and acquisition procedure of CelebV-Text to avoid potential misuse and abuse. In the future, we plan to use synthetic face generation frameworks to generate synthetic face videos to address the ethical shortcomings of existing real-world face video datasets. + +Acknowledgement. CelebV-Text is developed under OpenXDLab - an open platform for X-Dimension high-quality data. This study is supported by the RIE2020 Industry Alignment Fund Industry Collaboration Projects (IAF-ICP) Funding Initiative, in-kind contribution from the industry partner(s), and the MOE AcRF Tier 1 (RG16/21). + +# References + +[1] Niki Aifanti, Christos Papachristou, and Anastasios Delopoulos. The mug facial expression database. In WIAMIS, 2010. 2 +[2] Jean-Baptiste Alayrac, Piotr Bojanowski, Nishant Agrawal, Josef Sivic, Ivan Laptev, and Simon Lacoste-Julien. Unsupervised learning from narrated instruction videos. In CVPR, 2016. 1 +[3] Lisa Anne Hendricks, Oliver Wang, Eli Shechtman, Josef Sivic, Trevor Darrell, and Bryan Russell. Localizing moments in video with natural language. In ICCV, 2017. 2, 4 +[4] Max Bain, Arsha Nagrani, Gül Varol, and Andrew Zisserman. Frozen in time: A joint video and image encoder for end-to-end retrieval. In ICCV, 2021. 1, 4 +[5] Yogesh Balaji, Martin Renqiang Min, Bing Bai, Rama Chellappa, and Hans Peter Graf. Conditional gan with discriminative filter generation for text-to-video synthesis. In IJCAI, 2019. 1, 2, 4, 6, 7, 8 +[6] Alex Bewley, Zongyuan Ge, Lionel Ott, Fabio Ramos, and Ben Upcroft. Simple online and realtime tracking. In ICIP. IEEE, 2016. 3 +[7] Sergey Bezryadin, Pavel Bourov, and Dmitry Ilinih. Brightness calculation in digital image processing. In TDPF, 2007. 4 +[8] Steven Bird, Ewan Klein, and Edward Loper. Natural language processing with Python: analyzing text with the natural language toolkit. "O'Reilly Media, Inc.", 2009. 4 +[9] David Chen and William B Dolan. Collecting highly parallel data for paraphrase evaluation. In ACL, 2011. 2, 4 +[10] Danqi Chen and Christopher D Manning. A fast and accurate dependency parser using neural networks. In EMNLP, 2014. 2, 4 +[11] Joon Son Chung, Arsha Nagrani, and Andrew Zisserman. Voxceleb2: Deep speaker recognition. In INTERSPEECH, 2018. 1, 2, 3, 5 +[12] Karl Cobbe, Oleg Klimov, Chris Hesse, Taehoon Kim, and John Schulman. Quantifying generalization in reinforcement learning. In ICML, 2019. 2 +[13] Dima Damen, Hazel Doughty, Giovanni Maria Farinella, Antonino Furnari, Evangelos Kazakos, Jian Ma, Davide Moltisanti, Jonathan Munro, Toby Perrett, Will Price, et al. Rescaling egocentric vision: collection, pipeline and challenges for epic-kitchens-100. In IJCV, 2022. 2 +[14] Jiankang Deng, Jia Guo, Niannan Xue, and Stefanos Zafeiriou. Arcface: Additive angular margin loss for deep face recognition. In CVPR, 2019. 3 +[15] Kangle Deng, Tianyi Fei, Xin Huang, and Yuxin Peng. IRCAN: Introspective recurrent convolutional gan for text-to-video generation. In IJCAI, pages 2216-2222, 2019. 2 +[16] Patrick Esser, Robin Rombach, and Bjorn Ommer. Taming transformers for high-resolution image synthesis. In CVPR, 2021. 2 +[17] Han Fang, Pengfei Xiong, Luhui Xu, and Yu Chen. Clip2video: Mastering video-text retrieval via image clip. arXiv preprint arXiv:2106.11097, 2021. 2, 6 +[18] Tanmay Gupta, Dustin Schwenk, Ali Farhadi, Derek Hoiem, and Aniruddha Kembhavi. Imagine this! scripts to compositions to videos. In ECCV, 2018. 1 +[19] Ligong Han, Jian Ren, Hsin-Ying Lee, Francesco Barbieri, Kyle Olszewski, Shervin Minaee, Dimitris Metaxas, and + +Sergey Tulyakov. Show me what and tell me how: Video synthesis via multimodal conditioning. In CVPR, 2022. 2, 3, 4, 5, 6, 7, 8 +[20] William Harvey, Saeid Naderiparizi, Vaden Masrani, Christian Weilbach, and Frank Wood. Flexible diffusion modeling of long videos. arXiv preprint arXiv:2205.11495, 2022. 2 +[21] Thomas Hayes, Songyang Zhang, Xi Yin, Guan Pang, Sasha Sheng, Harry Yang, Songwei Ge, Isabelle Hu, and Devi Parikh. Mugen: A playground for video-audio-text multimodal understanding and generation. arXiv preprint arXiv:2204.08058, 2022. 2, 4 +[22] Javier Hernandez-Andres, Raymond L Lee, and Javier Romero. Calculating correlated color temperatures across the entire gamut of daylight and skylight chromaticities. In Applied optics, 1999. 4 +[23] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. NeurIPS, 2017. 8 +[24] Jonathan Ho, William Chan, Chitwan Sahara, Jay Whang, Ruiqi Gao, Alexey Gritsenko, Diederik P Kingma, Ben Poole, Mohammad Norouzi, David J Fleet, et al. Imagen video: High definition video generation with diffusion models. arXiv preprint arXiv:2210.02303, 2022. 2, 6, 7 +[25] Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan, Mohammad Norouzi, and David J Fleet. Video diffusion models. arXiv preprint arXiv:2204.03458, 2022. 2 +[26] Wenyi Hong, Ming Ding, Wendi Zheng, Xinghan Liu, and Jie Tang. Cogvideo: Large-scale pretraining for text-to-video generation via transformers. arXiv preprint arXiv:2205.15868, 2022. 2, 6, 7 +[27] Yaosi Hu, Chong Luo, and Zhenzhong Chen. Make it move: controllable image-to-video generation with text descriptions. In CVPR, 2022. 2, 4, 6 +[28] Justin Johnson, Bharath Hariharan, Laurens Van Der Maaten, Li Fei-Fei, C Lawrence Zitnick, and Ross Girshick. Clevr: A diagnostic dataset for compositional language and elementary visual reasoning. In CVPR, 2017. 2 +[29] Peter Kán and Hannes Kafumann. Deep light: light source estimation for augmented reality using deep learning. The Visual Computer, 2019. 4 +[30] Doyeon Kim, Donggyu Joo, and Junmo Kim. Tivgan: Text to image to video generation with step-by-step evolutionary generator. In IEEE Access, 2020. 3 +[31] Dan Klein and Christopher D Manning. Accurate unlexicalized parsing. In ACL, 2003. 4 +[32] Ranjay Krishna, Kenji Hata, Frederic Ren, Li Fei-Fei, and Juan Carlos Niebles. Dense-captioning events in videos. In ICCV, 2017. 2, 4 +[33] Dingquan Li, Tingting Jiang, and Ming Jiang. Quality assessment of in-the-wild videos. In ACM MM, pages 2351-2359, 2019. 5 +[34] Yitong Li, Martin Min, Dinghan Shen, David Carlson, and Lawrence Carin. Video generation from text. In Proceedings of the AAAI conference on artificial intelligence, 2018. 1 +[35] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dálár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In ECCV, 2014. 2 +[36] Wen Liu, Zhixin Piao, Jie Min, Wenhan Luo, Lin Ma, and Shenghua Gao. Liquid warping gan: A unified framework + +for human motion imitation, appearance transfer and novel view synthesis. In ICCV, 2019. 2 +[37] Yue Liu, Xin Wang, Yitian Yuan, and Wenwu Zhu. Cross-modal dual learning for sentence-to-video generation. In ACM MM, 2019. 1 +[38] Ziwei Liu, Ping Luo, Xiaogang Wang, and Xiaou Tang. Deep learning face attributes in the wild. In ICCV, 2015. 3, 4 +[39] Jonathan Malmaud, Jonathan Huang, Vivek Rathod, Nick Johnston, Andrew Rabinovich, and Kevin Murphy. What's cookin'? interpreting cooking videos using text, speech and vision. In ACL, 2015. 1 +[40] Antoine Miech, Dimitri Zhukov, Jean-Baptiste Alayrac, Makarand Tapaswi, Ivan Laptev, and Josef Sivic. Howto100m: Learning a text-video embedding by watching hundred million narrated video clips. In CVPR, 2019. 2, 4 +[41] Niluthpol Chowdhury Mithun, Juncheng Li, Florian Metze, and Amit K Roy-Chowdhury. Learning joint embedding with multimodal cues for cross-modal video-text retrieval. In ICMR, 2018. 6 +[42] Anish Mittal, Anush Krishna Moorthy, and Alan Conrad Bovik. No-reference image quality assessment in the spatial domain. In TIP, 2012. 5 +[43] Gaurav Mittal, Tanya Marwah, and Vineeth N Balasubramanian. Sync画: Automatic video generation using deep recurrent attentive architectures. In ACM MM, 2017. 1, 2 +[44] Ali Mollahosseini, Behzad Hasani, and Mohammad H Ma-hoor. Affectnet: A database for facial expression, valence, and arousal computing in the wild. In TAC, 2017. 4 +[45] A. Nagrani, J. S. Chung, and A. Zisserman. Voxceleb: a large-scale speaker identification dataset. In INTER-SPEECH, 2017. 1, 3 +[46] Yingwei Pan, Zhaofan Qiu, Ting Yao, Houqiang Li, and Tao Mei. To create what you tell: Generating videos from captions. In ACM MM, 2017. 2 +[47] MarcAurelio Ranzato, Arthur Szlam, Joan Bruna, Michael Mathieu, Ronan Collobert, and Sumit Chopra. Video (language) modeling: a baseline for generative models of natural videos. arXiv preprint arXiv:1412.6604, 2014. 1 +[48] Andreas Rössler, Davide Cozzolino, Luisa Verdoliva, Christian Riess, Justus Thies, and Matthias Nießner. Faceforensics: A large-scale video dataset for forgery detection in human faces. arXiv preprint arXiv:1803.09179, 2018. 1 +[49] Christoph Schuhmann, Richard Vencu, Romain Beaumont, Robert Kaczmarczyk, Clayton Mullis, Aarush Katta, Theo Coombes, Jenia Jitsev, and Aran Komatsuzaki. Laion-400m: Open dataset of clip-filtered 400 million image-text pairs. arXiv preprint arXiv:2111.02114, 2021. 2 +[50] Piyush Sharma, Nan Ding, Sebastian Goodman, and Radu Soricut. Conceptual captions: A cleaned, hypernymed, image alt-text dataset for automatic image captioning. In ACL, 2018. 2 +[51] Uriel Singer, Adam Polyak, Thomas Hayes, Xi Yin, Jie An, Songyang Zhang, Qiyuan Hu, Harry Yang, Oron Ashual, Oran Gafni, et al. Make-a-video: Text-to-video generation without text-video data. arXiv preprint arXiv:2209.14792, 2022. 2, 6, 7 +[52] David Stap, Maurits Bleeker, Sarah Ibrahimi, and Maartje ter Hoeve. Conditional image generation and manipulation for user-specified content. arXiv preprint arXiv:2005.04909, 2020. 2, 4 + +[53] Thomas Unterthiner, Sjoerd van Steenkiste, Karol Kurach, Raphael Marinier, Marcin Michalski, and Sylvain Gelly. Towards accurate generative models of video: A new metric & challenges. arXiv preprint arXiv:1812.01717, 2018. 8 +[54] Aaron Van Den Oord, Oriol Vinyals, et al. Neural discrete representation learning. NeurIPS, 2017. 2 +[55] Ruben Villegas, Mohammad Babaeizadeh, Pieter-Jan Kindermans, Hernan Moraldo, Han Zhang, Mohammad Taghi Saffar, Santiago Castro, Julius Kunze, and Dumitru Erhan. Phenaki: Variable length video generation from open domain textual description. arXiv preprint arXiv:2210.02399, 2022. 2, 6, 7 +[56] Xin Wang, Jiawei Wu, Junkun Chen, Lei Li, Yuan-Fang Wang, and William Yang Wang. Vatex: A large-scale, high-quality multilingual dataset for video-and-language research. In ICCV, 2019. 4, 5, 6 +[57] Chenfei Wu, Lun Huang, Qianxi Zhang, Binyang Li, Lei Ji, Fan Yang, Guillermo Sapiro, and Nan Duan. Godiva: Generating open-domain videos from natural descriptions. arXiv preprint arXiv:2104.14806, 2021. 2, 8 +[58] Chenfei Wu, Jian Liang, Xiaowei Hu, Zhe Gan, Jianfeng Wang, Lijuan Wang, Zicheng Liu, Yuejian Fang, and Nan Duan. Nuwa-infinity: Autoregressive over autoregressive generation for infinite visual synthesis. In NeurIPS, 2022. 2 +[59] Chenfei Wu, Jian Liang, Lei Ji, Fan Yang, Yuejian Fang, Daxin Jiang, and Nan Duan. N\`uwa: Visual synthesis pretraining for neural visual world creation. In ECCV, 2022. 2 +[60] Wayne Wu, Yunxuan Zhang, Cheng Li, Chen Qian, and Chen Change Loy. Reenactgan: Learning to reenact faces via boundary transfer. In ECCV, 2018. 1, 3, 5 +[61] Weihao Xia, Yujiu Yang, Jing-Hao Xue, and Baoyuan Wu. Tedigan: Text-guided diverse face image generation and manipulation. In CVPR, 2021. 3, 4 +[62] Jun Xu, Tao Mei, Ting Yao, and Yong Rui. Msr-vtt: A large video description dataset for bridging video and language. In CVPR, 2016. 2, 4 +[63] Gilbert Youmans. Measuring lexical style and competence: The type-token vocabulary curve. Style, 1990. 6 +[64] Bowen Zhang, Hexiang Hu, and Fei Sha. Cross-modal and hierarchical modeling of video and text. In ECCV, 2018. 6 +[65] Luowei Zhou, Chenliang Xu, and Jason J Corso. Towards automatic learning of procedures from web instructional videos. In AAAI, 2018. 2 +[66] Hao Zhu, Wayne Wu, Wentao Zhu, Liming Jiang, Siwei Tang, Li Zhang, Ziwei Liu, and Chen Change Loy. CelebV-HQ: A large-scale video facial attributes dataset. In ECCV, 2022. 2, 3, 4, 5, 6, 7 +[67] Dimitri Zhukov, Jean-Baptiste Alayrac, Ramadan Gokberk Cinbis, David Fouhey, Ivan Laptev, and Josef Sivic. Cross-task weakly supervised learning from instructional videos. In CVPR, 2019. 1, 4 \ No newline at end of file diff --git a/2023/CelebV-Text_ A Large-Scale Facial Text-Video Dataset/images.zip b/2023/CelebV-Text_ A Large-Scale Facial Text-Video Dataset/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..71457bc46089a7368217be1bcc684a502c1057f6 --- /dev/null +++ b/2023/CelebV-Text_ A Large-Scale Facial Text-Video Dataset/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2262673592864e643d2aef144c4f2fe65a23a5eaabab5ba0ed616e89f94778fc +size 752742 diff --git a/2023/CelebV-Text_ A Large-Scale Facial Text-Video Dataset/layout.json b/2023/CelebV-Text_ A Large-Scale Facial Text-Video Dataset/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..43eb1177051718abb4f21c7804c2c4a9292757f9 --- /dev/null +++ b/2023/CelebV-Text_ A Large-Scale Facial Text-Video Dataset/layout.json @@ -0,0 +1,8278 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 128, + 91, + 465, + 107 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 128, + 91, + 465, + 107 + ], + "spans": [ + { + "bbox": [ + 128, + 91, + 465, + 107 + ], + "type": "text", + "content": "CelebV-Text: A Large-Scale Facial Text-Video Dataset" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 59, + 118, + 532, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 118, + 532, + 133 + ], + "spans": [ + { + "bbox": [ + 59, + 118, + 532, + 133 + ], + "type": "text", + "content": "Jianhui " + }, + { + "bbox": [ + 59, + 118, + 532, + 133 + ], + "type": "inline_equation", + "content": "\\mathrm{Yu}^{1*}" + }, + { + "bbox": [ + 59, + 118, + 532, + 133 + ], + "type": "text", + "content": " Hao Zhu" + }, + { + "bbox": [ + 59, + 118, + 532, + 133 + ], + "type": "inline_equation", + "content": "^{2*}" + }, + { + "bbox": [ + 59, + 118, + 532, + 133 + ], + "type": "text", + "content": " Liming Jiang" + }, + { + "bbox": [ + 59, + 118, + 532, + 133 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 59, + 118, + 532, + 133 + ], + "type": "text", + "content": " Chen Change Loy" + }, + { + "bbox": [ + 59, + 118, + 532, + 133 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 59, + 118, + 532, + 133 + ], + "type": "text", + "content": " Weidong Cai" + }, + { + "bbox": [ + 59, + 118, + 532, + 133 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 59, + 118, + 532, + 133 + ], + "type": "text", + "content": " Wayne Wu" + }, + { + "bbox": [ + 59, + 118, + 532, + 133 + ], + "type": "inline_equation", + "content": "^{4}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 76, + 133, + 517, + 146 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 133, + 517, + 146 + ], + "spans": [ + { + "bbox": [ + 76, + 133, + 517, + 146 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 76, + 133, + 517, + 146 + ], + "type": "text", + "content": "University of Sydney " + }, + { + "bbox": [ + 76, + 133, + 517, + 146 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 76, + 133, + 517, + 146 + ], + "type": "text", + "content": "SenseTime Research " + }, + { + "bbox": [ + 76, + 133, + 517, + 146 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 76, + 133, + 517, + 146 + ], + "type": "text", + "content": "S-Lab, Nanyang Technological University " + }, + { + "bbox": [ + 76, + 133, + 517, + 146 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 76, + 133, + 517, + 146 + ], + "type": "text", + "content": "Shanghai AI Laboratory" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 91, + 147, + 497, + 158 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 147, + 497, + 158 + ], + "spans": [ + { + "bbox": [ + 91, + 147, + 497, + 158 + ], + "type": "text", + "content": "jianhui.yu@sydney.edu.au haozhu96@gmail.com {liming002,ccloy}@ntu.edu.sg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 171, + 159, + 419, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 171, + 159, + 419, + 171 + ], + "spans": [ + { + "bbox": [ + 171, + 159, + 419, + 171 + ], + "type": "text", + "content": "tom.cai@sydney.edu.au wuwenyan0503@gmail.com" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 49, + 191, + 391, + 392 + ], + "blocks": [ + { + "bbox": [ + 49, + 191, + 391, + 392 + ], + "lines": [ + { + "bbox": [ + 49, + 191, + 391, + 392 + ], + "spans": [ + { + "bbox": [ + 49, + 191, + 391, + 392 + ], + "type": "image", + "image_path": "f0109e28e042bfb7ecc4716cf4e58dae3e14efdf34943c697065303fe469de7e.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 94, + 392, + 346, + 400 + ], + "lines": [ + { + "bbox": [ + 94, + 392, + 346, + 400 + ], + "spans": [ + { + "bbox": [ + 94, + 392, + 346, + 400 + ], + "type": "text", + "content": "(a) Video Samples (left: general appearance, middle: detailed appearance, right: light conditions)" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 403, + 546, + 425 + ], + "lines": [ + { + "bbox": [ + 46, + 403, + 546, + 425 + ], + "spans": [ + { + "bbox": [ + 46, + 403, + 546, + 425 + ], + "type": "text", + "content": "Figure 1. Overview of CelebV-Text. CelebV-Text contains (a) 70,000 video samples and (b) 1,400,000 text descriptions. Each video sample is annotated with general appearance, detailed appearance, light conditions, action, emotion, and light directions." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 394, + 190, + 542, + 392 + ], + "blocks": [ + { + "bbox": [ + 394, + 190, + 542, + 392 + ], + "lines": [ + { + "bbox": [ + 394, + 190, + 542, + 392 + ], + "spans": [ + { + "bbox": [ + 394, + 190, + 542, + 392 + ], + "type": "image", + "image_path": "fcf8b3338799a55aa4546a38e787d33d2e42b686dbba211adb062ffd7cec9900.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 440, + 392, + 496, + 400 + ], + "lines": [ + { + "bbox": [ + 440, + 392, + 496, + 400 + ], + "spans": [ + { + "bbox": [ + 440, + 392, + 496, + 400 + ], + "type": "text", + "content": "(b) Text Descriptions" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 143, + 435, + 191, + 447 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 435, + 191, + 447 + ], + "spans": [ + { + "bbox": [ + 143, + 435, + 191, + 447 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 46, + 456, + 288, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 456, + 288, + 676 + ], + "spans": [ + { + "bbox": [ + 46, + 456, + 288, + 676 + ], + "type": "text", + "content": "Text-driven generation models are flourishing in video generation and editing. However, face-centric text-to-video generation remains a challenge due to the lack of a suitable dataset containing high-quality videos and highly relevant texts. This paper presents CelebV-Text, a large-scale, diverse, and high-quality dataset of facial text-video pairs, to facilitate research on facial text-to-video generation tasks. CelebV-Text comprises 70,000 in-the-wild face video clips with diverse visual content, each paired with 20 texts generated using the proposed semi-automatic text generation strategy. The provided texts are of high quality, describing both static and dynamic attributes precisely. The superiority of CelebV-Text over other datasets is demonstrated via comprehensive statistical analysis of the videos, texts, and text-video relevance. The effectiveness and potential of CelebV-Text are further shown through extensive self-evaluation. A benchmark is constructed with representative methods to standardize the evaluation of the facial text-to-video generation task. All data and models are publicly available1." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 307, + 435, + 386, + 447 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 435, + 386, + 447 + ], + "spans": [ + { + "bbox": [ + 307, + 435, + 386, + 447 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 456, + 545, + 589 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 456, + 545, + 589 + ], + "spans": [ + { + "bbox": [ + 304, + 456, + 545, + 589 + ], + "type": "text", + "content": "Text-driven video generation has recently garnered significant attention in the fields of computer vision and computer graphics. By using text as input, video content can be generated and controlled, inspiring numerous applications in both academia and industry [5,34,43,47]. However, text-to-video generation still faces many challenges, particularly in the face-centric scenario where generated video frames often lack quality [18, 34, 37] or have weak relevance to input texts [2,4,39,67]. We believe that one of the main issues is the absence of a well-suited facial text-video dataset containing high-quality video samples and text descriptions of various attributes highly relevant to videos." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 591, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 591, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 591, + 546, + 713 + ], + "type": "text", + "content": "Constructing a high-quality facial text-video dataset poses several challenges, mainly in three aspects. 1) Data collection. The quality and quantity of video samples largely determine the quality of generated videos [11, 45, 48, 60]. However, obtaining such a large-scale dataset with high-quality samples while maintaining a natural distribution and smooth video motion is challenging. 2) Data annotation. The relevance of text-video pairs needs to be ensured. This requires a comprehensive coverage of text for describing the content and motion appearing in the video, such as light conditions and head movements. 3) Text gen" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 693, + 126, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 693, + 126, + 703 + ], + "spans": [ + { + "bbox": [ + 58, + 693, + 126, + 703 + ], + "type": "text", + "content": "*Equal contribution." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 59, + 703, + 248, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 703, + 248, + 713 + ], + "spans": [ + { + "bbox": [ + 59, + 703, + 248, + 713 + ], + "type": "text", + "content": "1Project page: https://celebv-text.github.io" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14805" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 286, + 115 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 286, + 115 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 286, + 115 + ], + "type": "text", + "content": "eration. Producing diverse and natural texts are non-trivial. Manual text generation is expensive and not scalable. While auto-text generation is easily extensible, it is limited in naturalness." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 117, + 286, + 314 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 117, + 286, + 314 + ], + "spans": [ + { + "bbox": [ + 46, + 117, + 286, + 314 + ], + "type": "text", + "content": "To overcome the challenges mentioned above, we carefully design a comprehensive data construction pipeline that includes data collection and processing, data annotation, and semi-auto text generation. First, to obtain raw videos, we follow the data collection steps of CelebV-HQ, which has proven to be effective in [66]. We introduce a minor modification to the video processing step to improve the video's smoothness further. Next, to ensure highly relevant text-video pairs, we analyze videos from both temporal dynamics and static content and establish a set of attributes that may or may not change over time. Finally, we propose a semi-auto template-based method to generate texts that are diverse and natural. Our approach leverages the advantages of both auto- and manual-text methods. Specifically, we design a rich variety of grammar templates as [10,52] to parse annotation and manual texts, which are flexibly combined and modified to achieve high diversity, complexity, and naturalness." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 316, + 286, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 316, + 286, + 502 + ], + "spans": [ + { + "bbox": [ + 46, + 316, + 286, + 502 + ], + "type": "text", + "content": "With the proposed pipeline, we create CelebV-Text, a Large-Scale Facial Text-Video Dataset, which includes 70,000 in-the-wild video clips with a resolution of at least " + }, + { + "bbox": [ + 46, + 316, + 286, + 502 + ], + "type": "inline_equation", + "content": "512 \\times 512" + }, + { + "bbox": [ + 46, + 316, + 286, + 502 + ], + "type": "text", + "content": " and 1,400,000 text descriptions with 20 for each clip. As depicted in Figure 1, CelebV-Text consists of high-quality video samples and text descriptions for realistic face video generation. Each video is annotated with three types of static attributes (40 general appearances, 5 detailed appearances, and 6 light conditions) and three types of dynamic attributes (37 actions, 8 emotions, and 6 light directions). All dynamic attributes are densely annotated with start and end timestamps, while manual-texts are provided for labels that cannot be discretized. Furthermore, we have designed three templates for each attribute type, resulting in a total of 18 templates that can be flexibly combined. All attributes and manual-texts are naturally described in our generated texts." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 503, + 286, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 503, + 286, + 700 + ], + "spans": [ + { + "bbox": [ + 46, + 503, + 286, + 700 + ], + "type": "text", + "content": "CelebV-Text surpasses existing face video datasets [11] in terms of resolution (over 2 times higher), number of samples, and more diverse distribution. In addition, the texts in CelebV-Text exhibit higher diversity, complexity, and naturalness than those in text-video datasets [19, 66]. CelebV-Text also shows high relevance of text-video pairs, validated by our text-video retrieval experiments [17]. To further examine the effectiveness and potential of CelebV-Text, we evaluate it on a representative baseline [19] for facial text-to-video generation. Our results show better relevance between generated face videos and texts when compared to a state-of-the-art large-scale pretrained model [26]. Furthermore, we show that a simple modification of [19] with text interpolation can significantly improve temporal coherence. Finally, we present a new benchmark for text-to-video generation to standardize the facial text-to-video generation task, which includes representative models [5, 19] on three text-video datasets." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 59, + 701, + 286, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 701, + 286, + 712 + ], + "spans": [ + { + "bbox": [ + 59, + 701, + 286, + 712 + ], + "type": "text", + "content": "The main contributions of this work are summarized as" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 72, + 545, + 194 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 194 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 194 + ], + "type": "text", + "content": "follows: 1) We propose CelebV-Text, the first large-scale facial text-video dataset with high-quality videos, as well as rich and highly-relevant texts, to facilitate research in facial text-to-video generation. 2) Comprehensive statistical analyses are conducted to examine video/text quality and diversity, as well as text-video relevance, demonstrating the superiority of CelebV-Text. 3) A series of self-evaluations are performed to demonstrate the effectiveness and potential of CelebV-Text. 4) A new benchmark for text-to-video generation is constructed to promote the standardization of the facial text-to-video generation task." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 206, + 392, + 219 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 206, + 392, + 219 + ], + "spans": [ + { + "bbox": [ + 306, + 206, + 392, + 219 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 227, + 545, + 535 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 227, + 545, + 535 + ], + "spans": [ + { + "bbox": [ + 304, + 227, + 545, + 535 + ], + "type": "text", + "content": "Text-to-Video Generation. Text-driven video generation, which involves generating videos from text descriptions, has recently gained significant interest as a challenging task. Mittal et al. [43] first introduced this task to generate semantically consistent videos conditioned on encoded captions. Other studies, such as [5, 15, 46], attempt to generate video samples conditioned on encoded text inputs. However, due to the low richness of text descriptions and the small number of data samples, the generated video samples are often at low resolution or lack relevance with the input texts. More recently, several works [19, 26, 27, 55, 57–59] have employed discrete latent codes [16, 54] for more realistic video generation. Some of these works treat videos as a sequence of independent images [19, 26, 58, 59], while Phenaki [55] considers temporal relations between each frame for a more robust video decoding process. Another branch of studies leverage diffusion models for text-to-video generation [20, 24, 25, 51], which require millions or billions of samples to achieve high-quality generation. While text-to-video generation methods are rapidly evolving, they are generally designed for generating general videos. Among these methods, only MMVID [19] has conducted specific experiments with face-centric descriptions. One possible reason for this is that facial text-to-video generation requires more accurate and detailed text descriptions than general tasks. However, there is currently no suitable dataset available that provides such properties for face-centric text-to-video generation." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 536, + 545, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 536, + 545, + 700 + ], + "spans": [ + { + "bbox": [ + 304, + 536, + 545, + 700 + ], + "type": "text", + "content": "Multimodal Datasets. Existing multimodal datasets can be categorized into two classes: open-world and closed-world. Open-world datasets [3,9,13,32,35,36,40,43,49,50,62,65] are widely used for text-to-image/video generation tasks. Some of them have manual annotations [13, 32, 35, 50, 62] and part of them are directly collected from the Internet, such as subtitles [40, 49]. Closed-world datasets are mostly composed of images or videos collected in constrained environment with corresponding information such as text. CLEVR [28] is a synthetic text-image dataset produced by arranging 3D objects with different shapes under a controlled background. While MUGEN [21] is a video-audiotext dataset that was collected using CoinRun [12] by introducing audio and new interactions. The corresponding text is produced by human annotators and grammar templates." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 317, + 701, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 701, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 317, + 701, + 545, + 713 + ], + "type": "text", + "content": "Multimodal face datasets also exist. Modified MUG [1]" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "text", + "content": "14806" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 61, + 96, + 531, + 208 + ], + "blocks": [ + { + "bbox": [ + 47, + 61, + 545, + 93 + ], + "lines": [ + { + "bbox": [ + 47, + 61, + 545, + 93 + ], + "spans": [ + { + "bbox": [ + 47, + 61, + 545, + 93 + ], + "type": "text", + "content": "Table 1. In-the-wild face video dataset comparison. The symbol “#” indicates the number. The abbreviations “Res.”, “Dura.”, “App.”, “Cond.”, “Act.”, “Emo.”, and “Dir.” stand for Resolution, Duration, Appearance, Condition, Action, Emotion, and Direction, respectively. The “half checkmark” denotes that CelebV-HQ consists of action attributes with no timestamp." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 61, + 96, + 531, + 208 + ], + "lines": [ + { + "bbox": [ + 61, + 96, + 531, + 208 + ], + "spans": [ + { + "bbox": [ + 61, + 96, + 531, + 208 + ], + "type": "table", + "html": "
DatasetsMeta InformationAttribute LabelsText
StaticDynamic
#SamplesRes.Dura.General App.Detail App.Light Cond.Act.Emo.Light Dir.Auto Manual
CelebV [60]5256×2562hrs
VoxCeleb2 [11]150,480224×2242442hrs
CelebV-HQ [66]35,666512×51268hrs✓XX✓✓XX
MM-Vox [19]19,522224×224323hrs✓XXXXX✓
CelebV-Text70,000512×512+279hrs✓✓✓✓✓✓✓
", + "image_path": "97120df4bda02a1a091b792dc9856012e28a2c3c55b51fe4e5b3b1aa21522161.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 221, + 289, + 421 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 221, + 289, + 421 + ], + "spans": [ + { + "bbox": [ + 46, + 221, + 289, + 421 + ], + "type": "text", + "content": "is a closed-world text-video dataset that contains 1,039 videos with subjects showing different emotions, where the text descriptions are generated from facial emotions using a fixed template [30]. MM-Vox [19] contains 19,522 face videos from VoxCeleb [45], with 36 facial attributes manually labeled following CelebA [38] and text descriptions generated via Probabilistic Context-Free Grammar (PCFG) [61]. However, both datasets only contain language descriptions related to static facial attributes without considering the temporal state change (i.e., emotion or action) presented in the original face videos. Moreover, the limited label annotations restrict the diversity of the text descriptions, making them sub-optimal for studying the text-to-video generation task on the face domain. CelebV-HQ [66] is the latest high-quality face video dataset that covers facial annotations, including appearance, movement, and emotion. However, it only provides discrete labels and timestamps, with no text descriptions." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 433, + 171, + 445 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 433, + 171, + 445 + ], + "spans": [ + { + "bbox": [ + 47, + 433, + 171, + 445 + ], + "type": "text", + "content": "3. Dataset Construction" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 453, + 287, + 531 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 453, + 287, + 531 + ], + "spans": [ + { + "bbox": [ + 46, + 453, + 287, + 531 + ], + "type": "text", + "content": "In this work, we aim to build a facial text-video dataset, which requires not only large-scale video samples of high quality, but also natural and diverse text descriptions that are highly relevant to videos. To achieve so, we propose an efficient pipeline, as shown in Figure 2, to construct CelebV-Text, including Data Collection & Processing, Data Annotation, and Semi-auto Text Generation." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 539, + 208, + 553 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 539, + 208, + 553 + ], + "spans": [ + { + "bbox": [ + 47, + 539, + 208, + 553 + ], + "type": "text", + "content": "3.1. Data Collection & Processing" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 558, + 287, + 668 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 558, + 287, + 668 + ], + "spans": [ + { + "bbox": [ + 46, + 558, + 287, + 668 + ], + "type": "text", + "content": "Collection. We follow the same strategy as CelebV-HQ [66] due to its effectiveness in large-scale high-quality data collection. Specifically, we firstly generate a large number of queries, including human names, movie titles, vlogs and so on, to retrieve videos that contain human faces with temporally dynamic state changes and abundant facial attributes. Our data are collected from open world with videos downloaded from online resources. Videos with low resolution " + }, + { + "bbox": [ + 46, + 558, + 287, + 668 + ], + "type": "inline_equation", + "content": "(< 512^{2})" + }, + { + "bbox": [ + 46, + 558, + 287, + 668 + ], + "type": "text", + "content": ", low time duration " + }, + { + "bbox": [ + 46, + 558, + 287, + 668 + ], + "type": "inline_equation", + "content": "(< 5s)" + }, + { + "bbox": [ + 46, + 558, + 287, + 668 + ], + "type": "text", + "content": ", and having appeared in CelebV-HQ are filtered out." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 669, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 669, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 669, + 288, + 715 + ], + "type": "text", + "content": "Processing. To sample high-quality and diverse video clips from our raw collections, similar steps are followed as CelebV-HQ [66] with modifications. We first filter out video clips with bounding box regions less than " + }, + { + "bbox": [ + 46, + 669, + 288, + 715 + ], + "type": "inline_equation", + "content": "512^{2}" + }, + { + "bbox": [ + 46, + 669, + 288, + 715 + ], + "type": "text", + "content": " rather" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 310, + 220, + 545, + 371 + ], + "blocks": [ + { + "bbox": [ + 310, + 220, + 545, + 371 + ], + "lines": [ + { + "bbox": [ + 310, + 220, + 545, + 371 + ], + "spans": [ + { + "bbox": [ + 310, + 220, + 545, + 371 + ], + "type": "image", + "image_path": "eedc5f3f0b2e60ba783324e68f63307e42cfcdf071714f39f4be781746fcf053.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 381, + 545, + 414 + ], + "lines": [ + { + "bbox": [ + 305, + 381, + 545, + 414 + ], + "spans": [ + { + "bbox": [ + 305, + 381, + 545, + 414 + ], + "type": "text", + "content": "Figure 2. Pipeline of our dataset construction process. The pipeline includes data collection & processing, data annotation, and semi-auto text generation." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 421, + 546, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 421, + 546, + 533 + ], + "spans": [ + { + "bbox": [ + 305, + 421, + 546, + 533 + ], + "type": "text", + "content": "than resize them to the same resolution. In this way, clips are not upsampled or downsampled hence the video quality would not be affected, which leads to various resolutions of collected videos: " + }, + { + "bbox": [ + 305, + 421, + 546, + 533 + ], + "type": "inline_equation", + "content": "56.4\\%" + }, + { + "bbox": [ + 305, + 421, + 546, + 533 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 305, + 421, + 546, + 533 + ], + "type": "inline_equation", + "content": "512^2 \\sim 1024^2" + }, + { + "bbox": [ + 305, + 421, + 546, + 533 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 305, + 421, + 546, + 533 + ], + "type": "inline_equation", + "content": "43.6\\%" + }, + { + "bbox": [ + 305, + 421, + 546, + 533 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 305, + 421, + 546, + 533 + ], + "type": "inline_equation", + "content": "1024^2+" + }, + { + "bbox": [ + 305, + 421, + 546, + 533 + ], + "type": "text", + "content": ". To reduce the face area noise when the background changes, we further change the video splitting strategy. In addition to our focus on the same human motion [6] and identity [14] present in adjacent frames, we split the video into different clips when the background changes by a toolkit1." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 540, + 407, + 553 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 540, + 407, + 553 + ], + "spans": [ + { + "bbox": [ + 306, + 540, + 407, + 553 + ], + "type": "text", + "content": "3.2. Data Annotation" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 559, + 545, + 625 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 559, + 545, + 625 + ], + "spans": [ + { + "bbox": [ + 305, + 559, + 545, + 625 + ], + "type": "text", + "content": "The annotation process is a core part in CelebV-Text construction, which would greatly affect the relevance of text-video pairs, as our designed text templates heavily depend on the annotation results. Here, we first describe how we design attributes, and then give details about the annotation strategy for face videos." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 625, + 545, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 625, + 545, + 693 + ], + "spans": [ + { + "bbox": [ + 305, + 625, + 545, + 693 + ], + "type": "text", + "content": "Attributes Design. Temporal dynamic is the key difference between images and videos. However, as shown in Table 1, most face video datasets focus on static attributes where attribute information does not change over time, such as appearance. Dynamic attributes that change over time, such as emotion and face actions, are often neglected. In the fol" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 317, + 702, + 479, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 702, + 479, + 713 + ], + "spans": [ + { + "bbox": [ + 317, + 702, + 479, + 713 + ], + "type": "text", + "content": "1 https://github.com/Breakthrough/PySceneDetect" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14807" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 75, + 59, + 233, + 194 + ], + "blocks": [ + { + "bbox": [ + 75, + 59, + 233, + 194 + ], + "lines": [ + { + "bbox": [ + 75, + 59, + 233, + 194 + ], + "spans": [ + { + "bbox": [ + 75, + 59, + 233, + 194 + ], + "type": "image", + "image_path": "a913896e83a99c179f17204cbfab4c4efac3155f5fef60f086907c9c4feecf3c.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 69, + 196, + 523, + 207 + ], + "lines": [ + { + "bbox": [ + 69, + 196, + 523, + 207 + ], + "spans": [ + { + "bbox": [ + 69, + 196, + 523, + 207 + ], + "type": "text", + "content": "Figure 3. Dataset distribution comparison. The distributions of appearance attributes, action attributes, and light directions." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 232, + 59, + 389, + 194 + ], + "blocks": [ + { + "bbox": [ + 232, + 59, + 389, + 194 + ], + "lines": [ + { + "bbox": [ + 232, + 59, + 389, + 194 + ], + "spans": [ + { + "bbox": [ + 232, + 59, + 389, + 194 + ], + "type": "image", + "image_path": "28c5acb5920de6e0ddd8f704943dcbbe09d28a7eced1a6c54706017cf64a6290.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 389, + 59, + 511, + 194 + ], + "blocks": [ + { + "bbox": [ + 389, + 59, + 511, + 194 + ], + "lines": [ + { + "bbox": [ + 389, + 59, + 511, + 194 + ], + "spans": [ + { + "bbox": [ + 389, + 59, + 511, + 194 + ], + "type": "image", + "image_path": "744e07655ba3d485ce2a50ff5b2975ac45985578192d20caa4c48711963ed744.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 214, + 287, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 214, + 287, + 236 + ], + "spans": [ + { + "bbox": [ + 46, + 214, + 287, + 236 + ], + "type": "text", + "content": "lowing, we decouple face videos into static and dynamic categories and details are given as follows." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 236, + 287, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 236, + 287, + 357 + ], + "spans": [ + { + "bbox": [ + 46, + 236, + 287, + 357 + ], + "type": "text", + "content": "1) Static. The current dataset [19] only considers static information such as the appearance attribute, which includes 40 classes as CelebA [38]. In contrast, we define static information to include three types of attributes: general appearance, detailed appearance, and light conditions. General appearance attributes follow the same definition as CelebA [38]. Detailed appearance attributes including five classes are proposed for realistic face generation, i.e., scar, mole, freckle, dimple, and one-eyed. We define light conditions in a restricted manner to include light color temperature [22] and brightness [7], with a total of 6 classes." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 358, + 287, + 501 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 358, + 287, + 501 + ], + "spans": [ + { + "bbox": [ + 46, + 358, + 287, + 501 + ], + "type": "text", + "content": "2) Dynamic. Here, we design three dynamic attributes, i.e., action, emotion, and light directions. For action attributes, we follow CelebV-HQ [66] and expand their action list by two classes, i.e., squint and blink. For emotion attributes, we select the 8 emotion setting in Affectnet [44], including neutral, anger, contempt, disgust, fear, happiness, sadness, and surprise. For light direction attributes, we derive and modify classes from [29] and give 6 light direction classes. Complete lists are given in the Appendix. Moreover, as shown in Table 1, CelebV-HQ [66] is the only dataset giving timestamps of dynamic attributes. Following their idea, we densely annotate all dynamic attributes of CelebV-Text with the start and end time." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 502, + 287, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 502, + 287, + 578 + ], + "spans": [ + { + "bbox": [ + 46, + 502, + 287, + 578 + ], + "type": "text", + "content": "Automatic and Manual Annotation. Based on our attributes design, we find that some attributes can be annotated automatically (e.g., appearance) while some need manual annotations (e.g., timestamps of dynamic attributes). Considering the dataset quality and cost of expense, our annotation strategy includes both automatic and manual annotations." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 580, + 287, + 690 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 580, + 287, + 690 + ], + "spans": [ + { + "bbox": [ + 46, + 580, + 287, + 690 + ], + "type": "text", + "content": "For automatic annotation, we first investigate algorithms and select designed attributes that can be automatically annotated. We then test different algorithms on our dataset and keep those giving annotation accuracy of " + }, + { + "bbox": [ + 46, + 580, + 287, + 690 + ], + "type": "inline_equation", + "content": "85\\%" + }, + { + "bbox": [ + 46, + 580, + 287, + 690 + ], + "type": "text", + "content": " or higher. This process yields all light condition labels, all appearance labels, and all emotion labels suitable for automatic annotation. Algorithms for different labels we finally chose are reported in the Appendix. Automatic annotation results can be further revised by human workers to improve accuracy in a less costly way." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 691, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 691, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 691, + 287, + 713 + ], + "type": "text", + "content": "For manual annotation, we hire and train human workers following [66] to annotate attributes that are filtered out by" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 214, + 545, + 313 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 214, + 545, + 313 + ], + "spans": [ + { + "bbox": [ + 304, + 214, + 545, + 313 + ], + "type": "text", + "content": "an automatic annotation process. In this case, we manually annotate dynamic attributes, i.e., action and light directions, to give both class labels and exact timestamps. In addition, it is hard to represent detailed appearance attributes by the discrete label, e.g., the characteristics of scars or moles. We therefore ask annotators to give a natural description for each attribute, describing exact positions relative to face parts. These designs greatly enhance the relevance between the final text and the video." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 320, + 454, + 332 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 320, + 454, + 332 + ], + "spans": [ + { + "bbox": [ + 305, + 320, + 454, + 332 + ], + "type": "text", + "content": "3.3. Semi-auto Text Generation" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 338, + 545, + 537 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 338, + 545, + 537 + ], + "spans": [ + { + "bbox": [ + 304, + 338, + 545, + 537 + ], + "type": "text", + "content": "Multimodal text-video datasets collect texts via three common methods: subtitles [4, 40, 67], manual-text generation [3, 9, 32, 56, 62], and auto-text generation [5, 21, 27]. However, it is difficult for the individual method to generate texts with high relevance to videos, natural expression, and high diversity. Specifically, although subtitles are easy to obtain, they can pose weakly relevant text-video pairs and introduce noise, making the dataset quality hard to control. Moreover, manual-text generation method is time and cost consuming, as natural language descriptions are required for each video. In this case, increasing the data scale is quite hard as more workers are needed to describe new videos, which does not meet the efficiency and scalability of annotation. Finally, auto-text generation is flexible and scalable, as abundant texts can be simultaneously generated given annotation results of collected videos. However, the diversity, complexity, and naturalness of generated texts can be impacted by the designed grammar templates." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 537, + 545, + 603 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 537, + 545, + 603 + ], + "spans": [ + { + "bbox": [ + 304, + 537, + 545, + 603 + ], + "type": "text", + "content": "To this end, we propose a semi-auto template-based text generation strategy that combines both manual-text and auto-text generation methods. Specifically, as mentioned in Section 3.2, manual-texts are required to describe detailed appearance attributes. Annotated attribute information is fed into our designed template for auto-text generation." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 603, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 603, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 603, + 545, + 713 + ], + "type": "text", + "content": "To make our template as natural as possible, we first ask each annotator to describe 10 different face videos for each attribute. We then analyze the grammar structure (i.e., parse tree banks) along with online corpora following [10, 31], and find the most three common grammar structures for each attribute. Finally, we utilize probabilistic context-free grammar [52, 61] and modify the grammar structures to design our own templates. Texts are generated based on templates with synonym replacement using NLTK [8] to increase our generation diversity. Details of our template de" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14808" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 65, + 65, + 252, + 137 + ], + "blocks": [ + { + "bbox": [ + 65, + 65, + 252, + 137 + ], + "lines": [ + { + "bbox": [ + 65, + 65, + 252, + 137 + ], + "spans": [ + { + "bbox": [ + 65, + 65, + 252, + 137 + ], + "type": "image", + "image_path": "c5c729fb58a120a7c4b9b09e97194b28978b2f7c66e2e7beb3395396282b25e8.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 115, + 138, + 197, + 147 + ], + "lines": [ + { + "bbox": [ + 115, + 138, + 197, + 147 + ], + "spans": [ + { + "bbox": [ + 115, + 138, + 197, + 147 + ], + "type": "text", + "content": "(a) Image quality distribution" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 65, + 155, + 248, + 220 + ], + "blocks": [ + { + "bbox": [ + 65, + 155, + 248, + 220 + ], + "lines": [ + { + "bbox": [ + 65, + 155, + 248, + 220 + ], + "spans": [ + { + "bbox": [ + 65, + 155, + 248, + 220 + ], + "type": "image", + "image_path": "8ae208a61353f51058e45f49014aa986ea1c2b064485a4f62d0a05ee160c2b33.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 120, + 224, + 201, + 232 + ], + "lines": [ + { + "bbox": [ + 120, + 224, + 201, + 232 + ], + "spans": [ + { + "bbox": [ + 120, + 224, + 201, + 232 + ], + "type": "text", + "content": "(b) Video quality distribution" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 47, + 232, + 286, + 251 + ], + "lines": [ + { + "bbox": [ + 47, + 232, + 286, + 251 + ], + "spans": [ + { + "bbox": [ + 47, + 232, + 286, + 251 + ], + "type": "text", + "content": "Figure 4. Dataset quality distribution. The metrics used are BRISQUE [42] and VSFA [33] respectively." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 260, + 154, + 271 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 260, + 154, + 271 + ], + "spans": [ + { + "bbox": [ + 47, + 260, + 154, + 271 + ], + "type": "text", + "content": "signs are in the Appendix." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 281, + 238, + 295 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 281, + 238, + 295 + ], + "spans": [ + { + "bbox": [ + 47, + 281, + 238, + 295 + ], + "type": "text", + "content": "4. Statistical Analysis of CelebV-Text" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 301, + 287, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 301, + 287, + 389 + ], + "spans": [ + { + "bbox": [ + 46, + 301, + 287, + 389 + ], + "type": "text", + "content": "In this section, we compare CelebV-Text with the two most relevant and representative face video datasets [19,66]. We perform a comprehensive analysis of CelebV-Text in terms of video, text, and text-video relevance. To verify the effectiveness of our designed grammar templates, we generate text descriptions for CelebV-HQ based on its attributes for comparison. For simplicity, we use \"CelebV-HQ\" to denote this variant in the following." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 397, + 162, + 410 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 397, + 162, + 410 + ], + "spans": [ + { + "bbox": [ + 47, + 397, + 162, + 410 + ], + "type": "text", + "content": "4.1. Video Comparisons" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 415, + 287, + 591 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 415, + 287, + 591 + ], + "spans": [ + { + "bbox": [ + 46, + 415, + 287, + 591 + ], + "type": "text", + "content": "We briefly compare the overall statistics of existing face video datasets [11, 19, 60, 66] in Table 1. As reported, CelebV-Text contains 70,000 video clips with a total duration of around 279 hours. Each video is accompanied by 20 sentences describing all 6 designed attributes. Compared to CelebV [60], CelebV-Text has a larger scale and higher resolution. Although VoxCeleb2 [11] has more samples than CelebV-Text, its video distribution is limited as most videos are mainly talking faces. Moreover, video samples of both CelebV-HQ [66] and CelebV-Text are collected in open-world with diverse queries so that they are rich in distribution, while CelebV-Text has about 2 times video data, more video attributes, and highly relevant text descriptions. Finally, compared to the only existing facial text-video dataset MM-Vox [19], CelebV-Text overpasses MM-Vox in terms of scale and quality." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 591, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 591, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 591, + 287, + 713 + ], + "type": "text", + "content": "Attributes Distribution. In order to better present the distribution of different attributes in CelebV-Text, we pick and divide general appearance, action, and light direction attributes into groups. More distributions and division designs are provided in the Appendix. Specifically, all 40 general appearance classes are divided into 5 groups shown in Figure 3 (a). Facial features (e.g., double chin, big nose, and oval face) account for the most portion around " + }, + { + "bbox": [ + 46, + 591, + 287, + 713 + ], + "type": "inline_equation", + "content": "45\\%" + }, + { + "bbox": [ + 46, + 591, + 287, + 713 + ], + "type": "text", + "content": ". The elementary group is twice large than the beard type, accounting for around " + }, + { + "bbox": [ + 46, + 591, + 287, + 713 + ], + "type": "inline_equation", + "content": "25\\%" + }, + { + "bbox": [ + 46, + 591, + 287, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 591, + 287, + 713 + ], + "type": "inline_equation", + "content": "12\\%" + }, + { + "bbox": [ + 46, + 591, + 287, + 713 + ], + "type": "text", + "content": ", respectively. Fewer samples are located to the hairstyle and accessories groups," + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 316, + 67, + 424, + 143 + ], + "blocks": [ + { + "bbox": [ + 316, + 67, + 424, + 143 + ], + "lines": [ + { + "bbox": [ + 316, + 67, + 424, + 143 + ], + "spans": [ + { + "bbox": [ + 316, + 67, + 424, + 143 + ], + "type": "image", + "image_path": "8f1c27f9666ba9bebd95f0a8a9d2ec53224c929a45807271dd4d9d8d189191b1.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 333, + 144, + 417, + 151 + ], + "lines": [ + { + "bbox": [ + 333, + 144, + 417, + 151 + ], + "spans": [ + { + "bbox": [ + 333, + 144, + 417, + 151 + ], + "type": "text", + "content": "(a) Number of word distribution" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 427, + 68, + 536, + 143 + ], + "blocks": [ + { + "bbox": [ + 427, + 68, + 536, + 143 + ], + "lines": [ + { + "bbox": [ + 427, + 68, + 536, + 143 + ], + "spans": [ + { + "bbox": [ + 427, + 68, + 536, + 143 + ], + "type": "image", + "image_path": "0fcc4b2effa71e2482d61d55b7b5b1005caf599f0c78b81d51fb0e376a4f946f.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 459, + 144, + 522, + 152 + ], + "lines": [ + { + "bbox": [ + 459, + 144, + 522, + 152 + ], + "spans": [ + { + "bbox": [ + 459, + 144, + 522, + 152 + ], + "type": "text", + "content": "(b) 4-Grams distribution" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 495, + 222, + 518, + 228 + ], + "lines": [ + { + "bbox": [ + 495, + 222, + 518, + 228 + ], + "spans": [ + { + "bbox": [ + 495, + 222, + 518, + 228 + ], + "type": "text", + "content": "MM-Vox" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 322, + 154, + 359, + 191 + ], + "blocks": [ + { + "bbox": [ + 322, + 154, + 359, + 191 + ], + "lines": [ + { + "bbox": [ + 322, + 154, + 359, + 191 + ], + "spans": [ + { + "bbox": [ + 322, + 154, + 359, + 191 + ], + "type": "image", + "image_path": "a84628f0117593e033e8be4469d122b0db503a029981c50b4f5c21a92a69dffa.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 361, + 156, + 529, + 192 + ], + "lines": [ + { + "bbox": [ + 361, + 156, + 529, + 192 + ], + "spans": [ + { + "bbox": [ + 361, + 156, + 529, + 192 + ], + "type": "text", + "content": "This man with grey hair has a big nose and bags under eyes. He is wearing eyeglasses and earrings, with freckles above the eye areas. The man is firstly gazing for a short time and then he wags his head for a short time, and finally he keeps gazing for the rest of the time. He firstly remains neutral for a long time and then he turns angry for the rest of the time. The video is slightly dark in warm light. The light direction is front lighting for the whole time. CelebV-Text" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_footnote" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 323, + 196, + 361, + 232 + ], + "blocks": [ + { + "bbox": [ + 323, + 196, + 361, + 232 + ], + "lines": [ + { + "bbox": [ + 323, + 196, + 361, + 232 + ], + "spans": [ + { + "bbox": [ + 323, + 196, + 361, + 232 + ], + "type": "image", + "image_path": "c9689ecab2ebe6ca19b83c6809ae2af7c720137dc660b41a353b2d6903cfadb8.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 367, + 204, + 511, + 224 + ], + "lines": [ + { + "bbox": [ + 367, + 204, + 511, + 224 + ], + "spans": [ + { + "bbox": [ + 367, + 204, + 511, + 224 + ], + "type": "text", + "content": "a man has bags under eyes , receding hairline and straight hairhe is wearing goatee . he is chubby . he has beard , archedeyebrows and black hair ." + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 398, + 236, + 473, + 243 + ], + "lines": [ + { + "bbox": [ + 398, + 236, + 473, + 243 + ], + "spans": [ + { + "bbox": [ + 398, + 236, + 473, + 243 + ], + "type": "text", + "content": "(c) Video and text examples" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 306, + 244, + 544, + 265 + ], + "lines": [ + { + "bbox": [ + 306, + 244, + 544, + 265 + ], + "spans": [ + { + "bbox": [ + 306, + 244, + 544, + 265 + ], + "type": "text", + "content": "Figure 5. Text distribution. CelebV-Text achieves better performance in both 4-gram and number words distribution." + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 280, + 545, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 280, + 545, + 380 + ], + "spans": [ + { + "bbox": [ + 304, + 280, + 545, + 380 + ], + "type": "text", + "content": "taking around " + }, + { + "bbox": [ + 304, + 280, + 545, + 380 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 304, + 280, + 545, + 380 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 280, + 545, + 380 + ], + "type": "inline_equation", + "content": "8\\%" + }, + { + "bbox": [ + 304, + 280, + 545, + 380 + ], + "type": "text", + "content": ", respectively. Besides, action attributes are divided into 5 groups in Figure 3 (b), where it is clear that head-related actions account for the largest portion of around " + }, + { + "bbox": [ + 304, + 280, + 545, + 380 + ], + "type": "inline_equation", + "content": "60\\%" + }, + { + "bbox": [ + 304, + 280, + 545, + 380 + ], + "type": "text", + "content": ", followed by eyes-related actions of around " + }, + { + "bbox": [ + 304, + 280, + 545, + 380 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 304, + 280, + 545, + 380 + ], + "type": "text", + "content": ". The interaction group (e.g., eat), feeling group (e.g., smile), and daily group (e.g., sleep) account for around " + }, + { + "bbox": [ + 304, + 280, + 545, + 380 + ], + "type": "inline_equation", + "content": "9\\%" + }, + { + "bbox": [ + 304, + 280, + 545, + 380 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 280, + 545, + 380 + ], + "type": "inline_equation", + "content": "7\\%" + }, + { + "bbox": [ + 304, + 280, + 545, + 380 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 280, + 545, + 380 + ], + "type": "inline_equation", + "content": "4\\%" + }, + { + "bbox": [ + 304, + 280, + 545, + 380 + ], + "type": "text", + "content": ", respectively. Finally, for light directions (Figure 3 (c)), most samples contain the front lighting and the remaining ones are evenly distributed." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 304, + 380, + 545, + 513 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 380, + 545, + 513 + ], + "spans": [ + { + "bbox": [ + 304, + 380, + 545, + 513 + ], + "type": "text", + "content": "Video Quality Distribution. We follow [66] to analyze the quality of our collected videos. To demonstrate the superiority of CelebV-Text, we compare with MM-Vox [19] and CelebV-HQ [66], where mean BRISQUE [42] and VSFA [33] are used to evaluate the image and video quality, respectively. Image quality of all datasets is shown in Figure 4 (a), where CelebV-Text and CelebV-HQ achieve comparable quality, higher than MM-Vox by a large margin. Video quality of all datasets is shown in Figure 4 (b), where CelebV-Text has the best quality, which is due to the effect of the video split method mentioned in Section 3.1, alleviating the discontinuity during background transitions." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 306, + 519, + 413, + 532 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 519, + 413, + 532 + ], + "spans": [ + { + "bbox": [ + 306, + 519, + 413, + 532 + ], + "type": "text", + "content": "4.2. Text Comparisons" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 304, + 536, + 545, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 536, + 545, + 624 + ], + "spans": [ + { + "bbox": [ + 304, + 536, + 545, + 624 + ], + "type": "text", + "content": "In addition to a large number of video samples, text descriptions of CelebV-Text are longer and more detailed than those in MM-Vox [19] and CelebV-HQ [66] (see Figure 5 (a)), where the average text length of MM-Vox, CelebV-HQ, and CelebV-Text are 28.39, 31.06, and 67.15. Distributions of Celeb-HQ and MM-Vox are close, but there are more words in CelebV-Text to describe a video due to the comprehensive annotation." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 304, + 624, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 624, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 624, + 545, + 713 + ], + "type": "text", + "content": "To validate the linguistic diversity of the generated texts, comparisons are conducted among the three datasets following [56]. Specifically, we report the unique part-of-speech (POS) tags (i.e., verb, noun, adjective, and adverb) of the three datasets in Table 3. Obviously, due to our comprehensively designed attribute list and the number of templates, CelebV-Text presents a wider variety of text styles, covering a broader range of face attributes that are static or" + } + ] + } + ], + "index": 26 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14809" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 49, + 95, + 547, + 187 + ], + "blocks": [ + { + "bbox": [ + 47, + 70, + 547, + 91 + ], + "lines": [ + { + "bbox": [ + 47, + 70, + 547, + 91 + ], + "spans": [ + { + "bbox": [ + 47, + 70, + 547, + 91 + ], + "type": "text", + "content": "Table 2. Multimodal retrieval results. Clip2Video [17] is leveraged to measure the text-video relevance via retrieval experiments. Bold values indicate the best results, underlined ones indicate the second best." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 49, + 95, + 547, + 187 + ], + "lines": [ + { + "bbox": [ + 49, + 95, + 547, + 187 + ], + "spans": [ + { + "bbox": [ + 49, + 95, + 547, + 187 + ], + "type": "table", + "html": "
Text ⇒ VideoVideo ⇒ Text
DescriptionDatasetR@1(↑)R@5(↑)R@10(↑)MdR(↓)MnR(↓)R@1(↑)R@5(↑)R@10(↑)MdR(↓)MnR(↓)
(a) App.MM-Vox [19]1.59.015.752.068.82.09.214.643.057.8
CelebV-HQ [66]5.919.229.727.052.27.220.732.427.046.9
CelebV-Text6.121.335.526.349.17.420.729.926.648.3
(b) App.+Emo.CelebV-HQ [66]6.520.130.825.048.07.925.538.817.037.0
CelebV-Text6.623.437.126.047.68.127.234.718.238.3
(c) App.+Emo.+Act.CelebV-Text6.924.139.225.846.78.027.637.116.736.1
", + "image_path": "e053a57d299a3b0c6dfaa09f338e98fa83362fcd2c5eb7070d247701db3515de.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 69, + 220, + 268, + 271 + ], + "blocks": [ + { + "bbox": [ + 47, + 197, + 287, + 218 + ], + "lines": [ + { + "bbox": [ + 47, + 197, + 287, + 218 + ], + "spans": [ + { + "bbox": [ + 47, + 197, + 287, + 218 + ], + "type": "text", + "content": "Table 3. Number of unique POS tags. The numbers of unique POS tags for MM-Vox, CelebV-HQ, and CelebV-Text." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 69, + 220, + 268, + 271 + ], + "lines": [ + { + "bbox": [ + 69, + 220, + 268, + 271 + ], + "spans": [ + { + "bbox": [ + 69, + 220, + 268, + 271 + ], + "type": "table", + "html": "
Dataset#Verb#Adj.#Noun#Adv.
MM-Vox [19]520380
CelebV-HQ [66]1024506
CelebV-Text967817424
", + "image_path": "39f3b3b671181fbce26fb6f2f7293d5e439754ca18dad09257ba3b3c52aa18bd.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 284, + 183, + 296 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 284, + 183, + 296 + ], + "spans": [ + { + "bbox": [ + 47, + 284, + 183, + 296 + ], + "type": "text", + "content": "dynamic in the temporal domain." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 300, + 287, + 400 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 300, + 287, + 400 + ], + "spans": [ + { + "bbox": [ + 46, + 300, + 287, + 400 + ], + "type": "text", + "content": "In addition, we further examine the naturalness and complexity of our texts compared to MM-Vox, where we modify [63] to calculate the type-token vocabulary curve for all captions. As shown in Figure 5 (b) where unique 4-grams are selected as the types [56], it is evident that due to our grammar structures and synonym replacement, the linguistic naturalness (vocabulary use) and complexity (vocabulary size) of our CelebV-Text are much better. Please refer to Appendix for more " + }, + { + "bbox": [ + 46, + 300, + 287, + 400 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 46, + 300, + 287, + 400 + ], + "type": "text", + "content": "-grams results." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 421, + 171, + 432 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 421, + 171, + 432 + ], + "spans": [ + { + "bbox": [ + 47, + 421, + 171, + 432 + ], + "type": "text", + "content": "4.3. Text-Video Relevance" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 443, + 287, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 443, + 287, + 567 + ], + "spans": [ + { + "bbox": [ + 46, + 443, + 287, + 567 + ], + "type": "text", + "content": "To quantitatively validate our text-video relevance, we conduct text-video retrieval tasks on three datasets: MMVox [19], CelebV-HQ [66], and CelebV-Text. Rather than use conventional frame-wise clip score as most works [24, 51, 55], we follow [17] to compute feature similarities between texts and videos with the consideration of temporal dynamics, which reflects accurate multimodal interactions across the two modalities. Recall at rank K (R@K), median rank (MdR), and mean rank (MnR) [17, 41, 64] are used as evaluation metrics, where the higher R@K, the lower median rank and mean rank indicate better performance." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "type": "text", + "content": "We first examine the performance given texts with descriptions of general appearance in Table 2 (a). Results of CelebV-HQ and CelebV-Text are both better than MM-Vox for two retrieval tasks, which indicates our designed templates can produce texts more relevant to videos than MM-Vox. We further add descriptions about dynamic emotion changes to CelebV-HQ and CelebV-Text in Table 2 (b). Similar results are achieved in both datasets, which reflects that our annotation accuracy on static appearance attributes is as good as CelebV-HQ. Finally, we append action descriptions to CelebV-Text in Table 2 (c), which achieves the best performance on most metrics, verifying the relevance between our generated texts and video samples." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 198, + 383, + 212 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 198, + 383, + 212 + ], + "spans": [ + { + "bbox": [ + 305, + 198, + 383, + 212 + ], + "type": "text", + "content": "5. Experiment" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 217, + 547, + 262 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 217, + 547, + 262 + ], + "spans": [ + { + "bbox": [ + 304, + 217, + 547, + 262 + ], + "type": "text", + "content": "In this section, we first conduct facial text-to-video generation to validate the effectiveness of our CelebV-Text dataset. We then benchmark representative approaches on facial text-to-video generation task." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 266, + 522, + 280 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 266, + 522, + 280 + ], + "spans": [ + { + "bbox": [ + 305, + 266, + 522, + 280 + ], + "type": "text", + "content": "5.1. High-relevance Text-to-Video Generation" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 285, + 545, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 285, + 545, + 472 + ], + "spans": [ + { + "bbox": [ + 304, + 285, + 545, + 472 + ], + "type": "text", + "content": "To show the benefits brought by our text descriptions which depict both static and dynamic attributes, we conduct experiments to show the effectiveness of CelebV-Text. Experiments are mainly based on a recent open-sourced state-of-the-art method, MMVID [19], and compared with " + }, + { + "bbox": [ + 304, + 285, + 545, + 472 + ], + "type": "inline_equation", + "content": "\\mathrm{CogVideo}^2" + }, + { + "bbox": [ + 304, + 285, + 545, + 472 + ], + "type": "text", + "content": " [26], which is a large-scale pretrained text-to-video model, trained on millions of text-image/video pairs. Static Face Video Generation. To validate the effectiveness of our facial text-video dataset in static attributes, we use the models stated above to generate videos conditioned on general appearance, face details, and light conditions descriptions, respectively. Specifically, we first train MMVID [19] from scratch solely on CelebV-Text. We then generate 3 input texts including individual descriptions of each of the static attributes. Generated texts are fed into both MMVID [19] and " + }, + { + "bbox": [ + 304, + 285, + 545, + 472 + ], + "type": "inline_equation", + "content": "\\mathrm{CogVideo}" + }, + { + "bbox": [ + 304, + 285, + 545, + 472 + ], + "type": "text", + "content": " [26] and corresponding video outputs are examined." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 472, + 545, + 560 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 472, + 545, + 560 + ], + "spans": [ + { + "bbox": [ + 304, + 472, + 545, + 560 + ], + "type": "text", + "content": "Visualization results of general appearance are shown in Figure 6 (a), which prove the effectiveness of our dataset. We observe that although CogVideo can output the face video given a text description, the text-video pair is not quite relevant, such as \"bags under eyes\" and \"wavy hair\". However, MMVID [19] produces videos with high relevance to input texts, containing all attributes described in the text. More results are shown in the Appendix." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 560, + 545, + 648 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 560, + 545, + 648 + ], + "spans": [ + { + "bbox": [ + 304, + 560, + 545, + 648 + ], + "type": "text", + "content": "Dynamic Face Video Generation. We follow the above experimental setting to validate the effectiveness of our dataset with dynamic attribute changes (i.e., emotion, action and light direction). Due to the difficulty in modelling state change [27, 51], we follow [5] to apply test-time interpolation to MMVID [19], named MMVID-interp, to improve the text encoding and better understand the dynamics. Details of our modification are shown in the Appendix." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 648, + 547, + 671 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 648, + 547, + 671 + ], + "spans": [ + { + "bbox": [ + 305, + 648, + 547, + 671 + ], + "type": "text", + "content": "In Figure 6 (b), we observe that CogVideo fails to reflect the temporal change described in the input text, i.e., smile" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 305, + 676, + 547, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 676, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 676, + 547, + 713 + ], + "type": "text", + "content": "2We choose CogVideo [26] as the representative large-scale model for comparison, since the inference code and pretrained models of other large-scale methods (e.g., CogVideo [26], Phenaki [55],Imagen Video [24], and Make-A-Video [51]) are not public." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "14810" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 83, + 64, + 109, + 87 + ], + "blocks": [ + { + "bbox": [ + 83, + 64, + 109, + 87 + ], + "lines": [ + { + "bbox": [ + 83, + 64, + 109, + 87 + ], + "spans": [ + { + "bbox": [ + 83, + 64, + 109, + 87 + ], + "type": "image", + "image_path": "efae64ba66f72d93c5ff2220733fd9f97a67492b842e56d1eb24ba07da6c4fbf.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 110, + 70, + 445, + 81 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 110, + 70, + 445, + 81 + ], + "spans": [ + { + "bbox": [ + 110, + 70, + 445, + 81 + ], + "type": "text", + "content": "The man has black hair and bags under eyes. He has a beard, goatee and arched eyebrows." + } + ] + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 63, + 89, + 530, + 211 + ], + "blocks": [ + { + "bbox": [ + 63, + 89, + 530, + 211 + ], + "lines": [ + { + "bbox": [ + 63, + 89, + 530, + 211 + ], + "spans": [ + { + "bbox": [ + 63, + 89, + 530, + 211 + ], + "type": "image", + "image_path": "ba52a7552e850638ece4b6b6087f7e53b959ae5de42549044c9215ee4b796997.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 247, + 213, + 365, + 223 + ], + "lines": [ + { + "bbox": [ + 247, + 213, + 365, + 223 + ], + "spans": [ + { + "bbox": [ + 247, + 213, + 365, + 223 + ], + "type": "text", + "content": "(a) Static - General Appearance" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 82, + 222, + 107, + 244 + ], + "blocks": [ + { + "bbox": [ + 82, + 222, + 107, + 244 + ], + "lines": [ + { + "bbox": [ + 82, + 222, + 107, + 244 + ], + "spans": [ + { + "bbox": [ + 82, + 222, + 107, + 244 + ], + "type": "image", + "image_path": "74b2b38e3ce9c9cf64c6b6bbcf306e44b89fed6a6b604f4806da868eab81ce21.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 108, + 228, + 458, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 108, + 228, + 458, + 239 + ], + "spans": [ + { + "bbox": [ + 108, + 228, + 458, + 239 + ], + "type": "text", + "content": "The young woman is wearing lipstick and earrings. She firstly smiles and then turns her head." + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 62, + 245, + 529, + 423 + ], + "blocks": [ + { + "bbox": [ + 62, + 245, + 529, + 423 + ], + "lines": [ + { + "bbox": [ + 62, + 245, + 529, + 423 + ], + "spans": [ + { + "bbox": [ + 62, + 245, + 529, + 423 + ], + "type": "image", + "image_path": "3932e1044d7d61834cb037e35b9f0c6004423953232f3ed597b93f856772e966.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 266, + 425, + 346, + 434 + ], + "lines": [ + { + "bbox": [ + 266, + 425, + 346, + 434 + ], + "spans": [ + { + "bbox": [ + 266, + 425, + 346, + 434 + ], + "type": "text", + "content": "(b) Dynamic - Action" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 435, + 544, + 455 + ], + "lines": [ + { + "bbox": [ + 46, + 435, + 544, + 455 + ], + "spans": [ + { + "bbox": [ + 46, + 435, + 544, + 455 + ], + "type": "text", + "content": "Figure 6. Qualitative results of facial text-to-video generation. The generated samples are given texts describing (a) the static attribute and (b) dynamic attribute." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 49, + 484, + 282, + 546 + ], + "blocks": [ + { + "bbox": [ + 59, + 461, + 532, + 472 + ], + "lines": [ + { + "bbox": [ + 59, + 461, + 532, + 472 + ], + "spans": [ + { + "bbox": [ + 59, + 461, + 532, + 472 + ], + "type": "text", + "content": "Table 4. Benchmark of text-to-video generation on different datasets. " + }, + { + "bbox": [ + 59, + 461, + 532, + 472 + ], + "type": "inline_equation", + "content": "\\downarrow" + }, + { + "bbox": [ + 59, + 461, + 532, + 472 + ], + "type": "text", + "content": " means a lower value is better and " + }, + { + "bbox": [ + 59, + 461, + 532, + 472 + ], + "type": "inline_equation", + "content": "\\uparrow" + }, + { + "bbox": [ + 59, + 461, + 532, + 472 + ], + "type": "text", + "content": " means the opposite." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 69, + 473, + 260, + 482 + ], + "lines": [ + { + "bbox": [ + 69, + 473, + 260, + 482 + ], + "spans": [ + { + "bbox": [ + 69, + 473, + 260, + 482 + ], + "type": "text", + "content": "(a) Quantitative results on general appearance descriptions." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 49, + 484, + 282, + 546 + ], + "lines": [ + { + "bbox": [ + 49, + 484, + 282, + 546 + ], + "spans": [ + { + "bbox": [ + 49, + 484, + 282, + 546 + ], + "type": "table", + "html": "
DatasetMethodFVD(↓)FID(↓)CLIPSIM(↑)
MM-Vox [19]TFGAN [5]502.28 ± 1.66760.24 ± 16.010.165 ± 0.022
MMVID [19]65.79 ± 1.8138.81 ± 3.660.170 ± 0.020
CelebV-HQ [66]TFGAN [5]428.04 ± 1.76616.24 ± 17.450.168 ± 0.021
MMVID [19]73.65 ± 1.4363.86 ± 3.660.172 ± 0.019
CelebV-TextTFGAN [5]403.04 ± 1.34589.24 ± 16.460.177 ± 0.012
MMVID [19]66.69 ± 1.3558.70 ± 4.670.198 ± 0.014
", + "image_path": "2a5b0c7d0e2c1351f7e2b85915032c0ba9b614092494afbc88176411b50665db.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 554, + 286, + 643 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 554, + 286, + 643 + ], + "spans": [ + { + "bbox": [ + 46, + 554, + 286, + 643 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 46, + 554, + 286, + 643 + ], + "type": "text", + "content": " turn. However, both MMVID [19] and MMVID-interp trained on CelebV-Text can successfully model the dynamic attribute changes, which demonstrates the effectiveness of our dataset. In addition, we find that MMVID [19] cannot preserve some attributes well (e.g., earrings), while MMVID-interp can stabilize the sampling process, validating the effectiveness of our modification. More results are shown in the Appendix." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 46, + 646, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 646, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 646, + 286, + 713 + ], + "type": "text", + "content": "Note that CogVideo [26] has a much larger model size (" + }, + { + "bbox": [ + 46, + 646, + 286, + 713 + ], + "type": "inline_equation", + "content": "\\sim 100" + }, + { + "bbox": [ + 46, + 646, + 286, + 713 + ], + "type": "text", + "content": " times larger than MMVID [19]) and is trained on much large text-video data (" + }, + { + "bbox": [ + 46, + 646, + 286, + 713 + ], + "type": "inline_equation", + "content": "\\sim 75" + }, + { + "bbox": [ + 46, + 646, + 286, + 713 + ], + "type": "text", + "content": " times larger than CelebV-Text). However, video samples produced by CogVideo [26] shown in Figure 6 are of a lower quality than the ones by MMVID [19] trained solely on CelebV-Text, where gener" + } + ] + } + ], + "index": 13 + }, + { + "type": "table", + "bbox": [ + 310, + 484, + 541, + 546 + ], + "blocks": [ + { + "bbox": [ + 321, + 473, + 529, + 482 + ], + "lines": [ + { + "bbox": [ + 321, + 473, + 529, + 482 + ], + "spans": [ + { + "bbox": [ + 321, + 473, + 529, + 482 + ], + "type": "text", + "content": "(b) Quantitative results on dynamic descriptions of CelebV-Text." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 310, + 484, + 541, + 546 + ], + "lines": [ + { + "bbox": [ + 310, + 484, + 541, + 546 + ], + "spans": [ + { + "bbox": [ + 310, + 484, + 541, + 546 + ], + "type": "table", + "html": "
DatasetMethodFVD(↓)FID(↓)CLIPSIM(↑)
CelebV-TextTFGAN [5]442.30 ± 2.56623.17 ± 18.880.158 ± 0.024
App.+Emo.MMVID [19]82.78 ± 1.4761.58 ± 3.990.176 ± 0.008
MMVID-interp72.87 ± 1.2341.57 ± 3.560.182 ± 0.010
CelebV-TextTFGAN [5]571.34 ± 4.54784.93 ± 20.130.154 ± 0.028
App.+Act.MMVID [19]109.25 ± 2.1182.55 ± 4.370.174 ± 0.019
MMVID-interp80.81 ± 2.5570.88 ± 4.770.176 ± 0.020
", + "image_path": "2839e4c4536ba0d9f794d000acc62263c4f09bdaa3205a831d9b62686d925b6b.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "table_body" + } + ], + "index": 15 + }, + { + "bbox": [ + 305, + 555, + 544, + 577 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 555, + 544, + 577 + ], + "spans": [ + { + "bbox": [ + 305, + 555, + 544, + 577 + ], + "type": "text", + "content": "ated faces are not in a high relevance to input texts, demonstrating the effectiveness of our facial text-video dataset." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 582, + 545, + 594 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 582, + 545, + 594 + ], + "spans": [ + { + "bbox": [ + 306, + 582, + 545, + 594 + ], + "type": "text", + "content": "5.2. Benchmark on Facial Text-to-Video Generation" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 599, + 545, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 599, + 545, + 688 + ], + "spans": [ + { + "bbox": [ + 304, + 599, + 545, + 688 + ], + "type": "text", + "content": "As the domain of text-to-video generation is currently thriving, there exists only one benchmark in the face domain, MM-Vox [19]. We expand [19] and construct a benchmark of facial text-to-video generation tasks on three datasets: MM-Vox [19], CelebV-HQ [66] with texts generated by our templates, and CelebV-Text. We choose two representative methods3, TFGAN [5] and MMVID [19], to evaluate their performances on all datasets." + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 306, + 693, + 545, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 693, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 306, + 693, + 545, + 712 + ], + "type": "text", + "content": "3Other methods, e.g., CogVideo [26], Phenaki [55], Imagen Video [24], and Make-A-Video [51] are not included since their training codes are not" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14811" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 63, + 67, + 87, + 89 + ], + "blocks": [ + { + "bbox": [ + 63, + 67, + 87, + 89 + ], + "lines": [ + { + "bbox": [ + 63, + 67, + 87, + 89 + ], + "spans": [ + { + "bbox": [ + 63, + 67, + 87, + 89 + ], + "type": "image", + "image_path": "49b1fc43a2f5568f612a67090fb86cc277c7a252e1dda53770aee87931ba751f.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 88, + 67, + 510, + 89 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 88, + 67, + 510, + 89 + ], + "spans": [ + { + "bbox": [ + 88, + 67, + 510, + 89 + ], + "type": "text", + "content": "He has bushy eyebrows, beard and wavy hair. He has got 5 o'clock shadow and brown hair. He has bags under eyes and sideburns with mustache." + } + ] + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 52, + 91, + 521, + 269 + ], + "blocks": [ + { + "bbox": [ + 52, + 91, + 521, + 269 + ], + "lines": [ + { + "bbox": [ + 52, + 91, + 521, + 269 + ], + "spans": [ + { + "bbox": [ + 52, + 91, + 521, + 269 + ], + "type": "image", + "image_path": "abe76c6b731b63028efed25c276ff8711a484ae88577c50204f564b470ab8381.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 47, + 270, + 545, + 292 + ], + "lines": [ + { + "bbox": [ + 47, + 270, + 545, + 292 + ], + "spans": [ + { + "bbox": [ + 47, + 270, + 545, + 292 + ], + "type": "text", + "content": "Figure 7. Qualitative results on three facial text-video datasets. Red and yellow regions indicate the missing of \"bags under eyes\" and the existence of \"wavy hair\" and \"bags under eyes\"." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 300, + 287, + 543 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 300, + 287, + 543 + ], + "spans": [ + { + "bbox": [ + 46, + 300, + 287, + 543 + ], + "type": "text", + "content": "Quantitative Results. For thorough benchmark construction, we evaluate baseline methods given variant texts including static and dynamic attributes. We use FVD [53] (temporal consistency), FID [23] (individual frame quality), and CLIPSIM [57] (text-video relevance) as evaluation metrics following [19] and report detailed results for appearance, action, and emotion in Table 4. Evaluation steps are repeated over ten runs with mean values and standard errors reported as well. All other values are shown in the Appendix. It can be seen from Table 4 that MMVID [19] obtains good FVD/FID/CLIPSIM metrics over TFGAN [5] which fails to generate reasonable video outputs. In addition, when input texts contain descriptions about a dynamic state change in the temporal domain, the generated video quality by MMVID [19] decreases, which encourages future methods to focus more on cross-modal understanding and consistent video generation. Moreover, the performance of MMVID-interp is better than MMVID [19] on all metrics, validating the effectiveness of our modification mentioned in Section 5.1. Due to challenges posed by our dataset and text-to-video generation task, there is still considerable room to improve." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 543, + 287, + 643 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 543, + 287, + 643 + ], + "spans": [ + { + "bbox": [ + 46, + 543, + 287, + 643 + ], + "type": "text", + "content": "Qualitative Results. Video samples generated from MMVID [19] trained on different datasets are shown in Figure 7, where all video frames are of " + }, + { + "bbox": [ + 46, + 543, + 287, + 643 + ], + "type": "inline_equation", + "content": "128^2" + }, + { + "bbox": [ + 46, + 543, + 287, + 643 + ], + "type": "text", + "content": ". We can see that video samples generated by MMVID [19] trained on different datasets are of high quality with temporal consistency. However, MMVID [19] trained on MM-Vox [19] can sometimes fail to generate attributes mentioned in the input texts. More generated video samples with dynamic attribute changes are shown in the Appendix." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 653, + 117, + 665 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 653, + 117, + 665 + ], + "spans": [ + { + "bbox": [ + 47, + 653, + 117, + 665 + ], + "type": "text", + "content": "6. Discussion" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 672, + 287, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 672, + 287, + 696 + ], + "spans": [ + { + "bbox": [ + 47, + 672, + 287, + 696 + ], + "type": "text", + "content": "We have proposed CelebV-Text, a large-scale, high-quality, and diverse facial text-video dataset with static and" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 300, + 545, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 300, + 545, + 422 + ], + "spans": [ + { + "bbox": [ + 306, + 300, + 545, + 422 + ], + "type": "text", + "content": "dynamic attributes. CelebV-Text contains 70,000 video clips, each of which is accompanied by 20 individual sentences describing both static and dynamic factors. Through extensive statistical analysis and experiments, we have demonstrated the superiority and effectiveness of CelebV-Text. In the future, we plan to further enlarge CelebV-Text in both scale and diversity. We may further explore several new tasks based on CelebV-Text, such as fine-grained control of video face, adaptation of general pretrained models to the face domain, and text-driven 3D-aware facial video generation." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 429, + 546, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 429, + 546, + 639 + ], + "spans": [ + { + "bbox": [ + 304, + 429, + 546, + 639 + ], + "type": "text", + "content": "Ethical Consideration. CelebV-Text is intended for research purposes only. While the raw videos will not be released, the data annotations, links to raw videos, and data processing tools will be made available after undergoing a rigorous legality check procedure at our institution. It is worth noting that our data annotation does not include any personal biometric information such as identity. Only generic attribute information such as gender, hair color, and motion is annotated. Additionally, synthetic videos generated in this work do not exhibit bias or certain biometric information (e.g., big lips or big nose), alleviating ethical concerns. CelebV-Text may be used for deepfakes, but it can also be used for forgery detection tasks to prevent such issues. We will try our best to control the application and acquisition procedure of CelebV-Text to avoid potential misuse and abuse. In the future, we plan to use synthetic face generation frameworks to generate synthetic face videos to address the ethical shortcomings of existing real-world face video datasets." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 646, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 646, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 646, + 545, + 713 + ], + "type": "text", + "content": "Acknowledgement. CelebV-Text is developed under OpenXDLab - an open platform for X-Dimension high-quality data. This study is supported by the RIE2020 Industry Alignment Fund Industry Collaboration Projects (IAF-ICP) Funding Initiative, in-kind contribution from the industry partner(s), and the MOE AcRF Tier 1 (RG16/21)." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 47, + 703, + 91, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 703, + 91, + 712 + ], + "spans": [ + { + "bbox": [ + 47, + 703, + 91, + 712 + ], + "type": "text", + "content": "public so far." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14812" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 286, + 712 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 53, + 91, + 286, + 120 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 286, + 120 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 286, + 120 + ], + "type": "text", + "content": "[1] Niki Aifanti, Christos Papachristou, and Anastasios Delopoulos. The mug facial expression database. In WIAMIS, 2010. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 121, + 286, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 121, + 286, + 161 + ], + "spans": [ + { + "bbox": [ + 53, + 121, + 286, + 161 + ], + "type": "text", + "content": "[2] Jean-Baptiste Alayrac, Piotr Bojanowski, Nishant Agrawal, Josef Sivic, Ivan Laptev, and Simon Lacoste-Julien. Unsupervised learning from narrated instruction videos. In CVPR, 2016. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 163, + 286, + 201 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 163, + 286, + 201 + ], + "spans": [ + { + "bbox": [ + 53, + 163, + 286, + 201 + ], + "type": "text", + "content": "[3] Lisa Anne Hendricks, Oliver Wang, Eli Shechtman, Josef Sivic, Trevor Darrell, and Bryan Russell. Localizing moments in video with natural language. In ICCV, 2017. 2, 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 204, + 286, + 234 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 204, + 286, + 234 + ], + "spans": [ + { + "bbox": [ + 53, + 204, + 286, + 234 + ], + "type": "text", + "content": "[4] Max Bain, Arsha Nagrani, Gül Varol, and Andrew Zisserman. Frozen in time: A joint video and image encoder for end-to-end retrieval. In ICCV, 2021. 1, 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 236, + 286, + 275 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 236, + 286, + 275 + ], + "spans": [ + { + "bbox": [ + 53, + 236, + 286, + 275 + ], + "type": "text", + "content": "[5] Yogesh Balaji, Martin Renqiang Min, Bing Bai, Rama Chellappa, and Hans Peter Graf. Conditional gan with discriminative filter generation for text-to-video synthesis. In IJCAI, 2019. 1, 2, 4, 6, 7, 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 277, + 286, + 306 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 277, + 286, + 306 + ], + "spans": [ + { + "bbox": [ + 53, + 277, + 286, + 306 + ], + "type": "text", + "content": "[6] Alex Bewley, Zongyuan Ge, Lionel Ott, Fabio Ramos, and Ben Upcroft. Simple online and realtime tracking. In ICIP. IEEE, 2016. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 308, + 286, + 337 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 308, + 286, + 337 + ], + "spans": [ + { + "bbox": [ + 53, + 308, + 286, + 337 + ], + "type": "text", + "content": "[7] Sergey Bezryadin, Pavel Bourov, and Dmitry Ilinih. Brightness calculation in digital image processing. In TDPF, 2007. 4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 339, + 286, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 339, + 286, + 369 + ], + "spans": [ + { + "bbox": [ + 53, + 339, + 286, + 369 + ], + "type": "text", + "content": "[8] Steven Bird, Ewan Klein, and Edward Loper. Natural language processing with Python: analyzing text with the natural language toolkit. \"O'Reilly Media, Inc.\", 2009. 4" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 370, + 286, + 391 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 370, + 286, + 391 + ], + "spans": [ + { + "bbox": [ + 53, + 370, + 286, + 391 + ], + "type": "text", + "content": "[9] David Chen and William B Dolan. Collecting highly parallel data for paraphrase evaluation. In ACL, 2011. 2, 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 392, + 286, + 421 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 392, + 286, + 421 + ], + "spans": [ + { + "bbox": [ + 48, + 392, + 286, + 421 + ], + "type": "text", + "content": "[10] Danqi Chen and Christopher D Manning. A fast and accurate dependency parser using neural networks. In EMNLP, 2014. 2, 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 422, + 286, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 422, + 286, + 453 + ], + "spans": [ + { + "bbox": [ + 48, + 422, + 286, + 453 + ], + "type": "text", + "content": "[11] Joon Son Chung, Arsha Nagrani, and Andrew Zisserman. Voxceleb2: Deep speaker recognition. In INTERSPEECH, 2018. 1, 2, 3, 5" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 453, + 286, + 483 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 453, + 286, + 483 + ], + "spans": [ + { + "bbox": [ + 48, + 453, + 286, + 483 + ], + "type": "text", + "content": "[12] Karl Cobbe, Oleg Klimov, Chris Hesse, Taehoon Kim, and John Schulman. Quantifying generalization in reinforcement learning. In ICML, 2019. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 485, + 286, + 535 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 485, + 286, + 535 + ], + "spans": [ + { + "bbox": [ + 48, + 485, + 286, + 535 + ], + "type": "text", + "content": "[13] Dima Damen, Hazel Doughty, Giovanni Maria Farinella, Antonino Furnari, Evangelos Kazakos, Jian Ma, Davide Moltisanti, Jonathan Munro, Toby Perrett, Will Price, et al. Rescaling egocentric vision: collection, pipeline and challenges for epic-kitchens-100. In IJCV, 2022. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 536, + 286, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 536, + 286, + 567 + ], + "spans": [ + { + "bbox": [ + 48, + 536, + 286, + 567 + ], + "type": "text", + "content": "[14] Jiankang Deng, Jia Guo, Niannan Xue, and Stefanos Zafeiriou. Arcface: Additive angular margin loss for deep face recognition. In CVPR, 2019. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 567, + 286, + 597 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 567, + 286, + 597 + ], + "spans": [ + { + "bbox": [ + 48, + 567, + 286, + 597 + ], + "type": "text", + "content": "[15] Kangle Deng, Tianyi Fei, Xin Huang, and Yuxin Peng. IRCAN: Introspective recurrent convolutional gan for text-to-video generation. In IJCAI, pages 2216-2222, 2019. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 48, + 599, + 286, + 628 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 599, + 286, + 628 + ], + "spans": [ + { + "bbox": [ + 48, + 599, + 286, + 628 + ], + "type": "text", + "content": "[16] Patrick Esser, Robin Rombach, and Bjorn Ommer. Taming transformers for high-resolution image synthesis. In CVPR, 2021. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 48, + 630, + 286, + 660 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 630, + 286, + 660 + ], + "spans": [ + { + "bbox": [ + 48, + 630, + 286, + 660 + ], + "type": "text", + "content": "[17] Han Fang, Pengfei Xiong, Luhui Xu, and Yu Chen. Clip2video: Mastering video-text retrieval via image clip. arXiv preprint arXiv:2106.11097, 2021. 2, 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 48, + 662, + 286, + 691 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 662, + 286, + 691 + ], + "spans": [ + { + "bbox": [ + 48, + 662, + 286, + 691 + ], + "type": "text", + "content": "[18] Tanmay Gupta, Dustin Schwenk, Ali Farhadi, Derek Hoiem, and Aniruddha Kembhavi. Imagine this! scripts to compositions to videos. In ECCV, 2018. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 48, + 692, + 286, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 692, + 286, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 692, + 286, + 712 + ], + "type": "text", + "content": "[19] Ligong Han, Jian Ren, Hsin-Ying Lee, Francesco Barbieri, Kyle Olszewski, Shervin Minaee, Dimitris Metaxas, and" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 39, + "blocks": [ + { + "bbox": [ + 327, + 73, + 545, + 103 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 545, + 103 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 545, + 103 + ], + "type": "text", + "content": "Sergey Tulyakov. Show me what and tell me how: Video synthesis via multimodal conditioning. In CVPR, 2022. 2, 3, 4, 5, 6, 7, 8" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 105, + 545, + 134 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 105, + 545, + 134 + ], + "spans": [ + { + "bbox": [ + 308, + 105, + 545, + 134 + ], + "type": "text", + "content": "[20] William Harvey, Saeid Naderiparizi, Vaden Masrani, Christian Weilbach, and Frank Wood. Flexible diffusion modeling of long videos. arXiv preprint arXiv:2205.11495, 2022. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 136, + 545, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 136, + 545, + 185 + ], + "spans": [ + { + "bbox": [ + 308, + 136, + 545, + 185 + ], + "type": "text", + "content": "[21] Thomas Hayes, Songyang Zhang, Xi Yin, Guan Pang, Sasha Sheng, Harry Yang, Songwei Ge, Isabelle Hu, and Devi Parikh. Mugen: A playground for video-audio-text multimodal understanding and generation. arXiv preprint arXiv:2204.08058, 2022. 2, 4" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 186, + 545, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 186, + 545, + 227 + ], + "spans": [ + { + "bbox": [ + 308, + 186, + 545, + 227 + ], + "type": "text", + "content": "[22] Javier Hernandez-Andres, Raymond L Lee, and Javier Romero. Calculating correlated color temperatures across the entire gamut of daylight and skylight chromaticities. In Applied optics, 1999. 4" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 228, + 545, + 267 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 228, + 545, + 267 + ], + "spans": [ + { + "bbox": [ + 308, + 228, + 545, + 267 + ], + "type": "text", + "content": "[23] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. NeurIPS, 2017. 8" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 269, + 545, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 269, + 545, + 319 + ], + "spans": [ + { + "bbox": [ + 308, + 269, + 545, + 319 + ], + "type": "text", + "content": "[24] Jonathan Ho, William Chan, Chitwan Sahara, Jay Whang, Ruiqi Gao, Alexey Gritsenko, Diederik P Kingma, Ben Poole, Mohammad Norouzi, David J Fleet, et al. Imagen video: High definition video generation with diffusion models. arXiv preprint arXiv:2210.02303, 2022. 2, 6, 7" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 320, + 545, + 350 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 320, + 545, + 350 + ], + "spans": [ + { + "bbox": [ + 308, + 320, + 545, + 350 + ], + "type": "text", + "content": "[25] Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan, Mohammad Norouzi, and David J Fleet. Video diffusion models. arXiv preprint arXiv:2204.03458, 2022. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 352, + 545, + 391 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 352, + 545, + 391 + ], + "spans": [ + { + "bbox": [ + 308, + 352, + 545, + 391 + ], + "type": "text", + "content": "[26] Wenyi Hong, Ming Ding, Wendi Zheng, Xinghan Liu, and Jie Tang. Cogvideo: Large-scale pretraining for text-to-video generation via transformers. arXiv preprint arXiv:2205.15868, 2022. 2, 6, 7" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 308, + 392, + 545, + 422 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 392, + 545, + 422 + ], + "spans": [ + { + "bbox": [ + 308, + 392, + 545, + 422 + ], + "type": "text", + "content": "[27] Yaosi Hu, Chong Luo, and Zhenzhong Chen. Make it move: controllable image-to-video generation with text descriptions. In CVPR, 2022. 2, 4, 6" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 308, + 424, + 545, + 473 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 424, + 545, + 473 + ], + "spans": [ + { + "bbox": [ + 308, + 424, + 545, + 473 + ], + "type": "text", + "content": "[28] Justin Johnson, Bharath Hariharan, Laurens Van Der Maaten, Li Fei-Fei, C Lawrence Zitnick, and Ross Girshick. Clevr: A diagnostic dataset for compositional language and elementary visual reasoning. In CVPR, 2017. 2" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 308, + 475, + 545, + 505 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 475, + 545, + 505 + ], + "spans": [ + { + "bbox": [ + 308, + 475, + 545, + 505 + ], + "type": "text", + "content": "[29] Peter Kán and Hannes Kafumann. Deep light: light source estimation for augmented reality using deep learning. The Visual Computer, 2019. 4" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 308, + 506, + 545, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 506, + 545, + 536 + ], + "spans": [ + { + "bbox": [ + 308, + 506, + 545, + 536 + ], + "type": "text", + "content": "[30] Doyeon Kim, Donggyu Joo, and Junmo Kim. Tivgan: Text to image to video generation with step-by-step evolutionary generator. In IEEE Access, 2020. 3" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 308, + 537, + 545, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 537, + 545, + 557 + ], + "spans": [ + { + "bbox": [ + 308, + 537, + 545, + 557 + ], + "type": "text", + "content": "[31] Dan Klein and Christopher D Manning. Accurate unlexicalized parsing. In ACL, 2003. 4" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 308, + 558, + 545, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 558, + 545, + 588 + ], + "spans": [ + { + "bbox": [ + 308, + 558, + 545, + 588 + ], + "type": "text", + "content": "[32] Ranjay Krishna, Kenji Hata, Frederic Ren, Li Fei-Fei, and Juan Carlos Niebles. Dense-captioning events in videos. In ICCV, 2017. 2, 4" + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 308, + 590, + 545, + 619 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 590, + 545, + 619 + ], + "spans": [ + { + "bbox": [ + 308, + 590, + 545, + 619 + ], + "type": "text", + "content": "[33] Dingquan Li, Tingting Jiang, and Ming Jiang. Quality assessment of in-the-wild videos. In ACM MM, pages 2351-2359, 2019. 5" + } + ] + } + ], + "index": 35 + }, + { + "bbox": [ + 308, + 620, + 545, + 651 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 620, + 545, + 651 + ], + "spans": [ + { + "bbox": [ + 308, + 620, + 545, + 651 + ], + "type": "text", + "content": "[34] Yitong Li, Martin Min, Dinghan Shen, David Carlson, and Lawrence Carin. Video generation from text. In Proceedings of the AAAI conference on artificial intelligence, 2018. 1" + } + ] + } + ], + "index": 36 + }, + { + "bbox": [ + 308, + 651, + 545, + 691 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 651, + 545, + 691 + ], + "spans": [ + { + "bbox": [ + 308, + 651, + 545, + 691 + ], + "type": "text", + "content": "[35] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dálár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In ECCV, 2014. 2" + } + ] + } + ], + "index": 37 + }, + { + "bbox": [ + 308, + 692, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 692, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 308, + 692, + 545, + 712 + ], + "type": "text", + "content": "[36] Wen Liu, Zhixin Piao, Jie Min, Wenhan Luo, Lin Ma, and Shenghua Gao. Liquid warping gan: A unified framework" + } + ] + } + ], + "index": 38 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14813" + } + ] + } + ], + "index": 40 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 73, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 66, + 73, + 287, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 73, + 287, + 94 + ], + "spans": [ + { + "bbox": [ + 66, + 73, + 287, + 94 + ], + "type": "text", + "content": "for human motion imitation, appearance transfer and novel view synthesis. In ICCV, 2019. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 95, + 287, + 125 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 95, + 287, + 125 + ], + "spans": [ + { + "bbox": [ + 48, + 95, + 287, + 125 + ], + "type": "text", + "content": "[37] Yue Liu, Xin Wang, Yitian Yuan, and Wenwu Zhu. Cross-modal dual learning for sentence-to-video generation. In ACM MM, 2019. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 127, + 286, + 156 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 127, + 286, + 156 + ], + "spans": [ + { + "bbox": [ + 48, + 127, + 286, + 156 + ], + "type": "text", + "content": "[38] Ziwei Liu, Ping Luo, Xiaogang Wang, and Xiaou Tang. Deep learning face attributes in the wild. In ICCV, 2015. 3, 4" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 159, + 287, + 198 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 159, + 287, + 198 + ], + "spans": [ + { + "bbox": [ + 48, + 159, + 287, + 198 + ], + "type": "text", + "content": "[39] Jonathan Malmaud, Jonathan Huang, Vivek Rathod, Nick Johnston, Andrew Rabinovich, and Kevin Murphy. What's cookin'? interpreting cooking videos using text, speech and vision. In ACL, 2015. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 201, + 287, + 241 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 201, + 287, + 241 + ], + "spans": [ + { + "bbox": [ + 48, + 201, + 287, + 241 + ], + "type": "text", + "content": "[40] Antoine Miech, Dimitri Zhukov, Jean-Baptiste Alayrac, Makarand Tapaswi, Ivan Laptev, and Josef Sivic. Howto100m: Learning a text-video embedding by watching hundred million narrated video clips. In CVPR, 2019. 2, 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 243, + 287, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 243, + 287, + 282 + ], + "spans": [ + { + "bbox": [ + 48, + 243, + 287, + 282 + ], + "type": "text", + "content": "[41] Niluthpol Chowdhury Mithun, Juncheng Li, Florian Metze, and Amit K Roy-Chowdhury. Learning joint embedding with multimodal cues for cross-modal video-text retrieval. In ICMR, 2018. 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 285, + 287, + 314 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 285, + 287, + 314 + ], + "spans": [ + { + "bbox": [ + 48, + 285, + 287, + 314 + ], + "type": "text", + "content": "[42] Anish Mittal, Anush Krishna Moorthy, and Alan Conrad Bovik. No-reference image quality assessment in the spatial domain. In TIP, 2012. 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 316, + 287, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 316, + 287, + 346 + ], + "spans": [ + { + "bbox": [ + 48, + 316, + 287, + 346 + ], + "type": "text", + "content": "[43] Gaurav Mittal, Tanya Marwah, and Vineeth N Balasubramanian. Sync画: Automatic video generation using deep recurrent attentive architectures. In ACM MM, 2017. 1, 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 348, + 287, + 378 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 348, + 287, + 378 + ], + "spans": [ + { + "bbox": [ + 48, + 348, + 287, + 378 + ], + "type": "text", + "content": "[44] Ali Mollahosseini, Behzad Hasani, and Mohammad H Ma-hoor. Affectnet: A database for facial expression, valence, and arousal computing in the wild. In TAC, 2017. 4" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 380, + 287, + 410 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 380, + 287, + 410 + ], + "spans": [ + { + "bbox": [ + 48, + 380, + 287, + 410 + ], + "type": "text", + "content": "[45] A. Nagrani, J. S. Chung, and A. Zisserman. Voxceleb: a large-scale speaker identification dataset. In INTER-SPEECH, 2017. 1, 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 412, + 287, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 412, + 287, + 441 + ], + "spans": [ + { + "bbox": [ + 48, + 412, + 287, + 441 + ], + "type": "text", + "content": "[46] Yingwei Pan, Zhaofan Qiu, Ting Yao, Houqiang Li, and Tao Mei. To create what you tell: Generating videos from captions. In ACM MM, 2017. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 444, + 287, + 483 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 444, + 287, + 483 + ], + "spans": [ + { + "bbox": [ + 48, + 444, + 287, + 483 + ], + "type": "text", + "content": "[47] MarcAurelio Ranzato, Arthur Szlam, Joan Bruna, Michael Mathieu, Ronan Collobert, and Sumit Chopra. Video (language) modeling: a baseline for generative models of natural videos. arXiv preprint arXiv:1412.6604, 2014. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 485, + 287, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 485, + 287, + 525 + ], + "spans": [ + { + "bbox": [ + 48, + 485, + 287, + 525 + ], + "type": "text", + "content": "[48] Andreas Rössler, Davide Cozzolino, Luisa Verdoliva, Christian Riess, Justus Thies, and Matthias Nießner. Faceforensics: A large-scale video dataset for forgery detection in human faces. arXiv preprint arXiv:1803.09179, 2018. 1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 527, + 287, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 527, + 287, + 577 + ], + "spans": [ + { + "bbox": [ + 48, + 527, + 287, + 577 + ], + "type": "text", + "content": "[49] Christoph Schuhmann, Richard Vencu, Romain Beaumont, Robert Kaczmarczyk, Clayton Mullis, Aarush Katta, Theo Coombes, Jenia Jitsev, and Aran Komatsuzaki. Laion-400m: Open dataset of clip-filtered 400 million image-text pairs. arXiv preprint arXiv:2111.02114, 2021. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 579, + 287, + 618 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 579, + 287, + 618 + ], + "spans": [ + { + "bbox": [ + 48, + 579, + 287, + 618 + ], + "type": "text", + "content": "[50] Piyush Sharma, Nan Ding, Sebastian Goodman, and Radu Soricut. Conceptual captions: A cleaned, hypernymed, image alt-text dataset for automatic image captioning. In ACL, 2018. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 620, + 287, + 670 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 620, + 287, + 670 + ], + "spans": [ + { + "bbox": [ + 48, + 620, + 287, + 670 + ], + "type": "text", + "content": "[51] Uriel Singer, Adam Polyak, Thomas Hayes, Xi Yin, Jie An, Songyang Zhang, Qiyuan Hu, Harry Yang, Oron Ashual, Oran Gafni, et al. Make-a-video: Text-to-video generation without text-video data. arXiv preprint arXiv:2209.14792, 2022. 2, 6, 7" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 48, + 673, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 673, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 673, + 287, + 712 + ], + "type": "text", + "content": "[52] David Stap, Maurits Bleeker, Sarah Ibrahimi, and Maartje ter Hoeve. Conditional image generation and manipulation for user-specified content. arXiv preprint arXiv:2005.04909, 2020. 2, 4" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 622 + ], + "type": "list", + "angle": 0, + "index": 33, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 114 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 114 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 114 + ], + "type": "text", + "content": "[53] Thomas Unterthiner, Sjoerd van Steenkiste, Karol Kurach, Raphael Marinier, Marcin Michalski, and Sylvain Gelly. Towards accurate generative models of video: A new metric & challenges. arXiv preprint arXiv:1812.01717, 2018. 8" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 115, + 545, + 135 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 115, + 545, + 135 + ], + "spans": [ + { + "bbox": [ + 307, + 115, + 545, + 135 + ], + "type": "text", + "content": "[54] Aaron Van Den Oord, Oriol Vinyals, et al. Neural discrete representation learning. NeurIPS, 2017. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 137, + 545, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 137, + 545, + 196 + ], + "spans": [ + { + "bbox": [ + 308, + 137, + 545, + 196 + ], + "type": "text", + "content": "[55] Ruben Villegas, Mohammad Babaeizadeh, Pieter-Jan Kindermans, Hernan Moraldo, Han Zhang, Mohammad Taghi Saffar, Santiago Castro, Julius Kunze, and Dumitru Erhan. Phenaki: Variable length video generation from open domain textual description. arXiv preprint arXiv:2210.02399, 2022. 2, 6, 7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 198, + 545, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 198, + 545, + 237 + ], + "spans": [ + { + "bbox": [ + 308, + 198, + 545, + 237 + ], + "type": "text", + "content": "[56] Xin Wang, Jiawei Wu, Junkun Chen, Lei Li, Yuan-Fang Wang, and William Yang Wang. Vatex: A large-scale, high-quality multilingual dataset for video-and-language research. In ICCV, 2019. 4, 5, 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 239, + 545, + 279 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 239, + 545, + 279 + ], + "spans": [ + { + "bbox": [ + 308, + 239, + 545, + 279 + ], + "type": "text", + "content": "[57] Chenfei Wu, Lun Huang, Qianxi Zhang, Binyang Li, Lei Ji, Fan Yang, Guillermo Sapiro, and Nan Duan. Godiva: Generating open-domain videos from natural descriptions. arXiv preprint arXiv:2104.14806, 2021. 2, 8" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 281, + 545, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 281, + 545, + 330 + ], + "spans": [ + { + "bbox": [ + 308, + 281, + 545, + 330 + ], + "type": "text", + "content": "[58] Chenfei Wu, Jian Liang, Xiaowei Hu, Zhe Gan, Jianfeng Wang, Lijuan Wang, Zicheng Liu, Yuejian Fang, and Nan Duan. Nuwa-infinity: Autoregressive over autoregressive generation for infinite visual synthesis. In NeurIPS, 2022. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 332, + 545, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 332, + 545, + 370 + ], + "spans": [ + { + "bbox": [ + 308, + 332, + 545, + 370 + ], + "type": "text", + "content": "[59] Chenfei Wu, Jian Liang, Lei Ji, Fan Yang, Yuejian Fang, Daxin Jiang, and Nan Duan. N\\`uwa: Visual synthesis pretraining for neural visual world creation. In ECCV, 2022. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 373, + 545, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 373, + 545, + 403 + ], + "spans": [ + { + "bbox": [ + 308, + 373, + 545, + 403 + ], + "type": "text", + "content": "[60] Wayne Wu, Yunxuan Zhang, Cheng Li, Chen Qian, and Chen Change Loy. Reenactgan: Learning to reenact faces via boundary transfer. In ECCV, 2018. 1, 3, 5" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 404, + 545, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 404, + 545, + 435 + ], + "spans": [ + { + "bbox": [ + 308, + 404, + 545, + 435 + ], + "type": "text", + "content": "[61] Weihao Xia, Yujiu Yang, Jing-Hao Xue, and Baoyuan Wu. Tedigan: Text-guided diverse face image generation and manipulation. In CVPR, 2021. 3, 4" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 436, + 545, + 465 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 436, + 545, + 465 + ], + "spans": [ + { + "bbox": [ + 308, + 436, + 545, + 465 + ], + "type": "text", + "content": "[62] Jun Xu, Tao Mei, Ting Yao, and Yong Rui. Msr-vtt: A large video description dataset for bridging video and language. In CVPR, 2016. 2, 4" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 467, + 545, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 467, + 545, + 487 + ], + "spans": [ + { + "bbox": [ + 308, + 467, + 545, + 487 + ], + "type": "text", + "content": "[63] Gilbert Youmans. Measuring lexical style and competence: The type-token vocabulary curve. Style, 1990. 6" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 308, + 488, + 545, + 508 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 488, + 545, + 508 + ], + "spans": [ + { + "bbox": [ + 308, + 488, + 545, + 508 + ], + "type": "text", + "content": "[64] Bowen Zhang, Hexiang Hu, and Fei Sha. Cross-modal and hierarchical modeling of video and text. In ECCV, 2018. 6" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 308, + 509, + 545, + 539 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 509, + 545, + 539 + ], + "spans": [ + { + "bbox": [ + 308, + 509, + 545, + 539 + ], + "type": "text", + "content": "[65] Luowei Zhou, Chenliang Xu, and Jason J Corso. Towards automatic learning of procedures from web instructional videos. In AAAI, 2018. 2" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 308, + 540, + 545, + 581 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 540, + 545, + 581 + ], + "spans": [ + { + "bbox": [ + 308, + 540, + 545, + 581 + ], + "type": "text", + "content": "[66] Hao Zhu, Wayne Wu, Wentao Zhu, Liming Jiang, Siwei Tang, Li Zhang, Ziwei Liu, and Chen Change Loy. CelebV-HQ: A large-scale video facial attributes dataset. In ECCV, 2022. 2, 3, 4, 5, 6, 7" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 308, + 582, + 545, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 582, + 545, + 622 + ], + "spans": [ + { + "bbox": [ + 308, + 582, + 545, + 622 + ], + "type": "text", + "content": "[67] Dimitri Zhukov, Jean-Baptiste Alayrac, Ramadan Gokberk Cinbis, David Fouhey, Ivan Laptev, and Josef Sivic. Cross-task weakly supervised learning from instructional videos. In CVPR, 2019. 1, 4" + } + ] + } + ], + "index": 32 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "text", + "content": "14814" + } + ] + } + ], + "index": 34 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Center Focusing Network for Real-Time LiDAR Panoptic Segmentation/e749a0dd-41b1-4b7d-9325-ea72d56e45cb_content_list.json b/2023/Center Focusing Network for Real-Time LiDAR Panoptic Segmentation/e749a0dd-41b1-4b7d-9325-ea72d56e45cb_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..511dc90512aebe7413a170712794ff0385db615d --- /dev/null +++ b/2023/Center Focusing Network for Real-Time LiDAR Panoptic Segmentation/e749a0dd-41b1-4b7d-9325-ea72d56e45cb_content_list.json @@ -0,0 +1,1640 @@ +[ + { + "type": "text", + "text": "Center Focusing Network for Real-Time LiDAR Panoptic Segmentation", + "text_level": 1, + "bbox": [ + 122, + 130, + 846, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Xiaoyan Li $^{1,2}$ Gang Zhang $^{3}$ Boyue Wang $^{1,2}$ Yongli Hu $^{1,2}$ Baocai Yin $^{1,2}$", + "bbox": [ + 169, + 178, + 792, + 198 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1Beijing Municipal Key Lab of Multimedia and Intelligent Software Technology", + "bbox": [ + 168, + 198, + 800, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ Beijing Institute of Artificial Intelligence, Faculty of Information Technology, Beijing University of Technology, Beijing 100124, China", + "bbox": [ + 174, + 215, + 795, + 251 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3Mogo Auto Intelligence and Telemetics Information Technology Co. Ltd.", + "bbox": [ + 189, + 251, + 777, + 268 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{xiaoyan.li, wby, huyongli, ybc}@bjut.edu.cn zhanggang11021136@gmail.com", + "bbox": [ + 169, + 271, + 795, + 286 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 321, + 313, + 338 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "LiDAR panoptic segmentation facilitates an autonomous vehicle to comprehensively understand the surrounding objects and scenes and is required to run in real time. The recent proposal-free methods accelerate the algorithm, but their effectiveness and efficiency are still limited owing to the difficulty of modeling non-existent instance centers and the costly center-based clustering modules. To achieve accurate and real-time LiDAR panoptic segmentation, a novel center focusing network (CFNet) is introduced. Specifically, the center focusing feature encoding (CFFE) is proposed to explicitly understand the relationships between the original LiDAR points and virtual instance centers by shifting the LiDAR points and filling in the center points. Moreover, to leverage the redundantly detected centers, a fast center dedduplication module (CDM) is proposed to select only one center for each instance. Experiments on the SemanticKITTI and nuScenes panoptic segmentation benchmarks demonstrate that our CFNet outperforms all existing methods by a large margin and is 1.6 times faster than the most efficient method.", + "bbox": [ + 75, + 354, + 473, + 657 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 691, + 209, + 708 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Panoptic segmentation [18] combines both semantic segmentation and instance segmentation in a single framework. It predicts semantic labels for the uncountable stuff classes (e.g. road, sidewalk), while it simultaneously provides semantic labels and instance IDs for the countable things classes (e.g. car, pedestrian). The LiDAR panoptic segmentation is one of the bases for the safety of autonomous driving, which employs the point clouds collected by the Light Detection and Ranging (LiDAR) sensors to effectively depict the surroundings. Existing LiDAR panoptic segmentation methods first conduct semantic segmentation, and then achieve instance segmentation for the things categories in", + "bbox": [ + 75, + 719, + 470, + 901 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/bdf504294a4b4d31dbec88f296864e5e0cb8826f00a6b19e1001db7767b8cd69.jpg", + "image_caption": [ + "Figure 1. PQ vs. runtime on the SemanticKITTI test set. Runtime measurements are taken on a single NVIDIA RTX 3090 GPU. The panoptic quality (PQ) is introduced in section 4.1." + ], + "image_footnote": [], + "bbox": [ + 509, + 319, + 882, + 547 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "two ways, the proposal-based and proposal-free methods.", + "bbox": [ + 498, + 627, + 879, + 642 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The proposal-based methods [17, 31, 37] adopt a two-stage process similar to the well-known Mask R-CNN [14] in the image domain. It first generates object proposals for the things points by using 3D detection networks [19, 30] and then refines the instance segmentation results within each proposal. As shown in Fig. 1, these methods are usually complicated and hardly achieve real-time processing, owing to their sequential multi-stage pipelines.", + "bbox": [ + 496, + 643, + 890, + 763 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The proposal-free frameworks [13, 15, 21, 22, 29, 35, 39] are more compact. To associate the things points with instance IDs, these methods usually leverage the instance centers. Specifically, they regress the offsets from the points to their corresponding centers, and then adopt the class-agnostic center-based clustering modules [13, 15, 29] or the bird's-eye view (BEV) center heatmap [22, 35, 39]. However, two problems exist in these methods. First, for center feature extracting and center modeling, the non-existent in", + "bbox": [ + 496, + 763, + 890, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 44 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "13425", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/739d52840857770b48bf2f272685550328dd7f45992da9ede8d2413a38ad8818.jpg", + "image_caption": [ + "(a) CFNet without CFFE" + ], + "image_footnote": [], + "bbox": [ + 89, + 88, + 263, + 198 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/155e0ba77e4ce0beccb1409edc9db8247cd9a825ad52c636deb8ccd9172fd29d.jpg", + "image_caption": [ + "(b) CFNet with CFFE", + "Figure 2. Instance segmentation of a car. Without our CFFE, the car is split into parts (a), while the CFFE significantly alleviates this problem (b). Different colors represent different instances." + ], + "image_footnote": [], + "bbox": [ + 282, + 88, + 455, + 198 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "stance centers increase the difficulty, considering that the LiDAR points are usually surface-aggregated [35] and an instance center is imaginary in most cases. As shown in Fig. 2(a), the difficulty often results in the fault that one instance is incorrectly split into several parts. Second, for exploiting the redundantly detected centers, the clustering modules (e.g. MeanShift, DBSCAN) are too time-consuming to support the real-time autonomous driving perception systems, while the BEV center heatmap cannot distinguish objects with different altitudes in the same BEV grid.", + "bbox": [ + 75, + 292, + 467, + 459 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "For accurate and fast LiDAR panoptic segmentation, a proposal-free center focusing network (CFNet) is proposed. For better encoding center features, a novel center focusing feature encoding (CFFE) is proposed to generate center-focusing feature maps by shifting the things points to fill in the non-existent instance centers for more accurate predictions (as shown in Fig. 2(b)). For center modeling, the CFNet not only decomposes the panoptic segmentation task into the widely-used semantic segmentation and center offset regression, but also proposes a new confidence score prediction for indicating the accuracy of the center offset regression. Subsequently, for the detected centers exploiting, a novel center dedduplication module (CDM) is designed to select one center for a single instance. The CDM keeps the predicted centers with higher confidence scores, while suppressing the ones with lower confidence. Finally, instance segmentation is achieved by assigning the shifted things points to the closest center. For efficiency, the proposed CFNet is built on the 2D projection-based segmentation paradigm. Our contributions are as follows:", + "bbox": [ + 75, + 460, + 467, + 761 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- A proposal-free CFNet is proposed to achieve accurate and fast LiDAR panoptic segmentation by solving the bottleneck problems of center modeling and center-based clustering in previous methods.", + "- The CFFE is proposed to alleviate the difficulty of modeling the non-existent instance centers and the CDM is designed to efficiently keep one center for each instance." + ], + "bbox": [ + 94, + 770, + 467, + 898 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "- The proposed CFNet is evaluated on the nuScenes and SemanticKITTI LiDAR panoptic segmentation benchmarks. Our CFNet achieves the state-of-the-art performance with a real-time inference speed.", + "bbox": [ + 519, + 90, + 890, + 151 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 165, + 640, + 180 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The LiDAR panoptic segmentation methods usually adopt the LiDAR semantic segmentation networks as the backbone and jointly optimize the semantic segmentation and instance segmentation tasks. The results of the two tasks are fused to generate the final panoptic segmentation results. The LiDAR semantic segmentation backbone is first introduced. Then, the panoptic segmentation methods, achieving instance segmentation upon the semantic segmentation backbone, are grouped into the proposal-based and proposal-free methods, and are further discussed.", + "bbox": [ + 496, + 190, + 890, + 340 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "LiDAR semantic segmentation backbone. The backbone network for LiDAR semantic segmentation can be divided into three groups: point-based, voxel-based, and 2D projection-based methods. The point-based methods [16, 27, 28, 33] directly operate on the raw point clouds but are extremely time-consuming due to the expensive local neighbor searching. The voxel-based methods [7, 8, 32, 40] discretize the point clouds into structured voxels, where sparse 3D convolutions are applied. These methods are still difficult to meet the real-time applications, although they achieve the highest accuracy. The 2D projection-based methods are more efficient since most of their computation is done in the 2D space, such as range view (RV) [10,25,34], bird's-eye view (BEV) [19], polar view [38], and multiview fusion [1, 23]. For fast LiDAR panoptic segmentation, the Panoptic-PolarNet [39], SMAC-Seg [20], and LPSAD [24] all adopt the 2D projection-based backbone. A recent real-time semantic segmentation method, namely CPGNet [23], is a 2D projection-based one that explores an end-to-end multi-view fusion framework by fusing the features from the point, BEV, and RV.", + "bbox": [ + 496, + 340, + 892, + 657 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "**Proposal-based methods.** The proposal-based methods conduct instance segmentation through a two-stage complicated process. They first detect the foreground instances and then refine the instance segmentation results independently within each detected bounding box. Based on the Mask R-CNN [14] for instance segmentation, the MOPT [17] and EfficientLPS [31] insert a semantic branch to achieve panoptic segmentation on the range view (RV). Recently, LidarMultiNet [37] unifies LiDAR-based 3D object detection, semantic segmentation, and panoptic segmentation in a single framework to reduce the computation cost by sharing a strong voxel-based backbone.", + "bbox": [ + 496, + 659, + 890, + 839 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Proposal-free methods. The proposal-free methods usually apply the class-agnostic clustering to the things points to conduct instance segmentation. The LPSAD [24] clusters points into instances by regressing the center off", + "bbox": [ + 496, + 839, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "13426", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/6b477c5cd5f7519ef9d8618ecd027fb88a9c3a0aa05dd6413acf48af6c7613cb.jpg", + "image_caption": [ + "Figure 3. The overview of our CFNet. It consists of four steps: 1) the 2D projection-based backbone extracts features on the 2D space; 2) the proposed center focusing feature encoding (CFFE) mimics and enhances the non-existent instance center features; 3) the panoptic segmentation head predicts the output results; 4) the proposed center dedduplication module (CDM) achieves instance segmentation that is fused to generate the final panoptic segmentation results. The dashed lines indicate that the operations are only used during inference." + ], + "image_footnote": [], + "bbox": [ + 76, + 88, + 890, + 330 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/8493b7dd4cb04a7ae0b68c07950105b284750e2ac83566bf88e934fa0a7a78a5.jpg", + "image_caption": [], + "image_footnote": [ + ": Multi-Layer Perception", + "P2G: Point to Grid", + "FCN: Fully Convolutional Network", + ": Fully Connected Layer", + "G2P: Grid to Point", + "PF: Point Fusion" + ], + "bbox": [ + 80, + 333, + 447, + 415 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/c29088879b4f7a3e7e87ca8fb061349c93ad6c9c5e8abf4bd960e08545d9314d.jpg", + "image_caption": [], + "image_footnote": [ + "CFFE: Center Focusing Feature Encoding", + "CDM: Center Dedduplication Module" + ], + "bbox": [ + 478, + 333, + 888, + 415 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "sets. The Panoster [13] directly predicts the instance IDs from a learning-based clustering module, where the time-consuming DBSCAN [11] is used to refine the instance segmentation results. The DS-Net [15] proposes a learnable dynamic shifting module that adjusts the kernel functions of the MeanShift [9] to handle instances with various sizes. The GP-S3Net [29] constructs a graph convolutional network (GCN) on the over-segmentation clusters to identify instances. The SMAC-Seg [20] proposes a sparse multidirectional attention clustering and center-aware repel loss for instance segmentation. These methods adopt the time-consuming clustering methods. To further accelerate the algorithm, recently, Panoptic-PolarNet [39], SCAN [35], and Panoptic-PHNet [22] adopt the BEV center heatmap and center offsets for instance segmentation. However, the BEV center heatmap is also costly and confuses on $z$ -axis.", + "bbox": [ + 75, + 561, + 472, + 806 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Approach", + "text_level": 1, + "bbox": [ + 76, + 825, + 184, + 843 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The input of the LiDAR panoptic segmentation task is the LiDAR point clouds, given by a set of unordered points $\\pmb{P} = \\{(p_i,f_i)\\}_{i = 0}^{N - 1}$ (where $\\pmb {p}_i = (x,y,z)$ is the 3D coor", + "bbox": [ + 76, + 854, + 468, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "dinates in the cartesian space and $\\pmb{f}_i$ denotes additional LiDAR point features, e.g. intensity). The task aims to assign a set of labels $\\pmb{L} = \\{(s_i, inst_i)\\}_{i=0}^{N-1}$ to these points, where $s_i$ is the semantic label (e.g. road, building, car, pedestrian), and $inst_i$ is the instance ID of the $i^{th}$ point. Moreover, $s_i$ can be divided into the uncountable stuff classes (e.g. road, building) and countable things classes (e.g. car, pedestrian). The instance IDs of stuff points are set as 0.", + "bbox": [ + 496, + 561, + 890, + 684 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To predict the labels $L$ of the input LiDAR point clouds $P$ , our CFNet decomposes this process into four steps, as shown in Fig. 3: 1) an off-the-shelf 2D projection-based backbone is applied to efficiently extract features on the 2D spaces (e.g. BEV, RV); 2) a novel center focusing feature encoding (CFFE) is used to generate center-focusing feature maps for more accurate predictions of the instance centers; 3) the panoptic segmentation head fuses the features from the 3D points and 2D spaces to predict the semantic segmentation results, the center offsets, and the confidence scores of the center offsets, respectively; 4) during inference, a post-process is conducted to produce the panoptic segmentation results, where a novel center dedduplication module (CDM) operates on the shifted things points to se", + "bbox": [ + 496, + 688, + 892, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "13427", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/8bdd3a6609487f1e67f8756fb55ea40ea595a770e1b9ae4f7b544fb54786ca76.jpg", + "image_caption": [ + "Figure 4. The proposed center focusing feature encoding (CFFE). The \"Conv\" represents a 2D convolution with $3 \\times 3$ kernels, a batch normalization, and a ReLU layer. The details of the semantic branch and instance branch are shown in Fig. 3. The blue arrows are coordinate-related operations." + ], + "image_footnote": [], + "bbox": [ + 81, + 90, + 890, + 281 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "lect only one center for a single instance.", + "bbox": [ + 76, + 361, + 349, + 376 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The design of the backbone in the first step is beyond the scope of the paper, and the last three mentioned steps are illustrated in detail in the following subsections.", + "bbox": [ + 75, + 376, + 468, + 422 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1. Center Focusing Feature Encoding", + "text_level": 1, + "bbox": [ + 76, + 433, + 380, + 450 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As mentioned above, the LiDAR points of an object are usually surface-aggregated, especially for car and truck categories, resulting in the fact that the center of an object is imaginary and does not exist in the LiDAR point clouds. To encode the features of the non-existent centers, a novel center focusing feature encoding (CFFE) is proposed, which takes the 2D features from the backbone and the 3D point coordinates as inputs and generates the enhanced center-focusing feature maps as shown in Fig. 3.", + "bbox": [ + 75, + 457, + 468, + 593 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The CFFE module consists of three steps, including intermediate result prediction, center feature generation, and feature enhancement module, as shown in Fig. 4.", + "bbox": [ + 75, + 594, + 468, + 638 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Intermediate result prediction. In this step, the CFFE predicts intermediate results (including the semantic segmentation, center offset, and its confidence scores) according to the 2D features $\\pmb{F}_{m}$ and 3D point features $\\pmb{F}_{p}$ for subsequent center feature simulation. Specifically, two convolution layers are applied on the 2D features $\\pmb{F}_{m}$ independently, to generate semantic features $\\pmb{F}_{m}^{sem}$ and instance features $\\pmb{F}_{m}^{ins}$ ( $m$ is the specific 2D view, such as RV [10, 25, 34], BEV [19], and polar view [38].)", + "bbox": [ + 75, + 640, + 470, + 776 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {F} _ {m} ^ {\\text {s e m}} = \\operatorname {C o n v} \\left(\\boldsymbol {F} _ {m}; \\boldsymbol {\\theta} _ {1}\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 186, + 789, + 468, + 806 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {F} _ {m} ^ {\\text {i n s}} = \\operatorname {C o n v} \\left(\\boldsymbol {F} _ {m}; \\boldsymbol {\\theta} _ {2}\\right), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 192, + 809, + 468, + 827 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $Conv$ denotes sequential 2D convolution, batch normalization, and ReLU operations, and $\\theta_{1}$ and $\\theta_{2}$ are their learnable parameters. Then, the semantic branch generates per-point 3D semantic features $F_{p}^{sem}$ by fusing the point", + "bbox": [ + 75, + 839, + 470, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "features $F_{p}$ and 2D semantic features $\\pmb{F}_{m}^{sem}$", + "bbox": [ + 498, + 359, + 799, + 378 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {F} _ {p} ^ {\\text {s e m}} = \\operatorname {S e g} \\left(\\boldsymbol {F} _ {p}, \\boldsymbol {F} _ {m} ^ {\\text {s e m}}; \\boldsymbol {\\theta} _ {3}\\right), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 594, + 387, + 890, + 406 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $Seg$ is the semantic branch, and $\\theta_{3}$ is the parameters.", + "bbox": [ + 498, + 414, + 890, + 430 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Finally, the intermediate semantic results $\\hat{S}^{\\prime} = \\{s_{i}^{\\prime}\\}_{i=0}^{N-1}$ is produced based on $F_{p}^{sem}$ . The intermediate results of center offsets $\\hat{O}^{\\prime} = \\{o_{i}^{\\prime}\\}_{i=0}^{N-1}$ and confidence scores $\\hat{C}^{\\prime} = \\{c_{i}^{\\prime}\\}_{i=0}^{N-1}$ are predicted by the instance branch with the point features $F_{p}$ and 2D instance features $F_{m}^{ins}$ ,", + "bbox": [ + 498, + 431, + 892, + 513 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\boldsymbol {S}} ^ {\\prime} = F C \\left(\\boldsymbol {F} _ {p} ^ {s e m}; \\boldsymbol {\\theta} _ {4}\\right), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 620, + 535, + 890, + 556 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\boldsymbol {O}} ^ {\\prime}, \\hat {\\boldsymbol {C}} ^ {\\prime} = I n s \\left(\\boldsymbol {F} _ {p}, \\boldsymbol {F} _ {m} ^ {i n s}; \\boldsymbol {\\theta} _ {5}\\right), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 594, + 559, + 890, + 578 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $Ins$ is the instance branch, and $FC$ represents the fully-connected layer. The structure and the training objective of semantic and instance branches are the same as those in the panoptic segmentation head and are elucidated in Fig. 3 and section 3.2.", + "bbox": [ + 496, + 588, + 890, + 662 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Center feature generation (CFG). In this step, the CFFE generates the shifted center features $\\pmb{F}_{p\\rightarrow m}^{sem}$ by shifting the 3D semantic point features $\\pmb{F}_{p}^{sem}$ to the predicted centers according to the mentioned intermediate results. First, the coordinates of a predicted center $\\pmb{p}_i^{shift}$ are computed by,", + "bbox": [ + 496, + 664, + 890, + 755 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {p} _ {i} ^ {\\text {s h i f t}} = \\boldsymbol {p} _ {i} + \\hat {\\boldsymbol {o}} _ {i} ^ {\\prime} * \\mathbb {1} \\left[ \\hat {c} _ {i} ^ {\\prime} > \\delta \\right], \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 589, + 765, + 890, + 785 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\pmb{p}_i$ is the original 3D LiDAR point coordinates, and $\\mathbb{1}[\\hat{c}_i^{\\prime} > \\delta ]$ is a binary indicator of whether the confidence $\\hat{c}_i^\\prime$ is greater than $\\delta = 0.2$ . In other words, it does not shift the stuff points or the things points with low confidence scores.", + "bbox": [ + 498, + 794, + 890, + 854 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Then, the shifted 3D points $\\{\\pmb{p}_i^{shift}\\}_{i=0}^{N-1}$ are served as the new coordinates of the feature points, and the Point to Grid (P2G) operation in [23] re-projects the 3D semantic", + "bbox": [ + 496, + 854, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "13428", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "features $\\pmb{F}_{p}^{sem}$ onto the 2D projected feature maps $\\pmb{F}_{p\\rightarrow m}^{sem}$ with this new coordinates,", + "bbox": [ + 76, + 90, + 467, + 121 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {F} _ {p \\rightarrow m} ^ {s e m} = P 2 G (\\boldsymbol {F} _ {p} ^ {s e m}; \\{\\boldsymbol {p} _ {i} ^ {\\text {s h i f t}} \\} _ {i = 0} ^ {N - 1}). \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 143, + 128, + 467, + 150 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Compared with $F_{p}^{sem}$ , the feature maps $F_{p\\rightarrow m}^{sem}$ pay more attention to the imaginary centers since most of the things points have been shifted to their predicted centers.", + "bbox": [ + 75, + 156, + 467, + 200 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Feature enhancement module (FEM). The CFFE finally fuses the semantic feature maps $\\pmb{F}_{m}^{sem}$ and the reprojected shifted center feature maps $\\pmb{F}_{p\\rightarrow m}^{sem}$ to generate center-focusing semantic feature maps $\\pmb{F}_{m}^{CFsem}$ and instance feature maps $\\pmb{F}_{m}^{CFS}$ , which are used by the subsequent semantic and instance branches for more accurate predictions. The enhancement module is composed of a simple concatenation operation and several convolution layers, whose details are in the supplementary material.", + "bbox": [ + 75, + 202, + 467, + 340 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In addition, for application of the multi-view fusion backbone (e.g. [23] shown in the supplementary material), the feature maps $\\mathbf{F}_m$ , $\\mathbf{F}_m^{sem}$ , $\\mathbf{F}_m^{ins}$ , $\\mathbf{F}_{p\\rightarrow m}^{sem}$ (e.g. $m \\in \\{\\mathrm{RV},\\mathrm{BEV}\\}$ ) of each view are computed independently according to the above process. Then, they are fused in the point fusion (PF) module [23] to generate integrated 3D point features and per-point prediction.", + "bbox": [ + 76, + 340, + 467, + 446 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2. Panoptic Segmentation Head", + "text_level": 1, + "bbox": [ + 76, + 454, + 334, + 470 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For better modeling the instance center, the panoptic segmentation head uses a semantic branch to predict the semantic segmentation, and an instance branch to estimate both the center offsets and the newly introduced confidence scores, given the center-focusing semantic and instance feature maps. The architectures of both branches follow the segmentation head in [23] and are briefly introduced, while their training objectives are elaborated on.", + "bbox": [ + 75, + 478, + 467, + 598 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Semantic branch. For a per-point prediction, the semantic branch first applies the Grid to Point (G2P) operations to acquire the 3D point-wise features from the 2D semantic feature maps. Then, a PF module fuses the pointwise features from the G2P operation and the original 3D points to generate point-wise semantic features. The details of the G2P and PF can be found in [23]. After obtaining the point-wise semantic features, a fully connected (FC) layer is used to predict the final per-point semantic result $\\hat{\\psi}_i^j$ ( $\\hat{\\psi}_i^j \\in [0,1]$ ). $\\hat{\\psi}_i^j$ represents the probability of the $i^{th}$ LiDAR point belonging to the $j^{th}$ class. The predicted semantic label $\\hat{s}_i$ is acquired by selecting the most probable class $\\hat{s}_i = \\arg \\max_j \\hat{\\psi}_i^j$ .", + "bbox": [ + 75, + 599, + 467, + 795 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Referring to the CPGNet [23], the same loss functions are adopted, including weighted cross entropy (WCE) loss $\\mathcal{L}_{wce}$ , Lovász-Softmax loss $\\mathcal{L}_{ls}$ [4], and transformation consistency loss $\\mathcal{L}_{tc}$ .", + "bbox": [ + 76, + 796, + 467, + 854 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Instance branch. Similar to the semantic branch, the instance branch also adopts the G2P operation and a PF module to acquire point-wise instance features. An FC layer", + "bbox": [ + 76, + 854, + 467, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "is applied to predict the per-point center offsets $\\hat{o}_i$ . The ground-truth for $\\hat{o}_i$ is the offset vector $e_i - p_i$ from the $i^{th}$ point $p_i$ to its corresponding instance center $e_i$ .", + "bbox": [ + 496, + 90, + 890, + 136 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For center offset regression, the loss function for optimizing $\\hat{o}_i$ only considers the things classes and is formulated as follows,", + "bbox": [ + 496, + 136, + 890, + 180 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\boldsymbol {o} _ {i}} = \\left\\{ \\begin{array}{l l} | | \\hat {\\boldsymbol {o}} _ {i} - (\\boldsymbol {e} _ {i} - \\boldsymbol {p} _ {i}) | | _ {2} & \\text {i f} s _ {i} \\in t h i n g s \\text {c l a s s e s}, \\\\ 0 & \\text {o t h e r w i s e}, \\end{array} \\right. \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 511, + 190, + 890, + 231 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\pmb{e}_i$ is the axis-aligned center of the instance [15]. Then, the losses are summarized as the following,", + "bbox": [ + 496, + 239, + 888, + 270 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {o} = \\frac {1}{N ^ {\\text {t h i n g s}}} \\sum_ {i = 0} ^ {N - 1} \\mathcal {L} _ {\\boldsymbol {o} _ {i}}, \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 612, + 280, + 890, + 321 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $N$ and $N^{things}$ are the numbers of all points and things points, respectively.", + "bbox": [ + 496, + 332, + 888, + 362 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For confidence score regression, another FC layer is used to predict the per-point confidence score $\\hat{c}_i$ to indicate the accuracy degree of $\\hat{o}_i$ . $\\hat{c}_i$ is activated by a sigmoid function to ensure $\\hat{c}_i \\in [0,1]$ . The ground-truth label $c_i$ for supervising $\\hat{c}_i$ is generated by", + "bbox": [ + 496, + 363, + 890, + 438 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nc _ {i} = \\left\\{ \\begin{array}{l l} \\exp \\left(- \\frac {\\mathcal {L} _ {o _ {i}} ^ {2}}{2 \\sigma^ {2}}\\right) & \\text {i f} s _ {i} \\in \\text {t h i n g s c l a s s e s}, \\\\ 0 & \\text {o t h e r w i s e .} \\end{array} \\right. \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 537, + 448, + 890, + 491 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For the things points, the lower $\\mathcal{L}_{o_i}$ is, the higher $c_{i}$ is. It means that the point with a more accurate regression of the center offset has a higher confidence score.", + "bbox": [ + 496, + 500, + 890, + 544 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The weighted binary cross entropy loss $\\mathcal{L}_{wbce}$ is applied,", + "bbox": [ + 517, + 545, + 888, + 561 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {w b c e} = - \\frac {1}{N} \\sum_ {i = 0} ^ {N - 1} 6 c _ {i} \\log (\\hat {c} _ {i}) + (1 - c _ {i}) \\log (1 - \\hat {c} _ {i}), \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 506, + 583, + 890, + 625 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where the things points are manually emphasized since the number of them is much less than that of the stuff points.", + "bbox": [ + 496, + 633, + 888, + 662 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Finally, the loss for each group of results (from the CFNet or CFFE) is defined as", + "bbox": [ + 496, + 664, + 888, + 693 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\mathcal {L} _ {w c e} + 3 \\mathcal {L} _ {l s} + \\mathcal {L} _ {t c} + 2 \\mathcal {L} _ {o} + \\mathcal {L} _ {w b c e}. \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 557, + 705, + 890, + 720 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The total loss is the sum of the two losses sourced from the CFNet and CFFE.", + "bbox": [ + 496, + 732, + 888, + 761 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Center Dedduplication Module", + "text_level": 1, + "bbox": [ + 500, + 771, + 764, + 787 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Given the final predictions of the semantic segmentation $\\hat{S}$ , center offsets $\\hat{O}$ , and confidence scores $\\hat{C}$ , this section introduces how to exploit the detected centers to acquire the panoptic segmentation results during inference, and illustrates the key module, center dedduplication module (CDM).", + "bbox": [ + 496, + 795, + 890, + 869 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Post-process. For the panoptic segmentation results, instance segmentation is first generated and then the final", + "bbox": [ + 496, + 869, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "13429", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "code", + "sub_type": "algorithm", + "code_caption": [ + "Algorithm 1: Center dedduplication module" + ], + "code_body": "Input: Shifted things points $\\{\\tilde{p}_t^{shift}\\in \\mathbb{R}^3\\}_{t = 0}^{M - 1}$ confidence scores $\\{\\tilde{c}_t\\in [0,1]\\}_{t = 0}^{M - 1}$ distance threshold $d\\in \\mathbb{R}$ Output: Instance centers $\\{\\tilde{e}_j\\in \\mathbb{R}^3\\}_{j = 0}^{D - 1}$ \n1 order $=$ sort_index( $\\{\\tilde{c}_t\\}_{t = 0}^{M - 1}$ , decending=true) \n2 initialize a boolean vector $rej\\in \\mathbb{B}^{M}$ with all elements of false and an empty list keep \n3 for $k\\gets 0$ to $M - 1$ do \n4 if not rej[order[k]] then \n5 keep.append(order[k]) \n6 for $h\\gets k + 1$ to $M - 1$ do \n7 dist $= ||\\tilde{p}_{order[k]}^{shift} - \\tilde{p}_{order[h]}^{shift}||_2$ \n8 if dist $< d$ then \n9 | rej[order[h]] $=$ true \n10 end \n11 end \n12 end \n13 end \n14 $\\{\\tilde{e}_j\\} = \\tilde{p}_{keep}^{shift}$", + "bbox": [ + 81, + 112, + 468, + 428 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "panoptic segmentation is acquired by fusing the semantic and instance segmentation labels. There are five steps to achieve the final panoptic segmentation:", + "bbox": [ + 75, + 459, + 468, + 503 + ], + "page_idx": 5 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1) The things points $\\{\\tilde{p}_t\\}_{t=0}^{M-1} = \\{\\pmb{p}_i | \\hat{s}_i \\in$ things classes} are selected according to the predicted semantic labels, with their offsets $\\{\\tilde{o}_t\\}_{t=0}^{M-1}$ and confidence scores $\\{\\tilde{c}_t\\}_{t=0}^{M-1}$ (where $M$ is the number of things points).", + "2) Each shifted things point $\\tilde{\\pmb{p}}_t^{shift}$ is given by $\\tilde{\\pmb{p}}_t^{shift} = \\tilde{\\pmb{p}}_t + \\tilde{\\pmb{o}}_t$ and is severing as an instance center candidate.", + "3) A center $\\tilde{\\pmb{e}}_j$ for each instance $j$ is selected by the CDM based on the coordinates $\\tilde{\\pmb{p}}_t^{shift}$ and its confidence score $\\tilde{c}_t$ , while other candidates are suppressed.", + "4) The instance $\\mathrm{ID}\\widetilde{\\text{in}}\\widetilde{\\text{st}}_t$ is acquired by assigning the shifted things points $\\tilde{\\pmb{p}}_t^{\\text{shift}}$ to the closest center in all centers $\\{\\tilde{\\pmb{e}}_j\\}_{j=0}^{D-1}$ ( $D$ is the number of detected instances).", + "5) The majority voting [15, 39] reassigns the most frequent semantic label to all points of a predicted instance to further guarantee the consistency of semantic labels within a predicted instance." + ], + "bbox": [ + 75, + 503, + 468, + 748 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Center dedduplication module (CDM). The CDM takes the shifted points $\\{\\tilde{p}_t^{shift}\\}_{t=0}^{M-1}$ and confidence scores $\\{\\tilde{c}_t\\}_{t=0}^{M-1}$ as input to get an instance center $\\hat{e}_j$ for each instance $j$ . Inspired by the bounding box NMS, our CDM suppresses the candidate centers with lower scores within a euclidean distance threshold $d$ . The pseudo-code of our CDM is shown in Algorithm 1, where two centers with a distance less than $d$ are considered as the same instance. The process of our CDM is simple and can be easily implemented in CUDA.", + "bbox": [ + 75, + 750, + 470, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 500, + 89, + 633, + 107 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The proposed CFNet is compared with the existing methods on the public SemanticKITTI [2] and nuScenes [5] panoptic segmentation benchmarks.", + "bbox": [ + 496, + 114, + 893, + 160 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Experimental Setup", + "text_level": 1, + "bbox": [ + 498, + 167, + 689, + 184 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Datasets. The SemanticKITTI [2] consists of 22 sequences collected by a Velodyne HDL-64E $360^{\\circ}$ rotating LiDAR with 64 beams vertically. It contains 43,552 LiDAR scans in total and is split into a training set with 19,130 scans from sequences 00 to 10 except 08, a validation set containing sequence 08 with 4,071 scans, and a test set including the rest sequences with 20,351 scans. The test set is only provided with point clouds and used for the online leaderboards. For the panoptic segmentation, it has 19 valid categories, including 11 stuff and 8 things categories.", + "bbox": [ + 496, + 190, + 890, + 340 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The nuScenes [5] is a newly released benchmark with 1,000 scenes collected by a Velodyne HDL-32E $360^{\\circ}$ rotating LiDAR with 32 beams vertically. The dataset is collected in Boston and Singapore. It uses 28,130 scans for training, 6,019 for validation, and 6,008 for testing. The panoptic segmentation benchmark provides 16 categories, including 6 stuff and 10 things categories [12].", + "bbox": [ + 496, + 340, + 890, + 446 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation metric. As the official benchmarks [3, 12] suggest, the panoptic quality (PQ) is adopted to evaluate the performance of LiDAR panoptic segmentation. The PQ [3, 12] measures the overall quality of the panoptic segmentation, and can be decomposed into two terms: 1) the segmentation quality (SQ) is the average IoU of all matched pairs; 2) the recognition quality (RQ) measures the $F_{1}$ score. These three metrics are also reported separately on the stuff and things classes, including $\\mathrm{PQ}^{St}$ , $\\mathrm{SQ}^{St}$ , $\\mathrm{RQ}^{St}$ , and $\\mathrm{PQ}^{Th}$ , $\\mathrm{SQ}^{Th}$ , $\\mathrm{RQ}^{Th}$ , respectively. The $\\mathrm{PQ}^{\\dagger}$ [26] is also reported by replacing the PQ of each stuff class with its IoU. For evaluating the sub-task of semantic segmentation, mean Intersection over Union (mIoU) [2] is reported.", + "bbox": [ + 496, + 448, + 890, + 643 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Training details. In the experiments, two 2D projection-based backbones, namely PolarNet [38] and CPGNet [23], are used in our CFNet. For the backbone of PolarNet [38], the training schedules and data augmentation are the same as those of Panoptic-PolarNet [39] for a fair comparison. Besides, for efficiency, only one stage version of CPGNet [23] is used in our CFNet. It is trained from scratch for 48 epochs with a batch size of 8 and takes around 24 hours on 8 NVIDIA RTX 3090 GPUs. Stochastic gradient descent (SGD) is used as the optimizer, where the initial learning rate, weight decay, and momentum are 0.02, 0.001, and 0.9, respectively. The learning rate is decayed by 0.1 per 10 epochs. In training, the data augmentation is applied as a convention, including random flipping along the $x$ and $y$ axes, random global scale sampled from $[0.95, 1.05]$ , random rotation around the $z$ -axis, and random Gaussian noise $\\mathcal{N}(0, 0.02)$ .", + "bbox": [ + 496, + 643, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "13430", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/e260e27b7bc4b3b641d3fa4692ea841d1a6482f1a7fa2b02d8e1870f30713add.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
BackboneCFFEDeduplicationPQPQ†PQTthPQS\\( ^t \\)mIoURT(ms)
CFGFEM
a b c dPolarNet [38]✓BEV Center Heatmap [39]59.164.165.754.364.565.3 + 20.9
BEV Center Heatmap [39]59.564.366.254.464.669.7 + 20.9
✓✓BEV Center Heatmap [39]60.465.667.255.165.271.9 + 20.9
✓✓CDM [Ours]60.665.767.855.265.471.9 + 3.6
e f g h i jCPGNet [23]✓DBSCAN [11]59.564.163.856.364.631.6+24.9
HDBSCAN [6]58.362.960.956.364.931.6+48.7
MeanShift [9]59.964.564.756.364.931.6+84.6
CDM [Ours]60.565.666.156.566.332.7+2.3
✓CDM [Ours]60.865.966.356.966.537.4+2.3
✓✓CDM [Ours]62.767.570.057.367.441.2+2.3
", + "bbox": [ + 122, + 87, + 848, + 258 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. Ablation Studies", + "text_level": 1, + "bbox": [ + 76, + 310, + 240, + 325 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Ablative analyses are conducted on the SemanticKITTI validation set with the above experimental setup. As shown in Table 1, two 2D projection-based backbones, namely PolarNet [38] and CPGNet [23], are incorporated with our framework to validate the proposed CFFE and CDM. Row (a) shows the original Panoptic-Polarnet [39], while rows (b,c,d) are implemented based on the official code of Panoptic-Polarnet [39]. The running time is reported by the time of the model inference plus that of the post-process.", + "bbox": [ + 75, + 335, + 468, + 470 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Effects of the CFFE. The CFFE is configured in three ways: 1) the CFFE is not applied (a,e,f,g,h); 2) the center feature generation (CFG) in the CFFE is removed (b,i), which is equal to the setting of only imposing the intermediate supervision; 3) the whole CFFE is applied (c,d,j). For the two backbones, a common observation is that the improvements primarily come from the CFG rather than the intermediate supervision, especially for the things classes (a,b,c and h,i,j). In terms of the running time, The CFFE brings acceptable latency of $+6.6\\mathrm{ms}$ and $+8.5\\mathrm{ms}$ for these two backbones, respectively.", + "bbox": [ + 75, + 474, + 468, + 641 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Moreover, the errors in predicting things center offsets are calculated for both the CFNet with CFFE and its intermediate results as shown in Table 2. Equipped with the backbone of CPGNet [23], the proposed CFFE reduces the errors of predicting the things center offsets. The observations demonstrate that explicitly modeling center features helps in identifying an instance.", + "bbox": [ + 75, + 642, + 468, + 748 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Effects of the CDM. For the proposal-free meth", + "bbox": [ + 96, + 750, + 467, + 766 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/21cded0f81ff2478bc96ed6c364e7815938ad7308d1ab4711488b2ef87db081f.jpg", + "table_caption": [ + "Table 1. Ablation studies on the SemanticKITTI validation set. RT: running time." + ], + "table_footnote": [], + "table_body": "
training setvalidation set
Intermediate1.351.22
CFNet1.191.13
", + "bbox": [ + 140, + 792, + 406, + 843 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 2. Mean errors measured in meter (m) of things center offsets for the intermediate results and the CFNet with the CFFE on the SemanticKITTI training and validation set.", + "bbox": [ + 75, + 854, + 467, + 896 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "ods, there are primarily three ways to identify each instance: 1) the class-agnostic clustering methods (e.g. DBSCAN, HDBSCAN, MeanShift) (e,f,g), 2) the BEV center heatmap [39] (a,b,c), and 3) our CDM (d,h,i,j). For the backbone of PolarNet [38], our CDM outperforms the BEV center heatmap [39] by $+0.2$ for the PQ (c,d), while our CDM runs faster. For the backbone of CPGNet [23], compared with the class-agnostic clustering methods, our CDM also achieves the best performance on both accuracy $(+0.6\\sim +2.2$ for the PQ) and running time $(-21.5\\mathrm{ms}\\sim -$ $81.2\\mathrm{ms})$ , though it slows down the model inference by $+1.1~\\mathrm{ms}$ $\\mathrm{(g,h)}$ owing to the additional confidence prediction.", + "bbox": [ + 496, + 310, + 890, + 492 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "More ablation studies can be seen in the supplementary material.", + "bbox": [ + 496, + 492, + 890, + 521 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Evaluation Comparisons", + "text_level": 1, + "bbox": [ + 498, + 530, + 725, + 546 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Quantitative results. In Table 3 and Table 4, our CFNet with CPGNet as its backbone is compared with the existing methods on the SemanticKITTI test set and nuScenes validation set, respectively. From top to bottom, the methods are grouped as proposal-based and proposal-free methods. It is found that our proposal-free CFNet surpasses the previous methods by a large margin on most metrics. Specifically, on the SemanticKITTI test set, our CFNet outperforms the previous state-of-the-art SCAN [35] and Panoptic-PHNet [22] by $+1.9$ for the PQ, and $+1.5$ for the RQ. On the nuScenes validation set, our CFNet outperforms the most competitive Panoptic-PHNet [22] by $+0.4$ for the PQ, $+0.6$ for the SQ, and $+0.4$ for the RQ, respectively.", + "bbox": [ + 496, + 553, + 890, + 750 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Running time. The running time is reported on the SemanticKITTI. It is measured with PyTorch FP32 without any model acceleration tricks on a single NVIDIA RTX 3090 GPU. For clarity, the running time of our method is further divided into two parts, including the time costs of the model and post-process. Based on the same kind of devices as the existing methods, our CFNet runs $43.5\\mathrm{ms}$ (41.2 ms for model inference and 2.3 ms for post-process) and is 1.6 times faster than the most efficient Panoptic-PHNet [22].", + "bbox": [ + 496, + 750, + 890, + 885 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Qualitative results. The visualization results are", + "bbox": [ + 517, + 885, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "13431", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/c1d1530e659ba9bc197f27396204ab5a02a6f269878bb2dc63286363bf2dfe16.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodsPQPQ†SQRQPQThSQThRQThPQStSQStRQStmIoURT(ms)
Proposal-based Methods
RangeNet++ [25]&PointPillars [19]37.145.975.947.020.275.225.249.376.562.852.4416.6
KPConv [33]&PointPillars [19]44.552.580.054.432.781.538.753.179.065.958.8526.3
PanopticTrackNet [17]43.150.778.853.928.680.435.553.677.767.352.6147
EfficientLPS [31]57.463.283.068.753.187.860.560.579.574.661.4-
Proposal-free Methods
LPSAD [24]38.047.076.548.225.676.831.847.176.260.150.984.7
DS-Net [15]55.962.582.366.755.187.262.856.578.769.561.6294.1
Panoster [13]52.759.980.764.149.483.358.555.178.868.259.9-
GP-S3Net [29]60.069.082.072.165.086.674.556.478.770.470.8270.3
SMAC-Seg [20]56.162.582.067.953.085.661.858.479.372.363.399
Panoptic-PolarNet [39]54.160.781.465.053.387.260.654.877.268.159.586.2
SCAN [35]61.567.584.572.161.488.169.361.581.874.167.778.1
Panoptic-PHNet [22]61.567.984.872.163.890.770.459.980.573.366.069.3
CFNet [Ours] w/CPGNet [23]63.469.785.273.666.789.874.361.081.973.168.341.2 + 2.3
", + "bbox": [ + 122, + 87, + 849, + 315 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/96ece7ee8e70c4b00886032f46ccd7728fc8d3c5494baeff9561740f059db2d0.jpg", + "table_caption": [ + "Table 3. Comparison results on the SemanticKITTI test set." + ], + "table_footnote": [], + "table_body": "
MethodsPQ\\( PQ^† \\)SQRQ\\( PQ^{Th} \\)\\( SQ^{Th} \\)\\( RQ^{Th} \\)\\( PQ^{St} \\)\\( SQ^{St} \\)\\( RQ^{St} \\)mIoU
Proposal-based Methods
Cylinder3D [40]&PointPillars [19]36.044.583.343.023.383.727.057.282.769.652.3
Cylinder3D [40]&SECOND [36]40.148.484.247.329.084.433.658.583.770.158.5
PanopticTrackNet [17]51.456.280.263.345.881.455.960.478.375.558.0
EfficientLPS [31]62.065.683.473.956.883.268.070.683.883.665.6
Proposal-free Methods
LPSAD [24]50.457.779.462.443.280.253.257.578.571.762.5
DS-Net [15]42.551.083.650.332.583.138.359.284.470.370.7
GP-S3Net [29]61.067.584.172.056.085.365.266.082.978.775.8
SMAC-Seg [20]67.071.885.078.265.287.174.268.882.982.272.2
Panoptic-PolarNet [39]63.467.283.975.359.284.170.370.483.683.566.9
SCAN [35]65.168.985.775.360.685.770.272.585.783.877.4
Panoptic-PHNet [22]74.777.788.284.274.089.082.575.986.886.979.7
CFNet [Ours] w/CPGNet [23]75.178.088.884.674.889.882.976.687.187.379.3
", + "bbox": [ + 122, + 352, + 848, + 556 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 4. Comparison results on the nuScenes validation set.", + "bbox": [ + 305, + 566, + 661, + 580 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/b052111d925cc62768d07df341b90122b38eeabbf009657386e66860535c3e11.jpg", + "image_caption": [ + "Semantic Segmentation" + ], + "image_footnote": [], + "bbox": [ + 88, + 606, + 274, + 752 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/e5f441e2eb16aac27dfee5184afbbbaaa802c81f1a6a4849e6c66705f19ab3fd.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 274, + 606, + 457, + 678 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/d6176e70d804c2a52d065fda3f5ff5c752045d44377594bd42ab595d7a3169a3.jpg", + "image_caption": [ + "Instance Segmentation", + "Figure 5. Visualizations of our CFNet on the SemanticKITTI test set. Different colors represent different classes or instances." + ], + "image_footnote": [], + "bbox": [ + 274, + 679, + 457, + 751 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "shown in Fig. 5. Our CFNet can distinguish adjacent pedestrians or cars. Besides, the boundaries of instances can be accurately segmented. More qualitative results can be seen in the supplementary material.", + "bbox": [ + 75, + 839, + 470, + 902 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 607, + 619, + 622 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "A novel proposal-free CFNet is proposed for real-time LiDAR panoptic segmentation. For better modeling and exploiting the non-existent instance centers, a novel CFFE is proposed to generate enhanced center-focusing feature maps and a CDM is introduced to keep only one center for each instance and then assign the shifted things points to the closest center for acquiring instance IDs. From the experiments, it can be seen that center modeling and exploiting is a key problem in the proposal-free LiDAR panoptic segmentation methods, and mimicking the non-existent center features is promising and shows a clear benefit.", + "bbox": [ + 496, + 632, + 893, + 799 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 500, + 814, + 668, + 830 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The research project is partially supported by the National Key R&D Program of China (No.2021ZD0111902), and the National Natural Science Foundation of China (No.U21B2038, 62272015, U19B2039).", + "bbox": [ + 496, + 839, + 892, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "13432", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Yara Ali Alnaggar, Mohamed Afifi, Karim Amer, and Mohamed ElHelw. Multi projection fusion for real-time semantic segmentation of 3d lidar point clouds. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1800-1809, 2021. 2", + "[2] Jens Behley, Martin Garbade, Andres Milioto, Jan Quenzel, Sven Behnke, Cyril Stachniss, and Jurgen Gall. Semantickitti: A dataset for semantic scene understanding of lidar sequences. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9297-9307, 2019. 6", + "[3] Jens Behley, Andres Milioto, and Cyril Stachniss. A benchmark for lidar-based panoptic segmentation based on kitti. In 2021 IEEE International Conference on Robotics and Automation (ICRA), pages 13596-13603. IEEE, 2021. 6", + "[4] Maxim Berman, Amal Rannen Triki, and Matthew B Blaschko. The lovasz-softmax loss: A tractable surrogate for the optimization of the intersection-over-union measure in neural networks. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 4413-4421, 2018. 5", + "[5] Holger Caesar, Varun Bankiti, Alex H Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuscenes: A multimodal dataset for autonomous driving. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11621-11631, 2020. 6", + "[6] Ricardo JGB Campello, Davoud Moulavi, and Jörg Sander. Density-based clustering based on hierarchical density estimates. In Pacific-Asia conference on knowledge discovery and data mining, pages 160–172. Springer, 2013. 7", + "[7] Ran Cheng, Ryan Razani, Ehsan Taghavi, Enxu Li, and Bingbing Liu. 2-s3net: Attentive feature fusion with adaptive feature selection for sparse semantic segmentation network. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12547-12556, 2021. 2", + "[8] Christopher Choy, JunYoung Gwak, and Silvio Savarese. 4d spatio-temporal convnets: Minkowski convolutional neural networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3075-3084, 2019. 2", + "[9] Dorin Comaniciu and Peter Meer. Mean shift: A robust approach toward feature space analysis. IEEE Transactions on pattern analysis and machine intelligence, 24(5):603-619, 2002. 3, 7", + "[10] Tiago Cortinhal, George Tzelepis, and Eren Erdal Aksoy. Salsanext: Fast, uncertainty-aware semantic segmentation of lidar point clouds. In International Symposium on Visual Computing, pages 207-222. Springer, 2020. 2, 4", + "[11] Martin Ester, Hans-Peter Kriegel, Jörg Sander, Xiaowei Xu, et al. A density-based algorithm for discovering clusters in large spatial databases with noise. In kdd, volume 96, pages 226-231, 1996. 3, 7", + "[12] Whye Kit Fong, Rohit Mohan, Juana Valeria Hurtado, Lubing Zhou, Holger Caesar, Oscar Beijbom, and Abhinav Val-" + ], + "bbox": [ + 78, + 114, + 472, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "ada. Panoptic nuscenes: A large-scale benchmark for lidar panoptic segmentation and tracking. IEEE Robotics and Automation Letters, 7(2):3795-3802, 2022. 6", + "[13] Stefano Gasperini, Mohammad-Ali Nikouei Mahani, Alvaro Marcos-Ramiro, Nassir Navab, and Federico Tombari. Panoster: End-to-end panoptic segmentation of lidar point clouds. IEEE Robotics and Automation Letters, 6(2):3216-3223, 2021. 1, 3, 8", + "[14] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In Proceedings of the IEEE international conference on computer vision, pages 2961-2969, 2017. 1, 2", + "[15] Fangzhou Hong, Hui Zhou, Xinge Zhu, Hongsheng Li, and Ziwei Liu. Lidar-based panoptic segmentation via dynamic shifting network. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13090-13099, 2021. 1, 3, 5, 6, 8", + "[16] Qingyong Hu, Bo Yang, Linhai Xie, Stefano Rosa, Yulan Guo, Zhihua Wang, Niki Trigoni, and Andrew Markham. Randla-net: Efficient semantic segmentation of large-scale point clouds. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11108-11117, 2020. 2", + "[17] Juana Valeria Hurtado, Rohit Mohan, Wolfram Burgard, and Abhinav Valada. Mopt: Multi-object panoptic tracking. arXiv preprint arXiv:2004.08189, 2020. 1, 2, 8", + "[18] Alexander Kirillov, Kaiming He, Ross Girshick, Carsten Rother, and Piotr Dollar. Panoptic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9404-9413, 2019. 1", + "[19] Alex H Lang, Sourabh Vora, Holger Caesar, Lubing Zhou, Jiong Yang, and Oscar Beijbom. Pointpillars: Fast encoders for object detection from point clouds. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12697-12705, 2019. 1, 2, 4, 8", + "[20] Enxu Li, Ryan Razani, Yixuan Xu, and Liu Bingbing. Smac-seg: Lidar panoptic segmentation via sparse multi-directional attention clustering. arXiv preprint arXiv:2108.13588, 2021. 2, 3, 8", + "[21] Enxu Li, Ryan Razani, Yixuan Xu, and Bingbing Liu. Cpseg: Cluster-free panoptic segmentation of 3d lidar point clouds. arXiv preprint arXiv:2111.01723, 2021. 1", + "[22] Jinke Li, Xiao He, Yang Wen, Yuan Gao, Xiaogiang Cheng, and Dan Zhang. Panoptic-phnet: Towards real-time and high-precision lidar panoptic segmentation via clustering pseudo heatmap. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11809–11818, 2022. 1, 3, 7, 8", + "[23] Xiaoyan Li, Gang Zhang, Hongyu Pan, and Zhenhua Wang. Cpgnet: Cascade point-grid fusion network for real-time lidar semantic segmentation. arXiv preprint arXiv:2204.09914, 2022. 2, 4, 5, 6, 7, 8", + "[24] Andres Milioto, Jens Behley, Chris McCool, and Cyril Stachniss. Lidar panoptic segmentation for autonomous driving. In 2020 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 8505-8512. IEEE, 2020. 2, 8" + ], + "bbox": [ + 503, + 92, + 893, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "13433", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[25] Andres Milioto, Ignacio Vizzo, Jens Behley, and Cyril Stachniss. Rangenet++: Fast and accurate lidar semantic segmentation. In 2019 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 4213-4220. IEEE, 2019. 2, 4, 8", + "[26] Lorenzo Porzi, Samuel Rota Bulo, Aleksander Colovic, and Peter Kontschieder. Seamless scene segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8277-8286, 2019. 6", + "[27] Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. Pointnet: Deep learning on point sets for 3d classification and segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 652-660, 2017. 2", + "[28] Charles R Qi, Li Yi, Hao Su, and Leonidas J Guibas. Point-net++: Deep hierarchical feature learning on point sets in a metric space. arXiv preprint arXiv:1706.02413, 2017. 2", + "[29] Ryan Razani, Ran Cheng, Enxu Li, Ehsan Taghavi, Yuan Ren, and Liu Bingbing. Gp-s3net: Graph-based panoptic sparse semantic segmentation network. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 16076-16085, 2021. 1, 3, 8", + "[30] Shaoshuai Shi, Chaoxu Guo, Li Jiang, Zhe Wang, Jianping Shi, Xiaogang Wang, and Hongsheng Li. Pv-rcnn: Pointvoxel feature set abstraction for 3d object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10529–10538, 2020. 1", + "[31] Kshitij Sirohi, Rohit Mohan, Daniel Buscher, Wolfram Burgard, and Abhinav Valada. Efficienttps: Efficient lidar panoptic segmentation. IEEE Transactions on Robotics, 2021. 1, 2, 8", + "[32] Haotian Tang, Zhijian Liu, Shengyu Zhao, Yujun Lin, Ji Lin, Hanrui Wang, and Song Han. Searching efficient 3d architectures with sparse point-voxel convolution. In European Conference on Computer Vision, pages 685-702. Springer, 2020. 2", + "[33] Hugues Thomas, Charles R Qi, Jean-Emmanuel Deschaud, Beatrix Marcotegui, François Goulette, and Leonidas J Guibas. Kpconv: Flexible and deformable convolution for point clouds. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6411–6420, 2019. 2, 8", + "[34] Chenfeng Xu, Bichen Wu, Zining Wang, Wei Zhan, Peter Vajda, Kurt Keutzer, and Masayoshi Tomizuka. Squeeze-segv3: Spatially-adaptive convolution for efficient pointcloud segmentation. In European Conference on Computer Vision, pages 1-19. Springer, 2020. 2, 4", + "[35] Shuangjie Xu, Rui Wan, Maosheng Ye, Xiaoyi Zou, and Tongyi Cao. Sparse cross-scale attention network for efficient lidar panoptic segmentation. arXiv preprint arXiv:2201.05972, 2022. 1, 2, 3, 7, 8", + "[36] Yan Yan, Yuxing Mao, and Bo Li. Second: Sparsely embedded convolutional detection. Sensors, 18(10):3337, 2018. 8", + "[37] Dongqiangzi Ye, Zixiang Zhou, Weijia Chen, Yufei Xie, Yu Wang, Panqu Wang, and Hassan Foroosh. Lidarmultinet: Towards a unified multi-task network for lidar perception. arXiv preprint arXiv:2209.09385, 2022. 1, 2" + ], + "bbox": [ + 78, + 90, + 468, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[38] Yang Zhang, Zixiang Zhou, Philip David, Xiangyu Yue, Zerong Xi, Boqing Gong, and Hassan Foroosh. Polarnet: An improved grid representation for online lidar point clouds semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9601-9610, 2020. 2, 4, 6, 7", + "[39] Zixiang Zhou, Yang Zhang, and Hassan Foroosh. Panoptic-polarnet: Proposal-free lidar point cloud panoptic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13194-13203, 2021. 1, 2, 3, 6, 7, 8", + "[40] Xinge Zhu, Hui Zhou, Tai Wang, Fangzhou Hong, Yuexin Ma, Wei Li, Hongsheng Li, and Dahua Lin. Cylindrical and asymmetrical 3d convolution networks for lidar segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9939-9948, 2021. 2, 8" + ], + "bbox": [ + 501, + 90, + 892, + 329 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "13434", + "bbox": [ + 480, + 945, + 519, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2023/Center Focusing Network for Real-Time LiDAR Panoptic Segmentation/e749a0dd-41b1-4b7d-9325-ea72d56e45cb_model.json b/2023/Center Focusing Network for Real-Time LiDAR Panoptic Segmentation/e749a0dd-41b1-4b7d-9325-ea72d56e45cb_model.json new file mode 100644 index 0000000000000000000000000000000000000000..6d527bff92f8644c04dc898a63114976104bdf5d --- /dev/null +++ b/2023/Center Focusing Network for Real-Time LiDAR Panoptic Segmentation/e749a0dd-41b1-4b7d-9325-ea72d56e45cb_model.json @@ -0,0 +1,2255 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.0, + 0.812, + 0.045 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.124, + 0.131, + 0.848, + 0.154 + ], + "angle": 0, + "content": "Center Focusing Network for Real-Time LiDAR Panoptic Segmentation" + }, + { + "type": "text", + "bbox": [ + 0.17, + 0.179, + 0.793, + 0.199 + ], + "angle": 0, + "content": "Xiaoyan Li\\(^{1,2}\\) Gang Zhang\\(^{3}\\) Boyue Wang\\(^{1,2}\\) Yongli Hu\\(^{1,2}\\) Baocai Yin\\(^{1,2}\\)" + }, + { + "type": "text", + "bbox": [ + 0.169, + 0.199, + 0.802, + 0.217 + ], + "angle": 0, + "content": "1Beijing Municipal Key Lab of Multimedia and Intelligent Software Technology" + }, + { + "type": "text", + "bbox": [ + 0.175, + 0.217, + 0.796, + 0.252 + ], + "angle": 0, + "content": "\\(^{2}\\)Beijing Institute of Artificial Intelligence, Faculty of Information Technology, Beijing University of Technology, Beijing 100124, China" + }, + { + "type": "text", + "bbox": [ + 0.191, + 0.252, + 0.778, + 0.27 + ], + "angle": 0, + "content": "3Mogo Auto Intelligence and Telemetics Information Technology Co. Ltd." + }, + { + "type": "text", + "bbox": [ + 0.17, + 0.272, + 0.796, + 0.287 + ], + "angle": 0, + "content": "{xiaoyan.li, wby, huyongli, ybc}@bjut.edu.cn zhanggang11021136@gmail.com" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.322, + 0.314, + 0.339 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.356, + 0.474, + 0.659 + ], + "angle": 0, + "content": "LiDAR panoptic segmentation facilitates an autonomous vehicle to comprehensively understand the surrounding objects and scenes and is required to run in real time. The recent proposal-free methods accelerate the algorithm, but their effectiveness and efficiency are still limited owing to the difficulty of modeling non-existent instance centers and the costly center-based clustering modules. To achieve accurate and real-time LiDAR panoptic segmentation, a novel center focusing network (CFNet) is introduced. Specifically, the center focusing feature encoding (CFFE) is proposed to explicitly understand the relationships between the original LiDAR points and virtual instance centers by shifting the LiDAR points and filling in the center points. Moreover, to leverage the redundantly detected centers, a fast center dedduplication module (CDM) is proposed to select only one center for each instance. Experiments on the SemanticKITTI and nuScenes panoptic segmentation benchmarks demonstrate that our CFNet outperforms all existing methods by a large margin and is 1.6 times faster than the most efficient method." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.692, + 0.21, + 0.709 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.72, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Panoptic segmentation [18] combines both semantic segmentation and instance segmentation in a single framework. It predicts semantic labels for the uncountable stuff classes (e.g. road, sidewalk), while it simultaneously provides semantic labels and instance IDs for the countable things classes (e.g. car, pedestrian). The LiDAR panoptic segmentation is one of the bases for the safety of autonomous driving, which employs the point clouds collected by the Light Detection and Ranging (LiDAR) sensors to effectively depict the surroundings. Existing LiDAR panoptic segmentation methods first conduct semantic segmentation, and then achieve instance segmentation for the things categories in" + }, + { + "type": "image", + "bbox": [ + 0.51, + 0.32, + 0.883, + 0.549 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.558, + 0.894, + 0.601 + ], + "angle": 0, + "content": "Figure 1. PQ vs. runtime on the SemanticKITTI test set. Runtime measurements are taken on a single NVIDIA RTX 3090 GPU. The panoptic quality (PQ) is introduced in section 4.1." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.628, + 0.88, + 0.643 + ], + "angle": 0, + "content": "two ways, the proposal-based and proposal-free methods." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.644, + 0.892, + 0.765 + ], + "angle": 0, + "content": "The proposal-based methods [17, 31, 37] adopt a two-stage process similar to the well-known Mask R-CNN [14] in the image domain. It first generates object proposals for the things points by using 3D detection networks [19, 30] and then refines the instance segmentation results within each proposal. As shown in Fig. 1, these methods are usually complicated and hardly achieve real-time processing, owing to their sequential multi-stage pipelines." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.892, + 0.902 + ], + "angle": 0, + "content": "The proposal-free frameworks [13, 15, 21, 22, 29, 35, 39] are more compact. To associate the things points with instance IDs, these methods usually leverage the instance centers. Specifically, they regress the offsets from the points to their corresponding centers, and then adopt the class-agnostic center-based clustering modules [13, 15, 29] or the bird's-eye view (BEV) center heatmap [22, 35, 39]. However, two problems exist in these methods. First, for center feature extracting and center modeling, the non-existent in" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13425" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.09, + 0.089, + 0.264, + 0.199 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.098, + 0.203, + 0.256, + 0.216 + ], + "angle": 0, + "content": "(a) CFNet without CFFE" + }, + { + "type": "image", + "bbox": [ + 0.284, + 0.089, + 0.457, + 0.199 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.302, + 0.202, + 0.441, + 0.216 + ], + "angle": 0, + "content": "(b) CFNet with CFFE" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.228, + 0.47, + 0.271 + ], + "angle": 0, + "content": "Figure 2. Instance segmentation of a car. Without our CFFE, the car is split into parts (a), while the CFFE significantly alleviates this problem (b). Different colors represent different instances." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.294, + 0.468, + 0.46 + ], + "angle": 0, + "content": "stance centers increase the difficulty, considering that the LiDAR points are usually surface-aggregated [35] and an instance center is imaginary in most cases. As shown in Fig. 2(a), the difficulty often results in the fault that one instance is incorrectly split into several parts. Second, for exploiting the redundantly detected centers, the clustering modules (e.g. MeanShift, DBSCAN) are too time-consuming to support the real-time autonomous driving perception systems, while the BEV center heatmap cannot distinguish objects with different altitudes in the same BEV grid." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.461, + 0.468, + 0.762 + ], + "angle": 0, + "content": "For accurate and fast LiDAR panoptic segmentation, a proposal-free center focusing network (CFNet) is proposed. For better encoding center features, a novel center focusing feature encoding (CFFE) is proposed to generate center-focusing feature maps by shifting the things points to fill in the non-existent instance centers for more accurate predictions (as shown in Fig. 2(b)). For center modeling, the CFNet not only decomposes the panoptic segmentation task into the widely-used semantic segmentation and center offset regression, but also proposes a new confidence score prediction for indicating the accuracy of the center offset regression. Subsequently, for the detected centers exploiting, a novel center dedduplication module (CDM) is designed to select one center for a single instance. The CDM keeps the predicted centers with higher confidence scores, while suppressing the ones with lower confidence. Finally, instance segmentation is achieved by assigning the shifted things points to the closest center. For efficiency, the proposed CFNet is built on the 2D projection-based segmentation paradigm. Our contributions are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.771, + 0.468, + 0.832 + ], + "angle": 0, + "content": "- A proposal-free CFNet is proposed to achieve accurate and fast LiDAR panoptic segmentation by solving the bottleneck problems of center modeling and center-based clustering in previous methods." + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.84, + 0.468, + 0.9 + ], + "angle": 0, + "content": "- The CFFE is proposed to alleviate the difficulty of modeling the non-existent instance centers and the CDM is designed to efficiently keep one center for each instance." + }, + { + "type": "list", + "bbox": [ + 0.096, + 0.771, + 0.468, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.52, + 0.092, + 0.892, + 0.152 + ], + "angle": 0, + "content": "- The proposed CFNet is evaluated on the nuScenes and SemanticKITTI LiDAR panoptic segmentation benchmarks. Our CFNet achieves the state-of-the-art performance with a real-time inference speed." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.166, + 0.642, + 0.181 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.191, + 0.892, + 0.341 + ], + "angle": 0, + "content": "The LiDAR panoptic segmentation methods usually adopt the LiDAR semantic segmentation networks as the backbone and jointly optimize the semantic segmentation and instance segmentation tasks. The results of the two tasks are fused to generate the final panoptic segmentation results. The LiDAR semantic segmentation backbone is first introduced. Then, the panoptic segmentation methods, achieving instance segmentation upon the semantic segmentation backbone, are grouped into the proposal-based and proposal-free methods, and are further discussed." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.342, + 0.893, + 0.658 + ], + "angle": 0, + "content": "LiDAR semantic segmentation backbone. The backbone network for LiDAR semantic segmentation can be divided into three groups: point-based, voxel-based, and 2D projection-based methods. The point-based methods [16, 27, 28, 33] directly operate on the raw point clouds but are extremely time-consuming due to the expensive local neighbor searching. The voxel-based methods [7, 8, 32, 40] discretize the point clouds into structured voxels, where sparse 3D convolutions are applied. These methods are still difficult to meet the real-time applications, although they achieve the highest accuracy. The 2D projection-based methods are more efficient since most of their computation is done in the 2D space, such as range view (RV) [10,25,34], bird's-eye view (BEV) [19], polar view [38], and multiview fusion [1, 23]. For fast LiDAR panoptic segmentation, the Panoptic-PolarNet [39], SMAC-Seg [20], and LPSAD [24] all adopt the 2D projection-based backbone. A recent real-time semantic segmentation method, namely CPGNet [23], is a 2D projection-based one that explores an end-to-end multi-view fusion framework by fusing the features from the point, BEV, and RV." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.66, + 0.892, + 0.84 + ], + "angle": 0, + "content": "**Proposal-based methods.** The proposal-based methods conduct instance segmentation through a two-stage complicated process. They first detect the foreground instances and then refine the instance segmentation results independently within each detected bounding box. Based on the Mask R-CNN [14] for instance segmentation, the MOPT [17] and EfficientLPS [31] insert a semantic branch to achieve panoptic segmentation on the range view (RV). Recently, LidarMultiNet [37] unifies LiDAR-based 3D object detection, semantic segmentation, and panoptic segmentation in a single framework to reduce the computation cost by sharing a strong voxel-based backbone." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Proposal-free methods. The proposal-free methods usually apply the class-agnostic clustering to the things points to conduct instance segmentation. The LPSAD [24] clusters points into instances by regressing the center off" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13426" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.078, + 0.089, + 0.891, + 0.332 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.081, + 0.334, + 0.449, + 0.416 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.48, + 0.334, + 0.889, + 0.416 + ], + "angle": 0, + "content": null + }, + { + "type": "image_footnote", + "bbox": [ + 0.104, + 0.424, + 0.248, + 0.441 + ], + "angle": 0, + "content": ": Multi-Layer Perception" + }, + { + "type": "image_footnote", + "bbox": [ + 0.278, + 0.429, + 0.38, + 0.44 + ], + "angle": 0, + "content": "P2G: Point to Grid" + }, + { + "type": "image_footnote", + "bbox": [ + 0.406, + 0.428, + 0.592, + 0.44 + ], + "angle": 0, + "content": "FCN: Fully Convolutional Network" + }, + { + "type": "image_footnote", + "bbox": [ + 0.631, + 0.429, + 0.852, + 0.442 + ], + "angle": 0, + "content": "CFFE: Center Focusing Feature Encoding" + }, + { + "type": "image_footnote", + "bbox": [ + 0.104, + 0.446, + 0.245, + 0.464 + ], + "angle": 0, + "content": ": Fully Connected Layer" + }, + { + "type": "image_footnote", + "bbox": [ + 0.279, + 0.448, + 0.38, + 0.459 + ], + "angle": 0, + "content": "G2P: Grid to Point" + }, + { + "type": "image_footnote", + "bbox": [ + 0.407, + 0.447, + 0.496, + 0.458 + ], + "angle": 0, + "content": "PF: Point Fusion" + }, + { + "type": "image_footnote", + "bbox": [ + 0.631, + 0.447, + 0.823, + 0.459 + ], + "angle": 0, + "content": "CDM: Center Dedduplication Module" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.48, + 0.893, + 0.536 + ], + "angle": 0, + "content": "Figure 3. The overview of our CFNet. It consists of four steps: 1) the 2D projection-based backbone extracts features on the 2D space; 2) the proposed center focusing feature encoding (CFFE) mimics and enhances the non-existent instance center features; 3) the panoptic segmentation head predicts the output results; 4) the proposed center dedduplication module (CDM) achieves instance segmentation that is fused to generate the final panoptic segmentation results. The dashed lines indicate that the operations are only used during inference." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.562, + 0.473, + 0.807 + ], + "angle": 0, + "content": "sets. The Panoster [13] directly predicts the instance IDs from a learning-based clustering module, where the time-consuming DBSCAN [11] is used to refine the instance segmentation results. The DS-Net [15] proposes a learnable dynamic shifting module that adjusts the kernel functions of the MeanShift [9] to handle instances with various sizes. The GP-S3Net [29] constructs a graph convolutional network (GCN) on the over-segmentation clusters to identify instances. The SMAC-Seg [20] proposes a sparse multidirectional attention clustering and center-aware repel loss for instance segmentation. These methods adopt the time-consuming clustering methods. To further accelerate the algorithm, recently, Panoptic-PolarNet [39], SCAN [35], and Panoptic-PHNet [22] adopt the BEV center heatmap and center offsets for instance segmentation. However, the BEV center heatmap is also costly and confuses on \\(z\\)-axis." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.827, + 0.186, + 0.844 + ], + "angle": 0, + "content": "3. Approach" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.856, + 0.47, + 0.903 + ], + "angle": 0, + "content": "The input of the LiDAR panoptic segmentation task is the LiDAR point clouds, given by a set of unordered points \\(\\pmb{P} = \\{(p_i,f_i)\\}_{i = 0}^{N - 1}\\) (where \\(\\pmb {p}_i = (x,y,z)\\) is the 3D coor" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.562, + 0.892, + 0.685 + ], + "angle": 0, + "content": "dinates in the cartesian space and \\( \\pmb{f}_i \\) denotes additional LiDAR point features, e.g. intensity). The task aims to assign a set of labels \\( \\pmb{L} = \\{(s_i, inst_i)\\}_{i=0}^{N-1} \\) to these points, where \\( s_i \\) is the semantic label (e.g. road, building, car, pedestrian), and \\( inst_i \\) is the instance ID of the \\( i^{th} \\) point. Moreover, \\( s_i \\) can be divided into the uncountable stuff classes (e.g. road, building) and countable things classes (e.g. car, pedestrian). The instance IDs of stuff points are set as 0." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.689, + 0.893, + 0.901 + ], + "angle": 0, + "content": "To predict the labels \\(L\\) of the input LiDAR point clouds \\(P\\), our CFNet decomposes this process into four steps, as shown in Fig. 3: 1) an off-the-shelf 2D projection-based backbone is applied to efficiently extract features on the 2D spaces (e.g. BEV, RV); 2) a novel center focusing feature encoding (CFFE) is used to generate center-focusing feature maps for more accurate predictions of the instance centers; 3) the panoptic segmentation head fuses the features from the 3D points and 2D spaces to predict the semantic segmentation results, the center offsets, and the confidence scores of the center offsets, respectively; 4) during inference, a post-process is conducted to produce the panoptic segmentation results, where a novel center dedduplication module (CDM) operates on the shifted things points to se" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13427" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.092, + 0.891, + 0.282 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.293, + 0.896, + 0.337 + ], + "angle": 0, + "content": "Figure 4. The proposed center focusing feature encoding (CFFE). The \"Conv\" represents a 2D convolution with \\(3 \\times 3\\) kernels, a batch normalization, and a ReLU layer. The details of the semantic branch and instance branch are shown in Fig. 3. The blue arrows are coordinate-related operations." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.362, + 0.351, + 0.377 + ], + "angle": 0, + "content": "lect only one center for a single instance." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.377, + 0.47, + 0.424 + ], + "angle": 0, + "content": "The design of the backbone in the first step is beyond the scope of the paper, and the last three mentioned steps are illustrated in detail in the following subsections." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.434, + 0.382, + 0.451 + ], + "angle": 0, + "content": "3.1. Center Focusing Feature Encoding" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.458, + 0.47, + 0.594 + ], + "angle": 0, + "content": "As mentioned above, the LiDAR points of an object are usually surface-aggregated, especially for car and truck categories, resulting in the fact that the center of an object is imaginary and does not exist in the LiDAR point clouds. To encode the features of the non-existent centers, a novel center focusing feature encoding (CFFE) is proposed, which takes the 2D features from the backbone and the 3D point coordinates as inputs and generates the enhanced center-focusing feature maps as shown in Fig. 3." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.595, + 0.47, + 0.64 + ], + "angle": 0, + "content": "The CFFE module consists of three steps, including intermediate result prediction, center feature generation, and feature enhancement module, as shown in Fig. 4." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.641, + 0.471, + 0.777 + ], + "angle": 0, + "content": "Intermediate result prediction. In this step, the CFFE predicts intermediate results (including the semantic segmentation, center offset, and its confidence scores) according to the 2D features \\( \\pmb{F}_{m} \\) and 3D point features \\( \\pmb{F}_{p} \\) for subsequent center feature simulation. Specifically, two convolution layers are applied on the 2D features \\( \\pmb{F}_{m} \\) independently, to generate semantic features \\( \\pmb{F}_{m}^{sem} \\) and instance features \\( \\pmb{F}_{m}^{ins} \\) (\\( m \\) is the specific 2D view, such as RV [10, 25, 34], BEV [19], and polar view [38].)" + }, + { + "type": "equation", + "bbox": [ + 0.187, + 0.79, + 0.469, + 0.807 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {F} _ {m} ^ {\\text {s e m}} = \\operatorname {C o n v} \\left(\\boldsymbol {F} _ {m}; \\boldsymbol {\\theta} _ {1}\\right), \\tag {1}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.194, + 0.81, + 0.469, + 0.828 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {F} _ {m} ^ {\\text {i n s}} = \\operatorname {C o n v} \\left(\\boldsymbol {F} _ {m}; \\boldsymbol {\\theta} _ {2}\\right), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.471, + 0.904 + ], + "angle": 0, + "content": "where \\( Conv \\) denotes sequential 2D convolution, batch normalization, and ReLU operations, and \\( \\theta_{1} \\) and \\( \\theta_{2} \\) are their learnable parameters. Then, the semantic branch generates per-point 3D semantic features \\( F_{p}^{sem} \\) by fusing the point" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.361, + 0.8, + 0.379 + ], + "angle": 0, + "content": "features \\(F_{p}\\) and 2D semantic features \\(\\pmb{F}_{m}^{sem}\\)" + }, + { + "type": "equation", + "bbox": [ + 0.595, + 0.388, + 0.892, + 0.407 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {F} _ {p} ^ {\\text {s e m}} = \\operatorname {S e g} \\left(\\boldsymbol {F} _ {p}, \\boldsymbol {F} _ {m} ^ {\\text {s e m}}; \\boldsymbol {\\theta} _ {3}\\right), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.415, + 0.892, + 0.431 + ], + "angle": 0, + "content": "where \\(Seg\\) is the semantic branch, and \\(\\theta_{3}\\) is the parameters." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.432, + 0.893, + 0.514 + ], + "angle": 0, + "content": "Finally, the intermediate semantic results \\(\\hat{S}^{\\prime} = \\{s_{i}^{\\prime}\\}_{i=0}^{N-1}\\) is produced based on \\(F_{p}^{sem}\\). The intermediate results of center offsets \\(\\hat{O}^{\\prime} = \\{o_{i}^{\\prime}\\}_{i=0}^{N-1}\\) and confidence scores \\(\\hat{C}^{\\prime} = \\{c_{i}^{\\prime}\\}_{i=0}^{N-1}\\) are predicted by the instance branch with the point features \\(F_{p}\\) and 2D instance features \\(F_{m}^{ins}\\)," + }, + { + "type": "equation", + "bbox": [ + 0.622, + 0.536, + 0.891, + 0.557 + ], + "angle": 0, + "content": "\\[\n\\hat {\\boldsymbol {S}} ^ {\\prime} = F C \\left(\\boldsymbol {F} _ {p} ^ {s e m}; \\boldsymbol {\\theta} _ {4}\\right), \\tag {4}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.595, + 0.56, + 0.891, + 0.579 + ], + "angle": 0, + "content": "\\[\n\\hat {\\boldsymbol {O}} ^ {\\prime}, \\hat {\\boldsymbol {C}} ^ {\\prime} = I n s \\left(\\boldsymbol {F} _ {p}, \\boldsymbol {F} _ {m} ^ {i n s}; \\boldsymbol {\\theta} _ {5}\\right), \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.589, + 0.892, + 0.664 + ], + "angle": 0, + "content": "where \\(Ins\\) is the instance branch, and \\(FC\\) represents the fully-connected layer. The structure and the training objective of semantic and instance branches are the same as those in the panoptic segmentation head and are elucidated in Fig. 3 and section 3.2." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.665, + 0.892, + 0.756 + ], + "angle": 0, + "content": "Center feature generation (CFG). In this step, the CFFE generates the shifted center features \\( \\pmb{F}_{p\\rightarrow m}^{sem} \\) by shifting the 3D semantic point features \\( \\pmb{F}_{p}^{sem} \\) to the predicted centers according to the mentioned intermediate results. First, the coordinates of a predicted center \\( \\pmb{p}_i^{shift} \\) are computed by," + }, + { + "type": "equation", + "bbox": [ + 0.59, + 0.766, + 0.891, + 0.786 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {p} _ {i} ^ {\\text {s h i f t}} = \\boldsymbol {p} _ {i} + \\hat {\\boldsymbol {o}} _ {i} ^ {\\prime} * \\mathbb {1} \\left[ \\hat {c} _ {i} ^ {\\prime} > \\delta \\right], \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.795, + 0.892, + 0.855 + ], + "angle": 0, + "content": "where \\(\\pmb{p}_i\\) is the original 3D LiDAR point coordinates, and \\(\\mathbb{1}[\\hat{c}_i^{\\prime} > \\delta ]\\) is a binary indicator of whether the confidence \\(\\hat{c}_i^\\prime\\) is greater than \\(\\delta = 0.2\\). In other words, it does not shift the stuff points or the things points with low confidence scores." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.855, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Then, the shifted 3D points \\(\\{\\pmb{p}_i^{shift}\\}_{i=0}^{N-1}\\) are served as the new coordinates of the feature points, and the Point to Grid (P2G) operation in [23] re-projects the 3D semantic" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13428" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.091, + 0.468, + 0.122 + ], + "angle": 0, + "content": "features \\( \\pmb{F}_{p}^{sem} \\) onto the 2D projected feature maps \\( \\pmb{F}_{p\\rightarrow m}^{sem} \\) with this new coordinates," + }, + { + "type": "equation", + "bbox": [ + 0.145, + 0.129, + 0.468, + 0.151 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {F} _ {p \\rightarrow m} ^ {s e m} = P 2 G (\\boldsymbol {F} _ {p} ^ {s e m}; \\{\\boldsymbol {p} _ {i} ^ {\\text {s h i f t}} \\} _ {i = 0} ^ {N - 1}). \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.157, + 0.468, + 0.202 + ], + "angle": 0, + "content": "Compared with \\( F_{p}^{sem} \\), the feature maps \\( F_{p\\rightarrow m}^{sem} \\) pay more attention to the imaginary centers since most of the things points have been shifted to their predicted centers." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.203, + 0.468, + 0.341 + ], + "angle": 0, + "content": "Feature enhancement module (FEM). The CFFE finally fuses the semantic feature maps \\( \\pmb{F}_{m}^{sem} \\) and the reprojected shifted center feature maps \\( \\pmb{F}_{p\\rightarrow m}^{sem} \\) to generate center-focusing semantic feature maps \\( \\pmb{F}_{m}^{CFsem} \\) and instance feature maps \\( \\pmb{F}_{m}^{CFS} \\), which are used by the subsequent semantic and instance branches for more accurate predictions. The enhancement module is composed of a simple concatenation operation and several convolution layers, whose details are in the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.342, + 0.468, + 0.448 + ], + "angle": 0, + "content": "In addition, for application of the multi-view fusion backbone (e.g. [23] shown in the supplementary material), the feature maps \\( \\mathbf{F}_m \\), \\( \\mathbf{F}_m^{sem} \\), \\( \\mathbf{F}_m^{ins} \\), \\( \\mathbf{F}_{p\\rightarrow m}^{sem} \\) (e.g. \\( m \\in \\{\\mathrm{RV},\\mathrm{BEV}\\} \\)) of each view are computed independently according to the above process. Then, they are fused in the point fusion (PF) module [23] to generate integrated 3D point features and per-point prediction." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.455, + 0.336, + 0.472 + ], + "angle": 0, + "content": "3.2. Panoptic Segmentation Head" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.479, + 0.468, + 0.599 + ], + "angle": 0, + "content": "For better modeling the instance center, the panoptic segmentation head uses a semantic branch to predict the semantic segmentation, and an instance branch to estimate both the center offsets and the newly introduced confidence scores, given the center-focusing semantic and instance feature maps. The architectures of both branches follow the segmentation head in [23] and are briefly introduced, while their training objectives are elaborated on." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.6, + 0.468, + 0.796 + ], + "angle": 0, + "content": "Semantic branch. For a per-point prediction, the semantic branch first applies the Grid to Point (G2P) operations to acquire the 3D point-wise features from the 2D semantic feature maps. Then, a PF module fuses the pointwise features from the G2P operation and the original 3D points to generate point-wise semantic features. The details of the G2P and PF can be found in [23]. After obtaining the point-wise semantic features, a fully connected (FC) layer is used to predict the final per-point semantic result \\(\\hat{\\psi}_i^j\\) (\\(\\hat{\\psi}_i^j \\in [0,1]\\)). \\(\\hat{\\psi}_i^j\\) represents the probability of the \\(i^{th}\\) LiDAR point belonging to the \\(j^{th}\\) class. The predicted semantic label \\(\\hat{s}_i\\) is acquired by selecting the most probable class \\(\\hat{s}_i = \\arg \\max_j \\hat{\\psi}_i^j\\)." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.797, + 0.468, + 0.856 + ], + "angle": 0, + "content": "Referring to the CPGNet [23], the same loss functions are adopted, including weighted cross entropy (WCE) loss \\(\\mathcal{L}_{wce}\\), Lovász-Softmax loss \\(\\mathcal{L}_{ls}\\) [4], and transformation consistency loss \\(\\mathcal{L}_{tc}\\)." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.856, + 0.468, + 0.901 + ], + "angle": 0, + "content": "Instance branch. Similar to the semantic branch, the instance branch also adopts the G2P operation and a PF module to acquire point-wise instance features. An FC layer" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.891, + 0.137 + ], + "angle": 0, + "content": "is applied to predict the per-point center offsets \\(\\hat{o}_i\\). The ground-truth for \\(\\hat{o}_i\\) is the offset vector \\(e_i - p_i\\) from the \\(i^{th}\\) point \\(p_i\\) to its corresponding instance center \\(e_i\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.137, + 0.891, + 0.181 + ], + "angle": 0, + "content": "For center offset regression, the loss function for optimizing \\(\\hat{o}_i\\) only considers the things classes and is formulated as follows," + }, + { + "type": "equation", + "bbox": [ + 0.513, + 0.191, + 0.891, + 0.232 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\boldsymbol {o} _ {i}} = \\left\\{ \\begin{array}{l l} | | \\hat {\\boldsymbol {o}} _ {i} - (\\boldsymbol {e} _ {i} - \\boldsymbol {p} _ {i}) | | _ {2} & \\text {i f} s _ {i} \\in t h i n g s \\text {c l a s s e s}, \\\\ 0 & \\text {o t h e r w i s e}, \\end{array} \\right. \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.241, + 0.89, + 0.271 + ], + "angle": 0, + "content": "where \\( \\pmb{e}_i \\) is the axis-aligned center of the instance [15]. Then, the losses are summarized as the following," + }, + { + "type": "equation", + "bbox": [ + 0.613, + 0.281, + 0.891, + 0.322 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {o} = \\frac {1}{N ^ {\\text {t h i n g s}}} \\sum_ {i = 0} ^ {N - 1} \\mathcal {L} _ {\\boldsymbol {o} _ {i}}, \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.333, + 0.89, + 0.363 + ], + "angle": 0, + "content": "where \\(N\\) and \\(N^{things}\\) are the numbers of all points and things points, respectively." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.364, + 0.891, + 0.439 + ], + "angle": 0, + "content": "For confidence score regression, another FC layer is used to predict the per-point confidence score \\(\\hat{c}_i\\) to indicate the accuracy degree of \\(\\hat{o}_i\\). \\(\\hat{c}_i\\) is activated by a sigmoid function to ensure \\(\\hat{c}_i \\in [0,1]\\). The ground-truth label \\(c_i\\) for supervising \\(\\hat{c}_i\\) is generated by" + }, + { + "type": "equation", + "bbox": [ + 0.538, + 0.449, + 0.891, + 0.492 + ], + "angle": 0, + "content": "\\[\nc _ {i} = \\left\\{ \\begin{array}{l l} \\exp \\left(- \\frac {\\mathcal {L} _ {o _ {i}} ^ {2}}{2 \\sigma^ {2}}\\right) & \\text {i f} s _ {i} \\in \\text {t h i n g s c l a s s e s}, \\\\ 0 & \\text {o t h e r w i s e .} \\end{array} \\right. \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.501, + 0.891, + 0.545 + ], + "angle": 0, + "content": "For the things points, the lower \\(\\mathcal{L}_{o_i}\\) is, the higher \\(c_{i}\\) is. It means that the point with a more accurate regression of the center offset has a higher confidence score." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.546, + 0.89, + 0.562 + ], + "angle": 0, + "content": "The weighted binary cross entropy loss \\(\\mathcal{L}_{wbce}\\) is applied," + }, + { + "type": "equation", + "bbox": [ + 0.508, + 0.584, + 0.891, + 0.625 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {w b c e} = - \\frac {1}{N} \\sum_ {i = 0} ^ {N - 1} 6 c _ {i} \\log (\\hat {c} _ {i}) + (1 - c _ {i}) \\log (1 - \\hat {c} _ {i}), \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.635, + 0.89, + 0.664 + ], + "angle": 0, + "content": "where the things points are manually emphasized since the number of them is much less than that of the stuff points." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.665, + 0.89, + 0.694 + ], + "angle": 0, + "content": "Finally, the loss for each group of results (from the CFNet or CFFE) is defined as" + }, + { + "type": "equation", + "bbox": [ + 0.558, + 0.706, + 0.891, + 0.722 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\mathcal {L} _ {w c e} + 3 \\mathcal {L} _ {l s} + \\mathcal {L} _ {t c} + 2 \\mathcal {L} _ {o} + \\mathcal {L} _ {w b c e}. \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.733, + 0.89, + 0.762 + ], + "angle": 0, + "content": "The total loss is the sum of the two losses sourced from the CFNet and CFFE." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.772, + 0.765, + 0.789 + ], + "angle": 0, + "content": "3.3. Center Dedduplication Module" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.796, + 0.891, + 0.87 + ], + "angle": 0, + "content": "Given the final predictions of the semantic segmentation \\(\\hat{S}\\), center offsets \\(\\hat{O}\\), and confidence scores \\(\\hat{C}\\), this section introduces how to exploit the detected centers to acquire the panoptic segmentation results during inference, and illustrates the key module, center dedduplication module (CDM)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.871, + 0.891, + 0.901 + ], + "angle": 0, + "content": "Post-process. For the panoptic segmentation results, instance segmentation is first generated and then the final" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13429" + } + ], + [ + { + "type": "code_caption", + "bbox": [ + 0.088, + 0.095, + 0.377, + 0.111 + ], + "angle": 0, + "content": "Algorithm 1: Center dedduplication module" + }, + { + "type": "algorithm", + "bbox": [ + 0.082, + 0.113, + 0.47, + 0.429 + ], + "angle": 0, + "content": "Input: Shifted things points \\(\\{\\tilde{p}_t^{shift}\\in \\mathbb{R}^3\\}_{t = 0}^{M - 1}\\) confidence scores \\(\\{\\tilde{c}_t\\in [0,1]\\}_{t = 0}^{M - 1}\\) distance threshold \\(d\\in \\mathbb{R}\\) Output: Instance centers \\(\\{\\tilde{e}_j\\in \\mathbb{R}^3\\}_{j = 0}^{D - 1}\\) \n1 order \\(=\\) sort_index( \\(\\{\\tilde{c}_t\\}_{t = 0}^{M - 1}\\) , decending=true) \n2 initialize a boolean vector \\(rej\\in \\mathbb{B}^{M}\\) with all elements of false and an empty list keep \n3 for \\(k\\gets 0\\) to \\(M - 1\\) do \n4 if not rej[order[k]] then \n5 keep.append(order[k]) \n6 for \\(h\\gets k + 1\\) to \\(M - 1\\) do \n7 dist \\(= ||\\tilde{p}_{order[k]}^{shift} - \\tilde{p}_{order[h]}^{shift}||_2\\) \n8 if dist \\(< d\\) then \n9 | rej[order[h]] \\(=\\) true \n10 end \n11 end \n12 end \n13 end \n14 \\(\\{\\tilde{e}_j\\} = \\tilde{p}_{keep}^{shift}\\)" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.46, + 0.47, + 0.505 + ], + "angle": 0, + "content": "panoptic segmentation is acquired by fusing the semantic and instance segmentation labels. There are five steps to achieve the final panoptic segmentation:" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.504, + 0.47, + 0.567 + ], + "angle": 0, + "content": "1) The things points \\(\\{\\tilde{p}_t\\}_{t=0}^{M-1} = \\{\\pmb{p}_i | \\hat{s}_i \\in\\) things classes} are selected according to the predicted semantic labels, with their offsets \\(\\{\\tilde{o}_t\\}_{t=0}^{M-1}\\) and confidence scores \\(\\{\\tilde{c}_t\\}_{t=0}^{M-1}\\) (where \\(M\\) is the number of things points)." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.566, + 0.469, + 0.597 + ], + "angle": 0, + "content": "2) Each shifted things point \\(\\tilde{\\pmb{p}}_t^{shift}\\) is given by \\(\\tilde{\\pmb{p}}_t^{shift} = \\tilde{\\pmb{p}}_t + \\tilde{\\pmb{o}}_t\\) and is severing as an instance center candidate." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.597, + 0.469, + 0.644 + ], + "angle": 0, + "content": "3) A center \\(\\tilde{\\pmb{e}}_j\\) for each instance \\(j\\) is selected by the CDM based on the coordinates \\(\\tilde{\\pmb{p}}_t^{shift}\\) and its confidence score \\(\\tilde{c}_t\\), while other candidates are suppressed." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.645, + 0.469, + 0.69 + ], + "angle": 0, + "content": "4) The instance \\(\\mathrm{ID}\\widetilde{\\text{in}}\\widetilde{\\text{st}}_t\\) is acquired by assigning the shifted things points \\(\\tilde{\\pmb{p}}_t^{\\text{shift}}\\) to the closest center in all centers \\(\\{\\tilde{\\pmb{e}}_j\\}_{j=0}^{D-1}\\) (\\(D\\) is the number of detected instances)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.69, + 0.469, + 0.749 + ], + "angle": 0, + "content": "5) The majority voting [15, 39] reassigns the most frequent semantic label to all points of a predicted instance to further guarantee the consistency of semantic labels within a predicted instance." + }, + { + "type": "list", + "bbox": [ + 0.076, + 0.504, + 0.47, + 0.749 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.471, + 0.901 + ], + "angle": 0, + "content": "Center dedduplication module (CDM). The CDM takes the shifted points \\(\\{\\tilde{p}_t^{shift}\\}_{t=0}^{M-1}\\) and confidence scores \\(\\{\\tilde{c}_t\\}_{t=0}^{M-1}\\) as input to get an instance center \\(\\hat{e}_j\\) for each instance \\(j\\). Inspired by the bounding box NMS, our CDM suppresses the candidate centers with lower scores within a euclidean distance threshold \\(d\\). The pseudo-code of our CDM is shown in Algorithm 1, where two centers with a distance less than \\(d\\) are considered as the same instance. The process of our CDM is simple and can be easily implemented in CUDA." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.09, + 0.634, + 0.108 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.115, + 0.894, + 0.161 + ], + "angle": 0, + "content": "The proposed CFNet is compared with the existing methods on the public SemanticKITTI [2] and nuScenes [5] panoptic segmentation benchmarks." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.169, + 0.691, + 0.185 + ], + "angle": 0, + "content": "4.1. Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.191, + 0.892, + 0.342 + ], + "angle": 0, + "content": "Datasets. The SemanticKITTI [2] consists of 22 sequences collected by a Velodyne HDL-64E \\(360^{\\circ}\\) rotating LiDAR with 64 beams vertically. It contains 43,552 LiDAR scans in total and is split into a training set with 19,130 scans from sequences 00 to 10 except 08, a validation set containing sequence 08 with 4,071 scans, and a test set including the rest sequences with 20,351 scans. The test set is only provided with point clouds and used for the online leaderboards. For the panoptic segmentation, it has 19 valid categories, including 11 stuff and 8 things categories." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.342, + 0.892, + 0.448 + ], + "angle": 0, + "content": "The nuScenes [5] is a newly released benchmark with 1,000 scenes collected by a Velodyne HDL-32E \\(360^{\\circ}\\) rotating LiDAR with 32 beams vertically. The dataset is collected in Boston and Singapore. It uses 28,130 scans for training, 6,019 for validation, and 6,008 for testing. The panoptic segmentation benchmark provides 16 categories, including 6 stuff and 10 things categories [12]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.449, + 0.892, + 0.644 + ], + "angle": 0, + "content": "Evaluation metric. As the official benchmarks [3, 12] suggest, the panoptic quality (PQ) is adopted to evaluate the performance of LiDAR panoptic segmentation. The PQ [3, 12] measures the overall quality of the panoptic segmentation, and can be decomposed into two terms: 1) the segmentation quality (SQ) is the average IoU of all matched pairs; 2) the recognition quality (RQ) measures the \\(F_{1}\\) score. These three metrics are also reported separately on the stuff and things classes, including \\(\\mathrm{PQ}^{St}\\), \\(\\mathrm{SQ}^{St}\\), \\(\\mathrm{RQ}^{St}\\), and \\(\\mathrm{PQ}^{Th}\\), \\(\\mathrm{SQ}^{Th}\\), \\(\\mathrm{RQ}^{Th}\\), respectively. The \\(\\mathrm{PQ}^{\\dagger}\\) [26] is also reported by replacing the PQ of each stuff class with its IoU. For evaluating the sub-task of semantic segmentation, mean Intersection over Union (mIoU) [2] is reported." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.645, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Training details. In the experiments, two 2D projection-based backbones, namely PolarNet [38] and CPGNet [23], are used in our CFNet. For the backbone of PolarNet [38], the training schedules and data augmentation are the same as those of Panoptic-PolarNet [39] for a fair comparison. Besides, for efficiency, only one stage version of CPGNet [23] is used in our CFNet. It is trained from scratch for 48 epochs with a batch size of 8 and takes around 24 hours on 8 NVIDIA RTX 3090 GPUs. Stochastic gradient descent (SGD) is used as the optimizer, where the initial learning rate, weight decay, and momentum are 0.02, 0.001, and 0.9, respectively. The learning rate is decayed by 0.1 per 10 epochs. In training, the data augmentation is applied as a convention, including random flipping along the \\(x\\) and \\(y\\) axes, random global scale sampled from \\([0.95, 1.05]\\), random rotation around the \\(z\\)-axis, and random Gaussian noise \\(\\mathcal{N}(0, 0.02)\\)." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "13430" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.123, + 0.088, + 0.849, + 0.26 + ], + "angle": 0, + "content": "
BackboneCFFEDeduplicationPQPQ†PQTthPQS\\( ^t \\)mIoURT(ms)
CFGFEM
a b c dPolarNet [38]✓BEV Center Heatmap [39]59.164.165.754.364.565.3 + 20.9
BEV Center Heatmap [39]59.564.366.254.464.669.7 + 20.9
✓✓BEV Center Heatmap [39]60.465.667.255.165.271.9 + 20.9
✓✓CDM [Ours]60.665.767.855.265.471.9 + 3.6
e f g h i jCPGNet [23]✓DBSCAN [11]59.564.163.856.364.631.6+24.9
HDBSCAN [6]58.362.960.956.364.931.6+48.7
MeanShift [9]59.964.564.756.364.931.6+84.6
CDM [Ours]60.565.666.156.566.332.7+2.3
✓CDM [Ours]60.865.966.356.966.537.4+2.3
✓✓CDM [Ours]62.767.570.057.367.441.2+2.3
" + }, + { + "type": "table_caption", + "bbox": [ + 0.243, + 0.27, + 0.726, + 0.285 + ], + "angle": 0, + "content": "Table 1. Ablation studies on the SemanticKITTI validation set. RT: running time." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.311, + 0.241, + 0.326 + ], + "angle": 0, + "content": "4.2. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.336, + 0.47, + 0.472 + ], + "angle": 0, + "content": "Ablative analyses are conducted on the SemanticKITTI validation set with the above experimental setup. As shown in Table 1, two 2D projection-based backbones, namely PolarNet [38] and CPGNet [23], are incorporated with our framework to validate the proposed CFFE and CDM. Row (a) shows the original Panoptic-Polarnet [39], while rows (b,c,d) are implemented based on the official code of Panoptic-Polarnet [39]. The running time is reported by the time of the model inference plus that of the post-process." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.475, + 0.47, + 0.642 + ], + "angle": 0, + "content": "Effects of the CFFE. The CFFE is configured in three ways: 1) the CFFE is not applied (a,e,f,g,h); 2) the center feature generation (CFG) in the CFFE is removed (b,i), which is equal to the setting of only imposing the intermediate supervision; 3) the whole CFFE is applied (c,d,j). For the two backbones, a common observation is that the improvements primarily come from the CFG rather than the intermediate supervision, especially for the things classes (a,b,c and h,i,j). In terms of the running time, The CFFE brings acceptable latency of \\(+6.6\\mathrm{ms}\\) and \\(+8.5\\mathrm{ms}\\) for these two backbones, respectively." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.643, + 0.47, + 0.749 + ], + "angle": 0, + "content": "Moreover, the errors in predicting things center offsets are calculated for both the CFNet with CFFE and its intermediate results as shown in Table 2. Equipped with the backbone of CPGNet [23], the proposed CFFE reduces the errors of predicting the things center offsets. The observations demonstrate that explicitly modeling center features helps in identifying an instance." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.751, + 0.468, + 0.767 + ], + "angle": 0, + "content": "Effects of the CDM. For the proposal-free meth" + }, + { + "type": "table", + "bbox": [ + 0.141, + 0.793, + 0.407, + 0.844 + ], + "angle": 0, + "content": "
training setvalidation set
Intermediate1.351.22
CFNet1.191.13
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.855, + 0.468, + 0.897 + ], + "angle": 0, + "content": "Table 2. Mean errors measured in meter (m) of things center offsets for the intermediate results and the CFNet with the CFFE on the SemanticKITTI training and validation set." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.311, + 0.892, + 0.493 + ], + "angle": 0, + "content": "ods, there are primarily three ways to identify each instance: 1) the class-agnostic clustering methods (e.g. DBSCAN, HDBSCAN, MeanShift) (e,f,g), 2) the BEV center heatmap [39] (a,b,c), and 3) our CDM (d,h,i,j). For the backbone of PolarNet [38], our CDM outperforms the BEV center heatmap [39] by \\(+0.2\\) for the PQ (c,d), while our CDM runs faster. For the backbone of CPGNet [23], compared with the class-agnostic clustering methods, our CDM also achieves the best performance on both accuracy \\((+0.6\\sim +2.2\\) for the PQ) and running time \\((-21.5\\mathrm{ms}\\sim -\\) \\(81.2\\mathrm{ms})\\) , though it slows down the model inference by \\(+1.1~\\mathrm{ms}\\) \\(\\mathrm{(g,h)}\\) owing to the additional confidence prediction." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.493, + 0.892, + 0.522 + ], + "angle": 0, + "content": "More ablation studies can be seen in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.531, + 0.727, + 0.547 + ], + "angle": 0, + "content": "4.3. Evaluation Comparisons" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.554, + 0.892, + 0.75 + ], + "angle": 0, + "content": "Quantitative results. In Table 3 and Table 4, our CFNet with CPGNet as its backbone is compared with the existing methods on the SemanticKITTI test set and nuScenes validation set, respectively. From top to bottom, the methods are grouped as proposal-based and proposal-free methods. It is found that our proposal-free CFNet surpasses the previous methods by a large margin on most metrics. Specifically, on the SemanticKITTI test set, our CFNet outperforms the previous state-of-the-art SCAN [35] and Panoptic-PHNet [22] by \\(+1.9\\) for the PQ, and \\(+1.5\\) for the RQ. On the nuScenes validation set, our CFNet outperforms the most competitive Panoptic-PHNet [22] by \\(+0.4\\) for the PQ, \\(+0.6\\) for the SQ, and \\(+0.4\\) for the RQ, respectively." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.892, + 0.886 + ], + "angle": 0, + "content": "Running time. The running time is reported on the SemanticKITTI. It is measured with PyTorch FP32 without any model acceleration tricks on a single NVIDIA RTX 3090 GPU. For clarity, the running time of our method is further divided into two parts, including the time costs of the model and post-process. Based on the same kind of devices as the existing methods, our CFNet runs \\(43.5\\mathrm{ms}\\) (41.2 ms for model inference and 2.3 ms for post-process) and is 1.6 times faster than the most efficient Panoptic-PHNet [22]." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.886, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Qualitative results. The visualization results are" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "13431" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.124, + 0.088, + 0.851, + 0.316 + ], + "angle": 0, + "content": "
MethodsPQPQ†SQRQPQThSQThRQThPQStSQStRQStmIoURT(ms)
Proposal-based Methods
RangeNet++ [25]&PointPillars [19]37.145.975.947.020.275.225.249.376.562.852.4416.6
KPConv [33]&PointPillars [19]44.552.580.054.432.781.538.753.179.065.958.8526.3
PanopticTrackNet [17]43.150.778.853.928.680.435.553.677.767.352.6147
EfficientLPS [31]57.463.283.068.753.187.860.560.579.574.661.4-
Proposal-free Methods
LPSAD [24]38.047.076.548.225.676.831.847.176.260.150.984.7
DS-Net [15]55.962.582.366.755.187.262.856.578.769.561.6294.1
Panoster [13]52.759.980.764.149.483.358.555.178.868.259.9-
GP-S3Net [29]60.069.082.072.165.086.674.556.478.770.470.8270.3
SMAC-Seg [20]56.162.582.067.953.085.661.858.479.372.363.399
Panoptic-PolarNet [39]54.160.781.465.053.387.260.654.877.268.159.586.2
SCAN [35]61.567.584.572.161.488.169.361.581.874.167.778.1
Panoptic-PHNet [22]61.567.984.872.163.890.770.459.980.573.366.069.3
CFNet [Ours] w/CPGNet [23]63.469.785.273.666.789.874.361.081.973.168.341.2 + 2.3
" + }, + { + "type": "table_caption", + "bbox": [ + 0.307, + 0.326, + 0.663, + 0.34 + ], + "angle": 0, + "content": "Table 3. Comparison results on the SemanticKITTI test set." + }, + { + "type": "table", + "bbox": [ + 0.123, + 0.353, + 0.849, + 0.558 + ], + "angle": 0, + "content": "
MethodsPQ\\( PQ^† \\)SQRQ\\( PQ^{Th} \\)\\( SQ^{Th} \\)\\( RQ^{Th} \\)\\( PQ^{St} \\)\\( SQ^{St} \\)\\( RQ^{St} \\)mIoU
Proposal-based Methods
Cylinder3D [40]&PointPillars [19]36.044.583.343.023.383.727.057.282.769.652.3
Cylinder3D [40]&SECOND [36]40.148.484.247.329.084.433.658.583.770.158.5
PanopticTrackNet [17]51.456.280.263.345.881.455.960.478.375.558.0
EfficientLPS [31]62.065.683.473.956.883.268.070.683.883.665.6
Proposal-free Methods
LPSAD [24]50.457.779.462.443.280.253.257.578.571.762.5
DS-Net [15]42.551.083.650.332.583.138.359.284.470.370.7
GP-S3Net [29]61.067.584.172.056.085.365.266.082.978.775.8
SMAC-Seg [20]67.071.885.078.265.287.174.268.882.982.272.2
Panoptic-PolarNet [39]63.467.283.975.359.284.170.370.483.683.566.9
SCAN [35]65.168.985.775.360.685.770.272.585.783.877.4
Panoptic-PHNet [22]74.777.788.284.274.089.082.575.986.886.979.7
CFNet [Ours] w/CPGNet [23]75.178.088.884.674.889.882.976.687.187.379.3
" + }, + { + "type": "table_caption", + "bbox": [ + 0.307, + 0.568, + 0.663, + 0.582 + ], + "angle": 0, + "content": "Table 4. Comparison results on the nuScenes validation set." + }, + { + "type": "image", + "bbox": [ + 0.089, + 0.607, + 0.275, + 0.753 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.115, + 0.754, + 0.249, + 0.766 + ], + "angle": 0, + "content": "Semantic Segmentation" + }, + { + "type": "image", + "bbox": [ + 0.276, + 0.607, + 0.458, + 0.679 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.276, + 0.68, + 0.458, + 0.752 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.301, + 0.754, + 0.431, + 0.766 + ], + "angle": 0, + "content": "Instance Segmentation" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.777, + 0.47, + 0.806 + ], + "angle": 0, + "content": "Figure 5. Visualizations of our CFNet on the SemanticKITTI test set. Different colors represent different classes or instances." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.471, + 0.903 + ], + "angle": 0, + "content": "shown in Fig. 5. Our CFNet can distinguish adjacent pedestrians or cars. Besides, the boundaries of instances can be accurately segmented. More qualitative results can be seen in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.608, + 0.62, + 0.623 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.633, + 0.894, + 0.8 + ], + "angle": 0, + "content": "A novel proposal-free CFNet is proposed for real-time LiDAR panoptic segmentation. For better modeling and exploiting the non-existent instance centers, a novel CFFE is proposed to generate enhanced center-focusing feature maps and a CDM is introduced to keep only one center for each instance and then assign the shifted things points to the closest center for acquiring instance IDs. From the experiments, it can be seen that center modeling and exploiting is a key problem in the proposal-free LiDAR panoptic segmentation methods, and mimicking the non-existent center features is promising and shows a clear benefit." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.815, + 0.669, + 0.832 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.893, + 0.901 + ], + "angle": 0, + "content": "The research project is partially supported by the National Key R&D Program of China (No.2021ZD0111902), and the National Natural Science Foundation of China (No.U21B2038, 62272015, U19B2039)." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "13432" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.473, + 0.184 + ], + "angle": 0, + "content": "[1] Yara Ali Alnaggar, Mohamed Afifi, Karim Amer, and Mohamed ElHelw. Multi projection fusion for real-time semantic segmentation of 3d lidar point clouds. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1800-1809, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.186, + 0.472, + 0.268 + ], + "angle": 0, + "content": "[2] Jens Behley, Martin Garbade, Andres Milioto, Jan Quenzel, Sven Behnke, Cyril Stachniss, and Jurgen Gall. Semantickitti: A dataset for semantic scene understanding of lidar sequences. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9297-9307, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.27, + 0.472, + 0.325 + ], + "angle": 0, + "content": "[3] Jens Behley, Andres Milioto, and Cyril Stachniss. A benchmark for lidar-based panoptic segmentation based on kitti. In 2021 IEEE International Conference on Robotics and Automation (ICRA), pages 13596-13603. IEEE, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.326, + 0.472, + 0.408 + ], + "angle": 0, + "content": "[4] Maxim Berman, Amal Rannen Triki, and Matthew B Blaschko. The lovasz-softmax loss: A tractable surrogate for the optimization of the intersection-over-union measure in neural networks. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 4413-4421, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.41, + 0.472, + 0.493 + ], + "angle": 0, + "content": "[5] Holger Caesar, Varun Bankiti, Alex H Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuscenes: A multimodal dataset for autonomous driving. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11621-11631, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.494, + 0.472, + 0.55 + ], + "angle": 0, + "content": "[6] Ricardo JGB Campello, Davoud Moulavi, and Jörg Sander. Density-based clustering based on hierarchical density estimates. In Pacific-Asia conference on knowledge discovery and data mining, pages 160–172. Springer, 2013. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.55, + 0.472, + 0.632 + ], + "angle": 0, + "content": "[7] Ran Cheng, Ryan Razani, Ehsan Taghavi, Enxu Li, and Bingbing Liu. 2-s3net: Attentive feature fusion with adaptive feature selection for sparse semantic segmentation network. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12547-12556, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.634, + 0.472, + 0.702 + ], + "angle": 0, + "content": "[8] Christopher Choy, JunYoung Gwak, and Silvio Savarese. 4d spatio-temporal convnets: Minkowski convolutional neural networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3075-3084, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.704, + 0.472, + 0.759 + ], + "angle": 0, + "content": "[9] Dorin Comaniciu and Peter Meer. Mean shift: A robust approach toward feature space analysis. IEEE Transactions on pattern analysis and machine intelligence, 24(5):603-619, 2002. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.761, + 0.472, + 0.816 + ], + "angle": 0, + "content": "[10] Tiago Cortinhal, George Tzelepis, and Eren Erdal Aksoy. Salsanext: Fast, uncertainty-aware semantic segmentation of lidar point clouds. In International Symposium on Visual Computing, pages 207-222. Springer, 2020. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.817, + 0.472, + 0.871 + ], + "angle": 0, + "content": "[11] Martin Ester, Hans-Peter Kriegel, Jörg Sander, Xiaowei Xu, et al. A density-based algorithm for discovering clusters in large spatial databases with noise. In kdd, volume 96, pages 226-231, 1996. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.472, + 0.902 + ], + "angle": 0, + "content": "[12] Whye Kit Fong, Rohit Mohan, Juana Valeria Hurtado, Lubing Zhou, Holger Caesar, Oscar Beijbom, and Abhinav Val-" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.473, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.134 + ], + "angle": 0, + "content": "ada. Panoptic nuscenes: A large-scale benchmark for lidar panoptic segmentation and tracking. IEEE Robotics and Automation Letters, 7(2):3795-3802, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.136, + 0.894, + 0.204 + ], + "angle": 0, + "content": "[13] Stefano Gasperini, Mohammad-Ali Nikouei Mahani, Alvaro Marcos-Ramiro, Nassir Navab, and Federico Tombari. Panoster: End-to-end panoptic segmentation of lidar point clouds. IEEE Robotics and Automation Letters, 6(2):3216-3223, 2021. 1, 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.206, + 0.894, + 0.262 + ], + "angle": 0, + "content": "[14] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In Proceedings of the IEEE international conference on computer vision, pages 2961-2969, 2017. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.264, + 0.892, + 0.332 + ], + "angle": 0, + "content": "[15] Fangzhou Hong, Hui Zhou, Xinge Zhu, Hongsheng Li, and Ziwei Liu. Lidar-based panoptic segmentation via dynamic shifting network. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13090-13099, 2021. 1, 3, 5, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.334, + 0.892, + 0.417 + ], + "angle": 0, + "content": "[16] Qingyong Hu, Bo Yang, Linhai Xie, Stefano Rosa, Yulan Guo, Zhihua Wang, Niki Trigoni, and Andrew Markham. Randla-net: Efficient semantic segmentation of large-scale point clouds. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11108-11117, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.419, + 0.892, + 0.461 + ], + "angle": 0, + "content": "[17] Juana Valeria Hurtado, Rohit Mohan, Wolfram Burgard, and Abhinav Valada. Mopt: Multi-object panoptic tracking. arXiv preprint arXiv:2004.08189, 2020. 1, 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.462, + 0.892, + 0.518 + ], + "angle": 0, + "content": "[18] Alexander Kirillov, Kaiming He, Ross Girshick, Carsten Rother, and Piotr Dollar. Panoptic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9404-9413, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.519, + 0.892, + 0.589 + ], + "angle": 0, + "content": "[19] Alex H Lang, Sourabh Vora, Holger Caesar, Lubing Zhou, Jiong Yang, and Oscar Beijbom. Pointpillars: Fast encoders for object detection from point clouds. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12697-12705, 2019. 1, 2, 4, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.59, + 0.892, + 0.645 + ], + "angle": 0, + "content": "[20] Enxu Li, Ryan Razani, Yixuan Xu, and Liu Bingbing. Smac-seg: Lidar panoptic segmentation via sparse multi-directional attention clustering. arXiv preprint arXiv:2108.13588, 2021. 2, 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.647, + 0.892, + 0.689 + ], + "angle": 0, + "content": "[21] Enxu Li, Ryan Razani, Yixuan Xu, and Bingbing Liu. Cpseg: Cluster-free panoptic segmentation of 3d lidar point clouds. arXiv preprint arXiv:2111.01723, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.69, + 0.892, + 0.772 + ], + "angle": 0, + "content": "[22] Jinke Li, Xiao He, Yang Wen, Yuan Gao, Xiaogiang Cheng, and Dan Zhang. Panoptic-phnet: Towards real-time and high-precision lidar panoptic segmentation via clustering pseudo heatmap. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11809–11818, 2022. 1, 3, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.774, + 0.892, + 0.83 + ], + "angle": 0, + "content": "[23] Xiaoyan Li, Gang Zhang, Hongyu Pan, and Zhenhua Wang. Cpgnet: Cascade point-grid fusion network for real-time lidar semantic segmentation. arXiv preprint arXiv:2204.09914, 2022. 2, 4, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.832, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[24] Andres Milioto, Jens Behley, Chris McCool, and Cyril Stachniss. Lidar panoptic segmentation for autonomous driving. In 2020 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 8505-8512. IEEE, 2020. 2, 8" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.894, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13433" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.161 + ], + "angle": 0, + "content": "[25] Andres Milioto, Ignacio Vizzo, Jens Behley, and Cyril Stachniss. Rangenet++: Fast and accurate lidar semantic segmentation. In 2019 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 4213-4220. IEEE, 2019. 2, 4, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.163, + 0.47, + 0.22 + ], + "angle": 0, + "content": "[26] Lorenzo Porzi, Samuel Rota Bulo, Aleksander Colovic, and Peter Kontschieder. Seamless scene segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8277-8286, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.221, + 0.47, + 0.288 + ], + "angle": 0, + "content": "[27] Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. Pointnet: Deep learning on point sets for 3d classification and segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 652-660, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.291, + 0.47, + 0.332 + ], + "angle": 0, + "content": "[28] Charles R Qi, Li Yi, Hao Su, and Leonidas J Guibas. Point-net++: Deep hierarchical feature learning on point sets in a metric space. arXiv preprint arXiv:1706.02413, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.334, + 0.47, + 0.404 + ], + "angle": 0, + "content": "[29] Ryan Razani, Ran Cheng, Enxu Li, Ehsan Taghavi, Yuan Ren, and Liu Bingbing. Gp-s3net: Graph-based panoptic sparse semantic segmentation network. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 16076-16085, 2021. 1, 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.405, + 0.47, + 0.475 + ], + "angle": 0, + "content": "[30] Shaoshuai Shi, Chaoxu Guo, Li Jiang, Zhe Wang, Jianping Shi, Xiaogang Wang, and Hongsheng Li. Pv-rcnn: Pointvoxel feature set abstraction for 3d object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10529–10538, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.476, + 0.47, + 0.531 + ], + "angle": 0, + "content": "[31] Kshitij Sirohi, Rohit Mohan, Daniel Buscher, Wolfram Burgard, and Abhinav Valada. Efficienttps: Efficient lidar panoptic segmentation. IEEE Transactions on Robotics, 2021. 1, 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.533, + 0.47, + 0.601 + ], + "angle": 0, + "content": "[32] Haotian Tang, Zhijian Liu, Shengyu Zhao, Yujun Lin, Ji Lin, Hanrui Wang, and Song Han. Searching efficient 3d architectures with sparse point-voxel convolution. In European Conference on Computer Vision, pages 685-702. Springer, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.603, + 0.47, + 0.685 + ], + "angle": 0, + "content": "[33] Hugues Thomas, Charles R Qi, Jean-Emmanuel Deschaud, Beatrix Marcotegui, François Goulette, and Leonidas J Guibas. Kpconv: Flexible and deformable convolution for point clouds. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6411–6420, 2019. 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.688, + 0.47, + 0.758 + ], + "angle": 0, + "content": "[34] Chenfeng Xu, Bichen Wu, Zining Wang, Wei Zhan, Peter Vajda, Kurt Keutzer, and Masayoshi Tomizuka. Squeeze-segv3: Spatially-adaptive convolution for efficient pointcloud segmentation. In European Conference on Computer Vision, pages 1-19. Springer, 2020. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.759, + 0.47, + 0.814 + ], + "angle": 0, + "content": "[35] Shuangjie Xu, Rui Wan, Maosheng Ye, Xiaoyi Zou, and Tongyi Cao. Sparse cross-scale attention network for efficient lidar panoptic segmentation. arXiv preprint arXiv:2201.05972, 2022. 1, 2, 3, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.816, + 0.47, + 0.843 + ], + "angle": 0, + "content": "[36] Yan Yan, Yuxing Mao, and Bo Li. Second: Sparsely embedded convolutional detection. Sensors, 18(10):3337, 2018. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.845, + 0.47, + 0.901 + ], + "angle": 0, + "content": "[37] Dongqiangzi Ye, Zixiang Zhou, Weijia Chen, Yufei Xie, Yu Wang, Panqu Wang, and Hassan Foroosh. Lidarmultinet: Towards a unified multi-task network for lidar perception. arXiv preprint arXiv:2209.09385, 2022. 1, 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.092, + 0.893, + 0.175 + ], + "angle": 0, + "content": "[38] Yang Zhang, Zixiang Zhou, Philip David, Xiangyu Yue, Zerong Xi, Boqing Gong, and Hassan Foroosh. Polarnet: An improved grid representation for online lidar point clouds semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9601-9610, 2020. 2, 4, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.178, + 0.893, + 0.245 + ], + "angle": 0, + "content": "[39] Zixiang Zhou, Yang Zhang, and Hassan Foroosh. Panoptic-polarnet: Proposal-free lidar point cloud panoptic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13194-13203, 2021. 1, 2, 3, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.248, + 0.893, + 0.33 + ], + "angle": 0, + "content": "[40] Xinge Zhu, Hui Zhou, Tai Wang, Fangzhou Hong, Yuexin Ma, Wei Li, Hongsheng Li, and Dahua Lin. Cylindrical and asymmetrical 3d convolution networks for lidar segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9939-9948, 2021. 2, 8" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.092, + 0.893, + 0.33 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "13434" + } + ] +] \ No newline at end of file diff --git a/2023/Center Focusing Network for Real-Time LiDAR Panoptic Segmentation/e749a0dd-41b1-4b7d-9325-ea72d56e45cb_origin.pdf b/2023/Center Focusing Network for Real-Time LiDAR Panoptic Segmentation/e749a0dd-41b1-4b7d-9325-ea72d56e45cb_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..19e0f3a74015aa2576e6deea2ef73551193b331b --- /dev/null +++ b/2023/Center Focusing Network for Real-Time LiDAR Panoptic Segmentation/e749a0dd-41b1-4b7d-9325-ea72d56e45cb_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:309e518bdf72e9c9759de5b5e54697e756f920c92216b3181e5414ea8a852bb5 +size 2964757 diff --git a/2023/Center Focusing Network for Real-Time LiDAR Panoptic Segmentation/full.md b/2023/Center Focusing Network for Real-Time LiDAR Panoptic Segmentation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..7e645e678d7603afb2202376bf7c18450c23e0eb --- /dev/null +++ b/2023/Center Focusing Network for Real-Time LiDAR Panoptic Segmentation/full.md @@ -0,0 +1,347 @@ +# Center Focusing Network for Real-Time LiDAR Panoptic Segmentation + +Xiaoyan Li $^{1,2}$ Gang Zhang $^{3}$ Boyue Wang $^{1,2}$ Yongli Hu $^{1,2}$ Baocai Yin $^{1,2}$ + +1Beijing Municipal Key Lab of Multimedia and Intelligent Software Technology + +$^{2}$ Beijing Institute of Artificial Intelligence, Faculty of Information Technology, Beijing University of Technology, Beijing 100124, China + +3Mogo Auto Intelligence and Telemetics Information Technology Co. Ltd. + +{xiaoyan.li, wby, huyongli, ybc}@bjut.edu.cn zhanggang11021136@gmail.com + +# Abstract + +LiDAR panoptic segmentation facilitates an autonomous vehicle to comprehensively understand the surrounding objects and scenes and is required to run in real time. The recent proposal-free methods accelerate the algorithm, but their effectiveness and efficiency are still limited owing to the difficulty of modeling non-existent instance centers and the costly center-based clustering modules. To achieve accurate and real-time LiDAR panoptic segmentation, a novel center focusing network (CFNet) is introduced. Specifically, the center focusing feature encoding (CFFE) is proposed to explicitly understand the relationships between the original LiDAR points and virtual instance centers by shifting the LiDAR points and filling in the center points. Moreover, to leverage the redundantly detected centers, a fast center dedduplication module (CDM) is proposed to select only one center for each instance. Experiments on the SemanticKITTI and nuScenes panoptic segmentation benchmarks demonstrate that our CFNet outperforms all existing methods by a large margin and is 1.6 times faster than the most efficient method. + +# 1. Introduction + +Panoptic segmentation [18] combines both semantic segmentation and instance segmentation in a single framework. It predicts semantic labels for the uncountable stuff classes (e.g. road, sidewalk), while it simultaneously provides semantic labels and instance IDs for the countable things classes (e.g. car, pedestrian). The LiDAR panoptic segmentation is one of the bases for the safety of autonomous driving, which employs the point clouds collected by the Light Detection and Ranging (LiDAR) sensors to effectively depict the surroundings. Existing LiDAR panoptic segmentation methods first conduct semantic segmentation, and then achieve instance segmentation for the things categories in + +![](images/bdf504294a4b4d31dbec88f296864e5e0cb8826f00a6b19e1001db7767b8cd69.jpg) +Figure 1. PQ vs. runtime on the SemanticKITTI test set. Runtime measurements are taken on a single NVIDIA RTX 3090 GPU. The panoptic quality (PQ) is introduced in section 4.1. + +two ways, the proposal-based and proposal-free methods. + +The proposal-based methods [17, 31, 37] adopt a two-stage process similar to the well-known Mask R-CNN [14] in the image domain. It first generates object proposals for the things points by using 3D detection networks [19, 30] and then refines the instance segmentation results within each proposal. As shown in Fig. 1, these methods are usually complicated and hardly achieve real-time processing, owing to their sequential multi-stage pipelines. + +The proposal-free frameworks [13, 15, 21, 22, 29, 35, 39] are more compact. To associate the things points with instance IDs, these methods usually leverage the instance centers. Specifically, they regress the offsets from the points to their corresponding centers, and then adopt the class-agnostic center-based clustering modules [13, 15, 29] or the bird's-eye view (BEV) center heatmap [22, 35, 39]. However, two problems exist in these methods. First, for center feature extracting and center modeling, the non-existent in + +![](images/739d52840857770b48bf2f272685550328dd7f45992da9ede8d2413a38ad8818.jpg) +(a) CFNet without CFFE + +![](images/155e0ba77e4ce0beccb1409edc9db8247cd9a825ad52c636deb8ccd9172fd29d.jpg) +(b) CFNet with CFFE +Figure 2. Instance segmentation of a car. Without our CFFE, the car is split into parts (a), while the CFFE significantly alleviates this problem (b). Different colors represent different instances. + +stance centers increase the difficulty, considering that the LiDAR points are usually surface-aggregated [35] and an instance center is imaginary in most cases. As shown in Fig. 2(a), the difficulty often results in the fault that one instance is incorrectly split into several parts. Second, for exploiting the redundantly detected centers, the clustering modules (e.g. MeanShift, DBSCAN) are too time-consuming to support the real-time autonomous driving perception systems, while the BEV center heatmap cannot distinguish objects with different altitudes in the same BEV grid. + +For accurate and fast LiDAR panoptic segmentation, a proposal-free center focusing network (CFNet) is proposed. For better encoding center features, a novel center focusing feature encoding (CFFE) is proposed to generate center-focusing feature maps by shifting the things points to fill in the non-existent instance centers for more accurate predictions (as shown in Fig. 2(b)). For center modeling, the CFNet not only decomposes the panoptic segmentation task into the widely-used semantic segmentation and center offset regression, but also proposes a new confidence score prediction for indicating the accuracy of the center offset regression. Subsequently, for the detected centers exploiting, a novel center dedduplication module (CDM) is designed to select one center for a single instance. The CDM keeps the predicted centers with higher confidence scores, while suppressing the ones with lower confidence. Finally, instance segmentation is achieved by assigning the shifted things points to the closest center. For efficiency, the proposed CFNet is built on the 2D projection-based segmentation paradigm. Our contributions are as follows: + +- A proposal-free CFNet is proposed to achieve accurate and fast LiDAR panoptic segmentation by solving the bottleneck problems of center modeling and center-based clustering in previous methods. +- The CFFE is proposed to alleviate the difficulty of modeling the non-existent instance centers and the CDM is designed to efficiently keep one center for each instance. + +- The proposed CFNet is evaluated on the nuScenes and SemanticKITTI LiDAR panoptic segmentation benchmarks. Our CFNet achieves the state-of-the-art performance with a real-time inference speed. + +# 2. Related Work + +The LiDAR panoptic segmentation methods usually adopt the LiDAR semantic segmentation networks as the backbone and jointly optimize the semantic segmentation and instance segmentation tasks. The results of the two tasks are fused to generate the final panoptic segmentation results. The LiDAR semantic segmentation backbone is first introduced. Then, the panoptic segmentation methods, achieving instance segmentation upon the semantic segmentation backbone, are grouped into the proposal-based and proposal-free methods, and are further discussed. + +LiDAR semantic segmentation backbone. The backbone network for LiDAR semantic segmentation can be divided into three groups: point-based, voxel-based, and 2D projection-based methods. The point-based methods [16, 27, 28, 33] directly operate on the raw point clouds but are extremely time-consuming due to the expensive local neighbor searching. The voxel-based methods [7, 8, 32, 40] discretize the point clouds into structured voxels, where sparse 3D convolutions are applied. These methods are still difficult to meet the real-time applications, although they achieve the highest accuracy. The 2D projection-based methods are more efficient since most of their computation is done in the 2D space, such as range view (RV) [10,25,34], bird's-eye view (BEV) [19], polar view [38], and multiview fusion [1, 23]. For fast LiDAR panoptic segmentation, the Panoptic-PolarNet [39], SMAC-Seg [20], and LPSAD [24] all adopt the 2D projection-based backbone. A recent real-time semantic segmentation method, namely CPGNet [23], is a 2D projection-based one that explores an end-to-end multi-view fusion framework by fusing the features from the point, BEV, and RV. + +**Proposal-based methods.** The proposal-based methods conduct instance segmentation through a two-stage complicated process. They first detect the foreground instances and then refine the instance segmentation results independently within each detected bounding box. Based on the Mask R-CNN [14] for instance segmentation, the MOPT [17] and EfficientLPS [31] insert a semantic branch to achieve panoptic segmentation on the range view (RV). Recently, LidarMultiNet [37] unifies LiDAR-based 3D object detection, semantic segmentation, and panoptic segmentation in a single framework to reduce the computation cost by sharing a strong voxel-based backbone. + +Proposal-free methods. The proposal-free methods usually apply the class-agnostic clustering to the things points to conduct instance segmentation. The LPSAD [24] clusters points into instances by regressing the center off + +![](images/6b477c5cd5f7519ef9d8618ecd027fb88a9c3a0aa05dd6413acf48af6c7613cb.jpg) +Figure 3. The overview of our CFNet. It consists of four steps: 1) the 2D projection-based backbone extracts features on the 2D space; 2) the proposed center focusing feature encoding (CFFE) mimics and enhances the non-existent instance center features; 3) the panoptic segmentation head predicts the output results; 4) the proposed center dedduplication module (CDM) achieves instance segmentation that is fused to generate the final panoptic segmentation results. The dashed lines indicate that the operations are only used during inference. + +![](images/8493b7dd4cb04a7ae0b68c07950105b284750e2ac83566bf88e934fa0a7a78a5.jpg) +: Multi-Layer Perception +P2G: Point to Grid +FCN: Fully Convolutional Network +: Fully Connected Layer +G2P: Grid to Point +PF: Point Fusion + +![](images/c29088879b4f7a3e7e87ca8fb061349c93ad6c9c5e8abf4bd960e08545d9314d.jpg) +CFFE: Center Focusing Feature Encoding +CDM: Center Dedduplication Module + +sets. The Panoster [13] directly predicts the instance IDs from a learning-based clustering module, where the time-consuming DBSCAN [11] is used to refine the instance segmentation results. The DS-Net [15] proposes a learnable dynamic shifting module that adjusts the kernel functions of the MeanShift [9] to handle instances with various sizes. The GP-S3Net [29] constructs a graph convolutional network (GCN) on the over-segmentation clusters to identify instances. The SMAC-Seg [20] proposes a sparse multidirectional attention clustering and center-aware repel loss for instance segmentation. These methods adopt the time-consuming clustering methods. To further accelerate the algorithm, recently, Panoptic-PolarNet [39], SCAN [35], and Panoptic-PHNet [22] adopt the BEV center heatmap and center offsets for instance segmentation. However, the BEV center heatmap is also costly and confuses on $z$ -axis. + +# 3. Approach + +The input of the LiDAR panoptic segmentation task is the LiDAR point clouds, given by a set of unordered points $\pmb{P} = \{(p_i,f_i)\}_{i = 0}^{N - 1}$ (where $\pmb {p}_i = (x,y,z)$ is the 3D coor + +dinates in the cartesian space and $\pmb{f}_i$ denotes additional LiDAR point features, e.g. intensity). The task aims to assign a set of labels $\pmb{L} = \{(s_i, inst_i)\}_{i=0}^{N-1}$ to these points, where $s_i$ is the semantic label (e.g. road, building, car, pedestrian), and $inst_i$ is the instance ID of the $i^{th}$ point. Moreover, $s_i$ can be divided into the uncountable stuff classes (e.g. road, building) and countable things classes (e.g. car, pedestrian). The instance IDs of stuff points are set as 0. + +To predict the labels $L$ of the input LiDAR point clouds $P$ , our CFNet decomposes this process into four steps, as shown in Fig. 3: 1) an off-the-shelf 2D projection-based backbone is applied to efficiently extract features on the 2D spaces (e.g. BEV, RV); 2) a novel center focusing feature encoding (CFFE) is used to generate center-focusing feature maps for more accurate predictions of the instance centers; 3) the panoptic segmentation head fuses the features from the 3D points and 2D spaces to predict the semantic segmentation results, the center offsets, and the confidence scores of the center offsets, respectively; 4) during inference, a post-process is conducted to produce the panoptic segmentation results, where a novel center dedduplication module (CDM) operates on the shifted things points to se + +![](images/8bdd3a6609487f1e67f8756fb55ea40ea595a770e1b9ae4f7b544fb54786ca76.jpg) +Figure 4. The proposed center focusing feature encoding (CFFE). The "Conv" represents a 2D convolution with $3 \times 3$ kernels, a batch normalization, and a ReLU layer. The details of the semantic branch and instance branch are shown in Fig. 3. The blue arrows are coordinate-related operations. + +lect only one center for a single instance. + +The design of the backbone in the first step is beyond the scope of the paper, and the last three mentioned steps are illustrated in detail in the following subsections. + +# 3.1. Center Focusing Feature Encoding + +As mentioned above, the LiDAR points of an object are usually surface-aggregated, especially for car and truck categories, resulting in the fact that the center of an object is imaginary and does not exist in the LiDAR point clouds. To encode the features of the non-existent centers, a novel center focusing feature encoding (CFFE) is proposed, which takes the 2D features from the backbone and the 3D point coordinates as inputs and generates the enhanced center-focusing feature maps as shown in Fig. 3. + +The CFFE module consists of three steps, including intermediate result prediction, center feature generation, and feature enhancement module, as shown in Fig. 4. + +Intermediate result prediction. In this step, the CFFE predicts intermediate results (including the semantic segmentation, center offset, and its confidence scores) according to the 2D features $\pmb{F}_{m}$ and 3D point features $\pmb{F}_{p}$ for subsequent center feature simulation. Specifically, two convolution layers are applied on the 2D features $\pmb{F}_{m}$ independently, to generate semantic features $\pmb{F}_{m}^{sem}$ and instance features $\pmb{F}_{m}^{ins}$ ( $m$ is the specific 2D view, such as RV [10, 25, 34], BEV [19], and polar view [38].) + +$$ +\boldsymbol {F} _ {m} ^ {\text {s e m}} = \operatorname {C o n v} \left(\boldsymbol {F} _ {m}; \boldsymbol {\theta} _ {1}\right), \tag {1} +$$ + +$$ +\boldsymbol {F} _ {m} ^ {\text {i n s}} = \operatorname {C o n v} \left(\boldsymbol {F} _ {m}; \boldsymbol {\theta} _ {2}\right), \tag {2} +$$ + +where $Conv$ denotes sequential 2D convolution, batch normalization, and ReLU operations, and $\theta_{1}$ and $\theta_{2}$ are their learnable parameters. Then, the semantic branch generates per-point 3D semantic features $F_{p}^{sem}$ by fusing the point + +features $F_{p}$ and 2D semantic features $\pmb{F}_{m}^{sem}$ + +$$ +\boldsymbol {F} _ {p} ^ {\text {s e m}} = \operatorname {S e g} \left(\boldsymbol {F} _ {p}, \boldsymbol {F} _ {m} ^ {\text {s e m}}; \boldsymbol {\theta} _ {3}\right), \tag {3} +$$ + +where $Seg$ is the semantic branch, and $\theta_{3}$ is the parameters. + +Finally, the intermediate semantic results $\hat{S}^{\prime} = \{s_{i}^{\prime}\}_{i=0}^{N-1}$ is produced based on $F_{p}^{sem}$ . The intermediate results of center offsets $\hat{O}^{\prime} = \{o_{i}^{\prime}\}_{i=0}^{N-1}$ and confidence scores $\hat{C}^{\prime} = \{c_{i}^{\prime}\}_{i=0}^{N-1}$ are predicted by the instance branch with the point features $F_{p}$ and 2D instance features $F_{m}^{ins}$ , + +$$ +\hat {\boldsymbol {S}} ^ {\prime} = F C \left(\boldsymbol {F} _ {p} ^ {s e m}; \boldsymbol {\theta} _ {4}\right), \tag {4} +$$ + +$$ +\hat {\boldsymbol {O}} ^ {\prime}, \hat {\boldsymbol {C}} ^ {\prime} = I n s \left(\boldsymbol {F} _ {p}, \boldsymbol {F} _ {m} ^ {i n s}; \boldsymbol {\theta} _ {5}\right), \tag {5} +$$ + +where $Ins$ is the instance branch, and $FC$ represents the fully-connected layer. The structure and the training objective of semantic and instance branches are the same as those in the panoptic segmentation head and are elucidated in Fig. 3 and section 3.2. + +Center feature generation (CFG). In this step, the CFFE generates the shifted center features $\pmb{F}_{p\rightarrow m}^{sem}$ by shifting the 3D semantic point features $\pmb{F}_{p}^{sem}$ to the predicted centers according to the mentioned intermediate results. First, the coordinates of a predicted center $\pmb{p}_i^{shift}$ are computed by, + +$$ +\boldsymbol {p} _ {i} ^ {\text {s h i f t}} = \boldsymbol {p} _ {i} + \hat {\boldsymbol {o}} _ {i} ^ {\prime} * \mathbb {1} \left[ \hat {c} _ {i} ^ {\prime} > \delta \right], \tag {6} +$$ + +where $\pmb{p}_i$ is the original 3D LiDAR point coordinates, and $\mathbb{1}[\hat{c}_i^{\prime} > \delta ]$ is a binary indicator of whether the confidence $\hat{c}_i^\prime$ is greater than $\delta = 0.2$ . In other words, it does not shift the stuff points or the things points with low confidence scores. + +Then, the shifted 3D points $\{\pmb{p}_i^{shift}\}_{i=0}^{N-1}$ are served as the new coordinates of the feature points, and the Point to Grid (P2G) operation in [23] re-projects the 3D semantic + +features $\pmb{F}_{p}^{sem}$ onto the 2D projected feature maps $\pmb{F}_{p\rightarrow m}^{sem}$ with this new coordinates, + +$$ +\boldsymbol {F} _ {p \rightarrow m} ^ {s e m} = P 2 G (\boldsymbol {F} _ {p} ^ {s e m}; \{\boldsymbol {p} _ {i} ^ {\text {s h i f t}} \} _ {i = 0} ^ {N - 1}). \tag {7} +$$ + +Compared with $F_{p}^{sem}$ , the feature maps $F_{p\rightarrow m}^{sem}$ pay more attention to the imaginary centers since most of the things points have been shifted to their predicted centers. + +Feature enhancement module (FEM). The CFFE finally fuses the semantic feature maps $\pmb{F}_{m}^{sem}$ and the reprojected shifted center feature maps $\pmb{F}_{p\rightarrow m}^{sem}$ to generate center-focusing semantic feature maps $\pmb{F}_{m}^{CFsem}$ and instance feature maps $\pmb{F}_{m}^{CFS}$ , which are used by the subsequent semantic and instance branches for more accurate predictions. The enhancement module is composed of a simple concatenation operation and several convolution layers, whose details are in the supplementary material. + +In addition, for application of the multi-view fusion backbone (e.g. [23] shown in the supplementary material), the feature maps $\mathbf{F}_m$ , $\mathbf{F}_m^{sem}$ , $\mathbf{F}_m^{ins}$ , $\mathbf{F}_{p\rightarrow m}^{sem}$ (e.g. $m \in \{\mathrm{RV},\mathrm{BEV}\}$ ) of each view are computed independently according to the above process. Then, they are fused in the point fusion (PF) module [23] to generate integrated 3D point features and per-point prediction. + +# 3.2. Panoptic Segmentation Head + +For better modeling the instance center, the panoptic segmentation head uses a semantic branch to predict the semantic segmentation, and an instance branch to estimate both the center offsets and the newly introduced confidence scores, given the center-focusing semantic and instance feature maps. The architectures of both branches follow the segmentation head in [23] and are briefly introduced, while their training objectives are elaborated on. + +Semantic branch. For a per-point prediction, the semantic branch first applies the Grid to Point (G2P) operations to acquire the 3D point-wise features from the 2D semantic feature maps. Then, a PF module fuses the pointwise features from the G2P operation and the original 3D points to generate point-wise semantic features. The details of the G2P and PF can be found in [23]. After obtaining the point-wise semantic features, a fully connected (FC) layer is used to predict the final per-point semantic result $\hat{\psi}_i^j$ ( $\hat{\psi}_i^j \in [0,1]$ ). $\hat{\psi}_i^j$ represents the probability of the $i^{th}$ LiDAR point belonging to the $j^{th}$ class. The predicted semantic label $\hat{s}_i$ is acquired by selecting the most probable class $\hat{s}_i = \arg \max_j \hat{\psi}_i^j$ . + +Referring to the CPGNet [23], the same loss functions are adopted, including weighted cross entropy (WCE) loss $\mathcal{L}_{wce}$ , Lovász-Softmax loss $\mathcal{L}_{ls}$ [4], and transformation consistency loss $\mathcal{L}_{tc}$ . + +Instance branch. Similar to the semantic branch, the instance branch also adopts the G2P operation and a PF module to acquire point-wise instance features. An FC layer + +is applied to predict the per-point center offsets $\hat{o}_i$ . The ground-truth for $\hat{o}_i$ is the offset vector $e_i - p_i$ from the $i^{th}$ point $p_i$ to its corresponding instance center $e_i$ . + +For center offset regression, the loss function for optimizing $\hat{o}_i$ only considers the things classes and is formulated as follows, + +$$ +\mathcal {L} _ {\boldsymbol {o} _ {i}} = \left\{ \begin{array}{l l} | | \hat {\boldsymbol {o}} _ {i} - (\boldsymbol {e} _ {i} - \boldsymbol {p} _ {i}) | | _ {2} & \text {i f} s _ {i} \in t h i n g s \text {c l a s s e s}, \\ 0 & \text {o t h e r w i s e}, \end{array} \right. \tag {8} +$$ + +where $\pmb{e}_i$ is the axis-aligned center of the instance [15]. Then, the losses are summarized as the following, + +$$ +\mathcal {L} _ {o} = \frac {1}{N ^ {\text {t h i n g s}}} \sum_ {i = 0} ^ {N - 1} \mathcal {L} _ {\boldsymbol {o} _ {i}}, \tag {9} +$$ + +where $N$ and $N^{things}$ are the numbers of all points and things points, respectively. + +For confidence score regression, another FC layer is used to predict the per-point confidence score $\hat{c}_i$ to indicate the accuracy degree of $\hat{o}_i$ . $\hat{c}_i$ is activated by a sigmoid function to ensure $\hat{c}_i \in [0,1]$ . The ground-truth label $c_i$ for supervising $\hat{c}_i$ is generated by + +$$ +c _ {i} = \left\{ \begin{array}{l l} \exp \left(- \frac {\mathcal {L} _ {o _ {i}} ^ {2}}{2 \sigma^ {2}}\right) & \text {i f} s _ {i} \in \text {t h i n g s c l a s s e s}, \\ 0 & \text {o t h e r w i s e .} \end{array} \right. \tag {10} +$$ + +For the things points, the lower $\mathcal{L}_{o_i}$ is, the higher $c_{i}$ is. It means that the point with a more accurate regression of the center offset has a higher confidence score. + +The weighted binary cross entropy loss $\mathcal{L}_{wbce}$ is applied, + +$$ +\mathcal {L} _ {w b c e} = - \frac {1}{N} \sum_ {i = 0} ^ {N - 1} 6 c _ {i} \log (\hat {c} _ {i}) + (1 - c _ {i}) \log (1 - \hat {c} _ {i}), \tag {11} +$$ + +where the things points are manually emphasized since the number of them is much less than that of the stuff points. + +Finally, the loss for each group of results (from the CFNet or CFFE) is defined as + +$$ +\mathcal {L} = \mathcal {L} _ {w c e} + 3 \mathcal {L} _ {l s} + \mathcal {L} _ {t c} + 2 \mathcal {L} _ {o} + \mathcal {L} _ {w b c e}. \tag {12} +$$ + +The total loss is the sum of the two losses sourced from the CFNet and CFFE. + +# 3.3. Center Dedduplication Module + +Given the final predictions of the semantic segmentation $\hat{S}$ , center offsets $\hat{O}$ , and confidence scores $\hat{C}$ , this section introduces how to exploit the detected centers to acquire the panoptic segmentation results during inference, and illustrates the key module, center dedduplication module (CDM). + +Post-process. For the panoptic segmentation results, instance segmentation is first generated and then the final + +Algorithm 1: Center dedduplication module +Input: Shifted things points $\{\tilde{p}_t^{shift}\in \mathbb{R}^3\}_{t = 0}^{M - 1}$ confidence scores $\{\tilde{c}_t\in [0,1]\}_{t = 0}^{M - 1}$ distance threshold $d\in \mathbb{R}$ Output: Instance centers $\{\tilde{e}_j\in \mathbb{R}^3\}_{j = 0}^{D - 1}$ +1 order $=$ sort_index( $\{\tilde{c}_t\}_{t = 0}^{M - 1}$ , decending=true) +2 initialize a boolean vector $rej\in \mathbb{B}^{M}$ with all elements of false and an empty list keep +3 for $k\gets 0$ to $M - 1$ do +4 if not rej[order[k]] then +5 keep.append(order[k]) +6 for $h\gets k + 1$ to $M - 1$ do +7 dist $= ||\tilde{p}_{order[k]}^{shift} - \tilde{p}_{order[h]}^{shift}||_2$ +8 if dist $< d$ then +9 | rej[order[h]] $=$ true +10 end +11 end +12 end +13 end +14 $\{\tilde{e}_j\} = \tilde{p}_{keep}^{shift}$ + +panoptic segmentation is acquired by fusing the semantic and instance segmentation labels. There are five steps to achieve the final panoptic segmentation: + +1) The things points $\{\tilde{p}_t\}_{t=0}^{M-1} = \{\pmb{p}_i | \hat{s}_i \in$ things classes} are selected according to the predicted semantic labels, with their offsets $\{\tilde{o}_t\}_{t=0}^{M-1}$ and confidence scores $\{\tilde{c}_t\}_{t=0}^{M-1}$ (where $M$ is the number of things points). +2) Each shifted things point $\tilde{\pmb{p}}_t^{shift}$ is given by $\tilde{\pmb{p}}_t^{shift} = \tilde{\pmb{p}}_t + \tilde{\pmb{o}}_t$ and is severing as an instance center candidate. +3) A center $\tilde{\pmb{e}}_j$ for each instance $j$ is selected by the CDM based on the coordinates $\tilde{\pmb{p}}_t^{shift}$ and its confidence score $\tilde{c}_t$ , while other candidates are suppressed. +4) The instance $\mathrm{ID}\widetilde{\text{in}}\widetilde{\text{st}}_t$ is acquired by assigning the shifted things points $\tilde{\pmb{p}}_t^{\text{shift}}$ to the closest center in all centers $\{\tilde{\pmb{e}}_j\}_{j=0}^{D-1}$ ( $D$ is the number of detected instances). +5) The majority voting [15, 39] reassigns the most frequent semantic label to all points of a predicted instance to further guarantee the consistency of semantic labels within a predicted instance. + +Center dedduplication module (CDM). The CDM takes the shifted points $\{\tilde{p}_t^{shift}\}_{t=0}^{M-1}$ and confidence scores $\{\tilde{c}_t\}_{t=0}^{M-1}$ as input to get an instance center $\hat{e}_j$ for each instance $j$ . Inspired by the bounding box NMS, our CDM suppresses the candidate centers with lower scores within a euclidean distance threshold $d$ . The pseudo-code of our CDM is shown in Algorithm 1, where two centers with a distance less than $d$ are considered as the same instance. The process of our CDM is simple and can be easily implemented in CUDA. + +# 4. Experiments + +The proposed CFNet is compared with the existing methods on the public SemanticKITTI [2] and nuScenes [5] panoptic segmentation benchmarks. + +# 4.1. Experimental Setup + +Datasets. The SemanticKITTI [2] consists of 22 sequences collected by a Velodyne HDL-64E $360^{\circ}$ rotating LiDAR with 64 beams vertically. It contains 43,552 LiDAR scans in total and is split into a training set with 19,130 scans from sequences 00 to 10 except 08, a validation set containing sequence 08 with 4,071 scans, and a test set including the rest sequences with 20,351 scans. The test set is only provided with point clouds and used for the online leaderboards. For the panoptic segmentation, it has 19 valid categories, including 11 stuff and 8 things categories. + +The nuScenes [5] is a newly released benchmark with 1,000 scenes collected by a Velodyne HDL-32E $360^{\circ}$ rotating LiDAR with 32 beams vertically. The dataset is collected in Boston and Singapore. It uses 28,130 scans for training, 6,019 for validation, and 6,008 for testing. The panoptic segmentation benchmark provides 16 categories, including 6 stuff and 10 things categories [12]. + +Evaluation metric. As the official benchmarks [3, 12] suggest, the panoptic quality (PQ) is adopted to evaluate the performance of LiDAR panoptic segmentation. The PQ [3, 12] measures the overall quality of the panoptic segmentation, and can be decomposed into two terms: 1) the segmentation quality (SQ) is the average IoU of all matched pairs; 2) the recognition quality (RQ) measures the $F_{1}$ score. These three metrics are also reported separately on the stuff and things classes, including $\mathrm{PQ}^{St}$ , $\mathrm{SQ}^{St}$ , $\mathrm{RQ}^{St}$ , and $\mathrm{PQ}^{Th}$ , $\mathrm{SQ}^{Th}$ , $\mathrm{RQ}^{Th}$ , respectively. The $\mathrm{PQ}^{\dagger}$ [26] is also reported by replacing the PQ of each stuff class with its IoU. For evaluating the sub-task of semantic segmentation, mean Intersection over Union (mIoU) [2] is reported. + +Training details. In the experiments, two 2D projection-based backbones, namely PolarNet [38] and CPGNet [23], are used in our CFNet. For the backbone of PolarNet [38], the training schedules and data augmentation are the same as those of Panoptic-PolarNet [39] for a fair comparison. Besides, for efficiency, only one stage version of CPGNet [23] is used in our CFNet. It is trained from scratch for 48 epochs with a batch size of 8 and takes around 24 hours on 8 NVIDIA RTX 3090 GPUs. Stochastic gradient descent (SGD) is used as the optimizer, where the initial learning rate, weight decay, and momentum are 0.02, 0.001, and 0.9, respectively. The learning rate is decayed by 0.1 per 10 epochs. In training, the data augmentation is applied as a convention, including random flipping along the $x$ and $y$ axes, random global scale sampled from $[0.95, 1.05]$ , random rotation around the $z$ -axis, and random Gaussian noise $\mathcal{N}(0, 0.02)$ . + +
BackboneCFFEDeduplicationPQPQ†PQTthPQS\( ^t \)mIoURT(ms)
CFGFEM
a b c dPolarNet [38]✓BEV Center Heatmap [39]59.164.165.754.364.565.3 + 20.9
BEV Center Heatmap [39]59.564.366.254.464.669.7 + 20.9
✓✓BEV Center Heatmap [39]60.465.667.255.165.271.9 + 20.9
✓✓CDM [Ours]60.665.767.855.265.471.9 + 3.6
e f g h i jCPGNet [23]✓DBSCAN [11]59.564.163.856.364.631.6+24.9
HDBSCAN [6]58.362.960.956.364.931.6+48.7
MeanShift [9]59.964.564.756.364.931.6+84.6
CDM [Ours]60.565.666.156.566.332.7+2.3
✓CDM [Ours]60.865.966.356.966.537.4+2.3
✓✓CDM [Ours]62.767.570.057.367.441.2+2.3
+ +# 4.2. Ablation Studies + +Ablative analyses are conducted on the SemanticKITTI validation set with the above experimental setup. As shown in Table 1, two 2D projection-based backbones, namely PolarNet [38] and CPGNet [23], are incorporated with our framework to validate the proposed CFFE and CDM. Row (a) shows the original Panoptic-Polarnet [39], while rows (b,c,d) are implemented based on the official code of Panoptic-Polarnet [39]. The running time is reported by the time of the model inference plus that of the post-process. + +Effects of the CFFE. The CFFE is configured in three ways: 1) the CFFE is not applied (a,e,f,g,h); 2) the center feature generation (CFG) in the CFFE is removed (b,i), which is equal to the setting of only imposing the intermediate supervision; 3) the whole CFFE is applied (c,d,j). For the two backbones, a common observation is that the improvements primarily come from the CFG rather than the intermediate supervision, especially for the things classes (a,b,c and h,i,j). In terms of the running time, The CFFE brings acceptable latency of $+6.6\mathrm{ms}$ and $+8.5\mathrm{ms}$ for these two backbones, respectively. + +Moreover, the errors in predicting things center offsets are calculated for both the CFNet with CFFE and its intermediate results as shown in Table 2. Equipped with the backbone of CPGNet [23], the proposed CFFE reduces the errors of predicting the things center offsets. The observations demonstrate that explicitly modeling center features helps in identifying an instance. + +Effects of the CDM. For the proposal-free meth + +Table 1. Ablation studies on the SemanticKITTI validation set. RT: running time. + +
training setvalidation set
Intermediate1.351.22
CFNet1.191.13
+ +Table 2. Mean errors measured in meter (m) of things center offsets for the intermediate results and the CFNet with the CFFE on the SemanticKITTI training and validation set. + +ods, there are primarily three ways to identify each instance: 1) the class-agnostic clustering methods (e.g. DBSCAN, HDBSCAN, MeanShift) (e,f,g), 2) the BEV center heatmap [39] (a,b,c), and 3) our CDM (d,h,i,j). For the backbone of PolarNet [38], our CDM outperforms the BEV center heatmap [39] by $+0.2$ for the PQ (c,d), while our CDM runs faster. For the backbone of CPGNet [23], compared with the class-agnostic clustering methods, our CDM also achieves the best performance on both accuracy $(+0.6\sim +2.2$ for the PQ) and running time $(-21.5\mathrm{ms}\sim -$ $81.2\mathrm{ms})$ , though it slows down the model inference by $+1.1~\mathrm{ms}$ $\mathrm{(g,h)}$ owing to the additional confidence prediction. + +More ablation studies can be seen in the supplementary material. + +# 4.3. Evaluation Comparisons + +Quantitative results. In Table 3 and Table 4, our CFNet with CPGNet as its backbone is compared with the existing methods on the SemanticKITTI test set and nuScenes validation set, respectively. From top to bottom, the methods are grouped as proposal-based and proposal-free methods. It is found that our proposal-free CFNet surpasses the previous methods by a large margin on most metrics. Specifically, on the SemanticKITTI test set, our CFNet outperforms the previous state-of-the-art SCAN [35] and Panoptic-PHNet [22] by $+1.9$ for the PQ, and $+1.5$ for the RQ. On the nuScenes validation set, our CFNet outperforms the most competitive Panoptic-PHNet [22] by $+0.4$ for the PQ, $+0.6$ for the SQ, and $+0.4$ for the RQ, respectively. + +Running time. The running time is reported on the SemanticKITTI. It is measured with PyTorch FP32 without any model acceleration tricks on a single NVIDIA RTX 3090 GPU. For clarity, the running time of our method is further divided into two parts, including the time costs of the model and post-process. Based on the same kind of devices as the existing methods, our CFNet runs $43.5\mathrm{ms}$ (41.2 ms for model inference and 2.3 ms for post-process) and is 1.6 times faster than the most efficient Panoptic-PHNet [22]. + +Qualitative results. The visualization results are + +
MethodsPQPQ†SQRQPQThSQThRQThPQStSQStRQStmIoURT(ms)
Proposal-based Methods
RangeNet++ [25]&PointPillars [19]37.145.975.947.020.275.225.249.376.562.852.4416.6
KPConv [33]&PointPillars [19]44.552.580.054.432.781.538.753.179.065.958.8526.3
PanopticTrackNet [17]43.150.778.853.928.680.435.553.677.767.352.6147
EfficientLPS [31]57.463.283.068.753.187.860.560.579.574.661.4-
Proposal-free Methods
LPSAD [24]38.047.076.548.225.676.831.847.176.260.150.984.7
DS-Net [15]55.962.582.366.755.187.262.856.578.769.561.6294.1
Panoster [13]52.759.980.764.149.483.358.555.178.868.259.9-
GP-S3Net [29]60.069.082.072.165.086.674.556.478.770.470.8270.3
SMAC-Seg [20]56.162.582.067.953.085.661.858.479.372.363.399
Panoptic-PolarNet [39]54.160.781.465.053.387.260.654.877.268.159.586.2
SCAN [35]61.567.584.572.161.488.169.361.581.874.167.778.1
Panoptic-PHNet [22]61.567.984.872.163.890.770.459.980.573.366.069.3
CFNet [Ours] w/CPGNet [23]63.469.785.273.666.789.874.361.081.973.168.341.2 + 2.3
+ +Table 3. Comparison results on the SemanticKITTI test set. + +
MethodsPQ\( PQ^† \)SQRQ\( PQ^{Th} \)\( SQ^{Th} \)\( RQ^{Th} \)\( PQ^{St} \)\( SQ^{St} \)\( RQ^{St} \)mIoU
Proposal-based Methods
Cylinder3D [40]&PointPillars [19]36.044.583.343.023.383.727.057.282.769.652.3
Cylinder3D [40]&SECOND [36]40.148.484.247.329.084.433.658.583.770.158.5
PanopticTrackNet [17]51.456.280.263.345.881.455.960.478.375.558.0
EfficientLPS [31]62.065.683.473.956.883.268.070.683.883.665.6
Proposal-free Methods
LPSAD [24]50.457.779.462.443.280.253.257.578.571.762.5
DS-Net [15]42.551.083.650.332.583.138.359.284.470.370.7
GP-S3Net [29]61.067.584.172.056.085.365.266.082.978.775.8
SMAC-Seg [20]67.071.885.078.265.287.174.268.882.982.272.2
Panoptic-PolarNet [39]63.467.283.975.359.284.170.370.483.683.566.9
SCAN [35]65.168.985.775.360.685.770.272.585.783.877.4
Panoptic-PHNet [22]74.777.788.284.274.089.082.575.986.886.979.7
CFNet [Ours] w/CPGNet [23]75.178.088.884.674.889.882.976.687.187.379.3
+ +Table 4. Comparison results on the nuScenes validation set. + +![](images/b052111d925cc62768d07df341b90122b38eeabbf009657386e66860535c3e11.jpg) +Semantic Segmentation + +![](images/e5f441e2eb16aac27dfee5184afbbbaaa802c81f1a6a4849e6c66705f19ab3fd.jpg) + +![](images/d6176e70d804c2a52d065fda3f5ff5c752045d44377594bd42ab595d7a3169a3.jpg) +Instance Segmentation +Figure 5. Visualizations of our CFNet on the SemanticKITTI test set. Different colors represent different classes or instances. + +shown in Fig. 5. Our CFNet can distinguish adjacent pedestrians or cars. Besides, the boundaries of instances can be accurately segmented. More qualitative results can be seen in the supplementary material. + +# 5. Conclusion + +A novel proposal-free CFNet is proposed for real-time LiDAR panoptic segmentation. For better modeling and exploiting the non-existent instance centers, a novel CFFE is proposed to generate enhanced center-focusing feature maps and a CDM is introduced to keep only one center for each instance and then assign the shifted things points to the closest center for acquiring instance IDs. From the experiments, it can be seen that center modeling and exploiting is a key problem in the proposal-free LiDAR panoptic segmentation methods, and mimicking the non-existent center features is promising and shows a clear benefit. + +# Acknowledgements + +The research project is partially supported by the National Key R&D Program of China (No.2021ZD0111902), and the National Natural Science Foundation of China (No.U21B2038, 62272015, U19B2039). + +# References + +[1] Yara Ali Alnaggar, Mohamed Afifi, Karim Amer, and Mohamed ElHelw. Multi projection fusion for real-time semantic segmentation of 3d lidar point clouds. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1800-1809, 2021. 2 +[2] Jens Behley, Martin Garbade, Andres Milioto, Jan Quenzel, Sven Behnke, Cyril Stachniss, and Jurgen Gall. Semantickitti: A dataset for semantic scene understanding of lidar sequences. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9297-9307, 2019. 6 +[3] Jens Behley, Andres Milioto, and Cyril Stachniss. A benchmark for lidar-based panoptic segmentation based on kitti. In 2021 IEEE International Conference on Robotics and Automation (ICRA), pages 13596-13603. IEEE, 2021. 6 +[4] Maxim Berman, Amal Rannen Triki, and Matthew B Blaschko. The lovasz-softmax loss: A tractable surrogate for the optimization of the intersection-over-union measure in neural networks. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 4413-4421, 2018. 5 +[5] Holger Caesar, Varun Bankiti, Alex H Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuscenes: A multimodal dataset for autonomous driving. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11621-11631, 2020. 6 +[6] Ricardo JGB Campello, Davoud Moulavi, and Jörg Sander. Density-based clustering based on hierarchical density estimates. In Pacific-Asia conference on knowledge discovery and data mining, pages 160–172. Springer, 2013. 7 +[7] Ran Cheng, Ryan Razani, Ehsan Taghavi, Enxu Li, and Bingbing Liu. 2-s3net: Attentive feature fusion with adaptive feature selection for sparse semantic segmentation network. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12547-12556, 2021. 2 +[8] Christopher Choy, JunYoung Gwak, and Silvio Savarese. 4d spatio-temporal convnets: Minkowski convolutional neural networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3075-3084, 2019. 2 +[9] Dorin Comaniciu and Peter Meer. Mean shift: A robust approach toward feature space analysis. IEEE Transactions on pattern analysis and machine intelligence, 24(5):603-619, 2002. 3, 7 +[10] Tiago Cortinhal, George Tzelepis, and Eren Erdal Aksoy. Salsanext: Fast, uncertainty-aware semantic segmentation of lidar point clouds. In International Symposium on Visual Computing, pages 207-222. Springer, 2020. 2, 4 +[11] Martin Ester, Hans-Peter Kriegel, Jörg Sander, Xiaowei Xu, et al. A density-based algorithm for discovering clusters in large spatial databases with noise. In kdd, volume 96, pages 226-231, 1996. 3, 7 +[12] Whye Kit Fong, Rohit Mohan, Juana Valeria Hurtado, Lubing Zhou, Holger Caesar, Oscar Beijbom, and Abhinav Val- + +ada. Panoptic nuscenes: A large-scale benchmark for lidar panoptic segmentation and tracking. IEEE Robotics and Automation Letters, 7(2):3795-3802, 2022. 6 +[13] Stefano Gasperini, Mohammad-Ali Nikouei Mahani, Alvaro Marcos-Ramiro, Nassir Navab, and Federico Tombari. Panoster: End-to-end panoptic segmentation of lidar point clouds. IEEE Robotics and Automation Letters, 6(2):3216-3223, 2021. 1, 3, 8 +[14] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In Proceedings of the IEEE international conference on computer vision, pages 2961-2969, 2017. 1, 2 +[15] Fangzhou Hong, Hui Zhou, Xinge Zhu, Hongsheng Li, and Ziwei Liu. Lidar-based panoptic segmentation via dynamic shifting network. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13090-13099, 2021. 1, 3, 5, 6, 8 +[16] Qingyong Hu, Bo Yang, Linhai Xie, Stefano Rosa, Yulan Guo, Zhihua Wang, Niki Trigoni, and Andrew Markham. Randla-net: Efficient semantic segmentation of large-scale point clouds. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11108-11117, 2020. 2 +[17] Juana Valeria Hurtado, Rohit Mohan, Wolfram Burgard, and Abhinav Valada. Mopt: Multi-object panoptic tracking. arXiv preprint arXiv:2004.08189, 2020. 1, 2, 8 +[18] Alexander Kirillov, Kaiming He, Ross Girshick, Carsten Rother, and Piotr Dollar. Panoptic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9404-9413, 2019. 1 +[19] Alex H Lang, Sourabh Vora, Holger Caesar, Lubing Zhou, Jiong Yang, and Oscar Beijbom. Pointpillars: Fast encoders for object detection from point clouds. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12697-12705, 2019. 1, 2, 4, 8 +[20] Enxu Li, Ryan Razani, Yixuan Xu, and Liu Bingbing. Smac-seg: Lidar panoptic segmentation via sparse multi-directional attention clustering. arXiv preprint arXiv:2108.13588, 2021. 2, 3, 8 +[21] Enxu Li, Ryan Razani, Yixuan Xu, and Bingbing Liu. Cpseg: Cluster-free panoptic segmentation of 3d lidar point clouds. arXiv preprint arXiv:2111.01723, 2021. 1 +[22] Jinke Li, Xiao He, Yang Wen, Yuan Gao, Xiaogiang Cheng, and Dan Zhang. Panoptic-phnet: Towards real-time and high-precision lidar panoptic segmentation via clustering pseudo heatmap. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11809–11818, 2022. 1, 3, 7, 8 +[23] Xiaoyan Li, Gang Zhang, Hongyu Pan, and Zhenhua Wang. Cpgnet: Cascade point-grid fusion network for real-time lidar semantic segmentation. arXiv preprint arXiv:2204.09914, 2022. 2, 4, 5, 6, 7, 8 +[24] Andres Milioto, Jens Behley, Chris McCool, and Cyril Stachniss. Lidar panoptic segmentation for autonomous driving. In 2020 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 8505-8512. IEEE, 2020. 2, 8 + +[25] Andres Milioto, Ignacio Vizzo, Jens Behley, and Cyril Stachniss. Rangenet++: Fast and accurate lidar semantic segmentation. In 2019 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 4213-4220. IEEE, 2019. 2, 4, 8 +[26] Lorenzo Porzi, Samuel Rota Bulo, Aleksander Colovic, and Peter Kontschieder. Seamless scene segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8277-8286, 2019. 6 +[27] Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. Pointnet: Deep learning on point sets for 3d classification and segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 652-660, 2017. 2 +[28] Charles R Qi, Li Yi, Hao Su, and Leonidas J Guibas. Point-net++: Deep hierarchical feature learning on point sets in a metric space. arXiv preprint arXiv:1706.02413, 2017. 2 +[29] Ryan Razani, Ran Cheng, Enxu Li, Ehsan Taghavi, Yuan Ren, and Liu Bingbing. Gp-s3net: Graph-based panoptic sparse semantic segmentation network. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 16076-16085, 2021. 1, 3, 8 +[30] Shaoshuai Shi, Chaoxu Guo, Li Jiang, Zhe Wang, Jianping Shi, Xiaogang Wang, and Hongsheng Li. Pv-rcnn: Pointvoxel feature set abstraction for 3d object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10529–10538, 2020. 1 +[31] Kshitij Sirohi, Rohit Mohan, Daniel Buscher, Wolfram Burgard, and Abhinav Valada. Efficienttps: Efficient lidar panoptic segmentation. IEEE Transactions on Robotics, 2021. 1, 2, 8 +[32] Haotian Tang, Zhijian Liu, Shengyu Zhao, Yujun Lin, Ji Lin, Hanrui Wang, and Song Han. Searching efficient 3d architectures with sparse point-voxel convolution. In European Conference on Computer Vision, pages 685-702. Springer, 2020. 2 +[33] Hugues Thomas, Charles R Qi, Jean-Emmanuel Deschaud, Beatrix Marcotegui, François Goulette, and Leonidas J Guibas. Kpconv: Flexible and deformable convolution for point clouds. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6411–6420, 2019. 2, 8 +[34] Chenfeng Xu, Bichen Wu, Zining Wang, Wei Zhan, Peter Vajda, Kurt Keutzer, and Masayoshi Tomizuka. Squeeze-segv3: Spatially-adaptive convolution for efficient pointcloud segmentation. In European Conference on Computer Vision, pages 1-19. Springer, 2020. 2, 4 +[35] Shuangjie Xu, Rui Wan, Maosheng Ye, Xiaoyi Zou, and Tongyi Cao. Sparse cross-scale attention network for efficient lidar panoptic segmentation. arXiv preprint arXiv:2201.05972, 2022. 1, 2, 3, 7, 8 +[36] Yan Yan, Yuxing Mao, and Bo Li. Second: Sparsely embedded convolutional detection. Sensors, 18(10):3337, 2018. 8 +[37] Dongqiangzi Ye, Zixiang Zhou, Weijia Chen, Yufei Xie, Yu Wang, Panqu Wang, and Hassan Foroosh. Lidarmultinet: Towards a unified multi-task network for lidar perception. arXiv preprint arXiv:2209.09385, 2022. 1, 2 + +[38] Yang Zhang, Zixiang Zhou, Philip David, Xiangyu Yue, Zerong Xi, Boqing Gong, and Hassan Foroosh. Polarnet: An improved grid representation for online lidar point clouds semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9601-9610, 2020. 2, 4, 6, 7 +[39] Zixiang Zhou, Yang Zhang, and Hassan Foroosh. Panoptic-polarnet: Proposal-free lidar point cloud panoptic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13194-13203, 2021. 1, 2, 3, 6, 7, 8 +[40] Xinge Zhu, Hui Zhou, Tai Wang, Fangzhou Hong, Yuexin Ma, Wei Li, Hongsheng Li, and Dahua Lin. Cylindrical and asymmetrical 3d convolution networks for lidar segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9939-9948, 2021. 2, 8 \ No newline at end of file diff --git a/2023/Center Focusing Network for Real-Time LiDAR Panoptic Segmentation/images.zip b/2023/Center Focusing Network for Real-Time LiDAR Panoptic Segmentation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..ba4ac1059e0349f97daf118543e0fe63e275b5a1 --- /dev/null +++ b/2023/Center Focusing Network for Real-Time LiDAR Panoptic Segmentation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b862f0698b42d78c9f80c412dedf5164ecfb161e2646321894b80921d90412bd +size 630233 diff --git a/2023/Center Focusing Network for Real-Time LiDAR Panoptic Segmentation/layout.json b/2023/Center Focusing Network for Real-Time LiDAR Panoptic Segmentation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..2ed29cfd6c424e22bed52858b4a0ebd6edbcf042 --- /dev/null +++ b/2023/Center Focusing Network for Real-Time LiDAR Panoptic Segmentation/layout.json @@ -0,0 +1,9978 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 75, + 103, + 518, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 103, + 518, + 121 + ], + "spans": [ + { + "bbox": [ + 75, + 103, + 518, + 121 + ], + "type": "text", + "content": "Center Focusing Network for Real-Time LiDAR Panoptic Segmentation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 104, + 141, + 485, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 104, + 141, + 485, + 157 + ], + "spans": [ + { + "bbox": [ + 104, + 141, + 485, + 157 + ], + "type": "text", + "content": "Xiaoyan Li" + }, + { + "bbox": [ + 104, + 141, + 485, + 157 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 104, + 141, + 485, + 157 + ], + "type": "text", + "content": " Gang Zhang" + }, + { + "bbox": [ + 104, + 141, + 485, + 157 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 104, + 141, + 485, + 157 + ], + "type": "text", + "content": " Boyue Wang" + }, + { + "bbox": [ + 104, + 141, + 485, + 157 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 104, + 141, + 485, + 157 + ], + "type": "text", + "content": " Yongli Hu" + }, + { + "bbox": [ + 104, + 141, + 485, + 157 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 104, + 141, + 485, + 157 + ], + "type": "text", + "content": " Baocai Yin" + }, + { + "bbox": [ + 104, + 141, + 485, + 157 + ], + "type": "inline_equation", + "content": "^{1,2}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 103, + 157, + 490, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 103, + 157, + 490, + 171 + ], + "spans": [ + { + "bbox": [ + 103, + 157, + 490, + 171 + ], + "type": "text", + "content": "1Beijing Municipal Key Lab of Multimedia and Intelligent Software Technology" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 107, + 171, + 487, + 199 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 107, + 171, + 487, + 199 + ], + "spans": [ + { + "bbox": [ + 107, + 171, + 487, + 199 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 107, + 171, + 487, + 199 + ], + "type": "text", + "content": "Beijing Institute of Artificial Intelligence, Faculty of Information Technology, Beijing University of Technology, Beijing 100124, China" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 116, + 199, + 476, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 116, + 199, + 476, + 213 + ], + "spans": [ + { + "bbox": [ + 116, + 199, + 476, + 213 + ], + "type": "text", + "content": "3Mogo Auto Intelligence and Telemetics Information Technology Co. Ltd." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 104, + 215, + 487, + 227 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 104, + 215, + 487, + 227 + ], + "spans": [ + { + "bbox": [ + 104, + 215, + 487, + 227 + ], + "type": "text", + "content": "{xiaoyan.li, wby, huyongli, ybc}@bjut.edu.cn zhanggang11021136@gmail.com" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 143, + 255, + 192, + 268 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 255, + 192, + 268 + ], + "spans": [ + { + "bbox": [ + 143, + 255, + 192, + 268 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 281, + 290, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 281, + 290, + 521 + ], + "spans": [ + { + "bbox": [ + 46, + 281, + 290, + 521 + ], + "type": "text", + "content": "LiDAR panoptic segmentation facilitates an autonomous vehicle to comprehensively understand the surrounding objects and scenes and is required to run in real time. The recent proposal-free methods accelerate the algorithm, but their effectiveness and efficiency are still limited owing to the difficulty of modeling non-existent instance centers and the costly center-based clustering modules. To achieve accurate and real-time LiDAR panoptic segmentation, a novel center focusing network (CFNet) is introduced. Specifically, the center focusing feature encoding (CFFE) is proposed to explicitly understand the relationships between the original LiDAR points and virtual instance centers by shifting the LiDAR points and filling in the center points. Moreover, to leverage the redundantly detected centers, a fast center dedduplication module (CDM) is proposed to select only one center for each instance. Experiments on the SemanticKITTI and nuScenes panoptic segmentation benchmarks demonstrate that our CFNet outperforms all existing methods by a large margin and is 1.6 times faster than the most efficient method." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 548, + 128, + 561 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 548, + 128, + 561 + ], + "spans": [ + { + "bbox": [ + 47, + 548, + 128, + 561 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "type": "text", + "content": "Panoptic segmentation [18] combines both semantic segmentation and instance segmentation in a single framework. It predicts semantic labels for the uncountable stuff classes (e.g. road, sidewalk), while it simultaneously provides semantic labels and instance IDs for the countable things classes (e.g. car, pedestrian). The LiDAR panoptic segmentation is one of the bases for the safety of autonomous driving, which employs the point clouds collected by the Light Detection and Ranging (LiDAR) sensors to effectively depict the surroundings. Existing LiDAR panoptic segmentation methods first conduct semantic segmentation, and then achieve instance segmentation for the things categories in" + } + ] + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 312, + 253, + 540, + 434 + ], + "blocks": [ + { + "bbox": [ + 312, + 253, + 540, + 434 + ], + "lines": [ + { + "bbox": [ + 312, + 253, + 540, + 434 + ], + "spans": [ + { + "bbox": [ + 312, + 253, + 540, + 434 + ], + "type": "image", + "image_path": "bdf504294a4b4d31dbec88f296864e5e0cb8826f00a6b19e1001db7767b8cd69.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 441, + 547, + 475 + ], + "lines": [ + { + "bbox": [ + 305, + 441, + 547, + 475 + ], + "spans": [ + { + "bbox": [ + 305, + 441, + 547, + 475 + ], + "type": "text", + "content": "Figure 1. PQ vs. runtime on the SemanticKITTI test set. Runtime measurements are taken on a single NVIDIA RTX 3090 GPU. The panoptic quality (PQ) is introduced in section 4.1." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 497, + 538, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 497, + 538, + 509 + ], + "spans": [ + { + "bbox": [ + 305, + 497, + 538, + 509 + ], + "type": "text", + "content": "two ways, the proposal-based and proposal-free methods." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 510, + 545, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 510, + 545, + 605 + ], + "spans": [ + { + "bbox": [ + 304, + 510, + 545, + 605 + ], + "type": "text", + "content": "The proposal-based methods [17, 31, 37] adopt a two-stage process similar to the well-known Mask R-CNN [14] in the image domain. It first generates object proposals for the things points by using 3D detection networks [19, 30] and then refines the instance segmentation results within each proposal. As shown in Fig. 1, these methods are usually complicated and hardly achieve real-time processing, owing to their sequential multi-stage pipelines." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 605, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 545, + 714 + ], + "type": "text", + "content": "The proposal-free frameworks [13, 15, 21, 22, 29, 35, 39] are more compact. To associate the things points with instance IDs, these methods usually leverage the instance centers. Specifically, they regress the offsets from the points to their corresponding centers, and then adopt the class-agnostic center-based clustering modules [13, 15, 29] or the bird's-eye view (BEV) center heatmap [22, 35, 39]. However, two problems exist in these methods. First, for center feature extracting and center modeling, the non-existent in" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13425" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 55, + 70, + 161, + 157 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 161, + 157 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 161, + 157 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 161, + 157 + ], + "type": "image", + "image_path": "739d52840857770b48bf2f272685550328dd7f45992da9ede8d2413a38ad8818.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 59, + 160, + 156, + 171 + ], + "lines": [ + { + "bbox": [ + 59, + 160, + 156, + 171 + ], + "spans": [ + { + "bbox": [ + 59, + 160, + 156, + 171 + ], + "type": "text", + "content": "(a) CFNet without CFFE" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 173, + 70, + 279, + 157 + ], + "blocks": [ + { + "bbox": [ + 173, + 70, + 279, + 157 + ], + "lines": [ + { + "bbox": [ + 173, + 70, + 279, + 157 + ], + "spans": [ + { + "bbox": [ + 173, + 70, + 279, + 157 + ], + "type": "image", + "image_path": "155e0ba77e4ce0beccb1409edc9db8247cd9a825ad52c636deb8ccd9172fd29d.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 184, + 159, + 269, + 171 + ], + "lines": [ + { + "bbox": [ + 184, + 159, + 269, + 171 + ], + "spans": [ + { + "bbox": [ + 184, + 159, + 269, + 171 + ], + "type": "text", + "content": "(b) CFNet with CFFE" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 180, + 287, + 214 + ], + "lines": [ + { + "bbox": [ + 46, + 180, + 287, + 214 + ], + "spans": [ + { + "bbox": [ + 46, + 180, + 287, + 214 + ], + "type": "text", + "content": "Figure 2. Instance segmentation of a car. Without our CFFE, the car is split into parts (a), while the CFFE significantly alleviates this problem (b). Different colors represent different instances." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 232, + 286, + 364 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 232, + 286, + 364 + ], + "spans": [ + { + "bbox": [ + 46, + 232, + 286, + 364 + ], + "type": "text", + "content": "stance centers increase the difficulty, considering that the LiDAR points are usually surface-aggregated [35] and an instance center is imaginary in most cases. As shown in Fig. 2(a), the difficulty often results in the fault that one instance is incorrectly split into several parts. Second, for exploiting the redundantly detected centers, the clustering modules (e.g. MeanShift, DBSCAN) are too time-consuming to support the real-time autonomous driving perception systems, while the BEV center heatmap cannot distinguish objects with different altitudes in the same BEV grid." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 365, + 286, + 603 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 365, + 286, + 603 + ], + "spans": [ + { + "bbox": [ + 46, + 365, + 286, + 603 + ], + "type": "text", + "content": "For accurate and fast LiDAR panoptic segmentation, a proposal-free center focusing network (CFNet) is proposed. For better encoding center features, a novel center focusing feature encoding (CFFE) is proposed to generate center-focusing feature maps by shifting the things points to fill in the non-existent instance centers for more accurate predictions (as shown in Fig. 2(b)). For center modeling, the CFNet not only decomposes the panoptic segmentation task into the widely-used semantic segmentation and center offset regression, but also proposes a new confidence score prediction for indicating the accuracy of the center offset regression. Subsequently, for the detected centers exploiting, a novel center dedduplication module (CDM) is designed to select one center for a single instance. The CDM keeps the predicted centers with higher confidence scores, while suppressing the ones with lower confidence. Finally, instance segmentation is achieved by assigning the shifted things points to the closest center. For efficiency, the proposed CFNet is built on the 2D projection-based segmentation paradigm. Our contributions are as follows:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 58, + 610, + 286, + 712 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 58, + 610, + 286, + 658 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 610, + 286, + 658 + ], + "spans": [ + { + "bbox": [ + 58, + 610, + 286, + 658 + ], + "type": "text", + "content": "- A proposal-free CFNet is proposed to achieve accurate and fast LiDAR panoptic segmentation by solving the bottleneck problems of center modeling and center-based clustering in previous methods." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 58, + 665, + 286, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 665, + 286, + 712 + ], + "spans": [ + { + "bbox": [ + 58, + 665, + 286, + 712 + ], + "type": "text", + "content": "- The CFFE is proposed to alleviate the difficulty of modeling the non-existent instance centers and the CDM is designed to efficiently keep one center for each instance." + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 318, + 72, + 545, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 72, + 545, + 120 + ], + "spans": [ + { + "bbox": [ + 318, + 72, + 545, + 120 + ], + "type": "text", + "content": "- The proposed CFNet is evaluated on the nuScenes and SemanticKITTI LiDAR panoptic segmentation benchmarks. Our CFNet achieves the state-of-the-art performance with a real-time inference speed." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 131, + 392, + 143 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 131, + 392, + 143 + ], + "spans": [ + { + "bbox": [ + 306, + 131, + 392, + 143 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 151, + 545, + 270 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 151, + 545, + 270 + ], + "spans": [ + { + "bbox": [ + 304, + 151, + 545, + 270 + ], + "type": "text", + "content": "The LiDAR panoptic segmentation methods usually adopt the LiDAR semantic segmentation networks as the backbone and jointly optimize the semantic segmentation and instance segmentation tasks. The results of the two tasks are fused to generate the final panoptic segmentation results. The LiDAR semantic segmentation backbone is first introduced. Then, the panoptic segmentation methods, achieving instance segmentation upon the semantic segmentation backbone, are grouped into the proposal-based and proposal-free methods, and are further discussed." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 270, + 546, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 270, + 546, + 521 + ], + "spans": [ + { + "bbox": [ + 304, + 270, + 546, + 521 + ], + "type": "text", + "content": "LiDAR semantic segmentation backbone. The backbone network for LiDAR semantic segmentation can be divided into three groups: point-based, voxel-based, and 2D projection-based methods. The point-based methods [16, 27, 28, 33] directly operate on the raw point clouds but are extremely time-consuming due to the expensive local neighbor searching. The voxel-based methods [7, 8, 32, 40] discretize the point clouds into structured voxels, where sparse 3D convolutions are applied. These methods are still difficult to meet the real-time applications, although they achieve the highest accuracy. The 2D projection-based methods are more efficient since most of their computation is done in the 2D space, such as range view (RV) [10,25,34], bird's-eye view (BEV) [19], polar view [38], and multiview fusion [1, 23]. For fast LiDAR panoptic segmentation, the Panoptic-PolarNet [39], SMAC-Seg [20], and LPSAD [24] all adopt the 2D projection-based backbone. A recent real-time semantic segmentation method, namely CPGNet [23], is a 2D projection-based one that explores an end-to-end multi-view fusion framework by fusing the features from the point, BEV, and RV." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 522, + 545, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 522, + 545, + 665 + ], + "spans": [ + { + "bbox": [ + 304, + 522, + 545, + 665 + ], + "type": "text", + "content": "**Proposal-based methods.** The proposal-based methods conduct instance segmentation through a two-stage complicated process. They first detect the foreground instances and then refine the instance segmentation results independently within each detected bounding box. Based on the Mask R-CNN [14] for instance segmentation, the MOPT [17] and EfficientLPS [31] insert a semantic branch to achieve panoptic segmentation on the range view (RV). Recently, LidarMultiNet [37] unifies LiDAR-based 3D object detection, semantic segmentation, and panoptic segmentation in a single framework to reduce the computation cost by sharing a strong voxel-based backbone." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "content": "Proposal-free methods. The proposal-free methods usually apply the class-agnostic clustering to the things points to conduct instance segmentation. The LPSAD [24] clusters points into instances by regressing the center off" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13426" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 47, + 70, + 545, + 262 + ], + "blocks": [ + { + "bbox": [ + 47, + 70, + 545, + 262 + ], + "lines": [ + { + "bbox": [ + 47, + 70, + 545, + 262 + ], + "spans": [ + { + "bbox": [ + 47, + 70, + 545, + 262 + ], + "type": "image", + "image_path": "6b477c5cd5f7519ef9d8618ecd027fb88a9c3a0aa05dd6413acf48af6c7613cb.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 380, + 546, + 424 + ], + "lines": [ + { + "bbox": [ + 46, + 380, + 546, + 424 + ], + "spans": [ + { + "bbox": [ + 46, + 380, + 546, + 424 + ], + "type": "text", + "content": "Figure 3. The overview of our CFNet. It consists of four steps: 1) the 2D projection-based backbone extracts features on the 2D space; 2) the proposed center focusing feature encoding (CFFE) mimics and enhances the non-existent instance center features; 3) the panoptic segmentation head predicts the output results; 4) the proposed center dedduplication module (CDM) achieves instance segmentation that is fused to generate the final panoptic segmentation results. The dashed lines indicate that the operations are only used during inference." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 49, + 264, + 274, + 329 + ], + "blocks": [ + { + "bbox": [ + 49, + 264, + 274, + 329 + ], + "lines": [ + { + "bbox": [ + 49, + 264, + 274, + 329 + ], + "spans": [ + { + "bbox": [ + 49, + 264, + 274, + 329 + ], + "type": "image", + "image_path": "8493b7dd4cb04a7ae0b68c07950105b284750e2ac83566bf88e934fa0a7a78a5.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 63, + 335, + 151, + 349 + ], + "lines": [ + { + "bbox": [ + 63, + 335, + 151, + 349 + ], + "spans": [ + { + "bbox": [ + 63, + 335, + 151, + 349 + ], + "type": "text", + "content": ": Multi-Layer Perception" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 170, + 339, + 232, + 348 + ], + "lines": [ + { + "bbox": [ + 170, + 339, + 232, + 348 + ], + "spans": [ + { + "bbox": [ + 170, + 339, + 232, + 348 + ], + "type": "text", + "content": "P2G: Point to Grid" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 248, + 338, + 362, + 348 + ], + "lines": [ + { + "bbox": [ + 248, + 338, + 362, + 348 + ], + "spans": [ + { + "bbox": [ + 248, + 338, + 362, + 348 + ], + "type": "text", + "content": "FCN: Fully Convolutional Network" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 63, + 353, + 149, + 367 + ], + "lines": [ + { + "bbox": [ + 63, + 353, + 149, + 367 + ], + "spans": [ + { + "bbox": [ + 63, + 353, + 149, + 367 + ], + "type": "text", + "content": ": Fully Connected Layer" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 170, + 354, + 232, + 363 + ], + "lines": [ + { + "bbox": [ + 170, + 354, + 232, + 363 + ], + "spans": [ + { + "bbox": [ + 170, + 354, + 232, + 363 + ], + "type": "text", + "content": "G2P: Grid to Point" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 249, + 354, + 303, + 362 + ], + "lines": [ + { + "bbox": [ + 249, + 354, + 303, + 362 + ], + "spans": [ + { + "bbox": [ + 249, + 354, + 303, + 362 + ], + "type": "text", + "content": "PF: Point Fusion" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_footnote" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 293, + 264, + 544, + 329 + ], + "blocks": [ + { + "bbox": [ + 293, + 264, + 544, + 329 + ], + "lines": [ + { + "bbox": [ + 293, + 264, + 544, + 329 + ], + "spans": [ + { + "bbox": [ + 293, + 264, + 544, + 329 + ], + "type": "image", + "image_path": "c29088879b4f7a3e7e87ca8fb061349c93ad6c9c5e8abf4bd960e08545d9314d.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 386, + 339, + 521, + 350 + ], + "lines": [ + { + "bbox": [ + 386, + 339, + 521, + 350 + ], + "spans": [ + { + "bbox": [ + 386, + 339, + 521, + 350 + ], + "type": "text", + "content": "CFFE: Center Focusing Feature Encoding" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 386, + 354, + 503, + 363 + ], + "lines": [ + { + "bbox": [ + 386, + 354, + 503, + 363 + ], + "spans": [ + { + "bbox": [ + 386, + 354, + 503, + 363 + ], + "type": "text", + "content": "CDM: Center Dedduplication Module" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_footnote" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 445, + 289, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 445, + 289, + 639 + ], + "spans": [ + { + "bbox": [ + 46, + 445, + 289, + 639 + ], + "type": "text", + "content": "sets. The Panoster [13] directly predicts the instance IDs from a learning-based clustering module, where the time-consuming DBSCAN [11] is used to refine the instance segmentation results. The DS-Net [15] proposes a learnable dynamic shifting module that adjusts the kernel functions of the MeanShift [9] to handle instances with various sizes. The GP-S3Net [29] constructs a graph convolutional network (GCN) on the over-segmentation clusters to identify instances. The SMAC-Seg [20] proposes a sparse multidirectional attention clustering and center-aware repel loss for instance segmentation. These methods adopt the time-consuming clustering methods. To further accelerate the algorithm, recently, Panoptic-PolarNet [39], SCAN [35], and Panoptic-PHNet [22] adopt the BEV center heatmap and center offsets for instance segmentation. However, the BEV center heatmap is also costly and confuses on " + }, + { + "bbox": [ + 46, + 445, + 289, + 639 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 46, + 445, + 289, + 639 + ], + "type": "text", + "content": "-axis." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 47, + 654, + 113, + 668 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 654, + 113, + 668 + ], + "spans": [ + { + "bbox": [ + 47, + 654, + 113, + 668 + ], + "type": "text", + "content": "3. Approach" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 47, + 677, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 677, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 47, + 677, + 287, + 715 + ], + "type": "text", + "content": "The input of the LiDAR panoptic segmentation task is the LiDAR point clouds, given by a set of unordered points " + }, + { + "bbox": [ + 47, + 677, + 287, + 715 + ], + "type": "inline_equation", + "content": "\\pmb{P} = \\{(p_i,f_i)\\}_{i = 0}^{N - 1}" + }, + { + "bbox": [ + 47, + 677, + 287, + 715 + ], + "type": "text", + "content": " (where " + }, + { + "bbox": [ + 47, + 677, + 287, + 715 + ], + "type": "inline_equation", + "content": "\\pmb {p}_i = (x,y,z)" + }, + { + "bbox": [ + 47, + 677, + 287, + 715 + ], + "type": "text", + "content": " is the 3D coor" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 445, + 545, + 542 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 445, + 545, + 542 + ], + "spans": [ + { + "bbox": [ + 304, + 445, + 545, + 542 + ], + "type": "text", + "content": "dinates in the cartesian space and " + }, + { + "bbox": [ + 304, + 445, + 545, + 542 + ], + "type": "inline_equation", + "content": "\\pmb{f}_i" + }, + { + "bbox": [ + 304, + 445, + 545, + 542 + ], + "type": "text", + "content": " denotes additional LiDAR point features, e.g. intensity). The task aims to assign a set of labels " + }, + { + "bbox": [ + 304, + 445, + 545, + 542 + ], + "type": "inline_equation", + "content": "\\pmb{L} = \\{(s_i, inst_i)\\}_{i=0}^{N-1}" + }, + { + "bbox": [ + 304, + 445, + 545, + 542 + ], + "type": "text", + "content": " to these points, where " + }, + { + "bbox": [ + 304, + 445, + 545, + 542 + ], + "type": "inline_equation", + "content": "s_i" + }, + { + "bbox": [ + 304, + 445, + 545, + 542 + ], + "type": "text", + "content": " is the semantic label (e.g. road, building, car, pedestrian), and " + }, + { + "bbox": [ + 304, + 445, + 545, + 542 + ], + "type": "inline_equation", + "content": "inst_i" + }, + { + "bbox": [ + 304, + 445, + 545, + 542 + ], + "type": "text", + "content": " is the instance ID of the " + }, + { + "bbox": [ + 304, + 445, + 545, + 542 + ], + "type": "inline_equation", + "content": "i^{th}" + }, + { + "bbox": [ + 304, + 445, + 545, + 542 + ], + "type": "text", + "content": " point. Moreover, " + }, + { + "bbox": [ + 304, + 445, + 545, + 542 + ], + "type": "inline_equation", + "content": "s_i" + }, + { + "bbox": [ + 304, + 445, + 545, + 542 + ], + "type": "text", + "content": " can be divided into the uncountable stuff classes (e.g. road, building) and countable things classes (e.g. car, pedestrian). The instance IDs of stuff points are set as 0." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 545, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 545, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 545, + 546, + 713 + ], + "type": "text", + "content": "To predict the labels " + }, + { + "bbox": [ + 304, + 545, + 546, + 713 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 304, + 545, + 546, + 713 + ], + "type": "text", + "content": " of the input LiDAR point clouds " + }, + { + "bbox": [ + 304, + 545, + 546, + 713 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 304, + 545, + 546, + 713 + ], + "type": "text", + "content": ", our CFNet decomposes this process into four steps, as shown in Fig. 3: 1) an off-the-shelf 2D projection-based backbone is applied to efficiently extract features on the 2D spaces (e.g. BEV, RV); 2) a novel center focusing feature encoding (CFFE) is used to generate center-focusing feature maps for more accurate predictions of the instance centers; 3) the panoptic segmentation head fuses the features from the 3D points and 2D spaces to predict the semantic segmentation results, the center offsets, and the confidence scores of the center offsets, respectively; 4) during inference, a post-process is conducted to produce the panoptic segmentation results, where a novel center dedduplication module (CDM) operates on the shifted things points to se" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "13427" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 72, + 545, + 223 + ], + "blocks": [ + { + "bbox": [ + 50, + 72, + 545, + 223 + ], + "lines": [ + { + "bbox": [ + 50, + 72, + 545, + 223 + ], + "spans": [ + { + "bbox": [ + 50, + 72, + 545, + 223 + ], + "type": "image", + "image_path": "8bdd3a6609487f1e67f8756fb55ea40ea595a770e1b9ae4f7b544fb54786ca76.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 232, + 548, + 266 + ], + "lines": [ + { + "bbox": [ + 46, + 232, + 548, + 266 + ], + "spans": [ + { + "bbox": [ + 46, + 232, + 548, + 266 + ], + "type": "text", + "content": "Figure 4. The proposed center focusing feature encoding (CFFE). The \"Conv\" represents a 2D convolution with " + }, + { + "bbox": [ + 46, + 232, + 548, + 266 + ], + "type": "inline_equation", + "content": "3 \\times 3" + }, + { + "bbox": [ + 46, + 232, + 548, + 266 + ], + "type": "text", + "content": " kernels, a batch normalization, and a ReLU layer. The details of the semantic branch and instance branch are shown in Fig. 3. The blue arrows are coordinate-related operations." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 286, + 214, + 298 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 286, + 214, + 298 + ], + "spans": [ + { + "bbox": [ + 47, + 286, + 214, + 298 + ], + "type": "text", + "content": "lect only one center for a single instance." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 298, + 287, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 298, + 287, + 335 + ], + "spans": [ + { + "bbox": [ + 46, + 298, + 287, + 335 + ], + "type": "text", + "content": "The design of the backbone in the first step is beyond the scope of the paper, and the last three mentioned steps are illustrated in detail in the following subsections." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 343, + 233, + 357 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 343, + 233, + 357 + ], + "spans": [ + { + "bbox": [ + 47, + 343, + 233, + 357 + ], + "type": "text", + "content": "3.1. Center Focusing Feature Encoding" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 362, + 287, + 470 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 362, + 287, + 470 + ], + "spans": [ + { + "bbox": [ + 46, + 362, + 287, + 470 + ], + "type": "text", + "content": "As mentioned above, the LiDAR points of an object are usually surface-aggregated, especially for car and truck categories, resulting in the fact that the center of an object is imaginary and does not exist in the LiDAR point clouds. To encode the features of the non-existent centers, a novel center focusing feature encoding (CFFE) is proposed, which takes the 2D features from the backbone and the 3D point coordinates as inputs and generates the enhanced center-focusing feature maps as shown in Fig. 3." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 471, + 287, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 471, + 287, + 506 + ], + "spans": [ + { + "bbox": [ + 46, + 471, + 287, + 506 + ], + "type": "text", + "content": "The CFFE module consists of three steps, including intermediate result prediction, center feature generation, and feature enhancement module, as shown in Fig. 4." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 507, + 288, + 615 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 507, + 288, + 615 + ], + "spans": [ + { + "bbox": [ + 46, + 507, + 288, + 615 + ], + "type": "text", + "content": "Intermediate result prediction. In this step, the CFFE predicts intermediate results (including the semantic segmentation, center offset, and its confidence scores) according to the 2D features " + }, + { + "bbox": [ + 46, + 507, + 288, + 615 + ], + "type": "inline_equation", + "content": "\\pmb{F}_{m}" + }, + { + "bbox": [ + 46, + 507, + 288, + 615 + ], + "type": "text", + "content": " and 3D point features " + }, + { + "bbox": [ + 46, + 507, + 288, + 615 + ], + "type": "inline_equation", + "content": "\\pmb{F}_{p}" + }, + { + "bbox": [ + 46, + 507, + 288, + 615 + ], + "type": "text", + "content": " for subsequent center feature simulation. Specifically, two convolution layers are applied on the 2D features " + }, + { + "bbox": [ + 46, + 507, + 288, + 615 + ], + "type": "inline_equation", + "content": "\\pmb{F}_{m}" + }, + { + "bbox": [ + 46, + 507, + 288, + 615 + ], + "type": "text", + "content": " independently, to generate semantic features " + }, + { + "bbox": [ + 46, + 507, + 288, + 615 + ], + "type": "inline_equation", + "content": "\\pmb{F}_{m}^{sem}" + }, + { + "bbox": [ + 46, + 507, + 288, + 615 + ], + "type": "text", + "content": " and instance features " + }, + { + "bbox": [ + 46, + 507, + 288, + 615 + ], + "type": "inline_equation", + "content": "\\pmb{F}_{m}^{ins}" + }, + { + "bbox": [ + 46, + 507, + 288, + 615 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 46, + 507, + 288, + 615 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 46, + 507, + 288, + 615 + ], + "type": "text", + "content": " is the specific 2D view, such as RV [10, 25, 34], BEV [19], and polar view [38].)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 114, + 625, + 287, + 639 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 114, + 625, + 287, + 639 + ], + "spans": [ + { + "bbox": [ + 114, + 625, + 287, + 639 + ], + "type": "interline_equation", + "content": "\\boldsymbol {F} _ {m} ^ {\\text {s e m}} = \\operatorname {C o n v} \\left(\\boldsymbol {F} _ {m}; \\boldsymbol {\\theta} _ {1}\\right), \\tag {1}", + "image_path": "fc3567029ed6674b929913e83ef93447fb9dc7ea3579bded1dc2a6d3d433a067.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 118, + 641, + 287, + 655 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 118, + 641, + 287, + 655 + ], + "spans": [ + { + "bbox": [ + 118, + 641, + 287, + 655 + ], + "type": "interline_equation", + "content": "\\boldsymbol {F} _ {m} ^ {\\text {i n s}} = \\operatorname {C o n v} \\left(\\boldsymbol {F} _ {m}; \\boldsymbol {\\theta} _ {2}\\right), \\tag {2}", + "image_path": "54b562f8cf2daac8fe0804547cb7c281bb57767acb209ce2ce8d79f74b3cea63.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "type": "inline_equation", + "content": "Conv" + }, + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "type": "text", + "content": " denotes sequential 2D convolution, batch normalization, and ReLU operations, and " + }, + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\theta_{1}" + }, + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\theta_{2}" + }, + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "type": "text", + "content": " are their learnable parameters. Then, the semantic branch generates per-point 3D semantic features " + }, + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "type": "inline_equation", + "content": "F_{p}^{sem}" + }, + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "type": "text", + "content": " by fusing the point" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 285, + 489, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 285, + 489, + 300 + ], + "spans": [ + { + "bbox": [ + 305, + 285, + 489, + 300 + ], + "type": "text", + "content": "features " + }, + { + "bbox": [ + 305, + 285, + 489, + 300 + ], + "type": "inline_equation", + "content": "F_{p}" + }, + { + "bbox": [ + 305, + 285, + 489, + 300 + ], + "type": "text", + "content": " and 2D semantic features " + }, + { + "bbox": [ + 305, + 285, + 489, + 300 + ], + "type": "inline_equation", + "content": "\\pmb{F}_{m}^{sem}" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 364, + 307, + 545, + 322 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 364, + 307, + 545, + 322 + ], + "spans": [ + { + "bbox": [ + 364, + 307, + 545, + 322 + ], + "type": "interline_equation", + "content": "\\boldsymbol {F} _ {p} ^ {\\text {s e m}} = \\operatorname {S e g} \\left(\\boldsymbol {F} _ {p}, \\boldsymbol {F} _ {m} ^ {\\text {s e m}}; \\boldsymbol {\\theta} _ {3}\\right), \\tag {3}", + "image_path": "a5e9a527e294e4ada0093e240ce01e948b6ac5f55770f1bfe30e0e68afcc3a0a.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 328, + 545, + 341 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 328, + 545, + 341 + ], + "spans": [ + { + "bbox": [ + 305, + 328, + 545, + 341 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 328, + 545, + 341 + ], + "type": "inline_equation", + "content": "Seg" + }, + { + "bbox": [ + 305, + 328, + 545, + 341 + ], + "type": "text", + "content": " is the semantic branch, and " + }, + { + "bbox": [ + 305, + 328, + 545, + 341 + ], + "type": "inline_equation", + "content": "\\theta_{3}" + }, + { + "bbox": [ + 305, + 328, + 545, + 341 + ], + "type": "text", + "content": " is the parameters." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 342, + 546, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 342, + 546, + 407 + ], + "spans": [ + { + "bbox": [ + 305, + 342, + 546, + 407 + ], + "type": "text", + "content": "Finally, the intermediate semantic results " + }, + { + "bbox": [ + 305, + 342, + 546, + 407 + ], + "type": "inline_equation", + "content": "\\hat{S}^{\\prime} = \\{s_{i}^{\\prime}\\}_{i=0}^{N-1}" + }, + { + "bbox": [ + 305, + 342, + 546, + 407 + ], + "type": "text", + "content": " is produced based on " + }, + { + "bbox": [ + 305, + 342, + 546, + 407 + ], + "type": "inline_equation", + "content": "F_{p}^{sem}" + }, + { + "bbox": [ + 305, + 342, + 546, + 407 + ], + "type": "text", + "content": ". The intermediate results of center offsets " + }, + { + "bbox": [ + 305, + 342, + 546, + 407 + ], + "type": "inline_equation", + "content": "\\hat{O}^{\\prime} = \\{o_{i}^{\\prime}\\}_{i=0}^{N-1}" + }, + { + "bbox": [ + 305, + 342, + 546, + 407 + ], + "type": "text", + "content": " and confidence scores " + }, + { + "bbox": [ + 305, + 342, + 546, + 407 + ], + "type": "inline_equation", + "content": "\\hat{C}^{\\prime} = \\{c_{i}^{\\prime}\\}_{i=0}^{N-1}" + }, + { + "bbox": [ + 305, + 342, + 546, + 407 + ], + "type": "text", + "content": " are predicted by the instance branch with the point features " + }, + { + "bbox": [ + 305, + 342, + 546, + 407 + ], + "type": "inline_equation", + "content": "F_{p}" + }, + { + "bbox": [ + 305, + 342, + 546, + 407 + ], + "type": "text", + "content": " and 2D instance features " + }, + { + "bbox": [ + 305, + 342, + 546, + 407 + ], + "type": "inline_equation", + "content": "F_{m}^{ins}" + }, + { + "bbox": [ + 305, + 342, + 546, + 407 + ], + "type": "text", + "content": "," + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 380, + 424, + 545, + 441 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 380, + 424, + 545, + 441 + ], + "spans": [ + { + "bbox": [ + 380, + 424, + 545, + 441 + ], + "type": "interline_equation", + "content": "\\hat {\\boldsymbol {S}} ^ {\\prime} = F C \\left(\\boldsymbol {F} _ {p} ^ {s e m}; \\boldsymbol {\\theta} _ {4}\\right), \\tag {4}", + "image_path": "363e5570f77f093667e404ddd09cb2c8b92f8ed8fb0d7b288738f10725e5e088.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 364, + 443, + 545, + 458 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 364, + 443, + 545, + 458 + ], + "spans": [ + { + "bbox": [ + 364, + 443, + 545, + 458 + ], + "type": "interline_equation", + "content": "\\hat {\\boldsymbol {O}} ^ {\\prime}, \\hat {\\boldsymbol {C}} ^ {\\prime} = I n s \\left(\\boldsymbol {F} _ {p}, \\boldsymbol {F} _ {m} ^ {i n s}; \\boldsymbol {\\theta} _ {5}\\right), \\tag {5}", + "image_path": "a9e3c1a6974f453b9b6cb17cfbee4416430408c84ce3fb9b61ecef91133dcbe8.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 466, + 545, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 466, + 545, + 525 + ], + "spans": [ + { + "bbox": [ + 304, + 466, + 545, + 525 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 466, + 545, + 525 + ], + "type": "inline_equation", + "content": "Ins" + }, + { + "bbox": [ + 304, + 466, + 545, + 525 + ], + "type": "text", + "content": " is the instance branch, and " + }, + { + "bbox": [ + 304, + 466, + 545, + 525 + ], + "type": "inline_equation", + "content": "FC" + }, + { + "bbox": [ + 304, + 466, + 545, + 525 + ], + "type": "text", + "content": " represents the fully-connected layer. The structure and the training objective of semantic and instance branches are the same as those in the panoptic segmentation head and are elucidated in Fig. 3 and section 3.2." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 526, + 545, + 598 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 526, + 545, + 598 + ], + "spans": [ + { + "bbox": [ + 304, + 526, + 545, + 598 + ], + "type": "text", + "content": "Center feature generation (CFG). In this step, the CFFE generates the shifted center features " + }, + { + "bbox": [ + 304, + 526, + 545, + 598 + ], + "type": "inline_equation", + "content": "\\pmb{F}_{p\\rightarrow m}^{sem}" + }, + { + "bbox": [ + 304, + 526, + 545, + 598 + ], + "type": "text", + "content": " by shifting the 3D semantic point features " + }, + { + "bbox": [ + 304, + 526, + 545, + 598 + ], + "type": "inline_equation", + "content": "\\pmb{F}_{p}^{sem}" + }, + { + "bbox": [ + 304, + 526, + 545, + 598 + ], + "type": "text", + "content": " to the predicted centers according to the mentioned intermediate results. First, the coordinates of a predicted center " + }, + { + "bbox": [ + 304, + 526, + 545, + 598 + ], + "type": "inline_equation", + "content": "\\pmb{p}_i^{shift}" + }, + { + "bbox": [ + 304, + 526, + 545, + 598 + ], + "type": "text", + "content": " are computed by," + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 361, + 606, + 545, + 622 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 361, + 606, + 545, + 622 + ], + "spans": [ + { + "bbox": [ + 361, + 606, + 545, + 622 + ], + "type": "interline_equation", + "content": "\\boldsymbol {p} _ {i} ^ {\\text {s h i f t}} = \\boldsymbol {p} _ {i} + \\hat {\\boldsymbol {o}} _ {i} ^ {\\prime} * \\mathbb {1} \\left[ \\hat {c} _ {i} ^ {\\prime} > \\delta \\right], \\tag {6}", + "image_path": "a0d3bff063544c2c08e97e9603a199810112ef24d74f40cc54cf9595d40a476d.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 305, + 629, + 545, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 629, + 545, + 677 + ], + "spans": [ + { + "bbox": [ + 305, + 629, + 545, + 677 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 629, + 545, + 677 + ], + "type": "inline_equation", + "content": "\\pmb{p}_i" + }, + { + "bbox": [ + 305, + 629, + 545, + 677 + ], + "type": "text", + "content": " is the original 3D LiDAR point coordinates, and " + }, + { + "bbox": [ + 305, + 629, + 545, + 677 + ], + "type": "inline_equation", + "content": "\\mathbb{1}[\\hat{c}_i^{\\prime} > \\delta ]" + }, + { + "bbox": [ + 305, + 629, + 545, + 677 + ], + "type": "text", + "content": " is a binary indicator of whether the confidence " + }, + { + "bbox": [ + 305, + 629, + 545, + 677 + ], + "type": "inline_equation", + "content": "\\hat{c}_i^\\prime" + }, + { + "bbox": [ + 305, + 629, + 545, + 677 + ], + "type": "text", + "content": " is greater than " + }, + { + "bbox": [ + 305, + 629, + 545, + 677 + ], + "type": "inline_equation", + "content": "\\delta = 0.2" + }, + { + "bbox": [ + 305, + 629, + 545, + 677 + ], + "type": "text", + "content": ". In other words, it does not shift the stuff points or the things points with low confidence scores." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "content": "Then, the shifted 3D points " + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\{\\pmb{p}_i^{shift}\\}_{i=0}^{N-1}" + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "content": " are served as the new coordinates of the feature points, and the Point to Grid (P2G) operation in [23] re-projects the 3D semantic" + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "13428" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 286, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 286, + 96 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 286, + 96 + ], + "type": "text", + "content": "features " + }, + { + "bbox": [ + 47, + 72, + 286, + 96 + ], + "type": "inline_equation", + "content": "\\pmb{F}_{p}^{sem}" + }, + { + "bbox": [ + 47, + 72, + 286, + 96 + ], + "type": "text", + "content": " onto the 2D projected feature maps " + }, + { + "bbox": [ + 47, + 72, + 286, + 96 + ], + "type": "inline_equation", + "content": "\\pmb{F}_{p\\rightarrow m}^{sem}" + }, + { + "bbox": [ + 47, + 72, + 286, + 96 + ], + "type": "text", + "content": " with this new coordinates," + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 88, + 102, + 286, + 119 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 88, + 102, + 286, + 119 + ], + "spans": [ + { + "bbox": [ + 88, + 102, + 286, + 119 + ], + "type": "interline_equation", + "content": "\\boldsymbol {F} _ {p \\rightarrow m} ^ {s e m} = P 2 G (\\boldsymbol {F} _ {p} ^ {s e m}; \\{\\boldsymbol {p} _ {i} ^ {\\text {s h i f t}} \\} _ {i = 0} ^ {N - 1}). \\tag {7}", + "image_path": "878480705edf58da89f952f919d767c6bb1ae3beec47937fe1d20de2d481897e.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 124, + 286, + 159 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 124, + 286, + 159 + ], + "spans": [ + { + "bbox": [ + 46, + 124, + 286, + 159 + ], + "type": "text", + "content": "Compared with " + }, + { + "bbox": [ + 46, + 124, + 286, + 159 + ], + "type": "inline_equation", + "content": "F_{p}^{sem}" + }, + { + "bbox": [ + 46, + 124, + 286, + 159 + ], + "type": "text", + "content": ", the feature maps " + }, + { + "bbox": [ + 46, + 124, + 286, + 159 + ], + "type": "inline_equation", + "content": "F_{p\\rightarrow m}^{sem}" + }, + { + "bbox": [ + 46, + 124, + 286, + 159 + ], + "type": "text", + "content": " pay more attention to the imaginary centers since most of the things points have been shifted to their predicted centers." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 160, + 286, + 270 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 160, + 286, + 270 + ], + "spans": [ + { + "bbox": [ + 46, + 160, + 286, + 270 + ], + "type": "text", + "content": "Feature enhancement module (FEM). The CFFE finally fuses the semantic feature maps " + }, + { + "bbox": [ + 46, + 160, + 286, + 270 + ], + "type": "inline_equation", + "content": "\\pmb{F}_{m}^{sem}" + }, + { + "bbox": [ + 46, + 160, + 286, + 270 + ], + "type": "text", + "content": " and the reprojected shifted center feature maps " + }, + { + "bbox": [ + 46, + 160, + 286, + 270 + ], + "type": "inline_equation", + "content": "\\pmb{F}_{p\\rightarrow m}^{sem}" + }, + { + "bbox": [ + 46, + 160, + 286, + 270 + ], + "type": "text", + "content": " to generate center-focusing semantic feature maps " + }, + { + "bbox": [ + 46, + 160, + 286, + 270 + ], + "type": "inline_equation", + "content": "\\pmb{F}_{m}^{CFsem}" + }, + { + "bbox": [ + 46, + 160, + 286, + 270 + ], + "type": "text", + "content": " and instance feature maps " + }, + { + "bbox": [ + 46, + 160, + 286, + 270 + ], + "type": "inline_equation", + "content": "\\pmb{F}_{m}^{CFS}" + }, + { + "bbox": [ + 46, + 160, + 286, + 270 + ], + "type": "text", + "content": ", which are used by the subsequent semantic and instance branches for more accurate predictions. The enhancement module is composed of a simple concatenation operation and several convolution layers, whose details are in the supplementary material." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 270, + 286, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 270, + 286, + 354 + ], + "spans": [ + { + "bbox": [ + 47, + 270, + 286, + 354 + ], + "type": "text", + "content": "In addition, for application of the multi-view fusion backbone (e.g. [23] shown in the supplementary material), the feature maps " + }, + { + "bbox": [ + 47, + 270, + 286, + 354 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_m" + }, + { + "bbox": [ + 47, + 270, + 286, + 354 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 47, + 270, + 286, + 354 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_m^{sem}" + }, + { + "bbox": [ + 47, + 270, + 286, + 354 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 47, + 270, + 286, + 354 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_m^{ins}" + }, + { + "bbox": [ + 47, + 270, + 286, + 354 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 47, + 270, + 286, + 354 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{p\\rightarrow m}^{sem}" + }, + { + "bbox": [ + 47, + 270, + 286, + 354 + ], + "type": "text", + "content": " (e.g. " + }, + { + "bbox": [ + 47, + 270, + 286, + 354 + ], + "type": "inline_equation", + "content": "m \\in \\{\\mathrm{RV},\\mathrm{BEV}\\}" + }, + { + "bbox": [ + 47, + 270, + 286, + 354 + ], + "type": "text", + "content": ") of each view are computed independently according to the above process. Then, they are fused in the point fusion (PF) module [23] to generate integrated 3D point features and per-point prediction." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 360, + 205, + 373 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 360, + 205, + 373 + ], + "spans": [ + { + "bbox": [ + 47, + 360, + 205, + 373 + ], + "type": "text", + "content": "3.2. Panoptic Segmentation Head" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 379, + 286, + 474 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 379, + 286, + 474 + ], + "spans": [ + { + "bbox": [ + 46, + 379, + 286, + 474 + ], + "type": "text", + "content": "For better modeling the instance center, the panoptic segmentation head uses a semantic branch to predict the semantic segmentation, and an instance branch to estimate both the center offsets and the newly introduced confidence scores, given the center-focusing semantic and instance feature maps. The architectures of both branches follow the segmentation head in [23] and are briefly introduced, while their training objectives are elaborated on." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 475, + 286, + 630 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 475, + 286, + 630 + ], + "spans": [ + { + "bbox": [ + 46, + 475, + 286, + 630 + ], + "type": "text", + "content": "Semantic branch. For a per-point prediction, the semantic branch first applies the Grid to Point (G2P) operations to acquire the 3D point-wise features from the 2D semantic feature maps. Then, a PF module fuses the pointwise features from the G2P operation and the original 3D points to generate point-wise semantic features. The details of the G2P and PF can be found in [23]. After obtaining the point-wise semantic features, a fully connected (FC) layer is used to predict the final per-point semantic result " + }, + { + "bbox": [ + 46, + 475, + 286, + 630 + ], + "type": "inline_equation", + "content": "\\hat{\\psi}_i^j" + }, + { + "bbox": [ + 46, + 475, + 286, + 630 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 46, + 475, + 286, + 630 + ], + "type": "inline_equation", + "content": "\\hat{\\psi}_i^j \\in [0,1]" + }, + { + "bbox": [ + 46, + 475, + 286, + 630 + ], + "type": "text", + "content": "). " + }, + { + "bbox": [ + 46, + 475, + 286, + 630 + ], + "type": "inline_equation", + "content": "\\hat{\\psi}_i^j" + }, + { + "bbox": [ + 46, + 475, + 286, + 630 + ], + "type": "text", + "content": " represents the probability of the " + }, + { + "bbox": [ + 46, + 475, + 286, + 630 + ], + "type": "inline_equation", + "content": "i^{th}" + }, + { + "bbox": [ + 46, + 475, + 286, + 630 + ], + "type": "text", + "content": " LiDAR point belonging to the " + }, + { + "bbox": [ + 46, + 475, + 286, + 630 + ], + "type": "inline_equation", + "content": "j^{th}" + }, + { + "bbox": [ + 46, + 475, + 286, + 630 + ], + "type": "text", + "content": " class. The predicted semantic label " + }, + { + "bbox": [ + 46, + 475, + 286, + 630 + ], + "type": "inline_equation", + "content": "\\hat{s}_i" + }, + { + "bbox": [ + 46, + 475, + 286, + 630 + ], + "type": "text", + "content": " is acquired by selecting the most probable class " + }, + { + "bbox": [ + 46, + 475, + 286, + 630 + ], + "type": "inline_equation", + "content": "\\hat{s}_i = \\arg \\max_j \\hat{\\psi}_i^j" + }, + { + "bbox": [ + 46, + 475, + 286, + 630 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 631, + 286, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 631, + 286, + 677 + ], + "spans": [ + { + "bbox": [ + 47, + 631, + 286, + 677 + ], + "type": "text", + "content": "Referring to the CPGNet [23], the same loss functions are adopted, including weighted cross entropy (WCE) loss " + }, + { + "bbox": [ + 47, + 631, + 286, + 677 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{wce}" + }, + { + "bbox": [ + 47, + 631, + 286, + 677 + ], + "type": "text", + "content": ", Lovász-Softmax loss " + }, + { + "bbox": [ + 47, + 631, + 286, + 677 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{ls}" + }, + { + "bbox": [ + 47, + 631, + 286, + 677 + ], + "type": "text", + "content": " [4], and transformation consistency loss " + }, + { + "bbox": [ + 47, + 631, + 286, + 677 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{tc}" + }, + { + "bbox": [ + 47, + 631, + 286, + 677 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 677, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 677, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 677, + 286, + 713 + ], + "type": "text", + "content": "Instance branch. Similar to the semantic branch, the instance branch also adopts the G2P operation and a PF module to acquire point-wise instance features. An FC layer" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 72, + 545, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 108 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 108 + ], + "type": "text", + "content": "is applied to predict the per-point center offsets " + }, + { + "bbox": [ + 304, + 72, + 545, + 108 + ], + "type": "inline_equation", + "content": "\\hat{o}_i" + }, + { + "bbox": [ + 304, + 72, + 545, + 108 + ], + "type": "text", + "content": ". The ground-truth for " + }, + { + "bbox": [ + 304, + 72, + 545, + 108 + ], + "type": "inline_equation", + "content": "\\hat{o}_i" + }, + { + "bbox": [ + 304, + 72, + 545, + 108 + ], + "type": "text", + "content": " is the offset vector " + }, + { + "bbox": [ + 304, + 72, + 545, + 108 + ], + "type": "inline_equation", + "content": "e_i - p_i" + }, + { + "bbox": [ + 304, + 72, + 545, + 108 + ], + "type": "text", + "content": " from the " + }, + { + "bbox": [ + 304, + 72, + 545, + 108 + ], + "type": "inline_equation", + "content": "i^{th}" + }, + { + "bbox": [ + 304, + 72, + 545, + 108 + ], + "type": "text", + "content": " point " + }, + { + "bbox": [ + 304, + 72, + 545, + 108 + ], + "type": "inline_equation", + "content": "p_i" + }, + { + "bbox": [ + 304, + 72, + 545, + 108 + ], + "type": "text", + "content": " to its corresponding instance center " + }, + { + "bbox": [ + 304, + 72, + 545, + 108 + ], + "type": "inline_equation", + "content": "e_i" + }, + { + "bbox": [ + 304, + 72, + 545, + 108 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 108, + 545, + 143 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 108, + 545, + 143 + ], + "spans": [ + { + "bbox": [ + 304, + 108, + 545, + 143 + ], + "type": "text", + "content": "For center offset regression, the loss function for optimizing " + }, + { + "bbox": [ + 304, + 108, + 545, + 143 + ], + "type": "inline_equation", + "content": "\\hat{o}_i" + }, + { + "bbox": [ + 304, + 108, + 545, + 143 + ], + "type": "text", + "content": " only considers the things classes and is formulated as follows," + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 151, + 545, + 183 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 151, + 545, + 183 + ], + "spans": [ + { + "bbox": [ + 313, + 151, + 545, + 183 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\boldsymbol {o} _ {i}} = \\left\\{ \\begin{array}{l l} | | \\hat {\\boldsymbol {o}} _ {i} - (\\boldsymbol {e} _ {i} - \\boldsymbol {p} _ {i}) | | _ {2} & \\text {i f} s _ {i} \\in t h i n g s \\text {c l a s s e s}, \\\\ 0 & \\text {o t h e r w i s e}, \\end{array} \\right. \\tag {8}", + "image_path": "3d94e2c1201dcd21949b28a74ae40cb936fa9e3c50b86eb575ee305b56277d3d.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 190, + 544, + 214 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 190, + 544, + 214 + ], + "spans": [ + { + "bbox": [ + 304, + 190, + 544, + 214 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 190, + 544, + 214 + ], + "type": "inline_equation", + "content": "\\pmb{e}_i" + }, + { + "bbox": [ + 304, + 190, + 544, + 214 + ], + "type": "text", + "content": " is the axis-aligned center of the instance [15]. Then, the losses are summarized as the following," + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 375, + 222, + 545, + 255 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 375, + 222, + 545, + 255 + ], + "spans": [ + { + "bbox": [ + 375, + 222, + 545, + 255 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {o} = \\frac {1}{N ^ {\\text {t h i n g s}}} \\sum_ {i = 0} ^ {N - 1} \\mathcal {L} _ {\\boldsymbol {o} _ {i}}, \\tag {9}", + "image_path": "df620564561afd8e37536351ff9433e4a152302aab825a6e7050483db34eb7d5.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 263, + 544, + 287 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 263, + 544, + 287 + ], + "spans": [ + { + "bbox": [ + 304, + 263, + 544, + 287 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 263, + 544, + 287 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 263, + 544, + 287 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 263, + 544, + 287 + ], + "type": "inline_equation", + "content": "N^{things}" + }, + { + "bbox": [ + 304, + 263, + 544, + 287 + ], + "type": "text", + "content": " are the numbers of all points and things points, respectively." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 288, + 545, + 347 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 288, + 545, + 347 + ], + "spans": [ + { + "bbox": [ + 304, + 288, + 545, + 347 + ], + "type": "text", + "content": "For confidence score regression, another FC layer is used to predict the per-point confidence score " + }, + { + "bbox": [ + 304, + 288, + 545, + 347 + ], + "type": "inline_equation", + "content": "\\hat{c}_i" + }, + { + "bbox": [ + 304, + 288, + 545, + 347 + ], + "type": "text", + "content": " to indicate the accuracy degree of " + }, + { + "bbox": [ + 304, + 288, + 545, + 347 + ], + "type": "inline_equation", + "content": "\\hat{o}_i" + }, + { + "bbox": [ + 304, + 288, + 545, + 347 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 304, + 288, + 545, + 347 + ], + "type": "inline_equation", + "content": "\\hat{c}_i" + }, + { + "bbox": [ + 304, + 288, + 545, + 347 + ], + "type": "text", + "content": " is activated by a sigmoid function to ensure " + }, + { + "bbox": [ + 304, + 288, + 545, + 347 + ], + "type": "inline_equation", + "content": "\\hat{c}_i \\in [0,1]" + }, + { + "bbox": [ + 304, + 288, + 545, + 347 + ], + "type": "text", + "content": ". The ground-truth label " + }, + { + "bbox": [ + 304, + 288, + 545, + 347 + ], + "type": "inline_equation", + "content": "c_i" + }, + { + "bbox": [ + 304, + 288, + 545, + 347 + ], + "type": "text", + "content": " for supervising " + }, + { + "bbox": [ + 304, + 288, + 545, + 347 + ], + "type": "inline_equation", + "content": "\\hat{c}_i" + }, + { + "bbox": [ + 304, + 288, + 545, + 347 + ], + "type": "text", + "content": " is generated by" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 329, + 355, + 545, + 389 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 329, + 355, + 545, + 389 + ], + "spans": [ + { + "bbox": [ + 329, + 355, + 545, + 389 + ], + "type": "interline_equation", + "content": "c _ {i} = \\left\\{ \\begin{array}{l l} \\exp \\left(- \\frac {\\mathcal {L} _ {o _ {i}} ^ {2}}{2 \\sigma^ {2}}\\right) & \\text {i f} s _ {i} \\in \\text {t h i n g s c l a s s e s}, \\\\ 0 & \\text {o t h e r w i s e .} \\end{array} \\right. \\tag {10}", + "image_path": "34b32742c8efebfb61fea3d9035aeea0dce908ee344e93bd7f18aa203cd552ef.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 396, + 545, + 431 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 396, + 545, + 431 + ], + "spans": [ + { + "bbox": [ + 304, + 396, + 545, + 431 + ], + "type": "text", + "content": "For the things points, the lower " + }, + { + "bbox": [ + 304, + 396, + 545, + 431 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{o_i}" + }, + { + "bbox": [ + 304, + 396, + 545, + 431 + ], + "type": "text", + "content": " is, the higher " + }, + { + "bbox": [ + 304, + 396, + 545, + 431 + ], + "type": "inline_equation", + "content": "c_{i}" + }, + { + "bbox": [ + 304, + 396, + 545, + 431 + ], + "type": "text", + "content": " is. It means that the point with a more accurate regression of the center offset has a higher confidence score." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 432, + 544, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 432, + 544, + 445 + ], + "spans": [ + { + "bbox": [ + 317, + 432, + 544, + 445 + ], + "type": "text", + "content": "The weighted binary cross entropy loss " + }, + { + "bbox": [ + 317, + 432, + 544, + 445 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{wbce}" + }, + { + "bbox": [ + 317, + 432, + 544, + 445 + ], + "type": "text", + "content": " is applied," + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 310, + 462, + 545, + 495 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 310, + 462, + 545, + 495 + ], + "spans": [ + { + "bbox": [ + 310, + 462, + 545, + 495 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {w b c e} = - \\frac {1}{N} \\sum_ {i = 0} ^ {N - 1} 6 c _ {i} \\log (\\hat {c} _ {i}) + (1 - c _ {i}) \\log (1 - \\hat {c} _ {i}), \\tag {11}", + "image_path": "438c1a9ea41aafdd2c863e066c36fa8fbd25b40495b038400d0a809058f2b29c.jpg" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 502, + 544, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 502, + 544, + 525 + ], + "spans": [ + { + "bbox": [ + 304, + 502, + 544, + 525 + ], + "type": "text", + "content": "where the things points are manually emphasized since the number of them is much less than that of the stuff points." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 304, + 526, + 544, + 549 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 526, + 544, + 549 + ], + "spans": [ + { + "bbox": [ + 304, + 526, + 544, + 549 + ], + "type": "text", + "content": "Finally, the loss for each group of results (from the CFNet or CFFE) is defined as" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 341, + 559, + 545, + 571 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 341, + 559, + 545, + 571 + ], + "spans": [ + { + "bbox": [ + 341, + 559, + 545, + 571 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\mathcal {L} _ {w c e} + 3 \\mathcal {L} _ {l s} + \\mathcal {L} _ {t c} + 2 \\mathcal {L} _ {o} + \\mathcal {L} _ {w b c e}. \\tag {12}", + "image_path": "9f1d4d641737d689a3e46226767f8e7cfda43755fb6290b5abcca57247d4552d.jpg" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 304, + 580, + 544, + 603 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 580, + 544, + 603 + ], + "spans": [ + { + "bbox": [ + 304, + 580, + 544, + 603 + ], + "type": "text", + "content": "The total loss is the sum of the two losses sourced from the CFNet and CFFE." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 306, + 611, + 468, + 624 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 611, + 468, + 624 + ], + "spans": [ + { + "bbox": [ + 306, + 611, + 468, + 624 + ], + "type": "text", + "content": "3.3. Center Dedduplication Module" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 304, + 630, + 545, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 630, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 304, + 630, + 545, + 689 + ], + "type": "text", + "content": "Given the final predictions of the semantic segmentation " + }, + { + "bbox": [ + 304, + 630, + 545, + 689 + ], + "type": "inline_equation", + "content": "\\hat{S}" + }, + { + "bbox": [ + 304, + 630, + 545, + 689 + ], + "type": "text", + "content": ", center offsets " + }, + { + "bbox": [ + 304, + 630, + 545, + 689 + ], + "type": "inline_equation", + "content": "\\hat{O}" + }, + { + "bbox": [ + 304, + 630, + 545, + 689 + ], + "type": "text", + "content": ", and confidence scores " + }, + { + "bbox": [ + 304, + 630, + 545, + 689 + ], + "type": "inline_equation", + "content": "\\hat{C}" + }, + { + "bbox": [ + 304, + 630, + 545, + 689 + ], + "type": "text", + "content": ", this section introduces how to exploit the detected centers to acquire the panoptic segmentation results during inference, and illustrates the key module, center dedduplication module (CDM)." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 304, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 689, + 545, + 713 + ], + "type": "text", + "content": "Post-process. For the panoptic segmentation results, instance segmentation is first generated and then the final" + } + ] + } + ], + "index": 27 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13429" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "code", + "bbox": [ + 50, + 89, + 287, + 339 + ], + "blocks": [ + { + "bbox": [ + 53, + 75, + 230, + 87 + ], + "lines": [ + { + "bbox": [ + 53, + 75, + 230, + 87 + ], + "spans": [ + { + "bbox": [ + 53, + 75, + 230, + 87 + ], + "type": "text", + "content": "Algorithm 1: Center dedduplication module" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "code_caption" + }, + { + "bbox": [ + 50, + 89, + 287, + 339 + ], + "lines": [ + { + "bbox": [ + 50, + 89, + 287, + 339 + ], + "spans": [ + { + "bbox": [ + 50, + 89, + 287, + 339 + ], + "type": "text", + "content": "Input: Shifted things points " + }, + { + "bbox": [ + 50, + 89, + 287, + 339 + ], + "type": "inline_equation", + "content": "\\{\\tilde{p}_t^{shift}\\in \\mathbb{R}^3\\}_{t = 0}^{M - 1}" + }, + { + "bbox": [ + 50, + 89, + 287, + 339 + ], + "type": "text", + "content": " confidence scores " + }, + { + "bbox": [ + 50, + 89, + 287, + 339 + ], + "type": "inline_equation", + "content": "\\{\\tilde{c}_t\\in [0,1]\\}_{t = 0}^{M - 1}" + }, + { + "bbox": [ + 50, + 89, + 287, + 339 + ], + "type": "text", + "content": " distance threshold " + }, + { + "bbox": [ + 50, + 89, + 287, + 339 + ], + "type": "inline_equation", + "content": "d\\in \\mathbb{R}" + }, + { + "bbox": [ + 50, + 89, + 287, + 339 + ], + "type": "text", + "content": " Output: Instance centers " + }, + { + "bbox": [ + 50, + 89, + 287, + 339 + ], + "type": "inline_equation", + "content": "\\{\\tilde{e}_j\\in \\mathbb{R}^3\\}_{j = 0}^{D - 1}" + }, + { + "bbox": [ + 50, + 89, + 287, + 339 + ], + "type": "text", + "content": " \n1 order " + }, + { + "bbox": [ + 50, + 89, + 287, + 339 + ], + "type": "inline_equation", + "content": "=" + }, + { + "bbox": [ + 50, + 89, + 287, + 339 + ], + "type": "text", + "content": " sort_index( " + }, + { + "bbox": [ + 50, + 89, + 287, + 339 + ], + "type": "inline_equation", + "content": "\\{\\tilde{c}_t\\}_{t = 0}^{M - 1}" + }, + { + "bbox": [ + 50, + 89, + 287, + 339 + ], + "type": "text", + "content": " , decending=true) \n2 initialize a boolean vector " + }, + { + "bbox": [ + 50, + 89, + 287, + 339 + ], + "type": "inline_equation", + "content": "rej\\in \\mathbb{B}^{M}" + }, + { + "bbox": [ + 50, + 89, + 287, + 339 + ], + "type": "text", + "content": " with all elements of false and an empty list keep \n3 for " + }, + { + "bbox": [ + 50, + 89, + 287, + 339 + ], + "type": "inline_equation", + "content": "k\\gets 0" + }, + { + "bbox": [ + 50, + 89, + 287, + 339 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 50, + 89, + 287, + 339 + ], + "type": "inline_equation", + "content": "M - 1" + }, + { + "bbox": [ + 50, + 89, + 287, + 339 + ], + "type": "text", + "content": " do \n4 if not rej[order[k]] then \n5 keep.append(order[k]) \n6 for " + }, + { + "bbox": [ + 50, + 89, + 287, + 339 + ], + "type": "inline_equation", + "content": "h\\gets k + 1" + }, + { + "bbox": [ + 50, + 89, + 287, + 339 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 50, + 89, + 287, + 339 + ], + "type": "inline_equation", + "content": "M - 1" + }, + { + "bbox": [ + 50, + 89, + 287, + 339 + ], + "type": "text", + "content": " do \n7 dist " + }, + { + "bbox": [ + 50, + 89, + 287, + 339 + ], + "type": "inline_equation", + "content": "= ||\\tilde{p}_{order[k]}^{shift} - \\tilde{p}_{order[h]}^{shift}||_2" + }, + { + "bbox": [ + 50, + 89, + 287, + 339 + ], + "type": "text", + "content": " \n8 if dist " + }, + { + "bbox": [ + 50, + 89, + 287, + 339 + ], + "type": "inline_equation", + "content": "< d" + }, + { + "bbox": [ + 50, + 89, + 287, + 339 + ], + "type": "text", + "content": " then \n9 | rej[order[h]] " + }, + { + "bbox": [ + 50, + 89, + 287, + 339 + ], + "type": "inline_equation", + "content": "=" + }, + { + "bbox": [ + 50, + 89, + 287, + 339 + ], + "type": "text", + "content": " true \n10 end \n11 end \n12 end \n13 end \n14 " + }, + { + "bbox": [ + 50, + 89, + 287, + 339 + ], + "type": "inline_equation", + "content": "\\{\\tilde{e}_j\\} = \\tilde{p}_{keep}^{shift}" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "code_body" + } + ], + "index": 1, + "sub_type": "algorithm" + }, + { + "bbox": [ + 46, + 364, + 287, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 364, + 287, + 399 + ], + "spans": [ + { + "bbox": [ + 46, + 364, + 287, + 399 + ], + "type": "text", + "content": "panoptic segmentation is acquired by fusing the semantic and instance segmentation labels. There are five steps to achieve the final panoptic segmentation:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 399, + 287, + 593 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 47, + 399, + 287, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 399, + 287, + 449 + ], + "spans": [ + { + "bbox": [ + 47, + 399, + 287, + 449 + ], + "type": "text", + "content": "1) The things points " + }, + { + "bbox": [ + 47, + 399, + 287, + 449 + ], + "type": "inline_equation", + "content": "\\{\\tilde{p}_t\\}_{t=0}^{M-1} = \\{\\pmb{p}_i | \\hat{s}_i \\in" + }, + { + "bbox": [ + 47, + 399, + 287, + 449 + ], + "type": "text", + "content": " things classes} are selected according to the predicted semantic labels, with their offsets " + }, + { + "bbox": [ + 47, + 399, + 287, + 449 + ], + "type": "inline_equation", + "content": "\\{\\tilde{o}_t\\}_{t=0}^{M-1}" + }, + { + "bbox": [ + 47, + 399, + 287, + 449 + ], + "type": "text", + "content": " and confidence scores " + }, + { + "bbox": [ + 47, + 399, + 287, + 449 + ], + "type": "inline_equation", + "content": "\\{\\tilde{c}_t\\}_{t=0}^{M-1}" + }, + { + "bbox": [ + 47, + 399, + 287, + 449 + ], + "type": "text", + "content": " (where " + }, + { + "bbox": [ + 47, + 399, + 287, + 449 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 47, + 399, + 287, + 449 + ], + "type": "text", + "content": " is the number of things points)." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 448, + 287, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 448, + 287, + 472 + ], + "spans": [ + { + "bbox": [ + 47, + 448, + 287, + 472 + ], + "type": "text", + "content": "2) Each shifted things point " + }, + { + "bbox": [ + 47, + 448, + 287, + 472 + ], + "type": "inline_equation", + "content": "\\tilde{\\pmb{p}}_t^{shift}" + }, + { + "bbox": [ + 47, + 448, + 287, + 472 + ], + "type": "text", + "content": " is given by " + }, + { + "bbox": [ + 47, + 448, + 287, + 472 + ], + "type": "inline_equation", + "content": "\\tilde{\\pmb{p}}_t^{shift} = \\tilde{\\pmb{p}}_t + \\tilde{\\pmb{o}}_t" + }, + { + "bbox": [ + 47, + 448, + 287, + 472 + ], + "type": "text", + "content": " and is severing as an instance center candidate." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 472, + 287, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 472, + 287, + 510 + ], + "spans": [ + { + "bbox": [ + 47, + 472, + 287, + 510 + ], + "type": "text", + "content": "3) A center " + }, + { + "bbox": [ + 47, + 472, + 287, + 510 + ], + "type": "inline_equation", + "content": "\\tilde{\\pmb{e}}_j" + }, + { + "bbox": [ + 47, + 472, + 287, + 510 + ], + "type": "text", + "content": " for each instance " + }, + { + "bbox": [ + 47, + 472, + 287, + 510 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 47, + 472, + 287, + 510 + ], + "type": "text", + "content": " is selected by the CDM based on the coordinates " + }, + { + "bbox": [ + 47, + 472, + 287, + 510 + ], + "type": "inline_equation", + "content": "\\tilde{\\pmb{p}}_t^{shift}" + }, + { + "bbox": [ + 47, + 472, + 287, + 510 + ], + "type": "text", + "content": " and its confidence score " + }, + { + "bbox": [ + 47, + 472, + 287, + 510 + ], + "type": "inline_equation", + "content": "\\tilde{c}_t" + }, + { + "bbox": [ + 47, + 472, + 287, + 510 + ], + "type": "text", + "content": ", while other candidates are suppressed." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 510, + 287, + 546 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 510, + 287, + 546 + ], + "spans": [ + { + "bbox": [ + 47, + 510, + 287, + 546 + ], + "type": "text", + "content": "4) The instance " + }, + { + "bbox": [ + 47, + 510, + 287, + 546 + ], + "type": "inline_equation", + "content": "\\mathrm{ID}\\widetilde{\\text{in}}\\widetilde{\\text{st}}_t" + }, + { + "bbox": [ + 47, + 510, + 287, + 546 + ], + "type": "text", + "content": " is acquired by assigning the shifted things points " + }, + { + "bbox": [ + 47, + 510, + 287, + 546 + ], + "type": "inline_equation", + "content": "\\tilde{\\pmb{p}}_t^{\\text{shift}}" + }, + { + "bbox": [ + 47, + 510, + 287, + 546 + ], + "type": "text", + "content": " to the closest center in all centers " + }, + { + "bbox": [ + 47, + 510, + 287, + 546 + ], + "type": "inline_equation", + "content": "\\{\\tilde{\\pmb{e}}_j\\}_{j=0}^{D-1}" + }, + { + "bbox": [ + 47, + 510, + 287, + 546 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 47, + 510, + 287, + 546 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 47, + 510, + 287, + 546 + ], + "type": "text", + "content": " is the number of detected instances)." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 546, + 287, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 546, + 287, + 593 + ], + "spans": [ + { + "bbox": [ + 46, + 546, + 287, + 593 + ], + "type": "text", + "content": "5) The majority voting [15, 39] reassigns the most frequent semantic label to all points of a predicted instance to further guarantee the consistency of semantic labels within a predicted instance." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 46, + 594, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 288, + 713 + ], + "type": "text", + "content": "Center dedduplication module (CDM). The CDM takes the shifted points " + }, + { + "bbox": [ + 46, + 594, + 288, + 713 + ], + "type": "inline_equation", + "content": "\\{\\tilde{p}_t^{shift}\\}_{t=0}^{M-1}" + }, + { + "bbox": [ + 46, + 594, + 288, + 713 + ], + "type": "text", + "content": " and confidence scores " + }, + { + "bbox": [ + 46, + 594, + 288, + 713 + ], + "type": "inline_equation", + "content": "\\{\\tilde{c}_t\\}_{t=0}^{M-1}" + }, + { + "bbox": [ + 46, + 594, + 288, + 713 + ], + "type": "text", + "content": " as input to get an instance center " + }, + { + "bbox": [ + 46, + 594, + 288, + 713 + ], + "type": "inline_equation", + "content": "\\hat{e}_j" + }, + { + "bbox": [ + 46, + 594, + 288, + 713 + ], + "type": "text", + "content": " for each instance " + }, + { + "bbox": [ + 46, + 594, + 288, + 713 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 46, + 594, + 288, + 713 + ], + "type": "text", + "content": ". Inspired by the bounding box NMS, our CDM suppresses the candidate centers with lower scores within a euclidean distance threshold " + }, + { + "bbox": [ + 46, + 594, + 288, + 713 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 46, + 594, + 288, + 713 + ], + "type": "text", + "content": ". The pseudo-code of our CDM is shown in Algorithm 1, where two centers with a distance less than " + }, + { + "bbox": [ + 46, + 594, + 288, + 713 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 46, + 594, + 288, + 713 + ], + "type": "text", + "content": " are considered as the same instance. The process of our CDM is simple and can be easily implemented in CUDA." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 71, + 388, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 71, + 388, + 85 + ], + "spans": [ + { + "bbox": [ + 306, + 71, + 388, + 85 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 91, + 547, + 127 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 91, + 547, + 127 + ], + "spans": [ + { + "bbox": [ + 304, + 91, + 547, + 127 + ], + "type": "text", + "content": "The proposed CFNet is compared with the existing methods on the public SemanticKITTI [2] and nuScenes [5] panoptic segmentation benchmarks." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 133, + 422, + 146 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 133, + 422, + 146 + ], + "spans": [ + { + "bbox": [ + 305, + 133, + 422, + 146 + ], + "type": "text", + "content": "4.1. Experimental Setup" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 151, + 545, + 270 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 151, + 545, + 270 + ], + "spans": [ + { + "bbox": [ + 304, + 151, + 545, + 270 + ], + "type": "text", + "content": "Datasets. The SemanticKITTI [2] consists of 22 sequences collected by a Velodyne HDL-64E " + }, + { + "bbox": [ + 304, + 151, + 545, + 270 + ], + "type": "inline_equation", + "content": "360^{\\circ}" + }, + { + "bbox": [ + 304, + 151, + 545, + 270 + ], + "type": "text", + "content": " rotating LiDAR with 64 beams vertically. It contains 43,552 LiDAR scans in total and is split into a training set with 19,130 scans from sequences 00 to 10 except 08, a validation set containing sequence 08 with 4,071 scans, and a test set including the rest sequences with 20,351 scans. The test set is only provided with point clouds and used for the online leaderboards. For the panoptic segmentation, it has 19 valid categories, including 11 stuff and 8 things categories." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 270, + 545, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 270, + 545, + 354 + ], + "spans": [ + { + "bbox": [ + 304, + 270, + 545, + 354 + ], + "type": "text", + "content": "The nuScenes [5] is a newly released benchmark with 1,000 scenes collected by a Velodyne HDL-32E " + }, + { + "bbox": [ + 304, + 270, + 545, + 354 + ], + "type": "inline_equation", + "content": "360^{\\circ}" + }, + { + "bbox": [ + 304, + 270, + 545, + 354 + ], + "type": "text", + "content": " rotating LiDAR with 32 beams vertically. The dataset is collected in Boston and Singapore. It uses 28,130 scans for training, 6,019 for validation, and 6,008 for testing. The panoptic segmentation benchmark provides 16 categories, including 6 stuff and 10 things categories [12]." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 355, + 545, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 355, + 545, + 510 + ], + "spans": [ + { + "bbox": [ + 304, + 355, + 545, + 510 + ], + "type": "text", + "content": "Evaluation metric. As the official benchmarks [3, 12] suggest, the panoptic quality (PQ) is adopted to evaluate the performance of LiDAR panoptic segmentation. The PQ [3, 12] measures the overall quality of the panoptic segmentation, and can be decomposed into two terms: 1) the segmentation quality (SQ) is the average IoU of all matched pairs; 2) the recognition quality (RQ) measures the " + }, + { + "bbox": [ + 304, + 355, + 545, + 510 + ], + "type": "inline_equation", + "content": "F_{1}" + }, + { + "bbox": [ + 304, + 355, + 545, + 510 + ], + "type": "text", + "content": " score. These three metrics are also reported separately on the stuff and things classes, including " + }, + { + "bbox": [ + 304, + 355, + 545, + 510 + ], + "type": "inline_equation", + "content": "\\mathrm{PQ}^{St}" + }, + { + "bbox": [ + 304, + 355, + 545, + 510 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 355, + 545, + 510 + ], + "type": "inline_equation", + "content": "\\mathrm{SQ}^{St}" + }, + { + "bbox": [ + 304, + 355, + 545, + 510 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 355, + 545, + 510 + ], + "type": "inline_equation", + "content": "\\mathrm{RQ}^{St}" + }, + { + "bbox": [ + 304, + 355, + 545, + 510 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 355, + 545, + 510 + ], + "type": "inline_equation", + "content": "\\mathrm{PQ}^{Th}" + }, + { + "bbox": [ + 304, + 355, + 545, + 510 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 355, + 545, + 510 + ], + "type": "inline_equation", + "content": "\\mathrm{SQ}^{Th}" + }, + { + "bbox": [ + 304, + 355, + 545, + 510 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 355, + 545, + 510 + ], + "type": "inline_equation", + "content": "\\mathrm{RQ}^{Th}" + }, + { + "bbox": [ + 304, + 355, + 545, + 510 + ], + "type": "text", + "content": ", respectively. The " + }, + { + "bbox": [ + 304, + 355, + 545, + 510 + ], + "type": "inline_equation", + "content": "\\mathrm{PQ}^{\\dagger}" + }, + { + "bbox": [ + 304, + 355, + 545, + 510 + ], + "type": "text", + "content": " [26] is also reported by replacing the PQ of each stuff class with its IoU. For evaluating the sub-task of semantic segmentation, mean Intersection over Union (mIoU) [2] is reported." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "type": "text", + "content": "Training details. In the experiments, two 2D projection-based backbones, namely PolarNet [38] and CPGNet [23], are used in our CFNet. For the backbone of PolarNet [38], the training schedules and data augmentation are the same as those of Panoptic-PolarNet [39] for a fair comparison. Besides, for efficiency, only one stage version of CPGNet [23] is used in our CFNet. It is trained from scratch for 48 epochs with a batch size of 8 and takes around 24 hours on 8 NVIDIA RTX 3090 GPUs. Stochastic gradient descent (SGD) is used as the optimizer, where the initial learning rate, weight decay, and momentum are 0.02, 0.001, and 0.9, respectively. The learning rate is decayed by 0.1 per 10 epochs. In training, the data augmentation is applied as a convention, including random flipping along the " + }, + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "type": "text", + "content": " axes, random global scale sampled from " + }, + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "type": "inline_equation", + "content": "[0.95, 1.05]" + }, + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "type": "text", + "content": ", random rotation around the " + }, + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "type": "text", + "content": "-axis, and random Gaussian noise " + }, + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{N}(0, 0.02)" + }, + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "13430" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 75, + 69, + 519, + 205 + ], + "blocks": [ + { + "bbox": [ + 75, + 69, + 519, + 205 + ], + "lines": [ + { + "bbox": [ + 75, + 69, + 519, + 205 + ], + "spans": [ + { + "bbox": [ + 75, + 69, + 519, + 205 + ], + "type": "table", + "html": "
BackboneCFFEDeduplicationPQPQ†PQTthPQS\\( ^t \\)mIoURT(ms)
CFGFEM
a b c dPolarNet [38]✓BEV Center Heatmap [39]59.164.165.754.364.565.3 + 20.9
BEV Center Heatmap [39]59.564.366.254.464.669.7 + 20.9
✓✓BEV Center Heatmap [39]60.465.667.255.165.271.9 + 20.9
✓✓CDM [Ours]60.665.767.855.265.471.9 + 3.6
e f g h i jCPGNet [23]✓DBSCAN [11]59.564.163.856.364.631.6+24.9
HDBSCAN [6]58.362.960.956.364.931.6+48.7
MeanShift [9]59.964.564.756.364.931.6+84.6
CDM [Ours]60.565.666.156.566.332.7+2.3
✓CDM [Ours]60.865.966.356.966.537.4+2.3
✓✓CDM [Ours]62.767.570.057.367.441.2+2.3
", + "image_path": "e260e27b7bc4b3b641d3fa4692ea841d1a6482f1a7fa2b02d8e1870f30713add.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 246, + 147, + 258 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 246, + 147, + 258 + ], + "spans": [ + { + "bbox": [ + 47, + 246, + 147, + 258 + ], + "type": "text", + "content": "4.2. Ablation Studies" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 266, + 287, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 266, + 287, + 373 + ], + "spans": [ + { + "bbox": [ + 46, + 266, + 287, + 373 + ], + "type": "text", + "content": "Ablative analyses are conducted on the SemanticKITTI validation set with the above experimental setup. As shown in Table 1, two 2D projection-based backbones, namely PolarNet [38] and CPGNet [23], are incorporated with our framework to validate the proposed CFFE and CDM. Row (a) shows the original Panoptic-Polarnet [39], while rows (b,c,d) are implemented based on the official code of Panoptic-Polarnet [39]. The running time is reported by the time of the model inference plus that of the post-process." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 376, + 287, + 508 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 376, + 287, + 508 + ], + "spans": [ + { + "bbox": [ + 46, + 376, + 287, + 508 + ], + "type": "text", + "content": "Effects of the CFFE. The CFFE is configured in three ways: 1) the CFFE is not applied (a,e,f,g,h); 2) the center feature generation (CFG) in the CFFE is removed (b,i), which is equal to the setting of only imposing the intermediate supervision; 3) the whole CFFE is applied (c,d,j). For the two backbones, a common observation is that the improvements primarily come from the CFG rather than the intermediate supervision, especially for the things classes (a,b,c and h,i,j). In terms of the running time, The CFFE brings acceptable latency of " + }, + { + "bbox": [ + 46, + 376, + 287, + 508 + ], + "type": "inline_equation", + "content": "+6.6\\mathrm{ms}" + }, + { + "bbox": [ + 46, + 376, + 287, + 508 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 376, + 287, + 508 + ], + "type": "inline_equation", + "content": "+8.5\\mathrm{ms}" + }, + { + "bbox": [ + 46, + 376, + 287, + 508 + ], + "type": "text", + "content": " for these two backbones, respectively." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 509, + 287, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 509, + 287, + 593 + ], + "spans": [ + { + "bbox": [ + 46, + 509, + 287, + 593 + ], + "type": "text", + "content": "Moreover, the errors in predicting things center offsets are calculated for both the CFNet with CFFE and its intermediate results as shown in Table 2. Equipped with the backbone of CPGNet [23], the proposed CFFE reduces the errors of predicting the things center offsets. The observations demonstrate that explicitly modeling center features helps in identifying an instance." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 59, + 594, + 286, + 607 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 594, + 286, + 607 + ], + "spans": [ + { + "bbox": [ + 59, + 594, + 286, + 607 + ], + "type": "text", + "content": "Effects of the CDM. For the proposal-free meth" + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 86, + 628, + 249, + 668 + ], + "blocks": [ + { + "bbox": [ + 148, + 213, + 444, + 225 + ], + "lines": [ + { + "bbox": [ + 148, + 213, + 444, + 225 + ], + "spans": [ + { + "bbox": [ + 148, + 213, + 444, + 225 + ], + "type": "text", + "content": "Table 1. Ablation studies on the SemanticKITTI validation set. RT: running time." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 86, + 628, + 249, + 668 + ], + "lines": [ + { + "bbox": [ + 86, + 628, + 249, + 668 + ], + "spans": [ + { + "bbox": [ + 86, + 628, + 249, + 668 + ], + "type": "table", + "html": "
training setvalidation set
Intermediate1.351.22
CFNet1.191.13
", + "image_path": "21cded0f81ff2478bc96ed6c364e7815938ad7308d1ab4711488b2ef87db081f.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 677, + 286, + 710 + ], + "lines": [ + { + "bbox": [ + 46, + 677, + 286, + 710 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 286, + 710 + ], + "type": "text", + "content": "Table 2. Mean errors measured in meter (m) of things center offsets for the intermediate results and the CFNet with the CFFE on the SemanticKITTI training and validation set." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 246, + 545, + 390 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 246, + 545, + 390 + ], + "spans": [ + { + "bbox": [ + 304, + 246, + 545, + 390 + ], + "type": "text", + "content": "ods, there are primarily three ways to identify each instance: 1) the class-agnostic clustering methods (e.g. DBSCAN, HDBSCAN, MeanShift) (e,f,g), 2) the BEV center heatmap [39] (a,b,c), and 3) our CDM (d,h,i,j). For the backbone of PolarNet [38], our CDM outperforms the BEV center heatmap [39] by " + }, + { + "bbox": [ + 304, + 246, + 545, + 390 + ], + "type": "inline_equation", + "content": "+0.2" + }, + { + "bbox": [ + 304, + 246, + 545, + 390 + ], + "type": "text", + "content": " for the PQ (c,d), while our CDM runs faster. For the backbone of CPGNet [23], compared with the class-agnostic clustering methods, our CDM also achieves the best performance on both accuracy " + }, + { + "bbox": [ + 304, + 246, + 545, + 390 + ], + "type": "inline_equation", + "content": "(+0.6\\sim +2.2" + }, + { + "bbox": [ + 304, + 246, + 545, + 390 + ], + "type": "text", + "content": " for the PQ) and running time " + }, + { + "bbox": [ + 304, + 246, + 545, + 390 + ], + "type": "inline_equation", + "content": "(-21.5\\mathrm{ms}\\sim -" + }, + { + "bbox": [ + 304, + 246, + 545, + 390 + ], + "type": "inline_equation", + "content": "81.2\\mathrm{ms})" + }, + { + "bbox": [ + 304, + 246, + 545, + 390 + ], + "type": "text", + "content": " , though it slows down the model inference by " + }, + { + "bbox": [ + 304, + 246, + 545, + 390 + ], + "type": "inline_equation", + "content": "+1.1~\\mathrm{ms}" + }, + { + "bbox": [ + 304, + 246, + 545, + 390 + ], + "type": "inline_equation", + "content": "\\mathrm{(g,h)}" + }, + { + "bbox": [ + 304, + 246, + 545, + 390 + ], + "type": "text", + "content": " owing to the additional confidence prediction." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 390, + 545, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 390, + 545, + 413 + ], + "spans": [ + { + "bbox": [ + 304, + 390, + 545, + 413 + ], + "type": "text", + "content": "More ablation studies can be seen in the supplementary material." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 420, + 444, + 433 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 420, + 444, + 433 + ], + "spans": [ + { + "bbox": [ + 305, + 420, + 444, + 433 + ], + "type": "text", + "content": "4.3. Evaluation Comparisons" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 438, + 545, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 438, + 545, + 594 + ], + "spans": [ + { + "bbox": [ + 304, + 438, + 545, + 594 + ], + "type": "text", + "content": "Quantitative results. In Table 3 and Table 4, our CFNet with CPGNet as its backbone is compared with the existing methods on the SemanticKITTI test set and nuScenes validation set, respectively. From top to bottom, the methods are grouped as proposal-based and proposal-free methods. It is found that our proposal-free CFNet surpasses the previous methods by a large margin on most metrics. Specifically, on the SemanticKITTI test set, our CFNet outperforms the previous state-of-the-art SCAN [35] and Panoptic-PHNet [22] by " + }, + { + "bbox": [ + 304, + 438, + 545, + 594 + ], + "type": "inline_equation", + "content": "+1.9" + }, + { + "bbox": [ + 304, + 438, + 545, + 594 + ], + "type": "text", + "content": " for the PQ, and " + }, + { + "bbox": [ + 304, + 438, + 545, + 594 + ], + "type": "inline_equation", + "content": "+1.5" + }, + { + "bbox": [ + 304, + 438, + 545, + 594 + ], + "type": "text", + "content": " for the RQ. On the nuScenes validation set, our CFNet outperforms the most competitive Panoptic-PHNet [22] by " + }, + { + "bbox": [ + 304, + 438, + 545, + 594 + ], + "type": "inline_equation", + "content": "+0.4" + }, + { + "bbox": [ + 304, + 438, + 545, + 594 + ], + "type": "text", + "content": " for the PQ, " + }, + { + "bbox": [ + 304, + 438, + 545, + 594 + ], + "type": "inline_equation", + "content": "+0.6" + }, + { + "bbox": [ + 304, + 438, + 545, + 594 + ], + "type": "text", + "content": " for the SQ, and " + }, + { + "bbox": [ + 304, + 438, + 545, + 594 + ], + "type": "inline_equation", + "content": "+0.4" + }, + { + "bbox": [ + 304, + 438, + 545, + 594 + ], + "type": "text", + "content": " for the RQ, respectively." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 594, + 545, + 701 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 545, + 701 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 545, + 701 + ], + "type": "text", + "content": "Running time. The running time is reported on the SemanticKITTI. It is measured with PyTorch FP32 without any model acceleration tricks on a single NVIDIA RTX 3090 GPU. For clarity, the running time of our method is further divided into two parts, including the time costs of the model and post-process. Based on the same kind of devices as the existing methods, our CFNet runs " + }, + { + "bbox": [ + 304, + 594, + 545, + 701 + ], + "type": "inline_equation", + "content": "43.5\\mathrm{ms}" + }, + { + "bbox": [ + 304, + 594, + 545, + 701 + ], + "type": "text", + "content": " (41.2 ms for model inference and 2.3 ms for post-process) and is 1.6 times faster than the most efficient Panoptic-PHNet [22]." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 317, + 701, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 701, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 317, + 701, + 545, + 713 + ], + "type": "text", + "content": "Qualitative results. The visualization results are" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13431" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 75, + 69, + 520, + 250 + ], + "blocks": [ + { + "bbox": [ + 75, + 69, + 520, + 250 + ], + "lines": [ + { + "bbox": [ + 75, + 69, + 520, + 250 + ], + "spans": [ + { + "bbox": [ + 75, + 69, + 520, + 250 + ], + "type": "table", + "html": "
MethodsPQPQ†SQRQPQThSQThRQThPQStSQStRQStmIoURT(ms)
Proposal-based Methods
RangeNet++ [25]&PointPillars [19]37.145.975.947.020.275.225.249.376.562.852.4416.6
KPConv [33]&PointPillars [19]44.552.580.054.432.781.538.753.179.065.958.8526.3
PanopticTrackNet [17]43.150.778.853.928.680.435.553.677.767.352.6147
EfficientLPS [31]57.463.283.068.753.187.860.560.579.574.661.4-
Proposal-free Methods
LPSAD [24]38.047.076.548.225.676.831.847.176.260.150.984.7
DS-Net [15]55.962.582.366.755.187.262.856.578.769.561.6294.1
Panoster [13]52.759.980.764.149.483.358.555.178.868.259.9-
GP-S3Net [29]60.069.082.072.165.086.674.556.478.770.470.8270.3
SMAC-Seg [20]56.162.582.067.953.085.661.858.479.372.363.399
Panoptic-PolarNet [39]54.160.781.465.053.387.260.654.877.268.159.586.2
SCAN [35]61.567.584.572.161.488.169.361.581.874.167.778.1
Panoptic-PHNet [22]61.567.984.872.163.890.770.459.980.573.366.069.3
CFNet [Ours] w/CPGNet [23]63.469.785.273.666.789.874.361.081.973.168.341.2 + 2.3
", + "image_path": "c1d1530e659ba9bc197f27396204ab5a02a6f269878bb2dc63286363bf2dfe16.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 75, + 279, + 519, + 441 + ], + "blocks": [ + { + "bbox": [ + 187, + 258, + 405, + 269 + ], + "lines": [ + { + "bbox": [ + 187, + 258, + 405, + 269 + ], + "spans": [ + { + "bbox": [ + 187, + 258, + 405, + 269 + ], + "type": "text", + "content": "Table 3. Comparison results on the SemanticKITTI test set." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 75, + 279, + 519, + 441 + ], + "lines": [ + { + "bbox": [ + 75, + 279, + 519, + 441 + ], + "spans": [ + { + "bbox": [ + 75, + 279, + 519, + 441 + ], + "type": "table", + "html": "
MethodsPQ\\( PQ^† \\)SQRQ\\( PQ^{Th} \\)\\( SQ^{Th} \\)\\( RQ^{Th} \\)\\( PQ^{St} \\)\\( SQ^{St} \\)\\( RQ^{St} \\)mIoU
Proposal-based Methods
Cylinder3D [40]&PointPillars [19]36.044.583.343.023.383.727.057.282.769.652.3
Cylinder3D [40]&SECOND [36]40.148.484.247.329.084.433.658.583.770.158.5
PanopticTrackNet [17]51.456.280.263.345.881.455.960.478.375.558.0
EfficientLPS [31]62.065.683.473.956.883.268.070.683.883.665.6
Proposal-free Methods
LPSAD [24]50.457.779.462.443.280.253.257.578.571.762.5
DS-Net [15]42.551.083.650.332.583.138.359.284.470.370.7
GP-S3Net [29]61.067.584.172.056.085.365.266.082.978.775.8
SMAC-Seg [20]67.071.885.078.265.287.174.268.882.982.272.2
Panoptic-PolarNet [39]63.467.283.975.359.284.170.370.483.683.566.9
SCAN [35]65.168.985.775.360.685.770.272.585.783.877.4
Panoptic-PHNet [22]74.777.788.284.274.089.082.575.986.886.979.7
CFNet [Ours] w/CPGNet [23]75.178.088.884.674.889.882.976.687.187.379.3
", + "image_path": "96ece7ee8e70c4b00886032f46ccd7728fc8d3c5494baeff9561740f059db2d0.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 187, + 449, + 405, + 460 + ], + "lines": [ + { + "bbox": [ + 187, + 449, + 405, + 460 + ], + "spans": [ + { + "bbox": [ + 187, + 449, + 405, + 460 + ], + "type": "text", + "content": "Table 4. Comparison results on the nuScenes validation set." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 54, + 480, + 168, + 596 + ], + "blocks": [ + { + "bbox": [ + 54, + 480, + 168, + 596 + ], + "lines": [ + { + "bbox": [ + 54, + 480, + 168, + 596 + ], + "spans": [ + { + "bbox": [ + 54, + 480, + 168, + 596 + ], + "type": "image", + "image_path": "b052111d925cc62768d07df341b90122b38eeabbf009657386e66860535c3e11.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 70, + 597, + 152, + 606 + ], + "lines": [ + { + "bbox": [ + 70, + 597, + 152, + 606 + ], + "spans": [ + { + "bbox": [ + 70, + 597, + 152, + 606 + ], + "type": "text", + "content": "Semantic Segmentation" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 168, + 480, + 280, + 537 + ], + "blocks": [ + { + "bbox": [ + 168, + 480, + 280, + 537 + ], + "lines": [ + { + "bbox": [ + 168, + 480, + 280, + 537 + ], + "spans": [ + { + "bbox": [ + 168, + 480, + 280, + 537 + ], + "type": "image", + "image_path": "e5f441e2eb16aac27dfee5184afbbbaaa802c81f1a6a4849e6c66705f19ab3fd.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 168, + 538, + 280, + 595 + ], + "blocks": [ + { + "bbox": [ + 168, + 538, + 280, + 595 + ], + "lines": [ + { + "bbox": [ + 168, + 538, + 280, + 595 + ], + "spans": [ + { + "bbox": [ + 168, + 538, + 280, + 595 + ], + "type": "image", + "image_path": "d6176e70d804c2a52d065fda3f5ff5c752045d44377594bd42ab595d7a3169a3.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 184, + 597, + 263, + 606 + ], + "lines": [ + { + "bbox": [ + 184, + 597, + 263, + 606 + ], + "spans": [ + { + "bbox": [ + 184, + 597, + 263, + 606 + ], + "type": "text", + "content": "Instance Segmentation" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 615, + 287, + 638 + ], + "lines": [ + { + "bbox": [ + 46, + 615, + 287, + 638 + ], + "spans": [ + { + "bbox": [ + 46, + 615, + 287, + 638 + ], + "type": "text", + "content": "Figure 5. Visualizations of our CFNet on the SemanticKITTI test set. Different colors represent different classes or instances." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "type": "text", + "content": "shown in Fig. 5. Our CFNet can distinguish adjacent pedestrians or cars. Besides, the boundaries of instances can be accurately segmented. More qualitative results can be seen in the supplementary material." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 481, + 379, + 493 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 481, + 379, + 493 + ], + "spans": [ + { + "bbox": [ + 306, + 481, + 379, + 493 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 501, + 547, + 633 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 501, + 547, + 633 + ], + "spans": [ + { + "bbox": [ + 304, + 501, + 547, + 633 + ], + "type": "text", + "content": "A novel proposal-free CFNet is proposed for real-time LiDAR panoptic segmentation. For better modeling and exploiting the non-existent instance centers, a novel CFFE is proposed to generate enhanced center-focusing feature maps and a CDM is introduced to keep only one center for each instance and then assign the shifted things points to the closest center for acquiring instance IDs. From the experiments, it can be seen that center modeling and exploiting is a key problem in the proposal-free LiDAR panoptic segmentation methods, and mimicking the non-existent center features is promising and shows a clear benefit." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 645, + 409, + 658 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 645, + 409, + 658 + ], + "spans": [ + { + "bbox": [ + 306, + 645, + 409, + 658 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 665, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 546, + 713 + ], + "type": "text", + "content": "The research project is partially supported by the National Key R&D Program of China (No.2021ZD0111902), and the National Natural Science Foundation of China (No.U21B2038, 62272015, U19B2039)." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "13432" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 289, + 714 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 53, + 91, + 289, + 145 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 289, + 145 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 289, + 145 + ], + "type": "text", + "content": "[1] Yara Ali Alnaggar, Mohamed Afifi, Karim Amer, and Mohamed ElHelw. Multi projection fusion for real-time semantic segmentation of 3d lidar point clouds. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1800-1809, 2021. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 147, + 288, + 212 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 147, + 288, + 212 + ], + "spans": [ + { + "bbox": [ + 53, + 147, + 288, + 212 + ], + "type": "text", + "content": "[2] Jens Behley, Martin Garbade, Andres Milioto, Jan Quenzel, Sven Behnke, Cyril Stachniss, and Jurgen Gall. Semantickitti: A dataset for semantic scene understanding of lidar sequences. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9297-9307, 2019. 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 213, + 288, + 257 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 213, + 288, + 257 + ], + "spans": [ + { + "bbox": [ + 53, + 213, + 288, + 257 + ], + "type": "text", + "content": "[3] Jens Behley, Andres Milioto, and Cyril Stachniss. A benchmark for lidar-based panoptic segmentation based on kitti. In 2021 IEEE International Conference on Robotics and Automation (ICRA), pages 13596-13603. IEEE, 2021. 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 258, + 288, + 323 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 258, + 288, + 323 + ], + "spans": [ + { + "bbox": [ + 53, + 258, + 288, + 323 + ], + "type": "text", + "content": "[4] Maxim Berman, Amal Rannen Triki, and Matthew B Blaschko. The lovasz-softmax loss: A tractable surrogate for the optimization of the intersection-over-union measure in neural networks. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 4413-4421, 2018. 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 324, + 288, + 390 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 324, + 288, + 390 + ], + "spans": [ + { + "bbox": [ + 53, + 324, + 288, + 390 + ], + "type": "text", + "content": "[5] Holger Caesar, Varun Bankiti, Alex H Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuscenes: A multimodal dataset for autonomous driving. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11621-11631, 2020. 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 391, + 288, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 391, + 288, + 435 + ], + "spans": [ + { + "bbox": [ + 53, + 391, + 288, + 435 + ], + "type": "text", + "content": "[6] Ricardo JGB Campello, Davoud Moulavi, and Jörg Sander. Density-based clustering based on hierarchical density estimates. In Pacific-Asia conference on knowledge discovery and data mining, pages 160–172. Springer, 2013. 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 435, + 288, + 500 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 435, + 288, + 500 + ], + "spans": [ + { + "bbox": [ + 53, + 435, + 288, + 500 + ], + "type": "text", + "content": "[7] Ran Cheng, Ryan Razani, Ehsan Taghavi, Enxu Li, and Bingbing Liu. 2-s3net: Attentive feature fusion with adaptive feature selection for sparse semantic segmentation network. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12547-12556, 2021. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 502, + 288, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 502, + 288, + 555 + ], + "spans": [ + { + "bbox": [ + 53, + 502, + 288, + 555 + ], + "type": "text", + "content": "[8] Christopher Choy, JunYoung Gwak, and Silvio Savarese. 4d spatio-temporal convnets: Minkowski convolutional neural networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3075-3084, 2019. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 557, + 288, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 557, + 288, + 601 + ], + "spans": [ + { + "bbox": [ + 53, + 557, + 288, + 601 + ], + "type": "text", + "content": "[9] Dorin Comaniciu and Peter Meer. Mean shift: A robust approach toward feature space analysis. IEEE Transactions on pattern analysis and machine intelligence, 24(5):603-619, 2002. 3, 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 602, + 288, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 602, + 288, + 646 + ], + "spans": [ + { + "bbox": [ + 48, + 602, + 288, + 646 + ], + "type": "text", + "content": "[10] Tiago Cortinhal, George Tzelepis, and Eren Erdal Aksoy. Salsanext: Fast, uncertainty-aware semantic segmentation of lidar point clouds. In International Symposium on Visual Computing, pages 207-222. Springer, 2020. 2, 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 647, + 288, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 647, + 288, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 647, + 288, + 689 + ], + "type": "text", + "content": "[11] Martin Ester, Hans-Peter Kriegel, Jörg Sander, Xiaowei Xu, et al. A density-based algorithm for discovering clusters in large spatial databases with noise. In kdd, volume 96, pages 226-231, 1996. 3, 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 691, + 288, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 288, + 714 + ], + "type": "text", + "content": "[12] Whye Kit Fong, Rohit Mohan, Juana Valeria Hurtado, Lubing Zhou, Holger Caesar, Oscar Beijbom, and Abhinav Val-" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 547, + 712 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 106 + ], + "type": "text", + "content": "ada. Panoptic nuscenes: A large-scale benchmark for lidar panoptic segmentation and tracking. IEEE Robotics and Automation Letters, 7(2):3795-3802, 2022. 6" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 308, + 107, + 547, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 107, + 547, + 161 + ], + "spans": [ + { + "bbox": [ + 308, + 107, + 547, + 161 + ], + "type": "text", + "content": "[13] Stefano Gasperini, Mohammad-Ali Nikouei Mahani, Alvaro Marcos-Ramiro, Nassir Navab, and Federico Tombari. Panoster: End-to-end panoptic segmentation of lidar point clouds. IEEE Robotics and Automation Letters, 6(2):3216-3223, 2021. 1, 3, 8" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 163, + 547, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 163, + 547, + 207 + ], + "spans": [ + { + "bbox": [ + 308, + 163, + 547, + 207 + ], + "type": "text", + "content": "[14] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In Proceedings of the IEEE international conference on computer vision, pages 2961-2969, 2017. 1, 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 209, + 545, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 209, + 545, + 262 + ], + "spans": [ + { + "bbox": [ + 308, + 209, + 545, + 262 + ], + "type": "text", + "content": "[15] Fangzhou Hong, Hui Zhou, Xinge Zhu, Hongsheng Li, and Ziwei Liu. Lidar-based panoptic segmentation via dynamic shifting network. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13090-13099, 2021. 1, 3, 5, 6, 8" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 264, + 545, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 264, + 545, + 330 + ], + "spans": [ + { + "bbox": [ + 308, + 264, + 545, + 330 + ], + "type": "text", + "content": "[16] Qingyong Hu, Bo Yang, Linhai Xie, Stefano Rosa, Yulan Guo, Zhihua Wang, Niki Trigoni, and Andrew Markham. Randla-net: Efficient semantic segmentation of large-scale point clouds. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11108-11117, 2020. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 331, + 545, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 331, + 545, + 365 + ], + "spans": [ + { + "bbox": [ + 308, + 331, + 545, + 365 + ], + "type": "text", + "content": "[17] Juana Valeria Hurtado, Rohit Mohan, Wolfram Burgard, and Abhinav Valada. Mopt: Multi-object panoptic tracking. arXiv preprint arXiv:2004.08189, 2020. 1, 2, 8" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 365, + 545, + 410 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 365, + 545, + 410 + ], + "spans": [ + { + "bbox": [ + 308, + 365, + 545, + 410 + ], + "type": "text", + "content": "[18] Alexander Kirillov, Kaiming He, Ross Girshick, Carsten Rother, and Piotr Dollar. Panoptic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9404-9413, 2019. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 411, + 545, + 466 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 411, + 545, + 466 + ], + "spans": [ + { + "bbox": [ + 308, + 411, + 545, + 466 + ], + "type": "text", + "content": "[19] Alex H Lang, Sourabh Vora, Holger Caesar, Lubing Zhou, Jiong Yang, and Oscar Beijbom. Pointpillars: Fast encoders for object detection from point clouds. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12697-12705, 2019. 1, 2, 4, 8" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 467, + 545, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 467, + 545, + 510 + ], + "spans": [ + { + "bbox": [ + 308, + 467, + 545, + 510 + ], + "type": "text", + "content": "[20] Enxu Li, Ryan Razani, Yixuan Xu, and Liu Bingbing. Smac-seg: Lidar panoptic segmentation via sparse multi-directional attention clustering. arXiv preprint arXiv:2108.13588, 2021. 2, 3, 8" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 512, + 545, + 545 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 512, + 545, + 545 + ], + "spans": [ + { + "bbox": [ + 308, + 512, + 545, + 545 + ], + "type": "text", + "content": "[21] Enxu Li, Ryan Razani, Yixuan Xu, and Bingbing Liu. Cpseg: Cluster-free panoptic segmentation of 3d lidar point clouds. arXiv preprint arXiv:2111.01723, 2021. 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 546, + 545, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 546, + 545, + 611 + ], + "spans": [ + { + "bbox": [ + 308, + 546, + 545, + 611 + ], + "type": "text", + "content": "[22] Jinke Li, Xiao He, Yang Wen, Yuan Gao, Xiaogiang Cheng, and Dan Zhang. Panoptic-phnet: Towards real-time and high-precision lidar panoptic segmentation via clustering pseudo heatmap. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11809–11818, 2022. 1, 3, 7, 8" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 613, + 545, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 613, + 545, + 657 + ], + "spans": [ + { + "bbox": [ + 308, + 613, + 545, + 657 + ], + "type": "text", + "content": "[23] Xiaoyan Li, Gang Zhang, Hongyu Pan, and Zhenhua Wang. Cpgnet: Cascade point-grid fusion network for real-time lidar semantic segmentation. arXiv preprint arXiv:2204.09914, 2022. 2, 4, 5, 6, 7, 8" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 658, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 658, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 308, + 658, + 545, + 712 + ], + "type": "text", + "content": "[24] Andres Milioto, Jens Behley, Chris McCool, and Cyril Stachniss. Lidar panoptic segmentation for autonomous driving. In 2020 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 8505-8512. IEEE, 2020. 2, 8" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13433" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "text", + "content": "[25] Andres Milioto, Ignacio Vizzo, Jens Behley, and Cyril Stachniss. Rangenet++: Fast and accurate lidar semantic segmentation. In 2019 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 4213-4220. IEEE, 2019. 2, 4, 8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 129, + 287, + 174 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 129, + 287, + 174 + ], + "spans": [ + { + "bbox": [ + 48, + 129, + 287, + 174 + ], + "type": "text", + "content": "[26] Lorenzo Porzi, Samuel Rota Bulo, Aleksander Colovic, and Peter Kontschieder. Seamless scene segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8277-8286, 2019. 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 175, + 287, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 175, + 287, + 228 + ], + "spans": [ + { + "bbox": [ + 48, + 175, + 287, + 228 + ], + "type": "text", + "content": "[27] Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. Pointnet: Deep learning on point sets for 3d classification and segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 652-660, 2017. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 230, + 287, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 230, + 287, + 262 + ], + "spans": [ + { + "bbox": [ + 48, + 230, + 287, + 262 + ], + "type": "text", + "content": "[28] Charles R Qi, Li Yi, Hao Su, and Leonidas J Guibas. Point-net++: Deep hierarchical feature learning on point sets in a metric space. arXiv preprint arXiv:1706.02413, 2017. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 264, + 287, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 264, + 287, + 319 + ], + "spans": [ + { + "bbox": [ + 48, + 264, + 287, + 319 + ], + "type": "text", + "content": "[29] Ryan Razani, Ran Cheng, Enxu Li, Ehsan Taghavi, Yuan Ren, and Liu Bingbing. Gp-s3net: Graph-based panoptic sparse semantic segmentation network. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 16076-16085, 2021. 1, 3, 8" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 320, + 287, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 320, + 287, + 376 + ], + "spans": [ + { + "bbox": [ + 48, + 320, + 287, + 376 + ], + "type": "text", + "content": "[30] Shaoshuai Shi, Chaoxu Guo, Li Jiang, Zhe Wang, Jianping Shi, Xiaogang Wang, and Hongsheng Li. Pv-rcnn: Pointvoxel feature set abstraction for 3d object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10529–10538, 2020. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 376, + 287, + 420 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 376, + 287, + 420 + ], + "spans": [ + { + "bbox": [ + 48, + 376, + 287, + 420 + ], + "type": "text", + "content": "[31] Kshitij Sirohi, Rohit Mohan, Daniel Buscher, Wolfram Burgard, and Abhinav Valada. Efficienttps: Efficient lidar panoptic segmentation. IEEE Transactions on Robotics, 2021. 1, 2, 8" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 422, + 287, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 422, + 287, + 475 + ], + "spans": [ + { + "bbox": [ + 48, + 422, + 287, + 475 + ], + "type": "text", + "content": "[32] Haotian Tang, Zhijian Liu, Shengyu Zhao, Yujun Lin, Ji Lin, Hanrui Wang, and Song Han. Searching efficient 3d architectures with sparse point-voxel convolution. In European Conference on Computer Vision, pages 685-702. Springer, 2020. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 477, + 287, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 477, + 287, + 542 + ], + "spans": [ + { + "bbox": [ + 48, + 477, + 287, + 542 + ], + "type": "text", + "content": "[33] Hugues Thomas, Charles R Qi, Jean-Emmanuel Deschaud, Beatrix Marcotegui, François Goulette, and Leonidas J Guibas. Kpconv: Flexible and deformable convolution for point clouds. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6411–6420, 2019. 2, 8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 544, + 287, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 544, + 287, + 600 + ], + "spans": [ + { + "bbox": [ + 48, + 544, + 287, + 600 + ], + "type": "text", + "content": "[34] Chenfeng Xu, Bichen Wu, Zining Wang, Wei Zhan, Peter Vajda, Kurt Keutzer, and Masayoshi Tomizuka. Squeeze-segv3: Spatially-adaptive convolution for efficient pointcloud segmentation. In European Conference on Computer Vision, pages 1-19. Springer, 2020. 2, 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 601, + 287, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 601, + 287, + 644 + ], + "spans": [ + { + "bbox": [ + 48, + 601, + 287, + 644 + ], + "type": "text", + "content": "[35] Shuangjie Xu, Rui Wan, Maosheng Ye, Xiaoyi Zou, and Tongyi Cao. Sparse cross-scale attention network for efficient lidar panoptic segmentation. arXiv preprint arXiv:2201.05972, 2022. 1, 2, 3, 7, 8" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 646, + 287, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 646, + 287, + 667 + ], + "spans": [ + { + "bbox": [ + 48, + 646, + 287, + 667 + ], + "type": "text", + "content": "[36] Yan Yan, Yuxing Mao, and Bo Li. Second: Sparsely embedded convolutional detection. Sensors, 18(10):3337, 2018. 8" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 669, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 669, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 669, + 287, + 713 + ], + "type": "text", + "content": "[37] Dongqiangzi Ye, Zixiang Zhou, Weijia Chen, Yufei Xie, Yu Wang, Panqu Wang, and Hassan Foroosh. Lidarmultinet: Towards a unified multi-task network for lidar perception. arXiv preprint arXiv:2209.09385, 2022. 1, 2" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 72, + 546, + 261 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 307, + 72, + 546, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 72, + 546, + 138 + ], + "spans": [ + { + "bbox": [ + 307, + 72, + 546, + 138 + ], + "type": "text", + "content": "[38] Yang Zhang, Zixiang Zhou, Philip David, Xiangyu Yue, Zerong Xi, Boqing Gong, and Hassan Foroosh. Polarnet: An improved grid representation for online lidar point clouds semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9601-9610, 2020. 2, 4, 6, 7" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 140, + 546, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 140, + 546, + 194 + ], + "spans": [ + { + "bbox": [ + 307, + 140, + 546, + 194 + ], + "type": "text", + "content": "[39] Zixiang Zhou, Yang Zhang, and Hassan Foroosh. Panoptic-polarnet: Proposal-free lidar point cloud panoptic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13194-13203, 2021. 1, 2, 3, 6, 7, 8" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 196, + 546, + 261 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 196, + 546, + 261 + ], + "spans": [ + { + "bbox": [ + 307, + 196, + 546, + 261 + ], + "type": "text", + "content": "[40] Xinge Zhu, Hui Zhou, Tai Wang, Fangzhou Hong, Yuexin Ma, Wei Li, Hongsheng Li, and Dahua Lin. Cylindrical and asymmetrical 3d convolution networks for lidar segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9939-9948, 2021. 2, 8" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "text", + "content": "13434" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Change-Aware Sampling and Contrastive Learning for Satellite Images/ef28aefe-c0fe-45ce-8981-6fc9107e6b77_content_list.json b/2023/Change-Aware Sampling and Contrastive Learning for Satellite Images/ef28aefe-c0fe-45ce-8981-6fc9107e6b77_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..c9a09cb9f62e501bfd2202f24136334adfe00d3c --- /dev/null +++ b/2023/Change-Aware Sampling and Contrastive Learning for Satellite Images/ef28aefe-c0fe-45ce-8981-6fc9107e6b77_content_list.json @@ -0,0 +1,1711 @@ +[ + { + "type": "text", + "text": "Change-Aware Sampling and Contrastive Learning for Satellite Images", + "text_level": 1, + "bbox": [ + 125, + 130, + 843, + 151 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Utkarsh Mall", + "bbox": [ + 228, + 180, + 336, + 196 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Bharath Hariharan", + "bbox": [ + 415, + 181, + 563, + 196 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Cornell University", + "bbox": [ + 410, + 199, + 558, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{utkarshm, bharathh, kb}@cs.cornell.edu", + "bbox": [ + 310, + 218, + 653, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Kavita Bala", + "bbox": [ + 643, + 181, + 740, + 196 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 268, + 312, + 282 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Automatic remote sensing tools can help inform many large-scale challenges such as disaster management, climate change, etc. While a vast amount of spatio-temporal satellite image data is readily available, most of it remains unlabelled. Without labels, this data is not very useful for supervised learning algorithms. Self-supervised learning instead provides a way to learn effective representations for various downstream tasks without labels. In this work, we leverage characteristics unique to satellite images to learn better self-supervised features. Specifically, we use the temporal signal to contrast images with long-term and short-term differences, and we leverage the fact that satellite images do not change frequently. Using these characteristics, we formulate a new loss contrastive loss called Change-Aware Contrastive (CACo) Loss. Further, we also present a novel method of sampling different geographical regions. We show that leveraging these properties leads to better performance on diverse downstream tasks. For example, we see a $6.5\\%$ relative improvement for semantic segmentation and an $8.5\\%$ relative improvement for change detection over the best-performing baseline with our method.", + "bbox": [ + 75, + 301, + 473, + 618 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 647, + 209, + 662 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Our planet is surrounded by a large number of satellites constantly collecting images of the world. This massive trove of visual information can help monitor phenomena at the world-scale, and inform solutions to global problems such as climate change or loss of biodiversity. Automatic vision tools can help by, for example, monitoring land-use change over time [36] or the evolution of urban areas [6].", + "bbox": [ + 75, + 672, + 468, + 779 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "However, training all these models requires labeled data. Unfortunately, labeling the massive trove of satellite images is expensive, more so than internet images because of the expertise necessary. This issue is exacerbated by the different label requirements of many monitoring applications.", + "bbox": [ + 75, + 779, + 468, + 854 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "One way to alleviate this problem of limited labeled data is to use self-supervised learning techniques to learn a good feature representation from unlabeled satellite im", + "bbox": [ + 75, + 854, + 468, + 900 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/ab9b7b2dfaf29125f6acf27058333da77fb75d35c2c81fb3b1fc3b8c7cf51884.jpg", + "image_caption": [ + "Figure 1. Images of the same location at three different times. Changes from 2016 to 2020 are due to major urban development, while those from July to September are seasonal variations. Our approach, CACo, learns features that are sensitive to the former but invariant to the latter." + ], + "image_footnote": [], + "bbox": [ + 498, + 263, + 895, + 478 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "agery. This representation can then be further finetuned with much fewer labels for specific applications. Modern self-supervised learning approaches are based on contrastive learning. These techniques train a feature space so that each image in the dataset is embedded close to augmented versions of itself (e.g., with jittered colors) but far from other images of the dataset. A possible approach is to directly apply these techniques on satellite image datasets. However, the spatio-temporal structure of satellite imagery is much richer than the unstructured collections of internet images typically used in standard self-supervised learning. In this work, we ask, how can we best leverage the structure of satellite images for better self-supervised learning?", + "bbox": [ + 496, + 582, + 892, + 779 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The first important aspect of the spatio-temporal structure of satellite images is the availability of multiple temporally-spaced images for the same location. Past work has used this structure to sample images spread over a few weeks or months from each location to encourage invariance to seasonal variations [25]. However, we can access images not just over a few months, but over many years. Over such long time spans, we often see significant, per", + "bbox": [ + 496, + 780, + 890, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 807, + 44 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "5261", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "manent change, such as the construction of houses, the drying of lakes, or the logging of forests (Fig. 1). While we want feature representations to be invariant to temporary, seasonal change, we want the representation to be sensitive to permanent, long-term change. To capture this intuition, we sample multiple images from the same location covering both short and long time spans and encourage invariance to the former but sensitivity to the latter.", + "bbox": [ + 75, + 90, + 472, + 210 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "A second and more important aspect of satellite images is that permanent change is spatially rare. Change is concentrated near urban areas, but many parts of the planet see little change. When there is no change, we want our feature representation to be the same even over long time spans. We capture this insight with a novel strategy to robustly estimate whether or not there is a long-term change, even in the middle of training, by comparing long-term feature differences to short-term variations. We then design a loss function that is conditional on this change estimate: it encourages invariances to long-term differences depending on whether a change occurs or not. We call this novel loss function Change-aware Contrastive Loss (or CaCo).", + "bbox": [ + 75, + 212, + 472, + 407 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The above loss function uses the temporal structure of satellite images. We can also utilize geographical structure by carefully sampling the most informative locations on the planet. We provide an improvement over a previously proposed geographical sampling [25]. We show that sampling closer to cities, and ignoring samples completely in the ocean can result in a dataset much more useful for learning a general representation for various downstream tasks. We evaluate our new representation on a diverse set of downstream tasks such as landcover classification, semantic segmentation, and change detection. Our method achieves significant relative improvements (ranging from $6.5\\%$ to $8.5\\%$ ) over the state-of-the-art for a variety of tasks such as segmentation and change detection.", + "bbox": [ + 75, + 407, + 472, + 619 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To summarize, we make the following contributions:", + "bbox": [ + 96, + 619, + 444, + 633 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose a new self-supervised learning loss that uses long-term temporal information in satellite imagery to encourage invariance to seasonal variations but sensitivity to permanent, long-term changes.", + "- We introduce a novel approach to robustly estimate whether a location has undergone significant change by comparing long-term changes to seasonal variations. Our new change-aware loss function (CACo) uses this to decide when to encourage invariance.", + "- We use an improved geographical sampling that provides more diverse data for representation learning." + ], + "bbox": [ + 94, + 643, + 468, + 816 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 76, + 829, + 218, + 844 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Self-supervised Learning. Self-supervised learning methods can be used to learn a good general representation from a dataset without any labels. These methods obtain", + "bbox": [ + 75, + 854, + 470, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "supervisory signals or inductive bias from the unlabelled data itself. Earlier self-supervised works used \"pretext tasks\" such as rotation prediction [15], solving jigsaw puzzles [28], colorization [44], or missing data completion [17]. More recently contrastive learning methods have been shown to learn a better representation than pretext tasks. Methods for contrastive learning such as NPID [43], PIRL [26], MoCo [8, 18], SimCLR [7], and BYOL [16], use instance discrimination to learn a feature representation. While these methods can be applied even to satellite images, in this work we leverage the structure of satellite images that can be used for self-supervised learning.", + "bbox": [ + 496, + 90, + 893, + 273 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Self-supervised Learning in Remote Sensing. Increasingly, self-supervised learning is being applied to the area of remote sensing. Prior works have used signals from unlabeled data such as location [2, 20], seasonal variations [25], or texture [1], as a signal for self-supervision. Recent methods use transformers [12] and diffusion models [4] for self-supervised representation [10, 32].", + "bbox": [ + 496, + 291, + 893, + 398 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Remote Sensing Applications and Data Several remote sensing applications require vision tools for automatic recognition at scale. Applications such as landcover classification [19, 34] or segmentation [36, 40] benefit from the advances in vision algorithms for scene classification and semantic segmentation. Several object detection applications have also been developed using object detection algorithms such as for detecting constructions [9], buildings [38], trees [41], or floating objects [14]. Segmentation methods also help in extracting structural information like road networks [3, 22, 39], crop types [37], or clouds [27] from remote sensing images. Many applications such as change detection [11] require temporal information for regions as well. Temporal information and changes are also used to find long-term semantic events [24]. In this work, we aim to learn a representation without supervision, that leads to better performance on all these downstream tasks.", + "bbox": [ + 496, + 415, + 893, + 672 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The image data for these algorithms and datasets are obtained from a few different sources. Sentinel-2 [13] provides multispectral images (13 bands) at $10\\mathrm{m}$ resolution with a temporal revisit of 5 days. Several other satellites provide higher resolution information such as WorldView-3 and PlanetScope. We use Sentinel-2 satellite imagery as it provides frequent temporal information that we use in our self-supervised formulation. Furthermore, many of the downstream tasks [11, 19, 24, 36] use Sentinel-2 imagery.", + "bbox": [ + 496, + 672, + 893, + 809 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 500, + 821, + 591, + 835 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3.1. Overview", + "text_level": 1, + "bbox": [ + 500, + 845, + 609, + 861 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We propose a self-supervised method that can leverage properties that are unique to satellite images. One such", + "bbox": [ + 500, + 869, + 893, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "5262", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "property is the availability of long-term temporal information for any location on earth. For example, two images of the same location captured months apart will likely be the same except for seasonal variations; variations our features should be invariant to. In contrast, two images captured decades apart might show big, structural changes such as new buildings or sparser forests. Of course, this may not be true for all locations. Indeed, such long-term change is rare, and is concentrated in regions of human activity. But these long-term changes are important for many downstream applications such as modeling urban growth, or monitoring climate change. Our key insight is therefore to find these rare but meaningful long-term changes, and ensure that they are reflected as changes in the feature space as well.", + "bbox": [ + 75, + 90, + 472, + 301 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Past work on using temporal data for representation learning on satellite imagery only uses short-term seasonal variations but not long-term change [25]. We review this work, which is called Seasonal Contrast (or SeCo) [25], in Sec. 3.2. We then explain how we can use long-term temporal information in Sec. 3.3, and address the challenge of spatial rarity in Sec. 3.4. Finally, we look at how we can efficiently sample more informative locations when collecting unlabelled training data in Sec. 3.5.", + "bbox": [ + 75, + 301, + 472, + 439 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Background: Seasonal Contrast (SeCo)", + "text_level": 1, + "bbox": [ + 76, + 446, + 416, + 465 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Contrastive learning techniques train feature representations from unlabeled images by discriminating between individual instances. In particular, they pull augmentations of the same image closer together in feature space and push apart representations of two different images. SeCo is a contrastive learning-based representation learning framework for remote sensing images. SeCo uses MoCo v2 [8], to perform contrastive learning. In MoCo v2, in each iteration, one generates two views for an image using random augmentations. One is the \"query\" image $I_{q}$ and the other is the \"positive key\" $I_{k+}$ that is pulled closer to the query in the representation space. It also has a set of other images from the dataset, $\\mathcal{I}_{k-}$ , that are treated as negative keys and are pushed apart from query image $I_{q}$ in the latent space. The precise loss function in MoCo v2 called InfoNCE [29], for a representation function $f$ , can be written as:", + "bbox": [ + 75, + 470, + 472, + 714 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = - \\log \\frac {\\exp \\left(f \\left(I _ {q}\\right) \\cdot f \\left(I _ {k ^ {+}}\\right) / \\tau\\right)}{\\sum_ {I _ {k} \\in \\mathcal {I} _ {k ^ {-}} \\cup \\{I _ {k +} \\}} \\exp \\left(f \\left(I _ {q}\\right) \\cdot f \\left(I _ {k}\\right) / \\tau\\right)} \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 114, + 723, + 468, + 768 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The key idea in SeCo is to also use short-term temporal differences (over a few months) at a location as augmentations. Since these images are a few months apart, these images represent different seasons (hence the name seasonal contrast). The negative keys are then images from other locations. More specifically SeCo trains a representation with 3 subspaces, i) a subspace invariant to both season and artificial augmentations, ii) a subspace invariant to season only,", + "bbox": [ + 75, + 779, + 472, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "and iii) a subspace invariant to augmentation only. The additional invariance to seasonal differences results in a better representation than simply using MoCo v2 on this data.", + "bbox": [ + 496, + 90, + 892, + 137 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.3. Using Long-term Temporal Information", + "text_level": 1, + "bbox": [ + 498, + 154, + 844, + 171 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "SeCo introduces invariance to seasonal variations. However, invariance is not enough. We also want our feature representation to be informative. For this, simply pushing apart images from different locations (which already look quite different) may not be enough.", + "bbox": [ + 496, + 180, + 890, + 256 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "As discussed above, we argue that to produce a richer feature representation, we should use information from longer temporal spans. Over the span of years, we might observe big, structural changes such as new construction or urban growth. We want our feature representation to be sensitive to these major changes. Thus, while short-term temporal differences are typically seasonal and should be treated as augmentations as in SeCo, long-term temporal differences might correspond to permanent change, and should be reflected as a significant change in the underlying feature representation. This suggests that for a query image from a particular location, an image captured at the same location several years later should be treated as a negative in the contrastive learning objective. We formalize this below.", + "bbox": [ + 496, + 258, + 892, + 470 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Data: Let $\\{l_1, l_2 \\dots l_n\\}$ be $n$ different locations in our dataset. For each location, we have 2 sets of images captured around time points $t_1$ and $t_2$ that have a gap of 4 years between them. We represent these sets by $\\mathcal{I}_{l_i}^{t_1}$ and $\\mathcal{I}_{l_i}^{t_2}$ . Each set $\\mathcal{I}_{l_i}^{t_j}$ , contains images sampled over a few months around $t_j$ , and thus these images have only short-term or seasonal differences between them. See Fig. 2 (a) for a schematic of this data sampling strategy.", + "bbox": [ + 496, + 473, + 893, + 598 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Loss: Let $I_{l_i}^{t_j + \\Delta_1}, I_{l_i}^{t_j + \\Delta_2} \\sim \\mathcal{I}_{l_i}^{t_j}$ be two images sampled randomly from $\\mathcal{I}_{l_i}^{t_j}$ ( $j$ is 1 or 2). $I_{l_i}^{t_j + \\Delta_1}$ and $I_{l_i}^{t_j + \\Delta_2}$ are a few months apart. So like SeCo, we still want their representations to be pulled closer. However, images with larger time differences, for example, $I_{l_i}^{t_1 + \\Delta_1}$ and $I_{l_i}^{t_2 + \\Delta_1}$ should potentially be pushed apart. Finally, images from two different locations such as $I_{l_i}^{t_j + \\Delta_1}$ and $I_{l_j}^{t_j + \\Delta_1}$ , should always be pushed apart irrespective of time. This intuition is reflected in Fig. 2 (c).", + "bbox": [ + 496, + 599, + 893, + 743 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Following SeCo, we implement this loss function using MoCo v2 by adapting the positive key image $I_{k+}$ and negative key set $\\mathcal{I}_{k-}$ in Eq. (1). As discussed above, SeCo has 3 subspaces. We add our modifications to all 3 subspaces. However, for ease of explanation, we only describe the change to the subspace that is invariant to seasons. For this subspace, for a query image, $I_q = I_{l_i}^{t_1 + \\Delta_1}$ , the positive key is the same as in SeCo: the seasonal pair $(I_{k^+} = I_{l_i}^{t_1 + \\Delta_2})$ . But we modify the negative key set to not only include images from other locations but also images from the same", + "bbox": [ + 496, + 746, + 893, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "5263", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/e3ad505694a5a6acdc45320509ab95a7eaec1cdf354572ba08e1a76180b7b845.jpg", + "image_caption": [ + "Figure 2. Overview of our self-supervised framework. (a) Shows the notation for images and the temporal sampling strategy we use to create our dataset for training. (b) Shows the contrastive learning performed by SeCo in the temporally invariant subspace. The red arrows show examples that need to be pushed apart from the top left example. The green arrows indicate examples to be pulled closer. Different lightness (value) represents different short-term temporal pairs and different colors (hue) indicate different locations (see Sec. 3.2). (c) Shows our method when using long-term temporal pairs. Different shapes indicate a difference of 4 years between examples (see Sec. 3.3). (d) Shows the contrastive learning performed by our CACo loss term. The dashed red arrows indicate examples, that are pushed apart iff our method estimates a change has occurred between the long-term pairs (see Sec. 3.4)." + ], + "image_footnote": [], + "bbox": [ + 156, + 84, + 460, + 369 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/5083f2133d87bab85501d07caa55bca49e1d4029e352144e91d68b6572c31317.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 468, + 83, + 637, + 367 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/26811b45b3a8928490894dcb740b3cc094ddec8118ec3dc40bef9afc50ce1c1c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 645, + 85, + 812, + 366 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "location with a large time difference as follows:", + "bbox": [ + 76, + 503, + 392, + 517 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {I} _ {k ^ {-}} = \\left\\{I _ {l _ {i}} ^ {t _ {2} + \\Delta_ {k}}: k \\in \\{1, 2 \\} \\right\\} \\cup \\\\ \\left\\{I _ {l _ {h}} ^ {t _ {j} + \\Delta_ {k}}: h \\neq i, j \\in \\{1, 2 \\}, k \\in \\{1, 2 \\} \\right\\} \\quad (2) \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 91, + 530, + 467, + 573 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In each training iteration, we sample 4 images (long and short-time pairs) for each location instead of 2 like SeCo, so each batch contains half as many locations.", + "bbox": [ + 75, + 582, + 467, + 626 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.4. Change-aware Contrastive Learning", + "text_level": 1, + "bbox": [ + 76, + 635, + 393, + 651 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The training method in the previous section assumes that regions always change after a long time (a few years). However, this assumption is not always true: in remote areas away from human activity, there may be no change even over several years. For such locations, pushing apart features of temporally distant image pairs may be counterproductive and destroy needed invariances.", + "bbox": [ + 75, + 659, + 467, + 763 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "If we know whether a location has changed or not, we can make our loss function conditional on this information. If it has changed between $I_{l_i}^{t_1 + \\Delta_1}$ and $I_{l_i}^{t_2 + \\Delta_1}$ , we push them apart as described in Sec. 3.3. If not, we refrain from adding the temporally distant image to the negative set. This conditional loss function is illustrated in Fig. 2 (d).", + "bbox": [ + 75, + 765, + 467, + 854 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "However, the challenge is knowing which locations have changed. Change detection can be performed by looking at feature differences between temporally separated image", + "bbox": [ + 75, + 854, + 467, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "pairs [21, 23, 35, 42], but this requires a feature representation we do not have. We solve this chicken-and-egg problem by bootstrapping our feature representation. We start from a randomly initialized model, using which we extract features for different locations and use feature differences as an estimate for change. Using these change values, we use our conditional loss to train the model. We then keep alternating between training the model using changes and estimating the changes using the partially trained model. In the initial stages, the change estimates could be very poor as the model has random weights. However, as training progresses we expect better features, and, in turn, better estimates of changes. One might worry about degenerate solutions, but the other contrastive loss terms based on location (that do not depend on this inferred change) ensure that the feature representation avoids any pathological cases.", + "bbox": [ + 496, + 503, + 890, + 744 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Obtaining changes using features. The absolute distance between features scales differently for different types of locations, and does not directly reflect the magnitude of actual change. For example, urban locations with less vegetation look the same over the seasons, but temperate forests will change a lot with the seasons. So instead of using the absolute distance between two long-term images, we instead use the ratio of distances between long-term images and short-term images (images with a seasonal difference).", + "bbox": [ + 496, + 763, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "5264", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "This ratio normalizes the scaling differences between images from different locations. If $f_{k}$ is the partially trained feature extractor at epoch $k$ , we use the following ratio as an indicator of change for location $l_{i}$", + "bbox": [ + 75, + 90, + 468, + 151 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nR _ {l _ {i}} ^ {k} = \\frac {\\left| \\left| f _ {k} \\left(I _ {l _ {i}} ^ {t _ {1} + \\Delta_ {1}}\\right) - f _ {k} \\left(I _ {l _ {i}} ^ {t _ {2} + \\Delta_ {1}}\\right) \\right| \\right| ^ {2}}{\\left| \\left| f _ {k} \\left(I _ {l _ {i}} ^ {t _ {1} + \\Delta_ {1}}\\right) - f _ {k} \\left(I _ {l _ {i}} ^ {t _ {1} + \\Delta_ {2}}\\right) \\right| \\right| ^ {2}} \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 147, + 159, + 468, + 200 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Since, at each epoch, the seasonal images are selected randomly for each location, we use an exponential moving average over ratios from previous epochs to obtain a more stable estimate for change as follows", + "bbox": [ + 75, + 209, + 468, + 268 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {r} _ {l _ {i}} ^ {k} = (1 - \\beta) R _ {l _ {i}} ^ {k} + \\beta \\mathbf {r} _ {l _ {i}} ^ {k - 1}\n$$\n", + "text_format": "latex", + "bbox": [ + 181, + 277, + 361, + 299 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We use a Gaussian Mixture Model (GMM) on $\\mathbf{r}_{l_i}^k$ values to find our two clusters for change and no change. Locations with smaller $\\mathbf{r}_{l_i}^k$ are treated as having no long-term change, and those with larger $\\mathbf{r}_{l_i}^k$ values are deemed to have undergone significant change (see Fig. 3).", + "bbox": [ + 75, + 306, + 468, + 385 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For MoCo v2, we use features from the momentum encoder for images to calculate the ratio estimate. To maintain training speed, we compute and store the ratios during the training iterations themselves, even as the feature extractor is being updated. In preliminary experiments, we found that the exponential moving average was sufficient to regularize against the noise induced by the changing feature extractor.", + "bbox": [ + 75, + 386, + 468, + 491 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our change-aware formulation uses change information in long-term pairs. On the other hand, there could be short-term pairs with sudden changes such as due to disasters like landslides or earthquakes. However, we notice that in practice such changes are extremely rare. Out of the top 200 examples with the highest short-term change score only 3 pairs showed a sudden change. Because short-term changes are so rare, it is not useful to consider them in our pipeline.", + "bbox": [ + 75, + 491, + 468, + 612 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.5. Improving Geographical Sampling", + "text_level": 1, + "bbox": [ + 76, + 619, + 380, + 637 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Another property specific to satellite imagery is that we can control the geographic distribution of sampled images. Leveraging this freedom, SeCo collects data from a $100\\mathrm{km}$ radius around the 10k most populated cities of the world. However, we argue that 100 kms is too large a radius.", + "bbox": [ + 75, + 643, + 468, + 718 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Sampling far from cities results in uninformative images (such as images entirely over oceans, (see supplementary Fig. 3)). In fact, $22\\%$ of the dataset has such uninformative images and this hurts the quality of the representation.", + "bbox": [ + 75, + 719, + 468, + 779 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The median area of the 10k cities sampled is approximately $24\\mathrm{km}^2$ , which represents a radius of $\\sim 5\\mathrm{km}$ . Based on this observation, we use 2d-gaussian sampling with $\\sigma = 5\\mathrm{km}$ . This means that $95\\%$ of the data comes from a radius of $2\\sigma = 10\\mathrm{km}$ around the city center. Additionally, if the sampled location is completely in the ocean, we reject that sample and sample again (We use a landmass map of the earth to check this). Our improved sampling", + "bbox": [ + 75, + 780, + 468, + 901 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/14d85addedef9701805eb72ab54aaebe51556483a43dc3aab6932df45c543dd8.jpg", + "image_caption": [ + "Figure 3. The histogram of change estimates separated into two groups (blue and green) by the GMM. Pairs of examples with a higher estimate can be seen to have major construction and land-cover changes (2 pairs on the right). Whereas pairs with a lower estimate only have seasonal changes." + ], + "image_footnote": [], + "bbox": [ + 504, + 97, + 890, + 256 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "methods result in a dataset with a more diverse set of locations that are more informative. As we show in experiments, this additionally improves the quality of the representation.", + "bbox": [ + 498, + 364, + 890, + 411 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Results", + "text_level": 1, + "bbox": [ + 500, + 426, + 586, + 441 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Implementation details and Baselines", + "text_level": 1, + "bbox": [ + 500, + 452, + 821, + 468 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We evaluate our model by measuring its representation ability on various downstream tasks. We evaluate the following baselines for comparison.", + "bbox": [ + 498, + 476, + 890, + 521 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Random init.: is randomly initialized and is either fin-tuned or linear probed for downstream tasks.", + "bbox": [ + 498, + 521, + 888, + 550 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "ImageNet: is pre-trained on supervised ImageNet [31].", + "bbox": [ + 500, + 551, + 866, + 566 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "MoCo v2: uses self-supervised learning without the temporal or change information.", + "bbox": [ + 500, + 566, + 888, + 595 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "SeCo: is Seasonal Contrast [25] and uses only short-term changes as augmentations.", + "bbox": [ + 500, + 597, + 888, + 626 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "GSSL: uses temporal and geographical information for contrastive learning [2]. Since the data this model was trained on is of different resolution retrain it on our data.", + "bbox": [ + 500, + 627, + 888, + 671 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "SatMAE [10]: is a Masked autoencoder [17] that uses ViT-Large as backbone. The comparison is not apples-to-apples as the model has larger capacity. The model is trained on 713k images therefore it is comparable to our 1m images trained model.", + "bbox": [ + 500, + 672, + 888, + 746 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "DDPM-CD [4]: uses a denoising diffusion model as backbone. It is also trained on a million scale dataset.", + "bbox": [ + 500, + 748, + 888, + 777 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We show results with both a ResNet-18 and a ResNet-50 backbone. Similar to SeCo our models are trained with a batch size of 256 and with 16,384 negative embeddings. We also use the same optimizer, learning rate, schedule, and temperature scaling as in the SeCo implementation.", + "bbox": [ + 498, + 794, + 890, + 869 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We experiment with two datasets for self-supervised pretraining: one with 100k images and the other with 1 million", + "bbox": [ + 500, + 869, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5265", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/3753e9df59c1b5ef9003ed6b056d524b4a83c9df27aadd8e40f6d81d2e283c59.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
DataPre-trainingResNet-18ResNet-50
-Random init.64.2155.32
ImageNet.86.1689.08
100kMoCo v287.2289.75
GSSL87.7490.19
SeCo90.0593.12
CACo (ours)93.0894.48
1mSeCo93.9995.63
SatMAE (ViT-L)-93.03
DDPM-CD-87.67
CACo (ours)94.7295.90
", + "bbox": [ + 106, + 89, + 442, + 277 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "images. These datasets contain RGB images from Sentinel2 [13]. The long-time difference between the two sets $\\mathcal{I}_{l_i}^{t_1}$ and $\\mathcal{I}_{l_i}^{t_2}$ is 4 years. The maximum time difference within a set $\\mathcal{I}_{l_i}^{t_j}$ is 1 year. We train our model and baselines for 1000 and 200 epochs on the 100k and 1m dataset respectively.", + "bbox": [ + 75, + 359, + 468, + 441 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Landcover classification", + "text_level": 1, + "bbox": [ + 76, + 452, + 299, + 467 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Nine (out of seventeen) of the UN's Sustainable Development Goals require global monitoring of land cover [5]. Thus, improving the performance of models for landcover classification would result in better monitoring of these goals. We evaluate our method on two datasets for landcover classification: EuroSat [19] and BigEarthNet [34].", + "bbox": [ + 75, + 474, + 468, + 566 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "EuroSat has 10-classes with a total of $27\\mathrm{k}$ images $64\\times 64$ images from Sentinel-2. We use the same train/val split proposed by the dataset. We add a linear layer to the frozen pretrained backbone, to perform the linear evaluation. More details about the training are in the supplementary.", + "bbox": [ + 75, + 566, + 468, + 643 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Tab. 1 shows the top-1 classification accuracy of various pre-trained backbones on the EuroSat dataset. We first note that all satellite image-specific methods outperform generic pretraining on ImageNet, suggesting the importance of training a representation specifically for satellite imagery. They also outperform the MoCo v2 baseline trained on satellite images, indicating the importance of well-designed loss functions that use the structure of satellite imagery.", + "bbox": [ + 75, + 643, + 468, + 763 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Compared to SeCo, our approach (CACo) results in 3 points of improvement with the linear classifier (with ResNet-18 and 100k data). This validates our insight that invariance to seasonal changes is not enough; the features must also be trained to be sensitive to long-term change. In fact, CACo trained on 100k images produces a representation competitive with SeCo trained on 1m images. Thus our method can save an order of magnitude of images and training time over the SeCo baseline while producing rep", + "bbox": [ + 75, + 763, + 470, + 901 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/4822b7e3546351725b1b6f63d4cf9f29559b4ba3884077147217967eefacd12c.jpg", + "table_caption": [ + "Table 1. Performance of our representation on the EuroSat classification task with linear probing, in top-1 Accuracy. Our method provides a more accurate classification, with different backbones." + ], + "table_footnote": [], + "table_body": "
DataPre-trainingResNet-18ResNet-50
10%100%10%100%
-Random init.42.8745.9544.7645.22
ImageNet.65.4366.4070.3671.37
100kMoCo v265.4367.2070.3872.88
GSSL65.7867.3670.6572.86
SeCo65.8067.4370.6973.42
CACo (ours)67.8969.4371.5573.63
1mSeCo67.6869.9572.8974.82
SatMAE--67.9369.45
DDPM-CD--65.4767.31
CACo (ours)68.6470.4173.4074.98
", + "bbox": [ + 514, + 89, + 880, + 292 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 2. Performance of our method on BigEarthNet landcover classification, in mean Average Precision (mAP). The two columns use different percentages of data (10% and 100%) of BigEarthNet data for training the linear layer.", + "bbox": [ + 496, + 303, + 893, + 359 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "resentations of similar quality.", + "bbox": [ + 498, + 388, + 702, + 402 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We also evaluate our method on the BigEarthNet dataset, which is a multi-label classification dataset with 17 classes. The dataset is significantly larger than EuroSat, containing 590k sentinel-2 image patches. We use the RGB bands to perform classification on this dataset. Since this is a multi-label classification problem, we use mean Average Precision (mAP) to evaluate the performance.", + "bbox": [ + 496, + 404, + 890, + 508 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Tab. 2 compares the performance of our model to the baselines on BigEarthNet1. Even on BigEarthNet, our method performs better than the baselines.", + "bbox": [ + 496, + 510, + 890, + 556 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Change detection", + "text_level": 1, + "bbox": [ + 498, + 566, + 669, + 584 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We also evaluate our model on the OSCD change detection dataset [6]. It contains 24 pairs of images from the Sentinel-2 satellites between 2015 and 2018. The dataset is split into 14 training images and 10 testing images.", + "bbox": [ + 496, + 592, + 890, + 652 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Our model architecture follows past work [11, 25]. The input to the model is an image pair, each of which is fed into the frozen, pre-trained feature extractor (we use the ResNet-18 models). We extract feature maps after each downsampling layer for each image. We then take the absolute difference between the corresponding feature maps from the two images. These differences are passed as input to a U-Net Decoder [30]. The U-Net decoder is then trained with change supervision. More details about the training and architecture can be found in the supplementary.", + "bbox": [ + 496, + 652, + 890, + 804 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Results are shown in Tab. 3. We report the F1-score obtained by using a threshold of 0.5. Our method results in better features for change detection resulting in about 4", + "bbox": [ + 496, + 805, + 890, + 849 + ], + "page_idx": 5 + }, + { + "type": "page_footnote", + "text": "1We obtained the SeCo number by training a model using the authors' publicly released code and dataset, but this does not match the published number [25]. Repeated attempts to contact the authors were unanswered.", + "bbox": [ + 500, + 862, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "5266", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/cf58b9c9c7dfab0d653e18f356bed71a39842dee8392424dbce7a86881cd0da3.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
BackbonePre-trainingPrec.Rec.F1-score
ResNet-18Random init.69.7318.2428.91
ImageNet.70.2223.5835.30
MoCo v262.2127.5738.21
GSSL62.2934.0844.06
SeCo64.1536.8946.84
CACo (ours)60.6842.9450.29
ResNet-50Random init.65.4622.4333.41
ImageNet.70.0530.9642.94
GSSL62.1636.8346.26
SeCo63.2138.2647.67
CACo (ours)62.8744.4952.11
", + "bbox": [ + 93, + 89, + 455, + 292 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/82522e8a8d1e396aa7e05e63c985f0ed64ea69f920ce71a6c559352270f2fa4c.jpg", + "table_caption": [ + "Table 3. Performance of our method on the Change Detection Task. We report the precision, recall and F1-score at a threshold of 0.5 for various ResNet backbones, trained on the 100k dataset." + ], + "table_footnote": [], + "table_body": "
Pre-traininglinearfinetuning
Random init.41.5338.62
ImageNet.43.7543.78
MoCo v247.9747.15
GSSL46.7748.10
SeCo46.8348.18
CACo (ours)50.2051.29
", + "bbox": [ + 155, + 366, + 393, + 488 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "points of improvement in the F1 scores for both ResNet-18 and ResNet-50. This large improvement on change detection is likely because we explicitly train our representation to be sensitive to long-term change.", + "bbox": [ + 75, + 573, + 468, + 636 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Semantic segmentation", + "text_level": 1, + "bbox": [ + 76, + 647, + 290, + 662 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We also evaluate our method on a satellite images segmentation task. DynamicEarthNet [36] is a dataset used to evaluate landcover segmentation over 7 classes. The dataset contains images from Planet satellites for 65 locations. Each location has $241024\\times 1024$ images with labels. The locations are split into 55 for training and 10 for testing.", + "bbox": [ + 75, + 671, + 468, + 762 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Similar to change detection model, we use a U-net with ResNet-18 as the encoder. But unlike them, we input a single image and use feature maps instead of feature differences in the decoder. Due to the dataset imbalance, we use Dice Loss [33] for training (see supplementary for details).", + "bbox": [ + 75, + 763, + 468, + 839 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Tab. 4 shows the mIoU (intersection over union averaged over classes) for our method on DynamicEarthNet validation images. Our method is better than SeCo by more than 3 points, and ImageNet pre-training by more than 6 points.", + "bbox": [ + 75, + 839, + 468, + 901 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/16bce06da81c71473b4cee01c28f05c2ba5f0b2b70f23ac60e798deaf3abe496.jpg", + "table_caption": [ + "Table 4. Performance of our method on the DynamicEarthNet segmentation task. We report the mIoU score for the ResNet-18 backbone, trained on the 100k dataset." + ], + "table_footnote": [], + "table_body": "
Pre-trainingCaiRoadCalFire
AP@50AP@400AP@10AP@40
ImageNet41.7932.2652.6144.99
MoCo v241.7833.0254.8348.20
SeCo39.6334.7262.8751.87
CACo (ours)44.3835.9965.7153.44
", + "bbox": [ + 506, + 88, + 885, + 200 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 5. Performance of our method on change event retrieval for the CaiRoad and CalFire benchmarks. We show average precision@K for ResNet-18 backbones trained on the 100k dataset.", + "bbox": [ + 498, + 210, + 890, + 253 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b6f88e96e848e5a967ce929aac2eed91c96404b013669a4247bf41fdb9f72221.jpg", + "image_caption": [ + "Figure 4. Precision@K of our backbone on CaiRoad (Left) and CalFire (Right) retrieval benchmark. Our method has better Precision@K than baselines for most K values on both benchmarks." + ], + "image_footnote": [], + "bbox": [ + 503, + 267, + 694, + 412 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d39a771990eeefec49e023640f899f36b0b5d13ac493286762531d3a4e712431.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 696, + 268, + 888, + 412 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.5. Evaluation on Change Events", + "text_level": 1, + "bbox": [ + 500, + 481, + 764, + 498 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We evaluate our method on a recently released change events benchmark [24]. Change events are spatio-temporal semantic structures that are caused by a real-world event. The goal is to classify these change events. The benchmark contains two datasets: one from Cairo where we find road construction events (called CaiRoad), and the other from California where we find forest fire events (CalFire).", + "bbox": [ + 496, + 505, + 890, + 609 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We use the model proposed by Mall et al. [24], taking in a sequence of images and binary change masks for consecutive pairs. The model encodes image pairs using ResNet-18, average pooling the feature map with the downscaled change mask as weights. The final feature vector is obtained by temporal averaging. The representation is learned using SimCLR on change events (see supplementary).", + "bbox": [ + 496, + 611, + 890, + 717 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We follow the same procedure, but the ResNet-18 backbone is initialized with the trained weights, and then finetuned per [24]. We evaluate by measuring average precision@K (AP@K) for different backbone architectures.", + "bbox": [ + 496, + 717, + 890, + 776 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Tab. 5 shows the AP@K for different backbones, on the CaiRoad and CalFire benchmarks. Our method results in better retrieval for both types of events. Fig. 4 shows the precision@K for various methods on these benchmarks.", + "bbox": [ + 496, + 777, + 890, + 838 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.6. Ablation", + "text_level": 1, + "bbox": [ + 500, + 847, + 602, + 861 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We evaluate the design choices made by our method in Sec. 3. We look at improvements made by different com", + "bbox": [ + 498, + 869, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "5267", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/5bee307083074c6feef0ceb38ce97374bf22718bd396aaeb722b7fb709ec6de7.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Pre-trainingEuroSat (Acc.)DynamicEarth-Net (mIoU)
SeCo90.0546.83
+ Improved Sampling91.4247.24
+ Long-term temporal Contrast92.5649.68
+ Change-awareness (Ours)93.0850.20
", + "bbox": [ + 81, + 88, + 465, + 186 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/17abcf2b9a645010ccc6c0cc8b1ce28c1e7d275bd87886ca2e99d68d923d8bc5.jpg", + "table_caption": [ + "Table 6. Performance with different sub-components of CACo on EuroSat classification, with ResNet-18 pre-trained on the 100k dataset. All three of our novel contributions lead to an improvement in performance on downstream tasks." + ], + "table_footnote": [], + "table_body": "
Pre-trainingEuroSat (Acc.)OSCD (F1)
SeCo87.9744.05
CACo90.8347.50
", + "bbox": [ + 127, + 272, + 418, + 330 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "ponents such as better data, long-term temporal information, and change estimates. We also show that our method generalizes to other self-supervised frameworks.", + "bbox": [ + 75, + 414, + 468, + 460 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Are the new components of our method essential? In Tab. 6 we evaluate the impact of each component of our approach on the final performance as measured on EuroSat and DynamicEarthNet. We find that each component adds a significant improvement. It is telling that incorporating long-term temporal contrast has a particularly large impact on DynamicEarthNet, indicating that long-term contrast is especially needed for pixel-level localization tasks.", + "bbox": [ + 75, + 479, + 468, + 602 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Can our method work with other self-supervised frameworks? We replace the MoCo v2 framework with SimCLR [7]. SimCLR requires a large batch size to find more hard negative examples; we use a batch size of 512 instead of 256 for all the methods. Tab. 7 shows the performance of SeCo and CACo using SimCLR. Our model performs better than SeCo by 3 points when using SimCLR as well. This indicates that our insights are general and can be applied in the future to new self-supervision frameworks.", + "bbox": [ + 75, + 621, + 468, + 758 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "How useful is the ratio estimate? In Sec. 3.4 we proposed to estimate using $\\mathbf{r}_{l_i}^k$ to estimate changes. We now look at other estimates for change in Tab. 8. Directly using the feature distance between the long-term pair, instead of the ratio leads to an almost 2-point drop in performance. Removing the exponential moving average (i.e., using $R_{l_i}^k$ instead of $\\mathbf{r}_{l_i}^k$ ; see Eq. (3)) also reduces performance by about 1 point. Another alternative is to measure the ratio", + "bbox": [ + 75, + 777, + 468, + 902 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/5c931ec8dfbf712cae6771414891b044c9a3f37a89aba59a61e906bd3b6809aa.jpg", + "table_caption": [ + "Table 7. Accuracy on EuroSat when using a ResNet-18 backbone pre-trained using SimCLR instead of MoCo v2. Our insights generalize to other self-supervised frameworks like SimCLR as well." + ], + "table_footnote": [], + "table_body": "
Change EstimateEuroSat (Acc.)
long-term distance91.17
Rliki91.98
Align92.92
rkiLi (ours)93.08
", + "bbox": [ + 576, + 88, + 818, + 175 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 8. Performance of different change estimates with ResNet-18 on 100k dataset. $R_{l_i}^k$ (Eq. (3)) is the estimate without the moving average. Long-term distance is the numerator of the $R_{l_i}^k$ . Align uses seasonally aligned long-term pairs without smoothing.", + "bbox": [ + 498, + 188, + 890, + 247 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "with long-term pairs that are seasonally aligned. Align gives similar results to $\\mathbf{r}_{l_i}^k$ , but requires a more sophisticated sampling. In sum, these results suggest that how one identifies changed regions can have a big impact on performance.", + "bbox": [ + 498, + 272, + 890, + 334 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Discussion and Conclusions", + "text_level": 1, + "bbox": [ + 500, + 345, + 756, + 363 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Potential Negative Societal Impact and Limitations. As in all visual recognition, there is the possibility of negative impact through violations of privacy. To mitigate this concern we intentionally use low-resolution satellite images (1 pixel $\\sim 10\\mathrm{m}$ ). The use of our proposed techniques for surveillance should be appropriately regulated. Our work currently is limited in its ability to generalize to multiple resolutions. Though object detection tasks use higher-resolution images for better object delineation, our representation trained on medium-resolution images cannot directly generalize to them; although we can retrain our method on high-resolution images for these tasks.", + "bbox": [ + 498, + 372, + 890, + 554 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Conclusions. We present a novel self-supervised framework to learn representations well suited for remote sensing applications. We introduce a new loss that robustly leverages the long-term temporal information readily available for satellite images. We also propose a better location sampling method to provide more informative data. Our evaluation on diverse downstream tasks shows that our approach (CACo) is very versatile and leads to better downstream accuracy than prior art (with relative improvements of $8.5\\%$ on change detection and $6.5\\%$ on semantic segmentation).", + "bbox": [ + 496, + 571, + 890, + 723 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements. This research is based upon work supported in part by the ODNI (IARPA) via 2021-20111000006, NSF 1900783, NSF 2144117, and NSF 2212084. The views and conclusions contained herein are those of the authors and should not be interpreted as necessarily representing the official policies, either expressed or implied, of ODNI, IARPA, or the US Government. The US Government is authorized to reproduce and distribute reprints for governmental purposes notwithstanding any copyright annotation therein.", + "bbox": [ + 496, + 739, + 890, + 892 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "5268", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 174, + 106 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Peri Akiva, Matthew Purri, and Matthew Leotta. Self-supervised material and texture representation learning for remote sensing tasks. In CVPR, 2022. 2", + "[2] Kumar Ayush, Burak Uzkent, Chenlin Meng, Kumar Tanmay, Marshall Burke, David Lobell, and Stefano Ermon. Geography-aware self-supervised learning. In CVPR, 2021. 2, 5", + "[3] Gaetan Bahl, Mehdi Bahri, and Florent Lafarge. Single-shot end-to-end road graph extraction. In CVPR EarthVision, 2022. 2", + "[4] Wele Gedara Chaminda Bandara, Nithin Gopalakrishnan Nair, and Vishal M Patel. Ddpm-cd: Remote sensing change detection using denoising diffusion probabilistic models. CoRR, 2022. 2, 5", + "[5] Sarah Carter and Martin Herold. Specifications of land cover datasets for SDG indicator monitoring. https://ggim.un.org/ Accessed: 11-8-2022. 6", + "[6] R. Caye Daudt, B. Le Saux, A. Boulch, and Y. Gousseau. Urban change detection for multispectral earth observation using convolutional neural networks. In IGARSS, 2018. 1, 6", + "[7] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In ICML, 2020. 2, 8", + "[8] Xinlei Chen, Haoqi Fan, Ross Girshick, and Kaiming He. Improved baselines with momentum contrastive learning. CoRR, 2020. 2, 3", + "[9] Gordon Christie, Neil Fendley, James Wilson, and Ryan Mukherjee. Functional map of the world. In CVPR, 2018. 2", + "[10] Yezhen Cong, Samar Khanna, Chenlin Meng, Patrick Liu, Erik Rozi, Yutong He, Marshall Burke, David B Lobell, and Stefano Ermon. Satmae: Pre-training transformers for temporal and multi-spectral satellite imagery. In NeurIPS, 2022. 2, 5", + "[11] Rodrigo Caye Daudt, Bertr Le Saux, and Alexandre Boulch. Fully convolutional siamese networks for change detection. In ICIP, 2018. 2, 6", + "[12] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. CoRR, 2020. 2", + "[13] Matthias Drusch, Umberto Del Bello, Sébastien Carlier, Olivier Colin, Veronica Fernandez, Ferran Gascon, Bianca Hoersch, Claudia Isola, Paolo Laberinti, Philippe Martimort, et al. Sentinel-2: ESA's optical high-resolution mission for GMES operational services. Remote Sensing of Environment, 2012. 2, 6", + "[14] Jan Gąsienica-Józkowy, Mateusz Knapik, and Bogusław Cyganek. An ensemble deep learning method with optimized weights for drone-based water rescue and surveillance. *Integrated Computer-Aided Engineering*, 2021. 2", + "[15] Spyros Gidaris, Praveer Singh, and Nikos Komodakis. Unsupervised representation learning by predicting image rotations. In ICLR, 2018. 2", + "[16] Jean-Bastien Grill, Florian Strub, Florent Altché, Corentin Tallec, Pierre Richemond, Elena Buchatskaya, Carl Doersch," + ], + "bbox": [ + 78, + 114, + 468, + 902 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Bernardo Avila Pires, Zhaohan Guo, Mohammad Gheshlaghi Azar, et al. Bootstrap your own latent-a new approach to self-supervised learning. NeurIPS, 2020. 2", + "[17] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollár, and Ross Girshick. Masked autoencoders are scalable vision learners. In CVPR, 2022. 2, 5", + "[18] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In CVPR, 2020. 2", + "[19] Patrick Helber, Benjamin Bischke, Andreas Dengel, and Damian Borth. Eurosat: A novel dataset and deep learning benchmark for land use and land cover classification. JS-TAEORS, 2019. 2, 6", + "[20] Neal Jean, Sherrie Wang, Anshul Samar, George Azzari, David Lobell, and Stefano Ermon. Tile2vec: Unsupervised representation learning for spatially distributed data. In AAAI, 2019. 2", + "[21] Marrit Leenstra, Diego Marcos, Francesca Bovolo, and Devis Tuia. Self-supervised pre-training enhances change detection in sentinel-2 imagery. In ICPR, 2021. 4", + "[22] Yahui Liu, Jian Yao, Xiaohu Lu, Menghan Xia, Xingbo Wang, and Yuan Liu. RoadNet: Learning to comprehensively analyze road networks in complex urban scenes from high-resolution remotely sensed images. TGRS, 2018. 2", + "[23] William A Malila. Change vector analysis: an approach for detecting forest changes with landsat. In LARS symposia, 1980. 4", + "[24] Utkarsh Mall, Bharath Hariharan, and Kavita Bala. Change event dataset for discovery from spatio-temporal remote sensing imagery. In NeurIPS, 2022. 2, 7", + "[25] Oscar Manas, Alexandre Lacoste, Xavier Giroi-nieto, David Vazquez, and Pau Rodriguez. Seasonal contrast: Unsupervised pre-training from uncurated remote sensing data. In ICCV, 2021. 1, 2, 3, 5, 6", + "[26] Ishan Misra and Laurens van der Maaten. Self-supervised learning of pretext-invariant representations. In CVPR, 2020. 2", + "[27] Sorour Mohajerani and Parvaneh Saeedi. Cloud and cloud shadow segmentation for remote sensing imagery via filtered jaccard loss function and parametric augmentation. J-STARS, 2021. 2", + "[28] Mehdi Noroozi and Paolo Favaro. Unsupervised learning of visual representations by solving jigsaw puzzles. In ECCV, 2016. 2", + "[29] Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding. CoRR, 2018. 3", + "[30] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In ICMICCAI, pages 234-241. Springer, 2015. 6", + "[31] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael S. Bernstein, Alexander C. Berg, and Li Fei-Fei. Imagenet large scale visual recognition challenge. IJCV, 2015. 5", + "[32] Linus Scheibenreif, Joëlle Hanna, Michael Mommert, and Damian Borth. Self-supervised vision transformers for land-" + ], + "bbox": [ + 501, + 92, + 890, + 902 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "5269", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "cover segmentation and classification. In CVPR EarthVision, 2022. 2", + "[33] Carole H Sudre, Wenqi Li, Tom Vercauteren, Sebastien Ourselin, and M Jorge Cardoso. Generalised dice overlap as a deep learning loss function for highly unbalanced segmentations. In DLMIA, 2017. 7", + "[34] Gencer Sumbul, Marcela Charfuelan, Begüm Demir, and Volker Markl. Bigearthnet: A large-scale benchmark archive for remote sensing image understanding. In IGARSS, 2019. 2, 6", + "[35] Xu Tang, Huayu Zhang, Lichao Mou, Fang Liu, Xiangrong Zhang, Xiao Xiang Zhu, and Licheng Jiao. An unsupervised remote sensing change detection method based on multiscale graph convolutional network and metric learning. TGRSS, 2021. 4", + "[36] Aysim Token, Lukas Kondmann, Mark Weber, Marvin Eisenberger, Andrés Camero, Jingliang Hu, Ariadna Pregel Hoderlein, Caglar Senaras, Timothy Davis, Daniel Cremers, et al. Dynamicearthnet: Daily multi-spectral satellite dataset for semantic change segmentation. In CVPR, 2022. 1, 2, 7", + "[37] Gabriel Tseng, Ivan Zvonkov, Catherine Lilian Nakalembe, and Hannah Kerner. Cropharvest: A global dataset for croptype classification. In NeurIPS, 2021. 2", + "[38] Adam Van Etten, Dave Lindenbaum, and Todd M Bacastow. Spacenet: A remote sensing dataset and challenge series. CoRR, 2018. 2", + "[39] Adam Van Etten, Jacob Shermeyer, Daniel Hogan, Nicholas Weir, and Ryan Lewis. Road network and travel time extraction from multiple look angles with spacenet data. In IGARSS, 2020. 2", + "[40] Junjue Wang, Zhuo Zheng, Ailong Ma, Xiaoyan Lu, and Yanfei Zhong. Loveda: A remote sensing land-cover dataset for domain adaptive semantic segmentation. arXiv preprint arXiv:2110.08733, 2021. 2", + "[41] Ben G Weinstein, Sergio Marconi, Stephanie A Bohlman, Alina Zare, Aditya Singh, Sarah J Graves, and Ethan P White. A remote sensing derived data set of 100 million individual tree crowns for the national ecological observatory network. *Elife*, 2021. 2", + "[42] Chen Wu, Hongruixuan Chen, Bo Do, and Liangpei Zhang. Unsupervised change detection in multi-temporal vhr images based on deep kernel pca convolutional mapping network. IEEE Transactions on Cybernetics, 2019. 4", + "[43] Zhirong Wu, Yuanjun Xiong, Stella X. Yu, and Dahua Lin. Unsupervised feature learning via non-parametric instance discrimination. In CVPR, 2018. 2", + "[44] Richard Zhang, Phillip Isola, and Alexei A Efros. Colorful image colorization. In ECCV, 2016. 2" + ], + "bbox": [ + 78, + 90, + 468, + 771 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "5270", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2023/Change-Aware Sampling and Contrastive Learning for Satellite Images/ef28aefe-c0fe-45ce-8981-6fc9107e6b77_model.json b/2023/Change-Aware Sampling and Contrastive Learning for Satellite Images/ef28aefe-c0fe-45ce-8981-6fc9107e6b77_model.json new file mode 100644 index 0000000000000000000000000000000000000000..ddb5183a4c829bc1d9e1c6a0dc32dd3d0af90d0d --- /dev/null +++ b/2023/Change-Aware Sampling and Contrastive Learning for Satellite Images/ef28aefe-c0fe-45ce-8981-6fc9107e6b77_model.json @@ -0,0 +1,2233 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.808, + 0.045 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.127, + 0.131, + 0.844, + 0.152 + ], + "angle": 0, + "content": "Change-Aware Sampling and Contrastive Learning for Satellite Images" + }, + { + "type": "text", + "bbox": [ + 0.229, + 0.181, + 0.338, + 0.198 + ], + "angle": 0, + "content": "Utkarsh Mall" + }, + { + "type": "text", + "bbox": [ + 0.416, + 0.182, + 0.565, + 0.198 + ], + "angle": 0, + "content": "Bharath Hariharan" + }, + { + "type": "text", + "bbox": [ + 0.411, + 0.2, + 0.56, + 0.217 + ], + "angle": 0, + "content": "Cornell University" + }, + { + "type": "text", + "bbox": [ + 0.312, + 0.219, + 0.655, + 0.234 + ], + "angle": 0, + "content": "{utkarshm, bharathh, kb}@cs.cornell.edu" + }, + { + "type": "text", + "bbox": [ + 0.645, + 0.182, + 0.741, + 0.197 + ], + "angle": 0, + "content": "Kavita Bala" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.269, + 0.313, + 0.284 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.302, + 0.474, + 0.619 + ], + "angle": 0, + "content": "Automatic remote sensing tools can help inform many large-scale challenges such as disaster management, climate change, etc. While a vast amount of spatio-temporal satellite image data is readily available, most of it remains unlabelled. Without labels, this data is not very useful for supervised learning algorithms. Self-supervised learning instead provides a way to learn effective representations for various downstream tasks without labels. In this work, we leverage characteristics unique to satellite images to learn better self-supervised features. Specifically, we use the temporal signal to contrast images with long-term and short-term differences, and we leverage the fact that satellite images do not change frequently. Using these characteristics, we formulate a new loss contrastive loss called Change-Aware Contrastive (CACo) Loss. Further, we also present a novel method of sampling different geographical regions. We show that leveraging these properties leads to better performance on diverse downstream tasks. For example, we see a \\(6.5\\%\\) relative improvement for semantic segmentation and an \\(8.5\\%\\) relative improvement for change detection over the best-performing baseline with our method." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.648, + 0.21, + 0.663 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.673, + 0.47, + 0.78 + ], + "angle": 0, + "content": "Our planet is surrounded by a large number of satellites constantly collecting images of the world. This massive trove of visual information can help monitor phenomena at the world-scale, and inform solutions to global problems such as climate change or loss of biodiversity. Automatic vision tools can help by, for example, monitoring land-use change over time [36] or the evolution of urban areas [6]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.78, + 0.469, + 0.856 + ], + "angle": 0, + "content": "However, training all these models requires labeled data. Unfortunately, labeling the massive trove of satellite images is expensive, more so than internet images because of the expertise necessary. This issue is exacerbated by the different label requirements of many monitoring applications." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.47, + 0.901 + ], + "angle": 0, + "content": "One way to alleviate this problem of limited labeled data is to use self-supervised learning techniques to learn a good feature representation from unlabeled satellite im" + }, + { + "type": "image", + "bbox": [ + 0.499, + 0.264, + 0.897, + 0.479 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.487, + 0.895, + 0.556 + ], + "angle": 0, + "content": "Figure 1. Images of the same location at three different times. Changes from 2016 to 2020 are due to major urban development, while those from July to September are seasonal variations. Our approach, CACo, learns features that are sensitive to the former but invariant to the latter." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.583, + 0.893, + 0.78 + ], + "angle": 0, + "content": "agery. This representation can then be further finetuned with much fewer labels for specific applications. Modern self-supervised learning approaches are based on contrastive learning. These techniques train a feature space so that each image in the dataset is embedded close to augmented versions of itself (e.g., with jittered colors) but far from other images of the dataset. A possible approach is to directly apply these techniques on satellite image datasets. However, the spatio-temporal structure of satellite imagery is much richer than the unstructured collections of internet images typically used in standard self-supervised learning. In this work, we ask, how can we best leverage the structure of satellite images for better self-supervised learning?" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.781, + 0.892, + 0.901 + ], + "angle": 0, + "content": "The first important aspect of the spatio-temporal structure of satellite images is the availability of multiple temporally-spaced images for the same location. Past work has used this structure to sample images spread over a few weeks or months from each location to encourage invariance to seasonal variations [25]. However, we can access images not just over a few months, but over many years. Over such long time spans, we often see significant, per" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "5261" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.212 + ], + "angle": 0, + "content": "manent change, such as the construction of houses, the drying of lakes, or the logging of forests (Fig. 1). While we want feature representations to be invariant to temporary, seasonal change, we want the representation to be sensitive to permanent, long-term change. To capture this intuition, we sample multiple images from the same location covering both short and long time spans and encourage invariance to the former but sensitivity to the latter." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.213, + 0.473, + 0.408 + ], + "angle": 0, + "content": "A second and more important aspect of satellite images is that permanent change is spatially rare. Change is concentrated near urban areas, but many parts of the planet see little change. When there is no change, we want our feature representation to be the same even over long time spans. We capture this insight with a novel strategy to robustly estimate whether or not there is a long-term change, even in the middle of training, by comparing long-term feature differences to short-term variations. We then design a loss function that is conditional on this change estimate: it encourages invariances to long-term differences depending on whether a change occurs or not. We call this novel loss function Change-aware Contrastive Loss (or CaCo)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.409, + 0.473, + 0.62 + ], + "angle": 0, + "content": "The above loss function uses the temporal structure of satellite images. We can also utilize geographical structure by carefully sampling the most informative locations on the planet. We provide an improvement over a previously proposed geographical sampling [25]. We show that sampling closer to cities, and ignoring samples completely in the ocean can result in a dataset much more useful for learning a general representation for various downstream tasks. We evaluate our new representation on a diverse set of downstream tasks such as landcover classification, semantic segmentation, and change detection. Our method achieves significant relative improvements (ranging from \\(6.5\\%\\) to \\(8.5\\%\\)) over the state-of-the-art for a variety of tasks such as segmentation and change detection." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.62, + 0.446, + 0.635 + ], + "angle": 0, + "content": "To summarize, we make the following contributions:" + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.644, + 0.469, + 0.705 + ], + "angle": 0, + "content": "- We propose a new self-supervised learning loss that uses long-term temporal information in satellite imagery to encourage invariance to seasonal variations but sensitivity to permanent, long-term changes." + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.709, + 0.469, + 0.784 + ], + "angle": 0, + "content": "- We introduce a novel approach to robustly estimate whether a location has undergone significant change by comparing long-term changes to seasonal variations. Our new change-aware loss function (CACo) uses this to decide when to encourage invariance." + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.788, + 0.469, + 0.817 + ], + "angle": 0, + "content": "- We use an improved geographical sampling that provides more diverse data for representation learning." + }, + { + "type": "list", + "bbox": [ + 0.096, + 0.644, + 0.469, + 0.817 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.83, + 0.22, + 0.845 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.472, + 0.901 + ], + "angle": 0, + "content": "Self-supervised Learning. Self-supervised learning methods can be used to learn a good general representation from a dataset without any labels. These methods obtain" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.895, + 0.274 + ], + "angle": 0, + "content": "supervisory signals or inductive bias from the unlabelled data itself. Earlier self-supervised works used \"pretext tasks\" such as rotation prediction [15], solving jigsaw puzzles [28], colorization [44], or missing data completion [17]. More recently contrastive learning methods have been shown to learn a better representation than pretext tasks. Methods for contrastive learning such as NPID [43], PIRL [26], MoCo [8, 18], SimCLR [7], and BYOL [16], use instance discrimination to learn a feature representation. While these methods can be applied even to satellite images, in this work we leverage the structure of satellite images that can be used for self-supervised learning." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.292, + 0.895, + 0.399 + ], + "angle": 0, + "content": "Self-supervised Learning in Remote Sensing. Increasingly, self-supervised learning is being applied to the area of remote sensing. Prior works have used signals from unlabeled data such as location [2, 20], seasonal variations [25], or texture [1], as a signal for self-supervision. Recent methods use transformers [12] and diffusion models [4] for self-supervised representation [10, 32]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.416, + 0.895, + 0.673 + ], + "angle": 0, + "content": "Remote Sensing Applications and Data Several remote sensing applications require vision tools for automatic recognition at scale. Applications such as landcover classification [19, 34] or segmentation [36, 40] benefit from the advances in vision algorithms for scene classification and semantic segmentation. Several object detection applications have also been developed using object detection algorithms such as for detecting constructions [9], buildings [38], trees [41], or floating objects [14]. Segmentation methods also help in extracting structural information like road networks [3, 22, 39], crop types [37], or clouds [27] from remote sensing images. Many applications such as change detection [11] require temporal information for regions as well. Temporal information and changes are also used to find long-term semantic events [24]. In this work, we aim to learn a representation without supervision, that leads to better performance on all these downstream tasks." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.673, + 0.895, + 0.81 + ], + "angle": 0, + "content": "The image data for these algorithms and datasets are obtained from a few different sources. Sentinel-2 [13] provides multispectral images (13 bands) at \\(10\\mathrm{m}\\) resolution with a temporal revisit of 5 days. Several other satellites provide higher resolution information such as WorldView-3 and PlanetScope. We use Sentinel-2 satellite imagery as it provides frequent temporal information that we use in our self-supervised formulation. Furthermore, many of the downstream tasks [11, 19, 24, 36] use Sentinel-2 imagery." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.822, + 0.593, + 0.837 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.847, + 0.611, + 0.862 + ], + "angle": 0, + "content": "3.1. Overview" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.894, + 0.903 + ], + "angle": 0, + "content": "We propose a self-supervised method that can leverage properties that are unique to satellite images. One such" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5262" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.302 + ], + "angle": 0, + "content": "property is the availability of long-term temporal information for any location on earth. For example, two images of the same location captured months apart will likely be the same except for seasonal variations; variations our features should be invariant to. In contrast, two images captured decades apart might show big, structural changes such as new buildings or sparser forests. Of course, this may not be true for all locations. Indeed, such long-term change is rare, and is concentrated in regions of human activity. But these long-term changes are important for many downstream applications such as modeling urban growth, or monitoring climate change. Our key insight is therefore to find these rare but meaningful long-term changes, and ensure that they are reflected as changes in the feature space as well." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.303, + 0.473, + 0.44 + ], + "angle": 0, + "content": "Past work on using temporal data for representation learning on satellite imagery only uses short-term seasonal variations but not long-term change [25]. We review this work, which is called Seasonal Contrast (or SeCo) [25], in Sec. 3.2. We then explain how we can use long-term temporal information in Sec. 3.3, and address the challenge of spatial rarity in Sec. 3.4. Finally, we look at how we can efficiently sample more informative locations when collecting unlabelled training data in Sec. 3.5." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.448, + 0.418, + 0.466 + ], + "angle": 0, + "content": "3.2. Background: Seasonal Contrast (SeCo)" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.472, + 0.473, + 0.715 + ], + "angle": 0, + "content": "Contrastive learning techniques train feature representations from unlabeled images by discriminating between individual instances. In particular, they pull augmentations of the same image closer together in feature space and push apart representations of two different images. SeCo is a contrastive learning-based representation learning framework for remote sensing images. SeCo uses MoCo v2 [8], to perform contrastive learning. In MoCo v2, in each iteration, one generates two views for an image using random augmentations. One is the \"query\" image \\( I_{q} \\) and the other is the \"positive key\" \\( I_{k+} \\) that is pulled closer to the query in the representation space. It also has a set of other images from the dataset, \\( \\mathcal{I}_{k-} \\), that are treated as negative keys and are pushed apart from query image \\( I_{q} \\) in the latent space. The precise loss function in MoCo v2 called InfoNCE [29], for a representation function \\( f \\), can be written as:" + }, + { + "type": "equation", + "bbox": [ + 0.115, + 0.724, + 0.47, + 0.77 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = - \\log \\frac {\\exp \\left(f \\left(I _ {q}\\right) \\cdot f \\left(I _ {k ^ {+}}\\right) / \\tau\\right)}{\\sum_ {I _ {k} \\in \\mathcal {I} _ {k ^ {-}} \\cup \\{I _ {k +} \\}} \\exp \\left(f \\left(I _ {q}\\right) \\cdot f \\left(I _ {k}\\right) / \\tau\\right)} \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.78, + 0.473, + 0.903 + ], + "angle": 0, + "content": "The key idea in SeCo is to also use short-term temporal differences (over a few months) at a location as augmentations. Since these images are a few months apart, these images represent different seasons (hence the name seasonal contrast). The negative keys are then images from other locations. More specifically SeCo trains a representation with 3 subspaces, i) a subspace invariant to both season and artificial augmentations, ii) a subspace invariant to season only," + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.893, + 0.138 + ], + "angle": 0, + "content": "and iii) a subspace invariant to augmentation only. The additional invariance to seasonal differences results in a better representation than simply using MoCo v2 on this data." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.155, + 0.845, + 0.172 + ], + "angle": 0, + "content": "3.3. Using Long-term Temporal Information" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.181, + 0.892, + 0.257 + ], + "angle": 0, + "content": "SeCo introduces invariance to seasonal variations. However, invariance is not enough. We also want our feature representation to be informative. For this, simply pushing apart images from different locations (which already look quite different) may not be enough." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.259, + 0.893, + 0.472 + ], + "angle": 0, + "content": "As discussed above, we argue that to produce a richer feature representation, we should use information from longer temporal spans. Over the span of years, we might observe big, structural changes such as new construction or urban growth. We want our feature representation to be sensitive to these major changes. Thus, while short-term temporal differences are typically seasonal and should be treated as augmentations as in SeCo, long-term temporal differences might correspond to permanent change, and should be reflected as a significant change in the underlying feature representation. This suggests that for a query image from a particular location, an image captured at the same location several years later should be treated as a negative in the contrastive learning objective. We formalize this below." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.474, + 0.894, + 0.599 + ], + "angle": 0, + "content": "Data: Let \\(\\{l_1, l_2 \\dots l_n\\}\\) be \\(n\\) different locations in our dataset. For each location, we have 2 sets of images captured around time points \\(t_1\\) and \\(t_2\\) that have a gap of 4 years between them. We represent these sets by \\(\\mathcal{I}_{l_i}^{t_1}\\) and \\(\\mathcal{I}_{l_i}^{t_2}\\). Each set \\(\\mathcal{I}_{l_i}^{t_j}\\), contains images sampled over a few months around \\(t_j\\), and thus these images have only short-term or seasonal differences between them. See Fig. 2 (a) for a schematic of this data sampling strategy." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.6, + 0.894, + 0.744 + ], + "angle": 0, + "content": "Loss: Let \\( I_{l_i}^{t_j + \\Delta_1}, I_{l_i}^{t_j + \\Delta_2} \\sim \\mathcal{I}_{l_i}^{t_j} \\) be two images sampled randomly from \\( \\mathcal{I}_{l_i}^{t_j} \\) ( \\( j \\) is 1 or 2). \\( I_{l_i}^{t_j + \\Delta_1} \\) and \\( I_{l_i}^{t_j + \\Delta_2} \\) are a few months apart. So like SeCo, we still want their representations to be pulled closer. However, images with larger time differences, for example, \\( I_{l_i}^{t_1 + \\Delta_1} \\) and \\( I_{l_i}^{t_2 + \\Delta_1} \\) should potentially be pushed apart. Finally, images from two different locations such as \\( I_{l_i}^{t_j + \\Delta_1} \\) and \\( I_{l_j}^{t_j + \\Delta_1} \\), should always be pushed apart irrespective of time. This intuition is reflected in Fig. 2 (c)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.747, + 0.895, + 0.902 + ], + "angle": 0, + "content": "Following SeCo, we implement this loss function using MoCo v2 by adapting the positive key image \\( I_{k+} \\) and negative key set \\( \\mathcal{I}_{k-} \\) in Eq. (1). As discussed above, SeCo has 3 subspaces. We add our modifications to all 3 subspaces. However, for ease of explanation, we only describe the change to the subspace that is invariant to seasons. For this subspace, for a query image, \\( I_q = I_{l_i}^{t_1 + \\Delta_1} \\), the positive key is the same as in SeCo: the seasonal pair \\( (I_{k^+} = I_{l_i}^{t_1 + \\Delta_2}) \\). But we modify the negative key set to not only include images from other locations but also images from the same" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5263" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.158, + 0.085, + 0.462, + 0.37 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.469, + 0.084, + 0.638, + 0.368 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.647, + 0.087, + 0.813, + 0.367 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.379, + 0.892, + 0.477 + ], + "angle": 0, + "content": "Figure 2. Overview of our self-supervised framework. (a) Shows the notation for images and the temporal sampling strategy we use to create our dataset for training. (b) Shows the contrastive learning performed by SeCo in the temporally invariant subspace. The red arrows show examples that need to be pushed apart from the top left example. The green arrows indicate examples to be pulled closer. Different lightness (value) represents different short-term temporal pairs and different colors (hue) indicate different locations (see Sec. 3.2). (c) Shows our method when using long-term temporal pairs. Different shapes indicate a difference of 4 years between examples (see Sec. 3.3). (d) Shows the contrastive learning performed by our CACo loss term. The dashed red arrows indicate examples, that are pushed apart iff our method estimates a change has occurred between the long-term pairs (see Sec. 3.4)." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.504, + 0.393, + 0.518 + ], + "angle": 0, + "content": "location with a large time difference as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.092, + 0.531, + 0.468, + 0.574 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {I} _ {k ^ {-}} = \\left\\{I _ {l _ {i}} ^ {t _ {2} + \\Delta_ {k}}: k \\in \\{1, 2 \\} \\right\\} \\cup \\\\ \\left\\{I _ {l _ {h}} ^ {t _ {j} + \\Delta_ {k}}: h \\neq i, j \\in \\{1, 2 \\}, k \\in \\{1, 2 \\} \\right\\} \\quad (2) \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.583, + 0.468, + 0.627 + ], + "angle": 0, + "content": "In each training iteration, we sample 4 images (long and short-time pairs) for each location instead of 2 like SeCo, so each batch contains half as many locations." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.636, + 0.395, + 0.652 + ], + "angle": 0, + "content": "3.4. Change-aware Contrastive Learning" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.66, + 0.468, + 0.764 + ], + "angle": 0, + "content": "The training method in the previous section assumes that regions always change after a long time (a few years). However, this assumption is not always true: in remote areas away from human activity, there may be no change even over several years. For such locations, pushing apart features of temporally distant image pairs may be counterproductive and destroy needed invariances." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.766, + 0.468, + 0.855 + ], + "angle": 0, + "content": "If we know whether a location has changed or not, we can make our loss function conditional on this information. If it has changed between \\( I_{l_i}^{t_1 + \\Delta_1} \\) and \\( I_{l_i}^{t_2 + \\Delta_1} \\), we push them apart as described in Sec. 3.3. If not, we refrain from adding the temporally distant image to the negative set. This conditional loss function is illustrated in Fig. 2 (d)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.468, + 0.901 + ], + "angle": 0, + "content": "However, the challenge is knowing which locations have changed. Change detection can be performed by looking at feature differences between temporally separated image" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.504, + 0.892, + 0.745 + ], + "angle": 0, + "content": "pairs [21, 23, 35, 42], but this requires a feature representation we do not have. We solve this chicken-and-egg problem by bootstrapping our feature representation. We start from a randomly initialized model, using which we extract features for different locations and use feature differences as an estimate for change. Using these change values, we use our conditional loss to train the model. We then keep alternating between training the model using changes and estimating the changes using the partially trained model. In the initial stages, the change estimates could be very poor as the model has random weights. However, as training progresses we expect better features, and, in turn, better estimates of changes. One might worry about degenerate solutions, but the other contrastive loss terms based on location (that do not depend on this inferred change) ensure that the feature representation avoids any pathological cases." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Obtaining changes using features. The absolute distance between features scales differently for different types of locations, and does not directly reflect the magnitude of actual change. For example, urban locations with less vegetation look the same over the seasons, but temperate forests will change a lot with the seasons. So instead of using the absolute distance between two long-term images, we instead use the ratio of distances between long-term images and short-term images (images with a seasonal difference)." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "5264" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.152 + ], + "angle": 0, + "content": "This ratio normalizes the scaling differences between images from different locations. If \\( f_{k} \\) is the partially trained feature extractor at epoch \\( k \\), we use the following ratio as an indicator of change for location \\( l_{i} \\)" + }, + { + "type": "equation", + "bbox": [ + 0.148, + 0.16, + 0.469, + 0.202 + ], + "angle": 0, + "content": "\\[\nR _ {l _ {i}} ^ {k} = \\frac {\\left| \\left| f _ {k} \\left(I _ {l _ {i}} ^ {t _ {1} + \\Delta_ {1}}\\right) - f _ {k} \\left(I _ {l _ {i}} ^ {t _ {2} + \\Delta_ {1}}\\right) \\right| \\right| ^ {2}}{\\left| \\left| f _ {k} \\left(I _ {l _ {i}} ^ {t _ {1} + \\Delta_ {1}}\\right) - f _ {k} \\left(I _ {l _ {i}} ^ {t _ {1} + \\Delta_ {2}}\\right) \\right| \\right| ^ {2}} \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.21, + 0.469, + 0.27 + ], + "angle": 0, + "content": "Since, at each epoch, the seasonal images are selected randomly for each location, we use an exponential moving average over ratios from previous epochs to obtain a more stable estimate for change as follows" + }, + { + "type": "equation", + "bbox": [ + 0.182, + 0.279, + 0.362, + 0.3 + ], + "angle": 0, + "content": "\\[\n\\mathbf {r} _ {l _ {i}} ^ {k} = (1 - \\beta) R _ {l _ {i}} ^ {k} + \\beta \\mathbf {r} _ {l _ {i}} ^ {k - 1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.308, + 0.469, + 0.386 + ], + "angle": 0, + "content": "We use a Gaussian Mixture Model (GMM) on \\(\\mathbf{r}_{l_i}^k\\) values to find our two clusters for change and no change. Locations with smaller \\(\\mathbf{r}_{l_i}^k\\) are treated as having no long-term change, and those with larger \\(\\mathbf{r}_{l_i}^k\\) values are deemed to have undergone significant change (see Fig. 3)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.387, + 0.469, + 0.492 + ], + "angle": 0, + "content": "For MoCo v2, we use features from the momentum encoder for images to calculate the ratio estimate. To maintain training speed, we compute and store the ratios during the training iterations themselves, even as the feature extractor is being updated. In preliminary experiments, we found that the exponential moving average was sufficient to regularize against the noise induced by the changing feature extractor." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.492, + 0.469, + 0.613 + ], + "angle": 0, + "content": "Our change-aware formulation uses change information in long-term pairs. On the other hand, there could be short-term pairs with sudden changes such as due to disasters like landslides or earthquakes. However, we notice that in practice such changes are extremely rare. Out of the top 200 examples with the highest short-term change score only 3 pairs showed a sudden change. Because short-term changes are so rare, it is not useful to consider them in our pipeline." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.621, + 0.382, + 0.638 + ], + "angle": 0, + "content": "3.5. Improving Geographical Sampling" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.645, + 0.469, + 0.719 + ], + "angle": 0, + "content": "Another property specific to satellite imagery is that we can control the geographic distribution of sampled images. Leveraging this freedom, SeCo collects data from a \\(100\\mathrm{km}\\) radius around the 10k most populated cities of the world. However, we argue that 100 kms is too large a radius." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.72, + 0.469, + 0.78 + ], + "angle": 0, + "content": "Sampling far from cities results in uninformative images (such as images entirely over oceans, (see supplementary Fig. 3)). In fact, \\(22\\%\\) of the dataset has such uninformative images and this hurts the quality of the representation." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.781, + 0.469, + 0.902 + ], + "angle": 0, + "content": "The median area of the 10k cities sampled is approximately \\(24\\mathrm{km}^2\\), which represents a radius of \\(\\sim 5\\mathrm{km}\\). Based on this observation, we use 2d-gaussian sampling with \\(\\sigma = 5\\mathrm{km}\\). This means that \\(95\\%\\) of the data comes from a radius of \\(2\\sigma = 10\\mathrm{km}\\) around the city center. Additionally, if the sampled location is completely in the ocean, we reject that sample and sample again (We use a landmass map of the earth to check this). Our improved sampling" + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.098, + 0.892, + 0.257 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.268, + 0.892, + 0.338 + ], + "angle": 0, + "content": "Figure 3. The histogram of change estimates separated into two groups (blue and green) by the GMM. Pairs of examples with a higher estimate can be seen to have major construction and land-cover changes (2 pairs on the right). Whereas pairs with a lower estimate only have seasonal changes." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.366, + 0.892, + 0.412 + ], + "angle": 0, + "content": "methods result in a dataset with a more diverse set of locations that are more informative. As we show in experiments, this additionally improves the quality of the representation." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.427, + 0.587, + 0.442 + ], + "angle": 0, + "content": "4. Results" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.453, + 0.822, + 0.469 + ], + "angle": 0, + "content": "4.1. Implementation details and Baselines" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.477, + 0.891, + 0.522 + ], + "angle": 0, + "content": "We evaluate our model by measuring its representation ability on various downstream tasks. We evaluate the following baselines for comparison." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.522, + 0.89, + 0.551 + ], + "angle": 0, + "content": "Random init.: is randomly initialized and is either fin-tuned or linear probed for downstream tasks." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.552, + 0.867, + 0.568 + ], + "angle": 0, + "content": "ImageNet: is pre-trained on supervised ImageNet [31]." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.568, + 0.89, + 0.597 + ], + "angle": 0, + "content": "MoCo v2: uses self-supervised learning without the temporal or change information." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.598, + 0.89, + 0.627 + ], + "angle": 0, + "content": "SeCo: is Seasonal Contrast [25] and uses only short-term changes as augmentations." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.628, + 0.89, + 0.672 + ], + "angle": 0, + "content": "GSSL: uses temporal and geographical information for contrastive learning [2]. Since the data this model was trained on is of different resolution retrain it on our data." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.673, + 0.89, + 0.747 + ], + "angle": 0, + "content": "SatMAE [10]: is a Masked autoencoder [17] that uses ViT-Large as backbone. The comparison is not apples-to-apples as the model has larger capacity. The model is trained on 713k images therefore it is comparable to our 1m images trained model." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.749, + 0.89, + 0.779 + ], + "angle": 0, + "content": "DDPM-CD [4]: uses a denoising diffusion model as backbone. It is also trained on a million scale dataset." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.795, + 0.891, + 0.87 + ], + "angle": 0, + "content": "We show results with both a ResNet-18 and a ResNet-50 backbone. Similar to SeCo our models are trained with a batch size of 256 and with 16,384 negative embeddings. We also use the same optimizer, learning rate, schedule, and temperature scaling as in the SeCo implementation." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.891, + 0.901 + ], + "angle": 0, + "content": "We experiment with two datasets for self-supervised pretraining: one with 100k images and the other with 1 million" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5265" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.107, + 0.09, + 0.444, + 0.279 + ], + "angle": 0, + "content": "
DataPre-trainingResNet-18ResNet-50
-Random init.64.2155.32
ImageNet.86.1689.08
100kMoCo v287.2289.75
GSSL87.7490.19
SeCo90.0593.12
CACo (ours)93.0894.48
1mSeCo93.9995.63
SatMAE (ViT-L)-93.03
DDPM-CD-87.67
CACo (ours)94.7295.90
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.29, + 0.472, + 0.334 + ], + "angle": 0, + "content": "Table 1. Performance of our representation on the EuroSat classification task with linear probing, in top-1 Accuracy. Our method provides a more accurate classification, with different backbones." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.36, + 0.47, + 0.442 + ], + "angle": 0, + "content": "images. These datasets contain RGB images from Sentinel2 [13]. The long-time difference between the two sets \\(\\mathcal{I}_{l_i}^{t_1}\\) and \\(\\mathcal{I}_{l_i}^{t_2}\\) is 4 years. The maximum time difference within a set \\(\\mathcal{I}_{l_i}^{t_j}\\) is 1 year. We train our model and baselines for 1000 and 200 epochs on the 100k and 1m dataset respectively." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.453, + 0.3, + 0.468 + ], + "angle": 0, + "content": "4.2. Landcover classification" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.476, + 0.47, + 0.568 + ], + "angle": 0, + "content": "Nine (out of seventeen) of the UN's Sustainable Development Goals require global monitoring of land cover [5]. Thus, improving the performance of models for landcover classification would result in better monitoring of these goals. We evaluate our method on two datasets for landcover classification: EuroSat [19] and BigEarthNet [34]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.568, + 0.47, + 0.644 + ], + "angle": 0, + "content": "EuroSat has 10-classes with a total of \\(27\\mathrm{k}\\) images \\(64\\times 64\\) images from Sentinel-2. We use the same train/val split proposed by the dataset. We add a linear layer to the frozen pretrained backbone, to perform the linear evaluation. More details about the training are in the supplementary." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.644, + 0.47, + 0.765 + ], + "angle": 0, + "content": "Tab. 1 shows the top-1 classification accuracy of various pre-trained backbones on the EuroSat dataset. We first note that all satellite image-specific methods outperform generic pretraining on ImageNet, suggesting the importance of training a representation specifically for satellite imagery. They also outperform the MoCo v2 baseline trained on satellite images, indicating the importance of well-designed loss functions that use the structure of satellite imagery." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Compared to SeCo, our approach (CACo) results in 3 points of improvement with the linear classifier (with ResNet-18 and 100k data). This validates our insight that invariance to seasonal changes is not enough; the features must also be trained to be sensitive to long-term change. In fact, CACo trained on 100k images produces a representation competitive with SeCo trained on 1m images. Thus our method can save an order of magnitude of images and training time over the SeCo baseline while producing rep" + }, + { + "type": "table", + "bbox": [ + 0.515, + 0.09, + 0.882, + 0.293 + ], + "angle": 0, + "content": "
DataPre-trainingResNet-18ResNet-50
10%100%10%100%
-Random init.42.8745.9544.7645.22
ImageNet.65.4366.4070.3671.37
100kMoCo v265.4367.2070.3872.88
GSSL65.7867.3670.6572.86
SeCo65.8067.4370.6973.42
CACo (ours)67.8969.4371.5573.63
1mSeCo67.6869.9572.8974.82
SatMAE--67.9369.45
DDPM-CD--65.4767.31
CACo (ours)68.6470.4173.4074.98
" + }, + { + "type": "table_caption", + "bbox": [ + 0.498, + 0.304, + 0.895, + 0.361 + ], + "angle": 0, + "content": "Table 2. Performance of our method on BigEarthNet landcover classification, in mean Average Precision (mAP). The two columns use different percentages of data (10% and 100%) of BigEarthNet data for training the linear layer." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.389, + 0.704, + 0.404 + ], + "angle": 0, + "content": "resentations of similar quality." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.405, + 0.892, + 0.51 + ], + "angle": 0, + "content": "We also evaluate our method on the BigEarthNet dataset, which is a multi-label classification dataset with 17 classes. The dataset is significantly larger than EuroSat, containing 590k sentinel-2 image patches. We use the RGB bands to perform classification on this dataset. Since this is a multi-label classification problem, we use mean Average Precision (mAP) to evaluate the performance." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.511, + 0.892, + 0.557 + ], + "angle": 0, + "content": "Tab. 2 compares the performance of our model to the baselines on BigEarthNet1. Even on BigEarthNet, our method performs better than the baselines." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.568, + 0.671, + 0.585 + ], + "angle": 0, + "content": "4.3. Change detection" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.593, + 0.892, + 0.654 + ], + "angle": 0, + "content": "We also evaluate our model on the OSCD change detection dataset [6]. It contains 24 pairs of images from the Sentinel-2 satellites between 2015 and 2018. The dataset is split into 14 training images and 10 testing images." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.654, + 0.892, + 0.805 + ], + "angle": 0, + "content": "Our model architecture follows past work [11, 25]. The input to the model is an image pair, each of which is fed into the frozen, pre-trained feature extractor (we use the ResNet-18 models). We extract feature maps after each downsampling layer for each image. We then take the absolute difference between the corresponding feature maps from the two images. These differences are passed as input to a U-Net Decoder [30]. The U-Net decoder is then trained with change supervision. More details about the training and architecture can be found in the supplementary." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.806, + 0.892, + 0.851 + ], + "angle": 0, + "content": "Results are shown in Tab. 3. We report the F1-score obtained by using a threshold of 0.5. Our method results in better features for change detection resulting in about 4" + }, + { + "type": "page_footnote", + "bbox": [ + 0.5, + 0.863, + 0.892, + 0.901 + ], + "angle": 0, + "content": "1We obtained the SeCo number by training a model using the authors' publicly released code and dataset, but this does not match the published number [25]. Repeated attempts to contact the authors were unanswered." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5266" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.094, + 0.09, + 0.457, + 0.294 + ], + "angle": 0, + "content": "
BackbonePre-trainingPrec.Rec.F1-score
ResNet-18Random init.69.7318.2428.91
ImageNet.70.2223.5835.30
MoCo v262.2127.5738.21
GSSL62.2934.0844.06
SeCo64.1536.8946.84
CACo (ours)60.6842.9450.29
ResNet-50Random init.65.4622.4333.41
ImageNet.70.0530.9642.94
GSSL62.1636.8346.26
SeCo63.2138.2647.67
CACo (ours)62.8744.4952.11
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.307, + 0.472, + 0.351 + ], + "angle": 0, + "content": "Table 3. Performance of our method on the Change Detection Task. We report the precision, recall and F1-score at a threshold of 0.5 for various ResNet backbones, trained on the 100k dataset." + }, + { + "type": "table", + "bbox": [ + 0.156, + 0.367, + 0.394, + 0.489 + ], + "angle": 0, + "content": "
Pre-traininglinearfinetuning
Random init.41.5338.62
ImageNet.43.7543.78
MoCo v247.9747.15
GSSL46.7748.10
SeCo46.8348.18
CACo (ours)50.2051.29
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.502, + 0.47, + 0.545 + ], + "angle": 0, + "content": "Table 4. Performance of our method on the DynamicEarthNet segmentation task. We report the mIoU score for the ResNet-18 backbone, trained on the 100k dataset." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.574, + 0.47, + 0.637 + ], + "angle": 0, + "content": "points of improvement in the F1 scores for both ResNet-18 and ResNet-50. This large improvement on change detection is likely because we explicitly train our representation to be sensitive to long-term change." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.648, + 0.292, + 0.664 + ], + "angle": 0, + "content": "4.4. Semantic segmentation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.672, + 0.469, + 0.763 + ], + "angle": 0, + "content": "We also evaluate our method on a satellite images segmentation task. DynamicEarthNet [36] is a dataset used to evaluate landcover segmentation over 7 classes. The dataset contains images from Planet satellites for 65 locations. Each location has \\(241024\\times 1024\\) images with labels. The locations are split into 55 for training and 10 for testing." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.764, + 0.469, + 0.84 + ], + "angle": 0, + "content": "Similar to change detection model, we use a U-net with ResNet-18 as the encoder. But unlike them, we input a single image and use feature maps instead of feature differences in the decoder. Due to the dataset imbalance, we use Dice Loss [33] for training (see supplementary for details)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.469, + 0.902 + ], + "angle": 0, + "content": "Tab. 4 shows the mIoU (intersection over union averaged over classes) for our method on DynamicEarthNet validation images. Our method is better than SeCo by more than 3 points, and ImageNet pre-training by more than 6 points." + }, + { + "type": "table", + "bbox": [ + 0.508, + 0.089, + 0.887, + 0.201 + ], + "angle": 0, + "content": "
Pre-trainingCaiRoadCalFire
AP@50AP@400AP@10AP@40
ImageNet41.7932.2652.6144.99
MoCo v241.7833.0254.8348.20
SeCo39.6334.7262.8751.87
CACo (ours)44.3835.9965.7153.44
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.212, + 0.892, + 0.254 + ], + "angle": 0, + "content": "Table 5. Performance of our method on change event retrieval for the CaiRoad and CalFire benchmarks. We show average precision@K for ResNet-18 backbones trained on the 100k dataset." + }, + { + "type": "image", + "bbox": [ + 0.504, + 0.268, + 0.695, + 0.414 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.697, + 0.269, + 0.89, + 0.414 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.415, + 0.892, + 0.457 + ], + "angle": 0, + "content": "Figure 4. Precision@K of our backbone on CaiRoad (Left) and CalFire (Right) retrieval benchmark. Our method has better Precision@K than baselines for most K values on both benchmarks." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.482, + 0.765, + 0.499 + ], + "angle": 0, + "content": "4.5. Evaluation on Change Events" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.506, + 0.892, + 0.611 + ], + "angle": 0, + "content": "We evaluate our method on a recently released change events benchmark [24]. Change events are spatio-temporal semantic structures that are caused by a real-world event. The goal is to classify these change events. The benchmark contains two datasets: one from Cairo where we find road construction events (called CaiRoad), and the other from California where we find forest fire events (CalFire)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.612, + 0.892, + 0.718 + ], + "angle": 0, + "content": "We use the model proposed by Mall et al. [24], taking in a sequence of images and binary change masks for consecutive pairs. The model encodes image pairs using ResNet-18, average pooling the feature map with the downscaled change mask as weights. The final feature vector is obtained by temporal averaging. The representation is learned using SimCLR on change events (see supplementary)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.718, + 0.892, + 0.777 + ], + "angle": 0, + "content": "We follow the same procedure, but the ResNet-18 backbone is initialized with the trained weights, and then finetuned per [24]. We evaluate by measuring average precision@K (AP@K) for different backbone architectures." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.778, + 0.892, + 0.839 + ], + "angle": 0, + "content": "Tab. 5 shows the AP@K for different backbones, on the CaiRoad and CalFire benchmarks. Our method results in better retrieval for both types of events. Fig. 4 shows the precision@K for various methods on these benchmarks." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.848, + 0.604, + 0.862 + ], + "angle": 0, + "content": "4.6. Ablation" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.871, + 0.892, + 0.901 + ], + "angle": 0, + "content": "We evaluate the design choices made by our method in Sec. 3. We look at improvements made by different com" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5267" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.082, + 0.089, + 0.467, + 0.188 + ], + "angle": 0, + "content": "
Pre-trainingEuroSat (Acc.)DynamicEarth-Net (mIoU)
SeCo90.0546.83
+ Improved Sampling91.4247.24
+ Long-term temporal Contrast92.5649.68
+ Change-awareness (Ours)93.0850.20
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.202, + 0.47, + 0.258 + ], + "angle": 0, + "content": "Table 6. Performance with different sub-components of CACo on EuroSat classification, with ResNet-18 pre-trained on the 100k dataset. All three of our novel contributions lead to an improvement in performance on downstream tasks." + }, + { + "type": "table", + "bbox": [ + 0.129, + 0.273, + 0.419, + 0.332 + ], + "angle": 0, + "content": "
Pre-trainingEuroSat (Acc.)OSCD (F1)
SeCo87.9744.05
CACo90.8347.50
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.345, + 0.47, + 0.388 + ], + "angle": 0, + "content": "Table 7. Accuracy on EuroSat when using a ResNet-18 backbone pre-trained using SimCLR instead of MoCo v2. Our insights generalize to other self-supervised frameworks like SimCLR as well." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.415, + 0.47, + 0.461 + ], + "angle": 0, + "content": "ponents such as better data, long-term temporal information, and change estimates. We also show that our method generalizes to other self-supervised frameworks." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.481, + 0.47, + 0.603 + ], + "angle": 0, + "content": "Are the new components of our method essential? In Tab. 6 we evaluate the impact of each component of our approach on the final performance as measured on EuroSat and DynamicEarthNet. We find that each component adds a significant improvement. It is telling that incorporating long-term temporal contrast has a particularly large impact on DynamicEarthNet, indicating that long-term contrast is especially needed for pixel-level localization tasks." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.622, + 0.47, + 0.759 + ], + "angle": 0, + "content": "Can our method work with other self-supervised frameworks? We replace the MoCo v2 framework with SimCLR [7]. SimCLR requires a large batch size to find more hard negative examples; we use a batch size of 512 instead of 256 for all the methods. Tab. 7 shows the performance of SeCo and CACo using SimCLR. Our model performs better than SeCo by 3 points when using SimCLR as well. This indicates that our insights are general and can be applied in the future to new self-supervision frameworks." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.779, + 0.47, + 0.903 + ], + "angle": 0, + "content": "How useful is the ratio estimate? In Sec. 3.4 we proposed to estimate using \\(\\mathbf{r}_{l_i}^k\\) to estimate changes. We now look at other estimates for change in Tab. 8. Directly using the feature distance between the long-term pair, instead of the ratio leads to an almost 2-point drop in performance. Removing the exponential moving average (i.e., using \\(R_{l_i}^k\\) instead of \\(\\mathbf{r}_{l_i}^k\\); see Eq. (3)) also reduces performance by about 1 point. Another alternative is to measure the ratio" + }, + { + "type": "table", + "bbox": [ + 0.577, + 0.089, + 0.819, + 0.176 + ], + "angle": 0, + "content": "
Change EstimateEuroSat (Acc.)
long-term distance91.17
Rliki91.98
Align92.92
rkiLi (ours)93.08
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.189, + 0.892, + 0.248 + ], + "angle": 0, + "content": "Table 8. Performance of different change estimates with ResNet-18 on 100k dataset. \\( R_{l_i}^k \\) (Eq. (3)) is the estimate without the moving average. Long-term distance is the numerator of the \\( R_{l_i}^k \\). Align uses seasonally aligned long-term pairs without smoothing." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.273, + 0.892, + 0.335 + ], + "angle": 0, + "content": "with long-term pairs that are seasonally aligned. Align gives similar results to \\(\\mathbf{r}_{l_i}^k\\), but requires a more sophisticated sampling. In sum, these results suggest that how one identifies changed regions can have a big impact on performance." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.347, + 0.757, + 0.364 + ], + "angle": 0, + "content": "5. Discussion and Conclusions" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.373, + 0.892, + 0.555 + ], + "angle": 0, + "content": "Potential Negative Societal Impact and Limitations. As in all visual recognition, there is the possibility of negative impact through violations of privacy. To mitigate this concern we intentionally use low-resolution satellite images (1 pixel \\(\\sim 10\\mathrm{m}\\)). The use of our proposed techniques for surveillance should be appropriately regulated. Our work currently is limited in its ability to generalize to multiple resolutions. Though object detection tasks use higher-resolution images for better object delineation, our representation trained on medium-resolution images cannot directly generalize to them; although we can retrain our method on high-resolution images for these tasks." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.572, + 0.892, + 0.724 + ], + "angle": 0, + "content": "Conclusions. We present a novel self-supervised framework to learn representations well suited for remote sensing applications. We introduce a new loss that robustly leverages the long-term temporal information readily available for satellite images. We also propose a better location sampling method to provide more informative data. Our evaluation on diverse downstream tasks shows that our approach (CACo) is very versatile and leads to better downstream accuracy than prior art (with relative improvements of \\(8.5\\%\\) on change detection and \\(6.5\\%\\) on semantic segmentation)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.741, + 0.892, + 0.893 + ], + "angle": 0, + "content": "Acknowledgements. This research is based upon work supported in part by the ODNI (IARPA) via 2021-20111000006, NSF 1900783, NSF 2144117, and NSF 2212084. The views and conclusions contained herein are those of the authors and should not be interpreted as necessarily representing the official policies, either expressed or implied, of ODNI, IARPA, or the US Government. The US Government is authorized to reproduce and distribute reprints for governmental purposes notwithstanding any copyright annotation therein." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5268" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.079, + 0.09, + 0.176, + 0.107 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.156 + ], + "angle": 0, + "content": "[1] Peri Akiva, Matthew Purri, and Matthew Leotta. Self-supervised material and texture representation learning for remote sensing tasks. In CVPR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.157, + 0.47, + 0.212 + ], + "angle": 0, + "content": "[2] Kumar Ayush, Burak Uzkent, Chenlin Meng, Kumar Tanmay, Marshall Burke, David Lobell, and Stefano Ermon. Geography-aware self-supervised learning. In CVPR, 2021. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.213, + 0.47, + 0.255 + ], + "angle": 0, + "content": "[3] Gaetan Bahl, Mehdi Bahri, and Florent Lafarge. Single-shot end-to-end road graph extraction. In CVPR EarthVision, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.256, + 0.469, + 0.311 + ], + "angle": 0, + "content": "[4] Wele Gedara Chaminda Bandara, Nithin Gopalakrishnan Nair, and Vishal M Patel. Ddpm-cd: Remote sensing change detection using denoising diffusion probabilistic models. CoRR, 2022. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.312, + 0.469, + 0.353 + ], + "angle": 0, + "content": "[5] Sarah Carter and Martin Herold. Specifications of land cover datasets for SDG indicator monitoring. https://ggim.un.org/ Accessed: 11-8-2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.354, + 0.469, + 0.396 + ], + "angle": 0, + "content": "[6] R. Caye Daudt, B. Le Saux, A. Boulch, and Y. Gousseau. Urban change detection for multispectral earth observation using convolutional neural networks. In IGARSS, 2018. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.397, + 0.469, + 0.437 + ], + "angle": 0, + "content": "[7] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In ICML, 2020. 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.438, + 0.469, + 0.479 + ], + "angle": 0, + "content": "[8] Xinlei Chen, Haoqi Fan, Ross Girshick, and Kaiming He. Improved baselines with momentum contrastive learning. CoRR, 2020. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.48, + 0.469, + 0.509 + ], + "angle": 0, + "content": "[9] Gordon Christie, Neil Fendley, James Wilson, and Ryan Mukherjee. Functional map of the world. In CVPR, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.509, + 0.469, + 0.577 + ], + "angle": 0, + "content": "[10] Yezhen Cong, Samar Khanna, Chenlin Meng, Patrick Liu, Erik Rozi, Yutong He, Marshall Burke, David B Lobell, and Stefano Ermon. Satmae: Pre-training transformers for temporal and multi-spectral satellite imagery. In NeurIPS, 2022. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.579, + 0.469, + 0.62 + ], + "angle": 0, + "content": "[11] Rodrigo Caye Daudt, Bertr Le Saux, and Alexandre Boulch. Fully convolutional siamese networks for change detection. In ICIP, 2018. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.621, + 0.469, + 0.69 + ], + "angle": 0, + "content": "[12] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. CoRR, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.691, + 0.469, + 0.773 + ], + "angle": 0, + "content": "[13] Matthias Drusch, Umberto Del Bello, Sébastien Carlier, Olivier Colin, Veronica Fernandez, Ferran Gascon, Bianca Hoersch, Claudia Isola, Paolo Laberinti, Philippe Martimort, et al. Sentinel-2: ESA's optical high-resolution mission for GMES operational services. Remote Sensing of Environment, 2012. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.774, + 0.469, + 0.831 + ], + "angle": 0, + "content": "[14] Jan Gąsienica-Józkowy, Mateusz Knapik, and Bogusław Cyganek. An ensemble deep learning method with optimized weights for drone-based water rescue and surveillance. *Integrated Computer-Aided Engineering*, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.831, + 0.469, + 0.872 + ], + "angle": 0, + "content": "[15] Spyros Gidaris, Praveer Singh, and Nikos Komodakis. Unsupervised representation learning by predicting image rotations. In ICLR, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.873, + 0.469, + 0.903 + ], + "angle": 0, + "content": "[16] Jean-Bastien Grill, Florian Strub, Florent Altché, Corentin Tallec, Pierre Richemond, Elena Buchatskaya, Carl Doersch," + }, + { + "type": "list", + "bbox": [ + 0.079, + 0.115, + 0.47, + 0.903 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.135 + ], + "angle": 0, + "content": "Bernardo Avila Pires, Zhaohan Guo, Mohammad Gheshlaghi Azar, et al. Bootstrap your own latent-a new approach to self-supervised learning. NeurIPS, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.136, + 0.892, + 0.177 + ], + "angle": 0, + "content": "[17] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollár, and Ross Girshick. Masked autoencoders are scalable vision learners. In CVPR, 2022. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.179, + 0.892, + 0.22 + ], + "angle": 0, + "content": "[18] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In CVPR, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.221, + 0.892, + 0.275 + ], + "angle": 0, + "content": "[19] Patrick Helber, Benjamin Bischke, Andreas Dengel, and Damian Borth. Eurosat: A novel dataset and deep learning benchmark for land use and land cover classification. JS-TAEORS, 2019. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.277, + 0.892, + 0.332 + ], + "angle": 0, + "content": "[20] Neal Jean, Sherrie Wang, Anshul Samar, George Azzari, David Lobell, and Stefano Ermon. Tile2vec: Unsupervised representation learning for spatially distributed data. In AAAI, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.334, + 0.892, + 0.375 + ], + "angle": 0, + "content": "[21] Marrit Leenstra, Diego Marcos, Francesca Bovolo, and Devis Tuia. Self-supervised pre-training enhances change detection in sentinel-2 imagery. In ICPR, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.376, + 0.892, + 0.433 + ], + "angle": 0, + "content": "[22] Yahui Liu, Jian Yao, Xiaohu Lu, Menghan Xia, Xingbo Wang, and Yuan Liu. RoadNet: Learning to comprehensively analyze road networks in complex urban scenes from high-resolution remotely sensed images. TGRS, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.434, + 0.892, + 0.474 + ], + "angle": 0, + "content": "[23] William A Malila. Change vector analysis: an approach for detecting forest changes with landsat. In LARS symposia, 1980. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.476, + 0.892, + 0.518 + ], + "angle": 0, + "content": "[24] Utkarsh Mall, Bharath Hariharan, and Kavita Bala. Change event dataset for discovery from spatio-temporal remote sensing imagery. In NeurIPS, 2022. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.519, + 0.892, + 0.573 + ], + "angle": 0, + "content": "[25] Oscar Manas, Alexandre Lacoste, Xavier Giroi-nieto, David Vazquez, and Pau Rodriguez. Seasonal contrast: Unsupervised pre-training from uncurated remote sensing data. In ICCV, 2021. 1, 2, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.575, + 0.892, + 0.615 + ], + "angle": 0, + "content": "[26] Ishan Misra and Laurens van der Maaten. Self-supervised learning of pretext-invariant representations. In CVPR, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.617, + 0.892, + 0.672 + ], + "angle": 0, + "content": "[27] Sorour Mohajerani and Parvaneh Saeedi. Cloud and cloud shadow segmentation for remote sensing imagery via filtered jaccard loss function and parametric augmentation. J-STARS, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.674, + 0.892, + 0.715 + ], + "angle": 0, + "content": "[28] Mehdi Noroozi and Paolo Favaro. Unsupervised learning of visual representations by solving jigsaw puzzles. In ECCV, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.717, + 0.892, + 0.758 + ], + "angle": 0, + "content": "[29] Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding. CoRR, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.76, + 0.892, + 0.802 + ], + "angle": 0, + "content": "[30] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In ICMICCAI, pages 234-241. Springer, 2015. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.803, + 0.892, + 0.871 + ], + "angle": 0, + "content": "[31] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael S. Bernstein, Alexander C. Berg, and Li Fei-Fei. Imagenet large scale visual recognition challenge. IJCV, 2015. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.872, + 0.892, + 0.903 + ], + "angle": 0, + "content": "[32] Linus Scheibenreif, Joëlle Hanna, Michael Mommert, and Damian Borth. Self-supervised vision transformers for land-" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.903 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5269" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.092, + 0.468, + 0.119 + ], + "angle": 0, + "content": "cover segmentation and classification. In CVPR EarthVision, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.122, + 0.469, + 0.176 + ], + "angle": 0, + "content": "[33] Carole H Sudre, Wenqi Li, Tom Vercauteren, Sebastien Ourselin, and M Jorge Cardoso. Generalised dice overlap as a deep learning loss function for highly unbalanced segmentations. In DLMIA, 2017. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.179, + 0.469, + 0.233 + ], + "angle": 0, + "content": "[34] Gencer Sumbul, Marcela Charfuelan, Begüm Demir, and Volker Markl. Bigearthnet: A large-scale benchmark archive for remote sensing image understanding. In IGARSS, 2019. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.235, + 0.469, + 0.302 + ], + "angle": 0, + "content": "[35] Xu Tang, Huayu Zhang, Lichao Mou, Fang Liu, Xiangrong Zhang, Xiao Xiang Zhu, and Licheng Jiao. An unsupervised remote sensing change detection method based on multiscale graph convolutional network and metric learning. TGRSS, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.306, + 0.469, + 0.374 + ], + "angle": 0, + "content": "[36] Aysim Token, Lukas Kondmann, Mark Weber, Marvin Eisenberger, Andrés Camero, Jingliang Hu, Ariadna Pregel Hoderlein, Caglar Senaras, Timothy Davis, Daniel Cremers, et al. Dynamicearthnet: Daily multi-spectral satellite dataset for semantic change segmentation. In CVPR, 2022. 1, 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.376, + 0.469, + 0.417 + ], + "angle": 0, + "content": "[37] Gabriel Tseng, Ivan Zvonkov, Catherine Lilian Nakalembe, and Hannah Kerner. Cropharvest: A global dataset for croptype classification. In NeurIPS, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.42, + 0.468, + 0.459 + ], + "angle": 0, + "content": "[38] Adam Van Etten, Dave Lindenbaum, and Todd M Bacastow. Spacenet: A remote sensing dataset and challenge series. CoRR, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.462, + 0.469, + 0.515 + ], + "angle": 0, + "content": "[39] Adam Van Etten, Jacob Shermeyer, Daniel Hogan, Nicholas Weir, and Ryan Lewis. Road network and travel time extraction from multiple look angles with spacenet data. In IGARSS, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.519, + 0.469, + 0.572 + ], + "angle": 0, + "content": "[40] Junjue Wang, Zhuo Zheng, Ailong Ma, Xiaoyan Lu, and Yanfei Zhong. Loveda: A remote sensing land-cover dataset for domain adaptive semantic segmentation. arXiv preprint arXiv:2110.08733, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.576, + 0.469, + 0.643 + ], + "angle": 0, + "content": "[41] Ben G Weinstein, Sergio Marconi, Stephanie A Bohlman, Alina Zare, Aditya Singh, Sarah J Graves, and Ethan P White. A remote sensing derived data set of 100 million individual tree crowns for the national ecological observatory network. *Elife*, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.646, + 0.469, + 0.699 + ], + "angle": 0, + "content": "[42] Chen Wu, Hongruixuan Chen, Bo Do, and Liangpei Zhang. Unsupervised change detection in multi-temporal vhr images based on deep kernel pca convolutional mapping network. IEEE Transactions on Cybernetics, 2019. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.703, + 0.469, + 0.742 + ], + "angle": 0, + "content": "[43] Zhirong Wu, Yuanjun Xiong, Stella X. Yu, and Dahua Lin. Unsupervised feature learning via non-parametric instance discrimination. In CVPR, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.745, + 0.469, + 0.772 + ], + "angle": 0, + "content": "[44] Richard Zhang, Phillip Isola, and Alexei A Efros. Colorful image colorization. In ECCV, 2016. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.469, + 0.772 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "5270" + } + ] +] \ No newline at end of file diff --git a/2023/Change-Aware Sampling and Contrastive Learning for Satellite Images/ef28aefe-c0fe-45ce-8981-6fc9107e6b77_origin.pdf b/2023/Change-Aware Sampling and Contrastive Learning for Satellite Images/ef28aefe-c0fe-45ce-8981-6fc9107e6b77_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..34a7801ad843248f1ed5d045a75e7d24a5c782a1 --- /dev/null +++ b/2023/Change-Aware Sampling and Contrastive Learning for Satellite Images/ef28aefe-c0fe-45ce-8981-6fc9107e6b77_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5f7b38d89560cb9a63cb4c50455c6c45e97b01ca3cb1707f47b820e7fb2b1dc2 +size 1664051 diff --git a/2023/Change-Aware Sampling and Contrastive Learning for Satellite Images/full.md b/2023/Change-Aware Sampling and Contrastive Learning for Satellite Images/full.md new file mode 100644 index 0000000000000000000000000000000000000000..09cea3ce256c1bbff30950e1c8121e43009033f2 --- /dev/null +++ b/2023/Change-Aware Sampling and Contrastive Learning for Satellite Images/full.md @@ -0,0 +1,326 @@ +# Change-Aware Sampling and Contrastive Learning for Satellite Images + +Utkarsh Mall + +Bharath Hariharan + +Cornell University + +{utkarshm, bharathh, kb}@cs.cornell.edu + +Kavita Bala + +# Abstract + +Automatic remote sensing tools can help inform many large-scale challenges such as disaster management, climate change, etc. While a vast amount of spatio-temporal satellite image data is readily available, most of it remains unlabelled. Without labels, this data is not very useful for supervised learning algorithms. Self-supervised learning instead provides a way to learn effective representations for various downstream tasks without labels. In this work, we leverage characteristics unique to satellite images to learn better self-supervised features. Specifically, we use the temporal signal to contrast images with long-term and short-term differences, and we leverage the fact that satellite images do not change frequently. Using these characteristics, we formulate a new loss contrastive loss called Change-Aware Contrastive (CACo) Loss. Further, we also present a novel method of sampling different geographical regions. We show that leveraging these properties leads to better performance on diverse downstream tasks. For example, we see a $6.5\%$ relative improvement for semantic segmentation and an $8.5\%$ relative improvement for change detection over the best-performing baseline with our method. + +# 1. Introduction + +Our planet is surrounded by a large number of satellites constantly collecting images of the world. This massive trove of visual information can help monitor phenomena at the world-scale, and inform solutions to global problems such as climate change or loss of biodiversity. Automatic vision tools can help by, for example, monitoring land-use change over time [36] or the evolution of urban areas [6]. + +However, training all these models requires labeled data. Unfortunately, labeling the massive trove of satellite images is expensive, more so than internet images because of the expertise necessary. This issue is exacerbated by the different label requirements of many monitoring applications. + +One way to alleviate this problem of limited labeled data is to use self-supervised learning techniques to learn a good feature representation from unlabeled satellite im + +![](images/ab9b7b2dfaf29125f6acf27058333da77fb75d35c2c81fb3b1fc3b8c7cf51884.jpg) +Figure 1. Images of the same location at three different times. Changes from 2016 to 2020 are due to major urban development, while those from July to September are seasonal variations. Our approach, CACo, learns features that are sensitive to the former but invariant to the latter. + +agery. This representation can then be further finetuned with much fewer labels for specific applications. Modern self-supervised learning approaches are based on contrastive learning. These techniques train a feature space so that each image in the dataset is embedded close to augmented versions of itself (e.g., with jittered colors) but far from other images of the dataset. A possible approach is to directly apply these techniques on satellite image datasets. However, the spatio-temporal structure of satellite imagery is much richer than the unstructured collections of internet images typically used in standard self-supervised learning. In this work, we ask, how can we best leverage the structure of satellite images for better self-supervised learning? + +The first important aspect of the spatio-temporal structure of satellite images is the availability of multiple temporally-spaced images for the same location. Past work has used this structure to sample images spread over a few weeks or months from each location to encourage invariance to seasonal variations [25]. However, we can access images not just over a few months, but over many years. Over such long time spans, we often see significant, per + +manent change, such as the construction of houses, the drying of lakes, or the logging of forests (Fig. 1). While we want feature representations to be invariant to temporary, seasonal change, we want the representation to be sensitive to permanent, long-term change. To capture this intuition, we sample multiple images from the same location covering both short and long time spans and encourage invariance to the former but sensitivity to the latter. + +A second and more important aspect of satellite images is that permanent change is spatially rare. Change is concentrated near urban areas, but many parts of the planet see little change. When there is no change, we want our feature representation to be the same even over long time spans. We capture this insight with a novel strategy to robustly estimate whether or not there is a long-term change, even in the middle of training, by comparing long-term feature differences to short-term variations. We then design a loss function that is conditional on this change estimate: it encourages invariances to long-term differences depending on whether a change occurs or not. We call this novel loss function Change-aware Contrastive Loss (or CaCo). + +The above loss function uses the temporal structure of satellite images. We can also utilize geographical structure by carefully sampling the most informative locations on the planet. We provide an improvement over a previously proposed geographical sampling [25]. We show that sampling closer to cities, and ignoring samples completely in the ocean can result in a dataset much more useful for learning a general representation for various downstream tasks. We evaluate our new representation on a diverse set of downstream tasks such as landcover classification, semantic segmentation, and change detection. Our method achieves significant relative improvements (ranging from $6.5\%$ to $8.5\%$ ) over the state-of-the-art for a variety of tasks such as segmentation and change detection. + +To summarize, we make the following contributions: + +- We propose a new self-supervised learning loss that uses long-term temporal information in satellite imagery to encourage invariance to seasonal variations but sensitivity to permanent, long-term changes. +- We introduce a novel approach to robustly estimate whether a location has undergone significant change by comparing long-term changes to seasonal variations. Our new change-aware loss function (CACo) uses this to decide when to encourage invariance. +- We use an improved geographical sampling that provides more diverse data for representation learning. + +# 2. Related Work + +Self-supervised Learning. Self-supervised learning methods can be used to learn a good general representation from a dataset without any labels. These methods obtain + +supervisory signals or inductive bias from the unlabelled data itself. Earlier self-supervised works used "pretext tasks" such as rotation prediction [15], solving jigsaw puzzles [28], colorization [44], or missing data completion [17]. More recently contrastive learning methods have been shown to learn a better representation than pretext tasks. Methods for contrastive learning such as NPID [43], PIRL [26], MoCo [8, 18], SimCLR [7], and BYOL [16], use instance discrimination to learn a feature representation. While these methods can be applied even to satellite images, in this work we leverage the structure of satellite images that can be used for self-supervised learning. + +Self-supervised Learning in Remote Sensing. Increasingly, self-supervised learning is being applied to the area of remote sensing. Prior works have used signals from unlabeled data such as location [2, 20], seasonal variations [25], or texture [1], as a signal for self-supervision. Recent methods use transformers [12] and diffusion models [4] for self-supervised representation [10, 32]. + +Remote Sensing Applications and Data Several remote sensing applications require vision tools for automatic recognition at scale. Applications such as landcover classification [19, 34] or segmentation [36, 40] benefit from the advances in vision algorithms for scene classification and semantic segmentation. Several object detection applications have also been developed using object detection algorithms such as for detecting constructions [9], buildings [38], trees [41], or floating objects [14]. Segmentation methods also help in extracting structural information like road networks [3, 22, 39], crop types [37], or clouds [27] from remote sensing images. Many applications such as change detection [11] require temporal information for regions as well. Temporal information and changes are also used to find long-term semantic events [24]. In this work, we aim to learn a representation without supervision, that leads to better performance on all these downstream tasks. + +The image data for these algorithms and datasets are obtained from a few different sources. Sentinel-2 [13] provides multispectral images (13 bands) at $10\mathrm{m}$ resolution with a temporal revisit of 5 days. Several other satellites provide higher resolution information such as WorldView-3 and PlanetScope. We use Sentinel-2 satellite imagery as it provides frequent temporal information that we use in our self-supervised formulation. Furthermore, many of the downstream tasks [11, 19, 24, 36] use Sentinel-2 imagery. + +# 3. Method + +# 3.1. Overview + +We propose a self-supervised method that can leverage properties that are unique to satellite images. One such + +property is the availability of long-term temporal information for any location on earth. For example, two images of the same location captured months apart will likely be the same except for seasonal variations; variations our features should be invariant to. In contrast, two images captured decades apart might show big, structural changes such as new buildings or sparser forests. Of course, this may not be true for all locations. Indeed, such long-term change is rare, and is concentrated in regions of human activity. But these long-term changes are important for many downstream applications such as modeling urban growth, or monitoring climate change. Our key insight is therefore to find these rare but meaningful long-term changes, and ensure that they are reflected as changes in the feature space as well. + +Past work on using temporal data for representation learning on satellite imagery only uses short-term seasonal variations but not long-term change [25]. We review this work, which is called Seasonal Contrast (or SeCo) [25], in Sec. 3.2. We then explain how we can use long-term temporal information in Sec. 3.3, and address the challenge of spatial rarity in Sec. 3.4. Finally, we look at how we can efficiently sample more informative locations when collecting unlabelled training data in Sec. 3.5. + +# 3.2. Background: Seasonal Contrast (SeCo) + +Contrastive learning techniques train feature representations from unlabeled images by discriminating between individual instances. In particular, they pull augmentations of the same image closer together in feature space and push apart representations of two different images. SeCo is a contrastive learning-based representation learning framework for remote sensing images. SeCo uses MoCo v2 [8], to perform contrastive learning. In MoCo v2, in each iteration, one generates two views for an image using random augmentations. One is the "query" image $I_{q}$ and the other is the "positive key" $I_{k+}$ that is pulled closer to the query in the representation space. It also has a set of other images from the dataset, $\mathcal{I}_{k-}$ , that are treated as negative keys and are pushed apart from query image $I_{q}$ in the latent space. The precise loss function in MoCo v2 called InfoNCE [29], for a representation function $f$ , can be written as: + +$$ +\mathcal {L} = - \log \frac {\exp \left(f \left(I _ {q}\right) \cdot f \left(I _ {k ^ {+}}\right) / \tau\right)}{\sum_ {I _ {k} \in \mathcal {I} _ {k ^ {-}} \cup \{I _ {k +} \}} \exp \left(f \left(I _ {q}\right) \cdot f \left(I _ {k}\right) / \tau\right)} \tag {1} +$$ + +The key idea in SeCo is to also use short-term temporal differences (over a few months) at a location as augmentations. Since these images are a few months apart, these images represent different seasons (hence the name seasonal contrast). The negative keys are then images from other locations. More specifically SeCo trains a representation with 3 subspaces, i) a subspace invariant to both season and artificial augmentations, ii) a subspace invariant to season only, + +and iii) a subspace invariant to augmentation only. The additional invariance to seasonal differences results in a better representation than simply using MoCo v2 on this data. + +# 3.3. Using Long-term Temporal Information + +SeCo introduces invariance to seasonal variations. However, invariance is not enough. We also want our feature representation to be informative. For this, simply pushing apart images from different locations (which already look quite different) may not be enough. + +As discussed above, we argue that to produce a richer feature representation, we should use information from longer temporal spans. Over the span of years, we might observe big, structural changes such as new construction or urban growth. We want our feature representation to be sensitive to these major changes. Thus, while short-term temporal differences are typically seasonal and should be treated as augmentations as in SeCo, long-term temporal differences might correspond to permanent change, and should be reflected as a significant change in the underlying feature representation. This suggests that for a query image from a particular location, an image captured at the same location several years later should be treated as a negative in the contrastive learning objective. We formalize this below. + +Data: Let $\{l_1, l_2 \dots l_n\}$ be $n$ different locations in our dataset. For each location, we have 2 sets of images captured around time points $t_1$ and $t_2$ that have a gap of 4 years between them. We represent these sets by $\mathcal{I}_{l_i}^{t_1}$ and $\mathcal{I}_{l_i}^{t_2}$ . Each set $\mathcal{I}_{l_i}^{t_j}$ , contains images sampled over a few months around $t_j$ , and thus these images have only short-term or seasonal differences between them. See Fig. 2 (a) for a schematic of this data sampling strategy. + +Loss: Let $I_{l_i}^{t_j + \Delta_1}, I_{l_i}^{t_j + \Delta_2} \sim \mathcal{I}_{l_i}^{t_j}$ be two images sampled randomly from $\mathcal{I}_{l_i}^{t_j}$ ( $j$ is 1 or 2). $I_{l_i}^{t_j + \Delta_1}$ and $I_{l_i}^{t_j + \Delta_2}$ are a few months apart. So like SeCo, we still want their representations to be pulled closer. However, images with larger time differences, for example, $I_{l_i}^{t_1 + \Delta_1}$ and $I_{l_i}^{t_2 + \Delta_1}$ should potentially be pushed apart. Finally, images from two different locations such as $I_{l_i}^{t_j + \Delta_1}$ and $I_{l_j}^{t_j + \Delta_1}$ , should always be pushed apart irrespective of time. This intuition is reflected in Fig. 2 (c). + +Following SeCo, we implement this loss function using MoCo v2 by adapting the positive key image $I_{k+}$ and negative key set $\mathcal{I}_{k-}$ in Eq. (1). As discussed above, SeCo has 3 subspaces. We add our modifications to all 3 subspaces. However, for ease of explanation, we only describe the change to the subspace that is invariant to seasons. For this subspace, for a query image, $I_q = I_{l_i}^{t_1 + \Delta_1}$ , the positive key is the same as in SeCo: the seasonal pair $(I_{k^+} = I_{l_i}^{t_1 + \Delta_2})$ . But we modify the negative key set to not only include images from other locations but also images from the same + +![](images/e3ad505694a5a6acdc45320509ab95a7eaec1cdf354572ba08e1a76180b7b845.jpg) +Figure 2. Overview of our self-supervised framework. (a) Shows the notation for images and the temporal sampling strategy we use to create our dataset for training. (b) Shows the contrastive learning performed by SeCo in the temporally invariant subspace. The red arrows show examples that need to be pushed apart from the top left example. The green arrows indicate examples to be pulled closer. Different lightness (value) represents different short-term temporal pairs and different colors (hue) indicate different locations (see Sec. 3.2). (c) Shows our method when using long-term temporal pairs. Different shapes indicate a difference of 4 years between examples (see Sec. 3.3). (d) Shows the contrastive learning performed by our CACo loss term. The dashed red arrows indicate examples, that are pushed apart iff our method estimates a change has occurred between the long-term pairs (see Sec. 3.4). + +![](images/5083f2133d87bab85501d07caa55bca49e1d4029e352144e91d68b6572c31317.jpg) + +![](images/26811b45b3a8928490894dcb740b3cc094ddec8118ec3dc40bef9afc50ce1c1c.jpg) + +location with a large time difference as follows: + +$$ +\begin{array}{l} \mathcal {I} _ {k ^ {-}} = \left\{I _ {l _ {i}} ^ {t _ {2} + \Delta_ {k}}: k \in \{1, 2 \} \right\} \cup \\ \left\{I _ {l _ {h}} ^ {t _ {j} + \Delta_ {k}}: h \neq i, j \in \{1, 2 \}, k \in \{1, 2 \} \right\} \quad (2) \\ \end{array} +$$ + +In each training iteration, we sample 4 images (long and short-time pairs) for each location instead of 2 like SeCo, so each batch contains half as many locations. + +# 3.4. Change-aware Contrastive Learning + +The training method in the previous section assumes that regions always change after a long time (a few years). However, this assumption is not always true: in remote areas away from human activity, there may be no change even over several years. For such locations, pushing apart features of temporally distant image pairs may be counterproductive and destroy needed invariances. + +If we know whether a location has changed or not, we can make our loss function conditional on this information. If it has changed between $I_{l_i}^{t_1 + \Delta_1}$ and $I_{l_i}^{t_2 + \Delta_1}$ , we push them apart as described in Sec. 3.3. If not, we refrain from adding the temporally distant image to the negative set. This conditional loss function is illustrated in Fig. 2 (d). + +However, the challenge is knowing which locations have changed. Change detection can be performed by looking at feature differences between temporally separated image + +pairs [21, 23, 35, 42], but this requires a feature representation we do not have. We solve this chicken-and-egg problem by bootstrapping our feature representation. We start from a randomly initialized model, using which we extract features for different locations and use feature differences as an estimate for change. Using these change values, we use our conditional loss to train the model. We then keep alternating between training the model using changes and estimating the changes using the partially trained model. In the initial stages, the change estimates could be very poor as the model has random weights. However, as training progresses we expect better features, and, in turn, better estimates of changes. One might worry about degenerate solutions, but the other contrastive loss terms based on location (that do not depend on this inferred change) ensure that the feature representation avoids any pathological cases. + +Obtaining changes using features. The absolute distance between features scales differently for different types of locations, and does not directly reflect the magnitude of actual change. For example, urban locations with less vegetation look the same over the seasons, but temperate forests will change a lot with the seasons. So instead of using the absolute distance between two long-term images, we instead use the ratio of distances between long-term images and short-term images (images with a seasonal difference). + +This ratio normalizes the scaling differences between images from different locations. If $f_{k}$ is the partially trained feature extractor at epoch $k$ , we use the following ratio as an indicator of change for location $l_{i}$ + +$$ +R _ {l _ {i}} ^ {k} = \frac {\left| \left| f _ {k} \left(I _ {l _ {i}} ^ {t _ {1} + \Delta_ {1}}\right) - f _ {k} \left(I _ {l _ {i}} ^ {t _ {2} + \Delta_ {1}}\right) \right| \right| ^ {2}}{\left| \left| f _ {k} \left(I _ {l _ {i}} ^ {t _ {1} + \Delta_ {1}}\right) - f _ {k} \left(I _ {l _ {i}} ^ {t _ {1} + \Delta_ {2}}\right) \right| \right| ^ {2}} \tag {3} +$$ + +Since, at each epoch, the seasonal images are selected randomly for each location, we use an exponential moving average over ratios from previous epochs to obtain a more stable estimate for change as follows + +$$ +\mathbf {r} _ {l _ {i}} ^ {k} = (1 - \beta) R _ {l _ {i}} ^ {k} + \beta \mathbf {r} _ {l _ {i}} ^ {k - 1} +$$ + +We use a Gaussian Mixture Model (GMM) on $\mathbf{r}_{l_i}^k$ values to find our two clusters for change and no change. Locations with smaller $\mathbf{r}_{l_i}^k$ are treated as having no long-term change, and those with larger $\mathbf{r}_{l_i}^k$ values are deemed to have undergone significant change (see Fig. 3). + +For MoCo v2, we use features from the momentum encoder for images to calculate the ratio estimate. To maintain training speed, we compute and store the ratios during the training iterations themselves, even as the feature extractor is being updated. In preliminary experiments, we found that the exponential moving average was sufficient to regularize against the noise induced by the changing feature extractor. + +Our change-aware formulation uses change information in long-term pairs. On the other hand, there could be short-term pairs with sudden changes such as due to disasters like landslides or earthquakes. However, we notice that in practice such changes are extremely rare. Out of the top 200 examples with the highest short-term change score only 3 pairs showed a sudden change. Because short-term changes are so rare, it is not useful to consider them in our pipeline. + +# 3.5. Improving Geographical Sampling + +Another property specific to satellite imagery is that we can control the geographic distribution of sampled images. Leveraging this freedom, SeCo collects data from a $100\mathrm{km}$ radius around the 10k most populated cities of the world. However, we argue that 100 kms is too large a radius. + +Sampling far from cities results in uninformative images (such as images entirely over oceans, (see supplementary Fig. 3)). In fact, $22\%$ of the dataset has such uninformative images and this hurts the quality of the representation. + +The median area of the 10k cities sampled is approximately $24\mathrm{km}^2$ , which represents a radius of $\sim 5\mathrm{km}$ . Based on this observation, we use 2d-gaussian sampling with $\sigma = 5\mathrm{km}$ . This means that $95\%$ of the data comes from a radius of $2\sigma = 10\mathrm{km}$ around the city center. Additionally, if the sampled location is completely in the ocean, we reject that sample and sample again (We use a landmass map of the earth to check this). Our improved sampling + +![](images/14d85addedef9701805eb72ab54aaebe51556483a43dc3aab6932df45c543dd8.jpg) +Figure 3. The histogram of change estimates separated into two groups (blue and green) by the GMM. Pairs of examples with a higher estimate can be seen to have major construction and land-cover changes (2 pairs on the right). Whereas pairs with a lower estimate only have seasonal changes. + +methods result in a dataset with a more diverse set of locations that are more informative. As we show in experiments, this additionally improves the quality of the representation. + +# 4. Results + +# 4.1. Implementation details and Baselines + +We evaluate our model by measuring its representation ability on various downstream tasks. We evaluate the following baselines for comparison. + +Random init.: is randomly initialized and is either fin-tuned or linear probed for downstream tasks. + +ImageNet: is pre-trained on supervised ImageNet [31]. + +MoCo v2: uses self-supervised learning without the temporal or change information. + +SeCo: is Seasonal Contrast [25] and uses only short-term changes as augmentations. + +GSSL: uses temporal and geographical information for contrastive learning [2]. Since the data this model was trained on is of different resolution retrain it on our data. + +SatMAE [10]: is a Masked autoencoder [17] that uses ViT-Large as backbone. The comparison is not apples-to-apples as the model has larger capacity. The model is trained on 713k images therefore it is comparable to our 1m images trained model. + +DDPM-CD [4]: uses a denoising diffusion model as backbone. It is also trained on a million scale dataset. + +We show results with both a ResNet-18 and a ResNet-50 backbone. Similar to SeCo our models are trained with a batch size of 256 and with 16,384 negative embeddings. We also use the same optimizer, learning rate, schedule, and temperature scaling as in the SeCo implementation. + +We experiment with two datasets for self-supervised pretraining: one with 100k images and the other with 1 million + +
DataPre-trainingResNet-18ResNet-50
-Random init.64.2155.32
ImageNet.86.1689.08
100kMoCo v287.2289.75
GSSL87.7490.19
SeCo90.0593.12
CACo (ours)93.0894.48
1mSeCo93.9995.63
SatMAE (ViT-L)-93.03
DDPM-CD-87.67
CACo (ours)94.7295.90
+ +images. These datasets contain RGB images from Sentinel2 [13]. The long-time difference between the two sets $\mathcal{I}_{l_i}^{t_1}$ and $\mathcal{I}_{l_i}^{t_2}$ is 4 years. The maximum time difference within a set $\mathcal{I}_{l_i}^{t_j}$ is 1 year. We train our model and baselines for 1000 and 200 epochs on the 100k and 1m dataset respectively. + +# 4.2. Landcover classification + +Nine (out of seventeen) of the UN's Sustainable Development Goals require global monitoring of land cover [5]. Thus, improving the performance of models for landcover classification would result in better monitoring of these goals. We evaluate our method on two datasets for landcover classification: EuroSat [19] and BigEarthNet [34]. + +EuroSat has 10-classes with a total of $27\mathrm{k}$ images $64\times 64$ images from Sentinel-2. We use the same train/val split proposed by the dataset. We add a linear layer to the frozen pretrained backbone, to perform the linear evaluation. More details about the training are in the supplementary. + +Tab. 1 shows the top-1 classification accuracy of various pre-trained backbones on the EuroSat dataset. We first note that all satellite image-specific methods outperform generic pretraining on ImageNet, suggesting the importance of training a representation specifically for satellite imagery. They also outperform the MoCo v2 baseline trained on satellite images, indicating the importance of well-designed loss functions that use the structure of satellite imagery. + +Compared to SeCo, our approach (CACo) results in 3 points of improvement with the linear classifier (with ResNet-18 and 100k data). This validates our insight that invariance to seasonal changes is not enough; the features must also be trained to be sensitive to long-term change. In fact, CACo trained on 100k images produces a representation competitive with SeCo trained on 1m images. Thus our method can save an order of magnitude of images and training time over the SeCo baseline while producing rep + +Table 1. Performance of our representation on the EuroSat classification task with linear probing, in top-1 Accuracy. Our method provides a more accurate classification, with different backbones. + +
DataPre-trainingResNet-18ResNet-50
10%100%10%100%
-Random init.42.8745.9544.7645.22
ImageNet.65.4366.4070.3671.37
100kMoCo v265.4367.2070.3872.88
GSSL65.7867.3670.6572.86
SeCo65.8067.4370.6973.42
CACo (ours)67.8969.4371.5573.63
1mSeCo67.6869.9572.8974.82
SatMAE--67.9369.45
DDPM-CD--65.4767.31
CACo (ours)68.6470.4173.4074.98
+ +Table 2. Performance of our method on BigEarthNet landcover classification, in mean Average Precision (mAP). The two columns use different percentages of data (10% and 100%) of BigEarthNet data for training the linear layer. + +resentations of similar quality. + +We also evaluate our method on the BigEarthNet dataset, which is a multi-label classification dataset with 17 classes. The dataset is significantly larger than EuroSat, containing 590k sentinel-2 image patches. We use the RGB bands to perform classification on this dataset. Since this is a multi-label classification problem, we use mean Average Precision (mAP) to evaluate the performance. + +Tab. 2 compares the performance of our model to the baselines on BigEarthNet1. Even on BigEarthNet, our method performs better than the baselines. + +# 4.3. Change detection + +We also evaluate our model on the OSCD change detection dataset [6]. It contains 24 pairs of images from the Sentinel-2 satellites between 2015 and 2018. The dataset is split into 14 training images and 10 testing images. + +Our model architecture follows past work [11, 25]. The input to the model is an image pair, each of which is fed into the frozen, pre-trained feature extractor (we use the ResNet-18 models). We extract feature maps after each downsampling layer for each image. We then take the absolute difference between the corresponding feature maps from the two images. These differences are passed as input to a U-Net Decoder [30]. The U-Net decoder is then trained with change supervision. More details about the training and architecture can be found in the supplementary. + +Results are shown in Tab. 3. We report the F1-score obtained by using a threshold of 0.5. Our method results in better features for change detection resulting in about 4 + +
BackbonePre-trainingPrec.Rec.F1-score
ResNet-18Random init.69.7318.2428.91
ImageNet.70.2223.5835.30
MoCo v262.2127.5738.21
GSSL62.2934.0844.06
SeCo64.1536.8946.84
CACo (ours)60.6842.9450.29
ResNet-50Random init.65.4622.4333.41
ImageNet.70.0530.9642.94
GSSL62.1636.8346.26
SeCo63.2138.2647.67
CACo (ours)62.8744.4952.11
+ +Table 3. Performance of our method on the Change Detection Task. We report the precision, recall and F1-score at a threshold of 0.5 for various ResNet backbones, trained on the 100k dataset. + +
Pre-traininglinearfinetuning
Random init.41.5338.62
ImageNet.43.7543.78
MoCo v247.9747.15
GSSL46.7748.10
SeCo46.8348.18
CACo (ours)50.2051.29
+ +points of improvement in the F1 scores for both ResNet-18 and ResNet-50. This large improvement on change detection is likely because we explicitly train our representation to be sensitive to long-term change. + +# 4.4. Semantic segmentation + +We also evaluate our method on a satellite images segmentation task. DynamicEarthNet [36] is a dataset used to evaluate landcover segmentation over 7 classes. The dataset contains images from Planet satellites for 65 locations. Each location has $241024\times 1024$ images with labels. The locations are split into 55 for training and 10 for testing. + +Similar to change detection model, we use a U-net with ResNet-18 as the encoder. But unlike them, we input a single image and use feature maps instead of feature differences in the decoder. Due to the dataset imbalance, we use Dice Loss [33] for training (see supplementary for details). + +Tab. 4 shows the mIoU (intersection over union averaged over classes) for our method on DynamicEarthNet validation images. Our method is better than SeCo by more than 3 points, and ImageNet pre-training by more than 6 points. + +Table 4. Performance of our method on the DynamicEarthNet segmentation task. We report the mIoU score for the ResNet-18 backbone, trained on the 100k dataset. + +
Pre-trainingCaiRoadCalFire
AP@50AP@400AP@10AP@40
ImageNet41.7932.2652.6144.99
MoCo v241.7833.0254.8348.20
SeCo39.6334.7262.8751.87
CACo (ours)44.3835.9965.7153.44
+ +Table 5. Performance of our method on change event retrieval for the CaiRoad and CalFire benchmarks. We show average precision@K for ResNet-18 backbones trained on the 100k dataset. + +![](images/b6f88e96e848e5a967ce929aac2eed91c96404b013669a4247bf41fdb9f72221.jpg) +Figure 4. Precision@K of our backbone on CaiRoad (Left) and CalFire (Right) retrieval benchmark. Our method has better Precision@K than baselines for most K values on both benchmarks. + +![](images/d39a771990eeefec49e023640f899f36b0b5d13ac493286762531d3a4e712431.jpg) + +# 4.5. Evaluation on Change Events + +We evaluate our method on a recently released change events benchmark [24]. Change events are spatio-temporal semantic structures that are caused by a real-world event. The goal is to classify these change events. The benchmark contains two datasets: one from Cairo where we find road construction events (called CaiRoad), and the other from California where we find forest fire events (CalFire). + +We use the model proposed by Mall et al. [24], taking in a sequence of images and binary change masks for consecutive pairs. The model encodes image pairs using ResNet-18, average pooling the feature map with the downscaled change mask as weights. The final feature vector is obtained by temporal averaging. The representation is learned using SimCLR on change events (see supplementary). + +We follow the same procedure, but the ResNet-18 backbone is initialized with the trained weights, and then finetuned per [24]. We evaluate by measuring average precision@K (AP@K) for different backbone architectures. + +Tab. 5 shows the AP@K for different backbones, on the CaiRoad and CalFire benchmarks. Our method results in better retrieval for both types of events. Fig. 4 shows the precision@K for various methods on these benchmarks. + +# 4.6. Ablation + +We evaluate the design choices made by our method in Sec. 3. We look at improvements made by different com + +
Pre-trainingEuroSat (Acc.)DynamicEarth-Net (mIoU)
SeCo90.0546.83
+ Improved Sampling91.4247.24
+ Long-term temporal Contrast92.5649.68
+ Change-awareness (Ours)93.0850.20
+ +Table 6. Performance with different sub-components of CACo on EuroSat classification, with ResNet-18 pre-trained on the 100k dataset. All three of our novel contributions lead to an improvement in performance on downstream tasks. + +
Pre-trainingEuroSat (Acc.)OSCD (F1)
SeCo87.9744.05
CACo90.8347.50
+ +ponents such as better data, long-term temporal information, and change estimates. We also show that our method generalizes to other self-supervised frameworks. + +Are the new components of our method essential? In Tab. 6 we evaluate the impact of each component of our approach on the final performance as measured on EuroSat and DynamicEarthNet. We find that each component adds a significant improvement. It is telling that incorporating long-term temporal contrast has a particularly large impact on DynamicEarthNet, indicating that long-term contrast is especially needed for pixel-level localization tasks. + +Can our method work with other self-supervised frameworks? We replace the MoCo v2 framework with SimCLR [7]. SimCLR requires a large batch size to find more hard negative examples; we use a batch size of 512 instead of 256 for all the methods. Tab. 7 shows the performance of SeCo and CACo using SimCLR. Our model performs better than SeCo by 3 points when using SimCLR as well. This indicates that our insights are general and can be applied in the future to new self-supervision frameworks. + +How useful is the ratio estimate? In Sec. 3.4 we proposed to estimate using $\mathbf{r}_{l_i}^k$ to estimate changes. We now look at other estimates for change in Tab. 8. Directly using the feature distance between the long-term pair, instead of the ratio leads to an almost 2-point drop in performance. Removing the exponential moving average (i.e., using $R_{l_i}^k$ instead of $\mathbf{r}_{l_i}^k$ ; see Eq. (3)) also reduces performance by about 1 point. Another alternative is to measure the ratio + +Table 7. Accuracy on EuroSat when using a ResNet-18 backbone pre-trained using SimCLR instead of MoCo v2. Our insights generalize to other self-supervised frameworks like SimCLR as well. + +
Change EstimateEuroSat (Acc.)
long-term distance91.17
Rliki91.98
Align92.92
rkiLi (ours)93.08
+ +Table 8. Performance of different change estimates with ResNet-18 on 100k dataset. $R_{l_i}^k$ (Eq. (3)) is the estimate without the moving average. Long-term distance is the numerator of the $R_{l_i}^k$ . Align uses seasonally aligned long-term pairs without smoothing. + +with long-term pairs that are seasonally aligned. Align gives similar results to $\mathbf{r}_{l_i}^k$ , but requires a more sophisticated sampling. In sum, these results suggest that how one identifies changed regions can have a big impact on performance. + +# 5. Discussion and Conclusions + +Potential Negative Societal Impact and Limitations. As in all visual recognition, there is the possibility of negative impact through violations of privacy. To mitigate this concern we intentionally use low-resolution satellite images (1 pixel $\sim 10\mathrm{m}$ ). The use of our proposed techniques for surveillance should be appropriately regulated. Our work currently is limited in its ability to generalize to multiple resolutions. Though object detection tasks use higher-resolution images for better object delineation, our representation trained on medium-resolution images cannot directly generalize to them; although we can retrain our method on high-resolution images for these tasks. + +Conclusions. We present a novel self-supervised framework to learn representations well suited for remote sensing applications. We introduce a new loss that robustly leverages the long-term temporal information readily available for satellite images. We also propose a better location sampling method to provide more informative data. Our evaluation on diverse downstream tasks shows that our approach (CACo) is very versatile and leads to better downstream accuracy than prior art (with relative improvements of $8.5\%$ on change detection and $6.5\%$ on semantic segmentation). + +Acknowledgements. This research is based upon work supported in part by the ODNI (IARPA) via 2021-20111000006, NSF 1900783, NSF 2144117, and NSF 2212084. The views and conclusions contained herein are those of the authors and should not be interpreted as necessarily representing the official policies, either expressed or implied, of ODNI, IARPA, or the US Government. The US Government is authorized to reproduce and distribute reprints for governmental purposes notwithstanding any copyright annotation therein. + +# References + +[1] Peri Akiva, Matthew Purri, and Matthew Leotta. Self-supervised material and texture representation learning for remote sensing tasks. In CVPR, 2022. 2 +[2] Kumar Ayush, Burak Uzkent, Chenlin Meng, Kumar Tanmay, Marshall Burke, David Lobell, and Stefano Ermon. Geography-aware self-supervised learning. In CVPR, 2021. 2, 5 +[3] Gaetan Bahl, Mehdi Bahri, and Florent Lafarge. Single-shot end-to-end road graph extraction. In CVPR EarthVision, 2022. 2 +[4] Wele Gedara Chaminda Bandara, Nithin Gopalakrishnan Nair, and Vishal M Patel. Ddpm-cd: Remote sensing change detection using denoising diffusion probabilistic models. CoRR, 2022. 2, 5 +[5] Sarah Carter and Martin Herold. Specifications of land cover datasets for SDG indicator monitoring. https://ggim.un.org/ Accessed: 11-8-2022. 6 +[6] R. Caye Daudt, B. Le Saux, A. Boulch, and Y. Gousseau. Urban change detection for multispectral earth observation using convolutional neural networks. In IGARSS, 2018. 1, 6 +[7] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In ICML, 2020. 2, 8 +[8] Xinlei Chen, Haoqi Fan, Ross Girshick, and Kaiming He. Improved baselines with momentum contrastive learning. CoRR, 2020. 2, 3 +[9] Gordon Christie, Neil Fendley, James Wilson, and Ryan Mukherjee. Functional map of the world. In CVPR, 2018. 2 +[10] Yezhen Cong, Samar Khanna, Chenlin Meng, Patrick Liu, Erik Rozi, Yutong He, Marshall Burke, David B Lobell, and Stefano Ermon. Satmae: Pre-training transformers for temporal and multi-spectral satellite imagery. In NeurIPS, 2022. 2, 5 +[11] Rodrigo Caye Daudt, Bertr Le Saux, and Alexandre Boulch. Fully convolutional siamese networks for change detection. In ICIP, 2018. 2, 6 +[12] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. CoRR, 2020. 2 +[13] Matthias Drusch, Umberto Del Bello, Sébastien Carlier, Olivier Colin, Veronica Fernandez, Ferran Gascon, Bianca Hoersch, Claudia Isola, Paolo Laberinti, Philippe Martimort, et al. Sentinel-2: ESA's optical high-resolution mission for GMES operational services. Remote Sensing of Environment, 2012. 2, 6 +[14] Jan Gąsienica-Józkowy, Mateusz Knapik, and Bogusław Cyganek. An ensemble deep learning method with optimized weights for drone-based water rescue and surveillance. *Integrated Computer-Aided Engineering*, 2021. 2 +[15] Spyros Gidaris, Praveer Singh, and Nikos Komodakis. Unsupervised representation learning by predicting image rotations. In ICLR, 2018. 2 +[16] Jean-Bastien Grill, Florian Strub, Florent Altché, Corentin Tallec, Pierre Richemond, Elena Buchatskaya, Carl Doersch, + +Bernardo Avila Pires, Zhaohan Guo, Mohammad Gheshlaghi Azar, et al. Bootstrap your own latent-a new approach to self-supervised learning. NeurIPS, 2020. 2 +[17] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollár, and Ross Girshick. Masked autoencoders are scalable vision learners. In CVPR, 2022. 2, 5 +[18] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In CVPR, 2020. 2 +[19] Patrick Helber, Benjamin Bischke, Andreas Dengel, and Damian Borth. Eurosat: A novel dataset and deep learning benchmark for land use and land cover classification. JS-TAEORS, 2019. 2, 6 +[20] Neal Jean, Sherrie Wang, Anshul Samar, George Azzari, David Lobell, and Stefano Ermon. Tile2vec: Unsupervised representation learning for spatially distributed data. In AAAI, 2019. 2 +[21] Marrit Leenstra, Diego Marcos, Francesca Bovolo, and Devis Tuia. Self-supervised pre-training enhances change detection in sentinel-2 imagery. In ICPR, 2021. 4 +[22] Yahui Liu, Jian Yao, Xiaohu Lu, Menghan Xia, Xingbo Wang, and Yuan Liu. RoadNet: Learning to comprehensively analyze road networks in complex urban scenes from high-resolution remotely sensed images. TGRS, 2018. 2 +[23] William A Malila. Change vector analysis: an approach for detecting forest changes with landsat. In LARS symposia, 1980. 4 +[24] Utkarsh Mall, Bharath Hariharan, and Kavita Bala. Change event dataset for discovery from spatio-temporal remote sensing imagery. In NeurIPS, 2022. 2, 7 +[25] Oscar Manas, Alexandre Lacoste, Xavier Giroi-nieto, David Vazquez, and Pau Rodriguez. Seasonal contrast: Unsupervised pre-training from uncurated remote sensing data. In ICCV, 2021. 1, 2, 3, 5, 6 +[26] Ishan Misra and Laurens van der Maaten. Self-supervised learning of pretext-invariant representations. In CVPR, 2020. 2 +[27] Sorour Mohajerani and Parvaneh Saeedi. Cloud and cloud shadow segmentation for remote sensing imagery via filtered jaccard loss function and parametric augmentation. J-STARS, 2021. 2 +[28] Mehdi Noroozi and Paolo Favaro. Unsupervised learning of visual representations by solving jigsaw puzzles. In ECCV, 2016. 2 +[29] Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding. CoRR, 2018. 3 +[30] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In ICMICCAI, pages 234-241. Springer, 2015. 6 +[31] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael S. Bernstein, Alexander C. Berg, and Li Fei-Fei. Imagenet large scale visual recognition challenge. IJCV, 2015. 5 +[32] Linus Scheibenreif, Joëlle Hanna, Michael Mommert, and Damian Borth. Self-supervised vision transformers for land- + +cover segmentation and classification. In CVPR EarthVision, 2022. 2 +[33] Carole H Sudre, Wenqi Li, Tom Vercauteren, Sebastien Ourselin, and M Jorge Cardoso. Generalised dice overlap as a deep learning loss function for highly unbalanced segmentations. In DLMIA, 2017. 7 +[34] Gencer Sumbul, Marcela Charfuelan, Begüm Demir, and Volker Markl. Bigearthnet: A large-scale benchmark archive for remote sensing image understanding. In IGARSS, 2019. 2, 6 +[35] Xu Tang, Huayu Zhang, Lichao Mou, Fang Liu, Xiangrong Zhang, Xiao Xiang Zhu, and Licheng Jiao. An unsupervised remote sensing change detection method based on multiscale graph convolutional network and metric learning. TGRSS, 2021. 4 +[36] Aysim Token, Lukas Kondmann, Mark Weber, Marvin Eisenberger, Andrés Camero, Jingliang Hu, Ariadna Pregel Hoderlein, Caglar Senaras, Timothy Davis, Daniel Cremers, et al. Dynamicearthnet: Daily multi-spectral satellite dataset for semantic change segmentation. In CVPR, 2022. 1, 2, 7 +[37] Gabriel Tseng, Ivan Zvonkov, Catherine Lilian Nakalembe, and Hannah Kerner. Cropharvest: A global dataset for croptype classification. In NeurIPS, 2021. 2 +[38] Adam Van Etten, Dave Lindenbaum, and Todd M Bacastow. Spacenet: A remote sensing dataset and challenge series. CoRR, 2018. 2 +[39] Adam Van Etten, Jacob Shermeyer, Daniel Hogan, Nicholas Weir, and Ryan Lewis. Road network and travel time extraction from multiple look angles with spacenet data. In IGARSS, 2020. 2 +[40] Junjue Wang, Zhuo Zheng, Ailong Ma, Xiaoyan Lu, and Yanfei Zhong. Loveda: A remote sensing land-cover dataset for domain adaptive semantic segmentation. arXiv preprint arXiv:2110.08733, 2021. 2 +[41] Ben G Weinstein, Sergio Marconi, Stephanie A Bohlman, Alina Zare, Aditya Singh, Sarah J Graves, and Ethan P White. A remote sensing derived data set of 100 million individual tree crowns for the national ecological observatory network. *Elife*, 2021. 2 +[42] Chen Wu, Hongruixuan Chen, Bo Do, and Liangpei Zhang. Unsupervised change detection in multi-temporal vhr images based on deep kernel pca convolutional mapping network. IEEE Transactions on Cybernetics, 2019. 4 +[43] Zhirong Wu, Yuanjun Xiong, Stella X. Yu, and Dahua Lin. Unsupervised feature learning via non-parametric instance discrimination. In CVPR, 2018. 2 +[44] Richard Zhang, Phillip Isola, and Alexei A Efros. Colorful image colorization. In ECCV, 2016. 2 \ No newline at end of file diff --git a/2023/Change-Aware Sampling and Contrastive Learning for Satellite Images/images.zip b/2023/Change-Aware Sampling and Contrastive Learning for Satellite Images/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..35c57369c5e766464dd21e2d107b4afda1cce500 --- /dev/null +++ b/2023/Change-Aware Sampling and Contrastive Learning for Satellite Images/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:589ed6fdfe4f043f6ce0c4d0a2fbf4b605119e841a13268c4c6efb4ff15ab8ab +size 393535 diff --git a/2023/Change-Aware Sampling and Contrastive Learning for Satellite Images/layout.json b/2023/Change-Aware Sampling and Contrastive Learning for Satellite Images/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..3a2019bae0311178c675f63d6c20a748b0d12b9c --- /dev/null +++ b/2023/Change-Aware Sampling and Contrastive Learning for Satellite Images/layout.json @@ -0,0 +1,8075 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 77, + 103, + 516, + 120 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 103, + 516, + 120 + ], + "spans": [ + { + "bbox": [ + 77, + 103, + 516, + 120 + ], + "type": "text", + "content": "Change-Aware Sampling and Contrastive Learning for Satellite Images" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 140, + 143, + 206, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 140, + 143, + 206, + 156 + ], + "spans": [ + { + "bbox": [ + 140, + 143, + 206, + 156 + ], + "type": "text", + "content": "Utkarsh Mall" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 254, + 144, + 345, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 254, + 144, + 345, + 156 + ], + "spans": [ + { + "bbox": [ + 254, + 144, + 345, + 156 + ], + "type": "text", + "content": "Bharath Hariharan" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 251, + 158, + 342, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 251, + 158, + 342, + 171 + ], + "spans": [ + { + "bbox": [ + 251, + 158, + 342, + 171 + ], + "type": "text", + "content": "Cornell University" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 190, + 173, + 400, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 190, + 173, + 400, + 185 + ], + "spans": [ + { + "bbox": [ + 190, + 173, + 400, + 185 + ], + "type": "text", + "content": "{utkarshm, bharathh, kb}@cs.cornell.edu" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 394, + 144, + 453, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 394, + 144, + 453, + 156 + ], + "spans": [ + { + "bbox": [ + 394, + 144, + 453, + 156 + ], + "type": "text", + "content": "Kavita Bala" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 143, + 213, + 191, + 224 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 213, + 191, + 224 + ], + "spans": [ + { + "bbox": [ + 143, + 213, + 191, + 224 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 239, + 290, + 490 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 239, + 290, + 490 + ], + "spans": [ + { + "bbox": [ + 46, + 239, + 290, + 490 + ], + "type": "text", + "content": "Automatic remote sensing tools can help inform many large-scale challenges such as disaster management, climate change, etc. While a vast amount of spatio-temporal satellite image data is readily available, most of it remains unlabelled. Without labels, this data is not very useful for supervised learning algorithms. Self-supervised learning instead provides a way to learn effective representations for various downstream tasks without labels. In this work, we leverage characteristics unique to satellite images to learn better self-supervised features. Specifically, we use the temporal signal to contrast images with long-term and short-term differences, and we leverage the fact that satellite images do not change frequently. Using these characteristics, we formulate a new loss contrastive loss called Change-Aware Contrastive (CACo) Loss. Further, we also present a novel method of sampling different geographical regions. We show that leveraging these properties leads to better performance on diverse downstream tasks. For example, we see a " + }, + { + "bbox": [ + 46, + 239, + 290, + 490 + ], + "type": "inline_equation", + "content": "6.5\\%" + }, + { + "bbox": [ + 46, + 239, + 290, + 490 + ], + "type": "text", + "content": " relative improvement for semantic segmentation and an " + }, + { + "bbox": [ + 46, + 239, + 290, + 490 + ], + "type": "inline_equation", + "content": "8.5\\%" + }, + { + "bbox": [ + 46, + 239, + 290, + 490 + ], + "type": "text", + "content": " relative improvement for change detection over the best-performing baseline with our method." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 513, + 128, + 525 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 513, + 128, + 525 + ], + "spans": [ + { + "bbox": [ + 47, + 513, + 128, + 525 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 533, + 287, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 533, + 287, + 617 + ], + "spans": [ + { + "bbox": [ + 46, + 533, + 287, + 617 + ], + "type": "text", + "content": "Our planet is surrounded by a large number of satellites constantly collecting images of the world. This massive trove of visual information can help monitor phenomena at the world-scale, and inform solutions to global problems such as climate change or loss of biodiversity. Automatic vision tools can help by, for example, monitoring land-use change over time [36] or the evolution of urban areas [6]." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 617, + 287, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 617, + 287, + 677 + ], + "spans": [ + { + "bbox": [ + 46, + 617, + 287, + 677 + ], + "type": "text", + "content": "However, training all these models requires labeled data. Unfortunately, labeling the massive trove of satellite images is expensive, more so than internet images because of the expertise necessary. This issue is exacerbated by the different label requirements of many monitoring applications." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "text", + "content": "One way to alleviate this problem of limited labeled data is to use self-supervised learning techniques to learn a good feature representation from unlabeled satellite im" + } + ] + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 305, + 209, + 548, + 379 + ], + "blocks": [ + { + "bbox": [ + 305, + 209, + 548, + 379 + ], + "lines": [ + { + "bbox": [ + 305, + 209, + 548, + 379 + ], + "spans": [ + { + "bbox": [ + 305, + 209, + 548, + 379 + ], + "type": "image", + "image_path": "ab9b7b2dfaf29125f6acf27058333da77fb75d35c2c81fb3b1fc3b8c7cf51884.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 385, + 547, + 440 + ], + "lines": [ + { + "bbox": [ + 304, + 385, + 547, + 440 + ], + "spans": [ + { + "bbox": [ + 304, + 385, + 547, + 440 + ], + "type": "text", + "content": "Figure 1. Images of the same location at three different times. Changes from 2016 to 2020 are due to major urban development, while those from July to September are seasonal variations. Our approach, CACo, learns features that are sensitive to the former but invariant to the latter." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 461, + 546, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 461, + 546, + 617 + ], + "spans": [ + { + "bbox": [ + 304, + 461, + 546, + 617 + ], + "type": "text", + "content": "agery. This representation can then be further finetuned with much fewer labels for specific applications. Modern self-supervised learning approaches are based on contrastive learning. These techniques train a feature space so that each image in the dataset is embedded close to augmented versions of itself (e.g., with jittered colors) but far from other images of the dataset. A possible approach is to directly apply these techniques on satellite image datasets. However, the spatio-temporal structure of satellite imagery is much richer than the unstructured collections of internet images typically used in standard self-supervised learning. In this work, we ask, how can we best leverage the structure of satellite images for better self-supervised learning?" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 618, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 618, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 618, + 545, + 713 + ], + "type": "text", + "content": "The first important aspect of the spatio-temporal structure of satellite images is the availability of multiple temporally-spaced images for the same location. Past work has used this structure to sample images spread over a few weeks or months from each location to encourage invariance to seasonal variations [25]. However, we can access images not just over a few months, but over many years. Over such long time spans, we often see significant, per" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 494, + 35 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 494, + 35 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 494, + 35 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "5261" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 167 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 167 + ], + "type": "text", + "content": "manent change, such as the construction of houses, the drying of lakes, or the logging of forests (Fig. 1). While we want feature representations to be invariant to temporary, seasonal change, we want the representation to be sensitive to permanent, long-term change. To capture this intuition, we sample multiple images from the same location covering both short and long time spans and encourage invariance to the former but sensitivity to the latter." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 168, + 289, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 168, + 289, + 323 + ], + "spans": [ + { + "bbox": [ + 46, + 168, + 289, + 323 + ], + "type": "text", + "content": "A second and more important aspect of satellite images is that permanent change is spatially rare. Change is concentrated near urban areas, but many parts of the planet see little change. When there is no change, we want our feature representation to be the same even over long time spans. We capture this insight with a novel strategy to robustly estimate whether or not there is a long-term change, even in the middle of training, by comparing long-term feature differences to short-term variations. We then design a loss function that is conditional on this change estimate: it encourages invariances to long-term differences depending on whether a change occurs or not. We call this novel loss function Change-aware Contrastive Loss (or CaCo)." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 323, + 289, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 323, + 289, + 491 + ], + "spans": [ + { + "bbox": [ + 46, + 323, + 289, + 491 + ], + "type": "text", + "content": "The above loss function uses the temporal structure of satellite images. We can also utilize geographical structure by carefully sampling the most informative locations on the planet. We provide an improvement over a previously proposed geographical sampling [25]. We show that sampling closer to cities, and ignoring samples completely in the ocean can result in a dataset much more useful for learning a general representation for various downstream tasks. We evaluate our new representation on a diverse set of downstream tasks such as landcover classification, semantic segmentation, and change detection. Our method achieves significant relative improvements (ranging from " + }, + { + "bbox": [ + 46, + 323, + 289, + 491 + ], + "type": "inline_equation", + "content": "6.5\\%" + }, + { + "bbox": [ + 46, + 323, + 289, + 491 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 46, + 323, + 289, + 491 + ], + "type": "inline_equation", + "content": "8.5\\%" + }, + { + "bbox": [ + 46, + 323, + 289, + 491 + ], + "type": "text", + "content": ") over the state-of-the-art for a variety of tasks such as segmentation and change detection." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 59, + 491, + 272, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 491, + 272, + 502 + ], + "spans": [ + { + "bbox": [ + 59, + 491, + 272, + 502 + ], + "type": "text", + "content": "To summarize, we make the following contributions:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 58, + 510, + 287, + 647 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 58, + 510, + 287, + 558 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 510, + 287, + 558 + ], + "spans": [ + { + "bbox": [ + 58, + 510, + 287, + 558 + ], + "type": "text", + "content": "- We propose a new self-supervised learning loss that uses long-term temporal information in satellite imagery to encourage invariance to seasonal variations but sensitivity to permanent, long-term changes." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 58, + 561, + 287, + 620 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 561, + 287, + 620 + ], + "spans": [ + { + "bbox": [ + 58, + 561, + 287, + 620 + ], + "type": "text", + "content": "- We introduce a novel approach to robustly estimate whether a location has undergone significant change by comparing long-term changes to seasonal variations. Our new change-aware loss function (CACo) uses this to decide when to encourage invariance." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 58, + 624, + 287, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 624, + 287, + 647 + ], + "spans": [ + { + "bbox": [ + 58, + 624, + 287, + 647 + ], + "type": "text", + "content": "- We use an improved geographical sampling that provides more diverse data for representation learning." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 47, + 657, + 134, + 669 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 657, + 134, + 669 + ], + "spans": [ + { + "bbox": [ + 47, + 657, + 134, + 669 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 677, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 288, + 713 + ], + "type": "text", + "content": "Self-supervised Learning. Self-supervised learning methods can be used to learn a good general representation from a dataset without any labels. These methods obtain" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 72, + 547, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 217 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 217 + ], + "type": "text", + "content": "supervisory signals or inductive bias from the unlabelled data itself. Earlier self-supervised works used \"pretext tasks\" such as rotation prediction [15], solving jigsaw puzzles [28], colorization [44], or missing data completion [17]. More recently contrastive learning methods have been shown to learn a better representation than pretext tasks. Methods for contrastive learning such as NPID [43], PIRL [26], MoCo [8, 18], SimCLR [7], and BYOL [16], use instance discrimination to learn a feature representation. While these methods can be applied even to satellite images, in this work we leverage the structure of satellite images that can be used for self-supervised learning." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 231, + 547, + 316 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 231, + 547, + 316 + ], + "spans": [ + { + "bbox": [ + 304, + 231, + 547, + 316 + ], + "type": "text", + "content": "Self-supervised Learning in Remote Sensing. Increasingly, self-supervised learning is being applied to the area of remote sensing. Prior works have used signals from unlabeled data such as location [2, 20], seasonal variations [25], or texture [1], as a signal for self-supervision. Recent methods use transformers [12] and diffusion models [4] for self-supervised representation [10, 32]." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 329, + 547, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 329, + 547, + 533 + ], + "spans": [ + { + "bbox": [ + 304, + 329, + 547, + 533 + ], + "type": "text", + "content": "Remote Sensing Applications and Data Several remote sensing applications require vision tools for automatic recognition at scale. Applications such as landcover classification [19, 34] or segmentation [36, 40] benefit from the advances in vision algorithms for scene classification and semantic segmentation. Several object detection applications have also been developed using object detection algorithms such as for detecting constructions [9], buildings [38], trees [41], or floating objects [14]. Segmentation methods also help in extracting structural information like road networks [3, 22, 39], crop types [37], or clouds [27] from remote sensing images. Many applications such as change detection [11] require temporal information for regions as well. Temporal information and changes are also used to find long-term semantic events [24]. In this work, we aim to learn a representation without supervision, that leads to better performance on all these downstream tasks." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 533, + 547, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 533, + 547, + 641 + ], + "spans": [ + { + "bbox": [ + 304, + 533, + 547, + 641 + ], + "type": "text", + "content": "The image data for these algorithms and datasets are obtained from a few different sources. Sentinel-2 [13] provides multispectral images (13 bands) at " + }, + { + "bbox": [ + 304, + 533, + 547, + 641 + ], + "type": "inline_equation", + "content": "10\\mathrm{m}" + }, + { + "bbox": [ + 304, + 533, + 547, + 641 + ], + "type": "text", + "content": " resolution with a temporal revisit of 5 days. Several other satellites provide higher resolution information such as WorldView-3 and PlanetScope. We use Sentinel-2 satellite imagery as it provides frequent temporal information that we use in our self-supervised formulation. Furthermore, many of the downstream tasks [11, 19, 24, 36] use Sentinel-2 imagery." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 651, + 362, + 662 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 651, + 362, + 662 + ], + "spans": [ + { + "bbox": [ + 306, + 651, + 362, + 662 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 670, + 373, + 682 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 670, + 373, + 682 + ], + "spans": [ + { + "bbox": [ + 306, + 670, + 373, + 682 + ], + "type": "text", + "content": "3.1. Overview" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 689, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 547, + 715 + ], + "type": "text", + "content": "We propose a self-supervised method that can leverage properties that are unique to satellite images. One such" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5262" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 239 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 239 + ], + "type": "text", + "content": "property is the availability of long-term temporal information for any location on earth. For example, two images of the same location captured months apart will likely be the same except for seasonal variations; variations our features should be invariant to. In contrast, two images captured decades apart might show big, structural changes such as new buildings or sparser forests. Of course, this may not be true for all locations. Indeed, such long-term change is rare, and is concentrated in regions of human activity. But these long-term changes are important for many downstream applications such as modeling urban growth, or monitoring climate change. Our key insight is therefore to find these rare but meaningful long-term changes, and ensure that they are reflected as changes in the feature space as well." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 239, + 289, + 348 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 239, + 289, + 348 + ], + "spans": [ + { + "bbox": [ + 46, + 239, + 289, + 348 + ], + "type": "text", + "content": "Past work on using temporal data for representation learning on satellite imagery only uses short-term seasonal variations but not long-term change [25]. We review this work, which is called Seasonal Contrast (or SeCo) [25], in Sec. 3.2. We then explain how we can use long-term temporal information in Sec. 3.3, and address the challenge of spatial rarity in Sec. 3.4. Finally, we look at how we can efficiently sample more informative locations when collecting unlabelled training data in Sec. 3.5." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 354, + 255, + 369 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 354, + 255, + 369 + ], + "spans": [ + { + "bbox": [ + 47, + 354, + 255, + 369 + ], + "type": "text", + "content": "3.2. Background: Seasonal Contrast (SeCo)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 373, + 289, + 566 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 373, + 289, + 566 + ], + "spans": [ + { + "bbox": [ + 46, + 373, + 289, + 566 + ], + "type": "text", + "content": "Contrastive learning techniques train feature representations from unlabeled images by discriminating between individual instances. In particular, they pull augmentations of the same image closer together in feature space and push apart representations of two different images. SeCo is a contrastive learning-based representation learning framework for remote sensing images. SeCo uses MoCo v2 [8], to perform contrastive learning. In MoCo v2, in each iteration, one generates two views for an image using random augmentations. One is the \"query\" image " + }, + { + "bbox": [ + 46, + 373, + 289, + 566 + ], + "type": "inline_equation", + "content": "I_{q}" + }, + { + "bbox": [ + 46, + 373, + 289, + 566 + ], + "type": "text", + "content": " and the other is the \"positive key\" " + }, + { + "bbox": [ + 46, + 373, + 289, + 566 + ], + "type": "inline_equation", + "content": "I_{k+}" + }, + { + "bbox": [ + 46, + 373, + 289, + 566 + ], + "type": "text", + "content": " that is pulled closer to the query in the representation space. It also has a set of other images from the dataset, " + }, + { + "bbox": [ + 46, + 373, + 289, + 566 + ], + "type": "inline_equation", + "content": "\\mathcal{I}_{k-}" + }, + { + "bbox": [ + 46, + 373, + 289, + 566 + ], + "type": "text", + "content": ", that are treated as negative keys and are pushed apart from query image " + }, + { + "bbox": [ + 46, + 373, + 289, + 566 + ], + "type": "inline_equation", + "content": "I_{q}" + }, + { + "bbox": [ + 46, + 373, + 289, + 566 + ], + "type": "text", + "content": " in the latent space. The precise loss function in MoCo v2 called InfoNCE [29], for a representation function " + }, + { + "bbox": [ + 46, + 373, + 289, + 566 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 46, + 373, + 289, + 566 + ], + "type": "text", + "content": ", can be written as:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 70, + 573, + 287, + 609 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 573, + 287, + 609 + ], + "spans": [ + { + "bbox": [ + 70, + 573, + 287, + 609 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = - \\log \\frac {\\exp \\left(f \\left(I _ {q}\\right) \\cdot f \\left(I _ {k ^ {+}}\\right) / \\tau\\right)}{\\sum_ {I _ {k} \\in \\mathcal {I} _ {k ^ {-}} \\cup \\{I _ {k +} \\}} \\exp \\left(f \\left(I _ {q}\\right) \\cdot f \\left(I _ {k}\\right) / \\tau\\right)} \\tag {1}", + "image_path": "87e8c5f6623a4b8941007ce559dc7dfda42da6c2f219d74c9cf6988b496d1d55.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 617, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 617, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 617, + 289, + 715 + ], + "type": "text", + "content": "The key idea in SeCo is to also use short-term temporal differences (over a few months) at a location as augmentations. Since these images are a few months apart, these images represent different seasons (hence the name seasonal contrast). The negative keys are then images from other locations. More specifically SeCo trains a representation with 3 subspaces, i) a subspace invariant to both season and artificial augmentations, ii) a subspace invariant to season only," + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 72, + 546, + 109 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 546, + 109 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 546, + 109 + ], + "type": "text", + "content": "and iii) a subspace invariant to augmentation only. The additional invariance to seasonal differences results in a better representation than simply using MoCo v2 on this data." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 122, + 517, + 136 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 122, + 517, + 136 + ], + "spans": [ + { + "bbox": [ + 305, + 122, + 517, + 136 + ], + "type": "text", + "content": "3.3. Using Long-term Temporal Information" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 143, + 545, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 143, + 545, + 203 + ], + "spans": [ + { + "bbox": [ + 304, + 143, + 545, + 203 + ], + "type": "text", + "content": "SeCo introduces invariance to seasonal variations. However, invariance is not enough. We also want our feature representation to be informative. For this, simply pushing apart images from different locations (which already look quite different) may not be enough." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 205, + 546, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 205, + 546, + 373 + ], + "spans": [ + { + "bbox": [ + 304, + 205, + 546, + 373 + ], + "type": "text", + "content": "As discussed above, we argue that to produce a richer feature representation, we should use information from longer temporal spans. Over the span of years, we might observe big, structural changes such as new construction or urban growth. We want our feature representation to be sensitive to these major changes. Thus, while short-term temporal differences are typically seasonal and should be treated as augmentations as in SeCo, long-term temporal differences might correspond to permanent change, and should be reflected as a significant change in the underlying feature representation. This suggests that for a query image from a particular location, an image captured at the same location several years later should be treated as a negative in the contrastive learning objective. We formalize this below." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 375, + 547, + 474 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 375, + 547, + 474 + ], + "spans": [ + { + "bbox": [ + 304, + 375, + 547, + 474 + ], + "type": "text", + "content": "Data: Let " + }, + { + "bbox": [ + 304, + 375, + 547, + 474 + ], + "type": "inline_equation", + "content": "\\{l_1, l_2 \\dots l_n\\}" + }, + { + "bbox": [ + 304, + 375, + 547, + 474 + ], + "type": "text", + "content": " be " + }, + { + "bbox": [ + 304, + 375, + 547, + 474 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 304, + 375, + 547, + 474 + ], + "type": "text", + "content": " different locations in our dataset. For each location, we have 2 sets of images captured around time points " + }, + { + "bbox": [ + 304, + 375, + 547, + 474 + ], + "type": "inline_equation", + "content": "t_1" + }, + { + "bbox": [ + 304, + 375, + 547, + 474 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 375, + 547, + 474 + ], + "type": "inline_equation", + "content": "t_2" + }, + { + "bbox": [ + 304, + 375, + 547, + 474 + ], + "type": "text", + "content": " that have a gap of 4 years between them. We represent these sets by " + }, + { + "bbox": [ + 304, + 375, + 547, + 474 + ], + "type": "inline_equation", + "content": "\\mathcal{I}_{l_i}^{t_1}" + }, + { + "bbox": [ + 304, + 375, + 547, + 474 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 375, + 547, + 474 + ], + "type": "inline_equation", + "content": "\\mathcal{I}_{l_i}^{t_2}" + }, + { + "bbox": [ + 304, + 375, + 547, + 474 + ], + "type": "text", + "content": ". Each set " + }, + { + "bbox": [ + 304, + 375, + 547, + 474 + ], + "type": "inline_equation", + "content": "\\mathcal{I}_{l_i}^{t_j}" + }, + { + "bbox": [ + 304, + 375, + 547, + 474 + ], + "type": "text", + "content": ", contains images sampled over a few months around " + }, + { + "bbox": [ + 304, + 375, + 547, + 474 + ], + "type": "inline_equation", + "content": "t_j" + }, + { + "bbox": [ + 304, + 375, + 547, + 474 + ], + "type": "text", + "content": ", and thus these images have only short-term or seasonal differences between them. See Fig. 2 (a) for a schematic of this data sampling strategy." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 475, + 547, + 589 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 475, + 547, + 589 + ], + "spans": [ + { + "bbox": [ + 304, + 475, + 547, + 589 + ], + "type": "text", + "content": "Loss: Let " + }, + { + "bbox": [ + 304, + 475, + 547, + 589 + ], + "type": "inline_equation", + "content": "I_{l_i}^{t_j + \\Delta_1}, I_{l_i}^{t_j + \\Delta_2} \\sim \\mathcal{I}_{l_i}^{t_j}" + }, + { + "bbox": [ + 304, + 475, + 547, + 589 + ], + "type": "text", + "content": " be two images sampled randomly from " + }, + { + "bbox": [ + 304, + 475, + 547, + 589 + ], + "type": "inline_equation", + "content": "\\mathcal{I}_{l_i}^{t_j}" + }, + { + "bbox": [ + 304, + 475, + 547, + 589 + ], + "type": "text", + "content": " ( " + }, + { + "bbox": [ + 304, + 475, + 547, + 589 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 304, + 475, + 547, + 589 + ], + "type": "text", + "content": " is 1 or 2). " + }, + { + "bbox": [ + 304, + 475, + 547, + 589 + ], + "type": "inline_equation", + "content": "I_{l_i}^{t_j + \\Delta_1}" + }, + { + "bbox": [ + 304, + 475, + 547, + 589 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 475, + 547, + 589 + ], + "type": "inline_equation", + "content": "I_{l_i}^{t_j + \\Delta_2}" + }, + { + "bbox": [ + 304, + 475, + 547, + 589 + ], + "type": "text", + "content": " are a few months apart. So like SeCo, we still want their representations to be pulled closer. However, images with larger time differences, for example, " + }, + { + "bbox": [ + 304, + 475, + 547, + 589 + ], + "type": "inline_equation", + "content": "I_{l_i}^{t_1 + \\Delta_1}" + }, + { + "bbox": [ + 304, + 475, + 547, + 589 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 475, + 547, + 589 + ], + "type": "inline_equation", + "content": "I_{l_i}^{t_2 + \\Delta_1}" + }, + { + "bbox": [ + 304, + 475, + 547, + 589 + ], + "type": "text", + "content": " should potentially be pushed apart. Finally, images from two different locations such as " + }, + { + "bbox": [ + 304, + 475, + 547, + 589 + ], + "type": "inline_equation", + "content": "I_{l_i}^{t_j + \\Delta_1}" + }, + { + "bbox": [ + 304, + 475, + 547, + 589 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 475, + 547, + 589 + ], + "type": "inline_equation", + "content": "I_{l_j}^{t_j + \\Delta_1}" + }, + { + "bbox": [ + 304, + 475, + 547, + 589 + ], + "type": "text", + "content": ", should always be pushed apart irrespective of time. This intuition is reflected in Fig. 2 (c)." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 591, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 591, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 591, + 547, + 714 + ], + "type": "text", + "content": "Following SeCo, we implement this loss function using MoCo v2 by adapting the positive key image " + }, + { + "bbox": [ + 304, + 591, + 547, + 714 + ], + "type": "inline_equation", + "content": "I_{k+}" + }, + { + "bbox": [ + 304, + 591, + 547, + 714 + ], + "type": "text", + "content": " and negative key set " + }, + { + "bbox": [ + 304, + 591, + 547, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{I}_{k-}" + }, + { + "bbox": [ + 304, + 591, + 547, + 714 + ], + "type": "text", + "content": " in Eq. (1). As discussed above, SeCo has 3 subspaces. We add our modifications to all 3 subspaces. However, for ease of explanation, we only describe the change to the subspace that is invariant to seasons. For this subspace, for a query image, " + }, + { + "bbox": [ + 304, + 591, + 547, + 714 + ], + "type": "inline_equation", + "content": "I_q = I_{l_i}^{t_1 + \\Delta_1}" + }, + { + "bbox": [ + 304, + 591, + 547, + 714 + ], + "type": "text", + "content": ", the positive key is the same as in SeCo: the seasonal pair " + }, + { + "bbox": [ + 304, + 591, + 547, + 714 + ], + "type": "inline_equation", + "content": "(I_{k^+} = I_{l_i}^{t_1 + \\Delta_2})" + }, + { + "bbox": [ + 304, + 591, + 547, + 714 + ], + "type": "text", + "content": ". But we modify the negative key set to not only include images from other locations but also images from the same" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5263" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 96, + 67, + 282, + 293 + ], + "blocks": [ + { + "bbox": [ + 96, + 67, + 282, + 293 + ], + "lines": [ + { + "bbox": [ + 96, + 67, + 282, + 293 + ], + "spans": [ + { + "bbox": [ + 96, + 67, + 282, + 293 + ], + "type": "image", + "image_path": "e3ad505694a5a6acdc45320509ab95a7eaec1cdf354572ba08e1a76180b7b845.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 300, + 545, + 377 + ], + "lines": [ + { + "bbox": [ + 46, + 300, + 545, + 377 + ], + "spans": [ + { + "bbox": [ + 46, + 300, + 545, + 377 + ], + "type": "text", + "content": "Figure 2. Overview of our self-supervised framework. (a) Shows the notation for images and the temporal sampling strategy we use to create our dataset for training. (b) Shows the contrastive learning performed by SeCo in the temporally invariant subspace. The red arrows show examples that need to be pushed apart from the top left example. The green arrows indicate examples to be pulled closer. Different lightness (value) represents different short-term temporal pairs and different colors (hue) indicate different locations (see Sec. 3.2). (c) Shows our method when using long-term temporal pairs. Different shapes indicate a difference of 4 years between examples (see Sec. 3.3). (d) Shows the contrastive learning performed by our CACo loss term. The dashed red arrows indicate examples, that are pushed apart iff our method estimates a change has occurred between the long-term pairs (see Sec. 3.4)." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 287, + 66, + 390, + 291 + ], + "blocks": [ + { + "bbox": [ + 287, + 66, + 390, + 291 + ], + "lines": [ + { + "bbox": [ + 287, + 66, + 390, + 291 + ], + "spans": [ + { + "bbox": [ + 287, + 66, + 390, + 291 + ], + "type": "image", + "image_path": "5083f2133d87bab85501d07caa55bca49e1d4029e352144e91d68b6572c31317.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 395, + 68, + 497, + 290 + ], + "blocks": [ + { + "bbox": [ + 395, + 68, + 497, + 290 + ], + "lines": [ + { + "bbox": [ + 395, + 68, + 497, + 290 + ], + "spans": [ + { + "bbox": [ + 395, + 68, + 497, + 290 + ], + "type": "image", + "image_path": "26811b45b3a8928490894dcb740b3cc094ddec8118ec3dc40bef9afc50ce1c1c.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 399, + 240, + 410 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 399, + 240, + 410 + ], + "spans": [ + { + "bbox": [ + 47, + 399, + 240, + 410 + ], + "type": "text", + "content": "location with a large time difference as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 420, + 286, + 454 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 420, + 286, + 454 + ], + "spans": [ + { + "bbox": [ + 56, + 420, + 286, + 454 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {I} _ {k ^ {-}} = \\left\\{I _ {l _ {i}} ^ {t _ {2} + \\Delta_ {k}}: k \\in \\{1, 2 \\} \\right\\} \\cup \\\\ \\left\\{I _ {l _ {h}} ^ {t _ {j} + \\Delta_ {k}}: h \\neq i, j \\in \\{1, 2 \\}, k \\in \\{1, 2 \\} \\right\\} \\quad (2) \\\\ \\end{array}", + "image_path": "90bdc7fd350c38b7ddeb94ed717671d31aca4bee835905a630751f5410d3b0dc.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 461, + 286, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 461, + 286, + 496 + ], + "spans": [ + { + "bbox": [ + 46, + 461, + 286, + 496 + ], + "type": "text", + "content": "In each training iteration, we sample 4 images (long and short-time pairs) for each location instead of 2 like SeCo, so each batch contains half as many locations." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 503, + 241, + 516 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 503, + 241, + 516 + ], + "spans": [ + { + "bbox": [ + 47, + 503, + 241, + 516 + ], + "type": "text", + "content": "3.4. Change-aware Contrastive Learning" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 522, + 286, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 522, + 286, + 605 + ], + "spans": [ + { + "bbox": [ + 46, + 522, + 286, + 605 + ], + "type": "text", + "content": "The training method in the previous section assumes that regions always change after a long time (a few years). However, this assumption is not always true: in remote areas away from human activity, there may be no change even over several years. For such locations, pushing apart features of temporally distant image pairs may be counterproductive and destroy needed invariances." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 606, + 286, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 606, + 286, + 677 + ], + "spans": [ + { + "bbox": [ + 46, + 606, + 286, + 677 + ], + "type": "text", + "content": "If we know whether a location has changed or not, we can make our loss function conditional on this information. If it has changed between " + }, + { + "bbox": [ + 46, + 606, + 286, + 677 + ], + "type": "inline_equation", + "content": "I_{l_i}^{t_1 + \\Delta_1}" + }, + { + "bbox": [ + 46, + 606, + 286, + 677 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 606, + 286, + 677 + ], + "type": "inline_equation", + "content": "I_{l_i}^{t_2 + \\Delta_1}" + }, + { + "bbox": [ + 46, + 606, + 286, + 677 + ], + "type": "text", + "content": ", we push them apart as described in Sec. 3.3. If not, we refrain from adding the temporally distant image to the negative set. This conditional loss function is illustrated in Fig. 2 (d)." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 677, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 286, + 713 + ], + "type": "text", + "content": "However, the challenge is knowing which locations have changed. Change detection can be performed by looking at feature differences between temporally separated image" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 399, + 545, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 399, + 545, + 590 + ], + "spans": [ + { + "bbox": [ + 304, + 399, + 545, + 590 + ], + "type": "text", + "content": "pairs [21, 23, 35, 42], but this requires a feature representation we do not have. We solve this chicken-and-egg problem by bootstrapping our feature representation. We start from a randomly initialized model, using which we extract features for different locations and use feature differences as an estimate for change. Using these change values, we use our conditional loss to train the model. We then keep alternating between training the model using changes and estimating the changes using the partially trained model. In the initial stages, the change estimates could be very poor as the model has random weights. However, as training progresses we expect better features, and, in turn, better estimates of changes. One might worry about degenerate solutions, but the other contrastive loss terms based on location (that do not depend on this inferred change) ensure that the feature representation avoids any pathological cases." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": "Obtaining changes using features. The absolute distance between features scales differently for different types of locations, and does not directly reflect the magnitude of actual change. For example, urban locations with less vegetation look the same over the seasons, but temperate forests will change a lot with the seasons. So instead of using the absolute distance between two long-term images, we instead use the ratio of distances between long-term images and short-term images (images with a seasonal difference)." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5264" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "text", + "content": "This ratio normalizes the scaling differences between images from different locations. If " + }, + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "inline_equation", + "content": "f_{k}" + }, + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "text", + "content": " is the partially trained feature extractor at epoch " + }, + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "text", + "content": ", we use the following ratio as an indicator of change for location " + }, + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "inline_equation", + "content": "l_{i}" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 90, + 126, + 287, + 159 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 90, + 126, + 287, + 159 + ], + "spans": [ + { + "bbox": [ + 90, + 126, + 287, + 159 + ], + "type": "interline_equation", + "content": "R _ {l _ {i}} ^ {k} = \\frac {\\left| \\left| f _ {k} \\left(I _ {l _ {i}} ^ {t _ {1} + \\Delta_ {1}}\\right) - f _ {k} \\left(I _ {l _ {i}} ^ {t _ {2} + \\Delta_ {1}}\\right) \\right| \\right| ^ {2}}{\\left| \\left| f _ {k} \\left(I _ {l _ {i}} ^ {t _ {1} + \\Delta_ {1}}\\right) - f _ {k} \\left(I _ {l _ {i}} ^ {t _ {1} + \\Delta_ {2}}\\right) \\right| \\right| ^ {2}} \\tag {3}", + "image_path": "8c37b8437e4b7d63c72d350934a0fe10a3a9fa20b55be6e32cd3f01b046e5fbe.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 166, + 287, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 166, + 287, + 213 + ], + "spans": [ + { + "bbox": [ + 46, + 166, + 287, + 213 + ], + "type": "text", + "content": "Since, at each epoch, the seasonal images are selected randomly for each location, we use an exponential moving average over ratios from previous epochs to obtain a more stable estimate for change as follows" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 111, + 220, + 221, + 237 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 111, + 220, + 221, + 237 + ], + "spans": [ + { + "bbox": [ + 111, + 220, + 221, + 237 + ], + "type": "interline_equation", + "content": "\\mathbf {r} _ {l _ {i}} ^ {k} = (1 - \\beta) R _ {l _ {i}} ^ {k} + \\beta \\mathbf {r} _ {l _ {i}} ^ {k - 1}", + "image_path": "aee3ba65e3cab29b4430a958945141b3796f27b93e2a5f7eb5260a345c11ecea.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 243, + 287, + 305 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 243, + 287, + 305 + ], + "spans": [ + { + "bbox": [ + 46, + 243, + 287, + 305 + ], + "type": "text", + "content": "We use a Gaussian Mixture Model (GMM) on " + }, + { + "bbox": [ + 46, + 243, + 287, + 305 + ], + "type": "inline_equation", + "content": "\\mathbf{r}_{l_i}^k" + }, + { + "bbox": [ + 46, + 243, + 287, + 305 + ], + "type": "text", + "content": " values to find our two clusters for change and no change. Locations with smaller " + }, + { + "bbox": [ + 46, + 243, + 287, + 305 + ], + "type": "inline_equation", + "content": "\\mathbf{r}_{l_i}^k" + }, + { + "bbox": [ + 46, + 243, + 287, + 305 + ], + "type": "text", + "content": " are treated as having no long-term change, and those with larger " + }, + { + "bbox": [ + 46, + 243, + 287, + 305 + ], + "type": "inline_equation", + "content": "\\mathbf{r}_{l_i}^k" + }, + { + "bbox": [ + 46, + 243, + 287, + 305 + ], + "type": "text", + "content": " values are deemed to have undergone significant change (see Fig. 3)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 306, + 287, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 306, + 287, + 389 + ], + "spans": [ + { + "bbox": [ + 46, + 306, + 287, + 389 + ], + "type": "text", + "content": "For MoCo v2, we use features from the momentum encoder for images to calculate the ratio estimate. To maintain training speed, we compute and store the ratios during the training iterations themselves, even as the feature extractor is being updated. In preliminary experiments, we found that the exponential moving average was sufficient to regularize against the noise induced by the changing feature extractor." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 389, + 287, + 485 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 389, + 287, + 485 + ], + "spans": [ + { + "bbox": [ + 46, + 389, + 287, + 485 + ], + "type": "text", + "content": "Our change-aware formulation uses change information in long-term pairs. On the other hand, there could be short-term pairs with sudden changes such as due to disasters like landslides or earthquakes. However, we notice that in practice such changes are extremely rare. Out of the top 200 examples with the highest short-term change score only 3 pairs showed a sudden change. Because short-term changes are so rare, it is not useful to consider them in our pipeline." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 491, + 233, + 505 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 491, + 233, + 505 + ], + "spans": [ + { + "bbox": [ + 47, + 491, + 233, + 505 + ], + "type": "text", + "content": "3.5. Improving Geographical Sampling" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 510, + 287, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 510, + 287, + 569 + ], + "spans": [ + { + "bbox": [ + 46, + 510, + 287, + 569 + ], + "type": "text", + "content": "Another property specific to satellite imagery is that we can control the geographic distribution of sampled images. Leveraging this freedom, SeCo collects data from a " + }, + { + "bbox": [ + 46, + 510, + 287, + 569 + ], + "type": "inline_equation", + "content": "100\\mathrm{km}" + }, + { + "bbox": [ + 46, + 510, + 287, + 569 + ], + "type": "text", + "content": " radius around the 10k most populated cities of the world. However, we argue that 100 kms is too large a radius." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 570, + 287, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 570, + 287, + 617 + ], + "spans": [ + { + "bbox": [ + 46, + 570, + 287, + 617 + ], + "type": "text", + "content": "Sampling far from cities results in uninformative images (such as images entirely over oceans, (see supplementary Fig. 3)). In fact, " + }, + { + "bbox": [ + 46, + 570, + 287, + 617 + ], + "type": "inline_equation", + "content": "22\\%" + }, + { + "bbox": [ + 46, + 570, + 287, + 617 + ], + "type": "text", + "content": " of the dataset has such uninformative images and this hurts the quality of the representation." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 618, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 618, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 618, + 287, + 714 + ], + "type": "text", + "content": "The median area of the 10k cities sampled is approximately " + }, + { + "bbox": [ + 46, + 618, + 287, + 714 + ], + "type": "inline_equation", + "content": "24\\mathrm{km}^2" + }, + { + "bbox": [ + 46, + 618, + 287, + 714 + ], + "type": "text", + "content": ", which represents a radius of " + }, + { + "bbox": [ + 46, + 618, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\sim 5\\mathrm{km}" + }, + { + "bbox": [ + 46, + 618, + 287, + 714 + ], + "type": "text", + "content": ". Based on this observation, we use 2d-gaussian sampling with " + }, + { + "bbox": [ + 46, + 618, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\sigma = 5\\mathrm{km}" + }, + { + "bbox": [ + 46, + 618, + 287, + 714 + ], + "type": "text", + "content": ". This means that " + }, + { + "bbox": [ + 46, + 618, + 287, + 714 + ], + "type": "inline_equation", + "content": "95\\%" + }, + { + "bbox": [ + 46, + 618, + 287, + 714 + ], + "type": "text", + "content": " of the data comes from a radius of " + }, + { + "bbox": [ + 46, + 618, + 287, + 714 + ], + "type": "inline_equation", + "content": "2\\sigma = 10\\mathrm{km}" + }, + { + "bbox": [ + 46, + 618, + 287, + 714 + ], + "type": "text", + "content": " around the city center. Additionally, if the sampled location is completely in the ocean, we reject that sample and sample again (We use a landmass map of the earth to check this). Our improved sampling" + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 309, + 77, + 545, + 203 + ], + "blocks": [ + { + "bbox": [ + 309, + 77, + 545, + 203 + ], + "lines": [ + { + "bbox": [ + 309, + 77, + 545, + 203 + ], + "spans": [ + { + "bbox": [ + 309, + 77, + 545, + 203 + ], + "type": "image", + "image_path": "14d85addedef9701805eb72ab54aaebe51556483a43dc3aab6932df45c543dd8.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 212, + 545, + 267 + ], + "lines": [ + { + "bbox": [ + 305, + 212, + 545, + 267 + ], + "spans": [ + { + "bbox": [ + 305, + 212, + 545, + 267 + ], + "type": "text", + "content": "Figure 3. The histogram of change estimates separated into two groups (blue and green) by the GMM. Pairs of examples with a higher estimate can be seen to have major construction and land-cover changes (2 pairs on the right). Whereas pairs with a lower estimate only have seasonal changes." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 289, + 545, + 326 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 289, + 545, + 326 + ], + "spans": [ + { + "bbox": [ + 305, + 289, + 545, + 326 + ], + "type": "text", + "content": "methods result in a dataset with a more diverse set of locations that are more informative. As we show in experiments, this additionally improves the quality of the representation." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 338, + 359, + 350 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 338, + 359, + 350 + ], + "spans": [ + { + "bbox": [ + 306, + 338, + 359, + 350 + ], + "type": "text", + "content": "4. Results" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 358, + 503, + 371 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 358, + 503, + 371 + ], + "spans": [ + { + "bbox": [ + 306, + 358, + 503, + 371 + ], + "type": "text", + "content": "4.1. Implementation details and Baselines" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 305, + 377, + 545, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 377, + 545, + 413 + ], + "spans": [ + { + "bbox": [ + 305, + 377, + 545, + 413 + ], + "type": "text", + "content": "We evaluate our model by measuring its representation ability on various downstream tasks. We evaluate the following baselines for comparison." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 305, + 413, + 544, + 436 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 413, + 544, + 436 + ], + "spans": [ + { + "bbox": [ + 305, + 413, + 544, + 436 + ], + "type": "text", + "content": "Random init.: is randomly initialized and is either fin-tuned or linear probed for downstream tasks." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 306, + 437, + 530, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 437, + 530, + 449 + ], + "spans": [ + { + "bbox": [ + 306, + 437, + 530, + 449 + ], + "type": "text", + "content": "ImageNet: is pre-trained on supervised ImageNet [31]." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 306, + 449, + 544, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 449, + 544, + 472 + ], + "spans": [ + { + "bbox": [ + 306, + 449, + 544, + 472 + ], + "type": "text", + "content": "MoCo v2: uses self-supervised learning without the temporal or change information." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 306, + 473, + 544, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 473, + 544, + 496 + ], + "spans": [ + { + "bbox": [ + 306, + 473, + 544, + 496 + ], + "type": "text", + "content": "SeCo: is Seasonal Contrast [25] and uses only short-term changes as augmentations." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 306, + 497, + 544, + 532 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 497, + 544, + 532 + ], + "spans": [ + { + "bbox": [ + 306, + 497, + 544, + 532 + ], + "type": "text", + "content": "GSSL: uses temporal and geographical information for contrastive learning [2]. Since the data this model was trained on is of different resolution retrain it on our data." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 306, + 533, + 544, + 591 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 533, + 544, + 591 + ], + "spans": [ + { + "bbox": [ + 306, + 533, + 544, + 591 + ], + "type": "text", + "content": "SatMAE [10]: is a Masked autoencoder [17] that uses ViT-Large as backbone. The comparison is not apples-to-apples as the model has larger capacity. The model is trained on 713k images therefore it is comparable to our 1m images trained model." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 306, + 593, + 544, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 593, + 544, + 616 + ], + "spans": [ + { + "bbox": [ + 306, + 593, + 544, + 616 + ], + "type": "text", + "content": "DDPM-CD [4]: uses a denoising diffusion model as backbone. It is also trained on a million scale dataset." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 305, + 629, + 545, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 629, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 305, + 629, + 545, + 689 + ], + "type": "text", + "content": "We show results with both a ResNet-18 and a ResNet-50 backbone. Similar to SeCo our models are trained with a batch size of 256 and with 16,384 negative embeddings. We also use the same optimizer, learning rate, schedule, and temperature scaling as in the SeCo implementation." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "content": "We experiment with two datasets for self-supervised pretraining: one with 100k images and the other with 1 million" + } + ] + } + ], + "index": 25 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5265" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 65, + 71, + 271, + 220 + ], + "blocks": [ + { + "bbox": [ + 65, + 71, + 271, + 220 + ], + "lines": [ + { + "bbox": [ + 65, + 71, + 271, + 220 + ], + "spans": [ + { + "bbox": [ + 65, + 71, + 271, + 220 + ], + "type": "table", + "html": "
DataPre-trainingResNet-18ResNet-50
-Random init.64.2155.32
ImageNet.86.1689.08
100kMoCo v287.2289.75
GSSL87.7490.19
SeCo90.0593.12
CACo (ours)93.0894.48
1mSeCo93.9995.63
SatMAE (ViT-L)-93.03
DDPM-CD-87.67
CACo (ours)94.7295.90
", + "image_path": "3753e9df59c1b5ef9003ed6b056d524b4a83c9df27aadd8e40f6d81d2e283c59.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 285, + 287, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 285, + 287, + 350 + ], + "spans": [ + { + "bbox": [ + 46, + 285, + 287, + 350 + ], + "type": "text", + "content": "images. These datasets contain RGB images from Sentinel2 [13]. The long-time difference between the two sets " + }, + { + "bbox": [ + 46, + 285, + 287, + 350 + ], + "type": "inline_equation", + "content": "\\mathcal{I}_{l_i}^{t_1}" + }, + { + "bbox": [ + 46, + 285, + 287, + 350 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 285, + 287, + 350 + ], + "type": "inline_equation", + "content": "\\mathcal{I}_{l_i}^{t_2}" + }, + { + "bbox": [ + 46, + 285, + 287, + 350 + ], + "type": "text", + "content": " is 4 years. The maximum time difference within a set " + }, + { + "bbox": [ + 46, + 285, + 287, + 350 + ], + "type": "inline_equation", + "content": "\\mathcal{I}_{l_i}^{t_j}" + }, + { + "bbox": [ + 46, + 285, + 287, + 350 + ], + "type": "text", + "content": " is 1 year. We train our model and baselines for 1000 and 200 epochs on the 100k and 1m dataset respectively." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 358, + 183, + 370 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 358, + 183, + 370 + ], + "spans": [ + { + "bbox": [ + 47, + 358, + 183, + 370 + ], + "type": "text", + "content": "4.2. Landcover classification" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 376, + 287, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 376, + 287, + 449 + ], + "spans": [ + { + "bbox": [ + 46, + 376, + 287, + 449 + ], + "type": "text", + "content": "Nine (out of seventeen) of the UN's Sustainable Development Goals require global monitoring of land cover [5]. Thus, improving the performance of models for landcover classification would result in better monitoring of these goals. We evaluate our method on two datasets for landcover classification: EuroSat [19] and BigEarthNet [34]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 449, + 287, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 449, + 287, + 510 + ], + "spans": [ + { + "bbox": [ + 46, + 449, + 287, + 510 + ], + "type": "text", + "content": "EuroSat has 10-classes with a total of " + }, + { + "bbox": [ + 46, + 449, + 287, + 510 + ], + "type": "inline_equation", + "content": "27\\mathrm{k}" + }, + { + "bbox": [ + 46, + 449, + 287, + 510 + ], + "type": "text", + "content": " images " + }, + { + "bbox": [ + 46, + 449, + 287, + 510 + ], + "type": "inline_equation", + "content": "64\\times 64" + }, + { + "bbox": [ + 46, + 449, + 287, + 510 + ], + "type": "text", + "content": " images from Sentinel-2. We use the same train/val split proposed by the dataset. We add a linear layer to the frozen pretrained backbone, to perform the linear evaluation. More details about the training are in the supplementary." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 510, + 287, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 510, + 287, + 605 + ], + "spans": [ + { + "bbox": [ + 46, + 510, + 287, + 605 + ], + "type": "text", + "content": "Tab. 1 shows the top-1 classification accuracy of various pre-trained backbones on the EuroSat dataset. We first note that all satellite image-specific methods outperform generic pretraining on ImageNet, suggesting the importance of training a representation specifically for satellite imagery. They also outperform the MoCo v2 baseline trained on satellite images, indicating the importance of well-designed loss functions that use the structure of satellite imagery." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 605, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 288, + 714 + ], + "type": "text", + "content": "Compared to SeCo, our approach (CACo) results in 3 points of improvement with the linear classifier (with ResNet-18 and 100k data). This validates our insight that invariance to seasonal changes is not enough; the features must also be trained to be sensitive to long-term change. In fact, CACo trained on 100k images produces a representation competitive with SeCo trained on 1m images. Thus our method can save an order of magnitude of images and training time over the SeCo baseline while producing rep" + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 315, + 71, + 539, + 232 + ], + "blocks": [ + { + "bbox": [ + 46, + 229, + 288, + 264 + ], + "lines": [ + { + "bbox": [ + 46, + 229, + 288, + 264 + ], + "spans": [ + { + "bbox": [ + 46, + 229, + 288, + 264 + ], + "type": "text", + "content": "Table 1. Performance of our representation on the EuroSat classification task with linear probing, in top-1 Accuracy. Our method provides a more accurate classification, with different backbones." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 315, + 71, + 539, + 232 + ], + "lines": [ + { + "bbox": [ + 315, + 71, + 539, + 232 + ], + "spans": [ + { + "bbox": [ + 315, + 71, + 539, + 232 + ], + "type": "table", + "html": "
DataPre-trainingResNet-18ResNet-50
10%100%10%100%
-Random init.42.8745.9544.7645.22
ImageNet.65.4366.4070.3671.37
100kMoCo v265.4367.2070.3872.88
GSSL65.7867.3670.6572.86
SeCo65.8067.4370.6973.42
CACo (ours)67.8969.4371.5573.63
1mSeCo67.6869.9572.8974.82
SatMAE--67.9369.45
DDPM-CD--65.4767.31
CACo (ours)68.6470.4173.4074.98
", + "image_path": "4822b7e3546351725b1b6f63d4cf9f29559b4ba3884077147217967eefacd12c.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 240, + 547, + 285 + ], + "lines": [ + { + "bbox": [ + 304, + 240, + 547, + 285 + ], + "spans": [ + { + "bbox": [ + 304, + 240, + 547, + 285 + ], + "type": "text", + "content": "Table 2. Performance of our method on BigEarthNet landcover classification, in mean Average Precision (mAP). The two columns use different percentages of data (10% and 100%) of BigEarthNet data for training the linear layer." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 305, + 308, + 430, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 308, + 430, + 319 + ], + "spans": [ + { + "bbox": [ + 305, + 308, + 430, + 319 + ], + "type": "text", + "content": "resentations of similar quality." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 320, + 545, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 320, + 545, + 403 + ], + "spans": [ + { + "bbox": [ + 304, + 320, + 545, + 403 + ], + "type": "text", + "content": "We also evaluate our method on the BigEarthNet dataset, which is a multi-label classification dataset with 17 classes. The dataset is significantly larger than EuroSat, containing 590k sentinel-2 image patches. We use the RGB bands to perform classification on this dataset. Since this is a multi-label classification problem, we use mean Average Precision (mAP) to evaluate the performance." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 404, + 545, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 404, + 545, + 441 + ], + "spans": [ + { + "bbox": [ + 304, + 404, + 545, + 441 + ], + "type": "text", + "content": "Tab. 2 compares the performance of our model to the baselines on BigEarthNet1. Even on BigEarthNet, our method performs better than the baselines." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 449, + 410, + 463 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 449, + 410, + 463 + ], + "spans": [ + { + "bbox": [ + 305, + 449, + 410, + 463 + ], + "type": "text", + "content": "4.3. Change detection" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 469, + 545, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 469, + 545, + 517 + ], + "spans": [ + { + "bbox": [ + 304, + 469, + 545, + 517 + ], + "type": "text", + "content": "We also evaluate our model on the OSCD change detection dataset [6]. It contains 24 pairs of images from the Sentinel-2 satellites between 2015 and 2018. The dataset is split into 14 training images and 10 testing images." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 517, + 545, + 637 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 517, + 545, + 637 + ], + "spans": [ + { + "bbox": [ + 304, + 517, + 545, + 637 + ], + "type": "text", + "content": "Our model architecture follows past work [11, 25]. The input to the model is an image pair, each of which is fed into the frozen, pre-trained feature extractor (we use the ResNet-18 models). We extract feature maps after each downsampling layer for each image. We then take the absolute difference between the corresponding feature maps from the two images. These differences are passed as input to a U-Net Decoder [30]. The U-Net decoder is then trained with change supervision. More details about the training and architecture can be found in the supplementary." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 638, + 545, + 673 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 638, + 545, + 673 + ], + "spans": [ + { + "bbox": [ + 304, + 638, + 545, + 673 + ], + "type": "text", + "content": "Results are shown in Tab. 3. We report the F1-score obtained by using a threshold of 0.5. Our method results in better features for change detection resulting in about 4" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 306, + 683, + 545, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 683, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 683, + 545, + 713 + ], + "type": "text", + "content": "1We obtained the SeCo number by training a model using the authors' publicly released code and dataset, but this does not match the published number [25]. Repeated attempts to contact the authors were unanswered." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5266" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 57, + 71, + 279, + 232 + ], + "blocks": [ + { + "bbox": [ + 57, + 71, + 279, + 232 + ], + "lines": [ + { + "bbox": [ + 57, + 71, + 279, + 232 + ], + "spans": [ + { + "bbox": [ + 57, + 71, + 279, + 232 + ], + "type": "table", + "html": "
BackbonePre-trainingPrec.Rec.F1-score
ResNet-18Random init.69.7318.2428.91
ImageNet.70.2223.5835.30
MoCo v262.2127.5738.21
GSSL62.2934.0844.06
SeCo64.1536.8946.84
CACo (ours)60.6842.9450.29
ResNet-50Random init.65.4622.4333.41
ImageNet.70.0530.9642.94
GSSL62.1636.8346.26
SeCo63.2138.2647.67
CACo (ours)62.8744.4952.11
", + "image_path": "cf58b9c9c7dfab0d653e18f356bed71a39842dee8392424dbce7a86881cd0da3.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 95, + 290, + 241, + 387 + ], + "blocks": [ + { + "bbox": [ + 46, + 243, + 288, + 277 + ], + "lines": [ + { + "bbox": [ + 46, + 243, + 288, + 277 + ], + "spans": [ + { + "bbox": [ + 46, + 243, + 288, + 277 + ], + "type": "text", + "content": "Table 3. Performance of our method on the Change Detection Task. We report the precision, recall and F1-score at a threshold of 0.5 for various ResNet backbones, trained on the 100k dataset." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 95, + 290, + 241, + 387 + ], + "lines": [ + { + "bbox": [ + 95, + 290, + 241, + 387 + ], + "spans": [ + { + "bbox": [ + 95, + 290, + 241, + 387 + ], + "type": "table", + "html": "
Pre-traininglinearfinetuning
Random init.41.5338.62
ImageNet.43.7543.78
MoCo v247.9747.15
GSSL46.7748.10
SeCo46.8348.18
CACo (ours)50.2051.29
", + "image_path": "82522e8a8d1e396aa7e05e63c985f0ed64ea69f920ce71a6c559352270f2fa4c.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 454, + 287, + 504 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 454, + 287, + 504 + ], + "spans": [ + { + "bbox": [ + 46, + 454, + 287, + 504 + ], + "type": "text", + "content": "points of improvement in the F1 scores for both ResNet-18 and ResNet-50. This large improvement on change detection is likely because we explicitly train our representation to be sensitive to long-term change." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 513, + 178, + 525 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 513, + 178, + 525 + ], + "spans": [ + { + "bbox": [ + 47, + 513, + 178, + 525 + ], + "type": "text", + "content": "4.4. Semantic segmentation" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 532, + 287, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 532, + 287, + 604 + ], + "spans": [ + { + "bbox": [ + 46, + 532, + 287, + 604 + ], + "type": "text", + "content": "We also evaluate our method on a satellite images segmentation task. DynamicEarthNet [36] is a dataset used to evaluate landcover segmentation over 7 classes. The dataset contains images from Planet satellites for 65 locations. Each location has " + }, + { + "bbox": [ + 46, + 532, + 287, + 604 + ], + "type": "inline_equation", + "content": "241024\\times 1024" + }, + { + "bbox": [ + 46, + 532, + 287, + 604 + ], + "type": "text", + "content": " images with labels. The locations are split into 55 for training and 10 for testing." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 605, + 287, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 287, + 665 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 287, + 665 + ], + "type": "text", + "content": "Similar to change detection model, we use a U-net with ResNet-18 as the encoder. But unlike them, we input a single image and use feature maps instead of feature differences in the decoder. Due to the dataset imbalance, we use Dice Loss [33] for training (see supplementary for details)." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "text", + "content": "Tab. 4 shows the mIoU (intersection over union averaged over classes) for our method on DynamicEarthNet validation images. Our method is better than SeCo by more than 3 points, and ImageNet pre-training by more than 6 points." + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 310, + 70, + 542, + 159 + ], + "blocks": [ + { + "bbox": [ + 46, + 397, + 287, + 431 + ], + "lines": [ + { + "bbox": [ + 46, + 397, + 287, + 431 + ], + "spans": [ + { + "bbox": [ + 46, + 397, + 287, + 431 + ], + "type": "text", + "content": "Table 4. Performance of our method on the DynamicEarthNet segmentation task. We report the mIoU score for the ResNet-18 backbone, trained on the 100k dataset." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 310, + 70, + 542, + 159 + ], + "lines": [ + { + "bbox": [ + 310, + 70, + 542, + 159 + ], + "spans": [ + { + "bbox": [ + 310, + 70, + 542, + 159 + ], + "type": "table", + "html": "
Pre-trainingCaiRoadCalFire
AP@50AP@400AP@10AP@40
ImageNet41.7932.2652.6144.99
MoCo v241.7833.0254.8348.20
SeCo39.6334.7262.8751.87
CACo (ours)44.3835.9965.7153.44
", + "image_path": "16bce06da81c71473b4cee01c28f05c2ba5f0b2b70f23ac60e798deaf3abe496.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 167, + 545, + 201 + ], + "lines": [ + { + "bbox": [ + 305, + 167, + 545, + 201 + ], + "spans": [ + { + "bbox": [ + 305, + 167, + 545, + 201 + ], + "type": "text", + "content": "Table 5. Performance of our method on change event retrieval for the CaiRoad and CalFire benchmarks. We show average precision@K for ResNet-18 backbones trained on the 100k dataset." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 308, + 212, + 425, + 327 + ], + "blocks": [ + { + "bbox": [ + 308, + 212, + 425, + 327 + ], + "lines": [ + { + "bbox": [ + 308, + 212, + 425, + 327 + ], + "spans": [ + { + "bbox": [ + 308, + 212, + 425, + 327 + ], + "type": "image", + "image_path": "b6f88e96e848e5a967ce929aac2eed91c96404b013669a4247bf41fdb9f72221.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 328, + 545, + 361 + ], + "lines": [ + { + "bbox": [ + 305, + 328, + 545, + 361 + ], + "spans": [ + { + "bbox": [ + 305, + 328, + 545, + 361 + ], + "type": "text", + "content": "Figure 4. Precision@K of our backbone on CaiRoad (Left) and CalFire (Right) retrieval benchmark. Our method has better Precision@K than baselines for most K values on both benchmarks." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 426, + 213, + 544, + 327 + ], + "blocks": [ + { + "bbox": [ + 426, + 213, + 544, + 327 + ], + "lines": [ + { + "bbox": [ + 426, + 213, + 544, + 327 + ], + "spans": [ + { + "bbox": [ + 426, + 213, + 544, + 327 + ], + "type": "image", + "image_path": "d39a771990eeefec49e023640f899f36b0b5d13ac493286762531d3a4e712431.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 381, + 468, + 395 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 381, + 468, + 395 + ], + "spans": [ + { + "bbox": [ + 306, + 381, + 468, + 395 + ], + "type": "text", + "content": "4.5. Evaluation on Change Events" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 400, + 545, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 400, + 545, + 483 + ], + "spans": [ + { + "bbox": [ + 304, + 400, + 545, + 483 + ], + "type": "text", + "content": "We evaluate our method on a recently released change events benchmark [24]. Change events are spatio-temporal semantic structures that are caused by a real-world event. The goal is to classify these change events. The benchmark contains two datasets: one from Cairo where we find road construction events (called CaiRoad), and the other from California where we find forest fire events (CalFire)." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 484, + 545, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 484, + 545, + 568 + ], + "spans": [ + { + "bbox": [ + 304, + 484, + 545, + 568 + ], + "type": "text", + "content": "We use the model proposed by Mall et al. [24], taking in a sequence of images and binary change masks for consecutive pairs. The model encodes image pairs using ResNet-18, average pooling the feature map with the downscaled change mask as weights. The final feature vector is obtained by temporal averaging. The representation is learned using SimCLR on change events (see supplementary)." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 568, + 545, + 615 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 568, + 545, + 615 + ], + "spans": [ + { + "bbox": [ + 304, + 568, + 545, + 615 + ], + "type": "text", + "content": "We follow the same procedure, but the ResNet-18 backbone is initialized with the trained weights, and then finetuned per [24]. We evaluate by measuring average precision@K (AP@K) for different backbone architectures." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 616, + 545, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 616, + 545, + 664 + ], + "spans": [ + { + "bbox": [ + 304, + 616, + 545, + 664 + ], + "type": "text", + "content": "Tab. 5 shows the AP@K for different backbones, on the CaiRoad and CalFire benchmarks. Our method results in better retrieval for both types of events. Fig. 4 shows the precision@K for various methods on these benchmarks." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 306, + 671, + 369, + 682 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 671, + 369, + 682 + ], + "spans": [ + { + "bbox": [ + 306, + 671, + 369, + 682 + ], + "type": "text", + "content": "4.6. Ablation" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "text", + "content": "We evaluate the design choices made by our method in Sec. 3. We look at improvements made by different com" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5267" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 50, + 70, + 285, + 148 + ], + "blocks": [ + { + "bbox": [ + 50, + 70, + 285, + 148 + ], + "lines": [ + { + "bbox": [ + 50, + 70, + 285, + 148 + ], + "spans": [ + { + "bbox": [ + 50, + 70, + 285, + 148 + ], + "type": "table", + "html": "
Pre-trainingEuroSat (Acc.)DynamicEarth-Net (mIoU)
SeCo90.0546.83
+ Improved Sampling91.4247.24
+ Long-term temporal Contrast92.5649.68
+ Change-awareness (Ours)93.0850.20
", + "image_path": "5bee307083074c6feef0ceb38ce97374bf22718bd396aaeb722b7fb709ec6de7.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 78, + 216, + 256, + 262 + ], + "blocks": [ + { + "bbox": [ + 46, + 159, + 287, + 204 + ], + "lines": [ + { + "bbox": [ + 46, + 159, + 287, + 204 + ], + "spans": [ + { + "bbox": [ + 46, + 159, + 287, + 204 + ], + "type": "text", + "content": "Table 6. Performance with different sub-components of CACo on EuroSat classification, with ResNet-18 pre-trained on the 100k dataset. All three of our novel contributions lead to an improvement in performance on downstream tasks." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 78, + 216, + 256, + 262 + ], + "lines": [ + { + "bbox": [ + 78, + 216, + 256, + 262 + ], + "spans": [ + { + "bbox": [ + 78, + 216, + 256, + 262 + ], + "type": "table", + "html": "
Pre-trainingEuroSat (Acc.)OSCD (F1)
SeCo87.9744.05
CACo90.8347.50
", + "image_path": "17abcf2b9a645010ccc6c0cc8b1ce28c1e7d275bd87886ca2e99d68d923d8bc5.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 328, + 287, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 328, + 287, + 365 + ], + "spans": [ + { + "bbox": [ + 46, + 328, + 287, + 365 + ], + "type": "text", + "content": "ponents such as better data, long-term temporal information, and change estimates. We also show that our method generalizes to other self-supervised frameworks." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 380, + 287, + 477 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 380, + 287, + 477 + ], + "spans": [ + { + "bbox": [ + 46, + 380, + 287, + 477 + ], + "type": "text", + "content": "Are the new components of our method essential? In Tab. 6 we evaluate the impact of each component of our approach on the final performance as measured on EuroSat and DynamicEarthNet. We find that each component adds a significant improvement. It is telling that incorporating long-term temporal contrast has a particularly large impact on DynamicEarthNet, indicating that long-term contrast is especially needed for pixel-level localization tasks." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 492, + 287, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 492, + 287, + 601 + ], + "spans": [ + { + "bbox": [ + 46, + 492, + 287, + 601 + ], + "type": "text", + "content": "Can our method work with other self-supervised frameworks? We replace the MoCo v2 framework with SimCLR [7]. SimCLR requires a large batch size to find more hard negative examples; we use a batch size of 512 instead of 256 for all the methods. Tab. 7 shows the performance of SeCo and CACo using SimCLR. Our model performs better than SeCo by 3 points when using SimCLR as well. This indicates that our insights are general and can be applied in the future to new self-supervision frameworks." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 616, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 616, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 616, + 287, + 715 + ], + "type": "text", + "content": "How useful is the ratio estimate? In Sec. 3.4 we proposed to estimate using " + }, + { + "bbox": [ + 46, + 616, + 287, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{r}_{l_i}^k" + }, + { + "bbox": [ + 46, + 616, + 287, + 715 + ], + "type": "text", + "content": " to estimate changes. We now look at other estimates for change in Tab. 8. Directly using the feature distance between the long-term pair, instead of the ratio leads to an almost 2-point drop in performance. Removing the exponential moving average (i.e., using " + }, + { + "bbox": [ + 46, + 616, + 287, + 715 + ], + "type": "inline_equation", + "content": "R_{l_i}^k" + }, + { + "bbox": [ + 46, + 616, + 287, + 715 + ], + "type": "text", + "content": " instead of " + }, + { + "bbox": [ + 46, + 616, + 287, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{r}_{l_i}^k" + }, + { + "bbox": [ + 46, + 616, + 287, + 715 + ], + "type": "text", + "content": "; see Eq. (3)) also reduces performance by about 1 point. Another alternative is to measure the ratio" + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 353, + 70, + 501, + 139 + ], + "blocks": [ + { + "bbox": [ + 46, + 273, + 287, + 307 + ], + "lines": [ + { + "bbox": [ + 46, + 273, + 287, + 307 + ], + "spans": [ + { + "bbox": [ + 46, + 273, + 287, + 307 + ], + "type": "text", + "content": "Table 7. Accuracy on EuroSat when using a ResNet-18 backbone pre-trained using SimCLR instead of MoCo v2. Our insights generalize to other self-supervised frameworks like SimCLR as well." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 353, + 70, + 501, + 139 + ], + "lines": [ + { + "bbox": [ + 353, + 70, + 501, + 139 + ], + "spans": [ + { + "bbox": [ + 353, + 70, + 501, + 139 + ], + "type": "table", + "html": "
Change EstimateEuroSat (Acc.)
long-term distance91.17
Rliki91.98
Align92.92
rkiLi (ours)93.08
", + "image_path": "5c931ec8dfbf712cae6771414891b044c9a3f37a89aba59a61e906bd3b6809aa.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 149, + 545, + 196 + ], + "lines": [ + { + "bbox": [ + 305, + 149, + 545, + 196 + ], + "spans": [ + { + "bbox": [ + 305, + 149, + 545, + 196 + ], + "type": "text", + "content": "Table 8. Performance of different change estimates with ResNet-18 on 100k dataset. " + }, + { + "bbox": [ + 305, + 149, + 545, + 196 + ], + "type": "inline_equation", + "content": "R_{l_i}^k" + }, + { + "bbox": [ + 305, + 149, + 545, + 196 + ], + "type": "text", + "content": " (Eq. (3)) is the estimate without the moving average. Long-term distance is the numerator of the " + }, + { + "bbox": [ + 305, + 149, + 545, + 196 + ], + "type": "inline_equation", + "content": "R_{l_i}^k" + }, + { + "bbox": [ + 305, + 149, + 545, + 196 + ], + "type": "text", + "content": ". Align uses seasonally aligned long-term pairs without smoothing." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 305, + 216, + 545, + 265 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 216, + 545, + 265 + ], + "spans": [ + { + "bbox": [ + 305, + 216, + 545, + 265 + ], + "type": "text", + "content": "with long-term pairs that are seasonally aligned. Align gives similar results to " + }, + { + "bbox": [ + 305, + 216, + 545, + 265 + ], + "type": "inline_equation", + "content": "\\mathbf{r}_{l_i}^k" + }, + { + "bbox": [ + 305, + 216, + 545, + 265 + ], + "type": "text", + "content": ", but requires a more sophisticated sampling. In sum, these results suggest that how one identifies changed regions can have a big impact on performance." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 274, + 463, + 288 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 274, + 463, + 288 + ], + "spans": [ + { + "bbox": [ + 306, + 274, + 463, + 288 + ], + "type": "text", + "content": "5. Discussion and Conclusions" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 295, + 545, + 439 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 295, + 545, + 439 + ], + "spans": [ + { + "bbox": [ + 305, + 295, + 545, + 439 + ], + "type": "text", + "content": "Potential Negative Societal Impact and Limitations. As in all visual recognition, there is the possibility of negative impact through violations of privacy. To mitigate this concern we intentionally use low-resolution satellite images (1 pixel " + }, + { + "bbox": [ + 305, + 295, + 545, + 439 + ], + "type": "inline_equation", + "content": "\\sim 10\\mathrm{m}" + }, + { + "bbox": [ + 305, + 295, + 545, + 439 + ], + "type": "text", + "content": "). The use of our proposed techniques for surveillance should be appropriately regulated. Our work currently is limited in its ability to generalize to multiple resolutions. Though object detection tasks use higher-resolution images for better object delineation, our representation trained on medium-resolution images cannot directly generalize to them; although we can retrain our method on high-resolution images for these tasks." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 453, + 545, + 573 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 453, + 545, + 573 + ], + "spans": [ + { + "bbox": [ + 304, + 453, + 545, + 573 + ], + "type": "text", + "content": "Conclusions. We present a novel self-supervised framework to learn representations well suited for remote sensing applications. We introduce a new loss that robustly leverages the long-term temporal information readily available for satellite images. We also propose a better location sampling method to provide more informative data. Our evaluation on diverse downstream tasks shows that our approach (CACo) is very versatile and leads to better downstream accuracy than prior art (with relative improvements of " + }, + { + "bbox": [ + 304, + 453, + 545, + 573 + ], + "type": "inline_equation", + "content": "8.5\\%" + }, + { + "bbox": [ + 304, + 453, + 545, + 573 + ], + "type": "text", + "content": " on change detection and " + }, + { + "bbox": [ + 304, + 453, + 545, + 573 + ], + "type": "inline_equation", + "content": "6.5\\%" + }, + { + "bbox": [ + 304, + 453, + 545, + 573 + ], + "type": "text", + "content": " on semantic segmentation)." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 586, + 545, + 707 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 586, + 545, + 707 + ], + "spans": [ + { + "bbox": [ + 304, + 586, + 545, + 707 + ], + "type": "text", + "content": "Acknowledgements. This research is based upon work supported in part by the ODNI (IARPA) via 2021-20111000006, NSF 1900783, NSF 2144117, and NSF 2212084. The views and conclusions contained herein are those of the authors and should not be interpreted as necessarily representing the official policies, either expressed or implied, of ODNI, IARPA, or the US Government. The US Government is authorized to reproduce and distribute reprints for governmental purposes notwithstanding any copyright annotation therein." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5268" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 287, + 715 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 123 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 123 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 123 + ], + "type": "text", + "content": "[1] Peri Akiva, Matthew Purri, and Matthew Leotta. Self-supervised material and texture representation learning for remote sensing tasks. In CVPR, 2022. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 124, + 287, + 167 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 124, + 287, + 167 + ], + "spans": [ + { + "bbox": [ + 53, + 124, + 287, + 167 + ], + "type": "text", + "content": "[2] Kumar Ayush, Burak Uzkent, Chenlin Meng, Kumar Tanmay, Marshall Burke, David Lobell, and Stefano Ermon. Geography-aware self-supervised learning. In CVPR, 2021. 2, 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 168, + 287, + 201 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 168, + 287, + 201 + ], + "spans": [ + { + "bbox": [ + 53, + 168, + 287, + 201 + ], + "type": "text", + "content": "[3] Gaetan Bahl, Mehdi Bahri, and Florent Lafarge. Single-shot end-to-end road graph extraction. In CVPR EarthVision, 2022. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 202, + 287, + 246 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 202, + 287, + 246 + ], + "spans": [ + { + "bbox": [ + 53, + 202, + 287, + 246 + ], + "type": "text", + "content": "[4] Wele Gedara Chaminda Bandara, Nithin Gopalakrishnan Nair, and Vishal M Patel. Ddpm-cd: Remote sensing change detection using denoising diffusion probabilistic models. CoRR, 2022. 2, 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 247, + 287, + 279 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 247, + 287, + 279 + ], + "spans": [ + { + "bbox": [ + 53, + 247, + 287, + 279 + ], + "type": "text", + "content": "[5] Sarah Carter and Martin Herold. Specifications of land cover datasets for SDG indicator monitoring. https://ggim.un.org/ Accessed: 11-8-2022. 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 280, + 287, + 313 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 280, + 287, + 313 + ], + "spans": [ + { + "bbox": [ + 53, + 280, + 287, + 313 + ], + "type": "text", + "content": "[6] R. Caye Daudt, B. Le Saux, A. Boulch, and Y. Gousseau. Urban change detection for multispectral earth observation using convolutional neural networks. In IGARSS, 2018. 1, 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 314, + 287, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 314, + 287, + 346 + ], + "spans": [ + { + "bbox": [ + 53, + 314, + 287, + 346 + ], + "type": "text", + "content": "[7] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In ICML, 2020. 2, 8" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 346, + 287, + 379 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 346, + 287, + 379 + ], + "spans": [ + { + "bbox": [ + 53, + 346, + 287, + 379 + ], + "type": "text", + "content": "[8] Xinlei Chen, Haoqi Fan, Ross Girshick, and Kaiming He. Improved baselines with momentum contrastive learning. CoRR, 2020. 2, 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 380, + 287, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 380, + 287, + 403 + ], + "spans": [ + { + "bbox": [ + 53, + 380, + 287, + 403 + ], + "type": "text", + "content": "[9] Gordon Christie, Neil Fendley, James Wilson, and Ryan Mukherjee. Functional map of the world. In CVPR, 2018. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 403, + 287, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 403, + 287, + 456 + ], + "spans": [ + { + "bbox": [ + 48, + 403, + 287, + 456 + ], + "type": "text", + "content": "[10] Yezhen Cong, Samar Khanna, Chenlin Meng, Patrick Liu, Erik Rozi, Yutong He, Marshall Burke, David B Lobell, and Stefano Ermon. Satmae: Pre-training transformers for temporal and multi-spectral satellite imagery. In NeurIPS, 2022. 2, 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 458, + 287, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 458, + 287, + 491 + ], + "spans": [ + { + "bbox": [ + 48, + 458, + 287, + 491 + ], + "type": "text", + "content": "[11] Rodrigo Caye Daudt, Bertr Le Saux, and Alexandre Boulch. Fully convolutional siamese networks for change detection. In ICIP, 2018. 2, 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 491, + 287, + 546 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 491, + 287, + 546 + ], + "spans": [ + { + "bbox": [ + 48, + 491, + 287, + 546 + ], + "type": "text", + "content": "[12] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. CoRR, 2020. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 547, + 287, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 547, + 287, + 612 + ], + "spans": [ + { + "bbox": [ + 48, + 547, + 287, + 612 + ], + "type": "text", + "content": "[13] Matthias Drusch, Umberto Del Bello, Sébastien Carlier, Olivier Colin, Veronica Fernandez, Ferran Gascon, Bianca Hoersch, Claudia Isola, Paolo Laberinti, Philippe Martimort, et al. Sentinel-2: ESA's optical high-resolution mission for GMES operational services. Remote Sensing of Environment, 2012. 2, 6" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 613, + 287, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 613, + 287, + 658 + ], + "spans": [ + { + "bbox": [ + 48, + 613, + 287, + 658 + ], + "type": "text", + "content": "[14] Jan Gąsienica-Józkowy, Mateusz Knapik, and Bogusław Cyganek. An ensemble deep learning method with optimized weights for drone-based water rescue and surveillance. *Integrated Computer-Aided Engineering*, 2021. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 658, + 287, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 287, + 690 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 287, + 690 + ], + "type": "text", + "content": "[15] Spyros Gidaris, Praveer Singh, and Nikos Komodakis. Unsupervised representation learning by predicting image rotations. In ICLR, 2018. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 48, + 691, + 287, + 715 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 287, + 715 + ], + "type": "text", + "content": "[16] Jean-Bastien Grill, Florian Strub, Florent Altché, Corentin Tallec, Pierre Richemond, Elena Buchatskaya, Carl Doersch," + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 715 + ], + "type": "list", + "angle": 0, + "index": 35, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 106 + ], + "type": "text", + "content": "Bernardo Avila Pires, Zhaohan Guo, Mohammad Gheshlaghi Azar, et al. Bootstrap your own latent-a new approach to self-supervised learning. NeurIPS, 2020. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 107, + 545, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 107, + 545, + 140 + ], + "spans": [ + { + "bbox": [ + 307, + 107, + 545, + 140 + ], + "type": "text", + "content": "[17] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollár, and Ross Girshick. Masked autoencoders are scalable vision learners. In CVPR, 2022. 2, 5" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 141, + 545, + 174 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 141, + 545, + 174 + ], + "spans": [ + { + "bbox": [ + 307, + 141, + 545, + 174 + ], + "type": "text", + "content": "[18] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In CVPR, 2020. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 175, + 545, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 175, + 545, + 217 + ], + "spans": [ + { + "bbox": [ + 307, + 175, + 545, + 217 + ], + "type": "text", + "content": "[19] Patrick Helber, Benjamin Bischke, Andreas Dengel, and Damian Borth. Eurosat: A novel dataset and deep learning benchmark for land use and land cover classification. JS-TAEORS, 2019. 2, 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 219, + 545, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 219, + 545, + 262 + ], + "spans": [ + { + "bbox": [ + 307, + 219, + 545, + 262 + ], + "type": "text", + "content": "[20] Neal Jean, Sherrie Wang, Anshul Samar, George Azzari, David Lobell, and Stefano Ermon. Tile2vec: Unsupervised representation learning for spatially distributed data. In AAAI, 2019. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 264, + 545, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 264, + 545, + 297 + ], + "spans": [ + { + "bbox": [ + 307, + 264, + 545, + 297 + ], + "type": "text", + "content": "[21] Marrit Leenstra, Diego Marcos, Francesca Bovolo, and Devis Tuia. Self-supervised pre-training enhances change detection in sentinel-2 imagery. In ICPR, 2021. 4" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 297, + 545, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 297, + 545, + 342 + ], + "spans": [ + { + "bbox": [ + 307, + 297, + 545, + 342 + ], + "type": "text", + "content": "[22] Yahui Liu, Jian Yao, Xiaohu Lu, Menghan Xia, Xingbo Wang, and Yuan Liu. RoadNet: Learning to comprehensively analyze road networks in complex urban scenes from high-resolution remotely sensed images. TGRS, 2018. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 343, + 545, + 375 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 343, + 545, + 375 + ], + "spans": [ + { + "bbox": [ + 307, + 343, + 545, + 375 + ], + "type": "text", + "content": "[23] William A Malila. Change vector analysis: an approach for detecting forest changes with landsat. In LARS symposia, 1980. 4" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 376, + 545, + 410 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 376, + 545, + 410 + ], + "spans": [ + { + "bbox": [ + 307, + 376, + 545, + 410 + ], + "type": "text", + "content": "[24] Utkarsh Mall, Bharath Hariharan, and Kavita Bala. Change event dataset for discovery from spatio-temporal remote sensing imagery. In NeurIPS, 2022. 2, 7" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 411, + 545, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 411, + 545, + 453 + ], + "spans": [ + { + "bbox": [ + 307, + 411, + 545, + 453 + ], + "type": "text", + "content": "[25] Oscar Manas, Alexandre Lacoste, Xavier Giroi-nieto, David Vazquez, and Pau Rodriguez. Seasonal contrast: Unsupervised pre-training from uncurated remote sensing data. In ICCV, 2021. 1, 2, 3, 5, 6" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 455, + 545, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 455, + 545, + 487 + ], + "spans": [ + { + "bbox": [ + 307, + 455, + 545, + 487 + ], + "type": "text", + "content": "[26] Ishan Misra and Laurens van der Maaten. Self-supervised learning of pretext-invariant representations. In CVPR, 2020. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 488, + 545, + 532 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 488, + 545, + 532 + ], + "spans": [ + { + "bbox": [ + 307, + 488, + 545, + 532 + ], + "type": "text", + "content": "[27] Sorour Mohajerani and Parvaneh Saeedi. Cloud and cloud shadow segmentation for remote sensing imagery via filtered jaccard loss function and parametric augmentation. J-STARS, 2021. 2" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 307, + 533, + 545, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 533, + 545, + 566 + ], + "spans": [ + { + "bbox": [ + 307, + 533, + 545, + 566 + ], + "type": "text", + "content": "[28] Mehdi Noroozi and Paolo Favaro. Unsupervised learning of visual representations by solving jigsaw puzzles. In ECCV, 2016. 2" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 307, + 567, + 545, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 567, + 545, + 600 + ], + "spans": [ + { + "bbox": [ + 307, + 567, + 545, + 600 + ], + "type": "text", + "content": "[29] Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding. CoRR, 2018. 3" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 307, + 601, + 545, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 601, + 545, + 635 + ], + "spans": [ + { + "bbox": [ + 307, + 601, + 545, + 635 + ], + "type": "text", + "content": "[30] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In ICMICCAI, pages 234-241. Springer, 2015. 6" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 307, + 635, + 545, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 635, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 307, + 635, + 545, + 689 + ], + "type": "text", + "content": "[31] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael S. Bernstein, Alexander C. Berg, and Li Fei-Fei. Imagenet large scale visual recognition challenge. IJCV, 2015. 5" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 307, + 690, + 545, + 715 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 690, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 307, + 690, + 545, + 715 + ], + "type": "text", + "content": "[32] Linus Scheibenreif, Joëlle Hanna, Michael Mommert, and Damian Borth. Self-supervised vision transformers for land-" + } + ] + } + ], + "index": 34 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5269" + } + ] + } + ], + "index": 36 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 611 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 66, + 72, + 286, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 72, + 286, + 94 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 286, + 94 + ], + "type": "text", + "content": "cover segmentation and classification. In CVPR EarthVision, 2022. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 96, + 287, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 96, + 287, + 139 + ], + "spans": [ + { + "bbox": [ + 48, + 96, + 287, + 139 + ], + "type": "text", + "content": "[33] Carole H Sudre, Wenqi Li, Tom Vercauteren, Sebastien Ourselin, and M Jorge Cardoso. Generalised dice overlap as a deep learning loss function for highly unbalanced segmentations. In DLMIA, 2017. 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 141, + 287, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 141, + 287, + 184 + ], + "spans": [ + { + "bbox": [ + 48, + 141, + 287, + 184 + ], + "type": "text", + "content": "[34] Gencer Sumbul, Marcela Charfuelan, Begüm Demir, and Volker Markl. Bigearthnet: A large-scale benchmark archive for remote sensing image understanding. In IGARSS, 2019. 2, 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 186, + 287, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 186, + 287, + 239 + ], + "spans": [ + { + "bbox": [ + 48, + 186, + 287, + 239 + ], + "type": "text", + "content": "[35] Xu Tang, Huayu Zhang, Lichao Mou, Fang Liu, Xiangrong Zhang, Xiao Xiang Zhu, and Licheng Jiao. An unsupervised remote sensing change detection method based on multiscale graph convolutional network and metric learning. TGRSS, 2021. 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 242, + 287, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 242, + 287, + 296 + ], + "spans": [ + { + "bbox": [ + 48, + 242, + 287, + 296 + ], + "type": "text", + "content": "[36] Aysim Token, Lukas Kondmann, Mark Weber, Marvin Eisenberger, Andrés Camero, Jingliang Hu, Ariadna Pregel Hoderlein, Caglar Senaras, Timothy Davis, Daniel Cremers, et al. Dynamicearthnet: Daily multi-spectral satellite dataset for semantic change segmentation. In CVPR, 2022. 1, 2, 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 297, + 287, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 297, + 287, + 330 + ], + "spans": [ + { + "bbox": [ + 48, + 297, + 287, + 330 + ], + "type": "text", + "content": "[37] Gabriel Tseng, Ivan Zvonkov, Catherine Lilian Nakalembe, and Hannah Kerner. Cropharvest: A global dataset for croptype classification. In NeurIPS, 2021. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 332, + 286, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 332, + 286, + 363 + ], + "spans": [ + { + "bbox": [ + 48, + 332, + 286, + 363 + ], + "type": "text", + "content": "[38] Adam Van Etten, Dave Lindenbaum, and Todd M Bacastow. Spacenet: A remote sensing dataset and challenge series. CoRR, 2018. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 365, + 287, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 365, + 287, + 407 + ], + "spans": [ + { + "bbox": [ + 48, + 365, + 287, + 407 + ], + "type": "text", + "content": "[39] Adam Van Etten, Jacob Shermeyer, Daniel Hogan, Nicholas Weir, and Ryan Lewis. Road network and travel time extraction from multiple look angles with spacenet data. In IGARSS, 2020. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 411, + 287, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 411, + 287, + 453 + ], + "spans": [ + { + "bbox": [ + 48, + 411, + 287, + 453 + ], + "type": "text", + "content": "[40] Junjue Wang, Zhuo Zheng, Ailong Ma, Xiaoyan Lu, and Yanfei Zhong. Loveda: A remote sensing land-cover dataset for domain adaptive semantic segmentation. arXiv preprint arXiv:2110.08733, 2021. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 456, + 287, + 509 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 456, + 287, + 509 + ], + "spans": [ + { + "bbox": [ + 48, + 456, + 287, + 509 + ], + "type": "text", + "content": "[41] Ben G Weinstein, Sergio Marconi, Stephanie A Bohlman, Alina Zare, Aditya Singh, Sarah J Graves, and Ethan P White. A remote sensing derived data set of 100 million individual tree crowns for the national ecological observatory network. *Elife*, 2021. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 511, + 287, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 511, + 287, + 553 + ], + "spans": [ + { + "bbox": [ + 48, + 511, + 287, + 553 + ], + "type": "text", + "content": "[42] Chen Wu, Hongruixuan Chen, Bo Do, and Liangpei Zhang. Unsupervised change detection in multi-temporal vhr images based on deep kernel pca convolutional mapping network. IEEE Transactions on Cybernetics, 2019. 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 556, + 287, + 587 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 556, + 287, + 587 + ], + "spans": [ + { + "bbox": [ + 48, + 556, + 287, + 587 + ], + "type": "text", + "content": "[43] Zhirong Wu, Yuanjun Xiong, Stella X. Yu, and Dahua Lin. Unsupervised feature learning via non-parametric instance discrimination. In CVPR, 2018. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 590, + 287, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 590, + 287, + 611 + ], + "spans": [ + { + "bbox": [ + 48, + 590, + 287, + 611 + ], + "type": "text", + "content": "[44] Richard Zhang, Phillip Isola, and Alexei A Efros. Colorful image colorization. In ECCV, 2016. 2" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5270" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Chat2Map_ Efficient Scene Mapping From Multi-Ego Conversations/9562f06a-6203-4663-996c-df2336adbe79_content_list.json b/2023/Chat2Map_ Efficient Scene Mapping From Multi-Ego Conversations/9562f06a-6203-4663-996c-df2336adbe79_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..6a9914d613dab97e913895a557d5ee74df23eac8 --- /dev/null +++ b/2023/Chat2Map_ Efficient Scene Mapping From Multi-Ego Conversations/9562f06a-6203-4663-996c-df2336adbe79_content_list.json @@ -0,0 +1,1573 @@ +[ + { + "type": "text", + "text": "Chat2Map: Efficient Scene Mapping from Multi-Ego Conversations", + "text_level": 1, + "bbox": [ + 142, + 130, + 828, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Sagnik Majumder $^{1,2,3}$ Hao Jiang $^{2}$ Pierre Moulon $^{2}$ Ethan Henderson $^{2}$ Paul Calamia $^{2}$ Kristen Grauman $^{1,3*}$ Vamsi Krishna Ithapu $^{2*}$ UT Austin $^{2}$ Reality Labs Research, Meta FAIR", + "bbox": [ + 184, + 179, + 782, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 268, + 313, + 284 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Can conversational videos captured from multiple egocentric viewpoints reveal the map of a scene in a cost-efficient way? We seek to answer this question by proposing a new problem: efficiently building the map of a previously unseen 3D environment by exploiting shared information in the egocentric audio-visual observations of participants in a natural conversation. Our hypothesis is that as multiple people (\"egos\") move in a scene and talk among themselves, they receive rich audio-visual cues that can help uncover the unseen areas of the scene. Given the high cost of continuously processing egocentric visual streams, we further explore how to actively coordinate the sampling of visual information, so as to minimize redundancy and reduce power use. To that end, we present an audio-visual deep reinforcement learning approach that works with our shared scene mapper to selectively turn on the camera to efficiently chart out the space. We evaluate the approach using a state-of-the-art audio-visual simulator for 3D scenes as well as real-world video. Our model outperforms previous state-of-the-art mapping methods, and achieves an excellent cost-accuracy tradeoff. Project: http://vision.cs.utexas.edu/projects/chat2map.", + "bbox": [ + 75, + 301, + 473, + 632 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 654, + 209, + 669 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The spatial layout of the environment around us is fundamental to understanding our physical context. By representing the walls, furniture, and other major structures in a space, scene maps ground activity and objects in a persistent frame of reference, facilitating high-level reasoning for many downstream applications in augmented reality (AR) and robotics. For example, episodic memory [16, 28] aims to relocalize lost objects observed in first-person video (where are my keys?); floorplan estimation [10, 40, 47] aims to chart out the area and shapes of complex buildings; navigating agents try to discover routes in unfamiliar spaces [4, 11, 53].", + "bbox": [ + 75, + 679, + 470, + 845 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "While traditional computer vision approaches for mapping (e.g., visual SLAM) are highly effective when extensive", + "bbox": [ + 76, + 845, + 470, + 876 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/02b86ba0a5d0fef8f1d91542217a239f3c2a45bea27902b7cd21a1f522f16f87.jpg", + "image_caption": [ + "Figure 1. Given egocentric audio-visual observations from multiple people wearing AR glasses and moving and conversing (left), we aim to accurately map the scene (right). To mitigate cost, our model receives audio continuously but learns to selectively employ the ego cameras only when the visual input is expected to be informative." + ], + "image_footnote": [], + "bbox": [ + 503, + 266, + 893, + 511 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "exposure to the environment is possible, in many real-world scenarios only a fraction of the space is observed by the camera. Recent work shows the promise of sensing 3D spaces with both sight and sound [8, 13, 24, 26, 52]: listening to echoes bounce around the room can reveal the depth and shape of surrounding surfaces, and even help extrapolate a floorplan beyond the camera's field of view or behind occluded objects [52].", + "bbox": [ + 496, + 609, + 893, + 731 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "While we are inspired by these advances, they also have certain limitations. Often systems will emit sounds (e.g., a frequency sweep) into the environment to ping for spatial information [1, 13, 14, 22, 26, 39, 52, 61], which is intrusive if done around people. Furthermore, existing audio-visual models assume that the camera is always on grabbing new frames, which is wasteful if not intractable, particularly on lightweight, low-power computing devices in AR settings.", + "bbox": [ + 496, + 732, + 893, + 853 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We introduce Chat2Map, a new scene mapping task aimed at eliminating these challenges. In the proposed setting, multiple people converse as they move casually through", + "bbox": [ + 498, + 854, + 893, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Equal contribution", + "bbox": [ + 94, + 886, + 202, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "10554", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "the scene while wearing AR glasses equipped with an egocentric camera, microphones, and potentially other sensors (e.g., for odometry).1 Given their egocentric audio-visual data streams, the goal is to infer the ground-plane occupancy map for the larger environment around them. See Figure 1.", + "bbox": [ + 75, + 90, + 470, + 167 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We observe that audio-visual data from the egos' interactions will naturally reflect scene structure. First, as they walk and talk, their movements reveal spaces like corridors, doorways, and large rooms, in both modalities. Second, the speech captured by the device-wearer's cameras and microphones can be localized to different speakers, which, compared to active sound emission, is non-intrusive.", + "bbox": [ + 75, + 167, + 470, + 271 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To realize this vision, we develop a novel approach to efficient scene mapping from multi-ego conversations. Our approach has two key elements: a shared scene mapper and a visual sampling policy. For the former, we devise a transformer-based mapper that incorporates the multiple data streams to infer a map beyond the directly observed areas, and, most importantly, that enables communication among the egos about their observations and states in the 3D space to improve mapping accuracy. For the latter, our idea is to relax the common assumption of an \"always-on\" camera, and instead actively select when to sample visual frames from any one of the ego cameras. Intuitively, certain regions where the egos move will be more or less important for mapping (e.g., corners of the room, doors). We train a sampling policy with deep reinforcement learning that activates the visual feed only when it is anticipated to complement the continuous audio feed. This is a cost-conscious approach, mindful that switching on a camera is much more power consuming than sensing audio with microphones [2, 37].", + "bbox": [ + 75, + 272, + 470, + 559 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We demonstrate our approach using a state-of-the-art audio-visual simulator for 3D scenes [8] as well as real-world video input. We can successfully map an unfamiliar environment given only partial visibility via multiple conversing people moving about the scene. Compared to sampling all visual frames, our model reduces the visual capture and processing by $87.5\\%$ while the mapping accuracy declines only marginally $(\\sim 9\\%)$ .", + "bbox": [ + 75, + 560, + 470, + 679 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our main contributions are 1) we define the task of efficient and shared scene mapping from multi-ego conversations, a new direction for AR/VR research; 2) we present the first approach to begin tackling this task, specifically an RL-based framework that integrates an audio-visual scene mapper and a smart visual sampling policy; and 3) we rigorously experiment with a variety of environments (both in simulation and the real world), speech sounds, and use cases.", + "bbox": [ + 75, + 680, + 470, + 801 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 76, + 811, + 218, + 828 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Visual scene mapping. Past works tackle scene mapping using 3D Manhattan layouts [18, 64, 69, 74, 75], detailed", + "bbox": [ + 76, + 837, + 468, + 868 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "floorplans [10, 40, 47, 62, 67], occupancy [21, 34, 46, 54, 59, 60], and semantic maps [45]. Manhattan layouts include structured outputs like scene boundaries [64, 74, 75], corners [74, 75], and floor/ceilings [69, 75], but do not generalize to unseen environment regions. Floorplan estimation methods use dense scans of 3D scenes to predict geometric (walls, exterior/ interior) and semantic layouts (room type, object type, etc.)., rely on extensive human walkthroughs with RGB-D [10, 40] or 3D point cloud [47, 62] scans, and are usually limited to polygonal layouts [10, 40, 47, 62, 67]. Occupancy maps traditionally rely on wide field-of-view (FoV) LiDAR scanners [56] or evaluate on simple 2D environments without non-wall obstacles [21, 34, 60, 60]. More recent methods [4, 5, 11, 53] train an embodied agent to explore and build top-down maps of more complex scenes using RGB-D. On the contrary, our method uses both vision and audio from the observations of a group of conversing people for mapping. Rather than steer the camera of a robot to map the scene, our task requires processing passive video from human camera wearers.", + "bbox": [ + 500, + 90, + 893, + 393 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Audio-visual scene mapping. To our knowledge, the only prior work to translate audio-visual inputs into a general (arbitrarily shaped) floorplan maps is AV-Floorplan [52]. Unlike AV-Floorplan, our method infers maps from speech in natural human conversations, which avoids emitting intrusive frequency sweep signals to generate echoes. In addition, a key goal of our work is to reduce mapping cost by skipping redundant visual frames. Our experiments demonstrate the benefits of our model design over AV-Floorplan [52].", + "bbox": [ + 498, + 401, + 893, + 537 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Audio(-visual) spatial understanding. More broadly, beyond the mapping task, various methods leverage audio for geometric and material information about the 3D scene and its constituent objects. Prior work relies on acoustic reflections to estimate the shape of objects [39]. Echolocation is used in robotics to estimate proximity to surrounding surfaces [1, 14, 22, 61]. Together, vision and audio can better reveal the shape and materials of objects [48, 58, 73], self-supervise imagery [26], and improve depth sensing [35, 70]. Recent work exploits correlations between spatial audio and imagery to reason about scene acoustics [7, 44] or aid active embodied navigation [6, 9, 17, 25, 72] and source separation [42, 43]. No prior work intelligently selects images to capture during conversations to efficiently map a scene.", + "bbox": [ + 498, + 545, + 895, + 757 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Multi-agent spatial understanding. There is existing work [15, 29, 31, 32, 51] in the visual multi-agent reinforcement learning (MARL) community that learns collaborative agents for performing tasks like relocating furniture [31, 32], playing 3D multi-player games [30], coordinated scene exploration [29], or multi-object navigation [51]. In such settings, the collaborative agents actively interact with the environment to learn a shared scene representation for successfully completing their task. In contrast, we aim to learn a", + "bbox": [ + 498, + 763, + 895, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_footnote", + "text": "1 Throughout, we call each person participating in the conversation an \"ego\" for short.", + "bbox": [ + 73, + 875, + 468, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "10555", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "shared geometric map of a 3D scene given passive observations that come from the trajectories chosen by a group of people involved in a natural conversation.", + "bbox": [ + 75, + 90, + 470, + 136 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Efficient visual sampling in video. Efficient visual sampling has been studied in the context of video recognition [27, 36, 38, 68, 71] and summarization [12, 63] with the goal of selectively processing informative frames, which can both reduce computational cost and improve recognition. More closely related to our approach are methods that use audio for the decision-making [27, 36, 50]. Different from the above, we use efficient visual sampling in the context of mapping scenes. Furthermore, in our case an online sampling decision needs to be made at every step before looking at the current visual frame (or frames from future steps).", + "bbox": [ + 75, + 145, + 470, + 311 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Chat2Map Task Formulation", + "text_level": 1, + "bbox": [ + 76, + 327, + 352, + 343 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We propose a novel task: efficient and shared mapping of scenes from multi-ego conversations.", + "bbox": [ + 75, + 353, + 470, + 382 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Without loss of generality, we consider two egos, $E_{1}$ and $E_{2}$ , each wearing AR glasses equipped with an RGB-D camera and a multi-channel microphone array. The egos have a conversation and move around in an unmapped 3D environment. Each conversation is $T$ steps long. At each step $t$ , the ego $E_{i}$ 's glasses receive an observation $\\mathcal{O}_{i,t} = (\\mathcal{V}_{i,t}, S_{i,t}, P_{i,t}, S_{i,t}', P_{i,t}')$ . $\\mathcal{V}_{i,t}$ is the $90^{\\circ}$ FOV RGB-D image and $S_{i,t}$ is the speech waveform uttered by $E_{i}$ , as observed from its pose $P_{i,t} = (x_{i,t}, y_{i,t}, \\theta_{i,t})$ , where $(x_{i,t}, y_{i,t})$ denotes its location and $\\theta_{i,t}$ denotes its orientation in the 3D scene. $S_{i,t}'$ is the speech of the other ego $E_{i}'$ (the other person involved in the conversation), as perceived by $E_{i}$ (note, the speech sounds different depending on the listener position), and $P_{i,t}'$ is $E_{i}'$ 's pose relative to $E_{i}$ . Modern AR glasses, like Bose Frames or Facebook Aria already support capturing such multi-sensory observations, making it possible to have a real-world instantiation of our task.", + "bbox": [ + 75, + 383, + 470, + 642 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Given the real-time observation stream $\\mathcal{O}$ for the egos, where $\\mathcal{O} = \\{\\mathcal{O}_{i,t}:i = 1,\\dots ,2,t = 1\\dots T\\}$ and a total budget of visual frames $B$ we aim to learn a model that can accurately estimate the top-down occupancy map $M$ of the scene without exceeding the visual budget. The fixed budget represents allotting a fixed amount of battery to the energy-intensive visual capture [37], so that other on-device routines are not affected. We assume the first visual frames (at $t = 1$ ) for both egos to be used by the model. Thus we aim to learn a policy that samples $B$ frames from $2*(T - 1)$ choices—which are not considered a batch, but rather unfold in sequence—and a mapper that predicts the scene map given the sampled frames. Our goal is to build a model that samples the expensive visual frames only when absolutely needed for scene mapping. This is captured by the constraint $1\\leq B\\ll 2*(T - 1)$ .", + "bbox": [ + 75, + 643, + 470, + 883 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "There are three important aspects to our task. First, it", + "bbox": [ + 96, + 885, + 470, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "requires learning from both vision and audio. While the visual signal carries rich information about the local scene geometry, there can be a high amount of redundancy in the visual feed captured during a conversation (e.g., the egos may visit the same location more than once or change their viewpoint only marginally). Second, not only does the long-range nature of audio help uncover the global scene properties [19, 52] like shape and size—beyond what is visible in images—but also we can exploit audio to undersample the visual frames, thereby reducing the cost of capturing and processing sensory inputs for mapping. Third, shared mapping of a scene allows jointly leveraging the complementary information in the audio (speech) from the self and other egos, and the synergy of the audio-visual cues from multiple egos. These insights form the basis of our key hypothesis in this task—selectively sampling visual frames during a conversation involving egos that share information with each other can facilitate efficient mapping of a scene.", + "bbox": [ + 496, + 90, + 893, + 363 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4. Approach", + "text_level": 1, + "bbox": [ + 498, + 376, + 607, + 393 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We solve the task by learning a model that estimates the scene map given the egos' audio-visual observations and also sequentially decides when to sample visual frames for mapping given the audio stream, ego poses, and previously sampled frames, if any. Here, \"sampling\" refers to individually deciding for each ego whether to use its camera or not to capture the visuals at every step of its trajectory in the scene. The sampling is preemptive in nature, i.e. the policy selects or skips a frame without capturing it first.", + "bbox": [ + 496, + 402, + 893, + 537 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our model has two main components (see Fig. 2): (1) a shared scene mapper, and (2) a visual sampling policy. At every step $t$ , the shared mapper has two functions. First, it estimates the map of a previously unseen environment by exploiting the shared spatial cues in the audio-visual observations of the two egos. Second, it informs the policy about the utility of sampling a certain visual frame. Guided by the mapper, the policy samples only the most informative visual frames that are expected to boost mapping significantly over using just audio. Note that, unlike the visuals, we observe audio continuously as it is less resource-intensive vis-a-vis storage and power requirements for processing [2].", + "bbox": [ + 496, + 537, + 893, + 718 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We learn our task through the synergy of the mapper and the policy, such that under the constraint of a limited visual budget $B$ , our model implicitly understands which visual frames are critical for mapping.", + "bbox": [ + 496, + 719, + 890, + 779 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "First, we describe the steps involved to prepare our model inputs (Sec. 4.1). Next, we introduce our visual sampling policy (Sec. 4.2) and shared scene mapper (Sec. 4.3). Finally, we present model training details (Sec. 4.4). Through the rest of the text, we use separate notations to distinguish the egos' observations $\\mathcal{O}$ (i.e. what the egos receive from the environment) from our model inputs $O$ (i.e. what we capture and feed to our model for efficient mapping).", + "bbox": [ + 496, + 780, + 893, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "10556", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/b507832bb99458974534599042ac1c4d06072feaa28aab7988b17873014defa0.jpg", + "image_caption": [ + "Figure 2. Our model has two main components: a) a visual sampling policy (left), and b) a shared scene mapper (right). At each step, our policy receives the current audio along with the previous audio(-visual) observations for the egos and decides for each ego individually whether to capture its visual frame at the current step. As per the policy predictions, the shared mapper conditionally uses the current visual frame(s) and audio along with the past audio(-visual) observations to predict the occupancy map of the scene, a ground-plane map showing where obstacles and freespace are (shown in green and white)." + ], + "image_footnote": [], + "bbox": [ + 138, + 88, + 480, + 388 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/b6609000eeaff0df62be6fdbce6ef0677b96427cdf1fa89d7ea4912de78343be.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 488, + 88, + 830, + 388 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1. Model input preparation", + "text_level": 1, + "bbox": [ + 76, + 473, + 303, + 489 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We prepare our model inputs by separately preprocessing the visual and audio modalities. If our policy decides to sample an image $\\mathcal{V}$ , we transform it into $V = (V^{R}, V^{M})$ . $V^{R}$ denotes the normalized RGB image with pixel values $\\in [0,1]$ . $V^{M}$ denotes the $90^{\\circ}$ FoV top-down occupancy map created by projecting the depth image to the ground plane. To do the depth projection, we first backproject it into the world coordinates using the camera's intrinsic parameters to compute the local visible scene's 3D point cloud. Next, we project these points to obtain a two-channel binary top-down map of size $h \\times w \\times 2$ , where the first channel reveals occupied/free areas, and the second channel reveals seen/unseen areas. If our policy skips $\\mathcal{V}$ , we set $V^{R}$ and $V^{M}$ to all-zero matrices of the appropriate size.", + "bbox": [ + 75, + 498, + 470, + 709 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For a speech waveform $S$ , we calculate the short-time Fourier transform (STFT) magnitude spectrogram denoted by $S$ of size $F \\times T \\times C$ , where $F$ , $T$ , and $C$ are the number of frequency bins, time windows, and ambisonic microphone channels, respectively. Lastly, we normalize each pose $P_{i,t}$ to be relative to $P_{1,1}$ . See Sec. 5 and Supp. for more details.", + "bbox": [ + 75, + 710, + 470, + 801 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.2. Visual sampling policy", + "text_level": 1, + "bbox": [ + 76, + 815, + 285, + 832 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "At every step $t$ , our visual sampling policy $\\pi^V$ (Fig. 2 left) receives $O^\\pi(t)$ as input and makes the decision to either capture or skip the visual frame $\\mathcal{V}_{i,t}$ for each ego $E_i$ . $O^\\pi(t)$ comprises the visual cue from the last step along with the", + "bbox": [ + 75, + 839, + 470, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "speech cues, and the poses from the current step and the last step for both egos. Formally, $O^{\\pi}(t) = \\{O_{i}^{\\pi}(t):i = 1\\ldots 2\\}$ , where $O_{i}^{\\pi}(t) = \\{V_{i,t - 1},S_{i,j},P_{i,j},S_{i,j}^{\\prime},P_{i,j}^{\\prime}:j = t - 1\\ldots t\\}$ . The policy first uses an encoder network to generate a multi-modal embedding of $O^{\\pi}(t)$ , and then passes the embedding to a policy network that makes a sampling decision per ego. At $t = 1$ , as per our problem definition (Sec. 3), the policy always chooses to sample the visual frames for both egos, i.e., the cameras are initially on.", + "bbox": [ + 496, + 473, + 890, + 614 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Multi-modal policy embedding. To process ego $E_{i}$ 's visual input $V_{i,t-1}$ from the last step, we encode the RGB image $V_{i,t-1}^{R}$ and map $V_{i,t-1}^{M}$ with separate CNNs. We then concatenate the two features to generate the visual embedding $v_{i,t-1}$ . To encode the pose inputs $\\{P_{i,t-1}, P_{i,t-1}', P_{i,t}, P_{i,t}'\\}$ , we use a linear layer and generate pose embeddings $\\{p_{i,t-1}, p_{i,t-1}', p_{i,t}, p_{i,t}'\\}$ . We process the speech inputs $\\{S_{i,t-1}, S_{i,t-1}', S_{i,t}, S_{i,t}'\\}$ using another CNN and create speech embeddings $\\{s_{i,t-1}, s_{i,t-1}', s_{i,t}, s_{i,t}'\\}$ . Next, we fuse the visual, speech and pose embeddings using linear layers (see Fig. 2 left) to obtain the multi-modal policy embedding $e_{i,t}$ for $E_{i}$ . Finally, we fuse the policy embeddings for the two egos, $e_{1,t}$ and $e_{2,t}$ with a linear layer to produce the multi-modal policy embedding $e_t$ .", + "bbox": [ + 496, + 619, + 892, + 853 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The visual, audio, and pose inputs carry complementary cues required for efficient visual sampling. Whereas the pose inputs from the last and current steps explicitly reveal", + "bbox": [ + 496, + 854, + 893, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "10557", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "the viewpoint change between the steps, the previous and current speech inputs provide information about the changes in the local and global scene structures as a function of the previously sampled visual inputs, which together suggest the value of sampling a visual frame at the current step. Furthermore, guided by our training reward (below in Sec. 4.4), the previously observed visual frames and audio together enable our policy to anticipate the current frames and skip them if they are deemed redundant, thereby improving mapping accuracy for a low visual budget.", + "bbox": [ + 75, + 90, + 472, + 243 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Policy network. The policy network consists of a GRU that estimates an updated history $h_t$ along with the current state representation $g_t$ , using the fused embedding $e_t$ and the history of states $h_{t-1}$ . An actor-critic module takes $g_t$ and $h_{t-1}$ as inputs and predicts a policy distribution $\\pi_\\theta(a_{i,t}|g_t, h_{t-1})$ per ego along with the value of the state $H_\\theta(g_t, h_{t-1})$ ( $\\theta$ are policy parameters). The policy samples an action $a_{i,t} \\in \\{0,1\\}$ for every $E_i$ . $a_{i,t} = 1$ corresponds to selecting $\\mathcal{V}_{i,t}$ , whereas $a_{i,t} = 0$ means to skip that frame.", + "bbox": [ + 76, + 250, + 472, + 387 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.3. Shared scene mapper", + "text_level": 1, + "bbox": [ + 76, + 396, + 277, + 412 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Whereas $O^{\\pi}(t)$ denotes our policy input (Sec. 4.2), $O^{M}(t)$ denotes the input to our shared scene mapper $f^{M}$ at step $t$ , such that $O^{M}(t) = \\left\\{(V_{i,j}, S_{i,j}, S_{i,j}^{\\prime}, P_{i,j}, P_{i,j}^{\\prime}): i = 1 \\dots 2, j = 1 \\dots t\\right\\}$ . $f^{m}$ starts by embedding each component of $O^{M}(t)$ using a separate network. This is followed by a multi-modal memory that stores the embeddings since the start of the episode. Finally, a transformer [65] predicts an estimate $\\tilde{M}(t)$ of the scene map conditioned on the multi-modal memory and the egos' poses in the episode.", + "bbox": [ + 75, + 419, + 472, + 559 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Multi-modal mapper embedding. For the visual input $V_{i,j}$ , we encode $V_{i,j}^{R}$ and $V_{i,j}^{M}$ using separate CNNs and do a channel-wise concatenation to get visual features $\\hat{v}_{i,j}$ . Similarly, speech is encoded using separate CNNs to get $\\hat{s}_{i,j}$ and $\\hat{s}_{i,j}'$ . Each of $\\hat{v}$ , $\\hat{s}$ and $\\hat{s}'$ is of size $4 \\times 4 \\times 1024$ .", + "bbox": [ + 76, + 565, + 472, + 643 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For both vision and speech, we compute two positional embeddings, $p^{\\mathrm{I}}$ and $p^{\\mathrm{II}}$ . They encode the pose of the egos in the 3D space, and the index of each 1024-dimensional feature in the visual or speech features in the raster order respectively. Whereas $p^{\\mathrm{I}}$ helps discover spatial cues as a function of the egos' location in the 3D scene, $p^{\\mathrm{II}}$ enables our model to attend to different modalities in a more fine-grained manner. For both, we compute an 8-dimensional sinusoidal positional encoding [65] and then pass it through a linear layer to obtain a 1024-dimensional embedding. For $p^{\\mathrm{II}}$ , we additionally repeat this process for every feature index in the raster order. Lastly, we reshape $p^{\\mathrm{I}}$ and add it with $p^{\\mathrm{II}}$ to produce $4 \\times 4 \\times 1024$ -dimensional positional embeddings, $\\hat{p}_{i,j}$ for $\\hat{v}_{i,j}$ and $\\hat{s}_{i,j}$ , and $\\hat{p}_{i,j}'$ for $\\hat{s}_{i,j}'$ .", + "bbox": [ + 75, + 643, + 472, + 856 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Following [44], we also learn an embedding $\\hat{m}_{i,j} \\in \\{\\hat{m}_V, \\hat{m}_S, \\hat{m}_{S'}\\}$ to capture different modality types, where $\\hat{m}_V$ represents vision, and $\\hat{m}_S$ and $\\hat{m}_{S'}$ represent the speech", + "bbox": [ + 76, + 856, + 470, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "from self and that of the other ego, respectively. The modality-based embeddings help our model differentiate between different modalities and better map the scene by learning complementary spatial cues from them.", + "bbox": [ + 496, + 90, + 890, + 151 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Multi-modal memory. For the visual input $V_{i,j}$ , we add its embedding $\\hat{v}_{i,j}$ with its positional embedding $\\hat{p}_{i,j}$ and modality embedding $\\hat{m}_{i,j}^{V}$ , and flatten the sum to get a $16 \\times 1024$ -dimensional embedding. Similarly, we fuse the speech embeddings by taking their sum and flattening them. This generates a multi-modal memory of fused embeddings $o$ , such that $o = \\{o_{1,1}^{V}, \\ldots, o_{2,t}^{V}, o_{1,1}^{S}, \\ldots, o_{2,t}^{S}, o_{1,1}^{S'}, \\ldots, o_{2,t}^{S'}\\}$ .", + "bbox": [ + 496, + 155, + 893, + 263 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Occupancy prediction. To predict the underlying scene occupancy, we first use a transformer encoder [65] to attend to the embeddings in $o$ and capture short- and long-range correlations within and across modalities using a stack of self-attention layers. This generates an audio-visual representation that models the spatial layout of the 3D scene.", + "bbox": [ + 496, + 266, + 893, + 356 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Next, we use a transformer decoder [65] to perform cross-attention on the audio-visual representation of the scene conditioned on the embedding $\\hat{p}_{i,j}$ for every pose $P_{i,j}$ in $O^M (t)$ and generate an embedding $d_{i,j}$ for the pose. Finally, we upsample $d_{i,j}$ using a multi-layer network $U$ comprising transpose convolutions and a sigmoid layer at the end to predict an estimate $\\tilde{M}_{i,j}$ of the ground-truth local $360^{\\circ}$ FoV map for the pose, $M_{i,j}$ . Both $M_{i,j}$ and its estimate $\\tilde{M}_{i,j}$ are two-channel binary occupancy maps of size $H\\times W$ . To obtain the estimated map $\\tilde{M} (t)$ for the scene, we register each prediction $\\tilde{M}_{i,j}$ onto a larger shared map using the pose $P_{i,j}$ and threshold the final shared map at 0.5 (see Supp. for map registration details). Importantly, the shared map allows communication between both egos' data streams for more informed mapping and sampling, as we show in results.", + "bbox": [ + 496, + 357, + 893, + 585 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.4. Model training", + "text_level": 1, + "bbox": [ + 498, + 593, + 653, + 609 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Policy training. We propose a novel dense RL reward to train policy $\\pi^V$ :", + "bbox": [ + 496, + 617, + 890, + 647 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nr (t) = \\Delta Q (t) - \\eta * \\rho (t).\n$$\n", + "text_format": "latex", + "bbox": [ + 602, + 659, + 785, + 676 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "$\\Delta Q(t)$ measures the improvement in mapping from taking actions $\\{a_{i,t} : i = 1 \\dots 2\\}$ over not sampling any visual frame at step $t$ . $\\rho(t)$ is a penalty term to discourage sampling a frame from the same pose more than once, which we weight by $\\eta$ . We define $\\Delta Q(t)$ as", + "bbox": [ + 496, + 686, + 890, + 763 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\Delta Q (t) = Q \\left(\\tilde {M} (t) \\mid O ^ {M} (t)\\right) - Q \\left(\\tilde {M} (t) \\mid \\left(O ^ {M} (t) \\setminus V _ {t}\\right)\\right),\n$$\n", + "text_format": "latex", + "bbox": [ + 501, + 771, + 888, + 791 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $Q$ is a map quality measure, $Q(X|Y)$ represents the quality of map estimate $X$ given inputs $Y$ , and $(O^M(t) \\setminus V_t)$ denotes the mapper inputs devoid of any visual frame for the current step. We define $\\rho(t)$ as", + "bbox": [ + 496, + 800, + 893, + 861 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\rho (t) = \\sum_ {i = 1.. 2} a _ {i, t} * \\mathbb {1} (V _ {i, t} \\in O ^ {M} (t - 1)),\n$$\n", + "text_format": "latex", + "bbox": [ + 553, + 872, + 836, + 904 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "10558", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where the indicator function checks if $V_{i,t}$ was used in mapping before. While $\\Delta Q(t)$ incentivizes sampling frames that provide a big boost to the mapping accuracy over skipping them, $\\rho(t)$ penalizes wasting the visual budget on redundant sampling, thereby maximizing mapping performance within the constraints of a limited budget. We set $\\rho = 0.03$ in all our experiments and define $Q$ as the average F1 score over the occupied and free classes in a predicted occupancy map. We empirically observe that alternate penalty terms based on visual similarity, like the ORB [55] feature matching similarity, do not improve the model performance.", + "bbox": [ + 75, + 90, + 472, + 256 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We train $\\pi^V$ with Decentralized Distributed PPO (DDPPO) [66]. The DD-PPO loss consists of a value loss, policy loss and an entropy loss to promote exploration (see Supp. for details).", + "bbox": [ + 75, + 256, + 470, + 316 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Mapper training. At each step $t$ , we train the shared mapper $f^{m}$ with a loss $\\mathcal{L}^{M}(t)$ , such that", + "bbox": [ + 76, + 323, + 472, + 354 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} ^ {M} (t) = \\frac {1}{2 \\times t} \\sum_ {i = 1 \\ldots 2} \\sum_ {j = 1 \\ldots t} \\mathrm {B C E} (\\tilde {M} _ {i, j}, M _ {i, j}),\n$$\n", + "text_format": "latex", + "bbox": [ + 114, + 364, + 429, + 402 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\mathrm{BCE}(\\tilde{M}_{i,j}, M_{i,j})$ is the average binary cross entropy loss between $\\tilde{M}_{i,j}$ and $M_{i,j}$ .", + "bbox": [ + 75, + 414, + 468, + 449 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Training curriculum. To train our model, we first pretrain mapper $f^{m}$ in two phases and then train the policy $\\pi^{V}$ while keeping $f^{m}$ frozen. In phase 1, we train $f^{m}$ without visual sampling, i.e. all visual frames are provided at each step. In phase 2, we finetune the pretrained weights of $f^{m}$ from phase 1 on episodes where we randomly drop views to satisfy the budget $B$ . While phase 1 improves convergence when training with visual sampling, phase 2 helps with reward stationarity when training our RL policy. Both phases use the same training, validation and testing data.", + "bbox": [ + 75, + 452, + 472, + 604 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 76, + 618, + 209, + 633 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Experimental setup. For our main experiments, we use SoundSpaces [8] acoustic simulations with AI-Habitat [57] and Matterport3D [3] visual scenes. While Matterport3D provides dense 3D meshes and image scans of real-world houses and other indoor scenes, SoundSpaces provides room impulse responses (RIRs) at a spatial resolution of $1\\mathrm{m}$ for Matterport3D that model all real-world acoustic phenomena [8]. This setup allows us to evaluate with as many as 83 scenes, split in 56/10/17 for train/val/test, compare against relevant prior work [52, 53] and report reproducible results.", + "bbox": [ + 75, + 643, + 472, + 794 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We also collect real-world data in a mock-up apartment due to the absence of a publicly available alternative suited for our task. We capture a dense set of RGB images using a Samsung S22 camera and generate the corresponding depth images using monocular depth estimation [20, 33]. To compute the RIRs, following [23], we generate a sinusoidal sweep sound from $20\\mathrm{Hz} - 20\\mathrm{kHz}$ with a loudspeaker at source", + "bbox": [ + 75, + 795, + 472, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "location, capture it with an Eigenmike at a receiver location, and convolve the spatial sound with the inverse of the sweep sound to retrieve the RIR. All capturing devices are placed at a height of $1.5\\mathrm{m}$ . We generate occupancy maps by backprojecting the depth images (cf. Sec. 4.1) and register them onto a shared top-down map before taking egocentric crops to generate the local occupancy inputs and targets.", + "bbox": [ + 496, + 90, + 893, + 196 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Note that both datasets have real-world visuals as they are captured in the real environments; SoundSpaces has high-quality simulated audio while the apartment data has real-world collected audio RIRs.", + "bbox": [ + 496, + 196, + 893, + 257 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Conversation episode. For each episode,(simulated or real), we randomly place the egos in a scene. Episode length is $T = 16$ and 8 for simulation and real, respectively. At each step, the egos execute a movement from $A = \\{MoveForward, TurnLeft, TurnRight\\}$ , where $MoveForward$ moves an ego forward by $1\\mathrm{m}$ , and the $Turn$ actions rotate the ego by $90^{\\circ}$ , as required by SoundSpaces. Further, either of the egos speaks or both speak with equal probability of $\\frac{1}{3}$ at every step, i.e., there are no moments of total silence, but one ego may be silent. The egos stay between $1 - 3\\mathrm{m}$ from each other (typical in a conversation) so that they do not collide and each ego is audible by the other at all times. This results in train/val splits of 1,955,334/100 episodes in simulation, and a simulated/real-world test split of 1000/27 episodes. See Supp. for additional analysis on the model performance vs. training data size. We set the visual budget to $B = 2$ for our main experiments (see Supp. for $B = 4, 6$ evaluations). Note that these episodes are solely to generate video data; our task requires processing passive video, not controlling embodied agents.", + "bbox": [ + 496, + 268, + 895, + 571 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Observations and model output. For the occupancy maps, consistent with [4, 11, 53], we generate $(31 \\times 31 \\times 2)$ -dimensional input maps that cover $3.1 \\times 3.1 \\mathrm{~m}^2$ in area at a resolution of $0.1 \\mathrm{~m}$ , and set the local target map size to $H \\times W = 6.4 \\times 6.4 \\mathrm{~m}^2 (\\sim 41 \\mathrm{~m}^2)$ . For speech, we use 100 distinct speakers from LibriSpeech [49], split in 80/11 for heard/unheard, where unheard speech is only used in testing. We assume camera poses, as modern AR devices are equipped with motion sensors that can robustly estimate relative poses [41]. We test our robustness to ambient sounds that get mixed with the egos' speech, and incorporate odometry noise models [52, 53] in Supp.", + "bbox": [ + 496, + 575, + 895, + 758 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation settings. We evaluate our model in two settings: 1) passive mapping, the mapper has access to all visual frames in an episode (i.e., the camera is always on), and 2) active mapping, where the mapping agent has to actively sample frames to meet the visual budget $B$ . This helps disentangle our modeling contributions—whereas passive mapping lets us show improvements in the mapper $f^{M}$ over existing methods [52, 53], active mapping helps demonstrate the benefits of smart visual sampling.", + "bbox": [ + 496, + 763, + 895, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "10559", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/8d64f7683cf54910909e42e0ad498bf5e0d29481f8cfed33a002d27d3b78d204.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelSimulationReal world
F1 score ↑IoU ↑F1 score ↑IoU ↑
All-occupied63.448.836.223.8
Register-inputs72.660.150.835.0
OccAnt [53]74.562.753.938.3
AV-Floorplan [52]79.367.954.538.7
Ours81.871.455.539.2
Ours w/o vision72.860.350.835.0
Ours w/o audio78.166.754.138.0
Ours w/o \\( E_i' \\)'s speech81.570.955.439.1
Ours w/o shared mapping80.770.054.938.6
Ours w/o modality tag79.167.854.538.1
", + "bbox": [ + 102, + 88, + 444, + 256 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 1. Passive mapping performance $(\\%)$ . All our gains are statistically significant $(p \\leq 0.05)$ .", + "bbox": [ + 76, + 258, + 468, + 287 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We use standard evaluation metrics [53]: F1 score and IoU (intersection over union) between the predicted and target scene maps. For both metrics, we report the mean over the free and occupied classes. For active mapping, we average the metrics over 3 random seeds. We use the following baselines to compare our model's efficacy.", + "bbox": [ + 75, + 303, + 470, + 393 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Passive mapping:", + "bbox": [ + 76, + 396, + 196, + 410 + ], + "page_idx": 6 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- All-occupied: a naive baseline that predicts all locations in its map estimate as occupied", + "- Register-inputs: a naive baseline that registers the input maps onto a shared map and uses it as its prediction", + "- OccAnt [53]: a vision-only SOTA model that uses the RGB-D images at each step to anticipate the occupancy of the area around an ego that's outside its visible range.", + "- AV-Floorplan [52]: an audio-visual SOTA model that passively predicts the floorplan of a scene using a walk-through in it, where the audio is either self-generated or comes from semantic sources in the scene. We adapt the model for our occupancy prediction task and give it the exact same audio-visual observations as our model." + ], + "bbox": [ + 76, + 410, + 468, + 606 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Active mapping:", + "text_level": 1, + "bbox": [ + 75, + 608, + 187, + 622 + ], + "page_idx": 6 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Random: an agent that selects visual frames randomly for each ego as long as the budget allows", + "- Greedy: an agent that greedily uses up the visual budget by sampling frames as early as possible", + "- Unique-posed: an agent that samples a frame for every new ego pose in the episode" + ], + "bbox": [ + 76, + 623, + 468, + 712 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In active mapping, we use the model from the second pretraining phase (Sec. 4.4) as the mapper for all models for fair comparison. Thus, any difference in performance is due to the quality of each method's sampling decisions.", + "bbox": [ + 75, + 714, + 468, + 773 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "See Supp. for all other details like network architectures, training hyperparameters, and baseline implementation.", + "bbox": [ + 76, + 775, + 468, + 805 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.1. Map prediction results", + "text_level": 1, + "bbox": [ + 76, + 816, + 287, + 832 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Passive mapping. Table 1 (top) reports the prediction quality of all models in the passive mapping setting. Naive baselines (All-occupied, Register-inputs) perform worse than the learned models, showing the complexity of our map pre", + "bbox": [ + 75, + 839, + 470, + 901 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/4b8927fbc208d5e693d9e96cd47be53f85e258be176700e1909327fa1a4cd3f7.jpg", + "image_caption": [ + "(a) Simulation" + ], + "image_footnote": [], + "bbox": [ + 501, + 90, + 692, + 193 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/668d11c8601cda5195c65b3a29d2946ffeb20eae8e5fc6c9aa3fa8a2b3ea92a4.jpg", + "image_caption": [ + "(b) Real world" + ], + "image_footnote": [], + "bbox": [ + 697, + 89, + 890, + 194 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/5a69f564045c47faa92d760d19b7690880249e00dddc60a1ba3436c94afc154a.jpg", + "image_caption": [ + "Figure 3. Active mapping performance vs. episode step.", + "Figure 4. Impact of the other ego's speech on passive mapping vs. distance between the egos." + ], + "image_footnote": [], + "bbox": [ + 578, + 252, + 813, + 375 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "diction task. AV-Floorplan [52] fares the best among all baselines. Its improvement over OccAnt [53] demonstrates the benefits of exploiting the spatial cues in audio for mapping and using an attention-based model to leverage the long- and short-range correlations in the audio-visual inputs.", + "bbox": [ + 496, + 433, + 893, + 508 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Our method outperforms all baselines. Its improvement over AV-Floorplan [52] underlines the efficacy of performing attention at different granularities—across modalities, within a single modality and within a single input—guided by our positional and modality type embeddings. It also generalizes to the real-world setting and retains its benefits over the baselines, even without retraining on the real-world data. However, we do observe a drop in performance gains, probably due to the large sim-to-real gap.", + "bbox": [ + 496, + 510, + 893, + 647 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Active mapping. Fig. 3 shows the active mapping performance as a function of episode progress. Employing naive heuristics for sampling, like Random or Greedy, is insufficient for high-quality mapping, emphasizing the high levels of redundancy in the visual frames. Unique-pose improves over both Random and Greedy, showing that sampling diverse viewpoints provides more information about the underlying scene geometry.", + "bbox": [ + 496, + 657, + 893, + 779 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Although the baselines make initial progress, they flatten quickly and our model eventually outperforms them all, on both real-world and simulated data (gains are again statistically significant, $p \\leq 0.05$ ). This highlights the benefits of learning a smart policy that, given the audio streams and its past visual samples, understands the value of sampling a visual frame for mapping by taking cues from our novel reward. Besides, on the real-world data, we see better per", + "bbox": [ + 496, + 779, + 893, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "10560", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/cc04797f38573fc942b96e0d51e5c37a66bce51c35caabc21e0d304352baabe8.jpg", + "image_caption": [ + "Figure 5. Sample episodes for our active mapping model. While our policy samples only the salient visual frames, our mapper can both complete partially seen objects as well as anticipate objects never seen before in the sampled visuals (red boxes on the maps)." + ], + "image_footnote": [], + "bbox": [ + 78, + 88, + 893, + 289 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "formance margins over the baselines towards the end of episodes, showing that our policy can adaptively defer visual sampling to improve mapping. Owing to smart sampling, the per-episode reduction in the capture / processing cost for $B = 2$ is 74 Watts (W) / 7.2 GFLOPS in simulation and 37 W / 3.6 GFLOPS in the real world video. See Supp. for further analysis of our model's power and time cost.", + "bbox": [ + 75, + 345, + 472, + 450 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.2. Model analysis", + "text_level": 1, + "bbox": [ + 76, + 465, + 228, + 482 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Ablations. In Table 1 (bottom), we ablate the components of our model for passive mapping. Upon removing audio, our model experiences a large drop in mapping performance, which indicates that our model leverages complementary spatial cues in audio and vision. We also see a drop in the map quality when our model lacks access to the speech from the other ego $(E_i')$ . This shows that $E_i'$ 's speech can better reveal the more global scene geometry than $E_i$ 's own speech. Fig. 4 further shows that the impact of the other ego's speech becomes more prominent for larger inter-ego distances $(3 - 5\\mathrm{m}$ vs. $1 - 3\\mathrm{m})$ , in which case the two types of speech are dissimilar enough to carry complementary geometric cues, but reduces for even larger distances $(5\\mathrm{m}$ or more), in which case $E_i'$ is too far for its speech to carry useful cues about $E_i$ 's local scene geometry. Moreover, unlike the ablation that does not perform shared mapping, our model benefits significantly from jointly attending to the observations of the egos and exploiting the complementary information in them—even though both models use the exact same audio-visual observations, including both speech from self and the other ego. Finally, without our modality type embeddings, our model does worse, showing their role in learning complementary features from vision and audio.", + "bbox": [ + 75, + 491, + 470, + 837 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "For active mapping, Fig. 3 shows a drop in the mapping performance upon removing audio from the policy inputs. This implies that our policy exploits audio to reason about the level of redundancy in a new visual frame and improve", + "bbox": [ + 75, + 839, + 470, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "the mapping quality vs. visual budget tradeoff. On the more challenging real-world setting, audio plays an even bigger role, as shown by the larger performance drop in Fig. 3b.", + "bbox": [ + 496, + 345, + 890, + 390 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "See Supp. for similar results with 1) ambient sounds, 2) unheard speech, 3) higher budgets, 4) sensor noise, 5) larger target maps, 6) different ego initializations, and 7) multiple random data splits.", + "bbox": [ + 496, + 391, + 892, + 452 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Qualitative results. Fig. 5 shows two successful active mapping episodes of our method. Note how our model samples views that tend have to little visual overlap but are informative of the surrounding geometry (both occupied and free spaces). Besides, it is able to complete structures only partially visible in the sampled views, and more interestingly, leverage the synergy of audio and vision to anticipate unseen areas (red boxes on the occupancy maps in Fig. 5).", + "bbox": [ + 496, + 463, + 892, + 584 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Failure cases. We notice two common failure cases with active mapping: episodes where the egos stay at the same location, leading to very few informative visual frames to sample from; and episodes with highly unique visual samples at every trajectory step, in which case each sample is useful and our model behaves similar to Unique-pose or Greedy. For passive mapping, our model fails with very complex scenes that commonly have objects in spaces where both vision and audio cannot reach (e.g. narrow corners)", + "bbox": [ + 496, + 595, + 893, + 731 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 744, + 617, + 758 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We introduce Chat2Map, a new task aimed at scene mapping using audio-visual feeds from egocentric conversations. We develop a novel approach for Chat2Map comprising a shared scene mapper and a visual sampling policy based on a novel reinforcement learner that smartly samples the visuals only when necessary. We show promising performance on both simulated and real-world data from over 80 scenes.", + "bbox": [ + 496, + 766, + 893, + 869 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements: UT Austin is supported in part by the IFML NSF AI Institute. KG is paid as a research scientist by Meta.", + "bbox": [ + 500, + 872, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "10561", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Ego-Noise Predictions for Echolocation in Wheeled Robots, volume ALIFE 2019: The 2019 Conference on Artificial Life of ALIFE 2022: The 2022 Conference on Artificial Life, 07 2019. 1, 2", + "[2] Aaron Carroll and Gernot Heiser. An analysis of power consumption in a smartphone. In 2010 USENIX Annual Technical Conference (USENIX ATC 10), 2010. 2, 3", + "[3] Angel Chang, Angela Dai, Thomas Funkhouser, Maciej Halber, Matthias Niessner, Manolis Savva, Shuran Song, Andy Zeng, and Yinda Zhang. Matterport3d: Learning from rgb-d data in indoor environments. arXiv preprint arXiv:1709.06158, 2017. 6", + "[4] Devendra Singh Chaplot, Dhiraj Gandhi, Saurabh Gupta, Abhinav Gupta, and Ruslan Salakhutdinov. Learning to explore using active neural slam. In International Conference on Learning Representations, 2020. 1, 2, 6", + "[5] Devendra Singh Chaplot, Dhiraj Prakashchand Gandhi, Abhinav Gupta, and Russ R Salakhutdinov. Object goal navigation using goal-oriented semantic exploration. Advances in Neural Information Processing Systems, 33:4247-4258, 2020. 2", + "[6] Changan Chen, Ziad Al-Halah, and Kristen Grauman. Semantic audio-visual navigation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15516–15525, 2021. 2", + "[7] Changan Chen, Ruohan Gao, Paul T. Calamia, and Kristen Grauman. Visual acoustic matching. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 18836-18846, 2022. 2", + "[8] Changan Chen, Unnat Jain, Carl Schissler, Sebastia Vicenc Amengual Gari, Ziad Al-Halah, Vamsi Krishna Ithapu, Philip Robinson, and Kristen Grauman. Soundspaces: Audiovisual navigation in 3d environments. In ECCV, 2020. 1, 2, 6", + "[9] Changan Chen, Sagnik Majumder, Ziad Al-Halah, Ruohan Gao, Santhosh Kumar Ramakrishnan, and Kristen Grauman. Learning to set waypoints for audio-visual navigation. In International Conference on Learning Representations, 2021. 2", + "[10] Jiacheng Chen, Chen Liu, Jiaye Wu, and Yasutaka Furukawa. Floor-sp: Inverse cad for floorplans by sequential room-wise shortest path. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2661-2670, 2019. 1, 2", + "[11] Tao Chen, Saurabh Gupta, and Abhinav Gupta. Learning exploration policies for navigation. arXiv preprint arXiv:1903.01959, 2019. 1, 2, 6", + "[12] Yangyu Chen, Shuhui Wang, Weigang Zhang, and Qingming Huang. Less is more: Picking informative frames for video captioning. In Proceedings of the European conference on computer vision (ECCV), pages 358-373, 2018. 3", + "[13] Jesper Christensen, Sascha Hornauer, and Stella Yu. Batvision - learning to see 3d spatial layout with two ears. In ICRA, 2020. 1", + "[14] Jesper Haahr Christensen, Sascha Hornauer, and Stella X. Yu. Batvision: Learning to see 3d spatial layout with two ears. In 2020 IEEE International Conference on Robotics and Automation (ICRA), pages 1581-1587, 2020. 1, 2", + "[15] Abhishek Das, Théophile Gervet, Joshua Romoff, Dhruv" + ], + "bbox": [ + 78, + 114, + 470, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Batra, Devi Parikh, Michael G. Rabbat, and Joelle Pineau. Tarmac: Targeted multi-agent communication. ArXiv, abs/1810.11187, 2019. 2", + "[16] Samyak Datta, Sameer Dharur, Vincent Cartillier, Ruta Desai, Mukul Khanna, Dhruv Batra, and Devi Parikh. Episodic memory question answering. 2022. 1", + "[17] Victoria Dean, Shubham Tulsiani, and Abhinav Gupta. See, hear, explore: Curiosity via audio-visual association. Advances in Neural Information Processing Systems, 33:14961–14972, 2020. 2", + "[18] Helisa Dhamo, Nassir Navab, and Federico Tombari. Object-driven multi-layer scene decomposition from a single image. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5369-5378, 2019. 2", + "[19] Ivan Dokmanić, Reza Parhizkar, Andreas Walther, Yue M. Lu, and Martin Vetterli. Acoustic echoes reveal room shape. Proceedings of the National Academy of Sciences, 110(30):12186-12191, 2013. 3", + "[20] Ainaz Eftekhar, Alexander Sax, Jitendra Malik, and Amir Zamir. Omnidata: A scalable pipeline for making multi-task mid-level vision datasets from 3d scans. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10786-10796, 2021. 6", + "[21] Amine Elhafsi, Boris Ivanovic, Lucas Janson, and Marco Pavone. Map-predictive motion planning in unknown environments. In 2020 IEEE International Conference on Robotics and Automation (ICRA), pages 8552-8558. IEEE, 2020. 2", + "[22] Itamar Eliakim, Zahi Cohen, Gábor Kósa, and Yossi Yovel. A fully autonomous terrestrial bat-like acoustic robot. PLoS Computational Biology, 14, 2018. 1, 2", + "[23] Angelo Farina. Simultaneous measurement of impulse response and distortion with a swept-sine technique. Journal of The Audio Engineering Society, 2000. 6", + "[24] Chuang Gan, Jeremy Schwartz, Seth Alter, Martin Schrimpf, James Traer, Julian De Freitas, Jonas Kubilius, Abhishek Bhandwaldar, Nick Haber, Megumi Sano, Kuno Kim, Elias Wang, Damian Mrowca, Michael Lingelbach, Aidan Curtis, Kevin T. Feigelis, Daniel M. Bear, Dan Gutfreund, David D. Cox, James J. DiCarlo, Josh H. McDermott, Joshua B. Tenenbaum, and Daniel L. K. Yamins. Threadworld: A platform for interactive multi-modal physical simulation. In NeurIPS Track on Datasets and Benchmarks, 2021. 1", + "[25] Chuang Gan, Yiwei Zhang, Jiajun Wu, Boqing Gong, and Joshua B. Tenenbaum. Look, listen, and act: Towards audiovisual embodied navigation. In 2020 IEEE International Conference on Robotics and Automation (ICRA), pages 9701-9707, 2020. 2", + "[26] Ruohan Gao, Changan Chen, Ziad Al-Halah, Carl Schissler, and Kristen Grauman. Visualechoes: Spatial image representation learning through echolocation. In Andrea Vedaldi, Horst Bischof, Thomas Brox, and Jan-Michael Frahm, editors, Computer Vision - ECCV 2020, pages 658-676, Cham, 2020. Springer International Publishing. 1, 2", + "[27] Ruohan Gao, Tae-Hyun Oh, Kristen Grauman, and Lorenzo Torresani. Listen to look: Action recognition by previewing audio. 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10454-10464, 2020. 3", + "[28] Kristen Grauman, Andrew Westbury, Eugene Byrne, Zachary Chavis, Antonino Furnari, Rohit Girdhar, Jackson Hamburger," + ], + "bbox": [ + 503, + 92, + 893, + 892 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "10562", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Hao Jiang, Miao Liu, Xingyu Liu, Miguel Martin, Tushar Nagarajan, Ilija Radosavovic, Santhosh Kumar Ramakrishnan, Fiona Ryan, Jayant Sharma, Michael Wray, Mengmeng Xu, Eric Zhongcong Xu, Chen Zhao, Siddhant Bansal, Dhruv Batra, Vincent Cartillier, Sean Crane, Tien Do, Morrie Doulaty, Akshay Erapalli, Christoph Feichtenhofer, Adriano Fragomeni, Qichen Fu, Christian Fuegen, Abrham Gebreselasie, Cristina Gonzalez, James Hillis, Xuhua Huang, Yifei Huang, Wenqi Jia, Weslie Khoo, Jachym Kolar, Satwik Kottur, Anurag Kumar, Federico Landini, Chao Li, Yanghao Li, Zhenqiang Li, Karttikeya Mangalam, Raghava Modhugu, Jonathan Munro, Tullie Murrell, Takumi Nishiyasu, Will Price, Paola Ruiz Puentes, Merey Ramazanova, Leda Sari, Kiran Somasundaram, Audrey Southerland, Yusuke Sugano, Ruijie Tao, Minh Vo, Yuchen Wang, Xindi Wu, Takuma Yagi, Yunyi Zhu, Pablo Arbelaez, David Crandall, Dima Damen, Giovanni Maria Farinella, Bernard Ghanem, Vamsi Krishna Ithapu, C. V. Jawahar, Hanbyul Joo, Kris Kitani, Haizhou Li, Richard Newcombe, Aude Oliva, Hyun Soo Park, James M. Rehg, Yoichi Sato, Jianbo Shi, Mike Zheng Shou, Antonio Torralba, Lorenzo Torresani, Mingfei Yan, and Jitendra Malik. Ego4d: Around the World in 3,000 Hours of Egocentric Video. In IEEE/CVF Computer Vision and Pattern Recognition (CVPR), 2022. 1", + "[29] Shariq Iqbal and Fei Sha. Coordinated exploration via intrinsic rewards for multi-agent reinforcement learning. arXiv preprint arXiv:1905.12127, 2019. 2", + "[30] Max Jaderberg, Wojciech M. Czarnecki, Iain Dunning, Luke Marris, Guy Lever, Antonio Garcia Castañeda, Charlie Beattie, Neil C. Rabinowitz, Ari S. Morcos, Avraham Ruderman, Nicolas Sonnerat, Tim Green, Louise Deason, Joel Z. Leibo, David Silver, Demis Hassabis, Koray Kavukcuoglu, and Thore Graepel. Human-level performance in 3d multiplayer games with population-based reinforcement learning. Science, 364:859 - 865, 2019. 2", + "[31] Unnat Jain, Luca Weihs, Eric Kolve, Ali Farhadi, Svetlana Lazebnik, Aniruddha Kembhavi, and Alexander G. Schwing. A cordial sync: Going beyond marginal policies for multiagent embodied tasks. ArXiv, abs/2007.04979, 2020. 2", + "[32] Unnat Jain, Luca Weihs, Eric Kolve, Mohammad Rastegari, Svetlana Lazebnik, Ali Farhadi, Alexander G. Schwing, and Aniruddha Kembhavi. Two body problem: Collaborative visual task completion. 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6682-6692, 2019. 2", + "[33] Oğuzhan Fatih Kar, Teresa Yeo, Andrei Atanov, and Amir Zamir. 3d common corruptions and data augmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18963-18974, 2022. 6", + "[34] Kapil Katyal, Katie Popek, Chris Paxton, Phil Burlina, and Gregory D. Hager. Uncertainty-aware occupancy map prediction using generative networks for robot navigation. In 2019 International Conference on Robotics and Automation (ICRA), pages 5453-5459, 2019. 2", + "[35] Hansung Kim, Luca Remaggi, Philip JB Jackson, Filippo Maria Fazi, and Adrian Hilton. 3d room geometry reconstruction using audio-visual sensors. In 2017 International Conference on 3D Vision (3DV), pages 621–629, 2017. 2", + "[36] Bruno Korbar, Du Tran, and Lorenzo Torresani. Scsampler:" + ], + "bbox": [ + 78, + 90, + 472, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Sampling salient clips from video for efficient action recognition. 2019 IEEE/CVF International Conference on Computer Vision (ICCV), pages 6231-6241, 2019. 3", + "[37] Robert LiKamWa, Zhen Wang, Aaron Carroll, Felix Xiaozhu Lin, and Lin Zhong. Draining our glass: An energy and heat characterization of google glass. In Proceedings of 5th Asia-Pacific Workshop on Systems, pages 1-7, 2014. 2, 3", + "[38] Jintao Lin, Haodong Duan, Kai Chen, Dahua Lin, and Limin Wang. Ocsampler: Compressing videos to one clip with single-step sampling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13894-13903, 2022. 3", + "[39] David B. Lindell, Gordon Wetzstein, and Vladlen Koltun. Acoustic non-line-of-sight imaging. In 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6773-6782, 2019. 1, 2", + "[40] Chen Liu, Jiaye Wu, and Yasutaka Furukawa. Floornet: A unified framework for floorplan reconstruction from 3d scans. In Proceedings of the European conference on computer vision (ECCV), pages 201-217, 2018. 1, 2", + "[41] Wenxin Liu, David Caruso, Eddy Ig, Jing Dong, Anastasios I Mourikis, Kostas Daniilidis, Vijay Kumar, and Jakob Engel. Tlio: Tight learned inertial odometry. IEEE Robotics and Automation Letters, 5(4):5653-5660, 2020. 6", + "[42] Sagnik Majumder, Ziad Al-Halah, and Kristen Grauman. Move2hear: Active audio-visual source separation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 275–285, 2021. 2", + "[43] Sagnik Majumder, Ziad Al-Halah, and Kristen Grauman. Active audio-visual separation of dynamic sound sources. In European Conference on Computer Vision. Springer, 2022. 2", + "[44] Sagnik Majumder, Changan Chen, Ziad Al-Halah, and Kristen Grauman. Few-shot audio-visual learning of environment acoustics. In Thirty-Sixth Conference on Neural Information Processing Systems, 2022. 2, 5", + "[45] Medhini Narasimhan, Erik Wijmans, Xinlei Chen, Trevor Darrell, Dhruv Batra, Devi Parikh, and Amanpreet Singh. Seeing the un-scene: Learning amodal semantic maps for room navigation. In European Conference on Computer Vision, pages 513-529. Springer, 2020. 2", + "[46] Simon T O'Callaghan and Fabio T Ramos. Gaussian process occupancy maps*. Int. J. Rob. Res., 31(1):42-62, jan 2012. 2", + "[47] Brian Okorn, Xuehan Xiong, and Burcu Akinci. Toward automated modeling of floor plans. In 3D PVT, 2009. 1, 2", + "[48] Andrew Owens, Phillip Isola, Josh McDermott, Antonio Torralba, Edward H. Adelson, and William T. Freeman. Visually indicated sounds. In 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 2405-2413, 2016. 2", + "[49] Vassil Panayotov, Guoguo Chen, Daniel Povey, and Sanjeev Khudanpur. Librispeech: An asr corpus based on public domain audio books. In 2015 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 5206-5210, 2015. 6", + "[50] Rameswar Panda, Chun-Fu Richard Chen, Quanfu Fan, Ximeng Sun, Kate Saenko, Aude Oliva, and Rogerio Feris. Adamml: Adaptive multi-modal learning for efficient video recognition. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7576-7585, 2021. 3" + ], + "bbox": [ + 501, + 92, + 893, + 891 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10563", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[51] Shivansh Patel, Saim Wani, Unnat Jain, Alexander G. Schwing, Svetlana Lazebnik, Manolis Savva, and Angel X. Chang. Interpretation of emergent communication in heterogeneous collaborative embodied agents. 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 15993-15943, 2021. 2", + "[52] Senthil Purushwalkam, Sebastia Vicenc Amengual Gari, Vamsi Krishna Ithapu, Carl Schissler, Philip Robinson, Abhinav Gupta, and Kristen Grauman. Audio-visual floorplan reconstruction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1183–1192, 2021. 1, 2, 3, 6, 7", + "[53] Santhosh K Ramakrishnan, Ziad Al-Halah, and Kristen Grauman. Occupancy anticipation for efficient exploration and navigation. In European Conference on Computer Vision, pages 400-418. Springer, 2020. 1, 2, 6, 7", + "[54] Fabio Ramos and Lionel Ott. Hilbert maps: Scalable continuous occupancy mapping with stochastic gradient descent. The International Journal of Robotics Research, 35(14):1717-1730, 2016. 2", + "[55] Ethan Rublee, Vincent Rabaud, Kurt Konolige, and Gary Bradski. Orb: An efficient alternative to sift or surf. In 2011 International Conference on Computer Vision, pages 2564-2571, 2011. 6", + "[56] João Machado Santos, David Portugal, and Rui P. Rocha. An evaluation of 2d slam techniques available in robot operating system. In 2013 IEEE International Symposium on Safety, Security, and Rescue Robotics (SSRR), pages 1-6, 2013. 2", + "[57] Manolis Savva, Abhishek Kadian, Oleksandr Maksymets, Yili Zhao, Erik Wijmans, Bhavana Jain, Julian Straub, Jia Liu, Vladlen Koltun, Jitendra Malik, et al. Habitat: A platform for embodied ai research. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9339-9347, 2019. 6", + "[58] Carl Schissler, Christian Loftin, and Dinesh Manocha. Acoustic classification and optimization for multi-modal rendering of real-world scenes. IEEE Transactions on Visualization and Computer Graphics, 24:1246-1259, 2018. 2", + "[59] Ransalu Senanayake, Thushan Ganegedara, and Fabio Ramos. Deep occupancy maps: a continuous mapping technique for dynamic environments. In NIPS 2017 Workshop MLITS, 2017.", + "[60] Rakesh Shrestha, Fei-Peng Tian, Wei Feng, Ping Tan, and Richard Vaughan. Learned map prediction for enhanced mobile robot exploration. In 2019 International Conference on Robotics and Automation (ICRA), pages 1197-1204, 2019. 2", + "[61] Jascha Sohl-Dickstein, Santani Teng, Benjamin M. Gaub, Chris C. Rodgers, Crystal Li, Michael R. DeWeese, and Nicol S. Harper. A device for human ultrasonic echolocation. IEEE Transactions on Biomedical Engineering, 62(6):1526-1534, 2015. 1, 2", + "[62] Wei Sui, Lingfeng Wang, Bin Fan, Hongfei Xiao, Huaiyu Wu, and Chunhong Pan. Layer-wise floorplan extraction for automatic urban building reconstruction. IEEE Transactions on Visualization and Computer Graphics, 22(3):1261-1277, 2016. 2", + "[63] Maitreya Suin and A. N. Rajagopalan. An efficient framework for dense video captioning. Proceedings of the AAAI" + ], + "bbox": [ + 78, + 90, + 470, + 893 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Conference on Artificial Intelligence, 34(07):12039-12046, Apr. 2020. 3", + "[64] Cheng Sun, Chi-Wei Hsiao, Min Sun, and Hwann-Tzong Chen. Horizonnet: Learning room layout with 1d representation and piano stretch data augmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1047–1056, 2019. 2", + "[65] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 5", + "[66] Erik Wijmans, Abhishek Kadian, Ari S. Morcos, Stefan Lee, Irfan Essa, Devi Parikh, Manolis Savva, and Dhruv Batra. Dd-ppo: Learning near-perfect pointgoal navigators from 2.5 billion frames. In ICLR, 2020. 6", + "[67] Wenming Wu, Xiao-Ming Fu, Rui Tang, Yuhan Wang, YuHao Qi, and Ligang Liu. Data-driven interior plan generation for residential buildings. ACM Trans. Graph., 38(6), nov 2019. 2", + "[68] Zuxuan Wu, Caiming Xiong, Chih-Yao Ma, Richard Socher, and Larry S Davis. Adaframe: Adaptive frame selection for fast video recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1278-1287, 2019. 3", + "[69] Shang-Ta Yang, Fu-En Wang, Chi-Han Peng, Peter Wonka, Min Sun, and Hung-Kuo Chu. Dula-net: A dual-projection network for estimating room layouts from a single rgb panorama. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3363–3372, 2019. 2", + "[70] Mao Ye, Yu Zhang, Ruigang Yang, and Dinesh Manocha. 3d reconstruction in the presence of glasses by acoustic and stereo fusion. In 2015 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 4885-4893, 2015. 2", + "[71] Serena Yeung, Olga Russakovsky, Greg Mori, and Li Fei-Fei. End-to-end learning of action detection from frame glimpses in videos. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2678–2687, 2016. 3", + "[72] Yinfeng Yu, Wenbing Huang, Fuchun Sun, Changan Chen, Yikai Wang, and Xiaohong Liu. Sound adversarial audiovisual navigation. In International Conference on Learning Representations, 2022. 2", + "[73] Zhoutong Zhang, Jiajun Wu, Qiujia Li, Zhengjia Huang, James Traer, Josh H. McDermott, Joshua B. Tenenbaum, and William T. Freeman. Generative modeling of audible shapes for object perception. 2017 IEEE International Conference on Computer Vision (ICCV), pages 1260-1269, 2017. 2", + "[74] Chuhang Zou, Alex Colburn, Qi Shan, and Derek Hoiem. Layoutnet: Reconstructing the 3d room layout from a single rgb image. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2051-2059, 2018. 2", + "[75] Chuhang Zou, Jheng-Wei Su, Chi-Han Peng, Alex Colburn, Qi Shan, Peter Wonka, Hung-Kuo Chu, and Derek Hoiem. Manhattan room layout reconstruction from a single $360^{\\circ}$ image: A comparative study of state-of-the-art methods. International Journal of Computer Vision, 129(5):1410-1431, 2021. 2" + ], + "bbox": [ + 503, + 92, + 893, + 878 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "10564", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2023/Chat2Map_ Efficient Scene Mapping From Multi-Ego Conversations/9562f06a-6203-4663-996c-df2336adbe79_model.json b/2023/Chat2Map_ Efficient Scene Mapping From Multi-Ego Conversations/9562f06a-6203-4663-996c-df2336adbe79_model.json new file mode 100644 index 0000000000000000000000000000000000000000..54ee11152db877999c4b7f9816ba20bbfbde2e02 --- /dev/null +++ b/2023/Chat2Map_ Efficient Scene Mapping From Multi-Ego Conversations/9562f06a-6203-4663-996c-df2336adbe79_model.json @@ -0,0 +1,2466 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.143, + 0.131, + 0.83, + 0.154 + ], + "angle": 0, + "content": "Chat2Map: Efficient Scene Mapping from Multi-Ego Conversations" + }, + { + "type": "text", + "bbox": [ + 0.186, + 0.18, + 0.784, + 0.234 + ], + "angle": 0, + "content": "Sagnik Majumder\\(^{1,2,3}\\) Hao Jiang\\(^{2}\\) Pierre Moulon\\(^{2}\\) Ethan Henderson\\(^{2}\\) Paul Calamia\\(^{2}\\) Kristen Grauman\\(^{1,3*}\\) Vamsi Krishna Ithapu\\(^{2*}\\) UT Austin\\(^{2}\\) Reality Labs Research, Meta FAIR" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.269, + 0.314, + 0.285 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.302, + 0.474, + 0.633 + ], + "angle": 0, + "content": "Can conversational videos captured from multiple egocentric viewpoints reveal the map of a scene in a cost-efficient way? We seek to answer this question by proposing a new problem: efficiently building the map of a previously unseen 3D environment by exploiting shared information in the egocentric audio-visual observations of participants in a natural conversation. Our hypothesis is that as multiple people (\"egos\") move in a scene and talk among themselves, they receive rich audio-visual cues that can help uncover the unseen areas of the scene. Given the high cost of continuously processing egocentric visual streams, we further explore how to actively coordinate the sampling of visual information, so as to minimize redundancy and reduce power use. To that end, we present an audio-visual deep reinforcement learning approach that works with our shared scene mapper to selectively turn on the camera to efficiently chart out the space. We evaluate the approach using a state-of-the-art audio-visual simulator for 3D scenes as well as real-world video. Our model outperforms previous state-of-the-art mapping methods, and achieves an excellent cost-accuracy tradeoff. Project: http://vision.cs.utexas.edu/projects/chat2map." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.655, + 0.21, + 0.67 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.68, + 0.472, + 0.846 + ], + "angle": 0, + "content": "The spatial layout of the environment around us is fundamental to understanding our physical context. By representing the walls, furniture, and other major structures in a space, scene maps ground activity and objects in a persistent frame of reference, facilitating high-level reasoning for many downstream applications in augmented reality (AR) and robotics. For example, episodic memory [16, 28] aims to relocalize lost objects observed in first-person video (where are my keys?); floorplan estimation [10, 40, 47] aims to chart out the area and shapes of complex buildings; navigating agents try to discover routes in unfamiliar spaces [4, 11, 53]." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.847, + 0.471, + 0.877 + ], + "angle": 0, + "content": "While traditional computer vision approaches for mapping (e.g., visual SLAM) are highly effective when extensive" + }, + { + "type": "image", + "bbox": [ + 0.504, + 0.267, + 0.895, + 0.512 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.522, + 0.895, + 0.592 + ], + "angle": 0, + "content": "Figure 1. Given egocentric audio-visual observations from multiple people wearing AR glasses and moving and conversing (left), we aim to accurately map the scene (right). To mitigate cost, our model receives audio continuously but learns to selectively employ the ego cameras only when the visual input is expected to be informative." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.61, + 0.895, + 0.732 + ], + "angle": 0, + "content": "exposure to the environment is possible, in many real-world scenarios only a fraction of the space is observed by the camera. Recent work shows the promise of sensing 3D spaces with both sight and sound [8, 13, 24, 26, 52]: listening to echoes bounce around the room can reveal the depth and shape of surrounding surfaces, and even help extrapolate a floorplan beyond the camera's field of view or behind occluded objects [52]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.733, + 0.894, + 0.854 + ], + "angle": 0, + "content": "While we are inspired by these advances, they also have certain limitations. Often systems will emit sounds (e.g., a frequency sweep) into the environment to ping for spatial information [1, 13, 14, 22, 26, 39, 52, 61], which is intrusive if done around people. Furthermore, existing audio-visual models assume that the camera is always on grabbing new frames, which is wasteful if not intractable, particularly on lightweight, low-power computing devices in AR settings." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.856, + 0.895, + 0.901 + ], + "angle": 0, + "content": "We introduce Chat2Map, a new scene mapping task aimed at eliminating these challenges. In the proposed setting, multiple people converse as they move casually through" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.887, + 0.204, + 0.9 + ], + "angle": 0, + "content": "*Equal contribution" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "10554" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.471, + 0.168 + ], + "angle": 0, + "content": "the scene while wearing AR glasses equipped with an egocentric camera, microphones, and potentially other sensors (e.g., for odometry).1 Given their egocentric audio-visual data streams, the goal is to infer the ground-plane occupancy map for the larger environment around them. See Figure 1." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.168, + 0.472, + 0.272 + ], + "angle": 0, + "content": "We observe that audio-visual data from the egos' interactions will naturally reflect scene structure. First, as they walk and talk, their movements reveal spaces like corridors, doorways, and large rooms, in both modalities. Second, the speech captured by the device-wearer's cameras and microphones can be localized to different speakers, which, compared to active sound emission, is non-intrusive." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.273, + 0.472, + 0.56 + ], + "angle": 0, + "content": "To realize this vision, we develop a novel approach to efficient scene mapping from multi-ego conversations. Our approach has two key elements: a shared scene mapper and a visual sampling policy. For the former, we devise a transformer-based mapper that incorporates the multiple data streams to infer a map beyond the directly observed areas, and, most importantly, that enables communication among the egos about their observations and states in the 3D space to improve mapping accuracy. For the latter, our idea is to relax the common assumption of an \"always-on\" camera, and instead actively select when to sample visual frames from any one of the ego cameras. Intuitively, certain regions where the egos move will be more or less important for mapping (e.g., corners of the room, doors). We train a sampling policy with deep reinforcement learning that activates the visual feed only when it is anticipated to complement the continuous audio feed. This is a cost-conscious approach, mindful that switching on a camera is much more power consuming than sensing audio with microphones [2, 37]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.561, + 0.472, + 0.68 + ], + "angle": 0, + "content": "We demonstrate our approach using a state-of-the-art audio-visual simulator for 3D scenes [8] as well as real-world video input. We can successfully map an unfamiliar environment given only partial visibility via multiple conversing people moving about the scene. Compared to sampling all visual frames, our model reduces the visual capture and processing by \\(87.5\\%\\) while the mapping accuracy declines only marginally \\((\\sim 9\\%)\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.681, + 0.472, + 0.802 + ], + "angle": 0, + "content": "Our main contributions are 1) we define the task of efficient and shared scene mapping from multi-ego conversations, a new direction for AR/VR research; 2) we present the first approach to begin tackling this task, specifically an RL-based framework that integrates an audio-visual scene mapper and a smart visual sampling policy; and 3) we rigorously experiment with a variety of environments (both in simulation and the real world), speech sounds, and use cases." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.813, + 0.22, + 0.829 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.838, + 0.47, + 0.869 + ], + "angle": 0, + "content": "Visual scene mapping. Past works tackle scene mapping using 3D Manhattan layouts [18, 64, 69, 74, 75], detailed" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.092, + 0.895, + 0.394 + ], + "angle": 0, + "content": "floorplans [10, 40, 47, 62, 67], occupancy [21, 34, 46, 54, 59, 60], and semantic maps [45]. Manhattan layouts include structured outputs like scene boundaries [64, 74, 75], corners [74, 75], and floor/ceilings [69, 75], but do not generalize to unseen environment regions. Floorplan estimation methods use dense scans of 3D scenes to predict geometric (walls, exterior/ interior) and semantic layouts (room type, object type, etc.)., rely on extensive human walkthroughs with RGB-D [10, 40] or 3D point cloud [47, 62] scans, and are usually limited to polygonal layouts [10, 40, 47, 62, 67]. Occupancy maps traditionally rely on wide field-of-view (FoV) LiDAR scanners [56] or evaluate on simple 2D environments without non-wall obstacles [21, 34, 60, 60]. More recent methods [4, 5, 11, 53] train an embodied agent to explore and build top-down maps of more complex scenes using RGB-D. On the contrary, our method uses both vision and audio from the observations of a group of conversing people for mapping. Rather than steer the camera of a robot to map the scene, our task requires processing passive video from human camera wearers." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.402, + 0.895, + 0.538 + ], + "angle": 0, + "content": "Audio-visual scene mapping. To our knowledge, the only prior work to translate audio-visual inputs into a general (arbitrarily shaped) floorplan maps is AV-Floorplan [52]. Unlike AV-Floorplan, our method infers maps from speech in natural human conversations, which avoids emitting intrusive frequency sweep signals to generate echoes. In addition, a key goal of our work is to reduce mapping cost by skipping redundant visual frames. Our experiments demonstrate the benefits of our model design over AV-Floorplan [52]." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.546, + 0.897, + 0.758 + ], + "angle": 0, + "content": "Audio(-visual) spatial understanding. More broadly, beyond the mapping task, various methods leverage audio for geometric and material information about the 3D scene and its constituent objects. Prior work relies on acoustic reflections to estimate the shape of objects [39]. Echolocation is used in robotics to estimate proximity to surrounding surfaces [1, 14, 22, 61]. Together, vision and audio can better reveal the shape and materials of objects [48, 58, 73], self-supervise imagery [26], and improve depth sensing [35, 70]. Recent work exploits correlations between spatial audio and imagery to reason about scene acoustics [7, 44] or aid active embodied navigation [6, 9, 17, 25, 72] and source separation [42, 43]. No prior work intelligently selects images to capture during conversations to efficiently map a scene." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.765, + 0.897, + 0.903 + ], + "angle": 0, + "content": "Multi-agent spatial understanding. There is existing work [15, 29, 31, 32, 51] in the visual multi-agent reinforcement learning (MARL) community that learns collaborative agents for performing tasks like relocating furniture [31, 32], playing 3D multi-player games [30], coordinated scene exploration [29], or multi-object navigation [51]. In such settings, the collaborative agents actively interact with the environment to learn a shared scene representation for successfully completing their task. In contrast, we aim to learn a" + }, + { + "type": "page_footnote", + "bbox": [ + 0.075, + 0.875, + 0.47, + 0.901 + ], + "angle": 0, + "content": "1 Throughout, we call each person participating in the conversation an \"ego\" for short." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "10555" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.471, + 0.137 + ], + "angle": 0, + "content": "shared geometric map of a 3D scene given passive observations that come from the trajectories chosen by a group of people involved in a natural conversation." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.146, + 0.472, + 0.312 + ], + "angle": 0, + "content": "Efficient visual sampling in video. Efficient visual sampling has been studied in the context of video recognition [27, 36, 38, 68, 71] and summarization [12, 63] with the goal of selectively processing informative frames, which can both reduce computational cost and improve recognition. More closely related to our approach are methods that use audio for the decision-making [27, 36, 50]. Different from the above, we use efficient visual sampling in the context of mapping scenes. Furthermore, in our case an online sampling decision needs to be made at every step before looking at the current visual frame (or frames from future steps)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.328, + 0.353, + 0.344 + ], + "angle": 0, + "content": "3. Chat2Map Task Formulation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.354, + 0.471, + 0.383 + ], + "angle": 0, + "content": "We propose a novel task: efficient and shared mapping of scenes from multi-ego conversations." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.384, + 0.472, + 0.643 + ], + "angle": 0, + "content": "Without loss of generality, we consider two egos, \\( E_{1} \\) and \\( E_{2} \\), each wearing AR glasses equipped with an RGB-D camera and a multi-channel microphone array. The egos have a conversation and move around in an unmapped 3D environment. Each conversation is \\( T \\) steps long. At each step \\( t \\), the ego \\( E_{i} \\)'s glasses receive an observation \\( \\mathcal{O}_{i,t} = (\\mathcal{V}_{i,t}, S_{i,t}, P_{i,t}, S_{i,t}', P_{i,t}') \\). \\( \\mathcal{V}_{i,t} \\) is the \\( 90^{\\circ} \\) FOV RGB-D image and \\( S_{i,t} \\) is the speech waveform uttered by \\( E_{i} \\), as observed from its pose \\( P_{i,t} = (x_{i,t}, y_{i,t}, \\theta_{i,t}) \\), where \\( (x_{i,t}, y_{i,t}) \\) denotes its location and \\( \\theta_{i,t} \\) denotes its orientation in the 3D scene. \\( S_{i,t}' \\) is the speech of the other ego \\( E_{i}' \\) (the other person involved in the conversation), as perceived by \\( E_{i} \\) (note, the speech sounds different depending on the listener position), and \\( P_{i,t}' \\) is \\( E_{i}' \\)'s pose relative to \\( E_{i} \\). Modern AR glasses, like Bose Frames or Facebook Aria already support capturing such multi-sensory observations, making it possible to have a real-world instantiation of our task." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.644, + 0.472, + 0.884 + ], + "angle": 0, + "content": "Given the real-time observation stream \\(\\mathcal{O}\\) for the egos, where \\(\\mathcal{O} = \\{\\mathcal{O}_{i,t}:i = 1,\\dots ,2,t = 1\\dots T\\}\\) and a total budget of visual frames \\(B\\) we aim to learn a model that can accurately estimate the top-down occupancy map \\(M\\) of the scene without exceeding the visual budget. The fixed budget represents allotting a fixed amount of battery to the energy-intensive visual capture [37], so that other on-device routines are not affected. We assume the first visual frames (at \\(t = 1\\) ) for both egos to be used by the model. Thus we aim to learn a policy that samples \\(B\\) frames from \\(2*(T - 1)\\) choices—which are not considered a batch, but rather unfold in sequence—and a mapper that predicts the scene map given the sampled frames. Our goal is to build a model that samples the expensive visual frames only when absolutely needed for scene mapping. This is captured by the constraint \\(1\\leq B\\ll 2*(T - 1)\\)." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.886, + 0.471, + 0.901 + ], + "angle": 0, + "content": "There are three important aspects to our task. First, it" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.895, + 0.364 + ], + "angle": 0, + "content": "requires learning from both vision and audio. While the visual signal carries rich information about the local scene geometry, there can be a high amount of redundancy in the visual feed captured during a conversation (e.g., the egos may visit the same location more than once or change their viewpoint only marginally). Second, not only does the long-range nature of audio help uncover the global scene properties [19, 52] like shape and size—beyond what is visible in images—but also we can exploit audio to undersample the visual frames, thereby reducing the cost of capturing and processing sensory inputs for mapping. Third, shared mapping of a scene allows jointly leveraging the complementary information in the audio (speech) from the self and other egos, and the synergy of the audio-visual cues from multiple egos. These insights form the basis of our key hypothesis in this task—selectively sampling visual frames during a conversation involving egos that share information with each other can facilitate efficient mapping of a scene." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.377, + 0.608, + 0.394 + ], + "angle": 0, + "content": "4. Approach" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.403, + 0.895, + 0.538 + ], + "angle": 0, + "content": "We solve the task by learning a model that estimates the scene map given the egos' audio-visual observations and also sequentially decides when to sample visual frames for mapping given the audio stream, ego poses, and previously sampled frames, if any. Here, \"sampling\" refers to individually deciding for each ego whether to use its camera or not to capture the visuals at every step of its trajectory in the scene. The sampling is preemptive in nature, i.e. the policy selects or skips a frame without capturing it first." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.539, + 0.895, + 0.719 + ], + "angle": 0, + "content": "Our model has two main components (see Fig. 2): (1) a shared scene mapper, and (2) a visual sampling policy. At every step \\( t \\), the shared mapper has two functions. First, it estimates the map of a previously unseen environment by exploiting the shared spatial cues in the audio-visual observations of the two egos. Second, it informs the policy about the utility of sampling a certain visual frame. Guided by the mapper, the policy samples only the most informative visual frames that are expected to boost mapping significantly over using just audio. Note that, unlike the visuals, we observe audio continuously as it is less resource-intensive vis-a-vis storage and power requirements for processing [2]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.892, + 0.78 + ], + "angle": 0, + "content": "We learn our task through the synergy of the mapper and the policy, such that under the constraint of a limited visual budget \\( B \\), our model implicitly understands which visual frames are critical for mapping." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.781, + 0.895, + 0.901 + ], + "angle": 0, + "content": "First, we describe the steps involved to prepare our model inputs (Sec. 4.1). Next, we introduce our visual sampling policy (Sec. 4.2) and shared scene mapper (Sec. 4.3). Finally, we present model training details (Sec. 4.4). Through the rest of the text, we use separate notations to distinguish the egos' observations \\(\\mathcal{O}\\) (i.e. what the egos receive from the environment) from our model inputs \\(O\\) (i.e. what we capture and feed to our model for efficient mapping)." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "10556" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.139, + 0.089, + 0.481, + 0.39 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.489, + 0.089, + 0.831, + 0.39 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.392, + 0.895, + 0.462 + ], + "angle": 0, + "content": "Figure 2. Our model has two main components: a) a visual sampling policy (left), and b) a shared scene mapper (right). At each step, our policy receives the current audio along with the previous audio(-visual) observations for the egos and decides for each ego individually whether to capture its visual frame at the current step. As per the policy predictions, the shared mapper conditionally uses the current visual frame(s) and audio along with the past audio(-visual) observations to predict the occupancy map of the scene, a ground-plane map showing where obstacles and freespace are (shown in green and white)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.474, + 0.304, + 0.49 + ], + "angle": 0, + "content": "4.1. Model input preparation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.499, + 0.471, + 0.71 + ], + "angle": 0, + "content": "We prepare our model inputs by separately preprocessing the visual and audio modalities. If our policy decides to sample an image \\(\\mathcal{V}\\), we transform it into \\(V = (V^{R}, V^{M})\\). \\(V^{R}\\) denotes the normalized RGB image with pixel values \\(\\in [0,1]\\). \\(V^{M}\\) denotes the \\(90^{\\circ}\\) FoV top-down occupancy map created by projecting the depth image to the ground plane. To do the depth projection, we first backproject it into the world coordinates using the camera's intrinsic parameters to compute the local visible scene's 3D point cloud. Next, we project these points to obtain a two-channel binary top-down map of size \\(h \\times w \\times 2\\), where the first channel reveals occupied/free areas, and the second channel reveals seen/unseen areas. If our policy skips \\(\\mathcal{V}\\), we set \\(V^{R}\\) and \\(V^{M}\\) to all-zero matrices of the appropriate size." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.711, + 0.471, + 0.803 + ], + "angle": 0, + "content": "For a speech waveform \\( S \\), we calculate the short-time Fourier transform (STFT) magnitude spectrogram denoted by \\( S \\) of size \\( F \\times T \\times C \\), where \\( F \\), \\( T \\), and \\( C \\) are the number of frequency bins, time windows, and ambisonic microphone channels, respectively. Lastly, we normalize each pose \\( P_{i,t} \\) to be relative to \\( P_{1,1} \\). See Sec. 5 and Supp. for more details." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.816, + 0.287, + 0.833 + ], + "angle": 0, + "content": "4.2. Visual sampling policy" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.471, + 0.902 + ], + "angle": 0, + "content": "At every step \\( t \\), our visual sampling policy \\( \\pi^V \\) (Fig. 2 left) receives \\( O^\\pi(t) \\) as input and makes the decision to either capture or skip the visual frame \\( \\mathcal{V}_{i,t} \\) for each ego \\( E_i \\). \\( O^\\pi(t) \\) comprises the visual cue from the last step along with the" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.474, + 0.892, + 0.616 + ], + "angle": 0, + "content": "speech cues, and the poses from the current step and the last step for both egos. Formally, \\( O^{\\pi}(t) = \\{O_{i}^{\\pi}(t):i = 1\\ldots 2\\} \\), where \\( O_{i}^{\\pi}(t) = \\{V_{i,t - 1},S_{i,j},P_{i,j},S_{i,j}^{\\prime},P_{i,j}^{\\prime}:j = t - 1\\ldots t\\} \\). The policy first uses an encoder network to generate a multi-modal embedding of \\( O^{\\pi}(t) \\), and then passes the embedding to a policy network that makes a sampling decision per ego. At \\( t = 1 \\), as per our problem definition (Sec. 3), the policy always chooses to sample the visual frames for both egos, i.e., the cameras are initially on." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.621, + 0.893, + 0.854 + ], + "angle": 0, + "content": "Multi-modal policy embedding. To process ego \\( E_{i} \\)'s visual input \\( V_{i,t-1} \\) from the last step, we encode the RGB image \\( V_{i,t-1}^{R} \\) and map \\( V_{i,t-1}^{M} \\) with separate CNNs. We then concatenate the two features to generate the visual embedding \\( v_{i,t-1} \\). To encode the pose inputs \\( \\{P_{i,t-1}, P_{i,t-1}', P_{i,t}, P_{i,t}'\\} \\), we use a linear layer and generate pose embeddings \\( \\{p_{i,t-1}, p_{i,t-1}', p_{i,t}, p_{i,t}'\\} \\). We process the speech inputs \\( \\{S_{i,t-1}, S_{i,t-1}', S_{i,t}, S_{i,t}'\\} \\) using another CNN and create speech embeddings \\( \\{s_{i,t-1}, s_{i,t-1}', s_{i,t}, s_{i,t}'\\} \\). Next, we fuse the visual, speech and pose embeddings using linear layers (see Fig. 2 left) to obtain the multi-modal policy embedding \\( e_{i,t} \\) for \\( E_{i} \\). Finally, we fuse the policy embeddings for the two egos, \\( e_{1,t} \\) and \\( e_{2,t} \\) with a linear layer to produce the multi-modal policy embedding \\( e_t \\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.894, + 0.902 + ], + "angle": 0, + "content": "The visual, audio, and pose inputs carry complementary cues required for efficient visual sampling. Whereas the pose inputs from the last and current steps explicitly reveal" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "10557" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.244 + ], + "angle": 0, + "content": "the viewpoint change between the steps, the previous and current speech inputs provide information about the changes in the local and global scene structures as a function of the previously sampled visual inputs, which together suggest the value of sampling a visual frame at the current step. Furthermore, guided by our training reward (below in Sec. 4.4), the previously observed visual frames and audio together enable our policy to anticipate the current frames and skip them if they are deemed redundant, thereby improving mapping accuracy for a low visual budget." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.25, + 0.473, + 0.388 + ], + "angle": 0, + "content": "Policy network. The policy network consists of a GRU that estimates an updated history \\( h_t \\) along with the current state representation \\( g_t \\), using the fused embedding \\( e_t \\) and the history of states \\( h_{t-1} \\). An actor-critic module takes \\( g_t \\) and \\( h_{t-1} \\) as inputs and predicts a policy distribution \\( \\pi_\\theta(a_{i,t}|g_t, h_{t-1}) \\) per ego along with the value of the state \\( H_\\theta(g_t, h_{t-1}) \\) (\\( \\theta \\) are policy parameters). The policy samples an action \\( a_{i,t} \\in \\{0,1\\} \\) for every \\( E_i \\). \\( a_{i,t} = 1 \\) corresponds to selecting \\( \\mathcal{V}_{i,t} \\), whereas \\( a_{i,t} = 0 \\) means to skip that frame." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.397, + 0.279, + 0.414 + ], + "angle": 0, + "content": "4.3. Shared scene mapper" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.42, + 0.473, + 0.56 + ], + "angle": 0, + "content": "Whereas \\( O^{\\pi}(t) \\) denotes our policy input (Sec. 4.2), \\( O^{M}(t) \\) denotes the input to our shared scene mapper \\( f^{M} \\) at step \\( t \\), such that \\( O^{M}(t) = \\left\\{(V_{i,j}, S_{i,j}, S_{i,j}^{\\prime}, P_{i,j}, P_{i,j}^{\\prime}): i = 1 \\dots 2, j = 1 \\dots t\\right\\} \\). \\( f^{m} \\) starts by embedding each component of \\( O^{M}(t) \\) using a separate network. This is followed by a multi-modal memory that stores the embeddings since the start of the episode. Finally, a transformer [65] predicts an estimate \\( \\tilde{M}(t) \\) of the scene map conditioned on the multi-modal memory and the egos' poses in the episode." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.566, + 0.473, + 0.645 + ], + "angle": 0, + "content": "Multi-modal mapper embedding. For the visual input \\( V_{i,j} \\), we encode \\( V_{i,j}^{R} \\) and \\( V_{i,j}^{M} \\) using separate CNNs and do a channel-wise concatenation to get visual features \\( \\hat{v}_{i,j} \\). Similarly, speech is encoded using separate CNNs to get \\( \\hat{s}_{i,j} \\) and \\( \\hat{s}_{i,j}' \\). Each of \\( \\hat{v} \\), \\( \\hat{s} \\) and \\( \\hat{s}' \\) is of size \\( 4 \\times 4 \\times 1024 \\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.645, + 0.473, + 0.857 + ], + "angle": 0, + "content": "For both vision and speech, we compute two positional embeddings, \\( p^{\\mathrm{I}} \\) and \\( p^{\\mathrm{II}} \\). They encode the pose of the egos in the 3D space, and the index of each 1024-dimensional feature in the visual or speech features in the raster order respectively. Whereas \\( p^{\\mathrm{I}} \\) helps discover spatial cues as a function of the egos' location in the 3D scene, \\( p^{\\mathrm{II}} \\) enables our model to attend to different modalities in a more fine-grained manner. For both, we compute an 8-dimensional sinusoidal positional encoding [65] and then pass it through a linear layer to obtain a 1024-dimensional embedding. For \\( p^{\\mathrm{II}} \\), we additionally repeat this process for every feature index in the raster order. Lastly, we reshape \\( p^{\\mathrm{I}} \\) and add it with \\( p^{\\mathrm{II}} \\) to produce \\( 4 \\times 4 \\times 1024 \\)-dimensional positional embeddings, \\( \\hat{p}_{i,j} \\) for \\( \\hat{v}_{i,j} \\) and \\( \\hat{s}_{i,j} \\), and \\( \\hat{p}_{i,j}' \\) for \\( \\hat{s}_{i,j}' \\)." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.857, + 0.472, + 0.902 + ], + "angle": 0, + "content": "Following [44], we also learn an embedding \\(\\hat{m}_{i,j} \\in \\{\\hat{m}_V, \\hat{m}_S, \\hat{m}_{S'}\\}\\) to capture different modality types, where \\(\\hat{m}_V\\) represents vision, and \\(\\hat{m}_S\\) and \\(\\hat{m}_{S'}\\) represent the speech" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.152 + ], + "angle": 0, + "content": "from self and that of the other ego, respectively. The modality-based embeddings help our model differentiate between different modalities and better map the scene by learning complementary spatial cues from them." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.156, + 0.895, + 0.265 + ], + "angle": 0, + "content": "Multi-modal memory. For the visual input \\( V_{i,j} \\), we add its embedding \\( \\hat{v}_{i,j} \\) with its positional embedding \\( \\hat{p}_{i,j} \\) and modality embedding \\( \\hat{m}_{i,j}^{V} \\), and flatten the sum to get a \\( 16 \\times 1024 \\)-dimensional embedding. Similarly, we fuse the speech embeddings by taking their sum and flattening them. This generates a multi-modal memory of fused embeddings \\( o \\), such that \\( o = \\{o_{1,1}^{V}, \\ldots, o_{2,t}^{V}, o_{1,1}^{S}, \\ldots, o_{2,t}^{S}, o_{1,1}^{S'}, \\ldots, o_{2,t}^{S'}\\} \\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.267, + 0.895, + 0.357 + ], + "angle": 0, + "content": "Occupancy prediction. To predict the underlying scene occupancy, we first use a transformer encoder [65] to attend to the embeddings in \\(o\\) and capture short- and long-range correlations within and across modalities using a stack of self-attention layers. This generates an audio-visual representation that models the spatial layout of the 3D scene." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.358, + 0.895, + 0.587 + ], + "angle": 0, + "content": "Next, we use a transformer decoder [65] to perform cross-attention on the audio-visual representation of the scene conditioned on the embedding \\(\\hat{p}_{i,j}\\) for every pose \\(P_{i,j}\\) in \\(O^M (t)\\) and generate an embedding \\(d_{i,j}\\) for the pose. Finally, we upsample \\(d_{i,j}\\) using a multi-layer network \\(U\\) comprising transpose convolutions and a sigmoid layer at the end to predict an estimate \\(\\tilde{M}_{i,j}\\) of the ground-truth local \\(360^{\\circ}\\) FoV map for the pose, \\(M_{i,j}\\). Both \\(M_{i,j}\\) and its estimate \\(\\tilde{M}_{i,j}\\) are two-channel binary occupancy maps of size \\(H\\times W\\). To obtain the estimated map \\(\\tilde{M} (t)\\) for the scene, we register each prediction \\(\\tilde{M}_{i,j}\\) onto a larger shared map using the pose \\(P_{i,j}\\) and threshold the final shared map at 0.5 (see Supp. for map registration details). Importantly, the shared map allows communication between both egos' data streams for more informed mapping and sampling, as we show in results." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.594, + 0.654, + 0.611 + ], + "angle": 0, + "content": "4.4. Model training" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.618, + 0.892, + 0.648 + ], + "angle": 0, + "content": "Policy training. We propose a novel dense RL reward to train policy \\(\\pi^V\\):" + }, + { + "type": "equation", + "bbox": [ + 0.604, + 0.66, + 0.787, + 0.677 + ], + "angle": 0, + "content": "\\[\nr (t) = \\Delta Q (t) - \\eta * \\rho (t).\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.687, + 0.892, + 0.764 + ], + "angle": 0, + "content": "\\(\\Delta Q(t)\\) measures the improvement in mapping from taking actions \\(\\{a_{i,t} : i = 1 \\dots 2\\}\\) over not sampling any visual frame at step \\(t\\). \\(\\rho(t)\\) is a penalty term to discourage sampling a frame from the same pose more than once, which we weight by \\(\\eta\\). We define \\(\\Delta Q(t)\\) as" + }, + { + "type": "equation", + "bbox": [ + 0.502, + 0.772, + 0.89, + 0.792 + ], + "angle": 0, + "content": "\\[\n\\Delta Q (t) = Q \\left(\\tilde {M} (t) \\mid O ^ {M} (t)\\right) - Q \\left(\\tilde {M} (t) \\mid \\left(O ^ {M} (t) \\setminus V _ {t}\\right)\\right),\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.801, + 0.894, + 0.862 + ], + "angle": 0, + "content": "where \\( Q \\) is a map quality measure, \\( Q(X|Y) \\) represents the quality of map estimate \\( X \\) given inputs \\( Y \\), and \\( (O^M(t) \\setminus V_t) \\) denotes the mapper inputs devoid of any visual frame for the current step. We define \\( \\rho(t) \\) as" + }, + { + "type": "equation", + "bbox": [ + 0.554, + 0.873, + 0.838, + 0.905 + ], + "angle": 0, + "content": "\\[\n\\rho (t) = \\sum_ {i = 1.. 2} a _ {i, t} * \\mathbb {1} (V _ {i, t} \\in O ^ {M} (t - 1)),\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "10558" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.257 + ], + "angle": 0, + "content": "where the indicator function checks if \\( V_{i,t} \\) was used in mapping before. While \\( \\Delta Q(t) \\) incentivizes sampling frames that provide a big boost to the mapping accuracy over skipping them, \\( \\rho(t) \\) penalizes wasting the visual budget on redundant sampling, thereby maximizing mapping performance within the constraints of a limited budget. We set \\( \\rho = 0.03 \\) in all our experiments and define \\( Q \\) as the average F1 score over the occupied and free classes in a predicted occupancy map. We empirically observe that alternate penalty terms based on visual similarity, like the ORB [55] feature matching similarity, do not improve the model performance." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.257, + 0.472, + 0.318 + ], + "angle": 0, + "content": "We train \\(\\pi^V\\) with Decentralized Distributed PPO (DDPPO) [66]. The DD-PPO loss consists of a value loss, policy loss and an entropy loss to promote exploration (see Supp. for details)." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.324, + 0.473, + 0.356 + ], + "angle": 0, + "content": "Mapper training. At each step \\( t \\), we train the shared mapper \\( f^{m} \\) with a loss \\( \\mathcal{L}^{M}(t) \\), such that" + }, + { + "type": "equation", + "bbox": [ + 0.116, + 0.366, + 0.431, + 0.403 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} ^ {M} (t) = \\frac {1}{2 \\times t} \\sum_ {i = 1 \\ldots 2} \\sum_ {j = 1 \\ldots t} \\mathrm {B C E} (\\tilde {M} _ {i, j}, M _ {i, j}),\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.415, + 0.47, + 0.45 + ], + "angle": 0, + "content": "where \\(\\mathrm{BCE}(\\tilde{M}_{i,j}, M_{i,j})\\) is the average binary cross entropy loss between \\(\\tilde{M}_{i,j}\\) and \\(M_{i,j}\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.453, + 0.473, + 0.606 + ], + "angle": 0, + "content": "Training curriculum. To train our model, we first pretrain mapper \\( f^{m} \\) in two phases and then train the policy \\( \\pi^{V} \\) while keeping \\( f^{m} \\) frozen. In phase 1, we train \\( f^{m} \\) without visual sampling, i.e. all visual frames are provided at each step. In phase 2, we finetune the pretrained weights of \\( f^{m} \\) from phase 1 on episodes where we randomly drop views to satisfy the budget \\( B \\). While phase 1 improves convergence when training with visual sampling, phase 2 helps with reward stationarity when training our RL policy. Both phases use the same training, validation and testing data." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.619, + 0.21, + 0.635 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.644, + 0.473, + 0.795 + ], + "angle": 0, + "content": "Experimental setup. For our main experiments, we use SoundSpaces [8] acoustic simulations with AI-Habitat [57] and Matterport3D [3] visual scenes. While Matterport3D provides dense 3D meshes and image scans of real-world houses and other indoor scenes, SoundSpaces provides room impulse responses (RIRs) at a spatial resolution of \\(1\\mathrm{m}\\) for Matterport3D that model all real-world acoustic phenomena [8]. This setup allows us to evaluate with as many as 83 scenes, split in 56/10/17 for train/val/test, compare against relevant prior work [52, 53] and report reproducible results." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.796, + 0.473, + 0.902 + ], + "angle": 0, + "content": "We also collect real-world data in a mock-up apartment due to the absence of a publicly available alternative suited for our task. We capture a dense set of RGB images using a Samsung S22 camera and generate the corresponding depth images using monocular depth estimation [20, 33]. To compute the RIRs, following [23], we generate a sinusoidal sweep sound from \\(20\\mathrm{Hz} - 20\\mathrm{kHz}\\) with a loudspeaker at source" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.895, + 0.198 + ], + "angle": 0, + "content": "location, capture it with an Eigenmike at a receiver location, and convolve the spatial sound with the inverse of the sweep sound to retrieve the RIR. All capturing devices are placed at a height of \\(1.5\\mathrm{m}\\). We generate occupancy maps by backprojecting the depth images (cf. Sec. 4.1) and register them onto a shared top-down map before taking egocentric crops to generate the local occupancy inputs and targets." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.198, + 0.895, + 0.258 + ], + "angle": 0, + "content": "Note that both datasets have real-world visuals as they are captured in the real environments; SoundSpaces has high-quality simulated audio while the apartment data has real-world collected audio RIRs." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.269, + 0.897, + 0.572 + ], + "angle": 0, + "content": "Conversation episode. For each episode,(simulated or real), we randomly place the egos in a scene. Episode length is \\( T = 16 \\) and 8 for simulation and real, respectively. At each step, the egos execute a movement from \\( A = \\{MoveForward, TurnLeft, TurnRight\\} \\), where \\( MoveForward \\) moves an ego forward by \\( 1\\mathrm{m} \\), and the \\( Turn \\) actions rotate the ego by \\( 90^{\\circ} \\), as required by SoundSpaces. Further, either of the egos speaks or both speak with equal probability of \\( \\frac{1}{3} \\) at every step, i.e., there are no moments of total silence, but one ego may be silent. The egos stay between \\( 1 - 3\\mathrm{m} \\) from each other (typical in a conversation) so that they do not collide and each ego is audible by the other at all times. This results in train/val splits of 1,955,334/100 episodes in simulation, and a simulated/real-world test split of 1000/27 episodes. See Supp. for additional analysis on the model performance vs. training data size. We set the visual budget to \\( B = 2 \\) for our main experiments (see Supp. for \\( B = 4, 6 \\) evaluations). Note that these episodes are solely to generate video data; our task requires processing passive video, not controlling embodied agents." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.577, + 0.897, + 0.76 + ], + "angle": 0, + "content": "Observations and model output. For the occupancy maps, consistent with [4, 11, 53], we generate \\((31 \\times 31 \\times 2)\\)-dimensional input maps that cover \\(3.1 \\times 3.1 \\mathrm{~m}^2\\) in area at a resolution of \\(0.1 \\mathrm{~m}\\), and set the local target map size to \\(H \\times W = 6.4 \\times 6.4 \\mathrm{~m}^2 (\\sim 41 \\mathrm{~m}^2)\\). For speech, we use 100 distinct speakers from LibriSpeech [49], split in 80/11 for heard/unheard, where unheard speech is only used in testing. We assume camera poses, as modern AR devices are equipped with motion sensors that can robustly estimate relative poses [41]. We test our robustness to ambient sounds that get mixed with the egos' speech, and incorporate odometry noise models [52, 53] in Supp." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.897, + 0.903 + ], + "angle": 0, + "content": "Evaluation settings. We evaluate our model in two settings: 1) passive mapping, the mapper has access to all visual frames in an episode (i.e., the camera is always on), and 2) active mapping, where the mapping agent has to actively sample frames to meet the visual budget \\( B \\). This helps disentangle our modeling contributions—whereas passive mapping lets us show improvements in the mapper \\( f^{M} \\) over existing methods [52, 53], active mapping helps demonstrate the benefits of smart visual sampling." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "10559" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.103, + 0.089, + 0.445, + 0.257 + ], + "angle": 0, + "content": "
ModelSimulationReal world
F1 score ↑IoU ↑F1 score ↑IoU ↑
All-occupied63.448.836.223.8
Register-inputs72.660.150.835.0
OccAnt [53]74.562.753.938.3
AV-Floorplan [52]79.367.954.538.7
Ours81.871.455.539.2
Ours w/o vision72.860.350.835.0
Ours w/o audio78.166.754.138.0
Ours w/o \\( E_i' \\)'s speech81.570.955.439.1
Ours w/o shared mapping80.770.054.938.6
Ours w/o modality tag79.167.854.538.1
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.26, + 0.47, + 0.289 + ], + "angle": 0, + "content": "Table 1. Passive mapping performance \\((\\%)\\). All our gains are statistically significant \\((p \\leq 0.05)\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.304, + 0.471, + 0.395 + ], + "angle": 0, + "content": "We use standard evaluation metrics [53]: F1 score and IoU (intersection over union) between the predicted and target scene maps. For both metrics, we report the mean over the free and occupied classes. For active mapping, we average the metrics over 3 random seeds. We use the following baselines to compare our model's efficacy." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.397, + 0.197, + 0.411 + ], + "angle": 0, + "content": "Passive mapping:" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.411, + 0.469, + 0.441 + ], + "angle": 0, + "content": "- All-occupied: a naive baseline that predicts all locations in its map estimate as occupied" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.442, + 0.469, + 0.471 + ], + "angle": 0, + "content": "- Register-inputs: a naive baseline that registers the input maps onto a shared map and uses it as its prediction" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.472, + 0.47, + 0.517 + ], + "angle": 0, + "content": "- OccAnt [53]: a vision-only SOTA model that uses the RGB-D images at each step to anticipate the occupancy of the area around an ego that's outside its visible range." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.517, + 0.47, + 0.607 + ], + "angle": 0, + "content": "- AV-Floorplan [52]: an audio-visual SOTA model that passively predicts the floorplan of a scene using a walk-through in it, where the audio is either self-generated or comes from semantic sources in the scene. We adapt the model for our occupancy prediction task and give it the exact same audio-visual observations as our model." + }, + { + "type": "list", + "bbox": [ + 0.078, + 0.411, + 0.47, + 0.607 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.076, + 0.609, + 0.189, + 0.623 + ], + "angle": 0, + "content": "Active mapping:" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.624, + 0.47, + 0.653 + ], + "angle": 0, + "content": "- Random: an agent that selects visual frames randomly for each ego as long as the budget allows" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.654, + 0.469, + 0.684 + ], + "angle": 0, + "content": "- Greedy: an agent that greedily uses up the visual budget by sampling frames as early as possible" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.685, + 0.469, + 0.713 + ], + "angle": 0, + "content": "- Unique-posed: an agent that samples a frame for every new ego pose in the episode" + }, + { + "type": "list", + "bbox": [ + 0.078, + 0.624, + 0.47, + 0.713 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.715, + 0.47, + 0.775 + ], + "angle": 0, + "content": "In active mapping, we use the model from the second pretraining phase (Sec. 4.4) as the mapper for all models for fair comparison. Thus, any difference in performance is due to the quality of each method's sampling decisions." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.776, + 0.47, + 0.806 + ], + "angle": 0, + "content": "See Supp. for all other details like network architectures, training hyperparameters, and baseline implementation." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.817, + 0.288, + 0.833 + ], + "angle": 0, + "content": "5.1. Map prediction results" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.472, + 0.902 + ], + "angle": 0, + "content": "Passive mapping. Table 1 (top) reports the prediction quality of all models in the passive mapping setting. Naive baselines (All-occupied, Register-inputs) perform worse than the learned models, showing the complexity of our map pre" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.091, + 0.693, + 0.194 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.56, + 0.198, + 0.637, + 0.209 + ], + "angle": 0, + "content": "(a) Simulation" + }, + { + "type": "image", + "bbox": [ + 0.699, + 0.09, + 0.891, + 0.195 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.756, + 0.197, + 0.835, + 0.209 + ], + "angle": 0, + "content": "(b) Real world" + }, + { + "type": "image_caption", + "bbox": [ + 0.528, + 0.222, + 0.863, + 0.236 + ], + "angle": 0, + "content": "Figure 3. Active mapping performance vs. episode step." + }, + { + "type": "image", + "bbox": [ + 0.58, + 0.253, + 0.815, + 0.376 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.5, + 0.388, + 0.894, + 0.416 + ], + "angle": 0, + "content": "Figure 4. Impact of the other ego's speech on passive mapping vs. distance between the egos." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.434, + 0.894, + 0.51 + ], + "angle": 0, + "content": "diction task. AV-Floorplan [52] fares the best among all baselines. Its improvement over OccAnt [53] demonstrates the benefits of exploiting the spatial cues in audio for mapping and using an attention-based model to leverage the long- and short-range correlations in the audio-visual inputs." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.511, + 0.894, + 0.648 + ], + "angle": 0, + "content": "Our method outperforms all baselines. Its improvement over AV-Floorplan [52] underlines the efficacy of performing attention at different granularities—across modalities, within a single modality and within a single input—guided by our positional and modality type embeddings. It also generalizes to the real-world setting and retains its benefits over the baselines, even without retraining on the real-world data. However, we do observe a drop in performance gains, probably due to the large sim-to-real gap." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.658, + 0.894, + 0.78 + ], + "angle": 0, + "content": "Active mapping. Fig. 3 shows the active mapping performance as a function of episode progress. Employing naive heuristics for sampling, like Random or Greedy, is insufficient for high-quality mapping, emphasizing the high levels of redundancy in the visual frames. Unique-pose improves over both Random and Greedy, showing that sampling diverse viewpoints provides more information about the underlying scene geometry." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.78, + 0.895, + 0.901 + ], + "angle": 0, + "content": "Although the baselines make initial progress, they flatten quickly and our model eventually outperforms them all, on both real-world and simulated data (gains are again statistically significant, \\( p \\leq 0.05 \\)). This highlights the benefits of learning a smart policy that, given the audio streams and its past visual samples, understands the value of sampling a visual frame for mapping by taking cues from our novel reward. Besides, on the real-world data, we see better per" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "10560" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.08, + 0.089, + 0.894, + 0.29 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.3, + 0.893, + 0.329 + ], + "angle": 0, + "content": "Figure 5. Sample episodes for our active mapping model. While our policy samples only the salient visual frames, our mapper can both complete partially seen objects as well as anticipate objects never seen before in the sampled visuals (red boxes on the maps)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.346, + 0.473, + 0.452 + ], + "angle": 0, + "content": "formance margins over the baselines towards the end of episodes, showing that our policy can adaptively defer visual sampling to improve mapping. Owing to smart sampling, the per-episode reduction in the capture / processing cost for \\( B = 2 \\) is 74 Watts (W) / 7.2 GFLOPS in simulation and 37 W / 3.6 GFLOPS in the real world video. See Supp. for further analysis of our model's power and time cost." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.467, + 0.229, + 0.483 + ], + "angle": 0, + "content": "5.2. Model analysis" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.492, + 0.471, + 0.838 + ], + "angle": 0, + "content": "Ablations. In Table 1 (bottom), we ablate the components of our model for passive mapping. Upon removing audio, our model experiences a large drop in mapping performance, which indicates that our model leverages complementary spatial cues in audio and vision. We also see a drop in the map quality when our model lacks access to the speech from the other ego \\((E_i')\\). This shows that \\(E_i'\\)'s speech can better reveal the more global scene geometry than \\(E_i\\)'s own speech. Fig. 4 further shows that the impact of the other ego's speech becomes more prominent for larger inter-ego distances \\((3 - 5\\mathrm{m}\\) vs. \\(1 - 3\\mathrm{m})\\), in which case the two types of speech are dissimilar enough to carry complementary geometric cues, but reduces for even larger distances \\((5\\mathrm{m}\\) or more), in which case \\(E_i'\\) is too far for its speech to carry useful cues about \\(E_i\\)'s local scene geometry. Moreover, unlike the ablation that does not perform shared mapping, our model benefits significantly from jointly attending to the observations of the egos and exploiting the complementary information in them—even though both models use the exact same audio-visual observations, including both speech from self and the other ego. Finally, without our modality type embeddings, our model does worse, showing their role in learning complementary features from vision and audio." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.471, + 0.901 + ], + "angle": 0, + "content": "For active mapping, Fig. 3 shows a drop in the mapping performance upon removing audio from the policy inputs. This implies that our policy exploits audio to reason about the level of redundancy in a new visual frame and improve" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.346, + 0.892, + 0.391 + ], + "angle": 0, + "content": "the mapping quality vs. visual budget tradeoff. On the more challenging real-world setting, audio plays an even bigger role, as shown by the larger performance drop in Fig. 3b." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.392, + 0.893, + 0.453 + ], + "angle": 0, + "content": "See Supp. for similar results with 1) ambient sounds, 2) unheard speech, 3) higher budgets, 4) sensor noise, 5) larger target maps, 6) different ego initializations, and 7) multiple random data splits." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.464, + 0.893, + 0.585 + ], + "angle": 0, + "content": "Qualitative results. Fig. 5 shows two successful active mapping episodes of our method. Note how our model samples views that tend have to little visual overlap but are informative of the surrounding geometry (both occupied and free spaces). Besides, it is able to complete structures only partially visible in the sampled views, and more interestingly, leverage the synergy of audio and vision to anticipate unseen areas (red boxes on the occupancy maps in Fig. 5)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.596, + 0.894, + 0.732 + ], + "angle": 0, + "content": "Failure cases. We notice two common failure cases with active mapping: episodes where the egos stay at the same location, leading to very few informative visual frames to sample from; and episodes with highly unique visual samples at every trajectory step, in which case each sample is useful and our model behaves similar to Unique-pose or Greedy. For passive mapping, our model fails with very complex scenes that commonly have objects in spaces where both vision and audio cannot reach (e.g. narrow corners)" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.745, + 0.619, + 0.76 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.767, + 0.894, + 0.871 + ], + "angle": 0, + "content": "We introduce Chat2Map, a new task aimed at scene mapping using audio-visual feeds from egocentric conversations. We develop a novel approach for Chat2Map comprising a shared scene mapper and a visual sampling policy based on a novel reinforcement learner that smartly samples the visuals only when necessary. We show promising performance on both simulated and real-world data from over 80 scenes." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.873, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Acknowledgements: UT Austin is supported in part by the IFML NSF AI Institute. KG is paid as a research scientist by Meta." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "10561" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.169 + ], + "angle": 0, + "content": "[1] Ego-Noise Predictions for Echolocation in Wheeled Robots, volume ALIFE 2019: The 2019 Conference on Artificial Life of ALIFE 2022: The 2022 Conference on Artificial Life, 07 2019. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.17, + 0.472, + 0.21 + ], + "angle": 0, + "content": "[2] Aaron Carroll and Gernot Heiser. An analysis of power consumption in a smartphone. In 2010 USENIX Annual Technical Conference (USENIX ATC 10), 2010. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.21, + 0.471, + 0.278 + ], + "angle": 0, + "content": "[3] Angel Chang, Angela Dai, Thomas Funkhouser, Maciej Halber, Matthias Niessner, Manolis Savva, Shuran Song, Andy Zeng, and Yinda Zhang. Matterport3d: Learning from rgb-d data in indoor environments. arXiv preprint arXiv:1709.06158, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.278, + 0.471, + 0.332 + ], + "angle": 0, + "content": "[4] Devendra Singh Chaplot, Dhiraj Gandhi, Saurabh Gupta, Abhinav Gupta, and Ruslan Salakhutdinov. Learning to explore using active neural slam. In International Conference on Learning Representations, 2020. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.332, + 0.471, + 0.387 + ], + "angle": 0, + "content": "[5] Devendra Singh Chaplot, Dhiraj Prakashchand Gandhi, Abhinav Gupta, and Russ R Salakhutdinov. Object goal navigation using goal-oriented semantic exploration. Advances in Neural Information Processing Systems, 33:4247-4258, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.386, + 0.471, + 0.44 + ], + "angle": 0, + "content": "[6] Changan Chen, Ziad Al-Halah, and Kristen Grauman. Semantic audio-visual navigation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15516–15525, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.44, + 0.471, + 0.495 + ], + "angle": 0, + "content": "[7] Changan Chen, Ruohan Gao, Paul T. Calamia, and Kristen Grauman. Visual acoustic matching. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 18836-18846, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.495, + 0.471, + 0.562 + ], + "angle": 0, + "content": "[8] Changan Chen, Unnat Jain, Carl Schissler, Sebastia Vicenc Amengual Gari, Ziad Al-Halah, Vamsi Krishna Ithapu, Philip Robinson, and Kristen Grauman. Soundspaces: Audiovisual navigation in 3d environments. In ECCV, 2020. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.562, + 0.471, + 0.629 + ], + "angle": 0, + "content": "[9] Changan Chen, Sagnik Majumder, Ziad Al-Halah, Ruohan Gao, Santhosh Kumar Ramakrishnan, and Kristen Grauman. Learning to set waypoints for audio-visual navigation. In International Conference on Learning Representations, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.63, + 0.471, + 0.697 + ], + "angle": 0, + "content": "[10] Jiacheng Chen, Chen Liu, Jiaye Wu, and Yasutaka Furukawa. Floor-sp: Inverse cad for floorplans by sequential room-wise shortest path. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2661-2670, 2019. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.698, + 0.471, + 0.739 + ], + "angle": 0, + "content": "[11] Tao Chen, Saurabh Gupta, and Abhinav Gupta. Learning exploration policies for navigation. arXiv preprint arXiv:1903.01959, 2019. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.739, + 0.471, + 0.793 + ], + "angle": 0, + "content": "[12] Yangyu Chen, Shuhui Wang, Weigang Zhang, and Qingming Huang. Less is more: Picking informative frames for video captioning. In Proceedings of the European conference on computer vision (ECCV), pages 358-373, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.793, + 0.471, + 0.833 + ], + "angle": 0, + "content": "[13] Jesper Christensen, Sascha Hornauer, and Stella Yu. Batvision - learning to see 3d spatial layout with two ears. In ICRA, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.833, + 0.471, + 0.887 + ], + "angle": 0, + "content": "[14] Jesper Haahr Christensen, Sascha Hornauer, and Stella X. Yu. Batvision: Learning to see 3d spatial layout with two ears. In 2020 IEEE International Conference on Robotics and Automation (ICRA), pages 1581-1587, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.887, + 0.471, + 0.901 + ], + "angle": 0, + "content": "[15] Abhishek Das, Théophile Gervet, Joshua Romoff, Dhruv" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.894, + 0.134 + ], + "angle": 0, + "content": "Batra, Devi Parikh, Michael G. Rabbat, and Joelle Pineau. Tarmac: Targeted multi-agent communication. ArXiv, abs/1810.11187, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.134, + 0.894, + 0.174 + ], + "angle": 0, + "content": "[16] Samyak Datta, Sameer Dharur, Vincent Cartillier, Ruta Desai, Mukul Khanna, Dhruv Batra, and Devi Parikh. Episodic memory question answering. 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.174, + 0.894, + 0.227 + ], + "angle": 0, + "content": "[17] Victoria Dean, Shubham Tulsiani, and Abhinav Gupta. See, hear, explore: Curiosity via audio-visual association. Advances in Neural Information Processing Systems, 33:14961–14972, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.228, + 0.894, + 0.282 + ], + "angle": 0, + "content": "[18] Helisa Dhamo, Nassir Navab, and Federico Tombari. Object-driven multi-layer scene decomposition from a single image. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5369-5378, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.282, + 0.894, + 0.336 + ], + "angle": 0, + "content": "[19] Ivan Dokmanić, Reza Parhizkar, Andreas Walther, Yue M. Lu, and Martin Vetterli. Acoustic echoes reveal room shape. Proceedings of the National Academy of Sciences, 110(30):12186-12191, 2013. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.336, + 0.894, + 0.404 + ], + "angle": 0, + "content": "[20] Ainaz Eftekhar, Alexander Sax, Jitendra Malik, and Amir Zamir. Omnidata: A scalable pipeline for making multi-task mid-level vision datasets from 3d scans. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10786-10796, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.404, + 0.894, + 0.459 + ], + "angle": 0, + "content": "[21] Amine Elhafsi, Boris Ivanovic, Lucas Janson, and Marco Pavone. Map-predictive motion planning in unknown environments. In 2020 IEEE International Conference on Robotics and Automation (ICRA), pages 8552-8558. IEEE, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.459, + 0.894, + 0.499 + ], + "angle": 0, + "content": "[22] Itamar Eliakim, Zahi Cohen, Gábor Kósa, and Yossi Yovel. A fully autonomous terrestrial bat-like acoustic robot. PLoS Computational Biology, 14, 2018. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.499, + 0.894, + 0.539 + ], + "angle": 0, + "content": "[23] Angelo Farina. Simultaneous measurement of impulse response and distortion with a swept-sine technique. Journal of The Audio Engineering Society, 2000. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.539, + 0.894, + 0.662 + ], + "angle": 0, + "content": "[24] Chuang Gan, Jeremy Schwartz, Seth Alter, Martin Schrimpf, James Traer, Julian De Freitas, Jonas Kubilius, Abhishek Bhandwaldar, Nick Haber, Megumi Sano, Kuno Kim, Elias Wang, Damian Mrowca, Michael Lingelbach, Aidan Curtis, Kevin T. Feigelis, Daniel M. Bear, Dan Gutfreund, David D. Cox, James J. DiCarlo, Josh H. McDermott, Joshua B. Tenenbaum, and Daniel L. K. Yamins. Threadworld: A platform for interactive multi-modal physical simulation. In NeurIPS Track on Datasets and Benchmarks, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.662, + 0.894, + 0.73 + ], + "angle": 0, + "content": "[25] Chuang Gan, Yiwei Zhang, Jiajun Wu, Boqing Gong, and Joshua B. Tenenbaum. Look, listen, and act: Towards audiovisual embodied navigation. In 2020 IEEE International Conference on Robotics and Automation (ICRA), pages 9701-9707, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.73, + 0.894, + 0.812 + ], + "angle": 0, + "content": "[26] Ruohan Gao, Changan Chen, Ziad Al-Halah, Carl Schissler, and Kristen Grauman. Visualechoes: Spatial image representation learning through echolocation. In Andrea Vedaldi, Horst Bischof, Thomas Brox, and Jan-Michael Frahm, editors, Computer Vision - ECCV 2020, pages 658-676, Cham, 2020. Springer International Publishing. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.812, + 0.894, + 0.867 + ], + "angle": 0, + "content": "[27] Ruohan Gao, Tae-Hyun Oh, Kristen Grauman, and Lorenzo Torresani. Listen to look: Action recognition by previewing audio. 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10454-10464, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.867, + 0.894, + 0.893 + ], + "angle": 0, + "content": "[28] Kristen Grauman, Andrew Westbury, Eugene Byrne, Zachary Chavis, Antonino Furnari, Rohit Girdhar, Jackson Hamburger," + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.894, + 0.893 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "10562" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.11, + 0.092, + 0.473, + 0.424 + ], + "angle": 0, + "content": "Hao Jiang, Miao Liu, Xingyu Liu, Miguel Martin, Tushar Nagarajan, Ilija Radosavovic, Santhosh Kumar Ramakrishnan, Fiona Ryan, Jayant Sharma, Michael Wray, Mengmeng Xu, Eric Zhongcong Xu, Chen Zhao, Siddhant Bansal, Dhruv Batra, Vincent Cartillier, Sean Crane, Tien Do, Morrie Doulaty, Akshay Erapalli, Christoph Feichtenhofer, Adriano Fragomeni, Qichen Fu, Christian Fuegen, Abrham Gebreselasie, Cristina Gonzalez, James Hillis, Xuhua Huang, Yifei Huang, Wenqi Jia, Weslie Khoo, Jachym Kolar, Satwik Kottur, Anurag Kumar, Federico Landini, Chao Li, Yanghao Li, Zhenqiang Li, Karttikeya Mangalam, Raghava Modhugu, Jonathan Munro, Tullie Murrell, Takumi Nishiyasu, Will Price, Paola Ruiz Puentes, Merey Ramazanova, Leda Sari, Kiran Somasundaram, Audrey Southerland, Yusuke Sugano, Ruijie Tao, Minh Vo, Yuchen Wang, Xindi Wu, Takuma Yagi, Yunyi Zhu, Pablo Arbelaez, David Crandall, Dima Damen, Giovanni Maria Farinella, Bernard Ghanem, Vamsi Krishna Ithapu, C. V. Jawahar, Hanbyul Joo, Kris Kitani, Haizhou Li, Richard Newcombe, Aude Oliva, Hyun Soo Park, James M. Rehg, Yoichi Sato, Jianbo Shi, Mike Zheng Shou, Antonio Torralba, Lorenzo Torresani, Mingfei Yan, and Jitendra Malik. Ego4d: Around the World in 3,000 Hours of Egocentric Video. In IEEE/CVF Computer Vision and Pattern Recognition (CVPR), 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.424, + 0.472, + 0.464 + ], + "angle": 0, + "content": "[29] Shariq Iqbal and Fei Sha. Coordinated exploration via intrinsic rewards for multi-agent reinforcement learning. arXiv preprint arXiv:1905.12127, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.465, + 0.473, + 0.573 + ], + "angle": 0, + "content": "[30] Max Jaderberg, Wojciech M. Czarnecki, Iain Dunning, Luke Marris, Guy Lever, Antonio Garcia Castañeda, Charlie Beattie, Neil C. Rabinowitz, Ari S. Morcos, Avraham Ruderman, Nicolas Sonnerat, Tim Green, Louise Deason, Joel Z. Leibo, David Silver, Demis Hassabis, Koray Kavukcuoglu, and Thore Graepel. Human-level performance in 3d multiplayer games with population-based reinforcement learning. Science, 364:859 - 865, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.574, + 0.472, + 0.628 + ], + "angle": 0, + "content": "[31] Unnat Jain, Luca Weihs, Eric Kolve, Ali Farhadi, Svetlana Lazebnik, Aniruddha Kembhavi, and Alexander G. Schwing. A cordial sync: Going beyond marginal policies for multiagent embodied tasks. ArXiv, abs/2007.04979, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.629, + 0.472, + 0.709 + ], + "angle": 0, + "content": "[32] Unnat Jain, Luca Weihs, Eric Kolve, Mohammad Rastegari, Svetlana Lazebnik, Ali Farhadi, Alexander G. Schwing, and Aniruddha Kembhavi. Two body problem: Collaborative visual task completion. 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6682-6692, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.71, + 0.472, + 0.764 + ], + "angle": 0, + "content": "[33] Oğuzhan Fatih Kar, Teresa Yeo, Andrei Atanov, and Amir Zamir. 3d common corruptions and data augmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18963-18974, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.765, + 0.472, + 0.832 + ], + "angle": 0, + "content": "[34] Kapil Katyal, Katie Popek, Chris Paxton, Phil Burlina, and Gregory D. Hager. Uncertainty-aware occupancy map prediction using generative networks for robot navigation. In 2019 International Conference on Robotics and Automation (ICRA), pages 5453-5459, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.832, + 0.472, + 0.886 + ], + "angle": 0, + "content": "[35] Hansung Kim, Luca Remaggi, Philip JB Jackson, Filippo Maria Fazi, and Adrian Hilton. 3d room geometry reconstruction using audio-visual sensors. In 2017 International Conference on 3D Vision (3DV), pages 621–629, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.886, + 0.472, + 0.899 + ], + "angle": 0, + "content": "[36] Bruno Korbar, Du Tran, and Lorenzo Torresani. Scsampler:" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.473, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.894, + 0.134 + ], + "angle": 0, + "content": "Sampling salient clips from video for efficient action recognition. 2019 IEEE/CVF International Conference on Computer Vision (ICCV), pages 6231-6241, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.134, + 0.894, + 0.188 + ], + "angle": 0, + "content": "[37] Robert LiKamWa, Zhen Wang, Aaron Carroll, Felix Xiaozhu Lin, and Lin Zhong. Draining our glass: An energy and heat characterization of google glass. In Proceedings of 5th Asia-Pacific Workshop on Systems, pages 1-7, 2014. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.188, + 0.894, + 0.255 + ], + "angle": 0, + "content": "[38] Jintao Lin, Haodong Duan, Kai Chen, Dahua Lin, and Limin Wang. Ocsampler: Compressing videos to one clip with single-step sampling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13894-13903, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.255, + 0.894, + 0.31 + ], + "angle": 0, + "content": "[39] David B. Lindell, Gordon Wetzstein, and Vladlen Koltun. Acoustic non-line-of-sight imaging. In 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6773-6782, 2019. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.31, + 0.894, + 0.364 + ], + "angle": 0, + "content": "[40] Chen Liu, Jiaye Wu, and Yasutaka Furukawa. Floornet: A unified framework for floorplan reconstruction from 3d scans. In Proceedings of the European conference on computer vision (ECCV), pages 201-217, 2018. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.364, + 0.894, + 0.418 + ], + "angle": 0, + "content": "[41] Wenxin Liu, David Caruso, Eddy Ig, Jing Dong, Anastasios I Mourikis, Kostas Daniilidis, Vijay Kumar, and Jakob Engel. Tlio: Tight learned inertial odometry. IEEE Robotics and Automation Letters, 5(4):5653-5660, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.418, + 0.894, + 0.472 + ], + "angle": 0, + "content": "[42] Sagnik Majumder, Ziad Al-Halah, and Kristen Grauman. Move2hear: Active audio-visual source separation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 275–285, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.472, + 0.894, + 0.513 + ], + "angle": 0, + "content": "[43] Sagnik Majumder, Ziad Al-Halah, and Kristen Grauman. Active audio-visual separation of dynamic sound sources. In European Conference on Computer Vision. Springer, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.513, + 0.894, + 0.567 + ], + "angle": 0, + "content": "[44] Sagnik Majumder, Changan Chen, Ziad Al-Halah, and Kristen Grauman. Few-shot audio-visual learning of environment acoustics. In Thirty-Sixth Conference on Neural Information Processing Systems, 2022. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.567, + 0.894, + 0.635 + ], + "angle": 0, + "content": "[45] Medhini Narasimhan, Erik Wijmans, Xinlei Chen, Trevor Darrell, Dhruv Batra, Devi Parikh, and Amanpreet Singh. Seeing the un-scene: Learning amodal semantic maps for room navigation. In European Conference on Computer Vision, pages 513-529. Springer, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.634, + 0.894, + 0.662 + ], + "angle": 0, + "content": "[46] Simon T O'Callaghan and Fabio T Ramos. Gaussian process occupancy maps*. Int. J. Rob. Res., 31(1):42-62, jan 2012. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.662, + 0.894, + 0.688 + ], + "angle": 0, + "content": "[47] Brian Okorn, Xuehan Xiong, and Burcu Akinci. Toward automated modeling of floor plans. In 3D PVT, 2009. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.688, + 0.894, + 0.755 + ], + "angle": 0, + "content": "[48] Andrew Owens, Phillip Isola, Josh McDermott, Antonio Torralba, Edward H. Adelson, and William T. Freeman. Visually indicated sounds. In 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 2405-2413, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.755, + 0.894, + 0.822 + ], + "angle": 0, + "content": "[49] Vassil Panayotov, Guoguo Chen, Daniel Povey, and Sanjeev Khudanpur. Librispeech: An asr corpus based on public domain audio books. In 2015 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 5206-5210, 2015. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.822, + 0.894, + 0.892 + ], + "angle": 0, + "content": "[50] Rameswar Panda, Chun-Fu Richard Chen, Quanfu Fan, Ximeng Sun, Kate Saenko, Aude Oliva, and Rogerio Feris. Adamml: Adaptive multi-modal learning for efficient video recognition. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7576-7585, 2021. 3" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.892 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "10563" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.174 + ], + "angle": 0, + "content": "[51] Shivansh Patel, Saim Wani, Unnat Jain, Alexander G. Schwing, Svetlana Lazebnik, Manolis Savva, and Angel X. Chang. Interpretation of emergent communication in heterogeneous collaborative embodied agents. 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 15993-15943, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.175, + 0.472, + 0.256 + ], + "angle": 0, + "content": "[52] Senthil Purushwalkam, Sebastia Vicenc Amengual Gari, Vamsi Krishna Ithapu, Carl Schissler, Philip Robinson, Abhinav Gupta, and Kristen Grauman. Audio-visual floorplan reconstruction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1183–1192, 2021. 1, 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.257, + 0.472, + 0.311 + ], + "angle": 0, + "content": "[53] Santhosh K Ramakrishnan, Ziad Al-Halah, and Kristen Grauman. Occupancy anticipation for efficient exploration and navigation. In European Conference on Computer Vision, pages 400-418. Springer, 2020. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.311, + 0.472, + 0.364 + ], + "angle": 0, + "content": "[54] Fabio Ramos and Lionel Ott. Hilbert maps: Scalable continuous occupancy mapping with stochastic gradient descent. The International Journal of Robotics Research, 35(14):1717-1730, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.365, + 0.472, + 0.419 + ], + "angle": 0, + "content": "[55] Ethan Rublee, Vincent Rabaud, Kurt Konolige, and Gary Bradski. Orb: An efficient alternative to sift or surf. In 2011 International Conference on Computer Vision, pages 2564-2571, 2011. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.419, + 0.472, + 0.473 + ], + "angle": 0, + "content": "[56] João Machado Santos, David Portugal, and Rui P. Rocha. An evaluation of 2d slam techniques available in robot operating system. In 2013 IEEE International Symposium on Safety, Security, and Rescue Robotics (SSRR), pages 1-6, 2013. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.473, + 0.472, + 0.554 + ], + "angle": 0, + "content": "[57] Manolis Savva, Abhishek Kadian, Oleksandr Maksymets, Yili Zhao, Erik Wijmans, Bhavana Jain, Julian Straub, Jia Liu, Vladlen Koltun, Jitendra Malik, et al. Habitat: A platform for embodied ai research. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9339-9347, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.555, + 0.472, + 0.609 + ], + "angle": 0, + "content": "[58] Carl Schissler, Christian Loftin, and Dinesh Manocha. Acoustic classification and optimization for multi-modal rendering of real-world scenes. IEEE Transactions on Visualization and Computer Graphics, 24:1246-1259, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.609, + 0.472, + 0.662 + ], + "angle": 0, + "content": "[59] Ransalu Senanayake, Thushan Ganegedara, and Fabio Ramos. Deep occupancy maps: a continuous mapping technique for dynamic environments. In NIPS 2017 Workshop MLITS, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.663, + 0.472, + 0.73 + ], + "angle": 0, + "content": "[60] Rakesh Shrestha, Fei-Peng Tian, Wei Feng, Ping Tan, and Richard Vaughan. Learned map prediction for enhanced mobile robot exploration. In 2019 International Conference on Robotics and Automation (ICRA), pages 1197-1204, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.73, + 0.472, + 0.799 + ], + "angle": 0, + "content": "[61] Jascha Sohl-Dickstein, Santani Teng, Benjamin M. Gaub, Chris C. Rodgers, Crystal Li, Michael R. DeWeese, and Nicol S. Harper. A device for human ultrasonic echolocation. IEEE Transactions on Biomedical Engineering, 62(6):1526-1534, 2015. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.799, + 0.472, + 0.866 + ], + "angle": 0, + "content": "[62] Wei Sui, Lingfeng Wang, Bin Fan, Hongfei Xiao, Huaiyu Wu, and Chunhong Pan. Layer-wise floorplan extraction for automatic urban building reconstruction. IEEE Transactions on Visualization and Computer Graphics, 22(3):1261-1277, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.867, + 0.472, + 0.894 + ], + "angle": 0, + "content": "[63] Maitreya Suin and A. N. Rajagopalan. An efficient framework for dense video captioning. Proceedings of the AAAI" + }, + { + "type": "list", + "bbox": [ + 0.079, + 0.092, + 0.472, + 0.894 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.894, + 0.119 + ], + "angle": 0, + "content": "Conference on Artificial Intelligence, 34(07):12039-12046, Apr. 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.12, + 0.894, + 0.187 + ], + "angle": 0, + "content": "[64] Cheng Sun, Chi-Wei Hsiao, Min Sun, and Hwann-Tzong Chen. Horizonnet: Learning room layout with 1d representation and piano stretch data augmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1047–1056, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.187, + 0.894, + 0.242 + ], + "angle": 0, + "content": "[65] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.242, + 0.894, + 0.295 + ], + "angle": 0, + "content": "[66] Erik Wijmans, Abhishek Kadian, Ari S. Morcos, Stefan Lee, Irfan Essa, Devi Parikh, Manolis Savva, and Dhruv Batra. Dd-ppo: Learning near-perfect pointgoal navigators from 2.5 billion frames. In ICLR, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.295, + 0.894, + 0.349 + ], + "angle": 0, + "content": "[67] Wenming Wu, Xiao-Ming Fu, Rui Tang, Yuhan Wang, YuHao Qi, and Ligang Liu. Data-driven interior plan generation for residential buildings. ACM Trans. Graph., 38(6), nov 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.349, + 0.894, + 0.418 + ], + "angle": 0, + "content": "[68] Zuxuan Wu, Caiming Xiong, Chih-Yao Ma, Richard Socher, and Larry S Davis. Adaframe: Adaptive frame selection for fast video recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1278-1287, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.418, + 0.894, + 0.499 + ], + "angle": 0, + "content": "[69] Shang-Ta Yang, Fu-En Wang, Chi-Han Peng, Peter Wonka, Min Sun, and Hung-Kuo Chu. Dula-net: A dual-projection network for estimating room layouts from a single rgb panorama. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3363–3372, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.499, + 0.894, + 0.554 + ], + "angle": 0, + "content": "[70] Mao Ye, Yu Zhang, Ruigang Yang, and Dinesh Manocha. 3d reconstruction in the presence of glasses by acoustic and stereo fusion. In 2015 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 4885-4893, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.554, + 0.894, + 0.608 + ], + "angle": 0, + "content": "[71] Serena Yeung, Olga Russakovsky, Greg Mori, and Li Fei-Fei. End-to-end learning of action detection from frame glimpses in videos. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2678–2687, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.608, + 0.894, + 0.662 + ], + "angle": 0, + "content": "[72] Yinfeng Yu, Wenbing Huang, Fuchun Sun, Changan Chen, Yikai Wang, and Xiaohong Liu. Sound adversarial audiovisual navigation. In International Conference on Learning Representations, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.662, + 0.894, + 0.73 + ], + "angle": 0, + "content": "[73] Zhoutong Zhang, Jiajun Wu, Qiujia Li, Zhengjia Huang, James Traer, Josh H. McDermott, Joshua B. Tenenbaum, and William T. Freeman. Generative modeling of audible shapes for object perception. 2017 IEEE International Conference on Computer Vision (ICCV), pages 1260-1269, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.73, + 0.894, + 0.796 + ], + "angle": 0, + "content": "[74] Chuhang Zou, Alex Colburn, Qi Shan, and Derek Hoiem. Layoutnet: Reconstructing the 3d room layout from a single rgb image. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2051-2059, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.797, + 0.894, + 0.879 + ], + "angle": 0, + "content": "[75] Chuhang Zou, Jheng-Wei Su, Chi-Han Peng, Alex Colburn, Qi Shan, Peter Wonka, Hung-Kuo Chu, and Derek Hoiem. Manhattan room layout reconstruction from a single \\(360^{\\circ}\\) image: A comparative study of state-of-the-art methods. International Journal of Computer Vision, 129(5):1410-1431, 2021. 2" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.894, + 0.879 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "10564" + } + ] +] \ No newline at end of file diff --git a/2023/Chat2Map_ Efficient Scene Mapping From Multi-Ego Conversations/9562f06a-6203-4663-996c-df2336adbe79_origin.pdf b/2023/Chat2Map_ Efficient Scene Mapping From Multi-Ego Conversations/9562f06a-6203-4663-996c-df2336adbe79_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..fa565086a7958d712252889a9937c837c72545e5 --- /dev/null +++ b/2023/Chat2Map_ Efficient Scene Mapping From Multi-Ego Conversations/9562f06a-6203-4663-996c-df2336adbe79_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:dd797ab05ecb6f9bf9e75e70b946627eb2ccdf0f1dda211baf85c0f12b30b8bf +size 2087649 diff --git a/2023/Chat2Map_ Efficient Scene Mapping From Multi-Ego Conversations/full.md b/2023/Chat2Map_ Efficient Scene Mapping From Multi-Ego Conversations/full.md new file mode 100644 index 0000000000000000000000000000000000000000..574d8f3054c90d3b2a1855327b09959c8a888f43 --- /dev/null +++ b/2023/Chat2Map_ Efficient Scene Mapping From Multi-Ego Conversations/full.md @@ -0,0 +1,320 @@ +# Chat2Map: Efficient Scene Mapping from Multi-Ego Conversations + +Sagnik Majumder $^{1,2,3}$ Hao Jiang $^{2}$ Pierre Moulon $^{2}$ Ethan Henderson $^{2}$ Paul Calamia $^{2}$ Kristen Grauman $^{1,3*}$ Vamsi Krishna Ithapu $^{2*}$ UT Austin $^{2}$ Reality Labs Research, Meta FAIR + +# Abstract + +Can conversational videos captured from multiple egocentric viewpoints reveal the map of a scene in a cost-efficient way? We seek to answer this question by proposing a new problem: efficiently building the map of a previously unseen 3D environment by exploiting shared information in the egocentric audio-visual observations of participants in a natural conversation. Our hypothesis is that as multiple people ("egos") move in a scene and talk among themselves, they receive rich audio-visual cues that can help uncover the unseen areas of the scene. Given the high cost of continuously processing egocentric visual streams, we further explore how to actively coordinate the sampling of visual information, so as to minimize redundancy and reduce power use. To that end, we present an audio-visual deep reinforcement learning approach that works with our shared scene mapper to selectively turn on the camera to efficiently chart out the space. We evaluate the approach using a state-of-the-art audio-visual simulator for 3D scenes as well as real-world video. Our model outperforms previous state-of-the-art mapping methods, and achieves an excellent cost-accuracy tradeoff. Project: http://vision.cs.utexas.edu/projects/chat2map. + +# 1. Introduction + +The spatial layout of the environment around us is fundamental to understanding our physical context. By representing the walls, furniture, and other major structures in a space, scene maps ground activity and objects in a persistent frame of reference, facilitating high-level reasoning for many downstream applications in augmented reality (AR) and robotics. For example, episodic memory [16, 28] aims to relocalize lost objects observed in first-person video (where are my keys?); floorplan estimation [10, 40, 47] aims to chart out the area and shapes of complex buildings; navigating agents try to discover routes in unfamiliar spaces [4, 11, 53]. + +While traditional computer vision approaches for mapping (e.g., visual SLAM) are highly effective when extensive + +![](images/02b86ba0a5d0fef8f1d91542217a239f3c2a45bea27902b7cd21a1f522f16f87.jpg) +Figure 1. Given egocentric audio-visual observations from multiple people wearing AR glasses and moving and conversing (left), we aim to accurately map the scene (right). To mitigate cost, our model receives audio continuously but learns to selectively employ the ego cameras only when the visual input is expected to be informative. + +exposure to the environment is possible, in many real-world scenarios only a fraction of the space is observed by the camera. Recent work shows the promise of sensing 3D spaces with both sight and sound [8, 13, 24, 26, 52]: listening to echoes bounce around the room can reveal the depth and shape of surrounding surfaces, and even help extrapolate a floorplan beyond the camera's field of view or behind occluded objects [52]. + +While we are inspired by these advances, they also have certain limitations. Often systems will emit sounds (e.g., a frequency sweep) into the environment to ping for spatial information [1, 13, 14, 22, 26, 39, 52, 61], which is intrusive if done around people. Furthermore, existing audio-visual models assume that the camera is always on grabbing new frames, which is wasteful if not intractable, particularly on lightweight, low-power computing devices in AR settings. + +We introduce Chat2Map, a new scene mapping task aimed at eliminating these challenges. In the proposed setting, multiple people converse as they move casually through + +the scene while wearing AR glasses equipped with an egocentric camera, microphones, and potentially other sensors (e.g., for odometry).1 Given their egocentric audio-visual data streams, the goal is to infer the ground-plane occupancy map for the larger environment around them. See Figure 1. + +We observe that audio-visual data from the egos' interactions will naturally reflect scene structure. First, as they walk and talk, their movements reveal spaces like corridors, doorways, and large rooms, in both modalities. Second, the speech captured by the device-wearer's cameras and microphones can be localized to different speakers, which, compared to active sound emission, is non-intrusive. + +To realize this vision, we develop a novel approach to efficient scene mapping from multi-ego conversations. Our approach has two key elements: a shared scene mapper and a visual sampling policy. For the former, we devise a transformer-based mapper that incorporates the multiple data streams to infer a map beyond the directly observed areas, and, most importantly, that enables communication among the egos about their observations and states in the 3D space to improve mapping accuracy. For the latter, our idea is to relax the common assumption of an "always-on" camera, and instead actively select when to sample visual frames from any one of the ego cameras. Intuitively, certain regions where the egos move will be more or less important for mapping (e.g., corners of the room, doors). We train a sampling policy with deep reinforcement learning that activates the visual feed only when it is anticipated to complement the continuous audio feed. This is a cost-conscious approach, mindful that switching on a camera is much more power consuming than sensing audio with microphones [2, 37]. + +We demonstrate our approach using a state-of-the-art audio-visual simulator for 3D scenes [8] as well as real-world video input. We can successfully map an unfamiliar environment given only partial visibility via multiple conversing people moving about the scene. Compared to sampling all visual frames, our model reduces the visual capture and processing by $87.5\%$ while the mapping accuracy declines only marginally $(\sim 9\%)$ . + +Our main contributions are 1) we define the task of efficient and shared scene mapping from multi-ego conversations, a new direction for AR/VR research; 2) we present the first approach to begin tackling this task, specifically an RL-based framework that integrates an audio-visual scene mapper and a smart visual sampling policy; and 3) we rigorously experiment with a variety of environments (both in simulation and the real world), speech sounds, and use cases. + +# 2. Related Work + +Visual scene mapping. Past works tackle scene mapping using 3D Manhattan layouts [18, 64, 69, 74, 75], detailed + +floorplans [10, 40, 47, 62, 67], occupancy [21, 34, 46, 54, 59, 60], and semantic maps [45]. Manhattan layouts include structured outputs like scene boundaries [64, 74, 75], corners [74, 75], and floor/ceilings [69, 75], but do not generalize to unseen environment regions. Floorplan estimation methods use dense scans of 3D scenes to predict geometric (walls, exterior/ interior) and semantic layouts (room type, object type, etc.)., rely on extensive human walkthroughs with RGB-D [10, 40] or 3D point cloud [47, 62] scans, and are usually limited to polygonal layouts [10, 40, 47, 62, 67]. Occupancy maps traditionally rely on wide field-of-view (FoV) LiDAR scanners [56] or evaluate on simple 2D environments without non-wall obstacles [21, 34, 60, 60]. More recent methods [4, 5, 11, 53] train an embodied agent to explore and build top-down maps of more complex scenes using RGB-D. On the contrary, our method uses both vision and audio from the observations of a group of conversing people for mapping. Rather than steer the camera of a robot to map the scene, our task requires processing passive video from human camera wearers. + +Audio-visual scene mapping. To our knowledge, the only prior work to translate audio-visual inputs into a general (arbitrarily shaped) floorplan maps is AV-Floorplan [52]. Unlike AV-Floorplan, our method infers maps from speech in natural human conversations, which avoids emitting intrusive frequency sweep signals to generate echoes. In addition, a key goal of our work is to reduce mapping cost by skipping redundant visual frames. Our experiments demonstrate the benefits of our model design over AV-Floorplan [52]. + +Audio(-visual) spatial understanding. More broadly, beyond the mapping task, various methods leverage audio for geometric and material information about the 3D scene and its constituent objects. Prior work relies on acoustic reflections to estimate the shape of objects [39]. Echolocation is used in robotics to estimate proximity to surrounding surfaces [1, 14, 22, 61]. Together, vision and audio can better reveal the shape and materials of objects [48, 58, 73], self-supervise imagery [26], and improve depth sensing [35, 70]. Recent work exploits correlations between spatial audio and imagery to reason about scene acoustics [7, 44] or aid active embodied navigation [6, 9, 17, 25, 72] and source separation [42, 43]. No prior work intelligently selects images to capture during conversations to efficiently map a scene. + +Multi-agent spatial understanding. There is existing work [15, 29, 31, 32, 51] in the visual multi-agent reinforcement learning (MARL) community that learns collaborative agents for performing tasks like relocating furniture [31, 32], playing 3D multi-player games [30], coordinated scene exploration [29], or multi-object navigation [51]. In such settings, the collaborative agents actively interact with the environment to learn a shared scene representation for successfully completing their task. In contrast, we aim to learn a + +shared geometric map of a 3D scene given passive observations that come from the trajectories chosen by a group of people involved in a natural conversation. + +Efficient visual sampling in video. Efficient visual sampling has been studied in the context of video recognition [27, 36, 38, 68, 71] and summarization [12, 63] with the goal of selectively processing informative frames, which can both reduce computational cost and improve recognition. More closely related to our approach are methods that use audio for the decision-making [27, 36, 50]. Different from the above, we use efficient visual sampling in the context of mapping scenes. Furthermore, in our case an online sampling decision needs to be made at every step before looking at the current visual frame (or frames from future steps). + +# 3. Chat2Map Task Formulation + +We propose a novel task: efficient and shared mapping of scenes from multi-ego conversations. + +Without loss of generality, we consider two egos, $E_{1}$ and $E_{2}$ , each wearing AR glasses equipped with an RGB-D camera and a multi-channel microphone array. The egos have a conversation and move around in an unmapped 3D environment. Each conversation is $T$ steps long. At each step $t$ , the ego $E_{i}$ 's glasses receive an observation $\mathcal{O}_{i,t} = (\mathcal{V}_{i,t}, S_{i,t}, P_{i,t}, S_{i,t}', P_{i,t}')$ . $\mathcal{V}_{i,t}$ is the $90^{\circ}$ FOV RGB-D image and $S_{i,t}$ is the speech waveform uttered by $E_{i}$ , as observed from its pose $P_{i,t} = (x_{i,t}, y_{i,t}, \theta_{i,t})$ , where $(x_{i,t}, y_{i,t})$ denotes its location and $\theta_{i,t}$ denotes its orientation in the 3D scene. $S_{i,t}'$ is the speech of the other ego $E_{i}'$ (the other person involved in the conversation), as perceived by $E_{i}$ (note, the speech sounds different depending on the listener position), and $P_{i,t}'$ is $E_{i}'$ 's pose relative to $E_{i}$ . Modern AR glasses, like Bose Frames or Facebook Aria already support capturing such multi-sensory observations, making it possible to have a real-world instantiation of our task. + +Given the real-time observation stream $\mathcal{O}$ for the egos, where $\mathcal{O} = \{\mathcal{O}_{i,t}:i = 1,\dots ,2,t = 1\dots T\}$ and a total budget of visual frames $B$ we aim to learn a model that can accurately estimate the top-down occupancy map $M$ of the scene without exceeding the visual budget. The fixed budget represents allotting a fixed amount of battery to the energy-intensive visual capture [37], so that other on-device routines are not affected. We assume the first visual frames (at $t = 1$ ) for both egos to be used by the model. Thus we aim to learn a policy that samples $B$ frames from $2*(T - 1)$ choices—which are not considered a batch, but rather unfold in sequence—and a mapper that predicts the scene map given the sampled frames. Our goal is to build a model that samples the expensive visual frames only when absolutely needed for scene mapping. This is captured by the constraint $1\leq B\ll 2*(T - 1)$ . + +There are three important aspects to our task. First, it + +requires learning from both vision and audio. While the visual signal carries rich information about the local scene geometry, there can be a high amount of redundancy in the visual feed captured during a conversation (e.g., the egos may visit the same location more than once or change their viewpoint only marginally). Second, not only does the long-range nature of audio help uncover the global scene properties [19, 52] like shape and size—beyond what is visible in images—but also we can exploit audio to undersample the visual frames, thereby reducing the cost of capturing and processing sensory inputs for mapping. Third, shared mapping of a scene allows jointly leveraging the complementary information in the audio (speech) from the self and other egos, and the synergy of the audio-visual cues from multiple egos. These insights form the basis of our key hypothesis in this task—selectively sampling visual frames during a conversation involving egos that share information with each other can facilitate efficient mapping of a scene. + +# 4. Approach + +We solve the task by learning a model that estimates the scene map given the egos' audio-visual observations and also sequentially decides when to sample visual frames for mapping given the audio stream, ego poses, and previously sampled frames, if any. Here, "sampling" refers to individually deciding for each ego whether to use its camera or not to capture the visuals at every step of its trajectory in the scene. The sampling is preemptive in nature, i.e. the policy selects or skips a frame without capturing it first. + +Our model has two main components (see Fig. 2): (1) a shared scene mapper, and (2) a visual sampling policy. At every step $t$ , the shared mapper has two functions. First, it estimates the map of a previously unseen environment by exploiting the shared spatial cues in the audio-visual observations of the two egos. Second, it informs the policy about the utility of sampling a certain visual frame. Guided by the mapper, the policy samples only the most informative visual frames that are expected to boost mapping significantly over using just audio. Note that, unlike the visuals, we observe audio continuously as it is less resource-intensive vis-a-vis storage and power requirements for processing [2]. + +We learn our task through the synergy of the mapper and the policy, such that under the constraint of a limited visual budget $B$ , our model implicitly understands which visual frames are critical for mapping. + +First, we describe the steps involved to prepare our model inputs (Sec. 4.1). Next, we introduce our visual sampling policy (Sec. 4.2) and shared scene mapper (Sec. 4.3). Finally, we present model training details (Sec. 4.4). Through the rest of the text, we use separate notations to distinguish the egos' observations $\mathcal{O}$ (i.e. what the egos receive from the environment) from our model inputs $O$ (i.e. what we capture and feed to our model for efficient mapping). + +![](images/b507832bb99458974534599042ac1c4d06072feaa28aab7988b17873014defa0.jpg) +Figure 2. Our model has two main components: a) a visual sampling policy (left), and b) a shared scene mapper (right). At each step, our policy receives the current audio along with the previous audio(-visual) observations for the egos and decides for each ego individually whether to capture its visual frame at the current step. As per the policy predictions, the shared mapper conditionally uses the current visual frame(s) and audio along with the past audio(-visual) observations to predict the occupancy map of the scene, a ground-plane map showing where obstacles and freespace are (shown in green and white). + +![](images/b6609000eeaff0df62be6fdbce6ef0677b96427cdf1fa89d7ea4912de78343be.jpg) + +# 4.1. Model input preparation + +We prepare our model inputs by separately preprocessing the visual and audio modalities. If our policy decides to sample an image $\mathcal{V}$ , we transform it into $V = (V^{R}, V^{M})$ . $V^{R}$ denotes the normalized RGB image with pixel values $\in [0,1]$ . $V^{M}$ denotes the $90^{\circ}$ FoV top-down occupancy map created by projecting the depth image to the ground plane. To do the depth projection, we first backproject it into the world coordinates using the camera's intrinsic parameters to compute the local visible scene's 3D point cloud. Next, we project these points to obtain a two-channel binary top-down map of size $h \times w \times 2$ , where the first channel reveals occupied/free areas, and the second channel reveals seen/unseen areas. If our policy skips $\mathcal{V}$ , we set $V^{R}$ and $V^{M}$ to all-zero matrices of the appropriate size. + +For a speech waveform $S$ , we calculate the short-time Fourier transform (STFT) magnitude spectrogram denoted by $S$ of size $F \times T \times C$ , where $F$ , $T$ , and $C$ are the number of frequency bins, time windows, and ambisonic microphone channels, respectively. Lastly, we normalize each pose $P_{i,t}$ to be relative to $P_{1,1}$ . See Sec. 5 and Supp. for more details. + +# 4.2. Visual sampling policy + +At every step $t$ , our visual sampling policy $\pi^V$ (Fig. 2 left) receives $O^\pi(t)$ as input and makes the decision to either capture or skip the visual frame $\mathcal{V}_{i,t}$ for each ego $E_i$ . $O^\pi(t)$ comprises the visual cue from the last step along with the + +speech cues, and the poses from the current step and the last step for both egos. Formally, $O^{\pi}(t) = \{O_{i}^{\pi}(t):i = 1\ldots 2\}$ , where $O_{i}^{\pi}(t) = \{V_{i,t - 1},S_{i,j},P_{i,j},S_{i,j}^{\prime},P_{i,j}^{\prime}:j = t - 1\ldots t\}$ . The policy first uses an encoder network to generate a multi-modal embedding of $O^{\pi}(t)$ , and then passes the embedding to a policy network that makes a sampling decision per ego. At $t = 1$ , as per our problem definition (Sec. 3), the policy always chooses to sample the visual frames for both egos, i.e., the cameras are initially on. + +Multi-modal policy embedding. To process ego $E_{i}$ 's visual input $V_{i,t-1}$ from the last step, we encode the RGB image $V_{i,t-1}^{R}$ and map $V_{i,t-1}^{M}$ with separate CNNs. We then concatenate the two features to generate the visual embedding $v_{i,t-1}$ . To encode the pose inputs $\{P_{i,t-1}, P_{i,t-1}', P_{i,t}, P_{i,t}'\}$ , we use a linear layer and generate pose embeddings $\{p_{i,t-1}, p_{i,t-1}', p_{i,t}, p_{i,t}'\}$ . We process the speech inputs $\{S_{i,t-1}, S_{i,t-1}', S_{i,t}, S_{i,t}'\}$ using another CNN and create speech embeddings $\{s_{i,t-1}, s_{i,t-1}', s_{i,t}, s_{i,t}'\}$ . Next, we fuse the visual, speech and pose embeddings using linear layers (see Fig. 2 left) to obtain the multi-modal policy embedding $e_{i,t}$ for $E_{i}$ . Finally, we fuse the policy embeddings for the two egos, $e_{1,t}$ and $e_{2,t}$ with a linear layer to produce the multi-modal policy embedding $e_t$ . + +The visual, audio, and pose inputs carry complementary cues required for efficient visual sampling. Whereas the pose inputs from the last and current steps explicitly reveal + +the viewpoint change between the steps, the previous and current speech inputs provide information about the changes in the local and global scene structures as a function of the previously sampled visual inputs, which together suggest the value of sampling a visual frame at the current step. Furthermore, guided by our training reward (below in Sec. 4.4), the previously observed visual frames and audio together enable our policy to anticipate the current frames and skip them if they are deemed redundant, thereby improving mapping accuracy for a low visual budget. + +Policy network. The policy network consists of a GRU that estimates an updated history $h_t$ along with the current state representation $g_t$ , using the fused embedding $e_t$ and the history of states $h_{t-1}$ . An actor-critic module takes $g_t$ and $h_{t-1}$ as inputs and predicts a policy distribution $\pi_\theta(a_{i,t}|g_t, h_{t-1})$ per ego along with the value of the state $H_\theta(g_t, h_{t-1})$ ( $\theta$ are policy parameters). The policy samples an action $a_{i,t} \in \{0,1\}$ for every $E_i$ . $a_{i,t} = 1$ corresponds to selecting $\mathcal{V}_{i,t}$ , whereas $a_{i,t} = 0$ means to skip that frame. + +# 4.3. Shared scene mapper + +Whereas $O^{\pi}(t)$ denotes our policy input (Sec. 4.2), $O^{M}(t)$ denotes the input to our shared scene mapper $f^{M}$ at step $t$ , such that $O^{M}(t) = \left\{(V_{i,j}, S_{i,j}, S_{i,j}^{\prime}, P_{i,j}, P_{i,j}^{\prime}): i = 1 \dots 2, j = 1 \dots t\right\}$ . $f^{m}$ starts by embedding each component of $O^{M}(t)$ using a separate network. This is followed by a multi-modal memory that stores the embeddings since the start of the episode. Finally, a transformer [65] predicts an estimate $\tilde{M}(t)$ of the scene map conditioned on the multi-modal memory and the egos' poses in the episode. + +Multi-modal mapper embedding. For the visual input $V_{i,j}$ , we encode $V_{i,j}^{R}$ and $V_{i,j}^{M}$ using separate CNNs and do a channel-wise concatenation to get visual features $\hat{v}_{i,j}$ . Similarly, speech is encoded using separate CNNs to get $\hat{s}_{i,j}$ and $\hat{s}_{i,j}'$ . Each of $\hat{v}$ , $\hat{s}$ and $\hat{s}'$ is of size $4 \times 4 \times 1024$ . + +For both vision and speech, we compute two positional embeddings, $p^{\mathrm{I}}$ and $p^{\mathrm{II}}$ . They encode the pose of the egos in the 3D space, and the index of each 1024-dimensional feature in the visual or speech features in the raster order respectively. Whereas $p^{\mathrm{I}}$ helps discover spatial cues as a function of the egos' location in the 3D scene, $p^{\mathrm{II}}$ enables our model to attend to different modalities in a more fine-grained manner. For both, we compute an 8-dimensional sinusoidal positional encoding [65] and then pass it through a linear layer to obtain a 1024-dimensional embedding. For $p^{\mathrm{II}}$ , we additionally repeat this process for every feature index in the raster order. Lastly, we reshape $p^{\mathrm{I}}$ and add it with $p^{\mathrm{II}}$ to produce $4 \times 4 \times 1024$ -dimensional positional embeddings, $\hat{p}_{i,j}$ for $\hat{v}_{i,j}$ and $\hat{s}_{i,j}$ , and $\hat{p}_{i,j}'$ for $\hat{s}_{i,j}'$ . + +Following [44], we also learn an embedding $\hat{m}_{i,j} \in \{\hat{m}_V, \hat{m}_S, \hat{m}_{S'}\}$ to capture different modality types, where $\hat{m}_V$ represents vision, and $\hat{m}_S$ and $\hat{m}_{S'}$ represent the speech + +from self and that of the other ego, respectively. The modality-based embeddings help our model differentiate between different modalities and better map the scene by learning complementary spatial cues from them. + +Multi-modal memory. For the visual input $V_{i,j}$ , we add its embedding $\hat{v}_{i,j}$ with its positional embedding $\hat{p}_{i,j}$ and modality embedding $\hat{m}_{i,j}^{V}$ , and flatten the sum to get a $16 \times 1024$ -dimensional embedding. Similarly, we fuse the speech embeddings by taking their sum and flattening them. This generates a multi-modal memory of fused embeddings $o$ , such that $o = \{o_{1,1}^{V}, \ldots, o_{2,t}^{V}, o_{1,1}^{S}, \ldots, o_{2,t}^{S}, o_{1,1}^{S'}, \ldots, o_{2,t}^{S'}\}$ . + +Occupancy prediction. To predict the underlying scene occupancy, we first use a transformer encoder [65] to attend to the embeddings in $o$ and capture short- and long-range correlations within and across modalities using a stack of self-attention layers. This generates an audio-visual representation that models the spatial layout of the 3D scene. + +Next, we use a transformer decoder [65] to perform cross-attention on the audio-visual representation of the scene conditioned on the embedding $\hat{p}_{i,j}$ for every pose $P_{i,j}$ in $O^M (t)$ and generate an embedding $d_{i,j}$ for the pose. Finally, we upsample $d_{i,j}$ using a multi-layer network $U$ comprising transpose convolutions and a sigmoid layer at the end to predict an estimate $\tilde{M}_{i,j}$ of the ground-truth local $360^{\circ}$ FoV map for the pose, $M_{i,j}$ . Both $M_{i,j}$ and its estimate $\tilde{M}_{i,j}$ are two-channel binary occupancy maps of size $H\times W$ . To obtain the estimated map $\tilde{M} (t)$ for the scene, we register each prediction $\tilde{M}_{i,j}$ onto a larger shared map using the pose $P_{i,j}$ and threshold the final shared map at 0.5 (see Supp. for map registration details). Importantly, the shared map allows communication between both egos' data streams for more informed mapping and sampling, as we show in results. + +# 4.4. Model training + +Policy training. We propose a novel dense RL reward to train policy $\pi^V$ : + +$$ +r (t) = \Delta Q (t) - \eta * \rho (t). +$$ + +$\Delta Q(t)$ measures the improvement in mapping from taking actions $\{a_{i,t} : i = 1 \dots 2\}$ over not sampling any visual frame at step $t$ . $\rho(t)$ is a penalty term to discourage sampling a frame from the same pose more than once, which we weight by $\eta$ . We define $\Delta Q(t)$ as + +$$ +\Delta Q (t) = Q \left(\tilde {M} (t) \mid O ^ {M} (t)\right) - Q \left(\tilde {M} (t) \mid \left(O ^ {M} (t) \setminus V _ {t}\right)\right), +$$ + +where $Q$ is a map quality measure, $Q(X|Y)$ represents the quality of map estimate $X$ given inputs $Y$ , and $(O^M(t) \setminus V_t)$ denotes the mapper inputs devoid of any visual frame for the current step. We define $\rho(t)$ as + +$$ +\rho (t) = \sum_ {i = 1.. 2} a _ {i, t} * \mathbb {1} (V _ {i, t} \in O ^ {M} (t - 1)), +$$ + +where the indicator function checks if $V_{i,t}$ was used in mapping before. While $\Delta Q(t)$ incentivizes sampling frames that provide a big boost to the mapping accuracy over skipping them, $\rho(t)$ penalizes wasting the visual budget on redundant sampling, thereby maximizing mapping performance within the constraints of a limited budget. We set $\rho = 0.03$ in all our experiments and define $Q$ as the average F1 score over the occupied and free classes in a predicted occupancy map. We empirically observe that alternate penalty terms based on visual similarity, like the ORB [55] feature matching similarity, do not improve the model performance. + +We train $\pi^V$ with Decentralized Distributed PPO (DDPPO) [66]. The DD-PPO loss consists of a value loss, policy loss and an entropy loss to promote exploration (see Supp. for details). + +Mapper training. At each step $t$ , we train the shared mapper $f^{m}$ with a loss $\mathcal{L}^{M}(t)$ , such that + +$$ +\mathcal {L} ^ {M} (t) = \frac {1}{2 \times t} \sum_ {i = 1 \ldots 2} \sum_ {j = 1 \ldots t} \mathrm {B C E} (\tilde {M} _ {i, j}, M _ {i, j}), +$$ + +where $\mathrm{BCE}(\tilde{M}_{i,j}, M_{i,j})$ is the average binary cross entropy loss between $\tilde{M}_{i,j}$ and $M_{i,j}$ . + +Training curriculum. To train our model, we first pretrain mapper $f^{m}$ in two phases and then train the policy $\pi^{V}$ while keeping $f^{m}$ frozen. In phase 1, we train $f^{m}$ without visual sampling, i.e. all visual frames are provided at each step. In phase 2, we finetune the pretrained weights of $f^{m}$ from phase 1 on episodes where we randomly drop views to satisfy the budget $B$ . While phase 1 improves convergence when training with visual sampling, phase 2 helps with reward stationarity when training our RL policy. Both phases use the same training, validation and testing data. + +# 5. Experiments + +Experimental setup. For our main experiments, we use SoundSpaces [8] acoustic simulations with AI-Habitat [57] and Matterport3D [3] visual scenes. While Matterport3D provides dense 3D meshes and image scans of real-world houses and other indoor scenes, SoundSpaces provides room impulse responses (RIRs) at a spatial resolution of $1\mathrm{m}$ for Matterport3D that model all real-world acoustic phenomena [8]. This setup allows us to evaluate with as many as 83 scenes, split in 56/10/17 for train/val/test, compare against relevant prior work [52, 53] and report reproducible results. + +We also collect real-world data in a mock-up apartment due to the absence of a publicly available alternative suited for our task. We capture a dense set of RGB images using a Samsung S22 camera and generate the corresponding depth images using monocular depth estimation [20, 33]. To compute the RIRs, following [23], we generate a sinusoidal sweep sound from $20\mathrm{Hz} - 20\mathrm{kHz}$ with a loudspeaker at source + +location, capture it with an Eigenmike at a receiver location, and convolve the spatial sound with the inverse of the sweep sound to retrieve the RIR. All capturing devices are placed at a height of $1.5\mathrm{m}$ . We generate occupancy maps by backprojecting the depth images (cf. Sec. 4.1) and register them onto a shared top-down map before taking egocentric crops to generate the local occupancy inputs and targets. + +Note that both datasets have real-world visuals as they are captured in the real environments; SoundSpaces has high-quality simulated audio while the apartment data has real-world collected audio RIRs. + +Conversation episode. For each episode,(simulated or real), we randomly place the egos in a scene. Episode length is $T = 16$ and 8 for simulation and real, respectively. At each step, the egos execute a movement from $A = \{MoveForward, TurnLeft, TurnRight\}$ , where $MoveForward$ moves an ego forward by $1\mathrm{m}$ , and the $Turn$ actions rotate the ego by $90^{\circ}$ , as required by SoundSpaces. Further, either of the egos speaks or both speak with equal probability of $\frac{1}{3}$ at every step, i.e., there are no moments of total silence, but one ego may be silent. The egos stay between $1 - 3\mathrm{m}$ from each other (typical in a conversation) so that they do not collide and each ego is audible by the other at all times. This results in train/val splits of 1,955,334/100 episodes in simulation, and a simulated/real-world test split of 1000/27 episodes. See Supp. for additional analysis on the model performance vs. training data size. We set the visual budget to $B = 2$ for our main experiments (see Supp. for $B = 4, 6$ evaluations). Note that these episodes are solely to generate video data; our task requires processing passive video, not controlling embodied agents. + +Observations and model output. For the occupancy maps, consistent with [4, 11, 53], we generate $(31 \times 31 \times 2)$ -dimensional input maps that cover $3.1 \times 3.1 \mathrm{~m}^2$ in area at a resolution of $0.1 \mathrm{~m}$ , and set the local target map size to $H \times W = 6.4 \times 6.4 \mathrm{~m}^2 (\sim 41 \mathrm{~m}^2)$ . For speech, we use 100 distinct speakers from LibriSpeech [49], split in 80/11 for heard/unheard, where unheard speech is only used in testing. We assume camera poses, as modern AR devices are equipped with motion sensors that can robustly estimate relative poses [41]. We test our robustness to ambient sounds that get mixed with the egos' speech, and incorporate odometry noise models [52, 53] in Supp. + +Evaluation settings. We evaluate our model in two settings: 1) passive mapping, the mapper has access to all visual frames in an episode (i.e., the camera is always on), and 2) active mapping, where the mapping agent has to actively sample frames to meet the visual budget $B$ . This helps disentangle our modeling contributions—whereas passive mapping lets us show improvements in the mapper $f^{M}$ over existing methods [52, 53], active mapping helps demonstrate the benefits of smart visual sampling. + +
ModelSimulationReal world
F1 score ↑IoU ↑F1 score ↑IoU ↑
All-occupied63.448.836.223.8
Register-inputs72.660.150.835.0
OccAnt [53]74.562.753.938.3
AV-Floorplan [52]79.367.954.538.7
Ours81.871.455.539.2
Ours w/o vision72.860.350.835.0
Ours w/o audio78.166.754.138.0
Ours w/o \( E_i' \)'s speech81.570.955.439.1
Ours w/o shared mapping80.770.054.938.6
Ours w/o modality tag79.167.854.538.1
+ +Table 1. Passive mapping performance $(\%)$ . All our gains are statistically significant $(p \leq 0.05)$ . + +We use standard evaluation metrics [53]: F1 score and IoU (intersection over union) between the predicted and target scene maps. For both metrics, we report the mean over the free and occupied classes. For active mapping, we average the metrics over 3 random seeds. We use the following baselines to compare our model's efficacy. + +Passive mapping: + +- All-occupied: a naive baseline that predicts all locations in its map estimate as occupied +- Register-inputs: a naive baseline that registers the input maps onto a shared map and uses it as its prediction +- OccAnt [53]: a vision-only SOTA model that uses the RGB-D images at each step to anticipate the occupancy of the area around an ego that's outside its visible range. +- AV-Floorplan [52]: an audio-visual SOTA model that passively predicts the floorplan of a scene using a walk-through in it, where the audio is either self-generated or comes from semantic sources in the scene. We adapt the model for our occupancy prediction task and give it the exact same audio-visual observations as our model. + +# Active mapping: + +- Random: an agent that selects visual frames randomly for each ego as long as the budget allows +- Greedy: an agent that greedily uses up the visual budget by sampling frames as early as possible +- Unique-posed: an agent that samples a frame for every new ego pose in the episode + +In active mapping, we use the model from the second pretraining phase (Sec. 4.4) as the mapper for all models for fair comparison. Thus, any difference in performance is due to the quality of each method's sampling decisions. + +See Supp. for all other details like network architectures, training hyperparameters, and baseline implementation. + +# 5.1. Map prediction results + +Passive mapping. Table 1 (top) reports the prediction quality of all models in the passive mapping setting. Naive baselines (All-occupied, Register-inputs) perform worse than the learned models, showing the complexity of our map pre + +![](images/4b8927fbc208d5e693d9e96cd47be53f85e258be176700e1909327fa1a4cd3f7.jpg) +(a) Simulation + +![](images/668d11c8601cda5195c65b3a29d2946ffeb20eae8e5fc6c9aa3fa8a2b3ea92a4.jpg) +(b) Real world + +![](images/5a69f564045c47faa92d760d19b7690880249e00dddc60a1ba3436c94afc154a.jpg) +Figure 3. Active mapping performance vs. episode step. +Figure 4. Impact of the other ego's speech on passive mapping vs. distance between the egos. + +diction task. AV-Floorplan [52] fares the best among all baselines. Its improvement over OccAnt [53] demonstrates the benefits of exploiting the spatial cues in audio for mapping and using an attention-based model to leverage the long- and short-range correlations in the audio-visual inputs. + +Our method outperforms all baselines. Its improvement over AV-Floorplan [52] underlines the efficacy of performing attention at different granularities—across modalities, within a single modality and within a single input—guided by our positional and modality type embeddings. It also generalizes to the real-world setting and retains its benefits over the baselines, even without retraining on the real-world data. However, we do observe a drop in performance gains, probably due to the large sim-to-real gap. + +Active mapping. Fig. 3 shows the active mapping performance as a function of episode progress. Employing naive heuristics for sampling, like Random or Greedy, is insufficient for high-quality mapping, emphasizing the high levels of redundancy in the visual frames. Unique-pose improves over both Random and Greedy, showing that sampling diverse viewpoints provides more information about the underlying scene geometry. + +Although the baselines make initial progress, they flatten quickly and our model eventually outperforms them all, on both real-world and simulated data (gains are again statistically significant, $p \leq 0.05$ ). This highlights the benefits of learning a smart policy that, given the audio streams and its past visual samples, understands the value of sampling a visual frame for mapping by taking cues from our novel reward. Besides, on the real-world data, we see better per + +![](images/cc04797f38573fc942b96e0d51e5c37a66bce51c35caabc21e0d304352baabe8.jpg) +Figure 5. Sample episodes for our active mapping model. While our policy samples only the salient visual frames, our mapper can both complete partially seen objects as well as anticipate objects never seen before in the sampled visuals (red boxes on the maps). + +formance margins over the baselines towards the end of episodes, showing that our policy can adaptively defer visual sampling to improve mapping. Owing to smart sampling, the per-episode reduction in the capture / processing cost for $B = 2$ is 74 Watts (W) / 7.2 GFLOPS in simulation and 37 W / 3.6 GFLOPS in the real world video. See Supp. for further analysis of our model's power and time cost. + +# 5.2. Model analysis + +Ablations. In Table 1 (bottom), we ablate the components of our model for passive mapping. Upon removing audio, our model experiences a large drop in mapping performance, which indicates that our model leverages complementary spatial cues in audio and vision. We also see a drop in the map quality when our model lacks access to the speech from the other ego $(E_i')$ . This shows that $E_i'$ 's speech can better reveal the more global scene geometry than $E_i$ 's own speech. Fig. 4 further shows that the impact of the other ego's speech becomes more prominent for larger inter-ego distances $(3 - 5\mathrm{m}$ vs. $1 - 3\mathrm{m})$ , in which case the two types of speech are dissimilar enough to carry complementary geometric cues, but reduces for even larger distances $(5\mathrm{m}$ or more), in which case $E_i'$ is too far for its speech to carry useful cues about $E_i$ 's local scene geometry. Moreover, unlike the ablation that does not perform shared mapping, our model benefits significantly from jointly attending to the observations of the egos and exploiting the complementary information in them—even though both models use the exact same audio-visual observations, including both speech from self and the other ego. Finally, without our modality type embeddings, our model does worse, showing their role in learning complementary features from vision and audio. + +For active mapping, Fig. 3 shows a drop in the mapping performance upon removing audio from the policy inputs. This implies that our policy exploits audio to reason about the level of redundancy in a new visual frame and improve + +the mapping quality vs. visual budget tradeoff. On the more challenging real-world setting, audio plays an even bigger role, as shown by the larger performance drop in Fig. 3b. + +See Supp. for similar results with 1) ambient sounds, 2) unheard speech, 3) higher budgets, 4) sensor noise, 5) larger target maps, 6) different ego initializations, and 7) multiple random data splits. + +Qualitative results. Fig. 5 shows two successful active mapping episodes of our method. Note how our model samples views that tend have to little visual overlap but are informative of the surrounding geometry (both occupied and free spaces). Besides, it is able to complete structures only partially visible in the sampled views, and more interestingly, leverage the synergy of audio and vision to anticipate unseen areas (red boxes on the occupancy maps in Fig. 5). + +Failure cases. We notice two common failure cases with active mapping: episodes where the egos stay at the same location, leading to very few informative visual frames to sample from; and episodes with highly unique visual samples at every trajectory step, in which case each sample is useful and our model behaves similar to Unique-pose or Greedy. For passive mapping, our model fails with very complex scenes that commonly have objects in spaces where both vision and audio cannot reach (e.g. narrow corners) + +# 6. Conclusion + +We introduce Chat2Map, a new task aimed at scene mapping using audio-visual feeds from egocentric conversations. We develop a novel approach for Chat2Map comprising a shared scene mapper and a visual sampling policy based on a novel reinforcement learner that smartly samples the visuals only when necessary. We show promising performance on both simulated and real-world data from over 80 scenes. + +Acknowledgements: UT Austin is supported in part by the IFML NSF AI Institute. KG is paid as a research scientist by Meta. + +# References + +[1] Ego-Noise Predictions for Echolocation in Wheeled Robots, volume ALIFE 2019: The 2019 Conference on Artificial Life of ALIFE 2022: The 2022 Conference on Artificial Life, 07 2019. 1, 2 +[2] Aaron Carroll and Gernot Heiser. An analysis of power consumption in a smartphone. In 2010 USENIX Annual Technical Conference (USENIX ATC 10), 2010. 2, 3 +[3] Angel Chang, Angela Dai, Thomas Funkhouser, Maciej Halber, Matthias Niessner, Manolis Savva, Shuran Song, Andy Zeng, and Yinda Zhang. Matterport3d: Learning from rgb-d data in indoor environments. arXiv preprint arXiv:1709.06158, 2017. 6 +[4] Devendra Singh Chaplot, Dhiraj Gandhi, Saurabh Gupta, Abhinav Gupta, and Ruslan Salakhutdinov. Learning to explore using active neural slam. In International Conference on Learning Representations, 2020. 1, 2, 6 +[5] Devendra Singh Chaplot, Dhiraj Prakashchand Gandhi, Abhinav Gupta, and Russ R Salakhutdinov. Object goal navigation using goal-oriented semantic exploration. Advances in Neural Information Processing Systems, 33:4247-4258, 2020. 2 +[6] Changan Chen, Ziad Al-Halah, and Kristen Grauman. Semantic audio-visual navigation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15516–15525, 2021. 2 +[7] Changan Chen, Ruohan Gao, Paul T. Calamia, and Kristen Grauman. Visual acoustic matching. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 18836-18846, 2022. 2 +[8] Changan Chen, Unnat Jain, Carl Schissler, Sebastia Vicenc Amengual Gari, Ziad Al-Halah, Vamsi Krishna Ithapu, Philip Robinson, and Kristen Grauman. Soundspaces: Audiovisual navigation in 3d environments. In ECCV, 2020. 1, 2, 6 +[9] Changan Chen, Sagnik Majumder, Ziad Al-Halah, Ruohan Gao, Santhosh Kumar Ramakrishnan, and Kristen Grauman. Learning to set waypoints for audio-visual navigation. In International Conference on Learning Representations, 2021. 2 +[10] Jiacheng Chen, Chen Liu, Jiaye Wu, and Yasutaka Furukawa. Floor-sp: Inverse cad for floorplans by sequential room-wise shortest path. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2661-2670, 2019. 1, 2 +[11] Tao Chen, Saurabh Gupta, and Abhinav Gupta. Learning exploration policies for navigation. arXiv preprint arXiv:1903.01959, 2019. 1, 2, 6 +[12] Yangyu Chen, Shuhui Wang, Weigang Zhang, and Qingming Huang. Less is more: Picking informative frames for video captioning. In Proceedings of the European conference on computer vision (ECCV), pages 358-373, 2018. 3 +[13] Jesper Christensen, Sascha Hornauer, and Stella Yu. Batvision - learning to see 3d spatial layout with two ears. In ICRA, 2020. 1 +[14] Jesper Haahr Christensen, Sascha Hornauer, and Stella X. Yu. Batvision: Learning to see 3d spatial layout with two ears. In 2020 IEEE International Conference on Robotics and Automation (ICRA), pages 1581-1587, 2020. 1, 2 +[15] Abhishek Das, Théophile Gervet, Joshua Romoff, Dhruv + +Batra, Devi Parikh, Michael G. Rabbat, and Joelle Pineau. Tarmac: Targeted multi-agent communication. ArXiv, abs/1810.11187, 2019. 2 +[16] Samyak Datta, Sameer Dharur, Vincent Cartillier, Ruta Desai, Mukul Khanna, Dhruv Batra, and Devi Parikh. Episodic memory question answering. 2022. 1 +[17] Victoria Dean, Shubham Tulsiani, and Abhinav Gupta. See, hear, explore: Curiosity via audio-visual association. Advances in Neural Information Processing Systems, 33:14961–14972, 2020. 2 +[18] Helisa Dhamo, Nassir Navab, and Federico Tombari. Object-driven multi-layer scene decomposition from a single image. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5369-5378, 2019. 2 +[19] Ivan Dokmanić, Reza Parhizkar, Andreas Walther, Yue M. Lu, and Martin Vetterli. Acoustic echoes reveal room shape. Proceedings of the National Academy of Sciences, 110(30):12186-12191, 2013. 3 +[20] Ainaz Eftekhar, Alexander Sax, Jitendra Malik, and Amir Zamir. Omnidata: A scalable pipeline for making multi-task mid-level vision datasets from 3d scans. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10786-10796, 2021. 6 +[21] Amine Elhafsi, Boris Ivanovic, Lucas Janson, and Marco Pavone. Map-predictive motion planning in unknown environments. In 2020 IEEE International Conference on Robotics and Automation (ICRA), pages 8552-8558. IEEE, 2020. 2 +[22] Itamar Eliakim, Zahi Cohen, Gábor Kósa, and Yossi Yovel. A fully autonomous terrestrial bat-like acoustic robot. PLoS Computational Biology, 14, 2018. 1, 2 +[23] Angelo Farina. Simultaneous measurement of impulse response and distortion with a swept-sine technique. Journal of The Audio Engineering Society, 2000. 6 +[24] Chuang Gan, Jeremy Schwartz, Seth Alter, Martin Schrimpf, James Traer, Julian De Freitas, Jonas Kubilius, Abhishek Bhandwaldar, Nick Haber, Megumi Sano, Kuno Kim, Elias Wang, Damian Mrowca, Michael Lingelbach, Aidan Curtis, Kevin T. Feigelis, Daniel M. Bear, Dan Gutfreund, David D. Cox, James J. DiCarlo, Josh H. McDermott, Joshua B. Tenenbaum, and Daniel L. K. Yamins. Threadworld: A platform for interactive multi-modal physical simulation. In NeurIPS Track on Datasets and Benchmarks, 2021. 1 +[25] Chuang Gan, Yiwei Zhang, Jiajun Wu, Boqing Gong, and Joshua B. Tenenbaum. Look, listen, and act: Towards audiovisual embodied navigation. In 2020 IEEE International Conference on Robotics and Automation (ICRA), pages 9701-9707, 2020. 2 +[26] Ruohan Gao, Changan Chen, Ziad Al-Halah, Carl Schissler, and Kristen Grauman. Visualechoes: Spatial image representation learning through echolocation. In Andrea Vedaldi, Horst Bischof, Thomas Brox, and Jan-Michael Frahm, editors, Computer Vision - ECCV 2020, pages 658-676, Cham, 2020. Springer International Publishing. 1, 2 +[27] Ruohan Gao, Tae-Hyun Oh, Kristen Grauman, and Lorenzo Torresani. Listen to look: Action recognition by previewing audio. 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10454-10464, 2020. 3 +[28] Kristen Grauman, Andrew Westbury, Eugene Byrne, Zachary Chavis, Antonino Furnari, Rohit Girdhar, Jackson Hamburger, + +Hao Jiang, Miao Liu, Xingyu Liu, Miguel Martin, Tushar Nagarajan, Ilija Radosavovic, Santhosh Kumar Ramakrishnan, Fiona Ryan, Jayant Sharma, Michael Wray, Mengmeng Xu, Eric Zhongcong Xu, Chen Zhao, Siddhant Bansal, Dhruv Batra, Vincent Cartillier, Sean Crane, Tien Do, Morrie Doulaty, Akshay Erapalli, Christoph Feichtenhofer, Adriano Fragomeni, Qichen Fu, Christian Fuegen, Abrham Gebreselasie, Cristina Gonzalez, James Hillis, Xuhua Huang, Yifei Huang, Wenqi Jia, Weslie Khoo, Jachym Kolar, Satwik Kottur, Anurag Kumar, Federico Landini, Chao Li, Yanghao Li, Zhenqiang Li, Karttikeya Mangalam, Raghava Modhugu, Jonathan Munro, Tullie Murrell, Takumi Nishiyasu, Will Price, Paola Ruiz Puentes, Merey Ramazanova, Leda Sari, Kiran Somasundaram, Audrey Southerland, Yusuke Sugano, Ruijie Tao, Minh Vo, Yuchen Wang, Xindi Wu, Takuma Yagi, Yunyi Zhu, Pablo Arbelaez, David Crandall, Dima Damen, Giovanni Maria Farinella, Bernard Ghanem, Vamsi Krishna Ithapu, C. V. Jawahar, Hanbyul Joo, Kris Kitani, Haizhou Li, Richard Newcombe, Aude Oliva, Hyun Soo Park, James M. Rehg, Yoichi Sato, Jianbo Shi, Mike Zheng Shou, Antonio Torralba, Lorenzo Torresani, Mingfei Yan, and Jitendra Malik. Ego4d: Around the World in 3,000 Hours of Egocentric Video. In IEEE/CVF Computer Vision and Pattern Recognition (CVPR), 2022. 1 +[29] Shariq Iqbal and Fei Sha. Coordinated exploration via intrinsic rewards for multi-agent reinforcement learning. arXiv preprint arXiv:1905.12127, 2019. 2 +[30] Max Jaderberg, Wojciech M. Czarnecki, Iain Dunning, Luke Marris, Guy Lever, Antonio Garcia Castañeda, Charlie Beattie, Neil C. Rabinowitz, Ari S. Morcos, Avraham Ruderman, Nicolas Sonnerat, Tim Green, Louise Deason, Joel Z. Leibo, David Silver, Demis Hassabis, Koray Kavukcuoglu, and Thore Graepel. Human-level performance in 3d multiplayer games with population-based reinforcement learning. Science, 364:859 - 865, 2019. 2 +[31] Unnat Jain, Luca Weihs, Eric Kolve, Ali Farhadi, Svetlana Lazebnik, Aniruddha Kembhavi, and Alexander G. Schwing. A cordial sync: Going beyond marginal policies for multiagent embodied tasks. ArXiv, abs/2007.04979, 2020. 2 +[32] Unnat Jain, Luca Weihs, Eric Kolve, Mohammad Rastegari, Svetlana Lazebnik, Ali Farhadi, Alexander G. Schwing, and Aniruddha Kembhavi. Two body problem: Collaborative visual task completion. 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6682-6692, 2019. 2 +[33] Oğuzhan Fatih Kar, Teresa Yeo, Andrei Atanov, and Amir Zamir. 3d common corruptions and data augmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18963-18974, 2022. 6 +[34] Kapil Katyal, Katie Popek, Chris Paxton, Phil Burlina, and Gregory D. Hager. Uncertainty-aware occupancy map prediction using generative networks for robot navigation. In 2019 International Conference on Robotics and Automation (ICRA), pages 5453-5459, 2019. 2 +[35] Hansung Kim, Luca Remaggi, Philip JB Jackson, Filippo Maria Fazi, and Adrian Hilton. 3d room geometry reconstruction using audio-visual sensors. In 2017 International Conference on 3D Vision (3DV), pages 621–629, 2017. 2 +[36] Bruno Korbar, Du Tran, and Lorenzo Torresani. Scsampler: + +Sampling salient clips from video for efficient action recognition. 2019 IEEE/CVF International Conference on Computer Vision (ICCV), pages 6231-6241, 2019. 3 +[37] Robert LiKamWa, Zhen Wang, Aaron Carroll, Felix Xiaozhu Lin, and Lin Zhong. Draining our glass: An energy and heat characterization of google glass. In Proceedings of 5th Asia-Pacific Workshop on Systems, pages 1-7, 2014. 2, 3 +[38] Jintao Lin, Haodong Duan, Kai Chen, Dahua Lin, and Limin Wang. Ocsampler: Compressing videos to one clip with single-step sampling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13894-13903, 2022. 3 +[39] David B. Lindell, Gordon Wetzstein, and Vladlen Koltun. Acoustic non-line-of-sight imaging. In 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6773-6782, 2019. 1, 2 +[40] Chen Liu, Jiaye Wu, and Yasutaka Furukawa. Floornet: A unified framework for floorplan reconstruction from 3d scans. In Proceedings of the European conference on computer vision (ECCV), pages 201-217, 2018. 1, 2 +[41] Wenxin Liu, David Caruso, Eddy Ig, Jing Dong, Anastasios I Mourikis, Kostas Daniilidis, Vijay Kumar, and Jakob Engel. Tlio: Tight learned inertial odometry. IEEE Robotics and Automation Letters, 5(4):5653-5660, 2020. 6 +[42] Sagnik Majumder, Ziad Al-Halah, and Kristen Grauman. Move2hear: Active audio-visual source separation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 275–285, 2021. 2 +[43] Sagnik Majumder, Ziad Al-Halah, and Kristen Grauman. Active audio-visual separation of dynamic sound sources. In European Conference on Computer Vision. Springer, 2022. 2 +[44] Sagnik Majumder, Changan Chen, Ziad Al-Halah, and Kristen Grauman. Few-shot audio-visual learning of environment acoustics. In Thirty-Sixth Conference on Neural Information Processing Systems, 2022. 2, 5 +[45] Medhini Narasimhan, Erik Wijmans, Xinlei Chen, Trevor Darrell, Dhruv Batra, Devi Parikh, and Amanpreet Singh. Seeing the un-scene: Learning amodal semantic maps for room navigation. In European Conference on Computer Vision, pages 513-529. Springer, 2020. 2 +[46] Simon T O'Callaghan and Fabio T Ramos. Gaussian process occupancy maps*. Int. J. Rob. Res., 31(1):42-62, jan 2012. 2 +[47] Brian Okorn, Xuehan Xiong, and Burcu Akinci. Toward automated modeling of floor plans. In 3D PVT, 2009. 1, 2 +[48] Andrew Owens, Phillip Isola, Josh McDermott, Antonio Torralba, Edward H. Adelson, and William T. Freeman. Visually indicated sounds. In 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 2405-2413, 2016. 2 +[49] Vassil Panayotov, Guoguo Chen, Daniel Povey, and Sanjeev Khudanpur. Librispeech: An asr corpus based on public domain audio books. In 2015 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 5206-5210, 2015. 6 +[50] Rameswar Panda, Chun-Fu Richard Chen, Quanfu Fan, Ximeng Sun, Kate Saenko, Aude Oliva, and Rogerio Feris. Adamml: Adaptive multi-modal learning for efficient video recognition. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7576-7585, 2021. 3 + +[51] Shivansh Patel, Saim Wani, Unnat Jain, Alexander G. Schwing, Svetlana Lazebnik, Manolis Savva, and Angel X. Chang. Interpretation of emergent communication in heterogeneous collaborative embodied agents. 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 15993-15943, 2021. 2 +[52] Senthil Purushwalkam, Sebastia Vicenc Amengual Gari, Vamsi Krishna Ithapu, Carl Schissler, Philip Robinson, Abhinav Gupta, and Kristen Grauman. Audio-visual floorplan reconstruction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1183–1192, 2021. 1, 2, 3, 6, 7 +[53] Santhosh K Ramakrishnan, Ziad Al-Halah, and Kristen Grauman. Occupancy anticipation for efficient exploration and navigation. In European Conference on Computer Vision, pages 400-418. Springer, 2020. 1, 2, 6, 7 +[54] Fabio Ramos and Lionel Ott. Hilbert maps: Scalable continuous occupancy mapping with stochastic gradient descent. The International Journal of Robotics Research, 35(14):1717-1730, 2016. 2 +[55] Ethan Rublee, Vincent Rabaud, Kurt Konolige, and Gary Bradski. Orb: An efficient alternative to sift or surf. In 2011 International Conference on Computer Vision, pages 2564-2571, 2011. 6 +[56] João Machado Santos, David Portugal, and Rui P. Rocha. An evaluation of 2d slam techniques available in robot operating system. In 2013 IEEE International Symposium on Safety, Security, and Rescue Robotics (SSRR), pages 1-6, 2013. 2 +[57] Manolis Savva, Abhishek Kadian, Oleksandr Maksymets, Yili Zhao, Erik Wijmans, Bhavana Jain, Julian Straub, Jia Liu, Vladlen Koltun, Jitendra Malik, et al. Habitat: A platform for embodied ai research. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9339-9347, 2019. 6 +[58] Carl Schissler, Christian Loftin, and Dinesh Manocha. Acoustic classification and optimization for multi-modal rendering of real-world scenes. IEEE Transactions on Visualization and Computer Graphics, 24:1246-1259, 2018. 2 +[59] Ransalu Senanayake, Thushan Ganegedara, and Fabio Ramos. Deep occupancy maps: a continuous mapping technique for dynamic environments. In NIPS 2017 Workshop MLITS, 2017. +[60] Rakesh Shrestha, Fei-Peng Tian, Wei Feng, Ping Tan, and Richard Vaughan. Learned map prediction for enhanced mobile robot exploration. In 2019 International Conference on Robotics and Automation (ICRA), pages 1197-1204, 2019. 2 +[61] Jascha Sohl-Dickstein, Santani Teng, Benjamin M. Gaub, Chris C. Rodgers, Crystal Li, Michael R. DeWeese, and Nicol S. Harper. A device for human ultrasonic echolocation. IEEE Transactions on Biomedical Engineering, 62(6):1526-1534, 2015. 1, 2 +[62] Wei Sui, Lingfeng Wang, Bin Fan, Hongfei Xiao, Huaiyu Wu, and Chunhong Pan. Layer-wise floorplan extraction for automatic urban building reconstruction. IEEE Transactions on Visualization and Computer Graphics, 22(3):1261-1277, 2016. 2 +[63] Maitreya Suin and A. N. Rajagopalan. An efficient framework for dense video captioning. Proceedings of the AAAI + +Conference on Artificial Intelligence, 34(07):12039-12046, Apr. 2020. 3 +[64] Cheng Sun, Chi-Wei Hsiao, Min Sun, and Hwann-Tzong Chen. Horizonnet: Learning room layout with 1d representation and piano stretch data augmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1047–1056, 2019. 2 +[65] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 5 +[66] Erik Wijmans, Abhishek Kadian, Ari S. Morcos, Stefan Lee, Irfan Essa, Devi Parikh, Manolis Savva, and Dhruv Batra. Dd-ppo: Learning near-perfect pointgoal navigators from 2.5 billion frames. In ICLR, 2020. 6 +[67] Wenming Wu, Xiao-Ming Fu, Rui Tang, Yuhan Wang, YuHao Qi, and Ligang Liu. Data-driven interior plan generation for residential buildings. ACM Trans. Graph., 38(6), nov 2019. 2 +[68] Zuxuan Wu, Caiming Xiong, Chih-Yao Ma, Richard Socher, and Larry S Davis. Adaframe: Adaptive frame selection for fast video recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1278-1287, 2019. 3 +[69] Shang-Ta Yang, Fu-En Wang, Chi-Han Peng, Peter Wonka, Min Sun, and Hung-Kuo Chu. Dula-net: A dual-projection network for estimating room layouts from a single rgb panorama. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3363–3372, 2019. 2 +[70] Mao Ye, Yu Zhang, Ruigang Yang, and Dinesh Manocha. 3d reconstruction in the presence of glasses by acoustic and stereo fusion. In 2015 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 4885-4893, 2015. 2 +[71] Serena Yeung, Olga Russakovsky, Greg Mori, and Li Fei-Fei. End-to-end learning of action detection from frame glimpses in videos. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2678–2687, 2016. 3 +[72] Yinfeng Yu, Wenbing Huang, Fuchun Sun, Changan Chen, Yikai Wang, and Xiaohong Liu. Sound adversarial audiovisual navigation. In International Conference on Learning Representations, 2022. 2 +[73] Zhoutong Zhang, Jiajun Wu, Qiujia Li, Zhengjia Huang, James Traer, Josh H. McDermott, Joshua B. Tenenbaum, and William T. Freeman. Generative modeling of audible shapes for object perception. 2017 IEEE International Conference on Computer Vision (ICCV), pages 1260-1269, 2017. 2 +[74] Chuhang Zou, Alex Colburn, Qi Shan, and Derek Hoiem. Layoutnet: Reconstructing the 3d room layout from a single rgb image. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2051-2059, 2018. 2 +[75] Chuhang Zou, Jheng-Wei Su, Chi-Han Peng, Alex Colburn, Qi Shan, Peter Wonka, Hung-Kuo Chu, and Derek Hoiem. Manhattan room layout reconstruction from a single $360^{\circ}$ image: A comparative study of state-of-the-art methods. International Journal of Computer Vision, 129(5):1410-1431, 2021. 2 \ No newline at end of file diff --git a/2023/Chat2Map_ Efficient Scene Mapping From Multi-Ego Conversations/images.zip b/2023/Chat2Map_ Efficient Scene Mapping From Multi-Ego Conversations/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..db05c34959e0569b3bf05a09b61a83c7754209f8 --- /dev/null +++ b/2023/Chat2Map_ Efficient Scene Mapping From Multi-Ego Conversations/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d753871def9ba89e9bbf6f4f0867bad8c0d454b22b4740259a509a4e3a14b32d +size 372151 diff --git a/2023/Chat2Map_ Efficient Scene Mapping From Multi-Ego Conversations/layout.json b/2023/Chat2Map_ Efficient Scene Mapping From Multi-Ego Conversations/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..0136fff8644fdeba5cd39c1bfd29c87d4386f78f --- /dev/null +++ b/2023/Chat2Map_ Efficient Scene Mapping From Multi-Ego Conversations/layout.json @@ -0,0 +1,11752 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 87, + 103, + 507, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 87, + 103, + 507, + 121 + ], + "spans": [ + { + "bbox": [ + 87, + 103, + 507, + 121 + ], + "type": "text", + "content": "Chat2Map: Efficient Scene Mapping from Multi-Ego Conversations" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 113, + 142, + 479, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 113, + 142, + 479, + 185 + ], + "spans": [ + { + "bbox": [ + 113, + 142, + 479, + 185 + ], + "type": "text", + "content": "Sagnik Majumder" + }, + { + "bbox": [ + 113, + 142, + 479, + 185 + ], + "type": "inline_equation", + "content": "^{1,2,3}" + }, + { + "bbox": [ + 113, + 142, + 479, + 185 + ], + "type": "text", + "content": " Hao Jiang" + }, + { + "bbox": [ + 113, + 142, + 479, + 185 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 113, + 142, + 479, + 185 + ], + "type": "text", + "content": " Pierre Moulon" + }, + { + "bbox": [ + 113, + 142, + 479, + 185 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 113, + 142, + 479, + 185 + ], + "type": "text", + "content": " Ethan Henderson" + }, + { + "bbox": [ + 113, + 142, + 479, + 185 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 113, + 142, + 479, + 185 + ], + "type": "text", + "content": " Paul Calamia" + }, + { + "bbox": [ + 113, + 142, + 479, + 185 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 113, + 142, + 479, + 185 + ], + "type": "text", + "content": " Kristen Grauman" + }, + { + "bbox": [ + 113, + 142, + 479, + 185 + ], + "type": "inline_equation", + "content": "^{1,3*}" + }, + { + "bbox": [ + 113, + 142, + 479, + 185 + ], + "type": "text", + "content": " Vamsi Krishna Ithapu" + }, + { + "bbox": [ + 113, + 142, + 479, + 185 + ], + "type": "inline_equation", + "content": "^{2*}" + }, + { + "bbox": [ + 113, + 142, + 479, + 185 + ], + "type": "text", + "content": " UT Austin" + }, + { + "bbox": [ + 113, + 142, + 479, + 185 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 113, + 142, + 479, + 185 + ], + "type": "text", + "content": " Reality Labs Research, Meta FAIR" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "spans": [ + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 239, + 290, + 501 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 239, + 290, + 501 + ], + "spans": [ + { + "bbox": [ + 46, + 239, + 290, + 501 + ], + "type": "text", + "content": "Can conversational videos captured from multiple egocentric viewpoints reveal the map of a scene in a cost-efficient way? We seek to answer this question by proposing a new problem: efficiently building the map of a previously unseen 3D environment by exploiting shared information in the egocentric audio-visual observations of participants in a natural conversation. Our hypothesis is that as multiple people (\"egos\") move in a scene and talk among themselves, they receive rich audio-visual cues that can help uncover the unseen areas of the scene. Given the high cost of continuously processing egocentric visual streams, we further explore how to actively coordinate the sampling of visual information, so as to minimize redundancy and reduce power use. To that end, we present an audio-visual deep reinforcement learning approach that works with our shared scene mapper to selectively turn on the camera to efficiently chart out the space. We evaluate the approach using a state-of-the-art audio-visual simulator for 3D scenes as well as real-world video. Our model outperforms previous state-of-the-art mapping methods, and achieves an excellent cost-accuracy tradeoff. Project: http://vision.cs.utexas.edu/projects/chat2map." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 518, + 128, + 530 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 518, + 128, + 530 + ], + "spans": [ + { + "bbox": [ + 47, + 518, + 128, + 530 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 538, + 288, + 670 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 538, + 288, + 670 + ], + "spans": [ + { + "bbox": [ + 46, + 538, + 288, + 670 + ], + "type": "text", + "content": "The spatial layout of the environment around us is fundamental to understanding our physical context. By representing the walls, furniture, and other major structures in a space, scene maps ground activity and objects in a persistent frame of reference, facilitating high-level reasoning for many downstream applications in augmented reality (AR) and robotics. For example, episodic memory [16, 28] aims to relocalize lost objects observed in first-person video (where are my keys?); floorplan estimation [10, 40, 47] aims to chart out the area and shapes of complex buildings; navigating agents try to discover routes in unfamiliar spaces [4, 11, 53]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 670, + 288, + 694 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 670, + 288, + 694 + ], + "spans": [ + { + "bbox": [ + 47, + 670, + 288, + 694 + ], + "type": "text", + "content": "While traditional computer vision approaches for mapping (e.g., visual SLAM) are highly effective when extensive" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 308, + 211, + 547, + 405 + ], + "blocks": [ + { + "bbox": [ + 308, + 211, + 547, + 405 + ], + "lines": [ + { + "bbox": [ + 308, + 211, + 547, + 405 + ], + "spans": [ + { + "bbox": [ + 308, + 211, + 547, + 405 + ], + "type": "image", + "image_path": "02b86ba0a5d0fef8f1d91542217a239f3c2a45bea27902b7cd21a1f522f16f87.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 413, + 547, + 468 + ], + "lines": [ + { + "bbox": [ + 305, + 413, + 547, + 468 + ], + "spans": [ + { + "bbox": [ + 305, + 413, + 547, + 468 + ], + "type": "text", + "content": "Figure 1. Given egocentric audio-visual observations from multiple people wearing AR glasses and moving and conversing (left), we aim to accurately map the scene (right). To mitigate cost, our model receives audio continuously but learns to selectively employ the ego cameras only when the visual input is expected to be informative." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 483, + 547, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 483, + 547, + 579 + ], + "spans": [ + { + "bbox": [ + 304, + 483, + 547, + 579 + ], + "type": "text", + "content": "exposure to the environment is possible, in many real-world scenarios only a fraction of the space is observed by the camera. Recent work shows the promise of sensing 3D spaces with both sight and sound [8, 13, 24, 26, 52]: listening to echoes bounce around the room can reveal the depth and shape of surrounding surfaces, and even help extrapolate a floorplan beyond the camera's field of view or behind occluded objects [52]." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 580, + 547, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 580, + 547, + 676 + ], + "spans": [ + { + "bbox": [ + 304, + 580, + 547, + 676 + ], + "type": "text", + "content": "While we are inspired by these advances, they also have certain limitations. Often systems will emit sounds (e.g., a frequency sweep) into the environment to ping for spatial information [1, 13, 14, 22, 26, 39, 52, 61], which is intrusive if done around people. Furthermore, existing audio-visual models assume that the camera is always on grabbing new frames, which is wasteful if not intractable, particularly on lightweight, low-power computing devices in AR settings." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 677, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 677, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 677, + 547, + 713 + ], + "type": "text", + "content": "We introduce Chat2Map, a new scene mapping task aimed at eliminating these challenges. In the proposed setting, multiple people converse as they move casually through" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 702, + 124, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 702, + 124, + 712 + ], + "spans": [ + { + "bbox": [ + 58, + 702, + 124, + 712 + ], + "type": "text", + "content": "*Equal contribution" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "10554" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 288, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 288, + 133 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 288, + 133 + ], + "type": "text", + "content": "the scene while wearing AR glasses equipped with an egocentric camera, microphones, and potentially other sensors (e.g., for odometry).1 Given their egocentric audio-visual data streams, the goal is to infer the ground-plane occupancy map for the larger environment around them. See Figure 1." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 133, + 288, + 215 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 133, + 288, + 215 + ], + "spans": [ + { + "bbox": [ + 46, + 133, + 288, + 215 + ], + "type": "text", + "content": "We observe that audio-visual data from the egos' interactions will naturally reflect scene structure. First, as they walk and talk, their movements reveal spaces like corridors, doorways, and large rooms, in both modalities. Second, the speech captured by the device-wearer's cameras and microphones can be localized to different speakers, which, compared to active sound emission, is non-intrusive." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 216, + 288, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 216, + 288, + 443 + ], + "spans": [ + { + "bbox": [ + 46, + 216, + 288, + 443 + ], + "type": "text", + "content": "To realize this vision, we develop a novel approach to efficient scene mapping from multi-ego conversations. Our approach has two key elements: a shared scene mapper and a visual sampling policy. For the former, we devise a transformer-based mapper that incorporates the multiple data streams to infer a map beyond the directly observed areas, and, most importantly, that enables communication among the egos about their observations and states in the 3D space to improve mapping accuracy. For the latter, our idea is to relax the common assumption of an \"always-on\" camera, and instead actively select when to sample visual frames from any one of the ego cameras. Intuitively, certain regions where the egos move will be more or less important for mapping (e.g., corners of the room, doors). We train a sampling policy with deep reinforcement learning that activates the visual feed only when it is anticipated to complement the continuous audio feed. This is a cost-conscious approach, mindful that switching on a camera is much more power consuming than sensing audio with microphones [2, 37]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 444, + 288, + 538 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 444, + 288, + 538 + ], + "spans": [ + { + "bbox": [ + 46, + 444, + 288, + 538 + ], + "type": "text", + "content": "We demonstrate our approach using a state-of-the-art audio-visual simulator for 3D scenes [8] as well as real-world video input. We can successfully map an unfamiliar environment given only partial visibility via multiple conversing people moving about the scene. Compared to sampling all visual frames, our model reduces the visual capture and processing by " + }, + { + "bbox": [ + 46, + 444, + 288, + 538 + ], + "type": "inline_equation", + "content": "87.5\\%" + }, + { + "bbox": [ + 46, + 444, + 288, + 538 + ], + "type": "text", + "content": " while the mapping accuracy declines only marginally " + }, + { + "bbox": [ + 46, + 444, + 288, + 538 + ], + "type": "inline_equation", + "content": "(\\sim 9\\%)" + }, + { + "bbox": [ + 46, + 444, + 288, + 538 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 539, + 288, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 539, + 288, + 635 + ], + "spans": [ + { + "bbox": [ + 46, + 539, + 288, + 635 + ], + "type": "text", + "content": "Our main contributions are 1) we define the task of efficient and shared scene mapping from multi-ego conversations, a new direction for AR/VR research; 2) we present the first approach to begin tackling this task, specifically an RL-based framework that integrates an audio-visual scene mapper and a smart visual sampling policy; and 3) we rigorously experiment with a variety of environments (both in simulation and the real world), speech sounds, and use cases." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 643, + 134, + 656 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 643, + 134, + 656 + ], + "spans": [ + { + "bbox": [ + 47, + 643, + 134, + 656 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 663, + 287, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 663, + 287, + 688 + ], + "spans": [ + { + "bbox": [ + 47, + 663, + 287, + 688 + ], + "type": "text", + "content": "Visual scene mapping. Past works tackle scene mapping using 3D Manhattan layouts [18, 64, 69, 74, 75], detailed" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 72, + 547, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 547, + 312 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 547, + 312 + ], + "type": "text", + "content": "floorplans [10, 40, 47, 62, 67], occupancy [21, 34, 46, 54, 59, 60], and semantic maps [45]. Manhattan layouts include structured outputs like scene boundaries [64, 74, 75], corners [74, 75], and floor/ceilings [69, 75], but do not generalize to unseen environment regions. Floorplan estimation methods use dense scans of 3D scenes to predict geometric (walls, exterior/ interior) and semantic layouts (room type, object type, etc.)., rely on extensive human walkthroughs with RGB-D [10, 40] or 3D point cloud [47, 62] scans, and are usually limited to polygonal layouts [10, 40, 47, 62, 67]. Occupancy maps traditionally rely on wide field-of-view (FoV) LiDAR scanners [56] or evaluate on simple 2D environments without non-wall obstacles [21, 34, 60, 60]. More recent methods [4, 5, 11, 53] train an embodied agent to explore and build top-down maps of more complex scenes using RGB-D. On the contrary, our method uses both vision and audio from the observations of a group of conversing people for mapping. Rather than steer the camera of a robot to map the scene, our task requires processing passive video from human camera wearers." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 318, + 547, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 318, + 547, + 426 + ], + "spans": [ + { + "bbox": [ + 305, + 318, + 547, + 426 + ], + "type": "text", + "content": "Audio-visual scene mapping. To our knowledge, the only prior work to translate audio-visual inputs into a general (arbitrarily shaped) floorplan maps is AV-Floorplan [52]. Unlike AV-Floorplan, our method infers maps from speech in natural human conversations, which avoids emitting intrusive frequency sweep signals to generate echoes. In addition, a key goal of our work is to reduce mapping cost by skipping redundant visual frames. Our experiments demonstrate the benefits of our model design over AV-Floorplan [52]." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 432, + 548, + 600 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 432, + 548, + 600 + ], + "spans": [ + { + "bbox": [ + 305, + 432, + 548, + 600 + ], + "type": "text", + "content": "Audio(-visual) spatial understanding. More broadly, beyond the mapping task, various methods leverage audio for geometric and material information about the 3D scene and its constituent objects. Prior work relies on acoustic reflections to estimate the shape of objects [39]. Echolocation is used in robotics to estimate proximity to surrounding surfaces [1, 14, 22, 61]. Together, vision and audio can better reveal the shape and materials of objects [48, 58, 73], self-supervise imagery [26], and improve depth sensing [35, 70]. Recent work exploits correlations between spatial audio and imagery to reason about scene acoustics [7, 44] or aid active embodied navigation [6, 9, 17, 25, 72] and source separation [42, 43]. No prior work intelligently selects images to capture during conversations to efficiently map a scene." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 605, + 548, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 605, + 548, + 715 + ], + "spans": [ + { + "bbox": [ + 305, + 605, + 548, + 715 + ], + "type": "text", + "content": "Multi-agent spatial understanding. There is existing work [15, 29, 31, 32, 51] in the visual multi-agent reinforcement learning (MARL) community that learns collaborative agents for performing tasks like relocating furniture [31, 32], playing 3D multi-player games [30], coordinated scene exploration [29], or multi-object navigation [51]. In such settings, the collaborative agents actively interact with the environment to learn a shared scene representation for successfully completing their task. In contrast, we aim to learn a" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 45, + 693, + 287, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 693, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 45, + 693, + 287, + 713 + ], + "type": "text", + "content": "1 Throughout, we call each person participating in the conversation an \"ego\" for short." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10555" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 288, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 288, + 108 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 288, + 108 + ], + "type": "text", + "content": "shared geometric map of a 3D scene given passive observations that come from the trajectories chosen by a group of people involved in a natural conversation." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 115, + 288, + 247 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 115, + 288, + 247 + ], + "spans": [ + { + "bbox": [ + 46, + 115, + 288, + 247 + ], + "type": "text", + "content": "Efficient visual sampling in video. Efficient visual sampling has been studied in the context of video recognition [27, 36, 38, 68, 71] and summarization [12, 63] with the goal of selectively processing informative frames, which can both reduce computational cost and improve recognition. More closely related to our approach are methods that use audio for the decision-making [27, 36, 50]. Different from the above, we use efficient visual sampling in the context of mapping scenes. Furthermore, in our case an online sampling decision needs to be made at every step before looking at the current visual frame (or frames from future steps)." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 259, + 216, + 272 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 259, + 216, + 272 + ], + "spans": [ + { + "bbox": [ + 47, + 259, + 216, + 272 + ], + "type": "text", + "content": "3. Chat2Map Task Formulation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 280, + 288, + 303 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 280, + 288, + 303 + ], + "spans": [ + { + "bbox": [ + 46, + 280, + 288, + 303 + ], + "type": "text", + "content": "We propose a novel task: efficient and shared mapping of scenes from multi-ego conversations." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 304, + 288, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 304, + 288, + 509 + ], + "spans": [ + { + "bbox": [ + 46, + 304, + 288, + 509 + ], + "type": "text", + "content": "Without loss of generality, we consider two egos, " + }, + { + "bbox": [ + 46, + 304, + 288, + 509 + ], + "type": "inline_equation", + "content": "E_{1}" + }, + { + "bbox": [ + 46, + 304, + 288, + 509 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 304, + 288, + 509 + ], + "type": "inline_equation", + "content": "E_{2}" + }, + { + "bbox": [ + 46, + 304, + 288, + 509 + ], + "type": "text", + "content": ", each wearing AR glasses equipped with an RGB-D camera and a multi-channel microphone array. The egos have a conversation and move around in an unmapped 3D environment. Each conversation is " + }, + { + "bbox": [ + 46, + 304, + 288, + 509 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 46, + 304, + 288, + 509 + ], + "type": "text", + "content": " steps long. At each step " + }, + { + "bbox": [ + 46, + 304, + 288, + 509 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 304, + 288, + 509 + ], + "type": "text", + "content": ", the ego " + }, + { + "bbox": [ + 46, + 304, + 288, + 509 + ], + "type": "inline_equation", + "content": "E_{i}" + }, + { + "bbox": [ + 46, + 304, + 288, + 509 + ], + "type": "text", + "content": "'s glasses receive an observation " + }, + { + "bbox": [ + 46, + 304, + 288, + 509 + ], + "type": "inline_equation", + "content": "\\mathcal{O}_{i,t} = (\\mathcal{V}_{i,t}, S_{i,t}, P_{i,t}, S_{i,t}', P_{i,t}')" + }, + { + "bbox": [ + 46, + 304, + 288, + 509 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 46, + 304, + 288, + 509 + ], + "type": "inline_equation", + "content": "\\mathcal{V}_{i,t}" + }, + { + "bbox": [ + 46, + 304, + 288, + 509 + ], + "type": "text", + "content": " is the " + }, + { + "bbox": [ + 46, + 304, + 288, + 509 + ], + "type": "inline_equation", + "content": "90^{\\circ}" + }, + { + "bbox": [ + 46, + 304, + 288, + 509 + ], + "type": "text", + "content": " FOV RGB-D image and " + }, + { + "bbox": [ + 46, + 304, + 288, + 509 + ], + "type": "inline_equation", + "content": "S_{i,t}" + }, + { + "bbox": [ + 46, + 304, + 288, + 509 + ], + "type": "text", + "content": " is the speech waveform uttered by " + }, + { + "bbox": [ + 46, + 304, + 288, + 509 + ], + "type": "inline_equation", + "content": "E_{i}" + }, + { + "bbox": [ + 46, + 304, + 288, + 509 + ], + "type": "text", + "content": ", as observed from its pose " + }, + { + "bbox": [ + 46, + 304, + 288, + 509 + ], + "type": "inline_equation", + "content": "P_{i,t} = (x_{i,t}, y_{i,t}, \\theta_{i,t})" + }, + { + "bbox": [ + 46, + 304, + 288, + 509 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 304, + 288, + 509 + ], + "type": "inline_equation", + "content": "(x_{i,t}, y_{i,t})" + }, + { + "bbox": [ + 46, + 304, + 288, + 509 + ], + "type": "text", + "content": " denotes its location and " + }, + { + "bbox": [ + 46, + 304, + 288, + 509 + ], + "type": "inline_equation", + "content": "\\theta_{i,t}" + }, + { + "bbox": [ + 46, + 304, + 288, + 509 + ], + "type": "text", + "content": " denotes its orientation in the 3D scene. " + }, + { + "bbox": [ + 46, + 304, + 288, + 509 + ], + "type": "inline_equation", + "content": "S_{i,t}'" + }, + { + "bbox": [ + 46, + 304, + 288, + 509 + ], + "type": "text", + "content": " is the speech of the other ego " + }, + { + "bbox": [ + 46, + 304, + 288, + 509 + ], + "type": "inline_equation", + "content": "E_{i}'" + }, + { + "bbox": [ + 46, + 304, + 288, + 509 + ], + "type": "text", + "content": " (the other person involved in the conversation), as perceived by " + }, + { + "bbox": [ + 46, + 304, + 288, + 509 + ], + "type": "inline_equation", + "content": "E_{i}" + }, + { + "bbox": [ + 46, + 304, + 288, + 509 + ], + "type": "text", + "content": " (note, the speech sounds different depending on the listener position), and " + }, + { + "bbox": [ + 46, + 304, + 288, + 509 + ], + "type": "inline_equation", + "content": "P_{i,t}'" + }, + { + "bbox": [ + 46, + 304, + 288, + 509 + ], + "type": "text", + "content": " is " + }, + { + "bbox": [ + 46, + 304, + 288, + 509 + ], + "type": "inline_equation", + "content": "E_{i}'" + }, + { + "bbox": [ + 46, + 304, + 288, + 509 + ], + "type": "text", + "content": "'s pose relative to " + }, + { + "bbox": [ + 46, + 304, + 288, + 509 + ], + "type": "inline_equation", + "content": "E_{i}" + }, + { + "bbox": [ + 46, + 304, + 288, + 509 + ], + "type": "text", + "content": ". Modern AR glasses, like Bose Frames or Facebook Aria already support capturing such multi-sensory observations, making it possible to have a real-world instantiation of our task." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 510, + 288, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 510, + 288, + 700 + ], + "spans": [ + { + "bbox": [ + 46, + 510, + 288, + 700 + ], + "type": "text", + "content": "Given the real-time observation stream " + }, + { + "bbox": [ + 46, + 510, + 288, + 700 + ], + "type": "inline_equation", + "content": "\\mathcal{O}" + }, + { + "bbox": [ + 46, + 510, + 288, + 700 + ], + "type": "text", + "content": " for the egos, where " + }, + { + "bbox": [ + 46, + 510, + 288, + 700 + ], + "type": "inline_equation", + "content": "\\mathcal{O} = \\{\\mathcal{O}_{i,t}:i = 1,\\dots ,2,t = 1\\dots T\\}" + }, + { + "bbox": [ + 46, + 510, + 288, + 700 + ], + "type": "text", + "content": " and a total budget of visual frames " + }, + { + "bbox": [ + 46, + 510, + 288, + 700 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 46, + 510, + 288, + 700 + ], + "type": "text", + "content": " we aim to learn a model that can accurately estimate the top-down occupancy map " + }, + { + "bbox": [ + 46, + 510, + 288, + 700 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 46, + 510, + 288, + 700 + ], + "type": "text", + "content": " of the scene without exceeding the visual budget. The fixed budget represents allotting a fixed amount of battery to the energy-intensive visual capture [37], so that other on-device routines are not affected. We assume the first visual frames (at " + }, + { + "bbox": [ + 46, + 510, + 288, + 700 + ], + "type": "inline_equation", + "content": "t = 1" + }, + { + "bbox": [ + 46, + 510, + 288, + 700 + ], + "type": "text", + "content": " ) for both egos to be used by the model. Thus we aim to learn a policy that samples " + }, + { + "bbox": [ + 46, + 510, + 288, + 700 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 46, + 510, + 288, + 700 + ], + "type": "text", + "content": " frames from " + }, + { + "bbox": [ + 46, + 510, + 288, + 700 + ], + "type": "inline_equation", + "content": "2*(T - 1)" + }, + { + "bbox": [ + 46, + 510, + 288, + 700 + ], + "type": "text", + "content": " choices—which are not considered a batch, but rather unfold in sequence—and a mapper that predicts the scene map given the sampled frames. Our goal is to build a model that samples the expensive visual frames only when absolutely needed for scene mapping. This is captured by the constraint " + }, + { + "bbox": [ + 46, + 510, + 288, + 700 + ], + "type": "inline_equation", + "content": "1\\leq B\\ll 2*(T - 1)" + }, + { + "bbox": [ + 46, + 510, + 288, + 700 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 59, + 701, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 701, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 59, + 701, + 288, + 713 + ], + "type": "text", + "content": "There are three important aspects to our task. First, it" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 72, + 547, + 288 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 288 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 288 + ], + "type": "text", + "content": "requires learning from both vision and audio. While the visual signal carries rich information about the local scene geometry, there can be a high amount of redundancy in the visual feed captured during a conversation (e.g., the egos may visit the same location more than once or change their viewpoint only marginally). Second, not only does the long-range nature of audio help uncover the global scene properties [19, 52] like shape and size—beyond what is visible in images—but also we can exploit audio to undersample the visual frames, thereby reducing the cost of capturing and processing sensory inputs for mapping. Third, shared mapping of a scene allows jointly leveraging the complementary information in the audio (speech) from the self and other egos, and the synergy of the audio-visual cues from multiple egos. These insights form the basis of our key hypothesis in this task—selectively sampling visual frames during a conversation involving egos that share information with each other can facilitate efficient mapping of a scene." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 298, + 372, + 312 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 298, + 372, + 312 + ], + "spans": [ + { + "bbox": [ + 305, + 298, + 372, + 312 + ], + "type": "text", + "content": "4. Approach" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 319, + 547, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 319, + 547, + 426 + ], + "spans": [ + { + "bbox": [ + 304, + 319, + 547, + 426 + ], + "type": "text", + "content": "We solve the task by learning a model that estimates the scene map given the egos' audio-visual observations and also sequentially decides when to sample visual frames for mapping given the audio stream, ego poses, and previously sampled frames, if any. Here, \"sampling\" refers to individually deciding for each ego whether to use its camera or not to capture the visuals at every step of its trajectory in the scene. The sampling is preemptive in nature, i.e. the policy selects or skips a frame without capturing it first." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 426, + 547, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 426, + 547, + 569 + ], + "spans": [ + { + "bbox": [ + 304, + 426, + 547, + 569 + ], + "type": "text", + "content": "Our model has two main components (see Fig. 2): (1) a shared scene mapper, and (2) a visual sampling policy. At every step " + }, + { + "bbox": [ + 304, + 426, + 547, + 569 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 426, + 547, + 569 + ], + "type": "text", + "content": ", the shared mapper has two functions. First, it estimates the map of a previously unseen environment by exploiting the shared spatial cues in the audio-visual observations of the two egos. Second, it informs the policy about the utility of sampling a certain visual frame. Guided by the mapper, the policy samples only the most informative visual frames that are expected to boost mapping significantly over using just audio. Note that, unlike the visuals, we observe audio continuously as it is less resource-intensive vis-a-vis storage and power requirements for processing [2]." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 570, + 545, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 545, + 617 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 545, + 617 + ], + "type": "text", + "content": "We learn our task through the synergy of the mapper and the policy, such that under the constraint of a limited visual budget " + }, + { + "bbox": [ + 304, + 570, + 545, + 617 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 304, + 570, + 545, + 617 + ], + "type": "text", + "content": ", our model implicitly understands which visual frames are critical for mapping." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 618, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 618, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 618, + 547, + 713 + ], + "type": "text", + "content": "First, we describe the steps involved to prepare our model inputs (Sec. 4.1). Next, we introduce our visual sampling policy (Sec. 4.2) and shared scene mapper (Sec. 4.3). Finally, we present model training details (Sec. 4.4). Through the rest of the text, we use separate notations to distinguish the egos' observations " + }, + { + "bbox": [ + 304, + 618, + 547, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{O}" + }, + { + "bbox": [ + 304, + 618, + 547, + 713 + ], + "type": "text", + "content": " (i.e. what the egos receive from the environment) from our model inputs " + }, + { + "bbox": [ + 304, + 618, + 547, + 713 + ], + "type": "inline_equation", + "content": "O" + }, + { + "bbox": [ + 304, + 618, + 547, + 713 + ], + "type": "text", + "content": " (i.e. what we capture and feed to our model for efficient mapping)." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "10556" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 85, + 70, + 294, + 308 + ], + "blocks": [ + { + "bbox": [ + 85, + 70, + 294, + 308 + ], + "lines": [ + { + "bbox": [ + 85, + 70, + 294, + 308 + ], + "spans": [ + { + "bbox": [ + 85, + 70, + 294, + 308 + ], + "type": "image", + "image_path": "b507832bb99458974534599042ac1c4d06072feaa28aab7988b17873014defa0.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 310, + 547, + 365 + ], + "lines": [ + { + "bbox": [ + 46, + 310, + 547, + 365 + ], + "spans": [ + { + "bbox": [ + 46, + 310, + 547, + 365 + ], + "type": "text", + "content": "Figure 2. Our model has two main components: a) a visual sampling policy (left), and b) a shared scene mapper (right). At each step, our policy receives the current audio along with the previous audio(-visual) observations for the egos and decides for each ego individually whether to capture its visual frame at the current step. As per the policy predictions, the shared mapper conditionally uses the current visual frame(s) and audio along with the past audio(-visual) observations to predict the occupancy map of the scene, a ground-plane map showing where obstacles and freespace are (shown in green and white)." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 299, + 70, + 508, + 308 + ], + "blocks": [ + { + "bbox": [ + 299, + 70, + 508, + 308 + ], + "lines": [ + { + "bbox": [ + 299, + 70, + 508, + 308 + ], + "spans": [ + { + "bbox": [ + 299, + 70, + 508, + 308 + ], + "type": "image", + "image_path": "b6609000eeaff0df62be6fdbce6ef0677b96427cdf1fa89d7ea4912de78343be.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 375, + 186, + 388 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 375, + 186, + 388 + ], + "spans": [ + { + "bbox": [ + 47, + 375, + 186, + 388 + ], + "type": "text", + "content": "4.1. Model input preparation" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 395, + 288, + 562 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 395, + 288, + 562 + ], + "spans": [ + { + "bbox": [ + 46, + 395, + 288, + 562 + ], + "type": "text", + "content": "We prepare our model inputs by separately preprocessing the visual and audio modalities. If our policy decides to sample an image " + }, + { + "bbox": [ + 46, + 395, + 288, + 562 + ], + "type": "inline_equation", + "content": "\\mathcal{V}" + }, + { + "bbox": [ + 46, + 395, + 288, + 562 + ], + "type": "text", + "content": ", we transform it into " + }, + { + "bbox": [ + 46, + 395, + 288, + 562 + ], + "type": "inline_equation", + "content": "V = (V^{R}, V^{M})" + }, + { + "bbox": [ + 46, + 395, + 288, + 562 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 46, + 395, + 288, + 562 + ], + "type": "inline_equation", + "content": "V^{R}" + }, + { + "bbox": [ + 46, + 395, + 288, + 562 + ], + "type": "text", + "content": " denotes the normalized RGB image with pixel values " + }, + { + "bbox": [ + 46, + 395, + 288, + 562 + ], + "type": "inline_equation", + "content": "\\in [0,1]" + }, + { + "bbox": [ + 46, + 395, + 288, + 562 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 46, + 395, + 288, + 562 + ], + "type": "inline_equation", + "content": "V^{M}" + }, + { + "bbox": [ + 46, + 395, + 288, + 562 + ], + "type": "text", + "content": " denotes the " + }, + { + "bbox": [ + 46, + 395, + 288, + 562 + ], + "type": "inline_equation", + "content": "90^{\\circ}" + }, + { + "bbox": [ + 46, + 395, + 288, + 562 + ], + "type": "text", + "content": " FoV top-down occupancy map created by projecting the depth image to the ground plane. To do the depth projection, we first backproject it into the world coordinates using the camera's intrinsic parameters to compute the local visible scene's 3D point cloud. Next, we project these points to obtain a two-channel binary top-down map of size " + }, + { + "bbox": [ + 46, + 395, + 288, + 562 + ], + "type": "inline_equation", + "content": "h \\times w \\times 2" + }, + { + "bbox": [ + 46, + 395, + 288, + 562 + ], + "type": "text", + "content": ", where the first channel reveals occupied/free areas, and the second channel reveals seen/unseen areas. If our policy skips " + }, + { + "bbox": [ + 46, + 395, + 288, + 562 + ], + "type": "inline_equation", + "content": "\\mathcal{V}" + }, + { + "bbox": [ + 46, + 395, + 288, + 562 + ], + "type": "text", + "content": ", we set " + }, + { + "bbox": [ + 46, + 395, + 288, + 562 + ], + "type": "inline_equation", + "content": "V^{R}" + }, + { + "bbox": [ + 46, + 395, + 288, + 562 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 395, + 288, + 562 + ], + "type": "inline_equation", + "content": "V^{M}" + }, + { + "bbox": [ + 46, + 395, + 288, + 562 + ], + "type": "text", + "content": " to all-zero matrices of the appropriate size." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 563, + 288, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 563, + 288, + 635 + ], + "spans": [ + { + "bbox": [ + 46, + 563, + 288, + 635 + ], + "type": "text", + "content": "For a speech waveform " + }, + { + "bbox": [ + 46, + 563, + 288, + 635 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 46, + 563, + 288, + 635 + ], + "type": "text", + "content": ", we calculate the short-time Fourier transform (STFT) magnitude spectrogram denoted by " + }, + { + "bbox": [ + 46, + 563, + 288, + 635 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 46, + 563, + 288, + 635 + ], + "type": "text", + "content": " of size " + }, + { + "bbox": [ + 46, + 563, + 288, + 635 + ], + "type": "inline_equation", + "content": "F \\times T \\times C" + }, + { + "bbox": [ + 46, + 563, + 288, + 635 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 563, + 288, + 635 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 46, + 563, + 288, + 635 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 563, + 288, + 635 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 46, + 563, + 288, + 635 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 563, + 288, + 635 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 46, + 563, + 288, + 635 + ], + "type": "text", + "content": " are the number of frequency bins, time windows, and ambisonic microphone channels, respectively. Lastly, we normalize each pose " + }, + { + "bbox": [ + 46, + 563, + 288, + 635 + ], + "type": "inline_equation", + "content": "P_{i,t}" + }, + { + "bbox": [ + 46, + 563, + 288, + 635 + ], + "type": "text", + "content": " to be relative to " + }, + { + "bbox": [ + 46, + 563, + 288, + 635 + ], + "type": "inline_equation", + "content": "P_{1,1}" + }, + { + "bbox": [ + 46, + 563, + 288, + 635 + ], + "type": "text", + "content": ". See Sec. 5 and Supp. for more details." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 646, + 175, + 659 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 646, + 175, + 659 + ], + "spans": [ + { + "bbox": [ + 47, + 646, + 175, + 659 + ], + "type": "text", + "content": "4.2. Visual sampling policy" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 665, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 288, + 714 + ], + "type": "text", + "content": "At every step " + }, + { + "bbox": [ + 46, + 665, + 288, + 714 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 665, + 288, + 714 + ], + "type": "text", + "content": ", our visual sampling policy " + }, + { + "bbox": [ + 46, + 665, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\pi^V" + }, + { + "bbox": [ + 46, + 665, + 288, + 714 + ], + "type": "text", + "content": " (Fig. 2 left) receives " + }, + { + "bbox": [ + 46, + 665, + 288, + 714 + ], + "type": "inline_equation", + "content": "O^\\pi(t)" + }, + { + "bbox": [ + 46, + 665, + 288, + 714 + ], + "type": "text", + "content": " as input and makes the decision to either capture or skip the visual frame " + }, + { + "bbox": [ + 46, + 665, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{V}_{i,t}" + }, + { + "bbox": [ + 46, + 665, + 288, + 714 + ], + "type": "text", + "content": " for each ego " + }, + { + "bbox": [ + 46, + 665, + 288, + 714 + ], + "type": "inline_equation", + "content": "E_i" + }, + { + "bbox": [ + 46, + 665, + 288, + 714 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 46, + 665, + 288, + 714 + ], + "type": "inline_equation", + "content": "O^\\pi(t)" + }, + { + "bbox": [ + 46, + 665, + 288, + 714 + ], + "type": "text", + "content": " comprises the visual cue from the last step along with the" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 375, + 545, + 487 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 375, + 545, + 487 + ], + "spans": [ + { + "bbox": [ + 304, + 375, + 545, + 487 + ], + "type": "text", + "content": "speech cues, and the poses from the current step and the last step for both egos. Formally, " + }, + { + "bbox": [ + 304, + 375, + 545, + 487 + ], + "type": "inline_equation", + "content": "O^{\\pi}(t) = \\{O_{i}^{\\pi}(t):i = 1\\ldots 2\\}" + }, + { + "bbox": [ + 304, + 375, + 545, + 487 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 375, + 545, + 487 + ], + "type": "inline_equation", + "content": "O_{i}^{\\pi}(t) = \\{V_{i,t - 1},S_{i,j},P_{i,j},S_{i,j}^{\\prime},P_{i,j}^{\\prime}:j = t - 1\\ldots t\\}" + }, + { + "bbox": [ + 304, + 375, + 545, + 487 + ], + "type": "text", + "content": ". The policy first uses an encoder network to generate a multi-modal embedding of " + }, + { + "bbox": [ + 304, + 375, + 545, + 487 + ], + "type": "inline_equation", + "content": "O^{\\pi}(t)" + }, + { + "bbox": [ + 304, + 375, + 545, + 487 + ], + "type": "text", + "content": ", and then passes the embedding to a policy network that makes a sampling decision per ego. At " + }, + { + "bbox": [ + 304, + 375, + 545, + 487 + ], + "type": "inline_equation", + "content": "t = 1" + }, + { + "bbox": [ + 304, + 375, + 545, + 487 + ], + "type": "text", + "content": ", as per our problem definition (Sec. 3), the policy always chooses to sample the visual frames for both egos, i.e., the cameras are initially on." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 491, + 546, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 491, + 546, + 676 + ], + "spans": [ + { + "bbox": [ + 304, + 491, + 546, + 676 + ], + "type": "text", + "content": "Multi-modal policy embedding. To process ego " + }, + { + "bbox": [ + 304, + 491, + 546, + 676 + ], + "type": "inline_equation", + "content": "E_{i}" + }, + { + "bbox": [ + 304, + 491, + 546, + 676 + ], + "type": "text", + "content": "'s visual input " + }, + { + "bbox": [ + 304, + 491, + 546, + 676 + ], + "type": "inline_equation", + "content": "V_{i,t-1}" + }, + { + "bbox": [ + 304, + 491, + 546, + 676 + ], + "type": "text", + "content": " from the last step, we encode the RGB image " + }, + { + "bbox": [ + 304, + 491, + 546, + 676 + ], + "type": "inline_equation", + "content": "V_{i,t-1}^{R}" + }, + { + "bbox": [ + 304, + 491, + 546, + 676 + ], + "type": "text", + "content": " and map " + }, + { + "bbox": [ + 304, + 491, + 546, + 676 + ], + "type": "inline_equation", + "content": "V_{i,t-1}^{M}" + }, + { + "bbox": [ + 304, + 491, + 546, + 676 + ], + "type": "text", + "content": " with separate CNNs. We then concatenate the two features to generate the visual embedding " + }, + { + "bbox": [ + 304, + 491, + 546, + 676 + ], + "type": "inline_equation", + "content": "v_{i,t-1}" + }, + { + "bbox": [ + 304, + 491, + 546, + 676 + ], + "type": "text", + "content": ". To encode the pose inputs " + }, + { + "bbox": [ + 304, + 491, + 546, + 676 + ], + "type": "inline_equation", + "content": "\\{P_{i,t-1}, P_{i,t-1}', P_{i,t}, P_{i,t}'\\}" + }, + { + "bbox": [ + 304, + 491, + 546, + 676 + ], + "type": "text", + "content": ", we use a linear layer and generate pose embeddings " + }, + { + "bbox": [ + 304, + 491, + 546, + 676 + ], + "type": "inline_equation", + "content": "\\{p_{i,t-1}, p_{i,t-1}', p_{i,t}, p_{i,t}'\\}" + }, + { + "bbox": [ + 304, + 491, + 546, + 676 + ], + "type": "text", + "content": ". We process the speech inputs " + }, + { + "bbox": [ + 304, + 491, + 546, + 676 + ], + "type": "inline_equation", + "content": "\\{S_{i,t-1}, S_{i,t-1}', S_{i,t}, S_{i,t}'\\}" + }, + { + "bbox": [ + 304, + 491, + 546, + 676 + ], + "type": "text", + "content": " using another CNN and create speech embeddings " + }, + { + "bbox": [ + 304, + 491, + 546, + 676 + ], + "type": "inline_equation", + "content": "\\{s_{i,t-1}, s_{i,t-1}', s_{i,t}, s_{i,t}'\\}" + }, + { + "bbox": [ + 304, + 491, + 546, + 676 + ], + "type": "text", + "content": ". Next, we fuse the visual, speech and pose embeddings using linear layers (see Fig. 2 left) to obtain the multi-modal policy embedding " + }, + { + "bbox": [ + 304, + 491, + 546, + 676 + ], + "type": "inline_equation", + "content": "e_{i,t}" + }, + { + "bbox": [ + 304, + 491, + 546, + 676 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 304, + 491, + 546, + 676 + ], + "type": "inline_equation", + "content": "E_{i}" + }, + { + "bbox": [ + 304, + 491, + 546, + 676 + ], + "type": "text", + "content": ". Finally, we fuse the policy embeddings for the two egos, " + }, + { + "bbox": [ + 304, + 491, + 546, + 676 + ], + "type": "inline_equation", + "content": "e_{1,t}" + }, + { + "bbox": [ + 304, + 491, + 546, + 676 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 491, + 546, + 676 + ], + "type": "inline_equation", + "content": "e_{2,t}" + }, + { + "bbox": [ + 304, + 491, + 546, + 676 + ], + "type": "text", + "content": " with a linear layer to produce the multi-modal policy embedding " + }, + { + "bbox": [ + 304, + 491, + 546, + 676 + ], + "type": "inline_equation", + "content": "e_t" + }, + { + "bbox": [ + 304, + 491, + 546, + 676 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 677, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 547, + 714 + ], + "type": "text", + "content": "The visual, audio, and pose inputs carry complementary cues required for efficient visual sampling. Whereas the pose inputs from the last and current steps explicitly reveal" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10557" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 193 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 193 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 193 + ], + "type": "text", + "content": "the viewpoint change between the steps, the previous and current speech inputs provide information about the changes in the local and global scene structures as a function of the previously sampled visual inputs, which together suggest the value of sampling a visual frame at the current step. Furthermore, guided by our training reward (below in Sec. 4.4), the previously observed visual frames and audio together enable our policy to anticipate the current frames and skip them if they are deemed redundant, thereby improving mapping accuracy for a low visual budget." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 198, + 289, + 307 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 198, + 289, + 307 + ], + "spans": [ + { + "bbox": [ + 47, + 198, + 289, + 307 + ], + "type": "text", + "content": "Policy network. The policy network consists of a GRU that estimates an updated history " + }, + { + "bbox": [ + 47, + 198, + 289, + 307 + ], + "type": "inline_equation", + "content": "h_t" + }, + { + "bbox": [ + 47, + 198, + 289, + 307 + ], + "type": "text", + "content": " along with the current state representation " + }, + { + "bbox": [ + 47, + 198, + 289, + 307 + ], + "type": "inline_equation", + "content": "g_t" + }, + { + "bbox": [ + 47, + 198, + 289, + 307 + ], + "type": "text", + "content": ", using the fused embedding " + }, + { + "bbox": [ + 47, + 198, + 289, + 307 + ], + "type": "inline_equation", + "content": "e_t" + }, + { + "bbox": [ + 47, + 198, + 289, + 307 + ], + "type": "text", + "content": " and the history of states " + }, + { + "bbox": [ + 47, + 198, + 289, + 307 + ], + "type": "inline_equation", + "content": "h_{t-1}" + }, + { + "bbox": [ + 47, + 198, + 289, + 307 + ], + "type": "text", + "content": ". An actor-critic module takes " + }, + { + "bbox": [ + 47, + 198, + 289, + 307 + ], + "type": "inline_equation", + "content": "g_t" + }, + { + "bbox": [ + 47, + 198, + 289, + 307 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 198, + 289, + 307 + ], + "type": "inline_equation", + "content": "h_{t-1}" + }, + { + "bbox": [ + 47, + 198, + 289, + 307 + ], + "type": "text", + "content": " as inputs and predicts a policy distribution " + }, + { + "bbox": [ + 47, + 198, + 289, + 307 + ], + "type": "inline_equation", + "content": "\\pi_\\theta(a_{i,t}|g_t, h_{t-1})" + }, + { + "bbox": [ + 47, + 198, + 289, + 307 + ], + "type": "text", + "content": " per ego along with the value of the state " + }, + { + "bbox": [ + 47, + 198, + 289, + 307 + ], + "type": "inline_equation", + "content": "H_\\theta(g_t, h_{t-1})" + }, + { + "bbox": [ + 47, + 198, + 289, + 307 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 47, + 198, + 289, + 307 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 47, + 198, + 289, + 307 + ], + "type": "text", + "content": " are policy parameters). The policy samples an action " + }, + { + "bbox": [ + 47, + 198, + 289, + 307 + ], + "type": "inline_equation", + "content": "a_{i,t} \\in \\{0,1\\}" + }, + { + "bbox": [ + 47, + 198, + 289, + 307 + ], + "type": "text", + "content": " for every " + }, + { + "bbox": [ + 47, + 198, + 289, + 307 + ], + "type": "inline_equation", + "content": "E_i" + }, + { + "bbox": [ + 47, + 198, + 289, + 307 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 47, + 198, + 289, + 307 + ], + "type": "inline_equation", + "content": "a_{i,t} = 1" + }, + { + "bbox": [ + 47, + 198, + 289, + 307 + ], + "type": "text", + "content": " corresponds to selecting " + }, + { + "bbox": [ + 47, + 198, + 289, + 307 + ], + "type": "inline_equation", + "content": "\\mathcal{V}_{i,t}" + }, + { + "bbox": [ + 47, + 198, + 289, + 307 + ], + "type": "text", + "content": ", whereas " + }, + { + "bbox": [ + 47, + 198, + 289, + 307 + ], + "type": "inline_equation", + "content": "a_{i,t} = 0" + }, + { + "bbox": [ + 47, + 198, + 289, + 307 + ], + "type": "text", + "content": " means to skip that frame." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 314, + 170, + 327 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 314, + 170, + 327 + ], + "spans": [ + { + "bbox": [ + 47, + 314, + 170, + 327 + ], + "type": "text", + "content": "4.3. Shared scene mapper" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 332, + 289, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 332, + 289, + 443 + ], + "spans": [ + { + "bbox": [ + 46, + 332, + 289, + 443 + ], + "type": "text", + "content": "Whereas " + }, + { + "bbox": [ + 46, + 332, + 289, + 443 + ], + "type": "inline_equation", + "content": "O^{\\pi}(t)" + }, + { + "bbox": [ + 46, + 332, + 289, + 443 + ], + "type": "text", + "content": " denotes our policy input (Sec. 4.2), " + }, + { + "bbox": [ + 46, + 332, + 289, + 443 + ], + "type": "inline_equation", + "content": "O^{M}(t)" + }, + { + "bbox": [ + 46, + 332, + 289, + 443 + ], + "type": "text", + "content": " denotes the input to our shared scene mapper " + }, + { + "bbox": [ + 46, + 332, + 289, + 443 + ], + "type": "inline_equation", + "content": "f^{M}" + }, + { + "bbox": [ + 46, + 332, + 289, + 443 + ], + "type": "text", + "content": " at step " + }, + { + "bbox": [ + 46, + 332, + 289, + 443 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 332, + 289, + 443 + ], + "type": "text", + "content": ", such that " + }, + { + "bbox": [ + 46, + 332, + 289, + 443 + ], + "type": "inline_equation", + "content": "O^{M}(t) = \\left\\{(V_{i,j}, S_{i,j}, S_{i,j}^{\\prime}, P_{i,j}, P_{i,j}^{\\prime}): i = 1 \\dots 2, j = 1 \\dots t\\right\\}" + }, + { + "bbox": [ + 46, + 332, + 289, + 443 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 46, + 332, + 289, + 443 + ], + "type": "inline_equation", + "content": "f^{m}" + }, + { + "bbox": [ + 46, + 332, + 289, + 443 + ], + "type": "text", + "content": " starts by embedding each component of " + }, + { + "bbox": [ + 46, + 332, + 289, + 443 + ], + "type": "inline_equation", + "content": "O^{M}(t)" + }, + { + "bbox": [ + 46, + 332, + 289, + 443 + ], + "type": "text", + "content": " using a separate network. This is followed by a multi-modal memory that stores the embeddings since the start of the episode. Finally, a transformer [65] predicts an estimate " + }, + { + "bbox": [ + 46, + 332, + 289, + 443 + ], + "type": "inline_equation", + "content": "\\tilde{M}(t)" + }, + { + "bbox": [ + 46, + 332, + 289, + 443 + ], + "type": "text", + "content": " of the scene map conditioned on the multi-modal memory and the egos' poses in the episode." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 448, + 289, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 448, + 289, + 510 + ], + "spans": [ + { + "bbox": [ + 47, + 448, + 289, + 510 + ], + "type": "text", + "content": "Multi-modal mapper embedding. For the visual input " + }, + { + "bbox": [ + 47, + 448, + 289, + 510 + ], + "type": "inline_equation", + "content": "V_{i,j}" + }, + { + "bbox": [ + 47, + 448, + 289, + 510 + ], + "type": "text", + "content": ", we encode " + }, + { + "bbox": [ + 47, + 448, + 289, + 510 + ], + "type": "inline_equation", + "content": "V_{i,j}^{R}" + }, + { + "bbox": [ + 47, + 448, + 289, + 510 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 448, + 289, + 510 + ], + "type": "inline_equation", + "content": "V_{i,j}^{M}" + }, + { + "bbox": [ + 47, + 448, + 289, + 510 + ], + "type": "text", + "content": " using separate CNNs and do a channel-wise concatenation to get visual features " + }, + { + "bbox": [ + 47, + 448, + 289, + 510 + ], + "type": "inline_equation", + "content": "\\hat{v}_{i,j}" + }, + { + "bbox": [ + 47, + 448, + 289, + 510 + ], + "type": "text", + "content": ". Similarly, speech is encoded using separate CNNs to get " + }, + { + "bbox": [ + 47, + 448, + 289, + 510 + ], + "type": "inline_equation", + "content": "\\hat{s}_{i,j}" + }, + { + "bbox": [ + 47, + 448, + 289, + 510 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 448, + 289, + 510 + ], + "type": "inline_equation", + "content": "\\hat{s}_{i,j}'" + }, + { + "bbox": [ + 47, + 448, + 289, + 510 + ], + "type": "text", + "content": ". Each of " + }, + { + "bbox": [ + 47, + 448, + 289, + 510 + ], + "type": "inline_equation", + "content": "\\hat{v}" + }, + { + "bbox": [ + 47, + 448, + 289, + 510 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 47, + 448, + 289, + 510 + ], + "type": "inline_equation", + "content": "\\hat{s}" + }, + { + "bbox": [ + 47, + 448, + 289, + 510 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 448, + 289, + 510 + ], + "type": "inline_equation", + "content": "\\hat{s}'" + }, + { + "bbox": [ + 47, + 448, + 289, + 510 + ], + "type": "text", + "content": " is of size " + }, + { + "bbox": [ + 47, + 448, + 289, + 510 + ], + "type": "inline_equation", + "content": "4 \\times 4 \\times 1024" + }, + { + "bbox": [ + 47, + 448, + 289, + 510 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 510, + 289, + 678 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 510, + 289, + 678 + ], + "spans": [ + { + "bbox": [ + 46, + 510, + 289, + 678 + ], + "type": "text", + "content": "For both vision and speech, we compute two positional embeddings, " + }, + { + "bbox": [ + 46, + 510, + 289, + 678 + ], + "type": "inline_equation", + "content": "p^{\\mathrm{I}}" + }, + { + "bbox": [ + 46, + 510, + 289, + 678 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 510, + 289, + 678 + ], + "type": "inline_equation", + "content": "p^{\\mathrm{II}}" + }, + { + "bbox": [ + 46, + 510, + 289, + 678 + ], + "type": "text", + "content": ". They encode the pose of the egos in the 3D space, and the index of each 1024-dimensional feature in the visual or speech features in the raster order respectively. Whereas " + }, + { + "bbox": [ + 46, + 510, + 289, + 678 + ], + "type": "inline_equation", + "content": "p^{\\mathrm{I}}" + }, + { + "bbox": [ + 46, + 510, + 289, + 678 + ], + "type": "text", + "content": " helps discover spatial cues as a function of the egos' location in the 3D scene, " + }, + { + "bbox": [ + 46, + 510, + 289, + 678 + ], + "type": "inline_equation", + "content": "p^{\\mathrm{II}}" + }, + { + "bbox": [ + 46, + 510, + 289, + 678 + ], + "type": "text", + "content": " enables our model to attend to different modalities in a more fine-grained manner. For both, we compute an 8-dimensional sinusoidal positional encoding [65] and then pass it through a linear layer to obtain a 1024-dimensional embedding. For " + }, + { + "bbox": [ + 46, + 510, + 289, + 678 + ], + "type": "inline_equation", + "content": "p^{\\mathrm{II}}" + }, + { + "bbox": [ + 46, + 510, + 289, + 678 + ], + "type": "text", + "content": ", we additionally repeat this process for every feature index in the raster order. Lastly, we reshape " + }, + { + "bbox": [ + 46, + 510, + 289, + 678 + ], + "type": "inline_equation", + "content": "p^{\\mathrm{I}}" + }, + { + "bbox": [ + 46, + 510, + 289, + 678 + ], + "type": "text", + "content": " and add it with " + }, + { + "bbox": [ + 46, + 510, + 289, + 678 + ], + "type": "inline_equation", + "content": "p^{\\mathrm{II}}" + }, + { + "bbox": [ + 46, + 510, + 289, + 678 + ], + "type": "text", + "content": " to produce " + }, + { + "bbox": [ + 46, + 510, + 289, + 678 + ], + "type": "inline_equation", + "content": "4 \\times 4 \\times 1024" + }, + { + "bbox": [ + 46, + 510, + 289, + 678 + ], + "type": "text", + "content": "-dimensional positional embeddings, " + }, + { + "bbox": [ + 46, + 510, + 289, + 678 + ], + "type": "inline_equation", + "content": "\\hat{p}_{i,j}" + }, + { + "bbox": [ + 46, + 510, + 289, + 678 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 46, + 510, + 289, + 678 + ], + "type": "inline_equation", + "content": "\\hat{v}_{i,j}" + }, + { + "bbox": [ + 46, + 510, + 289, + 678 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 510, + 289, + 678 + ], + "type": "inline_equation", + "content": "\\hat{s}_{i,j}" + }, + { + "bbox": [ + 46, + 510, + 289, + 678 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 510, + 289, + 678 + ], + "type": "inline_equation", + "content": "\\hat{p}_{i,j}'" + }, + { + "bbox": [ + 46, + 510, + 289, + 678 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 46, + 510, + 289, + 678 + ], + "type": "inline_equation", + "content": "\\hat{s}_{i,j}'" + }, + { + "bbox": [ + 46, + 510, + 289, + 678 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 678, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 678, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 678, + 288, + 714 + ], + "type": "text", + "content": "Following [44], we also learn an embedding " + }, + { + "bbox": [ + 47, + 678, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\hat{m}_{i,j} \\in \\{\\hat{m}_V, \\hat{m}_S, \\hat{m}_{S'}\\}" + }, + { + "bbox": [ + 47, + 678, + 288, + 714 + ], + "type": "text", + "content": " to capture different modality types, where " + }, + { + "bbox": [ + 47, + 678, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\hat{m}_V" + }, + { + "bbox": [ + 47, + 678, + 288, + 714 + ], + "type": "text", + "content": " represents vision, and " + }, + { + "bbox": [ + 47, + 678, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\hat{m}_S" + }, + { + "bbox": [ + 47, + 678, + 288, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 678, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\hat{m}_{S'}" + }, + { + "bbox": [ + 47, + 678, + 288, + 714 + ], + "type": "text", + "content": " represent the speech" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "text", + "content": "from self and that of the other ego, respectively. The modality-based embeddings help our model differentiate between different modalities and better map the scene by learning complementary spatial cues from them." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 123, + 547, + 209 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 123, + 547, + 209 + ], + "spans": [ + { + "bbox": [ + 304, + 123, + 547, + 209 + ], + "type": "text", + "content": "Multi-modal memory. For the visual input " + }, + { + "bbox": [ + 304, + 123, + 547, + 209 + ], + "type": "inline_equation", + "content": "V_{i,j}" + }, + { + "bbox": [ + 304, + 123, + 547, + 209 + ], + "type": "text", + "content": ", we add its embedding " + }, + { + "bbox": [ + 304, + 123, + 547, + 209 + ], + "type": "inline_equation", + "content": "\\hat{v}_{i,j}" + }, + { + "bbox": [ + 304, + 123, + 547, + 209 + ], + "type": "text", + "content": " with its positional embedding " + }, + { + "bbox": [ + 304, + 123, + 547, + 209 + ], + "type": "inline_equation", + "content": "\\hat{p}_{i,j}" + }, + { + "bbox": [ + 304, + 123, + 547, + 209 + ], + "type": "text", + "content": " and modality embedding " + }, + { + "bbox": [ + 304, + 123, + 547, + 209 + ], + "type": "inline_equation", + "content": "\\hat{m}_{i,j}^{V}" + }, + { + "bbox": [ + 304, + 123, + 547, + 209 + ], + "type": "text", + "content": ", and flatten the sum to get a " + }, + { + "bbox": [ + 304, + 123, + 547, + 209 + ], + "type": "inline_equation", + "content": "16 \\times 1024" + }, + { + "bbox": [ + 304, + 123, + 547, + 209 + ], + "type": "text", + "content": "-dimensional embedding. Similarly, we fuse the speech embeddings by taking their sum and flattening them. This generates a multi-modal memory of fused embeddings " + }, + { + "bbox": [ + 304, + 123, + 547, + 209 + ], + "type": "inline_equation", + "content": "o" + }, + { + "bbox": [ + 304, + 123, + 547, + 209 + ], + "type": "text", + "content": ", such that " + }, + { + "bbox": [ + 304, + 123, + 547, + 209 + ], + "type": "inline_equation", + "content": "o = \\{o_{1,1}^{V}, \\ldots, o_{2,t}^{V}, o_{1,1}^{S}, \\ldots, o_{2,t}^{S}, o_{1,1}^{S'}, \\ldots, o_{2,t}^{S'}\\}" + }, + { + "bbox": [ + 304, + 123, + 547, + 209 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 211, + 547, + 282 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 211, + 547, + 282 + ], + "spans": [ + { + "bbox": [ + 304, + 211, + 547, + 282 + ], + "type": "text", + "content": "Occupancy prediction. To predict the underlying scene occupancy, we first use a transformer encoder [65] to attend to the embeddings in " + }, + { + "bbox": [ + 304, + 211, + 547, + 282 + ], + "type": "inline_equation", + "content": "o" + }, + { + "bbox": [ + 304, + 211, + 547, + 282 + ], + "type": "text", + "content": " and capture short- and long-range correlations within and across modalities using a stack of self-attention layers. This generates an audio-visual representation that models the spatial layout of the 3D scene." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 283, + 547, + 464 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 283, + 547, + 464 + ], + "spans": [ + { + "bbox": [ + 304, + 283, + 547, + 464 + ], + "type": "text", + "content": "Next, we use a transformer decoder [65] to perform cross-attention on the audio-visual representation of the scene conditioned on the embedding " + }, + { + "bbox": [ + 304, + 283, + 547, + 464 + ], + "type": "inline_equation", + "content": "\\hat{p}_{i,j}" + }, + { + "bbox": [ + 304, + 283, + 547, + 464 + ], + "type": "text", + "content": " for every pose " + }, + { + "bbox": [ + 304, + 283, + 547, + 464 + ], + "type": "inline_equation", + "content": "P_{i,j}" + }, + { + "bbox": [ + 304, + 283, + 547, + 464 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 304, + 283, + 547, + 464 + ], + "type": "inline_equation", + "content": "O^M (t)" + }, + { + "bbox": [ + 304, + 283, + 547, + 464 + ], + "type": "text", + "content": " and generate an embedding " + }, + { + "bbox": [ + 304, + 283, + 547, + 464 + ], + "type": "inline_equation", + "content": "d_{i,j}" + }, + { + "bbox": [ + 304, + 283, + 547, + 464 + ], + "type": "text", + "content": " for the pose. Finally, we upsample " + }, + { + "bbox": [ + 304, + 283, + 547, + 464 + ], + "type": "inline_equation", + "content": "d_{i,j}" + }, + { + "bbox": [ + 304, + 283, + 547, + 464 + ], + "type": "text", + "content": " using a multi-layer network " + }, + { + "bbox": [ + 304, + 283, + 547, + 464 + ], + "type": "inline_equation", + "content": "U" + }, + { + "bbox": [ + 304, + 283, + 547, + 464 + ], + "type": "text", + "content": " comprising transpose convolutions and a sigmoid layer at the end to predict an estimate " + }, + { + "bbox": [ + 304, + 283, + 547, + 464 + ], + "type": "inline_equation", + "content": "\\tilde{M}_{i,j}" + }, + { + "bbox": [ + 304, + 283, + 547, + 464 + ], + "type": "text", + "content": " of the ground-truth local " + }, + { + "bbox": [ + 304, + 283, + 547, + 464 + ], + "type": "inline_equation", + "content": "360^{\\circ}" + }, + { + "bbox": [ + 304, + 283, + 547, + 464 + ], + "type": "text", + "content": " FoV map for the pose, " + }, + { + "bbox": [ + 304, + 283, + 547, + 464 + ], + "type": "inline_equation", + "content": "M_{i,j}" + }, + { + "bbox": [ + 304, + 283, + 547, + 464 + ], + "type": "text", + "content": ". Both " + }, + { + "bbox": [ + 304, + 283, + 547, + 464 + ], + "type": "inline_equation", + "content": "M_{i,j}" + }, + { + "bbox": [ + 304, + 283, + 547, + 464 + ], + "type": "text", + "content": " and its estimate " + }, + { + "bbox": [ + 304, + 283, + 547, + 464 + ], + "type": "inline_equation", + "content": "\\tilde{M}_{i,j}" + }, + { + "bbox": [ + 304, + 283, + 547, + 464 + ], + "type": "text", + "content": " are two-channel binary occupancy maps of size " + }, + { + "bbox": [ + 304, + 283, + 547, + 464 + ], + "type": "inline_equation", + "content": "H\\times W" + }, + { + "bbox": [ + 304, + 283, + 547, + 464 + ], + "type": "text", + "content": ". To obtain the estimated map " + }, + { + "bbox": [ + 304, + 283, + 547, + 464 + ], + "type": "inline_equation", + "content": "\\tilde{M} (t)" + }, + { + "bbox": [ + 304, + 283, + 547, + 464 + ], + "type": "text", + "content": " for the scene, we register each prediction " + }, + { + "bbox": [ + 304, + 283, + 547, + 464 + ], + "type": "inline_equation", + "content": "\\tilde{M}_{i,j}" + }, + { + "bbox": [ + 304, + 283, + 547, + 464 + ], + "type": "text", + "content": " onto a larger shared map using the pose " + }, + { + "bbox": [ + 304, + 283, + 547, + 464 + ], + "type": "inline_equation", + "content": "P_{i,j}" + }, + { + "bbox": [ + 304, + 283, + 547, + 464 + ], + "type": "text", + "content": " and threshold the final shared map at 0.5 (see Supp. for map registration details). Importantly, the shared map allows communication between both egos' data streams for more informed mapping and sampling, as we show in results." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 470, + 400, + 483 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 470, + 400, + 483 + ], + "spans": [ + { + "bbox": [ + 305, + 470, + 400, + 483 + ], + "type": "text", + "content": "4.4. Model training" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 489, + 545, + 513 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 489, + 545, + 513 + ], + "spans": [ + { + "bbox": [ + 304, + 489, + 545, + 513 + ], + "type": "text", + "content": "Policy training. We propose a novel dense RL reward to train policy " + }, + { + "bbox": [ + 304, + 489, + 545, + 513 + ], + "type": "inline_equation", + "content": "\\pi^V" + }, + { + "bbox": [ + 304, + 489, + 545, + 513 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 369, + 522, + 481, + 536 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 369, + 522, + 481, + 536 + ], + "spans": [ + { + "bbox": [ + 369, + 522, + 481, + 536 + ], + "type": "interline_equation", + "content": "r (t) = \\Delta Q (t) - \\eta * \\rho (t).", + "image_path": "814aae703c533bed34519f4689dc3cedc2a7246b37b4cb21c2c0611954e9953f.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 544, + 545, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 544, + 545, + 605 + ], + "spans": [ + { + "bbox": [ + 304, + 544, + 545, + 605 + ], + "type": "inline_equation", + "content": "\\Delta Q(t)" + }, + { + "bbox": [ + 304, + 544, + 545, + 605 + ], + "type": "text", + "content": " measures the improvement in mapping from taking actions " + }, + { + "bbox": [ + 304, + 544, + 545, + 605 + ], + "type": "inline_equation", + "content": "\\{a_{i,t} : i = 1 \\dots 2\\}" + }, + { + "bbox": [ + 304, + 544, + 545, + 605 + ], + "type": "text", + "content": " over not sampling any visual frame at step " + }, + { + "bbox": [ + 304, + 544, + 545, + 605 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 544, + 545, + 605 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 304, + 544, + 545, + 605 + ], + "type": "inline_equation", + "content": "\\rho(t)" + }, + { + "bbox": [ + 304, + 544, + 545, + 605 + ], + "type": "text", + "content": " is a penalty term to discourage sampling a frame from the same pose more than once, which we weight by " + }, + { + "bbox": [ + 304, + 544, + 545, + 605 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 304, + 544, + 545, + 605 + ], + "type": "text", + "content": ". We define " + }, + { + "bbox": [ + 304, + 544, + 545, + 605 + ], + "type": "inline_equation", + "content": "\\Delta Q(t)" + }, + { + "bbox": [ + 304, + 544, + 545, + 605 + ], + "type": "text", + "content": " as" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 611, + 544, + 627 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 611, + 544, + 627 + ], + "spans": [ + { + "bbox": [ + 307, + 611, + 544, + 627 + ], + "type": "interline_equation", + "content": "\\Delta Q (t) = Q \\left(\\tilde {M} (t) \\mid O ^ {M} (t)\\right) - Q \\left(\\tilde {M} (t) \\mid \\left(O ^ {M} (t) \\setminus V _ {t}\\right)\\right),", + "image_path": "ff1068968fb9571f161eca6f61564d22d5effab6e822f49f9bf3a7e696d0e3b4.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 634, + 547, + 682 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 634, + 547, + 682 + ], + "spans": [ + { + "bbox": [ + 304, + 634, + 547, + 682 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 634, + 547, + 682 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 304, + 634, + 547, + 682 + ], + "type": "text", + "content": " is a map quality measure, " + }, + { + "bbox": [ + 304, + 634, + 547, + 682 + ], + "type": "inline_equation", + "content": "Q(X|Y)" + }, + { + "bbox": [ + 304, + 634, + 547, + 682 + ], + "type": "text", + "content": " represents the quality of map estimate " + }, + { + "bbox": [ + 304, + 634, + 547, + 682 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 304, + 634, + 547, + 682 + ], + "type": "text", + "content": " given inputs " + }, + { + "bbox": [ + 304, + 634, + 547, + 682 + ], + "type": "inline_equation", + "content": "Y" + }, + { + "bbox": [ + 304, + 634, + 547, + 682 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 634, + 547, + 682 + ], + "type": "inline_equation", + "content": "(O^M(t) \\setminus V_t)" + }, + { + "bbox": [ + 304, + 634, + 547, + 682 + ], + "type": "text", + "content": " denotes the mapper inputs devoid of any visual frame for the current step. We define " + }, + { + "bbox": [ + 304, + 634, + 547, + 682 + ], + "type": "inline_equation", + "content": "\\rho(t)" + }, + { + "bbox": [ + 304, + 634, + 547, + 682 + ], + "type": "text", + "content": " as" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 339, + 691, + 512, + 716 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 339, + 691, + 512, + 716 + ], + "spans": [ + { + "bbox": [ + 339, + 691, + 512, + 716 + ], + "type": "interline_equation", + "content": "\\rho (t) = \\sum_ {i = 1.. 2} a _ {i, t} * \\mathbb {1} (V _ {i, t} \\in O ^ {M} (t - 1)),", + "image_path": "209a1f7e2198c72dbd7d33c860fc760a2a037f0010b29a523d38ee5b946ecacd.jpg" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10558" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 203 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 203 + ], + "type": "text", + "content": "where the indicator function checks if " + }, + { + "bbox": [ + 46, + 72, + 289, + 203 + ], + "type": "inline_equation", + "content": "V_{i,t}" + }, + { + "bbox": [ + 46, + 72, + 289, + 203 + ], + "type": "text", + "content": " was used in mapping before. While " + }, + { + "bbox": [ + 46, + 72, + 289, + 203 + ], + "type": "inline_equation", + "content": "\\Delta Q(t)" + }, + { + "bbox": [ + 46, + 72, + 289, + 203 + ], + "type": "text", + "content": " incentivizes sampling frames that provide a big boost to the mapping accuracy over skipping them, " + }, + { + "bbox": [ + 46, + 72, + 289, + 203 + ], + "type": "inline_equation", + "content": "\\rho(t)" + }, + { + "bbox": [ + 46, + 72, + 289, + 203 + ], + "type": "text", + "content": " penalizes wasting the visual budget on redundant sampling, thereby maximizing mapping performance within the constraints of a limited budget. We set " + }, + { + "bbox": [ + 46, + 72, + 289, + 203 + ], + "type": "inline_equation", + "content": "\\rho = 0.03" + }, + { + "bbox": [ + 46, + 72, + 289, + 203 + ], + "type": "text", + "content": " in all our experiments and define " + }, + { + "bbox": [ + 46, + 72, + 289, + 203 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 46, + 72, + 289, + 203 + ], + "type": "text", + "content": " as the average F1 score over the occupied and free classes in a predicted occupancy map. We empirically observe that alternate penalty terms based on visual similarity, like the ORB [55] feature matching similarity, do not improve the model performance." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 203, + 288, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 203, + 288, + 251 + ], + "spans": [ + { + "bbox": [ + 46, + 203, + 288, + 251 + ], + "type": "text", + "content": "We train " + }, + { + "bbox": [ + 46, + 203, + 288, + 251 + ], + "type": "inline_equation", + "content": "\\pi^V" + }, + { + "bbox": [ + 46, + 203, + 288, + 251 + ], + "type": "text", + "content": " with Decentralized Distributed PPO (DDPPO) [66]. The DD-PPO loss consists of a value loss, policy loss and an entropy loss to promote exploration (see Supp. for details)." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 256, + 289, + 281 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 256, + 289, + 281 + ], + "spans": [ + { + "bbox": [ + 47, + 256, + 289, + 281 + ], + "type": "text", + "content": "Mapper training. At each step " + }, + { + "bbox": [ + 47, + 256, + 289, + 281 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 47, + 256, + 289, + 281 + ], + "type": "text", + "content": ", we train the shared mapper " + }, + { + "bbox": [ + 47, + 256, + 289, + 281 + ], + "type": "inline_equation", + "content": "f^{m}" + }, + { + "bbox": [ + 47, + 256, + 289, + 281 + ], + "type": "text", + "content": " with a loss " + }, + { + "bbox": [ + 47, + 256, + 289, + 281 + ], + "type": "inline_equation", + "content": "\\mathcal{L}^{M}(t)" + }, + { + "bbox": [ + 47, + 256, + 289, + 281 + ], + "type": "text", + "content": ", such that" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 70, + 289, + 263, + 319 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 289, + 263, + 319 + ], + "spans": [ + { + "bbox": [ + 70, + 289, + 263, + 319 + ], + "type": "interline_equation", + "content": "\\mathcal {L} ^ {M} (t) = \\frac {1}{2 \\times t} \\sum_ {i = 1 \\ldots 2} \\sum_ {j = 1 \\ldots t} \\mathrm {B C E} (\\tilde {M} _ {i, j}, M _ {i, j}),", + "image_path": "c0230b75cddc840c344ab97ee14461820c93b1fa1c1a4a6d64bb071e1feedc84.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 328, + 287, + 356 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 328, + 287, + 356 + ], + "spans": [ + { + "bbox": [ + 46, + 328, + 287, + 356 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 328, + 287, + 356 + ], + "type": "inline_equation", + "content": "\\mathrm{BCE}(\\tilde{M}_{i,j}, M_{i,j})" + }, + { + "bbox": [ + 46, + 328, + 287, + 356 + ], + "type": "text", + "content": " is the average binary cross entropy loss between " + }, + { + "bbox": [ + 46, + 328, + 287, + 356 + ], + "type": "inline_equation", + "content": "\\tilde{M}_{i,j}" + }, + { + "bbox": [ + 46, + 328, + 287, + 356 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 328, + 287, + 356 + ], + "type": "inline_equation", + "content": "M_{i,j}" + }, + { + "bbox": [ + 46, + 328, + 287, + 356 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 358, + 289, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 358, + 289, + 479 + ], + "spans": [ + { + "bbox": [ + 46, + 358, + 289, + 479 + ], + "type": "text", + "content": "Training curriculum. To train our model, we first pretrain mapper " + }, + { + "bbox": [ + 46, + 358, + 289, + 479 + ], + "type": "inline_equation", + "content": "f^{m}" + }, + { + "bbox": [ + 46, + 358, + 289, + 479 + ], + "type": "text", + "content": " in two phases and then train the policy " + }, + { + "bbox": [ + 46, + 358, + 289, + 479 + ], + "type": "inline_equation", + "content": "\\pi^{V}" + }, + { + "bbox": [ + 46, + 358, + 289, + 479 + ], + "type": "text", + "content": " while keeping " + }, + { + "bbox": [ + 46, + 358, + 289, + 479 + ], + "type": "inline_equation", + "content": "f^{m}" + }, + { + "bbox": [ + 46, + 358, + 289, + 479 + ], + "type": "text", + "content": " frozen. In phase 1, we train " + }, + { + "bbox": [ + 46, + 358, + 289, + 479 + ], + "type": "inline_equation", + "content": "f^{m}" + }, + { + "bbox": [ + 46, + 358, + 289, + 479 + ], + "type": "text", + "content": " without visual sampling, i.e. all visual frames are provided at each step. In phase 2, we finetune the pretrained weights of " + }, + { + "bbox": [ + 46, + 358, + 289, + 479 + ], + "type": "inline_equation", + "content": "f^{m}" + }, + { + "bbox": [ + 46, + 358, + 289, + 479 + ], + "type": "text", + "content": " from phase 1 on episodes where we randomly drop views to satisfy the budget " + }, + { + "bbox": [ + 46, + 358, + 289, + 479 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 46, + 358, + 289, + 479 + ], + "type": "text", + "content": ". While phase 1 improves convergence when training with visual sampling, phase 2 helps with reward stationarity when training our RL policy. Both phases use the same training, validation and testing data." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 490, + 128, + 502 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 490, + 128, + 502 + ], + "spans": [ + { + "bbox": [ + 47, + 490, + 128, + 502 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 510, + 289, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 510, + 289, + 629 + ], + "spans": [ + { + "bbox": [ + 46, + 510, + 289, + 629 + ], + "type": "text", + "content": "Experimental setup. For our main experiments, we use SoundSpaces [8] acoustic simulations with AI-Habitat [57] and Matterport3D [3] visual scenes. While Matterport3D provides dense 3D meshes and image scans of real-world houses and other indoor scenes, SoundSpaces provides room impulse responses (RIRs) at a spatial resolution of " + }, + { + "bbox": [ + 46, + 510, + 289, + 629 + ], + "type": "inline_equation", + "content": "1\\mathrm{m}" + }, + { + "bbox": [ + 46, + 510, + 289, + 629 + ], + "type": "text", + "content": " for Matterport3D that model all real-world acoustic phenomena [8]. This setup allows us to evaluate with as many as 83 scenes, split in 56/10/17 for train/val/test, compare against relevant prior work [52, 53] and report reproducible results." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 630, + 289, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 630, + 289, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 630, + 289, + 714 + ], + "type": "text", + "content": "We also collect real-world data in a mock-up apartment due to the absence of a publicly available alternative suited for our task. We capture a dense set of RGB images using a Samsung S22 camera and generate the corresponding depth images using monocular depth estimation [20, 33]. To compute the RIRs, following [23], we generate a sinusoidal sweep sound from " + }, + { + "bbox": [ + 46, + 630, + 289, + 714 + ], + "type": "inline_equation", + "content": "20\\mathrm{Hz} - 20\\mathrm{kHz}" + }, + { + "bbox": [ + 46, + 630, + 289, + 714 + ], + "type": "text", + "content": " with a loudspeaker at source" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 72, + 547, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 156 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 156 + ], + "type": "text", + "content": "location, capture it with an Eigenmike at a receiver location, and convolve the spatial sound with the inverse of the sweep sound to retrieve the RIR. All capturing devices are placed at a height of " + }, + { + "bbox": [ + 304, + 72, + 547, + 156 + ], + "type": "inline_equation", + "content": "1.5\\mathrm{m}" + }, + { + "bbox": [ + 304, + 72, + 547, + 156 + ], + "type": "text", + "content": ". We generate occupancy maps by backprojecting the depth images (cf. Sec. 4.1) and register them onto a shared top-down map before taking egocentric crops to generate the local occupancy inputs and targets." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 156, + 547, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 156, + 547, + 204 + ], + "spans": [ + { + "bbox": [ + 304, + 156, + 547, + 204 + ], + "type": "text", + "content": "Note that both datasets have real-world visuals as they are captured in the real environments; SoundSpaces has high-quality simulated audio while the apartment data has real-world collected audio RIRs." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 213, + 548, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 213, + 548, + 453 + ], + "spans": [ + { + "bbox": [ + 304, + 213, + 548, + 453 + ], + "type": "text", + "content": "Conversation episode. For each episode,(simulated or real), we randomly place the egos in a scene. Episode length is " + }, + { + "bbox": [ + 304, + 213, + 548, + 453 + ], + "type": "inline_equation", + "content": "T = 16" + }, + { + "bbox": [ + 304, + 213, + 548, + 453 + ], + "type": "text", + "content": " and 8 for simulation and real, respectively. At each step, the egos execute a movement from " + }, + { + "bbox": [ + 304, + 213, + 548, + 453 + ], + "type": "inline_equation", + "content": "A = \\{MoveForward, TurnLeft, TurnRight\\}" + }, + { + "bbox": [ + 304, + 213, + 548, + 453 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 213, + 548, + 453 + ], + "type": "inline_equation", + "content": "MoveForward" + }, + { + "bbox": [ + 304, + 213, + 548, + 453 + ], + "type": "text", + "content": " moves an ego forward by " + }, + { + "bbox": [ + 304, + 213, + 548, + 453 + ], + "type": "inline_equation", + "content": "1\\mathrm{m}" + }, + { + "bbox": [ + 304, + 213, + 548, + 453 + ], + "type": "text", + "content": ", and the " + }, + { + "bbox": [ + 304, + 213, + 548, + 453 + ], + "type": "inline_equation", + "content": "Turn" + }, + { + "bbox": [ + 304, + 213, + 548, + 453 + ], + "type": "text", + "content": " actions rotate the ego by " + }, + { + "bbox": [ + 304, + 213, + 548, + 453 + ], + "type": "inline_equation", + "content": "90^{\\circ}" + }, + { + "bbox": [ + 304, + 213, + 548, + 453 + ], + "type": "text", + "content": ", as required by SoundSpaces. Further, either of the egos speaks or both speak with equal probability of " + }, + { + "bbox": [ + 304, + 213, + 548, + 453 + ], + "type": "inline_equation", + "content": "\\frac{1}{3}" + }, + { + "bbox": [ + 304, + 213, + 548, + 453 + ], + "type": "text", + "content": " at every step, i.e., there are no moments of total silence, but one ego may be silent. The egos stay between " + }, + { + "bbox": [ + 304, + 213, + 548, + 453 + ], + "type": "inline_equation", + "content": "1 - 3\\mathrm{m}" + }, + { + "bbox": [ + 304, + 213, + 548, + 453 + ], + "type": "text", + "content": " from each other (typical in a conversation) so that they do not collide and each ego is audible by the other at all times. This results in train/val splits of 1,955,334/100 episodes in simulation, and a simulated/real-world test split of 1000/27 episodes. See Supp. for additional analysis on the model performance vs. training data size. We set the visual budget to " + }, + { + "bbox": [ + 304, + 213, + 548, + 453 + ], + "type": "inline_equation", + "content": "B = 2" + }, + { + "bbox": [ + 304, + 213, + 548, + 453 + ], + "type": "text", + "content": " for our main experiments (see Supp. for " + }, + { + "bbox": [ + 304, + 213, + 548, + 453 + ], + "type": "inline_equation", + "content": "B = 4, 6" + }, + { + "bbox": [ + 304, + 213, + 548, + 453 + ], + "type": "text", + "content": " evaluations). Note that these episodes are solely to generate video data; our task requires processing passive video, not controlling embodied agents." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 456, + 548, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 456, + 548, + 601 + ], + "spans": [ + { + "bbox": [ + 304, + 456, + 548, + 601 + ], + "type": "text", + "content": "Observations and model output. For the occupancy maps, consistent with [4, 11, 53], we generate " + }, + { + "bbox": [ + 304, + 456, + 548, + 601 + ], + "type": "inline_equation", + "content": "(31 \\times 31 \\times 2)" + }, + { + "bbox": [ + 304, + 456, + 548, + 601 + ], + "type": "text", + "content": "-dimensional input maps that cover " + }, + { + "bbox": [ + 304, + 456, + 548, + 601 + ], + "type": "inline_equation", + "content": "3.1 \\times 3.1 \\mathrm{~m}^2" + }, + { + "bbox": [ + 304, + 456, + 548, + 601 + ], + "type": "text", + "content": " in area at a resolution of " + }, + { + "bbox": [ + 304, + 456, + 548, + 601 + ], + "type": "inline_equation", + "content": "0.1 \\mathrm{~m}" + }, + { + "bbox": [ + 304, + 456, + 548, + 601 + ], + "type": "text", + "content": ", and set the local target map size to " + }, + { + "bbox": [ + 304, + 456, + 548, + 601 + ], + "type": "inline_equation", + "content": "H \\times W = 6.4 \\times 6.4 \\mathrm{~m}^2 (\\sim 41 \\mathrm{~m}^2)" + }, + { + "bbox": [ + 304, + 456, + 548, + 601 + ], + "type": "text", + "content": ". For speech, we use 100 distinct speakers from LibriSpeech [49], split in 80/11 for heard/unheard, where unheard speech is only used in testing. We assume camera poses, as modern AR devices are equipped with motion sensors that can robustly estimate relative poses [41]. We test our robustness to ambient sounds that get mixed with the egos' speech, and incorporate odometry noise models [52, 53] in Supp." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 605, + 548, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 548, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 548, + 715 + ], + "type": "text", + "content": "Evaluation settings. We evaluate our model in two settings: 1) passive mapping, the mapper has access to all visual frames in an episode (i.e., the camera is always on), and 2) active mapping, where the mapping agent has to actively sample frames to meet the visual budget " + }, + { + "bbox": [ + 304, + 605, + 548, + 715 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 304, + 605, + 548, + 715 + ], + "type": "text", + "content": ". This helps disentangle our modeling contributions—whereas passive mapping lets us show improvements in the mapper " + }, + { + "bbox": [ + 304, + 605, + 548, + 715 + ], + "type": "inline_equation", + "content": "f^{M}" + }, + { + "bbox": [ + 304, + 605, + 548, + 715 + ], + "type": "text", + "content": " over existing methods [52, 53], active mapping helps demonstrate the benefits of smart visual sampling." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "10559" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 63, + 70, + 272, + 203 + ], + "blocks": [ + { + "bbox": [ + 63, + 70, + 272, + 203 + ], + "lines": [ + { + "bbox": [ + 63, + 70, + 272, + 203 + ], + "spans": [ + { + "bbox": [ + 63, + 70, + 272, + 203 + ], + "type": "table", + "html": "
ModelSimulationReal world
F1 score ↑IoU ↑F1 score ↑IoU ↑
All-occupied63.448.836.223.8
Register-inputs72.660.150.835.0
OccAnt [53]74.562.753.938.3
AV-Floorplan [52]79.367.954.538.7
Ours81.871.455.539.2
Ours w/o vision72.860.350.835.0
Ours w/o audio78.166.754.138.0
Ours w/o \\( E_i' \\)'s speech81.570.955.439.1
Ours w/o shared mapping80.770.054.938.6
Ours w/o modality tag79.167.854.538.1
", + "image_path": "8d64f7683cf54910909e42e0ad498bf5e0d29481f8cfed33a002d27d3b78d204.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 205, + 287, + 228 + ], + "lines": [ + { + "bbox": [ + 47, + 205, + 287, + 228 + ], + "spans": [ + { + "bbox": [ + 47, + 205, + 287, + 228 + ], + "type": "text", + "content": "Table 1. Passive mapping performance " + }, + { + "bbox": [ + 47, + 205, + 287, + 228 + ], + "type": "inline_equation", + "content": "(\\%)" + }, + { + "bbox": [ + 47, + 205, + 287, + 228 + ], + "type": "text", + "content": ". All our gains are statistically significant " + }, + { + "bbox": [ + 47, + 205, + 287, + 228 + ], + "type": "inline_equation", + "content": "(p \\leq 0.05)" + }, + { + "bbox": [ + 47, + 205, + 287, + 228 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 240, + 288, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 240, + 288, + 312 + ], + "spans": [ + { + "bbox": [ + 46, + 240, + 288, + 312 + ], + "type": "text", + "content": "We use standard evaluation metrics [53]: F1 score and IoU (intersection over union) between the predicted and target scene maps. For both metrics, we report the mean over the free and occupied classes. For active mapping, we average the metrics over 3 random seeds. We use the following baselines to compare our model's efficacy." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 314, + 120, + 325 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 314, + 120, + 325 + ], + "spans": [ + { + "bbox": [ + 47, + 314, + 120, + 325 + ], + "type": "text", + "content": "Passive mapping:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 325, + 287, + 480 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 47, + 325, + 287, + 349 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 325, + 287, + 349 + ], + "spans": [ + { + "bbox": [ + 47, + 325, + 287, + 349 + ], + "type": "text", + "content": "- All-occupied: a naive baseline that predicts all locations in its map estimate as occupied" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 350, + 287, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 350, + 287, + 373 + ], + "spans": [ + { + "bbox": [ + 47, + 350, + 287, + 373 + ], + "type": "text", + "content": "- Register-inputs: a naive baseline that registers the input maps onto a shared map and uses it as its prediction" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 373, + 287, + 409 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 373, + 287, + 409 + ], + "spans": [ + { + "bbox": [ + 47, + 373, + 287, + 409 + ], + "type": "text", + "content": "- OccAnt [53]: a vision-only SOTA model that uses the RGB-D images at each step to anticipate the occupancy of the area around an ego that's outside its visible range." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 409, + 287, + 480 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 409, + 287, + 480 + ], + "spans": [ + { + "bbox": [ + 47, + 409, + 287, + 480 + ], + "type": "text", + "content": "- AV-Floorplan [52]: an audio-visual SOTA model that passively predicts the floorplan of a scene using a walk-through in it, where the audio is either self-generated or comes from semantic sources in the scene. We adapt the model for our occupancy prediction task and give it the exact same audio-visual observations as our model." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 46, + 482, + 115, + 493 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 482, + 115, + 493 + ], + "spans": [ + { + "bbox": [ + 46, + 482, + 115, + 493 + ], + "type": "text", + "content": "Active mapping:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 494, + 287, + 564 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 47, + 494, + 287, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 494, + 287, + 517 + ], + "spans": [ + { + "bbox": [ + 47, + 494, + 287, + 517 + ], + "type": "text", + "content": "- Random: an agent that selects visual frames randomly for each ego as long as the budget allows" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 517, + 287, + 541 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 517, + 287, + 541 + ], + "spans": [ + { + "bbox": [ + 47, + 517, + 287, + 541 + ], + "type": "text", + "content": "- Greedy: an agent that greedily uses up the visual budget by sampling frames as early as possible" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 47, + 542, + 287, + 564 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 542, + 287, + 564 + ], + "spans": [ + { + "bbox": [ + 47, + 542, + 287, + 564 + ], + "type": "text", + "content": "- Unique-posed: an agent that samples a frame for every new ego pose in the episode" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 46, + 566, + 287, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 566, + 287, + 613 + ], + "spans": [ + { + "bbox": [ + 46, + 566, + 287, + 613 + ], + "type": "text", + "content": "In active mapping, we use the model from the second pretraining phase (Sec. 4.4) as the mapper for all models for fair comparison. Thus, any difference in performance is due to the quality of each method's sampling decisions." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 47, + 614, + 287, + 638 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 614, + 287, + 638 + ], + "spans": [ + { + "bbox": [ + 47, + 614, + 287, + 638 + ], + "type": "text", + "content": "See Supp. for all other details like network architectures, training hyperparameters, and baseline implementation." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 47, + 647, + 176, + 659 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 647, + 176, + 659 + ], + "spans": [ + { + "bbox": [ + 47, + 647, + 176, + 659 + ], + "type": "text", + "content": "5.1. Map prediction results" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 46, + 665, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 288, + 714 + ], + "type": "text", + "content": "Passive mapping. Table 1 (top) reports the prediction quality of all models in the passive mapping setting. Naive baselines (All-occupied, Register-inputs) perform worse than the learned models, showing the complexity of our map pre" + } + ] + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 307, + 72, + 424, + 153 + ], + "blocks": [ + { + "bbox": [ + 307, + 72, + 424, + 153 + ], + "lines": [ + { + "bbox": [ + 307, + 72, + 424, + 153 + ], + "spans": [ + { + "bbox": [ + 307, + 72, + 424, + 153 + ], + "type": "image", + "image_path": "4b8927fbc208d5e693d9e96cd47be53f85e258be176700e1909327fa1a4cd3f7.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 342, + 156, + 389, + 165 + ], + "lines": [ + { + "bbox": [ + 342, + 156, + 389, + 165 + ], + "spans": [ + { + "bbox": [ + 342, + 156, + 389, + 165 + ], + "type": "text", + "content": "(a) Simulation" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 427, + 71, + 545, + 154 + ], + "blocks": [ + { + "bbox": [ + 427, + 71, + 545, + 154 + ], + "lines": [ + { + "bbox": [ + 427, + 71, + 545, + 154 + ], + "spans": [ + { + "bbox": [ + 427, + 71, + 545, + 154 + ], + "type": "image", + "image_path": "668d11c8601cda5195c65b3a29d2946ffeb20eae8e5fc6c9aa3fa8a2b3ea92a4.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 462, + 156, + 511, + 165 + ], + "lines": [ + { + "bbox": [ + 462, + 156, + 511, + 165 + ], + "spans": [ + { + "bbox": [ + 462, + 156, + 511, + 165 + ], + "type": "text", + "content": "(b) Real world" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 354, + 200, + 498, + 297 + ], + "blocks": [ + { + "bbox": [ + 323, + 175, + 528, + 186 + ], + "lines": [ + { + "bbox": [ + 323, + 175, + 528, + 186 + ], + "spans": [ + { + "bbox": [ + 323, + 175, + 528, + 186 + ], + "type": "text", + "content": "Figure 3. Active mapping performance vs. episode step." + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 354, + 200, + 498, + 297 + ], + "lines": [ + { + "bbox": [ + 354, + 200, + 498, + 297 + ], + "spans": [ + { + "bbox": [ + 354, + 200, + 498, + 297 + ], + "type": "image", + "image_path": "5a69f564045c47faa92d760d19b7690880249e00dddc60a1ba3436c94afc154a.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 306, + 307, + 547, + 329 + ], + "lines": [ + { + "bbox": [ + 306, + 307, + 547, + 329 + ], + "spans": [ + { + "bbox": [ + 306, + 307, + 547, + 329 + ], + "type": "text", + "content": "Figure 4. Impact of the other ego's speech on passive mapping vs. distance between the egos." + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + } + ], + "index": 23 + }, + { + "bbox": [ + 304, + 343, + 547, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 343, + 547, + 403 + ], + "spans": [ + { + "bbox": [ + 304, + 343, + 547, + 403 + ], + "type": "text", + "content": "diction task. AV-Floorplan [52] fares the best among all baselines. Its improvement over OccAnt [53] demonstrates the benefits of exploiting the spatial cues in audio for mapping and using an attention-based model to leverage the long- and short-range correlations in the audio-visual inputs." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 304, + 404, + 547, + 513 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 404, + 547, + 513 + ], + "spans": [ + { + "bbox": [ + 304, + 404, + 547, + 513 + ], + "type": "text", + "content": "Our method outperforms all baselines. Its improvement over AV-Floorplan [52] underlines the efficacy of performing attention at different granularities—across modalities, within a single modality and within a single input—guided by our positional and modality type embeddings. It also generalizes to the real-world setting and retains its benefits over the baselines, even without retraining on the real-world data. However, we do observe a drop in performance gains, probably due to the large sim-to-real gap." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 304, + 521, + 547, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 521, + 547, + 617 + ], + "spans": [ + { + "bbox": [ + 304, + 521, + 547, + 617 + ], + "type": "text", + "content": "Active mapping. Fig. 3 shows the active mapping performance as a function of episode progress. Employing naive heuristics for sampling, like Random or Greedy, is insufficient for high-quality mapping, emphasizing the high levels of redundancy in the visual frames. Unique-pose improves over both Random and Greedy, showing that sampling diverse viewpoints provides more information about the underlying scene geometry." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 304, + 617, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 617, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 617, + 547, + 713 + ], + "type": "text", + "content": "Although the baselines make initial progress, they flatten quickly and our model eventually outperforms them all, on both real-world and simulated data (gains are again statistically significant, " + }, + { + "bbox": [ + 304, + 617, + 547, + 713 + ], + "type": "inline_equation", + "content": "p \\leq 0.05" + }, + { + "bbox": [ + 304, + 617, + 547, + 713 + ], + "type": "text", + "content": "). This highlights the benefits of learning a smart policy that, given the audio streams and its past visual samples, understands the value of sampling a visual frame for mapping by taking cues from our novel reward. Besides, on the real-world data, we see better per" + } + ] + } + ], + "index": 28 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10560" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 48, + 70, + 547, + 229 + ], + "blocks": [ + { + "bbox": [ + 48, + 70, + 547, + 229 + ], + "lines": [ + { + "bbox": [ + 48, + 70, + 547, + 229 + ], + "spans": [ + { + "bbox": [ + 48, + 70, + 547, + 229 + ], + "type": "image", + "image_path": "cc04797f38573fc942b96e0d51e5c37a66bce51c35caabc21e0d304352baabe8.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 237, + 546, + 260 + ], + "lines": [ + { + "bbox": [ + 46, + 237, + 546, + 260 + ], + "spans": [ + { + "bbox": [ + 46, + 237, + 546, + 260 + ], + "type": "text", + "content": "Figure 5. Sample episodes for our active mapping model. While our policy samples only the salient visual frames, our mapper can both complete partially seen objects as well as anticipate objects never seen before in the sampled visuals (red boxes on the maps)." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 274, + 289, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 274, + 289, + 357 + ], + "spans": [ + { + "bbox": [ + 46, + 274, + 289, + 357 + ], + "type": "text", + "content": "formance margins over the baselines towards the end of episodes, showing that our policy can adaptively defer visual sampling to improve mapping. Owing to smart sampling, the per-episode reduction in the capture / processing cost for " + }, + { + "bbox": [ + 46, + 274, + 289, + 357 + ], + "type": "inline_equation", + "content": "B = 2" + }, + { + "bbox": [ + 46, + 274, + 289, + 357 + ], + "type": "text", + "content": " is 74 Watts (W) / 7.2 GFLOPS in simulation and 37 W / 3.6 GFLOPS in the real world video. See Supp. for further analysis of our model's power and time cost." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 369, + 140, + 382 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 369, + 140, + 382 + ], + "spans": [ + { + "bbox": [ + 47, + 369, + 140, + 382 + ], + "type": "text", + "content": "5.2. Model analysis" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 389, + 288, + 663 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 389, + 288, + 663 + ], + "spans": [ + { + "bbox": [ + 46, + 389, + 288, + 663 + ], + "type": "text", + "content": "Ablations. In Table 1 (bottom), we ablate the components of our model for passive mapping. Upon removing audio, our model experiences a large drop in mapping performance, which indicates that our model leverages complementary spatial cues in audio and vision. We also see a drop in the map quality when our model lacks access to the speech from the other ego " + }, + { + "bbox": [ + 46, + 389, + 288, + 663 + ], + "type": "inline_equation", + "content": "(E_i')" + }, + { + "bbox": [ + 46, + 389, + 288, + 663 + ], + "type": "text", + "content": ". This shows that " + }, + { + "bbox": [ + 46, + 389, + 288, + 663 + ], + "type": "inline_equation", + "content": "E_i'" + }, + { + "bbox": [ + 46, + 389, + 288, + 663 + ], + "type": "text", + "content": "'s speech can better reveal the more global scene geometry than " + }, + { + "bbox": [ + 46, + 389, + 288, + 663 + ], + "type": "inline_equation", + "content": "E_i" + }, + { + "bbox": [ + 46, + 389, + 288, + 663 + ], + "type": "text", + "content": "'s own speech. Fig. 4 further shows that the impact of the other ego's speech becomes more prominent for larger inter-ego distances " + }, + { + "bbox": [ + 46, + 389, + 288, + 663 + ], + "type": "inline_equation", + "content": "(3 - 5\\mathrm{m}" + }, + { + "bbox": [ + 46, + 389, + 288, + 663 + ], + "type": "text", + "content": " vs. " + }, + { + "bbox": [ + 46, + 389, + 288, + 663 + ], + "type": "inline_equation", + "content": "1 - 3\\mathrm{m})" + }, + { + "bbox": [ + 46, + 389, + 288, + 663 + ], + "type": "text", + "content": ", in which case the two types of speech are dissimilar enough to carry complementary geometric cues, but reduces for even larger distances " + }, + { + "bbox": [ + 46, + 389, + 288, + 663 + ], + "type": "inline_equation", + "content": "(5\\mathrm{m}" + }, + { + "bbox": [ + 46, + 389, + 288, + 663 + ], + "type": "text", + "content": " or more), in which case " + }, + { + "bbox": [ + 46, + 389, + 288, + 663 + ], + "type": "inline_equation", + "content": "E_i'" + }, + { + "bbox": [ + 46, + 389, + 288, + 663 + ], + "type": "text", + "content": " is too far for its speech to carry useful cues about " + }, + { + "bbox": [ + 46, + 389, + 288, + 663 + ], + "type": "inline_equation", + "content": "E_i" + }, + { + "bbox": [ + 46, + 389, + 288, + 663 + ], + "type": "text", + "content": "'s local scene geometry. Moreover, unlike the ablation that does not perform shared mapping, our model benefits significantly from jointly attending to the observations of the egos and exploiting the complementary information in them—even though both models use the exact same audio-visual observations, including both speech from self and the other ego. Finally, without our modality type embeddings, our model does worse, showing their role in learning complementary features from vision and audio." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 665, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 288, + 713 + ], + "type": "text", + "content": "For active mapping, Fig. 3 shows a drop in the mapping performance upon removing audio from the policy inputs. This implies that our policy exploits audio to reason about the level of redundancy in a new visual frame and improve" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 274, + 545, + 309 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 274, + 545, + 309 + ], + "spans": [ + { + "bbox": [ + 304, + 274, + 545, + 309 + ], + "type": "text", + "content": "the mapping quality vs. visual budget tradeoff. On the more challenging real-world setting, audio plays an even bigger role, as shown by the larger performance drop in Fig. 3b." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 310, + 546, + 358 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 310, + 546, + 358 + ], + "spans": [ + { + "bbox": [ + 304, + 310, + 546, + 358 + ], + "type": "text", + "content": "See Supp. for similar results with 1) ambient sounds, 2) unheard speech, 3) higher budgets, 4) sensor noise, 5) larger target maps, 6) different ego initializations, and 7) multiple random data splits." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 367, + 546, + 463 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 367, + 546, + 463 + ], + "spans": [ + { + "bbox": [ + 304, + 367, + 546, + 463 + ], + "type": "text", + "content": "Qualitative results. Fig. 5 shows two successful active mapping episodes of our method. Note how our model samples views that tend have to little visual overlap but are informative of the surrounding geometry (both occupied and free spaces). Besides, it is able to complete structures only partially visible in the sampled views, and more interestingly, leverage the synergy of audio and vision to anticipate unseen areas (red boxes on the occupancy maps in Fig. 5)." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 472, + 547, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 472, + 547, + 579 + ], + "spans": [ + { + "bbox": [ + 304, + 472, + 547, + 579 + ], + "type": "text", + "content": "Failure cases. We notice two common failure cases with active mapping: episodes where the egos stay at the same location, leading to very few informative visual frames to sample from; and episodes with highly unique visual samples at every trajectory step, in which case each sample is useful and our model behaves similar to Unique-pose or Greedy. For passive mapping, our model fails with very complex scenes that commonly have objects in spaces where both vision and audio cannot reach (e.g. narrow corners)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 590, + 378, + 601 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 590, + 378, + 601 + ], + "spans": [ + { + "bbox": [ + 306, + 590, + 378, + 601 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 607, + 547, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 607, + 547, + 689 + ], + "spans": [ + { + "bbox": [ + 304, + 607, + 547, + 689 + ], + "type": "text", + "content": "We introduce Chat2Map, a new task aimed at scene mapping using audio-visual feeds from egocentric conversations. We develop a novel approach for Chat2Map comprising a shared scene mapper and a visual sampling policy based on a novel reinforcement learner that smartly samples the visuals only when necessary. We show promising performance on both simulated and real-world data from over 80 scenes." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 691, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 691, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 691, + 545, + 713 + ], + "type": "text", + "content": "Acknowledgements: UT Austin is supported in part by the IFML NSF AI Institute. KG is paid as a research scientist by Meta." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10561" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 133 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 133 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 133 + ], + "type": "text", + "content": "[1] Ego-Noise Predictions for Echolocation in Wheeled Robots, volume ALIFE 2019: The 2019 Conference on Artificial Life of ALIFE 2022: The 2022 Conference on Artificial Life, 07 2019. 1, 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 134, + 288, + 166 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 134, + 288, + 166 + ], + "spans": [ + { + "bbox": [ + 53, + 134, + 288, + 166 + ], + "type": "text", + "content": "[2] Aaron Carroll and Gernot Heiser. An analysis of power consumption in a smartphone. In 2010 USENIX Annual Technical Conference (USENIX ATC 10), 2010. 2, 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 166, + 288, + 220 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 166, + 288, + 220 + ], + "spans": [ + { + "bbox": [ + 53, + 166, + 288, + 220 + ], + "type": "text", + "content": "[3] Angel Chang, Angela Dai, Thomas Funkhouser, Maciej Halber, Matthias Niessner, Manolis Savva, Shuran Song, Andy Zeng, and Yinda Zhang. Matterport3d: Learning from rgb-d data in indoor environments. arXiv preprint arXiv:1709.06158, 2017. 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 220, + 288, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 220, + 288, + 262 + ], + "spans": [ + { + "bbox": [ + 53, + 220, + 288, + 262 + ], + "type": "text", + "content": "[4] Devendra Singh Chaplot, Dhiraj Gandhi, Saurabh Gupta, Abhinav Gupta, and Ruslan Salakhutdinov. Learning to explore using active neural slam. In International Conference on Learning Representations, 2020. 1, 2, 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 262, + 288, + 306 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 262, + 288, + 306 + ], + "spans": [ + { + "bbox": [ + 53, + 262, + 288, + 306 + ], + "type": "text", + "content": "[5] Devendra Singh Chaplot, Dhiraj Prakashchand Gandhi, Abhinav Gupta, and Russ R Salakhutdinov. Object goal navigation using goal-oriented semantic exploration. Advances in Neural Information Processing Systems, 33:4247-4258, 2020. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 305, + 288, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 305, + 288, + 348 + ], + "spans": [ + { + "bbox": [ + 53, + 305, + 288, + 348 + ], + "type": "text", + "content": "[6] Changan Chen, Ziad Al-Halah, and Kristen Grauman. Semantic audio-visual navigation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15516–15525, 2021. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 348, + 288, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 348, + 288, + 392 + ], + "spans": [ + { + "bbox": [ + 53, + 348, + 288, + 392 + ], + "type": "text", + "content": "[7] Changan Chen, Ruohan Gao, Paul T. Calamia, and Kristen Grauman. Visual acoustic matching. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 18836-18846, 2022. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 392, + 288, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 392, + 288, + 445 + ], + "spans": [ + { + "bbox": [ + 53, + 392, + 288, + 445 + ], + "type": "text", + "content": "[8] Changan Chen, Unnat Jain, Carl Schissler, Sebastia Vicenc Amengual Gari, Ziad Al-Halah, Vamsi Krishna Ithapu, Philip Robinson, and Kristen Grauman. Soundspaces: Audiovisual navigation in 3d environments. In ECCV, 2020. 1, 2, 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 445, + 288, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 445, + 288, + 498 + ], + "spans": [ + { + "bbox": [ + 53, + 445, + 288, + 498 + ], + "type": "text", + "content": "[9] Changan Chen, Sagnik Majumder, Ziad Al-Halah, Ruohan Gao, Santhosh Kumar Ramakrishnan, and Kristen Grauman. Learning to set waypoints for audio-visual navigation. In International Conference on Learning Representations, 2021. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 498, + 288, + 552 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 498, + 288, + 552 + ], + "spans": [ + { + "bbox": [ + 48, + 498, + 288, + 552 + ], + "type": "text", + "content": "[10] Jiacheng Chen, Chen Liu, Jiaye Wu, and Yasutaka Furukawa. Floor-sp: Inverse cad for floorplans by sequential room-wise shortest path. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2661-2670, 2019. 1, 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 552, + 288, + 585 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 552, + 288, + 585 + ], + "spans": [ + { + "bbox": [ + 48, + 552, + 288, + 585 + ], + "type": "text", + "content": "[11] Tao Chen, Saurabh Gupta, and Abhinav Gupta. Learning exploration policies for navigation. arXiv preprint arXiv:1903.01959, 2019. 1, 2, 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 585, + 288, + 628 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 585, + 288, + 628 + ], + "spans": [ + { + "bbox": [ + 48, + 585, + 288, + 628 + ], + "type": "text", + "content": "[12] Yangyu Chen, Shuhui Wang, Weigang Zhang, and Qingming Huang. Less is more: Picking informative frames for video captioning. In Proceedings of the European conference on computer vision (ECCV), pages 358-373, 2018. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 628, + 288, + 659 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 628, + 288, + 659 + ], + "spans": [ + { + "bbox": [ + 48, + 628, + 288, + 659 + ], + "type": "text", + "content": "[13] Jesper Christensen, Sascha Hornauer, and Stella Yu. Batvision - learning to see 3d spatial layout with two ears. In ICRA, 2020. 1" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 659, + 288, + 702 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 659, + 288, + 702 + ], + "spans": [ + { + "bbox": [ + 48, + 659, + 288, + 702 + ], + "type": "text", + "content": "[14] Jesper Haahr Christensen, Sascha Hornauer, and Stella X. Yu. Batvision: Learning to see 3d spatial layout with two ears. In 2020 IEEE International Conference on Robotics and Automation (ICRA), pages 1581-1587, 2020. 1, 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 702, + 288, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 702, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 702, + 288, + 713 + ], + "type": "text", + "content": "[15] Abhishek Das, Théophile Gervet, Joshua Romoff, Dhruv" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 547, + 707 + ], + "type": "list", + "angle": 0, + "index": 31, + "blocks": [ + { + "bbox": [ + 326, + 73, + 547, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 547, + 106 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 547, + 106 + ], + "type": "text", + "content": "Batra, Devi Parikh, Michael G. Rabbat, and Joelle Pineau. Tarmac: Targeted multi-agent communication. ArXiv, abs/1810.11187, 2019. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 106, + 547, + 137 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 106, + 547, + 137 + ], + "spans": [ + { + "bbox": [ + 308, + 106, + 547, + 137 + ], + "type": "text", + "content": "[16] Samyak Datta, Sameer Dharur, Vincent Cartillier, Ruta Desai, Mukul Khanna, Dhruv Batra, and Devi Parikh. Episodic memory question answering. 2022. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 137, + 547, + 179 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 137, + 547, + 179 + ], + "spans": [ + { + "bbox": [ + 308, + 137, + 547, + 179 + ], + "type": "text", + "content": "[17] Victoria Dean, Shubham Tulsiani, and Abhinav Gupta. See, hear, explore: Curiosity via audio-visual association. Advances in Neural Information Processing Systems, 33:14961–14972, 2020. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 180, + 547, + 223 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 180, + 547, + 223 + ], + "spans": [ + { + "bbox": [ + 308, + 180, + 547, + 223 + ], + "type": "text", + "content": "[18] Helisa Dhamo, Nassir Navab, and Federico Tombari. Object-driven multi-layer scene decomposition from a single image. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5369-5378, 2019. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 223, + 547, + 266 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 223, + 547, + 266 + ], + "spans": [ + { + "bbox": [ + 308, + 223, + 547, + 266 + ], + "type": "text", + "content": "[19] Ivan Dokmanić, Reza Parhizkar, Andreas Walther, Yue M. Lu, and Martin Vetterli. Acoustic echoes reveal room shape. Proceedings of the National Academy of Sciences, 110(30):12186-12191, 2013. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 266, + 547, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 266, + 547, + 319 + ], + "spans": [ + { + "bbox": [ + 308, + 266, + 547, + 319 + ], + "type": "text", + "content": "[20] Ainaz Eftekhar, Alexander Sax, Jitendra Malik, and Amir Zamir. Omnidata: A scalable pipeline for making multi-task mid-level vision datasets from 3d scans. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10786-10796, 2021. 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 319, + 547, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 319, + 547, + 363 + ], + "spans": [ + { + "bbox": [ + 308, + 319, + 547, + 363 + ], + "type": "text", + "content": "[21] Amine Elhafsi, Boris Ivanovic, Lucas Janson, and Marco Pavone. Map-predictive motion planning in unknown environments. In 2020 IEEE International Conference on Robotics and Automation (ICRA), pages 8552-8558. IEEE, 2020. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 363, + 547, + 395 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 363, + 547, + 395 + ], + "spans": [ + { + "bbox": [ + 308, + 363, + 547, + 395 + ], + "type": "text", + "content": "[22] Itamar Eliakim, Zahi Cohen, Gábor Kósa, and Yossi Yovel. A fully autonomous terrestrial bat-like acoustic robot. PLoS Computational Biology, 14, 2018. 1, 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 395, + 547, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 395, + 547, + 426 + ], + "spans": [ + { + "bbox": [ + 308, + 395, + 547, + 426 + ], + "type": "text", + "content": "[23] Angelo Farina. Simultaneous measurement of impulse response and distortion with a swept-sine technique. Journal of The Audio Engineering Society, 2000. 6" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 426, + 547, + 524 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 426, + 547, + 524 + ], + "spans": [ + { + "bbox": [ + 308, + 426, + 547, + 524 + ], + "type": "text", + "content": "[24] Chuang Gan, Jeremy Schwartz, Seth Alter, Martin Schrimpf, James Traer, Julian De Freitas, Jonas Kubilius, Abhishek Bhandwaldar, Nick Haber, Megumi Sano, Kuno Kim, Elias Wang, Damian Mrowca, Michael Lingelbach, Aidan Curtis, Kevin T. Feigelis, Daniel M. Bear, Dan Gutfreund, David D. Cox, James J. DiCarlo, Josh H. McDermott, Joshua B. Tenenbaum, and Daniel L. K. Yamins. Threadworld: A platform for interactive multi-modal physical simulation. In NeurIPS Track on Datasets and Benchmarks, 2021. 1" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 524, + 547, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 524, + 547, + 578 + ], + "spans": [ + { + "bbox": [ + 308, + 524, + 547, + 578 + ], + "type": "text", + "content": "[25] Chuang Gan, Yiwei Zhang, Jiajun Wu, Boqing Gong, and Joshua B. Tenenbaum. Look, listen, and act: Towards audiovisual embodied navigation. In 2020 IEEE International Conference on Robotics and Automation (ICRA), pages 9701-9707, 2020. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 578, + 547, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 578, + 547, + 643 + ], + "spans": [ + { + "bbox": [ + 308, + 578, + 547, + 643 + ], + "type": "text", + "content": "[26] Ruohan Gao, Changan Chen, Ziad Al-Halah, Carl Schissler, and Kristen Grauman. Visualechoes: Spatial image representation learning through echolocation. In Andrea Vedaldi, Horst Bischof, Thomas Brox, and Jan-Michael Frahm, editors, Computer Vision - ECCV 2020, pages 658-676, Cham, 2020. Springer International Publishing. 1, 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 308, + 643, + 547, + 686 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 643, + 547, + 686 + ], + "spans": [ + { + "bbox": [ + 308, + 643, + 547, + 686 + ], + "type": "text", + "content": "[27] Ruohan Gao, Tae-Hyun Oh, Kristen Grauman, and Lorenzo Torresani. Listen to look: Action recognition by previewing audio. 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10454-10464, 2020. 3" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 308, + 686, + 547, + 707 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 686, + 547, + 707 + ], + "spans": [ + { + "bbox": [ + 308, + 686, + 547, + 707 + ], + "type": "text", + "content": "[28] Kristen Grauman, Andrew Westbury, Eugene Byrne, Zachary Chavis, Antonino Furnari, Rohit Girdhar, Jackson Hamburger," + } + ] + } + ], + "index": 30 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "10562" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 289, + 712 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 67, + 72, + 289, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 72, + 289, + 335 + ], + "spans": [ + { + "bbox": [ + 67, + 72, + 289, + 335 + ], + "type": "text", + "content": "Hao Jiang, Miao Liu, Xingyu Liu, Miguel Martin, Tushar Nagarajan, Ilija Radosavovic, Santhosh Kumar Ramakrishnan, Fiona Ryan, Jayant Sharma, Michael Wray, Mengmeng Xu, Eric Zhongcong Xu, Chen Zhao, Siddhant Bansal, Dhruv Batra, Vincent Cartillier, Sean Crane, Tien Do, Morrie Doulaty, Akshay Erapalli, Christoph Feichtenhofer, Adriano Fragomeni, Qichen Fu, Christian Fuegen, Abrham Gebreselasie, Cristina Gonzalez, James Hillis, Xuhua Huang, Yifei Huang, Wenqi Jia, Weslie Khoo, Jachym Kolar, Satwik Kottur, Anurag Kumar, Federico Landini, Chao Li, Yanghao Li, Zhenqiang Li, Karttikeya Mangalam, Raghava Modhugu, Jonathan Munro, Tullie Murrell, Takumi Nishiyasu, Will Price, Paola Ruiz Puentes, Merey Ramazanova, Leda Sari, Kiran Somasundaram, Audrey Southerland, Yusuke Sugano, Ruijie Tao, Minh Vo, Yuchen Wang, Xindi Wu, Takuma Yagi, Yunyi Zhu, Pablo Arbelaez, David Crandall, Dima Damen, Giovanni Maria Farinella, Bernard Ghanem, Vamsi Krishna Ithapu, C. V. Jawahar, Hanbyul Joo, Kris Kitani, Haizhou Li, Richard Newcombe, Aude Oliva, Hyun Soo Park, James M. Rehg, Yoichi Sato, Jianbo Shi, Mike Zheng Shou, Antonio Torralba, Lorenzo Torresani, Mingfei Yan, and Jitendra Malik. Ego4d: Around the World in 3,000 Hours of Egocentric Video. In IEEE/CVF Computer Vision and Pattern Recognition (CVPR), 2022. 1" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 335, + 288, + 367 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 335, + 288, + 367 + ], + "spans": [ + { + "bbox": [ + 48, + 335, + 288, + 367 + ], + "type": "text", + "content": "[29] Shariq Iqbal and Fei Sha. Coordinated exploration via intrinsic rewards for multi-agent reinforcement learning. arXiv preprint arXiv:1905.12127, 2019. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 368, + 289, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 368, + 289, + 453 + ], + "spans": [ + { + "bbox": [ + 48, + 368, + 289, + 453 + ], + "type": "text", + "content": "[30] Max Jaderberg, Wojciech M. Czarnecki, Iain Dunning, Luke Marris, Guy Lever, Antonio Garcia Castañeda, Charlie Beattie, Neil C. Rabinowitz, Ari S. Morcos, Avraham Ruderman, Nicolas Sonnerat, Tim Green, Louise Deason, Joel Z. Leibo, David Silver, Demis Hassabis, Koray Kavukcuoglu, and Thore Graepel. Human-level performance in 3d multiplayer games with population-based reinforcement learning. Science, 364:859 - 865, 2019. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 454, + 288, + 497 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 454, + 288, + 497 + ], + "spans": [ + { + "bbox": [ + 48, + 454, + 288, + 497 + ], + "type": "text", + "content": "[31] Unnat Jain, Luca Weihs, Eric Kolve, Ali Farhadi, Svetlana Lazebnik, Aniruddha Kembhavi, and Alexander G. Schwing. A cordial sync: Going beyond marginal policies for multiagent embodied tasks. ArXiv, abs/2007.04979, 2020. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 498, + 288, + 561 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 498, + 288, + 561 + ], + "spans": [ + { + "bbox": [ + 48, + 498, + 288, + 561 + ], + "type": "text", + "content": "[32] Unnat Jain, Luca Weihs, Eric Kolve, Mohammad Rastegari, Svetlana Lazebnik, Ali Farhadi, Alexander G. Schwing, and Aniruddha Kembhavi. Two body problem: Collaborative visual task completion. 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6682-6692, 2019. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 562, + 288, + 605 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 562, + 288, + 605 + ], + "spans": [ + { + "bbox": [ + 48, + 562, + 288, + 605 + ], + "type": "text", + "content": "[33] Oğuzhan Fatih Kar, Teresa Yeo, Andrei Atanov, and Amir Zamir. 3d common corruptions and data augmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18963-18974, 2022. 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 605, + 288, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 605, + 288, + 658 + ], + "spans": [ + { + "bbox": [ + 48, + 605, + 288, + 658 + ], + "type": "text", + "content": "[34] Kapil Katyal, Katie Popek, Chris Paxton, Phil Burlina, and Gregory D. Hager. Uncertainty-aware occupancy map prediction using generative networks for robot navigation. In 2019 International Conference on Robotics and Automation (ICRA), pages 5453-5459, 2019. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 658, + 288, + 701 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 288, + 701 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 288, + 701 + ], + "type": "text", + "content": "[35] Hansung Kim, Luca Remaggi, Philip JB Jackson, Filippo Maria Fazi, and Adrian Hilton. 3d room geometry reconstruction using audio-visual sensors. In 2017 International Conference on 3D Vision (3DV), pages 621–629, 2017. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 701, + 288, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 701, + 288, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 701, + 288, + 712 + ], + "type": "text", + "content": "[36] Bruno Korbar, Du Tran, and Lorenzo Torresani. Scsampler:" + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 706 + ], + "type": "list", + "angle": 0, + "index": 25, + "blocks": [ + { + "bbox": [ + 326, + 73, + 547, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 547, + 106 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 547, + 106 + ], + "type": "text", + "content": "Sampling salient clips from video for efficient action recognition. 2019 IEEE/CVF International Conference on Computer Vision (ICCV), pages 6231-6241, 2019. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 307, + 106, + 547, + 148 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 106, + 547, + 148 + ], + "spans": [ + { + "bbox": [ + 307, + 106, + 547, + 148 + ], + "type": "text", + "content": "[37] Robert LiKamWa, Zhen Wang, Aaron Carroll, Felix Xiaozhu Lin, and Lin Zhong. Draining our glass: An energy and heat characterization of google glass. In Proceedings of 5th Asia-Pacific Workshop on Systems, pages 1-7, 2014. 2, 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 307, + 148, + 547, + 201 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 148, + 547, + 201 + ], + "spans": [ + { + "bbox": [ + 307, + 148, + 547, + 201 + ], + "type": "text", + "content": "[38] Jintao Lin, Haodong Duan, Kai Chen, Dahua Lin, and Limin Wang. Ocsampler: Compressing videos to one clip with single-step sampling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13894-13903, 2022. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 307, + 201, + 547, + 245 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 201, + 547, + 245 + ], + "spans": [ + { + "bbox": [ + 307, + 201, + 547, + 245 + ], + "type": "text", + "content": "[39] David B. Lindell, Gordon Wetzstein, and Vladlen Koltun. Acoustic non-line-of-sight imaging. In 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6773-6782, 2019. 1, 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 307, + 245, + 547, + 288 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 245, + 547, + 288 + ], + "spans": [ + { + "bbox": [ + 307, + 245, + 547, + 288 + ], + "type": "text", + "content": "[40] Chen Liu, Jiaye Wu, and Yasutaka Furukawa. Floornet: A unified framework for floorplan reconstruction from 3d scans. In Proceedings of the European conference on computer vision (ECCV), pages 201-217, 2018. 1, 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 288, + 547, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 288, + 547, + 331 + ], + "spans": [ + { + "bbox": [ + 307, + 288, + 547, + 331 + ], + "type": "text", + "content": "[41] Wenxin Liu, David Caruso, Eddy Ig, Jing Dong, Anastasios I Mourikis, Kostas Daniilidis, Vijay Kumar, and Jakob Engel. Tlio: Tight learned inertial odometry. IEEE Robotics and Automation Letters, 5(4):5653-5660, 2020. 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 331, + 547, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 331, + 547, + 373 + ], + "spans": [ + { + "bbox": [ + 307, + 331, + 547, + 373 + ], + "type": "text", + "content": "[42] Sagnik Majumder, Ziad Al-Halah, and Kristen Grauman. Move2hear: Active audio-visual source separation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 275–285, 2021. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 373, + 547, + 406 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 373, + 547, + 406 + ], + "spans": [ + { + "bbox": [ + 307, + 373, + 547, + 406 + ], + "type": "text", + "content": "[43] Sagnik Majumder, Ziad Al-Halah, and Kristen Grauman. Active audio-visual separation of dynamic sound sources. In European Conference on Computer Vision. Springer, 2022. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 406, + 547, + 449 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 406, + 547, + 449 + ], + "spans": [ + { + "bbox": [ + 307, + 406, + 547, + 449 + ], + "type": "text", + "content": "[44] Sagnik Majumder, Changan Chen, Ziad Al-Halah, and Kristen Grauman. Few-shot audio-visual learning of environment acoustics. In Thirty-Sixth Conference on Neural Information Processing Systems, 2022. 2, 5" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 449, + 547, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 449, + 547, + 502 + ], + "spans": [ + { + "bbox": [ + 307, + 449, + 547, + 502 + ], + "type": "text", + "content": "[45] Medhini Narasimhan, Erik Wijmans, Xinlei Chen, Trevor Darrell, Dhruv Batra, Devi Parikh, and Amanpreet Singh. Seeing the un-scene: Learning amodal semantic maps for room navigation. In European Conference on Computer Vision, pages 513-529. Springer, 2020. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 502, + 547, + 524 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 502, + 547, + 524 + ], + "spans": [ + { + "bbox": [ + 307, + 502, + 547, + 524 + ], + "type": "text", + "content": "[46] Simon T O'Callaghan and Fabio T Ramos. Gaussian process occupancy maps*. Int. J. Rob. Res., 31(1):42-62, jan 2012. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 524, + 547, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 524, + 547, + 544 + ], + "spans": [ + { + "bbox": [ + 307, + 524, + 547, + 544 + ], + "type": "text", + "content": "[47] Brian Okorn, Xuehan Xiong, and Burcu Akinci. Toward automated modeling of floor plans. In 3D PVT, 2009. 1, 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 544, + 547, + 597 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 544, + 547, + 597 + ], + "spans": [ + { + "bbox": [ + 307, + 544, + 547, + 597 + ], + "type": "text", + "content": "[48] Andrew Owens, Phillip Isola, Josh McDermott, Antonio Torralba, Edward H. Adelson, and William T. Freeman. Visually indicated sounds. In 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 2405-2413, 2016. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 597, + 547, + 651 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 597, + 547, + 651 + ], + "spans": [ + { + "bbox": [ + 307, + 597, + 547, + 651 + ], + "type": "text", + "content": "[49] Vassil Panayotov, Guoguo Chen, Daniel Povey, and Sanjeev Khudanpur. Librispeech: An asr corpus based on public domain audio books. In 2015 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 5206-5210, 2015. 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 651, + 547, + 706 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 651, + 547, + 706 + ], + "spans": [ + { + "bbox": [ + 307, + 651, + 547, + 706 + ], + "type": "text", + "content": "[50] Rameswar Panda, Chun-Fu Richard Chen, Quanfu Fan, Ximeng Sun, Kate Saenko, Aude Oliva, and Rogerio Feris. Adamml: Adaptive multi-modal learning for efficient video recognition. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7576-7585, 2021. 3" + } + ] + } + ], + "index": 24 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "10563" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 708 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 137 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 288, + 137 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 288, + 137 + ], + "type": "text", + "content": "[51] Shivansh Patel, Saim Wani, Unnat Jain, Alexander G. Schwing, Svetlana Lazebnik, Manolis Savva, and Angel X. Chang. Interpretation of emergent communication in heterogeneous collaborative embodied agents. 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 15993-15943, 2021. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 138, + 288, + 202 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 138, + 288, + 202 + ], + "spans": [ + { + "bbox": [ + 48, + 138, + 288, + 202 + ], + "type": "text", + "content": "[52] Senthil Purushwalkam, Sebastia Vicenc Amengual Gari, Vamsi Krishna Ithapu, Carl Schissler, Philip Robinson, Abhinav Gupta, and Kristen Grauman. Audio-visual floorplan reconstruction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1183–1192, 2021. 1, 2, 3, 6, 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 203, + 288, + 246 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 203, + 288, + 246 + ], + "spans": [ + { + "bbox": [ + 48, + 203, + 288, + 246 + ], + "type": "text", + "content": "[53] Santhosh K Ramakrishnan, Ziad Al-Halah, and Kristen Grauman. Occupancy anticipation for efficient exploration and navigation. In European Conference on Computer Vision, pages 400-418. Springer, 2020. 1, 2, 6, 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 246, + 288, + 288 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 246, + 288, + 288 + ], + "spans": [ + { + "bbox": [ + 48, + 246, + 288, + 288 + ], + "type": "text", + "content": "[54] Fabio Ramos and Lionel Ott. Hilbert maps: Scalable continuous occupancy mapping with stochastic gradient descent. The International Journal of Robotics Research, 35(14):1717-1730, 2016. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 289, + 288, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 289, + 288, + 331 + ], + "spans": [ + { + "bbox": [ + 48, + 289, + 288, + 331 + ], + "type": "text", + "content": "[55] Ethan Rublee, Vincent Rabaud, Kurt Konolige, and Gary Bradski. Orb: An efficient alternative to sift or surf. In 2011 International Conference on Computer Vision, pages 2564-2571, 2011. 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 331, + 288, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 331, + 288, + 374 + ], + "spans": [ + { + "bbox": [ + 48, + 331, + 288, + 374 + ], + "type": "text", + "content": "[56] João Machado Santos, David Portugal, and Rui P. Rocha. An evaluation of 2d slam techniques available in robot operating system. In 2013 IEEE International Symposium on Safety, Security, and Rescue Robotics (SSRR), pages 1-6, 2013. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 374, + 288, + 438 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 374, + 288, + 438 + ], + "spans": [ + { + "bbox": [ + 48, + 374, + 288, + 438 + ], + "type": "text", + "content": "[57] Manolis Savva, Abhishek Kadian, Oleksandr Maksymets, Yili Zhao, Erik Wijmans, Bhavana Jain, Julian Straub, Jia Liu, Vladlen Koltun, Jitendra Malik, et al. Habitat: A platform for embodied ai research. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9339-9347, 2019. 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 439, + 288, + 482 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 439, + 288, + 482 + ], + "spans": [ + { + "bbox": [ + 48, + 439, + 288, + 482 + ], + "type": "text", + "content": "[58] Carl Schissler, Christian Loftin, and Dinesh Manocha. Acoustic classification and optimization for multi-modal rendering of real-world scenes. IEEE Transactions on Visualization and Computer Graphics, 24:1246-1259, 2018. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 482, + 288, + 524 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 482, + 288, + 524 + ], + "spans": [ + { + "bbox": [ + 48, + 482, + 288, + 524 + ], + "type": "text", + "content": "[59] Ransalu Senanayake, Thushan Ganegedara, and Fabio Ramos. Deep occupancy maps: a continuous mapping technique for dynamic environments. In NIPS 2017 Workshop MLITS, 2017." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 525, + 288, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 525, + 288, + 578 + ], + "spans": [ + { + "bbox": [ + 48, + 525, + 288, + 578 + ], + "type": "text", + "content": "[60] Rakesh Shrestha, Fei-Peng Tian, Wei Feng, Ping Tan, and Richard Vaughan. Learned map prediction for enhanced mobile robot exploration. In 2019 International Conference on Robotics and Automation (ICRA), pages 1197-1204, 2019. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 578, + 288, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 578, + 288, + 632 + ], + "spans": [ + { + "bbox": [ + 48, + 578, + 288, + 632 + ], + "type": "text", + "content": "[61] Jascha Sohl-Dickstein, Santani Teng, Benjamin M. Gaub, Chris C. Rodgers, Crystal Li, Michael R. DeWeese, and Nicol S. Harper. A device for human ultrasonic echolocation. IEEE Transactions on Biomedical Engineering, 62(6):1526-1534, 2015. 1, 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 632, + 288, + 685 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 632, + 288, + 685 + ], + "spans": [ + { + "bbox": [ + 48, + 632, + 288, + 685 + ], + "type": "text", + "content": "[62] Wei Sui, Lingfeng Wang, Bin Fan, Hongfei Xiao, Huaiyu Wu, and Chunhong Pan. Layer-wise floorplan extraction for automatic urban building reconstruction. IEEE Transactions on Visualization and Computer Graphics, 22(3):1261-1277, 2016. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 686, + 288, + 708 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 686, + 288, + 708 + ], + "spans": [ + { + "bbox": [ + 48, + 686, + 288, + 708 + ], + "type": "text", + "content": "[63] Maitreya Suin and A. N. Rajagopalan. An efficient framework for dense video captioning. Proceedings of the AAAI" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 547, + 696 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 326, + 73, + 547, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 547, + 94 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 547, + 94 + ], + "type": "text", + "content": "Conference on Artificial Intelligence, 34(07):12039-12046, Apr. 2020. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 308, + 95, + 547, + 148 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 95, + 547, + 148 + ], + "spans": [ + { + "bbox": [ + 308, + 95, + 547, + 148 + ], + "type": "text", + "content": "[64] Cheng Sun, Chi-Wei Hsiao, Min Sun, and Hwann-Tzong Chen. Horizonnet: Learning room layout with 1d representation and piano stretch data augmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1047–1056, 2019. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 148, + 547, + 191 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 148, + 547, + 191 + ], + "spans": [ + { + "bbox": [ + 308, + 148, + 547, + 191 + ], + "type": "text", + "content": "[65] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 5" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 191, + 547, + 233 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 191, + 547, + 233 + ], + "spans": [ + { + "bbox": [ + 308, + 191, + 547, + 233 + ], + "type": "text", + "content": "[66] Erik Wijmans, Abhishek Kadian, Ari S. Morcos, Stefan Lee, Irfan Essa, Devi Parikh, Manolis Savva, and Dhruv Batra. Dd-ppo: Learning near-perfect pointgoal navigators from 2.5 billion frames. In ICLR, 2020. 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 233, + 547, + 276 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 233, + 547, + 276 + ], + "spans": [ + { + "bbox": [ + 308, + 233, + 547, + 276 + ], + "type": "text", + "content": "[67] Wenming Wu, Xiao-Ming Fu, Rui Tang, Yuhan Wang, YuHao Qi, and Ligang Liu. Data-driven interior plan generation for residential buildings. ACM Trans. Graph., 38(6), nov 2019. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 276, + 547, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 276, + 547, + 331 + ], + "spans": [ + { + "bbox": [ + 308, + 276, + 547, + 331 + ], + "type": "text", + "content": "[68] Zuxuan Wu, Caiming Xiong, Chih-Yao Ma, Richard Socher, and Larry S Davis. Adaframe: Adaptive frame selection for fast video recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1278-1287, 2019. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 331, + 547, + 395 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 331, + 547, + 395 + ], + "spans": [ + { + "bbox": [ + 308, + 331, + 547, + 395 + ], + "type": "text", + "content": "[69] Shang-Ta Yang, Fu-En Wang, Chi-Han Peng, Peter Wonka, Min Sun, and Hung-Kuo Chu. Dula-net: A dual-projection network for estimating room layouts from a single rgb panorama. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3363–3372, 2019. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 395, + 547, + 438 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 395, + 547, + 438 + ], + "spans": [ + { + "bbox": [ + 308, + 395, + 547, + 438 + ], + "type": "text", + "content": "[70] Mao Ye, Yu Zhang, Ruigang Yang, and Dinesh Manocha. 3d reconstruction in the presence of glasses by acoustic and stereo fusion. In 2015 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 4885-4893, 2015. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 438, + 547, + 481 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 438, + 547, + 481 + ], + "spans": [ + { + "bbox": [ + 308, + 438, + 547, + 481 + ], + "type": "text", + "content": "[71] Serena Yeung, Olga Russakovsky, Greg Mori, and Li Fei-Fei. End-to-end learning of action detection from frame glimpses in videos. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2678–2687, 2016. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 481, + 547, + 524 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 481, + 547, + 524 + ], + "spans": [ + { + "bbox": [ + 308, + 481, + 547, + 524 + ], + "type": "text", + "content": "[72] Yinfeng Yu, Wenbing Huang, Fuchun Sun, Changan Chen, Yikai Wang, and Xiaohong Liu. Sound adversarial audiovisual navigation. In International Conference on Learning Representations, 2022. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 524, + 547, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 524, + 547, + 578 + ], + "spans": [ + { + "bbox": [ + 308, + 524, + 547, + 578 + ], + "type": "text", + "content": "[73] Zhoutong Zhang, Jiajun Wu, Qiujia Li, Zhengjia Huang, James Traer, Josh H. McDermott, Joshua B. Tenenbaum, and William T. Freeman. Generative modeling of audible shapes for object perception. 2017 IEEE International Conference on Computer Vision (ICCV), pages 1260-1269, 2017. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 578, + 547, + 630 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 578, + 547, + 630 + ], + "spans": [ + { + "bbox": [ + 308, + 578, + 547, + 630 + ], + "type": "text", + "content": "[74] Chuhang Zou, Alex Colburn, Qi Shan, and Derek Hoiem. Layoutnet: Reconstructing the 3d room layout from a single rgb image. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2051-2059, 2018. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 631, + 547, + 696 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 631, + 547, + 696 + ], + "spans": [ + { + "bbox": [ + 308, + 631, + 547, + 696 + ], + "type": "text", + "content": "[75] Chuhang Zou, Jheng-Wei Su, Chi-Han Peng, Alex Colburn, Qi Shan, Peter Wonka, Hung-Kuo Chu, and Derek Hoiem. Manhattan room layout reconstruction from a single " + }, + { + "bbox": [ + 308, + 631, + 547, + 696 + ], + "type": "inline_equation", + "content": "360^{\\circ}" + }, + { + "bbox": [ + 308, + 631, + 547, + 696 + ], + "type": "text", + "content": " image: A comparative study of state-of-the-art methods. International Journal of Computer Vision, 129(5):1410-1431, 2021. 2" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "10564" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/CiCo_ Domain-Aware Sign Language Retrieval via Cross-Lingual Contrastive Learning/047059b8-c0b2-4bcc-9546-ab9b6e4bc0e5_content_list.json b/2023/CiCo_ Domain-Aware Sign Language Retrieval via Cross-Lingual Contrastive Learning/047059b8-c0b2-4bcc-9546-ab9b6e4bc0e5_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..f1411587c52ad9d47a1003801bd78ffae77c1ffe --- /dev/null +++ b/2023/CiCo_ Domain-Aware Sign Language Retrieval via Cross-Lingual Contrastive Learning/047059b8-c0b2-4bcc-9546-ab9b6e4bc0e5_content_list.json @@ -0,0 +1,1494 @@ +[ + { + "type": "text", + "text": "CiCo: Domain-Aware Sign Language Retrieval via Cross-Lingual Contrastive Learning", + "text_level": 1, + "bbox": [ + 228, + 130, + 743, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yiting Cheng $^{1}$ , Fangyun Wei $^{2*}$ , Jianmin Bao $^{2}$ , Dong Chen $^{2}$ , Wenqiang Zhang $^{1*}$ $^{1}$ School of Computer Science, Fudan University, $^{2}$ Microsoft Research Asia \n{ytcheng18, wqzhang}@fudan.edu.cn, {fawe, jianbao, doch}@microsoft.com", + "bbox": [ + 150, + 202, + 825, + 257 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 291, + 313, + 306 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "This work focuses on sign language retrieval—a recently proposed task for sign language understanding. Sign language retrieval consists of two sub-tasks: text-to-sign-video (T2V) retrieval and sign-video-to-text (V2T) retrieval. Different from traditional video-text retrieval, sign language videos, not only contain visual signals but also carry abundant semantic meanings by themselves due to the fact that sign languages are also natural languages. Considering this character, we formulate sign language retrieval as a cross-lingual retrieval problem as well as a video-text retrieval task. Concretely, we take into account the linguistic properties of both sign languages and natural languages, and simultaneously identify the fine-grained cross-lingual (i.e., sign-to-word) mappings while contrasting the texts and the sign videos in a joint embedding space. This process is termed as cross-lingual contrastive learning. Another challenge is raised by the data scarcity issue—sign language datasets are orders of magnitude smaller in scale than that of speech recognition. We alleviate this issue by adopting a domain-agnostic sign encoder pre-trained on large-scale sign videos into the target domain via pseudolabeling. Our framework, termed as domain-aware sign language retrieval via Cross-lingual Contrastive learning or CiCo for short, outperforms the pioneering method by large margins on various datasets, e.g., +22.4 T2V and +28.0 V2TR@1 improvements on How2Sign dataset, and +13.7 T2V and +17.1 V2TR@1 improvements on PHOENIX-2014T dataset. Code and models are available at: https://github.com/FangyunWei/SLRT.", + "bbox": [ + 75, + 323, + 473, + 763 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 790, + 209, + 806 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Sign languages are the primary means of communication used by people who are deaf or hard of hearing. Sign language understanding [1, 10, 12-15, 18, 32, 33, 61, 73] is significant for overcoming the communication barrier between", + "bbox": [ + 76, + 816, + 468, + 878 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/01d2bcf4d4965d58c2397bc1dcf667b59d7f50fa64b981ec7ac2f54c7e5724a0.jpg", + "image_caption": [ + "(a) Text-to-sign-video (T2V) retrieval." + ], + "image_footnote": [], + "bbox": [ + 519, + 292, + 874, + 373 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/ed8f51b1773276f9834cc368dd8f8e36453c53712644d3230dadfc7314559f1a.jpg", + "image_caption": [ + "(b) Sign-video-to-text (V2T) retrieval.", + "Figure 1. Illustration of: (a) T2V retrieval; (b) V2T retrieval." + ], + "image_footnote": [], + "bbox": [ + 517, + 392, + 870, + 469 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "the hard-of-hearing and non-signers. Sign language recognition and translation (SLRT) has been extensively studied, with the goal of recognizing the arbitrary semantic meanings conveyed by sign languages. However, the lack of available data significantly limits the capability of SLRT. In this paper, we focus on developing a framework for a recently proposed sign language retrieval task [18]. Unlike SLRT, sign language retrieval focuses on retrieving the meanings that signers express from a closed-set, which can significantly reduce error rates in realistic deployment.", + "bbox": [ + 496, + 529, + 893, + 681 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Sign language retrieval is both similar to and distinct from the traditional video-text retrieval. On the one hand, like video-text retrieval, sign language retrieval is also composed of two sub-tasks, i.e., text-to-sign-video (T2V) retrieval and sign-video-to-text (V2T) retrieval. Given a free-form written query and a large collection of sign language videos, the objective of T2V is to find the video that best matches the written query (Figure 1a). In contrast, the goal of V2T is to identify the most relevant text description given a query of sign language video (Figure 1b).", + "bbox": [ + 496, + 684, + 893, + 837 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "On the other hand, different from the video-text retrieval, sign languages, like most natural languages, have their own grammars and linguistic properties. Therefore, sign language videos not only contain visual signals, but also carry", + "bbox": [ + 496, + 839, + 893, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author.", + "bbox": [ + 94, + 886, + 220, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "19016", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/2351c5086303c77f8edced1b80e17be1ae97267ee1ade4ecd4ccbd38d0dd1498.jpg", + "image_caption": [ + "(Sign Language)" + ], + "image_footnote": [ + "(a) While contrasting the sign videos and the texts in a joint embedding space, we simultaneously identify the fine-grained cross-lingual (sign-to-word) mappings of sign languages and natural languages via the proposed cross-lingual contrastive learning. Existing datasets do not annotate the sign-to-word mappings." + ], + "bbox": [ + 94, + 111, + 447, + 265 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/9ee6e99e7787c128052cf6936f6c4c66facc18ea1a046dff0fad964a778d1cf0.jpg", + "image_caption": [ + "Figure 2. Illustration of: (a) cross-lingual (sign-to-word) mapping; (b) sign-to-word mappings identified by our CiCo." + ], + "image_footnote": [ + "(b) We show four instances of the sign \"Book\" in How2Sign [19] dataset, which are identified by our approach. Please refer to the supplementary material for more examples." + ], + "bbox": [ + 94, + 340, + 269, + 439 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/fc11fdd1b44df7cb3628ab8f99db8c3b472c52356ddeae90c99d865bd42110a7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 272, + 340, + 447, + 439 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "semantics (i.e., sign1-to-word mappings between sign languages and natural languages) by themselves, which differentiates them from the general videos that merely contain visual information. Considering the linguistic characteristics of sign languages, we formulate sign language retrieval as a cross-lingual retrieval [6,34,59] problem in addition to a video-text retrieval [5,24,42-44,57,68] task.", + "bbox": [ + 75, + 531, + 468, + 636 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Sign language retrieval is extremely challenging due to the following reasons: (1) Sign languages are completely separate and distinct from natural languages since they have unique linguistic rules, word formation, and word order. The transcription between sign languages and natural languages is complicated, for instance, the word order is typically not preserved between sign languages and natural languages. It is necessary to automatically identify the sign-to-word mapping from the cross-lingual retrieval perspective; (2) In contrast to the text-video retrieval datasets [45, 48] which contain millions of training samples, sign language datasets are orders of magnitude smaller in scale—for example, there are only $30\\mathrm{K}$ video-text pairs in How2Sign [19] training set; (3) Sign languages convey information through the handshape, facial expression, and body movement, which requires models to distinguish fine", + "bbox": [ + 75, + 637, + 470, + 878 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "grained gestures and actions; (4) Sign language videos typically contain hundreds of frames. It is necessary to build efficient algorithms to lower the training cost and fit the long videos as well as the intermediate representations into limited GPU memory.", + "bbox": [ + 500, + 90, + 890, + 166 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this work, we concentrate on resolving the challenges listed above:", + "bbox": [ + 500, + 167, + 890, + 196 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We consider the linguistic rules (e.g., word order) of both sign languages and natural languages. We formulate sign language retrieval as a cross-lingual retrieval task as well as a video-text retrieval problem. While contrasting the sign videos and the texts in a joint embedding space as achieved in most vision-language pre-training frameworks [5, 44, 57], we simultaneously identify the fine-grained cross-lingual (sign-to-word) mappings between two types of languages via our proposed cross-lingual contrastive learning as shown in Figure 2.", + "- Data scarcity typically brings in the over-fitting issue. To alleviate this issue, we adopt transfer learning and adapt a recently released domain-agnostic sign encoder [61] pre-trained on large-scale sign-videos to the target domain. Although this encoder is capable of distinguishing the fine-grained signs, direct transferring may be suboptimal due to the unavoidable domain gap between the pre-training dataset and sign language retrieval datasets. To tackle this problem, we further fine-tune a domain-aware sign encoder on pseudo-labeled data from target datasets. The final sign encoder is composed of the well-optimized domain-aware sign encoder and the powerful domain-agnostic sign encoder.", + "- In order to effectively model long videos, we decouple our framework into two disjoint parts: (1) a sign encoder which adopts a sliding window on sign-videos to pre-extract their vision features; (2) a cross-lingual contrastive learning module which encodes the extracted vision features and their corresponding texts in a joint embedding space." + ], + "bbox": [ + 500, + 203, + 890, + 675 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our framework, called domain-aware sign language retrieval via Cross-lingual Contrastive learning or CiCo for short, outperforms the pioneer SPOT-ALIGN [18] by large margins on various datasets, achieving 56.6 (+22.4) T2V and 51.6 (+28.0) V2T R@1 accuracy (improvement) on How2Sign [19] dataset, and 69.5 (+13.7) T2V and 70.2 (+17.1) V2T R@1 accuracy (improvement) on PHOENIX-2014T [8] dataset. With its simplicity and strong performance, we hope our approach can serve as a solid baseline for future research.", + "bbox": [ + 496, + 681, + 890, + 832 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 844, + 640, + 859 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Sign Language Understanding. Sign language understanding aims at interpreting the semantic information con", + "bbox": [ + 500, + 869, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_footnote", + "text": "1We use sign to denote lexical item within a sign language vocabulary.", + "bbox": [ + 94, + 886, + 467, + 901 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "19017", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "veyed within sign videos. Researchers have explored such capability on various tasks including sign language recognition [13,20,21,30,38,58,73], sign spotting [1,18,47,61], sign language translation [8,9,12,37,72] and our focused sign language retrieval [18].", + "bbox": [ + 75, + 90, + 470, + 167 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "One of the fundamental tasks of sign language understanding is sign language recognition (SLR), which aims to transcribe a sign video into a gloss sequence. Previous works on SLR focus on designing carefully engineered features [20, 21, 58] or modeling temporal dependencies [55, 56]. Recently, the success of 3D convolutional neural networks in action related tasks [41, 60, 63] is transferred to SLR. In particular, the I3D [11] architecture has proven to be effective for this task [1, 30, 36, 38, 61]. In this work, we also adopt this network architecture in our sign encoder.", + "bbox": [ + 75, + 169, + 468, + 335 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Sign spotting is a particular variant of sign language recognition. It aims to localize all instances of a given sign within an untrimmed video. Recent works tackle this task with auxiliary cue of subtitles, introducing automatic annotation systems by using mouthing [1], dictionaries [47] and attention maps of Transformer [61]. SPOT-ALIGN [18] extends existing spotting methods [1, 47] with an iterative training schema, which alternates between repeated sign spotting and model fine-tuning. In this work, pseudo-labeling is served as our sign spotting approach to localize isolated signs in untrimmed videos from target sign language retrieval datasets. Compared with above efforts, our approach only employs a pre-trained sign encoder without utilizing additional auxiliary cue, which is proved to be simple yet efficient.", + "bbox": [ + 75, + 337, + 470, + 561 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Early works of sign language retrieval primarily investigate query-by-example searching [4, 70], which queries individual instances with given sign examples. Our work focuses on free-form textual retrieval—a recently introduced task by SPOT-ALIGN [18]. It symbolizes the real-world scenario of searching sign language videos with natural languages. The pioneer SPOT-ALIGN [18] purely formulates sign language retrieval as a video-text retrieval task, where the cross-modal alignment is modeled upon overall global embeddings of sign videos and texts. However, the linguistic properties of sign languages are ignored. In contrast, we formulate sign language retrieval as a joint task of text-video retrieval and cross-lingual retrieval.", + "bbox": [ + 75, + 566, + 468, + 762 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Vision-Language Models. Learning general-purpose representations for visual and textual modalities is a longstanding topic [7, 35]. The idea has been investigated decades ago [49]. Recently, the prominent success of CLIP [51], ALIGN [28] and ALBEF [39] has demonstrated the capability of learning joint cross-modal representations with large-scale web data using simple image-text contrastive learning. Similar idea has also been explored in video-text retrieval, which is the closest area of our work.", + "bbox": [ + 75, + 763, + 470, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The common practice in image-text retrieval [23, 25, 29, 51] and video-text retrieval [5, 24, 42-44, 57, 68] is to encode the images/videos and texts into the overall representations. It is reasonable since the visual signals of typical image-text and video-text retrieval datasets mainly describe the certain objects or describable events. In contrast, sign videos, as the carriers of sign languages, convey abundant semantics by themselves. There exist fine-grained mappings between sign videos and natural languages. We find that identifying such cross-lingual (sign-to-word) mappings significantly boosts the performance of sign language retrieval.", + "bbox": [ + 496, + 90, + 890, + 257 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Cross-Lingual Information Retrieval. In this work, we also formulate sign-language retrieval as a cross-lingual task. In natural language processing community, cross-lingual information retrieval (CLIR) [34, 59, 65] refers to the task of retrieving documents between different languages. One of the most common approaches of CLIR is to learn sentence-level embedding alignment by mapping pre-acquired monolingual embeddings [46, 50, 69] of different languages into a shared space. Recently, researchers have exploited fine-grained word-level mappings with self-training [2, 3, 59]. They find that mappings can be learned by initiating with a seed dictionary and alternating between alignment modeling and dictionary mining. Our work also exploits fine-grained word-level mappings, called sign-to-word mappings, in the context of sign language retrieval. Identifying sign-to-word mappings is challenging since sign languages are expressed in visual modality. The proposed cross-lingual contrastive learning tackles this challenge by exploiting the fine-grained cross-modal interactions.", + "bbox": [ + 496, + 258, + 890, + 544 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Methodology", + "text_level": 1, + "bbox": [ + 500, + 556, + 633, + 573 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we first formulate the task of sign language retrieval in Section 3.1. Next, we introduce our framework which is composed of two parts: 1) a sign encoder which extracts discriminative and domain-aligned features of sign videos (Section 3.2); 2) a cross-lingual contrastive learning framework, which contrasts sign-video-text pairs while concurrently identifying the fine-grained sign-to-word mappings (Section 3.3). At last, we explore text augmentations in Section 3.4.", + "bbox": [ + 496, + 582, + 890, + 717 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Task Formulation", + "text_level": 1, + "bbox": [ + 500, + 726, + 671, + 739 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Let $\\mathcal{V}$ and $\\mathcal{T}$ denote a set of sign videos and their corresponding texts (transcriptions), respectively. Sign language retrieval consists of two tasks namely text-to-sign-video retrieval (T2V), and sign-video-to-text retrieval (V2T), respectively. The objective of T2V is to find the sign video $v\\in \\mathcal{V}$ whose signing content best matches the text query. In contrast, the reverse task V2T requires the model to identify the most relevant text (transcription) $t\\in \\mathcal{T}$ given a query of sign video. We resolve sign language retrieval by learning a joint embedding space of sign videos and texts.", + "bbox": [ + 496, + 750, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "19018", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/b9c7644175c01d5251b0ebf06b002ac5650564f65aa85bae99313a5ceeff1e0e.jpg", + "image_caption": [ + "(a) Illustration of sign encoder." + ], + "image_footnote": [], + "bbox": [ + 99, + 89, + 869, + 304 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/177a9fdd10ab5bc16f025d961267330e3e5b63ba7832acf7ea46d6dcf0990232.jpg", + "image_caption": [ + "(b) Illustration of cross-lingual contrastive learning.", + "Figure 3. Overview of our framework. (a) Sign encoder is composed of a powerful domain-agnostic sign encoder pre-trained on large-scale sign videos, and a domain-aware sign encoder fine-tuned on pseudo-labeled data from target datasets. We adopt a sliding window manner to extract a discriminative and domain-aligned feature per clip. (b) Cross-lingual contrastive learning takes $N$ sign-video-text pairs as inputs and contrasts paired data in a shared embedding space while implicitly identifying the fine-grained sign-to-word mappings during training." + ], + "image_footnote": [], + "bbox": [ + 101, + 327, + 869, + 573 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Sign Encoder", + "text_level": 1, + "bbox": [ + 76, + 681, + 217, + 698 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Process Sign Videos with Sliding Window. Sign videos from sign language retrieval datasets typically contain hundreds of frames. To efficiently train our model and lower the usage of GPU memory, given a sign video, we adopt a sliding window manner with stride of 1 and window size of 16 to produce $M$ temporally overlapped clips. Next, we separately feed each clip into a sign encoder to extract its feature. The final sign-video feature is yielded by stacking features coming out of $M$ clips along temporal dimension. A powerful sign encoder is crucial.", + "bbox": [ + 75, + 710, + 470, + 863 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Overview of Sign Encoder. Recent advances in sign spotting [47, 61] greatly facilitate the collection of large-scale", + "bbox": [ + 75, + 869, + 470, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "sign language datasets, enabling powerful representation learning abilities of convolutional neural networks on the sign classification task. Previous methods [16, 18] have demonstrated the feasibility of transferring a sign encoder pre-trained on large-scale sign-spotting data into downstream tasks. We follow this practice and use an I3D network pre-trained on BSL-1K [61], a sign classification dataset collected via sign spotting, as our primary sign encoder. Due to its favorable transfer performance, we term this model as a domain-agnostic sign encoder. Nevertheless, the domain gap between BSL-1K and sign language retrieval datasets is non-negligible. To tackle this problem, we further fine-tune a domain-aware sign encoder, which has an identical architecture to the domain-agnostic sign en", + "bbox": [ + 496, + 681, + 890, + 893 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "19019", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "coder, on target datasets through pseudo-labeling. The final sign encoder is composed of the well-optimized domain-aware sign encoder and the powerful domain-agnostic sign encoder, as illustrated in Figure 3a.", + "bbox": [ + 75, + 90, + 468, + 151 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Pseudo-Labeling on Target Datasets. Now we describe the details of pseudo-labeling. Given a sign video from a target dataset, we adopt a sliding window with the stride of 1 and the window size of 16 to generate a set of temporally overlapped clips. For each clip, we first utilize the pretrained domain-agnostic sign encoder to produce its prediction. Then we binarize the prediction with a pre-defined threshold $\\lambda$ to generate the corresponding pseudo label. The invalid samples, whose maximum score is lower than $\\lambda$ , are filtered. We repeat the above process for all sign videos and eventually build a pseudo-labeled set. Our domain-aware sign encoder, which is initialized by the domain-agnostic sign encoder, is fine-tuned on the pseudo-labeled set via a standard cross-entropy loss for classification training.", + "bbox": [ + 75, + 152, + 470, + 364 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Feature Extraction with Sign Encoder. So far, we acquire a domain-aware sign encoder approximately aligned in target domain. Nevertheless, its capability is restricted by the unavoidable noises in pseudo-labels and the limited amount of pseudo-labeled samples. Recall that we already have a powerful domain-agnostic sign encoder pre-trained on large-scale dataset in hand, inspiring us to make use of both domain-agnostic sign encoder $h_{\\xi}(\\cdot)$ and domain-aware sign encoder $h_{\\theta}(\\cdot)$ to extract discriminative and domain-aligned features. Our final sign encoder $H(\\cdot)$ , as shown in Figure 3a, is a weighted combination of $h_{\\xi}(\\cdot)$ and $h_{\\theta}(\\cdot)$ with a trade-off hyper-parameter $\\alpha$ . As described above, $H(\\cdot)$ encodes sign videos in a sliding window manner. For simplicity, we use $H(v)$ to denote feature extraction on sign video $v$ , which is formulated as:", + "bbox": [ + 75, + 364, + 470, + 590 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nH (v) = \\alpha h _ {\\xi} (v) + (1 - \\alpha) h _ {\\theta} (v). \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 160, + 604, + 468, + 621 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Cross-Linguual Contrastive Learning", + "text_level": 1, + "bbox": [ + 76, + 633, + 393, + 650 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The objective of cross-lingual contrastive learning (CLCL) is to learn a joint embedding space of sign videos and texts while concurrently identifying the fine-grained sign-to-word mappings during training. An overview is shown in Figure 3b. CLCL takes a mini-batch $\\{(v_{n},t_{n})\\}_{n = 1}^{N}$ containing $N$ sign-video-text pairs as input, and contrasts paired data in a shared embedding space for sign language retrieval.", + "bbox": [ + 75, + 657, + 468, + 777 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Sign Features and Word Features. Given a sign video $v \\in \\{v_n\\}_{n=1}^N$ , we first adopt our sign encoder $H(\\cdot)$ described in Section 3.2 to extract its intermediate feature. Note $H(\\cdot)$ encodes sign videos in a sliding window manner, and thus there are no interactions among different clips. To facilitate information exchange, we further append a 12-layer Transformer [62] $F(\\cdot)$ onto $H(\\cdot)$ to extract sign features $S$ of sign video $v$ , which is formulated", + "bbox": [ + 75, + 779, + 468, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "as $S = F(H(v)) \\in \\mathbb{R}^{M \\times D}$ , where $M$ denotes the number of clips, and $D$ is the hidden dimension. Given a text $t \\in \\{t_n\\}_{n=1}^N$ , we convert $t$ into a lower-cased byte pair encoding (BPE) representation [53], which is subsequently fed into another 12-layer Transformer $G(\\cdot)$ to generate the word features $W = G(t) \\in \\mathbb{R}^{L \\times D}$ , where $L$ represents word number.", + "bbox": [ + 496, + 90, + 890, + 196 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Since CLIP [51] shows excellent transfer capability in various downstream tasks [17, 22, 27, 54, 66, 67], we initialize $F(\\cdot)$ and $G(\\cdot)$ with CLIP's image encoder (ViT-B) and text encoder, respectively, to ease the learning. Though CLIP's vision encoder takes image patches as inputs, we experimentally find that it generalizes well in our scenario where input data is in a different modality.", + "bbox": [ + 496, + 199, + 890, + 306 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Cross-Lingual Similarity. There exist inherent sign-to-word mappings between sign languages and natural languages. To incorporate this prior knowledge into learning, we introduce cross-lingual similarity—an indicator to identify sign-to-word mappings between $i$ -th sign video $v_{i}$ and $j$ -th text $t_j$ . Concretely, given sign features $\\pmb{S}_i \\in \\mathbb{R}^{M \\times D}$ of $v_i$ , and word features $\\pmb{W}_j \\in \\mathbb{R}^{L \\times D}$ of $t_j$ , we calculate a cross-lingual similarity matrix $\\pmb{E}_{(i,j)} = \\pmb{S}_i \\cdot \\pmb{W}_j^T \\in \\mathbb{R}^{M \\times L}$ . Each element in $\\pmb{E}_{(i,j)}$ represents the similarity of a sign clip in $v_i$ and a word in $t_j$ .", + "bbox": [ + 496, + 309, + 890, + 462 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Cross-Lingual Contrastive Learning. Directly apply supervisions on token-wise similarity matrix $\\pmb{E}_{(i,j)}$ is infeasible due to the absence of fine-grained sign-to-word annotations. Inspired by the recent progress of vision-language contrastive learning [23, 25, 29, 39, 40, 51], we turn to contrast the global representations of sign videos and texts. The underlying idea is to calculate a global similarity $z$ of $v_{i}$ and $t_j$ based on $\\pmb{E}_{(i,j)} \\in \\mathbb{R}^{M \\times L}$ .", + "bbox": [ + 496, + 464, + 890, + 587 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Sign-Video-to-Text Contrast. We first introduce sign-video-to-text contrast as shown in Figure 3b. To be specific, we utilize a Softmax operation to each row of $\\pmb{E}_{(i,j)}$ and multiply the resulting matrix with $\\pmb{E}_{(i,j)}$ to generate a sign-to-word similarity matrix $\\pmb{E}_{(i,j)}' \\in \\mathbb{R}^{M \\times L}$ , where each row represents the re-weighted similarities between a sign clip in $v_{i}$ and all words in $t_j$ . After that, we adopt a row-wise addition operation on $\\pmb{E}_{(i,j)}'$ to yield the sign-to-text similarity vector $e_{(i,j)} \\in \\mathbb{R}^M$ , where each element denotes a similarity of a sign clip in $v_{i}$ and whole text $t_j$ . At last, we average all elements in $e_{(i,j)}$ to produce the global similarity $z$ of sign video $v_{i}$ and text $t_j$ .", + "bbox": [ + 496, + 588, + 890, + 772 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In the same way, we can calculate the similarities for both positive pairs $\\{(v_i,t_i)\\}_{i = 1}^N$ and negative pairs $\\{(v_i,t_j)\\}_{i = 1,j = 1,i\\neq j}^N$ in a mini-batch, yielding a video-to-text similarity matrix $Z_{V2T}\\in \\mathbb{R}^{N\\times N}$ , where $Z_{V2T}^{(i,j)}$ denotes the global similarity of $v_{i}$ and $t_j$ . Following CLIP [51], we adopt InfoNCE loss [26] to pull the embeddings of matched image-text pairs together while pushing those of non-matched pairs apart, which is formulated as", + "bbox": [ + 496, + 775, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "19020", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "follows:", + "bbox": [ + 76, + 90, + 133, + 104 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} _ {V 2 T} = - \\frac {1}{2 N} \\sum_ {\\substack {i = 1 \\\\ N}} ^ {N} \\log \\frac {\\exp \\left(\\boldsymbol {Z} _ {V 2 T} ^ {(i , j)} / \\tau\\right)}{\\sum_ {j = 1} ^ {N} \\exp \\left(\\boldsymbol {Z} _ {V 2 T} ^ {(i , j)} / \\tau\\right)} \\tag{2} \\\\ - \\frac {1}{2 N} \\sum_ {j = 1} ^ {N} \\log \\frac {\\exp \\left(\\mathbf {Z} _ {V 2 T} ^ {(i , j)} / \\tau\\right)}{\\sum_ {i = 1} ^ {N} \\exp \\left(\\mathbf {Z} _ {V 2 T} ^ {(i , j)} / \\tau\\right)}, \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 120, + 114, + 468, + 203 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\tau$ is a trainable temperature parameter.", + "bbox": [ + 76, + 215, + 375, + 229 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Text-to-Sign-Video Contrast. Up to now, we have introduced sign-video-to-text contrast. A symmetrical version, termed text-to-sign-video contrast, shares the similar spirit as shown in Figure 3b. The implementation of text-to-sign-video contrast is extremely simple: we replace the row-wise operations (i.e., Softmax and addition) in sign-video-to-text contrast with column-wise ones and keep the remaining processes unchanged. We use $\\mathcal{L}_{T2V}$ to denote the loss function of text-to-sign-video contrast. In our implementation, we reuse the loss defined in Eq 2 but substitute the input with the text-to-video similarity matrix $Z_{T2V}$ .", + "bbox": [ + 75, + 229, + 468, + 393 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Loss Function. The overall loss for cross-lingual contrastive learning is a weighted sum of $\\mathcal{L}_{V2T}$ and $\\mathcal{L}_{T2V}$ with a trade-off hyper-parameter $\\beta$ :", + "bbox": [ + 76, + 396, + 468, + 441 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\beta \\mathcal {L} _ {V 2 T} + (1 - \\beta) \\mathcal {L} _ {T 2 V}. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 171, + 454, + 468, + 470 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.4. Text Augmentation", + "text_level": 1, + "bbox": [ + 76, + 481, + 261, + 498 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Considering that the datasets of sign language retrieval are typically small-scale, we explore text augmentations to improve the generalization of our approach. EDA [64] introduces three simple yet efficient data augmentations in text classification task: random delete randomly removes words in a sentence; synonym replacement randomly selects words from a sentence that are not stop words and replaces them with synonyms; random swap randomly chooses two words in a sentence and swaps their positions. The first two augmentations have been proven effective in text classification task. However, we experimentally find that our focused sign language retrieval is sensitive to random delete and synonym replacement augmentations. To guarantee that the augmented texts preserve the original semantic meanings, we only adopt the random swap augmentation in our approach. We suppose there are two reasons: 1) the word order of sign languages and natural languages are constitutionally distinct, and reordering does not affect semantic meanings; 2) the proposed cross-lingual contrastive learning is insensitive to word order.", + "bbox": [ + 75, + 505, + 468, + 808 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiment", + "text_level": 1, + "bbox": [ + 76, + 821, + 200, + 838 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Datasets and Implementation Details", + "text_level": 1, + "bbox": [ + 76, + 845, + 395, + 863 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Datasets. We primarily focus on How2Sign [19] dataset. Our model is also evaluated on PHOENIX-2014T [8] and", + "bbox": [ + 75, + 869, + 468, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "CSL-Daily [71], which are primarily used for sign language recognition and translation in previous works.", + "bbox": [ + 500, + 90, + 890, + 121 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "How2Sign is a large-scale continuous American sign language (ASL) dataset consisting of a parallel corpus of about 80 hours of sign videos with subtitle annotations. It covers a wide range of instructional videos corresponding to various categories. There are 31164, 1740 and 2356 sign-video-text pairs in training, validation and test sets, respectively. Following SPOT-ALIGN [18], we remove the invalid pairs where the subtitle alignment is detected to exceed the video duration, remaining 31085, 1739 and 2348 available pairs in training, validation and test sets, respectively. The resolution of sign videos is $1280 \\times 720$ , we crop the human bodies of signers with Faster R-CNN [52] to generate valid videos.", + "bbox": [ + 496, + 123, + 890, + 303 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "PHOENIX-2014T is a German sign language (Deutsche Gebärddensprache, DGS) dataset collected in the domain of weather forecast from TV broadcast, consisting of 7096, 519 and 642 video-text pairs in training, validation and test sets, respectively.", + "bbox": [ + 496, + 306, + 890, + 382 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "CSL-Daily is a recently released Chinese sign language (CSL) dataset. The topic of CSL-Daily revolves around people's daily lives, including 18401, 1077, 1176 parallel samples in training, validation and test set, respectively.", + "bbox": [ + 496, + 383, + 890, + 444 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation Metric. Following previous works [18, 43, 51, 57], retrieval performance is evaluated by recall at rank K (R@K, higher is better) and median rank (MedR, lower is better). We evaluate our approach on both text-to-sign-video (T2V) retrieval and sign-video-to-text (V2T) retrieval tasks. We report R@1, R@5, and R@10 in all experiments, and additionally report MedR when comparing with state-of-the-art approaches.", + "bbox": [ + 496, + 446, + 890, + 566 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation Details. The sign encoder takes videos of resolution of $256 \\times 256$ as input. The domain-agnostic sign encoder is a I3D [11] network pre-trained on BSL-1K [61]. In pseudo label generation, we set the threshold $\\lambda$ to 0.6 to filter samples with low-confidence. Non-maximum suppression (NMS) with a temporal window of 24 frames is utilized to remove the duplicates among the pseudo-labeled samples. A collection of approximate 64K pseudo-labeled samples covering a vocabulary of 1220 words is eventually generated. The domain-aware sign encoder is initialized with the domain-agnostic one and fine-tuned with a learning rate of $1 \\times 10^{-2}$ and batch size of 4 for 15 epochs. In the training of cross-lingual contrastive learning, the vision transformer and text transformer are initialized by the image encoder and text encoder in CLIP (ViT-B/32) [51]. The maximum length of sign clip features and text features are set to 64 and 32, respectively. The model is fine-tuned with Adam optimizer [31] with batch size of 512. The initial learning rate is set to $1 \\times 10^{-5}$ , which is decreased with a cosine schedule following the CLIP [51]. We set $\\alpha = 0.8$ in Eq. 1 and $\\beta = 0.5$ in Eq.3. The languages of PHOENIX-2014T and CSL-Daily are German and Chinese", + "bbox": [ + 496, + 568, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "19021", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/0d61485822ac79a3d779bc79c52ed8235b30f836ee46e1195bbbf08b4f79e876.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelT2VV2T
R@1↑R@5↑R@10↑MedR↓R@1↑R@5↑R@10↑MedR↓
SA-SR [18]18.932.136.562.011.627.432.569.0
SA-CM [18]24.340.746.516.017.940.146.914.0
SA-COMB [18]34.248.052.68.023.647.053.07.5
Ours56.669.974.71.051.664.870.11.0
", + "bbox": [ + 86, + 89, + 459, + 170 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/9f3573673a60cd914361f56931d0c810a46d15e8d0f25a98974762254dadd795.jpg", + "table_caption": [ + "Table 1. Comparison with the different variants of the pioneer SPOT-ALIGN (SA) [18] on How2Sign [19] dataset." + ], + "table_footnote": [], + "table_body": "
ModelT2VV2T
R@1↑R@5↑R@10↑MedR↓R@1↑R@5↑R@10↑MedR↓
Translation [10]30.253.163.44.528.852.060.856.1
SA-CM [18]48.676.584.62.050.378.484.41.0
SA-COMB [18]55.879.687.21.053.179.486.11.0
Ours69.586.692.11.070.288.092.81.0
", + "bbox": [ + 86, + 226, + 459, + 308 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/3b75778acbf62206832a51142494ab9f469493ea3d11fa8d81df64a2dbafb47d.jpg", + "table_caption": [ + "Table 2. Comparison with the different variants of the pioneer SPOT-ALIGN (SA) [18] on PHOENIX2014T [8] dataset." + ], + "table_footnote": [], + "table_body": "
ModelT2VV2T
R@1↑R@5↑R@10↑MedR↓R@1↑R@5↑R@10↑MedR↓
Ours75.388.291.91.074.789.492.21.0
", + "bbox": [ + 109, + 363, + 437, + 409 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "respectively. Since CLIP is trained on English corpus, to reuse CLIP's text encoder, we utilize Google translation to translate the texts of these two datasets into English.", + "bbox": [ + 76, + 464, + 468, + 510 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. Comparison with State-of-the-art Methods", + "text_level": 1, + "bbox": [ + 76, + 534, + 441, + 550 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We compare our method with different variants of the pioneer, called SPOT-ALIGN [18], on How2Sign and PHOENIX-2014T. We also provide the results on CSL-Daily as a baseline.", + "bbox": [ + 76, + 563, + 468, + 625 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 1 and Table 2 show the comparisons between our approach and SPOT-ALIGN [18] on How2Sign and PHOENIX-2014T, respectively. SPOT-ALIGN builds the final combination (COMB) model by integrating its primary cross-modal (CM) model with an auxiliary retrieval model (sign recognition (SR) model for How2Sign and Translation [10] for PHOENIX-2014T). Our method outperforms the COMB model, which achieves best results in SPOT-ALIGN, by large margins, achieving $+22.4$ T2V and $+28.0$ V2T R@1 improvements on How2Sign, $+13.7$ T2V and $+17.1$ V2T R@1 improvements on PHOENIX-2014T. It is worth mentioning that the SPOT-ALIGN conducts three rounds of sign spotting [1, 47] and encoder training. In contrast, we simplify the training of sign encoder and only perform a single round of training on pseudo-labeled data. We also provide a baseline on CSL-Daily dataset as shown in Table 3, demonstrating that our model can be generalized to various sign languages.", + "bbox": [ + 76, + 628, + 468, + 901 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/73f6a1930cfeb7bfd77f40c3f23941a8c0aba0e7befd5fc3879dc01ffe186fa4.jpg", + "table_caption": [ + "Table 3. We additionally provide a baseline for CSL-Daily [71] dataset." + ], + "table_footnote": [], + "table_body": "
MethodT2VV2T
R@1R@5R@10R@1R@5R@10
Baseline31.549.657.926.444.452.8
+SE33.351.959.830.948.856.4
+SE+CLCL54.067.171.950.063.268.1
+SE+CLCL+TA56.669.974.751.664.870.1
", + "bbox": [ + 532, + 89, + 859, + 170 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/1fb6164a0165fe1f17b3209729979dc63e4babc4954e7091a05d70e13392e84f.jpg", + "table_caption": [ + "Table 4. Ablation study of each proposed component. SE: sign encoder; CLCL: cross-lingual contrastive learning; TA: text augmentation." + ], + "table_footnote": [], + "table_body": "
EncoderT2VV2T
R@1R@5R@10R@1R@5R@10
Single-Ag [61]53.168.073.447.362.967.0
Single-Aw54.167.573.149.161.867.4
Fusion-Average54.768.773.849.663.769.0
Fusion-Weighted Sum56.669.974.751.664.870.1
", + "bbox": [ + 516, + 231, + 875, + 314 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 5. Results of domain-agnostic sign encoder (Single-Ag) and domain-aware sign encoder (Single-Aw). We also study different ways to integrate the features extracted by Single-Ag and Single-Aw, including average and weighted sum.", + "bbox": [ + 498, + 325, + 890, + 380 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Ablation Study", + "text_level": 1, + "bbox": [ + 500, + 390, + 653, + 407 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We conduct all ablation studies on the most challenging How2Sign dataset.", + "bbox": [ + 498, + 415, + 890, + 444 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The Effectiveness of Each Proposed Component. Table 4 shows the ablation of each component. We first build a baseline where the sign encoder is the domain-agnostic one and contrastive learning is trained with the standard contrastive loss [51]. Then we add the proposed sign encoder (SE), cross-lingual contrastive learning (CLCL) and text augmentation (TA) to the baseline step by step. SE encodes domain-relevant and discriminative features, yielding $+1.8$ T2V and $+4.5$ V2T R@1 gains. The proposed CLCL significantly boosts the performance by $+20.7$ T2V and $+19.1$ V2T R@1 improvements, demonstrating that identifying fine-grained sign-to-word mappings is essential for sign language retrieval. The introduction of TA further promotes the retrieval task, achieving $56.6$ R@1 on V2T and $51.6$ R@1 on T2V, respectively.", + "bbox": [ + 496, + 446, + 892, + 672 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Various Sign Encoders. As described in Section 3.2, our sign encoder is composed of a domain-agnostic sign encoder (Single-Ag) [61] and a domain-aware sign encoder (Single-Aw). We first report the results of each individual encoder in Table 5. Next, we study the different ways to integrate the features extracted by these two encoders, including average and weighted sum as defined in Eq. 1. The results are also shown in Table 5. We experimentally find that the weighted sum strategy yields best results.", + "bbox": [ + 496, + 672, + 890, + 808 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Variants of Cross-Lingual Contrastive Learning. As described in Section 3.3 and illustrated in Figure 3b, through the use of a combination of softmax, multiplication, sum, and average operations, we convert the fine-grained cross-lingual (sign-to-word) similarity to the coarse-grained sign-video-to-text similarity to enable contrastive learning. We", + "bbox": [ + 496, + 810, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "19022", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/096d60734936273dee4db1cc5cb4e2060739936fa2610ad0554c334de416b0ec.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
StrategyT2VV2T
R@1R@5R@10R@1R@5R@10
Mean33.152.559.729.847.855.3
Max42.259.766.038.555.162.0
Softmax56.669.974.751.664.870.1
", + "bbox": [ + 127, + 89, + 419, + 154 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/63ea8cb87b322705488ee5fcd88c48582643d7f56a1ab385c6aa4ed0a085235b.jpg", + "table_caption": [ + "Table 6. Study on different strategies to identify the fine-grained sign-to-word mappings in cross-lingual contrastive learning." + ], + "table_footnote": [], + "table_body": "
Text AugmentationT2VV2T
R@1R@5R@10R@1R@5R@10
None54.067.171.950.063.268.1
RD52.766.371.647.662.167.4
SR53.968.773.047.561.265.9
RS56.669.974.751.664.870.1
", + "bbox": [ + 116, + 203, + 431, + 279 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "refer to this process as \"Softmax\". Here we study two variants termed \"Mean\" and \"Max\". The \"Mean\" and \"Max\" strategies simply replace the combination of softmax, multiplication and sum operations with a simple mean operation and a max operation, respectively. The results are shown in Table 6. We observe that the \"Mean\" strategy performs worst since it merely evaluates the overall similarity of a text and a sign video and ignores the fine-grained sign-to-word mappings during training. The \"Max\" strategy identifies hard sign-to-word mappings, i.e., one sign is associated with the most similar word, and vice versa. Nevertheless, since we do not have the ground-truth of sign-to-word mappings, it is challenging for models to identify the confident one-to-one mappings in a weakly supervised manner, as discussed in previous works on multiple instance learning [47]. In contrast, our default \"Softmax\" strategy localizes the soft sign-to-word mappings and achieves the best results.", + "bbox": [ + 75, + 339, + 467, + 595 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Study of Text Augmentations. In Table 7, we study three text augmentations described in Section 3.4: random delete (RD), synonym replacement (SR) and random swap (RS). We observe a slight performance drop when introducing RD and SR into training. Sign language retrieval is not only a video-text retrieval task, but also a cross-lingual retrieval challenge, deletion and replacement may break the intrinsic sign-to-word mappings. In contrast, RS augmentation preserves the semantics of texts and we observe a +2.6 T2V and a +1.6 V2T R@1 improvements over the counterpart without any text augmentations.", + "bbox": [ + 75, + 597, + 467, + 762 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The Effects of CLIP Initialization. In our framework, the vision and text Transformer are initialized with the CLIP's vision encoder and text encoder. Note that the input modality of our vision Transformer and that of CLIP's vision encoder are different. Ours takes a feature sequence as input while the input of CLIP's image encoder is a set of image patches. In Table 8, we compare a randomly initialized baseline with the one initialized by CLIP. We find that CLIP-initialization significantly improves the perfor", + "bbox": [ + 75, + 763, + 467, + 900 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/e0ed1d145c38eb13d031044a6539d5c632f2147d9c5f02c9af73060d1da63e23.jpg", + "table_caption": [ + "Table 7. Ablation study on different text augmentations. RD: random delete; SR: synonym replacement; RS: random swap." + ], + "table_footnote": [], + "table_body": "
InitializationT2VV2T
R@1R@5R@10R@1R@5R@10
Random45.260.367.540.654.559.7
CLIP56.669.974.751.664.870.1
", + "bbox": [ + 542, + 89, + 849, + 143 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/d68ae9c2a50b495656d104b8026f8e7f66adda27835bf597822ea28ddb51e04e.jpg", + "table_caption": [ + "Table 8. Comparison between CLIP initialization and random initialization." + ], + "table_footnote": [], + "table_body": "
Frozen layersT2VV2T
R@1R@5R@10R@1R@5R@10
None56.669.974.751.664.870.1
154.468.273.449.763.568.8
253.868.673.348.562.867.9
651.766.671.146.260.966.6
1249.565.271.444.759.765.2
", + "bbox": [ + 540, + 191, + 851, + 284 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 9. Study of freezing different layers of text Transformer.", + "bbox": [ + 509, + 295, + 879, + 309 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "mance, yielding $+11.4$ T2V and $+11.0$ V2T R@1 improvements though the input originates from a completely distinct modality.", + "bbox": [ + 498, + 334, + 890, + 378 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Study on Text Transformer. The text Transformer in our model is initialized with the CLIP's text encoder. Considering the generalization capability of the pre-trained CLIP, we attempt to freeze the shallow layers of our text transformer. As shown in Table 9, the performance gradually decreases as the number of the frozen layers increases. CLIP is trained on large-scale image-text pairs, however, the domain gap between image-text data and sign-video-text data is non-negligible. Though CLIP shows promising transfer capacity, it is optimal to fine-tune the whole model on target datasets.", + "bbox": [ + 498, + 382, + 890, + 546 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 568, + 617, + 583 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we propose a novel framework named Cross-lingual Contrastive learning (CiCo) for recently introduced sign language retrieval task. We formulate sign language retrieval as a cross-lingual retrieval task as well as a video-text retrieval problem. CiCo models the fine-grained cross-lingual mappings between sign videos and texts via the proposed cross-lingual contrastive learning. We also introduce a sign encoder, which is composed of a domain-agnostic encoder and a domain-aware one, to extract discriminative and domain-aligned features. Our Cico outperforms the pioneer SPOT-ALIGN by large margins on How2Sign and PHOENIX-2014T benchmarks. We also provide a baseline on CSL-Daily. We hope our approach could serve as a solid baseline for future research.", + "bbox": [ + 498, + 595, + 890, + 806 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgement. This work was partially supported by National Natural Science Foundation of China (No. 62072112), National Key R&D Program of China (No. 2020AAA0108301), and Scientific and Technological Innovation Action Plan of Shanghai Science and Technology Committee (No. 22511101502, 22511102202).", + "bbox": [ + 498, + 809, + 890, + 898 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "19023", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Samuel Albanie, Gül Varol, Liliane Momeni, Triantafyllos Afouras, Joon Son Chung, Neil Fox, and Andrew Zisserman. Bsl-1k: Scaling up co-articulated sign language recognition using mouthing cues. In European Conference on Computer Vision, pages 35–53. Springer, 2020. 1, 3, 7", + "[2] Mikel Artetxe, Gorka Labaka, and Eneko Agirre. Learning bilingual word embeddings with (almost) no bilingual data. In *Meeting of the Association for Computational Linguistics*, pages 451–462, 2017. 3", + "[3] Mikel Artetxe, Gorka Labaka, and Eneko Agirre. A robust self-learning method for fully unsupervised cross-lingual mappings of word embeddings. arXiv preprint arXiv:1805.06297, 2018. 3", + "[4] Vassilis Athitsos, Carol Neidle, Stan Sclaroff, Joan Nash, Alexandra Stefan, Ashwin Thangali, Hajjing Wang, and Quan Yuan. Large lexicon project: American sign language video corpus and sign language indexing/retrieval algorithms. In Workshop on the Representation and Processing of Sign Languages: Corpora and Sign Language Technologies, volume 2, pages 11-14, 2010. 3", + "[5] Max Bain, Arsha Nagrani, Gül Varol, and Andrew Zisserman. Frozen in time: A joint video and image encoder for end-to-end retrieval. In IEEE/CVF International Conference on Computer Vision, pages 1728-1738, 2021. 2, 3", + "[6] Lisa Ballesteros and Bruce Croft. Dictionary methods for cross-lingual information retrieval. In International Conference on Database and Expert Systems Applications, pages 791–801. Springer, 1996. 2", + "[7] Yoshua Bengio, Aaron Courville, and Pascal Vincent. Representation learning: A review and new perspectives. IEEE Transactions on Pattern Analysis and Machine Intelligence, 35(8):1798-1828, 2013. 3", + "[8] Necati Cihan Camgoz, Simon Hadfield, Oscar Koller, Hermann Ney, and Richard Bowden. Neural sign language translation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7784-7793, 2018. 2, 3, 6, 7", + "[9] Necati Cihan Camgoz, Oscar Koller, Simon Hadfield, and Richard Bowden. Multi-channel transformers for multi-articulatory sign language translation. In European Conference on Computer Vision, pages 301-319. Springer, 2020. 3", + "[10] Necati Cihan Camgoz, Oscar Koller, Simon Hadfield, and Richard Bowden. Sign language transformers: Joint end-to-end sign language recognition and translation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10023-10033, 2020. 1, 7", + "[11] Joao Carreira and Andrew Zisserman. Quo vadis, action recognition? a new model and the kinetics dataset. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6299-6308, 2017. 3, 6", + "[12] Yutong Chen, Fangyun Wei, Xiao Sun, Zhirong Wu, and Stephen Lin. A simple multi-modality transfer learning baseline for sign language translation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5120-5130, 2022. 1, 3" + ], + "bbox": [ + 78, + 114, + 470, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[13] Yutong Chen, Ronglai Zuo, Fangyun Wei, Yu Wu, Shujie LIU, and Brian Mak. Two-stream network for sign language recognition and translation. In Alice H. Oh, Alekh Agarwal, Danielle Belgrave, and Kyunghyun Cho, editors, Advances in Neural Information Processing Systems, 2022. 1, 3", + "[14] Necati Cihan Camgoz, Simon Hadfield, Oscar Koller, and Richard Bowden. Subunets: End-to-end hand shape and continuous sign language recognition. In IEEE/CVF International Conference on Computer Vision, pages 3056-3065, 2017. 1", + "[15] Runpeng Cui, Hu Liu, and Changshui Zhang. A deep neural framework for continuous sign language recognition by iterative training. IEEE Transactions on Multimedia, 21(7):1880-1891, 2019. 1", + "[16] Tonni Das Jui, Gissella Maria Bejarano, and Pablo Rivas. A machine learning-based segmentation approach for measuring similarity between sign languages. In sign-lang@ LREC 2022, pages 94-101. European Language Resources Association (ELRA), 2022. 4", + "[17] Yu Du, Fangyun Wei, Zihe Zhang, Miaojing Shi, Yue Gao, and Guoqi Li. Learning to prompt for open-vocabulary object detection with vision-language model. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14084-14093, 2022. 5", + "[18] Amanda Duarte, Samuel Albanie, Xavier Giró-i Nieto, and Gül Varol. Sign language video retrieval with free-form textual queries. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14094-14104, 2022. 1, 2, 3, 4, 6, 7", + "[19] Amanda Duarte, Shruti Palaskar, Lucas Ventura, Deepti Ghadiyaram, Kenneth DeHaan, Florian Metze, Jordi Torres, and Xavier Giro-i Nieto. How2Sign: A Large-scale Multimodal Dataset for Continuous American Sign Language. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2021. 2, 6, 7", + "[20] Ali Farhadi, David Forsyth, and Ryan White. Transfer learning in sign language. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1-8. IEEE, 2007. 3", + "[21] Holger Fillbrandt, Suat Akyol, and K-F Kraiss. Extraction of 3d hand shape and posture from image sequences for sign language recognition. In 2003 IEEE International SOI Conference, pages 181-186. IEEE, 2003. 3", + "[22] Stanislav Fort, Jie Ren, and Balaji Lakshminarayanan. Exploring the limits of out-of-distribution detection. Advances in Neural Information Processing Systems, 34:7068-7081, 2021. 5", + "[23] Andrea Frome, Greg S Corrado, Jon Shlens, Samy Bengio, Jeff Dean, Marc'Aurelio Ranzato, and Tomas Mikolov. Devise: A deep visual-semantic embedding model. Advances in neural information processing systems, 26, 2013. 3, 5", + "[24] Valentin Gabeur, Chen Sun, Karteek Alahari, and Cordelia Schmid. Multi-modal transformer for video retrieval. In European Conference on Computer Vision, pages 214-229. Springer, 2020. 2, 3", + "[25] Yunchao Gong, Qifa Ke, Michael Isard, and Svetlana Lazebnik. A multi-view embedding space for modeling internet images, tags, and their semantics. International Journal of Computer Vision, 106(2):210-233, 2014. 3, 5" + ], + "bbox": [ + 501, + 92, + 893, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "19024", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[26] Michael Gutmann and Aapo Hyvarinen. Noise-contrastive estimation: A new estimation principle for unnormalized statistical models. In International Conference on Artificial Intelligence and Statistics, pages 297–304. JMLR Workshop and Conference Proceedings, 2010. 5", + "[27] Tony Huang, Jack Chu, and Fangyun Wei. Unsupervised prompt learning for vision-language models. arXiv preprint arXiv:2204.03649, 2022. 5", + "[28] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In International Conference on Machine Learning, pages 4904-4916. PMLR, 2021. 3", + "[29] Armand Joulin, Laurens van der Maaten, Allan Jabri, and Nicolas Vasilache. Learning visual features from large weakly supervised data. In European Conference on Computer Vision, pages 67-84. Springer, 2016. 3, 5", + "[30] Hamid Reza Vaezi Joze and Oscar Koller. Ms-asl: A large-scale data set and benchmark for understanding american sign language. arXiv preprint arXiv:1812.01053, 2018. 3", + "[31] Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In Yoshua Bengio and Yann LeCun, editors, 3rd International Conference on Learning Representations, ICLR 2015, San Diego, CA, USA, May 7-9, 2015, Conference Track Proceedings, 2015. 6", + "[32] Oscar Koller, Necati Cihan Camgoz, Hermann Ney, and Richard Bowden. Weakly supervised learning with multistream cnn-lstm-hmms to discover sequential parallelism in sign language videos. IEEE Transactions on Pattern Analysis and Machine Intelligence, 42(9):2306-2320, 2019. 1", + "[33] Oscar Koller, Jens Forster, and Hermann Ney. Continuous sign language recognition: Towards large vocabulary statistical recognition systems handling multiple signers. Computer Vision and Image Understanding, 141:108-125, 2015. 1", + "[34] Victor Lavrenko, Martin Choquette, and W Bruce Croft. Cross-lingual relevance models. In International ACM SIGIR Conference on Research and Development in Information Retrieval, pages 175–182, 2002. 2, 3", + "[35] Yann LeCun, Yoshua Bengio, and Geoffrey Hinton. Deep learning. Nature, 521(7553):436-444, 2015. 3", + "[36] Dongxu Li, Cristian Rodriguez, Xin Yu, and Hongdong Li. Word-level deep sign language recognition from video: A new large-scale dataset and methods comparison. In IEEE/CVF winter conference on applications of computer vision, pages 1459-1469, 2020. 3", + "[37] Dongxu Li, Chenchen Xu, Xin Yu, Kaihao Zhang, Benjamin Swift, Hanna Suominen, and Hongdong Li. Tspnet: Hierarchical feature learning via temporal semantic pyramid for sign language translation. Advances in Neural Information Processing Systems, 33:12034-12045, 2020. 3", + "[38] Dongxu Li, Xin Yu, Chenchen Xu, Lars Petersson, and Hongdong Li. Transferring cross-domain knowledge for video sign language recognition. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6205-6214, 2020. 3", + "[39] Junnan Li, Ramprasaath Selvaraju, Akhilesh Gotmare, Shafiq Joty, Caiming Xiong, and Steven Chu Hong Hoi." + ], + "bbox": [ + 78, + 90, + 470, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Align before fuse: Vision and language representation learning with momentum distillation. Advances in Neural Information Processing Systems, 34:9694-9705, 2021. 3, 5", + "[40] Yangguang Li, Feng Liang, Lichen Zhao, Yufeng Cui, Wanli Ouyang, Jing Shao, Fengwei Yu, and Junjie Yan. Supervision exists everywhere: A data efficient contrastive language-image pre-training paradigm. arXiv preprint arXiv:2110.05208, 2021. 5", + "[41] Ji Lin, Chuang Gan, and Song Han. Tsm: Temporal shift module for efficient video understanding. In IEEE/CVF International Conference on Computer Vision, pages 7083-7093, 2019. 3", + "[42] Song Liu, Haoqi Fan, Shengsheng Qian, Yiru Chen, Wenkui Ding, and Zhongyuan Wang. Hit: Hierarchical transformer with momentum contrast for video-text retrieval. In IEEE/CVF International Conference on Computer Vision, pages 11915-11925, 2021. 2, 3", + "[43] Yang Liu, Samuel Albanie, Arsha Nagrani, and Andrew Zisserman. Use what you have: Video retrieval using representations from collaborative experts. arXiv preprint arXiv:1907.13487, 2019. 2, 3, 6", + "[44] Huaishao Luo, Lei Ji, Ming Zhong, Yang Chen, Wen Lei, Nan Duan, and Tianrui Li. Clip4clip: An empirical study of clip for end to end video clip retrieval and captioning. Neurocomputing, 508:293-304, 2022. 2, 3", + "[45] Antoine Miech, Dimitri Zhukov, Jean-Baptiste Alayrac, Makarand Tapaswi, Ivan Laptev, and Josef Sivic. Howto100m: Learning a text-video embedding by watching hundred million narrated video clips. In IEEE/CVF International Conference on Computer Vision, pages 2630-2640, 2019. 2", + "[46] Tomas Mikolov, Quoc V Le, and Ilya Sutskever. Exploiting similarities among languages for machine translation. arXiv preprint arXiv:1309.4168, 2013. 3", + "[47] Liliane Momeni, Gül Varol, Samuel Albanie, Triantafyllos Afouras, and Andrew Zisserman. Watch, read and lookup: learning to spot signs from multiple supervisors. In Asian Conference on Computer Vision, 2020. 3, 4, 7, 8", + "[48] Mathew Monfort, SouYoung Jin, Alexander Liu, David Harwath, Rogerio Feris, James Glass, and Aude Oliva. Spoken moments: Learning joint audio-visual representations from video descriptions. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14871-14881, 2021. 2", + "[49] Yasuhide Mori, Hironobu Takahashi, and Ryuichi Oka. Image-to-word transformation based on dividing and vector quantizing images with words. In International Workshop on Multimedia Intelligent Storage and Retrieval Management, pages 1-9. Citeseer, 1999. 3", + "[50] Aitor Ormazabal, Mikel Artetxe, Gorka Labaka, Aitor Soroa, and Eneko Agirre. Analyzing the limitations of cross-lingual word embedding mappings. arXiv preprint arXiv:1906.05407, 2019. 3", + "[51] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language super" + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "19025", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "vision. In International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 3, 5, 6, 7", + "[52] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. Conference on Neural Information Processing Systems, 28, 2015. 6", + "[53] Rico Sennrich, Barry Haddow, and Alexandra Birch. Neural machine translation of rare words with subword units. arXiv preprint arXiv:1508.07909, 2015. 5", + "[54] Sheng Shen, Liunian Harold Li, Hao Tan, Mohit Bansal, Anna Rohrbach, Kai-Wei Chang, Zhewei Yao, and Kurt Keutzer. How much can clip benefit vision-and-language tasks? arXiv preprint arXiv:2107.06383, 2021. 5", + "[55] Thad Starner, Joshua Weaver, and Alex Pentland. Real-time american sign language recognition using desk and wearable computer based video. IEEE Transactions on Pattern Analysis and Machine Intelligence, 20(12):1371-1375, 1998. 3", + "[56] Thad E Starner. Visual recognition of american sign language using hidden markov models. Technical report, Massachusetts inst of tech Cambridge dept of brain and cognitive sciences, 1995. 3", + "[57] Chen Sun, Austin Myers, Carl Vondrick, Kevin Murphy, and Cordelia Schmid. Videobert: A joint model for video and language representation learning. In IEEE/CVF International Conference on Computer Vision, pages 7464-7473, 2019. 2, 3, 6", + "[58] Rachel Sutton-Spence. Mouthings and simultaneity in British sign language. Amsterdam Studies in the Theory and History of Linguistic Science Series, 281:147, 2007. 3", + "[59] Chau Tran, Yuqing Tang, Xian Li, and Jiatao Gu. Crosslingual retrieval for iterative self-supervised training. Advances in Neural Information Processing Systems, 33:2207-2219, 2020. 2, 3", + "[60] Du Tran, Heng Wang, Lorenzo Torresani, Jamie Ray, Yann LeCun, and Manohar Paluri. A closer look at spatiotemporal convolutions for action recognition. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6450-6459, 2018. 3", + "[61] Gul Varol, Liliane Momeni, Samuel Albanie, Triantafyllos Afouras, and Andrew Zisserman. Read and attend: Temporal localisation in sign language videos. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16857-16866, 2021. 1, 2, 3, 4, 6, 7", + "[62] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. 2017. 5", + "[63] Xiaolong Wang, Ross Girshick, Abhinav Gupta, and Kaiming He. Non-local neural networks. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7794-7803, 2018. 3", + "[64] Jason Wei and Kai Zou. EDA: Easy data augmentation techniques for boosting performance on text classification tasks. In 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages 6383-6389, Hong Kong, China, Nov. 2019. Association for Computational Linguistics. 6" + ], + "bbox": [ + 78, + 90, + 468, + 900 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[65] Jinxi Xu, Ralph Weischedel, and Chanh Nguyen. Evaluating a probabilistic model for cross-lingual information retrieval. In International ACM SIGIR Conference on Research and Development in Information Retrieval, pages 105–110, 2001. 3", + "[66] Mengde Xu, Zheng Zhang, Fangyun Wei, Han Hu, and Xiang Bai. Side adapter network for open-vocabulary semantic segmentation. arXiv preprint arXiv:2302.12242, 2023. 5", + "[67] Mengde Xu, Zheng Zhang, Fangyun Wei, Yutong Lin, Yue Cao, Han Hu, and Xiang Bai. A simple baseline for open-vocabulary semantic segmentation with pre-trained vision-language model. In European Conference on Computer Vision, pages 736-753. Springer, 2022. 5", + "[68] Youngjae Yu, Jongseok Kim, and Gunhee Kim. A joint sequence fusion model for video question answering and retrieval. In European Conference on Computer Vision, pages 471-487, 2018. 2, 3", + "[69] Mozhi Zhang, Keyulu Xu, Ken-ichi Kawarabayashi, Stefanie Jegelka, and Jordan Boyd-Graber. Are girls neko or shōjo? cross-lingual alignment of non-isomorphic embeddings with iterative normalization. arXiv preprint arXiv:1906.01622, 2019.3", + "[70] Shilin Zhang and Bo Zhang. Using revised string edit distance to sign language video retrieval. In International Conference on Computational Intelligence and Natural Computing, volume 1, pages 45-49. IEEE, 2010. 3", + "[71] Hao Zhou, Wengang Zhou, Weizhen Qi, Junfu Pu, and Houqiang Li. Improving sign language translation with monolingual data by sign back-translation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1316-1325, 2021. 6, 7", + "[72] Hao Zhou, Wengang Zhou, Yun Zhou, and Houqiang Li. Spatial-temporal multi-cue network for sign language recognition and translation. IEEE Transactions on Multimedia, 24:768-779, 2021. 3", + "[73] Ronglai Zuo, Fangyun Wei, and Brian Mak. Natural language-assisted sign language recognition. IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023. 1, 3" + ], + "bbox": [ + 501, + 92, + 890, + 640 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "19026", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2023/CiCo_ Domain-Aware Sign Language Retrieval via Cross-Lingual Contrastive Learning/047059b8-c0b2-4bcc-9546-ab9b6e4bc0e5_model.json b/2023/CiCo_ Domain-Aware Sign Language Retrieval via Cross-Lingual Contrastive Learning/047059b8-c0b2-4bcc-9546-ab9b6e4bc0e5_model.json new file mode 100644 index 0000000000000000000000000000000000000000..68c2213b2ee88d47a1fc4bf1b26435366a0ef091 --- /dev/null +++ b/2023/CiCo_ Domain-Aware Sign Language Retrieval via Cross-Lingual Contrastive Learning/047059b8-c0b2-4bcc-9546-ab9b6e4bc0e5_model.json @@ -0,0 +1,2378 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.23, + 0.131, + 0.744, + 0.177 + ], + "angle": 0, + "content": "CiCo: Domain-Aware Sign Language Retrieval via Cross-Lingual Contrastive Learning" + }, + { + "type": "text", + "bbox": [ + 0.151, + 0.203, + 0.826, + 0.258 + ], + "angle": 0, + "content": "Yiting Cheng\\(^{1}\\), Fangyun Wei\\(^{2*}\\), Jianmin Bao\\(^{2}\\), Dong Chen\\(^{2}\\), Wenqiang Zhang\\(^{1*}\\) \n\\(^{1}\\)School of Computer Science, Fudan University, \\(^{2}\\)Microsoft Research Asia \n{ytcheng18, wqzhang}@fudan.edu.cn, {fawe, jianbao, doch}@microsoft.com" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.292, + 0.314, + 0.308 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.324, + 0.474, + 0.765 + ], + "angle": 0, + "content": "This work focuses on sign language retrieval—a recently proposed task for sign language understanding. Sign language retrieval consists of two sub-tasks: text-to-sign-video (T2V) retrieval and sign-video-to-text (V2T) retrieval. Different from traditional video-text retrieval, sign language videos, not only contain visual signals but also carry abundant semantic meanings by themselves due to the fact that sign languages are also natural languages. Considering this character, we formulate sign language retrieval as a cross-lingual retrieval problem as well as a video-text retrieval task. Concretely, we take into account the linguistic properties of both sign languages and natural languages, and simultaneously identify the fine-grained cross-lingual (i.e., sign-to-word) mappings while contrasting the texts and the sign videos in a joint embedding space. This process is termed as cross-lingual contrastive learning. Another challenge is raised by the data scarcity issue—sign language datasets are orders of magnitude smaller in scale than that of speech recognition. We alleviate this issue by adopting a domain-agnostic sign encoder pre-trained on large-scale sign videos into the target domain via pseudolabeling. Our framework, termed as domain-aware sign language retrieval via Cross-lingual Contrastive learning or CiCo for short, outperforms the pioneering method by large margins on various datasets, e.g., +22.4 T2V and +28.0 V2TR@1 improvements on How2Sign dataset, and +13.7 T2V and +17.1 V2TR@1 improvements on PHOENIX-2014T dataset. Code and models are available at: https://github.com/FangyunWei/SLRT." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.791, + 0.21, + 0.807 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.817, + 0.47, + 0.879 + ], + "angle": 0, + "content": "Sign languages are the primary means of communication used by people who are deaf or hard of hearing. Sign language understanding [1, 10, 12-15, 18, 32, 33, 61, 73] is significant for overcoming the communication barrier between" + }, + { + "type": "image", + "bbox": [ + 0.52, + 0.294, + 0.875, + 0.374 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.595, + 0.377, + 0.798, + 0.39 + ], + "angle": 0, + "content": "(a) Text-to-sign-video (T2V) retrieval." + }, + { + "type": "image", + "bbox": [ + 0.519, + 0.393, + 0.872, + 0.47 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.594, + 0.474, + 0.795, + 0.486 + ], + "angle": 0, + "content": "(b) Sign-video-to-text (V2T) retrieval." + }, + { + "type": "image_caption", + "bbox": [ + 0.515, + 0.498, + 0.877, + 0.513 + ], + "angle": 0, + "content": "Figure 1. Illustration of: (a) T2V retrieval; (b) V2T retrieval." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.53, + 0.895, + 0.682 + ], + "angle": 0, + "content": "the hard-of-hearing and non-signers. Sign language recognition and translation (SLRT) has been extensively studied, with the goal of recognizing the arbitrary semantic meanings conveyed by sign languages. However, the lack of available data significantly limits the capability of SLRT. In this paper, we focus on developing a framework for a recently proposed sign language retrieval task [18]. Unlike SLRT, sign language retrieval focuses on retrieving the meanings that signers express from a closed-set, which can significantly reduce error rates in realistic deployment." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.685, + 0.895, + 0.838 + ], + "angle": 0, + "content": "Sign language retrieval is both similar to and distinct from the traditional video-text retrieval. On the one hand, like video-text retrieval, sign language retrieval is also composed of two sub-tasks, i.e., text-to-sign-video (T2V) retrieval and sign-video-to-text (V2T) retrieval. Given a free-form written query and a large collection of sign language videos, the objective of T2V is to find the video that best matches the written query (Figure 1a). In contrast, the goal of V2T is to identify the most relevant text description given a query of sign language video (Figure 1b)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.895, + 0.903 + ], + "angle": 0, + "content": "On the other hand, different from the video-text retrieval, sign languages, like most natural languages, have their own grammars and linguistic properties. Therefore, sign language videos not only contain visual signals, but also carry" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.887, + 0.222, + 0.9 + ], + "angle": 0, + "content": "*Corresponding author." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "19016" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.095, + 0.33, + 0.11 + ], + "angle": 0, + "content": "(Sign Language)" + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.112, + 0.449, + 0.266 + ], + "angle": 0, + "content": null + }, + { + "type": "image_footnote", + "bbox": [ + 0.086, + 0.276, + 0.457, + 0.338 + ], + "angle": 0, + "content": "(a) While contrasting the sign videos and the texts in a joint embedding space, we simultaneously identify the fine-grained cross-lingual (sign-to-word) mappings of sign languages and natural languages via the proposed cross-lingual contrastive learning. Existing datasets do not annotate the sign-to-word mappings." + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.342, + 0.27, + 0.44 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.274, + 0.342, + 0.449, + 0.44 + ], + "angle": 0, + "content": null + }, + { + "type": "image_footnote", + "bbox": [ + 0.086, + 0.446, + 0.457, + 0.484 + ], + "angle": 0, + "content": "(b) We show four instances of the sign \"Book\" in How2Sign [19] dataset, which are identified by our approach. Please refer to the supplementary material for more examples." + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.495, + 0.47, + 0.525 + ], + "angle": 0, + "content": "Figure 2. Illustration of: (a) cross-lingual (sign-to-word) mapping; (b) sign-to-word mappings identified by our CiCo." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.532, + 0.47, + 0.637 + ], + "angle": 0, + "content": "semantics (i.e., sign1-to-word mappings between sign languages and natural languages) by themselves, which differentiates them from the general videos that merely contain visual information. Considering the linguistic characteristics of sign languages, we formulate sign language retrieval as a cross-lingual retrieval [6,34,59] problem in addition to a video-text retrieval [5,24,42-44,57,68] task." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.638, + 0.471, + 0.879 + ], + "angle": 0, + "content": "Sign language retrieval is extremely challenging due to the following reasons: (1) Sign languages are completely separate and distinct from natural languages since they have unique linguistic rules, word formation, and word order. The transcription between sign languages and natural languages is complicated, for instance, the word order is typically not preserved between sign languages and natural languages. It is necessary to automatically identify the sign-to-word mapping from the cross-lingual retrieval perspective; (2) In contrast to the text-video retrieval datasets [45, 48] which contain millions of training samples, sign language datasets are orders of magnitude smaller in scale—for example, there are only \\(30\\mathrm{K}\\) video-text pairs in How2Sign [19] training set; (3) Sign languages convey information through the handshape, facial expression, and body movement, which requires models to distinguish fine" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.092, + 0.892, + 0.167 + ], + "angle": 0, + "content": "grained gestures and actions; (4) Sign language videos typically contain hundreds of frames. It is necessary to build efficient algorithms to lower the training cost and fit the long videos as well as the intermediate representations into limited GPU memory." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.168, + 0.892, + 0.197 + ], + "angle": 0, + "content": "In this work, we concentrate on resolving the challenges listed above:" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.204, + 0.892, + 0.356 + ], + "angle": 0, + "content": "- We consider the linguistic rules (e.g., word order) of both sign languages and natural languages. We formulate sign language retrieval as a cross-lingual retrieval task as well as a video-text retrieval problem. While contrasting the sign videos and the texts in a joint embedding space as achieved in most vision-language pre-training frameworks [5, 44, 57], we simultaneously identify the fine-grained cross-lingual (sign-to-word) mappings between two types of languages via our proposed cross-lingual contrastive learning as shown in Figure 2." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.364, + 0.892, + 0.561 + ], + "angle": 0, + "content": "- Data scarcity typically brings in the over-fitting issue. To alleviate this issue, we adopt transfer learning and adapt a recently released domain-agnostic sign encoder [61] pre-trained on large-scale sign-videos to the target domain. Although this encoder is capable of distinguishing the fine-grained signs, direct transferring may be suboptimal due to the unavoidable domain gap between the pre-training dataset and sign language retrieval datasets. To tackle this problem, we further fine-tune a domain-aware sign encoder on pseudo-labeled data from target datasets. The final sign encoder is composed of the well-optimized domain-aware sign encoder and the powerful domain-agnostic sign encoder." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.569, + 0.892, + 0.676 + ], + "angle": 0, + "content": "- In order to effectively model long videos, we decouple our framework into two disjoint parts: (1) a sign encoder which adopts a sliding window on sign-videos to pre-extract their vision features; (2) a cross-lingual contrastive learning module which encodes the extracted vision features and their corresponding texts in a joint embedding space." + }, + { + "type": "list", + "bbox": [ + 0.5, + 0.204, + 0.892, + 0.676 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.682, + 0.892, + 0.833 + ], + "angle": 0, + "content": "Our framework, called domain-aware sign language retrieval via Cross-lingual Contrastive learning or CiCo for short, outperforms the pioneer SPOT-ALIGN [18] by large margins on various datasets, achieving 56.6 (+22.4) T2V and 51.6 (+28.0) V2T R@1 accuracy (improvement) on How2Sign [19] dataset, and 69.5 (+13.7) T2V and 70.2 (+17.1) V2T R@1 accuracy (improvement) on PHOENIX-2014T [8] dataset. With its simplicity and strong performance, we hope our approach can serve as a solid baseline for future research." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.845, + 0.642, + 0.861 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Sign Language Understanding. Sign language understanding aims at interpreting the semantic information con" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.887, + 0.468, + 0.902 + ], + "angle": 0, + "content": "1We use sign to denote lexical item within a sign language vocabulary." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19017" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.471, + 0.168 + ], + "angle": 0, + "content": "veyed within sign videos. Researchers have explored such capability on various tasks including sign language recognition [13,20,21,30,38,58,73], sign spotting [1,18,47,61], sign language translation [8,9,12,37,72] and our focused sign language retrieval [18]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.17, + 0.47, + 0.336 + ], + "angle": 0, + "content": "One of the fundamental tasks of sign language understanding is sign language recognition (SLR), which aims to transcribe a sign video into a gloss sequence. Previous works on SLR focus on designing carefully engineered features [20, 21, 58] or modeling temporal dependencies [55, 56]. Recently, the success of 3D convolutional neural networks in action related tasks [41, 60, 63] is transferred to SLR. In particular, the I3D [11] architecture has proven to be effective for this task [1, 30, 36, 38, 61]. In this work, we also adopt this network architecture in our sign encoder." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.338, + 0.471, + 0.563 + ], + "angle": 0, + "content": "Sign spotting is a particular variant of sign language recognition. It aims to localize all instances of a given sign within an untrimmed video. Recent works tackle this task with auxiliary cue of subtitles, introducing automatic annotation systems by using mouthing [1], dictionaries [47] and attention maps of Transformer [61]. SPOT-ALIGN [18] extends existing spotting methods [1, 47] with an iterative training schema, which alternates between repeated sign spotting and model fine-tuning. In this work, pseudo-labeling is served as our sign spotting approach to localize isolated signs in untrimmed videos from target sign language retrieval datasets. Compared with above efforts, our approach only employs a pre-trained sign encoder without utilizing additional auxiliary cue, which is proved to be simple yet efficient." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.567, + 0.47, + 0.763 + ], + "angle": 0, + "content": "Early works of sign language retrieval primarily investigate query-by-example searching [4, 70], which queries individual instances with given sign examples. Our work focuses on free-form textual retrieval—a recently introduced task by SPOT-ALIGN [18]. It symbolizes the real-world scenario of searching sign language videos with natural languages. The pioneer SPOT-ALIGN [18] purely formulates sign language retrieval as a video-text retrieval task, where the cross-modal alignment is modeled upon overall global embeddings of sign videos and texts. However, the linguistic properties of sign languages are ignored. In contrast, we formulate sign language retrieval as a joint task of text-video retrieval and cross-lingual retrieval." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.471, + 0.901 + ], + "angle": 0, + "content": "Vision-Language Models. Learning general-purpose representations for visual and textual modalities is a longstanding topic [7, 35]. The idea has been investigated decades ago [49]. Recently, the prominent success of CLIP [51], ALIGN [28] and ALBEF [39] has demonstrated the capability of learning joint cross-modal representations with large-scale web data using simple image-text contrastive learning. Similar idea has also been explored in video-text retrieval, which is the closest area of our work." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.258 + ], + "angle": 0, + "content": "The common practice in image-text retrieval [23, 25, 29, 51] and video-text retrieval [5, 24, 42-44, 57, 68] is to encode the images/videos and texts into the overall representations. It is reasonable since the visual signals of typical image-text and video-text retrieval datasets mainly describe the certain objects or describable events. In contrast, sign videos, as the carriers of sign languages, convey abundant semantics by themselves. There exist fine-grained mappings between sign videos and natural languages. We find that identifying such cross-lingual (sign-to-word) mappings significantly boosts the performance of sign language retrieval." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.259, + 0.892, + 0.545 + ], + "angle": 0, + "content": "Cross-Lingual Information Retrieval. In this work, we also formulate sign-language retrieval as a cross-lingual task. In natural language processing community, cross-lingual information retrieval (CLIR) [34, 59, 65] refers to the task of retrieving documents between different languages. One of the most common approaches of CLIR is to learn sentence-level embedding alignment by mapping pre-acquired monolingual embeddings [46, 50, 69] of different languages into a shared space. Recently, researchers have exploited fine-grained word-level mappings with self-training [2, 3, 59]. They find that mappings can be learned by initiating with a seed dictionary and alternating between alignment modeling and dictionary mining. Our work also exploits fine-grained word-level mappings, called sign-to-word mappings, in the context of sign language retrieval. Identifying sign-to-word mappings is challenging since sign languages are expressed in visual modality. The proposed cross-lingual contrastive learning tackles this challenge by exploiting the fine-grained cross-modal interactions." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.558, + 0.634, + 0.574 + ], + "angle": 0, + "content": "3. Methodology" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.583, + 0.892, + 0.718 + ], + "angle": 0, + "content": "In this section, we first formulate the task of sign language retrieval in Section 3.1. Next, we introduce our framework which is composed of two parts: 1) a sign encoder which extracts discriminative and domain-aligned features of sign videos (Section 3.2); 2) a cross-lingual contrastive learning framework, which contrasts sign-video-text pairs while concurrently identifying the fine-grained sign-to-word mappings (Section 3.3). At last, we explore text augmentations in Section 3.4." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.727, + 0.673, + 0.741 + ], + "angle": 0, + "content": "3.1. Task Formulation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Let \\(\\mathcal{V}\\) and \\(\\mathcal{T}\\) denote a set of sign videos and their corresponding texts (transcriptions), respectively. Sign language retrieval consists of two tasks namely text-to-sign-video retrieval (T2V), and sign-video-to-text retrieval (V2T), respectively. The objective of T2V is to find the sign video \\(v\\in \\mathcal{V}\\) whose signing content best matches the text query. In contrast, the reverse task V2T requires the model to identify the most relevant text (transcription) \\(t\\in \\mathcal{T}\\) given a query of sign video. We resolve sign language retrieval by learning a joint embedding space of sign videos and texts." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19018" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.101, + 0.09, + 0.87, + 0.305 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.404, + 0.31, + 0.568, + 0.323 + ], + "angle": 0, + "content": "(a) Illustration of sign encoder." + }, + { + "type": "image", + "bbox": [ + 0.102, + 0.328, + 0.87, + 0.574 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.347, + 0.579, + 0.62, + 0.593 + ], + "angle": 0, + "content": "(b) Illustration of cross-lingual contrastive learning." + }, + { + "type": "image_caption", + "bbox": [ + 0.075, + 0.604, + 0.892, + 0.675 + ], + "angle": 0, + "content": "Figure 3. Overview of our framework. (a) Sign encoder is composed of a powerful domain-agnostic sign encoder pre-trained on large-scale sign videos, and a domain-aware sign encoder fine-tuned on pseudo-labeled data from target datasets. We adopt a sliding window manner to extract a discriminative and domain-aligned feature per clip. (b) Cross-lingual contrastive learning takes \\( N \\) sign-video-text pairs as inputs and contrasts paired data in a shared embedding space while implicitly identifying the fine-grained sign-to-word mappings during training." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.682, + 0.218, + 0.699 + ], + "angle": 0, + "content": "3.2. Sign Encoder" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.712, + 0.471, + 0.864 + ], + "angle": 0, + "content": "Process Sign Videos with Sliding Window. Sign videos from sign language retrieval datasets typically contain hundreds of frames. To efficiently train our model and lower the usage of GPU memory, given a sign video, we adopt a sliding window manner with stride of 1 and window size of 16 to produce \\( M \\) temporally overlapped clips. Next, we separately feed each clip into a sign encoder to extract its feature. The final sign-video feature is yielded by stacking features coming out of \\( M \\) clips along temporal dimension. A powerful sign encoder is crucial." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.871, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Overview of Sign Encoder. Recent advances in sign spotting [47, 61] greatly facilitate the collection of large-scale" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.683, + 0.892, + 0.895 + ], + "angle": 0, + "content": "sign language datasets, enabling powerful representation learning abilities of convolutional neural networks on the sign classification task. Previous methods [16, 18] have demonstrated the feasibility of transferring a sign encoder pre-trained on large-scale sign-spotting data into downstream tasks. We follow this practice and use an I3D network pre-trained on BSL-1K [61], a sign classification dataset collected via sign spotting, as our primary sign encoder. Due to its favorable transfer performance, we term this model as a domain-agnostic sign encoder. Nevertheless, the domain gap between BSL-1K and sign language retrieval datasets is non-negligible. To tackle this problem, we further fine-tune a domain-aware sign encoder, which has an identical architecture to the domain-agnostic sign en" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19019" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.152 + ], + "angle": 0, + "content": "coder, on target datasets through pseudo-labeling. The final sign encoder is composed of the well-optimized domain-aware sign encoder and the powerful domain-agnostic sign encoder, as illustrated in Figure 3a." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.153, + 0.471, + 0.365 + ], + "angle": 0, + "content": "Pseudo-Labeling on Target Datasets. Now we describe the details of pseudo-labeling. Given a sign video from a target dataset, we adopt a sliding window with the stride of 1 and the window size of 16 to generate a set of temporally overlapped clips. For each clip, we first utilize the pretrained domain-agnostic sign encoder to produce its prediction. Then we binarize the prediction with a pre-defined threshold \\(\\lambda\\) to generate the corresponding pseudo label. The invalid samples, whose maximum score is lower than \\(\\lambda\\), are filtered. We repeat the above process for all sign videos and eventually build a pseudo-labeled set. Our domain-aware sign encoder, which is initialized by the domain-agnostic sign encoder, is fine-tuned on the pseudo-labeled set via a standard cross-entropy loss for classification training." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.365, + 0.471, + 0.591 + ], + "angle": 0, + "content": "Feature Extraction with Sign Encoder. So far, we acquire a domain-aware sign encoder approximately aligned in target domain. Nevertheless, its capability is restricted by the unavoidable noises in pseudo-labels and the limited amount of pseudo-labeled samples. Recall that we already have a powerful domain-agnostic sign encoder pre-trained on large-scale dataset in hand, inspiring us to make use of both domain-agnostic sign encoder \\( h_{\\xi}(\\cdot) \\) and domain-aware sign encoder \\( h_{\\theta}(\\cdot) \\) to extract discriminative and domain-aligned features. Our final sign encoder \\( H(\\cdot) \\), as shown in Figure 3a, is a weighted combination of \\( h_{\\xi}(\\cdot) \\) and \\( h_{\\theta}(\\cdot) \\) with a trade-off hyper-parameter \\( \\alpha \\). As described above, \\( H(\\cdot) \\) encodes sign videos in a sliding window manner. For simplicity, we use \\( H(v) \\) to denote feature extraction on sign video \\( v \\), which is formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.161, + 0.605, + 0.469, + 0.622 + ], + "angle": 0, + "content": "\\[\nH (v) = \\alpha h _ {\\xi} (v) + (1 - \\alpha) h _ {\\theta} (v). \\tag {1}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.635, + 0.394, + 0.651 + ], + "angle": 0, + "content": "3.3. Cross-Linguual Contrastive Learning" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.659, + 0.469, + 0.779 + ], + "angle": 0, + "content": "The objective of cross-lingual contrastive learning (CLCL) is to learn a joint embedding space of sign videos and texts while concurrently identifying the fine-grained sign-to-word mappings during training. An overview is shown in Figure 3b. CLCL takes a mini-batch \\(\\{(v_{n},t_{n})\\}_{n = 1}^{N}\\) containing \\(N\\) sign-video-text pairs as input, and contrasts paired data in a shared embedding space for sign language retrieval." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.78, + 0.47, + 0.901 + ], + "angle": 0, + "content": "Sign Features and Word Features. Given a sign video \\( v \\in \\{v_n\\}_{n=1}^N \\), we first adopt our sign encoder \\( H(\\cdot) \\) described in Section 3.2 to extract its intermediate feature. Note \\( H(\\cdot) \\) encodes sign videos in a sliding window manner, and thus there are no interactions among different clips. To facilitate information exchange, we further append a 12-layer Transformer [62] \\( F(\\cdot) \\) onto \\( H(\\cdot) \\) to extract sign features \\( S \\) of sign video \\( v \\), which is formulated" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.091, + 0.892, + 0.198 + ], + "angle": 0, + "content": "as \\( S = F(H(v)) \\in \\mathbb{R}^{M \\times D} \\), where \\( M \\) denotes the number of clips, and \\( D \\) is the hidden dimension. Given a text \\( t \\in \\{t_n\\}_{n=1}^N \\), we convert \\( t \\) into a lower-cased byte pair encoding (BPE) representation [53], which is subsequently fed into another 12-layer Transformer \\( G(\\cdot) \\) to generate the word features \\( W = G(t) \\in \\mathbb{R}^{L \\times D} \\), where \\( L \\) represents word number." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.2, + 0.892, + 0.307 + ], + "angle": 0, + "content": "Since CLIP [51] shows excellent transfer capability in various downstream tasks [17, 22, 27, 54, 66, 67], we initialize \\( F(\\cdot) \\) and \\( G(\\cdot) \\) with CLIP's image encoder (ViT-B) and text encoder, respectively, to ease the learning. Though CLIP's vision encoder takes image patches as inputs, we experimentally find that it generalizes well in our scenario where input data is in a different modality." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.31, + 0.892, + 0.463 + ], + "angle": 0, + "content": "Cross-Lingual Similarity. There exist inherent sign-to-word mappings between sign languages and natural languages. To incorporate this prior knowledge into learning, we introduce cross-lingual similarity—an indicator to identify sign-to-word mappings between \\(i\\)-th sign video \\(v_{i}\\) and \\(j\\)-th text \\(t_j\\). Concretely, given sign features \\(\\pmb{S}_i \\in \\mathbb{R}^{M \\times D}\\) of \\(v_i\\), and word features \\(\\pmb{W}_j \\in \\mathbb{R}^{L \\times D}\\) of \\(t_j\\), we calculate a cross-lingual similarity matrix \\(\\pmb{E}_{(i,j)} = \\pmb{S}_i \\cdot \\pmb{W}_j^T \\in \\mathbb{R}^{M \\times L}\\). Each element in \\(\\pmb{E}_{(i,j)}\\) represents the similarity of a sign clip in \\(v_i\\) and a word in \\(t_j\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.465, + 0.892, + 0.588 + ], + "angle": 0, + "content": "Cross-Lingual Contrastive Learning. Directly apply supervisions on token-wise similarity matrix \\( \\pmb{E}_{(i,j)} \\) is infeasible due to the absence of fine-grained sign-to-word annotations. Inspired by the recent progress of vision-language contrastive learning [23, 25, 29, 39, 40, 51], we turn to contrast the global representations of sign videos and texts. The underlying idea is to calculate a global similarity \\( z \\) of \\( v_{i} \\) and \\( t_j \\) based on \\( \\pmb{E}_{(i,j)} \\in \\mathbb{R}^{M \\times L} \\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.589, + 0.892, + 0.773 + ], + "angle": 0, + "content": "Sign-Video-to-Text Contrast. We first introduce sign-video-to-text contrast as shown in Figure 3b. To be specific, we utilize a Softmax operation to each row of \\( \\pmb{E}_{(i,j)} \\) and multiply the resulting matrix with \\( \\pmb{E}_{(i,j)} \\) to generate a sign-to-word similarity matrix \\( \\pmb{E}_{(i,j)}' \\in \\mathbb{R}^{M \\times L} \\), where each row represents the re-weighted similarities between a sign clip in \\( v_{i} \\) and all words in \\( t_j \\). After that, we adopt a row-wise addition operation on \\( \\pmb{E}_{(i,j)}' \\) to yield the sign-to-text similarity vector \\( e_{(i,j)} \\in \\mathbb{R}^M \\), where each element denotes a similarity of a sign clip in \\( v_{i} \\) and whole text \\( t_j \\). At last, we average all elements in \\( e_{(i,j)} \\) to produce the global similarity \\( z \\) of sign video \\( v_{i} \\) and text \\( t_j \\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.776, + 0.892, + 0.901 + ], + "angle": 0, + "content": "In the same way, we can calculate the similarities for both positive pairs \\(\\{(v_i,t_i)\\}_{i = 1}^N\\) and negative pairs \\(\\{(v_i,t_j)\\}_{i = 1,j = 1,i\\neq j}^N\\) in a mini-batch, yielding a video-to-text similarity matrix \\(Z_{V2T}\\in \\mathbb{R}^{N\\times N}\\), where \\(Z_{V2T}^{(i,j)}\\) denotes the global similarity of \\(v_{i}\\) and \\(t_j\\). Following CLIP [51], we adopt InfoNCE loss [26] to pull the embeddings of matched image-text pairs together while pushing those of non-matched pairs apart, which is formulated as" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19020" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.078, + 0.092, + 0.135, + 0.105 + ], + "angle": 0, + "content": "follows:" + }, + { + "type": "equation", + "bbox": [ + 0.121, + 0.116, + 0.47, + 0.204 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} _ {V 2 T} = - \\frac {1}{2 N} \\sum_ {\\substack {i = 1 \\\\ N}} ^ {N} \\log \\frac {\\exp \\left(\\boldsymbol {Z} _ {V 2 T} ^ {(i , j)} / \\tau\\right)}{\\sum_ {j = 1} ^ {N} \\exp \\left(\\boldsymbol {Z} _ {V 2 T} ^ {(i , j)} / \\tau\\right)} \\tag{2} \\\\ - \\frac {1}{2 N} \\sum_ {j = 1} ^ {N} \\log \\frac {\\exp \\left(\\mathbf {Z} _ {V 2 T} ^ {(i , j)} / \\tau\\right)}{\\sum_ {i = 1} ^ {N} \\exp \\left(\\mathbf {Z} _ {V 2 T} ^ {(i , j)} / \\tau\\right)}, \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.216, + 0.377, + 0.23 + ], + "angle": 0, + "content": "where \\(\\tau\\) is a trainable temperature parameter." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.231, + 0.469, + 0.395 + ], + "angle": 0, + "content": "Text-to-Sign-Video Contrast. Up to now, we have introduced sign-video-to-text contrast. A symmetrical version, termed text-to-sign-video contrast, shares the similar spirit as shown in Figure 3b. The implementation of text-to-sign-video contrast is extremely simple: we replace the row-wise operations (i.e., Softmax and addition) in sign-video-to-text contrast with column-wise ones and keep the remaining processes unchanged. We use \\(\\mathcal{L}_{T2V}\\) to denote the loss function of text-to-sign-video contrast. In our implementation, we reuse the loss defined in Eq 2 but substitute the input with the text-to-video similarity matrix \\(Z_{T2V}\\)." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.397, + 0.469, + 0.443 + ], + "angle": 0, + "content": "Loss Function. The overall loss for cross-lingual contrastive learning is a weighted sum of \\(\\mathcal{L}_{V2T}\\) and \\(\\mathcal{L}_{T2V}\\) with a trade-off hyper-parameter \\(\\beta\\):" + }, + { + "type": "equation", + "bbox": [ + 0.173, + 0.455, + 0.469, + 0.471 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\beta \\mathcal {L} _ {V 2 T} + (1 - \\beta) \\mathcal {L} _ {T 2 V}. \\tag {3}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.482, + 0.262, + 0.499 + ], + "angle": 0, + "content": "3.4. Text Augmentation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.506, + 0.47, + 0.809 + ], + "angle": 0, + "content": "Considering that the datasets of sign language retrieval are typically small-scale, we explore text augmentations to improve the generalization of our approach. EDA [64] introduces three simple yet efficient data augmentations in text classification task: random delete randomly removes words in a sentence; synonym replacement randomly selects words from a sentence that are not stop words and replaces them with synonyms; random swap randomly chooses two words in a sentence and swaps their positions. The first two augmentations have been proven effective in text classification task. However, we experimentally find that our focused sign language retrieval is sensitive to random delete and synonym replacement augmentations. To guarantee that the augmented texts preserve the original semantic meanings, we only adopt the random swap augmentation in our approach. We suppose there are two reasons: 1) the word order of sign languages and natural languages are constitutionally distinct, and reordering does not affect semantic meanings; 2) the proposed cross-lingual contrastive learning is insensitive to word order." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.822, + 0.202, + 0.839 + ], + "angle": 0, + "content": "4. Experiment" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.847, + 0.397, + 0.864 + ], + "angle": 0, + "content": "4.1. Datasets and Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.871, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Datasets. We primarily focus on How2Sign [19] dataset. Our model is also evaluated on PHOENIX-2014T [8] and" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.092, + 0.892, + 0.122 + ], + "angle": 0, + "content": "CSL-Daily [71], which are primarily used for sign language recognition and translation in previous works." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.124, + 0.892, + 0.304 + ], + "angle": 0, + "content": "How2Sign is a large-scale continuous American sign language (ASL) dataset consisting of a parallel corpus of about 80 hours of sign videos with subtitle annotations. It covers a wide range of instructional videos corresponding to various categories. There are 31164, 1740 and 2356 sign-video-text pairs in training, validation and test sets, respectively. Following SPOT-ALIGN [18], we remove the invalid pairs where the subtitle alignment is detected to exceed the video duration, remaining 31085, 1739 and 2348 available pairs in training, validation and test sets, respectively. The resolution of sign videos is \\(1280 \\times 720\\), we crop the human bodies of signers with Faster R-CNN [52] to generate valid videos." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.307, + 0.892, + 0.383 + ], + "angle": 0, + "content": "PHOENIX-2014T is a German sign language (Deutsche Gebärddensprache, DGS) dataset collected in the domain of weather forecast from TV broadcast, consisting of 7096, 519 and 642 video-text pairs in training, validation and test sets, respectively." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.384, + 0.892, + 0.445 + ], + "angle": 0, + "content": "CSL-Daily is a recently released Chinese sign language (CSL) dataset. The topic of CSL-Daily revolves around people's daily lives, including 18401, 1077, 1176 parallel samples in training, validation and test set, respectively." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.447, + 0.892, + 0.567 + ], + "angle": 0, + "content": "Evaluation Metric. Following previous works [18, 43, 51, 57], retrieval performance is evaluated by recall at rank K (R@K, higher is better) and median rank (MedR, lower is better). We evaluate our approach on both text-to-sign-video (T2V) retrieval and sign-video-to-text (V2T) retrieval tasks. We report R@1, R@5, and R@10 in all experiments, and additionally report MedR when comparing with state-of-the-art approaches." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.569, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Implementation Details. The sign encoder takes videos of resolution of \\(256 \\times 256\\) as input. The domain-agnostic sign encoder is a I3D [11] network pre-trained on BSL-1K [61]. In pseudo label generation, we set the threshold \\(\\lambda\\) to 0.6 to filter samples with low-confidence. Non-maximum suppression (NMS) with a temporal window of 24 frames is utilized to remove the duplicates among the pseudo-labeled samples. A collection of approximate 64K pseudo-labeled samples covering a vocabulary of 1220 words is eventually generated. The domain-aware sign encoder is initialized with the domain-agnostic one and fine-tuned with a learning rate of \\(1 \\times 10^{-2}\\) and batch size of 4 for 15 epochs. In the training of cross-lingual contrastive learning, the vision transformer and text transformer are initialized by the image encoder and text encoder in CLIP (ViT-B/32) [51]. The maximum length of sign clip features and text features are set to 64 and 32, respectively. The model is fine-tuned with Adam optimizer [31] with batch size of 512. The initial learning rate is set to \\(1 \\times 10^{-5}\\), which is decreased with a cosine schedule following the CLIP [51]. We set \\(\\alpha = 0.8\\) in Eq. 1 and \\(\\beta = 0.5\\) in Eq.3. The languages of PHOENIX-2014T and CSL-Daily are German and Chinese" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "19021" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.088, + 0.09, + 0.46, + 0.171 + ], + "angle": 0, + "content": "
ModelT2VV2T
R@1↑R@5↑R@10↑MedR↓R@1↑R@5↑R@10↑MedR↓
SA-SR [18]18.932.136.562.011.627.432.569.0
SA-CM [18]24.340.746.516.017.940.146.914.0
SA-COMB [18]34.248.052.68.023.647.053.07.5
Ours56.669.974.71.051.664.870.11.0
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.183, + 0.47, + 0.211 + ], + "angle": 0, + "content": "Table 1. Comparison with the different variants of the pioneer SPOT-ALIGN (SA) [18] on How2Sign [19] dataset." + }, + { + "type": "table", + "bbox": [ + 0.088, + 0.227, + 0.46, + 0.309 + ], + "angle": 0, + "content": "
ModelT2VV2T
R@1↑R@5↑R@10↑MedR↓R@1↑R@5↑R@10↑MedR↓
Translation [10]30.253.163.44.528.852.060.856.1
SA-CM [18]48.676.584.62.050.378.484.41.0
SA-COMB [18]55.879.687.21.053.179.486.11.0
Ours69.586.692.11.070.288.092.81.0
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.32, + 0.47, + 0.349 + ], + "angle": 0, + "content": "Table 2. Comparison with the different variants of the pioneer SPOT-ALIGN (SA) [18] on PHOENIX2014T [8] dataset." + }, + { + "type": "table", + "bbox": [ + 0.11, + 0.364, + 0.438, + 0.41 + ], + "angle": 0, + "content": "
ModelT2VV2T
R@1↑R@5↑R@10↑MedR↓R@1↑R@5↑R@10↑MedR↓
Ours75.388.291.91.074.789.492.21.0
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.421, + 0.468, + 0.449 + ], + "angle": 0, + "content": "Table 3. We additionally provide a baseline for CSL-Daily [71] dataset." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.465, + 0.469, + 0.511 + ], + "angle": 0, + "content": "respectively. Since CLIP is trained on English corpus, to reuse CLIP's text encoder, we utilize Google translation to translate the texts of these two datasets into English." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.535, + 0.442, + 0.551 + ], + "angle": 0, + "content": "4.2. Comparison with State-of-the-art Methods" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.564, + 0.469, + 0.625 + ], + "angle": 0, + "content": "We compare our method with different variants of the pioneer, called SPOT-ALIGN [18], on How2Sign and PHOENIX-2014T. We also provide the results on CSL-Daily as a baseline." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.629, + 0.47, + 0.902 + ], + "angle": 0, + "content": "Table 1 and Table 2 show the comparisons between our approach and SPOT-ALIGN [18] on How2Sign and PHOENIX-2014T, respectively. SPOT-ALIGN builds the final combination (COMB) model by integrating its primary cross-modal (CM) model with an auxiliary retrieval model (sign recognition (SR) model for How2Sign and Translation [10] for PHOENIX-2014T). Our method outperforms the COMB model, which achieves best results in SPOT-ALIGN, by large margins, achieving \\(+22.4\\) T2V and \\(+28.0\\) V2T R@1 improvements on How2Sign, \\(+13.7\\) T2V and \\(+17.1\\) V2T R@1 improvements on PHOENIX-2014T. It is worth mentioning that the SPOT-ALIGN conducts three rounds of sign spotting [1, 47] and encoder training. In contrast, we simplify the training of sign encoder and only perform a single round of training on pseudo-labeled data. We also provide a baseline on CSL-Daily dataset as shown in Table 3, demonstrating that our model can be generalized to various sign languages." + }, + { + "type": "table", + "bbox": [ + 0.534, + 0.09, + 0.86, + 0.171 + ], + "angle": 0, + "content": "
MethodT2VV2T
R@1R@5R@10R@1R@5R@10
Baseline31.549.657.926.444.452.8
+SE33.351.959.830.948.856.4
+SE+CLCL54.067.171.950.063.268.1
+SE+CLCL+TA56.669.974.751.664.870.1
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.183, + 0.892, + 0.224 + ], + "angle": 0, + "content": "Table 4. Ablation study of each proposed component. SE: sign encoder; CLCL: cross-lingual contrastive learning; TA: text augmentation." + }, + { + "type": "table", + "bbox": [ + 0.517, + 0.232, + 0.876, + 0.315 + ], + "angle": 0, + "content": "
EncoderT2VV2T
R@1R@5R@10R@1R@5R@10
Single-Ag [61]53.168.073.447.362.967.0
Single-Aw54.167.573.149.161.867.4
Fusion-Average54.768.773.849.663.769.0
Fusion-Weighted Sum56.669.974.751.664.870.1
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.326, + 0.892, + 0.381 + ], + "angle": 0, + "content": "Table 5. Results of domain-agnostic sign encoder (Single-Ag) and domain-aware sign encoder (Single-Aw). We also study different ways to integrate the features extracted by Single-Ag and Single-Aw, including average and weighted sum." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.391, + 0.654, + 0.408 + ], + "angle": 0, + "content": "4.3. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.416, + 0.892, + 0.445 + ], + "angle": 0, + "content": "We conduct all ablation studies on the most challenging How2Sign dataset." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.447, + 0.893, + 0.673 + ], + "angle": 0, + "content": "The Effectiveness of Each Proposed Component. Table 4 shows the ablation of each component. We first build a baseline where the sign encoder is the domain-agnostic one and contrastive learning is trained with the standard contrastive loss [51]. Then we add the proposed sign encoder (SE), cross-lingual contrastive learning (CLCL) and text augmentation (TA) to the baseline step by step. SE encodes domain-relevant and discriminative features, yielding \\(+1.8\\) T2V and \\(+4.5\\) V2T R@1 gains. The proposed CLCL significantly boosts the performance by \\(+20.7\\) T2V and \\(+19.1\\) V2T R@1 improvements, demonstrating that identifying fine-grained sign-to-word mappings is essential for sign language retrieval. The introduction of TA further promotes the retrieval task, achieving \\(56.6\\) R@1 on V2T and \\(51.6\\) R@1 on T2V, respectively." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.674, + 0.892, + 0.809 + ], + "angle": 0, + "content": "Various Sign Encoders. As described in Section 3.2, our sign encoder is composed of a domain-agnostic sign encoder (Single-Ag) [61] and a domain-aware sign encoder (Single-Aw). We first report the results of each individual encoder in Table 5. Next, we study the different ways to integrate the features extracted by these two encoders, including average and weighted sum as defined in Eq. 1. The results are also shown in Table 5. We experimentally find that the weighted sum strategy yields best results." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.811, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Variants of Cross-Lingual Contrastive Learning. As described in Section 3.3 and illustrated in Figure 3b, through the use of a combination of softmax, multiplication, sum, and average operations, we convert the fine-grained cross-lingual (sign-to-word) similarity to the coarse-grained sign-video-to-text similarity to enable contrastive learning. We" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19022" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.128, + 0.09, + 0.421, + 0.155 + ], + "angle": 0, + "content": "
StrategyT2VV2T
R@1R@5R@10R@1R@5R@10
Mean33.152.559.729.847.855.3
Max42.259.766.038.555.162.0
Softmax56.669.974.751.664.870.1
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.167, + 0.47, + 0.195 + ], + "angle": 0, + "content": "Table 6. Study on different strategies to identify the fine-grained sign-to-word mappings in cross-lingual contrastive learning." + }, + { + "type": "table", + "bbox": [ + 0.117, + 0.204, + 0.432, + 0.28 + ], + "angle": 0, + "content": "
Text AugmentationT2VV2T
R@1R@5R@10R@1R@5R@10
None54.067.171.950.063.268.1
RD52.766.371.647.662.167.4
SR53.968.773.047.561.265.9
RS56.669.974.751.664.870.1
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.291, + 0.468, + 0.32 + ], + "angle": 0, + "content": "Table 7. Ablation study on different text augmentations. RD: random delete; SR: synonym replacement; RS: random swap." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.34, + 0.468, + 0.597 + ], + "angle": 0, + "content": "refer to this process as \"Softmax\". Here we study two variants termed \"Mean\" and \"Max\". The \"Mean\" and \"Max\" strategies simply replace the combination of softmax, multiplication and sum operations with a simple mean operation and a max operation, respectively. The results are shown in Table 6. We observe that the \"Mean\" strategy performs worst since it merely evaluates the overall similarity of a text and a sign video and ignores the fine-grained sign-to-word mappings during training. The \"Max\" strategy identifies hard sign-to-word mappings, i.e., one sign is associated with the most similar word, and vice versa. Nevertheless, since we do not have the ground-truth of sign-to-word mappings, it is challenging for models to identify the confident one-to-one mappings in a weakly supervised manner, as discussed in previous works on multiple instance learning [47]. In contrast, our default \"Softmax\" strategy localizes the soft sign-to-word mappings and achieves the best results." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.598, + 0.468, + 0.763 + ], + "angle": 0, + "content": "Study of Text Augmentations. In Table 7, we study three text augmentations described in Section 3.4: random delete (RD), synonym replacement (SR) and random swap (RS). We observe a slight performance drop when introducing RD and SR into training. Sign language retrieval is not only a video-text retrieval task, but also a cross-lingual retrieval challenge, deletion and replacement may break the intrinsic sign-to-word mappings. In contrast, RS augmentation preserves the semantics of texts and we observe a +2.6 T2V and a +1.6 V2T R@1 improvements over the counterpart without any text augmentations." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.468, + 0.901 + ], + "angle": 0, + "content": "The Effects of CLIP Initialization. In our framework, the vision and text Transformer are initialized with the CLIP's vision encoder and text encoder. Note that the input modality of our vision Transformer and that of CLIP's vision encoder are different. Ours takes a feature sequence as input while the input of CLIP's image encoder is a set of image patches. In Table 8, we compare a randomly initialized baseline with the one initialized by CLIP. We find that CLIP-initialization significantly improves the perfor" + }, + { + "type": "table", + "bbox": [ + 0.543, + 0.09, + 0.85, + 0.145 + ], + "angle": 0, + "content": "
InitializationT2VV2T
R@1R@5R@10R@1R@5R@10
Random45.260.367.540.654.559.7
CLIP56.669.974.751.664.870.1
" + }, + { + "type": "table_caption", + "bbox": [ + 0.5, + 0.156, + 0.89, + 0.184 + ], + "angle": 0, + "content": "Table 8. Comparison between CLIP initialization and random initialization." + }, + { + "type": "table", + "bbox": [ + 0.541, + 0.193, + 0.852, + 0.285 + ], + "angle": 0, + "content": "
Frozen layersT2VV2T
R@1R@5R@10R@1R@5R@10
None56.669.974.751.664.870.1
154.468.273.449.763.568.8
253.868.673.348.562.867.9
651.766.671.146.260.966.6
1249.565.271.444.759.765.2
" + }, + { + "type": "table_caption", + "bbox": [ + 0.51, + 0.296, + 0.88, + 0.31 + ], + "angle": 0, + "content": "Table 9. Study of freezing different layers of text Transformer." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.335, + 0.892, + 0.38 + ], + "angle": 0, + "content": "mance, yielding \\(+11.4\\) T2V and \\(+11.0\\) V2T R@1 improvements though the input originates from a completely distinct modality." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.383, + 0.892, + 0.547 + ], + "angle": 0, + "content": "Study on Text Transformer. The text Transformer in our model is initialized with the CLIP's text encoder. Considering the generalization capability of the pre-trained CLIP, we attempt to freeze the shallow layers of our text transformer. As shown in Table 9, the performance gradually decreases as the number of the frozen layers increases. CLIP is trained on large-scale image-text pairs, however, the domain gap between image-text data and sign-video-text data is non-negligible. Though CLIP shows promising transfer capacity, it is optimal to fine-tune the whole model on target datasets." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.569, + 0.618, + 0.584 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.597, + 0.892, + 0.807 + ], + "angle": 0, + "content": "In this paper, we propose a novel framework named Cross-lingual Contrastive learning (CiCo) for recently introduced sign language retrieval task. We formulate sign language retrieval as a cross-lingual retrieval task as well as a video-text retrieval problem. CiCo models the fine-grained cross-lingual mappings between sign videos and texts via the proposed cross-lingual contrastive learning. We also introduce a sign encoder, which is composed of a domain-agnostic encoder and a domain-aware one, to extract discriminative and domain-aligned features. Our Cico outperforms the pioneer SPOT-ALIGN by large margins on How2Sign and PHOENIX-2014T benchmarks. We also provide a baseline on CSL-Daily. We hope our approach could serve as a solid baseline for future research." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.81, + 0.892, + 0.9 + ], + "angle": 0, + "content": "Acknowledgement. This work was partially supported by National Natural Science Foundation of China (No. 62072112), National Key R&D Program of China (No. 2020AAA0108301), and Scientific and Technological Innovation Action Plan of Shanghai Science and Technology Committee (No. 22511101502, 22511102202)." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "19023" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.116, + 0.47, + 0.186 + ], + "angle": 0, + "content": "[1] Samuel Albanie, Gül Varol, Liliane Momeni, Triantafyllos Afouras, Joon Son Chung, Neil Fox, and Andrew Zisserman. Bsl-1k: Scaling up co-articulated sign language recognition using mouthing cues. In European Conference on Computer Vision, pages 35–53. Springer, 2020. 1, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.187, + 0.472, + 0.243 + ], + "angle": 0, + "content": "[2] Mikel Artetxe, Gorka Labaka, and Eneko Agirre. Learning bilingual word embeddings with (almost) no bilingual data. In *Meeting of the Association for Computational Linguistics*, pages 451–462, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.245, + 0.471, + 0.3 + ], + "angle": 0, + "content": "[3] Mikel Artetxe, Gorka Labaka, and Eneko Agirre. A robust self-learning method for fully unsupervised cross-lingual mappings of word embeddings. arXiv preprint arXiv:1805.06297, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.302, + 0.471, + 0.4 + ], + "angle": 0, + "content": "[4] Vassilis Athitsos, Carol Neidle, Stan Sclaroff, Joan Nash, Alexandra Stefan, Ashwin Thangali, Hajjing Wang, and Quan Yuan. Large lexicon project: American sign language video corpus and sign language indexing/retrieval algorithms. In Workshop on the Representation and Processing of Sign Languages: Corpora and Sign Language Technologies, volume 2, pages 11-14, 2010. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.402, + 0.471, + 0.457 + ], + "angle": 0, + "content": "[5] Max Bain, Arsha Nagrani, Gül Varol, and Andrew Zisserman. Frozen in time: A joint video and image encoder for end-to-end retrieval. In IEEE/CVF International Conference on Computer Vision, pages 1728-1738, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.459, + 0.47, + 0.514 + ], + "angle": 0, + "content": "[6] Lisa Ballesteros and Bruce Croft. Dictionary methods for cross-lingual information retrieval. In International Conference on Database and Expert Systems Applications, pages 791–801. Springer, 1996. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.517, + 0.47, + 0.572 + ], + "angle": 0, + "content": "[7] Yoshua Bengio, Aaron Courville, and Pascal Vincent. Representation learning: A review and new perspectives. IEEE Transactions on Pattern Analysis and Machine Intelligence, 35(8):1798-1828, 2013. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.574, + 0.47, + 0.63 + ], + "angle": 0, + "content": "[8] Necati Cihan Camgoz, Simon Hadfield, Oscar Koller, Hermann Ney, and Richard Bowden. Neural sign language translation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7784-7793, 2018. 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.632, + 0.47, + 0.699 + ], + "angle": 0, + "content": "[9] Necati Cihan Camgoz, Oscar Koller, Simon Hadfield, and Richard Bowden. Multi-channel transformers for multi-articulatory sign language translation. In European Conference on Computer Vision, pages 301-319. Springer, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.703, + 0.47, + 0.773 + ], + "angle": 0, + "content": "[10] Necati Cihan Camgoz, Oscar Koller, Simon Hadfield, and Richard Bowden. Sign language transformers: Joint end-to-end sign language recognition and translation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10023-10033, 2020. 1, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.775, + 0.47, + 0.83 + ], + "angle": 0, + "content": "[11] Joao Carreira and Andrew Zisserman. Quo vadis, action recognition? a new model and the kinetics dataset. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6299-6308, 2017. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.832, + 0.47, + 0.9 + ], + "angle": 0, + "content": "[12] Yutong Chen, Fangyun Wei, Xiao Sun, Zhirong Wu, and Stephen Lin. A simple multi-modality transfer learning baseline for sign language translation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5120-5130, 2022. 1, 3" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.116, + 0.472, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.162 + ], + "angle": 0, + "content": "[13] Yutong Chen, Ronglai Zuo, Fangyun Wei, Yu Wu, Shujie LIU, and Brian Mak. Two-stream network for sign language recognition and translation. In Alice H. Oh, Alekh Agarwal, Danielle Belgrave, and Kyunghyun Cho, editors, Advances in Neural Information Processing Systems, 2022. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.163, + 0.892, + 0.232 + ], + "angle": 0, + "content": "[14] Necati Cihan Camgoz, Simon Hadfield, Oscar Koller, and Richard Bowden. Subunets: End-to-end hand shape and continuous sign language recognition. In IEEE/CVF International Conference on Computer Vision, pages 3056-3065, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.233, + 0.892, + 0.287 + ], + "angle": 0, + "content": "[15] Runpeng Cui, Hu Liu, and Changshui Zhang. A deep neural framework for continuous sign language recognition by iterative training. IEEE Transactions on Multimedia, 21(7):1880-1891, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.288, + 0.892, + 0.357 + ], + "angle": 0, + "content": "[16] Tonni Das Jui, Gissella Maria Bejarano, and Pablo Rivas. A machine learning-based segmentation approach for measuring similarity between sign languages. In sign-lang@ LREC 2022, pages 94-101. European Language Resources Association (ELRA), 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.358, + 0.892, + 0.427 + ], + "angle": 0, + "content": "[17] Yu Du, Fangyun Wei, Zihe Zhang, Miaojing Shi, Yue Gao, and Guoqi Li. Learning to prompt for open-vocabulary object detection with vision-language model. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14084-14093, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.428, + 0.892, + 0.495 + ], + "angle": 0, + "content": "[18] Amanda Duarte, Samuel Albanie, Xavier Giró-i Nieto, and Gül Varol. Sign language video retrieval with free-form textual queries. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14094-14104, 2022. 1, 2, 3, 4, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.497, + 0.892, + 0.579 + ], + "angle": 0, + "content": "[19] Amanda Duarte, Shruti Palaskar, Lucas Ventura, Deepti Ghadiyaram, Kenneth DeHaan, Florian Metze, Jordi Torres, and Xavier Giro-i Nieto. How2Sign: A Large-scale Multimodal Dataset for Continuous American Sign Language. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2021. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.581, + 0.892, + 0.622 + ], + "angle": 0, + "content": "[20] Ali Farhadi, David Forsyth, and Ryan White. Transfer learning in sign language. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1-8. IEEE, 2007. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.623, + 0.892, + 0.678 + ], + "angle": 0, + "content": "[21] Holger Fillbrandt, Suat Akyol, and K-F Kraiss. Extraction of 3d hand shape and posture from image sequences for sign language recognition. In 2003 IEEE International SOI Conference, pages 181-186. IEEE, 2003. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.679, + 0.892, + 0.733 + ], + "angle": 0, + "content": "[22] Stanislav Fort, Jie Ren, and Balaji Lakshminarayanan. Exploring the limits of out-of-distribution detection. Advances in Neural Information Processing Systems, 34:7068-7081, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.735, + 0.892, + 0.789 + ], + "angle": 0, + "content": "[23] Andrea Frome, Greg S Corrado, Jon Shlens, Samy Bengio, Jeff Dean, Marc'Aurelio Ranzato, and Tomas Mikolov. Devise: A deep visual-semantic embedding model. Advances in neural information processing systems, 26, 2013. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.79, + 0.892, + 0.845 + ], + "angle": 0, + "content": "[24] Valentin Gabeur, Chen Sun, Karteek Alahari, and Cordelia Schmid. Multi-modal transformer for video retrieval. In European Conference on Computer Vision, pages 214-229. Springer, 2020. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.846, + 0.894, + 0.901 + ], + "angle": 0, + "content": "[25] Yunchao Gong, Qifa Ke, Michael Isard, and Svetlana Lazebnik. A multi-view embedding space for modeling internet images, tags, and their semantics. International Journal of Computer Vision, 106(2):210-233, 2014. 3, 5" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "19024" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.161 + ], + "angle": 0, + "content": "[26] Michael Gutmann and Aapo Hyvarinen. Noise-contrastive estimation: A new estimation principle for unnormalized statistical models. In International Conference on Artificial Intelligence and Statistics, pages 297–304. JMLR Workshop and Conference Proceedings, 2010. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.162, + 0.471, + 0.203 + ], + "angle": 0, + "content": "[27] Tony Huang, Jack Chu, and Fangyun Wei. Unsupervised prompt learning for vision-language models. arXiv preprint arXiv:2204.03649, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.204, + 0.471, + 0.285 + ], + "angle": 0, + "content": "[28] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In International Conference on Machine Learning, pages 4904-4916. PMLR, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.288, + 0.471, + 0.343 + ], + "angle": 0, + "content": "[29] Armand Joulin, Laurens van der Maaten, Allan Jabri, and Nicolas Vasilache. Learning visual features from large weakly supervised data. In European Conference on Computer Vision, pages 67-84. Springer, 2016. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.344, + 0.471, + 0.385 + ], + "angle": 0, + "content": "[30] Hamid Reza Vaezi Joze and Oscar Koller. Ms-asl: A large-scale data set and benchmark for understanding american sign language. arXiv preprint arXiv:1812.01053, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.386, + 0.471, + 0.455 + ], + "angle": 0, + "content": "[31] Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In Yoshua Bengio and Yann LeCun, editors, 3rd International Conference on Learning Representations, ICLR 2015, San Diego, CA, USA, May 7-9, 2015, Conference Track Proceedings, 2015. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.455, + 0.471, + 0.524 + ], + "angle": 0, + "content": "[32] Oscar Koller, Necati Cihan Camgoz, Hermann Ney, and Richard Bowden. Weakly supervised learning with multistream cnn-lstm-hmms to discover sequential parallelism in sign language videos. IEEE Transactions on Pattern Analysis and Machine Intelligence, 42(9):2306-2320, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.525, + 0.471, + 0.579 + ], + "angle": 0, + "content": "[33] Oscar Koller, Jens Forster, and Hermann Ney. Continuous sign language recognition: Towards large vocabulary statistical recognition systems handling multiple signers. Computer Vision and Image Understanding, 141:108-125, 2015. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.58, + 0.471, + 0.635 + ], + "angle": 0, + "content": "[34] Victor Lavrenko, Martin Choquette, and W Bruce Croft. Cross-lingual relevance models. In International ACM SIGIR Conference on Research and Development in Information Retrieval, pages 175–182, 2002. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.636, + 0.471, + 0.664 + ], + "angle": 0, + "content": "[35] Yann LeCun, Yoshua Bengio, and Geoffrey Hinton. Deep learning. Nature, 521(7553):436-444, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.665, + 0.471, + 0.733 + ], + "angle": 0, + "content": "[36] Dongxu Li, Cristian Rodriguez, Xin Yu, and Hongdong Li. Word-level deep sign language recognition from video: A new large-scale dataset and methods comparison. In IEEE/CVF winter conference on applications of computer vision, pages 1459-1469, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.734, + 0.471, + 0.803 + ], + "angle": 0, + "content": "[37] Dongxu Li, Chenchen Xu, Xin Yu, Kaihao Zhang, Benjamin Swift, Hanna Suominen, and Hongdong Li. Tspnet: Hierarchical feature learning via temporal semantic pyramid for sign language translation. Advances in Neural Information Processing Systems, 33:12034-12045, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.804, + 0.471, + 0.871 + ], + "angle": 0, + "content": "[38] Dongxu Li, Xin Yu, Chenchen Xu, Lars Petersson, and Hongdong Li. Transferring cross-domain knowledge for video sign language recognition. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6205-6214, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.872, + 0.471, + 0.901 + ], + "angle": 0, + "content": "[39] Junnan Li, Ramprasaath Selvaraju, Akhilesh Gotmare, Shafiq Joty, Caiming Xiong, and Steven Chu Hong Hoi." + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.471, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.892, + 0.134 + ], + "angle": 0, + "content": "Align before fuse: Vision and language representation learning with momentum distillation. Advances in Neural Information Processing Systems, 34:9694-9705, 2021. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.136, + 0.892, + 0.204 + ], + "angle": 0, + "content": "[40] Yangguang Li, Feng Liang, Lichen Zhao, Yufeng Cui, Wanli Ouyang, Jing Shao, Fengwei Yu, and Junjie Yan. Supervision exists everywhere: A data efficient contrastive language-image pre-training paradigm. arXiv preprint arXiv:2110.05208, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.206, + 0.892, + 0.261 + ], + "angle": 0, + "content": "[41] Ji Lin, Chuang Gan, and Song Han. Tsm: Temporal shift module for efficient video understanding. In IEEE/CVF International Conference on Computer Vision, pages 7083-7093, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.263, + 0.892, + 0.332 + ], + "angle": 0, + "content": "[42] Song Liu, Haoqi Fan, Shengsheng Qian, Yiru Chen, Wenkui Ding, and Zhongyuan Wang. Hit: Hierarchical transformer with momentum contrast for video-text retrieval. In IEEE/CVF International Conference on Computer Vision, pages 11915-11925, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.334, + 0.892, + 0.389 + ], + "angle": 0, + "content": "[43] Yang Liu, Samuel Albanie, Arsha Nagrani, and Andrew Zisserman. Use what you have: Video retrieval using representations from collaborative experts. arXiv preprint arXiv:1907.13487, 2019. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.391, + 0.892, + 0.446 + ], + "angle": 0, + "content": "[44] Huaishao Luo, Lei Ji, Ming Zhong, Yang Chen, Wen Lei, Nan Duan, and Tianrui Li. Clip4clip: An empirical study of clip for end to end video clip retrieval and captioning. Neurocomputing, 508:293-304, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.448, + 0.892, + 0.53 + ], + "angle": 0, + "content": "[45] Antoine Miech, Dimitri Zhukov, Jean-Baptiste Alayrac, Makarand Tapaswi, Ivan Laptev, and Josef Sivic. Howto100m: Learning a text-video embedding by watching hundred million narrated video clips. In IEEE/CVF International Conference on Computer Vision, pages 2630-2640, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.532, + 0.892, + 0.574 + ], + "angle": 0, + "content": "[46] Tomas Mikolov, Quoc V Le, and Ilya Sutskever. Exploiting similarities among languages for machine translation. arXiv preprint arXiv:1309.4168, 2013. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.576, + 0.892, + 0.631 + ], + "angle": 0, + "content": "[47] Liliane Momeni, Gül Varol, Samuel Albanie, Triantafyllos Afouras, and Andrew Zisserman. Watch, read and lookup: learning to spot signs from multiple supervisors. In Asian Conference on Computer Vision, 2020. 3, 4, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.633, + 0.892, + 0.715 + ], + "angle": 0, + "content": "[48] Mathew Monfort, SouYoung Jin, Alexander Liu, David Harwath, Rogerio Feris, James Glass, and Aude Oliva. Spoken moments: Learning joint audio-visual representations from video descriptions. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14871-14881, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.717, + 0.892, + 0.786 + ], + "angle": 0, + "content": "[49] Yasuhide Mori, Hironobu Takahashi, and Ryuichi Oka. Image-to-word transformation based on dividing and vector quantizing images with words. In International Workshop on Multimedia Intelligent Storage and Retrieval Management, pages 1-9. Citeseer, 1999. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.788, + 0.892, + 0.843 + ], + "angle": 0, + "content": "[50] Aitor Ormazabal, Mikel Artetxe, Gorka Labaka, Aitor Soroa, and Eneko Agirre. Analyzing the limitations of cross-lingual word embedding mappings. arXiv preprint arXiv:1906.05407, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.845, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[51] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language super" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "19025" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.092, + 0.468, + 0.12 + ], + "angle": 0, + "content": "vision. In International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 3, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.122, + 0.469, + 0.177 + ], + "angle": 0, + "content": "[52] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. Conference on Neural Information Processing Systems, 28, 2015. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.179, + 0.469, + 0.22 + ], + "angle": 0, + "content": "[53] Rico Sennrich, Barry Haddow, and Alexandra Birch. Neural machine translation of rare words with subword units. arXiv preprint arXiv:1508.07909, 2015. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.222, + 0.469, + 0.277 + ], + "angle": 0, + "content": "[54] Sheng Shen, Liunian Harold Li, Hao Tan, Mohit Bansal, Anna Rohrbach, Kai-Wei Chang, Zhewei Yao, and Kurt Keutzer. How much can clip benefit vision-and-language tasks? arXiv preprint arXiv:2107.06383, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.279, + 0.469, + 0.333 + ], + "angle": 0, + "content": "[55] Thad Starner, Joshua Weaver, and Alex Pentland. Real-time american sign language recognition using desk and wearable computer based video. IEEE Transactions on Pattern Analysis and Machine Intelligence, 20(12):1371-1375, 1998. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.335, + 0.469, + 0.39 + ], + "angle": 0, + "content": "[56] Thad E Starner. Visual recognition of american sign language using hidden markov models. Technical report, Massachusetts inst of tech Cambridge dept of brain and cognitive sciences, 1995. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.392, + 0.469, + 0.46 + ], + "angle": 0, + "content": "[57] Chen Sun, Austin Myers, Carl Vondrick, Kevin Murphy, and Cordelia Schmid. Videobert: A joint model for video and language representation learning. In IEEE/CVF International Conference on Computer Vision, pages 7464-7473, 2019. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.463, + 0.469, + 0.504 + ], + "angle": 0, + "content": "[58] Rachel Sutton-Spence. Mouthings and simultaneity in British sign language. Amsterdam Studies in the Theory and History of Linguistic Science Series, 281:147, 2007. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.507, + 0.469, + 0.56 + ], + "angle": 0, + "content": "[59] Chau Tran, Yuqing Tang, Xian Li, and Jiatao Gu. Crosslingual retrieval for iterative self-supervised training. Advances in Neural Information Processing Systems, 33:2207-2219, 2020. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.563, + 0.469, + 0.631 + ], + "angle": 0, + "content": "[60] Du Tran, Heng Wang, Lorenzo Torresani, Jamie Ray, Yann LeCun, and Manohar Paluri. A closer look at spatiotemporal convolutions for action recognition. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6450-6459, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.633, + 0.469, + 0.701 + ], + "angle": 0, + "content": "[61] Gul Varol, Liliane Momeni, Samuel Albanie, Triantafyllos Afouras, and Andrew Zisserman. Read and attend: Temporal localisation in sign language videos. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16857-16866, 2021. 1, 2, 3, 4, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.704, + 0.469, + 0.746 + ], + "angle": 0, + "content": "[62] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.747, + 0.469, + 0.802 + ], + "angle": 0, + "content": "[63] Xiaolong Wang, Ross Girshick, Abhinav Gupta, and Kaiming He. Non-local neural networks. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7794-7803, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.804, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[64] Jason Wei and Kai Zou. EDA: Easy data augmentation techniques for boosting performance on text classification tasks. In 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages 6383-6389, Hong Kong, China, Nov. 2019. Association for Computational Linguistics. 6" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.469, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.16 + ], + "angle": 0, + "content": "[65] Jinxi Xu, Ralph Weischedel, and Chanh Nguyen. Evaluating a probabilistic model for cross-lingual information retrieval. In International ACM SIGIR Conference on Research and Development in Information Retrieval, pages 105–110, 2001. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.164, + 0.892, + 0.205 + ], + "angle": 0, + "content": "[66] Mengde Xu, Zheng Zhang, Fangyun Wei, Han Hu, and Xiang Bai. Side adapter network for open-vocabulary semantic segmentation. arXiv preprint arXiv:2302.12242, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.207, + 0.892, + 0.275 + ], + "angle": 0, + "content": "[67] Mengde Xu, Zheng Zhang, Fangyun Wei, Yutong Lin, Yue Cao, Han Hu, and Xiang Bai. A simple baseline for open-vocabulary semantic segmentation with pre-trained vision-language model. In European Conference on Computer Vision, pages 736-753. Springer, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.277, + 0.892, + 0.33 + ], + "angle": 0, + "content": "[68] Youngjae Yu, Jongseok Kim, and Gunhee Kim. A joint sequence fusion model for video question answering and retrieval. In European Conference on Computer Vision, pages 471-487, 2018. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.333, + 0.892, + 0.4 + ], + "angle": 0, + "content": "[69] Mozhi Zhang, Keyulu Xu, Ken-ichi Kawarabayashi, Stefanie Jegelka, and Jordan Boyd-Graber. Are girls neko or shōjo? cross-lingual alignment of non-isomorphic embeddings with iterative normalization. arXiv preprint arXiv:1906.01622, 2019.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.403, + 0.892, + 0.459 + ], + "angle": 0, + "content": "[70] Shilin Zhang and Bo Zhang. Using revised string edit distance to sign language video retrieval. In International Conference on Computational Intelligence and Natural Computing, volume 1, pages 45-49. IEEE, 2010. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.461, + 0.892, + 0.529 + ], + "angle": 0, + "content": "[71] Hao Zhou, Wengang Zhou, Weizhen Qi, Junfu Pu, and Houqiang Li. Improving sign language translation with monolingual data by sign back-translation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1316-1325, 2021. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.531, + 0.892, + 0.584 + ], + "angle": 0, + "content": "[72] Hao Zhou, Wengang Zhou, Yun Zhou, and Houqiang Li. Spatial-temporal multi-cue network for sign language recognition and translation. IEEE Transactions on Multimedia, 24:768-779, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.587, + 0.892, + 0.641 + ], + "angle": 0, + "content": "[73] Ronglai Zuo, Fangyun Wei, and Brian Mak. Natural language-assisted sign language recognition. IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023. 1, 3" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.641 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "19026" + } + ] +] \ No newline at end of file diff --git a/2023/CiCo_ Domain-Aware Sign Language Retrieval via Cross-Lingual Contrastive Learning/047059b8-c0b2-4bcc-9546-ab9b6e4bc0e5_origin.pdf b/2023/CiCo_ Domain-Aware Sign Language Retrieval via Cross-Lingual Contrastive Learning/047059b8-c0b2-4bcc-9546-ab9b6e4bc0e5_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..01011296188f3b1f5117f66bee958b87e4ffc3f2 --- /dev/null +++ b/2023/CiCo_ Domain-Aware Sign Language Retrieval via Cross-Lingual Contrastive Learning/047059b8-c0b2-4bcc-9546-ab9b6e4bc0e5_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:607a33c12758faa9ddd2a3985be98ff7c9defa8495810a20fed298de59c3d01f +size 1593681 diff --git a/2023/CiCo_ Domain-Aware Sign Language Retrieval via Cross-Lingual Contrastive Learning/full.md b/2023/CiCo_ Domain-Aware Sign Language Retrieval via Cross-Lingual Contrastive Learning/full.md new file mode 100644 index 0000000000000000000000000000000000000000..2cb10adcf66045f10a6119af92c70db3d45fa727 --- /dev/null +++ b/2023/CiCo_ Domain-Aware Sign Language Retrieval via Cross-Lingual Contrastive Learning/full.md @@ -0,0 +1,309 @@ +# CiCo: Domain-Aware Sign Language Retrieval via Cross-Lingual Contrastive Learning + +Yiting Cheng $^{1}$ , Fangyun Wei $^{2*}$ , Jianmin Bao $^{2}$ , Dong Chen $^{2}$ , Wenqiang Zhang $^{1*}$ $^{1}$ School of Computer Science, Fudan University, $^{2}$ Microsoft Research Asia +{ytcheng18, wqzhang}@fudan.edu.cn, {fawe, jianbao, doch}@microsoft.com + +# Abstract + +This work focuses on sign language retrieval—a recently proposed task for sign language understanding. Sign language retrieval consists of two sub-tasks: text-to-sign-video (T2V) retrieval and sign-video-to-text (V2T) retrieval. Different from traditional video-text retrieval, sign language videos, not only contain visual signals but also carry abundant semantic meanings by themselves due to the fact that sign languages are also natural languages. Considering this character, we formulate sign language retrieval as a cross-lingual retrieval problem as well as a video-text retrieval task. Concretely, we take into account the linguistic properties of both sign languages and natural languages, and simultaneously identify the fine-grained cross-lingual (i.e., sign-to-word) mappings while contrasting the texts and the sign videos in a joint embedding space. This process is termed as cross-lingual contrastive learning. Another challenge is raised by the data scarcity issue—sign language datasets are orders of magnitude smaller in scale than that of speech recognition. We alleviate this issue by adopting a domain-agnostic sign encoder pre-trained on large-scale sign videos into the target domain via pseudolabeling. Our framework, termed as domain-aware sign language retrieval via Cross-lingual Contrastive learning or CiCo for short, outperforms the pioneering method by large margins on various datasets, e.g., +22.4 T2V and +28.0 V2TR@1 improvements on How2Sign dataset, and +13.7 T2V and +17.1 V2TR@1 improvements on PHOENIX-2014T dataset. Code and models are available at: https://github.com/FangyunWei/SLRT. + +# 1. Introduction + +Sign languages are the primary means of communication used by people who are deaf or hard of hearing. Sign language understanding [1, 10, 12-15, 18, 32, 33, 61, 73] is significant for overcoming the communication barrier between + +![](images/01d2bcf4d4965d58c2397bc1dcf667b59d7f50fa64b981ec7ac2f54c7e5724a0.jpg) +(a) Text-to-sign-video (T2V) retrieval. + +![](images/ed8f51b1773276f9834cc368dd8f8e36453c53712644d3230dadfc7314559f1a.jpg) +(b) Sign-video-to-text (V2T) retrieval. +Figure 1. Illustration of: (a) T2V retrieval; (b) V2T retrieval. + +the hard-of-hearing and non-signers. Sign language recognition and translation (SLRT) has been extensively studied, with the goal of recognizing the arbitrary semantic meanings conveyed by sign languages. However, the lack of available data significantly limits the capability of SLRT. In this paper, we focus on developing a framework for a recently proposed sign language retrieval task [18]. Unlike SLRT, sign language retrieval focuses on retrieving the meanings that signers express from a closed-set, which can significantly reduce error rates in realistic deployment. + +Sign language retrieval is both similar to and distinct from the traditional video-text retrieval. On the one hand, like video-text retrieval, sign language retrieval is also composed of two sub-tasks, i.e., text-to-sign-video (T2V) retrieval and sign-video-to-text (V2T) retrieval. Given a free-form written query and a large collection of sign language videos, the objective of T2V is to find the video that best matches the written query (Figure 1a). In contrast, the goal of V2T is to identify the most relevant text description given a query of sign language video (Figure 1b). + +On the other hand, different from the video-text retrieval, sign languages, like most natural languages, have their own grammars and linguistic properties. Therefore, sign language videos not only contain visual signals, but also carry + +(Sign Language) +![](images/2351c5086303c77f8edced1b80e17be1ae97267ee1ade4ecd4ccbd38d0dd1498.jpg) +(a) While contrasting the sign videos and the texts in a joint embedding space, we simultaneously identify the fine-grained cross-lingual (sign-to-word) mappings of sign languages and natural languages via the proposed cross-lingual contrastive learning. Existing datasets do not annotate the sign-to-word mappings. + +Figure 2. Illustration of: (a) cross-lingual (sign-to-word) mapping; (b) sign-to-word mappings identified by our CiCo. +![](images/9ee6e99e7787c128052cf6936f6c4c66facc18ea1a046dff0fad964a778d1cf0.jpg) +(b) We show four instances of the sign "Book" in How2Sign [19] dataset, which are identified by our approach. Please refer to the supplementary material for more examples. + +![](images/fc11fdd1b44df7cb3628ab8f99db8c3b472c52356ddeae90c99d865bd42110a7.jpg) + +semantics (i.e., sign1-to-word mappings between sign languages and natural languages) by themselves, which differentiates them from the general videos that merely contain visual information. Considering the linguistic characteristics of sign languages, we formulate sign language retrieval as a cross-lingual retrieval [6,34,59] problem in addition to a video-text retrieval [5,24,42-44,57,68] task. + +Sign language retrieval is extremely challenging due to the following reasons: (1) Sign languages are completely separate and distinct from natural languages since they have unique linguistic rules, word formation, and word order. The transcription between sign languages and natural languages is complicated, for instance, the word order is typically not preserved between sign languages and natural languages. It is necessary to automatically identify the sign-to-word mapping from the cross-lingual retrieval perspective; (2) In contrast to the text-video retrieval datasets [45, 48] which contain millions of training samples, sign language datasets are orders of magnitude smaller in scale—for example, there are only $30\mathrm{K}$ video-text pairs in How2Sign [19] training set; (3) Sign languages convey information through the handshape, facial expression, and body movement, which requires models to distinguish fine + +grained gestures and actions; (4) Sign language videos typically contain hundreds of frames. It is necessary to build efficient algorithms to lower the training cost and fit the long videos as well as the intermediate representations into limited GPU memory. + +In this work, we concentrate on resolving the challenges listed above: + +- We consider the linguistic rules (e.g., word order) of both sign languages and natural languages. We formulate sign language retrieval as a cross-lingual retrieval task as well as a video-text retrieval problem. While contrasting the sign videos and the texts in a joint embedding space as achieved in most vision-language pre-training frameworks [5, 44, 57], we simultaneously identify the fine-grained cross-lingual (sign-to-word) mappings between two types of languages via our proposed cross-lingual contrastive learning as shown in Figure 2. +- Data scarcity typically brings in the over-fitting issue. To alleviate this issue, we adopt transfer learning and adapt a recently released domain-agnostic sign encoder [61] pre-trained on large-scale sign-videos to the target domain. Although this encoder is capable of distinguishing the fine-grained signs, direct transferring may be suboptimal due to the unavoidable domain gap between the pre-training dataset and sign language retrieval datasets. To tackle this problem, we further fine-tune a domain-aware sign encoder on pseudo-labeled data from target datasets. The final sign encoder is composed of the well-optimized domain-aware sign encoder and the powerful domain-agnostic sign encoder. +- In order to effectively model long videos, we decouple our framework into two disjoint parts: (1) a sign encoder which adopts a sliding window on sign-videos to pre-extract their vision features; (2) a cross-lingual contrastive learning module which encodes the extracted vision features and their corresponding texts in a joint embedding space. + +Our framework, called domain-aware sign language retrieval via Cross-lingual Contrastive learning or CiCo for short, outperforms the pioneer SPOT-ALIGN [18] by large margins on various datasets, achieving 56.6 (+22.4) T2V and 51.6 (+28.0) V2T R@1 accuracy (improvement) on How2Sign [19] dataset, and 69.5 (+13.7) T2V and 70.2 (+17.1) V2T R@1 accuracy (improvement) on PHOENIX-2014T [8] dataset. With its simplicity and strong performance, we hope our approach can serve as a solid baseline for future research. + +# 2. Related Work + +Sign Language Understanding. Sign language understanding aims at interpreting the semantic information con + +veyed within sign videos. Researchers have explored such capability on various tasks including sign language recognition [13,20,21,30,38,58,73], sign spotting [1,18,47,61], sign language translation [8,9,12,37,72] and our focused sign language retrieval [18]. + +One of the fundamental tasks of sign language understanding is sign language recognition (SLR), which aims to transcribe a sign video into a gloss sequence. Previous works on SLR focus on designing carefully engineered features [20, 21, 58] or modeling temporal dependencies [55, 56]. Recently, the success of 3D convolutional neural networks in action related tasks [41, 60, 63] is transferred to SLR. In particular, the I3D [11] architecture has proven to be effective for this task [1, 30, 36, 38, 61]. In this work, we also adopt this network architecture in our sign encoder. + +Sign spotting is a particular variant of sign language recognition. It aims to localize all instances of a given sign within an untrimmed video. Recent works tackle this task with auxiliary cue of subtitles, introducing automatic annotation systems by using mouthing [1], dictionaries [47] and attention maps of Transformer [61]. SPOT-ALIGN [18] extends existing spotting methods [1, 47] with an iterative training schema, which alternates between repeated sign spotting and model fine-tuning. In this work, pseudo-labeling is served as our sign spotting approach to localize isolated signs in untrimmed videos from target sign language retrieval datasets. Compared with above efforts, our approach only employs a pre-trained sign encoder without utilizing additional auxiliary cue, which is proved to be simple yet efficient. + +Early works of sign language retrieval primarily investigate query-by-example searching [4, 70], which queries individual instances with given sign examples. Our work focuses on free-form textual retrieval—a recently introduced task by SPOT-ALIGN [18]. It symbolizes the real-world scenario of searching sign language videos with natural languages. The pioneer SPOT-ALIGN [18] purely formulates sign language retrieval as a video-text retrieval task, where the cross-modal alignment is modeled upon overall global embeddings of sign videos and texts. However, the linguistic properties of sign languages are ignored. In contrast, we formulate sign language retrieval as a joint task of text-video retrieval and cross-lingual retrieval. + +Vision-Language Models. Learning general-purpose representations for visual and textual modalities is a longstanding topic [7, 35]. The idea has been investigated decades ago [49]. Recently, the prominent success of CLIP [51], ALIGN [28] and ALBEF [39] has demonstrated the capability of learning joint cross-modal representations with large-scale web data using simple image-text contrastive learning. Similar idea has also been explored in video-text retrieval, which is the closest area of our work. + +The common practice in image-text retrieval [23, 25, 29, 51] and video-text retrieval [5, 24, 42-44, 57, 68] is to encode the images/videos and texts into the overall representations. It is reasonable since the visual signals of typical image-text and video-text retrieval datasets mainly describe the certain objects or describable events. In contrast, sign videos, as the carriers of sign languages, convey abundant semantics by themselves. There exist fine-grained mappings between sign videos and natural languages. We find that identifying such cross-lingual (sign-to-word) mappings significantly boosts the performance of sign language retrieval. + +Cross-Lingual Information Retrieval. In this work, we also formulate sign-language retrieval as a cross-lingual task. In natural language processing community, cross-lingual information retrieval (CLIR) [34, 59, 65] refers to the task of retrieving documents between different languages. One of the most common approaches of CLIR is to learn sentence-level embedding alignment by mapping pre-acquired monolingual embeddings [46, 50, 69] of different languages into a shared space. Recently, researchers have exploited fine-grained word-level mappings with self-training [2, 3, 59]. They find that mappings can be learned by initiating with a seed dictionary and alternating between alignment modeling and dictionary mining. Our work also exploits fine-grained word-level mappings, called sign-to-word mappings, in the context of sign language retrieval. Identifying sign-to-word mappings is challenging since sign languages are expressed in visual modality. The proposed cross-lingual contrastive learning tackles this challenge by exploiting the fine-grained cross-modal interactions. + +# 3. Methodology + +In this section, we first formulate the task of sign language retrieval in Section 3.1. Next, we introduce our framework which is composed of two parts: 1) a sign encoder which extracts discriminative and domain-aligned features of sign videos (Section 3.2); 2) a cross-lingual contrastive learning framework, which contrasts sign-video-text pairs while concurrently identifying the fine-grained sign-to-word mappings (Section 3.3). At last, we explore text augmentations in Section 3.4. + +# 3.1. Task Formulation + +Let $\mathcal{V}$ and $\mathcal{T}$ denote a set of sign videos and their corresponding texts (transcriptions), respectively. Sign language retrieval consists of two tasks namely text-to-sign-video retrieval (T2V), and sign-video-to-text retrieval (V2T), respectively. The objective of T2V is to find the sign video $v\in \mathcal{V}$ whose signing content best matches the text query. In contrast, the reverse task V2T requires the model to identify the most relevant text (transcription) $t\in \mathcal{T}$ given a query of sign video. We resolve sign language retrieval by learning a joint embedding space of sign videos and texts. + +![](images/b9c7644175c01d5251b0ebf06b002ac5650564f65aa85bae99313a5ceeff1e0e.jpg) +(a) Illustration of sign encoder. + +![](images/177a9fdd10ab5bc16f025d961267330e3e5b63ba7832acf7ea46d6dcf0990232.jpg) +(b) Illustration of cross-lingual contrastive learning. +Figure 3. Overview of our framework. (a) Sign encoder is composed of a powerful domain-agnostic sign encoder pre-trained on large-scale sign videos, and a domain-aware sign encoder fine-tuned on pseudo-labeled data from target datasets. We adopt a sliding window manner to extract a discriminative and domain-aligned feature per clip. (b) Cross-lingual contrastive learning takes $N$ sign-video-text pairs as inputs and contrasts paired data in a shared embedding space while implicitly identifying the fine-grained sign-to-word mappings during training. + +# 3.2. Sign Encoder + +Process Sign Videos with Sliding Window. Sign videos from sign language retrieval datasets typically contain hundreds of frames. To efficiently train our model and lower the usage of GPU memory, given a sign video, we adopt a sliding window manner with stride of 1 and window size of 16 to produce $M$ temporally overlapped clips. Next, we separately feed each clip into a sign encoder to extract its feature. The final sign-video feature is yielded by stacking features coming out of $M$ clips along temporal dimension. A powerful sign encoder is crucial. + +Overview of Sign Encoder. Recent advances in sign spotting [47, 61] greatly facilitate the collection of large-scale + +sign language datasets, enabling powerful representation learning abilities of convolutional neural networks on the sign classification task. Previous methods [16, 18] have demonstrated the feasibility of transferring a sign encoder pre-trained on large-scale sign-spotting data into downstream tasks. We follow this practice and use an I3D network pre-trained on BSL-1K [61], a sign classification dataset collected via sign spotting, as our primary sign encoder. Due to its favorable transfer performance, we term this model as a domain-agnostic sign encoder. Nevertheless, the domain gap between BSL-1K and sign language retrieval datasets is non-negligible. To tackle this problem, we further fine-tune a domain-aware sign encoder, which has an identical architecture to the domain-agnostic sign en + +coder, on target datasets through pseudo-labeling. The final sign encoder is composed of the well-optimized domain-aware sign encoder and the powerful domain-agnostic sign encoder, as illustrated in Figure 3a. + +Pseudo-Labeling on Target Datasets. Now we describe the details of pseudo-labeling. Given a sign video from a target dataset, we adopt a sliding window with the stride of 1 and the window size of 16 to generate a set of temporally overlapped clips. For each clip, we first utilize the pretrained domain-agnostic sign encoder to produce its prediction. Then we binarize the prediction with a pre-defined threshold $\lambda$ to generate the corresponding pseudo label. The invalid samples, whose maximum score is lower than $\lambda$ , are filtered. We repeat the above process for all sign videos and eventually build a pseudo-labeled set. Our domain-aware sign encoder, which is initialized by the domain-agnostic sign encoder, is fine-tuned on the pseudo-labeled set via a standard cross-entropy loss for classification training. + +Feature Extraction with Sign Encoder. So far, we acquire a domain-aware sign encoder approximately aligned in target domain. Nevertheless, its capability is restricted by the unavoidable noises in pseudo-labels and the limited amount of pseudo-labeled samples. Recall that we already have a powerful domain-agnostic sign encoder pre-trained on large-scale dataset in hand, inspiring us to make use of both domain-agnostic sign encoder $h_{\xi}(\cdot)$ and domain-aware sign encoder $h_{\theta}(\cdot)$ to extract discriminative and domain-aligned features. Our final sign encoder $H(\cdot)$ , as shown in Figure 3a, is a weighted combination of $h_{\xi}(\cdot)$ and $h_{\theta}(\cdot)$ with a trade-off hyper-parameter $\alpha$ . As described above, $H(\cdot)$ encodes sign videos in a sliding window manner. For simplicity, we use $H(v)$ to denote feature extraction on sign video $v$ , which is formulated as: + +$$ +H (v) = \alpha h _ {\xi} (v) + (1 - \alpha) h _ {\theta} (v). \tag {1} +$$ + +# 3.3. Cross-Linguual Contrastive Learning + +The objective of cross-lingual contrastive learning (CLCL) is to learn a joint embedding space of sign videos and texts while concurrently identifying the fine-grained sign-to-word mappings during training. An overview is shown in Figure 3b. CLCL takes a mini-batch $\{(v_{n},t_{n})\}_{n = 1}^{N}$ containing $N$ sign-video-text pairs as input, and contrasts paired data in a shared embedding space for sign language retrieval. + +Sign Features and Word Features. Given a sign video $v \in \{v_n\}_{n=1}^N$ , we first adopt our sign encoder $H(\cdot)$ described in Section 3.2 to extract its intermediate feature. Note $H(\cdot)$ encodes sign videos in a sliding window manner, and thus there are no interactions among different clips. To facilitate information exchange, we further append a 12-layer Transformer [62] $F(\cdot)$ onto $H(\cdot)$ to extract sign features $S$ of sign video $v$ , which is formulated + +as $S = F(H(v)) \in \mathbb{R}^{M \times D}$ , where $M$ denotes the number of clips, and $D$ is the hidden dimension. Given a text $t \in \{t_n\}_{n=1}^N$ , we convert $t$ into a lower-cased byte pair encoding (BPE) representation [53], which is subsequently fed into another 12-layer Transformer $G(\cdot)$ to generate the word features $W = G(t) \in \mathbb{R}^{L \times D}$ , where $L$ represents word number. + +Since CLIP [51] shows excellent transfer capability in various downstream tasks [17, 22, 27, 54, 66, 67], we initialize $F(\cdot)$ and $G(\cdot)$ with CLIP's image encoder (ViT-B) and text encoder, respectively, to ease the learning. Though CLIP's vision encoder takes image patches as inputs, we experimentally find that it generalizes well in our scenario where input data is in a different modality. + +Cross-Lingual Similarity. There exist inherent sign-to-word mappings between sign languages and natural languages. To incorporate this prior knowledge into learning, we introduce cross-lingual similarity—an indicator to identify sign-to-word mappings between $i$ -th sign video $v_{i}$ and $j$ -th text $t_j$ . Concretely, given sign features $\pmb{S}_i \in \mathbb{R}^{M \times D}$ of $v_i$ , and word features $\pmb{W}_j \in \mathbb{R}^{L \times D}$ of $t_j$ , we calculate a cross-lingual similarity matrix $\pmb{E}_{(i,j)} = \pmb{S}_i \cdot \pmb{W}_j^T \in \mathbb{R}^{M \times L}$ . Each element in $\pmb{E}_{(i,j)}$ represents the similarity of a sign clip in $v_i$ and a word in $t_j$ . + +Cross-Lingual Contrastive Learning. Directly apply supervisions on token-wise similarity matrix $\pmb{E}_{(i,j)}$ is infeasible due to the absence of fine-grained sign-to-word annotations. Inspired by the recent progress of vision-language contrastive learning [23, 25, 29, 39, 40, 51], we turn to contrast the global representations of sign videos and texts. The underlying idea is to calculate a global similarity $z$ of $v_{i}$ and $t_j$ based on $\pmb{E}_{(i,j)} \in \mathbb{R}^{M \times L}$ . + +Sign-Video-to-Text Contrast. We first introduce sign-video-to-text contrast as shown in Figure 3b. To be specific, we utilize a Softmax operation to each row of $\pmb{E}_{(i,j)}$ and multiply the resulting matrix with $\pmb{E}_{(i,j)}$ to generate a sign-to-word similarity matrix $\pmb{E}_{(i,j)}' \in \mathbb{R}^{M \times L}$ , where each row represents the re-weighted similarities between a sign clip in $v_{i}$ and all words in $t_j$ . After that, we adopt a row-wise addition operation on $\pmb{E}_{(i,j)}'$ to yield the sign-to-text similarity vector $e_{(i,j)} \in \mathbb{R}^M$ , where each element denotes a similarity of a sign clip in $v_{i}$ and whole text $t_j$ . At last, we average all elements in $e_{(i,j)}$ to produce the global similarity $z$ of sign video $v_{i}$ and text $t_j$ . + +In the same way, we can calculate the similarities for both positive pairs $\{(v_i,t_i)\}_{i = 1}^N$ and negative pairs $\{(v_i,t_j)\}_{i = 1,j = 1,i\neq j}^N$ in a mini-batch, yielding a video-to-text similarity matrix $Z_{V2T}\in \mathbb{R}^{N\times N}$ , where $Z_{V2T}^{(i,j)}$ denotes the global similarity of $v_{i}$ and $t_j$ . Following CLIP [51], we adopt InfoNCE loss [26] to pull the embeddings of matched image-text pairs together while pushing those of non-matched pairs apart, which is formulated as + +follows: + +$$ +\begin{array}{l} \mathcal {L} _ {V 2 T} = - \frac {1}{2 N} \sum_ {\substack {i = 1 \\ N}} ^ {N} \log \frac {\exp \left(\boldsymbol {Z} _ {V 2 T} ^ {(i , j)} / \tau\right)}{\sum_ {j = 1} ^ {N} \exp \left(\boldsymbol {Z} _ {V 2 T} ^ {(i , j)} / \tau\right)} \tag{2} \\ - \frac {1}{2 N} \sum_ {j = 1} ^ {N} \log \frac {\exp \left(\mathbf {Z} _ {V 2 T} ^ {(i , j)} / \tau\right)}{\sum_ {i = 1} ^ {N} \exp \left(\mathbf {Z} _ {V 2 T} ^ {(i , j)} / \tau\right)}, \\ \end{array} +$$ + +where $\tau$ is a trainable temperature parameter. + +Text-to-Sign-Video Contrast. Up to now, we have introduced sign-video-to-text contrast. A symmetrical version, termed text-to-sign-video contrast, shares the similar spirit as shown in Figure 3b. The implementation of text-to-sign-video contrast is extremely simple: we replace the row-wise operations (i.e., Softmax and addition) in sign-video-to-text contrast with column-wise ones and keep the remaining processes unchanged. We use $\mathcal{L}_{T2V}$ to denote the loss function of text-to-sign-video contrast. In our implementation, we reuse the loss defined in Eq 2 but substitute the input with the text-to-video similarity matrix $Z_{T2V}$ . + +Loss Function. The overall loss for cross-lingual contrastive learning is a weighted sum of $\mathcal{L}_{V2T}$ and $\mathcal{L}_{T2V}$ with a trade-off hyper-parameter $\beta$ : + +$$ +\mathcal {L} = \beta \mathcal {L} _ {V 2 T} + (1 - \beta) \mathcal {L} _ {T 2 V}. \tag {3} +$$ + +# 3.4. Text Augmentation + +Considering that the datasets of sign language retrieval are typically small-scale, we explore text augmentations to improve the generalization of our approach. EDA [64] introduces three simple yet efficient data augmentations in text classification task: random delete randomly removes words in a sentence; synonym replacement randomly selects words from a sentence that are not stop words and replaces them with synonyms; random swap randomly chooses two words in a sentence and swaps their positions. The first two augmentations have been proven effective in text classification task. However, we experimentally find that our focused sign language retrieval is sensitive to random delete and synonym replacement augmentations. To guarantee that the augmented texts preserve the original semantic meanings, we only adopt the random swap augmentation in our approach. We suppose there are two reasons: 1) the word order of sign languages and natural languages are constitutionally distinct, and reordering does not affect semantic meanings; 2) the proposed cross-lingual contrastive learning is insensitive to word order. + +# 4. Experiment + +# 4.1. Datasets and Implementation Details + +Datasets. We primarily focus on How2Sign [19] dataset. Our model is also evaluated on PHOENIX-2014T [8] and + +CSL-Daily [71], which are primarily used for sign language recognition and translation in previous works. + +How2Sign is a large-scale continuous American sign language (ASL) dataset consisting of a parallel corpus of about 80 hours of sign videos with subtitle annotations. It covers a wide range of instructional videos corresponding to various categories. There are 31164, 1740 and 2356 sign-video-text pairs in training, validation and test sets, respectively. Following SPOT-ALIGN [18], we remove the invalid pairs where the subtitle alignment is detected to exceed the video duration, remaining 31085, 1739 and 2348 available pairs in training, validation and test sets, respectively. The resolution of sign videos is $1280 \times 720$ , we crop the human bodies of signers with Faster R-CNN [52] to generate valid videos. + +PHOENIX-2014T is a German sign language (Deutsche Gebärddensprache, DGS) dataset collected in the domain of weather forecast from TV broadcast, consisting of 7096, 519 and 642 video-text pairs in training, validation and test sets, respectively. + +CSL-Daily is a recently released Chinese sign language (CSL) dataset. The topic of CSL-Daily revolves around people's daily lives, including 18401, 1077, 1176 parallel samples in training, validation and test set, respectively. + +Evaluation Metric. Following previous works [18, 43, 51, 57], retrieval performance is evaluated by recall at rank K (R@K, higher is better) and median rank (MedR, lower is better). We evaluate our approach on both text-to-sign-video (T2V) retrieval and sign-video-to-text (V2T) retrieval tasks. We report R@1, R@5, and R@10 in all experiments, and additionally report MedR when comparing with state-of-the-art approaches. + +Implementation Details. The sign encoder takes videos of resolution of $256 \times 256$ as input. The domain-agnostic sign encoder is a I3D [11] network pre-trained on BSL-1K [61]. In pseudo label generation, we set the threshold $\lambda$ to 0.6 to filter samples with low-confidence. Non-maximum suppression (NMS) with a temporal window of 24 frames is utilized to remove the duplicates among the pseudo-labeled samples. A collection of approximate 64K pseudo-labeled samples covering a vocabulary of 1220 words is eventually generated. The domain-aware sign encoder is initialized with the domain-agnostic one and fine-tuned with a learning rate of $1 \times 10^{-2}$ and batch size of 4 for 15 epochs. In the training of cross-lingual contrastive learning, the vision transformer and text transformer are initialized by the image encoder and text encoder in CLIP (ViT-B/32) [51]. The maximum length of sign clip features and text features are set to 64 and 32, respectively. The model is fine-tuned with Adam optimizer [31] with batch size of 512. The initial learning rate is set to $1 \times 10^{-5}$ , which is decreased with a cosine schedule following the CLIP [51]. We set $\alpha = 0.8$ in Eq. 1 and $\beta = 0.5$ in Eq.3. The languages of PHOENIX-2014T and CSL-Daily are German and Chinese + +
ModelT2VV2T
R@1↑R@5↑R@10↑MedR↓R@1↑R@5↑R@10↑MedR↓
SA-SR [18]18.932.136.562.011.627.432.569.0
SA-CM [18]24.340.746.516.017.940.146.914.0
SA-COMB [18]34.248.052.68.023.647.053.07.5
Ours56.669.974.71.051.664.870.11.0
+ +Table 1. Comparison with the different variants of the pioneer SPOT-ALIGN (SA) [18] on How2Sign [19] dataset. + +
ModelT2VV2T
R@1↑R@5↑R@10↑MedR↓R@1↑R@5↑R@10↑MedR↓
Translation [10]30.253.163.44.528.852.060.856.1
SA-CM [18]48.676.584.62.050.378.484.41.0
SA-COMB [18]55.879.687.21.053.179.486.11.0
Ours69.586.692.11.070.288.092.81.0
+ +Table 2. Comparison with the different variants of the pioneer SPOT-ALIGN (SA) [18] on PHOENIX2014T [8] dataset. + +
ModelT2VV2T
R@1↑R@5↑R@10↑MedR↓R@1↑R@5↑R@10↑MedR↓
Ours75.388.291.91.074.789.492.21.0
+ +respectively. Since CLIP is trained on English corpus, to reuse CLIP's text encoder, we utilize Google translation to translate the texts of these two datasets into English. + +# 4.2. Comparison with State-of-the-art Methods + +We compare our method with different variants of the pioneer, called SPOT-ALIGN [18], on How2Sign and PHOENIX-2014T. We also provide the results on CSL-Daily as a baseline. + +Table 1 and Table 2 show the comparisons between our approach and SPOT-ALIGN [18] on How2Sign and PHOENIX-2014T, respectively. SPOT-ALIGN builds the final combination (COMB) model by integrating its primary cross-modal (CM) model with an auxiliary retrieval model (sign recognition (SR) model for How2Sign and Translation [10] for PHOENIX-2014T). Our method outperforms the COMB model, which achieves best results in SPOT-ALIGN, by large margins, achieving $+22.4$ T2V and $+28.0$ V2T R@1 improvements on How2Sign, $+13.7$ T2V and $+17.1$ V2T R@1 improvements on PHOENIX-2014T. It is worth mentioning that the SPOT-ALIGN conducts three rounds of sign spotting [1, 47] and encoder training. In contrast, we simplify the training of sign encoder and only perform a single round of training on pseudo-labeled data. We also provide a baseline on CSL-Daily dataset as shown in Table 3, demonstrating that our model can be generalized to various sign languages. + +Table 3. We additionally provide a baseline for CSL-Daily [71] dataset. + +
MethodT2VV2T
R@1R@5R@10R@1R@5R@10
Baseline31.549.657.926.444.452.8
+SE33.351.959.830.948.856.4
+SE+CLCL54.067.171.950.063.268.1
+SE+CLCL+TA56.669.974.751.664.870.1
+ +Table 4. Ablation study of each proposed component. SE: sign encoder; CLCL: cross-lingual contrastive learning; TA: text augmentation. + +
EncoderT2VV2T
R@1R@5R@10R@1R@5R@10
Single-Ag [61]53.168.073.447.362.967.0
Single-Aw54.167.573.149.161.867.4
Fusion-Average54.768.773.849.663.769.0
Fusion-Weighted Sum56.669.974.751.664.870.1
+ +Table 5. Results of domain-agnostic sign encoder (Single-Ag) and domain-aware sign encoder (Single-Aw). We also study different ways to integrate the features extracted by Single-Ag and Single-Aw, including average and weighted sum. + +# 4.3. Ablation Study + +We conduct all ablation studies on the most challenging How2Sign dataset. + +The Effectiveness of Each Proposed Component. Table 4 shows the ablation of each component. We first build a baseline where the sign encoder is the domain-agnostic one and contrastive learning is trained with the standard contrastive loss [51]. Then we add the proposed sign encoder (SE), cross-lingual contrastive learning (CLCL) and text augmentation (TA) to the baseline step by step. SE encodes domain-relevant and discriminative features, yielding $+1.8$ T2V and $+4.5$ V2T R@1 gains. The proposed CLCL significantly boosts the performance by $+20.7$ T2V and $+19.1$ V2T R@1 improvements, demonstrating that identifying fine-grained sign-to-word mappings is essential for sign language retrieval. The introduction of TA further promotes the retrieval task, achieving $56.6$ R@1 on V2T and $51.6$ R@1 on T2V, respectively. + +Various Sign Encoders. As described in Section 3.2, our sign encoder is composed of a domain-agnostic sign encoder (Single-Ag) [61] and a domain-aware sign encoder (Single-Aw). We first report the results of each individual encoder in Table 5. Next, we study the different ways to integrate the features extracted by these two encoders, including average and weighted sum as defined in Eq. 1. The results are also shown in Table 5. We experimentally find that the weighted sum strategy yields best results. + +Variants of Cross-Lingual Contrastive Learning. As described in Section 3.3 and illustrated in Figure 3b, through the use of a combination of softmax, multiplication, sum, and average operations, we convert the fine-grained cross-lingual (sign-to-word) similarity to the coarse-grained sign-video-to-text similarity to enable contrastive learning. We + +
StrategyT2VV2T
R@1R@5R@10R@1R@5R@10
Mean33.152.559.729.847.855.3
Max42.259.766.038.555.162.0
Softmax56.669.974.751.664.870.1
+ +Table 6. Study on different strategies to identify the fine-grained sign-to-word mappings in cross-lingual contrastive learning. + +
Text AugmentationT2VV2T
R@1R@5R@10R@1R@5R@10
None54.067.171.950.063.268.1
RD52.766.371.647.662.167.4
SR53.968.773.047.561.265.9
RS56.669.974.751.664.870.1
+ +refer to this process as "Softmax". Here we study two variants termed "Mean" and "Max". The "Mean" and "Max" strategies simply replace the combination of softmax, multiplication and sum operations with a simple mean operation and a max operation, respectively. The results are shown in Table 6. We observe that the "Mean" strategy performs worst since it merely evaluates the overall similarity of a text and a sign video and ignores the fine-grained sign-to-word mappings during training. The "Max" strategy identifies hard sign-to-word mappings, i.e., one sign is associated with the most similar word, and vice versa. Nevertheless, since we do not have the ground-truth of sign-to-word mappings, it is challenging for models to identify the confident one-to-one mappings in a weakly supervised manner, as discussed in previous works on multiple instance learning [47]. In contrast, our default "Softmax" strategy localizes the soft sign-to-word mappings and achieves the best results. + +Study of Text Augmentations. In Table 7, we study three text augmentations described in Section 3.4: random delete (RD), synonym replacement (SR) and random swap (RS). We observe a slight performance drop when introducing RD and SR into training. Sign language retrieval is not only a video-text retrieval task, but also a cross-lingual retrieval challenge, deletion and replacement may break the intrinsic sign-to-word mappings. In contrast, RS augmentation preserves the semantics of texts and we observe a +2.6 T2V and a +1.6 V2T R@1 improvements over the counterpart without any text augmentations. + +The Effects of CLIP Initialization. In our framework, the vision and text Transformer are initialized with the CLIP's vision encoder and text encoder. Note that the input modality of our vision Transformer and that of CLIP's vision encoder are different. Ours takes a feature sequence as input while the input of CLIP's image encoder is a set of image patches. In Table 8, we compare a randomly initialized baseline with the one initialized by CLIP. We find that CLIP-initialization significantly improves the perfor + +Table 7. Ablation study on different text augmentations. RD: random delete; SR: synonym replacement; RS: random swap. + +
InitializationT2VV2T
R@1R@5R@10R@1R@5R@10
Random45.260.367.540.654.559.7
CLIP56.669.974.751.664.870.1
+ +Table 8. Comparison between CLIP initialization and random initialization. + +
Frozen layersT2VV2T
R@1R@5R@10R@1R@5R@10
None56.669.974.751.664.870.1
154.468.273.449.763.568.8
253.868.673.348.562.867.9
651.766.671.146.260.966.6
1249.565.271.444.759.765.2
+ +Table 9. Study of freezing different layers of text Transformer. + +mance, yielding $+11.4$ T2V and $+11.0$ V2T R@1 improvements though the input originates from a completely distinct modality. + +Study on Text Transformer. The text Transformer in our model is initialized with the CLIP's text encoder. Considering the generalization capability of the pre-trained CLIP, we attempt to freeze the shallow layers of our text transformer. As shown in Table 9, the performance gradually decreases as the number of the frozen layers increases. CLIP is trained on large-scale image-text pairs, however, the domain gap between image-text data and sign-video-text data is non-negligible. Though CLIP shows promising transfer capacity, it is optimal to fine-tune the whole model on target datasets. + +# 5. Conclusion + +In this paper, we propose a novel framework named Cross-lingual Contrastive learning (CiCo) for recently introduced sign language retrieval task. We formulate sign language retrieval as a cross-lingual retrieval task as well as a video-text retrieval problem. CiCo models the fine-grained cross-lingual mappings between sign videos and texts via the proposed cross-lingual contrastive learning. We also introduce a sign encoder, which is composed of a domain-agnostic encoder and a domain-aware one, to extract discriminative and domain-aligned features. Our Cico outperforms the pioneer SPOT-ALIGN by large margins on How2Sign and PHOENIX-2014T benchmarks. We also provide a baseline on CSL-Daily. We hope our approach could serve as a solid baseline for future research. + +Acknowledgement. This work was partially supported by National Natural Science Foundation of China (No. 62072112), National Key R&D Program of China (No. 2020AAA0108301), and Scientific and Technological Innovation Action Plan of Shanghai Science and Technology Committee (No. 22511101502, 22511102202). + +# References + +[1] Samuel Albanie, Gül Varol, Liliane Momeni, Triantafyllos Afouras, Joon Son Chung, Neil Fox, and Andrew Zisserman. Bsl-1k: Scaling up co-articulated sign language recognition using mouthing cues. In European Conference on Computer Vision, pages 35–53. Springer, 2020. 1, 3, 7 +[2] Mikel Artetxe, Gorka Labaka, and Eneko Agirre. Learning bilingual word embeddings with (almost) no bilingual data. In *Meeting of the Association for Computational Linguistics*, pages 451–462, 2017. 3 +[3] Mikel Artetxe, Gorka Labaka, and Eneko Agirre. A robust self-learning method for fully unsupervised cross-lingual mappings of word embeddings. arXiv preprint arXiv:1805.06297, 2018. 3 +[4] Vassilis Athitsos, Carol Neidle, Stan Sclaroff, Joan Nash, Alexandra Stefan, Ashwin Thangali, Hajjing Wang, and Quan Yuan. Large lexicon project: American sign language video corpus and sign language indexing/retrieval algorithms. In Workshop on the Representation and Processing of Sign Languages: Corpora and Sign Language Technologies, volume 2, pages 11-14, 2010. 3 +[5] Max Bain, Arsha Nagrani, Gül Varol, and Andrew Zisserman. Frozen in time: A joint video and image encoder for end-to-end retrieval. In IEEE/CVF International Conference on Computer Vision, pages 1728-1738, 2021. 2, 3 +[6] Lisa Ballesteros and Bruce Croft. Dictionary methods for cross-lingual information retrieval. In International Conference on Database and Expert Systems Applications, pages 791–801. Springer, 1996. 2 +[7] Yoshua Bengio, Aaron Courville, and Pascal Vincent. Representation learning: A review and new perspectives. IEEE Transactions on Pattern Analysis and Machine Intelligence, 35(8):1798-1828, 2013. 3 +[8] Necati Cihan Camgoz, Simon Hadfield, Oscar Koller, Hermann Ney, and Richard Bowden. Neural sign language translation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7784-7793, 2018. 2, 3, 6, 7 +[9] Necati Cihan Camgoz, Oscar Koller, Simon Hadfield, and Richard Bowden. Multi-channel transformers for multi-articulatory sign language translation. In European Conference on Computer Vision, pages 301-319. Springer, 2020. 3 +[10] Necati Cihan Camgoz, Oscar Koller, Simon Hadfield, and Richard Bowden. Sign language transformers: Joint end-to-end sign language recognition and translation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10023-10033, 2020. 1, 7 +[11] Joao Carreira and Andrew Zisserman. Quo vadis, action recognition? a new model and the kinetics dataset. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6299-6308, 2017. 3, 6 +[12] Yutong Chen, Fangyun Wei, Xiao Sun, Zhirong Wu, and Stephen Lin. A simple multi-modality transfer learning baseline for sign language translation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5120-5130, 2022. 1, 3 + +[13] Yutong Chen, Ronglai Zuo, Fangyun Wei, Yu Wu, Shujie LIU, and Brian Mak. Two-stream network for sign language recognition and translation. In Alice H. Oh, Alekh Agarwal, Danielle Belgrave, and Kyunghyun Cho, editors, Advances in Neural Information Processing Systems, 2022. 1, 3 +[14] Necati Cihan Camgoz, Simon Hadfield, Oscar Koller, and Richard Bowden. Subunets: End-to-end hand shape and continuous sign language recognition. In IEEE/CVF International Conference on Computer Vision, pages 3056-3065, 2017. 1 +[15] Runpeng Cui, Hu Liu, and Changshui Zhang. A deep neural framework for continuous sign language recognition by iterative training. IEEE Transactions on Multimedia, 21(7):1880-1891, 2019. 1 +[16] Tonni Das Jui, Gissella Maria Bejarano, and Pablo Rivas. A machine learning-based segmentation approach for measuring similarity between sign languages. In sign-lang@ LREC 2022, pages 94-101. European Language Resources Association (ELRA), 2022. 4 +[17] Yu Du, Fangyun Wei, Zihe Zhang, Miaojing Shi, Yue Gao, and Guoqi Li. Learning to prompt for open-vocabulary object detection with vision-language model. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14084-14093, 2022. 5 +[18] Amanda Duarte, Samuel Albanie, Xavier Giró-i Nieto, and Gül Varol. Sign language video retrieval with free-form textual queries. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14094-14104, 2022. 1, 2, 3, 4, 6, 7 +[19] Amanda Duarte, Shruti Palaskar, Lucas Ventura, Deepti Ghadiyaram, Kenneth DeHaan, Florian Metze, Jordi Torres, and Xavier Giro-i Nieto. How2Sign: A Large-scale Multimodal Dataset for Continuous American Sign Language. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2021. 2, 6, 7 +[20] Ali Farhadi, David Forsyth, and Ryan White. Transfer learning in sign language. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1-8. IEEE, 2007. 3 +[21] Holger Fillbrandt, Suat Akyol, and K-F Kraiss. Extraction of 3d hand shape and posture from image sequences for sign language recognition. In 2003 IEEE International SOI Conference, pages 181-186. IEEE, 2003. 3 +[22] Stanislav Fort, Jie Ren, and Balaji Lakshminarayanan. Exploring the limits of out-of-distribution detection. Advances in Neural Information Processing Systems, 34:7068-7081, 2021. 5 +[23] Andrea Frome, Greg S Corrado, Jon Shlens, Samy Bengio, Jeff Dean, Marc'Aurelio Ranzato, and Tomas Mikolov. Devise: A deep visual-semantic embedding model. Advances in neural information processing systems, 26, 2013. 3, 5 +[24] Valentin Gabeur, Chen Sun, Karteek Alahari, and Cordelia Schmid. Multi-modal transformer for video retrieval. In European Conference on Computer Vision, pages 214-229. Springer, 2020. 2, 3 +[25] Yunchao Gong, Qifa Ke, Michael Isard, and Svetlana Lazebnik. A multi-view embedding space for modeling internet images, tags, and their semantics. International Journal of Computer Vision, 106(2):210-233, 2014. 3, 5 + +[26] Michael Gutmann and Aapo Hyvarinen. Noise-contrastive estimation: A new estimation principle for unnormalized statistical models. In International Conference on Artificial Intelligence and Statistics, pages 297–304. JMLR Workshop and Conference Proceedings, 2010. 5 +[27] Tony Huang, Jack Chu, and Fangyun Wei. Unsupervised prompt learning for vision-language models. arXiv preprint arXiv:2204.03649, 2022. 5 +[28] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In International Conference on Machine Learning, pages 4904-4916. PMLR, 2021. 3 +[29] Armand Joulin, Laurens van der Maaten, Allan Jabri, and Nicolas Vasilache. Learning visual features from large weakly supervised data. In European Conference on Computer Vision, pages 67-84. Springer, 2016. 3, 5 +[30] Hamid Reza Vaezi Joze and Oscar Koller. Ms-asl: A large-scale data set and benchmark for understanding american sign language. arXiv preprint arXiv:1812.01053, 2018. 3 +[31] Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In Yoshua Bengio and Yann LeCun, editors, 3rd International Conference on Learning Representations, ICLR 2015, San Diego, CA, USA, May 7-9, 2015, Conference Track Proceedings, 2015. 6 +[32] Oscar Koller, Necati Cihan Camgoz, Hermann Ney, and Richard Bowden. Weakly supervised learning with multistream cnn-lstm-hmms to discover sequential parallelism in sign language videos. IEEE Transactions on Pattern Analysis and Machine Intelligence, 42(9):2306-2320, 2019. 1 +[33] Oscar Koller, Jens Forster, and Hermann Ney. Continuous sign language recognition: Towards large vocabulary statistical recognition systems handling multiple signers. Computer Vision and Image Understanding, 141:108-125, 2015. 1 +[34] Victor Lavrenko, Martin Choquette, and W Bruce Croft. Cross-lingual relevance models. In International ACM SIGIR Conference on Research and Development in Information Retrieval, pages 175–182, 2002. 2, 3 +[35] Yann LeCun, Yoshua Bengio, and Geoffrey Hinton. Deep learning. Nature, 521(7553):436-444, 2015. 3 +[36] Dongxu Li, Cristian Rodriguez, Xin Yu, and Hongdong Li. Word-level deep sign language recognition from video: A new large-scale dataset and methods comparison. In IEEE/CVF winter conference on applications of computer vision, pages 1459-1469, 2020. 3 +[37] Dongxu Li, Chenchen Xu, Xin Yu, Kaihao Zhang, Benjamin Swift, Hanna Suominen, and Hongdong Li. Tspnet: Hierarchical feature learning via temporal semantic pyramid for sign language translation. Advances in Neural Information Processing Systems, 33:12034-12045, 2020. 3 +[38] Dongxu Li, Xin Yu, Chenchen Xu, Lars Petersson, and Hongdong Li. Transferring cross-domain knowledge for video sign language recognition. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6205-6214, 2020. 3 +[39] Junnan Li, Ramprasaath Selvaraju, Akhilesh Gotmare, Shafiq Joty, Caiming Xiong, and Steven Chu Hong Hoi. + +Align before fuse: Vision and language representation learning with momentum distillation. Advances in Neural Information Processing Systems, 34:9694-9705, 2021. 3, 5 +[40] Yangguang Li, Feng Liang, Lichen Zhao, Yufeng Cui, Wanli Ouyang, Jing Shao, Fengwei Yu, and Junjie Yan. Supervision exists everywhere: A data efficient contrastive language-image pre-training paradigm. arXiv preprint arXiv:2110.05208, 2021. 5 +[41] Ji Lin, Chuang Gan, and Song Han. Tsm: Temporal shift module for efficient video understanding. In IEEE/CVF International Conference on Computer Vision, pages 7083-7093, 2019. 3 +[42] Song Liu, Haoqi Fan, Shengsheng Qian, Yiru Chen, Wenkui Ding, and Zhongyuan Wang. Hit: Hierarchical transformer with momentum contrast for video-text retrieval. In IEEE/CVF International Conference on Computer Vision, pages 11915-11925, 2021. 2, 3 +[43] Yang Liu, Samuel Albanie, Arsha Nagrani, and Andrew Zisserman. Use what you have: Video retrieval using representations from collaborative experts. arXiv preprint arXiv:1907.13487, 2019. 2, 3, 6 +[44] Huaishao Luo, Lei Ji, Ming Zhong, Yang Chen, Wen Lei, Nan Duan, and Tianrui Li. Clip4clip: An empirical study of clip for end to end video clip retrieval and captioning. Neurocomputing, 508:293-304, 2022. 2, 3 +[45] Antoine Miech, Dimitri Zhukov, Jean-Baptiste Alayrac, Makarand Tapaswi, Ivan Laptev, and Josef Sivic. Howto100m: Learning a text-video embedding by watching hundred million narrated video clips. In IEEE/CVF International Conference on Computer Vision, pages 2630-2640, 2019. 2 +[46] Tomas Mikolov, Quoc V Le, and Ilya Sutskever. Exploiting similarities among languages for machine translation. arXiv preprint arXiv:1309.4168, 2013. 3 +[47] Liliane Momeni, Gül Varol, Samuel Albanie, Triantafyllos Afouras, and Andrew Zisserman. Watch, read and lookup: learning to spot signs from multiple supervisors. In Asian Conference on Computer Vision, 2020. 3, 4, 7, 8 +[48] Mathew Monfort, SouYoung Jin, Alexander Liu, David Harwath, Rogerio Feris, James Glass, and Aude Oliva. Spoken moments: Learning joint audio-visual representations from video descriptions. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14871-14881, 2021. 2 +[49] Yasuhide Mori, Hironobu Takahashi, and Ryuichi Oka. Image-to-word transformation based on dividing and vector quantizing images with words. In International Workshop on Multimedia Intelligent Storage and Retrieval Management, pages 1-9. Citeseer, 1999. 3 +[50] Aitor Ormazabal, Mikel Artetxe, Gorka Labaka, Aitor Soroa, and Eneko Agirre. Analyzing the limitations of cross-lingual word embedding mappings. arXiv preprint arXiv:1906.05407, 2019. 3 +[51] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language super + +vision. In International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 3, 5, 6, 7 +[52] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. Conference on Neural Information Processing Systems, 28, 2015. 6 +[53] Rico Sennrich, Barry Haddow, and Alexandra Birch. Neural machine translation of rare words with subword units. arXiv preprint arXiv:1508.07909, 2015. 5 +[54] Sheng Shen, Liunian Harold Li, Hao Tan, Mohit Bansal, Anna Rohrbach, Kai-Wei Chang, Zhewei Yao, and Kurt Keutzer. How much can clip benefit vision-and-language tasks? arXiv preprint arXiv:2107.06383, 2021. 5 +[55] Thad Starner, Joshua Weaver, and Alex Pentland. Real-time american sign language recognition using desk and wearable computer based video. IEEE Transactions on Pattern Analysis and Machine Intelligence, 20(12):1371-1375, 1998. 3 +[56] Thad E Starner. Visual recognition of american sign language using hidden markov models. Technical report, Massachusetts inst of tech Cambridge dept of brain and cognitive sciences, 1995. 3 +[57] Chen Sun, Austin Myers, Carl Vondrick, Kevin Murphy, and Cordelia Schmid. Videobert: A joint model for video and language representation learning. In IEEE/CVF International Conference on Computer Vision, pages 7464-7473, 2019. 2, 3, 6 +[58] Rachel Sutton-Spence. Mouthings and simultaneity in British sign language. Amsterdam Studies in the Theory and History of Linguistic Science Series, 281:147, 2007. 3 +[59] Chau Tran, Yuqing Tang, Xian Li, and Jiatao Gu. Crosslingual retrieval for iterative self-supervised training. Advances in Neural Information Processing Systems, 33:2207-2219, 2020. 2, 3 +[60] Du Tran, Heng Wang, Lorenzo Torresani, Jamie Ray, Yann LeCun, and Manohar Paluri. A closer look at spatiotemporal convolutions for action recognition. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6450-6459, 2018. 3 +[61] Gul Varol, Liliane Momeni, Samuel Albanie, Triantafyllos Afouras, and Andrew Zisserman. Read and attend: Temporal localisation in sign language videos. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16857-16866, 2021. 1, 2, 3, 4, 6, 7 +[62] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. 2017. 5 +[63] Xiaolong Wang, Ross Girshick, Abhinav Gupta, and Kaiming He. Non-local neural networks. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7794-7803, 2018. 3 +[64] Jason Wei and Kai Zou. EDA: Easy data augmentation techniques for boosting performance on text classification tasks. In 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages 6383-6389, Hong Kong, China, Nov. 2019. Association for Computational Linguistics. 6 + +[65] Jinxi Xu, Ralph Weischedel, and Chanh Nguyen. Evaluating a probabilistic model for cross-lingual information retrieval. In International ACM SIGIR Conference on Research and Development in Information Retrieval, pages 105–110, 2001. 3 +[66] Mengde Xu, Zheng Zhang, Fangyun Wei, Han Hu, and Xiang Bai. Side adapter network for open-vocabulary semantic segmentation. arXiv preprint arXiv:2302.12242, 2023. 5 +[67] Mengde Xu, Zheng Zhang, Fangyun Wei, Yutong Lin, Yue Cao, Han Hu, and Xiang Bai. A simple baseline for open-vocabulary semantic segmentation with pre-trained vision-language model. In European Conference on Computer Vision, pages 736-753. Springer, 2022. 5 +[68] Youngjae Yu, Jongseok Kim, and Gunhee Kim. A joint sequence fusion model for video question answering and retrieval. In European Conference on Computer Vision, pages 471-487, 2018. 2, 3 +[69] Mozhi Zhang, Keyulu Xu, Ken-ichi Kawarabayashi, Stefanie Jegelka, and Jordan Boyd-Graber. Are girls neko or shōjo? cross-lingual alignment of non-isomorphic embeddings with iterative normalization. arXiv preprint arXiv:1906.01622, 2019.3 +[70] Shilin Zhang and Bo Zhang. Using revised string edit distance to sign language video retrieval. In International Conference on Computational Intelligence and Natural Computing, volume 1, pages 45-49. IEEE, 2010. 3 +[71] Hao Zhou, Wengang Zhou, Weizhen Qi, Junfu Pu, and Houqiang Li. Improving sign language translation with monolingual data by sign back-translation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1316-1325, 2021. 6, 7 +[72] Hao Zhou, Wengang Zhou, Yun Zhou, and Houqiang Li. Spatial-temporal multi-cue network for sign language recognition and translation. IEEE Transactions on Multimedia, 24:768-779, 2021. 3 +[73] Ronglai Zuo, Fangyun Wei, and Brian Mak. Natural language-assisted sign language recognition. IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023. 1, 3 \ No newline at end of file diff --git a/2023/CiCo_ Domain-Aware Sign Language Retrieval via Cross-Lingual Contrastive Learning/images.zip b/2023/CiCo_ Domain-Aware Sign Language Retrieval via Cross-Lingual Contrastive Learning/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..930a1470a5a10c9864ee930a734d76cf201cccc4 --- /dev/null +++ b/2023/CiCo_ Domain-Aware Sign Language Retrieval via Cross-Lingual Contrastive Learning/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a6be9d19a0415566eb532b87c97dbd1a9d4b8800c2418dd43019a52e96ef5b43 +size 429037 diff --git a/2023/CiCo_ Domain-Aware Sign Language Retrieval via Cross-Lingual Contrastive Learning/layout.json b/2023/CiCo_ Domain-Aware Sign Language Retrieval via Cross-Lingual Contrastive Learning/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..85b5c1c3b4923673066a614b6b5be1876604efed --- /dev/null +++ b/2023/CiCo_ Domain-Aware Sign Language Retrieval via Cross-Lingual Contrastive Learning/layout.json @@ -0,0 +1,9370 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 140, + 103, + 455, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 140, + 103, + 455, + 140 + ], + "spans": [ + { + "bbox": [ + 140, + 103, + 455, + 140 + ], + "type": "text", + "content": "CiCo: Domain-Aware Sign Language Retrieval via Cross-Lingual Contrastive Learning" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 92, + 160, + 505, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 92, + 160, + 505, + 204 + ], + "spans": [ + { + "bbox": [ + 92, + 160, + 505, + 204 + ], + "type": "text", + "content": "Yiting Cheng" + }, + { + "bbox": [ + 92, + 160, + 505, + 204 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 92, + 160, + 505, + 204 + ], + "type": "text", + "content": ", Fangyun Wei" + }, + { + "bbox": [ + 92, + 160, + 505, + 204 + ], + "type": "inline_equation", + "content": "^{2*}" + }, + { + "bbox": [ + 92, + 160, + 505, + 204 + ], + "type": "text", + "content": ", Jianmin Bao" + }, + { + "bbox": [ + 92, + 160, + 505, + 204 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 92, + 160, + 505, + 204 + ], + "type": "text", + "content": ", Dong Chen" + }, + { + "bbox": [ + 92, + 160, + 505, + 204 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 92, + 160, + 505, + 204 + ], + "type": "text", + "content": ", Wenqiang Zhang" + }, + { + "bbox": [ + 92, + 160, + 505, + 204 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 92, + 160, + 505, + 204 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 92, + 160, + 505, + 204 + ], + "type": "text", + "content": "School of Computer Science, Fudan University, " + }, + { + "bbox": [ + 92, + 160, + 505, + 204 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 92, + 160, + 505, + 204 + ], + "type": "text", + "content": "Microsoft Research Asia \n{ytcheng18, wqzhang}@fudan.edu.cn, {fawe, jianbao, doch}@microsoft.com" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 143, + 231, + 192, + 243 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 231, + 192, + 243 + ], + "spans": [ + { + "bbox": [ + 143, + 231, + 192, + 243 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 256, + 290, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 256, + 290, + 605 + ], + "spans": [ + { + "bbox": [ + 46, + 256, + 290, + 605 + ], + "type": "text", + "content": "This work focuses on sign language retrieval—a recently proposed task for sign language understanding. Sign language retrieval consists of two sub-tasks: text-to-sign-video (T2V) retrieval and sign-video-to-text (V2T) retrieval. Different from traditional video-text retrieval, sign language videos, not only contain visual signals but also carry abundant semantic meanings by themselves due to the fact that sign languages are also natural languages. Considering this character, we formulate sign language retrieval as a cross-lingual retrieval problem as well as a video-text retrieval task. Concretely, we take into account the linguistic properties of both sign languages and natural languages, and simultaneously identify the fine-grained cross-lingual (i.e., sign-to-word) mappings while contrasting the texts and the sign videos in a joint embedding space. This process is termed as cross-lingual contrastive learning. Another challenge is raised by the data scarcity issue—sign language datasets are orders of magnitude smaller in scale than that of speech recognition. We alleviate this issue by adopting a domain-agnostic sign encoder pre-trained on large-scale sign videos into the target domain via pseudolabeling. Our framework, termed as domain-aware sign language retrieval via Cross-lingual Contrastive learning or CiCo for short, outperforms the pioneering method by large margins on various datasets, e.g., +22.4 T2V and +28.0 V2TR@1 improvements on How2Sign dataset, and +13.7 T2V and +17.1 V2TR@1 improvements on PHOENIX-2014T dataset. Code and models are available at: https://github.com/FangyunWei/SLRT." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 626, + 128, + 639 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 626, + 128, + 639 + ], + "spans": [ + { + "bbox": [ + 47, + 626, + 128, + 639 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 647, + 287, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 647, + 287, + 696 + ], + "spans": [ + { + "bbox": [ + 47, + 647, + 287, + 696 + ], + "type": "text", + "content": "Sign languages are the primary means of communication used by people who are deaf or hard of hearing. Sign language understanding [1, 10, 12-15, 18, 32, 33, 61, 73] is significant for overcoming the communication barrier between" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 318, + 232, + 535, + 296 + ], + "blocks": [ + { + "bbox": [ + 318, + 232, + 535, + 296 + ], + "lines": [ + { + "bbox": [ + 318, + 232, + 535, + 296 + ], + "spans": [ + { + "bbox": [ + 318, + 232, + 535, + 296 + ], + "type": "image", + "image_path": "01d2bcf4d4965d58c2397bc1dcf667b59d7f50fa64b981ec7ac2f54c7e5724a0.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 364, + 298, + 488, + 308 + ], + "lines": [ + { + "bbox": [ + 364, + 298, + 488, + 308 + ], + "spans": [ + { + "bbox": [ + 364, + 298, + 488, + 308 + ], + "type": "text", + "content": "(a) Text-to-sign-video (T2V) retrieval." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 317, + 311, + 533, + 372 + ], + "blocks": [ + { + "bbox": [ + 317, + 311, + 533, + 372 + ], + "lines": [ + { + "bbox": [ + 317, + 311, + 533, + 372 + ], + "spans": [ + { + "bbox": [ + 317, + 311, + 533, + 372 + ], + "type": "image", + "image_path": "ed8f51b1773276f9834cc368dd8f8e36453c53712644d3230dadfc7314559f1a.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 363, + 375, + 486, + 384 + ], + "lines": [ + { + "bbox": [ + 363, + 375, + 486, + 384 + ], + "spans": [ + { + "bbox": [ + 363, + 375, + 486, + 384 + ], + "type": "text", + "content": "(b) Sign-video-to-text (V2T) retrieval." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 315, + 394, + 536, + 406 + ], + "lines": [ + { + "bbox": [ + 315, + 394, + 536, + 406 + ], + "spans": [ + { + "bbox": [ + 315, + 394, + 536, + 406 + ], + "type": "text", + "content": "Figure 1. Illustration of: (a) T2V retrieval; (b) V2T retrieval." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 419, + 547, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 419, + 547, + 540 + ], + "spans": [ + { + "bbox": [ + 304, + 419, + 547, + 540 + ], + "type": "text", + "content": "the hard-of-hearing and non-signers. Sign language recognition and translation (SLRT) has been extensively studied, with the goal of recognizing the arbitrary semantic meanings conveyed by sign languages. However, the lack of available data significantly limits the capability of SLRT. In this paper, we focus on developing a framework for a recently proposed sign language retrieval task [18]. Unlike SLRT, sign language retrieval focuses on retrieving the meanings that signers express from a closed-set, which can significantly reduce error rates in realistic deployment." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 542, + 547, + 663 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 542, + 547, + 663 + ], + "spans": [ + { + "bbox": [ + 304, + 542, + 547, + 663 + ], + "type": "text", + "content": "Sign language retrieval is both similar to and distinct from the traditional video-text retrieval. On the one hand, like video-text retrieval, sign language retrieval is also composed of two sub-tasks, i.e., text-to-sign-video (T2V) retrieval and sign-video-to-text (V2T) retrieval. Given a free-form written query and a large collection of sign language videos, the objective of T2V is to find the video that best matches the written query (Figure 1a). In contrast, the goal of V2T is to identify the most relevant text description given a query of sign language video (Figure 1b)." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 665, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 547, + 715 + ], + "type": "text", + "content": "On the other hand, different from the video-text retrieval, sign languages, like most natural languages, have their own grammars and linguistic properties. Therefore, sign language videos not only contain visual signals, but also carry" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 702, + 135, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 702, + 135, + 712 + ], + "spans": [ + { + "bbox": [ + 58, + 702, + 135, + 712 + ], + "type": "text", + "content": "*Corresponding author." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "19016" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 88, + 274, + 210 + ], + "blocks": [ + { + "bbox": [ + 130, + 75, + 201, + 87 + ], + "lines": [ + { + "bbox": [ + 130, + 75, + 201, + 87 + ], + "spans": [ + { + "bbox": [ + 130, + 75, + 201, + 87 + ], + "type": "text", + "content": "(Sign Language)" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 58, + 88, + 274, + 210 + ], + "lines": [ + { + "bbox": [ + 58, + 88, + 274, + 210 + ], + "spans": [ + { + "bbox": [ + 58, + 88, + 274, + 210 + ], + "type": "image", + "image_path": "2351c5086303c77f8edced1b80e17be1ae97267ee1ade4ecd4ccbd38d0dd1498.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 52, + 218, + 279, + 267 + ], + "lines": [ + { + "bbox": [ + 52, + 218, + 279, + 267 + ], + "spans": [ + { + "bbox": [ + 52, + 218, + 279, + 267 + ], + "type": "text", + "content": "(a) While contrasting the sign videos and the texts in a joint embedding space, we simultaneously identify the fine-grained cross-lingual (sign-to-word) mappings of sign languages and natural languages via the proposed cross-lingual contrastive learning. Existing datasets do not annotate the sign-to-word mappings." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_footnote" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 58, + 270, + 165, + 348 + ], + "blocks": [ + { + "bbox": [ + 58, + 270, + 165, + 348 + ], + "lines": [ + { + "bbox": [ + 58, + 270, + 165, + 348 + ], + "spans": [ + { + "bbox": [ + 58, + 270, + 165, + 348 + ], + "type": "image", + "image_path": "9ee6e99e7787c128052cf6936f6c4c66facc18ea1a046dff0fad964a778d1cf0.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 52, + 353, + 279, + 383 + ], + "lines": [ + { + "bbox": [ + 52, + 353, + 279, + 383 + ], + "spans": [ + { + "bbox": [ + 52, + 353, + 279, + 383 + ], + "type": "text", + "content": "(b) We show four instances of the sign \"Book\" in How2Sign [19] dataset, which are identified by our approach. Please refer to the supplementary material for more examples." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 46, + 392, + 287, + 415 + ], + "lines": [ + { + "bbox": [ + 46, + 392, + 287, + 415 + ], + "spans": [ + { + "bbox": [ + 46, + 392, + 287, + 415 + ], + "type": "text", + "content": "Figure 2. Illustration of: (a) cross-lingual (sign-to-word) mapping; (b) sign-to-word mappings identified by our CiCo." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 167, + 270, + 274, + 348 + ], + "blocks": [ + { + "bbox": [ + 167, + 270, + 274, + 348 + ], + "lines": [ + { + "bbox": [ + 167, + 270, + 274, + 348 + ], + "spans": [ + { + "bbox": [ + 167, + 270, + 274, + 348 + ], + "type": "image", + "image_path": "fc11fdd1b44df7cb3628ab8f99db8c3b472c52356ddeae90c99d865bd42110a7.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 421, + 287, + 504 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 421, + 287, + 504 + ], + "spans": [ + { + "bbox": [ + 46, + 421, + 287, + 504 + ], + "type": "text", + "content": "semantics (i.e., sign1-to-word mappings between sign languages and natural languages) by themselves, which differentiates them from the general videos that merely contain visual information. Considering the linguistic characteristics of sign languages, we formulate sign language retrieval as a cross-lingual retrieval [6,34,59] problem in addition to a video-text retrieval [5,24,42-44,57,68] task." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 505, + 288, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 505, + 288, + 696 + ], + "spans": [ + { + "bbox": [ + 46, + 505, + 288, + 696 + ], + "type": "text", + "content": "Sign language retrieval is extremely challenging due to the following reasons: (1) Sign languages are completely separate and distinct from natural languages since they have unique linguistic rules, word formation, and word order. The transcription between sign languages and natural languages is complicated, for instance, the word order is typically not preserved between sign languages and natural languages. It is necessary to automatically identify the sign-to-word mapping from the cross-lingual retrieval perspective; (2) In contrast to the text-video retrieval datasets [45, 48] which contain millions of training samples, sign language datasets are orders of magnitude smaller in scale—for example, there are only " + }, + { + "bbox": [ + 46, + 505, + 288, + 696 + ], + "type": "inline_equation", + "content": "30\\mathrm{K}" + }, + { + "bbox": [ + 46, + 505, + 288, + 696 + ], + "type": "text", + "content": " video-text pairs in How2Sign [19] training set; (3) Sign languages convey information through the handshape, facial expression, and body movement, which requires models to distinguish fine" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 72, + 545, + 132 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 545, + 132 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 545, + 132 + ], + "type": "text", + "content": "grained gestures and actions; (4) Sign language videos typically contain hundreds of frames. It is necessary to build efficient algorithms to lower the training cost and fit the long videos as well as the intermediate representations into limited GPU memory." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 133, + 545, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 133, + 545, + 156 + ], + "spans": [ + { + "bbox": [ + 306, + 133, + 545, + 156 + ], + "type": "text", + "content": "In this work, we concentrate on resolving the challenges listed above:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 161, + 545, + 535 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 306, + 161, + 545, + 281 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 161, + 545, + 281 + ], + "spans": [ + { + "bbox": [ + 306, + 161, + 545, + 281 + ], + "type": "text", + "content": "- We consider the linguistic rules (e.g., word order) of both sign languages and natural languages. We formulate sign language retrieval as a cross-lingual retrieval task as well as a video-text retrieval problem. While contrasting the sign videos and the texts in a joint embedding space as achieved in most vision-language pre-training frameworks [5, 44, 57], we simultaneously identify the fine-grained cross-lingual (sign-to-word) mappings between two types of languages via our proposed cross-lingual contrastive learning as shown in Figure 2." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 288, + 545, + 444 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 288, + 545, + 444 + ], + "spans": [ + { + "bbox": [ + 306, + 288, + 545, + 444 + ], + "type": "text", + "content": "- Data scarcity typically brings in the over-fitting issue. To alleviate this issue, we adopt transfer learning and adapt a recently released domain-agnostic sign encoder [61] pre-trained on large-scale sign-videos to the target domain. Although this encoder is capable of distinguishing the fine-grained signs, direct transferring may be suboptimal due to the unavoidable domain gap between the pre-training dataset and sign language retrieval datasets. To tackle this problem, we further fine-tune a domain-aware sign encoder on pseudo-labeled data from target datasets. The final sign encoder is composed of the well-optimized domain-aware sign encoder and the powerful domain-agnostic sign encoder." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 450, + 545, + 535 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 450, + 545, + 535 + ], + "spans": [ + { + "bbox": [ + 306, + 450, + 545, + 535 + ], + "type": "text", + "content": "- In order to effectively model long videos, we decouple our framework into two disjoint parts: (1) a sign encoder which adopts a sliding window on sign-videos to pre-extract their vision features; (2) a cross-lingual contrastive learning module which encodes the extracted vision features and their corresponding texts in a joint embedding space." + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 304, + 540, + 545, + 659 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 540, + 545, + 659 + ], + "spans": [ + { + "bbox": [ + 304, + 540, + 545, + 659 + ], + "type": "text", + "content": "Our framework, called domain-aware sign language retrieval via Cross-lingual Contrastive learning or CiCo for short, outperforms the pioneer SPOT-ALIGN [18] by large margins on various datasets, achieving 56.6 (+22.4) T2V and 51.6 (+28.0) V2T R@1 accuracy (improvement) on How2Sign [19] dataset, and 69.5 (+13.7) T2V and 70.2 (+17.1) V2T R@1 accuracy (improvement) on PHOENIX-2014T [8] dataset. With its simplicity and strong performance, we hope our approach can serve as a solid baseline for future research." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 669, + 392, + 681 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 669, + 392, + 681 + ], + "spans": [ + { + "bbox": [ + 306, + 669, + 392, + 681 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "content": "Sign Language Understanding. Sign language understanding aims at interpreting the semantic information con" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 58, + 702, + 286, + 714 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 702, + 286, + 714 + ], + "spans": [ + { + "bbox": [ + 58, + 702, + 286, + 714 + ], + "type": "text", + "content": "1We use sign to denote lexical item within a sign language vocabulary." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19017" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 288, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 288, + 133 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 288, + 133 + ], + "type": "text", + "content": "veyed within sign videos. Researchers have explored such capability on various tasks including sign language recognition [13,20,21,30,38,58,73], sign spotting [1,18,47,61], sign language translation [8,9,12,37,72] and our focused sign language retrieval [18]." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 134, + 287, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 134, + 287, + 266 + ], + "spans": [ + { + "bbox": [ + 46, + 134, + 287, + 266 + ], + "type": "text", + "content": "One of the fundamental tasks of sign language understanding is sign language recognition (SLR), which aims to transcribe a sign video into a gloss sequence. Previous works on SLR focus on designing carefully engineered features [20, 21, 58] or modeling temporal dependencies [55, 56]. Recently, the success of 3D convolutional neural networks in action related tasks [41, 60, 63] is transferred to SLR. In particular, the I3D [11] architecture has proven to be effective for this task [1, 30, 36, 38, 61]. In this work, we also adopt this network architecture in our sign encoder." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 267, + 288, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 267, + 288, + 445 + ], + "spans": [ + { + "bbox": [ + 46, + 267, + 288, + 445 + ], + "type": "text", + "content": "Sign spotting is a particular variant of sign language recognition. It aims to localize all instances of a given sign within an untrimmed video. Recent works tackle this task with auxiliary cue of subtitles, introducing automatic annotation systems by using mouthing [1], dictionaries [47] and attention maps of Transformer [61]. SPOT-ALIGN [18] extends existing spotting methods [1, 47] with an iterative training schema, which alternates between repeated sign spotting and model fine-tuning. In this work, pseudo-labeling is served as our sign spotting approach to localize isolated signs in untrimmed videos from target sign language retrieval datasets. Compared with above efforts, our approach only employs a pre-trained sign encoder without utilizing additional auxiliary cue, which is proved to be simple yet efficient." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 449, + 287, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 449, + 287, + 604 + ], + "spans": [ + { + "bbox": [ + 46, + 449, + 287, + 604 + ], + "type": "text", + "content": "Early works of sign language retrieval primarily investigate query-by-example searching [4, 70], which queries individual instances with given sign examples. Our work focuses on free-form textual retrieval—a recently introduced task by SPOT-ALIGN [18]. It symbolizes the real-world scenario of searching sign language videos with natural languages. The pioneer SPOT-ALIGN [18] purely formulates sign language retrieval as a video-text retrieval task, where the cross-modal alignment is modeled upon overall global embeddings of sign videos and texts. However, the linguistic properties of sign languages are ignored. In contrast, we formulate sign language retrieval as a joint task of text-video retrieval and cross-lingual retrieval." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 605, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 288, + 713 + ], + "type": "text", + "content": "Vision-Language Models. Learning general-purpose representations for visual and textual modalities is a longstanding topic [7, 35]. The idea has been investigated decades ago [49]. Recently, the prominent success of CLIP [51], ALIGN [28] and ALBEF [39] has demonstrated the capability of learning joint cross-modal representations with large-scale web data using simple image-text contrastive learning. Similar idea has also been explored in video-text retrieval, which is the closest area of our work." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 72, + 545, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 204 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 204 + ], + "type": "text", + "content": "The common practice in image-text retrieval [23, 25, 29, 51] and video-text retrieval [5, 24, 42-44, 57, 68] is to encode the images/videos and texts into the overall representations. It is reasonable since the visual signals of typical image-text and video-text retrieval datasets mainly describe the certain objects or describable events. In contrast, sign videos, as the carriers of sign languages, convey abundant semantics by themselves. There exist fine-grained mappings between sign videos and natural languages. We find that identifying such cross-lingual (sign-to-word) mappings significantly boosts the performance of sign language retrieval." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 205, + 545, + 431 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 205, + 545, + 431 + ], + "spans": [ + { + "bbox": [ + 304, + 205, + 545, + 431 + ], + "type": "text", + "content": "Cross-Lingual Information Retrieval. In this work, we also formulate sign-language retrieval as a cross-lingual task. In natural language processing community, cross-lingual information retrieval (CLIR) [34, 59, 65] refers to the task of retrieving documents between different languages. One of the most common approaches of CLIR is to learn sentence-level embedding alignment by mapping pre-acquired monolingual embeddings [46, 50, 69] of different languages into a shared space. Recently, researchers have exploited fine-grained word-level mappings with self-training [2, 3, 59]. They find that mappings can be learned by initiating with a seed dictionary and alternating between alignment modeling and dictionary mining. Our work also exploits fine-grained word-level mappings, called sign-to-word mappings, in the context of sign language retrieval. Identifying sign-to-word mappings is challenging since sign languages are expressed in visual modality. The proposed cross-lingual contrastive learning tackles this challenge by exploiting the fine-grained cross-modal interactions." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 441, + 388, + 454 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 441, + 388, + 454 + ], + "spans": [ + { + "bbox": [ + 306, + 441, + 388, + 454 + ], + "type": "text", + "content": "3. Methodology" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 461, + 545, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 461, + 545, + 568 + ], + "spans": [ + { + "bbox": [ + 304, + 461, + 545, + 568 + ], + "type": "text", + "content": "In this section, we first formulate the task of sign language retrieval in Section 3.1. Next, we introduce our framework which is composed of two parts: 1) a sign encoder which extracts discriminative and domain-aligned features of sign videos (Section 3.2); 2) a cross-lingual contrastive learning framework, which contrasts sign-video-text pairs while concurrently identifying the fine-grained sign-to-word mappings (Section 3.3). At last, we explore text augmentations in Section 3.4." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 575, + 411, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 575, + 411, + 586 + ], + "spans": [ + { + "bbox": [ + 306, + 575, + 411, + 586 + ], + "type": "text", + "content": "3.1. Task Formulation" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "text", + "content": "Let " + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{V}" + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{T}" + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "text", + "content": " denote a set of sign videos and their corresponding texts (transcriptions), respectively. Sign language retrieval consists of two tasks namely text-to-sign-video retrieval (T2V), and sign-video-to-text retrieval (V2T), respectively. The objective of T2V is to find the sign video " + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "inline_equation", + "content": "v\\in \\mathcal{V}" + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "text", + "content": " whose signing content best matches the text query. In contrast, the reverse task V2T requires the model to identify the most relevant text (transcription) " + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "inline_equation", + "content": "t\\in \\mathcal{T}" + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "text", + "content": " given a query of sign video. We resolve sign language retrieval by learning a joint embedding space of sign videos and texts." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19018" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 61, + 71, + 532, + 241 + ], + "blocks": [ + { + "bbox": [ + 61, + 71, + 532, + 241 + ], + "lines": [ + { + "bbox": [ + 61, + 71, + 532, + 241 + ], + "spans": [ + { + "bbox": [ + 61, + 71, + 532, + 241 + ], + "type": "image", + "image_path": "b9c7644175c01d5251b0ebf06b002ac5650564f65aa85bae99313a5ceeff1e0e.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 247, + 245, + 347, + 255 + ], + "lines": [ + { + "bbox": [ + 247, + 245, + 347, + 255 + ], + "spans": [ + { + "bbox": [ + 247, + 245, + 347, + 255 + ], + "type": "text", + "content": "(a) Illustration of sign encoder." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 62, + 259, + 532, + 454 + ], + "blocks": [ + { + "bbox": [ + 62, + 259, + 532, + 454 + ], + "lines": [ + { + "bbox": [ + 62, + 259, + 532, + 454 + ], + "spans": [ + { + "bbox": [ + 62, + 259, + 532, + 454 + ], + "type": "image", + "image_path": "177a9fdd10ab5bc16f025d961267330e3e5b63ba7832acf7ea46d6dcf0990232.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 212, + 458, + 379, + 469 + ], + "lines": [ + { + "bbox": [ + 212, + 458, + 379, + 469 + ], + "spans": [ + { + "bbox": [ + 212, + 458, + 379, + 469 + ], + "type": "text", + "content": "(b) Illustration of cross-lingual contrastive learning." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 45, + 478, + 545, + 534 + ], + "lines": [ + { + "bbox": [ + 45, + 478, + 545, + 534 + ], + "spans": [ + { + "bbox": [ + 45, + 478, + 545, + 534 + ], + "type": "text", + "content": "Figure 3. Overview of our framework. (a) Sign encoder is composed of a powerful domain-agnostic sign encoder pre-trained on large-scale sign videos, and a domain-aware sign encoder fine-tuned on pseudo-labeled data from target datasets. We adopt a sliding window manner to extract a discriminative and domain-aligned feature per clip. (b) Cross-lingual contrastive learning takes " + }, + { + "bbox": [ + 45, + 478, + 545, + 534 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 45, + 478, + 545, + 534 + ], + "type": "text", + "content": " sign-video-text pairs as inputs and contrasts paired data in a shared embedding space while implicitly identifying the fine-grained sign-to-word mappings during training." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 540, + 133, + 553 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 540, + 133, + 553 + ], + "spans": [ + { + "bbox": [ + 47, + 540, + 133, + 553 + ], + "type": "text", + "content": "3.2. Sign Encoder" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 563, + 288, + 684 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 563, + 288, + 684 + ], + "spans": [ + { + "bbox": [ + 46, + 563, + 288, + 684 + ], + "type": "text", + "content": "Process Sign Videos with Sliding Window. Sign videos from sign language retrieval datasets typically contain hundreds of frames. To efficiently train our model and lower the usage of GPU memory, given a sign video, we adopt a sliding window manner with stride of 1 and window size of 16 to produce " + }, + { + "bbox": [ + 46, + 563, + 288, + 684 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 46, + 563, + 288, + 684 + ], + "type": "text", + "content": " temporally overlapped clips. Next, we separately feed each clip into a sign encoder to extract its feature. The final sign-video feature is yielded by stacking features coming out of " + }, + { + "bbox": [ + 46, + 563, + 288, + 684 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 46, + 563, + 288, + 684 + ], + "type": "text", + "content": " clips along temporal dimension. A powerful sign encoder is crucial." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 689, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 689, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 689, + 288, + 715 + ], + "type": "text", + "content": "Overview of Sign Encoder. Recent advances in sign spotting [47, 61] greatly facilitate the collection of large-scale" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 540, + 545, + 708 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 540, + 545, + 708 + ], + "spans": [ + { + "bbox": [ + 304, + 540, + 545, + 708 + ], + "type": "text", + "content": "sign language datasets, enabling powerful representation learning abilities of convolutional neural networks on the sign classification task. Previous methods [16, 18] have demonstrated the feasibility of transferring a sign encoder pre-trained on large-scale sign-spotting data into downstream tasks. We follow this practice and use an I3D network pre-trained on BSL-1K [61], a sign classification dataset collected via sign spotting, as our primary sign encoder. Due to its favorable transfer performance, we term this model as a domain-agnostic sign encoder. Nevertheless, the domain gap between BSL-1K and sign language retrieval datasets is non-negligible. To tackle this problem, we further fine-tune a domain-aware sign encoder, which has an identical architecture to the domain-agnostic sign en" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19019" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "text", + "content": "coder, on target datasets through pseudo-labeling. The final sign encoder is composed of the well-optimized domain-aware sign encoder and the powerful domain-agnostic sign encoder, as illustrated in Figure 3a." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 121, + 288, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 121, + 288, + 289 + ], + "spans": [ + { + "bbox": [ + 46, + 121, + 288, + 289 + ], + "type": "text", + "content": "Pseudo-Labeling on Target Datasets. Now we describe the details of pseudo-labeling. Given a sign video from a target dataset, we adopt a sliding window with the stride of 1 and the window size of 16 to generate a set of temporally overlapped clips. For each clip, we first utilize the pretrained domain-agnostic sign encoder to produce its prediction. Then we binarize the prediction with a pre-defined threshold " + }, + { + "bbox": [ + 46, + 121, + 288, + 289 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 46, + 121, + 288, + 289 + ], + "type": "text", + "content": " to generate the corresponding pseudo label. The invalid samples, whose maximum score is lower than " + }, + { + "bbox": [ + 46, + 121, + 288, + 289 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 46, + 121, + 288, + 289 + ], + "type": "text", + "content": ", are filtered. We repeat the above process for all sign videos and eventually build a pseudo-labeled set. Our domain-aware sign encoder, which is initialized by the domain-agnostic sign encoder, is fine-tuned on the pseudo-labeled set via a standard cross-entropy loss for classification training." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 289, + 288, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 289, + 288, + 468 + ], + "spans": [ + { + "bbox": [ + 46, + 289, + 288, + 468 + ], + "type": "text", + "content": "Feature Extraction with Sign Encoder. So far, we acquire a domain-aware sign encoder approximately aligned in target domain. Nevertheless, its capability is restricted by the unavoidable noises in pseudo-labels and the limited amount of pseudo-labeled samples. Recall that we already have a powerful domain-agnostic sign encoder pre-trained on large-scale dataset in hand, inspiring us to make use of both domain-agnostic sign encoder " + }, + { + "bbox": [ + 46, + 289, + 288, + 468 + ], + "type": "inline_equation", + "content": "h_{\\xi}(\\cdot)" + }, + { + "bbox": [ + 46, + 289, + 288, + 468 + ], + "type": "text", + "content": " and domain-aware sign encoder " + }, + { + "bbox": [ + 46, + 289, + 288, + 468 + ], + "type": "inline_equation", + "content": "h_{\\theta}(\\cdot)" + }, + { + "bbox": [ + 46, + 289, + 288, + 468 + ], + "type": "text", + "content": " to extract discriminative and domain-aligned features. Our final sign encoder " + }, + { + "bbox": [ + 46, + 289, + 288, + 468 + ], + "type": "inline_equation", + "content": "H(\\cdot)" + }, + { + "bbox": [ + 46, + 289, + 288, + 468 + ], + "type": "text", + "content": ", as shown in Figure 3a, is a weighted combination of " + }, + { + "bbox": [ + 46, + 289, + 288, + 468 + ], + "type": "inline_equation", + "content": "h_{\\xi}(\\cdot)" + }, + { + "bbox": [ + 46, + 289, + 288, + 468 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 289, + 288, + 468 + ], + "type": "inline_equation", + "content": "h_{\\theta}(\\cdot)" + }, + { + "bbox": [ + 46, + 289, + 288, + 468 + ], + "type": "text", + "content": " with a trade-off hyper-parameter " + }, + { + "bbox": [ + 46, + 289, + 288, + 468 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 46, + 289, + 288, + 468 + ], + "type": "text", + "content": ". As described above, " + }, + { + "bbox": [ + 46, + 289, + 288, + 468 + ], + "type": "inline_equation", + "content": "H(\\cdot)" + }, + { + "bbox": [ + 46, + 289, + 288, + 468 + ], + "type": "text", + "content": " encodes sign videos in a sliding window manner. For simplicity, we use " + }, + { + "bbox": [ + 46, + 289, + 288, + 468 + ], + "type": "inline_equation", + "content": "H(v)" + }, + { + "bbox": [ + 46, + 289, + 288, + 468 + ], + "type": "text", + "content": " to denote feature extraction on sign video " + }, + { + "bbox": [ + 46, + 289, + 288, + 468 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 46, + 289, + 288, + 468 + ], + "type": "text", + "content": ", which is formulated as:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 98, + 479, + 287, + 492 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 98, + 479, + 287, + 492 + ], + "spans": [ + { + "bbox": [ + 98, + 479, + 287, + 492 + ], + "type": "interline_equation", + "content": "H (v) = \\alpha h _ {\\xi} (v) + (1 - \\alpha) h _ {\\theta} (v). \\tag {1}", + "image_path": "7dcfd7667335c6915eac0ee5c097c28639f7175cf0c6fef147147875e598fd59.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 502, + 241, + 515 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 502, + 241, + 515 + ], + "spans": [ + { + "bbox": [ + 47, + 502, + 241, + 515 + ], + "type": "text", + "content": "3.3. Cross-Linguual Contrastive Learning" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 521, + 287, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 521, + 287, + 616 + ], + "spans": [ + { + "bbox": [ + 46, + 521, + 287, + 616 + ], + "type": "text", + "content": "The objective of cross-lingual contrastive learning (CLCL) is to learn a joint embedding space of sign videos and texts while concurrently identifying the fine-grained sign-to-word mappings during training. An overview is shown in Figure 3b. CLCL takes a mini-batch " + }, + { + "bbox": [ + 46, + 521, + 287, + 616 + ], + "type": "inline_equation", + "content": "\\{(v_{n},t_{n})\\}_{n = 1}^{N}" + }, + { + "bbox": [ + 46, + 521, + 287, + 616 + ], + "type": "text", + "content": " containing " + }, + { + "bbox": [ + 46, + 521, + 287, + 616 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 521, + 287, + 616 + ], + "type": "text", + "content": " sign-video-text pairs as input, and contrasts paired data in a shared embedding space for sign language retrieval." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 617, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 617, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 617, + 287, + 713 + ], + "type": "text", + "content": "Sign Features and Word Features. Given a sign video " + }, + { + "bbox": [ + 46, + 617, + 287, + 713 + ], + "type": "inline_equation", + "content": "v \\in \\{v_n\\}_{n=1}^N" + }, + { + "bbox": [ + 46, + 617, + 287, + 713 + ], + "type": "text", + "content": ", we first adopt our sign encoder " + }, + { + "bbox": [ + 46, + 617, + 287, + 713 + ], + "type": "inline_equation", + "content": "H(\\cdot)" + }, + { + "bbox": [ + 46, + 617, + 287, + 713 + ], + "type": "text", + "content": " described in Section 3.2 to extract its intermediate feature. Note " + }, + { + "bbox": [ + 46, + 617, + 287, + 713 + ], + "type": "inline_equation", + "content": "H(\\cdot)" + }, + { + "bbox": [ + 46, + 617, + 287, + 713 + ], + "type": "text", + "content": " encodes sign videos in a sliding window manner, and thus there are no interactions among different clips. To facilitate information exchange, we further append a 12-layer Transformer [62] " + }, + { + "bbox": [ + 46, + 617, + 287, + 713 + ], + "type": "inline_equation", + "content": "F(\\cdot)" + }, + { + "bbox": [ + 46, + 617, + 287, + 713 + ], + "type": "text", + "content": " onto " + }, + { + "bbox": [ + 46, + 617, + 287, + 713 + ], + "type": "inline_equation", + "content": "H(\\cdot)" + }, + { + "bbox": [ + 46, + 617, + 287, + 713 + ], + "type": "text", + "content": " to extract sign features " + }, + { + "bbox": [ + 46, + 617, + 287, + 713 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 46, + 617, + 287, + 713 + ], + "type": "text", + "content": " of sign video " + }, + { + "bbox": [ + 46, + 617, + 287, + 713 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 46, + 617, + 287, + 713 + ], + "type": "text", + "content": ", which is formulated" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "text", + "content": "as " + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "inline_equation", + "content": "S = F(H(v)) \\in \\mathbb{R}^{M \\times D}" + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "text", + "content": " denotes the number of clips, and " + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "text", + "content": " is the hidden dimension. Given a text " + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "inline_equation", + "content": "t \\in \\{t_n\\}_{n=1}^N" + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "text", + "content": ", we convert " + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "text", + "content": " into a lower-cased byte pair encoding (BPE) representation [53], which is subsequently fed into another 12-layer Transformer " + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "inline_equation", + "content": "G(\\cdot)" + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "text", + "content": " to generate the word features " + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "inline_equation", + "content": "W = G(t) \\in \\mathbb{R}^{L \\times D}" + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "text", + "content": " represents word number." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 158, + 545, + 243 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 158, + 545, + 243 + ], + "spans": [ + { + "bbox": [ + 304, + 158, + 545, + 243 + ], + "type": "text", + "content": "Since CLIP [51] shows excellent transfer capability in various downstream tasks [17, 22, 27, 54, 66, 67], we initialize " + }, + { + "bbox": [ + 304, + 158, + 545, + 243 + ], + "type": "inline_equation", + "content": "F(\\cdot)" + }, + { + "bbox": [ + 304, + 158, + 545, + 243 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 158, + 545, + 243 + ], + "type": "inline_equation", + "content": "G(\\cdot)" + }, + { + "bbox": [ + 304, + 158, + 545, + 243 + ], + "type": "text", + "content": " with CLIP's image encoder (ViT-B) and text encoder, respectively, to ease the learning. Though CLIP's vision encoder takes image patches as inputs, we experimentally find that it generalizes well in our scenario where input data is in a different modality." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 245, + 545, + 366 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 245, + 545, + 366 + ], + "spans": [ + { + "bbox": [ + 304, + 245, + 545, + 366 + ], + "type": "text", + "content": "Cross-Lingual Similarity. There exist inherent sign-to-word mappings between sign languages and natural languages. To incorporate this prior knowledge into learning, we introduce cross-lingual similarity—an indicator to identify sign-to-word mappings between " + }, + { + "bbox": [ + 304, + 245, + 545, + 366 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 245, + 545, + 366 + ], + "type": "text", + "content": "-th sign video " + }, + { + "bbox": [ + 304, + 245, + 545, + 366 + ], + "type": "inline_equation", + "content": "v_{i}" + }, + { + "bbox": [ + 304, + 245, + 545, + 366 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 245, + 545, + 366 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 304, + 245, + 545, + 366 + ], + "type": "text", + "content": "-th text " + }, + { + "bbox": [ + 304, + 245, + 545, + 366 + ], + "type": "inline_equation", + "content": "t_j" + }, + { + "bbox": [ + 304, + 245, + 545, + 366 + ], + "type": "text", + "content": ". Concretely, given sign features " + }, + { + "bbox": [ + 304, + 245, + 545, + 366 + ], + "type": "inline_equation", + "content": "\\pmb{S}_i \\in \\mathbb{R}^{M \\times D}" + }, + { + "bbox": [ + 304, + 245, + 545, + 366 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 304, + 245, + 545, + 366 + ], + "type": "inline_equation", + "content": "v_i" + }, + { + "bbox": [ + 304, + 245, + 545, + 366 + ], + "type": "text", + "content": ", and word features " + }, + { + "bbox": [ + 304, + 245, + 545, + 366 + ], + "type": "inline_equation", + "content": "\\pmb{W}_j \\in \\mathbb{R}^{L \\times D}" + }, + { + "bbox": [ + 304, + 245, + 545, + 366 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 304, + 245, + 545, + 366 + ], + "type": "inline_equation", + "content": "t_j" + }, + { + "bbox": [ + 304, + 245, + 545, + 366 + ], + "type": "text", + "content": ", we calculate a cross-lingual similarity matrix " + }, + { + "bbox": [ + 304, + 245, + 545, + 366 + ], + "type": "inline_equation", + "content": "\\pmb{E}_{(i,j)} = \\pmb{S}_i \\cdot \\pmb{W}_j^T \\in \\mathbb{R}^{M \\times L}" + }, + { + "bbox": [ + 304, + 245, + 545, + 366 + ], + "type": "text", + "content": ". Each element in " + }, + { + "bbox": [ + 304, + 245, + 545, + 366 + ], + "type": "inline_equation", + "content": "\\pmb{E}_{(i,j)}" + }, + { + "bbox": [ + 304, + 245, + 545, + 366 + ], + "type": "text", + "content": " represents the similarity of a sign clip in " + }, + { + "bbox": [ + 304, + 245, + 545, + 366 + ], + "type": "inline_equation", + "content": "v_i" + }, + { + "bbox": [ + 304, + 245, + 545, + 366 + ], + "type": "text", + "content": " and a word in " + }, + { + "bbox": [ + 304, + 245, + 545, + 366 + ], + "type": "inline_equation", + "content": "t_j" + }, + { + "bbox": [ + 304, + 245, + 545, + 366 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 368, + 545, + 465 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 368, + 545, + 465 + ], + "spans": [ + { + "bbox": [ + 304, + 368, + 545, + 465 + ], + "type": "text", + "content": "Cross-Lingual Contrastive Learning. Directly apply supervisions on token-wise similarity matrix " + }, + { + "bbox": [ + 304, + 368, + 545, + 465 + ], + "type": "inline_equation", + "content": "\\pmb{E}_{(i,j)}" + }, + { + "bbox": [ + 304, + 368, + 545, + 465 + ], + "type": "text", + "content": " is infeasible due to the absence of fine-grained sign-to-word annotations. Inspired by the recent progress of vision-language contrastive learning [23, 25, 29, 39, 40, 51], we turn to contrast the global representations of sign videos and texts. The underlying idea is to calculate a global similarity " + }, + { + "bbox": [ + 304, + 368, + 545, + 465 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 304, + 368, + 545, + 465 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 304, + 368, + 545, + 465 + ], + "type": "inline_equation", + "content": "v_{i}" + }, + { + "bbox": [ + 304, + 368, + 545, + 465 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 368, + 545, + 465 + ], + "type": "inline_equation", + "content": "t_j" + }, + { + "bbox": [ + 304, + 368, + 545, + 465 + ], + "type": "text", + "content": " based on " + }, + { + "bbox": [ + 304, + 368, + 545, + 465 + ], + "type": "inline_equation", + "content": "\\pmb{E}_{(i,j)} \\in \\mathbb{R}^{M \\times L}" + }, + { + "bbox": [ + 304, + 368, + 545, + 465 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 466, + 545, + 612 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 466, + 545, + 612 + ], + "spans": [ + { + "bbox": [ + 304, + 466, + 545, + 612 + ], + "type": "text", + "content": "Sign-Video-to-Text Contrast. We first introduce sign-video-to-text contrast as shown in Figure 3b. To be specific, we utilize a Softmax operation to each row of " + }, + { + "bbox": [ + 304, + 466, + 545, + 612 + ], + "type": "inline_equation", + "content": "\\pmb{E}_{(i,j)}" + }, + { + "bbox": [ + 304, + 466, + 545, + 612 + ], + "type": "text", + "content": " and multiply the resulting matrix with " + }, + { + "bbox": [ + 304, + 466, + 545, + 612 + ], + "type": "inline_equation", + "content": "\\pmb{E}_{(i,j)}" + }, + { + "bbox": [ + 304, + 466, + 545, + 612 + ], + "type": "text", + "content": " to generate a sign-to-word similarity matrix " + }, + { + "bbox": [ + 304, + 466, + 545, + 612 + ], + "type": "inline_equation", + "content": "\\pmb{E}_{(i,j)}' \\in \\mathbb{R}^{M \\times L}" + }, + { + "bbox": [ + 304, + 466, + 545, + 612 + ], + "type": "text", + "content": ", where each row represents the re-weighted similarities between a sign clip in " + }, + { + "bbox": [ + 304, + 466, + 545, + 612 + ], + "type": "inline_equation", + "content": "v_{i}" + }, + { + "bbox": [ + 304, + 466, + 545, + 612 + ], + "type": "text", + "content": " and all words in " + }, + { + "bbox": [ + 304, + 466, + 545, + 612 + ], + "type": "inline_equation", + "content": "t_j" + }, + { + "bbox": [ + 304, + 466, + 545, + 612 + ], + "type": "text", + "content": ". After that, we adopt a row-wise addition operation on " + }, + { + "bbox": [ + 304, + 466, + 545, + 612 + ], + "type": "inline_equation", + "content": "\\pmb{E}_{(i,j)}'" + }, + { + "bbox": [ + 304, + 466, + 545, + 612 + ], + "type": "text", + "content": " to yield the sign-to-text similarity vector " + }, + { + "bbox": [ + 304, + 466, + 545, + 612 + ], + "type": "inline_equation", + "content": "e_{(i,j)} \\in \\mathbb{R}^M" + }, + { + "bbox": [ + 304, + 466, + 545, + 612 + ], + "type": "text", + "content": ", where each element denotes a similarity of a sign clip in " + }, + { + "bbox": [ + 304, + 466, + 545, + 612 + ], + "type": "inline_equation", + "content": "v_{i}" + }, + { + "bbox": [ + 304, + 466, + 545, + 612 + ], + "type": "text", + "content": " and whole text " + }, + { + "bbox": [ + 304, + 466, + 545, + 612 + ], + "type": "inline_equation", + "content": "t_j" + }, + { + "bbox": [ + 304, + 466, + 545, + 612 + ], + "type": "text", + "content": ". At last, we average all elements in " + }, + { + "bbox": [ + 304, + 466, + 545, + 612 + ], + "type": "inline_equation", + "content": "e_{(i,j)}" + }, + { + "bbox": [ + 304, + 466, + 545, + 612 + ], + "type": "text", + "content": " to produce the global similarity " + }, + { + "bbox": [ + 304, + 466, + 545, + 612 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 304, + 466, + 545, + 612 + ], + "type": "text", + "content": " of sign video " + }, + { + "bbox": [ + 304, + 466, + 545, + 612 + ], + "type": "inline_equation", + "content": "v_{i}" + }, + { + "bbox": [ + 304, + 466, + 545, + 612 + ], + "type": "text", + "content": " and text " + }, + { + "bbox": [ + 304, + 466, + 545, + 612 + ], + "type": "inline_equation", + "content": "t_j" + }, + { + "bbox": [ + 304, + 466, + 545, + 612 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 614, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 614, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 614, + 545, + 713 + ], + "type": "text", + "content": "In the same way, we can calculate the similarities for both positive pairs " + }, + { + "bbox": [ + 304, + 614, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\{(v_i,t_i)\\}_{i = 1}^N" + }, + { + "bbox": [ + 304, + 614, + 545, + 713 + ], + "type": "text", + "content": " and negative pairs " + }, + { + "bbox": [ + 304, + 614, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\{(v_i,t_j)\\}_{i = 1,j = 1,i\\neq j}^N" + }, + { + "bbox": [ + 304, + 614, + 545, + 713 + ], + "type": "text", + "content": " in a mini-batch, yielding a video-to-text similarity matrix " + }, + { + "bbox": [ + 304, + 614, + 545, + 713 + ], + "type": "inline_equation", + "content": "Z_{V2T}\\in \\mathbb{R}^{N\\times N}" + }, + { + "bbox": [ + 304, + 614, + 545, + 713 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 614, + 545, + 713 + ], + "type": "inline_equation", + "content": "Z_{V2T}^{(i,j)}" + }, + { + "bbox": [ + 304, + 614, + 545, + 713 + ], + "type": "text", + "content": " denotes the global similarity of " + }, + { + "bbox": [ + 304, + 614, + 545, + 713 + ], + "type": "inline_equation", + "content": "v_{i}" + }, + { + "bbox": [ + 304, + 614, + 545, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 614, + 545, + 713 + ], + "type": "inline_equation", + "content": "t_j" + }, + { + "bbox": [ + 304, + 614, + 545, + 713 + ], + "type": "text", + "content": ". Following CLIP [51], we adopt InfoNCE loss [26] to pull the embeddings of matched image-text pairs together while pushing those of non-matched pairs apart, which is formulated as" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19020" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 82, + 83 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 82, + 83 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 82, + 83 + ], + "type": "text", + "content": "follows:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 74, + 91, + 287, + 161 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 74, + 91, + 287, + 161 + ], + "spans": [ + { + "bbox": [ + 74, + 91, + 287, + 161 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} _ {V 2 T} = - \\frac {1}{2 N} \\sum_ {\\substack {i = 1 \\\\ N}} ^ {N} \\log \\frac {\\exp \\left(\\boldsymbol {Z} _ {V 2 T} ^ {(i , j)} / \\tau\\right)}{\\sum_ {j = 1} ^ {N} \\exp \\left(\\boldsymbol {Z} _ {V 2 T} ^ {(i , j)} / \\tau\\right)} \\tag{2} \\\\ - \\frac {1}{2 N} \\sum_ {j = 1} ^ {N} \\log \\frac {\\exp \\left(\\mathbf {Z} _ {V 2 T} ^ {(i , j)} / \\tau\\right)}{\\sum_ {i = 1} ^ {N} \\exp \\left(\\mathbf {Z} _ {V 2 T} ^ {(i , j)} / \\tau\\right)}, \\\\ \\end{array}", + "image_path": "ca1763d1ae0e35b87a0837f69ef68b7e27b585b9c4cf7d283bf1174f384e9c24.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 171, + 230, + 182 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 171, + 230, + 182 + ], + "spans": [ + { + "bbox": [ + 47, + 171, + 230, + 182 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 171, + 230, + 182 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 47, + 171, + 230, + 182 + ], + "type": "text", + "content": " is a trainable temperature parameter." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 182, + 287, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 182, + 287, + 312 + ], + "spans": [ + { + "bbox": [ + 46, + 182, + 287, + 312 + ], + "type": "text", + "content": "Text-to-Sign-Video Contrast. Up to now, we have introduced sign-video-to-text contrast. A symmetrical version, termed text-to-sign-video contrast, shares the similar spirit as shown in Figure 3b. The implementation of text-to-sign-video contrast is extremely simple: we replace the row-wise operations (i.e., Softmax and addition) in sign-video-to-text contrast with column-wise ones and keep the remaining processes unchanged. We use " + }, + { + "bbox": [ + 46, + 182, + 287, + 312 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{T2V}" + }, + { + "bbox": [ + 46, + 182, + 287, + 312 + ], + "type": "text", + "content": " to denote the loss function of text-to-sign-video contrast. In our implementation, we reuse the loss defined in Eq 2 but substitute the input with the text-to-video similarity matrix " + }, + { + "bbox": [ + 46, + 182, + 287, + 312 + ], + "type": "inline_equation", + "content": "Z_{T2V}" + }, + { + "bbox": [ + 46, + 182, + 287, + 312 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 314, + 287, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 314, + 287, + 350 + ], + "spans": [ + { + "bbox": [ + 47, + 314, + 287, + 350 + ], + "type": "text", + "content": "Loss Function. The overall loss for cross-lingual contrastive learning is a weighted sum of " + }, + { + "bbox": [ + 47, + 314, + 287, + 350 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{V2T}" + }, + { + "bbox": [ + 47, + 314, + 287, + 350 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 314, + 287, + 350 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{T2V}" + }, + { + "bbox": [ + 47, + 314, + 287, + 350 + ], + "type": "text", + "content": " with a trade-off hyper-parameter " + }, + { + "bbox": [ + 47, + 314, + 287, + 350 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 47, + 314, + 287, + 350 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 360, + 287, + 373 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 105, + 360, + 287, + 373 + ], + "spans": [ + { + "bbox": [ + 105, + 360, + 287, + 373 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\beta \\mathcal {L} _ {V 2 T} + (1 - \\beta) \\mathcal {L} _ {T 2 V}. \\tag {3}", + "image_path": "46b738db96a62a4aedfa72ba0b708560e81068fee9321a9de282948db08c1c92.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 381, + 160, + 395 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 381, + 160, + 395 + ], + "spans": [ + { + "bbox": [ + 47, + 381, + 160, + 395 + ], + "type": "text", + "content": "3.4. Text Augmentation" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 400, + 287, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 400, + 287, + 640 + ], + "spans": [ + { + "bbox": [ + 46, + 400, + 287, + 640 + ], + "type": "text", + "content": "Considering that the datasets of sign language retrieval are typically small-scale, we explore text augmentations to improve the generalization of our approach. EDA [64] introduces three simple yet efficient data augmentations in text classification task: random delete randomly removes words in a sentence; synonym replacement randomly selects words from a sentence that are not stop words and replaces them with synonyms; random swap randomly chooses two words in a sentence and swaps their positions. The first two augmentations have been proven effective in text classification task. However, we experimentally find that our focused sign language retrieval is sensitive to random delete and synonym replacement augmentations. To guarantee that the augmented texts preserve the original semantic meanings, we only adopt the random swap augmentation in our approach. We suppose there are two reasons: 1) the word order of sign languages and natural languages are constitutionally distinct, and reordering does not affect semantic meanings; 2) the proposed cross-lingual contrastive learning is insensitive to word order." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 651, + 123, + 664 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 651, + 123, + 664 + ], + "spans": [ + { + "bbox": [ + 47, + 651, + 123, + 664 + ], + "type": "text", + "content": "4. Experiment" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 670, + 242, + 684 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 670, + 242, + 684 + ], + "spans": [ + { + "bbox": [ + 47, + 670, + 242, + 684 + ], + "type": "text", + "content": "4.1. Datasets and Implementation Details" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 689, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 689, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 689, + 287, + 713 + ], + "type": "text", + "content": "Datasets. We primarily focus on How2Sign [19] dataset. Our model is also evaluated on PHOENIX-2014T [8] and" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 72, + 545, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 545, + 96 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 545, + 96 + ], + "type": "text", + "content": "CSL-Daily [71], which are primarily used for sign language recognition and translation in previous works." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 98, + 545, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 98, + 545, + 240 + ], + "spans": [ + { + "bbox": [ + 304, + 98, + 545, + 240 + ], + "type": "text", + "content": "How2Sign is a large-scale continuous American sign language (ASL) dataset consisting of a parallel corpus of about 80 hours of sign videos with subtitle annotations. It covers a wide range of instructional videos corresponding to various categories. There are 31164, 1740 and 2356 sign-video-text pairs in training, validation and test sets, respectively. Following SPOT-ALIGN [18], we remove the invalid pairs where the subtitle alignment is detected to exceed the video duration, remaining 31085, 1739 and 2348 available pairs in training, validation and test sets, respectively. The resolution of sign videos is " + }, + { + "bbox": [ + 304, + 98, + 545, + 240 + ], + "type": "inline_equation", + "content": "1280 \\times 720" + }, + { + "bbox": [ + 304, + 98, + 545, + 240 + ], + "type": "text", + "content": ", we crop the human bodies of signers with Faster R-CNN [52] to generate valid videos." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 243, + 545, + 303 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 243, + 545, + 303 + ], + "spans": [ + { + "bbox": [ + 304, + 243, + 545, + 303 + ], + "type": "text", + "content": "PHOENIX-2014T is a German sign language (Deutsche Gebärddensprache, DGS) dataset collected in the domain of weather forecast from TV broadcast, consisting of 7096, 519 and 642 video-text pairs in training, validation and test sets, respectively." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 304, + 545, + 352 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 304, + 545, + 352 + ], + "spans": [ + { + "bbox": [ + 304, + 304, + 545, + 352 + ], + "type": "text", + "content": "CSL-Daily is a recently released Chinese sign language (CSL) dataset. The topic of CSL-Daily revolves around people's daily lives, including 18401, 1077, 1176 parallel samples in training, validation and test set, respectively." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 354, + 545, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 354, + 545, + 449 + ], + "spans": [ + { + "bbox": [ + 304, + 354, + 545, + 449 + ], + "type": "text", + "content": "Evaluation Metric. Following previous works [18, 43, 51, 57], retrieval performance is evaluated by recall at rank K (R@K, higher is better) and median rank (MedR, lower is better). We evaluate our approach on both text-to-sign-video (T2V) retrieval and sign-video-to-text (V2T) retrieval tasks. We report R@1, R@5, and R@10 in all experiments, and additionally report MedR when comparing with state-of-the-art approaches." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 450, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 450, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 450, + 545, + 713 + ], + "type": "text", + "content": "Implementation Details. The sign encoder takes videos of resolution of " + }, + { + "bbox": [ + 304, + 450, + 545, + 713 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 304, + 450, + 545, + 713 + ], + "type": "text", + "content": " as input. The domain-agnostic sign encoder is a I3D [11] network pre-trained on BSL-1K [61]. In pseudo label generation, we set the threshold " + }, + { + "bbox": [ + 304, + 450, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 304, + 450, + 545, + 713 + ], + "type": "text", + "content": " to 0.6 to filter samples with low-confidence. Non-maximum suppression (NMS) with a temporal window of 24 frames is utilized to remove the duplicates among the pseudo-labeled samples. A collection of approximate 64K pseudo-labeled samples covering a vocabulary of 1220 words is eventually generated. The domain-aware sign encoder is initialized with the domain-agnostic one and fine-tuned with a learning rate of " + }, + { + "bbox": [ + 304, + 450, + 545, + 713 + ], + "type": "inline_equation", + "content": "1 \\times 10^{-2}" + }, + { + "bbox": [ + 304, + 450, + 545, + 713 + ], + "type": "text", + "content": " and batch size of 4 for 15 epochs. In the training of cross-lingual contrastive learning, the vision transformer and text transformer are initialized by the image encoder and text encoder in CLIP (ViT-B/32) [51]. The maximum length of sign clip features and text features are set to 64 and 32, respectively. The model is fine-tuned with Adam optimizer [31] with batch size of 512. The initial learning rate is set to " + }, + { + "bbox": [ + 304, + 450, + 545, + 713 + ], + "type": "inline_equation", + "content": "1 \\times 10^{-5}" + }, + { + "bbox": [ + 304, + 450, + 545, + 713 + ], + "type": "text", + "content": ", which is decreased with a cosine schedule following the CLIP [51]. We set " + }, + { + "bbox": [ + 304, + 450, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\alpha = 0.8" + }, + { + "bbox": [ + 304, + 450, + 545, + 713 + ], + "type": "text", + "content": " in Eq. 1 and " + }, + { + "bbox": [ + 304, + 450, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\beta = 0.5" + }, + { + "bbox": [ + 304, + 450, + 545, + 713 + ], + "type": "text", + "content": " in Eq.3. The languages of PHOENIX-2014T and CSL-Daily are German and Chinese" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19021" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 53, + 71, + 281, + 135 + ], + "blocks": [ + { + "bbox": [ + 53, + 71, + 281, + 135 + ], + "lines": [ + { + "bbox": [ + 53, + 71, + 281, + 135 + ], + "spans": [ + { + "bbox": [ + 53, + 71, + 281, + 135 + ], + "type": "table", + "html": "
ModelT2VV2T
R@1↑R@5↑R@10↑MedR↓R@1↑R@5↑R@10↑MedR↓
SA-SR [18]18.932.136.562.011.627.432.569.0
SA-CM [18]24.340.746.516.017.940.146.914.0
SA-COMB [18]34.248.052.68.023.647.053.07.5
Ours56.669.974.71.051.664.870.11.0
", + "image_path": "0d61485822ac79a3d779bc79c52ed8235b30f836ee46e1195bbbf08b4f79e876.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 53, + 179, + 281, + 244 + ], + "blocks": [ + { + "bbox": [ + 47, + 144, + 287, + 167 + ], + "lines": [ + { + "bbox": [ + 47, + 144, + 287, + 167 + ], + "spans": [ + { + "bbox": [ + 47, + 144, + 287, + 167 + ], + "type": "text", + "content": "Table 1. Comparison with the different variants of the pioneer SPOT-ALIGN (SA) [18] on How2Sign [19] dataset." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 53, + 179, + 281, + 244 + ], + "lines": [ + { + "bbox": [ + 53, + 179, + 281, + 244 + ], + "spans": [ + { + "bbox": [ + 53, + 179, + 281, + 244 + ], + "type": "table", + "html": "
ModelT2VV2T
R@1↑R@5↑R@10↑MedR↓R@1↑R@5↑R@10↑MedR↓
Translation [10]30.253.163.44.528.852.060.856.1
SA-CM [18]48.676.584.62.050.378.484.41.0
SA-COMB [18]55.879.687.21.053.179.486.11.0
Ours69.586.692.11.070.288.092.81.0
", + "image_path": "9f3573673a60cd914361f56931d0c810a46d15e8d0f25a98974762254dadd795.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 67, + 288, + 268, + 324 + ], + "blocks": [ + { + "bbox": [ + 47, + 253, + 287, + 276 + ], + "lines": [ + { + "bbox": [ + 47, + 253, + 287, + 276 + ], + "spans": [ + { + "bbox": [ + 47, + 253, + 287, + 276 + ], + "type": "text", + "content": "Table 2. Comparison with the different variants of the pioneer SPOT-ALIGN (SA) [18] on PHOENIX2014T [8] dataset." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 67, + 288, + 268, + 324 + ], + "lines": [ + { + "bbox": [ + 67, + 288, + 268, + 324 + ], + "spans": [ + { + "bbox": [ + 67, + 288, + 268, + 324 + ], + "type": "table", + "html": "
ModelT2VV2T
R@1↑R@5↑R@10↑MedR↓R@1↑R@5↑R@10↑MedR↓
Ours75.388.291.91.074.789.492.21.0
", + "image_path": "3b75778acbf62206832a51142494ab9f469493ea3d11fa8d81df64a2dbafb47d.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 368, + 287, + 404 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 368, + 287, + 404 + ], + "spans": [ + { + "bbox": [ + 47, + 368, + 287, + 404 + ], + "type": "text", + "content": "respectively. Since CLIP is trained on English corpus, to reuse CLIP's text encoder, we utilize Google translation to translate the texts of these two datasets into English." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 423, + 270, + 436 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 423, + 270, + 436 + ], + "spans": [ + { + "bbox": [ + 47, + 423, + 270, + 436 + ], + "type": "text", + "content": "4.2. Comparison with State-of-the-art Methods" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 446, + 287, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 446, + 287, + 495 + ], + "spans": [ + { + "bbox": [ + 47, + 446, + 287, + 495 + ], + "type": "text", + "content": "We compare our method with different variants of the pioneer, called SPOT-ALIGN [18], on How2Sign and PHOENIX-2014T. We also provide the results on CSL-Daily as a baseline." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 498, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 498, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 498, + 287, + 714 + ], + "type": "text", + "content": "Table 1 and Table 2 show the comparisons between our approach and SPOT-ALIGN [18] on How2Sign and PHOENIX-2014T, respectively. SPOT-ALIGN builds the final combination (COMB) model by integrating its primary cross-modal (CM) model with an auxiliary retrieval model (sign recognition (SR) model for How2Sign and Translation [10] for PHOENIX-2014T). Our method outperforms the COMB model, which achieves best results in SPOT-ALIGN, by large margins, achieving " + }, + { + "bbox": [ + 47, + 498, + 287, + 714 + ], + "type": "inline_equation", + "content": "+22.4" + }, + { + "bbox": [ + 47, + 498, + 287, + 714 + ], + "type": "text", + "content": " T2V and " + }, + { + "bbox": [ + 47, + 498, + 287, + 714 + ], + "type": "inline_equation", + "content": "+28.0" + }, + { + "bbox": [ + 47, + 498, + 287, + 714 + ], + "type": "text", + "content": " V2T R@1 improvements on How2Sign, " + }, + { + "bbox": [ + 47, + 498, + 287, + 714 + ], + "type": "inline_equation", + "content": "+13.7" + }, + { + "bbox": [ + 47, + 498, + 287, + 714 + ], + "type": "text", + "content": " T2V and " + }, + { + "bbox": [ + 47, + 498, + 287, + 714 + ], + "type": "inline_equation", + "content": "+17.1" + }, + { + "bbox": [ + 47, + 498, + 287, + 714 + ], + "type": "text", + "content": " V2T R@1 improvements on PHOENIX-2014T. It is worth mentioning that the SPOT-ALIGN conducts three rounds of sign spotting [1, 47] and encoder training. In contrast, we simplify the training of sign encoder and only perform a single round of training on pseudo-labeled data. We also provide a baseline on CSL-Daily dataset as shown in Table 3, demonstrating that our model can be generalized to various sign languages." + } + ] + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 326, + 71, + 526, + 135 + ], + "blocks": [ + { + "bbox": [ + 47, + 333, + 286, + 355 + ], + "lines": [ + { + "bbox": [ + 47, + 333, + 286, + 355 + ], + "spans": [ + { + "bbox": [ + 47, + 333, + 286, + 355 + ], + "type": "text", + "content": "Table 3. We additionally provide a baseline for CSL-Daily [71] dataset." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 326, + 71, + 526, + 135 + ], + "lines": [ + { + "bbox": [ + 326, + 71, + 526, + 135 + ], + "spans": [ + { + "bbox": [ + 326, + 71, + 526, + 135 + ], + "type": "table", + "html": "
MethodT2VV2T
R@1R@5R@10R@1R@5R@10
Baseline31.549.657.926.444.452.8
+SE33.351.959.830.948.856.4
+SE+CLCL54.067.171.950.063.268.1
+SE+CLCL+TA56.669.974.751.664.870.1
", + "image_path": "73f6a1930cfeb7bfd77f40c3f23941a8c0aba0e7befd5fc3879dc01ffe186fa4.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 316, + 183, + 536, + 249 + ], + "blocks": [ + { + "bbox": [ + 305, + 144, + 545, + 177 + ], + "lines": [ + { + "bbox": [ + 305, + 144, + 545, + 177 + ], + "spans": [ + { + "bbox": [ + 305, + 144, + 545, + 177 + ], + "type": "text", + "content": "Table 4. Ablation study of each proposed component. SE: sign encoder; CLCL: cross-lingual contrastive learning; TA: text augmentation." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 316, + 183, + 536, + 249 + ], + "lines": [ + { + "bbox": [ + 316, + 183, + 536, + 249 + ], + "spans": [ + { + "bbox": [ + 316, + 183, + 536, + 249 + ], + "type": "table", + "html": "
EncoderT2VV2T
R@1R@5R@10R@1R@5R@10
Single-Ag [61]53.168.073.447.362.967.0
Single-Aw54.167.573.149.161.867.4
Fusion-Average54.768.773.849.663.769.0
Fusion-Weighted Sum56.669.974.751.664.870.1
", + "image_path": "1fb6164a0165fe1f17b3209729979dc63e4babc4954e7091a05d70e13392e84f.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_body" + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 258, + 545, + 301 + ], + "lines": [ + { + "bbox": [ + 305, + 258, + 545, + 301 + ], + "spans": [ + { + "bbox": [ + 305, + 258, + 545, + 301 + ], + "type": "text", + "content": "Table 5. Results of domain-agnostic sign encoder (Single-Ag) and domain-aware sign encoder (Single-Aw). We also study different ways to integrate the features extracted by Single-Ag and Single-Aw, including average and weighted sum." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 306, + 309, + 400, + 323 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 309, + 400, + 323 + ], + "spans": [ + { + "bbox": [ + 306, + 309, + 400, + 323 + ], + "type": "text", + "content": "4.3. Ablation Study" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 329, + 545, + 352 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 329, + 545, + 352 + ], + "spans": [ + { + "bbox": [ + 305, + 329, + 545, + 352 + ], + "type": "text", + "content": "We conduct all ablation studies on the most challenging How2Sign dataset." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 354, + 546, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 354, + 546, + 533 + ], + "spans": [ + { + "bbox": [ + 304, + 354, + 546, + 533 + ], + "type": "text", + "content": "The Effectiveness of Each Proposed Component. Table 4 shows the ablation of each component. We first build a baseline where the sign encoder is the domain-agnostic one and contrastive learning is trained with the standard contrastive loss [51]. Then we add the proposed sign encoder (SE), cross-lingual contrastive learning (CLCL) and text augmentation (TA) to the baseline step by step. SE encodes domain-relevant and discriminative features, yielding " + }, + { + "bbox": [ + 304, + 354, + 546, + 533 + ], + "type": "inline_equation", + "content": "+1.8" + }, + { + "bbox": [ + 304, + 354, + 546, + 533 + ], + "type": "text", + "content": " T2V and " + }, + { + "bbox": [ + 304, + 354, + 546, + 533 + ], + "type": "inline_equation", + "content": "+4.5" + }, + { + "bbox": [ + 304, + 354, + 546, + 533 + ], + "type": "text", + "content": " V2T R@1 gains. The proposed CLCL significantly boosts the performance by " + }, + { + "bbox": [ + 304, + 354, + 546, + 533 + ], + "type": "inline_equation", + "content": "+20.7" + }, + { + "bbox": [ + 304, + 354, + 546, + 533 + ], + "type": "text", + "content": " T2V and " + }, + { + "bbox": [ + 304, + 354, + 546, + 533 + ], + "type": "inline_equation", + "content": "+19.1" + }, + { + "bbox": [ + 304, + 354, + 546, + 533 + ], + "type": "text", + "content": " V2T R@1 improvements, demonstrating that identifying fine-grained sign-to-word mappings is essential for sign language retrieval. The introduction of TA further promotes the retrieval task, achieving " + }, + { + "bbox": [ + 304, + 354, + 546, + 533 + ], + "type": "inline_equation", + "content": "56.6" + }, + { + "bbox": [ + 304, + 354, + 546, + 533 + ], + "type": "text", + "content": " R@1 on V2T and " + }, + { + "bbox": [ + 304, + 354, + 546, + 533 + ], + "type": "inline_equation", + "content": "51.6" + }, + { + "bbox": [ + 304, + 354, + 546, + 533 + ], + "type": "text", + "content": " R@1 on T2V, respectively." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 533, + 545, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 533, + 545, + 640 + ], + "spans": [ + { + "bbox": [ + 304, + 533, + 545, + 640 + ], + "type": "text", + "content": "Various Sign Encoders. As described in Section 3.2, our sign encoder is composed of a domain-agnostic sign encoder (Single-Ag) [61] and a domain-aware sign encoder (Single-Aw). We first report the results of each individual encoder in Table 5. Next, we study the different ways to integrate the features extracted by these two encoders, including average and weighted sum as defined in Eq. 1. The results are also shown in Table 5. We experimentally find that the weighted sum strategy yields best results." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "type": "text", + "content": "Variants of Cross-Lingual Contrastive Learning. As described in Section 3.3 and illustrated in Figure 3b, through the use of a combination of softmax, multiplication, sum, and average operations, we convert the fine-grained cross-lingual (sign-to-word) similarity to the coarse-grained sign-video-to-text similarity to enable contrastive learning. We" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19022" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 78, + 71, + 257, + 122 + ], + "blocks": [ + { + "bbox": [ + 78, + 71, + 257, + 122 + ], + "lines": [ + { + "bbox": [ + 78, + 71, + 257, + 122 + ], + "spans": [ + { + "bbox": [ + 78, + 71, + 257, + 122 + ], + "type": "table", + "html": "
StrategyT2VV2T
R@1R@5R@10R@1R@5R@10
Mean33.152.559.729.847.855.3
Max42.259.766.038.555.162.0
Softmax56.669.974.751.664.870.1
", + "image_path": "096d60734936273dee4db1cc5cb4e2060739936fa2610ad0554c334de416b0ec.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 71, + 161, + 264, + 221 + ], + "blocks": [ + { + "bbox": [ + 47, + 132, + 287, + 154 + ], + "lines": [ + { + "bbox": [ + 47, + 132, + 287, + 154 + ], + "spans": [ + { + "bbox": [ + 47, + 132, + 287, + 154 + ], + "type": "text", + "content": "Table 6. Study on different strategies to identify the fine-grained sign-to-word mappings in cross-lingual contrastive learning." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 71, + 161, + 264, + 221 + ], + "lines": [ + { + "bbox": [ + 71, + 161, + 264, + 221 + ], + "spans": [ + { + "bbox": [ + 71, + 161, + 264, + 221 + ], + "type": "table", + "html": "
Text AugmentationT2VV2T
R@1R@5R@10R@1R@5R@10
None54.067.171.950.063.268.1
RD52.766.371.647.662.167.4
SR53.968.773.047.561.265.9
RS56.669.974.751.664.870.1
", + "image_path": "63ea8cb87b322705488ee5fcd88c48582643d7f56a1ab385c6aa4ed0a085235b.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 269, + 286, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 269, + 286, + 472 + ], + "spans": [ + { + "bbox": [ + 46, + 269, + 286, + 472 + ], + "type": "text", + "content": "refer to this process as \"Softmax\". Here we study two variants termed \"Mean\" and \"Max\". The \"Mean\" and \"Max\" strategies simply replace the combination of softmax, multiplication and sum operations with a simple mean operation and a max operation, respectively. The results are shown in Table 6. We observe that the \"Mean\" strategy performs worst since it merely evaluates the overall similarity of a text and a sign video and ignores the fine-grained sign-to-word mappings during training. The \"Max\" strategy identifies hard sign-to-word mappings, i.e., one sign is associated with the most similar word, and vice versa. Nevertheless, since we do not have the ground-truth of sign-to-word mappings, it is challenging for models to identify the confident one-to-one mappings in a weakly supervised manner, as discussed in previous works on multiple instance learning [47]. In contrast, our default \"Softmax\" strategy localizes the soft sign-to-word mappings and achieves the best results." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 473, + 286, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 473, + 286, + 604 + ], + "spans": [ + { + "bbox": [ + 46, + 473, + 286, + 604 + ], + "type": "text", + "content": "Study of Text Augmentations. In Table 7, we study three text augmentations described in Section 3.4: random delete (RD), synonym replacement (SR) and random swap (RS). We observe a slight performance drop when introducing RD and SR into training. Sign language retrieval is not only a video-text retrieval task, but also a cross-lingual retrieval challenge, deletion and replacement may break the intrinsic sign-to-word mappings. In contrast, RS augmentation preserves the semantics of texts and we observe a +2.6 T2V and a +1.6 V2T R@1 improvements over the counterpart without any text augmentations." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 605, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 286, + 713 + ], + "type": "text", + "content": "The Effects of CLIP Initialization. In our framework, the vision and text Transformer are initialized with the CLIP's vision encoder and text encoder. Note that the input modality of our vision Transformer and that of CLIP's vision encoder are different. Ours takes a feature sequence as input while the input of CLIP's image encoder is a set of image patches. In Table 8, we compare a randomly initialized baseline with the one initialized by CLIP. We find that CLIP-initialization significantly improves the perfor" + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 332, + 71, + 520, + 114 + ], + "blocks": [ + { + "bbox": [ + 47, + 230, + 286, + 253 + ], + "lines": [ + { + "bbox": [ + 47, + 230, + 286, + 253 + ], + "spans": [ + { + "bbox": [ + 47, + 230, + 286, + 253 + ], + "type": "text", + "content": "Table 7. Ablation study on different text augmentations. RD: random delete; SR: synonym replacement; RS: random swap." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 332, + 71, + 520, + 114 + ], + "lines": [ + { + "bbox": [ + 332, + 71, + 520, + 114 + ], + "spans": [ + { + "bbox": [ + 332, + 71, + 520, + 114 + ], + "type": "table", + "html": "
InitializationT2VV2T
R@1R@5R@10R@1R@5R@10
Random45.260.367.540.654.559.7
CLIP56.669.974.751.664.870.1
", + "image_path": "e0ed1d145c38eb13d031044a6539d5c632f2147d9c5f02c9af73060d1da63e23.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 331, + 152, + 521, + 225 + ], + "blocks": [ + { + "bbox": [ + 306, + 123, + 544, + 145 + ], + "lines": [ + { + "bbox": [ + 306, + 123, + 544, + 145 + ], + "spans": [ + { + "bbox": [ + 306, + 123, + 544, + 145 + ], + "type": "text", + "content": "Table 8. Comparison between CLIP initialization and random initialization." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 331, + 152, + 521, + 225 + ], + "lines": [ + { + "bbox": [ + 331, + 152, + 521, + 225 + ], + "spans": [ + { + "bbox": [ + 331, + 152, + 521, + 225 + ], + "type": "table", + "html": "
Frozen layersT2VV2T
R@1R@5R@10R@1R@5R@10
None56.669.974.751.664.870.1
154.468.273.449.763.568.8
253.868.673.348.562.867.9
651.766.671.146.260.966.6
1249.565.271.444.759.765.2
", + "image_path": "d68ae9c2a50b495656d104b8026f8e7f66adda27835bf597822ea28ddb51e04e.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 312, + 234, + 538, + 245 + ], + "lines": [ + { + "bbox": [ + 312, + 234, + 538, + 245 + ], + "spans": [ + { + "bbox": [ + 312, + 234, + 538, + 245 + ], + "type": "text", + "content": "Table 9. Study of freezing different layers of text Transformer." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 305, + 265, + 545, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 265, + 545, + 300 + ], + "spans": [ + { + "bbox": [ + 305, + 265, + 545, + 300 + ], + "type": "text", + "content": "mance, yielding " + }, + { + "bbox": [ + 305, + 265, + 545, + 300 + ], + "type": "inline_equation", + "content": "+11.4" + }, + { + "bbox": [ + 305, + 265, + 545, + 300 + ], + "type": "text", + "content": " T2V and " + }, + { + "bbox": [ + 305, + 265, + 545, + 300 + ], + "type": "inline_equation", + "content": "+11.0" + }, + { + "bbox": [ + 305, + 265, + 545, + 300 + ], + "type": "text", + "content": " V2T R@1 improvements though the input originates from a completely distinct modality." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 303, + 545, + 433 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 303, + 545, + 433 + ], + "spans": [ + { + "bbox": [ + 305, + 303, + 545, + 433 + ], + "type": "text", + "content": "Study on Text Transformer. The text Transformer in our model is initialized with the CLIP's text encoder. Considering the generalization capability of the pre-trained CLIP, we attempt to freeze the shallow layers of our text transformer. As shown in Table 9, the performance gradually decreases as the number of the frozen layers increases. CLIP is trained on large-scale image-text pairs, however, the domain gap between image-text data and sign-video-text data is non-negligible. Though CLIP shows promising transfer capacity, it is optimal to fine-tune the whole model on target datasets." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 450, + 378, + 462 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 450, + 378, + 462 + ], + "spans": [ + { + "bbox": [ + 306, + 450, + 378, + 462 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 472, + 545, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 472, + 545, + 639 + ], + "spans": [ + { + "bbox": [ + 305, + 472, + 545, + 639 + ], + "type": "text", + "content": "In this paper, we propose a novel framework named Cross-lingual Contrastive learning (CiCo) for recently introduced sign language retrieval task. We formulate sign language retrieval as a cross-lingual retrieval task as well as a video-text retrieval problem. CiCo models the fine-grained cross-lingual mappings between sign videos and texts via the proposed cross-lingual contrastive learning. We also introduce a sign encoder, which is composed of a domain-agnostic encoder and a domain-aware one, to extract discriminative and domain-aligned features. Our Cico outperforms the pioneer SPOT-ALIGN by large margins on How2Sign and PHOENIX-2014T benchmarks. We also provide a baseline on CSL-Daily. We hope our approach could serve as a solid baseline for future research." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 641, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 641, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 305, + 641, + 545, + 712 + ], + "type": "text", + "content": "Acknowledgement. This work was partially supported by National Natural Science Foundation of China (No. 62072112), National Key R&D Program of China (No. 2020AAA0108301), and Scientific and Technological Innovation Action Plan of Shanghai Science and Technology Committee (No. 22511101502, 22511102202)." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19023" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 147 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 147 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 147 + ], + "type": "text", + "content": "[1] Samuel Albanie, Gül Varol, Liliane Momeni, Triantafyllos Afouras, Joon Son Chung, Neil Fox, and Andrew Zisserman. Bsl-1k: Scaling up co-articulated sign language recognition using mouthing cues. In European Conference on Computer Vision, pages 35–53. Springer, 2020. 1, 3, 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 148, + 288, + 192 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 148, + 288, + 192 + ], + "spans": [ + { + "bbox": [ + 53, + 148, + 288, + 192 + ], + "type": "text", + "content": "[2] Mikel Artetxe, Gorka Labaka, and Eneko Agirre. Learning bilingual word embeddings with (almost) no bilingual data. In *Meeting of the Association for Computational Linguistics*, pages 451–462, 2017. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 194, + 288, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 194, + 288, + 237 + ], + "spans": [ + { + "bbox": [ + 53, + 194, + 288, + 237 + ], + "type": "text", + "content": "[3] Mikel Artetxe, Gorka Labaka, and Eneko Agirre. A robust self-learning method for fully unsupervised cross-lingual mappings of word embeddings. arXiv preprint arXiv:1805.06297, 2018. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 239, + 288, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 239, + 288, + 316 + ], + "spans": [ + { + "bbox": [ + 53, + 239, + 288, + 316 + ], + "type": "text", + "content": "[4] Vassilis Athitsos, Carol Neidle, Stan Sclaroff, Joan Nash, Alexandra Stefan, Ashwin Thangali, Hajjing Wang, and Quan Yuan. Large lexicon project: American sign language video corpus and sign language indexing/retrieval algorithms. In Workshop on the Representation and Processing of Sign Languages: Corpora and Sign Language Technologies, volume 2, pages 11-14, 2010. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 318, + 288, + 361 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 318, + 288, + 361 + ], + "spans": [ + { + "bbox": [ + 53, + 318, + 288, + 361 + ], + "type": "text", + "content": "[5] Max Bain, Arsha Nagrani, Gül Varol, and Andrew Zisserman. Frozen in time: A joint video and image encoder for end-to-end retrieval. In IEEE/CVF International Conference on Computer Vision, pages 1728-1738, 2021. 2, 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 363, + 287, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 363, + 287, + 407 + ], + "spans": [ + { + "bbox": [ + 53, + 363, + 287, + 407 + ], + "type": "text", + "content": "[6] Lisa Ballesteros and Bruce Croft. Dictionary methods for cross-lingual information retrieval. In International Conference on Database and Expert Systems Applications, pages 791–801. Springer, 1996. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 409, + 287, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 409, + 287, + 453 + ], + "spans": [ + { + "bbox": [ + 53, + 409, + 287, + 453 + ], + "type": "text", + "content": "[7] Yoshua Bengio, Aaron Courville, and Pascal Vincent. Representation learning: A review and new perspectives. IEEE Transactions on Pattern Analysis and Machine Intelligence, 35(8):1798-1828, 2013. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 454, + 287, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 454, + 287, + 498 + ], + "spans": [ + { + "bbox": [ + 53, + 454, + 287, + 498 + ], + "type": "text", + "content": "[8] Necati Cihan Camgoz, Simon Hadfield, Oscar Koller, Hermann Ney, and Richard Bowden. Neural sign language translation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7784-7793, 2018. 2, 3, 6, 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 500, + 287, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 500, + 287, + 553 + ], + "spans": [ + { + "bbox": [ + 53, + 500, + 287, + 553 + ], + "type": "text", + "content": "[9] Necati Cihan Camgoz, Oscar Koller, Simon Hadfield, and Richard Bowden. Multi-channel transformers for multi-articulatory sign language translation. In European Conference on Computer Vision, pages 301-319. Springer, 2020. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 556, + 287, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 556, + 287, + 612 + ], + "spans": [ + { + "bbox": [ + 48, + 556, + 287, + 612 + ], + "type": "text", + "content": "[10] Necati Cihan Camgoz, Oscar Koller, Simon Hadfield, and Richard Bowden. Sign language transformers: Joint end-to-end sign language recognition and translation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10023-10033, 2020. 1, 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 613, + 287, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 613, + 287, + 657 + ], + "spans": [ + { + "bbox": [ + 48, + 613, + 287, + 657 + ], + "type": "text", + "content": "[11] Joao Carreira and Andrew Zisserman. Quo vadis, action recognition? a new model and the kinetics dataset. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6299-6308, 2017. 3, 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 658, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 287, + 712 + ], + "type": "text", + "content": "[12] Yutong Chen, Fangyun Wei, Xiao Sun, Zhirong Wu, and Stephen Lin. A simple multi-modality transfer learning baseline for sign language translation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5120-5130, 2022. 1, 3" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 713 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 128 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 128 + ], + "type": "text", + "content": "[13] Yutong Chen, Ronglai Zuo, Fangyun Wei, Yu Wu, Shujie LIU, and Brian Mak. Two-stream network for sign language recognition and translation. In Alice H. Oh, Alekh Agarwal, Danielle Belgrave, and Kyunghyun Cho, editors, Advances in Neural Information Processing Systems, 2022. 1, 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 129, + 545, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 129, + 545, + 183 + ], + "spans": [ + { + "bbox": [ + 307, + 129, + 545, + 183 + ], + "type": "text", + "content": "[14] Necati Cihan Camgoz, Simon Hadfield, Oscar Koller, and Richard Bowden. Subunets: End-to-end hand shape and continuous sign language recognition. In IEEE/CVF International Conference on Computer Vision, pages 3056-3065, 2017. 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 184, + 545, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 184, + 545, + 227 + ], + "spans": [ + { + "bbox": [ + 307, + 184, + 545, + 227 + ], + "type": "text", + "content": "[15] Runpeng Cui, Hu Liu, and Changshui Zhang. A deep neural framework for continuous sign language recognition by iterative training. IEEE Transactions on Multimedia, 21(7):1880-1891, 2019. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 228, + 545, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 228, + 545, + 282 + ], + "spans": [ + { + "bbox": [ + 308, + 228, + 545, + 282 + ], + "type": "text", + "content": "[16] Tonni Das Jui, Gissella Maria Bejarano, and Pablo Rivas. A machine learning-based segmentation approach for measuring similarity between sign languages. In sign-lang@ LREC 2022, pages 94-101. European Language Resources Association (ELRA), 2022. 4" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 283, + 545, + 338 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 283, + 545, + 338 + ], + "spans": [ + { + "bbox": [ + 308, + 283, + 545, + 338 + ], + "type": "text", + "content": "[17] Yu Du, Fangyun Wei, Zihe Zhang, Miaojing Shi, Yue Gao, and Guoqi Li. Learning to prompt for open-vocabulary object detection with vision-language model. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14084-14093, 2022. 5" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 338, + 545, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 338, + 545, + 392 + ], + "spans": [ + { + "bbox": [ + 308, + 338, + 545, + 392 + ], + "type": "text", + "content": "[18] Amanda Duarte, Samuel Albanie, Xavier Giró-i Nieto, and Gül Varol. Sign language video retrieval with free-form textual queries. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14094-14104, 2022. 1, 2, 3, 4, 6, 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 393, + 545, + 458 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 393, + 545, + 458 + ], + "spans": [ + { + "bbox": [ + 308, + 393, + 545, + 458 + ], + "type": "text", + "content": "[19] Amanda Duarte, Shruti Palaskar, Lucas Ventura, Deepti Ghadiyaram, Kenneth DeHaan, Florian Metze, Jordi Torres, and Xavier Giro-i Nieto. How2Sign: A Large-scale Multimodal Dataset for Continuous American Sign Language. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2021. 2, 6, 7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 460, + 545, + 492 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 460, + 545, + 492 + ], + "spans": [ + { + "bbox": [ + 308, + 460, + 545, + 492 + ], + "type": "text", + "content": "[20] Ali Farhadi, David Forsyth, and Ryan White. Transfer learning in sign language. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1-8. IEEE, 2007. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 493, + 545, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 493, + 545, + 536 + ], + "spans": [ + { + "bbox": [ + 308, + 493, + 545, + 536 + ], + "type": "text", + "content": "[21] Holger Fillbrandt, Suat Akyol, and K-F Kraiss. Extraction of 3d hand shape and posture from image sequences for sign language recognition. In 2003 IEEE International SOI Conference, pages 181-186. IEEE, 2003. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 537, + 545, + 580 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 537, + 545, + 580 + ], + "spans": [ + { + "bbox": [ + 308, + 537, + 545, + 580 + ], + "type": "text", + "content": "[22] Stanislav Fort, Jie Ren, and Balaji Lakshminarayanan. Exploring the limits of out-of-distribution detection. Advances in Neural Information Processing Systems, 34:7068-7081, 2021. 5" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 582, + 545, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 582, + 545, + 624 + ], + "spans": [ + { + "bbox": [ + 308, + 582, + 545, + 624 + ], + "type": "text", + "content": "[23] Andrea Frome, Greg S Corrado, Jon Shlens, Samy Bengio, Jeff Dean, Marc'Aurelio Ranzato, and Tomas Mikolov. Devise: A deep visual-semantic embedding model. Advances in neural information processing systems, 26, 2013. 3, 5" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 625, + 545, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 625, + 545, + 669 + ], + "spans": [ + { + "bbox": [ + 308, + 625, + 545, + 669 + ], + "type": "text", + "content": "[24] Valentin Gabeur, Chen Sun, Karteek Alahari, and Cordelia Schmid. Multi-modal transformer for video retrieval. In European Conference on Computer Vision, pages 214-229. Springer, 2020. 2, 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 670, + 547, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 670, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 670, + 547, + 713 + ], + "type": "text", + "content": "[25] Yunchao Gong, Qifa Ke, Michael Isard, and Svetlana Lazebnik. A multi-view embedding space for modeling internet images, tags, and their semantics. International Journal of Computer Vision, 106(2):210-233, 2014. 3, 5" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "19024" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "text", + "content": "[26] Michael Gutmann and Aapo Hyvarinen. Noise-contrastive estimation: A new estimation principle for unnormalized statistical models. In International Conference on Artificial Intelligence and Statistics, pages 297–304. JMLR Workshop and Conference Proceedings, 2010. 5" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 128, + 288, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 128, + 288, + 160 + ], + "spans": [ + { + "bbox": [ + 48, + 128, + 288, + 160 + ], + "type": "text", + "content": "[27] Tony Huang, Jack Chu, and Fangyun Wei. Unsupervised prompt learning for vision-language models. arXiv preprint arXiv:2204.03649, 2022. 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 161, + 288, + 225 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 161, + 288, + 225 + ], + "spans": [ + { + "bbox": [ + 48, + 161, + 288, + 225 + ], + "type": "text", + "content": "[28] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In International Conference on Machine Learning, pages 4904-4916. PMLR, 2021. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 228, + 288, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 228, + 288, + 271 + ], + "spans": [ + { + "bbox": [ + 48, + 228, + 288, + 271 + ], + "type": "text", + "content": "[29] Armand Joulin, Laurens van der Maaten, Allan Jabri, and Nicolas Vasilache. Learning visual features from large weakly supervised data. In European Conference on Computer Vision, pages 67-84. Springer, 2016. 3, 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 272, + 288, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 272, + 288, + 304 + ], + "spans": [ + { + "bbox": [ + 48, + 272, + 288, + 304 + ], + "type": "text", + "content": "[30] Hamid Reza Vaezi Joze and Oscar Koller. Ms-asl: A large-scale data set and benchmark for understanding american sign language. arXiv preprint arXiv:1812.01053, 2018. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 305, + 288, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 305, + 288, + 360 + ], + "spans": [ + { + "bbox": [ + 48, + 305, + 288, + 360 + ], + "type": "text", + "content": "[31] Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In Yoshua Bengio and Yann LeCun, editors, 3rd International Conference on Learning Representations, ICLR 2015, San Diego, CA, USA, May 7-9, 2015, Conference Track Proceedings, 2015. 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 360, + 288, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 360, + 288, + 415 + ], + "spans": [ + { + "bbox": [ + 48, + 360, + 288, + 415 + ], + "type": "text", + "content": "[32] Oscar Koller, Necati Cihan Camgoz, Hermann Ney, and Richard Bowden. Weakly supervised learning with multistream cnn-lstm-hmms to discover sequential parallelism in sign language videos. IEEE Transactions on Pattern Analysis and Machine Intelligence, 42(9):2306-2320, 2019. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 415, + 288, + 458 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 415, + 288, + 458 + ], + "spans": [ + { + "bbox": [ + 48, + 415, + 288, + 458 + ], + "type": "text", + "content": "[33] Oscar Koller, Jens Forster, and Hermann Ney. Continuous sign language recognition: Towards large vocabulary statistical recognition systems handling multiple signers. Computer Vision and Image Understanding, 141:108-125, 2015. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 459, + 288, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 459, + 288, + 502 + ], + "spans": [ + { + "bbox": [ + 48, + 459, + 288, + 502 + ], + "type": "text", + "content": "[34] Victor Lavrenko, Martin Choquette, and W Bruce Croft. Cross-lingual relevance models. In International ACM SIGIR Conference on Research and Development in Information Retrieval, pages 175–182, 2002. 2, 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 503, + 288, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 503, + 288, + 525 + ], + "spans": [ + { + "bbox": [ + 48, + 503, + 288, + 525 + ], + "type": "text", + "content": "[35] Yann LeCun, Yoshua Bengio, and Geoffrey Hinton. Deep learning. Nature, 521(7553):436-444, 2015. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 526, + 288, + 580 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 526, + 288, + 580 + ], + "spans": [ + { + "bbox": [ + 48, + 526, + 288, + 580 + ], + "type": "text", + "content": "[36] Dongxu Li, Cristian Rodriguez, Xin Yu, and Hongdong Li. Word-level deep sign language recognition from video: A new large-scale dataset and methods comparison. In IEEE/CVF winter conference on applications of computer vision, pages 1459-1469, 2020. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 581, + 288, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 581, + 288, + 635 + ], + "spans": [ + { + "bbox": [ + 48, + 581, + 288, + 635 + ], + "type": "text", + "content": "[37] Dongxu Li, Chenchen Xu, Xin Yu, Kaihao Zhang, Benjamin Swift, Hanna Suominen, and Hongdong Li. Tspnet: Hierarchical feature learning via temporal semantic pyramid for sign language translation. Advances in Neural Information Processing Systems, 33:12034-12045, 2020. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 636, + 288, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 636, + 288, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 636, + 288, + 689 + ], + "type": "text", + "content": "[38] Dongxu Li, Xin Yu, Chenchen Xu, Lars Petersson, and Hongdong Li. Transferring cross-domain knowledge for video sign language recognition. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6205-6214, 2020. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 690, + 288, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 690, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 690, + 288, + 713 + ], + "type": "text", + "content": "[39] Junnan Li, Ramprasaath Selvaraju, Akhilesh Gotmare, Shafiq Joty, Caiming Xiong, and Steven Chu Hong Hoi." + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 327, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 545, + 106 + ], + "type": "text", + "content": "Align before fuse: Vision and language representation learning with momentum distillation. Advances in Neural Information Processing Systems, 34:9694-9705, 2021. 3, 5" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 107, + 545, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 107, + 545, + 161 + ], + "spans": [ + { + "bbox": [ + 307, + 107, + 545, + 161 + ], + "type": "text", + "content": "[40] Yangguang Li, Feng Liang, Lichen Zhao, Yufeng Cui, Wanli Ouyang, Jing Shao, Fengwei Yu, and Junjie Yan. Supervision exists everywhere: A data efficient contrastive language-image pre-training paradigm. arXiv preprint arXiv:2110.05208, 2021. 5" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 163, + 545, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 163, + 545, + 206 + ], + "spans": [ + { + "bbox": [ + 307, + 163, + 545, + 206 + ], + "type": "text", + "content": "[41] Ji Lin, Chuang Gan, and Song Han. Tsm: Temporal shift module for efficient video understanding. In IEEE/CVF International Conference on Computer Vision, pages 7083-7093, 2019. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 208, + 545, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 208, + 545, + 262 + ], + "spans": [ + { + "bbox": [ + 307, + 208, + 545, + 262 + ], + "type": "text", + "content": "[42] Song Liu, Haoqi Fan, Shengsheng Qian, Yiru Chen, Wenkui Ding, and Zhongyuan Wang. Hit: Hierarchical transformer with momentum contrast for video-text retrieval. In IEEE/CVF International Conference on Computer Vision, pages 11915-11925, 2021. 2, 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 264, + 545, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 264, + 545, + 308 + ], + "spans": [ + { + "bbox": [ + 307, + 264, + 545, + 308 + ], + "type": "text", + "content": "[43] Yang Liu, Samuel Albanie, Arsha Nagrani, and Andrew Zisserman. Use what you have: Video retrieval using representations from collaborative experts. arXiv preprint arXiv:1907.13487, 2019. 2, 3, 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 309, + 545, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 309, + 545, + 353 + ], + "spans": [ + { + "bbox": [ + 307, + 309, + 545, + 353 + ], + "type": "text", + "content": "[44] Huaishao Luo, Lei Ji, Ming Zhong, Yang Chen, Wen Lei, Nan Duan, and Tianrui Li. Clip4clip: An empirical study of clip for end to end video clip retrieval and captioning. Neurocomputing, 508:293-304, 2022. 2, 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 354, + 545, + 419 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 354, + 545, + 419 + ], + "spans": [ + { + "bbox": [ + 307, + 354, + 545, + 419 + ], + "type": "text", + "content": "[45] Antoine Miech, Dimitri Zhukov, Jean-Baptiste Alayrac, Makarand Tapaswi, Ivan Laptev, and Josef Sivic. Howto100m: Learning a text-video embedding by watching hundred million narrated video clips. In IEEE/CVF International Conference on Computer Vision, pages 2630-2640, 2019. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 421, + 545, + 454 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 421, + 545, + 454 + ], + "spans": [ + { + "bbox": [ + 307, + 421, + 545, + 454 + ], + "type": "text", + "content": "[46] Tomas Mikolov, Quoc V Le, and Ilya Sutskever. Exploiting similarities among languages for machine translation. arXiv preprint arXiv:1309.4168, 2013. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 456, + 545, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 456, + 545, + 499 + ], + "spans": [ + { + "bbox": [ + 307, + 456, + 545, + 499 + ], + "type": "text", + "content": "[47] Liliane Momeni, Gül Varol, Samuel Albanie, Triantafyllos Afouras, and Andrew Zisserman. Watch, read and lookup: learning to spot signs from multiple supervisors. In Asian Conference on Computer Vision, 2020. 3, 4, 7, 8" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 501, + 545, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 501, + 545, + 566 + ], + "spans": [ + { + "bbox": [ + 307, + 501, + 545, + 566 + ], + "type": "text", + "content": "[48] Mathew Monfort, SouYoung Jin, Alexander Liu, David Harwath, Rogerio Feris, James Glass, and Aude Oliva. Spoken moments: Learning joint audio-visual representations from video descriptions. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14871-14881, 2021. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 567, + 545, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 567, + 545, + 622 + ], + "spans": [ + { + "bbox": [ + 307, + 567, + 545, + 622 + ], + "type": "text", + "content": "[49] Yasuhide Mori, Hironobu Takahashi, and Ryuichi Oka. Image-to-word transformation based on dividing and vector quantizing images with words. In International Workshop on Multimedia Intelligent Storage and Retrieval Management, pages 1-9. Citeseer, 1999. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 624, + 545, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 624, + 545, + 667 + ], + "spans": [ + { + "bbox": [ + 307, + 624, + 545, + 667 + ], + "type": "text", + "content": "[50] Aitor Ormazabal, Mikel Artetxe, Gorka Labaka, Aitor Soroa, and Eneko Agirre. Analyzing the limitations of cross-lingual word embedding mappings. arXiv preprint arXiv:1906.05407, 2019. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 669, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 669, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 669, + 545, + 713 + ], + "type": "text", + "content": "[51] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language super" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "19025" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 66, + 72, + 286, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 72, + 286, + 95 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 286, + 95 + ], + "type": "text", + "content": "vision. In International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 3, 5, 6, 7" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 96, + 287, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 96, + 287, + 140 + ], + "spans": [ + { + "bbox": [ + 48, + 96, + 287, + 140 + ], + "type": "text", + "content": "[52] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. Conference on Neural Information Processing Systems, 28, 2015. 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 141, + 287, + 174 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 141, + 287, + 174 + ], + "spans": [ + { + "bbox": [ + 48, + 141, + 287, + 174 + ], + "type": "text", + "content": "[53] Rico Sennrich, Barry Haddow, and Alexandra Birch. Neural machine translation of rare words with subword units. arXiv preprint arXiv:1508.07909, 2015. 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 175, + 287, + 219 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 175, + 287, + 219 + ], + "spans": [ + { + "bbox": [ + 48, + 175, + 287, + 219 + ], + "type": "text", + "content": "[54] Sheng Shen, Liunian Harold Li, Hao Tan, Mohit Bansal, Anna Rohrbach, Kai-Wei Chang, Zhewei Yao, and Kurt Keutzer. How much can clip benefit vision-and-language tasks? arXiv preprint arXiv:2107.06383, 2021. 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 220, + 287, + 263 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 220, + 287, + 263 + ], + "spans": [ + { + "bbox": [ + 48, + 220, + 287, + 263 + ], + "type": "text", + "content": "[55] Thad Starner, Joshua Weaver, and Alex Pentland. Real-time american sign language recognition using desk and wearable computer based video. IEEE Transactions on Pattern Analysis and Machine Intelligence, 20(12):1371-1375, 1998. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 265, + 287, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 265, + 287, + 308 + ], + "spans": [ + { + "bbox": [ + 48, + 265, + 287, + 308 + ], + "type": "text", + "content": "[56] Thad E Starner. Visual recognition of american sign language using hidden markov models. Technical report, Massachusetts inst of tech Cambridge dept of brain and cognitive sciences, 1995. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 310, + 287, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 310, + 287, + 364 + ], + "spans": [ + { + "bbox": [ + 48, + 310, + 287, + 364 + ], + "type": "text", + "content": "[57] Chen Sun, Austin Myers, Carl Vondrick, Kevin Murphy, and Cordelia Schmid. Videobert: A joint model for video and language representation learning. In IEEE/CVF International Conference on Computer Vision, pages 7464-7473, 2019. 2, 3, 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 366, + 287, + 399 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 366, + 287, + 399 + ], + "spans": [ + { + "bbox": [ + 48, + 366, + 287, + 399 + ], + "type": "text", + "content": "[58] Rachel Sutton-Spence. Mouthings and simultaneity in British sign language. Amsterdam Studies in the Theory and History of Linguistic Science Series, 281:147, 2007. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 401, + 287, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 401, + 287, + 443 + ], + "spans": [ + { + "bbox": [ + 48, + 401, + 287, + 443 + ], + "type": "text", + "content": "[59] Chau Tran, Yuqing Tang, Xian Li, and Jiatao Gu. Crosslingual retrieval for iterative self-supervised training. Advances in Neural Information Processing Systems, 33:2207-2219, 2020. 2, 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 445, + 287, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 445, + 287, + 499 + ], + "spans": [ + { + "bbox": [ + 48, + 445, + 287, + 499 + ], + "type": "text", + "content": "[60] Du Tran, Heng Wang, Lorenzo Torresani, Jamie Ray, Yann LeCun, and Manohar Paluri. A closer look at spatiotemporal convolutions for action recognition. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6450-6459, 2018. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 501, + 287, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 501, + 287, + 555 + ], + "spans": [ + { + "bbox": [ + 48, + 501, + 287, + 555 + ], + "type": "text", + "content": "[61] Gul Varol, Liliane Momeni, Samuel Albanie, Triantafyllos Afouras, and Andrew Zisserman. Read and attend: Temporal localisation in sign language videos. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16857-16866, 2021. 1, 2, 3, 4, 6, 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 557, + 287, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 557, + 287, + 590 + ], + "spans": [ + { + "bbox": [ + 48, + 557, + 287, + 590 + ], + "type": "text", + "content": "[62] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. 2017. 5" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 591, + 287, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 591, + 287, + 635 + ], + "spans": [ + { + "bbox": [ + 48, + 591, + 287, + 635 + ], + "type": "text", + "content": "[63] Xiaolong Wang, Ross Girshick, Abhinav Gupta, and Kaiming He. Non-local neural networks. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7794-7803, 2018. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 636, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 636, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 636, + 287, + 713 + ], + "type": "text", + "content": "[64] Jason Wei and Kai Zou. EDA: Easy data augmentation techniques for boosting performance on text classification tasks. In 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages 6383-6389, Hong Kong, China, Nov. 2019. Association for Computational Linguistics. 6" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 507 + ], + "type": "list", + "angle": 0, + "index": 24, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 126 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 126 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 126 + ], + "type": "text", + "content": "[65] Jinxi Xu, Ralph Weischedel, and Chanh Nguyen. Evaluating a probabilistic model for cross-lingual information retrieval. In International ACM SIGIR Conference on Research and Development in Information Retrieval, pages 105–110, 2001. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 129, + 545, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 129, + 545, + 162 + ], + "spans": [ + { + "bbox": [ + 307, + 129, + 545, + 162 + ], + "type": "text", + "content": "[66] Mengde Xu, Zheng Zhang, Fangyun Wei, Han Hu, and Xiang Bai. Side adapter network for open-vocabulary semantic segmentation. arXiv preprint arXiv:2302.12242, 2023. 5" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 163, + 545, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 163, + 545, + 217 + ], + "spans": [ + { + "bbox": [ + 307, + 163, + 545, + 217 + ], + "type": "text", + "content": "[67] Mengde Xu, Zheng Zhang, Fangyun Wei, Yutong Lin, Yue Cao, Han Hu, and Xiang Bai. A simple baseline for open-vocabulary semantic segmentation with pre-trained vision-language model. In European Conference on Computer Vision, pages 736-753. Springer, 2022. 5" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 219, + 545, + 261 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 219, + 545, + 261 + ], + "spans": [ + { + "bbox": [ + 307, + 219, + 545, + 261 + ], + "type": "text", + "content": "[68] Youngjae Yu, Jongseok Kim, and Gunhee Kim. A joint sequence fusion model for video question answering and retrieval. In European Conference on Computer Vision, pages 471-487, 2018. 2, 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 263, + 545, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 263, + 545, + 316 + ], + "spans": [ + { + "bbox": [ + 307, + 263, + 545, + 316 + ], + "type": "text", + "content": "[69] Mozhi Zhang, Keyulu Xu, Ken-ichi Kawarabayashi, Stefanie Jegelka, and Jordan Boyd-Graber. Are girls neko or shōjo? cross-lingual alignment of non-isomorphic embeddings with iterative normalization. arXiv preprint arXiv:1906.01622, 2019.3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 319, + 545, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 319, + 545, + 363 + ], + "spans": [ + { + "bbox": [ + 307, + 319, + 545, + 363 + ], + "type": "text", + "content": "[70] Shilin Zhang and Bo Zhang. Using revised string edit distance to sign language video retrieval. In International Conference on Computational Intelligence and Natural Computing, volume 1, pages 45-49. IEEE, 2010. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 365, + 545, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 365, + 545, + 418 + ], + "spans": [ + { + "bbox": [ + 307, + 365, + 545, + 418 + ], + "type": "text", + "content": "[71] Hao Zhou, Wengang Zhou, Weizhen Qi, Junfu Pu, and Houqiang Li. Improving sign language translation with monolingual data by sign back-translation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1316-1325, 2021. 6, 7" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 420, + 545, + 462 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 420, + 545, + 462 + ], + "spans": [ + { + "bbox": [ + 307, + 420, + 545, + 462 + ], + "type": "text", + "content": "[72] Hao Zhou, Wengang Zhou, Yun Zhou, and Houqiang Li. Spatial-temporal multi-cue network for sign language recognition and translation. IEEE Transactions on Multimedia, 24:768-779, 2021. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 464, + 545, + 507 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 464, + 545, + 507 + ], + "spans": [ + { + "bbox": [ + 307, + 464, + 545, + 507 + ], + "type": "text", + "content": "[73] Ronglai Zuo, Fangyun Wei, and Brian Mak. Natural language-assisted sign language recognition. IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023. 1, 3" + } + ] + } + ], + "index": 23 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "19026" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/CiaoSR_ Continuous Implicit Attention-in-Attention Network for Arbitrary-Scale Image Super-Resolution/dcf45018-bd80-4a92-ad40-a481bde62c68_content_list.json b/2023/CiaoSR_ Continuous Implicit Attention-in-Attention Network for Arbitrary-Scale Image Super-Resolution/dcf45018-bd80-4a92-ad40-a481bde62c68_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..9843bf150daed692100982e8c4934193f016b5f7 --- /dev/null +++ b/2023/CiaoSR_ Continuous Implicit Attention-in-Attention Network for Arbitrary-Scale Image Super-Resolution/dcf45018-bd80-4a92-ad40-a481bde62c68_content_list.json @@ -0,0 +1,1839 @@ +[ + { + "type": "text", + "text": "CiaoSR: Continuous Implicit Attention-in-Attention Network for Arbitrary-Scale Image Super-Resolution", + "text_level": 1, + "bbox": [ + 156, + 128, + 818, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Jiezhang Cao $^{1}$ Qin Wang $^{1}$ Yongqin Xian $^{1*}$ Yawei Li $^{1}$ Bingbing Ni $^{2}$ Zhiming Pi $^{2}$", + "bbox": [ + 133, + 202, + 834, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Kai Zhang $^{1\\dagger}$ Yulun Zhang $^{1\\dagger}$ Radu Timofte $^{1,3}$ Luc Van Gool $^{1,4}$", + "bbox": [ + 223, + 220, + 748, + 238 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ ETH Zürich $^{2}$ Huawei Inc. $^{3}$ University of Wurzburg $^{4}$ KU Leuven", + "bbox": [ + 205, + 238, + 764, + 256 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "https://github.com/caojiezhang/CiaoSR", + "bbox": [ + 316, + 258, + 650, + 273 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 308, + 313, + 325 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Learning continuous image representations is recently gaining popularity for image super-resolution (SR) because of its ability to reconstruct high-resolution images with arbitrary scales from low-resolution inputs. Existing methods mostly ensemble nearby features to predict the new pixel at any queried coordinate in the SR image. Such a local ensemble suffers from some limitations: i) it has no learnable parameters and it neglects the similarity of the visual features; ii) it has a limited receptive field and cannot ensemble relevant features in a large field which are important in an image. To address these issues, this paper proposes a continuous implicit attention-in-attention network, called CiaoSR. We explicitly design an implicit attention network to learn the ensemble weights for the nearby local features. Furthermore, we embed a scale-aware attention in this implicit attention network to exploit additional non-local information. Extensive experiments on benchmark datasets demonstrate CiaoSR significantly outperforms the existing single image SR methods with the same backbone. In addition, CiaoSR also achieves the state-of-the-art performance on the arbitrary-scale SR task. The effectiveness of the method is also demonstrated on the real-world SR setting. More importantly, CiaoSR can be flexibly integrated into any backbone to improve the SR performance.", + "bbox": [ + 75, + 340, + 473, + 705 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 714, + 209, + 729 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Single image super-resolution (SISR), which aims to reconstruct a high-resolution (HR) image from a low-resolution (LR) one, has been widely employed in many practical applications [24, 61, 91]. However, deep neural networks (DNN)-based SISR methods are facing some limitations in some real-world scenarios with arbitrary scales. For example, camera users may want to enhance the digital zoom quality by super-resolving a photo or a video to", + "bbox": [ + 75, + 732, + 470, + 853 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/ff01524b6deb00056d2d082586a728943e8e013d0bdd1e1107b95a47d23c9250.jpg", + "image_caption": [ + "Figure 1. Comparison of different backbones and implicit models. Our proposed implicit neural network on RDN [88] has better performance than SwinIR [40]. Check Section 5 for details." + ], + "image_footnote": [], + "bbox": [ + 500, + 308, + 890, + 405 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "continuous arbitrary scales. Most existing DNN-based SISR methods [40, 44, 86] need to train a series of models for all different scales separately. However, it can be impractical to store all these models on the device due to limited storage and computing power. Alternatively, arbitrary-scale image SR methods [13, 28, 39] aim to train a single network for all scales in a continuous manner.", + "bbox": [ + 496, + 458, + 893, + 565 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Most existing SISR methods [40,44,86] consist of a DNN and an upsampling module (e.g., pixel shuffling [60]) at a discrete scale. While substantial progresses have been made in the DNN backbones for SR, there is little attempt to study the upsampling module. A natural question to ask is: Does the pixel shuffling hinder the potential of SR models? One limitation of the pixel shuffling module is that it cannot synthesize SR images at large unseen and continuous scales. To tackle this, one can treat synthesizing different-scale SR images as a multi-task learning problem, and train a specific upsampling module for each scale [44]. However, these tasks are dependent and highly inter-related. Neglecting the correlation of different-scale SR tasks may lead to discrete representations and limited performance. Under a certain capacity of a network, training a model on multi-tasks may sacrifice the performance or have the comparable performance on each task. These above disadvantages limit its applicability and flexibility in the real-world scenarios.", + "bbox": [ + 496, + 566, + 895, + 838 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To address these, most existing arbitrary-scale SR methods [13, 28, 39] replace the upsampling method with an implicit neural function and boost the performance. These methods predict an RGB value at the query point in an image", + "bbox": [ + 496, + 839, + 895, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Currently with Google. This work was done at ETH Zürich.", + "bbox": [ + 94, + 862, + 418, + 875 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "†Corresponding Authors: Kai Zhang, cskaizhang@gmail.com; Yulun Zhang, yulun100@gmail.com", + "bbox": [ + 76, + 875, + 468, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "1796", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/eccbf2d2b3a66c58a8bff1e822218e41c9a651de862409ff0b1be2eabb7b8bdb.jpg", + "image_caption": [ + "Figure 2. Comparisons of different attention mechanisms. (a) Self-attention can predict pixel features on the grid, but it cannot be directly used in arbitrary-scale SR without considering coordinates. (b) Most existing methods can be treated as coordinate-based implicit attention since they calculate the distance between a key and query coordinate, and then use a function $g$ to aggregate with the value features. However, these methods ignore the distance between the features. (c) Our implicit attention not only considers the coordinate distance, but also the distance among features with visual information." + ], + "image_footnote": [], + "bbox": [ + 133, + 89, + 834, + 209 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "by ensembling features within a local region. However, the local ensemble methods have limitations in the ensemble weights and insufficient information (e.g., non-local information). The ensemble weights are often calculated by the area of the rectangle between the query point and each nearest point, which is equivalent to the bilinear interpolation. Thus, those methods cannot adaptively ensemble local features since there is no trainable parameter. These weights are only related to the coordinates of the local features, but independent of the local features. Ignoring both the coordinates and the local features lose visual information and result in blurry artifacts. It is important and necessary to design a new implicit network to predict the weights and exploit more information in the local ensemble.", + "bbox": [ + 75, + 289, + 472, + 500 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we propose a novel implicit attention model to enhance arbitrary-scale SR performance. Specifically, we use our attention to predict the ensemble weights by considering both the similarity and coordinate distance of local features, as shown in Figure 2. Based on such learnable weights, the implicit model can adaptively aggregate local features according to different inputs. To enrich more information, we introduce an attention in our implicit attention, which helps discover more features in a larger receptive field.", + "bbox": [ + 75, + 501, + 473, + 637 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our contributions are summarized as follows:", + "bbox": [ + 96, + 638, + 398, + 652 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose a novel continuous implicit attention-inattention network for arbitrary-scale image SR, called CiaoSR. Different from most existing local ensemble methods, our method explicitly learns the ensemble weights and exploits scale-aware non-local information.", + "- Our CiaoSR can be flexibly integrated into any backbone, allowing the network to super-resolve an image at arbitrary scales and improve the SR performance in Figure 1.", + "- Extensive experiments demonstrate CiaoSR achieves the state-of-the-art performance in both SISR and arbitrary-scale SR tasks. Besides, our CiaoSR has good generalization on both in-scale and out-of-scale distributions. Last, we extend our method to real-world SR settings to synthesize arbitrary-scale images." + ], + "bbox": [ + 76, + 660, + 470, + 885 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 287, + 640, + 303 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Single image super-resolution (SISR). SISR aims to synthesize high-resolution (HR) images from low-resolution (LR) images. Compared with DNN-based SR methods [6-8, 23, 27, 50, 66, 67], methods in recent years build on deep convolutional neural network (CNN) to improve the performance, such as SRCNN [19], SRResNet [38], EDSR [44], RDN [88] and RCAN [86]. To further improve SR performance, some methods design CNN with residual block [9, 34, 81], dense block [75, 88, 89] and others [14, 16, 18, 20, 21, 25, 31, 32, 35, 37, 41-43, 59, 65, 70-72, 77, 83, 84, 87, 90]. In addition, some SR methods are built based on attention mechanism [68], such as channel attention [17, 56, 86], self-attention (IPT [11] and SwinIR [40], HAT [12]), non-local attention [45, 48]. However, most methods focus on one specific scale, which limits the applicability and flexibility in arbitrary-scale.", + "bbox": [ + 496, + 306, + 895, + 549 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Arbitrary-scale super-resolution. To tackle this problem, very recently, the more practical setup of arbitrary-scale SR is considered, which aims to super-resolute images with arbitrary scales by a single model. MetaSR [28] makes the first attempt to propose an arbitrary-scale meta-upscale module. To improve the performance, many arbitrary-scale methods [64, 73] are proposed. With the help of implicit neural representation [2, 10, 15, 22, 33, 51, 52, 55, 57, 58, 62, 63], LIIF [13] predicts the RGB value at an arbitrary query coordinate by taking an image coordinate and features of backbone around the coordinate. The features are extracted by single image super-resolution methods, e.g., EDSR [44], RDN [88] and SwinIR [40]. To improve the performance, existing methods [39, 78] propose to integrate more features in SR models. For example, LTE [39] proposes a local texture estimator by characterizing image textures in the Fourier space. UltraSR [78] integrates spatial coordinates and periodic encoding in the implicit network. These methods use the bilinear interpolation to ensemble nearby features. However, such an ensemble way has no learnable parameters. Recently, ITSRN [79] learns the weights by taking the coordinate distance and scale token into a mapping. Most methods learn the ensemble without the feature similarity.", + "bbox": [ + 496, + 553, + 895, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "1797", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/a5f7fd21f2f55c2745db7f466b043c09a66a1933fc853e94ab22e48dda0502fa.jpg", + "image_caption": [ + "Figure 3. The architecture of our continuous implicit attention-in-attention network. Given an LR image, the encoder extracts features as latent codes. For a query point, we have a query feature and key features close to the query point, and the scale-aware non-local attention module extracts non-local features as value. Last, we use the triple of query, key and value to predict the RGB value at the query point." + ], + "image_footnote": [], + "bbox": [ + 106, + 89, + 861, + 325 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Preliminary and Motivation", + "text_level": 1, + "bbox": [ + 76, + 382, + 338, + 398 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Let $I$ be a continuous image, and $\\pmb{x}$ be a 2D coordinate of a pixel in the image $\\pmb{I}$ . Formally, given a 2D coordinate $\\pmb{x}$ in the continuous image domain and latent code $\\pmb{Z}$ extracted by deep neural networks, the RGB value can be predicted by an implicit image function which can be defined as follows,", + "bbox": [ + 75, + 404, + 470, + 479 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {I} (\\boldsymbol {x}) = f (\\boldsymbol {Z}, \\boldsymbol {x}), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 212, + 487, + 470, + 503 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where the implicit image function can be parameterized by a multilayer perceptron (MLP). Note that this implicit image function is shared by all images. For recent SISR methods [40, 86], the implicit image function $f$ can be implemented as a PixelShuffle [60] with convolutions with a specific scale. However, these methods are independent on the coordinates, leading to an issue that they only adapt to the specific scale and are inflexible to synthesize arbitrary-scale images.", + "bbox": [ + 75, + 507, + 468, + 627 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To predict the RGB value $I_{q}$ at an arbitrary query coordinate $x_{q}$ , most existing methods [13, 39, 79] propose to compute the RGB value at coordinate $x_{q}$ by directly ensembling its neighborhood information,", + "bbox": [ + 75, + 628, + 472, + 688 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\pmb {I} _ {q} := \\pmb {I} (\\pmb {x} _ {q}) = \\sum_ {(i, j) \\in \\mathcal {I}} w _ {i, j} \\cdot f (\\pmb {Z} _ {i, j} ^ {*}, \\pmb {x} _ {q} - \\pmb {x} _ {i, j} ^ {*}), (2)\n$$\n", + "text_format": "latex", + "bbox": [ + 89, + 695, + 472, + 722 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mathcal{I}$ is the local region centered at the query coordinate $x_{q}$ , e.g., $\\mathcal{I}$ can be top-left, top-right, bottom-left, bottom-right coordinates, and $w_{i}$ is the weight of the neighboring pixel $x_{i,j}^{*}$ , which is calculated w.r.t. the area of the rectangle between $x_{q}$ and $x_{i,j}^{*}$ , as shown in Figure 4. However, the performance improvement is limited because the ensemble weight $w_{i,j}$ is purely based on the coordinates. The visual similarities are completely ignored in the weight calculation. Besides, these methods only consider the nearest latent codes, leading to a limited receptive field. In this paper, our goal is to learn the weights adaptively by leveraging both visual information and coordinate information.", + "bbox": [ + 75, + 726, + 472, + 907 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4. Proposed Method", + "text_level": 1, + "bbox": [ + 498, + 382, + 671, + 398 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The above local ensemble Eqn. (2) has a similar form to the attention mechanism. Specifically, the weights $w_{i}$ in Eqn. (2) can be modeled as an attention map and the latter term $f$ can be the value in the attention mechanism. Such an attention map is related to the similarities of the latent code and coordinate, and it can be calculated using both query and key which can integrate the latent code and coordinate information. In this sense, attention models can be used to mitigate the drawbacks of previous methods (i.e., calculated purely based on coordinates without considering any visual information) by learning the soft weights from both visual and coordinate information. However, the use of attention in the implicit function is non-trivial because standard self-attention [68] and neighborhood attention [26] mechanisms are based on visual features, and not conditioned on the coordinate information. To exploit the continuous representation learning from both the coordinate information and visual features, we propose a new attention for arbitrary-scale SR.", + "bbox": [ + 496, + 407, + 893, + 679 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Continuous implicit attention-in-attention. The architecture of the proposed continuous implicit attention-in-attention for arbitrary-scale SR (called CiaoSR) is shown in Figure 3. It is called attention-in-attention because the value contains an additional embedded attention that additionally captures non-local information of repetitive patterns. Specifically, given an LR image, our attention is provided with both the visual features and coordinate information to learn good ensemble weights for the local neighboring latent codes for any query at any arbitrary-scale $s$ .", + "bbox": [ + 496, + 686, + 893, + 837 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Given a query coordinate $\\pmb{x}_q$ in the upsampled image, the known LR pixels in the local region centered at $\\pmb{x}_q$ are considered for the local ensemble. We define $\\pmb{x}_k$ and $\\pmb{x}_v$ as the key and value coordinates of the local neighbor. Let $Q, K$", + "bbox": [ + 496, + 839, + 893, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "1798", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "and $V$ be the latent code of query, key and value on the corresponding coordinates, respectively. Given pairs of coordinates and latent codes $(Q, x_{q})$ , $(K, x_{k})$ and $(V, x_{v})$ , we define a new implicit attention function (called i-Attention) to predict RGB values at the given query $x_{q}$ ,", + "bbox": [ + 75, + 90, + 470, + 167 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {I} _ {q} = \\mathrm {i} - \\text {A t t e n t i o n} (\\boldsymbol {Q}, \\boldsymbol {K}, \\boldsymbol {V}; \\boldsymbol {x} _ {q}, \\boldsymbol {x} _ {k}, \\boldsymbol {x} _ {v}), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 138, + 176, + 468, + 194 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where the value feature $V$ can be embedded with another attention to aggregate more non-local information.", + "bbox": [ + 75, + 203, + 468, + 233 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As shown in Figure 2, our implicit attention is different from existing attention mechanisms by considering the additional coordinate information. In particular, the self-attention mechanism [68] without considering coordinates cannot predict pixel features that are not on the grid. The coordinate-based implicit attention calculates the distance between a key $\\boldsymbol{x}_k$ and query coordinate $\\boldsymbol{x}_q$ , which ignores the visual information and similarity of the features.", + "bbox": [ + 75, + 234, + 470, + 354 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Implicit attention to learn weights for the local ensemble.", + "text_level": 1, + "bbox": [ + 75, + 357, + 470, + 372 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Unlike previous methods [13, 39] which directly use the normalized area as the ensemble weights, we propose to first use an attention to learn better weights for the ensemble of the local region. Our attention takes both the coordinate distance $r$ , the local features $F_{l}$ , as well as non-local features $G$ into account for the learning of the weight. Formally, the attention can be written as", + "bbox": [ + 75, + 372, + 468, + 477 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {I} _ {q} = \\phi_ {q} \\left(\\sum_ {(i, j) \\in \\mathcal {I}} \\underbrace {\\sigma \\left(\\boldsymbol {Q} ^ {\\top} \\boldsymbol {K} _ {i , j}\\right)} _ {\\text {e n s e m b l e w e i g h t}} \\boldsymbol {V} _ {i, j}\\right), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 140, + 484, + 468, + 525 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\phi_q$ is the query network, $\\sigma$ is the Softmax function, and $\\mathcal{I}$ is the local region centered at $\\boldsymbol{x}_q$ . The query, key, and value for this local ensemble attention are defined as", + "bbox": [ + 75, + 534, + 468, + 579 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\left\\{ \\begin{array}{l} Q = \\boldsymbol {F} ^ {*}, \\\\ \\boldsymbol {K} _ {i, j} = \\phi_ {k} \\left(\\left[ \\boldsymbol {F} _ {i, j}, \\left(\\boldsymbol {r} _ {k}\\right) _ {i, j}, \\boldsymbol {s} \\right]\\right), \\\\ \\boldsymbol {V} _ {i, j} = \\phi_ {v} \\left(\\left[ \\boldsymbol {F} _ {i, j}, \\tilde {\\boldsymbol {F}} _ {i, j} \\right], \\left(\\boldsymbol {r} _ {v}\\right) _ {i, j}, \\boldsymbol {s} \\right]), \\end{array} \\right. \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 148, + 585, + 468, + 643 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\phi_{k}$ and $\\phi_v$ are the key and value networks, respectively, $F^{*}$ are the features of nearest neighbors on the grid of the LR image w.r.t. the query coordinate $\\pmb{x}_q$ , $F_{i,j}$ is a local feature calculated by using the unfolding operator to the latent codes at the $(i,j)$ -th position, which is equivalent to concatenating the latent codes in neighboring $3\\times 3$ region. $\\tilde{F}_{i,j}$ are non-local features, which will be introduced below in detail. $s = [s_h;s_w]$ is a two-dimensional scale, and $\\boldsymbol{r}_k$ and $\\boldsymbol{r}_v$ are the relative distance between the coordinates in the local region, i.e.,", + "bbox": [ + 75, + 652, + 470, + 803 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {r} _ {k} = \\boldsymbol {x} _ {q} - \\left(\\boldsymbol {x} _ {k}\\right) _ {i, j}, \\quad \\boldsymbol {r} _ {v} = \\boldsymbol {x} _ {q} - \\left(\\boldsymbol {x} _ {v}\\right) _ {i, j}. \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 133, + 814, + 468, + 830 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Embedded scale-aware attention for non-local features.", + "text_level": 1, + "bbox": [ + 75, + 840, + 470, + 854 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Given a query coordinate, the local ensemble attention only considers the keys and values in the small local region, resulting in a limited receptive field. However, there may exist", + "bbox": [ + 75, + 854, + 470, + 900 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/60e71634d4e595d01a36073e2dab1f03ab62f979760454334f462fe5a02e4a7d.jpg", + "image_caption": [ + "(a) Bilinear-based ensemble" + ], + "image_footnote": [], + "bbox": [ + 522, + 88, + 684, + 213 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/d144375c24c9c5b05cac966f0cca9b68ea41524564f55af2d1041fac4f2b5690.jpg", + "image_caption": [ + "(b) Attention-based ensemble", + "Figure 4. Comparison of different ensemble methods. (a) Most existing methods calculate the weights related to the area of the rectangle between the query coordinate and the nearest point. Such weights are equivalent to the weight of the bilinear interpolation. (b) Our attention-based ensemble calculates the attention map as the weights for feature aggregation. Besides, the ensemble method has a large receptive field for exploiting more information." + ], + "image_footnote": [], + "bbox": [ + 702, + 88, + 864, + 213 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "similar textures with different scales within a single image in other positions. For example, repetitive patterns in different scales (e.g., facades, windows, etc. in a building, as shown in Figure 4) can exist in different locations within one image. To exploit the useful non-local information, we propose an embedded scale-aware non-local attention module, which aggregates the information from all coordinates on the grid of the LR image, motivated by [49]. Specifically, we obtain query $\\tilde{Q}$ and value features $\\tilde{V}$ from the latent code $F$ of the LR image, and downsample $F$ to a scale $s'$ as a key feature $\\tilde{K}$ . We first calculate the pixel-wise similarity between the query located at $(i,j)$ and the key on other coordinates, and then aggregate the weighted information from all of them. Lastly, we use a convolutional layer $\\varphi$ for downsampling with the scale $s'$ normalized similarity matrix to aggregate the features. Formally, the non-local feature $\\tilde{F}$ located at $(i,j)$ can be calculated by", + "bbox": [ + 496, + 337, + 893, + 595 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {\\boldsymbol {F}} _ {i, j} = \\varphi \\left(\\sum_ {u, v} \\frac {\\exp \\left(\\tilde {\\boldsymbol {Q}} _ {i , j} ^ {\\top} \\tilde {\\boldsymbol {K}} _ {u , v}\\right)}{\\sum_ {u ^ {\\prime} , v ^ {\\prime}} \\exp \\left(\\tilde {\\boldsymbol {Q}} _ {i , j} ^ {\\top} \\tilde {\\boldsymbol {K}} _ {u ^ {\\prime} , v ^ {\\prime}}\\right)} \\tilde {\\boldsymbol {V}} _ {s ^ {\\prime} u, s ^ {\\prime} v} ^ {s ^ {\\prime} p \\times s ^ {\\prime} p}\\right), \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 501, + 603, + 890, + 655 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\tilde{V}_{s'u,s'v}^{s'p\\times s'p}$ is the value feature patch of the size $s^{\\prime}p\\times s^{\\prime}p$ located at $(s^{\\prime}u,s^{\\prime}v)$ , $\\tilde{Q},\\tilde{K}$ and $\\tilde{V}$ are the query, key and value's non-local features, which can be calculated by", + "bbox": [ + 496, + 665, + 890, + 715 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\left\\{ \\begin{array}{l} \\tilde {\\boldsymbol {Q}} = \\varphi_ {q} (\\boldsymbol {F}), \\\\ \\tilde {\\boldsymbol {K}} = \\varphi_ {k} (\\boldsymbol {F} \\downarrow_ {s ^ {\\prime}}), \\\\ \\tilde {\\boldsymbol {V}} = \\varphi_ {v} (\\boldsymbol {F}), \\end{array} \\right. \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 630, + 724, + 890, + 785 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\downarrow_{s'}$ is the downsampling operation with a given scale $s'$ which can be $\\{2,3,4\\}$ , $\\varphi_q$ , $\\varphi_k$ and $\\varphi_v$ are query, key and value networks. Different from [49], our method can integrate multi-scale features. In the experiment, we can search for relevant features in a small window (e.g., $256\\times 256$ ), instead of the entire image. In this way, our method can guarantee a large receptive field, and reduce computational cost.", + "bbox": [ + 496, + 794, + 893, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "1799", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/8f1ae4c130e95e19759865a24818c0aa14da82e567bf7f879cfaf980c6524970.jpg", + "table_caption": [ + "Table 1. Quantitative comparison with state-of-the-art methods for arbitrary-scale SR on the DIV2K validation set (PSNR (dB)). Bold indicates the best performance. $\\dagger$ indicates the implementation of [39]." + ], + "table_footnote": [], + "table_body": "
BackbonesMethodsIn-scaleOut-of-scale
×2×3×4×6×12×18×24×30
-Bicubic31.0128.2226.6624.8222.2721.0020.1919.59
EDSR [44]EDSR-baseline [44]34.5530.9028.94-----
EDSR-baseline-MetaSR [28]34.6430.9328.9226.6123.5522.0321.0620.37
EDSR-baseline-LIIF [13]34.6730.9629.0026.7523.7122.1721.1820.48
EDSR-baseline-ITSRN† [79]34.7130.9529.0326.7723.7122.1721.1820.49
EDSR-baseline-LTE [39]34.7231.0229.0426.8123.7822.2321.2420.53
EDSR-baseline-CiaoSR (ours)34.9131.1529.2326.9523.8822.3221.3220.59
RDN [88]RDN-baseline [88]34.9431.2229.19-----
RDN-MetaSR [28]35.0031.2729.2526.8823.7322.1821.1720.47
RDN-LIIF [13]34.9931.2629.2726.9923.8922.3421.3120.59
RDN-ITSRN† [79]35.0931.3629.3827.0623.9322.3621.3220.61
RDN-LTE [39]35.0431.3229.3327.0423.9522.4021.3620.64
RDN-CiaoSR (ours)35.1531.4229.4527.1624.0622.4821.4320.70
SwinIR [40]SwinIR-baseline [40]34.9431.2229.19-----
SwinIR-MetaSR† [28]35.1531.4029.3326.9423.8022.2621.2620.54
SwinIR-LIIF† [13]35.1731.4629.4627.1524.0222.4321.4020.67
SwinIR-ITSRN† [79]35.1931.4229.4827.1323.8322.3121.3120.55
SwinIR-LTE [39]35.2431.5029.5127.2024.0922.5021.4720.73
SwinIR-CiaoSR (ours)35.2931.5529.5927.2824.1522.5421.5120.74
", + "bbox": [ + 80, + 119, + 888, + 393 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 76, + 400, + 209, + 417 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Datasets. For the arbitrary-scale SR task, we follow [13, 39] and use DIV2K [1] as the training set, which consists of 800 images in 2K resolution. In testing, we evaluate the models on the DIV2K validation set and a wide range of standard benchmark datasets, including Set5 [3], Set14 [80], B100 [46], Urban100 [29], and Manga109 [47].", + "bbox": [ + 75, + 422, + 470, + 513 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Evaluation metrics. We use PSNR to evaluate the quality of the synthesized SR images. Following [13, 39], the PSNR value is calculated on all RGBs channels for the DIV2K validation set, and additionally also calculated on the Y channel (i.e., luminance) of the transformed YCbCr space for other benchmark test sets. In addition, other evaluation metrics (such as SSIM [76] and LPIPS [85]) are provided to evaluate the image quality.", + "bbox": [ + 75, + 518, + 468, + 640 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Implementation details. We follow the same experimental setup as prior works [13, 39]. To synthesize paired training data, we first crop the $48s \\times 48s$ patches as ground-truth (GT) images, and use the Bicubic downsampling in PyTorch to have LR images with the patch size of $48 \\times 48$ , where $s$ is a scale factor sampled in the uniform distribution $\\mathcal{U}(1,4)$ . We sample $48^2$ pixels from both GT images and the corresponding coordinates. We use the existing SR models (e.g., EDSR [44], RDN [88] and SwinIR [40]) as backbones by removing their upsampling modules. The detailed architecture of our implicit network is provided in the supplementary material. For the training, we use Adam [36] as the optimizer, and use L1 loss to train all models for 1000 epochs with the batch size of 16 per GPU. We set the learning rate as $1e - 4$ at the beginning and decay the learning rate by a factor of 0.5 every 200 epochs. These experimental settings are identical to [13, 28, 39].", + "bbox": [ + 75, + 643, + 470, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.1. Comparisons with State-of-the-Art", + "text_level": 1, + "bbox": [ + 500, + 400, + 803, + 417 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Quantitative results. In Tables 1-3, CiaoSR achieves the best performance with the highest PSNR across all datasets with all backbones concerned on both in-scale and out-of-scale distributions. For the in-scales, in particular, our model has a remarkable PSNR gain of $0.3\\mathrm{dB}$ on Urban100 $(\\times 4)$ , compared with the previous SOTA method [39] under the same RDN backbone. It is worth noting that CiaoSR with the RDN backbone [88] can surpass the performance of the better backbone SwinIR [40] (with its original upsampling module). With the help of the continuous-scale training, a single SR model trained with our method has better generalization performance than vanilla SR models (including EDSR [44], RDN [88] and SwinIR [40] trained on specific scale. For the out-of-scale experiments, our model also achieves the best generalization performance on unseen scales. The vanilla EDSR-baseline [44], RDN [88] and SwinIR [40] cannot be applied on the out-of-scale experiments because their decoder can only predict the scale that it was trained on. We observe that our method yields the largest improvement on Urban100 over existing arbitrary-scale methods (e.g., LIIF [13] and LTE [39]).", + "bbox": [ + 496, + 424, + 893, + 742 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Qualitative results. In Figure 5, we provide qualitative comparisons with other arbitrary-scale SR methods. Our model is able to synthesize the SR images with sharper textures than other methods. Taking the second line as an example, the textures in the LR image are degraded, but CiaoSR is still able to restore the textures of the building. In contrast, other methods can only restore part of the textures possibly due to the limitations of the local ensemble and insufficient features. More visual results on more test sets are put in the supplementary materials.", + "bbox": [ + 496, + 750, + 893, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "1800", + "bbox": [ + 483, + 945, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/00a0cc65db05cffdb5255603a9bff595f2db2483a88dc037b0798c0b29de821c.jpg", + "table_caption": [ + "Table 2. Quantitative comparison with state-of-the-art methods for in-scale SR on benchmark datasets (PSNR (dB)). Bold indicates the best performance. † indicates our implementation." + ], + "table_footnote": [], + "table_body": "
MethodsSet5 [3]Set14 [80]B100 [46]Urban100 [29]Manga109 [47]
×2×3×4×2×3×4×2×3×4×2×3×4×2×3×4
RDN [88]38.2434.7132.4734.0130.5728.8132.3429.2627.7232.8928.8026.6139.1834.1331.00
RDN-MetaSR [28]38.2234.6332.3833.9830.5428.7832.3329.2627.7132.9228.8226.55---
RDN-LIIF [13]38.1734.6832.5033.9730.5328.8032.3229.2627.7432.8728.8226.6839.2634.2131.20
RDN-ITSRN† [79]38.2334.7632.5534.1930.5928.8832.3829.3227.7933.0728.9626.7739.3434.3931.37
RDN-LTE [39]38.2334.7232.6134.0930.5828.8832.3629.3027.7733.0428.9726.8139.2834.3231.30
RDN-CiaoSR (ours)38.2934.8532.6634.2230.6528.9332.4129.3427.8333.3029.1727.1139.5134.5731.57
SwinIR [40]38.3534.8932.7234.1430.7728.9432.4429.3727.8333.4029.2927.0739.6034.7431.67
SwinIR-MetaSR [28]38.2634.7732.4734.1430.6628.8532.3929.3127.7533.2929.1226.7639.4634.6231.37
SwinIR-LIIF [13]38.2834.8732.7334.1430.7528.9832.3929.3427.8433.3629.3327.1539.5734.6831.71
SwinIR-ITSRN† [79]38.2234.7532.6334.2630.7528.9732.4229.3827.8533.4629.3427.1239.6034.7531.74
SwinIR-LTE [39]38.3334.8932.8134.2530.8029.0632.4429.3927.8633.5029.4127.2439.6334.7931.79
SwinIR-CiaoSR (ours)38.3834.9132.8434.3330.8229.0832.4729.4227.9033.6529.5227.4239.6734.8431.91
", + "bbox": [ + 81, + 119, + 888, + 291 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/497ddbeb06f38210dad7a4a5fa41d3e53f06c7042ce7e509ef3fe8ab9bcbd2f4.jpg", + "table_caption": [ + "Table 3. Quantitative comparison with state-of-the-art methods for out-of-scale SR on benchmark datasets (PSNR (dB)). Bold indicates the best performance. † indicates our implementation." + ], + "table_footnote": [], + "table_body": "
MethodsSet5 [3]Set14 [80]B100 [46]Urban100 [29]Manga109 [47]
×6×8×12×6×8×12×6×8×12×6×8×12×6×8×12
RDN-MetaSR [28]29.0429.96-26.5124.97-25.9024.83-23.9922.59----
RDN-LIIF [13]29.1527.1424.8626.6425.1523.2425.9824.9123.5724.2022.7921.1527.3325.0422.36
RDN-ITSRN† [79]29.3227.2524.8626.6825.1723.2826.0124.9323.5824.2322.8121.1627.4525.0423.35
RDN-LTE [39]29.3227.2624.7926.7125.1623.3126.0124.9523.6024.2822.8821.2227.4925.1222.43
RDN-CiaoSR (ours)29.4627.3624.9226.7925.2823.3726.0725.0023.6424.5823.1321.4227.7025.4022.63
SwinIR-MetaSR [28]29.0927.0224.8226.5825.0923.3325.9424.8723.5924.1622.7521.3127.2924.9622.35
SwinIR-LIIF [13]29.4627.36-26.8225.34-26.0725.01-24.5923.14-27.6925.28-
SwinIR-ITSRN† [79]29.3127.2424.7926.7125.3223.3026.0524.9623.5724.5023.0621.3427.7225.2322.47
SwinIR-LTE [39]29.5027.35-26.8625.42-26.0925.03-24.6223.17-27.8325.42-
SwinIR-CiaoSR (ours)29.6227.4524.9626.8825.4223.3826.1325.0723.6824.8423.3421.6028.0125.6122.79
", + "bbox": [ + 81, + 330, + 888, + 479 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.2. Ablation Study", + "text_level": 1, + "bbox": [ + 76, + 503, + 228, + 518 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this section, we conduct ablation studies to investigate effect of each component in our architecture. Our CiaoSR consists of the attention-in-attention network and a scale-aware non-local attention module. Table 4 shows the contributions of each component on performance.", + "bbox": [ + 76, + 530, + 470, + 606 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Attention-in-attention. The attention-in-attention network is the main branch of our architecture. On the one hand, the network learns to ensemble weights with both the coordinates and the features. On the other hand, the network is able to aggregate the local and non-local information to improve the SR performance. To investigate how the attention mechanism affects the performance, we remove this module and replace it with an MLP to predict the weights of the ensemble. The coordinate information is still fed into the MLP to enable arbitrary scale prediction. As shown in Table 4, the method suffers a significant performance drop without the attention mechanism.", + "bbox": [ + 75, + 609, + 468, + 789 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Non-local attention. The embedded non-local attention is proposed to capture the scale-aware information from other locations of the images to improve the performance. We investigate the importance of the non-local attention in an implicit model by comparing the results with and without this module. As shown in Table 4, by introducing the nonlocal attention, a further performance gain is achieved. This", + "bbox": [ + 75, + 794, + 470, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "is because our module can exploit texture information from other scales in a large perceptual field.", + "bbox": [ + 498, + 503, + 890, + 534 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Training with different types of scales. We evaluate the effectiveness of our implicit model when trained with discrete (including single scale $\\{\\{2\\} /\\{3\\} /\\{4\\}\\}$ and multiple scales $\\{2,3,4\\}$ ) and continuous scales [1, 4], and show the results in Table 5. For the discrete type of scales, we observe that training our implicit model with a specific scale can also achieve significant improvement for that specific scale, compared with the original upsampling module in RDN [88]. Taking $\\times 2$ scale as an example, our method has an improvement of $0.24\\mathrm{dB}$ over RDN. However, training with a specific scale has poor generalization performance on unseen in-scale and out-of-scale. A few existing works [34, 44] consider the training with multiple discrete scales and found the performance of the joint learned model is comparable with single scale networks. Based on our implicit attention module, we are able to deal with more scales and exploit the correlation of multiple SR tasks at different scales. Unlike previous works [34, 44] which only achieve comparable performance compared with single scale models, CiaoSR achieves a better performance than training with one scale. We found our model has further improvement when trained on continuous scales. The improvement gain is attributed to our continuous representations design and our architecture, which enables efficient learning from cross-scale information.", + "bbox": [ + 496, + 537, + 893, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "1801", + "bbox": [ + 483, + 945, + 513, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/017229726e5527e13d8aef0b37d9453d1e3d2ebdfac7110c5c278dd744ee43ef.jpg", + "image_caption": [ + "004, $\\times 4$ Urban100 [29]" + ], + "image_footnote": [], + "bbox": [ + 84, + 88, + 326, + 231 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/2fe773b1ddcce8e0ee21e1300ad8e6b8977e0249c147971b975feee4dd033d43.jpg", + "image_caption": [ + "Bicubic" + ], + "image_footnote": [], + "bbox": [ + 330, + 88, + 464, + 152 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/c00f007273de11b08d3527a8eb7d8f8fef086df0bddd263fca24865ff1866cb3.jpg", + "image_caption": [ + "RDN [88]" + ], + "image_footnote": [], + "bbox": [ + 467, + 88, + 602, + 152 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/8abf6fcadcaebe33966e339674e949c22ddb0871faddbf91e1346c1131aa3610.jpg", + "image_caption": [ + "RDN-MetaSR [28]" + ], + "image_footnote": [], + "bbox": [ + 604, + 89, + 741, + 152 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/27887c0a76eec3bc040c897bd324471f33bcbfb139a7b7be652212f64caa3bd0.jpg", + "image_caption": [ + "RDN-LIIF [13]" + ], + "image_footnote": [], + "bbox": [ + 743, + 89, + 880, + 152 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/c0e51091767761fb3c9ceb37a57853b63679be7157190a5af1a4604e9cd3db20.jpg", + "image_caption": [ + "TaiyouNiSmash, $\\times 8$ ,Manga109 [47]" + ], + "image_footnote": [], + "bbox": [ + 84, + 246, + 325, + 391 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/2bcfdc0db4a5e5c0425a5ca1ec1ba3a26f7573bcc8d03650e461a3d6ca724f61.jpg", + "image_caption": [ + "RDN-ITSRN [79]" + ], + "image_footnote": [], + "bbox": [ + 328, + 167, + 465, + 231 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/2d6d904b3e21f45e9b65f82837ecd4bd3e955473a6aab3b075234602a51270eb.jpg", + "image_caption": [ + "RDN-LTE [39]" + ], + "image_footnote": [], + "bbox": [ + 467, + 167, + 602, + 231 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/292dfd17c80cfb1ae4a848b5470d1d40d7c46a5bd5ce26b8d8d5cae24925382e.jpg", + "image_caption": [ + "RDN-CiaoSR (ours)" + ], + "image_footnote": [], + "bbox": [ + 604, + 167, + 741, + 231 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d7397286c223ff8215569bb6039996a8daa3c70efc7bb33ab1c40a0020662355.jpg", + "image_caption": [ + "GT" + ], + "image_footnote": [], + "bbox": [ + 743, + 167, + 880, + 231 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/342bf68f4fc0645d74ce8570ea2952c88a231299ed6e76ac06ad1f6df0f173da.jpg", + "image_caption": [ + "Bicubic" + ], + "image_footnote": [], + "bbox": [ + 328, + 246, + 465, + 310 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/cf7e663ba7cca5a4992d8cc395c2a0052c25b682e29731ca6c9415eae3d95e67.jpg", + "image_caption": [ + "$\\mathbf{RDN}^*$ [88]" + ], + "image_footnote": [], + "bbox": [ + 467, + 246, + 602, + 310 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/abc35dd40df46d66c132bab6e4ac05ce6e573d5cbcb4baede219d2548093fa44.jpg", + "image_caption": [ + "RDN-MetaSR [28]" + ], + "image_footnote": [], + "bbox": [ + 604, + 246, + 741, + 310 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/0f7fed09bf0debfd75b82f8b4fe2f6962c8934750ba1dc1a27162702a614b926.jpg", + "image_caption": [ + "RDN-LIIF [13]" + ], + "image_footnote": [], + "bbox": [ + 743, + 246, + 880, + 310 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/86064fece19148d6164b92441003f33f93fac189a2422a55e68712c79c345c04.jpg", + "image_caption": [ + "RDN-ITSRN [79]" + ], + "image_footnote": [], + "bbox": [ + 328, + 325, + 465, + 388 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/17ec3b36a50b2eaf4cbcb4c3fcabed2ebe604e770b095ccbdf7059c4381ae2d2.jpg", + "image_caption": [ + "LTE [39]" + ], + "image_footnote": [], + "bbox": [ + 467, + 325, + 602, + 388 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/1e11f63b34a01d42b634a7dd52baf94259ad892dd975f767cbd09268ee9fae12.jpg", + "image_caption": [ + "RDN-CiaoSR (ours)" + ], + "image_footnote": [], + "bbox": [ + 604, + 325, + 741, + 388 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e21fae8d43d29ea55b5c3b1a07f9222489850a2263351299d40921bb91453c4e.jpg", + "image_caption": [ + "GT" + ], + "image_footnote": [], + "bbox": [ + 743, + 325, + 880, + 388 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/135999c514ff8e8d08702af099945753c294d6fad8758bd96d9ffd71c90544fa.jpg", + "image_caption": [ + "Figure 5. Visual comparison of different methods on benchmarks. \\*\\* means the model first synthesizes twice to $\\times 8$ images." + ], + "image_footnote": [], + "bbox": [ + 78, + 433, + 236, + 513 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/08a54ee2edf90283c42f6bc8f2cbe684813aa400c0992f721efc4e27ca3b1e79.jpg", + "image_caption": [ + "LR" + ], + "image_footnote": [], + "bbox": [ + 78, + 515, + 236, + 575 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/5c7802381b4a599dad06a3dac80a4420ab7247a6e35ae8d2d1ce514017533deb.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 243, + 433, + 398, + 513 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/db09f92a251b8f0eba18effab240fa418673056ba0ef04aeece7e330ca6c1dee.jpg", + "image_caption": [ + "RealSR" + ], + "image_footnote": [], + "bbox": [ + 243, + 515, + 398, + 577 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/c04769fb9e91c1b33811499688678a7f710e60c35f53268bffe7614b837e92f7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 406, + 433, + 562, + 513 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d6003f11db1ca87a94d12b24cec1ebbbe4093a71f6c072b2ffc2c7f2dcfdf131.jpg", + "image_caption": [ + "BSRGAN" + ], + "image_footnote": [], + "bbox": [ + 406, + 515, + 562, + 575 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/931cf6292228ee8a7a9ac79e2b6e75d129b6e19c535c142a1884e2e53e10e9f9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 568, + 433, + 725, + 513 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/69e488a7c5262c818a2f8769b4323953a87e785a943dbe0916f2f2345b604a2c.jpg", + "image_caption": [ + "Real-ESRGAN", + "Figure 6. Visual comparison of different methods on the RealSRSet [82] and DPED [30] dataset $(\\times 8)$ ." + ], + "image_footnote": [], + "bbox": [ + 568, + 515, + 725, + 575 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b256a9a1bc3a33ea924a66e938a317c36a9c18dc1321dc08dc1b38b723ccf74e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 733, + 433, + 861, + 513 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/3fffe2e75369a2f7440ec9ebff286f31cdcbafdc9a5716c67ea03276177ebfcb.jpg", + "image_caption": [ + "Ours" + ], + "image_footnote": [], + "bbox": [ + 733, + 515, + 861, + 575 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.3. Further Study", + "text_level": 1, + "bbox": [ + 76, + 616, + 223, + 632 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Effect on synthesis steps. Our trained model is able to synthesize SR images with any given arbitrary scale in one step. A natural alternative would be to synthesize a large scale in multiple steps. Taking a large scale $(e.g., \\times 12)$ as an example, SR images can be generated by two steps $(e.g., \\rightarrow \\times 2 \\rightarrow \\times 12)$ , or three steps $(e.g., \\rightarrow \\times 2 \\rightarrow \\times 4 \\rightarrow \\times 12)$ . We compare the performance of synthesizing a large scale in one and multiple steps, and the results on the Urban100 and Manga109 datasets are shown in Table 6. We observe that our way of synthesizing SR images in one step has the best performance under the PSNR and SSIM metrics. In contrast, the performance gets worse with more steps because errors can accumulate along the multiple synthesizing steps.", + "bbox": [ + 75, + 638, + 470, + 835 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Model size and inference time. To demonstrate the efficiency of our proposed method, we investigate the model size and inference time of different arbitrary-scale SR methods. Results are shown in Table 7. Specifically, we set", + "bbox": [ + 75, + 839, + 470, + 901 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "the input size as $192 \\times 192$ and calculate the inference time on an NVIDIA Quadro RTX 6000. Our implicit model has much fewer parameters than the previous SOTA methods (e.g., LTE [39]). Our model has the best performance although it requires a little more inference time due to exploiting the non-local features.", + "bbox": [ + 496, + 617, + 893, + 708 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "More evaluation metrics. To further demonstrate the effectiveness of our proposed method, we apply more evaluation metrics (e.g., SSIM [76] and LPIPS [85]) to evaluate the image quality across different methods on the Urban100 dataset. These evaluation metrics are widely used in the SR community. In general, higher SSIM and lower LPIPS correspond to better performance. As shown in Table 8, our method achieves the highest SSIM and lowest LPIPS, and thus has the best performance on both in-scale and out-of-scale distributions. Compared with other methods, our proposed CiaoSR is able to maintain more structural texture and perceptual information.", + "bbox": [ + 496, + 719, + 893, + 901 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "1802", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/728339d8915a08bb127c9a7f70cbc47d99b1e88ceb371f45cee0429910629cf7.jpg", + "table_caption": [ + "Table 4. Ablation study on each component of our networks on Urban100. We use RDN [88] as the backbone." + ], + "table_footnote": [], + "table_body": "
Attention-in-attention✗✓✓
Scale-aware Attention Network✗✗✓
In-scale×232.8733.2433.30
×328.8229.1029.17
×426.6926.9627.11
Out-of-scale×624.2224.5024.58
×822.8022.9823.13
", + "bbox": [ + 81, + 119, + 442, + 210 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/a13b0f46dea86ede9d1b23da663702165834a8f74220097ef178d7f6924f77af.jpg", + "table_caption": [ + "Table 6. Comparison of (PSNR and SSIM) for different synthesis steps on Urban100 [29] and Manga109 [47]." + ], + "table_footnote": [], + "table_body": "
TypeSynthesis stepsUrban100 [29]Manga109 [47]
PSNRSSIMPSNRSSIM
Multiple steps→×2→×4→×1221.280.55722.460.720
→×2→×1221.320.55822.530.721
One step→×1221.420.56122.630.723
", + "bbox": [ + 81, + 246, + 457, + 308 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/a98176b4c2a9c30e0183a7d2e9475a4696c592d0e2ef587a4cbd24a039fe565b.jpg", + "table_caption": [ + "Table 5. Ablation study on training our implicit model with different types of scales on Urban100." + ], + "table_footnote": [], + "table_body": "
TypeTraining scale sIn-scaleOut-of-scale
×2×3×4×6×8
Discretes ∈ {2}33.1327.0125.6022.2722.09
s ∈ {3}31.3929.0625.7723.4422.16
s ∈ {4}31.4227.8726.8824.2822.85
s ∈ {2,3,4}33.1529.1427.0224.4723.03
Continuouss ∈ [1,4]33.3029.1727.1124.5823.13
", + "bbox": [ + 455, + 119, + 897, + 210 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/70a26f1aae6c8588fa94d0277b1af4cf490ca018537965551de2989e8b4b8c38.jpg", + "table_caption": [ + "Table 7. Comparisons of model size, inference time and performance gain of different models." + ], + "table_footnote": [], + "table_body": "
Different modelsMeta-SR [28]LIIF [13]ITSRN [79]LTE [39]CiaoSR
Model size (M)1.71.60.71.71.4
Inference time (ms)237171343148528
PSNR (dB)26.5526.6826.7726.8127.11
Performance gain (dB)-0.060.070.160.20.5
", + "bbox": [ + 468, + 244, + 897, + 308 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/50a0dd81c943024c1b08292bf6bc853ee4a87d4a4df1f038acda8452604e3352.jpg", + "table_caption": [ + "Table 8. Comparison of more evaluation metrics (SSIM↑ [76] and LPIPS↓ [85]) for different methods on Urban100." + ], + "table_footnote": [], + "table_body": "
RDN-MethodIn-scaleOut-of-scale
×2×3×4×6×8
SSIMLPIPSSSIMLPIPSSSIMLPIPSSSIMLPIPSSSIMLPIPS
LIIF [13]0.9340.1040.8660.1970.8040.2670.7040.3600.6360.425
MetaSR [28]0.9350.1040.8660.1950.8050.2610.6980.3530.6240.423
ITSRN [79]0.9360.1020.8680.1930.8070.2590.7030.3500.6330.417
LTE [39]0.9360.1040.8680.1950.8070.2640.7070.3560.6390.421
CiaoSR0.9380.1000.8710.1880.8140.2550.7180.3460.6510.408
", + "bbox": [ + 81, + 344, + 580, + 450 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/81a3c1abd1c47c3f43a1d4b723589301e14339ac9eedafba785e43619b2d3cd4.jpg", + "table_caption": [ + "Table 9. Quantitative comparisons of different Realworld SR methods on RealSRSet [82]." + ], + "table_footnote": [], + "table_body": "
MethodsScaleNIQE↓BRISQUE↓PIQE↓
RealSR [5]×84.794923.322829.3976
×163.394621.639021.4116
BSRGAN [82]×84.700737.063836.1889
×164.140839.625237.9176
Real-ESRGAN [74]×84.528040.652159.8915
×165.644253.936484.3611
Ours-real×83.589430.921929.8554
×163.946141.814558.8244
", + "bbox": [ + 591, + 344, + 897, + 450 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.4. Real-World Arbitrary-Scale Image SR", + "text_level": 1, + "bbox": [ + 76, + 472, + 408, + 488 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We extend our proposed method to real-world applications with the ultimate goal of synthesizing real images with arbitrary scales. Specifically, we train our model on the DIV2K dataset [1], and test on the RealSRSet [82] and DPED [30] datasets. We apply the practical degradation models (including blur, noise, compression, etc.) of BSRGAN [82] and Real-ESRGAN [74] to synthesize paired training data on DIV2K [1]. We set the HR patch size of training data as 256, and the batch size as 48. We use Adam [36] as the optimizer, and adopt the exponential moving average (EMA). We first train our model with L1 loss for 1000K iterations with the learning rate of $2 \\times 10^{-4}$ and train with L1 loss, perceptual loss and GAN loss together for 400K iterations with the learning rate of $1 \\times 10^{-4}$ . More training details can be found in the supplementary materials.", + "bbox": [ + 75, + 491, + 468, + 717 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Since there are no arbitrary-scale SR methods applied to real-world images, we mainly compare with the state-of-the-art real-world image SR models, including RealSR [5], BSRGAN [82] and Real-ESRGAN [74]. Because there are no ground-truth (GT) real images, we apply non-reference image quality assessment (e.g., NIQE [54], BRISQUE [53] and PIQE [69]) to evaluate the image quality. As shown in Table 9, our model does not optimize for all metrics since these metrics cannot well match the actual human perceptual [4]. To address this, we further compare the visual results on the RealSRSet [82] and DPED [30] datasets in Figure 6. Our proposed method is able to synthesize real", + "bbox": [ + 75, + 719, + 470, + 901 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "SR images with realistic textures. In particular, our method is able to synthesize more details of a dog in the first line and more clear digit '5' in the second line, compared with other methods. In contrast, other methods often introduce unnatural artifacts within images, especially for out-of-scale distributions. To further investigate the effectiveness of our model, we provide more visual results of our real model in the supplementary materials.", + "bbox": [ + 496, + 473, + 890, + 594 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 607, + 617, + 622 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we have proposed a novel continuous implicit attention-in-attention network for arbitrary-scale image super-resolution, called CiaoSR. Specifically, our CiaoSR first introduces a scale-aware non-local attention in our attention to exploit more relevant features, and then predicts RGB values by locally ensembling features with the learnable attention weights. More importantly, CiaoSR has good flexibility and applicability since it can be used behind any SR backbone to boost the performance. Extensive experiments on all benchmark datasets demonstrate the superiority of our CiaoSR in both SISR and arbitrary-scale SR tasks. Besides, our CiaoSR has good generalization performance on both in-scale and out-of-scale distributions. Last, we extend our method to the real-world application to synthesize real SR images with arbitrary scale.", + "bbox": [ + 496, + 625, + 893, + 851 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements: This work was partly supported by the Huawei Fund, the ETH Zürich General Fund (OK) and the Alexander von Humboldt Foundation.", + "bbox": [ + 498, + 854, + 890, + 898 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "1803", + "bbox": [ + 483, + 945, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 174, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Eirikur Agustsson and Radu Timofte. Ntire 2017 challenge on single image super-resolution: Dataset and study. In IEEE Conference on Computer Vision and Pattern Recognition Workshops, pages 126-135, 2017. 5, 8", + "[2] Matan Atzmon and Yaron Lipman. Sal: Sign agnostic learning of shapes from raw data. In IEEE Conference on Computer Vision and Pattern Recognition, pages 2565-2574, 2020. 2", + "[3] Marco Bevilacqua, Aline Roumy, Christine Guillemot, and Marie line Alberi Morel. Low-complexity single-image super-resolution based on nonnegative neighbor embedding. In British Machine Vision Conference, pages 135.1-135.10, 2012. 5, 6", + "[4] Yochai Blau, Roey Mechrez, Radu Timofte, Tomer Michaeli, and Lihi Zelnik-Manor. The 2018 pirm challenge on perceptual image super-resolution. In European Conference on Computer Vision Workshops, 2018. 8", + "[5] Jianrui Cai, Hui Zeng, Hongwei Yong, Zisheng Cao, and Lei Zhang. Toward real-world single image super-resolution: A new benchmark and a new model. In IEEE Conference on International Conference on Computer Vision, pages 3086-3095, 2019. 8", + "[6] Jiezhang Cao, Yawei Li, Kai Zhang, and Luc Van Gool. Video super-resolution transformer. arXiv preprint arXiv:2106.06847, 2021. 2", + "[7] Jiezhang Cao, Jingyun Liang, Kai Zhang, Yawei Li, Yulun Zhang, Wenguan Wang, and Luc Van Gool. Reference-based image super-resolution with deformable attention transformer. In European Conference on Computer Vision, pages 325-342, 2022. 2", + "[8] Jiezhang Cao, Jingyun Liang, Kai Zhang, Wenguan Wang, Qin Wang, Yulun Zhang, Hao Tang, and Luc Van Gool. Towards interpretable video super-resolution via alternating optimization. In European Conference on Computer Vision, pages 393-411, 2022. 2", + "[9] Lukas Cavigelli, Pascal Hager, and Luca Benini. Cas-cnn: A deep convolutional neural network for image compression artifact suppression. In 2017 International Joint Conference on Neural Networks, pages 752-759, 2017. 2", + "[10] Rohan Chabra, Jan E Lenssen, Eddy Ig, Tanner Schmidt, Julian Straub, Steven Lovegrove, and Richard Newcombe. Deep local shapes: Learning local sdf priors for detailed 3d reconstruction. In European Conference on Computer Vision, pages 608-625, 2020. 2", + "[11] Hanting Chen, Yunhe Wang, Tianyu Guo, Chang Xu, Yiping Deng, Zhenhua Liu, Siwei Ma, Chunjing Xu, Chao Xu, and Wen Gao. Pre-trained image processing transformer. In IEEE Conference on Computer Vision and Pattern Recognition, pages 12299-12310, 2021. 2", + "[12] Xiangyu Chen, Xintao Wang, Jiantao Zhou, and Chao Dong. Activating more pixels in image super-resolution transformer. arXiv preprint arXiv:2205.04437, 2022. 2", + "[13] Yinbo Chen, Sifei Liu, and Xiaolong Wang. Learning continuous image representation with local implicit image function. In IEEE Conference on Computer Vision and Pattern Recognition, pages 8628-8638, 2021. 1, 2, 3, 4, 5, 6, 7, 8" + ], + "bbox": [ + 78, + 114, + 470, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[14] Yunjin Chen and Thomas Pock. Trainable nonlinear reaction diffusion: A flexible framework for fast and effective image restoration. IEEE Transactions on Pattern Analysis and Machine Intelligence, 39(6):1256-1272, 2016. 2", + "[15] Zhiqin Chen and Hao Zhang. Learning implicit fields for generative shape modeling. In IEEE Conference on Computer Vision and Pattern Recognition, pages 5939-5948, 2019. 2", + "[16] Wenlong Cheng, Mingbo Zhao, Zhiling Ye, and Shuhang Gu. Mfagan: A compression framework for memory-efficient on-device super-resolution gan. arXiv preprint arXiv:2107.12679, 2021. 2", + "[17] Tao Dai, Jianrui Cai, Yongbing Zhang, Shu-Tao Xia, and Lei Zhang. Second-order attention network for single image super-resolution. In IEEE Conference on Computer Vision and Pattern Recognition, pages 11065-11074, 2019. 2", + "[18] Xin Deng, Yutong Zhang, Mai Xu, Shuhang Gu, and Yiping Duan. Deep coupled feedback network for joint exposure fusion and image super-resolution. IEEE Transactions on Image Processing, 30:3098-3112, 2021. 2", + "[19] Chao Dong, Chen Change Loy, Kaiming He, and Xiaou Tang. Learning a deep convolutional network for image super-resolution. In European Conference on Computer Vision, pages 184–199, 2014. 2", + "[20] Xueyang Fu, Menglu Wang, Xiangyong Cao, Xinghao Ding, and Zheng-Jun Zha. A model-driven deep unfolding method for jpeg artifacts removal. IEEE Transactions on Neural Networks and Learning Systems, 2021. 2", + "[21] Xueyang Fu, Zheng-Jun Zha, Feng Wu, Xinghao Ding, and John Paisley. Jpeg artifacts reduction via deep convolutional sparse coding. In IEEE International Conference on Computer Vision, pages 2501-2510, 2019. 2", + "[22] Amos Gropp, Lior Yariv, Niv Haim, Matan Atzmon, and Yaron Lipman. Implicit geometric regularization for learning shapes. arXiv preprint arXiv:2002.10099, 2020. 2", + "[23] Shuhang Gu, Nong Sang, and Fan Ma. Fast image super resolution via local regression. In IEEE Conference on International Conference on Pattern Recognition, pages 3128-3131, 2012. 2", + "[24] Bahadir K Gunturk, Yucel Altunbasak, and Russell M Mersereau. Super-resolution reconstruction of compressed video using transform-domain statistics. IEEE Transactions on Image Processing, 13(1):33-43, 2004. 1", + "[25] Yong Guo, Jian Chen, Jingdong Wang, Qi Chen, Jiezhang Cao, Zeshuai Deng, Yanwu Xu, and Mingkui Tan. Closed-loop matters: Dual regression networks for single image super-resolution. In IEEE Conference on Computer Vision and Pattern Recognition, pages 5407-5416, 2020. 2", + "[26] Ali Hassani, Steven Walton, Jiachen Li, Shen Li, and Humphrey Shi. Neighborhood attention transformer. arXiv preprint arXiv:2204.07143, 2022. 3", + "[27] Kaiming He, Jian Sun, and Xiaou Tang. Single image haze removal using dark channel prior. IEEE Transactions on Pattern Analysis and Machine Intelligence, 33(12):2341-2353, 2010. 2", + "[28] Xuecai Hu, Haoyuan Mu, Xiangyu Zhang, Zilei Wang, Tieniu Tan, and Jian Sun. Meta-sr: A magnification-arbitrary network for super-resolution. In IEEE Conference on Computer" + ], + "bbox": [ + 501, + 92, + 893, + 901 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "1804", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Vision and Pattern Recognition, pages 1575-1584, 2019. 1, 2, 5, 6, 7, 8", + "[29] Jia-Bin Huang, Abhishek Singh, and Narendra Ahuja. Single image super-resolution from transformed self-exemplars. In IEEE Conference on Computer Vision and Pattern Recognition, pages 5197-5206, 2015. 5, 6, 7, 8", + "[30] Andrey Ignatov, Nikolay Kobyshev, Radu Timofte, Kenneth Vanhoey, and Luc Van Gool. Dslr-quality photos on mobile devices with deep convolutional networks. In IEEE International Conference on Computer Vision, pages 3277-3285, 2017. 7, 8", + "[31] Takashi Isobe, Xu Jia, Shuhang Gu, Songjiang Li, Shengjin Wang, and Qi Tian. Video super-resolution with recurrent structure-detail network. In European Conference on Computer Vision, pages 645-660, 2020. 2", + "[32] Xixi Jia, Sanyang Liu, Xiangchu Feng, and Lei Zhang. Focnet: A fractional optimal control network for image denoising. In IEEE Conference on Computer Vision and Pattern Recognition, pages 6054-6063, 2019. 2", + "[33] Chiyu Jiang, Avneesh Sud, Ameesh Makadia, Jingwei Huang, Matthias Nießner, Thomas Funkhouser, et al. Local implicit grid representations for 3d scenes. In IEEE Conference on Computer Vision and Pattern Recognition, pages 6001-6010, 2020. 2", + "[34] Jiwon Kim, Jung Kwon Lee, and Kyoung Mu Lee. Accurate image super-resolution using very deep convolutional networks. In IEEE Conference on Computer Vision and Pattern Recognition, pages 1646-1654, 2016. 2, 6", + "[35] Yoonsik Kim, Jae Woong Soh, Jaewoo Park, Byeongyong Ahn, Hyun-Seung Lee, Young-Su Moon, and Nam Ik Cho. A pseudo-blind convolutional neural network for the reduction of compression artifacts. IEEE Transactions on Circuits and Systems for Video Technology, 30(4):1121-1135, 2019. 2", + "[36] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In International Conference on Learning Representations, 2015. 5, 8", + "[37] Wei-Sheng Lai, Jia-Bin Huang, Narendra Ahuja, and Ming-Hsuan Yang. Deep laplacian pyramid networks for fast and accurate super-resolution. In IEEE Conference on Computer Vision and Pattern Recognition, pages 624-632, 2017. 2", + "[38] Christian Ledig, Lucas Theis, Ferenc Huszar, Jose Caballero, Andrew Cunningham, Alejandro Acosta, Andrew Aitken, Alykhan Tejani, Johannes Totz, Zehan Wang, et al. Photorealistic single image super-resolution using a generative adversarial network. In IEEE Conference on Computer Vision and Pattern Recognition, pages 4681-4690, 2017. 2", + "[39] Jaewon Lee and Kyong Hwan Jin. Local texture estimator for implicit representation function. In IEEE Conference on Computer Vision and Pattern Recognition, pages 1929-1938, 2022. 1, 2, 3, 4, 5, 6, 7, 8", + "[40] Jingyun Liang, Jiezhang Cao, Guolei Sun, Kai Zhang, Luc Van Gool, and Radu Timofte. Swinir: Image restoration using swin transformer. In IEEE International Conference on Computer Vision Workshops, pages 1833-1844, 2021. 1, 2, 3, 5, 6", + "[41] Jingyun Liang, Andreas Lugmayr, Kai Zhang, Martin Danelljan, Luc Van Gool, and Radu Timofte. Hierarchical conditional flow: A unified framework for image super-resolution" + ], + "bbox": [ + 78, + 90, + 470, + 901 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "and image rescaling. In IEEE Conference on International Conference on Computer Vision, 2021. 2", + "[42] Jingyun Liang, Guolei Sun, Kai Zhang, Luc Van Gool, and Radu Timofte. Mutual affine network for spatially variant kernel estimation in blind image super-resolution. In IEEE Conference on International Conference on Computer Vision, 2021. 2", + "[43] Jingyun Liang, Kai Zhang, Shuhang Gu, Luc Van Gool, and Radu Timofte. Flow-based kernel prior with application to blind super-resolution. In IEEE Conference on Computer Vision and Pattern Recognition, pages 10601-10610, 2021. 2", + "[44] Bee Lim, Sanghyun Son, Heewon Kim, Seungjun Nah, and Young Mu Lee. Enhanced deep residual networks for single image super-resolution. In IEEE Conference on Computer Vision and Pattern Recognition Workshops, pages 136-144, 2017. 1, 2, 5, 6", + "[45] Ding Liu, Bihan Wen, Yuchen Fan, Chen Change Loy, and Thomas S Huang. Non-local recurrent network for image restoration. arXiv preprint arXiv:1806.02919, 2018. 2", + "[46] David Martin, Charless Fowlkes, Doron Tal, and Jitendra Malik. A database of human segmented natural images and its application to evaluating segmentation algorithms and measuring ecological statistics. In IEEE Conference on International Conference on Computer Vision, pages 416-423, 2001. 5, 6", + "[47] Yusuke Matsui, Kota Ito, Yuji Aramaki, Azuma Fujimoto, Toru Ogawa, Toshihiko Yamasaki, and Kiyoharu Aizawa. Sketch-based manga retrieval using manga109 dataset. Multi-media Tools and Applications, 76(20):21811-21838, 2017. 5, 6, 7, 8", + "[48] Yiqun Mei, Yuchen Fan, and Yuqian Zhou. Image superresolution with non-local sparse attention. In IEEE Conference on Computer Vision and Pattern Recognition, pages 3517-3526, 2021. 2", + "[49] Yiqun Mei, Yuchen Fan, Yuqian Zhou, Lichao Huang, Thomas S Huang, and Honghui Shi. Image super-resolution with cross-scale non-local attention and exhaustive self-exemplars mining. In IEEE Conference on Computer Vision and Pattern Recognition, pages 5690-5699, 2020. 4", + "[50] Tomer Michaeli and Michal Irani. Nonparametric blind superresolution. In IEEE Conference on International Conference on Computer Vision, pages 945–952, 2013. 2", + "[51] Mateusz Michalkiewicz, Jhony K Pontes, Dominic Jack, Mahsa Baktashmotlagh, and Anders Eriksson. Implicit surface representations as layers in neural networks. In IEEE International Conference on Computer Vision, pages 4743-4752, 2019. 2", + "[52] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 2", + "[53] Anish Mittal, Anush K Moorthy, and Alan C Bovik. Blind/referenceless image spatial quality evaluator. In Asilomar Conference on Signals, Systems and Computers, pages 723-727, 2011. 8", + "[54] Anish Mittal, Rajiv Soundararajan, and Alan C Bovik. Making a \"completely blind\" image quality analyzer. IEEE Signal Processing Letters, 20(3):209-212, 2012. 8" + ], + "bbox": [ + 501, + 92, + 893, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "1805", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[55] Michael Niemeyer, Lars Mescheder, Michael Oechsle, and Andreas Geiger. Differentiable volumetric rendering: Learning implicit 3d representations without 3d supervision. In IEEE Conference on Computer Vision and Pattern Recognition, pages 3504-3515, 2020. 2", + "[56] Ben Niu, Weilei Wen, Wenqi Ren, Xiangde Zhang, Lianping Yang, Shuzhen Wang, Kaihao Zhang, Xiaochun Cao, and Haifeng Shen. Single image super-resolution via a holistic attention network. In European Conference on Computer Vision, pages 191-207, 2020. 2", + "[57] Michael Oechsle, Lars Mescheder, Michael Niemeyer, Thilo Strauss, and Andreas Geiger. Texture fields: Learning texture representations in function space. In IEEE International Conference on Computer Vision, pages 4531-4540, 2019. 2", + "[58] Songyou Peng, Michael Niemeyer, Lars Mescheder, Marc Pollefeys, and Andreas Geiger. Convolutional occupancy networks. In European Conference on Computer Vision, pages 523-540, 2020. 2", + "[59] Yali Peng, Lu Zhang, Shigang Liu, Xiaojun Wu, Yu Zhang, and Xili Wang. Dilated residual networks with symmetric skip connection for image denoising. Neurocomputing, 345:67-76, 2019. 2", + "[60] Wenzhe Shi, Jose Caballero, Ferenc Huszár, Johannes Totz, Andrew P Aitken, Rob Bishop, Daniel Rueckert, and Zehan Wang. Real-time single image and video super-resolution using an efficient sub-pixel convolutional neural network. In IEEE Conference on Computer Vision and Pattern Recognition, pages 1874-1883, 2016. 1, 3", + "[61] Wenzhe Shi, Jose Caballero, Christian Ledig, Xiahai Zhuang, Wenjia Bai, Kanwal Bhatia, Antonio M Marvao, Tim Dawes, Declan O'Regan, and Daniel Rueckert. Cardiac image superresolution with global correspondence using multi-atlas patchmatch. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 9-16, 2013. 1", + "[62] Vincent Sitzmann, Julien Martel, Alexander Bergman, David Lindell, and Gordon Wetzstein. Implicit neural representations with periodic activation functions. Advances in Neural Information Processing Systems, 33:7462-7473, 2020. 2", + "[63] Vincent Sitzmann, Michael Zollhöfer, and Gordon Wetzstein. Scene representation networks: Continuous 3d-structure-aware neural scene representations. Advances in Neural Information Processing Systems, 32, 2019. 2", + "[64] Sanghyun Son and Kyoung Mu Lee. Srwarp: Generalized image super-resolution under arbitrary transformation. In IEEE Conference on Computer Vision and Pattern Recognition, pages 7782-7791, 2021. 2", + "[65] Ying Tai, Jian Yang, Xiaoming Liu, and Chunyan Xu. Memnet: A persistent memory network for image restoration. In IEEE International Conference on Computer Vision, pages 4539-4547, 2017. 2", + "[66] Radu Timofte, Vincent De Smet, and Luc Van Gool. Anchored neighborhood regression for fast example-based superresolution. In IEEE Conference on International Conference on Computer Vision, pages 1920-1927, 2013. 2", + "[67] Radu Timofte, Vincent De Smet, and Luc Van Gool. A+: Adjusted anchored neighborhood regression for fast super-" + ], + "bbox": [ + 78, + 90, + 470, + 900 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "resolution. In Asian Conference on Computer Vision, pages 111-126, 2014. 2", + "[68] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Advances in Neural Information Processing Systems, 2017. 2, 3, 4", + "[69] N Venkatanath, D Praneeth, Maruthi Chandrasekhar Bh, Sumohana S Channappayya, and Swarup S Medasani. Blind image quality evaluation using perception based features. In National Conference on Communications, pages 1-6, 2015. 8", + "[70] Longguang Wang, Yingqian Wang, Xiaoyu Dong, Qingyu Xu, Jungang Yang, Wei An, and Yulan Guo. Unsupervised degradation representation learning for blind super-resolution. In IEEE Conference on Computer Vision and Pattern Recognition, pages 10581-10590, 2021. 2", + "[71] Longguang Wang, Yingqian Wang, Zhengfa Liang, Zaiping Lin, Jungang Yang, Wei An, and Yulan Guo. Learning parallel attention for stereo image super-resolution. In IEEE Conference on Computer Vision and Pattern Recognition, pages 12250-12259, 2019. 2", + "[72] Longguang Wang, Yingqian Wang, Zaiping Lin, Jungang Yang, Wei An, and Yulan Guo. Learning a single network for scale-arbitrary super-resolution. In IEEE Conference on International Conference on Computer Vision, pages 10581-10590, 2021. 2", + "[73] Longguang Wang, Yingqian Wang, Zaiping Lin, Jungang Yang, Wei An, and Yulan Guo. Learning a single network for scale-arbitrary super-resolution. In IEEE International Conference on Computer Vision, pages 4801-4810, 2021. 2", + "[74] Xintao Wang, Liangbin Xie, Chao Dong, and Ying Shan. Real-esrgan: Training real-world blind super-resolution with pure synthetic data. In International Conference on Computer Vision Workshops. 8", + "[75] Xintao Wang, Ke Yu, Shixiang Wu, Jinjin Gu, Yihao Liu, Chao Dong, Yu Qiao, and Chen Change Loy. Esrgan: Enhanced super-resolution generative adversarial networks. In European Conference on Computer Vision Workshops, pages 701-710, 2018. 2", + "[76] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE Transactions on Image Processing, 13(4):600-612, 2004. 5, 7, 8", + "[77] Yunxuan Wei, Shuhang Gu, Yawei Li, Radu Timofte, Longcun Jin, and Hengjie Song. Unsupervised real-world image super resolution via domain-distance aware training. In IEEE Conference on Computer Vision and Pattern Recognition, pages 13385-13394, 2021. 2", + "[78] Xingqian Xu, Zhangyang Wang, and Humphrey Shi. Ultrasr: Spatial encoding is a missing key for implicit image function-based arbitrary-scale super-resolution. arXiv preprint arXiv:2103.12716, 2021. 2", + "[79] Jingyu Yang, Sheng Shen, Huanjing Yue, and Kun Li. Implicit transformer network for screen content image continuous super-resolution. Advances in Neural Information Processing Systems, 34:13304-13315, 2021. 2, 3, 5, 6, 7, 8", + "[80] Roman Zeyde, Michael Elad, and Matan Protter. On single image scale-up using sparse-representations. In International" + ], + "bbox": [ + 503, + 92, + 893, + 900 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "1806", + "bbox": [ + 483, + 945, + 514, + 955 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Conference on Curves and Surfaces, pages 711-730, 2010. 5, 6", + "[81] Kai Zhang, Yawei Li, Wangmeng Zuo, Lei Zhang, Luc Van Gool, and Radu Timofte. Plug-and-play image restoration with deep denoiser prior. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2021. 2", + "[82] Kai Zhang, Jingyun Liang, Luc Van Gool, and Radu Timofte. Designing a practical degradation model for deep blind image super-resolution. In IEEE Conference on International Conference on Computer Vision, 2021. 7, 8", + "[83] Kai Zhang, Wangmeng Zuo, and Lei Zhang. Ffdnet: Toward a fast and flexible solution for cnn-based image denoising. IEEE Transactions on Image Processing, 27(9):4608-4622, 2018. 2", + "[84] Kai Zhang, Wangmeng Zuo, and Lei Zhang. Learning a single convolutional super-resolution network for multiple degradations. In IEEE Conference on Computer Vision and Pattern Recognition, pages 3262-3271, 2018. 2", + "[85] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In IEEE Conference on Computer Vision and Pattern Recognition, pages 586-595, 2018. 5, 7, 8", + "[86] Yulun Zhang, Kunpeng Li, Kai Li, Lichen Wang, Bineng Zhong, and Yun Fu. Image super-resolution using very deep residual channel attention networks. In European Conference on Computer Vision, pages 286-301, 2018. 1, 2, 3", + "[87] Yulun Zhang, Kunpeng Li, Kai Li, Bineng Zhong, and Yun Fu. Residual non-local attention networks for image restoration. arXiv preprint arXiv:1903.10082, 2019. 2", + "[88] Yulun Zhang, Yapeng Tian, Yu Kong, Bineng Zhong, and Yun Fu. Residual dense network for image super-resolution. In IEEE Conference on Computer Vision and Pattern Recognition, pages 2472-2481, 2018. 1, 2, 5, 6, 7, 8", + "[89] Yulun Zhang, Yapeng Tian, Yu Kong, Bineng Zhong, and Yun Fu. Residual dense network for image restoration. IEEE Transactions on Pattern Analysis and Machine Intelligence, 43(7):2480-2495, 2020. 2", + "[90] Shangchen Zhou, Jiawei Zhang, Wangmeng Zuo, and Chen Change Loy. Cross-scale internal graph neural network for image super-resolution. arXiv preprint arXiv:2006.16673, 2020.2", + "[91] Wilman WW Zou and Pong C Yuen. Very low resolution face recognition problem. IEEE Transactions on Image Processing, 21(1):327-340, 2011. 1" + ], + "bbox": [ + 78, + 90, + 470, + 728 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "1807", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 11 + } +] \ No newline at end of file diff --git a/2023/CiaoSR_ Continuous Implicit Attention-in-Attention Network for Arbitrary-Scale Image Super-Resolution/dcf45018-bd80-4a92-ad40-a481bde62c68_model.json b/2023/CiaoSR_ Continuous Implicit Attention-in-Attention Network for Arbitrary-Scale Image Super-Resolution/dcf45018-bd80-4a92-ad40-a481bde62c68_model.json new file mode 100644 index 0000000000000000000000000000000000000000..3a6edbb12142f858ff4d75ea42eea1ec8dcfb774 --- /dev/null +++ b/2023/CiaoSR_ Continuous Implicit Attention-in-Attention Network for Arbitrary-Scale Image Super-Resolution/dcf45018-bd80-4a92-ad40-a481bde62c68_model.json @@ -0,0 +1,3084 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.0, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.158, + 0.13, + 0.819, + 0.178 + ], + "angle": 0, + "content": "CiaoSR: Continuous Implicit Attention-in-Attention Network for Arbitrary-Scale Image Super-Resolution" + }, + { + "type": "text", + "bbox": [ + 0.135, + 0.203, + 0.836, + 0.222 + ], + "angle": 0, + "content": "Jiezhang Cao\\(^{1}\\) Qin Wang\\(^{1}\\) Yongqin Xian\\(^{1*}\\) Yawei Li\\(^{1}\\) Bingbing Ni\\(^{2}\\) Zhiming Pi\\(^{2}\\)" + }, + { + "type": "text", + "bbox": [ + 0.224, + 0.222, + 0.749, + 0.239 + ], + "angle": 0, + "content": "Kai Zhang\\(^{1\\dagger}\\) Yulun Zhang\\(^{1\\dagger}\\) Radu Timofte\\(^{1,3}\\) Luc Van Gool\\(^{1,4}\\)" + }, + { + "type": "text", + "bbox": [ + 0.207, + 0.239, + 0.765, + 0.257 + ], + "angle": 0, + "content": "\\(^{1}\\)ETH Zürich \\(^{2}\\)Huawei Inc. \\(^{3}\\)University of Wurzburg \\(^{4}\\)KU Leuven" + }, + { + "type": "text", + "bbox": [ + 0.317, + 0.26, + 0.651, + 0.275 + ], + "angle": 0, + "content": "https://github.com/caojiezhang/CiaoSR" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.309, + 0.314, + 0.327 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.342, + 0.474, + 0.706 + ], + "angle": 0, + "content": "Learning continuous image representations is recently gaining popularity for image super-resolution (SR) because of its ability to reconstruct high-resolution images with arbitrary scales from low-resolution inputs. Existing methods mostly ensemble nearby features to predict the new pixel at any queried coordinate in the SR image. Such a local ensemble suffers from some limitations: i) it has no learnable parameters and it neglects the similarity of the visual features; ii) it has a limited receptive field and cannot ensemble relevant features in a large field which are important in an image. To address these issues, this paper proposes a continuous implicit attention-in-attention network, called CiaoSR. We explicitly design an implicit attention network to learn the ensemble weights for the nearby local features. Furthermore, we embed a scale-aware attention in this implicit attention network to exploit additional non-local information. Extensive experiments on benchmark datasets demonstrate CiaoSR significantly outperforms the existing single image SR methods with the same backbone. In addition, CiaoSR also achieves the state-of-the-art performance on the arbitrary-scale SR task. The effectiveness of the method is also demonstrated on the real-world SR setting. More importantly, CiaoSR can be flexibly integrated into any backbone to improve the SR performance." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.715, + 0.21, + 0.73 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.733, + 0.472, + 0.854 + ], + "angle": 0, + "content": "Single image super-resolution (SISR), which aims to reconstruct a high-resolution (HR) image from a low-resolution (LR) one, has been widely employed in many practical applications [24, 61, 91]. However, deep neural networks (DNN)-based SISR methods are facing some limitations in some real-world scenarios with arbitrary scales. For example, camera users may want to enhance the digital zoom quality by super-resolving a photo or a video to" + }, + { + "type": "image", + "bbox": [ + 0.501, + 0.309, + 0.891, + 0.406 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.409, + 0.895, + 0.452 + ], + "angle": 0, + "content": "Figure 1. Comparison of different backbones and implicit models. Our proposed implicit neural network on RDN [88] has better performance than SwinIR [40]. Check Section 5 for details." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.459, + 0.895, + 0.566 + ], + "angle": 0, + "content": "continuous arbitrary scales. Most existing DNN-based SISR methods [40, 44, 86] need to train a series of models for all different scales separately. However, it can be impractical to store all these models on the device due to limited storage and computing power. Alternatively, arbitrary-scale image SR methods [13, 28, 39] aim to train a single network for all scales in a continuous manner." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.567, + 0.896, + 0.839 + ], + "angle": 0, + "content": "Most existing SISR methods [40,44,86] consist of a DNN and an upsampling module (e.g., pixel shuffling [60]) at a discrete scale. While substantial progresses have been made in the DNN backbones for SR, there is little attempt to study the upsampling module. A natural question to ask is: Does the pixel shuffling hinder the potential of SR models? One limitation of the pixel shuffling module is that it cannot synthesize SR images at large unseen and continuous scales. To tackle this, one can treat synthesizing different-scale SR images as a multi-task learning problem, and train a specific upsampling module for each scale [44]. However, these tasks are dependent and highly inter-related. Neglecting the correlation of different-scale SR tasks may lead to discrete representations and limited performance. Under a certain capacity of a network, training a model on multi-tasks may sacrifice the performance or have the comparable performance on each task. These above disadvantages limit its applicability and flexibility in the real-world scenarios." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.897, + 0.903 + ], + "angle": 0, + "content": "To address these, most existing arbitrary-scale SR methods [13, 28, 39] replace the upsampling method with an implicit neural function and boost the performance. These methods predict an RGB value at the query point in an image" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.863, + 0.419, + 0.876 + ], + "angle": 0, + "content": "*Currently with Google. This work was done at ETH Zürich." + }, + { + "type": "page_footnote", + "bbox": [ + 0.078, + 0.876, + 0.47, + 0.901 + ], + "angle": 0, + "content": "†Corresponding Authors: Kai Zhang, cskaizhang@gmail.com; Yulun Zhang, yulun100@gmail.com" + }, + { + "type": "list", + "bbox": [ + 0.078, + 0.863, + 0.47, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.957 + ], + "angle": 0, + "content": "1796" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.134, + 0.09, + 0.835, + 0.21 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.212, + 0.895, + 0.281 + ], + "angle": 0, + "content": "Figure 2. Comparisons of different attention mechanisms. (a) Self-attention can predict pixel features on the grid, but it cannot be directly used in arbitrary-scale SR without considering coordinates. (b) Most existing methods can be treated as coordinate-based implicit attention since they calculate the distance between a key and query coordinate, and then use a function \\( g \\) to aggregate with the value features. However, these methods ignore the distance between the features. (c) Our implicit attention not only considers the coordinate distance, but also the distance among features with visual information." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.29, + 0.473, + 0.501 + ], + "angle": 0, + "content": "by ensembling features within a local region. However, the local ensemble methods have limitations in the ensemble weights and insufficient information (e.g., non-local information). The ensemble weights are often calculated by the area of the rectangle between the query point and each nearest point, which is equivalent to the bilinear interpolation. Thus, those methods cannot adaptively ensemble local features since there is no trainable parameter. These weights are only related to the coordinates of the local features, but independent of the local features. Ignoring both the coordinates and the local features lose visual information and result in blurry artifacts. It is important and necessary to design a new implicit network to predict the weights and exploit more information in the local ensemble." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.502, + 0.474, + 0.638 + ], + "angle": 0, + "content": "In this paper, we propose a novel implicit attention model to enhance arbitrary-scale SR performance. Specifically, we use our attention to predict the ensemble weights by considering both the similarity and coordinate distance of local features, as shown in Figure 2. Based on such learnable weights, the implicit model can adaptively aggregate local features according to different inputs. To enrich more information, we introduce an attention in our implicit attention, which helps discover more features in a larger receptive field." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.639, + 0.4, + 0.653 + ], + "angle": 0, + "content": "Our contributions are summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.661, + 0.47, + 0.737 + ], + "angle": 0, + "content": "- We propose a novel continuous implicit attention-inattention network for arbitrary-scale image SR, called CiaoSR. Different from most existing local ensemble methods, our method explicitly learns the ensemble weights and exploits scale-aware non-local information." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.743, + 0.47, + 0.789 + ], + "angle": 0, + "content": "- Our CiaoSR can be flexibly integrated into any backbone, allowing the network to super-resolve an image at arbitrary scales and improve the SR performance in Figure 1." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.795, + 0.471, + 0.886 + ], + "angle": 0, + "content": "- Extensive experiments demonstrate CiaoSR achieves the state-of-the-art performance in both SISR and arbitrary-scale SR tasks. Besides, our CiaoSR has good generalization on both in-scale and out-of-scale distributions. Last, we extend our method to real-world SR settings to synthesize arbitrary-scale images." + }, + { + "type": "list", + "bbox": [ + 0.078, + 0.661, + 0.471, + 0.886 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.288, + 0.642, + 0.304 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.308, + 0.896, + 0.55 + ], + "angle": 0, + "content": "Single image super-resolution (SISR). SISR aims to synthesize high-resolution (HR) images from low-resolution (LR) images. Compared with DNN-based SR methods [6-8, 23, 27, 50, 66, 67], methods in recent years build on deep convolutional neural network (CNN) to improve the performance, such as SRCNN [19], SRResNet [38], EDSR [44], RDN [88] and RCAN [86]. To further improve SR performance, some methods design CNN with residual block [9, 34, 81], dense block [75, 88, 89] and others [14, 16, 18, 20, 21, 25, 31, 32, 35, 37, 41-43, 59, 65, 70-72, 77, 83, 84, 87, 90]. In addition, some SR methods are built based on attention mechanism [68], such as channel attention [17, 56, 86], self-attention (IPT [11] and SwinIR [40], HAT [12]), non-local attention [45, 48]. However, most methods focus on one specific scale, which limits the applicability and flexibility in arbitrary-scale." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.554, + 0.897, + 0.903 + ], + "angle": 0, + "content": "Arbitrary-scale super-resolution. To tackle this problem, very recently, the more practical setup of arbitrary-scale SR is considered, which aims to super-resolute images with arbitrary scales by a single model. MetaSR [28] makes the first attempt to propose an arbitrary-scale meta-upscale module. To improve the performance, many arbitrary-scale methods [64, 73] are proposed. With the help of implicit neural representation [2, 10, 15, 22, 33, 51, 52, 55, 57, 58, 62, 63], LIIF [13] predicts the RGB value at an arbitrary query coordinate by taking an image coordinate and features of backbone around the coordinate. The features are extracted by single image super-resolution methods, e.g., EDSR [44], RDN [88] and SwinIR [40]. To improve the performance, existing methods [39, 78] propose to integrate more features in SR models. For example, LTE [39] proposes a local texture estimator by characterizing image textures in the Fourier space. UltraSR [78] integrates spatial coordinates and periodic encoding in the implicit network. These methods use the bilinear interpolation to ensemble nearby features. However, such an ensemble way has no learnable parameters. Recently, ITSRN [79] learns the weights by taking the coordinate distance and scale token into a mapping. Most methods learn the ensemble without the feature similarity." + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1797" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.107, + 0.09, + 0.862, + 0.327 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.33, + 0.895, + 0.375 + ], + "angle": 0, + "content": "Figure 3. The architecture of our continuous implicit attention-in-attention network. Given an LR image, the encoder extracts features as latent codes. For a query point, we have a query feature and key features close to the query point, and the scale-aware non-local attention module extracts non-local features as value. Last, we use the triple of query, key and value to predict the RGB value at the query point." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.383, + 0.339, + 0.4 + ], + "angle": 0, + "content": "3. Preliminary and Motivation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.405, + 0.471, + 0.48 + ], + "angle": 0, + "content": "Let \\(I\\) be a continuous image, and \\(\\pmb{x}\\) be a 2D coordinate of a pixel in the image \\(\\pmb{I}\\). Formally, given a 2D coordinate \\(\\pmb{x}\\) in the continuous image domain and latent code \\(\\pmb{Z}\\) extracted by deep neural networks, the RGB value can be predicted by an implicit image function which can be defined as follows," + }, + { + "type": "equation", + "bbox": [ + 0.214, + 0.488, + 0.471, + 0.505 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {I} (\\boldsymbol {x}) = f (\\boldsymbol {Z}, \\boldsymbol {x}), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.508, + 0.47, + 0.628 + ], + "angle": 0, + "content": "where the implicit image function can be parameterized by a multilayer perceptron (MLP). Note that this implicit image function is shared by all images. For recent SISR methods [40, 86], the implicit image function \\( f \\) can be implemented as a PixelShuffle [60] with convolutions with a specific scale. However, these methods are independent on the coordinates, leading to an issue that they only adapt to the specific scale and are inflexible to synthesize arbitrary-scale images." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.629, + 0.473, + 0.689 + ], + "angle": 0, + "content": "To predict the RGB value \\( I_{q} \\) at an arbitrary query coordinate \\( x_{q} \\), most existing methods [13, 39, 79] propose to compute the RGB value at coordinate \\( x_{q} \\) by directly ensembling its neighborhood information," + }, + { + "type": "equation", + "bbox": [ + 0.091, + 0.696, + 0.473, + 0.723 + ], + "angle": 0, + "content": "\\[\n\\pmb {I} _ {q} := \\pmb {I} (\\pmb {x} _ {q}) = \\sum_ {(i, j) \\in \\mathcal {I}} w _ {i, j} \\cdot f (\\pmb {Z} _ {i, j} ^ {*}, \\pmb {x} _ {q} - \\pmb {x} _ {i, j} ^ {*}), (2)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.727, + 0.473, + 0.909 + ], + "angle": 0, + "content": "where \\(\\mathcal{I}\\) is the local region centered at the query coordinate \\(x_{q}\\), e.g., \\(\\mathcal{I}\\) can be top-left, top-right, bottom-left, bottom-right coordinates, and \\(w_{i}\\) is the weight of the neighboring pixel \\(x_{i,j}^{*}\\), which is calculated w.r.t. the area of the rectangle between \\(x_{q}\\) and \\(x_{i,j}^{*}\\), as shown in Figure 4. However, the performance improvement is limited because the ensemble weight \\(w_{i,j}\\) is purely based on the coordinates. The visual similarities are completely ignored in the weight calculation. Besides, these methods only consider the nearest latent codes, leading to a limited receptive field. In this paper, our goal is to learn the weights adaptively by leveraging both visual information and coordinate information." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.383, + 0.673, + 0.4 + ], + "angle": 0, + "content": "4. Proposed Method" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.408, + 0.895, + 0.68 + ], + "angle": 0, + "content": "The above local ensemble Eqn. (2) has a similar form to the attention mechanism. Specifically, the weights \\( w_{i} \\) in Eqn. (2) can be modeled as an attention map and the latter term \\( f \\) can be the value in the attention mechanism. Such an attention map is related to the similarities of the latent code and coordinate, and it can be calculated using both query and key which can integrate the latent code and coordinate information. In this sense, attention models can be used to mitigate the drawbacks of previous methods (i.e., calculated purely based on coordinates without considering any visual information) by learning the soft weights from both visual and coordinate information. However, the use of attention in the implicit function is non-trivial because standard self-attention [68] and neighborhood attention [26] mechanisms are based on visual features, and not conditioned on the coordinate information. To exploit the continuous representation learning from both the coordinate information and visual features, we propose a new attention for arbitrary-scale SR." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.687, + 0.895, + 0.838 + ], + "angle": 0, + "content": "Continuous implicit attention-in-attention. The architecture of the proposed continuous implicit attention-in-attention for arbitrary-scale SR (called CiaoSR) is shown in Figure 3. It is called attention-in-attention because the value contains an additional embedded attention that additionally captures non-local information of repetitive patterns. Specifically, given an LR image, our attention is provided with both the visual features and coordinate information to learn good ensemble weights for the local neighboring latent codes for any query at any arbitrary-scale \\( s \\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Given a query coordinate \\( \\pmb{x}_q \\) in the upsampled image, the known LR pixels in the local region centered at \\( \\pmb{x}_q \\) are considered for the local ensemble. We define \\( \\pmb{x}_k \\) and \\( \\pmb{x}_v \\) as the key and value coordinates of the local neighbor. Let \\( Q, K \\)" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1798" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.471, + 0.168 + ], + "angle": 0, + "content": "and \\(V\\) be the latent code of query, key and value on the corresponding coordinates, respectively. Given pairs of coordinates and latent codes \\((Q, x_{q})\\), \\((K, x_{k})\\) and \\((V, x_{v})\\), we define a new implicit attention function (called i-Attention) to predict RGB values at the given query \\(x_{q}\\)," + }, + { + "type": "equation", + "bbox": [ + 0.139, + 0.177, + 0.47, + 0.195 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {I} _ {q} = \\mathrm {i} - \\text {A t t e n t i o n} (\\boldsymbol {Q}, \\boldsymbol {K}, \\boldsymbol {V}; \\boldsymbol {x} _ {q}, \\boldsymbol {x} _ {k}, \\boldsymbol {x} _ {v}), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.204, + 0.47, + 0.234 + ], + "angle": 0, + "content": "where the value feature \\(V\\) can be embedded with another attention to aggregate more non-local information." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.235, + 0.471, + 0.355 + ], + "angle": 0, + "content": "As shown in Figure 2, our implicit attention is different from existing attention mechanisms by considering the additional coordinate information. In particular, the self-attention mechanism [68] without considering coordinates cannot predict pixel features that are not on the grid. The coordinate-based implicit attention calculates the distance between a key \\( \\boldsymbol{x}_k \\) and query coordinate \\( \\boldsymbol{x}_q \\), which ignores the visual information and similarity of the features." + }, + { + "type": "title", + "bbox": [ + 0.076, + 0.358, + 0.472, + 0.373 + ], + "angle": 0, + "content": "Implicit attention to learn weights for the local ensemble." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.373, + 0.47, + 0.478 + ], + "angle": 0, + "content": "Unlike previous methods [13, 39] which directly use the normalized area as the ensemble weights, we propose to first use an attention to learn better weights for the ensemble of the local region. Our attention takes both the coordinate distance \\( r \\), the local features \\( F_{l} \\), as well as non-local features \\( G \\) into account for the learning of the weight. Formally, the attention can be written as" + }, + { + "type": "equation", + "bbox": [ + 0.141, + 0.485, + 0.47, + 0.526 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {I} _ {q} = \\phi_ {q} \\left(\\sum_ {(i, j) \\in \\mathcal {I}} \\underbrace {\\sigma \\left(\\boldsymbol {Q} ^ {\\top} \\boldsymbol {K} _ {i , j}\\right)} _ {\\text {e n s e m b l e w e i g h t}} \\boldsymbol {V} _ {i, j}\\right), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.535, + 0.47, + 0.58 + ], + "angle": 0, + "content": "where \\(\\phi_q\\) is the query network, \\(\\sigma\\) is the Softmax function, and \\(\\mathcal{I}\\) is the local region centered at \\(\\boldsymbol{x}_q\\). The query, key, and value for this local ensemble attention are defined as" + }, + { + "type": "equation", + "bbox": [ + 0.149, + 0.587, + 0.47, + 0.645 + ], + "angle": 0, + "content": "\\[\n\\left\\{ \\begin{array}{l} Q = \\boldsymbol {F} ^ {*}, \\\\ \\boldsymbol {K} _ {i, j} = \\phi_ {k} \\left(\\left[ \\boldsymbol {F} _ {i, j}, \\left(\\boldsymbol {r} _ {k}\\right) _ {i, j}, \\boldsymbol {s} \\right]\\right), \\\\ \\boldsymbol {V} _ {i, j} = \\phi_ {v} \\left(\\left[ \\boldsymbol {F} _ {i, j}, \\tilde {\\boldsymbol {F}} _ {i, j} \\right], \\left(\\boldsymbol {r} _ {v}\\right) _ {i, j}, \\boldsymbol {s} \\right]), \\end{array} \\right. \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.654, + 0.471, + 0.804 + ], + "angle": 0, + "content": "where \\(\\phi_{k}\\) and \\(\\phi_v\\) are the key and value networks, respectively, \\(F^{*}\\) are the features of nearest neighbors on the grid of the LR image w.r.t. the query coordinate \\(\\pmb{x}_q\\), \\(F_{i,j}\\) is a local feature calculated by using the unfolding operator to the latent codes at the \\((i,j)\\)-th position, which is equivalent to concatenating the latent codes in neighboring \\(3\\times 3\\) region. \\(\\tilde{F}_{i,j}\\) are non-local features, which will be introduced below in detail. \\(s = [s_h;s_w]\\) is a two-dimensional scale, and \\(\\boldsymbol{r}_k\\) and \\(\\boldsymbol{r}_v\\) are the relative distance between the coordinates in the local region, i.e.," + }, + { + "type": "equation", + "bbox": [ + 0.134, + 0.815, + 0.47, + 0.832 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {r} _ {k} = \\boldsymbol {x} _ {q} - \\left(\\boldsymbol {x} _ {k}\\right) _ {i, j}, \\quad \\boldsymbol {r} _ {v} = \\boldsymbol {x} _ {q} - \\left(\\boldsymbol {x} _ {v}\\right) _ {i, j}. \\tag {6}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.076, + 0.841, + 0.471, + 0.855 + ], + "angle": 0, + "content": "Embedded scale-aware attention for non-local features." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.471, + 0.901 + ], + "angle": 0, + "content": "Given a query coordinate, the local ensemble attention only considers the keys and values in the small local region, resulting in a limited receptive field. However, there may exist" + }, + { + "type": "image", + "bbox": [ + 0.523, + 0.089, + 0.685, + 0.214 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.529, + 0.217, + 0.676, + 0.228 + ], + "angle": 0, + "content": "(a) Bilinear-based ensemble" + }, + { + "type": "image", + "bbox": [ + 0.703, + 0.089, + 0.865, + 0.214 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.707, + 0.217, + 0.861, + 0.228 + ], + "angle": 0, + "content": "(b) Attention-based ensemble" + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.23, + 0.894, + 0.327 + ], + "angle": 0, + "content": "Figure 4. Comparison of different ensemble methods. (a) Most existing methods calculate the weights related to the area of the rectangle between the query coordinate and the nearest point. Such weights are equivalent to the weight of the bilinear interpolation. (b) Our attention-based ensemble calculates the attention map as the weights for feature aggregation. Besides, the ensemble method has a large receptive field for exploiting more information." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.338, + 0.895, + 0.596 + ], + "angle": 0, + "content": "similar textures with different scales within a single image in other positions. For example, repetitive patterns in different scales (e.g., facades, windows, etc. in a building, as shown in Figure 4) can exist in different locations within one image. To exploit the useful non-local information, we propose an embedded scale-aware non-local attention module, which aggregates the information from all coordinates on the grid of the LR image, motivated by [49]. Specifically, we obtain query \\(\\tilde{Q}\\) and value features \\(\\tilde{V}\\) from the latent code \\(F\\) of the LR image, and downsample \\(F\\) to a scale \\(s'\\) as a key feature \\(\\tilde{K}\\). We first calculate the pixel-wise similarity between the query located at \\((i,j)\\) and the key on other coordinates, and then aggregate the weighted information from all of them. Lastly, we use a convolutional layer \\(\\varphi\\) for downsampling with the scale \\(s'\\) normalized similarity matrix to aggregate the features. Formally, the non-local feature \\(\\tilde{F}\\) located at \\((i,j)\\) can be calculated by" + }, + { + "type": "equation", + "bbox": [ + 0.503, + 0.604, + 0.892, + 0.656 + ], + "angle": 0, + "content": "\\[\n\\tilde {\\boldsymbol {F}} _ {i, j} = \\varphi \\left(\\sum_ {u, v} \\frac {\\exp \\left(\\tilde {\\boldsymbol {Q}} _ {i , j} ^ {\\top} \\tilde {\\boldsymbol {K}} _ {u , v}\\right)}{\\sum_ {u ^ {\\prime} , v ^ {\\prime}} \\exp \\left(\\tilde {\\boldsymbol {Q}} _ {i , j} ^ {\\top} \\tilde {\\boldsymbol {K}} _ {u ^ {\\prime} , v ^ {\\prime}}\\right)} \\tilde {\\boldsymbol {V}} _ {s ^ {\\prime} u, s ^ {\\prime} v} ^ {s ^ {\\prime} p \\times s ^ {\\prime} p}\\right), \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.666, + 0.892, + 0.717 + ], + "angle": 0, + "content": "where \\(\\tilde{V}_{s'u,s'v}^{s'p\\times s'p}\\) is the value feature patch of the size \\(s^{\\prime}p\\times s^{\\prime}p\\) located at \\((s^{\\prime}u,s^{\\prime}v)\\), \\(\\tilde{Q},\\tilde{K}\\) and \\(\\tilde{V}\\) are the query, key and value's non-local features, which can be calculated by" + }, + { + "type": "equation", + "bbox": [ + 0.631, + 0.726, + 0.892, + 0.786 + ], + "angle": 0, + "content": "\\[\n\\left\\{ \\begin{array}{l} \\tilde {\\boldsymbol {Q}} = \\varphi_ {q} (\\boldsymbol {F}), \\\\ \\tilde {\\boldsymbol {K}} = \\varphi_ {k} (\\boldsymbol {F} \\downarrow_ {s ^ {\\prime}}), \\\\ \\tilde {\\boldsymbol {V}} = \\varphi_ {v} (\\boldsymbol {F}), \\end{array} \\right. \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.795, + 0.895, + 0.901 + ], + "angle": 0, + "content": "where \\(\\downarrow_{s'}\\) is the downsampling operation with a given scale \\(s'\\) which can be \\(\\{2,3,4\\}\\), \\(\\varphi_q\\), \\(\\varphi_k\\) and \\(\\varphi_v\\) are query, key and value networks. Different from [49], our method can integrate multi-scale features. In the experiment, we can search for relevant features in a small window (e.g., \\(256\\times 256\\)), instead of the entire image. In this way, our method can guarantee a large receptive field, and reduce computational cost." + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1799" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.089, + 0.894, + 0.118 + ], + "angle": 0, + "content": "Table 1. Quantitative comparison with state-of-the-art methods for arbitrary-scale SR on the DIV2K validation set (PSNR (dB)). Bold indicates the best performance. \\(\\dagger\\) indicates the implementation of [39]." + }, + { + "type": "table", + "bbox": [ + 0.081, + 0.12, + 0.89, + 0.395 + ], + "angle": 0, + "content": "
BackbonesMethodsIn-scaleOut-of-scale
×2×3×4×6×12×18×24×30
-Bicubic31.0128.2226.6624.8222.2721.0020.1919.59
EDSR [44]EDSR-baseline [44]34.5530.9028.94-----
EDSR-baseline-MetaSR [28]34.6430.9328.9226.6123.5522.0321.0620.37
EDSR-baseline-LIIF [13]34.6730.9629.0026.7523.7122.1721.1820.48
EDSR-baseline-ITSRN† [79]34.7130.9529.0326.7723.7122.1721.1820.49
EDSR-baseline-LTE [39]34.7231.0229.0426.8123.7822.2321.2420.53
EDSR-baseline-CiaoSR (ours)34.9131.1529.2326.9523.8822.3221.3220.59
RDN [88]RDN-baseline [88]34.9431.2229.19-----
RDN-MetaSR [28]35.0031.2729.2526.8823.7322.1821.1720.47
RDN-LIIF [13]34.9931.2629.2726.9923.8922.3421.3120.59
RDN-ITSRN† [79]35.0931.3629.3827.0623.9322.3621.3220.61
RDN-LTE [39]35.0431.3229.3327.0423.9522.4021.3620.64
RDN-CiaoSR (ours)35.1531.4229.4527.1624.0622.4821.4320.70
SwinIR [40]SwinIR-baseline [40]34.9431.2229.19-----
SwinIR-MetaSR† [28]35.1531.4029.3326.9423.8022.2621.2620.54
SwinIR-LIIF† [13]35.1731.4629.4627.1524.0222.4321.4020.67
SwinIR-ITSRN† [79]35.1931.4229.4827.1323.8322.3121.3120.55
SwinIR-LTE [39]35.2431.5029.5127.2024.0922.5021.4720.73
SwinIR-CiaoSR (ours)35.2931.5529.5927.2824.1522.5421.5120.74
" + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.401, + 0.21, + 0.418 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.423, + 0.471, + 0.515 + ], + "angle": 0, + "content": "Datasets. For the arbitrary-scale SR task, we follow [13, 39] and use DIV2K [1] as the training set, which consists of 800 images in 2K resolution. In testing, we evaluate the models on the DIV2K validation set and a wide range of standard benchmark datasets, including Set5 [3], Set14 [80], B100 [46], Urban100 [29], and Manga109 [47]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.519, + 0.47, + 0.641 + ], + "angle": 0, + "content": "Evaluation metrics. We use PSNR to evaluate the quality of the synthesized SR images. Following [13, 39], the PSNR value is calculated on all RGBs channels for the DIV2K validation set, and additionally also calculated on the Y channel (i.e., luminance) of the transformed YCbCr space for other benchmark test sets. In addition, other evaluation metrics (such as SSIM [76] and LPIPS [85]) are provided to evaluate the image quality." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.645, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Implementation details. We follow the same experimental setup as prior works [13, 39]. To synthesize paired training data, we first crop the \\(48s \\times 48s\\) patches as ground-truth (GT) images, and use the Bicubic downsampling in PyTorch to have LR images with the patch size of \\(48 \\times 48\\), where \\(s\\) is a scale factor sampled in the uniform distribution \\(\\mathcal{U}(1,4)\\). We sample \\(48^2\\) pixels from both GT images and the corresponding coordinates. We use the existing SR models (e.g., EDSR [44], RDN [88] and SwinIR [40]) as backbones by removing their upsampling modules. The detailed architecture of our implicit network is provided in the supplementary material. For the training, we use Adam [36] as the optimizer, and use L1 loss to train all models for 1000 epochs with the batch size of 16 per GPU. We set the learning rate as \\(1e - 4\\) at the beginning and decay the learning rate by a factor of 0.5 every 200 epochs. These experimental settings are identical to [13, 28, 39]." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.401, + 0.805, + 0.418 + ], + "angle": 0, + "content": "5.1. Comparisons with State-of-the-Art" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.425, + 0.895, + 0.743 + ], + "angle": 0, + "content": "Quantitative results. In Tables 1-3, CiaoSR achieves the best performance with the highest PSNR across all datasets with all backbones concerned on both in-scale and out-of-scale distributions. For the in-scales, in particular, our model has a remarkable PSNR gain of \\(0.3\\mathrm{dB}\\) on Urban100 \\((\\times 4)\\), compared with the previous SOTA method [39] under the same RDN backbone. It is worth noting that CiaoSR with the RDN backbone [88] can surpass the performance of the better backbone SwinIR [40] (with its original upsampling module). With the help of the continuous-scale training, a single SR model trained with our method has better generalization performance than vanilla SR models (including EDSR [44], RDN [88] and SwinIR [40] trained on specific scale. For the out-of-scale experiments, our model also achieves the best generalization performance on unseen scales. The vanilla EDSR-baseline [44], RDN [88] and SwinIR [40] cannot be applied on the out-of-scale experiments because their decoder can only predict the scale that it was trained on. We observe that our method yields the largest improvement on Urban100 over existing arbitrary-scale methods (e.g., LIIF [13] and LTE [39])." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Qualitative results. In Figure 5, we provide qualitative comparisons with other arbitrary-scale SR methods. Our model is able to synthesize the SR images with sharper textures than other methods. Taking the second line as an example, the textures in the LR image are degraded, but CiaoSR is still able to restore the textures of the building. In contrast, other methods can only restore part of the textures possibly due to the limitations of the local ensemble and insufficient features. More visual results on more test sets are put in the supplementary materials." + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1800" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.089, + 0.892, + 0.117 + ], + "angle": 0, + "content": "Table 2. Quantitative comparison with state-of-the-art methods for in-scale SR on benchmark datasets (PSNR (dB)). Bold indicates the best performance. † indicates our implementation." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.12, + 0.89, + 0.292 + ], + "angle": 0, + "content": "
MethodsSet5 [3]Set14 [80]B100 [46]Urban100 [29]Manga109 [47]
×2×3×4×2×3×4×2×3×4×2×3×4×2×3×4
RDN [88]38.2434.7132.4734.0130.5728.8132.3429.2627.7232.8928.8026.6139.1834.1331.00
RDN-MetaSR [28]38.2234.6332.3833.9830.5428.7832.3329.2627.7132.9228.8226.55---
RDN-LIIF [13]38.1734.6832.5033.9730.5328.8032.3229.2627.7432.8728.8226.6839.2634.2131.20
RDN-ITSRN† [79]38.2334.7632.5534.1930.5928.8832.3829.3227.7933.0728.9626.7739.3434.3931.37
RDN-LTE [39]38.2334.7232.6134.0930.5828.8832.3629.3027.7733.0428.9726.8139.2834.3231.30
RDN-CiaoSR (ours)38.2934.8532.6634.2230.6528.9332.4129.3427.8333.3029.1727.1139.5134.5731.57
SwinIR [40]38.3534.8932.7234.1430.7728.9432.4429.3727.8333.4029.2927.0739.6034.7431.67
SwinIR-MetaSR [28]38.2634.7732.4734.1430.6628.8532.3929.3127.7533.2929.1226.7639.4634.6231.37
SwinIR-LIIF [13]38.2834.8732.7334.1430.7528.9832.3929.3427.8433.3629.3327.1539.5734.6831.71
SwinIR-ITSRN† [79]38.2234.7532.6334.2630.7528.9732.4229.3827.8533.4629.3427.1239.6034.7531.74
SwinIR-LTE [39]38.3334.8932.8134.2530.8029.0632.4429.3927.8633.5029.4127.2439.6334.7931.79
SwinIR-CiaoSR (ours)38.3834.9132.8434.3330.8229.0832.4729.4227.9033.6529.5227.4239.6734.8431.91
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.301, + 0.892, + 0.328 + ], + "angle": 0, + "content": "Table 3. Quantitative comparison with state-of-the-art methods for out-of-scale SR on benchmark datasets (PSNR (dB)). Bold indicates the best performance. † indicates our implementation." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.331, + 0.89, + 0.48 + ], + "angle": 0, + "content": "
MethodsSet5 [3]Set14 [80]B100 [46]Urban100 [29]Manga109 [47]
×6×8×12×6×8×12×6×8×12×6×8×12×6×8×12
RDN-MetaSR [28]29.0429.96-26.5124.97-25.9024.83-23.9922.59----
RDN-LIIF [13]29.1527.1424.8626.6425.1523.2425.9824.9123.5724.2022.7921.1527.3325.0422.36
RDN-ITSRN† [79]29.3227.2524.8626.6825.1723.2826.0124.9323.5824.2322.8121.1627.4525.0423.35
RDN-LTE [39]29.3227.2624.7926.7125.1623.3126.0124.9523.6024.2822.8821.2227.4925.1222.43
RDN-CiaoSR (ours)29.4627.3624.9226.7925.2823.3726.0725.0023.6424.5823.1321.4227.7025.4022.63
SwinIR-MetaSR [28]29.0927.0224.8226.5825.0923.3325.9424.8723.5924.1622.7521.3127.2924.9622.35
SwinIR-LIIF [13]29.4627.36-26.8225.34-26.0725.01-24.5923.14-27.6925.28-
SwinIR-ITSRN† [79]29.3127.2424.7926.7125.3223.3026.0524.9623.5724.5023.0621.3427.7225.2322.47
SwinIR-LTE [39]29.5027.35-26.8625.42-26.0925.03-24.6223.17-27.8325.42-
SwinIR-CiaoSR (ours)29.6227.4524.9626.8825.4223.3826.1325.0723.6824.8423.3421.6028.0125.6122.79
" + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.504, + 0.23, + 0.52 + ], + "angle": 0, + "content": "5.2. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.531, + 0.471, + 0.607 + ], + "angle": 0, + "content": "In this section, we conduct ablation studies to investigate effect of each component in our architecture. Our CiaoSR consists of the attention-in-attention network and a scale-aware non-local attention module. Table 4 shows the contributions of each component on performance." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.611, + 0.47, + 0.79 + ], + "angle": 0, + "content": "Attention-in-attention. The attention-in-attention network is the main branch of our architecture. On the one hand, the network learns to ensemble weights with both the coordinates and the features. On the other hand, the network is able to aggregate the local and non-local information to improve the SR performance. To investigate how the attention mechanism affects the performance, we remove this module and replace it with an MLP to predict the weights of the ensemble. The coordinate information is still fed into the MLP to enable arbitrary scale prediction. As shown in Table 4, the method suffers a significant performance drop without the attention mechanism." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.795, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Non-local attention. The embedded non-local attention is proposed to capture the scale-aware information from other locations of the images to improve the performance. We investigate the importance of the non-local attention in an implicit model by comparing the results with and without this module. As shown in Table 4, by introducing the nonlocal attention, a further performance gain is achieved. This" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.505, + 0.892, + 0.535 + ], + "angle": 0, + "content": "is because our module can exploit texture information from other scales in a large perceptual field." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.539, + 0.895, + 0.901 + ], + "angle": 0, + "content": "Training with different types of scales. We evaluate the effectiveness of our implicit model when trained with discrete (including single scale \\(\\{\\{2\\} /\\{3\\} /\\{4\\}\\}\\) and multiple scales \\(\\{2,3,4\\}\\)) and continuous scales [1, 4], and show the results in Table 5. For the discrete type of scales, we observe that training our implicit model with a specific scale can also achieve significant improvement for that specific scale, compared with the original upsampling module in RDN [88]. Taking \\(\\times 2\\) scale as an example, our method has an improvement of \\(0.24\\mathrm{dB}\\) over RDN. However, training with a specific scale has poor generalization performance on unseen in-scale and out-of-scale. A few existing works [34, 44] consider the training with multiple discrete scales and found the performance of the joint learned model is comparable with single scale networks. Based on our implicit attention module, we are able to deal with more scales and exploit the correlation of multiple SR tasks at different scales. Unlike previous works [34, 44] which only achieve comparable performance compared with single scale models, CiaoSR achieves a better performance than training with one scale. We found our model has further improvement when trained on continuous scales. The improvement gain is attributed to our continuous representations design and our architecture, which enables efficient learning from cross-scale information." + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.946, + 0.514, + 0.957 + ], + "angle": 0, + "content": "1801" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.086, + 0.089, + 0.327, + 0.232 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.138, + 0.234, + 0.271, + 0.246 + ], + "angle": 0, + "content": "004, \\(\\times 4\\) Urban100 [29]" + }, + { + "type": "image", + "bbox": [ + 0.331, + 0.089, + 0.465, + 0.153 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.375, + 0.156, + 0.418, + 0.166 + ], + "angle": 0, + "content": "Bicubic" + }, + { + "type": "image", + "bbox": [ + 0.468, + 0.089, + 0.604, + 0.153 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.508, + 0.155, + 0.564, + 0.167 + ], + "angle": 0, + "content": "RDN [88]" + }, + { + "type": "image", + "bbox": [ + 0.606, + 0.09, + 0.743, + 0.153 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.622, + 0.155, + 0.726, + 0.167 + ], + "angle": 0, + "content": "RDN-MetaSR [28]" + }, + { + "type": "image", + "bbox": [ + 0.745, + 0.09, + 0.881, + 0.153 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.769, + 0.155, + 0.856, + 0.167 + ], + "angle": 0, + "content": "RDN-LIIF [13]" + }, + { + "type": "image", + "bbox": [ + 0.086, + 0.247, + 0.326, + 0.392 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.104, + 0.394, + 0.306, + 0.406 + ], + "angle": 0, + "content": "TaiyouNiSmash, \\(\\times 8\\) ,Manga109 [47]" + }, + { + "type": "image", + "bbox": [ + 0.33, + 0.168, + 0.466, + 0.232 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.348, + 0.234, + 0.447, + 0.246 + ], + "angle": 0, + "content": "RDN-ITSRN [79]" + }, + { + "type": "image", + "bbox": [ + 0.468, + 0.168, + 0.604, + 0.232 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.494, + 0.234, + 0.577, + 0.246 + ], + "angle": 0, + "content": "RDN-LTE [39]" + }, + { + "type": "image", + "bbox": [ + 0.606, + 0.168, + 0.743, + 0.232 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.616, + 0.234, + 0.733, + 0.246 + ], + "angle": 0, + "content": "RDN-CiaoSR (ours)" + }, + { + "type": "image", + "bbox": [ + 0.744, + 0.168, + 0.881, + 0.232 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.802, + 0.234, + 0.823, + 0.245 + ], + "angle": 0, + "content": "GT" + }, + { + "type": "image", + "bbox": [ + 0.33, + 0.247, + 0.466, + 0.311 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.374, + 0.313, + 0.418, + 0.324 + ], + "angle": 0, + "content": "Bicubic" + }, + { + "type": "image", + "bbox": [ + 0.468, + 0.247, + 0.604, + 0.311 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.505, + 0.313, + 0.566, + 0.325 + ], + "angle": 0, + "content": "\\(\\mathbf{RDN}^*\\) [88]" + }, + { + "type": "image", + "bbox": [ + 0.606, + 0.247, + 0.743, + 0.311 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.622, + 0.313, + 0.726, + 0.325 + ], + "angle": 0, + "content": "RDN-MetaSR [28]" + }, + { + "type": "image", + "bbox": [ + 0.744, + 0.247, + 0.881, + 0.311 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.769, + 0.313, + 0.855, + 0.325 + ], + "angle": 0, + "content": "RDN-LIIF [13]" + }, + { + "type": "image", + "bbox": [ + 0.33, + 0.326, + 0.466, + 0.39 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.348, + 0.392, + 0.447, + 0.404 + ], + "angle": 0, + "content": "RDN-ITSRN [79]" + }, + { + "type": "image", + "bbox": [ + 0.468, + 0.326, + 0.604, + 0.39 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.51, + 0.392, + 0.561, + 0.404 + ], + "angle": 0, + "content": "LTE [39]" + }, + { + "type": "image", + "bbox": [ + 0.606, + 0.326, + 0.743, + 0.39 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.616, + 0.392, + 0.733, + 0.404 + ], + "angle": 0, + "content": "RDN-CiaoSR (ours)" + }, + { + "type": "image", + "bbox": [ + 0.744, + 0.326, + 0.881, + 0.39 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.802, + 0.392, + 0.823, + 0.403 + ], + "angle": 0, + "content": "GT" + }, + { + "type": "image_caption", + "bbox": [ + 0.115, + 0.413, + 0.853, + 0.427 + ], + "angle": 0, + "content": "Figure 5. Visual comparison of different methods on benchmarks. \\*\\* means the model first synthesizes twice to \\(\\times 8\\) images." + }, + { + "type": "image", + "bbox": [ + 0.08, + 0.434, + 0.237, + 0.515 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.08, + 0.516, + 0.237, + 0.577 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.149, + 0.578, + 0.166, + 0.587 + ], + "angle": 0, + "content": "LR" + }, + { + "type": "image", + "bbox": [ + 0.244, + 0.434, + 0.399, + 0.515 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.244, + 0.516, + 0.4, + 0.578 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.299, + 0.579, + 0.341, + 0.588 + ], + "angle": 0, + "content": "RealSR" + }, + { + "type": "image", + "bbox": [ + 0.407, + 0.434, + 0.563, + 0.515 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.407, + 0.516, + 0.563, + 0.577 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.459, + 0.579, + 0.51, + 0.588 + ], + "angle": 0, + "content": "BSRGAN" + }, + { + "type": "image", + "bbox": [ + 0.57, + 0.434, + 0.726, + 0.515 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.57, + 0.516, + 0.726, + 0.577 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.607, + 0.579, + 0.685, + 0.588 + ], + "angle": 0, + "content": "Real-ESRGAN" + }, + { + "type": "image", + "bbox": [ + 0.734, + 0.434, + 0.862, + 0.515 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.735, + 0.516, + 0.862, + 0.577 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.797, + 0.579, + 0.823, + 0.588 + ], + "angle": 0, + "content": "Ours" + }, + { + "type": "image_caption", + "bbox": [ + 0.181, + 0.595, + 0.787, + 0.609 + ], + "angle": 0, + "content": "Figure 6. Visual comparison of different methods on the RealSRSet [82] and DPED [30] dataset \\((\\times 8)\\)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.617, + 0.224, + 0.633 + ], + "angle": 0, + "content": "5.3. Further Study" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.639, + 0.471, + 0.836 + ], + "angle": 0, + "content": "Effect on synthesis steps. Our trained model is able to synthesize SR images with any given arbitrary scale in one step. A natural alternative would be to synthesize a large scale in multiple steps. Taking a large scale \\((e.g., \\times 12)\\) as an example, SR images can be generated by two steps \\((e.g., \\rightarrow \\times 2 \\rightarrow \\times 12)\\), or three steps \\((e.g., \\rightarrow \\times 2 \\rightarrow \\times 4 \\rightarrow \\times 12)\\). We compare the performance of synthesizing a large scale in one and multiple steps, and the results on the Urban100 and Manga109 datasets are shown in Table 6. We observe that our way of synthesizing SR images in one step has the best performance under the PSNR and SSIM metrics. In contrast, the performance gets worse with more steps because errors can accumulate along the multiple synthesizing steps." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Model size and inference time. To demonstrate the efficiency of our proposed method, we investigate the model size and inference time of different arbitrary-scale SR methods. Results are shown in Table 7. Specifically, we set" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.618, + 0.895, + 0.709 + ], + "angle": 0, + "content": "the input size as \\(192 \\times 192\\) and calculate the inference time on an NVIDIA Quadro RTX 6000. Our implicit model has much fewer parameters than the previous SOTA methods (e.g., LTE [39]). Our model has the best performance although it requires a little more inference time due to exploiting the non-local features." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.895, + 0.902 + ], + "angle": 0, + "content": "More evaluation metrics. To further demonstrate the effectiveness of our proposed method, we apply more evaluation metrics (e.g., SSIM [76] and LPIPS [85]) to evaluate the image quality across different methods on the Urban100 dataset. These evaluation metrics are widely used in the SR community. In general, higher SSIM and lower LPIPS correspond to better performance. As shown in Table 8, our method achieves the highest SSIM and lowest LPIPS, and thus has the best performance on both in-scale and out-of-scale distributions. Compared with other methods, our proposed CiaoSR is able to maintain more structural texture and perceptual information." + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1802" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.09, + 0.446, + 0.117 + ], + "angle": 0, + "content": "Table 4. Ablation study on each component of our networks on Urban100. We use RDN [88] as the backbone." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.12, + 0.443, + 0.212 + ], + "angle": 0, + "content": "
Attention-in-attention✗✓✓
Scale-aware Attention Network✗✗✓
In-scale×232.8733.2433.30
×328.8229.1029.17
×426.6926.9627.11
Out-of-scale×624.2224.5024.58
×822.8022.9823.13
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.216, + 0.462, + 0.244 + ], + "angle": 0, + "content": "Table 6. Comparison of (PSNR and SSIM) for different synthesis steps on Urban100 [29] and Manga109 [47]." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.247, + 0.458, + 0.309 + ], + "angle": 0, + "content": "
TypeSynthesis stepsUrban100 [29]Manga109 [47]
PSNRSSIMPSNRSSIM
Multiple steps→×2→×4→×1221.280.55722.460.720
→×2→×1221.320.55822.530.721
One step→×1221.420.56122.630.723
" + }, + { + "type": "table_caption", + "bbox": [ + 0.451, + 0.091, + 0.902, + 0.117 + ], + "angle": 0, + "content": "Table 5. Ablation study on training our implicit model with different types of scales on Urban100." + }, + { + "type": "table", + "bbox": [ + 0.456, + 0.121, + 0.898, + 0.212 + ], + "angle": 0, + "content": "
TypeTraining scale sIn-scaleOut-of-scale
×2×3×4×6×8
Discretes ∈ {2}33.1327.0125.6022.2722.09
s ∈ {3}31.3929.0625.7723.4422.16
s ∈ {4}31.4227.8726.8824.2822.85
s ∈ {2,3,4}33.1529.1427.0224.4723.03
Continuouss ∈ [1,4]33.3029.1727.1124.5823.13
" + }, + { + "type": "table_caption", + "bbox": [ + 0.466, + 0.216, + 0.901, + 0.242 + ], + "angle": 0, + "content": "Table 7. Comparisons of model size, inference time and performance gain of different models." + }, + { + "type": "table", + "bbox": [ + 0.47, + 0.246, + 0.898, + 0.309 + ], + "angle": 0, + "content": "
Different modelsMeta-SR [28]LIIF [13]ITSRN [79]LTE [39]CiaoSR
Model size (M)1.71.60.71.71.4
Inference time (ms)237171343148528
PSNR (dB)26.5526.6826.7726.8127.11
Performance gain (dB)-0.060.070.160.20.5
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.314, + 0.586, + 0.34 + ], + "angle": 0, + "content": "Table 8. Comparison of more evaluation metrics (SSIM↑ [76] and LPIPS↓ [85]) for different methods on Urban100." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.345, + 0.581, + 0.451 + ], + "angle": 0, + "content": "
RDN-MethodIn-scaleOut-of-scale
×2×3×4×6×8
SSIMLPIPSSSIMLPIPSSSIMLPIPSSSIMLPIPSSSIMLPIPS
LIIF [13]0.9340.1040.8660.1970.8040.2670.7040.3600.6360.425
MetaSR [28]0.9350.1040.8660.1950.8050.2610.6980.3530.6240.423
ITSRN [79]0.9360.1020.8680.1930.8070.2590.7030.3500.6330.417
LTE [39]0.9360.1040.8680.1950.8070.2640.7070.3560.6390.421
CiaoSR0.9380.1000.8710.1880.8140.2550.7180.3460.6510.408
" + }, + { + "type": "table_caption", + "bbox": [ + 0.589, + 0.315, + 0.902, + 0.341 + ], + "angle": 0, + "content": "Table 9. Quantitative comparisons of different Realworld SR methods on RealSRSet [82]." + }, + { + "type": "table", + "bbox": [ + 0.593, + 0.345, + 0.898, + 0.451 + ], + "angle": 0, + "content": "
MethodsScaleNIQE↓BRISQUE↓PIQE↓
RealSR [5]×84.794923.322829.3976
×163.394621.639021.4116
BSRGAN [82]×84.700737.063836.1889
×164.140839.625237.9176
Real-ESRGAN [74]×84.528040.652159.8915
×165.644253.936484.3611
Ours-real×83.589430.921929.8554
×163.946141.814558.8244
" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.473, + 0.409, + 0.489 + ], + "angle": 0, + "content": "5.4. Real-World Arbitrary-Scale Image SR" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.492, + 0.47, + 0.718 + ], + "angle": 0, + "content": "We extend our proposed method to real-world applications with the ultimate goal of synthesizing real images with arbitrary scales. Specifically, we train our model on the DIV2K dataset [1], and test on the RealSRSet [82] and DPED [30] datasets. We apply the practical degradation models (including blur, noise, compression, etc.) of BSRGAN [82] and Real-ESRGAN [74] to synthesize paired training data on DIV2K [1]. We set the HR patch size of training data as 256, and the batch size as 48. We use Adam [36] as the optimizer, and adopt the exponential moving average (EMA). We first train our model with L1 loss for 1000K iterations with the learning rate of \\(2 \\times 10^{-4}\\) and train with L1 loss, perceptual loss and GAN loss together for 400K iterations with the learning rate of \\(1 \\times 10^{-4}\\). More training details can be found in the supplementary materials." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.72, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Since there are no arbitrary-scale SR methods applied to real-world images, we mainly compare with the state-of-the-art real-world image SR models, including RealSR [5], BSRGAN [82] and Real-ESRGAN [74]. Because there are no ground-truth (GT) real images, we apply non-reference image quality assessment (e.g., NIQE [54], BRISQUE [53] and PIQE [69]) to evaluate the image quality. As shown in Table 9, our model does not optimize for all metrics since these metrics cannot well match the actual human perceptual [4]. To address this, we further compare the visual results on the RealSRSet [82] and DPED [30] datasets in Figure 6. Our proposed method is able to synthesize real" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.474, + 0.892, + 0.595 + ], + "angle": 0, + "content": "SR images with realistic textures. In particular, our method is able to synthesize more details of a dog in the first line and more clear digit '5' in the second line, compared with other methods. In contrast, other methods often introduce unnatural artifacts within images, especially for out-of-scale distributions. To further investigate the effectiveness of our model, we provide more visual results of our real model in the supplementary materials." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.608, + 0.619, + 0.623 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.626, + 0.895, + 0.852 + ], + "angle": 0, + "content": "In this paper, we have proposed a novel continuous implicit attention-in-attention network for arbitrary-scale image super-resolution, called CiaoSR. Specifically, our CiaoSR first introduces a scale-aware non-local attention in our attention to exploit more relevant features, and then predicts RGB values by locally ensembling features with the learnable attention weights. More importantly, CiaoSR has good flexibility and applicability since it can be used behind any SR backbone to boost the performance. Extensive experiments on all benchmark datasets demonstrate the superiority of our CiaoSR in both SISR and arbitrary-scale SR tasks. Besides, our CiaoSR has good generalization performance on both in-scale and out-of-scale distributions. Last, we extend our method to the real-world application to synthesize real SR images with arbitrary scale." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.856, + 0.892, + 0.9 + ], + "angle": 0, + "content": "Acknowledgements: This work was partly supported by the Huawei Fund, the ETH Zürich General Fund (OK) and the Alexander von Humboldt Foundation." + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1803" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.175, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.171 + ], + "angle": 0, + "content": "[1] Eirikur Agustsson and Radu Timofte. Ntire 2017 challenge on single image super-resolution: Dataset and study. In IEEE Conference on Computer Vision and Pattern Recognition Workshops, pages 126-135, 2017. 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.172, + 0.472, + 0.225 + ], + "angle": 0, + "content": "[2] Matan Atzmon and Yaron Lipman. Sal: Sign agnostic learning of shapes from raw data. In IEEE Conference on Computer Vision and Pattern Recognition, pages 2565-2574, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.228, + 0.472, + 0.296 + ], + "angle": 0, + "content": "[3] Marco Bevilacqua, Aline Roumy, Christine Guillemot, and Marie line Alberi Morel. Low-complexity single-image super-resolution based on nonnegative neighbor embedding. In British Machine Vision Conference, pages 135.1-135.10, 2012. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.298, + 0.472, + 0.354 + ], + "angle": 0, + "content": "[4] Yochai Blau, Roey Mechrez, Radu Timofte, Tomer Michaeli, and Lihi Zelnik-Manor. The 2018 pirm challenge on perceptual image super-resolution. In European Conference on Computer Vision Workshops, 2018. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.355, + 0.472, + 0.422 + ], + "angle": 0, + "content": "[5] Jianrui Cai, Hui Zeng, Hongwei Yong, Zisheng Cao, and Lei Zhang. Toward real-world single image super-resolution: A new benchmark and a new model. In IEEE Conference on International Conference on Computer Vision, pages 3086-3095, 2019. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.424, + 0.472, + 0.465 + ], + "angle": 0, + "content": "[6] Jiezhang Cao, Yawei Li, Kai Zhang, and Luc Van Gool. Video super-resolution transformer. arXiv preprint arXiv:2106.06847, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.467, + 0.472, + 0.535 + ], + "angle": 0, + "content": "[7] Jiezhang Cao, Jingyun Liang, Kai Zhang, Yawei Li, Yulun Zhang, Wenguan Wang, and Luc Van Gool. Reference-based image super-resolution with deformable attention transformer. In European Conference on Computer Vision, pages 325-342, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.537, + 0.472, + 0.606 + ], + "angle": 0, + "content": "[8] Jiezhang Cao, Jingyun Liang, Kai Zhang, Wenguan Wang, Qin Wang, Yulun Zhang, Hao Tang, and Luc Van Gool. Towards interpretable video super-resolution via alternating optimization. In European Conference on Computer Vision, pages 393-411, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.607, + 0.472, + 0.662 + ], + "angle": 0, + "content": "[9] Lukas Cavigelli, Pascal Hager, and Luca Benini. Cas-cnn: A deep convolutional neural network for image compression artifact suppression. In 2017 International Joint Conference on Neural Networks, pages 752-759, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.663, + 0.472, + 0.732 + ], + "angle": 0, + "content": "[10] Rohan Chabra, Jan E Lenssen, Eddy Ig, Tanner Schmidt, Julian Straub, Steven Lovegrove, and Richard Newcombe. Deep local shapes: Learning local sdf priors for detailed 3d reconstruction. In European Conference on Computer Vision, pages 608-625, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.733, + 0.472, + 0.802 + ], + "angle": 0, + "content": "[11] Hanting Chen, Yunhe Wang, Tianyu Guo, Chang Xu, Yiping Deng, Zhenhua Liu, Siwei Ma, Chunjing Xu, Chao Xu, and Wen Gao. Pre-trained image processing transformer. In IEEE Conference on Computer Vision and Pattern Recognition, pages 12299-12310, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.803, + 0.472, + 0.845 + ], + "angle": 0, + "content": "[12] Xiangyu Chen, Xintao Wang, Jiantao Zhou, and Chao Dong. Activating more pixels in image super-resolution transformer. arXiv preprint arXiv:2205.04437, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.846, + 0.472, + 0.901 + ], + "angle": 0, + "content": "[13] Yinbo Chen, Sifei Liu, and Xiaolong Wang. Learning continuous image representation with local implicit image function. In IEEE Conference on Computer Vision and Pattern Recognition, pages 8628-8638, 2021. 1, 2, 3, 4, 5, 6, 7, 8" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.148 + ], + "angle": 0, + "content": "[14] Yunjin Chen and Thomas Pock. Trainable nonlinear reaction diffusion: A flexible framework for fast and effective image restoration. IEEE Transactions on Pattern Analysis and Machine Intelligence, 39(6):1256-1272, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.149, + 0.894, + 0.191 + ], + "angle": 0, + "content": "[15] Zhiqin Chen and Hao Zhang. Learning implicit fields for generative shape modeling. In IEEE Conference on Computer Vision and Pattern Recognition, pages 5939-5948, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.192, + 0.894, + 0.247 + ], + "angle": 0, + "content": "[16] Wenlong Cheng, Mingbo Zhao, Zhiling Ye, and Shuhang Gu. Mfagan: A compression framework for memory-efficient on-device super-resolution gan. arXiv preprint arXiv:2107.12679, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.249, + 0.893, + 0.305 + ], + "angle": 0, + "content": "[17] Tao Dai, Jianrui Cai, Yongbing Zhang, Shu-Tao Xia, and Lei Zhang. Second-order attention network for single image super-resolution. In IEEE Conference on Computer Vision and Pattern Recognition, pages 11065-11074, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.306, + 0.893, + 0.361 + ], + "angle": 0, + "content": "[18] Xin Deng, Yutong Zhang, Mai Xu, Shuhang Gu, and Yiping Duan. Deep coupled feedback network for joint exposure fusion and image super-resolution. IEEE Transactions on Image Processing, 30:3098-3112, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.362, + 0.894, + 0.418 + ], + "angle": 0, + "content": "[19] Chao Dong, Chen Change Loy, Kaiming He, and Xiaou Tang. Learning a deep convolutional network for image super-resolution. In European Conference on Computer Vision, pages 184–199, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.42, + 0.894, + 0.475 + ], + "angle": 0, + "content": "[20] Xueyang Fu, Menglu Wang, Xiangyong Cao, Xinghao Ding, and Zheng-Jun Zha. A model-driven deep unfolding method for jpeg artifacts removal. IEEE Transactions on Neural Networks and Learning Systems, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.476, + 0.894, + 0.532 + ], + "angle": 0, + "content": "[21] Xueyang Fu, Zheng-Jun Zha, Feng Wu, Xinghao Ding, and John Paisley. Jpeg artifacts reduction via deep convolutional sparse coding. In IEEE International Conference on Computer Vision, pages 2501-2510, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.533, + 0.893, + 0.575 + ], + "angle": 0, + "content": "[22] Amos Gropp, Lior Yariv, Niv Haim, Matan Atzmon, and Yaron Lipman. Implicit geometric regularization for learning shapes. arXiv preprint arXiv:2002.10099, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.576, + 0.894, + 0.63 + ], + "angle": 0, + "content": "[23] Shuhang Gu, Nong Sang, and Fan Ma. Fast image super resolution via local regression. In IEEE Conference on International Conference on Pattern Recognition, pages 3128-3131, 2012. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.632, + 0.893, + 0.688 + ], + "angle": 0, + "content": "[24] Bahadir K Gunturk, Yucel Altunbasak, and Russell M Mersereau. Super-resolution reconstruction of compressed video using transform-domain statistics. IEEE Transactions on Image Processing, 13(1):33-43, 2004. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.689, + 0.894, + 0.758 + ], + "angle": 0, + "content": "[25] Yong Guo, Jian Chen, Jingdong Wang, Qi Chen, Jiezhang Cao, Zeshuai Deng, Yanwu Xu, and Mingkui Tan. Closed-loop matters: Dual regression networks for single image super-resolution. In IEEE Conference on Computer Vision and Pattern Recognition, pages 5407-5416, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.759, + 0.893, + 0.801 + ], + "angle": 0, + "content": "[26] Ali Hassani, Steven Walton, Jiachen Li, Shen Li, and Humphrey Shi. Neighborhood attention transformer. arXiv preprint arXiv:2204.07143, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.802, + 0.894, + 0.857 + ], + "angle": 0, + "content": "[27] Kaiming He, Jian Sun, and Xiaou Tang. Single image haze removal using dark channel prior. IEEE Transactions on Pattern Analysis and Machine Intelligence, 33(12):2341-2353, 2010. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.858, + 0.894, + 0.902 + ], + "angle": 0, + "content": "[28] Xuecai Hu, Haoyuan Mu, Xiangyu Zhang, Zilei Wang, Tieniu Tan, and Jian Sun. Meta-sr: A magnification-arbitrary network for super-resolution. In IEEE Conference on Computer" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1804" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.11, + 0.092, + 0.471, + 0.119 + ], + "angle": 0, + "content": "Vision and Pattern Recognition, pages 1575-1584, 2019. 1, 2, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.121, + 0.472, + 0.176 + ], + "angle": 0, + "content": "[29] Jia-Bin Huang, Abhishek Singh, and Narendra Ahuja. Single image super-resolution from transformed self-exemplars. In IEEE Conference on Computer Vision and Pattern Recognition, pages 5197-5206, 2015. 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.177, + 0.472, + 0.245 + ], + "angle": 0, + "content": "[30] Andrey Ignatov, Nikolay Kobyshev, Radu Timofte, Kenneth Vanhoey, and Luc Van Gool. Dslr-quality photos on mobile devices with deep convolutional networks. In IEEE International Conference on Computer Vision, pages 3277-3285, 2017. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.246, + 0.472, + 0.302 + ], + "angle": 0, + "content": "[31] Takashi Isobe, Xu Jia, Shuhang Gu, Songjiang Li, Shengjin Wang, and Qi Tian. Video super-resolution with recurrent structure-detail network. In European Conference on Computer Vision, pages 645-660, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.302, + 0.472, + 0.357 + ], + "angle": 0, + "content": "[32] Xixi Jia, Sanyang Liu, Xiangchu Feng, and Lei Zhang. Focnet: A fractional optimal control network for image denoising. In IEEE Conference on Computer Vision and Pattern Recognition, pages 6054-6063, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.359, + 0.472, + 0.426 + ], + "angle": 0, + "content": "[33] Chiyu Jiang, Avneesh Sud, Ameesh Makadia, Jingwei Huang, Matthias Nießner, Thomas Funkhouser, et al. Local implicit grid representations for 3d scenes. In IEEE Conference on Computer Vision and Pattern Recognition, pages 6001-6010, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.428, + 0.472, + 0.482 + ], + "angle": 0, + "content": "[34] Jiwon Kim, Jung Kwon Lee, and Kyoung Mu Lee. Accurate image super-resolution using very deep convolutional networks. In IEEE Conference on Computer Vision and Pattern Recognition, pages 1646-1654, 2016. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.484, + 0.472, + 0.552 + ], + "angle": 0, + "content": "[35] Yoonsik Kim, Jae Woong Soh, Jaewoo Park, Byeongyong Ahn, Hyun-Seung Lee, Young-Su Moon, and Nam Ik Cho. A pseudo-blind convolutional neural network for the reduction of compression artifacts. IEEE Transactions on Circuits and Systems for Video Technology, 30(4):1121-1135, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.553, + 0.472, + 0.594 + ], + "angle": 0, + "content": "[36] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In International Conference on Learning Representations, 2015. 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.595, + 0.472, + 0.65 + ], + "angle": 0, + "content": "[37] Wei-Sheng Lai, Jia-Bin Huang, Narendra Ahuja, and Ming-Hsuan Yang. Deep laplacian pyramid networks for fast and accurate super-resolution. In IEEE Conference on Computer Vision and Pattern Recognition, pages 624-632, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.651, + 0.472, + 0.734 + ], + "angle": 0, + "content": "[38] Christian Ledig, Lucas Theis, Ferenc Huszar, Jose Caballero, Andrew Cunningham, Alejandro Acosta, Andrew Aitken, Alykhan Tejani, Johannes Totz, Zehan Wang, et al. Photorealistic single image super-resolution using a generative adversarial network. In IEEE Conference on Computer Vision and Pattern Recognition, pages 4681-4690, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.735, + 0.472, + 0.789 + ], + "angle": 0, + "content": "[39] Jaewon Lee and Kyong Hwan Jin. Local texture estimator for implicit representation function. In IEEE Conference on Computer Vision and Pattern Recognition, pages 1929-1938, 2022. 1, 2, 3, 4, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.79, + 0.472, + 0.858 + ], + "angle": 0, + "content": "[40] Jingyun Liang, Jiezhang Cao, Guolei Sun, Kai Zhang, Luc Van Gool, and Radu Timofte. Swinir: Image restoration using swin transformer. In IEEE International Conference on Computer Vision Workshops, pages 1833-1844, 2021. 1, 2, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.859, + 0.472, + 0.902 + ], + "angle": 0, + "content": "[41] Jingyun Liang, Andreas Lugmayr, Kai Zhang, Martin Danelljan, Luc Van Gool, and Radu Timofte. Hierarchical conditional flow: A unified framework for image super-resolution" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.893, + 0.12 + ], + "angle": 0, + "content": "and image rescaling. In IEEE Conference on International Conference on Computer Vision, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.121, + 0.894, + 0.188 + ], + "angle": 0, + "content": "[42] Jingyun Liang, Guolei Sun, Kai Zhang, Luc Van Gool, and Radu Timofte. Mutual affine network for spatially variant kernel estimation in blind image super-resolution. In IEEE Conference on International Conference on Computer Vision, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.19, + 0.893, + 0.245 + ], + "angle": 0, + "content": "[43] Jingyun Liang, Kai Zhang, Shuhang Gu, Luc Van Gool, and Radu Timofte. Flow-based kernel prior with application to blind super-resolution. In IEEE Conference on Computer Vision and Pattern Recognition, pages 10601-10610, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.246, + 0.893, + 0.314 + ], + "angle": 0, + "content": "[44] Bee Lim, Sanghyun Son, Heewon Kim, Seungjun Nah, and Young Mu Lee. Enhanced deep residual networks for single image super-resolution. In IEEE Conference on Computer Vision and Pattern Recognition Workshops, pages 136-144, 2017. 1, 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.316, + 0.893, + 0.357 + ], + "angle": 0, + "content": "[45] Ding Liu, Bihan Wen, Yuchen Fan, Chen Change Loy, and Thomas S Huang. Non-local recurrent network for image restoration. arXiv preprint arXiv:1806.02919, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.359, + 0.894, + 0.426 + ], + "angle": 0, + "content": "[46] David Martin, Charless Fowlkes, Doron Tal, and Jitendra Malik. A database of human segmented natural images and its application to evaluating segmentation algorithms and measuring ecological statistics. In IEEE Conference on International Conference on Computer Vision, pages 416-423, 2001. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.428, + 0.894, + 0.495 + ], + "angle": 0, + "content": "[47] Yusuke Matsui, Kota Ito, Yuji Aramaki, Azuma Fujimoto, Toru Ogawa, Toshihiko Yamasaki, and Kiyoharu Aizawa. Sketch-based manga retrieval using manga109 dataset. Multi-media Tools and Applications, 76(20):21811-21838, 2017. 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.497, + 0.893, + 0.551 + ], + "angle": 0, + "content": "[48] Yiqun Mei, Yuchen Fan, and Yuqian Zhou. Image superresolution with non-local sparse attention. In IEEE Conference on Computer Vision and Pattern Recognition, pages 3517-3526, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.553, + 0.893, + 0.621 + ], + "angle": 0, + "content": "[49] Yiqun Mei, Yuchen Fan, Yuqian Zhou, Lichao Huang, Thomas S Huang, and Honghui Shi. Image super-resolution with cross-scale non-local attention and exhaustive self-exemplars mining. In IEEE Conference on Computer Vision and Pattern Recognition, pages 5690-5699, 2020. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.622, + 0.893, + 0.664 + ], + "angle": 0, + "content": "[50] Tomer Michaeli and Michal Irani. Nonparametric blind superresolution. In IEEE Conference on International Conference on Computer Vision, pages 945–952, 2013. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.665, + 0.894, + 0.732 + ], + "angle": 0, + "content": "[51] Mateusz Michalkiewicz, Jhony K Pontes, Dominic Jack, Mahsa Baktashmotlagh, and Anders Eriksson. Implicit surface representations as layers in neural networks. In IEEE International Conference on Computer Vision, pages 4743-4752, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.734, + 0.894, + 0.801 + ], + "angle": 0, + "content": "[52] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.803, + 0.893, + 0.858 + ], + "angle": 0, + "content": "[53] Anish Mittal, Anush K Moorthy, and Alan C Bovik. Blind/referenceless image spatial quality evaluator. In Asilomar Conference on Signals, Systems and Computers, pages 723-727, 2011. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.859, + 0.893, + 0.901 + ], + "angle": 0, + "content": "[54] Anish Mittal, Rajiv Soundararajan, and Alan C Bovik. Making a \"completely blind\" image quality analyzer. IEEE Signal Processing Letters, 20(3):209-212, 2012. 8" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1805" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.161 + ], + "angle": 0, + "content": "[55] Michael Niemeyer, Lars Mescheder, Michael Oechsle, and Andreas Geiger. Differentiable volumetric rendering: Learning implicit 3d representations without 3d supervision. In IEEE Conference on Computer Vision and Pattern Recognition, pages 3504-3515, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.164, + 0.471, + 0.233 + ], + "angle": 0, + "content": "[56] Ben Niu, Weilei Wen, Wenqi Ren, Xiangde Zhang, Lianping Yang, Shuzhen Wang, Kaihao Zhang, Xiaochun Cao, and Haifeng Shen. Single image super-resolution via a holistic attention network. In European Conference on Computer Vision, pages 191-207, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.234, + 0.471, + 0.29 + ], + "angle": 0, + "content": "[57] Michael Oechsle, Lars Mescheder, Michael Niemeyer, Thilo Strauss, and Andreas Geiger. Texture fields: Learning texture representations in function space. In IEEE International Conference on Computer Vision, pages 4531-4540, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.292, + 0.471, + 0.346 + ], + "angle": 0, + "content": "[58] Songyou Peng, Michael Niemeyer, Lars Mescheder, Marc Pollefeys, and Andreas Geiger. Convolutional occupancy networks. In European Conference on Computer Vision, pages 523-540, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.349, + 0.471, + 0.403 + ], + "angle": 0, + "content": "[59] Yali Peng, Lu Zhang, Shigang Liu, Xiaojun Wu, Yu Zhang, and Xili Wang. Dilated residual networks with symmetric skip connection for image denoising. Neurocomputing, 345:67-76, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.406, + 0.471, + 0.489 + ], + "angle": 0, + "content": "[60] Wenzhe Shi, Jose Caballero, Ferenc Huszár, Johannes Totz, Andrew P Aitken, Rob Bishop, Daniel Rueckert, and Zehan Wang. Real-time single image and video super-resolution using an efficient sub-pixel convolutional neural network. In IEEE Conference on Computer Vision and Pattern Recognition, pages 1874-1883, 2016. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.49, + 0.471, + 0.585 + ], + "angle": 0, + "content": "[61] Wenzhe Shi, Jose Caballero, Christian Ledig, Xiahai Zhuang, Wenjia Bai, Kanwal Bhatia, Antonio M Marvao, Tim Dawes, Declan O'Regan, and Daniel Rueckert. Cardiac image superresolution with global correspondence using multi-atlas patchmatch. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 9-16, 2013. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.588, + 0.471, + 0.643 + ], + "angle": 0, + "content": "[62] Vincent Sitzmann, Julien Martel, Alexander Bergman, David Lindell, and Gordon Wetzstein. Implicit neural representations with periodic activation functions. Advances in Neural Information Processing Systems, 33:7462-7473, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.646, + 0.471, + 0.701 + ], + "angle": 0, + "content": "[63] Vincent Sitzmann, Michael Zollhöfer, and Gordon Wetzstein. Scene representation networks: Continuous 3d-structure-aware neural scene representations. Advances in Neural Information Processing Systems, 32, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.703, + 0.471, + 0.758 + ], + "angle": 0, + "content": "[64] Sanghyun Son and Kyoung Mu Lee. Srwarp: Generalized image super-resolution under arbitrary transformation. In IEEE Conference on Computer Vision and Pattern Recognition, pages 7782-7791, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.76, + 0.471, + 0.814 + ], + "angle": 0, + "content": "[65] Ying Tai, Jian Yang, Xiaoming Liu, and Chunyan Xu. Memnet: A persistent memory network for image restoration. In IEEE International Conference on Computer Vision, pages 4539-4547, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.817, + 0.471, + 0.871 + ], + "angle": 0, + "content": "[66] Radu Timofte, Vincent De Smet, and Luc Van Gool. Anchored neighborhood regression for fast example-based superresolution. In IEEE Conference on International Conference on Computer Vision, pages 1920-1927, 2013. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.874, + 0.471, + 0.901 + ], + "angle": 0, + "content": "[67] Radu Timofte, Vincent De Smet, and Luc Van Gool. A+: Adjusted anchored neighborhood regression for fast super-" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "resolution. In Asian Conference on Computer Vision, pages 111-126, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.122, + 0.894, + 0.178 + ], + "angle": 0, + "content": "[68] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Advances in Neural Information Processing Systems, 2017. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.179, + 0.894, + 0.234 + ], + "angle": 0, + "content": "[69] N Venkatanath, D Praneeth, Maruthi Chandrasekhar Bh, Sumohana S Channappayya, and Swarup S Medasani. Blind image quality evaluation using perception based features. In National Conference on Communications, pages 1-6, 2015. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.236, + 0.894, + 0.305 + ], + "angle": 0, + "content": "[70] Longguang Wang, Yingqian Wang, Xiaoyu Dong, Qingyu Xu, Jungang Yang, Wei An, and Yulan Guo. Unsupervised degradation representation learning for blind super-resolution. In IEEE Conference on Computer Vision and Pattern Recognition, pages 10581-10590, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.307, + 0.894, + 0.376 + ], + "angle": 0, + "content": "[71] Longguang Wang, Yingqian Wang, Zhengfa Liang, Zaiping Lin, Jungang Yang, Wei An, and Yulan Guo. Learning parallel attention for stereo image super-resolution. In IEEE Conference on Computer Vision and Pattern Recognition, pages 12250-12259, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.378, + 0.894, + 0.445 + ], + "angle": 0, + "content": "[72] Longguang Wang, Yingqian Wang, Zaiping Lin, Jungang Yang, Wei An, and Yulan Guo. Learning a single network for scale-arbitrary super-resolution. In IEEE Conference on International Conference on Computer Vision, pages 10581-10590, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.448, + 0.892, + 0.504 + ], + "angle": 0, + "content": "[73] Longguang Wang, Yingqian Wang, Zaiping Lin, Jungang Yang, Wei An, and Yulan Guo. Learning a single network for scale-arbitrary super-resolution. In IEEE International Conference on Computer Vision, pages 4801-4810, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.505, + 0.892, + 0.56 + ], + "angle": 0, + "content": "[74] Xintao Wang, Liangbin Xie, Chao Dong, and Ying Shan. Real-esrgan: Training real-world blind super-resolution with pure synthetic data. In International Conference on Computer Vision Workshops. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.562, + 0.894, + 0.63 + ], + "angle": 0, + "content": "[75] Xintao Wang, Ke Yu, Shixiang Wu, Jinjin Gu, Yihao Liu, Chao Dong, Yu Qiao, and Chen Change Loy. Esrgan: Enhanced super-resolution generative adversarial networks. In European Conference on Computer Vision Workshops, pages 701-710, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.632, + 0.894, + 0.687 + ], + "angle": 0, + "content": "[76] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE Transactions on Image Processing, 13(4):600-612, 2004. 5, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.689, + 0.894, + 0.758 + ], + "angle": 0, + "content": "[77] Yunxuan Wei, Shuhang Gu, Yawei Li, Radu Timofte, Longcun Jin, and Hengjie Song. Unsupervised real-world image super resolution via domain-distance aware training. In IEEE Conference on Computer Vision and Pattern Recognition, pages 13385-13394, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.76, + 0.894, + 0.815 + ], + "angle": 0, + "content": "[78] Xingqian Xu, Zhangyang Wang, and Humphrey Shi. Ultrasr: Spatial encoding is a missing key for implicit image function-based arbitrary-scale super-resolution. arXiv preprint arXiv:2103.12716, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.817, + 0.892, + 0.871 + ], + "angle": 0, + "content": "[79] Jingyu Yang, Sheng Shen, Huanjing Yue, and Kun Li. Implicit transformer network for screen content image continuous super-resolution. Advances in Neural Information Processing Systems, 34:13304-13315, 2021. 2, 3, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.874, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[80] Roman Zeyde, Michael Elad, and Matan Protter. On single image scale-up using sparse-representations. In International" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.894, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.946, + 0.516, + 0.956 + ], + "angle": 0, + "content": "1806" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.11, + 0.092, + 0.472, + 0.119 + ], + "angle": 0, + "content": "Conference on Curves and Surfaces, pages 711-730, 2010. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.122, + 0.472, + 0.177 + ], + "angle": 0, + "content": "[81] Kai Zhang, Yawei Li, Wangmeng Zuo, Lei Zhang, Luc Van Gool, and Radu Timofte. Plug-and-play image restoration with deep denoiser prior. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.179, + 0.472, + 0.233 + ], + "angle": 0, + "content": "[82] Kai Zhang, Jingyun Liang, Luc Van Gool, and Radu Timofte. Designing a practical degradation model for deep blind image super-resolution. In IEEE Conference on International Conference on Computer Vision, 2021. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.235, + 0.472, + 0.289 + ], + "angle": 0, + "content": "[83] Kai Zhang, Wangmeng Zuo, and Lei Zhang. Ffdnet: Toward a fast and flexible solution for cnn-based image denoising. IEEE Transactions on Image Processing, 27(9):4608-4622, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.292, + 0.472, + 0.347 + ], + "angle": 0, + "content": "[84] Kai Zhang, Wangmeng Zuo, and Lei Zhang. Learning a single convolutional super-resolution network for multiple degradations. In IEEE Conference on Computer Vision and Pattern Recognition, pages 3262-3271, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.349, + 0.472, + 0.416 + ], + "angle": 0, + "content": "[85] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In IEEE Conference on Computer Vision and Pattern Recognition, pages 586-595, 2018. 5, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.419, + 0.472, + 0.473 + ], + "angle": 0, + "content": "[86] Yulun Zhang, Kunpeng Li, Kai Li, Lichen Wang, Bineng Zhong, and Yun Fu. Image super-resolution using very deep residual channel attention networks. In European Conference on Computer Vision, pages 286-301, 2018. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.476, + 0.472, + 0.516 + ], + "angle": 0, + "content": "[87] Yulun Zhang, Kunpeng Li, Kai Li, Bineng Zhong, and Yun Fu. Residual non-local attention networks for image restoration. arXiv preprint arXiv:1903.10082, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.518, + 0.472, + 0.572 + ], + "angle": 0, + "content": "[88] Yulun Zhang, Yapeng Tian, Yu Kong, Bineng Zhong, and Yun Fu. Residual dense network for image super-resolution. In IEEE Conference on Computer Vision and Pattern Recognition, pages 2472-2481, 2018. 1, 2, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.575, + 0.472, + 0.629 + ], + "angle": 0, + "content": "[89] Yulun Zhang, Yapeng Tian, Yu Kong, Bineng Zhong, and Yun Fu. Residual dense network for image restoration. IEEE Transactions on Pattern Analysis and Machine Intelligence, 43(7):2480-2495, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.632, + 0.472, + 0.685 + ], + "angle": 0, + "content": "[90] Shangchen Zhou, Jiawei Zhang, Wangmeng Zuo, and Chen Change Loy. Cross-scale internal graph neural network for image super-resolution. arXiv preprint arXiv:2006.16673, 2020.2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.688, + 0.472, + 0.729 + ], + "angle": 0, + "content": "[91] Wilman WW Zou and Pong C Yuen. Very low resolution face recognition problem. IEEE Transactions on Image Processing, 21(1):327-340, 2011. 1" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.729 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "1807" + } + ] +] \ No newline at end of file diff --git a/2023/CiaoSR_ Continuous Implicit Attention-in-Attention Network for Arbitrary-Scale Image Super-Resolution/dcf45018-bd80-4a92-ad40-a481bde62c68_origin.pdf b/2023/CiaoSR_ Continuous Implicit Attention-in-Attention Network for Arbitrary-Scale Image Super-Resolution/dcf45018-bd80-4a92-ad40-a481bde62c68_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..9f9e5adfe898ffa02f01a8a64ab0175d78fd12c5 --- /dev/null +++ b/2023/CiaoSR_ Continuous Implicit Attention-in-Attention Network for Arbitrary-Scale Image Super-Resolution/dcf45018-bd80-4a92-ad40-a481bde62c68_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bff965fa1a4bc3f845c0d55031a80b7231d406fca399fc6618b04628d488791d +size 3670516 diff --git a/2023/CiaoSR_ Continuous Implicit Attention-in-Attention Network for Arbitrary-Scale Image Super-Resolution/full.md b/2023/CiaoSR_ Continuous Implicit Attention-in-Attention Network for Arbitrary-Scale Image Super-Resolution/full.md new file mode 100644 index 0000000000000000000000000000000000000000..77e6d2115b6417023f7e28b521adcefed4c23eaa --- /dev/null +++ b/2023/CiaoSR_ Continuous Implicit Attention-in-Attention Network for Arbitrary-Scale Image Super-Resolution/full.md @@ -0,0 +1,400 @@ +# CiaoSR: Continuous Implicit Attention-in-Attention Network for Arbitrary-Scale Image Super-Resolution + +Jiezhang Cao $^{1}$ Qin Wang $^{1}$ Yongqin Xian $^{1*}$ Yawei Li $^{1}$ Bingbing Ni $^{2}$ Zhiming Pi $^{2}$ + +Kai Zhang $^{1\dagger}$ Yulun Zhang $^{1\dagger}$ Radu Timofte $^{1,3}$ Luc Van Gool $^{1,4}$ + +$^{1}$ ETH Zürich $^{2}$ Huawei Inc. $^{3}$ University of Wurzburg $^{4}$ KU Leuven + +https://github.com/caojiezhang/CiaoSR + +# Abstract + +Learning continuous image representations is recently gaining popularity for image super-resolution (SR) because of its ability to reconstruct high-resolution images with arbitrary scales from low-resolution inputs. Existing methods mostly ensemble nearby features to predict the new pixel at any queried coordinate in the SR image. Such a local ensemble suffers from some limitations: i) it has no learnable parameters and it neglects the similarity of the visual features; ii) it has a limited receptive field and cannot ensemble relevant features in a large field which are important in an image. To address these issues, this paper proposes a continuous implicit attention-in-attention network, called CiaoSR. We explicitly design an implicit attention network to learn the ensemble weights for the nearby local features. Furthermore, we embed a scale-aware attention in this implicit attention network to exploit additional non-local information. Extensive experiments on benchmark datasets demonstrate CiaoSR significantly outperforms the existing single image SR methods with the same backbone. In addition, CiaoSR also achieves the state-of-the-art performance on the arbitrary-scale SR task. The effectiveness of the method is also demonstrated on the real-world SR setting. More importantly, CiaoSR can be flexibly integrated into any backbone to improve the SR performance. + +# 1. Introduction + +Single image super-resolution (SISR), which aims to reconstruct a high-resolution (HR) image from a low-resolution (LR) one, has been widely employed in many practical applications [24, 61, 91]. However, deep neural networks (DNN)-based SISR methods are facing some limitations in some real-world scenarios with arbitrary scales. For example, camera users may want to enhance the digital zoom quality by super-resolving a photo or a video to + +![](images/ff01524b6deb00056d2d082586a728943e8e013d0bdd1e1107b95a47d23c9250.jpg) +Figure 1. Comparison of different backbones and implicit models. Our proposed implicit neural network on RDN [88] has better performance than SwinIR [40]. Check Section 5 for details. + +continuous arbitrary scales. Most existing DNN-based SISR methods [40, 44, 86] need to train a series of models for all different scales separately. However, it can be impractical to store all these models on the device due to limited storage and computing power. Alternatively, arbitrary-scale image SR methods [13, 28, 39] aim to train a single network for all scales in a continuous manner. + +Most existing SISR methods [40,44,86] consist of a DNN and an upsampling module (e.g., pixel shuffling [60]) at a discrete scale. While substantial progresses have been made in the DNN backbones for SR, there is little attempt to study the upsampling module. A natural question to ask is: Does the pixel shuffling hinder the potential of SR models? One limitation of the pixel shuffling module is that it cannot synthesize SR images at large unseen and continuous scales. To tackle this, one can treat synthesizing different-scale SR images as a multi-task learning problem, and train a specific upsampling module for each scale [44]. However, these tasks are dependent and highly inter-related. Neglecting the correlation of different-scale SR tasks may lead to discrete representations and limited performance. Under a certain capacity of a network, training a model on multi-tasks may sacrifice the performance or have the comparable performance on each task. These above disadvantages limit its applicability and flexibility in the real-world scenarios. + +To address these, most existing arbitrary-scale SR methods [13, 28, 39] replace the upsampling method with an implicit neural function and boost the performance. These methods predict an RGB value at the query point in an image + +![](images/eccbf2d2b3a66c58a8bff1e822218e41c9a651de862409ff0b1be2eabb7b8bdb.jpg) +Figure 2. Comparisons of different attention mechanisms. (a) Self-attention can predict pixel features on the grid, but it cannot be directly used in arbitrary-scale SR without considering coordinates. (b) Most existing methods can be treated as coordinate-based implicit attention since they calculate the distance between a key and query coordinate, and then use a function $g$ to aggregate with the value features. However, these methods ignore the distance between the features. (c) Our implicit attention not only considers the coordinate distance, but also the distance among features with visual information. + +by ensembling features within a local region. However, the local ensemble methods have limitations in the ensemble weights and insufficient information (e.g., non-local information). The ensemble weights are often calculated by the area of the rectangle between the query point and each nearest point, which is equivalent to the bilinear interpolation. Thus, those methods cannot adaptively ensemble local features since there is no trainable parameter. These weights are only related to the coordinates of the local features, but independent of the local features. Ignoring both the coordinates and the local features lose visual information and result in blurry artifacts. It is important and necessary to design a new implicit network to predict the weights and exploit more information in the local ensemble. + +In this paper, we propose a novel implicit attention model to enhance arbitrary-scale SR performance. Specifically, we use our attention to predict the ensemble weights by considering both the similarity and coordinate distance of local features, as shown in Figure 2. Based on such learnable weights, the implicit model can adaptively aggregate local features according to different inputs. To enrich more information, we introduce an attention in our implicit attention, which helps discover more features in a larger receptive field. + +Our contributions are summarized as follows: + +- We propose a novel continuous implicit attention-inattention network for arbitrary-scale image SR, called CiaoSR. Different from most existing local ensemble methods, our method explicitly learns the ensemble weights and exploits scale-aware non-local information. +- Our CiaoSR can be flexibly integrated into any backbone, allowing the network to super-resolve an image at arbitrary scales and improve the SR performance in Figure 1. +- Extensive experiments demonstrate CiaoSR achieves the state-of-the-art performance in both SISR and arbitrary-scale SR tasks. Besides, our CiaoSR has good generalization on both in-scale and out-of-scale distributions. Last, we extend our method to real-world SR settings to synthesize arbitrary-scale images. + +# 2. Related Work + +Single image super-resolution (SISR). SISR aims to synthesize high-resolution (HR) images from low-resolution (LR) images. Compared with DNN-based SR methods [6-8, 23, 27, 50, 66, 67], methods in recent years build on deep convolutional neural network (CNN) to improve the performance, such as SRCNN [19], SRResNet [38], EDSR [44], RDN [88] and RCAN [86]. To further improve SR performance, some methods design CNN with residual block [9, 34, 81], dense block [75, 88, 89] and others [14, 16, 18, 20, 21, 25, 31, 32, 35, 37, 41-43, 59, 65, 70-72, 77, 83, 84, 87, 90]. In addition, some SR methods are built based on attention mechanism [68], such as channel attention [17, 56, 86], self-attention (IPT [11] and SwinIR [40], HAT [12]), non-local attention [45, 48]. However, most methods focus on one specific scale, which limits the applicability and flexibility in arbitrary-scale. + +Arbitrary-scale super-resolution. To tackle this problem, very recently, the more practical setup of arbitrary-scale SR is considered, which aims to super-resolute images with arbitrary scales by a single model. MetaSR [28] makes the first attempt to propose an arbitrary-scale meta-upscale module. To improve the performance, many arbitrary-scale methods [64, 73] are proposed. With the help of implicit neural representation [2, 10, 15, 22, 33, 51, 52, 55, 57, 58, 62, 63], LIIF [13] predicts the RGB value at an arbitrary query coordinate by taking an image coordinate and features of backbone around the coordinate. The features are extracted by single image super-resolution methods, e.g., EDSR [44], RDN [88] and SwinIR [40]. To improve the performance, existing methods [39, 78] propose to integrate more features in SR models. For example, LTE [39] proposes a local texture estimator by characterizing image textures in the Fourier space. UltraSR [78] integrates spatial coordinates and periodic encoding in the implicit network. These methods use the bilinear interpolation to ensemble nearby features. However, such an ensemble way has no learnable parameters. Recently, ITSRN [79] learns the weights by taking the coordinate distance and scale token into a mapping. Most methods learn the ensemble without the feature similarity. + +![](images/a5f7fd21f2f55c2745db7f466b043c09a66a1933fc853e94ab22e48dda0502fa.jpg) +Figure 3. The architecture of our continuous implicit attention-in-attention network. Given an LR image, the encoder extracts features as latent codes. For a query point, we have a query feature and key features close to the query point, and the scale-aware non-local attention module extracts non-local features as value. Last, we use the triple of query, key and value to predict the RGB value at the query point. + +# 3. Preliminary and Motivation + +Let $I$ be a continuous image, and $\pmb{x}$ be a 2D coordinate of a pixel in the image $\pmb{I}$ . Formally, given a 2D coordinate $\pmb{x}$ in the continuous image domain and latent code $\pmb{Z}$ extracted by deep neural networks, the RGB value can be predicted by an implicit image function which can be defined as follows, + +$$ +\boldsymbol {I} (\boldsymbol {x}) = f (\boldsymbol {Z}, \boldsymbol {x}), \tag {1} +$$ + +where the implicit image function can be parameterized by a multilayer perceptron (MLP). Note that this implicit image function is shared by all images. For recent SISR methods [40, 86], the implicit image function $f$ can be implemented as a PixelShuffle [60] with convolutions with a specific scale. However, these methods are independent on the coordinates, leading to an issue that they only adapt to the specific scale and are inflexible to synthesize arbitrary-scale images. + +To predict the RGB value $I_{q}$ at an arbitrary query coordinate $x_{q}$ , most existing methods [13, 39, 79] propose to compute the RGB value at coordinate $x_{q}$ by directly ensembling its neighborhood information, + +$$ +\pmb {I} _ {q} := \pmb {I} (\pmb {x} _ {q}) = \sum_ {(i, j) \in \mathcal {I}} w _ {i, j} \cdot f (\pmb {Z} _ {i, j} ^ {*}, \pmb {x} _ {q} - \pmb {x} _ {i, j} ^ {*}), (2) +$$ + +where $\mathcal{I}$ is the local region centered at the query coordinate $x_{q}$ , e.g., $\mathcal{I}$ can be top-left, top-right, bottom-left, bottom-right coordinates, and $w_{i}$ is the weight of the neighboring pixel $x_{i,j}^{*}$ , which is calculated w.r.t. the area of the rectangle between $x_{q}$ and $x_{i,j}^{*}$ , as shown in Figure 4. However, the performance improvement is limited because the ensemble weight $w_{i,j}$ is purely based on the coordinates. The visual similarities are completely ignored in the weight calculation. Besides, these methods only consider the nearest latent codes, leading to a limited receptive field. In this paper, our goal is to learn the weights adaptively by leveraging both visual information and coordinate information. + +# 4. Proposed Method + +The above local ensemble Eqn. (2) has a similar form to the attention mechanism. Specifically, the weights $w_{i}$ in Eqn. (2) can be modeled as an attention map and the latter term $f$ can be the value in the attention mechanism. Such an attention map is related to the similarities of the latent code and coordinate, and it can be calculated using both query and key which can integrate the latent code and coordinate information. In this sense, attention models can be used to mitigate the drawbacks of previous methods (i.e., calculated purely based on coordinates without considering any visual information) by learning the soft weights from both visual and coordinate information. However, the use of attention in the implicit function is non-trivial because standard self-attention [68] and neighborhood attention [26] mechanisms are based on visual features, and not conditioned on the coordinate information. To exploit the continuous representation learning from both the coordinate information and visual features, we propose a new attention for arbitrary-scale SR. + +Continuous implicit attention-in-attention. The architecture of the proposed continuous implicit attention-in-attention for arbitrary-scale SR (called CiaoSR) is shown in Figure 3. It is called attention-in-attention because the value contains an additional embedded attention that additionally captures non-local information of repetitive patterns. Specifically, given an LR image, our attention is provided with both the visual features and coordinate information to learn good ensemble weights for the local neighboring latent codes for any query at any arbitrary-scale $s$ . + +Given a query coordinate $\pmb{x}_q$ in the upsampled image, the known LR pixels in the local region centered at $\pmb{x}_q$ are considered for the local ensemble. We define $\pmb{x}_k$ and $\pmb{x}_v$ as the key and value coordinates of the local neighbor. Let $Q, K$ + +and $V$ be the latent code of query, key and value on the corresponding coordinates, respectively. Given pairs of coordinates and latent codes $(Q, x_{q})$ , $(K, x_{k})$ and $(V, x_{v})$ , we define a new implicit attention function (called i-Attention) to predict RGB values at the given query $x_{q}$ , + +$$ +\boldsymbol {I} _ {q} = \mathrm {i} - \text {A t t e n t i o n} (\boldsymbol {Q}, \boldsymbol {K}, \boldsymbol {V}; \boldsymbol {x} _ {q}, \boldsymbol {x} _ {k}, \boldsymbol {x} _ {v}), \tag {3} +$$ + +where the value feature $V$ can be embedded with another attention to aggregate more non-local information. + +As shown in Figure 2, our implicit attention is different from existing attention mechanisms by considering the additional coordinate information. In particular, the self-attention mechanism [68] without considering coordinates cannot predict pixel features that are not on the grid. The coordinate-based implicit attention calculates the distance between a key $\boldsymbol{x}_k$ and query coordinate $\boldsymbol{x}_q$ , which ignores the visual information and similarity of the features. + +# Implicit attention to learn weights for the local ensemble. + +Unlike previous methods [13, 39] which directly use the normalized area as the ensemble weights, we propose to first use an attention to learn better weights for the ensemble of the local region. Our attention takes both the coordinate distance $r$ , the local features $F_{l}$ , as well as non-local features $G$ into account for the learning of the weight. Formally, the attention can be written as + +$$ +\boldsymbol {I} _ {q} = \phi_ {q} \left(\sum_ {(i, j) \in \mathcal {I}} \underbrace {\sigma \left(\boldsymbol {Q} ^ {\top} \boldsymbol {K} _ {i , j}\right)} _ {\text {e n s e m b l e w e i g h t}} \boldsymbol {V} _ {i, j}\right), \tag {4} +$$ + +where $\phi_q$ is the query network, $\sigma$ is the Softmax function, and $\mathcal{I}$ is the local region centered at $\boldsymbol{x}_q$ . The query, key, and value for this local ensemble attention are defined as + +$$ +\left\{ \begin{array}{l} Q = \boldsymbol {F} ^ {*}, \\ \boldsymbol {K} _ {i, j} = \phi_ {k} \left(\left[ \boldsymbol {F} _ {i, j}, \left(\boldsymbol {r} _ {k}\right) _ {i, j}, \boldsymbol {s} \right]\right), \\ \boldsymbol {V} _ {i, j} = \phi_ {v} \left(\left[ \boldsymbol {F} _ {i, j}, \tilde {\boldsymbol {F}} _ {i, j} \right], \left(\boldsymbol {r} _ {v}\right) _ {i, j}, \boldsymbol {s} \right]), \end{array} \right. \tag {5} +$$ + +where $\phi_{k}$ and $\phi_v$ are the key and value networks, respectively, $F^{*}$ are the features of nearest neighbors on the grid of the LR image w.r.t. the query coordinate $\pmb{x}_q$ , $F_{i,j}$ is a local feature calculated by using the unfolding operator to the latent codes at the $(i,j)$ -th position, which is equivalent to concatenating the latent codes in neighboring $3\times 3$ region. $\tilde{F}_{i,j}$ are non-local features, which will be introduced below in detail. $s = [s_h;s_w]$ is a two-dimensional scale, and $\boldsymbol{r}_k$ and $\boldsymbol{r}_v$ are the relative distance between the coordinates in the local region, i.e., + +$$ +\boldsymbol {r} _ {k} = \boldsymbol {x} _ {q} - \left(\boldsymbol {x} _ {k}\right) _ {i, j}, \quad \boldsymbol {r} _ {v} = \boldsymbol {x} _ {q} - \left(\boldsymbol {x} _ {v}\right) _ {i, j}. \tag {6} +$$ + +# Embedded scale-aware attention for non-local features. + +Given a query coordinate, the local ensemble attention only considers the keys and values in the small local region, resulting in a limited receptive field. However, there may exist + +![](images/60e71634d4e595d01a36073e2dab1f03ab62f979760454334f462fe5a02e4a7d.jpg) +(a) Bilinear-based ensemble + +![](images/d144375c24c9c5b05cac966f0cca9b68ea41524564f55af2d1041fac4f2b5690.jpg) +(b) Attention-based ensemble +Figure 4. Comparison of different ensemble methods. (a) Most existing methods calculate the weights related to the area of the rectangle between the query coordinate and the nearest point. Such weights are equivalent to the weight of the bilinear interpolation. (b) Our attention-based ensemble calculates the attention map as the weights for feature aggregation. Besides, the ensemble method has a large receptive field for exploiting more information. + +similar textures with different scales within a single image in other positions. For example, repetitive patterns in different scales (e.g., facades, windows, etc. in a building, as shown in Figure 4) can exist in different locations within one image. To exploit the useful non-local information, we propose an embedded scale-aware non-local attention module, which aggregates the information from all coordinates on the grid of the LR image, motivated by [49]. Specifically, we obtain query $\tilde{Q}$ and value features $\tilde{V}$ from the latent code $F$ of the LR image, and downsample $F$ to a scale $s'$ as a key feature $\tilde{K}$ . We first calculate the pixel-wise similarity between the query located at $(i,j)$ and the key on other coordinates, and then aggregate the weighted information from all of them. Lastly, we use a convolutional layer $\varphi$ for downsampling with the scale $s'$ normalized similarity matrix to aggregate the features. Formally, the non-local feature $\tilde{F}$ located at $(i,j)$ can be calculated by + +$$ +\tilde {\boldsymbol {F}} _ {i, j} = \varphi \left(\sum_ {u, v} \frac {\exp \left(\tilde {\boldsymbol {Q}} _ {i , j} ^ {\top} \tilde {\boldsymbol {K}} _ {u , v}\right)}{\sum_ {u ^ {\prime} , v ^ {\prime}} \exp \left(\tilde {\boldsymbol {Q}} _ {i , j} ^ {\top} \tilde {\boldsymbol {K}} _ {u ^ {\prime} , v ^ {\prime}}\right)} \tilde {\boldsymbol {V}} _ {s ^ {\prime} u, s ^ {\prime} v} ^ {s ^ {\prime} p \times s ^ {\prime} p}\right), \tag {7} +$$ + +where $\tilde{V}_{s'u,s'v}^{s'p\times s'p}$ is the value feature patch of the size $s^{\prime}p\times s^{\prime}p$ located at $(s^{\prime}u,s^{\prime}v)$ , $\tilde{Q},\tilde{K}$ and $\tilde{V}$ are the query, key and value's non-local features, which can be calculated by + +$$ +\left\{ \begin{array}{l} \tilde {\boldsymbol {Q}} = \varphi_ {q} (\boldsymbol {F}), \\ \tilde {\boldsymbol {K}} = \varphi_ {k} (\boldsymbol {F} \downarrow_ {s ^ {\prime}}), \\ \tilde {\boldsymbol {V}} = \varphi_ {v} (\boldsymbol {F}), \end{array} \right. \tag {8} +$$ + +where $\downarrow_{s'}$ is the downsampling operation with a given scale $s'$ which can be $\{2,3,4\}$ , $\varphi_q$ , $\varphi_k$ and $\varphi_v$ are query, key and value networks. Different from [49], our method can integrate multi-scale features. In the experiment, we can search for relevant features in a small window (e.g., $256\times 256$ ), instead of the entire image. In this way, our method can guarantee a large receptive field, and reduce computational cost. + +Table 1. Quantitative comparison with state-of-the-art methods for arbitrary-scale SR on the DIV2K validation set (PSNR (dB)). Bold indicates the best performance. $\dagger$ indicates the implementation of [39]. + +
BackbonesMethodsIn-scaleOut-of-scale
×2×3×4×6×12×18×24×30
-Bicubic31.0128.2226.6624.8222.2721.0020.1919.59
EDSR [44]EDSR-baseline [44]34.5530.9028.94-----
EDSR-baseline-MetaSR [28]34.6430.9328.9226.6123.5522.0321.0620.37
EDSR-baseline-LIIF [13]34.6730.9629.0026.7523.7122.1721.1820.48
EDSR-baseline-ITSRN† [79]34.7130.9529.0326.7723.7122.1721.1820.49
EDSR-baseline-LTE [39]34.7231.0229.0426.8123.7822.2321.2420.53
EDSR-baseline-CiaoSR (ours)34.9131.1529.2326.9523.8822.3221.3220.59
RDN [88]RDN-baseline [88]34.9431.2229.19-----
RDN-MetaSR [28]35.0031.2729.2526.8823.7322.1821.1720.47
RDN-LIIF [13]34.9931.2629.2726.9923.8922.3421.3120.59
RDN-ITSRN† [79]35.0931.3629.3827.0623.9322.3621.3220.61
RDN-LTE [39]35.0431.3229.3327.0423.9522.4021.3620.64
RDN-CiaoSR (ours)35.1531.4229.4527.1624.0622.4821.4320.70
SwinIR [40]SwinIR-baseline [40]34.9431.2229.19-----
SwinIR-MetaSR† [28]35.1531.4029.3326.9423.8022.2621.2620.54
SwinIR-LIIF† [13]35.1731.4629.4627.1524.0222.4321.4020.67
SwinIR-ITSRN† [79]35.1931.4229.4827.1323.8322.3121.3120.55
SwinIR-LTE [39]35.2431.5029.5127.2024.0922.5021.4720.73
SwinIR-CiaoSR (ours)35.2931.5529.5927.2824.1522.5421.5120.74
+ +# 5. Experiments + +Datasets. For the arbitrary-scale SR task, we follow [13, 39] and use DIV2K [1] as the training set, which consists of 800 images in 2K resolution. In testing, we evaluate the models on the DIV2K validation set and a wide range of standard benchmark datasets, including Set5 [3], Set14 [80], B100 [46], Urban100 [29], and Manga109 [47]. + +Evaluation metrics. We use PSNR to evaluate the quality of the synthesized SR images. Following [13, 39], the PSNR value is calculated on all RGBs channels for the DIV2K validation set, and additionally also calculated on the Y channel (i.e., luminance) of the transformed YCbCr space for other benchmark test sets. In addition, other evaluation metrics (such as SSIM [76] and LPIPS [85]) are provided to evaluate the image quality. + +Implementation details. We follow the same experimental setup as prior works [13, 39]. To synthesize paired training data, we first crop the $48s \times 48s$ patches as ground-truth (GT) images, and use the Bicubic downsampling in PyTorch to have LR images with the patch size of $48 \times 48$ , where $s$ is a scale factor sampled in the uniform distribution $\mathcal{U}(1,4)$ . We sample $48^2$ pixels from both GT images and the corresponding coordinates. We use the existing SR models (e.g., EDSR [44], RDN [88] and SwinIR [40]) as backbones by removing their upsampling modules. The detailed architecture of our implicit network is provided in the supplementary material. For the training, we use Adam [36] as the optimizer, and use L1 loss to train all models for 1000 epochs with the batch size of 16 per GPU. We set the learning rate as $1e - 4$ at the beginning and decay the learning rate by a factor of 0.5 every 200 epochs. These experimental settings are identical to [13, 28, 39]. + +# 5.1. Comparisons with State-of-the-Art + +Quantitative results. In Tables 1-3, CiaoSR achieves the best performance with the highest PSNR across all datasets with all backbones concerned on both in-scale and out-of-scale distributions. For the in-scales, in particular, our model has a remarkable PSNR gain of $0.3\mathrm{dB}$ on Urban100 $(\times 4)$ , compared with the previous SOTA method [39] under the same RDN backbone. It is worth noting that CiaoSR with the RDN backbone [88] can surpass the performance of the better backbone SwinIR [40] (with its original upsampling module). With the help of the continuous-scale training, a single SR model trained with our method has better generalization performance than vanilla SR models (including EDSR [44], RDN [88] and SwinIR [40] trained on specific scale. For the out-of-scale experiments, our model also achieves the best generalization performance on unseen scales. The vanilla EDSR-baseline [44], RDN [88] and SwinIR [40] cannot be applied on the out-of-scale experiments because their decoder can only predict the scale that it was trained on. We observe that our method yields the largest improvement on Urban100 over existing arbitrary-scale methods (e.g., LIIF [13] and LTE [39]). + +Qualitative results. In Figure 5, we provide qualitative comparisons with other arbitrary-scale SR methods. Our model is able to synthesize the SR images with sharper textures than other methods. Taking the second line as an example, the textures in the LR image are degraded, but CiaoSR is still able to restore the textures of the building. In contrast, other methods can only restore part of the textures possibly due to the limitations of the local ensemble and insufficient features. More visual results on more test sets are put in the supplementary materials. + +Table 2. Quantitative comparison with state-of-the-art methods for in-scale SR on benchmark datasets (PSNR (dB)). Bold indicates the best performance. † indicates our implementation. + +
MethodsSet5 [3]Set14 [80]B100 [46]Urban100 [29]Manga109 [47]
×2×3×4×2×3×4×2×3×4×2×3×4×2×3×4
RDN [88]38.2434.7132.4734.0130.5728.8132.3429.2627.7232.8928.8026.6139.1834.1331.00
RDN-MetaSR [28]38.2234.6332.3833.9830.5428.7832.3329.2627.7132.9228.8226.55---
RDN-LIIF [13]38.1734.6832.5033.9730.5328.8032.3229.2627.7432.8728.8226.6839.2634.2131.20
RDN-ITSRN† [79]38.2334.7632.5534.1930.5928.8832.3829.3227.7933.0728.9626.7739.3434.3931.37
RDN-LTE [39]38.2334.7232.6134.0930.5828.8832.3629.3027.7733.0428.9726.8139.2834.3231.30
RDN-CiaoSR (ours)38.2934.8532.6634.2230.6528.9332.4129.3427.8333.3029.1727.1139.5134.5731.57
SwinIR [40]38.3534.8932.7234.1430.7728.9432.4429.3727.8333.4029.2927.0739.6034.7431.67
SwinIR-MetaSR [28]38.2634.7732.4734.1430.6628.8532.3929.3127.7533.2929.1226.7639.4634.6231.37
SwinIR-LIIF [13]38.2834.8732.7334.1430.7528.9832.3929.3427.8433.3629.3327.1539.5734.6831.71
SwinIR-ITSRN† [79]38.2234.7532.6334.2630.7528.9732.4229.3827.8533.4629.3427.1239.6034.7531.74
SwinIR-LTE [39]38.3334.8932.8134.2530.8029.0632.4429.3927.8633.5029.4127.2439.6334.7931.79
SwinIR-CiaoSR (ours)38.3834.9132.8434.3330.8229.0832.4729.4227.9033.6529.5227.4239.6734.8431.91
+ +Table 3. Quantitative comparison with state-of-the-art methods for out-of-scale SR on benchmark datasets (PSNR (dB)). Bold indicates the best performance. † indicates our implementation. + +
MethodsSet5 [3]Set14 [80]B100 [46]Urban100 [29]Manga109 [47]
×6×8×12×6×8×12×6×8×12×6×8×12×6×8×12
RDN-MetaSR [28]29.0429.96-26.5124.97-25.9024.83-23.9922.59----
RDN-LIIF [13]29.1527.1424.8626.6425.1523.2425.9824.9123.5724.2022.7921.1527.3325.0422.36
RDN-ITSRN† [79]29.3227.2524.8626.6825.1723.2826.0124.9323.5824.2322.8121.1627.4525.0423.35
RDN-LTE [39]29.3227.2624.7926.7125.1623.3126.0124.9523.6024.2822.8821.2227.4925.1222.43
RDN-CiaoSR (ours)29.4627.3624.9226.7925.2823.3726.0725.0023.6424.5823.1321.4227.7025.4022.63
SwinIR-MetaSR [28]29.0927.0224.8226.5825.0923.3325.9424.8723.5924.1622.7521.3127.2924.9622.35
SwinIR-LIIF [13]29.4627.36-26.8225.34-26.0725.01-24.5923.14-27.6925.28-
SwinIR-ITSRN† [79]29.3127.2424.7926.7125.3223.3026.0524.9623.5724.5023.0621.3427.7225.2322.47
SwinIR-LTE [39]29.5027.35-26.8625.42-26.0925.03-24.6223.17-27.8325.42-
SwinIR-CiaoSR (ours)29.6227.4524.9626.8825.4223.3826.1325.0723.6824.8423.3421.6028.0125.6122.79
+ +# 5.2. Ablation Study + +In this section, we conduct ablation studies to investigate effect of each component in our architecture. Our CiaoSR consists of the attention-in-attention network and a scale-aware non-local attention module. Table 4 shows the contributions of each component on performance. + +Attention-in-attention. The attention-in-attention network is the main branch of our architecture. On the one hand, the network learns to ensemble weights with both the coordinates and the features. On the other hand, the network is able to aggregate the local and non-local information to improve the SR performance. To investigate how the attention mechanism affects the performance, we remove this module and replace it with an MLP to predict the weights of the ensemble. The coordinate information is still fed into the MLP to enable arbitrary scale prediction. As shown in Table 4, the method suffers a significant performance drop without the attention mechanism. + +Non-local attention. The embedded non-local attention is proposed to capture the scale-aware information from other locations of the images to improve the performance. We investigate the importance of the non-local attention in an implicit model by comparing the results with and without this module. As shown in Table 4, by introducing the nonlocal attention, a further performance gain is achieved. This + +is because our module can exploit texture information from other scales in a large perceptual field. + +Training with different types of scales. We evaluate the effectiveness of our implicit model when trained with discrete (including single scale $\{\{2\} /\{3\} /\{4\}\}$ and multiple scales $\{2,3,4\}$ ) and continuous scales [1, 4], and show the results in Table 5. For the discrete type of scales, we observe that training our implicit model with a specific scale can also achieve significant improvement for that specific scale, compared with the original upsampling module in RDN [88]. Taking $\times 2$ scale as an example, our method has an improvement of $0.24\mathrm{dB}$ over RDN. However, training with a specific scale has poor generalization performance on unseen in-scale and out-of-scale. A few existing works [34, 44] consider the training with multiple discrete scales and found the performance of the joint learned model is comparable with single scale networks. Based on our implicit attention module, we are able to deal with more scales and exploit the correlation of multiple SR tasks at different scales. Unlike previous works [34, 44] which only achieve comparable performance compared with single scale models, CiaoSR achieves a better performance than training with one scale. We found our model has further improvement when trained on continuous scales. The improvement gain is attributed to our continuous representations design and our architecture, which enables efficient learning from cross-scale information. + +![](images/017229726e5527e13d8aef0b37d9453d1e3d2ebdfac7110c5c278dd744ee43ef.jpg) +004, $\times 4$ Urban100 [29] + +![](images/2fe773b1ddcce8e0ee21e1300ad8e6b8977e0249c147971b975feee4dd033d43.jpg) +Bicubic + +![](images/c00f007273de11b08d3527a8eb7d8f8fef086df0bddd263fca24865ff1866cb3.jpg) +RDN [88] + +![](images/8abf6fcadcaebe33966e339674e949c22ddb0871faddbf91e1346c1131aa3610.jpg) +RDN-MetaSR [28] + +![](images/27887c0a76eec3bc040c897bd324471f33bcbfb139a7b7be652212f64caa3bd0.jpg) +RDN-LIIF [13] + +![](images/c0e51091767761fb3c9ceb37a57853b63679be7157190a5af1a4604e9cd3db20.jpg) +TaiyouNiSmash, $\times 8$ ,Manga109 [47] + +![](images/2bcfdc0db4a5e5c0425a5ca1ec1ba3a26f7573bcc8d03650e461a3d6ca724f61.jpg) +RDN-ITSRN [79] + +![](images/2d6d904b3e21f45e9b65f82837ecd4bd3e955473a6aab3b075234602a51270eb.jpg) +RDN-LTE [39] + +![](images/292dfd17c80cfb1ae4a848b5470d1d40d7c46a5bd5ce26b8d8d5cae24925382e.jpg) +RDN-CiaoSR (ours) + +![](images/d7397286c223ff8215569bb6039996a8daa3c70efc7bb33ab1c40a0020662355.jpg) +GT + +![](images/342bf68f4fc0645d74ce8570ea2952c88a231299ed6e76ac06ad1f6df0f173da.jpg) +Bicubic + +![](images/cf7e663ba7cca5a4992d8cc395c2a0052c25b682e29731ca6c9415eae3d95e67.jpg) +$\mathbf{RDN}^*$ [88] + +![](images/abc35dd40df46d66c132bab6e4ac05ce6e573d5cbcb4baede219d2548093fa44.jpg) +RDN-MetaSR [28] + +![](images/0f7fed09bf0debfd75b82f8b4fe2f6962c8934750ba1dc1a27162702a614b926.jpg) +RDN-LIIF [13] + +![](images/86064fece19148d6164b92441003f33f93fac189a2422a55e68712c79c345c04.jpg) +RDN-ITSRN [79] + +![](images/17ec3b36a50b2eaf4cbcb4c3fcabed2ebe604e770b095ccbdf7059c4381ae2d2.jpg) +LTE [39] + +![](images/1e11f63b34a01d42b634a7dd52baf94259ad892dd975f767cbd09268ee9fae12.jpg) +RDN-CiaoSR (ours) + +![](images/e21fae8d43d29ea55b5c3b1a07f9222489850a2263351299d40921bb91453c4e.jpg) +GT + +![](images/135999c514ff8e8d08702af099945753c294d6fad8758bd96d9ffd71c90544fa.jpg) +Figure 5. Visual comparison of different methods on benchmarks. \*\* means the model first synthesizes twice to $\times 8$ images. + +![](images/08a54ee2edf90283c42f6bc8f2cbe684813aa400c0992f721efc4e27ca3b1e79.jpg) +LR + +![](images/5c7802381b4a599dad06a3dac80a4420ab7247a6e35ae8d2d1ce514017533deb.jpg) + +![](images/db09f92a251b8f0eba18effab240fa418673056ba0ef04aeece7e330ca6c1dee.jpg) +RealSR + +![](images/c04769fb9e91c1b33811499688678a7f710e60c35f53268bffe7614b837e92f7.jpg) + +![](images/d6003f11db1ca87a94d12b24cec1ebbbe4093a71f6c072b2ffc2c7f2dcfdf131.jpg) +BSRGAN + +![](images/931cf6292228ee8a7a9ac79e2b6e75d129b6e19c535c142a1884e2e53e10e9f9.jpg) + +![](images/69e488a7c5262c818a2f8769b4323953a87e785a943dbe0916f2f2345b604a2c.jpg) +Real-ESRGAN +Figure 6. Visual comparison of different methods on the RealSRSet [82] and DPED [30] dataset $(\times 8)$ . + +![](images/b256a9a1bc3a33ea924a66e938a317c36a9c18dc1321dc08dc1b38b723ccf74e.jpg) + +![](images/3fffe2e75369a2f7440ec9ebff286f31cdcbafdc9a5716c67ea03276177ebfcb.jpg) +Ours + +# 5.3. Further Study + +Effect on synthesis steps. Our trained model is able to synthesize SR images with any given arbitrary scale in one step. A natural alternative would be to synthesize a large scale in multiple steps. Taking a large scale $(e.g., \times 12)$ as an example, SR images can be generated by two steps $(e.g., \rightarrow \times 2 \rightarrow \times 12)$ , or three steps $(e.g., \rightarrow \times 2 \rightarrow \times 4 \rightarrow \times 12)$ . We compare the performance of synthesizing a large scale in one and multiple steps, and the results on the Urban100 and Manga109 datasets are shown in Table 6. We observe that our way of synthesizing SR images in one step has the best performance under the PSNR and SSIM metrics. In contrast, the performance gets worse with more steps because errors can accumulate along the multiple synthesizing steps. + +Model size and inference time. To demonstrate the efficiency of our proposed method, we investigate the model size and inference time of different arbitrary-scale SR methods. Results are shown in Table 7. Specifically, we set + +the input size as $192 \times 192$ and calculate the inference time on an NVIDIA Quadro RTX 6000. Our implicit model has much fewer parameters than the previous SOTA methods (e.g., LTE [39]). Our model has the best performance although it requires a little more inference time due to exploiting the non-local features. + +More evaluation metrics. To further demonstrate the effectiveness of our proposed method, we apply more evaluation metrics (e.g., SSIM [76] and LPIPS [85]) to evaluate the image quality across different methods on the Urban100 dataset. These evaluation metrics are widely used in the SR community. In general, higher SSIM and lower LPIPS correspond to better performance. As shown in Table 8, our method achieves the highest SSIM and lowest LPIPS, and thus has the best performance on both in-scale and out-of-scale distributions. Compared with other methods, our proposed CiaoSR is able to maintain more structural texture and perceptual information. + +Table 4. Ablation study on each component of our networks on Urban100. We use RDN [88] as the backbone. + +
Attention-in-attention✗✓✓
Scale-aware Attention Network✗✗✓
In-scale×232.8733.2433.30
×328.8229.1029.17
×426.6926.9627.11
Out-of-scale×624.2224.5024.58
×822.8022.9823.13
+ +Table 6. Comparison of (PSNR and SSIM) for different synthesis steps on Urban100 [29] and Manga109 [47]. + +
TypeSynthesis stepsUrban100 [29]Manga109 [47]
PSNRSSIMPSNRSSIM
Multiple steps→×2→×4→×1221.280.55722.460.720
→×2→×1221.320.55822.530.721
One step→×1221.420.56122.630.723
+ +Table 5. Ablation study on training our implicit model with different types of scales on Urban100. + +
TypeTraining scale sIn-scaleOut-of-scale
×2×3×4×6×8
Discretes ∈ {2}33.1327.0125.6022.2722.09
s ∈ {3}31.3929.0625.7723.4422.16
s ∈ {4}31.4227.8726.8824.2822.85
s ∈ {2,3,4}33.1529.1427.0224.4723.03
Continuouss ∈ [1,4]33.3029.1727.1124.5823.13
+ +Table 7. Comparisons of model size, inference time and performance gain of different models. + +
Different modelsMeta-SR [28]LIIF [13]ITSRN [79]LTE [39]CiaoSR
Model size (M)1.71.60.71.71.4
Inference time (ms)237171343148528
PSNR (dB)26.5526.6826.7726.8127.11
Performance gain (dB)-0.060.070.160.20.5
+ +Table 8. Comparison of more evaluation metrics (SSIM↑ [76] and LPIPS↓ [85]) for different methods on Urban100. + +
RDN-MethodIn-scaleOut-of-scale
×2×3×4×6×8
SSIMLPIPSSSIMLPIPSSSIMLPIPSSSIMLPIPSSSIMLPIPS
LIIF [13]0.9340.1040.8660.1970.8040.2670.7040.3600.6360.425
MetaSR [28]0.9350.1040.8660.1950.8050.2610.6980.3530.6240.423
ITSRN [79]0.9360.1020.8680.1930.8070.2590.7030.3500.6330.417
LTE [39]0.9360.1040.8680.1950.8070.2640.7070.3560.6390.421
CiaoSR0.9380.1000.8710.1880.8140.2550.7180.3460.6510.408
+ +Table 9. Quantitative comparisons of different Realworld SR methods on RealSRSet [82]. + +
MethodsScaleNIQE↓BRISQUE↓PIQE↓
RealSR [5]×84.794923.322829.3976
×163.394621.639021.4116
BSRGAN [82]×84.700737.063836.1889
×164.140839.625237.9176
Real-ESRGAN [74]×84.528040.652159.8915
×165.644253.936484.3611
Ours-real×83.589430.921929.8554
×163.946141.814558.8244
+ +# 5.4. Real-World Arbitrary-Scale Image SR + +We extend our proposed method to real-world applications with the ultimate goal of synthesizing real images with arbitrary scales. Specifically, we train our model on the DIV2K dataset [1], and test on the RealSRSet [82] and DPED [30] datasets. We apply the practical degradation models (including blur, noise, compression, etc.) of BSRGAN [82] and Real-ESRGAN [74] to synthesize paired training data on DIV2K [1]. We set the HR patch size of training data as 256, and the batch size as 48. We use Adam [36] as the optimizer, and adopt the exponential moving average (EMA). We first train our model with L1 loss for 1000K iterations with the learning rate of $2 \times 10^{-4}$ and train with L1 loss, perceptual loss and GAN loss together for 400K iterations with the learning rate of $1 \times 10^{-4}$ . More training details can be found in the supplementary materials. + +Since there are no arbitrary-scale SR methods applied to real-world images, we mainly compare with the state-of-the-art real-world image SR models, including RealSR [5], BSRGAN [82] and Real-ESRGAN [74]. Because there are no ground-truth (GT) real images, we apply non-reference image quality assessment (e.g., NIQE [54], BRISQUE [53] and PIQE [69]) to evaluate the image quality. As shown in Table 9, our model does not optimize for all metrics since these metrics cannot well match the actual human perceptual [4]. To address this, we further compare the visual results on the RealSRSet [82] and DPED [30] datasets in Figure 6. Our proposed method is able to synthesize real + +SR images with realistic textures. In particular, our method is able to synthesize more details of a dog in the first line and more clear digit '5' in the second line, compared with other methods. In contrast, other methods often introduce unnatural artifacts within images, especially for out-of-scale distributions. To further investigate the effectiveness of our model, we provide more visual results of our real model in the supplementary materials. + +# 6. Conclusion + +In this paper, we have proposed a novel continuous implicit attention-in-attention network for arbitrary-scale image super-resolution, called CiaoSR. Specifically, our CiaoSR first introduces a scale-aware non-local attention in our attention to exploit more relevant features, and then predicts RGB values by locally ensembling features with the learnable attention weights. More importantly, CiaoSR has good flexibility and applicability since it can be used behind any SR backbone to boost the performance. Extensive experiments on all benchmark datasets demonstrate the superiority of our CiaoSR in both SISR and arbitrary-scale SR tasks. Besides, our CiaoSR has good generalization performance on both in-scale and out-of-scale distributions. Last, we extend our method to the real-world application to synthesize real SR images with arbitrary scale. + +Acknowledgements: This work was partly supported by the Huawei Fund, the ETH Zürich General Fund (OK) and the Alexander von Humboldt Foundation. + +# References + +[1] Eirikur Agustsson and Radu Timofte. Ntire 2017 challenge on single image super-resolution: Dataset and study. In IEEE Conference on Computer Vision and Pattern Recognition Workshops, pages 126-135, 2017. 5, 8 +[2] Matan Atzmon and Yaron Lipman. Sal: Sign agnostic learning of shapes from raw data. In IEEE Conference on Computer Vision and Pattern Recognition, pages 2565-2574, 2020. 2 +[3] Marco Bevilacqua, Aline Roumy, Christine Guillemot, and Marie line Alberi Morel. Low-complexity single-image super-resolution based on nonnegative neighbor embedding. In British Machine Vision Conference, pages 135.1-135.10, 2012. 5, 6 +[4] Yochai Blau, Roey Mechrez, Radu Timofte, Tomer Michaeli, and Lihi Zelnik-Manor. The 2018 pirm challenge on perceptual image super-resolution. In European Conference on Computer Vision Workshops, 2018. 8 +[5] Jianrui Cai, Hui Zeng, Hongwei Yong, Zisheng Cao, and Lei Zhang. Toward real-world single image super-resolution: A new benchmark and a new model. In IEEE Conference on International Conference on Computer Vision, pages 3086-3095, 2019. 8 +[6] Jiezhang Cao, Yawei Li, Kai Zhang, and Luc Van Gool. Video super-resolution transformer. arXiv preprint arXiv:2106.06847, 2021. 2 +[7] Jiezhang Cao, Jingyun Liang, Kai Zhang, Yawei Li, Yulun Zhang, Wenguan Wang, and Luc Van Gool. Reference-based image super-resolution with deformable attention transformer. In European Conference on Computer Vision, pages 325-342, 2022. 2 +[8] Jiezhang Cao, Jingyun Liang, Kai Zhang, Wenguan Wang, Qin Wang, Yulun Zhang, Hao Tang, and Luc Van Gool. Towards interpretable video super-resolution via alternating optimization. In European Conference on Computer Vision, pages 393-411, 2022. 2 +[9] Lukas Cavigelli, Pascal Hager, and Luca Benini. Cas-cnn: A deep convolutional neural network for image compression artifact suppression. In 2017 International Joint Conference on Neural Networks, pages 752-759, 2017. 2 +[10] Rohan Chabra, Jan E Lenssen, Eddy Ig, Tanner Schmidt, Julian Straub, Steven Lovegrove, and Richard Newcombe. Deep local shapes: Learning local sdf priors for detailed 3d reconstruction. In European Conference on Computer Vision, pages 608-625, 2020. 2 +[11] Hanting Chen, Yunhe Wang, Tianyu Guo, Chang Xu, Yiping Deng, Zhenhua Liu, Siwei Ma, Chunjing Xu, Chao Xu, and Wen Gao. Pre-trained image processing transformer. In IEEE Conference on Computer Vision and Pattern Recognition, pages 12299-12310, 2021. 2 +[12] Xiangyu Chen, Xintao Wang, Jiantao Zhou, and Chao Dong. Activating more pixels in image super-resolution transformer. arXiv preprint arXiv:2205.04437, 2022. 2 +[13] Yinbo Chen, Sifei Liu, and Xiaolong Wang. Learning continuous image representation with local implicit image function. In IEEE Conference on Computer Vision and Pattern Recognition, pages 8628-8638, 2021. 1, 2, 3, 4, 5, 6, 7, 8 + +[14] Yunjin Chen and Thomas Pock. Trainable nonlinear reaction diffusion: A flexible framework for fast and effective image restoration. IEEE Transactions on Pattern Analysis and Machine Intelligence, 39(6):1256-1272, 2016. 2 +[15] Zhiqin Chen and Hao Zhang. Learning implicit fields for generative shape modeling. In IEEE Conference on Computer Vision and Pattern Recognition, pages 5939-5948, 2019. 2 +[16] Wenlong Cheng, Mingbo Zhao, Zhiling Ye, and Shuhang Gu. Mfagan: A compression framework for memory-efficient on-device super-resolution gan. arXiv preprint arXiv:2107.12679, 2021. 2 +[17] Tao Dai, Jianrui Cai, Yongbing Zhang, Shu-Tao Xia, and Lei Zhang. Second-order attention network for single image super-resolution. In IEEE Conference on Computer Vision and Pattern Recognition, pages 11065-11074, 2019. 2 +[18] Xin Deng, Yutong Zhang, Mai Xu, Shuhang Gu, and Yiping Duan. Deep coupled feedback network for joint exposure fusion and image super-resolution. IEEE Transactions on Image Processing, 30:3098-3112, 2021. 2 +[19] Chao Dong, Chen Change Loy, Kaiming He, and Xiaou Tang. Learning a deep convolutional network for image super-resolution. In European Conference on Computer Vision, pages 184–199, 2014. 2 +[20] Xueyang Fu, Menglu Wang, Xiangyong Cao, Xinghao Ding, and Zheng-Jun Zha. A model-driven deep unfolding method for jpeg artifacts removal. IEEE Transactions on Neural Networks and Learning Systems, 2021. 2 +[21] Xueyang Fu, Zheng-Jun Zha, Feng Wu, Xinghao Ding, and John Paisley. Jpeg artifacts reduction via deep convolutional sparse coding. In IEEE International Conference on Computer Vision, pages 2501-2510, 2019. 2 +[22] Amos Gropp, Lior Yariv, Niv Haim, Matan Atzmon, and Yaron Lipman. Implicit geometric regularization for learning shapes. arXiv preprint arXiv:2002.10099, 2020. 2 +[23] Shuhang Gu, Nong Sang, and Fan Ma. Fast image super resolution via local regression. In IEEE Conference on International Conference on Pattern Recognition, pages 3128-3131, 2012. 2 +[24] Bahadir K Gunturk, Yucel Altunbasak, and Russell M Mersereau. Super-resolution reconstruction of compressed video using transform-domain statistics. IEEE Transactions on Image Processing, 13(1):33-43, 2004. 1 +[25] Yong Guo, Jian Chen, Jingdong Wang, Qi Chen, Jiezhang Cao, Zeshuai Deng, Yanwu Xu, and Mingkui Tan. Closed-loop matters: Dual regression networks for single image super-resolution. In IEEE Conference on Computer Vision and Pattern Recognition, pages 5407-5416, 2020. 2 +[26] Ali Hassani, Steven Walton, Jiachen Li, Shen Li, and Humphrey Shi. Neighborhood attention transformer. arXiv preprint arXiv:2204.07143, 2022. 3 +[27] Kaiming He, Jian Sun, and Xiaou Tang. Single image haze removal using dark channel prior. IEEE Transactions on Pattern Analysis and Machine Intelligence, 33(12):2341-2353, 2010. 2 +[28] Xuecai Hu, Haoyuan Mu, Xiangyu Zhang, Zilei Wang, Tieniu Tan, and Jian Sun. Meta-sr: A magnification-arbitrary network for super-resolution. In IEEE Conference on Computer + +Vision and Pattern Recognition, pages 1575-1584, 2019. 1, 2, 5, 6, 7, 8 +[29] Jia-Bin Huang, Abhishek Singh, and Narendra Ahuja. Single image super-resolution from transformed self-exemplars. In IEEE Conference on Computer Vision and Pattern Recognition, pages 5197-5206, 2015. 5, 6, 7, 8 +[30] Andrey Ignatov, Nikolay Kobyshev, Radu Timofte, Kenneth Vanhoey, and Luc Van Gool. Dslr-quality photos on mobile devices with deep convolutional networks. In IEEE International Conference on Computer Vision, pages 3277-3285, 2017. 7, 8 +[31] Takashi Isobe, Xu Jia, Shuhang Gu, Songjiang Li, Shengjin Wang, and Qi Tian. Video super-resolution with recurrent structure-detail network. In European Conference on Computer Vision, pages 645-660, 2020. 2 +[32] Xixi Jia, Sanyang Liu, Xiangchu Feng, and Lei Zhang. Focnet: A fractional optimal control network for image denoising. In IEEE Conference on Computer Vision and Pattern Recognition, pages 6054-6063, 2019. 2 +[33] Chiyu Jiang, Avneesh Sud, Ameesh Makadia, Jingwei Huang, Matthias Nießner, Thomas Funkhouser, et al. Local implicit grid representations for 3d scenes. In IEEE Conference on Computer Vision and Pattern Recognition, pages 6001-6010, 2020. 2 +[34] Jiwon Kim, Jung Kwon Lee, and Kyoung Mu Lee. Accurate image super-resolution using very deep convolutional networks. In IEEE Conference on Computer Vision and Pattern Recognition, pages 1646-1654, 2016. 2, 6 +[35] Yoonsik Kim, Jae Woong Soh, Jaewoo Park, Byeongyong Ahn, Hyun-Seung Lee, Young-Su Moon, and Nam Ik Cho. A pseudo-blind convolutional neural network for the reduction of compression artifacts. IEEE Transactions on Circuits and Systems for Video Technology, 30(4):1121-1135, 2019. 2 +[36] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In International Conference on Learning Representations, 2015. 5, 8 +[37] Wei-Sheng Lai, Jia-Bin Huang, Narendra Ahuja, and Ming-Hsuan Yang. Deep laplacian pyramid networks for fast and accurate super-resolution. In IEEE Conference on Computer Vision and Pattern Recognition, pages 624-632, 2017. 2 +[38] Christian Ledig, Lucas Theis, Ferenc Huszar, Jose Caballero, Andrew Cunningham, Alejandro Acosta, Andrew Aitken, Alykhan Tejani, Johannes Totz, Zehan Wang, et al. Photorealistic single image super-resolution using a generative adversarial network. In IEEE Conference on Computer Vision and Pattern Recognition, pages 4681-4690, 2017. 2 +[39] Jaewon Lee and Kyong Hwan Jin. Local texture estimator for implicit representation function. In IEEE Conference on Computer Vision and Pattern Recognition, pages 1929-1938, 2022. 1, 2, 3, 4, 5, 6, 7, 8 +[40] Jingyun Liang, Jiezhang Cao, Guolei Sun, Kai Zhang, Luc Van Gool, and Radu Timofte. Swinir: Image restoration using swin transformer. In IEEE International Conference on Computer Vision Workshops, pages 1833-1844, 2021. 1, 2, 3, 5, 6 +[41] Jingyun Liang, Andreas Lugmayr, Kai Zhang, Martin Danelljan, Luc Van Gool, and Radu Timofte. Hierarchical conditional flow: A unified framework for image super-resolution + +and image rescaling. In IEEE Conference on International Conference on Computer Vision, 2021. 2 +[42] Jingyun Liang, Guolei Sun, Kai Zhang, Luc Van Gool, and Radu Timofte. Mutual affine network for spatially variant kernel estimation in blind image super-resolution. In IEEE Conference on International Conference on Computer Vision, 2021. 2 +[43] Jingyun Liang, Kai Zhang, Shuhang Gu, Luc Van Gool, and Radu Timofte. Flow-based kernel prior with application to blind super-resolution. In IEEE Conference on Computer Vision and Pattern Recognition, pages 10601-10610, 2021. 2 +[44] Bee Lim, Sanghyun Son, Heewon Kim, Seungjun Nah, and Young Mu Lee. Enhanced deep residual networks for single image super-resolution. In IEEE Conference on Computer Vision and Pattern Recognition Workshops, pages 136-144, 2017. 1, 2, 5, 6 +[45] Ding Liu, Bihan Wen, Yuchen Fan, Chen Change Loy, and Thomas S Huang. Non-local recurrent network for image restoration. arXiv preprint arXiv:1806.02919, 2018. 2 +[46] David Martin, Charless Fowlkes, Doron Tal, and Jitendra Malik. A database of human segmented natural images and its application to evaluating segmentation algorithms and measuring ecological statistics. In IEEE Conference on International Conference on Computer Vision, pages 416-423, 2001. 5, 6 +[47] Yusuke Matsui, Kota Ito, Yuji Aramaki, Azuma Fujimoto, Toru Ogawa, Toshihiko Yamasaki, and Kiyoharu Aizawa. Sketch-based manga retrieval using manga109 dataset. Multi-media Tools and Applications, 76(20):21811-21838, 2017. 5, 6, 7, 8 +[48] Yiqun Mei, Yuchen Fan, and Yuqian Zhou. Image superresolution with non-local sparse attention. In IEEE Conference on Computer Vision and Pattern Recognition, pages 3517-3526, 2021. 2 +[49] Yiqun Mei, Yuchen Fan, Yuqian Zhou, Lichao Huang, Thomas S Huang, and Honghui Shi. Image super-resolution with cross-scale non-local attention and exhaustive self-exemplars mining. In IEEE Conference on Computer Vision and Pattern Recognition, pages 5690-5699, 2020. 4 +[50] Tomer Michaeli and Michal Irani. Nonparametric blind superresolution. In IEEE Conference on International Conference on Computer Vision, pages 945–952, 2013. 2 +[51] Mateusz Michalkiewicz, Jhony K Pontes, Dominic Jack, Mahsa Baktashmotlagh, and Anders Eriksson. Implicit surface representations as layers in neural networks. In IEEE International Conference on Computer Vision, pages 4743-4752, 2019. 2 +[52] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 2 +[53] Anish Mittal, Anush K Moorthy, and Alan C Bovik. Blind/referenceless image spatial quality evaluator. In Asilomar Conference on Signals, Systems and Computers, pages 723-727, 2011. 8 +[54] Anish Mittal, Rajiv Soundararajan, and Alan C Bovik. Making a "completely blind" image quality analyzer. IEEE Signal Processing Letters, 20(3):209-212, 2012. 8 + +[55] Michael Niemeyer, Lars Mescheder, Michael Oechsle, and Andreas Geiger. Differentiable volumetric rendering: Learning implicit 3d representations without 3d supervision. In IEEE Conference on Computer Vision and Pattern Recognition, pages 3504-3515, 2020. 2 +[56] Ben Niu, Weilei Wen, Wenqi Ren, Xiangde Zhang, Lianping Yang, Shuzhen Wang, Kaihao Zhang, Xiaochun Cao, and Haifeng Shen. Single image super-resolution via a holistic attention network. In European Conference on Computer Vision, pages 191-207, 2020. 2 +[57] Michael Oechsle, Lars Mescheder, Michael Niemeyer, Thilo Strauss, and Andreas Geiger. Texture fields: Learning texture representations in function space. In IEEE International Conference on Computer Vision, pages 4531-4540, 2019. 2 +[58] Songyou Peng, Michael Niemeyer, Lars Mescheder, Marc Pollefeys, and Andreas Geiger. Convolutional occupancy networks. In European Conference on Computer Vision, pages 523-540, 2020. 2 +[59] Yali Peng, Lu Zhang, Shigang Liu, Xiaojun Wu, Yu Zhang, and Xili Wang. Dilated residual networks with symmetric skip connection for image denoising. Neurocomputing, 345:67-76, 2019. 2 +[60] Wenzhe Shi, Jose Caballero, Ferenc Huszár, Johannes Totz, Andrew P Aitken, Rob Bishop, Daniel Rueckert, and Zehan Wang. Real-time single image and video super-resolution using an efficient sub-pixel convolutional neural network. In IEEE Conference on Computer Vision and Pattern Recognition, pages 1874-1883, 2016. 1, 3 +[61] Wenzhe Shi, Jose Caballero, Christian Ledig, Xiahai Zhuang, Wenjia Bai, Kanwal Bhatia, Antonio M Marvao, Tim Dawes, Declan O'Regan, and Daniel Rueckert. Cardiac image superresolution with global correspondence using multi-atlas patchmatch. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 9-16, 2013. 1 +[62] Vincent Sitzmann, Julien Martel, Alexander Bergman, David Lindell, and Gordon Wetzstein. Implicit neural representations with periodic activation functions. Advances in Neural Information Processing Systems, 33:7462-7473, 2020. 2 +[63] Vincent Sitzmann, Michael Zollhöfer, and Gordon Wetzstein. Scene representation networks: Continuous 3d-structure-aware neural scene representations. Advances in Neural Information Processing Systems, 32, 2019. 2 +[64] Sanghyun Son and Kyoung Mu Lee. Srwarp: Generalized image super-resolution under arbitrary transformation. In IEEE Conference on Computer Vision and Pattern Recognition, pages 7782-7791, 2021. 2 +[65] Ying Tai, Jian Yang, Xiaoming Liu, and Chunyan Xu. Memnet: A persistent memory network for image restoration. In IEEE International Conference on Computer Vision, pages 4539-4547, 2017. 2 +[66] Radu Timofte, Vincent De Smet, and Luc Van Gool. Anchored neighborhood regression for fast example-based superresolution. In IEEE Conference on International Conference on Computer Vision, pages 1920-1927, 2013. 2 +[67] Radu Timofte, Vincent De Smet, and Luc Van Gool. A+: Adjusted anchored neighborhood regression for fast super- + +resolution. In Asian Conference on Computer Vision, pages 111-126, 2014. 2 +[68] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Advances in Neural Information Processing Systems, 2017. 2, 3, 4 +[69] N Venkatanath, D Praneeth, Maruthi Chandrasekhar Bh, Sumohana S Channappayya, and Swarup S Medasani. Blind image quality evaluation using perception based features. In National Conference on Communications, pages 1-6, 2015. 8 +[70] Longguang Wang, Yingqian Wang, Xiaoyu Dong, Qingyu Xu, Jungang Yang, Wei An, and Yulan Guo. Unsupervised degradation representation learning for blind super-resolution. In IEEE Conference on Computer Vision and Pattern Recognition, pages 10581-10590, 2021. 2 +[71] Longguang Wang, Yingqian Wang, Zhengfa Liang, Zaiping Lin, Jungang Yang, Wei An, and Yulan Guo. Learning parallel attention for stereo image super-resolution. In IEEE Conference on Computer Vision and Pattern Recognition, pages 12250-12259, 2019. 2 +[72] Longguang Wang, Yingqian Wang, Zaiping Lin, Jungang Yang, Wei An, and Yulan Guo. Learning a single network for scale-arbitrary super-resolution. In IEEE Conference on International Conference on Computer Vision, pages 10581-10590, 2021. 2 +[73] Longguang Wang, Yingqian Wang, Zaiping Lin, Jungang Yang, Wei An, and Yulan Guo. Learning a single network for scale-arbitrary super-resolution. In IEEE International Conference on Computer Vision, pages 4801-4810, 2021. 2 +[74] Xintao Wang, Liangbin Xie, Chao Dong, and Ying Shan. Real-esrgan: Training real-world blind super-resolution with pure synthetic data. In International Conference on Computer Vision Workshops. 8 +[75] Xintao Wang, Ke Yu, Shixiang Wu, Jinjin Gu, Yihao Liu, Chao Dong, Yu Qiao, and Chen Change Loy. Esrgan: Enhanced super-resolution generative adversarial networks. In European Conference on Computer Vision Workshops, pages 701-710, 2018. 2 +[76] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE Transactions on Image Processing, 13(4):600-612, 2004. 5, 7, 8 +[77] Yunxuan Wei, Shuhang Gu, Yawei Li, Radu Timofte, Longcun Jin, and Hengjie Song. Unsupervised real-world image super resolution via domain-distance aware training. In IEEE Conference on Computer Vision and Pattern Recognition, pages 13385-13394, 2021. 2 +[78] Xingqian Xu, Zhangyang Wang, and Humphrey Shi. Ultrasr: Spatial encoding is a missing key for implicit image function-based arbitrary-scale super-resolution. arXiv preprint arXiv:2103.12716, 2021. 2 +[79] Jingyu Yang, Sheng Shen, Huanjing Yue, and Kun Li. Implicit transformer network for screen content image continuous super-resolution. Advances in Neural Information Processing Systems, 34:13304-13315, 2021. 2, 3, 5, 6, 7, 8 +[80] Roman Zeyde, Michael Elad, and Matan Protter. On single image scale-up using sparse-representations. In International + +Conference on Curves and Surfaces, pages 711-730, 2010. 5, 6 +[81] Kai Zhang, Yawei Li, Wangmeng Zuo, Lei Zhang, Luc Van Gool, and Radu Timofte. Plug-and-play image restoration with deep denoiser prior. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2021. 2 +[82] Kai Zhang, Jingyun Liang, Luc Van Gool, and Radu Timofte. Designing a practical degradation model for deep blind image super-resolution. In IEEE Conference on International Conference on Computer Vision, 2021. 7, 8 +[83] Kai Zhang, Wangmeng Zuo, and Lei Zhang. Ffdnet: Toward a fast and flexible solution for cnn-based image denoising. IEEE Transactions on Image Processing, 27(9):4608-4622, 2018. 2 +[84] Kai Zhang, Wangmeng Zuo, and Lei Zhang. Learning a single convolutional super-resolution network for multiple degradations. In IEEE Conference on Computer Vision and Pattern Recognition, pages 3262-3271, 2018. 2 +[85] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In IEEE Conference on Computer Vision and Pattern Recognition, pages 586-595, 2018. 5, 7, 8 +[86] Yulun Zhang, Kunpeng Li, Kai Li, Lichen Wang, Bineng Zhong, and Yun Fu. Image super-resolution using very deep residual channel attention networks. In European Conference on Computer Vision, pages 286-301, 2018. 1, 2, 3 +[87] Yulun Zhang, Kunpeng Li, Kai Li, Bineng Zhong, and Yun Fu. Residual non-local attention networks for image restoration. arXiv preprint arXiv:1903.10082, 2019. 2 +[88] Yulun Zhang, Yapeng Tian, Yu Kong, Bineng Zhong, and Yun Fu. Residual dense network for image super-resolution. In IEEE Conference on Computer Vision and Pattern Recognition, pages 2472-2481, 2018. 1, 2, 5, 6, 7, 8 +[89] Yulun Zhang, Yapeng Tian, Yu Kong, Bineng Zhong, and Yun Fu. Residual dense network for image restoration. IEEE Transactions on Pattern Analysis and Machine Intelligence, 43(7):2480-2495, 2020. 2 +[90] Shangchen Zhou, Jiawei Zhang, Wangmeng Zuo, and Chen Change Loy. Cross-scale internal graph neural network for image super-resolution. arXiv preprint arXiv:2006.16673, 2020.2 +[91] Wilman WW Zou and Pong C Yuen. Very low resolution face recognition problem. IEEE Transactions on Image Processing, 21(1):327-340, 2011. 1 \ No newline at end of file diff --git a/2023/CiaoSR_ Continuous Implicit Attention-in-Attention Network for Arbitrary-Scale Image Super-Resolution/images.zip b/2023/CiaoSR_ Continuous Implicit Attention-in-Attention Network for Arbitrary-Scale Image Super-Resolution/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..8b4a825691097e9990d71e3ad1502b7ef4fcd23d --- /dev/null +++ b/2023/CiaoSR_ Continuous Implicit Attention-in-Attention Network for Arbitrary-Scale Image Super-Resolution/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bac879a17b7520bb953f09621e6f4e1be8493efd583b5b49274fc2b18c2b1cdb +size 1017733 diff --git a/2023/CiaoSR_ Continuous Implicit Attention-in-Attention Network for Arbitrary-Scale Image Super-Resolution/layout.json b/2023/CiaoSR_ Continuous Implicit Attention-in-Attention Network for Arbitrary-Scale Image Super-Resolution/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..995f78d8261d1c32861ea18b24272e523d0566bd --- /dev/null +++ b/2023/CiaoSR_ Continuous Implicit Attention-in-Attention Network for Arbitrary-Scale Image Super-Resolution/layout.json @@ -0,0 +1,11858 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 96, + 102, + 501, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 96, + 102, + 501, + 140 + ], + "spans": [ + { + "bbox": [ + 96, + 102, + 501, + 140 + ], + "type": "text", + "content": "CiaoSR: Continuous Implicit Attention-in-Attention Network for Arbitrary-Scale Image Super-Resolution" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 82, + 160, + 511, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 82, + 160, + 511, + 175 + ], + "spans": [ + { + "bbox": [ + 82, + 160, + 511, + 175 + ], + "type": "text", + "content": "Jiezhang Cao" + }, + { + "bbox": [ + 82, + 160, + 511, + 175 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 82, + 160, + 511, + 175 + ], + "type": "text", + "content": " Qin Wang" + }, + { + "bbox": [ + 82, + 160, + 511, + 175 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 82, + 160, + 511, + 175 + ], + "type": "text", + "content": " Yongqin Xian" + }, + { + "bbox": [ + 82, + 160, + 511, + 175 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 82, + 160, + 511, + 175 + ], + "type": "text", + "content": " Yawei Li" + }, + { + "bbox": [ + 82, + 160, + 511, + 175 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 82, + 160, + 511, + 175 + ], + "type": "text", + "content": " Bingbing Ni" + }, + { + "bbox": [ + 82, + 160, + 511, + 175 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 82, + 160, + 511, + 175 + ], + "type": "text", + "content": " Zhiming Pi" + }, + { + "bbox": [ + 82, + 160, + 511, + 175 + ], + "type": "inline_equation", + "content": "^{2}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 137, + 175, + 458, + 189 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 175, + 458, + 189 + ], + "spans": [ + { + "bbox": [ + 137, + 175, + 458, + 189 + ], + "type": "text", + "content": "Kai Zhang" + }, + { + "bbox": [ + 137, + 175, + 458, + 189 + ], + "type": "inline_equation", + "content": "^{1\\dagger}" + }, + { + "bbox": [ + 137, + 175, + 458, + 189 + ], + "type": "text", + "content": " Yulun Zhang" + }, + { + "bbox": [ + 137, + 175, + 458, + 189 + ], + "type": "inline_equation", + "content": "^{1\\dagger}" + }, + { + "bbox": [ + 137, + 175, + 458, + 189 + ], + "type": "text", + "content": " Radu Timofte" + }, + { + "bbox": [ + 137, + 175, + 458, + 189 + ], + "type": "inline_equation", + "content": "^{1,3}" + }, + { + "bbox": [ + 137, + 175, + 458, + 189 + ], + "type": "text", + "content": " Luc Van Gool" + }, + { + "bbox": [ + 137, + 175, + 458, + 189 + ], + "type": "inline_equation", + "content": "^{1,4}" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 126, + 189, + 468, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 126, + 189, + 468, + 203 + ], + "spans": [ + { + "bbox": [ + 126, + 189, + 468, + 203 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 126, + 189, + 468, + 203 + ], + "type": "text", + "content": "ETH Zürich " + }, + { + "bbox": [ + 126, + 189, + 468, + 203 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 126, + 189, + 468, + 203 + ], + "type": "text", + "content": "Huawei Inc. " + }, + { + "bbox": [ + 126, + 189, + 468, + 203 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 126, + 189, + 468, + 203 + ], + "type": "text", + "content": "University of Wurzburg " + }, + { + "bbox": [ + 126, + 189, + 468, + 203 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 126, + 189, + 468, + 203 + ], + "type": "text", + "content": "KU Leuven" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 194, + 205, + 398, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 194, + 205, + 398, + 217 + ], + "spans": [ + { + "bbox": [ + 194, + 205, + 398, + 217 + ], + "type": "text", + "content": "https://github.com/caojiezhang/CiaoSR" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 143, + 244, + 192, + 258 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 244, + 192, + 258 + ], + "spans": [ + { + "bbox": [ + 143, + 244, + 192, + 258 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 270, + 290, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 270, + 290, + 559 + ], + "spans": [ + { + "bbox": [ + 46, + 270, + 290, + 559 + ], + "type": "text", + "content": "Learning continuous image representations is recently gaining popularity for image super-resolution (SR) because of its ability to reconstruct high-resolution images with arbitrary scales from low-resolution inputs. Existing methods mostly ensemble nearby features to predict the new pixel at any queried coordinate in the SR image. Such a local ensemble suffers from some limitations: i) it has no learnable parameters and it neglects the similarity of the visual features; ii) it has a limited receptive field and cannot ensemble relevant features in a large field which are important in an image. To address these issues, this paper proposes a continuous implicit attention-in-attention network, called CiaoSR. We explicitly design an implicit attention network to learn the ensemble weights for the nearby local features. Furthermore, we embed a scale-aware attention in this implicit attention network to exploit additional non-local information. Extensive experiments on benchmark datasets demonstrate CiaoSR significantly outperforms the existing single image SR methods with the same backbone. In addition, CiaoSR also achieves the state-of-the-art performance on the arbitrary-scale SR task. The effectiveness of the method is also demonstrated on the real-world SR setting. More importantly, CiaoSR can be flexibly integrated into any backbone to improve the SR performance." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 566, + 128, + 578 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 566, + 128, + 578 + ], + "spans": [ + { + "bbox": [ + 47, + 566, + 128, + 578 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 580, + 288, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 580, + 288, + 676 + ], + "spans": [ + { + "bbox": [ + 46, + 580, + 288, + 676 + ], + "type": "text", + "content": "Single image super-resolution (SISR), which aims to reconstruct a high-resolution (HR) image from a low-resolution (LR) one, has been widely employed in many practical applications [24, 61, 91]. However, deep neural networks (DNN)-based SISR methods are facing some limitations in some real-world scenarios with arbitrary scales. For example, camera users may want to enhance the digital zoom quality by super-resolving a photo or a video to" + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 306, + 244, + 545, + 321 + ], + "blocks": [ + { + "bbox": [ + 306, + 244, + 545, + 321 + ], + "lines": [ + { + "bbox": [ + 306, + 244, + 545, + 321 + ], + "spans": [ + { + "bbox": [ + 306, + 244, + 545, + 321 + ], + "type": "image", + "image_path": "ff01524b6deb00056d2d082586a728943e8e013d0bdd1e1107b95a47d23c9250.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 323, + 547, + 357 + ], + "lines": [ + { + "bbox": [ + 305, + 323, + 547, + 357 + ], + "spans": [ + { + "bbox": [ + 305, + 323, + 547, + 357 + ], + "type": "text", + "content": "Figure 1. Comparison of different backbones and implicit models. Our proposed implicit neural network on RDN [88] has better performance than SwinIR [40]. Check Section 5 for details." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 363, + 547, + 448 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 363, + 547, + 448 + ], + "spans": [ + { + "bbox": [ + 304, + 363, + 547, + 448 + ], + "type": "text", + "content": "continuous arbitrary scales. Most existing DNN-based SISR methods [40, 44, 86] need to train a series of models for all different scales separately. However, it can be impractical to store all these models on the device due to limited storage and computing power. Alternatively, arbitrary-scale image SR methods [13, 28, 39] aim to train a single network for all scales in a continuous manner." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 449, + 548, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 449, + 548, + 664 + ], + "spans": [ + { + "bbox": [ + 304, + 449, + 548, + 664 + ], + "type": "text", + "content": "Most existing SISR methods [40,44,86] consist of a DNN and an upsampling module (e.g., pixel shuffling [60]) at a discrete scale. While substantial progresses have been made in the DNN backbones for SR, there is little attempt to study the upsampling module. A natural question to ask is: Does the pixel shuffling hinder the potential of SR models? One limitation of the pixel shuffling module is that it cannot synthesize SR images at large unseen and continuous scales. To tackle this, one can treat synthesizing different-scale SR images as a multi-task learning problem, and train a specific upsampling module for each scale [44]. However, these tasks are dependent and highly inter-related. Neglecting the correlation of different-scale SR tasks may lead to discrete representations and limited performance. Under a certain capacity of a network, training a model on multi-tasks may sacrifice the performance or have the comparable performance on each task. These above disadvantages limit its applicability and flexibility in the real-world scenarios." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 665, + 548, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 548, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 548, + 715 + ], + "type": "text", + "content": "To address these, most existing arbitrary-scale SR methods [13, 28, 39] replace the upsampling method with an implicit neural function and boost the performance. These methods predict an RGB value at the query point in an image" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 683, + 256, + 693 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 683, + 256, + 693 + ], + "spans": [ + { + "bbox": [ + 58, + 683, + 256, + 693 + ], + "type": "text", + "content": "*Currently with Google. This work was done at ETH Zürich." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 47, + 693, + 287, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 693, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 693, + 287, + 713 + ], + "type": "text", + "content": "†Corresponding Authors: Kai Zhang, cskaizhang@gmail.com; Yulun Zhang, yulun100@gmail.com" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "1796" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 82, + 71, + 511, + 166 + ], + "blocks": [ + { + "bbox": [ + 82, + 71, + 511, + 166 + ], + "lines": [ + { + "bbox": [ + 82, + 71, + 511, + 166 + ], + "spans": [ + { + "bbox": [ + 82, + 71, + 511, + 166 + ], + "type": "image", + "image_path": "eccbf2d2b3a66c58a8bff1e822218e41c9a651de862409ff0b1be2eabb7b8bdb.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 167, + 547, + 222 + ], + "lines": [ + { + "bbox": [ + 46, + 167, + 547, + 222 + ], + "spans": [ + { + "bbox": [ + 46, + 167, + 547, + 222 + ], + "type": "text", + "content": "Figure 2. Comparisons of different attention mechanisms. (a) Self-attention can predict pixel features on the grid, but it cannot be directly used in arbitrary-scale SR without considering coordinates. (b) Most existing methods can be treated as coordinate-based implicit attention since they calculate the distance between a key and query coordinate, and then use a function " + }, + { + "bbox": [ + 46, + 167, + 547, + 222 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 46, + 167, + 547, + 222 + ], + "type": "text", + "content": " to aggregate with the value features. However, these methods ignore the distance between the features. (c) Our implicit attention not only considers the coordinate distance, but also the distance among features with visual information." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 229, + 289, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 229, + 289, + 396 + ], + "spans": [ + { + "bbox": [ + 46, + 229, + 289, + 396 + ], + "type": "text", + "content": "by ensembling features within a local region. However, the local ensemble methods have limitations in the ensemble weights and insufficient information (e.g., non-local information). The ensemble weights are often calculated by the area of the rectangle between the query point and each nearest point, which is equivalent to the bilinear interpolation. Thus, those methods cannot adaptively ensemble local features since there is no trainable parameter. These weights are only related to the coordinates of the local features, but independent of the local features. Ignoring both the coordinates and the local features lose visual information and result in blurry artifacts. It is important and necessary to design a new implicit network to predict the weights and exploit more information in the local ensemble." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 397, + 290, + 505 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 397, + 290, + 505 + ], + "spans": [ + { + "bbox": [ + 46, + 397, + 290, + 505 + ], + "type": "text", + "content": "In this paper, we propose a novel implicit attention model to enhance arbitrary-scale SR performance. Specifically, we use our attention to predict the ensemble weights by considering both the similarity and coordinate distance of local features, as shown in Figure 2. Based on such learnable weights, the implicit model can adaptively aggregate local features according to different inputs. To enrich more information, we introduce an attention in our implicit attention, which helps discover more features in a larger receptive field." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 59, + 506, + 244, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 506, + 244, + 517 + ], + "spans": [ + { + "bbox": [ + 59, + 506, + 244, + 517 + ], + "type": "text", + "content": "Our contributions are summarized as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 523, + 288, + 701 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 47, + 523, + 287, + 583 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 523, + 287, + 583 + ], + "spans": [ + { + "bbox": [ + 47, + 523, + 287, + 583 + ], + "type": "text", + "content": "- We propose a novel continuous implicit attention-inattention network for arbitrary-scale image SR, called CiaoSR. Different from most existing local ensemble methods, our method explicitly learns the ensemble weights and exploits scale-aware non-local information." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 588, + 287, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 588, + 287, + 624 + ], + "spans": [ + { + "bbox": [ + 47, + 588, + 287, + 624 + ], + "type": "text", + "content": "- Our CiaoSR can be flexibly integrated into any backbone, allowing the network to super-resolve an image at arbitrary scales and improve the SR performance in Figure 1." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 629, + 288, + 701 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 629, + 288, + 701 + ], + "spans": [ + { + "bbox": [ + 47, + 629, + 288, + 701 + ], + "type": "text", + "content": "- Extensive experiments demonstrate CiaoSR achieves the state-of-the-art performance in both SISR and arbitrary-scale SR tasks. Besides, our CiaoSR has good generalization on both in-scale and out-of-scale distributions. Last, we extend our method to real-world SR settings to synthesize arbitrary-scale images." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 228, + 392, + 240 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 228, + 392, + 240 + ], + "spans": [ + { + "bbox": [ + 306, + 228, + 392, + 240 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 243, + 548, + 435 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 243, + 548, + 435 + ], + "spans": [ + { + "bbox": [ + 304, + 243, + 548, + 435 + ], + "type": "text", + "content": "Single image super-resolution (SISR). SISR aims to synthesize high-resolution (HR) images from low-resolution (LR) images. Compared with DNN-based SR methods [6-8, 23, 27, 50, 66, 67], methods in recent years build on deep convolutional neural network (CNN) to improve the performance, such as SRCNN [19], SRResNet [38], EDSR [44], RDN [88] and RCAN [86]. To further improve SR performance, some methods design CNN with residual block [9, 34, 81], dense block [75, 88, 89] and others [14, 16, 18, 20, 21, 25, 31, 32, 35, 37, 41-43, 59, 65, 70-72, 77, 83, 84, 87, 90]. In addition, some SR methods are built based on attention mechanism [68], such as channel attention [17, 56, 86], self-attention (IPT [11] and SwinIR [40], HAT [12]), non-local attention [45, 48]. However, most methods focus on one specific scale, which limits the applicability and flexibility in arbitrary-scale." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 438, + 548, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 438, + 548, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 438, + 548, + 715 + ], + "type": "text", + "content": "Arbitrary-scale super-resolution. To tackle this problem, very recently, the more practical setup of arbitrary-scale SR is considered, which aims to super-resolute images with arbitrary scales by a single model. MetaSR [28] makes the first attempt to propose an arbitrary-scale meta-upscale module. To improve the performance, many arbitrary-scale methods [64, 73] are proposed. With the help of implicit neural representation [2, 10, 15, 22, 33, 51, 52, 55, 57, 58, 62, 63], LIIF [13] predicts the RGB value at an arbitrary query coordinate by taking an image coordinate and features of backbone around the coordinate. The features are extracted by single image super-resolution methods, e.g., EDSR [44], RDN [88] and SwinIR [40]. To improve the performance, existing methods [39, 78] propose to integrate more features in SR models. For example, LTE [39] proposes a local texture estimator by characterizing image textures in the Fourier space. UltraSR [78] integrates spatial coordinates and periodic encoding in the implicit network. These methods use the bilinear interpolation to ensemble nearby features. However, such an ensemble way has no learnable parameters. Recently, ITSRN [79] learns the weights by taking the coordinate distance and scale token into a mapping. Most methods learn the ensemble without the feature similarity." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1797" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 65, + 71, + 527, + 258 + ], + "blocks": [ + { + "bbox": [ + 65, + 71, + 527, + 258 + ], + "lines": [ + { + "bbox": [ + 65, + 71, + 527, + 258 + ], + "spans": [ + { + "bbox": [ + 65, + 71, + 527, + 258 + ], + "type": "image", + "image_path": "a5f7fd21f2f55c2745db7f466b043c09a66a1933fc853e94ab22e48dda0502fa.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 261, + 547, + 297 + ], + "lines": [ + { + "bbox": [ + 46, + 261, + 547, + 297 + ], + "spans": [ + { + "bbox": [ + 46, + 261, + 547, + 297 + ], + "type": "text", + "content": "Figure 3. The architecture of our continuous implicit attention-in-attention network. Given an LR image, the encoder extracts features as latent codes. For a query point, we have a query feature and key features close to the query point, and the scale-aware non-local attention module extracts non-local features as value. Last, we use the triple of query, key and value to predict the RGB value at the query point." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 303, + 207, + 316 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 303, + 207, + 316 + ], + "spans": [ + { + "bbox": [ + 47, + 303, + 207, + 316 + ], + "type": "text", + "content": "3. Preliminary and Motivation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 320, + 288, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 320, + 288, + 380 + ], + "spans": [ + { + "bbox": [ + 46, + 320, + 288, + 380 + ], + "type": "text", + "content": "Let " + }, + { + "bbox": [ + 46, + 320, + 288, + 380 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 46, + 320, + 288, + 380 + ], + "type": "text", + "content": " be a continuous image, and " + }, + { + "bbox": [ + 46, + 320, + 288, + 380 + ], + "type": "inline_equation", + "content": "\\pmb{x}" + }, + { + "bbox": [ + 46, + 320, + 288, + 380 + ], + "type": "text", + "content": " be a 2D coordinate of a pixel in the image " + }, + { + "bbox": [ + 46, + 320, + 288, + 380 + ], + "type": "inline_equation", + "content": "\\pmb{I}" + }, + { + "bbox": [ + 46, + 320, + 288, + 380 + ], + "type": "text", + "content": ". Formally, given a 2D coordinate " + }, + { + "bbox": [ + 46, + 320, + 288, + 380 + ], + "type": "inline_equation", + "content": "\\pmb{x}" + }, + { + "bbox": [ + 46, + 320, + 288, + 380 + ], + "type": "text", + "content": " in the continuous image domain and latent code " + }, + { + "bbox": [ + 46, + 320, + 288, + 380 + ], + "type": "inline_equation", + "content": "\\pmb{Z}" + }, + { + "bbox": [ + 46, + 320, + 288, + 380 + ], + "type": "text", + "content": " extracted by deep neural networks, the RGB value can be predicted by an implicit image function which can be defined as follows," + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 386, + 288, + 399 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 386, + 288, + 399 + ], + "spans": [ + { + "bbox": [ + 130, + 386, + 288, + 399 + ], + "type": "interline_equation", + "content": "\\boldsymbol {I} (\\boldsymbol {x}) = f (\\boldsymbol {Z}, \\boldsymbol {x}), \\tag {1}", + "image_path": "94bd8d8f94d7151e900e7702a642e2b7cabdabf9ccec5ca30f3a13c44b8476ae.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 402, + 287, + 497 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 402, + 287, + 497 + ], + "spans": [ + { + "bbox": [ + 46, + 402, + 287, + 497 + ], + "type": "text", + "content": "where the implicit image function can be parameterized by a multilayer perceptron (MLP). Note that this implicit image function is shared by all images. For recent SISR methods [40, 86], the implicit image function " + }, + { + "bbox": [ + 46, + 402, + 287, + 497 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 46, + 402, + 287, + 497 + ], + "type": "text", + "content": " can be implemented as a PixelShuffle [60] with convolutions with a specific scale. However, these methods are independent on the coordinates, leading to an issue that they only adapt to the specific scale and are inflexible to synthesize arbitrary-scale images." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 498, + 289, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 498, + 289, + 545 + ], + "spans": [ + { + "bbox": [ + 46, + 498, + 289, + 545 + ], + "type": "text", + "content": "To predict the RGB value " + }, + { + "bbox": [ + 46, + 498, + 289, + 545 + ], + "type": "inline_equation", + "content": "I_{q}" + }, + { + "bbox": [ + 46, + 498, + 289, + 545 + ], + "type": "text", + "content": " at an arbitrary query coordinate " + }, + { + "bbox": [ + 46, + 498, + 289, + 545 + ], + "type": "inline_equation", + "content": "x_{q}" + }, + { + "bbox": [ + 46, + 498, + 289, + 545 + ], + "type": "text", + "content": ", most existing methods [13, 39, 79] propose to compute the RGB value at coordinate " + }, + { + "bbox": [ + 46, + 498, + 289, + 545 + ], + "type": "inline_equation", + "content": "x_{q}" + }, + { + "bbox": [ + 46, + 498, + 289, + 545 + ], + "type": "text", + "content": " by directly ensembling its neighborhood information," + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 551, + 289, + 572 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 551, + 289, + 572 + ], + "spans": [ + { + "bbox": [ + 55, + 551, + 289, + 572 + ], + "type": "interline_equation", + "content": "\\pmb {I} _ {q} := \\pmb {I} (\\pmb {x} _ {q}) = \\sum_ {(i, j) \\in \\mathcal {I}} w _ {i, j} \\cdot f (\\pmb {Z} _ {i, j} ^ {*}, \\pmb {x} _ {q} - \\pmb {x} _ {i, j} ^ {*}), (2)", + "image_path": "07aa94d04ada154b0b028be257bcb3e896825a390cf6c63aad7ca131ab8efe06.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 575, + 289, + 719 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 575, + 289, + 719 + ], + "spans": [ + { + "bbox": [ + 46, + 575, + 289, + 719 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 575, + 289, + 719 + ], + "type": "inline_equation", + "content": "\\mathcal{I}" + }, + { + "bbox": [ + 46, + 575, + 289, + 719 + ], + "type": "text", + "content": " is the local region centered at the query coordinate " + }, + { + "bbox": [ + 46, + 575, + 289, + 719 + ], + "type": "inline_equation", + "content": "x_{q}" + }, + { + "bbox": [ + 46, + 575, + 289, + 719 + ], + "type": "text", + "content": ", e.g., " + }, + { + "bbox": [ + 46, + 575, + 289, + 719 + ], + "type": "inline_equation", + "content": "\\mathcal{I}" + }, + { + "bbox": [ + 46, + 575, + 289, + 719 + ], + "type": "text", + "content": " can be top-left, top-right, bottom-left, bottom-right coordinates, and " + }, + { + "bbox": [ + 46, + 575, + 289, + 719 + ], + "type": "inline_equation", + "content": "w_{i}" + }, + { + "bbox": [ + 46, + 575, + 289, + 719 + ], + "type": "text", + "content": " is the weight of the neighboring pixel " + }, + { + "bbox": [ + 46, + 575, + 289, + 719 + ], + "type": "inline_equation", + "content": "x_{i,j}^{*}" + }, + { + "bbox": [ + 46, + 575, + 289, + 719 + ], + "type": "text", + "content": ", which is calculated w.r.t. the area of the rectangle between " + }, + { + "bbox": [ + 46, + 575, + 289, + 719 + ], + "type": "inline_equation", + "content": "x_{q}" + }, + { + "bbox": [ + 46, + 575, + 289, + 719 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 575, + 289, + 719 + ], + "type": "inline_equation", + "content": "x_{i,j}^{*}" + }, + { + "bbox": [ + 46, + 575, + 289, + 719 + ], + "type": "text", + "content": ", as shown in Figure 4. However, the performance improvement is limited because the ensemble weight " + }, + { + "bbox": [ + 46, + 575, + 289, + 719 + ], + "type": "inline_equation", + "content": "w_{i,j}" + }, + { + "bbox": [ + 46, + 575, + 289, + 719 + ], + "type": "text", + "content": " is purely based on the coordinates. The visual similarities are completely ignored in the weight calculation. Besides, these methods only consider the nearest latent codes, leading to a limited receptive field. In this paper, our goal is to learn the weights adaptively by leveraging both visual information and coordinate information." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 303, + 411, + 316 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 303, + 411, + 316 + ], + "spans": [ + { + "bbox": [ + 305, + 303, + 411, + 316 + ], + "type": "text", + "content": "4. Proposed Method" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 323, + 547, + 538 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 323, + 547, + 538 + ], + "spans": [ + { + "bbox": [ + 304, + 323, + 547, + 538 + ], + "type": "text", + "content": "The above local ensemble Eqn. (2) has a similar form to the attention mechanism. Specifically, the weights " + }, + { + "bbox": [ + 304, + 323, + 547, + 538 + ], + "type": "inline_equation", + "content": "w_{i}" + }, + { + "bbox": [ + 304, + 323, + 547, + 538 + ], + "type": "text", + "content": " in Eqn. (2) can be modeled as an attention map and the latter term " + }, + { + "bbox": [ + 304, + 323, + 547, + 538 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 304, + 323, + 547, + 538 + ], + "type": "text", + "content": " can be the value in the attention mechanism. Such an attention map is related to the similarities of the latent code and coordinate, and it can be calculated using both query and key which can integrate the latent code and coordinate information. In this sense, attention models can be used to mitigate the drawbacks of previous methods (i.e., calculated purely based on coordinates without considering any visual information) by learning the soft weights from both visual and coordinate information. However, the use of attention in the implicit function is non-trivial because standard self-attention [68] and neighborhood attention [26] mechanisms are based on visual features, and not conditioned on the coordinate information. To exploit the continuous representation learning from both the coordinate information and visual features, we propose a new attention for arbitrary-scale SR." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 544, + 547, + 663 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 544, + 547, + 663 + ], + "spans": [ + { + "bbox": [ + 304, + 544, + 547, + 663 + ], + "type": "text", + "content": "Continuous implicit attention-in-attention. The architecture of the proposed continuous implicit attention-in-attention for arbitrary-scale SR (called CiaoSR) is shown in Figure 3. It is called attention-in-attention because the value contains an additional embedded attention that additionally captures non-local information of repetitive patterns. Specifically, given an LR image, our attention is provided with both the visual features and coordinate information to learn good ensemble weights for the local neighboring latent codes for any query at any arbitrary-scale " + }, + { + "bbox": [ + 304, + 544, + 547, + 663 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 304, + 544, + 547, + 663 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 665, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 547, + 715 + ], + "type": "text", + "content": "Given a query coordinate " + }, + { + "bbox": [ + 304, + 665, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\pmb{x}_q" + }, + { + "bbox": [ + 304, + 665, + 547, + 715 + ], + "type": "text", + "content": " in the upsampled image, the known LR pixels in the local region centered at " + }, + { + "bbox": [ + 304, + 665, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\pmb{x}_q" + }, + { + "bbox": [ + 304, + 665, + 547, + 715 + ], + "type": "text", + "content": " are considered for the local ensemble. We define " + }, + { + "bbox": [ + 304, + 665, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\pmb{x}_k" + }, + { + "bbox": [ + 304, + 665, + 547, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 665, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\pmb{x}_v" + }, + { + "bbox": [ + 304, + 665, + 547, + 715 + ], + "type": "text", + "content": " as the key and value coordinates of the local neighbor. Let " + }, + { + "bbox": [ + 304, + 665, + 547, + 715 + ], + "type": "inline_equation", + "content": "Q, K" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1798" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 288, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 288, + 133 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 288, + 133 + ], + "type": "text", + "content": "and " + }, + { + "bbox": [ + 46, + 72, + 288, + 133 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 46, + 72, + 288, + 133 + ], + "type": "text", + "content": " be the latent code of query, key and value on the corresponding coordinates, respectively. Given pairs of coordinates and latent codes " + }, + { + "bbox": [ + 46, + 72, + 288, + 133 + ], + "type": "inline_equation", + "content": "(Q, x_{q})" + }, + { + "bbox": [ + 46, + 72, + 288, + 133 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 72, + 288, + 133 + ], + "type": "inline_equation", + "content": "(K, x_{k})" + }, + { + "bbox": [ + 46, + 72, + 288, + 133 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 72, + 288, + 133 + ], + "type": "inline_equation", + "content": "(V, x_{v})" + }, + { + "bbox": [ + 46, + 72, + 288, + 133 + ], + "type": "text", + "content": ", we define a new implicit attention function (called i-Attention) to predict RGB values at the given query " + }, + { + "bbox": [ + 46, + 72, + 288, + 133 + ], + "type": "inline_equation", + "content": "x_{q}" + }, + { + "bbox": [ + 46, + 72, + 288, + 133 + ], + "type": "text", + "content": "," + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 85, + 140, + 287, + 154 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 85, + 140, + 287, + 154 + ], + "spans": [ + { + "bbox": [ + 85, + 140, + 287, + 154 + ], + "type": "interline_equation", + "content": "\\boldsymbol {I} _ {q} = \\mathrm {i} - \\text {A t t e n t i o n} (\\boldsymbol {Q}, \\boldsymbol {K}, \\boldsymbol {V}; \\boldsymbol {x} _ {q}, \\boldsymbol {x} _ {k}, \\boldsymbol {x} _ {v}), \\tag {3}", + "image_path": "762c7f20e17c565ac4ad459c702243595e1942ebffbec9a833f9b8971a3c8ebe.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 161, + 287, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 161, + 287, + 185 + ], + "spans": [ + { + "bbox": [ + 46, + 161, + 287, + 185 + ], + "type": "text", + "content": "where the value feature " + }, + { + "bbox": [ + 46, + 161, + 287, + 185 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 46, + 161, + 287, + 185 + ], + "type": "text", + "content": " can be embedded with another attention to aggregate more non-local information." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 186, + 288, + 281 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 186, + 288, + 281 + ], + "spans": [ + { + "bbox": [ + 46, + 186, + 288, + 281 + ], + "type": "text", + "content": "As shown in Figure 2, our implicit attention is different from existing attention mechanisms by considering the additional coordinate information. In particular, the self-attention mechanism [68] without considering coordinates cannot predict pixel features that are not on the grid. The coordinate-based implicit attention calculates the distance between a key " + }, + { + "bbox": [ + 46, + 186, + 288, + 281 + ], + "type": "inline_equation", + "content": "\\boldsymbol{x}_k" + }, + { + "bbox": [ + 46, + 186, + 288, + 281 + ], + "type": "text", + "content": " and query coordinate " + }, + { + "bbox": [ + 46, + 186, + 288, + 281 + ], + "type": "inline_equation", + "content": "\\boldsymbol{x}_q" + }, + { + "bbox": [ + 46, + 186, + 288, + 281 + ], + "type": "text", + "content": ", which ignores the visual information and similarity of the features." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 283, + 288, + 295 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 283, + 288, + 295 + ], + "spans": [ + { + "bbox": [ + 46, + 283, + 288, + 295 + ], + "type": "text", + "content": "Implicit attention to learn weights for the local ensemble." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 295, + 287, + 378 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 295, + 287, + 378 + ], + "spans": [ + { + "bbox": [ + 46, + 295, + 287, + 378 + ], + "type": "text", + "content": "Unlike previous methods [13, 39] which directly use the normalized area as the ensemble weights, we propose to first use an attention to learn better weights for the ensemble of the local region. Our attention takes both the coordinate distance " + }, + { + "bbox": [ + 46, + 295, + 287, + 378 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 46, + 295, + 287, + 378 + ], + "type": "text", + "content": ", the local features " + }, + { + "bbox": [ + 46, + 295, + 287, + 378 + ], + "type": "inline_equation", + "content": "F_{l}" + }, + { + "bbox": [ + 46, + 295, + 287, + 378 + ], + "type": "text", + "content": ", as well as non-local features " + }, + { + "bbox": [ + 46, + 295, + 287, + 378 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 46, + 295, + 287, + 378 + ], + "type": "text", + "content": " into account for the learning of the weight. Formally, the attention can be written as" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 86, + 384, + 287, + 416 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 86, + 384, + 287, + 416 + ], + "spans": [ + { + "bbox": [ + 86, + 384, + 287, + 416 + ], + "type": "interline_equation", + "content": "\\boldsymbol {I} _ {q} = \\phi_ {q} \\left(\\sum_ {(i, j) \\in \\mathcal {I}} \\underbrace {\\sigma \\left(\\boldsymbol {Q} ^ {\\top} \\boldsymbol {K} _ {i , j}\\right)} _ {\\text {e n s e m b l e w e i g h t}} \\boldsymbol {V} _ {i, j}\\right), \\tag {4}", + "image_path": "9c9186746bb38e51ae83e6e3151ae5cb5a1105dd409ae8aa692fad86101b19f4.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 423, + 287, + 459 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 423, + 287, + 459 + ], + "spans": [ + { + "bbox": [ + 46, + 423, + 287, + 459 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 423, + 287, + 459 + ], + "type": "inline_equation", + "content": "\\phi_q" + }, + { + "bbox": [ + 46, + 423, + 287, + 459 + ], + "type": "text", + "content": " is the query network, " + }, + { + "bbox": [ + 46, + 423, + 287, + 459 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 46, + 423, + 287, + 459 + ], + "type": "text", + "content": " is the Softmax function, and " + }, + { + "bbox": [ + 46, + 423, + 287, + 459 + ], + "type": "inline_equation", + "content": "\\mathcal{I}" + }, + { + "bbox": [ + 46, + 423, + 287, + 459 + ], + "type": "text", + "content": " is the local region centered at " + }, + { + "bbox": [ + 46, + 423, + 287, + 459 + ], + "type": "inline_equation", + "content": "\\boldsymbol{x}_q" + }, + { + "bbox": [ + 46, + 423, + 287, + 459 + ], + "type": "text", + "content": ". The query, key, and value for this local ensemble attention are defined as" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 91, + 464, + 287, + 510 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 464, + 287, + 510 + ], + "spans": [ + { + "bbox": [ + 91, + 464, + 287, + 510 + ], + "type": "interline_equation", + "content": "\\left\\{ \\begin{array}{l} Q = \\boldsymbol {F} ^ {*}, \\\\ \\boldsymbol {K} _ {i, j} = \\phi_ {k} \\left(\\left[ \\boldsymbol {F} _ {i, j}, \\left(\\boldsymbol {r} _ {k}\\right) _ {i, j}, \\boldsymbol {s} \\right]\\right), \\\\ \\boldsymbol {V} _ {i, j} = \\phi_ {v} \\left(\\left[ \\boldsymbol {F} _ {i, j}, \\tilde {\\boldsymbol {F}} _ {i, j} \\right], \\left(\\boldsymbol {r} _ {v}\\right) _ {i, j}, \\boldsymbol {s} \\right]), \\end{array} \\right. \\tag {5}", + "image_path": "1d2e8a59720b8ddbd712f831b102d9fd48f39f609f80b9b0666f940e68153154.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 517, + 288, + 636 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 517, + 288, + 636 + ], + "spans": [ + { + "bbox": [ + 46, + 517, + 288, + 636 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 517, + 288, + 636 + ], + "type": "inline_equation", + "content": "\\phi_{k}" + }, + { + "bbox": [ + 46, + 517, + 288, + 636 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 517, + 288, + 636 + ], + "type": "inline_equation", + "content": "\\phi_v" + }, + { + "bbox": [ + 46, + 517, + 288, + 636 + ], + "type": "text", + "content": " are the key and value networks, respectively, " + }, + { + "bbox": [ + 46, + 517, + 288, + 636 + ], + "type": "inline_equation", + "content": "F^{*}" + }, + { + "bbox": [ + 46, + 517, + 288, + 636 + ], + "type": "text", + "content": " are the features of nearest neighbors on the grid of the LR image w.r.t. the query coordinate " + }, + { + "bbox": [ + 46, + 517, + 288, + 636 + ], + "type": "inline_equation", + "content": "\\pmb{x}_q" + }, + { + "bbox": [ + 46, + 517, + 288, + 636 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 517, + 288, + 636 + ], + "type": "inline_equation", + "content": "F_{i,j}" + }, + { + "bbox": [ + 46, + 517, + 288, + 636 + ], + "type": "text", + "content": " is a local feature calculated by using the unfolding operator to the latent codes at the " + }, + { + "bbox": [ + 46, + 517, + 288, + 636 + ], + "type": "inline_equation", + "content": "(i,j)" + }, + { + "bbox": [ + 46, + 517, + 288, + 636 + ], + "type": "text", + "content": "-th position, which is equivalent to concatenating the latent codes in neighboring " + }, + { + "bbox": [ + 46, + 517, + 288, + 636 + ], + "type": "inline_equation", + "content": "3\\times 3" + }, + { + "bbox": [ + 46, + 517, + 288, + 636 + ], + "type": "text", + "content": " region. " + }, + { + "bbox": [ + 46, + 517, + 288, + 636 + ], + "type": "inline_equation", + "content": "\\tilde{F}_{i,j}" + }, + { + "bbox": [ + 46, + 517, + 288, + 636 + ], + "type": "text", + "content": " are non-local features, which will be introduced below in detail. " + }, + { + "bbox": [ + 46, + 517, + 288, + 636 + ], + "type": "inline_equation", + "content": "s = [s_h;s_w]" + }, + { + "bbox": [ + 46, + 517, + 288, + 636 + ], + "type": "text", + "content": " is a two-dimensional scale, and " + }, + { + "bbox": [ + 46, + 517, + 288, + 636 + ], + "type": "inline_equation", + "content": "\\boldsymbol{r}_k" + }, + { + "bbox": [ + 46, + 517, + 288, + 636 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 517, + 288, + 636 + ], + "type": "inline_equation", + "content": "\\boldsymbol{r}_v" + }, + { + "bbox": [ + 46, + 517, + 288, + 636 + ], + "type": "text", + "content": " are the relative distance between the coordinates in the local region, i.e.," + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 82, + 645, + 287, + 658 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 82, + 645, + 287, + 658 + ], + "spans": [ + { + "bbox": [ + 82, + 645, + 287, + 658 + ], + "type": "interline_equation", + "content": "\\boldsymbol {r} _ {k} = \\boldsymbol {x} _ {q} - \\left(\\boldsymbol {x} _ {k}\\right) _ {i, j}, \\quad \\boldsymbol {r} _ {v} = \\boldsymbol {x} _ {q} - \\left(\\boldsymbol {x} _ {v}\\right) _ {i, j}. \\tag {6}", + "image_path": "8d1f020cd7beb6b0b9c6697a847fa4fdddaa5a83af0ba0cde8f5981946e5f6f2.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 666, + 288, + 677 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 666, + 288, + 677 + ], + "spans": [ + { + "bbox": [ + 46, + 666, + 288, + 677 + ], + "type": "text", + "content": "Embedded scale-aware attention for non-local features." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 677, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 288, + 713 + ], + "type": "text", + "content": "Given a query coordinate, the local ensemble attention only considers the keys and values in the small local region, resulting in a limited receptive field. However, there may exist" + } + ] + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 320, + 70, + 419, + 169 + ], + "blocks": [ + { + "bbox": [ + 320, + 70, + 419, + 169 + ], + "lines": [ + { + "bbox": [ + 320, + 70, + 419, + 169 + ], + "spans": [ + { + "bbox": [ + 320, + 70, + 419, + 169 + ], + "type": "image", + "image_path": "60e71634d4e595d01a36073e2dab1f03ab62f979760454334f462fe5a02e4a7d.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 323, + 171, + 413, + 180 + ], + "lines": [ + { + "bbox": [ + 323, + 171, + 413, + 180 + ], + "spans": [ + { + "bbox": [ + 323, + 171, + 413, + 180 + ], + "type": "text", + "content": "(a) Bilinear-based ensemble" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 430, + 70, + 529, + 169 + ], + "blocks": [ + { + "bbox": [ + 430, + 70, + 529, + 169 + ], + "lines": [ + { + "bbox": [ + 430, + 70, + 529, + 169 + ], + "spans": [ + { + "bbox": [ + 430, + 70, + 529, + 169 + ], + "type": "image", + "image_path": "d144375c24c9c5b05cac966f0cca9b68ea41524564f55af2d1041fac4f2b5690.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 432, + 171, + 526, + 180 + ], + "lines": [ + { + "bbox": [ + 432, + 171, + 526, + 180 + ], + "spans": [ + { + "bbox": [ + 432, + 171, + 526, + 180 + ], + "type": "text", + "content": "(b) Attention-based ensemble" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 304, + 182, + 547, + 258 + ], + "lines": [ + { + "bbox": [ + 304, + 182, + 547, + 258 + ], + "spans": [ + { + "bbox": [ + 304, + 182, + 547, + 258 + ], + "type": "text", + "content": "Figure 4. Comparison of different ensemble methods. (a) Most existing methods calculate the weights related to the area of the rectangle between the query coordinate and the nearest point. Such weights are equivalent to the weight of the bilinear interpolation. (b) Our attention-based ensemble calculates the attention map as the weights for feature aggregation. Besides, the ensemble method has a large receptive field for exploiting more information." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 267, + 547, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 267, + 547, + 472 + ], + "spans": [ + { + "bbox": [ + 304, + 267, + 547, + 472 + ], + "type": "text", + "content": "similar textures with different scales within a single image in other positions. For example, repetitive patterns in different scales (e.g., facades, windows, etc. in a building, as shown in Figure 4) can exist in different locations within one image. To exploit the useful non-local information, we propose an embedded scale-aware non-local attention module, which aggregates the information from all coordinates on the grid of the LR image, motivated by [49]. Specifically, we obtain query " + }, + { + "bbox": [ + 304, + 267, + 547, + 472 + ], + "type": "inline_equation", + "content": "\\tilde{Q}" + }, + { + "bbox": [ + 304, + 267, + 547, + 472 + ], + "type": "text", + "content": " and value features " + }, + { + "bbox": [ + 304, + 267, + 547, + 472 + ], + "type": "inline_equation", + "content": "\\tilde{V}" + }, + { + "bbox": [ + 304, + 267, + 547, + 472 + ], + "type": "text", + "content": " from the latent code " + }, + { + "bbox": [ + 304, + 267, + 547, + 472 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 304, + 267, + 547, + 472 + ], + "type": "text", + "content": " of the LR image, and downsample " + }, + { + "bbox": [ + 304, + 267, + 547, + 472 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 304, + 267, + 547, + 472 + ], + "type": "text", + "content": " to a scale " + }, + { + "bbox": [ + 304, + 267, + 547, + 472 + ], + "type": "inline_equation", + "content": "s'" + }, + { + "bbox": [ + 304, + 267, + 547, + 472 + ], + "type": "text", + "content": " as a key feature " + }, + { + "bbox": [ + 304, + 267, + 547, + 472 + ], + "type": "inline_equation", + "content": "\\tilde{K}" + }, + { + "bbox": [ + 304, + 267, + 547, + 472 + ], + "type": "text", + "content": ". We first calculate the pixel-wise similarity between the query located at " + }, + { + "bbox": [ + 304, + 267, + 547, + 472 + ], + "type": "inline_equation", + "content": "(i,j)" + }, + { + "bbox": [ + 304, + 267, + 547, + 472 + ], + "type": "text", + "content": " and the key on other coordinates, and then aggregate the weighted information from all of them. Lastly, we use a convolutional layer " + }, + { + "bbox": [ + 304, + 267, + 547, + 472 + ], + "type": "inline_equation", + "content": "\\varphi" + }, + { + "bbox": [ + 304, + 267, + 547, + 472 + ], + "type": "text", + "content": " for downsampling with the scale " + }, + { + "bbox": [ + 304, + 267, + 547, + 472 + ], + "type": "inline_equation", + "content": "s'" + }, + { + "bbox": [ + 304, + 267, + 547, + 472 + ], + "type": "text", + "content": " normalized similarity matrix to aggregate the features. Formally, the non-local feature " + }, + { + "bbox": [ + 304, + 267, + 547, + 472 + ], + "type": "inline_equation", + "content": "\\tilde{F}" + }, + { + "bbox": [ + 304, + 267, + 547, + 472 + ], + "type": "text", + "content": " located at " + }, + { + "bbox": [ + 304, + 267, + 547, + 472 + ], + "type": "inline_equation", + "content": "(i,j)" + }, + { + "bbox": [ + 304, + 267, + 547, + 472 + ], + "type": "text", + "content": " can be calculated by" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 478, + 545, + 519 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 478, + 545, + 519 + ], + "spans": [ + { + "bbox": [ + 307, + 478, + 545, + 519 + ], + "type": "interline_equation", + "content": "\\tilde {\\boldsymbol {F}} _ {i, j} = \\varphi \\left(\\sum_ {u, v} \\frac {\\exp \\left(\\tilde {\\boldsymbol {Q}} _ {i , j} ^ {\\top} \\tilde {\\boldsymbol {K}} _ {u , v}\\right)}{\\sum_ {u ^ {\\prime} , v ^ {\\prime}} \\exp \\left(\\tilde {\\boldsymbol {Q}} _ {i , j} ^ {\\top} \\tilde {\\boldsymbol {K}} _ {u ^ {\\prime} , v ^ {\\prime}}\\right)} \\tilde {\\boldsymbol {V}} _ {s ^ {\\prime} u, s ^ {\\prime} v} ^ {s ^ {\\prime} p \\times s ^ {\\prime} p}\\right), \\tag {7}", + "image_path": "4ec19d8d34e612f074c0e56ed37022759c003a5f4b062b29069748fe392bd7e9.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 527, + 545, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 527, + 545, + 567 + ], + "spans": [ + { + "bbox": [ + 304, + 527, + 545, + 567 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 527, + 545, + 567 + ], + "type": "inline_equation", + "content": "\\tilde{V}_{s'u,s'v}^{s'p\\times s'p}" + }, + { + "bbox": [ + 304, + 527, + 545, + 567 + ], + "type": "text", + "content": " is the value feature patch of the size " + }, + { + "bbox": [ + 304, + 527, + 545, + 567 + ], + "type": "inline_equation", + "content": "s^{\\prime}p\\times s^{\\prime}p" + }, + { + "bbox": [ + 304, + 527, + 545, + 567 + ], + "type": "text", + "content": " located at " + }, + { + "bbox": [ + 304, + 527, + 545, + 567 + ], + "type": "inline_equation", + "content": "(s^{\\prime}u,s^{\\prime}v)" + }, + { + "bbox": [ + 304, + 527, + 545, + 567 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 527, + 545, + 567 + ], + "type": "inline_equation", + "content": "\\tilde{Q},\\tilde{K}" + }, + { + "bbox": [ + 304, + 527, + 545, + 567 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 527, + 545, + 567 + ], + "type": "inline_equation", + "content": "\\tilde{V}" + }, + { + "bbox": [ + 304, + 527, + 545, + 567 + ], + "type": "text", + "content": " are the query, key and value's non-local features, which can be calculated by" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 386, + 574, + 545, + 622 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 386, + 574, + 545, + 622 + ], + "spans": [ + { + "bbox": [ + 386, + 574, + 545, + 622 + ], + "type": "interline_equation", + "content": "\\left\\{ \\begin{array}{l} \\tilde {\\boldsymbol {Q}} = \\varphi_ {q} (\\boldsymbol {F}), \\\\ \\tilde {\\boldsymbol {K}} = \\varphi_ {k} (\\boldsymbol {F} \\downarrow_ {s ^ {\\prime}}), \\\\ \\tilde {\\boldsymbol {V}} = \\varphi_ {v} (\\boldsymbol {F}), \\end{array} \\right. \\tag {8}", + "image_path": "d6720394ac82ccc05fe0f2776cb123659e8a09cbc0bc29bd504f7d9036dccea8.jpg" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 304, + 629, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 629, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 629, + 547, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 629, + 547, + 713 + ], + "type": "inline_equation", + "content": "\\downarrow_{s'}" + }, + { + "bbox": [ + 304, + 629, + 547, + 713 + ], + "type": "text", + "content": " is the downsampling operation with a given scale " + }, + { + "bbox": [ + 304, + 629, + 547, + 713 + ], + "type": "inline_equation", + "content": "s'" + }, + { + "bbox": [ + 304, + 629, + 547, + 713 + ], + "type": "text", + "content": " which can be " + }, + { + "bbox": [ + 304, + 629, + 547, + 713 + ], + "type": "inline_equation", + "content": "\\{2,3,4\\}" + }, + { + "bbox": [ + 304, + 629, + 547, + 713 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 629, + 547, + 713 + ], + "type": "inline_equation", + "content": "\\varphi_q" + }, + { + "bbox": [ + 304, + 629, + 547, + 713 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 629, + 547, + 713 + ], + "type": "inline_equation", + "content": "\\varphi_k" + }, + { + "bbox": [ + 304, + 629, + 547, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 629, + 547, + 713 + ], + "type": "inline_equation", + "content": "\\varphi_v" + }, + { + "bbox": [ + 304, + 629, + 547, + 713 + ], + "type": "text", + "content": " are query, key and value networks. Different from [49], our method can integrate multi-scale features. In the experiment, we can search for relevant features in a small window (e.g., " + }, + { + "bbox": [ + 304, + 629, + 547, + 713 + ], + "type": "inline_equation", + "content": "256\\times 256" + }, + { + "bbox": [ + 304, + 629, + 547, + 713 + ], + "type": "text", + "content": "), instead of the entire image. In this way, our method can guarantee a large receptive field, and reduce computational cost." + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1799" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 49, + 95, + 544, + 312 + ], + "blocks": [ + { + "bbox": [ + 47, + 70, + 547, + 93 + ], + "lines": [ + { + "bbox": [ + 47, + 70, + 547, + 93 + ], + "spans": [ + { + "bbox": [ + 47, + 70, + 547, + 93 + ], + "type": "text", + "content": "Table 1. Quantitative comparison with state-of-the-art methods for arbitrary-scale SR on the DIV2K validation set (PSNR (dB)). Bold indicates the best performance. " + }, + { + "bbox": [ + 47, + 70, + 547, + 93 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 47, + 70, + 547, + 93 + ], + "type": "text", + "content": " indicates the implementation of [39]." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 49, + 95, + 544, + 312 + ], + "lines": [ + { + "bbox": [ + 49, + 95, + 544, + 312 + ], + "spans": [ + { + "bbox": [ + 49, + 95, + 544, + 312 + ], + "type": "table", + "html": "
BackbonesMethodsIn-scaleOut-of-scale
×2×3×4×6×12×18×24×30
-Bicubic31.0128.2226.6624.8222.2721.0020.1919.59
EDSR [44]EDSR-baseline [44]34.5530.9028.94-----
EDSR-baseline-MetaSR [28]34.6430.9328.9226.6123.5522.0321.0620.37
EDSR-baseline-LIIF [13]34.6730.9629.0026.7523.7122.1721.1820.48
EDSR-baseline-ITSRN† [79]34.7130.9529.0326.7723.7122.1721.1820.49
EDSR-baseline-LTE [39]34.7231.0229.0426.8123.7822.2321.2420.53
EDSR-baseline-CiaoSR (ours)34.9131.1529.2326.9523.8822.3221.3220.59
RDN [88]RDN-baseline [88]34.9431.2229.19-----
RDN-MetaSR [28]35.0031.2729.2526.8823.7322.1821.1720.47
RDN-LIIF [13]34.9931.2629.2726.9923.8922.3421.3120.59
RDN-ITSRN† [79]35.0931.3629.3827.0623.9322.3621.3220.61
RDN-LTE [39]35.0431.3229.3327.0423.9522.4021.3620.64
RDN-CiaoSR (ours)35.1531.4229.4527.1624.0622.4821.4320.70
SwinIR [40]SwinIR-baseline [40]34.9431.2229.19-----
SwinIR-MetaSR† [28]35.1531.4029.3326.9423.8022.2621.2620.54
SwinIR-LIIF† [13]35.1731.4629.4627.1524.0222.4321.4020.67
SwinIR-ITSRN† [79]35.1931.4229.4827.1323.8322.3121.3120.55
SwinIR-LTE [39]35.2431.5029.5127.2024.0922.5021.4720.73
SwinIR-CiaoSR (ours)35.2931.5529.5927.2824.1522.5421.5120.74
", + "image_path": "8f1ae4c130e95e19759865a24818c0aa14da82e567bf7f879cfaf980c6524970.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 317, + 128, + 331 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 317, + 128, + 331 + ], + "spans": [ + { + "bbox": [ + 47, + 317, + 128, + 331 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 335, + 288, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 335, + 288, + 407 + ], + "spans": [ + { + "bbox": [ + 46, + 335, + 288, + 407 + ], + "type": "text", + "content": "Datasets. For the arbitrary-scale SR task, we follow [13, 39] and use DIV2K [1] as the training set, which consists of 800 images in 2K resolution. In testing, we evaluate the models on the DIV2K validation set and a wide range of standard benchmark datasets, including Set5 [3], Set14 [80], B100 [46], Urban100 [29], and Manga109 [47]." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 411, + 287, + 507 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 411, + 287, + 507 + ], + "spans": [ + { + "bbox": [ + 46, + 411, + 287, + 507 + ], + "type": "text", + "content": "Evaluation metrics. We use PSNR to evaluate the quality of the synthesized SR images. Following [13, 39], the PSNR value is calculated on all RGBs channels for the DIV2K validation set, and additionally also calculated on the Y channel (i.e., luminance) of the transformed YCbCr space for other benchmark test sets. In addition, other evaluation metrics (such as SSIM [76] and LPIPS [85]) are provided to evaluate the image quality." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 510, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 510, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 510, + 288, + 714 + ], + "type": "text", + "content": "Implementation details. We follow the same experimental setup as prior works [13, 39]. To synthesize paired training data, we first crop the " + }, + { + "bbox": [ + 46, + 510, + 288, + 714 + ], + "type": "inline_equation", + "content": "48s \\times 48s" + }, + { + "bbox": [ + 46, + 510, + 288, + 714 + ], + "type": "text", + "content": " patches as ground-truth (GT) images, and use the Bicubic downsampling in PyTorch to have LR images with the patch size of " + }, + { + "bbox": [ + 46, + 510, + 288, + 714 + ], + "type": "inline_equation", + "content": "48 \\times 48" + }, + { + "bbox": [ + 46, + 510, + 288, + 714 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 510, + 288, + 714 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 46, + 510, + 288, + 714 + ], + "type": "text", + "content": " is a scale factor sampled in the uniform distribution " + }, + { + "bbox": [ + 46, + 510, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{U}(1,4)" + }, + { + "bbox": [ + 46, + 510, + 288, + 714 + ], + "type": "text", + "content": ". We sample " + }, + { + "bbox": [ + 46, + 510, + 288, + 714 + ], + "type": "inline_equation", + "content": "48^2" + }, + { + "bbox": [ + 46, + 510, + 288, + 714 + ], + "type": "text", + "content": " pixels from both GT images and the corresponding coordinates. We use the existing SR models (e.g., EDSR [44], RDN [88] and SwinIR [40]) as backbones by removing their upsampling modules. The detailed architecture of our implicit network is provided in the supplementary material. For the training, we use Adam [36] as the optimizer, and use L1 loss to train all models for 1000 epochs with the batch size of 16 per GPU. We set the learning rate as " + }, + { + "bbox": [ + 46, + 510, + 288, + 714 + ], + "type": "inline_equation", + "content": "1e - 4" + }, + { + "bbox": [ + 46, + 510, + 288, + 714 + ], + "type": "text", + "content": " at the beginning and decay the learning rate by a factor of 0.5 every 200 epochs. These experimental settings are identical to [13, 28, 39]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 317, + 492, + 331 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 317, + 492, + 331 + ], + "spans": [ + { + "bbox": [ + 306, + 317, + 492, + 331 + ], + "type": "text", + "content": "5.1. Comparisons with State-of-the-Art" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 336, + 547, + 588 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 336, + 547, + 588 + ], + "spans": [ + { + "bbox": [ + 304, + 336, + 547, + 588 + ], + "type": "text", + "content": "Quantitative results. In Tables 1-3, CiaoSR achieves the best performance with the highest PSNR across all datasets with all backbones concerned on both in-scale and out-of-scale distributions. For the in-scales, in particular, our model has a remarkable PSNR gain of " + }, + { + "bbox": [ + 304, + 336, + 547, + 588 + ], + "type": "inline_equation", + "content": "0.3\\mathrm{dB}" + }, + { + "bbox": [ + 304, + 336, + 547, + 588 + ], + "type": "text", + "content": " on Urban100 " + }, + { + "bbox": [ + 304, + 336, + 547, + 588 + ], + "type": "inline_equation", + "content": "(\\times 4)" + }, + { + "bbox": [ + 304, + 336, + 547, + 588 + ], + "type": "text", + "content": ", compared with the previous SOTA method [39] under the same RDN backbone. It is worth noting that CiaoSR with the RDN backbone [88] can surpass the performance of the better backbone SwinIR [40] (with its original upsampling module). With the help of the continuous-scale training, a single SR model trained with our method has better generalization performance than vanilla SR models (including EDSR [44], RDN [88] and SwinIR [40] trained on specific scale. For the out-of-scale experiments, our model also achieves the best generalization performance on unseen scales. The vanilla EDSR-baseline [44], RDN [88] and SwinIR [40] cannot be applied on the out-of-scale experiments because their decoder can only predict the scale that it was trained on. We observe that our method yields the largest improvement on Urban100 over existing arbitrary-scale methods (e.g., LIIF [13] and LTE [39])." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "type": "text", + "content": "Qualitative results. In Figure 5, we provide qualitative comparisons with other arbitrary-scale SR methods. Our model is able to synthesize the SR images with sharper textures than other methods. Taking the second line as an example, the textures in the LR image are degraded, but CiaoSR is still able to restore the textures of the building. In contrast, other methods can only restore part of the textures possibly due to the limitations of the local ensemble and insufficient features. More visual results on more test sets are put in the supplementary materials." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "type": "text", + "content": "1800" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 50, + 95, + 544, + 231 + ], + "blocks": [ + { + "bbox": [ + 47, + 70, + 545, + 92 + ], + "lines": [ + { + "bbox": [ + 47, + 70, + 545, + 92 + ], + "spans": [ + { + "bbox": [ + 47, + 70, + 545, + 92 + ], + "type": "text", + "content": "Table 2. Quantitative comparison with state-of-the-art methods for in-scale SR on benchmark datasets (PSNR (dB)). Bold indicates the best performance. † indicates our implementation." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 50, + 95, + 544, + 231 + ], + "lines": [ + { + "bbox": [ + 50, + 95, + 544, + 231 + ], + "spans": [ + { + "bbox": [ + 50, + 95, + 544, + 231 + ], + "type": "table", + "html": "
MethodsSet5 [3]Set14 [80]B100 [46]Urban100 [29]Manga109 [47]
×2×3×4×2×3×4×2×3×4×2×3×4×2×3×4
RDN [88]38.2434.7132.4734.0130.5728.8132.3429.2627.7232.8928.8026.6139.1834.1331.00
RDN-MetaSR [28]38.2234.6332.3833.9830.5428.7832.3329.2627.7132.9228.8226.55---
RDN-LIIF [13]38.1734.6832.5033.9730.5328.8032.3229.2627.7432.8728.8226.6839.2634.2131.20
RDN-ITSRN† [79]38.2334.7632.5534.1930.5928.8832.3829.3227.7933.0728.9626.7739.3434.3931.37
RDN-LTE [39]38.2334.7232.6134.0930.5828.8832.3629.3027.7733.0428.9726.8139.2834.3231.30
RDN-CiaoSR (ours)38.2934.8532.6634.2230.6528.9332.4129.3427.8333.3029.1727.1139.5134.5731.57
SwinIR [40]38.3534.8932.7234.1430.7728.9432.4429.3727.8333.4029.2927.0739.6034.7431.67
SwinIR-MetaSR [28]38.2634.7732.4734.1430.6628.8532.3929.3127.7533.2929.1226.7639.4634.6231.37
SwinIR-LIIF [13]38.2834.8732.7334.1430.7528.9832.3929.3427.8433.3629.3327.1539.5734.6831.71
SwinIR-ITSRN† [79]38.2234.7532.6334.2630.7528.9732.4229.3827.8533.4629.3427.1239.6034.7531.74
SwinIR-LTE [39]38.3334.8932.8134.2530.8029.0632.4429.3927.8633.5029.4127.2439.6334.7931.79
SwinIR-CiaoSR (ours)38.3834.9132.8434.3330.8229.0832.4729.4227.9033.6529.5227.4239.6734.8431.91
", + "image_path": "00a0cc65db05cffdb5255603a9bff595f2db2483a88dc037b0798c0b29de821c.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 50, + 262, + 544, + 380 + ], + "blocks": [ + { + "bbox": [ + 47, + 238, + 545, + 259 + ], + "lines": [ + { + "bbox": [ + 47, + 238, + 545, + 259 + ], + "spans": [ + { + "bbox": [ + 47, + 238, + 545, + 259 + ], + "type": "text", + "content": "Table 3. Quantitative comparison with state-of-the-art methods for out-of-scale SR on benchmark datasets (PSNR (dB)). Bold indicates the best performance. † indicates our implementation." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 50, + 262, + 544, + 380 + ], + "lines": [ + { + "bbox": [ + 50, + 262, + 544, + 380 + ], + "spans": [ + { + "bbox": [ + 50, + 262, + 544, + 380 + ], + "type": "table", + "html": "
MethodsSet5 [3]Set14 [80]B100 [46]Urban100 [29]Manga109 [47]
×6×8×12×6×8×12×6×8×12×6×8×12×6×8×12
RDN-MetaSR [28]29.0429.96-26.5124.97-25.9024.83-23.9922.59----
RDN-LIIF [13]29.1527.1424.8626.6425.1523.2425.9824.9123.5724.2022.7921.1527.3325.0422.36
RDN-ITSRN† [79]29.3227.2524.8626.6825.1723.2826.0124.9323.5824.2322.8121.1627.4525.0423.35
RDN-LTE [39]29.3227.2624.7926.7125.1623.3126.0124.9523.6024.2822.8821.2227.4925.1222.43
RDN-CiaoSR (ours)29.4627.3624.9226.7925.2823.3726.0725.0023.6424.5823.1321.4227.7025.4022.63
SwinIR-MetaSR [28]29.0927.0224.8226.5825.0923.3325.9424.8723.5924.1622.7521.3127.2924.9622.35
SwinIR-LIIF [13]29.4627.36-26.8225.34-26.0725.01-24.5923.14-27.6925.28-
SwinIR-ITSRN† [79]29.3127.2424.7926.7125.3223.3026.0524.9623.5724.5023.0621.3427.7225.2322.47
SwinIR-LTE [39]29.5027.35-26.8625.42-26.0925.03-24.6223.17-27.8325.42-
SwinIR-CiaoSR (ours)29.6227.4524.9626.8825.4223.3826.1325.0723.6824.8423.3421.6028.0125.6122.79
", + "image_path": "497ddbeb06f38210dad7a4a5fa41d3e53f06c7042ce7e509ef3fe8ab9bcbd2f4.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 399, + 140, + 411 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 399, + 140, + 411 + ], + "spans": [ + { + "bbox": [ + 47, + 399, + 140, + 411 + ], + "type": "text", + "content": "5.2. Ablation Study" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 420, + 288, + 480 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 420, + 288, + 480 + ], + "spans": [ + { + "bbox": [ + 47, + 420, + 288, + 480 + ], + "type": "text", + "content": "In this section, we conduct ablation studies to investigate effect of each component in our architecture. Our CiaoSR consists of the attention-in-attention network and a scale-aware non-local attention module. Table 4 shows the contributions of each component on performance." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 483, + 287, + 625 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 483, + 287, + 625 + ], + "spans": [ + { + "bbox": [ + 46, + 483, + 287, + 625 + ], + "type": "text", + "content": "Attention-in-attention. The attention-in-attention network is the main branch of our architecture. On the one hand, the network learns to ensemble weights with both the coordinates and the features. On the other hand, the network is able to aggregate the local and non-local information to improve the SR performance. To investigate how the attention mechanism affects the performance, we remove this module and replace it with an MLP to predict the weights of the ensemble. The coordinate information is still fed into the MLP to enable arbitrary scale prediction. As shown in Table 4, the method suffers a significant performance drop without the attention mechanism." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "type": "text", + "content": "Non-local attention. The embedded non-local attention is proposed to capture the scale-aware information from other locations of the images to improve the performance. We investigate the importance of the non-local attention in an implicit model by comparing the results with and without this module. As shown in Table 4, by introducing the nonlocal attention, a further performance gain is achieved. This" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 399, + 545, + 423 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 399, + 545, + 423 + ], + "spans": [ + { + "bbox": [ + 305, + 399, + 545, + 423 + ], + "type": "text", + "content": "is because our module can exploit texture information from other scales in a large perceptual field." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 426, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 426, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 426, + 547, + 713 + ], + "type": "text", + "content": "Training with different types of scales. We evaluate the effectiveness of our implicit model when trained with discrete (including single scale " + }, + { + "bbox": [ + 304, + 426, + 547, + 713 + ], + "type": "inline_equation", + "content": "\\{\\{2\\} /\\{3\\} /\\{4\\}\\}" + }, + { + "bbox": [ + 304, + 426, + 547, + 713 + ], + "type": "text", + "content": " and multiple scales " + }, + { + "bbox": [ + 304, + 426, + 547, + 713 + ], + "type": "inline_equation", + "content": "\\{2,3,4\\}" + }, + { + "bbox": [ + 304, + 426, + 547, + 713 + ], + "type": "text", + "content": ") and continuous scales [1, 4], and show the results in Table 5. For the discrete type of scales, we observe that training our implicit model with a specific scale can also achieve significant improvement for that specific scale, compared with the original upsampling module in RDN [88]. Taking " + }, + { + "bbox": [ + 304, + 426, + 547, + 713 + ], + "type": "inline_equation", + "content": "\\times 2" + }, + { + "bbox": [ + 304, + 426, + 547, + 713 + ], + "type": "text", + "content": " scale as an example, our method has an improvement of " + }, + { + "bbox": [ + 304, + 426, + 547, + 713 + ], + "type": "inline_equation", + "content": "0.24\\mathrm{dB}" + }, + { + "bbox": [ + 304, + 426, + 547, + 713 + ], + "type": "text", + "content": " over RDN. However, training with a specific scale has poor generalization performance on unseen in-scale and out-of-scale. A few existing works [34, 44] consider the training with multiple discrete scales and found the performance of the joint learned model is comparable with single scale networks. Based on our implicit attention module, we are able to deal with more scales and exploit the correlation of multiple SR tasks at different scales. Unlike previous works [34, 44] which only achieve comparable performance compared with single scale models, CiaoSR achieves a better performance than training with one scale. We found our model has further improvement when trained on continuous scales. The improvement gain is attributed to our continuous representations design and our architecture, which enables efficient learning from cross-scale information." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 749, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 749, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 749, + 314, + 757 + ], + "type": "text", + "content": "1801" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 52, + 70, + 200, + 183 + ], + "blocks": [ + { + "bbox": [ + 52, + 70, + 200, + 183 + ], + "lines": [ + { + "bbox": [ + 52, + 70, + 200, + 183 + ], + "spans": [ + { + "bbox": [ + 52, + 70, + 200, + 183 + ], + "type": "image", + "image_path": "017229726e5527e13d8aef0b37d9453d1e3d2ebdfac7110c5c278dd744ee43ef.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 84, + 185, + 165, + 194 + ], + "lines": [ + { + "bbox": [ + 84, + 185, + 165, + 194 + ], + "spans": [ + { + "bbox": [ + 84, + 185, + 165, + 194 + ], + "type": "text", + "content": "004, " + }, + { + "bbox": [ + 84, + 185, + 165, + 194 + ], + "type": "inline_equation", + "content": "\\times 4" + }, + { + "bbox": [ + 84, + 185, + 165, + 194 + ], + "type": "text", + "content": " Urban100 [29]" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 202, + 70, + 284, + 121 + ], + "blocks": [ + { + "bbox": [ + 202, + 70, + 284, + 121 + ], + "lines": [ + { + "bbox": [ + 202, + 70, + 284, + 121 + ], + "spans": [ + { + "bbox": [ + 202, + 70, + 284, + 121 + ], + "type": "image", + "image_path": "2fe773b1ddcce8e0ee21e1300ad8e6b8977e0249c147971b975feee4dd033d43.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 229, + 123, + 255, + 131 + ], + "lines": [ + { + "bbox": [ + 229, + 123, + 255, + 131 + ], + "spans": [ + { + "bbox": [ + 229, + 123, + 255, + 131 + ], + "type": "text", + "content": "Bicubic" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 286, + 70, + 369, + 121 + ], + "blocks": [ + { + "bbox": [ + 286, + 70, + 369, + 121 + ], + "lines": [ + { + "bbox": [ + 286, + 70, + 369, + 121 + ], + "spans": [ + { + "bbox": [ + 286, + 70, + 369, + 121 + ], + "type": "image", + "image_path": "c00f007273de11b08d3527a8eb7d8f8fef086df0bddd263fca24865ff1866cb3.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 310, + 122, + 345, + 132 + ], + "lines": [ + { + "bbox": [ + 310, + 122, + 345, + 132 + ], + "spans": [ + { + "bbox": [ + 310, + 122, + 345, + 132 + ], + "type": "text", + "content": "RDN [88]" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 370, + 71, + 454, + 121 + ], + "blocks": [ + { + "bbox": [ + 370, + 71, + 454, + 121 + ], + "lines": [ + { + "bbox": [ + 370, + 71, + 454, + 121 + ], + "spans": [ + { + "bbox": [ + 370, + 71, + 454, + 121 + ], + "type": "image", + "image_path": "8abf6fcadcaebe33966e339674e949c22ddb0871faddbf91e1346c1131aa3610.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 380, + 122, + 444, + 132 + ], + "lines": [ + { + "bbox": [ + 380, + 122, + 444, + 132 + ], + "spans": [ + { + "bbox": [ + 380, + 122, + 444, + 132 + ], + "type": "text", + "content": "RDN-MetaSR [28]" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 455, + 71, + 539, + 121 + ], + "blocks": [ + { + "bbox": [ + 455, + 71, + 539, + 121 + ], + "lines": [ + { + "bbox": [ + 455, + 71, + 539, + 121 + ], + "spans": [ + { + "bbox": [ + 455, + 71, + 539, + 121 + ], + "type": "image", + "image_path": "27887c0a76eec3bc040c897bd324471f33bcbfb139a7b7be652212f64caa3bd0.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 470, + 122, + 523, + 132 + ], + "lines": [ + { + "bbox": [ + 470, + 122, + 523, + 132 + ], + "spans": [ + { + "bbox": [ + 470, + 122, + 523, + 132 + ], + "type": "text", + "content": "RDN-LIIF [13]" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 52, + 195, + 199, + 310 + ], + "blocks": [ + { + "bbox": [ + 52, + 195, + 199, + 310 + ], + "lines": [ + { + "bbox": [ + 52, + 195, + 199, + 310 + ], + "spans": [ + { + "bbox": [ + 52, + 195, + 199, + 310 + ], + "type": "image", + "image_path": "c0e51091767761fb3c9ceb37a57853b63679be7157190a5af1a4604e9cd3db20.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 63, + 312, + 187, + 321 + ], + "lines": [ + { + "bbox": [ + 63, + 312, + 187, + 321 + ], + "spans": [ + { + "bbox": [ + 63, + 312, + 187, + 321 + ], + "type": "text", + "content": "TaiyouNiSmash, " + }, + { + "bbox": [ + 63, + 312, + 187, + 321 + ], + "type": "inline_equation", + "content": "\\times 8" + }, + { + "bbox": [ + 63, + 312, + 187, + 321 + ], + "type": "text", + "content": " ,Manga109 [47]" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 201, + 133, + 285, + 183 + ], + "blocks": [ + { + "bbox": [ + 201, + 133, + 285, + 183 + ], + "lines": [ + { + "bbox": [ + 201, + 133, + 285, + 183 + ], + "spans": [ + { + "bbox": [ + 201, + 133, + 285, + 183 + ], + "type": "image", + "image_path": "2bcfdc0db4a5e5c0425a5ca1ec1ba3a26f7573bcc8d03650e461a3d6ca724f61.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 212, + 185, + 273, + 194 + ], + "lines": [ + { + "bbox": [ + 212, + 185, + 273, + 194 + ], + "spans": [ + { + "bbox": [ + 212, + 185, + 273, + 194 + ], + "type": "text", + "content": "RDN-ITSRN [79]" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 286, + 133, + 369, + 183 + ], + "blocks": [ + { + "bbox": [ + 286, + 133, + 369, + 183 + ], + "lines": [ + { + "bbox": [ + 286, + 133, + 369, + 183 + ], + "spans": [ + { + "bbox": [ + 286, + 133, + 369, + 183 + ], + "type": "image", + "image_path": "2d6d904b3e21f45e9b65f82837ecd4bd3e955473a6aab3b075234602a51270eb.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 302, + 185, + 353, + 194 + ], + "lines": [ + { + "bbox": [ + 302, + 185, + 353, + 194 + ], + "spans": [ + { + "bbox": [ + 302, + 185, + 353, + 194 + ], + "type": "text", + "content": "RDN-LTE [39]" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 370, + 133, + 454, + 183 + ], + "blocks": [ + { + "bbox": [ + 370, + 133, + 454, + 183 + ], + "lines": [ + { + "bbox": [ + 370, + 133, + 454, + 183 + ], + "spans": [ + { + "bbox": [ + 370, + 133, + 454, + 183 + ], + "type": "image", + "image_path": "292dfd17c80cfb1ae4a848b5470d1d40d7c46a5bd5ce26b8d8d5cae24925382e.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 376, + 185, + 448, + 194 + ], + "lines": [ + { + "bbox": [ + 376, + 185, + 448, + 194 + ], + "spans": [ + { + "bbox": [ + 376, + 185, + 448, + 194 + ], + "type": "text", + "content": "RDN-CiaoSR (ours)" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 455, + 133, + 539, + 183 + ], + "blocks": [ + { + "bbox": [ + 455, + 133, + 539, + 183 + ], + "lines": [ + { + "bbox": [ + 455, + 133, + 539, + 183 + ], + "spans": [ + { + "bbox": [ + 455, + 133, + 539, + 183 + ], + "type": "image", + "image_path": "d7397286c223ff8215569bb6039996a8daa3c70efc7bb33ab1c40a0020662355.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 490, + 185, + 503, + 194 + ], + "lines": [ + { + "bbox": [ + 490, + 185, + 503, + 194 + ], + "spans": [ + { + "bbox": [ + 490, + 185, + 503, + 194 + ], + "type": "text", + "content": "GT" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 201, + 195, + 285, + 246 + ], + "blocks": [ + { + "bbox": [ + 201, + 195, + 285, + 246 + ], + "lines": [ + { + "bbox": [ + 201, + 195, + 285, + 246 + ], + "spans": [ + { + "bbox": [ + 201, + 195, + 285, + 246 + ], + "type": "image", + "image_path": "342bf68f4fc0645d74ce8570ea2952c88a231299ed6e76ac06ad1f6df0f173da.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 228, + 247, + 255, + 256 + ], + "lines": [ + { + "bbox": [ + 228, + 247, + 255, + 256 + ], + "spans": [ + { + "bbox": [ + 228, + 247, + 255, + 256 + ], + "type": "text", + "content": "Bicubic" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 286, + 195, + 369, + 246 + ], + "blocks": [ + { + "bbox": [ + 286, + 195, + 369, + 246 + ], + "lines": [ + { + "bbox": [ + 286, + 195, + 369, + 246 + ], + "spans": [ + { + "bbox": [ + 286, + 195, + 369, + 246 + ], + "type": "image", + "image_path": "cf7e663ba7cca5a4992d8cc395c2a0052c25b682e29731ca6c9415eae3d95e67.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 309, + 247, + 346, + 257 + ], + "lines": [ + { + "bbox": [ + 309, + 247, + 346, + 257 + ], + "spans": [ + { + "bbox": [ + 309, + 247, + 346, + 257 + ], + "type": "inline_equation", + "content": "\\mathbf{RDN}^*" + }, + { + "bbox": [ + 309, + 247, + 346, + 257 + ], + "type": "text", + "content": " [88]" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 370, + 195, + 454, + 246 + ], + "blocks": [ + { + "bbox": [ + 370, + 195, + 454, + 246 + ], + "lines": [ + { + "bbox": [ + 370, + 195, + 454, + 246 + ], + "spans": [ + { + "bbox": [ + 370, + 195, + 454, + 246 + ], + "type": "image", + "image_path": "abc35dd40df46d66c132bab6e4ac05ce6e573d5cbcb4baede219d2548093fa44.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 380, + 247, + 444, + 257 + ], + "lines": [ + { + "bbox": [ + 380, + 247, + 444, + 257 + ], + "spans": [ + { + "bbox": [ + 380, + 247, + 444, + 257 + ], + "type": "text", + "content": "RDN-MetaSR [28]" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 455, + 195, + 539, + 246 + ], + "blocks": [ + { + "bbox": [ + 455, + 195, + 539, + 246 + ], + "lines": [ + { + "bbox": [ + 455, + 195, + 539, + 246 + ], + "spans": [ + { + "bbox": [ + 455, + 195, + 539, + 246 + ], + "type": "image", + "image_path": "0f7fed09bf0debfd75b82f8b4fe2f6962c8934750ba1dc1a27162702a614b926.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 470, + 247, + 523, + 257 + ], + "lines": [ + { + "bbox": [ + 470, + 247, + 523, + 257 + ], + "spans": [ + { + "bbox": [ + 470, + 247, + 523, + 257 + ], + "type": "text", + "content": "RDN-LIIF [13]" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_caption" + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 201, + 258, + 285, + 308 + ], + "blocks": [ + { + "bbox": [ + 201, + 258, + 285, + 308 + ], + "lines": [ + { + "bbox": [ + 201, + 258, + 285, + 308 + ], + "spans": [ + { + "bbox": [ + 201, + 258, + 285, + 308 + ], + "type": "image", + "image_path": "86064fece19148d6164b92441003f33f93fac189a2422a55e68712c79c345c04.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 212, + 310, + 273, + 319 + ], + "lines": [ + { + "bbox": [ + 212, + 310, + 273, + 319 + ], + "spans": [ + { + "bbox": [ + 212, + 310, + 273, + 319 + ], + "type": "text", + "content": "RDN-ITSRN [79]" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_caption" + } + ], + "index": 28 + }, + { + "type": "image", + "bbox": [ + 286, + 258, + 369, + 308 + ], + "blocks": [ + { + "bbox": [ + 286, + 258, + 369, + 308 + ], + "lines": [ + { + "bbox": [ + 286, + 258, + 369, + 308 + ], + "spans": [ + { + "bbox": [ + 286, + 258, + 369, + 308 + ], + "type": "image", + "image_path": "17ec3b36a50b2eaf4cbcb4c3fcabed2ebe604e770b095ccbdf7059c4381ae2d2.jpg" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 312, + 310, + 343, + 319 + ], + "lines": [ + { + "bbox": [ + 312, + 310, + 343, + 319 + ], + "spans": [ + { + "bbox": [ + 312, + 310, + 343, + 319 + ], + "type": "text", + "content": "LTE [39]" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_caption" + } + ], + "index": 30 + }, + { + "type": "image", + "bbox": [ + 370, + 258, + 454, + 308 + ], + "blocks": [ + { + "bbox": [ + 370, + 258, + 454, + 308 + ], + "lines": [ + { + "bbox": [ + 370, + 258, + 454, + 308 + ], + "spans": [ + { + "bbox": [ + 370, + 258, + 454, + 308 + ], + "type": "image", + "image_path": "1e11f63b34a01d42b634a7dd52baf94259ad892dd975f767cbd09268ee9fae12.jpg" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 376, + 310, + 448, + 319 + ], + "lines": [ + { + "bbox": [ + 376, + 310, + 448, + 319 + ], + "spans": [ + { + "bbox": [ + 376, + 310, + 448, + 319 + ], + "type": "text", + "content": "RDN-CiaoSR (ours)" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_caption" + } + ], + "index": 32 + }, + { + "type": "image", + "bbox": [ + 455, + 258, + 539, + 308 + ], + "blocks": [ + { + "bbox": [ + 455, + 258, + 539, + 308 + ], + "lines": [ + { + "bbox": [ + 455, + 258, + 539, + 308 + ], + "spans": [ + { + "bbox": [ + 455, + 258, + 539, + 308 + ], + "type": "image", + "image_path": "e21fae8d43d29ea55b5c3b1a07f9222489850a2263351299d40921bb91453c4e.jpg" + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 490, + 310, + 503, + 319 + ], + "lines": [ + { + "bbox": [ + 490, + 310, + 503, + 319 + ], + "spans": [ + { + "bbox": [ + 490, + 310, + 503, + 319 + ], + "type": "text", + "content": "GT" + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "image_caption" + } + ], + "index": 34 + }, + { + "type": "image", + "bbox": [ + 48, + 343, + 145, + 407 + ], + "blocks": [ + { + "bbox": [ + 70, + 327, + 522, + 338 + ], + "lines": [ + { + "bbox": [ + 70, + 327, + 522, + 338 + ], + "spans": [ + { + "bbox": [ + 70, + 327, + 522, + 338 + ], + "type": "text", + "content": "Figure 5. Visual comparison of different methods on benchmarks. \\*\\* means the model first synthesizes twice to " + }, + { + "bbox": [ + 70, + 327, + 522, + 338 + ], + "type": "inline_equation", + "content": "\\times 8" + }, + { + "bbox": [ + 70, + 327, + 522, + 338 + ], + "type": "text", + "content": " images." + } + ] + } + ], + "index": 36, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 48, + 343, + 145, + 407 + ], + "lines": [ + { + "bbox": [ + 48, + 343, + 145, + 407 + ], + "spans": [ + { + "bbox": [ + 48, + 343, + 145, + 407 + ], + "type": "image", + "image_path": "135999c514ff8e8d08702af099945753c294d6fad8758bd96d9ffd71c90544fa.jpg" + } + ] + } + ], + "index": 37, + "angle": 0, + "type": "image_body" + } + ], + "index": 37 + }, + { + "type": "image", + "bbox": [ + 48, + 408, + 145, + 456 + ], + "blocks": [ + { + "bbox": [ + 48, + 408, + 145, + 456 + ], + "lines": [ + { + "bbox": [ + 48, + 408, + 145, + 456 + ], + "spans": [ + { + "bbox": [ + 48, + 408, + 145, + 456 + ], + "type": "image", + "image_path": "08a54ee2edf90283c42f6bc8f2cbe684813aa400c0992f721efc4e27ca3b1e79.jpg" + } + ] + } + ], + "index": 38, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 91, + 457, + 101, + 464 + ], + "lines": [ + { + "bbox": [ + 91, + 457, + 101, + 464 + ], + "spans": [ + { + "bbox": [ + 91, + 457, + 101, + 464 + ], + "type": "text", + "content": "LR" + } + ] + } + ], + "index": 39, + "angle": 0, + "type": "image_caption" + } + ], + "index": 38 + }, + { + "type": "image", + "bbox": [ + 149, + 343, + 244, + 407 + ], + "blocks": [ + { + "bbox": [ + 149, + 343, + 244, + 407 + ], + "lines": [ + { + "bbox": [ + 149, + 343, + 244, + 407 + ], + "spans": [ + { + "bbox": [ + 149, + 343, + 244, + 407 + ], + "type": "image", + "image_path": "5c7802381b4a599dad06a3dac80a4420ab7247a6e35ae8d2d1ce514017533deb.jpg" + } + ] + } + ], + "index": 40, + "angle": 0, + "type": "image_body" + } + ], + "index": 40 + }, + { + "type": "image", + "bbox": [ + 149, + 408, + 244, + 457 + ], + "blocks": [ + { + "bbox": [ + 149, + 408, + 244, + 457 + ], + "lines": [ + { + "bbox": [ + 149, + 408, + 244, + 457 + ], + "spans": [ + { + "bbox": [ + 149, + 408, + 244, + 457 + ], + "type": "image", + "image_path": "db09f92a251b8f0eba18effab240fa418673056ba0ef04aeece7e330ca6c1dee.jpg" + } + ] + } + ], + "index": 41, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 182, + 458, + 208, + 465 + ], + "lines": [ + { + "bbox": [ + 182, + 458, + 208, + 465 + ], + "spans": [ + { + "bbox": [ + 182, + 458, + 208, + 465 + ], + "type": "text", + "content": "RealSR" + } + ] + } + ], + "index": 42, + "angle": 0, + "type": "image_caption" + } + ], + "index": 41 + }, + { + "type": "image", + "bbox": [ + 249, + 343, + 344, + 407 + ], + "blocks": [ + { + "bbox": [ + 249, + 343, + 344, + 407 + ], + "lines": [ + { + "bbox": [ + 249, + 343, + 344, + 407 + ], + "spans": [ + { + "bbox": [ + 249, + 343, + 344, + 407 + ], + "type": "image", + "image_path": "c04769fb9e91c1b33811499688678a7f710e60c35f53268bffe7614b837e92f7.jpg" + } + ] + } + ], + "index": 43, + "angle": 0, + "type": "image_body" + } + ], + "index": 43 + }, + { + "type": "image", + "bbox": [ + 249, + 408, + 344, + 456 + ], + "blocks": [ + { + "bbox": [ + 249, + 408, + 344, + 456 + ], + "lines": [ + { + "bbox": [ + 249, + 408, + 344, + 456 + ], + "spans": [ + { + "bbox": [ + 249, + 408, + 344, + 456 + ], + "type": "image", + "image_path": "d6003f11db1ca87a94d12b24cec1ebbbe4093a71f6c072b2ffc2c7f2dcfdf131.jpg" + } + ] + } + ], + "index": 44, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 280, + 458, + 312, + 465 + ], + "lines": [ + { + "bbox": [ + 280, + 458, + 312, + 465 + ], + "spans": [ + { + "bbox": [ + 280, + 458, + 312, + 465 + ], + "type": "text", + "content": "BSRGAN" + } + ] + } + ], + "index": 45, + "angle": 0, + "type": "image_caption" + } + ], + "index": 44 + }, + { + "type": "image", + "bbox": [ + 348, + 343, + 444, + 407 + ], + "blocks": [ + { + "bbox": [ + 348, + 343, + 444, + 407 + ], + "lines": [ + { + "bbox": [ + 348, + 343, + 444, + 407 + ], + "spans": [ + { + "bbox": [ + 348, + 343, + 444, + 407 + ], + "type": "image", + "image_path": "931cf6292228ee8a7a9ac79e2b6e75d129b6e19c535c142a1884e2e53e10e9f9.jpg" + } + ] + } + ], + "index": 46, + "angle": 0, + "type": "image_body" + } + ], + "index": 46 + }, + { + "type": "image", + "bbox": [ + 348, + 408, + 444, + 456 + ], + "blocks": [ + { + "bbox": [ + 348, + 408, + 444, + 456 + ], + "lines": [ + { + "bbox": [ + 348, + 408, + 444, + 456 + ], + "spans": [ + { + "bbox": [ + 348, + 408, + 444, + 456 + ], + "type": "image", + "image_path": "69e488a7c5262c818a2f8769b4323953a87e785a943dbe0916f2f2345b604a2c.jpg" + } + ] + } + ], + "index": 47, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 371, + 458, + 419, + 465 + ], + "lines": [ + { + "bbox": [ + 371, + 458, + 419, + 465 + ], + "spans": [ + { + "bbox": [ + 371, + 458, + 419, + 465 + ], + "type": "text", + "content": "Real-ESRGAN" + } + ] + } + ], + "index": 48, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 110, + 471, + 481, + 482 + ], + "lines": [ + { + "bbox": [ + 110, + 471, + 481, + 482 + ], + "spans": [ + { + "bbox": [ + 110, + 471, + 481, + 482 + ], + "type": "text", + "content": "Figure 6. Visual comparison of different methods on the RealSRSet [82] and DPED [30] dataset " + }, + { + "bbox": [ + 110, + 471, + 481, + 482 + ], + "type": "inline_equation", + "content": "(\\times 8)" + }, + { + "bbox": [ + 110, + 471, + 481, + 482 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 52, + "angle": 0, + "type": "image_caption" + } + ], + "index": 47 + }, + { + "type": "image", + "bbox": [ + 449, + 343, + 527, + 407 + ], + "blocks": [ + { + "bbox": [ + 449, + 343, + 527, + 407 + ], + "lines": [ + { + "bbox": [ + 449, + 343, + 527, + 407 + ], + "spans": [ + { + "bbox": [ + 449, + 343, + 527, + 407 + ], + "type": "image", + "image_path": "b256a9a1bc3a33ea924a66e938a317c36a9c18dc1321dc08dc1b38b723ccf74e.jpg" + } + ] + } + ], + "index": 49, + "angle": 0, + "type": "image_body" + } + ], + "index": 49 + }, + { + "type": "image", + "bbox": [ + 449, + 408, + 527, + 456 + ], + "blocks": [ + { + "bbox": [ + 449, + 408, + 527, + 456 + ], + "lines": [ + { + "bbox": [ + 449, + 408, + 527, + 456 + ], + "spans": [ + { + "bbox": [ + 449, + 408, + 527, + 456 + ], + "type": "image", + "image_path": "3fffe2e75369a2f7440ec9ebff286f31cdcbafdc9a5716c67ea03276177ebfcb.jpg" + } + ] + } + ], + "index": 50, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 487, + 458, + 503, + 465 + ], + "lines": [ + { + "bbox": [ + 487, + 458, + 503, + 465 + ], + "spans": [ + { + "bbox": [ + 487, + 458, + 503, + 465 + ], + "type": "text", + "content": "Ours" + } + ] + } + ], + "index": 51, + "angle": 0, + "type": "image_caption" + } + ], + "index": 50 + }, + { + "bbox": [ + 47, + 488, + 137, + 501 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 488, + 137, + 501 + ], + "spans": [ + { + "bbox": [ + 47, + 488, + 137, + 501 + ], + "type": "text", + "content": "5.3. Further Study" + } + ] + } + ], + "index": 53 + }, + { + "bbox": [ + 46, + 506, + 288, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 506, + 288, + 662 + ], + "spans": [ + { + "bbox": [ + 46, + 506, + 288, + 662 + ], + "type": "text", + "content": "Effect on synthesis steps. Our trained model is able to synthesize SR images with any given arbitrary scale in one step. A natural alternative would be to synthesize a large scale in multiple steps. Taking a large scale " + }, + { + "bbox": [ + 46, + 506, + 288, + 662 + ], + "type": "inline_equation", + "content": "(e.g., \\times 12)" + }, + { + "bbox": [ + 46, + 506, + 288, + 662 + ], + "type": "text", + "content": " as an example, SR images can be generated by two steps " + }, + { + "bbox": [ + 46, + 506, + 288, + 662 + ], + "type": "inline_equation", + "content": "(e.g., \\rightarrow \\times 2 \\rightarrow \\times 12)" + }, + { + "bbox": [ + 46, + 506, + 288, + 662 + ], + "type": "text", + "content": ", or three steps " + }, + { + "bbox": [ + 46, + 506, + 288, + 662 + ], + "type": "inline_equation", + "content": "(e.g., \\rightarrow \\times 2 \\rightarrow \\times 4 \\rightarrow \\times 12)" + }, + { + "bbox": [ + 46, + 506, + 288, + 662 + ], + "type": "text", + "content": ". We compare the performance of synthesizing a large scale in one and multiple steps, and the results on the Urban100 and Manga109 datasets are shown in Table 6. We observe that our way of synthesizing SR images in one step has the best performance under the PSNR and SSIM metrics. In contrast, the performance gets worse with more steps because errors can accumulate along the multiple synthesizing steps." + } + ] + } + ], + "index": 54 + }, + { + "bbox": [ + 46, + 665, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 288, + 714 + ], + "type": "text", + "content": "Model size and inference time. To demonstrate the efficiency of our proposed method, we investigate the model size and inference time of different arbitrary-scale SR methods. Results are shown in Table 7. Specifically, we set" + } + ] + } + ], + "index": 55 + }, + { + "bbox": [ + 304, + 489, + 547, + 561 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 489, + 547, + 561 + ], + "spans": [ + { + "bbox": [ + 304, + 489, + 547, + 561 + ], + "type": "text", + "content": "the input size as " + }, + { + "bbox": [ + 304, + 489, + 547, + 561 + ], + "type": "inline_equation", + "content": "192 \\times 192" + }, + { + "bbox": [ + 304, + 489, + 547, + 561 + ], + "type": "text", + "content": " and calculate the inference time on an NVIDIA Quadro RTX 6000. Our implicit model has much fewer parameters than the previous SOTA methods (e.g., LTE [39]). Our model has the best performance although it requires a little more inference time due to exploiting the non-local features." + } + ] + } + ], + "index": 56 + }, + { + "bbox": [ + 304, + 570, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 547, + 714 + ], + "type": "text", + "content": "More evaluation metrics. To further demonstrate the effectiveness of our proposed method, we apply more evaluation metrics (e.g., SSIM [76] and LPIPS [85]) to evaluate the image quality across different methods on the Urban100 dataset. These evaluation metrics are widely used in the SR community. In general, higher SSIM and lower LPIPS correspond to better performance. As shown in Table 8, our method achieves the highest SSIM and lowest LPIPS, and thus has the best performance on both in-scale and out-of-scale distributions. Compared with other methods, our proposed CiaoSR is able to maintain more structural texture and perceptual information." + } + ] + } + ], + "index": 57 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1802" + } + ] + } + ], + "index": 58 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 50, + 95, + 271, + 167 + ], + "blocks": [ + { + "bbox": [ + 47, + 71, + 272, + 92 + ], + "lines": [ + { + "bbox": [ + 47, + 71, + 272, + 92 + ], + "spans": [ + { + "bbox": [ + 47, + 71, + 272, + 92 + ], + "type": "text", + "content": "Table 4. Ablation study on each component of our networks on Urban100. We use RDN [88] as the backbone." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 50, + 95, + 271, + 167 + ], + "lines": [ + { + "bbox": [ + 50, + 95, + 271, + 167 + ], + "spans": [ + { + "bbox": [ + 50, + 95, + 271, + 167 + ], + "type": "table", + "html": "
Attention-in-attention✗✓✓
Scale-aware Attention Network✗✗✓
In-scale×232.8733.2433.30
×328.8229.1029.17
×426.6926.9627.11
Out-of-scale×624.2224.5024.58
×822.8022.9823.13
", + "image_path": "728339d8915a08bb127c9a7f70cbc47d99b1e88ceb371f45cee0429910629cf7.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 50, + 195, + 280, + 244 + ], + "blocks": [ + { + "bbox": [ + 47, + 171, + 282, + 193 + ], + "lines": [ + { + "bbox": [ + 47, + 171, + 282, + 193 + ], + "spans": [ + { + "bbox": [ + 47, + 171, + 282, + 193 + ], + "type": "text", + "content": "Table 6. Comparison of (PSNR and SSIM) for different synthesis steps on Urban100 [29] and Manga109 [47]." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 50, + 195, + 280, + 244 + ], + "lines": [ + { + "bbox": [ + 50, + 195, + 280, + 244 + ], + "spans": [ + { + "bbox": [ + 50, + 195, + 280, + 244 + ], + "type": "table", + "html": "
TypeSynthesis stepsUrban100 [29]Manga109 [47]
PSNRSSIMPSNRSSIM
Multiple steps→×2→×4→×1221.280.55722.460.720
→×2→×1221.320.55822.530.721
One step→×1221.420.56122.630.723
", + "image_path": "a13b0f46dea86ede9d1b23da663702165834a8f74220097ef178d7f6924f77af.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 279, + 95, + 549, + 167 + ], + "blocks": [ + { + "bbox": [ + 276, + 72, + 552, + 92 + ], + "lines": [ + { + "bbox": [ + 276, + 72, + 552, + 92 + ], + "spans": [ + { + "bbox": [ + 276, + 72, + 552, + 92 + ], + "type": "text", + "content": "Table 5. Ablation study on training our implicit model with different types of scales on Urban100." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 279, + 95, + 549, + 167 + ], + "lines": [ + { + "bbox": [ + 279, + 95, + 549, + 167 + ], + "spans": [ + { + "bbox": [ + 279, + 95, + 549, + 167 + ], + "type": "table", + "html": "
TypeTraining scale sIn-scaleOut-of-scale
×2×3×4×6×8
Discretes ∈ {2}33.1327.0125.6022.2722.09
s ∈ {3}31.3929.0625.7723.4422.16
s ∈ {4}31.4227.8726.8824.2822.85
s ∈ {2,3,4}33.1529.1427.0224.4723.03
Continuouss ∈ [1,4]33.3029.1727.1124.5823.13
", + "image_path": "a98176b4c2a9c30e0183a7d2e9475a4696c592d0e2ef587a4cbd24a039fe565b.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 287, + 194, + 549, + 244 + ], + "blocks": [ + { + "bbox": [ + 285, + 171, + 551, + 191 + ], + "lines": [ + { + "bbox": [ + 285, + 171, + 551, + 191 + ], + "spans": [ + { + "bbox": [ + 285, + 171, + 551, + 191 + ], + "type": "text", + "content": "Table 7. Comparisons of model size, inference time and performance gain of different models." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 287, + 194, + 549, + 244 + ], + "lines": [ + { + "bbox": [ + 287, + 194, + 549, + 244 + ], + "spans": [ + { + "bbox": [ + 287, + 194, + 549, + 244 + ], + "type": "table", + "html": "
Different modelsMeta-SR [28]LIIF [13]ITSRN [79]LTE [39]CiaoSR
Model size (M)1.71.60.71.71.4
Inference time (ms)237171343148528
PSNR (dB)26.5526.6826.7726.8127.11
Performance gain (dB)-0.060.070.160.20.5
", + "image_path": "70a26f1aae6c8588fa94d0277b1af4cf490ca018537965551de2989e8b4b8c38.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 50, + 273, + 355, + 357 + ], + "blocks": [ + { + "bbox": [ + 47, + 248, + 358, + 269 + ], + "lines": [ + { + "bbox": [ + 47, + 248, + 358, + 269 + ], + "spans": [ + { + "bbox": [ + 47, + 248, + 358, + 269 + ], + "type": "text", + "content": "Table 8. Comparison of more evaluation metrics (SSIM↑ [76] and LPIPS↓ [85]) for different methods on Urban100." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 50, + 273, + 355, + 357 + ], + "lines": [ + { + "bbox": [ + 50, + 273, + 355, + 357 + ], + "spans": [ + { + "bbox": [ + 50, + 273, + 355, + 357 + ], + "type": "table", + "html": "
RDN-MethodIn-scaleOut-of-scale
×2×3×4×6×8
SSIMLPIPSSSIMLPIPSSSIMLPIPSSSIMLPIPSSSIMLPIPS
LIIF [13]0.9340.1040.8660.1970.8040.2670.7040.3600.6360.425
MetaSR [28]0.9350.1040.8660.1950.8050.2610.6980.3530.6240.423
ITSRN [79]0.9360.1020.8680.1930.8070.2590.7030.3500.6330.417
LTE [39]0.9360.1040.8680.1950.8070.2640.7070.3560.6390.421
CiaoSR0.9380.1000.8710.1880.8140.2550.7180.3460.6510.408
", + "image_path": "50a0dd81c943024c1b08292bf6bc853ee4a87d4a4df1f038acda8452604e3352.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 362, + 273, + 549, + 357 + ], + "blocks": [ + { + "bbox": [ + 360, + 249, + 552, + 270 + ], + "lines": [ + { + "bbox": [ + 360, + 249, + 552, + 270 + ], + "spans": [ + { + "bbox": [ + 360, + 249, + 552, + 270 + ], + "type": "text", + "content": "Table 9. Quantitative comparisons of different Realworld SR methods on RealSRSet [82]." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 362, + 273, + 549, + 357 + ], + "lines": [ + { + "bbox": [ + 362, + 273, + 549, + 357 + ], + "spans": [ + { + "bbox": [ + 362, + 273, + 549, + 357 + ], + "type": "table", + "html": "
MethodsScaleNIQE↓BRISQUE↓PIQE↓
RealSR [5]×84.794923.322829.3976
×163.394621.639021.4116
BSRGAN [82]×84.700737.063836.1889
×164.140839.625237.9176
Real-ESRGAN [74]×84.528040.652159.8915
×165.644253.936484.3611
Ours-real×83.589430.921929.8554
×163.946141.814558.8244
", + "image_path": "81a3c1abd1c47c3f43a1d4b723589301e14339ac9eedafba785e43619b2d3cd4.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 47, + 374, + 250, + 387 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 374, + 250, + 387 + ], + "spans": [ + { + "bbox": [ + 47, + 374, + 250, + 387 + ], + "type": "text", + "content": "5.4. Real-World Arbitrary-Scale Image SR" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 46, + 389, + 287, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 389, + 287, + 568 + ], + "spans": [ + { + "bbox": [ + 46, + 389, + 287, + 568 + ], + "type": "text", + "content": "We extend our proposed method to real-world applications with the ultimate goal of synthesizing real images with arbitrary scales. Specifically, we train our model on the DIV2K dataset [1], and test on the RealSRSet [82] and DPED [30] datasets. We apply the practical degradation models (including blur, noise, compression, etc.) of BSRGAN [82] and Real-ESRGAN [74] to synthesize paired training data on DIV2K [1]. We set the HR patch size of training data as 256, and the batch size as 48. We use Adam [36] as the optimizer, and adopt the exponential moving average (EMA). We first train our model with L1 loss for 1000K iterations with the learning rate of " + }, + { + "bbox": [ + 46, + 389, + 287, + 568 + ], + "type": "inline_equation", + "content": "2 \\times 10^{-4}" + }, + { + "bbox": [ + 46, + 389, + 287, + 568 + ], + "type": "text", + "content": " and train with L1 loss, perceptual loss and GAN loss together for 400K iterations with the learning rate of " + }, + { + "bbox": [ + 46, + 389, + 287, + 568 + ], + "type": "inline_equation", + "content": "1 \\times 10^{-4}" + }, + { + "bbox": [ + 46, + 389, + 287, + 568 + ], + "type": "text", + "content": ". More training details can be found in the supplementary materials." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "type": "text", + "content": "Since there are no arbitrary-scale SR methods applied to real-world images, we mainly compare with the state-of-the-art real-world image SR models, including RealSR [5], BSRGAN [82] and Real-ESRGAN [74]. Because there are no ground-truth (GT) real images, we apply non-reference image quality assessment (e.g., NIQE [54], BRISQUE [53] and PIQE [69]) to evaluate the image quality. As shown in Table 9, our model does not optimize for all metrics since these metrics cannot well match the actual human perceptual [4]. To address this, we further compare the visual results on the RealSRSet [82] and DPED [30] datasets in Figure 6. Our proposed method is able to synthesize real" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 375, + 545, + 471 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 375, + 545, + 471 + ], + "spans": [ + { + "bbox": [ + 304, + 375, + 545, + 471 + ], + "type": "text", + "content": "SR images with realistic textures. In particular, our method is able to synthesize more details of a dog in the first line and more clear digit '5' in the second line, compared with other methods. In contrast, other methods often introduce unnatural artifacts within images, especially for out-of-scale distributions. To further investigate the effectiveness of our model, we provide more visual results of our real model in the supplementary materials." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 481, + 378, + 493 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 481, + 378, + 493 + ], + "spans": [ + { + "bbox": [ + 306, + 481, + 378, + 493 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 495, + 547, + 674 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 495, + 547, + 674 + ], + "spans": [ + { + "bbox": [ + 304, + 495, + 547, + 674 + ], + "type": "text", + "content": "In this paper, we have proposed a novel continuous implicit attention-in-attention network for arbitrary-scale image super-resolution, called CiaoSR. Specifically, our CiaoSR first introduces a scale-aware non-local attention in our attention to exploit more relevant features, and then predicts RGB values by locally ensembling features with the learnable attention weights. More importantly, CiaoSR has good flexibility and applicability since it can be used behind any SR backbone to boost the performance. Extensive experiments on all benchmark datasets demonstrate the superiority of our CiaoSR in both SISR and arbitrary-scale SR tasks. Besides, our CiaoSR has good generalization performance on both in-scale and out-of-scale distributions. Last, we extend our method to the real-world application to synthesize real SR images with arbitrary scale." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 305, + 677, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 677, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 305, + 677, + 545, + 712 + ], + "type": "text", + "content": "Acknowledgements: This work was partly supported by the Huawei Fund, the ETH Zürich General Fund (OK) and the Alexander von Humboldt Foundation." + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "type": "text", + "content": "1803" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 135 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 135 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 135 + ], + "type": "text", + "content": "[1] Eirikur Agustsson and Radu Timofte. Ntire 2017 challenge on single image super-resolution: Dataset and study. In IEEE Conference on Computer Vision and Pattern Recognition Workshops, pages 126-135, 2017. 5, 8" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 136, + 288, + 178 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 136, + 288, + 178 + ], + "spans": [ + { + "bbox": [ + 53, + 136, + 288, + 178 + ], + "type": "text", + "content": "[2] Matan Atzmon and Yaron Lipman. Sal: Sign agnostic learning of shapes from raw data. In IEEE Conference on Computer Vision and Pattern Recognition, pages 2565-2574, 2020. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 180, + 288, + 234 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 180, + 288, + 234 + ], + "spans": [ + { + "bbox": [ + 53, + 180, + 288, + 234 + ], + "type": "text", + "content": "[3] Marco Bevilacqua, Aline Roumy, Christine Guillemot, and Marie line Alberi Morel. Low-complexity single-image super-resolution based on nonnegative neighbor embedding. In British Machine Vision Conference, pages 135.1-135.10, 2012. 5, 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 236, + 288, + 280 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 236, + 288, + 280 + ], + "spans": [ + { + "bbox": [ + 53, + 236, + 288, + 280 + ], + "type": "text", + "content": "[4] Yochai Blau, Roey Mechrez, Radu Timofte, Tomer Michaeli, and Lihi Zelnik-Manor. The 2018 pirm challenge on perceptual image super-resolution. In European Conference on Computer Vision Workshops, 2018. 8" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 281, + 288, + 334 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 281, + 288, + 334 + ], + "spans": [ + { + "bbox": [ + 53, + 281, + 288, + 334 + ], + "type": "text", + "content": "[5] Jianrui Cai, Hui Zeng, Hongwei Yong, Zisheng Cao, and Lei Zhang. Toward real-world single image super-resolution: A new benchmark and a new model. In IEEE Conference on International Conference on Computer Vision, pages 3086-3095, 2019. 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 335, + 288, + 368 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 335, + 288, + 368 + ], + "spans": [ + { + "bbox": [ + 53, + 335, + 288, + 368 + ], + "type": "text", + "content": "[6] Jiezhang Cao, Yawei Li, Kai Zhang, and Luc Van Gool. Video super-resolution transformer. arXiv preprint arXiv:2106.06847, 2021. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 369, + 288, + 423 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 369, + 288, + 423 + ], + "spans": [ + { + "bbox": [ + 53, + 369, + 288, + 423 + ], + "type": "text", + "content": "[7] Jiezhang Cao, Jingyun Liang, Kai Zhang, Yawei Li, Yulun Zhang, Wenguan Wang, and Luc Van Gool. Reference-based image super-resolution with deformable attention transformer. In European Conference on Computer Vision, pages 325-342, 2022. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 425, + 288, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 425, + 288, + 479 + ], + "spans": [ + { + "bbox": [ + 53, + 425, + 288, + 479 + ], + "type": "text", + "content": "[8] Jiezhang Cao, Jingyun Liang, Kai Zhang, Wenguan Wang, Qin Wang, Yulun Zhang, Hao Tang, and Luc Van Gool. Towards interpretable video super-resolution via alternating optimization. In European Conference on Computer Vision, pages 393-411, 2022. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 480, + 288, + 524 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 480, + 288, + 524 + ], + "spans": [ + { + "bbox": [ + 53, + 480, + 288, + 524 + ], + "type": "text", + "content": "[9] Lukas Cavigelli, Pascal Hager, and Luca Benini. Cas-cnn: A deep convolutional neural network for image compression artifact suppression. In 2017 International Joint Conference on Neural Networks, pages 752-759, 2017. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 525, + 288, + 579 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 525, + 288, + 579 + ], + "spans": [ + { + "bbox": [ + 48, + 525, + 288, + 579 + ], + "type": "text", + "content": "[10] Rohan Chabra, Jan E Lenssen, Eddy Ig, Tanner Schmidt, Julian Straub, Steven Lovegrove, and Richard Newcombe. Deep local shapes: Learning local sdf priors for detailed 3d reconstruction. In European Conference on Computer Vision, pages 608-625, 2020. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 580, + 288, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 580, + 288, + 635 + ], + "spans": [ + { + "bbox": [ + 48, + 580, + 288, + 635 + ], + "type": "text", + "content": "[11] Hanting Chen, Yunhe Wang, Tianyu Guo, Chang Xu, Yiping Deng, Zhenhua Liu, Siwei Ma, Chunjing Xu, Chao Xu, and Wen Gao. Pre-trained image processing transformer. In IEEE Conference on Computer Vision and Pattern Recognition, pages 12299-12310, 2021. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 635, + 288, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 635, + 288, + 669 + ], + "spans": [ + { + "bbox": [ + 48, + 635, + 288, + 669 + ], + "type": "text", + "content": "[12] Xiangyu Chen, Xintao Wang, Jiantao Zhou, and Chao Dong. Activating more pixels in image super-resolution transformer. arXiv preprint arXiv:2205.04437, 2022. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 670, + 288, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 670, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 670, + 288, + 713 + ], + "type": "text", + "content": "[13] Yinbo Chen, Sifei Liu, and Xiaolong Wang. Learning continuous image representation with local implicit image function. In IEEE Conference on Computer Vision and Pattern Recognition, pages 8628-8638, 2021. 1, 2, 3, 4, 5, 6, 7, 8" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 714 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 307, + 73, + 547, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 547, + 117 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 547, + 117 + ], + "type": "text", + "content": "[14] Yunjin Chen and Thomas Pock. Trainable nonlinear reaction diffusion: A flexible framework for fast and effective image restoration. IEEE Transactions on Pattern Analysis and Machine Intelligence, 39(6):1256-1272, 2016. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 118, + 547, + 151 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 547, + 151 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 547, + 151 + ], + "type": "text", + "content": "[15] Zhiqin Chen and Hao Zhang. Learning implicit fields for generative shape modeling. In IEEE Conference on Computer Vision and Pattern Recognition, pages 5939-5948, 2019. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 152, + 547, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 152, + 547, + 195 + ], + "spans": [ + { + "bbox": [ + 307, + 152, + 547, + 195 + ], + "type": "text", + "content": "[16] Wenlong Cheng, Mingbo Zhao, Zhiling Ye, and Shuhang Gu. Mfagan: A compression framework for memory-efficient on-device super-resolution gan. arXiv preprint arXiv:2107.12679, 2021. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 197, + 546, + 241 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 197, + 546, + 241 + ], + "spans": [ + { + "bbox": [ + 307, + 197, + 546, + 241 + ], + "type": "text", + "content": "[17] Tao Dai, Jianrui Cai, Yongbing Zhang, Shu-Tao Xia, and Lei Zhang. Second-order attention network for single image super-resolution. In IEEE Conference on Computer Vision and Pattern Recognition, pages 11065-11074, 2019. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 242, + 546, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 242, + 546, + 285 + ], + "spans": [ + { + "bbox": [ + 307, + 242, + 546, + 285 + ], + "type": "text", + "content": "[18] Xin Deng, Yutong Zhang, Mai Xu, Shuhang Gu, and Yiping Duan. Deep coupled feedback network for joint exposure fusion and image super-resolution. IEEE Transactions on Image Processing, 30:3098-3112, 2021. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 286, + 547, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 286, + 547, + 331 + ], + "spans": [ + { + "bbox": [ + 307, + 286, + 547, + 331 + ], + "type": "text", + "content": "[19] Chao Dong, Chen Change Loy, Kaiming He, and Xiaou Tang. Learning a deep convolutional network for image super-resolution. In European Conference on Computer Vision, pages 184–199, 2014. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 332, + 547, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 332, + 547, + 376 + ], + "spans": [ + { + "bbox": [ + 307, + 332, + 547, + 376 + ], + "type": "text", + "content": "[20] Xueyang Fu, Menglu Wang, Xiangyong Cao, Xinghao Ding, and Zheng-Jun Zha. A model-driven deep unfolding method for jpeg artifacts removal. IEEE Transactions on Neural Networks and Learning Systems, 2021. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 376, + 547, + 421 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 376, + 547, + 421 + ], + "spans": [ + { + "bbox": [ + 307, + 376, + 547, + 421 + ], + "type": "text", + "content": "[21] Xueyang Fu, Zheng-Jun Zha, Feng Wu, Xinghao Ding, and John Paisley. Jpeg artifacts reduction via deep convolutional sparse coding. In IEEE International Conference on Computer Vision, pages 2501-2510, 2019. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 422, + 546, + 455 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 422, + 546, + 455 + ], + "spans": [ + { + "bbox": [ + 307, + 422, + 546, + 455 + ], + "type": "text", + "content": "[22] Amos Gropp, Lior Yariv, Niv Haim, Matan Atzmon, and Yaron Lipman. Implicit geometric regularization for learning shapes. arXiv preprint arXiv:2002.10099, 2020. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 456, + 547, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 456, + 547, + 498 + ], + "spans": [ + { + "bbox": [ + 307, + 456, + 547, + 498 + ], + "type": "text", + "content": "[23] Shuhang Gu, Nong Sang, and Fan Ma. Fast image super resolution via local regression. In IEEE Conference on International Conference on Pattern Recognition, pages 3128-3131, 2012. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 500, + 546, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 500, + 546, + 544 + ], + "spans": [ + { + "bbox": [ + 307, + 500, + 546, + 544 + ], + "type": "text", + "content": "[24] Bahadir K Gunturk, Yucel Altunbasak, and Russell M Mersereau. Super-resolution reconstruction of compressed video using transform-domain statistics. IEEE Transactions on Image Processing, 13(1):33-43, 2004. 1" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 545, + 547, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 545, + 547, + 600 + ], + "spans": [ + { + "bbox": [ + 307, + 545, + 547, + 600 + ], + "type": "text", + "content": "[25] Yong Guo, Jian Chen, Jingdong Wang, Qi Chen, Jiezhang Cao, Zeshuai Deng, Yanwu Xu, and Mingkui Tan. Closed-loop matters: Dual regression networks for single image super-resolution. In IEEE Conference on Computer Vision and Pattern Recognition, pages 5407-5416, 2020. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 601, + 546, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 601, + 546, + 634 + ], + "spans": [ + { + "bbox": [ + 307, + 601, + 546, + 634 + ], + "type": "text", + "content": "[26] Ali Hassani, Steven Walton, Jiachen Li, Shen Li, and Humphrey Shi. Neighborhood attention transformer. arXiv preprint arXiv:2204.07143, 2022. 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 635, + 547, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 635, + 547, + 678 + ], + "spans": [ + { + "bbox": [ + 307, + 635, + 547, + 678 + ], + "type": "text", + "content": "[27] Kaiming He, Jian Sun, and Xiaou Tang. Single image haze removal using dark channel prior. IEEE Transactions on Pattern Analysis and Machine Intelligence, 33(12):2341-2353, 2010. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 679, + 547, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 679, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 307, + 679, + 547, + 714 + ], + "type": "text", + "content": "[28] Xuecai Hu, Haoyuan Mu, Xiangyu Zhang, Zilei Wang, Tieniu Tan, and Jian Sun. Meta-sr: A magnification-arbitrary network for super-resolution. In IEEE Conference on Computer" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1804" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 714 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 67, + 72, + 288, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 72, + 288, + 94 + ], + "spans": [ + { + "bbox": [ + 67, + 72, + 288, + 94 + ], + "type": "text", + "content": "Vision and Pattern Recognition, pages 1575-1584, 2019. 1, 2, 5, 6, 7, 8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 95, + 288, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 95, + 288, + 139 + ], + "spans": [ + { + "bbox": [ + 48, + 95, + 288, + 139 + ], + "type": "text", + "content": "[29] Jia-Bin Huang, Abhishek Singh, and Narendra Ahuja. Single image super-resolution from transformed self-exemplars. In IEEE Conference on Computer Vision and Pattern Recognition, pages 5197-5206, 2015. 5, 6, 7, 8" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 140, + 288, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 140, + 288, + 194 + ], + "spans": [ + { + "bbox": [ + 48, + 140, + 288, + 194 + ], + "type": "text", + "content": "[30] Andrey Ignatov, Nikolay Kobyshev, Radu Timofte, Kenneth Vanhoey, and Luc Van Gool. Dslr-quality photos on mobile devices with deep convolutional networks. In IEEE International Conference on Computer Vision, pages 3277-3285, 2017. 7, 8" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 194, + 288, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 194, + 288, + 239 + ], + "spans": [ + { + "bbox": [ + 48, + 194, + 288, + 239 + ], + "type": "text", + "content": "[31] Takashi Isobe, Xu Jia, Shuhang Gu, Songjiang Li, Shengjin Wang, and Qi Tian. Video super-resolution with recurrent structure-detail network. In European Conference on Computer Vision, pages 645-660, 2020. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 239, + 288, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 239, + 288, + 282 + ], + "spans": [ + { + "bbox": [ + 48, + 239, + 288, + 282 + ], + "type": "text", + "content": "[32] Xixi Jia, Sanyang Liu, Xiangchu Feng, and Lei Zhang. Focnet: A fractional optimal control network for image denoising. In IEEE Conference on Computer Vision and Pattern Recognition, pages 6054-6063, 2019. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 284, + 288, + 337 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 284, + 288, + 337 + ], + "spans": [ + { + "bbox": [ + 48, + 284, + 288, + 337 + ], + "type": "text", + "content": "[33] Chiyu Jiang, Avneesh Sud, Ameesh Makadia, Jingwei Huang, Matthias Nießner, Thomas Funkhouser, et al. Local implicit grid representations for 3d scenes. In IEEE Conference on Computer Vision and Pattern Recognition, pages 6001-6010, 2020. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 338, + 288, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 338, + 288, + 381 + ], + "spans": [ + { + "bbox": [ + 48, + 338, + 288, + 381 + ], + "type": "text", + "content": "[34] Jiwon Kim, Jung Kwon Lee, and Kyoung Mu Lee. Accurate image super-resolution using very deep convolutional networks. In IEEE Conference on Computer Vision and Pattern Recognition, pages 1646-1654, 2016. 2, 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 383, + 288, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 383, + 288, + 437 + ], + "spans": [ + { + "bbox": [ + 48, + 383, + 288, + 437 + ], + "type": "text", + "content": "[35] Yoonsik Kim, Jae Woong Soh, Jaewoo Park, Byeongyong Ahn, Hyun-Seung Lee, Young-Su Moon, and Nam Ik Cho. A pseudo-blind convolutional neural network for the reduction of compression artifacts. IEEE Transactions on Circuits and Systems for Video Technology, 30(4):1121-1135, 2019. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 437, + 288, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 437, + 288, + 470 + ], + "spans": [ + { + "bbox": [ + 48, + 437, + 288, + 470 + ], + "type": "text", + "content": "[36] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In International Conference on Learning Representations, 2015. 5, 8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 471, + 288, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 471, + 288, + 514 + ], + "spans": [ + { + "bbox": [ + 48, + 471, + 288, + 514 + ], + "type": "text", + "content": "[37] Wei-Sheng Lai, Jia-Bin Huang, Narendra Ahuja, and Ming-Hsuan Yang. Deep laplacian pyramid networks for fast and accurate super-resolution. In IEEE Conference on Computer Vision and Pattern Recognition, pages 624-632, 2017. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 515, + 288, + 581 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 515, + 288, + 581 + ], + "spans": [ + { + "bbox": [ + 48, + 515, + 288, + 581 + ], + "type": "text", + "content": "[38] Christian Ledig, Lucas Theis, Ferenc Huszar, Jose Caballero, Andrew Cunningham, Alejandro Acosta, Andrew Aitken, Alykhan Tejani, Johannes Totz, Zehan Wang, et al. Photorealistic single image super-resolution using a generative adversarial network. In IEEE Conference on Computer Vision and Pattern Recognition, pages 4681-4690, 2017. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 582, + 288, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 582, + 288, + 624 + ], + "spans": [ + { + "bbox": [ + 48, + 582, + 288, + 624 + ], + "type": "text", + "content": "[39] Jaewon Lee and Kyong Hwan Jin. Local texture estimator for implicit representation function. In IEEE Conference on Computer Vision and Pattern Recognition, pages 1929-1938, 2022. 1, 2, 3, 4, 5, 6, 7, 8" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 625, + 288, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 625, + 288, + 679 + ], + "spans": [ + { + "bbox": [ + 48, + 625, + 288, + 679 + ], + "type": "text", + "content": "[40] Jingyun Liang, Jiezhang Cao, Guolei Sun, Kai Zhang, Luc Van Gool, and Radu Timofte. Swinir: Image restoration using swin transformer. In IEEE International Conference on Computer Vision Workshops, pages 1833-1844, 2021. 1, 2, 3, 5, 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 680, + 288, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 680, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 48, + 680, + 288, + 714 + ], + "type": "text", + "content": "[41] Jingyun Liang, Andreas Lugmayr, Kai Zhang, Martin Danelljan, Luc Van Gool, and Radu Timofte. Hierarchical conditional flow: A unified framework for image super-resolution" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 713 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 326, + 73, + 546, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 546, + 95 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 546, + 95 + ], + "type": "text", + "content": "and image rescaling. In IEEE Conference on International Conference on Computer Vision, 2021. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 95, + 547, + 148 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 95, + 547, + 148 + ], + "spans": [ + { + "bbox": [ + 307, + 95, + 547, + 148 + ], + "type": "text", + "content": "[42] Jingyun Liang, Guolei Sun, Kai Zhang, Luc Van Gool, and Radu Timofte. Mutual affine network for spatially variant kernel estimation in blind image super-resolution. In IEEE Conference on International Conference on Computer Vision, 2021. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 150, + 546, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 150, + 546, + 194 + ], + "spans": [ + { + "bbox": [ + 307, + 150, + 546, + 194 + ], + "type": "text", + "content": "[43] Jingyun Liang, Kai Zhang, Shuhang Gu, Luc Van Gool, and Radu Timofte. Flow-based kernel prior with application to blind super-resolution. In IEEE Conference on Computer Vision and Pattern Recognition, pages 10601-10610, 2021. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 194, + 546, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 194, + 546, + 248 + ], + "spans": [ + { + "bbox": [ + 307, + 194, + 546, + 248 + ], + "type": "text", + "content": "[44] Bee Lim, Sanghyun Son, Heewon Kim, Seungjun Nah, and Young Mu Lee. Enhanced deep residual networks for single image super-resolution. In IEEE Conference on Computer Vision and Pattern Recognition Workshops, pages 136-144, 2017. 1, 2, 5, 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 250, + 546, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 250, + 546, + 282 + ], + "spans": [ + { + "bbox": [ + 307, + 250, + 546, + 282 + ], + "type": "text", + "content": "[45] Ding Liu, Bihan Wen, Yuchen Fan, Chen Change Loy, and Thomas S Huang. Non-local recurrent network for image restoration. arXiv preprint arXiv:1806.02919, 2018. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 284, + 547, + 337 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 284, + 547, + 337 + ], + "spans": [ + { + "bbox": [ + 307, + 284, + 547, + 337 + ], + "type": "text", + "content": "[46] David Martin, Charless Fowlkes, Doron Tal, and Jitendra Malik. A database of human segmented natural images and its application to evaluating segmentation algorithms and measuring ecological statistics. In IEEE Conference on International Conference on Computer Vision, pages 416-423, 2001. 5, 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 338, + 547, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 338, + 547, + 392 + ], + "spans": [ + { + "bbox": [ + 307, + 338, + 547, + 392 + ], + "type": "text", + "content": "[47] Yusuke Matsui, Kota Ito, Yuji Aramaki, Azuma Fujimoto, Toru Ogawa, Toshihiko Yamasaki, and Kiyoharu Aizawa. Sketch-based manga retrieval using manga109 dataset. Multi-media Tools and Applications, 76(20):21811-21838, 2017. 5, 6, 7, 8" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 393, + 546, + 436 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 393, + 546, + 436 + ], + "spans": [ + { + "bbox": [ + 307, + 393, + 546, + 436 + ], + "type": "text", + "content": "[48] Yiqun Mei, Yuchen Fan, and Yuqian Zhou. Image superresolution with non-local sparse attention. In IEEE Conference on Computer Vision and Pattern Recognition, pages 3517-3526, 2021. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 437, + 546, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 437, + 546, + 491 + ], + "spans": [ + { + "bbox": [ + 307, + 437, + 546, + 491 + ], + "type": "text", + "content": "[49] Yiqun Mei, Yuchen Fan, Yuqian Zhou, Lichao Huang, Thomas S Huang, and Honghui Shi. Image super-resolution with cross-scale non-local attention and exhaustive self-exemplars mining. In IEEE Conference on Computer Vision and Pattern Recognition, pages 5690-5699, 2020. 4" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 492, + 546, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 492, + 546, + 525 + ], + "spans": [ + { + "bbox": [ + 307, + 492, + 546, + 525 + ], + "type": "text", + "content": "[50] Tomer Michaeli and Michal Irani. Nonparametric blind superresolution. In IEEE Conference on International Conference on Computer Vision, pages 945–952, 2013. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 526, + 547, + 579 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 526, + 547, + 579 + ], + "spans": [ + { + "bbox": [ + 307, + 526, + 547, + 579 + ], + "type": "text", + "content": "[51] Mateusz Michalkiewicz, Jhony K Pontes, Dominic Jack, Mahsa Baktashmotlagh, and Anders Eriksson. Implicit surface representations as layers in neural networks. In IEEE International Conference on Computer Vision, pages 4743-4752, 2019. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 581, + 547, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 581, + 547, + 634 + ], + "spans": [ + { + "bbox": [ + 307, + 581, + 547, + 634 + ], + "type": "text", + "content": "[52] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 635, + 546, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 635, + 546, + 679 + ], + "spans": [ + { + "bbox": [ + 307, + 635, + 546, + 679 + ], + "type": "text", + "content": "[53] Anish Mittal, Anush K Moorthy, and Alan C Bovik. Blind/referenceless image spatial quality evaluator. In Asilomar Conference on Signals, Systems and Computers, pages 723-727, 2011. 8" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 680, + 546, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 680, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 680, + 546, + 713 + ], + "type": "text", + "content": "[54] Anish Mittal, Rajiv Soundararajan, and Alan C Bovik. Making a \"completely blind\" image quality analyzer. IEEE Signal Processing Letters, 20(3):209-212, 2012. 8" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1805" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 288, + 127 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 288, + 127 + ], + "type": "text", + "content": "[55] Michael Niemeyer, Lars Mescheder, Michael Oechsle, and Andreas Geiger. Differentiable volumetric rendering: Learning implicit 3d representations without 3d supervision. In IEEE Conference on Computer Vision and Pattern Recognition, pages 3504-3515, 2020. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 129, + 288, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 129, + 288, + 184 + ], + "spans": [ + { + "bbox": [ + 49, + 129, + 288, + 184 + ], + "type": "text", + "content": "[56] Ben Niu, Weilei Wen, Wenqi Ren, Xiangde Zhang, Lianping Yang, Shuzhen Wang, Kaihao Zhang, Xiaochun Cao, and Haifeng Shen. Single image super-resolution via a holistic attention network. In European Conference on Computer Vision, pages 191-207, 2020. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 185, + 288, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 185, + 288, + 229 + ], + "spans": [ + { + "bbox": [ + 49, + 185, + 288, + 229 + ], + "type": "text", + "content": "[57] Michael Oechsle, Lars Mescheder, Michael Niemeyer, Thilo Strauss, and Andreas Geiger. Texture fields: Learning texture representations in function space. In IEEE International Conference on Computer Vision, pages 4531-4540, 2019. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 231, + 288, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 231, + 288, + 274 + ], + "spans": [ + { + "bbox": [ + 49, + 231, + 288, + 274 + ], + "type": "text", + "content": "[58] Songyou Peng, Michael Niemeyer, Lars Mescheder, Marc Pollefeys, and Andreas Geiger. Convolutional occupancy networks. In European Conference on Computer Vision, pages 523-540, 2020. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 49, + 276, + 288, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 276, + 288, + 319 + ], + "spans": [ + { + "bbox": [ + 49, + 276, + 288, + 319 + ], + "type": "text", + "content": "[59] Yali Peng, Lu Zhang, Shigang Liu, Xiaojun Wu, Yu Zhang, and Xili Wang. Dilated residual networks with symmetric skip connection for image denoising. Neurocomputing, 345:67-76, 2019. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 49, + 321, + 288, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 321, + 288, + 387 + ], + "spans": [ + { + "bbox": [ + 49, + 321, + 288, + 387 + ], + "type": "text", + "content": "[60] Wenzhe Shi, Jose Caballero, Ferenc Huszár, Johannes Totz, Andrew P Aitken, Rob Bishop, Daniel Rueckert, and Zehan Wang. Real-time single image and video super-resolution using an efficient sub-pixel convolutional neural network. In IEEE Conference on Computer Vision and Pattern Recognition, pages 1874-1883, 2016. 1, 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 49, + 388, + 288, + 463 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 388, + 288, + 463 + ], + "spans": [ + { + "bbox": [ + 49, + 388, + 288, + 463 + ], + "type": "text", + "content": "[61] Wenzhe Shi, Jose Caballero, Christian Ledig, Xiahai Zhuang, Wenjia Bai, Kanwal Bhatia, Antonio M Marvao, Tim Dawes, Declan O'Regan, and Daniel Rueckert. Cardiac image superresolution with global correspondence using multi-atlas patchmatch. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 9-16, 2013. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 49, + 465, + 288, + 509 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 465, + 288, + 509 + ], + "spans": [ + { + "bbox": [ + 49, + 465, + 288, + 509 + ], + "type": "text", + "content": "[62] Vincent Sitzmann, Julien Martel, Alexander Bergman, David Lindell, and Gordon Wetzstein. Implicit neural representations with periodic activation functions. Advances in Neural Information Processing Systems, 33:7462-7473, 2020. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 49, + 511, + 288, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 511, + 288, + 555 + ], + "spans": [ + { + "bbox": [ + 49, + 511, + 288, + 555 + ], + "type": "text", + "content": "[63] Vincent Sitzmann, Michael Zollhöfer, and Gordon Wetzstein. Scene representation networks: Continuous 3d-structure-aware neural scene representations. Advances in Neural Information Processing Systems, 32, 2019. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 49, + 556, + 288, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 556, + 288, + 600 + ], + "spans": [ + { + "bbox": [ + 49, + 556, + 288, + 600 + ], + "type": "text", + "content": "[64] Sanghyun Son and Kyoung Mu Lee. Srwarp: Generalized image super-resolution under arbitrary transformation. In IEEE Conference on Computer Vision and Pattern Recognition, pages 7782-7791, 2021. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 49, + 601, + 288, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 601, + 288, + 644 + ], + "spans": [ + { + "bbox": [ + 49, + 601, + 288, + 644 + ], + "type": "text", + "content": "[65] Ying Tai, Jian Yang, Xiaoming Liu, and Chunyan Xu. Memnet: A persistent memory network for image restoration. In IEEE International Conference on Computer Vision, pages 4539-4547, 2017. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 49, + 647, + 288, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 647, + 288, + 689 + ], + "spans": [ + { + "bbox": [ + 49, + 647, + 288, + 689 + ], + "type": "text", + "content": "[66] Radu Timofte, Vincent De Smet, and Luc Van Gool. Anchored neighborhood regression for fast example-based superresolution. In IEEE Conference on International Conference on Computer Vision, pages 1920-1927, 2013. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 49, + 692, + 288, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 692, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 49, + 692, + 288, + 713 + ], + "type": "text", + "content": "[67] Radu Timofte, Vincent De Smet, and Luc Van Gool. A+: Adjusted anchored neighborhood regression for fast super-" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 547, + 713 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "type": "text", + "content": "resolution. In Asian Conference on Computer Vision, pages 111-126, 2014. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 308, + 96, + 547, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 96, + 547, + 140 + ], + "spans": [ + { + "bbox": [ + 308, + 96, + 547, + 140 + ], + "type": "text", + "content": "[68] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Advances in Neural Information Processing Systems, 2017. 2, 3, 4" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 141, + 547, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 141, + 547, + 185 + ], + "spans": [ + { + "bbox": [ + 308, + 141, + 547, + 185 + ], + "type": "text", + "content": "[69] N Venkatanath, D Praneeth, Maruthi Chandrasekhar Bh, Sumohana S Channappayya, and Swarup S Medasani. Blind image quality evaluation using perception based features. In National Conference on Communications, pages 1-6, 2015. 8" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 186, + 547, + 241 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 186, + 547, + 241 + ], + "spans": [ + { + "bbox": [ + 308, + 186, + 547, + 241 + ], + "type": "text", + "content": "[70] Longguang Wang, Yingqian Wang, Xiaoyu Dong, Qingyu Xu, Jungang Yang, Wei An, and Yulan Guo. Unsupervised degradation representation learning for blind super-resolution. In IEEE Conference on Computer Vision and Pattern Recognition, pages 10581-10590, 2021. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 243, + 547, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 243, + 547, + 297 + ], + "spans": [ + { + "bbox": [ + 308, + 243, + 547, + 297 + ], + "type": "text", + "content": "[71] Longguang Wang, Yingqian Wang, Zhengfa Liang, Zaiping Lin, Jungang Yang, Wei An, and Yulan Guo. Learning parallel attention for stereo image super-resolution. In IEEE Conference on Computer Vision and Pattern Recognition, pages 12250-12259, 2019. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 299, + 547, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 299, + 547, + 352 + ], + "spans": [ + { + "bbox": [ + 308, + 299, + 547, + 352 + ], + "type": "text", + "content": "[72] Longguang Wang, Yingqian Wang, Zaiping Lin, Jungang Yang, Wei An, and Yulan Guo. Learning a single network for scale-arbitrary super-resolution. In IEEE Conference on International Conference on Computer Vision, pages 10581-10590, 2021. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 354, + 545, + 399 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 354, + 545, + 399 + ], + "spans": [ + { + "bbox": [ + 308, + 354, + 545, + 399 + ], + "type": "text", + "content": "[73] Longguang Wang, Yingqian Wang, Zaiping Lin, Jungang Yang, Wei An, and Yulan Guo. Learning a single network for scale-arbitrary super-resolution. In IEEE International Conference on Computer Vision, pages 4801-4810, 2021. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 399, + 545, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 399, + 545, + 443 + ], + "spans": [ + { + "bbox": [ + 308, + 399, + 545, + 443 + ], + "type": "text", + "content": "[74] Xintao Wang, Liangbin Xie, Chao Dong, and Ying Shan. Real-esrgan: Training real-world blind super-resolution with pure synthetic data. In International Conference on Computer Vision Workshops. 8" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 445, + 547, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 445, + 547, + 498 + ], + "spans": [ + { + "bbox": [ + 308, + 445, + 547, + 498 + ], + "type": "text", + "content": "[75] Xintao Wang, Ke Yu, Shixiang Wu, Jinjin Gu, Yihao Liu, Chao Dong, Yu Qiao, and Chen Change Loy. Esrgan: Enhanced super-resolution generative adversarial networks. In European Conference on Computer Vision Workshops, pages 701-710, 2018. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 500, + 547, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 500, + 547, + 544 + ], + "spans": [ + { + "bbox": [ + 308, + 500, + 547, + 544 + ], + "type": "text", + "content": "[76] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE Transactions on Image Processing, 13(4):600-612, 2004. 5, 7, 8" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 545, + 547, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 545, + 547, + 600 + ], + "spans": [ + { + "bbox": [ + 308, + 545, + 547, + 600 + ], + "type": "text", + "content": "[77] Yunxuan Wei, Shuhang Gu, Yawei Li, Radu Timofte, Longcun Jin, and Hengjie Song. Unsupervised real-world image super resolution via domain-distance aware training. In IEEE Conference on Computer Vision and Pattern Recognition, pages 13385-13394, 2021. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 601, + 547, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 601, + 547, + 645 + ], + "spans": [ + { + "bbox": [ + 308, + 601, + 547, + 645 + ], + "type": "text", + "content": "[78] Xingqian Xu, Zhangyang Wang, and Humphrey Shi. Ultrasr: Spatial encoding is a missing key for implicit image function-based arbitrary-scale super-resolution. arXiv preprint arXiv:2103.12716, 2021. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 647, + 545, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 647, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 308, + 647, + 545, + 689 + ], + "type": "text", + "content": "[79] Jingyu Yang, Sheng Shen, Huanjing Yue, and Kun Li. Implicit transformer network for screen content image continuous super-resolution. Advances in Neural Information Processing Systems, 34:13304-13315, 2021. 2, 3, 5, 6, 7, 8" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 692, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 692, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 692, + 545, + 713 + ], + "type": "text", + "content": "[80] Roman Zeyde, Michael Elad, and Matan Protter. On single image scale-up using sparse-representations. In International" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "type": "text", + "content": "1806" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 577 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 67, + 72, + 288, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 72, + 288, + 94 + ], + "spans": [ + { + "bbox": [ + 67, + 72, + 288, + 94 + ], + "type": "text", + "content": "Conference on Curves and Surfaces, pages 711-730, 2010. 5, 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 96, + 288, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 96, + 288, + 140 + ], + "spans": [ + { + "bbox": [ + 48, + 96, + 288, + 140 + ], + "type": "text", + "content": "[81] Kai Zhang, Yawei Li, Wangmeng Zuo, Lei Zhang, Luc Van Gool, and Radu Timofte. Plug-and-play image restoration with deep denoiser prior. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2021. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 141, + 288, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 141, + 288, + 184 + ], + "spans": [ + { + "bbox": [ + 48, + 141, + 288, + 184 + ], + "type": "text", + "content": "[82] Kai Zhang, Jingyun Liang, Luc Van Gool, and Radu Timofte. Designing a practical degradation model for deep blind image super-resolution. In IEEE Conference on International Conference on Computer Vision, 2021. 7, 8" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 186, + 288, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 186, + 288, + 228 + ], + "spans": [ + { + "bbox": [ + 48, + 186, + 288, + 228 + ], + "type": "text", + "content": "[83] Kai Zhang, Wangmeng Zuo, and Lei Zhang. Ffdnet: Toward a fast and flexible solution for cnn-based image denoising. IEEE Transactions on Image Processing, 27(9):4608-4622, 2018. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 231, + 288, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 231, + 288, + 274 + ], + "spans": [ + { + "bbox": [ + 48, + 231, + 288, + 274 + ], + "type": "text", + "content": "[84] Kai Zhang, Wangmeng Zuo, and Lei Zhang. Learning a single convolutional super-resolution network for multiple degradations. In IEEE Conference on Computer Vision and Pattern Recognition, pages 3262-3271, 2018. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 276, + 288, + 329 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 276, + 288, + 329 + ], + "spans": [ + { + "bbox": [ + 48, + 276, + 288, + 329 + ], + "type": "text", + "content": "[85] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In IEEE Conference on Computer Vision and Pattern Recognition, pages 586-595, 2018. 5, 7, 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 331, + 288, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 331, + 288, + 374 + ], + "spans": [ + { + "bbox": [ + 48, + 331, + 288, + 374 + ], + "type": "text", + "content": "[86] Yulun Zhang, Kunpeng Li, Kai Li, Lichen Wang, Bineng Zhong, and Yun Fu. Image super-resolution using very deep residual channel attention networks. In European Conference on Computer Vision, pages 286-301, 2018. 1, 2, 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 376, + 288, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 376, + 288, + 408 + ], + "spans": [ + { + "bbox": [ + 48, + 376, + 288, + 408 + ], + "type": "text", + "content": "[87] Yulun Zhang, Kunpeng Li, Kai Li, Bineng Zhong, and Yun Fu. Residual non-local attention networks for image restoration. arXiv preprint arXiv:1903.10082, 2019. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 410, + 288, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 410, + 288, + 453 + ], + "spans": [ + { + "bbox": [ + 48, + 410, + 288, + 453 + ], + "type": "text", + "content": "[88] Yulun Zhang, Yapeng Tian, Yu Kong, Bineng Zhong, and Yun Fu. Residual dense network for image super-resolution. In IEEE Conference on Computer Vision and Pattern Recognition, pages 2472-2481, 2018. 1, 2, 5, 6, 7, 8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 455, + 288, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 455, + 288, + 498 + ], + "spans": [ + { + "bbox": [ + 48, + 455, + 288, + 498 + ], + "type": "text", + "content": "[89] Yulun Zhang, Yapeng Tian, Yu Kong, Bineng Zhong, and Yun Fu. Residual dense network for image restoration. IEEE Transactions on Pattern Analysis and Machine Intelligence, 43(7):2480-2495, 2020. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 500, + 288, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 500, + 288, + 542 + ], + "spans": [ + { + "bbox": [ + 48, + 500, + 288, + 542 + ], + "type": "text", + "content": "[90] Shangchen Zhou, Jiawei Zhang, Wangmeng Zuo, and Chen Change Loy. Cross-scale internal graph neural network for image super-resolution. arXiv preprint arXiv:2006.16673, 2020.2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 544, + 288, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 544, + 288, + 577 + ], + "spans": [ + { + "bbox": [ + 48, + 544, + 288, + 577 + ], + "type": "text", + "content": "[91] Wilman WW Zou and Pong C Yuen. Very low resolution face recognition problem. IEEE Transactions on Image Processing, 21(1):327-340, 2011. 1" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1807" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Class Adaptive Network Calibration/d3a59c4a-62e4-447c-b052-c6b11b8ed279_content_list.json b/2023/Class Adaptive Network Calibration/d3a59c4a-62e4-447c-b052-c6b11b8ed279_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..9cbcb3bbfb3948391ba5ff0f5f0d9e1dcb272111 --- /dev/null +++ b/2023/Class Adaptive Network Calibration/d3a59c4a-62e4-447c-b052-c6b11b8ed279_content_list.json @@ -0,0 +1,1629 @@ +[ + { + "type": "text", + "text": "Class Adaptive Network Calibration", + "text_level": 1, + "bbox": [ + 300, + 130, + 669, + 151 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Bingyuan Liu\\*, Jérôme Rony\\*, Adrian Galdran\\*, Jose Dolz\\*, Ismail Ben Ayed", + "bbox": [ + 166, + 179, + 799, + 199 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1ÉTS Montreal, Canada 2Universitat Pompeu Fabra, Barcelona, Spain", + "bbox": [ + 192, + 209, + 769, + 229 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 263, + 313, + 277 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recent studies have revealed that, beyond conventional accuracy, calibration should also be considered for training modern deep neural networks. To address miscalibration during learning, some methods have explored different penalty functions as part of the learning objective, alongside a standard classification loss, with a hyper-parameter controlling the relative contribution of each term. Nevertheless, these methods share two major drawbacks: 1) the scalar balancing weight is the same for all classes, hindering the ability to address different intrinsic difficulties or imbalance among classes; and 2) the balancing weight is usually fixed without an adaptive strategy, which may prevent from reaching the best compromise between accuracy and calibration, and requires hyper-parameter search for each application. We propose Class Adaptive Label Smoothing (CALS) for calibrating deep networks, which allows to learn class-wise multipliers during training, yielding a powerful alternative to common label smoothing penalties. Our method builds on a general Augmented Lagrangian approach, a well-established technique in constrained optimization, but we introduce several modifications to tailor it for large-scale, class-adaptive training. Comprehensive evaluation and multiple comparisons on a variety of benchmarks, including standard and long-tailed image classification, semantic segmentation, and text classification, demonstrate the superiority of the proposed method. The code is available at https://github.com/by-liu/CALS.", + "bbox": [ + 75, + 295, + 472, + 704 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 733, + 207, + 748 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Deep Neural Networks (DNNs) have become the prevailing model in machine learning, particularly for computer vision [13] and natural language processing applications [44]. Increasingly powerful architectures [3, 13, 24], learning methods [4, 12] and a large body of other techniques [15, 27] are constantly introduced. Nonetheless, recent studies [11, 31] have shown that regardless of their superior discriminative", + "bbox": [ + 75, + 758, + 470, + 864 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "performance, high-capacity modern DNNs are poorly calibrated, i.e. failing to produce reliable predictive confidences. Specifically, they tend to yield over-confident predictions, where the probability associated with the predicted class overestimates the actual likelihood. Since this is a critical issue in safety-sensitive applications like autonomous driving or computational medical diagnosis, the problem of DNN calibration has been attracting increasing attention in recent years [11, 31, 38].", + "bbox": [ + 500, + 263, + 893, + 400 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Current calibration methods can be categorized into two main families. The first family involves techniques that perform an additional post-processing parameterized operation on the output logits (or pre-softmax activations) [11], with the calibration parameters of that operation obtained from a validation set by either learning or grid-search. Despite the simplicity and low computational cost, these methods have empirically proven to be highly effective [8, 11]. However, their main drawback is that the choice of the optimal calibration parameters is highly sensitive to the trained model instance and validation set [22, 31].", + "bbox": [ + 496, + 401, + 893, + 566 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The second family of methods attempts to simultaneously optimize for accuracy and calibration during network training. This is achieved by introducing, explicitly or implicitly, a secondary optimization goal involving the model's predictive uncertainty, alongside the main training objective. As a result, a scalar balancing hyper-parameter is required to tune the relative contribution of each term in the overall loss function. Some examples of this type of approaches include: Explicit Confidence Penalty (ECP) [38], Label Smoothing (LS) [32], Focal Loss (FL) [21] and its variant, Sample-Dependent Focal Loss (FLSD) [31]. It has been recently demonstrated in [22] that all these methods can be formulated as different penalty terms that enforce the same equality constraint on the logits of the DNN: driving the logit distances towards zero. Here, logit distances refers to the vector of L1 distances between the highest logit value and the rest. Observing the non-informative nature of this equality constraint, [22] proposed to use a generalized inequality constraint, only penalizing those logits for which the distance is larger than a pre-defined margin, achieving state-of-the-art calibration performance on many different benchmarks.", + "bbox": [ + 496, + 566, + 895, + 883 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Although learning based methods achieve greater calibra", + "bbox": [ + 519, + 885, + 893, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Equal Contributions. Correspondence to: {liubingyuan1988@gmail.com, jerome.rony.1@etsmtl.net}", + "bbox": [ + 76, + 875, + 467, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "16070", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/efbbb40bcc5c3e8826b7a9fcd227fec317cc07d4a7186c95cc9452b851551f17.jpg", + "image_caption": [ + "Figure 1. Many techniques have been proposed for jointly improving accuracy and calibration during training [11,31], but they fail to consider uneven learning scenarios like high class imbalance or long-tail distributions. We show a comparison of the proposed CALS-ALM method and different learning approaches in terms of Calibration Error (ECE) vs Accuracy on the (a) ImageNet and (b) ImageNet-LT (long-tailed ImageNet) datasets. A lower ECE indicates better calibration: a better model should attain high ACC and low ECE. Among all the considered methods, CALS-ALM shows superior performance when considering both discriminative power and well-balanced probabilistic predictions, achieving best accuracy and calibration on ImageNet, and best calibration and second best accuracy on ImageNet-LT." + ], + "image_footnote": [], + "bbox": [ + 94, + 85, + 488, + 289 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/f3a4a14c56f26e6becfa638ee582811403c0d1e8809aa9037ba042ebd4abcdfd.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 496, + 85, + 874, + 289 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "tion performance [22, 31], they have two major limitations: 1) The scalar balancing weight is equal for all classes. This hinders the network performance when some classes are harder to learn or less represented than others, such as in datasets with a large number of categories (ImageNet) or considerable class imbalance (ImageNet-LT). 2) The balancing weight is usually fixed before network optimization, with no learning or adaptive strategy throughout training. This can prevent the model from reaching the best compromise between accuracy and calibration. To address the above issues, we introduce Class Adaptive Label Smoothing method based on an Augmented Lagrangian Multiplier algorithm, which we refer to as CALS-ALM. Our Contributions can be summarized as follows:", + "bbox": [ + 75, + 410, + 472, + 621 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose Class Adaptive Label Smoothing (CALS) for network calibration. Adaptive class-wise multipliers are introduced instead of the widely used single balancing weight, which addresses the above two issues: 1) CALS can handle a high number of classes with different intrinsic difficulties, e.g. ImageNet; 2) CALS can effectively learn from data suffering from class imbalance or a long-tailed distribution, e.g. ImageNet-LT.", + "- Different from previous penalty based methods, we solve the resulting constrained optimization problem by implementing a modified Augmented Lagrangian Multiplier (ALM) algorithm, which yields adaptive and optimal weights for the constraints. We make some critical design decisions in order to adapt ALM to the nature of modern learning techniques: 1) The inner convergence criterion in ALM is relaxed to a fixed number of iterations in each inner stage, which is amenable to mini-batch stochas" + ], + "bbox": [ + 76, + 632, + 472, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "tic gradient optimization in deep learning. 2) Popular techniques, such as data augmentation, batch normalization [15] and dropout [10], rule out the possibility of tracking original samples and applying sample-wise multipliers. To overcome this complication, we introduce class-wise multipliers, instead of sample-wise multipliers in the standard ALM. 3) The outer-step update for estimating optimal ALM multipliers is performed on the validation set, which is meaningful for training on large-scale training set and avoids potential overfitting.", + "bbox": [ + 511, + 410, + 893, + 561 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "- Comprehensive experiments over a variety of applications and benchmarks, including standard image classification (Tiny-ImageNet and ImageNet), long-tailed image classification (ImageNet-LT), semantic segmentation (PASCAL VOC 2012), and text classification (20 Newsgroups), demonstrate the effectiveness of our CALS-ALM method. As shown in Figure 1, CALS-ALM yields superior performance over baselines and state-of-the-art calibration losses when considering both accuracy and calibration, especially for more realistic large-scaled datasets with large number of classes or class imbalance.", + "bbox": [ + 500, + 568, + 895, + 734 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 746, + 640, + 762 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Problem Formulation", + "text_level": 1, + "bbox": [ + 500, + 771, + 702, + 786 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Given a dataset $\\mathcal{D} = \\{(x^{(i)},y^{(i)})\\}_{i = 1}^{N}$ with $N\\in \\mathbb{N}$ pairs of samples $\\pmb {x}\\in \\mathcal{X}$ and corresponding labels $y\\in \\mathcal{V}$ with $\\mathcal{V} = \\{1,\\dots ,K\\}$ , a deep neural network (DNN) $F_{\\theta}:X\\to \\mathbb{R}^{K}$ parameterized by $\\theta$ yields logits $l = F_{\\theta}(\\pmb {x}) = (F_{\\theta}(\\pmb {x})_k)_{1\\leq k\\leq K}\\in \\mathbb{R}^K$ . In a classification scenario, the output probability $s = (s_k)_{1\\leq k\\leq K}\\in \\Delta^{K - 1}$ where $\\Delta^{K - 1}\\subset [0,1]^K$ denotes the probability simplex, is", + "bbox": [ + 496, + 792, + 895, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "16071", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "obtained by applying the softmax function on the output log-its, i.e. $s = \\text{softmax}(l) = \\frac{\\exp l}{\\sum \\exp l}$ . Therefore, the predicted class $\\hat{y}$ is computed as $\\hat{y} = \\arg \\max_k s_k$ , and the predicted confidence is $\\hat{p} = s_{\\hat{y}} = \\max_k s_k$ . A perfectly calibrated model should satisfy that the predicted confidence of any input is equal to the accuracy of the model: $\\hat{p} = \\mathbb{P}(\\hat{y} = y|\\hat{p})$ . Hence, an over-confident model yields on average larger confidence than the associated accuracy, whereas an under-confident model yields lower confidence than its accuracy.", + "bbox": [ + 75, + 90, + 472, + 227 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "A number of recent studies [11, 22, 30, 31] have shown that DNNs tend to become over-confident during training as a result of minimizing the popular cross-entropy (CE) training loss:", + "bbox": [ + 75, + 229, + 470, + 291 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {C E}} (\\boldsymbol {x}, y) = - \\sum_ {k = 1} ^ {K} y _ {k} \\log s _ {k} \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 176, + 303, + 470, + 347 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\pmb{y} \\in \\{0,1\\}^{K}$ is the one-hot encoding of $y$ . This objective function is minimized when the predictions for all the training samples fully match the ground-truth labels $y$ , i.e. $s_{y} = 1$ and $\\forall k \\neq y, s_{k} = 0$ . The negative logarithmic term on the logit of the correct category renders the global minimization of the CE loss unreachable, as it keeps pushing the predicted probabilities $s$ towards the vertices of the $(K-1)$ -simplex even after the classification error is zero [31], resulting in over-confident models.", + "bbox": [ + 75, + 358, + 472, + 494 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. Post-processing methods", + "text_level": 1, + "bbox": [ + 76, + 507, + 302, + 523 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To address mis-calibration, different post-processing techniques applied after model training have been proposed recently [8, 11, 43]. The most popular of these strategies is Temperature Scaling (TS) [11], which applies a single scalar temperature parameter to manipulate logit outputs monotonely, resulting in softened prediction confidences without affecting predicted labels. Note that here the temperature parameter needs to be tuned on a separate validation set. Despite its simplicity, TS has been shown effective in fixing over-confidence predictions [11]. As a local alternative to TS, [8] proposed to train a regression model for learning position-specific temperature for semantic segmentation problems. Unfortunately, TS and its variants can be sensitive to both the model and the validation set, and do not work well under data distribution shifts [36]. Thus, some subsequent works [28, 43] have attempted to provide solutions for improving performance under domain shift.", + "bbox": [ + 75, + 531, + 472, + 789 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.3. Learning-based methods", + "text_level": 1, + "bbox": [ + 76, + 800, + 303, + 816 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Another popular direction is to directly address miscalibration during training by introducing an additional penalty or supervision regarding model calibration with the standard training loss. In [18], the authors introduced a trainable calibration measure based on RKHS kernels, while [16]", + "bbox": [ + 75, + 824, + 470, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "proposed a differential calibration loss based on a soft version of the binning operation in the ECE metric. In [5], two types of binary pairwise calibration constraints were proposed as additional penalty terms during training. Other methods try to decrease over-fitting on the cross-entropy loss, which has been demonstrated to be the main reason of over-confidence [11, 31]. In [38] an explicit confidence penalty (ECP) is proposed to maximize the entropy and reduce over-fitting, while Label Smoothing [42] has also been shown to implicitly improve the calibration [32] by softening the hard one-hot targets in the cross-entropy. The Focal Loss [21], originally proposed to tackle class imbalance, can also be effective for calibration [31], as it implicitly minimizes the Kullback-Leibler (KL) divergence between the uniform distribution and the network softmax probabilities, thereby increasing the entropy of the predictions. As an extension the Sample-Dependent Focal Loss (FLSD) was also proposed in [31] to further boost calibration performance.", + "bbox": [ + 496, + 90, + 893, + 364 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Margin-based Label Smoothing (MbLS). A unifying constrained-optimization formulation of loss functions promoting calibration has been recently presented in [22]. Specifically, the additional penalties integrated in these methods, including ECP [38], LS [32] and FL [31], can be viewed as different forms of approximations to the same constraint, i.e. enforcing the logit distances to be zero. Noticing that this constraint is non-informative (its solution being uniformly distributed probabilities), [22] further proposed a generalized formulation by relaxing the constraint to allow the logit distances being lower than a strictly positive margin.", + "bbox": [ + 496, + 387, + 893, + 553 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The specific formulation of MbLS [22] is as follows. Given a margin $m \\in \\mathbb{R}_+$ , the constrained optimization problem for network training is:", + "bbox": [ + 496, + 554, + 893, + 599 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {\\theta} \\sum_ {i = 1} ^ {N} \\mathcal {L} _ {\\mathrm {C E}} \\left(\\boldsymbol {x} ^ {(i)}, y ^ {(i)}\\right) \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 519, + 613, + 890, + 654 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\text {s . t .} \\quad \\max _ {k} \\left\\{\\boldsymbol {l} _ {k} ^ {(i)} \\right\\} - l ^ {(i)} \\preceq m \\mathbf {1} _ {K}, \\quad i = 1, \\dots , N,\n$$\n", + "text_format": "latex", + "bbox": [ + 522, + 656, + 851, + 681 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $l^{(i)} = F_{\\theta}(\\pmb{x}^{(i)})$ . The minimum can be approximated by penalty-based optimization methods, transforming the above formulation into an unconstrained problem by means of simple ReLU functions:", + "bbox": [ + 496, + 695, + 890, + 757 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {\\theta} \\sum_ {i = 1} ^ {N} \\mathcal {L} _ {\\mathrm {C E}} \\left(\\boldsymbol {x} ^ {(i)}, y ^ {(i)}\\right) + \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 539, + 770, + 890, + 820 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\lambda \\sum_ {i = 1} ^ {N} \\sum_ {j = 1} ^ {K} \\max \\{0, \\max _ {k} \\{l _ {k} ^ {(i)} \\} - l _ {j} ^ {(i)} - m \\},\n$$\n", + "text_format": "latex", + "bbox": [ + 575, + 815, + 851, + 857 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\lambda \\in \\mathbb{R}_+$ is a scalar weight balancing contributions of the CE loss and the corresponding penalty.", + "bbox": [ + 500, + 869, + 890, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "16072", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Sample-wise Constrained DNN Optimization", + "text_level": 1, + "bbox": [ + 76, + 89, + 468, + 107 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Although MbLS can significantly improve calibration, the associated constrained problem (2) is not solved accurately. It is approximated by an unconstrained problem with a single uniform penalty, regardless of the data sample or category. However, the samples and classes considered in a classification problem have different intrinsic learning difficulties. Therefore, an improved training scheme would involve considering distinct penalty weights $\\lambda$ for each sample and class. This would result in having to chose $N\\times K$ penalty weights $\\Lambda \\in \\mathbb{R}_{+}^{N\\times K}$ , with the resulting optimization problem being:", + "bbox": [ + 75, + 114, + 472, + 267 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\min _ {\\theta} \\sum_ {i = 1 \\atop N = K} ^ {N} \\mathcal {L} _ {\\mathrm {C E}} \\left(\\boldsymbol {x} ^ {(i)}, y ^ {(i)}\\right) + \\tag {4} \\\\ \\sum_ {i = 1} ^ {N} \\sum_ {j = 1} ^ {K} \\Lambda_ {i j} \\max \\{0, \\max _ {k} \\{l _ {k} ^ {(i)} \\} - l _ {j} ^ {(i)} - m \\}. \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 94, + 276, + 468, + 363 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "From an optimization perspective, supposing that optimal weights $\\theta^{\\star}$ exists for problem (2), there exists $\\Lambda^{\\star} \\in \\mathbb{R}_{+}^{N \\times K}$ such that $(\\theta^{\\star}, \\Lambda^{\\star})$ is a saddle point of the Lagrangian associated to (2). These $\\Lambda^{\\star}$ are the Lagrange multipliers of the problem. Therefore, using $\\Lambda = \\Lambda^{\\star}$ would be the best choice to solve (4).", + "bbox": [ + 75, + 372, + 468, + 462 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In practice, using the Lagrange multipliers of problem (2) as the weights for the penalties may not be computationally feasible, and it could even result in degraded performance. Indeed, in the context of machine learning, we optimize a model's weights $\\theta$ to solve (2) on a training set $\\mathcal{D}_{\\mathrm{train}}$ , and expect to generalize on a test set $\\mathcal{D}_{\\mathrm{test}}$ , which we do not have access to during training. Because of the bias-variance tradeoff, solving (2) optimally would likely result in overfitting, i.e. we may solve problem (2) accurately on the train set, but not generalize properly on the test set, resulting in poor calibration and classification performance overall. This suggests that it could be preferable to evaluate during training the quality of multipliers on a separate validation set $\\mathcal{D}_{\\mathrm{val}}$ . Additionally, several mechanisms for training DNNs are not compatible with a straightforward minimization. First, the use of batch normalization yields predictions that are not independent between samples in a minibatch. Second, the use of regularization techniques such as dropout may lead to virtually inaccurate predictions on certain training samples, impacting the correct estimation of multipliers. Third, data augmentation, which is standard in DNN training, would result in additional penalty weights for the augmented samples: they can be easier or harder to classify than the original ones.", + "bbox": [ + 75, + 463, + 472, + 824 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In addition to the above obstacles, applying a penalty weight per sample and per class (resulting in $N \\times K$ weights) would not scale well for large datasets and dense predictive tasks, such as semantic segmentation, which is typically formulated as a per-pixel classification task. Assuming that im", + "bbox": [ + 75, + 825, + 472, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "ages in the dataset have a size of $H \\times W$ , this would result in $N \\times H \\times W \\times K$ penalty weights. This rapidly becomes a limiting factor for moderately sized segmentation datasets. For instance, Pascal VOC 2012 [9] contains 21 classes and 1464 training images, amounting to $2.62 \\times 10^{8}$ pixels, or $5.5 \\times 10^{9}$ penalty weights, which, stored as float32, would use $\\sim 20$ GiB. For Cityscapes [6], containing approximately 3000 training images of size $2048 \\times 1024$ in 19 classes, this amounts to $\\sim 445$ GiB.", + "bbox": [ + 496, + 90, + 893, + 226 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Following the above observations, we introduce a relaxation of sample-wise penalties, and propose to solve the following problem:", + "bbox": [ + 496, + 227, + 893, + 273 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\min _ {\\theta} \\sum_ {i = 1 \\atop N, K} ^ {N} \\mathcal {L} _ {\\mathrm {C E}} \\left(\\boldsymbol {x} ^ {(i)}, y ^ {(i)}\\right) + \\tag {5} \\\\ \\sum_ {i = 1} ^ {N} \\sum_ {j = 1} ^ {K} \\lambda_ {j} \\max \\{0, \\max _ {k} \\{l _ {k} ^ {(i)} \\} - l _ {j} ^ {(i)} - m \\}, \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 519, + 282, + 890, + 371 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $(\\lambda_j)_{1\\leq j\\leq K}\\in \\mathbb{R}_+^K$ . Since penalties are now class-wise, we need $K$ penalty weights. This has the advantage to scale well to denser classification tasks such as segmentation. However, we still face a challenging optimization problem (5), since we still need to chose $K$ weights, which can be extremely complicated for large-scale datasets with many classes such as ImageNet, which contains 1000 classes. In the next section, we introduce a numerical technique to deal with this challenge.", + "bbox": [ + 496, + 383, + 893, + 521 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Class Adaptive Network Calibration", + "text_level": 1, + "bbox": [ + 498, + 535, + 828, + 551 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The challenge of the previous formulation stems from correctly choosing the weights $\\lambda \\in \\mathbb{R}_+^K$ , which can be cumbersome when $K$ is large. Therefore, we propose to use an Augmented Lagrangian Multiplier (ALM) method to adaptively learn the weights of the penalties.", + "bbox": [ + 496, + 560, + 893, + 636 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1. General ALM", + "text_level": 1, + "bbox": [ + 500, + 646, + 643, + 660 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "ALM methods combine penalties and primal-dual updates to solve a constrained problem. They have well-established advantages and enjoy widespread popularity in the general context of optimization [1,35,41]. Specifically, we have the following generic constrained optimization problem:", + "bbox": [ + 496, + 670, + 890, + 746 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {x} f (x) \\quad \\text {s . t .} \\quad h _ {i} (x) \\leq 0, \\quad i = 1, \\dots , n \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 535, + 758, + 890, + 779 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $f: \\mathbb{R}^d \\to \\mathbb{R}$ is the objective function and $h_i: \\mathbb{R}^d \\to \\mathbb{R}, i = 1, \\dots, n$ are the constraint functions. We tackle it by approximately solving a sequence $j \\in \\mathbb{N}$ of unconstrained problems:", + "bbox": [ + 496, + 791, + 890, + 853 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {x} \\quad \\mathcal {L} ^ {(j)} (x) = f (x) + \\sum_ {i = 1} ^ {n} P \\left(h _ {i} (x), \\rho_ {i} ^ {(j)}, \\lambda_ {i} ^ {(j)}\\right) \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 519, + 864, + 890, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "16073", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/77ae33c1b29ebad9fab4420309eee8c8a6c20490d3a1172073de4002ba940f03.jpg", + "image_caption": [ + "Figure 2. A penalty-Lagrangian function $P$ with varying values of $\\rho$ and $\\mu$ . Higher values of $\\rho$ bring $P$ closer to an ideal penalty. The multiplier $\\lambda$ is the derivative of $P$ w.r.t. the constraint at $z = 0$ ." + ], + "image_footnote": [], + "bbox": [ + 107, + 90, + 444, + 250 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "with $P: \\mathbb{R} \\times \\mathbb{R}_{++}^{n} \\times \\mathbb{R}_{++}^{n} \\to \\mathbb{R}$ a penalty-Lagrangian function, and $\\pmb{\\rho}^{(j)} = (\\rho_i)_{1 \\leq i \\leq n} \\in \\mathbb{R}_{++}^n$ , $\\pmb{\\lambda}^{(j)} = (\\lambda_i^{(j)})_{1 \\leq i \\leq n} \\in \\mathbb{R}_{++}^n$ the penalty parameters and multipliers associated to $P$ at the $j$ -th iteration. This sequence of unconstrained problems is called outer iterations, while the steps in the minimization of $\\mathcal{L}^{(j)}$ are called inner iterations.", + "bbox": [ + 75, + 329, + 470, + 421 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The main components of ALM methods are (i) the penalty-Lagrangian function $P$ , (ii) the update of the penalty multipliers $\\lambda^{(j)}$ and (iii) the increase of the penalty parameters $\\pmb{\\rho}^{(j)}$ . First, the penalty function $P$ needs to satisfy a set of axioms [2] (see Appendix A): these axioms constrain the function to be continuously differentiable w.r.t. its first variable and to have a non-negative derivative: $\\forall z \\in \\mathbb{R}, P'(z, \\rho, \\lambda) = \\frac{\\partial}{\\partial z} P(z, \\rho, \\lambda) \\geq 0$ , with $P'(0, \\rho, \\lambda) = \\lambda$ . Figure 2 gives an example of a penalty, and how $\\rho$ and $\\lambda$ affect it. The choice of penalty function is critical to the performance of ALM methods, especially for nonconvex problems [2]. Typical functions include PHR [14,39], $P_{2}$ [17] and $P_{3}$ [34] (see section 3.2 of [2]). Second, the penalty multipliers $\\lambda^{(j)}$ are updated to the derivative of $P$ w.r.t. the solution obtained during the last inner minimization. Formally, let $x^{(j)}$ be the approximate minimizer of $\\mathcal{L}^{(j)}$ , then $\\forall i \\in \\{1, \\dots, n\\}$ :", + "bbox": [ + 75, + 422, + 470, + 680 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\lambda_ {i} ^ {(j + 1)} = P ^ {\\prime} \\left(h _ {i} \\left(x ^ {(j)}\\right), \\rho_ {i} ^ {(j)}, \\lambda_ {i} ^ {(j)}\\right) \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 161, + 688, + 468, + 709 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "This update rule corresponds to a first-order multiplier estimate for the constrained problem. Third, the penalty parameters $\\rho^{(j)}$ are increased during the outer iterations if the constraints do not improve (i.e. is closer to being satisfied) compared to the previous outer iteration. Typically, $\\rho_{i}^{(j + 1)} = \\gamma \\rho_{i}^{(j)}$ if $h_i$ does not improve, with $\\gamma > 1$ .", + "bbox": [ + 75, + 717, + 468, + 810 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "When the problem is convex, alternating between the approximate minimization of (7) and the update of the multipliers (8) leads to a solution for the constrained problem. The inner minimization corresponds to minimizing the primal while the outer iterations correspond to solving the dual problem. The complete procedure is presented in Algorithm 1.", + "bbox": [ + 75, + 810, + 470, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Although guarantees exist only in the convex case, it is well-known that ALM methods can efficiently solve nonconvex problems as well [2]. In the context of deep learning, their use has been surprisingly underexplored [40, 41].", + "bbox": [ + 496, + 90, + 893, + 151 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Algorithm 1 Augmented Lagrangian Multiplier algorithm", + "text_level": 1, + "bbox": [ + 500, + 166, + 885, + 181 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Require: Objective function $f$", + "bbox": [ + 500, + 185, + 689, + 199 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Require: Constraint functions $h_i, i = 1, \\dots, n$", + "bbox": [ + 500, + 200, + 785, + 213 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Require: Penalty function $P$ , initial $\\pmb{\\lambda}^{(0)} \\in \\mathbb{R}_{++}^{n}$ , $\\pmb{\\rho}^{(0)} \\in \\mathbb{R}_{++}^{n}$", + "bbox": [ + 500, + 213, + 877, + 227 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Require: Initial variable $x^{(0)}$ , iterations $j = 1$", + "bbox": [ + 500, + 227, + 781, + 241 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} ^ {(j)} (x) = f (x) + \\sum_ {i = 1} ^ {n} P \\left(h _ {i} (x), \\rho_ {i} ^ {(j)}, \\lambda_ {i} ^ {(j)}\\right)\n$$\n", + "text_format": "latex", + "bbox": [ + 542, + 268, + 816, + 284 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1: while not converged do", + "2: Initialize with $x^{(j - 1)}$ and minimize (approximately):", + "3: $x^{(j)}\\gets$ (approximate) minimizer of $\\mathcal{L}^{(j)}$", + "4: for $i = 1,\\dots ,n$ do", + "5: $\\lambda_{i}^{(j + 1)}\\gets P^{\\prime}(h_{i}(x^{(j)}),\\rho_{i}^{(j)},\\lambda_{i}^{(j)})$", + "6: if the $i$ -th constraint does not improve then", + "7: $\\rho_{i}^{(j + 1)}\\gets \\gamma \\rho_{i}^{(j)}$", + "8: else", + "9: $\\rho_{i}^{(j + 1)}\\gets \\rho_{i}^{(j)}$", + "10: end if", + "11: end for", + "12: $j\\gets j + 1$", + "13: end while" + ], + "bbox": [ + 503, + 241, + 856, + 436 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. ALM for calibration", + "text_level": 1, + "bbox": [ + 500, + 467, + 692, + 481 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our goal now is to build an ALM method effective for calibration purposes. We can achieve this by reformulating problem (5) using a penalty function $P$ parameterized by $(\\pmb {\\rho},\\pmb {\\lambda})\\in \\mathbb{R}_{++}^{K}\\times \\mathbb{R}_{++}^{K}$ as follows:", + "bbox": [ + 496, + 489, + 890, + 551 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {\\theta} \\quad \\sum_ {i = 1} ^ {N} \\mathcal {L} _ {\\mathrm {C E}} \\left(\\boldsymbol {x} ^ {(i)}, y ^ {(i)}\\right) + \\sum_ {k = 1} ^ {K} P \\left(d _ {k} ^ {(i)} - m, \\rho_ {k}, \\lambda_ {k}\\right) \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 509, + 564, + 890, + 604 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $d_k^{(i)} = \\max \\{l^{(i)}\\} - l_k^{(i)} \\in \\mathbb{R}_+$ . With this formulation, it is natural to use a penalty-Lagrangian function for $P$ . To avoid numerical issues typically associated with non-linear penalties, we normalize the constraints by the margin $m > 0$ :", + "bbox": [ + 496, + 617, + 893, + 681 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nd _ {k} ^ {(i)} - m \\leq 0 \\Leftrightarrow \\frac {d _ {k} ^ {(i)}}{m} - 1 \\leq 0 \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 593, + 693, + 890, + 726 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "This leads to improved numerical stability for the ALM multiplier update as well. Additionally, we average the constraints instead of summing them. This makes the method independent of the number of classes, and eases the choice of initial penalty parameters $\\rho^{(0)}$ . The resulting loss is:", + "bbox": [ + 496, + 729, + 890, + 806 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\sum_ {i = 1} ^ {N} \\mathcal {L} _ {\\mathrm {C E}} \\left(\\boldsymbol {x} ^ {(i)}, y ^ {(i)}\\right) + \\frac {1}{K} \\sum_ {k = 1} ^ {K} P \\left(\\frac {d _ {k} ^ {(i)}}{m} - 1, \\rho_ {k}, \\lambda_ {k}\\right) \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 513, + 816, + 890, + 859 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As noted in Section 3, one of the main downsides of estimating Lagrange multipliers from the training set is that", + "bbox": [ + 500, + 869, + 892, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "16074", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "we could quickly overfit the data. Therefore, we propose to use the validation set to obtain a reliable estimate of the penalty multipliers at each epoch. We consider that an epoch of training corresponds to the approximate minimization of the loss function, and then compute the average penalty multiplier on the validation set. Formally after a training epoch $j$ , the penalty multipliers for epoch $j + 1$ will be, for all $k = 1, \\ldots, K$ :", + "bbox": [ + 75, + 90, + 472, + 210 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\lambda_ {k} ^ {(j + 1)} = \\frac {1}{| \\mathcal {D} _ {\\mathrm {v a l}} |} \\sum_ {(\\boldsymbol {x}, y) \\in \\mathcal {D} _ {\\mathrm {v a l}}} P ^ {\\prime} \\left(\\frac {d _ {k}}{m} - 1, \\rho_ {k} ^ {(j)}, \\lambda_ {k} ^ {(j)}\\right) \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 99, + 220, + 470, + 260 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Finally, the penalty multiplier is projected on a safeguarding interval $[\\lambda_{\\min}, \\lambda_{\\max}] = [10^{-6}, 10^{6}]$ in our case. To update the penalty parameters $\\rho$ , we compute the average constraint per class on the validation set. Then, for each class, if the average constraint is positive and has not decreased compared to the previous epoch, we multiply the corresponding penalty parameter by $\\gamma$ .", + "bbox": [ + 75, + 268, + 468, + 375 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Finally, as suggested by [2] and confirmed by our empirical results, we utilize the PHR function in our implementation, defined as follows:", + "bbox": [ + 75, + 375, + 470, + 419 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {P H R} (z, \\rho , \\lambda) = \\left\\{ \\begin{array}{l l} \\lambda z + \\frac {1}{2} \\rho z ^ {2} & \\text {i f} \\quad \\lambda + \\rho z \\geq 0; \\\\ - \\frac {\\lambda^ {2}}{2 \\rho} & \\text {o t h e r w i s e .} \\end{array} \\right. \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 96, + 428, + 468, + 469 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Overall, the proposed method, consolidated in Algorithm 2, corresponds to approximately solving the constrained problem (2), by learning class-wise penalty multipliers on the validation set, to avoid overfitting and training specificities (i.e. batch normalization, dropout, augmentations) which would result in unreliable penalty multipliers estimates.", + "bbox": [ + 75, + 477, + 472, + 568 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 76, + 580, + 209, + 598 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1. Experimental Setup", + "text_level": 1, + "bbox": [ + 76, + 604, + 267, + 622 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Datasets. We perform experiments on a variety of popular benchmarks. First, we include three widely used image classification benchmarks, including Tiny-ImageNet [7], ImageNet [7] and one long-tailed image classification, ImageNet-LT [25]. Tiny-ImageNet is widely used in the calibration literature [22, 31], with relatively small $64 \\times 64$ resolution, while ImageNet [7] is a large-scale benchmark consisting of 1000 categories and over 1M images. The main characteristic of ImageNet-LT is that the number of samples is extremely imbalanced across classes, ranging from 5 to 1280. To evaluate performance in dense prediction tasks, we include one semantic segmentation benchmark, PASCAL VOC2012 [9]. Furthermore, one benchmark from the NLP domain, 20 Newsgroups [19], is included to show the general applicability. For a detailed description of each dataset and the pre-processing settings, please refer to Appendix B.", + "bbox": [ + 75, + 628, + 472, + 869 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation Metrics. For calibration, we report the most widely used Expected Calibration Error (ECE) [33]. Samples", + "bbox": [ + 75, + 869, + 470, + 901 + ], + "page_idx": 5 + }, + { + "type": "code", + "sub_type": "algorithm", + "code_caption": [ + "Algorithm 2 CALS-ALM training" + ], + "code_body": "Require: DNN initial $\\theta^0$ , margin m \nRequire: Dataset: $\\mathcal{D}_{\\mathrm{train}}$ $\\mathcal{D}_{\\mathrm{val}}$ , batch size $B$ \nRequire: Penalty function $P,\\gamma >1,\\tau \\in (0,1)$ \nRequire: Initial $\\lambda^{(0)}\\in \\mathbb{R}_{++}^{n},\\pmb{\\rho}^{(0)}\\in \\mathbb{R}_{++}^{n}$ 1: for $j = 0,\\dots ,T$ do // Epochs of training \n2: for each mini-batch $\\{(x^{(i)},y^{(i)})\\}_{i = 1}^{B}$ in $\\mathcal{D}_{\\mathrm{train}}$ do \n3: $\\mathcal{L}_{\\mathrm{c}} = \\sum_{i = 1}^{B}\\mathcal{L}_{\\mathrm{CE}}(\\boldsymbol{x}^{(i)},y^{(i)})$ // Cross-entropy \n4: $\\mathcal{L}_{\\mathrm{p}} = \\sum_{i = 1}^{B}\\frac{1}{K}\\sum_{k = 1}^{K}P\\Big(\\frac{d_k^{(i)}}{m} -1,\\rho_k^{(j)},\\lambda_k^{(j)}\\Big)$ // Penalties \n5: $\\mathcal{L} = \\frac{1}{B} (\\mathcal{L}_{\\mathrm{c}} + \\mathcal{L}_{\\mathrm{p}})$ \n6: $\\theta^{(t + 1)}\\gets \\theta^{(t)} - \\alpha \\cdot \\nabla_{\\theta}\\mathcal{L}$ // Gradient descent \n7: end for \n8: for $k = 1,\\ldots ,K$ do // Adjust $\\lambda$ and $\\rho$ on validation \n9: $\\lambda_{k}^{(j + 1)} = \\frac{1}{|\\mathcal{D}_{\\mathrm{val}}|}\\sum_{(\\boldsymbol {x},\\boldsymbol {y})\\in \\mathcal{D}_{\\mathrm{val}}}P'\\left(\\frac{d_k}{m} -1,\\rho_k^{(j)},\\lambda_k^{(j)}\\right)$ \n10: $\\overline{d_k}^{(j)} = \\frac{1}{|\\mathcal{D}_{\\mathrm{val}}|}\\sum_{(\\boldsymbol {x},\\boldsymbol {y})\\in \\mathcal{D}_{\\mathrm{val}}}d_k - 1$ // Average constraint \n11: if $j\\geq 1$ and $\\overline{d_k}^{(j)} > \\tau \\max \\{0,\\overline{d_k}^{(j - 1)}\\}$ then \n12: $\\rho_{k}^{(j + 1)}\\gets \\gamma \\rho_{k}^{(j)}$ // Constraint has not improved \n13: else \n14: $\\rho_{k}^{(j + 1)}\\gets \\rho_{k}^{(j)}$ \n15: end if \n16: end for \n17: end for", + "bbox": [ + 500, + 107, + 890, + 479 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "are grouped into $M$ equi-spaced bins according to prediction confidence, and a weighted average of the absolute difference between accuracy and confidence in each bin is calculated:", + "bbox": [ + 496, + 508, + 893, + 555 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nE C E = \\sum_ {m = 1} ^ {M} \\frac {\\left| B _ {m} \\right|}{N} \\left| A _ {m} - C _ {m} \\right|, \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 586, + 577, + 890, + 619 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $M$ is the number of bins, $N$ the amount of test samples, $B_{m}$ the samples with prediction confidence in the $m^{th}$ bin, $A_{m}$ the accuracy and $C_{m}$ the mean confidence of samples in the $m^{th}$ bin. Note we fix $N$ to 15 according to [22, 31]. In accordance with [22], we also report Adaptive ECE (AECE), a variant of ECE where the bins are configured to evenly distribute the test samples across them. Additionally, Classwise Calibration Error (CWCE) [29], a classwise extension of ECE, is included in Appendix C. For discriminative performance, we use standard measures: accuracy (Acc) for classification, and intersection over union (mIoU) for segmentation.", + "bbox": [ + 496, + 628, + 893, + 809 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Compared methods. We compare our method to other learning based calibration losses, including (i) methods that impose constraints on predictions (either logits or softmax probabilities), i.e. Explicit Confidence Penalty (ECP) [38], Label Smoothing (LS) [42], Focal Loss (FL) [21] and its sample-dependent version (FLSD) [31], Margin-based Label", + "bbox": [ + 496, + 810, + 893, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "16075", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/c68ded947cb515613c0b2d9cbf5208d9b7b78442d8236cd16c164508913b4a10.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodTinyImageNetImageNetImageNet-LT
ResNet-50ResNet-50SwinV2-TResNet-50SwinV2-T
AccECEAECEAccECEAECEAccECEAECEAccECEAECEAccECEAECE
CE65.023.733.6975.169.199.1875.609.959.9437.9028.1228.1231.8231.8236.68
MMCE [18]65.342.812.6174.858.578.5676.689.079.0837.7928.4128.4033.1426.4126.41
ECP [38]64.904.003.9275.228.278.2675.829.889.8637.6928.1428.1331.2233.7033.70
LS [42]65.783.173.1676.042.572.8875.427.327.3337.8810.4610.3831.7011.4211.40
FL [31]63.092.963.1273.871.601.6575.603.193.1836.0418.3718.3630.7325.5025.50
FLSD [31]64.092.912.9573.972.082.0674.702.442.3736.1817.7717.7832.5625.1625.17
CPC [5]64.494.884.9176.333.663.5976.345.505.3338.9016.0015.9932.5413.2113.19
MbLS [22]64.741.641.7375.824.444.2677.181.951.7338.326.166.1632.057.657.64
CALS-HR65.092.502.4276.345.635.6977.583.062.9538.502.832.7834.312.372.45
CALS-ALM65.031.541.3876.441.461.3277.101.611.6938.562.152.3033.942.322.45
", + "bbox": [ + 122, + 88, + 849, + 244 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 1. Calibration performance for different approaches on three image classification benchmarks. We report two lower-is-better calibration metrics, i.e. ECE and AECE. Best method is highlighted in bold, while the second-best one is underlined.", + "bbox": [ + 75, + 253, + 893, + 282 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Smoothing (MbLS) [22] and CPC [5], and (ii) techniques that directly optimize calibration measures, i.e. MMCE [18]. We refer to the related literature [22, 31] to set the hyperparameters for various methods. For instance, the smoothing factor in LS and FL is set to 0.05 and 3 respectively, and we set margin to 10 in MbLS. A detailed description of hyper-parameter values can be found in Appendix F.", + "bbox": [ + 75, + 297, + 470, + 404 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Our methods. A simple alternative to the algorithm presented in Section 4.2 would be to heuristically tune multipliers by scaling them according to penalty values: if $P_{k}^{(j + 1)}$ increases we also increase $\\lambda_{k}^{j + 1}$ and vice versa. This strategy, akin to learning rate scheduling, can be formulated as:", + "bbox": [ + 75, + 405, + 470, + 484 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\lambda_ {k} ^ {(j + 1)} = \\left\\{ \\begin{array}{l l} \\mu \\lambda_ {k} ^ {(j)} & \\text {i f} P _ {k} ^ {(j + 1)} > \\tau P _ {k} ^ {(j)} \\\\ \\lambda_ {k} ^ {(j)} / \\mu & \\text {i f} P _ {k} ^ {(j)} > \\tau P _ {k} ^ {(j + 1)} \\\\ \\lambda_ {k} ^ {(j)} & \\text {o t h e r w i s e} \\end{array} \\right. \\tag {15}\n$$\n", + "text_format": "latex", + "bbox": [ + 130, + 496, + 468, + 564 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $\\mu > 1$ and $\\tau > 1$ are hyper-parameters that we fix to 1.1. We refer to our main algorithm as CALS-ALM and to this heuristic rule as CALS-HR in what follows.", + "bbox": [ + 75, + 575, + 468, + 619 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We fix the margin to $m = 10$ on vision tasks and $m = 6$ on the NLP benchmark, as in [22], for a fair comparison. We also perform an ablation study to investigate the impact of the margin value. For other hyper-parameters, we set $\\pmb{\\lambda}^{(0)} = 10^{-6} \\cdot \\mathbf{1}_K$ , $\\pmb{\\rho}^{(0)} = \\mathbf{1}_K$ , $\\gamma = 1.2$ , and we update the penalty parameters $\\pmb{\\rho}$ every 10 epochs. Please refer to Appendix F for a detailed description of all hyper-parameters.", + "bbox": [ + 75, + 622, + 470, + 728 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Implementation Details. For image classification, we experiment with ResNet [13] and a vision Transformer model, i.e. Swin Transformer V2 (SwinV2-T) [23]. DeepLabV3 [3] is employed for semantic segmentation on PASCAL VOC2012. Following [22, 31], we use the Global Pooling CNN (GPoolCNN) architecture [20] on the NLP recognition task. Further training details on each dataset can be found in Appendix B.", + "bbox": [ + 75, + 729, + 470, + 834 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.2. Results", + "text_level": 1, + "bbox": [ + 76, + 845, + 171, + 859 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Results on image classification. Table 1 presents the discriminative and calibration performance of our methods", + "bbox": [ + 75, + 869, + 472, + 900 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "across three widely used classification benchmarks, compared to baselines and related works. We can observe that our CALS-ALM approach consistently outperforms existing techniques in terms of calibration. Specifically, the results indicate that the standard CE loss and other approaches often lead to miscalibrated models, with the severity of miscalibration substantially increasing in correlation with dataset difficulty. This is particularly evident in large-scaled datasets with numerous classes, such as ImageNet, or those with long-tail class distributions, like ImageNet-LT. Although other learning-based methods could provide better calibrated networks, their performance is not stable across different settings. For example, while FL achieves a relatively low ECE of 1.60 with a Resnet50 trained on ImageNet, it only yields an ECE of 25.50 when training a SwinV2-T network on ImageNet-LT, revealing a limitation in adapting to different learning scenarios. In contrast, CALS-ALM attains the best calibration performance in all cases, frequently outperforming existing approaches by a significant margin, with minimal variations across datasets and architectures. This trends persists when compared to the most closely related technique, MbLS. It is noteworthy that the improvements on the long-tailed ImageNet-LT dataset are substantial; for example, we decrese ECE from 28.12 to 2.15 for ResNet50, and from 31.82 to 2.32 for SwinV2-T, validating the effectiveness of class adaptive learning. Another interesting finding is that CALS-HR, employing a naive update strategy for class-wise penalty weights, achieves nearly the second-best performance in the ImageNet-LT dataset. This further demonstrates the effectiveness of CALS for learning under class-imbalance scenarios. For the reliability diagrams of various models, please refer to Appendix E.", + "bbox": [ + 496, + 297, + 893, + 784 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In terms of model accuracy, our method delivers competitive performances, surpassing existing methods in certain cases. It is important to emphasize that, while the proposed method achieves discriminative results comparable to the best-performing approach for each dataset, the differences in calibration are considerable. This highlights the superiority of the proposed formulation for training highly discrimi", + "bbox": [ + 496, + 795, + 895, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "16076", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/62e78eca5eb382b7850f62eb708db706f17b5aa1ad3f60232d49924890d793ba.jpg", + "image_caption": [ + "(a) Validation ECE and multipliers during training." + ], + "image_footnote": [], + "bbox": [ + 81, + 90, + 356, + 268 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/c862e61f3d4cd7c3e9ee5bc0b8e22c20dcae1e6a6693b1aad89dba89c032ca97.jpg", + "image_caption": [ + "Figure 3. Ablation study on ImageNet-LT. (a) Evolution of ECE on validation and multipliers for CALS: ECE on the validation set for our method (CALS), CE and MbLS [22] and values of multipliers $\\lambda$ for CALS after each training epoch. (b) Effect of penalty functions and margin: ECE and accuracy on validation and test set are shown across different choices of penalty functions and margin values." + ], + "image_footnote": [], + "bbox": [ + 388, + 90, + 648, + 268 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/213ff12f720a013ab09eb9be14846ad88d70b6a413e2bbbbcd3a08b1a4fbca46.jpg", + "image_caption": [ + "(b) Effect of penalty function and margin on accuracy and ECE for CALS." + ], + "image_footnote": [], + "bbox": [ + 655, + 90, + 883, + 268 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "native and well-calibrated networks. Figure 1 provides a more visual comparison considering both accuracy and ECE. It is demonstrated that CALS-ALM provides the optimal compromise between accuracy and calibration performance.", + "bbox": [ + 75, + 367, + 470, + 428 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Ablation Analysis. Figure 3 illustrates the evolution of ECE in (a) and penalty multipliers $\\lambda$ in (b) during training. It is interesting to observe that the evolution of $\\lambda$ is consistent with the ECE. Specifically, the average penalty weight gradually increases while the ECE initially deteriorates because the model is focused on increasing accuracy. However, the value of the penalty weight begins to decline alongside the ECE, as the network starts to become better calibrated. For a visualization of classwise multipliers, please refer to Appendix D. Figure 3(c) highlights the impact of the choice of penalty functions and margin values. This demonstrates that the PHR penalty function is preferable over the other two options, P2 and P3, for both calibration and accuracy. Regarding the margin, the best performance is achieved with $m \\approx 10$ , which is consistent with the findings in [22].", + "bbox": [ + 75, + 429, + 468, + 655 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Semantic Segmentation. Table 2 presents the performance on Pascal VOC dataset. Note here CALS refers to our best method, i.e. CALS-ALM. It can be observed that the trend is consistent with image classification experiments: CALS outperforms counterparts in terms of ECE, and yields competitive results on discriminative performance, i.e. mIoU in segmentation. It is worth noting that some methods like MMCE, FLSD and CPC, are not included here because their computation demands were too heavy for pixel-wise segmentation tasks. In contrast, our method is unlimited in dense prediction tasks as the computation cost it adds is moderate.", + "bbox": [ + 75, + 657, + 468, + 823 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Text Classification. Last, we demonstrate the general applicability of the proposed method by analyzing its performance on a non-vision task, i.e. text classification on 20 Newsgroups dataset. The results are reported in Table 3. Remarkably, CALs again brings substantial improvement in", + "bbox": [ + 75, + 824, + 470, + 900 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/088b47b9b1de7d521cf3fbbc37275a9b2790c2cccd7a36ee11eb3a0157c92045.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
CEECPLSFLMbLSCALS
ECE14.755.156.9010.875.224.66
mIoU66.4665.5767.7364.2565.2966.77
", + "bbox": [ + 542, + 364, + 849, + 414 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/48dc62d79eb07fc70e53914917a3c3c087b38fe4e2c9697d4448621aa38b64f9.jpg", + "table_caption": [ + "Table 2. Segmentation results on the PASCAL VOC 2012." + ], + "table_footnote": [], + "table_body": "
CEMMCEECPLSFLFLSDCPCMbLSCALS
ECE (%)22.7523.0222.978.0710.8010.879.465.402.04
Acc (%)67.0166.2366.4867.1466.0865.8568.2767.8968.32
", + "bbox": [ + 503, + 450, + 888, + 500 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 3. Results on the text classification task, 20 Newsgroups.", + "bbox": [ + 506, + 508, + 880, + 523 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "terms of calibration, with ECE decreasing to $2.04\\%$ , while yielding the best accuracy $68.32\\%$ . This reveals that the proposed class adaptive learning method is also able to handle class differences in NLP applications and provide promising performance in terms of both accuracy and calibration.", + "bbox": [ + 496, + 537, + 893, + 613 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Limitations and Future Work", + "text_level": 1, + "bbox": [ + 500, + 625, + 771, + 641 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We have proposed Class Adaptive Label Smoothing for network calibration based on a modified Augmented Lagrangian Multiplier algorithm. Despite its superior performance over previous methods, there are potential limitations in this work. For instance, our method requires the validation set to have the same distribution as the training set. Although this is satisfied in nearly every benchmark, it will be interesting to investigate the impact of using non-independent and identically distributed (i.i.d) validation sets.", + "bbox": [ + 496, + 651, + 893, + 787 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 500, + 800, + 666, + 816 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This work is supported by National Science and Engineering Research Council of Canada (NSERC) and Prompt Quebec. We also thank Calcul Quebec and Compute Canada. Adrian Galdran is supported by a Marie Sklodowska-Curie Fellowship (No 892297).", + "bbox": [ + 496, + 824, + 893, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "16077", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 174, + 106 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Dimitri P. Bertsekas. Constrained Optimization and Lagrange Multiplier Methods (Optimization and Neural Computation Series). Athena Scientific, 1 edition, 1996. 4", + "[2] Ernesto G Birgin, Romulo A Castillo, and José Mario Martínez. Numerical comparison of augmented lagrangian algorithms for nonconvex problems. Computational Optimization and Applications, 31(1), 2005. 5, 6, 11", + "[3] Liang-Chieh Chen, George Papandreou, Florian Schroff, and Hartwig Adam. Rethinking atrous convolution for semantic image segmentation. In CVPR, 2017. 1, 7, 11", + "[4] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In ICML, 2020. 1", + "[5] Jiacheng Cheng and Nuno Vasconcelos. Calibrating deep neural networks by pairwise constraints. In CVPR, 2022. 3, 7, 13", + "[6] Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The cityscapes dataset for semantic urban scene understanding. In CVPR, June 2016. 4", + "[7] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In CVPR, 2009. 6, 11", + "[8] Zhipeng Ding, Xu Han, Peirong Liu, and Marc Niethammer. Local temperature scaling for probability calibration. In ICCV, 2021. 1, 3", + "[9] Mark Everingham, S. M. Eslami, Luc Gool, Christopher K. Williams, John Winn, and Andrew Zisserman. The pascal visual object classes challenge: A retrospective. *IJCV*, 111(1):98–136, 2015. 4, 6, 11", + "[10] Yarin Gal and Zoubin Ghahramani. Dropout as a bayesian approximation: Representing model uncertainty in deep learning. In ICML, 2016. 2", + "[11] Chuan Guo, Geoff Pleiss, Yu Sun, and Kilian Q Weinberger. On calibration of modern neural networks. In ICML, 2017. 1, 2, 3, 12", + "[12] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollar, and Ross Girshick. Masked autoencoders are scalable vision learners. In CVPR, 2022. 1", + "[13] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, 2016. 1, 7, 11", + "[14] Magnus R. Hestenes. Multiplier and gradient methods. Journal of Optimization Theory and Applications, 4:303-320, 1969. 5", + "[15] Sergey Ioffe and Christian Szegedy. Batch normalization: Accelerating deep network training by reducing internal covariate shift. In ICML, 2015. 1, 2", + "[16] Archit Karandikar, Nicholas Cain, Dustin Tran, Balaji Lakshminarayanan, Jonathon Shlens, Michael C Mozer, and Rebecca Roelofs. Soft calibration objectives for neural networks. In NeurIPS, 2021. 3", + "[17] Barry W. Kort and Dimitri P. Bertsekas. Combined primal-dual and penalty methods for convex programming. SIAM Journal on Control and Optimization, 14(2):268-294, 1976. 5" + ], + "bbox": [ + 78, + 114, + 470, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[18] Aviral Kumar, Sunita Sarawagi, and Ujjwal Jain. Trainable calibration measures for neural networks from kernel mean embeddings. In ICML, 2018. 3, 7, 13", + "[19] Ken Lang. Newsweeder: Learning to filter netnews. In ICML, 1995. 6, 12", + "[20] Min Lin, Qiang Chen, and Shuicheng Yan. Network in network. In ICLR, 2014. 7, 12", + "[21] Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dólar. Focal loss for dense object detection. In CVPR, 2017. 1, 3, 6", + "[22] Bingyuan Liu, Ismail Ben Ayed, Adrian Galdran, and Jose Dolz. The devil is in the margin: Margin-based label smoothing for network calibration. In CVPR, 2022. 1, 2, 3, 6, 7, 8, 11, 12, 13", + "[23] Ze Liu, Han Hu, Yutong Lin, Zhuliang Yao, Zhenda Xie, Yixuan Wei, Jia Ning, Yue Cao, Zheng Zhang, Li Dong, Furu Wei, and Baining Guo. Swin transformer v2: Scaling up capacity and resolution. In CVPR, 2022. 7, 11", + "[24] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In ICCV, 2021. 1", + "[25] Ziwei Liu, Zhongqi Miao, Xiaohang Zhan, Jiayun Wang, Boqing Gong, and Stella X. Yu. Large-scale long-tailed recognition in an open world. In CVPR, 2019. 6, 11", + "[26] Ilya Loshchilov and Frank Hutter. SGDR: Stochastic gradient descent with warm restarts. In *ICLR*, 2017. 11", + "[27] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In *ICLR*, 2019. 1, 11", + "[28] Xingchen Ma and Matthew B. Blaschko. Meta-cal: Well-controlled post-hoc calibration by ranking. In ICML, 2021. 3", + "[29] Lena Maier-Hein et al. Metrics reloaded: Pitfalls and recommendations for image analysis validation. arXiv preprint arXiv:2206.01653, 2022. 6, 12", + "[30] Matthias Minderer, Josip Djolonga, Rob Romijnders, Frances Hubis, Xiaohua Zhai, Neil Houlsby, Dustin Tran, and Mario Lucic. Revisiting the calibration of modern neural networks. In NeurIPS, 2021. 3, 11, 12", + "[31] Jishnu Mukhoti, Viveka Kulharia, Amartya Sanyal, Stuart Golodetz, Philip HS Torr, and Puneet K Dokania. Calibrating deep neural networks using focal loss. In NeurIPS, 2020. 1, 2, 3, 6, 7, 11, 12, 13", + "[32] Rafael Müller, Simon Kornblith, and Geoffrey Hinton. When does label smoothing help? In NeurIPS, 2019. 1, 3", + "[33] Mahdi Pakdaman Naeini, Gregory F. Cooper, and Milos Hauskrecht. Obtaining well calibrated probabilities using bayesian binning. In AAAI, 2015. 6", + "[34] H. Nakayama, H. Samaya, and H. Samaya. A generalized lagrangian function and multiplier method. Journal of Optimization Theory and Applications, 17:211-227, 1975. 5", + "[35] Jorge Nocedal and Stephen J. Wright. Numerical Optimization. Springer, New York, NY, USA, 2 edition, 2006. 4", + "[36] Yaniv Ovadia, Emily Fertig, Jie Ren, Zachary Nado, David Sculley, Sebastian Nowozin, Joshua V Dillon, Balaji Lakshminarayanan, and Jasper Snoek. Can you trust your model's uncertainty? evaluating predictive uncertainty under dataset shift. In NeurIPS, 2019. 3" + ], + "bbox": [ + 501, + 92, + 890, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "16078", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[37] Jeffrey Pennington, Richard Socher, and Christopher D Manning. Glove: Global vectors for word representation. In EMNLP, 2014. 12", + "[38] Gabriel Pereyra, George Tucker, Jan Chorowski, Lukasz Kaiser, and Geoffrey Hinton. Regularizing neural networks by penalizing confident output distributions. In ICLR, 2017. 1, 3, 6, 7, 13", + "[39] M. J. D. Powell. A method for nonlinear constraints in minimization problems. In *Fletcher, R., Ed., Optimization, Academic Press, New York, NY*, 1969. 5", + "[40] Jérôme Rony, Eric Granger, Marco Pedersoli, and Ismail Ben Ayed. Augmented lagrangian adversarial attacks. In ICCV, 2021. 5", + "[41] Sara Sangalli, Ertunc Erdil, Andeas Hotker, Olivio F Donati, and Ender Konukoglu. Constrained optimization to train neural networks on critical and under-represented classes. In NeurIPS, 2021. 4, 5", + "[42] Christian Szegedy, Vincent Vanhoucke, Sergey Ioffe, Jon Shlens, and Zbigniew Wojna. Rethinking the inception architecture for computer vision. In CVPR, 2016. 3, 6, 7, 13", + "[43] Christian Tomani, Sebastian Gruber, Muhammed Ebrar Erdem, Daniel Cremers, and Florian Buettner. Post-hoc uncertainty calibration for domain drift scenarios. In CVPR, 2021. 3", + "[44] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In NeurIPS, 2017. 1" + ], + "bbox": [ + 78, + 90, + 470, + 488 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "16079", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2023/Class Adaptive Network Calibration/d3a59c4a-62e4-447c-b052-c6b11b8ed279_model.json b/2023/Class Adaptive Network Calibration/d3a59c4a-62e4-447c-b052-c6b11b8ed279_model.json new file mode 100644 index 0000000000000000000000000000000000000000..125a5075e1f401534e8a0059bab802695d6dd899 --- /dev/null +++ b/2023/Class Adaptive Network Calibration/d3a59c4a-62e4-447c-b052-c6b11b8ed279_model.json @@ -0,0 +1,2233 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.301, + 0.131, + 0.67, + 0.152 + ], + "angle": 0, + "content": "Class Adaptive Network Calibration" + }, + { + "type": "text", + "bbox": [ + 0.167, + 0.18, + 0.8, + 0.2 + ], + "angle": 0, + "content": "Bingyuan Liu\\*, Jérôme Rony\\*, Adrian Galdran\\*, Jose Dolz\\*, Ismail Ben Ayed" + }, + { + "type": "text", + "bbox": [ + 0.194, + 0.21, + 0.77, + 0.23 + ], + "angle": 0, + "content": "1ÉTS Montreal, Canada 2Universitat Pompeu Fabra, Barcelona, Spain" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.264, + 0.314, + 0.279 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.296, + 0.473, + 0.705 + ], + "angle": 0, + "content": "Recent studies have revealed that, beyond conventional accuracy, calibration should also be considered for training modern deep neural networks. To address miscalibration during learning, some methods have explored different penalty functions as part of the learning objective, alongside a standard classification loss, with a hyper-parameter controlling the relative contribution of each term. Nevertheless, these methods share two major drawbacks: 1) the scalar balancing weight is the same for all classes, hindering the ability to address different intrinsic difficulties or imbalance among classes; and 2) the balancing weight is usually fixed without an adaptive strategy, which may prevent from reaching the best compromise between accuracy and calibration, and requires hyper-parameter search for each application. We propose Class Adaptive Label Smoothing (CALS) for calibrating deep networks, which allows to learn class-wise multipliers during training, yielding a powerful alternative to common label smoothing penalties. Our method builds on a general Augmented Lagrangian approach, a well-established technique in constrained optimization, but we introduce several modifications to tailor it for large-scale, class-adaptive training. Comprehensive evaluation and multiple comparisons on a variety of benchmarks, including standard and long-tailed image classification, semantic segmentation, and text classification, demonstrate the superiority of the proposed method. The code is available at https://github.com/by-liu/CALS." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.734, + 0.208, + 0.749 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.759, + 0.471, + 0.866 + ], + "angle": 0, + "content": "Deep Neural Networks (DNNs) have become the prevailing model in machine learning, particularly for computer vision [13] and natural language processing applications [44]. Increasingly powerful architectures [3, 13, 24], learning methods [4, 12] and a large body of other techniques [15, 27] are constantly introduced. Nonetheless, recent studies [11, 31] have shown that regardless of their superior discriminative" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.265, + 0.895, + 0.401 + ], + "angle": 0, + "content": "performance, high-capacity modern DNNs are poorly calibrated, i.e. failing to produce reliable predictive confidences. Specifically, they tend to yield over-confident predictions, where the probability associated with the predicted class overestimates the actual likelihood. Since this is a critical issue in safety-sensitive applications like autonomous driving or computational medical diagnosis, the problem of DNN calibration has been attracting increasing attention in recent years [11, 31, 38]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.402, + 0.895, + 0.567 + ], + "angle": 0, + "content": "Current calibration methods can be categorized into two main families. The first family involves techniques that perform an additional post-processing parameterized operation on the output logits (or pre-softmax activations) [11], with the calibration parameters of that operation obtained from a validation set by either learning or grid-search. Despite the simplicity and low computational cost, these methods have empirically proven to be highly effective [8, 11]. However, their main drawback is that the choice of the optimal calibration parameters is highly sensitive to the trained model instance and validation set [22, 31]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.568, + 0.896, + 0.884 + ], + "angle": 0, + "content": "The second family of methods attempts to simultaneously optimize for accuracy and calibration during network training. This is achieved by introducing, explicitly or implicitly, a secondary optimization goal involving the model's predictive uncertainty, alongside the main training objective. As a result, a scalar balancing hyper-parameter is required to tune the relative contribution of each term in the overall loss function. Some examples of this type of approaches include: Explicit Confidence Penalty (ECP) [38], Label Smoothing (LS) [32], Focal Loss (FL) [21] and its variant, Sample-Dependent Focal Loss (FLSD) [31]. It has been recently demonstrated in [22] that all these methods can be formulated as different penalty terms that enforce the same equality constraint on the logits of the DNN: driving the logit distances towards zero. Here, logit distances refers to the vector of L1 distances between the highest logit value and the rest. Observing the non-informative nature of this equality constraint, [22] proposed to use a generalized inequality constraint, only penalizing those logits for which the distance is larger than a pre-defined margin, achieving state-of-the-art calibration performance on many different benchmarks." + }, + { + "type": "text", + "bbox": [ + 0.52, + 0.886, + 0.895, + 0.902 + ], + "angle": 0, + "content": "Although learning based methods achieve greater calibra" + }, + { + "type": "page_footnote", + "bbox": [ + 0.078, + 0.875, + 0.468, + 0.901 + ], + "angle": 0, + "content": "*Equal Contributions. Correspondence to: {liubingyuan1988@gmail.com, jerome.rony.1@etsmtl.net}" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "16070" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.096, + 0.086, + 0.49, + 0.29 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.498, + 0.087, + 0.875, + 0.29 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.3, + 0.896, + 0.385 + ], + "angle": 0, + "content": "Figure 1. Many techniques have been proposed for jointly improving accuracy and calibration during training [11,31], but they fail to consider uneven learning scenarios like high class imbalance or long-tail distributions. We show a comparison of the proposed CALS-ALM method and different learning approaches in terms of Calibration Error (ECE) vs Accuracy on the (a) ImageNet and (b) ImageNet-LT (long-tailed ImageNet) datasets. A lower ECE indicates better calibration: a better model should attain high ACC and low ECE. Among all the considered methods, CALS-ALM shows superior performance when considering both discriminative power and well-balanced probabilistic predictions, achieving best accuracy and calibration on ImageNet, and best calibration and second best accuracy on ImageNet-LT." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.411, + 0.473, + 0.622 + ], + "angle": 0, + "content": "tion performance [22, 31], they have two major limitations: 1) The scalar balancing weight is equal for all classes. This hinders the network performance when some classes are harder to learn or less represented than others, such as in datasets with a large number of categories (ImageNet) or considerable class imbalance (ImageNet-LT). 2) The balancing weight is usually fixed before network optimization, with no learning or adaptive strategy throughout training. This can prevent the model from reaching the best compromise between accuracy and calibration. To address the above issues, we introduce Class Adaptive Label Smoothing method based on an Augmented Lagrangian Multiplier algorithm, which we refer to as CALS-ALM. Our Contributions can be summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.633, + 0.47, + 0.754 + ], + "angle": 0, + "content": "- We propose Class Adaptive Label Smoothing (CALS) for network calibration. Adaptive class-wise multipliers are introduced instead of the widely used single balancing weight, which addresses the above two issues: 1) CALS can handle a high number of classes with different intrinsic difficulties, e.g. ImageNet; 2) CALS can effectively learn from data suffering from class imbalance or a long-tailed distribution, e.g. ImageNet-LT." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.765, + 0.473, + 0.902 + ], + "angle": 0, + "content": "- Different from previous penalty based methods, we solve the resulting constrained optimization problem by implementing a modified Augmented Lagrangian Multiplier (ALM) algorithm, which yields adaptive and optimal weights for the constraints. We make some critical design decisions in order to adapt ALM to the nature of modern learning techniques: 1) The inner convergence criterion in ALM is relaxed to a fixed number of iterations in each inner stage, which is amenable to mini-batch stochas" + }, + { + "type": "list", + "bbox": [ + 0.078, + 0.633, + 0.473, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.411, + 0.895, + 0.562 + ], + "angle": 0, + "content": "tic gradient optimization in deep learning. 2) Popular techniques, such as data augmentation, batch normalization [15] and dropout [10], rule out the possibility of tracking original samples and applying sample-wise multipliers. To overcome this complication, we introduce class-wise multipliers, instead of sample-wise multipliers in the standard ALM. 3) The outer-step update for estimating optimal ALM multipliers is performed on the validation set, which is meaningful for training on large-scale training set and avoids potential overfitting." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.569, + 0.897, + 0.736 + ], + "angle": 0, + "content": "- Comprehensive experiments over a variety of applications and benchmarks, including standard image classification (Tiny-ImageNet and ImageNet), long-tailed image classification (ImageNet-LT), semantic segmentation (PASCAL VOC 2012), and text classification (20 Newsgroups), demonstrate the effectiveness of our CALS-ALM method. As shown in Figure 1, CALS-ALM yields superior performance over baselines and state-of-the-art calibration losses when considering both accuracy and calibration, especially for more realistic large-scaled datasets with large number of classes or class imbalance." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.747, + 0.642, + 0.763 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.772, + 0.704, + 0.787 + ], + "angle": 0, + "content": "2.1. Problem Formulation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.794, + 0.896, + 0.903 + ], + "angle": 0, + "content": "Given a dataset \\(\\mathcal{D} = \\{(x^{(i)},y^{(i)})\\}_{i = 1}^{N}\\) with \\(N\\in \\mathbb{N}\\) pairs of samples \\(\\pmb {x}\\in \\mathcal{X}\\) and corresponding labels \\(y\\in \\mathcal{V}\\) with \\(\\mathcal{V} = \\{1,\\dots ,K\\}\\), a deep neural network (DNN) \\(F_{\\theta}:X\\to \\mathbb{R}^{K}\\) parameterized by \\(\\theta\\) yields logits \\(l = F_{\\theta}(\\pmb {x}) = (F_{\\theta}(\\pmb {x})_k)_{1\\leq k\\leq K}\\in \\mathbb{R}^K\\). In a classification scenario, the output probability \\(s = (s_k)_{1\\leq k\\leq K}\\in \\Delta^{K - 1}\\) where \\(\\Delta^{K - 1}\\subset [0,1]^K\\) denotes the probability simplex, is" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "16071" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.228 + ], + "angle": 0, + "content": "obtained by applying the softmax function on the output log-its, i.e. \\( s = \\text{softmax}(l) = \\frac{\\exp l}{\\sum \\exp l} \\). Therefore, the predicted class \\( \\hat{y} \\) is computed as \\( \\hat{y} = \\arg \\max_k s_k \\), and the predicted confidence is \\( \\hat{p} = s_{\\hat{y}} = \\max_k s_k \\). A perfectly calibrated model should satisfy that the predicted confidence of any input is equal to the accuracy of the model: \\( \\hat{p} = \\mathbb{P}(\\hat{y} = y|\\hat{p}) \\). Hence, an over-confident model yields on average larger confidence than the associated accuracy, whereas an under-confident model yields lower confidence than its accuracy." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.23, + 0.472, + 0.292 + ], + "angle": 0, + "content": "A number of recent studies [11, 22, 30, 31] have shown that DNNs tend to become over-confident during training as a result of minimizing the popular cross-entropy (CE) training loss:" + }, + { + "type": "equation", + "bbox": [ + 0.177, + 0.304, + 0.472, + 0.348 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {C E}} (\\boldsymbol {x}, y) = - \\sum_ {k = 1} ^ {K} y _ {k} \\log s _ {k} \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.359, + 0.473, + 0.496 + ], + "angle": 0, + "content": "where \\(\\pmb{y} \\in \\{0,1\\}^{K}\\) is the one-hot encoding of \\(y\\). This objective function is minimized when the predictions for all the training samples fully match the ground-truth labels \\(y\\), i.e. \\(s_{y} = 1\\) and \\(\\forall k \\neq y, s_{k} = 0\\). The negative logarithmic term on the logit of the correct category renders the global minimization of the CE loss unreachable, as it keeps pushing the predicted probabilities \\(s\\) towards the vertices of the \\((K-1)\\)-simplex even after the classification error is zero [31], resulting in over-confident models." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.508, + 0.303, + 0.525 + ], + "angle": 0, + "content": "2.2. Post-processing methods" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.532, + 0.473, + 0.79 + ], + "angle": 0, + "content": "To address mis-calibration, different post-processing techniques applied after model training have been proposed recently [8, 11, 43]. The most popular of these strategies is Temperature Scaling (TS) [11], which applies a single scalar temperature parameter to manipulate logit outputs monotonely, resulting in softened prediction confidences without affecting predicted labels. Note that here the temperature parameter needs to be tuned on a separate validation set. Despite its simplicity, TS has been shown effective in fixing over-confidence predictions [11]. As a local alternative to TS, [8] proposed to train a regression model for learning position-specific temperature for semantic segmentation problems. Unfortunately, TS and its variants can be sensitive to both the model and the validation set, and do not work well under data distribution shifts [36]. Thus, some subsequent works [28, 43] have attempted to provide solutions for improving performance under domain shift." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.801, + 0.305, + 0.817 + ], + "angle": 0, + "content": "2.3. Learning-based methods" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.472, + 0.903 + ], + "angle": 0, + "content": "Another popular direction is to directly address miscalibration during training by introducing an additional penalty or supervision regarding model calibration with the standard training loss. In [18], the authors introduced a trainable calibration measure based on RKHS kernels, while [16]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.895, + 0.365 + ], + "angle": 0, + "content": "proposed a differential calibration loss based on a soft version of the binning operation in the ECE metric. In [5], two types of binary pairwise calibration constraints were proposed as additional penalty terms during training. Other methods try to decrease over-fitting on the cross-entropy loss, which has been demonstrated to be the main reason of over-confidence [11, 31]. In [38] an explicit confidence penalty (ECP) is proposed to maximize the entropy and reduce over-fitting, while Label Smoothing [42] has also been shown to implicitly improve the calibration [32] by softening the hard one-hot targets in the cross-entropy. The Focal Loss [21], originally proposed to tackle class imbalance, can also be effective for calibration [31], as it implicitly minimizes the Kullback-Leibler (KL) divergence between the uniform distribution and the network softmax probabilities, thereby increasing the entropy of the predictions. As an extension the Sample-Dependent Focal Loss (FLSD) was also proposed in [31] to further boost calibration performance." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.388, + 0.895, + 0.554 + ], + "angle": 0, + "content": "Margin-based Label Smoothing (MbLS). A unifying constrained-optimization formulation of loss functions promoting calibration has been recently presented in [22]. Specifically, the additional penalties integrated in these methods, including ECP [38], LS [32] and FL [31], can be viewed as different forms of approximations to the same constraint, i.e. enforcing the logit distances to be zero. Noticing that this constraint is non-informative (its solution being uniformly distributed probabilities), [22] further proposed a generalized formulation by relaxing the constraint to allow the logit distances being lower than a strictly positive margin." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.555, + 0.895, + 0.601 + ], + "angle": 0, + "content": "The specific formulation of MbLS [22] is as follows. Given a margin \\(m \\in \\mathbb{R}_+\\), the constrained optimization problem for network training is:" + }, + { + "type": "equation", + "bbox": [ + 0.52, + 0.614, + 0.892, + 0.655 + ], + "angle": 0, + "content": "\\[\n\\min _ {\\theta} \\sum_ {i = 1} ^ {N} \\mathcal {L} _ {\\mathrm {C E}} \\left(\\boldsymbol {x} ^ {(i)}, y ^ {(i)}\\right) \\tag {2}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.524, + 0.657, + 0.852, + 0.682 + ], + "angle": 0, + "content": "\\[\n\\text {s . t .} \\quad \\max _ {k} \\left\\{\\boldsymbol {l} _ {k} ^ {(i)} \\right\\} - l ^ {(i)} \\preceq m \\mathbf {1} _ {K}, \\quad i = 1, \\dots , N,\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.696, + 0.892, + 0.758 + ], + "angle": 0, + "content": "where \\( l^{(i)} = F_{\\theta}(\\pmb{x}^{(i)}) \\). The minimum can be approximated by penalty-based optimization methods, transforming the above formulation into an unconstrained problem by means of simple ReLU functions:" + }, + { + "type": "equation", + "bbox": [ + 0.54, + 0.771, + 0.892, + 0.821 + ], + "angle": 0, + "content": "\\[\n\\min _ {\\theta} \\sum_ {i = 1} ^ {N} \\mathcal {L} _ {\\mathrm {C E}} \\left(\\boldsymbol {x} ^ {(i)}, y ^ {(i)}\\right) + \\tag {3}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.576, + 0.816, + 0.852, + 0.858 + ], + "angle": 0, + "content": "\\[\n\\lambda \\sum_ {i = 1} ^ {N} \\sum_ {j = 1} ^ {K} \\max \\{0, \\max _ {k} \\{l _ {k} ^ {(i)} \\} - l _ {j} ^ {(i)} - m \\},\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.892, + 0.902 + ], + "angle": 0, + "content": "where \\(\\lambda \\in \\mathbb{R}_+\\) is a scalar weight balancing contributions of the CE loss and the corresponding penalty." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16072" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.078, + 0.09, + 0.47, + 0.108 + ], + "angle": 0, + "content": "3. Sample-wise Constrained DNN Optimization" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.115, + 0.473, + 0.268 + ], + "angle": 0, + "content": "Although MbLS can significantly improve calibration, the associated constrained problem (2) is not solved accurately. It is approximated by an unconstrained problem with a single uniform penalty, regardless of the data sample or category. However, the samples and classes considered in a classification problem have different intrinsic learning difficulties. Therefore, an improved training scheme would involve considering distinct penalty weights \\(\\lambda\\) for each sample and class. This would result in having to chose \\(N\\times K\\) penalty weights \\(\\Lambda \\in \\mathbb{R}_{+}^{N\\times K}\\), with the resulting optimization problem being:" + }, + { + "type": "equation", + "bbox": [ + 0.095, + 0.277, + 0.47, + 0.364 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\min _ {\\theta} \\sum_ {i = 1 \\atop N = K} ^ {N} \\mathcal {L} _ {\\mathrm {C E}} \\left(\\boldsymbol {x} ^ {(i)}, y ^ {(i)}\\right) + \\tag {4} \\\\ \\sum_ {i = 1} ^ {N} \\sum_ {j = 1} ^ {K} \\Lambda_ {i j} \\max \\{0, \\max _ {k} \\{l _ {k} ^ {(i)} \\} - l _ {j} ^ {(i)} - m \\}. \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.373, + 0.47, + 0.463 + ], + "angle": 0, + "content": "From an optimization perspective, supposing that optimal weights \\(\\theta^{\\star}\\) exists for problem (2), there exists \\(\\Lambda^{\\star} \\in \\mathbb{R}_{+}^{N \\times K}\\) such that \\((\\theta^{\\star}, \\Lambda^{\\star})\\) is a saddle point of the Lagrangian associated to (2). These \\(\\Lambda^{\\star}\\) are the Lagrange multipliers of the problem. Therefore, using \\(\\Lambda = \\Lambda^{\\star}\\) would be the best choice to solve (4)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.464, + 0.473, + 0.825 + ], + "angle": 0, + "content": "In practice, using the Lagrange multipliers of problem (2) as the weights for the penalties may not be computationally feasible, and it could even result in degraded performance. Indeed, in the context of machine learning, we optimize a model's weights \\(\\theta\\) to solve (2) on a training set \\(\\mathcal{D}_{\\mathrm{train}}\\), and expect to generalize on a test set \\(\\mathcal{D}_{\\mathrm{test}}\\), which we do not have access to during training. Because of the bias-variance tradeoff, solving (2) optimally would likely result in overfitting, i.e. we may solve problem (2) accurately on the train set, but not generalize properly on the test set, resulting in poor calibration and classification performance overall. This suggests that it could be preferable to evaluate during training the quality of multipliers on a separate validation set \\(\\mathcal{D}_{\\mathrm{val}}\\). Additionally, several mechanisms for training DNNs are not compatible with a straightforward minimization. First, the use of batch normalization yields predictions that are not independent between samples in a minibatch. Second, the use of regularization techniques such as dropout may lead to virtually inaccurate predictions on certain training samples, impacting the correct estimation of multipliers. Third, data augmentation, which is standard in DNN training, would result in additional penalty weights for the augmented samples: they can be easier or harder to classify than the original ones." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.826, + 0.473, + 0.902 + ], + "angle": 0, + "content": "In addition to the above obstacles, applying a penalty weight per sample and per class (resulting in \\( N \\times K \\) weights) would not scale well for large datasets and dense predictive tasks, such as semantic segmentation, which is typically formulated as a per-pixel classification task. Assuming that im" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.895, + 0.227 + ], + "angle": 0, + "content": "ages in the dataset have a size of \\( H \\times W \\), this would result in \\( N \\times H \\times W \\times K \\) penalty weights. This rapidly becomes a limiting factor for moderately sized segmentation datasets. For instance, Pascal VOC 2012 [9] contains 21 classes and 1464 training images, amounting to \\( 2.62 \\times 10^{8} \\) pixels, or \\( 5.5 \\times 10^{9} \\) penalty weights, which, stored as float32, would use \\( \\sim 20 \\) GiB. For Cityscapes [6], containing approximately 3000 training images of size \\( 2048 \\times 1024 \\) in 19 classes, this amounts to \\( \\sim 445 \\) GiB." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.228, + 0.895, + 0.274 + ], + "angle": 0, + "content": "Following the above observations, we introduce a relaxation of sample-wise penalties, and propose to solve the following problem:" + }, + { + "type": "equation", + "bbox": [ + 0.52, + 0.284, + 0.892, + 0.372 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\min _ {\\theta} \\sum_ {i = 1 \\atop N, K} ^ {N} \\mathcal {L} _ {\\mathrm {C E}} \\left(\\boldsymbol {x} ^ {(i)}, y ^ {(i)}\\right) + \\tag {5} \\\\ \\sum_ {i = 1} ^ {N} \\sum_ {j = 1} ^ {K} \\lambda_ {j} \\max \\{0, \\max _ {k} \\{l _ {k} ^ {(i)} \\} - l _ {j} ^ {(i)} - m \\}, \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.384, + 0.895, + 0.522 + ], + "angle": 0, + "content": "where \\((\\lambda_j)_{1\\leq j\\leq K}\\in \\mathbb{R}_+^K\\). Since penalties are now class-wise, we need \\(K\\) penalty weights. This has the advantage to scale well to denser classification tasks such as segmentation. However, we still face a challenging optimization problem (5), since we still need to chose \\(K\\) weights, which can be extremely complicated for large-scale datasets with many classes such as ImageNet, which contains 1000 classes. In the next section, we introduce a numerical technique to deal with this challenge." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.536, + 0.83, + 0.552 + ], + "angle": 0, + "content": "4. Class Adaptive Network Calibration" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.561, + 0.894, + 0.637 + ], + "angle": 0, + "content": "The challenge of the previous formulation stems from correctly choosing the weights \\(\\lambda \\in \\mathbb{R}_+^K\\), which can be cumbersome when \\(K\\) is large. Therefore, we propose to use an Augmented Lagrangian Multiplier (ALM) method to adaptively learn the weights of the penalties." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.647, + 0.645, + 0.661 + ], + "angle": 0, + "content": "4.1. General ALM" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.671, + 0.892, + 0.747 + ], + "angle": 0, + "content": "ALM methods combine penalties and primal-dual updates to solve a constrained problem. They have well-established advantages and enjoy widespread popularity in the general context of optimization [1,35,41]. Specifically, we have the following generic constrained optimization problem:" + }, + { + "type": "equation", + "bbox": [ + 0.536, + 0.759, + 0.892, + 0.78 + ], + "angle": 0, + "content": "\\[\n\\min _ {x} f (x) \\quad \\text {s . t .} \\quad h _ {i} (x) \\leq 0, \\quad i = 1, \\dots , n \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.792, + 0.892, + 0.854 + ], + "angle": 0, + "content": "where \\( f: \\mathbb{R}^d \\to \\mathbb{R} \\) is the objective function and \\( h_i: \\mathbb{R}^d \\to \\mathbb{R}, i = 1, \\dots, n \\) are the constraint functions. We tackle it by approximately solving a sequence \\( j \\in \\mathbb{N} \\) of unconstrained problems:" + }, + { + "type": "equation", + "bbox": [ + 0.521, + 0.865, + 0.892, + 0.904 + ], + "angle": 0, + "content": "\\[\n\\min _ {x} \\quad \\mathcal {L} ^ {(j)} (x) = f (x) + \\sum_ {i = 1} ^ {n} P \\left(h _ {i} (x), \\rho_ {i} ^ {(j)}, \\lambda_ {i} ^ {(j)}\\right) \\tag {7}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16073" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.109, + 0.092, + 0.445, + 0.25 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.264, + 0.471, + 0.306 + ], + "angle": 0, + "content": "Figure 2. A penalty-Lagrangian function \\( P \\) with varying values of \\( \\rho \\) and \\( \\mu \\). Higher values of \\( \\rho \\) bring \\( P \\) closer to an ideal penalty. The multiplier \\( \\lambda \\) is the derivative of \\( P \\) w.r.t. the constraint at \\( z = 0 \\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.33, + 0.471, + 0.422 + ], + "angle": 0, + "content": "with \\(P: \\mathbb{R} \\times \\mathbb{R}_{++}^{n} \\times \\mathbb{R}_{++}^{n} \\to \\mathbb{R}\\) a penalty-Lagrangian function, and \\(\\pmb{\\rho}^{(j)} = (\\rho_i)_{1 \\leq i \\leq n} \\in \\mathbb{R}_{++}^n\\), \\(\\pmb{\\lambda}^{(j)} = (\\lambda_i^{(j)})_{1 \\leq i \\leq n} \\in \\mathbb{R}_{++}^n\\) the penalty parameters and multipliers associated to \\(P\\) at the \\(j\\)-th iteration. This sequence of unconstrained problems is called outer iterations, while the steps in the minimization of \\(\\mathcal{L}^{(j)}\\) are called inner iterations." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.424, + 0.471, + 0.681 + ], + "angle": 0, + "content": "The main components of ALM methods are (i) the penalty-Lagrangian function \\( P \\), (ii) the update of the penalty multipliers \\( \\lambda^{(j)} \\) and (iii) the increase of the penalty parameters \\( \\pmb{\\rho}^{(j)} \\). First, the penalty function \\( P \\) needs to satisfy a set of axioms [2] (see Appendix A): these axioms constrain the function to be continuously differentiable w.r.t. its first variable and to have a non-negative derivative: \\( \\forall z \\in \\mathbb{R}, P'(z, \\rho, \\lambda) = \\frac{\\partial}{\\partial z} P(z, \\rho, \\lambda) \\geq 0 \\), with \\( P'(0, \\rho, \\lambda) = \\lambda \\). Figure 2 gives an example of a penalty, and how \\( \\rho \\) and \\( \\lambda \\) affect it. The choice of penalty function is critical to the performance of ALM methods, especially for nonconvex problems [2]. Typical functions include PHR [14,39], \\( P_{2} \\) [17] and \\( P_{3} \\) [34] (see section 3.2 of [2]). Second, the penalty multipliers \\( \\lambda^{(j)} \\) are updated to the derivative of \\( P \\) w.r.t. the solution obtained during the last inner minimization. Formally, let \\( x^{(j)} \\) be the approximate minimizer of \\( \\mathcal{L}^{(j)} \\), then \\( \\forall i \\in \\{1, \\dots, n\\} \\):" + }, + { + "type": "equation", + "bbox": [ + 0.163, + 0.689, + 0.47, + 0.71 + ], + "angle": 0, + "content": "\\[\n\\lambda_ {i} ^ {(j + 1)} = P ^ {\\prime} \\left(h _ {i} \\left(x ^ {(j)}\\right), \\rho_ {i} ^ {(j)}, \\lambda_ {i} ^ {(j)}\\right) \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.718, + 0.47, + 0.811 + ], + "angle": 0, + "content": "This update rule corresponds to a first-order multiplier estimate for the constrained problem. Third, the penalty parameters \\(\\rho^{(j)}\\) are increased during the outer iterations if the constraints do not improve (i.e. is closer to being satisfied) compared to the previous outer iteration. Typically, \\(\\rho_{i}^{(j + 1)} = \\gamma \\rho_{i}^{(j)}\\) if \\(h_i\\) does not improve, with \\(\\gamma > 1\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.811, + 0.471, + 0.901 + ], + "angle": 0, + "content": "When the problem is convex, alternating between the approximate minimization of (7) and the update of the multipliers (8) leads to a solution for the constrained problem. The inner minimization corresponds to minimizing the primal while the outer iterations correspond to solving the dual problem. The complete procedure is presented in Algorithm 1." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.894, + 0.152 + ], + "angle": 0, + "content": "Although guarantees exist only in the convex case, it is well-known that ALM methods can efficiently solve nonconvex problems as well [2]. In the context of deep learning, their use has been surprisingly underexplored [40, 41]." + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.167, + 0.887, + 0.183 + ], + "angle": 0, + "content": "Algorithm 1 Augmented Lagrangian Multiplier algorithm" + }, + { + "type": "text", + "bbox": [ + 0.501, + 0.186, + 0.69, + 0.2 + ], + "angle": 0, + "content": "Require: Objective function \\( f \\)" + }, + { + "type": "text", + "bbox": [ + 0.501, + 0.201, + 0.787, + 0.214 + ], + "angle": 0, + "content": "Require: Constraint functions \\(h_i, i = 1, \\dots, n\\)" + }, + { + "type": "text", + "bbox": [ + 0.501, + 0.214, + 0.878, + 0.228 + ], + "angle": 0, + "content": "Require: Penalty function \\(P\\), initial \\(\\pmb{\\lambda}^{(0)} \\in \\mathbb{R}_{++}^{n}\\), \\(\\pmb{\\rho}^{(0)} \\in \\mathbb{R}_{++}^{n}\\)" + }, + { + "type": "text", + "bbox": [ + 0.501, + 0.228, + 0.782, + 0.242 + ], + "angle": 0, + "content": "Require: Initial variable \\( x^{(0)} \\), iterations \\( j = 1 \\)" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.242, + 0.669, + 0.255 + ], + "angle": 0, + "content": "1: while not converged do" + }, + { + "type": "text", + "bbox": [ + 0.51, + 0.256, + 0.857, + 0.269 + ], + "angle": 0, + "content": "2: Initialize with \\( x^{(j - 1)} \\) and minimize (approximately):" + }, + { + "type": "equation", + "bbox": [ + 0.544, + 0.269, + 0.817, + 0.285 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} ^ {(j)} (x) = f (x) + \\sum_ {i = 1} ^ {n} P \\left(h _ {i} (x), \\rho_ {i} ^ {(j)}, \\lambda_ {i} ^ {(j)}\\right)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.51, + 0.285, + 0.785, + 0.3 + ], + "angle": 0, + "content": "3: \\(x^{(j)}\\gets\\) (approximate) minimizer of \\(\\mathcal{L}^{(j)}\\)" + }, + { + "type": "text", + "bbox": [ + 0.51, + 0.301, + 0.663, + 0.313 + ], + "angle": 0, + "content": "4: for \\(i = 1,\\dots ,n\\) do" + }, + { + "type": "text", + "bbox": [ + 0.51, + 0.313, + 0.764, + 0.328 + ], + "angle": 0, + "content": "5: \\(\\lambda_{i}^{(j + 1)}\\gets P^{\\prime}(h_{i}(x^{(j)}),\\rho_{i}^{(j)},\\lambda_{i}^{(j)})\\)" + }, + { + "type": "text", + "bbox": [ + 0.51, + 0.329, + 0.816, + 0.341 + ], + "angle": 0, + "content": "6: if the \\(i\\)-th constraint does not improve then" + }, + { + "type": "text", + "bbox": [ + 0.51, + 0.341, + 0.671, + 0.357 + ], + "angle": 0, + "content": "7: \\(\\rho_{i}^{(j + 1)}\\gets \\gamma \\rho_{i}^{(j)}\\)" + }, + { + "type": "text", + "bbox": [ + 0.51, + 0.358, + 0.585, + 0.369 + ], + "angle": 0, + "content": "8: else" + }, + { + "type": "text", + "bbox": [ + 0.51, + 0.37, + 0.663, + 0.385 + ], + "angle": 0, + "content": "9: \\(\\rho_{i}^{(j + 1)}\\gets \\rho_{i}^{(j)}\\)" + }, + { + "type": "text", + "bbox": [ + 0.504, + 0.386, + 0.6, + 0.396 + ], + "angle": 0, + "content": "10: end if" + }, + { + "type": "text", + "bbox": [ + 0.504, + 0.399, + 0.591, + 0.41 + ], + "angle": 0, + "content": "11: end for" + }, + { + "type": "text", + "bbox": [ + 0.504, + 0.412, + 0.609, + 0.426 + ], + "angle": 0, + "content": "12: \\(j\\gets j + 1\\)" + }, + { + "type": "text", + "bbox": [ + 0.504, + 0.428, + 0.59, + 0.438 + ], + "angle": 0, + "content": "13: end while" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.242, + 0.857, + 0.438 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.468, + 0.694, + 0.482 + ], + "angle": 0, + "content": "4.2. ALM for calibration" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.491, + 0.891, + 0.553 + ], + "angle": 0, + "content": "Our goal now is to build an ALM method effective for calibration purposes. We can achieve this by reformulating problem (5) using a penalty function \\(P\\) parameterized by \\((\\pmb {\\rho},\\pmb {\\lambda})\\in \\mathbb{R}_{++}^{K}\\times \\mathbb{R}_{++}^{K}\\) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.51, + 0.565, + 0.892, + 0.606 + ], + "angle": 0, + "content": "\\[\n\\min _ {\\theta} \\quad \\sum_ {i = 1} ^ {N} \\mathcal {L} _ {\\mathrm {C E}} \\left(\\boldsymbol {x} ^ {(i)}, y ^ {(i)}\\right) + \\sum_ {k = 1} ^ {K} P \\left(d _ {k} ^ {(i)} - m, \\rho_ {k}, \\lambda_ {k}\\right) \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.618, + 0.894, + 0.682 + ], + "angle": 0, + "content": "where \\( d_k^{(i)} = \\max \\{l^{(i)}\\} - l_k^{(i)} \\in \\mathbb{R}_+ \\). With this formulation, it is natural to use a penalty-Lagrangian function for \\( P \\). To avoid numerical issues typically associated with non-linear penalties, we normalize the constraints by the margin \\( m > 0 \\):" + }, + { + "type": "equation", + "bbox": [ + 0.594, + 0.694, + 0.892, + 0.727 + ], + "angle": 0, + "content": "\\[\nd _ {k} ^ {(i)} - m \\leq 0 \\Leftrightarrow \\frac {d _ {k} ^ {(i)}}{m} - 1 \\leq 0 \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.731, + 0.892, + 0.807 + ], + "angle": 0, + "content": "This leads to improved numerical stability for the ALM multiplier update as well. Additionally, we average the constraints instead of summing them. This makes the method independent of the number of classes, and eases the choice of initial penalty parameters \\(\\rho^{(0)}\\). The resulting loss is:" + }, + { + "type": "equation", + "bbox": [ + 0.514, + 0.818, + 0.892, + 0.86 + ], + "angle": 0, + "content": "\\[\n\\sum_ {i = 1} ^ {N} \\mathcal {L} _ {\\mathrm {C E}} \\left(\\boldsymbol {x} ^ {(i)}, y ^ {(i)}\\right) + \\frac {1}{K} \\sum_ {k = 1} ^ {K} P \\left(\\frac {d _ {k} ^ {(i)}}{m} - 1, \\rho_ {k}, \\lambda_ {k}\\right) \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.893, + 0.901 + ], + "angle": 0, + "content": "As noted in Section 3, one of the main downsides of estimating Lagrange multipliers from the training set is that" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16074" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.212 + ], + "angle": 0, + "content": "we could quickly overfit the data. Therefore, we propose to use the validation set to obtain a reliable estimate of the penalty multipliers at each epoch. We consider that an epoch of training corresponds to the approximate minimization of the loss function, and then compute the average penalty multiplier on the validation set. Formally after a training epoch \\( j \\), the penalty multipliers for epoch \\( j + 1 \\) will be, for all \\( k = 1, \\ldots, K \\):" + }, + { + "type": "equation", + "bbox": [ + 0.1, + 0.221, + 0.471, + 0.261 + ], + "angle": 0, + "content": "\\[\n\\lambda_ {k} ^ {(j + 1)} = \\frac {1}{| \\mathcal {D} _ {\\mathrm {v a l}} |} \\sum_ {(\\boldsymbol {x}, y) \\in \\mathcal {D} _ {\\mathrm {v a l}}} P ^ {\\prime} \\left(\\frac {d _ {k}}{m} - 1, \\rho_ {k} ^ {(j)}, \\lambda_ {k} ^ {(j)}\\right) \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.269, + 0.47, + 0.375 + ], + "angle": 0, + "content": "Finally, the penalty multiplier is projected on a safeguarding interval \\([\\lambda_{\\min}, \\lambda_{\\max}] = [10^{-6}, 10^{6}]\\) in our case. To update the penalty parameters \\(\\rho\\), we compute the average constraint per class on the validation set. Then, for each class, if the average constraint is positive and has not decreased compared to the previous epoch, we multiply the corresponding penalty parameter by \\(\\gamma\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.376, + 0.472, + 0.42 + ], + "angle": 0, + "content": "Finally, as suggested by [2] and confirmed by our empirical results, we utilize the PHR function in our implementation, defined as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.097, + 0.429, + 0.47, + 0.47 + ], + "angle": 0, + "content": "\\[\n\\operatorname {P H R} (z, \\rho , \\lambda) = \\left\\{ \\begin{array}{l l} \\lambda z + \\frac {1}{2} \\rho z ^ {2} & \\text {i f} \\quad \\lambda + \\rho z \\geq 0; \\\\ - \\frac {\\lambda^ {2}}{2 \\rho} & \\text {o t h e r w i s e .} \\end{array} \\right. \\tag {13}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.478, + 0.473, + 0.569 + ], + "angle": 0, + "content": "Overall, the proposed method, consolidated in Algorithm 2, corresponds to approximately solving the constrained problem (2), by learning class-wise penalty multipliers on the validation set, to avoid overfitting and training specificities (i.e. batch normalization, dropout, augmentations) which would result in unreliable penalty multipliers estimates." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.581, + 0.21, + 0.599 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.606, + 0.268, + 0.623 + ], + "angle": 0, + "content": "5.1. Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.629, + 0.473, + 0.871 + ], + "angle": 0, + "content": "Datasets. We perform experiments on a variety of popular benchmarks. First, we include three widely used image classification benchmarks, including Tiny-ImageNet [7], ImageNet [7] and one long-tailed image classification, ImageNet-LT [25]. Tiny-ImageNet is widely used in the calibration literature [22, 31], with relatively small \\(64 \\times 64\\) resolution, while ImageNet [7] is a large-scale benchmark consisting of 1000 categories and over 1M images. The main characteristic of ImageNet-LT is that the number of samples is extremely imbalanced across classes, ranging from 5 to 1280. To evaluate performance in dense prediction tasks, we include one semantic segmentation benchmark, PASCAL VOC2012 [9]. Furthermore, one benchmark from the NLP domain, 20 Newsgroups [19], is included to show the general applicability. For a detailed description of each dataset and the pre-processing settings, please refer to Appendix B." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.871, + 0.472, + 0.902 + ], + "angle": 0, + "content": "Evaluation Metrics. For calibration, we report the most widely used Expected Calibration Error (ECE) [33]. Samples" + }, + { + "type": "code_caption", + "bbox": [ + 0.501, + 0.092, + 0.732, + 0.107 + ], + "angle": 0, + "content": "Algorithm 2 CALS-ALM training" + }, + { + "type": "algorithm", + "bbox": [ + 0.5, + 0.108, + 0.892, + 0.481 + ], + "angle": 0, + "content": "Require: DNN initial \\(\\theta^0\\) , margin m \nRequire: Dataset: \\(\\mathcal{D}_{\\mathrm{train}}\\) \\(\\mathcal{D}_{\\mathrm{val}}\\) , batch size \\(B\\) \nRequire: Penalty function \\(P,\\gamma >1,\\tau \\in (0,1)\\) \nRequire: Initial \\(\\lambda^{(0)}\\in \\mathbb{R}_{++}^{n},\\pmb{\\rho}^{(0)}\\in \\mathbb{R}_{++}^{n}\\) 1: for \\(j = 0,\\dots ,T\\) do // Epochs of training \n2: for each mini-batch \\(\\{(x^{(i)},y^{(i)})\\}_{i = 1}^{B}\\) in \\(\\mathcal{D}_{\\mathrm{train}}\\) do \n3: \\(\\mathcal{L}_{\\mathrm{c}} = \\sum_{i = 1}^{B}\\mathcal{L}_{\\mathrm{CE}}(\\boldsymbol{x}^{(i)},y^{(i)})\\) // Cross-entropy \n4: \\(\\mathcal{L}_{\\mathrm{p}} = \\sum_{i = 1}^{B}\\frac{1}{K}\\sum_{k = 1}^{K}P\\Big(\\frac{d_k^{(i)}}{m} -1,\\rho_k^{(j)},\\lambda_k^{(j)}\\Big)\\) // Penalties \n5: \\(\\mathcal{L} = \\frac{1}{B} (\\mathcal{L}_{\\mathrm{c}} + \\mathcal{L}_{\\mathrm{p}})\\) \n6: \\(\\theta^{(t + 1)}\\gets \\theta^{(t)} - \\alpha \\cdot \\nabla_{\\theta}\\mathcal{L}\\) // Gradient descent \n7: end for \n8: for \\(k = 1,\\ldots ,K\\) do // Adjust \\(\\lambda\\) and \\(\\rho\\) on validation \n9: \\(\\lambda_{k}^{(j + 1)} = \\frac{1}{|\\mathcal{D}_{\\mathrm{val}}|}\\sum_{(\\boldsymbol {x},\\boldsymbol {y})\\in \\mathcal{D}_{\\mathrm{val}}}P'\\left(\\frac{d_k}{m} -1,\\rho_k^{(j)},\\lambda_k^{(j)}\\right)\\) \n10: \\(\\overline{d_k}^{(j)} = \\frac{1}{|\\mathcal{D}_{\\mathrm{val}}|}\\sum_{(\\boldsymbol {x},\\boldsymbol {y})\\in \\mathcal{D}_{\\mathrm{val}}}d_k - 1\\) // Average constraint \n11: if \\(j\\geq 1\\) and \\(\\overline{d_k}^{(j)} > \\tau \\max \\{0,\\overline{d_k}^{(j - 1)}\\}\\) then \n12: \\(\\rho_{k}^{(j + 1)}\\gets \\gamma \\rho_{k}^{(j)}\\) // Constraint has not improved \n13: else \n14: \\(\\rho_{k}^{(j + 1)}\\gets \\rho_{k}^{(j)}\\) \n15: end if \n16: end for \n17: end for" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.509, + 0.894, + 0.556 + ], + "angle": 0, + "content": "are grouped into \\( M \\) equi-spaced bins according to prediction confidence, and a weighted average of the absolute difference between accuracy and confidence in each bin is calculated:" + }, + { + "type": "equation", + "bbox": [ + 0.588, + 0.578, + 0.892, + 0.62 + ], + "angle": 0, + "content": "\\[\nE C E = \\sum_ {m = 1} ^ {M} \\frac {\\left| B _ {m} \\right|}{N} \\left| A _ {m} - C _ {m} \\right|, \\tag {14}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.629, + 0.895, + 0.81 + ], + "angle": 0, + "content": "where \\(M\\) is the number of bins, \\(N\\) the amount of test samples, \\(B_{m}\\) the samples with prediction confidence in the \\(m^{th}\\) bin, \\(A_{m}\\) the accuracy and \\(C_{m}\\) the mean confidence of samples in the \\(m^{th}\\) bin. Note we fix \\(N\\) to 15 according to [22, 31]. In accordance with [22], we also report Adaptive ECE (AECE), a variant of ECE where the bins are configured to evenly distribute the test samples across them. Additionally, Classwise Calibration Error (CWCE) [29], a classwise extension of ECE, is included in Appendix C. For discriminative performance, we use standard measures: accuracy (Acc) for classification, and intersection over union (mIoU) for segmentation." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.811, + 0.894, + 0.902 + ], + "angle": 0, + "content": "Compared methods. We compare our method to other learning based calibration losses, including (i) methods that impose constraints on predictions (either logits or softmax probabilities), i.e. Explicit Confidence Penalty (ECP) [38], Label Smoothing (LS) [42], Focal Loss (FL) [21] and its sample-dependent version (FLSD) [31], Margin-based Label" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16075" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.123, + 0.089, + 0.85, + 0.245 + ], + "angle": 0, + "content": "
MethodTinyImageNetImageNetImageNet-LT
ResNet-50ResNet-50SwinV2-TResNet-50SwinV2-T
AccECEAECEAccECEAECEAccECEAECEAccECEAECEAccECEAECE
CE65.023.733.6975.169.199.1875.609.959.9437.9028.1228.1231.8231.8236.68
MMCE [18]65.342.812.6174.858.578.5676.689.079.0837.7928.4128.4033.1426.4126.41
ECP [38]64.904.003.9275.228.278.2675.829.889.8637.6928.1428.1331.2233.7033.70
LS [42]65.783.173.1676.042.572.8875.427.327.3337.8810.4610.3831.7011.4211.40
FL [31]63.092.963.1273.871.601.6575.603.193.1836.0418.3718.3630.7325.5025.50
FLSD [31]64.092.912.9573.972.082.0674.702.442.3736.1817.7717.7832.5625.1625.17
CPC [5]64.494.884.9176.333.663.5976.345.505.3338.9016.0015.9932.5413.2113.19
MbLS [22]64.741.641.7375.824.444.2677.181.951.7338.326.166.1632.057.657.64
CALS-HR65.092.502.4276.345.635.6977.583.062.9538.502.832.7834.312.372.45
CALS-ALM65.031.541.3876.441.461.3277.101.611.6938.562.152.3033.942.322.45
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.255, + 0.894, + 0.283 + ], + "angle": 0, + "content": "Table 1. Calibration performance for different approaches on three image classification benchmarks. We report two lower-is-better calibration metrics, i.e. ECE and AECE. Best method is highlighted in bold, while the second-best one is underlined." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.298, + 0.471, + 0.405 + ], + "angle": 0, + "content": "Smoothing (MbLS) [22] and CPC [5], and (ii) techniques that directly optimize calibration measures, i.e. MMCE [18]. We refer to the related literature [22, 31] to set the hyperparameters for various methods. For instance, the smoothing factor in LS and FL is set to 0.05 and 3 respectively, and we set margin to 10 in MbLS. A detailed description of hyper-parameter values can be found in Appendix F." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.406, + 0.472, + 0.485 + ], + "angle": 0, + "content": "Our methods. A simple alternative to the algorithm presented in Section 4.2 would be to heuristically tune multipliers by scaling them according to penalty values: if \\( P_{k}^{(j + 1)} \\) increases we also increase \\( \\lambda_{k}^{j + 1} \\) and vice versa. This strategy, akin to learning rate scheduling, can be formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.131, + 0.497, + 0.47, + 0.565 + ], + "angle": 0, + "content": "\\[\n\\lambda_ {k} ^ {(j + 1)} = \\left\\{ \\begin{array}{l l} \\mu \\lambda_ {k} ^ {(j)} & \\text {i f} P _ {k} ^ {(j + 1)} > \\tau P _ {k} ^ {(j)} \\\\ \\lambda_ {k} ^ {(j)} / \\mu & \\text {i f} P _ {k} ^ {(j)} > \\tau P _ {k} ^ {(j + 1)} \\\\ \\lambda_ {k} ^ {(j)} & \\text {o t h e r w i s e} \\end{array} \\right. \\tag {15}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.577, + 0.47, + 0.621 + ], + "angle": 0, + "content": "where \\(\\mu > 1\\) and \\(\\tau > 1\\) are hyper-parameters that we fix to 1.1. We refer to our main algorithm as CALS-ALM and to this heuristic rule as CALS-HR in what follows." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.623, + 0.472, + 0.729 + ], + "angle": 0, + "content": "We fix the margin to \\( m = 10 \\) on vision tasks and \\( m = 6 \\) on the NLP benchmark, as in [22], for a fair comparison. We also perform an ablation study to investigate the impact of the margin value. For other hyper-parameters, we set \\( \\pmb{\\lambda}^{(0)} = 10^{-6} \\cdot \\mathbf{1}_K \\), \\( \\pmb{\\rho}^{(0)} = \\mathbf{1}_K \\), \\( \\gamma = 1.2 \\), and we update the penalty parameters \\( \\pmb{\\rho} \\) every 10 epochs. Please refer to Appendix F for a detailed description of all hyper-parameters." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.73, + 0.472, + 0.835 + ], + "angle": 0, + "content": "Implementation Details. For image classification, we experiment with ResNet [13] and a vision Transformer model, i.e. Swin Transformer V2 (SwinV2-T) [23]. DeepLabV3 [3] is employed for semantic segmentation on PASCAL VOC2012. Following [22, 31], we use the Global Pooling CNN (GPoolCNN) architecture [20] on the NLP recognition task. Further training details on each dataset can be found in Appendix B." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.847, + 0.172, + 0.86 + ], + "angle": 0, + "content": "5.2. Results" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.871, + 0.473, + 0.901 + ], + "angle": 0, + "content": "Results on image classification. Table 1 presents the discriminative and calibration performance of our methods" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.298, + 0.895, + 0.785 + ], + "angle": 0, + "content": "across three widely used classification benchmarks, compared to baselines and related works. We can observe that our CALS-ALM approach consistently outperforms existing techniques in terms of calibration. Specifically, the results indicate that the standard CE loss and other approaches often lead to miscalibrated models, with the severity of miscalibration substantially increasing in correlation with dataset difficulty. This is particularly evident in large-scaled datasets with numerous classes, such as ImageNet, or those with long-tail class distributions, like ImageNet-LT. Although other learning-based methods could provide better calibrated networks, their performance is not stable across different settings. For example, while FL achieves a relatively low ECE of 1.60 with a Resnet50 trained on ImageNet, it only yields an ECE of 25.50 when training a SwinV2-T network on ImageNet-LT, revealing a limitation in adapting to different learning scenarios. In contrast, CALS-ALM attains the best calibration performance in all cases, frequently outperforming existing approaches by a significant margin, with minimal variations across datasets and architectures. This trends persists when compared to the most closely related technique, MbLS. It is noteworthy that the improvements on the long-tailed ImageNet-LT dataset are substantial; for example, we decrese ECE from 28.12 to 2.15 for ResNet50, and from 31.82 to 2.32 for SwinV2-T, validating the effectiveness of class adaptive learning. Another interesting finding is that CALS-HR, employing a naive update strategy for class-wise penalty weights, achieves nearly the second-best performance in the ImageNet-LT dataset. This further demonstrates the effectiveness of CALS for learning under class-imbalance scenarios. For the reliability diagrams of various models, please refer to Appendix E." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.796, + 0.897, + 0.903 + ], + "angle": 0, + "content": "In terms of model accuracy, our method delivers competitive performances, surpassing existing methods in certain cases. It is important to emphasize that, while the proposed method achieves discriminative results comparable to the best-performing approach for each dataset, the differences in calibration are considerable. This highlights the superiority of the proposed formulation for training highly discrimi" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "16076" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.083, + 0.092, + 0.357, + 0.27 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.086, + 0.275, + 0.355, + 0.288 + ], + "angle": 0, + "content": "(a) Validation ECE and multipliers during training." + }, + { + "type": "image", + "bbox": [ + 0.389, + 0.091, + 0.65, + 0.269 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.443, + 0.275, + 0.835, + 0.288 + ], + "angle": 0, + "content": "(b) Effect of penalty function and margin on accuracy and ECE for CALS." + }, + { + "type": "image", + "bbox": [ + 0.656, + 0.092, + 0.885, + 0.27 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.3, + 0.894, + 0.342 + ], + "angle": 0, + "content": "Figure 3. Ablation study on ImageNet-LT. (a) Evolution of ECE on validation and multipliers for CALS: ECE on the validation set for our method (CALS), CE and MbLS [22] and values of multipliers \\(\\lambda\\) for CALS after each training epoch. (b) Effect of penalty functions and margin: ECE and accuracy on validation and test set are shown across different choices of penalty functions and margin values." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.368, + 0.471, + 0.429 + ], + "angle": 0, + "content": "native and well-calibrated networks. Figure 1 provides a more visual comparison considering both accuracy and ECE. It is demonstrated that CALS-ALM provides the optimal compromise between accuracy and calibration performance." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.43, + 0.47, + 0.656 + ], + "angle": 0, + "content": "Ablation Analysis. Figure 3 illustrates the evolution of ECE in (a) and penalty multipliers \\(\\lambda\\) in (b) during training. It is interesting to observe that the evolution of \\(\\lambda\\) is consistent with the ECE. Specifically, the average penalty weight gradually increases while the ECE initially deteriorates because the model is focused on increasing accuracy. However, the value of the penalty weight begins to decline alongside the ECE, as the network starts to become better calibrated. For a visualization of classwise multipliers, please refer to Appendix D. Figure 3(c) highlights the impact of the choice of penalty functions and margin values. This demonstrates that the PHR penalty function is preferable over the other two options, P2 and P3, for both calibration and accuracy. Regarding the margin, the best performance is achieved with \\(m \\approx 10\\), which is consistent with the findings in [22]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.658, + 0.47, + 0.824 + ], + "angle": 0, + "content": "Semantic Segmentation. Table 2 presents the performance on Pascal VOC dataset. Note here CALS refers to our best method, i.e. CALS-ALM. It can be observed that the trend is consistent with image classification experiments: CALS outperforms counterparts in terms of ECE, and yields competitive results on discriminative performance, i.e. mIoU in segmentation. It is worth noting that some methods like MMCE, FLSD and CPC, are not included here because their computation demands were too heavy for pixel-wise segmentation tasks. In contrast, our method is unlimited in dense prediction tasks as the computation cost it adds is moderate." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.471, + 0.901 + ], + "angle": 0, + "content": "Text Classification. Last, we demonstrate the general applicability of the proposed method by analyzing its performance on a non-vision task, i.e. text classification on 20 Newsgroups dataset. The results are reported in Table 3. Remarkably, CALs again brings substantial improvement in" + }, + { + "type": "table", + "bbox": [ + 0.543, + 0.365, + 0.851, + 0.415 + ], + "angle": 0, + "content": "
CEECPLSFLMbLSCALS
ECE14.755.156.9010.875.224.66
mIoU66.4665.5767.7364.2565.2966.77
" + }, + { + "type": "table_caption", + "bbox": [ + 0.522, + 0.425, + 0.869, + 0.44 + ], + "angle": 0, + "content": "Table 2. Segmentation results on the PASCAL VOC 2012." + }, + { + "type": "table", + "bbox": [ + 0.504, + 0.451, + 0.889, + 0.5 + ], + "angle": 0, + "content": "
CEMMCEECPLSFLFLSDCPCMbLSCALS
ECE (%)22.7523.0222.978.0710.8010.879.465.402.04
Acc (%)67.0166.2366.4867.1466.0865.8568.2767.8968.32
" + }, + { + "type": "table_caption", + "bbox": [ + 0.507, + 0.51, + 0.882, + 0.524 + ], + "angle": 0, + "content": "Table 3. Results on the text classification task, 20 Newsgroups." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.538, + 0.894, + 0.614 + ], + "angle": 0, + "content": "terms of calibration, with ECE decreasing to \\(2.04\\%\\), while yielding the best accuracy \\(68.32\\%\\). This reveals that the proposed class adaptive learning method is also able to handle class differences in NLP applications and provide promising performance in terms of both accuracy and calibration." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.626, + 0.772, + 0.642 + ], + "angle": 0, + "content": "6. Limitations and Future Work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.652, + 0.895, + 0.788 + ], + "angle": 0, + "content": "We have proposed Class Adaptive Label Smoothing for network calibration based on a modified Augmented Lagrangian Multiplier algorithm. Despite its superior performance over previous methods, there are potential limitations in this work. For instance, our method requires the validation set to have the same distribution as the training set. Although this is satisfied in nearly every benchmark, it will be interesting to investigate the impact of using non-independent and identically distributed (i.i.d) validation sets." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.801, + 0.667, + 0.817 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.825, + 0.894, + 0.901 + ], + "angle": 0, + "content": "This work is supported by National Science and Engineering Research Council of Canada (NSERC) and Prompt Quebec. We also thank Calcul Quebec and Compute Canada. Adrian Galdran is supported by a Marie Sklodowska-Curie Fellowship (No 892297)." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16077" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.175, + 0.107 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.156 + ], + "angle": 0, + "content": "[1] Dimitri P. Bertsekas. Constrained Optimization and Lagrange Multiplier Methods (Optimization and Neural Computation Series). Athena Scientific, 1 edition, 1996. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.157, + 0.472, + 0.212 + ], + "angle": 0, + "content": "[2] Ernesto G Birgin, Romulo A Castillo, and José Mario Martínez. Numerical comparison of augmented lagrangian algorithms for nonconvex problems. Computational Optimization and Applications, 31(1), 2005. 5, 6, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.213, + 0.47, + 0.255 + ], + "angle": 0, + "content": "[3] Liang-Chieh Chen, George Papandreou, Florian Schroff, and Hartwig Adam. Rethinking atrous convolution for semantic image segmentation. In CVPR, 2017. 1, 7, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.256, + 0.471, + 0.297 + ], + "angle": 0, + "content": "[4] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In ICML, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.298, + 0.471, + 0.338 + ], + "angle": 0, + "content": "[5] Jiacheng Cheng and Nuno Vasconcelos. Calibrating deep neural networks by pairwise constraints. In CVPR, 2022. 3, 7, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.34, + 0.471, + 0.395 + ], + "angle": 0, + "content": "[6] Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The cityscapes dataset for semantic urban scene understanding. In CVPR, June 2016. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.396, + 0.471, + 0.437 + ], + "angle": 0, + "content": "[7] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In CVPR, 2009. 6, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.438, + 0.471, + 0.479 + ], + "angle": 0, + "content": "[8] Zhipeng Ding, Xu Han, Peirong Liu, and Marc Niethammer. Local temperature scaling for probability calibration. In ICCV, 2021. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.48, + 0.471, + 0.535 + ], + "angle": 0, + "content": "[9] Mark Everingham, S. M. Eslami, Luc Gool, Christopher K. Williams, John Winn, and Andrew Zisserman. The pascal visual object classes challenge: A retrospective. *IJCV*, 111(1):98–136, 2015. 4, 6, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.537, + 0.471, + 0.578 + ], + "angle": 0, + "content": "[10] Yarin Gal and Zoubin Ghahramani. Dropout as a bayesian approximation: Representing model uncertainty in deep learning. In ICML, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.579, + 0.471, + 0.619 + ], + "angle": 0, + "content": "[11] Chuan Guo, Geoff Pleiss, Yu Sun, and Kilian Q Weinberger. On calibration of modern neural networks. In ICML, 2017. 1, 2, 3, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.621, + 0.471, + 0.661 + ], + "angle": 0, + "content": "[12] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollar, and Ross Girshick. Masked autoencoders are scalable vision learners. In CVPR, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.663, + 0.471, + 0.703 + ], + "angle": 0, + "content": "[13] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, 2016. 1, 7, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.705, + 0.471, + 0.746 + ], + "angle": 0, + "content": "[14] Magnus R. Hestenes. Multiplier and gradient methods. Journal of Optimization Theory and Applications, 4:303-320, 1969. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.747, + 0.471, + 0.788 + ], + "angle": 0, + "content": "[15] Sergey Ioffe and Christian Szegedy. Batch normalization: Accelerating deep network training by reducing internal covariate shift. In ICML, 2015. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.789, + 0.471, + 0.844 + ], + "angle": 0, + "content": "[16] Archit Karandikar, Nicholas Cain, Dustin Tran, Balaji Lakshminarayanan, Jonathon Shlens, Michael C Mozer, and Rebecca Roelofs. Soft calibration objectives for neural networks. In NeurIPS, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.845, + 0.471, + 0.899 + ], + "angle": 0, + "content": "[17] Barry W. Kort and Dimitri P. Bertsekas. Combined primal-dual and penalty methods for convex programming. SIAM Journal on Control and Optimization, 14(2):268-294, 1976. 5" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.472, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.134 + ], + "angle": 0, + "content": "[18] Aviral Kumar, Sunita Sarawagi, and Ujjwal Jain. Trainable calibration measures for neural networks from kernel mean embeddings. In ICML, 2018. 3, 7, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.136, + 0.892, + 0.16 + ], + "angle": 0, + "content": "[19] Ken Lang. Newsweeder: Learning to filter netnews. In ICML, 1995. 6, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.162, + 0.892, + 0.189 + ], + "angle": 0, + "content": "[20] Min Lin, Qiang Chen, and Shuicheng Yan. Network in network. In ICLR, 2014. 7, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.191, + 0.892, + 0.232 + ], + "angle": 0, + "content": "[21] Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dólar. Focal loss for dense object detection. In CVPR, 2017. 1, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.233, + 0.892, + 0.286 + ], + "angle": 0, + "content": "[22] Bingyuan Liu, Ismail Ben Ayed, Adrian Galdran, and Jose Dolz. The devil is in the margin: Margin-based label smoothing for network calibration. In CVPR, 2022. 1, 2, 3, 6, 7, 8, 11, 12, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.288, + 0.892, + 0.343 + ], + "angle": 0, + "content": "[23] Ze Liu, Han Hu, Yutong Lin, Zhuliang Yao, Zhenda Xie, Yixuan Wei, Jia Ning, Yue Cao, Zheng Zhang, Li Dong, Furu Wei, and Baining Guo. Swin transformer v2: Scaling up capacity and resolution. In CVPR, 2022. 7, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.344, + 0.892, + 0.397 + ], + "angle": 0, + "content": "[24] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In ICCV, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.399, + 0.892, + 0.44 + ], + "angle": 0, + "content": "[25] Ziwei Liu, Zhongqi Miao, Xiaohang Zhan, Jiayun Wang, Boqing Gong, and Stella X. Yu. Large-scale long-tailed recognition in an open world. In CVPR, 2019. 6, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.441, + 0.892, + 0.468 + ], + "angle": 0, + "content": "[26] Ilya Loshchilov and Frank Hutter. SGDR: Stochastic gradient descent with warm restarts. In *ICLR*, 2017. 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.469, + 0.892, + 0.496 + ], + "angle": 0, + "content": "[27] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In *ICLR*, 2019. 1, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.497, + 0.892, + 0.536 + ], + "angle": 0, + "content": "[28] Xingchen Ma and Matthew B. Blaschko. Meta-cal: Well-controlled post-hoc calibration by ranking. In ICML, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.538, + 0.892, + 0.579 + ], + "angle": 0, + "content": "[29] Lena Maier-Hein et al. Metrics reloaded: Pitfalls and recommendations for image analysis validation. arXiv preprint arXiv:2206.01653, 2022. 6, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.581, + 0.892, + 0.635 + ], + "angle": 0, + "content": "[30] Matthias Minderer, Josip Djolonga, Rob Romijnders, Frances Hubis, Xiaohua Zhai, Neil Houlsby, Dustin Tran, and Mario Lucic. Revisiting the calibration of modern neural networks. In NeurIPS, 2021. 3, 11, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.637, + 0.892, + 0.691 + ], + "angle": 0, + "content": "[31] Jishnu Mukhoti, Viveka Kulharia, Amartya Sanyal, Stuart Golodetz, Philip HS Torr, and Puneet K Dokania. Calibrating deep neural networks using focal loss. In NeurIPS, 2020. 1, 2, 3, 6, 7, 11, 12, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.692, + 0.892, + 0.72 + ], + "angle": 0, + "content": "[32] Rafael Müller, Simon Kornblith, and Geoffrey Hinton. When does label smoothing help? In NeurIPS, 2019. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.721, + 0.892, + 0.761 + ], + "angle": 0, + "content": "[33] Mahdi Pakdaman Naeini, Gregory F. Cooper, and Milos Hauskrecht. Obtaining well calibrated probabilities using bayesian binning. In AAAI, 2015. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.762, + 0.892, + 0.803 + ], + "angle": 0, + "content": "[34] H. Nakayama, H. Samaya, and H. Samaya. A generalized lagrangian function and multiplier method. Journal of Optimization Theory and Applications, 17:211-227, 1975. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.804, + 0.892, + 0.831 + ], + "angle": 0, + "content": "[35] Jorge Nocedal and Stephen J. Wright. Numerical Optimization. Springer, New York, NY, USA, 2 edition, 2006. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.832, + 0.892, + 0.899 + ], + "angle": 0, + "content": "[36] Yaniv Ovadia, Emily Fertig, Jie Ren, Zachary Nado, David Sculley, Sebastian Nowozin, Joshua V Dillon, Balaji Lakshminarayanan, and Jasper Snoek. Can you trust your model's uncertainty? evaluating predictive uncertainty under dataset shift. In NeurIPS, 2019. 3" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "16078" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.133 + ], + "angle": 0, + "content": "[37] Jeffrey Pennington, Richard Socher, and Christopher D Manning. Glove: Global vectors for word representation. In EMNLP, 2014. 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.136, + 0.472, + 0.19 + ], + "angle": 0, + "content": "[38] Gabriel Pereyra, George Tucker, Jan Chorowski, Lukasz Kaiser, and Geoffrey Hinton. Regularizing neural networks by penalizing confident output distributions. In ICLR, 2017. 1, 3, 6, 7, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.192, + 0.472, + 0.233 + ], + "angle": 0, + "content": "[39] M. J. D. Powell. A method for nonlinear constraints in minimization problems. In *Fletcher, R., Ed., Optimization, Academic Press, New York, NY*, 1969. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.235, + 0.472, + 0.275 + ], + "angle": 0, + "content": "[40] Jérôme Rony, Eric Granger, Marco Pedersoli, and Ismail Ben Ayed. Augmented lagrangian adversarial attacks. In ICCV, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.278, + 0.472, + 0.332 + ], + "angle": 0, + "content": "[41] Sara Sangalli, Ertunc Erdil, Andeas Hotker, Olivio F Donati, and Ender Konukoglu. Constrained optimization to train neural networks on critical and under-represented classes. In NeurIPS, 2021. 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.334, + 0.472, + 0.387 + ], + "angle": 0, + "content": "[42] Christian Szegedy, Vincent Vanhoucke, Sergey Ioffe, Jon Shlens, and Zbigniew Wojna. Rethinking the inception architecture for computer vision. In CVPR, 2016. 3, 6, 7, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.391, + 0.472, + 0.444 + ], + "angle": 0, + "content": "[43] Christian Tomani, Sebastian Gruber, Muhammed Ebrar Erdem, Daniel Cremers, and Florian Buettner. Post-hoc uncertainty calibration for domain drift scenarios. In CVPR, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.447, + 0.472, + 0.489 + ], + "angle": 0, + "content": "[44] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In NeurIPS, 2017. 1" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.489 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.956 + ], + "angle": 0, + "content": "16079" + } + ] +] \ No newline at end of file diff --git a/2023/Class Adaptive Network Calibration/d3a59c4a-62e4-447c-b052-c6b11b8ed279_origin.pdf b/2023/Class Adaptive Network Calibration/d3a59c4a-62e4-447c-b052-c6b11b8ed279_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..f16f1ee8ee57e9fcd53cbf1a7bfd758baad644f3 --- /dev/null +++ b/2023/Class Adaptive Network Calibration/d3a59c4a-62e4-447c-b052-c6b11b8ed279_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b717127575cddb08bc31f4ba57bb06e417a9662b3cc4fcf10781a132d32b0e6e +size 696989 diff --git a/2023/Class Adaptive Network Calibration/full.md b/2023/Class Adaptive Network Calibration/full.md new file mode 100644 index 0000000000000000000000000000000000000000..103d9ad54be909290f93ccf54f52344abba49383 --- /dev/null +++ b/2023/Class Adaptive Network Calibration/full.md @@ -0,0 +1,360 @@ +# Class Adaptive Network Calibration + +Bingyuan Liu\*, Jérôme Rony\*, Adrian Galdran\*, Jose Dolz\*, Ismail Ben Ayed + +1ÉTS Montreal, Canada 2Universitat Pompeu Fabra, Barcelona, Spain + +# Abstract + +Recent studies have revealed that, beyond conventional accuracy, calibration should also be considered for training modern deep neural networks. To address miscalibration during learning, some methods have explored different penalty functions as part of the learning objective, alongside a standard classification loss, with a hyper-parameter controlling the relative contribution of each term. Nevertheless, these methods share two major drawbacks: 1) the scalar balancing weight is the same for all classes, hindering the ability to address different intrinsic difficulties or imbalance among classes; and 2) the balancing weight is usually fixed without an adaptive strategy, which may prevent from reaching the best compromise between accuracy and calibration, and requires hyper-parameter search for each application. We propose Class Adaptive Label Smoothing (CALS) for calibrating deep networks, which allows to learn class-wise multipliers during training, yielding a powerful alternative to common label smoothing penalties. Our method builds on a general Augmented Lagrangian approach, a well-established technique in constrained optimization, but we introduce several modifications to tailor it for large-scale, class-adaptive training. Comprehensive evaluation and multiple comparisons on a variety of benchmarks, including standard and long-tailed image classification, semantic segmentation, and text classification, demonstrate the superiority of the proposed method. The code is available at https://github.com/by-liu/CALS. + +# 1. Introduction + +Deep Neural Networks (DNNs) have become the prevailing model in machine learning, particularly for computer vision [13] and natural language processing applications [44]. Increasingly powerful architectures [3, 13, 24], learning methods [4, 12] and a large body of other techniques [15, 27] are constantly introduced. Nonetheless, recent studies [11, 31] have shown that regardless of their superior discriminative + +performance, high-capacity modern DNNs are poorly calibrated, i.e. failing to produce reliable predictive confidences. Specifically, they tend to yield over-confident predictions, where the probability associated with the predicted class overestimates the actual likelihood. Since this is a critical issue in safety-sensitive applications like autonomous driving or computational medical diagnosis, the problem of DNN calibration has been attracting increasing attention in recent years [11, 31, 38]. + +Current calibration methods can be categorized into two main families. The first family involves techniques that perform an additional post-processing parameterized operation on the output logits (or pre-softmax activations) [11], with the calibration parameters of that operation obtained from a validation set by either learning or grid-search. Despite the simplicity and low computational cost, these methods have empirically proven to be highly effective [8, 11]. However, their main drawback is that the choice of the optimal calibration parameters is highly sensitive to the trained model instance and validation set [22, 31]. + +The second family of methods attempts to simultaneously optimize for accuracy and calibration during network training. This is achieved by introducing, explicitly or implicitly, a secondary optimization goal involving the model's predictive uncertainty, alongside the main training objective. As a result, a scalar balancing hyper-parameter is required to tune the relative contribution of each term in the overall loss function. Some examples of this type of approaches include: Explicit Confidence Penalty (ECP) [38], Label Smoothing (LS) [32], Focal Loss (FL) [21] and its variant, Sample-Dependent Focal Loss (FLSD) [31]. It has been recently demonstrated in [22] that all these methods can be formulated as different penalty terms that enforce the same equality constraint on the logits of the DNN: driving the logit distances towards zero. Here, logit distances refers to the vector of L1 distances between the highest logit value and the rest. Observing the non-informative nature of this equality constraint, [22] proposed to use a generalized inequality constraint, only penalizing those logits for which the distance is larger than a pre-defined margin, achieving state-of-the-art calibration performance on many different benchmarks. + +Although learning based methods achieve greater calibra + +![](images/efbbb40bcc5c3e8826b7a9fcd227fec317cc07d4a7186c95cc9452b851551f17.jpg) +Figure 1. Many techniques have been proposed for jointly improving accuracy and calibration during training [11,31], but they fail to consider uneven learning scenarios like high class imbalance or long-tail distributions. We show a comparison of the proposed CALS-ALM method and different learning approaches in terms of Calibration Error (ECE) vs Accuracy on the (a) ImageNet and (b) ImageNet-LT (long-tailed ImageNet) datasets. A lower ECE indicates better calibration: a better model should attain high ACC and low ECE. Among all the considered methods, CALS-ALM shows superior performance when considering both discriminative power and well-balanced probabilistic predictions, achieving best accuracy and calibration on ImageNet, and best calibration and second best accuracy on ImageNet-LT. + +![](images/f3a4a14c56f26e6becfa638ee582811403c0d1e8809aa9037ba042ebd4abcdfd.jpg) + +tion performance [22, 31], they have two major limitations: 1) The scalar balancing weight is equal for all classes. This hinders the network performance when some classes are harder to learn or less represented than others, such as in datasets with a large number of categories (ImageNet) or considerable class imbalance (ImageNet-LT). 2) The balancing weight is usually fixed before network optimization, with no learning or adaptive strategy throughout training. This can prevent the model from reaching the best compromise between accuracy and calibration. To address the above issues, we introduce Class Adaptive Label Smoothing method based on an Augmented Lagrangian Multiplier algorithm, which we refer to as CALS-ALM. Our Contributions can be summarized as follows: + +- We propose Class Adaptive Label Smoothing (CALS) for network calibration. Adaptive class-wise multipliers are introduced instead of the widely used single balancing weight, which addresses the above two issues: 1) CALS can handle a high number of classes with different intrinsic difficulties, e.g. ImageNet; 2) CALS can effectively learn from data suffering from class imbalance or a long-tailed distribution, e.g. ImageNet-LT. +- Different from previous penalty based methods, we solve the resulting constrained optimization problem by implementing a modified Augmented Lagrangian Multiplier (ALM) algorithm, which yields adaptive and optimal weights for the constraints. We make some critical design decisions in order to adapt ALM to the nature of modern learning techniques: 1) The inner convergence criterion in ALM is relaxed to a fixed number of iterations in each inner stage, which is amenable to mini-batch stochas + +tic gradient optimization in deep learning. 2) Popular techniques, such as data augmentation, batch normalization [15] and dropout [10], rule out the possibility of tracking original samples and applying sample-wise multipliers. To overcome this complication, we introduce class-wise multipliers, instead of sample-wise multipliers in the standard ALM. 3) The outer-step update for estimating optimal ALM multipliers is performed on the validation set, which is meaningful for training on large-scale training set and avoids potential overfitting. + +- Comprehensive experiments over a variety of applications and benchmarks, including standard image classification (Tiny-ImageNet and ImageNet), long-tailed image classification (ImageNet-LT), semantic segmentation (PASCAL VOC 2012), and text classification (20 Newsgroups), demonstrate the effectiveness of our CALS-ALM method. As shown in Figure 1, CALS-ALM yields superior performance over baselines and state-of-the-art calibration losses when considering both accuracy and calibration, especially for more realistic large-scaled datasets with large number of classes or class imbalance. + +# 2. Related Work + +# 2.1. Problem Formulation + +Given a dataset $\mathcal{D} = \{(x^{(i)},y^{(i)})\}_{i = 1}^{N}$ with $N\in \mathbb{N}$ pairs of samples $\pmb {x}\in \mathcal{X}$ and corresponding labels $y\in \mathcal{V}$ with $\mathcal{V} = \{1,\dots ,K\}$ , a deep neural network (DNN) $F_{\theta}:X\to \mathbb{R}^{K}$ parameterized by $\theta$ yields logits $l = F_{\theta}(\pmb {x}) = (F_{\theta}(\pmb {x})_k)_{1\leq k\leq K}\in \mathbb{R}^K$ . In a classification scenario, the output probability $s = (s_k)_{1\leq k\leq K}\in \Delta^{K - 1}$ where $\Delta^{K - 1}\subset [0,1]^K$ denotes the probability simplex, is + +obtained by applying the softmax function on the output log-its, i.e. $s = \text{softmax}(l) = \frac{\exp l}{\sum \exp l}$ . Therefore, the predicted class $\hat{y}$ is computed as $\hat{y} = \arg \max_k s_k$ , and the predicted confidence is $\hat{p} = s_{\hat{y}} = \max_k s_k$ . A perfectly calibrated model should satisfy that the predicted confidence of any input is equal to the accuracy of the model: $\hat{p} = \mathbb{P}(\hat{y} = y|\hat{p})$ . Hence, an over-confident model yields on average larger confidence than the associated accuracy, whereas an under-confident model yields lower confidence than its accuracy. + +A number of recent studies [11, 22, 30, 31] have shown that DNNs tend to become over-confident during training as a result of minimizing the popular cross-entropy (CE) training loss: + +$$ +\mathcal {L} _ {\mathrm {C E}} (\boldsymbol {x}, y) = - \sum_ {k = 1} ^ {K} y _ {k} \log s _ {k} \tag {1} +$$ + +where $\pmb{y} \in \{0,1\}^{K}$ is the one-hot encoding of $y$ . This objective function is minimized when the predictions for all the training samples fully match the ground-truth labels $y$ , i.e. $s_{y} = 1$ and $\forall k \neq y, s_{k} = 0$ . The negative logarithmic term on the logit of the correct category renders the global minimization of the CE loss unreachable, as it keeps pushing the predicted probabilities $s$ towards the vertices of the $(K-1)$ -simplex even after the classification error is zero [31], resulting in over-confident models. + +# 2.2. Post-processing methods + +To address mis-calibration, different post-processing techniques applied after model training have been proposed recently [8, 11, 43]. The most popular of these strategies is Temperature Scaling (TS) [11], which applies a single scalar temperature parameter to manipulate logit outputs monotonely, resulting in softened prediction confidences without affecting predicted labels. Note that here the temperature parameter needs to be tuned on a separate validation set. Despite its simplicity, TS has been shown effective in fixing over-confidence predictions [11]. As a local alternative to TS, [8] proposed to train a regression model for learning position-specific temperature for semantic segmentation problems. Unfortunately, TS and its variants can be sensitive to both the model and the validation set, and do not work well under data distribution shifts [36]. Thus, some subsequent works [28, 43] have attempted to provide solutions for improving performance under domain shift. + +# 2.3. Learning-based methods + +Another popular direction is to directly address miscalibration during training by introducing an additional penalty or supervision regarding model calibration with the standard training loss. In [18], the authors introduced a trainable calibration measure based on RKHS kernels, while [16] + +proposed a differential calibration loss based on a soft version of the binning operation in the ECE metric. In [5], two types of binary pairwise calibration constraints were proposed as additional penalty terms during training. Other methods try to decrease over-fitting on the cross-entropy loss, which has been demonstrated to be the main reason of over-confidence [11, 31]. In [38] an explicit confidence penalty (ECP) is proposed to maximize the entropy and reduce over-fitting, while Label Smoothing [42] has also been shown to implicitly improve the calibration [32] by softening the hard one-hot targets in the cross-entropy. The Focal Loss [21], originally proposed to tackle class imbalance, can also be effective for calibration [31], as it implicitly minimizes the Kullback-Leibler (KL) divergence between the uniform distribution and the network softmax probabilities, thereby increasing the entropy of the predictions. As an extension the Sample-Dependent Focal Loss (FLSD) was also proposed in [31] to further boost calibration performance. + +Margin-based Label Smoothing (MbLS). A unifying constrained-optimization formulation of loss functions promoting calibration has been recently presented in [22]. Specifically, the additional penalties integrated in these methods, including ECP [38], LS [32] and FL [31], can be viewed as different forms of approximations to the same constraint, i.e. enforcing the logit distances to be zero. Noticing that this constraint is non-informative (its solution being uniformly distributed probabilities), [22] further proposed a generalized formulation by relaxing the constraint to allow the logit distances being lower than a strictly positive margin. + +The specific formulation of MbLS [22] is as follows. Given a margin $m \in \mathbb{R}_+$ , the constrained optimization problem for network training is: + +$$ +\min _ {\theta} \sum_ {i = 1} ^ {N} \mathcal {L} _ {\mathrm {C E}} \left(\boldsymbol {x} ^ {(i)}, y ^ {(i)}\right) \tag {2} +$$ + +$$ +\text {s . t .} \quad \max _ {k} \left\{\boldsymbol {l} _ {k} ^ {(i)} \right\} - l ^ {(i)} \preceq m \mathbf {1} _ {K}, \quad i = 1, \dots , N, +$$ + +where $l^{(i)} = F_{\theta}(\pmb{x}^{(i)})$ . The minimum can be approximated by penalty-based optimization methods, transforming the above formulation into an unconstrained problem by means of simple ReLU functions: + +$$ +\min _ {\theta} \sum_ {i = 1} ^ {N} \mathcal {L} _ {\mathrm {C E}} \left(\boldsymbol {x} ^ {(i)}, y ^ {(i)}\right) + \tag {3} +$$ + +$$ +\lambda \sum_ {i = 1} ^ {N} \sum_ {j = 1} ^ {K} \max \{0, \max _ {k} \{l _ {k} ^ {(i)} \} - l _ {j} ^ {(i)} - m \}, +$$ + +where $\lambda \in \mathbb{R}_+$ is a scalar weight balancing contributions of the CE loss and the corresponding penalty. + +# 3. Sample-wise Constrained DNN Optimization + +Although MbLS can significantly improve calibration, the associated constrained problem (2) is not solved accurately. It is approximated by an unconstrained problem with a single uniform penalty, regardless of the data sample or category. However, the samples and classes considered in a classification problem have different intrinsic learning difficulties. Therefore, an improved training scheme would involve considering distinct penalty weights $\lambda$ for each sample and class. This would result in having to chose $N\times K$ penalty weights $\Lambda \in \mathbb{R}_{+}^{N\times K}$ , with the resulting optimization problem being: + +$$ +\begin{array}{l} \min _ {\theta} \sum_ {i = 1 \atop N = K} ^ {N} \mathcal {L} _ {\mathrm {C E}} \left(\boldsymbol {x} ^ {(i)}, y ^ {(i)}\right) + \tag {4} \\ \sum_ {i = 1} ^ {N} \sum_ {j = 1} ^ {K} \Lambda_ {i j} \max \{0, \max _ {k} \{l _ {k} ^ {(i)} \} - l _ {j} ^ {(i)} - m \}. \\ \end{array} +$$ + +From an optimization perspective, supposing that optimal weights $\theta^{\star}$ exists for problem (2), there exists $\Lambda^{\star} \in \mathbb{R}_{+}^{N \times K}$ such that $(\theta^{\star}, \Lambda^{\star})$ is a saddle point of the Lagrangian associated to (2). These $\Lambda^{\star}$ are the Lagrange multipliers of the problem. Therefore, using $\Lambda = \Lambda^{\star}$ would be the best choice to solve (4). + +In practice, using the Lagrange multipliers of problem (2) as the weights for the penalties may not be computationally feasible, and it could even result in degraded performance. Indeed, in the context of machine learning, we optimize a model's weights $\theta$ to solve (2) on a training set $\mathcal{D}_{\mathrm{train}}$ , and expect to generalize on a test set $\mathcal{D}_{\mathrm{test}}$ , which we do not have access to during training. Because of the bias-variance tradeoff, solving (2) optimally would likely result in overfitting, i.e. we may solve problem (2) accurately on the train set, but not generalize properly on the test set, resulting in poor calibration and classification performance overall. This suggests that it could be preferable to evaluate during training the quality of multipliers on a separate validation set $\mathcal{D}_{\mathrm{val}}$ . Additionally, several mechanisms for training DNNs are not compatible with a straightforward minimization. First, the use of batch normalization yields predictions that are not independent between samples in a minibatch. Second, the use of regularization techniques such as dropout may lead to virtually inaccurate predictions on certain training samples, impacting the correct estimation of multipliers. Third, data augmentation, which is standard in DNN training, would result in additional penalty weights for the augmented samples: they can be easier or harder to classify than the original ones. + +In addition to the above obstacles, applying a penalty weight per sample and per class (resulting in $N \times K$ weights) would not scale well for large datasets and dense predictive tasks, such as semantic segmentation, which is typically formulated as a per-pixel classification task. Assuming that im + +ages in the dataset have a size of $H \times W$ , this would result in $N \times H \times W \times K$ penalty weights. This rapidly becomes a limiting factor for moderately sized segmentation datasets. For instance, Pascal VOC 2012 [9] contains 21 classes and 1464 training images, amounting to $2.62 \times 10^{8}$ pixels, or $5.5 \times 10^{9}$ penalty weights, which, stored as float32, would use $\sim 20$ GiB. For Cityscapes [6], containing approximately 3000 training images of size $2048 \times 1024$ in 19 classes, this amounts to $\sim 445$ GiB. + +Following the above observations, we introduce a relaxation of sample-wise penalties, and propose to solve the following problem: + +$$ +\begin{array}{l} \min _ {\theta} \sum_ {i = 1 \atop N, K} ^ {N} \mathcal {L} _ {\mathrm {C E}} \left(\boldsymbol {x} ^ {(i)}, y ^ {(i)}\right) + \tag {5} \\ \sum_ {i = 1} ^ {N} \sum_ {j = 1} ^ {K} \lambda_ {j} \max \{0, \max _ {k} \{l _ {k} ^ {(i)} \} - l _ {j} ^ {(i)} - m \}, \\ \end{array} +$$ + +where $(\lambda_j)_{1\leq j\leq K}\in \mathbb{R}_+^K$ . Since penalties are now class-wise, we need $K$ penalty weights. This has the advantage to scale well to denser classification tasks such as segmentation. However, we still face a challenging optimization problem (5), since we still need to chose $K$ weights, which can be extremely complicated for large-scale datasets with many classes such as ImageNet, which contains 1000 classes. In the next section, we introduce a numerical technique to deal with this challenge. + +# 4. Class Adaptive Network Calibration + +The challenge of the previous formulation stems from correctly choosing the weights $\lambda \in \mathbb{R}_+^K$ , which can be cumbersome when $K$ is large. Therefore, we propose to use an Augmented Lagrangian Multiplier (ALM) method to adaptively learn the weights of the penalties. + +# 4.1. General ALM + +ALM methods combine penalties and primal-dual updates to solve a constrained problem. They have well-established advantages and enjoy widespread popularity in the general context of optimization [1,35,41]. Specifically, we have the following generic constrained optimization problem: + +$$ +\min _ {x} f (x) \quad \text {s . t .} \quad h _ {i} (x) \leq 0, \quad i = 1, \dots , n \tag {6} +$$ + +where $f: \mathbb{R}^d \to \mathbb{R}$ is the objective function and $h_i: \mathbb{R}^d \to \mathbb{R}, i = 1, \dots, n$ are the constraint functions. We tackle it by approximately solving a sequence $j \in \mathbb{N}$ of unconstrained problems: + +$$ +\min _ {x} \quad \mathcal {L} ^ {(j)} (x) = f (x) + \sum_ {i = 1} ^ {n} P \left(h _ {i} (x), \rho_ {i} ^ {(j)}, \lambda_ {i} ^ {(j)}\right) \tag {7} +$$ + +![](images/77ae33c1b29ebad9fab4420309eee8c8a6c20490d3a1172073de4002ba940f03.jpg) +Figure 2. A penalty-Lagrangian function $P$ with varying values of $\rho$ and $\mu$ . Higher values of $\rho$ bring $P$ closer to an ideal penalty. The multiplier $\lambda$ is the derivative of $P$ w.r.t. the constraint at $z = 0$ . + +with $P: \mathbb{R} \times \mathbb{R}_{++}^{n} \times \mathbb{R}_{++}^{n} \to \mathbb{R}$ a penalty-Lagrangian function, and $\pmb{\rho}^{(j)} = (\rho_i)_{1 \leq i \leq n} \in \mathbb{R}_{++}^n$ , $\pmb{\lambda}^{(j)} = (\lambda_i^{(j)})_{1 \leq i \leq n} \in \mathbb{R}_{++}^n$ the penalty parameters and multipliers associated to $P$ at the $j$ -th iteration. This sequence of unconstrained problems is called outer iterations, while the steps in the minimization of $\mathcal{L}^{(j)}$ are called inner iterations. + +The main components of ALM methods are (i) the penalty-Lagrangian function $P$ , (ii) the update of the penalty multipliers $\lambda^{(j)}$ and (iii) the increase of the penalty parameters $\pmb{\rho}^{(j)}$ . First, the penalty function $P$ needs to satisfy a set of axioms [2] (see Appendix A): these axioms constrain the function to be continuously differentiable w.r.t. its first variable and to have a non-negative derivative: $\forall z \in \mathbb{R}, P'(z, \rho, \lambda) = \frac{\partial}{\partial z} P(z, \rho, \lambda) \geq 0$ , with $P'(0, \rho, \lambda) = \lambda$ . Figure 2 gives an example of a penalty, and how $\rho$ and $\lambda$ affect it. The choice of penalty function is critical to the performance of ALM methods, especially for nonconvex problems [2]. Typical functions include PHR [14,39], $P_{2}$ [17] and $P_{3}$ [34] (see section 3.2 of [2]). Second, the penalty multipliers $\lambda^{(j)}$ are updated to the derivative of $P$ w.r.t. the solution obtained during the last inner minimization. Formally, let $x^{(j)}$ be the approximate minimizer of $\mathcal{L}^{(j)}$ , then $\forall i \in \{1, \dots, n\}$ : + +$$ +\lambda_ {i} ^ {(j + 1)} = P ^ {\prime} \left(h _ {i} \left(x ^ {(j)}\right), \rho_ {i} ^ {(j)}, \lambda_ {i} ^ {(j)}\right) \tag {8} +$$ + +This update rule corresponds to a first-order multiplier estimate for the constrained problem. Third, the penalty parameters $\rho^{(j)}$ are increased during the outer iterations if the constraints do not improve (i.e. is closer to being satisfied) compared to the previous outer iteration. Typically, $\rho_{i}^{(j + 1)} = \gamma \rho_{i}^{(j)}$ if $h_i$ does not improve, with $\gamma > 1$ . + +When the problem is convex, alternating between the approximate minimization of (7) and the update of the multipliers (8) leads to a solution for the constrained problem. The inner minimization corresponds to minimizing the primal while the outer iterations correspond to solving the dual problem. The complete procedure is presented in Algorithm 1. + +Although guarantees exist only in the convex case, it is well-known that ALM methods can efficiently solve nonconvex problems as well [2]. In the context of deep learning, their use has been surprisingly underexplored [40, 41]. + +# Algorithm 1 Augmented Lagrangian Multiplier algorithm + +Require: Objective function $f$ + +Require: Constraint functions $h_i, i = 1, \dots, n$ + +Require: Penalty function $P$ , initial $\pmb{\lambda}^{(0)} \in \mathbb{R}_{++}^{n}$ , $\pmb{\rho}^{(0)} \in \mathbb{R}_{++}^{n}$ + +Require: Initial variable $x^{(0)}$ , iterations $j = 1$ + +$$ +\mathcal {L} ^ {(j)} (x) = f (x) + \sum_ {i = 1} ^ {n} P \left(h _ {i} (x), \rho_ {i} ^ {(j)}, \lambda_ {i} ^ {(j)}\right) +$$ + +1: while not converged do +2: Initialize with $x^{(j - 1)}$ and minimize (approximately): +3: $x^{(j)}\gets$ (approximate) minimizer of $\mathcal{L}^{(j)}$ +4: for $i = 1,\dots ,n$ do +5: $\lambda_{i}^{(j + 1)}\gets P^{\prime}(h_{i}(x^{(j)}),\rho_{i}^{(j)},\lambda_{i}^{(j)})$ +6: if the $i$ -th constraint does not improve then +7: $\rho_{i}^{(j + 1)}\gets \gamma \rho_{i}^{(j)}$ +8: else +9: $\rho_{i}^{(j + 1)}\gets \rho_{i}^{(j)}$ +10: end if +11: end for +12: $j\gets j + 1$ +13: end while + +# 4.2. ALM for calibration + +Our goal now is to build an ALM method effective for calibration purposes. We can achieve this by reformulating problem (5) using a penalty function $P$ parameterized by $(\pmb {\rho},\pmb {\lambda})\in \mathbb{R}_{++}^{K}\times \mathbb{R}_{++}^{K}$ as follows: + +$$ +\min _ {\theta} \quad \sum_ {i = 1} ^ {N} \mathcal {L} _ {\mathrm {C E}} \left(\boldsymbol {x} ^ {(i)}, y ^ {(i)}\right) + \sum_ {k = 1} ^ {K} P \left(d _ {k} ^ {(i)} - m, \rho_ {k}, \lambda_ {k}\right) \tag {9} +$$ + +where $d_k^{(i)} = \max \{l^{(i)}\} - l_k^{(i)} \in \mathbb{R}_+$ . With this formulation, it is natural to use a penalty-Lagrangian function for $P$ . To avoid numerical issues typically associated with non-linear penalties, we normalize the constraints by the margin $m > 0$ : + +$$ +d _ {k} ^ {(i)} - m \leq 0 \Leftrightarrow \frac {d _ {k} ^ {(i)}}{m} - 1 \leq 0 \tag {10} +$$ + +This leads to improved numerical stability for the ALM multiplier update as well. Additionally, we average the constraints instead of summing them. This makes the method independent of the number of classes, and eases the choice of initial penalty parameters $\rho^{(0)}$ . The resulting loss is: + +$$ +\sum_ {i = 1} ^ {N} \mathcal {L} _ {\mathrm {C E}} \left(\boldsymbol {x} ^ {(i)}, y ^ {(i)}\right) + \frac {1}{K} \sum_ {k = 1} ^ {K} P \left(\frac {d _ {k} ^ {(i)}}{m} - 1, \rho_ {k}, \lambda_ {k}\right) \tag {11} +$$ + +As noted in Section 3, one of the main downsides of estimating Lagrange multipliers from the training set is that + +we could quickly overfit the data. Therefore, we propose to use the validation set to obtain a reliable estimate of the penalty multipliers at each epoch. We consider that an epoch of training corresponds to the approximate minimization of the loss function, and then compute the average penalty multiplier on the validation set. Formally after a training epoch $j$ , the penalty multipliers for epoch $j + 1$ will be, for all $k = 1, \ldots, K$ : + +$$ +\lambda_ {k} ^ {(j + 1)} = \frac {1}{| \mathcal {D} _ {\mathrm {v a l}} |} \sum_ {(\boldsymbol {x}, y) \in \mathcal {D} _ {\mathrm {v a l}}} P ^ {\prime} \left(\frac {d _ {k}}{m} - 1, \rho_ {k} ^ {(j)}, \lambda_ {k} ^ {(j)}\right) \tag {12} +$$ + +Finally, the penalty multiplier is projected on a safeguarding interval $[\lambda_{\min}, \lambda_{\max}] = [10^{-6}, 10^{6}]$ in our case. To update the penalty parameters $\rho$ , we compute the average constraint per class on the validation set. Then, for each class, if the average constraint is positive and has not decreased compared to the previous epoch, we multiply the corresponding penalty parameter by $\gamma$ . + +Finally, as suggested by [2] and confirmed by our empirical results, we utilize the PHR function in our implementation, defined as follows: + +$$ +\operatorname {P H R} (z, \rho , \lambda) = \left\{ \begin{array}{l l} \lambda z + \frac {1}{2} \rho z ^ {2} & \text {i f} \quad \lambda + \rho z \geq 0; \\ - \frac {\lambda^ {2}}{2 \rho} & \text {o t h e r w i s e .} \end{array} \right. \tag {13} +$$ + +Overall, the proposed method, consolidated in Algorithm 2, corresponds to approximately solving the constrained problem (2), by learning class-wise penalty multipliers on the validation set, to avoid overfitting and training specificities (i.e. batch normalization, dropout, augmentations) which would result in unreliable penalty multipliers estimates. + +# 5. Experiments + +# 5.1. Experimental Setup + +Datasets. We perform experiments on a variety of popular benchmarks. First, we include three widely used image classification benchmarks, including Tiny-ImageNet [7], ImageNet [7] and one long-tailed image classification, ImageNet-LT [25]. Tiny-ImageNet is widely used in the calibration literature [22, 31], with relatively small $64 \times 64$ resolution, while ImageNet [7] is a large-scale benchmark consisting of 1000 categories and over 1M images. The main characteristic of ImageNet-LT is that the number of samples is extremely imbalanced across classes, ranging from 5 to 1280. To evaluate performance in dense prediction tasks, we include one semantic segmentation benchmark, PASCAL VOC2012 [9]. Furthermore, one benchmark from the NLP domain, 20 Newsgroups [19], is included to show the general applicability. For a detailed description of each dataset and the pre-processing settings, please refer to Appendix B. + +Evaluation Metrics. For calibration, we report the most widely used Expected Calibration Error (ECE) [33]. Samples + +Algorithm 2 CALS-ALM training +Require: DNN initial $\theta^0$ , margin m +Require: Dataset: $\mathcal{D}_{\mathrm{train}}$ $\mathcal{D}_{\mathrm{val}}$ , batch size $B$ +Require: Penalty function $P,\gamma >1,\tau \in (0,1)$ +Require: Initial $\lambda^{(0)}\in \mathbb{R}_{++}^{n},\pmb{\rho}^{(0)}\in \mathbb{R}_{++}^{n}$ 1: for $j = 0,\dots ,T$ do // Epochs of training +2: for each mini-batch $\{(x^{(i)},y^{(i)})\}_{i = 1}^{B}$ in $\mathcal{D}_{\mathrm{train}}$ do +3: $\mathcal{L}_{\mathrm{c}} = \sum_{i = 1}^{B}\mathcal{L}_{\mathrm{CE}}(\boldsymbol{x}^{(i)},y^{(i)})$ // Cross-entropy +4: $\mathcal{L}_{\mathrm{p}} = \sum_{i = 1}^{B}\frac{1}{K}\sum_{k = 1}^{K}P\Big(\frac{d_k^{(i)}}{m} -1,\rho_k^{(j)},\lambda_k^{(j)}\Big)$ // Penalties +5: $\mathcal{L} = \frac{1}{B} (\mathcal{L}_{\mathrm{c}} + \mathcal{L}_{\mathrm{p}})$ +6: $\theta^{(t + 1)}\gets \theta^{(t)} - \alpha \cdot \nabla_{\theta}\mathcal{L}$ // Gradient descent +7: end for +8: for $k = 1,\ldots ,K$ do // Adjust $\lambda$ and $\rho$ on validation +9: $\lambda_{k}^{(j + 1)} = \frac{1}{|\mathcal{D}_{\mathrm{val}}|}\sum_{(\boldsymbol {x},\boldsymbol {y})\in \mathcal{D}_{\mathrm{val}}}P'\left(\frac{d_k}{m} -1,\rho_k^{(j)},\lambda_k^{(j)}\right)$ +10: $\overline{d_k}^{(j)} = \frac{1}{|\mathcal{D}_{\mathrm{val}}|}\sum_{(\boldsymbol {x},\boldsymbol {y})\in \mathcal{D}_{\mathrm{val}}}d_k - 1$ // Average constraint +11: if $j\geq 1$ and $\overline{d_k}^{(j)} > \tau \max \{0,\overline{d_k}^{(j - 1)}\}$ then +12: $\rho_{k}^{(j + 1)}\gets \gamma \rho_{k}^{(j)}$ // Constraint has not improved +13: else +14: $\rho_{k}^{(j + 1)}\gets \rho_{k}^{(j)}$ +15: end if +16: end for +17: end for + +are grouped into $M$ equi-spaced bins according to prediction confidence, and a weighted average of the absolute difference between accuracy and confidence in each bin is calculated: + +$$ +E C E = \sum_ {m = 1} ^ {M} \frac {\left| B _ {m} \right|}{N} \left| A _ {m} - C _ {m} \right|, \tag {14} +$$ + +where $M$ is the number of bins, $N$ the amount of test samples, $B_{m}$ the samples with prediction confidence in the $m^{th}$ bin, $A_{m}$ the accuracy and $C_{m}$ the mean confidence of samples in the $m^{th}$ bin. Note we fix $N$ to 15 according to [22, 31]. In accordance with [22], we also report Adaptive ECE (AECE), a variant of ECE where the bins are configured to evenly distribute the test samples across them. Additionally, Classwise Calibration Error (CWCE) [29], a classwise extension of ECE, is included in Appendix C. For discriminative performance, we use standard measures: accuracy (Acc) for classification, and intersection over union (mIoU) for segmentation. + +Compared methods. We compare our method to other learning based calibration losses, including (i) methods that impose constraints on predictions (either logits or softmax probabilities), i.e. Explicit Confidence Penalty (ECP) [38], Label Smoothing (LS) [42], Focal Loss (FL) [21] and its sample-dependent version (FLSD) [31], Margin-based Label + +
MethodTinyImageNetImageNetImageNet-LT
ResNet-50ResNet-50SwinV2-TResNet-50SwinV2-T
AccECEAECEAccECEAECEAccECEAECEAccECEAECEAccECEAECE
CE65.023.733.6975.169.199.1875.609.959.9437.9028.1228.1231.8231.8236.68
MMCE [18]65.342.812.6174.858.578.5676.689.079.0837.7928.4128.4033.1426.4126.41
ECP [38]64.904.003.9275.228.278.2675.829.889.8637.6928.1428.1331.2233.7033.70
LS [42]65.783.173.1676.042.572.8875.427.327.3337.8810.4610.3831.7011.4211.40
FL [31]63.092.963.1273.871.601.6575.603.193.1836.0418.3718.3630.7325.5025.50
FLSD [31]64.092.912.9573.972.082.0674.702.442.3736.1817.7717.7832.5625.1625.17
CPC [5]64.494.884.9176.333.663.5976.345.505.3338.9016.0015.9932.5413.2113.19
MbLS [22]64.741.641.7375.824.444.2677.181.951.7338.326.166.1632.057.657.64
CALS-HR65.092.502.4276.345.635.6977.583.062.9538.502.832.7834.312.372.45
CALS-ALM65.031.541.3876.441.461.3277.101.611.6938.562.152.3033.942.322.45
+ +Table 1. Calibration performance for different approaches on three image classification benchmarks. We report two lower-is-better calibration metrics, i.e. ECE and AECE. Best method is highlighted in bold, while the second-best one is underlined. + +Smoothing (MbLS) [22] and CPC [5], and (ii) techniques that directly optimize calibration measures, i.e. MMCE [18]. We refer to the related literature [22, 31] to set the hyperparameters for various methods. For instance, the smoothing factor in LS and FL is set to 0.05 and 3 respectively, and we set margin to 10 in MbLS. A detailed description of hyper-parameter values can be found in Appendix F. + +Our methods. A simple alternative to the algorithm presented in Section 4.2 would be to heuristically tune multipliers by scaling them according to penalty values: if $P_{k}^{(j + 1)}$ increases we also increase $\lambda_{k}^{j + 1}$ and vice versa. This strategy, akin to learning rate scheduling, can be formulated as: + +$$ +\lambda_ {k} ^ {(j + 1)} = \left\{ \begin{array}{l l} \mu \lambda_ {k} ^ {(j)} & \text {i f} P _ {k} ^ {(j + 1)} > \tau P _ {k} ^ {(j)} \\ \lambda_ {k} ^ {(j)} / \mu & \text {i f} P _ {k} ^ {(j)} > \tau P _ {k} ^ {(j + 1)} \\ \lambda_ {k} ^ {(j)} & \text {o t h e r w i s e} \end{array} \right. \tag {15} +$$ + +where $\mu > 1$ and $\tau > 1$ are hyper-parameters that we fix to 1.1. We refer to our main algorithm as CALS-ALM and to this heuristic rule as CALS-HR in what follows. + +We fix the margin to $m = 10$ on vision tasks and $m = 6$ on the NLP benchmark, as in [22], for a fair comparison. We also perform an ablation study to investigate the impact of the margin value. For other hyper-parameters, we set $\pmb{\lambda}^{(0)} = 10^{-6} \cdot \mathbf{1}_K$ , $\pmb{\rho}^{(0)} = \mathbf{1}_K$ , $\gamma = 1.2$ , and we update the penalty parameters $\pmb{\rho}$ every 10 epochs. Please refer to Appendix F for a detailed description of all hyper-parameters. + +Implementation Details. For image classification, we experiment with ResNet [13] and a vision Transformer model, i.e. Swin Transformer V2 (SwinV2-T) [23]. DeepLabV3 [3] is employed for semantic segmentation on PASCAL VOC2012. Following [22, 31], we use the Global Pooling CNN (GPoolCNN) architecture [20] on the NLP recognition task. Further training details on each dataset can be found in Appendix B. + +# 5.2. Results + +Results on image classification. Table 1 presents the discriminative and calibration performance of our methods + +across three widely used classification benchmarks, compared to baselines and related works. We can observe that our CALS-ALM approach consistently outperforms existing techniques in terms of calibration. Specifically, the results indicate that the standard CE loss and other approaches often lead to miscalibrated models, with the severity of miscalibration substantially increasing in correlation with dataset difficulty. This is particularly evident in large-scaled datasets with numerous classes, such as ImageNet, or those with long-tail class distributions, like ImageNet-LT. Although other learning-based methods could provide better calibrated networks, their performance is not stable across different settings. For example, while FL achieves a relatively low ECE of 1.60 with a Resnet50 trained on ImageNet, it only yields an ECE of 25.50 when training a SwinV2-T network on ImageNet-LT, revealing a limitation in adapting to different learning scenarios. In contrast, CALS-ALM attains the best calibration performance in all cases, frequently outperforming existing approaches by a significant margin, with minimal variations across datasets and architectures. This trends persists when compared to the most closely related technique, MbLS. It is noteworthy that the improvements on the long-tailed ImageNet-LT dataset are substantial; for example, we decrese ECE from 28.12 to 2.15 for ResNet50, and from 31.82 to 2.32 for SwinV2-T, validating the effectiveness of class adaptive learning. Another interesting finding is that CALS-HR, employing a naive update strategy for class-wise penalty weights, achieves nearly the second-best performance in the ImageNet-LT dataset. This further demonstrates the effectiveness of CALS for learning under class-imbalance scenarios. For the reliability diagrams of various models, please refer to Appendix E. + +In terms of model accuracy, our method delivers competitive performances, surpassing existing methods in certain cases. It is important to emphasize that, while the proposed method achieves discriminative results comparable to the best-performing approach for each dataset, the differences in calibration are considerable. This highlights the superiority of the proposed formulation for training highly discrimi + +![](images/62e78eca5eb382b7850f62eb708db706f17b5aa1ad3f60232d49924890d793ba.jpg) +(a) Validation ECE and multipliers during training. + +![](images/c862e61f3d4cd7c3e9ee5bc0b8e22c20dcae1e6a6693b1aad89dba89c032ca97.jpg) +Figure 3. Ablation study on ImageNet-LT. (a) Evolution of ECE on validation and multipliers for CALS: ECE on the validation set for our method (CALS), CE and MbLS [22] and values of multipliers $\lambda$ for CALS after each training epoch. (b) Effect of penalty functions and margin: ECE and accuracy on validation and test set are shown across different choices of penalty functions and margin values. + +![](images/213ff12f720a013ab09eb9be14846ad88d70b6a413e2bbbbcd3a08b1a4fbca46.jpg) +(b) Effect of penalty function and margin on accuracy and ECE for CALS. + +native and well-calibrated networks. Figure 1 provides a more visual comparison considering both accuracy and ECE. It is demonstrated that CALS-ALM provides the optimal compromise between accuracy and calibration performance. + +Ablation Analysis. Figure 3 illustrates the evolution of ECE in (a) and penalty multipliers $\lambda$ in (b) during training. It is interesting to observe that the evolution of $\lambda$ is consistent with the ECE. Specifically, the average penalty weight gradually increases while the ECE initially deteriorates because the model is focused on increasing accuracy. However, the value of the penalty weight begins to decline alongside the ECE, as the network starts to become better calibrated. For a visualization of classwise multipliers, please refer to Appendix D. Figure 3(c) highlights the impact of the choice of penalty functions and margin values. This demonstrates that the PHR penalty function is preferable over the other two options, P2 and P3, for both calibration and accuracy. Regarding the margin, the best performance is achieved with $m \approx 10$ , which is consistent with the findings in [22]. + +Semantic Segmentation. Table 2 presents the performance on Pascal VOC dataset. Note here CALS refers to our best method, i.e. CALS-ALM. It can be observed that the trend is consistent with image classification experiments: CALS outperforms counterparts in terms of ECE, and yields competitive results on discriminative performance, i.e. mIoU in segmentation. It is worth noting that some methods like MMCE, FLSD and CPC, are not included here because their computation demands were too heavy for pixel-wise segmentation tasks. In contrast, our method is unlimited in dense prediction tasks as the computation cost it adds is moderate. + +Text Classification. Last, we demonstrate the general applicability of the proposed method by analyzing its performance on a non-vision task, i.e. text classification on 20 Newsgroups dataset. The results are reported in Table 3. Remarkably, CALs again brings substantial improvement in + +
CEECPLSFLMbLSCALS
ECE14.755.156.9010.875.224.66
mIoU66.4665.5767.7364.2565.2966.77
+ +Table 2. Segmentation results on the PASCAL VOC 2012. + +
CEMMCEECPLSFLFLSDCPCMbLSCALS
ECE (%)22.7523.0222.978.0710.8010.879.465.402.04
Acc (%)67.0166.2366.4867.1466.0865.8568.2767.8968.32
+ +Table 3. Results on the text classification task, 20 Newsgroups. + +terms of calibration, with ECE decreasing to $2.04\%$ , while yielding the best accuracy $68.32\%$ . This reveals that the proposed class adaptive learning method is also able to handle class differences in NLP applications and provide promising performance in terms of both accuracy and calibration. + +# 6. Limitations and Future Work + +We have proposed Class Adaptive Label Smoothing for network calibration based on a modified Augmented Lagrangian Multiplier algorithm. Despite its superior performance over previous methods, there are potential limitations in this work. For instance, our method requires the validation set to have the same distribution as the training set. Although this is satisfied in nearly every benchmark, it will be interesting to investigate the impact of using non-independent and identically distributed (i.i.d) validation sets. + +# Acknowledgements + +This work is supported by National Science and Engineering Research Council of Canada (NSERC) and Prompt Quebec. We also thank Calcul Quebec and Compute Canada. Adrian Galdran is supported by a Marie Sklodowska-Curie Fellowship (No 892297). + +# References + +[1] Dimitri P. Bertsekas. Constrained Optimization and Lagrange Multiplier Methods (Optimization and Neural Computation Series). Athena Scientific, 1 edition, 1996. 4 +[2] Ernesto G Birgin, Romulo A Castillo, and José Mario Martínez. Numerical comparison of augmented lagrangian algorithms for nonconvex problems. Computational Optimization and Applications, 31(1), 2005. 5, 6, 11 +[3] Liang-Chieh Chen, George Papandreou, Florian Schroff, and Hartwig Adam. Rethinking atrous convolution for semantic image segmentation. In CVPR, 2017. 1, 7, 11 +[4] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In ICML, 2020. 1 +[5] Jiacheng Cheng and Nuno Vasconcelos. Calibrating deep neural networks by pairwise constraints. In CVPR, 2022. 3, 7, 13 +[6] Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The cityscapes dataset for semantic urban scene understanding. In CVPR, June 2016. 4 +[7] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In CVPR, 2009. 6, 11 +[8] Zhipeng Ding, Xu Han, Peirong Liu, and Marc Niethammer. Local temperature scaling for probability calibration. In ICCV, 2021. 1, 3 +[9] Mark Everingham, S. M. Eslami, Luc Gool, Christopher K. Williams, John Winn, and Andrew Zisserman. The pascal visual object classes challenge: A retrospective. *IJCV*, 111(1):98–136, 2015. 4, 6, 11 +[10] Yarin Gal and Zoubin Ghahramani. Dropout as a bayesian approximation: Representing model uncertainty in deep learning. In ICML, 2016. 2 +[11] Chuan Guo, Geoff Pleiss, Yu Sun, and Kilian Q Weinberger. On calibration of modern neural networks. In ICML, 2017. 1, 2, 3, 12 +[12] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollar, and Ross Girshick. Masked autoencoders are scalable vision learners. In CVPR, 2022. 1 +[13] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, 2016. 1, 7, 11 +[14] Magnus R. Hestenes. Multiplier and gradient methods. Journal of Optimization Theory and Applications, 4:303-320, 1969. 5 +[15] Sergey Ioffe and Christian Szegedy. Batch normalization: Accelerating deep network training by reducing internal covariate shift. In ICML, 2015. 1, 2 +[16] Archit Karandikar, Nicholas Cain, Dustin Tran, Balaji Lakshminarayanan, Jonathon Shlens, Michael C Mozer, and Rebecca Roelofs. Soft calibration objectives for neural networks. In NeurIPS, 2021. 3 +[17] Barry W. Kort and Dimitri P. Bertsekas. Combined primal-dual and penalty methods for convex programming. SIAM Journal on Control and Optimization, 14(2):268-294, 1976. 5 + +[18] Aviral Kumar, Sunita Sarawagi, and Ujjwal Jain. Trainable calibration measures for neural networks from kernel mean embeddings. In ICML, 2018. 3, 7, 13 +[19] Ken Lang. Newsweeder: Learning to filter netnews. In ICML, 1995. 6, 12 +[20] Min Lin, Qiang Chen, and Shuicheng Yan. Network in network. In ICLR, 2014. 7, 12 +[21] Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dólar. Focal loss for dense object detection. In CVPR, 2017. 1, 3, 6 +[22] Bingyuan Liu, Ismail Ben Ayed, Adrian Galdran, and Jose Dolz. The devil is in the margin: Margin-based label smoothing for network calibration. In CVPR, 2022. 1, 2, 3, 6, 7, 8, 11, 12, 13 +[23] Ze Liu, Han Hu, Yutong Lin, Zhuliang Yao, Zhenda Xie, Yixuan Wei, Jia Ning, Yue Cao, Zheng Zhang, Li Dong, Furu Wei, and Baining Guo. Swin transformer v2: Scaling up capacity and resolution. In CVPR, 2022. 7, 11 +[24] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In ICCV, 2021. 1 +[25] Ziwei Liu, Zhongqi Miao, Xiaohang Zhan, Jiayun Wang, Boqing Gong, and Stella X. Yu. Large-scale long-tailed recognition in an open world. In CVPR, 2019. 6, 11 +[26] Ilya Loshchilov and Frank Hutter. SGDR: Stochastic gradient descent with warm restarts. In *ICLR*, 2017. 11 +[27] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In *ICLR*, 2019. 1, 11 +[28] Xingchen Ma and Matthew B. Blaschko. Meta-cal: Well-controlled post-hoc calibration by ranking. In ICML, 2021. 3 +[29] Lena Maier-Hein et al. Metrics reloaded: Pitfalls and recommendations for image analysis validation. arXiv preprint arXiv:2206.01653, 2022. 6, 12 +[30] Matthias Minderer, Josip Djolonga, Rob Romijnders, Frances Hubis, Xiaohua Zhai, Neil Houlsby, Dustin Tran, and Mario Lucic. Revisiting the calibration of modern neural networks. In NeurIPS, 2021. 3, 11, 12 +[31] Jishnu Mukhoti, Viveka Kulharia, Amartya Sanyal, Stuart Golodetz, Philip HS Torr, and Puneet K Dokania. Calibrating deep neural networks using focal loss. In NeurIPS, 2020. 1, 2, 3, 6, 7, 11, 12, 13 +[32] Rafael Müller, Simon Kornblith, and Geoffrey Hinton. When does label smoothing help? In NeurIPS, 2019. 1, 3 +[33] Mahdi Pakdaman Naeini, Gregory F. Cooper, and Milos Hauskrecht. Obtaining well calibrated probabilities using bayesian binning. In AAAI, 2015. 6 +[34] H. Nakayama, H. Samaya, and H. Samaya. A generalized lagrangian function and multiplier method. Journal of Optimization Theory and Applications, 17:211-227, 1975. 5 +[35] Jorge Nocedal and Stephen J. Wright. Numerical Optimization. Springer, New York, NY, USA, 2 edition, 2006. 4 +[36] Yaniv Ovadia, Emily Fertig, Jie Ren, Zachary Nado, David Sculley, Sebastian Nowozin, Joshua V Dillon, Balaji Lakshminarayanan, and Jasper Snoek. Can you trust your model's uncertainty? evaluating predictive uncertainty under dataset shift. In NeurIPS, 2019. 3 + +[37] Jeffrey Pennington, Richard Socher, and Christopher D Manning. Glove: Global vectors for word representation. In EMNLP, 2014. 12 +[38] Gabriel Pereyra, George Tucker, Jan Chorowski, Lukasz Kaiser, and Geoffrey Hinton. Regularizing neural networks by penalizing confident output distributions. In ICLR, 2017. 1, 3, 6, 7, 13 +[39] M. J. D. Powell. A method for nonlinear constraints in minimization problems. In *Fletcher, R., Ed., Optimization, Academic Press, New York, NY*, 1969. 5 +[40] Jérôme Rony, Eric Granger, Marco Pedersoli, and Ismail Ben Ayed. Augmented lagrangian adversarial attacks. In ICCV, 2021. 5 +[41] Sara Sangalli, Ertunc Erdil, Andeas Hotker, Olivio F Donati, and Ender Konukoglu. Constrained optimization to train neural networks on critical and under-represented classes. In NeurIPS, 2021. 4, 5 +[42] Christian Szegedy, Vincent Vanhoucke, Sergey Ioffe, Jon Shlens, and Zbigniew Wojna. Rethinking the inception architecture for computer vision. In CVPR, 2016. 3, 6, 7, 13 +[43] Christian Tomani, Sebastian Gruber, Muhammed Ebrar Erdem, Daniel Cremers, and Florian Buettner. Post-hoc uncertainty calibration for domain drift scenarios. In CVPR, 2021. 3 +[44] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In NeurIPS, 2017. 1 \ No newline at end of file diff --git a/2023/Class Adaptive Network Calibration/images.zip b/2023/Class Adaptive Network Calibration/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..087fb5d1ab348ac22c3587bf6d7efa85a4c25393 --- /dev/null +++ b/2023/Class Adaptive Network Calibration/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:34d540121c70b66141edc50f87c81cb17798ac30598791241b182cfb8b1dfb89 +size 352838 diff --git a/2023/Class Adaptive Network Calibration/layout.json b/2023/Class Adaptive Network Calibration/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..bac972ec9fbfad7964fd4828d30fac7cf073ebf8 --- /dev/null +++ b/2023/Class Adaptive Network Calibration/layout.json @@ -0,0 +1,9936 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 184, + 103, + 410, + 120 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 184, + 103, + 410, + 120 + ], + "spans": [ + { + "bbox": [ + 184, + 103, + 410, + 120 + ], + "type": "text", + "content": "Class Adaptive Network Calibration" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 102, + 142, + 489, + 158 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 102, + 142, + 489, + 158 + ], + "spans": [ + { + "bbox": [ + 102, + 142, + 489, + 158 + ], + "type": "text", + "content": "Bingyuan Liu\\*, Jérôme Rony\\*, Adrian Galdran\\*, Jose Dolz\\*, Ismail Ben Ayed" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 118, + 166, + 471, + 182 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 118, + 166, + 471, + 182 + ], + "spans": [ + { + "bbox": [ + 118, + 166, + 471, + 182 + ], + "type": "text", + "content": "1ÉTS Montreal, Canada 2Universitat Pompeu Fabra, Barcelona, Spain" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 143, + 209, + 192, + 220 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 209, + 192, + 220 + ], + "spans": [ + { + "bbox": [ + 143, + 209, + 192, + 220 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 234, + 289, + 558 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 234, + 289, + 558 + ], + "spans": [ + { + "bbox": [ + 46, + 234, + 289, + 558 + ], + "type": "text", + "content": "Recent studies have revealed that, beyond conventional accuracy, calibration should also be considered for training modern deep neural networks. To address miscalibration during learning, some methods have explored different penalty functions as part of the learning objective, alongside a standard classification loss, with a hyper-parameter controlling the relative contribution of each term. Nevertheless, these methods share two major drawbacks: 1) the scalar balancing weight is the same for all classes, hindering the ability to address different intrinsic difficulties or imbalance among classes; and 2) the balancing weight is usually fixed without an adaptive strategy, which may prevent from reaching the best compromise between accuracy and calibration, and requires hyper-parameter search for each application. We propose Class Adaptive Label Smoothing (CALS) for calibrating deep networks, which allows to learn class-wise multipliers during training, yielding a powerful alternative to common label smoothing penalties. Our method builds on a general Augmented Lagrangian approach, a well-established technique in constrained optimization, but we introduce several modifications to tailor it for large-scale, class-adaptive training. Comprehensive evaluation and multiple comparisons on a variety of benchmarks, including standard and long-tailed image classification, semantic segmentation, and text classification, demonstrate the superiority of the proposed method. The code is available at https://github.com/by-liu/CALS." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 581, + 127, + 593 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 581, + 127, + 593 + ], + "spans": [ + { + "bbox": [ + 47, + 581, + 127, + 593 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 601, + 288, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 601, + 288, + 685 + ], + "spans": [ + { + "bbox": [ + 46, + 601, + 288, + 685 + ], + "type": "text", + "content": "Deep Neural Networks (DNNs) have become the prevailing model in machine learning, particularly for computer vision [13] and natural language processing applications [44]. Increasingly powerful architectures [3, 13, 24], learning methods [4, 12] and a large body of other techniques [15, 27] are constantly introduced. Nonetheless, recent studies [11, 31] have shown that regardless of their superior discriminative" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 209, + 547, + 317 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 209, + 547, + 317 + ], + "spans": [ + { + "bbox": [ + 306, + 209, + 547, + 317 + ], + "type": "text", + "content": "performance, high-capacity modern DNNs are poorly calibrated, i.e. failing to produce reliable predictive confidences. Specifically, they tend to yield over-confident predictions, where the probability associated with the predicted class overestimates the actual likelihood. Since this is a critical issue in safety-sensitive applications like autonomous driving or computational medical diagnosis, the problem of DNN calibration has been attracting increasing attention in recent years [11, 31, 38]." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 318, + 547, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 318, + 547, + 449 + ], + "spans": [ + { + "bbox": [ + 304, + 318, + 547, + 449 + ], + "type": "text", + "content": "Current calibration methods can be categorized into two main families. The first family involves techniques that perform an additional post-processing parameterized operation on the output logits (or pre-softmax activations) [11], with the calibration parameters of that operation obtained from a validation set by either learning or grid-search. Despite the simplicity and low computational cost, these methods have empirically proven to be highly effective [8, 11]. However, their main drawback is that the choice of the optimal calibration parameters is highly sensitive to the trained model instance and validation set [22, 31]." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 449, + 548, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 449, + 548, + 700 + ], + "spans": [ + { + "bbox": [ + 304, + 449, + 548, + 700 + ], + "type": "text", + "content": "The second family of methods attempts to simultaneously optimize for accuracy and calibration during network training. This is achieved by introducing, explicitly or implicitly, a secondary optimization goal involving the model's predictive uncertainty, alongside the main training objective. As a result, a scalar balancing hyper-parameter is required to tune the relative contribution of each term in the overall loss function. Some examples of this type of approaches include: Explicit Confidence Penalty (ECP) [38], Label Smoothing (LS) [32], Focal Loss (FL) [21] and its variant, Sample-Dependent Focal Loss (FLSD) [31]. It has been recently demonstrated in [22] that all these methods can be formulated as different penalty terms that enforce the same equality constraint on the logits of the DNN: driving the logit distances towards zero. Here, logit distances refers to the vector of L1 distances between the highest logit value and the rest. Observing the non-informative nature of this equality constraint, [22] proposed to use a generalized inequality constraint, only penalizing those logits for which the distance is larger than a pre-defined margin, achieving state-of-the-art calibration performance on many different benchmarks." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 318, + 701, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 701, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 318, + 701, + 547, + 714 + ], + "type": "text", + "content": "Although learning based methods achieve greater calibra" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 693, + 286, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 693, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 693, + 286, + 713 + ], + "type": "text", + "content": "*Equal Contributions. Correspondence to: {liubingyuan1988@gmail.com, jerome.rony.1@etsmtl.net}" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "16070" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 68, + 299, + 229 + ], + "blocks": [ + { + "bbox": [ + 58, + 68, + 299, + 229 + ], + "lines": [ + { + "bbox": [ + 58, + 68, + 299, + 229 + ], + "spans": [ + { + "bbox": [ + 58, + 68, + 299, + 229 + ], + "type": "image", + "image_path": "efbbb40bcc5c3e8826b7a9fcd227fec317cc07d4a7186c95cc9452b851551f17.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 237, + 548, + 304 + ], + "lines": [ + { + "bbox": [ + 46, + 237, + 548, + 304 + ], + "spans": [ + { + "bbox": [ + 46, + 237, + 548, + 304 + ], + "type": "text", + "content": "Figure 1. Many techniques have been proposed for jointly improving accuracy and calibration during training [11,31], but they fail to consider uneven learning scenarios like high class imbalance or long-tail distributions. We show a comparison of the proposed CALS-ALM method and different learning approaches in terms of Calibration Error (ECE) vs Accuracy on the (a) ImageNet and (b) ImageNet-LT (long-tailed ImageNet) datasets. A lower ECE indicates better calibration: a better model should attain high ACC and low ECE. Among all the considered methods, CALS-ALM shows superior performance when considering both discriminative power and well-balanced probabilistic predictions, achieving best accuracy and calibration on ImageNet, and best calibration and second best accuracy on ImageNet-LT." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 304, + 68, + 535, + 229 + ], + "blocks": [ + { + "bbox": [ + 304, + 68, + 535, + 229 + ], + "lines": [ + { + "bbox": [ + 304, + 68, + 535, + 229 + ], + "spans": [ + { + "bbox": [ + 304, + 68, + 535, + 229 + ], + "type": "image", + "image_path": "f3a4a14c56f26e6becfa638ee582811403c0d1e8809aa9037ba042ebd4abcdfd.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 325, + 289, + 492 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 325, + 289, + 492 + ], + "spans": [ + { + "bbox": [ + 46, + 325, + 289, + 492 + ], + "type": "text", + "content": "tion performance [22, 31], they have two major limitations: 1) The scalar balancing weight is equal for all classes. This hinders the network performance when some classes are harder to learn or less represented than others, such as in datasets with a large number of categories (ImageNet) or considerable class imbalance (ImageNet-LT). 2) The balancing weight is usually fixed before network optimization, with no learning or adaptive strategy throughout training. This can prevent the model from reaching the best compromise between accuracy and calibration. To address the above issues, we introduce Class Adaptive Label Smoothing method based on an Augmented Lagrangian Multiplier algorithm, which we refer to as CALS-ALM. Our Contributions can be summarized as follows:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 501, + 289, + 714 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 47, + 501, + 287, + 597 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 501, + 287, + 597 + ], + "spans": [ + { + "bbox": [ + 47, + 501, + 287, + 597 + ], + "type": "text", + "content": "- We propose Class Adaptive Label Smoothing (CALS) for network calibration. Adaptive class-wise multipliers are introduced instead of the widely used single balancing weight, which addresses the above two issues: 1) CALS can handle a high number of classes with different intrinsic difficulties, e.g. ImageNet; 2) CALS can effectively learn from data suffering from class imbalance or a long-tailed distribution, e.g. ImageNet-LT." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 605, + 289, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 605, + 289, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 605, + 289, + 714 + ], + "type": "text", + "content": "- Different from previous penalty based methods, we solve the resulting constrained optimization problem by implementing a modified Augmented Lagrangian Multiplier (ALM) algorithm, which yields adaptive and optimal weights for the constraints. We make some critical design decisions in order to adapt ALM to the nature of modern learning techniques: 1) The inner convergence criterion in ALM is relaxed to a fixed number of iterations in each inner stage, which is amenable to mini-batch stochas" + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 313, + 325, + 547, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 325, + 547, + 445 + ], + "spans": [ + { + "bbox": [ + 313, + 325, + 547, + 445 + ], + "type": "text", + "content": "tic gradient optimization in deep learning. 2) Popular techniques, such as data augmentation, batch normalization [15] and dropout [10], rule out the possibility of tracking original samples and applying sample-wise multipliers. To overcome this complication, we introduce class-wise multipliers, instead of sample-wise multipliers in the standard ALM. 3) The outer-step update for estimating optimal ALM multipliers is performed on the validation set, which is meaningful for training on large-scale training set and avoids potential overfitting." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 450, + 548, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 450, + 548, + 582 + ], + "spans": [ + { + "bbox": [ + 306, + 450, + 548, + 582 + ], + "type": "text", + "content": "- Comprehensive experiments over a variety of applications and benchmarks, including standard image classification (Tiny-ImageNet and ImageNet), long-tailed image classification (ImageNet-LT), semantic segmentation (PASCAL VOC 2012), and text classification (20 Newsgroups), demonstrate the effectiveness of our CALS-ALM method. As shown in Figure 1, CALS-ALM yields superior performance over baselines and state-of-the-art calibration losses when considering both accuracy and calibration, especially for more realistic large-scaled datasets with large number of classes or class imbalance." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 591, + 392, + 604 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 591, + 392, + 604 + ], + "spans": [ + { + "bbox": [ + 306, + 591, + 392, + 604 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 611, + 430, + 623 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 611, + 430, + 623 + ], + "spans": [ + { + "bbox": [ + 306, + 611, + 430, + 623 + ], + "type": "text", + "content": "2.1. Problem Formulation" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 628, + 548, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 628, + 548, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 628, + 548, + 715 + ], + "type": "text", + "content": "Given a dataset " + }, + { + "bbox": [ + 304, + 628, + 548, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{D} = \\{(x^{(i)},y^{(i)})\\}_{i = 1}^{N}" + }, + { + "bbox": [ + 304, + 628, + 548, + 715 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 304, + 628, + 548, + 715 + ], + "type": "inline_equation", + "content": "N\\in \\mathbb{N}" + }, + { + "bbox": [ + 304, + 628, + 548, + 715 + ], + "type": "text", + "content": " pairs of samples " + }, + { + "bbox": [ + 304, + 628, + 548, + 715 + ], + "type": "inline_equation", + "content": "\\pmb {x}\\in \\mathcal{X}" + }, + { + "bbox": [ + 304, + 628, + 548, + 715 + ], + "type": "text", + "content": " and corresponding labels " + }, + { + "bbox": [ + 304, + 628, + 548, + 715 + ], + "type": "inline_equation", + "content": "y\\in \\mathcal{V}" + }, + { + "bbox": [ + 304, + 628, + 548, + 715 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 304, + 628, + 548, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{V} = \\{1,\\dots ,K\\}" + }, + { + "bbox": [ + 304, + 628, + 548, + 715 + ], + "type": "text", + "content": ", a deep neural network (DNN) " + }, + { + "bbox": [ + 304, + 628, + 548, + 715 + ], + "type": "inline_equation", + "content": "F_{\\theta}:X\\to \\mathbb{R}^{K}" + }, + { + "bbox": [ + 304, + 628, + 548, + 715 + ], + "type": "text", + "content": " parameterized by " + }, + { + "bbox": [ + 304, + 628, + 548, + 715 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 304, + 628, + 548, + 715 + ], + "type": "text", + "content": " yields logits " + }, + { + "bbox": [ + 304, + 628, + 548, + 715 + ], + "type": "inline_equation", + "content": "l = F_{\\theta}(\\pmb {x}) = (F_{\\theta}(\\pmb {x})_k)_{1\\leq k\\leq K}\\in \\mathbb{R}^K" + }, + { + "bbox": [ + 304, + 628, + 548, + 715 + ], + "type": "text", + "content": ". In a classification scenario, the output probability " + }, + { + "bbox": [ + 304, + 628, + 548, + 715 + ], + "type": "inline_equation", + "content": "s = (s_k)_{1\\leq k\\leq K}\\in \\Delta^{K - 1}" + }, + { + "bbox": [ + 304, + 628, + 548, + 715 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 304, + 628, + 548, + 715 + ], + "type": "inline_equation", + "content": "\\Delta^{K - 1}\\subset [0,1]^K" + }, + { + "bbox": [ + 304, + 628, + 548, + 715 + ], + "type": "text", + "content": " denotes the probability simplex, is" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16071" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "type": "text", + "content": "obtained by applying the softmax function on the output log-its, i.e. " + }, + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "type": "inline_equation", + "content": "s = \\text{softmax}(l) = \\frac{\\exp l}{\\sum \\exp l}" + }, + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "type": "text", + "content": ". Therefore, the predicted class " + }, + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "type": "inline_equation", + "content": "\\hat{y}" + }, + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "type": "text", + "content": " is computed as " + }, + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "type": "inline_equation", + "content": "\\hat{y} = \\arg \\max_k s_k" + }, + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "type": "text", + "content": ", and the predicted confidence is " + }, + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "type": "inline_equation", + "content": "\\hat{p} = s_{\\hat{y}} = \\max_k s_k" + }, + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "type": "text", + "content": ". A perfectly calibrated model should satisfy that the predicted confidence of any input is equal to the accuracy of the model: " + }, + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "type": "inline_equation", + "content": "\\hat{p} = \\mathbb{P}(\\hat{y} = y|\\hat{p})" + }, + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "type": "text", + "content": ". Hence, an over-confident model yields on average larger confidence than the associated accuracy, whereas an under-confident model yields lower confidence than its accuracy." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 182, + 288, + 231 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 182, + 288, + 231 + ], + "spans": [ + { + "bbox": [ + 46, + 182, + 288, + 231 + ], + "type": "text", + "content": "A number of recent studies [11, 22, 30, 31] have shown that DNNs tend to become over-confident during training as a result of minimizing the popular cross-entropy (CE) training loss:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 108, + 240, + 288, + 275 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 108, + 240, + 288, + 275 + ], + "spans": [ + { + "bbox": [ + 108, + 240, + 288, + 275 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {C E}} (\\boldsymbol {x}, y) = - \\sum_ {k = 1} ^ {K} y _ {k} \\log s _ {k} \\tag {1}", + "image_path": "1fbb0f59cb413bc7f2e397da4907070c0283cf65f2f347c968f2801b5c1a1f52.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 284, + 289, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 284, + 289, + 392 + ], + "spans": [ + { + "bbox": [ + 46, + 284, + 289, + 392 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 284, + 289, + 392 + ], + "type": "inline_equation", + "content": "\\pmb{y} \\in \\{0,1\\}^{K}" + }, + { + "bbox": [ + 46, + 284, + 289, + 392 + ], + "type": "text", + "content": " is the one-hot encoding of " + }, + { + "bbox": [ + 46, + 284, + 289, + 392 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 46, + 284, + 289, + 392 + ], + "type": "text", + "content": ". This objective function is minimized when the predictions for all the training samples fully match the ground-truth labels " + }, + { + "bbox": [ + 46, + 284, + 289, + 392 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 46, + 284, + 289, + 392 + ], + "type": "text", + "content": ", i.e. " + }, + { + "bbox": [ + 46, + 284, + 289, + 392 + ], + "type": "inline_equation", + "content": "s_{y} = 1" + }, + { + "bbox": [ + 46, + 284, + 289, + 392 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 284, + 289, + 392 + ], + "type": "inline_equation", + "content": "\\forall k \\neq y, s_{k} = 0" + }, + { + "bbox": [ + 46, + 284, + 289, + 392 + ], + "type": "text", + "content": ". The negative logarithmic term on the logit of the correct category renders the global minimization of the CE loss unreachable, as it keeps pushing the predicted probabilities " + }, + { + "bbox": [ + 46, + 284, + 289, + 392 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 46, + 284, + 289, + 392 + ], + "type": "text", + "content": " towards the vertices of the " + }, + { + "bbox": [ + 46, + 284, + 289, + 392 + ], + "type": "inline_equation", + "content": "(K-1)" + }, + { + "bbox": [ + 46, + 284, + 289, + 392 + ], + "type": "text", + "content": "-simplex even after the classification error is zero [31], resulting in over-confident models." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 402, + 185, + 415 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 402, + 185, + 415 + ], + "spans": [ + { + "bbox": [ + 47, + 402, + 185, + 415 + ], + "type": "text", + "content": "2.2. Post-processing methods" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 421, + 289, + 625 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 421, + 289, + 625 + ], + "spans": [ + { + "bbox": [ + 46, + 421, + 289, + 625 + ], + "type": "text", + "content": "To address mis-calibration, different post-processing techniques applied after model training have been proposed recently [8, 11, 43]. The most popular of these strategies is Temperature Scaling (TS) [11], which applies a single scalar temperature parameter to manipulate logit outputs monotonely, resulting in softened prediction confidences without affecting predicted labels. Note that here the temperature parameter needs to be tuned on a separate validation set. Despite its simplicity, TS has been shown effective in fixing over-confidence predictions [11]. As a local alternative to TS, [8] proposed to train a regression model for learning position-specific temperature for semantic segmentation problems. Unfortunately, TS and its variants can be sensitive to both the model and the validation set, and do not work well under data distribution shifts [36]. Thus, some subsequent works [28, 43] have attempted to provide solutions for improving performance under domain shift." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 634, + 186, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 634, + 186, + 647 + ], + "spans": [ + { + "bbox": [ + 47, + 634, + 186, + 647 + ], + "type": "text", + "content": "2.3. Learning-based methods" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "type": "text", + "content": "Another popular direction is to directly address miscalibration during training by introducing an additional penalty or supervision regarding model calibration with the standard training loss. In [18], the authors introduced a trainable calibration measure based on RKHS kernels, while [16]" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 72, + 547, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 289 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 289 + ], + "type": "text", + "content": "proposed a differential calibration loss based on a soft version of the binning operation in the ECE metric. In [5], two types of binary pairwise calibration constraints were proposed as additional penalty terms during training. Other methods try to decrease over-fitting on the cross-entropy loss, which has been demonstrated to be the main reason of over-confidence [11, 31]. In [38] an explicit confidence penalty (ECP) is proposed to maximize the entropy and reduce over-fitting, while Label Smoothing [42] has also been shown to implicitly improve the calibration [32] by softening the hard one-hot targets in the cross-entropy. The Focal Loss [21], originally proposed to tackle class imbalance, can also be effective for calibration [31], as it implicitly minimizes the Kullback-Leibler (KL) divergence between the uniform distribution and the network softmax probabilities, thereby increasing the entropy of the predictions. As an extension the Sample-Dependent Focal Loss (FLSD) was also proposed in [31] to further boost calibration performance." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 307, + 547, + 438 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 307, + 547, + 438 + ], + "spans": [ + { + "bbox": [ + 304, + 307, + 547, + 438 + ], + "type": "text", + "content": "Margin-based Label Smoothing (MbLS). A unifying constrained-optimization formulation of loss functions promoting calibration has been recently presented in [22]. Specifically, the additional penalties integrated in these methods, including ECP [38], LS [32] and FL [31], can be viewed as different forms of approximations to the same constraint, i.e. enforcing the logit distances to be zero. Noticing that this constraint is non-informative (its solution being uniformly distributed probabilities), [22] further proposed a generalized formulation by relaxing the constraint to allow the logit distances being lower than a strictly positive margin." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 439, + 547, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 439, + 547, + 475 + ], + "spans": [ + { + "bbox": [ + 304, + 439, + 547, + 475 + ], + "type": "text", + "content": "The specific formulation of MbLS [22] is as follows. Given a margin " + }, + { + "bbox": [ + 304, + 439, + 547, + 475 + ], + "type": "inline_equation", + "content": "m \\in \\mathbb{R}_+" + }, + { + "bbox": [ + 304, + 439, + 547, + 475 + ], + "type": "text", + "content": ", the constrained optimization problem for network training is:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 318, + 486, + 545, + 518 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 486, + 545, + 518 + ], + "spans": [ + { + "bbox": [ + 318, + 486, + 545, + 518 + ], + "type": "interline_equation", + "content": "\\min _ {\\theta} \\sum_ {i = 1} ^ {N} \\mathcal {L} _ {\\mathrm {C E}} \\left(\\boldsymbol {x} ^ {(i)}, y ^ {(i)}\\right) \\tag {2}", + "image_path": "c890f44d78eebe88bb348054435f3958bd2d4135db659a417bb15a60f7a6e074.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 320, + 520, + 521, + 540 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 520, + 521, + 540 + ], + "spans": [ + { + "bbox": [ + 320, + 520, + 521, + 540 + ], + "type": "interline_equation", + "content": "\\text {s . t .} \\quad \\max _ {k} \\left\\{\\boldsymbol {l} _ {k} ^ {(i)} \\right\\} - l ^ {(i)} \\preceq m \\mathbf {1} _ {K}, \\quad i = 1, \\dots , N,", + "image_path": "dd2639ddd39bd7c152545d11a69326658edfae2a04ad0ed57131eea396d0f1b8.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 551, + 545, + 600 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 551, + 545, + 600 + ], + "spans": [ + { + "bbox": [ + 304, + 551, + 545, + 600 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 551, + 545, + 600 + ], + "type": "inline_equation", + "content": "l^{(i)} = F_{\\theta}(\\pmb{x}^{(i)})" + }, + { + "bbox": [ + 304, + 551, + 545, + 600 + ], + "type": "text", + "content": ". The minimum can be approximated by penalty-based optimization methods, transforming the above formulation into an unconstrained problem by means of simple ReLU functions:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 330, + 610, + 545, + 650 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 330, + 610, + 545, + 650 + ], + "spans": [ + { + "bbox": [ + 330, + 610, + 545, + 650 + ], + "type": "interline_equation", + "content": "\\min _ {\\theta} \\sum_ {i = 1} ^ {N} \\mathcal {L} _ {\\mathrm {C E}} \\left(\\boldsymbol {x} ^ {(i)}, y ^ {(i)}\\right) + \\tag {3}", + "image_path": "2b826fc42da02881870c96e4bf008f58528f1a941fddbf7e56363fb120d67b95.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 352, + 646, + 521, + 679 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 352, + 646, + 521, + 679 + ], + "spans": [ + { + "bbox": [ + 352, + 646, + 521, + 679 + ], + "type": "interline_equation", + "content": "\\lambda \\sum_ {i = 1} ^ {N} \\sum_ {j = 1} ^ {K} \\max \\{0, \\max _ {k} \\{l _ {k} ^ {(i)} \\} - l _ {j} ^ {(i)} - m \\},", + "image_path": "519372a2b1d85c7f1b005b35f858d4bbc38fec71effc04a17dac20bdcf8711ba.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 689, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 545, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 306, + 689, + 545, + 714 + ], + "type": "inline_equation", + "content": "\\lambda \\in \\mathbb{R}_+" + }, + { + "bbox": [ + 306, + 689, + 545, + 714 + ], + "type": "text", + "content": " is a scalar weight balancing contributions of the CE loss and the corresponding penalty." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16072" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 71, + 287, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 71, + 287, + 85 + ], + "spans": [ + { + "bbox": [ + 47, + 71, + 287, + 85 + ], + "type": "text", + "content": "3. Sample-wise Constrained DNN Optimization" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 91, + 289, + 212 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 91, + 289, + 212 + ], + "spans": [ + { + "bbox": [ + 46, + 91, + 289, + 212 + ], + "type": "text", + "content": "Although MbLS can significantly improve calibration, the associated constrained problem (2) is not solved accurately. It is approximated by an unconstrained problem with a single uniform penalty, regardless of the data sample or category. However, the samples and classes considered in a classification problem have different intrinsic learning difficulties. Therefore, an improved training scheme would involve considering distinct penalty weights " + }, + { + "bbox": [ + 46, + 91, + 289, + 212 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 46, + 91, + 289, + 212 + ], + "type": "text", + "content": " for each sample and class. This would result in having to chose " + }, + { + "bbox": [ + 46, + 91, + 289, + 212 + ], + "type": "inline_equation", + "content": "N\\times K" + }, + { + "bbox": [ + 46, + 91, + 289, + 212 + ], + "type": "text", + "content": " penalty weights " + }, + { + "bbox": [ + 46, + 91, + 289, + 212 + ], + "type": "inline_equation", + "content": "\\Lambda \\in \\mathbb{R}_{+}^{N\\times K}" + }, + { + "bbox": [ + 46, + 91, + 289, + 212 + ], + "type": "text", + "content": ", with the resulting optimization problem being:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 219, + 287, + 288 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 219, + 287, + 288 + ], + "spans": [ + { + "bbox": [ + 58, + 219, + 287, + 288 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\min _ {\\theta} \\sum_ {i = 1 \\atop N = K} ^ {N} \\mathcal {L} _ {\\mathrm {C E}} \\left(\\boldsymbol {x} ^ {(i)}, y ^ {(i)}\\right) + \\tag {4} \\\\ \\sum_ {i = 1} ^ {N} \\sum_ {j = 1} ^ {K} \\Lambda_ {i j} \\max \\{0, \\max _ {k} \\{l _ {k} ^ {(i)} \\} - l _ {j} ^ {(i)} - m \\}. \\\\ \\end{array}", + "image_path": "826b39d8d91f707d79eab0b2f9e9bfd30fce33c19be8b1b2c5fef78d5643cf2b.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 295, + 287, + 366 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 295, + 287, + 366 + ], + "spans": [ + { + "bbox": [ + 46, + 295, + 287, + 366 + ], + "type": "text", + "content": "From an optimization perspective, supposing that optimal weights " + }, + { + "bbox": [ + 46, + 295, + 287, + 366 + ], + "type": "inline_equation", + "content": "\\theta^{\\star}" + }, + { + "bbox": [ + 46, + 295, + 287, + 366 + ], + "type": "text", + "content": " exists for problem (2), there exists " + }, + { + "bbox": [ + 46, + 295, + 287, + 366 + ], + "type": "inline_equation", + "content": "\\Lambda^{\\star} \\in \\mathbb{R}_{+}^{N \\times K}" + }, + { + "bbox": [ + 46, + 295, + 287, + 366 + ], + "type": "text", + "content": " such that " + }, + { + "bbox": [ + 46, + 295, + 287, + 366 + ], + "type": "inline_equation", + "content": "(\\theta^{\\star}, \\Lambda^{\\star})" + }, + { + "bbox": [ + 46, + 295, + 287, + 366 + ], + "type": "text", + "content": " is a saddle point of the Lagrangian associated to (2). These " + }, + { + "bbox": [ + 46, + 295, + 287, + 366 + ], + "type": "inline_equation", + "content": "\\Lambda^{\\star}" + }, + { + "bbox": [ + 46, + 295, + 287, + 366 + ], + "type": "text", + "content": " are the Lagrange multipliers of the problem. Therefore, using " + }, + { + "bbox": [ + 46, + 295, + 287, + 366 + ], + "type": "inline_equation", + "content": "\\Lambda = \\Lambda^{\\star}" + }, + { + "bbox": [ + 46, + 295, + 287, + 366 + ], + "type": "text", + "content": " would be the best choice to solve (4)." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 367, + 289, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 367, + 289, + 653 + ], + "spans": [ + { + "bbox": [ + 46, + 367, + 289, + 653 + ], + "type": "text", + "content": "In practice, using the Lagrange multipliers of problem (2) as the weights for the penalties may not be computationally feasible, and it could even result in degraded performance. Indeed, in the context of machine learning, we optimize a model's weights " + }, + { + "bbox": [ + 46, + 367, + 289, + 653 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 46, + 367, + 289, + 653 + ], + "type": "text", + "content": " to solve (2) on a training set " + }, + { + "bbox": [ + 46, + 367, + 289, + 653 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{\\mathrm{train}}" + }, + { + "bbox": [ + 46, + 367, + 289, + 653 + ], + "type": "text", + "content": ", and expect to generalize on a test set " + }, + { + "bbox": [ + 46, + 367, + 289, + 653 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{\\mathrm{test}}" + }, + { + "bbox": [ + 46, + 367, + 289, + 653 + ], + "type": "text", + "content": ", which we do not have access to during training. Because of the bias-variance tradeoff, solving (2) optimally would likely result in overfitting, i.e. we may solve problem (2) accurately on the train set, but not generalize properly on the test set, resulting in poor calibration and classification performance overall. This suggests that it could be preferable to evaluate during training the quality of multipliers on a separate validation set " + }, + { + "bbox": [ + 46, + 367, + 289, + 653 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{\\mathrm{val}}" + }, + { + "bbox": [ + 46, + 367, + 289, + 653 + ], + "type": "text", + "content": ". Additionally, several mechanisms for training DNNs are not compatible with a straightforward minimization. First, the use of batch normalization yields predictions that are not independent between samples in a minibatch. Second, the use of regularization techniques such as dropout may lead to virtually inaccurate predictions on certain training samples, impacting the correct estimation of multipliers. Third, data augmentation, which is standard in DNN training, would result in additional penalty weights for the augmented samples: they can be easier or harder to classify than the original ones." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 654, + 289, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 654, + 289, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 654, + 289, + 714 + ], + "type": "text", + "content": "In addition to the above obstacles, applying a penalty weight per sample and per class (resulting in " + }, + { + "bbox": [ + 46, + 654, + 289, + 714 + ], + "type": "inline_equation", + "content": "N \\times K" + }, + { + "bbox": [ + 46, + 654, + 289, + 714 + ], + "type": "text", + "content": " weights) would not scale well for large datasets and dense predictive tasks, such as semantic segmentation, which is typically formulated as a per-pixel classification task. Assuming that im" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 72, + 547, + 179 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 179 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 179 + ], + "type": "text", + "content": "ages in the dataset have a size of " + }, + { + "bbox": [ + 304, + 72, + 547, + 179 + ], + "type": "inline_equation", + "content": "H \\times W" + }, + { + "bbox": [ + 304, + 72, + 547, + 179 + ], + "type": "text", + "content": ", this would result in " + }, + { + "bbox": [ + 304, + 72, + 547, + 179 + ], + "type": "inline_equation", + "content": "N \\times H \\times W \\times K" + }, + { + "bbox": [ + 304, + 72, + 547, + 179 + ], + "type": "text", + "content": " penalty weights. This rapidly becomes a limiting factor for moderately sized segmentation datasets. For instance, Pascal VOC 2012 [9] contains 21 classes and 1464 training images, amounting to " + }, + { + "bbox": [ + 304, + 72, + 547, + 179 + ], + "type": "inline_equation", + "content": "2.62 \\times 10^{8}" + }, + { + "bbox": [ + 304, + 72, + 547, + 179 + ], + "type": "text", + "content": " pixels, or " + }, + { + "bbox": [ + 304, + 72, + 547, + 179 + ], + "type": "inline_equation", + "content": "5.5 \\times 10^{9}" + }, + { + "bbox": [ + 304, + 72, + 547, + 179 + ], + "type": "text", + "content": " penalty weights, which, stored as float32, would use " + }, + { + "bbox": [ + 304, + 72, + 547, + 179 + ], + "type": "inline_equation", + "content": "\\sim 20" + }, + { + "bbox": [ + 304, + 72, + 547, + 179 + ], + "type": "text", + "content": " GiB. For Cityscapes [6], containing approximately 3000 training images of size " + }, + { + "bbox": [ + 304, + 72, + 547, + 179 + ], + "type": "inline_equation", + "content": "2048 \\times 1024" + }, + { + "bbox": [ + 304, + 72, + 547, + 179 + ], + "type": "text", + "content": " in 19 classes, this amounts to " + }, + { + "bbox": [ + 304, + 72, + 547, + 179 + ], + "type": "inline_equation", + "content": "\\sim 445" + }, + { + "bbox": [ + 304, + 72, + 547, + 179 + ], + "type": "text", + "content": " GiB." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 180, + 547, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 180, + 547, + 217 + ], + "spans": [ + { + "bbox": [ + 304, + 180, + 547, + 217 + ], + "type": "text", + "content": "Following the above observations, we introduce a relaxation of sample-wise penalties, and propose to solve the following problem:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 318, + 224, + 545, + 294 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 224, + 545, + 294 + ], + "spans": [ + { + "bbox": [ + 318, + 224, + 545, + 294 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\min _ {\\theta} \\sum_ {i = 1 \\atop N, K} ^ {N} \\mathcal {L} _ {\\mathrm {C E}} \\left(\\boldsymbol {x} ^ {(i)}, y ^ {(i)}\\right) + \\tag {5} \\\\ \\sum_ {i = 1} ^ {N} \\sum_ {j = 1} ^ {K} \\lambda_ {j} \\max \\{0, \\max _ {k} \\{l _ {k} ^ {(i)} \\} - l _ {j} ^ {(i)} - m \\}, \\\\ \\end{array}", + "image_path": "daa0b239f7666ba8aaff02871d222b96bc9a8b3cedeb1786e9d2345f9c7c3a25.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 304, + 547, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 304, + 547, + 413 + ], + "spans": [ + { + "bbox": [ + 304, + 304, + 547, + 413 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 304, + 547, + 413 + ], + "type": "inline_equation", + "content": "(\\lambda_j)_{1\\leq j\\leq K}\\in \\mathbb{R}_+^K" + }, + { + "bbox": [ + 304, + 304, + 547, + 413 + ], + "type": "text", + "content": ". Since penalties are now class-wise, we need " + }, + { + "bbox": [ + 304, + 304, + 547, + 413 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 304, + 304, + 547, + 413 + ], + "type": "text", + "content": " penalty weights. This has the advantage to scale well to denser classification tasks such as segmentation. However, we still face a challenging optimization problem (5), since we still need to chose " + }, + { + "bbox": [ + 304, + 304, + 547, + 413 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 304, + 304, + 547, + 413 + ], + "type": "text", + "content": " weights, which can be extremely complicated for large-scale datasets with many classes such as ImageNet, which contains 1000 classes. In the next section, we introduce a numerical technique to deal with this challenge." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 424, + 507, + 437 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 424, + 507, + 437 + ], + "spans": [ + { + "bbox": [ + 305, + 424, + 507, + 437 + ], + "type": "text", + "content": "4. Class Adaptive Network Calibration" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 444, + 547, + 504 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 444, + 547, + 504 + ], + "spans": [ + { + "bbox": [ + 304, + 444, + 547, + 504 + ], + "type": "text", + "content": "The challenge of the previous formulation stems from correctly choosing the weights " + }, + { + "bbox": [ + 304, + 444, + 547, + 504 + ], + "type": "inline_equation", + "content": "\\lambda \\in \\mathbb{R}_+^K" + }, + { + "bbox": [ + 304, + 444, + 547, + 504 + ], + "type": "text", + "content": ", which can be cumbersome when " + }, + { + "bbox": [ + 304, + 444, + 547, + 504 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 304, + 444, + 547, + 504 + ], + "type": "text", + "content": " is large. Therefore, we propose to use an Augmented Lagrangian Multiplier (ALM) method to adaptively learn the weights of the penalties." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 512, + 394, + 523 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 512, + 394, + 523 + ], + "spans": [ + { + "bbox": [ + 306, + 512, + 394, + 523 + ], + "type": "text", + "content": "4.1. General ALM" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 531, + 545, + 591 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 531, + 545, + 591 + ], + "spans": [ + { + "bbox": [ + 304, + 531, + 545, + 591 + ], + "type": "text", + "content": "ALM methods combine penalties and primal-dual updates to solve a constrained problem. They have well-established advantages and enjoy widespread popularity in the general context of optimization [1,35,41]. Specifically, we have the following generic constrained optimization problem:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 328, + 601, + 545, + 617 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 328, + 601, + 545, + 617 + ], + "spans": [ + { + "bbox": [ + 328, + 601, + 545, + 617 + ], + "type": "interline_equation", + "content": "\\min _ {x} f (x) \\quad \\text {s . t .} \\quad h _ {i} (x) \\leq 0, \\quad i = 1, \\dots , n \\tag {6}", + "image_path": "83007018beeb7d660695eb057aa1044bc78841361af1d82029130967cd2c5ec5.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 627, + 545, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 627, + 545, + 676 + ], + "spans": [ + { + "bbox": [ + 304, + 627, + 545, + 676 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 627, + 545, + 676 + ], + "type": "inline_equation", + "content": "f: \\mathbb{R}^d \\to \\mathbb{R}" + }, + { + "bbox": [ + 304, + 627, + 545, + 676 + ], + "type": "text", + "content": " is the objective function and " + }, + { + "bbox": [ + 304, + 627, + 545, + 676 + ], + "type": "inline_equation", + "content": "h_i: \\mathbb{R}^d \\to \\mathbb{R}, i = 1, \\dots, n" + }, + { + "bbox": [ + 304, + 627, + 545, + 676 + ], + "type": "text", + "content": " are the constraint functions. We tackle it by approximately solving a sequence " + }, + { + "bbox": [ + 304, + 627, + 545, + 676 + ], + "type": "inline_equation", + "content": "j \\in \\mathbb{N}" + }, + { + "bbox": [ + 304, + 627, + 545, + 676 + ], + "type": "text", + "content": " of unconstrained problems:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 318, + 685, + 545, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 685, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 318, + 685, + 545, + 715 + ], + "type": "interline_equation", + "content": "\\min _ {x} \\quad \\mathcal {L} ^ {(j)} (x) = f (x) + \\sum_ {i = 1} ^ {n} P \\left(h _ {i} (x), \\rho_ {i} ^ {(j)}, \\lambda_ {i} ^ {(j)}\\right) \\tag {7}", + "image_path": "74e9b5996e921f029f759011d215949225ca6bd709d4f8b9b9573f73eaf93cca.jpg" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16073" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 66, + 72, + 272, + 198 + ], + "blocks": [ + { + "bbox": [ + 66, + 72, + 272, + 198 + ], + "lines": [ + { + "bbox": [ + 66, + 72, + 272, + 198 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 272, + 198 + ], + "type": "image", + "image_path": "77ae33c1b29ebad9fab4420309eee8c8a6c20490d3a1172073de4002ba940f03.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 47, + 209, + 288, + 242 + ], + "lines": [ + { + "bbox": [ + 47, + 209, + 288, + 242 + ], + "spans": [ + { + "bbox": [ + 47, + 209, + 288, + 242 + ], + "type": "text", + "content": "Figure 2. A penalty-Lagrangian function " + }, + { + "bbox": [ + 47, + 209, + 288, + 242 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 47, + 209, + 288, + 242 + ], + "type": "text", + "content": " with varying values of " + }, + { + "bbox": [ + 47, + 209, + 288, + 242 + ], + "type": "inline_equation", + "content": "\\rho" + }, + { + "bbox": [ + 47, + 209, + 288, + 242 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 209, + 288, + 242 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 47, + 209, + 288, + 242 + ], + "type": "text", + "content": ". Higher values of " + }, + { + "bbox": [ + 47, + 209, + 288, + 242 + ], + "type": "inline_equation", + "content": "\\rho" + }, + { + "bbox": [ + 47, + 209, + 288, + 242 + ], + "type": "text", + "content": " bring " + }, + { + "bbox": [ + 47, + 209, + 288, + 242 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 47, + 209, + 288, + 242 + ], + "type": "text", + "content": " closer to an ideal penalty. The multiplier " + }, + { + "bbox": [ + 47, + 209, + 288, + 242 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 47, + 209, + 288, + 242 + ], + "type": "text", + "content": " is the derivative of " + }, + { + "bbox": [ + 47, + 209, + 288, + 242 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 47, + 209, + 288, + 242 + ], + "type": "text", + "content": " w.r.t. the constraint at " + }, + { + "bbox": [ + 47, + 209, + 288, + 242 + ], + "type": "inline_equation", + "content": "z = 0" + }, + { + "bbox": [ + 47, + 209, + 288, + 242 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 261, + 288, + 334 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 261, + 288, + 334 + ], + "spans": [ + { + "bbox": [ + 46, + 261, + 288, + 334 + ], + "type": "text", + "content": "with " + }, + { + "bbox": [ + 46, + 261, + 288, + 334 + ], + "type": "inline_equation", + "content": "P: \\mathbb{R} \\times \\mathbb{R}_{++}^{n} \\times \\mathbb{R}_{++}^{n} \\to \\mathbb{R}" + }, + { + "bbox": [ + 46, + 261, + 288, + 334 + ], + "type": "text", + "content": " a penalty-Lagrangian function, and " + }, + { + "bbox": [ + 46, + 261, + 288, + 334 + ], + "type": "inline_equation", + "content": "\\pmb{\\rho}^{(j)} = (\\rho_i)_{1 \\leq i \\leq n} \\in \\mathbb{R}_{++}^n" + }, + { + "bbox": [ + 46, + 261, + 288, + 334 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 261, + 288, + 334 + ], + "type": "inline_equation", + "content": "\\pmb{\\lambda}^{(j)} = (\\lambda_i^{(j)})_{1 \\leq i \\leq n} \\in \\mathbb{R}_{++}^n" + }, + { + "bbox": [ + 46, + 261, + 288, + 334 + ], + "type": "text", + "content": " the penalty parameters and multipliers associated to " + }, + { + "bbox": [ + 46, + 261, + 288, + 334 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 46, + 261, + 288, + 334 + ], + "type": "text", + "content": " at the " + }, + { + "bbox": [ + 46, + 261, + 288, + 334 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 46, + 261, + 288, + 334 + ], + "type": "text", + "content": "-th iteration. This sequence of unconstrained problems is called outer iterations, while the steps in the minimization of " + }, + { + "bbox": [ + 46, + 261, + 288, + 334 + ], + "type": "inline_equation", + "content": "\\mathcal{L}^{(j)}" + }, + { + "bbox": [ + 46, + 261, + 288, + 334 + ], + "type": "text", + "content": " are called inner iterations." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 335, + 288, + 539 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 335, + 288, + 539 + ], + "spans": [ + { + "bbox": [ + 46, + 335, + 288, + 539 + ], + "type": "text", + "content": "The main components of ALM methods are (i) the penalty-Lagrangian function " + }, + { + "bbox": [ + 46, + 335, + 288, + 539 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 46, + 335, + 288, + 539 + ], + "type": "text", + "content": ", (ii) the update of the penalty multipliers " + }, + { + "bbox": [ + 46, + 335, + 288, + 539 + ], + "type": "inline_equation", + "content": "\\lambda^{(j)}" + }, + { + "bbox": [ + 46, + 335, + 288, + 539 + ], + "type": "text", + "content": " and (iii) the increase of the penalty parameters " + }, + { + "bbox": [ + 46, + 335, + 288, + 539 + ], + "type": "inline_equation", + "content": "\\pmb{\\rho}^{(j)}" + }, + { + "bbox": [ + 46, + 335, + 288, + 539 + ], + "type": "text", + "content": ". First, the penalty function " + }, + { + "bbox": [ + 46, + 335, + 288, + 539 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 46, + 335, + 288, + 539 + ], + "type": "text", + "content": " needs to satisfy a set of axioms [2] (see Appendix A): these axioms constrain the function to be continuously differentiable w.r.t. its first variable and to have a non-negative derivative: " + }, + { + "bbox": [ + 46, + 335, + 288, + 539 + ], + "type": "inline_equation", + "content": "\\forall z \\in \\mathbb{R}, P'(z, \\rho, \\lambda) = \\frac{\\partial}{\\partial z} P(z, \\rho, \\lambda) \\geq 0" + }, + { + "bbox": [ + 46, + 335, + 288, + 539 + ], + "type": "text", + "content": ", with " + }, + { + "bbox": [ + 46, + 335, + 288, + 539 + ], + "type": "inline_equation", + "content": "P'(0, \\rho, \\lambda) = \\lambda" + }, + { + "bbox": [ + 46, + 335, + 288, + 539 + ], + "type": "text", + "content": ". Figure 2 gives an example of a penalty, and how " + }, + { + "bbox": [ + 46, + 335, + 288, + 539 + ], + "type": "inline_equation", + "content": "\\rho" + }, + { + "bbox": [ + 46, + 335, + 288, + 539 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 335, + 288, + 539 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 46, + 335, + 288, + 539 + ], + "type": "text", + "content": " affect it. The choice of penalty function is critical to the performance of ALM methods, especially for nonconvex problems [2]. Typical functions include PHR [14,39], " + }, + { + "bbox": [ + 46, + 335, + 288, + 539 + ], + "type": "inline_equation", + "content": "P_{2}" + }, + { + "bbox": [ + 46, + 335, + 288, + 539 + ], + "type": "text", + "content": " [17] and " + }, + { + "bbox": [ + 46, + 335, + 288, + 539 + ], + "type": "inline_equation", + "content": "P_{3}" + }, + { + "bbox": [ + 46, + 335, + 288, + 539 + ], + "type": "text", + "content": " [34] (see section 3.2 of [2]). Second, the penalty multipliers " + }, + { + "bbox": [ + 46, + 335, + 288, + 539 + ], + "type": "inline_equation", + "content": "\\lambda^{(j)}" + }, + { + "bbox": [ + 46, + 335, + 288, + 539 + ], + "type": "text", + "content": " are updated to the derivative of " + }, + { + "bbox": [ + 46, + 335, + 288, + 539 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 46, + 335, + 288, + 539 + ], + "type": "text", + "content": " w.r.t. the solution obtained during the last inner minimization. Formally, let " + }, + { + "bbox": [ + 46, + 335, + 288, + 539 + ], + "type": "inline_equation", + "content": "x^{(j)}" + }, + { + "bbox": [ + 46, + 335, + 288, + 539 + ], + "type": "text", + "content": " be the approximate minimizer of " + }, + { + "bbox": [ + 46, + 335, + 288, + 539 + ], + "type": "inline_equation", + "content": "\\mathcal{L}^{(j)}" + }, + { + "bbox": [ + 46, + 335, + 288, + 539 + ], + "type": "text", + "content": ", then " + }, + { + "bbox": [ + 46, + 335, + 288, + 539 + ], + "type": "inline_equation", + "content": "\\forall i \\in \\{1, \\dots, n\\}" + }, + { + "bbox": [ + 46, + 335, + 288, + 539 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 99, + 545, + 287, + 562 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 99, + 545, + 287, + 562 + ], + "spans": [ + { + "bbox": [ + 99, + 545, + 287, + 562 + ], + "type": "interline_equation", + "content": "\\lambda_ {i} ^ {(j + 1)} = P ^ {\\prime} \\left(h _ {i} \\left(x ^ {(j)}\\right), \\rho_ {i} ^ {(j)}, \\lambda_ {i} ^ {(j)}\\right) \\tag {8}", + "image_path": "11157b0f87b46e11ccc17b10a451a1eba357334f8f4d425e6430631a872b8e94.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 568, + 287, + 642 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 568, + 287, + 642 + ], + "spans": [ + { + "bbox": [ + 46, + 568, + 287, + 642 + ], + "type": "text", + "content": "This update rule corresponds to a first-order multiplier estimate for the constrained problem. Third, the penalty parameters " + }, + { + "bbox": [ + 46, + 568, + 287, + 642 + ], + "type": "inline_equation", + "content": "\\rho^{(j)}" + }, + { + "bbox": [ + 46, + 568, + 287, + 642 + ], + "type": "text", + "content": " are increased during the outer iterations if the constraints do not improve (i.e. is closer to being satisfied) compared to the previous outer iteration. Typically, " + }, + { + "bbox": [ + 46, + 568, + 287, + 642 + ], + "type": "inline_equation", + "content": "\\rho_{i}^{(j + 1)} = \\gamma \\rho_{i}^{(j)}" + }, + { + "bbox": [ + 46, + 568, + 287, + 642 + ], + "type": "text", + "content": " if " + }, + { + "bbox": [ + 46, + 568, + 287, + 642 + ], + "type": "inline_equation", + "content": "h_i" + }, + { + "bbox": [ + 46, + 568, + 287, + 642 + ], + "type": "text", + "content": " does not improve, with " + }, + { + "bbox": [ + 46, + 568, + 287, + 642 + ], + "type": "inline_equation", + "content": "\\gamma > 1" + }, + { + "bbox": [ + 46, + 568, + 287, + 642 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 642, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 642, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 642, + 288, + 713 + ], + "type": "text", + "content": "When the problem is convex, alternating between the approximate minimization of (7) and the update of the multipliers (8) leads to a solution for the constrained problem. The inner minimization corresponds to minimizing the primal while the outer iterations correspond to solving the dual problem. The complete procedure is presented in Algorithm 1." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 72, + 547, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 120 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 120 + ], + "type": "text", + "content": "Although guarantees exist only in the convex case, it is well-known that ALM methods can efficiently solve nonconvex problems as well [2]. In the context of deep learning, their use has been surprisingly underexplored [40, 41]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 132, + 542, + 144 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 132, + 542, + 144 + ], + "spans": [ + { + "bbox": [ + 306, + 132, + 542, + 144 + ], + "type": "text", + "content": "Algorithm 1 Augmented Lagrangian Multiplier algorithm" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 147, + 422, + 158 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 147, + 422, + 158 + ], + "spans": [ + { + "bbox": [ + 306, + 147, + 422, + 158 + ], + "type": "text", + "content": "Require: Objective function " + }, + { + "bbox": [ + 306, + 147, + 422, + 158 + ], + "type": "inline_equation", + "content": "f" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 159, + 481, + 169 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 159, + 481, + 169 + ], + "spans": [ + { + "bbox": [ + 306, + 159, + 481, + 169 + ], + "type": "text", + "content": "Require: Constraint functions " + }, + { + "bbox": [ + 306, + 159, + 481, + 169 + ], + "type": "inline_equation", + "content": "h_i, i = 1, \\dots, n" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 169, + 537, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 169, + 537, + 180 + ], + "spans": [ + { + "bbox": [ + 306, + 169, + 537, + 180 + ], + "type": "text", + "content": "Require: Penalty function " + }, + { + "bbox": [ + 306, + 169, + 537, + 180 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 306, + 169, + 537, + 180 + ], + "type": "text", + "content": ", initial " + }, + { + "bbox": [ + 306, + 169, + 537, + 180 + ], + "type": "inline_equation", + "content": "\\pmb{\\lambda}^{(0)} \\in \\mathbb{R}_{++}^{n}" + }, + { + "bbox": [ + 306, + 169, + 537, + 180 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 306, + 169, + 537, + 180 + ], + "type": "inline_equation", + "content": "\\pmb{\\rho}^{(0)} \\in \\mathbb{R}_{++}^{n}" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 180, + 478, + 191 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 180, + 478, + 191 + ], + "spans": [ + { + "bbox": [ + 306, + 180, + 478, + 191 + ], + "type": "text", + "content": "Require: Initial variable " + }, + { + "bbox": [ + 306, + 180, + 478, + 191 + ], + "type": "inline_equation", + "content": "x^{(0)}" + }, + { + "bbox": [ + 306, + 180, + 478, + 191 + ], + "type": "text", + "content": ", iterations " + }, + { + "bbox": [ + 306, + 180, + 478, + 191 + ], + "type": "inline_equation", + "content": "j = 1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 332, + 213, + 500, + 225 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 332, + 213, + 500, + 225 + ], + "spans": [ + { + "bbox": [ + 332, + 213, + 500, + 225 + ], + "type": "interline_equation", + "content": "\\mathcal {L} ^ {(j)} (x) = f (x) + \\sum_ {i = 1} ^ {n} P \\left(h _ {i} (x), \\rho_ {i} ^ {(j)}, \\lambda_ {i} ^ {(j)}\\right)", + "image_path": "0e4bb31bb489fbb37ba1f032e62aa0f21907ffa8847021646a756afc18862bac.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 191, + 524, + 346 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 312, + 191, + 409, + 201 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 191, + 409, + 201 + ], + "spans": [ + { + "bbox": [ + 312, + 191, + 409, + 201 + ], + "type": "text", + "content": "1: while not converged do" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 312, + 202, + 524, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 202, + 524, + 213 + ], + "spans": [ + { + "bbox": [ + 312, + 202, + 524, + 213 + ], + "type": "text", + "content": "2: Initialize with " + }, + { + "bbox": [ + 312, + 202, + 524, + 213 + ], + "type": "inline_equation", + "content": "x^{(j - 1)}" + }, + { + "bbox": [ + 312, + 202, + 524, + 213 + ], + "type": "text", + "content": " and minimize (approximately):" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 312, + 225, + 480, + 237 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 225, + 480, + 237 + ], + "spans": [ + { + "bbox": [ + 312, + 225, + 480, + 237 + ], + "type": "text", + "content": "3: " + }, + { + "bbox": [ + 312, + 225, + 480, + 237 + ], + "type": "inline_equation", + "content": "x^{(j)}\\gets" + }, + { + "bbox": [ + 312, + 225, + 480, + 237 + ], + "type": "text", + "content": " (approximate) minimizer of " + }, + { + "bbox": [ + 312, + 225, + 480, + 237 + ], + "type": "inline_equation", + "content": "\\mathcal{L}^{(j)}" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 312, + 238, + 405, + 247 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 238, + 405, + 247 + ], + "spans": [ + { + "bbox": [ + 312, + 238, + 405, + 247 + ], + "type": "text", + "content": "4: for " + }, + { + "bbox": [ + 312, + 238, + 405, + 247 + ], + "type": "inline_equation", + "content": "i = 1,\\dots ,n" + }, + { + "bbox": [ + 312, + 238, + 405, + 247 + ], + "type": "text", + "content": " do" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 312, + 247, + 467, + 259 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 247, + 467, + 259 + ], + "spans": [ + { + "bbox": [ + 312, + 247, + 467, + 259 + ], + "type": "text", + "content": "5: " + }, + { + "bbox": [ + 312, + 247, + 467, + 259 + ], + "type": "inline_equation", + "content": "\\lambda_{i}^{(j + 1)}\\gets P^{\\prime}(h_{i}(x^{(j)}),\\rho_{i}^{(j)},\\lambda_{i}^{(j)})" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 312, + 260, + 499, + 270 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 260, + 499, + 270 + ], + "spans": [ + { + "bbox": [ + 312, + 260, + 499, + 270 + ], + "type": "text", + "content": "6: if the " + }, + { + "bbox": [ + 312, + 260, + 499, + 270 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 312, + 260, + 499, + 270 + ], + "type": "text", + "content": "-th constraint does not improve then" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 312, + 270, + 410, + 282 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 270, + 410, + 282 + ], + "spans": [ + { + "bbox": [ + 312, + 270, + 410, + 282 + ], + "type": "text", + "content": "7: " + }, + { + "bbox": [ + 312, + 270, + 410, + 282 + ], + "type": "inline_equation", + "content": "\\rho_{i}^{(j + 1)}\\gets \\gamma \\rho_{i}^{(j)}" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 312, + 283, + 358, + 292 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 283, + 358, + 292 + ], + "spans": [ + { + "bbox": [ + 312, + 283, + 358, + 292 + ], + "type": "text", + "content": "8: else" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 312, + 293, + 405, + 304 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 293, + 405, + 304 + ], + "spans": [ + { + "bbox": [ + 312, + 293, + 405, + 304 + ], + "type": "text", + "content": "9: " + }, + { + "bbox": [ + 312, + 293, + 405, + 304 + ], + "type": "inline_equation", + "content": "\\rho_{i}^{(j + 1)}\\gets \\rho_{i}^{(j)}" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 305, + 367, + 313 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 305, + 367, + 313 + ], + "spans": [ + { + "bbox": [ + 308, + 305, + 367, + 313 + ], + "type": "text", + "content": "10: end if" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 316, + 361, + 324 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 316, + 361, + 324 + ], + "spans": [ + { + "bbox": [ + 308, + 316, + 361, + 324 + ], + "type": "text", + "content": "11: end for" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 326, + 372, + 337 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 326, + 372, + 337 + ], + "spans": [ + { + "bbox": [ + 308, + 326, + 372, + 337 + ], + "type": "text", + "content": "12: " + }, + { + "bbox": [ + 308, + 326, + 372, + 337 + ], + "type": "inline_equation", + "content": "j\\gets j + 1" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 338, + 361, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 338, + 361, + 346 + ], + "spans": [ + { + "bbox": [ + 308, + 338, + 361, + 346 + ], + "type": "text", + "content": "13: end while" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 370, + 424, + 381 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 370, + 424, + 381 + ], + "spans": [ + { + "bbox": [ + 306, + 370, + 424, + 381 + ], + "type": "text", + "content": "4.2. ALM for calibration" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 304, + 388, + 545, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 388, + 545, + 437 + ], + "spans": [ + { + "bbox": [ + 304, + 388, + 545, + 437 + ], + "type": "text", + "content": "Our goal now is to build an ALM method effective for calibration purposes. We can achieve this by reformulating problem (5) using a penalty function " + }, + { + "bbox": [ + 304, + 388, + 545, + 437 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 304, + 388, + 545, + 437 + ], + "type": "text", + "content": " parameterized by " + }, + { + "bbox": [ + 304, + 388, + 545, + 437 + ], + "type": "inline_equation", + "content": "(\\pmb {\\rho},\\pmb {\\lambda})\\in \\mathbb{R}_{++}^{K}\\times \\mathbb{R}_{++}^{K}" + }, + { + "bbox": [ + 304, + 388, + 545, + 437 + ], + "type": "text", + "content": " as follows:" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 312, + 447, + 545, + 479 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 447, + 545, + 479 + ], + "spans": [ + { + "bbox": [ + 312, + 447, + 545, + 479 + ], + "type": "interline_equation", + "content": "\\min _ {\\theta} \\quad \\sum_ {i = 1} ^ {N} \\mathcal {L} _ {\\mathrm {C E}} \\left(\\boldsymbol {x} ^ {(i)}, y ^ {(i)}\\right) + \\sum_ {k = 1} ^ {K} P \\left(d _ {k} ^ {(i)} - m, \\rho_ {k}, \\lambda_ {k}\\right) \\tag {9}", + "image_path": "4a257dbf221e5b7460425d192d71bea3321545ca1227cd72fe322cd2ed18cdf4.jpg" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 304, + 489, + 547, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 489, + 547, + 540 + ], + "spans": [ + { + "bbox": [ + 304, + 489, + 547, + 540 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 489, + 547, + 540 + ], + "type": "inline_equation", + "content": "d_k^{(i)} = \\max \\{l^{(i)}\\} - l_k^{(i)} \\in \\mathbb{R}_+" + }, + { + "bbox": [ + 304, + 489, + 547, + 540 + ], + "type": "text", + "content": ". With this formulation, it is natural to use a penalty-Lagrangian function for " + }, + { + "bbox": [ + 304, + 489, + 547, + 540 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 304, + 489, + 547, + 540 + ], + "type": "text", + "content": ". To avoid numerical issues typically associated with non-linear penalties, we normalize the constraints by the margin " + }, + { + "bbox": [ + 304, + 489, + 547, + 540 + ], + "type": "inline_equation", + "content": "m > 0" + }, + { + "bbox": [ + 304, + 489, + 547, + 540 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 363, + 549, + 545, + 575 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 363, + 549, + 545, + 575 + ], + "spans": [ + { + "bbox": [ + 363, + 549, + 545, + 575 + ], + "type": "interline_equation", + "content": "d _ {k} ^ {(i)} - m \\leq 0 \\Leftrightarrow \\frac {d _ {k} ^ {(i)}}{m} - 1 \\leq 0 \\tag {10}", + "image_path": "ffb7a0eadfd9e10a1edb2a74f59cb0b98e6d66eaa30c8df9cd831aeed3606b9a.jpg" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 304, + 578, + 545, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 578, + 545, + 639 + ], + "spans": [ + { + "bbox": [ + 304, + 578, + 545, + 639 + ], + "type": "text", + "content": "This leads to improved numerical stability for the ALM multiplier update as well. Additionally, we average the constraints instead of summing them. This makes the method independent of the number of classes, and eases the choice of initial penalty parameters " + }, + { + "bbox": [ + 304, + 578, + 545, + 639 + ], + "type": "inline_equation", + "content": "\\rho^{(0)}" + }, + { + "bbox": [ + 304, + 578, + 545, + 639 + ], + "type": "text", + "content": ". The resulting loss is:" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 314, + 647, + 545, + 681 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 647, + 545, + 681 + ], + "spans": [ + { + "bbox": [ + 314, + 647, + 545, + 681 + ], + "type": "interline_equation", + "content": "\\sum_ {i = 1} ^ {N} \\mathcal {L} _ {\\mathrm {C E}} \\left(\\boldsymbol {x} ^ {(i)}, y ^ {(i)}\\right) + \\frac {1}{K} \\sum_ {k = 1} ^ {K} P \\left(\\frac {d _ {k} ^ {(i)}}{m} - 1, \\rho_ {k}, \\lambda_ {k}\\right) \\tag {11}", + "image_path": "88e4d7102ad27fb7780fa7348a66d689fc40d77c612d001e1ccf0f64d5f1c4d0.jpg" + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 306, + 689, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 546, + 713 + ], + "type": "text", + "content": "As noted in Section 3, one of the main downsides of estimating Lagrange multipliers from the training set is that" + } + ] + } + ], + "index": 35 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16074" + } + ] + } + ], + "index": 36 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 167 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 167 + ], + "type": "text", + "content": "we could quickly overfit the data. Therefore, we propose to use the validation set to obtain a reliable estimate of the penalty multipliers at each epoch. We consider that an epoch of training corresponds to the approximate minimization of the loss function, and then compute the average penalty multiplier on the validation set. Formally after a training epoch " + }, + { + "bbox": [ + 46, + 72, + 289, + 167 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 46, + 72, + 289, + 167 + ], + "type": "text", + "content": ", the penalty multipliers for epoch " + }, + { + "bbox": [ + 46, + 72, + 289, + 167 + ], + "type": "inline_equation", + "content": "j + 1" + }, + { + "bbox": [ + 46, + 72, + 289, + 167 + ], + "type": "text", + "content": " will be, for all " + }, + { + "bbox": [ + 46, + 72, + 289, + 167 + ], + "type": "inline_equation", + "content": "k = 1, \\ldots, K" + }, + { + "bbox": [ + 46, + 72, + 289, + 167 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 61, + 175, + 288, + 206 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 175, + 288, + 206 + ], + "spans": [ + { + "bbox": [ + 61, + 175, + 288, + 206 + ], + "type": "interline_equation", + "content": "\\lambda_ {k} ^ {(j + 1)} = \\frac {1}{| \\mathcal {D} _ {\\mathrm {v a l}} |} \\sum_ {(\\boldsymbol {x}, y) \\in \\mathcal {D} _ {\\mathrm {v a l}}} P ^ {\\prime} \\left(\\frac {d _ {k}}{m} - 1, \\rho_ {k} ^ {(j)}, \\lambda_ {k} ^ {(j)}\\right) \\tag {12}", + "image_path": "17cedd248a46433158ce9662efce7ec932fbd7464f08f5a26c8550fd6934f90f.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 213, + 287, + 297 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 213, + 287, + 297 + ], + "spans": [ + { + "bbox": [ + 46, + 213, + 287, + 297 + ], + "type": "text", + "content": "Finally, the penalty multiplier is projected on a safeguarding interval " + }, + { + "bbox": [ + 46, + 213, + 287, + 297 + ], + "type": "inline_equation", + "content": "[\\lambda_{\\min}, \\lambda_{\\max}] = [10^{-6}, 10^{6}]" + }, + { + "bbox": [ + 46, + 213, + 287, + 297 + ], + "type": "text", + "content": " in our case. To update the penalty parameters " + }, + { + "bbox": [ + 46, + 213, + 287, + 297 + ], + "type": "inline_equation", + "content": "\\rho" + }, + { + "bbox": [ + 46, + 213, + 287, + 297 + ], + "type": "text", + "content": ", we compute the average constraint per class on the validation set. Then, for each class, if the average constraint is positive and has not decreased compared to the previous epoch, we multiply the corresponding penalty parameter by " + }, + { + "bbox": [ + 46, + 213, + 287, + 297 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 46, + 213, + 287, + 297 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 297, + 288, + 332 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 297, + 288, + 332 + ], + "spans": [ + { + "bbox": [ + 46, + 297, + 288, + 332 + ], + "type": "text", + "content": "Finally, as suggested by [2] and confirmed by our empirical results, we utilize the PHR function in our implementation, defined as follows:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 59, + 339, + 287, + 372 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 339, + 287, + 372 + ], + "spans": [ + { + "bbox": [ + 59, + 339, + 287, + 372 + ], + "type": "interline_equation", + "content": "\\operatorname {P H R} (z, \\rho , \\lambda) = \\left\\{ \\begin{array}{l l} \\lambda z + \\frac {1}{2} \\rho z ^ {2} & \\text {i f} \\quad \\lambda + \\rho z \\geq 0; \\\\ - \\frac {\\lambda^ {2}}{2 \\rho} & \\text {o t h e r w i s e .} \\end{array} \\right. \\tag {13}", + "image_path": "7f69361132c438f2ec9166a7e01cefd9ec46baa9d9442d255eeb3747cff4cc52.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 378, + 289, + 450 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 378, + 289, + 450 + ], + "spans": [ + { + "bbox": [ + 46, + 378, + 289, + 450 + ], + "type": "text", + "content": "Overall, the proposed method, consolidated in Algorithm 2, corresponds to approximately solving the constrained problem (2), by learning class-wise penalty multipliers on the validation set, to avoid overfitting and training specificities (i.e. batch normalization, dropout, augmentations) which would result in unreliable penalty multipliers estimates." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 460, + 128, + 474 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 460, + 128, + 474 + ], + "spans": [ + { + "bbox": [ + 47, + 460, + 128, + 474 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 479, + 164, + 493 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 479, + 164, + 493 + ], + "spans": [ + { + "bbox": [ + 47, + 479, + 164, + 493 + ], + "type": "text", + "content": "5.1. Experimental Setup" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 498, + 289, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 498, + 289, + 689 + ], + "spans": [ + { + "bbox": [ + 46, + 498, + 289, + 689 + ], + "type": "text", + "content": "Datasets. We perform experiments on a variety of popular benchmarks. First, we include three widely used image classification benchmarks, including Tiny-ImageNet [7], ImageNet [7] and one long-tailed image classification, ImageNet-LT [25]. Tiny-ImageNet is widely used in the calibration literature [22, 31], with relatively small " + }, + { + "bbox": [ + 46, + 498, + 289, + 689 + ], + "type": "inline_equation", + "content": "64 \\times 64" + }, + { + "bbox": [ + 46, + 498, + 289, + 689 + ], + "type": "text", + "content": " resolution, while ImageNet [7] is a large-scale benchmark consisting of 1000 categories and over 1M images. The main characteristic of ImageNet-LT is that the number of samples is extremely imbalanced across classes, ranging from 5 to 1280. To evaluate performance in dense prediction tasks, we include one semantic segmentation benchmark, PASCAL VOC2012 [9]. Furthermore, one benchmark from the NLP domain, 20 Newsgroups [19], is included to show the general applicability. For a detailed description of each dataset and the pre-processing settings, please refer to Appendix B." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 689, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 689, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 689, + 288, + 714 + ], + "type": "text", + "content": "Evaluation Metrics. For calibration, we report the most widely used Expected Calibration Error (ECE) [33]. Samples" + } + ] + } + ], + "index": 9 + }, + { + "type": "code", + "bbox": [ + 306, + 85, + 545, + 380 + ], + "blocks": [ + { + "bbox": [ + 306, + 72, + 447, + 84 + ], + "lines": [ + { + "bbox": [ + 306, + 72, + 447, + 84 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 447, + 84 + ], + "type": "text", + "content": "Algorithm 2 CALS-ALM training" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "code_caption" + }, + { + "bbox": [ + 306, + 85, + 545, + 380 + ], + "lines": [ + { + "bbox": [ + 306, + 85, + 545, + 380 + ], + "spans": [ + { + "bbox": [ + 306, + 85, + 545, + 380 + ], + "type": "text", + "content": "Require: DNN initial " + }, + { + "bbox": [ + 306, + 85, + 545, + 380 + ], + "type": "inline_equation", + "content": "\\theta^0" + }, + { + "bbox": [ + 306, + 85, + 545, + 380 + ], + "type": "text", + "content": " , margin m \nRequire: Dataset: " + }, + { + "bbox": [ + 306, + 85, + 545, + 380 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{\\mathrm{train}}" + }, + { + "bbox": [ + 306, + 85, + 545, + 380 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{\\mathrm{val}}" + }, + { + "bbox": [ + 306, + 85, + 545, + 380 + ], + "type": "text", + "content": " , batch size " + }, + { + "bbox": [ + 306, + 85, + 545, + 380 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 306, + 85, + 545, + 380 + ], + "type": "text", + "content": " \nRequire: Penalty function " + }, + { + "bbox": [ + 306, + 85, + 545, + 380 + ], + "type": "inline_equation", + "content": "P,\\gamma >1,\\tau \\in (0,1)" + }, + { + "bbox": [ + 306, + 85, + 545, + 380 + ], + "type": "text", + "content": " \nRequire: Initial " + }, + { + "bbox": [ + 306, + 85, + 545, + 380 + ], + "type": "inline_equation", + "content": "\\lambda^{(0)}\\in \\mathbb{R}_{++}^{n},\\pmb{\\rho}^{(0)}\\in \\mathbb{R}_{++}^{n}" + }, + { + "bbox": [ + 306, + 85, + 545, + 380 + ], + "type": "text", + "content": " 1: for " + }, + { + "bbox": [ + 306, + 85, + 545, + 380 + ], + "type": "inline_equation", + "content": "j = 0,\\dots ,T" + }, + { + "bbox": [ + 306, + 85, + 545, + 380 + ], + "type": "text", + "content": " do // Epochs of training \n2: for each mini-batch " + }, + { + "bbox": [ + 306, + 85, + 545, + 380 + ], + "type": "inline_equation", + "content": "\\{(x^{(i)},y^{(i)})\\}_{i = 1}^{B}" + }, + { + "bbox": [ + 306, + 85, + 545, + 380 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 306, + 85, + 545, + 380 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{\\mathrm{train}}" + }, + { + "bbox": [ + 306, + 85, + 545, + 380 + ], + "type": "text", + "content": " do \n3: " + }, + { + "bbox": [ + 306, + 85, + 545, + 380 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{c}} = \\sum_{i = 1}^{B}\\mathcal{L}_{\\mathrm{CE}}(\\boldsymbol{x}^{(i)},y^{(i)})" + }, + { + "bbox": [ + 306, + 85, + 545, + 380 + ], + "type": "text", + "content": " // Cross-entropy \n4: " + }, + { + "bbox": [ + 306, + 85, + 545, + 380 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{p}} = \\sum_{i = 1}^{B}\\frac{1}{K}\\sum_{k = 1}^{K}P\\Big(\\frac{d_k^{(i)}}{m} -1,\\rho_k^{(j)},\\lambda_k^{(j)}\\Big)" + }, + { + "bbox": [ + 306, + 85, + 545, + 380 + ], + "type": "text", + "content": " // Penalties \n5: " + }, + { + "bbox": [ + 306, + 85, + 545, + 380 + ], + "type": "inline_equation", + "content": "\\mathcal{L} = \\frac{1}{B} (\\mathcal{L}_{\\mathrm{c}} + \\mathcal{L}_{\\mathrm{p}})" + }, + { + "bbox": [ + 306, + 85, + 545, + 380 + ], + "type": "text", + "content": " \n6: " + }, + { + "bbox": [ + 306, + 85, + 545, + 380 + ], + "type": "inline_equation", + "content": "\\theta^{(t + 1)}\\gets \\theta^{(t)} - \\alpha \\cdot \\nabla_{\\theta}\\mathcal{L}" + }, + { + "bbox": [ + 306, + 85, + 545, + 380 + ], + "type": "text", + "content": " // Gradient descent \n7: end for \n8: for " + }, + { + "bbox": [ + 306, + 85, + 545, + 380 + ], + "type": "inline_equation", + "content": "k = 1,\\ldots ,K" + }, + { + "bbox": [ + 306, + 85, + 545, + 380 + ], + "type": "text", + "content": " do // Adjust " + }, + { + "bbox": [ + 306, + 85, + 545, + 380 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 306, + 85, + 545, + 380 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 306, + 85, + 545, + 380 + ], + "type": "inline_equation", + "content": "\\rho" + }, + { + "bbox": [ + 306, + 85, + 545, + 380 + ], + "type": "text", + "content": " on validation \n9: " + }, + { + "bbox": [ + 306, + 85, + 545, + 380 + ], + "type": "inline_equation", + "content": "\\lambda_{k}^{(j + 1)} = \\frac{1}{|\\mathcal{D}_{\\mathrm{val}}|}\\sum_{(\\boldsymbol {x},\\boldsymbol {y})\\in \\mathcal{D}_{\\mathrm{val}}}P'\\left(\\frac{d_k}{m} -1,\\rho_k^{(j)},\\lambda_k^{(j)}\\right)" + }, + { + "bbox": [ + 306, + 85, + 545, + 380 + ], + "type": "text", + "content": " \n10: " + }, + { + "bbox": [ + 306, + 85, + 545, + 380 + ], + "type": "inline_equation", + "content": "\\overline{d_k}^{(j)} = \\frac{1}{|\\mathcal{D}_{\\mathrm{val}}|}\\sum_{(\\boldsymbol {x},\\boldsymbol {y})\\in \\mathcal{D}_{\\mathrm{val}}}d_k - 1" + }, + { + "bbox": [ + 306, + 85, + 545, + 380 + ], + "type": "text", + "content": " // Average constraint \n11: if " + }, + { + "bbox": [ + 306, + 85, + 545, + 380 + ], + "type": "inline_equation", + "content": "j\\geq 1" + }, + { + "bbox": [ + 306, + 85, + 545, + 380 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 306, + 85, + 545, + 380 + ], + "type": "inline_equation", + "content": "\\overline{d_k}^{(j)} > \\tau \\max \\{0,\\overline{d_k}^{(j - 1)}\\}" + }, + { + "bbox": [ + 306, + 85, + 545, + 380 + ], + "type": "text", + "content": " then \n12: " + }, + { + "bbox": [ + 306, + 85, + 545, + 380 + ], + "type": "inline_equation", + "content": "\\rho_{k}^{(j + 1)}\\gets \\gamma \\rho_{k}^{(j)}" + }, + { + "bbox": [ + 306, + 85, + 545, + 380 + ], + "type": "text", + "content": " // Constraint has not improved \n13: else \n14: " + }, + { + "bbox": [ + 306, + 85, + 545, + 380 + ], + "type": "inline_equation", + "content": "\\rho_{k}^{(j + 1)}\\gets \\rho_{k}^{(j)}" + }, + { + "bbox": [ + 306, + 85, + 545, + 380 + ], + "type": "text", + "content": " \n15: end if \n16: end for \n17: end for" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "code_body" + } + ], + "index": 11, + "sub_type": "algorithm" + }, + { + "bbox": [ + 304, + 403, + 547, + 440 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 403, + 547, + 440 + ], + "spans": [ + { + "bbox": [ + 304, + 403, + 547, + 440 + ], + "type": "text", + "content": "are grouped into " + }, + { + "bbox": [ + 304, + 403, + 547, + 440 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 304, + 403, + 547, + 440 + ], + "type": "text", + "content": " equi-spaced bins according to prediction confidence, and a weighted average of the absolute difference between accuracy and confidence in each bin is calculated:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 359, + 457, + 545, + 491 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 359, + 457, + 545, + 491 + ], + "spans": [ + { + "bbox": [ + 359, + 457, + 545, + 491 + ], + "type": "interline_equation", + "content": "E C E = \\sum_ {m = 1} ^ {M} \\frac {\\left| B _ {m} \\right|}{N} \\left| A _ {m} - C _ {m} \\right|, \\tag {14}", + "image_path": "46b5018e916e7d5b3e72041ad95fe1d0bbf9f225d9ba52f25489cdf099fc2e16.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 498, + 547, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 498, + 547, + 641 + ], + "spans": [ + { + "bbox": [ + 304, + 498, + 547, + 641 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 498, + 547, + 641 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 304, + 498, + 547, + 641 + ], + "type": "text", + "content": " is the number of bins, " + }, + { + "bbox": [ + 304, + 498, + 547, + 641 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 498, + 547, + 641 + ], + "type": "text", + "content": " the amount of test samples, " + }, + { + "bbox": [ + 304, + 498, + 547, + 641 + ], + "type": "inline_equation", + "content": "B_{m}" + }, + { + "bbox": [ + 304, + 498, + 547, + 641 + ], + "type": "text", + "content": " the samples with prediction confidence in the " + }, + { + "bbox": [ + 304, + 498, + 547, + 641 + ], + "type": "inline_equation", + "content": "m^{th}" + }, + { + "bbox": [ + 304, + 498, + 547, + 641 + ], + "type": "text", + "content": " bin, " + }, + { + "bbox": [ + 304, + 498, + 547, + 641 + ], + "type": "inline_equation", + "content": "A_{m}" + }, + { + "bbox": [ + 304, + 498, + 547, + 641 + ], + "type": "text", + "content": " the accuracy and " + }, + { + "bbox": [ + 304, + 498, + 547, + 641 + ], + "type": "inline_equation", + "content": "C_{m}" + }, + { + "bbox": [ + 304, + 498, + 547, + 641 + ], + "type": "text", + "content": " the mean confidence of samples in the " + }, + { + "bbox": [ + 304, + 498, + 547, + 641 + ], + "type": "inline_equation", + "content": "m^{th}" + }, + { + "bbox": [ + 304, + 498, + 547, + 641 + ], + "type": "text", + "content": " bin. Note we fix " + }, + { + "bbox": [ + 304, + 498, + 547, + 641 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 498, + 547, + 641 + ], + "type": "text", + "content": " to 15 according to [22, 31]. In accordance with [22], we also report Adaptive ECE (AECE), a variant of ECE where the bins are configured to evenly distribute the test samples across them. Additionally, Classwise Calibration Error (CWCE) [29], a classwise extension of ECE, is included in Appendix C. For discriminative performance, we use standard measures: accuracy (Acc) for classification, and intersection over union (mIoU) for segmentation." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 642, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 642, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 642, + 547, + 714 + ], + "type": "text", + "content": "Compared methods. We compare our method to other learning based calibration losses, including (i) methods that impose constraints on predictions (either logits or softmax probabilities), i.e. Explicit Confidence Penalty (ECP) [38], Label Smoothing (LS) [42], Focal Loss (FL) [21] and its sample-dependent version (FLSD) [31], Margin-based Label" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16075" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 75, + 70, + 520, + 194 + ], + "blocks": [ + { + "bbox": [ + 75, + 70, + 520, + 194 + ], + "lines": [ + { + "bbox": [ + 75, + 70, + 520, + 194 + ], + "spans": [ + { + "bbox": [ + 75, + 70, + 520, + 194 + ], + "type": "table", + "html": "
MethodTinyImageNetImageNetImageNet-LT
ResNet-50ResNet-50SwinV2-TResNet-50SwinV2-T
AccECEAECEAccECEAECEAccECEAECEAccECEAECEAccECEAECE
CE65.023.733.6975.169.199.1875.609.959.9437.9028.1228.1231.8231.8236.68
MMCE [18]65.342.812.6174.858.578.5676.689.079.0837.7928.4128.4033.1426.4126.41
ECP [38]64.904.003.9275.228.278.2675.829.889.8637.6928.1428.1331.2233.7033.70
LS [42]65.783.173.1676.042.572.8875.427.327.3337.8810.4610.3831.7011.4211.40
FL [31]63.092.963.1273.871.601.6575.603.193.1836.0418.3718.3630.7325.5025.50
FLSD [31]64.092.912.9573.972.082.0674.702.442.3736.1817.7717.7832.5625.1625.17
CPC [5]64.494.884.9176.333.663.5976.345.505.3338.9016.0015.9932.5413.2113.19
MbLS [22]64.741.641.7375.824.444.2677.181.951.7338.326.166.1632.057.657.64
CALS-HR65.092.502.4276.345.635.6977.583.062.9538.502.832.7834.312.372.45
CALS-ALM65.031.541.3876.441.461.3277.101.611.6938.562.152.3033.942.322.45
", + "image_path": "c68ded947cb515613c0b2d9cbf5208d9b7b78442d8236cd16c164508913b4a10.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 201, + 547, + 224 + ], + "lines": [ + { + "bbox": [ + 46, + 201, + 547, + 224 + ], + "spans": [ + { + "bbox": [ + 46, + 201, + 547, + 224 + ], + "type": "text", + "content": "Table 1. Calibration performance for different approaches on three image classification benchmarks. We report two lower-is-better calibration metrics, i.e. ECE and AECE. Best method is highlighted in bold, while the second-best one is underlined." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 236, + 288, + 320 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 236, + 288, + 320 + ], + "spans": [ + { + "bbox": [ + 46, + 236, + 288, + 320 + ], + "type": "text", + "content": "Smoothing (MbLS) [22] and CPC [5], and (ii) techniques that directly optimize calibration measures, i.e. MMCE [18]. We refer to the related literature [22, 31] to set the hyperparameters for various methods. For instance, the smoothing factor in LS and FL is set to 0.05 and 3 respectively, and we set margin to 10 in MbLS. A detailed description of hyper-parameter values can be found in Appendix F." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 321, + 288, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 321, + 288, + 384 + ], + "spans": [ + { + "bbox": [ + 46, + 321, + 288, + 384 + ], + "type": "text", + "content": "Our methods. A simple alternative to the algorithm presented in Section 4.2 would be to heuristically tune multipliers by scaling them according to penalty values: if " + }, + { + "bbox": [ + 46, + 321, + 288, + 384 + ], + "type": "inline_equation", + "content": "P_{k}^{(j + 1)}" + }, + { + "bbox": [ + 46, + 321, + 288, + 384 + ], + "type": "text", + "content": " increases we also increase " + }, + { + "bbox": [ + 46, + 321, + 288, + 384 + ], + "type": "inline_equation", + "content": "\\lambda_{k}^{j + 1}" + }, + { + "bbox": [ + 46, + 321, + 288, + 384 + ], + "type": "text", + "content": " and vice versa. This strategy, akin to learning rate scheduling, can be formulated as:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 80, + 393, + 287, + 447 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 80, + 393, + 287, + 447 + ], + "spans": [ + { + "bbox": [ + 80, + 393, + 287, + 447 + ], + "type": "interline_equation", + "content": "\\lambda_ {k} ^ {(j + 1)} = \\left\\{ \\begin{array}{l l} \\mu \\lambda_ {k} ^ {(j)} & \\text {i f} P _ {k} ^ {(j + 1)} > \\tau P _ {k} ^ {(j)} \\\\ \\lambda_ {k} ^ {(j)} / \\mu & \\text {i f} P _ {k} ^ {(j)} > \\tau P _ {k} ^ {(j + 1)} \\\\ \\lambda_ {k} ^ {(j)} & \\text {o t h e r w i s e} \\end{array} \\right. \\tag {15}", + "image_path": "8a27c2d5623e1ff62d89db5d47de92545ef2a8c3af69399bb37a0c78937a7840.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 456, + 287, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 456, + 287, + 491 + ], + "spans": [ + { + "bbox": [ + 46, + 456, + 287, + 491 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 456, + 287, + 491 + ], + "type": "inline_equation", + "content": "\\mu > 1" + }, + { + "bbox": [ + 46, + 456, + 287, + 491 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 456, + 287, + 491 + ], + "type": "inline_equation", + "content": "\\tau > 1" + }, + { + "bbox": [ + 46, + 456, + 287, + 491 + ], + "type": "text", + "content": " are hyper-parameters that we fix to 1.1. We refer to our main algorithm as CALS-ALM and to this heuristic rule as CALS-HR in what follows." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 493, + 288, + 577 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 493, + 288, + 577 + ], + "spans": [ + { + "bbox": [ + 46, + 493, + 288, + 577 + ], + "type": "text", + "content": "We fix the margin to " + }, + { + "bbox": [ + 46, + 493, + 288, + 577 + ], + "type": "inline_equation", + "content": "m = 10" + }, + { + "bbox": [ + 46, + 493, + 288, + 577 + ], + "type": "text", + "content": " on vision tasks and " + }, + { + "bbox": [ + 46, + 493, + 288, + 577 + ], + "type": "inline_equation", + "content": "m = 6" + }, + { + "bbox": [ + 46, + 493, + 288, + 577 + ], + "type": "text", + "content": " on the NLP benchmark, as in [22], for a fair comparison. We also perform an ablation study to investigate the impact of the margin value. For other hyper-parameters, we set " + }, + { + "bbox": [ + 46, + 493, + 288, + 577 + ], + "type": "inline_equation", + "content": "\\pmb{\\lambda}^{(0)} = 10^{-6} \\cdot \\mathbf{1}_K" + }, + { + "bbox": [ + 46, + 493, + 288, + 577 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 493, + 288, + 577 + ], + "type": "inline_equation", + "content": "\\pmb{\\rho}^{(0)} = \\mathbf{1}_K" + }, + { + "bbox": [ + 46, + 493, + 288, + 577 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 493, + 288, + 577 + ], + "type": "inline_equation", + "content": "\\gamma = 1.2" + }, + { + "bbox": [ + 46, + 493, + 288, + 577 + ], + "type": "text", + "content": ", and we update the penalty parameters " + }, + { + "bbox": [ + 46, + 493, + 288, + 577 + ], + "type": "inline_equation", + "content": "\\pmb{\\rho}" + }, + { + "bbox": [ + 46, + 493, + 288, + 577 + ], + "type": "text", + "content": " every 10 epochs. Please refer to Appendix F for a detailed description of all hyper-parameters." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 578, + 288, + 661 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 578, + 288, + 661 + ], + "spans": [ + { + "bbox": [ + 46, + 578, + 288, + 661 + ], + "type": "text", + "content": "Implementation Details. For image classification, we experiment with ResNet [13] and a vision Transformer model, i.e. Swin Transformer V2 (SwinV2-T) [23]. DeepLabV3 [3] is employed for semantic segmentation on PASCAL VOC2012. Following [22, 31], we use the Global Pooling CNN (GPoolCNN) architecture [20] on the NLP recognition task. Further training details on each dataset can be found in Appendix B." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 670, + 105, + 681 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 670, + 105, + 681 + ], + "spans": [ + { + "bbox": [ + 47, + 670, + 105, + 681 + ], + "type": "text", + "content": "5.2. Results" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 689, + 289, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 689, + 289, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 689, + 289, + 713 + ], + "type": "text", + "content": "Results on image classification. Table 1 presents the discriminative and calibration performance of our methods" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 236, + 547, + 621 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 236, + 547, + 621 + ], + "spans": [ + { + "bbox": [ + 304, + 236, + 547, + 621 + ], + "type": "text", + "content": "across three widely used classification benchmarks, compared to baselines and related works. We can observe that our CALS-ALM approach consistently outperforms existing techniques in terms of calibration. Specifically, the results indicate that the standard CE loss and other approaches often lead to miscalibrated models, with the severity of miscalibration substantially increasing in correlation with dataset difficulty. This is particularly evident in large-scaled datasets with numerous classes, such as ImageNet, or those with long-tail class distributions, like ImageNet-LT. Although other learning-based methods could provide better calibrated networks, their performance is not stable across different settings. For example, while FL achieves a relatively low ECE of 1.60 with a Resnet50 trained on ImageNet, it only yields an ECE of 25.50 when training a SwinV2-T network on ImageNet-LT, revealing a limitation in adapting to different learning scenarios. In contrast, CALS-ALM attains the best calibration performance in all cases, frequently outperforming existing approaches by a significant margin, with minimal variations across datasets and architectures. This trends persists when compared to the most closely related technique, MbLS. It is noteworthy that the improvements on the long-tailed ImageNet-LT dataset are substantial; for example, we decrese ECE from 28.12 to 2.15 for ResNet50, and from 31.82 to 2.32 for SwinV2-T, validating the effectiveness of class adaptive learning. Another interesting finding is that CALS-HR, employing a naive update strategy for class-wise penalty weights, achieves nearly the second-best performance in the ImageNet-LT dataset. This further demonstrates the effectiveness of CALS for learning under class-imbalance scenarios. For the reliability diagrams of various models, please refer to Appendix E." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 630, + 548, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 630, + 548, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 630, + 548, + 715 + ], + "type": "text", + "content": "In terms of model accuracy, our method delivers competitive performances, surpassing existing methods in certain cases. It is important to emphasize that, while the proposed method achieves discriminative results comparable to the best-performing approach for each dataset, the differences in calibration are considerable. This highlights the superiority of the proposed formulation for training highly discrimi" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "16076" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 72, + 218, + 213 + ], + "blocks": [ + { + "bbox": [ + 50, + 72, + 218, + 213 + ], + "lines": [ + { + "bbox": [ + 50, + 72, + 218, + 213 + ], + "spans": [ + { + "bbox": [ + 50, + 72, + 218, + 213 + ], + "type": "image", + "image_path": "62e78eca5eb382b7850f62eb708db706f17b5aa1ad3f60232d49924890d793ba.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 52, + 217, + 217, + 228 + ], + "lines": [ + { + "bbox": [ + 52, + 217, + 217, + 228 + ], + "spans": [ + { + "bbox": [ + 52, + 217, + 217, + 228 + ], + "type": "text", + "content": "(a) Validation ECE and multipliers during training." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 238, + 72, + 397, + 213 + ], + "blocks": [ + { + "bbox": [ + 238, + 72, + 397, + 213 + ], + "lines": [ + { + "bbox": [ + 238, + 72, + 397, + 213 + ], + "spans": [ + { + "bbox": [ + 238, + 72, + 397, + 213 + ], + "type": "image", + "image_path": "c862e61f3d4cd7c3e9ee5bc0b8e22c20dcae1e6a6693b1aad89dba89c032ca97.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 237, + 547, + 270 + ], + "lines": [ + { + "bbox": [ + 46, + 237, + 547, + 270 + ], + "spans": [ + { + "bbox": [ + 46, + 237, + 547, + 270 + ], + "type": "text", + "content": "Figure 3. Ablation study on ImageNet-LT. (a) Evolution of ECE on validation and multipliers for CALS: ECE on the validation set for our method (CALS), CE and MbLS [22] and values of multipliers " + }, + { + "bbox": [ + 46, + 237, + 547, + 270 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 46, + 237, + 547, + 270 + ], + "type": "text", + "content": " for CALS after each training epoch. (b) Effect of penalty functions and margin: ECE and accuracy on validation and test set are shown across different choices of penalty functions and margin values." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 401, + 72, + 541, + 213 + ], + "blocks": [ + { + "bbox": [ + 271, + 217, + 511, + 228 + ], + "lines": [ + { + "bbox": [ + 271, + 217, + 511, + 228 + ], + "spans": [ + { + "bbox": [ + 271, + 217, + 511, + 228 + ], + "type": "text", + "content": "(b) Effect of penalty function and margin on accuracy and ECE for CALS." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 401, + 72, + 541, + 213 + ], + "lines": [ + { + "bbox": [ + 401, + 72, + 541, + 213 + ], + "spans": [ + { + "bbox": [ + 401, + 72, + 541, + 213 + ], + "type": "image", + "image_path": "213ff12f720a013ab09eb9be14846ad88d70b6a413e2bbbbcd3a08b1a4fbca46.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 291, + 288, + 339 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 291, + 288, + 339 + ], + "spans": [ + { + "bbox": [ + 46, + 291, + 288, + 339 + ], + "type": "text", + "content": "native and well-calibrated networks. Figure 1 provides a more visual comparison considering both accuracy and ECE. It is demonstrated that CALS-ALM provides the optimal compromise between accuracy and calibration performance." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 340, + 287, + 519 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 340, + 287, + 519 + ], + "spans": [ + { + "bbox": [ + 46, + 340, + 287, + 519 + ], + "type": "text", + "content": "Ablation Analysis. Figure 3 illustrates the evolution of ECE in (a) and penalty multipliers " + }, + { + "bbox": [ + 46, + 340, + 287, + 519 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 46, + 340, + 287, + 519 + ], + "type": "text", + "content": " in (b) during training. It is interesting to observe that the evolution of " + }, + { + "bbox": [ + 46, + 340, + 287, + 519 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 46, + 340, + 287, + 519 + ], + "type": "text", + "content": " is consistent with the ECE. Specifically, the average penalty weight gradually increases while the ECE initially deteriorates because the model is focused on increasing accuracy. However, the value of the penalty weight begins to decline alongside the ECE, as the network starts to become better calibrated. For a visualization of classwise multipliers, please refer to Appendix D. Figure 3(c) highlights the impact of the choice of penalty functions and margin values. This demonstrates that the PHR penalty function is preferable over the other two options, P2 and P3, for both calibration and accuracy. Regarding the margin, the best performance is achieved with " + }, + { + "bbox": [ + 46, + 340, + 287, + 519 + ], + "type": "inline_equation", + "content": "m \\approx 10" + }, + { + "bbox": [ + 46, + 340, + 287, + 519 + ], + "type": "text", + "content": ", which is consistent with the findings in [22]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 521, + 287, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 521, + 287, + 652 + ], + "spans": [ + { + "bbox": [ + 46, + 521, + 287, + 652 + ], + "type": "text", + "content": "Semantic Segmentation. Table 2 presents the performance on Pascal VOC dataset. Note here CALS refers to our best method, i.e. CALS-ALM. It can be observed that the trend is consistent with image classification experiments: CALS outperforms counterparts in terms of ECE, and yields competitive results on discriminative performance, i.e. mIoU in segmentation. It is worth noting that some methods like MMCE, FLSD and CPC, are not included here because their computation demands were too heavy for pixel-wise segmentation tasks. In contrast, our method is unlimited in dense prediction tasks as the computation cost it adds is moderate." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 653, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 288, + 713 + ], + "type": "text", + "content": "Text Classification. Last, we demonstrate the general applicability of the proposed method by analyzing its performance on a non-vision task, i.e. text classification on 20 Newsgroups dataset. The results are reported in Table 3. Remarkably, CALs again brings substantial improvement in" + } + ] + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 332, + 289, + 520, + 328 + ], + "blocks": [ + { + "bbox": [ + 332, + 289, + 520, + 328 + ], + "lines": [ + { + "bbox": [ + 332, + 289, + 520, + 328 + ], + "spans": [ + { + "bbox": [ + 332, + 289, + 520, + 328 + ], + "type": "table", + "html": "
CEECPLSFLMbLSCALS
ECE14.755.156.9010.875.224.66
mIoU66.4665.5767.7364.2565.2966.77
", + "image_path": "088b47b9b1de7d521cf3fbbc37275a9b2790c2cccd7a36ee11eb3a0157c92045.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 308, + 357, + 544, + 396 + ], + "blocks": [ + { + "bbox": [ + 319, + 336, + 531, + 348 + ], + "lines": [ + { + "bbox": [ + 319, + 336, + 531, + 348 + ], + "spans": [ + { + "bbox": [ + 319, + 336, + 531, + 348 + ], + "type": "text", + "content": "Table 2. Segmentation results on the PASCAL VOC 2012." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 308, + 357, + 544, + 396 + ], + "lines": [ + { + "bbox": [ + 308, + 357, + 544, + 396 + ], + "spans": [ + { + "bbox": [ + 308, + 357, + 544, + 396 + ], + "type": "table", + "html": "
CEMMCEECPLSFLFLSDCPCMbLSCALS
ECE (%)22.7523.0222.978.0710.8010.879.465.402.04
Acc (%)67.0166.2366.4867.1466.0865.8568.2767.8968.32
", + "image_path": "48dc62d79eb07fc70e53914917a3c3c087b38fe4e2c9697d4448621aa38b64f9.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_body" + } + ], + "index": 12 + }, + { + "bbox": [ + 310, + 403, + 539, + 415 + ], + "lines": [ + { + "bbox": [ + 310, + 403, + 539, + 415 + ], + "spans": [ + { + "bbox": [ + 310, + 403, + 539, + 415 + ], + "type": "text", + "content": "Table 3. Results on the text classification task, 20 Newsgroups." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 426, + 547, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 426, + 547, + 486 + ], + "spans": [ + { + "bbox": [ + 304, + 426, + 547, + 486 + ], + "type": "text", + "content": "terms of calibration, with ECE decreasing to " + }, + { + "bbox": [ + 304, + 426, + 547, + 486 + ], + "type": "inline_equation", + "content": "2.04\\%" + }, + { + "bbox": [ + 304, + 426, + 547, + 486 + ], + "type": "text", + "content": ", while yielding the best accuracy " + }, + { + "bbox": [ + 304, + 426, + 547, + 486 + ], + "type": "inline_equation", + "content": "68.32\\%" + }, + { + "bbox": [ + 304, + 426, + 547, + 486 + ], + "type": "text", + "content": ". This reveals that the proposed class adaptive learning method is also able to handle class differences in NLP applications and provide promising performance in terms of both accuracy and calibration." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 495, + 472, + 508 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 495, + 472, + 508 + ], + "spans": [ + { + "bbox": [ + 306, + 495, + 472, + 508 + ], + "type": "text", + "content": "6. Limitations and Future Work" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 516, + 547, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 516, + 547, + 624 + ], + "spans": [ + { + "bbox": [ + 304, + 516, + 547, + 624 + ], + "type": "text", + "content": "We have proposed Class Adaptive Label Smoothing for network calibration based on a modified Augmented Lagrangian Multiplier algorithm. Despite its superior performance over previous methods, there are potential limitations in this work. For instance, our method requires the validation set to have the same distribution as the training set. Although this is satisfied in nearly every benchmark, it will be interesting to investigate the impact of using non-independent and identically distributed (i.i.d) validation sets." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 634, + 408, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 634, + 408, + 647 + ], + "spans": [ + { + "bbox": [ + 306, + 634, + 408, + 647 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 653, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 653, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 653, + 547, + 713 + ], + "type": "text", + "content": "This work is supported by National Science and Engineering Research Council of Canada (NSERC) and Prompt Quebec. We also thank Calcul Quebec and Compute Canada. Adrian Galdran is supported by a Marie Sklodowska-Curie Fellowship (No 892297)." + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16077" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 123 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 123 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 123 + ], + "type": "text", + "content": "[1] Dimitri P. Bertsekas. Constrained Optimization and Lagrange Multiplier Methods (Optimization and Neural Computation Series). Athena Scientific, 1 edition, 1996. 4" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 124, + 288, + 167 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 124, + 288, + 167 + ], + "spans": [ + { + "bbox": [ + 53, + 124, + 288, + 167 + ], + "type": "text", + "content": "[2] Ernesto G Birgin, Romulo A Castillo, and José Mario Martínez. Numerical comparison of augmented lagrangian algorithms for nonconvex problems. Computational Optimization and Applications, 31(1), 2005. 5, 6, 11" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 168, + 287, + 201 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 168, + 287, + 201 + ], + "spans": [ + { + "bbox": [ + 53, + 168, + 287, + 201 + ], + "type": "text", + "content": "[3] Liang-Chieh Chen, George Papandreou, Florian Schroff, and Hartwig Adam. Rethinking atrous convolution for semantic image segmentation. In CVPR, 2017. 1, 7, 11" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 202, + 288, + 235 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 202, + 288, + 235 + ], + "spans": [ + { + "bbox": [ + 53, + 202, + 288, + 235 + ], + "type": "text", + "content": "[4] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In ICML, 2020. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 236, + 288, + 267 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 236, + 288, + 267 + ], + "spans": [ + { + "bbox": [ + 53, + 236, + 288, + 267 + ], + "type": "text", + "content": "[5] Jiacheng Cheng and Nuno Vasconcelos. Calibrating deep neural networks by pairwise constraints. In CVPR, 2022. 3, 7, 13" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 269, + 288, + 312 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 269, + 288, + 312 + ], + "spans": [ + { + "bbox": [ + 53, + 269, + 288, + 312 + ], + "type": "text", + "content": "[6] Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The cityscapes dataset for semantic urban scene understanding. In CVPR, June 2016. 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 313, + 288, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 313, + 288, + 346 + ], + "spans": [ + { + "bbox": [ + 53, + 313, + 288, + 346 + ], + "type": "text", + "content": "[7] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In CVPR, 2009. 6, 11" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 346, + 288, + 379 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 346, + 288, + 379 + ], + "spans": [ + { + "bbox": [ + 53, + 346, + 288, + 379 + ], + "type": "text", + "content": "[8] Zhipeng Ding, Xu Han, Peirong Liu, and Marc Niethammer. Local temperature scaling for probability calibration. In ICCV, 2021. 1, 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 380, + 288, + 423 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 380, + 288, + 423 + ], + "spans": [ + { + "bbox": [ + 53, + 380, + 288, + 423 + ], + "type": "text", + "content": "[9] Mark Everingham, S. M. Eslami, Luc Gool, Christopher K. Williams, John Winn, and Andrew Zisserman. The pascal visual object classes challenge: A retrospective. *IJCV*, 111(1):98–136, 2015. 4, 6, 11" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 425, + 288, + 457 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 425, + 288, + 457 + ], + "spans": [ + { + "bbox": [ + 48, + 425, + 288, + 457 + ], + "type": "text", + "content": "[10] Yarin Gal and Zoubin Ghahramani. Dropout as a bayesian approximation: Representing model uncertainty in deep learning. In ICML, 2016. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 458, + 288, + 490 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 458, + 288, + 490 + ], + "spans": [ + { + "bbox": [ + 48, + 458, + 288, + 490 + ], + "type": "text", + "content": "[11] Chuan Guo, Geoff Pleiss, Yu Sun, and Kilian Q Weinberger. On calibration of modern neural networks. In ICML, 2017. 1, 2, 3, 12" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 491, + 288, + 523 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 491, + 288, + 523 + ], + "spans": [ + { + "bbox": [ + 48, + 491, + 288, + 523 + ], + "type": "text", + "content": "[12] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollar, and Ross Girshick. Masked autoencoders are scalable vision learners. In CVPR, 2022. 1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 525, + 288, + 556 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 525, + 288, + 556 + ], + "spans": [ + { + "bbox": [ + 48, + 525, + 288, + 556 + ], + "type": "text", + "content": "[13] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, 2016. 1, 7, 11" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 558, + 288, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 558, + 288, + 590 + ], + "spans": [ + { + "bbox": [ + 48, + 558, + 288, + 590 + ], + "type": "text", + "content": "[14] Magnus R. Hestenes. Multiplier and gradient methods. Journal of Optimization Theory and Applications, 4:303-320, 1969. 5" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 591, + 288, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 591, + 288, + 624 + ], + "spans": [ + { + "bbox": [ + 48, + 591, + 288, + 624 + ], + "type": "text", + "content": "[15] Sergey Ioffe and Christian Szegedy. Batch normalization: Accelerating deep network training by reducing internal covariate shift. In ICML, 2015. 1, 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 48, + 624, + 288, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 624, + 288, + 668 + ], + "spans": [ + { + "bbox": [ + 48, + 624, + 288, + 668 + ], + "type": "text", + "content": "[16] Archit Karandikar, Nicholas Cain, Dustin Tran, Balaji Lakshminarayanan, Jonathon Shlens, Michael C Mozer, and Rebecca Roelofs. Soft calibration objectives for neural networks. In NeurIPS, 2021. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 48, + 669, + 288, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 669, + 288, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 669, + 288, + 712 + ], + "type": "text", + "content": "[17] Barry W. Kort and Dimitri P. Bertsekas. Combined primal-dual and penalty methods for convex programming. SIAM Journal on Control and Optimization, 14(2):268-294, 1976. 5" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 38, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "text", + "content": "[18] Aviral Kumar, Sunita Sarawagi, and Ujjwal Jain. Trainable calibration measures for neural networks from kernel mean embeddings. In ICML, 2018. 3, 7, 13" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 107, + 545, + 126 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 107, + 545, + 126 + ], + "spans": [ + { + "bbox": [ + 307, + 107, + 545, + 126 + ], + "type": "text", + "content": "[19] Ken Lang. Newsweeder: Learning to filter netnews. In ICML, 1995. 6, 12" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 128, + 545, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 128, + 545, + 149 + ], + "spans": [ + { + "bbox": [ + 307, + 128, + 545, + 149 + ], + "type": "text", + "content": "[20] Min Lin, Qiang Chen, and Shuicheng Yan. Network in network. In ICLR, 2014. 7, 12" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 151, + 545, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 151, + 545, + 183 + ], + "spans": [ + { + "bbox": [ + 307, + 151, + 545, + 183 + ], + "type": "text", + "content": "[21] Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dólar. Focal loss for dense object detection. In CVPR, 2017. 1, 3, 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 184, + 545, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 184, + 545, + 226 + ], + "spans": [ + { + "bbox": [ + 307, + 184, + 545, + 226 + ], + "type": "text", + "content": "[22] Bingyuan Liu, Ismail Ben Ayed, Adrian Galdran, and Jose Dolz. The devil is in the margin: Margin-based label smoothing for network calibration. In CVPR, 2022. 1, 2, 3, 6, 7, 8, 11, 12, 13" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 228, + 545, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 228, + 545, + 271 + ], + "spans": [ + { + "bbox": [ + 307, + 228, + 545, + 271 + ], + "type": "text", + "content": "[23] Ze Liu, Han Hu, Yutong Lin, Zhuliang Yao, Zhenda Xie, Yixuan Wei, Jia Ning, Yue Cao, Zheng Zhang, Li Dong, Furu Wei, and Baining Guo. Swin transformer v2: Scaling up capacity and resolution. In CVPR, 2022. 7, 11" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 272, + 545, + 314 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 272, + 545, + 314 + ], + "spans": [ + { + "bbox": [ + 307, + 272, + 545, + 314 + ], + "type": "text", + "content": "[24] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In ICCV, 2021. 1" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 316, + 545, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 316, + 545, + 348 + ], + "spans": [ + { + "bbox": [ + 307, + 316, + 545, + 348 + ], + "type": "text", + "content": "[25] Ziwei Liu, Zhongqi Miao, Xiaohang Zhan, Jiayun Wang, Boqing Gong, and Stella X. Yu. Large-scale long-tailed recognition in an open world. In CVPR, 2019. 6, 11" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 349, + 545, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 349, + 545, + 370 + ], + "spans": [ + { + "bbox": [ + 307, + 349, + 545, + 370 + ], + "type": "text", + "content": "[26] Ilya Loshchilov and Frank Hutter. SGDR: Stochastic gradient descent with warm restarts. In *ICLR*, 2017. 11" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 371, + 545, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 371, + 545, + 392 + ], + "spans": [ + { + "bbox": [ + 307, + 371, + 545, + 392 + ], + "type": "text", + "content": "[27] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In *ICLR*, 2019. 1, 11" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 393, + 545, + 424 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 393, + 545, + 424 + ], + "spans": [ + { + "bbox": [ + 307, + 393, + 545, + 424 + ], + "type": "text", + "content": "[28] Xingchen Ma and Matthew B. Blaschko. Meta-cal: Well-controlled post-hoc calibration by ranking. In ICML, 2021. 3" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 307, + 426, + 545, + 458 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 426, + 545, + 458 + ], + "spans": [ + { + "bbox": [ + 307, + 426, + 545, + 458 + ], + "type": "text", + "content": "[29] Lena Maier-Hein et al. Metrics reloaded: Pitfalls and recommendations for image analysis validation. arXiv preprint arXiv:2206.01653, 2022. 6, 12" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 307, + 460, + 545, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 460, + 545, + 502 + ], + "spans": [ + { + "bbox": [ + 307, + 460, + 545, + 502 + ], + "type": "text", + "content": "[30] Matthias Minderer, Josip Djolonga, Rob Romijnders, Frances Hubis, Xiaohua Zhai, Neil Houlsby, Dustin Tran, and Mario Lucic. Revisiting the calibration of modern neural networks. In NeurIPS, 2021. 3, 11, 12" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 307, + 504, + 545, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 504, + 545, + 547 + ], + "spans": [ + { + "bbox": [ + 307, + 504, + 545, + 547 + ], + "type": "text", + "content": "[31] Jishnu Mukhoti, Viveka Kulharia, Amartya Sanyal, Stuart Golodetz, Philip HS Torr, and Puneet K Dokania. Calibrating deep neural networks using focal loss. In NeurIPS, 2020. 1, 2, 3, 6, 7, 11, 12, 13" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 307, + 548, + 545, + 570 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 548, + 545, + 570 + ], + "spans": [ + { + "bbox": [ + 307, + 548, + 545, + 570 + ], + "type": "text", + "content": "[32] Rafael Müller, Simon Kornblith, and Geoffrey Hinton. When does label smoothing help? In NeurIPS, 2019. 1, 3" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 307, + 571, + 545, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 571, + 545, + 602 + ], + "spans": [ + { + "bbox": [ + 307, + 571, + 545, + 602 + ], + "type": "text", + "content": "[33] Mahdi Pakdaman Naeini, Gregory F. Cooper, and Milos Hauskrecht. Obtaining well calibrated probabilities using bayesian binning. In AAAI, 2015. 6" + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 307, + 603, + 545, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 603, + 545, + 635 + ], + "spans": [ + { + "bbox": [ + 307, + 603, + 545, + 635 + ], + "type": "text", + "content": "[34] H. Nakayama, H. Samaya, and H. Samaya. A generalized lagrangian function and multiplier method. Journal of Optimization Theory and Applications, 17:211-227, 1975. 5" + } + ] + } + ], + "index": 35 + }, + { + "bbox": [ + 307, + 636, + 545, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 636, + 545, + 658 + ], + "spans": [ + { + "bbox": [ + 307, + 636, + 545, + 658 + ], + "type": "text", + "content": "[35] Jorge Nocedal and Stephen J. Wright. Numerical Optimization. Springer, New York, NY, USA, 2 edition, 2006. 4" + } + ] + } + ], + "index": 36 + }, + { + "bbox": [ + 307, + 658, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 658, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 307, + 658, + 545, + 712 + ], + "type": "text", + "content": "[36] Yaniv Ovadia, Emily Fertig, Jie Ren, Zachary Nado, David Sculley, Sebastian Nowozin, Joshua V Dillon, Balaji Lakshminarayanan, and Jasper Snoek. Can you trust your model's uncertainty? evaluating predictive uncertainty under dataset shift. In NeurIPS, 2019. 3" + } + ] + } + ], + "index": 37 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "16078" + } + ] + } + ], + "index": 39 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 387 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 105 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 105 + ], + "type": "text", + "content": "[37] Jeffrey Pennington, Richard Socher, and Christopher D Manning. Glove: Global vectors for word representation. In EMNLP, 2014. 12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 107, + 288, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 107, + 288, + 150 + ], + "spans": [ + { + "bbox": [ + 49, + 107, + 288, + 150 + ], + "type": "text", + "content": "[38] Gabriel Pereyra, George Tucker, Jan Chorowski, Lukasz Kaiser, and Geoffrey Hinton. Regularizing neural networks by penalizing confident output distributions. In ICLR, 2017. 1, 3, 6, 7, 13" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 152, + 288, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 152, + 288, + 184 + ], + "spans": [ + { + "bbox": [ + 49, + 152, + 288, + 184 + ], + "type": "text", + "content": "[39] M. J. D. Powell. A method for nonlinear constraints in minimization problems. In *Fletcher, R., Ed., Optimization, Academic Press, New York, NY*, 1969. 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 186, + 288, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 186, + 288, + 217 + ], + "spans": [ + { + "bbox": [ + 49, + 186, + 288, + 217 + ], + "type": "text", + "content": "[40] Jérôme Rony, Eric Granger, Marco Pedersoli, and Ismail Ben Ayed. Augmented lagrangian adversarial attacks. In ICCV, 2021. 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 49, + 220, + 288, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 220, + 288, + 262 + ], + "spans": [ + { + "bbox": [ + 49, + 220, + 288, + 262 + ], + "type": "text", + "content": "[41] Sara Sangalli, Ertunc Erdil, Andeas Hotker, Olivio F Donati, and Ender Konukoglu. Constrained optimization to train neural networks on critical and under-represented classes. In NeurIPS, 2021. 4, 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 49, + 264, + 288, + 306 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 264, + 288, + 306 + ], + "spans": [ + { + "bbox": [ + 49, + 264, + 288, + 306 + ], + "type": "text", + "content": "[42] Christian Szegedy, Vincent Vanhoucke, Sergey Ioffe, Jon Shlens, and Zbigniew Wojna. Rethinking the inception architecture for computer vision. In CVPR, 2016. 3, 6, 7, 13" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 49, + 309, + 288, + 351 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 309, + 288, + 351 + ], + "spans": [ + { + "bbox": [ + 49, + 309, + 288, + 351 + ], + "type": "text", + "content": "[43] Christian Tomani, Sebastian Gruber, Muhammed Ebrar Erdem, Daniel Cremers, and Florian Buettner. Post-hoc uncertainty calibration for domain drift scenarios. In CVPR, 2021. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 49, + 354, + 288, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 354, + 288, + 387 + ], + "spans": [ + { + "bbox": [ + 49, + 354, + 288, + 387 + ], + "type": "text", + "content": "[44] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In NeurIPS, 2017. 1" + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "16079" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Class Attention Transfer Based Knowledge Distillation/a68b12fe-6c62-4a74-9758-0370df04b2ad_content_list.json b/2023/Class Attention Transfer Based Knowledge Distillation/a68b12fe-6c62-4a74-9758-0370df04b2ad_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..c5b5f3fa3810a7582640a926d4a7c2a8350a264d --- /dev/null +++ b/2023/Class Attention Transfer Based Knowledge Distillation/a68b12fe-6c62-4a74-9758-0370df04b2ad_content_list.json @@ -0,0 +1,1586 @@ +[ + { + "type": "text", + "text": "Class Attention Transfer Based Knowledge Distillation", + "text_level": 1, + "bbox": [ + 207, + 130, + 759, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Ziyao Guo $^{1}$ , Haonan Yan $^{1,2,*}$ , Hui Li $^{1,*}$ , Xiaodong Lin $^{2}$ $^{1}$ Xidian University, $^{2}$ University of Guelph", + "bbox": [ + 267, + 179, + 699, + 217 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "gzyaftermath@outlook.com", + "bbox": [ + 374, + 219, + 589, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 268, + 313, + 284 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Previous knowledge distillation methods have shown their impressive performance on model compression tasks, however, it is hard to explain how the knowledge they transferred helps to improve the performance of the student network. In this work, we focus on proposing a knowledge distillation method that has both high interpretability and competitive performance. We first revisit the structure of mainstream CNN models and reveal that possessing the capacity of identifying class discriminative regions of input is critical for CNN to perform classification. Furthermore, we demonstrate that this capacity can be obtained and enhanced by transferring class activation maps. Based on our findings, we propose class attention transfer based knowledge distillation (CAT-KD). Different from previous KD methods, we explore and present several properties of the knowledge transferred by our method, which not only improve the interpretability of CAT-KD but also contribute to a better understanding of CNN. While having high interpretability, CAT-KD achieves state-of-the-art performance on multiple benchmarks. Code is available at: https://github.com/GzyAftermath/CAT-KD.", + "bbox": [ + 75, + 301, + 473, + 619 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 650, + 209, + 666 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Knowledge distillation (KD) transfers knowledge distilled from the bigger teacher network to the smaller student network, aiming to improve the performance of the student network. Depending on the type of the transferred knowledge, previous KD methods can be divided into three categories: based on transferring logits [3, 6, 11, 16, 33], features [2, 10, 17-19, 23, 24, 28], and attention [29]. Although KD methods that are based on transferring logits and features have shown their promising performance [2, 33], it is hard to explain how the knowledge they transferred helps to improve the performance of the student network, due to the uninterpretability of logits and features. Relatively, the principle of attention-based KD methods is more intuitive:", + "bbox": [ + 75, + 676, + 468, + 875 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/31922e3db0993b7355024d0fc3ec40f96a6e3808c2348f59e5eb0e742ad2e2d5.jpg", + "image_caption": [ + "Figure 1. Illustration of the converted structure. After converting the FC layer into a convolutional layer with $1 \\times 1$ kernel and moving the position of the global average pooling layer, CAMs can be obtained during the forward propagation." + ], + "image_footnote": [], + "bbox": [ + 509, + 267, + 883, + 575 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "it aims at telling the student network which part of the input should it focus on during the classification, which is realized by forcing the student network to mimic the transferred attention maps during training. However, though previous work AT [29] has validated the effectiveness of transferring attention, it does not present what role attention plays during the classification. This makes it hard to explain why telling the trained model where should it focus could improve its performance on the classification mission. Besides, the performance of the previous attention-based KD method [29] is less competitive compared with the methods that are based on transferring logits and features [2, 33]. In this work, we focus on proposing an attention-based KD method that has higher interpretability and better performance.", + "bbox": [ + 496, + 672, + 893, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author", + "bbox": [ + 94, + 886, + 220, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "11868", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/99ad99410587f8d2adc94e067252483f53781f197df3f22eaa26bbb23742dc9a.jpg", + "image_caption": [ + "Figure 2. Visualization of CAMs corresponding to categories with Top 4 prediction scores for the given image. The predicted categories and their scores are reported in the picture." + ], + "image_footnote": [], + "bbox": [ + 78, + 88, + 238, + 181 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/bf4fc73385716d74922ecbed962a8fb5843852c6822c3c2c521099929f37fe90.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 241, + 88, + 401, + 181 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/45caf93496e785cc702fbe528421ea023f2d02ca15a5f732fe9b28068900f7bc.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 405, + 88, + 565, + 181 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/0ddae377e3ec9301f1704fb9d4de7cb24811433c83c1ea10d29270b1856b8ac1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 566, + 88, + 728, + 181 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/24231d24833314bc016b0d5355521e61a0f3d7342009bed46403c761b82b4db8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 730, + 88, + 890, + 181 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We start our work by exploring what role attention plays during classification. After revisiting the structure of the mainstream models, we find that with a little conversion (illustrated in Figure 1), class activation map (CAM) [34], a kind of class attention map which indicates the discriminative regions of input for a specific category, can be obtained during the classification. Without changing the parameters and outputs, the classification process of the converted model can be viewed in two steps: (1) the model exploits its capacity to identify class discriminative regions of input and generate CAM for each category contained in the classification mission, (2) the model outputs the prediction score of each category by computing the average activation of the corresponding CAM. Considering that the converted model makes predictions by simply comparing the average activation of CAMs, possessing the capacity to identify class discriminative regions of input is critical for CNN to perform classification. The question is: can we enhance this capacity by offering hints about class discriminative regions of input during training? To answer this question, we propose class attention transfer (CAT).", + "bbox": [ + 75, + 248, + 472, + 565 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "During CAT, the trained model is not required to predict the category of input, it is only forced to mimic the transferred CAMs, which are normalized to ensure they only contain hints about class discriminative regions of input. Through experiments with CAT, we reveal that transferring only CAMs can train a model with high accuracy on the classification task, reflecting the trained model obtains the capacity to identify class discriminative regions of input. Besides, the performance of the trained model is influenced by the accuracy of the model offering the transferred CAMs. This further demonstrates that the capacity of identifying class discriminative regions can be enhanced by transferring more precise CAMs.", + "bbox": [ + 75, + 566, + 468, + 763 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Based on our findings, we propose class attention transfer based knowledge distillation (CAT-KD), aiming to enable the student network to achieve better performance by improving its capacity of identifying class discriminative regions. Different from previous KD methods transferring dark knowledge, we present why transferring CAMs to the trained model can improve its performance on the classification task. Moreover, through experiments with CAT, we reveal several interesting properties of transferring CAMs,", + "bbox": [ + 75, + 765, + 470, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "which not only help to improve the performance and interpretability of CAT-KD but also contribute to a better understanding of CNN. While having high interpretability, CAT-KD achieves state-of-the-art performance on multiple benchmarks. Overall, the main contributions of our work are shown below:", + "bbox": [ + 496, + 248, + 893, + 338 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose class attention transfer and use it to demonstrate that the capacity of identifying class discriminative regions of input, which is critical for CNN to perform classification, can be obtained and enhanced by transferring CAMs.", + "- We present several interesting properties of transferring CAMs, which contribute to a better understanding of CNN.", + "- We apply CAT to knowledge distillation and name it CAT-KD. While having high Interpretability, CAT-KD achieves state-of-the-art performance on multiple benchmarks." + ], + "bbox": [ + 517, + 345, + 890, + 553 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Background", + "text_level": 1, + "bbox": [ + 500, + 570, + 627, + 588 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The concept of knowledge distillation was proposed in [11]. As a transfer learning method, KD aims to improve the performance of the smaller student network by transferring the dark knowledge distilled from the bigger teacher network. Previous KD methods can be divided into three types: distillation from logits [3, 6, 11, 16, 33], features [2, 10, 17-19, 23, 24, 28] and attention [29].", + "bbox": [ + 496, + 597, + 890, + 702 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To our knowledge, AT [29] is the only KD method based on transferring attention, which defines attention map as the spatial map indicating the area of input that the model focuses on most. In practice, they obtain attention maps by calculating the sum of feature maps while their values are absolutized. However, AT did not present what role attention plays during the classification and why transferring attention maps defined in this way can improve the performance of the student network.", + "bbox": [ + 496, + 703, + 890, + 838 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Previous works related to class attention originate from [34], where the authors propose to utilize high-level feature maps and the parameters of the fully connected layer to generate attention map for a specific category, which is named", + "bbox": [ + 496, + 840, + 893, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "11869", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "class activation map (CAM). According to [34], class discriminative regions of input are highlighted in the corresponding CAM. To facilitate understanding, we visualize several CAMs in Figure 2. The following works have successfully applied CAM in various weakly supervised visual tasks [14,27,31]. Besides, there are also many works focus on generalizing CAM [1,21,26] and improving the performance of models by exploiting the information contained in CAM during training [7,25].", + "bbox": [ + 75, + 90, + 472, + 227 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Previous works have not presented what role attention plays during classification and why transferring attention maps can improve the trained model's performance on the classification mission. In this paper, we focus on figuring out this question and try to propose an attention-based KD method that has both high interpretability and competitive performance.", + "bbox": [ + 75, + 229, + 472, + 335 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Our Method", + "text_level": 1, + "bbox": [ + 76, + 353, + 207, + 369 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we first analyze the structure of the mainstream CNN models and reveal that possessing the capacity of identifying class discriminative regions is critical for CNN to perform classification. Then we further propose class attention transfer to prove that this capacity can be obtained and enhanced by transferring CAMs. Finally, we apply CAT to knowledge distillation.", + "bbox": [ + 75, + 381, + 468, + 488 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Revisit the structure of CNN", + "text_level": 1, + "bbox": [ + 76, + 501, + 333, + 516 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In image classification tasks, mainstream models usually use CNN to extract features, the resulting high-level feature maps are then globally pooled and fed to a simple fully connected layer to perform classification [8,9,12]. Let $\\mathbf{F} = [F_1,F_2,\\dots,F_C]\\in \\mathbb{R}^{C\\times W\\times H}$ represents the feature maps generated by the last convolutional layer, where $C$ , $W$ , and $H$ indicate channel dimension, width, and height respectively. And $f_{j}(x,y)$ denotes the activation of $\\mathbf{F}$ in $j$ channel at spatial location $(x,y)$ , while $GAP$ is the global average pooling layer. Then the process of calculating logits for normal CNN models can be written as:", + "bbox": [ + 75, + 527, + 472, + 694 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} L _ {i} = \\sum_ {1 \\leq j \\leq C} \\omega_ {j} ^ {i} \\times G A P (F _ {j}) \\\\ = \\frac {1}{W \\times H} \\sum_ {x, y} \\sum_ {1 \\leq j \\leq C} \\omega_ {j} ^ {i} \\times f _ {j} (x, y), \\tag {1} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 137, + 705, + 468, + 779 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $L_{i}$ denotes the logit of $i$ -th class, $\\omega_{j}^{i}$ is the weight of the fully connected layer (FC layer) corresponding to class $i$ for $GAP(F_{j})$ . According to [34], we can obtain the CAM corresponding to category $i$ by:", + "bbox": [ + 76, + 794, + 470, + 856 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nC A M _ {i} (x, y) = \\sum_ {1 \\leq j \\leq C} \\omega_ {j} ^ {i} \\times f _ {j} (x, y). \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 148, + 869, + 468, + 902 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/1cedc01fa88552c4cbedf18a5871b9b87116bb34dd9d67c06402451fd18bd291.jpg", + "image_caption": [ + "Figure 3. Illustration of CAT. During CAT, the structure of teacher and student are converted to our style (Figure 1)." + ], + "image_footnote": [], + "bbox": [ + 509, + 89, + 885, + 255 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "According to Equation (1) and Equation (2), the calculation of $L_{i}$ can be written in another form:", + "bbox": [ + 498, + 316, + 890, + 345 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} L _ {i} = \\frac {1}{W \\times H} \\sum_ {x, y} C A M _ {i} (x, y) \\tag {3} \\\\ = G A P (C A M _ {i}). \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 588, + 349, + 890, + 406 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "As reflected in Equation (3), logits can be obtained by computing the average activation of CAMs. Inspired by it, as illustrated in Figure 1, we convert the FC layer into a $1 \\times 1$ convolutional layer and move the position of the GAP layer. Then $\\bar{L}_i$ , the logit of $i$ -th class generated by the converted model, can be obtained by:", + "bbox": [ + 498, + 419, + 890, + 510 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\bar {L} _ {i} = G A P \\left(\\operatorname {C o n v} _ {i} (\\mathbf {F})\\right) \\\\ = \\frac {1}{W \\times H} \\sum_ {x, y} \\left(\\sum_ {1 \\leq j \\leq C} \\omega_ {j} ^ {i} \\times f _ {j} (x, y)\\right) \\tag {4} \\\\ = G A P \\left(C A M _ {i}\\right), \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 558, + 515, + 890, + 592 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $Conv_{i}$ denotes the converted $1 \\times 1$ convolution kernel that used to separate features corresponding to $i$ -th class from $\\mathbf{F}$ , and $\\omega_{j}^{i}$ is its weight of $j$ channel. As reflected in Eqn(3) and Eqn(4), the conversion does not change the value of its prediction score (i.e., logits). And class activation maps can be obtained during the classification of the converted model.", + "bbox": [ + 496, + 598, + 890, + 702 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "As reflected in Eqn(4), the classification process of the converted model can be viewed in two steps: (1) the model exploits its capacity to identify class discriminative regions of input and generate CAMs, (2) the model outputs prediction score of each category by computing the average activation of the corresponding CAM. Considering that the model makes predictions by simply comparing the average activation of CAMs, possessing the capacity to identify class discriminative regions of input is critical for CNN to perform classification. To examine if this capacity can be obtained and enhanced by offering hints indicating class discriminative regions of input to the trained model, we propose class attention transfer.", + "bbox": [ + 496, + 704, + 892, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "11870", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Class Attention Transfer", + "text_level": 1, + "bbox": [ + 76, + 90, + 302, + 104 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The purpose of CAT is to examine if a model can obtain the capacity to identify class discriminative regions of input by transferring only CAMs. Thus, during CAT, the trained model is not required to perform classification, and any information related to the category of the training set data (e.g., ground-truth labels and logits) is not released to the trained model. In practice, we utilize a pre-trained model with the converted structure to generate the transferred CAMs. The illustration of the process of CAT is shown in Figure 3, while the formal description is shown below.", + "bbox": [ + 75, + 113, + 468, + 277 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For a given input, let $\\mathbf{A} \\in \\mathbb{R}^{K \\times W \\times H}$ denotes the CAMs generated by the converted structure, where $K$ is the number of categories contained in the classification task, $W$ and $H$ denote the width and height of the generated CAM respectively. $A_{i} \\in \\mathbb{R}^{W \\times H}$ represents the $i$ channel of $\\mathbf{A}$ , which is the CAM corresponding to category $i$ . And $S, T$ denote student and teacher correspondingly. Besides, we use the average pooling function $\\phi$ to reduce the resolution of the transferred CAMs, to improve the performance of CAT (Section 4.2). Then CAT's loss function can be defined as:", + "bbox": [ + 75, + 279, + 470, + 445 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {C A T} = \\sum_ {1 \\leq i \\leq K} \\frac {1}{K} \\| \\frac {\\phi \\left(A _ {i} ^ {T}\\right)}{\\| \\phi \\left(A _ {i} ^ {T}\\right) \\| _ {2}} - \\frac {\\phi \\left(A _ {i} ^ {S}\\right)}{\\| \\phi \\left(A _ {i} ^ {S}\\right) \\| _ {2}} \\| _ {2} ^ {2}. \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 99, + 454, + 468, + 494 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As can be seen, we perform $l_{2}$ normalization on $\\phi(A_i^T)$ and $\\phi(A_i^S)$ ( $l_{1}$ normalization can be used as well), to ensure that information related to the category of input is not released to the trained model during CAT, considering that the average activation of CAM indicates the prediction score (Equation (3)). Besides, note that here we transfer CAMs of all categories, which is based on our finding that CAMs of all categories both contain beneficial information for CAT (Section 4.2).", + "bbox": [ + 75, + 506, + 468, + 642 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our core findings through the experiments with CAT are presented as follows, while the corresponding experimental verification and their detailed analysis can be found in Section 4.2.", + "bbox": [ + 75, + 643, + 468, + 703 + ], + "page_idx": 3 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- The capacity to identify class discriminative regions of input can be obtained and enhanced by transferring CAMs.", + "- CAMs of all categories both contain beneficial information for CAT.", + "- Transferring smaller CAMs performs better.", + "- For CAT, the critical information contained in the transferred CAMs is the spatial location of the regions with high activation in them rather than their specific value." + ], + "bbox": [ + 94, + 715, + 468, + 898 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. CAT-KD", + "text_level": 1, + "bbox": [ + 500, + 90, + 604, + 104 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "After validating the effectiveness of CAT, we apply CAT to knowledge distillation and name it CAT-KD. The loss function of CAT-KD is:", + "bbox": [ + 498, + 114, + 890, + 160 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {K D} = \\mathcal {L} _ {C E} + \\beta \\mathcal {L} _ {C A T}, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 609, + 178, + 890, + 194 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathcal{L}_{CE}$ denotes the standard cross-entropy loss, and $\\beta$ is the factor used to balance the CE loss and CAT loss.", + "bbox": [ + 498, + 210, + 890, + 239 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Different from previous KD methods, we present how the knowledge transferred by CAT-KD helps to improve the performance of the student network: by improving its capacity of identifying class discriminative regions. Besides, through experiments with CAT, we analyze and reveal several properties of the knowledge transferred by our method. This further enhances the interpretability of CAT-KD.", + "bbox": [ + 496, + 243, + 890, + 349 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 498, + 369, + 633, + 386 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1. Datasets and Implementation Details", + "text_level": 1, + "bbox": [ + 498, + 397, + 818, + 412 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Datasets. In the following section we explore CAT and CAT-KD mainly on two image classification datasets:", + "bbox": [ + 498, + 422, + 890, + 452 + ], + "page_idx": 3 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "(1) CIFAR-100 [13] comprise $32 \\times 32$ pixel images of 100 categories, the training and validate sets contain 50K and 10K images.", + "(2) ImageNet [5] is a large-scale dataset for the classification of 1K categories, containing 1.2 million training and 50K validation images." + ], + "bbox": [ + 500, + 455, + 890, + 547 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Implementation details. Our implementation for CIFAR-100 and ImageNet strictly follows [2, 33]. Specifically, for CIFAR-100, we train all models for 240 epochs with batch size 64 using SGD. The initial learning rate is 0.05 (0.01 for ShuffleNet [15, 32] and MobileNet [20]), divided by 10 at 150, 180, and 210 epochs. For ImageNet, we train models for 100 epochs with batch size 512. The initial learning rate is 0.2 and divided by 10 for every 30 epochs. We experiment with various representative CNN network: VGG [22], ResNet [9], WideResNet [30], MobileNet [20], and ShuffleNet [15, 32].", + "bbox": [ + 498, + 563, + 890, + 729 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For fairness, all the results of previous methods are either reported in previous papers [2,33] (we keep our training setting the same as theirs) or obtained using codes released by the author with our training setting. All results on CIFAR-100 are the average over 5 trials, while that on ImageNet is the average over 3 trials.", + "bbox": [ + 496, + 732, + 890, + 823 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For all experiments reported in Section 4.2 and Section 4.3, without special specification, we pool the transferred CAMs into $2 \\times 2$ during CAT and CAT-KD. More implementation details such as the settings of $\\beta$ are attached in the appendix due to the page limits.", + "bbox": [ + 496, + 824, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "11871", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/6652ce4e1cdeeb81801e302608654780d43980f597ba701148f54f6465f2fe6d.jpg", + "image_caption": [ + "Figure 4. Accuracy of models trained with CAT on CIFAR-100. Left: Only CAMs of certain categories are transferred, which are selected by two strategies: (1) select categories with top n prediction scores, (2) select categories with the lowest n prediction scores. Right: The training set is reduced to contain data of partial categories only. T: test set of CIFAR-100. S: a subset of T which only contains data of classes that are not contained in the training set." + ], + "image_footnote": [], + "bbox": [ + 80, + 88, + 467, + 250 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Exploration of CAT", + "text_level": 1, + "bbox": [ + 76, + 393, + 264, + 409 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this section, we explore several properties of class attention transfer, which not only help to improve the performance and interpretability of CAT-KD but also contributes to a better understanding of CNN. Note that any information related to the category of the training set (e.g., ground-truth labels and logits) is not utilized in the experiments reported in this section.", + "bbox": [ + 75, + 417, + 470, + 523 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The capacity of identifying class discriminative regions can be obtained and enhanced by transferring CAMs. As revealed in Section 3.1, being able to identify class discriminative regions of input is critical for CNN to perform classification. Thus, the intensity of this capacity can be evaluated by the model's performance on the classification mission. We perform CAT on ShuffleNetV1, where the transferred CAMs are produced by different models with various accuracy. As the results reported in Table 1, transferring only CAMs can train a model with high accuracy on the classification mission, proving the capacity of identifying class discriminative regions can be obtained by transferring CAMs. Besides, the performance of the trained model is influenced by the accuracy of the model producing the transferred CAMs, indicating that this capacity can be enhanced by transferring more precise CAMs.", + "bbox": [ + 75, + 537, + 468, + 792 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "CAMs of all categories both contain beneficial information for CAT. For a given input, we can use the method of CAM [34] to generate class activation maps for any categories contained in the classification mission. However, though a few non-target categories may share certain similarities (e.g., shape and patterns) with the target", + "bbox": [ + 75, + 809, + 470, + 902 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/c5bcf1d0e42e4df24c38d9f7174afb12ef69db638438178f2952e3c8aeadd814.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 516, + 88, + 635, + 178 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/34a5b85f0474a65bdbced53260aa9381525d84fe82eb156cbe1a9bcf5cf51662.jpg", + "image_caption": [ + "Top1 prairie fowl" + ], + "image_footnote": [], + "bbox": [ + 516, + 179, + 635, + 271 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/54791102e3cd1e719f415854c16fb9a87a6368a066c7fb428fd10cc4371576af.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 637, + 88, + 754, + 178 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/6414cdd8ada81c8a2f4772c71b1bceea7afeb90737ee054c05097ad181744896.jpg", + "image_caption": [ + "Top1 otter" + ], + "image_footnote": [], + "bbox": [ + 637, + 179, + 754, + 271 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/3a73e068b6b6491437ece202e490c922a5599cd02ec1c2a7d53283373e7dbadb.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 759, + 88, + 875, + 178 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/8fab1b0c00593f69db1c2c5c834d8779cf84c11d62abf8cdec8c13f1c96c2486.jpg", + "image_caption": [ + "Top1basset", + "Figure 5. We set a pre-trained ResNet50 as CAMs producer to train another ResNet50 from scratch with CAT, CAMs are pooled into $2 \\times 2$ during the transfer. The first row shows the visualization of the CAMs generated by the producer, while the CAMs visualized in the second row come from the trained model." + ], + "image_footnote": [], + "bbox": [ + 759, + 179, + 875, + 271 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "category, most of them are completely irrelevant to the input from a human understanding. However, our experiments show that class activation maps of all categories both contain beneficial information for CAT.", + "bbox": [ + 496, + 400, + 890, + 460 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We first perform CAT on CIFAR-100 where only CAMs of certain categories are transferred. We designed two strategies to select the categories of the transferred CAMs: (1) select categories with the lowest n prediction scores. (2) select categories with top n prediction scores (the empirical assumption here we make is that the categories with higher prediction scores have more similarities with the target category). As the results reported in Figure 4 (left), while CAMs of classes with higher prediction scores bring more improvement, others are also beneficial for CAT. Besides, we further perform CAT on the reduced CIFAR-100, where CAMs of all classes are transferred but the training set is reduced to contain data of only partial categories. Then the trained model is evaluated on the complete test set and a subset of it which only contains data of classes that are not contained in the training set. As the results reported in Figure 4 (right), interestingly, the trained model achieves high accuracy on the subset, indicating that transferring CAMs enables the trained model to classify the categories that are not contained in the training set. This further proves that non-target CAMs contain beneficial information for CAT even if their categories seem to be irrelevant to the input from a human perspective.", + "bbox": [ + 496, + 462, + 892, + 811 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Transferring smaller CAMs performs better. Intuitively, larger CAM contains more detailed hints about the spatial location of the class discriminative regions, then transferring larger CAMs should perform better. However, insufficient accuracy of the model will result in deviations", + "bbox": [ + 496, + 824, + 893, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "11872", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/bdcc80d5c9958158942304ab20cf6ffaaa090671c389969a3c5c526e2303671e.jpg", + "image_caption": [ + "Top1 bee" + ], + "image_footnote": [], + "bbox": [ + 80, + 88, + 166, + 224 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/7cc4cfcf97f183829b521ccdc1a80400ef3df2f862d4f36ae39a034fbc3eec1e.jpg", + "image_caption": [ + "Top2 sunflower" + ], + "image_footnote": [], + "bbox": [ + 166, + 88, + 254, + 224 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/d85cffa7978b89543896f630f2cfd43d950bcddb42402095332d06088c3609ff.jpg", + "image_caption": [ + "Top3 tiger" + ], + "image_footnote": [], + "bbox": [ + 256, + 88, + 344, + 224 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/4b6e372d959922e6633655735a036a45cc23195774abfb074ed1cfb8920ecb0a.jpg", + "image_caption": [ + "Top1 kangaroo", + "Figure 6. The first row shows the visualization of the CAMs corresponding to the top 3 predicted categories, while the following row shows the visualization of them after binarization." + ], + "image_footnote": [], + "bbox": [ + 352, + 88, + 439, + 224 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/14bf39c2fff11381fa21e872653c3324b1dbe7ca968954e999dd25ab058802ae.jpg", + "image_caption": [ + "Top2 rabbit" + ], + "image_footnote": [], + "bbox": [ + 441, + 88, + 527, + 224 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/c87eb7d030dd71e46b5ecf56e77625fa0774d74ac29d3f641c787027a714fc2e.jpg", + "image_caption": [ + "Top3 fox" + ], + "image_footnote": [], + "bbox": [ + 529, + 88, + 617, + 224 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/f44d0fca783d9d4bc5bb7c17fd524620774a16d472a5ab0b13a1163110c458f5.jpg", + "image_caption": [ + "Top1 bear" + ], + "image_footnote": [], + "bbox": [ + 625, + 88, + 712, + 224 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/af8f982287d31039a84b247c226a44bf9899b028e2957c2f7f7ad36f357947fc.jpg", + "image_caption": [ + "Top2 cattle" + ], + "image_footnote": [], + "bbox": [ + 715, + 88, + 802, + 224 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/84805156fe67564ac1b1337fb9ce162b99295c718e83fe63192682fa2c530ff2.jpg", + "image_caption": [ + "Top3 otter" + ], + "image_footnote": [], + "bbox": [ + 803, + 88, + 890, + 224 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/e207e66180ebe62f8a9e0f65a6f195eda1da5a2a4c589b6441c93c7ef6d5753e.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
CAM ProducerResNet56ResNet110ResNet50
Acc72.3474.3179.34
CAT72.4774.4276.17
", + "bbox": [ + 81, + 309, + 465, + 357 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/4c3092481199014fb2627a83bbfc1495d2288bc982a782b7bb718b2fe08e693c.jpg", + "table_caption": [ + "Table 1. Accuracy $(\\%)$ of ShuffleNetV1 trained with CAT on CIFAR-100. The transferred CAMs are produced by different models with various accuracy." + ], + "table_footnote": [], + "table_body": "
BaselineModel AccResNet32×4 79.42ResNet8×4 72.5ResNet20 69.06
CAT8×879.6567.9266.21
4×479.8471.6166.43
2×279.7172.4566.84
", + "bbox": [ + 78, + 426, + 468, + 501 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "between the highlighted areas in its generated CAM and the actual class discriminative regions of the image (which can be observed in Figure 2). Besides, different models differ in their capacity to identify class discriminative regions, which will lead to subtle differences in the generated CAMs. Therefore transferring CAMs with a larger size does not necessarily improve the performance of CAT. Through experiments, we found that performing average pooling on the transferred CAMs, which will expand the highlighted areas of CAMs and reduce the bias between CAMs generated by different models, could alleviate the above issues. As the results reported in Table 2, though pooling blurs the details, transferring smaller CAMs always performs better. Besides, since the pooling operation expands the highlighted areas of CAM, which will make it encompass larger class discriminative regions, transferring pooled CAMs will force the trained model to pay attention to more discriminative regions, which can be observed in Figure 5. In practice, we pool the transferred CAMs into a smaller size to improve the performance of CAT and CAT-KD (normally $2 \\times 2$ ).", + "bbox": [ + 75, + 583, + 470, + 900 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/9e210c8a03dc6d28067ed1de250d1e03ebeb91144b2f8011bdf1afc051a81255.jpg", + "table_caption": [ + "Table 2. Accuracy $(\\%)$ of various models trained with CAT on the CIFAR-100 test set. During CAT, CAMs are pooled into various sizes. The transferred CAMs are produced by ResNet32 $\\times$ 4." + ], + "table_footnote": [], + "table_body": "
BaselineModelResNet32×4ResNet50
Acc79.4279.34
CATCAMs79.7180.45
Binarized CAMs79.3579.65
", + "bbox": [ + 500, + 304, + 890, + 369 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/527ffb2517f13bcee5b05486b998909ef069ca86094a5b97160c94aa28842d7e.jpg", + "table_caption": [ + "Table 3. Results of transferring binarized CAMs. The transferred CAMs are produced by ResNet $32 \\times 4$ ." + ], + "table_footnote": [], + "table_body": "
TeacherResNet32×4
Acc77.5179.4281.36
ReviewKD [2]76.4277.4577.91
DKD [33]76.5876.4577.29
CAT-KD76.3678.2678.84
Δ-0.22+0.81+0.93
", + "bbox": [ + 550, + 407, + 841, + 503 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 4. Comparison with two SOTA methods. The student network is ShuffleNetV1. $\\Delta$ represents the gap between CAT-KD and the best-performing method among ReviewKD and DKD (marked with underline).", + "bbox": [ + 500, + 513, + 890, + 569 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The exact value of the transferred CAMs is not important. To demonstrate that the role CAMs play in CAT is offering hints about the spatial location of the class discriminative regions of input, we binarize the values of the transferred CAMs to 0 and 1, using their average values as the thresholds. The regions of CAM with values above the threshold are considered as being highlighted, indicating the class discriminative regions of input. Thus, we set the values of these regions to 1 to keep them activated after the binarization. Other regions with values below the threshold are considered unhighlighted, and their values are set to 0. As shown in Figure 6, though the specific values of CAMs are lost during the binarization process, the binarized CAMs still contain hints about the spatial location of the class discriminative regions. Note that the threshold can also be specified in other ways (e.g., median).", + "bbox": [ + 496, + 597, + 890, + 839 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "As the results reported in Table 3, although the class discriminative regions obtained by our rudimentary binarization method are not precise, the accuracy of the resulting model dropped by less than one percent, proving that the", + "bbox": [ + 496, + 839, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "11873", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/fc74a4631d86d7b6af28b4ff7246c1c5da4aac958982fd08000375845d9f0606.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Distillation MechanismTeacher AccResNet32×4 79.42WRN40-2 75.61ResNet32×4 79.42ResNet50 79.34VGG13 74.64
Student AccShuffleNetV1 70.5ShuffleNetV1 70.5ShuffleNetV2 71.82MobileNetV2 64.6MobileNetV2 64.6
LogitsKD [11]74.0774.8374.4567.3567.37
DKD [33]76.4576.777.0770.3569.71
FeaturesCRD [23]75.1176.0575.6569.1169.73
OFD [10]75.9875.8576.8269.0469.48
FitNet [19]73.5973.7373.5463.1664.14
RKD [17]72.2872.2173.2164.4364.52
ReviewKD [2]77.4577.1477.7869.8970.37
AttentionAT [29]71.7373.3272.7358.5859.4
CAT-KD78.2677.3578.4171.3669.13
↑+6.53+4.03+5.68+12.78+9.73
", + "bbox": [ + 101, + 87, + 866, + 305 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/c495b02943812e5a7c4b3e5f009b963ce6feb8b8b7de0d7fb2291a48fc555ea4.jpg", + "table_caption": [ + "Table 5. Results on CIFAR-100. Teachers and students have different architectures. $\\uparrow$ represents the performance improvement of CAT-KD compared with AT." + ], + "table_footnote": [], + "table_body": "
Distillation MechanismTeacher AccResNet56 72.34ResNet110 74.31ResNet32×4 79.42WRN-40-2 75.61WRN-40-2 75.61VGG13 74.64
Student AccResNet20 69.06ResNet32 71.14ResNet8×4 72.5WRN-16-2 73.26WRN-40-1 71.98VGG8 70.36
LogitsKD [11]70.6673.0873.3374.9273.5472.98
DKD [33]71.9774.1176.3276.2474.8174.68
FeaturesCRD [23]71.1673.4875.5175.4874.1473.94
OFD [10]70.9873.2374.9575.2474.3373.95
FitNet [19]69.2171.0673.573.5872.2471.02
RKD [17]69.6171.8271.973.3572.2271.48
ReviewKD [2]71.8973.8975.6376.1275.0974.84
AttentionAT [29]70.5572.3173.4474.0872.7771.43
CAT-KD71.6273.6276.9175.674.8274.65
↑+1.07+1.31+3.47+1.52+2.05+3.22
", + "bbox": [ + 107, + 349, + 857, + 566 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 6. Results on CIFAR-100. Teachers and students have the same architecture. $\\uparrow$ represents the performance improvement of CAT-KD compared with AT.", + "bbox": [ + 75, + 573, + 892, + 603 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "critical information CAMs contained for CAT is the spatial location of class discriminative regions rather than its exact value. This strongly demonstrates that our method is based on transferring attention.", + "bbox": [ + 75, + 627, + 470, + 688 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Evaluation of CAT-KD", + "text_level": 1, + "bbox": [ + 76, + 696, + 289, + 710 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Consistent with previous works [2, 23, 33], we compare the performance of CAT-KD with several representative KD methods. Moreover, we further evaluate our method from two aspects: transferability and efficiency.", + "bbox": [ + 75, + 719, + 470, + 781 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Results on CIFAR-100. Table 5 reports the results on CIFAR-100 with the teachers and students having different architectures. Table 6 shows the results where teachers and students have architectures of the same style. Notably, our method outperforms the other attention-based method AT [29] with a large margin (1.07% ~ 12.78%). Moreover, CAT-KD achieves comparable or", + "bbox": [ + 75, + 794, + 470, + 902 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "even better performance compared with feature-based distillation method [2] which requires additional networks and multiple-layer information. Besides, consistent with CAT, the performance of CAT-KD is affected by the accuracy of the teacher: CAMs produced by teachers with lower accuracy contain more incorrect hints about the class discriminative regions of input. To verify this, we further evaluate the impact of the accuracy of the teacher on our method. As the results reported in Table 4, CAT-KD is relatively less effective when the teacher is weak. Thus, as can be observed in Table 6, the performance of CAT-KD is not the best when the teacher is weak.", + "bbox": [ + 496, + 628, + 893, + 809 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Results on ImageNet. Table 7 and Table 8 report the top-1 and top-5 accuracy of image classification on ImageNet. Though the performance of CAT-KD is restricted by the weakness of the teacher network in this setting, our method still outperforms most KD methods.", + "bbox": [ + 496, + 824, + 893, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "11874", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/5a28c679d216639377d8e86f469c4cc5d8ba79f7d56f30a4dc43957d095b514e.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
FeaturesLogitsAttention
TeacherStudentOFD [10]CRD [23]ReviewKD [2]KD [11]DKD [33]AT [29]CAT-KD
Top-173.3169.7570.8171.1771.6170.6671.770.6971.26
Top-591.4189.0789.9890.1390.5189.8890.4190.0190.45
", + "bbox": [ + 84, + 88, + 879, + 152 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/bd1eab9fbe0ee9b395e0db19b8677f7375080856ba9f72b35a4a97621be1b5c9.jpg", + "table_caption": [ + "Table 7. Results on ImageNet. In this group, we set ResNet34 as the teacher and ResNet18 as the student. The method with the second-best performance is marked with an underline." + ], + "table_footnote": [], + "table_body": "
FeaturesLogitsAttention
TeacherStudentOFD [10]CRD [23]ReviewKD [2]KD [11]DKD [33]AT [29]CAT-KD
Top-176.1668.8771.2571.3772.5668.5872.0569.5672.24
Top-592.8688.7690.3490.4191.0088.9891.0589.3391.13
", + "bbox": [ + 84, + 181, + 879, + 244 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/9822b0876194242460c143e9301034b8eb4421c248265276a0e5f023b6894aa5.jpg", + "table_caption": [ + "Table 8. Results on ImageNet. In this group, we set ResNet50 as the teacher and MobileNet as the student. The method with the second-best performance is marked with an underline." + ], + "table_footnote": [], + "table_body": "
Teacher\nStudentResNet32×4ResNet50
ShuffleNetV1MobileNetV2
DatasetSTLTISTLTI
Baseline69.0536.5464.3930.85
KD [11]66.6132.5667.8132.37
DKD [33]70.7336.7771.0536.48
CRD [23]70.6837.8571.4638.75
ReviewKD [2]71.4638.4666.1632.65
AT [29]71.3637.3665.129.13
CAT-KD74.4340.7373.239.87
", + "bbox": [ + 101, + 297, + 444, + 454 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 9. Comparison on transferring representations learned from CIFAR-100 to STL-10 (STL) and Tiny-ImageNet (TI).", + "bbox": [ + 75, + 465, + 467, + 493 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Transferability. We perform experiments to compare the transferability of representations to evaluate the generalizability of the knowledge transferred by various methods. We use ShuffleNetV1 and MobileNetV2 as the frozen representations extractors, which are either trained from scratch on CIFAR-100 [13] or distilled from ResNet $32 \\times 4$ and ResNet50 with various KD methods. Then linear probing tasks are performed on STL-10 [4] and TinyImageNet [5] to quantify their transferability. As the results reported in Table 9, CAT-KD outperforms other methods by a large margin, indicating the outstanding generalizability of the knowledge transferred by our method.", + "bbox": [ + 75, + 523, + 467, + 703 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Efficiency. We first compare the performance of multiple KD methods on CIFAR-100, where the training set is reduced at various ratios, to evaluate their dependence on the amount of training data. As the results reported in Figure 7 (left), CAT-KD is minimally affected by the decrease in the amount of training data, proving the outstanding distillation efficiency of our method. Besides, we further compare the training cost and performance of various KD methods. As reflected in the results reported in Figure 7 (right), CAT-KD has the highest training efficiency. Since CAT-KD does not require extra parameters, its computational cost is almost the same as logits-based", + "bbox": [ + 75, + 719, + 467, + 898 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/ac6edc14c512022db031a55c2c7ff30d9595670adf1d2157705fc3050e50188d.jpg", + "image_caption": [ + "Figure 7. We set ResNet32×4 as the teacher and ShuffleNetV1 as the student. Left: accuracy of students trained with various methods on CIFAR-100, where the training set is reduced at various ratios. Right: comparison of accuracy and training time (per epoch) on CIFAR-100." + ], + "image_footnote": [], + "bbox": [ + 500, + 285, + 890, + 454 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "methods. Relatively, feature-based methods require much more computational resources because most of them need additional auxiliary networks to distill features.", + "bbox": [ + 498, + 560, + 890, + 604 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 618, + 617, + 633 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we propose CAT-KD which has both high interpretability and competitive performance. More importantly, we demonstrate that the capacity of identifying class discriminative regions of input can be obtained and enhanced by transferring CAMs. Furthermore, we present several interesting properties of transferring CAMs, which contribute to a better understanding of CNN. We hope our findings will help future research on the interpretability of CNN and knowledge distillation.", + "bbox": [ + 496, + 643, + 890, + 779 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgement. We thank the reviewers for their constructive feedback. Part of Hui Li's work is supported by the National Natural Science Foundation of China (61932015), Shaanxi Innovation Team project (2018TD007), Higher Education Discipline Innovation 111 project (B16037). Part of Haonan Yan's work is done when he visits the University of Guelph.", + "bbox": [ + 496, + 794, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "11875", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 174, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Aditya Chattopadhyay, Anirban Sarkar, Prantik Howlader, and Vineeth N Balasubramanian. Grad-cam++: Generalized gradient-based visual explanations for deep convolutional networks. workshop on applications of computer vision, 2018. 3", + "[2] Pengguang Chen, Shu Liu, Hengshuang Zhao, and Jiaya Jia. Distilling knowledge via knowledge review. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5008-5017, 2021. 1, 2, 4, 6, 7, 8", + "[3] Jang Hyun Cho and Bharath Hariharan. On the efficacy of knowledge distillation. international conference on computer vision, 2019. 1, 2", + "[4] Adam Coates, Andrew Ng, and Honglak Lee. An analysis of single-layer networks in unsupervised feature learning. In Proceedings of the fourteenth international conference on artificial intelligence and statistics, pages 215-223. JMLR Workshop and Conference Proceedings, 2011. 8", + "[5] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. computer vision and pattern recognition, 2009. 4, 8", + "[6] Tommaso Furlanello, Zachary C. Lipton, Michael Tschannen, Laurent Itti, and Animashree Anandkumar. Born again neural networks. international conference on machine learning, 2018. 1, 2", + "[7] Hao Guo, Kang Zheng, Xiaochuan Fan, Hongkai Yu, and Song Wang. Visual attention consistency under image transforms for multi-label image classification. computer vision and pattern recognition, 2019. 3", + "[8] Dongyoon Han, Jiwhan Kim, and Junmo Kim. Deep pyramidal residual networks. computer vision and pattern recognition, 2016. 3", + "[9] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. arXiv: Computer Vision and Pattern Recognition, 2015. 3, 4", + "[10] Byeongho Heo, Jeesoo Kim, Sangdoo Yun, Hyojin Park, Nojun Kwak, and Jin Young Choi. A comprehensive overhaul of feature distillation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1921–1930, 2019. 1, 2, 7, 8", + "[11] Geoffrey E. Hinton, Oriol Vinyals, and Jeffrey Dean. Distilling the knowledge in a neural network. arXiv: Machine Learning, 2015. 1, 2, 7, 8", + "[12] Gao Huang, Zhuang Liu, Laurens van der Maaten, and Kilian Q. Weinberger. Densely connected convolutional networks. computer vision and pattern recognition, 2016. 3", + "[13] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009. 4, 8", + "[14] Kunpeng Li, Ziyan Wu, Kuan-Chuan Peng, Jan Ernst, and Yun Fu. Tell me where to look: Guided attention inference network. computer vision and pattern recognition, 2018. 3", + "[15] Ningning Ma, Xiangyu Zhang, Hai-Tao Zheng, and Jian Sun. Shufflenet v2: Practical guidelines for efficient cnn architecture design. *european conference on computer vision*, 2018. 4" + ], + "bbox": [ + 78, + 114, + 470, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[16] Seyed Iman Mirzadeh, Mehrdad Farajtabar, Ang Li, Nir Levine, Akihiro Matsukawa, and Hassan Ghasemzadeh. Improved knowledge distillation via teacher assistant. national conference on artificial intelligence, 2019. 1, 2", + "[17] Wonpyo Park, Dongju Kim, Yan Lu, and Minsu Cho. Relational knowledge distillation. computer vision and pattern recognition, 2019. 1, 2, 7", + "[18] Baoyun Peng, Xiao Jin, Dongsheng Li, Shunfeng Zhou, Yichao Wu, Jiaheng Liu, Zhaoning Zhang, and Yu Liu. Correlation congruence for knowledge distillation. international conference on computer vision, 2019. 1, 2", + "[19] Adriana Romero, Nicolas Ballas, Samira Ebrahimi Kahou, Antoine Chassang, Carlo Gatta, and Yoshua Bengio. Fitnets: Hints for thin deep nets. arXiv: Learning, 2014. 1, 2, 7", + "[20] Mark Sandler, Andrew Howard, Menglong Zhu, Andrey Zhoginov, and Liang-Chieh Chen. *Mobilenetv2: Inverted residuals and linear bottlenecks*. computer vision and pattern recognition, 2018. 4", + "[21] Ramprasaath R. Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra. Grad-cam: Visual explanations from deep networks via gradient-based localization. International Journal of Computer Vision, 2016. 3", + "[22] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. computer vision and pattern recognition, 2014. 4", + "[23] Yonglong Tian, Dilip Krishnan, and Phillip Isola. Contrastive representation distillation. arXiv preprint arXiv:1910.10699, 2019. 1, 2, 7, 8", + "[24] Frederick Tung and Greg Mori. Similarity-preserving knowledge distillation. arXiv: Computer Vision and Pattern Recognition, 2019. 1, 2", + "[25] Chaofei Wang, Jiayu Xiao, Yizeng Han, Qisen Yang, Shiji Song, and Gao Huang. Towards learning spatially discriminative feature representations. international conference on computer vision, 2021. 3", + "[26] Haofan Wang, Zifan Wang, Mengnan Du, Fan Yang, Zijian Zhang, Sirui Ding, Piotr Mardziel, and Xia Hu. Score-cam: Score-weighted visual explanations for convolutional neural networks. computer vision and pattern recognition, 2020. 3", + "[27] Seunghan Yang, YoonHyung Kim, Youngeun Kim, and Changick Kim. Combinational class activation maps for weakly supervised object localization. workshop on applications of computer vision, 2019. 3", + "[28] Junho Yim, Donggyu Joo, Ji-Hoon Bae, and Junmo Kim. A gift from knowledge distillation: Fast optimization, network minimization and transfer learning. computer vision and pattern recognition, 2017. 1, 2", + "[29] Sergey Zagoruyko and Nikos Komodakis. Paying more attention to attention: improving the performance of convolutional neural networks via attention transfer. Learning, 2016, 1, 2, 7, 8", + "[30] Sergey Zagoruyko and Nikos Komodakis. Wide residual networks. *british machine vision conference*, 2016. 4", + "[31] Xiaopeng Zhang, Jiashi Feng, Hongkai Xiong, and Qi Tian. Zigzag learning for weakly supervised object detection. computer vision and pattern recognition, 2018. 3" + ], + "bbox": [ + 501, + 92, + 892, + 901 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "11876", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[32] Xiangyu Zhang, Xinyu Zhou, Mengxiao Lin, and Jian Sun. Shufflenet: An extremely efficient convolutional neural network for mobile devices. computer vision and pattern recognition, 2017. 4", + "[33] Borui Zhao, Quan Cui, Renjie Song, Yiyu Qiu, and Jiajun Liang. Decoupled knowledge distillation. arXiv preprint arXiv:2203.08679, 2022. 1, 2, 4, 6, 7, 8", + "[34] Bolei Zhou, Aditya Khosla, Agata Lapedriza, Aude Oliva, and Antonio Torralba. Learning deep features for discriminative localization. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2921-2929, 2016. 2, 3, 5" + ], + "bbox": [ + 78, + 90, + 468, + 258 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "11877", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2023/Class Attention Transfer Based Knowledge Distillation/a68b12fe-6c62-4a74-9758-0370df04b2ad_model.json b/2023/Class Attention Transfer Based Knowledge Distillation/a68b12fe-6c62-4a74-9758-0370df04b2ad_model.json new file mode 100644 index 0000000000000000000000000000000000000000..d3cccbe25c13be6b3f2092869fa68f169b859636 --- /dev/null +++ b/2023/Class Attention Transfer Based Knowledge Distillation/a68b12fe-6c62-4a74-9758-0370df04b2ad_model.json @@ -0,0 +1,2156 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.209, + 0.131, + 0.761, + 0.154 + ], + "angle": 0, + "content": "Class Attention Transfer Based Knowledge Distillation" + }, + { + "type": "text", + "bbox": [ + 0.268, + 0.18, + 0.7, + 0.218 + ], + "angle": 0, + "content": "Ziyao Guo\\(^{1}\\), Haonan Yan\\(^{1,2,*}\\), Hui Li\\(^{1,*}\\), Xiaodong Lin\\(^{2}\\) \n\\(^{1}\\)Xidian University, \\(^{2}\\)University of Guelph" + }, + { + "type": "text", + "bbox": [ + 0.375, + 0.22, + 0.591, + 0.234 + ], + "angle": 0, + "content": "gzyaftermath@outlook.com" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.269, + 0.314, + 0.285 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.302, + 0.474, + 0.62 + ], + "angle": 0, + "content": "Previous knowledge distillation methods have shown their impressive performance on model compression tasks, however, it is hard to explain how the knowledge they transferred helps to improve the performance of the student network. In this work, we focus on proposing a knowledge distillation method that has both high interpretability and competitive performance. We first revisit the structure of mainstream CNN models and reveal that possessing the capacity of identifying class discriminative regions of input is critical for CNN to perform classification. Furthermore, we demonstrate that this capacity can be obtained and enhanced by transferring class activation maps. Based on our findings, we propose class attention transfer based knowledge distillation (CAT-KD). Different from previous KD methods, we explore and present several properties of the knowledge transferred by our method, which not only improve the interpretability of CAT-KD but also contribute to a better understanding of CNN. While having high interpretability, CAT-KD achieves state-of-the-art performance on multiple benchmarks. Code is available at: https://github.com/GzyAftermath/CAT-KD." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.651, + 0.21, + 0.667 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.677, + 0.47, + 0.875 + ], + "angle": 0, + "content": "Knowledge distillation (KD) transfers knowledge distilled from the bigger teacher network to the smaller student network, aiming to improve the performance of the student network. Depending on the type of the transferred knowledge, previous KD methods can be divided into three categories: based on transferring logits [3, 6, 11, 16, 33], features [2, 10, 17-19, 23, 24, 28], and attention [29]. Although KD methods that are based on transferring logits and features have shown their promising performance [2, 33], it is hard to explain how the knowledge they transferred helps to improve the performance of the student network, due to the uninterpretability of logits and features. Relatively, the principle of attention-based KD methods is more intuitive:" + }, + { + "type": "image", + "bbox": [ + 0.511, + 0.268, + 0.885, + 0.576 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.586, + 0.894, + 0.643 + ], + "angle": 0, + "content": "Figure 1. Illustration of the converted structure. After converting the FC layer into a convolutional layer with \\(1 \\times 1\\) kernel and moving the position of the global average pooling layer, CAMs can be obtained during the forward propagation." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.674, + 0.895, + 0.901 + ], + "angle": 0, + "content": "it aims at telling the student network which part of the input should it focus on during the classification, which is realized by forcing the student network to mimic the transferred attention maps during training. However, though previous work AT [29] has validated the effectiveness of transferring attention, it does not present what role attention plays during the classification. This makes it hard to explain why telling the trained model where should it focus could improve its performance on the classification mission. Besides, the performance of the previous attention-based KD method [29] is less competitive compared with the methods that are based on transferring logits and features [2, 33]. In this work, we focus on proposing an attention-based KD method that has higher interpretability and better performance." + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.887, + 0.221, + 0.9 + ], + "angle": 0, + "content": "*Corresponding author" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "11868" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.08, + 0.089, + 0.239, + 0.183 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.243, + 0.089, + 0.403, + 0.183 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.406, + 0.089, + 0.566, + 0.183 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.568, + 0.089, + 0.729, + 0.183 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.732, + 0.089, + 0.892, + 0.183 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.194, + 0.893, + 0.222 + ], + "angle": 0, + "content": "Figure 2. Visualization of CAMs corresponding to categories with Top 4 prediction scores for the given image. The predicted categories and their scores are reported in the picture." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.249, + 0.473, + 0.566 + ], + "angle": 0, + "content": "We start our work by exploring what role attention plays during classification. After revisiting the structure of the mainstream models, we find that with a little conversion (illustrated in Figure 1), class activation map (CAM) [34], a kind of class attention map which indicates the discriminative regions of input for a specific category, can be obtained during the classification. Without changing the parameters and outputs, the classification process of the converted model can be viewed in two steps: (1) the model exploits its capacity to identify class discriminative regions of input and generate CAM for each category contained in the classification mission, (2) the model outputs the prediction score of each category by computing the average activation of the corresponding CAM. Considering that the converted model makes predictions by simply comparing the average activation of CAMs, possessing the capacity to identify class discriminative regions of input is critical for CNN to perform classification. The question is: can we enhance this capacity by offering hints about class discriminative regions of input during training? To answer this question, we propose class attention transfer (CAT)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.568, + 0.47, + 0.764 + ], + "angle": 0, + "content": "During CAT, the trained model is not required to predict the category of input, it is only forced to mimic the transferred CAMs, which are normalized to ensure they only contain hints about class discriminative regions of input. Through experiments with CAT, we reveal that transferring only CAMs can train a model with high accuracy on the classification task, reflecting the trained model obtains the capacity to identify class discriminative regions of input. Besides, the performance of the trained model is influenced by the accuracy of the model offering the transferred CAMs. This further demonstrates that the capacity of identifying class discriminative regions can be enhanced by transferring more precise CAMs." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.766, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Based on our findings, we propose class attention transfer based knowledge distillation (CAT-KD), aiming to enable the student network to achieve better performance by improving its capacity of identifying class discriminative regions. Different from previous KD methods transferring dark knowledge, we present why transferring CAMs to the trained model can improve its performance on the classification task. Moreover, through experiments with CAT, we reveal several interesting properties of transferring CAMs," + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.249, + 0.895, + 0.339 + ], + "angle": 0, + "content": "which not only help to improve the performance and interpretability of CAT-KD but also contribute to a better understanding of CNN. While having high interpretability, CAT-KD achieves state-of-the-art performance on multiple benchmarks. Overall, the main contributions of our work are shown below:" + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.347, + 0.892, + 0.422 + ], + "angle": 0, + "content": "- We propose class attention transfer and use it to demonstrate that the capacity of identifying class discriminative regions of input, which is critical for CNN to perform classification, can be obtained and enhanced by transferring CAMs." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.436, + 0.892, + 0.48 + ], + "angle": 0, + "content": "- We present several interesting properties of transferring CAMs, which contribute to a better understanding of CNN." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.495, + 0.892, + 0.554 + ], + "angle": 0, + "content": "- We apply CAT to knowledge distillation and name it CAT-KD. While having high Interpretability, CAT-KD achieves state-of-the-art performance on multiple benchmarks." + }, + { + "type": "list", + "bbox": [ + 0.518, + 0.347, + 0.892, + 0.554 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.571, + 0.629, + 0.589 + ], + "angle": 0, + "content": "2. Background" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.598, + 0.892, + 0.703 + ], + "angle": 0, + "content": "The concept of knowledge distillation was proposed in [11]. As a transfer learning method, KD aims to improve the performance of the smaller student network by transferring the dark knowledge distilled from the bigger teacher network. Previous KD methods can be divided into three types: distillation from logits [3, 6, 11, 16, 33], features [2, 10, 17-19, 23, 24, 28] and attention [29]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.704, + 0.892, + 0.839 + ], + "angle": 0, + "content": "To our knowledge, AT [29] is the only KD method based on transferring attention, which defines attention map as the spatial map indicating the area of input that the model focuses on most. In practice, they obtain attention maps by calculating the sum of feature maps while their values are absolutized. However, AT did not present what role attention plays during the classification and why transferring attention maps defined in this way can improve the performance of the student network." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.841, + 0.894, + 0.901 + ], + "angle": 0, + "content": "Previous works related to class attention originate from [34], where the authors propose to utilize high-level feature maps and the parameters of the fully connected layer to generate attention map for a specific category, which is named" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "11869" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.228 + ], + "angle": 0, + "content": "class activation map (CAM). According to [34], class discriminative regions of input are highlighted in the corresponding CAM. To facilitate understanding, we visualize several CAMs in Figure 2. The following works have successfully applied CAM in various weakly supervised visual tasks [14,27,31]. Besides, there are also many works focus on generalizing CAM [1,21,26] and improving the performance of models by exploiting the information contained in CAM during training [7,25]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.23, + 0.473, + 0.336 + ], + "angle": 0, + "content": "Previous works have not presented what role attention plays during classification and why transferring attention maps can improve the trained model's performance on the classification mission. In this paper, we focus on figuring out this question and try to propose an attention-based KD method that has both high interpretability and competitive performance." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.354, + 0.208, + 0.37 + ], + "angle": 0, + "content": "3. Our Method" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.382, + 0.47, + 0.489 + ], + "angle": 0, + "content": "In this section, we first analyze the structure of the mainstream CNN models and reveal that possessing the capacity of identifying class discriminative regions is critical for CNN to perform classification. Then we further propose class attention transfer to prove that this capacity can be obtained and enhanced by transferring CAMs. Finally, we apply CAT to knowledge distillation." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.502, + 0.334, + 0.517 + ], + "angle": 0, + "content": "3.1. Revisit the structure of CNN" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.528, + 0.473, + 0.695 + ], + "angle": 0, + "content": "In image classification tasks, mainstream models usually use CNN to extract features, the resulting high-level feature maps are then globally pooled and fed to a simple fully connected layer to perform classification [8,9,12]. Let \\(\\mathbf{F} = [F_1,F_2,\\dots,F_C]\\in \\mathbb{R}^{C\\times W\\times H}\\) represents the feature maps generated by the last convolutional layer, where \\(C\\), \\(W\\), and \\(H\\) indicate channel dimension, width, and height respectively. And \\(f_{j}(x,y)\\) denotes the activation of \\(\\mathbf{F}\\) in \\(j\\) channel at spatial location \\((x,y)\\), while \\(GAP\\) is the global average pooling layer. Then the process of calculating logits for normal CNN models can be written as:" + }, + { + "type": "equation", + "bbox": [ + 0.138, + 0.706, + 0.469, + 0.78 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} L _ {i} = \\sum_ {1 \\leq j \\leq C} \\omega_ {j} ^ {i} \\times G A P (F _ {j}) \\\\ = \\frac {1}{W \\times H} \\sum_ {x, y} \\sum_ {1 \\leq j \\leq C} \\omega_ {j} ^ {i} \\times f _ {j} (x, y), \\tag {1} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.795, + 0.471, + 0.857 + ], + "angle": 0, + "content": "where \\(L_{i}\\) denotes the logit of \\(i\\)-th class, \\(\\omega_{j}^{i}\\) is the weight of the fully connected layer (FC layer) corresponding to class \\(i\\) for \\(GAP(F_{j})\\). According to [34], we can obtain the CAM corresponding to category \\(i\\) by:" + }, + { + "type": "equation", + "bbox": [ + 0.149, + 0.871, + 0.47, + 0.904 + ], + "angle": 0, + "content": "\\[\nC A M _ {i} (x, y) = \\sum_ {1 \\leq j \\leq C} \\omega_ {j} ^ {i} \\times f _ {j} (x, y). \\tag {2}\n\\]" + }, + { + "type": "image", + "bbox": [ + 0.51, + 0.09, + 0.887, + 0.256 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.265, + 0.892, + 0.295 + ], + "angle": 0, + "content": "Figure 3. Illustration of CAT. During CAT, the structure of teacher and student are converted to our style (Figure 1)." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.317, + 0.892, + 0.346 + ], + "angle": 0, + "content": "According to Equation (1) and Equation (2), the calculation of \\( L_{i} \\) can be written in another form:" + }, + { + "type": "equation", + "bbox": [ + 0.589, + 0.351, + 0.892, + 0.407 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} L _ {i} = \\frac {1}{W \\times H} \\sum_ {x, y} C A M _ {i} (x, y) \\tag {3} \\\\ = G A P (C A M _ {i}). \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.42, + 0.892, + 0.511 + ], + "angle": 0, + "content": "As reflected in Equation (3), logits can be obtained by computing the average activation of CAMs. Inspired by it, as illustrated in Figure 1, we convert the FC layer into a \\(1 \\times 1\\) convolutional layer and move the position of the GAP layer. Then \\(\\bar{L}_i\\), the logit of \\(i\\)-th class generated by the converted model, can be obtained by:" + }, + { + "type": "equation", + "bbox": [ + 0.559, + 0.516, + 0.892, + 0.593 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\bar {L} _ {i} = G A P \\left(\\operatorname {C o n v} _ {i} (\\mathbf {F})\\right) \\\\ = \\frac {1}{W \\times H} \\sum_ {x, y} \\left(\\sum_ {1 \\leq j \\leq C} \\omega_ {j} ^ {i} \\times f _ {j} (x, y)\\right) \\tag {4} \\\\ = G A P \\left(C A M _ {i}\\right), \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.599, + 0.892, + 0.703 + ], + "angle": 0, + "content": "where \\( Conv_{i} \\) denotes the converted \\( 1 \\times 1 \\) convolution kernel that used to separate features corresponding to \\( i \\)-th class from \\( \\mathbf{F} \\), and \\( \\omega_{j}^{i} \\) is its weight of \\( j \\) channel. As reflected in Eqn(3) and Eqn(4), the conversion does not change the value of its prediction score (i.e., logits). And class activation maps can be obtained during the classification of the converted model." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.705, + 0.893, + 0.901 + ], + "angle": 0, + "content": "As reflected in Eqn(4), the classification process of the converted model can be viewed in two steps: (1) the model exploits its capacity to identify class discriminative regions of input and generate CAMs, (2) the model outputs prediction score of each category by computing the average activation of the corresponding CAM. Considering that the model makes predictions by simply comparing the average activation of CAMs, possessing the capacity to identify class discriminative regions of input is critical for CNN to perform classification. To examine if this capacity can be obtained and enhanced by offering hints indicating class discriminative regions of input to the trained model, we propose class attention transfer." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "11870" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.078, + 0.091, + 0.303, + 0.105 + ], + "angle": 0, + "content": "3.2. Class Attention Transfer" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.114, + 0.47, + 0.278 + ], + "angle": 0, + "content": "The purpose of CAT is to examine if a model can obtain the capacity to identify class discriminative regions of input by transferring only CAMs. Thus, during CAT, the trained model is not required to perform classification, and any information related to the category of the training set data (e.g., ground-truth labels and logits) is not released to the trained model. In practice, we utilize a pre-trained model with the converted structure to generate the transferred CAMs. The illustration of the process of CAT is shown in Figure 3, while the formal description is shown below." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.28, + 0.471, + 0.446 + ], + "angle": 0, + "content": "For a given input, let \\(\\mathbf{A} \\in \\mathbb{R}^{K \\times W \\times H}\\) denotes the CAMs generated by the converted structure, where \\(K\\) is the number of categories contained in the classification task, \\(W\\) and \\(H\\) denote the width and height of the generated CAM respectively. \\(A_{i} \\in \\mathbb{R}^{W \\times H}\\) represents the \\(i\\) channel of \\(\\mathbf{A}\\), which is the CAM corresponding to category \\(i\\). And \\(S, T\\) denote student and teacher correspondingly. Besides, we use the average pooling function \\(\\phi\\) to reduce the resolution of the transferred CAMs, to improve the performance of CAT (Section 4.2). Then CAT's loss function can be defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.101, + 0.455, + 0.47, + 0.495 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {C A T} = \\sum_ {1 \\leq i \\leq K} \\frac {1}{K} \\| \\frac {\\phi \\left(A _ {i} ^ {T}\\right)}{\\| \\phi \\left(A _ {i} ^ {T}\\right) \\| _ {2}} - \\frac {\\phi \\left(A _ {i} ^ {S}\\right)}{\\| \\phi \\left(A _ {i} ^ {S}\\right) \\| _ {2}} \\| _ {2} ^ {2}. \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.507, + 0.47, + 0.643 + ], + "angle": 0, + "content": "As can be seen, we perform \\(l_{2}\\) normalization on \\(\\phi(A_i^T)\\) and \\(\\phi(A_i^S)\\) (\\(l_{1}\\) normalization can be used as well), to ensure that information related to the category of input is not released to the trained model during CAT, considering that the average activation of CAM indicates the prediction score (Equation (3)). Besides, note that here we transfer CAMs of all categories, which is based on our finding that CAMs of all categories both contain beneficial information for CAT (Section 4.2)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.644, + 0.47, + 0.704 + ], + "angle": 0, + "content": "Our core findings through the experiments with CAT are presented as follows, while the corresponding experimental verification and their detailed analysis can be found in Section 4.2." + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.717, + 0.469, + 0.761 + ], + "angle": 0, + "content": "- The capacity to identify class discriminative regions of input can be obtained and enhanced by transferring CAMs." + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.773, + 0.469, + 0.803 + ], + "angle": 0, + "content": "- CAMs of all categories both contain beneficial information for CAT." + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.814, + 0.403, + 0.83 + ], + "angle": 0, + "content": "- Transferring smaller CAMs performs better." + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.84, + 0.469, + 0.899 + ], + "angle": 0, + "content": "- For CAT, the critical information contained in the transferred CAMs is the spatial location of the regions with high activation in them rather than their specific value." + }, + { + "type": "list", + "bbox": [ + 0.096, + 0.717, + 0.469, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.091, + 0.605, + 0.105 + ], + "angle": 0, + "content": "3.3. CAT-KD" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.116, + 0.892, + 0.161 + ], + "angle": 0, + "content": "After validating the effectiveness of CAT, we apply CAT to knowledge distillation and name it CAT-KD. The loss function of CAT-KD is:" + }, + { + "type": "equation", + "bbox": [ + 0.611, + 0.179, + 0.892, + 0.195 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {K D} = \\mathcal {L} _ {C E} + \\beta \\mathcal {L} _ {C A T}, \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.212, + 0.892, + 0.241 + ], + "angle": 0, + "content": "where \\(\\mathcal{L}_{CE}\\) denotes the standard cross-entropy loss, and \\(\\beta\\) is the factor used to balance the CE loss and CAT loss." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.244, + 0.892, + 0.35 + ], + "angle": 0, + "content": "Different from previous KD methods, we present how the knowledge transferred by CAT-KD helps to improve the performance of the student network: by improving its capacity of identifying class discriminative regions. Besides, through experiments with CAT, we analyze and reveal several properties of the knowledge transferred by our method. This further enhances the interpretability of CAT-KD." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.371, + 0.634, + 0.387 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.398, + 0.82, + 0.414 + ], + "angle": 0, + "content": "4.1. Datasets and Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.424, + 0.892, + 0.453 + ], + "angle": 0, + "content": "Datasets. In the following section we explore CAT and CAT-KD mainly on two image classification datasets:" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.456, + 0.892, + 0.501 + ], + "angle": 0, + "content": "(1) CIFAR-100 [13] comprise \\(32 \\times 32\\) pixel images of 100 categories, the training and validate sets contain 50K and 10K images." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.504, + 0.892, + 0.549 + ], + "angle": 0, + "content": "(2) ImageNet [5] is a large-scale dataset for the classification of 1K categories, containing 1.2 million training and 50K validation images." + }, + { + "type": "list", + "bbox": [ + 0.5, + 0.456, + 0.892, + 0.549 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.564, + 0.892, + 0.731 + ], + "angle": 0, + "content": "Implementation details. Our implementation for CIFAR-100 and ImageNet strictly follows [2, 33]. Specifically, for CIFAR-100, we train all models for 240 epochs with batch size 64 using SGD. The initial learning rate is 0.05 (0.01 for ShuffleNet [15, 32] and MobileNet [20]), divided by 10 at 150, 180, and 210 epochs. For ImageNet, we train models for 100 epochs with batch size 512. The initial learning rate is 0.2 and divided by 10 for every 30 epochs. We experiment with various representative CNN network: VGG [22], ResNet [9], WideResNet [30], MobileNet [20], and ShuffleNet [15, 32]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.733, + 0.892, + 0.824 + ], + "angle": 0, + "content": "For fairness, all the results of previous methods are either reported in previous papers [2,33] (we keep our training setting the same as theirs) or obtained using codes released by the author with our training setting. All results on CIFAR-100 are the average over 5 trials, while that on ImageNet is the average over 3 trials." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.825, + 0.892, + 0.901 + ], + "angle": 0, + "content": "For all experiments reported in Section 4.2 and Section 4.3, without special specification, we pool the transferred CAMs into \\(2 \\times 2\\) during CAT and CAT-KD. More implementation details such as the settings of \\(\\beta\\) are attached in the appendix due to the page limits." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "11871" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.081, + 0.089, + 0.468, + 0.251 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.255, + 0.47, + 0.366 + ], + "angle": 0, + "content": "Figure 4. Accuracy of models trained with CAT on CIFAR-100. Left: Only CAMs of certain categories are transferred, which are selected by two strategies: (1) select categories with top n prediction scores, (2) select categories with the lowest n prediction scores. Right: The training set is reduced to contain data of partial categories only. T: test set of CIFAR-100. S: a subset of T which only contains data of classes that are not contained in the training set." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.394, + 0.266, + 0.41 + ], + "angle": 0, + "content": "4.2. Exploration of CAT" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.418, + 0.471, + 0.525 + ], + "angle": 0, + "content": "In this section, we explore several properties of class attention transfer, which not only help to improve the performance and interpretability of CAT-KD but also contributes to a better understanding of CNN. Note that any information related to the category of the training set (e.g., ground-truth labels and logits) is not utilized in the experiments reported in this section." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.538, + 0.47, + 0.794 + ], + "angle": 0, + "content": "The capacity of identifying class discriminative regions can be obtained and enhanced by transferring CAMs. As revealed in Section 3.1, being able to identify class discriminative regions of input is critical for CNN to perform classification. Thus, the intensity of this capacity can be evaluated by the model's performance on the classification mission. We perform CAT on ShuffleNetV1, where the transferred CAMs are produced by different models with various accuracy. As the results reported in Table 1, transferring only CAMs can train a model with high accuracy on the classification mission, proving the capacity of identifying class discriminative regions can be obtained by transferring CAMs. Besides, the performance of the trained model is influenced by the accuracy of the model producing the transferred CAMs, indicating that this capacity can be enhanced by transferring more precise CAMs." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.81, + 0.471, + 0.904 + ], + "angle": 0, + "content": "CAMs of all categories both contain beneficial information for CAT. For a given input, we can use the method of CAM [34] to generate class activation maps for any categories contained in the classification mission. However, though a few non-target categories may share certain similarities (e.g., shape and patterns) with the target" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.089, + 0.637, + 0.179 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.18, + 0.636, + 0.272 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.518, + 0.275, + 0.635, + 0.29 + ], + "angle": 0, + "content": "Top1 prairie fowl" + }, + { + "type": "image", + "bbox": [ + 0.638, + 0.089, + 0.755, + 0.179 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.638, + 0.18, + 0.755, + 0.272 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.662, + 0.275, + 0.732, + 0.29 + ], + "angle": 0, + "content": "Top1 otter" + }, + { + "type": "image", + "bbox": [ + 0.76, + 0.089, + 0.877, + 0.179 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.76, + 0.18, + 0.877, + 0.272 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.779, + 0.275, + 0.858, + 0.29 + ], + "angle": 0, + "content": "Top1basset" + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.3, + 0.895, + 0.37 + ], + "angle": 0, + "content": "Figure 5. We set a pre-trained ResNet50 as CAMs producer to train another ResNet50 from scratch with CAT, CAMs are pooled into \\(2 \\times 2\\) during the transfer. The first row shows the visualization of the CAMs generated by the producer, while the CAMs visualized in the second row come from the trained model." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.401, + 0.892, + 0.461 + ], + "angle": 0, + "content": "category, most of them are completely irrelevant to the input from a human understanding. However, our experiments show that class activation maps of all categories both contain beneficial information for CAT." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.463, + 0.893, + 0.812 + ], + "angle": 0, + "content": "We first perform CAT on CIFAR-100 where only CAMs of certain categories are transferred. We designed two strategies to select the categories of the transferred CAMs: (1) select categories with the lowest n prediction scores. (2) select categories with top n prediction scores (the empirical assumption here we make is that the categories with higher prediction scores have more similarities with the target category). As the results reported in Figure 4 (left), while CAMs of classes with higher prediction scores bring more improvement, others are also beneficial for CAT. Besides, we further perform CAT on the reduced CIFAR-100, where CAMs of all classes are transferred but the training set is reduced to contain data of only partial categories. Then the trained model is evaluated on the complete test set and a subset of it which only contains data of classes that are not contained in the training set. As the results reported in Figure 4 (right), interestingly, the trained model achieves high accuracy on the subset, indicating that transferring CAMs enables the trained model to classify the categories that are not contained in the training set. This further proves that non-target CAMs contain beneficial information for CAT even if their categories seem to be irrelevant to the input from a human perspective." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.825, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Transferring smaller CAMs performs better. Intuitively, larger CAM contains more detailed hints about the spatial location of the class discriminative regions, then transferring larger CAMs should perform better. However, insufficient accuracy of the model will result in deviations" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "11872" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.081, + 0.089, + 0.167, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.106, + 0.228, + 0.14, + 0.258 + ], + "angle": 0, + "content": "Top1 bee" + }, + { + "type": "image", + "bbox": [ + 0.168, + 0.089, + 0.255, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.179, + 0.228, + 0.246, + 0.257 + ], + "angle": 0, + "content": "Top2 sunflower" + }, + { + "type": "image", + "bbox": [ + 0.257, + 0.089, + 0.345, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.282, + 0.228, + 0.318, + 0.258 + ], + "angle": 0, + "content": "Top3 tiger" + }, + { + "type": "image", + "bbox": [ + 0.353, + 0.089, + 0.44, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.365, + 0.228, + 0.429, + 0.258 + ], + "angle": 0, + "content": "Top1 kangaroo" + }, + { + "type": "image", + "bbox": [ + 0.442, + 0.089, + 0.529, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.465, + 0.228, + 0.506, + 0.257 + ], + "angle": 0, + "content": "Top2 rabbit" + }, + { + "type": "image", + "bbox": [ + 0.53, + 0.089, + 0.618, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.556, + 0.228, + 0.588, + 0.256 + ], + "angle": 0, + "content": "Top3 fox" + }, + { + "type": "image", + "bbox": [ + 0.627, + 0.089, + 0.714, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.653, + 0.228, + 0.686, + 0.256 + ], + "angle": 0, + "content": "Top1 bear" + }, + { + "type": "image", + "bbox": [ + 0.716, + 0.089, + 0.803, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.741, + 0.228, + 0.778, + 0.256 + ], + "angle": 0, + "content": "Top2 cattle" + }, + { + "type": "image", + "bbox": [ + 0.805, + 0.089, + 0.892, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.831, + 0.228, + 0.865, + 0.256 + ], + "angle": 0, + "content": "Top3 otter" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.263, + 0.892, + 0.292 + ], + "angle": 0, + "content": "Figure 6. The first row shows the visualization of the CAMs corresponding to the top 3 predicted categories, while the following row shows the visualization of them after binarization." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.31, + 0.466, + 0.358 + ], + "angle": 0, + "content": "
CAM ProducerResNet56ResNet110ResNet50
Acc72.3474.3179.34
CAT72.4774.4276.17
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.369, + 0.47, + 0.412 + ], + "angle": 0, + "content": "Table 1. Accuracy \\((\\%)\\) of ShuffleNetV1 trained with CAT on CIFAR-100. The transferred CAMs are produced by different models with various accuracy." + }, + { + "type": "table", + "bbox": [ + 0.08, + 0.427, + 0.469, + 0.502 + ], + "angle": 0, + "content": "
BaselineModel AccResNet32×4 79.42ResNet8×4 72.5ResNet20 69.06
CAT8×879.6567.9266.21
4×479.8471.6166.43
2×279.7172.4566.84
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.512, + 0.47, + 0.556 + ], + "angle": 0, + "content": "Table 2. Accuracy \\((\\%)\\) of various models trained with CAT on the CIFAR-100 test set. During CAT, CAMs are pooled into various sizes. The transferred CAMs are produced by ResNet32 \\(\\times\\) 4." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.584, + 0.471, + 0.901 + ], + "angle": 0, + "content": "between the highlighted areas in its generated CAM and the actual class discriminative regions of the image (which can be observed in Figure 2). Besides, different models differ in their capacity to identify class discriminative regions, which will lead to subtle differences in the generated CAMs. Therefore transferring CAMs with a larger size does not necessarily improve the performance of CAT. Through experiments, we found that performing average pooling on the transferred CAMs, which will expand the highlighted areas of CAMs and reduce the bias between CAMs generated by different models, could alleviate the above issues. As the results reported in Table 2, though pooling blurs the details, transferring smaller CAMs always performs better. Besides, since the pooling operation expands the highlighted areas of CAM, which will make it encompass larger class discriminative regions, transferring pooled CAMs will force the trained model to pay attention to more discriminative regions, which can be observed in Figure 5. In practice, we pool the transferred CAMs into a smaller size to improve the performance of CAT and CAT-KD (normally \\(2 \\times 2\\))." + }, + { + "type": "table", + "bbox": [ + 0.501, + 0.305, + 0.892, + 0.37 + ], + "angle": 0, + "content": "
BaselineModelResNet32×4ResNet50
Acc79.4279.34
CATCAMs79.7180.45
Binarized CAMs79.3579.65
" + }, + { + "type": "table_caption", + "bbox": [ + 0.5, + 0.371, + 0.892, + 0.4 + ], + "angle": 0, + "content": "Table 3. Results of transferring binarized CAMs. The transferred CAMs are produced by ResNet \\(32 \\times 4\\)." + }, + { + "type": "table", + "bbox": [ + 0.552, + 0.408, + 0.842, + 0.504 + ], + "angle": 0, + "content": "
TeacherResNet32×4
Acc77.5179.4281.36
ReviewKD [2]76.4277.4577.91
DKD [33]76.5876.4577.29
CAT-KD76.3678.2678.84
Δ-0.22+0.81+0.93
" + }, + { + "type": "table_caption", + "bbox": [ + 0.5, + 0.514, + 0.892, + 0.57 + ], + "angle": 0, + "content": "Table 4. Comparison with two SOTA methods. The student network is ShuffleNetV1. \\(\\Delta\\) represents the gap between CAT-KD and the best-performing method among ReviewKD and DKD (marked with underline)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.598, + 0.892, + 0.84 + ], + "angle": 0, + "content": "The exact value of the transferred CAMs is not important. To demonstrate that the role CAMs play in CAT is offering hints about the spatial location of the class discriminative regions of input, we binarize the values of the transferred CAMs to 0 and 1, using their average values as the thresholds. The regions of CAM with values above the threshold are considered as being highlighted, indicating the class discriminative regions of input. Thus, we set the values of these regions to 1 to keep them activated after the binarization. Other regions with values below the threshold are considered unhighlighted, and their values are set to 0. As shown in Figure 6, though the specific values of CAMs are lost during the binarization process, the binarized CAMs still contain hints about the spatial location of the class discriminative regions. Note that the threshold can also be specified in other ways (e.g., median)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.892, + 0.901 + ], + "angle": 0, + "content": "As the results reported in Table 3, although the class discriminative regions obtained by our rudimentary binarization method are not precise, the accuracy of the resulting model dropped by less than one percent, proving that the" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "11873" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.102, + 0.088, + 0.867, + 0.306 + ], + "angle": 0, + "content": "
Distillation MechanismTeacher AccResNet32×4 79.42WRN40-2 75.61ResNet32×4 79.42ResNet50 79.34VGG13 74.64
Student AccShuffleNetV1 70.5ShuffleNetV1 70.5ShuffleNetV2 71.82MobileNetV2 64.6MobileNetV2 64.6
LogitsKD [11]74.0774.8374.4567.3567.37
DKD [33]76.4576.777.0770.3569.71
FeaturesCRD [23]75.1176.0575.6569.1169.73
OFD [10]75.9875.8576.8269.0469.48
FitNet [19]73.5973.7373.5463.1664.14
RKD [17]72.2872.2173.2164.4364.52
ReviewKD [2]77.4577.1477.7869.8970.37
AttentionAT [29]71.7373.3272.7358.5859.4
CAT-KD78.2677.3578.4171.3669.13
↑+6.53+4.03+5.68+12.78+9.73
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.312, + 0.895, + 0.34 + ], + "angle": 0, + "content": "Table 5. Results on CIFAR-100. Teachers and students have different architectures. \\( \\uparrow \\) represents the performance improvement of CAT-KD compared with AT." + }, + { + "type": "table", + "bbox": [ + 0.109, + 0.35, + 0.859, + 0.567 + ], + "angle": 0, + "content": "
Distillation MechanismTeacher AccResNet56 72.34ResNet110 74.31ResNet32×4 79.42WRN-40-2 75.61WRN-40-2 75.61VGG13 74.64
Student AccResNet20 69.06ResNet32 71.14ResNet8×4 72.5WRN-16-2 73.26WRN-40-1 71.98VGG8 70.36
LogitsKD [11]70.6673.0873.3374.9273.5472.98
DKD [33]71.9774.1176.3276.2474.8174.68
FeaturesCRD [23]71.1673.4875.5175.4874.1473.94
OFD [10]70.9873.2374.9575.2474.3373.95
FitNet [19]69.2171.0673.573.5872.2471.02
RKD [17]69.6171.8271.973.3572.2271.48
ReviewKD [2]71.8973.8975.6376.1275.0974.84
AttentionAT [29]70.5572.3173.4474.0872.7771.43
CAT-KD71.6273.6276.9175.674.8274.65
↑+1.07+1.31+3.47+1.52+2.05+3.22
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.574, + 0.893, + 0.604 + ], + "angle": 0, + "content": "Table 6. Results on CIFAR-100. Teachers and students have the same architecture. \\( \\uparrow \\) represents the performance improvement of CAT-KD compared with AT." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.628, + 0.471, + 0.689 + ], + "angle": 0, + "content": "critical information CAMs contained for CAT is the spatial location of class discriminative regions rather than its exact value. This strongly demonstrates that our method is based on transferring attention." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.697, + 0.29, + 0.712 + ], + "angle": 0, + "content": "4.3. Evaluation of CAT-KD" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.72, + 0.471, + 0.782 + ], + "angle": 0, + "content": "Consistent with previous works [2, 23, 33], we compare the performance of CAT-KD with several representative KD methods. Moreover, we further evaluate our method from two aspects: transferability and efficiency." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.795, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Results on CIFAR-100. Table 5 reports the results on CIFAR-100 with the teachers and students having different architectures. Table 6 shows the results where teachers and students have architectures of the same style. Notably, our method outperforms the other attention-based method AT [29] with a large margin (1.07% ~ 12.78%). Moreover, CAT-KD achieves comparable or" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.629, + 0.895, + 0.81 + ], + "angle": 0, + "content": "even better performance compared with feature-based distillation method [2] which requires additional networks and multiple-layer information. Besides, consistent with CAT, the performance of CAT-KD is affected by the accuracy of the teacher: CAMs produced by teachers with lower accuracy contain more incorrect hints about the class discriminative regions of input. To verify this, we further evaluate the impact of the accuracy of the teacher on our method. As the results reported in Table 4, CAT-KD is relatively less effective when the teacher is weak. Thus, as can be observed in Table 6, the performance of CAT-KD is not the best when the teacher is weak." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.825, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Results on ImageNet. Table 7 and Table 8 report the top-1 and top-5 accuracy of image classification on ImageNet. Though the performance of CAT-KD is restricted by the weakness of the teacher network in this setting, our method still outperforms most KD methods." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "11874" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.086, + 0.089, + 0.88, + 0.153 + ], + "angle": 0, + "content": "
FeaturesLogitsAttention
TeacherStudentOFD [10]CRD [23]ReviewKD [2]KD [11]DKD [33]AT [29]CAT-KD
Top-173.3169.7570.8171.1771.6170.6671.770.6971.26
Top-591.4189.0789.9890.1390.5189.8890.4190.0190.45
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.155, + 0.892, + 0.181 + ], + "angle": 0, + "content": "Table 7. Results on ImageNet. In this group, we set ResNet34 as the teacher and ResNet18 as the student. The method with the second-best performance is marked with an underline." + }, + { + "type": "table", + "bbox": [ + 0.086, + 0.183, + 0.88, + 0.245 + ], + "angle": 0, + "content": "
FeaturesLogitsAttention
TeacherStudentOFD [10]CRD [23]ReviewKD [2]KD [11]DKD [33]AT [29]CAT-KD
Top-176.1668.8771.2571.3772.5668.5872.0569.5672.24
Top-592.8688.7690.3490.4191.0088.9891.0589.3391.13
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.248, + 0.89, + 0.275 + ], + "angle": 0, + "content": "Table 8. Results on ImageNet. In this group, we set ResNet50 as the teacher and MobileNet as the student. The method with the second-best performance is marked with an underline." + }, + { + "type": "table", + "bbox": [ + 0.102, + 0.298, + 0.446, + 0.455 + ], + "angle": 0, + "content": "
Teacher\nStudentResNet32×4ResNet50
ShuffleNetV1MobileNetV2
DatasetSTLTISTLTI
Baseline69.0536.5464.3930.85
KD [11]66.6132.5667.8132.37
DKD [33]70.7336.7771.0536.48
CRD [23]70.6837.8571.4638.75
ReviewKD [2]71.4638.4666.1632.65
AT [29]71.3637.3665.129.13
CAT-KD74.4340.7373.239.87
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.466, + 0.468, + 0.494 + ], + "angle": 0, + "content": "Table 9. Comparison on transferring representations learned from CIFAR-100 to STL-10 (STL) and Tiny-ImageNet (TI)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.524, + 0.468, + 0.704 + ], + "angle": 0, + "content": "Transferability. We perform experiments to compare the transferability of representations to evaluate the generalizability of the knowledge transferred by various methods. We use ShuffleNetV1 and MobileNetV2 as the frozen representations extractors, which are either trained from scratch on CIFAR-100 [13] or distilled from ResNet \\(32 \\times 4\\) and ResNet50 with various KD methods. Then linear probing tasks are performed on STL-10 [4] and TinyImageNet [5] to quantify their transferability. As the results reported in Table 9, CAT-KD outperforms other methods by a large margin, indicating the outstanding generalizability of the knowledge transferred by our method." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.72, + 0.468, + 0.9 + ], + "angle": 0, + "content": "Efficiency. We first compare the performance of multiple KD methods on CIFAR-100, where the training set is reduced at various ratios, to evaluate their dependence on the amount of training data. As the results reported in Figure 7 (left), CAT-KD is minimally affected by the decrease in the amount of training data, proving the outstanding distillation efficiency of our method. Besides, we further compare the training cost and performance of various KD methods. As reflected in the results reported in Figure 7 (right), CAT-KD has the highest training efficiency. Since CAT-KD does not require extra parameters, its computational cost is almost the same as logits-based" + }, + { + "type": "image", + "bbox": [ + 0.501, + 0.286, + 0.891, + 0.455 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.466, + 0.892, + 0.534 + ], + "angle": 0, + "content": "Figure 7. We set ResNet32×4 as the teacher and ShuffleNetV1 as the student. Left: accuracy of students trained with various methods on CIFAR-100, where the training set is reduced at various ratios. Right: comparison of accuracy and training time (per epoch) on CIFAR-100." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.561, + 0.892, + 0.606 + ], + "angle": 0, + "content": "methods. Relatively, feature-based methods require much more computational resources because most of them need additional auxiliary networks to distill features." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.619, + 0.618, + 0.635 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.644, + 0.892, + 0.78 + ], + "angle": 0, + "content": "In this paper, we propose CAT-KD which has both high interpretability and competitive performance. More importantly, we demonstrate that the capacity of identifying class discriminative regions of input can be obtained and enhanced by transferring CAMs. Furthermore, we present several interesting properties of transferring CAMs, which contribute to a better understanding of CNN. We hope our findings will help future research on the interpretability of CNN and knowledge distillation." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.795, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Acknowledgement. We thank the reviewers for their constructive feedback. Part of Hui Li's work is supported by the National Natural Science Foundation of China (61932015), Shaanxi Innovation Team project (2018TD007), Higher Education Discipline Innovation 111 project (B16037). Part of Haonan Yan's work is done when he visits the University of Guelph." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.956 + ], + "angle": 0, + "content": "11875" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.175, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.116, + 0.47, + 0.184 + ], + "angle": 0, + "content": "[1] Aditya Chattopadhyay, Anirban Sarkar, Prantik Howlader, and Vineeth N Balasubramanian. Grad-cam++: Generalized gradient-based visual explanations for deep convolutional networks. workshop on applications of computer vision, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.187, + 0.472, + 0.243 + ], + "angle": 0, + "content": "[2] Pengguang Chen, Shu Liu, Hengshuang Zhao, and Jiaya Jia. Distilling knowledge via knowledge review. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5008-5017, 2021. 1, 2, 4, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.245, + 0.471, + 0.285 + ], + "angle": 0, + "content": "[3] Jang Hyun Cho and Bharath Hariharan. On the efficacy of knowledge distillation. international conference on computer vision, 2019. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.287, + 0.471, + 0.356 + ], + "angle": 0, + "content": "[4] Adam Coates, Andrew Ng, and Honglak Lee. An analysis of single-layer networks in unsupervised feature learning. In Proceedings of the fourteenth international conference on artificial intelligence and statistics, pages 215-223. JMLR Workshop and Conference Proceedings, 2011. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.358, + 0.47, + 0.412 + ], + "angle": 0, + "content": "[5] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. computer vision and pattern recognition, 2009. 4, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.415, + 0.47, + 0.47 + ], + "angle": 0, + "content": "[6] Tommaso Furlanello, Zachary C. Lipton, Michael Tschannen, Laurent Itti, and Animashree Anandkumar. Born again neural networks. international conference on machine learning, 2018. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.472, + 0.47, + 0.527 + ], + "angle": 0, + "content": "[7] Hao Guo, Kang Zheng, Xiaochuan Fan, Hongkai Yu, and Song Wang. Visual attention consistency under image transforms for multi-label image classification. computer vision and pattern recognition, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.529, + 0.47, + 0.57 + ], + "angle": 0, + "content": "[8] Dongyoon Han, Jiwhan Kim, and Junmo Kim. Deep pyramidal residual networks. computer vision and pattern recognition, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.572, + 0.47, + 0.614 + ], + "angle": 0, + "content": "[9] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. arXiv: Computer Vision and Pattern Recognition, 2015. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.616, + 0.47, + 0.684 + ], + "angle": 0, + "content": "[10] Byeongho Heo, Jeesoo Kim, Sangdoo Yun, Hyojin Park, Nojun Kwak, and Jin Young Choi. A comprehensive overhaul of feature distillation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1921–1930, 2019. 1, 2, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.687, + 0.47, + 0.728 + ], + "angle": 0, + "content": "[11] Geoffrey E. Hinton, Oriol Vinyals, and Jeffrey Dean. Distilling the knowledge in a neural network. arXiv: Machine Learning, 2015. 1, 2, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.73, + 0.47, + 0.771 + ], + "angle": 0, + "content": "[12] Gao Huang, Zhuang Liu, Laurens van der Maaten, and Kilian Q. Weinberger. Densely connected convolutional networks. computer vision and pattern recognition, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.773, + 0.47, + 0.801 + ], + "angle": 0, + "content": "[13] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009. 4, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.803, + 0.47, + 0.844 + ], + "angle": 0, + "content": "[14] Kunpeng Li, Ziyan Wu, Kuan-Chuan Peng, Jan Ernst, and Yun Fu. Tell me where to look: Guided attention inference network. computer vision and pattern recognition, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.846, + 0.469, + 0.899 + ], + "angle": 0, + "content": "[15] Ningning Ma, Xiangyu Zhang, Hai-Tao Zheng, and Jian Sun. Shufflenet v2: Practical guidelines for efficient cnn architecture design. *european conference on computer vision*, 2018. 4" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.116, + 0.472, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.893, + 0.148 + ], + "angle": 0, + "content": "[16] Seyed Iman Mirzadeh, Mehrdad Farajtabar, Ang Li, Nir Levine, Akihiro Matsukawa, and Hassan Ghasemzadeh. Improved knowledge distillation via teacher assistant. national conference on artificial intelligence, 2019. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.149, + 0.893, + 0.19 + ], + "angle": 0, + "content": "[17] Wonpyo Park, Dongju Kim, Yan Lu, and Minsu Cho. Relational knowledge distillation. computer vision and pattern recognition, 2019. 1, 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.192, + 0.893, + 0.247 + ], + "angle": 0, + "content": "[18] Baoyun Peng, Xiao Jin, Dongsheng Li, Shunfeng Zhou, Yichao Wu, Jiaheng Liu, Zhaoning Zhang, and Yu Liu. Correlation congruence for knowledge distillation. international conference on computer vision, 2019. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.249, + 0.893, + 0.29 + ], + "angle": 0, + "content": "[19] Adriana Romero, Nicolas Ballas, Samira Ebrahimi Kahou, Antoine Chassang, Carlo Gatta, and Yoshua Bengio. Fitnets: Hints for thin deep nets. arXiv: Learning, 2014. 1, 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.292, + 0.893, + 0.347 + ], + "angle": 0, + "content": "[20] Mark Sandler, Andrew Howard, Menglong Zhu, Andrey Zhoginov, and Liang-Chieh Chen. *Mobilenetv2: Inverted residuals and linear bottlenecks*. computer vision and pattern recognition, 2018. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.349, + 0.893, + 0.417 + ], + "angle": 0, + "content": "[21] Ramprasaath R. Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra. Grad-cam: Visual explanations from deep networks via gradient-based localization. International Journal of Computer Vision, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.419, + 0.893, + 0.461 + ], + "angle": 0, + "content": "[22] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. computer vision and pattern recognition, 2014. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.462, + 0.893, + 0.503 + ], + "angle": 0, + "content": "[23] Yonglong Tian, Dilip Krishnan, and Phillip Isola. Contrastive representation distillation. arXiv preprint arXiv:1910.10699, 2019. 1, 2, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.504, + 0.893, + 0.545 + ], + "angle": 0, + "content": "[24] Frederick Tung and Greg Mori. Similarity-preserving knowledge distillation. arXiv: Computer Vision and Pattern Recognition, 2019. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.547, + 0.893, + 0.603 + ], + "angle": 0, + "content": "[25] Chaofei Wang, Jiayu Xiao, Yizeng Han, Qisen Yang, Shiji Song, and Gao Huang. Towards learning spatially discriminative feature representations. international conference on computer vision, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.604, + 0.893, + 0.659 + ], + "angle": 0, + "content": "[26] Haofan Wang, Zifan Wang, Mengnan Du, Fan Yang, Zijian Zhang, Sirui Ding, Piotr Mardziel, and Xia Hu. Score-cam: Score-weighted visual explanations for convolutional neural networks. computer vision and pattern recognition, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.66, + 0.893, + 0.716 + ], + "angle": 0, + "content": "[27] Seunghan Yang, YoonHyung Kim, Youngeun Kim, and Changick Kim. Combinational class activation maps for weakly supervised object localization. workshop on applications of computer vision, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.717, + 0.893, + 0.773 + ], + "angle": 0, + "content": "[28] Junho Yim, Donggyu Joo, Ji-Hoon Bae, and Junmo Kim. A gift from knowledge distillation: Fast optimization, network minimization and transfer learning. computer vision and pattern recognition, 2017. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.774, + 0.893, + 0.828 + ], + "angle": 0, + "content": "[29] Sergey Zagoruyko and Nikos Komodakis. Paying more attention to attention: improving the performance of convolutional neural networks via attention transfer. Learning, 2016, 1, 2, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.83, + 0.893, + 0.858 + ], + "angle": 0, + "content": "[30] Sergey Zagoruyko and Nikos Komodakis. Wide residual networks. *british machine vision conference*, 2016. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.86, + 0.893, + 0.902 + ], + "angle": 0, + "content": "[31] Xiaopeng Zhang, Jiashi Feng, Hongkai Xiong, and Qi Tian. Zigzag learning for weakly supervised object detection. computer vision and pattern recognition, 2018. 3" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.893, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "11876" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.092, + 0.47, + 0.147 + ], + "angle": 0, + "content": "[32] Xiangyu Zhang, Xinyu Zhou, Mengxiao Lin, and Jian Sun. Shufflenet: An extremely efficient convolutional neural network for mobile devices. computer vision and pattern recognition, 2017. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.149, + 0.47, + 0.19 + ], + "angle": 0, + "content": "[33] Borui Zhao, Quan Cui, Renjie Song, Yiyu Qiu, and Jiajun Liang. Decoupled knowledge distillation. arXiv preprint arXiv:2203.08679, 2022. 1, 2, 4, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.192, + 0.469, + 0.26 + ], + "angle": 0, + "content": "[34] Bolei Zhou, Aditya Khosla, Agata Lapedriza, Aude Oliva, and Antonio Torralba. Learning deep features for discriminative localization. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2921-2929, 2016. 2, 3, 5" + }, + { + "type": "list", + "bbox": [ + 0.079, + 0.092, + 0.47, + 0.26 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.956 + ], + "angle": 0, + "content": "11877" + } + ] +] \ No newline at end of file diff --git a/2023/Class Attention Transfer Based Knowledge Distillation/a68b12fe-6c62-4a74-9758-0370df04b2ad_origin.pdf b/2023/Class Attention Transfer Based Knowledge Distillation/a68b12fe-6c62-4a74-9758-0370df04b2ad_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..42aab2a13b43406f31ee68960365e52d889106d3 --- /dev/null +++ b/2023/Class Attention Transfer Based Knowledge Distillation/a68b12fe-6c62-4a74-9758-0370df04b2ad_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c4b6349645b6910642bf6f8679f9ddc97e857150a287e42ef1085d67dfb74a11 +size 2221623 diff --git a/2023/Class Attention Transfer Based Knowledge Distillation/full.md b/2023/Class Attention Transfer Based Knowledge Distillation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..1ff9cb9879a35ab49b903ae22e25e6b68bda3ce2 --- /dev/null +++ b/2023/Class Attention Transfer Based Knowledge Distillation/full.md @@ -0,0 +1,306 @@ +# Class Attention Transfer Based Knowledge Distillation + +Ziyao Guo $^{1}$ , Haonan Yan $^{1,2,*}$ , Hui Li $^{1,*}$ , Xiaodong Lin $^{2}$ $^{1}$ Xidian University, $^{2}$ University of Guelph + +gzyaftermath@outlook.com + +# Abstract + +Previous knowledge distillation methods have shown their impressive performance on model compression tasks, however, it is hard to explain how the knowledge they transferred helps to improve the performance of the student network. In this work, we focus on proposing a knowledge distillation method that has both high interpretability and competitive performance. We first revisit the structure of mainstream CNN models and reveal that possessing the capacity of identifying class discriminative regions of input is critical for CNN to perform classification. Furthermore, we demonstrate that this capacity can be obtained and enhanced by transferring class activation maps. Based on our findings, we propose class attention transfer based knowledge distillation (CAT-KD). Different from previous KD methods, we explore and present several properties of the knowledge transferred by our method, which not only improve the interpretability of CAT-KD but also contribute to a better understanding of CNN. While having high interpretability, CAT-KD achieves state-of-the-art performance on multiple benchmarks. Code is available at: https://github.com/GzyAftermath/CAT-KD. + +# 1. Introduction + +Knowledge distillation (KD) transfers knowledge distilled from the bigger teacher network to the smaller student network, aiming to improve the performance of the student network. Depending on the type of the transferred knowledge, previous KD methods can be divided into three categories: based on transferring logits [3, 6, 11, 16, 33], features [2, 10, 17-19, 23, 24, 28], and attention [29]. Although KD methods that are based on transferring logits and features have shown their promising performance [2, 33], it is hard to explain how the knowledge they transferred helps to improve the performance of the student network, due to the uninterpretability of logits and features. Relatively, the principle of attention-based KD methods is more intuitive: + +![](images/31922e3db0993b7355024d0fc3ec40f96a6e3808c2348f59e5eb0e742ad2e2d5.jpg) +Figure 1. Illustration of the converted structure. After converting the FC layer into a convolutional layer with $1 \times 1$ kernel and moving the position of the global average pooling layer, CAMs can be obtained during the forward propagation. + +it aims at telling the student network which part of the input should it focus on during the classification, which is realized by forcing the student network to mimic the transferred attention maps during training. However, though previous work AT [29] has validated the effectiveness of transferring attention, it does not present what role attention plays during the classification. This makes it hard to explain why telling the trained model where should it focus could improve its performance on the classification mission. Besides, the performance of the previous attention-based KD method [29] is less competitive compared with the methods that are based on transferring logits and features [2, 33]. In this work, we focus on proposing an attention-based KD method that has higher interpretability and better performance. + +![](images/99ad99410587f8d2adc94e067252483f53781f197df3f22eaa26bbb23742dc9a.jpg) +Figure 2. Visualization of CAMs corresponding to categories with Top 4 prediction scores for the given image. The predicted categories and their scores are reported in the picture. + +![](images/bf4fc73385716d74922ecbed962a8fb5843852c6822c3c2c521099929f37fe90.jpg) + +![](images/45caf93496e785cc702fbe528421ea023f2d02ca15a5f732fe9b28068900f7bc.jpg) + +![](images/0ddae377e3ec9301f1704fb9d4de7cb24811433c83c1ea10d29270b1856b8ac1.jpg) + +![](images/24231d24833314bc016b0d5355521e61a0f3d7342009bed46403c761b82b4db8.jpg) + +We start our work by exploring what role attention plays during classification. After revisiting the structure of the mainstream models, we find that with a little conversion (illustrated in Figure 1), class activation map (CAM) [34], a kind of class attention map which indicates the discriminative regions of input for a specific category, can be obtained during the classification. Without changing the parameters and outputs, the classification process of the converted model can be viewed in two steps: (1) the model exploits its capacity to identify class discriminative regions of input and generate CAM for each category contained in the classification mission, (2) the model outputs the prediction score of each category by computing the average activation of the corresponding CAM. Considering that the converted model makes predictions by simply comparing the average activation of CAMs, possessing the capacity to identify class discriminative regions of input is critical for CNN to perform classification. The question is: can we enhance this capacity by offering hints about class discriminative regions of input during training? To answer this question, we propose class attention transfer (CAT). + +During CAT, the trained model is not required to predict the category of input, it is only forced to mimic the transferred CAMs, which are normalized to ensure they only contain hints about class discriminative regions of input. Through experiments with CAT, we reveal that transferring only CAMs can train a model with high accuracy on the classification task, reflecting the trained model obtains the capacity to identify class discriminative regions of input. Besides, the performance of the trained model is influenced by the accuracy of the model offering the transferred CAMs. This further demonstrates that the capacity of identifying class discriminative regions can be enhanced by transferring more precise CAMs. + +Based on our findings, we propose class attention transfer based knowledge distillation (CAT-KD), aiming to enable the student network to achieve better performance by improving its capacity of identifying class discriminative regions. Different from previous KD methods transferring dark knowledge, we present why transferring CAMs to the trained model can improve its performance on the classification task. Moreover, through experiments with CAT, we reveal several interesting properties of transferring CAMs, + +which not only help to improve the performance and interpretability of CAT-KD but also contribute to a better understanding of CNN. While having high interpretability, CAT-KD achieves state-of-the-art performance on multiple benchmarks. Overall, the main contributions of our work are shown below: + +- We propose class attention transfer and use it to demonstrate that the capacity of identifying class discriminative regions of input, which is critical for CNN to perform classification, can be obtained and enhanced by transferring CAMs. +- We present several interesting properties of transferring CAMs, which contribute to a better understanding of CNN. +- We apply CAT to knowledge distillation and name it CAT-KD. While having high Interpretability, CAT-KD achieves state-of-the-art performance on multiple benchmarks. + +# 2. Background + +The concept of knowledge distillation was proposed in [11]. As a transfer learning method, KD aims to improve the performance of the smaller student network by transferring the dark knowledge distilled from the bigger teacher network. Previous KD methods can be divided into three types: distillation from logits [3, 6, 11, 16, 33], features [2, 10, 17-19, 23, 24, 28] and attention [29]. + +To our knowledge, AT [29] is the only KD method based on transferring attention, which defines attention map as the spatial map indicating the area of input that the model focuses on most. In practice, they obtain attention maps by calculating the sum of feature maps while their values are absolutized. However, AT did not present what role attention plays during the classification and why transferring attention maps defined in this way can improve the performance of the student network. + +Previous works related to class attention originate from [34], where the authors propose to utilize high-level feature maps and the parameters of the fully connected layer to generate attention map for a specific category, which is named + +class activation map (CAM). According to [34], class discriminative regions of input are highlighted in the corresponding CAM. To facilitate understanding, we visualize several CAMs in Figure 2. The following works have successfully applied CAM in various weakly supervised visual tasks [14,27,31]. Besides, there are also many works focus on generalizing CAM [1,21,26] and improving the performance of models by exploiting the information contained in CAM during training [7,25]. + +Previous works have not presented what role attention plays during classification and why transferring attention maps can improve the trained model's performance on the classification mission. In this paper, we focus on figuring out this question and try to propose an attention-based KD method that has both high interpretability and competitive performance. + +# 3. Our Method + +In this section, we first analyze the structure of the mainstream CNN models and reveal that possessing the capacity of identifying class discriminative regions is critical for CNN to perform classification. Then we further propose class attention transfer to prove that this capacity can be obtained and enhanced by transferring CAMs. Finally, we apply CAT to knowledge distillation. + +# 3.1. Revisit the structure of CNN + +In image classification tasks, mainstream models usually use CNN to extract features, the resulting high-level feature maps are then globally pooled and fed to a simple fully connected layer to perform classification [8,9,12]. Let $\mathbf{F} = [F_1,F_2,\dots,F_C]\in \mathbb{R}^{C\times W\times H}$ represents the feature maps generated by the last convolutional layer, where $C$ , $W$ , and $H$ indicate channel dimension, width, and height respectively. And $f_{j}(x,y)$ denotes the activation of $\mathbf{F}$ in $j$ channel at spatial location $(x,y)$ , while $GAP$ is the global average pooling layer. Then the process of calculating logits for normal CNN models can be written as: + +$$ +\begin{array}{l} L _ {i} = \sum_ {1 \leq j \leq C} \omega_ {j} ^ {i} \times G A P (F _ {j}) \\ = \frac {1}{W \times H} \sum_ {x, y} \sum_ {1 \leq j \leq C} \omega_ {j} ^ {i} \times f _ {j} (x, y), \tag {1} \\ \end{array} +$$ + +where $L_{i}$ denotes the logit of $i$ -th class, $\omega_{j}^{i}$ is the weight of the fully connected layer (FC layer) corresponding to class $i$ for $GAP(F_{j})$ . According to [34], we can obtain the CAM corresponding to category $i$ by: + +$$ +C A M _ {i} (x, y) = \sum_ {1 \leq j \leq C} \omega_ {j} ^ {i} \times f _ {j} (x, y). \tag {2} +$$ + +![](images/1cedc01fa88552c4cbedf18a5871b9b87116bb34dd9d67c06402451fd18bd291.jpg) +Figure 3. Illustration of CAT. During CAT, the structure of teacher and student are converted to our style (Figure 1). + +According to Equation (1) and Equation (2), the calculation of $L_{i}$ can be written in another form: + +$$ +\begin{array}{l} L _ {i} = \frac {1}{W \times H} \sum_ {x, y} C A M _ {i} (x, y) \tag {3} \\ = G A P (C A M _ {i}). \\ \end{array} +$$ + +As reflected in Equation (3), logits can be obtained by computing the average activation of CAMs. Inspired by it, as illustrated in Figure 1, we convert the FC layer into a $1 \times 1$ convolutional layer and move the position of the GAP layer. Then $\bar{L}_i$ , the logit of $i$ -th class generated by the converted model, can be obtained by: + +$$ +\begin{array}{l} \bar {L} _ {i} = G A P \left(\operatorname {C o n v} _ {i} (\mathbf {F})\right) \\ = \frac {1}{W \times H} \sum_ {x, y} \left(\sum_ {1 \leq j \leq C} \omega_ {j} ^ {i} \times f _ {j} (x, y)\right) \tag {4} \\ = G A P \left(C A M _ {i}\right), \\ \end{array} +$$ + +where $Conv_{i}$ denotes the converted $1 \times 1$ convolution kernel that used to separate features corresponding to $i$ -th class from $\mathbf{F}$ , and $\omega_{j}^{i}$ is its weight of $j$ channel. As reflected in Eqn(3) and Eqn(4), the conversion does not change the value of its prediction score (i.e., logits). And class activation maps can be obtained during the classification of the converted model. + +As reflected in Eqn(4), the classification process of the converted model can be viewed in two steps: (1) the model exploits its capacity to identify class discriminative regions of input and generate CAMs, (2) the model outputs prediction score of each category by computing the average activation of the corresponding CAM. Considering that the model makes predictions by simply comparing the average activation of CAMs, possessing the capacity to identify class discriminative regions of input is critical for CNN to perform classification. To examine if this capacity can be obtained and enhanced by offering hints indicating class discriminative regions of input to the trained model, we propose class attention transfer. + +# 3.2. Class Attention Transfer + +The purpose of CAT is to examine if a model can obtain the capacity to identify class discriminative regions of input by transferring only CAMs. Thus, during CAT, the trained model is not required to perform classification, and any information related to the category of the training set data (e.g., ground-truth labels and logits) is not released to the trained model. In practice, we utilize a pre-trained model with the converted structure to generate the transferred CAMs. The illustration of the process of CAT is shown in Figure 3, while the formal description is shown below. + +For a given input, let $\mathbf{A} \in \mathbb{R}^{K \times W \times H}$ denotes the CAMs generated by the converted structure, where $K$ is the number of categories contained in the classification task, $W$ and $H$ denote the width and height of the generated CAM respectively. $A_{i} \in \mathbb{R}^{W \times H}$ represents the $i$ channel of $\mathbf{A}$ , which is the CAM corresponding to category $i$ . And $S, T$ denote student and teacher correspondingly. Besides, we use the average pooling function $\phi$ to reduce the resolution of the transferred CAMs, to improve the performance of CAT (Section 4.2). Then CAT's loss function can be defined as: + +$$ +\mathcal {L} _ {C A T} = \sum_ {1 \leq i \leq K} \frac {1}{K} \| \frac {\phi \left(A _ {i} ^ {T}\right)}{\| \phi \left(A _ {i} ^ {T}\right) \| _ {2}} - \frac {\phi \left(A _ {i} ^ {S}\right)}{\| \phi \left(A _ {i} ^ {S}\right) \| _ {2}} \| _ {2} ^ {2}. \tag {5} +$$ + +As can be seen, we perform $l_{2}$ normalization on $\phi(A_i^T)$ and $\phi(A_i^S)$ ( $l_{1}$ normalization can be used as well), to ensure that information related to the category of input is not released to the trained model during CAT, considering that the average activation of CAM indicates the prediction score (Equation (3)). Besides, note that here we transfer CAMs of all categories, which is based on our finding that CAMs of all categories both contain beneficial information for CAT (Section 4.2). + +Our core findings through the experiments with CAT are presented as follows, while the corresponding experimental verification and their detailed analysis can be found in Section 4.2. + +- The capacity to identify class discriminative regions of input can be obtained and enhanced by transferring CAMs. +- CAMs of all categories both contain beneficial information for CAT. +- Transferring smaller CAMs performs better. +- For CAT, the critical information contained in the transferred CAMs is the spatial location of the regions with high activation in them rather than their specific value. + +# 3.3. CAT-KD + +After validating the effectiveness of CAT, we apply CAT to knowledge distillation and name it CAT-KD. The loss function of CAT-KD is: + +$$ +\mathcal {L} _ {K D} = \mathcal {L} _ {C E} + \beta \mathcal {L} _ {C A T}, \tag {6} +$$ + +where $\mathcal{L}_{CE}$ denotes the standard cross-entropy loss, and $\beta$ is the factor used to balance the CE loss and CAT loss. + +Different from previous KD methods, we present how the knowledge transferred by CAT-KD helps to improve the performance of the student network: by improving its capacity of identifying class discriminative regions. Besides, through experiments with CAT, we analyze and reveal several properties of the knowledge transferred by our method. This further enhances the interpretability of CAT-KD. + +# 4. Experiments + +# 4.1. Datasets and Implementation Details + +Datasets. In the following section we explore CAT and CAT-KD mainly on two image classification datasets: + +(1) CIFAR-100 [13] comprise $32 \times 32$ pixel images of 100 categories, the training and validate sets contain 50K and 10K images. +(2) ImageNet [5] is a large-scale dataset for the classification of 1K categories, containing 1.2 million training and 50K validation images. + +Implementation details. Our implementation for CIFAR-100 and ImageNet strictly follows [2, 33]. Specifically, for CIFAR-100, we train all models for 240 epochs with batch size 64 using SGD. The initial learning rate is 0.05 (0.01 for ShuffleNet [15, 32] and MobileNet [20]), divided by 10 at 150, 180, and 210 epochs. For ImageNet, we train models for 100 epochs with batch size 512. The initial learning rate is 0.2 and divided by 10 for every 30 epochs. We experiment with various representative CNN network: VGG [22], ResNet [9], WideResNet [30], MobileNet [20], and ShuffleNet [15, 32]. + +For fairness, all the results of previous methods are either reported in previous papers [2,33] (we keep our training setting the same as theirs) or obtained using codes released by the author with our training setting. All results on CIFAR-100 are the average over 5 trials, while that on ImageNet is the average over 3 trials. + +For all experiments reported in Section 4.2 and Section 4.3, without special specification, we pool the transferred CAMs into $2 \times 2$ during CAT and CAT-KD. More implementation details such as the settings of $\beta$ are attached in the appendix due to the page limits. + +![](images/6652ce4e1cdeeb81801e302608654780d43980f597ba701148f54f6465f2fe6d.jpg) +Figure 4. Accuracy of models trained with CAT on CIFAR-100. Left: Only CAMs of certain categories are transferred, which are selected by two strategies: (1) select categories with top n prediction scores, (2) select categories with the lowest n prediction scores. Right: The training set is reduced to contain data of partial categories only. T: test set of CIFAR-100. S: a subset of T which only contains data of classes that are not contained in the training set. + +# 4.2. Exploration of CAT + +In this section, we explore several properties of class attention transfer, which not only help to improve the performance and interpretability of CAT-KD but also contributes to a better understanding of CNN. Note that any information related to the category of the training set (e.g., ground-truth labels and logits) is not utilized in the experiments reported in this section. + +The capacity of identifying class discriminative regions can be obtained and enhanced by transferring CAMs. As revealed in Section 3.1, being able to identify class discriminative regions of input is critical for CNN to perform classification. Thus, the intensity of this capacity can be evaluated by the model's performance on the classification mission. We perform CAT on ShuffleNetV1, where the transferred CAMs are produced by different models with various accuracy. As the results reported in Table 1, transferring only CAMs can train a model with high accuracy on the classification mission, proving the capacity of identifying class discriminative regions can be obtained by transferring CAMs. Besides, the performance of the trained model is influenced by the accuracy of the model producing the transferred CAMs, indicating that this capacity can be enhanced by transferring more precise CAMs. + +CAMs of all categories both contain beneficial information for CAT. For a given input, we can use the method of CAM [34] to generate class activation maps for any categories contained in the classification mission. However, though a few non-target categories may share certain similarities (e.g., shape and patterns) with the target + +![](images/c5bcf1d0e42e4df24c38d9f7174afb12ef69db638438178f2952e3c8aeadd814.jpg) + +![](images/34a5b85f0474a65bdbced53260aa9381525d84fe82eb156cbe1a9bcf5cf51662.jpg) +Top1 prairie fowl + +![](images/54791102e3cd1e719f415854c16fb9a87a6368a066c7fb428fd10cc4371576af.jpg) + +![](images/6414cdd8ada81c8a2f4772c71b1bceea7afeb90737ee054c05097ad181744896.jpg) +Top1 otter + +![](images/3a73e068b6b6491437ece202e490c922a5599cd02ec1c2a7d53283373e7dbadb.jpg) + +![](images/8fab1b0c00593f69db1c2c5c834d8779cf84c11d62abf8cdec8c13f1c96c2486.jpg) +Top1basset +Figure 5. We set a pre-trained ResNet50 as CAMs producer to train another ResNet50 from scratch with CAT, CAMs are pooled into $2 \times 2$ during the transfer. The first row shows the visualization of the CAMs generated by the producer, while the CAMs visualized in the second row come from the trained model. + +category, most of them are completely irrelevant to the input from a human understanding. However, our experiments show that class activation maps of all categories both contain beneficial information for CAT. + +We first perform CAT on CIFAR-100 where only CAMs of certain categories are transferred. We designed two strategies to select the categories of the transferred CAMs: (1) select categories with the lowest n prediction scores. (2) select categories with top n prediction scores (the empirical assumption here we make is that the categories with higher prediction scores have more similarities with the target category). As the results reported in Figure 4 (left), while CAMs of classes with higher prediction scores bring more improvement, others are also beneficial for CAT. Besides, we further perform CAT on the reduced CIFAR-100, where CAMs of all classes are transferred but the training set is reduced to contain data of only partial categories. Then the trained model is evaluated on the complete test set and a subset of it which only contains data of classes that are not contained in the training set. As the results reported in Figure 4 (right), interestingly, the trained model achieves high accuracy on the subset, indicating that transferring CAMs enables the trained model to classify the categories that are not contained in the training set. This further proves that non-target CAMs contain beneficial information for CAT even if their categories seem to be irrelevant to the input from a human perspective. + +Transferring smaller CAMs performs better. Intuitively, larger CAM contains more detailed hints about the spatial location of the class discriminative regions, then transferring larger CAMs should perform better. However, insufficient accuracy of the model will result in deviations + +![](images/bdcc80d5c9958158942304ab20cf6ffaaa090671c389969a3c5c526e2303671e.jpg) +Top1 bee + +![](images/7cc4cfcf97f183829b521ccdc1a80400ef3df2f862d4f36ae39a034fbc3eec1e.jpg) +Top2 sunflower + +![](images/d85cffa7978b89543896f630f2cfd43d950bcddb42402095332d06088c3609ff.jpg) +Top3 tiger + +![](images/4b6e372d959922e6633655735a036a45cc23195774abfb074ed1cfb8920ecb0a.jpg) +Top1 kangaroo +Figure 6. The first row shows the visualization of the CAMs corresponding to the top 3 predicted categories, while the following row shows the visualization of them after binarization. + +![](images/14bf39c2fff11381fa21e872653c3324b1dbe7ca968954e999dd25ab058802ae.jpg) +Top2 rabbit + +![](images/c87eb7d030dd71e46b5ecf56e77625fa0774d74ac29d3f641c787027a714fc2e.jpg) +Top3 fox + +![](images/f44d0fca783d9d4bc5bb7c17fd524620774a16d472a5ab0b13a1163110c458f5.jpg) +Top1 bear + +![](images/af8f982287d31039a84b247c226a44bf9899b028e2957c2f7f7ad36f357947fc.jpg) +Top2 cattle + +![](images/84805156fe67564ac1b1337fb9ce162b99295c718e83fe63192682fa2c530ff2.jpg) +Top3 otter + +
CAM ProducerResNet56ResNet110ResNet50
Acc72.3474.3179.34
CAT72.4774.4276.17
+ +Table 1. Accuracy $(\%)$ of ShuffleNetV1 trained with CAT on CIFAR-100. The transferred CAMs are produced by different models with various accuracy. + +
BaselineModel AccResNet32×4 79.42ResNet8×4 72.5ResNet20 69.06
CAT8×879.6567.9266.21
4×479.8471.6166.43
2×279.7172.4566.84
+ +between the highlighted areas in its generated CAM and the actual class discriminative regions of the image (which can be observed in Figure 2). Besides, different models differ in their capacity to identify class discriminative regions, which will lead to subtle differences in the generated CAMs. Therefore transferring CAMs with a larger size does not necessarily improve the performance of CAT. Through experiments, we found that performing average pooling on the transferred CAMs, which will expand the highlighted areas of CAMs and reduce the bias between CAMs generated by different models, could alleviate the above issues. As the results reported in Table 2, though pooling blurs the details, transferring smaller CAMs always performs better. Besides, since the pooling operation expands the highlighted areas of CAM, which will make it encompass larger class discriminative regions, transferring pooled CAMs will force the trained model to pay attention to more discriminative regions, which can be observed in Figure 5. In practice, we pool the transferred CAMs into a smaller size to improve the performance of CAT and CAT-KD (normally $2 \times 2$ ). + +Table 2. Accuracy $(\%)$ of various models trained with CAT on the CIFAR-100 test set. During CAT, CAMs are pooled into various sizes. The transferred CAMs are produced by ResNet32 $\times$ 4. + +
BaselineModelResNet32×4ResNet50
Acc79.4279.34
CATCAMs79.7180.45
Binarized CAMs79.3579.65
+ +Table 3. Results of transferring binarized CAMs. The transferred CAMs are produced by ResNet $32 \times 4$ . + +
TeacherResNet32×4
Acc77.5179.4281.36
ReviewKD [2]76.4277.4577.91
DKD [33]76.5876.4577.29
CAT-KD76.3678.2678.84
Δ-0.22+0.81+0.93
+ +Table 4. Comparison with two SOTA methods. The student network is ShuffleNetV1. $\Delta$ represents the gap between CAT-KD and the best-performing method among ReviewKD and DKD (marked with underline). + +The exact value of the transferred CAMs is not important. To demonstrate that the role CAMs play in CAT is offering hints about the spatial location of the class discriminative regions of input, we binarize the values of the transferred CAMs to 0 and 1, using their average values as the thresholds. The regions of CAM with values above the threshold are considered as being highlighted, indicating the class discriminative regions of input. Thus, we set the values of these regions to 1 to keep them activated after the binarization. Other regions with values below the threshold are considered unhighlighted, and their values are set to 0. As shown in Figure 6, though the specific values of CAMs are lost during the binarization process, the binarized CAMs still contain hints about the spatial location of the class discriminative regions. Note that the threshold can also be specified in other ways (e.g., median). + +As the results reported in Table 3, although the class discriminative regions obtained by our rudimentary binarization method are not precise, the accuracy of the resulting model dropped by less than one percent, proving that the + +
Distillation MechanismTeacher AccResNet32×4 79.42WRN40-2 75.61ResNet32×4 79.42ResNet50 79.34VGG13 74.64
Student AccShuffleNetV1 70.5ShuffleNetV1 70.5ShuffleNetV2 71.82MobileNetV2 64.6MobileNetV2 64.6
LogitsKD [11]74.0774.8374.4567.3567.37
DKD [33]76.4576.777.0770.3569.71
FeaturesCRD [23]75.1176.0575.6569.1169.73
OFD [10]75.9875.8576.8269.0469.48
FitNet [19]73.5973.7373.5463.1664.14
RKD [17]72.2872.2173.2164.4364.52
ReviewKD [2]77.4577.1477.7869.8970.37
AttentionAT [29]71.7373.3272.7358.5859.4
CAT-KD78.2677.3578.4171.3669.13
↑+6.53+4.03+5.68+12.78+9.73
+ +Table 5. Results on CIFAR-100. Teachers and students have different architectures. $\uparrow$ represents the performance improvement of CAT-KD compared with AT. + +
Distillation MechanismTeacher AccResNet56 72.34ResNet110 74.31ResNet32×4 79.42WRN-40-2 75.61WRN-40-2 75.61VGG13 74.64
Student AccResNet20 69.06ResNet32 71.14ResNet8×4 72.5WRN-16-2 73.26WRN-40-1 71.98VGG8 70.36
LogitsKD [11]70.6673.0873.3374.9273.5472.98
DKD [33]71.9774.1176.3276.2474.8174.68
FeaturesCRD [23]71.1673.4875.5175.4874.1473.94
OFD [10]70.9873.2374.9575.2474.3373.95
FitNet [19]69.2171.0673.573.5872.2471.02
RKD [17]69.6171.8271.973.3572.2271.48
ReviewKD [2]71.8973.8975.6376.1275.0974.84
AttentionAT [29]70.5572.3173.4474.0872.7771.43
CAT-KD71.6273.6276.9175.674.8274.65
↑+1.07+1.31+3.47+1.52+2.05+3.22
+ +Table 6. Results on CIFAR-100. Teachers and students have the same architecture. $\uparrow$ represents the performance improvement of CAT-KD compared with AT. + +critical information CAMs contained for CAT is the spatial location of class discriminative regions rather than its exact value. This strongly demonstrates that our method is based on transferring attention. + +# 4.3. Evaluation of CAT-KD + +Consistent with previous works [2, 23, 33], we compare the performance of CAT-KD with several representative KD methods. Moreover, we further evaluate our method from two aspects: transferability and efficiency. + +Results on CIFAR-100. Table 5 reports the results on CIFAR-100 with the teachers and students having different architectures. Table 6 shows the results where teachers and students have architectures of the same style. Notably, our method outperforms the other attention-based method AT [29] with a large margin (1.07% ~ 12.78%). Moreover, CAT-KD achieves comparable or + +even better performance compared with feature-based distillation method [2] which requires additional networks and multiple-layer information. Besides, consistent with CAT, the performance of CAT-KD is affected by the accuracy of the teacher: CAMs produced by teachers with lower accuracy contain more incorrect hints about the class discriminative regions of input. To verify this, we further evaluate the impact of the accuracy of the teacher on our method. As the results reported in Table 4, CAT-KD is relatively less effective when the teacher is weak. Thus, as can be observed in Table 6, the performance of CAT-KD is not the best when the teacher is weak. + +Results on ImageNet. Table 7 and Table 8 report the top-1 and top-5 accuracy of image classification on ImageNet. Though the performance of CAT-KD is restricted by the weakness of the teacher network in this setting, our method still outperforms most KD methods. + +
FeaturesLogitsAttention
TeacherStudentOFD [10]CRD [23]ReviewKD [2]KD [11]DKD [33]AT [29]CAT-KD
Top-173.3169.7570.8171.1771.6170.6671.770.6971.26
Top-591.4189.0789.9890.1390.5189.8890.4190.0190.45
+ +Table 7. Results on ImageNet. In this group, we set ResNet34 as the teacher and ResNet18 as the student. The method with the second-best performance is marked with an underline. + +
FeaturesLogitsAttention
TeacherStudentOFD [10]CRD [23]ReviewKD [2]KD [11]DKD [33]AT [29]CAT-KD
Top-176.1668.8771.2571.3772.5668.5872.0569.5672.24
Top-592.8688.7690.3490.4191.0088.9891.0589.3391.13
+ +Table 8. Results on ImageNet. In this group, we set ResNet50 as the teacher and MobileNet as the student. The method with the second-best performance is marked with an underline. + +
Teacher +StudentResNet32×4ResNet50
ShuffleNetV1MobileNetV2
DatasetSTLTISTLTI
Baseline69.0536.5464.3930.85
KD [11]66.6132.5667.8132.37
DKD [33]70.7336.7771.0536.48
CRD [23]70.6837.8571.4638.75
ReviewKD [2]71.4638.4666.1632.65
AT [29]71.3637.3665.129.13
CAT-KD74.4340.7373.239.87
+ +Table 9. Comparison on transferring representations learned from CIFAR-100 to STL-10 (STL) and Tiny-ImageNet (TI). + +Transferability. We perform experiments to compare the transferability of representations to evaluate the generalizability of the knowledge transferred by various methods. We use ShuffleNetV1 and MobileNetV2 as the frozen representations extractors, which are either trained from scratch on CIFAR-100 [13] or distilled from ResNet $32 \times 4$ and ResNet50 with various KD methods. Then linear probing tasks are performed on STL-10 [4] and TinyImageNet [5] to quantify their transferability. As the results reported in Table 9, CAT-KD outperforms other methods by a large margin, indicating the outstanding generalizability of the knowledge transferred by our method. + +Efficiency. We first compare the performance of multiple KD methods on CIFAR-100, where the training set is reduced at various ratios, to evaluate their dependence on the amount of training data. As the results reported in Figure 7 (left), CAT-KD is minimally affected by the decrease in the amount of training data, proving the outstanding distillation efficiency of our method. Besides, we further compare the training cost and performance of various KD methods. As reflected in the results reported in Figure 7 (right), CAT-KD has the highest training efficiency. Since CAT-KD does not require extra parameters, its computational cost is almost the same as logits-based + +![](images/ac6edc14c512022db031a55c2c7ff30d9595670adf1d2157705fc3050e50188d.jpg) +Figure 7. We set ResNet32×4 as the teacher and ShuffleNetV1 as the student. Left: accuracy of students trained with various methods on CIFAR-100, where the training set is reduced at various ratios. Right: comparison of accuracy and training time (per epoch) on CIFAR-100. + +methods. Relatively, feature-based methods require much more computational resources because most of them need additional auxiliary networks to distill features. + +# 5. Conclusion + +In this paper, we propose CAT-KD which has both high interpretability and competitive performance. More importantly, we demonstrate that the capacity of identifying class discriminative regions of input can be obtained and enhanced by transferring CAMs. Furthermore, we present several interesting properties of transferring CAMs, which contribute to a better understanding of CNN. We hope our findings will help future research on the interpretability of CNN and knowledge distillation. + +Acknowledgement. We thank the reviewers for their constructive feedback. Part of Hui Li's work is supported by the National Natural Science Foundation of China (61932015), Shaanxi Innovation Team project (2018TD007), Higher Education Discipline Innovation 111 project (B16037). Part of Haonan Yan's work is done when he visits the University of Guelph. + +# References + +[1] Aditya Chattopadhyay, Anirban Sarkar, Prantik Howlader, and Vineeth N Balasubramanian. Grad-cam++: Generalized gradient-based visual explanations for deep convolutional networks. workshop on applications of computer vision, 2018. 3 +[2] Pengguang Chen, Shu Liu, Hengshuang Zhao, and Jiaya Jia. Distilling knowledge via knowledge review. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5008-5017, 2021. 1, 2, 4, 6, 7, 8 +[3] Jang Hyun Cho and Bharath Hariharan. On the efficacy of knowledge distillation. international conference on computer vision, 2019. 1, 2 +[4] Adam Coates, Andrew Ng, and Honglak Lee. An analysis of single-layer networks in unsupervised feature learning. In Proceedings of the fourteenth international conference on artificial intelligence and statistics, pages 215-223. JMLR Workshop and Conference Proceedings, 2011. 8 +[5] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. computer vision and pattern recognition, 2009. 4, 8 +[6] Tommaso Furlanello, Zachary C. Lipton, Michael Tschannen, Laurent Itti, and Animashree Anandkumar. Born again neural networks. international conference on machine learning, 2018. 1, 2 +[7] Hao Guo, Kang Zheng, Xiaochuan Fan, Hongkai Yu, and Song Wang. Visual attention consistency under image transforms for multi-label image classification. computer vision and pattern recognition, 2019. 3 +[8] Dongyoon Han, Jiwhan Kim, and Junmo Kim. Deep pyramidal residual networks. computer vision and pattern recognition, 2016. 3 +[9] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. arXiv: Computer Vision and Pattern Recognition, 2015. 3, 4 +[10] Byeongho Heo, Jeesoo Kim, Sangdoo Yun, Hyojin Park, Nojun Kwak, and Jin Young Choi. A comprehensive overhaul of feature distillation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1921–1930, 2019. 1, 2, 7, 8 +[11] Geoffrey E. Hinton, Oriol Vinyals, and Jeffrey Dean. Distilling the knowledge in a neural network. arXiv: Machine Learning, 2015. 1, 2, 7, 8 +[12] Gao Huang, Zhuang Liu, Laurens van der Maaten, and Kilian Q. Weinberger. Densely connected convolutional networks. computer vision and pattern recognition, 2016. 3 +[13] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009. 4, 8 +[14] Kunpeng Li, Ziyan Wu, Kuan-Chuan Peng, Jan Ernst, and Yun Fu. Tell me where to look: Guided attention inference network. computer vision and pattern recognition, 2018. 3 +[15] Ningning Ma, Xiangyu Zhang, Hai-Tao Zheng, and Jian Sun. Shufflenet v2: Practical guidelines for efficient cnn architecture design. *european conference on computer vision*, 2018. 4 + +[16] Seyed Iman Mirzadeh, Mehrdad Farajtabar, Ang Li, Nir Levine, Akihiro Matsukawa, and Hassan Ghasemzadeh. Improved knowledge distillation via teacher assistant. national conference on artificial intelligence, 2019. 1, 2 +[17] Wonpyo Park, Dongju Kim, Yan Lu, and Minsu Cho. Relational knowledge distillation. computer vision and pattern recognition, 2019. 1, 2, 7 +[18] Baoyun Peng, Xiao Jin, Dongsheng Li, Shunfeng Zhou, Yichao Wu, Jiaheng Liu, Zhaoning Zhang, and Yu Liu. Correlation congruence for knowledge distillation. international conference on computer vision, 2019. 1, 2 +[19] Adriana Romero, Nicolas Ballas, Samira Ebrahimi Kahou, Antoine Chassang, Carlo Gatta, and Yoshua Bengio. Fitnets: Hints for thin deep nets. arXiv: Learning, 2014. 1, 2, 7 +[20] Mark Sandler, Andrew Howard, Menglong Zhu, Andrey Zhoginov, and Liang-Chieh Chen. *Mobilenetv2: Inverted residuals and linear bottlenecks*. computer vision and pattern recognition, 2018. 4 +[21] Ramprasaath R. Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra. Grad-cam: Visual explanations from deep networks via gradient-based localization. International Journal of Computer Vision, 2016. 3 +[22] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. computer vision and pattern recognition, 2014. 4 +[23] Yonglong Tian, Dilip Krishnan, and Phillip Isola. Contrastive representation distillation. arXiv preprint arXiv:1910.10699, 2019. 1, 2, 7, 8 +[24] Frederick Tung and Greg Mori. Similarity-preserving knowledge distillation. arXiv: Computer Vision and Pattern Recognition, 2019. 1, 2 +[25] Chaofei Wang, Jiayu Xiao, Yizeng Han, Qisen Yang, Shiji Song, and Gao Huang. Towards learning spatially discriminative feature representations. international conference on computer vision, 2021. 3 +[26] Haofan Wang, Zifan Wang, Mengnan Du, Fan Yang, Zijian Zhang, Sirui Ding, Piotr Mardziel, and Xia Hu. Score-cam: Score-weighted visual explanations for convolutional neural networks. computer vision and pattern recognition, 2020. 3 +[27] Seunghan Yang, YoonHyung Kim, Youngeun Kim, and Changick Kim. Combinational class activation maps for weakly supervised object localization. workshop on applications of computer vision, 2019. 3 +[28] Junho Yim, Donggyu Joo, Ji-Hoon Bae, and Junmo Kim. A gift from knowledge distillation: Fast optimization, network minimization and transfer learning. computer vision and pattern recognition, 2017. 1, 2 +[29] Sergey Zagoruyko and Nikos Komodakis. Paying more attention to attention: improving the performance of convolutional neural networks via attention transfer. Learning, 2016, 1, 2, 7, 8 +[30] Sergey Zagoruyko and Nikos Komodakis. Wide residual networks. *british machine vision conference*, 2016. 4 +[31] Xiaopeng Zhang, Jiashi Feng, Hongkai Xiong, and Qi Tian. Zigzag learning for weakly supervised object detection. computer vision and pattern recognition, 2018. 3 + +[32] Xiangyu Zhang, Xinyu Zhou, Mengxiao Lin, and Jian Sun. Shufflenet: An extremely efficient convolutional neural network for mobile devices. computer vision and pattern recognition, 2017. 4 +[33] Borui Zhao, Quan Cui, Renjie Song, Yiyu Qiu, and Jiajun Liang. Decoupled knowledge distillation. arXiv preprint arXiv:2203.08679, 2022. 1, 2, 4, 6, 7, 8 +[34] Bolei Zhou, Aditya Khosla, Agata Lapedriza, Aude Oliva, and Antonio Torralba. Learning deep features for discriminative localization. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2921-2929, 2016. 2, 3, 5 \ No newline at end of file diff --git a/2023/Class Attention Transfer Based Knowledge Distillation/images.zip b/2023/Class Attention Transfer Based Knowledge Distillation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..18ea85c95ec0a3a36dc6e5534be76d4be9c1d4ee --- /dev/null +++ b/2023/Class Attention Transfer Based Knowledge Distillation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:92b704c062ac6417ba6247a34f899b6d957774ad84f71902c070b225e6199247 +size 669597 diff --git a/2023/Class Attention Transfer Based Knowledge Distillation/layout.json b/2023/Class Attention Transfer Based Knowledge Distillation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..e2948bdde5146d22958f76361a44b09690d21183 --- /dev/null +++ b/2023/Class Attention Transfer Based Knowledge Distillation/layout.json @@ -0,0 +1,8025 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 127, + 103, + 465, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 127, + 103, + 465, + 121 + ], + "spans": [ + { + "bbox": [ + 127, + 103, + 465, + 121 + ], + "type": "text", + "content": "Class Attention Transfer Based Knowledge Distillation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 164, + 142, + 428, + 172 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 164, + 142, + 428, + 172 + ], + "spans": [ + { + "bbox": [ + 164, + 142, + 428, + 172 + ], + "type": "text", + "content": "Ziyao Guo" + }, + { + "bbox": [ + 164, + 142, + 428, + 172 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 164, + 142, + 428, + 172 + ], + "type": "text", + "content": ", Haonan Yan" + }, + { + "bbox": [ + 164, + 142, + 428, + 172 + ], + "type": "inline_equation", + "content": "^{1,2,*}" + }, + { + "bbox": [ + 164, + 142, + 428, + 172 + ], + "type": "text", + "content": ", Hui Li" + }, + { + "bbox": [ + 164, + 142, + 428, + 172 + ], + "type": "inline_equation", + "content": "^{1,*}" + }, + { + "bbox": [ + 164, + 142, + 428, + 172 + ], + "type": "text", + "content": ", Xiaodong Lin" + }, + { + "bbox": [ + 164, + 142, + 428, + 172 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 164, + 142, + 428, + 172 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 164, + 142, + 428, + 172 + ], + "type": "text", + "content": "Xidian University, " + }, + { + "bbox": [ + 164, + 142, + 428, + 172 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 164, + 142, + 428, + 172 + ], + "type": "text", + "content": "University of Guelph" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 229, + 174, + 361, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 229, + 174, + 361, + 185 + ], + "spans": [ + { + "bbox": [ + 229, + 174, + 361, + 185 + ], + "type": "text", + "content": "gzyaftermath@outlook.com" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "spans": [ + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 239, + 290, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 239, + 290, + 491 + ], + "spans": [ + { + "bbox": [ + 46, + 239, + 290, + 491 + ], + "type": "text", + "content": "Previous knowledge distillation methods have shown their impressive performance on model compression tasks, however, it is hard to explain how the knowledge they transferred helps to improve the performance of the student network. In this work, we focus on proposing a knowledge distillation method that has both high interpretability and competitive performance. We first revisit the structure of mainstream CNN models and reveal that possessing the capacity of identifying class discriminative regions of input is critical for CNN to perform classification. Furthermore, we demonstrate that this capacity can be obtained and enhanced by transferring class activation maps. Based on our findings, we propose class attention transfer based knowledge distillation (CAT-KD). Different from previous KD methods, we explore and present several properties of the knowledge transferred by our method, which not only improve the interpretability of CAT-KD but also contribute to a better understanding of CNN. While having high interpretability, CAT-KD achieves state-of-the-art performance on multiple benchmarks. Code is available at: https://github.com/GzyAftermath/CAT-KD." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 515, + 128, + 528 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 515, + 128, + 528 + ], + "spans": [ + { + "bbox": [ + 47, + 515, + 128, + 528 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 536, + 287, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 536, + 287, + 693 + ], + "spans": [ + { + "bbox": [ + 46, + 536, + 287, + 693 + ], + "type": "text", + "content": "Knowledge distillation (KD) transfers knowledge distilled from the bigger teacher network to the smaller student network, aiming to improve the performance of the student network. Depending on the type of the transferred knowledge, previous KD methods can be divided into three categories: based on transferring logits [3, 6, 11, 16, 33], features [2, 10, 17-19, 23, 24, 28], and attention [29]. Although KD methods that are based on transferring logits and features have shown their promising performance [2, 33], it is hard to explain how the knowledge they transferred helps to improve the performance of the student network, due to the uninterpretability of logits and features. Relatively, the principle of attention-based KD methods is more intuitive:" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 312, + 212, + 541, + 456 + ], + "blocks": [ + { + "bbox": [ + 312, + 212, + 541, + 456 + ], + "lines": [ + { + "bbox": [ + 312, + 212, + 541, + 456 + ], + "spans": [ + { + "bbox": [ + 312, + 212, + 541, + 456 + ], + "type": "image", + "image_path": "31922e3db0993b7355024d0fc3ec40f96a6e3808c2348f59e5eb0e742ad2e2d5.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 464, + 547, + 509 + ], + "lines": [ + { + "bbox": [ + 305, + 464, + 547, + 509 + ], + "spans": [ + { + "bbox": [ + 305, + 464, + 547, + 509 + ], + "type": "text", + "content": "Figure 1. Illustration of the converted structure. After converting the FC layer into a convolutional layer with " + }, + { + "bbox": [ + 305, + 464, + 547, + 509 + ], + "type": "inline_equation", + "content": "1 \\times 1" + }, + { + "bbox": [ + 305, + 464, + 547, + 509 + ], + "type": "text", + "content": " kernel and moving the position of the global average pooling layer, CAMs can be obtained during the forward propagation." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 533, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 533, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 533, + 547, + 713 + ], + "type": "text", + "content": "it aims at telling the student network which part of the input should it focus on during the classification, which is realized by forcing the student network to mimic the transferred attention maps during training. However, though previous work AT [29] has validated the effectiveness of transferring attention, it does not present what role attention plays during the classification. This makes it hard to explain why telling the trained model where should it focus could improve its performance on the classification mission. Besides, the performance of the previous attention-based KD method [29] is less competitive compared with the methods that are based on transferring logits and features [2, 33]. In this work, we focus on proposing an attention-based KD method that has higher interpretability and better performance." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 702, + 135, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 702, + 135, + 712 + ], + "spans": [ + { + "bbox": [ + 58, + 702, + 135, + 712 + ], + "type": "text", + "content": "*Corresponding author" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "11868" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 48, + 70, + 146, + 144 + ], + "blocks": [ + { + "bbox": [ + 48, + 70, + 146, + 144 + ], + "lines": [ + { + "bbox": [ + 48, + 70, + 146, + 144 + ], + "spans": [ + { + "bbox": [ + 48, + 70, + 146, + 144 + ], + "type": "image", + "image_path": "99ad99410587f8d2adc94e067252483f53781f197df3f22eaa26bbb23742dc9a.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 153, + 546, + 175 + ], + "lines": [ + { + "bbox": [ + 46, + 153, + 546, + 175 + ], + "spans": [ + { + "bbox": [ + 46, + 153, + 546, + 175 + ], + "type": "text", + "content": "Figure 2. Visualization of CAMs corresponding to categories with Top 4 prediction scores for the given image. The predicted categories and their scores are reported in the picture." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 148, + 70, + 246, + 144 + ], + "blocks": [ + { + "bbox": [ + 148, + 70, + 246, + 144 + ], + "lines": [ + { + "bbox": [ + 148, + 70, + 246, + 144 + ], + "spans": [ + { + "bbox": [ + 148, + 70, + 246, + 144 + ], + "type": "image", + "image_path": "bf4fc73385716d74922ecbed962a8fb5843852c6822c3c2c521099929f37fe90.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 248, + 70, + 346, + 144 + ], + "blocks": [ + { + "bbox": [ + 248, + 70, + 346, + 144 + ], + "lines": [ + { + "bbox": [ + 248, + 70, + 346, + 144 + ], + "spans": [ + { + "bbox": [ + 248, + 70, + 346, + 144 + ], + "type": "image", + "image_path": "45caf93496e785cc702fbe528421ea023f2d02ca15a5f732fe9b28068900f7bc.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 347, + 70, + 446, + 144 + ], + "blocks": [ + { + "bbox": [ + 347, + 70, + 446, + 144 + ], + "lines": [ + { + "bbox": [ + 347, + 70, + 446, + 144 + ], + "spans": [ + { + "bbox": [ + 347, + 70, + 446, + 144 + ], + "type": "image", + "image_path": "0ddae377e3ec9301f1704fb9d4de7cb24811433c83c1ea10d29270b1856b8ac1.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 447, + 70, + 545, + 144 + ], + "blocks": [ + { + "bbox": [ + 447, + 70, + 545, + 144 + ], + "lines": [ + { + "bbox": [ + 447, + 70, + 545, + 144 + ], + "spans": [ + { + "bbox": [ + 447, + 70, + 545, + 144 + ], + "type": "image", + "image_path": "24231d24833314bc016b0d5355521e61a0f3d7342009bed46403c761b82b4db8.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 197, + 289, + 448 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 197, + 289, + 448 + ], + "spans": [ + { + "bbox": [ + 46, + 197, + 289, + 448 + ], + "type": "text", + "content": "We start our work by exploring what role attention plays during classification. After revisiting the structure of the mainstream models, we find that with a little conversion (illustrated in Figure 1), class activation map (CAM) [34], a kind of class attention map which indicates the discriminative regions of input for a specific category, can be obtained during the classification. Without changing the parameters and outputs, the classification process of the converted model can be viewed in two steps: (1) the model exploits its capacity to identify class discriminative regions of input and generate CAM for each category contained in the classification mission, (2) the model outputs the prediction score of each category by computing the average activation of the corresponding CAM. Considering that the converted model makes predictions by simply comparing the average activation of CAMs, possessing the capacity to identify class discriminative regions of input is critical for CNN to perform classification. The question is: can we enhance this capacity by offering hints about class discriminative regions of input during training? To answer this question, we propose class attention transfer (CAT)." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 449, + 287, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 449, + 287, + 605 + ], + "spans": [ + { + "bbox": [ + 46, + 449, + 287, + 605 + ], + "type": "text", + "content": "During CAT, the trained model is not required to predict the category of input, it is only forced to mimic the transferred CAMs, which are normalized to ensure they only contain hints about class discriminative regions of input. Through experiments with CAT, we reveal that transferring only CAMs can train a model with high accuracy on the classification task, reflecting the trained model obtains the capacity to identify class discriminative regions of input. Besides, the performance of the trained model is influenced by the accuracy of the model offering the transferred CAMs. This further demonstrates that the capacity of identifying class discriminative regions can be enhanced by transferring more precise CAMs." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 606, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 606, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 606, + 288, + 715 + ], + "type": "text", + "content": "Based on our findings, we propose class attention transfer based knowledge distillation (CAT-KD), aiming to enable the student network to achieve better performance by improving its capacity of identifying class discriminative regions. Different from previous KD methods transferring dark knowledge, we present why transferring CAMs to the trained model can improve its performance on the classification task. Moreover, through experiments with CAT, we reveal several interesting properties of transferring CAMs," + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 197, + 547, + 268 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 197, + 547, + 268 + ], + "spans": [ + { + "bbox": [ + 304, + 197, + 547, + 268 + ], + "type": "text", + "content": "which not only help to improve the performance and interpretability of CAT-KD but also contribute to a better understanding of CNN. While having high interpretability, CAT-KD achieves state-of-the-art performance on multiple benchmarks. Overall, the main contributions of our work are shown below:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 317, + 274, + 545, + 438 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 317, + 274, + 545, + 334 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 274, + 545, + 334 + ], + "spans": [ + { + "bbox": [ + 317, + 274, + 545, + 334 + ], + "type": "text", + "content": "- We propose class attention transfer and use it to demonstrate that the capacity of identifying class discriminative regions of input, which is critical for CNN to perform classification, can be obtained and enhanced by transferring CAMs." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 317, + 345, + 545, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 345, + 545, + 380 + ], + "spans": [ + { + "bbox": [ + 317, + 345, + 545, + 380 + ], + "type": "text", + "content": "- We present several interesting properties of transferring CAMs, which contribute to a better understanding of CNN." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 317, + 392, + 545, + 438 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 392, + 545, + 438 + ], + "spans": [ + { + "bbox": [ + 317, + 392, + 545, + 438 + ], + "type": "text", + "content": "- We apply CAT to knowledge distillation and name it CAT-KD. While having high Interpretability, CAT-KD achieves state-of-the-art performance on multiple benchmarks." + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 452, + 384, + 466 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 452, + 384, + 466 + ], + "spans": [ + { + "bbox": [ + 306, + 452, + 384, + 466 + ], + "type": "text", + "content": "2. Background" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 473, + 545, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 473, + 545, + 556 + ], + "spans": [ + { + "bbox": [ + 304, + 473, + 545, + 556 + ], + "type": "text", + "content": "The concept of knowledge distillation was proposed in [11]. As a transfer learning method, KD aims to improve the performance of the smaller student network by transferring the dark knowledge distilled from the bigger teacher network. Previous KD methods can be divided into three types: distillation from logits [3, 6, 11, 16, 33], features [2, 10, 17-19, 23, 24, 28] and attention [29]." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 557, + 545, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 557, + 545, + 664 + ], + "spans": [ + { + "bbox": [ + 304, + 557, + 545, + 664 + ], + "type": "text", + "content": "To our knowledge, AT [29] is the only KD method based on transferring attention, which defines attention map as the spatial map indicating the area of input that the model focuses on most. In practice, they obtain attention maps by calculating the sum of feature maps while their values are absolutized. However, AT did not present what role attention plays during the classification and why transferring attention maps defined in this way can improve the performance of the student network." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 666, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 666, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 666, + 547, + 713 + ], + "type": "text", + "content": "Previous works related to class attention originate from [34], where the authors propose to utilize high-level feature maps and the parameters of the fully connected layer to generate attention map for a specific category, which is named" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "11869" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "type": "text", + "content": "class activation map (CAM). According to [34], class discriminative regions of input are highlighted in the corresponding CAM. To facilitate understanding, we visualize several CAMs in Figure 2. The following works have successfully applied CAM in various weakly supervised visual tasks [14,27,31]. Besides, there are also many works focus on generalizing CAM [1,21,26] and improving the performance of models by exploiting the information contained in CAM during training [7,25]." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 182, + 289, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 182, + 289, + 266 + ], + "spans": [ + { + "bbox": [ + 46, + 182, + 289, + 266 + ], + "type": "text", + "content": "Previous works have not presented what role attention plays during classification and why transferring attention maps can improve the trained model's performance on the classification mission. In this paper, we focus on figuring out this question and try to propose an attention-based KD method that has both high interpretability and competitive performance." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 280, + 127, + 293 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 280, + 127, + 293 + ], + "spans": [ + { + "bbox": [ + 47, + 280, + 127, + 293 + ], + "type": "text", + "content": "3. Our Method" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 302, + 287, + 387 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 302, + 287, + 387 + ], + "spans": [ + { + "bbox": [ + 46, + 302, + 287, + 387 + ], + "type": "text", + "content": "In this section, we first analyze the structure of the mainstream CNN models and reveal that possessing the capacity of identifying class discriminative regions is critical for CNN to perform classification. Then we further propose class attention transfer to prove that this capacity can be obtained and enhanced by transferring CAMs. Finally, we apply CAT to knowledge distillation." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 397, + 204, + 409 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 397, + 204, + 409 + ], + "spans": [ + { + "bbox": [ + 47, + 397, + 204, + 409 + ], + "type": "text", + "content": "3.1. Revisit the structure of CNN" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 418, + 289, + 550 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 418, + 289, + 550 + ], + "spans": [ + { + "bbox": [ + 46, + 418, + 289, + 550 + ], + "type": "text", + "content": "In image classification tasks, mainstream models usually use CNN to extract features, the resulting high-level feature maps are then globally pooled and fed to a simple fully connected layer to perform classification [8,9,12]. Let " + }, + { + "bbox": [ + 46, + 418, + 289, + 550 + ], + "type": "inline_equation", + "content": "\\mathbf{F} = [F_1,F_2,\\dots,F_C]\\in \\mathbb{R}^{C\\times W\\times H}" + }, + { + "bbox": [ + 46, + 418, + 289, + 550 + ], + "type": "text", + "content": " represents the feature maps generated by the last convolutional layer, where " + }, + { + "bbox": [ + 46, + 418, + 289, + 550 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 46, + 418, + 289, + 550 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 418, + 289, + 550 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 46, + 418, + 289, + 550 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 418, + 289, + 550 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 46, + 418, + 289, + 550 + ], + "type": "text", + "content": " indicate channel dimension, width, and height respectively. And " + }, + { + "bbox": [ + 46, + 418, + 289, + 550 + ], + "type": "inline_equation", + "content": "f_{j}(x,y)" + }, + { + "bbox": [ + 46, + 418, + 289, + 550 + ], + "type": "text", + "content": " denotes the activation of " + }, + { + "bbox": [ + 46, + 418, + 289, + 550 + ], + "type": "inline_equation", + "content": "\\mathbf{F}" + }, + { + "bbox": [ + 46, + 418, + 289, + 550 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 46, + 418, + 289, + 550 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 46, + 418, + 289, + 550 + ], + "type": "text", + "content": " channel at spatial location " + }, + { + "bbox": [ + 46, + 418, + 289, + 550 + ], + "type": "inline_equation", + "content": "(x,y)" + }, + { + "bbox": [ + 46, + 418, + 289, + 550 + ], + "type": "text", + "content": ", while " + }, + { + "bbox": [ + 46, + 418, + 289, + 550 + ], + "type": "inline_equation", + "content": "GAP" + }, + { + "bbox": [ + 46, + 418, + 289, + 550 + ], + "type": "text", + "content": " is the global average pooling layer. Then the process of calculating logits for normal CNN models can be written as:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 84, + 559, + 287, + 617 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 84, + 559, + 287, + 617 + ], + "spans": [ + { + "bbox": [ + 84, + 559, + 287, + 617 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} L _ {i} = \\sum_ {1 \\leq j \\leq C} \\omega_ {j} ^ {i} \\times G A P (F _ {j}) \\\\ = \\frac {1}{W \\times H} \\sum_ {x, y} \\sum_ {1 \\leq j \\leq C} \\omega_ {j} ^ {i} \\times f _ {j} (x, y), \\tag {1} \\\\ \\end{array}", + "image_path": "2c9727f0d72de9b5163fd88bf602eff84ecc8d7769da33d4c575a8dc3fc08dc5.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 629, + 288, + 678 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 629, + 288, + 678 + ], + "spans": [ + { + "bbox": [ + 47, + 629, + 288, + 678 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 629, + 288, + 678 + ], + "type": "inline_equation", + "content": "L_{i}" + }, + { + "bbox": [ + 47, + 629, + 288, + 678 + ], + "type": "text", + "content": " denotes the logit of " + }, + { + "bbox": [ + 47, + 629, + 288, + 678 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 47, + 629, + 288, + 678 + ], + "type": "text", + "content": "-th class, " + }, + { + "bbox": [ + 47, + 629, + 288, + 678 + ], + "type": "inline_equation", + "content": "\\omega_{j}^{i}" + }, + { + "bbox": [ + 47, + 629, + 288, + 678 + ], + "type": "text", + "content": " is the weight of the fully connected layer (FC layer) corresponding to class " + }, + { + "bbox": [ + 47, + 629, + 288, + 678 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 47, + 629, + 288, + 678 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 47, + 629, + 288, + 678 + ], + "type": "inline_equation", + "content": "GAP(F_{j})" + }, + { + "bbox": [ + 47, + 629, + 288, + 678 + ], + "type": "text", + "content": ". According to [34], we can obtain the CAM corresponding to category " + }, + { + "bbox": [ + 47, + 629, + 288, + 678 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 47, + 629, + 288, + 678 + ], + "type": "text", + "content": " by:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 91, + 689, + 287, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 689, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 91, + 689, + 287, + 715 + ], + "type": "interline_equation", + "content": "C A M _ {i} (x, y) = \\sum_ {1 \\leq j \\leq C} \\omega_ {j} ^ {i} \\times f _ {j} (x, y). \\tag {2}", + "image_path": "d4c8c8a820f1ae5ec50b0f537aed584ab1644bf6c09d7f3debf101065a6da7b7.jpg" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 312, + 71, + 542, + 202 + ], + "blocks": [ + { + "bbox": [ + 312, + 71, + 542, + 202 + ], + "lines": [ + { + "bbox": [ + 312, + 71, + 542, + 202 + ], + "spans": [ + { + "bbox": [ + 312, + 71, + 542, + 202 + ], + "type": "image", + "image_path": "1cedc01fa88552c4cbedf18a5871b9b87116bb34dd9d67c06402451fd18bd291.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 209, + 545, + 233 + ], + "lines": [ + { + "bbox": [ + 305, + 209, + 545, + 233 + ], + "spans": [ + { + "bbox": [ + 305, + 209, + 545, + 233 + ], + "type": "text", + "content": "Figure 3. Illustration of CAT. During CAT, the structure of teacher and student are converted to our style (Figure 1)." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 251, + 545, + 274 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 251, + 545, + 274 + ], + "spans": [ + { + "bbox": [ + 305, + 251, + 545, + 274 + ], + "type": "text", + "content": "According to Equation (1) and Equation (2), the calculation of " + }, + { + "bbox": [ + 305, + 251, + 545, + 274 + ], + "type": "inline_equation", + "content": "L_{i}" + }, + { + "bbox": [ + 305, + 251, + 545, + 274 + ], + "type": "text", + "content": " can be written in another form:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 360, + 277, + 545, + 322 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 360, + 277, + 545, + 322 + ], + "spans": [ + { + "bbox": [ + 360, + 277, + 545, + 322 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} L _ {i} = \\frac {1}{W \\times H} \\sum_ {x, y} C A M _ {i} (x, y) \\tag {3} \\\\ = G A P (C A M _ {i}). \\\\ \\end{array}", + "image_path": "8e0f57143703e3a7b2c8917ea46c8439449ac0fbe817950ca57fb09bad8d45c0.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 332, + 545, + 404 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 332, + 545, + 404 + ], + "spans": [ + { + "bbox": [ + 305, + 332, + 545, + 404 + ], + "type": "text", + "content": "As reflected in Equation (3), logits can be obtained by computing the average activation of CAMs. Inspired by it, as illustrated in Figure 1, we convert the FC layer into a " + }, + { + "bbox": [ + 305, + 332, + 545, + 404 + ], + "type": "inline_equation", + "content": "1 \\times 1" + }, + { + "bbox": [ + 305, + 332, + 545, + 404 + ], + "type": "text", + "content": " convolutional layer and move the position of the GAP layer. Then " + }, + { + "bbox": [ + 305, + 332, + 545, + 404 + ], + "type": "inline_equation", + "content": "\\bar{L}_i" + }, + { + "bbox": [ + 305, + 332, + 545, + 404 + ], + "type": "text", + "content": ", the logit of " + }, + { + "bbox": [ + 305, + 332, + 545, + 404 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 305, + 332, + 545, + 404 + ], + "type": "text", + "content": "-th class generated by the converted model, can be obtained by:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 342, + 408, + 545, + 469 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 342, + 408, + 545, + 469 + ], + "spans": [ + { + "bbox": [ + 342, + 408, + 545, + 469 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\bar {L} _ {i} = G A P \\left(\\operatorname {C o n v} _ {i} (\\mathbf {F})\\right) \\\\ = \\frac {1}{W \\times H} \\sum_ {x, y} \\left(\\sum_ {1 \\leq j \\leq C} \\omega_ {j} ^ {i} \\times f _ {j} (x, y)\\right) \\tag {4} \\\\ = G A P \\left(C A M _ {i}\\right), \\\\ \\end{array}", + "image_path": "eb8cbb4eb8134bcd050376433879c7e6bf50cdaa056a5b2da1b67cf4daa3738f.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 474, + 545, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 474, + 545, + 556 + ], + "spans": [ + { + "bbox": [ + 304, + 474, + 545, + 556 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 474, + 545, + 556 + ], + "type": "inline_equation", + "content": "Conv_{i}" + }, + { + "bbox": [ + 304, + 474, + 545, + 556 + ], + "type": "text", + "content": " denotes the converted " + }, + { + "bbox": [ + 304, + 474, + 545, + 556 + ], + "type": "inline_equation", + "content": "1 \\times 1" + }, + { + "bbox": [ + 304, + 474, + 545, + 556 + ], + "type": "text", + "content": " convolution kernel that used to separate features corresponding to " + }, + { + "bbox": [ + 304, + 474, + 545, + 556 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 474, + 545, + 556 + ], + "type": "text", + "content": "-th class from " + }, + { + "bbox": [ + 304, + 474, + 545, + 556 + ], + "type": "inline_equation", + "content": "\\mathbf{F}" + }, + { + "bbox": [ + 304, + 474, + 545, + 556 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 474, + 545, + 556 + ], + "type": "inline_equation", + "content": "\\omega_{j}^{i}" + }, + { + "bbox": [ + 304, + 474, + 545, + 556 + ], + "type": "text", + "content": " is its weight of " + }, + { + "bbox": [ + 304, + 474, + 545, + 556 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 304, + 474, + 545, + 556 + ], + "type": "text", + "content": " channel. As reflected in Eqn(3) and Eqn(4), the conversion does not change the value of its prediction score (i.e., logits). And class activation maps can be obtained during the classification of the converted model." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 558, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 558, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 558, + 546, + 713 + ], + "type": "text", + "content": "As reflected in Eqn(4), the classification process of the converted model can be viewed in two steps: (1) the model exploits its capacity to identify class discriminative regions of input and generate CAMs, (2) the model outputs prediction score of each category by computing the average activation of the corresponding CAM. Considering that the model makes predictions by simply comparing the average activation of CAMs, possessing the capacity to identify class discriminative regions of input is critical for CNN to perform classification. To examine if this capacity can be obtained and enhanced by offering hints indicating class discriminative regions of input to the trained model, we propose class attention transfer." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "11870" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 185, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 185, + 83 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 185, + 83 + ], + "type": "text", + "content": "3.2. Class Attention Transfer" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 90, + 287, + 220 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 90, + 287, + 220 + ], + "spans": [ + { + "bbox": [ + 46, + 90, + 287, + 220 + ], + "type": "text", + "content": "The purpose of CAT is to examine if a model can obtain the capacity to identify class discriminative regions of input by transferring only CAMs. Thus, during CAT, the trained model is not required to perform classification, and any information related to the category of the training set data (e.g., ground-truth labels and logits) is not released to the trained model. In practice, we utilize a pre-trained model with the converted structure to generate the transferred CAMs. The illustration of the process of CAT is shown in Figure 3, while the formal description is shown below." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 221, + 288, + 353 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 221, + 288, + 353 + ], + "spans": [ + { + "bbox": [ + 46, + 221, + 288, + 353 + ], + "type": "text", + "content": "For a given input, let " + }, + { + "bbox": [ + 46, + 221, + 288, + 353 + ], + "type": "inline_equation", + "content": "\\mathbf{A} \\in \\mathbb{R}^{K \\times W \\times H}" + }, + { + "bbox": [ + 46, + 221, + 288, + 353 + ], + "type": "text", + "content": " denotes the CAMs generated by the converted structure, where " + }, + { + "bbox": [ + 46, + 221, + 288, + 353 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 46, + 221, + 288, + 353 + ], + "type": "text", + "content": " is the number of categories contained in the classification task, " + }, + { + "bbox": [ + 46, + 221, + 288, + 353 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 46, + 221, + 288, + 353 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 221, + 288, + 353 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 46, + 221, + 288, + 353 + ], + "type": "text", + "content": " denote the width and height of the generated CAM respectively. " + }, + { + "bbox": [ + 46, + 221, + 288, + 353 + ], + "type": "inline_equation", + "content": "A_{i} \\in \\mathbb{R}^{W \\times H}" + }, + { + "bbox": [ + 46, + 221, + 288, + 353 + ], + "type": "text", + "content": " represents the " + }, + { + "bbox": [ + 46, + 221, + 288, + 353 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 221, + 288, + 353 + ], + "type": "text", + "content": " channel of " + }, + { + "bbox": [ + 46, + 221, + 288, + 353 + ], + "type": "inline_equation", + "content": "\\mathbf{A}" + }, + { + "bbox": [ + 46, + 221, + 288, + 353 + ], + "type": "text", + "content": ", which is the CAM corresponding to category " + }, + { + "bbox": [ + 46, + 221, + 288, + 353 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 221, + 288, + 353 + ], + "type": "text", + "content": ". And " + }, + { + "bbox": [ + 46, + 221, + 288, + 353 + ], + "type": "inline_equation", + "content": "S, T" + }, + { + "bbox": [ + 46, + 221, + 288, + 353 + ], + "type": "text", + "content": " denote student and teacher correspondingly. Besides, we use the average pooling function " + }, + { + "bbox": [ + 46, + 221, + 288, + 353 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 46, + 221, + 288, + 353 + ], + "type": "text", + "content": " to reduce the resolution of the transferred CAMs, to improve the performance of CAT (Section 4.2). Then CAT's loss function can be defined as:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 360, + 287, + 392 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 360, + 287, + 392 + ], + "spans": [ + { + "bbox": [ + 61, + 360, + 287, + 392 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {C A T} = \\sum_ {1 \\leq i \\leq K} \\frac {1}{K} \\| \\frac {\\phi \\left(A _ {i} ^ {T}\\right)}{\\| \\phi \\left(A _ {i} ^ {T}\\right) \\| _ {2}} - \\frac {\\phi \\left(A _ {i} ^ {S}\\right)}{\\| \\phi \\left(A _ {i} ^ {S}\\right) \\| _ {2}} \\| _ {2} ^ {2}. \\tag {5}", + "image_path": "cf6f42cb8e6c72bb4872fc125e2115dd51f5226a0939b62b4cf1bd7a769dfe85.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 401, + 287, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 401, + 287, + 509 + ], + "spans": [ + { + "bbox": [ + 46, + 401, + 287, + 509 + ], + "type": "text", + "content": "As can be seen, we perform " + }, + { + "bbox": [ + 46, + 401, + 287, + 509 + ], + "type": "inline_equation", + "content": "l_{2}" + }, + { + "bbox": [ + 46, + 401, + 287, + 509 + ], + "type": "text", + "content": " normalization on " + }, + { + "bbox": [ + 46, + 401, + 287, + 509 + ], + "type": "inline_equation", + "content": "\\phi(A_i^T)" + }, + { + "bbox": [ + 46, + 401, + 287, + 509 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 401, + 287, + 509 + ], + "type": "inline_equation", + "content": "\\phi(A_i^S)" + }, + { + "bbox": [ + 46, + 401, + 287, + 509 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 46, + 401, + 287, + 509 + ], + "type": "inline_equation", + "content": "l_{1}" + }, + { + "bbox": [ + 46, + 401, + 287, + 509 + ], + "type": "text", + "content": " normalization can be used as well), to ensure that information related to the category of input is not released to the trained model during CAT, considering that the average activation of CAM indicates the prediction score (Equation (3)). Besides, note that here we transfer CAMs of all categories, which is based on our finding that CAMs of all categories both contain beneficial information for CAT (Section 4.2)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 510, + 287, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 510, + 287, + 557 + ], + "spans": [ + { + "bbox": [ + 46, + 510, + 287, + 557 + ], + "type": "text", + "content": "Our core findings through the experiments with CAT are presented as follows, while the corresponding experimental verification and their detailed analysis can be found in Section 4.2." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 58, + 567, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 58, + 567, + 287, + 602 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 567, + 287, + 602 + ], + "spans": [ + { + "bbox": [ + 58, + 567, + 287, + 602 + ], + "type": "text", + "content": "- The capacity to identify class discriminative regions of input can be obtained and enhanced by transferring CAMs." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 58, + 612, + 287, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 612, + 287, + 635 + ], + "spans": [ + { + "bbox": [ + 58, + 612, + 287, + 635 + ], + "type": "text", + "content": "- CAMs of all categories both contain beneficial information for CAT." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 58, + 644, + 246, + 657 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 644, + 246, + 657 + ], + "spans": [ + { + "bbox": [ + 58, + 644, + 246, + 657 + ], + "type": "text", + "content": "- Transferring smaller CAMs performs better." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 58, + 665, + 287, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 665, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 58, + 665, + 287, + 712 + ], + "type": "text", + "content": "- For CAT, the critical information contained in the transferred CAMs is the spatial location of the regions with high activation in them rather than their specific value." + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 72, + 370, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 370, + 83 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 370, + 83 + ], + "type": "text", + "content": "3.3. CAT-KD" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 91, + 545, + 127 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 91, + 545, + 127 + ], + "spans": [ + { + "bbox": [ + 305, + 91, + 545, + 127 + ], + "type": "text", + "content": "After validating the effectiveness of CAT, we apply CAT to knowledge distillation and name it CAT-KD. The loss function of CAT-KD is:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 373, + 141, + 545, + 154 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 373, + 141, + 545, + 154 + ], + "spans": [ + { + "bbox": [ + 373, + 141, + 545, + 154 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {K D} = \\mathcal {L} _ {C E} + \\beta \\mathcal {L} _ {C A T}, \\tag {6}", + "image_path": "a5089feec6c0f653b75daf6641a34d9390ae99ba094a39f75b5482bd954a7cd5.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 167, + 545, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 167, + 545, + 190 + ], + "spans": [ + { + "bbox": [ + 305, + 167, + 545, + 190 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 167, + 545, + 190 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{CE}" + }, + { + "bbox": [ + 305, + 167, + 545, + 190 + ], + "type": "text", + "content": " denotes the standard cross-entropy loss, and " + }, + { + "bbox": [ + 305, + 167, + 545, + 190 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 305, + 167, + 545, + 190 + ], + "type": "text", + "content": " is the factor used to balance the CE loss and CAT loss." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 193, + 545, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 193, + 545, + 277 + ], + "spans": [ + { + "bbox": [ + 304, + 193, + 545, + 277 + ], + "type": "text", + "content": "Different from previous KD methods, we present how the knowledge transferred by CAT-KD helps to improve the performance of the student network: by improving its capacity of identifying class discriminative regions. Besides, through experiments with CAT, we analyze and reveal several properties of the knowledge transferred by our method. This further enhances the interpretability of CAT-KD." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 305, + 293, + 388, + 306 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 293, + 388, + 306 + ], + "spans": [ + { + "bbox": [ + 305, + 293, + 388, + 306 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 305, + 315, + 501, + 327 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 315, + 501, + 327 + ], + "spans": [ + { + "bbox": [ + 305, + 315, + 501, + 327 + ], + "type": "text", + "content": "4.1. Datasets and Implementation Details" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 305, + 335, + 545, + 358 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 335, + 545, + 358 + ], + "spans": [ + { + "bbox": [ + 305, + 335, + 545, + 358 + ], + "type": "text", + "content": "Datasets. In the following section we explore CAT and CAT-KD mainly on two image classification datasets:" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 306, + 361, + 545, + 434 + ], + "type": "list", + "angle": 0, + "index": 21, + "blocks": [ + { + "bbox": [ + 306, + 361, + 545, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 361, + 545, + 396 + ], + "spans": [ + { + "bbox": [ + 306, + 361, + 545, + 396 + ], + "type": "text", + "content": "(1) CIFAR-100 [13] comprise " + }, + { + "bbox": [ + 306, + 361, + 545, + 396 + ], + "type": "inline_equation", + "content": "32 \\times 32" + }, + { + "bbox": [ + 306, + 361, + 545, + 396 + ], + "type": "text", + "content": " pixel images of 100 categories, the training and validate sets contain 50K and 10K images." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 306, + 399, + 545, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 399, + 545, + 434 + ], + "spans": [ + { + "bbox": [ + 306, + 399, + 545, + 434 + ], + "type": "text", + "content": "(2) ImageNet [5] is a large-scale dataset for the classification of 1K categories, containing 1.2 million training and 50K validation images." + } + ] + } + ], + "index": 20 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 305, + 446, + 545, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 446, + 545, + 578 + ], + "spans": [ + { + "bbox": [ + 305, + 446, + 545, + 578 + ], + "type": "text", + "content": "Implementation details. Our implementation for CIFAR-100 and ImageNet strictly follows [2, 33]. Specifically, for CIFAR-100, we train all models for 240 epochs with batch size 64 using SGD. The initial learning rate is 0.05 (0.01 for ShuffleNet [15, 32] and MobileNet [20]), divided by 10 at 150, 180, and 210 epochs. For ImageNet, we train models for 100 epochs with batch size 512. The initial learning rate is 0.2 and divided by 10 for every 30 epochs. We experiment with various representative CNN network: VGG [22], ResNet [9], WideResNet [30], MobileNet [20], and ShuffleNet [15, 32]." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 304, + 580, + 545, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 580, + 545, + 652 + ], + "spans": [ + { + "bbox": [ + 304, + 580, + 545, + 652 + ], + "type": "text", + "content": "For fairness, all the results of previous methods are either reported in previous papers [2,33] (we keep our training setting the same as theirs) or obtained using codes released by the author with our training setting. All results on CIFAR-100 are the average over 5 trials, while that on ImageNet is the average over 3 trials." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "type": "text", + "content": "For all experiments reported in Section 4.2 and Section 4.3, without special specification, we pool the transferred CAMs into " + }, + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "type": "inline_equation", + "content": "2 \\times 2" + }, + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "type": "text", + "content": " during CAT and CAT-KD. More implementation details such as the settings of " + }, + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "type": "text", + "content": " are attached in the appendix due to the page limits." + } + ] + } + ], + "index": 24 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "11871" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 49, + 70, + 286, + 198 + ], + "blocks": [ + { + "bbox": [ + 49, + 70, + 286, + 198 + ], + "lines": [ + { + "bbox": [ + 49, + 70, + 286, + 198 + ], + "spans": [ + { + "bbox": [ + 49, + 70, + 286, + 198 + ], + "type": "image", + "image_path": "6652ce4e1cdeeb81801e302608654780d43980f597ba701148f54f6465f2fe6d.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 201, + 287, + 289 + ], + "lines": [ + { + "bbox": [ + 46, + 201, + 287, + 289 + ], + "spans": [ + { + "bbox": [ + 46, + 201, + 287, + 289 + ], + "type": "text", + "content": "Figure 4. Accuracy of models trained with CAT on CIFAR-100. Left: Only CAMs of certain categories are transferred, which are selected by two strategies: (1) select categories with top n prediction scores, (2) select categories with the lowest n prediction scores. Right: The training set is reduced to contain data of partial categories only. T: test set of CIFAR-100. S: a subset of T which only contains data of classes that are not contained in the training set." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 312, + 162, + 324 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 312, + 162, + 324 + ], + "spans": [ + { + "bbox": [ + 47, + 312, + 162, + 324 + ], + "type": "text", + "content": "4.2. Exploration of CAT" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 331, + 288, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 331, + 288, + 415 + ], + "spans": [ + { + "bbox": [ + 46, + 331, + 288, + 415 + ], + "type": "text", + "content": "In this section, we explore several properties of class attention transfer, which not only help to improve the performance and interpretability of CAT-KD but also contributes to a better understanding of CNN. Note that any information related to the category of the training set (e.g., ground-truth labels and logits) is not utilized in the experiments reported in this section." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 426, + 287, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 426, + 287, + 628 + ], + "spans": [ + { + "bbox": [ + 46, + 426, + 287, + 628 + ], + "type": "text", + "content": "The capacity of identifying class discriminative regions can be obtained and enhanced by transferring CAMs. As revealed in Section 3.1, being able to identify class discriminative regions of input is critical for CNN to perform classification. Thus, the intensity of this capacity can be evaluated by the model's performance on the classification mission. We perform CAT on ShuffleNetV1, where the transferred CAMs are produced by different models with various accuracy. As the results reported in Table 1, transferring only CAMs can train a model with high accuracy on the classification mission, proving the capacity of identifying class discriminative regions can be obtained by transferring CAMs. Besides, the performance of the trained model is influenced by the accuracy of the model producing the transferred CAMs, indicating that this capacity can be enhanced by transferring more precise CAMs." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 641, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 641, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 641, + 288, + 715 + ], + "type": "text", + "content": "CAMs of all categories both contain beneficial information for CAT. For a given input, we can use the method of CAM [34] to generate class activation maps for any categories contained in the classification mission. However, though a few non-target categories may share certain similarities (e.g., shape and patterns) with the target" + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 316, + 70, + 389, + 141 + ], + "blocks": [ + { + "bbox": [ + 316, + 70, + 389, + 141 + ], + "lines": [ + { + "bbox": [ + 316, + 70, + 389, + 141 + ], + "spans": [ + { + "bbox": [ + 316, + 70, + 389, + 141 + ], + "type": "image", + "image_path": "c5bcf1d0e42e4df24c38d9f7174afb12ef69db638438178f2952e3c8aeadd814.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 316, + 142, + 389, + 215 + ], + "blocks": [ + { + "bbox": [ + 316, + 142, + 389, + 215 + ], + "lines": [ + { + "bbox": [ + 316, + 142, + 389, + 215 + ], + "spans": [ + { + "bbox": [ + 316, + 142, + 389, + 215 + ], + "type": "image", + "image_path": "34a5b85f0474a65bdbced53260aa9381525d84fe82eb156cbe1a9bcf5cf51662.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 317, + 217, + 388, + 229 + ], + "lines": [ + { + "bbox": [ + 317, + 217, + 388, + 229 + ], + "spans": [ + { + "bbox": [ + 317, + 217, + 388, + 229 + ], + "type": "text", + "content": "Top1 prairie fowl" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 390, + 70, + 462, + 141 + ], + "blocks": [ + { + "bbox": [ + 390, + 70, + 462, + 141 + ], + "lines": [ + { + "bbox": [ + 390, + 70, + 462, + 141 + ], + "spans": [ + { + "bbox": [ + 390, + 70, + 462, + 141 + ], + "type": "image", + "image_path": "54791102e3cd1e719f415854c16fb9a87a6368a066c7fb428fd10cc4371576af.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 390, + 142, + 462, + 215 + ], + "blocks": [ + { + "bbox": [ + 390, + 142, + 462, + 215 + ], + "lines": [ + { + "bbox": [ + 390, + 142, + 462, + 215 + ], + "spans": [ + { + "bbox": [ + 390, + 142, + 462, + 215 + ], + "type": "image", + "image_path": "6414cdd8ada81c8a2f4772c71b1bceea7afeb90737ee054c05097ad181744896.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 405, + 217, + 447, + 229 + ], + "lines": [ + { + "bbox": [ + 405, + 217, + 447, + 229 + ], + "spans": [ + { + "bbox": [ + 405, + 217, + 447, + 229 + ], + "type": "text", + "content": "Top1 otter" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 465, + 70, + 536, + 141 + ], + "blocks": [ + { + "bbox": [ + 465, + 70, + 536, + 141 + ], + "lines": [ + { + "bbox": [ + 465, + 70, + 536, + 141 + ], + "spans": [ + { + "bbox": [ + 465, + 70, + 536, + 141 + ], + "type": "image", + "image_path": "3a73e068b6b6491437ece202e490c922a5599cd02ec1c2a7d53283373e7dbadb.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 465, + 142, + 536, + 215 + ], + "blocks": [ + { + "bbox": [ + 465, + 142, + 536, + 215 + ], + "lines": [ + { + "bbox": [ + 465, + 142, + 536, + 215 + ], + "spans": [ + { + "bbox": [ + 465, + 142, + 536, + 215 + ], + "type": "image", + "image_path": "8fab1b0c00593f69db1c2c5c834d8779cf84c11d62abf8cdec8c13f1c96c2486.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 476, + 217, + 525, + 229 + ], + "lines": [ + { + "bbox": [ + 476, + 217, + 525, + 229 + ], + "spans": [ + { + "bbox": [ + 476, + 217, + 525, + 229 + ], + "type": "text", + "content": "Top1basset" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 304, + 237, + 547, + 293 + ], + "lines": [ + { + "bbox": [ + 304, + 237, + 547, + 293 + ], + "spans": [ + { + "bbox": [ + 304, + 237, + 547, + 293 + ], + "type": "text", + "content": "Figure 5. We set a pre-trained ResNet50 as CAMs producer to train another ResNet50 from scratch with CAT, CAMs are pooled into " + }, + { + "bbox": [ + 304, + 237, + 547, + 293 + ], + "type": "inline_equation", + "content": "2 \\times 2" + }, + { + "bbox": [ + 304, + 237, + 547, + 293 + ], + "type": "text", + "content": " during the transfer. The first row shows the visualization of the CAMs generated by the producer, while the CAMs visualized in the second row come from the trained model." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 317, + 545, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 317, + 545, + 365 + ], + "spans": [ + { + "bbox": [ + 304, + 317, + 545, + 365 + ], + "type": "text", + "content": "category, most of them are completely irrelevant to the input from a human understanding. However, our experiments show that class activation maps of all categories both contain beneficial information for CAT." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 366, + 546, + 643 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 366, + 546, + 643 + ], + "spans": [ + { + "bbox": [ + 304, + 366, + 546, + 643 + ], + "type": "text", + "content": "We first perform CAT on CIFAR-100 where only CAMs of certain categories are transferred. We designed two strategies to select the categories of the transferred CAMs: (1) select categories with the lowest n prediction scores. (2) select categories with top n prediction scores (the empirical assumption here we make is that the categories with higher prediction scores have more similarities with the target category). As the results reported in Figure 4 (left), while CAMs of classes with higher prediction scores bring more improvement, others are also beneficial for CAT. Besides, we further perform CAT on the reduced CIFAR-100, where CAMs of all classes are transferred but the training set is reduced to contain data of only partial categories. Then the trained model is evaluated on the complete test set and a subset of it which only contains data of classes that are not contained in the training set. As the results reported in Figure 4 (right), interestingly, the trained model achieves high accuracy on the subset, indicating that transferring CAMs enables the trained model to classify the categories that are not contained in the training set. This further proves that non-target CAMs contain beneficial information for CAT even if their categories seem to be irrelevant to the input from a human perspective." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 653, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 653, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 653, + 547, + 715 + ], + "type": "text", + "content": "Transferring smaller CAMs performs better. Intuitively, larger CAM contains more detailed hints about the spatial location of the class discriminative regions, then transferring larger CAMs should perform better. However, insufficient accuracy of the model will result in deviations" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "11872" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 49, + 70, + 102, + 178 + ], + "blocks": [ + { + "bbox": [ + 49, + 70, + 102, + 178 + ], + "lines": [ + { + "bbox": [ + 49, + 70, + 102, + 178 + ], + "spans": [ + { + "bbox": [ + 49, + 70, + 102, + 178 + ], + "type": "image", + "image_path": "bdcc80d5c9958158942304ab20cf6ffaaa090671c389969a3c5c526e2303671e.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 64, + 180, + 85, + 204 + ], + "lines": [ + { + "bbox": [ + 64, + 180, + 85, + 204 + ], + "spans": [ + { + "bbox": [ + 64, + 180, + 85, + 204 + ], + "type": "text", + "content": "Top1 bee" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 102, + 70, + 156, + 178 + ], + "blocks": [ + { + "bbox": [ + 102, + 70, + 156, + 178 + ], + "lines": [ + { + "bbox": [ + 102, + 70, + 156, + 178 + ], + "spans": [ + { + "bbox": [ + 102, + 70, + 156, + 178 + ], + "type": "image", + "image_path": "7cc4cfcf97f183829b521ccdc1a80400ef3df2f862d4f36ae39a034fbc3eec1e.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 109, + 180, + 150, + 203 + ], + "lines": [ + { + "bbox": [ + 109, + 180, + 150, + 203 + ], + "spans": [ + { + "bbox": [ + 109, + 180, + 150, + 203 + ], + "type": "text", + "content": "Top2 sunflower" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 157, + 70, + 211, + 178 + ], + "blocks": [ + { + "bbox": [ + 157, + 70, + 211, + 178 + ], + "lines": [ + { + "bbox": [ + 157, + 70, + 211, + 178 + ], + "spans": [ + { + "bbox": [ + 157, + 70, + 211, + 178 + ], + "type": "image", + "image_path": "d85cffa7978b89543896f630f2cfd43d950bcddb42402095332d06088c3609ff.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 172, + 180, + 194, + 204 + ], + "lines": [ + { + "bbox": [ + 172, + 180, + 194, + 204 + ], + "spans": [ + { + "bbox": [ + 172, + 180, + 194, + 204 + ], + "type": "text", + "content": "Top3 tiger" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 216, + 70, + 269, + 178 + ], + "blocks": [ + { + "bbox": [ + 216, + 70, + 269, + 178 + ], + "lines": [ + { + "bbox": [ + 216, + 70, + 269, + 178 + ], + "spans": [ + { + "bbox": [ + 216, + 70, + 269, + 178 + ], + "type": "image", + "image_path": "4b6e372d959922e6633655735a036a45cc23195774abfb074ed1cfb8920ecb0a.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 223, + 180, + 262, + 204 + ], + "lines": [ + { + "bbox": [ + 223, + 180, + 262, + 204 + ], + "spans": [ + { + "bbox": [ + 223, + 180, + 262, + 204 + ], + "type": "text", + "content": "Top1 kangaroo" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 208, + 545, + 231 + ], + "lines": [ + { + "bbox": [ + 46, + 208, + 545, + 231 + ], + "spans": [ + { + "bbox": [ + 46, + 208, + 545, + 231 + ], + "type": "text", + "content": "Figure 6. The first row shows the visualization of the CAMs corresponding to the top 3 predicted categories, while the following row shows the visualization of them after binarization." + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 270, + 70, + 323, + 178 + ], + "blocks": [ + { + "bbox": [ + 270, + 70, + 323, + 178 + ], + "lines": [ + { + "bbox": [ + 270, + 70, + 323, + 178 + ], + "spans": [ + { + "bbox": [ + 270, + 70, + 323, + 178 + ], + "type": "image", + "image_path": "14bf39c2fff11381fa21e872653c3324b1dbe7ca968954e999dd25ab058802ae.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 284, + 180, + 309, + 203 + ], + "lines": [ + { + "bbox": [ + 284, + 180, + 309, + 203 + ], + "spans": [ + { + "bbox": [ + 284, + 180, + 309, + 203 + ], + "type": "text", + "content": "Top2 rabbit" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 324, + 70, + 378, + 178 + ], + "blocks": [ + { + "bbox": [ + 324, + 70, + 378, + 178 + ], + "lines": [ + { + "bbox": [ + 324, + 70, + 378, + 178 + ], + "spans": [ + { + "bbox": [ + 324, + 70, + 378, + 178 + ], + "type": "image", + "image_path": "c87eb7d030dd71e46b5ecf56e77625fa0774d74ac29d3f641c787027a714fc2e.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 340, + 180, + 359, + 202 + ], + "lines": [ + { + "bbox": [ + 340, + 180, + 359, + 202 + ], + "spans": [ + { + "bbox": [ + 340, + 180, + 359, + 202 + ], + "type": "text", + "content": "Top3 fox" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 383, + 70, + 436, + 178 + ], + "blocks": [ + { + "bbox": [ + 383, + 70, + 436, + 178 + ], + "lines": [ + { + "bbox": [ + 383, + 70, + 436, + 178 + ], + "spans": [ + { + "bbox": [ + 383, + 70, + 436, + 178 + ], + "type": "image", + "image_path": "f44d0fca783d9d4bc5bb7c17fd524620774a16d472a5ab0b13a1163110c458f5.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 399, + 180, + 419, + 202 + ], + "lines": [ + { + "bbox": [ + 399, + 180, + 419, + 202 + ], + "spans": [ + { + "bbox": [ + 399, + 180, + 419, + 202 + ], + "type": "text", + "content": "Top1 bear" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 438, + 70, + 491, + 178 + ], + "blocks": [ + { + "bbox": [ + 438, + 70, + 491, + 178 + ], + "lines": [ + { + "bbox": [ + 438, + 70, + 491, + 178 + ], + "spans": [ + { + "bbox": [ + 438, + 70, + 491, + 178 + ], + "type": "image", + "image_path": "af8f982287d31039a84b247c226a44bf9899b028e2957c2f7f7ad36f357947fc.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 453, + 180, + 476, + 202 + ], + "lines": [ + { + "bbox": [ + 453, + 180, + 476, + 202 + ], + "spans": [ + { + "bbox": [ + 453, + 180, + 476, + 202 + ], + "type": "text", + "content": "Top2 cattle" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 492, + 70, + 545, + 178 + ], + "blocks": [ + { + "bbox": [ + 492, + 70, + 545, + 178 + ], + "lines": [ + { + "bbox": [ + 492, + 70, + 545, + 178 + ], + "spans": [ + { + "bbox": [ + 492, + 70, + 545, + 178 + ], + "type": "image", + "image_path": "84805156fe67564ac1b1337fb9ce162b99295c718e83fe63192682fa2c530ff2.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 508, + 180, + 529, + 202 + ], + "lines": [ + { + "bbox": [ + 508, + 180, + 529, + 202 + ], + "spans": [ + { + "bbox": [ + 508, + 180, + 529, + 202 + ], + "type": "text", + "content": "Top3 otter" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "table", + "bbox": [ + 50, + 245, + 285, + 283 + ], + "blocks": [ + { + "bbox": [ + 50, + 245, + 285, + 283 + ], + "lines": [ + { + "bbox": [ + 50, + 245, + 285, + 283 + ], + "spans": [ + { + "bbox": [ + 50, + 245, + 285, + 283 + ], + "type": "table", + "html": "
CAM ProducerResNet56ResNet110ResNet50
Acc72.3474.3179.34
CAT72.4774.4276.17
", + "image_path": "e207e66180ebe62f8a9e0f65a6f195eda1da5a2a4c589b6441c93c7ef6d5753e.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "table_body" + } + ], + "index": 19 + }, + { + "type": "table", + "bbox": [ + 48, + 338, + 287, + 397 + ], + "blocks": [ + { + "bbox": [ + 46, + 292, + 287, + 326 + ], + "lines": [ + { + "bbox": [ + 46, + 292, + 287, + 326 + ], + "spans": [ + { + "bbox": [ + 46, + 292, + 287, + 326 + ], + "type": "text", + "content": "Table 1. Accuracy " + }, + { + "bbox": [ + 46, + 292, + 287, + 326 + ], + "type": "inline_equation", + "content": "(\\%)" + }, + { + "bbox": [ + 46, + 292, + 287, + 326 + ], + "type": "text", + "content": " of ShuffleNetV1 trained with CAT on CIFAR-100. The transferred CAMs are produced by different models with various accuracy." + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 48, + 338, + 287, + 397 + ], + "lines": [ + { + "bbox": [ + 48, + 338, + 287, + 397 + ], + "spans": [ + { + "bbox": [ + 48, + 338, + 287, + 397 + ], + "type": "table", + "html": "
BaselineModel AccResNet32×4 79.42ResNet8×4 72.5ResNet20 69.06
CAT8×879.6567.9266.21
4×479.8471.6166.43
2×279.7172.4566.84
", + "image_path": "4c3092481199014fb2627a83bbfc1495d2288bc982a782b7bb718b2fe08e693c.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "table_body" + } + ], + "index": 21 + }, + { + "bbox": [ + 46, + 462, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 462, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 462, + 288, + 713 + ], + "type": "text", + "content": "between the highlighted areas in its generated CAM and the actual class discriminative regions of the image (which can be observed in Figure 2). Besides, different models differ in their capacity to identify class discriminative regions, which will lead to subtle differences in the generated CAMs. Therefore transferring CAMs with a larger size does not necessarily improve the performance of CAT. Through experiments, we found that performing average pooling on the transferred CAMs, which will expand the highlighted areas of CAMs and reduce the bias between CAMs generated by different models, could alleviate the above issues. As the results reported in Table 2, though pooling blurs the details, transferring smaller CAMs always performs better. Besides, since the pooling operation expands the highlighted areas of CAM, which will make it encompass larger class discriminative regions, transferring pooled CAMs will force the trained model to pay attention to more discriminative regions, which can be observed in Figure 5. In practice, we pool the transferred CAMs into a smaller size to improve the performance of CAT and CAT-KD (normally " + }, + { + "bbox": [ + 46, + 462, + 288, + 713 + ], + "type": "inline_equation", + "content": "2 \\times 2" + }, + { + "bbox": [ + 46, + 462, + 288, + 713 + ], + "type": "text", + "content": ")." + } + ] + } + ], + "index": 23 + }, + { + "type": "table", + "bbox": [ + 306, + 241, + 545, + 293 + ], + "blocks": [ + { + "bbox": [ + 46, + 405, + 287, + 440 + ], + "lines": [ + { + "bbox": [ + 46, + 405, + 287, + 440 + ], + "spans": [ + { + "bbox": [ + 46, + 405, + 287, + 440 + ], + "type": "text", + "content": "Table 2. Accuracy " + }, + { + "bbox": [ + 46, + 405, + 287, + 440 + ], + "type": "inline_equation", + "content": "(\\%)" + }, + { + "bbox": [ + 46, + 405, + 287, + 440 + ], + "type": "text", + "content": " of various models trained with CAT on the CIFAR-100 test set. During CAT, CAMs are pooled into various sizes. The transferred CAMs are produced by ResNet32 " + }, + { + "bbox": [ + 46, + 405, + 287, + 440 + ], + "type": "inline_equation", + "content": "\\times" + }, + { + "bbox": [ + 46, + 405, + 287, + 440 + ], + "type": "text", + "content": " 4." + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 306, + 241, + 545, + 293 + ], + "lines": [ + { + "bbox": [ + 306, + 241, + 545, + 293 + ], + "spans": [ + { + "bbox": [ + 306, + 241, + 545, + 293 + ], + "type": "table", + "html": "
BaselineModelResNet32×4ResNet50
Acc79.4279.34
CATCAMs79.7180.45
Binarized CAMs79.3579.65
", + "image_path": "9e210c8a03dc6d28067ed1de250d1e03ebeb91144b2f8011bdf1afc051a81255.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "table_body" + } + ], + "index": 24 + }, + { + "type": "table", + "bbox": [ + 337, + 323, + 515, + 399 + ], + "blocks": [ + { + "bbox": [ + 306, + 293, + 545, + 316 + ], + "lines": [ + { + "bbox": [ + 306, + 293, + 545, + 316 + ], + "spans": [ + { + "bbox": [ + 306, + 293, + 545, + 316 + ], + "type": "text", + "content": "Table 3. Results of transferring binarized CAMs. The transferred CAMs are produced by ResNet " + }, + { + "bbox": [ + 306, + 293, + 545, + 316 + ], + "type": "inline_equation", + "content": "32 \\times 4" + }, + { + "bbox": [ + 306, + 293, + 545, + 316 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 337, + 323, + 515, + 399 + ], + "lines": [ + { + "bbox": [ + 337, + 323, + 515, + 399 + ], + "spans": [ + { + "bbox": [ + 337, + 323, + 515, + 399 + ], + "type": "table", + "html": "
TeacherResNet32×4
Acc77.5179.4281.36
ReviewKD [2]76.4277.4577.91
DKD [33]76.5876.4577.29
CAT-KD76.3678.2678.84
Δ-0.22+0.81+0.93
", + "image_path": "527ffb2517f13bcee5b05486b998909ef069ca86094a5b97160c94aa28842d7e.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "table_body" + } + ], + "index": 26 + }, + { + "bbox": [ + 306, + 407, + 545, + 451 + ], + "lines": [ + { + "bbox": [ + 306, + 407, + 545, + 451 + ], + "spans": [ + { + "bbox": [ + 306, + 407, + 545, + 451 + ], + "type": "text", + "content": "Table 4. Comparison with two SOTA methods. The student network is ShuffleNetV1. " + }, + { + "bbox": [ + 306, + 407, + 545, + 451 + ], + "type": "inline_equation", + "content": "\\Delta" + }, + { + "bbox": [ + 306, + 407, + 545, + 451 + ], + "type": "text", + "content": " represents the gap between CAT-KD and the best-performing method among ReviewKD and DKD (marked with underline)." + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 473, + 545, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 473, + 545, + 665 + ], + "spans": [ + { + "bbox": [ + 304, + 473, + 545, + 665 + ], + "type": "text", + "content": "The exact value of the transferred CAMs is not important. To demonstrate that the role CAMs play in CAT is offering hints about the spatial location of the class discriminative regions of input, we binarize the values of the transferred CAMs to 0 and 1, using their average values as the thresholds. The regions of CAM with values above the threshold are considered as being highlighted, indicating the class discriminative regions of input. Thus, we set the values of these regions to 1 to keep them activated after the binarization. Other regions with values below the threshold are considered unhighlighted, and their values are set to 0. As shown in Figure 6, though the specific values of CAMs are lost during the binarization process, the binarized CAMs still contain hints about the spatial location of the class discriminative regions. Note that the threshold can also be specified in other ways (e.g., median)." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "content": "As the results reported in Table 3, although the class discriminative regions obtained by our rudimentary binarization method are not precise, the accuracy of the resulting model dropped by less than one percent, proving that the" + } + ] + } + ], + "index": 29 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "11873" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 62, + 69, + 530, + 242 + ], + "blocks": [ + { + "bbox": [ + 62, + 69, + 530, + 242 + ], + "lines": [ + { + "bbox": [ + 62, + 69, + 530, + 242 + ], + "spans": [ + { + "bbox": [ + 62, + 69, + 530, + 242 + ], + "type": "table", + "html": "
Distillation MechanismTeacher AccResNet32×4 79.42WRN40-2 75.61ResNet32×4 79.42ResNet50 79.34VGG13 74.64
Student AccShuffleNetV1 70.5ShuffleNetV1 70.5ShuffleNetV2 71.82MobileNetV2 64.6MobileNetV2 64.6
LogitsKD [11]74.0774.8374.4567.3567.37
DKD [33]76.4576.777.0770.3569.71
FeaturesCRD [23]75.1176.0575.6569.1169.73
OFD [10]75.9875.8576.8269.0469.48
FitNet [19]73.5973.7373.5463.1664.14
RKD [17]72.2872.2173.2164.4364.52
ReviewKD [2]77.4577.1477.7869.8970.37
AttentionAT [29]71.7373.3272.7358.5859.4
CAT-KD78.2677.3578.4171.3669.13
↑+6.53+4.03+5.68+12.78+9.73
", + "image_path": "fc74a4631d86d7b6af28b4ff7246c1c5da4aac958982fd08000375845d9f0606.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 66, + 277, + 525, + 449 + ], + "blocks": [ + { + "bbox": [ + 46, + 247, + 547, + 269 + ], + "lines": [ + { + "bbox": [ + 46, + 247, + 547, + 269 + ], + "spans": [ + { + "bbox": [ + 46, + 247, + 547, + 269 + ], + "type": "text", + "content": "Table 5. Results on CIFAR-100. Teachers and students have different architectures. " + }, + { + "bbox": [ + 46, + 247, + 547, + 269 + ], + "type": "inline_equation", + "content": "\\uparrow" + }, + { + "bbox": [ + 46, + 247, + 547, + 269 + ], + "type": "text", + "content": " represents the performance improvement of CAT-KD compared with AT." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 66, + 277, + 525, + 449 + ], + "lines": [ + { + "bbox": [ + 66, + 277, + 525, + 449 + ], + "spans": [ + { + "bbox": [ + 66, + 277, + 525, + 449 + ], + "type": "table", + "html": "
Distillation MechanismTeacher AccResNet56 72.34ResNet110 74.31ResNet32×4 79.42WRN-40-2 75.61WRN-40-2 75.61VGG13 74.64
Student AccResNet20 69.06ResNet32 71.14ResNet8×4 72.5WRN-16-2 73.26WRN-40-1 71.98VGG8 70.36
LogitsKD [11]70.6673.0873.3374.9273.5472.98
DKD [33]71.9774.1176.3276.2474.8174.68
FeaturesCRD [23]71.1673.4875.5175.4874.1473.94
OFD [10]70.9873.2374.9575.2474.3373.95
FitNet [19]69.2171.0673.573.5872.2471.02
RKD [17]69.6171.8271.973.3572.2271.48
ReviewKD [2]71.8973.8975.6376.1275.0974.84
AttentionAT [29]70.5572.3173.4474.0872.7771.43
CAT-KD71.6273.6276.9175.674.8274.65
↑+1.07+1.31+3.47+1.52+2.05+3.22
", + "image_path": "c495b02943812e5a7c4b3e5f009b963ce6feb8b8b7de0d7fb2291a48fc555ea4.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 454, + 546, + 478 + ], + "lines": [ + { + "bbox": [ + 46, + 454, + 546, + 478 + ], + "spans": [ + { + "bbox": [ + 46, + 454, + 546, + 478 + ], + "type": "text", + "content": "Table 6. Results on CIFAR-100. Teachers and students have the same architecture. " + }, + { + "bbox": [ + 46, + 454, + 546, + 478 + ], + "type": "inline_equation", + "content": "\\uparrow" + }, + { + "bbox": [ + 46, + 454, + 546, + 478 + ], + "type": "text", + "content": " represents the performance improvement of CAT-KD compared with AT." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 497, + 288, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 497, + 288, + 545 + ], + "spans": [ + { + "bbox": [ + 46, + 497, + 288, + 545 + ], + "type": "text", + "content": "critical information CAMs contained for CAT is the spatial location of class discriminative regions rather than its exact value. This strongly demonstrates that our method is based on transferring attention." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 552, + 177, + 563 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 552, + 177, + 563 + ], + "spans": [ + { + "bbox": [ + 47, + 552, + 177, + 563 + ], + "type": "text", + "content": "4.3. Evaluation of CAT-KD" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 570, + 288, + 619 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 570, + 288, + 619 + ], + "spans": [ + { + "bbox": [ + 46, + 570, + 288, + 619 + ], + "type": "text", + "content": "Consistent with previous works [2, 23, 33], we compare the performance of CAT-KD with several representative KD methods. Moreover, we further evaluate our method from two aspects: transferability and efficiency." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "type": "text", + "content": "Results on CIFAR-100. Table 5 reports the results on CIFAR-100 with the teachers and students having different architectures. Table 6 shows the results where teachers and students have architectures of the same style. Notably, our method outperforms the other attention-based method AT [29] with a large margin (1.07% ~ 12.78%). Moreover, CAT-KD achieves comparable or" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 498, + 547, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 498, + 547, + 641 + ], + "spans": [ + { + "bbox": [ + 304, + 498, + 547, + 641 + ], + "type": "text", + "content": "even better performance compared with feature-based distillation method [2] which requires additional networks and multiple-layer information. Besides, consistent with CAT, the performance of CAT-KD is affected by the accuracy of the teacher: CAMs produced by teachers with lower accuracy contain more incorrect hints about the class discriminative regions of input. To verify this, we further evaluate the impact of the accuracy of the teacher on our method. As the results reported in Table 4, CAT-KD is relatively less effective when the teacher is weak. Thus, as can be observed in Table 6, the performance of CAT-KD is not the best when the teacher is weak." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 653, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 653, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 653, + 547, + 715 + ], + "type": "text", + "content": "Results on ImageNet. Table 7 and Table 8 report the top-1 and top-5 accuracy of image classification on ImageNet. Though the performance of CAT-KD is restricted by the weakness of the teacher network in this setting, our method still outperforms most KD methods." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "11874" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 52, + 70, + 538, + 121 + ], + "blocks": [ + { + "bbox": [ + 52, + 70, + 538, + 121 + ], + "lines": [ + { + "bbox": [ + 52, + 70, + 538, + 121 + ], + "spans": [ + { + "bbox": [ + 52, + 70, + 538, + 121 + ], + "type": "table", + "html": "
FeaturesLogitsAttention
TeacherStudentOFD [10]CRD [23]ReviewKD [2]KD [11]DKD [33]AT [29]CAT-KD
Top-173.3169.7570.8171.1771.6170.6671.770.6971.26
Top-591.4189.0789.9890.1390.5189.8890.4190.0190.45
", + "image_path": "5a28c679d216639377d8e86f469c4cc5d8ba79f7d56f30a4dc43957d095b514e.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 52, + 144, + 538, + 194 + ], + "blocks": [ + { + "bbox": [ + 46, + 122, + 545, + 143 + ], + "lines": [ + { + "bbox": [ + 46, + 122, + 545, + 143 + ], + "spans": [ + { + "bbox": [ + 46, + 122, + 545, + 143 + ], + "type": "text", + "content": "Table 7. Results on ImageNet. In this group, we set ResNet34 as the teacher and ResNet18 as the student. The method with the second-best performance is marked with an underline." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 52, + 144, + 538, + 194 + ], + "lines": [ + { + "bbox": [ + 52, + 144, + 538, + 194 + ], + "spans": [ + { + "bbox": [ + 52, + 144, + 538, + 194 + ], + "type": "table", + "html": "
FeaturesLogitsAttention
TeacherStudentOFD [10]CRD [23]ReviewKD [2]KD [11]DKD [33]AT [29]CAT-KD
Top-176.1668.8771.2571.3772.5668.5872.0569.5672.24
Top-592.8688.7690.3490.4191.0088.9891.0589.3391.13
", + "image_path": "bd1eab9fbe0ee9b395e0db19b8677f7375080856ba9f72b35a4a97621be1b5c9.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 62, + 236, + 272, + 360 + ], + "blocks": [ + { + "bbox": [ + 46, + 196, + 544, + 217 + ], + "lines": [ + { + "bbox": [ + 46, + 196, + 544, + 217 + ], + "spans": [ + { + "bbox": [ + 46, + 196, + 544, + 217 + ], + "type": "text", + "content": "Table 8. Results on ImageNet. In this group, we set ResNet50 as the teacher and MobileNet as the student. The method with the second-best performance is marked with an underline." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 62, + 236, + 272, + 360 + ], + "lines": [ + { + "bbox": [ + 62, + 236, + 272, + 360 + ], + "spans": [ + { + "bbox": [ + 62, + 236, + 272, + 360 + ], + "type": "table", + "html": "
Teacher\nStudentResNet32×4ResNet50
ShuffleNetV1MobileNetV2
DatasetSTLTISTLTI
Baseline69.0536.5464.3930.85
KD [11]66.6132.5667.8132.37
DKD [33]70.7336.7771.0536.48
CRD [23]70.6837.8571.4638.75
ReviewKD [2]71.4638.4666.1632.65
AT [29]71.3637.3665.129.13
CAT-KD74.4340.7373.239.87
", + "image_path": "9822b0876194242460c143e9301034b8eb4421c248265276a0e5f023b6894aa5.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 369, + 286, + 391 + ], + "lines": [ + { + "bbox": [ + 46, + 369, + 286, + 391 + ], + "spans": [ + { + "bbox": [ + 46, + 369, + 286, + 391 + ], + "type": "text", + "content": "Table 9. Comparison on transferring representations learned from CIFAR-100 to STL-10 (STL) and Tiny-ImageNet (TI)." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 415, + 286, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 415, + 286, + 557 + ], + "spans": [ + { + "bbox": [ + 46, + 415, + 286, + 557 + ], + "type": "text", + "content": "Transferability. We perform experiments to compare the transferability of representations to evaluate the generalizability of the knowledge transferred by various methods. We use ShuffleNetV1 and MobileNetV2 as the frozen representations extractors, which are either trained from scratch on CIFAR-100 [13] or distilled from ResNet " + }, + { + "bbox": [ + 46, + 415, + 286, + 557 + ], + "type": "inline_equation", + "content": "32 \\times 4" + }, + { + "bbox": [ + 46, + 415, + 286, + 557 + ], + "type": "text", + "content": " and ResNet50 with various KD methods. Then linear probing tasks are performed on STL-10 [4] and TinyImageNet [5] to quantify their transferability. As the results reported in Table 9, CAT-KD outperforms other methods by a large margin, indicating the outstanding generalizability of the knowledge transferred by our method." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 570, + 286, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 570, + 286, + 712 + ], + "spans": [ + { + "bbox": [ + 46, + 570, + 286, + 712 + ], + "type": "text", + "content": "Efficiency. We first compare the performance of multiple KD methods on CIFAR-100, where the training set is reduced at various ratios, to evaluate their dependence on the amount of training data. As the results reported in Figure 7 (left), CAT-KD is minimally affected by the decrease in the amount of training data, proving the outstanding distillation efficiency of our method. Besides, we further compare the training cost and performance of various KD methods. As reflected in the results reported in Figure 7 (right), CAT-KD has the highest training efficiency. Since CAT-KD does not require extra parameters, its computational cost is almost the same as logits-based" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 306, + 226, + 545, + 360 + ], + "blocks": [ + { + "bbox": [ + 306, + 226, + 545, + 360 + ], + "lines": [ + { + "bbox": [ + 306, + 226, + 545, + 360 + ], + "spans": [ + { + "bbox": [ + 306, + 226, + 545, + 360 + ], + "type": "image", + "image_path": "ac6edc14c512022db031a55c2c7ff30d9595670adf1d2157705fc3050e50188d.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 369, + 545, + 422 + ], + "lines": [ + { + "bbox": [ + 305, + 369, + 545, + 422 + ], + "spans": [ + { + "bbox": [ + 305, + 369, + 545, + 422 + ], + "type": "text", + "content": "Figure 7. We set ResNet32×4 as the teacher and ShuffleNetV1 as the student. Left: accuracy of students trained with various methods on CIFAR-100, where the training set is reduced at various ratios. Right: comparison of accuracy and training time (per epoch) on CIFAR-100." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 444, + 545, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 444, + 545, + 479 + ], + "spans": [ + { + "bbox": [ + 305, + 444, + 545, + 479 + ], + "type": "text", + "content": "methods. Relatively, feature-based methods require much more computational resources because most of them need additional auxiliary networks to distill features." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 490, + 378, + 502 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 490, + 378, + 502 + ], + "spans": [ + { + "bbox": [ + 306, + 490, + 378, + 502 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 510, + 545, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 510, + 545, + 617 + ], + "spans": [ + { + "bbox": [ + 304, + 510, + 545, + 617 + ], + "type": "text", + "content": "In this paper, we propose CAT-KD which has both high interpretability and competitive performance. More importantly, we demonstrate that the capacity of identifying class discriminative regions of input can be obtained and enhanced by transferring CAMs. Furthermore, we present several interesting properties of transferring CAMs, which contribute to a better understanding of CNN. We hope our findings will help future research on the interpretability of CNN and knowledge distillation." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "text", + "content": "Acknowledgement. We thank the reviewers for their constructive feedback. Part of Hui Li's work is supported by the National Natural Science Foundation of China (61932015), Shaanxi Innovation Team project (2018TD007), Higher Education Discipline Innovation 111 project (B16037). Part of Haonan Yan's work is done when he visits the University of Guelph." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "11875" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "type": "text", + "content": "[1] Aditya Chattopadhyay, Anirban Sarkar, Prantik Howlader, and Vineeth N Balasubramanian. Grad-cam++: Generalized gradient-based visual explanations for deep convolutional networks. workshop on applications of computer vision, 2018. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 148, + 288, + 192 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 148, + 288, + 192 + ], + "spans": [ + { + "bbox": [ + 53, + 148, + 288, + 192 + ], + "type": "text", + "content": "[2] Pengguang Chen, Shu Liu, Hengshuang Zhao, and Jiaya Jia. Distilling knowledge via knowledge review. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5008-5017, 2021. 1, 2, 4, 6, 7, 8" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 194, + 288, + 225 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 194, + 288, + 225 + ], + "spans": [ + { + "bbox": [ + 53, + 194, + 288, + 225 + ], + "type": "text", + "content": "[3] Jang Hyun Cho and Bharath Hariharan. On the efficacy of knowledge distillation. international conference on computer vision, 2019. 1, 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 227, + 288, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 227, + 288, + 281 + ], + "spans": [ + { + "bbox": [ + 53, + 227, + 288, + 281 + ], + "type": "text", + "content": "[4] Adam Coates, Andrew Ng, and Honglak Lee. An analysis of single-layer networks in unsupervised feature learning. In Proceedings of the fourteenth international conference on artificial intelligence and statistics, pages 215-223. JMLR Workshop and Conference Proceedings, 2011. 8" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 283, + 287, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 283, + 287, + 326 + ], + "spans": [ + { + "bbox": [ + 53, + 283, + 287, + 326 + ], + "type": "text", + "content": "[5] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. computer vision and pattern recognition, 2009. 4, 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 328, + 287, + 372 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 328, + 287, + 372 + ], + "spans": [ + { + "bbox": [ + 53, + 328, + 287, + 372 + ], + "type": "text", + "content": "[6] Tommaso Furlanello, Zachary C. Lipton, Michael Tschannen, Laurent Itti, and Animashree Anandkumar. Born again neural networks. international conference on machine learning, 2018. 1, 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 373, + 287, + 417 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 373, + 287, + 417 + ], + "spans": [ + { + "bbox": [ + 53, + 373, + 287, + 417 + ], + "type": "text", + "content": "[7] Hao Guo, Kang Zheng, Xiaochuan Fan, Hongkai Yu, and Song Wang. Visual attention consistency under image transforms for multi-label image classification. computer vision and pattern recognition, 2019. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 418, + 287, + 451 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 418, + 287, + 451 + ], + "spans": [ + { + "bbox": [ + 53, + 418, + 287, + 451 + ], + "type": "text", + "content": "[8] Dongyoon Han, Jiwhan Kim, and Junmo Kim. Deep pyramidal residual networks. computer vision and pattern recognition, 2016. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 453, + 287, + 486 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 453, + 287, + 486 + ], + "spans": [ + { + "bbox": [ + 53, + 453, + 287, + 486 + ], + "type": "text", + "content": "[9] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. arXiv: Computer Vision and Pattern Recognition, 2015. 3, 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 487, + 287, + 541 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 487, + 287, + 541 + ], + "spans": [ + { + "bbox": [ + 48, + 487, + 287, + 541 + ], + "type": "text", + "content": "[10] Byeongho Heo, Jeesoo Kim, Sangdoo Yun, Hyojin Park, Nojun Kwak, and Jin Young Choi. A comprehensive overhaul of feature distillation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1921–1930, 2019. 1, 2, 7, 8" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 544, + 287, + 576 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 544, + 287, + 576 + ], + "spans": [ + { + "bbox": [ + 48, + 544, + 287, + 576 + ], + "type": "text", + "content": "[11] Geoffrey E. Hinton, Oriol Vinyals, and Jeffrey Dean. Distilling the knowledge in a neural network. arXiv: Machine Learning, 2015. 1, 2, 7, 8" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 578, + 287, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 578, + 287, + 610 + ], + "spans": [ + { + "bbox": [ + 48, + 578, + 287, + 610 + ], + "type": "text", + "content": "[12] Gao Huang, Zhuang Liu, Laurens van der Maaten, and Kilian Q. Weinberger. Densely connected convolutional networks. computer vision and pattern recognition, 2016. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 612, + 287, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 612, + 287, + 634 + ], + "spans": [ + { + "bbox": [ + 48, + 612, + 287, + 634 + ], + "type": "text", + "content": "[13] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009. 4, 8" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 635, + 287, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 635, + 287, + 668 + ], + "spans": [ + { + "bbox": [ + 48, + 635, + 287, + 668 + ], + "type": "text", + "content": "[14] Kunpeng Li, Ziyan Wu, Kuan-Chuan Peng, Jan Ernst, and Yun Fu. Tell me where to look: Guided attention inference network. computer vision and pattern recognition, 2018. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "type": "text", + "content": "[15] Ningning Ma, Xiangyu Zhang, Hai-Tao Zheng, and Jian Sun. Shufflenet v2: Practical guidelines for efficient cnn architecture design. *european conference on computer vision*, 2018. 4" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 546, + 714 + ], + "type": "list", + "angle": 0, + "index": 33, + "blocks": [ + { + "bbox": [ + 307, + 73, + 546, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 546, + 117 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 546, + 117 + ], + "type": "text", + "content": "[16] Seyed Iman Mirzadeh, Mehrdad Farajtabar, Ang Li, Nir Levine, Akihiro Matsukawa, and Hassan Ghasemzadeh. Improved knowledge distillation via teacher assistant. national conference on artificial intelligence, 2019. 1, 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 118, + 546, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 546, + 150 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 546, + 150 + ], + "type": "text", + "content": "[17] Wonpyo Park, Dongju Kim, Yan Lu, and Minsu Cho. Relational knowledge distillation. computer vision and pattern recognition, 2019. 1, 2, 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 152, + 546, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 152, + 546, + 195 + ], + "spans": [ + { + "bbox": [ + 307, + 152, + 546, + 195 + ], + "type": "text", + "content": "[18] Baoyun Peng, Xiao Jin, Dongsheng Li, Shunfeng Zhou, Yichao Wu, Jiaheng Liu, Zhaoning Zhang, and Yu Liu. Correlation congruence for knowledge distillation. international conference on computer vision, 2019. 1, 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 197, + 546, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 197, + 546, + 229 + ], + "spans": [ + { + "bbox": [ + 307, + 197, + 546, + 229 + ], + "type": "text", + "content": "[19] Adriana Romero, Nicolas Ballas, Samira Ebrahimi Kahou, Antoine Chassang, Carlo Gatta, and Yoshua Bengio. Fitnets: Hints for thin deep nets. arXiv: Learning, 2014. 1, 2, 7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 231, + 546, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 231, + 546, + 274 + ], + "spans": [ + { + "bbox": [ + 307, + 231, + 546, + 274 + ], + "type": "text", + "content": "[20] Mark Sandler, Andrew Howard, Menglong Zhu, Andrey Zhoginov, and Liang-Chieh Chen. *Mobilenetv2: Inverted residuals and linear bottlenecks*. computer vision and pattern recognition, 2018. 4" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 276, + 546, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 276, + 546, + 330 + ], + "spans": [ + { + "bbox": [ + 307, + 276, + 546, + 330 + ], + "type": "text", + "content": "[21] Ramprasaath R. Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra. Grad-cam: Visual explanations from deep networks via gradient-based localization. International Journal of Computer Vision, 2016. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 331, + 546, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 331, + 546, + 365 + ], + "spans": [ + { + "bbox": [ + 307, + 331, + 546, + 365 + ], + "type": "text", + "content": "[22] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. computer vision and pattern recognition, 2014. 4" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 365, + 546, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 365, + 546, + 398 + ], + "spans": [ + { + "bbox": [ + 307, + 365, + 546, + 398 + ], + "type": "text", + "content": "[23] Yonglong Tian, Dilip Krishnan, and Phillip Isola. Contrastive representation distillation. arXiv preprint arXiv:1910.10699, 2019. 1, 2, 7, 8" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 399, + 546, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 399, + 546, + 431 + ], + "spans": [ + { + "bbox": [ + 307, + 399, + 546, + 431 + ], + "type": "text", + "content": "[24] Frederick Tung and Greg Mori. Similarity-preserving knowledge distillation. arXiv: Computer Vision and Pattern Recognition, 2019. 1, 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 433, + 546, + 477 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 433, + 546, + 477 + ], + "spans": [ + { + "bbox": [ + 307, + 433, + 546, + 477 + ], + "type": "text", + "content": "[25] Chaofei Wang, Jiayu Xiao, Yizeng Han, Qisen Yang, Shiji Song, and Gao Huang. Towards learning spatially discriminative feature representations. international conference on computer vision, 2021. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 478, + 546, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 478, + 546, + 521 + ], + "spans": [ + { + "bbox": [ + 307, + 478, + 546, + 521 + ], + "type": "text", + "content": "[26] Haofan Wang, Zifan Wang, Mengnan Du, Fan Yang, Zijian Zhang, Sirui Ding, Piotr Mardziel, and Xia Hu. Score-cam: Score-weighted visual explanations for convolutional neural networks. computer vision and pattern recognition, 2020. 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 522, + 546, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 522, + 546, + 567 + ], + "spans": [ + { + "bbox": [ + 307, + 522, + 546, + 567 + ], + "type": "text", + "content": "[27] Seunghan Yang, YoonHyung Kim, Youngeun Kim, and Changick Kim. Combinational class activation maps for weakly supervised object localization. workshop on applications of computer vision, 2019. 3" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 567, + 546, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 567, + 546, + 612 + ], + "spans": [ + { + "bbox": [ + 307, + 567, + 546, + 612 + ], + "type": "text", + "content": "[28] Junho Yim, Donggyu Joo, Ji-Hoon Bae, and Junmo Kim. A gift from knowledge distillation: Fast optimization, network minimization and transfer learning. computer vision and pattern recognition, 2017. 1, 2" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 307, + 613, + 546, + 655 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 613, + 546, + 655 + ], + "spans": [ + { + "bbox": [ + 307, + 613, + 546, + 655 + ], + "type": "text", + "content": "[29] Sergey Zagoruyko and Nikos Komodakis. Paying more attention to attention: improving the performance of convolutional neural networks via attention transfer. Learning, 2016, 1, 2, 7, 8" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 307, + 657, + 546, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 657, + 546, + 679 + ], + "spans": [ + { + "bbox": [ + 307, + 657, + 546, + 679 + ], + "type": "text", + "content": "[30] Sergey Zagoruyko and Nikos Komodakis. Wide residual networks. *british machine vision conference*, 2016. 4" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 307, + 681, + 546, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 681, + 546, + 714 + ], + "spans": [ + { + "bbox": [ + 307, + 681, + 546, + 714 + ], + "type": "text", + "content": "[31] Xiaopeng Zhang, Jiashi Feng, Hongkai Xiong, and Qi Tian. Zigzag learning for weakly supervised object detection. computer vision and pattern recognition, 2018. 3" + } + ] + } + ], + "index": 32 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "11876" + } + ] + } + ], + "index": 34 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 205 + ], + "type": "list", + "angle": 0, + "index": 3, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "text", + "content": "[32] Xiangyu Zhang, Xinyu Zhou, Mengxiao Lin, and Jian Sun. Shufflenet: An extremely efficient convolutional neural network for mobile devices. computer vision and pattern recognition, 2017. 4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 118, + 287, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 118, + 287, + 150 + ], + "spans": [ + { + "bbox": [ + 48, + 118, + 287, + 150 + ], + "type": "text", + "content": "[33] Borui Zhao, Quan Cui, Renjie Song, Yiyu Qiu, and Jiajun Liang. Decoupled knowledge distillation. arXiv preprint arXiv:2203.08679, 2022. 1, 2, 4, 6, 7, 8" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 152, + 287, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 152, + 287, + 205 + ], + "spans": [ + { + "bbox": [ + 48, + 152, + 287, + 205 + ], + "type": "text", + "content": "[34] Bolei Zhou, Aditya Khosla, Agata Lapedriza, Aude Oliva, and Antonio Torralba. Learning deep features for discriminative localization. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2921-2929, 2016. 2, 3, 5" + } + ] + } + ], + "index": 2 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "11877" + } + ] + } + ], + "index": 4 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Class Balanced Adaptive Pseudo Labeling for Federated Semi-Supervised Learning/27b4466b-7a94-415b-9a3c-c8f4f39a02f2_content_list.json b/2023/Class Balanced Adaptive Pseudo Labeling for Federated Semi-Supervised Learning/27b4466b-7a94-415b-9a3c-c8f4f39a02f2_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..be88bde4e01444222a234582d98cf30bd7296371 --- /dev/null +++ b/2023/Class Balanced Adaptive Pseudo Labeling for Federated Semi-Supervised Learning/27b4466b-7a94-415b-9a3c-c8f4f39a02f2_content_list.json @@ -0,0 +1,1744 @@ +[ + { + "type": "text", + "text": "Class Balanced Adaptive Pseudo Labeling for Federated Semi-Supervised Learning", + "text_level": 1, + "bbox": [ + 114, + 128, + 854, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Ming Li Qingli Li Yan Wang*", + "bbox": [ + 349, + 203, + 630, + 222 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Shanghai Key Laboratory of Multidimensional Information Processing, East China Normal University", + "bbox": [ + 84, + 224, + 892, + 244 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "lm1640362161@gmail.com, qlli@cs.ecnu.edu.cn, ywang@cee.ecnu.edu.cn", + "bbox": [ + 197, + 246, + 781, + 260 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 295, + 313, + 311 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "This paper focuses on federated semi-supervised learning (FSSL), assuming that few clients have fully labeled data (labeled clients) and the training datasets in other clients are fully unlabeled (unlabeled clients). Existing methods attempt to deal with the challenges caused by not independent and identically distributed data (Non-IID) setting. Though methods such as sub-consensus models have been proposed, they usually adopt standard pseudo labeling or consistency regularization on unlabeled clients which can be easily influenced by imbalanced class distribution. Thus, problems in FSSL are still yet to be solved. To seek for a fundamental solution to this problem, we present Class Balanced Adaptive Pseudo Labeling (CBAFed), to study FSSL from the perspective of pseudo labeling. In CBAFed, the first key element is a fixed pseudo labeling strategy to handle the catastrophic forgetting problem, where we keep a fixed set by letting pass information of unlabeled data at the beginning of the unlabeled client training in each communication round. The second key element is that we design class balanced adaptive thresholds via considering the empirical distribution of all training data in local clients, to encourage a balanced training process. To make the model reach a better optimum, we further propose a residual weight connection in local supervised training and global model aggregation. Extensive experiments on five datasets demonstrate the superiority of CBAFed. Code will be available at https://github.com/minglllli/CBAFed.", + "bbox": [ + 75, + 327, + 472, + 750 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 777, + 209, + 792 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Federated learning (FL) aims to train machine learning models on a decentralized manner while preserving data privacy, i.e., separate local models are trained on separate local training datasets independently. In recent years, FL has received much attention for privacy protection reasons", + "bbox": [ + 75, + 803, + 468, + 878 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "[32]. However, most of FL works focused on supervised learning with fully labeled data. But in practice, labeling process of large-scale training data is laborious and expensive. Due to the lack of funds or experts, large labeled training dataset is difficult for many companies and institutions to obtain. These may hinder applicability of FL.", + "bbox": [ + 500, + 296, + 892, + 388 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To handle this problem, federated semi-supervised learning (FSSL) has been explored by many researchers recently [8, 14, 15]. There are broadly three lines of FSSL methods by considering different places and status of labeled data. The first two lines consider that there are only limited labeled data in the central server [8] or each client has partially labeled data [15]. The third line assumes that few clients have fully labeled data and the training datasets in other clients are fully unlabeled [14, 18, 31]. Our paper mainly focuses on the third line of FSSL, and we call local clients with fully labeled data as labeled clients and the other clients as unlabeled clients.", + "bbox": [ + 496, + 392, + 892, + 571 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The main difficulties to train a third line FSSL model lie in three folds: 1) There are no labeled data in unlabeled clients. Thus, the training can be easily biased without label guidance. 2) Due to the divergent class distribution of labeled and unlabeled clients, namely the not independent and identically distributed data (Non-IID) setting, inaccurate supervisory signals may be generated in unlabeled clients via employing the model trained in labeled clients by either pseudo labeling or consistency regularization framework. 3) Due to the catastrophic forgetting problems in CNNs, with the training process of unlabeled clients going on, models may forget the knowledge learned on labeled clients and so decrease the prediction accuracy drastically.", + "bbox": [ + 496, + 577, + 893, + 776 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "RSCFed [14], the state-of-the-art method significantly boosts the FSSL performance by first distilling sub-consensus models, and then aggregating the sub-consensus models to the global model. The sub-consensus models can handle the Non-IID setting to some extent, but the meanteacher based consistency regularization framework in unlabeled clients inevitably causes the accuracy degradation when the classes are imbalanced distributed. RSCFed at", + "bbox": [ + 496, + 779, + 893, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding Author.", + "bbox": [ + 94, + 886, + 225, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "16292", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "tempt to address the catastrophic forgetting problem by adjusting aggregation weights for labeled and unlabeled clients. But, it seems just alleviates the negative impact on the unlabeled clients and the problem is still yet to be tackled. Besides, the sub-consensus models may occur unstable training problems, and increase the communication cost. Other methods like FedConsist [31] and FedIRM [18] achieve promising results by utilizing pseudo labeling methods to produce artificial labels, but they do not consider the Non-IID setting among local clients.", + "bbox": [ + 75, + 90, + 472, + 241 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Pseudo labeling methods are shown to be effective in semi-supervised learning (SSL) [13,24,33,37], which generate pseudo-labels for unlabeled images and propose several strategies to ensure the quality of pseudo-labels. Existing FSSL works only adopt standard pseudo labeling or consistency regularization on FSSL. Since the mentioned difficulties hamper the usage of these methods, other remedy is usually proposed to alleviate the difficulties, which is palliative. Thus, we propose to study FSSL from the perspective of pseudo labeling, seeking for a fundamental solution to this problem.", + "bbox": [ + 75, + 242, + 472, + 409 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Concretely, we present Class Balanced Adaptive Pseudo Labeling, namely CBAFed, by rethinking standard pseudo labeling methods in SSL. To handle the catastrophic forgetting problem, we propose a fixed pseudo labeling strategy, which builds a fixed set by letting pass informative unlabeled data and their pseudo labels at the beginning of the unlabeled client training. Due to the Non-IID and heterogeneous data partition problems in FL, training distribution of unlabeled data can be highly imbalanced, so existing thresholds are not suitable in FSSL. We design class balanced adaptive thresholds via considering the empirical distribution of all training data in local clients at the previous communication round. Analysis proves that our method sets a reasonably high threshold for extremely scarce classes and encourages a balanced training process. To enhance the learning ability and discover unlabeled data from tail classes, we propose to leverage information from so-called \"not informative\" unlabeled data. Besides, we also explore a novel training strategy for labeled clients and the central server, termed as residual weight connection, skip connecting weights from previous epochs (for labeled clients) or previous global models (for the central server). It can help the model reach better optimum, when the training distribution is imbalanced and training amount is small. We conduct extensive experiments on five datasets to show the effectiveness of CBAFed. Overall, our main contributions can be summarized as follows:", + "bbox": [ + 75, + 409, + 472, + 816 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "- We present a CBAFed method to deal with the catastrophic forgetting problems in federated learning. Unlike existing FSSL frameworks that directly adopts pseudo labeling or consistency regularization methods, CBAFed explores a fundamental solution to FSSL via", + "bbox": [ + 94, + 824, + 470, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "designing a novel pseudo labeling method.", + "bbox": [ + 531, + 90, + 815, + 107 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We introduce a residual weight connection method, to improve the robustness of the models in labeled clients and the central server, which skip connects weights from previous epoch or communication round to finally reach better optimum.", + "- Experiments are conducted on five datasets: four natural datasets CIFAR-10/100, SVHN, fashion MNIST and one medical dataset ISIC 2018 Skin. CBAFed outperforms state-of-the-art FSSL methods by a large margin, with $11.3\\%$ improvements on SVHN dataset." + ], + "bbox": [ + 517, + 114, + 893, + 276 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related work", + "text_level": 1, + "bbox": [ + 500, + 300, + 637, + 316 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Federated Learning", + "text_level": 1, + "bbox": [ + 500, + 325, + 689, + 340 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recently, federated learning is becoming more and more popular for privacy protection reasons. Statistical data heterogeneity is one of the most important research problems. FedAvg [19], one of the pioneer works to deal with this problem, performs weighted-averaging on local weights according to the local training size, has been used as the most widely recognized FL baseline. To deal with data heterogeneity, two research directions have been mostly explored: model aggregation [12, 27, 28] and local training [9, 11, 12, 17]. Recently, [23] demonstrates that self-attention-based architectures, e.g., vision transformers(ViT) are more robust to distribution shifts and can converge to better optimum over heterogeneous data.", + "bbox": [ + 496, + 349, + 890, + 546 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Semi-supervised Leaning", + "text_level": 1, + "bbox": [ + 500, + 556, + 730, + 571 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The goal of standard semi-supervised learning (SSL) is to train a good model with limited labeled data and much more unlabeled data. Popular research directions involves consistency regularization [4, 26] and pseudo labeling [3, 10, 16]. Many recent works use the combinations of consistency regularization and pseudo labeling and achieves state-of-the-art performance [1, 2, 25, 30]. In this paper, we focus on using pseudo labeling method to deal with SSL. While these methods achieve much progress, they only consider pre-defined fixed threshold for pseudo labeling. FlexMatch [34] demonstrates that leveraging unlabeled data according to the model's learning status to flexibly adjust thresholds for different classes can boost performance. While these methods perform well in centralized SSL, they all update pseudo labels after every batch's update of the model, which is not suitable in FSSL as shown in later section.", + "bbox": [ + 496, + 579, + 890, + 835 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.3. Federated Semi-supervised Learning", + "text_level": 1, + "bbox": [ + 500, + 845, + 820, + 863 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recently, with the development of federated learning, FSSL becomes popular. There are broadly three categories", + "bbox": [ + 500, + 869, + 890, + 901 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "16293", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/2a2c958fb80ec8b46cc44c4281bb2b9c801e7cbfc89bebfd2e65e7f2a988a7bd.jpg", + "image_caption": [ + "Figure 1. An overview of our CBAFed. In the central server (left side), the global model is aggregated with the returned local models (step ④) and the adaptive thresholds are calculated by the returned training data statistics (step ⑤). Then central server passes the global model, adaptive thresholds and class distribution to all local clients (step ①). After downloading these data, local clients perform local training on the right side (step ②). Labeled clients use labeled data to train the model with residual weight connection. Unlabeled clients obtain the new training dataset by adaptive pseudo labeling and tail class data discovery and use it to train the model. After local training, local clients return trained models and number of data in each class back to central server (step ③)." + ], + "image_footnote": [], + "bbox": [ + 166, + 88, + 477, + 286 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/e91d5849fb9ef1c8e6afb98aeea5656fcef12b04af4e280f30f12c879f9f5576.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 480, + 88, + 805, + 286 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "of FSSL. The first two categories assume that every local client contains partially labeled data or there exists some labeled data in central server [8, 15]. The third category considers that few clients are fully labeled while other clients contain only unlabeled data. This paper focuses on the third category. To deal with this problem, Fed-Consist [31] and FedIRM [18] have been proposed. But, these two methods do not consider data heterogeneity in federated learning. Recently, RSCFed [14] proposes to perform random sub-sampling to reach consensus over clients. However, it uses standard consistency regularization for unlabeled data, which still suffers from the Non-IID setting. We emphasize that all the aforementioned approaches just use traditional pseudo labeling strategy, and none of them discuss how to apply pseudo labeling to FSSL.", + "bbox": [ + 75, + 407, + 472, + 636 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Methodology", + "text_level": 1, + "bbox": [ + 76, + 646, + 212, + 662 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We first introduce the problem setting of federated semi-supervised learning. We consider decentralized training with labeled clients and unlabeled clients. Let $m$ (usually $m = 1$ ) labeled clients be $\\{C_1, C_2, \\dots, C_m\\}$ and $n$ (usually $n > 1$ ) unlabeled clients be $\\{C_{m+1}, C_{m+2}, \\dots, C_{m+n}\\}$ . For labeled client $C_\\ell$ , the labeled local training dataset is denoted as $D_\\ell = \\{(X_i^\\ell, y_i^\\ell)\\}_{i=1}^{N_\\ell}$ , $\\ell = 1, 2, \\dots, m$ , where $N_\\ell$ denotes the data number in $D_\\ell$ . Similarly, for unlabeled client $C_\\mu$ , the unlabeled local training dataset is denoted as $D_\\mu = \\{X_i^\\mu\\}_{i=1}^{N_\\mu}$ , $\\mu = m + 1, m + 2, \\dots, n + m$ .", + "bbox": [ + 76, + 671, + 468, + 840 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The overall pipeline of the proposed CBAFed is shown in Fig. 1, which performs the following steps: (1) Warm up stage: train fully supervised models on only labeled clients using residual weight connection in a normal federated", + "bbox": [ + 75, + 840, + 470, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "learning manner (Sec. 3.1) (2) The central server computes the empirical class distribution and obtains the class balanced adaptive thresholds, then passes them to local clients (Sec. 3.3); (3) All local clients update local models, adaptive threshold and class distribution. Labeled clients train local models on all the data using proposed residual weight connection; Unlabeled clients acquire the fixed training set (Sec. 3.2.2) by the threshold and the tail class datasets (Sec. 3.4), and train local models on the newly obtained training dataset. After local training, all local clients pass the trained model, the number of data in each class and the amount of training data to the central server; (4) The central server aggregates a new model with residual weight connection, computes the class distribution, and obtains the class balanced adaptive threshold. Then, the central server passes them to local clients; (5) Repeat step (3)-(4) until the specified number of communication round is reached. The overall training procedure of our CBAFed is provided in Algorithm 1 in the supplementary material.", + "bbox": [ + 496, + 407, + 893, + 696 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Residual Weight Connection", + "text_level": 1, + "bbox": [ + 500, + 720, + 756, + 736 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We propose a novel training method in local supervised training and global model aggregation called residual weight connection. Our idea of res-weight is simple and similar to ResNet. In ResNet, there is a skip connection between every layer. For our res-weight connection, there is a skip connection of model's parameters between training epochs (or communication rounds). Let $\\theta^{E}$ be the parameters of model after the training epoch (communication round) $E(E > 0)$ , then the final parameters of model after this training epoch by using residual weight connection will", + "bbox": [ + 496, + 750, + 893, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "16294", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "be", + "bbox": [ + 76, + 92, + 98, + 104 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\theta^ {E} = \\left\\{ \\begin{array}{l l} \\theta^ {E} & E \\% s \\neq 0 \\\\ \\alpha_ {1} \\theta^ {E - s} + (1 - \\alpha_ {1}) \\theta^ {E} & E \\% s = 0 \\end{array} \\right. \\tag{1}\n$$\n", + "text_format": "latex", + "bbox": [ + 125, + 102, + 468, + 142 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\alpha_{1}$ is the parameter to scale the size of $\\theta^{E - s}$ and $\\theta^{E}$ and $s$ is the number of skip epoch. In [22], authors show that averaging model weights over training steps tends to produce a more accurate model than using the final weights directly. In our experiments, we will show that this training method can effectively make the model more robust and finally reach better optimum, enhancing the predicted accuracy of pseudo labeling. Note that although the formula of our residual weight connection is similar to mean teachers [26], the update strategy and usage scenario is very different from it. Please refer to supplementary material for more detailed comparison and discussions.", + "bbox": [ + 75, + 150, + 472, + 330 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Pseudo Labeling Methods", + "text_level": 1, + "bbox": [ + 76, + 342, + 313, + 357 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2.1 Review of Classic Batch-based Pseudo Labeling", + "text_level": 1, + "bbox": [ + 76, + 364, + 464, + 381 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In semi-supervised learning, for unlabeled data, traditional approaches [2, 25, 34], such as Fixmatch [25] and Flexmatch [34], use original data or their weak augmented version in one batch to generate pseudo labels. These labels are adopted to supervise model's training. For detailed review, please refer to supplementary materials.", + "bbox": [ + 75, + 388, + 468, + 479 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2.2 Fixed Pseudo Labeling in FSSL", + "text_level": 1, + "bbox": [ + 76, + 500, + 349, + 513 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Warm up stage. To conduct pseudo labeling for unlabeled data, in the warm up stage, we train our model only with labeled data. In the first $P$ communication round, training only conducted in labeled clients. In local training on labeled client $C_{\\ell}$ , the training loss is defined as:", + "bbox": [ + 75, + 523, + 468, + 599 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\ell} = \\frac {1}{| D _ {\\ell} |} \\sum_ {(X ^ {\\ell}, y ^ {\\ell}) \\in D _ {\\ell}} H \\left(y ^ {\\ell}, p _ {m} \\left(y \\mid \\theta_ {t} ^ {\\ell} \\left(X ^ {\\ell}\\right)\\right)\\right), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 122, + 609, + 468, + 648 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\theta_t^\\ell$ means local model in labeled client $C_\\ell$ at communication round $t$ . The local training epoch is set to $J(J > 1)$ with residual weight convection. Then we use FedAVG [19] with residual weight connection for model aggregation:", + "bbox": [ + 76, + 661, + 468, + 719 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\theta_ {t + 1} ^ {G} = \\left\\{ \\begin{array}{l l} \\sum_ {\\ell = 1} ^ {m} \\frac {| D _ {\\ell} |}{\\sum_ {i = 1} ^ {m} | D _ {i} |} \\theta_ {t} ^ {\\ell} & t \\% s \\neq 0 \\\\ \\alpha_ {1} \\theta_ {t + 1 - s} ^ {G} + (1 - \\alpha_ {1}) \\sum_ {\\ell = 1} ^ {m} \\frac {| D _ {\\ell} |}{\\sum_ {i = 1} ^ {m} | D _ {i} |} \\theta_ {t} ^ {\\ell} & t \\% s = 0 \\end{array} \\right. \\tag{3}\n$$\n", + "text_format": "latex", + "bbox": [ + 81, + 728, + 468, + 777 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\theta_{t + 1}^{G}$ is the global model of communication round $t+$ 1. Note that, usually, the number of labeled clients is only 1, so we can finish warm up stage in only one communication round(training can be completed on one labeled client).", + "bbox": [ + 76, + 779, + 468, + 839 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Different from traditional semi-supervised learning, in federated semi-supervised learning, Non-IID data partitions in clients can easily lead to catastrophic forgetting [7, 21, 23]. Thus, local models trained on unlabeled clients", + "bbox": [ + 76, + 839, + 468, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "usually forget the former learned knowledge by batch-based pseudo labeling method, which will lead to abrupt degradation and sabotage the training of unlabeled clients. To handle heterogeneous data partitions and catastrophic forgetting, we propose a new pseudo labeling method called fixed pseudo labeling. For the unlabeled training dataset $D_{\\mu}$ in an unlabeled client $C_{\\mu}$ , after initializing the local model $\\theta_t^\\mu$ with the global model $\\theta_t^G$ , we first compute the pseudo label for each data $X_i^\\mu$ in $D_{\\mu}$ via:", + "bbox": [ + 496, + 90, + 890, + 227 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {y} _ {i} ^ {\\mu} = \\arg \\max p _ {m} (y | \\theta_ {t} ^ {\\mu} \\left(X _ {i} ^ {\\mu}\\right), i = 1, 2, \\dots , N _ {\\mu}. \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 524, + 233, + 890, + 251 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We obtain a subset of unlabeled data with their pseudo labels as supervisory signals during local client training:", + "bbox": [ + 496, + 255, + 890, + 285 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\widetilde {\\mathcal {D}} _ {\\mu} = \\left\\{\\left(X _ {i} ^ {\\mu}, \\hat {y} _ {i} ^ {\\mu}\\right) \\mid X _ {i} ^ {\\mu} \\in D _ {\\mu} \\wedge \\max \\left(p _ {m} \\left(y \\mid \\theta_ {t} ^ {\\mu} \\left(X _ {i} ^ {\\mu}\\right)\\right)\\right) > \\tau \\right\\} _ {i = 1} ^ {N _ {\\mu}}. \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 500, + 289, + 890, + 321 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Then we use this fixed set $\\widetilde{\\mathcal{D}}_{\\mu}$ as the training dataset and define the loss to train our unlabeled client as: $\\mathcal{L}_{\\mu} = \\frac{1}{|\\widetilde{\\mathcal{D}}_{\\mu}|}\\sum_{(X^{\\mu},\\hat{y}^{\\mu})\\in \\widetilde{\\mathcal{D}}_{\\mu}}H(\\hat{y}^{\\mu},p_m(y|\\theta_t^\\mu (X^\\mu))).$", + "bbox": [ + 500, + 320, + 890, + 372 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Class Balanced Adaptive Threshold for Pseudo Labeling", + "text_level": 1, + "bbox": [ + 500, + 377, + 890, + 409 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In semi-supervised learning, the value of threshold $\\tau$ is non-trivial for selecting confident pseudo labels, namely informative unlabeled data. Traditionally, a fixed threshold is pre-defined [25], but setting a fixed threshold usually makes the model fail to consider different learning status and learning difficulties of different classes. In one of the most recent works, Curriculum Pseudo Labeling [34] is proposed to achieve state-of-the-art results on most semi-supervised learning benchmarks, where a flexible threshold for class $c$ at time step $t$ is calculated by: $\\mathcal{T}_t(c) = \\beta_t(c)\\cdot \\tau$ , where $\\beta_{t}(c)$ is the ratio of the number of selected pseudo labels in class $c$ and the maximum number of selected pseudo labels of all classes. $\\beta_{t}(c)$ is used to scale the fixed threshold $\\tau$ . Curriculum Pseudo Labeling significantly boosts the accuracy and convergence performance of various semi-supervised learning algorithms. But in FSSL, there exist heterogeneous data partition problems. Due to the Non-IID partition, the labeled data are not balanced, so purely using the number of selected unlabeled data to design threshold is improper. Moreover, the data of some classes can be extremely rare in certain clients, which will lead to a very low $\\beta_{t}(c)$ , and thus, a very low threshold $(\\mathcal{T}_t(c)\\approx 0)$ . In other words, directly applying Curriculum Pseudo Labeling will introduce many noisy labels into training.", + "bbox": [ + 496, + 416, + 892, + 777 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To this end, we propose a novel Class Balanced Adaptive threshold for Pseudo Labeling (CBAPL). For an unlabeled client $C_{\\mu}$ in the $t$ th communication round, let the number of selected pseudo labels in class $c$ be:", + "bbox": [ + 496, + 777, + 890, + 838 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\sigma_ {t} ^ {\\mu} (c) = \\sum_ {i = 1} ^ {N _ {\\mu}} \\mathbf {1} \\left(\\max \\left(p _ {m} \\left(y \\mid \\theta_ {t} ^ {\\mu} \\left(X _ {i} ^ {\\mu}\\right)\\right)\\right) > \\mathcal {T} _ {t} (c)\\right) \\mathbf {1} \\left(\\hat {y} _ {i} ^ {\\mu} = c\\right), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 501, + 844, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "16295", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathbf{1}(\\cdot)$ is the indicator function. The data number of class $c$ in labeled client $C_{\\ell}$ is calculated by: $\\sigma_t^\\ell (c) = \\sum_{i = 1}^{N_\\ell}\\mathbf{1}(y_i^\\ell = c)$ . After local training, $\\sigma_t^\\mu (c),\\sigma_t^\\ell (c)$ and weights of local models are returned to central server. Then the central server calculates the total number of training data of class $c$ at communication round $t$ :", + "bbox": [ + 76, + 90, + 472, + 183 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\sigma_ {t} (c) = \\sum_ {\\ell = 1} ^ {m} \\sigma_ {t} ^ {\\ell} (c) + \\sum_ {\\mu = m + 1} ^ {n + m} \\sigma_ {t} ^ {\\mu} (c). \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 153, + 193, + 468, + 234 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We then normalize the training numbers and acquire the empirical distribution of training data:", + "bbox": [ + 76, + 244, + 468, + 276 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {p} _ {t} (c) = \\frac {\\sigma_ {t} (c)}{\\sum_ {i = 1} ^ {C} \\sigma_ {t} (i)}. \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 199, + 286, + 468, + 324 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The standard deviation of the empirical distribution is calculated by:", + "bbox": [ + 76, + 333, + 468, + 364 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {s t d} \\left(\\widetilde {p} _ {t}\\right) = \\sqrt {\\frac {1}{C - 1} \\sum_ {c = 1} ^ {C} \\left(\\widetilde {p} _ {t} (c) - \\bar {p} _ {t}\\right) ^ {2}}, \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 143, + 375, + 468, + 422 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\overline{p}_t = \\frac{1}{C}\\sum_{c = 1}^{C}\\widetilde{p}_t(c)$ . Lastly, the threshold of class $c$ is calculated by", + "bbox": [ + 76, + 436, + 468, + 468 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\tau_ {t, c} = \\widetilde {p} _ {t} (c) + \\tau - s t d (\\widetilde {p} _ {t}), \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 179, + 479, + 468, + 500 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\tau$ is the pre-defined threshold base. We then set an upper bound of threshold as:", + "bbox": [ + 76, + 508, + 468, + 539 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {T} _ {t + 1} (c) = \\left\\{ \\begin{array}{l l} \\tau_ {t, c}, & \\tau_ {t, c} < \\tau_ {h} \\\\ \\tau_ {h}, & \\tau_ {t, c} \\geq \\tau_ {h} \\end{array} , \\right. \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 169, + 550, + 468, + 590 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\tau_h$ is the pre-defined upper bound of threshold (usually $\\tau_h = 0.95$ for ensuring a certain amount of training data in unlabeled clients). If our model is trained with more data in one class, the threshold of that class will be higher since $\\widetilde{p}_t(c)$ is higher. $std(\\widetilde{p}_t)$ is important for balancing the empirical distribution of training data. Please refer to the supplementary material for more discussions.", + "bbox": [ + 76, + 602, + 468, + 707 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We argue that for scarce classes, unlike Curriculum Pseudo Labeling [34] whose threshold is extremely low, our CBAPL calculates a threshold with a relatively high lower bound. We have theorem below:", + "bbox": [ + 76, + 708, + 468, + 767 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Theorem 3.1.", + "text_level": 1, + "bbox": [ + 76, + 779, + 174, + 792 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\tau + \\widetilde {p} _ {t} (c) - \\sqrt {\\frac {1}{C}} \\leq \\mathcal {T} _ {t} (c) \\leq \\tau + \\widetilde {p} _ {t} (c), \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 138, + 803, + 468, + 837 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Proof is provided in the supplementary material.", + "bbox": [ + 96, + 847, + 416, + 862 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Since $\\tau >>\\sqrt{\\frac{1}{C}}$ $\\mathcal{T}_t(c)$ will have a high lower bound $(>>0)$ . To balance the trade-off between $\\widetilde{p}_t(t)$ and", + "bbox": [ + 76, + 864, + 468, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "$\\tau - \\sqrt{1 / C}$ , a normalization factor is applied to consider the class number. Thus, we propose a modified $\\widetilde{p}_t(c)$ as below:", + "bbox": [ + 498, + 90, + 890, + 122 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\widetilde {p} _ {t} (c) = \\frac {\\sigma_ {t} (c)}{\\sum_ {i = 1} ^ {C} \\sigma_ {t} (i)} \\times \\frac {C}{1 0}. \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 601, + 131, + 890, + 167 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "At the first $P$ communication rounds, the models are trained on only labeled clients. Thus, when $t \\in \\{1, \\dots, P\\}$ , $\\sigma_t^\\mu(c) = 0$ for all classes. After calculating the thresholds for all classes via Eq. 11, the central server will pass the thresholds to local clients. Unlabeled clients further obtain the fixed pseudo label training dataset by:", + "bbox": [ + 498, + 176, + 890, + 267 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\widetilde {\\mathcal {D}} _ {t + 1, \\mu} = \\left\\{\\left(X _ {i} ^ {\\mu}, \\hat {y} _ {i} ^ {\\mu}\\right) \\mid X _ {i} ^ {\\mu} \\in D _ {\\mu} \\right. \\tag {14} \\\\ \\wedge \\max (p _ {m} (y | \\theta_ {t} ^ {\\mu} (X _ {i} ^ {\\mu}))) > \\mathcal {T} _ {t + 1} (\\hat {y} _ {i} ^ {\\mu}) \\} _ {i = 1} ^ {N _ {\\mu}}. \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 516, + 276, + 890, + 318 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Discovery of Unlabeled Data from Tail Classes", + "text_level": 1, + "bbox": [ + 500, + 325, + 890, + 343 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In FSSL, heterogeneous data partitions will lead to class imbalanced distribution in local clients and there are rare or even no data from some classes in certain clients. For warm up stage in labeled clients, it is similar to long-tailed classification, so the problems in long-tailed classification will also exist in our pseudo labeling process, i.e., models tend to classify tail (rare) classes as head (common) classes [29, 35, 36]. Although adaptive pseudo labeling can lower the threshold of tail classes, much less data will be selected by PL since very few data from tailed classes will be classified as correct classes. To enhance the learning ability and discover unlabeled data from these classes, instead of directly dropping the data whose maximum confidences are low, namely, \"not informative\" unlabeled data, we propose to leverage information from them. For simplicity, we drop subscript $t$ and default all representations occur in $t$ th communication round in Sec. 3.4. Firstly, we define a mask function $\\mathcal{M}(\\cdot): \\mathcal{R}^C \\to \\mathcal{R}^C$ for a probability simplex $p \\in \\mathcal{R}^C$ :", + "bbox": [ + 498, + 349, + 890, + 637 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {M} _ {i} (p) = \\left\\{ \\begin{array}{l l} p _ {i} & i \\neq \\arg \\max p \\\\ 0 & i = \\arg \\max p \\end{array} , \\right. \\tag {15}\n$$\n", + "text_format": "latex", + "bbox": [ + 584, + 646, + 890, + 686 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $p_i$ is the $i$ th element in $p$ . The effect of mask function $\\mathcal{M}(\\cdot)$ is to set the maximum of a probability simplex to 0. Then, we turn to analyze the second largest confidence score via:", + "bbox": [ + 498, + 696, + 890, + 755 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} D _ {\\mu} ^ {t a i l} = \\left\\{\\left(X _ {i} ^ {\\mu}, \\hat {y} _ {i} ^ {\\mu^ {\\prime}}\\right) \\mid X _ {i} ^ {\\mu} \\in D _ {\\mu} \\right. \\\\ \\left. \\wedge \\max \\left(p _ {m} \\left(y \\mid \\theta_ {t} ^ {\\mu} \\left(X _ {i} ^ {\\mu}\\right)\\right)\\right) \\leq \\mathcal {T} _ {t} \\left(\\hat {y} _ {i} ^ {\\mu}\\right) \\wedge \\widetilde {p} _ {t} \\left(\\hat {y} _ {i} ^ {\\mu^ {\\prime}}\\right) < \\frac {\\beta}{C} \\right\\}, \\tag {16} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 519, + 762, + 890, + 828 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\hat{y}_i^{u'} = \\arg \\max \\mathcal{M}(p_m(y|\\theta_t^\\mu(X_i^\\mu)))$ and $\\beta$ is a hyperparameter. In Eq. 15, the parameter $\\beta$ is to find the tail classes (\"not informative\" unlabeled data). For unlabeled data $X_i^\\mu$ , if its largest confidence is smaller than threshold", + "bbox": [ + 498, + 839, + 890, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "16296", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "(not informative or confident, i.e., $\\max(p_m(y|\\theta_t^\\mu(X_i^\\mu))) \\leq \\mathcal{T}_t(\\hat{y}_i^\\mu)$ ) and the class of second largest confidence is tail class (i.e., $\\widetilde{p}_t(\\hat{y}_i^{\\mu \\prime}) < \\frac{\\beta}{C}$ ), we regard it as misclassified data and include it in training with the class of second largest confidence as its label. Finally, the new training dataset for unlabeled client $C_\\mu$ is defined as", + "bbox": [ + 76, + 90, + 472, + 183 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nD _ {\\mu} ^ {t r a i n} = D _ {\\mu} ^ {t a i l} \\cup \\widetilde {D} _ {\\mu}, \\tag {17}\n$$\n", + "text_format": "latex", + "bbox": [ + 196, + 194, + 468, + 215 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "and the final training loss in $C_{\\mu}$ is rewritten as", + "bbox": [ + 76, + 224, + 385, + 242 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mu} = \\frac {1}{| D _ {\\mu} ^ {t r a i n} |} \\sum_ {(X ^ {\\mu}, \\hat {y} ^ {\\mu}) \\in D _ {\\mu} ^ {t r a i n}} H \\left(\\hat {y} ^ {\\mu}, p _ {m} \\left(y \\mid \\theta_ {t} ^ {\\mu} \\left(X ^ {\\mu}\\right)\\right)\\right). \\tag {18}\n$$\n", + "text_format": "latex", + "bbox": [ + 89, + 252, + 468, + 304 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The calculation of $\\sigma_t^u (c)$ is revised as:", + "bbox": [ + 76, + 305, + 331, + 321 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\sigma_ {t} ^ {\\mu} (c) = \\sum_ {(X ^ {\\mu}, \\hat {y} ^ {\\mu}) \\in D _ {\\mu} ^ {t r a i n}} \\mathbf {1} (\\hat {y} ^ {\\mu} = c). \\tag {19}\n$$\n", + "text_format": "latex", + "bbox": [ + 153, + 332, + 468, + 367 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.5. Aggregation of local models", + "text_level": 1, + "bbox": [ + 76, + 377, + 326, + 393 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Since not all data in local clients are included into training, we propose to use only data included in training for model aggregation. The scale factor $w_{t}^{i}$ is computed by", + "bbox": [ + 76, + 400, + 468, + 446 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nw _ {t} ^ {i} = \\left\\{ \\begin{array}{l l} \\frac {\\left| D _ {i} \\right|}{\\left| D _ {t} ^ {t r a i n} \\right|} & \\text {i f} i \\in \\{1, \\dots , m \\} \\\\ \\frac {\\left| D _ {t , i} ^ {t r a i n} \\right|}{\\left| D _ {t} ^ {t r a i n} \\right|} & \\text {i f} i \\in \\{m + 1, \\dots , m + n \\}, \\end{array} \\right. \\tag {20}\n$$\n", + "text_format": "latex", + "bbox": [ + 101, + 455, + 468, + 513 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $|D_t^{train}| = \\sum_{\\ell=1}^m |D_\\ell| + \\sum_{\\mu=m+1}^{m+n} |D_{t,\\mu}^{train}|$ . Then at $(t+1)$ th synchronization round, the global model $\\theta_{t+1}^G$ is computed by", + "bbox": [ + 76, + 523, + 468, + 571 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\theta_ {t + 1} ^ {G} = \\left\\{ \\begin{array}{l l} \\sum_ {i = 1} ^ {m + n} w _ {t} ^ {i} \\theta_ {t} ^ {i} & t \\% s \\neq 0 \\\\ \\alpha_ {2} \\theta_ {t + 1 - s} ^ {G} + (1 - \\alpha_ {2}) \\sum_ {i = 1} ^ {m + n} w _ {t} ^ {i} \\theta_ {t} ^ {i} & t \\% s = 0, \\end{array} \\right. \\tag{21}\n$$\n", + "text_format": "latex", + "bbox": [ + 84, + 580, + 468, + 619 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\alpha_{2}$ is a hyperparameter.", + "bbox": [ + 81, + 627, + 282, + 642 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 76, + 656, + 209, + 672 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Experimental Setup", + "text_level": 1, + "bbox": [ + 76, + 680, + 266, + 696 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Benchmark Datasets To evaluate the effectiveness of our proposed method, we conduct extensive experiments on four image classification datasets, i.e., SVHN, CIFAR-10, CIFAR-100, Fashion MNIST and one medical image classification dataset: ISIC 2018 (Skin Lesion Analysis Towards Melanoma Detection). Dataset splitting and pre-processing are provided in the supplementary material.", + "bbox": [ + 76, + 704, + 468, + 809 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "FSSL Setting Simulation Following [14], training datasets contain 10 clients: one labeled and nine unlabeled. We use a Dirichlet distribution $\\text{Dir}(\\gamma)$ , where $\\gamma = 0.8$ for five datasets [14] to generate Non-IID data partition in clients.", + "bbox": [ + 76, + 810, + 468, + 869 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation Details We utilize the SGD optimizer with a momentum of 0.9, and implement our method with", + "bbox": [ + 76, + 869, + 468, + 900 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/6ba1612a951d4ca351d73178f6854ebf22e02b0228ccf57159f3b72aa4092cb1.jpg", + "image_caption": [ + "Figure 2. Test accuracy curves in local training of SVHN dataset w/ and w/o residual weight connection. ResNet18 (a) and ViT (b) are adopted as the backbones. W/ res-weight* indicates we only show test accuracy on epochs (communication rounds, since local training epoch for labeled client is 1) with skip weight connection. Best viewed electronically." + ], + "image_footnote": [], + "bbox": [ + 501, + 90, + 700, + 239 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/115cb75ad70581d87bb29fd31b0f256a92835927fed7dd12377d5a66dc5f13ba.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 700, + 90, + 885, + 239 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "PyTorch. We adopt ResNet18 [6] from PyTorch for all datasets. For fair comparison, we use the same network architecture and training protocol, including the optimizer, data preprocessing, etc. across all FSSL methods. The local training epoch is set to 11 for labeled clients and 1 for unlabeled clients. More implementation details are provided in supplementary material.", + "bbox": [ + 498, + 349, + 890, + 455 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Comparisons We compare our method against state-of-the-art FSSL methods, including RSCFed [14], FedIRM [18], and Fed-Consist [31]. Since FedIRM [18] and Fed-Consist [31] are not desgined for dealing with Non-IID FSSL, following [14], we enlarge the weight of labeled client to about $50\\%$ and other nine unlabeled clients share the remaining $50\\%$ weight in each FSSL synchronization round, when implementing these methods. We also compare our network against FedAvg [19] trained with all clients as an upper bound and trained with only one labeled clients as the lower bound. Besides, to show the effectiveness of our residual weight connection, we report the result of FedAvg [19] trained with only one labeled clients using residual weight connection. Since Fed-Consist [31] utilizes traditional batch-based pseudo labeling method, we report the result of Fed-Consist [31] using our proposed fixed pseudo labeling without enlarging the weight of labeled client.", + "bbox": [ + 498, + 455, + 892, + 714 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Comparisons with State-of-the-arts", + "text_level": 1, + "bbox": [ + 500, + 724, + 805, + 742 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Comparison Results. Table 1 reports the quantitative results of our CBAFed and other state-of-the-art methods on five benchmarks. We can observe that our proposed CBAFed achieves superior performances over all competitors on the five benchmark datasets. For some easy tasks (e.g., SVHN and Fashion-MNIST), the performance of our CBAFed is approaching upper bound obtained by FedAVG [19] and surpasses all state-of-the-arts by a large margin, showing a strong power of fixed pseudo labeling and class balanced threshold. Note that if replacing the traditional", + "bbox": [ + 498, + 750, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "16297", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/4821af0147cba5088d02e38a90651647115ead6d80cfee580e48bec9ea0fd922.jpg", + "table_caption": [ + "Table 1. Results on SVHN, CIFAR-10/100, Fashion MNIST and ISIC 2018 datasets under heterogeneous data partition with ResNet18. FedAVG $^+$ means FedAvg [19] trained with all one labeled clients using our residual weight connection. Fed-consist $^+$ means Fed-Consist [31] using our proposed fixed pseudo labeling without enlarging the weight of labeled client." + ], + "table_footnote": [], + "table_body": "
Labeling StrategyMethodClient Num.Dataset
labeledunlabeledSVHNCIFAR10CIFAR100Fashion-MNISTISIC 2018
Fully supervisedFedAvg [19](upper-bound)10091.8380.8951.3890.1481.32
FedAvg [19](lower-bound)1067.7154.6620.4974.8765.13
FedAvg+ [19]1076.9858.2124.8478.2666.69
Semi supervisedFedIRM [18]1969.2252.8420.2076.8364.85
Fed-Consist [31]1970.5654.2321.8176.5765.20
Fed-Consist+ [31]1986.5756.3523.2578.3565.50
RSCFed [14]1976.7457.0728.4678.4067.21
CBAFed(ours)1988.0767.0830.1885.4968.29
", + "bbox": [ + 81, + 141, + 893, + 266 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "batch-based pseudo labeling by our fixed pseudo labeling in Fed-Consist [31], the performance is largely increased. It shows that fixed pseudo labeling method is well suited for Non-IID FSSL.", + "bbox": [ + 75, + 291, + 468, + 349 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Trained on one labeled client, FedAVG [19] w/ residual weight connection achieves much better performance compared with that w/o residual weight connection. Fig. 2 (a) shows the test accuracy curves during training. Due to the imbalanced training data distribution in labeled client, the test accuracy is unstable during training for FedAVG (w/o res-weight in the figure). But, if training w/ our residual weight connection, the test accuracy curve is much more stable and the performance is also enhanced (w/ res-weight and w/ res-weight*). To sum up, residual weight connection can improve the robustness of the models when the training amount is small and the training distribution is imbalanced. Since long tailed vision recognition task is similar to local training in labeled client, it is a potential direction to use our residual weight connection in long tailed vision. Besides, we also conduct experiments on partially labeled clients, results are shown in the supplementary material.", + "bbox": [ + 75, + 353, + 468, + 609 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "ViT Backbone. Vision Transformer(ViT) has shown to be more robust to heterogeneous data and distribution shifts, and has been demonstrated to converge faster with better optimum in Federated learning [23]. To further study the effect of our method, we conduct experiments on SVHN dataset using ViT-Tiny as backbone for all competitors. The implementation details are provided in the supplementary material. Table 2 shows the comparison results. Our method can outperform all other methods. Similarly, trained on one labeled client, FedAVG [19] w/ residual weight connection surpasses the one w/o residual weight connection, meaning that our residual weight connection is also effective on ViT, not only CNNs, as shown in Fig. 2 (b).", + "bbox": [ + 75, + 611, + 468, + 808 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Two Labeled Clients. To better understand the effectiveness of our CBAFed and study how to use residual weight connection when the number of labeled clients is more than one, following [14, 18], we conduct experiments on CIFAR-10 by dividing the whole training data into 10 clients, with two labeled clients and eight unlabeled clients. Results are", + "bbox": [ + 75, + 810, + 468, + 900 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/42f2ce88d940c39e81c63f133180f8ddef449260a7c44db3a3e8ea9ac9c06f48.jpg", + "table_caption": [ + "Table 2. Comparison of our method against RSCFed [14], FedConsist [31] and FedAVG [19] in SVHN dataset on ViT [5] as the backbone, with one labeled and nine unlabeled clients." + ], + "table_footnote": [], + "table_body": "
MethodClient Num.Accuracy
labeledunlabeled
FedAVG [19](upper bound)10096.81
FedAVG [19](lower bound)1081.68
FedAVG+ [19]1088.93
FedIRM [18]1979.44
Fed-Consist [31]1985.91
Fed-Consist+ [31]1993.21
RSCFed [14]1989.43
CBAFed(ours)1995.09
", + "bbox": [ + 522, + 342, + 870, + 450 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/6ac072582755d096deaa5a9114ed69a97aff9e1b6d3f7d3e7f441f98958e1bab.jpg", + "table_caption": [ + "Table 3. Comparison of our method against RSCFed [14], FedConsist [31], FedIRM [18] and FedAVG [19] with the number of labeled and unlabeled client set to 2 and 8." + ], + "table_footnote": [], + "table_body": "
MethodClient Num.Accuracy
labeledunlabeled
FedAVG [19](upper bound)10080.89
FedAVG [19](lower bound)2061.85
FedAVG+ [19]2066.55
FedIRM [18]2862.62
Fed-Consist [31]2861.67
Fed-Consist+ [31]2868.04
RSCFed [14]2864.25
CBAFed(ours)2872.01
", + "bbox": [ + 504, + 513, + 890, + 633 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "reported in Table 3. Our method has a better performance compared with all state-of-the-art FSSL methods. Adding our residual weight connection and fixed pseudo labeling lead to much performance gain. In the supplementary material, we also discuss different training strategies used for labeled clients w.r.t. number of local epochs and the usage of residual weight connection.", + "bbox": [ + 496, + 656, + 890, + 761 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Communication Cost. Since the statistics of local training data can be nearly neglected (very low compared with model's parameter), our method does not add any burden in communication cost, which is only $66\\%$ of the state-of-the-art method RSCFed [14].", + "bbox": [ + 496, + 762, + 890, + 837 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Study of Pseudo Labeling Strategies in FSSL", + "text_level": 1, + "bbox": [ + 500, + 847, + 879, + 864 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In this section, we discuss how to perform pseudo labeling in FSSL and why traditional batch-based pseudo la", + "bbox": [ + 496, + 869, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "16298", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/be35e86cc4f6b643dbdc12187d6d6b5889ddc0ba345357c0159fee5347d44c7d.jpg", + "image_caption": [ + "Figure 3. Left: Test accuracy of all 5 strategies after every communication round. Note that the test accuracy of communication round 0 is the test accuracy of model trained on labeled client. Right: Accuracy of pseudo labels in local training epoch of one randomly selected unlabeled client. Best viewed electronically." + ], + "image_footnote": [], + "bbox": [ + 81, + 90, + 269, + 227 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/f9c462a8aac5664fcfbd97999ec2daa599ebc85feffbf2427b5a2650514a3d61.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 272, + 90, + 464, + 228 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "beling fails. We conduct a simple experiment on SVHN dataset [20] using ResNet18 [6]. The experimental setting is the same as illustrated in Sec. 4.1 except that we only use pseudo labeling without other proposed modules. We consider five training strategies: (1) Batched pseudo labeling (1 local training epoch). (2) Batched pseudo labeling (5 local training epochs). (3) Fixed pseudo labeling (1 local training epoch). (4) Fixed pseudo labeling (5 local training epochs), and we do not update the fixed training dataset in all local training epochs after pseudo labeling. (5) Fixed pseudo labeling (5 local training epochs), and we update the fixed training dataset at the beginning of every training epoch. For labeled client, the local training epoch is set to 1 for all 5 strategies. The total communication round is set to 200 for all strategies. The test accuracy after communication round is shown in the left figure of Fig. 3. We can observe the following: i), Batch-based pseudo labeling performs poorly in FSSL. It is known that CNNs often do not work well on out-of-distribution data [23]. Thus, a few update can result in catastrophic forgetting on Non-IID data partitions. The right figure of Fig. 3 shows the predicted accuracy of communication round one in five local training epochs of strategy (2) and (5). ii), Our fixed pseudo labeling method is capable of dealing with catastrophic forgetting. Strategy (3) performs the best and strategy (5) performs much worse than strategy (3) and (4), indicating that one local training epoch is optimum. iii) Comparing strategy (3) with (4), the former has better performance but the latter has faster convergence speed. It means more local training epochs will speed up convergence but may degrade performance, similar phenomenon is also found in [23].", + "bbox": [ + 75, + 324, + 472, + 792 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4. Ablation Study", + "text_level": 1, + "bbox": [ + 76, + 801, + 232, + 816 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We conduct ablation study to demonstrate the effectiveness of the main components of our method: residual weight connection, class balanced pseudo labeling, and tail class data discovery. Table 4 shows the results on three datasets: CIFAR-10/100 and Fashion-MNIST. It can be", + "bbox": [ + 75, + 825, + 468, + 900 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/c94f903c635d2cb8a4e4047fbbe73981580e998910daa0b4431af2aca24c072c.jpg", + "table_caption": [ + "Table 4. Ablation Study of CBAFed in CIFAR-10/100 and Fashion MNIST Datasets. Fixed PL: fixed pseudo labeling, CBA: class balanced adaptive pseudo labeling, DD: tail class data discovery." + ], + "table_footnote": [], + "table_body": "
DatasetFixed PLCBADDRes-WeightAccuracy
CIFAR-10✓59.16
✓✓64.29
✓✓✓65.15
✓✓✓✓67.08
CIFAR-100✓27.64
✓✓29.41
✓✓✓29.86
✓✓✓✓30.18
Fashion-MNIST✓79.99
✓✓80.87
✓✓✓84.37
✓✓✓✓85.49
", + "bbox": [ + 521, + 141, + 870, + 256 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/5ea2888668a874d4a89787ac9194eb9374b2690d579f8ba6a281056f3067b40d.jpg", + "image_caption": [ + "Figure 4. Performance changes on Fashion-MNIST by varying (a) threshold base $\\tau$ and (b) upper bound threshold $\\tau_{h}$ , and (c) parameter for selecting tail class data $\\beta$ ." + ], + "image_footnote": [], + "bbox": [ + 521, + 266, + 870, + 362 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "seen that all three modules are important for boosting performance. Then, we aim at discussing hyper-parameters of threshold base $\\tau$ , upper bound threshold $\\tau_h$ , and parameters for selecting tail class data $\\beta$ . Experiments are conducted on Fashion-MNIST. As shown in Fig. 4, performances are not sensitive within certain ranges.", + "bbox": [ + 498, + 439, + 890, + 530 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 545, + 619, + 561 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We present Class Balanced Adaptive pseudo labeling (CBAFed) for Federated Semi-Supervised Learning (FSSL). In CBAFed, a fixed pseudo labeling strategy is proposed to handle the catastrophic forgetting problem. To deal with the Non-IID setting of FSSL, we propose a class balanced adaptive thresholds selection method to choose better pseudo labels. Furthermore, a residual weight connection method is designed to make the model reach better optimum. We evaluate CBAFed on five datasets, whose performances show the superiority of our method.", + "bbox": [ + 498, + 571, + 890, + 720 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Limitations. Like other FSSL methods, a good model in warm-up cannot be guaranteed if the number of data in labeled clients are extremely few, so the final global model may not perform well.", + "bbox": [ + 498, + 723, + 890, + 782 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Acknowledgement", + "text_level": 1, + "bbox": [ + 500, + 799, + 679, + 816 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This work was supported by the National Natural Science Foundation of China (Grant No. 62101191), Shanghai Natural Science Foundation (Grant No. 21ZR1420800), and the Science and Technology Commission of Shanghai Municipality (Grant No. 22DZ2229004).", + "bbox": [ + 498, + 825, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "16299", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 174, + 106 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] David Berthelot, Nicholas Carlini, Ekin D Cubuk, Alex Kurakin, Kihyuk Sohn, Han Zhang, and Colin Raffel. Remixmatch: Semi-supervised learning with distribution alignment and augmentation anchoring. arXiv preprint arXiv:1911.09785, 2019. 2", + "[2] David Berthelot, Nicholas Carlini, Ian Goodfellow, Nicolas Papernot, Avital Oliver, and Colin A Raffel. Mixmatch: A holistic approach to semi-supervised learning. Advances in neural information processing systems, 32, 2019. 2, 4", + "[3] Xiaokang Chen, Yuhui Yuan, Gang Zeng, and Jingdong Wang. Semi-supervised semantic segmentation with cross pseudo supervision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2613-2622, 2021. 2", + "[4] Wenhui Cui, Yanlin Liu, Yuxing Li, Menghao Guo, Yiming Li, Xiuli Li, Tianle Wang, Xiangzhu Zeng, and Chuyang Ye. Semi-supervised brain lesion segmentation with an adapted mean teacher model. In International Conference on Information Processing in Medical Imaging, pages 554-565. Springer, 2019. 2", + "[5] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 7", + "[6] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 6, 8", + "[7] Wenke Huang, Mang Ye, and Bo Du. Learn from others and be yourself in heterogeneous federated learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10143-10153, 2022. 4", + "[8] Wonyong Jeong, Jaehong Yoon, Eunho Yang, and Sung Ju Hwang. Federated semi-supervised learning with inter-client consistency & disjoint learning. In Proc. ICLR, 2021. 1, 3", + "[9] Sai Praneeth Karimireddy, Satyen Kale, Mehryar Mohri, Sashank Reddi, Sebastian Stich, and Ananda Theertha Suresh. Scaffold: Stochastic controlled averaging for federated learning. In International Conference on Machine Learning, pages 5132-5143. PMLR, 2020. 2", + "[10] Dong-Hyun Lee et al. Pseudo-label: The simple and efficient semi-supervised learning method for deep neural networks. In Workshop on challenges in representation learning, ICML, volume 3, page 896, 2013. 2", + "[11] Qinbin Li, Bingsheng He, and Dawn Song. Model-contrastive federated learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10713–10722, 2021. 2", + "[12] Tian Li, Anit Kumar Sahu, Manzil Zaheer, Maziar Sanjabi, Ameet Talwalkar, and Virginia Smith. Federated optimization in heterogeneous networks. Proceedings of Machine Learning and Systems, 2:429-450, 2020. 2", + "[13] Xiaomeng Li, Lequan Yu, Hao Chen, Chi-Wing Fu, and Pheng-Ann Heng. Transformation consistent self" + ], + "bbox": [ + 76, + 114, + 470, + 902 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "ensembling model for semi-supervised medical image segmentation. IEEE Trans. Neural Networks and Learning Systems, 2020. 2", + "[14] Xiaoxiao Liang, Yiqun Lin, Huazhu Fu, Lei Zhu, and Xiaomeng Li. Rscfed: Random sampling consensus federated semi-supervised learning. In Proc. CVPR, 2022. 1, 3, 6, 7", + "[15] Haowen Lin, Jian Lou, Li Xiong, and Cyrus Shahabi. Semifed: Semi-supervised federated learning with consistency and pseudo-labeling. arXiv preprint arXiv:2108.09412, 2021. 1, 3", + "[16] Fengbei Liu, Yu Tian, Yuanhong Chen, Yuyuan Liu, Vasileios Belagiannis, and Gustavo Carneiro. Acpl: Anticurriculum pseudo-labelling for semi-supervised medical image classification. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20697-20706, 2022. 2", + "[17] Quande Liu, Cheng Chen, Jing Qin, Qi Dou, and Pheng-Ann Heng. Feddg: Federated domain generalization on medical image segmentation via episodic learning in continuous frequency space. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1013–1023, 2021. 2", + "[18] Quande Liu, Hongzheng Yang, Qi Dou, and Pheng-Ann Heng. Federated semi-supervised medical image classification via inter-client relation matching. In Proc. MICCAI, 2021. 1, 2, 3, 6, 7", + "[19] Brendan McMahan, Eider Moore, Daniel Ramage, Seth Hampson, and Blaise Aguera y Arcas. Communication-efficient learning of deep networks from decentralized data. In Artificial intelligence and statistics, pages 1273–1282. PMLR, 2017. 2, 4, 6, 7", + "[20] Yuval Netzer, Tao Wang, Adam Coates, Alessandro Bissacco, Bo Wu, and Andrew Y Ng. Reading digits in natural images with unsupervised feature learning. 2011. 8", + "[21] Krishna Pillutla, Kshitiz Malik, Abdel-Rahman Mohamed, Mike Rabbat, Maziar Sanjabi, and Lin Xiao. Federated learning with partial model personalization. In International Conference on Machine Learning, pages 17716-17758. PMLR, 2022. 4", + "[22] Boris T Polyak and Anatoli B Juditsky. Acceleration of stochastic approximation by averaging. SIAM journal on control and optimization, 30(4):838-855, 1992. 4", + "[23] Liangqiong Qu, Yuyin Zhou, Paul Pu Liang, Yingda Xia, Feifei Wang, Ehsan Adeli, Li Fei-Fei, and Daniel Rubin. Rethinking architecture design for tackling data heterogeneity in federated learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10061–10071, 2022. 2, 4, 7, 8", + "[24] Yinghuan Shi, Jian Zhang, Tong Ling, Jiwen Lu, Yefeng Zheng, Qian Yu, Lei Qi, and Yang Gao. Inconsistency-aware uncertainty estimation for semi-supervised medical image segmentation. IEEE Trans. Medical Imaging, 41(3):608-620, 2022. 2", + "[25] Kihyuk Sohn, David Berthelot, Nicholas Carlini, Zizhao Zhang, Han Zhang, Colin A Raffel, Ekin Dogus Cubuk, Alexey Kurakin, and Chun-Liang Li. Fixmatch: Simplifying semi-supervised learning with consistency and confidence." + ], + "bbox": [ + 501, + 92, + 890, + 902 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "16300", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Advances in neural information processing systems, 33:596-608, 2020. 2, 4", + "[26] Antti Tarvainen and Harri Valpola. Mean teachers are better role models: Weight-averaged consistency targets improve semi-supervised deep learning results. Advances in neural information processing systems, 30, 2017. 2, 4", + "[27] Hongyi Wang, Mikhail Yurochkin, Yuekai Sun, Dimitris Papailiopoulos, and Yasaman Khazaeni. Federated learning with matched averaging. arXiv preprint arXiv:2002.06440, 2020. 2", + "[28] Jianyu Wang, Qinghua Liu, Hao Liang, Gauri Joshi, and H Vincent Poor. Tackling the objective inconsistency problem in heterogeneous federated optimization. Advances in neural information processing systems, 33:7611-7623, 2020. 2", + "[29] Yidong Wang, Bowen Zhang, Wenxin Hou, Zhen Wu, Jindong Wang, and Takahiro Shinozaki. Margin calibration for long-tailed visual recognition. arXiv preprint arXiv:2112.07225, 2021. 5", + "[30] Qizhe Xie, Zihang Dai, Eduard Hovy, Thang Luong, and Quoc Le. Unsupervised data augmentation for consistency training. Advances in Neural Information Processing Systems, 33:6256-6268, 2020. 2", + "[31] Dong Yang, Ziyue Xu, Wenqi Li, Andriy Myronenko, Holger R Roth, Stephanie Harmon, Sheng Xu, Baris Turkbey, Evrim Turkbey, Xiaosong Wang, et al. Federated semi-supervised learning for pandemic region segmentation in chest CT using multi-national data from china, Italy, japan. Medical image analysis, 70:101992, 2021. 1, 2, 3, 6, 7", + "[32] Qiang Yang, Yang Liu, Tianjian Chen, and Yongxin Tong. Federated machine learning: Concept and applications. ACM Transactions on Intelligent Systems and Technology (TIST), 10(2):1-19, 2019. 1", + "[33] Lequan Yu, Shujun Wang, Xiaomeng Li, Chi-Wing Fu, and Pheng-Ann Heng. Uncertainty-aware self-ensembling model for semi-supervised 3d left atrium segmentation. In Proc. MICCAI, 2019. 2", + "[34] Bowen Zhang, Yidong Wang, Wenxin Hou, Hao Wu, Jindong Wang, Manabu Okumura, and Takahiro Shinozaki. Flexmatch: Boosting semi-supervised learning with curriculum pseudo labeling. Advances in Neural Information Processing Systems, 34:18408-18419, 2021. 2, 4, 5", + "[35] Songyang Zhang, Zeming Li, Shipeng Yan, Xuming He, and Jian Sun. Distribution alignment: A unified framework for long-tail visual recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2361-2370, 2021. 5", + "[36] Boyan Zhou, Quan Cui, Xiu-Shen Wei, and Zhao-Min Chen. Bbn: Bilateral-branch network with cumulative learning for long-tailed visual recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9719–9728, 2020. 5", + "[37] Yuyin Zhou, Yan Wang, Peng Tang, Song Bai, Wei Shen, Elliot K. Fishman, and Alan L. Yuille. Semi-supervised 3d abdominal multi-organ segmentation via deep multi-planar co-training. In Proc. WACV, 2019. 2" + ], + "bbox": [ + 78, + 92, + 468, + 881 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "16301", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2023/Class Balanced Adaptive Pseudo Labeling for Federated Semi-Supervised Learning/27b4466b-7a94-415b-9a3c-c8f4f39a02f2_model.json b/2023/Class Balanced Adaptive Pseudo Labeling for Federated Semi-Supervised Learning/27b4466b-7a94-415b-9a3c-c8f4f39a02f2_model.json new file mode 100644 index 0000000000000000000000000000000000000000..56b88189bb0e634c33f67cff2040d025d4783028 --- /dev/null +++ b/2023/Class Balanced Adaptive Pseudo Labeling for Federated Semi-Supervised Learning/27b4466b-7a94-415b-9a3c-c8f4f39a02f2_model.json @@ -0,0 +1,2167 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.116, + 0.13, + 0.856, + 0.177 + ], + "angle": 0, + "content": "Class Balanced Adaptive Pseudo Labeling for Federated Semi-Supervised Learning" + }, + { + "type": "text", + "bbox": [ + 0.35, + 0.204, + 0.631, + 0.223 + ], + "angle": 0, + "content": "Ming Li Qingli Li Yan Wang*" + }, + { + "type": "text", + "bbox": [ + 0.086, + 0.226, + 0.893, + 0.245 + ], + "angle": 0, + "content": "Shanghai Key Laboratory of Multidimensional Information Processing, East China Normal University" + }, + { + "type": "text", + "bbox": [ + 0.198, + 0.247, + 0.782, + 0.261 + ], + "angle": 0, + "content": "lm1640362161@gmail.com, qlli@cs.ecnu.edu.cn, ywang@cee.ecnu.edu.cn" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.296, + 0.314, + 0.312 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.328, + 0.473, + 0.751 + ], + "angle": 0, + "content": "This paper focuses on federated semi-supervised learning (FSSL), assuming that few clients have fully labeled data (labeled clients) and the training datasets in other clients are fully unlabeled (unlabeled clients). Existing methods attempt to deal with the challenges caused by not independent and identically distributed data (Non-IID) setting. Though methods such as sub-consensus models have been proposed, they usually adopt standard pseudo labeling or consistency regularization on unlabeled clients which can be easily influenced by imbalanced class distribution. Thus, problems in FSSL are still yet to be solved. To seek for a fundamental solution to this problem, we present Class Balanced Adaptive Pseudo Labeling (CBAFed), to study FSSL from the perspective of pseudo labeling. In CBAFed, the first key element is a fixed pseudo labeling strategy to handle the catastrophic forgetting problem, where we keep a fixed set by letting pass information of unlabeled data at the beginning of the unlabeled client training in each communication round. The second key element is that we design class balanced adaptive thresholds via considering the empirical distribution of all training data in local clients, to encourage a balanced training process. To make the model reach a better optimum, we further propose a residual weight connection in local supervised training and global model aggregation. Extensive experiments on five datasets demonstrate the superiority of CBAFed. Code will be available at https://github.com/minglllli/CBAFed." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.778, + 0.21, + 0.794 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.804, + 0.47, + 0.88 + ], + "angle": 0, + "content": "Federated learning (FL) aims to train machine learning models on a decentralized manner while preserving data privacy, i.e., separate local models are trained on separate local training datasets independently. In recent years, FL has received much attention for privacy protection reasons" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.297, + 0.893, + 0.389 + ], + "angle": 0, + "content": "[32]. However, most of FL works focused on supervised learning with fully labeled data. But in practice, labeling process of large-scale training data is laborious and expensive. Due to the lack of funds or experts, large labeled training dataset is difficult for many companies and institutions to obtain. These may hinder applicability of FL." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.393, + 0.893, + 0.573 + ], + "angle": 0, + "content": "To handle this problem, federated semi-supervised learning (FSSL) has been explored by many researchers recently [8, 14, 15]. There are broadly three lines of FSSL methods by considering different places and status of labeled data. The first two lines consider that there are only limited labeled data in the central server [8] or each client has partially labeled data [15]. The third line assumes that few clients have fully labeled data and the training datasets in other clients are fully unlabeled [14, 18, 31]. Our paper mainly focuses on the third line of FSSL, and we call local clients with fully labeled data as labeled clients and the other clients as unlabeled clients." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.578, + 0.895, + 0.777 + ], + "angle": 0, + "content": "The main difficulties to train a third line FSSL model lie in three folds: 1) There are no labeled data in unlabeled clients. Thus, the training can be easily biased without label guidance. 2) Due to the divergent class distribution of labeled and unlabeled clients, namely the not independent and identically distributed data (Non-IID) setting, inaccurate supervisory signals may be generated in unlabeled clients via employing the model trained in labeled clients by either pseudo labeling or consistency regularization framework. 3) Due to the catastrophic forgetting problems in CNNs, with the training process of unlabeled clients going on, models may forget the knowledge learned on labeled clients and so decrease the prediction accuracy drastically." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.78, + 0.895, + 0.901 + ], + "angle": 0, + "content": "RSCFed [14], the state-of-the-art method significantly boosts the FSSL performance by first distilling sub-consensus models, and then aggregating the sub-consensus models to the global model. The sub-consensus models can handle the Non-IID setting to some extent, but the meanteacher based consistency regularization framework in unlabeled clients inevitably causes the accuracy degradation when the classes are imbalanced distributed. RSCFed at" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.887, + 0.227, + 0.9 + ], + "angle": 0, + "content": "*Corresponding Author." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "16292" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.242 + ], + "angle": 0, + "content": "tempt to address the catastrophic forgetting problem by adjusting aggregation weights for labeled and unlabeled clients. But, it seems just alleviates the negative impact on the unlabeled clients and the problem is still yet to be tackled. Besides, the sub-consensus models may occur unstable training problems, and increase the communication cost. Other methods like FedConsist [31] and FedIRM [18] achieve promising results by utilizing pseudo labeling methods to produce artificial labels, but they do not consider the Non-IID setting among local clients." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.243, + 0.473, + 0.41 + ], + "angle": 0, + "content": "Pseudo labeling methods are shown to be effective in semi-supervised learning (SSL) [13,24,33,37], which generate pseudo-labels for unlabeled images and propose several strategies to ensure the quality of pseudo-labels. Existing FSSL works only adopt standard pseudo labeling or consistency regularization on FSSL. Since the mentioned difficulties hamper the usage of these methods, other remedy is usually proposed to alleviate the difficulties, which is palliative. Thus, we propose to study FSSL from the perspective of pseudo labeling, seeking for a fundamental solution to this problem." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.41, + 0.473, + 0.817 + ], + "angle": 0, + "content": "Concretely, we present Class Balanced Adaptive Pseudo Labeling, namely CBAFed, by rethinking standard pseudo labeling methods in SSL. To handle the catastrophic forgetting problem, we propose a fixed pseudo labeling strategy, which builds a fixed set by letting pass informative unlabeled data and their pseudo labels at the beginning of the unlabeled client training. Due to the Non-IID and heterogeneous data partition problems in FL, training distribution of unlabeled data can be highly imbalanced, so existing thresholds are not suitable in FSSL. We design class balanced adaptive thresholds via considering the empirical distribution of all training data in local clients at the previous communication round. Analysis proves that our method sets a reasonably high threshold for extremely scarce classes and encourages a balanced training process. To enhance the learning ability and discover unlabeled data from tail classes, we propose to leverage information from so-called \"not informative\" unlabeled data. Besides, we also explore a novel training strategy for labeled clients and the central server, termed as residual weight connection, skip connecting weights from previous epochs (for labeled clients) or previous global models (for the central server). It can help the model reach better optimum, when the training distribution is imbalanced and training amount is small. We conduct extensive experiments on five datasets to show the effectiveness of CBAFed. Overall, our main contributions can be summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.825, + 0.472, + 0.902 + ], + "angle": 0, + "content": "- We present a CBAFed method to deal with the catastrophic forgetting problems in federated learning. Unlike existing FSSL frameworks that directly adopts pseudo labeling or consistency regularization methods, CBAFed explores a fundamental solution to FSSL via" + }, + { + "type": "text", + "bbox": [ + 0.532, + 0.092, + 0.816, + 0.108 + ], + "angle": 0, + "content": "designing a novel pseudo labeling method." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.116, + 0.892, + 0.191 + ], + "angle": 0, + "content": "- We introduce a residual weight connection method, to improve the robustness of the models in labeled clients and the central server, which skip connects weights from previous epoch or communication round to finally reach better optimum." + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.2, + 0.894, + 0.277 + ], + "angle": 0, + "content": "- Experiments are conducted on five datasets: four natural datasets CIFAR-10/100, SVHN, fashion MNIST and one medical dataset ISIC 2018 Skin. CBAFed outperforms state-of-the-art FSSL methods by a large margin, with \\(11.3\\%\\) improvements on SVHN dataset." + }, + { + "type": "list", + "bbox": [ + 0.518, + 0.116, + 0.894, + 0.277 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.301, + 0.638, + 0.317 + ], + "angle": 0, + "content": "2. Related work" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.326, + 0.691, + 0.342 + ], + "angle": 0, + "content": "2.1. Federated Learning" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.35, + 0.892, + 0.547 + ], + "angle": 0, + "content": "Recently, federated learning is becoming more and more popular for privacy protection reasons. Statistical data heterogeneity is one of the most important research problems. FedAvg [19], one of the pioneer works to deal with this problem, performs weighted-averaging on local weights according to the local training size, has been used as the most widely recognized FL baseline. To deal with data heterogeneity, two research directions have been mostly explored: model aggregation [12, 27, 28] and local training [9, 11, 12, 17]. Recently, [23] demonstrates that self-attention-based architectures, e.g., vision transformers(ViT) are more robust to distribution shifts and can converge to better optimum over heterogeneous data." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.557, + 0.731, + 0.573 + ], + "angle": 0, + "content": "2.2. Semi-supervised Leaning" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.58, + 0.892, + 0.836 + ], + "angle": 0, + "content": "The goal of standard semi-supervised learning (SSL) is to train a good model with limited labeled data and much more unlabeled data. Popular research directions involves consistency regularization [4, 26] and pseudo labeling [3, 10, 16]. Many recent works use the combinations of consistency regularization and pseudo labeling and achieves state-of-the-art performance [1, 2, 25, 30]. In this paper, we focus on using pseudo labeling method to deal with SSL. While these methods achieve much progress, they only consider pre-defined fixed threshold for pseudo labeling. FlexMatch [34] demonstrates that leveraging unlabeled data according to the model's learning status to flexibly adjust thresholds for different classes can boost performance. While these methods perform well in centralized SSL, they all update pseudo labels after every batch's update of the model, which is not suitable in FSSL as shown in later section." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.847, + 0.821, + 0.864 + ], + "angle": 0, + "content": "2.3. Federated Semi-supervised Learning" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.892, + 0.902 + ], + "angle": 0, + "content": "Recently, with the development of federated learning, FSSL becomes popular. There are broadly three categories" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16293" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.168, + 0.089, + 0.478, + 0.287 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.481, + 0.089, + 0.807, + 0.287 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.304, + 0.895, + 0.39 + ], + "angle": 0, + "content": "Figure 1. An overview of our CBAFed. In the central server (left side), the global model is aggregated with the returned local models (step ④) and the adaptive thresholds are calculated by the returned training data statistics (step ⑤). Then central server passes the global model, adaptive thresholds and class distribution to all local clients (step ①). After downloading these data, local clients perform local training on the right side (step ②). Labeled clients use labeled data to train the model with residual weight connection. Unlabeled clients obtain the new training dataset by adaptive pseudo labeling and tail class data discovery and use it to train the model. After local training, local clients return trained models and number of data in each class back to central server (step ③)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.409, + 0.473, + 0.637 + ], + "angle": 0, + "content": "of FSSL. The first two categories assume that every local client contains partially labeled data or there exists some labeled data in central server [8, 15]. The third category considers that few clients are fully labeled while other clients contain only unlabeled data. This paper focuses on the third category. To deal with this problem, Fed-Consist [31] and FedIRM [18] have been proposed. But, these two methods do not consider data heterogeneity in federated learning. Recently, RSCFed [14] proposes to perform random sub-sampling to reach consensus over clients. However, it uses standard consistency regularization for unlabeled data, which still suffers from the Non-IID setting. We emphasize that all the aforementioned approaches just use traditional pseudo labeling strategy, and none of them discuss how to apply pseudo labeling to FSSL." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.647, + 0.214, + 0.664 + ], + "angle": 0, + "content": "3. Methodology" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.672, + 0.47, + 0.841 + ], + "angle": 0, + "content": "We first introduce the problem setting of federated semi-supervised learning. We consider decentralized training with labeled clients and unlabeled clients. Let \\(m\\) (usually \\(m = 1\\)) labeled clients be \\(\\{C_1, C_2, \\dots, C_m\\}\\) and \\(n\\) (usually \\(n > 1\\)) unlabeled clients be \\(\\{C_{m+1}, C_{m+2}, \\dots, C_{m+n}\\}\\). For labeled client \\(C_\\ell\\), the labeled local training dataset is denoted as \\(D_\\ell = \\{(X_i^\\ell, y_i^\\ell)\\}_{i=1}^{N_\\ell}\\), \\(\\ell = 1, 2, \\dots, m\\), where \\(N_\\ell\\) denotes the data number in \\(D_\\ell\\). Similarly, for unlabeled client \\(C_\\mu\\), the unlabeled local training dataset is denoted as \\(D_\\mu = \\{X_i^\\mu\\}_{i=1}^{N_\\mu}\\), \\(\\mu = m + 1, m + 2, \\dots, n + m\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.841, + 0.471, + 0.902 + ], + "angle": 0, + "content": "The overall pipeline of the proposed CBAFed is shown in Fig. 1, which performs the following steps: (1) Warm up stage: train fully supervised models on only labeled clients using residual weight connection in a normal federated" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.409, + 0.895, + 0.697 + ], + "angle": 0, + "content": "learning manner (Sec. 3.1) (2) The central server computes the empirical class distribution and obtains the class balanced adaptive thresholds, then passes them to local clients (Sec. 3.3); (3) All local clients update local models, adaptive threshold and class distribution. Labeled clients train local models on all the data using proposed residual weight connection; Unlabeled clients acquire the fixed training set (Sec. 3.2.2) by the threshold and the tail class datasets (Sec. 3.4), and train local models on the newly obtained training dataset. After local training, all local clients pass the trained model, the number of data in each class and the amount of training data to the central server; (4) The central server aggregates a new model with residual weight connection, computes the class distribution, and obtains the class balanced adaptive threshold. Then, the central server passes them to local clients; (5) Repeat step (3)-(4) until the specified number of communication round is reached. The overall training procedure of our CBAFed is provided in Algorithm 1 in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.721, + 0.757, + 0.737 + ], + "angle": 0, + "content": "3.1. Residual Weight Connection" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.895, + 0.903 + ], + "angle": 0, + "content": "We propose a novel training method in local supervised training and global model aggregation called residual weight connection. Our idea of res-weight is simple and similar to ResNet. In ResNet, there is a skip connection between every layer. For our res-weight connection, there is a skip connection of model's parameters between training epochs (or communication rounds). Let \\(\\theta^{E}\\) be the parameters of model after the training epoch (communication round) \\(E(E > 0)\\), then the final parameters of model after this training epoch by using residual weight connection will" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "16294" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.078, + 0.093, + 0.099, + 0.105 + ], + "angle": 0, + "content": "be" + }, + { + "type": "equation", + "bbox": [ + 0.127, + 0.103, + 0.47, + 0.143 + ], + "angle": 0, + "content": "\\[\n\\theta^ {E} = \\left\\{ \\begin{array}{l l} \\theta^ {E} & E \\% s \\neq 0 \\\\ \\alpha_ {1} \\theta^ {E - s} + (1 - \\alpha_ {1}) \\theta^ {E} & E \\% s = 0 \\end{array} \\right. \\tag{1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.151, + 0.473, + 0.332 + ], + "angle": 0, + "content": "where \\(\\alpha_{1}\\) is the parameter to scale the size of \\(\\theta^{E - s}\\) and \\(\\theta^{E}\\) and \\(s\\) is the number of skip epoch. In [22], authors show that averaging model weights over training steps tends to produce a more accurate model than using the final weights directly. In our experiments, we will show that this training method can effectively make the model more robust and finally reach better optimum, enhancing the predicted accuracy of pseudo labeling. Note that although the formula of our residual weight connection is similar to mean teachers [26], the update strategy and usage scenario is very different from it. Please refer to supplementary material for more detailed comparison and discussions." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.343, + 0.314, + 0.358 + ], + "angle": 0, + "content": "3.2. Pseudo Labeling Methods" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.366, + 0.465, + 0.382 + ], + "angle": 0, + "content": "3.2.1 Review of Classic Batch-based Pseudo Labeling" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.39, + 0.47, + 0.481 + ], + "angle": 0, + "content": "In semi-supervised learning, for unlabeled data, traditional approaches [2, 25, 34], such as Fixmatch [25] and Flexmatch [34], use original data or their weak augmented version in one batch to generate pseudo labels. These labels are adopted to supervise model's training. For detailed review, please refer to supplementary materials." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.5, + 0.351, + 0.515 + ], + "angle": 0, + "content": "3.2.2 Fixed Pseudo Labeling in FSSL" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.524, + 0.47, + 0.6 + ], + "angle": 0, + "content": "Warm up stage. To conduct pseudo labeling for unlabeled data, in the warm up stage, we train our model only with labeled data. In the first \\(P\\) communication round, training only conducted in labeled clients. In local training on labeled client \\(C_{\\ell}\\), the training loss is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.123, + 0.61, + 0.47, + 0.649 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\ell} = \\frac {1}{| D _ {\\ell} |} \\sum_ {(X ^ {\\ell}, y ^ {\\ell}) \\in D _ {\\ell}} H \\left(y ^ {\\ell}, p _ {m} \\left(y \\mid \\theta_ {t} ^ {\\ell} \\left(X ^ {\\ell}\\right)\\right)\\right), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.662, + 0.469, + 0.72 + ], + "angle": 0, + "content": "where \\(\\theta_t^\\ell\\) means local model in labeled client \\(C_\\ell\\) at communication round \\(t\\). The local training epoch is set to \\(J(J > 1)\\) with residual weight convection. Then we use FedAVG [19] with residual weight connection for model aggregation:" + }, + { + "type": "equation", + "bbox": [ + 0.082, + 0.729, + 0.469, + 0.779 + ], + "angle": 0, + "content": "\\[\n\\theta_ {t + 1} ^ {G} = \\left\\{ \\begin{array}{l l} \\sum_ {\\ell = 1} ^ {m} \\frac {| D _ {\\ell} |}{\\sum_ {i = 1} ^ {m} | D _ {i} |} \\theta_ {t} ^ {\\ell} & t \\% s \\neq 0 \\\\ \\alpha_ {1} \\theta_ {t + 1 - s} ^ {G} + (1 - \\alpha_ {1}) \\sum_ {\\ell = 1} ^ {m} \\frac {| D _ {\\ell} |}{\\sum_ {i = 1} ^ {m} | D _ {i} |} \\theta_ {t} ^ {\\ell} & t \\% s = 0 \\end{array} \\right. \\tag{3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.78, + 0.469, + 0.84 + ], + "angle": 0, + "content": "where \\(\\theta_{t + 1}^{G}\\) is the global model of communication round \\(t+\\) 1. Note that, usually, the number of labeled clients is only 1, so we can finish warm up stage in only one communication round(training can be completed on one labeled client)." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.84, + 0.47, + 0.901 + ], + "angle": 0, + "content": "Different from traditional semi-supervised learning, in federated semi-supervised learning, Non-IID data partitions in clients can easily lead to catastrophic forgetting [7, 21, 23]. Thus, local models trained on unlabeled clients" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.228 + ], + "angle": 0, + "content": "usually forget the former learned knowledge by batch-based pseudo labeling method, which will lead to abrupt degradation and sabotage the training of unlabeled clients. To handle heterogeneous data partitions and catastrophic forgetting, we propose a new pseudo labeling method called fixed pseudo labeling. For the unlabeled training dataset \\(D_{\\mu}\\) in an unlabeled client \\(C_{\\mu}\\), after initializing the local model \\(\\theta_t^\\mu\\) with the global model \\(\\theta_t^G\\), we first compute the pseudo label for each data \\(X_i^\\mu\\) in \\(D_{\\mu}\\) via:" + }, + { + "type": "equation", + "bbox": [ + 0.526, + 0.234, + 0.892, + 0.252 + ], + "angle": 0, + "content": "\\[\n\\hat {y} _ {i} ^ {\\mu} = \\arg \\max p _ {m} (y | \\theta_ {t} ^ {\\mu} \\left(X _ {i} ^ {\\mu}\\right), i = 1, 2, \\dots , N _ {\\mu}. \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.256, + 0.892, + 0.286 + ], + "angle": 0, + "content": "We obtain a subset of unlabeled data with their pseudo labels as supervisory signals during local client training:" + }, + { + "type": "equation", + "bbox": [ + 0.5, + 0.29, + 0.892, + 0.322 + ], + "angle": 0, + "content": "\\[\n\\widetilde {\\mathcal {D}} _ {\\mu} = \\left\\{\\left(X _ {i} ^ {\\mu}, \\hat {y} _ {i} ^ {\\mu}\\right) \\mid X _ {i} ^ {\\mu} \\in D _ {\\mu} \\wedge \\max \\left(p _ {m} \\left(y \\mid \\theta_ {t} ^ {\\mu} \\left(X _ {i} ^ {\\mu}\\right)\\right)\\right) > \\tau \\right\\} _ {i = 1} ^ {N _ {\\mu}}. \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.321, + 0.892, + 0.373 + ], + "angle": 0, + "content": "Then we use this fixed set \\(\\widetilde{\\mathcal{D}}_{\\mu}\\) as the training dataset and define the loss to train our unlabeled client as: \\(\\mathcal{L}_{\\mu} = \\frac{1}{|\\widetilde{\\mathcal{D}}_{\\mu}|}\\sum_{(X^{\\mu},\\hat{y}^{\\mu})\\in \\widetilde{\\mathcal{D}}_{\\mu}}H(\\hat{y}^{\\mu},p_m(y|\\theta_t^\\mu (X^\\mu))).\\)" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.378, + 0.892, + 0.41 + ], + "angle": 0, + "content": "3.3. Class Balanced Adaptive Threshold for Pseudo Labeling" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.417, + 0.893, + 0.779 + ], + "angle": 0, + "content": "In semi-supervised learning, the value of threshold \\(\\tau\\) is non-trivial for selecting confident pseudo labels, namely informative unlabeled data. Traditionally, a fixed threshold is pre-defined [25], but setting a fixed threshold usually makes the model fail to consider different learning status and learning difficulties of different classes. In one of the most recent works, Curriculum Pseudo Labeling [34] is proposed to achieve state-of-the-art results on most semi-supervised learning benchmarks, where a flexible threshold for class \\(c\\) at time step \\(t\\) is calculated by: \\(\\mathcal{T}_t(c) = \\beta_t(c)\\cdot \\tau\\), where \\(\\beta_{t}(c)\\) is the ratio of the number of selected pseudo labels in class \\(c\\) and the maximum number of selected pseudo labels of all classes. \\(\\beta_{t}(c)\\) is used to scale the fixed threshold \\(\\tau\\). Curriculum Pseudo Labeling significantly boosts the accuracy and convergence performance of various semi-supervised learning algorithms. But in FSSL, there exist heterogeneous data partition problems. Due to the Non-IID partition, the labeled data are not balanced, so purely using the number of selected unlabeled data to design threshold is improper. Moreover, the data of some classes can be extremely rare in certain clients, which will lead to a very low \\(\\beta_{t}(c)\\), and thus, a very low threshold \\((\\mathcal{T}_t(c)\\approx 0)\\). In other words, directly applying Curriculum Pseudo Labeling will introduce many noisy labels into training." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.779, + 0.892, + 0.839 + ], + "angle": 0, + "content": "To this end, we propose a novel Class Balanced Adaptive threshold for Pseudo Labeling (CBAPL). For an unlabeled client \\( C_{\\mu} \\) in the \\( t \\)th communication round, let the number of selected pseudo labels in class \\( c \\) be:" + }, + { + "type": "equation", + "bbox": [ + 0.502, + 0.845, + 0.892, + 0.901 + ], + "angle": 0, + "content": "\\[\n\\sigma_ {t} ^ {\\mu} (c) = \\sum_ {i = 1} ^ {N _ {\\mu}} \\mathbf {1} \\left(\\max \\left(p _ {m} \\left(y \\mid \\theta_ {t} ^ {\\mu} \\left(X _ {i} ^ {\\mu}\\right)\\right)\\right) > \\mathcal {T} _ {t} (c)\\right) \\mathbf {1} \\left(\\hat {y} _ {i} ^ {\\mu} = c\\right), \\tag {6}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16295" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.473, + 0.184 + ], + "angle": 0, + "content": "where \\(\\mathbf{1}(\\cdot)\\) is the indicator function. The data number of class \\(c\\) in labeled client \\(C_{\\ell}\\) is calculated by: \\(\\sigma_t^\\ell (c) = \\sum_{i = 1}^{N_\\ell}\\mathbf{1}(y_i^\\ell = c)\\). After local training, \\(\\sigma_t^\\mu (c),\\sigma_t^\\ell (c)\\) and weights of local models are returned to central server. Then the central server calculates the total number of training data of class \\(c\\) at communication round \\(t\\):" + }, + { + "type": "equation", + "bbox": [ + 0.155, + 0.194, + 0.47, + 0.236 + ], + "angle": 0, + "content": "\\[\n\\sigma_ {t} (c) = \\sum_ {\\ell = 1} ^ {m} \\sigma_ {t} ^ {\\ell} (c) + \\sum_ {\\mu = m + 1} ^ {n + m} \\sigma_ {t} ^ {\\mu} (c). \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.246, + 0.47, + 0.277 + ], + "angle": 0, + "content": "We then normalize the training numbers and acquire the empirical distribution of training data:" + }, + { + "type": "equation", + "bbox": [ + 0.2, + 0.287, + 0.47, + 0.325 + ], + "angle": 0, + "content": "\\[\n\\tilde {p} _ {t} (c) = \\frac {\\sigma_ {t} (c)}{\\sum_ {i = 1} ^ {C} \\sigma_ {t} (i)}. \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.334, + 0.47, + 0.366 + ], + "angle": 0, + "content": "The standard deviation of the empirical distribution is calculated by:" + }, + { + "type": "equation", + "bbox": [ + 0.145, + 0.376, + 0.47, + 0.424 + ], + "angle": 0, + "content": "\\[\n\\operatorname {s t d} \\left(\\widetilde {p} _ {t}\\right) = \\sqrt {\\frac {1}{C - 1} \\sum_ {c = 1} ^ {C} \\left(\\widetilde {p} _ {t} (c) - \\bar {p} _ {t}\\right) ^ {2}}, \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.437, + 0.47, + 0.469 + ], + "angle": 0, + "content": "where \\(\\overline{p}_t = \\frac{1}{C}\\sum_{c = 1}^{C}\\widetilde{p}_t(c)\\). Lastly, the threshold of class \\(c\\) is calculated by" + }, + { + "type": "equation", + "bbox": [ + 0.18, + 0.481, + 0.469, + 0.5 + ], + "angle": 0, + "content": "\\[\n\\tau_ {t, c} = \\widetilde {p} _ {t} (c) + \\tau - s t d (\\widetilde {p} _ {t}), \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.51, + 0.47, + 0.54 + ], + "angle": 0, + "content": "where \\(\\tau\\) is the pre-defined threshold base. We then set an upper bound of threshold as:" + }, + { + "type": "equation", + "bbox": [ + 0.171, + 0.551, + 0.469, + 0.592 + ], + "angle": 0, + "content": "\\[\n\\mathcal {T} _ {t + 1} (c) = \\left\\{ \\begin{array}{l l} \\tau_ {t, c}, & \\tau_ {t, c} < \\tau_ {h} \\\\ \\tau_ {h}, & \\tau_ {t, c} \\geq \\tau_ {h} \\end{array} , \\right. \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.603, + 0.47, + 0.708 + ], + "angle": 0, + "content": "where \\(\\tau_h\\) is the pre-defined upper bound of threshold (usually \\(\\tau_h = 0.95\\) for ensuring a certain amount of training data in unlabeled clients). If our model is trained with more data in one class, the threshold of that class will be higher since \\(\\widetilde{p}_t(c)\\) is higher. \\(std(\\widetilde{p}_t)\\) is important for balancing the empirical distribution of training data. Please refer to the supplementary material for more discussions." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.709, + 0.47, + 0.768 + ], + "angle": 0, + "content": "We argue that for scarce classes, unlike Curriculum Pseudo Labeling [34] whose threshold is extremely low, our CBAPL calculates a threshold with a relatively high lower bound. We have theorem below:" + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.78, + 0.175, + 0.793 + ], + "angle": 0, + "content": "Theorem 3.1." + }, + { + "type": "equation", + "bbox": [ + 0.14, + 0.804, + 0.47, + 0.838 + ], + "angle": 0, + "content": "\\[\n\\tau + \\widetilde {p} _ {t} (c) - \\sqrt {\\frac {1}{C}} \\leq \\mathcal {T} _ {t} (c) \\leq \\tau + \\widetilde {p} _ {t} (c), \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.848, + 0.418, + 0.863 + ], + "angle": 0, + "content": "Proof is provided in the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.865, + 0.47, + 0.902 + ], + "angle": 0, + "content": "Since \\(\\tau >>\\sqrt{\\frac{1}{C}}\\) \\(\\mathcal{T}_t(c)\\) will have a high lower bound \\((>>0)\\). To balance the trade-off between \\(\\widetilde{p}_t(t)\\) and" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.091, + 0.892, + 0.123 + ], + "angle": 0, + "content": "\\(\\tau - \\sqrt{1 / C}\\), a normalization factor is applied to consider the class number. Thus, we propose a modified \\(\\widetilde{p}_t(c)\\) as below:" + }, + { + "type": "equation", + "bbox": [ + 0.602, + 0.132, + 0.892, + 0.168 + ], + "angle": 0, + "content": "\\[\n\\widetilde {p} _ {t} (c) = \\frac {\\sigma_ {t} (c)}{\\sum_ {i = 1} ^ {C} \\sigma_ {t} (i)} \\times \\frac {C}{1 0}. \\tag {13}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.177, + 0.892, + 0.268 + ], + "angle": 0, + "content": "At the first \\(P\\) communication rounds, the models are trained on only labeled clients. Thus, when \\(t \\in \\{1, \\dots, P\\}\\), \\(\\sigma_t^\\mu(c) = 0\\) for all classes. After calculating the thresholds for all classes via Eq. 11, the central server will pass the thresholds to local clients. Unlabeled clients further obtain the fixed pseudo label training dataset by:" + }, + { + "type": "equation", + "bbox": [ + 0.517, + 0.277, + 0.892, + 0.319 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\widetilde {\\mathcal {D}} _ {t + 1, \\mu} = \\left\\{\\left(X _ {i} ^ {\\mu}, \\hat {y} _ {i} ^ {\\mu}\\right) \\mid X _ {i} ^ {\\mu} \\in D _ {\\mu} \\right. \\tag {14} \\\\ \\wedge \\max (p _ {m} (y | \\theta_ {t} ^ {\\mu} (X _ {i} ^ {\\mu}))) > \\mathcal {T} _ {t + 1} (\\hat {y} _ {i} ^ {\\mu}) \\} _ {i = 1} ^ {N _ {\\mu}}. \\\\ \\end{array}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.327, + 0.892, + 0.344 + ], + "angle": 0, + "content": "3.4. Discovery of Unlabeled Data from Tail Classes" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.351, + 0.892, + 0.638 + ], + "angle": 0, + "content": "In FSSL, heterogeneous data partitions will lead to class imbalanced distribution in local clients and there are rare or even no data from some classes in certain clients. For warm up stage in labeled clients, it is similar to long-tailed classification, so the problems in long-tailed classification will also exist in our pseudo labeling process, i.e., models tend to classify tail (rare) classes as head (common) classes [29, 35, 36]. Although adaptive pseudo labeling can lower the threshold of tail classes, much less data will be selected by PL since very few data from tailed classes will be classified as correct classes. To enhance the learning ability and discover unlabeled data from these classes, instead of directly dropping the data whose maximum confidences are low, namely, \"not informative\" unlabeled data, we propose to leverage information from them. For simplicity, we drop subscript \\(t\\) and default all representations occur in \\(t\\)th communication round in Sec. 3.4. Firstly, we define a mask function \\(\\mathcal{M}(\\cdot): \\mathcal{R}^C \\to \\mathcal{R}^C\\) for a probability simplex \\(p \\in \\mathcal{R}^C\\):" + }, + { + "type": "equation", + "bbox": [ + 0.586, + 0.647, + 0.892, + 0.688 + ], + "angle": 0, + "content": "\\[\n\\mathcal {M} _ {i} (p) = \\left\\{ \\begin{array}{l l} p _ {i} & i \\neq \\arg \\max p \\\\ 0 & i = \\arg \\max p \\end{array} , \\right. \\tag {15}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.697, + 0.892, + 0.756 + ], + "angle": 0, + "content": "where \\( p_i \\) is the \\( i \\)th element in \\( p \\). The effect of mask function \\( \\mathcal{M}(\\cdot) \\) is to set the maximum of a probability simplex to 0. Then, we turn to analyze the second largest confidence score via:" + }, + { + "type": "equation", + "bbox": [ + 0.521, + 0.763, + 0.892, + 0.829 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} D _ {\\mu} ^ {t a i l} = \\left\\{\\left(X _ {i} ^ {\\mu}, \\hat {y} _ {i} ^ {\\mu^ {\\prime}}\\right) \\mid X _ {i} ^ {\\mu} \\in D _ {\\mu} \\right. \\\\ \\left. \\wedge \\max \\left(p _ {m} \\left(y \\mid \\theta_ {t} ^ {\\mu} \\left(X _ {i} ^ {\\mu}\\right)\\right)\\right) \\leq \\mathcal {T} _ {t} \\left(\\hat {y} _ {i} ^ {\\mu}\\right) \\wedge \\widetilde {p} _ {t} \\left(\\hat {y} _ {i} ^ {\\mu^ {\\prime}}\\right) < \\frac {\\beta}{C} \\right\\}, \\tag {16} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.84, + 0.892, + 0.902 + ], + "angle": 0, + "content": "where \\(\\hat{y}_i^{u'} = \\arg \\max \\mathcal{M}(p_m(y|\\theta_t^\\mu(X_i^\\mu)))\\) and \\(\\beta\\) is a hyperparameter. In Eq. 15, the parameter \\(\\beta\\) is to find the tail classes (\"not informative\" unlabeled data). For unlabeled data \\(X_i^\\mu\\), if its largest confidence is smaller than threshold" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "16296" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.091, + 0.473, + 0.184 + ], + "angle": 0, + "content": "(not informative or confident, i.e., \\(\\max(p_m(y|\\theta_t^\\mu(X_i^\\mu))) \\leq \\mathcal{T}_t(\\hat{y}_i^\\mu)\\)) and the class of second largest confidence is tail class (i.e., \\(\\widetilde{p}_t(\\hat{y}_i^{\\mu \\prime}) < \\frac{\\beta}{C}\\)), we regard it as misclassified data and include it in training with the class of second largest confidence as its label. Finally, the new training dataset for unlabeled client \\(C_\\mu\\) is defined as" + }, + { + "type": "equation", + "bbox": [ + 0.197, + 0.195, + 0.47, + 0.216 + ], + "angle": 0, + "content": "\\[\nD _ {\\mu} ^ {t r a i n} = D _ {\\mu} ^ {t a i l} \\cup \\widetilde {D} _ {\\mu}, \\tag {17}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.226, + 0.387, + 0.243 + ], + "angle": 0, + "content": "and the final training loss in \\( C_{\\mu} \\) is rewritten as" + }, + { + "type": "equation", + "bbox": [ + 0.09, + 0.253, + 0.47, + 0.305 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mu} = \\frac {1}{| D _ {\\mu} ^ {t r a i n} |} \\sum_ {(X ^ {\\mu}, \\hat {y} ^ {\\mu}) \\in D _ {\\mu} ^ {t r a i n}} H \\left(\\hat {y} ^ {\\mu}, p _ {m} \\left(y \\mid \\theta_ {t} ^ {\\mu} \\left(X ^ {\\mu}\\right)\\right)\\right). \\tag {18}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.306, + 0.332, + 0.322 + ], + "angle": 0, + "content": "The calculation of \\(\\sigma_t^u (c)\\) is revised as:" + }, + { + "type": "equation", + "bbox": [ + 0.154, + 0.333, + 0.469, + 0.368 + ], + "angle": 0, + "content": "\\[\n\\sigma_ {t} ^ {\\mu} (c) = \\sum_ {(X ^ {\\mu}, \\hat {y} ^ {\\mu}) \\in D _ {\\mu} ^ {t r a i n}} \\mathbf {1} (\\hat {y} ^ {\\mu} = c). \\tag {19}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.378, + 0.327, + 0.394 + ], + "angle": 0, + "content": "3.5. Aggregation of local models" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.401, + 0.47, + 0.448 + ], + "angle": 0, + "content": "Since not all data in local clients are included into training, we propose to use only data included in training for model aggregation. The scale factor \\( w_{t}^{i} \\) is computed by" + }, + { + "type": "equation", + "bbox": [ + 0.102, + 0.457, + 0.47, + 0.515 + ], + "angle": 0, + "content": "\\[\nw _ {t} ^ {i} = \\left\\{ \\begin{array}{l l} \\frac {\\left| D _ {i} \\right|}{\\left| D _ {t} ^ {t r a i n} \\right|} & \\text {i f} i \\in \\{1, \\dots , m \\} \\\\ \\frac {\\left| D _ {t , i} ^ {t r a i n} \\right|}{\\left| D _ {t} ^ {t r a i n} \\right|} & \\text {i f} i \\in \\{m + 1, \\dots , m + n \\}, \\end{array} \\right. \\tag {20}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.525, + 0.47, + 0.573 + ], + "angle": 0, + "content": "where \\( |D_t^{train}| = \\sum_{\\ell=1}^m |D_\\ell| + \\sum_{\\mu=m+1}^{m+n} |D_{t,\\mu}^{train}| \\). Then at \\( (t+1) \\)th synchronization round, the global model \\( \\theta_{t+1}^G \\) is computed by" + }, + { + "type": "equation", + "bbox": [ + 0.085, + 0.582, + 0.469, + 0.62 + ], + "angle": 0, + "content": "\\[\n\\theta_ {t + 1} ^ {G} = \\left\\{ \\begin{array}{l l} \\sum_ {i = 1} ^ {m + n} w _ {t} ^ {i} \\theta_ {t} ^ {i} & t \\% s \\neq 0 \\\\ \\alpha_ {2} \\theta_ {t + 1 - s} ^ {G} + (1 - \\alpha_ {2}) \\sum_ {i = 1} ^ {m + n} w _ {t} ^ {i} \\theta_ {t} ^ {i} & t \\% s = 0, \\end{array} \\right. \\tag{21}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.082, + 0.628, + 0.284, + 0.643 + ], + "angle": 0, + "content": "where \\(\\alpha_{2}\\) is a hyperparameter." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.657, + 0.21, + 0.673 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.681, + 0.267, + 0.698 + ], + "angle": 0, + "content": "4.1. Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.705, + 0.469, + 0.81 + ], + "angle": 0, + "content": "Benchmark Datasets To evaluate the effectiveness of our proposed method, we conduct extensive experiments on four image classification datasets, i.e., SVHN, CIFAR-10, CIFAR-100, Fashion MNIST and one medical image classification dataset: ISIC 2018 (Skin Lesion Analysis Towards Melanoma Detection). Dataset splitting and pre-processing are provided in the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.811, + 0.469, + 0.871 + ], + "angle": 0, + "content": "FSSL Setting Simulation Following [14], training datasets contain 10 clients: one labeled and nine unlabeled. We use a Dirichlet distribution \\( \\text{Dir}(\\gamma) \\), where \\( \\gamma = 0.8 \\) for five datasets [14] to generate Non-IID data partition in clients." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Implementation Details We utilize the SGD optimizer with a momentum of 0.9, and implement our method with" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.091, + 0.701, + 0.24 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.701, + 0.091, + 0.887, + 0.24 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.255, + 0.892, + 0.34 + ], + "angle": 0, + "content": "Figure 2. Test accuracy curves in local training of SVHN dataset w/ and w/o residual weight connection. ResNet18 (a) and ViT (b) are adopted as the backbones. W/ res-weight* indicates we only show test accuracy on epochs (communication rounds, since local training epoch for labeled client is 1) with skip weight connection. Best viewed electronically." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.351, + 0.892, + 0.457 + ], + "angle": 0, + "content": "PyTorch. We adopt ResNet18 [6] from PyTorch for all datasets. For fair comparison, we use the same network architecture and training protocol, including the optimizer, data preprocessing, etc. across all FSSL methods. The local training epoch is set to 11 for labeled clients and 1 for unlabeled clients. More implementation details are provided in supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.457, + 0.893, + 0.715 + ], + "angle": 0, + "content": "Comparisons We compare our method against state-of-the-art FSSL methods, including RSCFed [14], FedIRM [18], and Fed-Consist [31]. Since FedIRM [18] and Fed-Consist [31] are not desgined for dealing with Non-IID FSSL, following [14], we enlarge the weight of labeled client to about \\(50\\%\\) and other nine unlabeled clients share the remaining \\(50\\%\\) weight in each FSSL synchronization round, when implementing these methods. We also compare our network against FedAvg [19] trained with all clients as an upper bound and trained with only one labeled clients as the lower bound. Besides, to show the effectiveness of our residual weight connection, we report the result of FedAvg [19] trained with only one labeled clients using residual weight connection. Since Fed-Consist [31] utilizes traditional batch-based pseudo labeling method, we report the result of Fed-Consist [31] using our proposed fixed pseudo labeling without enlarging the weight of labeled client." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.726, + 0.807, + 0.743 + ], + "angle": 0, + "content": "4.2. Comparisons with State-of-the-arts" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.75, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Comparison Results. Table 1 reports the quantitative results of our CBAFed and other state-of-the-art methods on five benchmarks. We can observe that our proposed CBAFed achieves superior performances over all competitors on the five benchmark datasets. For some easy tasks (e.g., SVHN and Fashion-MNIST), the performance of our CBAFed is approaching upper bound obtained by FedAVG [19] and surpasses all state-of-the-arts by a large margin, showing a strong power of fixed pseudo labeling and class balanced threshold. Note that if replacing the traditional" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16297" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.089, + 0.892, + 0.132 + ], + "angle": 0, + "content": "Table 1. Results on SVHN, CIFAR-10/100, Fashion MNIST and ISIC 2018 datasets under heterogeneous data partition with ResNet18. FedAVG \\(^+\\) means FedAvg [19] trained with all one labeled clients using our residual weight connection. Fed-consist \\(^+\\) means Fed-Consist [31] using our proposed fixed pseudo labeling without enlarging the weight of labeled client." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.142, + 0.894, + 0.267 + ], + "angle": 0, + "content": "
Labeling StrategyMethodClient Num.Dataset
labeledunlabeledSVHNCIFAR10CIFAR100Fashion-MNISTISIC 2018
Fully supervisedFedAvg [19](upper-bound)10091.8380.8951.3890.1481.32
FedAvg [19](lower-bound)1067.7154.6620.4974.8765.13
FedAvg+ [19]1076.9858.2124.8478.2666.69
Semi supervisedFedIRM [18]1969.2252.8420.2076.8364.85
Fed-Consist [31]1970.5654.2321.8176.5765.20
Fed-Consist+ [31]1986.5756.3523.2578.3565.50
RSCFed [14]1976.7457.0728.4678.4067.21
CBAFed(ours)1988.0767.0830.1885.4968.29
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.292, + 0.47, + 0.351 + ], + "angle": 0, + "content": "batch-based pseudo labeling by our fixed pseudo labeling in Fed-Consist [31], the performance is largely increased. It shows that fixed pseudo labeling method is well suited for Non-IID FSSL." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.354, + 0.47, + 0.61 + ], + "angle": 0, + "content": "Trained on one labeled client, FedAVG [19] w/ residual weight connection achieves much better performance compared with that w/o residual weight connection. Fig. 2 (a) shows the test accuracy curves during training. Due to the imbalanced training data distribution in labeled client, the test accuracy is unstable during training for FedAVG (w/o res-weight in the figure). But, if training w/ our residual weight connection, the test accuracy curve is much more stable and the performance is also enhanced (w/ res-weight and w/ res-weight*). To sum up, residual weight connection can improve the robustness of the models when the training amount is small and the training distribution is imbalanced. Since long tailed vision recognition task is similar to local training in labeled client, it is a potential direction to use our residual weight connection in long tailed vision. Besides, we also conduct experiments on partially labeled clients, results are shown in the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.612, + 0.47, + 0.809 + ], + "angle": 0, + "content": "ViT Backbone. Vision Transformer(ViT) has shown to be more robust to heterogeneous data and distribution shifts, and has been demonstrated to converge faster with better optimum in Federated learning [23]. To further study the effect of our method, we conduct experiments on SVHN dataset using ViT-Tiny as backbone for all competitors. The implementation details are provided in the supplementary material. Table 2 shows the comparison results. Our method can outperform all other methods. Similarly, trained on one labeled client, FedAVG [19] w/ residual weight connection surpasses the one w/o residual weight connection, meaning that our residual weight connection is also effective on ViT, not only CNNs, as shown in Fig. 2 (b)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.811, + 0.47, + 0.901 + ], + "angle": 0, + "content": "Two Labeled Clients. To better understand the effectiveness of our CBAFed and study how to use residual weight connection when the number of labeled clients is more than one, following [14, 18], we conduct experiments on CIFAR-10 by dividing the whole training data into 10 clients, with two labeled clients and eight unlabeled clients. Results are" + }, + { + "type": "table_caption", + "bbox": [ + 0.498, + 0.29, + 0.892, + 0.332 + ], + "angle": 0, + "content": "Table 2. Comparison of our method against RSCFed [14], FedConsist [31] and FedAVG [19] in SVHN dataset on ViT [5] as the backbone, with one labeled and nine unlabeled clients." + }, + { + "type": "table", + "bbox": [ + 0.523, + 0.343, + 0.871, + 0.452 + ], + "angle": 0, + "content": "
MethodClient Num.Accuracy
labeledunlabeled
FedAVG [19](upper bound)10096.81
FedAVG [19](lower bound)1081.68
FedAVG+ [19]1088.93
FedIRM [18]1979.44
Fed-Consist [31]1985.91
Fed-Consist+ [31]1993.21
RSCFed [14]1989.43
CBAFed(ours)1995.09
" + }, + { + "type": "table_caption", + "bbox": [ + 0.498, + 0.46, + 0.894, + 0.502 + ], + "angle": 0, + "content": "Table 3. Comparison of our method against RSCFed [14], FedConsist [31], FedIRM [18] and FedAVG [19] with the number of labeled and unlabeled client set to 2 and 8." + }, + { + "type": "table", + "bbox": [ + 0.505, + 0.514, + 0.892, + 0.635 + ], + "angle": 0, + "content": "
MethodClient Num.Accuracy
labeledunlabeled
FedAVG [19](upper bound)10080.89
FedAVG [19](lower bound)2061.85
FedAVG+ [19]2066.55
FedIRM [18]2862.62
Fed-Consist [31]2861.67
Fed-Consist+ [31]2868.04
RSCFed [14]2864.25
CBAFed(ours)2872.01
" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.657, + 0.892, + 0.762 + ], + "angle": 0, + "content": "reported in Table 3. Our method has a better performance compared with all state-of-the-art FSSL methods. Adding our residual weight connection and fixed pseudo labeling lead to much performance gain. In the supplementary material, we also discuss different training strategies used for labeled clients w.r.t. number of local epochs and the usage of residual weight connection." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.763, + 0.892, + 0.838 + ], + "angle": 0, + "content": "Communication Cost. Since the statistics of local training data can be nearly neglected (very low compared with model's parameter), our method does not add any burden in communication cost, which is only \\(66\\%\\) of the state-of-the-art method RSCFed [14]." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.848, + 0.88, + 0.865 + ], + "angle": 0, + "content": "4.3. Study of Pseudo Labeling Strategies in FSSL" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.871, + 0.892, + 0.901 + ], + "angle": 0, + "content": "In this section, we discuss how to perform pseudo labeling in FSSL and why traditional batch-based pseudo la" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16298" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.092, + 0.27, + 0.228 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.274, + 0.092, + 0.465, + 0.229 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.243, + 0.47, + 0.314 + ], + "angle": 0, + "content": "Figure 3. Left: Test accuracy of all 5 strategies after every communication round. Note that the test accuracy of communication round 0 is the test accuracy of model trained on labeled client. Right: Accuracy of pseudo labels in local training epoch of one randomly selected unlabeled client. Best viewed electronically." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.325, + 0.473, + 0.794 + ], + "angle": 0, + "content": "beling fails. We conduct a simple experiment on SVHN dataset [20] using ResNet18 [6]. The experimental setting is the same as illustrated in Sec. 4.1 except that we only use pseudo labeling without other proposed modules. We consider five training strategies: (1) Batched pseudo labeling (1 local training epoch). (2) Batched pseudo labeling (5 local training epochs). (3) Fixed pseudo labeling (1 local training epoch). (4) Fixed pseudo labeling (5 local training epochs), and we do not update the fixed training dataset in all local training epochs after pseudo labeling. (5) Fixed pseudo labeling (5 local training epochs), and we update the fixed training dataset at the beginning of every training epoch. For labeled client, the local training epoch is set to 1 for all 5 strategies. The total communication round is set to 200 for all strategies. The test accuracy after communication round is shown in the left figure of Fig. 3. We can observe the following: i), Batch-based pseudo labeling performs poorly in FSSL. It is known that CNNs often do not work well on out-of-distribution data [23]. Thus, a few update can result in catastrophic forgetting on Non-IID data partitions. The right figure of Fig. 3 shows the predicted accuracy of communication round one in five local training epochs of strategy (2) and (5). ii), Our fixed pseudo labeling method is capable of dealing with catastrophic forgetting. Strategy (3) performs the best and strategy (5) performs much worse than strategy (3) and (4), indicating that one local training epoch is optimum. iii) Comparing strategy (3) with (4), the former has better performance but the latter has faster convergence speed. It means more local training epochs will speed up convergence but may degrade performance, similar phenomenon is also found in [23]." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.802, + 0.233, + 0.818 + ], + "angle": 0, + "content": "4.4. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.826, + 0.47, + 0.901 + ], + "angle": 0, + "content": "We conduct ablation study to demonstrate the effectiveness of the main components of our method: residual weight connection, class balanced pseudo labeling, and tail class data discovery. Table 4 shows the results on three datasets: CIFAR-10/100 and Fashion-MNIST. It can be" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.089, + 0.892, + 0.132 + ], + "angle": 0, + "content": "Table 4. Ablation Study of CBAFed in CIFAR-10/100 and Fashion MNIST Datasets. Fixed PL: fixed pseudo labeling, CBA: class balanced adaptive pseudo labeling, DD: tail class data discovery." + }, + { + "type": "table", + "bbox": [ + 0.522, + 0.142, + 0.872, + 0.257 + ], + "angle": 0, + "content": "
DatasetFixed PLCBADDRes-WeightAccuracy
CIFAR-10✓59.16
✓✓64.29
✓✓✓65.15
✓✓✓✓67.08
CIFAR-100✓27.64
✓✓29.41
✓✓✓29.86
✓✓✓✓30.18
Fashion-MNIST✓79.99
✓✓80.87
✓✓✓84.37
✓✓✓✓85.49
" + }, + { + "type": "image", + "bbox": [ + 0.522, + 0.267, + 0.872, + 0.363 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.377, + 0.892, + 0.42 + ], + "angle": 0, + "content": "Figure 4. Performance changes on Fashion-MNIST by varying (a) threshold base \\(\\tau\\) and (b) upper bound threshold \\(\\tau_{h}\\), and (c) parameter for selecting tail class data \\(\\beta\\)." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.44, + 0.892, + 0.531 + ], + "angle": 0, + "content": "seen that all three modules are important for boosting performance. Then, we aim at discussing hyper-parameters of threshold base \\(\\tau\\), upper bound threshold \\(\\tau_h\\), and parameters for selecting tail class data \\(\\beta\\). Experiments are conducted on Fashion-MNIST. As shown in Fig. 4, performances are not sensitive within certain ranges." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.546, + 0.62, + 0.563 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.572, + 0.892, + 0.722 + ], + "angle": 0, + "content": "We present Class Balanced Adaptive pseudo labeling (CBAFed) for Federated Semi-Supervised Learning (FSSL). In CBAFed, a fixed pseudo labeling strategy is proposed to handle the catastrophic forgetting problem. To deal with the Non-IID setting of FSSL, we propose a class balanced adaptive thresholds selection method to choose better pseudo labels. Furthermore, a residual weight connection method is designed to make the model reach better optimum. We evaluate CBAFed on five datasets, whose performances show the superiority of our method." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.724, + 0.892, + 0.784 + ], + "angle": 0, + "content": "Limitations. Like other FSSL methods, a good model in warm-up cannot be guaranteed if the number of data in labeled clients are extremely few, so the final global model may not perform well." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.8, + 0.681, + 0.817 + ], + "angle": 0, + "content": "6. Acknowledgement" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.826, + 0.892, + 0.901 + ], + "angle": 0, + "content": "This work was supported by the National Natural Science Foundation of China (Grant No. 62101191), Shanghai Natural Science Foundation (Grant No. 21ZR1420800), and the Science and Technology Commission of Shanghai Municipality (Grant No. 22DZ2229004)." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "16299" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.079, + 0.09, + 0.176, + 0.107 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.184 + ], + "angle": 0, + "content": "[1] David Berthelot, Nicholas Carlini, Ekin D Cubuk, Alex Kurakin, Kihyuk Sohn, Han Zhang, and Colin Raffel. Remixmatch: Semi-supervised learning with distribution alignment and augmentation anchoring. arXiv preprint arXiv:1911.09785, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.185, + 0.472, + 0.242 + ], + "angle": 0, + "content": "[2] David Berthelot, Nicholas Carlini, Ian Goodfellow, Nicolas Papernot, Avital Oliver, and Colin A Raffel. Mixmatch: A holistic approach to semi-supervised learning. Advances in neural information processing systems, 32, 2019. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.242, + 0.472, + 0.311 + ], + "angle": 0, + "content": "[3] Xiaokang Chen, Yuhui Yuan, Gang Zeng, and Jingdong Wang. Semi-supervised semantic segmentation with cross pseudo supervision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2613-2622, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.311, + 0.472, + 0.395 + ], + "angle": 0, + "content": "[4] Wenhui Cui, Yanlin Liu, Yuxing Li, Menghao Guo, Yiming Li, Xiuli Li, Tianle Wang, Xiangzhu Zeng, and Chuyang Ye. Semi-supervised brain lesion segmentation with an adapted mean teacher model. In International Conference on Information Processing in Medical Imaging, pages 554-565. Springer, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.396, + 0.472, + 0.478 + ], + "angle": 0, + "content": "[5] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.48, + 0.472, + 0.535 + ], + "angle": 0, + "content": "[6] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.536, + 0.472, + 0.592 + ], + "angle": 0, + "content": "[7] Wenke Huang, Mang Ye, and Bo Du. Learn from others and be yourself in heterogeneous federated learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10143-10153, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.592, + 0.472, + 0.635 + ], + "angle": 0, + "content": "[8] Wonyong Jeong, Jaehong Yoon, Eunho Yang, and Sung Ju Hwang. Federated semi-supervised learning with inter-client consistency & disjoint learning. In Proc. ICLR, 2021. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.634, + 0.472, + 0.704 + ], + "angle": 0, + "content": "[9] Sai Praneeth Karimireddy, Satyen Kale, Mehryar Mohri, Sashank Reddi, Sebastian Stich, and Ananda Theertha Suresh. Scaffold: Stochastic controlled averaging for federated learning. In International Conference on Machine Learning, pages 5132-5143. PMLR, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.078, + 0.704, + 0.472, + 0.761 + ], + "angle": 0, + "content": "[10] Dong-Hyun Lee et al. Pseudo-label: The simple and efficient semi-supervised learning method for deep neural networks. In Workshop on challenges in representation learning, ICML, volume 3, page 896, 2013. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.078, + 0.761, + 0.472, + 0.817 + ], + "angle": 0, + "content": "[11] Qinbin Li, Bingsheng He, and Dawn Song. Model-contrastive federated learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10713–10722, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.078, + 0.818, + 0.472, + 0.873 + ], + "angle": 0, + "content": "[12] Tian Li, Anit Kumar Sahu, Manzil Zaheer, Maziar Sanjabi, Ameet Talwalkar, and Virginia Smith. Federated optimization in heterogeneous networks. Proceedings of Machine Learning and Systems, 2:429-450, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.078, + 0.873, + 0.472, + 0.903 + ], + "angle": 0, + "content": "[13] Xiaomeng Li, Lequan Yu, Hao Chen, Chi-Wing Fu, and Pheng-Ann Heng. Transformation consistent self" + }, + { + "type": "list", + "bbox": [ + 0.078, + 0.115, + 0.472, + 0.903 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.134 + ], + "angle": 0, + "content": "ensembling model for semi-supervised medical image segmentation. IEEE Trans. Neural Networks and Learning Systems, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.136, + 0.892, + 0.178 + ], + "angle": 0, + "content": "[14] Xiaoxiao Liang, Yiqun Lin, Huazhu Fu, Lei Zhu, and Xiaomeng Li. Rscfed: Random sampling consensus federated semi-supervised learning. In Proc. CVPR, 2022. 1, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.179, + 0.892, + 0.234 + ], + "angle": 0, + "content": "[15] Haowen Lin, Jian Lou, Li Xiong, and Cyrus Shahabi. Semifed: Semi-supervised federated learning with consistency and pseudo-labeling. arXiv preprint arXiv:2108.09412, 2021. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.235, + 0.892, + 0.319 + ], + "angle": 0, + "content": "[16] Fengbei Liu, Yu Tian, Yuanhong Chen, Yuyuan Liu, Vasileios Belagiannis, and Gustavo Carneiro. Acpl: Anticurriculum pseudo-labelling for semi-supervised medical image classification. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20697-20706, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.32, + 0.892, + 0.403 + ], + "angle": 0, + "content": "[17] Quande Liu, Cheng Chen, Jing Qin, Qi Dou, and Pheng-Ann Heng. Feddg: Federated domain generalization on medical image segmentation via episodic learning in continuous frequency space. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1013–1023, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.405, + 0.892, + 0.461 + ], + "angle": 0, + "content": "[18] Quande Liu, Hongzheng Yang, Qi Dou, and Pheng-Ann Heng. Federated semi-supervised medical image classification via inter-client relation matching. In Proc. MICCAI, 2021. 1, 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.461, + 0.892, + 0.532 + ], + "angle": 0, + "content": "[19] Brendan McMahan, Eider Moore, Daniel Ramage, Seth Hampson, and Blaise Aguera y Arcas. Communication-efficient learning of deep networks from decentralized data. In Artificial intelligence and statistics, pages 1273–1282. PMLR, 2017. 2, 4, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.532, + 0.892, + 0.575 + ], + "angle": 0, + "content": "[20] Yuval Netzer, Tao Wang, Adam Coates, Alessandro Bissacco, Bo Wu, and Andrew Y Ng. Reading digits in natural images with unsupervised feature learning. 2011. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.576, + 0.892, + 0.645 + ], + "angle": 0, + "content": "[21] Krishna Pillutla, Kshitiz Malik, Abdel-Rahman Mohamed, Mike Rabbat, Maziar Sanjabi, and Lin Xiao. Federated learning with partial model personalization. In International Conference on Machine Learning, pages 17716-17758. PMLR, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.646, + 0.892, + 0.689 + ], + "angle": 0, + "content": "[22] Boris T Polyak and Anatoli B Juditsky. Acceleration of stochastic approximation by averaging. SIAM journal on control and optimization, 30(4):838-855, 1992. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.69, + 0.892, + 0.773 + ], + "angle": 0, + "content": "[23] Liangqiong Qu, Yuyin Zhou, Paul Pu Liang, Yingda Xia, Feifei Wang, Ehsan Adeli, Li Fei-Fei, and Daniel Rubin. Rethinking architecture design for tackling data heterogeneity in federated learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10061–10071, 2022. 2, 4, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.774, + 0.892, + 0.845 + ], + "angle": 0, + "content": "[24] Yinghuan Shi, Jian Zhang, Tong Ling, Jiwen Lu, Yefeng Zheng, Qian Yu, Lei Qi, and Yang Gao. Inconsistency-aware uncertainty estimation for semi-supervised medical image segmentation. IEEE Trans. Medical Imaging, 41(3):608-620, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.845, + 0.892, + 0.903 + ], + "angle": 0, + "content": "[25] Kihyuk Sohn, David Berthelot, Nicholas Carlini, Zizhao Zhang, Han Zhang, Colin A Raffel, Ekin Dogus Cubuk, Alexey Kurakin, and Chun-Liang Li. Fixmatch: Simplifying semi-supervised learning with consistency and confidence." + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.903 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "16300" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.11, + 0.093, + 0.469, + 0.119 + ], + "angle": 0, + "content": "Advances in neural information processing systems, 33:596-608, 2020. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.122, + 0.47, + 0.177 + ], + "angle": 0, + "content": "[26] Antti Tarvainen and Harri Valpola. Mean teachers are better role models: Weight-averaged consistency targets improve semi-supervised deep learning results. Advances in neural information processing systems, 30, 2017. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.179, + 0.469, + 0.233 + ], + "angle": 0, + "content": "[27] Hongyi Wang, Mikhail Yurochkin, Yuekai Sun, Dimitris Papailiopoulos, and Yasaman Khazaeni. Federated learning with matched averaging. arXiv preprint arXiv:2002.06440, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.235, + 0.469, + 0.302 + ], + "angle": 0, + "content": "[28] Jianyu Wang, Qinghua Liu, Hao Liang, Gauri Joshi, and H Vincent Poor. Tackling the objective inconsistency problem in heterogeneous federated optimization. Advances in neural information processing systems, 33:7611-7623, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.305, + 0.469, + 0.359 + ], + "angle": 0, + "content": "[29] Yidong Wang, Bowen Zhang, Wenxin Hou, Zhen Wu, Jindong Wang, and Takahiro Shinozaki. Margin calibration for long-tailed visual recognition. arXiv preprint arXiv:2112.07225, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.362, + 0.469, + 0.416 + ], + "angle": 0, + "content": "[30] Qizhe Xie, Zihang Dai, Eduard Hovy, Thang Luong, and Quoc Le. Unsupervised data augmentation for consistency training. Advances in Neural Information Processing Systems, 33:6256-6268, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.419, + 0.469, + 0.501 + ], + "angle": 0, + "content": "[31] Dong Yang, Ziyue Xu, Wenqi Li, Andriy Myronenko, Holger R Roth, Stephanie Harmon, Sheng Xu, Baris Turkbey, Evrim Turkbey, Xiaosong Wang, et al. Federated semi-supervised learning for pandemic region segmentation in chest CT using multi-national data from china, Italy, japan. Medical image analysis, 70:101992, 2021. 1, 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.504, + 0.469, + 0.557 + ], + "angle": 0, + "content": "[32] Qiang Yang, Yang Liu, Tianjian Chen, and Yongxin Tong. Federated machine learning: Concept and applications. ACM Transactions on Intelligent Systems and Technology (TIST), 10(2):1-19, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.56, + 0.469, + 0.613 + ], + "angle": 0, + "content": "[33] Lequan Yu, Shujun Wang, Xiaomeng Li, Chi-Wing Fu, and Pheng-Ann Heng. Uncertainty-aware self-ensembling model for semi-supervised 3d left atrium segmentation. In Proc. MICCAI, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.616, + 0.469, + 0.685 + ], + "angle": 0, + "content": "[34] Bowen Zhang, Yidong Wang, Wenxin Hou, Hao Wu, Jindong Wang, Manabu Okumura, and Takahiro Shinozaki. Flexmatch: Boosting semi-supervised learning with curriculum pseudo labeling. Advances in Neural Information Processing Systems, 34:18408-18419, 2021. 2, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.687, + 0.469, + 0.755 + ], + "angle": 0, + "content": "[35] Songyang Zhang, Zeming Li, Shipeng Yan, Xuming He, and Jian Sun. Distribution alignment: A unified framework for long-tail visual recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2361-2370, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.757, + 0.469, + 0.826 + ], + "angle": 0, + "content": "[36] Boyan Zhou, Quan Cui, Xiu-Shen Wei, and Zhao-Min Chen. Bbn: Bilateral-branch network with cumulative learning for long-tailed visual recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9719–9728, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.828, + 0.469, + 0.882 + ], + "angle": 0, + "content": "[37] Yuyin Zhou, Yan Wang, Peng Tang, Song Bai, Wei Shen, Elliot K. Fishman, and Alan L. Yuille. Semi-supervised 3d abdominal multi-organ segmentation via deep multi-planar co-training. In Proc. WACV, 2019. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.093, + 0.47, + 0.882 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.956 + ], + "angle": 0, + "content": "16301" + } + ] +] \ No newline at end of file diff --git a/2023/Class Balanced Adaptive Pseudo Labeling for Federated Semi-Supervised Learning/27b4466b-7a94-415b-9a3c-c8f4f39a02f2_origin.pdf b/2023/Class Balanced Adaptive Pseudo Labeling for Federated Semi-Supervised Learning/27b4466b-7a94-415b-9a3c-c8f4f39a02f2_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..13300bc6b4a889c4d0a7338e18883ded67b22e34 --- /dev/null +++ b/2023/Class Balanced Adaptive Pseudo Labeling for Federated Semi-Supervised Learning/27b4466b-7a94-415b-9a3c-c8f4f39a02f2_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:577d7701874cd5c33afdefd247d06714f1132a8924a908accddf603899b25c72 +size 1283548 diff --git a/2023/Class Balanced Adaptive Pseudo Labeling for Federated Semi-Supervised Learning/full.md b/2023/Class Balanced Adaptive Pseudo Labeling for Federated Semi-Supervised Learning/full.md new file mode 100644 index 0000000000000000000000000000000000000000..efe94e9c4ce3290910ad189a73e39bb50b233c44 --- /dev/null +++ b/2023/Class Balanced Adaptive Pseudo Labeling for Federated Semi-Supervised Learning/full.md @@ -0,0 +1,356 @@ +# Class Balanced Adaptive Pseudo Labeling for Federated Semi-Supervised Learning + +Ming Li Qingli Li Yan Wang* + +Shanghai Key Laboratory of Multidimensional Information Processing, East China Normal University + +lm1640362161@gmail.com, qlli@cs.ecnu.edu.cn, ywang@cee.ecnu.edu.cn + +# Abstract + +This paper focuses on federated semi-supervised learning (FSSL), assuming that few clients have fully labeled data (labeled clients) and the training datasets in other clients are fully unlabeled (unlabeled clients). Existing methods attempt to deal with the challenges caused by not independent and identically distributed data (Non-IID) setting. Though methods such as sub-consensus models have been proposed, they usually adopt standard pseudo labeling or consistency regularization on unlabeled clients which can be easily influenced by imbalanced class distribution. Thus, problems in FSSL are still yet to be solved. To seek for a fundamental solution to this problem, we present Class Balanced Adaptive Pseudo Labeling (CBAFed), to study FSSL from the perspective of pseudo labeling. In CBAFed, the first key element is a fixed pseudo labeling strategy to handle the catastrophic forgetting problem, where we keep a fixed set by letting pass information of unlabeled data at the beginning of the unlabeled client training in each communication round. The second key element is that we design class balanced adaptive thresholds via considering the empirical distribution of all training data in local clients, to encourage a balanced training process. To make the model reach a better optimum, we further propose a residual weight connection in local supervised training and global model aggregation. Extensive experiments on five datasets demonstrate the superiority of CBAFed. Code will be available at https://github.com/minglllli/CBAFed. + +# 1. Introduction + +Federated learning (FL) aims to train machine learning models on a decentralized manner while preserving data privacy, i.e., separate local models are trained on separate local training datasets independently. In recent years, FL has received much attention for privacy protection reasons + +[32]. However, most of FL works focused on supervised learning with fully labeled data. But in practice, labeling process of large-scale training data is laborious and expensive. Due to the lack of funds or experts, large labeled training dataset is difficult for many companies and institutions to obtain. These may hinder applicability of FL. + +To handle this problem, federated semi-supervised learning (FSSL) has been explored by many researchers recently [8, 14, 15]. There are broadly three lines of FSSL methods by considering different places and status of labeled data. The first two lines consider that there are only limited labeled data in the central server [8] or each client has partially labeled data [15]. The third line assumes that few clients have fully labeled data and the training datasets in other clients are fully unlabeled [14, 18, 31]. Our paper mainly focuses on the third line of FSSL, and we call local clients with fully labeled data as labeled clients and the other clients as unlabeled clients. + +The main difficulties to train a third line FSSL model lie in three folds: 1) There are no labeled data in unlabeled clients. Thus, the training can be easily biased without label guidance. 2) Due to the divergent class distribution of labeled and unlabeled clients, namely the not independent and identically distributed data (Non-IID) setting, inaccurate supervisory signals may be generated in unlabeled clients via employing the model trained in labeled clients by either pseudo labeling or consistency regularization framework. 3) Due to the catastrophic forgetting problems in CNNs, with the training process of unlabeled clients going on, models may forget the knowledge learned on labeled clients and so decrease the prediction accuracy drastically. + +RSCFed [14], the state-of-the-art method significantly boosts the FSSL performance by first distilling sub-consensus models, and then aggregating the sub-consensus models to the global model. The sub-consensus models can handle the Non-IID setting to some extent, but the meanteacher based consistency regularization framework in unlabeled clients inevitably causes the accuracy degradation when the classes are imbalanced distributed. RSCFed at + +tempt to address the catastrophic forgetting problem by adjusting aggregation weights for labeled and unlabeled clients. But, it seems just alleviates the negative impact on the unlabeled clients and the problem is still yet to be tackled. Besides, the sub-consensus models may occur unstable training problems, and increase the communication cost. Other methods like FedConsist [31] and FedIRM [18] achieve promising results by utilizing pseudo labeling methods to produce artificial labels, but they do not consider the Non-IID setting among local clients. + +Pseudo labeling methods are shown to be effective in semi-supervised learning (SSL) [13,24,33,37], which generate pseudo-labels for unlabeled images and propose several strategies to ensure the quality of pseudo-labels. Existing FSSL works only adopt standard pseudo labeling or consistency regularization on FSSL. Since the mentioned difficulties hamper the usage of these methods, other remedy is usually proposed to alleviate the difficulties, which is palliative. Thus, we propose to study FSSL from the perspective of pseudo labeling, seeking for a fundamental solution to this problem. + +Concretely, we present Class Balanced Adaptive Pseudo Labeling, namely CBAFed, by rethinking standard pseudo labeling methods in SSL. To handle the catastrophic forgetting problem, we propose a fixed pseudo labeling strategy, which builds a fixed set by letting pass informative unlabeled data and their pseudo labels at the beginning of the unlabeled client training. Due to the Non-IID and heterogeneous data partition problems in FL, training distribution of unlabeled data can be highly imbalanced, so existing thresholds are not suitable in FSSL. We design class balanced adaptive thresholds via considering the empirical distribution of all training data in local clients at the previous communication round. Analysis proves that our method sets a reasonably high threshold for extremely scarce classes and encourages a balanced training process. To enhance the learning ability and discover unlabeled data from tail classes, we propose to leverage information from so-called "not informative" unlabeled data. Besides, we also explore a novel training strategy for labeled clients and the central server, termed as residual weight connection, skip connecting weights from previous epochs (for labeled clients) or previous global models (for the central server). It can help the model reach better optimum, when the training distribution is imbalanced and training amount is small. We conduct extensive experiments on five datasets to show the effectiveness of CBAFed. Overall, our main contributions can be summarized as follows: + +- We present a CBAFed method to deal with the catastrophic forgetting problems in federated learning. Unlike existing FSSL frameworks that directly adopts pseudo labeling or consistency regularization methods, CBAFed explores a fundamental solution to FSSL via + +designing a novel pseudo labeling method. + +- We introduce a residual weight connection method, to improve the robustness of the models in labeled clients and the central server, which skip connects weights from previous epoch or communication round to finally reach better optimum. +- Experiments are conducted on five datasets: four natural datasets CIFAR-10/100, SVHN, fashion MNIST and one medical dataset ISIC 2018 Skin. CBAFed outperforms state-of-the-art FSSL methods by a large margin, with $11.3\%$ improvements on SVHN dataset. + +# 2. Related work + +# 2.1. Federated Learning + +Recently, federated learning is becoming more and more popular for privacy protection reasons. Statistical data heterogeneity is one of the most important research problems. FedAvg [19], one of the pioneer works to deal with this problem, performs weighted-averaging on local weights according to the local training size, has been used as the most widely recognized FL baseline. To deal with data heterogeneity, two research directions have been mostly explored: model aggregation [12, 27, 28] and local training [9, 11, 12, 17]. Recently, [23] demonstrates that self-attention-based architectures, e.g., vision transformers(ViT) are more robust to distribution shifts and can converge to better optimum over heterogeneous data. + +# 2.2. Semi-supervised Leaning + +The goal of standard semi-supervised learning (SSL) is to train a good model with limited labeled data and much more unlabeled data. Popular research directions involves consistency regularization [4, 26] and pseudo labeling [3, 10, 16]. Many recent works use the combinations of consistency regularization and pseudo labeling and achieves state-of-the-art performance [1, 2, 25, 30]. In this paper, we focus on using pseudo labeling method to deal with SSL. While these methods achieve much progress, they only consider pre-defined fixed threshold for pseudo labeling. FlexMatch [34] demonstrates that leveraging unlabeled data according to the model's learning status to flexibly adjust thresholds for different classes can boost performance. While these methods perform well in centralized SSL, they all update pseudo labels after every batch's update of the model, which is not suitable in FSSL as shown in later section. + +# 2.3. Federated Semi-supervised Learning + +Recently, with the development of federated learning, FSSL becomes popular. There are broadly three categories + +![](images/2a2c958fb80ec8b46cc44c4281bb2b9c801e7cbfc89bebfd2e65e7f2a988a7bd.jpg) +Figure 1. An overview of our CBAFed. In the central server (left side), the global model is aggregated with the returned local models (step ④) and the adaptive thresholds are calculated by the returned training data statistics (step ⑤). Then central server passes the global model, adaptive thresholds and class distribution to all local clients (step ①). After downloading these data, local clients perform local training on the right side (step ②). Labeled clients use labeled data to train the model with residual weight connection. Unlabeled clients obtain the new training dataset by adaptive pseudo labeling and tail class data discovery and use it to train the model. After local training, local clients return trained models and number of data in each class back to central server (step ③). + +![](images/e91d5849fb9ef1c8e6afb98aeea5656fcef12b04af4e280f30f12c879f9f5576.jpg) + +of FSSL. The first two categories assume that every local client contains partially labeled data or there exists some labeled data in central server [8, 15]. The third category considers that few clients are fully labeled while other clients contain only unlabeled data. This paper focuses on the third category. To deal with this problem, Fed-Consist [31] and FedIRM [18] have been proposed. But, these two methods do not consider data heterogeneity in federated learning. Recently, RSCFed [14] proposes to perform random sub-sampling to reach consensus over clients. However, it uses standard consistency regularization for unlabeled data, which still suffers from the Non-IID setting. We emphasize that all the aforementioned approaches just use traditional pseudo labeling strategy, and none of them discuss how to apply pseudo labeling to FSSL. + +# 3. Methodology + +We first introduce the problem setting of federated semi-supervised learning. We consider decentralized training with labeled clients and unlabeled clients. Let $m$ (usually $m = 1$ ) labeled clients be $\{C_1, C_2, \dots, C_m\}$ and $n$ (usually $n > 1$ ) unlabeled clients be $\{C_{m+1}, C_{m+2}, \dots, C_{m+n}\}$ . For labeled client $C_\ell$ , the labeled local training dataset is denoted as $D_\ell = \{(X_i^\ell, y_i^\ell)\}_{i=1}^{N_\ell}$ , $\ell = 1, 2, \dots, m$ , where $N_\ell$ denotes the data number in $D_\ell$ . Similarly, for unlabeled client $C_\mu$ , the unlabeled local training dataset is denoted as $D_\mu = \{X_i^\mu\}_{i=1}^{N_\mu}$ , $\mu = m + 1, m + 2, \dots, n + m$ . + +The overall pipeline of the proposed CBAFed is shown in Fig. 1, which performs the following steps: (1) Warm up stage: train fully supervised models on only labeled clients using residual weight connection in a normal federated + +learning manner (Sec. 3.1) (2) The central server computes the empirical class distribution and obtains the class balanced adaptive thresholds, then passes them to local clients (Sec. 3.3); (3) All local clients update local models, adaptive threshold and class distribution. Labeled clients train local models on all the data using proposed residual weight connection; Unlabeled clients acquire the fixed training set (Sec. 3.2.2) by the threshold and the tail class datasets (Sec. 3.4), and train local models on the newly obtained training dataset. After local training, all local clients pass the trained model, the number of data in each class and the amount of training data to the central server; (4) The central server aggregates a new model with residual weight connection, computes the class distribution, and obtains the class balanced adaptive threshold. Then, the central server passes them to local clients; (5) Repeat step (3)-(4) until the specified number of communication round is reached. The overall training procedure of our CBAFed is provided in Algorithm 1 in the supplementary material. + +# 3.1. Residual Weight Connection + +We propose a novel training method in local supervised training and global model aggregation called residual weight connection. Our idea of res-weight is simple and similar to ResNet. In ResNet, there is a skip connection between every layer. For our res-weight connection, there is a skip connection of model's parameters between training epochs (or communication rounds). Let $\theta^{E}$ be the parameters of model after the training epoch (communication round) $E(E > 0)$ , then the final parameters of model after this training epoch by using residual weight connection will + +be + +$$ +\theta^ {E} = \left\{ \begin{array}{l l} \theta^ {E} & E \% s \neq 0 \\ \alpha_ {1} \theta^ {E - s} + (1 - \alpha_ {1}) \theta^ {E} & E \% s = 0 \end{array} \right. \tag{1} +$$ + +where $\alpha_{1}$ is the parameter to scale the size of $\theta^{E - s}$ and $\theta^{E}$ and $s$ is the number of skip epoch. In [22], authors show that averaging model weights over training steps tends to produce a more accurate model than using the final weights directly. In our experiments, we will show that this training method can effectively make the model more robust and finally reach better optimum, enhancing the predicted accuracy of pseudo labeling. Note that although the formula of our residual weight connection is similar to mean teachers [26], the update strategy and usage scenario is very different from it. Please refer to supplementary material for more detailed comparison and discussions. + +# 3.2. Pseudo Labeling Methods + +# 3.2.1 Review of Classic Batch-based Pseudo Labeling + +In semi-supervised learning, for unlabeled data, traditional approaches [2, 25, 34], such as Fixmatch [25] and Flexmatch [34], use original data or their weak augmented version in one batch to generate pseudo labels. These labels are adopted to supervise model's training. For detailed review, please refer to supplementary materials. + +# 3.2.2 Fixed Pseudo Labeling in FSSL + +Warm up stage. To conduct pseudo labeling for unlabeled data, in the warm up stage, we train our model only with labeled data. In the first $P$ communication round, training only conducted in labeled clients. In local training on labeled client $C_{\ell}$ , the training loss is defined as: + +$$ +\mathcal {L} _ {\ell} = \frac {1}{| D _ {\ell} |} \sum_ {(X ^ {\ell}, y ^ {\ell}) \in D _ {\ell}} H \left(y ^ {\ell}, p _ {m} \left(y \mid \theta_ {t} ^ {\ell} \left(X ^ {\ell}\right)\right)\right), \tag {2} +$$ + +where $\theta_t^\ell$ means local model in labeled client $C_\ell$ at communication round $t$ . The local training epoch is set to $J(J > 1)$ with residual weight convection. Then we use FedAVG [19] with residual weight connection for model aggregation: + +$$ +\theta_ {t + 1} ^ {G} = \left\{ \begin{array}{l l} \sum_ {\ell = 1} ^ {m} \frac {| D _ {\ell} |}{\sum_ {i = 1} ^ {m} | D _ {i} |} \theta_ {t} ^ {\ell} & t \% s \neq 0 \\ \alpha_ {1} \theta_ {t + 1 - s} ^ {G} + (1 - \alpha_ {1}) \sum_ {\ell = 1} ^ {m} \frac {| D _ {\ell} |}{\sum_ {i = 1} ^ {m} | D _ {i} |} \theta_ {t} ^ {\ell} & t \% s = 0 \end{array} \right. \tag{3} +$$ + +where $\theta_{t + 1}^{G}$ is the global model of communication round $t+$ 1. Note that, usually, the number of labeled clients is only 1, so we can finish warm up stage in only one communication round(training can be completed on one labeled client). + +Different from traditional semi-supervised learning, in federated semi-supervised learning, Non-IID data partitions in clients can easily lead to catastrophic forgetting [7, 21, 23]. Thus, local models trained on unlabeled clients + +usually forget the former learned knowledge by batch-based pseudo labeling method, which will lead to abrupt degradation and sabotage the training of unlabeled clients. To handle heterogeneous data partitions and catastrophic forgetting, we propose a new pseudo labeling method called fixed pseudo labeling. For the unlabeled training dataset $D_{\mu}$ in an unlabeled client $C_{\mu}$ , after initializing the local model $\theta_t^\mu$ with the global model $\theta_t^G$ , we first compute the pseudo label for each data $X_i^\mu$ in $D_{\mu}$ via: + +$$ +\hat {y} _ {i} ^ {\mu} = \arg \max p _ {m} (y | \theta_ {t} ^ {\mu} \left(X _ {i} ^ {\mu}\right), i = 1, 2, \dots , N _ {\mu}. \tag {4} +$$ + +We obtain a subset of unlabeled data with their pseudo labels as supervisory signals during local client training: + +$$ +\widetilde {\mathcal {D}} _ {\mu} = \left\{\left(X _ {i} ^ {\mu}, \hat {y} _ {i} ^ {\mu}\right) \mid X _ {i} ^ {\mu} \in D _ {\mu} \wedge \max \left(p _ {m} \left(y \mid \theta_ {t} ^ {\mu} \left(X _ {i} ^ {\mu}\right)\right)\right) > \tau \right\} _ {i = 1} ^ {N _ {\mu}}. \tag {5} +$$ + +Then we use this fixed set $\widetilde{\mathcal{D}}_{\mu}$ as the training dataset and define the loss to train our unlabeled client as: $\mathcal{L}_{\mu} = \frac{1}{|\widetilde{\mathcal{D}}_{\mu}|}\sum_{(X^{\mu},\hat{y}^{\mu})\in \widetilde{\mathcal{D}}_{\mu}}H(\hat{y}^{\mu},p_m(y|\theta_t^\mu (X^\mu))).$ + +# 3.3. Class Balanced Adaptive Threshold for Pseudo Labeling + +In semi-supervised learning, the value of threshold $\tau$ is non-trivial for selecting confident pseudo labels, namely informative unlabeled data. Traditionally, a fixed threshold is pre-defined [25], but setting a fixed threshold usually makes the model fail to consider different learning status and learning difficulties of different classes. In one of the most recent works, Curriculum Pseudo Labeling [34] is proposed to achieve state-of-the-art results on most semi-supervised learning benchmarks, where a flexible threshold for class $c$ at time step $t$ is calculated by: $\mathcal{T}_t(c) = \beta_t(c)\cdot \tau$ , where $\beta_{t}(c)$ is the ratio of the number of selected pseudo labels in class $c$ and the maximum number of selected pseudo labels of all classes. $\beta_{t}(c)$ is used to scale the fixed threshold $\tau$ . Curriculum Pseudo Labeling significantly boosts the accuracy and convergence performance of various semi-supervised learning algorithms. But in FSSL, there exist heterogeneous data partition problems. Due to the Non-IID partition, the labeled data are not balanced, so purely using the number of selected unlabeled data to design threshold is improper. Moreover, the data of some classes can be extremely rare in certain clients, which will lead to a very low $\beta_{t}(c)$ , and thus, a very low threshold $(\mathcal{T}_t(c)\approx 0)$ . In other words, directly applying Curriculum Pseudo Labeling will introduce many noisy labels into training. + +To this end, we propose a novel Class Balanced Adaptive threshold for Pseudo Labeling (CBAPL). For an unlabeled client $C_{\mu}$ in the $t$ th communication round, let the number of selected pseudo labels in class $c$ be: + +$$ +\sigma_ {t} ^ {\mu} (c) = \sum_ {i = 1} ^ {N _ {\mu}} \mathbf {1} \left(\max \left(p _ {m} \left(y \mid \theta_ {t} ^ {\mu} \left(X _ {i} ^ {\mu}\right)\right)\right) > \mathcal {T} _ {t} (c)\right) \mathbf {1} \left(\hat {y} _ {i} ^ {\mu} = c\right), \tag {6} +$$ + +where $\mathbf{1}(\cdot)$ is the indicator function. The data number of class $c$ in labeled client $C_{\ell}$ is calculated by: $\sigma_t^\ell (c) = \sum_{i = 1}^{N_\ell}\mathbf{1}(y_i^\ell = c)$ . After local training, $\sigma_t^\mu (c),\sigma_t^\ell (c)$ and weights of local models are returned to central server. Then the central server calculates the total number of training data of class $c$ at communication round $t$ : + +$$ +\sigma_ {t} (c) = \sum_ {\ell = 1} ^ {m} \sigma_ {t} ^ {\ell} (c) + \sum_ {\mu = m + 1} ^ {n + m} \sigma_ {t} ^ {\mu} (c). \tag {7} +$$ + +We then normalize the training numbers and acquire the empirical distribution of training data: + +$$ +\tilde {p} _ {t} (c) = \frac {\sigma_ {t} (c)}{\sum_ {i = 1} ^ {C} \sigma_ {t} (i)}. \tag {8} +$$ + +The standard deviation of the empirical distribution is calculated by: + +$$ +\operatorname {s t d} \left(\widetilde {p} _ {t}\right) = \sqrt {\frac {1}{C - 1} \sum_ {c = 1} ^ {C} \left(\widetilde {p} _ {t} (c) - \bar {p} _ {t}\right) ^ {2}}, \tag {9} +$$ + +where $\overline{p}_t = \frac{1}{C}\sum_{c = 1}^{C}\widetilde{p}_t(c)$ . Lastly, the threshold of class $c$ is calculated by + +$$ +\tau_ {t, c} = \widetilde {p} _ {t} (c) + \tau - s t d (\widetilde {p} _ {t}), \tag {10} +$$ + +where $\tau$ is the pre-defined threshold base. We then set an upper bound of threshold as: + +$$ +\mathcal {T} _ {t + 1} (c) = \left\{ \begin{array}{l l} \tau_ {t, c}, & \tau_ {t, c} < \tau_ {h} \\ \tau_ {h}, & \tau_ {t, c} \geq \tau_ {h} \end{array} , \right. \tag {11} +$$ + +where $\tau_h$ is the pre-defined upper bound of threshold (usually $\tau_h = 0.95$ for ensuring a certain amount of training data in unlabeled clients). If our model is trained with more data in one class, the threshold of that class will be higher since $\widetilde{p}_t(c)$ is higher. $std(\widetilde{p}_t)$ is important for balancing the empirical distribution of training data. Please refer to the supplementary material for more discussions. + +We argue that for scarce classes, unlike Curriculum Pseudo Labeling [34] whose threshold is extremely low, our CBAPL calculates a threshold with a relatively high lower bound. We have theorem below: + +# Theorem 3.1. + +$$ +\tau + \widetilde {p} _ {t} (c) - \sqrt {\frac {1}{C}} \leq \mathcal {T} _ {t} (c) \leq \tau + \widetilde {p} _ {t} (c), \tag {12} +$$ + +Proof is provided in the supplementary material. + +Since $\tau >>\sqrt{\frac{1}{C}}$ $\mathcal{T}_t(c)$ will have a high lower bound $(>>0)$ . To balance the trade-off between $\widetilde{p}_t(t)$ and + +$\tau - \sqrt{1 / C}$ , a normalization factor is applied to consider the class number. Thus, we propose a modified $\widetilde{p}_t(c)$ as below: + +$$ +\widetilde {p} _ {t} (c) = \frac {\sigma_ {t} (c)}{\sum_ {i = 1} ^ {C} \sigma_ {t} (i)} \times \frac {C}{1 0}. \tag {13} +$$ + +At the first $P$ communication rounds, the models are trained on only labeled clients. Thus, when $t \in \{1, \dots, P\}$ , $\sigma_t^\mu(c) = 0$ for all classes. After calculating the thresholds for all classes via Eq. 11, the central server will pass the thresholds to local clients. Unlabeled clients further obtain the fixed pseudo label training dataset by: + +$$ +\begin{array}{l} \widetilde {\mathcal {D}} _ {t + 1, \mu} = \left\{\left(X _ {i} ^ {\mu}, \hat {y} _ {i} ^ {\mu}\right) \mid X _ {i} ^ {\mu} \in D _ {\mu} \right. \tag {14} \\ \wedge \max (p _ {m} (y | \theta_ {t} ^ {\mu} (X _ {i} ^ {\mu}))) > \mathcal {T} _ {t + 1} (\hat {y} _ {i} ^ {\mu}) \} _ {i = 1} ^ {N _ {\mu}}. \\ \end{array} +$$ + +# 3.4. Discovery of Unlabeled Data from Tail Classes + +In FSSL, heterogeneous data partitions will lead to class imbalanced distribution in local clients and there are rare or even no data from some classes in certain clients. For warm up stage in labeled clients, it is similar to long-tailed classification, so the problems in long-tailed classification will also exist in our pseudo labeling process, i.e., models tend to classify tail (rare) classes as head (common) classes [29, 35, 36]. Although adaptive pseudo labeling can lower the threshold of tail classes, much less data will be selected by PL since very few data from tailed classes will be classified as correct classes. To enhance the learning ability and discover unlabeled data from these classes, instead of directly dropping the data whose maximum confidences are low, namely, "not informative" unlabeled data, we propose to leverage information from them. For simplicity, we drop subscript $t$ and default all representations occur in $t$ th communication round in Sec. 3.4. Firstly, we define a mask function $\mathcal{M}(\cdot): \mathcal{R}^C \to \mathcal{R}^C$ for a probability simplex $p \in \mathcal{R}^C$ : + +$$ +\mathcal {M} _ {i} (p) = \left\{ \begin{array}{l l} p _ {i} & i \neq \arg \max p \\ 0 & i = \arg \max p \end{array} , \right. \tag {15} +$$ + +where $p_i$ is the $i$ th element in $p$ . The effect of mask function $\mathcal{M}(\cdot)$ is to set the maximum of a probability simplex to 0. Then, we turn to analyze the second largest confidence score via: + +$$ +\begin{array}{l} D _ {\mu} ^ {t a i l} = \left\{\left(X _ {i} ^ {\mu}, \hat {y} _ {i} ^ {\mu^ {\prime}}\right) \mid X _ {i} ^ {\mu} \in D _ {\mu} \right. \\ \left. \wedge \max \left(p _ {m} \left(y \mid \theta_ {t} ^ {\mu} \left(X _ {i} ^ {\mu}\right)\right)\right) \leq \mathcal {T} _ {t} \left(\hat {y} _ {i} ^ {\mu}\right) \wedge \widetilde {p} _ {t} \left(\hat {y} _ {i} ^ {\mu^ {\prime}}\right) < \frac {\beta}{C} \right\}, \tag {16} \\ \end{array} +$$ + +where $\hat{y}_i^{u'} = \arg \max \mathcal{M}(p_m(y|\theta_t^\mu(X_i^\mu)))$ and $\beta$ is a hyperparameter. In Eq. 15, the parameter $\beta$ is to find the tail classes ("not informative" unlabeled data). For unlabeled data $X_i^\mu$ , if its largest confidence is smaller than threshold + +(not informative or confident, i.e., $\max(p_m(y|\theta_t^\mu(X_i^\mu))) \leq \mathcal{T}_t(\hat{y}_i^\mu)$ ) and the class of second largest confidence is tail class (i.e., $\widetilde{p}_t(\hat{y}_i^{\mu \prime}) < \frac{\beta}{C}$ ), we regard it as misclassified data and include it in training with the class of second largest confidence as its label. Finally, the new training dataset for unlabeled client $C_\mu$ is defined as + +$$ +D _ {\mu} ^ {t r a i n} = D _ {\mu} ^ {t a i l} \cup \widetilde {D} _ {\mu}, \tag {17} +$$ + +and the final training loss in $C_{\mu}$ is rewritten as + +$$ +\mathcal {L} _ {\mu} = \frac {1}{| D _ {\mu} ^ {t r a i n} |} \sum_ {(X ^ {\mu}, \hat {y} ^ {\mu}) \in D _ {\mu} ^ {t r a i n}} H \left(\hat {y} ^ {\mu}, p _ {m} \left(y \mid \theta_ {t} ^ {\mu} \left(X ^ {\mu}\right)\right)\right). \tag {18} +$$ + +The calculation of $\sigma_t^u (c)$ is revised as: + +$$ +\sigma_ {t} ^ {\mu} (c) = \sum_ {(X ^ {\mu}, \hat {y} ^ {\mu}) \in D _ {\mu} ^ {t r a i n}} \mathbf {1} (\hat {y} ^ {\mu} = c). \tag {19} +$$ + +# 3.5. Aggregation of local models + +Since not all data in local clients are included into training, we propose to use only data included in training for model aggregation. The scale factor $w_{t}^{i}$ is computed by + +$$ +w _ {t} ^ {i} = \left\{ \begin{array}{l l} \frac {\left| D _ {i} \right|}{\left| D _ {t} ^ {t r a i n} \right|} & \text {i f} i \in \{1, \dots , m \} \\ \frac {\left| D _ {t , i} ^ {t r a i n} \right|}{\left| D _ {t} ^ {t r a i n} \right|} & \text {i f} i \in \{m + 1, \dots , m + n \}, \end{array} \right. \tag {20} +$$ + +where $|D_t^{train}| = \sum_{\ell=1}^m |D_\ell| + \sum_{\mu=m+1}^{m+n} |D_{t,\mu}^{train}|$ . Then at $(t+1)$ th synchronization round, the global model $\theta_{t+1}^G$ is computed by + +$$ +\theta_ {t + 1} ^ {G} = \left\{ \begin{array}{l l} \sum_ {i = 1} ^ {m + n} w _ {t} ^ {i} \theta_ {t} ^ {i} & t \% s \neq 0 \\ \alpha_ {2} \theta_ {t + 1 - s} ^ {G} + (1 - \alpha_ {2}) \sum_ {i = 1} ^ {m + n} w _ {t} ^ {i} \theta_ {t} ^ {i} & t \% s = 0, \end{array} \right. \tag{21} +$$ + +where $\alpha_{2}$ is a hyperparameter. + +# 4. Experiments + +# 4.1. Experimental Setup + +Benchmark Datasets To evaluate the effectiveness of our proposed method, we conduct extensive experiments on four image classification datasets, i.e., SVHN, CIFAR-10, CIFAR-100, Fashion MNIST and one medical image classification dataset: ISIC 2018 (Skin Lesion Analysis Towards Melanoma Detection). Dataset splitting and pre-processing are provided in the supplementary material. + +FSSL Setting Simulation Following [14], training datasets contain 10 clients: one labeled and nine unlabeled. We use a Dirichlet distribution $\text{Dir}(\gamma)$ , where $\gamma = 0.8$ for five datasets [14] to generate Non-IID data partition in clients. + +Implementation Details We utilize the SGD optimizer with a momentum of 0.9, and implement our method with + +![](images/6ba1612a951d4ca351d73178f6854ebf22e02b0228ccf57159f3b72aa4092cb1.jpg) +Figure 2. Test accuracy curves in local training of SVHN dataset w/ and w/o residual weight connection. ResNet18 (a) and ViT (b) are adopted as the backbones. W/ res-weight* indicates we only show test accuracy on epochs (communication rounds, since local training epoch for labeled client is 1) with skip weight connection. Best viewed electronically. + +![](images/115cb75ad70581d87bb29fd31b0f256a92835927fed7dd12377d5a66dc5f13ba.jpg) + +PyTorch. We adopt ResNet18 [6] from PyTorch for all datasets. For fair comparison, we use the same network architecture and training protocol, including the optimizer, data preprocessing, etc. across all FSSL methods. The local training epoch is set to 11 for labeled clients and 1 for unlabeled clients. More implementation details are provided in supplementary material. + +Comparisons We compare our method against state-of-the-art FSSL methods, including RSCFed [14], FedIRM [18], and Fed-Consist [31]. Since FedIRM [18] and Fed-Consist [31] are not desgined for dealing with Non-IID FSSL, following [14], we enlarge the weight of labeled client to about $50\%$ and other nine unlabeled clients share the remaining $50\%$ weight in each FSSL synchronization round, when implementing these methods. We also compare our network against FedAvg [19] trained with all clients as an upper bound and trained with only one labeled clients as the lower bound. Besides, to show the effectiveness of our residual weight connection, we report the result of FedAvg [19] trained with only one labeled clients using residual weight connection. Since Fed-Consist [31] utilizes traditional batch-based pseudo labeling method, we report the result of Fed-Consist [31] using our proposed fixed pseudo labeling without enlarging the weight of labeled client. + +# 4.2. Comparisons with State-of-the-arts + +Comparison Results. Table 1 reports the quantitative results of our CBAFed and other state-of-the-art methods on five benchmarks. We can observe that our proposed CBAFed achieves superior performances over all competitors on the five benchmark datasets. For some easy tasks (e.g., SVHN and Fashion-MNIST), the performance of our CBAFed is approaching upper bound obtained by FedAVG [19] and surpasses all state-of-the-arts by a large margin, showing a strong power of fixed pseudo labeling and class balanced threshold. Note that if replacing the traditional + +Table 1. Results on SVHN, CIFAR-10/100, Fashion MNIST and ISIC 2018 datasets under heterogeneous data partition with ResNet18. FedAVG $^+$ means FedAvg [19] trained with all one labeled clients using our residual weight connection. Fed-consist $^+$ means Fed-Consist [31] using our proposed fixed pseudo labeling without enlarging the weight of labeled client. + +
Labeling StrategyMethodClient Num.Dataset
labeledunlabeledSVHNCIFAR10CIFAR100Fashion-MNISTISIC 2018
Fully supervisedFedAvg [19](upper-bound)10091.8380.8951.3890.1481.32
FedAvg [19](lower-bound)1067.7154.6620.4974.8765.13
FedAvg+ [19]1076.9858.2124.8478.2666.69
Semi supervisedFedIRM [18]1969.2252.8420.2076.8364.85
Fed-Consist [31]1970.5654.2321.8176.5765.20
Fed-Consist+ [31]1986.5756.3523.2578.3565.50
RSCFed [14]1976.7457.0728.4678.4067.21
CBAFed(ours)1988.0767.0830.1885.4968.29
+ +batch-based pseudo labeling by our fixed pseudo labeling in Fed-Consist [31], the performance is largely increased. It shows that fixed pseudo labeling method is well suited for Non-IID FSSL. + +Trained on one labeled client, FedAVG [19] w/ residual weight connection achieves much better performance compared with that w/o residual weight connection. Fig. 2 (a) shows the test accuracy curves during training. Due to the imbalanced training data distribution in labeled client, the test accuracy is unstable during training for FedAVG (w/o res-weight in the figure). But, if training w/ our residual weight connection, the test accuracy curve is much more stable and the performance is also enhanced (w/ res-weight and w/ res-weight*). To sum up, residual weight connection can improve the robustness of the models when the training amount is small and the training distribution is imbalanced. Since long tailed vision recognition task is similar to local training in labeled client, it is a potential direction to use our residual weight connection in long tailed vision. Besides, we also conduct experiments on partially labeled clients, results are shown in the supplementary material. + +ViT Backbone. Vision Transformer(ViT) has shown to be more robust to heterogeneous data and distribution shifts, and has been demonstrated to converge faster with better optimum in Federated learning [23]. To further study the effect of our method, we conduct experiments on SVHN dataset using ViT-Tiny as backbone for all competitors. The implementation details are provided in the supplementary material. Table 2 shows the comparison results. Our method can outperform all other methods. Similarly, trained on one labeled client, FedAVG [19] w/ residual weight connection surpasses the one w/o residual weight connection, meaning that our residual weight connection is also effective on ViT, not only CNNs, as shown in Fig. 2 (b). + +Two Labeled Clients. To better understand the effectiveness of our CBAFed and study how to use residual weight connection when the number of labeled clients is more than one, following [14, 18], we conduct experiments on CIFAR-10 by dividing the whole training data into 10 clients, with two labeled clients and eight unlabeled clients. Results are + +Table 2. Comparison of our method against RSCFed [14], FedConsist [31] and FedAVG [19] in SVHN dataset on ViT [5] as the backbone, with one labeled and nine unlabeled clients. + +
MethodClient Num.Accuracy
labeledunlabeled
FedAVG [19](upper bound)10096.81
FedAVG [19](lower bound)1081.68
FedAVG+ [19]1088.93
FedIRM [18]1979.44
Fed-Consist [31]1985.91
Fed-Consist+ [31]1993.21
RSCFed [14]1989.43
CBAFed(ours)1995.09
+ +Table 3. Comparison of our method against RSCFed [14], FedConsist [31], FedIRM [18] and FedAVG [19] with the number of labeled and unlabeled client set to 2 and 8. + +
MethodClient Num.Accuracy
labeledunlabeled
FedAVG [19](upper bound)10080.89
FedAVG [19](lower bound)2061.85
FedAVG+ [19]2066.55
FedIRM [18]2862.62
Fed-Consist [31]2861.67
Fed-Consist+ [31]2868.04
RSCFed [14]2864.25
CBAFed(ours)2872.01
+ +reported in Table 3. Our method has a better performance compared with all state-of-the-art FSSL methods. Adding our residual weight connection and fixed pseudo labeling lead to much performance gain. In the supplementary material, we also discuss different training strategies used for labeled clients w.r.t. number of local epochs and the usage of residual weight connection. + +Communication Cost. Since the statistics of local training data can be nearly neglected (very low compared with model's parameter), our method does not add any burden in communication cost, which is only $66\%$ of the state-of-the-art method RSCFed [14]. + +# 4.3. Study of Pseudo Labeling Strategies in FSSL + +In this section, we discuss how to perform pseudo labeling in FSSL and why traditional batch-based pseudo la + +![](images/be35e86cc4f6b643dbdc12187d6d6b5889ddc0ba345357c0159fee5347d44c7d.jpg) +Figure 3. Left: Test accuracy of all 5 strategies after every communication round. Note that the test accuracy of communication round 0 is the test accuracy of model trained on labeled client. Right: Accuracy of pseudo labels in local training epoch of one randomly selected unlabeled client. Best viewed electronically. + +![](images/f9c462a8aac5664fcfbd97999ec2daa599ebc85feffbf2427b5a2650514a3d61.jpg) + +beling fails. We conduct a simple experiment on SVHN dataset [20] using ResNet18 [6]. The experimental setting is the same as illustrated in Sec. 4.1 except that we only use pseudo labeling without other proposed modules. We consider five training strategies: (1) Batched pseudo labeling (1 local training epoch). (2) Batched pseudo labeling (5 local training epochs). (3) Fixed pseudo labeling (1 local training epoch). (4) Fixed pseudo labeling (5 local training epochs), and we do not update the fixed training dataset in all local training epochs after pseudo labeling. (5) Fixed pseudo labeling (5 local training epochs), and we update the fixed training dataset at the beginning of every training epoch. For labeled client, the local training epoch is set to 1 for all 5 strategies. The total communication round is set to 200 for all strategies. The test accuracy after communication round is shown in the left figure of Fig. 3. We can observe the following: i), Batch-based pseudo labeling performs poorly in FSSL. It is known that CNNs often do not work well on out-of-distribution data [23]. Thus, a few update can result in catastrophic forgetting on Non-IID data partitions. The right figure of Fig. 3 shows the predicted accuracy of communication round one in five local training epochs of strategy (2) and (5). ii), Our fixed pseudo labeling method is capable of dealing with catastrophic forgetting. Strategy (3) performs the best and strategy (5) performs much worse than strategy (3) and (4), indicating that one local training epoch is optimum. iii) Comparing strategy (3) with (4), the former has better performance but the latter has faster convergence speed. It means more local training epochs will speed up convergence but may degrade performance, similar phenomenon is also found in [23]. + +# 4.4. Ablation Study + +We conduct ablation study to demonstrate the effectiveness of the main components of our method: residual weight connection, class balanced pseudo labeling, and tail class data discovery. Table 4 shows the results on three datasets: CIFAR-10/100 and Fashion-MNIST. It can be + +Table 4. Ablation Study of CBAFed in CIFAR-10/100 and Fashion MNIST Datasets. Fixed PL: fixed pseudo labeling, CBA: class balanced adaptive pseudo labeling, DD: tail class data discovery. + +
DatasetFixed PLCBADDRes-WeightAccuracy
CIFAR-10✓59.16
✓✓64.29
✓✓✓65.15
✓✓✓✓67.08
CIFAR-100✓27.64
✓✓29.41
✓✓✓29.86
✓✓✓✓30.18
Fashion-MNIST✓79.99
✓✓80.87
✓✓✓84.37
✓✓✓✓85.49
+ +![](images/5ea2888668a874d4a89787ac9194eb9374b2690d579f8ba6a281056f3067b40d.jpg) +Figure 4. Performance changes on Fashion-MNIST by varying (a) threshold base $\tau$ and (b) upper bound threshold $\tau_{h}$ , and (c) parameter for selecting tail class data $\beta$ . + +seen that all three modules are important for boosting performance. Then, we aim at discussing hyper-parameters of threshold base $\tau$ , upper bound threshold $\tau_h$ , and parameters for selecting tail class data $\beta$ . Experiments are conducted on Fashion-MNIST. As shown in Fig. 4, performances are not sensitive within certain ranges. + +# 5. Conclusion + +We present Class Balanced Adaptive pseudo labeling (CBAFed) for Federated Semi-Supervised Learning (FSSL). In CBAFed, a fixed pseudo labeling strategy is proposed to handle the catastrophic forgetting problem. To deal with the Non-IID setting of FSSL, we propose a class balanced adaptive thresholds selection method to choose better pseudo labels. Furthermore, a residual weight connection method is designed to make the model reach better optimum. We evaluate CBAFed on five datasets, whose performances show the superiority of our method. + +Limitations. Like other FSSL methods, a good model in warm-up cannot be guaranteed if the number of data in labeled clients are extremely few, so the final global model may not perform well. + +# 6. Acknowledgement + +This work was supported by the National Natural Science Foundation of China (Grant No. 62101191), Shanghai Natural Science Foundation (Grant No. 21ZR1420800), and the Science and Technology Commission of Shanghai Municipality (Grant No. 22DZ2229004). + +# References + +[1] David Berthelot, Nicholas Carlini, Ekin D Cubuk, Alex Kurakin, Kihyuk Sohn, Han Zhang, and Colin Raffel. Remixmatch: Semi-supervised learning with distribution alignment and augmentation anchoring. arXiv preprint arXiv:1911.09785, 2019. 2 +[2] David Berthelot, Nicholas Carlini, Ian Goodfellow, Nicolas Papernot, Avital Oliver, and Colin A Raffel. Mixmatch: A holistic approach to semi-supervised learning. Advances in neural information processing systems, 32, 2019. 2, 4 +[3] Xiaokang Chen, Yuhui Yuan, Gang Zeng, and Jingdong Wang. Semi-supervised semantic segmentation with cross pseudo supervision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2613-2622, 2021. 2 +[4] Wenhui Cui, Yanlin Liu, Yuxing Li, Menghao Guo, Yiming Li, Xiuli Li, Tianle Wang, Xiangzhu Zeng, and Chuyang Ye. Semi-supervised brain lesion segmentation with an adapted mean teacher model. In International Conference on Information Processing in Medical Imaging, pages 554-565. Springer, 2019. 2 +[5] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 7 +[6] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 6, 8 +[7] Wenke Huang, Mang Ye, and Bo Du. Learn from others and be yourself in heterogeneous federated learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10143-10153, 2022. 4 +[8] Wonyong Jeong, Jaehong Yoon, Eunho Yang, and Sung Ju Hwang. Federated semi-supervised learning with inter-client consistency & disjoint learning. In Proc. ICLR, 2021. 1, 3 +[9] Sai Praneeth Karimireddy, Satyen Kale, Mehryar Mohri, Sashank Reddi, Sebastian Stich, and Ananda Theertha Suresh. Scaffold: Stochastic controlled averaging for federated learning. In International Conference on Machine Learning, pages 5132-5143. PMLR, 2020. 2 +[10] Dong-Hyun Lee et al. Pseudo-label: The simple and efficient semi-supervised learning method for deep neural networks. In Workshop on challenges in representation learning, ICML, volume 3, page 896, 2013. 2 +[11] Qinbin Li, Bingsheng He, and Dawn Song. Model-contrastive federated learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10713–10722, 2021. 2 +[12] Tian Li, Anit Kumar Sahu, Manzil Zaheer, Maziar Sanjabi, Ameet Talwalkar, and Virginia Smith. Federated optimization in heterogeneous networks. Proceedings of Machine Learning and Systems, 2:429-450, 2020. 2 +[13] Xiaomeng Li, Lequan Yu, Hao Chen, Chi-Wing Fu, and Pheng-Ann Heng. Transformation consistent self + +ensembling model for semi-supervised medical image segmentation. IEEE Trans. Neural Networks and Learning Systems, 2020. 2 +[14] Xiaoxiao Liang, Yiqun Lin, Huazhu Fu, Lei Zhu, and Xiaomeng Li. Rscfed: Random sampling consensus federated semi-supervised learning. In Proc. CVPR, 2022. 1, 3, 6, 7 +[15] Haowen Lin, Jian Lou, Li Xiong, and Cyrus Shahabi. Semifed: Semi-supervised federated learning with consistency and pseudo-labeling. arXiv preprint arXiv:2108.09412, 2021. 1, 3 +[16] Fengbei Liu, Yu Tian, Yuanhong Chen, Yuyuan Liu, Vasileios Belagiannis, and Gustavo Carneiro. Acpl: Anticurriculum pseudo-labelling for semi-supervised medical image classification. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20697-20706, 2022. 2 +[17] Quande Liu, Cheng Chen, Jing Qin, Qi Dou, and Pheng-Ann Heng. Feddg: Federated domain generalization on medical image segmentation via episodic learning in continuous frequency space. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1013–1023, 2021. 2 +[18] Quande Liu, Hongzheng Yang, Qi Dou, and Pheng-Ann Heng. Federated semi-supervised medical image classification via inter-client relation matching. In Proc. MICCAI, 2021. 1, 2, 3, 6, 7 +[19] Brendan McMahan, Eider Moore, Daniel Ramage, Seth Hampson, and Blaise Aguera y Arcas. Communication-efficient learning of deep networks from decentralized data. In Artificial intelligence and statistics, pages 1273–1282. PMLR, 2017. 2, 4, 6, 7 +[20] Yuval Netzer, Tao Wang, Adam Coates, Alessandro Bissacco, Bo Wu, and Andrew Y Ng. Reading digits in natural images with unsupervised feature learning. 2011. 8 +[21] Krishna Pillutla, Kshitiz Malik, Abdel-Rahman Mohamed, Mike Rabbat, Maziar Sanjabi, and Lin Xiao. Federated learning with partial model personalization. In International Conference on Machine Learning, pages 17716-17758. PMLR, 2022. 4 +[22] Boris T Polyak and Anatoli B Juditsky. Acceleration of stochastic approximation by averaging. SIAM journal on control and optimization, 30(4):838-855, 1992. 4 +[23] Liangqiong Qu, Yuyin Zhou, Paul Pu Liang, Yingda Xia, Feifei Wang, Ehsan Adeli, Li Fei-Fei, and Daniel Rubin. Rethinking architecture design for tackling data heterogeneity in federated learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10061–10071, 2022. 2, 4, 7, 8 +[24] Yinghuan Shi, Jian Zhang, Tong Ling, Jiwen Lu, Yefeng Zheng, Qian Yu, Lei Qi, and Yang Gao. Inconsistency-aware uncertainty estimation for semi-supervised medical image segmentation. IEEE Trans. Medical Imaging, 41(3):608-620, 2022. 2 +[25] Kihyuk Sohn, David Berthelot, Nicholas Carlini, Zizhao Zhang, Han Zhang, Colin A Raffel, Ekin Dogus Cubuk, Alexey Kurakin, and Chun-Liang Li. Fixmatch: Simplifying semi-supervised learning with consistency and confidence. + +Advances in neural information processing systems, 33:596-608, 2020. 2, 4 +[26] Antti Tarvainen and Harri Valpola. Mean teachers are better role models: Weight-averaged consistency targets improve semi-supervised deep learning results. Advances in neural information processing systems, 30, 2017. 2, 4 +[27] Hongyi Wang, Mikhail Yurochkin, Yuekai Sun, Dimitris Papailiopoulos, and Yasaman Khazaeni. Federated learning with matched averaging. arXiv preprint arXiv:2002.06440, 2020. 2 +[28] Jianyu Wang, Qinghua Liu, Hao Liang, Gauri Joshi, and H Vincent Poor. Tackling the objective inconsistency problem in heterogeneous federated optimization. Advances in neural information processing systems, 33:7611-7623, 2020. 2 +[29] Yidong Wang, Bowen Zhang, Wenxin Hou, Zhen Wu, Jindong Wang, and Takahiro Shinozaki. Margin calibration for long-tailed visual recognition. arXiv preprint arXiv:2112.07225, 2021. 5 +[30] Qizhe Xie, Zihang Dai, Eduard Hovy, Thang Luong, and Quoc Le. Unsupervised data augmentation for consistency training. Advances in Neural Information Processing Systems, 33:6256-6268, 2020. 2 +[31] Dong Yang, Ziyue Xu, Wenqi Li, Andriy Myronenko, Holger R Roth, Stephanie Harmon, Sheng Xu, Baris Turkbey, Evrim Turkbey, Xiaosong Wang, et al. Federated semi-supervised learning for pandemic region segmentation in chest CT using multi-national data from china, Italy, japan. Medical image analysis, 70:101992, 2021. 1, 2, 3, 6, 7 +[32] Qiang Yang, Yang Liu, Tianjian Chen, and Yongxin Tong. Federated machine learning: Concept and applications. ACM Transactions on Intelligent Systems and Technology (TIST), 10(2):1-19, 2019. 1 +[33] Lequan Yu, Shujun Wang, Xiaomeng Li, Chi-Wing Fu, and Pheng-Ann Heng. Uncertainty-aware self-ensembling model for semi-supervised 3d left atrium segmentation. In Proc. MICCAI, 2019. 2 +[34] Bowen Zhang, Yidong Wang, Wenxin Hou, Hao Wu, Jindong Wang, Manabu Okumura, and Takahiro Shinozaki. Flexmatch: Boosting semi-supervised learning with curriculum pseudo labeling. Advances in Neural Information Processing Systems, 34:18408-18419, 2021. 2, 4, 5 +[35] Songyang Zhang, Zeming Li, Shipeng Yan, Xuming He, and Jian Sun. Distribution alignment: A unified framework for long-tail visual recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2361-2370, 2021. 5 +[36] Boyan Zhou, Quan Cui, Xiu-Shen Wei, and Zhao-Min Chen. Bbn: Bilateral-branch network with cumulative learning for long-tailed visual recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9719–9728, 2020. 5 +[37] Yuyin Zhou, Yan Wang, Peng Tang, Song Bai, Wei Shen, Elliot K. Fishman, and Alan L. Yuille. Semi-supervised 3d abdominal multi-organ segmentation via deep multi-planar co-training. In Proc. WACV, 2019. 2 \ No newline at end of file diff --git a/2023/Class Balanced Adaptive Pseudo Labeling for Federated Semi-Supervised Learning/images.zip b/2023/Class Balanced Adaptive Pseudo Labeling for Federated Semi-Supervised Learning/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..bb450ebd9328ec4de72ec6e6dad50c2786af9121 --- /dev/null +++ b/2023/Class Balanced Adaptive Pseudo Labeling for Federated Semi-Supervised Learning/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:67f8db0ddcd1220feafdbee03b8819edaccae96b021d1f059956aeaaa20d04ca +size 411146 diff --git a/2023/Class Balanced Adaptive Pseudo Labeling for Federated Semi-Supervised Learning/layout.json b/2023/Class Balanced Adaptive Pseudo Labeling for Federated Semi-Supervised Learning/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..11634e2addf279456c58a81a616c56480ebd5dd3 --- /dev/null +++ b/2023/Class Balanced Adaptive Pseudo Labeling for Federated Semi-Supervised Learning/layout.json @@ -0,0 +1,8762 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 70, + 102, + 523, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 102, + 523, + 140 + ], + "spans": [ + { + "bbox": [ + 70, + 102, + 523, + 140 + ], + "type": "text", + "content": "Class Balanced Adaptive Pseudo Labeling for Federated Semi-Supervised Learning" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 214, + 161, + 386, + 176 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 214, + 161, + 386, + 176 + ], + "spans": [ + { + "bbox": [ + 214, + 161, + 386, + 176 + ], + "type": "text", + "content": "Ming Li Qingli Li Yan Wang*" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 52, + 178, + 546, + 194 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 52, + 178, + 546, + 194 + ], + "spans": [ + { + "bbox": [ + 52, + 178, + 546, + 194 + ], + "type": "text", + "content": "Shanghai Key Laboratory of Multidimensional Information Processing, East China Normal University" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 121, + 195, + 478, + 206 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 121, + 195, + 478, + 206 + ], + "spans": [ + { + "bbox": [ + 121, + 195, + 478, + 206 + ], + "type": "text", + "content": "lm1640362161@gmail.com, qlli@cs.ecnu.edu.cn, ywang@cee.ecnu.edu.cn" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 143, + 234, + 192, + 247 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 234, + 192, + 247 + ], + "spans": [ + { + "bbox": [ + 143, + 234, + 192, + 247 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 259, + 289, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 259, + 289, + 594 + ], + "spans": [ + { + "bbox": [ + 46, + 259, + 289, + 594 + ], + "type": "text", + "content": "This paper focuses on federated semi-supervised learning (FSSL), assuming that few clients have fully labeled data (labeled clients) and the training datasets in other clients are fully unlabeled (unlabeled clients). Existing methods attempt to deal with the challenges caused by not independent and identically distributed data (Non-IID) setting. Though methods such as sub-consensus models have been proposed, they usually adopt standard pseudo labeling or consistency regularization on unlabeled clients which can be easily influenced by imbalanced class distribution. Thus, problems in FSSL are still yet to be solved. To seek for a fundamental solution to this problem, we present Class Balanced Adaptive Pseudo Labeling (CBAFed), to study FSSL from the perspective of pseudo labeling. In CBAFed, the first key element is a fixed pseudo labeling strategy to handle the catastrophic forgetting problem, where we keep a fixed set by letting pass information of unlabeled data at the beginning of the unlabeled client training in each communication round. The second key element is that we design class balanced adaptive thresholds via considering the empirical distribution of all training data in local clients, to encourage a balanced training process. To make the model reach a better optimum, we further propose a residual weight connection in local supervised training and global model aggregation. Extensive experiments on five datasets demonstrate the superiority of CBAFed. Code will be available at https://github.com/minglllli/CBAFed." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 616, + 128, + 628 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 616, + 128, + 628 + ], + "spans": [ + { + "bbox": [ + 47, + 616, + 128, + 628 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 636, + 287, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 636, + 287, + 696 + ], + "spans": [ + { + "bbox": [ + 46, + 636, + 287, + 696 + ], + "type": "text", + "content": "Federated learning (FL) aims to train machine learning models on a decentralized manner while preserving data privacy, i.e., separate local models are trained on separate local training datasets independently. In recent years, FL has received much attention for privacy protection reasons" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 235, + 546, + 308 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 235, + 546, + 308 + ], + "spans": [ + { + "bbox": [ + 306, + 235, + 546, + 308 + ], + "type": "text", + "content": "[32]. However, most of FL works focused on supervised learning with fully labeled data. But in practice, labeling process of large-scale training data is laborious and expensive. Due to the lack of funds or experts, large labeled training dataset is difficult for many companies and institutions to obtain. These may hinder applicability of FL." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 311, + 546, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 311, + 546, + 453 + ], + "spans": [ + { + "bbox": [ + 304, + 311, + 546, + 453 + ], + "type": "text", + "content": "To handle this problem, federated semi-supervised learning (FSSL) has been explored by many researchers recently [8, 14, 15]. There are broadly three lines of FSSL methods by considering different places and status of labeled data. The first two lines consider that there are only limited labeled data in the central server [8] or each client has partially labeled data [15]. The third line assumes that few clients have fully labeled data and the training datasets in other clients are fully unlabeled [14, 18, 31]. Our paper mainly focuses on the third line of FSSL, and we call local clients with fully labeled data as labeled clients and the other clients as unlabeled clients." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 457, + 547, + 615 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 457, + 547, + 615 + ], + "spans": [ + { + "bbox": [ + 304, + 457, + 547, + 615 + ], + "type": "text", + "content": "The main difficulties to train a third line FSSL model lie in three folds: 1) There are no labeled data in unlabeled clients. Thus, the training can be easily biased without label guidance. 2) Due to the divergent class distribution of labeled and unlabeled clients, namely the not independent and identically distributed data (Non-IID) setting, inaccurate supervisory signals may be generated in unlabeled clients via employing the model trained in labeled clients by either pseudo labeling or consistency regularization framework. 3) Due to the catastrophic forgetting problems in CNNs, with the training process of unlabeled clients going on, models may forget the knowledge learned on labeled clients and so decrease the prediction accuracy drastically." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 617, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 617, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 617, + 547, + 713 + ], + "type": "text", + "content": "RSCFed [14], the state-of-the-art method significantly boosts the FSSL performance by first distilling sub-consensus models, and then aggregating the sub-consensus models to the global model. The sub-consensus models can handle the Non-IID setting to some extent, but the meanteacher based consistency regularization framework in unlabeled clients inevitably causes the accuracy degradation when the classes are imbalanced distributed. RSCFed at" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 702, + 138, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 702, + 138, + 712 + ], + "spans": [ + { + "bbox": [ + 58, + 702, + 138, + 712 + ], + "type": "text", + "content": "*Corresponding Author." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "16292" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 191 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 191 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 191 + ], + "type": "text", + "content": "tempt to address the catastrophic forgetting problem by adjusting aggregation weights for labeled and unlabeled clients. But, it seems just alleviates the negative impact on the unlabeled clients and the problem is still yet to be tackled. Besides, the sub-consensus models may occur unstable training problems, and increase the communication cost. Other methods like FedConsist [31] and FedIRM [18] achieve promising results by utilizing pseudo labeling methods to produce artificial labels, but they do not consider the Non-IID setting among local clients." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 192, + 289, + 324 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 192, + 289, + 324 + ], + "spans": [ + { + "bbox": [ + 46, + 192, + 289, + 324 + ], + "type": "text", + "content": "Pseudo labeling methods are shown to be effective in semi-supervised learning (SSL) [13,24,33,37], which generate pseudo-labels for unlabeled images and propose several strategies to ensure the quality of pseudo-labels. Existing FSSL works only adopt standard pseudo labeling or consistency regularization on FSSL. Since the mentioned difficulties hamper the usage of these methods, other remedy is usually proposed to alleviate the difficulties, which is palliative. Thus, we propose to study FSSL from the perspective of pseudo labeling, seeking for a fundamental solution to this problem." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 324, + 289, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 324, + 289, + 647 + ], + "spans": [ + { + "bbox": [ + 46, + 324, + 289, + 647 + ], + "type": "text", + "content": "Concretely, we present Class Balanced Adaptive Pseudo Labeling, namely CBAFed, by rethinking standard pseudo labeling methods in SSL. To handle the catastrophic forgetting problem, we propose a fixed pseudo labeling strategy, which builds a fixed set by letting pass informative unlabeled data and their pseudo labels at the beginning of the unlabeled client training. Due to the Non-IID and heterogeneous data partition problems in FL, training distribution of unlabeled data can be highly imbalanced, so existing thresholds are not suitable in FSSL. We design class balanced adaptive thresholds via considering the empirical distribution of all training data in local clients at the previous communication round. Analysis proves that our method sets a reasonably high threshold for extremely scarce classes and encourages a balanced training process. To enhance the learning ability and discover unlabeled data from tail classes, we propose to leverage information from so-called \"not informative\" unlabeled data. Besides, we also explore a novel training strategy for labeled clients and the central server, termed as residual weight connection, skip connecting weights from previous epochs (for labeled clients) or previous global models (for the central server). It can help the model reach better optimum, when the training distribution is imbalanced and training amount is small. We conduct extensive experiments on five datasets to show the effectiveness of CBAFed. Overall, our main contributions can be summarized as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 58, + 653, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 653, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 58, + 653, + 288, + 714 + ], + "type": "text", + "content": "- We present a CBAFed method to deal with the catastrophic forgetting problems in federated learning. Unlike existing FSSL frameworks that directly adopts pseudo labeling or consistency regularization methods, CBAFed explores a fundamental solution to FSSL via" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 325, + 72, + 499, + 85 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 72, + 499, + 85 + ], + "spans": [ + { + "bbox": [ + 325, + 72, + 499, + 85 + ], + "type": "text", + "content": "designing a novel pseudo labeling method." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 317, + 91, + 547, + 219 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 317, + 91, + 545, + 151 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 91, + 545, + 151 + ], + "spans": [ + { + "bbox": [ + 317, + 91, + 545, + 151 + ], + "type": "text", + "content": "- We introduce a residual weight connection method, to improve the robustness of the models in labeled clients and the central server, which skip connects weights from previous epoch or communication round to finally reach better optimum." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 317, + 158, + 547, + 219 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 158, + 547, + 219 + ], + "spans": [ + { + "bbox": [ + 317, + 158, + 547, + 219 + ], + "type": "text", + "content": "- Experiments are conducted on five datasets: four natural datasets CIFAR-10/100, SVHN, fashion MNIST and one medical dataset ISIC 2018 Skin. CBAFed outperforms state-of-the-art FSSL methods by a large margin, with " + }, + { + "bbox": [ + 317, + 158, + 547, + 219 + ], + "type": "inline_equation", + "content": "11.3\\%" + }, + { + "bbox": [ + 317, + 158, + 547, + 219 + ], + "type": "text", + "content": " improvements on SVHN dataset." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 238, + 390, + 251 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 238, + 390, + 251 + ], + "spans": [ + { + "bbox": [ + 306, + 238, + 390, + 251 + ], + "type": "text", + "content": "2. Related work" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 258, + 422, + 270 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 258, + 422, + 270 + ], + "spans": [ + { + "bbox": [ + 306, + 258, + 422, + 270 + ], + "type": "text", + "content": "2.1. Federated Learning" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 277, + 545, + 433 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 277, + 545, + 433 + ], + "spans": [ + { + "bbox": [ + 304, + 277, + 545, + 433 + ], + "type": "text", + "content": "Recently, federated learning is becoming more and more popular for privacy protection reasons. Statistical data heterogeneity is one of the most important research problems. FedAvg [19], one of the pioneer works to deal with this problem, performs weighted-averaging on local weights according to the local training size, has been used as the most widely recognized FL baseline. To deal with data heterogeneity, two research directions have been mostly explored: model aggregation [12, 27, 28] and local training [9, 11, 12, 17]. Recently, [23] demonstrates that self-attention-based architectures, e.g., vision transformers(ViT) are more robust to distribution shifts and can converge to better optimum over heterogeneous data." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 441, + 447, + 453 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 441, + 447, + 453 + ], + "spans": [ + { + "bbox": [ + 306, + 441, + 447, + 453 + ], + "type": "text", + "content": "2.2. Semi-supervised Leaning" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 459, + 545, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 459, + 545, + 662 + ], + "spans": [ + { + "bbox": [ + 304, + 459, + 545, + 662 + ], + "type": "text", + "content": "The goal of standard semi-supervised learning (SSL) is to train a good model with limited labeled data and much more unlabeled data. Popular research directions involves consistency regularization [4, 26] and pseudo labeling [3, 10, 16]. Many recent works use the combinations of consistency regularization and pseudo labeling and achieves state-of-the-art performance [1, 2, 25, 30]. In this paper, we focus on using pseudo labeling method to deal with SSL. While these methods achieve much progress, they only consider pre-defined fixed threshold for pseudo labeling. FlexMatch [34] demonstrates that leveraging unlabeled data according to the model's learning status to flexibly adjust thresholds for different classes can boost performance. While these methods perform well in centralized SSL, they all update pseudo labels after every batch's update of the model, which is not suitable in FSSL as shown in later section." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 670, + 502, + 684 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 670, + 502, + 684 + ], + "spans": [ + { + "bbox": [ + 306, + 670, + 502, + 684 + ], + "type": "text", + "content": "2.3. Federated Semi-supervised Learning" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 689, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 545, + 714 + ], + "type": "text", + "content": "Recently, with the development of federated learning, FSSL becomes popular. There are broadly three categories" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16293" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 102, + 70, + 292, + 227 + ], + "blocks": [ + { + "bbox": [ + 102, + 70, + 292, + 227 + ], + "lines": [ + { + "bbox": [ + 102, + 70, + 292, + 227 + ], + "spans": [ + { + "bbox": [ + 102, + 70, + 292, + 227 + ], + "type": "image", + "image_path": "2a2c958fb80ec8b46cc44c4281bb2b9c801e7cbfc89bebfd2e65e7f2a988a7bd.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 240, + 547, + 308 + ], + "lines": [ + { + "bbox": [ + 46, + 240, + 547, + 308 + ], + "spans": [ + { + "bbox": [ + 46, + 240, + 547, + 308 + ], + "type": "text", + "content": "Figure 1. An overview of our CBAFed. In the central server (left side), the global model is aggregated with the returned local models (step ④) and the adaptive thresholds are calculated by the returned training data statistics (step ⑤). Then central server passes the global model, adaptive thresholds and class distribution to all local clients (step ①). After downloading these data, local clients perform local training on the right side (step ②). Labeled clients use labeled data to train the model with residual weight connection. Unlabeled clients obtain the new training dataset by adaptive pseudo labeling and tail class data discovery and use it to train the model. After local training, local clients return trained models and number of data in each class back to central server (step ③)." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 294, + 70, + 493, + 227 + ], + "blocks": [ + { + "bbox": [ + 294, + 70, + 493, + 227 + ], + "lines": [ + { + "bbox": [ + 294, + 70, + 493, + 227 + ], + "spans": [ + { + "bbox": [ + 294, + 70, + 493, + 227 + ], + "type": "image", + "image_path": "e91d5849fb9ef1c8e6afb98aeea5656fcef12b04af4e280f30f12c879f9f5576.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 323, + 289, + 504 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 323, + 289, + 504 + ], + "spans": [ + { + "bbox": [ + 46, + 323, + 289, + 504 + ], + "type": "text", + "content": "of FSSL. The first two categories assume that every local client contains partially labeled data or there exists some labeled data in central server [8, 15]. The third category considers that few clients are fully labeled while other clients contain only unlabeled data. This paper focuses on the third category. To deal with this problem, Fed-Consist [31] and FedIRM [18] have been proposed. But, these two methods do not consider data heterogeneity in federated learning. Recently, RSCFed [14] proposes to perform random sub-sampling to reach consensus over clients. However, it uses standard consistency regularization for unlabeled data, which still suffers from the Non-IID setting. We emphasize that all the aforementioned approaches just use traditional pseudo labeling strategy, and none of them discuss how to apply pseudo labeling to FSSL." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 512, + 130, + 525 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 512, + 130, + 525 + ], + "spans": [ + { + "bbox": [ + 47, + 512, + 130, + 525 + ], + "type": "text", + "content": "3. Methodology" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 532, + 287, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 532, + 287, + 666 + ], + "spans": [ + { + "bbox": [ + 47, + 532, + 287, + 666 + ], + "type": "text", + "content": "We first introduce the problem setting of federated semi-supervised learning. We consider decentralized training with labeled clients and unlabeled clients. Let " + }, + { + "bbox": [ + 47, + 532, + 287, + 666 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 47, + 532, + 287, + 666 + ], + "type": "text", + "content": " (usually " + }, + { + "bbox": [ + 47, + 532, + 287, + 666 + ], + "type": "inline_equation", + "content": "m = 1" + }, + { + "bbox": [ + 47, + 532, + 287, + 666 + ], + "type": "text", + "content": ") labeled clients be " + }, + { + "bbox": [ + 47, + 532, + 287, + 666 + ], + "type": "inline_equation", + "content": "\\{C_1, C_2, \\dots, C_m\\}" + }, + { + "bbox": [ + 47, + 532, + 287, + 666 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 532, + 287, + 666 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 47, + 532, + 287, + 666 + ], + "type": "text", + "content": " (usually " + }, + { + "bbox": [ + 47, + 532, + 287, + 666 + ], + "type": "inline_equation", + "content": "n > 1" + }, + { + "bbox": [ + 47, + 532, + 287, + 666 + ], + "type": "text", + "content": ") unlabeled clients be " + }, + { + "bbox": [ + 47, + 532, + 287, + 666 + ], + "type": "inline_equation", + "content": "\\{C_{m+1}, C_{m+2}, \\dots, C_{m+n}\\}" + }, + { + "bbox": [ + 47, + 532, + 287, + 666 + ], + "type": "text", + "content": ". For labeled client " + }, + { + "bbox": [ + 47, + 532, + 287, + 666 + ], + "type": "inline_equation", + "content": "C_\\ell" + }, + { + "bbox": [ + 47, + 532, + 287, + 666 + ], + "type": "text", + "content": ", the labeled local training dataset is denoted as " + }, + { + "bbox": [ + 47, + 532, + 287, + 666 + ], + "type": "inline_equation", + "content": "D_\\ell = \\{(X_i^\\ell, y_i^\\ell)\\}_{i=1}^{N_\\ell}" + }, + { + "bbox": [ + 47, + 532, + 287, + 666 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 47, + 532, + 287, + 666 + ], + "type": "inline_equation", + "content": "\\ell = 1, 2, \\dots, m" + }, + { + "bbox": [ + 47, + 532, + 287, + 666 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 47, + 532, + 287, + 666 + ], + "type": "inline_equation", + "content": "N_\\ell" + }, + { + "bbox": [ + 47, + 532, + 287, + 666 + ], + "type": "text", + "content": " denotes the data number in " + }, + { + "bbox": [ + 47, + 532, + 287, + 666 + ], + "type": "inline_equation", + "content": "D_\\ell" + }, + { + "bbox": [ + 47, + 532, + 287, + 666 + ], + "type": "text", + "content": ". Similarly, for unlabeled client " + }, + { + "bbox": [ + 47, + 532, + 287, + 666 + ], + "type": "inline_equation", + "content": "C_\\mu" + }, + { + "bbox": [ + 47, + 532, + 287, + 666 + ], + "type": "text", + "content": ", the unlabeled local training dataset is denoted as " + }, + { + "bbox": [ + 47, + 532, + 287, + 666 + ], + "type": "inline_equation", + "content": "D_\\mu = \\{X_i^\\mu\\}_{i=1}^{N_\\mu}" + }, + { + "bbox": [ + 47, + 532, + 287, + 666 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 47, + 532, + 287, + 666 + ], + "type": "inline_equation", + "content": "\\mu = m + 1, m + 2, \\dots, n + m" + }, + { + "bbox": [ + 47, + 532, + 287, + 666 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 666, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 666, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 666, + 288, + 714 + ], + "type": "text", + "content": "The overall pipeline of the proposed CBAFed is shown in Fig. 1, which performs the following steps: (1) Warm up stage: train fully supervised models on only labeled clients using residual weight connection in a normal federated" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 323, + 547, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 323, + 547, + 552 + ], + "spans": [ + { + "bbox": [ + 304, + 323, + 547, + 552 + ], + "type": "text", + "content": "learning manner (Sec. 3.1) (2) The central server computes the empirical class distribution and obtains the class balanced adaptive thresholds, then passes them to local clients (Sec. 3.3); (3) All local clients update local models, adaptive threshold and class distribution. Labeled clients train local models on all the data using proposed residual weight connection; Unlabeled clients acquire the fixed training set (Sec. 3.2.2) by the threshold and the tail class datasets (Sec. 3.4), and train local models on the newly obtained training dataset. After local training, all local clients pass the trained model, the number of data in each class and the amount of training data to the central server; (4) The central server aggregates a new model with residual weight connection, computes the class distribution, and obtains the class balanced adaptive threshold. Then, the central server passes them to local clients; (5) Repeat step (3)-(4) until the specified number of communication round is reached. The overall training procedure of our CBAFed is provided in Algorithm 1 in the supplementary material." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 571, + 463, + 583 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 571, + 463, + 583 + ], + "spans": [ + { + "bbox": [ + 306, + 571, + 463, + 583 + ], + "type": "text", + "content": "3.1. Residual Weight Connection" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "type": "text", + "content": "We propose a novel training method in local supervised training and global model aggregation called residual weight connection. Our idea of res-weight is simple and similar to ResNet. In ResNet, there is a skip connection between every layer. For our res-weight connection, there is a skip connection of model's parameters between training epochs (or communication rounds). Let " + }, + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\theta^{E}" + }, + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "type": "text", + "content": " be the parameters of model after the training epoch (communication round) " + }, + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "type": "inline_equation", + "content": "E(E > 0)" + }, + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "type": "text", + "content": ", then the final parameters of model after this training epoch by using residual weight connection will" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "16294" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 73, + 60, + 83 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 73, + 60, + 83 + ], + "spans": [ + { + "bbox": [ + 47, + 73, + 60, + 83 + ], + "type": "text", + "content": "be" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 77, + 81, + 287, + 113 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 81, + 287, + 113 + ], + "spans": [ + { + "bbox": [ + 77, + 81, + 287, + 113 + ], + "type": "interline_equation", + "content": "\\theta^ {E} = \\left\\{ \\begin{array}{l l} \\theta^ {E} & E \\% s \\neq 0 \\\\ \\alpha_ {1} \\theta^ {E - s} + (1 - \\alpha_ {1}) \\theta^ {E} & E \\% s = 0 \\end{array} \\right. \\tag{1}", + "image_path": "247d5c2b71bffffcc66e6b739a2a7592598b9237264bd1edec520c107667285d.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 119, + 289, + 262 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 119, + 289, + 262 + ], + "spans": [ + { + "bbox": [ + 46, + 119, + 289, + 262 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 119, + 289, + 262 + ], + "type": "inline_equation", + "content": "\\alpha_{1}" + }, + { + "bbox": [ + 46, + 119, + 289, + 262 + ], + "type": "text", + "content": " is the parameter to scale the size of " + }, + { + "bbox": [ + 46, + 119, + 289, + 262 + ], + "type": "inline_equation", + "content": "\\theta^{E - s}" + }, + { + "bbox": [ + 46, + 119, + 289, + 262 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 119, + 289, + 262 + ], + "type": "inline_equation", + "content": "\\theta^{E}" + }, + { + "bbox": [ + 46, + 119, + 289, + 262 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 119, + 289, + 262 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 46, + 119, + 289, + 262 + ], + "type": "text", + "content": " is the number of skip epoch. In [22], authors show that averaging model weights over training steps tends to produce a more accurate model than using the final weights directly. In our experiments, we will show that this training method can effectively make the model more robust and finally reach better optimum, enhancing the predicted accuracy of pseudo labeling. Note that although the formula of our residual weight connection is similar to mean teachers [26], the update strategy and usage scenario is very different from it. Please refer to supplementary material for more detailed comparison and discussions." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 271, + 192, + 283 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 271, + 192, + 283 + ], + "spans": [ + { + "bbox": [ + 47, + 271, + 192, + 283 + ], + "type": "text", + "content": "3.2. Pseudo Labeling Methods" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 289, + 284, + 302 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 289, + 284, + 302 + ], + "spans": [ + { + "bbox": [ + 47, + 289, + 284, + 302 + ], + "type": "text", + "content": "3.2.1 Review of Classic Batch-based Pseudo Labeling" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 308, + 287, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 308, + 287, + 380 + ], + "spans": [ + { + "bbox": [ + 46, + 308, + 287, + 380 + ], + "type": "text", + "content": "In semi-supervised learning, for unlabeled data, traditional approaches [2, 25, 34], such as Fixmatch [25] and Flexmatch [34], use original data or their weak augmented version in one batch to generate pseudo labels. These labels are adopted to supervise model's training. For detailed review, please refer to supplementary materials." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 396, + 214, + 407 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 396, + 214, + 407 + ], + "spans": [ + { + "bbox": [ + 47, + 396, + 214, + 407 + ], + "type": "text", + "content": "3.2.2 Fixed Pseudo Labeling in FSSL" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 415, + 287, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 415, + 287, + 475 + ], + "spans": [ + { + "bbox": [ + 46, + 415, + 287, + 475 + ], + "type": "text", + "content": "Warm up stage. To conduct pseudo labeling for unlabeled data, in the warm up stage, we train our model only with labeled data. In the first " + }, + { + "bbox": [ + 46, + 415, + 287, + 475 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 46, + 415, + 287, + 475 + ], + "type": "text", + "content": " communication round, training only conducted in labeled clients. In local training on labeled client " + }, + { + "bbox": [ + 46, + 415, + 287, + 475 + ], + "type": "inline_equation", + "content": "C_{\\ell}" + }, + { + "bbox": [ + 46, + 415, + 287, + 475 + ], + "type": "text", + "content": ", the training loss is defined as:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 75, + 483, + 287, + 514 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 483, + 287, + 514 + ], + "spans": [ + { + "bbox": [ + 75, + 483, + 287, + 514 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\ell} = \\frac {1}{| D _ {\\ell} |} \\sum_ {(X ^ {\\ell}, y ^ {\\ell}) \\in D _ {\\ell}} H \\left(y ^ {\\ell}, p _ {m} \\left(y \\mid \\theta_ {t} ^ {\\ell} \\left(X ^ {\\ell}\\right)\\right)\\right), \\tag {2}", + "image_path": "de1b4e4db2bd93c3d7104fa11251260111fdbd1a3c3537bf86a4c62e58a80bf5.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 524, + 287, + 570 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 524, + 287, + 570 + ], + "spans": [ + { + "bbox": [ + 47, + 524, + 287, + 570 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 524, + 287, + 570 + ], + "type": "inline_equation", + "content": "\\theta_t^\\ell" + }, + { + "bbox": [ + 47, + 524, + 287, + 570 + ], + "type": "text", + "content": " means local model in labeled client " + }, + { + "bbox": [ + 47, + 524, + 287, + 570 + ], + "type": "inline_equation", + "content": "C_\\ell" + }, + { + "bbox": [ + 47, + 524, + 287, + 570 + ], + "type": "text", + "content": " at communication round " + }, + { + "bbox": [ + 47, + 524, + 287, + 570 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 47, + 524, + 287, + 570 + ], + "type": "text", + "content": ". The local training epoch is set to " + }, + { + "bbox": [ + 47, + 524, + 287, + 570 + ], + "type": "inline_equation", + "content": "J(J > 1)" + }, + { + "bbox": [ + 47, + 524, + 287, + 570 + ], + "type": "text", + "content": " with residual weight convection. Then we use FedAVG [19] with residual weight connection for model aggregation:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 50, + 577, + 287, + 616 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 50, + 577, + 287, + 616 + ], + "spans": [ + { + "bbox": [ + 50, + 577, + 287, + 616 + ], + "type": "interline_equation", + "content": "\\theta_ {t + 1} ^ {G} = \\left\\{ \\begin{array}{l l} \\sum_ {\\ell = 1} ^ {m} \\frac {| D _ {\\ell} |}{\\sum_ {i = 1} ^ {m} | D _ {i} |} \\theta_ {t} ^ {\\ell} & t \\% s \\neq 0 \\\\ \\alpha_ {1} \\theta_ {t + 1 - s} ^ {G} + (1 - \\alpha_ {1}) \\sum_ {\\ell = 1} ^ {m} \\frac {| D _ {\\ell} |}{\\sum_ {i = 1} ^ {m} | D _ {i} |} \\theta_ {t} ^ {\\ell} & t \\% s = 0 \\end{array} \\right. \\tag{3}", + "image_path": "65a67575a0851babb3aa7886c503fbb76ca78a2eb394bd687891ab2cef55f51d.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 617, + 287, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 617, + 287, + 665 + ], + "spans": [ + { + "bbox": [ + 47, + 617, + 287, + 665 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 617, + 287, + 665 + ], + "type": "inline_equation", + "content": "\\theta_{t + 1}^{G}" + }, + { + "bbox": [ + 47, + 617, + 287, + 665 + ], + "type": "text", + "content": " is the global model of communication round " + }, + { + "bbox": [ + 47, + 617, + 287, + 665 + ], + "type": "inline_equation", + "content": "t+" + }, + { + "bbox": [ + 47, + 617, + 287, + 665 + ], + "type": "text", + "content": " 1. Note that, usually, the number of labeled clients is only 1, so we can finish warm up stage in only one communication round(training can be completed on one labeled client)." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 47, + 665, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 665, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 665, + 287, + 713 + ], + "type": "text", + "content": "Different from traditional semi-supervised learning, in federated semi-supervised learning, Non-IID data partitions in clients can easily lead to catastrophic forgetting [7, 21, 23]. Thus, local models trained on unlabeled clients" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 72, + 545, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 180 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 180 + ], + "type": "text", + "content": "usually forget the former learned knowledge by batch-based pseudo labeling method, which will lead to abrupt degradation and sabotage the training of unlabeled clients. To handle heterogeneous data partitions and catastrophic forgetting, we propose a new pseudo labeling method called fixed pseudo labeling. For the unlabeled training dataset " + }, + { + "bbox": [ + 304, + 72, + 545, + 180 + ], + "type": "inline_equation", + "content": "D_{\\mu}" + }, + { + "bbox": [ + 304, + 72, + 545, + 180 + ], + "type": "text", + "content": " in an unlabeled client " + }, + { + "bbox": [ + 304, + 72, + 545, + 180 + ], + "type": "inline_equation", + "content": "C_{\\mu}" + }, + { + "bbox": [ + 304, + 72, + 545, + 180 + ], + "type": "text", + "content": ", after initializing the local model " + }, + { + "bbox": [ + 304, + 72, + 545, + 180 + ], + "type": "inline_equation", + "content": "\\theta_t^\\mu" + }, + { + "bbox": [ + 304, + 72, + 545, + 180 + ], + "type": "text", + "content": " with the global model " + }, + { + "bbox": [ + 304, + 72, + 545, + 180 + ], + "type": "inline_equation", + "content": "\\theta_t^G" + }, + { + "bbox": [ + 304, + 72, + 545, + 180 + ], + "type": "text", + "content": ", we first compute the pseudo label for each data " + }, + { + "bbox": [ + 304, + 72, + 545, + 180 + ], + "type": "inline_equation", + "content": "X_i^\\mu" + }, + { + "bbox": [ + 304, + 72, + 545, + 180 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 304, + 72, + 545, + 180 + ], + "type": "inline_equation", + "content": "D_{\\mu}" + }, + { + "bbox": [ + 304, + 72, + 545, + 180 + ], + "type": "text", + "content": " via:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 321, + 185, + 545, + 199 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 185, + 545, + 199 + ], + "spans": [ + { + "bbox": [ + 321, + 185, + 545, + 199 + ], + "type": "interline_equation", + "content": "\\hat {y} _ {i} ^ {\\mu} = \\arg \\max p _ {m} (y | \\theta_ {t} ^ {\\mu} \\left(X _ {i} ^ {\\mu}\\right), i = 1, 2, \\dots , N _ {\\mu}. \\tag {4}", + "image_path": "87b433d3a9eb81f4efbaceca23ae22d043b85613dc2bac07cb93c015cbe6f807.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 202, + 545, + 226 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 202, + 545, + 226 + ], + "spans": [ + { + "bbox": [ + 304, + 202, + 545, + 226 + ], + "type": "text", + "content": "We obtain a subset of unlabeled data with their pseudo labels as supervisory signals during local client training:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 229, + 545, + 255 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 229, + 545, + 255 + ], + "spans": [ + { + "bbox": [ + 306, + 229, + 545, + 255 + ], + "type": "interline_equation", + "content": "\\widetilde {\\mathcal {D}} _ {\\mu} = \\left\\{\\left(X _ {i} ^ {\\mu}, \\hat {y} _ {i} ^ {\\mu}\\right) \\mid X _ {i} ^ {\\mu} \\in D _ {\\mu} \\wedge \\max \\left(p _ {m} \\left(y \\mid \\theta_ {t} ^ {\\mu} \\left(X _ {i} ^ {\\mu}\\right)\\right)\\right) > \\tau \\right\\} _ {i = 1} ^ {N _ {\\mu}}. \\tag {5}", + "image_path": "a21d74890fdb7fd48bb93d7c7c8dd3f94306f9ca15c6323d9cee80a5e75c872f.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 254, + 545, + 295 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 254, + 545, + 295 + ], + "spans": [ + { + "bbox": [ + 306, + 254, + 545, + 295 + ], + "type": "text", + "content": "Then we use this fixed set " + }, + { + "bbox": [ + 306, + 254, + 545, + 295 + ], + "type": "inline_equation", + "content": "\\widetilde{\\mathcal{D}}_{\\mu}" + }, + { + "bbox": [ + 306, + 254, + 545, + 295 + ], + "type": "text", + "content": " as the training dataset and define the loss to train our unlabeled client as: " + }, + { + "bbox": [ + 306, + 254, + 545, + 295 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mu} = \\frac{1}{|\\widetilde{\\mathcal{D}}_{\\mu}|}\\sum_{(X^{\\mu},\\hat{y}^{\\mu})\\in \\widetilde{\\mathcal{D}}_{\\mu}}H(\\hat{y}^{\\mu},p_m(y|\\theta_t^\\mu (X^\\mu)))." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 306, + 299, + 545, + 324 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 299, + 545, + 324 + ], + "spans": [ + { + "bbox": [ + 306, + 299, + 545, + 324 + ], + "type": "text", + "content": "3.3. Class Balanced Adaptive Threshold for Pseudo Labeling" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 330, + 546, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 330, + 546, + 616 + ], + "spans": [ + { + "bbox": [ + 304, + 330, + 546, + 616 + ], + "type": "text", + "content": "In semi-supervised learning, the value of threshold " + }, + { + "bbox": [ + 304, + 330, + 546, + 616 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 304, + 330, + 546, + 616 + ], + "type": "text", + "content": " is non-trivial for selecting confident pseudo labels, namely informative unlabeled data. Traditionally, a fixed threshold is pre-defined [25], but setting a fixed threshold usually makes the model fail to consider different learning status and learning difficulties of different classes. In one of the most recent works, Curriculum Pseudo Labeling [34] is proposed to achieve state-of-the-art results on most semi-supervised learning benchmarks, where a flexible threshold for class " + }, + { + "bbox": [ + 304, + 330, + 546, + 616 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 304, + 330, + 546, + 616 + ], + "type": "text", + "content": " at time step " + }, + { + "bbox": [ + 304, + 330, + 546, + 616 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 330, + 546, + 616 + ], + "type": "text", + "content": " is calculated by: " + }, + { + "bbox": [ + 304, + 330, + 546, + 616 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_t(c) = \\beta_t(c)\\cdot \\tau" + }, + { + "bbox": [ + 304, + 330, + 546, + 616 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 330, + 546, + 616 + ], + "type": "inline_equation", + "content": "\\beta_{t}(c)" + }, + { + "bbox": [ + 304, + 330, + 546, + 616 + ], + "type": "text", + "content": " is the ratio of the number of selected pseudo labels in class " + }, + { + "bbox": [ + 304, + 330, + 546, + 616 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 304, + 330, + 546, + 616 + ], + "type": "text", + "content": " and the maximum number of selected pseudo labels of all classes. " + }, + { + "bbox": [ + 304, + 330, + 546, + 616 + ], + "type": "inline_equation", + "content": "\\beta_{t}(c)" + }, + { + "bbox": [ + 304, + 330, + 546, + 616 + ], + "type": "text", + "content": " is used to scale the fixed threshold " + }, + { + "bbox": [ + 304, + 330, + 546, + 616 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 304, + 330, + 546, + 616 + ], + "type": "text", + "content": ". Curriculum Pseudo Labeling significantly boosts the accuracy and convergence performance of various semi-supervised learning algorithms. But in FSSL, there exist heterogeneous data partition problems. Due to the Non-IID partition, the labeled data are not balanced, so purely using the number of selected unlabeled data to design threshold is improper. Moreover, the data of some classes can be extremely rare in certain clients, which will lead to a very low " + }, + { + "bbox": [ + 304, + 330, + 546, + 616 + ], + "type": "inline_equation", + "content": "\\beta_{t}(c)" + }, + { + "bbox": [ + 304, + 330, + 546, + 616 + ], + "type": "text", + "content": ", and thus, a very low threshold " + }, + { + "bbox": [ + 304, + 330, + 546, + 616 + ], + "type": "inline_equation", + "content": "(\\mathcal{T}_t(c)\\approx 0)" + }, + { + "bbox": [ + 304, + 330, + 546, + 616 + ], + "type": "text", + "content": ". In other words, directly applying Curriculum Pseudo Labeling will introduce many noisy labels into training." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 616, + 545, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 616, + 545, + 664 + ], + "spans": [ + { + "bbox": [ + 304, + 616, + 545, + 664 + ], + "type": "text", + "content": "To this end, we propose a novel Class Balanced Adaptive threshold for Pseudo Labeling (CBAPL). For an unlabeled client " + }, + { + "bbox": [ + 304, + 616, + 545, + 664 + ], + "type": "inline_equation", + "content": "C_{\\mu}" + }, + { + "bbox": [ + 304, + 616, + 545, + 664 + ], + "type": "text", + "content": " in the " + }, + { + "bbox": [ + 304, + 616, + 545, + 664 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 616, + 545, + 664 + ], + "type": "text", + "content": "th communication round, let the number of selected pseudo labels in class " + }, + { + "bbox": [ + 304, + 616, + 545, + 664 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 304, + 616, + 545, + 664 + ], + "type": "text", + "content": " be:" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 669, + 545, + 713 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 669, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 669, + 545, + 713 + ], + "type": "interline_equation", + "content": "\\sigma_ {t} ^ {\\mu} (c) = \\sum_ {i = 1} ^ {N _ {\\mu}} \\mathbf {1} \\left(\\max \\left(p _ {m} \\left(y \\mid \\theta_ {t} ^ {\\mu} \\left(X _ {i} ^ {\\mu}\\right)\\right)\\right) > \\mathcal {T} _ {t} (c)\\right) \\mathbf {1} \\left(\\hat {y} _ {i} ^ {\\mu} = c\\right), \\tag {6}", + "image_path": "74040e71b138666344de9c7a2166a96ab2e3b7ab1ec37ee91dc829e0fbced47b.jpg" + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16295" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 289, + 145 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 289, + 145 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 289, + 145 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 72, + 289, + 145 + ], + "type": "inline_equation", + "content": "\\mathbf{1}(\\cdot)" + }, + { + "bbox": [ + 47, + 72, + 289, + 145 + ], + "type": "text", + "content": " is the indicator function. The data number of class " + }, + { + "bbox": [ + 47, + 72, + 289, + 145 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 47, + 72, + 289, + 145 + ], + "type": "text", + "content": " in labeled client " + }, + { + "bbox": [ + 47, + 72, + 289, + 145 + ], + "type": "inline_equation", + "content": "C_{\\ell}" + }, + { + "bbox": [ + 47, + 72, + 289, + 145 + ], + "type": "text", + "content": " is calculated by: " + }, + { + "bbox": [ + 47, + 72, + 289, + 145 + ], + "type": "inline_equation", + "content": "\\sigma_t^\\ell (c) = \\sum_{i = 1}^{N_\\ell}\\mathbf{1}(y_i^\\ell = c)" + }, + { + "bbox": [ + 47, + 72, + 289, + 145 + ], + "type": "text", + "content": ". After local training, " + }, + { + "bbox": [ + 47, + 72, + 289, + 145 + ], + "type": "inline_equation", + "content": "\\sigma_t^\\mu (c),\\sigma_t^\\ell (c)" + }, + { + "bbox": [ + 47, + 72, + 289, + 145 + ], + "type": "text", + "content": " and weights of local models are returned to central server. Then the central server calculates the total number of training data of class " + }, + { + "bbox": [ + 47, + 72, + 289, + 145 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 47, + 72, + 289, + 145 + ], + "type": "text", + "content": " at communication round " + }, + { + "bbox": [ + 47, + 72, + 289, + 145 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 47, + 72, + 289, + 145 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 94, + 153, + 287, + 186 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 94, + 153, + 287, + 186 + ], + "spans": [ + { + "bbox": [ + 94, + 153, + 287, + 186 + ], + "type": "interline_equation", + "content": "\\sigma_ {t} (c) = \\sum_ {\\ell = 1} ^ {m} \\sigma_ {t} ^ {\\ell} (c) + \\sum_ {\\mu = m + 1} ^ {n + m} \\sigma_ {t} ^ {\\mu} (c). \\tag {7}", + "image_path": "1757f035343a583fb77f13e19e4610121ad07674fb626d253fa9496257c71632.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 194, + 287, + 219 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 194, + 287, + 219 + ], + "spans": [ + { + "bbox": [ + 47, + 194, + 287, + 219 + ], + "type": "text", + "content": "We then normalize the training numbers and acquire the empirical distribution of training data:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 122, + 227, + 287, + 257 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 122, + 227, + 287, + 257 + ], + "spans": [ + { + "bbox": [ + 122, + 227, + 287, + 257 + ], + "type": "interline_equation", + "content": "\\tilde {p} _ {t} (c) = \\frac {\\sigma_ {t} (c)}{\\sum_ {i = 1} ^ {C} \\sigma_ {t} (i)}. \\tag {8}", + "image_path": "fb15ad2c506a81f432e33638e3b6f1367dbe5ea6d3f8c9b31aad33f0dbc2c64b.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 264, + 287, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 264, + 287, + 289 + ], + "spans": [ + { + "bbox": [ + 47, + 264, + 287, + 289 + ], + "type": "text", + "content": "The standard deviation of the empirical distribution is calculated by:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 88, + 297, + 287, + 335 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 88, + 297, + 287, + 335 + ], + "spans": [ + { + "bbox": [ + 88, + 297, + 287, + 335 + ], + "type": "interline_equation", + "content": "\\operatorname {s t d} \\left(\\widetilde {p} _ {t}\\right) = \\sqrt {\\frac {1}{C - 1} \\sum_ {c = 1} ^ {C} \\left(\\widetilde {p} _ {t} (c) - \\bar {p} _ {t}\\right) ^ {2}}, \\tag {9}", + "image_path": "a53df1350027573daeebc597f6d65fb38cf85386ef6777d39d4c82408947a627.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 346, + 287, + 371 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 346, + 287, + 371 + ], + "spans": [ + { + "bbox": [ + 47, + 346, + 287, + 371 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 346, + 287, + 371 + ], + "type": "inline_equation", + "content": "\\overline{p}_t = \\frac{1}{C}\\sum_{c = 1}^{C}\\widetilde{p}_t(c)" + }, + { + "bbox": [ + 47, + 346, + 287, + 371 + ], + "type": "text", + "content": ". Lastly, the threshold of class " + }, + { + "bbox": [ + 47, + 346, + 287, + 371 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 47, + 346, + 287, + 371 + ], + "type": "text", + "content": " is calculated by" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 110, + 380, + 287, + 396 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 110, + 380, + 287, + 396 + ], + "spans": [ + { + "bbox": [ + 110, + 380, + 287, + 396 + ], + "type": "interline_equation", + "content": "\\tau_ {t, c} = \\widetilde {p} _ {t} (c) + \\tau - s t d (\\widetilde {p} _ {t}), \\tag {10}", + "image_path": "e8e369a6377dd7327fa12d3a4b256a88cd685189070cf0b198a4704299ed4a26.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 403, + 287, + 427 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 403, + 287, + 427 + ], + "spans": [ + { + "bbox": [ + 47, + 403, + 287, + 427 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 403, + 287, + 427 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 47, + 403, + 287, + 427 + ], + "type": "text", + "content": " is the pre-defined threshold base. We then set an upper bound of threshold as:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 104, + 436, + 287, + 468 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 104, + 436, + 287, + 468 + ], + "spans": [ + { + "bbox": [ + 104, + 436, + 287, + 468 + ], + "type": "interline_equation", + "content": "\\mathcal {T} _ {t + 1} (c) = \\left\\{ \\begin{array}{l l} \\tau_ {t, c}, & \\tau_ {t, c} < \\tau_ {h} \\\\ \\tau_ {h}, & \\tau_ {t, c} \\geq \\tau_ {h} \\end{array} , \\right. \\tag {11}", + "image_path": "2ec46e7604ea9a11f6b5dbdea42ff04812d822524bb9e3f705870d3272deb56f.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 477, + 287, + 560 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 477, + 287, + 560 + ], + "spans": [ + { + "bbox": [ + 47, + 477, + 287, + 560 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 477, + 287, + 560 + ], + "type": "inline_equation", + "content": "\\tau_h" + }, + { + "bbox": [ + 47, + 477, + 287, + 560 + ], + "type": "text", + "content": " is the pre-defined upper bound of threshold (usually " + }, + { + "bbox": [ + 47, + 477, + 287, + 560 + ], + "type": "inline_equation", + "content": "\\tau_h = 0.95" + }, + { + "bbox": [ + 47, + 477, + 287, + 560 + ], + "type": "text", + "content": " for ensuring a certain amount of training data in unlabeled clients). If our model is trained with more data in one class, the threshold of that class will be higher since " + }, + { + "bbox": [ + 47, + 477, + 287, + 560 + ], + "type": "inline_equation", + "content": "\\widetilde{p}_t(c)" + }, + { + "bbox": [ + 47, + 477, + 287, + 560 + ], + "type": "text", + "content": " is higher. " + }, + { + "bbox": [ + 47, + 477, + 287, + 560 + ], + "type": "inline_equation", + "content": "std(\\widetilde{p}_t)" + }, + { + "bbox": [ + 47, + 477, + 287, + 560 + ], + "type": "text", + "content": " is important for balancing the empirical distribution of training data. Please refer to the supplementary material for more discussions." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 561, + 287, + 608 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 561, + 287, + 608 + ], + "spans": [ + { + "bbox": [ + 47, + 561, + 287, + 608 + ], + "type": "text", + "content": "We argue that for scarce classes, unlike Curriculum Pseudo Labeling [34] whose threshold is extremely low, our CBAPL calculates a threshold with a relatively high lower bound. We have theorem below:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 47, + 617, + 107, + 628 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 617, + 107, + 628 + ], + "spans": [ + { + "bbox": [ + 47, + 617, + 107, + 628 + ], + "type": "text", + "content": "Theorem 3.1." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 85, + 636, + 287, + 663 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 85, + 636, + 287, + 663 + ], + "spans": [ + { + "bbox": [ + 85, + 636, + 287, + 663 + ], + "type": "interline_equation", + "content": "\\tau + \\widetilde {p} _ {t} (c) - \\sqrt {\\frac {1}{C}} \\leq \\mathcal {T} _ {t} (c) \\leq \\tau + \\widetilde {p} _ {t} (c), \\tag {12}", + "image_path": "f9341ca3d788379603329550c8176ca66d2ce42f8f8a9625a699b4a8de25cb82.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 59, + 671, + 255, + 683 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 671, + 255, + 683 + ], + "spans": [ + { + "bbox": [ + 59, + 671, + 255, + 683 + ], + "type": "text", + "content": "Proof is provided in the supplementary material." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 47, + 685, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 685, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 685, + 287, + 714 + ], + "type": "text", + "content": "Since " + }, + { + "bbox": [ + 47, + 685, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\tau >>\\sqrt{\\frac{1}{C}}" + }, + { + "bbox": [ + 47, + 685, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_t(c)" + }, + { + "bbox": [ + 47, + 685, + 287, + 714 + ], + "type": "text", + "content": " will have a high lower bound " + }, + { + "bbox": [ + 47, + 685, + 287, + 714 + ], + "type": "inline_equation", + "content": "(>>0)" + }, + { + "bbox": [ + 47, + 685, + 287, + 714 + ], + "type": "text", + "content": ". To balance the trade-off between " + }, + { + "bbox": [ + 47, + 685, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\widetilde{p}_t(t)" + }, + { + "bbox": [ + 47, + 685, + 287, + 714 + ], + "type": "text", + "content": " and" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 305, + 72, + 545, + 97 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 545, + 97 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 545, + 97 + ], + "type": "inline_equation", + "content": "\\tau - \\sqrt{1 / C}" + }, + { + "bbox": [ + 305, + 72, + 545, + 97 + ], + "type": "text", + "content": ", a normalization factor is applied to consider the class number. Thus, we propose a modified " + }, + { + "bbox": [ + 305, + 72, + 545, + 97 + ], + "type": "inline_equation", + "content": "\\widetilde{p}_t(c)" + }, + { + "bbox": [ + 305, + 72, + 545, + 97 + ], + "type": "text", + "content": " as below:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 368, + 104, + 545, + 133 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 368, + 104, + 545, + 133 + ], + "spans": [ + { + "bbox": [ + 368, + 104, + 545, + 133 + ], + "type": "interline_equation", + "content": "\\widetilde {p} _ {t} (c) = \\frac {\\sigma_ {t} (c)}{\\sum_ {i = 1} ^ {C} \\sigma_ {t} (i)} \\times \\frac {C}{1 0}. \\tag {13}", + "image_path": "e1cc2ed8950decb8ee83269fa11585ba7d3b4950ce94da40a57059902f7bc860.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 305, + 140, + 545, + 212 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 140, + 545, + 212 + ], + "spans": [ + { + "bbox": [ + 305, + 140, + 545, + 212 + ], + "type": "text", + "content": "At the first " + }, + { + "bbox": [ + 305, + 140, + 545, + 212 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 305, + 140, + 545, + 212 + ], + "type": "text", + "content": " communication rounds, the models are trained on only labeled clients. Thus, when " + }, + { + "bbox": [ + 305, + 140, + 545, + 212 + ], + "type": "inline_equation", + "content": "t \\in \\{1, \\dots, P\\}" + }, + { + "bbox": [ + 305, + 140, + 545, + 212 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 305, + 140, + 545, + 212 + ], + "type": "inline_equation", + "content": "\\sigma_t^\\mu(c) = 0" + }, + { + "bbox": [ + 305, + 140, + 545, + 212 + ], + "type": "text", + "content": " for all classes. After calculating the thresholds for all classes via Eq. 11, the central server will pass the thresholds to local clients. Unlabeled clients further obtain the fixed pseudo label training dataset by:" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 219, + 545, + 252 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 219, + 545, + 252 + ], + "spans": [ + { + "bbox": [ + 316, + 219, + 545, + 252 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\widetilde {\\mathcal {D}} _ {t + 1, \\mu} = \\left\\{\\left(X _ {i} ^ {\\mu}, \\hat {y} _ {i} ^ {\\mu}\\right) \\mid X _ {i} ^ {\\mu} \\in D _ {\\mu} \\right. \\tag {14} \\\\ \\wedge \\max (p _ {m} (y | \\theta_ {t} ^ {\\mu} (X _ {i} ^ {\\mu}))) > \\mathcal {T} _ {t + 1} (\\hat {y} _ {i} ^ {\\mu}) \\} _ {i = 1} ^ {N _ {\\mu}}. \\\\ \\end{array}", + "image_path": "463a3ac6581d10e422051395a2a9f3d8c4f521a793138adedb7ca5722b04f986.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 306, + 258, + 545, + 272 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 258, + 545, + 272 + ], + "spans": [ + { + "bbox": [ + 306, + 258, + 545, + 272 + ], + "type": "text", + "content": "3.4. Discovery of Unlabeled Data from Tail Classes" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 305, + 277, + 545, + 505 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 277, + 545, + 505 + ], + "spans": [ + { + "bbox": [ + 305, + 277, + 545, + 505 + ], + "type": "text", + "content": "In FSSL, heterogeneous data partitions will lead to class imbalanced distribution in local clients and there are rare or even no data from some classes in certain clients. For warm up stage in labeled clients, it is similar to long-tailed classification, so the problems in long-tailed classification will also exist in our pseudo labeling process, i.e., models tend to classify tail (rare) classes as head (common) classes [29, 35, 36]. Although adaptive pseudo labeling can lower the threshold of tail classes, much less data will be selected by PL since very few data from tailed classes will be classified as correct classes. To enhance the learning ability and discover unlabeled data from these classes, instead of directly dropping the data whose maximum confidences are low, namely, \"not informative\" unlabeled data, we propose to leverage information from them. For simplicity, we drop subscript " + }, + { + "bbox": [ + 305, + 277, + 545, + 505 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 305, + 277, + 545, + 505 + ], + "type": "text", + "content": " and default all representations occur in " + }, + { + "bbox": [ + 305, + 277, + 545, + 505 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 305, + 277, + 545, + 505 + ], + "type": "text", + "content": "th communication round in Sec. 3.4. Firstly, we define a mask function " + }, + { + "bbox": [ + 305, + 277, + 545, + 505 + ], + "type": "inline_equation", + "content": "\\mathcal{M}(\\cdot): \\mathcal{R}^C \\to \\mathcal{R}^C" + }, + { + "bbox": [ + 305, + 277, + 545, + 505 + ], + "type": "text", + "content": " for a probability simplex " + }, + { + "bbox": [ + 305, + 277, + 545, + 505 + ], + "type": "inline_equation", + "content": "p \\in \\mathcal{R}^C" + }, + { + "bbox": [ + 305, + 277, + 545, + 505 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 358, + 512, + 545, + 544 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 358, + 512, + 545, + 544 + ], + "spans": [ + { + "bbox": [ + 358, + 512, + 545, + 544 + ], + "type": "interline_equation", + "content": "\\mathcal {M} _ {i} (p) = \\left\\{ \\begin{array}{l l} p _ {i} & i \\neq \\arg \\max p \\\\ 0 & i = \\arg \\max p \\end{array} , \\right. \\tag {15}", + "image_path": "647cd6560a98d40b457b58da3cdd2be2a4beed17252cc9528ede153b85b546dc.jpg" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 305, + 552, + 545, + 598 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 552, + 545, + 598 + ], + "spans": [ + { + "bbox": [ + 305, + 552, + 545, + 598 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 552, + 545, + 598 + ], + "type": "inline_equation", + "content": "p_i" + }, + { + "bbox": [ + 305, + 552, + 545, + 598 + ], + "type": "text", + "content": " is the " + }, + { + "bbox": [ + 305, + 552, + 545, + 598 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 305, + 552, + 545, + 598 + ], + "type": "text", + "content": "th element in " + }, + { + "bbox": [ + 305, + 552, + 545, + 598 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 305, + 552, + 545, + 598 + ], + "type": "text", + "content": ". The effect of mask function " + }, + { + "bbox": [ + 305, + 552, + 545, + 598 + ], + "type": "inline_equation", + "content": "\\mathcal{M}(\\cdot)" + }, + { + "bbox": [ + 305, + 552, + 545, + 598 + ], + "type": "text", + "content": " is to set the maximum of a probability simplex to 0. Then, we turn to analyze the second largest confidence score via:" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 318, + 604, + 545, + 656 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 604, + 545, + 656 + ], + "spans": [ + { + "bbox": [ + 318, + 604, + 545, + 656 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} D _ {\\mu} ^ {t a i l} = \\left\\{\\left(X _ {i} ^ {\\mu}, \\hat {y} _ {i} ^ {\\mu^ {\\prime}}\\right) \\mid X _ {i} ^ {\\mu} \\in D _ {\\mu} \\right. \\\\ \\left. \\wedge \\max \\left(p _ {m} \\left(y \\mid \\theta_ {t} ^ {\\mu} \\left(X _ {i} ^ {\\mu}\\right)\\right)\\right) \\leq \\mathcal {T} _ {t} \\left(\\hat {y} _ {i} ^ {\\mu}\\right) \\wedge \\widetilde {p} _ {t} \\left(\\hat {y} _ {i} ^ {\\mu^ {\\prime}}\\right) < \\frac {\\beta}{C} \\right\\}, \\tag {16} \\\\ \\end{array}", + "image_path": "2fc252c7e95b355f26d7c4da35617411a6e3caba3802c3bf5751ca6a68eb0e2d.jpg" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 305, + 665, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 665, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 305, + 665, + 545, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 665, + 545, + 714 + ], + "type": "inline_equation", + "content": "\\hat{y}_i^{u'} = \\arg \\max \\mathcal{M}(p_m(y|\\theta_t^\\mu(X_i^\\mu)))" + }, + { + "bbox": [ + 305, + 665, + 545, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 665, + 545, + 714 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 305, + 665, + 545, + 714 + ], + "type": "text", + "content": " is a hyperparameter. In Eq. 15, the parameter " + }, + { + "bbox": [ + 305, + 665, + 545, + 714 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 305, + 665, + 545, + 714 + ], + "type": "text", + "content": " is to find the tail classes (\"not informative\" unlabeled data). For unlabeled data " + }, + { + "bbox": [ + 305, + 665, + 545, + 714 + ], + "type": "inline_equation", + "content": "X_i^\\mu" + }, + { + "bbox": [ + 305, + 665, + 545, + 714 + ], + "type": "text", + "content": ", if its largest confidence is smaller than threshold" + } + ] + } + ], + "index": 25 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "16296" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 289, + 145 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 289, + 145 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 289, + 145 + ], + "type": "text", + "content": "(not informative or confident, i.e., " + }, + { + "bbox": [ + 47, + 72, + 289, + 145 + ], + "type": "inline_equation", + "content": "\\max(p_m(y|\\theta_t^\\mu(X_i^\\mu))) \\leq \\mathcal{T}_t(\\hat{y}_i^\\mu)" + }, + { + "bbox": [ + 47, + 72, + 289, + 145 + ], + "type": "text", + "content": ") and the class of second largest confidence is tail class (i.e., " + }, + { + "bbox": [ + 47, + 72, + 289, + 145 + ], + "type": "inline_equation", + "content": "\\widetilde{p}_t(\\hat{y}_i^{\\mu \\prime}) < \\frac{\\beta}{C}" + }, + { + "bbox": [ + 47, + 72, + 289, + 145 + ], + "type": "text", + "content": "), we regard it as misclassified data and include it in training with the class of second largest confidence as its label. Finally, the new training dataset for unlabeled client " + }, + { + "bbox": [ + 47, + 72, + 289, + 145 + ], + "type": "inline_equation", + "content": "C_\\mu" + }, + { + "bbox": [ + 47, + 72, + 289, + 145 + ], + "type": "text", + "content": " is defined as" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 120, + 154, + 287, + 171 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 120, + 154, + 287, + 171 + ], + "spans": [ + { + "bbox": [ + 120, + 154, + 287, + 171 + ], + "type": "interline_equation", + "content": "D _ {\\mu} ^ {t r a i n} = D _ {\\mu} ^ {t a i l} \\cup \\widetilde {D} _ {\\mu}, \\tag {17}", + "image_path": "a89195a76e41496f895299908db694aad9736e63146e6cc34d0ab780d90b09e5.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 178, + 236, + 192 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 178, + 236, + 192 + ], + "spans": [ + { + "bbox": [ + 47, + 178, + 236, + 192 + ], + "type": "text", + "content": "and the final training loss in " + }, + { + "bbox": [ + 47, + 178, + 236, + 192 + ], + "type": "inline_equation", + "content": "C_{\\mu}" + }, + { + "bbox": [ + 47, + 178, + 236, + 192 + ], + "type": "text", + "content": " is rewritten as" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 200, + 287, + 241 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 200, + 287, + 241 + ], + "spans": [ + { + "bbox": [ + 55, + 200, + 287, + 241 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mu} = \\frac {1}{| D _ {\\mu} ^ {t r a i n} |} \\sum_ {(X ^ {\\mu}, \\hat {y} ^ {\\mu}) \\in D _ {\\mu} ^ {t r a i n}} H \\left(\\hat {y} ^ {\\mu}, p _ {m} \\left(y \\mid \\theta_ {t} ^ {\\mu} \\left(X ^ {\\mu}\\right)\\right)\\right). \\tag {18}", + "image_path": "41b5b4c5f86e275fd81ee2e27ff42dc3537754d9a1ff097441f6fc771f029114.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 242, + 203, + 255 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 242, + 203, + 255 + ], + "spans": [ + { + "bbox": [ + 47, + 242, + 203, + 255 + ], + "type": "text", + "content": "The calculation of " + }, + { + "bbox": [ + 47, + 242, + 203, + 255 + ], + "type": "inline_equation", + "content": "\\sigma_t^u (c)" + }, + { + "bbox": [ + 47, + 242, + 203, + 255 + ], + "type": "text", + "content": " is revised as:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 94, + 263, + 287, + 291 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 94, + 263, + 287, + 291 + ], + "spans": [ + { + "bbox": [ + 94, + 263, + 287, + 291 + ], + "type": "interline_equation", + "content": "\\sigma_ {t} ^ {\\mu} (c) = \\sum_ {(X ^ {\\mu}, \\hat {y} ^ {\\mu}) \\in D _ {\\mu} ^ {t r a i n}} \\mathbf {1} (\\hat {y} ^ {\\mu} = c). \\tag {19}", + "image_path": "37b3e5e2c18ae9d82af89a8b2a70dec1596a43646d305b4851ba3d40d47d068b.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 299, + 200, + 312 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 299, + 200, + 312 + ], + "spans": [ + { + "bbox": [ + 47, + 299, + 200, + 312 + ], + "type": "text", + "content": "3.5. Aggregation of local models" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 317, + 287, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 317, + 287, + 354 + ], + "spans": [ + { + "bbox": [ + 47, + 317, + 287, + 354 + ], + "type": "text", + "content": "Since not all data in local clients are included into training, we propose to use only data included in training for model aggregation. The scale factor " + }, + { + "bbox": [ + 47, + 317, + 287, + 354 + ], + "type": "inline_equation", + "content": "w_{t}^{i}" + }, + { + "bbox": [ + 47, + 317, + 287, + 354 + ], + "type": "text", + "content": " is computed by" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 361, + 287, + 407 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 361, + 287, + 407 + ], + "spans": [ + { + "bbox": [ + 62, + 361, + 287, + 407 + ], + "type": "interline_equation", + "content": "w _ {t} ^ {i} = \\left\\{ \\begin{array}{l l} \\frac {\\left| D _ {i} \\right|}{\\left| D _ {t} ^ {t r a i n} \\right|} & \\text {i f} i \\in \\{1, \\dots , m \\} \\\\ \\frac {\\left| D _ {t , i} ^ {t r a i n} \\right|}{\\left| D _ {t} ^ {t r a i n} \\right|} & \\text {i f} i \\in \\{m + 1, \\dots , m + n \\}, \\end{array} \\right. \\tag {20}", + "image_path": "0e54aba48f19598681f44c5beea13c4fefc28873997fc00350baca903b24e0a9.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 415, + 287, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 415, + 287, + 453 + ], + "spans": [ + { + "bbox": [ + 47, + 415, + 287, + 453 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 415, + 287, + 453 + ], + "type": "inline_equation", + "content": "|D_t^{train}| = \\sum_{\\ell=1}^m |D_\\ell| + \\sum_{\\mu=m+1}^{m+n} |D_{t,\\mu}^{train}|" + }, + { + "bbox": [ + 47, + 415, + 287, + 453 + ], + "type": "text", + "content": ". Then at " + }, + { + "bbox": [ + 47, + 415, + 287, + 453 + ], + "type": "inline_equation", + "content": "(t+1)" + }, + { + "bbox": [ + 47, + 415, + 287, + 453 + ], + "type": "text", + "content": "th synchronization round, the global model " + }, + { + "bbox": [ + 47, + 415, + 287, + 453 + ], + "type": "inline_equation", + "content": "\\theta_{t+1}^G" + }, + { + "bbox": [ + 47, + 415, + 287, + 453 + ], + "type": "text", + "content": " is computed by" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 52, + 460, + 287, + 491 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 52, + 460, + 287, + 491 + ], + "spans": [ + { + "bbox": [ + 52, + 460, + 287, + 491 + ], + "type": "interline_equation", + "content": "\\theta_ {t + 1} ^ {G} = \\left\\{ \\begin{array}{l l} \\sum_ {i = 1} ^ {m + n} w _ {t} ^ {i} \\theta_ {t} ^ {i} & t \\% s \\neq 0 \\\\ \\alpha_ {2} \\theta_ {t + 1 - s} ^ {G} + (1 - \\alpha_ {2}) \\sum_ {i = 1} ^ {m + n} w _ {t} ^ {i} \\theta_ {t} ^ {i} & t \\% s = 0, \\end{array} \\right. \\tag{21}", + "image_path": "d86653f960c793e637e4f382229591eba75ee15aefb64188d19d93e837efb21e.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 50, + 497, + 173, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 50, + 497, + 173, + 509 + ], + "spans": [ + { + "bbox": [ + 50, + 497, + 173, + 509 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 50, + 497, + 173, + 509 + ], + "type": "inline_equation", + "content": "\\alpha_{2}" + }, + { + "bbox": [ + 50, + 497, + 173, + 509 + ], + "type": "text", + "content": " is a hyperparameter." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 47, + 520, + 128, + 533 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 520, + 128, + 533 + ], + "spans": [ + { + "bbox": [ + 47, + 520, + 128, + 533 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 47, + 539, + 163, + 552 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 539, + 163, + 552 + ], + "spans": [ + { + "bbox": [ + 47, + 539, + 163, + 552 + ], + "type": "text", + "content": "4.1. Experimental Setup" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 47, + 558, + 287, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 558, + 287, + 641 + ], + "spans": [ + { + "bbox": [ + 47, + 558, + 287, + 641 + ], + "type": "text", + "content": "Benchmark Datasets To evaluate the effectiveness of our proposed method, we conduct extensive experiments on four image classification datasets, i.e., SVHN, CIFAR-10, CIFAR-100, Fashion MNIST and one medical image classification dataset: ISIC 2018 (Skin Lesion Analysis Towards Melanoma Detection). Dataset splitting and pre-processing are provided in the supplementary material." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 47, + 642, + 287, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 642, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 47, + 642, + 287, + 689 + ], + "type": "text", + "content": "FSSL Setting Simulation Following [14], training datasets contain 10 clients: one labeled and nine unlabeled. We use a Dirichlet distribution " + }, + { + "bbox": [ + 47, + 642, + 287, + 689 + ], + "type": "inline_equation", + "content": "\\text{Dir}(\\gamma)" + }, + { + "bbox": [ + 47, + 642, + 287, + 689 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 47, + 642, + 287, + 689 + ], + "type": "inline_equation", + "content": "\\gamma = 0.8" + }, + { + "bbox": [ + 47, + 642, + 287, + 689 + ], + "type": "text", + "content": " for five datasets [14] to generate Non-IID data partition in clients." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "content": "Implementation Details We utilize the SGD optimizer with a momentum of 0.9, and implement our method with" + } + ] + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 307, + 72, + 429, + 190 + ], + "blocks": [ + { + "bbox": [ + 307, + 72, + 429, + 190 + ], + "lines": [ + { + "bbox": [ + 307, + 72, + 429, + 190 + ], + "spans": [ + { + "bbox": [ + 307, + 72, + 429, + 190 + ], + "type": "image", + "image_path": "6ba1612a951d4ca351d73178f6854ebf22e02b0228ccf57159f3b72aa4092cb1.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 201, + 545, + 269 + ], + "lines": [ + { + "bbox": [ + 305, + 201, + 545, + 269 + ], + "spans": [ + { + "bbox": [ + 305, + 201, + 545, + 269 + ], + "type": "text", + "content": "Figure 2. Test accuracy curves in local training of SVHN dataset w/ and w/o residual weight connection. ResNet18 (a) and ViT (b) are adopted as the backbones. W/ res-weight* indicates we only show test accuracy on epochs (communication rounds, since local training epoch for labeled client is 1) with skip weight connection. Best viewed electronically." + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 429, + 72, + 542, + 190 + ], + "blocks": [ + { + "bbox": [ + 429, + 72, + 542, + 190 + ], + "lines": [ + { + "bbox": [ + 429, + 72, + 542, + 190 + ], + "spans": [ + { + "bbox": [ + 429, + 72, + 542, + 190 + ], + "type": "image", + "image_path": "115cb75ad70581d87bb29fd31b0f256a92835927fed7dd12377d5a66dc5f13ba.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "bbox": [ + 305, + 277, + 545, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 277, + 545, + 361 + ], + "spans": [ + { + "bbox": [ + 305, + 277, + 545, + 361 + ], + "type": "text", + "content": "PyTorch. We adopt ResNet18 [6] from PyTorch for all datasets. For fair comparison, we use the same network architecture and training protocol, including the optimizer, data preprocessing, etc. across all FSSL methods. The local training epoch is set to 11 for labeled clients and 1 for unlabeled clients. More implementation details are provided in supplementary material." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 305, + 361, + 546, + 566 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 361, + 546, + 566 + ], + "spans": [ + { + "bbox": [ + 305, + 361, + 546, + 566 + ], + "type": "text", + "content": "Comparisons We compare our method against state-of-the-art FSSL methods, including RSCFed [14], FedIRM [18], and Fed-Consist [31]. Since FedIRM [18] and Fed-Consist [31] are not desgined for dealing with Non-IID FSSL, following [14], we enlarge the weight of labeled client to about " + }, + { + "bbox": [ + 305, + 361, + 546, + 566 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 305, + 361, + 546, + 566 + ], + "type": "text", + "content": " and other nine unlabeled clients share the remaining " + }, + { + "bbox": [ + 305, + 361, + 546, + 566 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 305, + 361, + 546, + 566 + ], + "type": "text", + "content": " weight in each FSSL synchronization round, when implementing these methods. We also compare our network against FedAvg [19] trained with all clients as an upper bound and trained with only one labeled clients as the lower bound. Besides, to show the effectiveness of our residual weight connection, we report the result of FedAvg [19] trained with only one labeled clients using residual weight connection. Since Fed-Consist [31] utilizes traditional batch-based pseudo labeling method, we report the result of Fed-Consist [31] using our proposed fixed pseudo labeling without enlarging the weight of labeled client." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 306, + 574, + 493, + 588 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 574, + 493, + 588 + ], + "spans": [ + { + "bbox": [ + 306, + 574, + 493, + 588 + ], + "type": "text", + "content": "4.2. Comparisons with State-of-the-arts" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 305, + 594, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 594, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 594, + 545, + 713 + ], + "type": "text", + "content": "Comparison Results. Table 1 reports the quantitative results of our CBAFed and other state-of-the-art methods on five benchmarks. We can observe that our proposed CBAFed achieves superior performances over all competitors on the five benchmark datasets. For some easy tasks (e.g., SVHN and Fashion-MNIST), the performance of our CBAFed is approaching upper bound obtained by FedAVG [19] and surpasses all state-of-the-arts by a large margin, showing a strong power of fixed pseudo labeling and class balanced threshold. Note that if replacing the traditional" + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16297" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 50, + 112, + 547, + 211 + ], + "blocks": [ + { + "bbox": [ + 47, + 70, + 545, + 104 + ], + "lines": [ + { + "bbox": [ + 47, + 70, + 545, + 104 + ], + "spans": [ + { + "bbox": [ + 47, + 70, + 545, + 104 + ], + "type": "text", + "content": "Table 1. Results on SVHN, CIFAR-10/100, Fashion MNIST and ISIC 2018 datasets under heterogeneous data partition with ResNet18. FedAVG " + }, + { + "bbox": [ + 47, + 70, + 545, + 104 + ], + "type": "inline_equation", + "content": "^+" + }, + { + "bbox": [ + 47, + 70, + 545, + 104 + ], + "type": "text", + "content": " means FedAvg [19] trained with all one labeled clients using our residual weight connection. Fed-consist " + }, + { + "bbox": [ + 47, + 70, + 545, + 104 + ], + "type": "inline_equation", + "content": "^+" + }, + { + "bbox": [ + 47, + 70, + 545, + 104 + ], + "type": "text", + "content": " means Fed-Consist [31] using our proposed fixed pseudo labeling without enlarging the weight of labeled client." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 50, + 112, + 547, + 211 + ], + "lines": [ + { + "bbox": [ + 50, + 112, + 547, + 211 + ], + "spans": [ + { + "bbox": [ + 50, + 112, + 547, + 211 + ], + "type": "table", + "html": "
Labeling StrategyMethodClient Num.Dataset
labeledunlabeledSVHNCIFAR10CIFAR100Fashion-MNISTISIC 2018
Fully supervisedFedAvg [19](upper-bound)10091.8380.8951.3890.1481.32
FedAvg [19](lower-bound)1067.7154.6620.4974.8765.13
FedAvg+ [19]1076.9858.2124.8478.2666.69
Semi supervisedFedIRM [18]1969.2252.8420.2076.8364.85
Fed-Consist [31]1970.5654.2321.8176.5765.20
Fed-Consist+ [31]1986.5756.3523.2578.3565.50
RSCFed [14]1976.7457.0728.4678.4067.21
CBAFed(ours)1988.0767.0830.1885.4968.29
", + "image_path": "4821af0147cba5088d02e38a90651647115ead6d80cfee580e48bec9ea0fd922.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 231, + 287, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 231, + 287, + 277 + ], + "spans": [ + { + "bbox": [ + 46, + 231, + 287, + 277 + ], + "type": "text", + "content": "batch-based pseudo labeling by our fixed pseudo labeling in Fed-Consist [31], the performance is largely increased. It shows that fixed pseudo labeling method is well suited for Non-IID FSSL." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 280, + 287, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 280, + 287, + 483 + ], + "spans": [ + { + "bbox": [ + 46, + 280, + 287, + 483 + ], + "type": "text", + "content": "Trained on one labeled client, FedAVG [19] w/ residual weight connection achieves much better performance compared with that w/o residual weight connection. Fig. 2 (a) shows the test accuracy curves during training. Due to the imbalanced training data distribution in labeled client, the test accuracy is unstable during training for FedAVG (w/o res-weight in the figure). But, if training w/ our residual weight connection, the test accuracy curve is much more stable and the performance is also enhanced (w/ res-weight and w/ res-weight*). To sum up, residual weight connection can improve the robustness of the models when the training amount is small and the training distribution is imbalanced. Since long tailed vision recognition task is similar to local training in labeled client, it is a potential direction to use our residual weight connection in long tailed vision. Besides, we also conduct experiments on partially labeled clients, results are shown in the supplementary material." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 484, + 287, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 484, + 287, + 640 + ], + "spans": [ + { + "bbox": [ + 46, + 484, + 287, + 640 + ], + "type": "text", + "content": "ViT Backbone. Vision Transformer(ViT) has shown to be more robust to heterogeneous data and distribution shifts, and has been demonstrated to converge faster with better optimum in Federated learning [23]. To further study the effect of our method, we conduct experiments on SVHN dataset using ViT-Tiny as backbone for all competitors. The implementation details are provided in the supplementary material. Table 2 shows the comparison results. Our method can outperform all other methods. Similarly, trained on one labeled client, FedAVG [19] w/ residual weight connection surpasses the one w/o residual weight connection, meaning that our residual weight connection is also effective on ViT, not only CNNs, as shown in Fig. 2 (b)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 642, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 642, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 642, + 287, + 713 + ], + "type": "text", + "content": "Two Labeled Clients. To better understand the effectiveness of our CBAFed and study how to use residual weight connection when the number of labeled clients is more than one, following [14, 18], we conduct experiments on CIFAR-10 by dividing the whole training data into 10 clients, with two labeled clients and eight unlabeled clients. Results are" + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 320, + 271, + 533, + 357 + ], + "blocks": [ + { + "bbox": [ + 304, + 229, + 545, + 262 + ], + "lines": [ + { + "bbox": [ + 304, + 229, + 545, + 262 + ], + "spans": [ + { + "bbox": [ + 304, + 229, + 545, + 262 + ], + "type": "text", + "content": "Table 2. Comparison of our method against RSCFed [14], FedConsist [31] and FedAVG [19] in SVHN dataset on ViT [5] as the backbone, with one labeled and nine unlabeled clients." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 320, + 271, + 533, + 357 + ], + "lines": [ + { + "bbox": [ + 320, + 271, + 533, + 357 + ], + "spans": [ + { + "bbox": [ + 320, + 271, + 533, + 357 + ], + "type": "table", + "html": "
MethodClient Num.Accuracy
labeledunlabeled
FedAVG [19](upper bound)10096.81
FedAVG [19](lower bound)1081.68
FedAVG+ [19]1088.93
FedIRM [18]1979.44
Fed-Consist [31]1985.91
Fed-Consist+ [31]1993.21
RSCFed [14]1989.43
CBAFed(ours)1995.09
", + "image_path": "42f2ce88d940c39e81c63f133180f8ddef449260a7c44db3a3e8ea9ac9c06f48.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 309, + 407, + 545, + 502 + ], + "blocks": [ + { + "bbox": [ + 304, + 364, + 547, + 397 + ], + "lines": [ + { + "bbox": [ + 304, + 364, + 547, + 397 + ], + "spans": [ + { + "bbox": [ + 304, + 364, + 547, + 397 + ], + "type": "text", + "content": "Table 3. Comparison of our method against RSCFed [14], FedConsist [31], FedIRM [18] and FedAVG [19] with the number of labeled and unlabeled client set to 2 and 8." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 309, + 407, + 545, + 502 + ], + "lines": [ + { + "bbox": [ + 309, + 407, + 545, + 502 + ], + "spans": [ + { + "bbox": [ + 309, + 407, + 545, + 502 + ], + "type": "table", + "html": "
MethodClient Num.Accuracy
labeledunlabeled
FedAVG [19](upper bound)10080.89
FedAVG [19](lower bound)2061.85
FedAVG+ [19]2066.55
FedIRM [18]2862.62
Fed-Consist [31]2861.67
Fed-Consist+ [31]2868.04
RSCFed [14]2864.25
CBAFed(ours)2872.01
", + "image_path": "6ac072582755d096deaa5a9114ed69a97aff9e1b6d3f7d3e7f441f98958e1bab.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 520, + 545, + 603 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 520, + 545, + 603 + ], + "spans": [ + { + "bbox": [ + 304, + 520, + 545, + 603 + ], + "type": "text", + "content": "reported in Table 3. Our method has a better performance compared with all state-of-the-art FSSL methods. Adding our residual weight connection and fixed pseudo labeling lead to much performance gain. In the supplementary material, we also discuss different training strategies used for labeled clients w.r.t. number of local epochs and the usage of residual weight connection." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 604, + 545, + 663 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 604, + 545, + 663 + ], + "spans": [ + { + "bbox": [ + 304, + 604, + 545, + 663 + ], + "type": "text", + "content": "Communication Cost. Since the statistics of local training data can be nearly neglected (very low compared with model's parameter), our method does not add any burden in communication cost, which is only " + }, + { + "bbox": [ + 304, + 604, + 545, + 663 + ], + "type": "inline_equation", + "content": "66\\%" + }, + { + "bbox": [ + 304, + 604, + 545, + 663 + ], + "type": "text", + "content": " of the state-of-the-art method RSCFed [14]." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 671, + 538, + 685 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 671, + 538, + 685 + ], + "spans": [ + { + "bbox": [ + 306, + 671, + 538, + 685 + ], + "type": "text", + "content": "4.3. Study of Pseudo Labeling Strategies in FSSL" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 689, + 545, + 713 + ], + "type": "text", + "content": "In this section, we discuss how to perform pseudo labeling in FSSL and why traditional batch-based pseudo la" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16298" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 72, + 165, + 180 + ], + "blocks": [ + { + "bbox": [ + 50, + 72, + 165, + 180 + ], + "lines": [ + { + "bbox": [ + 50, + 72, + 165, + 180 + ], + "spans": [ + { + "bbox": [ + 50, + 72, + 165, + 180 + ], + "type": "image", + "image_path": "be35e86cc4f6b643dbdc12187d6d6b5889ddc0ba345357c0159fee5347d44c7d.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 192, + 287, + 248 + ], + "lines": [ + { + "bbox": [ + 46, + 192, + 287, + 248 + ], + "spans": [ + { + "bbox": [ + 46, + 192, + 287, + 248 + ], + "type": "text", + "content": "Figure 3. Left: Test accuracy of all 5 strategies after every communication round. Note that the test accuracy of communication round 0 is the test accuracy of model trained on labeled client. Right: Accuracy of pseudo labels in local training epoch of one randomly selected unlabeled client. Best viewed electronically." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 167, + 72, + 284, + 181 + ], + "blocks": [ + { + "bbox": [ + 167, + 72, + 284, + 181 + ], + "lines": [ + { + "bbox": [ + 167, + 72, + 284, + 181 + ], + "spans": [ + { + "bbox": [ + 167, + 72, + 284, + 181 + ], + "type": "image", + "image_path": "f9c462a8aac5664fcfbd97999ec2daa599ebc85feffbf2427b5a2650514a3d61.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 257, + 289, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 257, + 289, + 628 + ], + "spans": [ + { + "bbox": [ + 46, + 257, + 289, + 628 + ], + "type": "text", + "content": "beling fails. We conduct a simple experiment on SVHN dataset [20] using ResNet18 [6]. The experimental setting is the same as illustrated in Sec. 4.1 except that we only use pseudo labeling without other proposed modules. We consider five training strategies: (1) Batched pseudo labeling (1 local training epoch). (2) Batched pseudo labeling (5 local training epochs). (3) Fixed pseudo labeling (1 local training epoch). (4) Fixed pseudo labeling (5 local training epochs), and we do not update the fixed training dataset in all local training epochs after pseudo labeling. (5) Fixed pseudo labeling (5 local training epochs), and we update the fixed training dataset at the beginning of every training epoch. For labeled client, the local training epoch is set to 1 for all 5 strategies. The total communication round is set to 200 for all strategies. The test accuracy after communication round is shown in the left figure of Fig. 3. We can observe the following: i), Batch-based pseudo labeling performs poorly in FSSL. It is known that CNNs often do not work well on out-of-distribution data [23]. Thus, a few update can result in catastrophic forgetting on Non-IID data partitions. The right figure of Fig. 3 shows the predicted accuracy of communication round one in five local training epochs of strategy (2) and (5). ii), Our fixed pseudo labeling method is capable of dealing with catastrophic forgetting. Strategy (3) performs the best and strategy (5) performs much worse than strategy (3) and (4), indicating that one local training epoch is optimum. iii) Comparing strategy (3) with (4), the former has better performance but the latter has faster convergence speed. It means more local training epochs will speed up convergence but may degrade performance, similar phenomenon is also found in [23]." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 635, + 142, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 635, + 142, + 647 + ], + "spans": [ + { + "bbox": [ + 47, + 635, + 142, + 647 + ], + "type": "text", + "content": "4.4. Ablation Study" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 654, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 654, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 654, + 287, + 713 + ], + "type": "text", + "content": "We conduct ablation study to demonstrate the effectiveness of the main components of our method: residual weight connection, class balanced pseudo labeling, and tail class data discovery. Table 4 shows the results on three datasets: CIFAR-10/100 and Fashion-MNIST. It can be" + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 319, + 112, + 533, + 203 + ], + "blocks": [ + { + "bbox": [ + 305, + 70, + 545, + 104 + ], + "lines": [ + { + "bbox": [ + 305, + 70, + 545, + 104 + ], + "spans": [ + { + "bbox": [ + 305, + 70, + 545, + 104 + ], + "type": "text", + "content": "Table 4. Ablation Study of CBAFed in CIFAR-10/100 and Fashion MNIST Datasets. Fixed PL: fixed pseudo labeling, CBA: class balanced adaptive pseudo labeling, DD: tail class data discovery." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 319, + 112, + 533, + 203 + ], + "lines": [ + { + "bbox": [ + 319, + 112, + 533, + 203 + ], + "spans": [ + { + "bbox": [ + 319, + 112, + 533, + 203 + ], + "type": "table", + "html": "
DatasetFixed PLCBADDRes-WeightAccuracy
CIFAR-10✓59.16
✓✓64.29
✓✓✓65.15
✓✓✓✓67.08
CIFAR-100✓27.64
✓✓29.41
✓✓✓29.86
✓✓✓✓30.18
Fashion-MNIST✓79.99
✓✓80.87
✓✓✓84.37
✓✓✓✓85.49
", + "image_path": "c94f903c635d2cb8a4e4047fbbe73981580e998910daa0b4431af2aca24c072c.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 319, + 211, + 533, + 287 + ], + "blocks": [ + { + "bbox": [ + 319, + 211, + 533, + 287 + ], + "lines": [ + { + "bbox": [ + 319, + 211, + 533, + 287 + ], + "spans": [ + { + "bbox": [ + 319, + 211, + 533, + 287 + ], + "type": "image", + "image_path": "5ea2888668a874d4a89787ac9194eb9374b2690d579f8ba6a281056f3067b40d.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 298, + 545, + 332 + ], + "lines": [ + { + "bbox": [ + 305, + 298, + 545, + 332 + ], + "spans": [ + { + "bbox": [ + 305, + 298, + 545, + 332 + ], + "type": "text", + "content": "Figure 4. Performance changes on Fashion-MNIST by varying (a) threshold base " + }, + { + "bbox": [ + 305, + 298, + 545, + 332 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 305, + 298, + 545, + 332 + ], + "type": "text", + "content": " and (b) upper bound threshold " + }, + { + "bbox": [ + 305, + 298, + 545, + 332 + ], + "type": "inline_equation", + "content": "\\tau_{h}" + }, + { + "bbox": [ + 305, + 298, + 545, + 332 + ], + "type": "text", + "content": ", and (c) parameter for selecting tail class data " + }, + { + "bbox": [ + 305, + 298, + 545, + 332 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 305, + 298, + 545, + 332 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 348, + 545, + 420 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 348, + 545, + 420 + ], + "spans": [ + { + "bbox": [ + 305, + 348, + 545, + 420 + ], + "type": "text", + "content": "seen that all three modules are important for boosting performance. Then, we aim at discussing hyper-parameters of threshold base " + }, + { + "bbox": [ + 305, + 348, + 545, + 420 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 305, + 348, + 545, + 420 + ], + "type": "text", + "content": ", upper bound threshold " + }, + { + "bbox": [ + 305, + 348, + 545, + 420 + ], + "type": "inline_equation", + "content": "\\tau_h" + }, + { + "bbox": [ + 305, + 348, + 545, + 420 + ], + "type": "text", + "content": ", and parameters for selecting tail class data " + }, + { + "bbox": [ + 305, + 348, + 545, + 420 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 305, + 348, + 545, + 420 + ], + "type": "text", + "content": ". Experiments are conducted on Fashion-MNIST. As shown in Fig. 4, performances are not sensitive within certain ranges." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 432, + 379, + 445 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 432, + 379, + 445 + ], + "spans": [ + { + "bbox": [ + 306, + 432, + 379, + 445 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 453, + 545, + 571 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 453, + 545, + 571 + ], + "spans": [ + { + "bbox": [ + 305, + 453, + 545, + 571 + ], + "type": "text", + "content": "We present Class Balanced Adaptive pseudo labeling (CBAFed) for Federated Semi-Supervised Learning (FSSL). In CBAFed, a fixed pseudo labeling strategy is proposed to handle the catastrophic forgetting problem. To deal with the Non-IID setting of FSSL, we propose a class balanced adaptive thresholds selection method to choose better pseudo labels. Furthermore, a residual weight connection method is designed to make the model reach better optimum. We evaluate CBAFed on five datasets, whose performances show the superiority of our method." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 573, + 545, + 620 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 573, + 545, + 620 + ], + "spans": [ + { + "bbox": [ + 305, + 573, + 545, + 620 + ], + "type": "text", + "content": "Limitations. Like other FSSL methods, a good model in warm-up cannot be guaranteed if the number of data in labeled clients are extremely few, so the final global model may not perform well." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 633, + 416, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 633, + 416, + 647 + ], + "spans": [ + { + "bbox": [ + 306, + 633, + 416, + 647 + ], + "type": "text", + "content": "6. Acknowledgement" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 654, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 654, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 654, + 545, + 713 + ], + "type": "text", + "content": "This work was supported by the National Natural Science Foundation of China (Grant No. 62101191), Shanghai Natural Science Foundation (Grant No. 21ZR1420800), and the Science and Technology Commission of Shanghai Municipality (Grant No. 22DZ2229004)." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "16299" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 91, + 288, + 715 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "type": "text", + "content": "[1] David Berthelot, Nicholas Carlini, Ekin D Cubuk, Alex Kurakin, Kihyuk Sohn, Han Zhang, and Colin Raffel. Remixmatch: Semi-supervised learning with distribution alignment and augmentation anchoring. arXiv preprint arXiv:1911.09785, 2019. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 146, + 288, + 191 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 146, + 288, + 191 + ], + "spans": [ + { + "bbox": [ + 53, + 146, + 288, + 191 + ], + "type": "text", + "content": "[2] David Berthelot, Nicholas Carlini, Ian Goodfellow, Nicolas Papernot, Avital Oliver, and Colin A Raffel. Mixmatch: A holistic approach to semi-supervised learning. Advances in neural information processing systems, 32, 2019. 2, 4" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 191, + 288, + 246 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 191, + 288, + 246 + ], + "spans": [ + { + "bbox": [ + 53, + 191, + 288, + 246 + ], + "type": "text", + "content": "[3] Xiaokang Chen, Yuhui Yuan, Gang Zeng, and Jingdong Wang. Semi-supervised semantic segmentation with cross pseudo supervision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2613-2622, 2021. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 246, + 288, + 312 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 246, + 288, + 312 + ], + "spans": [ + { + "bbox": [ + 53, + 246, + 288, + 312 + ], + "type": "text", + "content": "[4] Wenhui Cui, Yanlin Liu, Yuxing Li, Menghao Guo, Yiming Li, Xiuli Li, Tianle Wang, Xiangzhu Zeng, and Chuyang Ye. Semi-supervised brain lesion segmentation with an adapted mean teacher model. In International Conference on Information Processing in Medical Imaging, pages 554-565. Springer, 2019. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 313, + 288, + 378 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 313, + 288, + 378 + ], + "spans": [ + { + "bbox": [ + 53, + 313, + 288, + 378 + ], + "type": "text", + "content": "[5] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 380, + 288, + 423 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 380, + 288, + 423 + ], + "spans": [ + { + "bbox": [ + 53, + 380, + 288, + 423 + ], + "type": "text", + "content": "[6] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 6, 8" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 424, + 288, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 424, + 288, + 468 + ], + "spans": [ + { + "bbox": [ + 53, + 424, + 288, + 468 + ], + "type": "text", + "content": "[7] Wenke Huang, Mang Ye, and Bo Du. Learn from others and be yourself in heterogeneous federated learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10143-10153, 2022. 4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 468, + 288, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 468, + 288, + 502 + ], + "spans": [ + { + "bbox": [ + 53, + 468, + 288, + 502 + ], + "type": "text", + "content": "[8] Wonyong Jeong, Jaehong Yoon, Eunho Yang, and Sung Ju Hwang. Federated semi-supervised learning with inter-client consistency & disjoint learning. In Proc. ICLR, 2021. 1, 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 502, + 288, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 502, + 288, + 557 + ], + "spans": [ + { + "bbox": [ + 53, + 502, + 288, + 557 + ], + "type": "text", + "content": "[9] Sai Praneeth Karimireddy, Satyen Kale, Mehryar Mohri, Sashank Reddi, Sebastian Stich, and Ananda Theertha Suresh. Scaffold: Stochastic controlled averaging for federated learning. In International Conference on Machine Learning, pages 5132-5143. PMLR, 2020. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 557, + 288, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 557, + 288, + 602 + ], + "spans": [ + { + "bbox": [ + 47, + 557, + 288, + 602 + ], + "type": "text", + "content": "[10] Dong-Hyun Lee et al. Pseudo-label: The simple and efficient semi-supervised learning method for deep neural networks. In Workshop on challenges in representation learning, ICML, volume 3, page 896, 2013. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 602, + 288, + 647 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 602, + 288, + 647 + ], + "spans": [ + { + "bbox": [ + 47, + 602, + 288, + 647 + ], + "type": "text", + "content": "[11] Qinbin Li, Bingsheng He, and Dawn Song. Model-contrastive federated learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10713–10722, 2021. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 47, + 647, + 288, + 691 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 647, + 288, + 691 + ], + "spans": [ + { + "bbox": [ + 47, + 647, + 288, + 691 + ], + "type": "text", + "content": "[12] Tian Li, Anit Kumar Sahu, Manzil Zaheer, Maziar Sanjabi, Ameet Talwalkar, and Virginia Smith. Federated optimization in heterogeneous networks. Proceedings of Machine Learning and Systems, 2:429-450, 2020. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 47, + 691, + 288, + 715 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 691, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 47, + 691, + 288, + 715 + ], + "type": "text", + "content": "[13] Xiaomeng Li, Lequan Yu, Hao Chen, Chi-Wing Fu, and Pheng-Ann Heng. Transformation consistent self" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 715 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 106 + ], + "type": "text", + "content": "ensembling model for semi-supervised medical image segmentation. IEEE Trans. Neural Networks and Learning Systems, 2020. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 107, + 545, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 107, + 545, + 140 + ], + "spans": [ + { + "bbox": [ + 307, + 107, + 545, + 140 + ], + "type": "text", + "content": "[14] Xiaoxiao Liang, Yiqun Lin, Huazhu Fu, Lei Zhu, and Xiaomeng Li. Rscfed: Random sampling consensus federated semi-supervised learning. In Proc. CVPR, 2022. 1, 3, 6, 7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 141, + 545, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 141, + 545, + 185 + ], + "spans": [ + { + "bbox": [ + 307, + 141, + 545, + 185 + ], + "type": "text", + "content": "[15] Haowen Lin, Jian Lou, Li Xiong, and Cyrus Shahabi. Semifed: Semi-supervised federated learning with consistency and pseudo-labeling. arXiv preprint arXiv:2108.09412, 2021. 1, 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 186, + 545, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 186, + 545, + 252 + ], + "spans": [ + { + "bbox": [ + 307, + 186, + 545, + 252 + ], + "type": "text", + "content": "[16] Fengbei Liu, Yu Tian, Yuanhong Chen, Yuyuan Liu, Vasileios Belagiannis, and Gustavo Carneiro. Acpl: Anticurriculum pseudo-labelling for semi-supervised medical image classification. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20697-20706, 2022. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 253, + 545, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 253, + 545, + 319 + ], + "spans": [ + { + "bbox": [ + 307, + 253, + 545, + 319 + ], + "type": "text", + "content": "[17] Quande Liu, Cheng Chen, Jing Qin, Qi Dou, and Pheng-Ann Heng. Feddg: Federated domain generalization on medical image segmentation via episodic learning in continuous frequency space. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1013–1023, 2021. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 320, + 545, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 320, + 545, + 365 + ], + "spans": [ + { + "bbox": [ + 307, + 320, + 545, + 365 + ], + "type": "text", + "content": "[18] Quande Liu, Hongzheng Yang, Qi Dou, and Pheng-Ann Heng. Federated semi-supervised medical image classification via inter-client relation matching. In Proc. MICCAI, 2021. 1, 2, 3, 6, 7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 365, + 545, + 421 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 365, + 545, + 421 + ], + "spans": [ + { + "bbox": [ + 307, + 365, + 545, + 421 + ], + "type": "text", + "content": "[19] Brendan McMahan, Eider Moore, Daniel Ramage, Seth Hampson, and Blaise Aguera y Arcas. Communication-efficient learning of deep networks from decentralized data. In Artificial intelligence and statistics, pages 1273–1282. PMLR, 2017. 2, 4, 6, 7" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 421, + 545, + 455 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 421, + 545, + 455 + ], + "spans": [ + { + "bbox": [ + 307, + 421, + 545, + 455 + ], + "type": "text", + "content": "[20] Yuval Netzer, Tao Wang, Adam Coates, Alessandro Bissacco, Bo Wu, and Andrew Y Ng. Reading digits in natural images with unsupervised feature learning. 2011. 8" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 456, + 545, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 456, + 545, + 510 + ], + "spans": [ + { + "bbox": [ + 307, + 456, + 545, + 510 + ], + "type": "text", + "content": "[21] Krishna Pillutla, Kshitiz Malik, Abdel-Rahman Mohamed, Mike Rabbat, Maziar Sanjabi, and Lin Xiao. Federated learning with partial model personalization. In International Conference on Machine Learning, pages 17716-17758. PMLR, 2022. 4" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 511, + 545, + 545 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 511, + 545, + 545 + ], + "spans": [ + { + "bbox": [ + 307, + 511, + 545, + 545 + ], + "type": "text", + "content": "[22] Boris T Polyak and Anatoli B Juditsky. Acceleration of stochastic approximation by averaging. SIAM journal on control and optimization, 30(4):838-855, 1992. 4" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 546, + 545, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 546, + 545, + 612 + ], + "spans": [ + { + "bbox": [ + 307, + 546, + 545, + 612 + ], + "type": "text", + "content": "[23] Liangqiong Qu, Yuyin Zhou, Paul Pu Liang, Yingda Xia, Feifei Wang, Ehsan Adeli, Li Fei-Fei, and Daniel Rubin. Rethinking architecture design for tackling data heterogeneity in federated learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10061–10071, 2022. 2, 4, 7, 8" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 613, + 545, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 613, + 545, + 669 + ], + "spans": [ + { + "bbox": [ + 307, + 613, + 545, + 669 + ], + "type": "text", + "content": "[24] Yinghuan Shi, Jian Zhang, Tong Ling, Jiwen Lu, Yefeng Zheng, Qian Yu, Lei Qi, and Yang Gao. Inconsistency-aware uncertainty estimation for semi-supervised medical image segmentation. IEEE Trans. Medical Imaging, 41(3):608-620, 2022. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 669, + 545, + 715 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 669, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 307, + 669, + 545, + 715 + ], + "type": "text", + "content": "[25] Kihyuk Sohn, David Berthelot, Nicholas Carlini, Zizhao Zhang, Han Zhang, Colin A Raffel, Ekin Dogus Cubuk, Alexey Kurakin, and Chun-Liang Li. Fixmatch: Simplifying semi-supervised learning with consistency and confidence." + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "16300" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 73, + 287, + 698 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 67, + 73, + 287, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 73, + 287, + 94 + ], + "spans": [ + { + "bbox": [ + 67, + 73, + 287, + 94 + ], + "type": "text", + "content": "Advances in neural information processing systems, 33:596-608, 2020. 2, 4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 96, + 287, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 96, + 287, + 140 + ], + "spans": [ + { + "bbox": [ + 48, + 96, + 287, + 140 + ], + "type": "text", + "content": "[26] Antti Tarvainen and Harri Valpola. Mean teachers are better role models: Weight-averaged consistency targets improve semi-supervised deep learning results. Advances in neural information processing systems, 30, 2017. 2, 4" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 141, + 287, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 141, + 287, + 184 + ], + "spans": [ + { + "bbox": [ + 48, + 141, + 287, + 184 + ], + "type": "text", + "content": "[27] Hongyi Wang, Mikhail Yurochkin, Yuekai Sun, Dimitris Papailiopoulos, and Yasaman Khazaeni. Federated learning with matched averaging. arXiv preprint arXiv:2002.06440, 2020. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 186, + 287, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 186, + 287, + 239 + ], + "spans": [ + { + "bbox": [ + 48, + 186, + 287, + 239 + ], + "type": "text", + "content": "[28] Jianyu Wang, Qinghua Liu, Hao Liang, Gauri Joshi, and H Vincent Poor. Tackling the objective inconsistency problem in heterogeneous federated optimization. Advances in neural information processing systems, 33:7611-7623, 2020. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 241, + 287, + 284 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 241, + 287, + 284 + ], + "spans": [ + { + "bbox": [ + 48, + 241, + 287, + 284 + ], + "type": "text", + "content": "[29] Yidong Wang, Bowen Zhang, Wenxin Hou, Zhen Wu, Jindong Wang, and Takahiro Shinozaki. Margin calibration for long-tailed visual recognition. arXiv preprint arXiv:2112.07225, 2021. 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 286, + 287, + 329 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 286, + 287, + 329 + ], + "spans": [ + { + "bbox": [ + 48, + 286, + 287, + 329 + ], + "type": "text", + "content": "[30] Qizhe Xie, Zihang Dai, Eduard Hovy, Thang Luong, and Quoc Le. Unsupervised data augmentation for consistency training. Advances in Neural Information Processing Systems, 33:6256-6268, 2020. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 331, + 287, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 331, + 287, + 396 + ], + "spans": [ + { + "bbox": [ + 48, + 331, + 287, + 396 + ], + "type": "text", + "content": "[31] Dong Yang, Ziyue Xu, Wenqi Li, Andriy Myronenko, Holger R Roth, Stephanie Harmon, Sheng Xu, Baris Turkbey, Evrim Turkbey, Xiaosong Wang, et al. Federated semi-supervised learning for pandemic region segmentation in chest CT using multi-national data from china, Italy, japan. Medical image analysis, 70:101992, 2021. 1, 2, 3, 6, 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 399, + 287, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 399, + 287, + 441 + ], + "spans": [ + { + "bbox": [ + 48, + 399, + 287, + 441 + ], + "type": "text", + "content": "[32] Qiang Yang, Yang Liu, Tianjian Chen, and Yongxin Tong. Federated machine learning: Concept and applications. ACM Transactions on Intelligent Systems and Technology (TIST), 10(2):1-19, 2019. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 443, + 287, + 485 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 443, + 287, + 485 + ], + "spans": [ + { + "bbox": [ + 48, + 443, + 287, + 485 + ], + "type": "text", + "content": "[33] Lequan Yu, Shujun Wang, Xiaomeng Li, Chi-Wing Fu, and Pheng-Ann Heng. Uncertainty-aware self-ensembling model for semi-supervised 3d left atrium segmentation. In Proc. MICCAI, 2019. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 487, + 287, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 487, + 287, + 542 + ], + "spans": [ + { + "bbox": [ + 48, + 487, + 287, + 542 + ], + "type": "text", + "content": "[34] Bowen Zhang, Yidong Wang, Wenxin Hou, Hao Wu, Jindong Wang, Manabu Okumura, and Takahiro Shinozaki. Flexmatch: Boosting semi-supervised learning with curriculum pseudo labeling. Advances in Neural Information Processing Systems, 34:18408-18419, 2021. 2, 4, 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 544, + 287, + 597 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 544, + 287, + 597 + ], + "spans": [ + { + "bbox": [ + 48, + 544, + 287, + 597 + ], + "type": "text", + "content": "[35] Songyang Zhang, Zeming Li, Shipeng Yan, Xuming He, and Jian Sun. Distribution alignment: A unified framework for long-tail visual recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2361-2370, 2021. 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 599, + 287, + 654 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 599, + 287, + 654 + ], + "spans": [ + { + "bbox": [ + 48, + 599, + 287, + 654 + ], + "type": "text", + "content": "[36] Boyan Zhou, Quan Cui, Xiu-Shen Wei, and Zhao-Min Chen. Bbn: Bilateral-branch network with cumulative learning for long-tailed visual recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9719–9728, 2020. 5" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 655, + 287, + 698 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 655, + 287, + 698 + ], + "spans": [ + { + "bbox": [ + 48, + 655, + 287, + 698 + ], + "type": "text", + "content": "[37] Yuyin Zhou, Yan Wang, Peng Tang, Song Bai, Wei Shen, Elliot K. Fishman, and Alan L. Yuille. Semi-supervised 3d abdominal multi-organ segmentation via deep multi-planar co-training. In Proc. WACV, 2019. 2" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16301" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Class Prototypes Based Contrastive Learning for Classifying Multi-Label and Fine-Grained Educational Videos/bbbb89aa-912d-4dd5-8424-10e3f0db2ed4_content_list.json b/2023/Class Prototypes Based Contrastive Learning for Classifying Multi-Label and Fine-Grained Educational Videos/bbbb89aa-912d-4dd5-8424-10e3f0db2ed4_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..2782cc44aed70ef9b6137346fdcd8c69fa7e8400 --- /dev/null +++ b/2023/Class Prototypes Based Contrastive Learning for Classifying Multi-Label and Fine-Grained Educational Videos/bbbb89aa-912d-4dd5-8424-10e3f0db2ed4_content_list.json @@ -0,0 +1,1770 @@ +[ + { + "type": "text", + "text": "Class Prototypes based Contrastive Learning for Classifying Multi-Label and Fine-Grained Educational Videos", + "text_level": 1, + "bbox": [ + 96, + 128, + 875, + 174 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Rohit Gupta $^{1}$ * Anirban Roy $^{2}$ Claire Christensen $^{2}$ Sukeong Kim $^{2}$ Sarah Gerard $^{2}$ Madeline Cincebeaux $^{2}$ Ajay Divakaran $^{2}$ Todd Grindal $^{2}$ Mubarak Shah $^{1}$", + "bbox": [ + 109, + 200, + 857, + 247 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Center for Research in Computer Vision, University of Central Florida rohitg@knights.ucf.edu, shah@crcv.ucf.edu", + "bbox": [ + 161, + 253, + 594, + 289 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "2 SRI International anirban.roy@sri.com", + "bbox": [ + 629, + 253, + 800, + 289 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 323, + 313, + 339 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The recent growth in the consumption of online media by children during early childhood necessitates data-driven tools enabling educators to filter out appropriate educational content for young learners. This paper presents an approach for detecting educational content in online videos. We focus on two widely used educational content classes: literacy and math. For each class, we choose prominent codes (sub-classes) based on the Common Core Standards. For example, literacy codes include 'letter names', 'letter sounds', and math codes include 'counting', 'sorting'. We pose this as a fine-grained multilabel classification problem as videos can contain multiple types of educational content and the content classes can get visually similar (e.g., 'letter names' vs 'letter sounds'). We propose a novel class prototypes based supervised contrastive learning approach that can handle fine-grained samples associated with multiple labels. We learn a class prototype for each class and a loss function is employed to minimize the distances between a class prototype and the samples from the class. Similarly, distances between a class prototype and the samples from other classes are maximized. As the alignment between visual and audio cues are crucial for effective comprehension, we consider a multimodal transformer network to capture the interaction between visual and audio cues in videos while learning the embedding for videos. For evaluation, we present a dataset, APPROVE, employing educational videos from YouTube labeled with fine-grained education classes by education researchers. APPROVE consists of 193 hours of expert-annotated videos with 19 classes. The proposed approach outperforms strong baselines on APPROVE and other benchmarks such as Youtube-8M, and COIN. The dataset is available at https://nusci.csl.sri.com/project/APPROVE.", + "bbox": [ + 75, + 357, + 473, + 857 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 500, + 323, + 630, + 339 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "With the expansion of internet access and the ubiquitous availability of smart devices, children increasingly spend a significant amount of time watching online videos. A recent nationally representative survey reported that $89\\%$ of parents of children aged 11 or younger say their child watches videos on YouTube [4]. Moreover, it is estimated that young children in the age range of two to four years consume 2.5 hours and five to eight years consume 3.0 hours per day on average [45, 46]. Childhood is typically a key period for education, especially for learning basic skills such as literacy and math [20, 25]. Unlike generic online videos, watching appropriate educational videos supports healthy child development and learning [7, 22, 23]. Thus, analyzing the content of these videos may help parents, teachers, and media developers increase young children's exposure to high-quality education videos, which has been shown to produce meaningful learning gains [22]. As the amount of online content produced grows exponentially, automated content understanding methods are essential to facilitate this.", + "bbox": [ + 496, + 349, + 893, + 636 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this work, given a video, our goal is to determine whether the video contains any educational content and characterize the content. Detecting educational content requires identifying multiple distinct types of content in a video while distinguishing between similar content types. The task is challenging as the education codes by Common Core Standards [3, 40] can be similar such as 'letter names' and 'letter sounds', where the former focuses on the name of the letter and the latter is based on the phonetic sound of the letter. Also, understanding education content requires analyzing both visual and audio cues simultaneously as both signals are to be present to ensure effective learning [3, 40]. This is in contrast to standard video classification benchmarks such as the sports or generic YouTube videos in UCF101 [54] Kinetics400 [53], YouTube-8M [1], where visual cues are often sufficient to detect the different classes. Finally, unlike standard well-known action videos,", + "bbox": [ + 496, + 643, + 895, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Work partly done during an internship at SRI International.", + "bbox": [ + 99, + 886, + 419, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "19923", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "education codes are more structured and not accessible to common users. Thus, it requires a carefully curated set of videos and expert annotations to create a dataset to enable a data-driven approach. In this work, we focus on two widely used educational content classes: literacy and math. For each class, we choose prominent codes (sub-classes) based on the Common Core Standards that outline age-appropriate learning standards [3,40]. For example, literacy codes include 'letter names', 'letter sounds', 'rhyming', and math codes include 'counting', 'addition subtraction', 'sorting', 'analyze shapes'.", + "bbox": [ + 75, + 90, + 472, + 257 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We formulate the problem as a multilabel fine-grained video classification task as a video may contain multiple types of content that can be similar. We employ multimodal cues since besides visual cues, audio cues provide important cues to distinguish between similar types of educational content. We propose a class prototypes based supervised contrastive learning approach to address the above-mentioned challenges. We learn a prototype embedding for each class. Then a loss function is employed to minimize the distance between a class prototype and the samples associated with the class label. Similarly, the distance between a class prototype and the samples without that class label is maximized. This is unlike the standard supervised contrastive learning setup where inter-class distance is maximized and intra-class distance is minimized by considering classwise positive and negative samples. This approach is shown to be effective for single-label setups [26]. However, it is not straightforward to extend this for the proposed multilabel setup as samples cannot be identified as positive or negative due to the multiple labels. We jointly learn the embedding of the class prototypes and the samples. The embeddings are learned by a multimodal transformer network (MTN) that captures the interaction between visual and audio cues in videos. We employ automatic speech recognition (ASR) to transcribe text from the audio. The MTN consists of video and text encoders that learn modality-specific embedding and a cross-attention mechanism is employed to capture the interaction between them. The MTN is end-to-end learned through the contrastive loss.", + "bbox": [ + 75, + 261, + 472, + 699 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Due to the lack of suitable datasets for evaluating fine-grained classification of education videos, we propose a new dataset, called APPROVE, of curated YouTube videos annotated with educational content. We follow Common Core Standards [3, 40] to select education content suitable for the kindergarten level. We consider two high-level classes of educational content: literacy and math. For each of these content classes, we select a set of codes. For the literacy class, we select 7 codes and for the math class, we select 11 codes. Each video is associated with multiple labels corresponding to these codes. The videos are annotated by trained education researchers following standard validation protocol [41] to ensure correctness. APPROVE also con", + "bbox": [ + 75, + 704, + 472, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "sists of carefully chosen background videos, i.e., without educational content, that are visually similar to the videos with educational content. APPROVE consists of 193 hours of expert-annotated videos with 19 classes (7 literacy codes, 11 math codes, and a background) where each video has 3 labels on average.", + "bbox": [ + 496, + 90, + 890, + 181 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our contributions can be summarized as follows:", + "bbox": [ + 500, + 181, + 823, + 195 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- APPROVE, a fine-grained multi-label dataset of education videos, to promote exploration in this field.", + "- Class prototypes based contrastive learning framework along with a multi-modal fusion transformer suitable for the problem where videos have multiple fine-grained labels.", + "- Outperforming relevant baselines on three datasets: APPROVE, YouTube-8M [1] and COIN [55]." + ], + "bbox": [ + 517, + 196, + 890, + 316 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 500, + 330, + 648, + 347 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Self-Supervised Contastive Learning (CL) has been an effective paradigm for visual representation learning. Methods such as SimCLR [8], MoCo (Momentum Contrastive learning) [18], Augmented Multiscale Deep InfoMax (AMDIM) [5], Contrastive Predictive Coding (CPC) [39], MoCov2 [10], MoCov3 [12] and SimCLRv2 [9] have achieved strong performance on image classification benchmarks. The shared property between these CL frameworks is that data augmentation is used to generate positive pairs for CL from a single instance, where other data instances are treated as negatives. Prototypical Contrastive Learning (PCL) [30] extends self-supervised contrastive learning with the idea of clustering data representations during training to generate unsupervised prototypes which represent intra-class variation. We utilize class prototypes instead in the supervised setting, to learn fine-grained distinctions between classes.", + "bbox": [ + 496, + 353, + 890, + 608 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Supervised CL methods such as SupCon [26] utilize labels to enhance contrastive learning by forming positive and negative pairs using labels instead of data augmentation. Supervised Contrastive Learning has also been used for other tasks such as image segmentation [57] and classification in the presence of noisy labels [32]. Hierarchical CL [63] extends SupCon to the hierarchical classification case. However, SupCon cannot be extended to the multi-label case in a straightforward manner, as pairs of data samples with multiple labels cannot be clearly classified just into positives and negatives.", + "bbox": [ + 496, + 609, + 890, + 776 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Weakly-Supervised Multi-Modal CL: Weakly aligned text-image/video datasets scraped from the web such as Conceptual Captions [50] and WebVid-10M [6] enable learning of multi-modal representations. CLIP [42] applies a cross-modal contrastive loss to train individual text and image encoders. Everything at Once [51] is able to additionally utilize the audio modality and incorporates a pairwise fusion encoder which encodes pairs of modalities, as a re", + "bbox": [ + 496, + 779, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "19924", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/de2d6f94ec35f2ce55da96d7a453b2c8fcad43a3d8731826aa7bab93a2c70265.jpg", + "image_caption": [ + "sounds in words" + ], + "image_footnote": [], + "bbox": [ + 83, + 89, + 212, + 140 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/a9c25ce02359740175cd7f7eee47e60b0a53f03399eb2ea44ca9dd4204fbc637.jpg", + "image_caption": [ + "rhyming" + ], + "image_footnote": [], + "bbox": [ + 217, + 90, + 344, + 140 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/2064b423610fdf6bbbee26e16f03ca45424ccd5be0bc30dff4a45f20c8910f04.jpg", + "image_caption": [ + "letter names" + ], + "image_footnote": [], + "bbox": [ + 81, + 157, + 210, + 209 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/4086dc0c4c105f6f56f5ac5d849b3fe58dcfb1760398a816516daa493b62a1bc.jpg", + "image_caption": [ + "letter sounds" + ], + "image_footnote": [], + "bbox": [ + 214, + 157, + 344, + 209 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/b6b683cb29e18ec994045cb67ed3c6ee9cbe9f52794b9d989b719338e25e7d44.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 356, + 90, + 486, + 140 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/d96e217df87c94004be40ca45ff179bb041727d2ed795cb749d96ac4354823b7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 488, + 90, + 612, + 141 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/b22019975ef426b83d90d2f01b2c0c9fdb05cb798da0a8a91b1866a7903ae153.jpg", + "image_caption": [ + "(a) Frames from literacy videos", + "analyze compare shape", + "addition subtraction" + ], + "image_footnote": [], + "bbox": [ + 356, + 162, + 485, + 214 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/0a14f1dd2d4cc2e27758e64586bee9b116c3a0ff6b152f9ede720953b4e072c0.jpg", + "image_caption": [ + "building drawing shapes", + "counting", + "(b) Frames from math videos", + "Figure 1. Sample video frames from the APPROVE dataset. Videos belong to the (a) literacy classes, (b) math classes, and (c) background. Background videos do not contain educational content but share visual similarities with educational videos. The videos are labeled with fine-grained sub-classes, e.g., letter names vs letter sounds." + ], + "image_footnote": [], + "bbox": [ + 488, + 162, + 616, + 215 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/f62656d7c4adaed62e4b85bbbcb7f94c0ad53fc9ccb942b08b498ada61f7095a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 624, + 90, + 753, + 142 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/6b631189dbbd0d5105d952fae59eb31b78b2bfc77933254287061d68634f4626.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 756, + 90, + 883, + 142 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/3705b586c96f1539d4e5c1329c024fd30fde3a47044aad6c1116e5faf67649ba.jpg", + "image_caption": [ + "(c) Frames from background videos" + ], + "image_footnote": [], + "bbox": [ + 624, + 157, + 753, + 209 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/42fb2181f46789989d6dad2ec6d93796eaff5df97870038314eab38607521bf6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 756, + 157, + 883, + 209 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "sult, 6 forward passes of the fusion model are required for 3 modalities. Frozen in Time [6] is able to utilize both image-text and video-text datasets through the use of a Space-Time Transformer Visual Encoder. Visual Conditioned GPT [37] uses a single cross-attention fusion layer to combine pretrained CLIP text and visual features. Flamingo [2] adds cross-attention layers interleaved with language decoder layers to fuse visual information into text generation. MERLOT [60, 61] and Triple Contrastive Learning [58] combine contrastive learning and generative language modeling to learn aligned text-image representations.", + "bbox": [ + 75, + 291, + 467, + 455 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Supervised Multi-Modal Learning: Supervised MultiModal Learning typically relies on crowd-captioned datasets such as Flickr30k [59] and MS-COCO Captions [11]. Some prior works such as OSCAR [33] and VinVL [62] have utilized pre-trained object detectors and multi-modal transformers to learn image captioning using supervised aligned datasets. BLIP [28] takes a hybrid approach where it bootstraps an image captioner using a labeled dataset and uses it to generate captions for web images. This generated corpus is then filtered and used for learning an aligned representation. ALign BBefore Fuse [29] highlights the importance of aligning text and image tokens before fusing them using a multi-modal transformer.", + "bbox": [ + 75, + 462, + 467, + 657 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this paper, we focus on the fine-grained classification of multilabel educational videos. Due to the lack of suitable datasets, we propose a new dataset, APPROVE, which is described next.", + "bbox": [ + 75, + 659, + 467, + 718 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. APPROVE Dataset", + "text_level": 1, + "bbox": [ + 76, + 729, + 261, + 744 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We propose a dataset, called APPROVE, of curated YouTube videos annotated with educational content. APPROVE consists of 193 hours of expert-annotated videos with 19 classes (7 literacy codes, 11 math, and background) and each video is associated with approximately 3 labels on average. We follow the Common Core Standards [3,40] to select education content suitable for kindergarten level. The Common Core Standards outline what students are expected to know and do at various age ranges and grades. This is a widely accepted standard followed by a range of", + "bbox": [ + 75, + 750, + 467, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "educators. We consider two high-level classes of educational content: literacy and math. For each of these content classes, we select a set of codes. For the literacy class, we select 7 codes including letter names, letter sounds, follow words, sight words, letters in words, sounds in words, and rhyming. For the math class, we select 11 codes including counting, individual number, comparing groups, addition subtraction, measurable attributes, sorting, spatial language, shape identification, building drawing, analyzing and comparing shapes. More details about the standard and the description of the codes are provided in the supplementary material. APPROVE also consists of carefully chosen background videos, i.e., without educational content, that are visually similar to the videos with educational content. We present frames corresponding to these classes in Fig. 1.", + "bbox": [ + 496, + 291, + 890, + 518 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/eaa0b7ba26fca96539e7a0d27187f4efcb0b95745bdacfb14065959e8898d39f.jpg", + "image_caption": [ + "Frequency", + "Figure 2. Frequency of the classes in APPROVE. Math codes are in Orange and literacy codes in Blue." + ], + "image_footnote": [], + "bbox": [ + 542, + 537, + 846, + 743 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To ensure the quality and correctness of the annotations, we consider educational researchers to annotate the videos and follow a standard validation protocol [41]. Each annotator is trained by an expert and annotations on a selected set are examined before engaging the annotator for the final annotation. Annotators start once they reach more than $90\\%$ agreement with the expert. Further, we estimate inter", + "bbox": [ + 496, + 795, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "19925", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/1dcf09c212e2f1b060f8e7bb90aeda0cb2d49912691ed473f98715cb3b2bcd91.jpg", + "image_caption": [ + "Figure 3. Distribution of the number of labels per video." + ], + "image_footnote": [], + "bbox": [ + 138, + 90, + 403, + 186 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "annotator consistency to filter out anomalies. Details about the validation process are provided in the supplementary material. It takes a month to train an education researcher to match expert-level coding accuracy. On average, it takes the trained annotators 1 min to annotate 1 min of video.", + "bbox": [ + 75, + 220, + 468, + 295 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The videos are curated from YouTube and are annotated by the trained annotators to determine educational content in them. Each video can have multiple class labels that are quite similar making the task a multi-label and fine-grained classification problem. For example 'letter names' and 'letter sounds' where visual letters are shown in both but in 'letter sounds', the phonetic sound on the letter is emphasized (Fig. 1 (a)). Similarly, in both 'build and draw shapes' and 'analyzing and comparing shapes', multiple shapes can appear but the latter focuses on comparing multiple shapes by shape and size (Fig. 1 (b)). Class-wise stats are presented in Fig. 2. Note that the task is different from common video classification setups where either multi-label or fine-grained aspects are dealt separately. Single-label datasets such as HMDB51 [27], UCF101 [54], Kinetics700 [53] and multi-label ones such as Charades [52] are widely used benchmarks for this problem. YouTube-Birds and YouTube-Cars [64] are analogous datasets for object recognition from videos. Multi-Sports [34] and FineGym [49] label fine-grained action classes for sports. HVU [15] also adds scenes and attributes annotations along with action and objects. However, action, object and scene recognition are not enough for fine-grained video understanding. For instance, videos from a given education provider might share similar objects (person, chalkboard, etc.) and actions (writing on chalkboard) while covering different topics (counting, shape recognition etc.) in each video.", + "bbox": [ + 75, + 296, + 470, + 705 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Proposed Approach", + "text_level": 1, + "bbox": [ + 76, + 719, + 266, + 736 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this section, we first describe the proposed class prototypes based contrastive learning framework suitable for videos containing multiple educational codes. Then we present the approach to learning the class prototypes and finally describe the multimodal transformer network that learns features by fusing visual and text cues from videos.", + "bbox": [ + 75, + 737, + 468, + 829 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1. Class prototypes based contrastive learning", + "text_level": 1, + "bbox": [ + 76, + 839, + 447, + 854 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In a contrastive learning framework, feature representations are typically learned by simultaneously minimizing the distance between positive samples and maximizing", + "bbox": [ + 75, + 854, + 468, + 901 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/63a71a5d6491e16d6c302e93ef351e6033286e631b1e206cac5575dc14429c20.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
DatasetSize (in hr)Multi-LabelFine GrainedTypeAnnotators
Action Recognition
HMDB5XXVAuthors
UCF27XXVAuthors
Kinetics800XXV+ACrowd
Video Classification
COIN476XXV+ACrowd
YT-8M-✓XFMachine
APPROVE193✓✓V+T+AExperts
", + "bbox": [ + 501, + 80, + 880, + 247 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Table 1. APPROVE dataset compared with selected prior datasets. V→Video Frames, A→Audio, T→Text, F→Features only.", + "bbox": [ + 498, + 258, + 890, + 287 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "the distance between negative samples (See Figure 4.(a)). The positive and negative samples are determined with respect to an anchor sample usually based on the class labels. For example, supervised contrastive learning (SupCon) [26] learns a representation to minimize the intra-class distances and maximize inter-class distances. We denote $\\boldsymbol{x}_i$ and $y_i$ as the $i$ th sample and its label, respectively. Let's define $z_i$ as the representation of the $i$ th sample in a batch $A$ , and $sim(z_i, z_j) = \\frac{z_i \\cdot z_j}{|z_i||z_j|}$ the cosine similarity, then the SupCon loss [26] is defined as:", + "bbox": [ + 498, + 301, + 890, + 455 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\operatorname {S u p C o n}} = \\sum_ {i \\in A} \\frac {- 1}{| P (i) |} \\sum_ {p \\in P (i)} \\log \\frac {\\exp \\left(\\operatorname {s i m} \\left(\\boldsymbol {z} _ {i} , \\boldsymbol {z} _ {p}\\right) / \\tau\\right)}{\\sum_ {a \\in A \\backslash i} \\exp \\left(\\operatorname {s i m} \\left(\\boldsymbol {z} _ {i} , \\boldsymbol {z} _ {a}\\right) / \\tau\\right)}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 500, + 465, + 901, + 523 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $P(i)$ is the set of positive samples, i.e., with the same label as $z_{i}$ , in the batch excluding $i$ and $a \\in A \\setminus i$ is the index of all samples in the batch excluding the $i$ th sample. $\\tau$ is a scalar temperature parameter used for scaling similarity values. The positive pairs are grouped into the numerator and minimizing the loss minimizes their distance in the learned representation and vice versa for negative pairs. SupCon is known to be effective for classifying samples with a single label. However, it is not straightforward to extend this for the multilabel setup as beyond positive samples, where all labels are the same, and negative samples, where none of the labels is the same, there can be a third scenario where labels are partially overlapping. Though SupCon has been extended to hierarchical classification [63], it cannot be directly extended to the true multi-label case.", + "bbox": [ + 496, + 523, + 890, + 748 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To address this issue, we learn class prototypes as the representative for each class and consider these as anchors while determining positive and negative samples. Specifically, for a specific class prototype, a representation is learned to minimize distances between the prototype and samples with this class label and maximize the distances between the prototype and samples without this class label. We compare the proposed approach with the standard single-label contrastive learning in figure 4. We iteratively update the class prototypes while learning the feature rep", + "bbox": [ + 496, + 750, + 890, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "19926", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/0e76b14c48ebc3c7bcf29f384f22d0146c1256dba49800c43dd96ec2465ec366.jpg", + "image_caption": [ + "Figure 4. Contrastive learning operates on the feature space by bringing the representations of similar samples close and pushing distinct samples apart. Prior work in (a) Supervised Contrastive Learning [26] trains the network by treating instances from the same class as positive pairs and instances from different classes as negative pairs. This approach doesn't generalize to multi-label classification tasks, as some instance pairs have partially overlapping labels. We propose the use of class prototypes to enable (b) Multi-Label Prototypes Contrastive Learning. Each sample and the class prototypes corresponding to the labels associated with the sample are treated as positive pairs. Similarly, negative pairs are determined based on the missing class labels. Prototypes are represented by stars $(\\star)$ and inputs as circles $(\\odot)$ colored with all their relevant labels. We discuss strategies for initializing and learning the label prototypes in Sec. 4.2." + ], + "image_footnote": [], + "bbox": [ + 140, + 90, + 478, + 253 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/593ec9dba574c1a181e71408ab42f8c7927ae739864e3d963c10245ecd348819.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 493, + 93, + 823, + 253 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "resentations. We define $C = \\{c_1, \\ldots, c_K\\}$ as the set of classes where $K$ is the number of classes. For a sample $\\pmb{x}$ , let's define $P_{ml}(\\pmb{x}) = \\{c_k^+\\}, c_k^+ \\in C$ as the set of multiple class labels associated with $\\pmb{x}$ (positive classes) and $c_k^- \\in C \\setminus P_{ml}(\\pmb{x})$ denotes the missing classes (negative classes). We define $CP = \\{cp_1, \\ldots, cp_K\\}$ as the set of class prototypes. Considering $z$ is the representation for the sample $\\pmb{x}$ , the class prototypes based multilabel contrastive loss is defined as:", + "bbox": [ + 75, + 372, + 472, + 507 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} _ {m l c} (\\pmb {x}) = \\\\ \\frac {- 1}{| P _ {m l} (\\boldsymbol {x}) |} \\sum_ {c _ {k} ^ {+} \\in P _ {m l} (\\boldsymbol {x})} \\left[ \\log \\frac {\\exp \\left(\\operatorname {s i m} \\left(\\boldsymbol {z} , \\boldsymbol {c p} _ {k}\\right) / \\tau\\right)}{\\sum_ {c _ {j} ^ {-} \\in C \\backslash P _ {m l} (\\boldsymbol {x})} \\exp \\left(\\operatorname {s i m} \\left(\\boldsymbol {z} , \\boldsymbol {c p} _ {j}\\right) / \\tau\\right)} \\right]. \\tag {2} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 76, + 513, + 478, + 593 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Here, minimizing the loss of the positive class prototype and instance pairs in the numerator minimizes the distance between the representation $z$ and the class prototypes corresponding to the sample, and vice versa for negative classes. We also utilize negative sampling to account for the class imbalance between positives and negatives.", + "bbox": [ + 75, + 594, + 468, + 684 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Learning class prototypes", + "text_level": 1, + "bbox": [ + 76, + 688, + 310, + 704 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We aim to learn class-specific prototypes such that the multilabel samples can be thought of as the combinations of the class prototypes selected based on the associated labels. Lets assume $Z_{t}$ is an $N\\times d$ matrix of $d$ -dimensional representations, i.e., $\\mathbf{z}\\mathbf{s})$ , of $N$ samples and $L\\in \\{0,1\\}^{N\\times K}$ is corresponding labels matrix with $K$ classes. Let's denote $CP_{t}$ is a matrix of size $K\\times d$ of $K$ class prototypes at a training iteration $t$ . Then, $Z_{t} = L\\times CP_{t} + \\varepsilon$ , where $\\varepsilon$ is the residual noise term. Assuming a Gaussian noise that is unbiased and uncorrelated with the labels $L$ , the class prototypes can be approximated as $CP_{t}^{*}\\approx (L^{T}L)^{-1}L^{T}Z_{t}$ where operation $(L^T L)$ results in a square matrix amenable to inversion. Note that for single labels, this implies averag-", + "bbox": [ + 75, + 704, + 470, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "ing the features of the instances belonging to a given class as the prototype for that class. In a multi-label setup, additionally, the co-occurrence between the labels is considered. Finally, the class prototypes are updated with learning iterations as:", + "bbox": [ + 496, + 372, + 890, + 446 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nC P _ {t + 1} = \\beta \\cdot C P _ {t} + (1 - \\beta) \\cdot C P _ {t} ^ {*}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 573, + 454, + 890, + 470 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\beta$ is the decay parameter for the exponential moving averaging. The moving averaging avoids collapsing prototypes across the training iterations [31].", + "bbox": [ + 496, + 483, + 890, + 529 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.3. Multi-Modal Fusion Transformer", + "text_level": 1, + "bbox": [ + 498, + 536, + 795, + 550 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Given a sample video $\\mathbf{x}$ , we consider multimodal cues such as visual and audio cues to learn its representation $\\mathbf{z}$ . Considering multimodal cues is crucial for content recognition, specifically for education videos as effective comprehension requires attending to both visual demonstration and audio explaining the educational content [3, 40]. To capture audio cues, we consider the audio track of the video and extract speech by removing the background such as tunes or instruments [21]. Then we employ the automatic speech recognition (ASR) technique to transcribe text from the speech [43]. We notice that separating speech from the background is important for an accurate ASR transcription. Given the video frames and text transcription, we propose a multimodal transformer network (MTN) to fuse these cues using cross-modal attention.", + "bbox": [ + 496, + 551, + 890, + 777 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our MTN (Figure 5) has three components: image encoder, text encoder, and fusion encoder to learn visual $(z^v)$ , text $(z^t)$ , and fusion $(z^f)$ representations, respectively. The sample representation is comprised of these three representations $z = \\{z^v, z^t, z^f\\}$ . The image encoder is implemented by a vision transformer (ViT) [16] that learns frame embeddings from the video frames along with a special CLS for each frame [14]. We pool the CLS tokens across", + "bbox": [ + 496, + 780, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "19927", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/bafa208d7a462b2edd8c2a532a34c4f605924eafb2a27298c3ff635cceab7ab3.jpg", + "image_caption": [ + "Figure 5. Multi-Modal Classification Network. A text encoder is used to encode ASR text from the video, while an Image Encoder is used to get tokens representing each frame of the video. Unimodal pre-training is carried out on the text & image encoders respectively. Multi-label contrastive loss is used along with shared prototypes to align the representations across both modalities. This is followed by joint end-to-end learning of the whole multimodal network including the fusion encoder which applies multi-head self-attention within and across the modalities. The prototypes are further refined during the multi-modal training phase." + ], + "image_footnote": [], + "bbox": [ + 99, + 89, + 446, + 496 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "the frames and consider this as a compact video representation. Similarly, we consider the BERT-based text transformer [24] to learn the word embeddings along with the text $CLS$ token for the entire transcription. We consider the text $CLS$ as the text representation. Finally, the fusion encoder fuses the visual and text cues by leveraging cross-modal attention between frame and word embeddings.", + "bbox": [ + 75, + 648, + 468, + 755 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.4. Inference using prototypes", + "text_level": 1, + "bbox": [ + 76, + 762, + 316, + 777 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Existing approaches in contrastive learning [8, 26] usually discard the MLP layer after the contrastive training, and a linear classifier is trained on the frozen backbone network. We rely on the class prototypes to carry out inference by utilizing the cosine distance between the learned prototypes and test features. Given our prototype loss-based training, the estimated probability of a given class should be proportional to the normalized temperature scaled cosine distance", + "bbox": [ + 75, + 779, + 468, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "(Equation 4). In practice, we normalize the cosine distance such that $-1$ and $1$ correspond to a confidence of $0\\& 1$ respectively. Thus the prediction is made following:", + "bbox": [ + 498, + 90, + 890, + 137 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {p} (k | \\boldsymbol {x}) \\propto \\exp (\\operatorname {s i m} (\\boldsymbol {z}, \\boldsymbol {c p} _ {k}) / \\tau), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 586, + 147, + 890, + 162 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $z$ is the multimodal representation of the sample $\\pmb{x}$", + "bbox": [ + 498, + 176, + 880, + 191 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.5. Implementation Details", + "text_level": 1, + "bbox": [ + 500, + 198, + 718, + 213 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Training strategy for the multimodal fusion: We follow a two-stage training process: during the initial unimodal training phase, we utilize fixed prototypes in each modality to align the representations. Then in the second stage, we train the unimodal encoders and the multi-modal fusion encoder end-to-end. The cross-modal alignment learned during the first stage improves the learning of the multi-modal representation. The multi-modal learning phase consists of alternating optimization steps of training the network using our contrastive loss per Sec. 4.1 and refining the class prototypes as described in Sec. 4.2.", + "bbox": [ + 496, + 218, + 890, + 383 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Image encoder: For video frames, Random Resized Crop and RandAugment [13] augmentations are used from torchvision. We use ImageNet pretrained vision encoders ResNet50 [19], ViT-B/32 (224×224 resolution) and ViT-B/16 (384×384 resolution) [17].", + "bbox": [ + 496, + 386, + 890, + 460 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Text encoder: We generate text from the videos using Whisper [43], an open source ASR model. For data augmentation, we generate four versions of the ASR text by back-translation using the Helsinki-NLP/opus-mt-{en-de, en-nl, en-fr} models through the nlpaug library [38]. Synonym replacement, text span removal and random word swapping augmentations are also used for the text data. We use DistilBERT-Base-uncased [48], and t5-small [44] from HuggingFace transformers library as the text encoder.", + "bbox": [ + 496, + 462, + 892, + 613 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Optimizer: We employ AdamW optimizer [36] for training with a learning rate of 0.0005. Weight decay of 1e-6 is utilized only on the MLP head during contrastive training and the classifier during BCE/Focal/Asym. loss. Since pre-trained vision and text backbones are used, the backbone learning rate is set to 1/10th of the learning rate for the head. Exponential Moving Averaging every 10 steps with a decay of 0.999 was used for the model parameters.", + "bbox": [ + 496, + 614, + 890, + 734 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 500, + 744, + 632, + 760 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Datasets. In addition to the proposed APPROVE dataset, we evaluate our approach on a subset of Youtube-8M [1] and COIN [55] datasets. YT-8M consists of a diverse set of YouTube videos with video and audio modalities. We consider a subset of YT-8M dataset with 46K videos and 165 classes. COIN consists of instructional videos covering a wide variety of domains and spanning over 180 classes.", + "bbox": [ + 496, + 763, + 890, + 868 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Baselines. We compare the efficacy of our multilabel classification framework against the following baselines:", + "bbox": [ + 498, + 869, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "19928", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Binary cross-entropy: In this baseline, loss for multiple labels is computed by combining the binary cross-entropy losses for individual classes.", + "bbox": [ + 75, + 90, + 468, + 135 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Focal loss [35]: This considers a modified binary cross-entropy to assign a higher weight to hard samples by adjusting a focusing parameter $\\gamma$ . Negative samples can also be down-weighted by using a weight $\\alpha$ . The focal loss for a positive label is given as $\\mathcal{L}_{focal}(p) = -\\alpha (1 - p)^{\\gamma}\\log (p)$ . We set $\\gamma = 2$ and $\\alpha = 0.2$ in our experiments.", + "bbox": [ + 75, + 136, + 467, + 226 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Asymmetric loss [47]: This builds upon the focal loss by utilizing different focusing parameters $\\gamma_{+}$ and $\\gamma_{-}$ for positive and negative labels. It also ignores the negative samples with a prediction probability lower than a margin $m$ . Asymmetric Loss for prediction $p$ corresponding to label $y$ is given as: $\\mathcal{L}_{asym}(p,y) = -yL_{+} - (1 - y)L_{-}$ , where $L_{+} = (1 - p)^{\\gamma_{+}}\\cdot \\log (p)$ and", + "bbox": [ + 75, + 227, + 467, + 332 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "$L_{-} = (\\max (p - m,0))^{\\gamma_{-}}\\cdot \\log (1 - \\max (p - m,0))$ We follow the 5-step procedure recommended by the original authors to train this baseline. We experimentally set $\\{\\gamma_{-} = 2,\\gamma_{+} = 1,m = 0.1\\}$ corresponding to the best performance on APPROVE.", + "bbox": [ + 76, + 333, + 467, + 406 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Metrics. In order to develop a reliable education content detection framework, achieving high precision is crucial. Thus we consider Recall @ 80% Precision (R@80) as the primary metric. We also consider the standard area under the precision-recall curve (AUPR) that is not sensitive to a specific threshold for making the final prediction. We also consider a label ranking average precision (LRAP) [56] metric that is more suitable for the multilabel setup. This estimates whether the ground truth classes are predicted with higher scores than the rest:", + "bbox": [ + 75, + 407, + 467, + 556 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {L R A P} = \\frac {1}{n} \\sum_ {i = 1} ^ {m} \\frac {1}{| Y _ {i} |} \\sum_ {\\lambda \\in Y _ {i}} \\frac {| \\lambda^ {\\prime} \\in Y _ {i} : r n k _ {i} (\\lambda^ {\\prime}) \\leq r n k _ {i} (\\lambda) |}{r n k _ {i} (\\lambda)},\n$$\n", + "text_format": "latex", + "bbox": [ + 76, + 561, + 465, + 602 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $r n k_{i}(\\lambda)$ is the predicted rank of class $\\lambda$ for sample $i$ . LRAP is a ranking metric and in independent of a threshold.", + "bbox": [ + 75, + 613, + 467, + 643 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Results on APPROVE. We compare the proposed approach with the baselines in Tab. 2. Our approach outperforms the strongest baselines by $3.1\\%$ and $2.3\\%$ with respect to R@80 and AUPR, respectively. We also present results for separate models trained on Math and Literacy subsets of Approve respectively. Results on the Math subset are higher compared to the Literacy subset, which indicates the literacy classes are harder to distinguish mostly due to the high inter-class similarity. The top three hardest classes are follow_words, letters_in_words, and sounds_in_words and these are from the literacy set.", + "bbox": [ + 75, + 643, + 467, + 809 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Results on YT-46K and COIN. Beyond the APPROVE dataset, we also test our approach on two public datasets: YT-46K and COIN. Here we provide the results for the multi-modal models and the results for the single-modality models are in Section D of the Supplementary Material.", + "bbox": [ + 75, + 810, + 467, + 883 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Results on YT-46K are provided in Table 3. As YT-8M", + "bbox": [ + 96, + 885, + 467, + 900 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/b2ae9e9ab761df3b30aa8ebb37bb4254e1b8d522c55f8de26b2ca9440c266f3a.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
SubsetModalityMethodAUPRLRAPR@80
AllVBCE45.554.36.9
Focal45.956.615.0
Ours46.757.919.6
TBCE79.885.163.3
Focal79.985.772.8
Ours82.587.475.4
V+TBCE84.388.476.3
Focal [35]86.189.182.2
Asym. [47]86.089.282.4
Ours88.4 +2.390.7 +1.585.5 +3.1
MTHV+TBCE86.392.480.3
Focal87.292.182.4
Ours88.4 +1.293.2 +1.183.2 +0.8
LITV+TBCE72.182.950.7
Focal72.783.550.9
Ours73.6 +0.984.7 +1.254.7 +3.8
", + "bbox": [ + 501, + 88, + 888, + 385 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/4a0f1f9cc9113d0b184287386fd9b63848684c5d10fcd00742ab586deebcb0f6.jpg", + "table_caption": [ + "Table 2. Results on APPROVE dataset. All metrics in %. V $\\rightarrow$ Video & T $\\rightarrow$ Text. M $\\rightarrow$ Math & L $\\rightarrow$ Literacy Subsets." + ], + "table_footnote": [], + "table_body": "
ModalityMethodAUPRLRAPR@80
V+TBCE64.670.242.3
V+TFocal [35]69.772.744.6
V+TOurs70.9 +1.274.9 +2.249.1 +4.5
", + "bbox": [ + 527, + 426, + 857, + 503 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/638c50aa613c23c0c2d3033d1d8468c59619aa9e2ef3cb802abd4f6e5108dc38.jpg", + "table_caption": [ + "Table 3. Results on YT-46K. V→Video Frames and T→Text." + ], + "table_footnote": [], + "table_body": "
ModalityMethodTop-1 Accuracy
V+TCE53.7
V+TBCE54.9
V+TFocal [35]56.1
V+TSupCon [26]54.7
V+TOurs57.5 +1.4
", + "bbox": [ + 553, + 529, + 834, + 636 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 4. Results on COIN. V→Video Frames and T→Text.", + "bbox": [ + 517, + 638, + 870, + 652 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "was primarily collected with the intention of visual classification, the additional use of text data leads to a smaller improvement compared to APPROVE. Results on COIN are provided in Table 4. As each video from COIN is mapped to a single task. Thus, we consider the Top-1 accuracy as the metric. On COIN we compare our approach with SupCon [26] which is effective for single labels. Note that our approach outperforms SupCon and this justifies the effectiveness of the class prototypes based training in a generic contrastive learning framework.", + "bbox": [ + 496, + 660, + 890, + 811 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.1. Ablation Studies", + "text_level": 1, + "bbox": [ + 500, + 823, + 663, + 837 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We perform the following ablation studies to quantify the impact of our approach for learning class prototypes, the multimodal fusion module, and the choice of visual and text encoding frameworks.", + "bbox": [ + 496, + 839, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "19929", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/29c0e89f6d30afb11a9c6157a69403333a5d506332927448b7c1f7379e0ec04f.jpg", + "table_caption": [ + "(a) Class Prototypes" + ], + "table_footnote": [], + "table_body": "
VariantAPPROVECOIN
Random84.156.6
Orthogonal84.857.0
Learned85.557.5
Hierarchical86.057.8
", + "bbox": [ + 83, + 106, + 269, + 172 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/49e8f88cc45ff949c1ec559eae8bd49fbf92962d5fb3c9b61af02754b79a5f0b.jpg", + "table_caption": [ + "(b) Fusion Encoder Size" + ], + "table_footnote": [], + "table_body": "
LayersAPPROVECOIN
184.957.1
285.557.5
485.357.6
", + "bbox": [ + 297, + 106, + 454, + 160 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/360771c02d65046a0d3bd00d5c78ece9f4cbcbc634c64ad063a88fd4b9561ffc.jpg", + "table_caption": [ + "(c) Vision Encoder" + ], + "table_footnote": [], + "table_body": "
Vision ModelAPPROVECOIN
R5084.855.2
ViT-B/3285.557.5
ViT-B/1683.857.8
", + "bbox": [ + 475, + 106, + 669, + 160 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/d8ac3947169d80e31ccfbdf04d2c0afe6268cf07ea8cbbf9f4ef77952a36b98a.jpg", + "table_caption": [ + "(d) Text Encoder" + ], + "table_footnote": [], + "table_body": "
Text ModelAPPROVECOIN
DistilBert-B85.557.5
t5-S87.357.9
", + "bbox": [ + 679, + 106, + 864, + 147 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/4004cb77dcd211dc99d5efe025e7c0274bcfb0e4cd2ef25ac4df780bc2566bc7.jpg", + "table_caption": [ + "Table 5. Ablation studies. R@80% Precision for APPROVE and Top-1 Accuracy for COIN. Default setup is underlined.", + "(a) Noisy modalities." + ], + "table_footnote": [], + "table_body": "
% missingAPPROVECOIN
0%85.557.5
10% V80.153.3
10% T75.857.2
30% T68.942.8
", + "bbox": [ + 129, + 210, + 305, + 276 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/e092c12ed9ddde6c20ad00d26ea9add9f51b2d89f54d966dd400fc5bfa83cc80.jpg", + "table_caption": [ + "(b) Run-to-Run Variance" + ], + "table_footnote": [], + "table_body": "
MethodAPPROVECOIN
BCE76.3 ± 0.754.9 ± 0.6
Focal82.2 ± 0.556.1 ± 0.3
Ours85.5 ± 0.557.5 ± 0.8
", + "bbox": [ + 387, + 210, + 575, + 263 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/1fba4160f622a13f9468404c14142b02de63724245cc05aaa49d265a20be6c82.jpg", + "table_caption": [ + "(c) Initialization" + ], + "table_footnote": [], + "table_body": "
MethodAPPROVECOIN
ImageNet & Wiki-en+TBC85.557.5
CLIP86.763.5
", + "bbox": [ + 648, + 210, + 846, + 270 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 6. Robustness analysis. Our method is robust to partially missing modality and has similar run-to-run variance as baselines.", + "bbox": [ + 99, + 280, + 864, + 294 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Learning class prototypes: We compare the two strategies where after initializing the class prototypes, we 1) keep them fixed and learn only the multimodal embedding of the samples, and 2) class prototypes and sample embedding are learned iteratively. The initialization can be done either randomly, or with orthogonal constraints. We note that orthogonal initialization performs best in our experiments and iterative adjusting the class prototypes achieves better performance as shown in Table 5 (a). We also consider hierarchical prototypes, for APPROVE, using a 2-level hierarchy where the first level consists of 18 classes, and the second level is the 3 super-classes: math, literacy, and background. The 180 task categories of COIN are organized into 12 domains in the taxonomy provided with the dataset. This hierarchy imposes an additional constraint on learning the embeddings during training.", + "bbox": [ + 75, + 304, + 468, + 545 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Fusion Encoder: This evaluates the effect of the number of layers in the fusion encoder on the final performance. As expected, the performance improves with more layers and saturates around 4 layers (Table 5 (b)).", + "bbox": [ + 75, + 546, + 468, + 606 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Vision Encoder: This evaluates the effect of the image encoder on the final performance. We consider ResNet [19] and ViT variants [16]. We notice that the ViT-B/16-384 encoder works well for the larger COIN dataset, whereas the ViT-B/32-224 encoder works best for APPROVE (Table 5 (c)).", + "bbox": [ + 75, + 607, + 468, + 696 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Text Encoder: This evaluates the effect of the text encoder on the final performance. We test DistilBERT and T5 backbones for the text encoder. BERT is trained to predict masked spans of text. T5's unsupervised objective is similar, however, it trains on predicting the entire sequence instead of just the masked spans. GPT2 takes an auto-regressive approach to language modeling (Table 5 (d)).", + "bbox": [ + 75, + 698, + 468, + 804 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.2. Robustness analysis", + "text_level": 1, + "bbox": [ + 76, + 808, + 264, + 823 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We evaluate the robustness of our approach as follows:", + "bbox": [ + 76, + 825, + 439, + 839 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Noisy modality: YouTube videos may have noisy modalities where some of the video frames are missing or ASR transcription is noisy. We show that our approach is robust against the cases where a percentage of video frames or text", + "bbox": [ + 75, + 839, + 468, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "words are missing as shown in Tab 6(a).", + "bbox": [ + 500, + 304, + 764, + 319 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Run-to-Run variance: The low variance across runs (Tab 6(b)) indicates that our approach is not sensitive to random initialization of the class prototypes.", + "bbox": [ + 498, + 320, + 890, + 364 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Initializing the encoders: We consider the ImageNet pretraining for the image encoder, while English Wikipedia + Toronto Book Corpus is used to pre-train the text encoder. We provide results where the backbones are initialized with CLIP [42], which provides a more aligned vision-text representation. As expected, the results are better with the CLIP initialization (Tab 6(c)). The improvements are more significant on COIN than APPROVE as CLIP models may not be exposed to educational videos.", + "bbox": [ + 496, + 366, + 892, + 502 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 511, + 617, + 526 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We have proposed an approach for detecting educational content in online videos. The problem is formulated as a fine-grained multilabel video classification task and we have considered class-prototypes based contrastive learning to address this. We have employed a multimodal transformer network to fuse visual and audio cues. This is crucial for comprehending educational content as both visual and audio cues are to be aligned to ensure effective comprehension. Our approach is shown to be effective in distinguishing fine-grained educational content with high interclass similarity. We have introduced APPROVE - a dataset with 193 hours of expert-annotated educational videos. Beyond APPROVE, we have evaluated our approach on COIN and YouTube-8M datasets where our approach outperforms the competitive baselines.", + "bbox": [ + 496, + 530, + 890, + 756 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 500, + 761, + 666, + 776 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This work is supported by NSF grant # 2139219 and SRI R&D award. Rohit Gupta is supported by ARO grant W911NF-19-1-0356.", + "bbox": [ + 498, + 777, + 890, + 816 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Views and conclusions contained herein are those of the authors and should not be interpreted as representing the official policies or endorsements, either expressed or implied, of NSF, ARO, IARPA, DOI/IBC, or the U.S. Government. The U.S. Government is authorized to reproduce and distribute reprints for Governmental purposes notwithstanding any copyright annotation thereon.", + "bbox": [ + 496, + 818, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "19930", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Sami Abu-El-Haija, Nisarg Kothari, Joonseok Lee, Paul Natsev, George Toderici, Balakrishnan Varadarajan, and Sudheendra Vijayanarasimhan. YouTube-8M: A large-scale video classification benchmark. arXiv preprint arXiv:1609.08675, 2016. 1, 2, 6", + "[2] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katie Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. arXiv preprint arXiv:2204.14198, 2022. 3", + "[3] National Governors Association et al. Common core state standards. Washington, DC, 2010. 1, 2, 3, 5", + "[4] Brooke Auxier, Monica Anderson Andrew Perrin, and Erica Turner. Parenting children in the age of screens. Technical report, Pew Research Center, 2020. 1", + "[5] Philip Bachman, R Devon Hjelm, and William Buchwalter. Learning representations by maximizing mutual information across views. In H. Wallach, H. Larochelle, A. Beygelzimer, F. d Alché-Buc, E. Fox, and R. Garnett, editors, Advances in Neural Information Processing Systems, volume 32. Curran Associates, Inc., 2019. 2", + "[6] Max Bain, Arsha Nagrani, Gül Varol, and Andrew Zisserman. Frozen in time: A joint video and image encoder for end-to-end retrieval. In IEEE International Conference on Computer Vision, 2021. 2, 3", + "[7] Johanna Burkhardt and Wolfgang Lenhard. A meta-analysis on the longitudinal, age-dependent effects of violent video games on aggression. Media Psychology, 25(3):499-512, 2022. 1", + "[8] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International conference on machine learning, pages 1597-1607. PMLR, 2020. 2, 6", + "[9] Ting Chen, Simon Kornblith, Kevin Swersky, Mohammad Norouzi, and Geoffrey E Hinton. Big self-supervised models are strong semi-supervised learners. In H. Larochelle, M. Ranzato, R. Hadsell, M. F. Balcan, and H. Lin, editors, Advances in Neural Information Processing Systems, volume 33, pages 22243-22255. Curran Associates, Inc., 2020. 2", + "[10] Xinlei Chen, Haoqi Fan, Ross Girshick, and Kaiming He. Improved baselines with momentum contrastive learning, 2020. 2", + "[11] Xinlei Chen, Hao Fang, Tsung-Yi Lin, Ramakrishna Vedantam, Saurabh Gupta, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco captions: Data collection and evaluation server. arXiv preprint arXiv:1504.00325, 2015. 3", + "[12] Xinlei Chen, Saining Xie, and Kaiming He. An empirical study of training self-supervised vision transformers, 2021. 2", + "[13] Ekin D Cubuk, Barret Zoph, Jonathon Shlens, and Quoc V Le. Randaugment: Practical automated data augmentation with a reduced search space. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops, pages 702-703, 2020. 6" + ], + "bbox": [ + 78, + 114, + 470, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[14] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018. 5", + "[15] Ali Diba, Mohsen Fayyaz, Vivek Sharma, Manohar Paluri, Jürgen Gall, Rainer Stiefelhagen, and Luc Van Gool. Large scale holistic video understanding. In European Conference on Computer Vision, pages 593-610. Springer, 2020. 4", + "[16] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. ICLR, 2021. 5, 8", + "[17] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations, 2021. 6", + "[18] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9729-9738, 2020. 2", + "[19] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 6, 8", + "[20] Lowry Hemphill and Terrence Tivnan. The importance of early vocabulary for literacy achievement in high-poverty schools. Journal of Education for Students Placed at Risk, 13(4):426-451, 2008. 1", + "[21] Romain Hennequin, Anis Khlif, Felix Voituret, and Manuel Moussallam. Spleeter: a fast and efficient music source separation tool with pre-trained models. Journal of Open Source Software, 5(50):2154, 2020. Deezer Research. 5", + "[22] Lisa B Hurwitz. Getting a read on ready to learn media: A meta-analytic review of effects on literacy. *Child Development*, 90(5):1754–1771, 2019. 1", + "[23] Lisa B Hurwitz and Kelly L Schmitt. Raising readers with ready to learn: A six-year follow-up to an early educational computer game intervention. Computers in Human Behavior, 104:106176, 2020. 1", + "[24] Andrew Jaegle, Felix Gimeno, Andy Brock, Oriol Vinyals, Andrew Zisserman, and Joao Carreira. Perceiver: General perception with iterative attention. In International conference on machine learning, pages 4651-4664. PMLR, 2021. 6", + "[25] Nancy C Jordan, David Kaplan, Chaitanya Ramineni, and Maria N Locuniak. Early math matters: kindergarten number competence and later mathematics outcomes. Developmental psychology, 45(3):850, 2009. 1", + "[26] Prannay Khosla, Piotr Teterwak, Chen Wang, Aaron Sarna, Yonglong Tian, Phillip Isola, Aaron Maschinot, Ce Liu, and Dilip Krishnan. Supervised contrastive learning. Advances in Neural Information Processing Systems, 33:18661-18673, 2020. 2, 4, 5, 6, 7" + ], + "bbox": [ + 501, + 92, + 890, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "19931", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[27] Hildegard Kuehne, Hueihan Jhuang, Estíbaliz Garrote, Tomaso Poggio, and Thomas Serre. Hmdb: a large video database for human motion recognition. In 2011 International conference on computer vision, pages 2556-2563. IEEE, 2011. 4", + "[28] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In ICML, 2022. 3", + "[29] Junnan Li, Ramprasaath R. Selvaraju, Akhilesh Deepak Gotmare, Shafiq Joty, Caiming Xiong, and Steven Hoi. Align before fuse: Vision and language representation learning with momentum distillation. In NeurIPS, 2021. 3", + "[30] Junnan Li, Pan Zhou, Caiming Xiong, and Steven Hoi. Prototypical contrastive learning of unsupervised representations. In International Conference on Learning Representations, 2021. 2", + "[31] Junnan Li, Pan Zhou, Caiming Xiong, and Steven C.H. Hoi. Prototypical contrastive learning of unsupervised representations. In ICLR, 2021. 5", + "[32] Shikun Li, Xiaobo Xia, Shiming Ge, and Tongliang Liu. Selective-supervised contrastive learning with noisy labels. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 316-325, June 2022. 2", + "[33] Xiujun Li, Xi Yin, Chunyuan Li, Xiaowei Hu, Pengchuan Zhang, Lei Zhang, Lijuan Wang, Houdong Hu, Li Dong, Furu Wei, Yejin Choi, and Jianfeng Gao. Oscar: Object-semantics aligned pre-training for vision-language tasks. ECCV 2020, 2020. 3", + "[34] Yixuan Li, Lei Chen, Runyu He, Zhenzhi Wang, Gangshan Wu, and Limin Wang. Multisports: A multi-person video dataset of spatio-temporally localized sports actions. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13536-13545, 2021. 4", + "[35] Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dólár. Focal loss for dense object detection. In Proceedings of the IEEE international conference on computer vision, pages 2980-2988, 2017. 7", + "[36] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In International Conference on Learning Representations, 2019. 6", + "[37] Ziyang Luo, Yadong Xi, Rongsheng Zhang, and Jing Ma. A frustratingly simple approach for end-to-end image captioning, 2022. 3", + "[38] Edward Ma. Nlp augmentation. https://github.com/makcedward/nlpaug, 2019.6", + "[39] Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748, 2018. 2", + "[40] Andrew Porter, Jennifer McMaken, Jun Hwang, and Rui Yang. Common core standards: The new us intended curriculum. Educational researcher, 40(3):103-116, 2011. 1, 2, 3, 5", + "[41] J. S. Radesky, A. Schaller, S. L. Yeo, H. M. Weeks, and M. B. Robb. Young kids and youtube: How ads, toys, and games dominate viewing. common sense media. Technical report, Common Sense Media, 2020. 2, 3" + ], + "bbox": [ + 78, + 90, + 468, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[42] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 2, 8", + "[43] Alec Radford, Jong Wook Kim, Tao Xu, Greg Brockman, Christine McLeavey, and Ilya Sutskever. Robust speech recognition via large-scale weak supervision. Technical report, Tech. Rep., Technical report, OpenAI, 2022. 5, 6", + "[44] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, Peter J Liu, et al. Exploring the limits of transfer learning with a unified text-to-text transformer. J. Mach. Learn. Res., 21(140):1-67, 2020. 6", + "[45] Victoria Rideout and Michael B. Robb. The common sense census: Media use by tweens and teens, Oct 2019. 1", + "[46] V Rideout and M. B. Robb. The common sense census: Media use by tweens and teens. common sense media. Technical report, Common Sense Media, 2019. 1", + "[47] Tal Ridnik, Emanuel Ben-Baruch, Nadav Zamir, Asaf Noy, Itamar Friedman, Matan Protter, and Lihi Zelnik-Manor. Asymmetric loss for multi-label classification. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 82-91, 2021. 7", + "[48] Victor Sanh, Lysandre Debut, Julien Chaumont, and Thomas Wolf. Distilbert, a distilled version of bert: smaller, faster, cheaper and lighter. arXiv preprint arXiv:1910.01108, 2019.6", + "[49] Dian Shao, Yue Zhao, Bo Dai, and Dahua Lin. Finegym: A hierarchical video dataset for fine-grained action understanding. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2616-2625, 2020. 4", + "[50] Piyush Sharma, Nan Ding, Sebastian Goodman, and Radu Soricut. Conceptual captions: A cleaned, hypernymed, image alt-text dataset for automatic image captioning. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 2556-2565, Melbourne, Australia, July 2018. Association for Computational Linguistics. 2", + "[51] Nina Shvetsova, Brian Chen, Andrew Rouditchenko, Samuel Thomas, Brian Kingsbury, Rogerio S Feris, David Harwath, James Glass, and Hilde Kuehne. Everything at once-multimodal fusion transformer for video retrieval. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20020-20029, 2022. 2", + "[52] Gunnar A. Sigurdsson, Gül Varol, Xiaolong Wang, Ali Farhadi, Ivan Laptev, and Abhinav Gupta. Hollywood in homes: Crowdsourcing data collection for activity understanding. In Bastian Leibe, Jiri Matas, Nicu Sebe, and Max Welling, editors, Computer Vision – ECCV 2016, pages 510–526, Cham, 2016. Springer International Publishing. 4", + "[53] Lucas Smaira, João Carreira, Eric Noland, Ellen Clancy, Amy Wu, and Andrew Zisserman. A short note on the kinetics-700-2020 human action dataset, 2020. 1, 4" + ], + "bbox": [ + 501, + 90, + 892, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "19932", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[54] Khurram Soomro, Amir Roshan Zamir, and Mubarak Shah. Ucf101: A dataset of 101 human actions classes from videos in the wild. arXiv preprint arXiv:1212.0402, 2012. 1, 4", + "[55] Yansong Tang, Dajun Ding, Yongming Rao, Yu Zheng, Danyang Zhang, Lili Zhao, Jiwen Lu, and Jie Zhou. Coin: A large-scale dataset for comprehensive instructional video analysis. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 2, 6", + "[56] Grigorios Tsoumakas, Ioannis Katakis, and Ioannis Vla-havas. Mining Multi-label Data, pages 667-685. Springer US, Boston, MA, 2010. 7", + "[57] Wenguan Wang, Tianfei Zhou, Fisher Yu, Jifeng Dai, Ender Konukoglu, and Luc Van Gool. Exploring cross-image pixel contrast for semantic segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 7303-7313, 2021. 2", + "[58] Jinyu Yang, Jiali Duan, Son Tran, Yi Xu, Sampath Chanda, Liquun Chen, Belinda Zeng, Trishul Chilimbi, and Junzhou Huang. Vision-language pre-training with triple contrastive learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15671-15680, 2022. 3", + "[59] Peter Young, Alice Lai, Micah Hodosh, and Julia Hockenmaier. From image descriptions to visual denotations: New similarity metrics for semantic inference over event descriptions. Transactions of the Association for Computational Linguistics, 2:67-78, 2014. 3" + ], + "bbox": [ + 78, + 90, + 468, + 465 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[60] Rowan Zellers, Jiasen Lu, Ximing Lu, Youngjae Yu, Yanpeng Zhao, Mohammadreza Salehi, Aditya Kusupati, Jack Hessel, Ali Farhadi, and Yejin Choi. Merlot reserve: Neural script knowledge through vision and language and sound. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 16375-16387, June 2022. 3", + "[61] Rowan Zellers, Ximing Lu, Jack Hessel, Youngjae Yu, Jae Sung Park, Jize Cao, Ali Farhadi, and Yejin Choi. Merlot: Multimodal neural script knowledge models. Advances in Neural Information Processing Systems, 34:23634-23651, 2021. 3", + "[62] Pengchuan Zhang, Xiujun Li, Xiaowei Hu, Jianwei Yang, Lei Zhang, Lijuan Wang, Yejin Choi, and Jianfeng Gao. Vinvl: Making visual representations matter in vision-language models. CVPR 2021, 2021. 3", + "[63] Shu Zhang, Ran Xu, Caiming Xiong, and Chetan Ramaiah. Use all the labels: A hierarchical multi-label contrastive learning framework. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 16660-16669, June 2022. 2, 4", + "[64] Chen Zhu, Xiao Tan, Feng Zhou, Xiao Liu, Kaiyu Yue, Errui Ding, and Yi Ma. Fine-grained video categorization with redundancy reduction attention. In Proceedings of the European Conference on Computer Vision (ECCV), pages 136-152, 2018. 4" + ], + "bbox": [ + 501, + 90, + 890, + 464 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "19933", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2023/Class Prototypes Based Contrastive Learning for Classifying Multi-Label and Fine-Grained Educational Videos/bbbb89aa-912d-4dd5-8424-10e3f0db2ed4_model.json b/2023/Class Prototypes Based Contrastive Learning for Classifying Multi-Label and Fine-Grained Educational Videos/bbbb89aa-912d-4dd5-8424-10e3f0db2ed4_model.json new file mode 100644 index 0000000000000000000000000000000000000000..d22681c7f0f14e80a40bee3f3d1e72433f2c016f --- /dev/null +++ b/2023/Class Prototypes Based Contrastive Learning for Classifying Multi-Label and Fine-Grained Educational Videos/bbbb89aa-912d-4dd5-8424-10e3f0db2ed4_model.json @@ -0,0 +1,2609 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.0, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.098, + 0.13, + 0.877, + 0.175 + ], + "angle": 0, + "content": "Class Prototypes based Contrastive Learning for Classifying Multi-Label and Fine-Grained Educational Videos" + }, + { + "type": "text", + "bbox": [ + 0.111, + 0.202, + 0.859, + 0.248 + ], + "angle": 0, + "content": "Rohit Gupta \\(^{1}\\) * Anirban Roy \\(^{2}\\) Claire Christensen \\(^{2}\\) Sukeong Kim \\(^{2}\\) Sarah Gerard \\(^{2}\\) Madeline Cincebeaux \\(^{2}\\) Ajay Divakaran \\(^{2}\\) Todd Grindal \\(^{2}\\) Mubarak Shah \\(^{1}\\)" + }, + { + "type": "text", + "bbox": [ + 0.163, + 0.254, + 0.596, + 0.29 + ], + "angle": 0, + "content": "1 Center for Research in Computer Vision, University of Central Florida rohitg@knights.ucf.edu, shah@crcv.ucf.edu" + }, + { + "type": "text", + "bbox": [ + 0.63, + 0.254, + 0.802, + 0.29 + ], + "angle": 0, + "content": "2 SRI International anirban.roy@sri.com" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.324, + 0.314, + 0.34 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.358, + 0.474, + 0.858 + ], + "angle": 0, + "content": "The recent growth in the consumption of online media by children during early childhood necessitates data-driven tools enabling educators to filter out appropriate educational content for young learners. This paper presents an approach for detecting educational content in online videos. We focus on two widely used educational content classes: literacy and math. For each class, we choose prominent codes (sub-classes) based on the Common Core Standards. For example, literacy codes include 'letter names', 'letter sounds', and math codes include 'counting', 'sorting'. We pose this as a fine-grained multilabel classification problem as videos can contain multiple types of educational content and the content classes can get visually similar (e.g., 'letter names' vs 'letter sounds'). We propose a novel class prototypes based supervised contrastive learning approach that can handle fine-grained samples associated with multiple labels. We learn a class prototype for each class and a loss function is employed to minimize the distances between a class prototype and the samples from the class. Similarly, distances between a class prototype and the samples from other classes are maximized. As the alignment between visual and audio cues are crucial for effective comprehension, we consider a multimodal transformer network to capture the interaction between visual and audio cues in videos while learning the embedding for videos. For evaluation, we present a dataset, APPROVE, employing educational videos from YouTube labeled with fine-grained education classes by education researchers. APPROVE consists of 193 hours of expert-annotated videos with 19 classes. The proposed approach outperforms strong baselines on APPROVE and other benchmarks such as Youtube-8M, and COIN. The dataset is available at https://nusci.csl.sri.com/project/APPROVE." + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.324, + 0.632, + 0.34 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.35, + 0.895, + 0.637 + ], + "angle": 0, + "content": "With the expansion of internet access and the ubiquitous availability of smart devices, children increasingly spend a significant amount of time watching online videos. A recent nationally representative survey reported that \\(89\\%\\) of parents of children aged 11 or younger say their child watches videos on YouTube [4]. Moreover, it is estimated that young children in the age range of two to four years consume 2.5 hours and five to eight years consume 3.0 hours per day on average [45, 46]. Childhood is typically a key period for education, especially for learning basic skills such as literacy and math [20, 25]. Unlike generic online videos, watching appropriate educational videos supports healthy child development and learning [7, 22, 23]. Thus, analyzing the content of these videos may help parents, teachers, and media developers increase young children's exposure to high-quality education videos, which has been shown to produce meaningful learning gains [22]. As the amount of online content produced grows exponentially, automated content understanding methods are essential to facilitate this." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.644, + 0.896, + 0.903 + ], + "angle": 0, + "content": "In this work, given a video, our goal is to determine whether the video contains any educational content and characterize the content. Detecting educational content requires identifying multiple distinct types of content in a video while distinguishing between similar content types. The task is challenging as the education codes by Common Core Standards [3, 40] can be similar such as 'letter names' and 'letter sounds', where the former focuses on the name of the letter and the latter is based on the phonetic sound of the letter. Also, understanding education content requires analyzing both visual and audio cues simultaneously as both signals are to be present to ensure effective learning [3, 40]. This is in contrast to standard video classification benchmarks such as the sports or generic YouTube videos in UCF101 [54] Kinetics400 [53], YouTube-8M [1], where visual cues are often sufficient to detect the different classes. Finally, unlike standard well-known action videos," + }, + { + "type": "page_footnote", + "bbox": [ + 0.101, + 0.887, + 0.421, + 0.901 + ], + "angle": 0, + "content": "*Work partly done during an internship at SRI International." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19923" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.258 + ], + "angle": 0, + "content": "education codes are more structured and not accessible to common users. Thus, it requires a carefully curated set of videos and expert annotations to create a dataset to enable a data-driven approach. In this work, we focus on two widely used educational content classes: literacy and math. For each class, we choose prominent codes (sub-classes) based on the Common Core Standards that outline age-appropriate learning standards [3,40]. For example, literacy codes include 'letter names', 'letter sounds', 'rhyming', and math codes include 'counting', 'addition subtraction', 'sorting', 'analyze shapes'." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.262, + 0.473, + 0.7 + ], + "angle": 0, + "content": "We formulate the problem as a multilabel fine-grained video classification task as a video may contain multiple types of content that can be similar. We employ multimodal cues since besides visual cues, audio cues provide important cues to distinguish between similar types of educational content. We propose a class prototypes based supervised contrastive learning approach to address the above-mentioned challenges. We learn a prototype embedding for each class. Then a loss function is employed to minimize the distance between a class prototype and the samples associated with the class label. Similarly, the distance between a class prototype and the samples without that class label is maximized. This is unlike the standard supervised contrastive learning setup where inter-class distance is maximized and intra-class distance is minimized by considering classwise positive and negative samples. This approach is shown to be effective for single-label setups [26]. However, it is not straightforward to extend this for the proposed multilabel setup as samples cannot be identified as positive or negative due to the multiple labels. We jointly learn the embedding of the class prototypes and the samples. The embeddings are learned by a multimodal transformer network (MTN) that captures the interaction between visual and audio cues in videos. We employ automatic speech recognition (ASR) to transcribe text from the audio. The MTN consists of video and text encoders that learn modality-specific embedding and a cross-attention mechanism is employed to capture the interaction between them. The MTN is end-to-end learned through the contrastive loss." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.705, + 0.473, + 0.903 + ], + "angle": 0, + "content": "Due to the lack of suitable datasets for evaluating fine-grained classification of education videos, we propose a new dataset, called APPROVE, of curated YouTube videos annotated with educational content. We follow Common Core Standards [3, 40] to select education content suitable for the kindergarten level. We consider two high-level classes of educational content: literacy and math. For each of these content classes, we select a set of codes. For the literacy class, we select 7 codes and for the math class, we select 11 codes. Each video is associated with multiple labels corresponding to these codes. The videos are annotated by trained education researchers following standard validation protocol [41] to ensure correctness. APPROVE also con" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.182 + ], + "angle": 0, + "content": "sists of carefully chosen background videos, i.e., without educational content, that are visually similar to the videos with educational content. APPROVE consists of 193 hours of expert-annotated videos with 19 classes (7 literacy codes, 11 math codes, and a background) where each video has 3 labels on average." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.183, + 0.825, + 0.196 + ], + "angle": 0, + "content": "Our contributions can be summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.198, + 0.89, + 0.227 + ], + "angle": 0, + "content": "- APPROVE, a fine-grained multi-label dataset of education videos, to promote exploration in this field." + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.228, + 0.891, + 0.287 + ], + "angle": 0, + "content": "- Class prototypes based contrastive learning framework along with a multi-modal fusion transformer suitable for the problem where videos have multiple fine-grained labels." + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.289, + 0.89, + 0.318 + ], + "angle": 0, + "content": "- Outperforming relevant baselines on three datasets: APPROVE, YouTube-8M [1] and COIN [55]." + }, + { + "type": "list", + "bbox": [ + 0.518, + 0.198, + 0.891, + 0.318 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.332, + 0.65, + 0.348 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.354, + 0.892, + 0.609 + ], + "angle": 0, + "content": "Self-Supervised Contastive Learning (CL) has been an effective paradigm for visual representation learning. Methods such as SimCLR [8], MoCo (Momentum Contrastive learning) [18], Augmented Multiscale Deep InfoMax (AMDIM) [5], Contrastive Predictive Coding (CPC) [39], MoCov2 [10], MoCov3 [12] and SimCLRv2 [9] have achieved strong performance on image classification benchmarks. The shared property between these CL frameworks is that data augmentation is used to generate positive pairs for CL from a single instance, where other data instances are treated as negatives. Prototypical Contrastive Learning (PCL) [30] extends self-supervised contrastive learning with the idea of clustering data representations during training to generate unsupervised prototypes which represent intra-class variation. We utilize class prototypes instead in the supervised setting, to learn fine-grained distinctions between classes." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.611, + 0.892, + 0.777 + ], + "angle": 0, + "content": "Supervised CL methods such as SupCon [26] utilize labels to enhance contrastive learning by forming positive and negative pairs using labels instead of data augmentation. Supervised Contrastive Learning has also been used for other tasks such as image segmentation [57] and classification in the presence of noisy labels [32]. Hierarchical CL [63] extends SupCon to the hierarchical classification case. However, SupCon cannot be extended to the multi-label case in a straightforward manner, as pairs of data samples with multiple labels cannot be clearly classified just into positives and negatives." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.78, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Weakly-Supervised Multi-Modal CL: Weakly aligned text-image/video datasets scraped from the web such as Conceptual Captions [50] and WebVid-10M [6] enable learning of multi-modal representations. CLIP [42] applies a cross-modal contrastive loss to train individual text and image encoders. Everything at Once [51] is able to additionally utilize the audio modality and incorporates a pairwise fusion encoder which encodes pairs of modalities, as a re" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "19924" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.084, + 0.09, + 0.214, + 0.141 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.097, + 0.143, + 0.193, + 0.152 + ], + "angle": 0, + "content": "sounds in words" + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.091, + 0.345, + 0.141 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.261, + 0.144, + 0.306, + 0.153 + ], + "angle": 0, + "content": "rhyming" + }, + { + "type": "image", + "bbox": [ + 0.083, + 0.158, + 0.212, + 0.21 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.105, + 0.211, + 0.182, + 0.219 + ], + "angle": 0, + "content": "letter names" + }, + { + "type": "image", + "bbox": [ + 0.215, + 0.158, + 0.345, + 0.21 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.237, + 0.211, + 0.321, + 0.219 + ], + "angle": 0, + "content": "letter sounds" + }, + { + "type": "image_caption", + "bbox": [ + 0.128, + 0.228, + 0.323, + 0.24 + ], + "angle": 0, + "content": "(a) Frames from literacy videos" + }, + { + "type": "image", + "bbox": [ + 0.357, + 0.091, + 0.488, + 0.141 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.374, + 0.143, + 0.471, + 0.163 + ], + "angle": 0, + "content": "analyze compare shape" + }, + { + "type": "image", + "bbox": [ + 0.49, + 0.091, + 0.613, + 0.142 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.5, + 0.144, + 0.603, + 0.163 + ], + "angle": 0, + "content": "building drawing shapes" + }, + { + "type": "image", + "bbox": [ + 0.357, + 0.164, + 0.486, + 0.215 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.356, + 0.216, + 0.484, + 0.226 + ], + "angle": 0, + "content": "addition subtraction" + }, + { + "type": "image", + "bbox": [ + 0.489, + 0.164, + 0.617, + 0.216 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.527, + 0.217, + 0.579, + 0.226 + ], + "angle": 0, + "content": "counting" + }, + { + "type": "image_caption", + "bbox": [ + 0.381, + 0.228, + 0.547, + 0.24 + ], + "angle": 0, + "content": "(b) Frames from math videos" + }, + { + "type": "image", + "bbox": [ + 0.625, + 0.092, + 0.754, + 0.143 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.757, + 0.092, + 0.885, + 0.143 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.625, + 0.158, + 0.754, + 0.21 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.757, + 0.158, + 0.885, + 0.21 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.651, + 0.228, + 0.86, + 0.24 + ], + "angle": 0, + "content": "(c) Frames from background videos" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.245, + 0.892, + 0.286 + ], + "angle": 0, + "content": "Figure 1. Sample video frames from the APPROVE dataset. Videos belong to the (a) literacy classes, (b) math classes, and (c) background. Background videos do not contain educational content but share visual similarities with educational videos. The videos are labeled with fine-grained sub-classes, e.g., letter names vs letter sounds." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.292, + 0.468, + 0.457 + ], + "angle": 0, + "content": "sult, 6 forward passes of the fusion model are required for 3 modalities. Frozen in Time [6] is able to utilize both image-text and video-text datasets through the use of a Space-Time Transformer Visual Encoder. Visual Conditioned GPT [37] uses a single cross-attention fusion layer to combine pretrained CLIP text and visual features. Flamingo [2] adds cross-attention layers interleaved with language decoder layers to fuse visual information into text generation. MERLOT [60, 61] and Triple Contrastive Learning [58] combine contrastive learning and generative language modeling to learn aligned text-image representations." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.463, + 0.468, + 0.658 + ], + "angle": 0, + "content": "Supervised Multi-Modal Learning: Supervised MultiModal Learning typically relies on crowd-captioned datasets such as Flickr30k [59] and MS-COCO Captions [11]. Some prior works such as OSCAR [33] and VinVL [62] have utilized pre-trained object detectors and multi-modal transformers to learn image captioning using supervised aligned datasets. BLIP [28] takes a hybrid approach where it bootstraps an image captioner using a labeled dataset and uses it to generate captions for web images. This generated corpus is then filtered and used for learning an aligned representation. ALign BBefore Fuse [29] highlights the importance of aligning text and image tokens before fusing them using a multi-modal transformer." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.66, + 0.468, + 0.719 + ], + "angle": 0, + "content": "In this paper, we focus on the fine-grained classification of multilabel educational videos. Due to the lack of suitable datasets, we propose a new dataset, APPROVE, which is described next." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.731, + 0.263, + 0.745 + ], + "angle": 0, + "content": "3. APPROVE Dataset" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.468, + 0.901 + ], + "angle": 0, + "content": "We propose a dataset, called APPROVE, of curated YouTube videos annotated with educational content. APPROVE consists of 193 hours of expert-annotated videos with 19 classes (7 literacy codes, 11 math, and background) and each video is associated with approximately 3 labels on average. We follow the Common Core Standards [3,40] to select education content suitable for kindergarten level. The Common Core Standards outline what students are expected to know and do at various age ranges and grades. This is a widely accepted standard followed by a range of" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.292, + 0.892, + 0.519 + ], + "angle": 0, + "content": "educators. We consider two high-level classes of educational content: literacy and math. For each of these content classes, we select a set of codes. For the literacy class, we select 7 codes including letter names, letter sounds, follow words, sight words, letters in words, sounds in words, and rhyming. For the math class, we select 11 codes including counting, individual number, comparing groups, addition subtraction, measurable attributes, sorting, spatial language, shape identification, building drawing, analyzing and comparing shapes. More details about the standard and the description of the codes are provided in the supplementary material. APPROVE also consists of carefully chosen background videos, i.e., without educational content, that are visually similar to the videos with educational content. We present frames corresponding to these classes in Fig. 1." + }, + { + "type": "image", + "bbox": [ + 0.544, + 0.539, + 0.848, + 0.744 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.734, + 0.747, + 0.786, + 0.758 + ], + "angle": 0, + "content": "Frequency" + }, + { + "type": "image_caption", + "bbox": [ + 0.5, + 0.764, + 0.819, + 0.793 + ], + "angle": 0, + "content": "Figure 2. Frequency of the classes in APPROVE. Math codes are in Orange and literacy codes in Blue." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.796, + 0.892, + 0.901 + ], + "angle": 0, + "content": "To ensure the quality and correctness of the annotations, we consider educational researchers to annotate the videos and follow a standard validation protocol [41]. Each annotator is trained by an expert and annotations on a selected set are examined before engaging the annotator for the final annotation. Annotators start once they reach more than \\(90\\%\\) agreement with the expert. Further, we estimate inter" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "19925" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.14, + 0.092, + 0.404, + 0.187 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.105, + 0.191, + 0.441, + 0.205 + ], + "angle": 0, + "content": "Figure 3. Distribution of the number of labels per video." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.222, + 0.47, + 0.296 + ], + "angle": 0, + "content": "annotator consistency to filter out anomalies. Details about the validation process are provided in the supplementary material. It takes a month to train an education researcher to match expert-level coding accuracy. On average, it takes the trained annotators 1 min to annotate 1 min of video." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.297, + 0.471, + 0.706 + ], + "angle": 0, + "content": "The videos are curated from YouTube and are annotated by the trained annotators to determine educational content in them. Each video can have multiple class labels that are quite similar making the task a multi-label and fine-grained classification problem. For example 'letter names' and 'letter sounds' where visual letters are shown in both but in 'letter sounds', the phonetic sound on the letter is emphasized (Fig. 1 (a)). Similarly, in both 'build and draw shapes' and 'analyzing and comparing shapes', multiple shapes can appear but the latter focuses on comparing multiple shapes by shape and size (Fig. 1 (b)). Class-wise stats are presented in Fig. 2. Note that the task is different from common video classification setups where either multi-label or fine-grained aspects are dealt separately. Single-label datasets such as HMDB51 [27], UCF101 [54], Kinetics700 [53] and multi-label ones such as Charades [52] are widely used benchmarks for this problem. YouTube-Birds and YouTube-Cars [64] are analogous datasets for object recognition from videos. Multi-Sports [34] and FineGym [49] label fine-grained action classes for sports. HVU [15] also adds scenes and attributes annotations along with action and objects. However, action, object and scene recognition are not enough for fine-grained video understanding. For instance, videos from a given education provider might share similar objects (person, chalkboard, etc.) and actions (writing on chalkboard) while covering different topics (counting, shape recognition etc.) in each video." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.72, + 0.267, + 0.737 + ], + "angle": 0, + "content": "4. Proposed Approach" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.738, + 0.47, + 0.83 + ], + "angle": 0, + "content": "In this section, we first describe the proposed class prototypes based contrastive learning framework suitable for videos containing multiple educational codes. Then we present the approach to learning the class prototypes and finally describe the multimodal transformer network that learns features by fusing visual and text cues from videos." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.84, + 0.449, + 0.856 + ], + "angle": 0, + "content": "4.1. Class prototypes based contrastive learning" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.47, + 0.902 + ], + "angle": 0, + "content": "In a contrastive learning framework, feature representations are typically learned by simultaneously minimizing the distance between positive samples and maximizing" + }, + { + "type": "table", + "bbox": [ + 0.502, + 0.082, + 0.882, + 0.248 + ], + "angle": 0, + "content": "
DatasetSize (in hr)Multi-LabelFine GrainedTypeAnnotators
Action Recognition
HMDB5XXVAuthors
UCF27XXVAuthors
Kinetics800XXV+ACrowd
Video Classification
COIN476XXV+ACrowd
YT-8M-✓XFMachine
APPROVE193✓✓V+T+AExperts
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.259, + 0.892, + 0.288 + ], + "angle": 0, + "content": "Table 1. APPROVE dataset compared with selected prior datasets. V→Video Frames, A→Audio, T→Text, F→Features only." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.303, + 0.892, + 0.456 + ], + "angle": 0, + "content": "the distance between negative samples (See Figure 4.(a)). The positive and negative samples are determined with respect to an anchor sample usually based on the class labels. For example, supervised contrastive learning (SupCon) [26] learns a representation to minimize the intra-class distances and maximize inter-class distances. We denote \\( \\boldsymbol{x}_i \\) and \\( y_i \\) as the \\( i \\)th sample and its label, respectively. Let's define \\( z_i \\) as the representation of the \\( i \\)th sample in a batch \\( A \\), and \\( sim(z_i, z_j) = \\frac{z_i \\cdot z_j}{|z_i||z_j|} \\) the cosine similarity, then the SupCon loss [26] is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.5, + 0.466, + 0.903, + 0.524 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\operatorname {S u p C o n}} = \\sum_ {i \\in A} \\frac {- 1}{| P (i) |} \\sum_ {p \\in P (i)} \\log \\frac {\\exp \\left(\\operatorname {s i m} \\left(\\boldsymbol {z} _ {i} , \\boldsymbol {z} _ {p}\\right) / \\tau\\right)}{\\sum_ {a \\in A \\backslash i} \\exp \\left(\\operatorname {s i m} \\left(\\boldsymbol {z} _ {i} , \\boldsymbol {z} _ {a}\\right) / \\tau\\right)}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.524, + 0.892, + 0.749 + ], + "angle": 0, + "content": "where \\( P(i) \\) is the set of positive samples, i.e., with the same label as \\( z_{i} \\), in the batch excluding \\( i \\) and \\( a \\in A \\setminus i \\) is the index of all samples in the batch excluding the \\( i \\)th sample. \\( \\tau \\) is a scalar temperature parameter used for scaling similarity values. The positive pairs are grouped into the numerator and minimizing the loss minimizes their distance in the learned representation and vice versa for negative pairs. SupCon is known to be effective for classifying samples with a single label. However, it is not straightforward to extend this for the multilabel setup as beyond positive samples, where all labels are the same, and negative samples, where none of the labels is the same, there can be a third scenario where labels are partially overlapping. Though SupCon has been extended to hierarchical classification [63], it cannot be directly extended to the true multi-label case." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.892, + 0.902 + ], + "angle": 0, + "content": "To address this issue, we learn class prototypes as the representative for each class and consider these as anchors while determining positive and negative samples. Specifically, for a specific class prototype, a representation is learned to minimize distances between the prototype and samples with this class label and maximize the distances between the prototype and samples without this class label. We compare the proposed approach with the standard single-label contrastive learning in figure 4. We iteratively update the class prototypes while learning the feature rep" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "19926" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.142, + 0.091, + 0.48, + 0.254 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.495, + 0.094, + 0.825, + 0.254 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.256, + 0.892, + 0.354 + ], + "angle": 0, + "content": "Figure 4. Contrastive learning operates on the feature space by bringing the representations of similar samples close and pushing distinct samples apart. Prior work in (a) Supervised Contrastive Learning [26] trains the network by treating instances from the same class as positive pairs and instances from different classes as negative pairs. This approach doesn't generalize to multi-label classification tasks, as some instance pairs have partially overlapping labels. We propose the use of class prototypes to enable (b) Multi-Label Prototypes Contrastive Learning. Each sample and the class prototypes corresponding to the labels associated with the sample are treated as positive pairs. Similarly, negative pairs are determined based on the missing class labels. Prototypes are represented by stars \\((\\star)\\) and inputs as circles \\((\\odot)\\) colored with all their relevant labels. We discuss strategies for initializing and learning the label prototypes in Sec. 4.2." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.373, + 0.473, + 0.508 + ], + "angle": 0, + "content": "resentations. We define \\( C = \\{c_1, \\ldots, c_K\\} \\) as the set of classes where \\( K \\) is the number of classes. For a sample \\( \\pmb{x} \\), let's define \\( P_{ml}(\\pmb{x}) = \\{c_k^+\\}, c_k^+ \\in C \\) as the set of multiple class labels associated with \\( \\pmb{x} \\) (positive classes) and \\( c_k^- \\in C \\setminus P_{ml}(\\pmb{x}) \\) denotes the missing classes (negative classes). We define \\( CP = \\{cp_1, \\ldots, cp_K\\} \\) as the set of class prototypes. Considering \\( z \\) is the representation for the sample \\( \\pmb{x} \\), the class prototypes based multilabel contrastive loss is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.078, + 0.515, + 0.48, + 0.594 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} _ {m l c} (\\pmb {x}) = \\\\ \\frac {- 1}{| P _ {m l} (\\boldsymbol {x}) |} \\sum_ {c _ {k} ^ {+} \\in P _ {m l} (\\boldsymbol {x})} \\left[ \\log \\frac {\\exp \\left(\\operatorname {s i m} \\left(\\boldsymbol {z} , \\boldsymbol {c p} _ {k}\\right) / \\tau\\right)}{\\sum_ {c _ {j} ^ {-} \\in C \\backslash P _ {m l} (\\boldsymbol {x})} \\exp \\left(\\operatorname {s i m} \\left(\\boldsymbol {z} , \\boldsymbol {c p} _ {j}\\right) / \\tau\\right)} \\right]. \\tag {2} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.595, + 0.47, + 0.685 + ], + "angle": 0, + "content": "Here, minimizing the loss of the positive class prototype and instance pairs in the numerator minimizes the distance between the representation \\( z \\) and the class prototypes corresponding to the sample, and vice versa for negative classes. We also utilize negative sampling to account for the class imbalance between positives and negatives." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.689, + 0.312, + 0.705 + ], + "angle": 0, + "content": "4.2. Learning class prototypes" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.705, + 0.471, + 0.902 + ], + "angle": 0, + "content": "We aim to learn class-specific prototypes such that the multilabel samples can be thought of as the combinations of the class prototypes selected based on the associated labels. Lets assume \\( Z_{t} \\) is an \\( N\\times d \\) matrix of \\( d \\)-dimensional representations, i.e., \\( \\mathbf{z}\\mathbf{s}) \\), of \\( N \\) samples and \\( L\\in \\{0,1\\}^{N\\times K} \\) is corresponding labels matrix with \\( K \\) classes. Let's denote \\( CP_{t} \\) is a matrix of size \\( K\\times d \\) of \\( K \\) class prototypes at a training iteration \\( t \\). Then, \\( Z_{t} = L\\times CP_{t} + \\varepsilon \\), where \\( \\varepsilon \\) is the residual noise term. Assuming a Gaussian noise that is unbiased and uncorrelated with the labels \\( L \\), the class prototypes can be approximated as \\( CP_{t}^{*}\\approx (L^{T}L)^{-1}L^{T}Z_{t} \\) where operation \\( (L^T L) \\) results in a square matrix amenable to inversion. Note that for single labels, this implies averag-" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.373, + 0.892, + 0.448 + ], + "angle": 0, + "content": "ing the features of the instances belonging to a given class as the prototype for that class. In a multi-label setup, additionally, the co-occurrence between the labels is considered. Finally, the class prototypes are updated with learning iterations as:" + }, + { + "type": "equation", + "bbox": [ + 0.574, + 0.455, + 0.892, + 0.471 + ], + "angle": 0, + "content": "\\[\nC P _ {t + 1} = \\beta \\cdot C P _ {t} + (1 - \\beta) \\cdot C P _ {t} ^ {*}, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.484, + 0.892, + 0.53 + ], + "angle": 0, + "content": "where \\(\\beta\\) is the decay parameter for the exponential moving averaging. The moving averaging avoids collapsing prototypes across the training iterations [31]." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.537, + 0.796, + 0.551 + ], + "angle": 0, + "content": "4.3. Multi-Modal Fusion Transformer" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.553, + 0.892, + 0.779 + ], + "angle": 0, + "content": "Given a sample video \\( \\mathbf{x} \\), we consider multimodal cues such as visual and audio cues to learn its representation \\( \\mathbf{z} \\). Considering multimodal cues is crucial for content recognition, specifically for education videos as effective comprehension requires attending to both visual demonstration and audio explaining the educational content [3, 40]. To capture audio cues, we consider the audio track of the video and extract speech by removing the background such as tunes or instruments [21]. Then we employ the automatic speech recognition (ASR) technique to transcribe text from the speech [43]. We notice that separating speech from the background is important for an accurate ASR transcription. Given the video frames and text transcription, we propose a multimodal transformer network (MTN) to fuse these cues using cross-modal attention." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.781, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Our MTN (Figure 5) has three components: image encoder, text encoder, and fusion encoder to learn visual \\((z^v)\\), text \\((z^t)\\), and fusion \\((z^f)\\) representations, respectively. The sample representation is comprised of these three representations \\(z = \\{z^v, z^t, z^f\\}\\). The image encoder is implemented by a vision transformer (ViT) [16] that learns frame embeddings from the video frames along with a special CLS for each frame [14]. We pool the CLS tokens across" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19927" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.1, + 0.09, + 0.447, + 0.497 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.503, + 0.47, + 0.643 + ], + "angle": 0, + "content": "Figure 5. Multi-Modal Classification Network. A text encoder is used to encode ASR text from the video, while an Image Encoder is used to get tokens representing each frame of the video. Unimodal pre-training is carried out on the text & image encoders respectively. Multi-label contrastive loss is used along with shared prototypes to align the representations across both modalities. This is followed by joint end-to-end learning of the whole multimodal network including the fusion encoder which applies multi-head self-attention within and across the modalities. The prototypes are further refined during the multi-modal training phase." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.65, + 0.47, + 0.756 + ], + "angle": 0, + "content": "the frames and consider this as a compact video representation. Similarly, we consider the BERT-based text transformer [24] to learn the word embeddings along with the text \\(CLS\\) token for the entire transcription. We consider the text \\(CLS\\) as the text representation. Finally, the fusion encoder fuses the visual and text cues by leveraging cross-modal attention between frame and word embeddings." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.763, + 0.318, + 0.779 + ], + "angle": 0, + "content": "4.4. Inference using prototypes" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.78, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Existing approaches in contrastive learning [8, 26] usually discard the MLP layer after the contrastive training, and a linear classifier is trained on the frozen backbone network. We rely on the class prototypes to carry out inference by utilizing the cosine distance between the learned prototypes and test features. Given our prototype loss-based training, the estimated probability of a given class should be proportional to the normalized temperature scaled cosine distance" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.892, + 0.138 + ], + "angle": 0, + "content": "(Equation 4). In practice, we normalize the cosine distance such that \\(-1\\) and \\(1\\) correspond to a confidence of \\(0\\& 1\\) respectively. Thus the prediction is made following:" + }, + { + "type": "equation", + "bbox": [ + 0.587, + 0.148, + 0.892, + 0.164 + ], + "angle": 0, + "content": "\\[\n\\hat {p} (k | \\boldsymbol {x}) \\propto \\exp (\\operatorname {s i m} (\\boldsymbol {z}, \\boldsymbol {c p} _ {k}) / \\tau), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.178, + 0.882, + 0.193 + ], + "angle": 0, + "content": "where \\(z\\) is the multimodal representation of the sample \\(\\pmb{x}\\)" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.199, + 0.719, + 0.214 + ], + "angle": 0, + "content": "4.5. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.219, + 0.892, + 0.385 + ], + "angle": 0, + "content": "Training strategy for the multimodal fusion: We follow a two-stage training process: during the initial unimodal training phase, we utilize fixed prototypes in each modality to align the representations. Then in the second stage, we train the unimodal encoders and the multi-modal fusion encoder end-to-end. The cross-modal alignment learned during the first stage improves the learning of the multi-modal representation. The multi-modal learning phase consists of alternating optimization steps of training the network using our contrastive loss per Sec. 4.1 and refining the class prototypes as described in Sec. 4.2." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.387, + 0.892, + 0.461 + ], + "angle": 0, + "content": "Image encoder: For video frames, Random Resized Crop and RandAugment [13] augmentations are used from torchvision. We use ImageNet pretrained vision encoders ResNet50 [19], ViT-B/32 (224×224 resolution) and ViT-B/16 (384×384 resolution) [17]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.463, + 0.893, + 0.614 + ], + "angle": 0, + "content": "Text encoder: We generate text from the videos using Whisper [43], an open source ASR model. For data augmentation, we generate four versions of the ASR text by back-translation using the Helsinki-NLP/opus-mt-{en-de, en-nl, en-fr} models through the nlpaug library [38]. Synonym replacement, text span removal and random word swapping augmentations are also used for the text data. We use DistilBERT-Base-uncased [48], and t5-small [44] from HuggingFace transformers library as the text encoder." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.615, + 0.892, + 0.735 + ], + "angle": 0, + "content": "Optimizer: We employ AdamW optimizer [36] for training with a learning rate of 0.0005. Weight decay of 1e-6 is utilized only on the MLP head during contrastive training and the classifier during BCE/Focal/Asym. loss. Since pre-trained vision and text backbones are used, the backbone learning rate is set to 1/10th of the learning rate for the head. Exponential Moving Averaging every 10 steps with a decay of 0.999 was used for the model parameters." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.745, + 0.633, + 0.761 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.764, + 0.892, + 0.869 + ], + "angle": 0, + "content": "Datasets. In addition to the proposed APPROVE dataset, we evaluate our approach on a subset of Youtube-8M [1] and COIN [55] datasets. YT-8M consists of a diverse set of YouTube videos with video and audio modalities. We consider a subset of YT-8M dataset with 46K videos and 165 classes. COIN consists of instructional videos covering a wide variety of domains and spanning over 180 classes." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.871, + 0.891, + 0.901 + ], + "angle": 0, + "content": "Baselines. We compare the efficacy of our multilabel classification framework against the following baselines:" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19928" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.136 + ], + "angle": 0, + "content": "Binary cross-entropy: In this baseline, loss for multiple labels is computed by combining the binary cross-entropy losses for individual classes." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.137, + 0.468, + 0.227 + ], + "angle": 0, + "content": "Focal loss [35]: This considers a modified binary cross-entropy to assign a higher weight to hard samples by adjusting a focusing parameter \\(\\gamma\\). Negative samples can also be down-weighted by using a weight \\(\\alpha\\). The focal loss for a positive label is given as \\(\\mathcal{L}_{focal}(p) = -\\alpha (1 - p)^{\\gamma}\\log (p)\\). We set \\(\\gamma = 2\\) and \\(\\alpha = 0.2\\) in our experiments." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.228, + 0.468, + 0.333 + ], + "angle": 0, + "content": "Asymmetric loss [47]: This builds upon the focal loss by utilizing different focusing parameters \\(\\gamma_{+}\\) and \\(\\gamma_{-}\\) for positive and negative labels. It also ignores the negative samples with a prediction probability lower than a margin \\(m\\). Asymmetric Loss for prediction \\(p\\) corresponding to label \\(y\\) is given as: \\(\\mathcal{L}_{asym}(p,y) = -yL_{+} - (1 - y)L_{-}\\), where \\(L_{+} = (1 - p)^{\\gamma_{+}}\\cdot \\log (p)\\) and" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.334, + 0.468, + 0.407 + ], + "angle": 0, + "content": "\\(L_{-} = (\\max (p - m,0))^{\\gamma_{-}}\\cdot \\log (1 - \\max (p - m,0))\\) We follow the 5-step procedure recommended by the original authors to train this baseline. We experimentally set \\(\\{\\gamma_{-} = 2,\\gamma_{+} = 1,m = 0.1\\}\\) corresponding to the best performance on APPROVE." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.409, + 0.468, + 0.558 + ], + "angle": 0, + "content": "Metrics. In order to develop a reliable education content detection framework, achieving high precision is crucial. Thus we consider Recall @ 80% Precision (R@80) as the primary metric. We also consider the standard area under the precision-recall curve (AUPR) that is not sensitive to a specific threshold for making the final prediction. We also consider a label ranking average precision (LRAP) [56] metric that is more suitable for the multilabel setup. This estimates whether the ground truth classes are predicted with higher scores than the rest:" + }, + { + "type": "equation", + "bbox": [ + 0.078, + 0.563, + 0.466, + 0.603 + ], + "angle": 0, + "content": "\\[\n\\mathrm {L R A P} = \\frac {1}{n} \\sum_ {i = 1} ^ {m} \\frac {1}{| Y _ {i} |} \\sum_ {\\lambda \\in Y _ {i}} \\frac {| \\lambda^ {\\prime} \\in Y _ {i} : r n k _ {i} (\\lambda^ {\\prime}) \\leq r n k _ {i} (\\lambda) |}{r n k _ {i} (\\lambda)},\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.614, + 0.468, + 0.644 + ], + "angle": 0, + "content": "where \\( r n k_{i}(\\lambda) \\) is the predicted rank of class \\( \\lambda \\) for sample \\( i \\). LRAP is a ranking metric and in independent of a threshold." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.644, + 0.468, + 0.81 + ], + "angle": 0, + "content": "Results on APPROVE. We compare the proposed approach with the baselines in Tab. 2. Our approach outperforms the strongest baselines by \\(3.1\\%\\) and \\(2.3\\%\\) with respect to R@80 and AUPR, respectively. We also present results for separate models trained on Math and Literacy subsets of Approve respectively. Results on the Math subset are higher compared to the Literacy subset, which indicates the literacy classes are harder to distinguish mostly due to the high inter-class similarity. The top three hardest classes are follow_words, letters_in_words, and sounds_in_words and these are from the literacy set." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.811, + 0.468, + 0.885 + ], + "angle": 0, + "content": "Results on YT-46K and COIN. Beyond the APPROVE dataset, we also test our approach on two public datasets: YT-46K and COIN. Here we provide the results for the multi-modal models and the results for the single-modality models are in Section D of the Supplementary Material." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.886, + 0.468, + 0.901 + ], + "angle": 0, + "content": "Results on YT-46K are provided in Table 3. As YT-8M" + }, + { + "type": "table", + "bbox": [ + 0.503, + 0.089, + 0.889, + 0.386 + ], + "angle": 0, + "content": "
SubsetModalityMethodAUPRLRAPR@80
AllVBCE45.554.36.9
Focal45.956.615.0
Ours46.757.919.6
TBCE79.885.163.3
Focal79.985.772.8
Ours82.587.475.4
V+TBCE84.388.476.3
Focal [35]86.189.182.2
Asym. [47]86.089.282.4
Ours88.4 +2.390.7 +1.585.5 +3.1
MTHV+TBCE86.392.480.3
Focal87.292.182.4
Ours88.4 +1.293.2 +1.183.2 +0.8
LITV+TBCE72.182.950.7
Focal72.783.550.9
Ours73.6 +0.984.7 +1.254.7 +3.8
" + }, + { + "type": "table_caption", + "bbox": [ + 0.5, + 0.39, + 0.851, + 0.419 + ], + "angle": 0, + "content": "Table 2. Results on APPROVE dataset. All metrics in %. V \\(\\rightarrow\\) Video & T \\(\\rightarrow\\) Text. M \\(\\rightarrow\\) Math & L \\(\\rightarrow\\) Literacy Subsets." + }, + { + "type": "table", + "bbox": [ + 0.529, + 0.427, + 0.858, + 0.504 + ], + "angle": 0, + "content": "
ModalityMethodAUPRLRAPR@80
V+TBCE64.670.242.3
V+TFocal [35]69.772.744.6
V+TOurs70.9 +1.274.9 +2.249.1 +4.5
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.507, + 0.878, + 0.521 + ], + "angle": 0, + "content": "Table 3. Results on YT-46K. V→Video Frames and T→Text." + }, + { + "type": "table", + "bbox": [ + 0.554, + 0.53, + 0.835, + 0.637 + ], + "angle": 0, + "content": "
ModalityMethodTop-1 Accuracy
V+TCE53.7
V+TBCE54.9
V+TFocal [35]56.1
V+TSupCon [26]54.7
V+TOurs57.5 +1.4
" + }, + { + "type": "table_caption", + "bbox": [ + 0.518, + 0.64, + 0.872, + 0.654 + ], + "angle": 0, + "content": "Table 4. Results on COIN. V→Video Frames and T→Text." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.661, + 0.892, + 0.812 + ], + "angle": 0, + "content": "was primarily collected with the intention of visual classification, the additional use of text data leads to a smaller improvement compared to APPROVE. Results on COIN are provided in Table 4. As each video from COIN is mapped to a single task. Thus, we consider the Top-1 accuracy as the metric. On COIN we compare our approach with SupCon [26] which is effective for single labels. Note that our approach outperforms SupCon and this justifies the effectiveness of the class prototypes based training in a generic contrastive learning framework." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.824, + 0.665, + 0.838 + ], + "angle": 0, + "content": "5.1. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.892, + 0.901 + ], + "angle": 0, + "content": "We perform the following ablation studies to quantify the impact of our approach for learning class prototypes, the multimodal fusion module, and the choice of visual and text encoding frameworks." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19929" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.121, + 0.094, + 0.235, + 0.106 + ], + "angle": 0, + "content": "(a) Class Prototypes" + }, + { + "type": "table", + "bbox": [ + 0.084, + 0.107, + 0.27, + 0.173 + ], + "angle": 0, + "content": "
VariantAPPROVECOIN
Random84.156.6
Orthogonal84.857.0
Learned85.557.5
Hierarchical86.057.8
" + }, + { + "type": "table_caption", + "bbox": [ + 0.308, + 0.094, + 0.444, + 0.105 + ], + "angle": 0, + "content": "(b) Fusion Encoder Size" + }, + { + "type": "table", + "bbox": [ + 0.299, + 0.107, + 0.455, + 0.161 + ], + "angle": 0, + "content": "
LayersAPPROVECOIN
184.957.1
285.557.5
485.357.6
" + }, + { + "type": "table_caption", + "bbox": [ + 0.521, + 0.094, + 0.627, + 0.105 + ], + "angle": 0, + "content": "(c) Vision Encoder" + }, + { + "type": "table", + "bbox": [ + 0.477, + 0.107, + 0.671, + 0.161 + ], + "angle": 0, + "content": "
Vision ModelAPPROVECOIN
R5084.855.2
ViT-B/3285.557.5
ViT-B/1683.857.8
" + }, + { + "type": "table_caption", + "bbox": [ + 0.725, + 0.094, + 0.821, + 0.105 + ], + "angle": 0, + "content": "(d) Text Encoder" + }, + { + "type": "table", + "bbox": [ + 0.68, + 0.107, + 0.865, + 0.148 + ], + "angle": 0, + "content": "
Text ModelAPPROVECOIN
DistilBert-B85.557.5
t5-S87.357.9
" + }, + { + "type": "table_caption", + "bbox": [ + 0.13, + 0.175, + 0.839, + 0.188 + ], + "angle": 0, + "content": "Table 5. Ablation studies. R@80% Precision for APPROVE and Top-1 Accuracy for COIN. Default setup is underlined." + }, + { + "type": "table_caption", + "bbox": [ + 0.16, + 0.197, + 0.276, + 0.209 + ], + "angle": 0, + "content": "(a) Noisy modalities." + }, + { + "type": "table", + "bbox": [ + 0.13, + 0.211, + 0.307, + 0.277 + ], + "angle": 0, + "content": "
% missingAPPROVECOIN
0%85.557.5
10% V80.153.3
10% T75.857.2
30% T68.942.8
" + }, + { + "type": "table_caption", + "bbox": [ + 0.411, + 0.197, + 0.553, + 0.209 + ], + "angle": 0, + "content": "(b) Run-to-Run Variance" + }, + { + "type": "table", + "bbox": [ + 0.388, + 0.211, + 0.576, + 0.265 + ], + "angle": 0, + "content": "
MethodAPPROVECOIN
BCE76.3 ± 0.754.9 ± 0.6
Focal82.2 ± 0.556.1 ± 0.3
Ours85.5 ± 0.557.5 ± 0.8
" + }, + { + "type": "table_caption", + "bbox": [ + 0.702, + 0.197, + 0.795, + 0.209 + ], + "angle": 0, + "content": "(c) Initialization" + }, + { + "type": "table", + "bbox": [ + 0.65, + 0.211, + 0.847, + 0.271 + ], + "angle": 0, + "content": "
MethodAPPROVECOIN
ImageNet & Wiki-en+TBC85.557.5
CLIP86.763.5
" + }, + { + "type": "table_caption", + "bbox": [ + 0.101, + 0.281, + 0.866, + 0.295 + ], + "angle": 0, + "content": "Table 6. Robustness analysis. Our method is robust to partially missing modality and has similar run-to-run variance as baselines." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.305, + 0.47, + 0.546 + ], + "angle": 0, + "content": "Learning class prototypes: We compare the two strategies where after initializing the class prototypes, we 1) keep them fixed and learn only the multimodal embedding of the samples, and 2) class prototypes and sample embedding are learned iteratively. The initialization can be done either randomly, or with orthogonal constraints. We note that orthogonal initialization performs best in our experiments and iterative adjusting the class prototypes achieves better performance as shown in Table 5 (a). We also consider hierarchical prototypes, for APPROVE, using a 2-level hierarchy where the first level consists of 18 classes, and the second level is the 3 super-classes: math, literacy, and background. The 180 task categories of COIN are organized into 12 domains in the taxonomy provided with the dataset. This hierarchy imposes an additional constraint on learning the embeddings during training." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.547, + 0.47, + 0.607 + ], + "angle": 0, + "content": "Fusion Encoder: This evaluates the effect of the number of layers in the fusion encoder on the final performance. As expected, the performance improves with more layers and saturates around 4 layers (Table 5 (b))." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.608, + 0.469, + 0.698 + ], + "angle": 0, + "content": "Vision Encoder: This evaluates the effect of the image encoder on the final performance. We consider ResNet [19] and ViT variants [16]. We notice that the ViT-B/16-384 encoder works well for the larger COIN dataset, whereas the ViT-B/32-224 encoder works best for APPROVE (Table 5 (c))." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.699, + 0.469, + 0.805 + ], + "angle": 0, + "content": "Text Encoder: This evaluates the effect of the text encoder on the final performance. We test DistilBERT and T5 backbones for the text encoder. BERT is trained to predict masked spans of text. T5's unsupervised objective is similar, however, it trains on predicting the entire sequence instead of just the masked spans. GPT2 takes an auto-regressive approach to language modeling (Table 5 (d))." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.809, + 0.266, + 0.824 + ], + "angle": 0, + "content": "5.2. Robustness analysis" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.826, + 0.44, + 0.84 + ], + "angle": 0, + "content": "We evaluate the robustness of our approach as follows:" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.47, + 0.901 + ], + "angle": 0, + "content": "Noisy modality: YouTube videos may have noisy modalities where some of the video frames are missing or ASR transcription is noisy. We show that our approach is robust against the cases where a percentage of video frames or text" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.305, + 0.766, + 0.32 + ], + "angle": 0, + "content": "words are missing as shown in Tab 6(a)." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.321, + 0.892, + 0.366 + ], + "angle": 0, + "content": "Run-to-Run variance: The low variance across runs (Tab 6(b)) indicates that our approach is not sensitive to random initialization of the class prototypes." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.367, + 0.893, + 0.503 + ], + "angle": 0, + "content": "Initializing the encoders: We consider the ImageNet pretraining for the image encoder, while English Wikipedia + Toronto Book Corpus is used to pre-train the text encoder. We provide results where the backbones are initialized with CLIP [42], which provides a more aligned vision-text representation. As expected, the results are better with the CLIP initialization (Tab 6(c)). The improvements are more significant on COIN than APPROVE as CLIP models may not be exposed to educational videos." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.512, + 0.619, + 0.527 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.531, + 0.892, + 0.757 + ], + "angle": 0, + "content": "We have proposed an approach for detecting educational content in online videos. The problem is formulated as a fine-grained multilabel video classification task and we have considered class-prototypes based contrastive learning to address this. We have employed a multimodal transformer network to fuse visual and audio cues. This is crucial for comprehending educational content as both visual and audio cues are to be aligned to ensure effective comprehension. Our approach is shown to be effective in distinguishing fine-grained educational content with high interclass similarity. We have introduced APPROVE - a dataset with 193 hours of expert-annotated educational videos. Beyond APPROVE, we have evaluated our approach on COIN and YouTube-8M datasets where our approach outperforms the competitive baselines." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.762, + 0.667, + 0.777 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.778, + 0.892, + 0.818 + ], + "angle": 0, + "content": "This work is supported by NSF grant # 2139219 and SRI R&D award. Rohit Gupta is supported by ARO grant W911NF-19-1-0356." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.819, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Views and conclusions contained herein are those of the authors and should not be interpreted as representing the official policies or endorsements, either expressed or implied, of NSF, ARO, IARPA, DOI/IBC, or the U.S. Government. The U.S. Government is authorized to reproduce and distribute reprints for Governmental purposes notwithstanding any copyright annotation thereon." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19930" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.116, + 0.47, + 0.185 + ], + "angle": 0, + "content": "[1] Sami Abu-El-Haija, Nisarg Kothari, Joonseok Lee, Paul Natsev, George Toderici, Balakrishnan Varadarajan, and Sudheendra Vijayanarasimhan. YouTube-8M: A large-scale video classification benchmark. arXiv preprint arXiv:1609.08675, 2016. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.187, + 0.472, + 0.257 + ], + "angle": 0, + "content": "[2] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katie Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. arXiv preprint arXiv:2204.14198, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.259, + 0.471, + 0.286 + ], + "angle": 0, + "content": "[3] National Governors Association et al. Common core state standards. Washington, DC, 2010. 1, 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.288, + 0.471, + 0.33 + ], + "angle": 0, + "content": "[4] Brooke Auxier, Monica Anderson Andrew Perrin, and Erica Turner. Parenting children in the age of screens. Technical report, Pew Research Center, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.332, + 0.47, + 0.415 + ], + "angle": 0, + "content": "[5] Philip Bachman, R Devon Hjelm, and William Buchwalter. Learning representations by maximizing mutual information across views. In H. Wallach, H. Larochelle, A. Beygelzimer, F. d Alché-Buc, E. Fox, and R. Garnett, editors, Advances in Neural Information Processing Systems, volume 32. Curran Associates, Inc., 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.417, + 0.47, + 0.472 + ], + "angle": 0, + "content": "[6] Max Bain, Arsha Nagrani, Gül Varol, and Andrew Zisserman. Frozen in time: A joint video and image encoder for end-to-end retrieval. In IEEE International Conference on Computer Vision, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.474, + 0.47, + 0.528 + ], + "angle": 0, + "content": "[7] Johanna Burkhardt and Wolfgang Lenhard. A meta-analysis on the longitudinal, age-dependent effects of violent video games on aggression. Media Psychology, 25(3):499-512, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.531, + 0.47, + 0.587 + ], + "angle": 0, + "content": "[8] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International conference on machine learning, pages 1597-1607. PMLR, 2020. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.589, + 0.47, + 0.683 + ], + "angle": 0, + "content": "[9] Ting Chen, Simon Kornblith, Kevin Swersky, Mohammad Norouzi, and Geoffrey E Hinton. Big self-supervised models are strong semi-supervised learners. In H. Larochelle, M. Ranzato, R. Hadsell, M. F. Balcan, and H. Lin, editors, Advances in Neural Information Processing Systems, volume 33, pages 22243-22255. Curran Associates, Inc., 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.687, + 0.468, + 0.728 + ], + "angle": 0, + "content": "[10] Xinlei Chen, Haoqi Fan, Ross Girshick, and Kaiming He. Improved baselines with momentum contrastive learning, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.731, + 0.468, + 0.786 + ], + "angle": 0, + "content": "[11] Xinlei Chen, Hao Fang, Tsung-Yi Lin, Ramakrishna Vedantam, Saurabh Gupta, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco captions: Data collection and evaluation server. arXiv preprint arXiv:1504.00325, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.788, + 0.468, + 0.829 + ], + "angle": 0, + "content": "[12] Xinlei Chen, Saining Xie, and Kaiming He. An empirical study of training self-supervised vision transformers, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.831, + 0.471, + 0.902 + ], + "angle": 0, + "content": "[13] Ekin D Cubuk, Barret Zoph, Jonathon Shlens, and Quoc V Le. Randaugment: Practical automated data augmentation with a reduced search space. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops, pages 702-703, 2020. 6" + }, + { + "type": "list", + "bbox": [ + 0.079, + 0.116, + 0.472, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.147 + ], + "angle": 0, + "content": "[14] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.149, + 0.892, + 0.204 + ], + "angle": 0, + "content": "[15] Ali Diba, Mohsen Fayyaz, Vivek Sharma, Manohar Paluri, Jürgen Gall, Rainer Stiefelhagen, and Luc Van Gool. Large scale holistic video understanding. In European Conference on Computer Vision, pages 593-610. Springer, 2020. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.205, + 0.892, + 0.274 + ], + "angle": 0, + "content": "[16] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. ICLR, 2021. 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.275, + 0.892, + 0.37 + ], + "angle": 0, + "content": "[17] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.372, + 0.892, + 0.44 + ], + "angle": 0, + "content": "[18] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9729-9738, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.442, + 0.892, + 0.496 + ], + "angle": 0, + "content": "[19] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.498, + 0.892, + 0.552 + ], + "angle": 0, + "content": "[20] Lowry Hemphill and Terrence Tivnan. The importance of early vocabulary for literacy achievement in high-poverty schools. Journal of Education for Students Placed at Risk, 13(4):426-451, 2008. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.553, + 0.892, + 0.608 + ], + "angle": 0, + "content": "[21] Romain Hennequin, Anis Khlif, Felix Voituret, and Manuel Moussallam. Spleeter: a fast and efficient music source separation tool with pre-trained models. Journal of Open Source Software, 5(50):2154, 2020. Deezer Research. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.609, + 0.892, + 0.65 + ], + "angle": 0, + "content": "[22] Lisa B Hurwitz. Getting a read on ready to learn media: A meta-analytic review of effects on literacy. *Child Development*, 90(5):1754–1771, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.651, + 0.892, + 0.705 + ], + "angle": 0, + "content": "[23] Lisa B Hurwitz and Kelly L Schmitt. Raising readers with ready to learn: A six-year follow-up to an early educational computer game intervention. Computers in Human Behavior, 104:106176, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.707, + 0.892, + 0.774 + ], + "angle": 0, + "content": "[24] Andrew Jaegle, Felix Gimeno, Andy Brock, Oriol Vinyals, Andrew Zisserman, and Joao Carreira. Perceiver: General perception with iterative attention. In International conference on machine learning, pages 4651-4664. PMLR, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.776, + 0.892, + 0.831 + ], + "angle": 0, + "content": "[25] Nancy C Jordan, David Kaplan, Chaitanya Ramineni, and Maria N Locuniak. Early math matters: kindergarten number competence and later mathematics outcomes. Developmental psychology, 45(3):850, 2009. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.832, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[26] Prannay Khosla, Piotr Teterwak, Chen Wang, Aaron Sarna, Yonglong Tian, Phillip Isola, Aaron Maschinot, Ce Liu, and Dilip Krishnan. Supervised contrastive learning. Advances in Neural Information Processing Systems, 33:18661-18673, 2020. 2, 4, 5, 6, 7" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "19931" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.161 + ], + "angle": 0, + "content": "[27] Hildegard Kuehne, Hueihan Jhuang, Estíbaliz Garrote, Tomaso Poggio, and Thomas Serre. Hmdb: a large video database for human motion recognition. In 2011 International conference on computer vision, pages 2556-2563. IEEE, 2011. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.162, + 0.47, + 0.217 + ], + "angle": 0, + "content": "[28] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In ICML, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.218, + 0.47, + 0.273 + ], + "angle": 0, + "content": "[29] Junnan Li, Ramprasaath R. Selvaraju, Akhilesh Deepak Gotmare, Shafiq Joty, Caiming Xiong, and Steven Hoi. Align before fuse: Vision and language representation learning with momentum distillation. In NeurIPS, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.274, + 0.47, + 0.328 + ], + "angle": 0, + "content": "[30] Junnan Li, Pan Zhou, Caiming Xiong, and Steven Hoi. Prototypical contrastive learning of unsupervised representations. In International Conference on Learning Representations, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.33, + 0.47, + 0.371 + ], + "angle": 0, + "content": "[31] Junnan Li, Pan Zhou, Caiming Xiong, and Steven C.H. Hoi. Prototypical contrastive learning of unsupervised representations. In ICLR, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.371, + 0.47, + 0.44 + ], + "angle": 0, + "content": "[32] Shikun Li, Xiaobo Xia, Shiming Ge, and Tongliang Liu. Selective-supervised contrastive learning with noisy labels. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 316-325, June 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.441, + 0.47, + 0.51 + ], + "angle": 0, + "content": "[33] Xiujun Li, Xi Yin, Chunyuan Li, Xiaowei Hu, Pengchuan Zhang, Lei Zhang, Lijuan Wang, Houdong Hu, Li Dong, Furu Wei, Yejin Choi, and Jianfeng Gao. Oscar: Object-semantics aligned pre-training for vision-language tasks. ECCV 2020, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.511, + 0.47, + 0.58 + ], + "angle": 0, + "content": "[34] Yixuan Li, Lei Chen, Runyu He, Zhenzhi Wang, Gangshan Wu, and Limin Wang. Multisports: A multi-person video dataset of spatio-temporally localized sports actions. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13536-13545, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.58, + 0.47, + 0.636 + ], + "angle": 0, + "content": "[35] Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dólár. Focal loss for dense object detection. In Proceedings of the IEEE international conference on computer vision, pages 2980-2988, 2017. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.636, + 0.47, + 0.678 + ], + "angle": 0, + "content": "[36] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In International Conference on Learning Representations, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.678, + 0.47, + 0.719 + ], + "angle": 0, + "content": "[37] Ziyang Luo, Yadong Xi, Rongsheng Zhang, and Jing Ma. A frustratingly simple approach for end-to-end image captioning, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.72, + 0.47, + 0.748 + ], + "angle": 0, + "content": "[38] Edward Ma. Nlp augmentation. https://github.com/makcedward/nlpaug, 2019.6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.749, + 0.47, + 0.789 + ], + "angle": 0, + "content": "[39] Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.79, + 0.47, + 0.844 + ], + "angle": 0, + "content": "[40] Andrew Porter, Jennifer McMaken, Jun Hwang, and Rui Yang. Common core standards: The new us intended curriculum. Educational researcher, 40(3):103-116, 2011. 1, 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.845, + 0.47, + 0.901 + ], + "angle": 0, + "content": "[41] J. S. Radesky, A. Schaller, S. L. Yeo, H. M. Weeks, and M. B. Robb. Young kids and youtube: How ads, toys, and games dominate viewing. common sense media. Technical report, Common Sense Media, 2020. 2, 3" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.092, + 0.893, + 0.176 + ], + "angle": 0, + "content": "[42] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.179, + 0.893, + 0.235 + ], + "angle": 0, + "content": "[43] Alec Radford, Jong Wook Kim, Tao Xu, Greg Brockman, Christine McLeavey, and Ilya Sutskever. Robust speech recognition via large-scale weak supervision. Technical report, Tech. Rep., Technical report, OpenAI, 2022. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.236, + 0.893, + 0.306 + ], + "angle": 0, + "content": "[44] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, Peter J Liu, et al. Exploring the limits of transfer learning with a unified text-to-text transformer. J. Mach. Learn. Res., 21(140):1-67, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.309, + 0.893, + 0.338 + ], + "angle": 0, + "content": "[45] Victoria Rideout and Michael B. Robb. The common sense census: Media use by tweens and teens, Oct 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.34, + 0.893, + 0.382 + ], + "angle": 0, + "content": "[46] V Rideout and M. B. Robb. The common sense census: Media use by tweens and teens. common sense media. Technical report, Common Sense Media, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.384, + 0.893, + 0.455 + ], + "angle": 0, + "content": "[47] Tal Ridnik, Emanuel Ben-Baruch, Nadav Zamir, Asaf Noy, Itamar Friedman, Matan Protter, and Lihi Zelnik-Manor. Asymmetric loss for multi-label classification. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 82-91, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.457, + 0.893, + 0.512 + ], + "angle": 0, + "content": "[48] Victor Sanh, Lysandre Debut, Julien Chaumont, and Thomas Wolf. Distilbert, a distilled version of bert: smaller, faster, cheaper and lighter. arXiv preprint arXiv:1910.01108, 2019.6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.515, + 0.893, + 0.583 + ], + "angle": 0, + "content": "[49] Dian Shao, Yue Zhao, Bo Dai, and Dahua Lin. Finegym: A hierarchical video dataset for fine-grained action understanding. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2616-2625, 2020. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.587, + 0.893, + 0.685 + ], + "angle": 0, + "content": "[50] Piyush Sharma, Nan Ding, Sebastian Goodman, and Radu Soricut. Conceptual captions: A cleaned, hypernymed, image alt-text dataset for automatic image captioning. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 2556-2565, Melbourne, Australia, July 2018. Association for Computational Linguistics. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.687, + 0.893, + 0.771 + ], + "angle": 0, + "content": "[51] Nina Shvetsova, Brian Chen, Andrew Rouditchenko, Samuel Thomas, Brian Kingsbury, Rogerio S Feris, David Harwath, James Glass, and Hilde Kuehne. Everything at once-multimodal fusion transformer for video retrieval. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20020-20029, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.773, + 0.893, + 0.857 + ], + "angle": 0, + "content": "[52] Gunnar A. Sigurdsson, Gül Varol, Xiaolong Wang, Ali Farhadi, Ivan Laptev, and Abhinav Gupta. Hollywood in homes: Crowdsourcing data collection for activity understanding. In Bastian Leibe, Jiri Matas, Nicu Sebe, and Max Welling, editors, Computer Vision – ECCV 2016, pages 510–526, Cham, 2016. Springer International Publishing. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.859, + 0.893, + 0.901 + ], + "angle": 0, + "content": "[53] Lucas Smaira, João Carreira, Eric Noland, Ellen Clancy, Amy Wu, and Andrew Zisserman. A short note on the kinetics-700-2020 human action dataset, 2020. 1, 4" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.092, + 0.893, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "19932" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.468, + 0.134 + ], + "angle": 0, + "content": "[54] Khurram Soomro, Amir Roshan Zamir, and Mubarak Shah. Ucf101: A dataset of 101 human actions classes from videos in the wild. arXiv preprint arXiv:1212.0402, 2012. 1, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.134, + 0.468, + 0.203 + ], + "angle": 0, + "content": "[55] Yansong Tang, Dajun Ding, Yongming Rao, Yu Zheng, Danyang Zhang, Lili Zhao, Jiwen Lu, and Jie Zhou. Coin: A large-scale dataset for comprehensive instructional video analysis. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.203, + 0.468, + 0.244 + ], + "angle": 0, + "content": "[56] Grigorios Tsoumakas, Ioannis Katakis, and Ioannis Vla-havas. Mining Multi-label Data, pages 667-685. Springer US, Boston, MA, 2010. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.245, + 0.469, + 0.313 + ], + "angle": 0, + "content": "[57] Wenguan Wang, Tianfei Zhou, Fisher Yu, Jifeng Dai, Ender Konukoglu, and Luc Van Gool. Exploring cross-image pixel contrast for semantic segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 7303-7313, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.314, + 0.469, + 0.396 + ], + "angle": 0, + "content": "[58] Jinyu Yang, Jiali Duan, Son Tran, Yi Xu, Sampath Chanda, Liquun Chen, Belinda Zeng, Trishul Chilimbi, and Junzhou Huang. Vision-language pre-training with triple contrastive learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15671-15680, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.397, + 0.469, + 0.466 + ], + "angle": 0, + "content": "[59] Peter Young, Alice Lai, Micah Hodosh, and Julia Hockenmaier. From image descriptions to visual denotations: New similarity metrics for semantic inference over event descriptions. Transactions of the Association for Computational Linguistics, 2:67-78, 2014. 3" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.469, + 0.466 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.092, + 0.892, + 0.188 + ], + "angle": 0, + "content": "[60] Rowan Zellers, Jiasen Lu, Ximing Lu, Youngjae Yu, Yanpeng Zhao, Mohammadreza Salehi, Aditya Kusupati, Jack Hessel, Ali Farhadi, and Yejin Choi. Merlot reserve: Neural script knowledge through vision and language and sound. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 16375-16387, June 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.192, + 0.892, + 0.261 + ], + "angle": 0, + "content": "[61] Rowan Zellers, Ximing Lu, Jack Hessel, Youngjae Yu, Jae Sung Park, Jize Cao, Ali Farhadi, and Yejin Choi. Merlot: Multimodal neural script knowledge models. Advances in Neural Information Processing Systems, 34:23634-23651, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.265, + 0.892, + 0.321 + ], + "angle": 0, + "content": "[62] Pengchuan Zhang, Xiujun Li, Xiaowei Hu, Jianwei Yang, Lei Zhang, Lijuan Wang, Yejin Choi, and Jianfeng Gao. Vinvl: Making visual representations matter in vision-language models. CVPR 2021, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.324, + 0.892, + 0.393 + ], + "angle": 0, + "content": "[63] Shu Zhang, Ran Xu, Caiming Xiong, and Chetan Ramaiah. Use all the labels: A hierarchical multi-label contrastive learning framework. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 16660-16669, June 2022. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.396, + 0.892, + 0.465 + ], + "angle": 0, + "content": "[64] Chen Zhu, Xiao Tan, Feng Zhou, Xiao Liu, Kaiyu Yue, Errui Ding, and Yi Ma. Fine-grained video categorization with redundancy reduction attention. In Proceedings of the European Conference on Computer Vision (ECCV), pages 136-152, 2018. 4" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.092, + 0.892, + 0.465 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.956 + ], + "angle": 0, + "content": "19933" + } + ] +] \ No newline at end of file diff --git a/2023/Class Prototypes Based Contrastive Learning for Classifying Multi-Label and Fine-Grained Educational Videos/bbbb89aa-912d-4dd5-8424-10e3f0db2ed4_origin.pdf b/2023/Class Prototypes Based Contrastive Learning for Classifying Multi-Label and Fine-Grained Educational Videos/bbbb89aa-912d-4dd5-8424-10e3f0db2ed4_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..c6f0214e09fa30b49079a97dd64c2d1978f10783 --- /dev/null +++ b/2023/Class Prototypes Based Contrastive Learning for Classifying Multi-Label and Fine-Grained Educational Videos/bbbb89aa-912d-4dd5-8424-10e3f0db2ed4_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c7c0f4320d8ae2b1de268c7d8c7ec67877c59614c7386dbcb6aec7cec17c045c +size 870281 diff --git a/2023/Class Prototypes Based Contrastive Learning for Classifying Multi-Label and Fine-Grained Educational Videos/full.md b/2023/Class Prototypes Based Contrastive Learning for Classifying Multi-Label and Fine-Grained Educational Videos/full.md new file mode 100644 index 0000000000000000000000000000000000000000..1affdc3c740cba3ab224c36d50635b2a5a73006a --- /dev/null +++ b/2023/Class Prototypes Based Contrastive Learning for Classifying Multi-Label and Fine-Grained Educational Videos/full.md @@ -0,0 +1,359 @@ +# Class Prototypes based Contrastive Learning for Classifying Multi-Label and Fine-Grained Educational Videos + +Rohit Gupta $^{1}$ * Anirban Roy $^{2}$ Claire Christensen $^{2}$ Sukeong Kim $^{2}$ Sarah Gerard $^{2}$ Madeline Cincebeaux $^{2}$ Ajay Divakaran $^{2}$ Todd Grindal $^{2}$ Mubarak Shah $^{1}$ + +1 Center for Research in Computer Vision, University of Central Florida rohitg@knights.ucf.edu, shah@crcv.ucf.edu + +2 SRI International anirban.roy@sri.com + +# Abstract + +The recent growth in the consumption of online media by children during early childhood necessitates data-driven tools enabling educators to filter out appropriate educational content for young learners. This paper presents an approach for detecting educational content in online videos. We focus on two widely used educational content classes: literacy and math. For each class, we choose prominent codes (sub-classes) based on the Common Core Standards. For example, literacy codes include 'letter names', 'letter sounds', and math codes include 'counting', 'sorting'. We pose this as a fine-grained multilabel classification problem as videos can contain multiple types of educational content and the content classes can get visually similar (e.g., 'letter names' vs 'letter sounds'). We propose a novel class prototypes based supervised contrastive learning approach that can handle fine-grained samples associated with multiple labels. We learn a class prototype for each class and a loss function is employed to minimize the distances between a class prototype and the samples from the class. Similarly, distances between a class prototype and the samples from other classes are maximized. As the alignment between visual and audio cues are crucial for effective comprehension, we consider a multimodal transformer network to capture the interaction between visual and audio cues in videos while learning the embedding for videos. For evaluation, we present a dataset, APPROVE, employing educational videos from YouTube labeled with fine-grained education classes by education researchers. APPROVE consists of 193 hours of expert-annotated videos with 19 classes. The proposed approach outperforms strong baselines on APPROVE and other benchmarks such as Youtube-8M, and COIN. The dataset is available at https://nusci.csl.sri.com/project/APPROVE. + +# 1. Introduction + +With the expansion of internet access and the ubiquitous availability of smart devices, children increasingly spend a significant amount of time watching online videos. A recent nationally representative survey reported that $89\%$ of parents of children aged 11 or younger say their child watches videos on YouTube [4]. Moreover, it is estimated that young children in the age range of two to four years consume 2.5 hours and five to eight years consume 3.0 hours per day on average [45, 46]. Childhood is typically a key period for education, especially for learning basic skills such as literacy and math [20, 25]. Unlike generic online videos, watching appropriate educational videos supports healthy child development and learning [7, 22, 23]. Thus, analyzing the content of these videos may help parents, teachers, and media developers increase young children's exposure to high-quality education videos, which has been shown to produce meaningful learning gains [22]. As the amount of online content produced grows exponentially, automated content understanding methods are essential to facilitate this. + +In this work, given a video, our goal is to determine whether the video contains any educational content and characterize the content. Detecting educational content requires identifying multiple distinct types of content in a video while distinguishing between similar content types. The task is challenging as the education codes by Common Core Standards [3, 40] can be similar such as 'letter names' and 'letter sounds', where the former focuses on the name of the letter and the latter is based on the phonetic sound of the letter. Also, understanding education content requires analyzing both visual and audio cues simultaneously as both signals are to be present to ensure effective learning [3, 40]. This is in contrast to standard video classification benchmarks such as the sports or generic YouTube videos in UCF101 [54] Kinetics400 [53], YouTube-8M [1], where visual cues are often sufficient to detect the different classes. Finally, unlike standard well-known action videos, + +education codes are more structured and not accessible to common users. Thus, it requires a carefully curated set of videos and expert annotations to create a dataset to enable a data-driven approach. In this work, we focus on two widely used educational content classes: literacy and math. For each class, we choose prominent codes (sub-classes) based on the Common Core Standards that outline age-appropriate learning standards [3,40]. For example, literacy codes include 'letter names', 'letter sounds', 'rhyming', and math codes include 'counting', 'addition subtraction', 'sorting', 'analyze shapes'. + +We formulate the problem as a multilabel fine-grained video classification task as a video may contain multiple types of content that can be similar. We employ multimodal cues since besides visual cues, audio cues provide important cues to distinguish between similar types of educational content. We propose a class prototypes based supervised contrastive learning approach to address the above-mentioned challenges. We learn a prototype embedding for each class. Then a loss function is employed to minimize the distance between a class prototype and the samples associated with the class label. Similarly, the distance between a class prototype and the samples without that class label is maximized. This is unlike the standard supervised contrastive learning setup where inter-class distance is maximized and intra-class distance is minimized by considering classwise positive and negative samples. This approach is shown to be effective for single-label setups [26]. However, it is not straightforward to extend this for the proposed multilabel setup as samples cannot be identified as positive or negative due to the multiple labels. We jointly learn the embedding of the class prototypes and the samples. The embeddings are learned by a multimodal transformer network (MTN) that captures the interaction between visual and audio cues in videos. We employ automatic speech recognition (ASR) to transcribe text from the audio. The MTN consists of video and text encoders that learn modality-specific embedding and a cross-attention mechanism is employed to capture the interaction between them. The MTN is end-to-end learned through the contrastive loss. + +Due to the lack of suitable datasets for evaluating fine-grained classification of education videos, we propose a new dataset, called APPROVE, of curated YouTube videos annotated with educational content. We follow Common Core Standards [3, 40] to select education content suitable for the kindergarten level. We consider two high-level classes of educational content: literacy and math. For each of these content classes, we select a set of codes. For the literacy class, we select 7 codes and for the math class, we select 11 codes. Each video is associated with multiple labels corresponding to these codes. The videos are annotated by trained education researchers following standard validation protocol [41] to ensure correctness. APPROVE also con + +sists of carefully chosen background videos, i.e., without educational content, that are visually similar to the videos with educational content. APPROVE consists of 193 hours of expert-annotated videos with 19 classes (7 literacy codes, 11 math codes, and a background) where each video has 3 labels on average. + +Our contributions can be summarized as follows: + +- APPROVE, a fine-grained multi-label dataset of education videos, to promote exploration in this field. +- Class prototypes based contrastive learning framework along with a multi-modal fusion transformer suitable for the problem where videos have multiple fine-grained labels. +- Outperforming relevant baselines on three datasets: APPROVE, YouTube-8M [1] and COIN [55]. + +# 2. Related Works + +Self-Supervised Contastive Learning (CL) has been an effective paradigm for visual representation learning. Methods such as SimCLR [8], MoCo (Momentum Contrastive learning) [18], Augmented Multiscale Deep InfoMax (AMDIM) [5], Contrastive Predictive Coding (CPC) [39], MoCov2 [10], MoCov3 [12] and SimCLRv2 [9] have achieved strong performance on image classification benchmarks. The shared property between these CL frameworks is that data augmentation is used to generate positive pairs for CL from a single instance, where other data instances are treated as negatives. Prototypical Contrastive Learning (PCL) [30] extends self-supervised contrastive learning with the idea of clustering data representations during training to generate unsupervised prototypes which represent intra-class variation. We utilize class prototypes instead in the supervised setting, to learn fine-grained distinctions between classes. + +Supervised CL methods such as SupCon [26] utilize labels to enhance contrastive learning by forming positive and negative pairs using labels instead of data augmentation. Supervised Contrastive Learning has also been used for other tasks such as image segmentation [57] and classification in the presence of noisy labels [32]. Hierarchical CL [63] extends SupCon to the hierarchical classification case. However, SupCon cannot be extended to the multi-label case in a straightforward manner, as pairs of data samples with multiple labels cannot be clearly classified just into positives and negatives. + +Weakly-Supervised Multi-Modal CL: Weakly aligned text-image/video datasets scraped from the web such as Conceptual Captions [50] and WebVid-10M [6] enable learning of multi-modal representations. CLIP [42] applies a cross-modal contrastive loss to train individual text and image encoders. Everything at Once [51] is able to additionally utilize the audio modality and incorporates a pairwise fusion encoder which encodes pairs of modalities, as a re + +![](images/de2d6f94ec35f2ce55da96d7a453b2c8fcad43a3d8731826aa7bab93a2c70265.jpg) +sounds in words + +![](images/a9c25ce02359740175cd7f7eee47e60b0a53f03399eb2ea44ca9dd4204fbc637.jpg) +rhyming + +![](images/2064b423610fdf6bbbee26e16f03ca45424ccd5be0bc30dff4a45f20c8910f04.jpg) +letter names + +![](images/4086dc0c4c105f6f56f5ac5d849b3fe58dcfb1760398a816516daa493b62a1bc.jpg) +letter sounds + +![](images/b6b683cb29e18ec994045cb67ed3c6ee9cbe9f52794b9d989b719338e25e7d44.jpg) + +![](images/d96e217df87c94004be40ca45ff179bb041727d2ed795cb749d96ac4354823b7.jpg) + +![](images/b22019975ef426b83d90d2f01b2c0c9fdb05cb798da0a8a91b1866a7903ae153.jpg) +(a) Frames from literacy videos +analyze compare shape +addition subtraction + +![](images/0a14f1dd2d4cc2e27758e64586bee9b116c3a0ff6b152f9ede720953b4e072c0.jpg) +building drawing shapes +counting +(b) Frames from math videos +Figure 1. Sample video frames from the APPROVE dataset. Videos belong to the (a) literacy classes, (b) math classes, and (c) background. Background videos do not contain educational content but share visual similarities with educational videos. The videos are labeled with fine-grained sub-classes, e.g., letter names vs letter sounds. + +![](images/f62656d7c4adaed62e4b85bbbcb7f94c0ad53fc9ccb942b08b498ada61f7095a.jpg) + +![](images/6b631189dbbd0d5105d952fae59eb31b78b2bfc77933254287061d68634f4626.jpg) + +![](images/3705b586c96f1539d4e5c1329c024fd30fde3a47044aad6c1116e5faf67649ba.jpg) +(c) Frames from background videos + +![](images/42fb2181f46789989d6dad2ec6d93796eaff5df97870038314eab38607521bf6.jpg) + +sult, 6 forward passes of the fusion model are required for 3 modalities. Frozen in Time [6] is able to utilize both image-text and video-text datasets through the use of a Space-Time Transformer Visual Encoder. Visual Conditioned GPT [37] uses a single cross-attention fusion layer to combine pretrained CLIP text and visual features. Flamingo [2] adds cross-attention layers interleaved with language decoder layers to fuse visual information into text generation. MERLOT [60, 61] and Triple Contrastive Learning [58] combine contrastive learning and generative language modeling to learn aligned text-image representations. + +Supervised Multi-Modal Learning: Supervised MultiModal Learning typically relies on crowd-captioned datasets such as Flickr30k [59] and MS-COCO Captions [11]. Some prior works such as OSCAR [33] and VinVL [62] have utilized pre-trained object detectors and multi-modal transformers to learn image captioning using supervised aligned datasets. BLIP [28] takes a hybrid approach where it bootstraps an image captioner using a labeled dataset and uses it to generate captions for web images. This generated corpus is then filtered and used for learning an aligned representation. ALign BBefore Fuse [29] highlights the importance of aligning text and image tokens before fusing them using a multi-modal transformer. + +In this paper, we focus on the fine-grained classification of multilabel educational videos. Due to the lack of suitable datasets, we propose a new dataset, APPROVE, which is described next. + +# 3. APPROVE Dataset + +We propose a dataset, called APPROVE, of curated YouTube videos annotated with educational content. APPROVE consists of 193 hours of expert-annotated videos with 19 classes (7 literacy codes, 11 math, and background) and each video is associated with approximately 3 labels on average. We follow the Common Core Standards [3,40] to select education content suitable for kindergarten level. The Common Core Standards outline what students are expected to know and do at various age ranges and grades. This is a widely accepted standard followed by a range of + +educators. We consider two high-level classes of educational content: literacy and math. For each of these content classes, we select a set of codes. For the literacy class, we select 7 codes including letter names, letter sounds, follow words, sight words, letters in words, sounds in words, and rhyming. For the math class, we select 11 codes including counting, individual number, comparing groups, addition subtraction, measurable attributes, sorting, spatial language, shape identification, building drawing, analyzing and comparing shapes. More details about the standard and the description of the codes are provided in the supplementary material. APPROVE also consists of carefully chosen background videos, i.e., without educational content, that are visually similar to the videos with educational content. We present frames corresponding to these classes in Fig. 1. + +![](images/eaa0b7ba26fca96539e7a0d27187f4efcb0b95745bdacfb14065959e8898d39f.jpg) +Frequency +Figure 2. Frequency of the classes in APPROVE. Math codes are in Orange and literacy codes in Blue. + +To ensure the quality and correctness of the annotations, we consider educational researchers to annotate the videos and follow a standard validation protocol [41]. Each annotator is trained by an expert and annotations on a selected set are examined before engaging the annotator for the final annotation. Annotators start once they reach more than $90\%$ agreement with the expert. Further, we estimate inter + +![](images/1dcf09c212e2f1b060f8e7bb90aeda0cb2d49912691ed473f98715cb3b2bcd91.jpg) +Figure 3. Distribution of the number of labels per video. + +annotator consistency to filter out anomalies. Details about the validation process are provided in the supplementary material. It takes a month to train an education researcher to match expert-level coding accuracy. On average, it takes the trained annotators 1 min to annotate 1 min of video. + +The videos are curated from YouTube and are annotated by the trained annotators to determine educational content in them. Each video can have multiple class labels that are quite similar making the task a multi-label and fine-grained classification problem. For example 'letter names' and 'letter sounds' where visual letters are shown in both but in 'letter sounds', the phonetic sound on the letter is emphasized (Fig. 1 (a)). Similarly, in both 'build and draw shapes' and 'analyzing and comparing shapes', multiple shapes can appear but the latter focuses on comparing multiple shapes by shape and size (Fig. 1 (b)). Class-wise stats are presented in Fig. 2. Note that the task is different from common video classification setups where either multi-label or fine-grained aspects are dealt separately. Single-label datasets such as HMDB51 [27], UCF101 [54], Kinetics700 [53] and multi-label ones such as Charades [52] are widely used benchmarks for this problem. YouTube-Birds and YouTube-Cars [64] are analogous datasets for object recognition from videos. Multi-Sports [34] and FineGym [49] label fine-grained action classes for sports. HVU [15] also adds scenes and attributes annotations along with action and objects. However, action, object and scene recognition are not enough for fine-grained video understanding. For instance, videos from a given education provider might share similar objects (person, chalkboard, etc.) and actions (writing on chalkboard) while covering different topics (counting, shape recognition etc.) in each video. + +# 4. Proposed Approach + +In this section, we first describe the proposed class prototypes based contrastive learning framework suitable for videos containing multiple educational codes. Then we present the approach to learning the class prototypes and finally describe the multimodal transformer network that learns features by fusing visual and text cues from videos. + +# 4.1. Class prototypes based contrastive learning + +In a contrastive learning framework, feature representations are typically learned by simultaneously minimizing the distance between positive samples and maximizing + +
DatasetSize (in hr)Multi-LabelFine GrainedTypeAnnotators
Action Recognition
HMDB5XXVAuthors
UCF27XXVAuthors
Kinetics800XXV+ACrowd
Video Classification
COIN476XXV+ACrowd
YT-8M-✓XFMachine
APPROVE193✓✓V+T+AExperts
+ +Table 1. APPROVE dataset compared with selected prior datasets. V→Video Frames, A→Audio, T→Text, F→Features only. + +the distance between negative samples (See Figure 4.(a)). The positive and negative samples are determined with respect to an anchor sample usually based on the class labels. For example, supervised contrastive learning (SupCon) [26] learns a representation to minimize the intra-class distances and maximize inter-class distances. We denote $\boldsymbol{x}_i$ and $y_i$ as the $i$ th sample and its label, respectively. Let's define $z_i$ as the representation of the $i$ th sample in a batch $A$ , and $sim(z_i, z_j) = \frac{z_i \cdot z_j}{|z_i||z_j|}$ the cosine similarity, then the SupCon loss [26] is defined as: + +$$ +\mathcal {L} _ {\operatorname {S u p C o n}} = \sum_ {i \in A} \frac {- 1}{| P (i) |} \sum_ {p \in P (i)} \log \frac {\exp \left(\operatorname {s i m} \left(\boldsymbol {z} _ {i} , \boldsymbol {z} _ {p}\right) / \tau\right)}{\sum_ {a \in A \backslash i} \exp \left(\operatorname {s i m} \left(\boldsymbol {z} _ {i} , \boldsymbol {z} _ {a}\right) / \tau\right)}, \tag {1} +$$ + +where $P(i)$ is the set of positive samples, i.e., with the same label as $z_{i}$ , in the batch excluding $i$ and $a \in A \setminus i$ is the index of all samples in the batch excluding the $i$ th sample. $\tau$ is a scalar temperature parameter used for scaling similarity values. The positive pairs are grouped into the numerator and minimizing the loss minimizes their distance in the learned representation and vice versa for negative pairs. SupCon is known to be effective for classifying samples with a single label. However, it is not straightforward to extend this for the multilabel setup as beyond positive samples, where all labels are the same, and negative samples, where none of the labels is the same, there can be a third scenario where labels are partially overlapping. Though SupCon has been extended to hierarchical classification [63], it cannot be directly extended to the true multi-label case. + +To address this issue, we learn class prototypes as the representative for each class and consider these as anchors while determining positive and negative samples. Specifically, for a specific class prototype, a representation is learned to minimize distances between the prototype and samples with this class label and maximize the distances between the prototype and samples without this class label. We compare the proposed approach with the standard single-label contrastive learning in figure 4. We iteratively update the class prototypes while learning the feature rep + +![](images/0e76b14c48ebc3c7bcf29f384f22d0146c1256dba49800c43dd96ec2465ec366.jpg) +Figure 4. Contrastive learning operates on the feature space by bringing the representations of similar samples close and pushing distinct samples apart. Prior work in (a) Supervised Contrastive Learning [26] trains the network by treating instances from the same class as positive pairs and instances from different classes as negative pairs. This approach doesn't generalize to multi-label classification tasks, as some instance pairs have partially overlapping labels. We propose the use of class prototypes to enable (b) Multi-Label Prototypes Contrastive Learning. Each sample and the class prototypes corresponding to the labels associated with the sample are treated as positive pairs. Similarly, negative pairs are determined based on the missing class labels. Prototypes are represented by stars $(\star)$ and inputs as circles $(\odot)$ colored with all their relevant labels. We discuss strategies for initializing and learning the label prototypes in Sec. 4.2. + +![](images/593ec9dba574c1a181e71408ab42f8c7927ae739864e3d963c10245ecd348819.jpg) + +resentations. We define $C = \{c_1, \ldots, c_K\}$ as the set of classes where $K$ is the number of classes. For a sample $\pmb{x}$ , let's define $P_{ml}(\pmb{x}) = \{c_k^+\}, c_k^+ \in C$ as the set of multiple class labels associated with $\pmb{x}$ (positive classes) and $c_k^- \in C \setminus P_{ml}(\pmb{x})$ denotes the missing classes (negative classes). We define $CP = \{cp_1, \ldots, cp_K\}$ as the set of class prototypes. Considering $z$ is the representation for the sample $\pmb{x}$ , the class prototypes based multilabel contrastive loss is defined as: + +$$ +\begin{array}{l} \mathcal {L} _ {m l c} (\pmb {x}) = \\ \frac {- 1}{| P _ {m l} (\boldsymbol {x}) |} \sum_ {c _ {k} ^ {+} \in P _ {m l} (\boldsymbol {x})} \left[ \log \frac {\exp \left(\operatorname {s i m} \left(\boldsymbol {z} , \boldsymbol {c p} _ {k}\right) / \tau\right)}{\sum_ {c _ {j} ^ {-} \in C \backslash P _ {m l} (\boldsymbol {x})} \exp \left(\operatorname {s i m} \left(\boldsymbol {z} , \boldsymbol {c p} _ {j}\right) / \tau\right)} \right]. \tag {2} \\ \end{array} +$$ + +Here, minimizing the loss of the positive class prototype and instance pairs in the numerator minimizes the distance between the representation $z$ and the class prototypes corresponding to the sample, and vice versa for negative classes. We also utilize negative sampling to account for the class imbalance between positives and negatives. + +# 4.2. Learning class prototypes + +We aim to learn class-specific prototypes such that the multilabel samples can be thought of as the combinations of the class prototypes selected based on the associated labels. Lets assume $Z_{t}$ is an $N\times d$ matrix of $d$ -dimensional representations, i.e., $\mathbf{z}\mathbf{s})$ , of $N$ samples and $L\in \{0,1\}^{N\times K}$ is corresponding labels matrix with $K$ classes. Let's denote $CP_{t}$ is a matrix of size $K\times d$ of $K$ class prototypes at a training iteration $t$ . Then, $Z_{t} = L\times CP_{t} + \varepsilon$ , where $\varepsilon$ is the residual noise term. Assuming a Gaussian noise that is unbiased and uncorrelated with the labels $L$ , the class prototypes can be approximated as $CP_{t}^{*}\approx (L^{T}L)^{-1}L^{T}Z_{t}$ where operation $(L^T L)$ results in a square matrix amenable to inversion. Note that for single labels, this implies averag- + +ing the features of the instances belonging to a given class as the prototype for that class. In a multi-label setup, additionally, the co-occurrence between the labels is considered. Finally, the class prototypes are updated with learning iterations as: + +$$ +C P _ {t + 1} = \beta \cdot C P _ {t} + (1 - \beta) \cdot C P _ {t} ^ {*}, \tag {3} +$$ + +where $\beta$ is the decay parameter for the exponential moving averaging. The moving averaging avoids collapsing prototypes across the training iterations [31]. + +# 4.3. Multi-Modal Fusion Transformer + +Given a sample video $\mathbf{x}$ , we consider multimodal cues such as visual and audio cues to learn its representation $\mathbf{z}$ . Considering multimodal cues is crucial for content recognition, specifically for education videos as effective comprehension requires attending to both visual demonstration and audio explaining the educational content [3, 40]. To capture audio cues, we consider the audio track of the video and extract speech by removing the background such as tunes or instruments [21]. Then we employ the automatic speech recognition (ASR) technique to transcribe text from the speech [43]. We notice that separating speech from the background is important for an accurate ASR transcription. Given the video frames and text transcription, we propose a multimodal transformer network (MTN) to fuse these cues using cross-modal attention. + +Our MTN (Figure 5) has three components: image encoder, text encoder, and fusion encoder to learn visual $(z^v)$ , text $(z^t)$ , and fusion $(z^f)$ representations, respectively. The sample representation is comprised of these three representations $z = \{z^v, z^t, z^f\}$ . The image encoder is implemented by a vision transformer (ViT) [16] that learns frame embeddings from the video frames along with a special CLS for each frame [14]. We pool the CLS tokens across + +![](images/bafa208d7a462b2edd8c2a532a34c4f605924eafb2a27298c3ff635cceab7ab3.jpg) +Figure 5. Multi-Modal Classification Network. A text encoder is used to encode ASR text from the video, while an Image Encoder is used to get tokens representing each frame of the video. Unimodal pre-training is carried out on the text & image encoders respectively. Multi-label contrastive loss is used along with shared prototypes to align the representations across both modalities. This is followed by joint end-to-end learning of the whole multimodal network including the fusion encoder which applies multi-head self-attention within and across the modalities. The prototypes are further refined during the multi-modal training phase. + +the frames and consider this as a compact video representation. Similarly, we consider the BERT-based text transformer [24] to learn the word embeddings along with the text $CLS$ token for the entire transcription. We consider the text $CLS$ as the text representation. Finally, the fusion encoder fuses the visual and text cues by leveraging cross-modal attention between frame and word embeddings. + +# 4.4. Inference using prototypes + +Existing approaches in contrastive learning [8, 26] usually discard the MLP layer after the contrastive training, and a linear classifier is trained on the frozen backbone network. We rely on the class prototypes to carry out inference by utilizing the cosine distance between the learned prototypes and test features. Given our prototype loss-based training, the estimated probability of a given class should be proportional to the normalized temperature scaled cosine distance + +(Equation 4). In practice, we normalize the cosine distance such that $-1$ and $1$ correspond to a confidence of $0\& 1$ respectively. Thus the prediction is made following: + +$$ +\hat {p} (k | \boldsymbol {x}) \propto \exp (\operatorname {s i m} (\boldsymbol {z}, \boldsymbol {c p} _ {k}) / \tau), \tag {4} +$$ + +where $z$ is the multimodal representation of the sample $\pmb{x}$ + +# 4.5. Implementation Details + +Training strategy for the multimodal fusion: We follow a two-stage training process: during the initial unimodal training phase, we utilize fixed prototypes in each modality to align the representations. Then in the second stage, we train the unimodal encoders and the multi-modal fusion encoder end-to-end. The cross-modal alignment learned during the first stage improves the learning of the multi-modal representation. The multi-modal learning phase consists of alternating optimization steps of training the network using our contrastive loss per Sec. 4.1 and refining the class prototypes as described in Sec. 4.2. + +Image encoder: For video frames, Random Resized Crop and RandAugment [13] augmentations are used from torchvision. We use ImageNet pretrained vision encoders ResNet50 [19], ViT-B/32 (224×224 resolution) and ViT-B/16 (384×384 resolution) [17]. + +Text encoder: We generate text from the videos using Whisper [43], an open source ASR model. For data augmentation, we generate four versions of the ASR text by back-translation using the Helsinki-NLP/opus-mt-{en-de, en-nl, en-fr} models through the nlpaug library [38]. Synonym replacement, text span removal and random word swapping augmentations are also used for the text data. We use DistilBERT-Base-uncased [48], and t5-small [44] from HuggingFace transformers library as the text encoder. + +Optimizer: We employ AdamW optimizer [36] for training with a learning rate of 0.0005. Weight decay of 1e-6 is utilized only on the MLP head during contrastive training and the classifier during BCE/Focal/Asym. loss. Since pre-trained vision and text backbones are used, the backbone learning rate is set to 1/10th of the learning rate for the head. Exponential Moving Averaging every 10 steps with a decay of 0.999 was used for the model parameters. + +# 5. Experiments + +Datasets. In addition to the proposed APPROVE dataset, we evaluate our approach on a subset of Youtube-8M [1] and COIN [55] datasets. YT-8M consists of a diverse set of YouTube videos with video and audio modalities. We consider a subset of YT-8M dataset with 46K videos and 165 classes. COIN consists of instructional videos covering a wide variety of domains and spanning over 180 classes. + +Baselines. We compare the efficacy of our multilabel classification framework against the following baselines: + +Binary cross-entropy: In this baseline, loss for multiple labels is computed by combining the binary cross-entropy losses for individual classes. + +Focal loss [35]: This considers a modified binary cross-entropy to assign a higher weight to hard samples by adjusting a focusing parameter $\gamma$ . Negative samples can also be down-weighted by using a weight $\alpha$ . The focal loss for a positive label is given as $\mathcal{L}_{focal}(p) = -\alpha (1 - p)^{\gamma}\log (p)$ . We set $\gamma = 2$ and $\alpha = 0.2$ in our experiments. + +Asymmetric loss [47]: This builds upon the focal loss by utilizing different focusing parameters $\gamma_{+}$ and $\gamma_{-}$ for positive and negative labels. It also ignores the negative samples with a prediction probability lower than a margin $m$ . Asymmetric Loss for prediction $p$ corresponding to label $y$ is given as: $\mathcal{L}_{asym}(p,y) = -yL_{+} - (1 - y)L_{-}$ , where $L_{+} = (1 - p)^{\gamma_{+}}\cdot \log (p)$ and + +$L_{-} = (\max (p - m,0))^{\gamma_{-}}\cdot \log (1 - \max (p - m,0))$ We follow the 5-step procedure recommended by the original authors to train this baseline. We experimentally set $\{\gamma_{-} = 2,\gamma_{+} = 1,m = 0.1\}$ corresponding to the best performance on APPROVE. + +Metrics. In order to develop a reliable education content detection framework, achieving high precision is crucial. Thus we consider Recall @ 80% Precision (R@80) as the primary metric. We also consider the standard area under the precision-recall curve (AUPR) that is not sensitive to a specific threshold for making the final prediction. We also consider a label ranking average precision (LRAP) [56] metric that is more suitable for the multilabel setup. This estimates whether the ground truth classes are predicted with higher scores than the rest: + +$$ +\mathrm {L R A P} = \frac {1}{n} \sum_ {i = 1} ^ {m} \frac {1}{| Y _ {i} |} \sum_ {\lambda \in Y _ {i}} \frac {| \lambda^ {\prime} \in Y _ {i} : r n k _ {i} (\lambda^ {\prime}) \leq r n k _ {i} (\lambda) |}{r n k _ {i} (\lambda)}, +$$ + +where $r n k_{i}(\lambda)$ is the predicted rank of class $\lambda$ for sample $i$ . LRAP is a ranking metric and in independent of a threshold. + +Results on APPROVE. We compare the proposed approach with the baselines in Tab. 2. Our approach outperforms the strongest baselines by $3.1\%$ and $2.3\%$ with respect to R@80 and AUPR, respectively. We also present results for separate models trained on Math and Literacy subsets of Approve respectively. Results on the Math subset are higher compared to the Literacy subset, which indicates the literacy classes are harder to distinguish mostly due to the high inter-class similarity. The top three hardest classes are follow_words, letters_in_words, and sounds_in_words and these are from the literacy set. + +Results on YT-46K and COIN. Beyond the APPROVE dataset, we also test our approach on two public datasets: YT-46K and COIN. Here we provide the results for the multi-modal models and the results for the single-modality models are in Section D of the Supplementary Material. + +Results on YT-46K are provided in Table 3. As YT-8M + +
SubsetModalityMethodAUPRLRAPR@80
AllVBCE45.554.36.9
Focal45.956.615.0
Ours46.757.919.6
TBCE79.885.163.3
Focal79.985.772.8
Ours82.587.475.4
V+TBCE84.388.476.3
Focal [35]86.189.182.2
Asym. [47]86.089.282.4
Ours88.4 +2.390.7 +1.585.5 +3.1
MTHV+TBCE86.392.480.3
Focal87.292.182.4
Ours88.4 +1.293.2 +1.183.2 +0.8
LITV+TBCE72.182.950.7
Focal72.783.550.9
Ours73.6 +0.984.7 +1.254.7 +3.8
+ +Table 2. Results on APPROVE dataset. All metrics in %. V $\rightarrow$ Video & T $\rightarrow$ Text. M $\rightarrow$ Math & L $\rightarrow$ Literacy Subsets. + +
ModalityMethodAUPRLRAPR@80
V+TBCE64.670.242.3
V+TFocal [35]69.772.744.6
V+TOurs70.9 +1.274.9 +2.249.1 +4.5
+ +Table 3. Results on YT-46K. V→Video Frames and T→Text. + +
ModalityMethodTop-1 Accuracy
V+TCE53.7
V+TBCE54.9
V+TFocal [35]56.1
V+TSupCon [26]54.7
V+TOurs57.5 +1.4
+ +Table 4. Results on COIN. V→Video Frames and T→Text. + +was primarily collected with the intention of visual classification, the additional use of text data leads to a smaller improvement compared to APPROVE. Results on COIN are provided in Table 4. As each video from COIN is mapped to a single task. Thus, we consider the Top-1 accuracy as the metric. On COIN we compare our approach with SupCon [26] which is effective for single labels. Note that our approach outperforms SupCon and this justifies the effectiveness of the class prototypes based training in a generic contrastive learning framework. + +# 5.1. Ablation Studies + +We perform the following ablation studies to quantify the impact of our approach for learning class prototypes, the multimodal fusion module, and the choice of visual and text encoding frameworks. + +(a) Class Prototypes + +
VariantAPPROVECOIN
Random84.156.6
Orthogonal84.857.0
Learned85.557.5
Hierarchical86.057.8
+ +(b) Fusion Encoder Size + +
LayersAPPROVECOIN
184.957.1
285.557.5
485.357.6
+ +(c) Vision Encoder + +
Vision ModelAPPROVECOIN
R5084.855.2
ViT-B/3285.557.5
ViT-B/1683.857.8
+ +(d) Text Encoder + +
Text ModelAPPROVECOIN
DistilBert-B85.557.5
t5-S87.357.9
+ +Table 5. Ablation studies. R@80% Precision for APPROVE and Top-1 Accuracy for COIN. Default setup is underlined. +(a) Noisy modalities. + +
% missingAPPROVECOIN
0%85.557.5
10% V80.153.3
10% T75.857.2
30% T68.942.8
+ +(b) Run-to-Run Variance + +
MethodAPPROVECOIN
BCE76.3 ± 0.754.9 ± 0.6
Focal82.2 ± 0.556.1 ± 0.3
Ours85.5 ± 0.557.5 ± 0.8
+ +(c) Initialization + +
MethodAPPROVECOIN
ImageNet & Wiki-en+TBC85.557.5
CLIP86.763.5
+ +Table 6. Robustness analysis. Our method is robust to partially missing modality and has similar run-to-run variance as baselines. + +Learning class prototypes: We compare the two strategies where after initializing the class prototypes, we 1) keep them fixed and learn only the multimodal embedding of the samples, and 2) class prototypes and sample embedding are learned iteratively. The initialization can be done either randomly, or with orthogonal constraints. We note that orthogonal initialization performs best in our experiments and iterative adjusting the class prototypes achieves better performance as shown in Table 5 (a). We also consider hierarchical prototypes, for APPROVE, using a 2-level hierarchy where the first level consists of 18 classes, and the second level is the 3 super-classes: math, literacy, and background. The 180 task categories of COIN are organized into 12 domains in the taxonomy provided with the dataset. This hierarchy imposes an additional constraint on learning the embeddings during training. + +Fusion Encoder: This evaluates the effect of the number of layers in the fusion encoder on the final performance. As expected, the performance improves with more layers and saturates around 4 layers (Table 5 (b)). + +Vision Encoder: This evaluates the effect of the image encoder on the final performance. We consider ResNet [19] and ViT variants [16]. We notice that the ViT-B/16-384 encoder works well for the larger COIN dataset, whereas the ViT-B/32-224 encoder works best for APPROVE (Table 5 (c)). + +Text Encoder: This evaluates the effect of the text encoder on the final performance. We test DistilBERT and T5 backbones for the text encoder. BERT is trained to predict masked spans of text. T5's unsupervised objective is similar, however, it trains on predicting the entire sequence instead of just the masked spans. GPT2 takes an auto-regressive approach to language modeling (Table 5 (d)). + +# 5.2. Robustness analysis + +We evaluate the robustness of our approach as follows: + +Noisy modality: YouTube videos may have noisy modalities where some of the video frames are missing or ASR transcription is noisy. We show that our approach is robust against the cases where a percentage of video frames or text + +words are missing as shown in Tab 6(a). + +Run-to-Run variance: The low variance across runs (Tab 6(b)) indicates that our approach is not sensitive to random initialization of the class prototypes. + +Initializing the encoders: We consider the ImageNet pretraining for the image encoder, while English Wikipedia + Toronto Book Corpus is used to pre-train the text encoder. We provide results where the backbones are initialized with CLIP [42], which provides a more aligned vision-text representation. As expected, the results are better with the CLIP initialization (Tab 6(c)). The improvements are more significant on COIN than APPROVE as CLIP models may not be exposed to educational videos. + +# 6. Conclusion + +We have proposed an approach for detecting educational content in online videos. The problem is formulated as a fine-grained multilabel video classification task and we have considered class-prototypes based contrastive learning to address this. We have employed a multimodal transformer network to fuse visual and audio cues. This is crucial for comprehending educational content as both visual and audio cues are to be aligned to ensure effective comprehension. Our approach is shown to be effective in distinguishing fine-grained educational content with high interclass similarity. We have introduced APPROVE - a dataset with 193 hours of expert-annotated educational videos. Beyond APPROVE, we have evaluated our approach on COIN and YouTube-8M datasets where our approach outperforms the competitive baselines. + +# Acknowledgements + +This work is supported by NSF grant # 2139219 and SRI R&D award. Rohit Gupta is supported by ARO grant W911NF-19-1-0356. + +Views and conclusions contained herein are those of the authors and should not be interpreted as representing the official policies or endorsements, either expressed or implied, of NSF, ARO, IARPA, DOI/IBC, or the U.S. Government. The U.S. Government is authorized to reproduce and distribute reprints for Governmental purposes notwithstanding any copyright annotation thereon. + +# References + +[1] Sami Abu-El-Haija, Nisarg Kothari, Joonseok Lee, Paul Natsev, George Toderici, Balakrishnan Varadarajan, and Sudheendra Vijayanarasimhan. YouTube-8M: A large-scale video classification benchmark. arXiv preprint arXiv:1609.08675, 2016. 1, 2, 6 +[2] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katie Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. arXiv preprint arXiv:2204.14198, 2022. 3 +[3] National Governors Association et al. Common core state standards. Washington, DC, 2010. 1, 2, 3, 5 +[4] Brooke Auxier, Monica Anderson Andrew Perrin, and Erica Turner. Parenting children in the age of screens. Technical report, Pew Research Center, 2020. 1 +[5] Philip Bachman, R Devon Hjelm, and William Buchwalter. Learning representations by maximizing mutual information across views. In H. Wallach, H. Larochelle, A. Beygelzimer, F. d Alché-Buc, E. Fox, and R. Garnett, editors, Advances in Neural Information Processing Systems, volume 32. Curran Associates, Inc., 2019. 2 +[6] Max Bain, Arsha Nagrani, Gül Varol, and Andrew Zisserman. Frozen in time: A joint video and image encoder for end-to-end retrieval. In IEEE International Conference on Computer Vision, 2021. 2, 3 +[7] Johanna Burkhardt and Wolfgang Lenhard. A meta-analysis on the longitudinal, age-dependent effects of violent video games on aggression. Media Psychology, 25(3):499-512, 2022. 1 +[8] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International conference on machine learning, pages 1597-1607. PMLR, 2020. 2, 6 +[9] Ting Chen, Simon Kornblith, Kevin Swersky, Mohammad Norouzi, and Geoffrey E Hinton. Big self-supervised models are strong semi-supervised learners. In H. Larochelle, M. Ranzato, R. Hadsell, M. F. Balcan, and H. Lin, editors, Advances in Neural Information Processing Systems, volume 33, pages 22243-22255. Curran Associates, Inc., 2020. 2 +[10] Xinlei Chen, Haoqi Fan, Ross Girshick, and Kaiming He. Improved baselines with momentum contrastive learning, 2020. 2 +[11] Xinlei Chen, Hao Fang, Tsung-Yi Lin, Ramakrishna Vedantam, Saurabh Gupta, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco captions: Data collection and evaluation server. arXiv preprint arXiv:1504.00325, 2015. 3 +[12] Xinlei Chen, Saining Xie, and Kaiming He. An empirical study of training self-supervised vision transformers, 2021. 2 +[13] Ekin D Cubuk, Barret Zoph, Jonathon Shlens, and Quoc V Le. Randaugment: Practical automated data augmentation with a reduced search space. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops, pages 702-703, 2020. 6 + +[14] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018. 5 +[15] Ali Diba, Mohsen Fayyaz, Vivek Sharma, Manohar Paluri, Jürgen Gall, Rainer Stiefelhagen, and Luc Van Gool. Large scale holistic video understanding. In European Conference on Computer Vision, pages 593-610. Springer, 2020. 4 +[16] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. ICLR, 2021. 5, 8 +[17] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations, 2021. 6 +[18] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9729-9738, 2020. 2 +[19] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 6, 8 +[20] Lowry Hemphill and Terrence Tivnan. The importance of early vocabulary for literacy achievement in high-poverty schools. Journal of Education for Students Placed at Risk, 13(4):426-451, 2008. 1 +[21] Romain Hennequin, Anis Khlif, Felix Voituret, and Manuel Moussallam. Spleeter: a fast and efficient music source separation tool with pre-trained models. Journal of Open Source Software, 5(50):2154, 2020. Deezer Research. 5 +[22] Lisa B Hurwitz. Getting a read on ready to learn media: A meta-analytic review of effects on literacy. *Child Development*, 90(5):1754–1771, 2019. 1 +[23] Lisa B Hurwitz and Kelly L Schmitt. Raising readers with ready to learn: A six-year follow-up to an early educational computer game intervention. Computers in Human Behavior, 104:106176, 2020. 1 +[24] Andrew Jaegle, Felix Gimeno, Andy Brock, Oriol Vinyals, Andrew Zisserman, and Joao Carreira. Perceiver: General perception with iterative attention. In International conference on machine learning, pages 4651-4664. PMLR, 2021. 6 +[25] Nancy C Jordan, David Kaplan, Chaitanya Ramineni, and Maria N Locuniak. Early math matters: kindergarten number competence and later mathematics outcomes. Developmental psychology, 45(3):850, 2009. 1 +[26] Prannay Khosla, Piotr Teterwak, Chen Wang, Aaron Sarna, Yonglong Tian, Phillip Isola, Aaron Maschinot, Ce Liu, and Dilip Krishnan. Supervised contrastive learning. Advances in Neural Information Processing Systems, 33:18661-18673, 2020. 2, 4, 5, 6, 7 + +[27] Hildegard Kuehne, Hueihan Jhuang, Estíbaliz Garrote, Tomaso Poggio, and Thomas Serre. Hmdb: a large video database for human motion recognition. In 2011 International conference on computer vision, pages 2556-2563. IEEE, 2011. 4 +[28] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In ICML, 2022. 3 +[29] Junnan Li, Ramprasaath R. Selvaraju, Akhilesh Deepak Gotmare, Shafiq Joty, Caiming Xiong, and Steven Hoi. Align before fuse: Vision and language representation learning with momentum distillation. In NeurIPS, 2021. 3 +[30] Junnan Li, Pan Zhou, Caiming Xiong, and Steven Hoi. Prototypical contrastive learning of unsupervised representations. In International Conference on Learning Representations, 2021. 2 +[31] Junnan Li, Pan Zhou, Caiming Xiong, and Steven C.H. Hoi. Prototypical contrastive learning of unsupervised representations. In ICLR, 2021. 5 +[32] Shikun Li, Xiaobo Xia, Shiming Ge, and Tongliang Liu. Selective-supervised contrastive learning with noisy labels. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 316-325, June 2022. 2 +[33] Xiujun Li, Xi Yin, Chunyuan Li, Xiaowei Hu, Pengchuan Zhang, Lei Zhang, Lijuan Wang, Houdong Hu, Li Dong, Furu Wei, Yejin Choi, and Jianfeng Gao. Oscar: Object-semantics aligned pre-training for vision-language tasks. ECCV 2020, 2020. 3 +[34] Yixuan Li, Lei Chen, Runyu He, Zhenzhi Wang, Gangshan Wu, and Limin Wang. Multisports: A multi-person video dataset of spatio-temporally localized sports actions. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13536-13545, 2021. 4 +[35] Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dólár. Focal loss for dense object detection. In Proceedings of the IEEE international conference on computer vision, pages 2980-2988, 2017. 7 +[36] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In International Conference on Learning Representations, 2019. 6 +[37] Ziyang Luo, Yadong Xi, Rongsheng Zhang, and Jing Ma. A frustratingly simple approach for end-to-end image captioning, 2022. 3 +[38] Edward Ma. Nlp augmentation. https://github.com/makcedward/nlpaug, 2019.6 +[39] Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748, 2018. 2 +[40] Andrew Porter, Jennifer McMaken, Jun Hwang, and Rui Yang. Common core standards: The new us intended curriculum. Educational researcher, 40(3):103-116, 2011. 1, 2, 3, 5 +[41] J. S. Radesky, A. Schaller, S. L. Yeo, H. M. Weeks, and M. B. Robb. Young kids and youtube: How ads, toys, and games dominate viewing. common sense media. Technical report, Common Sense Media, 2020. 2, 3 + +[42] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 2, 8 +[43] Alec Radford, Jong Wook Kim, Tao Xu, Greg Brockman, Christine McLeavey, and Ilya Sutskever. Robust speech recognition via large-scale weak supervision. Technical report, Tech. Rep., Technical report, OpenAI, 2022. 5, 6 +[44] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, Peter J Liu, et al. Exploring the limits of transfer learning with a unified text-to-text transformer. J. Mach. Learn. Res., 21(140):1-67, 2020. 6 +[45] Victoria Rideout and Michael B. Robb. The common sense census: Media use by tweens and teens, Oct 2019. 1 +[46] V Rideout and M. B. Robb. The common sense census: Media use by tweens and teens. common sense media. Technical report, Common Sense Media, 2019. 1 +[47] Tal Ridnik, Emanuel Ben-Baruch, Nadav Zamir, Asaf Noy, Itamar Friedman, Matan Protter, and Lihi Zelnik-Manor. Asymmetric loss for multi-label classification. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 82-91, 2021. 7 +[48] Victor Sanh, Lysandre Debut, Julien Chaumont, and Thomas Wolf. Distilbert, a distilled version of bert: smaller, faster, cheaper and lighter. arXiv preprint arXiv:1910.01108, 2019.6 +[49] Dian Shao, Yue Zhao, Bo Dai, and Dahua Lin. Finegym: A hierarchical video dataset for fine-grained action understanding. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2616-2625, 2020. 4 +[50] Piyush Sharma, Nan Ding, Sebastian Goodman, and Radu Soricut. Conceptual captions: A cleaned, hypernymed, image alt-text dataset for automatic image captioning. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 2556-2565, Melbourne, Australia, July 2018. Association for Computational Linguistics. 2 +[51] Nina Shvetsova, Brian Chen, Andrew Rouditchenko, Samuel Thomas, Brian Kingsbury, Rogerio S Feris, David Harwath, James Glass, and Hilde Kuehne. Everything at once-multimodal fusion transformer for video retrieval. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20020-20029, 2022. 2 +[52] Gunnar A. Sigurdsson, Gül Varol, Xiaolong Wang, Ali Farhadi, Ivan Laptev, and Abhinav Gupta. Hollywood in homes: Crowdsourcing data collection for activity understanding. In Bastian Leibe, Jiri Matas, Nicu Sebe, and Max Welling, editors, Computer Vision – ECCV 2016, pages 510–526, Cham, 2016. Springer International Publishing. 4 +[53] Lucas Smaira, João Carreira, Eric Noland, Ellen Clancy, Amy Wu, and Andrew Zisserman. A short note on the kinetics-700-2020 human action dataset, 2020. 1, 4 + +[54] Khurram Soomro, Amir Roshan Zamir, and Mubarak Shah. Ucf101: A dataset of 101 human actions classes from videos in the wild. arXiv preprint arXiv:1212.0402, 2012. 1, 4 +[55] Yansong Tang, Dajun Ding, Yongming Rao, Yu Zheng, Danyang Zhang, Lili Zhao, Jiwen Lu, and Jie Zhou. Coin: A large-scale dataset for comprehensive instructional video analysis. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 2, 6 +[56] Grigorios Tsoumakas, Ioannis Katakis, and Ioannis Vla-havas. Mining Multi-label Data, pages 667-685. Springer US, Boston, MA, 2010. 7 +[57] Wenguan Wang, Tianfei Zhou, Fisher Yu, Jifeng Dai, Ender Konukoglu, and Luc Van Gool. Exploring cross-image pixel contrast for semantic segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 7303-7313, 2021. 2 +[58] Jinyu Yang, Jiali Duan, Son Tran, Yi Xu, Sampath Chanda, Liquun Chen, Belinda Zeng, Trishul Chilimbi, and Junzhou Huang. Vision-language pre-training with triple contrastive learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15671-15680, 2022. 3 +[59] Peter Young, Alice Lai, Micah Hodosh, and Julia Hockenmaier. From image descriptions to visual denotations: New similarity metrics for semantic inference over event descriptions. Transactions of the Association for Computational Linguistics, 2:67-78, 2014. 3 + +[60] Rowan Zellers, Jiasen Lu, Ximing Lu, Youngjae Yu, Yanpeng Zhao, Mohammadreza Salehi, Aditya Kusupati, Jack Hessel, Ali Farhadi, and Yejin Choi. Merlot reserve: Neural script knowledge through vision and language and sound. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 16375-16387, June 2022. 3 +[61] Rowan Zellers, Ximing Lu, Jack Hessel, Youngjae Yu, Jae Sung Park, Jize Cao, Ali Farhadi, and Yejin Choi. Merlot: Multimodal neural script knowledge models. Advances in Neural Information Processing Systems, 34:23634-23651, 2021. 3 +[62] Pengchuan Zhang, Xiujun Li, Xiaowei Hu, Jianwei Yang, Lei Zhang, Lijuan Wang, Yejin Choi, and Jianfeng Gao. Vinvl: Making visual representations matter in vision-language models. CVPR 2021, 2021. 3 +[63] Shu Zhang, Ran Xu, Caiming Xiong, and Chetan Ramaiah. Use all the labels: A hierarchical multi-label contrastive learning framework. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 16660-16669, June 2022. 2, 4 +[64] Chen Zhu, Xiao Tan, Feng Zhou, Xiao Liu, Kaiyu Yue, Errui Ding, and Yi Ma. Fine-grained video categorization with redundancy reduction attention. In Proceedings of the European Conference on Computer Vision (ECCV), pages 136-152, 2018. 4 \ No newline at end of file diff --git a/2023/Class Prototypes Based Contrastive Learning for Classifying Multi-Label and Fine-Grained Educational Videos/images.zip b/2023/Class Prototypes Based Contrastive Learning for Classifying Multi-Label and Fine-Grained Educational Videos/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..4a5f98cc7c73a6db1de7beefefa3233372b3346d --- /dev/null +++ b/2023/Class Prototypes Based Contrastive Learning for Classifying Multi-Label and Fine-Grained Educational Videos/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1e385364b7e1b8547666563799ad6a21106ab141565a356dce5f8df10e208639 +size 449051 diff --git a/2023/Class Prototypes Based Contrastive Learning for Classifying Multi-Label and Fine-Grained Educational Videos/layout.json b/2023/Class Prototypes Based Contrastive Learning for Classifying Multi-Label and Fine-Grained Educational Videos/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..59d5f5442fd93ef165eb6637864ee7daf952fc67 --- /dev/null +++ b/2023/Class Prototypes Based Contrastive Learning for Classifying Multi-Label and Fine-Grained Educational Videos/layout.json @@ -0,0 +1,9851 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 59, + 102, + 536, + 138 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 102, + 536, + 138 + ], + "spans": [ + { + "bbox": [ + 59, + 102, + 536, + 138 + ], + "type": "text", + "content": "Class Prototypes based Contrastive Learning for Classifying Multi-Label and Fine-Grained Educational Videos" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 67, + 159, + 525, + 196 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 159, + 525, + 196 + ], + "spans": [ + { + "bbox": [ + 67, + 159, + 525, + 196 + ], + "type": "text", + "content": "Rohit Gupta " + }, + { + "bbox": [ + 67, + 159, + 525, + 196 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 67, + 159, + 525, + 196 + ], + "type": "text", + "content": " * Anirban Roy " + }, + { + "bbox": [ + 67, + 159, + 525, + 196 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 67, + 159, + 525, + 196 + ], + "type": "text", + "content": " Claire Christensen " + }, + { + "bbox": [ + 67, + 159, + 525, + 196 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 67, + 159, + 525, + 196 + ], + "type": "text", + "content": " Sukeong Kim " + }, + { + "bbox": [ + 67, + 159, + 525, + 196 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 67, + 159, + 525, + 196 + ], + "type": "text", + "content": " Sarah Gerard " + }, + { + "bbox": [ + 67, + 159, + 525, + 196 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 67, + 159, + 525, + 196 + ], + "type": "text", + "content": " Madeline Cincebeaux " + }, + { + "bbox": [ + 67, + 159, + 525, + 196 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 67, + 159, + 525, + 196 + ], + "type": "text", + "content": " Ajay Divakaran " + }, + { + "bbox": [ + 67, + 159, + 525, + 196 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 67, + 159, + 525, + 196 + ], + "type": "text", + "content": " Todd Grindal " + }, + { + "bbox": [ + 67, + 159, + 525, + 196 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 67, + 159, + 525, + 196 + ], + "type": "text", + "content": " Mubarak Shah " + }, + { + "bbox": [ + 67, + 159, + 525, + 196 + ], + "type": "inline_equation", + "content": "^{1}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 99, + 201, + 364, + 229 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 99, + 201, + 364, + 229 + ], + "spans": [ + { + "bbox": [ + 99, + 201, + 364, + 229 + ], + "type": "text", + "content": "1 Center for Research in Computer Vision, University of Central Florida rohitg@knights.ucf.edu, shah@crcv.ucf.edu" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 385, + 201, + 490, + 229 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 385, + 201, + 490, + 229 + ], + "spans": [ + { + "bbox": [ + 385, + 201, + 490, + 229 + ], + "type": "text", + "content": "2 SRI International anirban.roy@sri.com" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 143, + 256, + 192, + 269 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 256, + 192, + 269 + ], + "spans": [ + { + "bbox": [ + 143, + 256, + 192, + 269 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 283, + 290, + 679 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 283, + 290, + 679 + ], + "spans": [ + { + "bbox": [ + 46, + 283, + 290, + 679 + ], + "type": "text", + "content": "The recent growth in the consumption of online media by children during early childhood necessitates data-driven tools enabling educators to filter out appropriate educational content for young learners. This paper presents an approach for detecting educational content in online videos. We focus on two widely used educational content classes: literacy and math. For each class, we choose prominent codes (sub-classes) based on the Common Core Standards. For example, literacy codes include 'letter names', 'letter sounds', and math codes include 'counting', 'sorting'. We pose this as a fine-grained multilabel classification problem as videos can contain multiple types of educational content and the content classes can get visually similar (e.g., 'letter names' vs 'letter sounds'). We propose a novel class prototypes based supervised contrastive learning approach that can handle fine-grained samples associated with multiple labels. We learn a class prototype for each class and a loss function is employed to minimize the distances between a class prototype and the samples from the class. Similarly, distances between a class prototype and the samples from other classes are maximized. As the alignment between visual and audio cues are crucial for effective comprehension, we consider a multimodal transformer network to capture the interaction between visual and audio cues in videos while learning the embedding for videos. For evaluation, we present a dataset, APPROVE, employing educational videos from YouTube labeled with fine-grained education classes by education researchers. APPROVE consists of 193 hours of expert-annotated videos with 19 classes. The proposed approach outperforms strong baselines on APPROVE and other benchmarks such as Youtube-8M, and COIN. The dataset is available at https://nusci.csl.sri.com/project/APPROVE." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 256, + 386, + 269 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 256, + 386, + 269 + ], + "spans": [ + { + "bbox": [ + 306, + 256, + 386, + 269 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 277, + 547, + 504 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 277, + 547, + 504 + ], + "spans": [ + { + "bbox": [ + 304, + 277, + 547, + 504 + ], + "type": "text", + "content": "With the expansion of internet access and the ubiquitous availability of smart devices, children increasingly spend a significant amount of time watching online videos. A recent nationally representative survey reported that " + }, + { + "bbox": [ + 304, + 277, + 547, + 504 + ], + "type": "inline_equation", + "content": "89\\%" + }, + { + "bbox": [ + 304, + 277, + 547, + 504 + ], + "type": "text", + "content": " of parents of children aged 11 or younger say their child watches videos on YouTube [4]. Moreover, it is estimated that young children in the age range of two to four years consume 2.5 hours and five to eight years consume 3.0 hours per day on average [45, 46]. Childhood is typically a key period for education, especially for learning basic skills such as literacy and math [20, 25]. Unlike generic online videos, watching appropriate educational videos supports healthy child development and learning [7, 22, 23]. Thus, analyzing the content of these videos may help parents, teachers, and media developers increase young children's exposure to high-quality education videos, which has been shown to produce meaningful learning gains [22]. As the amount of online content produced grows exponentially, automated content understanding methods are essential to facilitate this." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 510, + 548, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 510, + 548, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 510, + 548, + 715 + ], + "type": "text", + "content": "In this work, given a video, our goal is to determine whether the video contains any educational content and characterize the content. Detecting educational content requires identifying multiple distinct types of content in a video while distinguishing between similar content types. The task is challenging as the education codes by Common Core Standards [3, 40] can be similar such as 'letter names' and 'letter sounds', where the former focuses on the name of the letter and the latter is based on the phonetic sound of the letter. Also, understanding education content requires analyzing both visual and audio cues simultaneously as both signals are to be present to ensure effective learning [3, 40]. This is in contrast to standard video classification benchmarks such as the sports or generic YouTube videos in UCF101 [54] Kinetics400 [53], YouTube-8M [1], where visual cues are often sufficient to detect the different classes. Finally, unlike standard well-known action videos," + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 702, + 257, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 702, + 257, + 713 + ], + "spans": [ + { + "bbox": [ + 61, + 702, + 257, + 713 + ], + "type": "text", + "content": "*Work partly done during an internship at SRI International." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19923" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 204 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 204 + ], + "type": "text", + "content": "education codes are more structured and not accessible to common users. Thus, it requires a carefully curated set of videos and expert annotations to create a dataset to enable a data-driven approach. In this work, we focus on two widely used educational content classes: literacy and math. For each class, we choose prominent codes (sub-classes) based on the Common Core Standards that outline age-appropriate learning standards [3,40]. For example, literacy codes include 'letter names', 'letter sounds', 'rhyming', and math codes include 'counting', 'addition subtraction', 'sorting', 'analyze shapes'." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 207, + 289, + 554 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 207, + 289, + 554 + ], + "spans": [ + { + "bbox": [ + 46, + 207, + 289, + 554 + ], + "type": "text", + "content": "We formulate the problem as a multilabel fine-grained video classification task as a video may contain multiple types of content that can be similar. We employ multimodal cues since besides visual cues, audio cues provide important cues to distinguish between similar types of educational content. We propose a class prototypes based supervised contrastive learning approach to address the above-mentioned challenges. We learn a prototype embedding for each class. Then a loss function is employed to minimize the distance between a class prototype and the samples associated with the class label. Similarly, the distance between a class prototype and the samples without that class label is maximized. This is unlike the standard supervised contrastive learning setup where inter-class distance is maximized and intra-class distance is minimized by considering classwise positive and negative samples. This approach is shown to be effective for single-label setups [26]. However, it is not straightforward to extend this for the proposed multilabel setup as samples cannot be identified as positive or negative due to the multiple labels. We jointly learn the embedding of the class prototypes and the samples. The embeddings are learned by a multimodal transformer network (MTN) that captures the interaction between visual and audio cues in videos. We employ automatic speech recognition (ASR) to transcribe text from the audio. The MTN consists of video and text encoders that learn modality-specific embedding and a cross-attention mechanism is employed to capture the interaction between them. The MTN is end-to-end learned through the contrastive loss." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 558, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 558, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 558, + 289, + 715 + ], + "type": "text", + "content": "Due to the lack of suitable datasets for evaluating fine-grained classification of education videos, we propose a new dataset, called APPROVE, of curated YouTube videos annotated with educational content. We follow Common Core Standards [3, 40] to select education content suitable for the kindergarten level. We consider two high-level classes of educational content: literacy and math. For each of these content classes, we select a set of codes. For the literacy class, we select 7 codes and for the math class, we select 11 codes. Each video is associated with multiple labels corresponding to these codes. The videos are annotated by trained education researchers following standard validation protocol [41] to ensure correctness. APPROVE also con" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "text", + "content": "sists of carefully chosen background videos, i.e., without educational content, that are visually similar to the videos with educational content. APPROVE consists of 193 hours of expert-annotated videos with 19 classes (7 literacy codes, 11 math codes, and a background) where each video has 3 labels on average." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 306, + 144, + 504, + 155 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 144, + 504, + 155 + ], + "spans": [ + { + "bbox": [ + 306, + 144, + 504, + 155 + ], + "type": "text", + "content": "Our contributions can be summarized as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 317, + 156, + 545, + 251 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 317, + 156, + 544, + 179 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 156, + 544, + 179 + ], + "spans": [ + { + "bbox": [ + 317, + 156, + 544, + 179 + ], + "type": "text", + "content": "- APPROVE, a fine-grained multi-label dataset of education videos, to promote exploration in this field." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 317, + 180, + 545, + 227 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 180, + 545, + 227 + ], + "spans": [ + { + "bbox": [ + 317, + 180, + 545, + 227 + ], + "type": "text", + "content": "- Class prototypes based contrastive learning framework along with a multi-modal fusion transformer suitable for the problem where videos have multiple fine-grained labels." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 317, + 228, + 544, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 228, + 544, + 251 + ], + "spans": [ + { + "bbox": [ + 317, + 228, + 544, + 251 + ], + "type": "text", + "content": "- Outperforming relevant baselines on three datasets: APPROVE, YouTube-8M [1] and COIN [55]." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 262, + 397, + 275 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 262, + 397, + 275 + ], + "spans": [ + { + "bbox": [ + 306, + 262, + 397, + 275 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 280, + 545, + 482 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 280, + 545, + 482 + ], + "spans": [ + { + "bbox": [ + 304, + 280, + 545, + 482 + ], + "type": "text", + "content": "Self-Supervised Contastive Learning (CL) has been an effective paradigm for visual representation learning. Methods such as SimCLR [8], MoCo (Momentum Contrastive learning) [18], Augmented Multiscale Deep InfoMax (AMDIM) [5], Contrastive Predictive Coding (CPC) [39], MoCov2 [10], MoCov3 [12] and SimCLRv2 [9] have achieved strong performance on image classification benchmarks. The shared property between these CL frameworks is that data augmentation is used to generate positive pairs for CL from a single instance, where other data instances are treated as negatives. Prototypical Contrastive Learning (PCL) [30] extends self-supervised contrastive learning with the idea of clustering data representations during training to generate unsupervised prototypes which represent intra-class variation. We utilize class prototypes instead in the supervised setting, to learn fine-grained distinctions between classes." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 483, + 545, + 615 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 483, + 545, + 615 + ], + "spans": [ + { + "bbox": [ + 304, + 483, + 545, + 615 + ], + "type": "text", + "content": "Supervised CL methods such as SupCon [26] utilize labels to enhance contrastive learning by forming positive and negative pairs using labels instead of data augmentation. Supervised Contrastive Learning has also been used for other tasks such as image segmentation [57] and classification in the presence of noisy labels [32]. Hierarchical CL [63] extends SupCon to the hierarchical classification case. However, SupCon cannot be extended to the multi-label case in a straightforward manner, as pairs of data samples with multiple labels cannot be clearly classified just into positives and negatives." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "content": "Weakly-Supervised Multi-Modal CL: Weakly aligned text-image/video datasets scraped from the web such as Conceptual Captions [50] and WebVid-10M [6] enable learning of multi-modal representations. CLIP [42] applies a cross-modal contrastive loss to train individual text and image encoders. Everything at Once [51] is able to additionally utilize the audio modality and incorporates a pairwise fusion encoder which encodes pairs of modalities, as a re" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "19924" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 51, + 71, + 130, + 111 + ], + "blocks": [ + { + "bbox": [ + 51, + 71, + 130, + 111 + ], + "lines": [ + { + "bbox": [ + 51, + 71, + 130, + 111 + ], + "spans": [ + { + "bbox": [ + 51, + 71, + 130, + 111 + ], + "type": "image", + "image_path": "de2d6f94ec35f2ce55da96d7a453b2c8fcad43a3d8731826aa7bab93a2c70265.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 59, + 113, + 118, + 120 + ], + "lines": [ + { + "bbox": [ + 59, + 113, + 118, + 120 + ], + "spans": [ + { + "bbox": [ + 59, + 113, + 118, + 120 + ], + "type": "text", + "content": "sounds in words" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 133, + 72, + 211, + 111 + ], + "blocks": [ + { + "bbox": [ + 133, + 72, + 211, + 111 + ], + "lines": [ + { + "bbox": [ + 133, + 72, + 211, + 111 + ], + "spans": [ + { + "bbox": [ + 133, + 72, + 211, + 111 + ], + "type": "image", + "image_path": "a9c25ce02359740175cd7f7eee47e60b0a53f03399eb2ea44ca9dd4204fbc637.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 159, + 114, + 187, + 121 + ], + "lines": [ + { + "bbox": [ + 159, + 114, + 187, + 121 + ], + "spans": [ + { + "bbox": [ + 159, + 114, + 187, + 121 + ], + "type": "text", + "content": "rhyming" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 50, + 125, + 129, + 166 + ], + "blocks": [ + { + "bbox": [ + 50, + 125, + 129, + 166 + ], + "lines": [ + { + "bbox": [ + 50, + 125, + 129, + 166 + ], + "spans": [ + { + "bbox": [ + 50, + 125, + 129, + 166 + ], + "type": "image", + "image_path": "2064b423610fdf6bbbee26e16f03ca45424ccd5be0bc30dff4a45f20c8910f04.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 64, + 167, + 111, + 173 + ], + "lines": [ + { + "bbox": [ + 64, + 167, + 111, + 173 + ], + "spans": [ + { + "bbox": [ + 64, + 167, + 111, + 173 + ], + "type": "text", + "content": "letter names" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 131, + 125, + 211, + 166 + ], + "blocks": [ + { + "bbox": [ + 131, + 125, + 211, + 166 + ], + "lines": [ + { + "bbox": [ + 131, + 125, + 211, + 166 + ], + "spans": [ + { + "bbox": [ + 131, + 125, + 211, + 166 + ], + "type": "image", + "image_path": "4086dc0c4c105f6f56f5ac5d849b3fe58dcfb1760398a816516daa493b62a1bc.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 145, + 167, + 196, + 173 + ], + "lines": [ + { + "bbox": [ + 145, + 167, + 196, + 173 + ], + "spans": [ + { + "bbox": [ + 145, + 167, + 196, + 173 + ], + "type": "text", + "content": "letter sounds" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 218, + 72, + 298, + 111 + ], + "blocks": [ + { + "bbox": [ + 218, + 72, + 298, + 111 + ], + "lines": [ + { + "bbox": [ + 218, + 72, + 298, + 111 + ], + "spans": [ + { + "bbox": [ + 218, + 72, + 298, + 111 + ], + "type": "image", + "image_path": "b6b683cb29e18ec994045cb67ed3c6ee9cbe9f52794b9d989b719338e25e7d44.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 299, + 72, + 375, + 112 + ], + "blocks": [ + { + "bbox": [ + 299, + 72, + 375, + 112 + ], + "lines": [ + { + "bbox": [ + 299, + 72, + 375, + 112 + ], + "spans": [ + { + "bbox": [ + 299, + 72, + 375, + 112 + ], + "type": "image", + "image_path": "d96e217df87c94004be40ca45ff179bb041727d2ed795cb749d96ac4354823b7.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 218, + 129, + 297, + 170 + ], + "blocks": [ + { + "bbox": [ + 78, + 180, + 197, + 190 + ], + "lines": [ + { + "bbox": [ + 78, + 180, + 197, + 190 + ], + "spans": [ + { + "bbox": [ + 78, + 180, + 197, + 190 + ], + "type": "text", + "content": "(a) Frames from literacy videos" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 228, + 113, + 288, + 129 + ], + "lines": [ + { + "bbox": [ + 228, + 113, + 288, + 129 + ], + "spans": [ + { + "bbox": [ + 228, + 113, + 288, + 129 + ], + "type": "text", + "content": "analyze compare shape" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 218, + 129, + 297, + 170 + ], + "lines": [ + { + "bbox": [ + 218, + 129, + 297, + 170 + ], + "spans": [ + { + "bbox": [ + 218, + 129, + 297, + 170 + ], + "type": "image", + "image_path": "b22019975ef426b83d90d2f01b2c0c9fdb05cb798da0a8a91b1866a7903ae153.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 217, + 171, + 296, + 178 + ], + "lines": [ + { + "bbox": [ + 217, + 171, + 296, + 178 + ], + "spans": [ + { + "bbox": [ + 217, + 171, + 296, + 178 + ], + "type": "text", + "content": "addition subtraction" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 299, + 129, + 377, + 171 + ], + "blocks": [ + { + "bbox": [ + 306, + 114, + 369, + 129 + ], + "lines": [ + { + "bbox": [ + 306, + 114, + 369, + 129 + ], + "spans": [ + { + "bbox": [ + 306, + 114, + 369, + 129 + ], + "type": "text", + "content": "building drawing shapes" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 299, + 129, + 377, + 171 + ], + "lines": [ + { + "bbox": [ + 299, + 129, + 377, + 171 + ], + "spans": [ + { + "bbox": [ + 299, + 129, + 377, + 171 + ], + "type": "image", + "image_path": "0a14f1dd2d4cc2e27758e64586bee9b116c3a0ff6b152f9ede720953b4e072c0.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 322, + 171, + 354, + 178 + ], + "lines": [ + { + "bbox": [ + 322, + 171, + 354, + 178 + ], + "spans": [ + { + "bbox": [ + 322, + 171, + 354, + 178 + ], + "type": "text", + "content": "counting" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 233, + 180, + 334, + 190 + ], + "lines": [ + { + "bbox": [ + 233, + 180, + 334, + 190 + ], + "spans": [ + { + "bbox": [ + 233, + 180, + 334, + 190 + ], + "type": "text", + "content": "(b) Frames from math videos" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 194, + 545, + 226 + ], + "lines": [ + { + "bbox": [ + 46, + 194, + 545, + 226 + ], + "spans": [ + { + "bbox": [ + 46, + 194, + 545, + 226 + ], + "type": "text", + "content": "Figure 1. Sample video frames from the APPROVE dataset. Videos belong to the (a) literacy classes, (b) math classes, and (c) background. Background videos do not contain educational content but share visual similarities with educational videos. The videos are labeled with fine-grained sub-classes, e.g., letter names vs letter sounds." + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 382, + 72, + 461, + 113 + ], + "blocks": [ + { + "bbox": [ + 382, + 72, + 461, + 113 + ], + "lines": [ + { + "bbox": [ + 382, + 72, + 461, + 113 + ], + "spans": [ + { + "bbox": [ + 382, + 72, + 461, + 113 + ], + "type": "image", + "image_path": "f62656d7c4adaed62e4b85bbbcb7f94c0ad53fc9ccb942b08b498ada61f7095a.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 463, + 72, + 541, + 113 + ], + "blocks": [ + { + "bbox": [ + 463, + 72, + 541, + 113 + ], + "lines": [ + { + "bbox": [ + 463, + 72, + 541, + 113 + ], + "spans": [ + { + "bbox": [ + 463, + 72, + 541, + 113 + ], + "type": "image", + "image_path": "6b631189dbbd0d5105d952fae59eb31b78b2bfc77933254287061d68634f4626.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 382, + 125, + 461, + 166 + ], + "blocks": [ + { + "bbox": [ + 382, + 125, + 461, + 166 + ], + "lines": [ + { + "bbox": [ + 382, + 125, + 461, + 166 + ], + "spans": [ + { + "bbox": [ + 382, + 125, + 461, + 166 + ], + "type": "image", + "image_path": "3705b586c96f1539d4e5c1329c024fd30fde3a47044aad6c1116e5faf67649ba.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 398, + 180, + 526, + 190 + ], + "lines": [ + { + "bbox": [ + 398, + 180, + 526, + 190 + ], + "spans": [ + { + "bbox": [ + 398, + 180, + 526, + 190 + ], + "type": "text", + "content": "(c) Frames from background videos" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 463, + 125, + 541, + 166 + ], + "blocks": [ + { + "bbox": [ + 463, + 125, + 541, + 166 + ], + "lines": [ + { + "bbox": [ + 463, + 125, + 541, + 166 + ], + "spans": [ + { + "bbox": [ + 463, + 125, + 541, + 166 + ], + "type": "image", + "image_path": "42fb2181f46789989d6dad2ec6d93796eaff5df97870038314eab38607521bf6.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + } + ], + "index": 21 + }, + { + "bbox": [ + 46, + 231, + 286, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 231, + 286, + 361 + ], + "spans": [ + { + "bbox": [ + 46, + 231, + 286, + 361 + ], + "type": "text", + "content": "sult, 6 forward passes of the fusion model are required for 3 modalities. Frozen in Time [6] is able to utilize both image-text and video-text datasets through the use of a Space-Time Transformer Visual Encoder. Visual Conditioned GPT [37] uses a single cross-attention fusion layer to combine pretrained CLIP text and visual features. Flamingo [2] adds cross-attention layers interleaved with language decoder layers to fuse visual information into text generation. MERLOT [60, 61] and Triple Contrastive Learning [58] combine contrastive learning and generative language modeling to learn aligned text-image representations." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 46, + 366, + 286, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 366, + 286, + 521 + ], + "spans": [ + { + "bbox": [ + 46, + 366, + 286, + 521 + ], + "type": "text", + "content": "Supervised Multi-Modal Learning: Supervised MultiModal Learning typically relies on crowd-captioned datasets such as Flickr30k [59] and MS-COCO Captions [11]. Some prior works such as OSCAR [33] and VinVL [62] have utilized pre-trained object detectors and multi-modal transformers to learn image captioning using supervised aligned datasets. BLIP [28] takes a hybrid approach where it bootstraps an image captioner using a labeled dataset and uses it to generate captions for web images. This generated corpus is then filtered and used for learning an aligned representation. ALign BBefore Fuse [29] highlights the importance of aligning text and image tokens before fusing them using a multi-modal transformer." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 46, + 522, + 286, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 522, + 286, + 569 + ], + "spans": [ + { + "bbox": [ + 46, + 522, + 286, + 569 + ], + "type": "text", + "content": "In this paper, we focus on the fine-grained classification of multilabel educational videos. Due to the lack of suitable datasets, we propose a new dataset, APPROVE, which is described next." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 47, + 578, + 160, + 590 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 578, + 160, + 590 + ], + "spans": [ + { + "bbox": [ + 47, + 578, + 160, + 590 + ], + "type": "text", + "content": "3. APPROVE Dataset" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 46, + 594, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 286, + 713 + ], + "type": "text", + "content": "We propose a dataset, called APPROVE, of curated YouTube videos annotated with educational content. APPROVE consists of 193 hours of expert-annotated videos with 19 classes (7 literacy codes, 11 math, and background) and each video is associated with approximately 3 labels on average. We follow the Common Core Standards [3,40] to select education content suitable for kindergarten level. The Common Core Standards outline what students are expected to know and do at various age ranges and grades. This is a widely accepted standard followed by a range of" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 304, + 231, + 545, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 231, + 545, + 411 + ], + "spans": [ + { + "bbox": [ + 304, + 231, + 545, + 411 + ], + "type": "text", + "content": "educators. We consider two high-level classes of educational content: literacy and math. For each of these content classes, we select a set of codes. For the literacy class, we select 7 codes including letter names, letter sounds, follow words, sight words, letters in words, sounds in words, and rhyming. For the math class, we select 11 codes including counting, individual number, comparing groups, addition subtraction, measurable attributes, sorting, spatial language, shape identification, building drawing, analyzing and comparing shapes. More details about the standard and the description of the codes are provided in the supplementary material. APPROVE also consists of carefully chosen background videos, i.e., without educational content, that are visually similar to the videos with educational content. We present frames corresponding to these classes in Fig. 1." + } + ] + } + ], + "index": 29 + }, + { + "type": "image", + "bbox": [ + 332, + 426, + 518, + 589 + ], + "blocks": [ + { + "bbox": [ + 332, + 426, + 518, + 589 + ], + "lines": [ + { + "bbox": [ + 332, + 426, + 518, + 589 + ], + "spans": [ + { + "bbox": [ + 332, + 426, + 518, + 589 + ], + "type": "image", + "image_path": "eaa0b7ba26fca96539e7a0d27187f4efcb0b95745bdacfb14065959e8898d39f.jpg" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 449, + 591, + 481, + 600 + ], + "lines": [ + { + "bbox": [ + 449, + 591, + 481, + 600 + ], + "spans": [ + { + "bbox": [ + 449, + 591, + 481, + 600 + ], + "type": "text", + "content": "Frequency" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 306, + 605, + 501, + 628 + ], + "lines": [ + { + "bbox": [ + 306, + 605, + 501, + 628 + ], + "spans": [ + { + "bbox": [ + 306, + 605, + 501, + 628 + ], + "type": "text", + "content": "Figure 2. Frequency of the classes in APPROVE. Math codes are in Orange and literacy codes in Blue." + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_caption" + } + ], + "index": 30 + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "content": "To ensure the quality and correctness of the annotations, we consider educational researchers to annotate the videos and follow a standard validation protocol [41]. Each annotator is trained by an expert and annotations on a selected set are examined before engaging the annotator for the final annotation. Annotators start once they reach more than " + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "inline_equation", + "content": "90\\%" + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "content": " agreement with the expert. Further, we estimate inter" + } + ] + } + ], + "index": 33 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19925" + } + ] + } + ], + "index": 34 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 85, + 72, + 247, + 148 + ], + "blocks": [ + { + "bbox": [ + 85, + 72, + 247, + 148 + ], + "lines": [ + { + "bbox": [ + 85, + 72, + 247, + 148 + ], + "spans": [ + { + "bbox": [ + 85, + 72, + 247, + 148 + ], + "type": "image", + "image_path": "1dcf09c212e2f1b060f8e7bb90aeda0cb2d49912691ed473f98715cb3b2bcd91.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 64, + 151, + 269, + 162 + ], + "lines": [ + { + "bbox": [ + 64, + 151, + 269, + 162 + ], + "spans": [ + { + "bbox": [ + 64, + 151, + 269, + 162 + ], + "type": "text", + "content": "Figure 3. Distribution of the number of labels per video." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 175, + 287, + 234 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 175, + 287, + 234 + ], + "spans": [ + { + "bbox": [ + 46, + 175, + 287, + 234 + ], + "type": "text", + "content": "annotator consistency to filter out anomalies. Details about the validation process are provided in the supplementary material. It takes a month to train an education researcher to match expert-level coding accuracy. On average, it takes the trained annotators 1 min to annotate 1 min of video." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 235, + 288, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 235, + 288, + 559 + ], + "spans": [ + { + "bbox": [ + 46, + 235, + 288, + 559 + ], + "type": "text", + "content": "The videos are curated from YouTube and are annotated by the trained annotators to determine educational content in them. Each video can have multiple class labels that are quite similar making the task a multi-label and fine-grained classification problem. For example 'letter names' and 'letter sounds' where visual letters are shown in both but in 'letter sounds', the phonetic sound on the letter is emphasized (Fig. 1 (a)). Similarly, in both 'build and draw shapes' and 'analyzing and comparing shapes', multiple shapes can appear but the latter focuses on comparing multiple shapes by shape and size (Fig. 1 (b)). Class-wise stats are presented in Fig. 2. Note that the task is different from common video classification setups where either multi-label or fine-grained aspects are dealt separately. Single-label datasets such as HMDB51 [27], UCF101 [54], Kinetics700 [53] and multi-label ones such as Charades [52] are widely used benchmarks for this problem. YouTube-Birds and YouTube-Cars [64] are analogous datasets for object recognition from videos. Multi-Sports [34] and FineGym [49] label fine-grained action classes for sports. HVU [15] also adds scenes and attributes annotations along with action and objects. However, action, object and scene recognition are not enough for fine-grained video understanding. For instance, videos from a given education provider might share similar objects (person, chalkboard, etc.) and actions (writing on chalkboard) while covering different topics (counting, shape recognition etc.) in each video." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 570, + 163, + 583 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 570, + 163, + 583 + ], + "spans": [ + { + "bbox": [ + 47, + 570, + 163, + 583 + ], + "type": "text", + "content": "4. Proposed Approach" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 584, + 287, + 657 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 584, + 287, + 657 + ], + "spans": [ + { + "bbox": [ + 46, + 584, + 287, + 657 + ], + "type": "text", + "content": "In this section, we first describe the proposed class prototypes based contrastive learning framework suitable for videos containing multiple educational codes. Then we present the approach to learning the class prototypes and finally describe the multimodal transformer network that learns features by fusing visual and text cues from videos." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 665, + 274, + 677 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 665, + 274, + 677 + ], + "spans": [ + { + "bbox": [ + 47, + 665, + 274, + 677 + ], + "type": "text", + "content": "4.1. Class prototypes based contrastive learning" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 677, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 287, + 714 + ], + "type": "text", + "content": "In a contrastive learning framework, feature representations are typically learned by simultaneously minimizing the distance between positive samples and maximizing" + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 307, + 64, + 539, + 196 + ], + "blocks": [ + { + "bbox": [ + 307, + 64, + 539, + 196 + ], + "lines": [ + { + "bbox": [ + 307, + 64, + 539, + 196 + ], + "spans": [ + { + "bbox": [ + 307, + 64, + 539, + 196 + ], + "type": "table", + "html": "
DatasetSize (in hr)Multi-LabelFine GrainedTypeAnnotators
Action Recognition
HMDB5XXVAuthors
UCF27XXVAuthors
Kinetics800XXV+ACrowd
Video Classification
COIN476XXV+ACrowd
YT-8M-✓XFMachine
APPROVE193✓✓V+T+AExperts
", + "image_path": "63a71a5d6491e16d6c302e93ef351e6033286e631b1e206cac5575dc14429c20.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 205, + 545, + 228 + ], + "lines": [ + { + "bbox": [ + 305, + 205, + 545, + 228 + ], + "spans": [ + { + "bbox": [ + 305, + 205, + 545, + 228 + ], + "type": "text", + "content": "Table 1. APPROVE dataset compared with selected prior datasets. V→Video Frames, A→Audio, T→Text, F→Features only." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 305, + 239, + 545, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 239, + 545, + 361 + ], + "spans": [ + { + "bbox": [ + 305, + 239, + 545, + 361 + ], + "type": "text", + "content": "the distance between negative samples (See Figure 4.(a)). The positive and negative samples are determined with respect to an anchor sample usually based on the class labels. For example, supervised contrastive learning (SupCon) [26] learns a representation to minimize the intra-class distances and maximize inter-class distances. We denote " + }, + { + "bbox": [ + 305, + 239, + 545, + 361 + ], + "type": "inline_equation", + "content": "\\boldsymbol{x}_i" + }, + { + "bbox": [ + 305, + 239, + 545, + 361 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 239, + 545, + 361 + ], + "type": "inline_equation", + "content": "y_i" + }, + { + "bbox": [ + 305, + 239, + 545, + 361 + ], + "type": "text", + "content": " as the " + }, + { + "bbox": [ + 305, + 239, + 545, + 361 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 305, + 239, + 545, + 361 + ], + "type": "text", + "content": "th sample and its label, respectively. Let's define " + }, + { + "bbox": [ + 305, + 239, + 545, + 361 + ], + "type": "inline_equation", + "content": "z_i" + }, + { + "bbox": [ + 305, + 239, + 545, + 361 + ], + "type": "text", + "content": " as the representation of the " + }, + { + "bbox": [ + 305, + 239, + 545, + 361 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 305, + 239, + 545, + 361 + ], + "type": "text", + "content": "th sample in a batch " + }, + { + "bbox": [ + 305, + 239, + 545, + 361 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 305, + 239, + 545, + 361 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 305, + 239, + 545, + 361 + ], + "type": "inline_equation", + "content": "sim(z_i, z_j) = \\frac{z_i \\cdot z_j}{|z_i||z_j|}" + }, + { + "bbox": [ + 305, + 239, + 545, + 361 + ], + "type": "text", + "content": " the cosine similarity, then the SupCon loss [26] is defined as:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 369, + 552, + 415 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 369, + 552, + 415 + ], + "spans": [ + { + "bbox": [ + 306, + 369, + 552, + 415 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\operatorname {S u p C o n}} = \\sum_ {i \\in A} \\frac {- 1}{| P (i) |} \\sum_ {p \\in P (i)} \\log \\frac {\\exp \\left(\\operatorname {s i m} \\left(\\boldsymbol {z} _ {i} , \\boldsymbol {z} _ {p}\\right) / \\tau\\right)}{\\sum_ {a \\in A \\backslash i} \\exp \\left(\\operatorname {s i m} \\left(\\boldsymbol {z} _ {i} , \\boldsymbol {z} _ {a}\\right) / \\tau\\right)}, \\tag {1}", + "image_path": "5255949584d3209516610a7d1c1d3943615a756b7dc7ba856a6b238630ad87e9.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 415, + 545, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 415, + 545, + 593 + ], + "spans": [ + { + "bbox": [ + 304, + 415, + 545, + 593 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 415, + 545, + 593 + ], + "type": "inline_equation", + "content": "P(i)" + }, + { + "bbox": [ + 304, + 415, + 545, + 593 + ], + "type": "text", + "content": " is the set of positive samples, i.e., with the same label as " + }, + { + "bbox": [ + 304, + 415, + 545, + 593 + ], + "type": "inline_equation", + "content": "z_{i}" + }, + { + "bbox": [ + 304, + 415, + 545, + 593 + ], + "type": "text", + "content": ", in the batch excluding " + }, + { + "bbox": [ + 304, + 415, + 545, + 593 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 415, + 545, + 593 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 415, + 545, + 593 + ], + "type": "inline_equation", + "content": "a \\in A \\setminus i" + }, + { + "bbox": [ + 304, + 415, + 545, + 593 + ], + "type": "text", + "content": " is the index of all samples in the batch excluding the " + }, + { + "bbox": [ + 304, + 415, + 545, + 593 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 415, + 545, + 593 + ], + "type": "text", + "content": "th sample. " + }, + { + "bbox": [ + 304, + 415, + 545, + 593 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 304, + 415, + 545, + 593 + ], + "type": "text", + "content": " is a scalar temperature parameter used for scaling similarity values. The positive pairs are grouped into the numerator and minimizing the loss minimizes their distance in the learned representation and vice versa for negative pairs. SupCon is known to be effective for classifying samples with a single label. However, it is not straightforward to extend this for the multilabel setup as beyond positive samples, where all labels are the same, and negative samples, where none of the labels is the same, there can be a third scenario where labels are partially overlapping. Though SupCon has been extended to hierarchical classification [63], it cannot be directly extended to the true multi-label case." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 594, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 545, + 714 + ], + "type": "text", + "content": "To address this issue, we learn class prototypes as the representative for each class and consider these as anchors while determining positive and negative samples. Specifically, for a specific class prototype, a representation is learned to minimize distances between the prototype and samples with this class label and maximize the distances between the prototype and samples without this class label. We compare the proposed approach with the standard single-label contrastive learning in figure 4. We iteratively update the class prototypes while learning the feature rep" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "19926" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 86, + 72, + 293, + 201 + ], + "blocks": [ + { + "bbox": [ + 86, + 72, + 293, + 201 + ], + "lines": [ + { + "bbox": [ + 86, + 72, + 293, + 201 + ], + "spans": [ + { + "bbox": [ + 86, + 72, + 293, + 201 + ], + "type": "image", + "image_path": "0e76b14c48ebc3c7bcf29f384f22d0146c1256dba49800c43dd96ec2465ec366.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 202, + 545, + 280 + ], + "lines": [ + { + "bbox": [ + 46, + 202, + 545, + 280 + ], + "spans": [ + { + "bbox": [ + 46, + 202, + 545, + 280 + ], + "type": "text", + "content": "Figure 4. Contrastive learning operates on the feature space by bringing the representations of similar samples close and pushing distinct samples apart. Prior work in (a) Supervised Contrastive Learning [26] trains the network by treating instances from the same class as positive pairs and instances from different classes as negative pairs. This approach doesn't generalize to multi-label classification tasks, as some instance pairs have partially overlapping labels. We propose the use of class prototypes to enable (b) Multi-Label Prototypes Contrastive Learning. Each sample and the class prototypes corresponding to the labels associated with the sample are treated as positive pairs. Similarly, negative pairs are determined based on the missing class labels. Prototypes are represented by stars " + }, + { + "bbox": [ + 46, + 202, + 545, + 280 + ], + "type": "inline_equation", + "content": "(\\star)" + }, + { + "bbox": [ + 46, + 202, + 545, + 280 + ], + "type": "text", + "content": " and inputs as circles " + }, + { + "bbox": [ + 46, + 202, + 545, + 280 + ], + "type": "inline_equation", + "content": "(\\odot)" + }, + { + "bbox": [ + 46, + 202, + 545, + 280 + ], + "type": "text", + "content": " colored with all their relevant labels. We discuss strategies for initializing and learning the label prototypes in Sec. 4.2." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 302, + 74, + 504, + 201 + ], + "blocks": [ + { + "bbox": [ + 302, + 74, + 504, + 201 + ], + "lines": [ + { + "bbox": [ + 302, + 74, + 504, + 201 + ], + "spans": [ + { + "bbox": [ + 302, + 74, + 504, + 201 + ], + "type": "image", + "image_path": "593ec9dba574c1a181e71408ab42f8c7927ae739864e3d963c10245ecd348819.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 295, + 289, + 402 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 295, + 289, + 402 + ], + "spans": [ + { + "bbox": [ + 46, + 295, + 289, + 402 + ], + "type": "text", + "content": "resentations. We define " + }, + { + "bbox": [ + 46, + 295, + 289, + 402 + ], + "type": "inline_equation", + "content": "C = \\{c_1, \\ldots, c_K\\}" + }, + { + "bbox": [ + 46, + 295, + 289, + 402 + ], + "type": "text", + "content": " as the set of classes where " + }, + { + "bbox": [ + 46, + 295, + 289, + 402 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 46, + 295, + 289, + 402 + ], + "type": "text", + "content": " is the number of classes. For a sample " + }, + { + "bbox": [ + 46, + 295, + 289, + 402 + ], + "type": "inline_equation", + "content": "\\pmb{x}" + }, + { + "bbox": [ + 46, + 295, + 289, + 402 + ], + "type": "text", + "content": ", let's define " + }, + { + "bbox": [ + 46, + 295, + 289, + 402 + ], + "type": "inline_equation", + "content": "P_{ml}(\\pmb{x}) = \\{c_k^+\\}, c_k^+ \\in C" + }, + { + "bbox": [ + 46, + 295, + 289, + 402 + ], + "type": "text", + "content": " as the set of multiple class labels associated with " + }, + { + "bbox": [ + 46, + 295, + 289, + 402 + ], + "type": "inline_equation", + "content": "\\pmb{x}" + }, + { + "bbox": [ + 46, + 295, + 289, + 402 + ], + "type": "text", + "content": " (positive classes) and " + }, + { + "bbox": [ + 46, + 295, + 289, + 402 + ], + "type": "inline_equation", + "content": "c_k^- \\in C \\setminus P_{ml}(\\pmb{x})" + }, + { + "bbox": [ + 46, + 295, + 289, + 402 + ], + "type": "text", + "content": " denotes the missing classes (negative classes). We define " + }, + { + "bbox": [ + 46, + 295, + 289, + 402 + ], + "type": "inline_equation", + "content": "CP = \\{cp_1, \\ldots, cp_K\\}" + }, + { + "bbox": [ + 46, + 295, + 289, + 402 + ], + "type": "text", + "content": " as the set of class prototypes. Considering " + }, + { + "bbox": [ + 46, + 295, + 289, + 402 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 46, + 295, + 289, + 402 + ], + "type": "text", + "content": " is the representation for the sample " + }, + { + "bbox": [ + 46, + 295, + 289, + 402 + ], + "type": "inline_equation", + "content": "\\pmb{x}" + }, + { + "bbox": [ + 46, + 295, + 289, + 402 + ], + "type": "text", + "content": ", the class prototypes based multilabel contrastive loss is defined as:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 407, + 293, + 470 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 407, + 293, + 470 + ], + "spans": [ + { + "bbox": [ + 47, + 407, + 293, + 470 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} _ {m l c} (\\pmb {x}) = \\\\ \\frac {- 1}{| P _ {m l} (\\boldsymbol {x}) |} \\sum_ {c _ {k} ^ {+} \\in P _ {m l} (\\boldsymbol {x})} \\left[ \\log \\frac {\\exp \\left(\\operatorname {s i m} \\left(\\boldsymbol {z} , \\boldsymbol {c p} _ {k}\\right) / \\tau\\right)}{\\sum_ {c _ {j} ^ {-} \\in C \\backslash P _ {m l} (\\boldsymbol {x})} \\exp \\left(\\operatorname {s i m} \\left(\\boldsymbol {z} , \\boldsymbol {c p} _ {j}\\right) / \\tau\\right)} \\right]. \\tag {2} \\\\ \\end{array}", + "image_path": "4f3ebf0ec9ab0a80c69fc6373bb2300a8388ab221ab71d10974fb420e1ba1c34.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 471, + 287, + 542 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 471, + 287, + 542 + ], + "spans": [ + { + "bbox": [ + 46, + 471, + 287, + 542 + ], + "type": "text", + "content": "Here, minimizing the loss of the positive class prototype and instance pairs in the numerator minimizes the distance between the representation " + }, + { + "bbox": [ + 46, + 471, + 287, + 542 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 46, + 471, + 287, + 542 + ], + "type": "text", + "content": " and the class prototypes corresponding to the sample, and vice versa for negative classes. We also utilize negative sampling to account for the class imbalance between positives and negatives." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 545, + 190, + 558 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 545, + 190, + 558 + ], + "spans": [ + { + "bbox": [ + 47, + 545, + 190, + 558 + ], + "type": "text", + "content": "4.2. Learning class prototypes" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 558, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 558, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 558, + 288, + 714 + ], + "type": "text", + "content": "We aim to learn class-specific prototypes such that the multilabel samples can be thought of as the combinations of the class prototypes selected based on the associated labels. Lets assume " + }, + { + "bbox": [ + 46, + 558, + 288, + 714 + ], + "type": "inline_equation", + "content": "Z_{t}" + }, + { + "bbox": [ + 46, + 558, + 288, + 714 + ], + "type": "text", + "content": " is an " + }, + { + "bbox": [ + 46, + 558, + 288, + 714 + ], + "type": "inline_equation", + "content": "N\\times d" + }, + { + "bbox": [ + 46, + 558, + 288, + 714 + ], + "type": "text", + "content": " matrix of " + }, + { + "bbox": [ + 46, + 558, + 288, + 714 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 46, + 558, + 288, + 714 + ], + "type": "text", + "content": "-dimensional representations, i.e., " + }, + { + "bbox": [ + 46, + 558, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{z}\\mathbf{s})" + }, + { + "bbox": [ + 46, + 558, + 288, + 714 + ], + "type": "text", + "content": ", of " + }, + { + "bbox": [ + 46, + 558, + 288, + 714 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 558, + 288, + 714 + ], + "type": "text", + "content": " samples and " + }, + { + "bbox": [ + 46, + 558, + 288, + 714 + ], + "type": "inline_equation", + "content": "L\\in \\{0,1\\}^{N\\times K}" + }, + { + "bbox": [ + 46, + 558, + 288, + 714 + ], + "type": "text", + "content": " is corresponding labels matrix with " + }, + { + "bbox": [ + 46, + 558, + 288, + 714 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 46, + 558, + 288, + 714 + ], + "type": "text", + "content": " classes. Let's denote " + }, + { + "bbox": [ + 46, + 558, + 288, + 714 + ], + "type": "inline_equation", + "content": "CP_{t}" + }, + { + "bbox": [ + 46, + 558, + 288, + 714 + ], + "type": "text", + "content": " is a matrix of size " + }, + { + "bbox": [ + 46, + 558, + 288, + 714 + ], + "type": "inline_equation", + "content": "K\\times d" + }, + { + "bbox": [ + 46, + 558, + 288, + 714 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 46, + 558, + 288, + 714 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 46, + 558, + 288, + 714 + ], + "type": "text", + "content": " class prototypes at a training iteration " + }, + { + "bbox": [ + 46, + 558, + 288, + 714 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 558, + 288, + 714 + ], + "type": "text", + "content": ". Then, " + }, + { + "bbox": [ + 46, + 558, + 288, + 714 + ], + "type": "inline_equation", + "content": "Z_{t} = L\\times CP_{t} + \\varepsilon" + }, + { + "bbox": [ + 46, + 558, + 288, + 714 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 558, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\varepsilon" + }, + { + "bbox": [ + 46, + 558, + 288, + 714 + ], + "type": "text", + "content": " is the residual noise term. Assuming a Gaussian noise that is unbiased and uncorrelated with the labels " + }, + { + "bbox": [ + 46, + 558, + 288, + 714 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 46, + 558, + 288, + 714 + ], + "type": "text", + "content": ", the class prototypes can be approximated as " + }, + { + "bbox": [ + 46, + 558, + 288, + 714 + ], + "type": "inline_equation", + "content": "CP_{t}^{*}\\approx (L^{T}L)^{-1}L^{T}Z_{t}" + }, + { + "bbox": [ + 46, + 558, + 288, + 714 + ], + "type": "text", + "content": " where operation " + }, + { + "bbox": [ + 46, + 558, + 288, + 714 + ], + "type": "inline_equation", + "content": "(L^T L)" + }, + { + "bbox": [ + 46, + 558, + 288, + 714 + ], + "type": "text", + "content": " results in a square matrix amenable to inversion. Note that for single labels, this implies averag-" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 295, + 545, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 295, + 545, + 354 + ], + "spans": [ + { + "bbox": [ + 304, + 295, + 545, + 354 + ], + "type": "text", + "content": "ing the features of the instances belonging to a given class as the prototype for that class. In a multi-label setup, additionally, the co-occurrence between the labels is considered. Finally, the class prototypes are updated with learning iterations as:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 351, + 360, + 545, + 373 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 351, + 360, + 545, + 373 + ], + "spans": [ + { + "bbox": [ + 351, + 360, + 545, + 373 + ], + "type": "interline_equation", + "content": "C P _ {t + 1} = \\beta \\cdot C P _ {t} + (1 - \\beta) \\cdot C P _ {t} ^ {*}, \\tag {3}", + "image_path": "f2a4e0e84283682bce83d812c63de7a3c632b782f83e57ce3586c6e5ad5ce317.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 383, + 545, + 419 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 383, + 545, + 419 + ], + "spans": [ + { + "bbox": [ + 304, + 383, + 545, + 419 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 383, + 545, + 419 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 304, + 383, + 545, + 419 + ], + "type": "text", + "content": " is the decay parameter for the exponential moving averaging. The moving averaging avoids collapsing prototypes across the training iterations [31]." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 425, + 487, + 436 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 425, + 487, + 436 + ], + "spans": [ + { + "bbox": [ + 305, + 425, + 487, + 436 + ], + "type": "text", + "content": "4.3. Multi-Modal Fusion Transformer" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 437, + 545, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 437, + 545, + 616 + ], + "spans": [ + { + "bbox": [ + 304, + 437, + 545, + 616 + ], + "type": "text", + "content": "Given a sample video " + }, + { + "bbox": [ + 304, + 437, + 545, + 616 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 304, + 437, + 545, + 616 + ], + "type": "text", + "content": ", we consider multimodal cues such as visual and audio cues to learn its representation " + }, + { + "bbox": [ + 304, + 437, + 545, + 616 + ], + "type": "inline_equation", + "content": "\\mathbf{z}" + }, + { + "bbox": [ + 304, + 437, + 545, + 616 + ], + "type": "text", + "content": ". Considering multimodal cues is crucial for content recognition, specifically for education videos as effective comprehension requires attending to both visual demonstration and audio explaining the educational content [3, 40]. To capture audio cues, we consider the audio track of the video and extract speech by removing the background such as tunes or instruments [21]. Then we employ the automatic speech recognition (ASR) technique to transcribe text from the speech [43]. We notice that separating speech from the background is important for an accurate ASR transcription. Given the video frames and text transcription, we propose a multimodal transformer network (MTN) to fuse these cues using cross-modal attention." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 618, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 618, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 618, + 545, + 713 + ], + "type": "text", + "content": "Our MTN (Figure 5) has three components: image encoder, text encoder, and fusion encoder to learn visual " + }, + { + "bbox": [ + 304, + 618, + 545, + 713 + ], + "type": "inline_equation", + "content": "(z^v)" + }, + { + "bbox": [ + 304, + 618, + 545, + 713 + ], + "type": "text", + "content": ", text " + }, + { + "bbox": [ + 304, + 618, + 545, + 713 + ], + "type": "inline_equation", + "content": "(z^t)" + }, + { + "bbox": [ + 304, + 618, + 545, + 713 + ], + "type": "text", + "content": ", and fusion " + }, + { + "bbox": [ + 304, + 618, + 545, + 713 + ], + "type": "inline_equation", + "content": "(z^f)" + }, + { + "bbox": [ + 304, + 618, + 545, + 713 + ], + "type": "text", + "content": " representations, respectively. The sample representation is comprised of these three representations " + }, + { + "bbox": [ + 304, + 618, + 545, + 713 + ], + "type": "inline_equation", + "content": "z = \\{z^v, z^t, z^f\\}" + }, + { + "bbox": [ + 304, + 618, + 545, + 713 + ], + "type": "text", + "content": ". The image encoder is implemented by a vision transformer (ViT) [16] that learns frame embeddings from the video frames along with a special CLS for each frame [14]. We pool the CLS tokens across" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19927" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 61, + 71, + 273, + 393 + ], + "blocks": [ + { + "bbox": [ + 61, + 71, + 273, + 393 + ], + "lines": [ + { + "bbox": [ + 61, + 71, + 273, + 393 + ], + "spans": [ + { + "bbox": [ + 61, + 71, + 273, + 393 + ], + "type": "image", + "image_path": "bafa208d7a462b2edd8c2a532a34c4f605924eafb2a27298c3ff635cceab7ab3.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 398, + 287, + 509 + ], + "lines": [ + { + "bbox": [ + 46, + 398, + 287, + 509 + ], + "spans": [ + { + "bbox": [ + 46, + 398, + 287, + 509 + ], + "type": "text", + "content": "Figure 5. Multi-Modal Classification Network. A text encoder is used to encode ASR text from the video, while an Image Encoder is used to get tokens representing each frame of the video. Unimodal pre-training is carried out on the text & image encoders respectively. Multi-label contrastive loss is used along with shared prototypes to align the representations across both modalities. This is followed by joint end-to-end learning of the whole multimodal network including the fusion encoder which applies multi-head self-attention within and across the modalities. The prototypes are further refined during the multi-modal training phase." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 514, + 287, + 598 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 514, + 287, + 598 + ], + "spans": [ + { + "bbox": [ + 46, + 514, + 287, + 598 + ], + "type": "text", + "content": "the frames and consider this as a compact video representation. Similarly, we consider the BERT-based text transformer [24] to learn the word embeddings along with the text " + }, + { + "bbox": [ + 46, + 514, + 287, + 598 + ], + "type": "inline_equation", + "content": "CLS" + }, + { + "bbox": [ + 46, + 514, + 287, + 598 + ], + "type": "text", + "content": " token for the entire transcription. We consider the text " + }, + { + "bbox": [ + 46, + 514, + 287, + 598 + ], + "type": "inline_equation", + "content": "CLS" + }, + { + "bbox": [ + 46, + 514, + 287, + 598 + ], + "type": "text", + "content": " as the text representation. Finally, the fusion encoder fuses the visual and text cues by leveraging cross-modal attention between frame and word embeddings." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 604, + 194, + 616 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 604, + 194, + 616 + ], + "spans": [ + { + "bbox": [ + 47, + 604, + 194, + 616 + ], + "type": "text", + "content": "4.4. Inference using prototypes" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 617, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 617, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 617, + 287, + 713 + ], + "type": "text", + "content": "Existing approaches in contrastive learning [8, 26] usually discard the MLP layer after the contrastive training, and a linear classifier is trained on the frozen backbone network. We rely on the class prototypes to carry out inference by utilizing the cosine distance between the learned prototypes and test features. Given our prototype loss-based training, the estimated probability of a given class should be proportional to the normalized temperature scaled cosine distance" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 305, + 72, + 545, + 109 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 545, + 109 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 545, + 109 + ], + "type": "text", + "content": "(Equation 4). In practice, we normalize the cosine distance such that " + }, + { + "bbox": [ + 305, + 72, + 545, + 109 + ], + "type": "inline_equation", + "content": "-1" + }, + { + "bbox": [ + 305, + 72, + 545, + 109 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 72, + 545, + 109 + ], + "type": "inline_equation", + "content": "1" + }, + { + "bbox": [ + 305, + 72, + 545, + 109 + ], + "type": "text", + "content": " correspond to a confidence of " + }, + { + "bbox": [ + 305, + 72, + 545, + 109 + ], + "type": "inline_equation", + "content": "0\\& 1" + }, + { + "bbox": [ + 305, + 72, + 545, + 109 + ], + "type": "text", + "content": " respectively. Thus the prediction is made following:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 359, + 117, + 545, + 129 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 359, + 117, + 545, + 129 + ], + "spans": [ + { + "bbox": [ + 359, + 117, + 545, + 129 + ], + "type": "interline_equation", + "content": "\\hat {p} (k | \\boldsymbol {x}) \\propto \\exp (\\operatorname {s i m} (\\boldsymbol {z}, \\boldsymbol {c p} _ {k}) / \\tau), \\tag {4}", + "image_path": "be66700ebf73dd3e98ca642c26c1fa0db1633733c036ebaa4041685b97f1b846.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 140, + 539, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 140, + 539, + 152 + ], + "spans": [ + { + "bbox": [ + 305, + 140, + 539, + 152 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 140, + 539, + 152 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 305, + 140, + 539, + 152 + ], + "type": "text", + "content": " is the multimodal representation of the sample " + }, + { + "bbox": [ + 305, + 140, + 539, + 152 + ], + "type": "inline_equation", + "content": "\\pmb{x}" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 157, + 440, + 169 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 157, + 440, + 169 + ], + "spans": [ + { + "bbox": [ + 306, + 157, + 440, + 169 + ], + "type": "text", + "content": "4.5. Implementation Details" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 173, + 545, + 304 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 173, + 545, + 304 + ], + "spans": [ + { + "bbox": [ + 304, + 173, + 545, + 304 + ], + "type": "text", + "content": "Training strategy for the multimodal fusion: We follow a two-stage training process: during the initial unimodal training phase, we utilize fixed prototypes in each modality to align the representations. Then in the second stage, we train the unimodal encoders and the multi-modal fusion encoder end-to-end. The cross-modal alignment learned during the first stage improves the learning of the multi-modal representation. The multi-modal learning phase consists of alternating optimization steps of training the network using our contrastive loss per Sec. 4.1 and refining the class prototypes as described in Sec. 4.2." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 306, + 545, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 306, + 545, + 365 + ], + "spans": [ + { + "bbox": [ + 304, + 306, + 545, + 365 + ], + "type": "text", + "content": "Image encoder: For video frames, Random Resized Crop and RandAugment [13] augmentations are used from torchvision. We use ImageNet pretrained vision encoders ResNet50 [19], ViT-B/32 (224×224 resolution) and ViT-B/16 (384×384 resolution) [17]." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 366, + 546, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 366, + 546, + 486 + ], + "spans": [ + { + "bbox": [ + 304, + 366, + 546, + 486 + ], + "type": "text", + "content": "Text encoder: We generate text from the videos using Whisper [43], an open source ASR model. For data augmentation, we generate four versions of the ASR text by back-translation using the Helsinki-NLP/opus-mt-{en-de, en-nl, en-fr} models through the nlpaug library [38]. Synonym replacement, text span removal and random word swapping augmentations are also used for the text data. We use DistilBERT-Base-uncased [48], and t5-small [44] from HuggingFace transformers library as the text encoder." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 487, + 545, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 487, + 545, + 582 + ], + "spans": [ + { + "bbox": [ + 304, + 487, + 545, + 582 + ], + "type": "text", + "content": "Optimizer: We employ AdamW optimizer [36] for training with a learning rate of 0.0005. Weight decay of 1e-6 is utilized only on the MLP head during contrastive training and the classifier during BCE/Focal/Asym. loss. Since pre-trained vision and text backbones are used, the backbone learning rate is set to 1/10th of the learning rate for the head. Exponential Moving Averaging every 10 steps with a decay of 0.999 was used for the model parameters." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 590, + 387, + 602 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 590, + 387, + 602 + ], + "spans": [ + { + "bbox": [ + 306, + 590, + 387, + 602 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 605, + 545, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 545, + 688 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 545, + 688 + ], + "type": "text", + "content": "Datasets. In addition to the proposed APPROVE dataset, we evaluate our approach on a subset of Youtube-8M [1] and COIN [55] datasets. YT-8M consists of a diverse set of YouTube videos with video and audio modalities. We consider a subset of YT-8M dataset with 46K videos and 165 classes. COIN consists of instructional videos covering a wide variety of domains and spanning over 180 classes." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "text", + "content": "Baselines. We compare the efficacy of our multilabel classification framework against the following baselines:" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19928" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 107 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 107 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 107 + ], + "type": "text", + "content": "Binary cross-entropy: In this baseline, loss for multiple labels is computed by combining the binary cross-entropy losses for individual classes." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 108, + 286, + 179 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 108, + 286, + 179 + ], + "spans": [ + { + "bbox": [ + 46, + 108, + 286, + 179 + ], + "type": "text", + "content": "Focal loss [35]: This considers a modified binary cross-entropy to assign a higher weight to hard samples by adjusting a focusing parameter " + }, + { + "bbox": [ + 46, + 108, + 286, + 179 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 46, + 108, + 286, + 179 + ], + "type": "text", + "content": ". Negative samples can also be down-weighted by using a weight " + }, + { + "bbox": [ + 46, + 108, + 286, + 179 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 46, + 108, + 286, + 179 + ], + "type": "text", + "content": ". The focal loss for a positive label is given as " + }, + { + "bbox": [ + 46, + 108, + 286, + 179 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{focal}(p) = -\\alpha (1 - p)^{\\gamma}\\log (p)" + }, + { + "bbox": [ + 46, + 108, + 286, + 179 + ], + "type": "text", + "content": ". We set " + }, + { + "bbox": [ + 46, + 108, + 286, + 179 + ], + "type": "inline_equation", + "content": "\\gamma = 2" + }, + { + "bbox": [ + 46, + 108, + 286, + 179 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 108, + 286, + 179 + ], + "type": "inline_equation", + "content": "\\alpha = 0.2" + }, + { + "bbox": [ + 46, + 108, + 286, + 179 + ], + "type": "text", + "content": " in our experiments." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 180, + 286, + 263 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 180, + 286, + 263 + ], + "spans": [ + { + "bbox": [ + 46, + 180, + 286, + 263 + ], + "type": "text", + "content": "Asymmetric loss [47]: This builds upon the focal loss by utilizing different focusing parameters " + }, + { + "bbox": [ + 46, + 180, + 286, + 263 + ], + "type": "inline_equation", + "content": "\\gamma_{+}" + }, + { + "bbox": [ + 46, + 180, + 286, + 263 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 180, + 286, + 263 + ], + "type": "inline_equation", + "content": "\\gamma_{-}" + }, + { + "bbox": [ + 46, + 180, + 286, + 263 + ], + "type": "text", + "content": " for positive and negative labels. It also ignores the negative samples with a prediction probability lower than a margin " + }, + { + "bbox": [ + 46, + 180, + 286, + 263 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 46, + 180, + 286, + 263 + ], + "type": "text", + "content": ". Asymmetric Loss for prediction " + }, + { + "bbox": [ + 46, + 180, + 286, + 263 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 46, + 180, + 286, + 263 + ], + "type": "text", + "content": " corresponding to label " + }, + { + "bbox": [ + 46, + 180, + 286, + 263 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 46, + 180, + 286, + 263 + ], + "type": "text", + "content": " is given as: " + }, + { + "bbox": [ + 46, + 180, + 286, + 263 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{asym}(p,y) = -yL_{+} - (1 - y)L_{-}" + }, + { + "bbox": [ + 46, + 180, + 286, + 263 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 180, + 286, + 263 + ], + "type": "inline_equation", + "content": "L_{+} = (1 - p)^{\\gamma_{+}}\\cdot \\log (p)" + }, + { + "bbox": [ + 46, + 180, + 286, + 263 + ], + "type": "text", + "content": " and" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 264, + 286, + 322 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 264, + 286, + 322 + ], + "spans": [ + { + "bbox": [ + 47, + 264, + 286, + 322 + ], + "type": "inline_equation", + "content": "L_{-} = (\\max (p - m,0))^{\\gamma_{-}}\\cdot \\log (1 - \\max (p - m,0))" + }, + { + "bbox": [ + 47, + 264, + 286, + 322 + ], + "type": "text", + "content": " We follow the 5-step procedure recommended by the original authors to train this baseline. We experimentally set " + }, + { + "bbox": [ + 47, + 264, + 286, + 322 + ], + "type": "inline_equation", + "content": "\\{\\gamma_{-} = 2,\\gamma_{+} = 1,m = 0.1\\}" + }, + { + "bbox": [ + 47, + 264, + 286, + 322 + ], + "type": "text", + "content": " corresponding to the best performance on APPROVE." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 323, + 286, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 323, + 286, + 441 + ], + "spans": [ + { + "bbox": [ + 46, + 323, + 286, + 441 + ], + "type": "text", + "content": "Metrics. In order to develop a reliable education content detection framework, achieving high precision is crucial. Thus we consider Recall @ 80% Precision (R@80) as the primary metric. We also consider the standard area under the precision-recall curve (AUPR) that is not sensitive to a specific threshold for making the final prediction. We also consider a label ranking average precision (LRAP) [56] metric that is more suitable for the multilabel setup. This estimates whether the ground truth classes are predicted with higher scores than the rest:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 445, + 285, + 477 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 445, + 285, + 477 + ], + "spans": [ + { + "bbox": [ + 47, + 445, + 285, + 477 + ], + "type": "interline_equation", + "content": "\\mathrm {L R A P} = \\frac {1}{n} \\sum_ {i = 1} ^ {m} \\frac {1}{| Y _ {i} |} \\sum_ {\\lambda \\in Y _ {i}} \\frac {| \\lambda^ {\\prime} \\in Y _ {i} : r n k _ {i} (\\lambda^ {\\prime}) \\leq r n k _ {i} (\\lambda) |}{r n k _ {i} (\\lambda)},", + "image_path": "689bf6cf660ba42a9679a5d369a96685315bdd6048a6ed07e055ea910cf2b12a.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 486, + 286, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 486, + 286, + 510 + ], + "spans": [ + { + "bbox": [ + 46, + 486, + 286, + 510 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 486, + 286, + 510 + ], + "type": "inline_equation", + "content": "r n k_{i}(\\lambda)" + }, + { + "bbox": [ + 46, + 486, + 286, + 510 + ], + "type": "text", + "content": " is the predicted rank of class " + }, + { + "bbox": [ + 46, + 486, + 286, + 510 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 46, + 486, + 286, + 510 + ], + "type": "text", + "content": " for sample " + }, + { + "bbox": [ + 46, + 486, + 286, + 510 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 486, + 286, + 510 + ], + "type": "text", + "content": ". LRAP is a ranking metric and in independent of a threshold." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 510, + 286, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 510, + 286, + 641 + ], + "spans": [ + { + "bbox": [ + 46, + 510, + 286, + 641 + ], + "type": "text", + "content": "Results on APPROVE. We compare the proposed approach with the baselines in Tab. 2. Our approach outperforms the strongest baselines by " + }, + { + "bbox": [ + 46, + 510, + 286, + 641 + ], + "type": "inline_equation", + "content": "3.1\\%" + }, + { + "bbox": [ + 46, + 510, + 286, + 641 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 510, + 286, + 641 + ], + "type": "inline_equation", + "content": "2.3\\%" + }, + { + "bbox": [ + 46, + 510, + 286, + 641 + ], + "type": "text", + "content": " with respect to R@80 and AUPR, respectively. We also present results for separate models trained on Math and Literacy subsets of Approve respectively. Results on the Math subset are higher compared to the Literacy subset, which indicates the literacy classes are harder to distinguish mostly due to the high inter-class similarity. The top three hardest classes are follow_words, letters_in_words, and sounds_in_words and these are from the literacy set." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 642, + 286, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 642, + 286, + 700 + ], + "spans": [ + { + "bbox": [ + 46, + 642, + 286, + 700 + ], + "type": "text", + "content": "Results on YT-46K and COIN. Beyond the APPROVE dataset, we also test our approach on two public datasets: YT-46K and COIN. Here we provide the results for the multi-modal models and the results for the single-modality models are in Section D of the Supplementary Material." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 59, + 701, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 701, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 59, + 701, + 286, + 713 + ], + "type": "text", + "content": "Results on YT-46K are provided in Table 3. As YT-8M" + } + ] + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 307, + 70, + 544, + 305 + ], + "blocks": [ + { + "bbox": [ + 307, + 70, + 544, + 305 + ], + "lines": [ + { + "bbox": [ + 307, + 70, + 544, + 305 + ], + "spans": [ + { + "bbox": [ + 307, + 70, + 544, + 305 + ], + "type": "table", + "html": "
SubsetModalityMethodAUPRLRAPR@80
AllVBCE45.554.36.9
Focal45.956.615.0
Ours46.757.919.6
TBCE79.885.163.3
Focal79.985.772.8
Ours82.587.475.4
V+TBCE84.388.476.3
Focal [35]86.189.182.2
Asym. [47]86.089.282.4
Ours88.4 +2.390.7 +1.585.5 +3.1
MTHV+TBCE86.392.480.3
Focal87.292.182.4
Ours88.4 +1.293.2 +1.183.2 +0.8
LITV+TBCE72.182.950.7
Focal72.783.550.9
Ours73.6 +0.984.7 +1.254.7 +3.8
", + "image_path": "b2ae9e9ab761df3b30aa8ebb37bb4254e1b8d522c55f8de26b2ca9440c266f3a.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 323, + 338, + 525, + 399 + ], + "blocks": [ + { + "bbox": [ + 306, + 308, + 520, + 331 + ], + "lines": [ + { + "bbox": [ + 306, + 308, + 520, + 331 + ], + "spans": [ + { + "bbox": [ + 306, + 308, + 520, + 331 + ], + "type": "text", + "content": "Table 2. Results on APPROVE dataset. All metrics in %. V " + }, + { + "bbox": [ + 306, + 308, + 520, + 331 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 306, + 308, + 520, + 331 + ], + "type": "text", + "content": " Video & T " + }, + { + "bbox": [ + 306, + 308, + 520, + 331 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 306, + 308, + 520, + 331 + ], + "type": "text", + "content": " Text. M " + }, + { + "bbox": [ + 306, + 308, + 520, + 331 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 306, + 308, + 520, + 331 + ], + "type": "text", + "content": " Math & L " + }, + { + "bbox": [ + 306, + 308, + 520, + 331 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 306, + 308, + 520, + 331 + ], + "type": "text", + "content": " Literacy Subsets." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 323, + 338, + 525, + 399 + ], + "lines": [ + { + "bbox": [ + 323, + 338, + 525, + 399 + ], + "spans": [ + { + "bbox": [ + 323, + 338, + 525, + 399 + ], + "type": "table", + "html": "
ModalityMethodAUPRLRAPR@80
V+TBCE64.670.242.3
V+TFocal [35]69.772.744.6
V+TOurs70.9 +1.274.9 +2.249.1 +4.5
", + "image_path": "4a0f1f9cc9113d0b184287386fd9b63848684c5d10fcd00742ab586deebcb0f6.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_body" + } + ], + "index": 12 + }, + { + "type": "table", + "bbox": [ + 339, + 419, + 511, + 504 + ], + "blocks": [ + { + "bbox": [ + 313, + 401, + 537, + 412 + ], + "lines": [ + { + "bbox": [ + 313, + 401, + 537, + 412 + ], + "spans": [ + { + "bbox": [ + 313, + 401, + 537, + 412 + ], + "type": "text", + "content": "Table 3. Results on YT-46K. V→Video Frames and T→Text." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 339, + 419, + 511, + 504 + ], + "lines": [ + { + "bbox": [ + 339, + 419, + 511, + 504 + ], + "spans": [ + { + "bbox": [ + 339, + 419, + 511, + 504 + ], + "type": "table", + "html": "
ModalityMethodTop-1 Accuracy
V+TCE53.7
V+TBCE54.9
V+TFocal [35]56.1
V+TSupCon [26]54.7
V+TOurs57.5 +1.4
", + "image_path": "638c50aa613c23c0c2d3033d1d8468c59619aa9e2ef3cb802abd4f6e5108dc38.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "table_body" + } + ], + "index": 14 + }, + { + "bbox": [ + 317, + 506, + 533, + 517 + ], + "lines": [ + { + "bbox": [ + 317, + 506, + 533, + 517 + ], + "spans": [ + { + "bbox": [ + 317, + 506, + 533, + 517 + ], + "type": "text", + "content": "Table 4. Results on COIN. V→Video Frames and T→Text." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 523, + 545, + 643 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 523, + 545, + 643 + ], + "spans": [ + { + "bbox": [ + 304, + 523, + 545, + 643 + ], + "type": "text", + "content": "was primarily collected with the intention of visual classification, the additional use of text data leads to a smaller improvement compared to APPROVE. Results on COIN are provided in Table 4. As each video from COIN is mapped to a single task. Thus, we consider the Top-1 accuracy as the metric. On COIN we compare our approach with SupCon [26] which is effective for single labels. Note that our approach outperforms SupCon and this justifies the effectiveness of the class prototypes based training in a generic contrastive learning framework." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 652, + 406, + 663 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 652, + 406, + 663 + ], + "spans": [ + { + "bbox": [ + 306, + 652, + 406, + 663 + ], + "type": "text", + "content": "5.1. Ablation Studies" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "content": "We perform the following ablation studies to quantify the impact of our approach for learning class prototypes, the multimodal fusion module, and the choice of visual and text encoding frameworks." + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19929" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 51, + 84, + 165, + 137 + ], + "blocks": [ + { + "bbox": [ + 74, + 74, + 143, + 83 + ], + "lines": [ + { + "bbox": [ + 74, + 74, + 143, + 83 + ], + "spans": [ + { + "bbox": [ + 74, + 74, + 143, + 83 + ], + "type": "text", + "content": "(a) Class Prototypes" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 51, + 84, + 165, + 137 + ], + "lines": [ + { + "bbox": [ + 51, + 84, + 165, + 137 + ], + "spans": [ + { + "bbox": [ + 51, + 84, + 165, + 137 + ], + "type": "table", + "html": "
VariantAPPROVECOIN
Random84.156.6
Orthogonal84.857.0
Learned85.557.5
Hierarchical86.057.8
", + "image_path": "29c0e89f6d30afb11a9c6157a69403333a5d506332927448b7c1f7379e0ec04f.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 182, + 84, + 278, + 127 + ], + "blocks": [ + { + "bbox": [ + 188, + 74, + 271, + 83 + ], + "lines": [ + { + "bbox": [ + 188, + 74, + 271, + 83 + ], + "spans": [ + { + "bbox": [ + 188, + 74, + 271, + 83 + ], + "type": "text", + "content": "(b) Fusion Encoder Size" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 182, + 84, + 278, + 127 + ], + "lines": [ + { + "bbox": [ + 182, + 84, + 278, + 127 + ], + "spans": [ + { + "bbox": [ + 182, + 84, + 278, + 127 + ], + "type": "table", + "html": "
LayersAPPROVECOIN
184.957.1
285.557.5
485.357.6
", + "image_path": "49e8f88cc45ff949c1ec559eae8bd49fbf92962d5fb3c9b61af02754b79a5f0b.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 291, + 84, + 410, + 127 + ], + "blocks": [ + { + "bbox": [ + 318, + 74, + 383, + 83 + ], + "lines": [ + { + "bbox": [ + 318, + 74, + 383, + 83 + ], + "spans": [ + { + "bbox": [ + 318, + 74, + 383, + 83 + ], + "type": "text", + "content": "(c) Vision Encoder" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 291, + 84, + 410, + 127 + ], + "lines": [ + { + "bbox": [ + 291, + 84, + 410, + 127 + ], + "spans": [ + { + "bbox": [ + 291, + 84, + 410, + 127 + ], + "type": "table", + "html": "
Vision ModelAPPROVECOIN
R5084.855.2
ViT-B/3285.557.5
ViT-B/1683.857.8
", + "image_path": "360771c02d65046a0d3bd00d5c78ece9f4cbcbc634c64ad063a88fd4b9561ffc.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 416, + 84, + 529, + 117 + ], + "blocks": [ + { + "bbox": [ + 443, + 74, + 502, + 83 + ], + "lines": [ + { + "bbox": [ + 443, + 74, + 502, + 83 + ], + "spans": [ + { + "bbox": [ + 443, + 74, + 502, + 83 + ], + "type": "text", + "content": "(d) Text Encoder" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 416, + 84, + 529, + 117 + ], + "lines": [ + { + "bbox": [ + 416, + 84, + 529, + 117 + ], + "spans": [ + { + "bbox": [ + 416, + 84, + 529, + 117 + ], + "type": "table", + "html": "
Text ModelAPPROVECOIN
DistilBert-B85.557.5
t5-S87.357.9
", + "image_path": "d8ac3947169d80e31ccfbdf04d2c0afe6268cf07ea8cbbf9f4ef77952a36b98a.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 79, + 167, + 187, + 219 + ], + "blocks": [ + { + "bbox": [ + 79, + 138, + 513, + 148 + ], + "lines": [ + { + "bbox": [ + 79, + 138, + 513, + 148 + ], + "spans": [ + { + "bbox": [ + 79, + 138, + 513, + 148 + ], + "type": "text", + "content": "Table 5. Ablation studies. R@80% Precision for APPROVE and Top-1 Accuracy for COIN. Default setup is underlined." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 97, + 156, + 168, + 165 + ], + "lines": [ + { + "bbox": [ + 97, + 156, + 168, + 165 + ], + "spans": [ + { + "bbox": [ + 97, + 156, + 168, + 165 + ], + "type": "text", + "content": "(a) Noisy modalities." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 79, + 167, + 187, + 219 + ], + "lines": [ + { + "bbox": [ + 79, + 167, + 187, + 219 + ], + "spans": [ + { + "bbox": [ + 79, + 167, + 187, + 219 + ], + "type": "table", + "html": "
% missingAPPROVECOIN
0%85.557.5
10% V80.153.3
10% T75.857.2
30% T68.942.8
", + "image_path": "4004cb77dcd211dc99d5efe025e7c0274bcfb0e4cd2ef25ac4df780bc2566bc7.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 237, + 167, + 352, + 209 + ], + "blocks": [ + { + "bbox": [ + 251, + 156, + 338, + 165 + ], + "lines": [ + { + "bbox": [ + 251, + 156, + 338, + 165 + ], + "spans": [ + { + "bbox": [ + 251, + 156, + 338, + 165 + ], + "type": "text", + "content": "(b) Run-to-Run Variance" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 237, + 167, + 352, + 209 + ], + "lines": [ + { + "bbox": [ + 237, + 167, + 352, + 209 + ], + "spans": [ + { + "bbox": [ + 237, + 167, + 352, + 209 + ], + "type": "table", + "html": "
MethodAPPROVECOIN
BCE76.3 ± 0.754.9 ± 0.6
Focal82.2 ± 0.556.1 ± 0.3
Ours85.5 ± 0.557.5 ± 0.8
", + "image_path": "e092c12ed9ddde6c20ad00d26ea9add9f51b2d89f54d966dd400fc5bfa83cc80.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_body" + } + ], + "index": 12 + }, + { + "type": "table", + "bbox": [ + 397, + 167, + 518, + 214 + ], + "blocks": [ + { + "bbox": [ + 429, + 156, + 486, + 165 + ], + "lines": [ + { + "bbox": [ + 429, + 156, + 486, + 165 + ], + "spans": [ + { + "bbox": [ + 429, + 156, + 486, + 165 + ], + "type": "text", + "content": "(c) Initialization" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 397, + 167, + 518, + 214 + ], + "lines": [ + { + "bbox": [ + 397, + 167, + 518, + 214 + ], + "spans": [ + { + "bbox": [ + 397, + 167, + 518, + 214 + ], + "type": "table", + "html": "
MethodAPPROVECOIN
ImageNet & Wiki-en+TBC85.557.5
CLIP86.763.5
", + "image_path": "1fba4160f622a13f9468404c14142b02de63724245cc05aaa49d265a20be6c82.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "table_body" + } + ], + "index": 14 + }, + { + "bbox": [ + 61, + 222, + 529, + 233 + ], + "lines": [ + { + "bbox": [ + 61, + 222, + 529, + 233 + ], + "spans": [ + { + "bbox": [ + 61, + 222, + 529, + 233 + ], + "type": "text", + "content": "Table 6. Robustness analysis. Our method is robust to partially missing modality and has similar run-to-run variance as baselines." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 241, + 287, + 432 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 241, + 287, + 432 + ], + "spans": [ + { + "bbox": [ + 46, + 241, + 287, + 432 + ], + "type": "text", + "content": "Learning class prototypes: We compare the two strategies where after initializing the class prototypes, we 1) keep them fixed and learn only the multimodal embedding of the samples, and 2) class prototypes and sample embedding are learned iteratively. The initialization can be done either randomly, or with orthogonal constraints. We note that orthogonal initialization performs best in our experiments and iterative adjusting the class prototypes achieves better performance as shown in Table 5 (a). We also consider hierarchical prototypes, for APPROVE, using a 2-level hierarchy where the first level consists of 18 classes, and the second level is the 3 super-classes: math, literacy, and background. The 180 task categories of COIN are organized into 12 domains in the taxonomy provided with the dataset. This hierarchy imposes an additional constraint on learning the embeddings during training." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 46, + 433, + 287, + 480 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 433, + 287, + 480 + ], + "spans": [ + { + "bbox": [ + 46, + 433, + 287, + 480 + ], + "type": "text", + "content": "Fusion Encoder: This evaluates the effect of the number of layers in the fusion encoder on the final performance. As expected, the performance improves with more layers and saturates around 4 layers (Table 5 (b))." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 46, + 481, + 287, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 481, + 287, + 552 + ], + "spans": [ + { + "bbox": [ + 46, + 481, + 287, + 552 + ], + "type": "text", + "content": "Vision Encoder: This evaluates the effect of the image encoder on the final performance. We consider ResNet [19] and ViT variants [16]. We notice that the ViT-B/16-384 encoder works well for the larger COIN dataset, whereas the ViT-B/32-224 encoder works best for APPROVE (Table 5 (c))." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 46, + 553, + 287, + 637 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 553, + 287, + 637 + ], + "spans": [ + { + "bbox": [ + 46, + 553, + 287, + 637 + ], + "type": "text", + "content": "Text Encoder: This evaluates the effect of the text encoder on the final performance. We test DistilBERT and T5 backbones for the text encoder. BERT is trained to predict masked spans of text. T5's unsupervised objective is similar, however, it trains on predicting the entire sequence instead of just the masked spans. GPT2 takes an auto-regressive approach to language modeling (Table 5 (d))." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 47, + 640, + 162, + 652 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 640, + 162, + 652 + ], + "spans": [ + { + "bbox": [ + 47, + 640, + 162, + 652 + ], + "type": "text", + "content": "5.2. Robustness analysis" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 47, + 654, + 269, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 654, + 269, + 665 + ], + "spans": [ + { + "bbox": [ + 47, + 654, + 269, + 665 + ], + "type": "text", + "content": "We evaluate the robustness of our approach as follows:" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 46, + 665, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 287, + 713 + ], + "type": "text", + "content": "Noisy modality: YouTube videos may have noisy modalities where some of the video frames are missing or ASR transcription is noisy. We show that our approach is robust against the cases where a percentage of video frames or text" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 306, + 241, + 468, + 253 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 241, + 468, + 253 + ], + "spans": [ + { + "bbox": [ + 306, + 241, + 468, + 253 + ], + "type": "text", + "content": "words are missing as shown in Tab 6(a)." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 305, + 254, + 545, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 254, + 545, + 289 + ], + "spans": [ + { + "bbox": [ + 305, + 254, + 545, + 289 + ], + "type": "text", + "content": "Run-to-Run variance: The low variance across runs (Tab 6(b)) indicates that our approach is not sensitive to random initialization of the class prototypes." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 304, + 290, + 546, + 398 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 290, + 546, + 398 + ], + "spans": [ + { + "bbox": [ + 304, + 290, + 546, + 398 + ], + "type": "text", + "content": "Initializing the encoders: We consider the ImageNet pretraining for the image encoder, while English Wikipedia + Toronto Book Corpus is used to pre-train the text encoder. We provide results where the backbones are initialized with CLIP [42], which provides a more aligned vision-text representation. As expected, the results are better with the CLIP initialization (Tab 6(c)). The improvements are more significant on COIN than APPROVE as CLIP models may not be exposed to educational videos." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 306, + 405, + 378, + 417 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 405, + 378, + 417 + ], + "spans": [ + { + "bbox": [ + 306, + 405, + 378, + 417 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 304, + 420, + 545, + 599 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 420, + 545, + 599 + ], + "spans": [ + { + "bbox": [ + 304, + 420, + 545, + 599 + ], + "type": "text", + "content": "We have proposed an approach for detecting educational content in online videos. The problem is formulated as a fine-grained multilabel video classification task and we have considered class-prototypes based contrastive learning to address this. We have employed a multimodal transformer network to fuse visual and audio cues. This is crucial for comprehending educational content as both visual and audio cues are to be aligned to ensure effective comprehension. Our approach is shown to be effective in distinguishing fine-grained educational content with high interclass similarity. We have introduced APPROVE - a dataset with 193 hours of expert-annotated educational videos. Beyond APPROVE, we have evaluated our approach on COIN and YouTube-8M datasets where our approach outperforms the competitive baselines." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 306, + 603, + 408, + 615 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 603, + 408, + 615 + ], + "spans": [ + { + "bbox": [ + 306, + 603, + 408, + 615 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 305, + 616, + 545, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 616, + 545, + 647 + ], + "spans": [ + { + "bbox": [ + 305, + 616, + 545, + 647 + ], + "type": "text", + "content": "This work is supported by NSF grant # 2139219 and SRI R&D award. Rohit Gupta is supported by ARO grant W911NF-19-1-0356." + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 304, + 648, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 648, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 648, + 545, + 713 + ], + "type": "text", + "content": "Views and conclusions contained herein are those of the authors and should not be interpreted as representing the official policies or endorsements, either expressed or implied, of NSF, ARO, IARPA, DOI/IBC, or the U.S. Government. The U.S. Government is authorized to reproduce and distribute reprints for Governmental purposes notwithstanding any copyright annotation thereon." + } + ] + } + ], + "index": 30 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19930" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 714 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 146 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 146 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 146 + ], + "type": "text", + "content": "[1] Sami Abu-El-Haija, Nisarg Kothari, Joonseok Lee, Paul Natsev, George Toderici, Balakrishnan Varadarajan, and Sudheendra Vijayanarasimhan. YouTube-8M: A large-scale video classification benchmark. arXiv preprint arXiv:1609.08675, 2016. 1, 2, 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 148, + 288, + 203 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 148, + 288, + 203 + ], + "spans": [ + { + "bbox": [ + 53, + 148, + 288, + 203 + ], + "type": "text", + "content": "[2] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katie Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. arXiv preprint arXiv:2204.14198, 2022. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 205, + 288, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 205, + 288, + 226 + ], + "spans": [ + { + "bbox": [ + 53, + 205, + 288, + 226 + ], + "type": "text", + "content": "[3] National Governors Association et al. Common core state standards. Washington, DC, 2010. 1, 2, 3, 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 228, + 288, + 261 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 228, + 288, + 261 + ], + "spans": [ + { + "bbox": [ + 53, + 228, + 288, + 261 + ], + "type": "text", + "content": "[4] Brooke Auxier, Monica Anderson Andrew Perrin, and Erica Turner. Parenting children in the age of screens. Technical report, Pew Research Center, 2020. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 262, + 287, + 328 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 262, + 287, + 328 + ], + "spans": [ + { + "bbox": [ + 53, + 262, + 287, + 328 + ], + "type": "text", + "content": "[5] Philip Bachman, R Devon Hjelm, and William Buchwalter. Learning representations by maximizing mutual information across views. In H. Wallach, H. Larochelle, A. Beygelzimer, F. d Alché-Buc, E. Fox, and R. Garnett, editors, Advances in Neural Information Processing Systems, volume 32. Curran Associates, Inc., 2019. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 330, + 287, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 330, + 287, + 373 + ], + "spans": [ + { + "bbox": [ + 53, + 330, + 287, + 373 + ], + "type": "text", + "content": "[6] Max Bain, Arsha Nagrani, Gül Varol, and Andrew Zisserman. Frozen in time: A joint video and image encoder for end-to-end retrieval. In IEEE International Conference on Computer Vision, 2021. 2, 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 375, + 287, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 375, + 287, + 418 + ], + "spans": [ + { + "bbox": [ + 53, + 375, + 287, + 418 + ], + "type": "text", + "content": "[7] Johanna Burkhardt and Wolfgang Lenhard. A meta-analysis on the longitudinal, age-dependent effects of violent video games on aggression. Media Psychology, 25(3):499-512, 2022. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 420, + 287, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 420, + 287, + 464 + ], + "spans": [ + { + "bbox": [ + 53, + 420, + 287, + 464 + ], + "type": "text", + "content": "[8] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International conference on machine learning, pages 1597-1607. PMLR, 2020. 2, 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 466, + 287, + 540 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 466, + 287, + 540 + ], + "spans": [ + { + "bbox": [ + 53, + 466, + 287, + 540 + ], + "type": "text", + "content": "[9] Ting Chen, Simon Kornblith, Kevin Swersky, Mohammad Norouzi, and Geoffrey E Hinton. Big self-supervised models are strong semi-supervised learners. In H. Larochelle, M. Ranzato, R. Hadsell, M. F. Balcan, and H. Lin, editors, Advances in Neural Information Processing Systems, volume 33, pages 22243-22255. Curran Associates, Inc., 2020. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 544, + 286, + 576 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 544, + 286, + 576 + ], + "spans": [ + { + "bbox": [ + 48, + 544, + 286, + 576 + ], + "type": "text", + "content": "[10] Xinlei Chen, Haoqi Fan, Ross Girshick, and Kaiming He. Improved baselines with momentum contrastive learning, 2020. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 578, + 286, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 578, + 286, + 622 + ], + "spans": [ + { + "bbox": [ + 48, + 578, + 286, + 622 + ], + "type": "text", + "content": "[11] Xinlei Chen, Hao Fang, Tsung-Yi Lin, Ramakrishna Vedantam, Saurabh Gupta, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco captions: Data collection and evaluation server. arXiv preprint arXiv:1504.00325, 2015. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 624, + 286, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 624, + 286, + 656 + ], + "spans": [ + { + "bbox": [ + 48, + 624, + 286, + 656 + ], + "type": "text", + "content": "[12] Xinlei Chen, Saining Xie, and Kaiming He. An empirical study of training self-supervised vision transformers, 2021. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 658, + 288, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 288, + 714 + ], + "type": "text", + "content": "[13] Ekin D Cubuk, Barret Zoph, Jonathon Shlens, and Quoc V Le. Randaugment: Practical automated data augmentation with a reduced search space. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops, pages 702-703, 2020. 6" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "text", + "content": "[14] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018. 5" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 118, + 545, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 545, + 161 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 545, + 161 + ], + "type": "text", + "content": "[15] Ali Diba, Mohsen Fayyaz, Vivek Sharma, Manohar Paluri, Jürgen Gall, Rainer Stiefelhagen, and Luc Van Gool. Large scale holistic video understanding. In European Conference on Computer Vision, pages 593-610. Springer, 2020. 4" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 162, + 545, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 162, + 545, + 217 + ], + "spans": [ + { + "bbox": [ + 308, + 162, + 545, + 217 + ], + "type": "text", + "content": "[16] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. ICLR, 2021. 5, 8" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 217, + 545, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 217, + 545, + 293 + ], + "spans": [ + { + "bbox": [ + 308, + 217, + 545, + 293 + ], + "type": "text", + "content": "[17] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations, 2021. 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 294, + 545, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 294, + 545, + 348 + ], + "spans": [ + { + "bbox": [ + 308, + 294, + 545, + 348 + ], + "type": "text", + "content": "[18] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9729-9738, 2020. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 350, + 545, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 350, + 545, + 392 + ], + "spans": [ + { + "bbox": [ + 308, + 350, + 545, + 392 + ], + "type": "text", + "content": "[19] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 6, 8" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 394, + 545, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 394, + 545, + 437 + ], + "spans": [ + { + "bbox": [ + 308, + 394, + 545, + 437 + ], + "type": "text", + "content": "[20] Lowry Hemphill and Terrence Tivnan. The importance of early vocabulary for literacy achievement in high-poverty schools. Journal of Education for Students Placed at Risk, 13(4):426-451, 2008. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 437, + 545, + 481 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 437, + 545, + 481 + ], + "spans": [ + { + "bbox": [ + 308, + 437, + 545, + 481 + ], + "type": "text", + "content": "[21] Romain Hennequin, Anis Khlif, Felix Voituret, and Manuel Moussallam. Spleeter: a fast and efficient music source separation tool with pre-trained models. Journal of Open Source Software, 5(50):2154, 2020. Deezer Research. 5" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 482, + 545, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 482, + 545, + 514 + ], + "spans": [ + { + "bbox": [ + 308, + 482, + 545, + 514 + ], + "type": "text", + "content": "[22] Lisa B Hurwitz. Getting a read on ready to learn media: A meta-analytic review of effects on literacy. *Child Development*, 90(5):1754–1771, 2019. 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 515, + 545, + 558 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 515, + 545, + 558 + ], + "spans": [ + { + "bbox": [ + 308, + 515, + 545, + 558 + ], + "type": "text", + "content": "[23] Lisa B Hurwitz and Kelly L Schmitt. Raising readers with ready to learn: A six-year follow-up to an early educational computer game intervention. Computers in Human Behavior, 104:106176, 2020. 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 559, + 545, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 559, + 545, + 613 + ], + "spans": [ + { + "bbox": [ + 308, + 559, + 545, + 613 + ], + "type": "text", + "content": "[24] Andrew Jaegle, Felix Gimeno, Andy Brock, Oriol Vinyals, Andrew Zisserman, and Joao Carreira. Perceiver: General perception with iterative attention. In International conference on machine learning, pages 4651-4664. PMLR, 2021. 6" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 614, + 545, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 614, + 545, + 658 + ], + "spans": [ + { + "bbox": [ + 308, + 614, + 545, + 658 + ], + "type": "text", + "content": "[25] Nancy C Jordan, David Kaplan, Chaitanya Ramineni, and Maria N Locuniak. Early math matters: kindergarten number competence and later mathematics outcomes. Developmental psychology, 45(3):850, 2009. 1" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 658, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 658, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 308, + 658, + 545, + 712 + ], + "type": "text", + "content": "[26] Prannay Khosla, Piotr Teterwak, Chen Wang, Aaron Sarna, Yonglong Tian, Phillip Isola, Aaron Maschinot, Ce Liu, and Dilip Krishnan. Supervised contrastive learning. Advances in Neural Information Processing Systems, 33:18661-18673, 2020. 2, 4, 5, 6, 7" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19931" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "text", + "content": "[27] Hildegard Kuehne, Hueihan Jhuang, Estíbaliz Garrote, Tomaso Poggio, and Thomas Serre. Hmdb: a large video database for human motion recognition. In 2011 International conference on computer vision, pages 2556-2563. IEEE, 2011. 4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 128, + 287, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 128, + 287, + 171 + ], + "spans": [ + { + "bbox": [ + 48, + 128, + 287, + 171 + ], + "type": "text", + "content": "[28] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In ICML, 2022. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 172, + 287, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 172, + 287, + 216 + ], + "spans": [ + { + "bbox": [ + 48, + 172, + 287, + 216 + ], + "type": "text", + "content": "[29] Junnan Li, Ramprasaath R. Selvaraju, Akhilesh Deepak Gotmare, Shafiq Joty, Caiming Xiong, and Steven Hoi. Align before fuse: Vision and language representation learning with momentum distillation. In NeurIPS, 2021. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 217, + 287, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 217, + 287, + 259 + ], + "spans": [ + { + "bbox": [ + 48, + 217, + 287, + 259 + ], + "type": "text", + "content": "[30] Junnan Li, Pan Zhou, Caiming Xiong, and Steven Hoi. Prototypical contrastive learning of unsupervised representations. In International Conference on Learning Representations, 2021. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 261, + 287, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 261, + 287, + 293 + ], + "spans": [ + { + "bbox": [ + 48, + 261, + 287, + 293 + ], + "type": "text", + "content": "[31] Junnan Li, Pan Zhou, Caiming Xiong, and Steven C.H. Hoi. Prototypical contrastive learning of unsupervised representations. In ICLR, 2021. 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 293, + 287, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 293, + 287, + 348 + ], + "spans": [ + { + "bbox": [ + 48, + 293, + 287, + 348 + ], + "type": "text", + "content": "[32] Shikun Li, Xiaobo Xia, Shiming Ge, and Tongliang Liu. Selective-supervised contrastive learning with noisy labels. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 316-325, June 2022. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 349, + 287, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 349, + 287, + 403 + ], + "spans": [ + { + "bbox": [ + 48, + 349, + 287, + 403 + ], + "type": "text", + "content": "[33] Xiujun Li, Xi Yin, Chunyuan Li, Xiaowei Hu, Pengchuan Zhang, Lei Zhang, Lijuan Wang, Houdong Hu, Li Dong, Furu Wei, Yejin Choi, and Jianfeng Gao. Oscar: Object-semantics aligned pre-training for vision-language tasks. ECCV 2020, 2020. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 404, + 287, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 404, + 287, + 459 + ], + "spans": [ + { + "bbox": [ + 48, + 404, + 287, + 459 + ], + "type": "text", + "content": "[34] Yixuan Li, Lei Chen, Runyu He, Zhenzhi Wang, Gangshan Wu, and Limin Wang. Multisports: A multi-person video dataset of spatio-temporally localized sports actions. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13536-13545, 2021. 4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 459, + 287, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 459, + 287, + 503 + ], + "spans": [ + { + "bbox": [ + 48, + 459, + 287, + 503 + ], + "type": "text", + "content": "[35] Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dólár. Focal loss for dense object detection. In Proceedings of the IEEE international conference on computer vision, pages 2980-2988, 2017. 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 503, + 287, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 503, + 287, + 536 + ], + "spans": [ + { + "bbox": [ + 48, + 503, + 287, + 536 + ], + "type": "text", + "content": "[36] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In International Conference on Learning Representations, 2019. 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 536, + 287, + 569 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 536, + 287, + 569 + ], + "spans": [ + { + "bbox": [ + 48, + 536, + 287, + 569 + ], + "type": "text", + "content": "[37] Ziyang Luo, Yadong Xi, Rongsheng Zhang, and Jing Ma. A frustratingly simple approach for end-to-end image captioning, 2022. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 570, + 287, + 592 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 570, + 287, + 592 + ], + "spans": [ + { + "bbox": [ + 48, + 570, + 287, + 592 + ], + "type": "text", + "content": "[38] Edward Ma. Nlp augmentation. https://github.com/makcedward/nlpaug, 2019.6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 593, + 287, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 593, + 287, + 624 + ], + "spans": [ + { + "bbox": [ + 48, + 593, + 287, + 624 + ], + "type": "text", + "content": "[39] Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748, 2018. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 625, + 287, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 625, + 287, + 668 + ], + "spans": [ + { + "bbox": [ + 48, + 625, + 287, + 668 + ], + "type": "text", + "content": "[40] Andrew Porter, Jennifer McMaken, Jun Hwang, and Rui Yang. Common core standards: The new us intended curriculum. Educational researcher, 40(3):103-116, 2011. 1, 2, 3, 5" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 669, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 669, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 669, + 287, + 713 + ], + "type": "text", + "content": "[41] J. S. Radesky, A. Schaller, S. L. Yeo, H. M. Weeks, and M. B. Robb. Young kids and youtube: How ads, toys, and games dominate viewing. common sense media. Technical report, Common Sense Media, 2020. 2, 3" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 72, + 546, + 713 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 307, + 72, + 546, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 72, + 546, + 139 + ], + "spans": [ + { + "bbox": [ + 307, + 72, + 546, + 139 + ], + "type": "text", + "content": "[42] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 2, 8" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 141, + 546, + 186 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 141, + 546, + 186 + ], + "spans": [ + { + "bbox": [ + 307, + 141, + 546, + 186 + ], + "type": "text", + "content": "[43] Alec Radford, Jong Wook Kim, Tao Xu, Greg Brockman, Christine McLeavey, and Ilya Sutskever. Robust speech recognition via large-scale weak supervision. Technical report, Tech. Rep., Technical report, OpenAI, 2022. 5, 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 186, + 546, + 242 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 186, + 546, + 242 + ], + "spans": [ + { + "bbox": [ + 307, + 186, + 546, + 242 + ], + "type": "text", + "content": "[44] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, Peter J Liu, et al. Exploring the limits of transfer learning with a unified text-to-text transformer. J. Mach. Learn. Res., 21(140):1-67, 2020. 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 244, + 546, + 267 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 244, + 546, + 267 + ], + "spans": [ + { + "bbox": [ + 307, + 244, + 546, + 267 + ], + "type": "text", + "content": "[45] Victoria Rideout and Michael B. Robb. The common sense census: Media use by tweens and teens, Oct 2019. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 269, + 546, + 302 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 269, + 546, + 302 + ], + "spans": [ + { + "bbox": [ + 307, + 269, + 546, + 302 + ], + "type": "text", + "content": "[46] V Rideout and M. B. Robb. The common sense census: Media use by tweens and teens. common sense media. Technical report, Common Sense Media, 2019. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 304, + 546, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 304, + 546, + 360 + ], + "spans": [ + { + "bbox": [ + 307, + 304, + 546, + 360 + ], + "type": "text", + "content": "[47] Tal Ridnik, Emanuel Ben-Baruch, Nadav Zamir, Asaf Noy, Itamar Friedman, Matan Protter, and Lihi Zelnik-Manor. Asymmetric loss for multi-label classification. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 82-91, 2021. 7" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 361, + 546, + 405 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 361, + 546, + 405 + ], + "spans": [ + { + "bbox": [ + 307, + 361, + 546, + 405 + ], + "type": "text", + "content": "[48] Victor Sanh, Lysandre Debut, Julien Chaumont, and Thomas Wolf. Distilbert, a distilled version of bert: smaller, faster, cheaper and lighter. arXiv preprint arXiv:1910.01108, 2019.6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 407, + 546, + 461 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 407, + 546, + 461 + ], + "spans": [ + { + "bbox": [ + 307, + 407, + 546, + 461 + ], + "type": "text", + "content": "[49] Dian Shao, Yue Zhao, Bo Dai, and Dahua Lin. Finegym: A hierarchical video dataset for fine-grained action understanding. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2616-2625, 2020. 4" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 464, + 546, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 464, + 546, + 542 + ], + "spans": [ + { + "bbox": [ + 307, + 464, + 546, + 542 + ], + "type": "text", + "content": "[50] Piyush Sharma, Nan Ding, Sebastian Goodman, and Radu Soricut. Conceptual captions: A cleaned, hypernymed, image alt-text dataset for automatic image captioning. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 2556-2565, Melbourne, Australia, July 2018. Association for Computational Linguistics. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 544, + 546, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 544, + 546, + 610 + ], + "spans": [ + { + "bbox": [ + 307, + 544, + 546, + 610 + ], + "type": "text", + "content": "[51] Nina Shvetsova, Brian Chen, Andrew Rouditchenko, Samuel Thomas, Brian Kingsbury, Rogerio S Feris, David Harwath, James Glass, and Hilde Kuehne. Everything at once-multimodal fusion transformer for video retrieval. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20020-20029, 2022. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 612, + 546, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 612, + 546, + 678 + ], + "spans": [ + { + "bbox": [ + 307, + 612, + 546, + 678 + ], + "type": "text", + "content": "[52] Gunnar A. Sigurdsson, Gül Varol, Xiaolong Wang, Ali Farhadi, Ivan Laptev, and Abhinav Gupta. Hollywood in homes: Crowdsourcing data collection for activity understanding. In Bastian Leibe, Jiri Matas, Nicu Sebe, and Max Welling, editors, Computer Vision – ECCV 2016, pages 510–526, Cham, 2016. Springer International Publishing. 4" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 680, + 546, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 680, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 680, + 546, + 713 + ], + "type": "text", + "content": "[53] Lucas Smaira, João Carreira, Eric Noland, Ellen Clancy, Amy Wu, and Andrew Zisserman. A short note on the kinetics-700-2020 human action dataset, 2020. 1, 4" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "19932" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 369 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 48, + 72, + 286, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 286, + 106 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 286, + 106 + ], + "type": "text", + "content": "[54] Khurram Soomro, Amir Roshan Zamir, and Mubarak Shah. Ucf101: A dataset of 101 human actions classes from videos in the wild. arXiv preprint arXiv:1212.0402, 2012. 1, 4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 106, + 286, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 106, + 286, + 160 + ], + "spans": [ + { + "bbox": [ + 49, + 106, + 286, + 160 + ], + "type": "text", + "content": "[55] Yansong Tang, Dajun Ding, Yongming Rao, Yu Zheng, Danyang Zhang, Lili Zhao, Jiwen Lu, and Jie Zhou. Coin: A large-scale dataset for comprehensive instructional video analysis. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 2, 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 160, + 286, + 193 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 160, + 286, + 193 + ], + "spans": [ + { + "bbox": [ + 49, + 160, + 286, + 193 + ], + "type": "text", + "content": "[56] Grigorios Tsoumakas, Ioannis Katakis, and Ioannis Vla-havas. Mining Multi-label Data, pages 667-685. Springer US, Boston, MA, 2010. 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 194, + 287, + 247 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 194, + 287, + 247 + ], + "spans": [ + { + "bbox": [ + 49, + 194, + 287, + 247 + ], + "type": "text", + "content": "[57] Wenguan Wang, Tianfei Zhou, Fisher Yu, Jifeng Dai, Ender Konukoglu, and Luc Van Gool. Exploring cross-image pixel contrast for semantic segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 7303-7313, 2021. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 49, + 248, + 287, + 313 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 248, + 287, + 313 + ], + "spans": [ + { + "bbox": [ + 49, + 248, + 287, + 313 + ], + "type": "text", + "content": "[58] Jinyu Yang, Jiali Duan, Son Tran, Yi Xu, Sampath Chanda, Liquun Chen, Belinda Zeng, Trishul Chilimbi, and Junzhou Huang. Vision-language pre-training with triple contrastive learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15671-15680, 2022. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 49, + 314, + 287, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 314, + 287, + 369 + ], + "spans": [ + { + "bbox": [ + 49, + 314, + 287, + 369 + ], + "type": "text", + "content": "[59] Peter Young, Alice Lai, Micah Hodosh, and Julia Hockenmaier. From image descriptions to visual denotations: New similarity metrics for semantic inference over event descriptions. Transactions of the Association for Computational Linguistics, 2:67-78, 2014. 3" + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 72, + 545, + 368 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 307, + 72, + 545, + 148 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 72, + 545, + 148 + ], + "spans": [ + { + "bbox": [ + 307, + 72, + 545, + 148 + ], + "type": "text", + "content": "[60] Rowan Zellers, Jiasen Lu, Ximing Lu, Youngjae Yu, Yanpeng Zhao, Mohammadreza Salehi, Aditya Kusupati, Jack Hessel, Ali Farhadi, and Yejin Choi. Merlot reserve: Neural script knowledge through vision and language and sound. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 16375-16387, June 2022. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 307, + 152, + 545, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 152, + 545, + 206 + ], + "spans": [ + { + "bbox": [ + 307, + 152, + 545, + 206 + ], + "type": "text", + "content": "[61] Rowan Zellers, Ximing Lu, Jack Hessel, Youngjae Yu, Jae Sung Park, Jize Cao, Ali Farhadi, and Yejin Choi. Merlot: Multimodal neural script knowledge models. Advances in Neural Information Processing Systems, 34:23634-23651, 2021. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 308, + 209, + 545, + 254 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 209, + 545, + 254 + ], + "spans": [ + { + "bbox": [ + 308, + 209, + 545, + 254 + ], + "type": "text", + "content": "[62] Pengchuan Zhang, Xiujun Li, Xiaowei Hu, Jianwei Yang, Lei Zhang, Lijuan Wang, Yejin Choi, and Jianfeng Gao. Vinvl: Making visual representations matter in vision-language models. CVPR 2021, 2021. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 308, + 256, + 545, + 311 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 256, + 545, + 311 + ], + "spans": [ + { + "bbox": [ + 308, + 256, + 545, + 311 + ], + "type": "text", + "content": "[63] Shu Zhang, Ran Xu, Caiming Xiong, and Chetan Ramaiah. Use all the labels: A hierarchical multi-label contrastive learning framework. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 16660-16669, June 2022. 2, 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 308, + 313, + 545, + 368 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 313, + 545, + 368 + ], + "spans": [ + { + "bbox": [ + 308, + 313, + 545, + 368 + ], + "type": "text", + "content": "[64] Chen Zhu, Xiao Tan, Feng Zhou, Xiao Liu, Kaiyu Yue, Errui Ding, and Yi Ma. Fine-grained video categorization with redundancy reduction attention. In Proceedings of the European Conference on Computer Vision (ECCV), pages 136-152, 2018. 4" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19933" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Class Relationship Embedded Learning for Source-Free Unsupervised Domain Adaptation/a0503f36-a779-42c8-98b9-ccfd34b3ed5c_content_list.json b/2023/Class Relationship Embedded Learning for Source-Free Unsupervised Domain Adaptation/a0503f36-a779-42c8-98b9-ccfd34b3ed5c_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..301e611b3000d551bbbfc8cf34aa3606175ce82f --- /dev/null +++ b/2023/Class Relationship Embedded Learning for Source-Free Unsupervised Domain Adaptation/a0503f36-a779-42c8-98b9-ccfd34b3ed5c_content_list.json @@ -0,0 +1,1623 @@ +[ + { + "type": "text", + "text": "Class Relationship Embedded Learning for Source-Free Unsupervised Domain Adaptation", + "text_level": 1, + "bbox": [ + 228, + 130, + 738, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yixin Zhang $^{1,2}$ Zilei Wang $^{*2}$ Weinan He $^{2}$", + "bbox": [ + 303, + 202, + 666, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Institute of Artificial Intelligence, Hefei Comprehensive National Science Center", + "bbox": [ + 161, + 220, + 808, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ University of Science and Technology of China", + "bbox": [ + 292, + 238, + 678, + 256 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{zhyx12, zlwang}@ustc.edu.cn, hwn2018@mail.ustc.edu.cn", + "bbox": [ + 243, + 258, + 720, + 273 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 308, + 313, + 324 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "This work focuses on a practical knowledge transfer task defined as Source-Free Unsupervised Domain Adaptation (SFUDA), where only a well-trained source model and unlabeled target data are available. To fully utilize source knowledge, we propose to transfer the class relationship, which is domain-invariant but still under-explored in previous works. To this end, we first regard the classifier weights of the source model as class prototypes to compute class relationship, and then propose a novel probability-based similarity between target-domain samples by embedding the source-domain class relationship, resulting in Class Relationship embedded Similarity (CRS). Here the inter-class term is particularly considered in order to more accurately represent the similarity between two samples, in which the source prior of class relationship is utilized by weighting. Finally, we propose to embed CRS into contrastive learning in a unified form. Here both class-aware and instance discrimination contrastive losses are employed, which are complementary to each other. We combine the proposed method with existing representative methods to evaluate its efficacy in multiple SFUDA settings. Extensive experimental results reveal that our method can achieve state-of-the-art performance due to the transfer of domain-invariant class relationship.", + "bbox": [ + 76, + 340, + 473, + 705 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 733, + 209, + 750 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Benefiting from the large amount of labeled training data, deep neural networks have achieved promising results in many computer vision tasks [7, 15, 19, 95]. To reduce the annotation cost, Unsupervised Domain Adaptation (UDA) has been devised by transferring knowledge from a label-rich source domain to a label-scarce target domain. Currently, many UDA methods have been proposed that jointly", + "bbox": [ + 75, + 758, + 468, + 864 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/01f28aac75a59ee55de6ab8029a003a01d3a372089872d6649859ddaecae9d84.jpg", + "image_caption": [ + "Figure 1. Illustration of our proposed method. The upper left represents the target-domain featural distribution by source pre-trained model. The upper right is the feature distribution obtained by our method. The bottom is the process of class relationship embedded contrastive learning. Best viewed in color." + ], + "image_footnote": [], + "bbox": [ + 506, + 305, + 890, + 547 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "learn on the source and target data. But they would be unapplicable for some real-world scenarios involving privacy (e.g., medical images, surveillance videos) because the source-domain data cannot be accessed. Thus, more recent methods [36, 42, 71, 79, 87, 88] focus on Source-Free Unsupervised Domain Adaptation (SFUDA). Under this setting, the labeled source data are not accessible any more when training the target model, but the pre-trained model in the source domain is provided. Then a natural question arises, i.e., what knowledge should we transfer to facilitate the learning of unlabeled target-domain data?", + "bbox": [ + 496, + 657, + 892, + 824 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Some methods [21, 36, 42] assume that the source hypothesis (i.e., classifier [42] or whole model [21, 36]) contains sufficient knowledge for target domain. Then they transfer source knowledge by directly aligning features with the fixed source classifier [42], resorting to historical mod", + "bbox": [ + 496, + 824, + 893, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author", + "bbox": [ + 94, + 875, + 222, + 887 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "$^{1}$ Code is available at https://github.com/zhyx12/CRCo", + "bbox": [ + 94, + 887, + 447, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "7619", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "els [21], or weight regularization [36]. Another line of works [87, 89, 90] assume the source model already forms some semantic structure, and then utilize the local information (i.e., neighborhood) of feature space to enforce the similarity in the output space. Despite these progress, what knowledge to be transferred remains an open question.", + "bbox": [ + 75, + 90, + 470, + 181 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this work, we propose to transfer the class relationship represented by the similarity of classifier weights. Actually, the class relationship is domain-invariant [14], since the same class in the source and target domains essentially represents the same semantic in spite of domain discrepancy. For example, the computers are always more similar with the TV than the scissors. Thus, it is reasonable to guide the target domain learning using class relationship prior. Unlike previous methods that learn class distribution by pseudo labeling [21, 42] and local aggregation of neighborhood predictions [79, 87, 89, 90], here we explicitly model the source-domain class relationship. To be specific, we regard each weight of classifier as the class prototype [90], and then compute a class relationship matrix $A^s$ by cosine similarity, as shown in Figure 1.", + "bbox": [ + 75, + 185, + 468, + 412 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Before explaining how to use this matrix, we illustrate the purpose of representation learning in the target domain using Figure 1, where three classes are adopted for clarity. The top left shows the target-domain feature distribution together with the class weights learned from the source domain. It can be seen that such a situation makes it difficult to perform correct classification. In fact, it is expected that the learned features are discriminative and compact around the corresponding class weight, as shown on the top right. To this end, an intuitive way is to make the relationship of learned class prototypes in the target domain consistent with that in the source domain. However, it has a very limited effect on training the target-domain model as such a prototype-level constraint is too weak for optimization.", + "bbox": [ + 75, + 414, + 468, + 626 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this work, we propose to embed the source-domain class relationship in contrastive learning, which has been shown to be outstanding in representation learning [9, 18]. Here we design a novel sample similarity by taking into account $A^s$ . Specifically, we compute the similarity between two target domain samples in the output space, represented by the prediction probabilities (i.e., $p$ and $p'$ ), as shown in Figure 1. In particular, we consider the inter-class term (i.e. $\\sum_{i \\neq j} p_i p_j'$ ) in addition to the traditional intra-class term (i.e. $\\sum_i p_i p_i'$ ). Note that the sum of these two terms equals one, in which the intra-class term measures the similarity that two samples come from the same class, while the inter-class term measures that from different classes. Considering the relationship between classes, we weight the inter-class term by the non-diagonal elements in $A^s$ . In particular, if the class $i$ is closer to $j$ than $k$ (i.e., $A_{ij}^s > A_{ik}^s$ ), $p_i p_j'$ would be more important in calculating the similarity of $p$ and $p'$ . By the above design, our proposed similarity", + "bbox": [ + 75, + 628, + 468, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "can more accurately express the relationship of two samples based on output space. For example, among three classes 1, 2, 3 in Figure 1, the classes 1 and 2 are closer. Given three samples $x_{1}, x_{2}, x_{3}$ belonging to them with the probabilities [0.9, 0.05, 0.05], [0.05, 0.9, 0.05], and [0.05, 0.05, 0.9]. If we only use the intra-class term, all three samples have the same similarity. But for our designed similarity, $x_{1}$ is closer to $x_{2}$ than $x_{3}$ , which is more reasonable.", + "bbox": [ + 496, + 90, + 890, + 210 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "On the basis of the proposed similarity, we further propose to perform contrastive learning in a unified form. Inspired by recent success in semi-supervised learning [67, 85, 91] and unsupervised learning [9, 18], we propose two types of contrastive losses. As shown in Figure 1, the first one is Class Relationship embedded Class-Aware Contrastive (CR-CACo) loss where the high-confident samples are enforced to be close to the corresponding prototype and away from other prototypes. Due to embedding the prior class relationship, our CR-CACo loss is more robust to label noise caused by domain shift. The second one is Class Relationship embedded Instance Discrimination Contrastive (CR-IDCo) loss where two views of the same sample are encouraged to be close and away from all other samples. Benefited from our designed accurate similarity, the CR-IDCo loss would more effectively learn discriminative features [9, 18, 75]. Actually, these two losses are complementary to each other, and their combination can achieve better performance.", + "bbox": [ + 496, + 212, + 892, + 498 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our contributions are summarized as follows:", + "bbox": [ + 519, + 500, + 821, + 512 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose to explicitly transfer class relationship for SFUDA which is more domain-invariant. And we propose to embed the class relationship into contrastive learning in order to effectively perform representation learning in the target domain.", + "- We propose a novel class relationship embedded similarity which can more accurately express the sample relationship in the output space. Furthermore, we propose two contrastive losses (i.e., CR-CACo and CR-IDCo) exploiting our designed similarity.", + "- We conduct extensive experiments to evaluate the proposed method, and the results validate the effectiveness of our method, which achieves the state-of-the-art performance on multiple SFUDA benchmarks." + ], + "bbox": [ + 519, + 521, + 890, + 750 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 762, + 640, + 777 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Unsupervised Domain Adaptation", + "text_level": 1, + "bbox": [ + 500, + 786, + 797, + 803 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The conventional approaches in UDA [2, 3] are to learn domain-invariant representations, and can be classified into two coarse types. The first one decreases some distribution discrepancy metrics [49, 57, 68, 72, 84]. Another common line is adversarial training [11, 12, 45, 48, 64, 77, 96]. To increase discriminability, more recent DA methods attempt", + "bbox": [ + 496, + 810, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "7620", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "to investigate the target domain structure. Self-training as a typical approach generates target domain pseudo labels [16, 17, 27, 28, 40, 46, 51, 54, 94, 101]. Another category is to construct prototypes [8, 34, 55, 80, 81, 93, 97, 99] or cluster centers [13, 26, 47, 70] across domains and then perform class-wise alignment. Among these methods, most of them rely on source-labeled data during target training, while pseudo label based methods can be directly applied in SFUDA. In our CR-CACo loss, pseudo label is also used and our method can alleviate the negative impact of label noise due to the usage of source class relationship.", + "bbox": [ + 75, + 90, + 472, + 257 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. Source-free Domain Adaptation", + "text_level": 1, + "bbox": [ + 76, + 268, + 359, + 284 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Early SFUDA methods [31, 32, 36] resort to synthesize extra training samples to get compact decision boundaries. Some recent methods [10, 21, 33, 42, 61] use pseudolabel based self-training. SHOT [42] proposes to freeze the source classifier and it clusters target features by IM loss [20] along with clustering based pseudo labeling. HCL [21] adopts feature space Instance Discrimination [78] from current and historical models, together with pseudo label learning conditioned on historical consistency. D-MCD [10] proposes denoised MCD [64] to mitigate the impact of sample selection bias and label noise. CoWA-JMD [33] uses the joint model-data structure score as sample-wise weights, along with weight mixup to exploit more target knowledge. BMD [61] proposes a class-balanced multicentric dynamic prototype strategy to obtain more accurate pseudo labels. Another line of methods (i.e. NRC [87], G-SFDA [89] and AaD [90]) propose neighborhood clustering which enforces prediction consistency between local neighbors.", + "bbox": [ + 75, + 291, + 472, + 579 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In addition to the above methods, others solve SFUDA from different aspects. $A^2$ Net [79] proposes to learn an additional target classifier and uses a contrastive category-wise matching module to learn compact features. DIPE [74] and VMP [25] explore transferability of source model parameters. Sub-sup [29] proposes novel subsidiary pretext tasks to assist domain adaptation. U-SFAN [62] uses the uncertainty quantified by the predictions to guide the target adaptation. Feat-mix [30] proposes image-level and feature-level mixup to enhances the discriminability-transferability tradeoff. Among these methods, few of them explicitly consider the usage of class relationship.", + "bbox": [ + 75, + 579, + 472, + 762 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 76, + 775, + 169, + 791 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Preliminaries", + "text_level": 1, + "bbox": [ + 76, + 800, + 217, + 816 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "For source-free unsupervised domain adaptation (SFUDA), we are given source pre-trained model and an unlabeled target domain with $N_{t}$ samples as $\\mathcal{D}_t = \\{x_i^t\\}_{i = 1}^{N_t}$ . Target domain have same $C$ classes as source domain (for closed-set setting). The goal of SFUDA is to adapt the", + "bbox": [ + 75, + 825, + 470, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "model to target domain without source data. We divide the model into two parts: the feature extractor $f$ , and the classifier $g$ . The output of classifier is denoted as $\\pmb{p} = \\delta(g(f(x))) \\in \\mathbb{R}^C$ where $\\delta$ is the softmax function.", + "bbox": [ + 496, + 90, + 890, + 151 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We adopt a teacher-student framework following [18] as shown in Figure 2. The teacher feature extractor (i.e. $\\tilde{f}$ ) is continuously updated by exponential moving average of the student feature extractor (i.e. $f$ ), and the classifiers (i.e. $g$ , $\\tilde{g}$ ) are shared and frozen as in SHOT [42]. We use sample probabilities as input for contrastive loss, the query probability is obtained from the student model, and the positive and negative probabilities are obtained through the teacher model. During inference, we directly use the student model. In the following, we first introduce class relationship embedded similarity, and then elaborate on how it is used in the proposed contrastive learning process.", + "bbox": [ + 496, + 152, + 893, + 335 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Class Relationship Embedded Similarity", + "text_level": 1, + "bbox": [ + 498, + 347, + 846, + 364 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The inherent semantic and visual relationships among different classes are consistent across domains, regardless of distribution discrepancy. Inspired by this, we propose to explicitly transfer class similarities. Here we regard the normalized classifier weights $\\pmb{W}^{s} = [\\pmb{w}_{1}^{s},\\pmb{w}_{2}^{s},\\dots \\pmb{w}_{C}^{s}]^{T}\\in$ $\\mathbb{R}^{C\\times D}$ as class prototypes where $D$ is the dimension of the last feature. Thus, the class similarity matrix $\\pmb{A}^{s}$ can be obtained by $\\pmb{A}^{s} = \\pmb{W}^{s}(\\pmb{W}^{s})^{T}\\in \\mathbb{R}^{C\\times C}$ .", + "bbox": [ + 496, + 372, + 890, + 493 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "An natural way to utilize $\\mathbf{A}^s$ is to obtain target class similarity matrix $\\mathbf{A}^t$ by target prototypes and then enforce consistency between $\\mathbf{A}^s$ and $\\mathbf{A}^t$ . However, such a prototype-level constraint is coarse-grained and not strong enough to benefit the training. Instead of prototype-level, we propose a probability based sample-level similarity that considers not only the traditional intra-class term, but also the inter-class term. The latter contains probability products of all different classes, thus we can embed the class relationship matrix $\\mathbf{A}^s$ into it as coefficients. Specifically, given the probabilities $\\mathbf{p}, \\mathbf{p}' \\in \\mathbb{R}^{C \\times 1}$ of two samples, our Class Relationship embedded Similarity (CRS) can be presented as follows:", + "bbox": [ + 496, + 494, + 893, + 690 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} s _ {i n t r a} \\left(\\boldsymbol {p}, \\boldsymbol {p} ^ {\\prime}\\right) = \\sum_ {i = 1} ^ {C} p _ {i} \\times p _ {i} ^ {\\prime} \\\\ s _ {i n t e r} \\left(\\boldsymbol {p}, \\boldsymbol {p} ^ {\\prime}\\right) = \\sum_ {i = 1} ^ {C} \\sum_ {j = 1, j \\neq i} ^ {C} A _ {i, j} ^ {s} \\times p _ {i} \\times p _ {j} ^ {\\prime} \\tag {1} \\\\ s _ {c r} (\\pmb {p}, \\pmb {p} ^ {\\prime}) = s _ {i n t r a} (\\pmb {p}, \\pmb {p} ^ {\\prime}) + \\lambda_ {i n t e r} s _ {i n t e r} (\\pmb {p}, \\pmb {p} ^ {\\prime}) \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 527, + 700, + 890, + 810 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $s_{\\text{intra}}$ is the traditional intra-class similarity, and $s_{\\text{inter}}$ is the inter-class similarity, $\\lambda_{\\text{inter}}$ is a trade-off parameter between these two terms.", + "bbox": [ + 496, + 823, + 890, + 867 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Given $\\lambda_{inter} = 1$ , our similarity can be presented in a simpler version as follows:", + "bbox": [ + 498, + 869, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "7621", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/3417d54a80c94ccb5351e10eafe9ce0f715c02ffc9a1b50d8ae263e9f104dde4.jpg", + "image_caption": [ + "Figure 2. The framework of our proposed method. We use a teacher-student architecture to conduct contrastive learning. Class relationship matrix is extracted from fixed classifier and embedded in two types of contrastive loss. Best viewed in color." + ], + "image_footnote": [], + "bbox": [ + 102, + 78, + 867, + 300 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} s _ {c r} (\\boldsymbol {p}, \\boldsymbol {p} ^ {\\prime}) = \\sum_ {i = 1} ^ {C} \\sum_ {j = 1} ^ {C} A _ {i, j} \\times p _ {i} \\times p _ {j} ^ {\\prime} \\\\ = \\sum_ {i = 1} ^ {C} \\sum_ {j = 1} ^ {C} \\left(p _ {i} \\boldsymbol {w} _ {i} ^ {s}\\right) ^ {T} \\left(p _ {j} ^ {\\prime} \\boldsymbol {w} _ {j} ^ {s}\\right) = \\left(\\boldsymbol {p} ^ {T} \\boldsymbol {W} ^ {s}\\right) \\left(\\boldsymbol {p} ^ {\\prime T} \\boldsymbol {W} ^ {s}\\right) ^ {T} \\tag {2} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 91, + 375, + 468, + 474 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\pmb{p}^T\\pmb{W}^s \\in \\mathbb{R}^{1\\times D}$ can be viewed as a new feature which is probability weighted sum of different class weights.", + "bbox": [ + 75, + 476, + 468, + 523 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Class Relationship Embedded Contrastive Loss", + "text_level": 1, + "bbox": [ + 76, + 532, + 468, + 550 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Inspired by the success of contrastive learning in different tasks [9, 18, 21, 26, 85], we propose to use our CRS for contrastive loss. Following previous work [9, 18], the contrastive loss can be presented as follows:", + "bbox": [ + 75, + 556, + 468, + 617 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} h (\\pmb {p}, \\pmb {p} ^ {\\prime}) = \\exp (s _ {c r} (\\pmb {p}, \\pmb {p} ^ {\\prime}) / T _ {c o}), \\\\ \\ell_ {c o} = - \\log \\frac {h \\left(\\boldsymbol {p} , \\boldsymbol {p} _ {+}\\right)}{h \\left(\\boldsymbol {p} , \\boldsymbol {p} _ {+}\\right) + \\sum_ {\\boldsymbol {p} _ {-} \\in M} h \\left(\\boldsymbol {p} , \\boldsymbol {p} _ {-}\\right)}, \\tag {3} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 135, + 622, + 468, + 686 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $T_{co}$ is a temperature parameter, and $h$ denotes the exponent of scaled similarity. $M$ is the memory bank [18] storing probabilities of negative samples. Since our CRS is built on probability, here we use probabilities as inputs instead of original features [18] or projected features [9].", + "bbox": [ + 75, + 693, + 468, + 768 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "With this unified form, we further instantiate it to two losses: class-aware contrastive loss and instance discrimination contrastive loss. The framework of our proposed method is shown in Figure 2.", + "bbox": [ + 75, + 768, + 468, + 830 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Class-Aware Contrastive Loss To achieve class-aware learning, it is important to select proper samples and obtain more accurate pseudo labels for them. Inspired by previous efforts in SSL [67, 85] and UDA [41, 93], we consider", + "bbox": [ + 75, + 839, + 468, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "constructing positive pairs from strongly and weakly augmented views of the same image. Specifically, given probabilities $p_s$ and $\\tilde{p}_w$ (tilde superscript denotes output by the teacher model) of strongly and weakly augmented views, we first compare $Max(\\tilde{p}_w)$ with a predefined threshold $\\tau$ to select high-confident samples. Due to domain shift, instead of directly using $argmax(p_w)$ as pseudo label, we adopt the clustering-based pseudo-labeling method in SHOT [42]. Based on pseudo label, we can get the positive samples $p^*$ with the one-hot form. For the negative samples $p_-$ , we use one-hot probabilities of all classes except the pseudo label class. Therefore, both positive and negative samples (i.e. $p^*$ and $p_-$ ) are generated from the pseudo label of weak target sample through teacher model. Finally, the Class Relationship embedded Class-Aware Contrastive (CR-CACo) loss is presented as", + "bbox": [ + 496, + 349, + 893, + 592 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\ell_ {c r \\_ c a c o} = - \\log \\frac {h \\left(\\boldsymbol {p} _ {s} , \\boldsymbol {p} ^ {*}\\right)}{h \\left(\\boldsymbol {p} _ {s} , \\boldsymbol {p} ^ {*}\\right) + \\sum_ {\\boldsymbol {p} _ {-} \\in M _ {c a c o}} h \\left(\\boldsymbol {p} _ {s} , \\boldsymbol {p} _ {-}\\right)}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 509, + 603, + 890, + 648 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $h$ follows the definition in Equation (3), $M_{caco}$ stores negative samples and is different for each sample according to its pseudo label.", + "bbox": [ + 496, + 657, + 890, + 703 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Regardless of pseudo label generation, using only $s_{intra}$ and replacing $p_s$ with features and $p^*, p_-$ with classifier weights, the above Equation (4) is actually the FixMatch loss [67]. Although FixMatch can improve the performance, our CR-CACo loss can greatly outperform it.", + "bbox": [ + 496, + 704, + 890, + 780 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our CR-CACo loss is more robust to label noise due to the class relationship embedded inter-class similarity. We compare our CRS with the traditional semi-supervised FixMatch loss. The results are shown in Figure 3. It can be seen that the simple IM loss [42] has lower entropy since it has an entropy minimization term. When FixMatch loss is used, the entropy will increase since the class relationship is learned through high-confidence target samples. Our CRS", + "bbox": [ + 496, + 780, + 892, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "7622", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/4f671f0b27492789c3e7ac64a9867ab38a621a4cf81ed33177ba8713d682b514.jpg", + "image_caption": [ + "Figure 3. Entropy density plots under SUDA setting Office-Home $\\mathrm{Rw} \\rightarrow \\mathrm{Cl}$ . Our class relationship embedded similarity $s_{cr}$ has higher entropy but better accuracy. Best viewed in color." + ], + "image_footnote": [], + "bbox": [ + 89, + 85, + 457, + 275 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "has larger entropy since we explicitly consider the class relationship and can alleviate the negative impact of label noise.", + "bbox": [ + 75, + 337, + 470, + 383 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Instance Discrimination Contrastive Loss The instance discrimination contrastive (IDCo) loss has shown remarkable success in self-supervised learning [9, 18]. It can learn discriminative features by regrading the two views of the same sample as positive pairs and all other samples as negative samples. The traditional IDCo loss uses features as input, but it only makes the feature discriminative and cannot guarantee that the learned feature will be close to corresponding prototypes (i.e. classifier weights). Conducting IDCo loss in the output space (i.e. computing similarity using $s_{\\text{intra}}$ ) can alleviate this problem [35], but the more informative relationship between samples in the original feature space can not be fully expressed. This can be addressed by our CRS since we embed the source prior class similarity and can reflect more accurate relationship of samples.", + "bbox": [ + 75, + 402, + 468, + 628 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Here we use the probabilities $\\pmb{p}_{s1}$ and $\\tilde{\\pmb{p}}_{s2}'$ of two strong augmented views from the same image as positive pairs, and probabilities of all other samples in the same batch as negative samples. Finally, the Class Relationship embedded Instance Discrimination Contrastive (CR-IDCo) loss is presented as follows:", + "bbox": [ + 75, + 630, + 468, + 720 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\ell_ {c r. i d c o} = - \\log \\frac {h \\left(\\boldsymbol {p} _ {s 1} , \\tilde {\\boldsymbol {p}} _ {s 2} ^ {\\prime}\\right)}{h \\left(\\boldsymbol {p} _ {s 1} , \\tilde {\\boldsymbol {p}} _ {s 2} ^ {\\prime}\\right) + \\sum_ {\\tilde {\\boldsymbol {p}} _ {-} \\in M _ {i d c o}} h \\left(\\boldsymbol {p} _ {s 1} , \\tilde {\\boldsymbol {p}} _ {-}\\right)}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 86, + 733, + 467, + 790 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $h$ follows the definition in Equation (3), $M_{idco}$ contains the probabilities of other samples within the same batch.", + "bbox": [ + 75, + 791, + 468, + 835 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Overall Training Objective", + "text_level": 1, + "bbox": [ + 76, + 845, + 320, + 863 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The proposed two losses are complementary to each other and can be trained end-to-end. Since our CR-CACo", + "bbox": [ + 75, + 869, + 470, + 898 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "uses a high prob-threshold (i.e. 0.95) to select confident samples, it should be combined with other methods that can generate confident predictions. Otherwise, it will be dominated by noise labels in the early training stage. Although it performs normally when combined with our CR-IDCo loss, both of them focus on transferring source class relationship, and other methods that encourage confident predictions are still needed and complementary to ours. As a result, we combine our method with existing representative methods (i.e. SHOT-IM [42], AaD [90]) to validate the effectiveness. As a result, the overall loss is presented as:", + "bbox": [ + 496, + 90, + 893, + 258 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\ell_ {\\text {o u r s}} = \\ell_ {\\text {b a s e l i n e}} + \\lambda_ {\\text {c a c o}} \\ell_ {\\text {c r} - \\text {c a c o}} + \\lambda_ {\\text {i d c o}} \\ell_ {\\text {c r} - \\text {i d c o}} \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 524, + 270, + 890, + 287 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\lambda_{caco},\\lambda_{idco}$ are the trade-off hyperparameters. $l_{baseline}$ is the loss of baseline (i.e., IM or AaD). It contains two terms: discriminability and diversity. The former produces high-confident outputs, and the latter avoids collapse of predictions. IM uses entropy of single sample and average of all samples, and AaD uses neighbor information.", + "bbox": [ + 496, + 299, + 893, + 390 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 498, + 402, + 633, + 421 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Experimental Setup", + "text_level": 1, + "bbox": [ + 500, + 429, + 689, + 446 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Datasets. We evaluate the effectiveness of our approach on four standard DA benchmarks. Office-31 [63] benchmark consists of three domains in office environments: Amazon (A), DSLR (D), and Webcam (W), each with 31 categories. Office-Home [73] is a more challenging dataset. It comprises of images of commonplace objects divided into four domains: Artistic (Ar), Clipart (Cl), Product (Pr), and Real-World (Rw), each with 65 classes. VisDA [58] is a large-scale dataset for synthetic-to-real domain adaptation. The source domain has 152,397 synthetic images, while the target domain has 55,388 real-world images. DomainNet [57] is the most challenging dataset involving 6 domains: Clipart (C), Real (R), Infograph (I), Painting (P), Sketch (S) and Quickdraw (Q) with 345 classes each.", + "bbox": [ + 496, + 452, + 893, + 664 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Implementation details. To ensure fair comparison with related methods, we adopt the backbone of a ResNet-50 [19] for Office-31, Office-Home, and DomainNet, and a ResNet-101 for VisDA. We use the same network architecture as SHOT [42] where the final part of the network is changed to suit the SFUDA task. For the classifier, we find that fixing it or not have little impact on the final performance, and we choose to fix it in all experiments. We adopt SGD with momentum 0.9 and batch size 64 for all datasets where each image has three views (i.e. one weak and two strong augmentations).", + "bbox": [ + 496, + 672, + 893, + 839 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For the hyperparameters, we empirically set the probability threshold $\\tau$ to 0.95 following FixMatch [67]. We find that $\\tau = 0.95$ works well across all settings and tasks. For the temperature $T_{co}$ in both contrastive losses,", + "bbox": [ + 496, + 840, + 893, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "7623", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/c87709867c41b2deef8453b51d8fe983bbefa6b1e63235f9c73fa354e8cadb60.jpg", + "table_caption": [ + "Table 1. Single-Source Unsupervised DA (SUDA) on Office-Home." + ], + "table_footnote": [], + "table_body": "
MethodSFOffice-Home
Ar→ClAr→PrAr→RwCl→ArCl→PrCl→RwPr→ArPr→ClPr→RwRw→ArRw→ClRw→PrAvg
SENTRY (ICCV'21) [59]X61.877.480.166.371.674.766.863.080.974.066.384.172.2
FixBi (CVPR'21) [52]X58.177.380.467.779.578.165.857.981.776.462.986.772.7
SCDA (ICCV'21) [38]X60.776.482.869.877.578.468.959.082.774.961.884.573.1
A²Net (ICCV'21) [79]✓58.479.082.467.579.378.968.056.282.974.160.585.072.8
NRC (NeurIPS'21) [87]✓57.780.382.068.179.878.665.356.483.071.058.685.672.2
D-MCD (AAAI'22) [10]✓59.478.980.267.279.378.665.355.682.273.362.883.972.2
DIPE (CVPR'22) [74]✓56.579.280.770.179.878.867.955.183.574.159.384.872.5
Sub-Sup (ECCV'22) [29]✓61.080.482.569.179.979.569.157.882.774.565.186.474.0
BMD (ECCV'22) [61]✓58.179.782.669.381.080.770.857.683.674.060.085.973.6
U-SFAN (ECCV'22) [62]✓57.877.881.667.977.379.267.254.781.273.360.383.971.9
CoWA-JMDS (ICML'22) [33]✓56.978.481.069.180.079.967.757.282.472.860.584.572.5
Feat-Mixup (ICML'22) [30]✓61.881.283.068.580.679.467.861.585.173.764.186.574.5
VMP (NeurIPS'22) [25]✓57.977.682.568.679.480.668.455.683.175.259.684.772.8
DaC (NeurIPS'22) [98]✓59.179.581.269.378.979.267.456.482.474.061.484.472.8
SHOT-IM (ICML'20) [42]✓55.476.680.466.974.375.465.654.880.773.758.483.470.5
Ours+ SHOT-IM✓62.882.084.370.980.882.670.061.183.676.265.187.075.5 (+5.0)
AaD (NeurIPS'22) [90]✓59.379.382.168.979.879.567.257.483.172.158.585.472.7
Ours+ AaD✓63.582.185.073.082.782.469.562.982.674.265.787.375.9 (+3.2)
", + "bbox": [ + 76, + 102, + 893, + 361 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/b994a1ed63dca05df72116415ba61cbb956dbcf77fd335155d9c9c321342a446.jpg", + "table_caption": [ + "Table 2. Single-Source Unsupervised Domain Adaptation (SUDA) on Office-31 and VisDA benchmarks." + ], + "table_footnote": [], + "table_body": "
MethodSFOffice-31VisDA
A→DA→WD→WW→DD→AW→A AvgS→R
BCDM (AAAI'20) [37]✗93.895.498.6100.073.173.089.083.4
MCC (ECCV'20) [24]✗94.495.598.6100.072.974.989.478.8
FixBi (CVPR'21) [52]✗95.096.199.3100.078.779.491.487.2
RADA (ICCV'21) [23]✗96.196.299.3100.077.577.491.176.3
FAA (ICCV'21) [22]✗94.492.399.299.780.578.790.882.7
SCDA (ICCV'21) [38]✗95.495.399.0100.077.275.990.5-
SHOT (ICML'20) [42]✓94.090.198.499.974.774.388.682.9
CPGA (IJCAI'21) [60]✓94.494.198.499.876.076.689.984.1
VDM-DA (TCSVT'21) [71]✓93.294.198.0100.075.877.189.785.1
A2Net (ICCV'21) [79]✓94.594.099.2100.076.776.190.184.3
HCL (NeurIPS'21) [21]✓90.891.398.2100.072.772.787.683.5
NRC (NeurIPS'21) [87]✓96.090.899.0100.075.375.089.485.9
SHOT++ (TPAMI'21) [43]✓94.390.498.799.976.275.889.287.3
D-MCD (AAAI'22) [10]✓94.193.598.8100.076.476.489.987.5
DIPE (CVPR'22) [74]✓96.693.198.499.675.577.290.183.1
Sub-Sup (ECCV'22) [29]✓95.694.699.299.877.077.790.788.2
BMD (ECCV'22) [61]✓96.294.298.0100.076.076.090.188.7
CoWA-JMDS (ICML'22) [33]✓94.495.298.5100.076.277.690.386.9
Feat-Mixup (ICML'22) [30]✓94.693.298.9100.078.378.990.787.8
VMP (NeurIPS'22) [25]✓93.396.298.6100.075.476.990.0-
SHOT-IM (ICML'20) [42]✓90.691.298.399.972.571.487.380.4
Ours + SHOT-IM✓95.895.199.0100.076.678.390.8(+3.5)89.1(+8.7)
AaD (NeurIPS'22) [90]✓96.492.199.1100.075.076.589.988.0
Ours + AaD✓96.695.599.1100.076.978.391.1(+1.2)89.6(+1.6)
", + "bbox": [ + 78, + 412, + 468, + 672 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "we set it to 0.07 following previous contrastive learning methods [18, 21, 85]. For the $\\lambda_{inter}$ in our similarity, we use 1.0 for simplicity. For the trade-off hyperparameters $\\lambda_{caco}, \\lambda_{idco}$ , we directly set them to 1.0. More training details can be found in the supplementary material.", + "bbox": [ + 75, + 696, + 468, + 773 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Results", + "text_level": 1, + "bbox": [ + 76, + 796, + 171, + 811 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The results are shown in Table 1, Table 2, Table 3 and Table 4. In each table, SF indicates source-free, * indicates results we reproduced from the released code. (+x.x) indicates gains over SHOT-IM [42] and AaD [90] respectively. Ours means combination of our proposed two losses.", + "bbox": [ + 75, + 824, + 470, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Single Source Domain Adaptation Table 1 and Table 2 show the classification accuracy for single-source DA on each dataset: Office-Home, Office-31, and VisDA. Based on the simple SHOT-IM and stronger AaD, our method can achieve state-of-the-art performance. UDA methods BCDM [37] and MCC [24] also explore the inter-class relationship by directly suppressing the inter-class term to obtain determined outputs. Our method surpasses these methods even without source domain supervision.", + "bbox": [ + 496, + 383, + 890, + 518 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The two proposed contrastive losses exploit different views of the same image. Thus our method is complementary to those not utilizing information of different views. Both baselines only consider weak views, and so our method can boost them. AaD surpasses IM by regarding local neighbors as positive pairs. In our CR-IDCo loss, the strong views of the same image can be regarded as neighbors since they are very close in feature space, which has overlapped effect with AaD. When our method is combined with AaD, therefore, the improvement is relatively smaller.", + "bbox": [ + 496, + 520, + 892, + 672 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Multi Source Domain Adaptation Table 3 shows the results for multi-source DA on Office-Home and the largescale DomainNet benchmark. We adopt two baselines without using domain labels, that is, all data from different source domains is treated as belonging to a single domain and only one source model is trained and transferred. We observe improvements of $4.1\\%$ and $3.4\\%$ in DomainNet, $4.5\\%$ and $3.0\\%$ in Office-Home. CAiDA [14] considers class relationship by enforcing consistency of different source model outputs. It achieves $76.2\\%$ on Office-Home by combining IM loss, a novel confident-anchor-induced pseudo label method and the class-relation-space-aware consistency loss. Based on the same IM loss, ours achieves $78.1\\%$ and exceeds it by $1.9\\%$ .", + "bbox": [ + 496, + 688, + 893, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "7624", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/0b1ccb0ca9182052a039c3f2aed7f994245ed89173fc36247861f0b78a8f283b.jpg", + "table_caption": [ + "Table 3. Multi-Source Unsupervised Domain Adaptation (MUDA) on DomainNet and Office-Home." + ], + "table_footnote": [], + "table_body": "
MethodSFw/o Domain \nLabelsDomainNetOffice-Home
→C→I→P→Q→R→SAvg→Ar→Cl→Pr→RwAvg
MCC (ECCV'20) [24]XX65.526.056.616.568.052.747.6-----
CMSDA (BMVC'21) [66]XX70.926.557.521.368.159.450.471.567.784.182.976.6
DRT (CVPR'21) [39]XX71.031.661.012.371.460.751.3-----
STEM (ICCV'21) [53]XX72.028.261.525.772.660.253.4-----
SourcecombineX✓57.023.454.114.667.250.344.458.057.374.277.966.9
DECISION (CVPR'21) [1]✓X61.521.654.618.967.551.045.974.559.484.483.675.5
CAiDA (NeurIPS'21) [14]✓X-------75.260.584.784.276.2
Sub-Sup (ECCV'22) [29]✓✓70.325.757.317.169.957.149.675.164.186.684.477.6
Feat-Mixup (ICML'22) [30]✓✓75.424.657.823.665.858.551.072.667.485.983.677.4
SHOT-IM* (ICML'20) [42]✓✓64.524.255.215.567.053.246.671.558.681.982.373.6
Ours + SHOT-IM✓✓69.827.059.021.270.257.550.7(+4.1)74.966.485.185.878.1 (+4.5)
AaD* (NeurIPS'22) [90]✓✓65.725.556.417.068.154.647.972.160.685.384.475.6
Ours +AaD✓✓71.227.859.322.569.357.851.3(+3.4)75.466.986.485.978.6 (+3.0)
", + "bbox": [ + 78, + 99, + 890, + 275 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/f962328e11faae001b096f914b93baa83f19c8543ac910e3a8fd455ac3146ee1.jpg", + "table_caption": [ + "Table 4. Partial-set and open-set Domain Adaptation (PDA and ODA) on Office-Home." + ], + "table_footnote": [], + "table_body": "
Partial-set DASF Avg.Open-set DASF Avg.
ResNet-50 [19]× 61.3ResNet [19]× 65.3
IWAN (CVPR'18) [92]× 63.6OpenMax (CVPR'16) [4]× 66.7
SAN (CVPR'18) [5]× 65.3ATI-λ (ICCV'17) [56]× 66.1
ETN (CVPR'19) [6]× 70.5OSBP (ECCV'18) [65]× 65.7
SAFN (ICCV'19) [82]× 71.8STA (CVPR'19) [44]× 69.5
Source model only✓ 62.8Source model only✓ 46.6
SHOT (ICML'20) [42]✓ 79.3SHOT (ICML'20) [42]✓ 72.8
SHOT+HCL (NeurIPS'21) [21]✓ 80.1SHOT+HCL (NeurIPS'21) [21]✓ 73.2
CoWA-JMDS (ICML'22) [33]✓ 83.2CoWA-JMDS (ICML'22) [33]✓ 73.2
SHOT-IM (ICML'20) [42]✓ 76.8SHOT-IM (ICML'20) [42]✓ 71.5
Ours +SHOT-IM✓ 80.6 (+3.8)Ours +SHOT-IM✓ 73.2(+1.7)
AaD* (NeurIPS'22) [90]✓ 79.7AaD* (NeurIPS'22) [90]✓ 71.8
Ours +AaD✓ 82.4 (+2.7)Ours +AaD✓ 73.3(+1.5)
", + "bbox": [ + 78, + 323, + 464, + 459 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Domain Adaptation Beyond Vanilla Closed-set We provide additional results under source-free partial-set and open-set DA (PDA and ODA) setting on Office-Home. For open-set detection in ODA, we follow the same protocol to detect unseen categories as SHOT [42]. Results are shown in Table 4. It can be seen that our method can achieve performance comparable to that of the state-of-the-arts.", + "bbox": [ + 75, + 472, + 468, + 580 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Analysis", + "text_level": 1, + "bbox": [ + 76, + 587, + 181, + 603 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Component-wise ablations Here we validate the effectiveness of different proposed losses. As described in Sec. 3.4, we do not use CR-CACo loss alone. The results are shown in Table 5. It can be seen that when using our method alone (line #2) can achieve comparable performance. When combined with SHOT-IM, our both losses are complementary, and the best performance can be achieved by combining baseline and our losses. Note that although CR-IDCo (line #1) outperforms IM (line #0), it is worse than IM when combined with CR-CACo (i.e. line #2 < line #3). This is caused by the fact that CR-IDCo cannot produce enough confident samples as IM for the learning of CR-CACo loss.", + "bbox": [ + 75, + 609, + 470, + 806 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Effect of pseudo labels in CR-CACo In our proposed CR-CACo loss, the pseudo label is generated from clustering-based method in SHOT [42] instead of predictions of weakly augmented images. The results of the comparison are shown in Table 6. It can be seen that our", + "bbox": [ + 75, + 824, + 468, + 902 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/15d144692b73da0e54e1c931d2647ab5c0ccae4334c670002dfe64ef9f0a4726.jpg", + "table_caption": [ + "Table 5. Component-wise ablation studies of the proposed methods under the SUDA setting." + ], + "table_footnote": [], + "table_body": "
#IMCR-CACoCR-IDCoOffice-31Office-HomeVisDA
0✓87.370.580.4
1✓89.272.583.0
2✓✓89.674.085.3
3✓✓89.874.887.0
4✓✓90.274.385.8
5✓✓✓90.875.589.1
", + "bbox": [ + 503, + 329, + 888, + 422 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "CR-CACo can bring significant improvement when using pseudo label from model prediction. The clustering based pseudo-label can bring slight but consistent improvements.", + "bbox": [ + 498, + 449, + 890, + 494 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/720436a42396060a1f628f55f06242983d3f1b94f2e6db180be5739a245a69d1.jpg", + "table_caption": [ + "Table 6. Comparison of different pseudo label methods." + ], + "table_footnote": [], + "table_body": "
#BaselinePseudo-label in CR-CACoOffice-31Office-HomeVisDA
PredictionClustering
087.370.580.4
1IM✓89.6(+2.3)74.5(+4.0)86.6(+6.2)
2✓89.8(+2.5)74.8(+4.3)87.0(+6.6)
389.972.788.0
4AaD✓90.6(+0.7)75.0(+2.3)88.9(+0.9)
5✓90.7(+0.8)75.2(+2.5)89.1(+1.1)
", + "bbox": [ + 503, + 523, + 888, + 636 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Different similarities and disentangled improvements. To validate the superiority of our class relationship embedded similarity in contrastive loss, we compare it with other two choices. The first one uses only $s_{intra}$ based on features, and the second one uses only $s_{intra}$ based on probabilities which equals $\\lambda_{inter} = 0$ in our case. The results are shown in Table 7. Using features (i.e. i.e. $s_{intra}$ w/ feature) rather than probabilities is a more natural way. Indeed, using features in CACo Loss would result in FixMatch loss, and using features in IDCo loss would fall in the traditional self-supervised way (e.g., MoCo). Only using the probability based intra-class term (i.e. $s_{intra}$ w/ probability) is still special for similarity in contrastive losses. And it can bring consistent improvements. In both CACo and IDCo losses, our $s_{cr}$ can achieve the best result, thus validating the effectiveness of adopting source prior class similarities.", + "bbox": [ + 496, + 662, + 890, + 905 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7625", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/6e4869d08700b71dfd0cd985adbe7be697cdc04a3ca686a873cf159aada6e5d9.jpg", + "table_caption": [ + "Table 7. Comparison with different similarities." + ], + "table_footnote": [], + "table_body": "
#LossSimilarityOffice-31Office-HomeVisDA
0IM—87.370.580.4
1IM+CACosintra w/ feature88.972.583.5
2sintra w/ probability89.172.984.0
3scr w/ probability89.874.887.0
4IM+IDCosintra w/ feature87.570.980.8
5sintra w/ probability88.271.481.5
6scr w/ probability90.274.385.8
", + "bbox": [ + 81, + 107, + 465, + 217 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Comparison with fixed coefficient within $s_{inter}$ In the definition of inter-class term in Equation (1), we use the source class similarity matrix $A^s$ as coefficient. Instead of this prior knowledge, we can also set the coefficient to a small fixed value. Considering the positive pairs in contrastive loss, the negative coefficient (i.e. -0.1) works to suppress the inter-class term, thus obtaining more determined predictions as in [24, 37] while the positive value (i.e. 0.1) is more like a smooth regularization as KLD loss in [100]. The results are shown in Figure 4. It can be seen that both positive and negative values can bring some improvements. However, our CR-CACo loss can achieve better performance, which validates the importance of source prior class relationship.", + "bbox": [ + 75, + 226, + 468, + 438 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/40eee282579e5154d2c9b30b11ecb5601bfe9e42b8e5d5cc9267255ae70d646e.jpg", + "image_caption": [ + "Figure 4. Using fixed coefficients in $s_{inter}$ in both CACo and IDCo losses under SUDA setting Office-Home Rw→Cl. Black dotted line shows the performance of baseline IM. Red dotted line shows the performance of our losses." + ], + "image_footnote": [], + "bbox": [ + 88, + 448, + 267, + 547 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/97f9aace554f4bf3963be3b5a63656b1c009b8a3cacc3cc5828033531f0184ed.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 272, + 448, + 454, + 547 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Parameter sensitivity Here we analyze the hyperparameter sensitivity under the setting of SUDA Office-Home with SHOT-IM baseline. We consider four hyperparameters, as shown in Figure 5. It is evident that they are stable within specific ranges.", + "bbox": [ + 75, + 619, + 468, + 696 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/fd46b8a5fb321a60c20f354ffc3153494d624b1ad57f0ea342e5abb3b0b397f5.jpg", + "image_caption": [ + "Figure 5. Sensitivity analysis of different hyperparameters. The first two use our two losses. The last two use corresponding loss." + ], + "image_footnote": [], + "bbox": [ + 81, + 709, + 178, + 768 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/e2ec62eae2eee709ec6a4f0d714a9e827c2bab0869a754c906934091a6e6a6c3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 181, + 710, + 271, + 768 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/3b5ebe708294f823b942c00d64035aaaeafae24c2449cd2d7241bea804eb5702.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 272, + 710, + 372, + 768 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/5adb7af2bb021a2f1e0209afda563b0f312537b89fe91b92ac2309e8dbadb225.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 372, + 710, + 460, + 768 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Visualization of features We visualize the target features and classifier weights in Figure 6. It can be seen that our method has less misclassified samples for classes 0, 1, 7 compared with IM. Thus our method can learn more discriminative features.", + "bbox": [ + 75, + 824, + 467, + 898 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/a96140f46635e00fcd21fda7f2e65eac519197eb71608c757f8a972d12758aff.jpg", + "image_caption": [ + "Source Pre-trained Model" + ], + "image_footnote": [], + "bbox": [ + 509, + 88, + 633, + 181 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/c8de4fefb7c9b6cb96f3beda508aebde9f7d424866d32e8f17d93ad2fd091e9b.jpg", + "image_caption": [ + "IM", + "Figure 6. The t-SNE visualization of classifier weights (i.e. star) and target domain features (i.e. dot) of different methods under the setting of SUDA Office-Home $\\mathrm{Rw} \\rightarrow \\mathrm{Cl}$ . We directly use the first 10 classes (i.e. different colors), and the coordinates of classifier weights are the same across different methods since source classifier is frozen. Best viewed in color." + ], + "image_footnote": [], + "bbox": [ + 635, + 88, + 759, + 181 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/7dc7c4a2f18f0de05efdcba2655e03b7e7cea50aada2308289fbccfad77834e3.jpg", + "image_caption": [ + "IM+Ours" + ], + "image_footnote": [], + "bbox": [ + 763, + 89, + 885, + 181 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Transformer backbone We further apply our method with Transformer-based backbone under the SFUDA setting, as shown in Table 8. We choose the same network (i.e. ViT-B [15]) following SSRT [69]. For the training of the source model, we follow SHOT [42] which obtains better performance than original ViT-B (78.2% vs. 75.5%). Adding IM loss achieves 84.0%. Based on IM loss, ours achieves 88.3% which significantly outperforms SSRT.", + "bbox": [ + 496, + 311, + 890, + 434 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/7ddc7ae76cbd9514fba365541210a7737d45d790ab4409a4aa532da9f2ecbe96.jpg", + "table_caption": [ + "Table 8. Single-Source Unsupervised DA (SUDA) on Office-Home with ViT-B backbone." + ], + "table_footnote": [], + "table_body": "
MethodSFAvg.MethodSFAvg.
ViT-B✗75.5CDTrans (ICLR'22) [83]✗80.5
TVT (WACV'23) [86]✗83.6DOT-B (ACMMM'22) [50]✗84.3
SSRT (CVPR'22) [69]✗85.4BCAT-DTF (Arxiv'22) [76]✗86.6
Source model✓78.2SHOT-IM✓84.0
SHOT-IM+CR-CACo✓87.0(+3.0)SHOT-IM+Ours✓88.3(+4.3)
", + "bbox": [ + 503, + 484, + 883, + 556 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 565, + 617, + 580 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we propose to explicitly transfer the class relationship for SFUDA which is more domain-invariant. We propose a novel class relationship embedded similarity that can more accurately express the sample relationship in the output space. Furthermore, we propose two contrastive losses (i.e., CR-CACo and CR-IDCo) that exploit our designed similarity. These two losses are complementary, and their combination can better explore the target distribution. We combine our method with existing representative baselines in multiple SFUDA settings. Extensive experiments show the effectiveness of the proposed method, which achieves state-of-the-art performance.", + "bbox": [ + 496, + 590, + 890, + 772 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Acknowledgements", + "text_level": 1, + "bbox": [ + 500, + 784, + 686, + 801 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This work is supported by the National Natural Science Foundation of China under Grant 62176246 and Grant 61836008. This work is also supported by Anhui Provincial Natural Science Foundation 2208085UD17 and the Fundamental Research Funds for the Central Universities (WK3490000006).", + "bbox": [ + 496, + 809, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "7626", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 174, + 106 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Sk. Miraj Ahmed, Dripta S. Raychaudhuri, S. Paul, Samet Oymak, and Amit K. Roy-Chowdhury. Unsupervised multi-source domain adaptation without access to source data. In CVPR, 2021. 7", + "[2] Shai Ben-David, John Blitzer, Koby Crammer, Alex Kulesza, Fernando Pereira, and Jennifer Wortman Vaughan. A theory of learning from different domains. Machine learning, 79(1-2):151-175, 2010. 2", + "[3] Shai Ben-David, John Blitzer, Koby Crammer, and Fernando Pereira. Analysis of representations for domain adaptation. In NeurIPS, 2006. 2", + "[4] Abhijit Bendale and Terrance E Boult. Towards open set deep networks. In CVPR, 2016. 7", + "[5] Zhangjie Cao, Mingsheng Long, Jianmin Wang, and Michael I Jordan. Partial transfer learning with selective adversarial networks. In CVPR, 2018. 7", + "[6] Zhangjie Cao, Kaichao You, Mingsheng Long, Jianmin Wang, and Qiang Yang. Learning to transfer examples for partial domain adaptation. In CVPR, 2019. 7", + "[7] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In ECCV, 2020. 1", + "[8] Chaoqi Chen, Weiping Xie, Tingyang Xu, Wenbing Huang, Yu Rong, Xinghao Ding, Yue Huang, and Junzhou Huang. Progressive feature alignment for unsupervised domain adaptation. In CVPR, 2019. 3", + "[9] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In ICML, 2020. 2, 4, 5", + "[10] Tong Chu, Yahao Liu, Jinhong Deng, Wen Li, and Lixin Duan. Denoised maximum classifier discrepancy for sourcefree unsupervised domain adaptation. In AAAI, 2022. 3, 6", + "[11] Safa Cicek and Stefano Soatto. Unsupervised domain adaptation via regularized conditional alignment. In ICCV, 2019. 2", + "[12] Shuhao Cui, Shuhui Wang, Junbao Zhuo, Chi Su, Qingming Huang, and Qi Tian. Gradually vanishing bridge for adversarial domain adaptation. In CVPR, 2020. 2", + "[13] Zhijie Deng, Yucen Luo, and Jun Zhu. Cluster alignment with a teacher for unsupervised domain adaptation. In ICCV, 2019. 3", + "[14] Jiahua Dong, Zhen Fang, Anjin Liu, Gan Sun, and Tongliang Liu. Confident anchor-induced multi-source free domain adaptation. In NeurIPS, 2021. 2, 6, 7", + "[15] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2020. 1, 8", + "[16] Geoffrey French, Michal Mackiewicz, and Mark Fisher. Self-ensembling for visual domain adaptation. In ICLR, 2018. 3" + ], + "bbox": [ + 86, + 114, + 470, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[17] Yuan Gao, Zilei Wang, Jiafan Zhuang, Yixin Zhang, and Junjie Li. Exploit domain-robust optical flow in domain adaptive video semantic segmentation. In AAAI, 2023. 3", + "[18] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In CVPR, 2020. 2, 3, 4, 5, 6", + "[19] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, 2016. 1, 5, 7", + "[20] Weihua Hu, Takeru Miyato, Seiya Tokui, Eiichi Matsumoto, and Masashi Sugiyama. Learning discrete representations via information maximizing self-augmented training. In ICML, 2017. 3", + "[21] Jiaxing Huang, Dayan Guan, Aoran Xiao, and Shijian Lu. Model adaptation: Historical contrastive learning for unsupervised domain adaptation without source data. In NeurIPS, 2021. 1, 2, 3, 4, 6, 7", + "[22] Jiaxing Huang, Dayan Guan, Aoran Xiao, and Shijian Lu. RDA: Robust domain adaptation via fourier adversarial attacking. In ICCV, 2021. 6", + "[23] Xin Jin, Cuiling Lan, Wenjun Zeng, and Zhibo Chen. Re-energizing domain discriminator with sample relabeling for adversarial domain adaptation. In ICCV, 2021. 6", + "[24] Ying Jin, Ximei Wang, Mingsheng Long, and Jianmin Wang. Minimum class confusion for versatile domain adaptation. In ECCV, 2020. 6, 7, 8", + "[25] Mengmeng Jing, Xiantong Zhen, Jingjing Li, and Cees GM Snoek. Variational model perturbation for source-free domain adaptation. In NeurIPS, 2022. 3, 6", + "[26] Guoliang Kang, Lu Jiang, Yi Yang, and Alexander G Hauptmann. Contrastive adaptation network for unsupervised domain adaptation. In CVPR, 2019. 3, 4", + "[27] Seunghyeon Kim, Jaehoon Choi, Taekyung Kim, and Changick Kim. Self-training and adversarial background regularization for unsupervised domain adaptive one-stage object detection. In CVPR, 2019. 3", + "[28] Ananya Kumar, Tengyu Ma, and Percy Liang. Understanding self-training for gradual domain adaptation. In ICML, 2020. 3", + "[29] Jogendra Nath Kundu, Suvaansh Bhambri, Akshay Kulkarni, Hiran Sarkar, Varun Jampani, and R Venkatesh Babu. Concurrent subsidiary supervision for unsupervised source-free domain adaptation. In ECCV, 2022. 3, 6, 7", + "[30] Jogendra Nath Kundu, Akshay R Kulkarni, Suvaansh Bhambri, Deepesh Mehta, Shreyas Anand Kulkarni, Varun Jampani, and Venkatesh Babu Radhakrishnan. Balancing discriminability and transferability for source-free domain adaptation. In ICML, 2022. 3, 6, 7", + "[31] Jogendra Nath Kundu, Naveen Venkat, and R Venkatesh Babu. Universal source-free domain adaptation. In CVPR, 2020. 3", + "[32] Jogendra Nath Kundu, Naveen Venkat, Ambareesh Revanur, Rahul M V, and R. Venkatesh Babu. Towards inheritable models for open-set domain adaptation. In CVPR, 2020. 3", + "[33] Jonghyun Lee, Dahuin Jung, Junho Yim, and Sungroh Yoon. Confidence score for source-free unsupervised domain adaptation. In ICML, 2022. 3, 6, 7" + ], + "bbox": [ + 509, + 92, + 890, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "7627", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[34] Junjie Li, Zilei Wang, Yuan Gao, and Xiaoming Hu. Exploring high-quality target domain information for unsupervised domain adaptive semantic segmentation. In ACMMM, 2022. 3", + "[35] Junjie Li, Yixin Zhang, Zilei Wang, and Keyu Tu. Semantic-aware representation learning via probability contrastive loss. arXiv preprint arXiv:2111.06021, 2021. 5", + "[36] Rui Li, Qianfen Jiao, Wenming Cao, Hau-San Wong, and Si Wu. Model adaptation: Unsupervised domain adaptation without source data. In CVPR, 2020. 1, 2, 3", + "[37] Shuang Li, Fangrui Lv, Binhui Xie, Chi Harold Liu, Jian Liang, and Chen Qin. Bi-classifier determinacy maximization for unsupervised domain adaptation. In AAAI, 2021. 6, 8", + "[38] Shuang Li, Mixue Xie, Fangrui Lv, Chi Harold Liu, Jian Liang, Chen Qin, and Wei Li. Semantic concentration for domain adaptation. In ICCV, 2021. 6", + "[39] Yunsheng Li, Lu Yuan, Yinpeng Chen, Pei Wang, and N. Vasconcelos. Dynamic transfer for multi-source domain adaptation. In CVPR, 2021. 7", + "[40] Yunsheng Li, Lu Yuan, and Nuno Vasconcelos. Bidirectional learning for domain adaptation of semantic segmentation. In CVPR, 2019. 3", + "[41] Yu-Jhe Li, Xiaoliang Dai, Chih-Yao Ma, Yen-Cheng Liu, Kan Chen, Bichen Wu, Zijian He, Kris Kitani, and Peter Vajda. Cross-domain adaptive teacher for object detection. In CVPR, 2022. 4", + "[42] Jian Liang, Dapeng Hu, and Jiashi Feng. Do we really need to access the source data? source hypothesis transfer for unsupervised domain adaptation. In ICML, 2020. 1, 2, 3, 4, 5, 6, 7, 8", + "[43] Jian Liang, Dapeng Hu, Yunbo Wang, Ran He, and Jiashi Feng. Source data-absent unsupervised domain adaptation through hypothesis transfer and labeling transfer. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2021. 6", + "[44] Hong Liu, Zhangjie Cao, Mingsheng Long, Jianmin Wang, and Qiang Yang. Separate to adapt: Open set domain adaptation via progressive separation. In CVPR, 2019. 7", + "[45] Hong Liu, Mingsheng Long, Jianmin Wang, and Michael Jordan. Transferable adversarial training: A general approach to adapting deep classifiers. In ICML, 2019. 2", + "[46] Hong Liu, Jianmin Wang, and Mingsheng Long. Cycle self-training for domain adaptation. In NeurIPS, 2021. 3", + "[47] Qinying Liu and Zilei Wang. Collaborating domain-shared and target-specific feature clustering for cross-domain 3d action recognition. In ECCV, 2022. 3", + "[48] Mingsheng Long, Zhangjie Cao, Jianmin Wang, and Michael I Jordan. Conditional adversarial domain adaptation. In NeurIPS, 2018. 2", + "[49] Mingsheng Long, Han Zhu, Jianmin Wang, and Michael I Jordan. Deep transfer learning with joint adaptation networks. In ICML, 2017. 2", + "[50] Wenxuan Ma, Jinming Zhang, Shuang Li, Chi Harold Liu, Yulin Wang, and Wei Li. Making the best of both worlds: A domain-oriented transformer for unsupervised domain adaptation. In ACMMM, 2022. 8" + ], + "bbox": [ + 86, + 90, + 470, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[51] Ke Mei, Chuang Zhu, Jiaqi Zou, and Shanghang Zhang. Instance adaptive self-training for unsupervised domain adaptation. In ECCV, 2020. 3", + "[52] Jaemin Na, Heechul Jung, Hyung Jin Chang, and Wonjun Hwang. FixBi: Bridging domain spaces for unsupervised domain adaptation. In CVPR, 2021. 6", + "[53] Van-Anh Nguyen, Tuan Nguyen, Trung Le, Quan Hung Tran, and Dinh Phung. STEM: An approach to multi-source domain adaptation with guarantees. In ICCV, 2021. 7", + "[54] Fei Pan, Inkyu Shin, Francois Rameau, Seokju Lee, and In So Kweon. Unsupervised intra-domain adaptation for semantic segmentation through self-supervision. In CVPR, 2020. 3", + "[55] Yingwei Pan, Ting Yao, Yehao Li, Yu Wang, Chong-Wah Ngo, and Tao Mei. Transferrable prototypical networks for unsupervised domain adaptation. In CVPR, 2019. 3", + "[56] Pau Panareda Busto and Juergen Gall. Open set domain adaptation. In ICCV, 2017. 7", + "[57] Xingchao Peng, Qinxun Bai, Xide Xia, Zijun Huang, Kate Saenko, and Bo Wang. Moment matching for multi-source domain adaptation. In ICCV, 2019. 2, 5", + "[58] Xingchao Peng, Ben Usman, Neela Kaushik, Judy Hoffman, Dequan Wang, and Kate Saenko. VisDA: The visual domain adaptation challenge. In CVPRW, 2018. 5", + "[59] Viraj Prabhu, Shivam Khare, Deeksha Kartik, and Judy Hoffman. SENTRY: Selective entropy optimization via committee consistency for unsupervised domain adaptation. In ICCV, 2021. 6", + "[60] Zhen Qiu, Yifan Zhang, Hongbin Lin, Shuaicheng Niu, Yanxia Liu, Qing Du, and Mingkui Tan. Source-free domain adaptation via avatar prototype generation and adaptation. In IJCAI, 2021. 6", + "[61] Sanqing Qu, Guang Chen, Jing Zhang, Zhijun Li, Wei He, and Dacheng Tao. Bmd: A general class-balanced multicentric dynamic prototype strategy for source-free domain adaptation. In ECCV, 2022. 3, 6", + "[62] Subhankar Roy, Martin Trapp, Andrea Pilzer, Juho Kannala, Nicu Sebe, Elisa Ricci, and Arno Solin. Uncertainty-guided source-free domain adaptation. In ECCV, 2022. 3, 6", + "[63] Kate Saenko, Brian Kulis, Mario Fritz, and Trevor Darrell. Adapting visual category models to new domains. In ECCV, 2010. 5", + "[64] Kuniaki Saito, Kohei Watanabe, Yoshitaka Ushiku, and Tatsuya Harada. Maximum classifier discrepancy for unsupervised domain adaptation. In CVPR, 2018. 2, 3", + "[65] Kuniaki Saito, Shohei Yamamoto, Yoshitaka Ushiku, and Tatsuya Harada. Open set domain adaptation by backpropagation. In ECCV, 2018. 7", + "[66] Marin Scalbert, Maria Vakalopoulou, and Florent Couzini'e-Devy. Multi-source domain adaptation via supervised contrastive learning and confident consistency regularization. In BMVC, 2021. 7", + "[67] Kihyuk Sohn, David Berthelot, Chun-Liang Li, Zizhao Zhang, Nicholas Carlini, Ekin D Cubuk, Alex Kurakin, Han Zhang, and Colin Raffel. Fixmatch: Simplifying semi-supervised learning with consistency and confidence. In NeurIPS, 2020. 2, 4, 5" + ], + "bbox": [ + 509, + 90, + 890, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "7628", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[68] Baochen Sun and Kate Saenko. Deep coral: Correlation alignment for deep domain adaptation. In ECCV, 2016. 2", + "[69] Tao Sun, Cheng Lu, Tianshuo Zhang, and Haibin Ling. Safe self-refinement for transformer-based domain adaptation. In CVPR, 2022. 8", + "[70] Hui Tang, Ke Chen, and Kui Jia. Unsupervised domain adaptation via structurally regularized deep clustering. In CVPR, 2020. 3", + "[71] Jiayi Tian, Jing Zhang, Wen Li, and Dong Xu. VDMDA: Virtual domain modeling for source data-free domain adaptation. IEEE Transactions on Circuits and Systems for Video Technology, 2021. 1, 6", + "[72] Eric Tzeng, Judy Hoffman, Ning Zhang, Kate Saenko, and Trevor Darrell. Deep domain confusion: Maximizing for domain invariance. arXiv preprint arXiv:1412.3474, 2014. 2", + "[73] Hemanth Venkateswara, Jose Eusebio, Shayok Chakraborty, and Sethuraman Panchanathan. Deep hashing network for unsupervised domain adaptation. In CVPR, 2017. 5", + "[74] Fan Wang, Zhongyi Han, Yongshun Gong, and Yilong Yin. Exploring domain-invariant parameters for source free domain adaptation. In CVPR, 2022. 3, 6", + "[75] Feng Wang and Huaping Liu. Understanding the behaviour of contrastive loss. In CVPR, 2021. 2", + "[76] Xiyu Wang, Pengxin Guo, and Yu Zhang. Domain adaptation via bidirectional cross-attention transformer. arXiv preprint arXiv:2201.05887, 2022. 8", + "[77] Ximei Wang, Liang Li, Weirui Ye, Mingsheng Long, and Jianmin Wang. Transferable attention for domain adaptation. In AAAI, 2019. 2", + "[78] Zhirong Wu, Yuanjun Xiong, Stella X Yu, and Dahua Lin. Unsupervised feature learning via non-parametric instance discrimination. In CVPR, 2018. 3", + "[79] Haifeng Xia, Handong Zhao, and Zhengming Ding. Adaptive adversarial network for source-free domain adaptation. In ICCV, 2021. 1, 2, 3, 6", + "[80] Shaoan Xie, Zibin Zheng, Liang Chen, and Chuan Chen. Learning semantic representations for unsupervised domain adaptation. In ICML, 2018. 3", + "[81] Minghao Xu, Hang Wang, Bingbing Ni, Qi Tian, and Wenjun Zhang. Cross-domain detection via graph-induced prototype alignment. In CVPR, 2020. 3", + "[82] Ruijia Xu, Guanbin Li, Jihan Yang, and Liang Lin. Larger norm more transferable: An adaptive feature norm approach for unsupervised domain adaptation. In ICCV, 2019. 7", + "[83] Tongkun Xu, Weihua Chen, WANG Pichao, Fan Wang, Hao Li, and Rong Jin. Cdtrans: Cross-domain transformer for unsupervised domain adaptation. In International Conference on Learning Representations, 2021. 8", + "[84] Hongliang Yan, Yukang Ding, Peihua Li, Qilong Wang, Yong Xu, and Wangmeng Zuo. Mind the class weight bias: Weighted maximum mean discrepancy for unsupervised domain adaptation. In CVPR, 2017. 2", + "[85] Fan Yang, Kai Wu, Shuyi Zhang, Guannan Jiang, Yong Liu, Feng Zheng, Wei Zhang, Chengjie Wang, and Long" + ], + "bbox": [ + 86, + 92, + 467, + 900 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Zeng. Class-aware contrastive semi-supervised learning. In CVPR, 2022. 2, 4, 6", + "[86] Jinyu Yang, Jingjing Liu, Ning Xu, and Junzhou Huang. Tvt: Transferable vision transformer for unsupervised domain adaptation. In WACV, 2023. 8", + "[87] Shiqi Yang, Joost van de Weijer, Luis Herranz, Shangling Jui, et al. Exploiting the intrinsic neighborhood structure for source-free domain adaptation. In NeurIPS, 2021. 1, 2, 3, 6", + "[88] Shiqi Yang, Yaxing Wang, Joost van de Weijer, Luis Herranz, and Shangling Jui. Unsupervised domain adaptation without source data by casting a bait. arXiv preprint arXiv:2010.12427, 2020. 1", + "[89] Shiqi Yang, Yaxing Wang, Joost van de Weijer, Luis Herranz, and Shangling Jui. Generalized source-free domain adaptation. In ICCV, 2021. 2, 3", + "[90] Shiqi Yang, Yaxing Wang, Kai Wang, Shangling Jui, and Joost van de Weijer. Attracting and dispersing: A simple approach for source-free domain adaptation. In NeurIPS, 2022. 2, 3, 5, 6, 7", + "[91] Bowen Zhang, Yidong Wang, Wenxin Hou, Hao Wu, Jindong Wang, Manabu Okumura, and Takahiro Shinozaki. Flexmatch: Boosting semi-supervised learning with curriculum pseudo labeling. NeurIPS, 2021. 2", + "[92] Jing Zhang, Zewei Ding, Wanqing Li, and Philip Ogunbona. Importance weighted adversarial nets for partial domain adaptation. In CVPR, 2018. 7", + "[93] Pan Zhang, Bo Zhang, Ting Zhang, Dong Chen, Yong Wang, and Fang Wen. Prototypical pseudo label denoising and target structure learning for domain adaptive semantic segmentation. In CVPR, 2021. 3, 4", + "[94] Weichen Zhang, Wanli Ouyang, Wen Li, and Dong Xu. Collaborative and adversarial network for unsupervised domain adaptation. In CVPR, 2018. 3", + "[95] Yiheng Zhang, Zhaofan Qiu, Ting Yao, Dong Liu, and Tao Mei. Fully convolutional adaptation networks for semantic segmentation. In CVPR, 2018. 1", + "[96] Yixin Zhang and Zilei Wang. Joint adversarial learning for domain adaptation in semantic segmentation. In AAAI, 2020. 2", + "[97] Yixin Zhang, Zilei Wang, and Yushi Mao. Rpn prototype alignment for domain adaptive object detector. In CVPR, 2021. 3", + "[98] Ziyi Zhang, Weikai Chen, Hui Cheng, Zhen Li, Siyuan Li, Liang Lin, and Guanbin Li. Divide and contrast: Source-free domain adaptation via adaptive contrastive learning. In NeurIPS, 2022. 6", + "[99] Yangtao Zheng, Di Huang, Songtao Liu, and Yunhong Wang. Cross-domain object detection through coarse-to-fine feature adaptation. In CVPR, 2020. 3", + "[100] Yang Zou, Zhiding Yu, Xiaofeng Liu, BVK Kumar, and Jinsong Wang. Confidence regularized self-training. In ICCV, 2019. 8", + "[101] Yang Zou, Zhiding Yu, BVK Vijaya Kumar, and Jinsong Wang. Unsupervised domain adaptation for semantic segmentation via class-balanced self-training. In ECCV, 2018. 3" + ], + "bbox": [ + 503, + 92, + 890, + 891 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "7629", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2023/Class Relationship Embedded Learning for Source-Free Unsupervised Domain Adaptation/a0503f36-a779-42c8-98b9-ccfd34b3ed5c_model.json b/2023/Class Relationship Embedded Learning for Source-Free Unsupervised Domain Adaptation/a0503f36-a779-42c8-98b9-ccfd34b3ed5c_model.json new file mode 100644 index 0000000000000000000000000000000000000000..7f6aec5202586e6f3ab33aaf89d1d9f9878a7c81 --- /dev/null +++ b/2023/Class Relationship Embedded Learning for Source-Free Unsupervised Domain Adaptation/a0503f36-a779-42c8-98b9-ccfd34b3ed5c_model.json @@ -0,0 +1,2774 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.23, + 0.131, + 0.74, + 0.177 + ], + "angle": 0, + "content": "Class Relationship Embedded Learning for Source-Free Unsupervised Domain Adaptation" + }, + { + "type": "text", + "bbox": [ + 0.304, + 0.203, + 0.667, + 0.221 + ], + "angle": 0, + "content": "Yixin Zhang\\(^{1,2}\\) Zilei Wang\\(^{*2}\\) Weinan He\\(^{2}\\)" + }, + { + "type": "text", + "bbox": [ + 0.162, + 0.221, + 0.81, + 0.24 + ], + "angle": 0, + "content": "\\(^{1}\\) Institute of Artificial Intelligence, Hefei Comprehensive National Science Center" + }, + { + "type": "text", + "bbox": [ + 0.293, + 0.239, + 0.679, + 0.257 + ], + "angle": 0, + "content": "\\(^{2}\\) University of Science and Technology of China" + }, + { + "type": "text", + "bbox": [ + 0.244, + 0.259, + 0.721, + 0.275 + ], + "angle": 0, + "content": "{zhyx12, zlwang}@ustc.edu.cn, hwn2018@mail.ustc.edu.cn" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.309, + 0.314, + 0.325 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.341, + 0.474, + 0.706 + ], + "angle": 0, + "content": "This work focuses on a practical knowledge transfer task defined as Source-Free Unsupervised Domain Adaptation (SFUDA), where only a well-trained source model and unlabeled target data are available. To fully utilize source knowledge, we propose to transfer the class relationship, which is domain-invariant but still under-explored in previous works. To this end, we first regard the classifier weights of the source model as class prototypes to compute class relationship, and then propose a novel probability-based similarity between target-domain samples by embedding the source-domain class relationship, resulting in Class Relationship embedded Similarity (CRS). Here the inter-class term is particularly considered in order to more accurately represent the similarity between two samples, in which the source prior of class relationship is utilized by weighting. Finally, we propose to embed CRS into contrastive learning in a unified form. Here both class-aware and instance discrimination contrastive losses are employed, which are complementary to each other. We combine the proposed method with existing representative methods to evaluate its efficacy in multiple SFUDA settings. Extensive experimental results reveal that our method can achieve state-of-the-art performance due to the transfer of domain-invariant class relationship." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.734, + 0.21, + 0.75 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.759, + 0.47, + 0.866 + ], + "angle": 0, + "content": "Benefiting from the large amount of labeled training data, deep neural networks have achieved promising results in many computer vision tasks [7, 15, 19, 95]. To reduce the annotation cost, Unsupervised Domain Adaptation (UDA) has been devised by transferring knowledge from a label-rich source domain to a label-scarce target domain. Currently, many UDA methods have been proposed that jointly" + }, + { + "type": "image", + "bbox": [ + 0.507, + 0.306, + 0.891, + 0.548 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.561, + 0.894, + 0.631 + ], + "angle": 0, + "content": "Figure 1. Illustration of our proposed method. The upper left represents the target-domain featural distribution by source pre-trained model. The upper right is the feature distribution obtained by our method. The bottom is the process of class relationship embedded contrastive learning. Best viewed in color." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.659, + 0.893, + 0.825 + ], + "angle": 0, + "content": "learn on the source and target data. But they would be unapplicable for some real-world scenarios involving privacy (e.g., medical images, surveillance videos) because the source-domain data cannot be accessed. Thus, more recent methods [36, 42, 71, 79, 87, 88] focus on Source-Free Unsupervised Domain Adaptation (SFUDA). Under this setting, the labeled source data are not accessible any more when training the target model, but the pre-trained model in the source domain is provided. Then a natural question arises, i.e., what knowledge should we transfer to facilitate the learning of unlabeled target-domain data?" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.825, + 0.894, + 0.901 + ], + "angle": 0, + "content": "Some methods [21, 36, 42] assume that the source hypothesis (i.e., classifier [42] or whole model [21, 36]) contains sufficient knowledge for target domain. Then they transfer source knowledge by directly aligning features with the fixed source classifier [42], resorting to historical mod" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.875, + 0.223, + 0.888 + ], + "angle": 0, + "content": "*Corresponding author" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.888, + 0.449, + 0.9 + ], + "angle": 0, + "content": "\\(^{1}\\)Code is available at https://github.com/zhyx12/CRCo" + }, + { + "type": "list", + "bbox": [ + 0.096, + 0.875, + 0.449, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7619" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.471, + 0.183 + ], + "angle": 0, + "content": "els [21], or weight regularization [36]. Another line of works [87, 89, 90] assume the source model already forms some semantic structure, and then utilize the local information (i.e., neighborhood) of feature space to enforce the similarity in the output space. Despite these progress, what knowledge to be transferred remains an open question." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.186, + 0.47, + 0.413 + ], + "angle": 0, + "content": "In this work, we propose to transfer the class relationship represented by the similarity of classifier weights. Actually, the class relationship is domain-invariant [14], since the same class in the source and target domains essentially represents the same semantic in spite of domain discrepancy. For example, the computers are always more similar with the TV than the scissors. Thus, it is reasonable to guide the target domain learning using class relationship prior. Unlike previous methods that learn class distribution by pseudo labeling [21, 42] and local aggregation of neighborhood predictions [79, 87, 89, 90], here we explicitly model the source-domain class relationship. To be specific, we regard each weight of classifier as the class prototype [90], and then compute a class relationship matrix \\( A^s \\) by cosine similarity, as shown in Figure 1." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.415, + 0.47, + 0.627 + ], + "angle": 0, + "content": "Before explaining how to use this matrix, we illustrate the purpose of representation learning in the target domain using Figure 1, where three classes are adopted for clarity. The top left shows the target-domain feature distribution together with the class weights learned from the source domain. It can be seen that such a situation makes it difficult to perform correct classification. In fact, it is expected that the learned features are discriminative and compact around the corresponding class weight, as shown on the top right. To this end, an intuitive way is to make the relationship of learned class prototypes in the target domain consistent with that in the source domain. However, it has a very limited effect on training the target-domain model as such a prototype-level constraint is too weak for optimization." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.629, + 0.47, + 0.903 + ], + "angle": 0, + "content": "In this work, we propose to embed the source-domain class relationship in contrastive learning, which has been shown to be outstanding in representation learning [9, 18]. Here we design a novel sample similarity by taking into account \\( A^s \\). Specifically, we compute the similarity between two target domain samples in the output space, represented by the prediction probabilities (i.e., \\( p \\) and \\( p' \\)), as shown in Figure 1. In particular, we consider the inter-class term (i.e. \\( \\sum_{i \\neq j} p_i p_j' \\)) in addition to the traditional intra-class term (i.e. \\( \\sum_i p_i p_i' \\)). Note that the sum of these two terms equals one, in which the intra-class term measures the similarity that two samples come from the same class, while the inter-class term measures that from different classes. Considering the relationship between classes, we weight the inter-class term by the non-diagonal elements in \\( A^s \\). In particular, if the class \\( i \\) is closer to \\( j \\) than \\( k \\) (i.e., \\( A_{ij}^s > A_{ik}^s \\)), \\( p_i p_j' \\) would be more important in calculating the similarity of \\( p \\) and \\( p' \\). By the above design, our proposed similarity" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.212 + ], + "angle": 0, + "content": "can more accurately express the relationship of two samples based on output space. For example, among three classes 1, 2, 3 in Figure 1, the classes 1 and 2 are closer. Given three samples \\( x_{1}, x_{2}, x_{3} \\) belonging to them with the probabilities [0.9, 0.05, 0.05], [0.05, 0.9, 0.05], and [0.05, 0.05, 0.9]. If we only use the intra-class term, all three samples have the same similarity. But for our designed similarity, \\( x_{1} \\) is closer to \\( x_{2} \\) than \\( x_{3} \\), which is more reasonable." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.213, + 0.893, + 0.499 + ], + "angle": 0, + "content": "On the basis of the proposed similarity, we further propose to perform contrastive learning in a unified form. Inspired by recent success in semi-supervised learning [67, 85, 91] and unsupervised learning [9, 18], we propose two types of contrastive losses. As shown in Figure 1, the first one is Class Relationship embedded Class-Aware Contrastive (CR-CACo) loss where the high-confident samples are enforced to be close to the corresponding prototype and away from other prototypes. Due to embedding the prior class relationship, our CR-CACo loss is more robust to label noise caused by domain shift. The second one is Class Relationship embedded Instance Discrimination Contrastive (CR-IDCo) loss where two views of the same sample are encouraged to be close and away from all other samples. Benefited from our designed accurate similarity, the CR-IDCo loss would more effectively learn discriminative features [9, 18, 75]. Actually, these two losses are complementary to each other, and their combination can achieve better performance." + }, + { + "type": "text", + "bbox": [ + 0.52, + 0.5, + 0.822, + 0.513 + ], + "angle": 0, + "content": "Our contributions are summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.52, + 0.522, + 0.892, + 0.597 + ], + "angle": 0, + "content": "- We propose to explicitly transfer class relationship for SFUDA which is more domain-invariant. And we propose to embed the class relationship into contrastive learning in order to effectively perform representation learning in the target domain." + }, + { + "type": "text", + "bbox": [ + 0.52, + 0.606, + 0.891, + 0.681 + ], + "angle": 0, + "content": "- We propose a novel class relationship embedded similarity which can more accurately express the sample relationship in the output space. Furthermore, we propose two contrastive losses (i.e., CR-CACo and CR-IDCo) exploiting our designed similarity." + }, + { + "type": "text", + "bbox": [ + 0.52, + 0.69, + 0.892, + 0.75 + ], + "angle": 0, + "content": "- We conduct extensive experiments to evaluate the proposed method, and the results validate the effectiveness of our method, which achieves the state-of-the-art performance on multiple SFUDA benchmarks." + }, + { + "type": "list", + "bbox": [ + 0.52, + 0.522, + 0.892, + 0.75 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.763, + 0.642, + 0.779 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.787, + 0.798, + 0.804 + ], + "angle": 0, + "content": "2.1. Unsupervised Domain Adaptation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.811, + 0.892, + 0.901 + ], + "angle": 0, + "content": "The conventional approaches in UDA [2, 3] are to learn domain-invariant representations, and can be classified into two coarse types. The first one decreases some distribution discrepancy metrics [49, 57, 68, 72, 84]. Another common line is adversarial training [11, 12, 45, 48, 64, 77, 96]. To increase discriminability, more recent DA methods attempt" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7620" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.258 + ], + "angle": 0, + "content": "to investigate the target domain structure. Self-training as a typical approach generates target domain pseudo labels [16, 17, 27, 28, 40, 46, 51, 54, 94, 101]. Another category is to construct prototypes [8, 34, 55, 80, 81, 93, 97, 99] or cluster centers [13, 26, 47, 70] across domains and then perform class-wise alignment. Among these methods, most of them rely on source-labeled data during target training, while pseudo label based methods can be directly applied in SFUDA. In our CR-CACo loss, pseudo label is also used and our method can alleviate the negative impact of label noise due to the usage of source class relationship." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.269, + 0.36, + 0.285 + ], + "angle": 0, + "content": "2.2. Source-free Domain Adaptation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.292, + 0.473, + 0.58 + ], + "angle": 0, + "content": "Early SFUDA methods [31, 32, 36] resort to synthesize extra training samples to get compact decision boundaries. Some recent methods [10, 21, 33, 42, 61] use pseudolabel based self-training. SHOT [42] proposes to freeze the source classifier and it clusters target features by IM loss [20] along with clustering based pseudo labeling. HCL [21] adopts feature space Instance Discrimination [78] from current and historical models, together with pseudo label learning conditioned on historical consistency. D-MCD [10] proposes denoised MCD [64] to mitigate the impact of sample selection bias and label noise. CoWA-JMD [33] uses the joint model-data structure score as sample-wise weights, along with weight mixup to exploit more target knowledge. BMD [61] proposes a class-balanced multicentric dynamic prototype strategy to obtain more accurate pseudo labels. Another line of methods (i.e. NRC [87], G-SFDA [89] and AaD [90]) propose neighborhood clustering which enforces prediction consistency between local neighbors." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.58, + 0.473, + 0.763 + ], + "angle": 0, + "content": "In addition to the above methods, others solve SFUDA from different aspects. \\( A^2 \\)Net [79] proposes to learn an additional target classifier and uses a contrastive category-wise matching module to learn compact features. DIPE [74] and VMP [25] explore transferability of source model parameters. Sub-sup [29] proposes novel subsidiary pretext tasks to assist domain adaptation. U-SFAN [62] uses the uncertainty quantified by the predictions to guide the target adaptation. Feat-mix [30] proposes image-level and feature-level mixup to enhances the discriminability-transferability tradeoff. Among these methods, few of them explicitly consider the usage of class relationship." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.776, + 0.17, + 0.792 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.801, + 0.218, + 0.817 + ], + "angle": 0, + "content": "3.1. Preliminaries" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.826, + 0.472, + 0.903 + ], + "angle": 0, + "content": "For source-free unsupervised domain adaptation (SFUDA), we are given source pre-trained model and an unlabeled target domain with \\(N_{t}\\) samples as \\(\\mathcal{D}_t = \\{x_i^t\\}_{i = 1}^{N_t}\\). Target domain have same \\(C\\) classes as source domain (for closed-set setting). The goal of SFUDA is to adapt the" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.152 + ], + "angle": 0, + "content": "model to target domain without source data. We divide the model into two parts: the feature extractor \\( f \\), and the classifier \\( g \\). The output of classifier is denoted as \\( \\pmb{p} = \\delta(g(f(x))) \\in \\mathbb{R}^C \\) where \\( \\delta \\) is the softmax function." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.153, + 0.895, + 0.336 + ], + "angle": 0, + "content": "We adopt a teacher-student framework following [18] as shown in Figure 2. The teacher feature extractor (i.e. \\(\\tilde{f}\\)) is continuously updated by exponential moving average of the student feature extractor (i.e. \\(f\\)), and the classifiers (i.e. \\(g\\), \\(\\tilde{g}\\)) are shared and frozen as in SHOT [42]. We use sample probabilities as input for contrastive loss, the query probability is obtained from the student model, and the positive and negative probabilities are obtained through the teacher model. During inference, we directly use the student model. In the following, we first introduce class relationship embedded similarity, and then elaborate on how it is used in the proposed contrastive learning process." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.348, + 0.847, + 0.365 + ], + "angle": 0, + "content": "3.2. Class Relationship Embedded Similarity" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.373, + 0.892, + 0.494 + ], + "angle": 0, + "content": "The inherent semantic and visual relationships among different classes are consistent across domains, regardless of distribution discrepancy. Inspired by this, we propose to explicitly transfer class similarities. Here we regard the normalized classifier weights \\(\\pmb{W}^{s} = [\\pmb{w}_{1}^{s},\\pmb{w}_{2}^{s},\\dots \\pmb{w}_{C}^{s}]^{T}\\in\\) \\(\\mathbb{R}^{C\\times D}\\) as class prototypes where \\(D\\) is the dimension of the last feature. Thus, the class similarity matrix \\(\\pmb{A}^{s}\\) can be obtained by \\(\\pmb{A}^{s} = \\pmb{W}^{s}(\\pmb{W}^{s})^{T}\\in \\mathbb{R}^{C\\times C}\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.495, + 0.895, + 0.691 + ], + "angle": 0, + "content": "An natural way to utilize \\( \\mathbf{A}^s \\) is to obtain target class similarity matrix \\( \\mathbf{A}^t \\) by target prototypes and then enforce consistency between \\( \\mathbf{A}^s \\) and \\( \\mathbf{A}^t \\). However, such a prototype-level constraint is coarse-grained and not strong enough to benefit the training. Instead of prototype-level, we propose a probability based sample-level similarity that considers not only the traditional intra-class term, but also the inter-class term. The latter contains probability products of all different classes, thus we can embed the class relationship matrix \\( \\mathbf{A}^s \\) into it as coefficients. Specifically, given the probabilities \\( \\mathbf{p}, \\mathbf{p}' \\in \\mathbb{R}^{C \\times 1} \\) of two samples, our Class Relationship embedded Similarity (CRS) can be presented as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.528, + 0.701, + 0.892, + 0.811 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} s _ {i n t r a} \\left(\\boldsymbol {p}, \\boldsymbol {p} ^ {\\prime}\\right) = \\sum_ {i = 1} ^ {C} p _ {i} \\times p _ {i} ^ {\\prime} \\\\ s _ {i n t e r} \\left(\\boldsymbol {p}, \\boldsymbol {p} ^ {\\prime}\\right) = \\sum_ {i = 1} ^ {C} \\sum_ {j = 1, j \\neq i} ^ {C} A _ {i, j} ^ {s} \\times p _ {i} \\times p _ {j} ^ {\\prime} \\tag {1} \\\\ s _ {c r} (\\pmb {p}, \\pmb {p} ^ {\\prime}) = s _ {i n t r a} (\\pmb {p}, \\pmb {p} ^ {\\prime}) + \\lambda_ {i n t e r} s _ {i n t e r} (\\pmb {p}, \\pmb {p} ^ {\\prime}) \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.824, + 0.892, + 0.868 + ], + "angle": 0, + "content": "where \\( s_{\\text{intra}} \\) is the traditional intra-class similarity, and \\( s_{\\text{inter}} \\) is the inter-class similarity, \\( \\lambda_{\\text{inter}} \\) is a trade-off parameter between these two terms." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.871, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Given \\(\\lambda_{inter} = 1\\), our similarity can be presented in a simpler version as follows:" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "7621" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.104, + 0.079, + 0.868, + 0.301 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.306, + 0.895, + 0.337 + ], + "angle": 0, + "content": "Figure 2. The framework of our proposed method. We use a teacher-student architecture to conduct contrastive learning. Class relationship matrix is extracted from fixed classifier and embedded in two types of contrastive loss. Best viewed in color." + }, + { + "type": "equation", + "bbox": [ + 0.093, + 0.375, + 0.469, + 0.476 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} s _ {c r} (\\boldsymbol {p}, \\boldsymbol {p} ^ {\\prime}) = \\sum_ {i = 1} ^ {C} \\sum_ {j = 1} ^ {C} A _ {i, j} \\times p _ {i} \\times p _ {j} ^ {\\prime} \\\\ = \\sum_ {i = 1} ^ {C} \\sum_ {j = 1} ^ {C} \\left(p _ {i} \\boldsymbol {w} _ {i} ^ {s}\\right) ^ {T} \\left(p _ {j} ^ {\\prime} \\boldsymbol {w} _ {j} ^ {s}\\right) = \\left(\\boldsymbol {p} ^ {T} \\boldsymbol {W} ^ {s}\\right) \\left(\\boldsymbol {p} ^ {\\prime T} \\boldsymbol {W} ^ {s}\\right) ^ {T} \\tag {2} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.477, + 0.47, + 0.524 + ], + "angle": 0, + "content": "where \\(\\pmb{p}^T\\pmb{W}^s \\in \\mathbb{R}^{1\\times D}\\) can be viewed as a new feature which is probability weighted sum of different class weights." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.534, + 0.47, + 0.551 + ], + "angle": 0, + "content": "3.3. Class Relationship Embedded Contrastive Loss" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.557, + 0.47, + 0.618 + ], + "angle": 0, + "content": "Inspired by the success of contrastive learning in different tasks [9, 18, 21, 26, 85], we propose to use our CRS for contrastive loss. Following previous work [9, 18], the contrastive loss can be presented as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.137, + 0.623, + 0.469, + 0.687 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} h (\\pmb {p}, \\pmb {p} ^ {\\prime}) = \\exp (s _ {c r} (\\pmb {p}, \\pmb {p} ^ {\\prime}) / T _ {c o}), \\\\ \\ell_ {c o} = - \\log \\frac {h \\left(\\boldsymbol {p} , \\boldsymbol {p} _ {+}\\right)}{h \\left(\\boldsymbol {p} , \\boldsymbol {p} _ {+}\\right) + \\sum_ {\\boldsymbol {p} _ {-} \\in M} h \\left(\\boldsymbol {p} , \\boldsymbol {p} _ {-}\\right)}, \\tag {3} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.694, + 0.469, + 0.77 + ], + "angle": 0, + "content": "where \\( T_{co} \\) is a temperature parameter, and \\( h \\) denotes the exponent of scaled similarity. \\( M \\) is the memory bank [18] storing probabilities of negative samples. Since our CRS is built on probability, here we use probabilities as inputs instead of original features [18] or projected features [9]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.77, + 0.47, + 0.832 + ], + "angle": 0, + "content": "With this unified form, we further instantiate it to two losses: class-aware contrastive loss and instance discrimination contrastive loss. The framework of our proposed method is shown in Figure 2." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.47, + 0.903 + ], + "angle": 0, + "content": "Class-Aware Contrastive Loss To achieve class-aware learning, it is important to select proper samples and obtain more accurate pseudo labels for them. Inspired by previous efforts in SSL [67, 85] and UDA [41, 93], we consider" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.351, + 0.895, + 0.593 + ], + "angle": 0, + "content": "constructing positive pairs from strongly and weakly augmented views of the same image. Specifically, given probabilities \\( p_s \\) and \\( \\tilde{p}_w \\) (tilde superscript denotes output by the teacher model) of strongly and weakly augmented views, we first compare \\( Max(\\tilde{p}_w) \\) with a predefined threshold \\( \\tau \\) to select high-confident samples. Due to domain shift, instead of directly using \\( argmax(p_w) \\) as pseudo label, we adopt the clustering-based pseudo-labeling method in SHOT [42]. Based on pseudo label, we can get the positive samples \\( p^* \\) with the one-hot form. For the negative samples \\( p_- \\), we use one-hot probabilities of all classes except the pseudo label class. Therefore, both positive and negative samples (i.e. \\( p^* \\) and \\( p_- \\)) are generated from the pseudo label of weak target sample through teacher model. Finally, the Class Relationship embedded Class-Aware Contrastive (CR-CACo) loss is presented as" + }, + { + "type": "equation", + "bbox": [ + 0.51, + 0.604, + 0.892, + 0.649 + ], + "angle": 0, + "content": "\\[\n\\ell_ {c r \\_ c a c o} = - \\log \\frac {h \\left(\\boldsymbol {p} _ {s} , \\boldsymbol {p} ^ {*}\\right)}{h \\left(\\boldsymbol {p} _ {s} , \\boldsymbol {p} ^ {*}\\right) + \\sum_ {\\boldsymbol {p} _ {-} \\in M _ {c a c o}} h \\left(\\boldsymbol {p} _ {s} , \\boldsymbol {p} _ {-}\\right)}, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.659, + 0.892, + 0.704 + ], + "angle": 0, + "content": "where \\( h \\) follows the definition in Equation (3), \\( M_{caco} \\) stores negative samples and is different for each sample according to its pseudo label." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.705, + 0.892, + 0.781 + ], + "angle": 0, + "content": "Regardless of pseudo label generation, using only \\( s_{intra} \\) and replacing \\( p_s \\) with features and \\( p^*, p_- \\) with classifier weights, the above Equation (4) is actually the FixMatch loss [67]. Although FixMatch can improve the performance, our CR-CACo loss can greatly outperform it." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.781, + 0.893, + 0.902 + ], + "angle": 0, + "content": "Our CR-CACo loss is more robust to label noise due to the class relationship embedded inter-class similarity. We compare our CRS with the traditional semi-supervised FixMatch loss. The results are shown in Figure 3. It can be seen that the simple IM loss [42] has lower entropy since it has an entropy minimization term. When FixMatch loss is used, the entropy will increase since the class relationship is learned through high-confidence target samples. Our CRS" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7622" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.091, + 0.086, + 0.458, + 0.276 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.28, + 0.47, + 0.322 + ], + "angle": 0, + "content": "Figure 3. Entropy density plots under SUDA setting Office-Home \\(\\mathrm{Rw} \\rightarrow \\mathrm{Cl}\\). Our class relationship embedded similarity \\(s_{cr}\\) has higher entropy but better accuracy. Best viewed in color." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.338, + 0.471, + 0.385 + ], + "angle": 0, + "content": "has larger entropy since we explicitly consider the class relationship and can alleviate the negative impact of label noise." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.403, + 0.47, + 0.629 + ], + "angle": 0, + "content": "Instance Discrimination Contrastive Loss The instance discrimination contrastive (IDCo) loss has shown remarkable success in self-supervised learning [9, 18]. It can learn discriminative features by regrading the two views of the same sample as positive pairs and all other samples as negative samples. The traditional IDCo loss uses features as input, but it only makes the feature discriminative and cannot guarantee that the learned feature will be close to corresponding prototypes (i.e. classifier weights). Conducting IDCo loss in the output space (i.e. computing similarity using \\( s_{\\text{intra}} \\)) can alleviate this problem [35], but the more informative relationship between samples in the original feature space can not be fully expressed. This can be addressed by our CRS since we embed the source prior class similarity and can reflect more accurate relationship of samples." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.631, + 0.47, + 0.722 + ], + "angle": 0, + "content": "Here we use the probabilities \\( \\pmb{p}_{s1} \\) and \\( \\tilde{\\pmb{p}}_{s2}' \\) of two strong augmented views from the same image as positive pairs, and probabilities of all other samples in the same batch as negative samples. Finally, the Class Relationship embedded Instance Discrimination Contrastive (CR-IDCo) loss is presented as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.088, + 0.734, + 0.468, + 0.791 + ], + "angle": 0, + "content": "\\[\n\\ell_ {c r. i d c o} = - \\log \\frac {h \\left(\\boldsymbol {p} _ {s 1} , \\tilde {\\boldsymbol {p}} _ {s 2} ^ {\\prime}\\right)}{h \\left(\\boldsymbol {p} _ {s 1} , \\tilde {\\boldsymbol {p}} _ {s 2} ^ {\\prime}\\right) + \\sum_ {\\tilde {\\boldsymbol {p}} _ {-} \\in M _ {i d c o}} h \\left(\\boldsymbol {p} _ {s 1} , \\tilde {\\boldsymbol {p}} _ {-}\\right)}, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.792, + 0.47, + 0.837 + ], + "angle": 0, + "content": "where \\( h \\) follows the definition in Equation (3), \\( M_{idco} \\) contains the probabilities of other samples within the same batch." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.847, + 0.321, + 0.864 + ], + "angle": 0, + "content": "3.4. Overall Training Objective" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.871, + 0.471, + 0.9 + ], + "angle": 0, + "content": "The proposed two losses are complementary to each other and can be trained end-to-end. Since our CR-CACo" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.895, + 0.259 + ], + "angle": 0, + "content": "uses a high prob-threshold (i.e. 0.95) to select confident samples, it should be combined with other methods that can generate confident predictions. Otherwise, it will be dominated by noise labels in the early training stage. Although it performs normally when combined with our CR-IDCo loss, both of them focus on transferring source class relationship, and other methods that encourage confident predictions are still needed and complementary to ours. As a result, we combine our method with existing representative methods (i.e. SHOT-IM [42], AaD [90]) to validate the effectiveness. As a result, the overall loss is presented as:" + }, + { + "type": "equation", + "bbox": [ + 0.525, + 0.271, + 0.892, + 0.288 + ], + "angle": 0, + "content": "\\[\n\\ell_ {\\text {o u r s}} = \\ell_ {\\text {b a s e l i n e}} + \\lambda_ {\\text {c a c o}} \\ell_ {\\text {c r} - \\text {c a c o}} + \\lambda_ {\\text {i d c o}} \\ell_ {\\text {c r} - \\text {i d c o}} \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.3, + 0.895, + 0.391 + ], + "angle": 0, + "content": "where \\(\\lambda_{caco},\\lambda_{idco}\\) are the trade-off hyperparameters. \\(l_{baseline}\\) is the loss of baseline (i.e., IM or AaD). It contains two terms: discriminability and diversity. The former produces high-confident outputs, and the latter avoids collapse of predictions. IM uses entropy of single sample and average of all samples, and AaD uses neighbor information." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.404, + 0.634, + 0.422 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.43, + 0.691, + 0.447 + ], + "angle": 0, + "content": "4.1. Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.453, + 0.895, + 0.665 + ], + "angle": 0, + "content": "Datasets. We evaluate the effectiveness of our approach on four standard DA benchmarks. Office-31 [63] benchmark consists of three domains in office environments: Amazon (A), DSLR (D), and Webcam (W), each with 31 categories. Office-Home [73] is a more challenging dataset. It comprises of images of commonplace objects divided into four domains: Artistic (Ar), Clipart (Cl), Product (Pr), and Real-World (Rw), each with 65 classes. VisDA [58] is a large-scale dataset for synthetic-to-real domain adaptation. The source domain has 152,397 synthetic images, while the target domain has 55,388 real-world images. DomainNet [57] is the most challenging dataset involving 6 domains: Clipart (C), Real (R), Infograph (I), Painting (P), Sketch (S) and Quickdraw (Q) with 345 classes each." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.674, + 0.895, + 0.84 + ], + "angle": 0, + "content": "Implementation details. To ensure fair comparison with related methods, we adopt the backbone of a ResNet-50 [19] for Office-31, Office-Home, and DomainNet, and a ResNet-101 for VisDA. We use the same network architecture as SHOT [42] where the final part of the network is changed to suit the SFUDA task. For the classifier, we find that fixing it or not have little impact on the final performance, and we choose to fix it in all experiments. We adopt SGD with momentum 0.9 and batch size 64 for all datasets where each image has three views (i.e. one weak and two strong augmentations)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.841, + 0.894, + 0.901 + ], + "angle": 0, + "content": "For the hyperparameters, we empirically set the probability threshold \\(\\tau\\) to 0.95 following FixMatch [67]. We find that \\(\\tau = 0.95\\) works well across all settings and tasks. For the temperature \\(T_{co}\\) in both contrastive losses," + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7623" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.28, + 0.086, + 0.69, + 0.1 + ], + "angle": 0, + "content": "Table 1. Single-Source Unsupervised DA (SUDA) on Office-Home." + }, + { + "type": "table", + "bbox": [ + 0.078, + 0.103, + 0.895, + 0.362 + ], + "angle": 0, + "content": "
MethodSFOffice-Home
Ar→ClAr→PrAr→RwCl→ArCl→PrCl→RwPr→ArPr→ClPr→RwRw→ArRw→ClRw→PrAvg
SENTRY (ICCV'21) [59]X61.877.480.166.371.674.766.863.080.974.066.384.172.2
FixBi (CVPR'21) [52]X58.177.380.467.779.578.165.857.981.776.462.986.772.7
SCDA (ICCV'21) [38]X60.776.482.869.877.578.468.959.082.774.961.884.573.1
A²Net (ICCV'21) [79]✓58.479.082.467.579.378.968.056.282.974.160.585.072.8
NRC (NeurIPS'21) [87]✓57.780.382.068.179.878.665.356.483.071.058.685.672.2
D-MCD (AAAI'22) [10]✓59.478.980.267.279.378.665.355.682.273.362.883.972.2
DIPE (CVPR'22) [74]✓56.579.280.770.179.878.867.955.183.574.159.384.872.5
Sub-Sup (ECCV'22) [29]✓61.080.482.569.179.979.569.157.882.774.565.186.474.0
BMD (ECCV'22) [61]✓58.179.782.669.381.080.770.857.683.674.060.085.973.6
U-SFAN (ECCV'22) [62]✓57.877.881.667.977.379.267.254.781.273.360.383.971.9
CoWA-JMDS (ICML'22) [33]✓56.978.481.069.180.079.967.757.282.472.860.584.572.5
Feat-Mixup (ICML'22) [30]✓61.881.283.068.580.679.467.861.585.173.764.186.574.5
VMP (NeurIPS'22) [25]✓57.977.682.568.679.480.668.455.683.175.259.684.772.8
DaC (NeurIPS'22) [98]✓59.179.581.269.378.979.267.456.482.474.061.484.472.8
SHOT-IM (ICML'20) [42]✓55.476.680.466.974.375.465.654.880.773.758.483.470.5
Ours+ SHOT-IM✓62.882.084.370.980.882.670.061.183.676.265.187.075.5 (+5.0)
AaD (NeurIPS'22) [90]✓59.379.382.168.979.879.567.257.483.172.158.585.472.7
Ours+ AaD✓63.582.185.073.082.782.469.562.982.674.265.787.375.9 (+3.2)
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.375, + 0.47, + 0.404 + ], + "angle": 0, + "content": "Table 2. Single-Source Unsupervised Domain Adaptation (SUDA) on Office-31 and VisDA benchmarks." + }, + { + "type": "table", + "bbox": [ + 0.08, + 0.414, + 0.47, + 0.673 + ], + "angle": 0, + "content": "
MethodSFOffice-31VisDA
A→DA→WD→WW→DD→AW→A AvgS→R
BCDM (AAAI'20) [37]✗93.895.498.6100.073.173.089.083.4
MCC (ECCV'20) [24]✗94.495.598.6100.072.974.989.478.8
FixBi (CVPR'21) [52]✗95.096.199.3100.078.779.491.487.2
RADA (ICCV'21) [23]✗96.196.299.3100.077.577.491.176.3
FAA (ICCV'21) [22]✗94.492.399.299.780.578.790.882.7
SCDA (ICCV'21) [38]✗95.495.399.0100.077.275.990.5-
SHOT (ICML'20) [42]✓94.090.198.499.974.774.388.682.9
CPGA (IJCAI'21) [60]✓94.494.198.499.876.076.689.984.1
VDM-DA (TCSVT'21) [71]✓93.294.198.0100.075.877.189.785.1
A2Net (ICCV'21) [79]✓94.594.099.2100.076.776.190.184.3
HCL (NeurIPS'21) [21]✓90.891.398.2100.072.772.787.683.5
NRC (NeurIPS'21) [87]✓96.090.899.0100.075.375.089.485.9
SHOT++ (TPAMI'21) [43]✓94.390.498.799.976.275.889.287.3
D-MCD (AAAI'22) [10]✓94.193.598.8100.076.476.489.987.5
DIPE (CVPR'22) [74]✓96.693.198.499.675.577.290.183.1
Sub-Sup (ECCV'22) [29]✓95.694.699.299.877.077.790.788.2
BMD (ECCV'22) [61]✓96.294.298.0100.076.076.090.188.7
CoWA-JMDS (ICML'22) [33]✓94.495.298.5100.076.277.690.386.9
Feat-Mixup (ICML'22) [30]✓94.693.298.9100.078.378.990.787.8
VMP (NeurIPS'22) [25]✓93.396.298.6100.075.476.990.0-
SHOT-IM (ICML'20) [42]✓90.691.298.399.972.571.487.380.4
Ours + SHOT-IM✓95.895.199.0100.076.678.390.8(+3.5)89.1(+8.7)
AaD (NeurIPS'22) [90]✓96.492.199.1100.075.076.589.988.0
Ours + AaD✓96.695.599.1100.076.978.391.1(+1.2)89.6(+1.6)
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.697, + 0.47, + 0.775 + ], + "angle": 0, + "content": "we set it to 0.07 following previous contrastive learning methods [18, 21, 85]. For the \\(\\lambda_{inter}\\) in our similarity, we use 1.0 for simplicity. For the trade-off hyperparameters \\(\\lambda_{caco}, \\lambda_{idco}\\), we directly set them to 1.0. More training details can be found in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.797, + 0.172, + 0.812 + ], + "angle": 0, + "content": "4.2. Results" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.471, + 0.903 + ], + "angle": 0, + "content": "The results are shown in Table 1, Table 2, Table 3 and Table 4. In each table, SF indicates source-free, * indicates results we reproduced from the released code. (+x.x) indicates gains over SHOT-IM [42] and AaD [90] respectively. Ours means combination of our proposed two losses." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.384, + 0.892, + 0.52 + ], + "angle": 0, + "content": "Single Source Domain Adaptation Table 1 and Table 2 show the classification accuracy for single-source DA on each dataset: Office-Home, Office-31, and VisDA. Based on the simple SHOT-IM and stronger AaD, our method can achieve state-of-the-art performance. UDA methods BCDM [37] and MCC [24] also explore the inter-class relationship by directly suppressing the inter-class term to obtain determined outputs. Our method surpasses these methods even without source domain supervision." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.521, + 0.893, + 0.673 + ], + "angle": 0, + "content": "The two proposed contrastive losses exploit different views of the same image. Thus our method is complementary to those not utilizing information of different views. Both baselines only consider weak views, and so our method can boost them. AaD surpasses IM by regarding local neighbors as positive pairs. In our CR-IDCo loss, the strong views of the same image can be regarded as neighbors since they are very close in feature space, which has overlapped effect with AaD. When our method is combined with AaD, therefore, the improvement is relatively smaller." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.689, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Multi Source Domain Adaptation Table 3 shows the results for multi-source DA on Office-Home and the largescale DomainNet benchmark. We adopt two baselines without using domain labels, that is, all data from different source domains is treated as belonging to a single domain and only one source model is trained and transferred. We observe improvements of \\(4.1\\%\\) and \\(3.4\\%\\) in DomainNet, \\(4.5\\%\\) and \\(3.0\\%\\) in Office-Home. CAiDA [14] considers class relationship by enforcing consistency of different source model outputs. It achieves \\(76.2\\%\\) on Office-Home by combining IM loss, a novel confident-anchor-induced pseudo label method and the class-relation-space-aware consistency loss. Based on the same IM loss, ours achieves \\(78.1\\%\\) and exceeds it by \\(1.9\\%\\)." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7624" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.178, + 0.086, + 0.791, + 0.1 + ], + "angle": 0, + "content": "Table 3. Multi-Source Unsupervised Domain Adaptation (MUDA) on DomainNet and Office-Home." + }, + { + "type": "table", + "bbox": [ + 0.079, + 0.101, + 0.891, + 0.276 + ], + "angle": 0, + "content": "
MethodSFw/o Domain \nLabelsDomainNetOffice-Home
→C→I→P→Q→R→SAvg→Ar→Cl→Pr→RwAvg
MCC (ECCV'20) [24]XX65.526.056.616.568.052.747.6-----
CMSDA (BMVC'21) [66]XX70.926.557.521.368.159.450.471.567.784.182.976.6
DRT (CVPR'21) [39]XX71.031.661.012.371.460.751.3-----
STEM (ICCV'21) [53]XX72.028.261.525.772.660.253.4-----
SourcecombineX✓57.023.454.114.667.250.344.458.057.374.277.966.9
DECISION (CVPR'21) [1]✓X61.521.654.618.967.551.045.974.559.484.483.675.5
CAiDA (NeurIPS'21) [14]✓X-------75.260.584.784.276.2
Sub-Sup (ECCV'22) [29]✓✓70.325.757.317.169.957.149.675.164.186.684.477.6
Feat-Mixup (ICML'22) [30]✓✓75.424.657.823.665.858.551.072.667.485.983.677.4
SHOT-IM* (ICML'20) [42]✓✓64.524.255.215.567.053.246.671.558.681.982.373.6
Ours + SHOT-IM✓✓69.827.059.021.270.257.550.7(+4.1)74.966.485.185.878.1 (+4.5)
AaD* (NeurIPS'22) [90]✓✓65.725.556.417.068.154.647.972.160.685.384.475.6
Ours +AaD✓✓71.227.859.322.569.357.851.3(+3.4)75.466.986.485.978.6 (+3.0)
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.292, + 0.47, + 0.319 + ], + "angle": 0, + "content": "Table 4. Partial-set and open-set Domain Adaptation (PDA and ODA) on Office-Home." + }, + { + "type": "table", + "bbox": [ + 0.08, + 0.324, + 0.465, + 0.46 + ], + "angle": 0, + "content": "
Partial-set DASF Avg.Open-set DASF Avg.
ResNet-50 [19]× 61.3ResNet [19]× 65.3
IWAN (CVPR'18) [92]× 63.6OpenMax (CVPR'16) [4]× 66.7
SAN (CVPR'18) [5]× 65.3ATI-λ (ICCV'17) [56]× 66.1
ETN (CVPR'19) [6]× 70.5OSBP (ECCV'18) [65]× 65.7
SAFN (ICCV'19) [82]× 71.8STA (CVPR'19) [44]× 69.5
Source model only✓ 62.8Source model only✓ 46.6
SHOT (ICML'20) [42]✓ 79.3SHOT (ICML'20) [42]✓ 72.8
SHOT+HCL (NeurIPS'21) [21]✓ 80.1SHOT+HCL (NeurIPS'21) [21]✓ 73.2
CoWA-JMDS (ICML'22) [33]✓ 83.2CoWA-JMDS (ICML'22) [33]✓ 73.2
SHOT-IM (ICML'20) [42]✓ 76.8SHOT-IM (ICML'20) [42]✓ 71.5
Ours +SHOT-IM✓ 80.6 (+3.8)Ours +SHOT-IM✓ 73.2(+1.7)
AaD* (NeurIPS'22) [90]✓ 79.7AaD* (NeurIPS'22) [90]✓ 71.8
Ours +AaD✓ 82.4 (+2.7)Ours +AaD✓ 73.3(+1.5)
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.473, + 0.47, + 0.581 + ], + "angle": 0, + "content": "Domain Adaptation Beyond Vanilla Closed-set We provide additional results under source-free partial-set and open-set DA (PDA and ODA) setting on Office-Home. For open-set detection in ODA, we follow the same protocol to detect unseen categories as SHOT [42]. Results are shown in Table 4. It can be seen that our method can achieve performance comparable to that of the state-of-the-arts." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.588, + 0.182, + 0.604 + ], + "angle": 0, + "content": "4.3. Analysis" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.611, + 0.471, + 0.807 + ], + "angle": 0, + "content": "Component-wise ablations Here we validate the effectiveness of different proposed losses. As described in Sec. 3.4, we do not use CR-CACo loss alone. The results are shown in Table 5. It can be seen that when using our method alone (line #2) can achieve comparable performance. When combined with SHOT-IM, our both losses are complementary, and the best performance can be achieved by combining baseline and our losses. Note that although CR-IDCo (line #1) outperforms IM (line #0), it is worse than IM when combined with CR-CACo (i.e. line #2 < line #3). This is caused by the fact that CR-IDCo cannot produce enough confident samples as IM for the learning of CR-CACo loss." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.47, + 0.903 + ], + "angle": 0, + "content": "Effect of pseudo labels in CR-CACo In our proposed CR-CACo loss, the pseudo label is generated from clustering-based method in SHOT [42] instead of predictions of weakly augmented images. The results of the comparison are shown in Table 6. It can be seen that our" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.292, + 0.892, + 0.321 + ], + "angle": 0, + "content": "Table 5. Component-wise ablation studies of the proposed methods under the SUDA setting." + }, + { + "type": "table", + "bbox": [ + 0.504, + 0.33, + 0.89, + 0.424 + ], + "angle": 0, + "content": "
#IMCR-CACoCR-IDCoOffice-31Office-HomeVisDA
0✓87.370.580.4
1✓89.272.583.0
2✓✓89.674.085.3
3✓✓89.874.887.0
4✓✓90.274.385.8
5✓✓✓90.875.589.1
" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.45, + 0.892, + 0.496 + ], + "angle": 0, + "content": "CR-CACo can bring significant improvement when using pseudo label from model prediction. The clustering based pseudo-label can bring slight but consistent improvements." + }, + { + "type": "table_caption", + "bbox": [ + 0.529, + 0.506, + 0.862, + 0.521 + ], + "angle": 0, + "content": "Table 6. Comparison of different pseudo label methods." + }, + { + "type": "table", + "bbox": [ + 0.504, + 0.524, + 0.89, + 0.637 + ], + "angle": 0, + "content": "
#BaselinePseudo-label in CR-CACoOffice-31Office-HomeVisDA
PredictionClustering
087.370.580.4
1IM✓89.6(+2.3)74.5(+4.0)86.6(+6.2)
2✓89.8(+2.5)74.8(+4.3)87.0(+6.6)
389.972.788.0
4AaD✓90.6(+0.7)75.0(+2.3)88.9(+0.9)
5✓90.7(+0.8)75.2(+2.5)89.1(+1.1)
" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.664, + 0.892, + 0.906 + ], + "angle": 0, + "content": "Different similarities and disentangled improvements. To validate the superiority of our class relationship embedded similarity in contrastive loss, we compare it with other two choices. The first one uses only \\( s_{intra} \\) based on features, and the second one uses only \\( s_{intra} \\) based on probabilities which equals \\( \\lambda_{inter} = 0 \\) in our case. The results are shown in Table 7. Using features (i.e. i.e. \\( s_{intra} \\) w/ feature) rather than probabilities is a more natural way. Indeed, using features in CACo Loss would result in FixMatch loss, and using features in IDCo loss would fall in the traditional self-supervised way (e.g., MoCo). Only using the probability based intra-class term (i.e. \\( s_{intra} \\) w/ probability) is still special for similarity in contrastive losses. And it can bring consistent improvements. In both CACo and IDCo losses, our \\( s_{cr} \\) can achieve the best result, thus validating the effectiveness of adopting source prior class similarities." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7625" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.131, + 0.09, + 0.414, + 0.103 + ], + "angle": 0, + "content": "Table 7. Comparison with different similarities." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.108, + 0.467, + 0.218 + ], + "angle": 0, + "content": "
#LossSimilarityOffice-31Office-HomeVisDA
0IM—87.370.580.4
1IM+CACosintra w/ feature88.972.583.5
2sintra w/ probability89.172.984.0
3scr w/ probability89.874.887.0
4IM+IDCosintra w/ feature87.570.980.8
5sintra w/ probability88.271.481.5
6scr w/ probability90.274.385.8
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.227, + 0.47, + 0.439 + ], + "angle": 0, + "content": "Comparison with fixed coefficient within \\( s_{inter} \\) In the definition of inter-class term in Equation (1), we use the source class similarity matrix \\( A^s \\) as coefficient. Instead of this prior knowledge, we can also set the coefficient to a small fixed value. Considering the positive pairs in contrastive loss, the negative coefficient (i.e. -0.1) works to suppress the inter-class term, thus obtaining more determined predictions as in [24, 37] while the positive value (i.e. 0.1) is more like a smooth regularization as KLD loss in [100]. The results are shown in Figure 4. It can be seen that both positive and negative values can bring some improvements. However, our CR-CACo loss can achieve better performance, which validates the importance of source prior class relationship." + }, + { + "type": "image", + "bbox": [ + 0.089, + 0.449, + 0.269, + 0.549 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.274, + 0.449, + 0.455, + 0.548 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.555, + 0.47, + 0.611 + ], + "angle": 0, + "content": "Figure 4. Using fixed coefficients in \\( s_{inter} \\) in both CACo and IDCo losses under SUDA setting Office-Home Rw→Cl. Black dotted line shows the performance of baseline IM. Red dotted line shows the performance of our losses." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.621, + 0.47, + 0.697 + ], + "angle": 0, + "content": "Parameter sensitivity Here we analyze the hyperparameter sensitivity under the setting of SUDA Office-Home with SHOT-IM baseline. We consider four hyperparameters, as shown in Figure 5. It is evident that they are stable within specific ranges." + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.71, + 0.179, + 0.769 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.182, + 0.711, + 0.272, + 0.77 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.274, + 0.711, + 0.373, + 0.77 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.374, + 0.711, + 0.462, + 0.77 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.777, + 0.468, + 0.804 + ], + "angle": 0, + "content": "Figure 5. Sensitivity analysis of different hyperparameters. The first two use our two losses. The last two use corresponding loss." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.468, + 0.9 + ], + "angle": 0, + "content": "Visualization of features We visualize the target features and classifier weights in Figure 6. It can be seen that our method has less misclassified samples for classes 0, 1, 7 compared with IM. Thus our method can learn more discriminative features." + }, + { + "type": "image", + "bbox": [ + 0.51, + 0.089, + 0.635, + 0.182 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.516, + 0.185, + 0.625, + 0.194 + ], + "angle": 0, + "content": "Source Pre-trained Model" + }, + { + "type": "image", + "bbox": [ + 0.636, + 0.089, + 0.761, + 0.182 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.692, + 0.185, + 0.708, + 0.194 + ], + "angle": 0, + "content": "IM" + }, + { + "type": "image", + "bbox": [ + 0.764, + 0.09, + 0.886, + 0.182 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.811, + 0.185, + 0.853, + 0.194 + ], + "angle": 0, + "content": "IM+Ours" + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.206, + 0.892, + 0.289 + ], + "angle": 0, + "content": "Figure 6. The t-SNE visualization of classifier weights (i.e. star) and target domain features (i.e. dot) of different methods under the setting of SUDA Office-Home \\(\\mathrm{Rw} \\rightarrow \\mathrm{Cl}\\). We directly use the first 10 classes (i.e. different colors), and the coordinates of classifier weights are the same across different methods since source classifier is frozen. Best viewed in color." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.313, + 0.892, + 0.435 + ], + "angle": 0, + "content": "Transformer backbone We further apply our method with Transformer-based backbone under the SFUDA setting, as shown in Table 8. We choose the same network (i.e. ViT-B [15]) following SSRT [69]. For the training of the source model, we follow SHOT [42] which obtains better performance than original ViT-B (78.2% vs. 75.5%). Adding IM loss achieves 84.0%. Based on IM loss, ours achieves 88.3% which significantly outperforms SSRT." + }, + { + "type": "table_caption", + "bbox": [ + 0.5, + 0.446, + 0.891, + 0.473 + ], + "angle": 0, + "content": "Table 8. Single-Source Unsupervised DA (SUDA) on Office-Home with ViT-B backbone." + }, + { + "type": "table", + "bbox": [ + 0.504, + 0.485, + 0.885, + 0.558 + ], + "angle": 0, + "content": "
MethodSFAvg.MethodSFAvg.
ViT-B✗75.5CDTrans (ICLR'22) [83]✗80.5
TVT (WACV'23) [86]✗83.6DOT-B (ACMMM'22) [50]✗84.3
SSRT (CVPR'22) [69]✗85.4BCAT-DTF (Arxiv'22) [76]✗86.6
Source model✓78.2SHOT-IM✓84.0
SHOT-IM+CR-CACo✓87.0(+3.0)SHOT-IM+Ours✓88.3(+4.3)
" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.566, + 0.619, + 0.582 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.592, + 0.892, + 0.773 + ], + "angle": 0, + "content": "In this paper, we propose to explicitly transfer the class relationship for SFUDA which is more domain-invariant. We propose a novel class relationship embedded similarity that can more accurately express the sample relationship in the output space. Furthermore, we propose two contrastive losses (i.e., CR-CACo and CR-IDCo) that exploit our designed similarity. These two losses are complementary, and their combination can better explore the target distribution. We combine our method with existing representative baselines in multiple SFUDA settings. Extensive experiments show the effectiveness of the proposed method, which achieves state-of-the-art performance." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.785, + 0.687, + 0.803 + ], + "angle": 0, + "content": "6. Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.81, + 0.892, + 0.901 + ], + "angle": 0, + "content": "This work is supported by the National Natural Science Foundation of China under Grant 62176246 and Grant 61836008. This work is also supported by Anhui Provincial Natural Science Foundation 2208085UD17 and the Fundamental Research Funds for the Central Universities (WK3490000006)." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7626" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.079, + 0.09, + 0.176, + 0.107 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.116, + 0.47, + 0.17 + ], + "angle": 0, + "content": "[1] Sk. Miraj Ahmed, Dripta S. Raychaudhuri, S. Paul, Samet Oymak, and Amit K. Roy-Chowdhury. Unsupervised multi-source domain adaptation without access to source data. In CVPR, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.173, + 0.472, + 0.227 + ], + "angle": 0, + "content": "[2] Shai Ben-David, John Blitzer, Koby Crammer, Alex Kulesza, Fernando Pereira, and Jennifer Wortman Vaughan. A theory of learning from different domains. Machine learning, 79(1-2):151-175, 2010. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.23, + 0.47, + 0.271 + ], + "angle": 0, + "content": "[3] Shai Ben-David, John Blitzer, Koby Crammer, and Fernando Pereira. Analysis of representations for domain adaptation. In NeurIPS, 2006. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.273, + 0.47, + 0.3 + ], + "angle": 0, + "content": "[4] Abhijit Bendale and Terrance E Boult. Towards open set deep networks. In CVPR, 2016. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.302, + 0.47, + 0.343 + ], + "angle": 0, + "content": "[5] Zhangjie Cao, Mingsheng Long, Jianmin Wang, and Michael I Jordan. Partial transfer learning with selective adversarial networks. In CVPR, 2018. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.345, + 0.47, + 0.386 + ], + "angle": 0, + "content": "[6] Zhangjie Cao, Kaichao You, Mingsheng Long, Jianmin Wang, and Qiang Yang. Learning to transfer examples for partial domain adaptation. In CVPR, 2019. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.388, + 0.47, + 0.442 + ], + "angle": 0, + "content": "[7] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In ECCV, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.445, + 0.469, + 0.5 + ], + "angle": 0, + "content": "[8] Chaoqi Chen, Weiping Xie, Tingyang Xu, Wenbing Huang, Yu Rong, Xinghao Ding, Yue Huang, and Junzhou Huang. Progressive feature alignment for unsupervised domain adaptation. In CVPR, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.502, + 0.469, + 0.543 + ], + "angle": 0, + "content": "[9] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In ICML, 2020. 2, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.546, + 0.469, + 0.599 + ], + "angle": 0, + "content": "[10] Tong Chu, Yahao Liu, Jinhong Deng, Wen Li, and Lixin Duan. Denoised maximum classifier discrepancy for sourcefree unsupervised domain adaptation. In AAAI, 2022. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.603, + 0.469, + 0.642 + ], + "angle": 0, + "content": "[11] Safa Cicek and Stefano Soatto. Unsupervised domain adaptation via regularized conditional alignment. In ICCV, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.645, + 0.469, + 0.686 + ], + "angle": 0, + "content": "[12] Shuhao Cui, Shuhui Wang, Junbao Zhuo, Chi Su, Qingming Huang, and Qi Tian. Gradually vanishing bridge for adversarial domain adaptation. In CVPR, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.688, + 0.469, + 0.729 + ], + "angle": 0, + "content": "[13] Zhijie Deng, Yucen Luo, and Jun Zhu. Cluster alignment with a teacher for unsupervised domain adaptation. In ICCV, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.732, + 0.469, + 0.772 + ], + "angle": 0, + "content": "[14] Jiahua Dong, Zhen Fang, Anjin Liu, Gan Sun, and Tongliang Liu. Confident anchor-induced multi-source free domain adaptation. In NeurIPS, 2021. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.775, + 0.469, + 0.855 + ], + "angle": 0, + "content": "[15] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2020. 1, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.859, + 0.469, + 0.899 + ], + "angle": 0, + "content": "[16] Geoffrey French, Michal Mackiewicz, and Mark Fisher. Self-ensembling for visual domain adaptation. In ICLR, 2018. 3" + }, + { + "type": "list", + "bbox": [ + 0.088, + 0.116, + 0.472, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.093, + 0.892, + 0.134 + ], + "angle": 0, + "content": "[17] Yuan Gao, Zilei Wang, Jiafan Zhuang, Yixin Zhang, and Junjie Li. Exploit domain-robust optical flow in domain adaptive video semantic segmentation. In AAAI, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.136, + 0.892, + 0.175 + ], + "angle": 0, + "content": "[18] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In CVPR, 2020. 2, 3, 4, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.178, + 0.892, + 0.217 + ], + "angle": 0, + "content": "[19] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, 2016. 1, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.219, + 0.892, + 0.273 + ], + "angle": 0, + "content": "[20] Weihua Hu, Takeru Miyato, Seiya Tokui, Eiichi Matsumoto, and Masashi Sugiyama. Learning discrete representations via information maximizing self-augmented training. In ICML, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.275, + 0.892, + 0.328 + ], + "angle": 0, + "content": "[21] Jiaxing Huang, Dayan Guan, Aoran Xiao, and Shijian Lu. Model adaptation: Historical contrastive learning for unsupervised domain adaptation without source data. In NeurIPS, 2021. 1, 2, 3, 4, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.33, + 0.892, + 0.37 + ], + "angle": 0, + "content": "[22] Jiaxing Huang, Dayan Guan, Aoran Xiao, and Shijian Lu. RDA: Robust domain adaptation via fourier adversarial attacking. In ICCV, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.372, + 0.892, + 0.412 + ], + "angle": 0, + "content": "[23] Xin Jin, Cuiling Lan, Wenjun Zeng, and Zhibo Chen. Re-energizing domain discriminator with sample relabeling for adversarial domain adaptation. In ICCV, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.414, + 0.892, + 0.453 + ], + "angle": 0, + "content": "[24] Ying Jin, Ximei Wang, Mingsheng Long, and Jianmin Wang. Minimum class confusion for versatile domain adaptation. In ECCV, 2020. 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.456, + 0.892, + 0.496 + ], + "angle": 0, + "content": "[25] Mengmeng Jing, Xiantong Zhen, Jingjing Li, and Cees GM Snoek. Variational model perturbation for source-free domain adaptation. In NeurIPS, 2022. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.498, + 0.892, + 0.537 + ], + "angle": 0, + "content": "[26] Guoliang Kang, Lu Jiang, Yi Yang, and Alexander G Hauptmann. Contrastive adaptation network for unsupervised domain adaptation. In CVPR, 2019. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.539, + 0.892, + 0.593 + ], + "angle": 0, + "content": "[27] Seunghyeon Kim, Jaehoon Choi, Taekyung Kim, and Changick Kim. Self-training and adversarial background regularization for unsupervised domain adaptive one-stage object detection. In CVPR, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.596, + 0.892, + 0.635 + ], + "angle": 0, + "content": "[28] Ananya Kumar, Tengyu Ma, and Percy Liang. Understanding self-training for gradual domain adaptation. In ICML, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.637, + 0.892, + 0.691 + ], + "angle": 0, + "content": "[29] Jogendra Nath Kundu, Suvaansh Bhambri, Akshay Kulkarni, Hiran Sarkar, Varun Jampani, and R Venkatesh Babu. Concurrent subsidiary supervision for unsupervised source-free domain adaptation. In ECCV, 2022. 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.693, + 0.892, + 0.76 + ], + "angle": 0, + "content": "[30] Jogendra Nath Kundu, Akshay R Kulkarni, Suvaansh Bhambri, Deepesh Mehta, Shreyas Anand Kulkarni, Varun Jampani, and Venkatesh Babu Radhakrishnan. Balancing discriminability and transferability for source-free domain adaptation. In ICML, 2022. 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.763, + 0.892, + 0.801 + ], + "angle": 0, + "content": "[31] Jogendra Nath Kundu, Naveen Venkat, and R Venkatesh Babu. Universal source-free domain adaptation. In CVPR, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.804, + 0.892, + 0.857 + ], + "angle": 0, + "content": "[32] Jogendra Nath Kundu, Naveen Venkat, Ambareesh Revanur, Rahul M V, and R. Venkatesh Babu. Towards inheritable models for open-set domain adaptation. In CVPR, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.859, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[33] Jonghyun Lee, Dahuin Jung, Junho Yim, and Sungroh Yoon. Confidence score for source-free unsupervised domain adaptation. In ICML, 2022. 3, 6, 7" + }, + { + "type": "list", + "bbox": [ + 0.511, + 0.093, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7627" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.092, + 0.47, + 0.146 + ], + "angle": 0, + "content": "[34] Junjie Li, Zilei Wang, Yuan Gao, and Xiaoming Hu. Exploring high-quality target domain information for unsupervised domain adaptive semantic segmentation. In ACMMM, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.148, + 0.47, + 0.202 + ], + "angle": 0, + "content": "[35] Junjie Li, Yixin Zhang, Zilei Wang, and Keyu Tu. Semantic-aware representation learning via probability contrastive loss. arXiv preprint arXiv:2111.06021, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.204, + 0.471, + 0.245 + ], + "angle": 0, + "content": "[36] Rui Li, Qianfen Jiao, Wenming Cao, Hau-San Wong, and Si Wu. Model adaptation: Unsupervised domain adaptation without source data. In CVPR, 2020. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.247, + 0.47, + 0.299 + ], + "angle": 0, + "content": "[37] Shuang Li, Fangrui Lv, Binhui Xie, Chi Harold Liu, Jian Liang, and Chen Qin. Bi-classifier determinacy maximization for unsupervised domain adaptation. In AAAI, 2021. 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.301, + 0.47, + 0.342 + ], + "angle": 0, + "content": "[38] Shuang Li, Mixue Xie, Fangrui Lv, Chi Harold Liu, Jian Liang, Chen Qin, and Wei Li. Semantic concentration for domain adaptation. In ICCV, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.343, + 0.47, + 0.384 + ], + "angle": 0, + "content": "[39] Yunsheng Li, Lu Yuan, Yinpeng Chen, Pei Wang, and N. Vasconcelos. Dynamic transfer for multi-source domain adaptation. In CVPR, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.385, + 0.47, + 0.426 + ], + "angle": 0, + "content": "[40] Yunsheng Li, Lu Yuan, and Nuno Vasconcelos. Bidirectional learning for domain adaptation of semantic segmentation. In CVPR, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.427, + 0.47, + 0.481 + ], + "angle": 0, + "content": "[41] Yu-Jhe Li, Xiaoliang Dai, Chih-Yao Ma, Yen-Cheng Liu, Kan Chen, Bichen Wu, Zijian He, Kris Kitani, and Peter Vajda. Cross-domain adaptive teacher for object detection. In CVPR, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.483, + 0.47, + 0.537 + ], + "angle": 0, + "content": "[42] Jian Liang, Dapeng Hu, and Jiashi Feng. Do we really need to access the source data? source hypothesis transfer for unsupervised domain adaptation. In ICML, 2020. 1, 2, 3, 4, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.538, + 0.47, + 0.606 + ], + "angle": 0, + "content": "[43] Jian Liang, Dapeng Hu, Yunbo Wang, Ran He, and Jiashi Feng. Source data-absent unsupervised domain adaptation through hypothesis transfer and labeling transfer. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.608, + 0.47, + 0.65 + ], + "angle": 0, + "content": "[44] Hong Liu, Zhangjie Cao, Mingsheng Long, Jianmin Wang, and Qiang Yang. Separate to adapt: Open set domain adaptation via progressive separation. In CVPR, 2019. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.65, + 0.47, + 0.692 + ], + "angle": 0, + "content": "[45] Hong Liu, Mingsheng Long, Jianmin Wang, and Michael Jordan. Transferable adversarial training: A general approach to adapting deep classifiers. In ICML, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.692, + 0.47, + 0.719 + ], + "angle": 0, + "content": "[46] Hong Liu, Jianmin Wang, and Mingsheng Long. Cycle self-training for domain adaptation. In NeurIPS, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.72, + 0.47, + 0.761 + ], + "angle": 0, + "content": "[47] Qinying Liu and Zilei Wang. Collaborating domain-shared and target-specific feature clustering for cross-domain 3d action recognition. In ECCV, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.762, + 0.47, + 0.803 + ], + "angle": 0, + "content": "[48] Mingsheng Long, Zhangjie Cao, Jianmin Wang, and Michael I Jordan. Conditional adversarial domain adaptation. In NeurIPS, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.804, + 0.47, + 0.844 + ], + "angle": 0, + "content": "[49] Mingsheng Long, Han Zhu, Jianmin Wang, and Michael I Jordan. Deep transfer learning with joint adaptation networks. In ICML, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.845, + 0.47, + 0.9 + ], + "angle": 0, + "content": "[50] Wenxuan Ma, Jinming Zhang, Shuang Li, Chi Harold Liu, Yulin Wang, and Wei Li. Making the best of both worlds: A domain-oriented transformer for unsupervised domain adaptation. In ACMMM, 2022. 8" + }, + { + "type": "list", + "bbox": [ + 0.087, + 0.092, + 0.471, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.092, + 0.892, + 0.133 + ], + "angle": 0, + "content": "[51] Ke Mei, Chuang Zhu, Jiaqi Zou, and Shanghang Zhang. Instance adaptive self-training for unsupervised domain adaptation. In ECCV, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.134, + 0.892, + 0.176 + ], + "angle": 0, + "content": "[52] Jaemin Na, Heechul Jung, Hyung Jin Chang, and Wonjun Hwang. FixBi: Bridging domain spaces for unsupervised domain adaptation. In CVPR, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.177, + 0.892, + 0.218 + ], + "angle": 0, + "content": "[53] Van-Anh Nguyen, Tuan Nguyen, Trung Le, Quan Hung Tran, and Dinh Phung. STEM: An approach to multi-source domain adaptation with guarantees. In ICCV, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.219, + 0.892, + 0.272 + ], + "angle": 0, + "content": "[54] Fei Pan, Inkyu Shin, Francois Rameau, Seokju Lee, and In So Kweon. Unsupervised intra-domain adaptation for semantic segmentation through self-supervision. In CVPR, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.273, + 0.892, + 0.315 + ], + "angle": 0, + "content": "[55] Yingwei Pan, Ting Yao, Yehao Li, Yu Wang, Chong-Wah Ngo, and Tao Mei. Transferrable prototypical networks for unsupervised domain adaptation. In CVPR, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.316, + 0.892, + 0.343 + ], + "angle": 0, + "content": "[56] Pau Panareda Busto and Juergen Gall. Open set domain adaptation. In ICCV, 2017. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.343, + 0.892, + 0.385 + ], + "angle": 0, + "content": "[57] Xingchao Peng, Qinxun Bai, Xide Xia, Zijun Huang, Kate Saenko, and Bo Wang. Moment matching for multi-source domain adaptation. In ICCV, 2019. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.385, + 0.892, + 0.427 + ], + "angle": 0, + "content": "[58] Xingchao Peng, Ben Usman, Neela Kaushik, Judy Hoffman, Dequan Wang, and Kate Saenko. VisDA: The visual domain adaptation challenge. In CVPRW, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.427, + 0.892, + 0.482 + ], + "angle": 0, + "content": "[59] Viraj Prabhu, Shivam Khare, Deeksha Kartik, and Judy Hoffman. SENTRY: Selective entropy optimization via committee consistency for unsupervised domain adaptation. In ICCV, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.483, + 0.892, + 0.537 + ], + "angle": 0, + "content": "[60] Zhen Qiu, Yifan Zhang, Hongbin Lin, Shuaicheng Niu, Yanxia Liu, Qing Du, and Mingkui Tan. Source-free domain adaptation via avatar prototype generation and adaptation. In IJCAI, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.538, + 0.892, + 0.594 + ], + "angle": 0, + "content": "[61] Sanqing Qu, Guang Chen, Jing Zhang, Zhijun Li, Wei He, and Dacheng Tao. Bmd: A general class-balanced multicentric dynamic prototype strategy for source-free domain adaptation. In ECCV, 2022. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.595, + 0.892, + 0.648 + ], + "angle": 0, + "content": "[62] Subhankar Roy, Martin Trapp, Andrea Pilzer, Juho Kannala, Nicu Sebe, Elisa Ricci, and Arno Solin. Uncertainty-guided source-free domain adaptation. In ECCV, 2022. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.649, + 0.892, + 0.691 + ], + "angle": 0, + "content": "[63] Kate Saenko, Brian Kulis, Mario Fritz, and Trevor Darrell. Adapting visual category models to new domains. In ECCV, 2010. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.692, + 0.892, + 0.733 + ], + "angle": 0, + "content": "[64] Kuniaki Saito, Kohei Watanabe, Yoshitaka Ushiku, and Tatsuya Harada. Maximum classifier discrepancy for unsupervised domain adaptation. In CVPR, 2018. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.734, + 0.892, + 0.775 + ], + "angle": 0, + "content": "[65] Kuniaki Saito, Shohei Yamamoto, Yoshitaka Ushiku, and Tatsuya Harada. Open set domain adaptation by backpropagation. In ECCV, 2018. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.776, + 0.892, + 0.83 + ], + "angle": 0, + "content": "[66] Marin Scalbert, Maria Vakalopoulou, and Florent Couzini'e-Devy. Multi-source domain adaptation via supervised contrastive learning and confident consistency regularization. In BMVC, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.831, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[67] Kihyuk Sohn, David Berthelot, Chun-Liang Li, Zizhao Zhang, Nicholas Carlini, Ekin D Cubuk, Alex Kurakin, Han Zhang, and Colin Raffel. Fixmatch: Simplifying semi-supervised learning with consistency and confidence. In NeurIPS, 2020. 2, 4, 5" + }, + { + "type": "list", + "bbox": [ + 0.51, + 0.092, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "7628" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.093, + 0.468, + 0.12 + ], + "angle": 0, + "content": "[68] Baochen Sun and Kate Saenko. Deep coral: Correlation alignment for deep domain adaptation. In ECCV, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.122, + 0.468, + 0.162 + ], + "angle": 0, + "content": "[69] Tao Sun, Cheng Lu, Tianshuo Zhang, and Haibin Ling. Safe self-refinement for transformer-based domain adaptation. In CVPR, 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.165, + 0.468, + 0.205 + ], + "angle": 0, + "content": "[70] Hui Tang, Ke Chen, and Kui Jia. Unsupervised domain adaptation via structurally regularized deep clustering. In CVPR, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.208, + 0.468, + 0.262 + ], + "angle": 0, + "content": "[71] Jiayi Tian, Jing Zhang, Wen Li, and Dong Xu. VDMDA: Virtual domain modeling for source data-free domain adaptation. IEEE Transactions on Circuits and Systems for Video Technology, 2021. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.264, + 0.468, + 0.317 + ], + "angle": 0, + "content": "[72] Eric Tzeng, Judy Hoffman, Ning Zhang, Kate Saenko, and Trevor Darrell. Deep domain confusion: Maximizing for domain invariance. arXiv preprint arXiv:1412.3474, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.32, + 0.468, + 0.374 + ], + "angle": 0, + "content": "[73] Hemanth Venkateswara, Jose Eusebio, Shayok Chakraborty, and Sethuraman Panchanathan. Deep hashing network for unsupervised domain adaptation. In CVPR, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.377, + 0.468, + 0.417 + ], + "angle": 0, + "content": "[74] Fan Wang, Zhongyi Han, Yongshun Gong, and Yilong Yin. Exploring domain-invariant parameters for source free domain adaptation. In CVPR, 2022. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.42, + 0.468, + 0.446 + ], + "angle": 0, + "content": "[75] Feng Wang and Huaping Liu. Understanding the behaviour of contrastive loss. In CVPR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.449, + 0.468, + 0.489 + ], + "angle": 0, + "content": "[76] Xiyu Wang, Pengxin Guo, and Yu Zhang. Domain adaptation via bidirectional cross-attention transformer. arXiv preprint arXiv:2201.05887, 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.491, + 0.468, + 0.531 + ], + "angle": 0, + "content": "[77] Ximei Wang, Liang Li, Weirui Ye, Mingsheng Long, and Jianmin Wang. Transferable attention for domain adaptation. In AAAI, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.533, + 0.468, + 0.573 + ], + "angle": 0, + "content": "[78] Zhirong Wu, Yuanjun Xiong, Stella X Yu, and Dahua Lin. Unsupervised feature learning via non-parametric instance discrimination. In CVPR, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.576, + 0.468, + 0.616 + ], + "angle": 0, + "content": "[79] Haifeng Xia, Handong Zhao, and Zhengming Ding. Adaptive adversarial network for source-free domain adaptation. In ICCV, 2021. 1, 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.619, + 0.468, + 0.659 + ], + "angle": 0, + "content": "[80] Shaoan Xie, Zibin Zheng, Liang Chen, and Chuan Chen. Learning semantic representations for unsupervised domain adaptation. In ICML, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.662, + 0.468, + 0.702 + ], + "angle": 0, + "content": "[81] Minghao Xu, Hang Wang, Bingbing Ni, Qi Tian, and Wenjun Zhang. Cross-domain detection via graph-induced prototype alignment. In CVPR, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.704, + 0.468, + 0.757 + ], + "angle": 0, + "content": "[82] Ruijia Xu, Guanbin Li, Jihan Yang, and Liang Lin. Larger norm more transferable: An adaptive feature norm approach for unsupervised domain adaptation. In ICCV, 2019. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.761, + 0.468, + 0.815 + ], + "angle": 0, + "content": "[83] Tongkun Xu, Weihua Chen, WANG Pichao, Fan Wang, Hao Li, and Rong Jin. Cdtrans: Cross-domain transformer for unsupervised domain adaptation. In International Conference on Learning Representations, 2021. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.817, + 0.468, + 0.871 + ], + "angle": 0, + "content": "[84] Hongliang Yan, Yukang Ding, Peihua Li, Qilong Wang, Yong Xu, and Wangmeng Zuo. Mind the class weight bias: Weighted maximum mean discrepancy for unsupervised domain adaptation. In CVPR, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.874, + 0.468, + 0.901 + ], + "angle": 0, + "content": "[85] Fan Yang, Kai Wu, Shuyi Zhang, Guannan Jiang, Yong Liu, Feng Zheng, Wei Zhang, Chengjie Wang, and Long" + }, + { + "type": "list", + "bbox": [ + 0.087, + 0.093, + 0.468, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.541, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "Zeng. Class-aware contrastive semi-supervised learning. In CVPR, 2022. 2, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.122, + 0.892, + 0.162 + ], + "angle": 0, + "content": "[86] Jinyu Yang, Jingjing Liu, Ning Xu, and Junzhou Huang. Tvt: Transferable vision transformer for unsupervised domain adaptation. In WACV, 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.164, + 0.892, + 0.218 + ], + "angle": 0, + "content": "[87] Shiqi Yang, Joost van de Weijer, Luis Herranz, Shangling Jui, et al. Exploiting the intrinsic neighborhood structure for source-free domain adaptation. In NeurIPS, 2021. 1, 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.221, + 0.892, + 0.274 + ], + "angle": 0, + "content": "[88] Shiqi Yang, Yaxing Wang, Joost van de Weijer, Luis Herranz, and Shangling Jui. Unsupervised domain adaptation without source data by casting a bait. arXiv preprint arXiv:2010.12427, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.277, + 0.892, + 0.317 + ], + "angle": 0, + "content": "[89] Shiqi Yang, Yaxing Wang, Joost van de Weijer, Luis Herranz, and Shangling Jui. Generalized source-free domain adaptation. In ICCV, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.319, + 0.892, + 0.372 + ], + "angle": 0, + "content": "[90] Shiqi Yang, Yaxing Wang, Kai Wang, Shangling Jui, and Joost van de Weijer. Attracting and dispersing: A simple approach for source-free domain adaptation. In NeurIPS, 2022. 2, 3, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.375, + 0.892, + 0.43 + ], + "angle": 0, + "content": "[91] Bowen Zhang, Yidong Wang, Wenxin Hou, Hao Wu, Jindong Wang, Manabu Okumura, and Takahiro Shinozaki. Flexmatch: Boosting semi-supervised learning with curriculum pseudo labeling. NeurIPS, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.432, + 0.892, + 0.471 + ], + "angle": 0, + "content": "[92] Jing Zhang, Zewei Ding, Wanqing Li, and Philip Ogunbona. Importance weighted adversarial nets for partial domain adaptation. In CVPR, 2018. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.474, + 0.892, + 0.528 + ], + "angle": 0, + "content": "[93] Pan Zhang, Bo Zhang, Ting Zhang, Dong Chen, Yong Wang, and Fang Wen. Prototypical pseudo label denoising and target structure learning for domain adaptive semantic segmentation. In CVPR, 2021. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.53, + 0.892, + 0.57 + ], + "angle": 0, + "content": "[94] Weichen Zhang, Wanli Ouyang, Wen Li, and Dong Xu. Collaborative and adversarial network for unsupervised domain adaptation. In CVPR, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.572, + 0.892, + 0.612 + ], + "angle": 0, + "content": "[95] Yiheng Zhang, Zhaofan Qiu, Ting Yao, Dong Liu, and Tao Mei. Fully convolutional adaptation networks for semantic segmentation. In CVPR, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.614, + 0.892, + 0.654 + ], + "angle": 0, + "content": "[96] Yixin Zhang and Zilei Wang. Joint adversarial learning for domain adaptation in semantic segmentation. In AAAI, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.656, + 0.892, + 0.696 + ], + "angle": 0, + "content": "[97] Yixin Zhang, Zilei Wang, and Yushi Mao. Rpn prototype alignment for domain adaptive object detector. In CVPR, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.699, + 0.892, + 0.753 + ], + "angle": 0, + "content": "[98] Ziyi Zhang, Weikai Chen, Hui Cheng, Zhen Li, Siyuan Li, Liang Lin, and Guanbin Li. Divide and contrast: Source-free domain adaptation via adaptive contrastive learning. In NeurIPS, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.756, + 0.892, + 0.796 + ], + "angle": 0, + "content": "[99] Yangtao Zheng, Di Huang, Songtao Liu, and Yunhong Wang. Cross-domain object detection through coarse-to-fine feature adaptation. In CVPR, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.798, + 0.892, + 0.837 + ], + "angle": 0, + "content": "[100] Yang Zou, Zhiding Yu, Xiaofeng Liu, BVK Kumar, and Jinsong Wang. Confidence regularized self-training. In ICCV, 2019. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.84, + 0.892, + 0.892 + ], + "angle": 0, + "content": "[101] Yang Zou, Zhiding Yu, BVK Vijaya Kumar, and Jinsong Wang. Unsupervised domain adaptation for semantic segmentation via class-balanced self-training. In ECCV, 2018. 3" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.892, + 0.892 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.956 + ], + "angle": 0, + "content": "7629" + } + ] +] \ No newline at end of file diff --git a/2023/Class Relationship Embedded Learning for Source-Free Unsupervised Domain Adaptation/a0503f36-a779-42c8-98b9-ccfd34b3ed5c_origin.pdf b/2023/Class Relationship Embedded Learning for Source-Free Unsupervised Domain Adaptation/a0503f36-a779-42c8-98b9-ccfd34b3ed5c_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..1105dbb99836eb49640385d2eb03eb396fe48cf7 --- /dev/null +++ b/2023/Class Relationship Embedded Learning for Source-Free Unsupervised Domain Adaptation/a0503f36-a779-42c8-98b9-ccfd34b3ed5c_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3b74d0312d5304329e6d267e6aa2f8b5cfef93abbddcb938748685e981c45cd6 +size 1183456 diff --git a/2023/Class Relationship Embedded Learning for Source-Free Unsupervised Domain Adaptation/full.md b/2023/Class Relationship Embedded Learning for Source-Free Unsupervised Domain Adaptation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..053c43e9bd7541e1a8b1dd5c1dad089def3e7511 --- /dev/null +++ b/2023/Class Relationship Embedded Learning for Source-Free Unsupervised Domain Adaptation/full.md @@ -0,0 +1,359 @@ +# Class Relationship Embedded Learning for Source-Free Unsupervised Domain Adaptation + +Yixin Zhang $^{1,2}$ Zilei Wang $^{*2}$ Weinan He $^{2}$ + +$^{1}$ Institute of Artificial Intelligence, Hefei Comprehensive National Science Center + +$^{2}$ University of Science and Technology of China + +{zhyx12, zlwang}@ustc.edu.cn, hwn2018@mail.ustc.edu.cn + +# Abstract + +This work focuses on a practical knowledge transfer task defined as Source-Free Unsupervised Domain Adaptation (SFUDA), where only a well-trained source model and unlabeled target data are available. To fully utilize source knowledge, we propose to transfer the class relationship, which is domain-invariant but still under-explored in previous works. To this end, we first regard the classifier weights of the source model as class prototypes to compute class relationship, and then propose a novel probability-based similarity between target-domain samples by embedding the source-domain class relationship, resulting in Class Relationship embedded Similarity (CRS). Here the inter-class term is particularly considered in order to more accurately represent the similarity between two samples, in which the source prior of class relationship is utilized by weighting. Finally, we propose to embed CRS into contrastive learning in a unified form. Here both class-aware and instance discrimination contrastive losses are employed, which are complementary to each other. We combine the proposed method with existing representative methods to evaluate its efficacy in multiple SFUDA settings. Extensive experimental results reveal that our method can achieve state-of-the-art performance due to the transfer of domain-invariant class relationship. + +# 1. Introduction + +Benefiting from the large amount of labeled training data, deep neural networks have achieved promising results in many computer vision tasks [7, 15, 19, 95]. To reduce the annotation cost, Unsupervised Domain Adaptation (UDA) has been devised by transferring knowledge from a label-rich source domain to a label-scarce target domain. Currently, many UDA methods have been proposed that jointly + +![](images/01f28aac75a59ee55de6ab8029a003a01d3a372089872d6649859ddaecae9d84.jpg) +Figure 1. Illustration of our proposed method. The upper left represents the target-domain featural distribution by source pre-trained model. The upper right is the feature distribution obtained by our method. The bottom is the process of class relationship embedded contrastive learning. Best viewed in color. + +learn on the source and target data. But they would be unapplicable for some real-world scenarios involving privacy (e.g., medical images, surveillance videos) because the source-domain data cannot be accessed. Thus, more recent methods [36, 42, 71, 79, 87, 88] focus on Source-Free Unsupervised Domain Adaptation (SFUDA). Under this setting, the labeled source data are not accessible any more when training the target model, but the pre-trained model in the source domain is provided. Then a natural question arises, i.e., what knowledge should we transfer to facilitate the learning of unlabeled target-domain data? + +Some methods [21, 36, 42] assume that the source hypothesis (i.e., classifier [42] or whole model [21, 36]) contains sufficient knowledge for target domain. Then they transfer source knowledge by directly aligning features with the fixed source classifier [42], resorting to historical mod + +els [21], or weight regularization [36]. Another line of works [87, 89, 90] assume the source model already forms some semantic structure, and then utilize the local information (i.e., neighborhood) of feature space to enforce the similarity in the output space. Despite these progress, what knowledge to be transferred remains an open question. + +In this work, we propose to transfer the class relationship represented by the similarity of classifier weights. Actually, the class relationship is domain-invariant [14], since the same class in the source and target domains essentially represents the same semantic in spite of domain discrepancy. For example, the computers are always more similar with the TV than the scissors. Thus, it is reasonable to guide the target domain learning using class relationship prior. Unlike previous methods that learn class distribution by pseudo labeling [21, 42] and local aggregation of neighborhood predictions [79, 87, 89, 90], here we explicitly model the source-domain class relationship. To be specific, we regard each weight of classifier as the class prototype [90], and then compute a class relationship matrix $A^s$ by cosine similarity, as shown in Figure 1. + +Before explaining how to use this matrix, we illustrate the purpose of representation learning in the target domain using Figure 1, where three classes are adopted for clarity. The top left shows the target-domain feature distribution together with the class weights learned from the source domain. It can be seen that such a situation makes it difficult to perform correct classification. In fact, it is expected that the learned features are discriminative and compact around the corresponding class weight, as shown on the top right. To this end, an intuitive way is to make the relationship of learned class prototypes in the target domain consistent with that in the source domain. However, it has a very limited effect on training the target-domain model as such a prototype-level constraint is too weak for optimization. + +In this work, we propose to embed the source-domain class relationship in contrastive learning, which has been shown to be outstanding in representation learning [9, 18]. Here we design a novel sample similarity by taking into account $A^s$ . Specifically, we compute the similarity between two target domain samples in the output space, represented by the prediction probabilities (i.e., $p$ and $p'$ ), as shown in Figure 1. In particular, we consider the inter-class term (i.e. $\sum_{i \neq j} p_i p_j'$ ) in addition to the traditional intra-class term (i.e. $\sum_i p_i p_i'$ ). Note that the sum of these two terms equals one, in which the intra-class term measures the similarity that two samples come from the same class, while the inter-class term measures that from different classes. Considering the relationship between classes, we weight the inter-class term by the non-diagonal elements in $A^s$ . In particular, if the class $i$ is closer to $j$ than $k$ (i.e., $A_{ij}^s > A_{ik}^s$ ), $p_i p_j'$ would be more important in calculating the similarity of $p$ and $p'$ . By the above design, our proposed similarity + +can more accurately express the relationship of two samples based on output space. For example, among three classes 1, 2, 3 in Figure 1, the classes 1 and 2 are closer. Given three samples $x_{1}, x_{2}, x_{3}$ belonging to them with the probabilities [0.9, 0.05, 0.05], [0.05, 0.9, 0.05], and [0.05, 0.05, 0.9]. If we only use the intra-class term, all three samples have the same similarity. But for our designed similarity, $x_{1}$ is closer to $x_{2}$ than $x_{3}$ , which is more reasonable. + +On the basis of the proposed similarity, we further propose to perform contrastive learning in a unified form. Inspired by recent success in semi-supervised learning [67, 85, 91] and unsupervised learning [9, 18], we propose two types of contrastive losses. As shown in Figure 1, the first one is Class Relationship embedded Class-Aware Contrastive (CR-CACo) loss where the high-confident samples are enforced to be close to the corresponding prototype and away from other prototypes. Due to embedding the prior class relationship, our CR-CACo loss is more robust to label noise caused by domain shift. The second one is Class Relationship embedded Instance Discrimination Contrastive (CR-IDCo) loss where two views of the same sample are encouraged to be close and away from all other samples. Benefited from our designed accurate similarity, the CR-IDCo loss would more effectively learn discriminative features [9, 18, 75]. Actually, these two losses are complementary to each other, and their combination can achieve better performance. + +Our contributions are summarized as follows: + +- We propose to explicitly transfer class relationship for SFUDA which is more domain-invariant. And we propose to embed the class relationship into contrastive learning in order to effectively perform representation learning in the target domain. +- We propose a novel class relationship embedded similarity which can more accurately express the sample relationship in the output space. Furthermore, we propose two contrastive losses (i.e., CR-CACo and CR-IDCo) exploiting our designed similarity. +- We conduct extensive experiments to evaluate the proposed method, and the results validate the effectiveness of our method, which achieves the state-of-the-art performance on multiple SFUDA benchmarks. + +# 2. Related Work + +# 2.1. Unsupervised Domain Adaptation + +The conventional approaches in UDA [2, 3] are to learn domain-invariant representations, and can be classified into two coarse types. The first one decreases some distribution discrepancy metrics [49, 57, 68, 72, 84]. Another common line is adversarial training [11, 12, 45, 48, 64, 77, 96]. To increase discriminability, more recent DA methods attempt + +to investigate the target domain structure. Self-training as a typical approach generates target domain pseudo labels [16, 17, 27, 28, 40, 46, 51, 54, 94, 101]. Another category is to construct prototypes [8, 34, 55, 80, 81, 93, 97, 99] or cluster centers [13, 26, 47, 70] across domains and then perform class-wise alignment. Among these methods, most of them rely on source-labeled data during target training, while pseudo label based methods can be directly applied in SFUDA. In our CR-CACo loss, pseudo label is also used and our method can alleviate the negative impact of label noise due to the usage of source class relationship. + +# 2.2. Source-free Domain Adaptation + +Early SFUDA methods [31, 32, 36] resort to synthesize extra training samples to get compact decision boundaries. Some recent methods [10, 21, 33, 42, 61] use pseudolabel based self-training. SHOT [42] proposes to freeze the source classifier and it clusters target features by IM loss [20] along with clustering based pseudo labeling. HCL [21] adopts feature space Instance Discrimination [78] from current and historical models, together with pseudo label learning conditioned on historical consistency. D-MCD [10] proposes denoised MCD [64] to mitigate the impact of sample selection bias and label noise. CoWA-JMD [33] uses the joint model-data structure score as sample-wise weights, along with weight mixup to exploit more target knowledge. BMD [61] proposes a class-balanced multicentric dynamic prototype strategy to obtain more accurate pseudo labels. Another line of methods (i.e. NRC [87], G-SFDA [89] and AaD [90]) propose neighborhood clustering which enforces prediction consistency between local neighbors. + +In addition to the above methods, others solve SFUDA from different aspects. $A^2$ Net [79] proposes to learn an additional target classifier and uses a contrastive category-wise matching module to learn compact features. DIPE [74] and VMP [25] explore transferability of source model parameters. Sub-sup [29] proposes novel subsidiary pretext tasks to assist domain adaptation. U-SFAN [62] uses the uncertainty quantified by the predictions to guide the target adaptation. Feat-mix [30] proposes image-level and feature-level mixup to enhances the discriminability-transferability tradeoff. Among these methods, few of them explicitly consider the usage of class relationship. + +# 3. Method + +# 3.1. Preliminaries + +For source-free unsupervised domain adaptation (SFUDA), we are given source pre-trained model and an unlabeled target domain with $N_{t}$ samples as $\mathcal{D}_t = \{x_i^t\}_{i = 1}^{N_t}$ . Target domain have same $C$ classes as source domain (for closed-set setting). The goal of SFUDA is to adapt the + +model to target domain without source data. We divide the model into two parts: the feature extractor $f$ , and the classifier $g$ . The output of classifier is denoted as $\pmb{p} = \delta(g(f(x))) \in \mathbb{R}^C$ where $\delta$ is the softmax function. + +We adopt a teacher-student framework following [18] as shown in Figure 2. The teacher feature extractor (i.e. $\tilde{f}$ ) is continuously updated by exponential moving average of the student feature extractor (i.e. $f$ ), and the classifiers (i.e. $g$ , $\tilde{g}$ ) are shared and frozen as in SHOT [42]. We use sample probabilities as input for contrastive loss, the query probability is obtained from the student model, and the positive and negative probabilities are obtained through the teacher model. During inference, we directly use the student model. In the following, we first introduce class relationship embedded similarity, and then elaborate on how it is used in the proposed contrastive learning process. + +# 3.2. Class Relationship Embedded Similarity + +The inherent semantic and visual relationships among different classes are consistent across domains, regardless of distribution discrepancy. Inspired by this, we propose to explicitly transfer class similarities. Here we regard the normalized classifier weights $\pmb{W}^{s} = [\pmb{w}_{1}^{s},\pmb{w}_{2}^{s},\dots \pmb{w}_{C}^{s}]^{T}\in$ $\mathbb{R}^{C\times D}$ as class prototypes where $D$ is the dimension of the last feature. Thus, the class similarity matrix $\pmb{A}^{s}$ can be obtained by $\pmb{A}^{s} = \pmb{W}^{s}(\pmb{W}^{s})^{T}\in \mathbb{R}^{C\times C}$ . + +An natural way to utilize $\mathbf{A}^s$ is to obtain target class similarity matrix $\mathbf{A}^t$ by target prototypes and then enforce consistency between $\mathbf{A}^s$ and $\mathbf{A}^t$ . However, such a prototype-level constraint is coarse-grained and not strong enough to benefit the training. Instead of prototype-level, we propose a probability based sample-level similarity that considers not only the traditional intra-class term, but also the inter-class term. The latter contains probability products of all different classes, thus we can embed the class relationship matrix $\mathbf{A}^s$ into it as coefficients. Specifically, given the probabilities $\mathbf{p}, \mathbf{p}' \in \mathbb{R}^{C \times 1}$ of two samples, our Class Relationship embedded Similarity (CRS) can be presented as follows: + +$$ +\begin{array}{l} s _ {i n t r a} \left(\boldsymbol {p}, \boldsymbol {p} ^ {\prime}\right) = \sum_ {i = 1} ^ {C} p _ {i} \times p _ {i} ^ {\prime} \\ s _ {i n t e r} \left(\boldsymbol {p}, \boldsymbol {p} ^ {\prime}\right) = \sum_ {i = 1} ^ {C} \sum_ {j = 1, j \neq i} ^ {C} A _ {i, j} ^ {s} \times p _ {i} \times p _ {j} ^ {\prime} \tag {1} \\ s _ {c r} (\pmb {p}, \pmb {p} ^ {\prime}) = s _ {i n t r a} (\pmb {p}, \pmb {p} ^ {\prime}) + \lambda_ {i n t e r} s _ {i n t e r} (\pmb {p}, \pmb {p} ^ {\prime}) \\ \end{array} +$$ + +where $s_{\text{intra}}$ is the traditional intra-class similarity, and $s_{\text{inter}}$ is the inter-class similarity, $\lambda_{\text{inter}}$ is a trade-off parameter between these two terms. + +Given $\lambda_{inter} = 1$ , our similarity can be presented in a simpler version as follows: + +![](images/3417d54a80c94ccb5351e10eafe9ce0f715c02ffc9a1b50d8ae263e9f104dde4.jpg) +Figure 2. The framework of our proposed method. We use a teacher-student architecture to conduct contrastive learning. Class relationship matrix is extracted from fixed classifier and embedded in two types of contrastive loss. Best viewed in color. + +$$ +\begin{array}{l} s _ {c r} (\boldsymbol {p}, \boldsymbol {p} ^ {\prime}) = \sum_ {i = 1} ^ {C} \sum_ {j = 1} ^ {C} A _ {i, j} \times p _ {i} \times p _ {j} ^ {\prime} \\ = \sum_ {i = 1} ^ {C} \sum_ {j = 1} ^ {C} \left(p _ {i} \boldsymbol {w} _ {i} ^ {s}\right) ^ {T} \left(p _ {j} ^ {\prime} \boldsymbol {w} _ {j} ^ {s}\right) = \left(\boldsymbol {p} ^ {T} \boldsymbol {W} ^ {s}\right) \left(\boldsymbol {p} ^ {\prime T} \boldsymbol {W} ^ {s}\right) ^ {T} \tag {2} \\ \end{array} +$$ + +where $\pmb{p}^T\pmb{W}^s \in \mathbb{R}^{1\times D}$ can be viewed as a new feature which is probability weighted sum of different class weights. + +# 3.3. Class Relationship Embedded Contrastive Loss + +Inspired by the success of contrastive learning in different tasks [9, 18, 21, 26, 85], we propose to use our CRS for contrastive loss. Following previous work [9, 18], the contrastive loss can be presented as follows: + +$$ +\begin{array}{l} h (\pmb {p}, \pmb {p} ^ {\prime}) = \exp (s _ {c r} (\pmb {p}, \pmb {p} ^ {\prime}) / T _ {c o}), \\ \ell_ {c o} = - \log \frac {h \left(\boldsymbol {p} , \boldsymbol {p} _ {+}\right)}{h \left(\boldsymbol {p} , \boldsymbol {p} _ {+}\right) + \sum_ {\boldsymbol {p} _ {-} \in M} h \left(\boldsymbol {p} , \boldsymbol {p} _ {-}\right)}, \tag {3} \\ \end{array} +$$ + +where $T_{co}$ is a temperature parameter, and $h$ denotes the exponent of scaled similarity. $M$ is the memory bank [18] storing probabilities of negative samples. Since our CRS is built on probability, here we use probabilities as inputs instead of original features [18] or projected features [9]. + +With this unified form, we further instantiate it to two losses: class-aware contrastive loss and instance discrimination contrastive loss. The framework of our proposed method is shown in Figure 2. + +Class-Aware Contrastive Loss To achieve class-aware learning, it is important to select proper samples and obtain more accurate pseudo labels for them. Inspired by previous efforts in SSL [67, 85] and UDA [41, 93], we consider + +constructing positive pairs from strongly and weakly augmented views of the same image. Specifically, given probabilities $p_s$ and $\tilde{p}_w$ (tilde superscript denotes output by the teacher model) of strongly and weakly augmented views, we first compare $Max(\tilde{p}_w)$ with a predefined threshold $\tau$ to select high-confident samples. Due to domain shift, instead of directly using $argmax(p_w)$ as pseudo label, we adopt the clustering-based pseudo-labeling method in SHOT [42]. Based on pseudo label, we can get the positive samples $p^*$ with the one-hot form. For the negative samples $p_-$ , we use one-hot probabilities of all classes except the pseudo label class. Therefore, both positive and negative samples (i.e. $p^*$ and $p_-$ ) are generated from the pseudo label of weak target sample through teacher model. Finally, the Class Relationship embedded Class-Aware Contrastive (CR-CACo) loss is presented as + +$$ +\ell_ {c r \_ c a c o} = - \log \frac {h \left(\boldsymbol {p} _ {s} , \boldsymbol {p} ^ {*}\right)}{h \left(\boldsymbol {p} _ {s} , \boldsymbol {p} ^ {*}\right) + \sum_ {\boldsymbol {p} _ {-} \in M _ {c a c o}} h \left(\boldsymbol {p} _ {s} , \boldsymbol {p} _ {-}\right)}, \tag {4} +$$ + +where $h$ follows the definition in Equation (3), $M_{caco}$ stores negative samples and is different for each sample according to its pseudo label. + +Regardless of pseudo label generation, using only $s_{intra}$ and replacing $p_s$ with features and $p^*, p_-$ with classifier weights, the above Equation (4) is actually the FixMatch loss [67]. Although FixMatch can improve the performance, our CR-CACo loss can greatly outperform it. + +Our CR-CACo loss is more robust to label noise due to the class relationship embedded inter-class similarity. We compare our CRS with the traditional semi-supervised FixMatch loss. The results are shown in Figure 3. It can be seen that the simple IM loss [42] has lower entropy since it has an entropy minimization term. When FixMatch loss is used, the entropy will increase since the class relationship is learned through high-confidence target samples. Our CRS + +![](images/4f671f0b27492789c3e7ac64a9867ab38a621a4cf81ed33177ba8713d682b514.jpg) +Figure 3. Entropy density plots under SUDA setting Office-Home $\mathrm{Rw} \rightarrow \mathrm{Cl}$ . Our class relationship embedded similarity $s_{cr}$ has higher entropy but better accuracy. Best viewed in color. + +has larger entropy since we explicitly consider the class relationship and can alleviate the negative impact of label noise. + +Instance Discrimination Contrastive Loss The instance discrimination contrastive (IDCo) loss has shown remarkable success in self-supervised learning [9, 18]. It can learn discriminative features by regrading the two views of the same sample as positive pairs and all other samples as negative samples. The traditional IDCo loss uses features as input, but it only makes the feature discriminative and cannot guarantee that the learned feature will be close to corresponding prototypes (i.e. classifier weights). Conducting IDCo loss in the output space (i.e. computing similarity using $s_{\text{intra}}$ ) can alleviate this problem [35], but the more informative relationship between samples in the original feature space can not be fully expressed. This can be addressed by our CRS since we embed the source prior class similarity and can reflect more accurate relationship of samples. + +Here we use the probabilities $\pmb{p}_{s1}$ and $\tilde{\pmb{p}}_{s2}'$ of two strong augmented views from the same image as positive pairs, and probabilities of all other samples in the same batch as negative samples. Finally, the Class Relationship embedded Instance Discrimination Contrastive (CR-IDCo) loss is presented as follows: + +$$ +\ell_ {c r. i d c o} = - \log \frac {h \left(\boldsymbol {p} _ {s 1} , \tilde {\boldsymbol {p}} _ {s 2} ^ {\prime}\right)}{h \left(\boldsymbol {p} _ {s 1} , \tilde {\boldsymbol {p}} _ {s 2} ^ {\prime}\right) + \sum_ {\tilde {\boldsymbol {p}} _ {-} \in M _ {i d c o}} h \left(\boldsymbol {p} _ {s 1} , \tilde {\boldsymbol {p}} _ {-}\right)}, \tag {5} +$$ + +where $h$ follows the definition in Equation (3), $M_{idco}$ contains the probabilities of other samples within the same batch. + +# 3.4. Overall Training Objective + +The proposed two losses are complementary to each other and can be trained end-to-end. Since our CR-CACo + +uses a high prob-threshold (i.e. 0.95) to select confident samples, it should be combined with other methods that can generate confident predictions. Otherwise, it will be dominated by noise labels in the early training stage. Although it performs normally when combined with our CR-IDCo loss, both of them focus on transferring source class relationship, and other methods that encourage confident predictions are still needed and complementary to ours. As a result, we combine our method with existing representative methods (i.e. SHOT-IM [42], AaD [90]) to validate the effectiveness. As a result, the overall loss is presented as: + +$$ +\ell_ {\text {o u r s}} = \ell_ {\text {b a s e l i n e}} + \lambda_ {\text {c a c o}} \ell_ {\text {c r} - \text {c a c o}} + \lambda_ {\text {i d c o}} \ell_ {\text {c r} - \text {i d c o}} \tag {6} +$$ + +where $\lambda_{caco},\lambda_{idco}$ are the trade-off hyperparameters. $l_{baseline}$ is the loss of baseline (i.e., IM or AaD). It contains two terms: discriminability and diversity. The former produces high-confident outputs, and the latter avoids collapse of predictions. IM uses entropy of single sample and average of all samples, and AaD uses neighbor information. + +# 4. Experiments + +# 4.1. Experimental Setup + +Datasets. We evaluate the effectiveness of our approach on four standard DA benchmarks. Office-31 [63] benchmark consists of three domains in office environments: Amazon (A), DSLR (D), and Webcam (W), each with 31 categories. Office-Home [73] is a more challenging dataset. It comprises of images of commonplace objects divided into four domains: Artistic (Ar), Clipart (Cl), Product (Pr), and Real-World (Rw), each with 65 classes. VisDA [58] is a large-scale dataset for synthetic-to-real domain adaptation. The source domain has 152,397 synthetic images, while the target domain has 55,388 real-world images. DomainNet [57] is the most challenging dataset involving 6 domains: Clipart (C), Real (R), Infograph (I), Painting (P), Sketch (S) and Quickdraw (Q) with 345 classes each. + +Implementation details. To ensure fair comparison with related methods, we adopt the backbone of a ResNet-50 [19] for Office-31, Office-Home, and DomainNet, and a ResNet-101 for VisDA. We use the same network architecture as SHOT [42] where the final part of the network is changed to suit the SFUDA task. For the classifier, we find that fixing it or not have little impact on the final performance, and we choose to fix it in all experiments. We adopt SGD with momentum 0.9 and batch size 64 for all datasets where each image has three views (i.e. one weak and two strong augmentations). + +For the hyperparameters, we empirically set the probability threshold $\tau$ to 0.95 following FixMatch [67]. We find that $\tau = 0.95$ works well across all settings and tasks. For the temperature $T_{co}$ in both contrastive losses, + +Table 1. Single-Source Unsupervised DA (SUDA) on Office-Home. + +
MethodSFOffice-Home
Ar→ClAr→PrAr→RwCl→ArCl→PrCl→RwPr→ArPr→ClPr→RwRw→ArRw→ClRw→PrAvg
SENTRY (ICCV'21) [59]X61.877.480.166.371.674.766.863.080.974.066.384.172.2
FixBi (CVPR'21) [52]X58.177.380.467.779.578.165.857.981.776.462.986.772.7
SCDA (ICCV'21) [38]X60.776.482.869.877.578.468.959.082.774.961.884.573.1
A²Net (ICCV'21) [79]✓58.479.082.467.579.378.968.056.282.974.160.585.072.8
NRC (NeurIPS'21) [87]✓57.780.382.068.179.878.665.356.483.071.058.685.672.2
D-MCD (AAAI'22) [10]✓59.478.980.267.279.378.665.355.682.273.362.883.972.2
DIPE (CVPR'22) [74]✓56.579.280.770.179.878.867.955.183.574.159.384.872.5
Sub-Sup (ECCV'22) [29]✓61.080.482.569.179.979.569.157.882.774.565.186.474.0
BMD (ECCV'22) [61]✓58.179.782.669.381.080.770.857.683.674.060.085.973.6
U-SFAN (ECCV'22) [62]✓57.877.881.667.977.379.267.254.781.273.360.383.971.9
CoWA-JMDS (ICML'22) [33]✓56.978.481.069.180.079.967.757.282.472.860.584.572.5
Feat-Mixup (ICML'22) [30]✓61.881.283.068.580.679.467.861.585.173.764.186.574.5
VMP (NeurIPS'22) [25]✓57.977.682.568.679.480.668.455.683.175.259.684.772.8
DaC (NeurIPS'22) [98]✓59.179.581.269.378.979.267.456.482.474.061.484.472.8
SHOT-IM (ICML'20) [42]✓55.476.680.466.974.375.465.654.880.773.758.483.470.5
Ours+ SHOT-IM✓62.882.084.370.980.882.670.061.183.676.265.187.075.5 (+5.0)
AaD (NeurIPS'22) [90]✓59.379.382.168.979.879.567.257.483.172.158.585.472.7
Ours+ AaD✓63.582.185.073.082.782.469.562.982.674.265.787.375.9 (+3.2)
+ +Table 2. Single-Source Unsupervised Domain Adaptation (SUDA) on Office-31 and VisDA benchmarks. + +
MethodSFOffice-31VisDA
A→DA→WD→WW→DD→AW→A AvgS→R
BCDM (AAAI'20) [37]✗93.895.498.6100.073.173.089.083.4
MCC (ECCV'20) [24]✗94.495.598.6100.072.974.989.478.8
FixBi (CVPR'21) [52]✗95.096.199.3100.078.779.491.487.2
RADA (ICCV'21) [23]✗96.196.299.3100.077.577.491.176.3
FAA (ICCV'21) [22]✗94.492.399.299.780.578.790.882.7
SCDA (ICCV'21) [38]✗95.495.399.0100.077.275.990.5-
SHOT (ICML'20) [42]✓94.090.198.499.974.774.388.682.9
CPGA (IJCAI'21) [60]✓94.494.198.499.876.076.689.984.1
VDM-DA (TCSVT'21) [71]✓93.294.198.0100.075.877.189.785.1
A2Net (ICCV'21) [79]✓94.594.099.2100.076.776.190.184.3
HCL (NeurIPS'21) [21]✓90.891.398.2100.072.772.787.683.5
NRC (NeurIPS'21) [87]✓96.090.899.0100.075.375.089.485.9
SHOT++ (TPAMI'21) [43]✓94.390.498.799.976.275.889.287.3
D-MCD (AAAI'22) [10]✓94.193.598.8100.076.476.489.987.5
DIPE (CVPR'22) [74]✓96.693.198.499.675.577.290.183.1
Sub-Sup (ECCV'22) [29]✓95.694.699.299.877.077.790.788.2
BMD (ECCV'22) [61]✓96.294.298.0100.076.076.090.188.7
CoWA-JMDS (ICML'22) [33]✓94.495.298.5100.076.277.690.386.9
Feat-Mixup (ICML'22) [30]✓94.693.298.9100.078.378.990.787.8
VMP (NeurIPS'22) [25]✓93.396.298.6100.075.476.990.0-
SHOT-IM (ICML'20) [42]✓90.691.298.399.972.571.487.380.4
Ours + SHOT-IM✓95.895.199.0100.076.678.390.8(+3.5)89.1(+8.7)
AaD (NeurIPS'22) [90]✓96.492.199.1100.075.076.589.988.0
Ours + AaD✓96.695.599.1100.076.978.391.1(+1.2)89.6(+1.6)
+ +we set it to 0.07 following previous contrastive learning methods [18, 21, 85]. For the $\lambda_{inter}$ in our similarity, we use 1.0 for simplicity. For the trade-off hyperparameters $\lambda_{caco}, \lambda_{idco}$ , we directly set them to 1.0. More training details can be found in the supplementary material. + +# 4.2. Results + +The results are shown in Table 1, Table 2, Table 3 and Table 4. In each table, SF indicates source-free, * indicates results we reproduced from the released code. (+x.x) indicates gains over SHOT-IM [42] and AaD [90] respectively. Ours means combination of our proposed two losses. + +Single Source Domain Adaptation Table 1 and Table 2 show the classification accuracy for single-source DA on each dataset: Office-Home, Office-31, and VisDA. Based on the simple SHOT-IM and stronger AaD, our method can achieve state-of-the-art performance. UDA methods BCDM [37] and MCC [24] also explore the inter-class relationship by directly suppressing the inter-class term to obtain determined outputs. Our method surpasses these methods even without source domain supervision. + +The two proposed contrastive losses exploit different views of the same image. Thus our method is complementary to those not utilizing information of different views. Both baselines only consider weak views, and so our method can boost them. AaD surpasses IM by regarding local neighbors as positive pairs. In our CR-IDCo loss, the strong views of the same image can be regarded as neighbors since they are very close in feature space, which has overlapped effect with AaD. When our method is combined with AaD, therefore, the improvement is relatively smaller. + +Multi Source Domain Adaptation Table 3 shows the results for multi-source DA on Office-Home and the largescale DomainNet benchmark. We adopt two baselines without using domain labels, that is, all data from different source domains is treated as belonging to a single domain and only one source model is trained and transferred. We observe improvements of $4.1\%$ and $3.4\%$ in DomainNet, $4.5\%$ and $3.0\%$ in Office-Home. CAiDA [14] considers class relationship by enforcing consistency of different source model outputs. It achieves $76.2\%$ on Office-Home by combining IM loss, a novel confident-anchor-induced pseudo label method and the class-relation-space-aware consistency loss. Based on the same IM loss, ours achieves $78.1\%$ and exceeds it by $1.9\%$ . + +Table 3. Multi-Source Unsupervised Domain Adaptation (MUDA) on DomainNet and Office-Home. + +
MethodSFw/o Domain +LabelsDomainNetOffice-Home
→C→I→P→Q→R→SAvg→Ar→Cl→Pr→RwAvg
MCC (ECCV'20) [24]XX65.526.056.616.568.052.747.6-----
CMSDA (BMVC'21) [66]XX70.926.557.521.368.159.450.471.567.784.182.976.6
DRT (CVPR'21) [39]XX71.031.661.012.371.460.751.3-----
STEM (ICCV'21) [53]XX72.028.261.525.772.660.253.4-----
SourcecombineX✓57.023.454.114.667.250.344.458.057.374.277.966.9
DECISION (CVPR'21) [1]✓X61.521.654.618.967.551.045.974.559.484.483.675.5
CAiDA (NeurIPS'21) [14]✓X-------75.260.584.784.276.2
Sub-Sup (ECCV'22) [29]✓✓70.325.757.317.169.957.149.675.164.186.684.477.6
Feat-Mixup (ICML'22) [30]✓✓75.424.657.823.665.858.551.072.667.485.983.677.4
SHOT-IM* (ICML'20) [42]✓✓64.524.255.215.567.053.246.671.558.681.982.373.6
Ours + SHOT-IM✓✓69.827.059.021.270.257.550.7(+4.1)74.966.485.185.878.1 (+4.5)
AaD* (NeurIPS'22) [90]✓✓65.725.556.417.068.154.647.972.160.685.384.475.6
Ours +AaD✓✓71.227.859.322.569.357.851.3(+3.4)75.466.986.485.978.6 (+3.0)
+ +Table 4. Partial-set and open-set Domain Adaptation (PDA and ODA) on Office-Home. + +
Partial-set DASF Avg.Open-set DASF Avg.
ResNet-50 [19]× 61.3ResNet [19]× 65.3
IWAN (CVPR'18) [92]× 63.6OpenMax (CVPR'16) [4]× 66.7
SAN (CVPR'18) [5]× 65.3ATI-λ (ICCV'17) [56]× 66.1
ETN (CVPR'19) [6]× 70.5OSBP (ECCV'18) [65]× 65.7
SAFN (ICCV'19) [82]× 71.8STA (CVPR'19) [44]× 69.5
Source model only✓ 62.8Source model only✓ 46.6
SHOT (ICML'20) [42]✓ 79.3SHOT (ICML'20) [42]✓ 72.8
SHOT+HCL (NeurIPS'21) [21]✓ 80.1SHOT+HCL (NeurIPS'21) [21]✓ 73.2
CoWA-JMDS (ICML'22) [33]✓ 83.2CoWA-JMDS (ICML'22) [33]✓ 73.2
SHOT-IM (ICML'20) [42]✓ 76.8SHOT-IM (ICML'20) [42]✓ 71.5
Ours +SHOT-IM✓ 80.6 (+3.8)Ours +SHOT-IM✓ 73.2(+1.7)
AaD* (NeurIPS'22) [90]✓ 79.7AaD* (NeurIPS'22) [90]✓ 71.8
Ours +AaD✓ 82.4 (+2.7)Ours +AaD✓ 73.3(+1.5)
+ +Domain Adaptation Beyond Vanilla Closed-set We provide additional results under source-free partial-set and open-set DA (PDA and ODA) setting on Office-Home. For open-set detection in ODA, we follow the same protocol to detect unseen categories as SHOT [42]. Results are shown in Table 4. It can be seen that our method can achieve performance comparable to that of the state-of-the-arts. + +# 4.3. Analysis + +Component-wise ablations Here we validate the effectiveness of different proposed losses. As described in Sec. 3.4, we do not use CR-CACo loss alone. The results are shown in Table 5. It can be seen that when using our method alone (line #2) can achieve comparable performance. When combined with SHOT-IM, our both losses are complementary, and the best performance can be achieved by combining baseline and our losses. Note that although CR-IDCo (line #1) outperforms IM (line #0), it is worse than IM when combined with CR-CACo (i.e. line #2 < line #3). This is caused by the fact that CR-IDCo cannot produce enough confident samples as IM for the learning of CR-CACo loss. + +Effect of pseudo labels in CR-CACo In our proposed CR-CACo loss, the pseudo label is generated from clustering-based method in SHOT [42] instead of predictions of weakly augmented images. The results of the comparison are shown in Table 6. It can be seen that our + +Table 5. Component-wise ablation studies of the proposed methods under the SUDA setting. + +
#IMCR-CACoCR-IDCoOffice-31Office-HomeVisDA
0✓87.370.580.4
1✓89.272.583.0
2✓✓89.674.085.3
3✓✓89.874.887.0
4✓✓90.274.385.8
5✓✓✓90.875.589.1
+ +CR-CACo can bring significant improvement when using pseudo label from model prediction. The clustering based pseudo-label can bring slight but consistent improvements. + +Table 6. Comparison of different pseudo label methods. + +
#BaselinePseudo-label in CR-CACoOffice-31Office-HomeVisDA
PredictionClustering
087.370.580.4
1IM✓89.6(+2.3)74.5(+4.0)86.6(+6.2)
2✓89.8(+2.5)74.8(+4.3)87.0(+6.6)
389.972.788.0
4AaD✓90.6(+0.7)75.0(+2.3)88.9(+0.9)
5✓90.7(+0.8)75.2(+2.5)89.1(+1.1)
+ +Different similarities and disentangled improvements. To validate the superiority of our class relationship embedded similarity in contrastive loss, we compare it with other two choices. The first one uses only $s_{intra}$ based on features, and the second one uses only $s_{intra}$ based on probabilities which equals $\lambda_{inter} = 0$ in our case. The results are shown in Table 7. Using features (i.e. i.e. $s_{intra}$ w/ feature) rather than probabilities is a more natural way. Indeed, using features in CACo Loss would result in FixMatch loss, and using features in IDCo loss would fall in the traditional self-supervised way (e.g., MoCo). Only using the probability based intra-class term (i.e. $s_{intra}$ w/ probability) is still special for similarity in contrastive losses. And it can bring consistent improvements. In both CACo and IDCo losses, our $s_{cr}$ can achieve the best result, thus validating the effectiveness of adopting source prior class similarities. + +Table 7. Comparison with different similarities. + +
#LossSimilarityOffice-31Office-HomeVisDA
0IM—87.370.580.4
1IM+CACosintra w/ feature88.972.583.5
2sintra w/ probability89.172.984.0
3scr w/ probability89.874.887.0
4IM+IDCosintra w/ feature87.570.980.8
5sintra w/ probability88.271.481.5
6scr w/ probability90.274.385.8
+ +Comparison with fixed coefficient within $s_{inter}$ In the definition of inter-class term in Equation (1), we use the source class similarity matrix $A^s$ as coefficient. Instead of this prior knowledge, we can also set the coefficient to a small fixed value. Considering the positive pairs in contrastive loss, the negative coefficient (i.e. -0.1) works to suppress the inter-class term, thus obtaining more determined predictions as in [24, 37] while the positive value (i.e. 0.1) is more like a smooth regularization as KLD loss in [100]. The results are shown in Figure 4. It can be seen that both positive and negative values can bring some improvements. However, our CR-CACo loss can achieve better performance, which validates the importance of source prior class relationship. + +![](images/40eee282579e5154d2c9b30b11ecb5601bfe9e42b8e5d5cc9267255ae70d646e.jpg) +Figure 4. Using fixed coefficients in $s_{inter}$ in both CACo and IDCo losses under SUDA setting Office-Home Rw→Cl. Black dotted line shows the performance of baseline IM. Red dotted line shows the performance of our losses. + +![](images/97f9aace554f4bf3963be3b5a63656b1c009b8a3cacc3cc5828033531f0184ed.jpg) + +Parameter sensitivity Here we analyze the hyperparameter sensitivity under the setting of SUDA Office-Home with SHOT-IM baseline. We consider four hyperparameters, as shown in Figure 5. It is evident that they are stable within specific ranges. + +![](images/fd46b8a5fb321a60c20f354ffc3153494d624b1ad57f0ea342e5abb3b0b397f5.jpg) +Figure 5. Sensitivity analysis of different hyperparameters. The first two use our two losses. The last two use corresponding loss. + +![](images/e2ec62eae2eee709ec6a4f0d714a9e827c2bab0869a754c906934091a6e6a6c3.jpg) + +![](images/3b5ebe708294f823b942c00d64035aaaeafae24c2449cd2d7241bea804eb5702.jpg) + +![](images/5adb7af2bb021a2f1e0209afda563b0f312537b89fe91b92ac2309e8dbadb225.jpg) + +Visualization of features We visualize the target features and classifier weights in Figure 6. It can be seen that our method has less misclassified samples for classes 0, 1, 7 compared with IM. Thus our method can learn more discriminative features. + +![](images/a96140f46635e00fcd21fda7f2e65eac519197eb71608c757f8a972d12758aff.jpg) +Source Pre-trained Model + +![](images/c8de4fefb7c9b6cb96f3beda508aebde9f7d424866d32e8f17d93ad2fd091e9b.jpg) +IM +Figure 6. The t-SNE visualization of classifier weights (i.e. star) and target domain features (i.e. dot) of different methods under the setting of SUDA Office-Home $\mathrm{Rw} \rightarrow \mathrm{Cl}$ . We directly use the first 10 classes (i.e. different colors), and the coordinates of classifier weights are the same across different methods since source classifier is frozen. Best viewed in color. + +![](images/7dc7c4a2f18f0de05efdcba2655e03b7e7cea50aada2308289fbccfad77834e3.jpg) +IM+Ours + +Transformer backbone We further apply our method with Transformer-based backbone under the SFUDA setting, as shown in Table 8. We choose the same network (i.e. ViT-B [15]) following SSRT [69]. For the training of the source model, we follow SHOT [42] which obtains better performance than original ViT-B (78.2% vs. 75.5%). Adding IM loss achieves 84.0%. Based on IM loss, ours achieves 88.3% which significantly outperforms SSRT. + +Table 8. Single-Source Unsupervised DA (SUDA) on Office-Home with ViT-B backbone. + +
MethodSFAvg.MethodSFAvg.
ViT-B✗75.5CDTrans (ICLR'22) [83]✗80.5
TVT (WACV'23) [86]✗83.6DOT-B (ACMMM'22) [50]✗84.3
SSRT (CVPR'22) [69]✗85.4BCAT-DTF (Arxiv'22) [76]✗86.6
Source model✓78.2SHOT-IM✓84.0
SHOT-IM+CR-CACo✓87.0(+3.0)SHOT-IM+Ours✓88.3(+4.3)
+ +# 5. Conclusion + +In this paper, we propose to explicitly transfer the class relationship for SFUDA which is more domain-invariant. We propose a novel class relationship embedded similarity that can more accurately express the sample relationship in the output space. Furthermore, we propose two contrastive losses (i.e., CR-CACo and CR-IDCo) that exploit our designed similarity. These two losses are complementary, and their combination can better explore the target distribution. We combine our method with existing representative baselines in multiple SFUDA settings. Extensive experiments show the effectiveness of the proposed method, which achieves state-of-the-art performance. + +# 6. Acknowledgements + +This work is supported by the National Natural Science Foundation of China under Grant 62176246 and Grant 61836008. This work is also supported by Anhui Provincial Natural Science Foundation 2208085UD17 and the Fundamental Research Funds for the Central Universities (WK3490000006). + +# References + +[1] Sk. Miraj Ahmed, Dripta S. Raychaudhuri, S. Paul, Samet Oymak, and Amit K. Roy-Chowdhury. Unsupervised multi-source domain adaptation without access to source data. In CVPR, 2021. 7 +[2] Shai Ben-David, John Blitzer, Koby Crammer, Alex Kulesza, Fernando Pereira, and Jennifer Wortman Vaughan. A theory of learning from different domains. Machine learning, 79(1-2):151-175, 2010. 2 +[3] Shai Ben-David, John Blitzer, Koby Crammer, and Fernando Pereira. Analysis of representations for domain adaptation. In NeurIPS, 2006. 2 +[4] Abhijit Bendale and Terrance E Boult. Towards open set deep networks. In CVPR, 2016. 7 +[5] Zhangjie Cao, Mingsheng Long, Jianmin Wang, and Michael I Jordan. Partial transfer learning with selective adversarial networks. In CVPR, 2018. 7 +[6] Zhangjie Cao, Kaichao You, Mingsheng Long, Jianmin Wang, and Qiang Yang. Learning to transfer examples for partial domain adaptation. In CVPR, 2019. 7 +[7] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In ECCV, 2020. 1 +[8] Chaoqi Chen, Weiping Xie, Tingyang Xu, Wenbing Huang, Yu Rong, Xinghao Ding, Yue Huang, and Junzhou Huang. Progressive feature alignment for unsupervised domain adaptation. In CVPR, 2019. 3 +[9] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In ICML, 2020. 2, 4, 5 +[10] Tong Chu, Yahao Liu, Jinhong Deng, Wen Li, and Lixin Duan. Denoised maximum classifier discrepancy for sourcefree unsupervised domain adaptation. In AAAI, 2022. 3, 6 +[11] Safa Cicek and Stefano Soatto. Unsupervised domain adaptation via regularized conditional alignment. In ICCV, 2019. 2 +[12] Shuhao Cui, Shuhui Wang, Junbao Zhuo, Chi Su, Qingming Huang, and Qi Tian. Gradually vanishing bridge for adversarial domain adaptation. In CVPR, 2020. 2 +[13] Zhijie Deng, Yucen Luo, and Jun Zhu. Cluster alignment with a teacher for unsupervised domain adaptation. In ICCV, 2019. 3 +[14] Jiahua Dong, Zhen Fang, Anjin Liu, Gan Sun, and Tongliang Liu. Confident anchor-induced multi-source free domain adaptation. In NeurIPS, 2021. 2, 6, 7 +[15] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2020. 1, 8 +[16] Geoffrey French, Michal Mackiewicz, and Mark Fisher. Self-ensembling for visual domain adaptation. In ICLR, 2018. 3 + +[17] Yuan Gao, Zilei Wang, Jiafan Zhuang, Yixin Zhang, and Junjie Li. Exploit domain-robust optical flow in domain adaptive video semantic segmentation. In AAAI, 2023. 3 +[18] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In CVPR, 2020. 2, 3, 4, 5, 6 +[19] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, 2016. 1, 5, 7 +[20] Weihua Hu, Takeru Miyato, Seiya Tokui, Eiichi Matsumoto, and Masashi Sugiyama. Learning discrete representations via information maximizing self-augmented training. In ICML, 2017. 3 +[21] Jiaxing Huang, Dayan Guan, Aoran Xiao, and Shijian Lu. Model adaptation: Historical contrastive learning for unsupervised domain adaptation without source data. In NeurIPS, 2021. 1, 2, 3, 4, 6, 7 +[22] Jiaxing Huang, Dayan Guan, Aoran Xiao, and Shijian Lu. RDA: Robust domain adaptation via fourier adversarial attacking. In ICCV, 2021. 6 +[23] Xin Jin, Cuiling Lan, Wenjun Zeng, and Zhibo Chen. Re-energizing domain discriminator with sample relabeling for adversarial domain adaptation. In ICCV, 2021. 6 +[24] Ying Jin, Ximei Wang, Mingsheng Long, and Jianmin Wang. Minimum class confusion for versatile domain adaptation. In ECCV, 2020. 6, 7, 8 +[25] Mengmeng Jing, Xiantong Zhen, Jingjing Li, and Cees GM Snoek. Variational model perturbation for source-free domain adaptation. In NeurIPS, 2022. 3, 6 +[26] Guoliang Kang, Lu Jiang, Yi Yang, and Alexander G Hauptmann. Contrastive adaptation network for unsupervised domain adaptation. In CVPR, 2019. 3, 4 +[27] Seunghyeon Kim, Jaehoon Choi, Taekyung Kim, and Changick Kim. Self-training and adversarial background regularization for unsupervised domain adaptive one-stage object detection. In CVPR, 2019. 3 +[28] Ananya Kumar, Tengyu Ma, and Percy Liang. Understanding self-training for gradual domain adaptation. In ICML, 2020. 3 +[29] Jogendra Nath Kundu, Suvaansh Bhambri, Akshay Kulkarni, Hiran Sarkar, Varun Jampani, and R Venkatesh Babu. Concurrent subsidiary supervision for unsupervised source-free domain adaptation. In ECCV, 2022. 3, 6, 7 +[30] Jogendra Nath Kundu, Akshay R Kulkarni, Suvaansh Bhambri, Deepesh Mehta, Shreyas Anand Kulkarni, Varun Jampani, and Venkatesh Babu Radhakrishnan. Balancing discriminability and transferability for source-free domain adaptation. In ICML, 2022. 3, 6, 7 +[31] Jogendra Nath Kundu, Naveen Venkat, and R Venkatesh Babu. Universal source-free domain adaptation. In CVPR, 2020. 3 +[32] Jogendra Nath Kundu, Naveen Venkat, Ambareesh Revanur, Rahul M V, and R. Venkatesh Babu. Towards inheritable models for open-set domain adaptation. In CVPR, 2020. 3 +[33] Jonghyun Lee, Dahuin Jung, Junho Yim, and Sungroh Yoon. Confidence score for source-free unsupervised domain adaptation. In ICML, 2022. 3, 6, 7 + +[34] Junjie Li, Zilei Wang, Yuan Gao, and Xiaoming Hu. Exploring high-quality target domain information for unsupervised domain adaptive semantic segmentation. In ACMMM, 2022. 3 +[35] Junjie Li, Yixin Zhang, Zilei Wang, and Keyu Tu. Semantic-aware representation learning via probability contrastive loss. arXiv preprint arXiv:2111.06021, 2021. 5 +[36] Rui Li, Qianfen Jiao, Wenming Cao, Hau-San Wong, and Si Wu. Model adaptation: Unsupervised domain adaptation without source data. In CVPR, 2020. 1, 2, 3 +[37] Shuang Li, Fangrui Lv, Binhui Xie, Chi Harold Liu, Jian Liang, and Chen Qin. Bi-classifier determinacy maximization for unsupervised domain adaptation. In AAAI, 2021. 6, 8 +[38] Shuang Li, Mixue Xie, Fangrui Lv, Chi Harold Liu, Jian Liang, Chen Qin, and Wei Li. Semantic concentration for domain adaptation. In ICCV, 2021. 6 +[39] Yunsheng Li, Lu Yuan, Yinpeng Chen, Pei Wang, and N. Vasconcelos. Dynamic transfer for multi-source domain adaptation. In CVPR, 2021. 7 +[40] Yunsheng Li, Lu Yuan, and Nuno Vasconcelos. Bidirectional learning for domain adaptation of semantic segmentation. In CVPR, 2019. 3 +[41] Yu-Jhe Li, Xiaoliang Dai, Chih-Yao Ma, Yen-Cheng Liu, Kan Chen, Bichen Wu, Zijian He, Kris Kitani, and Peter Vajda. Cross-domain adaptive teacher for object detection. In CVPR, 2022. 4 +[42] Jian Liang, Dapeng Hu, and Jiashi Feng. Do we really need to access the source data? source hypothesis transfer for unsupervised domain adaptation. In ICML, 2020. 1, 2, 3, 4, 5, 6, 7, 8 +[43] Jian Liang, Dapeng Hu, Yunbo Wang, Ran He, and Jiashi Feng. Source data-absent unsupervised domain adaptation through hypothesis transfer and labeling transfer. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2021. 6 +[44] Hong Liu, Zhangjie Cao, Mingsheng Long, Jianmin Wang, and Qiang Yang. Separate to adapt: Open set domain adaptation via progressive separation. In CVPR, 2019. 7 +[45] Hong Liu, Mingsheng Long, Jianmin Wang, and Michael Jordan. Transferable adversarial training: A general approach to adapting deep classifiers. In ICML, 2019. 2 +[46] Hong Liu, Jianmin Wang, and Mingsheng Long. Cycle self-training for domain adaptation. In NeurIPS, 2021. 3 +[47] Qinying Liu and Zilei Wang. Collaborating domain-shared and target-specific feature clustering for cross-domain 3d action recognition. In ECCV, 2022. 3 +[48] Mingsheng Long, Zhangjie Cao, Jianmin Wang, and Michael I Jordan. Conditional adversarial domain adaptation. In NeurIPS, 2018. 2 +[49] Mingsheng Long, Han Zhu, Jianmin Wang, and Michael I Jordan. Deep transfer learning with joint adaptation networks. In ICML, 2017. 2 +[50] Wenxuan Ma, Jinming Zhang, Shuang Li, Chi Harold Liu, Yulin Wang, and Wei Li. Making the best of both worlds: A domain-oriented transformer for unsupervised domain adaptation. In ACMMM, 2022. 8 + +[51] Ke Mei, Chuang Zhu, Jiaqi Zou, and Shanghang Zhang. Instance adaptive self-training for unsupervised domain adaptation. In ECCV, 2020. 3 +[52] Jaemin Na, Heechul Jung, Hyung Jin Chang, and Wonjun Hwang. FixBi: Bridging domain spaces for unsupervised domain adaptation. In CVPR, 2021. 6 +[53] Van-Anh Nguyen, Tuan Nguyen, Trung Le, Quan Hung Tran, and Dinh Phung. STEM: An approach to multi-source domain adaptation with guarantees. In ICCV, 2021. 7 +[54] Fei Pan, Inkyu Shin, Francois Rameau, Seokju Lee, and In So Kweon. Unsupervised intra-domain adaptation for semantic segmentation through self-supervision. In CVPR, 2020. 3 +[55] Yingwei Pan, Ting Yao, Yehao Li, Yu Wang, Chong-Wah Ngo, and Tao Mei. Transferrable prototypical networks for unsupervised domain adaptation. In CVPR, 2019. 3 +[56] Pau Panareda Busto and Juergen Gall. Open set domain adaptation. In ICCV, 2017. 7 +[57] Xingchao Peng, Qinxun Bai, Xide Xia, Zijun Huang, Kate Saenko, and Bo Wang. Moment matching for multi-source domain adaptation. In ICCV, 2019. 2, 5 +[58] Xingchao Peng, Ben Usman, Neela Kaushik, Judy Hoffman, Dequan Wang, and Kate Saenko. VisDA: The visual domain adaptation challenge. In CVPRW, 2018. 5 +[59] Viraj Prabhu, Shivam Khare, Deeksha Kartik, and Judy Hoffman. SENTRY: Selective entropy optimization via committee consistency for unsupervised domain adaptation. In ICCV, 2021. 6 +[60] Zhen Qiu, Yifan Zhang, Hongbin Lin, Shuaicheng Niu, Yanxia Liu, Qing Du, and Mingkui Tan. Source-free domain adaptation via avatar prototype generation and adaptation. In IJCAI, 2021. 6 +[61] Sanqing Qu, Guang Chen, Jing Zhang, Zhijun Li, Wei He, and Dacheng Tao. Bmd: A general class-balanced multicentric dynamic prototype strategy for source-free domain adaptation. In ECCV, 2022. 3, 6 +[62] Subhankar Roy, Martin Trapp, Andrea Pilzer, Juho Kannala, Nicu Sebe, Elisa Ricci, and Arno Solin. Uncertainty-guided source-free domain adaptation. In ECCV, 2022. 3, 6 +[63] Kate Saenko, Brian Kulis, Mario Fritz, and Trevor Darrell. Adapting visual category models to new domains. In ECCV, 2010. 5 +[64] Kuniaki Saito, Kohei Watanabe, Yoshitaka Ushiku, and Tatsuya Harada. Maximum classifier discrepancy for unsupervised domain adaptation. In CVPR, 2018. 2, 3 +[65] Kuniaki Saito, Shohei Yamamoto, Yoshitaka Ushiku, and Tatsuya Harada. Open set domain adaptation by backpropagation. In ECCV, 2018. 7 +[66] Marin Scalbert, Maria Vakalopoulou, and Florent Couzini'e-Devy. Multi-source domain adaptation via supervised contrastive learning and confident consistency regularization. In BMVC, 2021. 7 +[67] Kihyuk Sohn, David Berthelot, Chun-Liang Li, Zizhao Zhang, Nicholas Carlini, Ekin D Cubuk, Alex Kurakin, Han Zhang, and Colin Raffel. Fixmatch: Simplifying semi-supervised learning with consistency and confidence. In NeurIPS, 2020. 2, 4, 5 + +[68] Baochen Sun and Kate Saenko. Deep coral: Correlation alignment for deep domain adaptation. In ECCV, 2016. 2 +[69] Tao Sun, Cheng Lu, Tianshuo Zhang, and Haibin Ling. Safe self-refinement for transformer-based domain adaptation. In CVPR, 2022. 8 +[70] Hui Tang, Ke Chen, and Kui Jia. Unsupervised domain adaptation via structurally regularized deep clustering. In CVPR, 2020. 3 +[71] Jiayi Tian, Jing Zhang, Wen Li, and Dong Xu. VDMDA: Virtual domain modeling for source data-free domain adaptation. IEEE Transactions on Circuits and Systems for Video Technology, 2021. 1, 6 +[72] Eric Tzeng, Judy Hoffman, Ning Zhang, Kate Saenko, and Trevor Darrell. Deep domain confusion: Maximizing for domain invariance. arXiv preprint arXiv:1412.3474, 2014. 2 +[73] Hemanth Venkateswara, Jose Eusebio, Shayok Chakraborty, and Sethuraman Panchanathan. Deep hashing network for unsupervised domain adaptation. In CVPR, 2017. 5 +[74] Fan Wang, Zhongyi Han, Yongshun Gong, and Yilong Yin. Exploring domain-invariant parameters for source free domain adaptation. In CVPR, 2022. 3, 6 +[75] Feng Wang and Huaping Liu. Understanding the behaviour of contrastive loss. In CVPR, 2021. 2 +[76] Xiyu Wang, Pengxin Guo, and Yu Zhang. Domain adaptation via bidirectional cross-attention transformer. arXiv preprint arXiv:2201.05887, 2022. 8 +[77] Ximei Wang, Liang Li, Weirui Ye, Mingsheng Long, and Jianmin Wang. Transferable attention for domain adaptation. In AAAI, 2019. 2 +[78] Zhirong Wu, Yuanjun Xiong, Stella X Yu, and Dahua Lin. Unsupervised feature learning via non-parametric instance discrimination. In CVPR, 2018. 3 +[79] Haifeng Xia, Handong Zhao, and Zhengming Ding. Adaptive adversarial network for source-free domain adaptation. In ICCV, 2021. 1, 2, 3, 6 +[80] Shaoan Xie, Zibin Zheng, Liang Chen, and Chuan Chen. Learning semantic representations for unsupervised domain adaptation. In ICML, 2018. 3 +[81] Minghao Xu, Hang Wang, Bingbing Ni, Qi Tian, and Wenjun Zhang. Cross-domain detection via graph-induced prototype alignment. In CVPR, 2020. 3 +[82] Ruijia Xu, Guanbin Li, Jihan Yang, and Liang Lin. Larger norm more transferable: An adaptive feature norm approach for unsupervised domain adaptation. In ICCV, 2019. 7 +[83] Tongkun Xu, Weihua Chen, WANG Pichao, Fan Wang, Hao Li, and Rong Jin. Cdtrans: Cross-domain transformer for unsupervised domain adaptation. In International Conference on Learning Representations, 2021. 8 +[84] Hongliang Yan, Yukang Ding, Peihua Li, Qilong Wang, Yong Xu, and Wangmeng Zuo. Mind the class weight bias: Weighted maximum mean discrepancy for unsupervised domain adaptation. In CVPR, 2017. 2 +[85] Fan Yang, Kai Wu, Shuyi Zhang, Guannan Jiang, Yong Liu, Feng Zheng, Wei Zhang, Chengjie Wang, and Long + +Zeng. Class-aware contrastive semi-supervised learning. In CVPR, 2022. 2, 4, 6 +[86] Jinyu Yang, Jingjing Liu, Ning Xu, and Junzhou Huang. Tvt: Transferable vision transformer for unsupervised domain adaptation. In WACV, 2023. 8 +[87] Shiqi Yang, Joost van de Weijer, Luis Herranz, Shangling Jui, et al. Exploiting the intrinsic neighborhood structure for source-free domain adaptation. In NeurIPS, 2021. 1, 2, 3, 6 +[88] Shiqi Yang, Yaxing Wang, Joost van de Weijer, Luis Herranz, and Shangling Jui. Unsupervised domain adaptation without source data by casting a bait. arXiv preprint arXiv:2010.12427, 2020. 1 +[89] Shiqi Yang, Yaxing Wang, Joost van de Weijer, Luis Herranz, and Shangling Jui. Generalized source-free domain adaptation. In ICCV, 2021. 2, 3 +[90] Shiqi Yang, Yaxing Wang, Kai Wang, Shangling Jui, and Joost van de Weijer. Attracting and dispersing: A simple approach for source-free domain adaptation. In NeurIPS, 2022. 2, 3, 5, 6, 7 +[91] Bowen Zhang, Yidong Wang, Wenxin Hou, Hao Wu, Jindong Wang, Manabu Okumura, and Takahiro Shinozaki. Flexmatch: Boosting semi-supervised learning with curriculum pseudo labeling. NeurIPS, 2021. 2 +[92] Jing Zhang, Zewei Ding, Wanqing Li, and Philip Ogunbona. Importance weighted adversarial nets for partial domain adaptation. In CVPR, 2018. 7 +[93] Pan Zhang, Bo Zhang, Ting Zhang, Dong Chen, Yong Wang, and Fang Wen. Prototypical pseudo label denoising and target structure learning for domain adaptive semantic segmentation. In CVPR, 2021. 3, 4 +[94] Weichen Zhang, Wanli Ouyang, Wen Li, and Dong Xu. Collaborative and adversarial network for unsupervised domain adaptation. In CVPR, 2018. 3 +[95] Yiheng Zhang, Zhaofan Qiu, Ting Yao, Dong Liu, and Tao Mei. Fully convolutional adaptation networks for semantic segmentation. In CVPR, 2018. 1 +[96] Yixin Zhang and Zilei Wang. Joint adversarial learning for domain adaptation in semantic segmentation. In AAAI, 2020. 2 +[97] Yixin Zhang, Zilei Wang, and Yushi Mao. Rpn prototype alignment for domain adaptive object detector. In CVPR, 2021. 3 +[98] Ziyi Zhang, Weikai Chen, Hui Cheng, Zhen Li, Siyuan Li, Liang Lin, and Guanbin Li. Divide and contrast: Source-free domain adaptation via adaptive contrastive learning. In NeurIPS, 2022. 6 +[99] Yangtao Zheng, Di Huang, Songtao Liu, and Yunhong Wang. Cross-domain object detection through coarse-to-fine feature adaptation. In CVPR, 2020. 3 +[100] Yang Zou, Zhiding Yu, Xiaofeng Liu, BVK Kumar, and Jinsong Wang. Confidence regularized self-training. In ICCV, 2019. 8 +[101] Yang Zou, Zhiding Yu, BVK Vijaya Kumar, and Jinsong Wang. Unsupervised domain adaptation for semantic segmentation via class-balanced self-training. In ECCV, 2018. 3 \ No newline at end of file diff --git a/2023/Class Relationship Embedded Learning for Source-Free Unsupervised Domain Adaptation/images.zip b/2023/Class Relationship Embedded Learning for Source-Free Unsupervised Domain Adaptation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..13efd729811f1f7a8159329eb8413fda5805790b --- /dev/null +++ b/2023/Class Relationship Embedded Learning for Source-Free Unsupervised Domain Adaptation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4dd1cc09b97648f072361a74a39770a00e9be2197941fc234273c9c2436bb635 +size 724519 diff --git a/2023/Class Relationship Embedded Learning for Source-Free Unsupervised Domain Adaptation/layout.json b/2023/Class Relationship Embedded Learning for Source-Free Unsupervised Domain Adaptation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..b134af2a56b9047181433e4986bed60a7c7a2808 --- /dev/null +++ b/2023/Class Relationship Embedded Learning for Source-Free Unsupervised Domain Adaptation/layout.json @@ -0,0 +1,10415 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 140, + 103, + 452, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 140, + 103, + 452, + 140 + ], + "spans": [ + { + "bbox": [ + 140, + 103, + 452, + 140 + ], + "type": "text", + "content": "Class Relationship Embedded Learning for Source-Free Unsupervised Domain Adaptation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 186, + 160, + 408, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 186, + 160, + 408, + 175 + ], + "spans": [ + { + "bbox": [ + 186, + 160, + 408, + 175 + ], + "type": "text", + "content": "Yixin Zhang" + }, + { + "bbox": [ + 186, + 160, + 408, + 175 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 186, + 160, + 408, + 175 + ], + "type": "text", + "content": " Zilei Wang" + }, + { + "bbox": [ + 186, + 160, + 408, + 175 + ], + "type": "inline_equation", + "content": "^{*2}" + }, + { + "bbox": [ + 186, + 160, + 408, + 175 + ], + "type": "text", + "content": " Weinan He" + }, + { + "bbox": [ + 186, + 160, + 408, + 175 + ], + "type": "inline_equation", + "content": "^{2}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 99, + 175, + 495, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 99, + 175, + 495, + 190 + ], + "spans": [ + { + "bbox": [ + 99, + 175, + 495, + 190 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 99, + 175, + 495, + 190 + ], + "type": "text", + "content": " Institute of Artificial Intelligence, Hefei Comprehensive National Science Center" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 179, + 189, + 415, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 179, + 189, + 415, + 203 + ], + "spans": [ + { + "bbox": [ + 179, + 189, + 415, + 203 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 179, + 189, + 415, + 203 + ], + "type": "text", + "content": " University of Science and Technology of China" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 149, + 205, + 441, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 149, + 205, + 441, + 217 + ], + "spans": [ + { + "bbox": [ + 149, + 205, + 441, + 217 + ], + "type": "text", + "content": "{zhyx12, zlwang}@ustc.edu.cn, hwn2018@mail.ustc.edu.cn" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 143, + 244, + 192, + 257 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 244, + 192, + 257 + ], + "spans": [ + { + "bbox": [ + 143, + 244, + 192, + 257 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 270, + 290, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 270, + 290, + 559 + ], + "spans": [ + { + "bbox": [ + 47, + 270, + 290, + 559 + ], + "type": "text", + "content": "This work focuses on a practical knowledge transfer task defined as Source-Free Unsupervised Domain Adaptation (SFUDA), where only a well-trained source model and unlabeled target data are available. To fully utilize source knowledge, we propose to transfer the class relationship, which is domain-invariant but still under-explored in previous works. To this end, we first regard the classifier weights of the source model as class prototypes to compute class relationship, and then propose a novel probability-based similarity between target-domain samples by embedding the source-domain class relationship, resulting in Class Relationship embedded Similarity (CRS). Here the inter-class term is particularly considered in order to more accurately represent the similarity between two samples, in which the source prior of class relationship is utilized by weighting. Finally, we propose to embed CRS into contrastive learning in a unified form. Here both class-aware and instance discrimination contrastive losses are employed, which are complementary to each other. We combine the proposed method with existing representative methods to evaluate its efficacy in multiple SFUDA settings. Extensive experimental results reveal that our method can achieve state-of-the-art performance due to the transfer of domain-invariant class relationship." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 581, + 128, + 594 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 581, + 128, + 594 + ], + "spans": [ + { + "bbox": [ + 47, + 581, + 128, + 594 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 601, + 287, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 601, + 287, + 685 + ], + "spans": [ + { + "bbox": [ + 46, + 601, + 287, + 685 + ], + "type": "text", + "content": "Benefiting from the large amount of labeled training data, deep neural networks have achieved promising results in many computer vision tasks [7, 15, 19, 95]. To reduce the annotation cost, Unsupervised Domain Adaptation (UDA) has been devised by transferring knowledge from a label-rich source domain to a label-scarce target domain. Currently, many UDA methods have been proposed that jointly" + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 310, + 242, + 545, + 434 + ], + "blocks": [ + { + "bbox": [ + 310, + 242, + 545, + 434 + ], + "lines": [ + { + "bbox": [ + 310, + 242, + 545, + 434 + ], + "spans": [ + { + "bbox": [ + 310, + 242, + 545, + 434 + ], + "type": "image", + "image_path": "01f28aac75a59ee55de6ab8029a003a01d3a372089872d6649859ddaecae9d84.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 444, + 547, + 499 + ], + "lines": [ + { + "bbox": [ + 305, + 444, + 547, + 499 + ], + "spans": [ + { + "bbox": [ + 305, + 444, + 547, + 499 + ], + "type": "text", + "content": "Figure 1. Illustration of our proposed method. The upper left represents the target-domain featural distribution by source pre-trained model. The upper right is the feature distribution obtained by our method. The bottom is the process of class relationship embedded contrastive learning. Best viewed in color." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 521, + 546, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 521, + 546, + 653 + ], + "spans": [ + { + "bbox": [ + 304, + 521, + 546, + 653 + ], + "type": "text", + "content": "learn on the source and target data. But they would be unapplicable for some real-world scenarios involving privacy (e.g., medical images, surveillance videos) because the source-domain data cannot be accessed. Thus, more recent methods [36, 42, 71, 79, 87, 88] focus on Source-Free Unsupervised Domain Adaptation (SFUDA). Under this setting, the labeled source data are not accessible any more when training the target model, but the pre-trained model in the source domain is provided. Then a natural question arises, i.e., what knowledge should we transfer to facilitate the learning of unlabeled target-domain data?" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 653, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 653, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 653, + 547, + 713 + ], + "type": "text", + "content": "Some methods [21, 36, 42] assume that the source hypothesis (i.e., classifier [42] or whole model [21, 36]) contains sufficient knowledge for target domain. Then they transfer source knowledge by directly aligning features with the fixed source classifier [42], resorting to historical mod" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 693, + 136, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 693, + 136, + 703 + ], + "spans": [ + { + "bbox": [ + 58, + 693, + 136, + 703 + ], + "type": "text", + "content": "*Corresponding author" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 58, + 703, + 274, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 703, + 274, + 712 + ], + "spans": [ + { + "bbox": [ + 58, + 703, + 274, + 712 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 58, + 703, + 274, + 712 + ], + "type": "text", + "content": "Code is available at https://github.com/zhyx12/CRCo" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7619" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 288, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 288, + 144 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 288, + 144 + ], + "type": "text", + "content": "els [21], or weight regularization [36]. Another line of works [87, 89, 90] assume the source model already forms some semantic structure, and then utilize the local information (i.e., neighborhood) of feature space to enforce the similarity in the output space. Despite these progress, what knowledge to be transferred remains an open question." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 147, + 287, + 327 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 147, + 287, + 327 + ], + "spans": [ + { + "bbox": [ + 46, + 147, + 287, + 327 + ], + "type": "text", + "content": "In this work, we propose to transfer the class relationship represented by the similarity of classifier weights. Actually, the class relationship is domain-invariant [14], since the same class in the source and target domains essentially represents the same semantic in spite of domain discrepancy. For example, the computers are always more similar with the TV than the scissors. Thus, it is reasonable to guide the target domain learning using class relationship prior. Unlike previous methods that learn class distribution by pseudo labeling [21, 42] and local aggregation of neighborhood predictions [79, 87, 89, 90], here we explicitly model the source-domain class relationship. To be specific, we regard each weight of classifier as the class prototype [90], and then compute a class relationship matrix " + }, + { + "bbox": [ + 46, + 147, + 287, + 327 + ], + "type": "inline_equation", + "content": "A^s" + }, + { + "bbox": [ + 46, + 147, + 287, + 327 + ], + "type": "text", + "content": " by cosine similarity, as shown in Figure 1." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 328, + 287, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 328, + 287, + 496 + ], + "spans": [ + { + "bbox": [ + 46, + 328, + 287, + 496 + ], + "type": "text", + "content": "Before explaining how to use this matrix, we illustrate the purpose of representation learning in the target domain using Figure 1, where three classes are adopted for clarity. The top left shows the target-domain feature distribution together with the class weights learned from the source domain. It can be seen that such a situation makes it difficult to perform correct classification. In fact, it is expected that the learned features are discriminative and compact around the corresponding class weight, as shown on the top right. To this end, an intuitive way is to make the relationship of learned class prototypes in the target domain consistent with that in the source domain. However, it has a very limited effect on training the target-domain model as such a prototype-level constraint is too weak for optimization." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 498, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 498, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 498, + 287, + 715 + ], + "type": "text", + "content": "In this work, we propose to embed the source-domain class relationship in contrastive learning, which has been shown to be outstanding in representation learning [9, 18]. Here we design a novel sample similarity by taking into account " + }, + { + "bbox": [ + 46, + 498, + 287, + 715 + ], + "type": "inline_equation", + "content": "A^s" + }, + { + "bbox": [ + 46, + 498, + 287, + 715 + ], + "type": "text", + "content": ". Specifically, we compute the similarity between two target domain samples in the output space, represented by the prediction probabilities (i.e., " + }, + { + "bbox": [ + 46, + 498, + 287, + 715 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 46, + 498, + 287, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 498, + 287, + 715 + ], + "type": "inline_equation", + "content": "p'" + }, + { + "bbox": [ + 46, + 498, + 287, + 715 + ], + "type": "text", + "content": "), as shown in Figure 1. In particular, we consider the inter-class term (i.e. " + }, + { + "bbox": [ + 46, + 498, + 287, + 715 + ], + "type": "inline_equation", + "content": "\\sum_{i \\neq j} p_i p_j'" + }, + { + "bbox": [ + 46, + 498, + 287, + 715 + ], + "type": "text", + "content": ") in addition to the traditional intra-class term (i.e. " + }, + { + "bbox": [ + 46, + 498, + 287, + 715 + ], + "type": "inline_equation", + "content": "\\sum_i p_i p_i'" + }, + { + "bbox": [ + 46, + 498, + 287, + 715 + ], + "type": "text", + "content": "). Note that the sum of these two terms equals one, in which the intra-class term measures the similarity that two samples come from the same class, while the inter-class term measures that from different classes. Considering the relationship between classes, we weight the inter-class term by the non-diagonal elements in " + }, + { + "bbox": [ + 46, + 498, + 287, + 715 + ], + "type": "inline_equation", + "content": "A^s" + }, + { + "bbox": [ + 46, + 498, + 287, + 715 + ], + "type": "text", + "content": ". In particular, if the class " + }, + { + "bbox": [ + 46, + 498, + 287, + 715 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 498, + 287, + 715 + ], + "type": "text", + "content": " is closer to " + }, + { + "bbox": [ + 46, + 498, + 287, + 715 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 46, + 498, + 287, + 715 + ], + "type": "text", + "content": " than " + }, + { + "bbox": [ + 46, + 498, + 287, + 715 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 498, + 287, + 715 + ], + "type": "text", + "content": " (i.e., " + }, + { + "bbox": [ + 46, + 498, + 287, + 715 + ], + "type": "inline_equation", + "content": "A_{ij}^s > A_{ik}^s" + }, + { + "bbox": [ + 46, + 498, + 287, + 715 + ], + "type": "text", + "content": "), " + }, + { + "bbox": [ + 46, + 498, + 287, + 715 + ], + "type": "inline_equation", + "content": "p_i p_j'" + }, + { + "bbox": [ + 46, + 498, + 287, + 715 + ], + "type": "text", + "content": " would be more important in calculating the similarity of " + }, + { + "bbox": [ + 46, + 498, + 287, + 715 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 46, + 498, + 287, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 498, + 287, + 715 + ], + "type": "inline_equation", + "content": "p'" + }, + { + "bbox": [ + 46, + 498, + 287, + 715 + ], + "type": "text", + "content": ". By the above design, our proposed similarity" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 304, + 72, + 545, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 167 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 167 + ], + "type": "text", + "content": "can more accurately express the relationship of two samples based on output space. For example, among three classes 1, 2, 3 in Figure 1, the classes 1 and 2 are closer. Given three samples " + }, + { + "bbox": [ + 304, + 72, + 545, + 167 + ], + "type": "inline_equation", + "content": "x_{1}, x_{2}, x_{3}" + }, + { + "bbox": [ + 304, + 72, + 545, + 167 + ], + "type": "text", + "content": " belonging to them with the probabilities [0.9, 0.05, 0.05], [0.05, 0.9, 0.05], and [0.05, 0.05, 0.9]. If we only use the intra-class term, all three samples have the same similarity. But for our designed similarity, " + }, + { + "bbox": [ + 304, + 72, + 545, + 167 + ], + "type": "inline_equation", + "content": "x_{1}" + }, + { + "bbox": [ + 304, + 72, + 545, + 167 + ], + "type": "text", + "content": " is closer to " + }, + { + "bbox": [ + 304, + 72, + 545, + 167 + ], + "type": "inline_equation", + "content": "x_{2}" + }, + { + "bbox": [ + 304, + 72, + 545, + 167 + ], + "type": "text", + "content": " than " + }, + { + "bbox": [ + 304, + 72, + 545, + 167 + ], + "type": "inline_equation", + "content": "x_{3}" + }, + { + "bbox": [ + 304, + 72, + 545, + 167 + ], + "type": "text", + "content": ", which is more reasonable." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 168, + 546, + 395 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 168, + 546, + 395 + ], + "spans": [ + { + "bbox": [ + 304, + 168, + 546, + 395 + ], + "type": "text", + "content": "On the basis of the proposed similarity, we further propose to perform contrastive learning in a unified form. Inspired by recent success in semi-supervised learning [67, 85, 91] and unsupervised learning [9, 18], we propose two types of contrastive losses. As shown in Figure 1, the first one is Class Relationship embedded Class-Aware Contrastive (CR-CACo) loss where the high-confident samples are enforced to be close to the corresponding prototype and away from other prototypes. Due to embedding the prior class relationship, our CR-CACo loss is more robust to label noise caused by domain shift. The second one is Class Relationship embedded Instance Discrimination Contrastive (CR-IDCo) loss where two views of the same sample are encouraged to be close and away from all other samples. Benefited from our designed accurate similarity, the CR-IDCo loss would more effectively learn discriminative features [9, 18, 75]. Actually, these two losses are complementary to each other, and their combination can achieve better performance." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 318, + 396, + 503, + 406 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 396, + 503, + 406 + ], + "spans": [ + { + "bbox": [ + 318, + 396, + 503, + 406 + ], + "type": "text", + "content": "Our contributions are summarized as follows:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 318, + 413, + 545, + 594 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 318, + 413, + 545, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 413, + 545, + 472 + ], + "spans": [ + { + "bbox": [ + 318, + 413, + 545, + 472 + ], + "type": "text", + "content": "- We propose to explicitly transfer class relationship for SFUDA which is more domain-invariant. And we propose to embed the class relationship into contrastive learning in order to effectively perform representation learning in the target domain." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 318, + 479, + 545, + 539 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 479, + 545, + 539 + ], + "spans": [ + { + "bbox": [ + 318, + 479, + 545, + 539 + ], + "type": "text", + "content": "- We propose a novel class relationship embedded similarity which can more accurately express the sample relationship in the output space. Furthermore, we propose two contrastive losses (i.e., CR-CACo and CR-IDCo) exploiting our designed similarity." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 318, + 546, + 545, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 546, + 545, + 594 + ], + "spans": [ + { + "bbox": [ + 318, + 546, + 545, + 594 + ], + "type": "text", + "content": "- We conduct extensive experiments to evaluate the proposed method, and the results validate the effectiveness of our method, which achieves the state-of-the-art performance on multiple SFUDA benchmarks." + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 604, + 392, + 616 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 604, + 392, + 616 + ], + "spans": [ + { + "bbox": [ + 306, + 604, + 392, + 616 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 623, + 488, + 636 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 623, + 488, + 636 + ], + "spans": [ + { + "bbox": [ + 306, + 623, + 488, + 636 + ], + "type": "text", + "content": "2.1. Unsupervised Domain Adaptation" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "type": "text", + "content": "The conventional approaches in UDA [2, 3] are to learn domain-invariant representations, and can be classified into two coarse types. The first one decreases some distribution discrepancy metrics [49, 57, 68, 72, 84]. Another common line is adversarial training [11, 12, 45, 48, 64, 77, 96]. To increase discriminability, more recent DA methods attempt" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7620" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 204 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 204 + ], + "type": "text", + "content": "to investigate the target domain structure. Self-training as a typical approach generates target domain pseudo labels [16, 17, 27, 28, 40, 46, 51, 54, 94, 101]. Another category is to construct prototypes [8, 34, 55, 80, 81, 93, 97, 99] or cluster centers [13, 26, 47, 70] across domains and then perform class-wise alignment. Among these methods, most of them rely on source-labeled data during target training, while pseudo label based methods can be directly applied in SFUDA. In our CR-CACo loss, pseudo label is also used and our method can alleviate the negative impact of label noise due to the usage of source class relationship." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 213, + 220, + 225 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 213, + 220, + 225 + ], + "spans": [ + { + "bbox": [ + 47, + 213, + 220, + 225 + ], + "type": "text", + "content": "2.2. Source-free Domain Adaptation" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 231, + 289, + 459 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 231, + 289, + 459 + ], + "spans": [ + { + "bbox": [ + 46, + 231, + 289, + 459 + ], + "type": "text", + "content": "Early SFUDA methods [31, 32, 36] resort to synthesize extra training samples to get compact decision boundaries. Some recent methods [10, 21, 33, 42, 61] use pseudolabel based self-training. SHOT [42] proposes to freeze the source classifier and it clusters target features by IM loss [20] along with clustering based pseudo labeling. HCL [21] adopts feature space Instance Discrimination [78] from current and historical models, together with pseudo label learning conditioned on historical consistency. D-MCD [10] proposes denoised MCD [64] to mitigate the impact of sample selection bias and label noise. CoWA-JMD [33] uses the joint model-data structure score as sample-wise weights, along with weight mixup to exploit more target knowledge. BMD [61] proposes a class-balanced multicentric dynamic prototype strategy to obtain more accurate pseudo labels. Another line of methods (i.e. NRC [87], G-SFDA [89] and AaD [90]) propose neighborhood clustering which enforces prediction consistency between local neighbors." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 459, + 289, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 459, + 289, + 604 + ], + "spans": [ + { + "bbox": [ + 46, + 459, + 289, + 604 + ], + "type": "text", + "content": "In addition to the above methods, others solve SFUDA from different aspects. " + }, + { + "bbox": [ + 46, + 459, + 289, + 604 + ], + "type": "inline_equation", + "content": "A^2" + }, + { + "bbox": [ + 46, + 459, + 289, + 604 + ], + "type": "text", + "content": "Net [79] proposes to learn an additional target classifier and uses a contrastive category-wise matching module to learn compact features. DIPE [74] and VMP [25] explore transferability of source model parameters. Sub-sup [29] proposes novel subsidiary pretext tasks to assist domain adaptation. U-SFAN [62] uses the uncertainty quantified by the predictions to guide the target adaptation. Feat-mix [30] proposes image-level and feature-level mixup to enhances the discriminability-transferability tradeoff. Among these methods, few of them explicitly consider the usage of class relationship." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 614, + 104, + 627 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 614, + 104, + 627 + ], + "spans": [ + { + "bbox": [ + 47, + 614, + 104, + 627 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 634, + 133, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 634, + 133, + 647 + ], + "spans": [ + { + "bbox": [ + 47, + 634, + 133, + 647 + ], + "type": "text", + "content": "3.1. Preliminaries" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 654, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 654, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 654, + 288, + 715 + ], + "type": "text", + "content": "For source-free unsupervised domain adaptation (SFUDA), we are given source pre-trained model and an unlabeled target domain with " + }, + { + "bbox": [ + 46, + 654, + 288, + 715 + ], + "type": "inline_equation", + "content": "N_{t}" + }, + { + "bbox": [ + 46, + 654, + 288, + 715 + ], + "type": "text", + "content": " samples as " + }, + { + "bbox": [ + 46, + 654, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_t = \\{x_i^t\\}_{i = 1}^{N_t}" + }, + { + "bbox": [ + 46, + 654, + 288, + 715 + ], + "type": "text", + "content": ". Target domain have same " + }, + { + "bbox": [ + 46, + 654, + 288, + 715 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 46, + 654, + 288, + 715 + ], + "type": "text", + "content": " classes as source domain (for closed-set setting). The goal of SFUDA is to adapt the" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "text", + "content": "model to target domain without source data. We divide the model into two parts: the feature extractor " + }, + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "text", + "content": ", and the classifier " + }, + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "text", + "content": ". The output of classifier is denoted as " + }, + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "inline_equation", + "content": "\\pmb{p} = \\delta(g(f(x))) \\in \\mathbb{R}^C" + }, + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "text", + "content": " is the softmax function." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 121, + 547, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 121, + 547, + 266 + ], + "spans": [ + { + "bbox": [ + 304, + 121, + 547, + 266 + ], + "type": "text", + "content": "We adopt a teacher-student framework following [18] as shown in Figure 2. The teacher feature extractor (i.e. " + }, + { + "bbox": [ + 304, + 121, + 547, + 266 + ], + "type": "inline_equation", + "content": "\\tilde{f}" + }, + { + "bbox": [ + 304, + 121, + 547, + 266 + ], + "type": "text", + "content": ") is continuously updated by exponential moving average of the student feature extractor (i.e. " + }, + { + "bbox": [ + 304, + 121, + 547, + 266 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 304, + 121, + 547, + 266 + ], + "type": "text", + "content": "), and the classifiers (i.e. " + }, + { + "bbox": [ + 304, + 121, + 547, + 266 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 304, + 121, + 547, + 266 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 121, + 547, + 266 + ], + "type": "inline_equation", + "content": "\\tilde{g}" + }, + { + "bbox": [ + 304, + 121, + 547, + 266 + ], + "type": "text", + "content": ") are shared and frozen as in SHOT [42]. We use sample probabilities as input for contrastive loss, the query probability is obtained from the student model, and the positive and negative probabilities are obtained through the teacher model. During inference, we directly use the student model. In the following, we first introduce class relationship embedded similarity, and then elaborate on how it is used in the proposed contrastive learning process." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 275, + 518, + 289 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 275, + 518, + 289 + ], + "spans": [ + { + "bbox": [ + 305, + 275, + 518, + 289 + ], + "type": "text", + "content": "3.2. Class Relationship Embedded Similarity" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 295, + 545, + 391 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 295, + 545, + 391 + ], + "spans": [ + { + "bbox": [ + 304, + 295, + 545, + 391 + ], + "type": "text", + "content": "The inherent semantic and visual relationships among different classes are consistent across domains, regardless of distribution discrepancy. Inspired by this, we propose to explicitly transfer class similarities. Here we regard the normalized classifier weights " + }, + { + "bbox": [ + 304, + 295, + 545, + 391 + ], + "type": "inline_equation", + "content": "\\pmb{W}^{s} = [\\pmb{w}_{1}^{s},\\pmb{w}_{2}^{s},\\dots \\pmb{w}_{C}^{s}]^{T}\\in" + }, + { + "bbox": [ + 304, + 295, + 545, + 391 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^{C\\times D}" + }, + { + "bbox": [ + 304, + 295, + 545, + 391 + ], + "type": "text", + "content": " as class prototypes where " + }, + { + "bbox": [ + 304, + 295, + 545, + 391 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 304, + 295, + 545, + 391 + ], + "type": "text", + "content": " is the dimension of the last feature. Thus, the class similarity matrix " + }, + { + "bbox": [ + 304, + 295, + 545, + 391 + ], + "type": "inline_equation", + "content": "\\pmb{A}^{s}" + }, + { + "bbox": [ + 304, + 295, + 545, + 391 + ], + "type": "text", + "content": " can be obtained by " + }, + { + "bbox": [ + 304, + 295, + 545, + 391 + ], + "type": "inline_equation", + "content": "\\pmb{A}^{s} = \\pmb{W}^{s}(\\pmb{W}^{s})^{T}\\in \\mathbb{R}^{C\\times C}" + }, + { + "bbox": [ + 304, + 295, + 545, + 391 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 392, + 547, + 547 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 392, + 547, + 547 + ], + "spans": [ + { + "bbox": [ + 304, + 392, + 547, + 547 + ], + "type": "text", + "content": "An natural way to utilize " + }, + { + "bbox": [ + 304, + 392, + 547, + 547 + ], + "type": "inline_equation", + "content": "\\mathbf{A}^s" + }, + { + "bbox": [ + 304, + 392, + 547, + 547 + ], + "type": "text", + "content": " is to obtain target class similarity matrix " + }, + { + "bbox": [ + 304, + 392, + 547, + 547 + ], + "type": "inline_equation", + "content": "\\mathbf{A}^t" + }, + { + "bbox": [ + 304, + 392, + 547, + 547 + ], + "type": "text", + "content": " by target prototypes and then enforce consistency between " + }, + { + "bbox": [ + 304, + 392, + 547, + 547 + ], + "type": "inline_equation", + "content": "\\mathbf{A}^s" + }, + { + "bbox": [ + 304, + 392, + 547, + 547 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 392, + 547, + 547 + ], + "type": "inline_equation", + "content": "\\mathbf{A}^t" + }, + { + "bbox": [ + 304, + 392, + 547, + 547 + ], + "type": "text", + "content": ". However, such a prototype-level constraint is coarse-grained and not strong enough to benefit the training. Instead of prototype-level, we propose a probability based sample-level similarity that considers not only the traditional intra-class term, but also the inter-class term. The latter contains probability products of all different classes, thus we can embed the class relationship matrix " + }, + { + "bbox": [ + 304, + 392, + 547, + 547 + ], + "type": "inline_equation", + "content": "\\mathbf{A}^s" + }, + { + "bbox": [ + 304, + 392, + 547, + 547 + ], + "type": "text", + "content": " into it as coefficients. Specifically, given the probabilities " + }, + { + "bbox": [ + 304, + 392, + 547, + 547 + ], + "type": "inline_equation", + "content": "\\mathbf{p}, \\mathbf{p}' \\in \\mathbb{R}^{C \\times 1}" + }, + { + "bbox": [ + 304, + 392, + 547, + 547 + ], + "type": "text", + "content": " of two samples, our Class Relationship embedded Similarity (CRS) can be presented as follows:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 323, + 555, + 545, + 642 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 323, + 555, + 545, + 642 + ], + "spans": [ + { + "bbox": [ + 323, + 555, + 545, + 642 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} s _ {i n t r a} \\left(\\boldsymbol {p}, \\boldsymbol {p} ^ {\\prime}\\right) = \\sum_ {i = 1} ^ {C} p _ {i} \\times p _ {i} ^ {\\prime} \\\\ s _ {i n t e r} \\left(\\boldsymbol {p}, \\boldsymbol {p} ^ {\\prime}\\right) = \\sum_ {i = 1} ^ {C} \\sum_ {j = 1, j \\neq i} ^ {C} A _ {i, j} ^ {s} \\times p _ {i} \\times p _ {j} ^ {\\prime} \\tag {1} \\\\ s _ {c r} (\\pmb {p}, \\pmb {p} ^ {\\prime}) = s _ {i n t r a} (\\pmb {p}, \\pmb {p} ^ {\\prime}) + \\lambda_ {i n t e r} s _ {i n t e r} (\\pmb {p}, \\pmb {p} ^ {\\prime}) \\\\ \\end{array}", + "image_path": "eeb695f9bad3f08e84ba7fb1849206a9f0376b5f239148ec392e20c4237d62f3.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 652, + 545, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 652, + 545, + 687 + ], + "spans": [ + { + "bbox": [ + 304, + 652, + 545, + 687 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 652, + 545, + 687 + ], + "type": "inline_equation", + "content": "s_{\\text{intra}}" + }, + { + "bbox": [ + 304, + 652, + 545, + 687 + ], + "type": "text", + "content": " is the traditional intra-class similarity, and " + }, + { + "bbox": [ + 304, + 652, + 545, + 687 + ], + "type": "inline_equation", + "content": "s_{\\text{inter}}" + }, + { + "bbox": [ + 304, + 652, + 545, + 687 + ], + "type": "text", + "content": " is the inter-class similarity, " + }, + { + "bbox": [ + 304, + 652, + 545, + 687 + ], + "type": "inline_equation", + "content": "\\lambda_{\\text{inter}}" + }, + { + "bbox": [ + 304, + 652, + 545, + 687 + ], + "type": "text", + "content": " is a trade-off parameter between these two terms." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "text", + "content": "Given " + }, + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\lambda_{inter} = 1" + }, + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "text", + "content": ", our similarity can be presented in a simpler version as follows:" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "7621" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 63, + 62, + 531, + 238 + ], + "blocks": [ + { + "bbox": [ + 63, + 62, + 531, + 238 + ], + "lines": [ + { + "bbox": [ + 63, + 62, + 531, + 238 + ], + "spans": [ + { + "bbox": [ + 63, + 62, + 531, + 238 + ], + "type": "image", + "image_path": "3417d54a80c94ccb5351e10eafe9ce0f715c02ffc9a1b50d8ae263e9f104dde4.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 242, + 547, + 266 + ], + "lines": [ + { + "bbox": [ + 46, + 242, + 547, + 266 + ], + "spans": [ + { + "bbox": [ + 46, + 242, + 547, + 266 + ], + "type": "text", + "content": "Figure 2. The framework of our proposed method. We use a teacher-student architecture to conduct contrastive learning. Class relationship matrix is extracted from fixed classifier and embedded in two types of contrastive loss. Best viewed in color." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 297, + 287, + 376 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 297, + 287, + 376 + ], + "spans": [ + { + "bbox": [ + 56, + 297, + 287, + 376 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} s _ {c r} (\\boldsymbol {p}, \\boldsymbol {p} ^ {\\prime}) = \\sum_ {i = 1} ^ {C} \\sum_ {j = 1} ^ {C} A _ {i, j} \\times p _ {i} \\times p _ {j} ^ {\\prime} \\\\ = \\sum_ {i = 1} ^ {C} \\sum_ {j = 1} ^ {C} \\left(p _ {i} \\boldsymbol {w} _ {i} ^ {s}\\right) ^ {T} \\left(p _ {j} ^ {\\prime} \\boldsymbol {w} _ {j} ^ {s}\\right) = \\left(\\boldsymbol {p} ^ {T} \\boldsymbol {W} ^ {s}\\right) \\left(\\boldsymbol {p} ^ {\\prime T} \\boldsymbol {W} ^ {s}\\right) ^ {T} \\tag {2} \\\\ \\end{array}", + "image_path": "866af3ab9a4dec635f98e14c99ecbdcf0ee1bb2926b7eb2c8684e254c366c305.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 377, + 287, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 377, + 287, + 415 + ], + "spans": [ + { + "bbox": [ + 46, + 377, + 287, + 415 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 377, + 287, + 415 + ], + "type": "inline_equation", + "content": "\\pmb{p}^T\\pmb{W}^s \\in \\mathbb{R}^{1\\times D}" + }, + { + "bbox": [ + 46, + 377, + 287, + 415 + ], + "type": "text", + "content": " can be viewed as a new feature which is probability weighted sum of different class weights." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 422, + 287, + 436 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 422, + 287, + 436 + ], + "spans": [ + { + "bbox": [ + 47, + 422, + 287, + 436 + ], + "type": "text", + "content": "3.3. Class Relationship Embedded Contrastive Loss" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 441, + 287, + 489 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 441, + 287, + 489 + ], + "spans": [ + { + "bbox": [ + 46, + 441, + 287, + 489 + ], + "type": "text", + "content": "Inspired by the success of contrastive learning in different tasks [9, 18, 21, 26, 85], we propose to use our CRS for contrastive loss. Following previous work [9, 18], the contrastive loss can be presented as follows:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 83, + 493, + 287, + 544 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 83, + 493, + 287, + 544 + ], + "spans": [ + { + "bbox": [ + 83, + 493, + 287, + 544 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} h (\\pmb {p}, \\pmb {p} ^ {\\prime}) = \\exp (s _ {c r} (\\pmb {p}, \\pmb {p} ^ {\\prime}) / T _ {c o}), \\\\ \\ell_ {c o} = - \\log \\frac {h \\left(\\boldsymbol {p} , \\boldsymbol {p} _ {+}\\right)}{h \\left(\\boldsymbol {p} , \\boldsymbol {p} _ {+}\\right) + \\sum_ {\\boldsymbol {p} _ {-} \\in M} h \\left(\\boldsymbol {p} , \\boldsymbol {p} _ {-}\\right)}, \\tag {3} \\\\ \\end{array}", + "image_path": "d860592b0856bb4bfdbdb21c607ed830d33e833e94053c93f55194df46736cbf.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 549, + 287, + 609 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 549, + 287, + 609 + ], + "spans": [ + { + "bbox": [ + 46, + 549, + 287, + 609 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 549, + 287, + 609 + ], + "type": "inline_equation", + "content": "T_{co}" + }, + { + "bbox": [ + 46, + 549, + 287, + 609 + ], + "type": "text", + "content": " is a temperature parameter, and " + }, + { + "bbox": [ + 46, + 549, + 287, + 609 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 46, + 549, + 287, + 609 + ], + "type": "text", + "content": " denotes the exponent of scaled similarity. " + }, + { + "bbox": [ + 46, + 549, + 287, + 609 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 46, + 549, + 287, + 609 + ], + "type": "text", + "content": " is the memory bank [18] storing probabilities of negative samples. Since our CRS is built on probability, here we use probabilities as inputs instead of original features [18] or projected features [9]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 609, + 287, + 658 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 609, + 287, + 658 + ], + "spans": [ + { + "bbox": [ + 46, + 609, + 287, + 658 + ], + "type": "text", + "content": "With this unified form, we further instantiate it to two losses: class-aware contrastive loss and instance discrimination contrastive loss. The framework of our proposed method is shown in Figure 2." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 665, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 287, + 715 + ], + "type": "text", + "content": "Class-Aware Contrastive Loss To achieve class-aware learning, it is important to select proper samples and obtain more accurate pseudo labels for them. Inspired by previous efforts in SSL [67, 85] and UDA [41, 93], we consider" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 277, + 547, + 469 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 277, + 547, + 469 + ], + "spans": [ + { + "bbox": [ + 304, + 277, + 547, + 469 + ], + "type": "text", + "content": "constructing positive pairs from strongly and weakly augmented views of the same image. Specifically, given probabilities " + }, + { + "bbox": [ + 304, + 277, + 547, + 469 + ], + "type": "inline_equation", + "content": "p_s" + }, + { + "bbox": [ + 304, + 277, + 547, + 469 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 277, + 547, + 469 + ], + "type": "inline_equation", + "content": "\\tilde{p}_w" + }, + { + "bbox": [ + 304, + 277, + 547, + 469 + ], + "type": "text", + "content": " (tilde superscript denotes output by the teacher model) of strongly and weakly augmented views, we first compare " + }, + { + "bbox": [ + 304, + 277, + 547, + 469 + ], + "type": "inline_equation", + "content": "Max(\\tilde{p}_w)" + }, + { + "bbox": [ + 304, + 277, + 547, + 469 + ], + "type": "text", + "content": " with a predefined threshold " + }, + { + "bbox": [ + 304, + 277, + 547, + 469 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 304, + 277, + 547, + 469 + ], + "type": "text", + "content": " to select high-confident samples. Due to domain shift, instead of directly using " + }, + { + "bbox": [ + 304, + 277, + 547, + 469 + ], + "type": "inline_equation", + "content": "argmax(p_w)" + }, + { + "bbox": [ + 304, + 277, + 547, + 469 + ], + "type": "text", + "content": " as pseudo label, we adopt the clustering-based pseudo-labeling method in SHOT [42]. Based on pseudo label, we can get the positive samples " + }, + { + "bbox": [ + 304, + 277, + 547, + 469 + ], + "type": "inline_equation", + "content": "p^*" + }, + { + "bbox": [ + 304, + 277, + 547, + 469 + ], + "type": "text", + "content": " with the one-hot form. For the negative samples " + }, + { + "bbox": [ + 304, + 277, + 547, + 469 + ], + "type": "inline_equation", + "content": "p_-" + }, + { + "bbox": [ + 304, + 277, + 547, + 469 + ], + "type": "text", + "content": ", we use one-hot probabilities of all classes except the pseudo label class. Therefore, both positive and negative samples (i.e. " + }, + { + "bbox": [ + 304, + 277, + 547, + 469 + ], + "type": "inline_equation", + "content": "p^*" + }, + { + "bbox": [ + 304, + 277, + 547, + 469 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 277, + 547, + 469 + ], + "type": "inline_equation", + "content": "p_-" + }, + { + "bbox": [ + 304, + 277, + 547, + 469 + ], + "type": "text", + "content": ") are generated from the pseudo label of weak target sample through teacher model. Finally, the Class Relationship embedded Class-Aware Contrastive (CR-CACo) loss is presented as" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 312, + 478, + 545, + 514 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 478, + 545, + 514 + ], + "spans": [ + { + "bbox": [ + 312, + 478, + 545, + 514 + ], + "type": "interline_equation", + "content": "\\ell_ {c r \\_ c a c o} = - \\log \\frac {h \\left(\\boldsymbol {p} _ {s} , \\boldsymbol {p} ^ {*}\\right)}{h \\left(\\boldsymbol {p} _ {s} , \\boldsymbol {p} ^ {*}\\right) + \\sum_ {\\boldsymbol {p} _ {-} \\in M _ {c a c o}} h \\left(\\boldsymbol {p} _ {s} , \\boldsymbol {p} _ {-}\\right)}, \\tag {4}", + "image_path": "4e103b7d4244a3bbfc0a995d3308b5612b0823a3c1492300fdd2d0bf627d8834.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 521, + 545, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 521, + 545, + 557 + ], + "spans": [ + { + "bbox": [ + 304, + 521, + 545, + 557 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 521, + 545, + 557 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 304, + 521, + 545, + 557 + ], + "type": "text", + "content": " follows the definition in Equation (3), " + }, + { + "bbox": [ + 304, + 521, + 545, + 557 + ], + "type": "inline_equation", + "content": "M_{caco}" + }, + { + "bbox": [ + 304, + 521, + 545, + 557 + ], + "type": "text", + "content": " stores negative samples and is different for each sample according to its pseudo label." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 558, + 545, + 618 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 558, + 545, + 618 + ], + "spans": [ + { + "bbox": [ + 304, + 558, + 545, + 618 + ], + "type": "text", + "content": "Regardless of pseudo label generation, using only " + }, + { + "bbox": [ + 304, + 558, + 545, + 618 + ], + "type": "inline_equation", + "content": "s_{intra}" + }, + { + "bbox": [ + 304, + 558, + 545, + 618 + ], + "type": "text", + "content": " and replacing " + }, + { + "bbox": [ + 304, + 558, + 545, + 618 + ], + "type": "inline_equation", + "content": "p_s" + }, + { + "bbox": [ + 304, + 558, + 545, + 618 + ], + "type": "text", + "content": " with features and " + }, + { + "bbox": [ + 304, + 558, + 545, + 618 + ], + "type": "inline_equation", + "content": "p^*, p_-" + }, + { + "bbox": [ + 304, + 558, + 545, + 618 + ], + "type": "text", + "content": " with classifier weights, the above Equation (4) is actually the FixMatch loss [67]. Although FixMatch can improve the performance, our CR-CACo loss can greatly outperform it." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 618, + 546, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 618, + 546, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 618, + 546, + 714 + ], + "type": "text", + "content": "Our CR-CACo loss is more robust to label noise due to the class relationship embedded inter-class similarity. We compare our CRS with the traditional semi-supervised FixMatch loss. The results are shown in Figure 3. It can be seen that the simple IM loss [42] has lower entropy since it has an entropy minimization term. When FixMatch loss is used, the entropy will increase since the class relationship is learned through high-confidence target samples. Our CRS" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7622" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 55, + 68, + 280, + 218 + ], + "blocks": [ + { + "bbox": [ + 55, + 68, + 280, + 218 + ], + "lines": [ + { + "bbox": [ + 55, + 68, + 280, + 218 + ], + "spans": [ + { + "bbox": [ + 55, + 68, + 280, + 218 + ], + "type": "image", + "image_path": "4f671f0b27492789c3e7ac64a9867ab38a621a4cf81ed33177ba8713d682b514.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 221, + 287, + 255 + ], + "lines": [ + { + "bbox": [ + 46, + 221, + 287, + 255 + ], + "spans": [ + { + "bbox": [ + 46, + 221, + 287, + 255 + ], + "type": "text", + "content": "Figure 3. Entropy density plots under SUDA setting Office-Home " + }, + { + "bbox": [ + 46, + 221, + 287, + 255 + ], + "type": "inline_equation", + "content": "\\mathrm{Rw} \\rightarrow \\mathrm{Cl}" + }, + { + "bbox": [ + 46, + 221, + 287, + 255 + ], + "type": "text", + "content": ". Our class relationship embedded similarity " + }, + { + "bbox": [ + 46, + 221, + 287, + 255 + ], + "type": "inline_equation", + "content": "s_{cr}" + }, + { + "bbox": [ + 46, + 221, + 287, + 255 + ], + "type": "text", + "content": " has higher entropy but better accuracy. Best viewed in color." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 267, + 288, + 304 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 267, + 288, + 304 + ], + "spans": [ + { + "bbox": [ + 46, + 267, + 288, + 304 + ], + "type": "text", + "content": "has larger entropy since we explicitly consider the class relationship and can alleviate the negative impact of label noise." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 319, + 287, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 319, + 287, + 498 + ], + "spans": [ + { + "bbox": [ + 46, + 319, + 287, + 498 + ], + "type": "text", + "content": "Instance Discrimination Contrastive Loss The instance discrimination contrastive (IDCo) loss has shown remarkable success in self-supervised learning [9, 18]. It can learn discriminative features by regrading the two views of the same sample as positive pairs and all other samples as negative samples. The traditional IDCo loss uses features as input, but it only makes the feature discriminative and cannot guarantee that the learned feature will be close to corresponding prototypes (i.e. classifier weights). Conducting IDCo loss in the output space (i.e. computing similarity using " + }, + { + "bbox": [ + 46, + 319, + 287, + 498 + ], + "type": "inline_equation", + "content": "s_{\\text{intra}}" + }, + { + "bbox": [ + 46, + 319, + 287, + 498 + ], + "type": "text", + "content": ") can alleviate this problem [35], but the more informative relationship between samples in the original feature space can not be fully expressed. This can be addressed by our CRS since we embed the source prior class similarity and can reflect more accurate relationship of samples." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 499, + 287, + 571 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 499, + 287, + 571 + ], + "spans": [ + { + "bbox": [ + 46, + 499, + 287, + 571 + ], + "type": "text", + "content": "Here we use the probabilities " + }, + { + "bbox": [ + 46, + 499, + 287, + 571 + ], + "type": "inline_equation", + "content": "\\pmb{p}_{s1}" + }, + { + "bbox": [ + 46, + 499, + 287, + 571 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 499, + 287, + 571 + ], + "type": "inline_equation", + "content": "\\tilde{\\pmb{p}}_{s2}'" + }, + { + "bbox": [ + 46, + 499, + 287, + 571 + ], + "type": "text", + "content": " of two strong augmented views from the same image as positive pairs, and probabilities of all other samples in the same batch as negative samples. Finally, the Class Relationship embedded Instance Discrimination Contrastive (CR-IDCo) loss is presented as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 581, + 286, + 626 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 581, + 286, + 626 + ], + "spans": [ + { + "bbox": [ + 53, + 581, + 286, + 626 + ], + "type": "interline_equation", + "content": "\\ell_ {c r. i d c o} = - \\log \\frac {h \\left(\\boldsymbol {p} _ {s 1} , \\tilde {\\boldsymbol {p}} _ {s 2} ^ {\\prime}\\right)}{h \\left(\\boldsymbol {p} _ {s 1} , \\tilde {\\boldsymbol {p}} _ {s 2} ^ {\\prime}\\right) + \\sum_ {\\tilde {\\boldsymbol {p}} _ {-} \\in M _ {i d c o}} h \\left(\\boldsymbol {p} _ {s 1} , \\tilde {\\boldsymbol {p}} _ {-}\\right)}, \\tag {5}", + "image_path": "cd0876724b8d6145399dc0719d71dff9e2f35962d3d5ec4f2217f4292024d2d9.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 627, + 287, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 627, + 287, + 662 + ], + "spans": [ + { + "bbox": [ + 46, + 627, + 287, + 662 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 627, + 287, + 662 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 46, + 627, + 287, + 662 + ], + "type": "text", + "content": " follows the definition in Equation (3), " + }, + { + "bbox": [ + 46, + 627, + 287, + 662 + ], + "type": "inline_equation", + "content": "M_{idco}" + }, + { + "bbox": [ + 46, + 627, + 287, + 662 + ], + "type": "text", + "content": " contains the probabilities of other samples within the same batch." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 670, + 196, + 684 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 670, + 196, + 684 + ], + "spans": [ + { + "bbox": [ + 47, + 670, + 196, + 684 + ], + "type": "text", + "content": "3.4. Overall Training Objective" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 689, + 288, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 689, + 288, + 712 + ], + "spans": [ + { + "bbox": [ + 46, + 689, + 288, + 712 + ], + "type": "text", + "content": "The proposed two losses are complementary to each other and can be trained end-to-end. Since our CR-CACo" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 72, + 547, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 205 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 205 + ], + "type": "text", + "content": "uses a high prob-threshold (i.e. 0.95) to select confident samples, it should be combined with other methods that can generate confident predictions. Otherwise, it will be dominated by noise labels in the early training stage. Although it performs normally when combined with our CR-IDCo loss, both of them focus on transferring source class relationship, and other methods that encourage confident predictions are still needed and complementary to ours. As a result, we combine our method with existing representative methods (i.e. SHOT-IM [42], AaD [90]) to validate the effectiveness. As a result, the overall loss is presented as:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 321, + 214, + 545, + 228 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 214, + 545, + 228 + ], + "spans": [ + { + "bbox": [ + 321, + 214, + 545, + 228 + ], + "type": "interline_equation", + "content": "\\ell_ {\\text {o u r s}} = \\ell_ {\\text {b a s e l i n e}} + \\lambda_ {\\text {c a c o}} \\ell_ {\\text {c r} - \\text {c a c o}} + \\lambda_ {\\text {i d c o}} \\ell_ {\\text {c r} - \\text {i d c o}} \\tag {6}", + "image_path": "de14eb903e53eed020ab8d3dc778a760b23227f02d63d2e4f744a4d10cc87abb.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 237, + 547, + 309 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 237, + 547, + 309 + ], + "spans": [ + { + "bbox": [ + 304, + 237, + 547, + 309 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 237, + 547, + 309 + ], + "type": "inline_equation", + "content": "\\lambda_{caco},\\lambda_{idco}" + }, + { + "bbox": [ + 304, + 237, + 547, + 309 + ], + "type": "text", + "content": " are the trade-off hyperparameters. " + }, + { + "bbox": [ + 304, + 237, + 547, + 309 + ], + "type": "inline_equation", + "content": "l_{baseline}" + }, + { + "bbox": [ + 304, + 237, + 547, + 309 + ], + "type": "text", + "content": " is the loss of baseline (i.e., IM or AaD). It contains two terms: discriminability and diversity. The former produces high-confident outputs, and the latter avoids collapse of predictions. IM uses entropy of single sample and average of all samples, and AaD uses neighbor information." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 319, + 388, + 334 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 319, + 388, + 334 + ], + "spans": [ + { + "bbox": [ + 305, + 319, + 388, + 334 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 340, + 422, + 354 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 340, + 422, + 354 + ], + "spans": [ + { + "bbox": [ + 306, + 340, + 422, + 354 + ], + "type": "text", + "content": "4.1. Experimental Setup" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 358, + 547, + 526 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 358, + 547, + 526 + ], + "spans": [ + { + "bbox": [ + 304, + 358, + 547, + 526 + ], + "type": "text", + "content": "Datasets. We evaluate the effectiveness of our approach on four standard DA benchmarks. Office-31 [63] benchmark consists of three domains in office environments: Amazon (A), DSLR (D), and Webcam (W), each with 31 categories. Office-Home [73] is a more challenging dataset. It comprises of images of commonplace objects divided into four domains: Artistic (Ar), Clipart (Cl), Product (Pr), and Real-World (Rw), each with 65 classes. VisDA [58] is a large-scale dataset for synthetic-to-real domain adaptation. The source domain has 152,397 synthetic images, while the target domain has 55,388 real-world images. DomainNet [57] is the most challenging dataset involving 6 domains: Clipart (C), Real (R), Infograph (I), Painting (P), Sketch (S) and Quickdraw (Q) with 345 classes each." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 533, + 547, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 533, + 547, + 665 + ], + "spans": [ + { + "bbox": [ + 304, + 533, + 547, + 665 + ], + "type": "text", + "content": "Implementation details. To ensure fair comparison with related methods, we adopt the backbone of a ResNet-50 [19] for Office-31, Office-Home, and DomainNet, and a ResNet-101 for VisDA. We use the same network architecture as SHOT [42] where the final part of the network is changed to suit the SFUDA task. For the classifier, we find that fixing it or not have little impact on the final performance, and we choose to fix it in all experiments. We adopt SGD with momentum 0.9 and batch size 64 for all datasets where each image has three views (i.e. one weak and two strong augmentations)." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 666, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 666, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 666, + 547, + 713 + ], + "type": "text", + "content": "For the hyperparameters, we empirically set the probability threshold " + }, + { + "bbox": [ + 304, + 666, + 547, + 713 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 304, + 666, + 547, + 713 + ], + "type": "text", + "content": " to 0.95 following FixMatch [67]. We find that " + }, + { + "bbox": [ + 304, + 666, + 547, + 713 + ], + "type": "inline_equation", + "content": "\\tau = 0.95" + }, + { + "bbox": [ + 304, + 666, + 547, + 713 + ], + "type": "text", + "content": " works well across all settings and tasks. For the temperature " + }, + { + "bbox": [ + 304, + 666, + 547, + 713 + ], + "type": "inline_equation", + "content": "T_{co}" + }, + { + "bbox": [ + 304, + 666, + 547, + 713 + ], + "type": "text", + "content": " in both contrastive losses," + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7623" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 47, + 81, + 547, + 286 + ], + "blocks": [ + { + "bbox": [ + 171, + 68, + 422, + 79 + ], + "lines": [ + { + "bbox": [ + 171, + 68, + 422, + 79 + ], + "spans": [ + { + "bbox": [ + 171, + 68, + 422, + 79 + ], + "type": "text", + "content": "Table 1. Single-Source Unsupervised DA (SUDA) on Office-Home." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 47, + 81, + 547, + 286 + ], + "lines": [ + { + "bbox": [ + 47, + 81, + 547, + 286 + ], + "spans": [ + { + "bbox": [ + 47, + 81, + 547, + 286 + ], + "type": "table", + "html": "
MethodSFOffice-Home
Ar→ClAr→PrAr→RwCl→ArCl→PrCl→RwPr→ArPr→ClPr→RwRw→ArRw→ClRw→PrAvg
SENTRY (ICCV'21) [59]X61.877.480.166.371.674.766.863.080.974.066.384.172.2
FixBi (CVPR'21) [52]X58.177.380.467.779.578.165.857.981.776.462.986.772.7
SCDA (ICCV'21) [38]X60.776.482.869.877.578.468.959.082.774.961.884.573.1
A²Net (ICCV'21) [79]✓58.479.082.467.579.378.968.056.282.974.160.585.072.8
NRC (NeurIPS'21) [87]✓57.780.382.068.179.878.665.356.483.071.058.685.672.2
D-MCD (AAAI'22) [10]✓59.478.980.267.279.378.665.355.682.273.362.883.972.2
DIPE (CVPR'22) [74]✓56.579.280.770.179.878.867.955.183.574.159.384.872.5
Sub-Sup (ECCV'22) [29]✓61.080.482.569.179.979.569.157.882.774.565.186.474.0
BMD (ECCV'22) [61]✓58.179.782.669.381.080.770.857.683.674.060.085.973.6
U-SFAN (ECCV'22) [62]✓57.877.881.667.977.379.267.254.781.273.360.383.971.9
CoWA-JMDS (ICML'22) [33]✓56.978.481.069.180.079.967.757.282.472.860.584.572.5
Feat-Mixup (ICML'22) [30]✓61.881.283.068.580.679.467.861.585.173.764.186.574.5
VMP (NeurIPS'22) [25]✓57.977.682.568.679.480.668.455.683.175.259.684.772.8
DaC (NeurIPS'22) [98]✓59.179.581.269.378.979.267.456.482.474.061.484.472.8
SHOT-IM (ICML'20) [42]✓55.476.680.466.974.375.465.654.880.773.758.483.470.5
Ours+ SHOT-IM✓62.882.084.370.980.882.670.061.183.676.265.187.075.5 (+5.0)
AaD (NeurIPS'22) [90]✓59.379.382.168.979.879.567.257.483.172.158.585.472.7
Ours+ AaD✓63.582.185.073.082.782.469.562.982.674.265.787.375.9 (+3.2)
", + "image_path": "c87709867c41b2deef8453b51d8fe983bbefa6b1e63235f9c73fa354e8cadb60.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 48, + 327, + 287, + 533 + ], + "blocks": [ + { + "bbox": [ + 47, + 297, + 287, + 319 + ], + "lines": [ + { + "bbox": [ + 47, + 297, + 287, + 319 + ], + "spans": [ + { + "bbox": [ + 47, + 297, + 287, + 319 + ], + "type": "text", + "content": "Table 2. Single-Source Unsupervised Domain Adaptation (SUDA) on Office-31 and VisDA benchmarks." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 48, + 327, + 287, + 533 + ], + "lines": [ + { + "bbox": [ + 48, + 327, + 287, + 533 + ], + "spans": [ + { + "bbox": [ + 48, + 327, + 287, + 533 + ], + "type": "table", + "html": "
MethodSFOffice-31VisDA
A→DA→WD→WW→DD→AW→A AvgS→R
BCDM (AAAI'20) [37]✗93.895.498.6100.073.173.089.083.4
MCC (ECCV'20) [24]✗94.495.598.6100.072.974.989.478.8
FixBi (CVPR'21) [52]✗95.096.199.3100.078.779.491.487.2
RADA (ICCV'21) [23]✗96.196.299.3100.077.577.491.176.3
FAA (ICCV'21) [22]✗94.492.399.299.780.578.790.882.7
SCDA (ICCV'21) [38]✗95.495.399.0100.077.275.990.5-
SHOT (ICML'20) [42]✓94.090.198.499.974.774.388.682.9
CPGA (IJCAI'21) [60]✓94.494.198.499.876.076.689.984.1
VDM-DA (TCSVT'21) [71]✓93.294.198.0100.075.877.189.785.1
A2Net (ICCV'21) [79]✓94.594.099.2100.076.776.190.184.3
HCL (NeurIPS'21) [21]✓90.891.398.2100.072.772.787.683.5
NRC (NeurIPS'21) [87]✓96.090.899.0100.075.375.089.485.9
SHOT++ (TPAMI'21) [43]✓94.390.498.799.976.275.889.287.3
D-MCD (AAAI'22) [10]✓94.193.598.8100.076.476.489.987.5
DIPE (CVPR'22) [74]✓96.693.198.499.675.577.290.183.1
Sub-Sup (ECCV'22) [29]✓95.694.699.299.877.077.790.788.2
BMD (ECCV'22) [61]✓96.294.298.0100.076.076.090.188.7
CoWA-JMDS (ICML'22) [33]✓94.495.298.5100.076.277.690.386.9
Feat-Mixup (ICML'22) [30]✓94.693.298.9100.078.378.990.787.8
VMP (NeurIPS'22) [25]✓93.396.298.6100.075.476.990.0-
SHOT-IM (ICML'20) [42]✓90.691.298.399.972.571.487.380.4
Ours + SHOT-IM✓95.895.199.0100.076.678.390.8(+3.5)89.1(+8.7)
AaD (NeurIPS'22) [90]✓96.492.199.1100.075.076.589.988.0
Ours + AaD✓96.695.599.1100.076.978.391.1(+1.2)89.6(+1.6)
", + "image_path": "b994a1ed63dca05df72116415ba61cbb956dbcf77fd335155d9c9c321342a446.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 552, + 287, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 552, + 287, + 613 + ], + "spans": [ + { + "bbox": [ + 46, + 552, + 287, + 613 + ], + "type": "text", + "content": "we set it to 0.07 following previous contrastive learning methods [18, 21, 85]. For the " + }, + { + "bbox": [ + 46, + 552, + 287, + 613 + ], + "type": "inline_equation", + "content": "\\lambda_{inter}" + }, + { + "bbox": [ + 46, + 552, + 287, + 613 + ], + "type": "text", + "content": " in our similarity, we use 1.0 for simplicity. For the trade-off hyperparameters " + }, + { + "bbox": [ + 46, + 552, + 287, + 613 + ], + "type": "inline_equation", + "content": "\\lambda_{caco}, \\lambda_{idco}" + }, + { + "bbox": [ + 46, + 552, + 287, + 613 + ], + "type": "text", + "content": ", we directly set them to 1.0. More training details can be found in the supplementary material." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 631, + 105, + 643 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 631, + 105, + 643 + ], + "spans": [ + { + "bbox": [ + 47, + 631, + 105, + 643 + ], + "type": "text", + "content": "4.2. Results" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "type": "text", + "content": "The results are shown in Table 1, Table 2, Table 3 and Table 4. In each table, SF indicates source-free, * indicates results we reproduced from the released code. (+x.x) indicates gains over SHOT-IM [42] and AaD [90] respectively. Ours means combination of our proposed two losses." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 304, + 545, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 304, + 545, + 411 + ], + "spans": [ + { + "bbox": [ + 304, + 304, + 545, + 411 + ], + "type": "text", + "content": "Single Source Domain Adaptation Table 1 and Table 2 show the classification accuracy for single-source DA on each dataset: Office-Home, Office-31, and VisDA. Based on the simple SHOT-IM and stronger AaD, our method can achieve state-of-the-art performance. UDA methods BCDM [37] and MCC [24] also explore the inter-class relationship by directly suppressing the inter-class term to obtain determined outputs. Our method surpasses these methods even without source domain supervision." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 412, + 546, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 412, + 546, + 533 + ], + "spans": [ + { + "bbox": [ + 304, + 412, + 546, + 533 + ], + "type": "text", + "content": "The two proposed contrastive losses exploit different views of the same image. Thus our method is complementary to those not utilizing information of different views. Both baselines only consider weak views, and so our method can boost them. AaD surpasses IM by regarding local neighbors as positive pairs. In our CR-IDCo loss, the strong views of the same image can be regarded as neighbors since they are very close in feature space, which has overlapped effect with AaD. When our method is combined with AaD, therefore, the improvement is relatively smaller." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 545, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 545, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 545, + 547, + 715 + ], + "type": "text", + "content": "Multi Source Domain Adaptation Table 3 shows the results for multi-source DA on Office-Home and the largescale DomainNet benchmark. We adopt two baselines without using domain labels, that is, all data from different source domains is treated as belonging to a single domain and only one source model is trained and transferred. We observe improvements of " + }, + { + "bbox": [ + 304, + 545, + 547, + 715 + ], + "type": "inline_equation", + "content": "4.1\\%" + }, + { + "bbox": [ + 304, + 545, + 547, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 545, + 547, + 715 + ], + "type": "inline_equation", + "content": "3.4\\%" + }, + { + "bbox": [ + 304, + 545, + 547, + 715 + ], + "type": "text", + "content": " in DomainNet, " + }, + { + "bbox": [ + 304, + 545, + 547, + 715 + ], + "type": "inline_equation", + "content": "4.5\\%" + }, + { + "bbox": [ + 304, + 545, + 547, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 545, + 547, + 715 + ], + "type": "inline_equation", + "content": "3.0\\%" + }, + { + "bbox": [ + 304, + 545, + 547, + 715 + ], + "type": "text", + "content": " in Office-Home. CAiDA [14] considers class relationship by enforcing consistency of different source model outputs. It achieves " + }, + { + "bbox": [ + 304, + 545, + 547, + 715 + ], + "type": "inline_equation", + "content": "76.2\\%" + }, + { + "bbox": [ + 304, + 545, + 547, + 715 + ], + "type": "text", + "content": " on Office-Home by combining IM loss, a novel confident-anchor-induced pseudo label method and the class-relation-space-aware consistency loss. Based on the same IM loss, ours achieves " + }, + { + "bbox": [ + 304, + 545, + 547, + 715 + ], + "type": "inline_equation", + "content": "78.1\\%" + }, + { + "bbox": [ + 304, + 545, + 547, + 715 + ], + "type": "text", + "content": " and exceeds it by " + }, + { + "bbox": [ + 304, + 545, + 547, + 715 + ], + "type": "inline_equation", + "content": "1.9\\%" + }, + { + "bbox": [ + 304, + 545, + 547, + 715 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7624" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 48, + 79, + 545, + 218 + ], + "blocks": [ + { + "bbox": [ + 108, + 68, + 484, + 79 + ], + "lines": [ + { + "bbox": [ + 108, + 68, + 484, + 79 + ], + "spans": [ + { + "bbox": [ + 108, + 68, + 484, + 79 + ], + "type": "text", + "content": "Table 3. Multi-Source Unsupervised Domain Adaptation (MUDA) on DomainNet and Office-Home." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 48, + 79, + 545, + 218 + ], + "lines": [ + { + "bbox": [ + 48, + 79, + 545, + 218 + ], + "spans": [ + { + "bbox": [ + 48, + 79, + 545, + 218 + ], + "type": "table", + "html": "
MethodSFw/o Domain \nLabelsDomainNetOffice-Home
→C→I→P→Q→R→SAvg→Ar→Cl→Pr→RwAvg
MCC (ECCV'20) [24]XX65.526.056.616.568.052.747.6-----
CMSDA (BMVC'21) [66]XX70.926.557.521.368.159.450.471.567.784.182.976.6
DRT (CVPR'21) [39]XX71.031.661.012.371.460.751.3-----
STEM (ICCV'21) [53]XX72.028.261.525.772.660.253.4-----
SourcecombineX✓57.023.454.114.667.250.344.458.057.374.277.966.9
DECISION (CVPR'21) [1]✓X61.521.654.618.967.551.045.974.559.484.483.675.5
CAiDA (NeurIPS'21) [14]✓X-------75.260.584.784.276.2
Sub-Sup (ECCV'22) [29]✓✓70.325.757.317.169.957.149.675.164.186.684.477.6
Feat-Mixup (ICML'22) [30]✓✓75.424.657.823.665.858.551.072.667.485.983.677.4
SHOT-IM* (ICML'20) [42]✓✓64.524.255.215.567.053.246.671.558.681.982.373.6
Ours + SHOT-IM✓✓69.827.059.021.270.257.550.7(+4.1)74.966.485.185.878.1 (+4.5)
AaD* (NeurIPS'22) [90]✓✓65.725.556.417.068.154.647.972.160.685.384.475.6
Ours +AaD✓✓71.227.859.322.569.357.851.3(+3.4)75.466.986.485.978.6 (+3.0)
", + "image_path": "0b1ccb0ca9182052a039c3f2aed7f994245ed89173fc36247861f0b78a8f283b.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 48, + 256, + 284, + 364 + ], + "blocks": [ + { + "bbox": [ + 47, + 231, + 287, + 252 + ], + "lines": [ + { + "bbox": [ + 47, + 231, + 287, + 252 + ], + "spans": [ + { + "bbox": [ + 47, + 231, + 287, + 252 + ], + "type": "text", + "content": "Table 4. Partial-set and open-set Domain Adaptation (PDA and ODA) on Office-Home." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 48, + 256, + 284, + 364 + ], + "lines": [ + { + "bbox": [ + 48, + 256, + 284, + 364 + ], + "spans": [ + { + "bbox": [ + 48, + 256, + 284, + 364 + ], + "type": "table", + "html": "
Partial-set DASF Avg.Open-set DASF Avg.
ResNet-50 [19]× 61.3ResNet [19]× 65.3
IWAN (CVPR'18) [92]× 63.6OpenMax (CVPR'16) [4]× 66.7
SAN (CVPR'18) [5]× 65.3ATI-λ (ICCV'17) [56]× 66.1
ETN (CVPR'19) [6]× 70.5OSBP (ECCV'18) [65]× 65.7
SAFN (ICCV'19) [82]× 71.8STA (CVPR'19) [44]× 69.5
Source model only✓ 62.8Source model only✓ 46.6
SHOT (ICML'20) [42]✓ 79.3SHOT (ICML'20) [42]✓ 72.8
SHOT+HCL (NeurIPS'21) [21]✓ 80.1SHOT+HCL (NeurIPS'21) [21]✓ 73.2
CoWA-JMDS (ICML'22) [33]✓ 83.2CoWA-JMDS (ICML'22) [33]✓ 73.2
SHOT-IM (ICML'20) [42]✓ 76.8SHOT-IM (ICML'20) [42]✓ 71.5
Ours +SHOT-IM✓ 80.6 (+3.8)Ours +SHOT-IM✓ 73.2(+1.7)
AaD* (NeurIPS'22) [90]✓ 79.7AaD* (NeurIPS'22) [90]✓ 71.8
Ours +AaD✓ 82.4 (+2.7)Ours +AaD✓ 73.3(+1.5)
", + "image_path": "f962328e11faae001b096f914b93baa83f19c8543ac910e3a8fd455ac3146ee1.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 374, + 287, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 374, + 287, + 460 + ], + "spans": [ + { + "bbox": [ + 46, + 374, + 287, + 460 + ], + "type": "text", + "content": "Domain Adaptation Beyond Vanilla Closed-set We provide additional results under source-free partial-set and open-set DA (PDA and ODA) setting on Office-Home. For open-set detection in ODA, we follow the same protocol to detect unseen categories as SHOT [42]. Results are shown in Table 4. It can be seen that our method can achieve performance comparable to that of the state-of-the-arts." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 465, + 111, + 478 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 465, + 111, + 478 + ], + "spans": [ + { + "bbox": [ + 47, + 465, + 111, + 478 + ], + "type": "text", + "content": "4.3. Analysis" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 483, + 288, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 483, + 288, + 639 + ], + "spans": [ + { + "bbox": [ + 46, + 483, + 288, + 639 + ], + "type": "text", + "content": "Component-wise ablations Here we validate the effectiveness of different proposed losses. As described in Sec. 3.4, we do not use CR-CACo loss alone. The results are shown in Table 5. It can be seen that when using our method alone (line #2) can achieve comparable performance. When combined with SHOT-IM, our both losses are complementary, and the best performance can be achieved by combining baseline and our losses. Note that although CR-IDCo (line #1) outperforms IM (line #0), it is worse than IM when combined with CR-CACo (i.e. line #2 < line #3). This is caused by the fact that CR-IDCo cannot produce enough confident samples as IM for the learning of CR-CACo loss." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 653, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 287, + 715 + ], + "type": "text", + "content": "Effect of pseudo labels in CR-CACo In our proposed CR-CACo loss, the pseudo label is generated from clustering-based method in SHOT [42] instead of predictions of weakly augmented images. The results of the comparison are shown in Table 6. It can be seen that our" + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 308, + 261, + 544, + 335 + ], + "blocks": [ + { + "bbox": [ + 305, + 231, + 545, + 254 + ], + "lines": [ + { + "bbox": [ + 305, + 231, + 545, + 254 + ], + "spans": [ + { + "bbox": [ + 305, + 231, + 545, + 254 + ], + "type": "text", + "content": "Table 5. Component-wise ablation studies of the proposed methods under the SUDA setting." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 308, + 261, + 544, + 335 + ], + "lines": [ + { + "bbox": [ + 308, + 261, + 544, + 335 + ], + "spans": [ + { + "bbox": [ + 308, + 261, + 544, + 335 + ], + "type": "table", + "html": "
#IMCR-CACoCR-IDCoOffice-31Office-HomeVisDA
0✓87.370.580.4
1✓89.272.583.0
2✓✓89.674.085.3
3✓✓89.874.887.0
4✓✓90.274.385.8
5✓✓✓90.875.589.1
", + "image_path": "15d144692b73da0e54e1c931d2647ab5c0ccae4334c670002dfe64ef9f0a4726.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 356, + 545, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 356, + 545, + 392 + ], + "spans": [ + { + "bbox": [ + 305, + 356, + 545, + 392 + ], + "type": "text", + "content": "CR-CACo can bring significant improvement when using pseudo label from model prediction. The clustering based pseudo-label can bring slight but consistent improvements." + } + ] + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 308, + 415, + 544, + 504 + ], + "blocks": [ + { + "bbox": [ + 323, + 400, + 527, + 412 + ], + "lines": [ + { + "bbox": [ + 323, + 400, + 527, + 412 + ], + "spans": [ + { + "bbox": [ + 323, + 400, + 527, + 412 + ], + "type": "text", + "content": "Table 6. Comparison of different pseudo label methods." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 308, + 415, + 544, + 504 + ], + "lines": [ + { + "bbox": [ + 308, + 415, + 544, + 504 + ], + "spans": [ + { + "bbox": [ + 308, + 415, + 544, + 504 + ], + "type": "table", + "html": "
#BaselinePseudo-label in CR-CACoOffice-31Office-HomeVisDA
PredictionClustering
087.370.580.4
1IM✓89.6(+2.3)74.5(+4.0)86.6(+6.2)
2✓89.8(+2.5)74.8(+4.3)87.0(+6.6)
389.972.788.0
4AaD✓90.6(+0.7)75.0(+2.3)88.9(+0.9)
5✓90.7(+0.8)75.2(+2.5)89.1(+1.1)
", + "image_path": "720436a42396060a1f628f55f06242983d3f1b94f2e6db180be5739a245a69d1.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_body" + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 525, + 545, + 717 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 525, + 545, + 717 + ], + "spans": [ + { + "bbox": [ + 304, + 525, + 545, + 717 + ], + "type": "text", + "content": "Different similarities and disentangled improvements. To validate the superiority of our class relationship embedded similarity in contrastive loss, we compare it with other two choices. The first one uses only " + }, + { + "bbox": [ + 304, + 525, + 545, + 717 + ], + "type": "inline_equation", + "content": "s_{intra}" + }, + { + "bbox": [ + 304, + 525, + 545, + 717 + ], + "type": "text", + "content": " based on features, and the second one uses only " + }, + { + "bbox": [ + 304, + 525, + 545, + 717 + ], + "type": "inline_equation", + "content": "s_{intra}" + }, + { + "bbox": [ + 304, + 525, + 545, + 717 + ], + "type": "text", + "content": " based on probabilities which equals " + }, + { + "bbox": [ + 304, + 525, + 545, + 717 + ], + "type": "inline_equation", + "content": "\\lambda_{inter} = 0" + }, + { + "bbox": [ + 304, + 525, + 545, + 717 + ], + "type": "text", + "content": " in our case. The results are shown in Table 7. Using features (i.e. i.e. " + }, + { + "bbox": [ + 304, + 525, + 545, + 717 + ], + "type": "inline_equation", + "content": "s_{intra}" + }, + { + "bbox": [ + 304, + 525, + 545, + 717 + ], + "type": "text", + "content": " w/ feature) rather than probabilities is a more natural way. Indeed, using features in CACo Loss would result in FixMatch loss, and using features in IDCo loss would fall in the traditional self-supervised way (e.g., MoCo). Only using the probability based intra-class term (i.e. " + }, + { + "bbox": [ + 304, + 525, + 545, + 717 + ], + "type": "inline_equation", + "content": "s_{intra}" + }, + { + "bbox": [ + 304, + 525, + 545, + 717 + ], + "type": "text", + "content": " w/ probability) is still special for similarity in contrastive losses. And it can bring consistent improvements. In both CACo and IDCo losses, our " + }, + { + "bbox": [ + 304, + 525, + 545, + 717 + ], + "type": "inline_equation", + "content": "s_{cr}" + }, + { + "bbox": [ + 304, + 525, + 545, + 717 + ], + "type": "text", + "content": " can achieve the best result, thus validating the effectiveness of adopting source prior class similarities." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7625" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 50, + 85, + 285, + 172 + ], + "blocks": [ + { + "bbox": [ + 80, + 71, + 253, + 81 + ], + "lines": [ + { + "bbox": [ + 80, + 71, + 253, + 81 + ], + "spans": [ + { + "bbox": [ + 80, + 71, + 253, + 81 + ], + "type": "text", + "content": "Table 7. Comparison with different similarities." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 50, + 85, + 285, + 172 + ], + "lines": [ + { + "bbox": [ + 50, + 85, + 285, + 172 + ], + "spans": [ + { + "bbox": [ + 50, + 85, + 285, + 172 + ], + "type": "table", + "html": "
#LossSimilarityOffice-31Office-HomeVisDA
0IM—87.370.580.4
1IM+CACosintra w/ feature88.972.583.5
2sintra w/ probability89.172.984.0
3scr w/ probability89.874.887.0
4IM+IDCosintra w/ feature87.570.980.8
5sintra w/ probability88.271.481.5
6scr w/ probability90.274.385.8
", + "image_path": "6e4869d08700b71dfd0cd985adbe7be697cdc04a3ca686a873cf159aada6e5d9.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 179, + 287, + 347 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 179, + 287, + 347 + ], + "spans": [ + { + "bbox": [ + 46, + 179, + 287, + 347 + ], + "type": "text", + "content": "Comparison with fixed coefficient within " + }, + { + "bbox": [ + 46, + 179, + 287, + 347 + ], + "type": "inline_equation", + "content": "s_{inter}" + }, + { + "bbox": [ + 46, + 179, + 287, + 347 + ], + "type": "text", + "content": " In the definition of inter-class term in Equation (1), we use the source class similarity matrix " + }, + { + "bbox": [ + 46, + 179, + 287, + 347 + ], + "type": "inline_equation", + "content": "A^s" + }, + { + "bbox": [ + 46, + 179, + 287, + 347 + ], + "type": "text", + "content": " as coefficient. Instead of this prior knowledge, we can also set the coefficient to a small fixed value. Considering the positive pairs in contrastive loss, the negative coefficient (i.e. -0.1) works to suppress the inter-class term, thus obtaining more determined predictions as in [24, 37] while the positive value (i.e. 0.1) is more like a smooth regularization as KLD loss in [100]. The results are shown in Figure 4. It can be seen that both positive and negative values can bring some improvements. However, our CR-CACo loss can achieve better performance, which validates the importance of source prior class relationship." + } + ] + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 54, + 355, + 164, + 434 + ], + "blocks": [ + { + "bbox": [ + 54, + 355, + 164, + 434 + ], + "lines": [ + { + "bbox": [ + 54, + 355, + 164, + 434 + ], + "spans": [ + { + "bbox": [ + 54, + 355, + 164, + 434 + ], + "type": "image", + "image_path": "40eee282579e5154d2c9b30b11ecb5601bfe9e42b8e5d5cc9267255ae70d646e.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 439, + 287, + 483 + ], + "lines": [ + { + "bbox": [ + 46, + 439, + 287, + 483 + ], + "spans": [ + { + "bbox": [ + 46, + 439, + 287, + 483 + ], + "type": "text", + "content": "Figure 4. Using fixed coefficients in " + }, + { + "bbox": [ + 46, + 439, + 287, + 483 + ], + "type": "inline_equation", + "content": "s_{inter}" + }, + { + "bbox": [ + 46, + 439, + 287, + 483 + ], + "type": "text", + "content": " in both CACo and IDCo losses under SUDA setting Office-Home Rw→Cl. Black dotted line shows the performance of baseline IM. Red dotted line shows the performance of our losses." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 167, + 355, + 278, + 434 + ], + "blocks": [ + { + "bbox": [ + 167, + 355, + 278, + 434 + ], + "lines": [ + { + "bbox": [ + 167, + 355, + 278, + 434 + ], + "spans": [ + { + "bbox": [ + 167, + 355, + 278, + 434 + ], + "type": "image", + "image_path": "97f9aace554f4bf3963be3b5a63656b1c009b8a3cacc3cc5828033531f0184ed.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 491, + 287, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 491, + 287, + 552 + ], + "spans": [ + { + "bbox": [ + 46, + 491, + 287, + 552 + ], + "type": "text", + "content": "Parameter sensitivity Here we analyze the hyperparameter sensitivity under the setting of SUDA Office-Home with SHOT-IM baseline. We consider four hyperparameters, as shown in Figure 5. It is evident that they are stable within specific ranges." + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 50, + 562, + 109, + 609 + ], + "blocks": [ + { + "bbox": [ + 50, + 562, + 109, + 609 + ], + "lines": [ + { + "bbox": [ + 50, + 562, + 109, + 609 + ], + "spans": [ + { + "bbox": [ + 50, + 562, + 109, + 609 + ], + "type": "image", + "image_path": "fd46b8a5fb321a60c20f354ffc3153494d624b1ad57f0ea342e5abb3b0b397f5.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 615, + 286, + 636 + ], + "lines": [ + { + "bbox": [ + 46, + 615, + 286, + 636 + ], + "spans": [ + { + "bbox": [ + 46, + 615, + 286, + 636 + ], + "type": "text", + "content": "Figure 5. Sensitivity analysis of different hyperparameters. The first two use our two losses. The last two use corresponding loss." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 111, + 563, + 166, + 609 + ], + "blocks": [ + { + "bbox": [ + 111, + 563, + 166, + 609 + ], + "lines": [ + { + "bbox": [ + 111, + 563, + 166, + 609 + ], + "spans": [ + { + "bbox": [ + 111, + 563, + 166, + 609 + ], + "type": "image", + "image_path": "e2ec62eae2eee709ec6a4f0d714a9e827c2bab0869a754c906934091a6e6a6c3.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 167, + 563, + 228, + 609 + ], + "blocks": [ + { + "bbox": [ + 167, + 563, + 228, + 609 + ], + "lines": [ + { + "bbox": [ + 167, + 563, + 228, + 609 + ], + "spans": [ + { + "bbox": [ + 167, + 563, + 228, + 609 + ], + "type": "image", + "image_path": "3b5ebe708294f823b942c00d64035aaaeafae24c2449cd2d7241bea804eb5702.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 228, + 563, + 282, + 609 + ], + "blocks": [ + { + "bbox": [ + 228, + 563, + 282, + 609 + ], + "lines": [ + { + "bbox": [ + 228, + 563, + 282, + 609 + ], + "spans": [ + { + "bbox": [ + 228, + 563, + 282, + 609 + ], + "type": "image", + "image_path": "5adb7af2bb021a2f1e0209afda563b0f312537b89fe91b92ac2309e8dbadb225.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 653, + 286, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 286, + 712 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 286, + 712 + ], + "type": "text", + "content": "Visualization of features We visualize the target features and classifier weights in Figure 6. It can be seen that our method has less misclassified samples for classes 0, 1, 7 compared with IM. Thus our method can learn more discriminative features." + } + ] + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 312, + 70, + 388, + 144 + ], + "blocks": [ + { + "bbox": [ + 312, + 70, + 388, + 144 + ], + "lines": [ + { + "bbox": [ + 312, + 70, + 388, + 144 + ], + "spans": [ + { + "bbox": [ + 312, + 70, + 388, + 144 + ], + "type": "image", + "image_path": "a96140f46635e00fcd21fda7f2e65eac519197eb71608c757f8a972d12758aff.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 315, + 146, + 382, + 153 + ], + "lines": [ + { + "bbox": [ + 315, + 146, + 382, + 153 + ], + "spans": [ + { + "bbox": [ + 315, + 146, + 382, + 153 + ], + "type": "text", + "content": "Source Pre-trained Model" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 389, + 70, + 465, + 144 + ], + "blocks": [ + { + "bbox": [ + 389, + 70, + 465, + 144 + ], + "lines": [ + { + "bbox": [ + 389, + 70, + 465, + 144 + ], + "spans": [ + { + "bbox": [ + 389, + 70, + 465, + 144 + ], + "type": "image", + "image_path": "c8de4fefb7c9b6cb96f3beda508aebde9f7d424866d32e8f17d93ad2fd091e9b.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 423, + 146, + 433, + 153 + ], + "lines": [ + { + "bbox": [ + 423, + 146, + 433, + 153 + ], + "spans": [ + { + "bbox": [ + 423, + 146, + 433, + 153 + ], + "type": "text", + "content": "IM" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 304, + 163, + 545, + 228 + ], + "lines": [ + { + "bbox": [ + 304, + 163, + 545, + 228 + ], + "spans": [ + { + "bbox": [ + 304, + 163, + 545, + 228 + ], + "type": "text", + "content": "Figure 6. The t-SNE visualization of classifier weights (i.e. star) and target domain features (i.e. dot) of different methods under the setting of SUDA Office-Home " + }, + { + "bbox": [ + 304, + 163, + 545, + 228 + ], + "type": "inline_equation", + "content": "\\mathrm{Rw} \\rightarrow \\mathrm{Cl}" + }, + { + "bbox": [ + 304, + 163, + 545, + 228 + ], + "type": "text", + "content": ". We directly use the first 10 classes (i.e. different colors), and the coordinates of classifier weights are the same across different methods since source classifier is frozen. Best viewed in color." + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 467, + 71, + 542, + 144 + ], + "blocks": [ + { + "bbox": [ + 467, + 71, + 542, + 144 + ], + "lines": [ + { + "bbox": [ + 467, + 71, + 542, + 144 + ], + "spans": [ + { + "bbox": [ + 467, + 71, + 542, + 144 + ], + "type": "image", + "image_path": "7dc7c4a2f18f0de05efdcba2655e03b7e7cea50aada2308289fbccfad77834e3.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 496, + 146, + 522, + 153 + ], + "lines": [ + { + "bbox": [ + 496, + 146, + 522, + 153 + ], + "spans": [ + { + "bbox": [ + 496, + 146, + 522, + 153 + ], + "type": "text", + "content": "IM+Ours" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 247, + 545, + 344 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 247, + 545, + 344 + ], + "spans": [ + { + "bbox": [ + 304, + 247, + 545, + 344 + ], + "type": "text", + "content": "Transformer backbone We further apply our method with Transformer-based backbone under the SFUDA setting, as shown in Table 8. We choose the same network (i.e. ViT-B [15]) following SSRT [69]. For the training of the source model, we follow SHOT [42] which obtains better performance than original ViT-B (78.2% vs. 75.5%). Adding IM loss achieves 84.0%. Based on IM loss, ours achieves 88.3% which significantly outperforms SSRT." + } + ] + } + ], + "index": 20 + }, + { + "type": "table", + "bbox": [ + 308, + 384, + 541, + 441 + ], + "blocks": [ + { + "bbox": [ + 306, + 353, + 545, + 374 + ], + "lines": [ + { + "bbox": [ + 306, + 353, + 545, + 374 + ], + "spans": [ + { + "bbox": [ + 306, + 353, + 545, + 374 + ], + "type": "text", + "content": "Table 8. Single-Source Unsupervised DA (SUDA) on Office-Home with ViT-B backbone." + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 308, + 384, + 541, + 441 + ], + "lines": [ + { + "bbox": [ + 308, + 384, + 541, + 441 + ], + "spans": [ + { + "bbox": [ + 308, + 384, + 541, + 441 + ], + "type": "table", + "html": "
MethodSFAvg.MethodSFAvg.
ViT-B✗75.5CDTrans (ICLR'22) [83]✗80.5
TVT (WACV'23) [86]✗83.6DOT-B (ACMMM'22) [50]✗84.3
SSRT (CVPR'22) [69]✗85.4BCAT-DTF (Arxiv'22) [76]✗86.6
Source model✓78.2SHOT-IM✓84.0
SHOT-IM+CR-CACo✓87.0(+3.0)SHOT-IM+Ours✓88.3(+4.3)
", + "image_path": "7ddc7ae76cbd9514fba365541210a7737d45d790ab4409a4aa532da9f2ecbe96.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "table_body" + } + ], + "index": 22 + }, + { + "bbox": [ + 306, + 448, + 378, + 460 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 448, + 378, + 460 + ], + "spans": [ + { + "bbox": [ + 306, + 448, + 378, + 460 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 304, + 468, + 545, + 612 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 468, + 545, + 612 + ], + "spans": [ + { + "bbox": [ + 304, + 468, + 545, + 612 + ], + "type": "text", + "content": "In this paper, we propose to explicitly transfer the class relationship for SFUDA which is more domain-invariant. We propose a novel class relationship embedded similarity that can more accurately express the sample relationship in the output space. Furthermore, we propose two contrastive losses (i.e., CR-CACo and CR-IDCo) that exploit our designed similarity. These two losses are complementary, and their combination can better explore the target distribution. We combine our method with existing representative baselines in multiple SFUDA settings. Extensive experiments show the effectiveness of the proposed method, which achieves state-of-the-art performance." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 306, + 621, + 420, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 621, + 420, + 635 + ], + "spans": [ + { + "bbox": [ + 306, + 621, + 420, + 635 + ], + "type": "text", + "content": "6. Acknowledgements" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 304, + 641, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 641, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 641, + 545, + 713 + ], + "type": "text", + "content": "This work is supported by the National Natural Science Foundation of China under Grant 62176246 and Grant 61836008. This work is also supported by Anhui Provincial Natural Science Foundation 2208085UD17 and the Fundamental Research Funds for the Central Universities (WK3490000006)." + } + ] + } + ], + "index": 26 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7626" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 53, + 91, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 56, + 91, + 287, + 134 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 91, + 287, + 134 + ], + "spans": [ + { + "bbox": [ + 56, + 91, + 287, + 134 + ], + "type": "text", + "content": "[1] Sk. Miraj Ahmed, Dripta S. Raychaudhuri, S. Paul, Samet Oymak, and Amit K. Roy-Chowdhury. Unsupervised multi-source domain adaptation without access to source data. In CVPR, 2021. 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 137, + 288, + 179 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 137, + 288, + 179 + ], + "spans": [ + { + "bbox": [ + 56, + 137, + 288, + 179 + ], + "type": "text", + "content": "[2] Shai Ben-David, John Blitzer, Koby Crammer, Alex Kulesza, Fernando Pereira, and Jennifer Wortman Vaughan. A theory of learning from different domains. Machine learning, 79(1-2):151-175, 2010. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 182, + 287, + 214 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 182, + 287, + 214 + ], + "spans": [ + { + "bbox": [ + 56, + 182, + 287, + 214 + ], + "type": "text", + "content": "[3] Shai Ben-David, John Blitzer, Koby Crammer, and Fernando Pereira. Analysis of representations for domain adaptation. In NeurIPS, 2006. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 57, + 216, + 287, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 216, + 287, + 237 + ], + "spans": [ + { + "bbox": [ + 57, + 216, + 287, + 237 + ], + "type": "text", + "content": "[4] Abhijit Bendale and Terrance E Boult. Towards open set deep networks. In CVPR, 2016. 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 57, + 239, + 287, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 239, + 287, + 271 + ], + "spans": [ + { + "bbox": [ + 57, + 239, + 287, + 271 + ], + "type": "text", + "content": "[5] Zhangjie Cao, Mingsheng Long, Jianmin Wang, and Michael I Jordan. Partial transfer learning with selective adversarial networks. In CVPR, 2018. 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 57, + 273, + 287, + 305 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 273, + 287, + 305 + ], + "spans": [ + { + "bbox": [ + 57, + 273, + 287, + 305 + ], + "type": "text", + "content": "[6] Zhangjie Cao, Kaichao You, Mingsheng Long, Jianmin Wang, and Qiang Yang. Learning to transfer examples for partial domain adaptation. In CVPR, 2019. 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 57, + 307, + 287, + 350 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 307, + 287, + 350 + ], + "spans": [ + { + "bbox": [ + 57, + 307, + 287, + 350 + ], + "type": "text", + "content": "[7] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In ECCV, 2020. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 57, + 352, + 287, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 352, + 287, + 396 + ], + "spans": [ + { + "bbox": [ + 57, + 352, + 287, + 396 + ], + "type": "text", + "content": "[8] Chaoqi Chen, Weiping Xie, Tingyang Xu, Wenbing Huang, Yu Rong, Xinghao Ding, Yue Huang, and Junzhou Huang. Progressive feature alignment for unsupervised domain adaptation. In CVPR, 2019. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 57, + 397, + 287, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 397, + 287, + 430 + ], + "spans": [ + { + "bbox": [ + 57, + 397, + 287, + 430 + ], + "type": "text", + "content": "[9] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In ICML, 2020. 2, 4, 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 53, + 432, + 287, + 474 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 432, + 287, + 474 + ], + "spans": [ + { + "bbox": [ + 53, + 432, + 287, + 474 + ], + "type": "text", + "content": "[10] Tong Chu, Yahao Liu, Jinhong Deng, Wen Li, and Lixin Duan. Denoised maximum classifier discrepancy for sourcefree unsupervised domain adaptation. In AAAI, 2022. 3, 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 53, + 477, + 287, + 508 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 477, + 287, + 508 + ], + "spans": [ + { + "bbox": [ + 53, + 477, + 287, + 508 + ], + "type": "text", + "content": "[11] Safa Cicek and Stefano Soatto. Unsupervised domain adaptation via regularized conditional alignment. In ICCV, 2019. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 53, + 510, + 287, + 543 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 510, + 287, + 543 + ], + "spans": [ + { + "bbox": [ + 53, + 510, + 287, + 543 + ], + "type": "text", + "content": "[12] Shuhao Cui, Shuhui Wang, Junbao Zhuo, Chi Su, Qingming Huang, and Qi Tian. Gradually vanishing bridge for adversarial domain adaptation. In CVPR, 2020. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 53, + 544, + 287, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 544, + 287, + 577 + ], + "spans": [ + { + "bbox": [ + 53, + 544, + 287, + 577 + ], + "type": "text", + "content": "[13] Zhijie Deng, Yucen Luo, and Jun Zhu. Cluster alignment with a teacher for unsupervised domain adaptation. In ICCV, 2019. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 53, + 579, + 287, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 579, + 287, + 611 + ], + "spans": [ + { + "bbox": [ + 53, + 579, + 287, + 611 + ], + "type": "text", + "content": "[14] Jiahua Dong, Zhen Fang, Anjin Liu, Gan Sun, and Tongliang Liu. Confident anchor-induced multi-source free domain adaptation. In NeurIPS, 2021. 2, 6, 7" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 53, + 613, + 287, + 677 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 613, + 287, + 677 + ], + "spans": [ + { + "bbox": [ + 53, + 613, + 287, + 677 + ], + "type": "text", + "content": "[15] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2020. 1, 8" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 53, + 680, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 680, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 53, + 680, + 287, + 712 + ], + "type": "text", + "content": "[16] Geoffrey French, Michal Mackiewicz, and Mark Fisher. Self-ensembling for visual domain adaptation. In ICLR, 2018. 3" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 312, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 35, + "blocks": [ + { + "bbox": [ + 312, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 312, + 73, + 545, + 106 + ], + "type": "text", + "content": "[17] Yuan Gao, Zilei Wang, Jiafan Zhuang, Yixin Zhang, and Junjie Li. Exploit domain-robust optical flow in domain adaptive video semantic segmentation. In AAAI, 2023. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 312, + 107, + 545, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 107, + 545, + 138 + ], + "spans": [ + { + "bbox": [ + 312, + 107, + 545, + 138 + ], + "type": "text", + "content": "[18] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In CVPR, 2020. 2, 3, 4, 5, 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 312, + 140, + 545, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 140, + 545, + 171 + ], + "spans": [ + { + "bbox": [ + 312, + 140, + 545, + 171 + ], + "type": "text", + "content": "[19] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, 2016. 1, 5, 7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 312, + 173, + 545, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 173, + 545, + 216 + ], + "spans": [ + { + "bbox": [ + 312, + 173, + 545, + 216 + ], + "type": "text", + "content": "[20] Weihua Hu, Takeru Miyato, Seiya Tokui, Eiichi Matsumoto, and Masashi Sugiyama. Learning discrete representations via information maximizing self-augmented training. In ICML, 2017. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 312, + 217, + 545, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 217, + 545, + 259 + ], + "spans": [ + { + "bbox": [ + 312, + 217, + 545, + 259 + ], + "type": "text", + "content": "[21] Jiaxing Huang, Dayan Guan, Aoran Xiao, and Shijian Lu. Model adaptation: Historical contrastive learning for unsupervised domain adaptation without source data. In NeurIPS, 2021. 1, 2, 3, 4, 6, 7" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 312, + 261, + 545, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 261, + 545, + 293 + ], + "spans": [ + { + "bbox": [ + 312, + 261, + 545, + 293 + ], + "type": "text", + "content": "[22] Jiaxing Huang, Dayan Guan, Aoran Xiao, and Shijian Lu. RDA: Robust domain adaptation via fourier adversarial attacking. In ICCV, 2021. 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 312, + 294, + 545, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 294, + 545, + 326 + ], + "spans": [ + { + "bbox": [ + 312, + 294, + 545, + 326 + ], + "type": "text", + "content": "[23] Xin Jin, Cuiling Lan, Wenjun Zeng, and Zhibo Chen. Re-energizing domain discriminator with sample relabeling for adversarial domain adaptation. In ICCV, 2021. 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 312, + 327, + 545, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 327, + 545, + 358 + ], + "spans": [ + { + "bbox": [ + 312, + 327, + 545, + 358 + ], + "type": "text", + "content": "[24] Ying Jin, Ximei Wang, Mingsheng Long, and Jianmin Wang. Minimum class confusion for versatile domain adaptation. In ECCV, 2020. 6, 7, 8" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 312, + 361, + 545, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 361, + 545, + 392 + ], + "spans": [ + { + "bbox": [ + 312, + 361, + 545, + 392 + ], + "type": "text", + "content": "[25] Mengmeng Jing, Xiantong Zhen, Jingjing Li, and Cees GM Snoek. Variational model perturbation for source-free domain adaptation. In NeurIPS, 2022. 3, 6" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 312, + 394, + 545, + 425 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 394, + 545, + 425 + ], + "spans": [ + { + "bbox": [ + 312, + 394, + 545, + 425 + ], + "type": "text", + "content": "[26] Guoliang Kang, Lu Jiang, Yi Yang, and Alexander G Hauptmann. Contrastive adaptation network for unsupervised domain adaptation. In CVPR, 2019. 3, 4" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 312, + 426, + 545, + 469 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 426, + 545, + 469 + ], + "spans": [ + { + "bbox": [ + 312, + 426, + 545, + 469 + ], + "type": "text", + "content": "[27] Seunghyeon Kim, Jaehoon Choi, Taekyung Kim, and Changick Kim. Self-training and adversarial background regularization for unsupervised domain adaptive one-stage object detection. In CVPR, 2019. 3" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 312, + 472, + 545, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 472, + 545, + 502 + ], + "spans": [ + { + "bbox": [ + 312, + 472, + 545, + 502 + ], + "type": "text", + "content": "[28] Ananya Kumar, Tengyu Ma, and Percy Liang. Understanding self-training for gradual domain adaptation. In ICML, 2020. 3" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 312, + 504, + 545, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 504, + 545, + 547 + ], + "spans": [ + { + "bbox": [ + 312, + 504, + 545, + 547 + ], + "type": "text", + "content": "[29] Jogendra Nath Kundu, Suvaansh Bhambri, Akshay Kulkarni, Hiran Sarkar, Varun Jampani, and R Venkatesh Babu. Concurrent subsidiary supervision for unsupervised source-free domain adaptation. In ECCV, 2022. 3, 6, 7" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 312, + 548, + 545, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 548, + 545, + 601 + ], + "spans": [ + { + "bbox": [ + 312, + 548, + 545, + 601 + ], + "type": "text", + "content": "[30] Jogendra Nath Kundu, Akshay R Kulkarni, Suvaansh Bhambri, Deepesh Mehta, Shreyas Anand Kulkarni, Varun Jampani, and Venkatesh Babu Radhakrishnan. Balancing discriminability and transferability for source-free domain adaptation. In ICML, 2022. 3, 6, 7" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 312, + 604, + 545, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 604, + 545, + 634 + ], + "spans": [ + { + "bbox": [ + 312, + 604, + 545, + 634 + ], + "type": "text", + "content": "[31] Jogendra Nath Kundu, Naveen Venkat, and R Venkatesh Babu. Universal source-free domain adaptation. In CVPR, 2020. 3" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 312, + 636, + 545, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 636, + 545, + 678 + ], + "spans": [ + { + "bbox": [ + 312, + 636, + 545, + 678 + ], + "type": "text", + "content": "[32] Jogendra Nath Kundu, Naveen Venkat, Ambareesh Revanur, Rahul M V, and R. Venkatesh Babu. Towards inheritable models for open-set domain adaptation. In CVPR, 2020. 3" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 312, + 680, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 680, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 312, + 680, + 545, + 712 + ], + "type": "text", + "content": "[33] Jonghyun Lee, Dahuin Jung, Junho Yim, and Sungroh Yoon. Confidence score for source-free unsupervised domain adaptation. In ICML, 2022. 3, 6, 7" + } + ] + } + ], + "index": 34 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7627" + } + ] + } + ], + "index": 36 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 53, + 72, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 53, + 72, + 287, + 115 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 72, + 287, + 115 + ], + "spans": [ + { + "bbox": [ + 53, + 72, + 287, + 115 + ], + "type": "text", + "content": "[34] Junjie Li, Zilei Wang, Yuan Gao, and Xiaoming Hu. Exploring high-quality target domain information for unsupervised domain adaptive semantic segmentation. In ACMMM, 2022. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 53, + 117, + 287, + 159 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 117, + 287, + 159 + ], + "spans": [ + { + "bbox": [ + 53, + 117, + 287, + 159 + ], + "type": "text", + "content": "[35] Junjie Li, Yixin Zhang, Zilei Wang, and Keyu Tu. Semantic-aware representation learning via probability contrastive loss. arXiv preprint arXiv:2111.06021, 2021. 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 161, + 288, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 161, + 288, + 194 + ], + "spans": [ + { + "bbox": [ + 53, + 161, + 288, + 194 + ], + "type": "text", + "content": "[36] Rui Li, Qianfen Jiao, Wenming Cao, Hau-San Wong, and Si Wu. Model adaptation: Unsupervised domain adaptation without source data. In CVPR, 2020. 1, 2, 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 195, + 287, + 236 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 195, + 287, + 236 + ], + "spans": [ + { + "bbox": [ + 53, + 195, + 287, + 236 + ], + "type": "text", + "content": "[37] Shuang Li, Fangrui Lv, Binhui Xie, Chi Harold Liu, Jian Liang, and Chen Qin. Bi-classifier determinacy maximization for unsupervised domain adaptation. In AAAI, 2021. 6, 8" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 238, + 287, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 238, + 287, + 270 + ], + "spans": [ + { + "bbox": [ + 53, + 238, + 287, + 270 + ], + "type": "text", + "content": "[38] Shuang Li, Mixue Xie, Fangrui Lv, Chi Harold Liu, Jian Liang, Chen Qin, and Wei Li. Semantic concentration for domain adaptation. In ICCV, 2021. 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 271, + 287, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 271, + 287, + 304 + ], + "spans": [ + { + "bbox": [ + 53, + 271, + 287, + 304 + ], + "type": "text", + "content": "[39] Yunsheng Li, Lu Yuan, Yinpeng Chen, Pei Wang, and N. Vasconcelos. Dynamic transfer for multi-source domain adaptation. In CVPR, 2021. 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 304, + 287, + 337 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 304, + 287, + 337 + ], + "spans": [ + { + "bbox": [ + 53, + 304, + 287, + 337 + ], + "type": "text", + "content": "[40] Yunsheng Li, Lu Yuan, and Nuno Vasconcelos. Bidirectional learning for domain adaptation of semantic segmentation. In CVPR, 2019. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 338, + 287, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 338, + 287, + 380 + ], + "spans": [ + { + "bbox": [ + 53, + 338, + 287, + 380 + ], + "type": "text", + "content": "[41] Yu-Jhe Li, Xiaoliang Dai, Chih-Yao Ma, Yen-Cheng Liu, Kan Chen, Bichen Wu, Zijian He, Kris Kitani, and Peter Vajda. Cross-domain adaptive teacher for object detection. In CVPR, 2022. 4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 382, + 287, + 425 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 382, + 287, + 425 + ], + "spans": [ + { + "bbox": [ + 53, + 382, + 287, + 425 + ], + "type": "text", + "content": "[42] Jian Liang, Dapeng Hu, and Jiashi Feng. Do we really need to access the source data? source hypothesis transfer for unsupervised domain adaptation. In ICML, 2020. 1, 2, 3, 4, 5, 6, 7, 8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 426, + 287, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 426, + 287, + 479 + ], + "spans": [ + { + "bbox": [ + 53, + 426, + 287, + 479 + ], + "type": "text", + "content": "[43] Jian Liang, Dapeng Hu, Yunbo Wang, Ran He, and Jiashi Feng. Source data-absent unsupervised domain adaptation through hypothesis transfer and labeling transfer. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2021. 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 53, + 481, + 287, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 481, + 287, + 514 + ], + "spans": [ + { + "bbox": [ + 53, + 481, + 287, + 514 + ], + "type": "text", + "content": "[44] Hong Liu, Zhangjie Cao, Mingsheng Long, Jianmin Wang, and Qiang Yang. Separate to adapt: Open set domain adaptation via progressive separation. In CVPR, 2019. 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 53, + 514, + 287, + 548 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 514, + 287, + 548 + ], + "spans": [ + { + "bbox": [ + 53, + 514, + 287, + 548 + ], + "type": "text", + "content": "[45] Hong Liu, Mingsheng Long, Jianmin Wang, and Michael Jordan. Transferable adversarial training: A general approach to adapting deep classifiers. In ICML, 2019. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 53, + 548, + 287, + 569 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 548, + 287, + 569 + ], + "spans": [ + { + "bbox": [ + 53, + 548, + 287, + 569 + ], + "type": "text", + "content": "[46] Hong Liu, Jianmin Wang, and Mingsheng Long. Cycle self-training for domain adaptation. In NeurIPS, 2021. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 53, + 570, + 287, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 570, + 287, + 602 + ], + "spans": [ + { + "bbox": [ + 53, + 570, + 287, + 602 + ], + "type": "text", + "content": "[47] Qinying Liu and Zilei Wang. Collaborating domain-shared and target-specific feature clustering for cross-domain 3d action recognition. In ECCV, 2022. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 53, + 603, + 287, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 603, + 287, + 635 + ], + "spans": [ + { + "bbox": [ + 53, + 603, + 287, + 635 + ], + "type": "text", + "content": "[48] Mingsheng Long, Zhangjie Cao, Jianmin Wang, and Michael I Jordan. Conditional adversarial domain adaptation. In NeurIPS, 2018. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 53, + 636, + 287, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 636, + 287, + 668 + ], + "spans": [ + { + "bbox": [ + 53, + 636, + 287, + 668 + ], + "type": "text", + "content": "[49] Mingsheng Long, Han Zhu, Jianmin Wang, and Michael I Jordan. Deep transfer learning with joint adaptation networks. In ICML, 2017. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 53, + 669, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 669, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 53, + 669, + 287, + 712 + ], + "type": "text", + "content": "[50] Wenxuan Ma, Jinming Zhang, Shuang Li, Chi Harold Liu, Yulin Wang, and Wei Li. Making the best of both worlds: A domain-oriented transformer for unsupervised domain adaptation. In ACMMM, 2022. 8" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 312, + 72, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 35, + "blocks": [ + { + "bbox": [ + 312, + 72, + 545, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 72, + 545, + 105 + ], + "spans": [ + { + "bbox": [ + 312, + 72, + 545, + 105 + ], + "type": "text", + "content": "[51] Ke Mei, Chuang Zhu, Jiaqi Zou, and Shanghang Zhang. Instance adaptive self-training for unsupervised domain adaptation. In ECCV, 2020. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 312, + 106, + 545, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 106, + 545, + 139 + ], + "spans": [ + { + "bbox": [ + 312, + 106, + 545, + 139 + ], + "type": "text", + "content": "[52] Jaemin Na, Heechul Jung, Hyung Jin Chang, and Wonjun Hwang. FixBi: Bridging domain spaces for unsupervised domain adaptation. In CVPR, 2021. 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 312, + 140, + 545, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 140, + 545, + 172 + ], + "spans": [ + { + "bbox": [ + 312, + 140, + 545, + 172 + ], + "type": "text", + "content": "[53] Van-Anh Nguyen, Tuan Nguyen, Trung Le, Quan Hung Tran, and Dinh Phung. STEM: An approach to multi-source domain adaptation with guarantees. In ICCV, 2021. 7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 312, + 173, + 545, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 173, + 545, + 215 + ], + "spans": [ + { + "bbox": [ + 312, + 173, + 545, + 215 + ], + "type": "text", + "content": "[54] Fei Pan, Inkyu Shin, Francois Rameau, Seokju Lee, and In So Kweon. Unsupervised intra-domain adaptation for semantic segmentation through self-supervision. In CVPR, 2020. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 312, + 216, + 545, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 216, + 545, + 249 + ], + "spans": [ + { + "bbox": [ + 312, + 216, + 545, + 249 + ], + "type": "text", + "content": "[55] Yingwei Pan, Ting Yao, Yehao Li, Yu Wang, Chong-Wah Ngo, and Tao Mei. Transferrable prototypical networks for unsupervised domain adaptation. In CVPR, 2019. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 312, + 250, + 545, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 250, + 545, + 271 + ], + "spans": [ + { + "bbox": [ + 312, + 250, + 545, + 271 + ], + "type": "text", + "content": "[56] Pau Panareda Busto and Juergen Gall. Open set domain adaptation. In ICCV, 2017. 7" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 312, + 271, + 545, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 271, + 545, + 304 + ], + "spans": [ + { + "bbox": [ + 312, + 271, + 545, + 304 + ], + "type": "text", + "content": "[57] Xingchao Peng, Qinxun Bai, Xide Xia, Zijun Huang, Kate Saenko, and Bo Wang. Moment matching for multi-source domain adaptation. In ICCV, 2019. 2, 5" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 312, + 304, + 545, + 338 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 304, + 545, + 338 + ], + "spans": [ + { + "bbox": [ + 312, + 304, + 545, + 338 + ], + "type": "text", + "content": "[58] Xingchao Peng, Ben Usman, Neela Kaushik, Judy Hoffman, Dequan Wang, and Kate Saenko. VisDA: The visual domain adaptation challenge. In CVPRW, 2018. 5" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 312, + 338, + 545, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 338, + 545, + 381 + ], + "spans": [ + { + "bbox": [ + 312, + 338, + 545, + 381 + ], + "type": "text", + "content": "[59] Viraj Prabhu, Shivam Khare, Deeksha Kartik, and Judy Hoffman. SENTRY: Selective entropy optimization via committee consistency for unsupervised domain adaptation. In ICCV, 2021. 6" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 312, + 382, + 545, + 425 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 382, + 545, + 425 + ], + "spans": [ + { + "bbox": [ + 312, + 382, + 545, + 425 + ], + "type": "text", + "content": "[60] Zhen Qiu, Yifan Zhang, Hongbin Lin, Shuaicheng Niu, Yanxia Liu, Qing Du, and Mingkui Tan. Source-free domain adaptation via avatar prototype generation and adaptation. In IJCAI, 2021. 6" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 312, + 426, + 545, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 426, + 545, + 470 + ], + "spans": [ + { + "bbox": [ + 312, + 426, + 545, + 470 + ], + "type": "text", + "content": "[61] Sanqing Qu, Guang Chen, Jing Zhang, Zhijun Li, Wei He, and Dacheng Tao. Bmd: A general class-balanced multicentric dynamic prototype strategy for source-free domain adaptation. In ECCV, 2022. 3, 6" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 312, + 471, + 545, + 513 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 471, + 545, + 513 + ], + "spans": [ + { + "bbox": [ + 312, + 471, + 545, + 513 + ], + "type": "text", + "content": "[62] Subhankar Roy, Martin Trapp, Andrea Pilzer, Juho Kannala, Nicu Sebe, Elisa Ricci, and Arno Solin. Uncertainty-guided source-free domain adaptation. In ECCV, 2022. 3, 6" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 312, + 514, + 545, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 514, + 545, + 547 + ], + "spans": [ + { + "bbox": [ + 312, + 514, + 545, + 547 + ], + "type": "text", + "content": "[63] Kate Saenko, Brian Kulis, Mario Fritz, and Trevor Darrell. Adapting visual category models to new domains. In ECCV, 2010. 5" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 312, + 548, + 545, + 580 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 548, + 545, + 580 + ], + "spans": [ + { + "bbox": [ + 312, + 548, + 545, + 580 + ], + "type": "text", + "content": "[64] Kuniaki Saito, Kohei Watanabe, Yoshitaka Ushiku, and Tatsuya Harada. Maximum classifier discrepancy for unsupervised domain adaptation. In CVPR, 2018. 2, 3" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 312, + 581, + 545, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 581, + 545, + 613 + ], + "spans": [ + { + "bbox": [ + 312, + 581, + 545, + 613 + ], + "type": "text", + "content": "[65] Kuniaki Saito, Shohei Yamamoto, Yoshitaka Ushiku, and Tatsuya Harada. Open set domain adaptation by backpropagation. In ECCV, 2018. 7" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 312, + 614, + 545, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 614, + 545, + 657 + ], + "spans": [ + { + "bbox": [ + 312, + 614, + 545, + 657 + ], + "type": "text", + "content": "[66] Marin Scalbert, Maria Vakalopoulou, and Florent Couzini'e-Devy. Multi-source domain adaptation via supervised contrastive learning and confident consistency regularization. In BMVC, 2021. 7" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 312, + 658, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 658, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 312, + 658, + 545, + 712 + ], + "type": "text", + "content": "[67] Kihyuk Sohn, David Berthelot, Chun-Liang Li, Zizhao Zhang, Nicholas Carlini, Ekin D Cubuk, Alex Kurakin, Han Zhang, and Colin Raffel. Fixmatch: Simplifying semi-supervised learning with consistency and confidence. In NeurIPS, 2020. 2, 4, 5" + } + ] + } + ], + "index": 34 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7628" + } + ] + } + ], + "index": 36 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 53, + 73, + 286, + 713 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 53, + 73, + 286, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 73, + 286, + 95 + ], + "spans": [ + { + "bbox": [ + 53, + 73, + 286, + 95 + ], + "type": "text", + "content": "[68] Baochen Sun and Kate Saenko. Deep coral: Correlation alignment for deep domain adaptation. In ECCV, 2016. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 53, + 96, + 286, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 96, + 286, + 128 + ], + "spans": [ + { + "bbox": [ + 53, + 96, + 286, + 128 + ], + "type": "text", + "content": "[69] Tao Sun, Cheng Lu, Tianshuo Zhang, and Haibin Ling. Safe self-refinement for transformer-based domain adaptation. In CVPR, 2022. 8" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 130, + 286, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 130, + 286, + 162 + ], + "spans": [ + { + "bbox": [ + 53, + 130, + 286, + 162 + ], + "type": "text", + "content": "[70] Hui Tang, Ke Chen, and Kui Jia. Unsupervised domain adaptation via structurally regularized deep clustering. In CVPR, 2020. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 164, + 286, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 164, + 286, + 207 + ], + "spans": [ + { + "bbox": [ + 53, + 164, + 286, + 207 + ], + "type": "text", + "content": "[71] Jiayi Tian, Jing Zhang, Wen Li, and Dong Xu. VDMDA: Virtual domain modeling for source data-free domain adaptation. IEEE Transactions on Circuits and Systems for Video Technology, 2021. 1, 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 209, + 286, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 209, + 286, + 251 + ], + "spans": [ + { + "bbox": [ + 53, + 209, + 286, + 251 + ], + "type": "text", + "content": "[72] Eric Tzeng, Judy Hoffman, Ning Zhang, Kate Saenko, and Trevor Darrell. Deep domain confusion: Maximizing for domain invariance. arXiv preprint arXiv:1412.3474, 2014. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 253, + 286, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 253, + 286, + 296 + ], + "spans": [ + { + "bbox": [ + 53, + 253, + 286, + 296 + ], + "type": "text", + "content": "[73] Hemanth Venkateswara, Jose Eusebio, Shayok Chakraborty, and Sethuraman Panchanathan. Deep hashing network for unsupervised domain adaptation. In CVPR, 2017. 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 298, + 286, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 298, + 286, + 330 + ], + "spans": [ + { + "bbox": [ + 53, + 298, + 286, + 330 + ], + "type": "text", + "content": "[74] Fan Wang, Zhongyi Han, Yongshun Gong, and Yilong Yin. Exploring domain-invariant parameters for source free domain adaptation. In CVPR, 2022. 3, 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 332, + 286, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 332, + 286, + 353 + ], + "spans": [ + { + "bbox": [ + 53, + 332, + 286, + 353 + ], + "type": "text", + "content": "[75] Feng Wang and Huaping Liu. Understanding the behaviour of contrastive loss. In CVPR, 2021. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 355, + 286, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 355, + 286, + 387 + ], + "spans": [ + { + "bbox": [ + 53, + 355, + 286, + 387 + ], + "type": "text", + "content": "[76] Xiyu Wang, Pengxin Guo, and Yu Zhang. Domain adaptation via bidirectional cross-attention transformer. arXiv preprint arXiv:2201.05887, 2022. 8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 388, + 286, + 420 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 388, + 286, + 420 + ], + "spans": [ + { + "bbox": [ + 53, + 388, + 286, + 420 + ], + "type": "text", + "content": "[77] Ximei Wang, Liang Li, Weirui Ye, Mingsheng Long, and Jianmin Wang. Transferable attention for domain adaptation. In AAAI, 2019. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 53, + 422, + 286, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 422, + 286, + 453 + ], + "spans": [ + { + "bbox": [ + 53, + 422, + 286, + 453 + ], + "type": "text", + "content": "[78] Zhirong Wu, Yuanjun Xiong, Stella X Yu, and Dahua Lin. Unsupervised feature learning via non-parametric instance discrimination. In CVPR, 2018. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 53, + 456, + 286, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 456, + 286, + 487 + ], + "spans": [ + { + "bbox": [ + 53, + 456, + 286, + 487 + ], + "type": "text", + "content": "[79] Haifeng Xia, Handong Zhao, and Zhengming Ding. Adaptive adversarial network for source-free domain adaptation. In ICCV, 2021. 1, 2, 3, 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 53, + 490, + 286, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 490, + 286, + 521 + ], + "spans": [ + { + "bbox": [ + 53, + 490, + 286, + 521 + ], + "type": "text", + "content": "[80] Shaoan Xie, Zibin Zheng, Liang Chen, and Chuan Chen. Learning semantic representations for unsupervised domain adaptation. In ICML, 2018. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 53, + 524, + 286, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 524, + 286, + 555 + ], + "spans": [ + { + "bbox": [ + 53, + 524, + 286, + 555 + ], + "type": "text", + "content": "[81] Minghao Xu, Hang Wang, Bingbing Ni, Qi Tian, and Wenjun Zhang. Cross-domain detection via graph-induced prototype alignment. In CVPR, 2020. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 53, + 557, + 286, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 557, + 286, + 599 + ], + "spans": [ + { + "bbox": [ + 53, + 557, + 286, + 599 + ], + "type": "text", + "content": "[82] Ruijia Xu, Guanbin Li, Jihan Yang, and Liang Lin. Larger norm more transferable: An adaptive feature norm approach for unsupervised domain adaptation. In ICCV, 2019. 7" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 53, + 602, + 286, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 602, + 286, + 645 + ], + "spans": [ + { + "bbox": [ + 53, + 602, + 286, + 645 + ], + "type": "text", + "content": "[83] Tongkun Xu, Weihua Chen, WANG Pichao, Fan Wang, Hao Li, and Rong Jin. Cdtrans: Cross-domain transformer for unsupervised domain adaptation. In International Conference on Learning Representations, 2021. 8" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 53, + 647, + 286, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 647, + 286, + 689 + ], + "spans": [ + { + "bbox": [ + 53, + 647, + 286, + 689 + ], + "type": "text", + "content": "[84] Hongliang Yan, Yukang Ding, Peihua Li, Qilong Wang, Yong Xu, and Wangmeng Zuo. Mind the class weight bias: Weighted maximum mean discrepancy for unsupervised domain adaptation. In CVPR, 2017. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 53, + 692, + 286, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 692, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 53, + 692, + 286, + 713 + ], + "type": "text", + "content": "[85] Fan Yang, Kai Wu, Shuyi Zhang, Guannan Jiang, Yong Liu, Feng Zheng, Wei Zhang, Chengjie Wang, and Long" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 545, + 706 + ], + "type": "list", + "angle": 0, + "index": 36, + "blocks": [ + { + "bbox": [ + 331, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 331, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 331, + 73, + 545, + 95 + ], + "type": "text", + "content": "Zeng. Class-aware contrastive semi-supervised learning. In CVPR, 2022. 2, 4, 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 312, + 96, + 545, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 96, + 545, + 128 + ], + "spans": [ + { + "bbox": [ + 312, + 96, + 545, + 128 + ], + "type": "text", + "content": "[86] Jinyu Yang, Jingjing Liu, Ning Xu, and Junzhou Huang. Tvt: Transferable vision transformer for unsupervised domain adaptation. In WACV, 2023. 8" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 312, + 129, + 545, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 129, + 545, + 172 + ], + "spans": [ + { + "bbox": [ + 312, + 129, + 545, + 172 + ], + "type": "text", + "content": "[87] Shiqi Yang, Joost van de Weijer, Luis Herranz, Shangling Jui, et al. Exploiting the intrinsic neighborhood structure for source-free domain adaptation. In NeurIPS, 2021. 1, 2, 3, 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 312, + 175, + 545, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 175, + 545, + 217 + ], + "spans": [ + { + "bbox": [ + 312, + 175, + 545, + 217 + ], + "type": "text", + "content": "[88] Shiqi Yang, Yaxing Wang, Joost van de Weijer, Luis Herranz, and Shangling Jui. Unsupervised domain adaptation without source data by casting a bait. arXiv preprint arXiv:2010.12427, 2020. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 312, + 219, + 545, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 219, + 545, + 251 + ], + "spans": [ + { + "bbox": [ + 312, + 219, + 545, + 251 + ], + "type": "text", + "content": "[89] Shiqi Yang, Yaxing Wang, Joost van de Weijer, Luis Herranz, and Shangling Jui. Generalized source-free domain adaptation. In ICCV, 2021. 2, 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 312, + 252, + 545, + 294 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 252, + 545, + 294 + ], + "spans": [ + { + "bbox": [ + 312, + 252, + 545, + 294 + ], + "type": "text", + "content": "[90] Shiqi Yang, Yaxing Wang, Kai Wang, Shangling Jui, and Joost van de Weijer. Attracting and dispersing: A simple approach for source-free domain adaptation. In NeurIPS, 2022. 2, 3, 5, 6, 7" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 312, + 297, + 545, + 340 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 297, + 545, + 340 + ], + "spans": [ + { + "bbox": [ + 312, + 297, + 545, + 340 + ], + "type": "text", + "content": "[91] Bowen Zhang, Yidong Wang, Wenxin Hou, Hao Wu, Jindong Wang, Manabu Okumura, and Takahiro Shinozaki. Flexmatch: Boosting semi-supervised learning with curriculum pseudo labeling. NeurIPS, 2021. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 312, + 342, + 545, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 342, + 545, + 373 + ], + "spans": [ + { + "bbox": [ + 312, + 342, + 545, + 373 + ], + "type": "text", + "content": "[92] Jing Zhang, Zewei Ding, Wanqing Li, and Philip Ogunbona. Importance weighted adversarial nets for partial domain adaptation. In CVPR, 2018. 7" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 312, + 375, + 545, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 375, + 545, + 418 + ], + "spans": [ + { + "bbox": [ + 312, + 375, + 545, + 418 + ], + "type": "text", + "content": "[93] Pan Zhang, Bo Zhang, Ting Zhang, Dong Chen, Yong Wang, and Fang Wen. Prototypical pseudo label denoising and target structure learning for domain adaptive semantic segmentation. In CVPR, 2021. 3, 4" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 312, + 419, + 545, + 451 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 419, + 545, + 451 + ], + "spans": [ + { + "bbox": [ + 312, + 419, + 545, + 451 + ], + "type": "text", + "content": "[94] Weichen Zhang, Wanli Ouyang, Wen Li, and Dong Xu. Collaborative and adversarial network for unsupervised domain adaptation. In CVPR, 2018. 3" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 312, + 453, + 545, + 484 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 453, + 545, + 484 + ], + "spans": [ + { + "bbox": [ + 312, + 453, + 545, + 484 + ], + "type": "text", + "content": "[95] Yiheng Zhang, Zhaofan Qiu, Ting Yao, Dong Liu, and Tao Mei. Fully convolutional adaptation networks for semantic segmentation. In CVPR, 2018. 1" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 312, + 486, + 545, + 517 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 486, + 545, + 517 + ], + "spans": [ + { + "bbox": [ + 312, + 486, + 545, + 517 + ], + "type": "text", + "content": "[96] Yixin Zhang and Zilei Wang. Joint adversarial learning for domain adaptation in semantic segmentation. In AAAI, 2020. 2" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 312, + 519, + 545, + 551 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 519, + 545, + 551 + ], + "spans": [ + { + "bbox": [ + 312, + 519, + 545, + 551 + ], + "type": "text", + "content": "[97] Yixin Zhang, Zilei Wang, and Yushi Mao. Rpn prototype alignment for domain adaptive object detector. In CVPR, 2021. 3" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 312, + 553, + 545, + 596 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 553, + 545, + 596 + ], + "spans": [ + { + "bbox": [ + 312, + 553, + 545, + 596 + ], + "type": "text", + "content": "[98] Ziyi Zhang, Weikai Chen, Hui Cheng, Zhen Li, Siyuan Li, Liang Lin, and Guanbin Li. Divide and contrast: Source-free domain adaptation via adaptive contrastive learning. In NeurIPS, 2022. 6" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 312, + 598, + 545, + 630 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 598, + 545, + 630 + ], + "spans": [ + { + "bbox": [ + 312, + 598, + 545, + 630 + ], + "type": "text", + "content": "[99] Yangtao Zheng, Di Huang, Songtao Liu, and Yunhong Wang. Cross-domain object detection through coarse-to-fine feature adaptation. In CVPR, 2020. 3" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 308, + 632, + 545, + 662 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 632, + 545, + 662 + ], + "spans": [ + { + "bbox": [ + 308, + 632, + 545, + 662 + ], + "type": "text", + "content": "[100] Yang Zou, Zhiding Yu, Xiaofeng Liu, BVK Kumar, and Jinsong Wang. Confidence regularized self-training. In ICCV, 2019. 8" + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 308, + 665, + 545, + 706 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 665, + 545, + 706 + ], + "spans": [ + { + "bbox": [ + 308, + 665, + 545, + 706 + ], + "type": "text", + "content": "[101] Yang Zou, Zhiding Yu, BVK Vijaya Kumar, and Jinsong Wang. Unsupervised domain adaptation for semantic segmentation via class-balanced self-training. In ECCV, 2018. 3" + } + ] + } + ], + "index": 35 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "7629" + } + ] + } + ], + "index": 37 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Class-Balancing Diffusion Models/eae79914-520a-4687-a1d4-1e04f53e44f8_content_list.json b/2023/Class-Balancing Diffusion Models/eae79914-520a-4687-a1d4-1e04f53e44f8_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..ee6fd85330553a7eae564ca8ebc0f0d257b702e7 --- /dev/null +++ b/2023/Class-Balancing Diffusion Models/eae79914-520a-4687-a1d4-1e04f53e44f8_content_list.json @@ -0,0 +1,1518 @@ +[ + { + "type": "text", + "text": "Class-Balancing Diffusion Models", + "text_level": 1, + "bbox": [ + 313, + 130, + 656, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yiming Qin", + "bbox": [ + 94, + 188, + 199, + 208 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Huangjie Zheng", + "bbox": [ + 233, + 188, + 372, + 208 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Jiangchao Yao $^{1,3}$", + "bbox": [ + 408, + 188, + 563, + 208 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Mingyuan Zhou", + "bbox": [ + 596, + 188, + 733, + 208 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Ya Zhang $^{1,3}$", + "bbox": [ + 772, + 188, + 870, + 208 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Cooperative Medianet Innovation Center, Shanghai Jiao-Tong University", + "bbox": [ + 240, + 228, + 727, + 244 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ University of Texas, Austin", + "bbox": [ + 390, + 244, + 578, + 258 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{3}$ Shanghai AI Laboratory", + "bbox": [ + 400, + 258, + 566, + 275 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 286, + 313, + 301 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Diffusion-based models have shown the merits of generating high-quality visual data while preserving better diversity in recent studies. However, such observation is only justified with curated data distribution, where the data samples are nicely pre-processed to be uniformly distributed in terms of their labels. In practice, a long-tailed data distribution appears more common and how diffusion models perform on such class-imbalanced data remains unknown. In this work, we first investigate this problem and observe significant degradation in both diversity and fidelity when the diffusion model is trained on datasets with class-imbalanced distributions. Especially in tail classes, the generations largely lose diversity and we observe severe mode-collapse issues. To tackle this problem, we set from the hypothesis that the data distribution is not class-balanced, and propose Class-Balancing Diffusion Models (CBDM) that are trained with a distribution adjustment regularizer as a solution. Experiments show that images generated by CBDM exhibit higher diversity and quality in both quantitative and qualitative ways. Our method benchmarked the generation results on CIFAR100/CIFAR100LT dataset and shows outstanding performance on the downstream recognition task.", + "bbox": [ + 75, + 316, + 473, + 650 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 679, + 209, + 695 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In recent years, log-likelihood-based diffusion models have evolved rapidly and established new benchmarks on a range of generation tasks [1,7]. Based on them, researchers have been able to further control the model generation process and the generation quality. This improves the applications of generative models in numerous domains including text-image generation [31], image editing [28, 38], speech synthesis [17], medical imaging [26, 44], video generation [13] and adversarial learning [18, 33], etc.", + "bbox": [ + 75, + 703, + 468, + 839 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Although diffusion models are known for the power of high fidelity and diversity in generation, most of the existing diffusion models are trained with the hypothesis that the data are uniformly distributed w.r.t. their labels. However,", + "bbox": [ + 75, + 839, + 470, + 900 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/b52f1f93a407fa27ecdca458a5edea70ef052a41c50f7baf264ddbde5122540d.jpg", + "image_caption": [ + "Figure 1. Generation degrades along with class frequency. Semantics of generated images become less recognizable when class frequency decreases, while the FID score increases significantly." + ], + "image_footnote": [], + "bbox": [ + 511, + 292, + 883, + 484 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "in the real world, the distribution is often very skewed. Especially for many domain-specific generation tasks such as medical images [14], fine-grained dataset for taxology [15] and data grabbed from the web [24], it is difficult to collect large amounts of data for each class equally, and the size of the training set for head and tail categories can differ by a factor of hundred or more. For such datasets, unconditional diffusion models tend to produce a significant portion of low-quality images. Conditional models, as shown in Figure 1, generate head class images with satisfying performance, while conversely the generated images on tail classes are very likely to show unrecognizable semantics. Concerning training generative models with limited data, there already exist several methods [21, 37, 50] based on GAN models [2]. However, quite few studies examine the impact of imbalanced class distribution [34] especially on diffusion models, which is practical yet under-explored.", + "bbox": [ + 496, + 579, + 892, + 837 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Our work first introduces diffusion models to imbalance generation tasks on several long-tailed datasets [19], and then build some straightforward baselines according to the common methods used in long-tailed recognition [27, 29].", + "bbox": [ + 496, + 839, + 893, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 44 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "18434", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To overcome the potential degeneration induced by the skewed distribution, we propose a novel Class-Balancing Diffusion Model (CBDM). Theoretically, CBDM resorts to adjusting the conditional transfer probability during sampling in order to implicitly force generated images to have a balanced prior distribution during every sampling step. Technically, the adjusted transfer probability of CBDM results in an additional MSE-form loss for a conditional diffusion model, which functions as a regularizer. Intuitively, this loss augments the similarity of generated images conditioned on different classes, and turns out to be an effective approach to transfer common information from head classes to tail classes without hurting the model's expressiveness on head classes. CBDM can be implemented within several flines of codes, and its lighter version admits fine-tuning an existing conditional model. We conducted extensive experiments on CIFAR10/CIFAR100 and their corresponding long-tailed dataset to show the promise of CBDM over existing state-of-the-art methods. In a nutshell, the contributions of this work can be summarized as follows:", + "bbox": [ + 76, + 90, + 472, + 393 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We identify the severe degeneration problem of diffusion models in long-tailed generation tasks and benchmark some straightforward baselines in this direction.", + "- We propose a new perspective to handle the generation quality collapse on tail classes, and derive a novel Class-Balancing Diffusion Model, which is effective and lightweight as a regularizer to existing methods.", + "- We validate that CBDM is capable of generating more diverse images with convincing fidelity, especially for datasets with large number of categories. In addition, CBDM is robust with accelerating algorithms such as DDIM [43], and can be transplanted to different conditional diffusion-based backbones easily." + ], + "bbox": [ + 96, + 405, + 468, + 622 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 76, + 648, + 225, + 664 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Diffusion models Diffusion models are recently proposed generative models [42] based on non-equilibrium thermodynamics. Conditional diffusion models [7] encode label information into the generation process and improve largely the generation performance. The guidance structure proposed in [7] makes it possible to control the generation process through an external module. Based on a similar intuition, researchers arrive to realize diverse functions, such as guided adversarial purification [33], few-shot generation [9] and so on [36, 41]. The drawback of classifier guidance (noted as CG) [7] lies in its requirement of training another auxiliary classifier. To address the issue, classifier-free guidance (noted as CFG) [12] proposed a mechanism that uses the generator itself to express the class guidance information. CFG is proved to be a resource efficient method", + "bbox": [ + 76, + 672, + 472, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "and achieves outstanding performance on large models [31]. Moreover, CFG only requires to add one line in training, which can be easily transplanted on different models.", + "bbox": [ + 498, + 90, + 890, + 137 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Long-tail recognition The problem of long-tailed distribution is a common dilemma in machine learning and have been widely explored in the area of discriminative models, i.e., long-tailed recognition. There mainly exist three paradigms in this domain, namely Class Re-balancing [27, 29], Information Augmentation [25,47,49] and Module Improvement [16, 32]. Among them, Class Re-balancing provides the best explainability, and its most common practice is re-sampling [27]. Thereafter, stemming from modifying the objective function from the global error rate to the class average one, [29] propose logit adjustment which has shown an impressive performance. Another effective method is based on Information Augmentation, which uses head class feature information to augment tail classes [5, 25, 47, 49]. However, discriminative models map data from higher to lower dimensions, while generative models map images from lower to higher dimensions. Thus, the mechanism of class rebalancing between them may be completely different and how to design the balancing method remains under-explored.", + "bbox": [ + 496, + 148, + 893, + 452 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Generative models based on limited data GAN [2, 6] is the most dominant model in the field of image synthesis in recent years. Given the requirement of large-scale data to train the generative models, a part of researchers focus on improving the performance of GAN models under the small datasets. To address the overfitting issue of the discriminator, a number of regularization methods [22, 37, 45] has been proposed. An alternative solution is Data Augmentation. As augmentation information is very prone to leak to the generator, researchers proposed improved augmentation strategies such as Differentiable Augmentation (DiffAug) [50] and Adaptive Augmentation (ADA) [21]. These augmentation ways can be transferred to diffusion models as baselines to help the generation of tail-class samples.", + "bbox": [ + 496, + 465, + 893, + 676 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 500, + 696, + 591, + 712 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this section, we first give the basic notations following the classical DDPM model, and then introduce the class-imbalanced generation setting. In the third part, we introduce our CBDM algorithm and present its training details including implementation and hyper-parameter settings.", + "bbox": [ + 496, + 723, + 890, + 800 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3.1. Preliminary", + "text_level": 1, + "bbox": [ + 500, + 814, + 629, + 830 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Diffusion models leverage a pre-defined forward process in training, where a clean image distribution $q(\\pmb{x}_0)$ can be corrupted to a noisy distribution $q(\\pmb{x}_t|\\pmb{x}_0)$ at a specified timestep $t$ . Given a pre-defined variance schedule $\\{\\beta_t\\}_{1:T}$ ,", + "bbox": [ + 496, + 839, + 890, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "18435", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/93689ce378ac70f873db28d10f3497b9583aa537a599b4fc905f992fe5c9857e.jpg", + "image_caption": [ + "Figure 2. Algorithm (left) and generation (right) visualization. In the left figure, we show that an extra regularization loss $gL_{r}$ proportional to the diffusion step $t$ is added during training. This loss function pushes the sampling distribution (curves on the surface) to wider region while preventing it to be excessively distorted compared to the ground truth distribution (gradient color on the background)." + ], + "image_footnote": [], + "bbox": [ + 84, + 99, + 638, + 273 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/5020ec0e967463607da0afac543321a5b90bdf43930986ac8b6b71e2fd04bdb2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 656, + 119, + 888, + 272 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "the noisy distribution at any intermediate timestep is", + "bbox": [ + 76, + 352, + 423, + 368 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nq \\left(\\boldsymbol {x} _ {t} \\mid \\boldsymbol {x} _ {0}\\right) = \\mathcal {N} \\left(\\sqrt {\\bar {\\alpha} _ {t}} \\boldsymbol {x} _ {0}, \\left(1 - \\bar {\\alpha} _ {t}\\right) \\boldsymbol {I}\\right); \\quad \\bar {\\alpha} _ {t} = \\prod_ {i = 1} ^ {t} \\left(1 - \\beta_ {i}\\right).\n$$\n", + "text_format": "latex", + "bbox": [ + 86, + 378, + 455, + 419 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To reverse such forward process, a generative model $\\theta$ learns to estimate the analytical true posterior in order to recover $x_{t - 1}$ from $x_{t}$ as follows:", + "bbox": [ + 76, + 429, + 468, + 474 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {\\theta} D _ {\\mathrm {K L}} [ q (\\pmb {x} _ {t - 1} | \\pmb {x} _ {t}, \\pmb {x} _ {0}) | | p _ {\\theta} (\\pmb {x} _ {t - 1} | \\pmb {x} _ {t}) ]; \\forall t \\in \\{1, \\dots , T \\},\n$$\n", + "text_format": "latex", + "bbox": [ + 78, + 486, + 465, + 508 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "and such an objective can be reduced to a simple denoising estimation loss [11]:", + "bbox": [ + 76, + 518, + 468, + 550 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {D D P M}} = \\mathbb {E} _ {t, \\boldsymbol {x} _ {0} \\sim q (\\boldsymbol {x} _ {0}), \\epsilon \\sim \\mathcal {N} (\\boldsymbol {0}, I)} \\left[ \\| \\boldsymbol {\\epsilon} - \\boldsymbol {\\epsilon} _ {\\theta} \\left(\\sqrt {\\bar {\\alpha}} \\boldsymbol {x} _ {0} + \\sqrt {1 - \\bar {\\alpha} _ {t}} \\boldsymbol {\\epsilon}, t\\right) \\| ^ {2} \\right] \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 76, + 561, + 467, + 590 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "For the case where label information is available, the model is trained to estimate the noise as above in both conditional cases $\\epsilon_{\\theta}(\\pmb{x}_t,y,t)$ with data-label pairs $(x_0,y)$ and unconditional case $\\epsilon_{\\theta}(\\pmb{x}_t,t)$ . In the sampling, the label-guided model estimates the noise with a linear interpolation $\\hat{\\epsilon} = (1 + \\omega)\\pmb{\\epsilon}_{\\theta}(\\pmb{x}_t,y,t) - \\omega \\pmb{\\epsilon}_{\\theta}(\\pmb{x}_t,t)$ to recover $\\pmb{x}_{t-1}$ , which is often referred as Classifier-Free Guidance (CFG) [12].", + "bbox": [ + 76, + 592, + 468, + 698 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Class-Balancing Diffusion Models", + "text_level": 1, + "bbox": [ + 76, + 707, + 372, + 723 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Current diffusion models assume the data distribution to be uniform in class, and thus equally treat samples in the training stage. However, based on our observation, such training strategy leads to degradation in generation quality. Below, we provide an analysis that motivates our Class-Balancing Diffusion Models (CBDM).", + "bbox": [ + 76, + 729, + 468, + 820 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Suppose $q(\\pmb{x}, y)$ is the data distribution that we need to match with the joint distribution $p_{\\theta}(\\pmb{x}, y)$ predicted by a generative model. We analyze their difference from the density ratio $r = \\frac{q(\\pmb{x}, y)}{p_{\\theta}(\\pmb{x}, y)} = \\frac{q(\\pmb{x} \\mid y)}{p_{\\theta}(\\pmb{x} \\mid y)} \\cdot \\frac{q(y)}{p_{\\theta}(y)}$ . When the true label distribution $q(y)$ is the same as the prior $p_{\\theta}(y)$ ,", + "bbox": [ + 75, + 821, + 468, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "which is usually assumed to be uniform, the density ratio $r$ is reduced the conditional term to learn a conditional model $p_{\\theta}(\\boldsymbol{x} \\mid y)$ . However, when such a hypothesis is violated, for head classes, $\\frac{q(y)}{p_{\\theta}(y)}$ would result in a larger weight that makes the model biased and hurt tail classes, and vice versa. Empirically, we observe that the generation degrades more on tail classes, as illustrated in Figure 1. Moreover, as shown in Figure 2a, compared to head classes, DDPM cannot well capture the tail-class data distribution and the mode is poorly covered during the sampling process. As a result, generations of tail classes often have poor quality and diversity, shown in Figure 2b.", + "bbox": [ + 496, + 352, + 890, + 536 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To tackle this issue, the most intuitive approach lies in adjusting the prior label distribution through a class balanced re-sampling. However, such abrupt adjustment easily leads to negative improvement in experiments. The step-by-step sampling nature of diffusion models provides another aspect to adjust this distribution more softly. In this spirit, we propose to calibrate the learning process through the conditional transfer probability $p_{\\theta}(\\pmb{x}_{t-1}|\\pmb{x}_t,y)$ when there exists a gap between the class distribution and the prior.", + "bbox": [ + 496, + 537, + 890, + 672 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Let $p_{\\theta}^{\\star}\\left(\\boldsymbol{x}_{t-1} \\mid \\boldsymbol{x}_t, y\\right)$ be the optimum trained in the case that $\\frac{q(y)}{p_{\\theta}(y)}$ is correctly estimated, and $p_{\\theta}\\left(\\boldsymbol{x}_{t-1} \\mid \\boldsymbol{x}_t, y\\right)$ be the one trained in a class-imbalanced case. The relation between such two generative distributions can be described as the following proposition.", + "bbox": [ + 496, + 672, + 890, + 752 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Proposition 1. When training a diffusion model parameterized with $\\theta$ on a class-imbalanced dataset, its conditional reverse distribution $p_{\\theta}\\left(\\boldsymbol{x}_{t-1}|\\boldsymbol{x}_t,y\\right)$ can be corrected with an adjustment schema:", + "bbox": [ + 496, + 760, + 890, + 820 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\np _ {\\theta} ^ {\\star} \\left(\\boldsymbol {x} _ {t - 1} \\mid \\boldsymbol {x} _ {t}, y\\right) = p _ {\\theta} \\left(\\boldsymbol {x} _ {t - 1} \\mid \\boldsymbol {x} _ {t}, y\\right) \\frac {p _ {\\theta} \\left(\\boldsymbol {x} _ {t - 1}\\right)}{p _ {\\theta} ^ {\\star} \\left(\\boldsymbol {x} _ {t - 1}\\right)} \\frac {q ^ {\\star} \\left(\\boldsymbol {x} _ {t}\\right)}{q \\left(\\boldsymbol {x} _ {t}\\right)} \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 514, + 828, + 890, + 862 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The proposition above shows that, when trained on a", + "bbox": [ + 519, + 885, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "18436", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "class-imbalanced dataset, a diffusion model can still approach the true data distribution by applying a distribution adjustment schema $\\frac{p_{\\theta}(\\boldsymbol{x}_{t-1})}{p_{\\theta}^*(\\boldsymbol{x}_{t-1})}\\frac{q^*(\\boldsymbol{x}_t)}{q(\\boldsymbol{x}_t)}$ at every step $t$ . However, approximating this schema is not feasible at every sampling step, so CBDM incorporates it into the training loss function to achieve an equivalent objective, and thus gets rid of the model-free part $\\frac{q^*(\\boldsymbol{x}_t)}{q(\\boldsymbol{x}_t)}$ . By further decomposing $p_{\\theta}(\\boldsymbol{x}_{t-1})$ and $p_{\\theta}^*(\\boldsymbol{x}_{t-1})$ to the expectation of the conditional probability $p_{\\theta}^*(\\boldsymbol{x}_{t-1}|\\boldsymbol{x}_{t:T},y)$ , we present a upper bound to approximate this probability in Proposition 2.", + "bbox": [ + 75, + 90, + 472, + 250 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Proposition 2. For the adjusted loss $\\mathcal{L}_{DM}^{\\star} = \\sum_{t=1}^{T} \\mathcal{L}_{t-1}^{\\star}$ , an upper-bound of the target training objective to calibrate at timestep $t$ (i.e., $\\mathcal{L}_{t-1}^{\\star}$ ) can be derived as:", + "bbox": [ + 75, + 258, + 470, + 306 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\sum_ {t \\geq 1} \\mathcal {L} _ {t - 1} ^ {\\star} = \\sum_ {t \\geq 1} D _ {\\mathrm {K L}} [ q (\\pmb {x} _ {t - 1} | \\pmb {x} _ {t}, \\pmb {x} _ {0}) | | p _ {\\theta} ^ {\\star} (\\pmb {x} _ {t - 1} | \\pmb {x} _ {t}, y) ] \\\\ \\leq \\sum_ {t \\geq 1} [ \\underbrace {D _ {\\mathrm {K L}} [ q (\\boldsymbol {x} _ {t - 1} | \\boldsymbol {x} _ {t} , \\boldsymbol {x} _ {0}) | | p _ {\\theta} (\\boldsymbol {x} _ {t - 1} | \\boldsymbol {x} _ {t} , y) ]} _ {\\text {D i f f u s i o n m o d e l l o s s} \\mathcal {L} _ {D M}} \\\\ + \\underbrace {t \\mathbb {E} _ {y ^ {\\prime} \\sim q _ {y} ^ {\\star}} [ D _ {\\mathrm {K L}} [ p _ {\\theta} (\\boldsymbol {x} _ {t - 1} | \\boldsymbol {x} _ {t}) | | p _ {\\theta} (\\boldsymbol {x} _ {t - 1} | \\boldsymbol {x} _ {t} , y ^ {\\prime}) ] ] ]} _ {D i s t r i b u t i o n a d j u s t m e n t l o s s \\mathcal {L} _ {r}}, \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 76, + 318, + 468, + 435 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The upper bound in the above proposition can be considered as two parts. The first term $\\mathcal{L}_{\\mathrm{DM}}$ corresponds to an ordinary DDPM loss [11] e.g., Eqn. (1), and the second loss $\\mathcal{L}_{\\mathrm{r}}$ is used to adjust the distribution as a regularization term. Roughly speaking, $\\mathcal{L}_{\\mathrm{r}}$ increases the similarity between the model's output and a random target class. Thus, it reduces the risk of overfitting on the head classes, and enlarges the generation diversity for tail class through knowledge obtained from other classes. When $q_{\\mathcal{Y}}^{\\star}$ is less longtailed than the dataset, this loss also increases the probability for underrepresented tail samples to be chosen during training.", + "bbox": [ + 75, + 463, + 468, + 630 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Training algorithm", + "text_level": 1, + "bbox": [ + 76, + 638, + 261, + 656 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The detailed training algorithm of CBDM is presented in Alg. (1). In the algorithm, we reduce the distribution adjustment loss $\\mathcal{L}_r$ as a square error loss with Monte-Carlo samples as indicated by Eqn.(3), where $\\mathcal{V}$ is a set of samples that drawn following the distribution $q_{\\mathcal{V}}^{\\star}$ and $y$ denotes the image label. Note for CFG [12], there is a fixed probability (usually $10\\%$ ) to drop the condition, i.e., $y = \\text{None}$ .", + "bbox": [ + 75, + 662, + 468, + 768 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {r}} \\left(\\boldsymbol {x} _ {t}, y, t\\right) = \\frac {1}{| \\mathcal {Y} |} \\sum_ {y ^ {\\prime} \\in \\mathcal {Y}} [ t \\| \\epsilon_ {\\theta} \\left(\\boldsymbol {x} _ {t}, y\\right) - \\epsilon_ {\\theta} \\left(\\boldsymbol {x} _ {t}, y ^ {\\prime}\\right) \\| ^ {2} ], \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 88, + 791, + 468, + 829 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For the implementation in practice, CBDM can be plugged into any existing conditional diffusion models by adopting their model architecture and adjusting the training loss $\\mathcal{L}_{\\mathrm{DM}}$ following lines 6 - 11. Specifically, the choice", + "bbox": [ + 75, + 839, + 470, + 902 + ], + "page_idx": 3 + }, + { + "type": "code", + "sub_type": "algorithm", + "code_caption": [ + "Algorithm 1 Training algorithm of CBDM." + ], + "code_body": "1: for Every batch of size N do \n2: for $(\\pmb{x}_0^{(i)}, y^{(i)})$ in this batch do \n3: Sample $\\pmb{\\epsilon}^{(i)} \\sim \\mathcal{N}(\\pmb{0}, \\pmb{I}), t \\sim \\mathcal{U}(\\{0, 1, \\dots, T\\})$ \n4: $x_{t}^{(i)} = \\sqrt{\\bar{\\alpha}_{t}} x_{0}^{(i)} + \\sqrt{1 - \\bar{\\alpha}_{t}} \\epsilon^{(i)}$ \n5: Calculate $\\mathcal{L}_{\\mathrm{DM}} = \\| \\pmb{\\epsilon}^{(i)} - \\pmb{\\epsilon}_{\\theta}(\\pmb{x}_{t}^{(i)}, y^{(i)}) \\|^{2}$ \n6: Sample $y'(i)$ from $q_{\\gamma}^{\\star}$ \n7: Calculate the first regularization term \n8: $\\mathcal{L}_{r} = t\\tau \\| \\pmb{\\epsilon}_{\\theta}(\\pmb{x}_{t}^{(i)}, y^{(i)}) - \\mathrm{sg}(\\pmb{\\epsilon}_{\\theta}(\\pmb{x}_{t}^{(i)}, y'^{(i)})) \\|^{2}$ \n9: Calculate the regularization commitment term \n10: $\\mathcal{L}_{rc} = t\\tau \\| \\mathrm{sg}(\\pmb{\\epsilon}_{\\theta}(\\pmb{x}_{t}^{(i)}, y^{(i)})) - \\pmb{\\epsilon}_{\\theta}(\\pmb{x}_{t}^{(i)}, y'^{(i)}) \\|^{2}$ \n11: Update with $\\mathcal{L}_{\\mathrm{CBDM}} = \\mathcal{L}_{\\mathrm{DM}} + \\mathcal{L}_{r} + \\gamma \\mathcal{L}_{rc}$ \n12: end for \n13: end for", + "bbox": [ + 506, + 109, + 890, + 318 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "of the regularization weight $\\tau$ affects the sharpness of the density ratio $\\frac{p_{\\theta}(\\boldsymbol{x}_t)}{p_{\\theta}^*(\\boldsymbol{x}_t)}$ . For the theoretical analysis, please refer to our proof of Prop. (2) in the Appendix. In addition, the choice of the sampling set $\\mathcal{V}$ is another important perspective of CBDM, which depends on the target distribution we wish to adjust. Without loss of generality, we discuss two cases here. On the one hand, we can adjust the label distribution to a class-balanced label distribution, where we sample labels to construct $\\mathcal{V}^{bal}$ . On the other hand, if the data distribution is heavily long-tailed, we can also target the adjusted distribution to a relatively less class-imbalanced distribution for stabilized training. In our experiments, we show CBDM can work well for both cases in different mechanisms.", + "bbox": [ + 496, + 366, + 890, + 580 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Moreover, we observe that naively optimizing with this loss could make the model collapse to some trivial solutions, where the model outputs the same result regardless of the condition $y$ and thus degenerates the conditional generation performance. Therefore, we follow previous works to apply a stop gradient operation [4, 46] to prevent this issue. The final loss of CBDM is", + "bbox": [ + 496, + 592, + 890, + 699 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} _ {\\mathrm {C B D M}} ^ {\\tau , \\gamma , \\mathcal {Y}} (\\pmb {x} _ {t}, y, t, \\pmb {\\epsilon}) = \\| \\pmb {\\epsilon} _ {\\theta} (\\pmb {x} _ {t}, y) - \\pmb {\\epsilon} \\| ^ {2} \\\\ + \\frac {\\tau t}{| \\mathcal {Y} |} \\sum_ {y ^ {\\prime} \\in \\mathcal {Y}} \\left(\\left\\| \\boldsymbol {\\epsilon} _ {\\theta} \\left(\\boldsymbol {x} _ {t}, y\\right) - \\operatorname {s g} \\left(\\boldsymbol {\\epsilon} _ {\\theta} \\left(\\boldsymbol {x} _ {t}, y ^ {\\prime}\\right)\\right) \\right\\| ^ {2} \\right. \\\\ + \\gamma \\| \\mathrm {s g} \\big (\\pmb {\\epsilon} _ {\\theta} (\\pmb {x} _ {t}, y) \\big) - \\pmb {\\epsilon} _ {\\theta} (\\pmb {x} _ {t}, y ^ {\\prime}) \\| ^ {2}), (4) \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 509, + 729, + 890, + 809 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathrm{sg}(\\cdot)$ denotes the stop gradient operation; $\\tau, \\gamma$ are weights for the regularization and the commitment term respectively with $\\gamma$ set to $\\frac{1}{4}$ in a default setting, and $|\\mathcal{V}|$ denotes the number of elements in the label set.", + "bbox": [ + 496, + 839, + 890, + 898 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "18437", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/dd0ccc3c990d101184872c016a8f0d46a73b9ecb4f44c1ece8ca8aedd15ec652.jpg", + "image_caption": [ + "(a) DDPM / Tail Class 94" + ], + "image_footnote": [], + "bbox": [ + 83, + 90, + 267, + 233 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/a5e275ecd61e93b3ee7ff383f1185faf0a4e67d4a25b8eb536ef5ccfc200bed5.jpg", + "image_caption": [ + "(b) CBDM / Tail Class 94" + ], + "image_footnote": [], + "bbox": [ + 277, + 90, + 464, + 233 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/946995c13751da3c24e493e2e4d8a3126d1ab1943c5615a9938206a54b8d6319.jpg", + "image_caption": [ + "(c) DDPM / Body Class 62", + "Figure 3. Comparison of image generation on heavily tail-distributed (94) and mild tail-distributed (62) classes between DDPM and CBDM." + ], + "image_footnote": [], + "bbox": [ + 83, + 252, + 267, + 395 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/1f6736bb9992078b5667558852d87a159b20deba9231a1a4a87b0ab4651d0ef8.jpg", + "image_caption": [ + "(d) CBDM / Body Class 62" + ], + "image_footnote": [], + "bbox": [ + 276, + 252, + 460, + 395 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experimental results", + "text_level": 1, + "bbox": [ + 76, + 496, + 274, + 513 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Experimental setup", + "text_level": 1, + "bbox": [ + 76, + 523, + 264, + 540 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Datasets We first chose two common datasets that are widely used in the domain of image synthesis, CIFAR10/CIFAR100, and their corresponding long-tailed versions CIFAR10LT and CIFAR100LT. The construction of CIFAR10LT and CIFAR100LT follows [3], where the size decreases exponentially with its category index according to the imbalance factor $imb = 0.01$ . We also conduct experiments on 3 higher-resolution datasets, whose details and results could be found in the Appendix.", + "bbox": [ + 75, + 549, + 468, + 686 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Implementation details We strictly follow the training configurations of the baseline models. For DDPM, we set diffusion schedule $\\beta_{1} = 10^{-4}$ and $\\beta_{T} = 0.02$ with $T = 1,000$ , and optimize the network with an Adam optimizer whose learning rate is 0.0002 after 5,000 epochs of warmup. Considering that the size and semantic complexity of the datasets vary greatly, we choose appropriate epochs for each dataset. In the ablation studies, we also include EDM [20] and Diffusion-ViT [48] as our backbones to show the compatibility with different architectures and score-matching methods. Note that, when a model does not use conditional input, we follow [7] to slightly modify the backbone [8] by adding an extra embedding layer.", + "bbox": [ + 75, + 703, + 470, + 902 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Baselines We first adopt following classic methods to build baselines: re-sampling (RS), soft re-sampling methods (RS-SQRT) [27], and augmentation-based methods including Differentiable Augmentation (DiffAug) [50] and Adaptive Augmentation (ADA) [21]. Besides diffusion-based baselines, previous state-of-the-art generative models that study the long-tailed distribution, e.g., CBGAN [35], and group spectral regularization for GANs [34], are also included in the comparison. Precisely, RS follows the uniform class distribution and RS-SQRT [27] samples with a probability determined by the square root of class frequency. And DiffAug [50] is directly applied to all training images following their default threshold. Following [20], ADA [21] is not only applied to images, the augmentation pipeline is also encoded as a condition via an additional embedding layer in the U-Net.", + "bbox": [ + 496, + 90, + 890, + 332 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Metrics CBDM and corresponding baselines are evaluated in terms of both generation diversity and fidelity via Frechet Inception Distance (FID) [10], Inception Score (IS) [40], Recall [23] and $F_{\\beta}$ [39]. We measure Recall and $F_{\\beta}$ using Inception-V3 features, and take respectively $K = 5$ for Recall, 1/8 and 8 for the threshold in $F_{\\beta}$ , and 20 times of class number as the clustering number of $F_{\\beta}$ to capture the inner class variance. To this end, Recall and $F_{8}$ can be regarded as diversity metrics and IS and $F_{1/8}$ are more inclined to measure fidelity. In evaluation, we take the class-balanced version corresponding to those used in training, and all metrics are measured with 50k generated images (10k for ablation experiments). The classifier-free guidance is adapted in sampling and we tune the guidance strength $\\omega$ for both baselines and our method to ensure the best performance. Specifically, we take $\\omega = 1.6, 0.8, 1.0, 0.8$ respectively for the four CIFAR datasets mentioned above.", + "bbox": [ + 496, + 337, + 890, + 595 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Main results", + "text_level": 1, + "bbox": [ + 500, + 604, + 633, + 619 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this part, we consider DDPM as the most direct baseline. As shown in Table 1, CBDM overall outperforms DDPM on all datasets except the IS is slightly lower on CIFAR10. For datasets with more classes, the improvements are more significant in terms of both diversity and fidelity, as on CIFAR100LT and CIFAR100. We further investigate the performance of diffusion models when combined with augmentation-based and re-sampling methods. Surprisingly, except for ADA, we consistently observe a performance degradation when these methods are compared with the vanilla DDPM. Moreover, adding the ADA augmentation with CBDM training leads to a further improvement in terms of both diversity and fidelity.", + "bbox": [ + 496, + 628, + 890, + 824 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As a qualitative justification, Figure 3 provides the visualization comparison between DDPM and CBDM on a relatively mild tail-distributed class (62) and a tail-distributed class (94). We remark that CBDM generates more diverse images. For example, on the tail class 94, the cabinets", + "bbox": [ + 496, + 825, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "18438", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/1be50f1d86ff41330a3e8ee5dfcf49a9944ca71724c33c293240408a9bd12304.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
DatasetModelFID↓F8↑Recall↑IS↑F1/8↑
CIFAR100LTDDPM [11]7.380.850.5213.110.88
+ADA [21]6.160.910.5712.710.90
+DiffAug [50]9.190.880.4711.560.86
+RS [27]10.500.650.4912.600.83
+SQRT-RS [27]9.720.660.4713.470.83
CBDM (ours)6.26 (-1.12)0.91 (+0.06)0.57 (+0.05)13.24 (+0.13)0.89 (+0.01)
+ADA [21]5.81 (-1.57)0.91 (+0.06)0.57 (+0.05)13.34 (+0.23)0.90 (+0.02)
CIFAR100DDPM3.110.970.6513.650.96
CBDM (ours)2.72 (-0.39)0.97 (±0)0.67 (+0.02)14.03 (+0.38)0.96 (±0)
CIFAR10LTDDPM5.760.970.579.170.95
CBDM (ours)5.46 (-0.30)0.97 (±0)0.59 (+0.02)9.28 (+0.11)0.95 (±0)
CIFAR10DDPM3.160.990.649.800.98
CBDM (ours)3.03 (-0.13)0.99 (±0)0.65 (+0.01)9.63 (-0.17)0.98 (±0)
", + "bbox": [ + 112, + 88, + 857, + 337 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/82052a5dd0c0971d426f54adb1988032f9171d2827b02b09542b47a3e57b2ac1.jpg", + "table_caption": [ + "Table 1. CBDM performance on different datasets. In the table, the first three columns are diversity-related, and we mark the best results in bold. As the comparison between CBDM and DDPM is more straightforward, we mark the performance gain in parentheses next to the CBDM results, using blue and red to indicate improvements and degradation respectively." + ], + "table_footnote": [], + "table_body": "
PTFTyFIDF8RecallIS
CIFAR 100LT-ytrain6.260.910.5713.24
ybal6.100.940.6412.29
ysqrt6.000.930.5913.00
CIFAR 100LTCIFAR 100LT-7.380.850.5213.11
ytrain6.200.880.5413.36
ybal5.850.920.6013.29
CIFAR 100CIFAR 10LT-5.390.660.589.43
ytrain5.050.670.609.48
ybal5.900.660.609.49
", + "bbox": [ + 81, + 412, + 470, + 575 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 2. CBDM performance for 3 mechanisms under different regularization sampling set $\\mathcal{V}$ . Column $PT$ indicates the dataset used in (pre)training, column $FT$ indicates the dataset used for fine-tuning. The rows where $\\mathcal{V}$ marked as “-” represent the results with DDPM, and rows where $FT$ marked as “-” represent the results of CBDM trained from scratch. Three diversity-preferred metrics and the best results are in bold for emphasis.", + "bbox": [ + 75, + 585, + 468, + 686 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "shown in Figure 3b have more color and texture, which justifies the improvements of diversity-related metrics in Table 1. On the contrary, DDPM only generates images that are highly similar to the training data.", + "bbox": [ + 75, + 715, + 468, + 777 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Case-by-case study. To better understand the generation conditioned on each class, we compare the FID case-by-case between DDPM and CBDM on each class. The results are shown in Figure 4. Compared to DDPM, CBDM shows more consistent improvements for the class index greater than 40. In the tail classes, CBDM outperforms than DDPM the FID by a more significant margin.", + "bbox": [ + 75, + 794, + 470, + 902 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/c558ad75a1e4270f6c288b5f82a37507320ddcc04703d71f4d45644f2e9edd1c.jpg", + "image_caption": [ + "Figure 4. FID improvement per class compared to DDPM. The curve is smoothened by a moving average of 5 unit." + ], + "image_footnote": [], + "bbox": [ + 521, + 417, + 870, + 559 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The choice of label set $\\mathcal{V}$ We investigate the effects of label distribution for CBDM under three mechanisms. The first mechanism is the vanilla CBDM given in Algorithm 1 which trains a diffusion model with regularization from scratch. Besides, we also study CBDM with two fine-tuning configurations. The first setting concerns using CBDM to adjust an existing model pre-trained on a long-tailed dataset to improve its performance. The second one refers to fine-tuning a pre-trained model to adapt to a smaller dataset with an imbalanced class distribution using CBDM. For those mechanisms, we consider three settings for sampling the label set: (1) a label distribution similar to the training set (denoted $\\mathcal{V}^{train}$ ); (2) a totally uniform label distribution (denoted $\\mathcal{V}^{bal}$ ); (3) a less long-tailed distribution compared to $\\mathcal{V}^{train}$ , whose class frequency is the square root of the original one (denoted $\\mathcal{V}^{sqrt}$ ).", + "bbox": [ + 496, + 627, + 892, + 868 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The results are shown in Table 2, where we found that CBDM performs well on all these configurations. As we", + "bbox": [ + 498, + 869, + 893, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "18439", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/b65227766b8c1e45b4a402d90bb701d283a2bcccb2a9deb9bb76887d446f3dff.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
DatasetModelFIDIS
CIFAR 100BigGAN+DA+RLC[45](2021)2.99-
DDPM3.1113.65
CBDM (ours)2.7214.03
CIFAR 100LTCBGAN [35](2021)28.17-
DDPM12.3112.69
CBDM (ours)+ADA [21]10.5012.81
CIFAR 10LTCBGAN [35](2021)32.93-
SNGAN [30]+gSR [34](2022)18.587.80
DDPM9.689.00
CBDM (ours)9.389.12
", + "bbox": [ + 81, + 88, + 472, + 277 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/a2d689151eed7765c7c0e65372750e3754f36c4b6164a7b666338c541ace9310.jpg", + "table_caption": [ + "Table 3. Comparison with long-tailed SoTAs on CIFAR. Following their setting, all methods are evaluated with $10\\mathrm{k}$ generated images and ground truth images are from their validation dataset. The publish year of every baselines are marked in next to the citation." + ], + "table_footnote": [], + "table_body": "
Training dataPrecisionRecall
CIFAR1000.700.67
CIFAR100LT + DDPMgens (50k)0.450.39
0.48 (+0.03)0.44 (+0.05)
+ CBDMgens (50k)0.49 (+0.04)0.47 (+0.08)
", + "bbox": [ + 94, + 358, + 452, + 450 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "analyzed in section 3.3, when the empirical distribution of $\\mathcal{V}$ has a significant difference from the label distribution of the dataset, adjusting with $\\mathcal{V}^{bal}$ could hurt training stability and result in performance degradation. We observe that $\\mathcal{V}^{bal}$ only shows better performance regarding generation diversity, while the generation fidelity has an obvious gap than the other settings. Using a relatively mild set $\\mathcal{V}^{sqrt}$ has better performance in terms of IS and preserves the generation diversity, which leads to the best FID among all these settings. On the contrary, when a pre-trained model is available, fine-tuning on top of it ensures better stability, and we can observe in such case using $\\mathcal{V}^{bal}$ has significant improvement than the vanilla DDPM or fine-tuning using $\\mathcal{V}^{train}$ . When adapting to a different dataset, we also observe using $\\mathcal{V}^{bal}$ causes stability issue and produce undesired results.", + "bbox": [ + 75, + 531, + 472, + 758 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Enhancement of training classifiers on long-tailed data As training models on long-tailed data usually leads to undesired classification results, we investigate whether the generated data could help improve the classifier trained on long-tailed data as a complementary evaluation metrics. Here, we train ResNet-32 models, respectively with CIFAR100, CIFAR100LT and CIFAR100LT augmented with DDPM and CBDM generations (50k samples). Table 4 shows that the training on long-tailed data results in se", + "bbox": [ + 75, + 763, + 470, + 902 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/689bcfd12a16d911291d6f4b25c019a809a4648edb10d65e6a3864fea5b65a23.jpg", + "table_caption": [ + "Table 4. Recognition results of different training data. All configurations are evaluated on the testing set of normal CIFAR100. Gains based on CIFAR100LT dataset is noted in blue." + ], + "table_footnote": [], + "table_body": "
BackboneFID\\( F_8 \\)RecallIS\\( F_{1/8} \\)
EDM [20]8.640.820.4812.160.86
+CBDM (ours)7.970.880.5212.010.86
Diffuse-ViT [48]37.60.760.457.660.62
+CBDM (ours)30.00.810.518.000.65
", + "bbox": [ + 501, + 88, + 901, + 186 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/4ceee53c06ef23e876299db383d03160c65542e78c798aa6603a7e7117fd2ff7.jpg", + "table_caption": [ + "Table 5. CBDM performance using different backbones on CIFAR100LT dataset. Three diversity-biased metrics." + ], + "table_footnote": [], + "table_body": "
DatasetFID↓F8↑Recall↑IS↑F1/8↑
CIFAR3.440.970.6913.110.96
100(+0.72)(±0)(+0.02)(-0.92)(±0)
CIFAR6.270.930.6112.560.89
100LT(+0.01)(+0.02)(+0.04)(-0.68)(±0)
", + "bbox": [ + 504, + 238, + 888, + 316 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 6. Sampling with 100 DDIM steps on CIFAR100 and CIFAR100LT. The difference compared with those using DDPM steps are reported below each result.", + "bbox": [ + 498, + 325, + 890, + 369 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "vere performance degradation. When augmenting the long-tailed data with the generated data, we observe fairly improvements in the trained classifier. Moreover, comparing the results between DDPM and CBDM, we observe that the improvements by using CBDM are more significant, which validates the effectiveness of our approach. Especially, the gain on recall is more remarkable than the gain on precision, which means that the generated images are more diverse than DDPM and justifies the results in Table 1.", + "bbox": [ + 496, + 393, + 890, + 530 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Comparison with other benchmarks We include the representative state-of-the-art long-tailed generative modeling approaches in the comparison, most of which are based on GANs since they suffer more severe problems when the training data is skewed. For a fair comparison, we strictly follow their evaluation setting and the results are reported in Table 3. From the results, DDPM surpasses the performance of the baselines on all the datasets, and CBDM exhibits even stronger performance.", + "bbox": [ + 496, + 532, + 892, + 670 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Ablations", + "text_level": 1, + "bbox": [ + 500, + 678, + 609, + 691 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Compatibility with different backbones We first investigate the performance of our method with different diffusion models. Without loss of generality, we adopt EDM [20] and Diffusion-ViT [48], which use a different score-matching method and different denoising network backbone. As shown in Table 5, compared with the corresponding baselines, CBDM improves the generation result on CIFAR100LT dataset for both backbones, demonstrating the compatibility with a variety of backbones and the effectiveness of handling long-tailed data.", + "bbox": [ + 496, + 700, + 890, + 852 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "CBDM with DDIM sampling Apart from DDPM reverse sampling and neural SDE methods, recent deterministic ODE methods such as DDIM [43] are widely used in the", + "bbox": [ + 498, + 854, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "18440", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/75a37acadd05bc48854fe94ad9c7471bb083e68cbc4bf9dbf366ed162efb6aee.jpg", + "image_caption": [ + "Figure 5. FID/IS score under different regularization weight $\\tau$" + ], + "image_footnote": [], + "bbox": [ + 81, + 88, + 269, + 190 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/fe780cded2078a75ff6b1eceef772b462c1b30bc7180e3e0bcce931b98a322b5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 279, + 88, + 464, + 190 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/9069478d6c2608070b1710d63bfb23e68e15d906fe96c02ba312399978efd9cc.jpg", + "image_caption": [ + "Figure 6. FID/IS score under different guidance strength $\\omega$" + ], + "image_footnote": [], + "bbox": [ + 81, + 223, + 269, + 323 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/c1e02173e8bc13d89d5bce211f5e972fcd8a07c83970705729f9b7c8898e749e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 277, + 223, + 464, + 324 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "generation process of diffusion models. We also conduct experiments to study whether CBDM is compatible with DDIM steps. Table 6 shows the comparison of using DDPM and DDIM reverse steps when the generation steps are compressed to $1/10$ . We note on normal CIFAR100, FID and IS become slightly worse than using 1000 DDPM steps. On CIFAR100LT, except for IS, all metrics remain almost the same as using DDPM steps.", + "bbox": [ + 75, + 368, + 468, + 489 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effects of hyperparameters We tested the effects of regularization weight $\\tau$ in CBDM, which is set in default $\\tau = \\tau_0 = 0.001$ so that the weight does not surpass 1 at all steps. As shown in Figure 5, we search weights across different scales and found that the optimal weight is $\\tau = \\tau_0$ . The results also point out that $\\tau$ should not be too large or too small. We also tested the difference between CBDM models trained using commitment loss $\\mathcal{L}_{rc}$ and those trained without it. Our experiments show that the FID score of the models using the commitment loss (8.30) are significantly lower than those of the models not using it (8.84).", + "bbox": [ + 75, + 494, + 468, + 661 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Guidance strength $\\omega$ As we adapt the Classifier-Free Guidance (CFG) during sampling, we also show the effects of guidance strength $\\omega$ . In Figure 6, we searched $\\omega$ from 0 to 2 with an interval of 0.2 and compare FID and IS score of DDPM and CBDM models. We observe that the FID of CBDM remains decreasing at larger guidance strengths compared to DDPM; and although the IS of CBDM is consistently weaker than that of DDPM, they reach the same level when the guidance strength reaches 2.0, where the FID of CBDM (6.75) is still significantly lower than the best FID of DDPM (7.38).", + "bbox": [ + 75, + 667, + 467, + 833 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Fidelity-diversity control [12] describes the phenomenon where excessive guidance strength $\\omega$ tends to lead to better fidelity at the price of overfitting and diversity degradation, as a fidelity-diversity tradeoff. CBDM", + "bbox": [ + 75, + 839, + 468, + 901 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/3012b664f7380c1364ed1b785a1aa3ee14e747273724618fe884593f33f0e987.jpg", + "image_caption": [ + "Figure 7. Image fidelity and diversity controlled by guidance strength $\\omega$ and regularization weight $\\tau$ ." + ], + "image_footnote": [], + "bbox": [ + 516, + 95, + 875, + 251 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "additionally provides another hyperparameter: the regularization weight $\\tau$ that promotes model diversity through cross-class information interaction. Figure 7 shows the body class (53) generation results under different guidance strengths and different regularization weight. We observe that a DDPM model with high guidance strength is able to generate very realistic results, but its images overlap almost exactly with a sample in the training set. In contrast, the guidance term of CBDM does not drastically change the image content, but rather refines the image content to be closer to the selected class than the unguided result. This means that CBDM retains more of the diversity of the unguided case and effectively enhances the category information with guide terms.", + "bbox": [ + 496, + 309, + 893, + 521 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 534, + 617, + 550 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we focus on the problem when the training data is imbalanced and the diffusion model drops in generation quality on tail classes. We first establish a low baseline for this task by examining the most common paradigms in long-tail scenarios and in training generative models with limited data. Thereafter, we propose the Class-Balancing Diffusion Model (CBDM) through theoretical analysis. This approach can be implemented very cleanly in the training of any conditional diffusion model and thus has the potential to be widely used in other fields. Our experiments show that the CBDM approach significantly improves model generation diversity with high fidelity, on both class-balanced and class-imbalanced datasets.", + "bbox": [ + 496, + 560, + 890, + 755 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Acknowledgement", + "text_level": 1, + "bbox": [ + 500, + 768, + 679, + 787 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This work is supported by the National Key R&D Program of China (No. 2022ZD0160703, No. 2022ZD0160702), National Natural Science Foundation of China (62271308), STCSM (No. 22511106101, No. 18DZ2270700, No. 22511105700, No. 21DZ1100100), 111 plan (No. BP0719010), and State Key Laboratory of UHD Video and Audio Production and Presentation.", + "bbox": [ + 496, + 794, + 893, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "18441", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Jan Kautz Arash Vahdat, Karsten Kreis. Score-based generative modeling in latent space. In NeurIPS, 2021. 1", + "[2] Andrew Brock, Jeff Donahue, and Karen Simonyan. Large scale GAN training for high fidelity natural image synthesis. In ICLR, 2019. 1, 2", + "[3] Kaidi Cao, Colin Wei, Adrien Gaidon, Nikos Arechiga, and Tengyu Ma. Learning imbalanced datasets with label-distribution-aware margin loss. NeurIPS, 32, 2019. 5", + "[4] Xinlei Chen and Kaiming He. Exploring simple siamese representation learning. In CVPR, pages 15750-15758, 2021. 4", + "[5] Peng Chu, Xiao Bian, Shaopeng Liu, and Haibin Ling. Feature space augmentation for long-tailed data. In Andrea Vedaldi, Horst Bischof, Thomas Brox, and Jan-Michael Frahm, editors, ECCV, pages 694-710, Cham, 2020. Springer International Publishing. 2", + "[6] Ishan Deshpande, Ziyu Zhang, and Alexander G Schwing. Generative modeling using the sliced Wasserstein distance. In CVPR, pages 3483-3491, 2018. 2", + "[7] Prafulla Dhariwal and Alexander Quinn Nichol. Diffusion models beat GANs on image synthesis. In A. Beygelzimer, Y. Dauphin, P. Liang, and J. Wortman Vaughan, editors, NeurIPS, 2021. 1, 2, 5", + "[8] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020.5", + "[9] Didrik Nielsen Giorgio Giannone and Ole Winther. Few-shot diffusion models. arXiv preprint arXiv:2205.15463, 2021. 2", + "[10] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. In NeurIPS, page 6629-6640, Red Hook, NY, USA, 2017. Curran Associates Inc. 5", + "[11] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In NeurIPS, 2020. 3, 4, 6", + "[12] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598, 2022. 2, 3, 4, 8", + "[13] Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan, Mohammad Norouzi, and David J Fleet. Video diffusion models. arXiv preprint arXiv:2204.03458, 2022. 1", + "[14] Greg Holste, Song Wang, Ziyu Jiang, Thomas C. Shen, George L. Shih, Ronald M. Summers, Yifan Peng, and Zhangyang Wang. Long-tailed classification of thorax diseases on chest x-ray: A new benchmark study. Data augmentation, labelling, and imperfections : second MICCAI workshop, DALI 2022, held in conjunction with MICCAI 2022, Singapore, September 22, 2022, proceedings. DALI (Workshop), 13567:22-32, 2022. 1", + "[15] Grant Van Horn, Oisin Mac Aodha, Yang Song, Yin Cui, Chen Sun, Alexander Shepard, Hartwig Adam, Pietro Perona, and Serge J. Belongie. The inaturalist species classification and detection dataset. In CVPR, pages 8769-8778, 2018. 1" + ], + "bbox": [ + 78, + 114, + 467, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[16] Chen Huang, Yining Li, Chen Change Loy, and Xiaou Tang. Learning deep representation for imbalanced classification. In CVPR, pages 5375-5384, 2016. 2", + "[17] Myeonghun Jeong, Hyeongju Kim, Sung Jun Cheon, Byoung Jin Choi, and Nam Soo Kim. Diff-TTS: A Denoising Diffusion Model for Text-to-Speech. In Proc. Interspeech 2021, pages 3605-3609, 2021. 1", + "[18] Alexia Jolicoeur-Martineau, Remi Piche-Taillefer, Rémi Tachet des Combes, and Ioannis Mitliagkas. Adversarial score matching and improved sampling for image generation. arXiv preprint arXiv:2009.05475. 1", + "[19] Haeyong Kang, Thang Vu, and Chang D. Yoo. Learning imbalanced datasets with maximum margin loss. In ICIP, pages 1269-1273, 2021. 1", + "[20] Tero Karras, Miika Aittala, Timo Aila, and Samuli Laine. Elucidating the design space of diffusion-based generative models. arXiv preprint arXiv:2206.00364, 2022. 5, 7", + "[21] Tero Karras, Miika Aittala, Janne Hellsten, Samuli Laine, Jaakko Lehtinen, and Timo Aila. Training generative adversarial networks with limited data. In NeurIPS, Red Hook, NY, USA, 2020. Curran Associates Inc. 1, 2, 5, 6, 7", + "[22] Minsu Ko, Eunju Cha, Sungjoo Suh, Huijin Lee, Jae-Joon Han, Jinwoo Shin, and Bohyung Han. Self-supervised dense consistency regularization for image-to-image translation. In CVPR, pages 18280-18289, 2022. 2", + "[23] Tuomas Kynkänniemi, Tero Karras, Samuli Laine, Jaakko Lehtinen, and Timo Aila. Improved precision and recall metric for assessing generative models. NeurIPS, 32, 2019. 5", + "[24] Wen Li, Limin Wang, Wei Li, Eirikur Agustsson, and Luc Van Gool. Webvision database: Visual learning and understanding from web data. arXiv preprint arXiv:1708.02862, 2017. 1", + "[25] Jialun Liu, Yifan Sun, Chuchu Han, Zhaopeng Dou, and Wenhui Li. Deep representation learning on long-tailed data: A learnable embedding augmentation perspective. In CVPR, pages 2967-2976, 2020. 2", + "[26] Guanxiong Luo, Martin Heide, and Martin Uecker. MRI reconstruction via data driven markov chain with joint uncertainty estimation. CoRR, abs/2202.01479, 2022. 1", + "[27] Dhruv Mahajan, Ross Girshick, Vignesh Ramanathan, Kaiming He, Manohar Paluri, Yixuan Li, Ashwin Bharambe, and Laurens van der Maaten. Exploring the limits of weakly supervised pretraining. In Vittorio Ferrari, Martial Hebert, Cristian Sminchisescu, and Yair Weiss, editors, ECCV, pages 185-201, Cham, 2018. Springer International Publishing. 1, 2, 5, 6", + "[28] Chenlin Meng, Yang Song, Jiaming Song, Jiajun Wu, JunYan Zhu, and Stefano Ermon. Sdedit: Image synthesis and editing with stochastic differential equations. arXiv preprint arXiv:2108.01073, 2021. 1", + "[29] Aditya Krishna Menon, Sadeep Jayasumana, Ankit Singh Rawat, Himanshu Jain, Andreas Veit, and Sanjiv Kumar. Long-tail learning via logit adjustment. In ICLR. OpenReview.net, 2021. 1, 2", + "[30] Takeru Miyato, Toshiki Kataoka, Masanori Koyama, and Yuichi Yoshida. Spectral normalization for generative adversarial networks. In ICLR, 2018. 7" + ], + "bbox": [ + 501, + 92, + 890, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "18442", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[31] Alexander Quinn Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. GLIDE: Towards photorealistic image generation and editing with text-guided diffusion models. In Kamalika Chaudhuri, Stefanie Jegelka, Le Song, Csaba Szepesvari, Gang Niu, and Sivan Sabato, editors, International Conference on Machine Learning (ICML), volume 162 of Proceedings of Machine Learning Research, pages 16784-16804. PMLR, 17-23 Jul 2022. 1, 2", + "[32] Wanli Ouyang, Xiaogang Wang, Cong Zhang, and Xiaokang Yang. Factors in finetuning deep model for object detection with long-tail distribution. In CVPR, pages 864-873, 2016. 2", + "[33] Yuntian Gu Quanlin Wu, Hang Ye. Guided diffusion model for adversarial purification from random noise. arXiv preprint arXiv:2206.10875, 2022. 1, 2", + "[34] Harsh Rangwani, Naman Jaswani, Tejan Karmali, Varun Jampani, and R. Venkatesh Babu. Improving gans for long-tailed data through group spectral regularization. In Shai Avidan, Gabriel Brostow, Moustapha Cisse, Giovanni Maria Farinella, and Tal Hassner, editors, ECCV, pages 426-442, Cham, 2022. Springer Nature Switzerland. 1, 5, 7", + "[35] Harsh Rangwani, Konda Reddy Mopuri, and R. Venkatesh Babu. Class balancing gan with a classifier in the loop. In Cassio de Campos and Marloes H. Maathuis, editors, Proceedings of the Thirty-Seventh Conference on Uncertainty in Artificial Intelligence, volume 161 of Proceedings of Machine Learning Research, pages 1618-1627. PMLR, 27-30 Jul 2021. 5, 7", + "[36] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, pages 10684-10695, 2022. 2", + "[37] Kevin Roth, Aurelien Lucchi, Sebastian Nowozin, and Thomas Hofmann. Stabilizing training of generative adversarial networks through regularization. In NeurIPS, page 2015-2025, Red Hook, NY, USA, 2017. Curran Associates Inc. 1, 2", + "[38] Chitwan Sahara, William Chan, Huiwen Chang, Chris A. Lee, Jonathan Ho, Tim Salimans, David J. Fleet, and Mohammad Norouzi. Palette: Image-to-image diffusion models. ACM SIGGRAPH 2022 Conference Proceedings, 2022. 1", + "[39] Mehdi S. M. Sajjadi, Olivier Bachem, Mario Lucic, Olivier Bousquet, and Sylvain Gelly. Assessing generative models via precision and recall. In NeurIPS, page 5234-5243, Red Hook, NY, USA, 2018. Curran Associates Inc. 5", + "[40] Tim Salimans, Ian Goodfellow, Wojciech Zaremba, Vicki Cheung, Alec Radford, Xi Chen, and Xi Chen. Improved techniques for training gans. In D. Lee, M. Sugiyama, U. Luxburg, I. Guyon, and R. Garnett, editors, NeurIPS, volume 29. Curran Associates, Inc., 2016. 5", + "[41] Vikash Sehwag, Caner Hazirbas, Albert Gordo, First Ozgenel, and Cristian Canton. Generating high fidelity data from low-density regions using diffusion models. In CVPR, pages 11492-11501, 2022. 2", + "[42] Jascha Sohl-Dickstein, Eric A. Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised" + ], + "bbox": [ + 78, + 90, + 468, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "learning using nonequilibrium thermodynamics. arXiv preprint arXiv:1503.03585, 2015. 2", + "[43] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models, 2021. 2, 7", + "[44] Yang Song, Liyue Shen, Lei Xing, and Stefano Ermon. Solving inverse problems in medical imaging with score-based generative models. arXiv preprint arXiv:2111.08005, 2022. 1", + "[45] Hung-Yu Tseng, Lu Jiang, Ce Liu, Ming-Hsuan Yang, and Weilong Yang. Regularizing generative adversarial networks under limited data. In CVPR, pages 7917-7927, 2021. 2, 7", + "[46] Aaron van den Oord, Oriol Vinyals, and Koray Kavukcuoglu. Neural discrete representation learning. In NeurIPS, NIPS'17, page 6309-6318, Red Hook, NY, USA, 2017. Curran Associates Inc. 4", + "[47] Jianfeng Wang, Thomas Lukasiewicz, Xiaolin Hu, Jianfei Cai, and Zhenghua Xu. Rsg: A simple but effective module for learning imbalanced datasets. In CVPR, pages 3783-3792, 2021. 2", + "[48] Xiulong Yang, Sheng-Min Shih, Yinlin Fu, Xiaoting Zhao, and Shihao Ji. Your vit is secretly a hybrid discriminative-generative diffusion model. ArXiv, abs/2208.07791, 2022. 5, 7", + "[49] Xi Yin, Xiang Yu, Kihyuk Sohn, Xiaoming Liu, and Manmohan Chandraker. Feature transfer learning for face recognition with under-represented data. In CVPR, pages 5697-5706, 2019. 2", + "[50] Shengyu Zhao, Zhijian Liu, Ji Lin, Jun-Yan Zhu, and Song Han. Differentiable augmentation for data-efficient gan training. In NeurlPS, Red Hook, NY, USA, 2020. Curran Associates Inc. 1, 2, 5, 6" + ], + "bbox": [ + 501, + 92, + 890, + 530 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "18443", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2023/Class-Balancing Diffusion Models/eae79914-520a-4687-a1d4-1e04f53e44f8_model.json b/2023/Class-Balancing Diffusion Models/eae79914-520a-4687-a1d4-1e04f53e44f8_model.json new file mode 100644 index 0000000000000000000000000000000000000000..72be67479a431a9b580ce9a67e3a9557d8685314 --- /dev/null +++ b/2023/Class-Balancing Diffusion Models/eae79914-520a-4687-a1d4-1e04f53e44f8_model.json @@ -0,0 +1,2145 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.0, + 0.812, + 0.045 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.314, + 0.131, + 0.658, + 0.153 + ], + "angle": 0, + "content": "Class-Balancing Diffusion Models" + }, + { + "type": "text", + "bbox": [ + 0.095, + 0.189, + 0.2, + 0.209 + ], + "angle": 0, + "content": "Yiming Qin" + }, + { + "type": "text", + "bbox": [ + 0.235, + 0.189, + 0.374, + 0.209 + ], + "angle": 0, + "content": "Huangjie Zheng" + }, + { + "type": "text", + "bbox": [ + 0.41, + 0.189, + 0.564, + 0.209 + ], + "angle": 0, + "content": "Jiangchao Yao\\(^{1,3}\\)" + }, + { + "type": "text", + "bbox": [ + 0.597, + 0.189, + 0.735, + 0.209 + ], + "angle": 0, + "content": "Mingyuan Zhou" + }, + { + "type": "text", + "bbox": [ + 0.773, + 0.189, + 0.871, + 0.209 + ], + "angle": 0, + "content": "Ya Zhang\\(^{1,3}\\)" + }, + { + "type": "text", + "bbox": [ + 0.241, + 0.229, + 0.728, + 0.246 + ], + "angle": 0, + "content": "\\(^{1}\\)Cooperative Medianet Innovation Center, Shanghai Jiao-Tong University" + }, + { + "type": "text", + "bbox": [ + 0.392, + 0.246, + 0.58, + 0.26 + ], + "angle": 0, + "content": "\\(^{2}\\)University of Texas, Austin" + }, + { + "type": "text", + "bbox": [ + 0.401, + 0.26, + 0.568, + 0.276 + ], + "angle": 0, + "content": "\\(^{3}\\)Shanghai AI Laboratory" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.287, + 0.314, + 0.303 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.318, + 0.474, + 0.651 + ], + "angle": 0, + "content": "Diffusion-based models have shown the merits of generating high-quality visual data while preserving better diversity in recent studies. However, such observation is only justified with curated data distribution, where the data samples are nicely pre-processed to be uniformly distributed in terms of their labels. In practice, a long-tailed data distribution appears more common and how diffusion models perform on such class-imbalanced data remains unknown. In this work, we first investigate this problem and observe significant degradation in both diversity and fidelity when the diffusion model is trained on datasets with class-imbalanced distributions. Especially in tail classes, the generations largely lose diversity and we observe severe mode-collapse issues. To tackle this problem, we set from the hypothesis that the data distribution is not class-balanced, and propose Class-Balancing Diffusion Models (CBDM) that are trained with a distribution adjustment regularizer as a solution. Experiments show that images generated by CBDM exhibit higher diversity and quality in both quantitative and qualitative ways. Our method benchmarked the generation results on CIFAR100/CIFAR100LT dataset and shows outstanding performance on the downstream recognition task." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.68, + 0.21, + 0.696 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.704, + 0.47, + 0.84 + ], + "angle": 0, + "content": "In recent years, log-likelihood-based diffusion models have evolved rapidly and established new benchmarks on a range of generation tasks [1,7]. Based on them, researchers have been able to further control the model generation process and the generation quality. This improves the applications of generative models in numerous domains including text-image generation [31], image editing [28, 38], speech synthesis [17], medical imaging [26, 44], video generation [13] and adversarial learning [18, 33], etc." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.471, + 0.901 + ], + "angle": 0, + "content": "Although diffusion models are known for the power of high fidelity and diversity in generation, most of the existing diffusion models are trained with the hypothesis that the data are uniformly distributed w.r.t. their labels. However," + }, + { + "type": "image", + "bbox": [ + 0.513, + 0.294, + 0.885, + 0.485 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.505, + 0.892, + 0.548 + ], + "angle": 0, + "content": "Figure 1. Generation degrades along with class frequency. Semantics of generated images become less recognizable when class frequency decreases, while the FID score increases significantly." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.58, + 0.893, + 0.838 + ], + "angle": 0, + "content": "in the real world, the distribution is often very skewed. Especially for many domain-specific generation tasks such as medical images [14], fine-grained dataset for taxology [15] and data grabbed from the web [24], it is difficult to collect large amounts of data for each class equally, and the size of the training set for head and tail categories can differ by a factor of hundred or more. For such datasets, unconditional diffusion models tend to produce a significant portion of low-quality images. Conditional models, as shown in Figure 1, generate head class images with satisfying performance, while conversely the generated images on tail classes are very likely to show unrecognizable semantics. Concerning training generative models with limited data, there already exist several methods [21, 37, 50] based on GAN models [2]. However, quite few studies examine the impact of imbalanced class distribution [34] especially on diffusion models, which is practical yet under-explored." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.894, + 0.901 + ], + "angle": 0, + "content": "Our work first introduces diffusion models to imbalance generation tasks on several long-tailed datasets [19], and then build some straightforward baselines according to the common methods used in long-tailed recognition [27, 29]." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "18434" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.473, + 0.394 + ], + "angle": 0, + "content": "To overcome the potential degeneration induced by the skewed distribution, we propose a novel Class-Balancing Diffusion Model (CBDM). Theoretically, CBDM resorts to adjusting the conditional transfer probability during sampling in order to implicitly force generated images to have a balanced prior distribution during every sampling step. Technically, the adjusted transfer probability of CBDM results in an additional MSE-form loss for a conditional diffusion model, which functions as a regularizer. Intuitively, this loss augments the similarity of generated images conditioned on different classes, and turns out to be an effective approach to transfer common information from head classes to tail classes without hurting the model's expressiveness on head classes. CBDM can be implemented within several flines of codes, and its lighter version admits fine-tuning an existing conditional model. We conducted extensive experiments on CIFAR10/CIFAR100 and their corresponding long-tailed dataset to show the promise of CBDM over existing state-of-the-art methods. In a nutshell, the contributions of this work can be summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.406, + 0.468, + 0.452 + ], + "angle": 0, + "content": "- We identify the severe degeneration problem of diffusion models in long-tailed generation tasks and benchmark some straightforward baselines in this direction." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.462, + 0.469, + 0.523 + ], + "angle": 0, + "content": "- We propose a new perspective to handle the generation quality collapse on tail classes, and derive a novel Class-Balancing Diffusion Model, which is effective and lightweight as a regularizer to existing methods." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.532, + 0.469, + 0.623 + ], + "angle": 0, + "content": "- We validate that CBDM is capable of generating more diverse images with convincing fidelity, especially for datasets with large number of categories. In addition, CBDM is robust with accelerating algorithms such as DDIM [43], and can be transplanted to different conditional diffusion-based backbones easily." + }, + { + "type": "list", + "bbox": [ + 0.097, + 0.406, + 0.469, + 0.623 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.649, + 0.227, + 0.665 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.674, + 0.473, + 0.901 + ], + "angle": 0, + "content": "Diffusion models Diffusion models are recently proposed generative models [42] based on non-equilibrium thermodynamics. Conditional diffusion models [7] encode label information into the generation process and improve largely the generation performance. The guidance structure proposed in [7] makes it possible to control the generation process through an external module. Based on a similar intuition, researchers arrive to realize diverse functions, such as guided adversarial purification [33], few-shot generation [9] and so on [36, 41]. The drawback of classifier guidance (noted as CG) [7] lies in its requirement of training another auxiliary classifier. To address the issue, classifier-free guidance (noted as CFG) [12] proposed a mechanism that uses the generator itself to express the class guidance information. CFG is proved to be a resource efficient method" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.892, + 0.138 + ], + "angle": 0, + "content": "and achieves outstanding performance on large models [31]. Moreover, CFG only requires to add one line in training, which can be easily transplanted on different models." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.15, + 0.895, + 0.453 + ], + "angle": 0, + "content": "Long-tail recognition The problem of long-tailed distribution is a common dilemma in machine learning and have been widely explored in the area of discriminative models, i.e., long-tailed recognition. There mainly exist three paradigms in this domain, namely Class Re-balancing [27, 29], Information Augmentation [25,47,49] and Module Improvement [16, 32]. Among them, Class Re-balancing provides the best explainability, and its most common practice is re-sampling [27]. Thereafter, stemming from modifying the objective function from the global error rate to the class average one, [29] propose logit adjustment which has shown an impressive performance. Another effective method is based on Information Augmentation, which uses head class feature information to augment tail classes [5, 25, 47, 49]. However, discriminative models map data from higher to lower dimensions, while generative models map images from lower to higher dimensions. Thus, the mechanism of class rebalancing between them may be completely different and how to design the balancing method remains under-explored." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.466, + 0.895, + 0.678 + ], + "angle": 0, + "content": "Generative models based on limited data GAN [2, 6] is the most dominant model in the field of image synthesis in recent years. Given the requirement of large-scale data to train the generative models, a part of researchers focus on improving the performance of GAN models under the small datasets. To address the overfitting issue of the discriminator, a number of regularization methods [22, 37, 45] has been proposed. An alternative solution is Data Augmentation. As augmentation information is very prone to leak to the generator, researchers proposed improved augmentation strategies such as Differentiable Augmentation (DiffAug) [50] and Adaptive Augmentation (ADA) [21]. These augmentation ways can be transferred to diffusion models as baselines to help the generation of tail-class samples." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.697, + 0.593, + 0.713 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.724, + 0.892, + 0.801 + ], + "angle": 0, + "content": "In this section, we first give the basic notations following the classical DDPM model, and then introduce the class-imbalanced generation setting. In the third part, we introduce our CBDM algorithm and present its training details including implementation and hyper-parameter settings." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.815, + 0.63, + 0.832 + ], + "angle": 0, + "content": "3.1. Preliminary" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.892, + 0.903 + ], + "angle": 0, + "content": "Diffusion models leverage a pre-defined forward process in training, where a clean image distribution \\( q(\\pmb{x}_0) \\) can be corrupted to a noisy distribution \\( q(\\pmb{x}_t|\\pmb{x}_0) \\) at a specified timestep \\( t \\). Given a pre-defined variance schedule \\( \\{\\beta_t\\}_{1:T} \\)," + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18435" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.085, + 0.1, + 0.64, + 0.274 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.658, + 0.121, + 0.889, + 0.273 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.284, + 0.893, + 0.327 + ], + "angle": 0, + "content": "Figure 2. Algorithm (left) and generation (right) visualization. In the left figure, we show that an extra regularization loss \\( gL_{r} \\) proportional to the diffusion step \\( t \\) is added during training. This loss function pushes the sampling distribution (curves on the surface) to wider region while preventing it to be excessively distorted compared to the ground truth distribution (gradient color on the background)." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.353, + 0.424, + 0.369 + ], + "angle": 0, + "content": "the noisy distribution at any intermediate timestep is" + }, + { + "type": "equation", + "bbox": [ + 0.088, + 0.379, + 0.457, + 0.42 + ], + "angle": 0, + "content": "\\[\nq \\left(\\boldsymbol {x} _ {t} \\mid \\boldsymbol {x} _ {0}\\right) = \\mathcal {N} \\left(\\sqrt {\\bar {\\alpha} _ {t}} \\boldsymbol {x} _ {0}, \\left(1 - \\bar {\\alpha} _ {t}\\right) \\boldsymbol {I}\\right); \\quad \\bar {\\alpha} _ {t} = \\prod_ {i = 1} ^ {t} \\left(1 - \\beta_ {i}\\right).\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.43, + 0.47, + 0.476 + ], + "angle": 0, + "content": "To reverse such forward process, a generative model \\(\\theta\\) learns to estimate the analytical true posterior in order to recover \\(x_{t - 1}\\) from \\(x_{t}\\) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.079, + 0.487, + 0.467, + 0.51 + ], + "angle": 0, + "content": "\\[\n\\min _ {\\theta} D _ {\\mathrm {K L}} [ q (\\pmb {x} _ {t - 1} | \\pmb {x} _ {t}, \\pmb {x} _ {0}) | | p _ {\\theta} (\\pmb {x} _ {t - 1} | \\pmb {x} _ {t}) ]; \\forall t \\in \\{1, \\dots , T \\},\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.52, + 0.469, + 0.551 + ], + "angle": 0, + "content": "and such an objective can be reduced to a simple denoising estimation loss [11]:" + }, + { + "type": "equation", + "bbox": [ + 0.078, + 0.562, + 0.468, + 0.592 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {D D P M}} = \\mathbb {E} _ {t, \\boldsymbol {x} _ {0} \\sim q (\\boldsymbol {x} _ {0}), \\epsilon \\sim \\mathcal {N} (\\boldsymbol {0}, I)} \\left[ \\| \\boldsymbol {\\epsilon} - \\boldsymbol {\\epsilon} _ {\\theta} \\left(\\sqrt {\\bar {\\alpha}} \\boldsymbol {x} _ {0} + \\sqrt {1 - \\bar {\\alpha} _ {t}} \\boldsymbol {\\epsilon}, t\\right) \\| ^ {2} \\right] \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.593, + 0.469, + 0.699 + ], + "angle": 0, + "content": "For the case where label information is available, the model is trained to estimate the noise as above in both conditional cases \\(\\epsilon_{\\theta}(\\pmb{x}_t,y,t)\\) with data-label pairs \\((x_0,y)\\) and unconditional case \\(\\epsilon_{\\theta}(\\pmb{x}_t,t)\\). In the sampling, the label-guided model estimates the noise with a linear interpolation \\(\\hat{\\epsilon} = (1 + \\omega)\\pmb{\\epsilon}_{\\theta}(\\pmb{x}_t,y,t) - \\omega \\pmb{\\epsilon}_{\\theta}(\\pmb{x}_t,t)\\) to recover \\(\\pmb{x}_{t-1}\\), which is often referred as Classifier-Free Guidance (CFG) [12]." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.708, + 0.373, + 0.724 + ], + "angle": 0, + "content": "3.2. Class-Balancing Diffusion Models" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.731, + 0.469, + 0.821 + ], + "angle": 0, + "content": "Current diffusion models assume the data distribution to be uniform in class, and thus equally treat samples in the training stage. However, based on our observation, such training strategy leads to degradation in generation quality. Below, we provide an analysis that motivates our Class-Balancing Diffusion Models (CBDM)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.822, + 0.469, + 0.902 + ], + "angle": 0, + "content": "Suppose \\( q(\\pmb{x}, y) \\) is the data distribution that we need to match with the joint distribution \\( p_{\\theta}(\\pmb{x}, y) \\) predicted by a generative model. We analyze their difference from the density ratio \\( r = \\frac{q(\\pmb{x}, y)}{p_{\\theta}(\\pmb{x}, y)} = \\frac{q(\\pmb{x} \\mid y)}{p_{\\theta}(\\pmb{x} \\mid y)} \\cdot \\frac{q(y)}{p_{\\theta}(y)} \\). When the true label distribution \\( q(y) \\) is the same as the prior \\( p_{\\theta}(y) \\)," + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.353, + 0.892, + 0.537 + ], + "angle": 0, + "content": "which is usually assumed to be uniform, the density ratio \\( r \\) is reduced the conditional term to learn a conditional model \\( p_{\\theta}(\\boldsymbol{x} \\mid y) \\). However, when such a hypothesis is violated, for head classes, \\( \\frac{q(y)}{p_{\\theta}(y)} \\) would result in a larger weight that makes the model biased and hurt tail classes, and vice versa. Empirically, we observe that the generation degrades more on tail classes, as illustrated in Figure 1. Moreover, as shown in Figure 2a, compared to head classes, DDPM cannot well capture the tail-class data distribution and the mode is poorly covered during the sampling process. As a result, generations of tail classes often have poor quality and diversity, shown in Figure 2b." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.538, + 0.892, + 0.674 + ], + "angle": 0, + "content": "To tackle this issue, the most intuitive approach lies in adjusting the prior label distribution through a class balanced re-sampling. However, such abrupt adjustment easily leads to negative improvement in experiments. The step-by-step sampling nature of diffusion models provides another aspect to adjust this distribution more softly. In this spirit, we propose to calibrate the learning process through the conditional transfer probability \\( p_{\\theta}(\\pmb{x}_{t-1}|\\pmb{x}_t,y) \\) when there exists a gap between the class distribution and the prior." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.673, + 0.892, + 0.753 + ], + "angle": 0, + "content": "Let \\( p_{\\theta}^{\\star}\\left(\\boldsymbol{x}_{t-1} \\mid \\boldsymbol{x}_t, y\\right) \\) be the optimum trained in the case that \\( \\frac{q(y)}{p_{\\theta}(y)} \\) is correctly estimated, and \\( p_{\\theta}\\left(\\boldsymbol{x}_{t-1} \\mid \\boldsymbol{x}_t, y\\right) \\) be the one trained in a class-imbalanced case. The relation between such two generative distributions can be described as the following proposition." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.761, + 0.892, + 0.821 + ], + "angle": 0, + "content": "Proposition 1. When training a diffusion model parameterized with \\(\\theta\\) on a class-imbalanced dataset, its conditional reverse distribution \\(p_{\\theta}\\left(\\boldsymbol{x}_{t-1}|\\boldsymbol{x}_t,y\\right)\\) can be corrected with an adjustment schema:" + }, + { + "type": "equation", + "bbox": [ + 0.516, + 0.829, + 0.892, + 0.863 + ], + "angle": 0, + "content": "\\[\np _ {\\theta} ^ {\\star} \\left(\\boldsymbol {x} _ {t - 1} \\mid \\boldsymbol {x} _ {t}, y\\right) = p _ {\\theta} \\left(\\boldsymbol {x} _ {t - 1} \\mid \\boldsymbol {x} _ {t}, y\\right) \\frac {p _ {\\theta} \\left(\\boldsymbol {x} _ {t - 1}\\right)}{p _ {\\theta} ^ {\\star} \\left(\\boldsymbol {x} _ {t - 1}\\right)} \\frac {q ^ {\\star} \\left(\\boldsymbol {x} _ {t}\\right)}{q \\left(\\boldsymbol {x} _ {t}\\right)} \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.52, + 0.886, + 0.892, + 0.901 + ], + "angle": 0, + "content": "The proposition above shows that, when trained on a" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18436" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.251 + ], + "angle": 0, + "content": "class-imbalanced dataset, a diffusion model can still approach the true data distribution by applying a distribution adjustment schema \\(\\frac{p_{\\theta}(\\boldsymbol{x}_{t-1})}{p_{\\theta}^*(\\boldsymbol{x}_{t-1})}\\frac{q^*(\\boldsymbol{x}_t)}{q(\\boldsymbol{x}_t)}\\) at every step \\(t\\). However, approximating this schema is not feasible at every sampling step, so CBDM incorporates it into the training loss function to achieve an equivalent objective, and thus gets rid of the model-free part \\(\\frac{q^*(\\boldsymbol{x}_t)}{q(\\boldsymbol{x}_t)}\\). By further decomposing \\(p_{\\theta}(\\boldsymbol{x}_{t-1})\\) and \\(p_{\\theta}^*(\\boldsymbol{x}_{t-1})\\) to the expectation of the conditional probability \\(p_{\\theta}^*(\\boldsymbol{x}_{t-1}|\\boldsymbol{x}_{t:T},y)\\), we present a upper bound to approximate this probability in Proposition 2." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.259, + 0.472, + 0.307 + ], + "angle": 0, + "content": "Proposition 2. For the adjusted loss \\(\\mathcal{L}_{DM}^{\\star} = \\sum_{t=1}^{T} \\mathcal{L}_{t-1}^{\\star}\\), an upper-bound of the target training objective to calibrate at timestep \\(t\\) (i.e., \\(\\mathcal{L}_{t-1}^{\\star}\\)) can be derived as:" + }, + { + "type": "equation", + "bbox": [ + 0.077, + 0.319, + 0.47, + 0.436 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\sum_ {t \\geq 1} \\mathcal {L} _ {t - 1} ^ {\\star} = \\sum_ {t \\geq 1} D _ {\\mathrm {K L}} [ q (\\pmb {x} _ {t - 1} | \\pmb {x} _ {t}, \\pmb {x} _ {0}) | | p _ {\\theta} ^ {\\star} (\\pmb {x} _ {t - 1} | \\pmb {x} _ {t}, y) ] \\\\ \\leq \\sum_ {t \\geq 1} [ \\underbrace {D _ {\\mathrm {K L}} [ q (\\boldsymbol {x} _ {t - 1} | \\boldsymbol {x} _ {t} , \\boldsymbol {x} _ {0}) | | p _ {\\theta} (\\boldsymbol {x} _ {t - 1} | \\boldsymbol {x} _ {t} , y) ]} _ {\\text {D i f f u s i o n m o d e l l o s s} \\mathcal {L} _ {D M}} \\\\ + \\underbrace {t \\mathbb {E} _ {y ^ {\\prime} \\sim q _ {y} ^ {\\star}} [ D _ {\\mathrm {K L}} [ p _ {\\theta} (\\boldsymbol {x} _ {t - 1} | \\boldsymbol {x} _ {t}) | | p _ {\\theta} (\\boldsymbol {x} _ {t - 1} | \\boldsymbol {x} _ {t} , y ^ {\\prime}) ] ] ]} _ {D i s t r i b u t i o n a d j u s t m e n t l o s s \\mathcal {L} _ {r}}, \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.464, + 0.47, + 0.631 + ], + "angle": 0, + "content": "The upper bound in the above proposition can be considered as two parts. The first term \\(\\mathcal{L}_{\\mathrm{DM}}\\) corresponds to an ordinary DDPM loss [11] e.g., Eqn. (1), and the second loss \\(\\mathcal{L}_{\\mathrm{r}}\\) is used to adjust the distribution as a regularization term. Roughly speaking, \\(\\mathcal{L}_{\\mathrm{r}}\\) increases the similarity between the model's output and a random target class. Thus, it reduces the risk of overfitting on the head classes, and enlarges the generation diversity for tail class through knowledge obtained from other classes. When \\(q_{\\mathcal{Y}}^{\\star}\\) is less longtailed than the dataset, this loss also increases the probability for underrepresented tail samples to be chosen during training." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.639, + 0.262, + 0.657 + ], + "angle": 0, + "content": "3.3. Training algorithm" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.663, + 0.47, + 0.77 + ], + "angle": 0, + "content": "The detailed training algorithm of CBDM is presented in Alg. (1). In the algorithm, we reduce the distribution adjustment loss \\(\\mathcal{L}_r\\) as a square error loss with Monte-Carlo samples as indicated by Eqn.(3), where \\(\\mathcal{V}\\) is a set of samples that drawn following the distribution \\(q_{\\mathcal{V}}^{\\star}\\) and \\(y\\) denotes the image label. Note for CFG [12], there is a fixed probability (usually \\(10\\%\\)) to drop the condition, i.e., \\(y = \\text{None}\\)." + }, + { + "type": "equation", + "bbox": [ + 0.089, + 0.792, + 0.47, + 0.83 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {r}} \\left(\\boldsymbol {x} _ {t}, y, t\\right) = \\frac {1}{| \\mathcal {Y} |} \\sum_ {y ^ {\\prime} \\in \\mathcal {Y}} [ t \\| \\epsilon_ {\\theta} \\left(\\boldsymbol {x} _ {t}, y\\right) - \\epsilon_ {\\theta} \\left(\\boldsymbol {x} _ {t}, y ^ {\\prime}\\right) \\| ^ {2} ], \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.471, + 0.903 + ], + "angle": 0, + "content": "For the implementation in practice, CBDM can be plugged into any existing conditional diffusion models by adopting their model architecture and adjusting the training loss \\(\\mathcal{L}_{\\mathrm{DM}}\\) following lines 6 - 11. Specifically, the choice" + }, + { + "type": "code_caption", + "bbox": [ + 0.502, + 0.091, + 0.791, + 0.107 + ], + "angle": 0, + "content": "Algorithm 1 Training algorithm of CBDM." + }, + { + "type": "algorithm", + "bbox": [ + 0.507, + 0.111, + 0.892, + 0.319 + ], + "angle": 0, + "content": "1: for Every batch of size N do \n2: for \\((\\pmb{x}_0^{(i)}, y^{(i)})\\) in this batch do \n3: Sample \\(\\pmb{\\epsilon}^{(i)} \\sim \\mathcal{N}(\\pmb{0}, \\pmb{I}), t \\sim \\mathcal{U}(\\{0, 1, \\dots, T\\})\\) \n4: \\(x_{t}^{(i)} = \\sqrt{\\bar{\\alpha}_{t}} x_{0}^{(i)} + \\sqrt{1 - \\bar{\\alpha}_{t}} \\epsilon^{(i)}\\) \n5: Calculate \\(\\mathcal{L}_{\\mathrm{DM}} = \\| \\pmb{\\epsilon}^{(i)} - \\pmb{\\epsilon}_{\\theta}(\\pmb{x}_{t}^{(i)}, y^{(i)}) \\|^{2}\\) \n6: Sample \\(y'(i)\\) from \\(q_{\\gamma}^{\\star}\\) \n7: Calculate the first regularization term \n8: \\(\\mathcal{L}_{r} = t\\tau \\| \\pmb{\\epsilon}_{\\theta}(\\pmb{x}_{t}^{(i)}, y^{(i)}) - \\mathrm{sg}(\\pmb{\\epsilon}_{\\theta}(\\pmb{x}_{t}^{(i)}, y'^{(i)})) \\|^{2}\\) \n9: Calculate the regularization commitment term \n10: \\(\\mathcal{L}_{rc} = t\\tau \\| \\mathrm{sg}(\\pmb{\\epsilon}_{\\theta}(\\pmb{x}_{t}^{(i)}, y^{(i)})) - \\pmb{\\epsilon}_{\\theta}(\\pmb{x}_{t}^{(i)}, y'^{(i)}) \\|^{2}\\) \n11: Update with \\(\\mathcal{L}_{\\mathrm{CBDM}} = \\mathcal{L}_{\\mathrm{DM}} + \\mathcal{L}_{r} + \\gamma \\mathcal{L}_{rc}\\) \n12: end for \n13: end for" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.367, + 0.892, + 0.581 + ], + "angle": 0, + "content": "of the regularization weight \\(\\tau\\) affects the sharpness of the density ratio \\(\\frac{p_{\\theta}(\\boldsymbol{x}_t)}{p_{\\theta}^*(\\boldsymbol{x}_t)}\\). For the theoretical analysis, please refer to our proof of Prop. (2) in the Appendix. In addition, the choice of the sampling set \\(\\mathcal{V}\\) is another important perspective of CBDM, which depends on the target distribution we wish to adjust. Without loss of generality, we discuss two cases here. On the one hand, we can adjust the label distribution to a class-balanced label distribution, where we sample labels to construct \\(\\mathcal{V}^{bal}\\). On the other hand, if the data distribution is heavily long-tailed, we can also target the adjusted distribution to a relatively less class-imbalanced distribution for stabilized training. In our experiments, we show CBDM can work well for both cases in different mechanisms." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.593, + 0.892, + 0.7 + ], + "angle": 0, + "content": "Moreover, we observe that naively optimizing with this loss could make the model collapse to some trivial solutions, where the model outputs the same result regardless of the condition \\( y \\) and thus degenerates the conditional generation performance. Therefore, we follow previous works to apply a stop gradient operation [4, 46] to prevent this issue. The final loss of CBDM is" + }, + { + "type": "equation", + "bbox": [ + 0.511, + 0.73, + 0.892, + 0.81 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} _ {\\mathrm {C B D M}} ^ {\\tau , \\gamma , \\mathcal {Y}} (\\pmb {x} _ {t}, y, t, \\pmb {\\epsilon}) = \\| \\pmb {\\epsilon} _ {\\theta} (\\pmb {x} _ {t}, y) - \\pmb {\\epsilon} \\| ^ {2} \\\\ + \\frac {\\tau t}{| \\mathcal {Y} |} \\sum_ {y ^ {\\prime} \\in \\mathcal {Y}} \\left(\\left\\| \\boldsymbol {\\epsilon} _ {\\theta} \\left(\\boldsymbol {x} _ {t}, y\\right) - \\operatorname {s g} \\left(\\boldsymbol {\\epsilon} _ {\\theta} \\left(\\boldsymbol {x} _ {t}, y ^ {\\prime}\\right)\\right) \\right\\| ^ {2} \\right. \\\\ + \\gamma \\| \\mathrm {s g} \\big (\\pmb {\\epsilon} _ {\\theta} (\\pmb {x} _ {t}, y) \\big) - \\pmb {\\epsilon} _ {\\theta} (\\pmb {x} _ {t}, y ^ {\\prime}) \\| ^ {2}), (4) \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.892, + 0.9 + ], + "angle": 0, + "content": "where \\(\\mathrm{sg}(\\cdot)\\) denotes the stop gradient operation; \\(\\tau, \\gamma\\) are weights for the regularization and the commitment term respectively with \\(\\gamma\\) set to \\(\\frac{1}{4}\\) in a default setting, and \\(|\\mathcal{V}|\\) denotes the number of elements in the label set." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18437" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.084, + 0.091, + 0.268, + 0.234 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.108, + 0.238, + 0.243, + 0.249 + ], + "angle": 0, + "content": "(a) DDPM / Tail Class 94" + }, + { + "type": "image", + "bbox": [ + 0.279, + 0.091, + 0.465, + 0.234 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.305, + 0.238, + 0.441, + 0.249 + ], + "angle": 0, + "content": "(b) CBDM / Tail Class 94" + }, + { + "type": "image", + "bbox": [ + 0.084, + 0.253, + 0.268, + 0.396 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.104, + 0.4, + 0.248, + 0.412 + ], + "angle": 0, + "content": "(c) DDPM / Body Class 62" + }, + { + "type": "image", + "bbox": [ + 0.277, + 0.253, + 0.461, + 0.396 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.296, + 0.4, + 0.442, + 0.412 + ], + "angle": 0, + "content": "(d) CBDM / Body Class 62" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.424, + 0.47, + 0.466 + ], + "angle": 0, + "content": "Figure 3. Comparison of image generation on heavily tail-distributed (94) and mild tail-distributed (62) classes between DDPM and CBDM." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.497, + 0.276, + 0.514 + ], + "angle": 0, + "content": "4. Experimental results" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.524, + 0.265, + 0.541 + ], + "angle": 0, + "content": "4.1. Experimental setup" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.55, + 0.47, + 0.687 + ], + "angle": 0, + "content": "Datasets We first chose two common datasets that are widely used in the domain of image synthesis, CIFAR10/CIFAR100, and their corresponding long-tailed versions CIFAR10LT and CIFAR100LT. The construction of CIFAR10LT and CIFAR100LT follows [3], where the size decreases exponentially with its category index according to the imbalance factor \\(imb = 0.01\\). We also conduct experiments on 3 higher-resolution datasets, whose details and results could be found in the Appendix." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.704, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Implementation details We strictly follow the training configurations of the baseline models. For DDPM, we set diffusion schedule \\(\\beta_{1} = 10^{-4}\\) and \\(\\beta_{T} = 0.02\\) with \\(T = 1,000\\), and optimize the network with an Adam optimizer whose learning rate is 0.0002 after 5,000 epochs of warmup. Considering that the size and semantic complexity of the datasets vary greatly, we choose appropriate epochs for each dataset. In the ablation studies, we also include EDM [20] and Diffusion-ViT [48] as our backbones to show the compatibility with different architectures and score-matching methods. Note that, when a model does not use conditional input, we follow [7] to slightly modify the backbone [8] by adding an extra embedding layer." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.091, + 0.892, + 0.333 + ], + "angle": 0, + "content": "Baselines We first adopt following classic methods to build baselines: re-sampling (RS), soft re-sampling methods (RS-SQRT) [27], and augmentation-based methods including Differentiable Augmentation (DiffAug) [50] and Adaptive Augmentation (ADA) [21]. Besides diffusion-based baselines, previous state-of-the-art generative models that study the long-tailed distribution, e.g., CBGAN [35], and group spectral regularization for GANs [34], are also included in the comparison. Precisely, RS follows the uniform class distribution and RS-SQRT [27] samples with a probability determined by the square root of class frequency. And DiffAug [50] is directly applied to all training images following their default threshold. Following [20], ADA [21] is not only applied to images, the augmentation pipeline is also encoded as a condition via an additional embedding layer in the U-Net." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.338, + 0.892, + 0.596 + ], + "angle": 0, + "content": "Metrics CBDM and corresponding baselines are evaluated in terms of both generation diversity and fidelity via Frechet Inception Distance (FID) [10], Inception Score (IS) [40], Recall [23] and \\( F_{\\beta} \\) [39]. We measure Recall and \\( F_{\\beta} \\) using Inception-V3 features, and take respectively \\( K = 5 \\) for Recall, 1/8 and 8 for the threshold in \\( F_{\\beta} \\), and 20 times of class number as the clustering number of \\( F_{\\beta} \\) to capture the inner class variance. To this end, Recall and \\( F_{8} \\) can be regarded as diversity metrics and IS and \\( F_{1/8} \\) are more inclined to measure fidelity. In evaluation, we take the class-balanced version corresponding to those used in training, and all metrics are measured with 50k generated images (10k for ablation experiments). The classifier-free guidance is adapted in sampling and we tune the guidance strength \\( \\omega \\) for both baselines and our method to ensure the best performance. Specifically, we take \\( \\omega = 1.6, 0.8, 1.0, 0.8 \\) respectively for the four CIFAR datasets mentioned above." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.605, + 0.634, + 0.62 + ], + "angle": 0, + "content": "4.2. Main results" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.629, + 0.892, + 0.825 + ], + "angle": 0, + "content": "In this part, we consider DDPM as the most direct baseline. As shown in Table 1, CBDM overall outperforms DDPM on all datasets except the IS is slightly lower on CIFAR10. For datasets with more classes, the improvements are more significant in terms of both diversity and fidelity, as on CIFAR100LT and CIFAR100. We further investigate the performance of diffusion models when combined with augmentation-based and re-sampling methods. Surprisingly, except for ADA, we consistently observe a performance degradation when these methods are compared with the vanilla DDPM. Moreover, adding the ADA augmentation with CBDM training leads to a further improvement in terms of both diversity and fidelity." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.826, + 0.892, + 0.901 + ], + "angle": 0, + "content": "As a qualitative justification, Figure 3 provides the visualization comparison between DDPM and CBDM on a relatively mild tail-distributed class (62) and a tail-distributed class (94). We remark that CBDM generates more diverse images. For example, on the tail class 94, the cabinets" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18438" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.114, + 0.089, + 0.859, + 0.338 + ], + "angle": 0, + "content": "
DatasetModelFID↓F8↑Recall↑IS↑F1/8↑
CIFAR100LTDDPM [11]7.380.850.5213.110.88
+ADA [21]6.160.910.5712.710.90
+DiffAug [50]9.190.880.4711.560.86
+RS [27]10.500.650.4912.600.83
+SQRT-RS [27]9.720.660.4713.470.83
CBDM (ours)6.26 (-1.12)0.91 (+0.06)0.57 (+0.05)13.24 (+0.13)0.89 (+0.01)
+ADA [21]5.81 (-1.57)0.91 (+0.06)0.57 (+0.05)13.34 (+0.23)0.90 (+0.02)
CIFAR100DDPM3.110.970.6513.650.96
CBDM (ours)2.72 (-0.39)0.97 (±0)0.67 (+0.02)14.03 (+0.38)0.96 (±0)
CIFAR10LTDDPM5.760.970.579.170.95
CBDM (ours)5.46 (-0.30)0.97 (±0)0.59 (+0.02)9.28 (+0.11)0.95 (±0)
CIFAR10DDPM3.160.990.649.800.98
CBDM (ours)3.03 (-0.13)0.99 (±0)0.65 (+0.01)9.63 (-0.17)0.98 (±0)
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.348, + 0.895, + 0.391 + ], + "angle": 0, + "content": "Table 1. CBDM performance on different datasets. In the table, the first three columns are diversity-related, and we mark the best results in bold. As the comparison between CBDM and DDPM is more straightforward, we mark the performance gain in parentheses next to the CBDM results, using blue and red to indicate improvements and degradation respectively." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.414, + 0.472, + 0.577 + ], + "angle": 0, + "content": "
PTFTyFIDF8RecallIS
CIFAR 100LT-ytrain6.260.910.5713.24
ybal6.100.940.6412.29
ysqrt6.000.930.5913.00
CIFAR 100LTCIFAR 100LT-7.380.850.5213.11
ytrain6.200.880.5413.36
ybal5.850.920.6013.29
CIFAR 100CIFAR 10LT-5.390.660.589.43
ytrain5.050.670.609.48
ybal5.900.660.609.49
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.587, + 0.47, + 0.687 + ], + "angle": 0, + "content": "Table 2. CBDM performance for 3 mechanisms under different regularization sampling set \\(\\mathcal{V}\\). Column \\(PT\\) indicates the dataset used in (pre)training, column \\(FT\\) indicates the dataset used for fine-tuning. The rows where \\(\\mathcal{V}\\) marked as “-” represent the results with DDPM, and rows where \\(FT\\) marked as “-” represent the results of CBDM trained from scratch. Three diversity-preferred metrics and the best results are in bold for emphasis." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.717, + 0.47, + 0.779 + ], + "angle": 0, + "content": "shown in Figure 3b have more color and texture, which justifies the improvements of diversity-related metrics in Table 1. On the contrary, DDPM only generates images that are highly similar to the training data." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.795, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Case-by-case study. To better understand the generation conditioned on each class, we compare the FID case-by-case between DDPM and CBDM on each class. The results are shown in Figure 4. Compared to DDPM, CBDM shows more consistent improvements for the class index greater than 40. In the tail classes, CBDM outperforms than DDPM the FID by a more significant margin." + }, + { + "type": "image", + "bbox": [ + 0.522, + 0.418, + 0.871, + 0.56 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.572, + 0.892, + 0.603 + ], + "angle": 0, + "content": "Figure 4. FID improvement per class compared to DDPM. The curve is smoothened by a moving average of 5 unit." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.628, + 0.893, + 0.869 + ], + "angle": 0, + "content": "The choice of label set \\(\\mathcal{V}\\) We investigate the effects of label distribution for CBDM under three mechanisms. The first mechanism is the vanilla CBDM given in Algorithm 1 which trains a diffusion model with regularization from scratch. Besides, we also study CBDM with two fine-tuning configurations. The first setting concerns using CBDM to adjust an existing model pre-trained on a long-tailed dataset to improve its performance. The second one refers to fine-tuning a pre-trained model to adapt to a smaller dataset with an imbalanced class distribution using CBDM. For those mechanisms, we consider three settings for sampling the label set: (1) a label distribution similar to the training set (denoted \\(\\mathcal{V}^{train}\\)); (2) a totally uniform label distribution (denoted \\(\\mathcal{V}^{bal}\\)); (3) a less long-tailed distribution compared to \\(\\mathcal{V}^{train}\\), whose class frequency is the square root of the original one (denoted \\(\\mathcal{V}^{sqrt}\\))." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.871, + 0.894, + 0.902 + ], + "angle": 0, + "content": "The results are shown in Table 2, where we found that CBDM performs well on all these configurations. As we" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "18439" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.082, + 0.089, + 0.473, + 0.278 + ], + "angle": 0, + "content": "
DatasetModelFIDIS
CIFAR 100BigGAN+DA+RLC[45](2021)2.99-
DDPM3.1113.65
CBDM (ours)2.7214.03
CIFAR 100LTCBGAN [35](2021)28.17-
DDPM12.3112.69
CBDM (ours)+ADA [21]10.5012.81
CIFAR 10LTCBGAN [35](2021)32.93-
SNGAN [30]+gSR [34](2022)18.587.80
DDPM9.689.00
CBDM (ours)9.389.12
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.287, + 0.472, + 0.345 + ], + "angle": 0, + "content": "Table 3. Comparison with long-tailed SoTAs on CIFAR. Following their setting, all methods are evaluated with \\(10\\mathrm{k}\\) generated images and ground truth images are from their validation dataset. The publish year of every baselines are marked in next to the citation." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.359, + 0.454, + 0.452 + ], + "angle": 0, + "content": "
Training dataPrecisionRecall
CIFAR1000.700.67
CIFAR100LT + DDPMgens (50k)0.450.39
0.48 (+0.03)0.44 (+0.05)
+ CBDMgens (50k)0.49 (+0.04)0.47 (+0.08)
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.462, + 0.47, + 0.504 + ], + "angle": 0, + "content": "Table 4. Recognition results of different training data. All configurations are evaluated on the testing set of normal CIFAR100. Gains based on CIFAR100LT dataset is noted in blue." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.532, + 0.473, + 0.76 + ], + "angle": 0, + "content": "analyzed in section 3.3, when the empirical distribution of \\(\\mathcal{V}\\) has a significant difference from the label distribution of the dataset, adjusting with \\(\\mathcal{V}^{bal}\\) could hurt training stability and result in performance degradation. We observe that \\(\\mathcal{V}^{bal}\\) only shows better performance regarding generation diversity, while the generation fidelity has an obvious gap than the other settings. Using a relatively mild set \\(\\mathcal{V}^{sqrt}\\) has better performance in terms of IS and preserves the generation diversity, which leads to the best FID among all these settings. On the contrary, when a pre-trained model is available, fine-tuning on top of it ensures better stability, and we can observe in such case using \\(\\mathcal{V}^{bal}\\) has significant improvement than the vanilla DDPM or fine-tuning using \\(\\mathcal{V}^{train}\\). When adapting to a different dataset, we also observe using \\(\\mathcal{V}^{bal}\\) causes stability issue and produce undesired results." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.472, + 0.903 + ], + "angle": 0, + "content": "Enhancement of training classifiers on long-tailed data As training models on long-tailed data usually leads to undesired classification results, we investigate whether the generated data could help improve the classifier trained on long-tailed data as a complementary evaluation metrics. Here, we train ResNet-32 models, respectively with CIFAR100, CIFAR100LT and CIFAR100LT augmented with DDPM and CBDM generations (50k samples). Table 4 shows that the training on long-tailed data results in se" + }, + { + "type": "table", + "bbox": [ + 0.502, + 0.089, + 0.903, + 0.188 + ], + "angle": 0, + "content": "
BackboneFID\\( F_8 \\)RecallIS\\( F_{1/8} \\)
EDM [20]8.640.820.4812.160.86
+CBDM (ours)7.970.880.5212.010.86
Diffuse-ViT [48]37.60.760.457.660.62
+CBDM (ours)30.00.810.518.000.65
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.198, + 0.892, + 0.227 + ], + "angle": 0, + "content": "Table 5. CBDM performance using different backbones on CIFAR100LT dataset. Three diversity-biased metrics." + }, + { + "type": "table", + "bbox": [ + 0.505, + 0.239, + 0.89, + 0.318 + ], + "angle": 0, + "content": "
DatasetFID↓F8↑Recall↑IS↑F1/8↑
CIFAR3.440.970.6913.110.96
100(+0.72)(±0)(+0.02)(-0.92)(±0)
CIFAR6.270.930.6112.560.89
100LT(+0.01)(+0.02)(+0.04)(-0.68)(±0)
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.327, + 0.892, + 0.37 + ], + "angle": 0, + "content": "Table 6. Sampling with 100 DDIM steps on CIFAR100 and CIFAR100LT. The difference compared with those using DDPM steps are reported below each result." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.395, + 0.892, + 0.531 + ], + "angle": 0, + "content": "vere performance degradation. When augmenting the long-tailed data with the generated data, we observe fairly improvements in the trained classifier. Moreover, comparing the results between DDPM and CBDM, we observe that the improvements by using CBDM are more significant, which validates the effectiveness of our approach. Especially, the gain on recall is more remarkable than the gain on precision, which means that the generated images are more diverse than DDPM and justifies the results in Table 1." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.534, + 0.893, + 0.671 + ], + "angle": 0, + "content": "Comparison with other benchmarks We include the representative state-of-the-art long-tailed generative modeling approaches in the comparison, most of which are based on GANs since they suffer more severe problems when the training data is skewed. For a fair comparison, we strictly follow their evaluation setting and the results are reported in Table 3. From the results, DDPM surpasses the performance of the baselines on all the datasets, and CBDM exhibits even stronger performance." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.679, + 0.611, + 0.693 + ], + "angle": 0, + "content": "4.3. Ablations" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.702, + 0.892, + 0.853 + ], + "angle": 0, + "content": "Compatibility with different backbones We first investigate the performance of our method with different diffusion models. Without loss of generality, we adopt EDM [20] and Diffusion-ViT [48], which use a different score-matching method and different denoising network backbone. As shown in Table 5, compared with the corresponding baselines, CBDM improves the generation result on CIFAR100LT dataset for both backbones, demonstrating the compatibility with a variety of backbones and the effectiveness of handling long-tailed data." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.856, + 0.892, + 0.901 + ], + "angle": 0, + "content": "CBDM with DDIM sampling Apart from DDPM reverse sampling and neural SDE methods, recent deterministic ODE methods such as DDIM [43] are widely used in the" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "18440" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.089, + 0.271, + 0.191 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.28, + 0.089, + 0.465, + 0.191 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.088, + 0.195, + 0.458, + 0.21 + ], + "angle": 0, + "content": "Figure 5. FID/IS score under different regularization weight \\(\\tau\\)" + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.224, + 0.27, + 0.324 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.279, + 0.224, + 0.465, + 0.325 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.098, + 0.327, + 0.449, + 0.341 + ], + "angle": 0, + "content": "Figure 6. FID/IS score under different guidance strength \\(\\omega\\)" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.369, + 0.47, + 0.49 + ], + "angle": 0, + "content": "generation process of diffusion models. We also conduct experiments to study whether CBDM is compatible with DDIM steps. Table 6 shows the comparison of using DDPM and DDIM reverse steps when the generation steps are compressed to \\(1/10\\). We note on normal CIFAR100, FID and IS become slightly worse than using 1000 DDPM steps. On CIFAR100LT, except for IS, all metrics remain almost the same as using DDPM steps." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.496, + 0.47, + 0.662 + ], + "angle": 0, + "content": "Effects of hyperparameters We tested the effects of regularization weight \\(\\tau\\) in CBDM, which is set in default \\(\\tau = \\tau_0 = 0.001\\) so that the weight does not surpass 1 at all steps. As shown in Figure 5, we search weights across different scales and found that the optimal weight is \\(\\tau = \\tau_0\\). The results also point out that \\(\\tau\\) should not be too large or too small. We also tested the difference between CBDM models trained using commitment loss \\(\\mathcal{L}_{rc}\\) and those trained without it. Our experiments show that the FID score of the models using the commitment loss (8.30) are significantly lower than those of the models not using it (8.84)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.668, + 0.468, + 0.834 + ], + "angle": 0, + "content": "Guidance strength \\(\\omega\\) As we adapt the Classifier-Free Guidance (CFG) during sampling, we also show the effects of guidance strength \\(\\omega\\). In Figure 6, we searched \\(\\omega\\) from 0 to 2 with an interval of 0.2 and compare FID and IS score of DDPM and CBDM models. We observe that the FID of CBDM remains decreasing at larger guidance strengths compared to DDPM; and although the IS of CBDM is consistently weaker than that of DDPM, they reach the same level when the guidance strength reaches 2.0, where the FID of CBDM (6.75) is still significantly lower than the best FID of DDPM (7.38)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.47, + 0.902 + ], + "angle": 0, + "content": "Fidelity-diversity control [12] describes the phenomenon where excessive guidance strength \\(\\omega\\) tends to lead to better fidelity at the price of overfitting and diversity degradation, as a fidelity-diversity tradeoff. CBDM" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.096, + 0.877, + 0.252 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.256, + 0.892, + 0.286 + ], + "angle": 0, + "content": "Figure 7. Image fidelity and diversity controlled by guidance strength \\(\\omega\\) and regularization weight \\(\\tau\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.31, + 0.895, + 0.522 + ], + "angle": 0, + "content": "additionally provides another hyperparameter: the regularization weight \\(\\tau\\) that promotes model diversity through cross-class information interaction. Figure 7 shows the body class (53) generation results under different guidance strengths and different regularization weight. We observe that a DDPM model with high guidance strength is able to generate very realistic results, but its images overlap almost exactly with a sample in the training set. In contrast, the guidance term of CBDM does not drastically change the image content, but rather refines the image content to be closer to the selected class than the unguided result. This means that CBDM retains more of the diversity of the unguided case and effectively enhances the category information with guide terms." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.535, + 0.619, + 0.551 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.561, + 0.892, + 0.756 + ], + "angle": 0, + "content": "In this paper, we focus on the problem when the training data is imbalanced and the diffusion model drops in generation quality on tail classes. We first establish a low baseline for this task by examining the most common paradigms in long-tail scenarios and in training generative models with limited data. Thereafter, we propose the Class-Balancing Diffusion Model (CBDM) through theoretical analysis. This approach can be implemented very cleanly in the training of any conditional diffusion model and thus has the potential to be widely used in other fields. Our experiments show that the CBDM approach significantly improves model generation diversity with high fidelity, on both class-balanced and class-imbalanced datasets." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.77, + 0.681, + 0.788 + ], + "angle": 0, + "content": "6. Acknowledgement" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.795, + 0.894, + 0.901 + ], + "angle": 0, + "content": "This work is supported by the National Key R&D Program of China (No. 2022ZD0160703, No. 2022ZD0160702), National Natural Science Foundation of China (62271308), STCSM (No. 22511106101, No. 18DZ2270700, No. 22511105700, No. 21DZ1100100), 111 plan (No. BP0719010), and State Key Laboratory of UHD Video and Audio Production and Presentation." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "18441" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.116, + 0.468, + 0.143 + ], + "angle": 0, + "content": "[1] Jan Kautz Arash Vahdat, Karsten Kreis. Score-based generative modeling in latent space. In NeurIPS, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.146, + 0.468, + 0.186 + ], + "angle": 0, + "content": "[2] Andrew Brock, Jeff Donahue, and Karen Simonyan. Large scale GAN training for high fidelity natural image synthesis. In ICLR, 2019. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.188, + 0.468, + 0.23 + ], + "angle": 0, + "content": "[3] Kaidi Cao, Colin Wei, Adrien Gaidon, Nikos Arechiga, and Tengyu Ma. Learning imbalanced datasets with label-distribution-aware margin loss. NeurIPS, 32, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.232, + 0.468, + 0.26 + ], + "angle": 0, + "content": "[4] Xinlei Chen and Kaiming He. Exploring simple siamese representation learning. In CVPR, pages 15750-15758, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.261, + 0.468, + 0.33 + ], + "angle": 0, + "content": "[5] Peng Chu, Xiao Bian, Shaopeng Liu, and Haibin Ling. Feature space augmentation for long-tailed data. In Andrea Vedaldi, Horst Bischof, Thomas Brox, and Jan-Michael Frahm, editors, ECCV, pages 694-710, Cham, 2020. Springer International Publishing. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.332, + 0.468, + 0.374 + ], + "angle": 0, + "content": "[6] Ishan Deshpande, Ziyu Zhang, and Alexander G Schwing. Generative modeling using the sliced Wasserstein distance. In CVPR, pages 3483-3491, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.375, + 0.468, + 0.43 + ], + "angle": 0, + "content": "[7] Prafulla Dhariwal and Alexander Quinn Nichol. Diffusion models beat GANs on image synthesis. In A. Beygelzimer, Y. Dauphin, P. Liang, and J. Wortman Vaughan, editors, NeurIPS, 2021. 1, 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.432, + 0.468, + 0.514 + ], + "angle": 0, + "content": "[8] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020.5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.517, + 0.468, + 0.544 + ], + "angle": 0, + "content": "[9] Didrik Nielsen Giorgio Giannone and Ole Winther. Few-shot diffusion models. arXiv preprint arXiv:2205.15463, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.547, + 0.468, + 0.614 + ], + "angle": 0, + "content": "[10] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. In NeurIPS, page 6629-6640, Red Hook, NY, USA, 2017. Curran Associates Inc. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.617, + 0.468, + 0.644 + ], + "angle": 0, + "content": "[11] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In NeurIPS, 2020. 3, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.647, + 0.468, + 0.674 + ], + "angle": 0, + "content": "[12] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598, 2022. 2, 3, 4, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.676, + 0.468, + 0.717 + ], + "angle": 0, + "content": "[13] Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan, Mohammad Norouzi, and David J Fleet. Video diffusion models. arXiv preprint arXiv:2204.03458, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.719, + 0.468, + 0.829 + ], + "angle": 0, + "content": "[14] Greg Holste, Song Wang, Ziyu Jiang, Thomas C. Shen, George L. Shih, Ronald M. Summers, Yifan Peng, and Zhangyang Wang. Long-tailed classification of thorax diseases on chest x-ray: A new benchmark study. Data augmentation, labelling, and imperfections : second MICCAI workshop, DALI 2022, held in conjunction with MICCAI 2022, Singapore, September 22, 2022, proceedings. DALI (Workshop), 13567:22-32, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.832, + 0.468, + 0.899 + ], + "angle": 0, + "content": "[15] Grant Van Horn, Oisin Mac Aodha, Yang Song, Yin Cui, Chen Sun, Alexander Shepard, Hartwig Adam, Pietro Perona, and Serge J. Belongie. The inaturalist species classification and detection dataset. In CVPR, pages 8769-8778, 2018. 1" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.116, + 0.468, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.134 + ], + "angle": 0, + "content": "[16] Chen Huang, Yining Li, Chen Change Loy, and Xiaou Tang. Learning deep representation for imbalanced classification. In CVPR, pages 5375-5384, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.136, + 0.892, + 0.19 + ], + "angle": 0, + "content": "[17] Myeonghun Jeong, Hyeongju Kim, Sung Jun Cheon, Byoung Jin Choi, and Nam Soo Kim. Diff-TTS: A Denoising Diffusion Model for Text-to-Speech. In Proc. Interspeech 2021, pages 3605-3609, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.193, + 0.892, + 0.247 + ], + "angle": 0, + "content": "[18] Alexia Jolicoeur-Martineau, Remi Piche-Taillefer, Rémi Tachet des Combes, and Ioannis Mitliagkas. Adversarial score matching and improved sampling for image generation. arXiv preprint arXiv:2009.05475. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.249, + 0.892, + 0.29 + ], + "angle": 0, + "content": "[19] Haeyong Kang, Thang Vu, and Chang D. Yoo. Learning imbalanced datasets with maximum margin loss. In ICIP, pages 1269-1273, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.292, + 0.892, + 0.333 + ], + "angle": 0, + "content": "[20] Tero Karras, Miika Aittala, Timo Aila, and Samuli Laine. Elucidating the design space of diffusion-based generative models. arXiv preprint arXiv:2206.00364, 2022. 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.335, + 0.892, + 0.389 + ], + "angle": 0, + "content": "[21] Tero Karras, Miika Aittala, Janne Hellsten, Samuli Laine, Jaakko Lehtinen, and Timo Aila. Training generative adversarial networks with limited data. In NeurIPS, Red Hook, NY, USA, 2020. Curran Associates Inc. 1, 2, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.391, + 0.892, + 0.447 + ], + "angle": 0, + "content": "[22] Minsu Ko, Eunju Cha, Sungjoo Suh, Huijin Lee, Jae-Joon Han, Jinwoo Shin, and Bohyung Han. Self-supervised dense consistency regularization for image-to-image translation. In CVPR, pages 18280-18289, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.449, + 0.892, + 0.49 + ], + "angle": 0, + "content": "[23] Tuomas Kynkänniemi, Tero Karras, Samuli Laine, Jaakko Lehtinen, and Timo Aila. Improved precision and recall metric for assessing generative models. NeurIPS, 32, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.491, + 0.892, + 0.544 + ], + "angle": 0, + "content": "[24] Wen Li, Limin Wang, Wei Li, Eirikur Agustsson, and Luc Van Gool. Webvision database: Visual learning and understanding from web data. arXiv preprint arXiv:1708.02862, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.547, + 0.892, + 0.604 + ], + "angle": 0, + "content": "[25] Jialun Liu, Yifan Sun, Chuchu Han, Zhaopeng Dou, and Wenhui Li. Deep representation learning on long-tailed data: A learnable embedding augmentation perspective. In CVPR, pages 2967-2976, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.605, + 0.892, + 0.646 + ], + "angle": 0, + "content": "[26] Guanxiong Luo, Martin Heide, and Martin Uecker. MRI reconstruction via data driven markov chain with joint uncertainty estimation. CoRR, abs/2202.01479, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.647, + 0.892, + 0.743 + ], + "angle": 0, + "content": "[27] Dhruv Mahajan, Ross Girshick, Vignesh Ramanathan, Kaiming He, Manohar Paluri, Yixuan Li, Ashwin Bharambe, and Laurens van der Maaten. Exploring the limits of weakly supervised pretraining. In Vittorio Ferrari, Martial Hebert, Cristian Sminchisescu, and Yair Weiss, editors, ECCV, pages 185-201, Cham, 2018. Springer International Publishing. 1, 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.746, + 0.892, + 0.8 + ], + "angle": 0, + "content": "[28] Chenlin Meng, Yang Song, Jiaming Song, Jiajun Wu, JunYan Zhu, and Stefano Ermon. Sdedit: Image synthesis and editing with stochastic differential equations. arXiv preprint arXiv:2108.01073, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.802, + 0.892, + 0.857 + ], + "angle": 0, + "content": "[29] Aditya Krishna Menon, Sadeep Jayasumana, Ankit Singh Rawat, Himanshu Jain, Andreas Veit, and Sanjiv Kumar. Long-tail learning via logit adjustment. In ICLR. OpenReview.net, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.859, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[30] Takeru Miyato, Toshiki Kataoka, Masanori Koyama, and Yuichi Yoshida. Spectral normalization for generative adversarial networks. In ICLR, 2018. 7" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "18442" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.217 + ], + "angle": 0, + "content": "[31] Alexander Quinn Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. GLIDE: Towards photorealistic image generation and editing with text-guided diffusion models. In Kamalika Chaudhuri, Stefanie Jegelka, Le Song, Csaba Szepesvari, Gang Niu, and Sivan Sabato, editors, International Conference on Machine Learning (ICML), volume 162 of Proceedings of Machine Learning Research, pages 16784-16804. PMLR, 17-23 Jul 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.218, + 0.47, + 0.272 + ], + "angle": 0, + "content": "[32] Wanli Ouyang, Xiaogang Wang, Cong Zhang, and Xiaokang Yang. Factors in finetuning deep model for object detection with long-tail distribution. In CVPR, pages 864-873, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.274, + 0.469, + 0.315 + ], + "angle": 0, + "content": "[33] Yuntian Gu Quanlin Wu, Hang Ye. Guided diffusion model for adversarial purification from random noise. arXiv preprint arXiv:2206.10875, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.316, + 0.469, + 0.398 + ], + "angle": 0, + "content": "[34] Harsh Rangwani, Naman Jaswani, Tejan Karmali, Varun Jampani, and R. Venkatesh Babu. Improving gans for long-tailed data through group spectral regularization. In Shai Avidan, Gabriel Brostow, Moustapha Cisse, Giovanni Maria Farinella, and Tal Hassner, editors, ECCV, pages 426-442, Cham, 2022. Springer Nature Switzerland. 1, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.4, + 0.469, + 0.495 + ], + "angle": 0, + "content": "[35] Harsh Rangwani, Konda Reddy Mopuri, and R. Venkatesh Babu. Class balancing gan with a classifier in the loop. In Cassio de Campos and Marloes H. Maathuis, editors, Proceedings of the Thirty-Seventh Conference on Uncertainty in Artificial Intelligence, volume 161 of Proceedings of Machine Learning Research, pages 1618-1627. PMLR, 27-30 Jul 2021. 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.497, + 0.469, + 0.55 + ], + "angle": 0, + "content": "[36] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, pages 10684-10695, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.552, + 0.469, + 0.62 + ], + "angle": 0, + "content": "[37] Kevin Roth, Aurelien Lucchi, Sebastian Nowozin, and Thomas Hofmann. Stabilizing training of generative adversarial networks through regularization. In NeurIPS, page 2015-2025, Red Hook, NY, USA, 2017. Curran Associates Inc. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.622, + 0.469, + 0.689 + ], + "angle": 0, + "content": "[38] Chitwan Sahara, William Chan, Huiwen Chang, Chris A. Lee, Jonathan Ho, Tim Salimans, David J. Fleet, and Mohammad Norouzi. Palette: Image-to-image diffusion models. ACM SIGGRAPH 2022 Conference Proceedings, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.692, + 0.469, + 0.746 + ], + "angle": 0, + "content": "[39] Mehdi S. M. Sajjadi, Olivier Bachem, Mario Lucic, Olivier Bousquet, and Sylvain Gelly. Assessing generative models via precision and recall. In NeurIPS, page 5234-5243, Red Hook, NY, USA, 2018. Curran Associates Inc. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.748, + 0.469, + 0.815 + ], + "angle": 0, + "content": "[40] Tim Salimans, Ian Goodfellow, Wojciech Zaremba, Vicki Cheung, Alec Radford, Xi Chen, and Xi Chen. Improved techniques for training gans. In D. Lee, M. Sugiyama, U. Luxburg, I. Guyon, and R. Garnett, editors, NeurIPS, volume 29. Curran Associates, Inc., 2016. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.818, + 0.469, + 0.872 + ], + "angle": 0, + "content": "[41] Vikash Sehwag, Caner Hazirbas, Albert Gordo, First Ozgenel, and Cristian Canton. Generating high fidelity data from low-density regions using diffusion models. In CVPR, pages 11492-11501, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.874, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[42] Jascha Sohl-Dickstein, Eric A. Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "learning using nonequilibrium thermodynamics. arXiv preprint arXiv:1503.03585, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.122, + 0.892, + 0.149 + ], + "angle": 0, + "content": "[43] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models, 2021. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.151, + 0.892, + 0.205 + ], + "angle": 0, + "content": "[44] Yang Song, Liyue Shen, Lei Xing, and Stefano Ermon. Solving inverse problems in medical imaging with score-based generative models. arXiv preprint arXiv:2111.08005, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.208, + 0.892, + 0.248 + ], + "angle": 0, + "content": "[45] Hung-Yu Tseng, Lu Jiang, Ce Liu, Ming-Hsuan Yang, and Weilong Yang. Regularizing generative adversarial networks under limited data. In CVPR, pages 7917-7927, 2021. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.25, + 0.892, + 0.305 + ], + "angle": 0, + "content": "[46] Aaron van den Oord, Oriol Vinyals, and Koray Kavukcuoglu. Neural discrete representation learning. In NeurIPS, NIPS'17, page 6309-6318, Red Hook, NY, USA, 2017. Curran Associates Inc. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.307, + 0.892, + 0.36 + ], + "angle": 0, + "content": "[47] Jianfeng Wang, Thomas Lukasiewicz, Xiaolin Hu, Jianfei Cai, and Zhenghua Xu. Rsg: A simple but effective module for learning imbalanced datasets. In CVPR, pages 3783-3792, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.363, + 0.892, + 0.416 + ], + "angle": 0, + "content": "[48] Xiulong Yang, Sheng-Min Shih, Yinlin Fu, Xiaoting Zhao, and Shihao Ji. Your vit is secretly a hybrid discriminative-generative diffusion model. ArXiv, abs/2208.07791, 2022. 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.42, + 0.892, + 0.474 + ], + "angle": 0, + "content": "[49] Xi Yin, Xiang Yu, Kihyuk Sohn, Xiaoming Liu, and Manmohan Chandraker. Feature transfer learning for face recognition with under-represented data. In CVPR, pages 5697-5706, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.476, + 0.892, + 0.531 + ], + "angle": 0, + "content": "[50] Shengyu Zhao, Zhijian Liu, Ji Lin, Jun-Yan Zhu, and Song Han. Differentiable augmentation for data-efficient gan training. In NeurlPS, Red Hook, NY, USA, 2020. Curran Associates Inc. 1, 2, 5, 6" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.531 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.956 + ], + "angle": 0, + "content": "18443" + } + ] +] \ No newline at end of file diff --git a/2023/Class-Balancing Diffusion Models/eae79914-520a-4687-a1d4-1e04f53e44f8_origin.pdf b/2023/Class-Balancing Diffusion Models/eae79914-520a-4687-a1d4-1e04f53e44f8_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..6b68556523b8d1f3ed8ee7746ddf4b3005e60a62 --- /dev/null +++ b/2023/Class-Balancing Diffusion Models/eae79914-520a-4687-a1d4-1e04f53e44f8_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d6f7d026c1354d908c123bbd1fd6a75b1c90dbbbfeeabbd460192e1bc97ab1ae +size 2237572 diff --git a/2023/Class-Balancing Diffusion Models/full.md b/2023/Class-Balancing Diffusion Models/full.md new file mode 100644 index 0000000000000000000000000000000000000000..a60fbc859858fcf997e8dca25505a7b2e8cc2bfa --- /dev/null +++ b/2023/Class-Balancing Diffusion Models/full.md @@ -0,0 +1,316 @@ +# Class-Balancing Diffusion Models + +Yiming Qin + +Huangjie Zheng + +Jiangchao Yao $^{1,3}$ + +Mingyuan Zhou + +Ya Zhang $^{1,3}$ + +$^{1}$ Cooperative Medianet Innovation Center, Shanghai Jiao-Tong University + +$^{2}$ University of Texas, Austin + +$^{3}$ Shanghai AI Laboratory + +# Abstract + +Diffusion-based models have shown the merits of generating high-quality visual data while preserving better diversity in recent studies. However, such observation is only justified with curated data distribution, where the data samples are nicely pre-processed to be uniformly distributed in terms of their labels. In practice, a long-tailed data distribution appears more common and how diffusion models perform on such class-imbalanced data remains unknown. In this work, we first investigate this problem and observe significant degradation in both diversity and fidelity when the diffusion model is trained on datasets with class-imbalanced distributions. Especially in tail classes, the generations largely lose diversity and we observe severe mode-collapse issues. To tackle this problem, we set from the hypothesis that the data distribution is not class-balanced, and propose Class-Balancing Diffusion Models (CBDM) that are trained with a distribution adjustment regularizer as a solution. Experiments show that images generated by CBDM exhibit higher diversity and quality in both quantitative and qualitative ways. Our method benchmarked the generation results on CIFAR100/CIFAR100LT dataset and shows outstanding performance on the downstream recognition task. + +# 1. Introduction + +In recent years, log-likelihood-based diffusion models have evolved rapidly and established new benchmarks on a range of generation tasks [1,7]. Based on them, researchers have been able to further control the model generation process and the generation quality. This improves the applications of generative models in numerous domains including text-image generation [31], image editing [28, 38], speech synthesis [17], medical imaging [26, 44], video generation [13] and adversarial learning [18, 33], etc. + +Although diffusion models are known for the power of high fidelity and diversity in generation, most of the existing diffusion models are trained with the hypothesis that the data are uniformly distributed w.r.t. their labels. However, + +![](images/b52f1f93a407fa27ecdca458a5edea70ef052a41c50f7baf264ddbde5122540d.jpg) +Figure 1. Generation degrades along with class frequency. Semantics of generated images become less recognizable when class frequency decreases, while the FID score increases significantly. + +in the real world, the distribution is often very skewed. Especially for many domain-specific generation tasks such as medical images [14], fine-grained dataset for taxology [15] and data grabbed from the web [24], it is difficult to collect large amounts of data for each class equally, and the size of the training set for head and tail categories can differ by a factor of hundred or more. For such datasets, unconditional diffusion models tend to produce a significant portion of low-quality images. Conditional models, as shown in Figure 1, generate head class images with satisfying performance, while conversely the generated images on tail classes are very likely to show unrecognizable semantics. Concerning training generative models with limited data, there already exist several methods [21, 37, 50] based on GAN models [2]. However, quite few studies examine the impact of imbalanced class distribution [34] especially on diffusion models, which is practical yet under-explored. + +Our work first introduces diffusion models to imbalance generation tasks on several long-tailed datasets [19], and then build some straightforward baselines according to the common methods used in long-tailed recognition [27, 29]. + +To overcome the potential degeneration induced by the skewed distribution, we propose a novel Class-Balancing Diffusion Model (CBDM). Theoretically, CBDM resorts to adjusting the conditional transfer probability during sampling in order to implicitly force generated images to have a balanced prior distribution during every sampling step. Technically, the adjusted transfer probability of CBDM results in an additional MSE-form loss for a conditional diffusion model, which functions as a regularizer. Intuitively, this loss augments the similarity of generated images conditioned on different classes, and turns out to be an effective approach to transfer common information from head classes to tail classes without hurting the model's expressiveness on head classes. CBDM can be implemented within several flines of codes, and its lighter version admits fine-tuning an existing conditional model. We conducted extensive experiments on CIFAR10/CIFAR100 and their corresponding long-tailed dataset to show the promise of CBDM over existing state-of-the-art methods. In a nutshell, the contributions of this work can be summarized as follows: + +- We identify the severe degeneration problem of diffusion models in long-tailed generation tasks and benchmark some straightforward baselines in this direction. +- We propose a new perspective to handle the generation quality collapse on tail classes, and derive a novel Class-Balancing Diffusion Model, which is effective and lightweight as a regularizer to existing methods. +- We validate that CBDM is capable of generating more diverse images with convincing fidelity, especially for datasets with large number of categories. In addition, CBDM is robust with accelerating algorithms such as DDIM [43], and can be transplanted to different conditional diffusion-based backbones easily. + +# 2. Related Works + +Diffusion models Diffusion models are recently proposed generative models [42] based on non-equilibrium thermodynamics. Conditional diffusion models [7] encode label information into the generation process and improve largely the generation performance. The guidance structure proposed in [7] makes it possible to control the generation process through an external module. Based on a similar intuition, researchers arrive to realize diverse functions, such as guided adversarial purification [33], few-shot generation [9] and so on [36, 41]. The drawback of classifier guidance (noted as CG) [7] lies in its requirement of training another auxiliary classifier. To address the issue, classifier-free guidance (noted as CFG) [12] proposed a mechanism that uses the generator itself to express the class guidance information. CFG is proved to be a resource efficient method + +and achieves outstanding performance on large models [31]. Moreover, CFG only requires to add one line in training, which can be easily transplanted on different models. + +Long-tail recognition The problem of long-tailed distribution is a common dilemma in machine learning and have been widely explored in the area of discriminative models, i.e., long-tailed recognition. There mainly exist three paradigms in this domain, namely Class Re-balancing [27, 29], Information Augmentation [25,47,49] and Module Improvement [16, 32]. Among them, Class Re-balancing provides the best explainability, and its most common practice is re-sampling [27]. Thereafter, stemming from modifying the objective function from the global error rate to the class average one, [29] propose logit adjustment which has shown an impressive performance. Another effective method is based on Information Augmentation, which uses head class feature information to augment tail classes [5, 25, 47, 49]. However, discriminative models map data from higher to lower dimensions, while generative models map images from lower to higher dimensions. Thus, the mechanism of class rebalancing between them may be completely different and how to design the balancing method remains under-explored. + +Generative models based on limited data GAN [2, 6] is the most dominant model in the field of image synthesis in recent years. Given the requirement of large-scale data to train the generative models, a part of researchers focus on improving the performance of GAN models under the small datasets. To address the overfitting issue of the discriminator, a number of regularization methods [22, 37, 45] has been proposed. An alternative solution is Data Augmentation. As augmentation information is very prone to leak to the generator, researchers proposed improved augmentation strategies such as Differentiable Augmentation (DiffAug) [50] and Adaptive Augmentation (ADA) [21]. These augmentation ways can be transferred to diffusion models as baselines to help the generation of tail-class samples. + +# 3. Method + +In this section, we first give the basic notations following the classical DDPM model, and then introduce the class-imbalanced generation setting. In the third part, we introduce our CBDM algorithm and present its training details including implementation and hyper-parameter settings. + +# 3.1. Preliminary + +Diffusion models leverage a pre-defined forward process in training, where a clean image distribution $q(\pmb{x}_0)$ can be corrupted to a noisy distribution $q(\pmb{x}_t|\pmb{x}_0)$ at a specified timestep $t$ . Given a pre-defined variance schedule $\{\beta_t\}_{1:T}$ , + +![](images/93689ce378ac70f873db28d10f3497b9583aa537a599b4fc905f992fe5c9857e.jpg) +Figure 2. Algorithm (left) and generation (right) visualization. In the left figure, we show that an extra regularization loss $gL_{r}$ proportional to the diffusion step $t$ is added during training. This loss function pushes the sampling distribution (curves on the surface) to wider region while preventing it to be excessively distorted compared to the ground truth distribution (gradient color on the background). + +![](images/5020ec0e967463607da0afac543321a5b90bdf43930986ac8b6b71e2fd04bdb2.jpg) + +the noisy distribution at any intermediate timestep is + +$$ +q \left(\boldsymbol {x} _ {t} \mid \boldsymbol {x} _ {0}\right) = \mathcal {N} \left(\sqrt {\bar {\alpha} _ {t}} \boldsymbol {x} _ {0}, \left(1 - \bar {\alpha} _ {t}\right) \boldsymbol {I}\right); \quad \bar {\alpha} _ {t} = \prod_ {i = 1} ^ {t} \left(1 - \beta_ {i}\right). +$$ + +To reverse such forward process, a generative model $\theta$ learns to estimate the analytical true posterior in order to recover $x_{t - 1}$ from $x_{t}$ as follows: + +$$ +\min _ {\theta} D _ {\mathrm {K L}} [ q (\pmb {x} _ {t - 1} | \pmb {x} _ {t}, \pmb {x} _ {0}) | | p _ {\theta} (\pmb {x} _ {t - 1} | \pmb {x} _ {t}) ]; \forall t \in \{1, \dots , T \}, +$$ + +and such an objective can be reduced to a simple denoising estimation loss [11]: + +$$ +\mathcal {L} _ {\mathrm {D D P M}} = \mathbb {E} _ {t, \boldsymbol {x} _ {0} \sim q (\boldsymbol {x} _ {0}), \epsilon \sim \mathcal {N} (\boldsymbol {0}, I)} \left[ \| \boldsymbol {\epsilon} - \boldsymbol {\epsilon} _ {\theta} \left(\sqrt {\bar {\alpha}} \boldsymbol {x} _ {0} + \sqrt {1 - \bar {\alpha} _ {t}} \boldsymbol {\epsilon}, t\right) \| ^ {2} \right] \tag {1} +$$ + +For the case where label information is available, the model is trained to estimate the noise as above in both conditional cases $\epsilon_{\theta}(\pmb{x}_t,y,t)$ with data-label pairs $(x_0,y)$ and unconditional case $\epsilon_{\theta}(\pmb{x}_t,t)$ . In the sampling, the label-guided model estimates the noise with a linear interpolation $\hat{\epsilon} = (1 + \omega)\pmb{\epsilon}_{\theta}(\pmb{x}_t,y,t) - \omega \pmb{\epsilon}_{\theta}(\pmb{x}_t,t)$ to recover $\pmb{x}_{t-1}$ , which is often referred as Classifier-Free Guidance (CFG) [12]. + +# 3.2. Class-Balancing Diffusion Models + +Current diffusion models assume the data distribution to be uniform in class, and thus equally treat samples in the training stage. However, based on our observation, such training strategy leads to degradation in generation quality. Below, we provide an analysis that motivates our Class-Balancing Diffusion Models (CBDM). + +Suppose $q(\pmb{x}, y)$ is the data distribution that we need to match with the joint distribution $p_{\theta}(\pmb{x}, y)$ predicted by a generative model. We analyze their difference from the density ratio $r = \frac{q(\pmb{x}, y)}{p_{\theta}(\pmb{x}, y)} = \frac{q(\pmb{x} \mid y)}{p_{\theta}(\pmb{x} \mid y)} \cdot \frac{q(y)}{p_{\theta}(y)}$ . When the true label distribution $q(y)$ is the same as the prior $p_{\theta}(y)$ , + +which is usually assumed to be uniform, the density ratio $r$ is reduced the conditional term to learn a conditional model $p_{\theta}(\boldsymbol{x} \mid y)$ . However, when such a hypothesis is violated, for head classes, $\frac{q(y)}{p_{\theta}(y)}$ would result in a larger weight that makes the model biased and hurt tail classes, and vice versa. Empirically, we observe that the generation degrades more on tail classes, as illustrated in Figure 1. Moreover, as shown in Figure 2a, compared to head classes, DDPM cannot well capture the tail-class data distribution and the mode is poorly covered during the sampling process. As a result, generations of tail classes often have poor quality and diversity, shown in Figure 2b. + +To tackle this issue, the most intuitive approach lies in adjusting the prior label distribution through a class balanced re-sampling. However, such abrupt adjustment easily leads to negative improvement in experiments. The step-by-step sampling nature of diffusion models provides another aspect to adjust this distribution more softly. In this spirit, we propose to calibrate the learning process through the conditional transfer probability $p_{\theta}(\pmb{x}_{t-1}|\pmb{x}_t,y)$ when there exists a gap between the class distribution and the prior. + +Let $p_{\theta}^{\star}\left(\boldsymbol{x}_{t-1} \mid \boldsymbol{x}_t, y\right)$ be the optimum trained in the case that $\frac{q(y)}{p_{\theta}(y)}$ is correctly estimated, and $p_{\theta}\left(\boldsymbol{x}_{t-1} \mid \boldsymbol{x}_t, y\right)$ be the one trained in a class-imbalanced case. The relation between such two generative distributions can be described as the following proposition. + +Proposition 1. When training a diffusion model parameterized with $\theta$ on a class-imbalanced dataset, its conditional reverse distribution $p_{\theta}\left(\boldsymbol{x}_{t-1}|\boldsymbol{x}_t,y\right)$ can be corrected with an adjustment schema: + +$$ +p _ {\theta} ^ {\star} \left(\boldsymbol {x} _ {t - 1} \mid \boldsymbol {x} _ {t}, y\right) = p _ {\theta} \left(\boldsymbol {x} _ {t - 1} \mid \boldsymbol {x} _ {t}, y\right) \frac {p _ {\theta} \left(\boldsymbol {x} _ {t - 1}\right)}{p _ {\theta} ^ {\star} \left(\boldsymbol {x} _ {t - 1}\right)} \frac {q ^ {\star} \left(\boldsymbol {x} _ {t}\right)}{q \left(\boldsymbol {x} _ {t}\right)} \tag {2} +$$ + +The proposition above shows that, when trained on a + +class-imbalanced dataset, a diffusion model can still approach the true data distribution by applying a distribution adjustment schema $\frac{p_{\theta}(\boldsymbol{x}_{t-1})}{p_{\theta}^*(\boldsymbol{x}_{t-1})}\frac{q^*(\boldsymbol{x}_t)}{q(\boldsymbol{x}_t)}$ at every step $t$ . However, approximating this schema is not feasible at every sampling step, so CBDM incorporates it into the training loss function to achieve an equivalent objective, and thus gets rid of the model-free part $\frac{q^*(\boldsymbol{x}_t)}{q(\boldsymbol{x}_t)}$ . By further decomposing $p_{\theta}(\boldsymbol{x}_{t-1})$ and $p_{\theta}^*(\boldsymbol{x}_{t-1})$ to the expectation of the conditional probability $p_{\theta}^*(\boldsymbol{x}_{t-1}|\boldsymbol{x}_{t:T},y)$ , we present a upper bound to approximate this probability in Proposition 2. + +Proposition 2. For the adjusted loss $\mathcal{L}_{DM}^{\star} = \sum_{t=1}^{T} \mathcal{L}_{t-1}^{\star}$ , an upper-bound of the target training objective to calibrate at timestep $t$ (i.e., $\mathcal{L}_{t-1}^{\star}$ ) can be derived as: + +$$ +\begin{array}{l} \sum_ {t \geq 1} \mathcal {L} _ {t - 1} ^ {\star} = \sum_ {t \geq 1} D _ {\mathrm {K L}} [ q (\pmb {x} _ {t - 1} | \pmb {x} _ {t}, \pmb {x} _ {0}) | | p _ {\theta} ^ {\star} (\pmb {x} _ {t - 1} | \pmb {x} _ {t}, y) ] \\ \leq \sum_ {t \geq 1} [ \underbrace {D _ {\mathrm {K L}} [ q (\boldsymbol {x} _ {t - 1} | \boldsymbol {x} _ {t} , \boldsymbol {x} _ {0}) | | p _ {\theta} (\boldsymbol {x} _ {t - 1} | \boldsymbol {x} _ {t} , y) ]} _ {\text {D i f f u s i o n m o d e l l o s s} \mathcal {L} _ {D M}} \\ + \underbrace {t \mathbb {E} _ {y ^ {\prime} \sim q _ {y} ^ {\star}} [ D _ {\mathrm {K L}} [ p _ {\theta} (\boldsymbol {x} _ {t - 1} | \boldsymbol {x} _ {t}) | | p _ {\theta} (\boldsymbol {x} _ {t - 1} | \boldsymbol {x} _ {t} , y ^ {\prime}) ] ] ]} _ {D i s t r i b u t i o n a d j u s t m e n t l o s s \mathcal {L} _ {r}}, \\ \end{array} +$$ + +The upper bound in the above proposition can be considered as two parts. The first term $\mathcal{L}_{\mathrm{DM}}$ corresponds to an ordinary DDPM loss [11] e.g., Eqn. (1), and the second loss $\mathcal{L}_{\mathrm{r}}$ is used to adjust the distribution as a regularization term. Roughly speaking, $\mathcal{L}_{\mathrm{r}}$ increases the similarity between the model's output and a random target class. Thus, it reduces the risk of overfitting on the head classes, and enlarges the generation diversity for tail class through knowledge obtained from other classes. When $q_{\mathcal{Y}}^{\star}$ is less longtailed than the dataset, this loss also increases the probability for underrepresented tail samples to be chosen during training. + +# 3.3. Training algorithm + +The detailed training algorithm of CBDM is presented in Alg. (1). In the algorithm, we reduce the distribution adjustment loss $\mathcal{L}_r$ as a square error loss with Monte-Carlo samples as indicated by Eqn.(3), where $\mathcal{V}$ is a set of samples that drawn following the distribution $q_{\mathcal{V}}^{\star}$ and $y$ denotes the image label. Note for CFG [12], there is a fixed probability (usually $10\%$ ) to drop the condition, i.e., $y = \text{None}$ . + +$$ +\mathcal {L} _ {\mathrm {r}} \left(\boldsymbol {x} _ {t}, y, t\right) = \frac {1}{| \mathcal {Y} |} \sum_ {y ^ {\prime} \in \mathcal {Y}} [ t \| \epsilon_ {\theta} \left(\boldsymbol {x} _ {t}, y\right) - \epsilon_ {\theta} \left(\boldsymbol {x} _ {t}, y ^ {\prime}\right) \| ^ {2} ], \tag {3} +$$ + +For the implementation in practice, CBDM can be plugged into any existing conditional diffusion models by adopting their model architecture and adjusting the training loss $\mathcal{L}_{\mathrm{DM}}$ following lines 6 - 11. Specifically, the choice + +Algorithm 1 Training algorithm of CBDM. +1: for Every batch of size N do +2: for $(\pmb{x}_0^{(i)}, y^{(i)})$ in this batch do +3: Sample $\pmb{\epsilon}^{(i)} \sim \mathcal{N}(\pmb{0}, \pmb{I}), t \sim \mathcal{U}(\{0, 1, \dots, T\})$ +4: $x_{t}^{(i)} = \sqrt{\bar{\alpha}_{t}} x_{0}^{(i)} + \sqrt{1 - \bar{\alpha}_{t}} \epsilon^{(i)}$ +5: Calculate $\mathcal{L}_{\mathrm{DM}} = \| \pmb{\epsilon}^{(i)} - \pmb{\epsilon}_{\theta}(\pmb{x}_{t}^{(i)}, y^{(i)}) \|^{2}$ +6: Sample $y'(i)$ from $q_{\gamma}^{\star}$ +7: Calculate the first regularization term +8: $\mathcal{L}_{r} = t\tau \| \pmb{\epsilon}_{\theta}(\pmb{x}_{t}^{(i)}, y^{(i)}) - \mathrm{sg}(\pmb{\epsilon}_{\theta}(\pmb{x}_{t}^{(i)}, y'^{(i)})) \|^{2}$ +9: Calculate the regularization commitment term +10: $\mathcal{L}_{rc} = t\tau \| \mathrm{sg}(\pmb{\epsilon}_{\theta}(\pmb{x}_{t}^{(i)}, y^{(i)})) - \pmb{\epsilon}_{\theta}(\pmb{x}_{t}^{(i)}, y'^{(i)}) \|^{2}$ +11: Update with $\mathcal{L}_{\mathrm{CBDM}} = \mathcal{L}_{\mathrm{DM}} + \mathcal{L}_{r} + \gamma \mathcal{L}_{rc}$ +12: end for +13: end for + +of the regularization weight $\tau$ affects the sharpness of the density ratio $\frac{p_{\theta}(\boldsymbol{x}_t)}{p_{\theta}^*(\boldsymbol{x}_t)}$ . For the theoretical analysis, please refer to our proof of Prop. (2) in the Appendix. In addition, the choice of the sampling set $\mathcal{V}$ is another important perspective of CBDM, which depends on the target distribution we wish to adjust. Without loss of generality, we discuss two cases here. On the one hand, we can adjust the label distribution to a class-balanced label distribution, where we sample labels to construct $\mathcal{V}^{bal}$ . On the other hand, if the data distribution is heavily long-tailed, we can also target the adjusted distribution to a relatively less class-imbalanced distribution for stabilized training. In our experiments, we show CBDM can work well for both cases in different mechanisms. + +Moreover, we observe that naively optimizing with this loss could make the model collapse to some trivial solutions, where the model outputs the same result regardless of the condition $y$ and thus degenerates the conditional generation performance. Therefore, we follow previous works to apply a stop gradient operation [4, 46] to prevent this issue. The final loss of CBDM is + +$$ +\begin{array}{l} \mathcal {L} _ {\mathrm {C B D M}} ^ {\tau , \gamma , \mathcal {Y}} (\pmb {x} _ {t}, y, t, \pmb {\epsilon}) = \| \pmb {\epsilon} _ {\theta} (\pmb {x} _ {t}, y) - \pmb {\epsilon} \| ^ {2} \\ + \frac {\tau t}{| \mathcal {Y} |} \sum_ {y ^ {\prime} \in \mathcal {Y}} \left(\left\| \boldsymbol {\epsilon} _ {\theta} \left(\boldsymbol {x} _ {t}, y\right) - \operatorname {s g} \left(\boldsymbol {\epsilon} _ {\theta} \left(\boldsymbol {x} _ {t}, y ^ {\prime}\right)\right) \right\| ^ {2} \right. \\ + \gamma \| \mathrm {s g} \big (\pmb {\epsilon} _ {\theta} (\pmb {x} _ {t}, y) \big) - \pmb {\epsilon} _ {\theta} (\pmb {x} _ {t}, y ^ {\prime}) \| ^ {2}), (4) \\ \end{array} +$$ + +where $\mathrm{sg}(\cdot)$ denotes the stop gradient operation; $\tau, \gamma$ are weights for the regularization and the commitment term respectively with $\gamma$ set to $\frac{1}{4}$ in a default setting, and $|\mathcal{V}|$ denotes the number of elements in the label set. + +![](images/dd0ccc3c990d101184872c016a8f0d46a73b9ecb4f44c1ece8ca8aedd15ec652.jpg) +(a) DDPM / Tail Class 94 + +![](images/a5e275ecd61e93b3ee7ff383f1185faf0a4e67d4a25b8eb536ef5ccfc200bed5.jpg) +(b) CBDM / Tail Class 94 + +![](images/946995c13751da3c24e493e2e4d8a3126d1ab1943c5615a9938206a54b8d6319.jpg) +(c) DDPM / Body Class 62 +Figure 3. Comparison of image generation on heavily tail-distributed (94) and mild tail-distributed (62) classes between DDPM and CBDM. + +![](images/1f6736bb9992078b5667558852d87a159b20deba9231a1a4a87b0ab4651d0ef8.jpg) +(d) CBDM / Body Class 62 + +# 4. Experimental results + +# 4.1. Experimental setup + +Datasets We first chose two common datasets that are widely used in the domain of image synthesis, CIFAR10/CIFAR100, and their corresponding long-tailed versions CIFAR10LT and CIFAR100LT. The construction of CIFAR10LT and CIFAR100LT follows [3], where the size decreases exponentially with its category index according to the imbalance factor $imb = 0.01$ . We also conduct experiments on 3 higher-resolution datasets, whose details and results could be found in the Appendix. + +Implementation details We strictly follow the training configurations of the baseline models. For DDPM, we set diffusion schedule $\beta_{1} = 10^{-4}$ and $\beta_{T} = 0.02$ with $T = 1,000$ , and optimize the network with an Adam optimizer whose learning rate is 0.0002 after 5,000 epochs of warmup. Considering that the size and semantic complexity of the datasets vary greatly, we choose appropriate epochs for each dataset. In the ablation studies, we also include EDM [20] and Diffusion-ViT [48] as our backbones to show the compatibility with different architectures and score-matching methods. Note that, when a model does not use conditional input, we follow [7] to slightly modify the backbone [8] by adding an extra embedding layer. + +Baselines We first adopt following classic methods to build baselines: re-sampling (RS), soft re-sampling methods (RS-SQRT) [27], and augmentation-based methods including Differentiable Augmentation (DiffAug) [50] and Adaptive Augmentation (ADA) [21]. Besides diffusion-based baselines, previous state-of-the-art generative models that study the long-tailed distribution, e.g., CBGAN [35], and group spectral regularization for GANs [34], are also included in the comparison. Precisely, RS follows the uniform class distribution and RS-SQRT [27] samples with a probability determined by the square root of class frequency. And DiffAug [50] is directly applied to all training images following their default threshold. Following [20], ADA [21] is not only applied to images, the augmentation pipeline is also encoded as a condition via an additional embedding layer in the U-Net. + +Metrics CBDM and corresponding baselines are evaluated in terms of both generation diversity and fidelity via Frechet Inception Distance (FID) [10], Inception Score (IS) [40], Recall [23] and $F_{\beta}$ [39]. We measure Recall and $F_{\beta}$ using Inception-V3 features, and take respectively $K = 5$ for Recall, 1/8 and 8 for the threshold in $F_{\beta}$ , and 20 times of class number as the clustering number of $F_{\beta}$ to capture the inner class variance. To this end, Recall and $F_{8}$ can be regarded as diversity metrics and IS and $F_{1/8}$ are more inclined to measure fidelity. In evaluation, we take the class-balanced version corresponding to those used in training, and all metrics are measured with 50k generated images (10k for ablation experiments). The classifier-free guidance is adapted in sampling and we tune the guidance strength $\omega$ for both baselines and our method to ensure the best performance. Specifically, we take $\omega = 1.6, 0.8, 1.0, 0.8$ respectively for the four CIFAR datasets mentioned above. + +# 4.2. Main results + +In this part, we consider DDPM as the most direct baseline. As shown in Table 1, CBDM overall outperforms DDPM on all datasets except the IS is slightly lower on CIFAR10. For datasets with more classes, the improvements are more significant in terms of both diversity and fidelity, as on CIFAR100LT and CIFAR100. We further investigate the performance of diffusion models when combined with augmentation-based and re-sampling methods. Surprisingly, except for ADA, we consistently observe a performance degradation when these methods are compared with the vanilla DDPM. Moreover, adding the ADA augmentation with CBDM training leads to a further improvement in terms of both diversity and fidelity. + +As a qualitative justification, Figure 3 provides the visualization comparison between DDPM and CBDM on a relatively mild tail-distributed class (62) and a tail-distributed class (94). We remark that CBDM generates more diverse images. For example, on the tail class 94, the cabinets + +
DatasetModelFID↓F8↑Recall↑IS↑F1/8↑
CIFAR100LTDDPM [11]7.380.850.5213.110.88
+ADA [21]6.160.910.5712.710.90
+DiffAug [50]9.190.880.4711.560.86
+RS [27]10.500.650.4912.600.83
+SQRT-RS [27]9.720.660.4713.470.83
CBDM (ours)6.26 (-1.12)0.91 (+0.06)0.57 (+0.05)13.24 (+0.13)0.89 (+0.01)
+ADA [21]5.81 (-1.57)0.91 (+0.06)0.57 (+0.05)13.34 (+0.23)0.90 (+0.02)
CIFAR100DDPM3.110.970.6513.650.96
CBDM (ours)2.72 (-0.39)0.97 (±0)0.67 (+0.02)14.03 (+0.38)0.96 (±0)
CIFAR10LTDDPM5.760.970.579.170.95
CBDM (ours)5.46 (-0.30)0.97 (±0)0.59 (+0.02)9.28 (+0.11)0.95 (±0)
CIFAR10DDPM3.160.990.649.800.98
CBDM (ours)3.03 (-0.13)0.99 (±0)0.65 (+0.01)9.63 (-0.17)0.98 (±0)
+ +Table 1. CBDM performance on different datasets. In the table, the first three columns are diversity-related, and we mark the best results in bold. As the comparison between CBDM and DDPM is more straightforward, we mark the performance gain in parentheses next to the CBDM results, using blue and red to indicate improvements and degradation respectively. + +
PTFTyFIDF8RecallIS
CIFAR 100LT-ytrain6.260.910.5713.24
ybal6.100.940.6412.29
ysqrt6.000.930.5913.00
CIFAR 100LTCIFAR 100LT-7.380.850.5213.11
ytrain6.200.880.5413.36
ybal5.850.920.6013.29
CIFAR 100CIFAR 10LT-5.390.660.589.43
ytrain5.050.670.609.48
ybal5.900.660.609.49
+ +Table 2. CBDM performance for 3 mechanisms under different regularization sampling set $\mathcal{V}$ . Column $PT$ indicates the dataset used in (pre)training, column $FT$ indicates the dataset used for fine-tuning. The rows where $\mathcal{V}$ marked as “-” represent the results with DDPM, and rows where $FT$ marked as “-” represent the results of CBDM trained from scratch. Three diversity-preferred metrics and the best results are in bold for emphasis. + +shown in Figure 3b have more color and texture, which justifies the improvements of diversity-related metrics in Table 1. On the contrary, DDPM only generates images that are highly similar to the training data. + +Case-by-case study. To better understand the generation conditioned on each class, we compare the FID case-by-case between DDPM and CBDM on each class. The results are shown in Figure 4. Compared to DDPM, CBDM shows more consistent improvements for the class index greater than 40. In the tail classes, CBDM outperforms than DDPM the FID by a more significant margin. + +![](images/c558ad75a1e4270f6c288b5f82a37507320ddcc04703d71f4d45644f2e9edd1c.jpg) +Figure 4. FID improvement per class compared to DDPM. The curve is smoothened by a moving average of 5 unit. + +The choice of label set $\mathcal{V}$ We investigate the effects of label distribution for CBDM under three mechanisms. The first mechanism is the vanilla CBDM given in Algorithm 1 which trains a diffusion model with regularization from scratch. Besides, we also study CBDM with two fine-tuning configurations. The first setting concerns using CBDM to adjust an existing model pre-trained on a long-tailed dataset to improve its performance. The second one refers to fine-tuning a pre-trained model to adapt to a smaller dataset with an imbalanced class distribution using CBDM. For those mechanisms, we consider three settings for sampling the label set: (1) a label distribution similar to the training set (denoted $\mathcal{V}^{train}$ ); (2) a totally uniform label distribution (denoted $\mathcal{V}^{bal}$ ); (3) a less long-tailed distribution compared to $\mathcal{V}^{train}$ , whose class frequency is the square root of the original one (denoted $\mathcal{V}^{sqrt}$ ). + +The results are shown in Table 2, where we found that CBDM performs well on all these configurations. As we + +
DatasetModelFIDIS
CIFAR 100BigGAN+DA+RLC[45](2021)2.99-
DDPM3.1113.65
CBDM (ours)2.7214.03
CIFAR 100LTCBGAN [35](2021)28.17-
DDPM12.3112.69
CBDM (ours)+ADA [21]10.5012.81
CIFAR 10LTCBGAN [35](2021)32.93-
SNGAN [30]+gSR [34](2022)18.587.80
DDPM9.689.00
CBDM (ours)9.389.12
+ +Table 3. Comparison with long-tailed SoTAs on CIFAR. Following their setting, all methods are evaluated with $10\mathrm{k}$ generated images and ground truth images are from their validation dataset. The publish year of every baselines are marked in next to the citation. + +
Training dataPrecisionRecall
CIFAR1000.700.67
CIFAR100LT + DDPMgens (50k)0.450.39
0.48 (+0.03)0.44 (+0.05)
+ CBDMgens (50k)0.49 (+0.04)0.47 (+0.08)
+ +analyzed in section 3.3, when the empirical distribution of $\mathcal{V}$ has a significant difference from the label distribution of the dataset, adjusting with $\mathcal{V}^{bal}$ could hurt training stability and result in performance degradation. We observe that $\mathcal{V}^{bal}$ only shows better performance regarding generation diversity, while the generation fidelity has an obvious gap than the other settings. Using a relatively mild set $\mathcal{V}^{sqrt}$ has better performance in terms of IS and preserves the generation diversity, which leads to the best FID among all these settings. On the contrary, when a pre-trained model is available, fine-tuning on top of it ensures better stability, and we can observe in such case using $\mathcal{V}^{bal}$ has significant improvement than the vanilla DDPM or fine-tuning using $\mathcal{V}^{train}$ . When adapting to a different dataset, we also observe using $\mathcal{V}^{bal}$ causes stability issue and produce undesired results. + +Enhancement of training classifiers on long-tailed data As training models on long-tailed data usually leads to undesired classification results, we investigate whether the generated data could help improve the classifier trained on long-tailed data as a complementary evaluation metrics. Here, we train ResNet-32 models, respectively with CIFAR100, CIFAR100LT and CIFAR100LT augmented with DDPM and CBDM generations (50k samples). Table 4 shows that the training on long-tailed data results in se + +Table 4. Recognition results of different training data. All configurations are evaluated on the testing set of normal CIFAR100. Gains based on CIFAR100LT dataset is noted in blue. + +
BackboneFID\( F_8 \)RecallIS\( F_{1/8} \)
EDM [20]8.640.820.4812.160.86
+CBDM (ours)7.970.880.5212.010.86
Diffuse-ViT [48]37.60.760.457.660.62
+CBDM (ours)30.00.810.518.000.65
+ +Table 5. CBDM performance using different backbones on CIFAR100LT dataset. Three diversity-biased metrics. + +
DatasetFID↓F8↑Recall↑IS↑F1/8↑
CIFAR3.440.970.6913.110.96
100(+0.72)(±0)(+0.02)(-0.92)(±0)
CIFAR6.270.930.6112.560.89
100LT(+0.01)(+0.02)(+0.04)(-0.68)(±0)
+ +Table 6. Sampling with 100 DDIM steps on CIFAR100 and CIFAR100LT. The difference compared with those using DDPM steps are reported below each result. + +vere performance degradation. When augmenting the long-tailed data with the generated data, we observe fairly improvements in the trained classifier. Moreover, comparing the results between DDPM and CBDM, we observe that the improvements by using CBDM are more significant, which validates the effectiveness of our approach. Especially, the gain on recall is more remarkable than the gain on precision, which means that the generated images are more diverse than DDPM and justifies the results in Table 1. + +Comparison with other benchmarks We include the representative state-of-the-art long-tailed generative modeling approaches in the comparison, most of which are based on GANs since they suffer more severe problems when the training data is skewed. For a fair comparison, we strictly follow their evaluation setting and the results are reported in Table 3. From the results, DDPM surpasses the performance of the baselines on all the datasets, and CBDM exhibits even stronger performance. + +# 4.3. Ablations + +Compatibility with different backbones We first investigate the performance of our method with different diffusion models. Without loss of generality, we adopt EDM [20] and Diffusion-ViT [48], which use a different score-matching method and different denoising network backbone. As shown in Table 5, compared with the corresponding baselines, CBDM improves the generation result on CIFAR100LT dataset for both backbones, demonstrating the compatibility with a variety of backbones and the effectiveness of handling long-tailed data. + +CBDM with DDIM sampling Apart from DDPM reverse sampling and neural SDE methods, recent deterministic ODE methods such as DDIM [43] are widely used in the + +![](images/75a37acadd05bc48854fe94ad9c7471bb083e68cbc4bf9dbf366ed162efb6aee.jpg) +Figure 5. FID/IS score under different regularization weight $\tau$ + +![](images/fe780cded2078a75ff6b1eceef772b462c1b30bc7180e3e0bcce931b98a322b5.jpg) + +![](images/9069478d6c2608070b1710d63bfb23e68e15d906fe96c02ba312399978efd9cc.jpg) +Figure 6. FID/IS score under different guidance strength $\omega$ + +![](images/c1e02173e8bc13d89d5bce211f5e972fcd8a07c83970705729f9b7c8898e749e.jpg) + +generation process of diffusion models. We also conduct experiments to study whether CBDM is compatible with DDIM steps. Table 6 shows the comparison of using DDPM and DDIM reverse steps when the generation steps are compressed to $1/10$ . We note on normal CIFAR100, FID and IS become slightly worse than using 1000 DDPM steps. On CIFAR100LT, except for IS, all metrics remain almost the same as using DDPM steps. + +Effects of hyperparameters We tested the effects of regularization weight $\tau$ in CBDM, which is set in default $\tau = \tau_0 = 0.001$ so that the weight does not surpass 1 at all steps. As shown in Figure 5, we search weights across different scales and found that the optimal weight is $\tau = \tau_0$ . The results also point out that $\tau$ should not be too large or too small. We also tested the difference between CBDM models trained using commitment loss $\mathcal{L}_{rc}$ and those trained without it. Our experiments show that the FID score of the models using the commitment loss (8.30) are significantly lower than those of the models not using it (8.84). + +Guidance strength $\omega$ As we adapt the Classifier-Free Guidance (CFG) during sampling, we also show the effects of guidance strength $\omega$ . In Figure 6, we searched $\omega$ from 0 to 2 with an interval of 0.2 and compare FID and IS score of DDPM and CBDM models. We observe that the FID of CBDM remains decreasing at larger guidance strengths compared to DDPM; and although the IS of CBDM is consistently weaker than that of DDPM, they reach the same level when the guidance strength reaches 2.0, where the FID of CBDM (6.75) is still significantly lower than the best FID of DDPM (7.38). + +Fidelity-diversity control [12] describes the phenomenon where excessive guidance strength $\omega$ tends to lead to better fidelity at the price of overfitting and diversity degradation, as a fidelity-diversity tradeoff. CBDM + +![](images/3012b664f7380c1364ed1b785a1aa3ee14e747273724618fe884593f33f0e987.jpg) +Figure 7. Image fidelity and diversity controlled by guidance strength $\omega$ and regularization weight $\tau$ . + +additionally provides another hyperparameter: the regularization weight $\tau$ that promotes model diversity through cross-class information interaction. Figure 7 shows the body class (53) generation results under different guidance strengths and different regularization weight. We observe that a DDPM model with high guidance strength is able to generate very realistic results, but its images overlap almost exactly with a sample in the training set. In contrast, the guidance term of CBDM does not drastically change the image content, but rather refines the image content to be closer to the selected class than the unguided result. This means that CBDM retains more of the diversity of the unguided case and effectively enhances the category information with guide terms. + +# 5. Conclusion + +In this paper, we focus on the problem when the training data is imbalanced and the diffusion model drops in generation quality on tail classes. We first establish a low baseline for this task by examining the most common paradigms in long-tail scenarios and in training generative models with limited data. Thereafter, we propose the Class-Balancing Diffusion Model (CBDM) through theoretical analysis. This approach can be implemented very cleanly in the training of any conditional diffusion model and thus has the potential to be widely used in other fields. Our experiments show that the CBDM approach significantly improves model generation diversity with high fidelity, on both class-balanced and class-imbalanced datasets. + +# 6. Acknowledgement + +This work is supported by the National Key R&D Program of China (No. 2022ZD0160703, No. 2022ZD0160702), National Natural Science Foundation of China (62271308), STCSM (No. 22511106101, No. 18DZ2270700, No. 22511105700, No. 21DZ1100100), 111 plan (No. BP0719010), and State Key Laboratory of UHD Video and Audio Production and Presentation. + +# References + +[1] Jan Kautz Arash Vahdat, Karsten Kreis. Score-based generative modeling in latent space. In NeurIPS, 2021. 1 +[2] Andrew Brock, Jeff Donahue, and Karen Simonyan. Large scale GAN training for high fidelity natural image synthesis. In ICLR, 2019. 1, 2 +[3] Kaidi Cao, Colin Wei, Adrien Gaidon, Nikos Arechiga, and Tengyu Ma. Learning imbalanced datasets with label-distribution-aware margin loss. NeurIPS, 32, 2019. 5 +[4] Xinlei Chen and Kaiming He. Exploring simple siamese representation learning. In CVPR, pages 15750-15758, 2021. 4 +[5] Peng Chu, Xiao Bian, Shaopeng Liu, and Haibin Ling. Feature space augmentation for long-tailed data. In Andrea Vedaldi, Horst Bischof, Thomas Brox, and Jan-Michael Frahm, editors, ECCV, pages 694-710, Cham, 2020. Springer International Publishing. 2 +[6] Ishan Deshpande, Ziyu Zhang, and Alexander G Schwing. Generative modeling using the sliced Wasserstein distance. In CVPR, pages 3483-3491, 2018. 2 +[7] Prafulla Dhariwal and Alexander Quinn Nichol. Diffusion models beat GANs on image synthesis. In A. Beygelzimer, Y. Dauphin, P. Liang, and J. Wortman Vaughan, editors, NeurIPS, 2021. 1, 2, 5 +[8] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020.5 +[9] Didrik Nielsen Giorgio Giannone and Ole Winther. Few-shot diffusion models. arXiv preprint arXiv:2205.15463, 2021. 2 +[10] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. In NeurIPS, page 6629-6640, Red Hook, NY, USA, 2017. Curran Associates Inc. 5 +[11] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In NeurIPS, 2020. 3, 4, 6 +[12] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598, 2022. 2, 3, 4, 8 +[13] Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan, Mohammad Norouzi, and David J Fleet. Video diffusion models. arXiv preprint arXiv:2204.03458, 2022. 1 +[14] Greg Holste, Song Wang, Ziyu Jiang, Thomas C. Shen, George L. Shih, Ronald M. Summers, Yifan Peng, and Zhangyang Wang. Long-tailed classification of thorax diseases on chest x-ray: A new benchmark study. Data augmentation, labelling, and imperfections : second MICCAI workshop, DALI 2022, held in conjunction with MICCAI 2022, Singapore, September 22, 2022, proceedings. DALI (Workshop), 13567:22-32, 2022. 1 +[15] Grant Van Horn, Oisin Mac Aodha, Yang Song, Yin Cui, Chen Sun, Alexander Shepard, Hartwig Adam, Pietro Perona, and Serge J. Belongie. The inaturalist species classification and detection dataset. In CVPR, pages 8769-8778, 2018. 1 + +[16] Chen Huang, Yining Li, Chen Change Loy, and Xiaou Tang. Learning deep representation for imbalanced classification. In CVPR, pages 5375-5384, 2016. 2 +[17] Myeonghun Jeong, Hyeongju Kim, Sung Jun Cheon, Byoung Jin Choi, and Nam Soo Kim. Diff-TTS: A Denoising Diffusion Model for Text-to-Speech. In Proc. Interspeech 2021, pages 3605-3609, 2021. 1 +[18] Alexia Jolicoeur-Martineau, Remi Piche-Taillefer, Rémi Tachet des Combes, and Ioannis Mitliagkas. Adversarial score matching and improved sampling for image generation. arXiv preprint arXiv:2009.05475. 1 +[19] Haeyong Kang, Thang Vu, and Chang D. Yoo. Learning imbalanced datasets with maximum margin loss. In ICIP, pages 1269-1273, 2021. 1 +[20] Tero Karras, Miika Aittala, Timo Aila, and Samuli Laine. Elucidating the design space of diffusion-based generative models. arXiv preprint arXiv:2206.00364, 2022. 5, 7 +[21] Tero Karras, Miika Aittala, Janne Hellsten, Samuli Laine, Jaakko Lehtinen, and Timo Aila. Training generative adversarial networks with limited data. In NeurIPS, Red Hook, NY, USA, 2020. Curran Associates Inc. 1, 2, 5, 6, 7 +[22] Minsu Ko, Eunju Cha, Sungjoo Suh, Huijin Lee, Jae-Joon Han, Jinwoo Shin, and Bohyung Han. Self-supervised dense consistency regularization for image-to-image translation. In CVPR, pages 18280-18289, 2022. 2 +[23] Tuomas Kynkänniemi, Tero Karras, Samuli Laine, Jaakko Lehtinen, and Timo Aila. Improved precision and recall metric for assessing generative models. NeurIPS, 32, 2019. 5 +[24] Wen Li, Limin Wang, Wei Li, Eirikur Agustsson, and Luc Van Gool. Webvision database: Visual learning and understanding from web data. arXiv preprint arXiv:1708.02862, 2017. 1 +[25] Jialun Liu, Yifan Sun, Chuchu Han, Zhaopeng Dou, and Wenhui Li. Deep representation learning on long-tailed data: A learnable embedding augmentation perspective. In CVPR, pages 2967-2976, 2020. 2 +[26] Guanxiong Luo, Martin Heide, and Martin Uecker. MRI reconstruction via data driven markov chain with joint uncertainty estimation. CoRR, abs/2202.01479, 2022. 1 +[27] Dhruv Mahajan, Ross Girshick, Vignesh Ramanathan, Kaiming He, Manohar Paluri, Yixuan Li, Ashwin Bharambe, and Laurens van der Maaten. Exploring the limits of weakly supervised pretraining. In Vittorio Ferrari, Martial Hebert, Cristian Sminchisescu, and Yair Weiss, editors, ECCV, pages 185-201, Cham, 2018. Springer International Publishing. 1, 2, 5, 6 +[28] Chenlin Meng, Yang Song, Jiaming Song, Jiajun Wu, JunYan Zhu, and Stefano Ermon. Sdedit: Image synthesis and editing with stochastic differential equations. arXiv preprint arXiv:2108.01073, 2021. 1 +[29] Aditya Krishna Menon, Sadeep Jayasumana, Ankit Singh Rawat, Himanshu Jain, Andreas Veit, and Sanjiv Kumar. Long-tail learning via logit adjustment. In ICLR. OpenReview.net, 2021. 1, 2 +[30] Takeru Miyato, Toshiki Kataoka, Masanori Koyama, and Yuichi Yoshida. Spectral normalization for generative adversarial networks. In ICLR, 2018. 7 + +[31] Alexander Quinn Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. GLIDE: Towards photorealistic image generation and editing with text-guided diffusion models. In Kamalika Chaudhuri, Stefanie Jegelka, Le Song, Csaba Szepesvari, Gang Niu, and Sivan Sabato, editors, International Conference on Machine Learning (ICML), volume 162 of Proceedings of Machine Learning Research, pages 16784-16804. PMLR, 17-23 Jul 2022. 1, 2 +[32] Wanli Ouyang, Xiaogang Wang, Cong Zhang, and Xiaokang Yang. Factors in finetuning deep model for object detection with long-tail distribution. In CVPR, pages 864-873, 2016. 2 +[33] Yuntian Gu Quanlin Wu, Hang Ye. Guided diffusion model for adversarial purification from random noise. arXiv preprint arXiv:2206.10875, 2022. 1, 2 +[34] Harsh Rangwani, Naman Jaswani, Tejan Karmali, Varun Jampani, and R. Venkatesh Babu. Improving gans for long-tailed data through group spectral regularization. In Shai Avidan, Gabriel Brostow, Moustapha Cisse, Giovanni Maria Farinella, and Tal Hassner, editors, ECCV, pages 426-442, Cham, 2022. Springer Nature Switzerland. 1, 5, 7 +[35] Harsh Rangwani, Konda Reddy Mopuri, and R. Venkatesh Babu. Class balancing gan with a classifier in the loop. In Cassio de Campos and Marloes H. Maathuis, editors, Proceedings of the Thirty-Seventh Conference on Uncertainty in Artificial Intelligence, volume 161 of Proceedings of Machine Learning Research, pages 1618-1627. PMLR, 27-30 Jul 2021. 5, 7 +[36] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, pages 10684-10695, 2022. 2 +[37] Kevin Roth, Aurelien Lucchi, Sebastian Nowozin, and Thomas Hofmann. Stabilizing training of generative adversarial networks through regularization. In NeurIPS, page 2015-2025, Red Hook, NY, USA, 2017. Curran Associates Inc. 1, 2 +[38] Chitwan Sahara, William Chan, Huiwen Chang, Chris A. Lee, Jonathan Ho, Tim Salimans, David J. Fleet, and Mohammad Norouzi. Palette: Image-to-image diffusion models. ACM SIGGRAPH 2022 Conference Proceedings, 2022. 1 +[39] Mehdi S. M. Sajjadi, Olivier Bachem, Mario Lucic, Olivier Bousquet, and Sylvain Gelly. Assessing generative models via precision and recall. In NeurIPS, page 5234-5243, Red Hook, NY, USA, 2018. Curran Associates Inc. 5 +[40] Tim Salimans, Ian Goodfellow, Wojciech Zaremba, Vicki Cheung, Alec Radford, Xi Chen, and Xi Chen. Improved techniques for training gans. In D. Lee, M. Sugiyama, U. Luxburg, I. Guyon, and R. Garnett, editors, NeurIPS, volume 29. Curran Associates, Inc., 2016. 5 +[41] Vikash Sehwag, Caner Hazirbas, Albert Gordo, First Ozgenel, and Cristian Canton. Generating high fidelity data from low-density regions using diffusion models. In CVPR, pages 11492-11501, 2022. 2 +[42] Jascha Sohl-Dickstein, Eric A. Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised + +learning using nonequilibrium thermodynamics. arXiv preprint arXiv:1503.03585, 2015. 2 +[43] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models, 2021. 2, 7 +[44] Yang Song, Liyue Shen, Lei Xing, and Stefano Ermon. Solving inverse problems in medical imaging with score-based generative models. arXiv preprint arXiv:2111.08005, 2022. 1 +[45] Hung-Yu Tseng, Lu Jiang, Ce Liu, Ming-Hsuan Yang, and Weilong Yang. Regularizing generative adversarial networks under limited data. In CVPR, pages 7917-7927, 2021. 2, 7 +[46] Aaron van den Oord, Oriol Vinyals, and Koray Kavukcuoglu. Neural discrete representation learning. In NeurIPS, NIPS'17, page 6309-6318, Red Hook, NY, USA, 2017. Curran Associates Inc. 4 +[47] Jianfeng Wang, Thomas Lukasiewicz, Xiaolin Hu, Jianfei Cai, and Zhenghua Xu. Rsg: A simple but effective module for learning imbalanced datasets. In CVPR, pages 3783-3792, 2021. 2 +[48] Xiulong Yang, Sheng-Min Shih, Yinlin Fu, Xiaoting Zhao, and Shihao Ji. Your vit is secretly a hybrid discriminative-generative diffusion model. ArXiv, abs/2208.07791, 2022. 5, 7 +[49] Xi Yin, Xiang Yu, Kihyuk Sohn, Xiaoming Liu, and Manmohan Chandraker. Feature transfer learning for face recognition with under-represented data. In CVPR, pages 5697-5706, 2019. 2 +[50] Shengyu Zhao, Zhijian Liu, Ji Lin, Jun-Yan Zhu, and Song Han. Differentiable augmentation for data-efficient gan training. In NeurlPS, Red Hook, NY, USA, 2020. Curran Associates Inc. 1, 2, 5, 6 \ No newline at end of file diff --git a/2023/Class-Balancing Diffusion Models/images.zip b/2023/Class-Balancing Diffusion Models/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..47c1eb51637c332aa238c69713ff867be1f40042 --- /dev/null +++ b/2023/Class-Balancing Diffusion Models/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b87c78c4f62c45cf221ceaaf4934c37e0f79e44ff0cd64db09fd17bc0cc61131 +size 602967 diff --git a/2023/Class-Balancing Diffusion Models/layout.json b/2023/Class-Balancing Diffusion Models/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..a5ea034ed660ab72d19cb9d063bd39ddfb25dba4 --- /dev/null +++ b/2023/Class-Balancing Diffusion Models/layout.json @@ -0,0 +1,8934 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 192, + 103, + 402, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 192, + 103, + 402, + 121 + ], + "spans": [ + { + "bbox": [ + 192, + 103, + 402, + 121 + ], + "type": "text", + "content": "Class-Balancing Diffusion Models" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 58, + 149, + 122, + 165 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 149, + 122, + 165 + ], + "spans": [ + { + "bbox": [ + 58, + 149, + 122, + 165 + ], + "type": "text", + "content": "Yiming Qin" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 143, + 149, + 228, + 165 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 149, + 228, + 165 + ], + "spans": [ + { + "bbox": [ + 143, + 149, + 228, + 165 + ], + "type": "text", + "content": "Huangjie Zheng" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 250, + 149, + 345, + 165 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 250, + 149, + 345, + 165 + ], + "spans": [ + { + "bbox": [ + 250, + 149, + 345, + 165 + ], + "type": "text", + "content": "Jiangchao Yao" + }, + { + "bbox": [ + 250, + 149, + 345, + 165 + ], + "type": "inline_equation", + "content": "^{1,3}" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 365, + 149, + 449, + 165 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 365, + 149, + 449, + 165 + ], + "spans": [ + { + "bbox": [ + 365, + 149, + 449, + 165 + ], + "type": "text", + "content": "Mingyuan Zhou" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 473, + 149, + 533, + 165 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 473, + 149, + 533, + 165 + ], + "spans": [ + { + "bbox": [ + 473, + 149, + 533, + 165 + ], + "type": "text", + "content": "Ya Zhang" + }, + { + "bbox": [ + 473, + 149, + 533, + 165 + ], + "type": "inline_equation", + "content": "^{1,3}" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 147, + 181, + 445, + 194 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 147, + 181, + 445, + 194 + ], + "spans": [ + { + "bbox": [ + 147, + 181, + 445, + 194 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 147, + 181, + 445, + 194 + ], + "type": "text", + "content": "Cooperative Medianet Innovation Center, Shanghai Jiao-Tong University" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 239, + 194, + 354, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 239, + 194, + 354, + 205 + ], + "spans": [ + { + "bbox": [ + 239, + 194, + 354, + 205 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 239, + 194, + 354, + 205 + ], + "type": "text", + "content": "University of Texas, Austin" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 245, + 205, + 347, + 218 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 245, + 205, + 347, + 218 + ], + "spans": [ + { + "bbox": [ + 245, + 205, + 347, + 218 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 245, + 205, + 347, + 218 + ], + "type": "text", + "content": "Shanghai AI Laboratory" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 143, + 227, + 192, + 239 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 227, + 192, + 239 + ], + "spans": [ + { + "bbox": [ + 143, + 227, + 192, + 239 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 251, + 290, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 251, + 290, + 515 + ], + "spans": [ + { + "bbox": [ + 46, + 251, + 290, + 515 + ], + "type": "text", + "content": "Diffusion-based models have shown the merits of generating high-quality visual data while preserving better diversity in recent studies. However, such observation is only justified with curated data distribution, where the data samples are nicely pre-processed to be uniformly distributed in terms of their labels. In practice, a long-tailed data distribution appears more common and how diffusion models perform on such class-imbalanced data remains unknown. In this work, we first investigate this problem and observe significant degradation in both diversity and fidelity when the diffusion model is trained on datasets with class-imbalanced distributions. Especially in tail classes, the generations largely lose diversity and we observe severe mode-collapse issues. To tackle this problem, we set from the hypothesis that the data distribution is not class-balanced, and propose Class-Balancing Diffusion Models (CBDM) that are trained with a distribution adjustment regularizer as a solution. Experiments show that images generated by CBDM exhibit higher diversity and quality in both quantitative and qualitative ways. Our method benchmarked the generation results on CIFAR100/CIFAR100LT dataset and shows outstanding performance on the downstream recognition task." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 47, + 538, + 128, + 551 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 538, + 128, + 551 + ], + "spans": [ + { + "bbox": [ + 47, + 538, + 128, + 551 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 46, + 557, + 287, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 557, + 287, + 665 + ], + "spans": [ + { + "bbox": [ + 46, + 557, + 287, + 665 + ], + "type": "text", + "content": "In recent years, log-likelihood-based diffusion models have evolved rapidly and established new benchmarks on a range of generation tasks [1,7]. Based on them, researchers have been able to further control the model generation process and the generation quality. This improves the applications of generative models in numerous domains including text-image generation [31], image editing [28, 38], speech synthesis [17], medical imaging [26, 44], video generation [13] and adversarial learning [18, 33], etc." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 46, + 665, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 288, + 713 + ], + "type": "text", + "content": "Although diffusion models are known for the power of high fidelity and diversity in generation, most of the existing diffusion models are trained with the hypothesis that the data are uniformly distributed w.r.t. their labels. However," + } + ] + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 313, + 232, + 541, + 384 + ], + "blocks": [ + { + "bbox": [ + 313, + 232, + 541, + 384 + ], + "lines": [ + { + "bbox": [ + 313, + 232, + 541, + 384 + ], + "spans": [ + { + "bbox": [ + 313, + 232, + 541, + 384 + ], + "type": "image", + "image_path": "b52f1f93a407fa27ecdca458a5edea70ef052a41c50f7baf264ddbde5122540d.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 399, + 545, + 434 + ], + "lines": [ + { + "bbox": [ + 305, + 399, + 545, + 434 + ], + "spans": [ + { + "bbox": [ + 305, + 399, + 545, + 434 + ], + "type": "text", + "content": "Figure 1. Generation degrades along with class frequency. Semantics of generated images become less recognizable when class frequency decreases, while the FID score increases significantly." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 459, + 546, + 663 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 459, + 546, + 663 + ], + "spans": [ + { + "bbox": [ + 304, + 459, + 546, + 663 + ], + "type": "text", + "content": "in the real world, the distribution is often very skewed. Especially for many domain-specific generation tasks such as medical images [14], fine-grained dataset for taxology [15] and data grabbed from the web [24], it is difficult to collect large amounts of data for each class equally, and the size of the training set for head and tail categories can differ by a factor of hundred or more. For such datasets, unconditional diffusion models tend to produce a significant portion of low-quality images. Conditional models, as shown in Figure 1, generate head class images with satisfying performance, while conversely the generated images on tail classes are very likely to show unrecognizable semantics. Concerning training generative models with limited data, there already exist several methods [21, 37, 50] based on GAN models [2]. However, quite few studies examine the impact of imbalanced class distribution [34] especially on diffusion models, which is practical yet under-explored." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "type": "text", + "content": "Our work first introduces diffusion models to imbalance generation tasks on several long-tailed datasets [19], and then build some straightforward baselines according to the common methods used in long-tailed recognition [27, 29]." + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "18434" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 289, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 289, + 312 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 289, + 312 + ], + "type": "text", + "content": "To overcome the potential degeneration induced by the skewed distribution, we propose a novel Class-Balancing Diffusion Model (CBDM). Theoretically, CBDM resorts to adjusting the conditional transfer probability during sampling in order to implicitly force generated images to have a balanced prior distribution during every sampling step. Technically, the adjusted transfer probability of CBDM results in an additional MSE-form loss for a conditional diffusion model, which functions as a regularizer. Intuitively, this loss augments the similarity of generated images conditioned on different classes, and turns out to be an effective approach to transfer common information from head classes to tail classes without hurting the model's expressiveness on head classes. CBDM can be implemented within several flines of codes, and its lighter version admits fine-tuning an existing conditional model. We conducted extensive experiments on CIFAR10/CIFAR100 and their corresponding long-tailed dataset to show the promise of CBDM over existing state-of-the-art methods. In a nutshell, the contributions of this work can be summarized as follows:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 59, + 321, + 287, + 493 + ], + "type": "list", + "angle": 0, + "index": 4, + "blocks": [ + { + "bbox": [ + 59, + 321, + 286, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 321, + 286, + 357 + ], + "spans": [ + { + "bbox": [ + 59, + 321, + 286, + 357 + ], + "type": "text", + "content": "- We identify the severe degeneration problem of diffusion models in long-tailed generation tasks and benchmark some straightforward baselines in this direction." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 59, + 365, + 287, + 414 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 365, + 287, + 414 + ], + "spans": [ + { + "bbox": [ + 59, + 365, + 287, + 414 + ], + "type": "text", + "content": "- We propose a new perspective to handle the generation quality collapse on tail classes, and derive a novel Class-Balancing Diffusion Model, which is effective and lightweight as a regularizer to existing methods." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 59, + 421, + 287, + 493 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 421, + 287, + 493 + ], + "spans": [ + { + "bbox": [ + 59, + 421, + 287, + 493 + ], + "type": "text", + "content": "- We validate that CBDM is capable of generating more diverse images with convincing fidelity, especially for datasets with large number of categories. In addition, CBDM is robust with accelerating algorithms such as DDIM [43], and can be transplanted to different conditional diffusion-based backbones easily." + } + ] + } + ], + "index": 3 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 47, + 514, + 138, + 526 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 514, + 138, + 526 + ], + "spans": [ + { + "bbox": [ + 47, + 514, + 138, + 526 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 533, + 289, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 533, + 289, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 533, + 289, + 713 + ], + "type": "text", + "content": "Diffusion models Diffusion models are recently proposed generative models [42] based on non-equilibrium thermodynamics. Conditional diffusion models [7] encode label information into the generation process and improve largely the generation performance. The guidance structure proposed in [7] makes it possible to control the generation process through an external module. Based on a similar intuition, researchers arrive to realize diverse functions, such as guided adversarial purification [33], few-shot generation [9] and so on [36, 41]. The drawback of classifier guidance (noted as CG) [7] lies in its requirement of training another auxiliary classifier. To address the issue, classifier-free guidance (noted as CFG) [12] proposed a mechanism that uses the generator itself to express the class guidance information. CFG is proved to be a resource efficient method" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 72, + 545, + 109 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 545, + 109 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 545, + 109 + ], + "type": "text", + "content": "and achieves outstanding performance on large models [31]. Moreover, CFG only requires to add one line in training, which can be easily transplanted on different models." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 118, + 547, + 358 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 118, + 547, + 358 + ], + "spans": [ + { + "bbox": [ + 304, + 118, + 547, + 358 + ], + "type": "text", + "content": "Long-tail recognition The problem of long-tailed distribution is a common dilemma in machine learning and have been widely explored in the area of discriminative models, i.e., long-tailed recognition. There mainly exist three paradigms in this domain, namely Class Re-balancing [27, 29], Information Augmentation [25,47,49] and Module Improvement [16, 32]. Among them, Class Re-balancing provides the best explainability, and its most common practice is re-sampling [27]. Thereafter, stemming from modifying the objective function from the global error rate to the class average one, [29] propose logit adjustment which has shown an impressive performance. Another effective method is based on Information Augmentation, which uses head class feature information to augment tail classes [5, 25, 47, 49]. However, discriminative models map data from higher to lower dimensions, while generative models map images from lower to higher dimensions. Thus, the mechanism of class rebalancing between them may be completely different and how to design the balancing method remains under-explored." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 369, + 547, + 536 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 369, + 547, + 536 + ], + "spans": [ + { + "bbox": [ + 304, + 369, + 547, + 536 + ], + "type": "text", + "content": "Generative models based on limited data GAN [2, 6] is the most dominant model in the field of image synthesis in recent years. Given the requirement of large-scale data to train the generative models, a part of researchers focus on improving the performance of GAN models under the small datasets. To address the overfitting issue of the discriminator, a number of regularization methods [22, 37, 45] has been proposed. An alternative solution is Data Augmentation. As augmentation information is very prone to leak to the generator, researchers proposed improved augmentation strategies such as Differentiable Augmentation (DiffAug) [50] and Adaptive Augmentation (ADA) [21]. These augmentation ways can be transferred to diffusion models as baselines to help the generation of tail-class samples." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 552, + 362, + 564 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 552, + 362, + 564 + ], + "spans": [ + { + "bbox": [ + 306, + 552, + 362, + 564 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 573, + 545, + 634 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 573, + 545, + 634 + ], + "spans": [ + { + "bbox": [ + 304, + 573, + 545, + 634 + ], + "type": "text", + "content": "In this section, we first give the basic notations following the classical DDPM model, and then introduce the class-imbalanced generation setting. In the third part, we introduce our CBDM algorithm and present its training details including implementation and hyper-parameter settings." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 645, + 385, + 658 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 645, + 385, + 658 + ], + "spans": [ + { + "bbox": [ + 306, + 645, + 385, + 658 + ], + "type": "text", + "content": "3.1. Preliminary" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 665, + 545, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 545, + 715 + ], + "type": "text", + "content": "Diffusion models leverage a pre-defined forward process in training, where a clean image distribution " + }, + { + "bbox": [ + 304, + 665, + 545, + 715 + ], + "type": "inline_equation", + "content": "q(\\pmb{x}_0)" + }, + { + "bbox": [ + 304, + 665, + 545, + 715 + ], + "type": "text", + "content": " can be corrupted to a noisy distribution " + }, + { + "bbox": [ + 304, + 665, + 545, + 715 + ], + "type": "inline_equation", + "content": "q(\\pmb{x}_t|\\pmb{x}_0)" + }, + { + "bbox": [ + 304, + 665, + 545, + 715 + ], + "type": "text", + "content": " at a specified timestep " + }, + { + "bbox": [ + 304, + 665, + 545, + 715 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 665, + 545, + 715 + ], + "type": "text", + "content": ". Given a pre-defined variance schedule " + }, + { + "bbox": [ + 304, + 665, + 545, + 715 + ], + "type": "inline_equation", + "content": "\\{\\beta_t\\}_{1:T}" + }, + { + "bbox": [ + 304, + 665, + 545, + 715 + ], + "type": "text", + "content": "," + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18435" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 52, + 79, + 391, + 217 + ], + "blocks": [ + { + "bbox": [ + 52, + 79, + 391, + 217 + ], + "lines": [ + { + "bbox": [ + 52, + 79, + 391, + 217 + ], + "spans": [ + { + "bbox": [ + 52, + 79, + 391, + 217 + ], + "type": "image", + "image_path": "93689ce378ac70f873db28d10f3497b9583aa537a599b4fc905f992fe5c9857e.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 224, + 546, + 258 + ], + "lines": [ + { + "bbox": [ + 46, + 224, + 546, + 258 + ], + "spans": [ + { + "bbox": [ + 46, + 224, + 546, + 258 + ], + "type": "text", + "content": "Figure 2. Algorithm (left) and generation (right) visualization. In the left figure, we show that an extra regularization loss " + }, + { + "bbox": [ + 46, + 224, + 546, + 258 + ], + "type": "inline_equation", + "content": "gL_{r}" + }, + { + "bbox": [ + 46, + 224, + 546, + 258 + ], + "type": "text", + "content": " proportional to the diffusion step " + }, + { + "bbox": [ + 46, + 224, + 546, + 258 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 224, + 546, + 258 + ], + "type": "text", + "content": " is added during training. This loss function pushes the sampling distribution (curves on the surface) to wider region while preventing it to be excessively distorted compared to the ground truth distribution (gradient color on the background)." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 402, + 95, + 544, + 216 + ], + "blocks": [ + { + "bbox": [ + 402, + 95, + 544, + 216 + ], + "lines": [ + { + "bbox": [ + 402, + 95, + 544, + 216 + ], + "spans": [ + { + "bbox": [ + 402, + 95, + 544, + 216 + ], + "type": "image", + "image_path": "5020ec0e967463607da0afac543321a5b90bdf43930986ac8b6b71e2fd04bdb2.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 279, + 259, + 292 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 279, + 259, + 292 + ], + "spans": [ + { + "bbox": [ + 47, + 279, + 259, + 292 + ], + "type": "text", + "content": "the noisy distribution at any intermediate timestep is" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 300, + 279, + 332 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 300, + 279, + 332 + ], + "spans": [ + { + "bbox": [ + 53, + 300, + 279, + 332 + ], + "type": "interline_equation", + "content": "q \\left(\\boldsymbol {x} _ {t} \\mid \\boldsymbol {x} _ {0}\\right) = \\mathcal {N} \\left(\\sqrt {\\bar {\\alpha} _ {t}} \\boldsymbol {x} _ {0}, \\left(1 - \\bar {\\alpha} _ {t}\\right) \\boldsymbol {I}\\right); \\quad \\bar {\\alpha} _ {t} = \\prod_ {i = 1} ^ {t} \\left(1 - \\beta_ {i}\\right).", + "image_path": "bbb4a5a63282cc5d08998b85f1745b77081fea7ae37ebe4ff057da65faa58a71.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 340, + 287, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 340, + 287, + 376 + ], + "spans": [ + { + "bbox": [ + 47, + 340, + 287, + 376 + ], + "type": "text", + "content": "To reverse such forward process, a generative model " + }, + { + "bbox": [ + 47, + 340, + 287, + 376 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 47, + 340, + 287, + 376 + ], + "type": "text", + "content": " learns to estimate the analytical true posterior in order to recover " + }, + { + "bbox": [ + 47, + 340, + 287, + 376 + ], + "type": "inline_equation", + "content": "x_{t - 1}" + }, + { + "bbox": [ + 47, + 340, + 287, + 376 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 47, + 340, + 287, + 376 + ], + "type": "inline_equation", + "content": "x_{t}" + }, + { + "bbox": [ + 47, + 340, + 287, + 376 + ], + "type": "text", + "content": " as follows:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 385, + 285, + 403 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 385, + 285, + 403 + ], + "spans": [ + { + "bbox": [ + 48, + 385, + 285, + 403 + ], + "type": "interline_equation", + "content": "\\min _ {\\theta} D _ {\\mathrm {K L}} [ q (\\pmb {x} _ {t - 1} | \\pmb {x} _ {t}, \\pmb {x} _ {0}) | | p _ {\\theta} (\\pmb {x} _ {t - 1} | \\pmb {x} _ {t}) ]; \\forall t \\in \\{1, \\dots , T \\},", + "image_path": "c289d5056c4481c25d0fda3221fbb0df05047b789d068d02165d250e6c2df92a.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 411, + 287, + 436 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 411, + 287, + 436 + ], + "spans": [ + { + "bbox": [ + 47, + 411, + 287, + 436 + ], + "type": "text", + "content": "and such an objective can be reduced to a simple denoising estimation loss [11]:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 445, + 286, + 468 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 445, + 286, + 468 + ], + "spans": [ + { + "bbox": [ + 47, + 445, + 286, + 468 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {D D P M}} = \\mathbb {E} _ {t, \\boldsymbol {x} _ {0} \\sim q (\\boldsymbol {x} _ {0}), \\epsilon \\sim \\mathcal {N} (\\boldsymbol {0}, I)} \\left[ \\| \\boldsymbol {\\epsilon} - \\boldsymbol {\\epsilon} _ {\\theta} \\left(\\sqrt {\\bar {\\alpha}} \\boldsymbol {x} _ {0} + \\sqrt {1 - \\bar {\\alpha} _ {t}} \\boldsymbol {\\epsilon}, t\\right) \\| ^ {2} \\right] \\tag {1}", + "image_path": "a7911785bfc06c5cddedf1d30f070a15df63066caa6d9af0b58cb4e994c2a45b.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 469, + 287, + 553 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 469, + 287, + 553 + ], + "spans": [ + { + "bbox": [ + 47, + 469, + 287, + 553 + ], + "type": "text", + "content": "For the case where label information is available, the model is trained to estimate the noise as above in both conditional cases " + }, + { + "bbox": [ + 47, + 469, + 287, + 553 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}(\\pmb{x}_t,y,t)" + }, + { + "bbox": [ + 47, + 469, + 287, + 553 + ], + "type": "text", + "content": " with data-label pairs " + }, + { + "bbox": [ + 47, + 469, + 287, + 553 + ], + "type": "inline_equation", + "content": "(x_0,y)" + }, + { + "bbox": [ + 47, + 469, + 287, + 553 + ], + "type": "text", + "content": " and unconditional case " + }, + { + "bbox": [ + 47, + 469, + 287, + 553 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}(\\pmb{x}_t,t)" + }, + { + "bbox": [ + 47, + 469, + 287, + 553 + ], + "type": "text", + "content": ". In the sampling, the label-guided model estimates the noise with a linear interpolation " + }, + { + "bbox": [ + 47, + 469, + 287, + 553 + ], + "type": "inline_equation", + "content": "\\hat{\\epsilon} = (1 + \\omega)\\pmb{\\epsilon}_{\\theta}(\\pmb{x}_t,y,t) - \\omega \\pmb{\\epsilon}_{\\theta}(\\pmb{x}_t,t)" + }, + { + "bbox": [ + 47, + 469, + 287, + 553 + ], + "type": "text", + "content": " to recover " + }, + { + "bbox": [ + 47, + 469, + 287, + 553 + ], + "type": "inline_equation", + "content": "\\pmb{x}_{t-1}" + }, + { + "bbox": [ + 47, + 469, + 287, + 553 + ], + "type": "text", + "content": ", which is often referred as Classifier-Free Guidance (CFG) [12]." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 560, + 228, + 573 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 560, + 228, + 573 + ], + "spans": [ + { + "bbox": [ + 47, + 560, + 228, + 573 + ], + "type": "text", + "content": "3.2. Class-Balancing Diffusion Models" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 578, + 287, + 650 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 578, + 287, + 650 + ], + "spans": [ + { + "bbox": [ + 47, + 578, + 287, + 650 + ], + "type": "text", + "content": "Current diffusion models assume the data distribution to be uniform in class, and thus equally treat samples in the training stage. However, based on our observation, such training strategy leads to degradation in generation quality. Below, we provide an analysis that motivates our Class-Balancing Diffusion Models (CBDM)." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 651, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 651, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 651, + 287, + 714 + ], + "type": "text", + "content": "Suppose " + }, + { + "bbox": [ + 46, + 651, + 287, + 714 + ], + "type": "inline_equation", + "content": "q(\\pmb{x}, y)" + }, + { + "bbox": [ + 46, + 651, + 287, + 714 + ], + "type": "text", + "content": " is the data distribution that we need to match with the joint distribution " + }, + { + "bbox": [ + 46, + 651, + 287, + 714 + ], + "type": "inline_equation", + "content": "p_{\\theta}(\\pmb{x}, y)" + }, + { + "bbox": [ + 46, + 651, + 287, + 714 + ], + "type": "text", + "content": " predicted by a generative model. We analyze their difference from the density ratio " + }, + { + "bbox": [ + 46, + 651, + 287, + 714 + ], + "type": "inline_equation", + "content": "r = \\frac{q(\\pmb{x}, y)}{p_{\\theta}(\\pmb{x}, y)} = \\frac{q(\\pmb{x} \\mid y)}{p_{\\theta}(\\pmb{x} \\mid y)} \\cdot \\frac{q(y)}{p_{\\theta}(y)}" + }, + { + "bbox": [ + 46, + 651, + 287, + 714 + ], + "type": "text", + "content": ". When the true label distribution " + }, + { + "bbox": [ + 46, + 651, + 287, + 714 + ], + "type": "inline_equation", + "content": "q(y)" + }, + { + "bbox": [ + 46, + 651, + 287, + 714 + ], + "type": "text", + "content": " is the same as the prior " + }, + { + "bbox": [ + 46, + 651, + 287, + 714 + ], + "type": "inline_equation", + "content": "p_{\\theta}(y)" + }, + { + "bbox": [ + 46, + 651, + 287, + 714 + ], + "type": "text", + "content": "," + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 279, + 545, + 425 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 279, + 545, + 425 + ], + "spans": [ + { + "bbox": [ + 304, + 279, + 545, + 425 + ], + "type": "text", + "content": "which is usually assumed to be uniform, the density ratio " + }, + { + "bbox": [ + 304, + 279, + 545, + 425 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 304, + 279, + 545, + 425 + ], + "type": "text", + "content": " is reduced the conditional term to learn a conditional model " + }, + { + "bbox": [ + 304, + 279, + 545, + 425 + ], + "type": "inline_equation", + "content": "p_{\\theta}(\\boldsymbol{x} \\mid y)" + }, + { + "bbox": [ + 304, + 279, + 545, + 425 + ], + "type": "text", + "content": ". However, when such a hypothesis is violated, for head classes, " + }, + { + "bbox": [ + 304, + 279, + 545, + 425 + ], + "type": "inline_equation", + "content": "\\frac{q(y)}{p_{\\theta}(y)}" + }, + { + "bbox": [ + 304, + 279, + 545, + 425 + ], + "type": "text", + "content": " would result in a larger weight that makes the model biased and hurt tail classes, and vice versa. Empirically, we observe that the generation degrades more on tail classes, as illustrated in Figure 1. Moreover, as shown in Figure 2a, compared to head classes, DDPM cannot well capture the tail-class data distribution and the mode is poorly covered during the sampling process. As a result, generations of tail classes often have poor quality and diversity, shown in Figure 2b." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 426, + 545, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 426, + 545, + 533 + ], + "spans": [ + { + "bbox": [ + 304, + 426, + 545, + 533 + ], + "type": "text", + "content": "To tackle this issue, the most intuitive approach lies in adjusting the prior label distribution through a class balanced re-sampling. However, such abrupt adjustment easily leads to negative improvement in experiments. The step-by-step sampling nature of diffusion models provides another aspect to adjust this distribution more softly. In this spirit, we propose to calibrate the learning process through the conditional transfer probability " + }, + { + "bbox": [ + 304, + 426, + 545, + 533 + ], + "type": "inline_equation", + "content": "p_{\\theta}(\\pmb{x}_{t-1}|\\pmb{x}_t,y)" + }, + { + "bbox": [ + 304, + 426, + 545, + 533 + ], + "type": "text", + "content": " when there exists a gap between the class distribution and the prior." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 533, + 545, + 596 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 533, + 545, + 596 + ], + "spans": [ + { + "bbox": [ + 304, + 533, + 545, + 596 + ], + "type": "text", + "content": "Let " + }, + { + "bbox": [ + 304, + 533, + 545, + 596 + ], + "type": "inline_equation", + "content": "p_{\\theta}^{\\star}\\left(\\boldsymbol{x}_{t-1} \\mid \\boldsymbol{x}_t, y\\right)" + }, + { + "bbox": [ + 304, + 533, + 545, + 596 + ], + "type": "text", + "content": " be the optimum trained in the case that " + }, + { + "bbox": [ + 304, + 533, + 545, + 596 + ], + "type": "inline_equation", + "content": "\\frac{q(y)}{p_{\\theta}(y)}" + }, + { + "bbox": [ + 304, + 533, + 545, + 596 + ], + "type": "text", + "content": " is correctly estimated, and " + }, + { + "bbox": [ + 304, + 533, + 545, + 596 + ], + "type": "inline_equation", + "content": "p_{\\theta}\\left(\\boldsymbol{x}_{t-1} \\mid \\boldsymbol{x}_t, y\\right)" + }, + { + "bbox": [ + 304, + 533, + 545, + 596 + ], + "type": "text", + "content": " be the one trained in a class-imbalanced case. The relation between such two generative distributions can be described as the following proposition." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 602, + 545, + 650 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 602, + 545, + 650 + ], + "spans": [ + { + "bbox": [ + 304, + 602, + 545, + 650 + ], + "type": "text", + "content": "Proposition 1. When training a diffusion model parameterized with " + }, + { + "bbox": [ + 304, + 602, + 545, + 650 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 304, + 602, + 545, + 650 + ], + "type": "text", + "content": " on a class-imbalanced dataset, its conditional reverse distribution " + }, + { + "bbox": [ + 304, + 602, + 545, + 650 + ], + "type": "inline_equation", + "content": "p_{\\theta}\\left(\\boldsymbol{x}_{t-1}|\\boldsymbol{x}_t,y\\right)" + }, + { + "bbox": [ + 304, + 602, + 545, + 650 + ], + "type": "text", + "content": " can be corrected with an adjustment schema:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 315, + 656, + 545, + 683 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 656, + 545, + 683 + ], + "spans": [ + { + "bbox": [ + 315, + 656, + 545, + 683 + ], + "type": "interline_equation", + "content": "p _ {\\theta} ^ {\\star} \\left(\\boldsymbol {x} _ {t - 1} \\mid \\boldsymbol {x} _ {t}, y\\right) = p _ {\\theta} \\left(\\boldsymbol {x} _ {t - 1} \\mid \\boldsymbol {x} _ {t}, y\\right) \\frac {p _ {\\theta} \\left(\\boldsymbol {x} _ {t - 1}\\right)}{p _ {\\theta} ^ {\\star} \\left(\\boldsymbol {x} _ {t - 1}\\right)} \\frac {q ^ {\\star} \\left(\\boldsymbol {x} _ {t}\\right)}{q \\left(\\boldsymbol {x} _ {t}\\right)} \\tag {2}", + "image_path": "2f03740b3f9a4c6c29dfb573ab1110bfd8cec5b0ef9d450c20cb7bc800a5b30e.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 318, + 701, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 701, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 318, + 701, + 545, + 713 + ], + "type": "text", + "content": "The proposition above shows that, when trained on a" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18436" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 198 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 198 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 198 + ], + "type": "text", + "content": "class-imbalanced dataset, a diffusion model can still approach the true data distribution by applying a distribution adjustment schema " + }, + { + "bbox": [ + 46, + 72, + 289, + 198 + ], + "type": "inline_equation", + "content": "\\frac{p_{\\theta}(\\boldsymbol{x}_{t-1})}{p_{\\theta}^*(\\boldsymbol{x}_{t-1})}\\frac{q^*(\\boldsymbol{x}_t)}{q(\\boldsymbol{x}_t)}" + }, + { + "bbox": [ + 46, + 72, + 289, + 198 + ], + "type": "text", + "content": " at every step " + }, + { + "bbox": [ + 46, + 72, + 289, + 198 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 72, + 289, + 198 + ], + "type": "text", + "content": ". However, approximating this schema is not feasible at every sampling step, so CBDM incorporates it into the training loss function to achieve an equivalent objective, and thus gets rid of the model-free part " + }, + { + "bbox": [ + 46, + 72, + 289, + 198 + ], + "type": "inline_equation", + "content": "\\frac{q^*(\\boldsymbol{x}_t)}{q(\\boldsymbol{x}_t)}" + }, + { + "bbox": [ + 46, + 72, + 289, + 198 + ], + "type": "text", + "content": ". By further decomposing " + }, + { + "bbox": [ + 46, + 72, + 289, + 198 + ], + "type": "inline_equation", + "content": "p_{\\theta}(\\boldsymbol{x}_{t-1})" + }, + { + "bbox": [ + 46, + 72, + 289, + 198 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 72, + 289, + 198 + ], + "type": "inline_equation", + "content": "p_{\\theta}^*(\\boldsymbol{x}_{t-1})" + }, + { + "bbox": [ + 46, + 72, + 289, + 198 + ], + "type": "text", + "content": " to the expectation of the conditional probability " + }, + { + "bbox": [ + 46, + 72, + 289, + 198 + ], + "type": "inline_equation", + "content": "p_{\\theta}^*(\\boldsymbol{x}_{t-1}|\\boldsymbol{x}_{t:T},y)" + }, + { + "bbox": [ + 46, + 72, + 289, + 198 + ], + "type": "text", + "content": ", we present a upper bound to approximate this probability in Proposition 2." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 205, + 288, + 243 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 205, + 288, + 243 + ], + "spans": [ + { + "bbox": [ + 46, + 205, + 288, + 243 + ], + "type": "text", + "content": "Proposition 2. For the adjusted loss " + }, + { + "bbox": [ + 46, + 205, + 288, + 243 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{DM}^{\\star} = \\sum_{t=1}^{T} \\mathcal{L}_{t-1}^{\\star}" + }, + { + "bbox": [ + 46, + 205, + 288, + 243 + ], + "type": "text", + "content": ", an upper-bound of the target training objective to calibrate at timestep " + }, + { + "bbox": [ + 46, + 205, + 288, + 243 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 205, + 288, + 243 + ], + "type": "text", + "content": " (i.e., " + }, + { + "bbox": [ + 46, + 205, + 288, + 243 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{t-1}^{\\star}" + }, + { + "bbox": [ + 46, + 205, + 288, + 243 + ], + "type": "text", + "content": ") can be derived as:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 252, + 287, + 345 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 252, + 287, + 345 + ], + "spans": [ + { + "bbox": [ + 47, + 252, + 287, + 345 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\sum_ {t \\geq 1} \\mathcal {L} _ {t - 1} ^ {\\star} = \\sum_ {t \\geq 1} D _ {\\mathrm {K L}} [ q (\\pmb {x} _ {t - 1} | \\pmb {x} _ {t}, \\pmb {x} _ {0}) | | p _ {\\theta} ^ {\\star} (\\pmb {x} _ {t - 1} | \\pmb {x} _ {t}, y) ] \\\\ \\leq \\sum_ {t \\geq 1} [ \\underbrace {D _ {\\mathrm {K L}} [ q (\\boldsymbol {x} _ {t - 1} | \\boldsymbol {x} _ {t} , \\boldsymbol {x} _ {0}) | | p _ {\\theta} (\\boldsymbol {x} _ {t - 1} | \\boldsymbol {x} _ {t} , y) ]} _ {\\text {D i f f u s i o n m o d e l l o s s} \\mathcal {L} _ {D M}} \\\\ + \\underbrace {t \\mathbb {E} _ {y ^ {\\prime} \\sim q _ {y} ^ {\\star}} [ D _ {\\mathrm {K L}} [ p _ {\\theta} (\\boldsymbol {x} _ {t - 1} | \\boldsymbol {x} _ {t}) | | p _ {\\theta} (\\boldsymbol {x} _ {t - 1} | \\boldsymbol {x} _ {t} , y ^ {\\prime}) ] ] ]} _ {D i s t r i b u t i o n a d j u s t m e n t l o s s \\mathcal {L} _ {r}}, \\\\ \\end{array}", + "image_path": "6e6b97ce8b1844e67e0692bd7dd3a6226c0cefe5567243fcaca6dd6f748c04e0.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 367, + 287, + 499 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 367, + 287, + 499 + ], + "spans": [ + { + "bbox": [ + 46, + 367, + 287, + 499 + ], + "type": "text", + "content": "The upper bound in the above proposition can be considered as two parts. The first term " + }, + { + "bbox": [ + 46, + 367, + 287, + 499 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{DM}}" + }, + { + "bbox": [ + 46, + 367, + 287, + 499 + ], + "type": "text", + "content": " corresponds to an ordinary DDPM loss [11] e.g., Eqn. (1), and the second loss " + }, + { + "bbox": [ + 46, + 367, + 287, + 499 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{r}}" + }, + { + "bbox": [ + 46, + 367, + 287, + 499 + ], + "type": "text", + "content": " is used to adjust the distribution as a regularization term. Roughly speaking, " + }, + { + "bbox": [ + 46, + 367, + 287, + 499 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{r}}" + }, + { + "bbox": [ + 46, + 367, + 287, + 499 + ], + "type": "text", + "content": " increases the similarity between the model's output and a random target class. Thus, it reduces the risk of overfitting on the head classes, and enlarges the generation diversity for tail class through knowledge obtained from other classes. When " + }, + { + "bbox": [ + 46, + 367, + 287, + 499 + ], + "type": "inline_equation", + "content": "q_{\\mathcal{Y}}^{\\star}" + }, + { + "bbox": [ + 46, + 367, + 287, + 499 + ], + "type": "text", + "content": " is less longtailed than the dataset, this loss also increases the probability for underrepresented tail samples to be chosen during training." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 506, + 160, + 520 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 506, + 160, + 520 + ], + "spans": [ + { + "bbox": [ + 47, + 506, + 160, + 520 + ], + "type": "text", + "content": "3.3. Training algorithm" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 525, + 287, + 609 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 525, + 287, + 609 + ], + "spans": [ + { + "bbox": [ + 46, + 525, + 287, + 609 + ], + "type": "text", + "content": "The detailed training algorithm of CBDM is presented in Alg. (1). In the algorithm, we reduce the distribution adjustment loss " + }, + { + "bbox": [ + 46, + 525, + 287, + 609 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_r" + }, + { + "bbox": [ + 46, + 525, + 287, + 609 + ], + "type": "text", + "content": " as a square error loss with Monte-Carlo samples as indicated by Eqn.(3), where " + }, + { + "bbox": [ + 46, + 525, + 287, + 609 + ], + "type": "inline_equation", + "content": "\\mathcal{V}" + }, + { + "bbox": [ + 46, + 525, + 287, + 609 + ], + "type": "text", + "content": " is a set of samples that drawn following the distribution " + }, + { + "bbox": [ + 46, + 525, + 287, + 609 + ], + "type": "inline_equation", + "content": "q_{\\mathcal{V}}^{\\star}" + }, + { + "bbox": [ + 46, + 525, + 287, + 609 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 525, + 287, + 609 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 46, + 525, + 287, + 609 + ], + "type": "text", + "content": " denotes the image label. Note for CFG [12], there is a fixed probability (usually " + }, + { + "bbox": [ + 46, + 525, + 287, + 609 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 46, + 525, + 287, + 609 + ], + "type": "text", + "content": ") to drop the condition, i.e., " + }, + { + "bbox": [ + 46, + 525, + 287, + 609 + ], + "type": "inline_equation", + "content": "y = \\text{None}" + }, + { + "bbox": [ + 46, + 525, + 287, + 609 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 627, + 287, + 657 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 627, + 287, + 657 + ], + "spans": [ + { + "bbox": [ + 54, + 627, + 287, + 657 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {r}} \\left(\\boldsymbol {x} _ {t}, y, t\\right) = \\frac {1}{| \\mathcal {Y} |} \\sum_ {y ^ {\\prime} \\in \\mathcal {Y}} [ t \\| \\epsilon_ {\\theta} \\left(\\boldsymbol {x} _ {t}, y\\right) - \\epsilon_ {\\theta} \\left(\\boldsymbol {x} _ {t}, y ^ {\\prime}\\right) \\| ^ {2} ], \\tag {3}", + "image_path": "d14198bce12154159299b4b9260e848a241b1b9c5b2bc11f955b38f3e285cfd2.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "type": "text", + "content": "For the implementation in practice, CBDM can be plugged into any existing conditional diffusion models by adopting their model architecture and adjusting the training loss " + }, + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{DM}}" + }, + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "type": "text", + "content": " following lines 6 - 11. Specifically, the choice" + } + ] + } + ], + "index": 7 + }, + { + "type": "code", + "bbox": [ + 310, + 87, + 545, + 252 + ], + "blocks": [ + { + "bbox": [ + 307, + 72, + 484, + 84 + ], + "lines": [ + { + "bbox": [ + 307, + 72, + 484, + 84 + ], + "spans": [ + { + "bbox": [ + 307, + 72, + 484, + 84 + ], + "type": "text", + "content": "Algorithm 1 Training algorithm of CBDM." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "code_caption" + }, + { + "bbox": [ + 310, + 87, + 545, + 252 + ], + "lines": [ + { + "bbox": [ + 310, + 87, + 545, + 252 + ], + "spans": [ + { + "bbox": [ + 310, + 87, + 545, + 252 + ], + "type": "text", + "content": "1: for Every batch of size N do \n2: for " + }, + { + "bbox": [ + 310, + 87, + 545, + 252 + ], + "type": "inline_equation", + "content": "(\\pmb{x}_0^{(i)}, y^{(i)})" + }, + { + "bbox": [ + 310, + 87, + 545, + 252 + ], + "type": "text", + "content": " in this batch do \n3: Sample " + }, + { + "bbox": [ + 310, + 87, + 545, + 252 + ], + "type": "inline_equation", + "content": "\\pmb{\\epsilon}^{(i)} \\sim \\mathcal{N}(\\pmb{0}, \\pmb{I}), t \\sim \\mathcal{U}(\\{0, 1, \\dots, T\\})" + }, + { + "bbox": [ + 310, + 87, + 545, + 252 + ], + "type": "text", + "content": " \n4: " + }, + { + "bbox": [ + 310, + 87, + 545, + 252 + ], + "type": "inline_equation", + "content": "x_{t}^{(i)} = \\sqrt{\\bar{\\alpha}_{t}} x_{0}^{(i)} + \\sqrt{1 - \\bar{\\alpha}_{t}} \\epsilon^{(i)}" + }, + { + "bbox": [ + 310, + 87, + 545, + 252 + ], + "type": "text", + "content": " \n5: Calculate " + }, + { + "bbox": [ + 310, + 87, + 545, + 252 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{DM}} = \\| \\pmb{\\epsilon}^{(i)} - \\pmb{\\epsilon}_{\\theta}(\\pmb{x}_{t}^{(i)}, y^{(i)}) \\|^{2}" + }, + { + "bbox": [ + 310, + 87, + 545, + 252 + ], + "type": "text", + "content": " \n6: Sample " + }, + { + "bbox": [ + 310, + 87, + 545, + 252 + ], + "type": "inline_equation", + "content": "y'(i)" + }, + { + "bbox": [ + 310, + 87, + 545, + 252 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 310, + 87, + 545, + 252 + ], + "type": "inline_equation", + "content": "q_{\\gamma}^{\\star}" + }, + { + "bbox": [ + 310, + 87, + 545, + 252 + ], + "type": "text", + "content": " \n7: Calculate the first regularization term \n8: " + }, + { + "bbox": [ + 310, + 87, + 545, + 252 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{r} = t\\tau \\| \\pmb{\\epsilon}_{\\theta}(\\pmb{x}_{t}^{(i)}, y^{(i)}) - \\mathrm{sg}(\\pmb{\\epsilon}_{\\theta}(\\pmb{x}_{t}^{(i)}, y'^{(i)})) \\|^{2}" + }, + { + "bbox": [ + 310, + 87, + 545, + 252 + ], + "type": "text", + "content": " \n9: Calculate the regularization commitment term \n10: " + }, + { + "bbox": [ + 310, + 87, + 545, + 252 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{rc} = t\\tau \\| \\mathrm{sg}(\\pmb{\\epsilon}_{\\theta}(\\pmb{x}_{t}^{(i)}, y^{(i)})) - \\pmb{\\epsilon}_{\\theta}(\\pmb{x}_{t}^{(i)}, y'^{(i)}) \\|^{2}" + }, + { + "bbox": [ + 310, + 87, + 545, + 252 + ], + "type": "text", + "content": " \n11: Update with " + }, + { + "bbox": [ + 310, + 87, + 545, + 252 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{CBDM}} = \\mathcal{L}_{\\mathrm{DM}} + \\mathcal{L}_{r} + \\gamma \\mathcal{L}_{rc}" + }, + { + "bbox": [ + 310, + 87, + 545, + 252 + ], + "type": "text", + "content": " \n12: end for \n13: end for" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "code_body" + } + ], + "index": 9, + "sub_type": "algorithm" + }, + { + "bbox": [ + 304, + 290, + 545, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 290, + 545, + 460 + ], + "spans": [ + { + "bbox": [ + 304, + 290, + 545, + 460 + ], + "type": "text", + "content": "of the regularization weight " + }, + { + "bbox": [ + 304, + 290, + 545, + 460 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 304, + 290, + 545, + 460 + ], + "type": "text", + "content": " affects the sharpness of the density ratio " + }, + { + "bbox": [ + 304, + 290, + 545, + 460 + ], + "type": "inline_equation", + "content": "\\frac{p_{\\theta}(\\boldsymbol{x}_t)}{p_{\\theta}^*(\\boldsymbol{x}_t)}" + }, + { + "bbox": [ + 304, + 290, + 545, + 460 + ], + "type": "text", + "content": ". For the theoretical analysis, please refer to our proof of Prop. (2) in the Appendix. In addition, the choice of the sampling set " + }, + { + "bbox": [ + 304, + 290, + 545, + 460 + ], + "type": "inline_equation", + "content": "\\mathcal{V}" + }, + { + "bbox": [ + 304, + 290, + 545, + 460 + ], + "type": "text", + "content": " is another important perspective of CBDM, which depends on the target distribution we wish to adjust. Without loss of generality, we discuss two cases here. On the one hand, we can adjust the label distribution to a class-balanced label distribution, where we sample labels to construct " + }, + { + "bbox": [ + 304, + 290, + 545, + 460 + ], + "type": "inline_equation", + "content": "\\mathcal{V}^{bal}" + }, + { + "bbox": [ + 304, + 290, + 545, + 460 + ], + "type": "text", + "content": ". On the other hand, if the data distribution is heavily long-tailed, we can also target the adjusted distribution to a relatively less class-imbalanced distribution for stabilized training. In our experiments, we show CBDM can work well for both cases in different mechanisms." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 469, + 545, + 554 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 469, + 545, + 554 + ], + "spans": [ + { + "bbox": [ + 304, + 469, + 545, + 554 + ], + "type": "text", + "content": "Moreover, we observe that naively optimizing with this loss could make the model collapse to some trivial solutions, where the model outputs the same result regardless of the condition " + }, + { + "bbox": [ + 304, + 469, + 545, + 554 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 304, + 469, + 545, + 554 + ], + "type": "text", + "content": " and thus degenerates the conditional generation performance. Therefore, we follow previous works to apply a stop gradient operation [4, 46] to prevent this issue. The final loss of CBDM is" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 312, + 578, + 545, + 641 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 578, + 545, + 641 + ], + "spans": [ + { + "bbox": [ + 312, + 578, + 545, + 641 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} _ {\\mathrm {C B D M}} ^ {\\tau , \\gamma , \\mathcal {Y}} (\\pmb {x} _ {t}, y, t, \\pmb {\\epsilon}) = \\| \\pmb {\\epsilon} _ {\\theta} (\\pmb {x} _ {t}, y) - \\pmb {\\epsilon} \\| ^ {2} \\\\ + \\frac {\\tau t}{| \\mathcal {Y} |} \\sum_ {y ^ {\\prime} \\in \\mathcal {Y}} \\left(\\left\\| \\boldsymbol {\\epsilon} _ {\\theta} \\left(\\boldsymbol {x} _ {t}, y\\right) - \\operatorname {s g} \\left(\\boldsymbol {\\epsilon} _ {\\theta} \\left(\\boldsymbol {x} _ {t}, y ^ {\\prime}\\right)\\right) \\right\\| ^ {2} \\right. \\\\ + \\gamma \\| \\mathrm {s g} \\big (\\pmb {\\epsilon} _ {\\theta} (\\pmb {x} _ {t}, y) \\big) - \\pmb {\\epsilon} _ {\\theta} (\\pmb {x} _ {t}, y ^ {\\prime}) \\| ^ {2}), (4) \\\\ \\end{array}", + "image_path": "be921de692a3e199335f1b0e8828d63c8961cc73879c0eb39897fc5db72900cc.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 665, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 545, + 712 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 665, + 545, + 712 + ], + "type": "inline_equation", + "content": "\\mathrm{sg}(\\cdot)" + }, + { + "bbox": [ + 304, + 665, + 545, + 712 + ], + "type": "text", + "content": " denotes the stop gradient operation; " + }, + { + "bbox": [ + 304, + 665, + 545, + 712 + ], + "type": "inline_equation", + "content": "\\tau, \\gamma" + }, + { + "bbox": [ + 304, + 665, + 545, + 712 + ], + "type": "text", + "content": " are weights for the regularization and the commitment term respectively with " + }, + { + "bbox": [ + 304, + 665, + 545, + 712 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 304, + 665, + 545, + 712 + ], + "type": "text", + "content": " set to " + }, + { + "bbox": [ + 304, + 665, + 545, + 712 + ], + "type": "inline_equation", + "content": "\\frac{1}{4}" + }, + { + "bbox": [ + 304, + 665, + 545, + 712 + ], + "type": "text", + "content": " in a default setting, and " + }, + { + "bbox": [ + 304, + 665, + 545, + 712 + ], + "type": "inline_equation", + "content": "|\\mathcal{V}|" + }, + { + "bbox": [ + 304, + 665, + 545, + 712 + ], + "type": "text", + "content": " denotes the number of elements in the label set." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18437" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 51, + 72, + 164, + 185 + ], + "blocks": [ + { + "bbox": [ + 51, + 72, + 164, + 185 + ], + "lines": [ + { + "bbox": [ + 51, + 72, + 164, + 185 + ], + "spans": [ + { + "bbox": [ + 51, + 72, + 164, + 185 + ], + "type": "image", + "image_path": "dd0ccc3c990d101184872c016a8f0d46a73b9ecb4f44c1ece8ca8aedd15ec652.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 66, + 188, + 148, + 197 + ], + "lines": [ + { + "bbox": [ + 66, + 188, + 148, + 197 + ], + "spans": [ + { + "bbox": [ + 66, + 188, + 148, + 197 + ], + "type": "text", + "content": "(a) DDPM / Tail Class 94" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 170, + 72, + 284, + 185 + ], + "blocks": [ + { + "bbox": [ + 170, + 72, + 284, + 185 + ], + "lines": [ + { + "bbox": [ + 170, + 72, + 284, + 185 + ], + "spans": [ + { + "bbox": [ + 170, + 72, + 284, + 185 + ], + "type": "image", + "image_path": "a5e275ecd61e93b3ee7ff383f1185faf0a4e67d4a25b8eb536ef5ccfc200bed5.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 186, + 188, + 269, + 197 + ], + "lines": [ + { + "bbox": [ + 186, + 188, + 269, + 197 + ], + "spans": [ + { + "bbox": [ + 186, + 188, + 269, + 197 + ], + "type": "text", + "content": "(b) CBDM / Tail Class 94" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 51, + 200, + 164, + 313 + ], + "blocks": [ + { + "bbox": [ + 51, + 200, + 164, + 313 + ], + "lines": [ + { + "bbox": [ + 51, + 200, + 164, + 313 + ], + "spans": [ + { + "bbox": [ + 51, + 200, + 164, + 313 + ], + "type": "image", + "image_path": "946995c13751da3c24e493e2e4d8a3126d1ab1943c5615a9938206a54b8d6319.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 63, + 316, + 151, + 326 + ], + "lines": [ + { + "bbox": [ + 63, + 316, + 151, + 326 + ], + "spans": [ + { + "bbox": [ + 63, + 316, + 151, + 326 + ], + "type": "text", + "content": "(c) DDPM / Body Class 62" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 335, + 287, + 369 + ], + "lines": [ + { + "bbox": [ + 46, + 335, + 287, + 369 + ], + "spans": [ + { + "bbox": [ + 46, + 335, + 287, + 369 + ], + "type": "text", + "content": "Figure 3. Comparison of image generation on heavily tail-distributed (94) and mild tail-distributed (62) classes between DDPM and CBDM." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 169, + 200, + 282, + 313 + ], + "blocks": [ + { + "bbox": [ + 169, + 200, + 282, + 313 + ], + "lines": [ + { + "bbox": [ + 169, + 200, + 282, + 313 + ], + "spans": [ + { + "bbox": [ + 169, + 200, + 282, + 313 + ], + "type": "image", + "image_path": "1f6736bb9992078b5667558852d87a159b20deba9231a1a4a87b0ab4651d0ef8.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 181, + 316, + 270, + 326 + ], + "lines": [ + { + "bbox": [ + 181, + 316, + 270, + 326 + ], + "spans": [ + { + "bbox": [ + 181, + 316, + 270, + 326 + ], + "type": "text", + "content": "(d) CBDM / Body Class 62" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 393, + 168, + 407 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 393, + 168, + 407 + ], + "spans": [ + { + "bbox": [ + 47, + 393, + 168, + 407 + ], + "type": "text", + "content": "4. Experimental results" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 415, + 162, + 428 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 415, + 162, + 428 + ], + "spans": [ + { + "bbox": [ + 47, + 415, + 162, + 428 + ], + "type": "text", + "content": "4.1. Experimental setup" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 435, + 287, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 435, + 287, + 544 + ], + "spans": [ + { + "bbox": [ + 46, + 435, + 287, + 544 + ], + "type": "text", + "content": "Datasets We first chose two common datasets that are widely used in the domain of image synthesis, CIFAR10/CIFAR100, and their corresponding long-tailed versions CIFAR10LT and CIFAR100LT. The construction of CIFAR10LT and CIFAR100LT follows [3], where the size decreases exponentially with its category index according to the imbalance factor " + }, + { + "bbox": [ + 46, + 435, + 287, + 544 + ], + "type": "inline_equation", + "content": "imb = 0.01" + }, + { + "bbox": [ + 46, + 435, + 287, + 544 + ], + "type": "text", + "content": ". We also conduct experiments on 3 higher-resolution datasets, whose details and results could be found in the Appendix." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 557, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 557, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 557, + 288, + 715 + ], + "type": "text", + "content": "Implementation details We strictly follow the training configurations of the baseline models. For DDPM, we set diffusion schedule " + }, + { + "bbox": [ + 46, + 557, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\beta_{1} = 10^{-4}" + }, + { + "bbox": [ + 46, + 557, + 288, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 557, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\beta_{T} = 0.02" + }, + { + "bbox": [ + 46, + 557, + 288, + 715 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 46, + 557, + 288, + 715 + ], + "type": "inline_equation", + "content": "T = 1,000" + }, + { + "bbox": [ + 46, + 557, + 288, + 715 + ], + "type": "text", + "content": ", and optimize the network with an Adam optimizer whose learning rate is 0.0002 after 5,000 epochs of warmup. Considering that the size and semantic complexity of the datasets vary greatly, we choose appropriate epochs for each dataset. In the ablation studies, we also include EDM [20] and Diffusion-ViT [48] as our backbones to show the compatibility with different architectures and score-matching methods. Note that, when a model does not use conditional input, we follow [7] to slightly modify the backbone [8] by adding an extra embedding layer." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 72, + 545, + 263 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 263 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 263 + ], + "type": "text", + "content": "Baselines We first adopt following classic methods to build baselines: re-sampling (RS), soft re-sampling methods (RS-SQRT) [27], and augmentation-based methods including Differentiable Augmentation (DiffAug) [50] and Adaptive Augmentation (ADA) [21]. Besides diffusion-based baselines, previous state-of-the-art generative models that study the long-tailed distribution, e.g., CBGAN [35], and group spectral regularization for GANs [34], are also included in the comparison. Precisely, RS follows the uniform class distribution and RS-SQRT [27] samples with a probability determined by the square root of class frequency. And DiffAug [50] is directly applied to all training images following their default threshold. Following [20], ADA [21] is not only applied to images, the augmentation pipeline is also encoded as a condition via an additional embedding layer in the U-Net." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 267, + 545, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 267, + 545, + 472 + ], + "spans": [ + { + "bbox": [ + 304, + 267, + 545, + 472 + ], + "type": "text", + "content": "Metrics CBDM and corresponding baselines are evaluated in terms of both generation diversity and fidelity via Frechet Inception Distance (FID) [10], Inception Score (IS) [40], Recall [23] and " + }, + { + "bbox": [ + 304, + 267, + 545, + 472 + ], + "type": "inline_equation", + "content": "F_{\\beta}" + }, + { + "bbox": [ + 304, + 267, + 545, + 472 + ], + "type": "text", + "content": " [39]. We measure Recall and " + }, + { + "bbox": [ + 304, + 267, + 545, + 472 + ], + "type": "inline_equation", + "content": "F_{\\beta}" + }, + { + "bbox": [ + 304, + 267, + 545, + 472 + ], + "type": "text", + "content": " using Inception-V3 features, and take respectively " + }, + { + "bbox": [ + 304, + 267, + 545, + 472 + ], + "type": "inline_equation", + "content": "K = 5" + }, + { + "bbox": [ + 304, + 267, + 545, + 472 + ], + "type": "text", + "content": " for Recall, 1/8 and 8 for the threshold in " + }, + { + "bbox": [ + 304, + 267, + 545, + 472 + ], + "type": "inline_equation", + "content": "F_{\\beta}" + }, + { + "bbox": [ + 304, + 267, + 545, + 472 + ], + "type": "text", + "content": ", and 20 times of class number as the clustering number of " + }, + { + "bbox": [ + 304, + 267, + 545, + 472 + ], + "type": "inline_equation", + "content": "F_{\\beta}" + }, + { + "bbox": [ + 304, + 267, + 545, + 472 + ], + "type": "text", + "content": " to capture the inner class variance. To this end, Recall and " + }, + { + "bbox": [ + 304, + 267, + 545, + 472 + ], + "type": "inline_equation", + "content": "F_{8}" + }, + { + "bbox": [ + 304, + 267, + 545, + 472 + ], + "type": "text", + "content": " can be regarded as diversity metrics and IS and " + }, + { + "bbox": [ + 304, + 267, + 545, + 472 + ], + "type": "inline_equation", + "content": "F_{1/8}" + }, + { + "bbox": [ + 304, + 267, + 545, + 472 + ], + "type": "text", + "content": " are more inclined to measure fidelity. In evaluation, we take the class-balanced version corresponding to those used in training, and all metrics are measured with 50k generated images (10k for ablation experiments). The classifier-free guidance is adapted in sampling and we tune the guidance strength " + }, + { + "bbox": [ + 304, + 267, + 545, + 472 + ], + "type": "inline_equation", + "content": "\\omega" + }, + { + "bbox": [ + 304, + 267, + 545, + 472 + ], + "type": "text", + "content": " for both baselines and our method to ensure the best performance. Specifically, we take " + }, + { + "bbox": [ + 304, + 267, + 545, + 472 + ], + "type": "inline_equation", + "content": "\\omega = 1.6, 0.8, 1.0, 0.8" + }, + { + "bbox": [ + 304, + 267, + 545, + 472 + ], + "type": "text", + "content": " respectively for the four CIFAR datasets mentioned above." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 479, + 388, + 491 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 479, + 388, + 491 + ], + "spans": [ + { + "bbox": [ + 306, + 479, + 388, + 491 + ], + "type": "text", + "content": "4.2. Main results" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 498, + 545, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 498, + 545, + 653 + ], + "spans": [ + { + "bbox": [ + 304, + 498, + 545, + 653 + ], + "type": "text", + "content": "In this part, we consider DDPM as the most direct baseline. As shown in Table 1, CBDM overall outperforms DDPM on all datasets except the IS is slightly lower on CIFAR10. For datasets with more classes, the improvements are more significant in terms of both diversity and fidelity, as on CIFAR100LT and CIFAR100. We further investigate the performance of diffusion models when combined with augmentation-based and re-sampling methods. Surprisingly, except for ADA, we consistently observe a performance degradation when these methods are compared with the vanilla DDPM. Moreover, adding the ADA augmentation with CBDM training leads to a further improvement in terms of both diversity and fidelity." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "content": "As a qualitative justification, Figure 3 provides the visualization comparison between DDPM and CBDM on a relatively mild tail-distributed class (62) and a tail-distributed class (94). We remark that CBDM generates more diverse images. For example, on the tail class 94, the cabinets" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18438" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 69, + 70, + 525, + 267 + ], + "blocks": [ + { + "bbox": [ + 69, + 70, + 525, + 267 + ], + "lines": [ + { + "bbox": [ + 69, + 70, + 525, + 267 + ], + "spans": [ + { + "bbox": [ + 69, + 70, + 525, + 267 + ], + "type": "table", + "html": "
DatasetModelFID↓F8↑Recall↑IS↑F1/8↑
CIFAR100LTDDPM [11]7.380.850.5213.110.88
+ADA [21]6.160.910.5712.710.90
+DiffAug [50]9.190.880.4711.560.86
+RS [27]10.500.650.4912.600.83
+SQRT-RS [27]9.720.660.4713.470.83
CBDM (ours)6.26 (-1.12)0.91 (+0.06)0.57 (+0.05)13.24 (+0.13)0.89 (+0.01)
+ADA [21]5.81 (-1.57)0.91 (+0.06)0.57 (+0.05)13.34 (+0.23)0.90 (+0.02)
CIFAR100DDPM3.110.970.6513.650.96
CBDM (ours)2.72 (-0.39)0.97 (±0)0.67 (+0.02)14.03 (+0.38)0.96 (±0)
CIFAR10LTDDPM5.760.970.579.170.95
CBDM (ours)5.46 (-0.30)0.97 (±0)0.59 (+0.02)9.28 (+0.11)0.95 (±0)
CIFAR10DDPM3.160.990.649.800.98
CBDM (ours)3.03 (-0.13)0.99 (±0)0.65 (+0.01)9.63 (-0.17)0.98 (±0)
", + "image_path": "1be50f1d86ff41330a3e8ee5dfcf49a9944ca71724c33c293240408a9bd12304.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 50, + 327, + 288, + 456 + ], + "blocks": [ + { + "bbox": [ + 46, + 275, + 547, + 309 + ], + "lines": [ + { + "bbox": [ + 46, + 275, + 547, + 309 + ], + "spans": [ + { + "bbox": [ + 46, + 275, + 547, + 309 + ], + "type": "text", + "content": "Table 1. CBDM performance on different datasets. In the table, the first three columns are diversity-related, and we mark the best results in bold. As the comparison between CBDM and DDPM is more straightforward, we mark the performance gain in parentheses next to the CBDM results, using blue and red to indicate improvements and degradation respectively." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 50, + 327, + 288, + 456 + ], + "lines": [ + { + "bbox": [ + 50, + 327, + 288, + 456 + ], + "spans": [ + { + "bbox": [ + 50, + 327, + 288, + 456 + ], + "type": "table", + "html": "
PTFTyFIDF8RecallIS
CIFAR 100LT-ytrain6.260.910.5713.24
ybal6.100.940.6412.29
ysqrt6.000.930.5913.00
CIFAR 100LTCIFAR 100LT-7.380.850.5213.11
ytrain6.200.880.5413.36
ybal5.850.920.6013.29
CIFAR 100CIFAR 10LT-5.390.660.589.43
ytrain5.050.670.609.48
ybal5.900.660.609.49
", + "image_path": "82052a5dd0c0971d426f54adb1988032f9171d2827b02b09542b47a3e57b2ac1.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 464, + 287, + 544 + ], + "lines": [ + { + "bbox": [ + 46, + 464, + 287, + 544 + ], + "spans": [ + { + "bbox": [ + 46, + 464, + 287, + 544 + ], + "type": "text", + "content": "Table 2. CBDM performance for 3 mechanisms under different regularization sampling set " + }, + { + "bbox": [ + 46, + 464, + 287, + 544 + ], + "type": "inline_equation", + "content": "\\mathcal{V}" + }, + { + "bbox": [ + 46, + 464, + 287, + 544 + ], + "type": "text", + "content": ". Column " + }, + { + "bbox": [ + 46, + 464, + 287, + 544 + ], + "type": "inline_equation", + "content": "PT" + }, + { + "bbox": [ + 46, + 464, + 287, + 544 + ], + "type": "text", + "content": " indicates the dataset used in (pre)training, column " + }, + { + "bbox": [ + 46, + 464, + 287, + 544 + ], + "type": "inline_equation", + "content": "FT" + }, + { + "bbox": [ + 46, + 464, + 287, + 544 + ], + "type": "text", + "content": " indicates the dataset used for fine-tuning. The rows where " + }, + { + "bbox": [ + 46, + 464, + 287, + 544 + ], + "type": "inline_equation", + "content": "\\mathcal{V}" + }, + { + "bbox": [ + 46, + 464, + 287, + 544 + ], + "type": "text", + "content": " marked as “-” represent the results with DDPM, and rows where " + }, + { + "bbox": [ + 46, + 464, + 287, + 544 + ], + "type": "inline_equation", + "content": "FT" + }, + { + "bbox": [ + 46, + 464, + 287, + 544 + ], + "type": "text", + "content": " marked as “-” represent the results of CBDM trained from scratch. Three diversity-preferred metrics and the best results are in bold for emphasis." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 567, + 287, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 567, + 287, + 616 + ], + "spans": [ + { + "bbox": [ + 46, + 567, + 287, + 616 + ], + "type": "text", + "content": "shown in Figure 3b have more color and texture, which justifies the improvements of diversity-related metrics in Table 1. On the contrary, DDPM only generates images that are highly similar to the training data." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "type": "text", + "content": "Case-by-case study. To better understand the generation conditioned on each class, we compare the FID case-by-case between DDPM and CBDM on each class. The results are shown in Figure 4. Compared to DDPM, CBDM shows more consistent improvements for the class index greater than 40. In the tail classes, CBDM outperforms than DDPM the FID by a more significant margin." + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 319, + 331, + 533, + 443 + ], + "blocks": [ + { + "bbox": [ + 319, + 331, + 533, + 443 + ], + "lines": [ + { + "bbox": [ + 319, + 331, + 533, + 443 + ], + "spans": [ + { + "bbox": [ + 319, + 331, + 533, + 443 + ], + "type": "image", + "image_path": "c558ad75a1e4270f6c288b5f82a37507320ddcc04703d71f4d45644f2e9edd1c.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 453, + 545, + 477 + ], + "lines": [ + { + "bbox": [ + 305, + 453, + 545, + 477 + ], + "spans": [ + { + "bbox": [ + 305, + 453, + 545, + 477 + ], + "type": "text", + "content": "Figure 4. FID improvement per class compared to DDPM. The curve is smoothened by a moving average of 5 unit." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 497, + 546, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 497, + 546, + 688 + ], + "spans": [ + { + "bbox": [ + 304, + 497, + 546, + 688 + ], + "type": "text", + "content": "The choice of label set " + }, + { + "bbox": [ + 304, + 497, + 546, + 688 + ], + "type": "inline_equation", + "content": "\\mathcal{V}" + }, + { + "bbox": [ + 304, + 497, + 546, + 688 + ], + "type": "text", + "content": " We investigate the effects of label distribution for CBDM under three mechanisms. The first mechanism is the vanilla CBDM given in Algorithm 1 which trains a diffusion model with regularization from scratch. Besides, we also study CBDM with two fine-tuning configurations. The first setting concerns using CBDM to adjust an existing model pre-trained on a long-tailed dataset to improve its performance. The second one refers to fine-tuning a pre-trained model to adapt to a smaller dataset with an imbalanced class distribution using CBDM. For those mechanisms, we consider three settings for sampling the label set: (1) a label distribution similar to the training set (denoted " + }, + { + "bbox": [ + 304, + 497, + 546, + 688 + ], + "type": "inline_equation", + "content": "\\mathcal{V}^{train}" + }, + { + "bbox": [ + 304, + 497, + 546, + 688 + ], + "type": "text", + "content": "); (2) a totally uniform label distribution (denoted " + }, + { + "bbox": [ + 304, + 497, + 546, + 688 + ], + "type": "inline_equation", + "content": "\\mathcal{V}^{bal}" + }, + { + "bbox": [ + 304, + 497, + 546, + 688 + ], + "type": "text", + "content": "); (3) a less long-tailed distribution compared to " + }, + { + "bbox": [ + 304, + 497, + 546, + 688 + ], + "type": "inline_equation", + "content": "\\mathcal{V}^{train}" + }, + { + "bbox": [ + 304, + 497, + 546, + 688 + ], + "type": "text", + "content": ", whose class frequency is the square root of the original one (denoted " + }, + { + "bbox": [ + 304, + 497, + 546, + 688 + ], + "type": "inline_equation", + "content": "\\mathcal{V}^{sqrt}" + }, + { + "bbox": [ + 304, + 497, + 546, + 688 + ], + "type": "text", + "content": ")." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 689, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 689, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 305, + 689, + 547, + 714 + ], + "type": "text", + "content": "The results are shown in Table 2, where we found that CBDM performs well on all these configurations. As we" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "18439" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 50, + 70, + 289, + 220 + ], + "blocks": [ + { + "bbox": [ + 50, + 70, + 289, + 220 + ], + "lines": [ + { + "bbox": [ + 50, + 70, + 289, + 220 + ], + "spans": [ + { + "bbox": [ + 50, + 70, + 289, + 220 + ], + "type": "table", + "html": "
DatasetModelFIDIS
CIFAR 100BigGAN+DA+RLC[45](2021)2.99-
DDPM3.1113.65
CBDM (ours)2.7214.03
CIFAR 100LTCBGAN [35](2021)28.17-
DDPM12.3112.69
CBDM (ours)+ADA [21]10.5012.81
CIFAR 10LTCBGAN [35](2021)32.93-
SNGAN [30]+gSR [34](2022)18.587.80
DDPM9.689.00
CBDM (ours)9.389.12
", + "image_path": "b65227766b8c1e45b4a402d90bb701d283a2bcccb2a9deb9bb76887d446f3dff.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 58, + 284, + 277, + 357 + ], + "blocks": [ + { + "bbox": [ + 46, + 227, + 288, + 273 + ], + "lines": [ + { + "bbox": [ + 46, + 227, + 288, + 273 + ], + "spans": [ + { + "bbox": [ + 46, + 227, + 288, + 273 + ], + "type": "text", + "content": "Table 3. Comparison with long-tailed SoTAs on CIFAR. Following their setting, all methods are evaluated with " + }, + { + "bbox": [ + 46, + 227, + 288, + 273 + ], + "type": "inline_equation", + "content": "10\\mathrm{k}" + }, + { + "bbox": [ + 46, + 227, + 288, + 273 + ], + "type": "text", + "content": " generated images and ground truth images are from their validation dataset. The publish year of every baselines are marked in next to the citation." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 284, + 277, + 357 + ], + "lines": [ + { + "bbox": [ + 58, + 284, + 277, + 357 + ], + "spans": [ + { + "bbox": [ + 58, + 284, + 277, + 357 + ], + "type": "table", + "html": "
Training dataPrecisionRecall
CIFAR1000.700.67
CIFAR100LT + DDPMgens (50k)0.450.39
0.48 (+0.03)0.44 (+0.05)
+ CBDMgens (50k)0.49 (+0.04)0.47 (+0.08)
", + "image_path": "a2d689151eed7765c7c0e65372750e3754f36c4b6164a7b666338c541ace9310.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 421, + 289, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 421, + 289, + 601 + ], + "spans": [ + { + "bbox": [ + 46, + 421, + 289, + 601 + ], + "type": "text", + "content": "analyzed in section 3.3, when the empirical distribution of " + }, + { + "bbox": [ + 46, + 421, + 289, + 601 + ], + "type": "inline_equation", + "content": "\\mathcal{V}" + }, + { + "bbox": [ + 46, + 421, + 289, + 601 + ], + "type": "text", + "content": " has a significant difference from the label distribution of the dataset, adjusting with " + }, + { + "bbox": [ + 46, + 421, + 289, + 601 + ], + "type": "inline_equation", + "content": "\\mathcal{V}^{bal}" + }, + { + "bbox": [ + 46, + 421, + 289, + 601 + ], + "type": "text", + "content": " could hurt training stability and result in performance degradation. We observe that " + }, + { + "bbox": [ + 46, + 421, + 289, + 601 + ], + "type": "inline_equation", + "content": "\\mathcal{V}^{bal}" + }, + { + "bbox": [ + 46, + 421, + 289, + 601 + ], + "type": "text", + "content": " only shows better performance regarding generation diversity, while the generation fidelity has an obvious gap than the other settings. Using a relatively mild set " + }, + { + "bbox": [ + 46, + 421, + 289, + 601 + ], + "type": "inline_equation", + "content": "\\mathcal{V}^{sqrt}" + }, + { + "bbox": [ + 46, + 421, + 289, + 601 + ], + "type": "text", + "content": " has better performance in terms of IS and preserves the generation diversity, which leads to the best FID among all these settings. On the contrary, when a pre-trained model is available, fine-tuning on top of it ensures better stability, and we can observe in such case using " + }, + { + "bbox": [ + 46, + 421, + 289, + 601 + ], + "type": "inline_equation", + "content": "\\mathcal{V}^{bal}" + }, + { + "bbox": [ + 46, + 421, + 289, + 601 + ], + "type": "text", + "content": " has significant improvement than the vanilla DDPM or fine-tuning using " + }, + { + "bbox": [ + 46, + 421, + 289, + 601 + ], + "type": "inline_equation", + "content": "\\mathcal{V}^{train}" + }, + { + "bbox": [ + 46, + 421, + 289, + 601 + ], + "type": "text", + "content": ". When adapting to a different dataset, we also observe using " + }, + { + "bbox": [ + 46, + 421, + 289, + 601 + ], + "type": "inline_equation", + "content": "\\mathcal{V}^{bal}" + }, + { + "bbox": [ + 46, + 421, + 289, + 601 + ], + "type": "text", + "content": " causes stability issue and produce undesired results." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "text", + "content": "Enhancement of training classifiers on long-tailed data As training models on long-tailed data usually leads to undesired classification results, we investigate whether the generated data could help improve the classifier trained on long-tailed data as a complementary evaluation metrics. Here, we train ResNet-32 models, respectively with CIFAR100, CIFAR100LT and CIFAR100LT augmented with DDPM and CBDM generations (50k samples). Table 4 shows that the training on long-tailed data results in se" + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 307, + 70, + 552, + 148 + ], + "blocks": [ + { + "bbox": [ + 46, + 365, + 287, + 399 + ], + "lines": [ + { + "bbox": [ + 46, + 365, + 287, + 399 + ], + "spans": [ + { + "bbox": [ + 46, + 365, + 287, + 399 + ], + "type": "text", + "content": "Table 4. Recognition results of different training data. All configurations are evaluated on the testing set of normal CIFAR100. Gains based on CIFAR100LT dataset is noted in blue." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 307, + 70, + 552, + 148 + ], + "lines": [ + { + "bbox": [ + 307, + 70, + 552, + 148 + ], + "spans": [ + { + "bbox": [ + 307, + 70, + 552, + 148 + ], + "type": "table", + "html": "
BackboneFID\\( F_8 \\)RecallIS\\( F_{1/8} \\)
EDM [20]8.640.820.4812.160.86
+CBDM (ours)7.970.880.5212.010.86
Diffuse-ViT [48]37.60.760.457.660.62
+CBDM (ours)30.00.810.518.000.65
", + "image_path": "689bcfd12a16d911291d6f4b25c019a809a4648edb10d65e6a3864fea5b65a23.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 309, + 189, + 544, + 251 + ], + "blocks": [ + { + "bbox": [ + 305, + 156, + 545, + 179 + ], + "lines": [ + { + "bbox": [ + 305, + 156, + 545, + 179 + ], + "spans": [ + { + "bbox": [ + 305, + 156, + 545, + 179 + ], + "type": "text", + "content": "Table 5. CBDM performance using different backbones on CIFAR100LT dataset. Three diversity-biased metrics." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 309, + 189, + 544, + 251 + ], + "lines": [ + { + "bbox": [ + 309, + 189, + 544, + 251 + ], + "spans": [ + { + "bbox": [ + 309, + 189, + 544, + 251 + ], + "type": "table", + "html": "
DatasetFID↓F8↑Recall↑IS↑F1/8↑
CIFAR3.440.970.6913.110.96
100(+0.72)(±0)(+0.02)(-0.92)(±0)
CIFAR6.270.930.6112.560.89
100LT(+0.01)(+0.02)(+0.04)(-0.68)(±0)
", + "image_path": "4ceee53c06ef23e876299db383d03160c65542e78c798aa6603a7e7117fd2ff7.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 258, + 545, + 293 + ], + "lines": [ + { + "bbox": [ + 305, + 258, + 545, + 293 + ], + "spans": [ + { + "bbox": [ + 305, + 258, + 545, + 293 + ], + "type": "text", + "content": "Table 6. Sampling with 100 DDIM steps on CIFAR100 and CIFAR100LT. The difference compared with those using DDPM steps are reported below each result." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 312, + 545, + 420 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 312, + 545, + 420 + ], + "spans": [ + { + "bbox": [ + 304, + 312, + 545, + 420 + ], + "type": "text", + "content": "vere performance degradation. When augmenting the long-tailed data with the generated data, we observe fairly improvements in the trained classifier. Moreover, comparing the results between DDPM and CBDM, we observe that the improvements by using CBDM are more significant, which validates the effectiveness of our approach. Especially, the gain on recall is more remarkable than the gain on precision, which means that the generated images are more diverse than DDPM and justifies the results in Table 1." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 422, + 546, + 531 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 422, + 546, + 531 + ], + "spans": [ + { + "bbox": [ + 304, + 422, + 546, + 531 + ], + "type": "text", + "content": "Comparison with other benchmarks We include the representative state-of-the-art long-tailed generative modeling approaches in the comparison, most of which are based on GANs since they suffer more severe problems when the training data is skewed. For a fair comparison, we strictly follow their evaluation setting and the results are reported in Table 3. From the results, DDPM surpasses the performance of the baselines on all the datasets, and CBDM exhibits even stronger performance." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 537, + 373, + 548 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 537, + 373, + 548 + ], + "spans": [ + { + "bbox": [ + 306, + 537, + 373, + 548 + ], + "type": "text", + "content": "4.3. Ablations" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 555, + 545, + 675 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 555, + 545, + 675 + ], + "spans": [ + { + "bbox": [ + 304, + 555, + 545, + 675 + ], + "type": "text", + "content": "Compatibility with different backbones We first investigate the performance of our method with different diffusion models. Without loss of generality, we adopt EDM [20] and Diffusion-ViT [48], which use a different score-matching method and different denoising network backbone. As shown in Table 5, compared with the corresponding baselines, CBDM improves the generation result on CIFAR100LT dataset for both backbones, demonstrating the compatibility with a variety of backbones and the effectiveness of handling long-tailed data." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 677, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 677, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 677, + 545, + 713 + ], + "type": "text", + "content": "CBDM with DDIM sampling Apart from DDPM reverse sampling and neural SDE methods, recent deterministic ODE methods such as DDIM [43] are widely used in the" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "18440" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 70, + 165, + 151 + ], + "blocks": [ + { + "bbox": [ + 50, + 70, + 165, + 151 + ], + "lines": [ + { + "bbox": [ + 50, + 70, + 165, + 151 + ], + "spans": [ + { + "bbox": [ + 50, + 70, + 165, + 151 + ], + "type": "image", + "image_path": "75a37acadd05bc48854fe94ad9c7471bb083e68cbc4bf9dbf366ed162efb6aee.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 53, + 154, + 280, + 166 + ], + "lines": [ + { + "bbox": [ + 53, + 154, + 280, + 166 + ], + "spans": [ + { + "bbox": [ + 53, + 154, + 280, + 166 + ], + "type": "text", + "content": "Figure 5. FID/IS score under different regularization weight " + }, + { + "bbox": [ + 53, + 154, + 280, + 166 + ], + "type": "inline_equation", + "content": "\\tau" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 171, + 70, + 284, + 151 + ], + "blocks": [ + { + "bbox": [ + 171, + 70, + 284, + 151 + ], + "lines": [ + { + "bbox": [ + 171, + 70, + 284, + 151 + ], + "spans": [ + { + "bbox": [ + 171, + 70, + 284, + 151 + ], + "type": "image", + "image_path": "fe780cded2078a75ff6b1eceef772b462c1b30bc7180e3e0bcce931b98a322b5.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 50, + 177, + 165, + 256 + ], + "blocks": [ + { + "bbox": [ + 50, + 177, + 165, + 256 + ], + "lines": [ + { + "bbox": [ + 50, + 177, + 165, + 256 + ], + "spans": [ + { + "bbox": [ + 50, + 177, + 165, + 256 + ], + "type": "image", + "image_path": "9069478d6c2608070b1710d63bfb23e68e15d906fe96c02ba312399978efd9cc.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 59, + 258, + 274, + 270 + ], + "lines": [ + { + "bbox": [ + 59, + 258, + 274, + 270 + ], + "spans": [ + { + "bbox": [ + 59, + 258, + 274, + 270 + ], + "type": "text", + "content": "Figure 6. FID/IS score under different guidance strength " + }, + { + "bbox": [ + 59, + 258, + 274, + 270 + ], + "type": "inline_equation", + "content": "\\omega" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 170, + 177, + 284, + 257 + ], + "blocks": [ + { + "bbox": [ + 170, + 177, + 284, + 257 + ], + "lines": [ + { + "bbox": [ + 170, + 177, + 284, + 257 + ], + "spans": [ + { + "bbox": [ + 170, + 177, + 284, + 257 + ], + "type": "image", + "image_path": "c1e02173e8bc13d89d5bce211f5e972fcd8a07c83970705729f9b7c8898e749e.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 292, + 287, + 388 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 292, + 287, + 388 + ], + "spans": [ + { + "bbox": [ + 46, + 292, + 287, + 388 + ], + "type": "text", + "content": "generation process of diffusion models. We also conduct experiments to study whether CBDM is compatible with DDIM steps. Table 6 shows the comparison of using DDPM and DDIM reverse steps when the generation steps are compressed to " + }, + { + "bbox": [ + 46, + 292, + 287, + 388 + ], + "type": "inline_equation", + "content": "1/10" + }, + { + "bbox": [ + 46, + 292, + 287, + 388 + ], + "type": "text", + "content": ". We note on normal CIFAR100, FID and IS become slightly worse than using 1000 DDPM steps. On CIFAR100LT, except for IS, all metrics remain almost the same as using DDPM steps." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 392, + 287, + 524 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 392, + 287, + 524 + ], + "spans": [ + { + "bbox": [ + 46, + 392, + 287, + 524 + ], + "type": "text", + "content": "Effects of hyperparameters We tested the effects of regularization weight " + }, + { + "bbox": [ + 46, + 392, + 287, + 524 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 46, + 392, + 287, + 524 + ], + "type": "text", + "content": " in CBDM, which is set in default " + }, + { + "bbox": [ + 46, + 392, + 287, + 524 + ], + "type": "inline_equation", + "content": "\\tau = \\tau_0 = 0.001" + }, + { + "bbox": [ + 46, + 392, + 287, + 524 + ], + "type": "text", + "content": " so that the weight does not surpass 1 at all steps. As shown in Figure 5, we search weights across different scales and found that the optimal weight is " + }, + { + "bbox": [ + 46, + 392, + 287, + 524 + ], + "type": "inline_equation", + "content": "\\tau = \\tau_0" + }, + { + "bbox": [ + 46, + 392, + 287, + 524 + ], + "type": "text", + "content": ". The results also point out that " + }, + { + "bbox": [ + 46, + 392, + 287, + 524 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 46, + 392, + 287, + 524 + ], + "type": "text", + "content": " should not be too large or too small. We also tested the difference between CBDM models trained using commitment loss " + }, + { + "bbox": [ + 46, + 392, + 287, + 524 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{rc}" + }, + { + "bbox": [ + 46, + 392, + 287, + 524 + ], + "type": "text", + "content": " and those trained without it. Our experiments show that the FID score of the models using the commitment loss (8.30) are significantly lower than those of the models not using it (8.84)." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 529, + 286, + 660 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 529, + 286, + 660 + ], + "spans": [ + { + "bbox": [ + 46, + 529, + 286, + 660 + ], + "type": "text", + "content": "Guidance strength " + }, + { + "bbox": [ + 46, + 529, + 286, + 660 + ], + "type": "inline_equation", + "content": "\\omega" + }, + { + "bbox": [ + 46, + 529, + 286, + 660 + ], + "type": "text", + "content": " As we adapt the Classifier-Free Guidance (CFG) during sampling, we also show the effects of guidance strength " + }, + { + "bbox": [ + 46, + 529, + 286, + 660 + ], + "type": "inline_equation", + "content": "\\omega" + }, + { + "bbox": [ + 46, + 529, + 286, + 660 + ], + "type": "text", + "content": ". In Figure 6, we searched " + }, + { + "bbox": [ + 46, + 529, + 286, + 660 + ], + "type": "inline_equation", + "content": "\\omega" + }, + { + "bbox": [ + 46, + 529, + 286, + 660 + ], + "type": "text", + "content": " from 0 to 2 with an interval of 0.2 and compare FID and IS score of DDPM and CBDM models. We observe that the FID of CBDM remains decreasing at larger guidance strengths compared to DDPM; and although the IS of CBDM is consistently weaker than that of DDPM, they reach the same level when the guidance strength reaches 2.0, where the FID of CBDM (6.75) is still significantly lower than the best FID of DDPM (7.38)." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "text", + "content": "Fidelity-diversity control [12] describes the phenomenon where excessive guidance strength " + }, + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\omega" + }, + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "text", + "content": " tends to lead to better fidelity at the price of overfitting and diversity degradation, as a fidelity-diversity tradeoff. CBDM" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 316, + 76, + 536, + 199 + ], + "blocks": [ + { + "bbox": [ + 316, + 76, + 536, + 199 + ], + "lines": [ + { + "bbox": [ + 316, + 76, + 536, + 199 + ], + "spans": [ + { + "bbox": [ + 316, + 76, + 536, + 199 + ], + "type": "image", + "image_path": "3012b664f7380c1364ed1b785a1aa3ee14e747273724618fe884593f33f0e987.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 202, + 545, + 226 + ], + "lines": [ + { + "bbox": [ + 305, + 202, + 545, + 226 + ], + "spans": [ + { + "bbox": [ + 305, + 202, + 545, + 226 + ], + "type": "text", + "content": "Figure 7. Image fidelity and diversity controlled by guidance strength " + }, + { + "bbox": [ + 305, + 202, + 545, + 226 + ], + "type": "inline_equation", + "content": "\\omega" + }, + { + "bbox": [ + 305, + 202, + 545, + 226 + ], + "type": "text", + "content": " and regularization weight " + }, + { + "bbox": [ + 305, + 202, + 545, + 226 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 305, + 202, + 545, + 226 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 245, + 547, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 245, + 547, + 413 + ], + "spans": [ + { + "bbox": [ + 304, + 245, + 547, + 413 + ], + "type": "text", + "content": "additionally provides another hyperparameter: the regularization weight " + }, + { + "bbox": [ + 304, + 245, + 547, + 413 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 304, + 245, + 547, + 413 + ], + "type": "text", + "content": " that promotes model diversity through cross-class information interaction. Figure 7 shows the body class (53) generation results under different guidance strengths and different regularization weight. We observe that a DDPM model with high guidance strength is able to generate very realistic results, but its images overlap almost exactly with a sample in the training set. In contrast, the guidance term of CBDM does not drastically change the image content, but rather refines the image content to be closer to the selected class than the unguided result. This means that CBDM retains more of the diversity of the unguided case and effectively enhances the category information with guide terms." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 423, + 378, + 436 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 423, + 378, + 436 + ], + "spans": [ + { + "bbox": [ + 306, + 423, + 378, + 436 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 444, + 545, + 598 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 444, + 545, + 598 + ], + "spans": [ + { + "bbox": [ + 304, + 444, + 545, + 598 + ], + "type": "text", + "content": "In this paper, we focus on the problem when the training data is imbalanced and the diffusion model drops in generation quality on tail classes. We first establish a low baseline for this task by examining the most common paradigms in long-tail scenarios and in training generative models with limited data. Thereafter, we propose the Class-Balancing Diffusion Model (CBDM) through theoretical analysis. This approach can be implemented very cleanly in the training of any conditional diffusion model and thus has the potential to be widely used in other fields. Our experiments show that the CBDM approach significantly improves model generation diversity with high fidelity, on both class-balanced and class-imbalanced datasets." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 609, + 416, + 624 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 609, + 416, + 624 + ], + "spans": [ + { + "bbox": [ + 306, + 609, + 416, + 624 + ], + "type": "text", + "content": "6. Acknowledgement" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 629, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 629, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 629, + 547, + 713 + ], + "type": "text", + "content": "This work is supported by the National Key R&D Program of China (No. 2022ZD0160703, No. 2022ZD0160702), National Natural Science Foundation of China (62271308), STCSM (No. 22511106101, No. 18DZ2270700, No. 22511105700, No. 21DZ1100100), 111 plan (No. BP0719010), and State Key Laboratory of UHD Video and Audio Production and Presentation." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18441" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 286, + 712 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 53, + 91, + 286, + 113 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 286, + 113 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 286, + 113 + ], + "type": "text", + "content": "[1] Jan Kautz Arash Vahdat, Karsten Kreis. Score-based generative modeling in latent space. In NeurIPS, 2021. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 115, + 286, + 147 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 115, + 286, + 147 + ], + "spans": [ + { + "bbox": [ + 53, + 115, + 286, + 147 + ], + "type": "text", + "content": "[2] Andrew Brock, Jeff Donahue, and Karen Simonyan. Large scale GAN training for high fidelity natural image synthesis. In ICLR, 2019. 1, 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 148, + 286, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 148, + 286, + 182 + ], + "spans": [ + { + "bbox": [ + 53, + 148, + 286, + 182 + ], + "type": "text", + "content": "[3] Kaidi Cao, Colin Wei, Adrien Gaidon, Nikos Arechiga, and Tengyu Ma. Learning imbalanced datasets with label-distribution-aware margin loss. NeurIPS, 32, 2019. 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 183, + 286, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 183, + 286, + 205 + ], + "spans": [ + { + "bbox": [ + 53, + 183, + 286, + 205 + ], + "type": "text", + "content": "[4] Xinlei Chen and Kaiming He. Exploring simple siamese representation learning. In CVPR, pages 15750-15758, 2021. 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 206, + 286, + 261 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 206, + 286, + 261 + ], + "spans": [ + { + "bbox": [ + 53, + 206, + 286, + 261 + ], + "type": "text", + "content": "[5] Peng Chu, Xiao Bian, Shaopeng Liu, and Haibin Ling. Feature space augmentation for long-tailed data. In Andrea Vedaldi, Horst Bischof, Thomas Brox, and Jan-Michael Frahm, editors, ECCV, pages 694-710, Cham, 2020. Springer International Publishing. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 262, + 286, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 262, + 286, + 296 + ], + "spans": [ + { + "bbox": [ + 53, + 262, + 286, + 296 + ], + "type": "text", + "content": "[6] Ishan Deshpande, Ziyu Zhang, and Alexander G Schwing. Generative modeling using the sliced Wasserstein distance. In CVPR, pages 3483-3491, 2018. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 297, + 286, + 340 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 297, + 286, + 340 + ], + "spans": [ + { + "bbox": [ + 53, + 297, + 286, + 340 + ], + "type": "text", + "content": "[7] Prafulla Dhariwal and Alexander Quinn Nichol. Diffusion models beat GANs on image synthesis. In A. Beygelzimer, Y. Dauphin, P. Liang, and J. Wortman Vaughan, editors, NeurIPS, 2021. 1, 2, 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 342, + 286, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 342, + 286, + 407 + ], + "spans": [ + { + "bbox": [ + 53, + 342, + 286, + 407 + ], + "type": "text", + "content": "[8] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020.5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 409, + 286, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 409, + 286, + 430 + ], + "spans": [ + { + "bbox": [ + 53, + 409, + 286, + 430 + ], + "type": "text", + "content": "[9] Didrik Nielsen Giorgio Giannone and Ole Winther. Few-shot diffusion models. arXiv preprint arXiv:2205.15463, 2021. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 433, + 286, + 486 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 433, + 286, + 486 + ], + "spans": [ + { + "bbox": [ + 48, + 433, + 286, + 486 + ], + "type": "text", + "content": "[10] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. In NeurIPS, page 6629-6640, Red Hook, NY, USA, 2017. Curran Associates Inc. 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 488, + 286, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 488, + 286, + 510 + ], + "spans": [ + { + "bbox": [ + 48, + 488, + 286, + 510 + ], + "type": "text", + "content": "[11] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In NeurIPS, 2020. 3, 4, 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 512, + 286, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 512, + 286, + 533 + ], + "spans": [ + { + "bbox": [ + 48, + 512, + 286, + 533 + ], + "type": "text", + "content": "[12] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598, 2022. 2, 3, 4, 8" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 535, + 286, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 535, + 286, + 567 + ], + "spans": [ + { + "bbox": [ + 48, + 535, + 286, + 567 + ], + "type": "text", + "content": "[13] Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan, Mohammad Norouzi, and David J Fleet. Video diffusion models. arXiv preprint arXiv:2204.03458, 2022. 1" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 569, + 286, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 569, + 286, + 656 + ], + "spans": [ + { + "bbox": [ + 48, + 569, + 286, + 656 + ], + "type": "text", + "content": "[14] Greg Holste, Song Wang, Ziyu Jiang, Thomas C. Shen, George L. Shih, Ronald M. Summers, Yifan Peng, and Zhangyang Wang. Long-tailed classification of thorax diseases on chest x-ray: A new benchmark study. Data augmentation, labelling, and imperfections : second MICCAI workshop, DALI 2022, held in conjunction with MICCAI 2022, Singapore, September 22, 2022, proceedings. DALI (Workshop), 13567:22-32, 2022. 1" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 658, + 286, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 286, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 286, + 712 + ], + "type": "text", + "content": "[15] Grant Van Horn, Oisin Mac Aodha, Yang Song, Yin Cui, Chen Sun, Alexander Shepard, Hartwig Adam, Pietro Perona, and Serge J. Belongie. The inaturalist species classification and detection dataset. In CVPR, pages 8769-8778, 2018. 1" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 32, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "text", + "content": "[16] Chen Huang, Yining Li, Chen Change Loy, and Xiaou Tang. Learning deep representation for imbalanced classification. In CVPR, pages 5375-5384, 2016. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 107, + 545, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 107, + 545, + 150 + ], + "spans": [ + { + "bbox": [ + 307, + 107, + 545, + 150 + ], + "type": "text", + "content": "[17] Myeonghun Jeong, Hyeongju Kim, Sung Jun Cheon, Byoung Jin Choi, and Nam Soo Kim. Diff-TTS: A Denoising Diffusion Model for Text-to-Speech. In Proc. Interspeech 2021, pages 3605-3609, 2021. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 152, + 545, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 152, + 545, + 195 + ], + "spans": [ + { + "bbox": [ + 307, + 152, + 545, + 195 + ], + "type": "text", + "content": "[18] Alexia Jolicoeur-Martineau, Remi Piche-Taillefer, Rémi Tachet des Combes, and Ioannis Mitliagkas. Adversarial score matching and improved sampling for image generation. arXiv preprint arXiv:2009.05475. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 197, + 545, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 197, + 545, + 229 + ], + "spans": [ + { + "bbox": [ + 307, + 197, + 545, + 229 + ], + "type": "text", + "content": "[19] Haeyong Kang, Thang Vu, and Chang D. Yoo. Learning imbalanced datasets with maximum margin loss. In ICIP, pages 1269-1273, 2021. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 231, + 545, + 263 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 231, + 545, + 263 + ], + "spans": [ + { + "bbox": [ + 307, + 231, + 545, + 263 + ], + "type": "text", + "content": "[20] Tero Karras, Miika Aittala, Timo Aila, and Samuli Laine. Elucidating the design space of diffusion-based generative models. arXiv preprint arXiv:2206.00364, 2022. 5, 7" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 265, + 545, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 265, + 545, + 308 + ], + "spans": [ + { + "bbox": [ + 307, + 265, + 545, + 308 + ], + "type": "text", + "content": "[21] Tero Karras, Miika Aittala, Janne Hellsten, Samuli Laine, Jaakko Lehtinen, and Timo Aila. Training generative adversarial networks with limited data. In NeurIPS, Red Hook, NY, USA, 2020. Curran Associates Inc. 1, 2, 5, 6, 7" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 309, + 545, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 309, + 545, + 354 + ], + "spans": [ + { + "bbox": [ + 307, + 309, + 545, + 354 + ], + "type": "text", + "content": "[22] Minsu Ko, Eunju Cha, Sungjoo Suh, Huijin Lee, Jae-Joon Han, Jinwoo Shin, and Bohyung Han. Self-supervised dense consistency regularization for image-to-image translation. In CVPR, pages 18280-18289, 2022. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 355, + 545, + 388 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 355, + 545, + 388 + ], + "spans": [ + { + "bbox": [ + 307, + 355, + 545, + 388 + ], + "type": "text", + "content": "[23] Tuomas Kynkänniemi, Tero Karras, Samuli Laine, Jaakko Lehtinen, and Timo Aila. Improved precision and recall metric for assessing generative models. NeurIPS, 32, 2019. 5" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 388, + 545, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 388, + 545, + 430 + ], + "spans": [ + { + "bbox": [ + 307, + 388, + 545, + 430 + ], + "type": "text", + "content": "[24] Wen Li, Limin Wang, Wei Li, Eirikur Agustsson, and Luc Van Gool. Webvision database: Visual learning and understanding from web data. arXiv preprint arXiv:1708.02862, 2017. 1" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 433, + 545, + 478 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 433, + 545, + 478 + ], + "spans": [ + { + "bbox": [ + 307, + 433, + 545, + 478 + ], + "type": "text", + "content": "[25] Jialun Liu, Yifan Sun, Chuchu Han, Zhaopeng Dou, and Wenhui Li. Deep representation learning on long-tailed data: A learnable embedding augmentation perspective. In CVPR, pages 2967-2976, 2020. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 479, + 545, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 479, + 545, + 511 + ], + "spans": [ + { + "bbox": [ + 307, + 479, + 545, + 511 + ], + "type": "text", + "content": "[26] Guanxiong Luo, Martin Heide, and Martin Uecker. MRI reconstruction via data driven markov chain with joint uncertainty estimation. CoRR, abs/2202.01479, 2022. 1" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 512, + 545, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 512, + 545, + 588 + ], + "spans": [ + { + "bbox": [ + 307, + 512, + 545, + 588 + ], + "type": "text", + "content": "[27] Dhruv Mahajan, Ross Girshick, Vignesh Ramanathan, Kaiming He, Manohar Paluri, Yixuan Li, Ashwin Bharambe, and Laurens van der Maaten. Exploring the limits of weakly supervised pretraining. In Vittorio Ferrari, Martial Hebert, Cristian Sminchisescu, and Yair Weiss, editors, ECCV, pages 185-201, Cham, 2018. Springer International Publishing. 1, 2, 5, 6" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 590, + 545, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 590, + 545, + 633 + ], + "spans": [ + { + "bbox": [ + 307, + 590, + 545, + 633 + ], + "type": "text", + "content": "[28] Chenlin Meng, Yang Song, Jiaming Song, Jiajun Wu, JunYan Zhu, and Stefano Ermon. Sdedit: Image synthesis and editing with stochastic differential equations. arXiv preprint arXiv:2108.01073, 2021. 1" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 307, + 635, + 545, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 635, + 545, + 678 + ], + "spans": [ + { + "bbox": [ + 307, + 635, + 545, + 678 + ], + "type": "text", + "content": "[29] Aditya Krishna Menon, Sadeep Jayasumana, Ankit Singh Rawat, Himanshu Jain, Andreas Veit, and Sanjiv Kumar. Long-tail learning via logit adjustment. In ICLR. OpenReview.net, 2021. 1, 2" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 307, + 680, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 680, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 307, + 680, + 545, + 712 + ], + "type": "text", + "content": "[30] Takeru Miyato, Toshiki Kataoka, Masanori Koyama, and Yuichi Yoshida. Spectral normalization for generative adversarial networks. In ICLR, 2018. 7" + } + ] + } + ], + "index": 31 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "18442" + } + ] + } + ], + "index": 33 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 171 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 171 + ], + "type": "text", + "content": "[31] Alexander Quinn Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. GLIDE: Towards photorealistic image generation and editing with text-guided diffusion models. In Kamalika Chaudhuri, Stefanie Jegelka, Le Song, Csaba Szepesvari, Gang Niu, and Sivan Sabato, editors, International Conference on Machine Learning (ICML), volume 162 of Proceedings of Machine Learning Research, pages 16784-16804. PMLR, 17-23 Jul 2022. 1, 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 172, + 287, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 172, + 287, + 215 + ], + "spans": [ + { + "bbox": [ + 48, + 172, + 287, + 215 + ], + "type": "text", + "content": "[32] Wanli Ouyang, Xiaogang Wang, Cong Zhang, and Xiaokang Yang. Factors in finetuning deep model for object detection with long-tail distribution. In CVPR, pages 864-873, 2016. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 217, + 287, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 217, + 287, + 249 + ], + "spans": [ + { + "bbox": [ + 48, + 217, + 287, + 249 + ], + "type": "text", + "content": "[33] Yuntian Gu Quanlin Wu, Hang Ye. Guided diffusion model for adversarial purification from random noise. arXiv preprint arXiv:2206.10875, 2022. 1, 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 250, + 287, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 250, + 287, + 315 + ], + "spans": [ + { + "bbox": [ + 48, + 250, + 287, + 315 + ], + "type": "text", + "content": "[34] Harsh Rangwani, Naman Jaswani, Tejan Karmali, Varun Jampani, and R. Venkatesh Babu. Improving gans for long-tailed data through group spectral regularization. In Shai Avidan, Gabriel Brostow, Moustapha Cisse, Giovanni Maria Farinella, and Tal Hassner, editors, ECCV, pages 426-442, Cham, 2022. Springer Nature Switzerland. 1, 5, 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 316, + 287, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 316, + 287, + 392 + ], + "spans": [ + { + "bbox": [ + 48, + 316, + 287, + 392 + ], + "type": "text", + "content": "[35] Harsh Rangwani, Konda Reddy Mopuri, and R. Venkatesh Babu. Class balancing gan with a classifier in the loop. In Cassio de Campos and Marloes H. Maathuis, editors, Proceedings of the Thirty-Seventh Conference on Uncertainty in Artificial Intelligence, volume 161 of Proceedings of Machine Learning Research, pages 1618-1627. PMLR, 27-30 Jul 2021. 5, 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 393, + 287, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 393, + 287, + 435 + ], + "spans": [ + { + "bbox": [ + 48, + 393, + 287, + 435 + ], + "type": "text", + "content": "[36] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, pages 10684-10695, 2022. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 437, + 287, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 437, + 287, + 491 + ], + "spans": [ + { + "bbox": [ + 48, + 437, + 287, + 491 + ], + "type": "text", + "content": "[37] Kevin Roth, Aurelien Lucchi, Sebastian Nowozin, and Thomas Hofmann. Stabilizing training of generative adversarial networks through regularization. In NeurIPS, page 2015-2025, Red Hook, NY, USA, 2017. Curran Associates Inc. 1, 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 492, + 287, + 545 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 492, + 287, + 545 + ], + "spans": [ + { + "bbox": [ + 48, + 492, + 287, + 545 + ], + "type": "text", + "content": "[38] Chitwan Sahara, William Chan, Huiwen Chang, Chris A. Lee, Jonathan Ho, Tim Salimans, David J. Fleet, and Mohammad Norouzi. Palette: Image-to-image diffusion models. ACM SIGGRAPH 2022 Conference Proceedings, 2022. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 548, + 287, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 548, + 287, + 590 + ], + "spans": [ + { + "bbox": [ + 48, + 548, + 287, + 590 + ], + "type": "text", + "content": "[39] Mehdi S. M. Sajjadi, Olivier Bachem, Mario Lucic, Olivier Bousquet, and Sylvain Gelly. Assessing generative models via precision and recall. In NeurIPS, page 5234-5243, Red Hook, NY, USA, 2018. Curran Associates Inc. 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 592, + 287, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 592, + 287, + 645 + ], + "spans": [ + { + "bbox": [ + 48, + 592, + 287, + 645 + ], + "type": "text", + "content": "[40] Tim Salimans, Ian Goodfellow, Wojciech Zaremba, Vicki Cheung, Alec Radford, Xi Chen, and Xi Chen. Improved techniques for training gans. In D. Lee, M. Sugiyama, U. Luxburg, I. Guyon, and R. Garnett, editors, NeurIPS, volume 29. Curran Associates, Inc., 2016. 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 647, + 287, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 647, + 287, + 690 + ], + "spans": [ + { + "bbox": [ + 48, + 647, + 287, + 690 + ], + "type": "text", + "content": "[41] Vikash Sehwag, Caner Hazirbas, Albert Gordo, First Ozgenel, and Cristian Canton. Generating high fidelity data from low-density regions using diffusion models. In CVPR, pages 11492-11501, 2022. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 692, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 692, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 692, + 287, + 713 + ], + "type": "text", + "content": "[42] Jascha Sohl-Dickstein, Eric A. Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 420 + ], + "type": "list", + "angle": 0, + "index": 22, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "text", + "content": "learning using nonequilibrium thermodynamics. arXiv preprint arXiv:1503.03585, 2015. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 307, + 96, + 545, + 118 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 96, + 545, + 118 + ], + "spans": [ + { + "bbox": [ + 307, + 96, + 545, + 118 + ], + "type": "text", + "content": "[43] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models, 2021. 2, 7" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 119, + 545, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 119, + 545, + 162 + ], + "spans": [ + { + "bbox": [ + 307, + 119, + 545, + 162 + ], + "type": "text", + "content": "[44] Yang Song, Liyue Shen, Lei Xing, and Stefano Ermon. Solving inverse problems in medical imaging with score-based generative models. arXiv preprint arXiv:2111.08005, 2022. 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 164, + 545, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 164, + 545, + 196 + ], + "spans": [ + { + "bbox": [ + 307, + 164, + 545, + 196 + ], + "type": "text", + "content": "[45] Hung-Yu Tseng, Lu Jiang, Ce Liu, Ming-Hsuan Yang, and Weilong Yang. Regularizing generative adversarial networks under limited data. In CVPR, pages 7917-7927, 2021. 2, 7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 198, + 545, + 241 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 198, + 545, + 241 + ], + "spans": [ + { + "bbox": [ + 307, + 198, + 545, + 241 + ], + "type": "text", + "content": "[46] Aaron van den Oord, Oriol Vinyals, and Koray Kavukcuoglu. Neural discrete representation learning. In NeurIPS, NIPS'17, page 6309-6318, Red Hook, NY, USA, 2017. Curran Associates Inc. 4" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 243, + 545, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 243, + 545, + 285 + ], + "spans": [ + { + "bbox": [ + 307, + 243, + 545, + 285 + ], + "type": "text", + "content": "[47] Jianfeng Wang, Thomas Lukasiewicz, Xiaolin Hu, Jianfei Cai, and Zhenghua Xu. Rsg: A simple but effective module for learning imbalanced datasets. In CVPR, pages 3783-3792, 2021. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 287, + 545, + 329 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 287, + 545, + 329 + ], + "spans": [ + { + "bbox": [ + 307, + 287, + 545, + 329 + ], + "type": "text", + "content": "[48] Xiulong Yang, Sheng-Min Shih, Yinlin Fu, Xiaoting Zhao, and Shihao Ji. Your vit is secretly a hybrid discriminative-generative diffusion model. ArXiv, abs/2208.07791, 2022. 5, 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 332, + 545, + 375 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 332, + 545, + 375 + ], + "spans": [ + { + "bbox": [ + 307, + 332, + 545, + 375 + ], + "type": "text", + "content": "[49] Xi Yin, Xiang Yu, Kihyuk Sohn, Xiaoming Liu, and Manmohan Chandraker. Feature transfer learning for face recognition with under-represented data. In CVPR, pages 5697-5706, 2019. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 376, + 545, + 420 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 376, + 545, + 420 + ], + "spans": [ + { + "bbox": [ + 307, + 376, + 545, + 420 + ], + "type": "text", + "content": "[50] Shengyu Zhao, Zhijian Liu, Ji Lin, Jun-Yan Zhu, and Song Han. Differentiable augmentation for data-efficient gan training. In NeurlPS, Red Hook, NY, USA, 2020. Curran Associates Inc. 1, 2, 5, 6" + } + ] + } + ], + "index": 21 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18443" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Class-Conditional Sharpness-Aware Minimization for Deep Long-Tailed Recognition/5537ac60-2d54-48e8-9091-a4f2d7a57e6f_content_list.json b/2023/Class-Conditional Sharpness-Aware Minimization for Deep Long-Tailed Recognition/5537ac60-2d54-48e8-9091-a4f2d7a57e6f_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..0a941cd8d382c16b5dec18eab8d2eca51b5892dd --- /dev/null +++ b/2023/Class-Conditional Sharpness-Aware Minimization for Deep Long-Tailed Recognition/5537ac60-2d54-48e8-9091-a4f2d7a57e6f_content_list.json @@ -0,0 +1,1864 @@ +[ + { + "type": "text", + "text": "Class-Conditional Sharpness-Aware Minimization for Deep Long-Tailed Recognition", + "text_level": 1, + "bbox": [ + 122, + 161, + 848, + 204 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Zhipeng Zhou $^{1,\\dagger}$ , Lanqing Li $^{2,4,\\dagger}$ , Peilin Zhao $^{3,\\star}$ , Pheng-Ann Heng $^{2}$ , Wei Gong $^{1,\\star}$", + "bbox": [ + 122, + 238, + 843, + 258 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1University of Science and Technology of China, 2The Chinese University of Hong Kong, 3Tencent AI Lab, 4Zhejiang Lab", + "bbox": [ + 121, + 266, + 855, + 300 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "zzpl1994@mail.ustc.edu.cn, lanqingli1993@gmail.com, masonzhao@tencent.com,", + "bbox": [ + 159, + 302, + 806, + 315 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "pheng@cse.cuhk.edu.hk, weigong@ustc.edu.cn", + "bbox": [ + 295, + 319, + 672, + 333 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 372, + 312, + 387 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "It's widely acknowledged that deep learning models with flatter minima in its loss landscape tend to generalize better. However, such property is under-explored in deep long-tailed recognition (DLTR), a practical problem where the model is required to generalize equally well across all classes when trained on highly imbalanced label distribution. In this paper, through empirical observations, we argue that sharp minima are in fact prevalent in deep long-tailed models, whereas naive integration of existing flattening operations into long-tailed learning algorithms brings little improvement. Instead, we propose an effective two-stage sharpness-aware optimization approach based on the decoupling paradigm in DLTR. In the first stage, both the feature extractor and classifier are trained under parameter perturbations at a class-conditioned scale, which is theoretically motivated by the characteristic radius of flat minima under the PAC-Bayesian framework. In the second stage, we generate adversarial features with class-balanced sampling to further robustify the classifier with the backbone frozen. Extensive experiments on multiple long-tailed visual recognition benchmarks show that, our proposed Class-Conditional Sharpness-Aware Minimization (CC-SAM), achieves competitive performance compared to the state-of-the-arts. Code is available at https://github.com/zzpustc/CC-SAM.", + "bbox": [ + 75, + 401, + 472, + 748 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 77, + 771, + 208, + 786 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Modern deep learning models, composed of multiple neural network layers with millions of parameters, have achieved remarkable successes in computer vision [24, 33,", + "bbox": [ + 75, + 794, + 468, + 837 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "† Equal contribution. * Corresponding authors. Work was primarily done when Z. Zhou worked as L. Li's intern at Tencent AI Lab, Shenzhen.", + "bbox": [ + 75, + 843, + 468, + 866 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "41, 48]. A key enabler of deep learning is the collection of large-scale datasets [29, 42, 64], which are normally split into training and testing sets with presumably i.i.d. samples. However, such scenario provides relatively trivial tests for the generalization of machine learning models. In practice, label [17, 25, 56] and domain [14, 18, 23] distribution shifts are prevalent, due to the disparity between the data preparation and evaluation protocols. A classical example is imbalanced [15] or long-tailed recognition [60], where a model is trained on highly imbalanced source label distribution $p_{s}(\\pmb{y})$ while evaluated on a uniform target label distribution $p_{t}(\\pmb{y})$ .", + "bbox": [ + 497, + 372, + 890, + 541 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this paper, we focus on the practical yet challenging deep long-tailed recognition (DLTR) problem, which is inherent in the visual world [32, 60] with fundamental connections to many disciplines such as the power-law scaling in network science [2] and the Pareto principle in economics [39]. In computer vision, numerous deep long-tailed learning studies have emerged in recent years, which mainly belong to 5 categories: class rebalancing [7, 10, 28, 40, 45, 52, 54], information augmentation [22, 27, 31, 50, 55], decoupled training [20, 58, 62], representation learning [9, 32, 51, 59, 65] and ensemble learning [6, 53, 63].", + "bbox": [ + 497, + 543, + 892, + 710 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this work, we propose a novel approach to DLTR from a distinct angle, by seeking out flat minima in the loss landscape of modern neural networks to ensure model robustness under parameter perturbation. Such optimization strategy, termed flattening in our context, have been shown in a myriad of literature to effectively improve generalization of deep learning models in terms of supervised learning [13, 21, 35, 43], self-supervised learning [30] and continual learning [11, 44]. However, application and adaptation of flattening in the context of DLTR remain under-explored. To fill this gap, we first show later in this paper", + "bbox": [ + 497, + 713, + 894, + 866 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 127, + 49, + 200, + 84 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 252, + 48, + 791, + 87 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "3499", + "bbox": [ + 468, + 885, + 502, + 897 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "(Section 2.2.2) that existing flattening methods are ineffective for long-tailed learning, consistent with the observation from a very recent paper [49], due to severe label distribution shifts. Accordingly, we present a new efficient variant of the sharpness-aware minimization (SAM) [13] technique based on the Decoupling paradigm [20] of DLTR, which leverages the invariance of the class conditional distribution between the source and the target domain. In a nutshell, our contributions are three-fold:", + "bbox": [ + 75, + 125, + 472, + 249 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Through the lens of flattening, we corroborate a very recent observation [49] that existing sharpness-aware minimization techniques are suboptimal for deep long-tailed learning, justifying the need of more effective approaches to this important and practical issue.", + "- We introduce Class-Conditional Sharpness-Aware Minimization (CC-SAM), a novel algorithm tailored for DLTR that improves model generalization by robust training against class-conditioned parameter perturbation. This technique is motivated by the characteristic radius of flat minima we derive under the PAC-Bayesian framework and can be implemented efficiently.", + "- We integrate CC-SAM with the two-stage decoupling paradigm of DLTR. Extensive experiments demonstrate that our method can achieve competitive performance on multiple public DLTR benchmarks, with remarkable robustness to out-of-distribution samples for open long-tailed recognition [32]." + ], + "bbox": [ + 95, + 262, + 468, + 546 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Preliminaries", + "text_level": 1, + "bbox": [ + 77, + 559, + 215, + 572 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Problem Setup", + "text_level": 1, + "bbox": [ + 77, + 581, + 226, + 596 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Throughout the paper, we denote scalars as $s$ , vectors as $\\mathbf{s}$ , sets as $\\mathcal{S}$ and equality by definition as :=. For a typical $d$ -way classification task in DLTR, suppose we are given a training dataset $\\mathcal{S} = \\bigcup_{i=1}^{n} \\{(x_i, y_i)\\}$ , where $n$ is the total number of samples. Let's denote $\\{n_1, n_2, \\ldots, n_d\\}$ as the sample number of each class, where $n = \\sum_{i} n_i$ . Without loss of generality, we assume $n_i < n_j$ if $i < j$ , and usually $n_d \\gg n_1$ , following a highly imbalanced class distribution. We further write the training (source) distribution as $p_s(\\boldsymbol{x}, \\boldsymbol{y}) = p_s(\\boldsymbol{x}|\\boldsymbol{y}) p_s(\\boldsymbol{y})$ and the testing (target) distribution as $p_t(\\boldsymbol{x}, \\boldsymbol{y}) = p_t(\\boldsymbol{x}|\\boldsymbol{y}) p_t(\\boldsymbol{y})$ . Consider a family of models parameterized by $\\boldsymbol{w} \\in \\mathcal{W} \\subseteq \\mathbb{R}^k$ ; given a loss function $l\\colon \\mathcal{W} \\times \\mathcal{X} \\times \\mathcal{Y} \\to \\mathbb{R}_+$ , we define the empirical training loss $L_{\\mathcal{S}}(\\boldsymbol{w}) := \\frac{1}{n} \\sum_{i=1}^{n} l(\\boldsymbol{w}, \\boldsymbol{x}_i, \\boldsymbol{y}_i)$ and the population testing loss $L_{\\mathcal{T}}(\\boldsymbol{w}) := \\mathbb{E}_{(\\boldsymbol{x}, \\boldsymbol{y}) \\sim p_t(\\boldsymbol{x}, \\boldsymbol{y})} [l(\\boldsymbol{w}, \\boldsymbol{x}, \\boldsymbol{y})]$ . Having observed only $\\mathcal{S}$ , the goal is to optimize for model parameters $\\boldsymbol{w}$ having lowest risk $L_{\\mathcal{T}}(\\boldsymbol{w})$ at test time.", + "bbox": [ + 75, + 603, + 468, + 838 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In DLTR, a key challenge is the label distribution shift between the training and testing sets, i.e., $p_{s}(\\pmb{y}) \\neq$", + "bbox": [ + 75, + 838, + 468, + 866 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/097dad271cd8c36d9333a4d2611060f1c097d69889acdcb19d328b6d8b747e71.jpg", + "image_caption": [ + "(a) 1D global loss landscape." + ], + "image_footnote": [], + "bbox": [ + 500, + 124, + 684, + 231 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/b04f3c48a2472bb0ac3d7f3f578e5fcf36da42dad6729f005c8049c354002a3c.jpg", + "image_caption": [ + "(b) 1D local loss landscape.", + "Figure 1. Loss value vs. noise norm ratio $||\\epsilon ||_2 / ||\\pmb {\\theta}||_2$ . All experiments are conducted on the CIFAR-10-LT with an imbalance ratio of 100, implemented with the same backbone ResNet-32 and averaged over 5 random seeds." + ], + "image_footnote": [], + "bbox": [ + 700, + 124, + 889, + 231 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "$p_t(\\boldsymbol{y})$ [17]. However, it's logical to assume no distribution shift within each individual class. That is, the training and testing samples are drawn i.i.d. from the same class-conditional distribution, which indicates $p_s(\\boldsymbol{x}|\\boldsymbol{y}) = p_t(\\boldsymbol{x}|\\boldsymbol{y})$ . Here, we build our flattening algorithm upon the decoupling paradigm of DLTR [20], where the feature extractor $f(\\boldsymbol{x};\\boldsymbol{\\varphi})$ and classifier $h(\\boldsymbol{z};\\boldsymbol{\\theta})$ are trained by distinct sampling strategies in a two-stage manner, with $\\boldsymbol{w} \\coloneqq (\\varphi, \\boldsymbol{\\theta})$ .", + "bbox": [ + 497, + 332, + 890, + 457 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Motivation", + "text_level": 1, + "bbox": [ + 499, + 465, + 621, + 479 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Previous studies have shown strong connection between the geometry of the loss landscape and generalization of deep learning models [1, 16, 19, 21]. In this section, we first empirically demonstrate that many deep long-tailed models exhibit sharp minima, and then show that naive application of classical flattening operations results in limited improvement. Motivated by these observations, we theoretically analyze the characteristic radius of flat minima in DLTR to show that a class-conditional flattening procedure is required.", + "bbox": [ + 497, + 486, + 892, + 625 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2.1 Sharpness in Deep Long-Tailed Models", + "text_level": 1, + "bbox": [ + 499, + 643, + 825, + 657 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To show the pervasive sharpness in current DLTR models, we select two baselines (empirical risk minimization of cross-entropy (CE), LDAM-DRW [7]) and two advanced methods (MiSLAS [62], GCL [26]) from recent literature to conduct experiments1. By perturbing the parameter $\\theta$ of the classifier $h$ with increasing noise $\\epsilon$ , we observe the loss of each model on CIFAR-10-LT in Figure 1.", + "bbox": [ + 497, + 663, + 890, + 761 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Figure 1(a) shows that, at global scale, the loss of CE is the sharpest across a wide noise range. However, due to the fact that DLTR methods usually employ distinct loss functions, it's hard to compare the sharpness of their loss landscapes in a fair and meaningful way. Moreover, large noise ratio can easily flip the sign of parameters and deteriorate", + "bbox": [ + 497, + 761, + 890, + 845 + ], + "page_idx": 1 + }, + { + "type": "page_footnote", + "text": "1We re-implement all models via their publicly released code.", + "bbox": [ + 515, + 853, + 845, + 865 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "3500", + "bbox": [ + 468, + 885, + 500, + 897 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/2efcd9cad09ebde024e5a1e3177fc0bee2e08ee77876811779f0a30cee234875.jpg", + "image_caption": [ + "Figure 2. Performance gain of naive integration of flattening operations with DLTR algorithms. The number represents the stage where flattening applies. For example, \"SN.1.2\" means we apply spectral normalization on both stages. Only the first-stage result of LDAM is reported since it's a one-stage approach. All experiments are conducted on CIFAR-100-LT with imbalance ratio of 100." + ], + "image_footnote": [], + "bbox": [ + 90, + 122, + 453, + 251 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "the classifier decision boundary. Therefore in this paper, we only consider locally flat minima, i.e., in the perturbative $(||\\epsilon ||_2\\ll ||\\boldsymbol {w}||_2)$ regime shown in Figure 1(b). From this local view, only MiSLAS is observed to have a comparable or flatter minimum than CE, suggesting that the existing DLTR methods have poor generalization and robustness against model parameter perturbation. More detailed 2D loss landscapes covering overall, head and tail classes are depicted in Appendix E, and corroborate our observation in 1D.", + "bbox": [ + 75, + 374, + 468, + 511 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2.2 Naive Integration of Flattening Procedures", + "text_level": 1, + "bbox": [ + 75, + 530, + 428, + 543 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To improve the generalization of deep long-tailed models, we experiment to navigate flatter minima by integrating the existing flattening operations (stochastic weight averaging (SWA) [16], spectral normalization (SN) [4], gradients penalization (GP) [61], and model perturbation (MP) [44]) into five representative deep long-tailed learning baselines to see if they are generally beneficial for DLTR. Illustrated in Figure 2, these flattening methods bring little or negative gains in most cases. Even though in some cases they are beneficial, the corresponding baselines (mainly LDAM-DRW) are too weak so that the final performance are still sub-optimal, which calls for more effective flattening procedures for DLTR. More implementation details are provided in the Appendix C.", + "bbox": [ + 75, + 551, + 468, + 745 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2.3 Characteristic Radius of Flat Minima", + "text_level": 1, + "bbox": [ + 75, + 761, + 393, + 774 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Let's re-consider flattening operations in DLTR from a theoretical perspective. The empirical loss $L_{S}(\\boldsymbol{w})$ is typically non-convex for deep neural network models, whose landscape may exhibit multiple local or global minima with similar values of $L_{S}(\\boldsymbol{w})$ while having drastically different generalization performance (i.e., significantly different values", + "bbox": [ + 75, + 783, + 468, + 866 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "of $L_{\\mathcal{T}}(\\boldsymbol{w})$ . In order to find a more effective approach for generalization in DLTR, motivated by the high correlation between sharpness of the loss landscape and model generalization [19], we follow the sharpness-aware minimization (SAM) framework [13] by optimizing an upper bound of $L_{\\mathcal{T}}(\\boldsymbol{w})$ derived from the PAC-Bayesian framework [34]. Given a prior over the parameters before observing any data and a posterior over the parameters dependent on the training set and learning algorithm, the PAC-Bayesian framework bounds the generalization error of any model in terms of the KL divergence between the two probability distributions. Assuming a Gaussian prior and posterior, we arrive at the following generalization bound (formal proof in Appendix D):", + "bbox": [ + 497, + 125, + 892, + 319 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Theorem 1 (Perturbative PAC-Bayesian Generalization Bound). For any $\\rho >0,0 < \\delta < 1$ , number of samples $n\\in \\mathbb{N}^+$ , $k\\coloneqq \\dim (\\boldsymbol {w})$ , with probability at least $1 - \\delta$ over a training set $\\mathcal{S}$ sampled i.i.d. from distribution $\\mathcal{T}$", + "bbox": [ + 497, + 330, + 892, + 385 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nL_{\\mathcal{T}}(\\boldsymbol {w})\\leq \\max_{||\\boldsymbol {\\epsilon}||_{2}\\leq \\sqrt{k}\\rho}L_{\\mathcal{S}}(\\boldsymbol {w} + \\boldsymbol {\\epsilon}) + \\sqrt{\\frac{\\|\\boldsymbol{w}\\|_{2}^{2}}{4\\rho^{2}} + \\log(\\frac{n}{\\delta}) + \\mathcal{O}(1)}\\\\ \\hskip 142.26378ptn - 1.\n$$\n", + "text_format": "latex", + "bbox": [ + 499, + 407, + 899, + 447 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Intuitively, optimizing the perturbative bound as a surrogate function effectively seeks out parameter solutions whose neighborhoods have uniformly low empirical training loss, thus achieving a flat minimum. Now, our key insight is that the bound is convex with respect to the radius $\\rho$ of the flat minima, hence there exists an optimal $\\rho^{*}$ which gives the tightest generalization bound:", + "bbox": [ + 497, + 456, + 892, + 552 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\rho^ {*} := \\arg \\min _ {\\rho} \\left[ \\max _ {| | \\boldsymbol {\\epsilon} | | _ {2} \\leq \\sqrt {k} \\rho} L _ {\\mathcal {S}} (\\boldsymbol {w} + \\boldsymbol {\\epsilon}) + \\sqrt {\\frac {\\frac {| | \\boldsymbol {w} | | _ {2} ^ {2}}{4 \\rho^ {2}} + \\log \\left(\\frac {n}{\\delta}\\right)}{n - 1}} \\right]. \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 497, + 562, + 910, + 619 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We denote $\\rho^{*}$ as the characteristic radius of the flat minima centered at $w$ . It's worth mentioning that such optimal perturbation radius is overlooked by previous literature on sharpness-aware optimization, mainly due to arguments that the PAC-Bayesian bound is too loose to capture the real generalization error [35]. However, our Theorem 1 inspired from [8, 12] ensures a non-vacuous bound (i.e., the square root term falls below 1, see empirical evidence in Appendix B) in the over-parameterized \"deep learning\" regime, rendering the characteristic radius relevant for optimization.", + "bbox": [ + 497, + 619, + 890, + 757 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Moreover, the perturbative bound can be approximated by a first-order Taylor expansion, yielding", + "bbox": [ + 497, + 757, + 890, + 786 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\max _ {\\| \\boldsymbol {\\epsilon} \\| _ {2} \\leq \\sqrt {k} \\rho} L _ {\\mathcal {S}} (\\boldsymbol {w} + \\boldsymbol {\\epsilon}) \\approx \\max _ {\\| \\boldsymbol {\\epsilon} \\| _ {2} \\leq \\sqrt {k} \\rho} \\left[ L _ {\\mathcal {S}} (\\boldsymbol {w}) + \\boldsymbol {\\epsilon} ^ {T} \\nabla_ {\\boldsymbol {w}} L _ {\\mathcal {S}} (\\boldsymbol {w}) \\right] \\\\ = L _ {\\mathcal {S}} (\\boldsymbol {w}) + \\sqrt {k} \\rho | | \\nabla_ {\\boldsymbol {w}} L _ {\\mathcal {S}} (\\boldsymbol {w}) | | _ {2} \\tag {2} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 500, + 807, + 899, + 866 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3501", + "bbox": [ + 468, + 885, + 500, + 897 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/16752340528c6eccc9ada79837d80a2dbbe5fcb032792dafd304ada06f4e15d3.jpg", + "image_caption": [ + "(a) Stage 1: Class-conditional sharpness-aware minimization. We take three classes and classifier perturbation as examples for illustration." + ], + "image_footnote": [], + "bbox": [ + 77, + 122, + 505, + 254 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/11f0154026619240555f56b110bab0278f9a51eafa90e8353aa96b036bbf8ec5.jpg", + "image_caption": [ + "(b) Stage 2: Robust training of the classifier by progressively generating adversarial features.", + "Figure 3. The overall framework of CC-SAM." + ], + "image_footnote": [], + "bbox": [ + 527, + 140, + 894, + 255 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "and the optimal perturbation vector", + "bbox": [ + 77, + 328, + 312, + 342 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\hat{\\boldsymbol{\\epsilon}}^{*}(\\boldsymbol {w})\\approx \\operatorname *{arg max}_{||\\boldsymbol {\\epsilon}||_{2}\\leq \\sqrt{k}\\rho}\\left[L_{\\mathcal{S}}(\\boldsymbol {w}) + \\boldsymbol{\\epsilon}^{T}\\nabla_{\\boldsymbol{w}}L_{\\mathcal{S}}(\\boldsymbol {w})\\right] \\\\ = \\sqrt {k} \\rho \\frac {\\nabla_ {\\boldsymbol {w}} L _ {\\mathcal {S}} (\\boldsymbol {w})}{\\| \\nabla_ {\\boldsymbol {w}} L _ {\\mathcal {S}} (\\boldsymbol {w}) \\| _ {2}}. \\tag {3} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 117, + 361, + 468, + 422 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Substitute Eqn 2 into 1, and ignore $\\log \\left(\\frac{n}{\\delta}\\right)$ by assuming $\\rho \\ll ||\\pmb{w}||_2$ as we discussed in Sec 2.2.1, obtaining", + "bbox": [ + 77, + 428, + 468, + 456 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\rho^ {*} \\approx \\left(\\frac {\\| \\boldsymbol {w} \\| _ {2}}{2 \\| \\nabla_ {\\boldsymbol {w}} L _ {\\mathcal {S}} (\\boldsymbol {w}) \\| _ {2}}\\right) ^ {\\frac {1}{2}} k ^ {- \\frac {1}{4}} (n - 1) ^ {- \\frac {1}{4}} \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 126, + 473, + 468, + 508 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "with the approximated optimal generalization bound", + "bbox": [ + 77, + 514, + 425, + 527 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\widehat {L} _ {\\mathcal {T}} (\\boldsymbol {w}) \\approx \\max _ {| | \\boldsymbol {\\epsilon} | | _ {2} \\leq \\sqrt {k} \\rho^ {*}} L _ {\\mathcal {S}} (\\boldsymbol {w} + \\boldsymbol {\\epsilon}) + \\frac {1}{2 \\sqrt {n - 1}} \\cdot \\frac {| | \\boldsymbol {w} | | _ {2}}{\\rho^ {*}}. \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 84, + 543, + 468, + 576 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The first term captures the sharpness of $L_{\\mathcal{S}}$ at $\\pmb{w}$ , and the second term serves as a regularizer on the magnitude of $\\pmb{w}$ akin to the L2 regularization in SAM [13]. A key difference here is that our regularization term is well motivated by theoretic interpretation from the PAC-Bayesian bound.", + "bbox": [ + 75, + 581, + 468, + 650 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3. Methodology", + "text_level": 1, + "bbox": [ + 77, + 662, + 211, + 678 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this section, we introduce our two-stage sharpness-aware optimization algorithm based on Decoupling [20] in detail. In the first stage of the decoupled training, both the feature extractor and classifier are trained under parameter perturbations at a class-conditioned scale. In the second stage, we generate adversarial features to further robustify the classifier while freezing the backbone. The complete pseudo-code is shown in Algorithm 1.", + "bbox": [ + 75, + 686, + 468, + 796 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1. Stage 1: Class-Conditional Sharpness-Aware Minimization (CC-SAM)", + "text_level": 1, + "bbox": [ + 77, + 802, + 468, + 830 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Note that a key assumption we made in Theorem 1 when deriving the characteristic radius of flat minima is that the", + "bbox": [ + 77, + 838, + 468, + 866 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "training data $S$ are drawn i.i.d. from the target distribution $\\mathcal{T}$ . However, as we explained in Sec 2.1, such assumption is invalid due to the severe label distribution shift between training and testing sets in DLTR. To resolve this fundamental conflict, we turn to the class conditional distribution $p_{s}(\\pmb{x}|\\pmb{y})$ and $p_{t}(\\pmb{x}|\\pmb{y})$ instead, for which the i.i.d. assumption reasonably holds. Accordingly, we decompose the total loss $L_{\\mathcal{T}}(\\pmb{w})$ and derive a generalization bound for each class separately, obtaining", + "bbox": [ + 497, + 328, + 890, + 453 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nL _ {\\mathcal {T}} (\\boldsymbol {w}) = \\sum_ {c = 1} ^ {k} L _ {\\mathcal {T}} ^ {c} (\\boldsymbol {w}) = \\sum_ {c = 1} ^ {k} \\mathbb {E} _ {\\boldsymbol {x} \\sim p _ {t} (\\boldsymbol {x} | c)} [ l (\\boldsymbol {w}, \\boldsymbol {x}, c) ], \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 509, + 461, + 890, + 500 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\widehat {L} _ {\\mathcal {T}} ^ {c} (\\boldsymbol {w}) = \\max _ {| | \\boldsymbol {\\epsilon} | | _ {2} \\leq \\sqrt {k} \\rho_ {c} ^ {*}} L _ {\\mathcal {S}} ^ {c} (\\boldsymbol {w} + \\boldsymbol {\\epsilon}) + \\frac {1}{2 \\sqrt {n - 1}} \\cdot \\frac {| | \\boldsymbol {w} | | _ {2}}{\\rho_ {c} ^ {*}}, \\\\ \\approx \\left(2 \\| \\boldsymbol {w} \\| _ {2} \\| \\nabla_ {w} L _ {\\mathcal {S}} ^ {c} (\\boldsymbol {w}) \\| _ {2}\\right) ^ {\\frac {1}{2}} k ^ {\\frac {1}{4}} \\left(n _ {c} - 1\\right) ^ {- \\frac {1}{4}} \\tag {7} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 510, + 501, + 890, + 555 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\rho_ {c} ^ {*} = \\left(\\frac {\\| \\boldsymbol {w} \\| _ {2}}{2 \\| \\nabla_ {w} L _ {\\mathcal {S}} ^ {c} (\\boldsymbol {w}) \\| _ {2}}\\right) ^ {\\frac {1}{2}} k ^ {- \\frac {1}{4}} \\left(n _ {c} - 1\\right) ^ {- \\frac {1}{4}}, \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 542, + 557, + 890, + 591 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\boldsymbol {\\epsilon}} _ {c} ^ {*} (\\boldsymbol {w}) \\approx \\sqrt {k} p _ {c} ^ {*} \\frac {\\nabla_ {w} L _ {\\mathcal {S}} ^ {c} (\\boldsymbol {w})}{\\| \\nabla_ {w} L _ {\\mathcal {S}} ^ {c} (\\boldsymbol {w}) \\| _ {2}}. \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 515, + 593, + 890, + 624 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "An immediate observation of Eqn 8 is that the characteristic radius $\\rho_c^*$ is class-dependent, more specifically, negatively correlated with the label frequency $n_c$ . One possible interpretation is that the model is more confident regarding the majority classes with higher $n_c$ , due to non-asymptotic estimation error given limited samples, hence it requires a smaller perturbative region in the parameter space to ensure flat loss landscape with respect to that class. Additionally, the class-wise generalization bound $\\widehat{L}_{\\mathcal{T}}^{c}(\\boldsymbol{w})$ in Eqn 7 is positively related to the class-wise gradient norm $\\|\\nabla_{\\boldsymbol{w}} L_{\\mathcal{S}}^{c}(\\boldsymbol{w})\\|_2$ , which effectively enforces hard example mining.", + "bbox": [ + 497, + 650, + 890, + 817 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "It follows that optimizing the approximated generalization bound $\\widehat{L}_{\\mathcal{T}}(\\boldsymbol{w}) = \\sum_{c=1}^{k} \\widehat{L}_{\\mathcal{T}}^{c}(\\boldsymbol{w})$ yields an effective algorithm in the first stage, namely Class-Conditional", + "bbox": [ + 497, + 823, + 890, + 866 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "3502", + "bbox": [ + 468, + 885, + 500, + 897 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Algorithm 1: Training Paradigm of CC-SAM", + "text_level": 1, + "bbox": [ + 87, + 128, + 391, + 142 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Input: Training Dataset", + "bbox": [ + 95, + 144, + 260, + 157 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {S} \\sim p _ {s} (\\boldsymbol {x}, \\boldsymbol {y}) = p _ {s} (\\boldsymbol {x} | \\boldsymbol {y}) p _ {s} (\\boldsymbol {y})\n$$\n", + "text_format": "latex", + "bbox": [ + 144, + 159, + 352, + 173 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Output: Model trained with CC-SAM", + "bbox": [ + 95, + 173, + 354, + 186 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Stage 1:", + "bbox": [ + 95, + 191, + 156, + 204 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Initialize $\\pmb{w} = \\{\\varphi, \\theta\\}$ randomly", + "bbox": [ + 95, + 205, + 310, + 219 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "while not converged do", + "bbox": [ + 95, + 219, + 257, + 232 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "foreach batch $\\mathcal{B}_i$ in $S$ do", + "bbox": [ + 122, + 233, + 292, + 245 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Compute empirical loss $L_{S}$ with $\\mathcal{B}_i$", + "bbox": [ + 147, + 245, + 383, + 258 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Estimate the class-specific gradient set", + "bbox": [ + 147, + 258, + 401, + 273 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "$G = \\{g_1,g_2,\\dots,g_k\\}$ with respect to $L_{S}$", + "bbox": [ + 154, + 273, + 421, + 287 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Perturb $\\pmb{w}$ with $G$ according to Eqn 8 and Eqn 9", + "bbox": [ + 149, + 287, + 423, + 314 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Update $\\pmb{w}$ via Eqn 10 and Eqn 11", + "bbox": [ + 149, + 314, + 369, + 328 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Stage 2:", + "bbox": [ + 95, + 333, + 156, + 346 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Freeze $\\varphi$", + "bbox": [ + 95, + 347, + 161, + 359 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "while not converged do", + "bbox": [ + 95, + 361, + 255, + 374 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Sample batches via class-balanced sampler", + "bbox": [ + 122, + 374, + 406, + 387 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {S} ^ {\\prime} = \\left\\{\\mathcal {B} _ {1} ^ {\\prime}, \\mathcal {B} _ {2} ^ {\\prime}, \\dots , \\mathcal {B} _ {m} ^ {\\prime} \\right\\} \\sim p _ {s} (\\boldsymbol {x} | \\boldsymbol {y}) \\cdot \\operatorname {U n i f o r m} (\\boldsymbol {y})\n$$\n", + "text_format": "latex", + "bbox": [ + 131, + 387, + 442, + 401 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "foreach batch $\\mathcal{B}_j^\\prime$ in $S^{\\prime}$ do", + "bbox": [ + 122, + 401, + 297, + 413 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Computing empirical loss $L_{\\mathcal{S}}$ with $\\mathcal{B}_i^\\prime$", + "bbox": [ + 147, + 413, + 396, + 427 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Obtain the adversarial feature via Eqn 12", + "bbox": [ + 147, + 427, + 418, + 440 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Evaluate overall loss according to Eqn 13 and Eqn 14", + "bbox": [ + 147, + 441, + 421, + 467 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Update $\\theta$ accordingly", + "bbox": [ + 147, + 467, + 292, + 482 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Sharpness-Aware Minimization (CC-SAM):", + "bbox": [ + 75, + 524, + 371, + 539 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Given learning rate $\\eta$", + "bbox": [ + 87, + 555, + 236, + 570 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {w} \\leftarrow \\boldsymbol {w} - \\eta \\nabla_ {\\boldsymbol {w}} L _ {\\mathcal {S}} ^ {\\mathrm {C C - S A M}} (\\boldsymbol {w}) \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 174, + 572, + 468, + 590 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\approx \\boldsymbol {w} - \\eta \\sum_ {c = 1} ^ {k} \\nabla_ {\\boldsymbol {w}} \\widehat {L} _ {\\mathcal {T}} ^ {c} (\\boldsymbol {w}) | _ {\\boldsymbol {w} + \\hat {\\epsilon} _ {c} ^ {*} (\\boldsymbol {w})}, \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 194, + 593, + 468, + 630 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "which is computationally efficient since it only involves first-order gradients. Even when $\\boldsymbol{w}$ is close to the optimal $\\boldsymbol{w}^*$ , where $\\nabla_{\\boldsymbol{w}}L_{S}(\\boldsymbol{w})|_{\\boldsymbol{w}^*} \\approx 0$ , the gradient for each class $\\nabla_{\\boldsymbol{w}}L_{S}^{c}(\\boldsymbol{w})|_{\\boldsymbol{w}^*}$ for estimating the optimal perturbation vector $\\hat{\\epsilon}_{c}^{*}(\\boldsymbol{w})$ in Eqn 9 is most likely far from zero, circumventing the need for computing high-order terms.", + "bbox": [ + 75, + 646, + 468, + 730 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2. Stage 2: Robust Training of the Classifier", + "text_level": 1, + "bbox": [ + 77, + 745, + 430, + 760 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For the second stage, we freeze the backbone to maintain feature representation and concentrate on refining the decision boundary of the classifier. Following Decoupling, we adopt the class-balanced sampling strategy in this stage, to rectify the classifier with uniform label distribution. As depicted in the Figure 3(b), when taking original feature $z$ as the input of classifier $h(z; \\theta)$ , we can generate adversarial", + "bbox": [ + 75, + 769, + 468, + 866 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "features in forward pass as follows:", + "bbox": [ + 499, + 125, + 734, + 138 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nz _ {a d v} = z + \\lambda \\frac {\\nabla_ {z} L _ {\\mathcal {S}} (z ; \\theta)}{| | \\nabla_ {z} L _ {\\mathcal {S}} (z ; \\theta) | | _ {2}} \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 589, + 150, + 890, + 181 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nL _ {\\mathcal {A}} = h \\left(\\boldsymbol {z} _ {\\mathbf {a d v}}; \\boldsymbol {\\theta}\\right) \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 635, + 200, + 890, + 214 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\lambda$ is the hyper-parameter to scale the adversarial gradient $\\nabla_{z}L_{\\mathcal{S}}(z;\\pmb {\\theta}) / ||\\nabla_{z}L_{\\mathcal{S}}(z;\\pmb {\\theta})||_{2}$ . Moreover, we employ a progressive strategy to balance $L_{\\mathcal{S}}$ and $L_{\\mathcal{A}}$ . At epoch $t$ ,", + "bbox": [ + 497, + 223, + 890, + 266 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nL = \\left(1 - \\frac {t}{T}\\right) L _ {\\mathcal {S}} + \\frac {t}{T} L _ {\\mathcal {A}} \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 608, + 275, + 890, + 302 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $T$ is the total number of epochs in this stage, and $L_{\\mathcal{A}}$ dominates the training progressively.", + "bbox": [ + 497, + 311, + 890, + 339 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Evaluation", + "text_level": 1, + "bbox": [ + 499, + 353, + 615, + 366 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this section, we evaluate our method on multiple mainstream DLTR datasets and compare it with popular baselines including the state-of-the-art (SOTA) methods. Moreover, we also report the result on open-set recognition [32, 47] tasks to demonstrate the excellent robustness of CC-SAM to out-of-distribution samples. In the end, an ablation study is presented to verify the effectiveness of our design choices.", + "bbox": [ + 497, + 377, + 890, + 488 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Datasets and Baselines", + "text_level": 1, + "bbox": [ + 499, + 498, + 709, + 511 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Following the mainstream evaluation protocol [7, 63], we conduct experiments on five major long-tailed datasets, CIFAR-10-LT, CIFAR-100-LT, Places-LT [32], ImageNet-LT [32], and iNaturalist 2018 [46].", + "bbox": [ + 497, + 520, + 890, + 574 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "CIFAR-10-LT/CIFAR-100-LT: These two datasets are sampled from the original CIFAR with different imbalance ratios $\\beta = N_{max} / N_{min}$ , where $N_{max}$ and $N_{min}$ are the corresponding number of the most and least frequent classes. Following [26], we set the imbalance ratio as $\\{200, 100, 50\\}$ for evaluation.", + "bbox": [ + 497, + 574, + 890, + 657 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Places-LT & ImageNet-LT: Both datasets were first proposed by OLTR [32]. Places-LT contains 62.5K training images spanning 365 classes in total, with imbalance ratio 996. ImageNet-LT has 115.8K training images covering 1000 categories, with imbalance ratio being 256.", + "bbox": [ + 497, + 657, + 890, + 726 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "iNaturalist 2018: As a naturally long-tailed classification dataset, iNaturalist 2018 contains 437.5K trainin images from 8142 categories, and its imbalance ratio is 512. We follow the official spilt in our evaluations.", + "bbox": [ + 497, + 726, + 890, + 782 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For fair comparison, we exclude ensemble or pretraining models in our experiments. The baselines and state-of-the-art methods evaluated include (1) class rebalancing: LDAM-DRW [7], LDAM-DRW + SAM [38], De-confound-TDE [45], Lifted Loss [36], Focal Loss [28], OpenMax [3], BBN [63], LADE [17], DisAlign [58],", + "bbox": [ + 497, + 783, + 890, + 866 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "3503", + "bbox": [ + 468, + 885, + 500, + 897 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/5612fbf3b20118985c67a1a642ae836d2e77e14a228a5ceb801c51710dc08f32.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
CIFAR-10-LTCIFAR-100-LT
Imbalance Ratio2001005020010050
CE65.6870.7074.8134.8438.4343.90
CE + Mixup [57]65.8472.9679.4835.8440.0145.16
LDAM-DRW [7]73.5277.0381.0338.9142.0447.62
De-confound-TDE [45]-80.6083.60-44.1550.31
CE + Mixup + cRT [20]73.0679.1584.2141.7345.1250.86
BBN [63]73.4779.8281.1837.2142.5647.02
Contrastive Learning [51]-81.4085.36-46.7251.87
BGP [49]---41.2045.2050.50
MiSLAS [62]77.3182.0685.1642.3347.5052.62
VS + SAM [38]-82.40--46.60-
GCL [26]79.0382.6885.4644.8848.7153.55
CC-SAM80.9483.9286.2245.6650.8353.91
", + "bbox": [ + 218, + 123, + 749, + 350 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/76060512f14fcbed006e9aa189bd555403581257e72d8a744a71b76af082d93e.jpg", + "table_caption": [ + "Table 1. Top-1 Accuracy on CIFAR-10-LT and CIFAR-100-LT. All methods take ResNet-32 as the backbone. All baseline results except BGP are directly adopted from [26]. \"--\" means the original paper didn't report the corresponding results." + ], + "table_footnote": [], + "table_body": "
cRTStage 1 + dirStage 1 + magStage 2Acc
✓37.8
✓✓38.9
✓✓37.5
✓✓✓40.1
✓✓✓✓40.6
", + "bbox": [ + 78, + 408, + 467, + 514 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 2. Ablation studies on Places-LT. \"Stage $1 + \\mathrm{dir}$ enforces parameter perturbation along the recommended direction $(\\nabla_w L_S^c (\\pmb {w}) / ||\\nabla_w L_S^c (\\pmb {w})||_2$ in Eqn 9) with magnitude of 1, whereas \"Stage $1+$ mag\" enforces perturbation with the recommended magnitude $(\\rho_c^*$ in Eqn 9) in a random direction.", + "bbox": [ + 75, + 524, + 468, + 589 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "LUNA [5], BGP [49], VS+SAM [38] (2) information augmentation: RSG [50], (3) decoupled training: Decouple- $\\tau$ norm [20], cRT [20], MisLAS [62], GCL [26], (4) representation learning: Range Loss [59], OLTR [32], IEM [65], Contrastive Learning [51], ResLT [9].", + "bbox": [ + 75, + 615, + 467, + 684 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Our code is implemented with Pytorch 1.4.0 and all experiments are carried out on Tesla V100 GPUs. We train each model with batch size of 64 (for CIFAR-10-LT and CIFAR-100-LT) / 128 (for Places-LT) / 256 (for ImageNet-LT) / 512 (for iNaturalist 2018), SGD optimizer with momentum of 0.9. We apply similar tricks in Balanced Softmax [40] as the mainstream methods have done. Although we implemented CC-SAM under a two-stage framework, it serves as a general technique to improve the existing DLTR methods without model perturbation (Appendix A.3).", + "bbox": [ + 75, + 685, + 467, + 823 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Intuitively, CC-SAM brings more computation overhead due to additional gradient descents. But we only perturbed the last several layers as an efficient version of CC-SAM in", + "bbox": [ + 75, + 824, + 467, + 866 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "our evaluation. A simple training time comparison is presented in Appendix A.5.", + "bbox": [ + 497, + 410, + 890, + 438 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Main Results", + "text_level": 1, + "bbox": [ + 499, + 448, + 636, + 461 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We evaluate CC-SAM on the five mainstream public benchmarks mentioned above, and the corresponding results are reported in Table 1 and Table 3. The best method is bolded and the second best is underlined. According to the evaluations, we observe that CC-SAM unanimously attains the best ranking on CIFAR-LT datasets. On large scale datasets (Places-LT, ImageNet-LT, and iNaturalist 2018), CC-SAM shows competitive performance compared to other advanced methods, too. Specifically, it consistently brings better gains to the medium and tail classes, demonstrating the effectiveness of class-conditional flattening for deep long-tailed recognition.", + "bbox": [ + 497, + 472, + 890, + 638 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Open Long-Tailed Recognition", + "text_level": 1, + "bbox": [ + 497, + 650, + 769, + 665 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The ability to detect out-of-distribution samples from the open world, namely open-set recognition [47], provides a unique dimension for evaluating the robustness of deep learning models. Open-set long-tailed recognition, or OLTR [32], combines the open-set problems with deep long-tailed learning to offer even more challenging tasks for model generalization. Following the setting of OLTR [32], we enable CC-SAM to distinguish the close-set and open-set samples by applying a simple, non-learnable prototype-based metric. Results are presented in Table 4. We find that CC-SAM achieves the top F-measure performance and outperforms LUNA [5] on Places-LT, a SOTA with a sophisticated open-set detection method based on hierarchical metrics. The experiments demonstrate the remarkable gen", + "bbox": [ + 497, + 672, + 890, + 866 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "3504", + "bbox": [ + 468, + 885, + 500, + 897 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/78d973e9d5ee350abd3cc6f7f9c391092985b430903664dcc446c2817902631b.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
DatasetMethodBackboneManyMediumFewOverall
ImageNet-LTCEResNeXt-5065.937.57.744.4
Decouple-τ-norm [20]ResNet-5056.644.227.446.7
Balanced Softmax [40]ResNeXt-5064.148.233.452.3
LADE [17]ResNeXt-5064.447.734.352.3
RSG [50]ResNeXt-5063.248.232.251.8
DisAlign [58]ResNet-5061.352.231.452.9
ResNeXt-5062.752.131.453.4
ResLT [9]ResNeXt-5063.053.335.552.9
BGP [49]ResNet-50---51.5
MiSLAS [62]ResNet-50---52.7
LDAM-DRW + SAM [38]ResNet-5062.052.134.853.1
GCL [26]ResNet-50---54.9
CC-SAMResNet-5061.449.537.152.4
ResNeXt-5063.153.441.155.4
Places-LTCEResNet-15245.727.38.230.2
Decouple-τ-norm [20]ResNet-15237.840.731.837.9
Balanced Softmax [40]ResNet-15242.039.330.538.6
LADE [17]ResNet-15242.839.031.238.8
RSG [50]ResNet-15241.941.432.039.3
DisAlign [58]ResNet-15240.442.430.139.3
ResLT [9]ResNet-15239.843.631.439.8
MiSLAS [62]ResNet-152---40.2
GCL [26]ResNet-152---40.6
CC-SAMResNet-15241.242.136.440.6
iNaturalist 2018CEResNet-5072.263.057.261.7
Decouple-τ-norm [20]ResNet-5065.665.365.965.6
Balanced Softmax [40]ResNet-50---70.6
LADE [17]ResNet-50---70.0
RSG [50]ResNet-50---70.3
DisAlign [58]ResNet-50---70.6
ResLT [9]ResNet-50---70.2
BGP [49]ResNet-5070.069.969.670.5
MiSLAS [62]ResNet-50---71.6
LDAM-DRW + SAM [38]ResNet-5064.170.571.270.1
GCL [26]ResNet-50---72.0
CC-SAMResNet-5065.470.972.270.9
", + "bbox": [ + 144, + 123, + 823, + 661 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 3. Top-1 Accuracy on Places-LT, ImageNet-LT and iNaturalist 2018. As for Places-LT, we take a pre-trained ResNet-152 as the backbone for a fair comparison.", + "bbox": [ + 75, + 671, + 890, + 697 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "eralization of CC-SAM due to its capacity to learn highly robust representations.", + "bbox": [ + 75, + 720, + 468, + 749 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Ablation Study", + "text_level": 1, + "bbox": [ + 75, + 760, + 230, + 776 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Since CC-SAM consists of multiple components, here we provide an ablation study to demonstrate their effectiveness. The result of experiments conducted on Places-LT are shown in Table 2. In particular, CC-SAM degrades to cRT without any of our proposed operations, which we take as a baseline. It is observed that CC-SAM benefits from the", + "bbox": [ + 75, + 783, + 468, + 866 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "design choices of each individual stage, and the integration of both stages shows the best performance.", + "bbox": [ + 497, + 720, + 890, + 750 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We also study the class-conditional perturbation by differentiating the impacts of choosing the right direction $\\nabla_{\\boldsymbol{w}}L_{S}^{c}(\\boldsymbol{w}) / ||\\nabla_{\\boldsymbol{w}}L_{S}^{c}(\\boldsymbol{w})||_{2}$ and the right magnitude $\\rho_c^*$ in Eqn 9, shown as \"stage $1 + \\mathrm{dir}$ \" and \"stage $1 + \\mathrm{mag}$ \" in Table 2 respectively. For the former, we choose a perturbation scale of 1. For the latter, we perturb the model parameters by a random direction sampled from a zero-mean Gaussian. It turns out the perturbation direction contributes more, and", + "bbox": [ + 497, + 755, + 892, + 866 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "3505", + "bbox": [ + 468, + 885, + 500, + 897 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/27bf662d3d688c8baa8e5dbca283a5679d147a398a958d40f142734ad013089d.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ImageNet-LTPlaces-LT
MethodManyMediumFewF-measureManyMediumFewF-measure
CE40.110.40.40.29545.922.40.40.366
Lifted Loss [36]34.829.317.40.37441.035.223.80.459
Focal Loss [28]35.729.315.60.37141.034.822.30.453
Range Loss [59]34.729.417.20.37341.035.323.10.457
OpenMax [3]35.830.017.60.36841.135.423.20.458
OLTR [32]41.933.917.40.47444.636.825.20.464
IEM [65]46.142.320.10.52548.842.428.90.486
LUNA [5]48.244.723.60.57948.141.629.00.491
CC-SAM61.449.537.10.55241.241.836.40.510
", + "bbox": [ + 147, + 123, + 823, + 302 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 4. Open long-tail performance of top-1 accuracy on ImageNet-LT and Places-LT. F-measure is a balanced treatment of precision and recall. The backbone of CC-SAM is ResNet-50 for ImageNet-LT, while other compared methods are equipped with ResNet-10.", + "bbox": [ + 75, + 312, + 892, + 339 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "applying both attains the best improvement. For more ablation studies, please refer to Appendix A.1.", + "bbox": [ + 75, + 362, + 470, + 391 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Related Work and Discussion", + "text_level": 1, + "bbox": [ + 75, + 403, + 347, + 418 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.1. Deep Long-Tailed Recognition", + "text_level": 1, + "bbox": [ + 75, + 426, + 347, + 441 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "A latest survey [60] systematically studies up-to-date DLTR algorithms published at the top-tier conferences and introduces a new taxonomy, dividing DLTR methodologies into three categories: class re-balancing, information augmentation, and module improvement. For class rebalancing, a branch of sampling strategies [20, 40, 52, 54] have been proposed to re-balance the training distribution. Meanwhile, some studies design dedicated loss functions [7, 10, 28, 45] to dynamically re-weight the gradients for optimization. As one representative of information augmentation, RSG [50] enriches the feature space of tail classes via dynamical prototypes and a maximized vector loss. Decoupled training [20, 58, 62] propose to learn effective representation and robust classifier according to different sampling strategies in two stages.", + "bbox": [ + 75, + 448, + 468, + 654 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Our method follows the decoupled training paradigm. Moreover, to combat label distribution shift, our method seeks out flat minima by perturbing the model parameters at a class-conditional scale, which is unseen in existing DLTR literature.", + "bbox": [ + 75, + 655, + 468, + 723 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.2. Flattening for Deep Learning Models", + "text_level": 1, + "bbox": [ + 75, + 733, + 398, + 749 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Sharpness as a generalization measure of deep learning models has been extensively explored in previous literature [1, 13, 16, 19]. Keskar et al. [21] empirically found that a large batch SGD optimizer might lead to sharp minima and poor generalization. Inspired by this, He et al. [16] implicitly averaged along the SGD trajectory to find the asymmetric valleys, namely asymmetric flat minima. Due to the power of flattening, it has also been applied to many", + "bbox": [ + 75, + 755, + 468, + 866 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "learning tasks to achieve better generalization. Rangwani et al. [37] introduced a domain adversarial training framework by adding an auxiliary loss to smooth the local minima for domain adaptation. Shi et al. [44] demonstrated that a flattened model was resilient to catastrophic forgetting in class-incremental few-shot learning.", + "bbox": [ + 497, + 362, + 892, + 447 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this work, we investigate flattening in the context of DLTR. To our best knowledge, there are only a few works which succeed in applying flattening to improve generalization in DLTR/OLTR. GBP [49] is built upon the observation of high correlation between logits and gradient norm, and proposes a gradient penalty loss as regularization. However, it lacks rigorous theoretical interpretation by employing a very loose (vacuous) PAC-Bayesian bound and meanwhile under-performs compared to our approach (Table 1, 3). Liu et al. [30] develops a variant of SAM that re-weights the perturbation at instance level to combat data imbalance, without specifying an optimal perturbation or giving any theoretical justification; thus, we regard it as a simplistic and incomplete version of our method. Besides, Rangwani et al. [38] observes that re-weighting techniques in DLTR lead to saddle points for tail classes, and directly leverages SAM to alleviate this problem. In contrast, CC-SAM further advances [38] by providing a theoretically motivated class-conditional SAM design for better generalization and robustness to label distribution shift.", + "bbox": [ + 497, + 461, + 892, + 738 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgment We thank anonymous reviewers for their valuable comments. This work was supported by the NSFC under Grants 61932017 and 61971390, and the following grants from the Research Grants Council of the Hong Kong Special Administrative Region, China (Project Reference Number: 14201620 and 14201321). Lanqing Li would like to thank Guangyong Chen and Danruo Deng for their helpful comments on the PAC-Bayesian framework.", + "bbox": [ + 497, + 755, + 892, + 866 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "3506", + "bbox": [ + 468, + 885, + 502, + 897 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 124, + 171, + 138 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Carlo Baldassi, Fabrizio Pittorino, and Riccardo Zecchina. Shaping the learning landscape in neural networks around wide flat minima. Proceedings of the National Academy of Sciences, 117(1):161-170, 2020. 2, 8", + "[2] Albert-László Barabási and Réka Albert. Emergence of scaling in random networks. science, 286(5439):509-512, 1999. 1", + "[3] Abhijit Bendale and Terrance E Boult. Towards open set deep networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1563-1572, 2016. 5, 8", + "[4] Andrew Brock, Jeff Donahue, and Karen Simonyan. Large scale gan training for high fidelity natural image synthesis. arXiv preprint arXiv:1809.11096, 2018. 3", + "[5] Jiarui Cai, Yizhou Wang, Hung-Min Hsu, Jenq-Neng Hwang, Kelsey Magrane, and Craig S Rose. Luna: Localizing unfamiliarity near acquaintance for open-set long-tailed recognition. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 36, pages 131-139, 2022. 6, 8", + "[6] Jiarui Cai, Yizhou Wang, and Jenq-Neng Hwang. Ace: Ally complementary experts for solving long-tailed recognition in one-shot. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 112-121, 2021. 1", + "[7] Kaidi Cao, Colin Wei, Adrien Gaidon, Nikos Arechiga, and Tengyu Ma. Learning imbalanced datasets with label-distribution-aware margin loss. Advances in neural information processing systems, 32, 2019. 1, 2, 5, 6, 8", + "[8] Niladri S. Chatterji, Behnam Neyshabur, and Hanie Sedghi. The intriguing role of module criticality in the generalization of deep networks. In 8th International Conference on Learning Representations, 2020. 3", + "[9] Jiequan Cui, Shu Liu, Zhuotao Tian, Zhisheng Zhong, and Jiaya Jia. Result: Residual learning for long-tailed recognition. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2022. 1, 6, 7", + "[10] Yin Cui, Menglin Jia, Tsung-Yi Lin, Yang Song, and Serge Belongie. Class-balanced loss based on effective number of samples. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9268-9277, 2019. 1, 8", + "[11] Danruo Deng, Guangyong Chen, Jianye Hao, Qiong Wang, and Pheng-Ann Heng. Flattening sharpness for dynamic gradient projection memory benefits continual learning. Advances in Neural Information Processing Systems, 34:18710-18721, 2021. 1", + "[12] Gintare Karolina Dziugaite and Daniel M. Roy. Computing nonvacuous generalization bounds for deep (stochastic) neural networks with many more parameters than training data. In Proceedings of the Thirty-Third Conference on Uncertainty in Artificial Intelligence, 2017. 3", + "[13] Pierre Foret, Ariel Kleiner, Hossein Mobahi, and Behnam Neyshabur. Sharpness-aware minimization for efficiently improving generalization. In 9th International Conference on Learning Representations, 2021. 1, 2, 3, 4, 8" + ], + "bbox": [ + 78, + 148, + 468, + 865 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[14] Ishaan Gulrajani and David Lopez-Paz. In search of lost domain generalization. In 9th International Conference on Learning Representations, 2021. 1", + "[15] Haibo He and Edwardo A Garcia. Learning from imbalanced data. IEEE Transactions on knowledge and data engineering, 21(9):1263-1284, 2009. 1", + "[16] Haowei He, Gao Huang, and Yang Yuan. Asymmetric valleys: Beyond sharp and flat local minima. Advances in neural information processing systems, 32, 2019. 2, 3, 8", + "[17] Youngkyu Hong, Seungju Han, Kwanghee Choi, Seokjun Seo, Beomsu Kim, and Buru Chang. Disentangling label distribution for long-tailed visual recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6626-6636, 2021. 1, 2, 5, 7", + "[18] Yuanfeng Ji, Lu Zhang, Jiaxiang Wu, Bingzhe Wu, Long-Kai Huang, Tingyang Xu, Yu Rong, Lanqing Li, Jie Ren, Ding Xue, et al. Drugood: Out-of-distribution (ood) dataset curator and benchmark for ai-aided drug discovery-a focus on affinity prediction problems with noise annotations. arXiv preprint arXiv:2201.09637, 2022. 1", + "[19] Yiding Jiang, Behnam Neyshabur, Hossein Mobahi, Dilip Krishnan, and Samy Bengio. Fantastic generalization measures and where to find them. In 8th International Conference on Learning Representations, 2020. 2, 3, 8", + "[20] Bingyi Kang, Saining Xie, Marcus Rohrbach, Zhicheng Yan, Albert Gordo, Jiashi Feng, and Yannis Kalantidis. Decoupling representation and classifier for long-tailed recognition. arXiv preprint arXiv:1910.09217, 2019. 1, 2, 4, 6, 7, 8", + "[21] Nitish Shirish Keskar, Dheevatsa Mudigere, Jorge Nocedal, Mikhail Smelyanskiy, and Ping Tak Peter Tang. On large-batch training for deep learning: Generalization gap and sharp minima. In 5th International Conference on Learning Representations, 2017. 1, 2, 8", + "[22] Jaehyung Kim, Jongheon Jeong, and Jinwoo Shin. M2m: Imbalanced classification via major-to-minor translation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 13896-13905, 2020. 1", + "[23] Pang Wei Koh, Shiori Sagawa, Henrik Marklund, Sang Michael Xie, Marvin Zhang, Akshay Balsubramani, Weihua Hu, Michihiro Yasunaga, Richard Lanas Phillips, Irena Gao, et al. Wilds: A benchmark of in-the-wild distribution shifts. In International Conference on Machine Learning, pages 5637-5664. PMLR, 2021. 1", + "[24] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E. Hinton. Imagenet classification with deep convolutional neural networks. In Advances in Neural Information Processing Systems 25: 26th Annual Conference on Neural Information Processing Systems, pages 1106-1114, 2012. 1", + "[25] Lanqing Li, Liang Zeng, Ziqi Gao, Shen Yuan, Yatao Bian, Bingzhe Wu, Hengtong Zhang, Chan Lu, Yang Yu, Wei Liu, et al. Imdrug: A benchmark for deep imbalanced learning in ai-aided drug discovery. arXiv preprint arXiv:2209.07921, 2022.1", + "[26] Mengke Li, Yiu-ming Cheung, and Yang Lu. Long-tailed visual recognition via gaussian clouded logit adjustment. In Proceedings of the IEEE/CVF Conference on Computer Vi" + ], + "bbox": [ + 502, + 127, + 890, + 865 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "3507", + "bbox": [ + 468, + 885, + 500, + 897 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "sion and Pattern Recognition, pages 6929-6938, 2022. 2, 5, 6, 7", + "[27] Shuang Li, Kaixiong Gong, Chi Harold Liu, Yulin Wang, Feng Qiao, and Xinjing Cheng. Metasaug: Meta semantic augmentation for long-tailed visual recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5212-5221, 2021. 1", + "[28] Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dollar. Focal loss for dense object detection. In Proceedings of the IEEE international conference on computer vision, pages 2980-2988, 2017. 1, 5, 8", + "[29] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In European conference on computer vision, pages 740-755. Springer, 2014. 1", + "[30] Hong Liu, Jeff Z HaoChen, Adrien Gaidon, and Tengyu Ma. Self-supervised learning is more robust to dataset imbalance. In International Conference on Learning Representations, 2021. 1, 8", + "[31] Jialun Liu, Yifan Sun, Chuchu Han, Zhaopeng Dou, and Wenhui Li. Deep representation learning on long-tailed data: A learnable embedding augmentation perspective. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2970-2979, 2020. 1", + "[32] Ziwei Liu, Zhongqi Miao, Xiaohang Zhan, Jiayun Wang, Boqing Gong, and Stella X Yu. Large-scale long-tailed recognition in an open world. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2537-2546, 2019. 1, 2, 5, 6, 8", + "[33] Jonathan Long, Evan Shelhamer, and Trevor Darrell. Fully convolutional networks for semantic segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3431-3440, 2015. 1", + "[34] David A McAllester. Pac-bayesian model averaging. In Proceedings of the twelfth annual conference on Computational learning theory, pages 164-170, 1999. 3", + "[35] Behnam Neyshabur, Srinadh Bhojanapalli, David McAllester, and Nati Srebro. Exploring generalization in deep learning. Advances in neural information processing systems, 30, 2017. 1, 3", + "[36] Hyun Oh Song, Yu Xiang, Stefanie Jegelka, and Silvio Savarese. Deep metric learning via lifted structured feature embedding. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4004-4012, 2016. 5, 8", + "[37] Harsh Rangwani, Sumukh K Aithal, Mayank Mishra, Arihant Jain, and Venkatesh Babu Radhakrishnan. A closer look at smoothness in domain adversarial training. In International Conference on Machine Learning, pages 18378-18399. PMLR, 2022. 8", + "[38] Harsh Rangwani, Sumukh K Aithal, Mayank Mishra, and Venkatesh Babu Radhakrishnan. Escaping saddle points for effective generalization on class-imbalanced data. In Advances in Neural Information Processing Systems. 5, 6, 7, 8", + "[39] William J Reed. The pareto, zipf and other power laws. Economics letters, 74(1):15-19, 2001. 1" + ], + "bbox": [ + 80, + 125, + 468, + 865 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[40] Jiawei Ren, Cunjun Yu, Xiao Ma, Haiyu Zhao, Shuai Yi, et al. Balanced meta-softmax for long-tailed visual recognition. Advances in neural information processing systems, 33:4175-4186, 2020. 1, 6, 7, 8", + "[41] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. Advances in neural information processing systems, 28, 2015. 1", + "[42] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, et al. Imagenet large scale visual recognition challenge. International journal of computer vision, 115(3):211-252, 2015. 1", + "[43] Ikuro Sato, Yamada Ryota, Masayuki Tanaka, Nakamasa Inoue, and Rei Kawakami. Pof: Post-training of feature extractor for improving generalization. In International Conference on Machine Learning, pages 19221-19230. PMLR, 2022. 1", + "[44] Guangyuan Shi, Jiaxin Chen, Wenlong Zhang, Li-Ming Zhan, and Xiao-Ming Wu. Overcoming catastrophic forgetting in incremental few-shot learning by finding flat minima. Advances in Neural Information Processing Systems, 34:6747-6761, 2021. 1, 3, 8", + "[45] Jingru Tan, Changbao Wang, Buyu Li, Quanquan Li, Wanli Ouyang, Changqing Yin, and Junjie Yan. Equalization loss for long-tailed object recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11662-11671, 2020. 1, 5, 6, 8", + "[46] Grant Van Horn, Oisin Mac Aodha, Yang Song, Yin Cui, Chen Sun, Alex Shepard, Hartwig Adam, Pietro Perona, and Serge Belongie. The inaturalist species classification and detection dataset. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 8769-8778, 2018. 5", + "[47] Sagar Vaze, Kai Han, Andrea Vedaldi, and Andrew Zisserman. Open-set recognition: A good closed-set classifier is all you need. arXiv preprint arXiv:2110.06207, 2021. 5, 6", + "[48] Athanasios Voulodimos, Nikolaos Doulamis, Anastasios Doulamis, and Eftychios Protopapadakis. Deep learning for computer vision: A brief review. Computational intelligence and neuroscience, 2018. 1", + "[49] Dong Wang, Yicheng Liu, Liangji Fang, Fanhua Shang, Yuanyuan Liu, and Hongying Liu. Balanced gradient penalty improves deep long-tailed learning. In Proceedings of the 30th ACM International Conference on Multimedia, pages 5093-5101, 2022. 2, 6, 7, 8", + "[50] Jianfeng Wang, Thomas Lukasiewicz, Xiaolin Hu, Jianfei Cai, and Zhenghua Xu. Rsg: A simple but effective module for learning imbalanced datasets. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3784-3793, 2021. 1, 6, 7, 8", + "[51] Peng Wang, Kai Han, Xiu-Shen Wei, Lei Zhang, and Lei Wang. Contrastive learning based hybrid networks for long-tailed image classification. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 943-952, 2021. 1, 6", + "[52] Tao Wang, Yu Li, Bingyi Kang, Junnan Li, Junhao Liew, Sheng Tang, Steven Hoi, and Jiashi Feng. The devil is" + ], + "bbox": [ + 502, + 127, + 892, + 865 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "3508", + "bbox": [ + 468, + 885, + 500, + 897 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "in classification: A simple framework for long-tail instance segmentation. In European conference on computer vision, pages 728-744. Springer, 2020. 1, 8", + "[53] Xudong Wang, Long Lian, Zhongqi Miao, Ziwei Liu, and Stella Yu. Long-tailed recognition by routing diverse distribution-aware experts. In International Conference on Learning Representations, 2020. 1", + "[54] Yiru Wang, Weihao Gan, Jie Yang, Wei Wu, and Junjie Yan. Dynamic curriculum learning for imbalanced data classification. In Proceedings of the IEEE/CVF international conference on computer vision, pages 5017-5026, 2019. 1, 8", + "[55] Xi Yin, Xiang Yu, Kihyuk Sohn, Xiaoming Liu, and Manmohan Chandraker. Feature transfer learning for face recognition with under-represented data. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5704-5713, 2019. 1", + "[56] Liang Zeng, Lanqing Li, Ziqi Gao, Peilin Zhao, and Jian Li. Imgcl: Revisiting graph contrastive learning on imbalanced node classification. arXiv preprint arXiv:2205.11332, 2022. 1", + "[57] Hongyi Zhang, Moustapha Cisse, Yann N Dauphin, and David Lopez-Paz. mixup: Beyond empirical risk minimization. arXiv preprint arXiv:1710.09412, 2017. 6", + "[58] Songyang Zhang, Zeming Li, Shipeng Yan, Xuming He, and Jian Sun. Distribution alignment: A unified framework for long-tail visual recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2361-2370, 2021. 1, 5, 7, 8", + "[59] Xiao Zhang, Zhiyuan Fang, Yandong Wen, Zhifeng Li, and Yu Qiao. Range loss for deep face recognition with long-tailed training data. In Proceedings of the IEEE International Conference on Computer Vision, pages 5409-5418, 2017. 1, 6, 8", + "[60] Yifan Zhang, Bingyi Kang, Bryan Hooi, Shuicheng Yan, and Jiashi Feng. Deep long-tailed learning: A survey. arXiv preprint arXiv:2110.04596, 2021. 1, 8", + "[61] Yang Zhao, Hao Zhang, and Xiuyuan Hu. Penalizing gradient norm for efficiently improving generalization in deep learning. arXiv preprint arXiv:2202.03599, 2022. 3", + "[62] Zhisheng Zhong, Jiequan Cui, Shu Liu, and Jiaya Jia. Improving calibration for long-tailed recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16489-16498, 2021. 1, 2, 6, 7, 8", + "[63] Boyan Zhou, Quan Cui, Xiu-Shen Wei, and Zhao-Min Chen. Bbn: Bilateral-branch network with cumulative learning for long-tailed visual recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9719–9728, 2020. 1, 5, 6", + "[64] Bolei Zhou, Agata Lapedriza, Aditya Khosla, Aude Oliva, and Antonio Torralba. Places: A 10 million image database for scene recognition. IEEE transactions on pattern analysis and machine intelligence, 40(6):1452-1464, 2017. 1", + "[65] Linchao Zhu and Yi Yang. Inflated episodic memory with region self-attention for long-tailed visual recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4344–4353, 2020. 1, 6, 8" + ], + "bbox": [ + 78, + 125, + 468, + 858 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "3509", + "bbox": [ + 468, + 885, + 500, + 897 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2023/Class-Conditional Sharpness-Aware Minimization for Deep Long-Tailed Recognition/5537ac60-2d54-48e8-9091-a4f2d7a57e6f_model.json b/2023/Class-Conditional Sharpness-Aware Minimization for Deep Long-Tailed Recognition/5537ac60-2d54-48e8-9091-a4f2d7a57e6f_model.json new file mode 100644 index 0000000000000000000000000000000000000000..b7f559d978f93d238d80f77cb9cf2f0bea202969 --- /dev/null +++ b/2023/Class-Conditional Sharpness-Aware Minimization for Deep Long-Tailed Recognition/5537ac60-2d54-48e8-9091-a4f2d7a57e6f_model.json @@ -0,0 +1,2587 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.129, + 0.051, + 0.201, + 0.085 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.253, + 0.049, + 0.793, + 0.088 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.123, + 0.162, + 0.849, + 0.205 + ], + "angle": 0, + "content": "Class-Conditional Sharpness-Aware Minimization for Deep Long-Tailed Recognition" + }, + { + "type": "text", + "bbox": [ + 0.124, + 0.239, + 0.844, + 0.259 + ], + "angle": 0, + "content": "Zhipeng Zhou\\(^{1,\\dagger}\\), Lanqing Li\\(^{2,4,\\dagger}\\), Peilin Zhao\\(^{3,\\star}\\), Pheng-Ann Heng\\(^{2}\\), Wei Gong\\(^{1,\\star}\\)" + }, + { + "type": "text", + "bbox": [ + 0.122, + 0.267, + 0.856, + 0.301 + ], + "angle": 0, + "content": "1University of Science and Technology of China, 2The Chinese University of Hong Kong, 3Tencent AI Lab, 4Zhejiang Lab" + }, + { + "type": "text", + "bbox": [ + 0.16, + 0.303, + 0.808, + 0.317 + ], + "angle": 0, + "content": "zzpl1994@mail.ustc.edu.cn, lanqingli1993@gmail.com, masonzhao@tencent.com," + }, + { + "type": "text", + "bbox": [ + 0.297, + 0.32, + 0.673, + 0.334 + ], + "angle": 0, + "content": "pheng@cse.cuhk.edu.hk, weigong@ustc.edu.cn" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.373, + 0.314, + 0.388 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.402, + 0.473, + 0.749 + ], + "angle": 0, + "content": "It's widely acknowledged that deep learning models with flatter minima in its loss landscape tend to generalize better. However, such property is under-explored in deep long-tailed recognition (DLTR), a practical problem where the model is required to generalize equally well across all classes when trained on highly imbalanced label distribution. In this paper, through empirical observations, we argue that sharp minima are in fact prevalent in deep long-tailed models, whereas naive integration of existing flattening operations into long-tailed learning algorithms brings little improvement. Instead, we propose an effective two-stage sharpness-aware optimization approach based on the decoupling paradigm in DLTR. In the first stage, both the feature extractor and classifier are trained under parameter perturbations at a class-conditioned scale, which is theoretically motivated by the characteristic radius of flat minima under the PAC-Bayesian framework. In the second stage, we generate adversarial features with class-balanced sampling to further robustify the classifier with the backbone frozen. Extensive experiments on multiple long-tailed visual recognition benchmarks show that, our proposed Class-Conditional Sharpness-Aware Minimization (CC-SAM), achieves competitive performance compared to the state-of-the-arts. Code is available at https://github.com/zzpustc/CC-SAM." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.772, + 0.21, + 0.787 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.795, + 0.47, + 0.838 + ], + "angle": 0, + "content": "Modern deep learning models, composed of multiple neural network layers with millions of parameters, have achieved remarkable successes in computer vision [24, 33," + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.844, + 0.47, + 0.867 + ], + "angle": 0, + "content": "† Equal contribution. * Corresponding authors. Work was primarily done when Z. Zhou worked as L. Li's intern at Tencent AI Lab, Shenzhen." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.374, + 0.892, + 0.542 + ], + "angle": 0, + "content": "41, 48]. A key enabler of deep learning is the collection of large-scale datasets [29, 42, 64], which are normally split into training and testing sets with presumably i.i.d. samples. However, such scenario provides relatively trivial tests for the generalization of machine learning models. In practice, label [17, 25, 56] and domain [14, 18, 23] distribution shifts are prevalent, due to the disparity between the data preparation and evaluation protocols. A classical example is imbalanced [15] or long-tailed recognition [60], where a model is trained on highly imbalanced source label distribution \\( p_{s}(\\pmb{y}) \\) while evaluated on a uniform target label distribution \\( p_{t}(\\pmb{y}) \\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.544, + 0.893, + 0.711 + ], + "angle": 0, + "content": "In this paper, we focus on the practical yet challenging deep long-tailed recognition (DLTR) problem, which is inherent in the visual world [32, 60] with fundamental connections to many disciplines such as the power-law scaling in network science [2] and the Pareto principle in economics [39]. In computer vision, numerous deep long-tailed learning studies have emerged in recent years, which mainly belong to 5 categories: class rebalancing [7, 10, 28, 40, 45, 52, 54], information augmentation [22, 27, 31, 50, 55], decoupled training [20, 58, 62], representation learning [9, 32, 51, 59, 65] and ensemble learning [6, 53, 63]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.714, + 0.895, + 0.868 + ], + "angle": 0, + "content": "In this work, we propose a novel approach to DLTR from a distinct angle, by seeking out flat minima in the loss landscape of modern neural networks to ensure model robustness under parameter perturbation. Such optimization strategy, termed flattening in our context, have been shown in a myriad of literature to effectively improve generalization of deep learning models in terms of supervised learning [13, 21, 35, 43], self-supervised learning [30] and continual learning [11, 44]. However, application and adaptation of flattening in the context of DLTR remain under-explored. To fill this gap, we first show later in this paper" + }, + { + "type": "page_number", + "bbox": [ + 0.469, + 0.887, + 0.503, + 0.898 + ], + "angle": 0, + "content": "3499" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.126, + 0.473, + 0.25 + ], + "angle": 0, + "content": "(Section 2.2.2) that existing flattening methods are ineffective for long-tailed learning, consistent with the observation from a very recent paper [49], due to severe label distribution shifts. Accordingly, we present a new efficient variant of the sharpness-aware minimization (SAM) [13] technique based on the Decoupling paradigm [20] of DLTR, which leverages the invariance of the class conditional distribution between the source and the target domain. In a nutshell, our contributions are three-fold:" + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.263, + 0.47, + 0.332 + ], + "angle": 0, + "content": "- Through the lens of flattening, we corroborate a very recent observation [49] that existing sharpness-aware minimization techniques are suboptimal for deep long-tailed learning, justifying the need of more effective approaches to this important and practical issue." + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.342, + 0.47, + 0.455 + ], + "angle": 0, + "content": "- We introduce Class-Conditional Sharpness-Aware Minimization (CC-SAM), a novel algorithm tailored for DLTR that improves model generalization by robust training against class-conditioned parameter perturbation. This technique is motivated by the characteristic radius of flat minima we derive under the PAC-Bayesian framework and can be implemented efficiently." + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.463, + 0.47, + 0.547 + ], + "angle": 0, + "content": "- We integrate CC-SAM with the two-stage decoupling paradigm of DLTR. Extensive experiments demonstrate that our method can achieve competitive performance on multiple public DLTR benchmarks, with remarkable robustness to out-of-distribution samples for open long-tailed recognition [32]." + }, + { + "type": "list", + "bbox": [ + 0.096, + 0.263, + 0.47, + 0.547 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.56, + 0.216, + 0.573 + ], + "angle": 0, + "content": "2. Preliminaries" + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.582, + 0.228, + 0.597 + ], + "angle": 0, + "content": "2.1. Problem Setup" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.604, + 0.47, + 0.839 + ], + "angle": 0, + "content": "Throughout the paper, we denote scalars as \\(s\\), vectors as \\(\\mathbf{s}\\), sets as \\(\\mathcal{S}\\) and equality by definition as :=. For a typical \\(d\\)-way classification task in DLTR, suppose we are given a training dataset \\(\\mathcal{S} = \\bigcup_{i=1}^{n} \\{(x_i, y_i)\\}\\), where \\(n\\) is the total number of samples. Let's denote \\(\\{n_1, n_2, \\ldots, n_d\\}\\) as the sample number of each class, where \\(n = \\sum_{i} n_i\\). Without loss of generality, we assume \\(n_i < n_j\\) if \\(i < j\\), and usually \\(n_d \\gg n_1\\), following a highly imbalanced class distribution. We further write the training (source) distribution as \\(p_s(\\boldsymbol{x}, \\boldsymbol{y}) = p_s(\\boldsymbol{x}|\\boldsymbol{y}) p_s(\\boldsymbol{y})\\) and the testing (target) distribution as \\(p_t(\\boldsymbol{x}, \\boldsymbol{y}) = p_t(\\boldsymbol{x}|\\boldsymbol{y}) p_t(\\boldsymbol{y})\\). Consider a family of models parameterized by \\(\\boldsymbol{w} \\in \\mathcal{W} \\subseteq \\mathbb{R}^k\\); given a loss function \\(l\\colon \\mathcal{W} \\times \\mathcal{X} \\times \\mathcal{Y} \\to \\mathbb{R}_+\\), we define the empirical training loss \\(L_{\\mathcal{S}}(\\boldsymbol{w}) := \\frac{1}{n} \\sum_{i=1}^{n} l(\\boldsymbol{w}, \\boldsymbol{x}_i, \\boldsymbol{y}_i)\\) and the population testing loss \\(L_{\\mathcal{T}}(\\boldsymbol{w}) := \\mathbb{E}_{(\\boldsymbol{x}, \\boldsymbol{y}) \\sim p_t(\\boldsymbol{x}, \\boldsymbol{y})} [l(\\boldsymbol{w}, \\boldsymbol{x}, \\boldsymbol{y})]\\). Having observed only \\(\\mathcal{S}\\), the goal is to optimize for model parameters \\(\\boldsymbol{w}\\) having lowest risk \\(L_{\\mathcal{T}}(\\boldsymbol{w})\\) at test time." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.839, + 0.469, + 0.868 + ], + "angle": 0, + "content": "In DLTR, a key challenge is the label distribution shift between the training and testing sets, i.e., \\( p_{s}(\\pmb{y}) \\neq \\)" + }, + { + "type": "image", + "bbox": [ + 0.501, + 0.125, + 0.685, + 0.232 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.515, + 0.234, + 0.669, + 0.247 + ], + "angle": 0, + "content": "(a) 1D global loss landscape." + }, + { + "type": "image", + "bbox": [ + 0.702, + 0.125, + 0.89, + 0.232 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.724, + 0.234, + 0.869, + 0.247 + ], + "angle": 0, + "content": "(b) 1D local loss landscape." + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.257, + 0.893, + 0.308 + ], + "angle": 0, + "content": "Figure 1. Loss value vs. noise norm ratio \\(||\\epsilon ||_2 / ||\\pmb {\\theta}||_2\\). All experiments are conducted on the CIFAR-10-LT with an imbalance ratio of 100, implemented with the same backbone ResNet-32 and averaged over 5 random seeds." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.333, + 0.892, + 0.458 + ], + "angle": 0, + "content": "\\(p_t(\\boldsymbol{y})\\) [17]. However, it's logical to assume no distribution shift within each individual class. That is, the training and testing samples are drawn i.i.d. from the same class-conditional distribution, which indicates \\(p_s(\\boldsymbol{x}|\\boldsymbol{y}) = p_t(\\boldsymbol{x}|\\boldsymbol{y})\\). Here, we build our flattening algorithm upon the decoupling paradigm of DLTR [20], where the feature extractor \\(f(\\boldsymbol{x};\\boldsymbol{\\varphi})\\) and classifier \\(h(\\boldsymbol{z};\\boldsymbol{\\theta})\\) are trained by distinct sampling strategies in a two-stage manner, with \\(\\boldsymbol{w} \\coloneqq (\\varphi, \\boldsymbol{\\theta})\\)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.466, + 0.622, + 0.48 + ], + "angle": 0, + "content": "2.2. Motivation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.488, + 0.893, + 0.626 + ], + "angle": 0, + "content": "Previous studies have shown strong connection between the geometry of the loss landscape and generalization of deep learning models [1, 16, 19, 21]. In this section, we first empirically demonstrate that many deep long-tailed models exhibit sharp minima, and then show that naive application of classical flattening operations results in limited improvement. Motivated by these observations, we theoretically analyze the characteristic radius of flat minima in DLTR to show that a class-conditional flattening procedure is required." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.644, + 0.826, + 0.658 + ], + "angle": 0, + "content": "2.2.1 Sharpness in Deep Long-Tailed Models" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.665, + 0.892, + 0.762 + ], + "angle": 0, + "content": "To show the pervasive sharpness in current DLTR models, we select two baselines (empirical risk minimization of cross-entropy (CE), LDAM-DRW [7]) and two advanced methods (MiSLAS [62], GCL [26]) from recent literature to conduct experiments1. By perturbing the parameter \\(\\theta\\) of the classifier \\(h\\) with increasing noise \\(\\epsilon\\), we observe the loss of each model on CIFAR-10-LT in Figure 1." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.762, + 0.892, + 0.846 + ], + "angle": 0, + "content": "Figure 1(a) shows that, at global scale, the loss of CE is the sharpest across a wide noise range. However, due to the fact that DLTR methods usually employ distinct loss functions, it's hard to compare the sharpness of their loss landscapes in a fair and meaningful way. Moreover, large noise ratio can easily flip the sign of parameters and deteriorate" + }, + { + "type": "page_footnote", + "bbox": [ + 0.516, + 0.854, + 0.846, + 0.866 + ], + "angle": 0, + "content": "1We re-implement all models via their publicly released code." + }, + { + "type": "page_number", + "bbox": [ + 0.469, + 0.887, + 0.502, + 0.898 + ], + "angle": 0, + "content": "3500" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.091, + 0.123, + 0.454, + 0.252 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.263, + 0.471, + 0.351 + ], + "angle": 0, + "content": "Figure 2. Performance gain of naive integration of flattening operations with DLTR algorithms. The number represents the stage where flattening applies. For example, \"SN.1.2\" means we apply spectral normalization on both stages. Only the first-stage result of LDAM is reported since it's a one-stage approach. All experiments are conducted on CIFAR-100-LT with imbalance ratio of 100." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.375, + 0.469, + 0.513 + ], + "angle": 0, + "content": "the classifier decision boundary. Therefore in this paper, we only consider locally flat minima, i.e., in the perturbative \\((||\\epsilon ||_2\\ll ||\\boldsymbol {w}||_2)\\) regime shown in Figure 1(b). From this local view, only MiSLAS is observed to have a comparable or flatter minimum than CE, suggesting that the existing DLTR methods have poor generalization and robustness against model parameter perturbation. More detailed 2D loss landscapes covering overall, head and tail classes are depicted in Appendix E, and corroborate our observation in 1D." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.531, + 0.43, + 0.545 + ], + "angle": 0, + "content": "2.2.2 Naive Integration of Flattening Procedures" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.552, + 0.47, + 0.746 + ], + "angle": 0, + "content": "To improve the generalization of deep long-tailed models, we experiment to navigate flatter minima by integrating the existing flattening operations (stochastic weight averaging (SWA) [16], spectral normalization (SN) [4], gradients penalization (GP) [61], and model perturbation (MP) [44]) into five representative deep long-tailed learning baselines to see if they are generally beneficial for DLTR. Illustrated in Figure 2, these flattening methods bring little or negative gains in most cases. Even though in some cases they are beneficial, the corresponding baselines (mainly LDAM-DRW) are too weak so that the final performance are still sub-optimal, which calls for more effective flattening procedures for DLTR. More implementation details are provided in the Appendix C." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.762, + 0.394, + 0.775 + ], + "angle": 0, + "content": "2.2.3 Characteristic Radius of Flat Minima" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.784, + 0.469, + 0.867 + ], + "angle": 0, + "content": "Let's re-consider flattening operations in DLTR from a theoretical perspective. The empirical loss \\( L_{S}(\\boldsymbol{w}) \\) is typically non-convex for deep neural network models, whose landscape may exhibit multiple local or global minima with similar values of \\( L_{S}(\\boldsymbol{w}) \\) while having drastically different generalization performance (i.e., significantly different values" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.127, + 0.893, + 0.32 + ], + "angle": 0, + "content": "of \\( L_{\\mathcal{T}}(\\boldsymbol{w}) \\). In order to find a more effective approach for generalization in DLTR, motivated by the high correlation between sharpness of the loss landscape and model generalization [19], we follow the sharpness-aware minimization (SAM) framework [13] by optimizing an upper bound of \\( L_{\\mathcal{T}}(\\boldsymbol{w}) \\) derived from the PAC-Bayesian framework [34]. Given a prior over the parameters before observing any data and a posterior over the parameters dependent on the training set and learning algorithm, the PAC-Bayesian framework bounds the generalization error of any model in terms of the KL divergence between the two probability distributions. Assuming a Gaussian prior and posterior, we arrive at the following generalization bound (formal proof in Appendix D):" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.331, + 0.893, + 0.387 + ], + "angle": 0, + "content": "Theorem 1 (Perturbative PAC-Bayesian Generalization Bound). For any \\(\\rho >0,0 < \\delta < 1\\), number of samples \\(n\\in \\mathbb{N}^+\\), \\(k\\coloneqq \\dim (\\boldsymbol {w})\\), with probability at least \\(1 - \\delta\\) over a training set \\(\\mathcal{S}\\) sampled i.i.d. from distribution \\(\\mathcal{T}\\)" + }, + { + "type": "equation", + "bbox": [ + 0.5, + 0.408, + 0.9, + 0.448 + ], + "angle": 0, + "content": "\\[\nL_{\\mathcal{T}}(\\boldsymbol {w})\\leq \\max_{||\\boldsymbol {\\epsilon}||_{2}\\leq \\sqrt{k}\\rho}L_{\\mathcal{S}}(\\boldsymbol {w} + \\boldsymbol {\\epsilon}) + \\sqrt{\\frac{\\|\\boldsymbol{w}\\|_{2}^{2}}{4\\rho^{2}} + \\log(\\frac{n}{\\delta}) + \\mathcal{O}(1)}\\\\ \\hskip 142.26378ptn - 1.\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.457, + 0.893, + 0.553 + ], + "angle": 0, + "content": "Intuitively, optimizing the perturbative bound as a surrogate function effectively seeks out parameter solutions whose neighborhoods have uniformly low empirical training loss, thus achieving a flat minimum. Now, our key insight is that the bound is convex with respect to the radius \\(\\rho\\) of the flat minima, hence there exists an optimal \\(\\rho^{*}\\) which gives the tightest generalization bound:" + }, + { + "type": "equation", + "bbox": [ + 0.499, + 0.563, + 0.912, + 0.62 + ], + "angle": 0, + "content": "\\[\n\\rho^ {*} := \\arg \\min _ {\\rho} \\left[ \\max _ {| | \\boldsymbol {\\epsilon} | | _ {2} \\leq \\sqrt {k} \\rho} L _ {\\mathcal {S}} (\\boldsymbol {w} + \\boldsymbol {\\epsilon}) + \\sqrt {\\frac {\\frac {| | \\boldsymbol {w} | | _ {2} ^ {2}}{4 \\rho^ {2}} + \\log \\left(\\frac {n}{\\delta}\\right)}{n - 1}} \\right]. \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.62, + 0.892, + 0.758 + ], + "angle": 0, + "content": "We denote \\(\\rho^{*}\\) as the characteristic radius of the flat minima centered at \\(w\\). It's worth mentioning that such optimal perturbation radius is overlooked by previous literature on sharpness-aware optimization, mainly due to arguments that the PAC-Bayesian bound is too loose to capture the real generalization error [35]. However, our Theorem 1 inspired from [8, 12] ensures a non-vacuous bound (i.e., the square root term falls below 1, see empirical evidence in Appendix B) in the over-parameterized \"deep learning\" regime, rendering the characteristic radius relevant for optimization." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.758, + 0.892, + 0.787 + ], + "angle": 0, + "content": "Moreover, the perturbative bound can be approximated by a first-order Taylor expansion, yielding" + }, + { + "type": "equation", + "bbox": [ + 0.501, + 0.808, + 0.9, + 0.867 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\max _ {\\| \\boldsymbol {\\epsilon} \\| _ {2} \\leq \\sqrt {k} \\rho} L _ {\\mathcal {S}} (\\boldsymbol {w} + \\boldsymbol {\\epsilon}) \\approx \\max _ {\\| \\boldsymbol {\\epsilon} \\| _ {2} \\leq \\sqrt {k} \\rho} \\left[ L _ {\\mathcal {S}} (\\boldsymbol {w}) + \\boldsymbol {\\epsilon} ^ {T} \\nabla_ {\\boldsymbol {w}} L _ {\\mathcal {S}} (\\boldsymbol {w}) \\right] \\\\ = L _ {\\mathcal {S}} (\\boldsymbol {w}) + \\sqrt {k} \\rho | | \\nabla_ {\\boldsymbol {w}} L _ {\\mathcal {S}} (\\boldsymbol {w}) | | _ {2} \\tag {2} \\\\ \\end{array}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.469, + 0.887, + 0.501, + 0.898 + ], + "angle": 0, + "content": "3501" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.078, + 0.123, + 0.506, + 0.255 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.257, + 0.505, + 0.281 + ], + "angle": 0, + "content": "(a) Stage 1: Class-conditional sharpness-aware minimization. We take three classes and classifier perturbation as examples for illustration." + }, + { + "type": "image", + "bbox": [ + 0.529, + 0.141, + 0.895, + 0.256 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.529, + 0.257, + 0.892, + 0.281 + ], + "angle": 0, + "content": "(b) Stage 2: Robust training of the classifier by progressively generating adversarial features." + }, + { + "type": "image_caption", + "bbox": [ + 0.346, + 0.291, + 0.622, + 0.305 + ], + "angle": 0, + "content": "Figure 3. The overall framework of CC-SAM." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.329, + 0.314, + 0.343 + ], + "angle": 0, + "content": "and the optimal perturbation vector" + }, + { + "type": "equation", + "bbox": [ + 0.118, + 0.362, + 0.469, + 0.423 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\hat{\\boldsymbol{\\epsilon}}^{*}(\\boldsymbol {w})\\approx \\operatorname *{arg max}_{||\\boldsymbol {\\epsilon}||_{2}\\leq \\sqrt{k}\\rho}\\left[L_{\\mathcal{S}}(\\boldsymbol {w}) + \\boldsymbol{\\epsilon}^{T}\\nabla_{\\boldsymbol{w}}L_{\\mathcal{S}}(\\boldsymbol {w})\\right] \\\\ = \\sqrt {k} \\rho \\frac {\\nabla_ {\\boldsymbol {w}} L _ {\\mathcal {S}} (\\boldsymbol {w})}{\\| \\nabla_ {\\boldsymbol {w}} L _ {\\mathcal {S}} (\\boldsymbol {w}) \\| _ {2}}. \\tag {3} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.429, + 0.47, + 0.457 + ], + "angle": 0, + "content": "Substitute Eqn 2 into 1, and ignore \\(\\log \\left(\\frac{n}{\\delta}\\right)\\) by assuming \\(\\rho \\ll ||\\pmb{w}||_2\\) as we discussed in Sec 2.2.1, obtaining" + }, + { + "type": "equation", + "bbox": [ + 0.127, + 0.474, + 0.47, + 0.509 + ], + "angle": 0, + "content": "\\[\n\\rho^ {*} \\approx \\left(\\frac {\\| \\boldsymbol {w} \\| _ {2}}{2 \\| \\nabla_ {\\boldsymbol {w}} L _ {\\mathcal {S}} (\\boldsymbol {w}) \\| _ {2}}\\right) ^ {\\frac {1}{2}} k ^ {- \\frac {1}{4}} (n - 1) ^ {- \\frac {1}{4}} \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.515, + 0.426, + 0.528 + ], + "angle": 0, + "content": "with the approximated optimal generalization bound" + }, + { + "type": "equation", + "bbox": [ + 0.085, + 0.545, + 0.469, + 0.577 + ], + "angle": 0, + "content": "\\[\n\\widehat {L} _ {\\mathcal {T}} (\\boldsymbol {w}) \\approx \\max _ {| | \\boldsymbol {\\epsilon} | | _ {2} \\leq \\sqrt {k} \\rho^ {*}} L _ {\\mathcal {S}} (\\boldsymbol {w} + \\boldsymbol {\\epsilon}) + \\frac {1}{2 \\sqrt {n - 1}} \\cdot \\frac {| | \\boldsymbol {w} | | _ {2}}{\\rho^ {*}}. \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.583, + 0.47, + 0.652 + ], + "angle": 0, + "content": "The first term captures the sharpness of \\(L_{\\mathcal{S}}\\) at \\(\\pmb{w}\\), and the second term serves as a regularizer on the magnitude of \\(\\pmb{w}\\) akin to the L2 regularization in SAM [13]. A key difference here is that our regularization term is well motivated by theoretic interpretation from the PAC-Bayesian bound." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.663, + 0.212, + 0.679 + ], + "angle": 0, + "content": "3. Methodology" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.687, + 0.469, + 0.797 + ], + "angle": 0, + "content": "In this section, we introduce our two-stage sharpness-aware optimization algorithm based on Decoupling [20] in detail. In the first stage of the decoupled training, both the feature extractor and classifier are trained under parameter perturbations at a class-conditioned scale. In the second stage, we generate adversarial features to further robustify the classifier while freezing the backbone. The complete pseudo-code is shown in Algorithm 1." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.804, + 0.469, + 0.831 + ], + "angle": 0, + "content": "3.1. Stage 1: Class-Conditional Sharpness-Aware Minimization (CC-SAM)" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.839, + 0.469, + 0.867 + ], + "angle": 0, + "content": "Note that a key assumption we made in Theorem 1 when deriving the characteristic radius of flat minima is that the" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.329, + 0.892, + 0.454 + ], + "angle": 0, + "content": "training data \\( S \\) are drawn i.i.d. from the target distribution \\( \\mathcal{T} \\). However, as we explained in Sec 2.1, such assumption is invalid due to the severe label distribution shift between training and testing sets in DLTR. To resolve this fundamental conflict, we turn to the class conditional distribution \\( p_{s}(\\pmb{x}|\\pmb{y}) \\) and \\( p_{t}(\\pmb{x}|\\pmb{y}) \\) instead, for which the i.i.d. assumption reasonably holds. Accordingly, we decompose the total loss \\( L_{\\mathcal{T}}(\\pmb{w}) \\) and derive a generalization bound for each class separately, obtaining" + }, + { + "type": "equation", + "bbox": [ + 0.51, + 0.462, + 0.892, + 0.5 + ], + "angle": 0, + "content": "\\[\nL _ {\\mathcal {T}} (\\boldsymbol {w}) = \\sum_ {c = 1} ^ {k} L _ {\\mathcal {T}} ^ {c} (\\boldsymbol {w}) = \\sum_ {c = 1} ^ {k} \\mathbb {E} _ {\\boldsymbol {x} \\sim p _ {t} (\\boldsymbol {x} | c)} [ l (\\boldsymbol {w}, \\boldsymbol {x}, c) ], \\tag {6}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.511, + 0.502, + 0.891, + 0.556 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\widehat {L} _ {\\mathcal {T}} ^ {c} (\\boldsymbol {w}) = \\max _ {| | \\boldsymbol {\\epsilon} | | _ {2} \\leq \\sqrt {k} \\rho_ {c} ^ {*}} L _ {\\mathcal {S}} ^ {c} (\\boldsymbol {w} + \\boldsymbol {\\epsilon}) + \\frac {1}{2 \\sqrt {n - 1}} \\cdot \\frac {| | \\boldsymbol {w} | | _ {2}}{\\rho_ {c} ^ {*}}, \\\\ \\approx \\left(2 \\| \\boldsymbol {w} \\| _ {2} \\| \\nabla_ {w} L _ {\\mathcal {S}} ^ {c} (\\boldsymbol {w}) \\| _ {2}\\right) ^ {\\frac {1}{2}} k ^ {\\frac {1}{4}} \\left(n _ {c} - 1\\right) ^ {- \\frac {1}{4}} \\tag {7} \\\\ \\end{array}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.543, + 0.558, + 0.891, + 0.592 + ], + "angle": 0, + "content": "\\[\n\\rho_ {c} ^ {*} = \\left(\\frac {\\| \\boldsymbol {w} \\| _ {2}}{2 \\| \\nabla_ {w} L _ {\\mathcal {S}} ^ {c} (\\boldsymbol {w}) \\| _ {2}}\\right) ^ {\\frac {1}{2}} k ^ {- \\frac {1}{4}} \\left(n _ {c} - 1\\right) ^ {- \\frac {1}{4}}, \\tag {8}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.517, + 0.594, + 0.891, + 0.625 + ], + "angle": 0, + "content": "\\[\n\\hat {\\boldsymbol {\\epsilon}} _ {c} ^ {*} (\\boldsymbol {w}) \\approx \\sqrt {k} p _ {c} ^ {*} \\frac {\\nabla_ {w} L _ {\\mathcal {S}} ^ {c} (\\boldsymbol {w})}{\\| \\nabla_ {w} L _ {\\mathcal {S}} ^ {c} (\\boldsymbol {w}) \\| _ {2}}. \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.652, + 0.892, + 0.818 + ], + "angle": 0, + "content": "An immediate observation of Eqn 8 is that the characteristic radius \\(\\rho_c^*\\) is class-dependent, more specifically, negatively correlated with the label frequency \\(n_c\\). One possible interpretation is that the model is more confident regarding the majority classes with higher \\(n_c\\), due to non-asymptotic estimation error given limited samples, hence it requires a smaller perturbative region in the parameter space to ensure flat loss landscape with respect to that class. Additionally, the class-wise generalization bound \\(\\widehat{L}_{\\mathcal{T}}^{c}(\\boldsymbol{w})\\) in Eqn 7 is positively related to the class-wise gradient norm \\(\\|\\nabla_{\\boldsymbol{w}} L_{\\mathcal{S}}^{c}(\\boldsymbol{w})\\|_2\\), which effectively enforces hard example mining." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.824, + 0.892, + 0.867 + ], + "angle": 0, + "content": "It follows that optimizing the approximated generalization bound \\(\\widehat{L}_{\\mathcal{T}}(\\boldsymbol{w}) = \\sum_{c=1}^{k} \\widehat{L}_{\\mathcal{T}}^{c}(\\boldsymbol{w})\\) yields an effective algorithm in the first stage, namely Class-Conditional" + }, + { + "type": "page_number", + "bbox": [ + 0.469, + 0.887, + 0.502, + 0.898 + ], + "angle": 0, + "content": "3502" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.088, + 0.129, + 0.393, + 0.143 + ], + "angle": 0, + "content": "Algorithm 1: Training Paradigm of CC-SAM" + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.146, + 0.262, + 0.159 + ], + "angle": 0, + "content": "Input: Training Dataset" + }, + { + "type": "equation", + "bbox": [ + 0.146, + 0.16, + 0.354, + 0.174 + ], + "angle": 0, + "content": "\\[\n\\mathcal {S} \\sim p _ {s} (\\boldsymbol {x}, \\boldsymbol {y}) = p _ {s} (\\boldsymbol {x} | \\boldsymbol {y}) p _ {s} (\\boldsymbol {y})\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.174, + 0.355, + 0.187 + ], + "angle": 0, + "content": "Output: Model trained with CC-SAM" + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.192, + 0.157, + 0.205 + ], + "angle": 0, + "content": "Stage 1:" + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.206, + 0.312, + 0.22 + ], + "angle": 0, + "content": "Initialize \\(\\pmb{w} = \\{\\varphi, \\theta\\}\\) randomly" + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.22, + 0.258, + 0.233 + ], + "angle": 0, + "content": "while not converged do" + }, + { + "type": "text", + "bbox": [ + 0.123, + 0.234, + 0.293, + 0.246 + ], + "angle": 0, + "content": "foreach batch \\(\\mathcal{B}_i\\) in \\(S\\) do" + }, + { + "type": "text", + "bbox": [ + 0.148, + 0.246, + 0.384, + 0.26 + ], + "angle": 0, + "content": "Compute empirical loss \\(L_{S}\\) with \\(\\mathcal{B}_i\\)" + }, + { + "type": "text", + "bbox": [ + 0.148, + 0.26, + 0.403, + 0.274 + ], + "angle": 0, + "content": "Estimate the class-specific gradient set" + }, + { + "type": "text", + "bbox": [ + 0.155, + 0.274, + 0.423, + 0.288 + ], + "angle": 0, + "content": "\\(G = \\{g_1,g_2,\\dots,g_k\\}\\) with respect to \\(L_{S}\\)" + }, + { + "type": "text", + "bbox": [ + 0.15, + 0.288, + 0.424, + 0.315 + ], + "angle": 0, + "content": "Perturb \\(\\pmb{w}\\) with \\(G\\) according to Eqn 8 and Eqn 9" + }, + { + "type": "text", + "bbox": [ + 0.15, + 0.315, + 0.37, + 0.33 + ], + "angle": 0, + "content": "Update \\(\\pmb{w}\\) via Eqn 10 and Eqn 11" + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.334, + 0.157, + 0.347 + ], + "angle": 0, + "content": "Stage 2:" + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.348, + 0.162, + 0.361 + ], + "angle": 0, + "content": "Freeze \\(\\varphi\\)" + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.362, + 0.257, + 0.375 + ], + "angle": 0, + "content": "while not converged do" + }, + { + "type": "text", + "bbox": [ + 0.124, + 0.375, + 0.408, + 0.388 + ], + "angle": 0, + "content": "Sample batches via class-balanced sampler" + }, + { + "type": "equation", + "bbox": [ + 0.132, + 0.388, + 0.443, + 0.402 + ], + "angle": 0, + "content": "\\[\n\\mathcal {S} ^ {\\prime} = \\left\\{\\mathcal {B} _ {1} ^ {\\prime}, \\mathcal {B} _ {2} ^ {\\prime}, \\dots , \\mathcal {B} _ {m} ^ {\\prime} \\right\\} \\sim p _ {s} (\\boldsymbol {x} | \\boldsymbol {y}) \\cdot \\operatorname {U n i f o r m} (\\boldsymbol {y})\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.124, + 0.402, + 0.298, + 0.414 + ], + "angle": 0, + "content": "foreach batch \\(\\mathcal{B}_j^\\prime\\) in \\(S^{\\prime}\\) do" + }, + { + "type": "text", + "bbox": [ + 0.148, + 0.414, + 0.398, + 0.428 + ], + "angle": 0, + "content": "Computing empirical loss \\(L_{\\mathcal{S}}\\) with \\(\\mathcal{B}_i^\\prime\\)" + }, + { + "type": "text", + "bbox": [ + 0.148, + 0.428, + 0.42, + 0.441 + ], + "angle": 0, + "content": "Obtain the adversarial feature via Eqn 12" + }, + { + "type": "text", + "bbox": [ + 0.148, + 0.442, + 0.422, + 0.469 + ], + "angle": 0, + "content": "Evaluate overall loss according to Eqn 13 and Eqn 14" + }, + { + "type": "text", + "bbox": [ + 0.148, + 0.469, + 0.294, + 0.483 + ], + "angle": 0, + "content": "Update \\(\\theta\\) accordingly" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.526, + 0.373, + 0.54 + ], + "angle": 0, + "content": "Sharpness-Aware Minimization (CC-SAM):" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.557, + 0.238, + 0.571 + ], + "angle": 0, + "content": "Given learning rate \\(\\eta\\)" + }, + { + "type": "equation", + "bbox": [ + 0.176, + 0.573, + 0.469, + 0.591 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {w} \\leftarrow \\boldsymbol {w} - \\eta \\nabla_ {\\boldsymbol {w}} L _ {\\mathcal {S}} ^ {\\mathrm {C C - S A M}} (\\boldsymbol {w}) \\tag {10}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.196, + 0.594, + 0.469, + 0.631 + ], + "angle": 0, + "content": "\\[\n\\approx \\boldsymbol {w} - \\eta \\sum_ {c = 1} ^ {k} \\nabla_ {\\boldsymbol {w}} \\widehat {L} _ {\\mathcal {T}} ^ {c} (\\boldsymbol {w}) | _ {\\boldsymbol {w} + \\hat {\\epsilon} _ {c} ^ {*} (\\boldsymbol {w})}, \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.647, + 0.47, + 0.731 + ], + "angle": 0, + "content": "which is computationally efficient since it only involves first-order gradients. Even when \\(\\boldsymbol{w}\\) is close to the optimal \\(\\boldsymbol{w}^*\\), where \\(\\nabla_{\\boldsymbol{w}}L_{S}(\\boldsymbol{w})|_{\\boldsymbol{w}^*} \\approx 0\\), the gradient for each class \\(\\nabla_{\\boldsymbol{w}}L_{S}^{c}(\\boldsymbol{w})|_{\\boldsymbol{w}^*}\\) for estimating the optimal perturbation vector \\(\\hat{\\epsilon}_{c}^{*}(\\boldsymbol{w})\\) in Eqn 9 is most likely far from zero, circumventing the need for computing high-order terms." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.746, + 0.431, + 0.761 + ], + "angle": 0, + "content": "3.2. Stage 2: Robust Training of the Classifier" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.77, + 0.47, + 0.868 + ], + "angle": 0, + "content": "For the second stage, we freeze the backbone to maintain feature representation and concentrate on refining the decision boundary of the classifier. Following Decoupling, we adopt the class-balanced sampling strategy in this stage, to rectify the classifier with uniform label distribution. As depicted in the Figure 3(b), when taking original feature \\( z \\) as the input of classifier \\( h(z; \\theta) \\), we can generate adversarial" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.127, + 0.736, + 0.14 + ], + "angle": 0, + "content": "features in forward pass as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.59, + 0.151, + 0.892, + 0.182 + ], + "angle": 0, + "content": "\\[\nz _ {a d v} = z + \\lambda \\frac {\\nabla_ {z} L _ {\\mathcal {S}} (z ; \\theta)}{| | \\nabla_ {z} L _ {\\mathcal {S}} (z ; \\theta) | | _ {2}} \\tag {12}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.636, + 0.201, + 0.892, + 0.216 + ], + "angle": 0, + "content": "\\[\nL _ {\\mathcal {A}} = h \\left(\\boldsymbol {z} _ {\\mathbf {a d v}}; \\boldsymbol {\\theta}\\right) \\tag {13}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.224, + 0.892, + 0.267 + ], + "angle": 0, + "content": "where \\(\\lambda\\) is the hyper-parameter to scale the adversarial gradient \\(\\nabla_{z}L_{\\mathcal{S}}(z;\\pmb {\\theta}) / ||\\nabla_{z}L_{\\mathcal{S}}(z;\\pmb {\\theta})||_{2}\\). Moreover, we employ a progressive strategy to balance \\(L_{\\mathcal{S}}\\) and \\(L_{\\mathcal{A}}\\). At epoch \\(t\\)," + }, + { + "type": "equation", + "bbox": [ + 0.609, + 0.276, + 0.892, + 0.303 + ], + "angle": 0, + "content": "\\[\nL = \\left(1 - \\frac {t}{T}\\right) L _ {\\mathcal {S}} + \\frac {t}{T} L _ {\\mathcal {A}} \\tag {14}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.312, + 0.892, + 0.34 + ], + "angle": 0, + "content": "where \\(T\\) is the total number of epochs in this stage, and \\(L_{\\mathcal{A}}\\) dominates the training progressively." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.354, + 0.616, + 0.368 + ], + "angle": 0, + "content": "4. Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.378, + 0.892, + 0.489 + ], + "angle": 0, + "content": "In this section, we evaluate our method on multiple mainstream DLTR datasets and compare it with popular baselines including the state-of-the-art (SOTA) methods. Moreover, we also report the result on open-set recognition [32, 47] tasks to demonstrate the excellent robustness of CC-SAM to out-of-distribution samples. In the end, an ablation study is presented to verify the effectiveness of our design choices." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.499, + 0.71, + 0.513 + ], + "angle": 0, + "content": "4.1. Datasets and Baselines" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.521, + 0.892, + 0.576 + ], + "angle": 0, + "content": "Following the mainstream evaluation protocol [7, 63], we conduct experiments on five major long-tailed datasets, CIFAR-10-LT, CIFAR-100-LT, Places-LT [32], ImageNet-LT [32], and iNaturalist 2018 [46]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.576, + 0.892, + 0.658 + ], + "angle": 0, + "content": "CIFAR-10-LT/CIFAR-100-LT: These two datasets are sampled from the original CIFAR with different imbalance ratios \\(\\beta = N_{max} / N_{min}\\), where \\(N_{max}\\) and \\(N_{min}\\) are the corresponding number of the most and least frequent classes. Following [26], we set the imbalance ratio as \\(\\{200, 100, 50\\}\\) for evaluation." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.659, + 0.892, + 0.728 + ], + "angle": 0, + "content": "Places-LT & ImageNet-LT: Both datasets were first proposed by OLTR [32]. Places-LT contains 62.5K training images spanning 365 classes in total, with imbalance ratio 996. ImageNet-LT has 115.8K training images covering 1000 categories, with imbalance ratio being 256." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.728, + 0.892, + 0.783 + ], + "angle": 0, + "content": "iNaturalist 2018: As a naturally long-tailed classification dataset, iNaturalist 2018 contains 437.5K trainin images from 8142 categories, and its imbalance ratio is 512. We follow the official spilt in our evaluations." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.784, + 0.892, + 0.868 + ], + "angle": 0, + "content": "For fair comparison, we exclude ensemble or pretraining models in our experiments. The baselines and state-of-the-art methods evaluated include (1) class rebalancing: LDAM-DRW [7], LDAM-DRW + SAM [38], De-confound-TDE [45], Lifted Loss [36], Focal Loss [28], OpenMax [3], BBN [63], LADE [17], DisAlign [58]," + }, + { + "type": "page_number", + "bbox": [ + 0.469, + 0.887, + 0.501, + 0.898 + ], + "angle": 0, + "content": "3503" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.22, + 0.124, + 0.751, + 0.351 + ], + "angle": 0, + "content": "
CIFAR-10-LTCIFAR-100-LT
Imbalance Ratio2001005020010050
CE65.6870.7074.8134.8438.4343.90
CE + Mixup [57]65.8472.9679.4835.8440.0145.16
LDAM-DRW [7]73.5277.0381.0338.9142.0447.62
De-confound-TDE [45]-80.6083.60-44.1550.31
CE + Mixup + cRT [20]73.0679.1584.2141.7345.1250.86
BBN [63]73.4779.8281.1837.2142.5647.02
Contrastive Learning [51]-81.4085.36-46.7251.87
BGP [49]---41.2045.2050.50
MiSLAS [62]77.3182.0685.1642.3347.5052.62
VS + SAM [38]-82.40--46.60-
GCL [26]79.0382.6885.4644.8848.7153.55
CC-SAM80.9483.9286.2245.6650.8353.91
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.36, + 0.893, + 0.389 + ], + "angle": 0, + "content": "Table 1. Top-1 Accuracy on CIFAR-10-LT and CIFAR-100-LT. All methods take ResNet-32 as the backbone. All baseline results except BGP are directly adopted from [26]. \"--\" means the original paper didn't report the corresponding results." + }, + { + "type": "table", + "bbox": [ + 0.079, + 0.409, + 0.468, + 0.515 + ], + "angle": 0, + "content": "
cRTStage 1 + dirStage 1 + magStage 2Acc
✓37.8
✓✓38.9
✓✓37.5
✓✓✓40.1
✓✓✓✓40.6
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.526, + 0.47, + 0.59 + ], + "angle": 0, + "content": "Table 2. Ablation studies on Places-LT. \"Stage \\(1 + \\mathrm{dir}\\) enforces parameter perturbation along the recommended direction \\((\\nabla_w L_S^c (\\pmb {w}) / ||\\nabla_w L_S^c (\\pmb {w})||_2\\) in Eqn 9) with magnitude of 1, whereas \"Stage \\(1+\\) mag\" enforces perturbation with the recommended magnitude \\((\\rho_c^*\\) in Eqn 9) in a random direction." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.616, + 0.468, + 0.685 + ], + "angle": 0, + "content": "LUNA [5], BGP [49], VS+SAM [38] (2) information augmentation: RSG [50], (3) decoupled training: Decouple- \\(\\tau\\) norm [20], cRT [20], MisLAS [62], GCL [26], (4) representation learning: Range Loss [59], OLTR [32], IEM [65], Contrastive Learning [51], ResLT [9]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.686, + 0.468, + 0.824 + ], + "angle": 0, + "content": "Our code is implemented with Pytorch 1.4.0 and all experiments are carried out on Tesla V100 GPUs. We train each model with batch size of 64 (for CIFAR-10-LT and CIFAR-100-LT) / 128 (for Places-LT) / 256 (for ImageNet-LT) / 512 (for iNaturalist 2018), SGD optimizer with momentum of 0.9. We apply similar tricks in Balanced Softmax [40] as the mainstream methods have done. Although we implemented CC-SAM under a two-stage framework, it serves as a general technique to improve the existing DLTR methods without model perturbation (Appendix A.3)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.468, + 0.867 + ], + "angle": 0, + "content": "Intuitively, CC-SAM brings more computation overhead due to additional gradient descents. But we only perturbed the last several layers as an efficient version of CC-SAM in" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.411, + 0.892, + 0.439 + ], + "angle": 0, + "content": "our evaluation. A simple training time comparison is presented in Appendix A.5." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.45, + 0.638, + 0.463 + ], + "angle": 0, + "content": "4.2. Main Results" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.473, + 0.892, + 0.639 + ], + "angle": 0, + "content": "We evaluate CC-SAM on the five mainstream public benchmarks mentioned above, and the corresponding results are reported in Table 1 and Table 3. The best method is bolded and the second best is underlined. According to the evaluations, we observe that CC-SAM unanimously attains the best ranking on CIFAR-LT datasets. On large scale datasets (Places-LT, ImageNet-LT, and iNaturalist 2018), CC-SAM shows competitive performance compared to other advanced methods, too. Specifically, it consistently brings better gains to the medium and tail classes, demonstrating the effectiveness of class-conditional flattening for deep long-tailed recognition." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.651, + 0.771, + 0.666 + ], + "angle": 0, + "content": "4.3. Open Long-Tailed Recognition" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.673, + 0.892, + 0.867 + ], + "angle": 0, + "content": "The ability to detect out-of-distribution samples from the open world, namely open-set recognition [47], provides a unique dimension for evaluating the robustness of deep learning models. Open-set long-tailed recognition, or OLTR [32], combines the open-set problems with deep long-tailed learning to offer even more challenging tasks for model generalization. Following the setting of OLTR [32], we enable CC-SAM to distinguish the close-set and open-set samples by applying a simple, non-learnable prototype-based metric. Results are presented in Table 4. We find that CC-SAM achieves the top F-measure performance and outperforms LUNA [5] on Places-LT, a SOTA with a sophisticated open-set detection method based on hierarchical metrics. The experiments demonstrate the remarkable gen" + }, + { + "type": "page_number", + "bbox": [ + 0.469, + 0.887, + 0.501, + 0.898 + ], + "angle": 0, + "content": "3504" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.146, + 0.124, + 0.825, + 0.662 + ], + "angle": 0, + "content": "
DatasetMethodBackboneManyMediumFewOverall
ImageNet-LTCEResNeXt-5065.937.57.744.4
Decouple-τ-norm [20]ResNet-5056.644.227.446.7
Balanced Softmax [40]ResNeXt-5064.148.233.452.3
LADE [17]ResNeXt-5064.447.734.352.3
RSG [50]ResNeXt-5063.248.232.251.8
DisAlign [58]ResNet-5061.352.231.452.9
ResNeXt-5062.752.131.453.4
ResLT [9]ResNeXt-5063.053.335.552.9
BGP [49]ResNet-50---51.5
MiSLAS [62]ResNet-50---52.7
LDAM-DRW + SAM [38]ResNet-5062.052.134.853.1
GCL [26]ResNet-50---54.9
CC-SAMResNet-5061.449.537.152.4
ResNeXt-5063.153.441.155.4
Places-LTCEResNet-15245.727.38.230.2
Decouple-τ-norm [20]ResNet-15237.840.731.837.9
Balanced Softmax [40]ResNet-15242.039.330.538.6
LADE [17]ResNet-15242.839.031.238.8
RSG [50]ResNet-15241.941.432.039.3
DisAlign [58]ResNet-15240.442.430.139.3
ResLT [9]ResNet-15239.843.631.439.8
MiSLAS [62]ResNet-152---40.2
GCL [26]ResNet-152---40.6
CC-SAMResNet-15241.242.136.440.6
iNaturalist 2018CEResNet-5072.263.057.261.7
Decouple-τ-norm [20]ResNet-5065.665.365.965.6
Balanced Softmax [40]ResNet-50---70.6
LADE [17]ResNet-50---70.0
RSG [50]ResNet-50---70.3
DisAlign [58]ResNet-50---70.6
ResLT [9]ResNet-50---70.2
BGP [49]ResNet-5070.069.969.670.5
MiSLAS [62]ResNet-50---71.6
LDAM-DRW + SAM [38]ResNet-5064.170.571.270.1
GCL [26]ResNet-50---72.0
CC-SAMResNet-5065.470.972.270.9
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.672, + 0.891, + 0.698 + ], + "angle": 0, + "content": "Table 3. Top-1 Accuracy on Places-LT, ImageNet-LT and iNaturalist 2018. As for Places-LT, we take a pre-trained ResNet-152 as the backbone for a fair comparison." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.722, + 0.47, + 0.75 + ], + "angle": 0, + "content": "eralization of CC-SAM due to its capacity to learn highly robust representations." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.761, + 0.231, + 0.777 + ], + "angle": 0, + "content": "4.4. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.784, + 0.47, + 0.867 + ], + "angle": 0, + "content": "Since CC-SAM consists of multiple components, here we provide an ablation study to demonstrate their effectiveness. The result of experiments conducted on Places-LT are shown in Table 2. In particular, CC-SAM degrades to cRT without any of our proposed operations, which we take as a baseline. It is observed that CC-SAM benefits from the" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.722, + 0.892, + 0.751 + ], + "angle": 0, + "content": "design choices of each individual stage, and the integration of both stages shows the best performance." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.756, + 0.893, + 0.867 + ], + "angle": 0, + "content": "We also study the class-conditional perturbation by differentiating the impacts of choosing the right direction \\(\\nabla_{\\boldsymbol{w}}L_{S}^{c}(\\boldsymbol{w}) / ||\\nabla_{\\boldsymbol{w}}L_{S}^{c}(\\boldsymbol{w})||_{2}\\) and the right magnitude \\(\\rho_c^*\\) in Eqn 9, shown as \"stage \\(1 + \\mathrm{dir}\\)\" and \"stage \\(1 + \\mathrm{mag}\\)\" in Table 2 respectively. For the former, we choose a perturbation scale of 1. For the latter, we perturb the model parameters by a random direction sampled from a zero-mean Gaussian. It turns out the perturbation direction contributes more, and" + }, + { + "type": "page_number", + "bbox": [ + 0.469, + 0.887, + 0.502, + 0.898 + ], + "angle": 0, + "content": "3505" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.148, + 0.124, + 0.825, + 0.304 + ], + "angle": 0, + "content": "
ImageNet-LTPlaces-LT
MethodManyMediumFewF-measureManyMediumFewF-measure
CE40.110.40.40.29545.922.40.40.366
Lifted Loss [36]34.829.317.40.37441.035.223.80.459
Focal Loss [28]35.729.315.60.37141.034.822.30.453
Range Loss [59]34.729.417.20.37341.035.323.10.457
OpenMax [3]35.830.017.60.36841.135.423.20.458
OLTR [32]41.933.917.40.47444.636.825.20.464
IEM [65]46.142.320.10.52548.842.428.90.486
LUNA [5]48.244.723.60.57948.141.629.00.491
CC-SAM61.449.537.10.55241.241.836.40.510
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.313, + 0.893, + 0.34 + ], + "angle": 0, + "content": "Table 4. Open long-tail performance of top-1 accuracy on ImageNet-LT and Places-LT. F-measure is a balanced treatment of precision and recall. The backbone of CC-SAM is ResNet-50 for ImageNet-LT, while other compared methods are equipped with ResNet-10." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.363, + 0.471, + 0.392 + ], + "angle": 0, + "content": "applying both attains the best improvement. For more ablation studies, please refer to Appendix A.1." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.404, + 0.349, + 0.419 + ], + "angle": 0, + "content": "5. Related Work and Discussion" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.427, + 0.349, + 0.442 + ], + "angle": 0, + "content": "5.1. Deep Long-Tailed Recognition" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.449, + 0.47, + 0.655 + ], + "angle": 0, + "content": "A latest survey [60] systematically studies up-to-date DLTR algorithms published at the top-tier conferences and introduces a new taxonomy, dividing DLTR methodologies into three categories: class re-balancing, information augmentation, and module improvement. For class rebalancing, a branch of sampling strategies [20, 40, 52, 54] have been proposed to re-balance the training distribution. Meanwhile, some studies design dedicated loss functions [7, 10, 28, 45] to dynamically re-weight the gradients for optimization. As one representative of information augmentation, RSG [50] enriches the feature space of tail classes via dynamical prototypes and a maximized vector loss. Decoupled training [20, 58, 62] propose to learn effective representation and robust classifier according to different sampling strategies in two stages." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.656, + 0.47, + 0.724 + ], + "angle": 0, + "content": "Our method follows the decoupled training paradigm. Moreover, to combat label distribution shift, our method seeks out flat minima by perturbing the model parameters at a class-conditional scale, which is unseen in existing DLTR literature." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.735, + 0.399, + 0.75 + ], + "angle": 0, + "content": "5.2. Flattening for Deep Learning Models" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.756, + 0.47, + 0.868 + ], + "angle": 0, + "content": "Sharpness as a generalization measure of deep learning models has been extensively explored in previous literature [1, 13, 16, 19]. Keskar et al. [21] empirically found that a large batch SGD optimizer might lead to sharp minima and poor generalization. Inspired by this, He et al. [16] implicitly averaged along the SGD trajectory to find the asymmetric valleys, namely asymmetric flat minima. Due to the power of flattening, it has also been applied to many" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.363, + 0.893, + 0.448 + ], + "angle": 0, + "content": "learning tasks to achieve better generalization. Rangwani et al. [37] introduced a domain adversarial training framework by adding an auxiliary loss to smooth the local minima for domain adaptation. Shi et al. [44] demonstrated that a flattened model was resilient to catastrophic forgetting in class-incremental few-shot learning." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.463, + 0.893, + 0.739 + ], + "angle": 0, + "content": "In this work, we investigate flattening in the context of DLTR. To our best knowledge, there are only a few works which succeed in applying flattening to improve generalization in DLTR/OLTR. GBP [49] is built upon the observation of high correlation between logits and gradient norm, and proposes a gradient penalty loss as regularization. However, it lacks rigorous theoretical interpretation by employing a very loose (vacuous) PAC-Bayesian bound and meanwhile under-performs compared to our approach (Table 1, 3). Liu et al. [30] develops a variant of SAM that re-weights the perturbation at instance level to combat data imbalance, without specifying an optimal perturbation or giving any theoretical justification; thus, we regard it as a simplistic and incomplete version of our method. Besides, Rangwani et al. [38] observes that re-weighting techniques in DLTR lead to saddle points for tail classes, and directly leverages SAM to alleviate this problem. In contrast, CC-SAM further advances [38] by providing a theoretically motivated class-conditional SAM design for better generalization and robustness to label distribution shift." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.756, + 0.894, + 0.868 + ], + "angle": 0, + "content": "Acknowledgment We thank anonymous reviewers for their valuable comments. This work was supported by the NSFC under Grants 61932017 and 61971390, and the following grants from the Research Grants Council of the Hong Kong Special Administrative Region, China (Project Reference Number: 14201620 and 14201321). Lanqing Li would like to thank Guangyong Chen and Danruo Deng for their helpful comments on the PAC-Bayesian framework." + }, + { + "type": "page_number", + "bbox": [ + 0.469, + 0.887, + 0.503, + 0.898 + ], + "angle": 0, + "content": "3506" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.125, + 0.173, + 0.139 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.149, + 0.47, + 0.2 + ], + "angle": 0, + "content": "[1] Carlo Baldassi, Fabrizio Pittorino, and Riccardo Zecchina. Shaping the learning landscape in neural networks around wide flat minima. Proceedings of the National Academy of Sciences, 117(1):161-170, 2020. 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.203, + 0.47, + 0.24 + ], + "angle": 0, + "content": "[2] Albert-László Barabási and Réka Albert. Emergence of scaling in random networks. science, 286(5439):509-512, 1999. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.244, + 0.469, + 0.293 + ], + "angle": 0, + "content": "[3] Abhijit Bendale and Terrance E Boult. Towards open set deep networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1563-1572, 2016. 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.297, + 0.469, + 0.335 + ], + "angle": 0, + "content": "[4] Andrew Brock, Jeff Donahue, and Karen Simonyan. Large scale gan training for high fidelity natural image synthesis. arXiv preprint arXiv:1809.11096, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.338, + 0.469, + 0.401 + ], + "angle": 0, + "content": "[5] Jiarui Cai, Yizhou Wang, Hung-Min Hsu, Jenq-Neng Hwang, Kelsey Magrane, and Craig S Rose. Luna: Localizing unfamiliarity near acquaintance for open-set long-tailed recognition. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 36, pages 131-139, 2022. 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.404, + 0.469, + 0.455 + ], + "angle": 0, + "content": "[6] Jiarui Cai, Yizhou Wang, and Jenq-Neng Hwang. Ace: Ally complementary experts for solving long-tailed recognition in one-shot. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 112-121, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.458, + 0.469, + 0.508 + ], + "angle": 0, + "content": "[7] Kaidi Cao, Colin Wei, Adrien Gaidon, Nikos Arechiga, and Tengyu Ma. Learning imbalanced datasets with label-distribution-aware margin loss. Advances in neural information processing systems, 32, 2019. 1, 2, 5, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.511, + 0.469, + 0.561 + ], + "angle": 0, + "content": "[8] Niladri S. Chatterji, Behnam Neyshabur, and Hanie Sedghi. The intriguing role of module criticality in the generalization of deep networks. In 8th International Conference on Learning Representations, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.565, + 0.469, + 0.614 + ], + "angle": 0, + "content": "[9] Jiequan Cui, Shu Liu, Zhuotao Tian, Zhisheng Zhong, and Jiaya Jia. Result: Residual learning for long-tailed recognition. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2022. 1, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.618, + 0.469, + 0.68 + ], + "angle": 0, + "content": "[10] Yin Cui, Menglin Jia, Tsung-Yi Lin, Yang Song, and Serge Belongie. Class-balanced loss based on effective number of samples. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9268-9277, 2019. 1, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.684, + 0.469, + 0.746 + ], + "angle": 0, + "content": "[11] Danruo Deng, Guangyong Chen, Jianye Hao, Qiong Wang, and Pheng-Ann Heng. Flattening sharpness for dynamic gradient projection memory benefits continual learning. Advances in Neural Information Processing Systems, 34:18710-18721, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.75, + 0.469, + 0.813 + ], + "angle": 0, + "content": "[12] Gintare Karolina Dziugaite and Daniel M. Roy. Computing nonvacuous generalization bounds for deep (stochastic) neural networks with many more parameters than training data. In Proceedings of the Thirty-Third Conference on Uncertainty in Artificial Intelligence, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.816, + 0.469, + 0.866 + ], + "angle": 0, + "content": "[13] Pierre Foret, Ariel Kleiner, Hossein Mobahi, and Behnam Neyshabur. Sharpness-aware minimization for efficiently improving generalization. In 9th International Conference on Learning Representations, 2021. 1, 2, 3, 4, 8" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.149, + 0.47, + 0.866 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.128, + 0.892, + 0.166 + ], + "angle": 0, + "content": "[14] Ishaan Gulrajani and David Lopez-Paz. In search of lost domain generalization. In 9th International Conference on Learning Representations, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.167, + 0.892, + 0.205 + ], + "angle": 0, + "content": "[15] Haibo He and Edwardo A Garcia. Learning from imbalanced data. IEEE Transactions on knowledge and data engineering, 21(9):1263-1284, 2009. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.207, + 0.892, + 0.244 + ], + "angle": 0, + "content": "[16] Haowei He, Gao Huang, and Yang Yuan. Asymmetric valleys: Beyond sharp and flat local minima. Advances in neural information processing systems, 32, 2019. 2, 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.246, + 0.892, + 0.309 + ], + "angle": 0, + "content": "[17] Youngkyu Hong, Seungju Han, Kwanghee Choi, Seokjun Seo, Beomsu Kim, and Buru Chang. Disentangling label distribution for long-tailed visual recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6626-6636, 2021. 1, 2, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.311, + 0.892, + 0.386 + ], + "angle": 0, + "content": "[18] Yuanfeng Ji, Lu Zhang, Jiaxiang Wu, Bingzhe Wu, Long-Kai Huang, Tingyang Xu, Yu Rong, Lanqing Li, Jie Ren, Ding Xue, et al. Drugood: Out-of-distribution (ood) dataset curator and benchmark for ai-aided drug discovery-a focus on affinity prediction problems with noise annotations. arXiv preprint arXiv:2201.09637, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.388, + 0.892, + 0.438 + ], + "angle": 0, + "content": "[19] Yiding Jiang, Behnam Neyshabur, Hossein Mobahi, Dilip Krishnan, and Samy Bengio. Fantastic generalization measures and where to find them. In 8th International Conference on Learning Representations, 2020. 2, 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.44, + 0.892, + 0.502 + ], + "angle": 0, + "content": "[20] Bingyi Kang, Saining Xie, Marcus Rohrbach, Zhicheng Yan, Albert Gordo, Jiashi Feng, and Yannis Kalantidis. Decoupling representation and classifier for long-tailed recognition. arXiv preprint arXiv:1910.09217, 2019. 1, 2, 4, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.505, + 0.892, + 0.568 + ], + "angle": 0, + "content": "[21] Nitish Shirish Keskar, Dheevatsa Mudigere, Jorge Nocedal, Mikhail Smelyanskiy, and Ping Tak Peter Tang. On large-batch training for deep learning: Generalization gap and sharp minima. In 5th International Conference on Learning Representations, 2017. 1, 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.57, + 0.892, + 0.62 + ], + "angle": 0, + "content": "[22] Jaehyung Kim, Jongheon Jeong, and Jinwoo Shin. M2m: Imbalanced classification via major-to-minor translation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 13896-13905, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.622, + 0.892, + 0.697 + ], + "angle": 0, + "content": "[23] Pang Wei Koh, Shiori Sagawa, Henrik Marklund, Sang Michael Xie, Marvin Zhang, Akshay Balsubramani, Weihua Hu, Michihiro Yasunaga, Richard Lanas Phillips, Irena Gao, et al. Wilds: A benchmark of in-the-wild distribution shifts. In International Conference on Machine Learning, pages 5637-5664. PMLR, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.7, + 0.892, + 0.762 + ], + "angle": 0, + "content": "[24] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E. Hinton. Imagenet classification with deep convolutional neural networks. In Advances in Neural Information Processing Systems 25: 26th Annual Conference on Neural Information Processing Systems, pages 1106-1114, 2012. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.764, + 0.892, + 0.826 + ], + "angle": 0, + "content": "[25] Lanqing Li, Liang Zeng, Ziqi Gao, Shen Yuan, Yatao Bian, Bingzhe Wu, Hengtong Zhang, Chan Lu, Yang Yu, Wei Liu, et al. Imdrug: A benchmark for deep imbalanced learning in ai-aided drug discovery. arXiv preprint arXiv:2209.07921, 2022.1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.829, + 0.892, + 0.866 + ], + "angle": 0, + "content": "[26] Mengke Li, Yiu-ming Cheung, and Yang Lu. Long-tailed visual recognition via gaussian clouded logit adjustment. In Proceedings of the IEEE/CVF Conference on Computer Vi" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.128, + 0.892, + 0.866 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.469, + 0.887, + 0.501, + 0.898 + ], + "angle": 0, + "content": "3507" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.11, + 0.127, + 0.469, + 0.152 + ], + "angle": 0, + "content": "sion and Pattern Recognition, pages 6929-6938, 2022. 2, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.153, + 0.47, + 0.217 + ], + "angle": 0, + "content": "[27] Shuang Li, Kaixiong Gong, Chi Harold Liu, Yulin Wang, Feng Qiao, and Xinjing Cheng. Metasaug: Meta semantic augmentation for long-tailed visual recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5212-5221, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.218, + 0.469, + 0.267 + ], + "angle": 0, + "content": "[28] Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dollar. Focal loss for dense object detection. In Proceedings of the IEEE international conference on computer vision, pages 2980-2988, 2017. 1, 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.268, + 0.469, + 0.331 + ], + "angle": 0, + "content": "[29] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In European conference on computer vision, pages 740-755. Springer, 2014. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.332, + 0.469, + 0.381 + ], + "angle": 0, + "content": "[30] Hong Liu, Jeff Z HaoChen, Adrien Gaidon, and Tengyu Ma. Self-supervised learning is more robust to dataset imbalance. In International Conference on Learning Representations, 2021. 1, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.383, + 0.469, + 0.445 + ], + "angle": 0, + "content": "[31] Jialun Liu, Yifan Sun, Chuchu Han, Zhaopeng Dou, and Wenhui Li. Deep representation learning on long-tailed data: A learnable embedding augmentation perspective. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2970-2979, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.447, + 0.469, + 0.509 + ], + "angle": 0, + "content": "[32] Ziwei Liu, Zhongqi Miao, Xiaohang Zhan, Jiayun Wang, Boqing Gong, and Stella X Yu. Large-scale long-tailed recognition in an open world. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2537-2546, 2019. 1, 2, 5, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.511, + 0.469, + 0.56 + ], + "angle": 0, + "content": "[33] Jonathan Long, Evan Shelhamer, and Trevor Darrell. Fully convolutional networks for semantic segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3431-3440, 2015. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.562, + 0.469, + 0.598 + ], + "angle": 0, + "content": "[34] David A McAllester. Pac-bayesian model averaging. In Proceedings of the twelfth annual conference on Computational learning theory, pages 164-170, 1999. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.6, + 0.469, + 0.649 + ], + "angle": 0, + "content": "[35] Behnam Neyshabur, Srinadh Bhojanapalli, David McAllester, and Nati Srebro. Exploring generalization in deep learning. Advances in neural information processing systems, 30, 2017. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.651, + 0.469, + 0.712 + ], + "angle": 0, + "content": "[36] Hyun Oh Song, Yu Xiang, Stefanie Jegelka, and Silvio Savarese. Deep metric learning via lifted structured feature embedding. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4004-4012, 2016. 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.715, + 0.469, + 0.776 + ], + "angle": 0, + "content": "[37] Harsh Rangwani, Sumukh K Aithal, Mayank Mishra, Arihant Jain, and Venkatesh Babu Radhakrishnan. A closer look at smoothness in domain adversarial training. In International Conference on Machine Learning, pages 18378-18399. PMLR, 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.778, + 0.469, + 0.839 + ], + "angle": 0, + "content": "[38] Harsh Rangwani, Sumukh K Aithal, Mayank Mishra, and Venkatesh Babu Radhakrishnan. Escaping saddle points for effective generalization on class-imbalanced data. In Advances in Neural Information Processing Systems. 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.842, + 0.469, + 0.866 + ], + "angle": 0, + "content": "[39] William J Reed. The pareto, zipf and other power laws. Economics letters, 74(1):15-19, 2001. 1" + }, + { + "type": "list", + "bbox": [ + 0.081, + 0.127, + 0.47, + 0.866 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.128, + 0.892, + 0.177 + ], + "angle": 0, + "content": "[40] Jiawei Ren, Cunjun Yu, Xiao Ma, Haiyu Zhao, Shuai Yi, et al. Balanced meta-softmax for long-tailed visual recognition. Advances in neural information processing systems, 33:4175-4186, 2020. 1, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.179, + 0.892, + 0.228 + ], + "angle": 0, + "content": "[41] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. Advances in neural information processing systems, 28, 2015. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.23, + 0.893, + 0.292 + ], + "angle": 0, + "content": "[42] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, et al. Imagenet large scale visual recognition challenge. International journal of computer vision, 115(3):211-252, 2015. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.294, + 0.892, + 0.355 + ], + "angle": 0, + "content": "[43] Ikuro Sato, Yamada Ryota, Masayuki Tanaka, Nakamasa Inoue, and Rei Kawakami. Pof: Post-training of feature extractor for improving generalization. In International Conference on Machine Learning, pages 19221-19230. PMLR, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.357, + 0.892, + 0.419 + ], + "angle": 0, + "content": "[44] Guangyuan Shi, Jiaxin Chen, Wenlong Zhang, Li-Ming Zhan, and Xiao-Ming Wu. Overcoming catastrophic forgetting in incremental few-shot learning by finding flat minima. Advances in Neural Information Processing Systems, 34:6747-6761, 2021. 1, 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.421, + 0.893, + 0.483 + ], + "angle": 0, + "content": "[45] Jingru Tan, Changbao Wang, Buyu Li, Quanquan Li, Wanli Ouyang, Changqing Yin, and Junjie Yan. Equalization loss for long-tailed object recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11662-11671, 2020. 1, 5, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.485, + 0.892, + 0.559 + ], + "angle": 0, + "content": "[46] Grant Van Horn, Oisin Mac Aodha, Yang Song, Yin Cui, Chen Sun, Alex Shepard, Hartwig Adam, Pietro Perona, and Serge Belongie. The inaturalist species classification and detection dataset. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 8769-8778, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.561, + 0.892, + 0.599 + ], + "angle": 0, + "content": "[47] Sagar Vaze, Kai Han, Andrea Vedaldi, and Andrew Zisserman. Open-set recognition: A good closed-set classifier is all you need. arXiv preprint arXiv:2110.06207, 2021. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.6, + 0.892, + 0.649 + ], + "angle": 0, + "content": "[48] Athanasios Voulodimos, Nikolaos Doulamis, Anastasios Doulamis, and Eftychios Protopapadakis. Deep learning for computer vision: A brief review. Computational intelligence and neuroscience, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.651, + 0.892, + 0.712 + ], + "angle": 0, + "content": "[49] Dong Wang, Yicheng Liu, Liangji Fang, Fanhua Shang, Yuanyuan Liu, and Hongying Liu. Balanced gradient penalty improves deep long-tailed learning. In Proceedings of the 30th ACM International Conference on Multimedia, pages 5093-5101, 2022. 2, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.714, + 0.893, + 0.776 + ], + "angle": 0, + "content": "[50] Jianfeng Wang, Thomas Lukasiewicz, Xiaolin Hu, Jianfei Cai, and Zhenghua Xu. Rsg: A simple but effective module for learning imbalanced datasets. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3784-3793, 2021. 1, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.778, + 0.892, + 0.841 + ], + "angle": 0, + "content": "[51] Peng Wang, Kai Han, Xiu-Shen Wei, Lei Zhang, and Lei Wang. Contrastive learning based hybrid networks for long-tailed image classification. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 943-952, 2021. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.842, + 0.892, + 0.866 + ], + "angle": 0, + "content": "[52] Tao Wang, Yu Li, Bingyi Kang, Junnan Li, Junhao Liew, Sheng Tang, Steven Hoi, and Jiashi Feng. The devil is" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.128, + 0.893, + 0.866 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.469, + 0.887, + 0.501, + 0.898 + ], + "angle": 0, + "content": "3508" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.127, + 0.469, + 0.166 + ], + "angle": 0, + "content": "in classification: A simple framework for long-tail instance segmentation. In European conference on computer vision, pages 728-744. Springer, 2020. 1, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.167, + 0.469, + 0.217 + ], + "angle": 0, + "content": "[53] Xudong Wang, Long Lian, Zhongqi Miao, Ziwei Liu, and Stella Yu. Long-tailed recognition by routing diverse distribution-aware experts. In International Conference on Learning Representations, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.218, + 0.469, + 0.268 + ], + "angle": 0, + "content": "[54] Yiru Wang, Weihao Gan, Jie Yang, Wei Wu, and Junjie Yan. Dynamic curriculum learning for imbalanced data classification. In Proceedings of the IEEE/CVF international conference on computer vision, pages 5017-5026, 2019. 1, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.269, + 0.469, + 0.332 + ], + "angle": 0, + "content": "[55] Xi Yin, Xiang Yu, Kihyuk Sohn, Xiaoming Liu, and Manmohan Chandraker. Feature transfer learning for face recognition with under-represented data. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5704-5713, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.334, + 0.469, + 0.383 + ], + "angle": 0, + "content": "[56] Liang Zeng, Lanqing Li, Ziqi Gao, Peilin Zhao, and Jian Li. Imgcl: Revisiting graph contrastive learning on imbalanced node classification. arXiv preprint arXiv:2205.11332, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.385, + 0.469, + 0.423 + ], + "angle": 0, + "content": "[57] Hongyi Zhang, Moustapha Cisse, Yann N Dauphin, and David Lopez-Paz. mixup: Beyond empirical risk minimization. arXiv preprint arXiv:1710.09412, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.424, + 0.469, + 0.487 + ], + "angle": 0, + "content": "[58] Songyang Zhang, Zeming Li, Shipeng Yan, Xuming He, and Jian Sun. Distribution alignment: A unified framework for long-tail visual recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2361-2370, 2021. 1, 5, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.488, + 0.469, + 0.55 + ], + "angle": 0, + "content": "[59] Xiao Zhang, Zhiyuan Fang, Yandong Wen, Zhifeng Li, and Yu Qiao. Range loss for deep face recognition with long-tailed training data. In Proceedings of the IEEE International Conference on Computer Vision, pages 5409-5418, 2017. 1, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.552, + 0.469, + 0.59 + ], + "angle": 0, + "content": "[60] Yifan Zhang, Bingyi Kang, Bryan Hooi, Shuicheng Yan, and Jiashi Feng. Deep long-tailed learning: A survey. arXiv preprint arXiv:2110.04596, 2021. 1, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.591, + 0.469, + 0.629 + ], + "angle": 0, + "content": "[61] Yang Zhao, Hao Zhang, and Xiuyuan Hu. Penalizing gradient norm for efficiently improving generalization in deep learning. arXiv preprint arXiv:2202.03599, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.63, + 0.469, + 0.68 + ], + "angle": 0, + "content": "[62] Zhisheng Zhong, Jiequan Cui, Shu Liu, and Jiaya Jia. Improving calibration for long-tailed recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16489-16498, 2021. 1, 2, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.681, + 0.469, + 0.744 + ], + "angle": 0, + "content": "[63] Boyan Zhou, Quan Cui, Xiu-Shen Wei, and Zhao-Min Chen. Bbn: Bilateral-branch network with cumulative learning for long-tailed visual recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9719–9728, 2020. 1, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.745, + 0.469, + 0.796 + ], + "angle": 0, + "content": "[64] Bolei Zhou, Agata Lapedriza, Aditya Khosla, Aude Oliva, and Antonio Torralba. Places: A 10 million image database for scene recognition. IEEE transactions on pattern analysis and machine intelligence, 40(6):1452-1464, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.797, + 0.469, + 0.859 + ], + "angle": 0, + "content": "[65] Linchao Zhu and Yi Yang. Inflated episodic memory with region self-attention for long-tailed visual recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4344–4353, 2020. 1, 6, 8" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.127, + 0.469, + 0.859 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.469, + 0.887, + 0.501, + 0.898 + ], + "angle": 0, + "content": "3509" + } + ] +] \ No newline at end of file diff --git a/2023/Class-Conditional Sharpness-Aware Minimization for Deep Long-Tailed Recognition/5537ac60-2d54-48e8-9091-a4f2d7a57e6f_origin.pdf b/2023/Class-Conditional Sharpness-Aware Minimization for Deep Long-Tailed Recognition/5537ac60-2d54-48e8-9091-a4f2d7a57e6f_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..eaadcae7fed548e143005a4bd45902e247068b08 --- /dev/null +++ b/2023/Class-Conditional Sharpness-Aware Minimization for Deep Long-Tailed Recognition/5537ac60-2d54-48e8-9091-a4f2d7a57e6f_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d68e861169e7f07366f1d3df0ea0aeb4ddd2c25646b703d476da113c444a5168 +size 372483 diff --git a/2023/Class-Conditional Sharpness-Aware Minimization for Deep Long-Tailed Recognition/full.md b/2023/Class-Conditional Sharpness-Aware Minimization for Deep Long-Tailed Recognition/full.md new file mode 100644 index 0000000000000000000000000000000000000000..1430e99701b1a2ecec4843189ebef3b94918aa25 --- /dev/null +++ b/2023/Class-Conditional Sharpness-Aware Minimization for Deep Long-Tailed Recognition/full.md @@ -0,0 +1,388 @@ +# Class-Conditional Sharpness-Aware Minimization for Deep Long-Tailed Recognition + +Zhipeng Zhou $^{1,\dagger}$ , Lanqing Li $^{2,4,\dagger}$ , Peilin Zhao $^{3,\star}$ , Pheng-Ann Heng $^{2}$ , Wei Gong $^{1,\star}$ + +1University of Science and Technology of China, 2The Chinese University of Hong Kong, 3Tencent AI Lab, 4Zhejiang Lab + +zzpl1994@mail.ustc.edu.cn, lanqingli1993@gmail.com, masonzhao@tencent.com, + +pheng@cse.cuhk.edu.hk, weigong@ustc.edu.cn + +# Abstract + +It's widely acknowledged that deep learning models with flatter minima in its loss landscape tend to generalize better. However, such property is under-explored in deep long-tailed recognition (DLTR), a practical problem where the model is required to generalize equally well across all classes when trained on highly imbalanced label distribution. In this paper, through empirical observations, we argue that sharp minima are in fact prevalent in deep long-tailed models, whereas naive integration of existing flattening operations into long-tailed learning algorithms brings little improvement. Instead, we propose an effective two-stage sharpness-aware optimization approach based on the decoupling paradigm in DLTR. In the first stage, both the feature extractor and classifier are trained under parameter perturbations at a class-conditioned scale, which is theoretically motivated by the characteristic radius of flat minima under the PAC-Bayesian framework. In the second stage, we generate adversarial features with class-balanced sampling to further robustify the classifier with the backbone frozen. Extensive experiments on multiple long-tailed visual recognition benchmarks show that, our proposed Class-Conditional Sharpness-Aware Minimization (CC-SAM), achieves competitive performance compared to the state-of-the-arts. Code is available at https://github.com/zzpustc/CC-SAM. + +# 1. Introduction + +Modern deep learning models, composed of multiple neural network layers with millions of parameters, have achieved remarkable successes in computer vision [24, 33, + +† Equal contribution. * Corresponding authors. Work was primarily done when Z. Zhou worked as L. Li's intern at Tencent AI Lab, Shenzhen. + +41, 48]. A key enabler of deep learning is the collection of large-scale datasets [29, 42, 64], which are normally split into training and testing sets with presumably i.i.d. samples. However, such scenario provides relatively trivial tests for the generalization of machine learning models. In practice, label [17, 25, 56] and domain [14, 18, 23] distribution shifts are prevalent, due to the disparity between the data preparation and evaluation protocols. A classical example is imbalanced [15] or long-tailed recognition [60], where a model is trained on highly imbalanced source label distribution $p_{s}(\pmb{y})$ while evaluated on a uniform target label distribution $p_{t}(\pmb{y})$ . + +In this paper, we focus on the practical yet challenging deep long-tailed recognition (DLTR) problem, which is inherent in the visual world [32, 60] with fundamental connections to many disciplines such as the power-law scaling in network science [2] and the Pareto principle in economics [39]. In computer vision, numerous deep long-tailed learning studies have emerged in recent years, which mainly belong to 5 categories: class rebalancing [7, 10, 28, 40, 45, 52, 54], information augmentation [22, 27, 31, 50, 55], decoupled training [20, 58, 62], representation learning [9, 32, 51, 59, 65] and ensemble learning [6, 53, 63]. + +In this work, we propose a novel approach to DLTR from a distinct angle, by seeking out flat minima in the loss landscape of modern neural networks to ensure model robustness under parameter perturbation. Such optimization strategy, termed flattening in our context, have been shown in a myriad of literature to effectively improve generalization of deep learning models in terms of supervised learning [13, 21, 35, 43], self-supervised learning [30] and continual learning [11, 44]. However, application and adaptation of flattening in the context of DLTR remain under-explored. To fill this gap, we first show later in this paper + +(Section 2.2.2) that existing flattening methods are ineffective for long-tailed learning, consistent with the observation from a very recent paper [49], due to severe label distribution shifts. Accordingly, we present a new efficient variant of the sharpness-aware minimization (SAM) [13] technique based on the Decoupling paradigm [20] of DLTR, which leverages the invariance of the class conditional distribution between the source and the target domain. In a nutshell, our contributions are three-fold: + +- Through the lens of flattening, we corroborate a very recent observation [49] that existing sharpness-aware minimization techniques are suboptimal for deep long-tailed learning, justifying the need of more effective approaches to this important and practical issue. +- We introduce Class-Conditional Sharpness-Aware Minimization (CC-SAM), a novel algorithm tailored for DLTR that improves model generalization by robust training against class-conditioned parameter perturbation. This technique is motivated by the characteristic radius of flat minima we derive under the PAC-Bayesian framework and can be implemented efficiently. +- We integrate CC-SAM with the two-stage decoupling paradigm of DLTR. Extensive experiments demonstrate that our method can achieve competitive performance on multiple public DLTR benchmarks, with remarkable robustness to out-of-distribution samples for open long-tailed recognition [32]. + +# 2. Preliminaries + +# 2.1. Problem Setup + +Throughout the paper, we denote scalars as $s$ , vectors as $\mathbf{s}$ , sets as $\mathcal{S}$ and equality by definition as :=. For a typical $d$ -way classification task in DLTR, suppose we are given a training dataset $\mathcal{S} = \bigcup_{i=1}^{n} \{(x_i, y_i)\}$ , where $n$ is the total number of samples. Let's denote $\{n_1, n_2, \ldots, n_d\}$ as the sample number of each class, where $n = \sum_{i} n_i$ . Without loss of generality, we assume $n_i < n_j$ if $i < j$ , and usually $n_d \gg n_1$ , following a highly imbalanced class distribution. We further write the training (source) distribution as $p_s(\boldsymbol{x}, \boldsymbol{y}) = p_s(\boldsymbol{x}|\boldsymbol{y}) p_s(\boldsymbol{y})$ and the testing (target) distribution as $p_t(\boldsymbol{x}, \boldsymbol{y}) = p_t(\boldsymbol{x}|\boldsymbol{y}) p_t(\boldsymbol{y})$ . Consider a family of models parameterized by $\boldsymbol{w} \in \mathcal{W} \subseteq \mathbb{R}^k$ ; given a loss function $l\colon \mathcal{W} \times \mathcal{X} \times \mathcal{Y} \to \mathbb{R}_+$ , we define the empirical training loss $L_{\mathcal{S}}(\boldsymbol{w}) := \frac{1}{n} \sum_{i=1}^{n} l(\boldsymbol{w}, \boldsymbol{x}_i, \boldsymbol{y}_i)$ and the population testing loss $L_{\mathcal{T}}(\boldsymbol{w}) := \mathbb{E}_{(\boldsymbol{x}, \boldsymbol{y}) \sim p_t(\boldsymbol{x}, \boldsymbol{y})} [l(\boldsymbol{w}, \boldsymbol{x}, \boldsymbol{y})]$ . Having observed only $\mathcal{S}$ , the goal is to optimize for model parameters $\boldsymbol{w}$ having lowest risk $L_{\mathcal{T}}(\boldsymbol{w})$ at test time. + +In DLTR, a key challenge is the label distribution shift between the training and testing sets, i.e., $p_{s}(\pmb{y}) \neq$ + +![](images/097dad271cd8c36d9333a4d2611060f1c097d69889acdcb19d328b6d8b747e71.jpg) +(a) 1D global loss landscape. + +![](images/b04f3c48a2472bb0ac3d7f3f578e5fcf36da42dad6729f005c8049c354002a3c.jpg) +(b) 1D local loss landscape. +Figure 1. Loss value vs. noise norm ratio $||\epsilon ||_2 / ||\pmb {\theta}||_2$ . All experiments are conducted on the CIFAR-10-LT with an imbalance ratio of 100, implemented with the same backbone ResNet-32 and averaged over 5 random seeds. + +$p_t(\boldsymbol{y})$ [17]. However, it's logical to assume no distribution shift within each individual class. That is, the training and testing samples are drawn i.i.d. from the same class-conditional distribution, which indicates $p_s(\boldsymbol{x}|\boldsymbol{y}) = p_t(\boldsymbol{x}|\boldsymbol{y})$ . Here, we build our flattening algorithm upon the decoupling paradigm of DLTR [20], where the feature extractor $f(\boldsymbol{x};\boldsymbol{\varphi})$ and classifier $h(\boldsymbol{z};\boldsymbol{\theta})$ are trained by distinct sampling strategies in a two-stage manner, with $\boldsymbol{w} \coloneqq (\varphi, \boldsymbol{\theta})$ . + +# 2.2. Motivation + +Previous studies have shown strong connection between the geometry of the loss landscape and generalization of deep learning models [1, 16, 19, 21]. In this section, we first empirically demonstrate that many deep long-tailed models exhibit sharp minima, and then show that naive application of classical flattening operations results in limited improvement. Motivated by these observations, we theoretically analyze the characteristic radius of flat minima in DLTR to show that a class-conditional flattening procedure is required. + +# 2.2.1 Sharpness in Deep Long-Tailed Models + +To show the pervasive sharpness in current DLTR models, we select two baselines (empirical risk minimization of cross-entropy (CE), LDAM-DRW [7]) and two advanced methods (MiSLAS [62], GCL [26]) from recent literature to conduct experiments1. By perturbing the parameter $\theta$ of the classifier $h$ with increasing noise $\epsilon$ , we observe the loss of each model on CIFAR-10-LT in Figure 1. + +Figure 1(a) shows that, at global scale, the loss of CE is the sharpest across a wide noise range. However, due to the fact that DLTR methods usually employ distinct loss functions, it's hard to compare the sharpness of their loss landscapes in a fair and meaningful way. Moreover, large noise ratio can easily flip the sign of parameters and deteriorate + +![](images/2efcd9cad09ebde024e5a1e3177fc0bee2e08ee77876811779f0a30cee234875.jpg) +Figure 2. Performance gain of naive integration of flattening operations with DLTR algorithms. The number represents the stage where flattening applies. For example, "SN.1.2" means we apply spectral normalization on both stages. Only the first-stage result of LDAM is reported since it's a one-stage approach. All experiments are conducted on CIFAR-100-LT with imbalance ratio of 100. + +the classifier decision boundary. Therefore in this paper, we only consider locally flat minima, i.e., in the perturbative $(||\epsilon ||_2\ll ||\boldsymbol {w}||_2)$ regime shown in Figure 1(b). From this local view, only MiSLAS is observed to have a comparable or flatter minimum than CE, suggesting that the existing DLTR methods have poor generalization and robustness against model parameter perturbation. More detailed 2D loss landscapes covering overall, head and tail classes are depicted in Appendix E, and corroborate our observation in 1D. + +# 2.2.2 Naive Integration of Flattening Procedures + +To improve the generalization of deep long-tailed models, we experiment to navigate flatter minima by integrating the existing flattening operations (stochastic weight averaging (SWA) [16], spectral normalization (SN) [4], gradients penalization (GP) [61], and model perturbation (MP) [44]) into five representative deep long-tailed learning baselines to see if they are generally beneficial for DLTR. Illustrated in Figure 2, these flattening methods bring little or negative gains in most cases. Even though in some cases they are beneficial, the corresponding baselines (mainly LDAM-DRW) are too weak so that the final performance are still sub-optimal, which calls for more effective flattening procedures for DLTR. More implementation details are provided in the Appendix C. + +# 2.2.3 Characteristic Radius of Flat Minima + +Let's re-consider flattening operations in DLTR from a theoretical perspective. The empirical loss $L_{S}(\boldsymbol{w})$ is typically non-convex for deep neural network models, whose landscape may exhibit multiple local or global minima with similar values of $L_{S}(\boldsymbol{w})$ while having drastically different generalization performance (i.e., significantly different values + +of $L_{\mathcal{T}}(\boldsymbol{w})$ . In order to find a more effective approach for generalization in DLTR, motivated by the high correlation between sharpness of the loss landscape and model generalization [19], we follow the sharpness-aware minimization (SAM) framework [13] by optimizing an upper bound of $L_{\mathcal{T}}(\boldsymbol{w})$ derived from the PAC-Bayesian framework [34]. Given a prior over the parameters before observing any data and a posterior over the parameters dependent on the training set and learning algorithm, the PAC-Bayesian framework bounds the generalization error of any model in terms of the KL divergence between the two probability distributions. Assuming a Gaussian prior and posterior, we arrive at the following generalization bound (formal proof in Appendix D): + +Theorem 1 (Perturbative PAC-Bayesian Generalization Bound). For any $\rho >0,0 < \delta < 1$ , number of samples $n\in \mathbb{N}^+$ , $k\coloneqq \dim (\boldsymbol {w})$ , with probability at least $1 - \delta$ over a training set $\mathcal{S}$ sampled i.i.d. from distribution $\mathcal{T}$ + +$$ +L_{\mathcal{T}}(\boldsymbol {w})\leq \max_{||\boldsymbol {\epsilon}||_{2}\leq \sqrt{k}\rho}L_{\mathcal{S}}(\boldsymbol {w} + \boldsymbol {\epsilon}) + \sqrt{\frac{\|\boldsymbol{w}\|_{2}^{2}}{4\rho^{2}} + \log(\frac{n}{\delta}) + \mathcal{O}(1)}\\ \hskip 142.26378ptn - 1. +$$ + +Intuitively, optimizing the perturbative bound as a surrogate function effectively seeks out parameter solutions whose neighborhoods have uniformly low empirical training loss, thus achieving a flat minimum. Now, our key insight is that the bound is convex with respect to the radius $\rho$ of the flat minima, hence there exists an optimal $\rho^{*}$ which gives the tightest generalization bound: + +$$ +\rho^ {*} := \arg \min _ {\rho} \left[ \max _ {| | \boldsymbol {\epsilon} | | _ {2} \leq \sqrt {k} \rho} L _ {\mathcal {S}} (\boldsymbol {w} + \boldsymbol {\epsilon}) + \sqrt {\frac {\frac {| | \boldsymbol {w} | | _ {2} ^ {2}}{4 \rho^ {2}} + \log \left(\frac {n}{\delta}\right)}{n - 1}} \right]. \tag {1} +$$ + +We denote $\rho^{*}$ as the characteristic radius of the flat minima centered at $w$ . It's worth mentioning that such optimal perturbation radius is overlooked by previous literature on sharpness-aware optimization, mainly due to arguments that the PAC-Bayesian bound is too loose to capture the real generalization error [35]. However, our Theorem 1 inspired from [8, 12] ensures a non-vacuous bound (i.e., the square root term falls below 1, see empirical evidence in Appendix B) in the over-parameterized "deep learning" regime, rendering the characteristic radius relevant for optimization. + +Moreover, the perturbative bound can be approximated by a first-order Taylor expansion, yielding + +$$ +\begin{array}{l} \max _ {\| \boldsymbol {\epsilon} \| _ {2} \leq \sqrt {k} \rho} L _ {\mathcal {S}} (\boldsymbol {w} + \boldsymbol {\epsilon}) \approx \max _ {\| \boldsymbol {\epsilon} \| _ {2} \leq \sqrt {k} \rho} \left[ L _ {\mathcal {S}} (\boldsymbol {w}) + \boldsymbol {\epsilon} ^ {T} \nabla_ {\boldsymbol {w}} L _ {\mathcal {S}} (\boldsymbol {w}) \right] \\ = L _ {\mathcal {S}} (\boldsymbol {w}) + \sqrt {k} \rho | | \nabla_ {\boldsymbol {w}} L _ {\mathcal {S}} (\boldsymbol {w}) | | _ {2} \tag {2} \\ \end{array} +$$ + +![](images/16752340528c6eccc9ada79837d80a2dbbe5fcb032792dafd304ada06f4e15d3.jpg) +(a) Stage 1: Class-conditional sharpness-aware minimization. We take three classes and classifier perturbation as examples for illustration. + +![](images/11f0154026619240555f56b110bab0278f9a51eafa90e8353aa96b036bbf8ec5.jpg) +(b) Stage 2: Robust training of the classifier by progressively generating adversarial features. +Figure 3. The overall framework of CC-SAM. + +and the optimal perturbation vector + +$$ +\begin{array}{l} \hat{\boldsymbol{\epsilon}}^{*}(\boldsymbol {w})\approx \operatorname *{arg max}_{||\boldsymbol {\epsilon}||_{2}\leq \sqrt{k}\rho}\left[L_{\mathcal{S}}(\boldsymbol {w}) + \boldsymbol{\epsilon}^{T}\nabla_{\boldsymbol{w}}L_{\mathcal{S}}(\boldsymbol {w})\right] \\ = \sqrt {k} \rho \frac {\nabla_ {\boldsymbol {w}} L _ {\mathcal {S}} (\boldsymbol {w})}{\| \nabla_ {\boldsymbol {w}} L _ {\mathcal {S}} (\boldsymbol {w}) \| _ {2}}. \tag {3} \\ \end{array} +$$ + +Substitute Eqn 2 into 1, and ignore $\log \left(\frac{n}{\delta}\right)$ by assuming $\rho \ll ||\pmb{w}||_2$ as we discussed in Sec 2.2.1, obtaining + +$$ +\rho^ {*} \approx \left(\frac {\| \boldsymbol {w} \| _ {2}}{2 \| \nabla_ {\boldsymbol {w}} L _ {\mathcal {S}} (\boldsymbol {w}) \| _ {2}}\right) ^ {\frac {1}{2}} k ^ {- \frac {1}{4}} (n - 1) ^ {- \frac {1}{4}} \tag {4} +$$ + +with the approximated optimal generalization bound + +$$ +\widehat {L} _ {\mathcal {T}} (\boldsymbol {w}) \approx \max _ {| | \boldsymbol {\epsilon} | | _ {2} \leq \sqrt {k} \rho^ {*}} L _ {\mathcal {S}} (\boldsymbol {w} + \boldsymbol {\epsilon}) + \frac {1}{2 \sqrt {n - 1}} \cdot \frac {| | \boldsymbol {w} | | _ {2}}{\rho^ {*}}. \tag {5} +$$ + +The first term captures the sharpness of $L_{\mathcal{S}}$ at $\pmb{w}$ , and the second term serves as a regularizer on the magnitude of $\pmb{w}$ akin to the L2 regularization in SAM [13]. A key difference here is that our regularization term is well motivated by theoretic interpretation from the PAC-Bayesian bound. + +# 3. Methodology + +In this section, we introduce our two-stage sharpness-aware optimization algorithm based on Decoupling [20] in detail. In the first stage of the decoupled training, both the feature extractor and classifier are trained under parameter perturbations at a class-conditioned scale. In the second stage, we generate adversarial features to further robustify the classifier while freezing the backbone. The complete pseudo-code is shown in Algorithm 1. + +# 3.1. Stage 1: Class-Conditional Sharpness-Aware Minimization (CC-SAM) + +Note that a key assumption we made in Theorem 1 when deriving the characteristic radius of flat minima is that the + +training data $S$ are drawn i.i.d. from the target distribution $\mathcal{T}$ . However, as we explained in Sec 2.1, such assumption is invalid due to the severe label distribution shift between training and testing sets in DLTR. To resolve this fundamental conflict, we turn to the class conditional distribution $p_{s}(\pmb{x}|\pmb{y})$ and $p_{t}(\pmb{x}|\pmb{y})$ instead, for which the i.i.d. assumption reasonably holds. Accordingly, we decompose the total loss $L_{\mathcal{T}}(\pmb{w})$ and derive a generalization bound for each class separately, obtaining + +$$ +L _ {\mathcal {T}} (\boldsymbol {w}) = \sum_ {c = 1} ^ {k} L _ {\mathcal {T}} ^ {c} (\boldsymbol {w}) = \sum_ {c = 1} ^ {k} \mathbb {E} _ {\boldsymbol {x} \sim p _ {t} (\boldsymbol {x} | c)} [ l (\boldsymbol {w}, \boldsymbol {x}, c) ], \tag {6} +$$ + +$$ +\begin{array}{l} \widehat {L} _ {\mathcal {T}} ^ {c} (\boldsymbol {w}) = \max _ {| | \boldsymbol {\epsilon} | | _ {2} \leq \sqrt {k} \rho_ {c} ^ {*}} L _ {\mathcal {S}} ^ {c} (\boldsymbol {w} + \boldsymbol {\epsilon}) + \frac {1}{2 \sqrt {n - 1}} \cdot \frac {| | \boldsymbol {w} | | _ {2}}{\rho_ {c} ^ {*}}, \\ \approx \left(2 \| \boldsymbol {w} \| _ {2} \| \nabla_ {w} L _ {\mathcal {S}} ^ {c} (\boldsymbol {w}) \| _ {2}\right) ^ {\frac {1}{2}} k ^ {\frac {1}{4}} \left(n _ {c} - 1\right) ^ {- \frac {1}{4}} \tag {7} \\ \end{array} +$$ + +$$ +\rho_ {c} ^ {*} = \left(\frac {\| \boldsymbol {w} \| _ {2}}{2 \| \nabla_ {w} L _ {\mathcal {S}} ^ {c} (\boldsymbol {w}) \| _ {2}}\right) ^ {\frac {1}{2}} k ^ {- \frac {1}{4}} \left(n _ {c} - 1\right) ^ {- \frac {1}{4}}, \tag {8} +$$ + +$$ +\hat {\boldsymbol {\epsilon}} _ {c} ^ {*} (\boldsymbol {w}) \approx \sqrt {k} p _ {c} ^ {*} \frac {\nabla_ {w} L _ {\mathcal {S}} ^ {c} (\boldsymbol {w})}{\| \nabla_ {w} L _ {\mathcal {S}} ^ {c} (\boldsymbol {w}) \| _ {2}}. \tag {9} +$$ + +An immediate observation of Eqn 8 is that the characteristic radius $\rho_c^*$ is class-dependent, more specifically, negatively correlated with the label frequency $n_c$ . One possible interpretation is that the model is more confident regarding the majority classes with higher $n_c$ , due to non-asymptotic estimation error given limited samples, hence it requires a smaller perturbative region in the parameter space to ensure flat loss landscape with respect to that class. Additionally, the class-wise generalization bound $\widehat{L}_{\mathcal{T}}^{c}(\boldsymbol{w})$ in Eqn 7 is positively related to the class-wise gradient norm $\|\nabla_{\boldsymbol{w}} L_{\mathcal{S}}^{c}(\boldsymbol{w})\|_2$ , which effectively enforces hard example mining. + +It follows that optimizing the approximated generalization bound $\widehat{L}_{\mathcal{T}}(\boldsymbol{w}) = \sum_{c=1}^{k} \widehat{L}_{\mathcal{T}}^{c}(\boldsymbol{w})$ yields an effective algorithm in the first stage, namely Class-Conditional + +# Algorithm 1: Training Paradigm of CC-SAM + +Input: Training Dataset + +$$ +\mathcal {S} \sim p _ {s} (\boldsymbol {x}, \boldsymbol {y}) = p _ {s} (\boldsymbol {x} | \boldsymbol {y}) p _ {s} (\boldsymbol {y}) +$$ + +Output: Model trained with CC-SAM + +Stage 1: + +Initialize $\pmb{w} = \{\varphi, \theta\}$ randomly + +while not converged do + +foreach batch $\mathcal{B}_i$ in $S$ do + +Compute empirical loss $L_{S}$ with $\mathcal{B}_i$ + +Estimate the class-specific gradient set + +$G = \{g_1,g_2,\dots,g_k\}$ with respect to $L_{S}$ + +Perturb $\pmb{w}$ with $G$ according to Eqn 8 and Eqn 9 + +Update $\pmb{w}$ via Eqn 10 and Eqn 11 + +Stage 2: + +Freeze $\varphi$ + +while not converged do + +Sample batches via class-balanced sampler + +$$ +\mathcal {S} ^ {\prime} = \left\{\mathcal {B} _ {1} ^ {\prime}, \mathcal {B} _ {2} ^ {\prime}, \dots , \mathcal {B} _ {m} ^ {\prime} \right\} \sim p _ {s} (\boldsymbol {x} | \boldsymbol {y}) \cdot \operatorname {U n i f o r m} (\boldsymbol {y}) +$$ + +foreach batch $\mathcal{B}_j^\prime$ in $S^{\prime}$ do + +Computing empirical loss $L_{\mathcal{S}}$ with $\mathcal{B}_i^\prime$ + +Obtain the adversarial feature via Eqn 12 + +Evaluate overall loss according to Eqn 13 and Eqn 14 + +Update $\theta$ accordingly + +Sharpness-Aware Minimization (CC-SAM): + +Given learning rate $\eta$ + +$$ +\boldsymbol {w} \leftarrow \boldsymbol {w} - \eta \nabla_ {\boldsymbol {w}} L _ {\mathcal {S}} ^ {\mathrm {C C - S A M}} (\boldsymbol {w}) \tag {10} +$$ + +$$ +\approx \boldsymbol {w} - \eta \sum_ {c = 1} ^ {k} \nabla_ {\boldsymbol {w}} \widehat {L} _ {\mathcal {T}} ^ {c} (\boldsymbol {w}) | _ {\boldsymbol {w} + \hat {\epsilon} _ {c} ^ {*} (\boldsymbol {w})}, \tag {11} +$$ + +which is computationally efficient since it only involves first-order gradients. Even when $\boldsymbol{w}$ is close to the optimal $\boldsymbol{w}^*$ , where $\nabla_{\boldsymbol{w}}L_{S}(\boldsymbol{w})|_{\boldsymbol{w}^*} \approx 0$ , the gradient for each class $\nabla_{\boldsymbol{w}}L_{S}^{c}(\boldsymbol{w})|_{\boldsymbol{w}^*}$ for estimating the optimal perturbation vector $\hat{\epsilon}_{c}^{*}(\boldsymbol{w})$ in Eqn 9 is most likely far from zero, circumventing the need for computing high-order terms. + +# 3.2. Stage 2: Robust Training of the Classifier + +For the second stage, we freeze the backbone to maintain feature representation and concentrate on refining the decision boundary of the classifier. Following Decoupling, we adopt the class-balanced sampling strategy in this stage, to rectify the classifier with uniform label distribution. As depicted in the Figure 3(b), when taking original feature $z$ as the input of classifier $h(z; \theta)$ , we can generate adversarial + +features in forward pass as follows: + +$$ +z _ {a d v} = z + \lambda \frac {\nabla_ {z} L _ {\mathcal {S}} (z ; \theta)}{| | \nabla_ {z} L _ {\mathcal {S}} (z ; \theta) | | _ {2}} \tag {12} +$$ + +$$ +L _ {\mathcal {A}} = h \left(\boldsymbol {z} _ {\mathbf {a d v}}; \boldsymbol {\theta}\right) \tag {13} +$$ + +where $\lambda$ is the hyper-parameter to scale the adversarial gradient $\nabla_{z}L_{\mathcal{S}}(z;\pmb {\theta}) / ||\nabla_{z}L_{\mathcal{S}}(z;\pmb {\theta})||_{2}$ . Moreover, we employ a progressive strategy to balance $L_{\mathcal{S}}$ and $L_{\mathcal{A}}$ . At epoch $t$ , + +$$ +L = \left(1 - \frac {t}{T}\right) L _ {\mathcal {S}} + \frac {t}{T} L _ {\mathcal {A}} \tag {14} +$$ + +where $T$ is the total number of epochs in this stage, and $L_{\mathcal{A}}$ dominates the training progressively. + +# 4. Evaluation + +In this section, we evaluate our method on multiple mainstream DLTR datasets and compare it with popular baselines including the state-of-the-art (SOTA) methods. Moreover, we also report the result on open-set recognition [32, 47] tasks to demonstrate the excellent robustness of CC-SAM to out-of-distribution samples. In the end, an ablation study is presented to verify the effectiveness of our design choices. + +# 4.1. Datasets and Baselines + +Following the mainstream evaluation protocol [7, 63], we conduct experiments on five major long-tailed datasets, CIFAR-10-LT, CIFAR-100-LT, Places-LT [32], ImageNet-LT [32], and iNaturalist 2018 [46]. + +CIFAR-10-LT/CIFAR-100-LT: These two datasets are sampled from the original CIFAR with different imbalance ratios $\beta = N_{max} / N_{min}$ , where $N_{max}$ and $N_{min}$ are the corresponding number of the most and least frequent classes. Following [26], we set the imbalance ratio as $\{200, 100, 50\}$ for evaluation. + +Places-LT & ImageNet-LT: Both datasets were first proposed by OLTR [32]. Places-LT contains 62.5K training images spanning 365 classes in total, with imbalance ratio 996. ImageNet-LT has 115.8K training images covering 1000 categories, with imbalance ratio being 256. + +iNaturalist 2018: As a naturally long-tailed classification dataset, iNaturalist 2018 contains 437.5K trainin images from 8142 categories, and its imbalance ratio is 512. We follow the official spilt in our evaluations. + +For fair comparison, we exclude ensemble or pretraining models in our experiments. The baselines and state-of-the-art methods evaluated include (1) class rebalancing: LDAM-DRW [7], LDAM-DRW + SAM [38], De-confound-TDE [45], Lifted Loss [36], Focal Loss [28], OpenMax [3], BBN [63], LADE [17], DisAlign [58], + +
CIFAR-10-LTCIFAR-100-LT
Imbalance Ratio2001005020010050
CE65.6870.7074.8134.8438.4343.90
CE + Mixup [57]65.8472.9679.4835.8440.0145.16
LDAM-DRW [7]73.5277.0381.0338.9142.0447.62
De-confound-TDE [45]-80.6083.60-44.1550.31
CE + Mixup + cRT [20]73.0679.1584.2141.7345.1250.86
BBN [63]73.4779.8281.1837.2142.5647.02
Contrastive Learning [51]-81.4085.36-46.7251.87
BGP [49]---41.2045.2050.50
MiSLAS [62]77.3182.0685.1642.3347.5052.62
VS + SAM [38]-82.40--46.60-
GCL [26]79.0382.6885.4644.8848.7153.55
CC-SAM80.9483.9286.2245.6650.8353.91
+ +Table 1. Top-1 Accuracy on CIFAR-10-LT and CIFAR-100-LT. All methods take ResNet-32 as the backbone. All baseline results except BGP are directly adopted from [26]. "--" means the original paper didn't report the corresponding results. + +
cRTStage 1 + dirStage 1 + magStage 2Acc
✓37.8
✓✓38.9
✓✓37.5
✓✓✓40.1
✓✓✓✓40.6
+ +Table 2. Ablation studies on Places-LT. "Stage $1 + \mathrm{dir}$ enforces parameter perturbation along the recommended direction $(\nabla_w L_S^c (\pmb {w}) / ||\nabla_w L_S^c (\pmb {w})||_2$ in Eqn 9) with magnitude of 1, whereas "Stage $1+$ mag" enforces perturbation with the recommended magnitude $(\rho_c^*$ in Eqn 9) in a random direction. + +LUNA [5], BGP [49], VS+SAM [38] (2) information augmentation: RSG [50], (3) decoupled training: Decouple- $\tau$ norm [20], cRT [20], MisLAS [62], GCL [26], (4) representation learning: Range Loss [59], OLTR [32], IEM [65], Contrastive Learning [51], ResLT [9]. + +Our code is implemented with Pytorch 1.4.0 and all experiments are carried out on Tesla V100 GPUs. We train each model with batch size of 64 (for CIFAR-10-LT and CIFAR-100-LT) / 128 (for Places-LT) / 256 (for ImageNet-LT) / 512 (for iNaturalist 2018), SGD optimizer with momentum of 0.9. We apply similar tricks in Balanced Softmax [40] as the mainstream methods have done. Although we implemented CC-SAM under a two-stage framework, it serves as a general technique to improve the existing DLTR methods without model perturbation (Appendix A.3). + +Intuitively, CC-SAM brings more computation overhead due to additional gradient descents. But we only perturbed the last several layers as an efficient version of CC-SAM in + +our evaluation. A simple training time comparison is presented in Appendix A.5. + +# 4.2. Main Results + +We evaluate CC-SAM on the five mainstream public benchmarks mentioned above, and the corresponding results are reported in Table 1 and Table 3. The best method is bolded and the second best is underlined. According to the evaluations, we observe that CC-SAM unanimously attains the best ranking on CIFAR-LT datasets. On large scale datasets (Places-LT, ImageNet-LT, and iNaturalist 2018), CC-SAM shows competitive performance compared to other advanced methods, too. Specifically, it consistently brings better gains to the medium and tail classes, demonstrating the effectiveness of class-conditional flattening for deep long-tailed recognition. + +# 4.3. Open Long-Tailed Recognition + +The ability to detect out-of-distribution samples from the open world, namely open-set recognition [47], provides a unique dimension for evaluating the robustness of deep learning models. Open-set long-tailed recognition, or OLTR [32], combines the open-set problems with deep long-tailed learning to offer even more challenging tasks for model generalization. Following the setting of OLTR [32], we enable CC-SAM to distinguish the close-set and open-set samples by applying a simple, non-learnable prototype-based metric. Results are presented in Table 4. We find that CC-SAM achieves the top F-measure performance and outperforms LUNA [5] on Places-LT, a SOTA with a sophisticated open-set detection method based on hierarchical metrics. The experiments demonstrate the remarkable gen + +
DatasetMethodBackboneManyMediumFewOverall
ImageNet-LTCEResNeXt-5065.937.57.744.4
Decouple-τ-norm [20]ResNet-5056.644.227.446.7
Balanced Softmax [40]ResNeXt-5064.148.233.452.3
LADE [17]ResNeXt-5064.447.734.352.3
RSG [50]ResNeXt-5063.248.232.251.8
DisAlign [58]ResNet-5061.352.231.452.9
ResNeXt-5062.752.131.453.4
ResLT [9]ResNeXt-5063.053.335.552.9
BGP [49]ResNet-50---51.5
MiSLAS [62]ResNet-50---52.7
LDAM-DRW + SAM [38]ResNet-5062.052.134.853.1
GCL [26]ResNet-50---54.9
CC-SAMResNet-5061.449.537.152.4
ResNeXt-5063.153.441.155.4
Places-LTCEResNet-15245.727.38.230.2
Decouple-τ-norm [20]ResNet-15237.840.731.837.9
Balanced Softmax [40]ResNet-15242.039.330.538.6
LADE [17]ResNet-15242.839.031.238.8
RSG [50]ResNet-15241.941.432.039.3
DisAlign [58]ResNet-15240.442.430.139.3
ResLT [9]ResNet-15239.843.631.439.8
MiSLAS [62]ResNet-152---40.2
GCL [26]ResNet-152---40.6
CC-SAMResNet-15241.242.136.440.6
iNaturalist 2018CEResNet-5072.263.057.261.7
Decouple-τ-norm [20]ResNet-5065.665.365.965.6
Balanced Softmax [40]ResNet-50---70.6
LADE [17]ResNet-50---70.0
RSG [50]ResNet-50---70.3
DisAlign [58]ResNet-50---70.6
ResLT [9]ResNet-50---70.2
BGP [49]ResNet-5070.069.969.670.5
MiSLAS [62]ResNet-50---71.6
LDAM-DRW + SAM [38]ResNet-5064.170.571.270.1
GCL [26]ResNet-50---72.0
CC-SAMResNet-5065.470.972.270.9
+ +Table 3. Top-1 Accuracy on Places-LT, ImageNet-LT and iNaturalist 2018. As for Places-LT, we take a pre-trained ResNet-152 as the backbone for a fair comparison. + +eralization of CC-SAM due to its capacity to learn highly robust representations. + +# 4.4. Ablation Study + +Since CC-SAM consists of multiple components, here we provide an ablation study to demonstrate their effectiveness. The result of experiments conducted on Places-LT are shown in Table 2. In particular, CC-SAM degrades to cRT without any of our proposed operations, which we take as a baseline. It is observed that CC-SAM benefits from the + +design choices of each individual stage, and the integration of both stages shows the best performance. + +We also study the class-conditional perturbation by differentiating the impacts of choosing the right direction $\nabla_{\boldsymbol{w}}L_{S}^{c}(\boldsymbol{w}) / ||\nabla_{\boldsymbol{w}}L_{S}^{c}(\boldsymbol{w})||_{2}$ and the right magnitude $\rho_c^*$ in Eqn 9, shown as "stage $1 + \mathrm{dir}$ " and "stage $1 + \mathrm{mag}$ " in Table 2 respectively. For the former, we choose a perturbation scale of 1. For the latter, we perturb the model parameters by a random direction sampled from a zero-mean Gaussian. It turns out the perturbation direction contributes more, and + +
ImageNet-LTPlaces-LT
MethodManyMediumFewF-measureManyMediumFewF-measure
CE40.110.40.40.29545.922.40.40.366
Lifted Loss [36]34.829.317.40.37441.035.223.80.459
Focal Loss [28]35.729.315.60.37141.034.822.30.453
Range Loss [59]34.729.417.20.37341.035.323.10.457
OpenMax [3]35.830.017.60.36841.135.423.20.458
OLTR [32]41.933.917.40.47444.636.825.20.464
IEM [65]46.142.320.10.52548.842.428.90.486
LUNA [5]48.244.723.60.57948.141.629.00.491
CC-SAM61.449.537.10.55241.241.836.40.510
+ +Table 4. Open long-tail performance of top-1 accuracy on ImageNet-LT and Places-LT. F-measure is a balanced treatment of precision and recall. The backbone of CC-SAM is ResNet-50 for ImageNet-LT, while other compared methods are equipped with ResNet-10. + +applying both attains the best improvement. For more ablation studies, please refer to Appendix A.1. + +# 5. Related Work and Discussion + +# 5.1. Deep Long-Tailed Recognition + +A latest survey [60] systematically studies up-to-date DLTR algorithms published at the top-tier conferences and introduces a new taxonomy, dividing DLTR methodologies into three categories: class re-balancing, information augmentation, and module improvement. For class rebalancing, a branch of sampling strategies [20, 40, 52, 54] have been proposed to re-balance the training distribution. Meanwhile, some studies design dedicated loss functions [7, 10, 28, 45] to dynamically re-weight the gradients for optimization. As one representative of information augmentation, RSG [50] enriches the feature space of tail classes via dynamical prototypes and a maximized vector loss. Decoupled training [20, 58, 62] propose to learn effective representation and robust classifier according to different sampling strategies in two stages. + +Our method follows the decoupled training paradigm. Moreover, to combat label distribution shift, our method seeks out flat minima by perturbing the model parameters at a class-conditional scale, which is unseen in existing DLTR literature. + +# 5.2. Flattening for Deep Learning Models + +Sharpness as a generalization measure of deep learning models has been extensively explored in previous literature [1, 13, 16, 19]. Keskar et al. [21] empirically found that a large batch SGD optimizer might lead to sharp minima and poor generalization. Inspired by this, He et al. [16] implicitly averaged along the SGD trajectory to find the asymmetric valleys, namely asymmetric flat minima. Due to the power of flattening, it has also been applied to many + +learning tasks to achieve better generalization. Rangwani et al. [37] introduced a domain adversarial training framework by adding an auxiliary loss to smooth the local minima for domain adaptation. Shi et al. [44] demonstrated that a flattened model was resilient to catastrophic forgetting in class-incremental few-shot learning. + +In this work, we investigate flattening in the context of DLTR. To our best knowledge, there are only a few works which succeed in applying flattening to improve generalization in DLTR/OLTR. GBP [49] is built upon the observation of high correlation between logits and gradient norm, and proposes a gradient penalty loss as regularization. However, it lacks rigorous theoretical interpretation by employing a very loose (vacuous) PAC-Bayesian bound and meanwhile under-performs compared to our approach (Table 1, 3). Liu et al. [30] develops a variant of SAM that re-weights the perturbation at instance level to combat data imbalance, without specifying an optimal perturbation or giving any theoretical justification; thus, we regard it as a simplistic and incomplete version of our method. Besides, Rangwani et al. [38] observes that re-weighting techniques in DLTR lead to saddle points for tail classes, and directly leverages SAM to alleviate this problem. In contrast, CC-SAM further advances [38] by providing a theoretically motivated class-conditional SAM design for better generalization and robustness to label distribution shift. + +Acknowledgment We thank anonymous reviewers for their valuable comments. This work was supported by the NSFC under Grants 61932017 and 61971390, and the following grants from the Research Grants Council of the Hong Kong Special Administrative Region, China (Project Reference Number: 14201620 and 14201321). Lanqing Li would like to thank Guangyong Chen and Danruo Deng for their helpful comments on the PAC-Bayesian framework. + +# References + +[1] Carlo Baldassi, Fabrizio Pittorino, and Riccardo Zecchina. Shaping the learning landscape in neural networks around wide flat minima. Proceedings of the National Academy of Sciences, 117(1):161-170, 2020. 2, 8 +[2] Albert-László Barabási and Réka Albert. Emergence of scaling in random networks. science, 286(5439):509-512, 1999. 1 +[3] Abhijit Bendale and Terrance E Boult. Towards open set deep networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1563-1572, 2016. 5, 8 +[4] Andrew Brock, Jeff Donahue, and Karen Simonyan. Large scale gan training for high fidelity natural image synthesis. arXiv preprint arXiv:1809.11096, 2018. 3 +[5] Jiarui Cai, Yizhou Wang, Hung-Min Hsu, Jenq-Neng Hwang, Kelsey Magrane, and Craig S Rose. Luna: Localizing unfamiliarity near acquaintance for open-set long-tailed recognition. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 36, pages 131-139, 2022. 6, 8 +[6] Jiarui Cai, Yizhou Wang, and Jenq-Neng Hwang. Ace: Ally complementary experts for solving long-tailed recognition in one-shot. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 112-121, 2021. 1 +[7] Kaidi Cao, Colin Wei, Adrien Gaidon, Nikos Arechiga, and Tengyu Ma. Learning imbalanced datasets with label-distribution-aware margin loss. Advances in neural information processing systems, 32, 2019. 1, 2, 5, 6, 8 +[8] Niladri S. Chatterji, Behnam Neyshabur, and Hanie Sedghi. The intriguing role of module criticality in the generalization of deep networks. In 8th International Conference on Learning Representations, 2020. 3 +[9] Jiequan Cui, Shu Liu, Zhuotao Tian, Zhisheng Zhong, and Jiaya Jia. Result: Residual learning for long-tailed recognition. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2022. 1, 6, 7 +[10] Yin Cui, Menglin Jia, Tsung-Yi Lin, Yang Song, and Serge Belongie. Class-balanced loss based on effective number of samples. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9268-9277, 2019. 1, 8 +[11] Danruo Deng, Guangyong Chen, Jianye Hao, Qiong Wang, and Pheng-Ann Heng. Flattening sharpness for dynamic gradient projection memory benefits continual learning. Advances in Neural Information Processing Systems, 34:18710-18721, 2021. 1 +[12] Gintare Karolina Dziugaite and Daniel M. Roy. Computing nonvacuous generalization bounds for deep (stochastic) neural networks with many more parameters than training data. In Proceedings of the Thirty-Third Conference on Uncertainty in Artificial Intelligence, 2017. 3 +[13] Pierre Foret, Ariel Kleiner, Hossein Mobahi, and Behnam Neyshabur. Sharpness-aware minimization for efficiently improving generalization. In 9th International Conference on Learning Representations, 2021. 1, 2, 3, 4, 8 + +[14] Ishaan Gulrajani and David Lopez-Paz. In search of lost domain generalization. In 9th International Conference on Learning Representations, 2021. 1 +[15] Haibo He and Edwardo A Garcia. Learning from imbalanced data. IEEE Transactions on knowledge and data engineering, 21(9):1263-1284, 2009. 1 +[16] Haowei He, Gao Huang, and Yang Yuan. Asymmetric valleys: Beyond sharp and flat local minima. Advances in neural information processing systems, 32, 2019. 2, 3, 8 +[17] Youngkyu Hong, Seungju Han, Kwanghee Choi, Seokjun Seo, Beomsu Kim, and Buru Chang. Disentangling label distribution for long-tailed visual recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6626-6636, 2021. 1, 2, 5, 7 +[18] Yuanfeng Ji, Lu Zhang, Jiaxiang Wu, Bingzhe Wu, Long-Kai Huang, Tingyang Xu, Yu Rong, Lanqing Li, Jie Ren, Ding Xue, et al. Drugood: Out-of-distribution (ood) dataset curator and benchmark for ai-aided drug discovery-a focus on affinity prediction problems with noise annotations. arXiv preprint arXiv:2201.09637, 2022. 1 +[19] Yiding Jiang, Behnam Neyshabur, Hossein Mobahi, Dilip Krishnan, and Samy Bengio. Fantastic generalization measures and where to find them. In 8th International Conference on Learning Representations, 2020. 2, 3, 8 +[20] Bingyi Kang, Saining Xie, Marcus Rohrbach, Zhicheng Yan, Albert Gordo, Jiashi Feng, and Yannis Kalantidis. Decoupling representation and classifier for long-tailed recognition. arXiv preprint arXiv:1910.09217, 2019. 1, 2, 4, 6, 7, 8 +[21] Nitish Shirish Keskar, Dheevatsa Mudigere, Jorge Nocedal, Mikhail Smelyanskiy, and Ping Tak Peter Tang. On large-batch training for deep learning: Generalization gap and sharp minima. In 5th International Conference on Learning Representations, 2017. 1, 2, 8 +[22] Jaehyung Kim, Jongheon Jeong, and Jinwoo Shin. M2m: Imbalanced classification via major-to-minor translation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 13896-13905, 2020. 1 +[23] Pang Wei Koh, Shiori Sagawa, Henrik Marklund, Sang Michael Xie, Marvin Zhang, Akshay Balsubramani, Weihua Hu, Michihiro Yasunaga, Richard Lanas Phillips, Irena Gao, et al. Wilds: A benchmark of in-the-wild distribution shifts. In International Conference on Machine Learning, pages 5637-5664. PMLR, 2021. 1 +[24] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E. Hinton. Imagenet classification with deep convolutional neural networks. In Advances in Neural Information Processing Systems 25: 26th Annual Conference on Neural Information Processing Systems, pages 1106-1114, 2012. 1 +[25] Lanqing Li, Liang Zeng, Ziqi Gao, Shen Yuan, Yatao Bian, Bingzhe Wu, Hengtong Zhang, Chan Lu, Yang Yu, Wei Liu, et al. Imdrug: A benchmark for deep imbalanced learning in ai-aided drug discovery. arXiv preprint arXiv:2209.07921, 2022.1 +[26] Mengke Li, Yiu-ming Cheung, and Yang Lu. Long-tailed visual recognition via gaussian clouded logit adjustment. In Proceedings of the IEEE/CVF Conference on Computer Vi + +sion and Pattern Recognition, pages 6929-6938, 2022. 2, 5, 6, 7 +[27] Shuang Li, Kaixiong Gong, Chi Harold Liu, Yulin Wang, Feng Qiao, and Xinjing Cheng. Metasaug: Meta semantic augmentation for long-tailed visual recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5212-5221, 2021. 1 +[28] Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dollar. Focal loss for dense object detection. In Proceedings of the IEEE international conference on computer vision, pages 2980-2988, 2017. 1, 5, 8 +[29] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In European conference on computer vision, pages 740-755. Springer, 2014. 1 +[30] Hong Liu, Jeff Z HaoChen, Adrien Gaidon, and Tengyu Ma. Self-supervised learning is more robust to dataset imbalance. In International Conference on Learning Representations, 2021. 1, 8 +[31] Jialun Liu, Yifan Sun, Chuchu Han, Zhaopeng Dou, and Wenhui Li. Deep representation learning on long-tailed data: A learnable embedding augmentation perspective. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2970-2979, 2020. 1 +[32] Ziwei Liu, Zhongqi Miao, Xiaohang Zhan, Jiayun Wang, Boqing Gong, and Stella X Yu. Large-scale long-tailed recognition in an open world. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2537-2546, 2019. 1, 2, 5, 6, 8 +[33] Jonathan Long, Evan Shelhamer, and Trevor Darrell. Fully convolutional networks for semantic segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3431-3440, 2015. 1 +[34] David A McAllester. Pac-bayesian model averaging. In Proceedings of the twelfth annual conference on Computational learning theory, pages 164-170, 1999. 3 +[35] Behnam Neyshabur, Srinadh Bhojanapalli, David McAllester, and Nati Srebro. Exploring generalization in deep learning. Advances in neural information processing systems, 30, 2017. 1, 3 +[36] Hyun Oh Song, Yu Xiang, Stefanie Jegelka, and Silvio Savarese. Deep metric learning via lifted structured feature embedding. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4004-4012, 2016. 5, 8 +[37] Harsh Rangwani, Sumukh K Aithal, Mayank Mishra, Arihant Jain, and Venkatesh Babu Radhakrishnan. A closer look at smoothness in domain adversarial training. In International Conference on Machine Learning, pages 18378-18399. PMLR, 2022. 8 +[38] Harsh Rangwani, Sumukh K Aithal, Mayank Mishra, and Venkatesh Babu Radhakrishnan. Escaping saddle points for effective generalization on class-imbalanced data. In Advances in Neural Information Processing Systems. 5, 6, 7, 8 +[39] William J Reed. The pareto, zipf and other power laws. Economics letters, 74(1):15-19, 2001. 1 + +[40] Jiawei Ren, Cunjun Yu, Xiao Ma, Haiyu Zhao, Shuai Yi, et al. Balanced meta-softmax for long-tailed visual recognition. Advances in neural information processing systems, 33:4175-4186, 2020. 1, 6, 7, 8 +[41] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. Advances in neural information processing systems, 28, 2015. 1 +[42] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, et al. Imagenet large scale visual recognition challenge. International journal of computer vision, 115(3):211-252, 2015. 1 +[43] Ikuro Sato, Yamada Ryota, Masayuki Tanaka, Nakamasa Inoue, and Rei Kawakami. Pof: Post-training of feature extractor for improving generalization. In International Conference on Machine Learning, pages 19221-19230. PMLR, 2022. 1 +[44] Guangyuan Shi, Jiaxin Chen, Wenlong Zhang, Li-Ming Zhan, and Xiao-Ming Wu. Overcoming catastrophic forgetting in incremental few-shot learning by finding flat minima. Advances in Neural Information Processing Systems, 34:6747-6761, 2021. 1, 3, 8 +[45] Jingru Tan, Changbao Wang, Buyu Li, Quanquan Li, Wanli Ouyang, Changqing Yin, and Junjie Yan. Equalization loss for long-tailed object recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11662-11671, 2020. 1, 5, 6, 8 +[46] Grant Van Horn, Oisin Mac Aodha, Yang Song, Yin Cui, Chen Sun, Alex Shepard, Hartwig Adam, Pietro Perona, and Serge Belongie. The inaturalist species classification and detection dataset. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 8769-8778, 2018. 5 +[47] Sagar Vaze, Kai Han, Andrea Vedaldi, and Andrew Zisserman. Open-set recognition: A good closed-set classifier is all you need. arXiv preprint arXiv:2110.06207, 2021. 5, 6 +[48] Athanasios Voulodimos, Nikolaos Doulamis, Anastasios Doulamis, and Eftychios Protopapadakis. Deep learning for computer vision: A brief review. Computational intelligence and neuroscience, 2018. 1 +[49] Dong Wang, Yicheng Liu, Liangji Fang, Fanhua Shang, Yuanyuan Liu, and Hongying Liu. Balanced gradient penalty improves deep long-tailed learning. In Proceedings of the 30th ACM International Conference on Multimedia, pages 5093-5101, 2022. 2, 6, 7, 8 +[50] Jianfeng Wang, Thomas Lukasiewicz, Xiaolin Hu, Jianfei Cai, and Zhenghua Xu. Rsg: A simple but effective module for learning imbalanced datasets. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3784-3793, 2021. 1, 6, 7, 8 +[51] Peng Wang, Kai Han, Xiu-Shen Wei, Lei Zhang, and Lei Wang. Contrastive learning based hybrid networks for long-tailed image classification. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 943-952, 2021. 1, 6 +[52] Tao Wang, Yu Li, Bingyi Kang, Junnan Li, Junhao Liew, Sheng Tang, Steven Hoi, and Jiashi Feng. The devil is + +in classification: A simple framework for long-tail instance segmentation. In European conference on computer vision, pages 728-744. Springer, 2020. 1, 8 +[53] Xudong Wang, Long Lian, Zhongqi Miao, Ziwei Liu, and Stella Yu. Long-tailed recognition by routing diverse distribution-aware experts. In International Conference on Learning Representations, 2020. 1 +[54] Yiru Wang, Weihao Gan, Jie Yang, Wei Wu, and Junjie Yan. Dynamic curriculum learning for imbalanced data classification. In Proceedings of the IEEE/CVF international conference on computer vision, pages 5017-5026, 2019. 1, 8 +[55] Xi Yin, Xiang Yu, Kihyuk Sohn, Xiaoming Liu, and Manmohan Chandraker. Feature transfer learning for face recognition with under-represented data. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5704-5713, 2019. 1 +[56] Liang Zeng, Lanqing Li, Ziqi Gao, Peilin Zhao, and Jian Li. Imgcl: Revisiting graph contrastive learning on imbalanced node classification. arXiv preprint arXiv:2205.11332, 2022. 1 +[57] Hongyi Zhang, Moustapha Cisse, Yann N Dauphin, and David Lopez-Paz. mixup: Beyond empirical risk minimization. arXiv preprint arXiv:1710.09412, 2017. 6 +[58] Songyang Zhang, Zeming Li, Shipeng Yan, Xuming He, and Jian Sun. Distribution alignment: A unified framework for long-tail visual recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2361-2370, 2021. 1, 5, 7, 8 +[59] Xiao Zhang, Zhiyuan Fang, Yandong Wen, Zhifeng Li, and Yu Qiao. Range loss for deep face recognition with long-tailed training data. In Proceedings of the IEEE International Conference on Computer Vision, pages 5409-5418, 2017. 1, 6, 8 +[60] Yifan Zhang, Bingyi Kang, Bryan Hooi, Shuicheng Yan, and Jiashi Feng. Deep long-tailed learning: A survey. arXiv preprint arXiv:2110.04596, 2021. 1, 8 +[61] Yang Zhao, Hao Zhang, and Xiuyuan Hu. Penalizing gradient norm for efficiently improving generalization in deep learning. arXiv preprint arXiv:2202.03599, 2022. 3 +[62] Zhisheng Zhong, Jiequan Cui, Shu Liu, and Jiaya Jia. Improving calibration for long-tailed recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16489-16498, 2021. 1, 2, 6, 7, 8 +[63] Boyan Zhou, Quan Cui, Xiu-Shen Wei, and Zhao-Min Chen. Bbn: Bilateral-branch network with cumulative learning for long-tailed visual recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9719–9728, 2020. 1, 5, 6 +[64] Bolei Zhou, Agata Lapedriza, Aditya Khosla, Aude Oliva, and Antonio Torralba. Places: A 10 million image database for scene recognition. IEEE transactions on pattern analysis and machine intelligence, 40(6):1452-1464, 2017. 1 +[65] Linchao Zhu and Yi Yang. Inflated episodic memory with region self-attention for long-tailed visual recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4344–4353, 2020. 1, 6, 8 \ No newline at end of file diff --git a/2023/Class-Conditional Sharpness-Aware Minimization for Deep Long-Tailed Recognition/images.zip b/2023/Class-Conditional Sharpness-Aware Minimization for Deep Long-Tailed Recognition/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..05a87b4250f580aef366f5ca99b0e13d53057056 --- /dev/null +++ b/2023/Class-Conditional Sharpness-Aware Minimization for Deep Long-Tailed Recognition/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:330c6b71a460a5b3909fbd61982700aaaeeffae1b9e6570dbb1f61657c8d72aa +size 509573 diff --git a/2023/Class-Conditional Sharpness-Aware Minimization for Deep Long-Tailed Recognition/layout.json b/2023/Class-Conditional Sharpness-Aware Minimization for Deep Long-Tailed Recognition/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..6f659d501e57f808e943cd477a366d5012c5e054 --- /dev/null +++ b/2023/Class-Conditional Sharpness-Aware Minimization for Deep Long-Tailed Recognition/layout.json @@ -0,0 +1,9971 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 73, + 136, + 505, + 172 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 73, + 136, + 505, + 172 + ], + "spans": [ + { + "bbox": [ + 73, + 136, + 505, + 172 + ], + "type": "text", + "content": "Class-Conditional Sharpness-Aware Minimization for Deep Long-Tailed Recognition" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 73, + 201, + 502, + 218 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 73, + 201, + 502, + 218 + ], + "spans": [ + { + "bbox": [ + 73, + 201, + 502, + 218 + ], + "type": "text", + "content": "Zhipeng Zhou" + }, + { + "bbox": [ + 73, + 201, + 502, + 218 + ], + "type": "inline_equation", + "content": "^{1,\\dagger}" + }, + { + "bbox": [ + 73, + 201, + 502, + 218 + ], + "type": "text", + "content": ", Lanqing Li" + }, + { + "bbox": [ + 73, + 201, + 502, + 218 + ], + "type": "inline_equation", + "content": "^{2,4,\\dagger}" + }, + { + "bbox": [ + 73, + 201, + 502, + 218 + ], + "type": "text", + "content": ", Peilin Zhao" + }, + { + "bbox": [ + 73, + 201, + 502, + 218 + ], + "type": "inline_equation", + "content": "^{3,\\star}" + }, + { + "bbox": [ + 73, + 201, + 502, + 218 + ], + "type": "text", + "content": ", Pheng-Ann Heng" + }, + { + "bbox": [ + 73, + 201, + 502, + 218 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 73, + 201, + 502, + 218 + ], + "type": "text", + "content": ", Wei Gong" + }, + { + "bbox": [ + 73, + 201, + 502, + 218 + ], + "type": "inline_equation", + "content": "^{1,\\star}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 72, + 224, + 509, + 253 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 72, + 224, + 509, + 253 + ], + "spans": [ + { + "bbox": [ + 72, + 224, + 509, + 253 + ], + "type": "text", + "content": "1University of Science and Technology of China, 2The Chinese University of Hong Kong, 3Tencent AI Lab, 4Zhejiang Lab" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 95, + 255, + 480, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 95, + 255, + 480, + 266 + ], + "spans": [ + { + "bbox": [ + 95, + 255, + 480, + 266 + ], + "type": "text", + "content": "zzpl1994@mail.ustc.edu.cn, lanqingli1993@gmail.com, masonzhao@tencent.com," + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 176, + 269, + 400, + 281 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 176, + 269, + 400, + 281 + ], + "spans": [ + { + "bbox": [ + 176, + 269, + 400, + 281 + ], + "type": "text", + "content": "pheng@cse.cuhk.edu.hk, weigong@ustc.edu.cn" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 139, + 314, + 186, + 326 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 139, + 314, + 186, + 326 + ], + "spans": [ + { + "bbox": [ + 139, + 314, + 186, + 326 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 45, + 338, + 281, + 630 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 338, + 281, + 630 + ], + "spans": [ + { + "bbox": [ + 45, + 338, + 281, + 630 + ], + "type": "text", + "content": "It's widely acknowledged that deep learning models with flatter minima in its loss landscape tend to generalize better. However, such property is under-explored in deep long-tailed recognition (DLTR), a practical problem where the model is required to generalize equally well across all classes when trained on highly imbalanced label distribution. In this paper, through empirical observations, we argue that sharp minima are in fact prevalent in deep long-tailed models, whereas naive integration of existing flattening operations into long-tailed learning algorithms brings little improvement. Instead, we propose an effective two-stage sharpness-aware optimization approach based on the decoupling paradigm in DLTR. In the first stage, both the feature extractor and classifier are trained under parameter perturbations at a class-conditioned scale, which is theoretically motivated by the characteristic radius of flat minima under the PAC-Bayesian framework. In the second stage, we generate adversarial features with class-balanced sampling to further robustify the classifier with the backbone frozen. Extensive experiments on multiple long-tailed visual recognition benchmarks show that, our proposed Class-Conditional Sharpness-Aware Minimization (CC-SAM), achieves competitive performance compared to the state-of-the-arts. Code is available at https://github.com/zzpustc/CC-SAM." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 650, + 124, + 662 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 650, + 124, + 662 + ], + "spans": [ + { + "bbox": [ + 46, + 650, + 124, + 662 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 45, + 669, + 279, + 705 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 669, + 279, + 705 + ], + "spans": [ + { + "bbox": [ + 45, + 669, + 279, + 705 + ], + "type": "text", + "content": "Modern deep learning models, composed of multiple neural network layers with millions of parameters, have achieved remarkable successes in computer vision [24, 33," + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 45, + 710, + 279, + 730 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 710, + 279, + 730 + ], + "spans": [ + { + "bbox": [ + 45, + 710, + 279, + 730 + ], + "type": "text", + "content": "† Equal contribution. * Corresponding authors. Work was primarily done when Z. Zhou worked as L. Li's intern at Tencent AI Lab, Shenzhen." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 296, + 314, + 530, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 314, + 530, + 456 + ], + "spans": [ + { + "bbox": [ + 296, + 314, + 530, + 456 + ], + "type": "text", + "content": "41, 48]. A key enabler of deep learning is the collection of large-scale datasets [29, 42, 64], which are normally split into training and testing sets with presumably i.i.d. samples. However, such scenario provides relatively trivial tests for the generalization of machine learning models. In practice, label [17, 25, 56] and domain [14, 18, 23] distribution shifts are prevalent, due to the disparity between the data preparation and evaluation protocols. A classical example is imbalanced [15] or long-tailed recognition [60], where a model is trained on highly imbalanced source label distribution " + }, + { + "bbox": [ + 296, + 314, + 530, + 456 + ], + "type": "inline_equation", + "content": "p_{s}(\\pmb{y})" + }, + { + "bbox": [ + 296, + 314, + 530, + 456 + ], + "type": "text", + "content": " while evaluated on a uniform target label distribution " + }, + { + "bbox": [ + 296, + 314, + 530, + 456 + ], + "type": "inline_equation", + "content": "p_{t}(\\pmb{y})" + }, + { + "bbox": [ + 296, + 314, + 530, + 456 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 296, + 458, + 531, + 598 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 458, + 531, + 598 + ], + "spans": [ + { + "bbox": [ + 296, + 458, + 531, + 598 + ], + "type": "text", + "content": "In this paper, we focus on the practical yet challenging deep long-tailed recognition (DLTR) problem, which is inherent in the visual world [32, 60] with fundamental connections to many disciplines such as the power-law scaling in network science [2] and the Pareto principle in economics [39]. In computer vision, numerous deep long-tailed learning studies have emerged in recent years, which mainly belong to 5 categories: class rebalancing [7, 10, 28, 40, 45, 52, 54], information augmentation [22, 27, 31, 50, 55], decoupled training [20, 58, 62], representation learning [9, 32, 51, 59, 65] and ensemble learning [6, 53, 63]." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 296, + 601, + 532, + 730 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 601, + 532, + 730 + ], + "spans": [ + { + "bbox": [ + 296, + 601, + 532, + 730 + ], + "type": "text", + "content": "In this work, we propose a novel approach to DLTR from a distinct angle, by seeking out flat minima in the loss landscape of modern neural networks to ensure model robustness under parameter perturbation. Such optimization strategy, termed flattening in our context, have been shown in a myriad of literature to effectively improve generalization of deep learning models in terms of supervised learning [13, 21, 35, 43], self-supervised learning [30] and continual learning [11, 44]. However, application and adaptation of flattening in the context of DLTR remain under-explored. To fill this gap, we first show later in this paper" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 76, + 42, + 119, + 71 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 42, + 119, + 71 + ], + "spans": [ + { + "bbox": [ + 76, + 42, + 119, + 71 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 150, + 41, + 471, + 74 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 150, + 41, + 471, + 74 + ], + "spans": [ + { + "bbox": [ + 150, + 41, + 471, + 74 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 279, + 746, + 299, + 756 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 279, + 746, + 299, + 756 + ], + "spans": [ + { + "bbox": [ + 279, + 746, + 299, + 756 + ], + "type": "text", + "content": "3499" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 595, + 842 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 45, + 106, + 281, + 210 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 106, + 281, + 210 + ], + "spans": [ + { + "bbox": [ + 45, + 106, + 281, + 210 + ], + "type": "text", + "content": "(Section 2.2.2) that existing flattening methods are ineffective for long-tailed learning, consistent with the observation from a very recent paper [49], due to severe label distribution shifts. Accordingly, we present a new efficient variant of the sharpness-aware minimization (SAM) [13] technique based on the Decoupling paradigm [20] of DLTR, which leverages the invariance of the class conditional distribution between the source and the target domain. In a nutshell, our contributions are three-fold:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 221, + 279, + 460 + ], + "type": "list", + "angle": 0, + "index": 4, + "blocks": [ + { + "bbox": [ + 57, + 221, + 279, + 279 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 221, + 279, + 279 + ], + "spans": [ + { + "bbox": [ + 57, + 221, + 279, + 279 + ], + "type": "text", + "content": "- Through the lens of flattening, we corroborate a very recent observation [49] that existing sharpness-aware minimization techniques are suboptimal for deep long-tailed learning, justifying the need of more effective approaches to this important and practical issue." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 57, + 287, + 279, + 383 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 287, + 279, + 383 + ], + "spans": [ + { + "bbox": [ + 57, + 287, + 279, + 383 + ], + "type": "text", + "content": "- We introduce Class-Conditional Sharpness-Aware Minimization (CC-SAM), a novel algorithm tailored for DLTR that improves model generalization by robust training against class-conditioned parameter perturbation. This technique is motivated by the characteristic radius of flat minima we derive under the PAC-Bayesian framework and can be implemented efficiently." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 389, + 279, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 389, + 279, + 460 + ], + "spans": [ + { + "bbox": [ + 57, + 389, + 279, + 460 + ], + "type": "text", + "content": "- We integrate CC-SAM with the two-stage decoupling paradigm of DLTR. Extensive experiments demonstrate that our method can achieve competitive performance on multiple public DLTR benchmarks, with remarkable robustness to out-of-distribution samples for open long-tailed recognition [32]." + } + ] + } + ], + "index": 3 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 46, + 471, + 128, + 482 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 471, + 128, + 482 + ], + "spans": [ + { + "bbox": [ + 46, + 471, + 128, + 482 + ], + "type": "text", + "content": "2. Preliminaries" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 490, + 135, + 502 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 490, + 135, + 502 + ], + "spans": [ + { + "bbox": [ + 46, + 490, + 135, + 502 + ], + "type": "text", + "content": "2.1. Problem Setup" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 45, + 508, + 279, + 706 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 508, + 279, + 706 + ], + "spans": [ + { + "bbox": [ + 45, + 508, + 279, + 706 + ], + "type": "text", + "content": "Throughout the paper, we denote scalars as " + }, + { + "bbox": [ + 45, + 508, + 279, + 706 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 45, + 508, + 279, + 706 + ], + "type": "text", + "content": ", vectors as " + }, + { + "bbox": [ + 45, + 508, + 279, + 706 + ], + "type": "inline_equation", + "content": "\\mathbf{s}" + }, + { + "bbox": [ + 45, + 508, + 279, + 706 + ], + "type": "text", + "content": ", sets as " + }, + { + "bbox": [ + 45, + 508, + 279, + 706 + ], + "type": "inline_equation", + "content": "\\mathcal{S}" + }, + { + "bbox": [ + 45, + 508, + 279, + 706 + ], + "type": "text", + "content": " and equality by definition as :=. For a typical " + }, + { + "bbox": [ + 45, + 508, + 279, + 706 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 45, + 508, + 279, + 706 + ], + "type": "text", + "content": "-way classification task in DLTR, suppose we are given a training dataset " + }, + { + "bbox": [ + 45, + 508, + 279, + 706 + ], + "type": "inline_equation", + "content": "\\mathcal{S} = \\bigcup_{i=1}^{n} \\{(x_i, y_i)\\}" + }, + { + "bbox": [ + 45, + 508, + 279, + 706 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 45, + 508, + 279, + 706 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 45, + 508, + 279, + 706 + ], + "type": "text", + "content": " is the total number of samples. Let's denote " + }, + { + "bbox": [ + 45, + 508, + 279, + 706 + ], + "type": "inline_equation", + "content": "\\{n_1, n_2, \\ldots, n_d\\}" + }, + { + "bbox": [ + 45, + 508, + 279, + 706 + ], + "type": "text", + "content": " as the sample number of each class, where " + }, + { + "bbox": [ + 45, + 508, + 279, + 706 + ], + "type": "inline_equation", + "content": "n = \\sum_{i} n_i" + }, + { + "bbox": [ + 45, + 508, + 279, + 706 + ], + "type": "text", + "content": ". Without loss of generality, we assume " + }, + { + "bbox": [ + 45, + 508, + 279, + 706 + ], + "type": "inline_equation", + "content": "n_i < n_j" + }, + { + "bbox": [ + 45, + 508, + 279, + 706 + ], + "type": "text", + "content": " if " + }, + { + "bbox": [ + 45, + 508, + 279, + 706 + ], + "type": "inline_equation", + "content": "i < j" + }, + { + "bbox": [ + 45, + 508, + 279, + 706 + ], + "type": "text", + "content": ", and usually " + }, + { + "bbox": [ + 45, + 508, + 279, + 706 + ], + "type": "inline_equation", + "content": "n_d \\gg n_1" + }, + { + "bbox": [ + 45, + 508, + 279, + 706 + ], + "type": "text", + "content": ", following a highly imbalanced class distribution. We further write the training (source) distribution as " + }, + { + "bbox": [ + 45, + 508, + 279, + 706 + ], + "type": "inline_equation", + "content": "p_s(\\boldsymbol{x}, \\boldsymbol{y}) = p_s(\\boldsymbol{x}|\\boldsymbol{y}) p_s(\\boldsymbol{y})" + }, + { + "bbox": [ + 45, + 508, + 279, + 706 + ], + "type": "text", + "content": " and the testing (target) distribution as " + }, + { + "bbox": [ + 45, + 508, + 279, + 706 + ], + "type": "inline_equation", + "content": "p_t(\\boldsymbol{x}, \\boldsymbol{y}) = p_t(\\boldsymbol{x}|\\boldsymbol{y}) p_t(\\boldsymbol{y})" + }, + { + "bbox": [ + 45, + 508, + 279, + 706 + ], + "type": "text", + "content": ". Consider a family of models parameterized by " + }, + { + "bbox": [ + 45, + 508, + 279, + 706 + ], + "type": "inline_equation", + "content": "\\boldsymbol{w} \\in \\mathcal{W} \\subseteq \\mathbb{R}^k" + }, + { + "bbox": [ + 45, + 508, + 279, + 706 + ], + "type": "text", + "content": "; given a loss function " + }, + { + "bbox": [ + 45, + 508, + 279, + 706 + ], + "type": "inline_equation", + "content": "l\\colon \\mathcal{W} \\times \\mathcal{X} \\times \\mathcal{Y} \\to \\mathbb{R}_+" + }, + { + "bbox": [ + 45, + 508, + 279, + 706 + ], + "type": "text", + "content": ", we define the empirical training loss " + }, + { + "bbox": [ + 45, + 508, + 279, + 706 + ], + "type": "inline_equation", + "content": "L_{\\mathcal{S}}(\\boldsymbol{w}) := \\frac{1}{n} \\sum_{i=1}^{n} l(\\boldsymbol{w}, \\boldsymbol{x}_i, \\boldsymbol{y}_i)" + }, + { + "bbox": [ + 45, + 508, + 279, + 706 + ], + "type": "text", + "content": " and the population testing loss " + }, + { + "bbox": [ + 45, + 508, + 279, + 706 + ], + "type": "inline_equation", + "content": "L_{\\mathcal{T}}(\\boldsymbol{w}) := \\mathbb{E}_{(\\boldsymbol{x}, \\boldsymbol{y}) \\sim p_t(\\boldsymbol{x}, \\boldsymbol{y})} [l(\\boldsymbol{w}, \\boldsymbol{x}, \\boldsymbol{y})]" + }, + { + "bbox": [ + 45, + 508, + 279, + 706 + ], + "type": "text", + "content": ". Having observed only " + }, + { + "bbox": [ + 45, + 508, + 279, + 706 + ], + "type": "inline_equation", + "content": "\\mathcal{S}" + }, + { + "bbox": [ + 45, + 508, + 279, + 706 + ], + "type": "text", + "content": ", the goal is to optimize for model parameters " + }, + { + "bbox": [ + 45, + 508, + 279, + 706 + ], + "type": "inline_equation", + "content": "\\boldsymbol{w}" + }, + { + "bbox": [ + 45, + 508, + 279, + 706 + ], + "type": "text", + "content": " having lowest risk " + }, + { + "bbox": [ + 45, + 508, + 279, + 706 + ], + "type": "inline_equation", + "content": "L_{\\mathcal{T}}(\\boldsymbol{w})" + }, + { + "bbox": [ + 45, + 508, + 279, + 706 + ], + "type": "text", + "content": " at test time." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 45, + 706, + 279, + 730 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 706, + 279, + 730 + ], + "spans": [ + { + "bbox": [ + 45, + 706, + 279, + 730 + ], + "type": "text", + "content": "In DLTR, a key challenge is the label distribution shift between the training and testing sets, i.e., " + }, + { + "bbox": [ + 45, + 706, + 279, + 730 + ], + "type": "inline_equation", + "content": "p_{s}(\\pmb{y}) \\neq" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 298, + 105, + 407, + 195 + ], + "blocks": [ + { + "bbox": [ + 298, + 105, + 407, + 195 + ], + "lines": [ + { + "bbox": [ + 298, + 105, + 407, + 195 + ], + "spans": [ + { + "bbox": [ + 298, + 105, + 407, + 195 + ], + "type": "image", + "image_path": "097dad271cd8c36d9333a4d2611060f1c097d69889acdcb19d328b6d8b747e71.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 306, + 197, + 398, + 207 + ], + "lines": [ + { + "bbox": [ + 306, + 197, + 398, + 207 + ], + "spans": [ + { + "bbox": [ + 306, + 197, + 398, + 207 + ], + "type": "text", + "content": "(a) 1D global loss landscape." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 417, + 105, + 529, + 195 + ], + "blocks": [ + { + "bbox": [ + 417, + 105, + 529, + 195 + ], + "lines": [ + { + "bbox": [ + 417, + 105, + 529, + 195 + ], + "spans": [ + { + "bbox": [ + 417, + 105, + 529, + 195 + ], + "type": "image", + "image_path": "b04f3c48a2472bb0ac3d7f3f578e5fcf36da42dad6729f005c8049c354002a3c.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 430, + 197, + 517, + 207 + ], + "lines": [ + { + "bbox": [ + 430, + 197, + 517, + 207 + ], + "spans": [ + { + "bbox": [ + 430, + 197, + 517, + 207 + ], + "type": "text", + "content": "(b) 1D local loss landscape." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 296, + 216, + 531, + 259 + ], + "lines": [ + { + "bbox": [ + 296, + 216, + 531, + 259 + ], + "spans": [ + { + "bbox": [ + 296, + 216, + 531, + 259 + ], + "type": "text", + "content": "Figure 1. Loss value vs. noise norm ratio " + }, + { + "bbox": [ + 296, + 216, + 531, + 259 + ], + "type": "inline_equation", + "content": "||\\epsilon ||_2 / ||\\pmb {\\theta}||_2" + }, + { + "bbox": [ + 296, + 216, + 531, + 259 + ], + "type": "text", + "content": ". All experiments are conducted on the CIFAR-10-LT with an imbalance ratio of 100, implemented with the same backbone ResNet-32 and averaged over 5 random seeds." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 296, + 280, + 530, + 385 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 280, + 530, + 385 + ], + "spans": [ + { + "bbox": [ + 296, + 280, + 530, + 385 + ], + "type": "inline_equation", + "content": "p_t(\\boldsymbol{y})" + }, + { + "bbox": [ + 296, + 280, + 530, + 385 + ], + "type": "text", + "content": " [17]. However, it's logical to assume no distribution shift within each individual class. That is, the training and testing samples are drawn i.i.d. from the same class-conditional distribution, which indicates " + }, + { + "bbox": [ + 296, + 280, + 530, + 385 + ], + "type": "inline_equation", + "content": "p_s(\\boldsymbol{x}|\\boldsymbol{y}) = p_t(\\boldsymbol{x}|\\boldsymbol{y})" + }, + { + "bbox": [ + 296, + 280, + 530, + 385 + ], + "type": "text", + "content": ". Here, we build our flattening algorithm upon the decoupling paradigm of DLTR [20], where the feature extractor " + }, + { + "bbox": [ + 296, + 280, + 530, + 385 + ], + "type": "inline_equation", + "content": "f(\\boldsymbol{x};\\boldsymbol{\\varphi})" + }, + { + "bbox": [ + 296, + 280, + 530, + 385 + ], + "type": "text", + "content": " and classifier " + }, + { + "bbox": [ + 296, + 280, + 530, + 385 + ], + "type": "inline_equation", + "content": "h(\\boldsymbol{z};\\boldsymbol{\\theta})" + }, + { + "bbox": [ + 296, + 280, + 530, + 385 + ], + "type": "text", + "content": " are trained by distinct sampling strategies in a two-stage manner, with " + }, + { + "bbox": [ + 296, + 280, + 530, + 385 + ], + "type": "inline_equation", + "content": "\\boldsymbol{w} \\coloneqq (\\varphi, \\boldsymbol{\\theta})" + }, + { + "bbox": [ + 296, + 280, + 530, + 385 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 297, + 392, + 370, + 404 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 392, + 370, + 404 + ], + "spans": [ + { + "bbox": [ + 297, + 392, + 370, + 404 + ], + "type": "text", + "content": "2.2. Motivation" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 296, + 410, + 531, + 527 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 410, + 531, + 527 + ], + "spans": [ + { + "bbox": [ + 296, + 410, + 531, + 527 + ], + "type": "text", + "content": "Previous studies have shown strong connection between the geometry of the loss landscape and generalization of deep learning models [1, 16, 19, 21]. In this section, we first empirically demonstrate that many deep long-tailed models exhibit sharp minima, and then show that naive application of classical flattening operations results in limited improvement. Motivated by these observations, we theoretically analyze the characteristic radius of flat minima in DLTR to show that a class-conditional flattening procedure is required." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 297, + 542, + 491, + 554 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 542, + 491, + 554 + ], + "spans": [ + { + "bbox": [ + 297, + 542, + 491, + 554 + ], + "type": "text", + "content": "2.2.1 Sharpness in Deep Long-Tailed Models" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 296, + 559, + 530, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 559, + 530, + 641 + ], + "spans": [ + { + "bbox": [ + 296, + 559, + 530, + 641 + ], + "type": "text", + "content": "To show the pervasive sharpness in current DLTR models, we select two baselines (empirical risk minimization of cross-entropy (CE), LDAM-DRW [7]) and two advanced methods (MiSLAS [62], GCL [26]) from recent literature to conduct experiments1. By perturbing the parameter " + }, + { + "bbox": [ + 296, + 559, + 530, + 641 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 296, + 559, + 530, + 641 + ], + "type": "text", + "content": " of the classifier " + }, + { + "bbox": [ + 296, + 559, + 530, + 641 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 296, + 559, + 530, + 641 + ], + "type": "text", + "content": " with increasing noise " + }, + { + "bbox": [ + 296, + 559, + 530, + 641 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 296, + 559, + 530, + 641 + ], + "type": "text", + "content": ", we observe the loss of each model on CIFAR-10-LT in Figure 1." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 296, + 641, + 530, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 641, + 530, + 712 + ], + "spans": [ + { + "bbox": [ + 296, + 641, + 530, + 712 + ], + "type": "text", + "content": "Figure 1(a) shows that, at global scale, the loss of CE is the sharpest across a wide noise range. However, due to the fact that DLTR methods usually employ distinct loss functions, it's hard to compare the sharpness of their loss landscapes in a fair and meaningful way. Moreover, large noise ratio can easily flip the sign of parameters and deteriorate" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 307, + 719, + 503, + 729 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 719, + 503, + 729 + ], + "spans": [ + { + "bbox": [ + 307, + 719, + 503, + 729 + ], + "type": "text", + "content": "1We re-implement all models via their publicly released code." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 279, + 746, + 298, + 756 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 279, + 746, + 298, + 756 + ], + "spans": [ + { + "bbox": [ + 279, + 746, + 298, + 756 + ], + "type": "text", + "content": "3500" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 595, + 842 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 54, + 103, + 270, + 212 + ], + "blocks": [ + { + "bbox": [ + 54, + 103, + 270, + 212 + ], + "lines": [ + { + "bbox": [ + 54, + 103, + 270, + 212 + ], + "spans": [ + { + "bbox": [ + 54, + 103, + 270, + 212 + ], + "type": "image", + "image_path": "2efcd9cad09ebde024e5a1e3177fc0bee2e08ee77876811779f0a30cee234875.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 45, + 221, + 280, + 295 + ], + "lines": [ + { + "bbox": [ + 45, + 221, + 280, + 295 + ], + "spans": [ + { + "bbox": [ + 45, + 221, + 280, + 295 + ], + "type": "text", + "content": "Figure 2. Performance gain of naive integration of flattening operations with DLTR algorithms. The number represents the stage where flattening applies. For example, \"SN.1.2\" means we apply spectral normalization on both stages. Only the first-stage result of LDAM is reported since it's a one-stage approach. All experiments are conducted on CIFAR-100-LT with imbalance ratio of 100." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 45, + 315, + 279, + 431 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 315, + 279, + 431 + ], + "spans": [ + { + "bbox": [ + 45, + 315, + 279, + 431 + ], + "type": "text", + "content": "the classifier decision boundary. Therefore in this paper, we only consider locally flat minima, i.e., in the perturbative " + }, + { + "bbox": [ + 45, + 315, + 279, + 431 + ], + "type": "inline_equation", + "content": "(||\\epsilon ||_2\\ll ||\\boldsymbol {w}||_2)" + }, + { + "bbox": [ + 45, + 315, + 279, + 431 + ], + "type": "text", + "content": " regime shown in Figure 1(b). From this local view, only MiSLAS is observed to have a comparable or flatter minimum than CE, suggesting that the existing DLTR methods have poor generalization and robustness against model parameter perturbation. More detailed 2D loss landscapes covering overall, head and tail classes are depicted in Appendix E, and corroborate our observation in 1D." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 45, + 447, + 255, + 458 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 447, + 255, + 458 + ], + "spans": [ + { + "bbox": [ + 45, + 447, + 255, + 458 + ], + "type": "text", + "content": "2.2.2 Naive Integration of Flattening Procedures" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 45, + 464, + 279, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 464, + 279, + 628 + ], + "spans": [ + { + "bbox": [ + 45, + 464, + 279, + 628 + ], + "type": "text", + "content": "To improve the generalization of deep long-tailed models, we experiment to navigate flatter minima by integrating the existing flattening operations (stochastic weight averaging (SWA) [16], spectral normalization (SN) [4], gradients penalization (GP) [61], and model perturbation (MP) [44]) into five representative deep long-tailed learning baselines to see if they are generally beneficial for DLTR. Illustrated in Figure 2, these flattening methods bring little or negative gains in most cases. Even though in some cases they are beneficial, the corresponding baselines (mainly LDAM-DRW) are too weak so that the final performance are still sub-optimal, which calls for more effective flattening procedures for DLTR. More implementation details are provided in the Appendix C." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 45, + 641, + 234, + 652 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 641, + 234, + 652 + ], + "spans": [ + { + "bbox": [ + 45, + 641, + 234, + 652 + ], + "type": "text", + "content": "2.2.3 Characteristic Radius of Flat Minima" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 45, + 660, + 279, + 730 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 660, + 279, + 730 + ], + "spans": [ + { + "bbox": [ + 45, + 660, + 279, + 730 + ], + "type": "text", + "content": "Let's re-consider flattening operations in DLTR from a theoretical perspective. The empirical loss " + }, + { + "bbox": [ + 45, + 660, + 279, + 730 + ], + "type": "inline_equation", + "content": "L_{S}(\\boldsymbol{w})" + }, + { + "bbox": [ + 45, + 660, + 279, + 730 + ], + "type": "text", + "content": " is typically non-convex for deep neural network models, whose landscape may exhibit multiple local or global minima with similar values of " + }, + { + "bbox": [ + 45, + 660, + 279, + 730 + ], + "type": "inline_equation", + "content": "L_{S}(\\boldsymbol{w})" + }, + { + "bbox": [ + 45, + 660, + 279, + 730 + ], + "type": "text", + "content": " while having drastically different generalization performance (i.e., significantly different values" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 296, + 106, + 531, + 269 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 106, + 531, + 269 + ], + "spans": [ + { + "bbox": [ + 296, + 106, + 531, + 269 + ], + "type": "text", + "content": "of " + }, + { + "bbox": [ + 296, + 106, + 531, + 269 + ], + "type": "inline_equation", + "content": "L_{\\mathcal{T}}(\\boldsymbol{w})" + }, + { + "bbox": [ + 296, + 106, + 531, + 269 + ], + "type": "text", + "content": ". In order to find a more effective approach for generalization in DLTR, motivated by the high correlation between sharpness of the loss landscape and model generalization [19], we follow the sharpness-aware minimization (SAM) framework [13] by optimizing an upper bound of " + }, + { + "bbox": [ + 296, + 106, + 531, + 269 + ], + "type": "inline_equation", + "content": "L_{\\mathcal{T}}(\\boldsymbol{w})" + }, + { + "bbox": [ + 296, + 106, + 531, + 269 + ], + "type": "text", + "content": " derived from the PAC-Bayesian framework [34]. Given a prior over the parameters before observing any data and a posterior over the parameters dependent on the training set and learning algorithm, the PAC-Bayesian framework bounds the generalization error of any model in terms of the KL divergence between the two probability distributions. Assuming a Gaussian prior and posterior, we arrive at the following generalization bound (formal proof in Appendix D):" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 296, + 278, + 531, + 325 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 278, + 531, + 325 + ], + "spans": [ + { + "bbox": [ + 296, + 278, + 531, + 325 + ], + "type": "text", + "content": "Theorem 1 (Perturbative PAC-Bayesian Generalization Bound). For any " + }, + { + "bbox": [ + 296, + 278, + 531, + 325 + ], + "type": "inline_equation", + "content": "\\rho >0,0 < \\delta < 1" + }, + { + "bbox": [ + 296, + 278, + 531, + 325 + ], + "type": "text", + "content": ", number of samples " + }, + { + "bbox": [ + 296, + 278, + 531, + 325 + ], + "type": "inline_equation", + "content": "n\\in \\mathbb{N}^+" + }, + { + "bbox": [ + 296, + 278, + 531, + 325 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 296, + 278, + 531, + 325 + ], + "type": "inline_equation", + "content": "k\\coloneqq \\dim (\\boldsymbol {w})" + }, + { + "bbox": [ + 296, + 278, + 531, + 325 + ], + "type": "text", + "content": ", with probability at least " + }, + { + "bbox": [ + 296, + 278, + 531, + 325 + ], + "type": "inline_equation", + "content": "1 - \\delta" + }, + { + "bbox": [ + 296, + 278, + 531, + 325 + ], + "type": "text", + "content": " over a training set " + }, + { + "bbox": [ + 296, + 278, + 531, + 325 + ], + "type": "inline_equation", + "content": "\\mathcal{S}" + }, + { + "bbox": [ + 296, + 278, + 531, + 325 + ], + "type": "text", + "content": " sampled i.i.d. from distribution " + }, + { + "bbox": [ + 296, + 278, + 531, + 325 + ], + "type": "inline_equation", + "content": "\\mathcal{T}" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 297, + 343, + 535, + 377 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 343, + 535, + 377 + ], + "spans": [ + { + "bbox": [ + 297, + 343, + 535, + 377 + ], + "type": "interline_equation", + "content": "L_{\\mathcal{T}}(\\boldsymbol {w})\\leq \\max_{||\\boldsymbol {\\epsilon}||_{2}\\leq \\sqrt{k}\\rho}L_{\\mathcal{S}}(\\boldsymbol {w} + \\boldsymbol {\\epsilon}) + \\sqrt{\\frac{\\|\\boldsymbol{w}\\|_{2}^{2}}{4\\rho^{2}} + \\log(\\frac{n}{\\delta}) + \\mathcal{O}(1)}\\\\ \\hskip 142.26378ptn - 1.", + "image_path": "504f2795f575384664b44a72a5079c3d3e9ab6633eba780714973c7d19b008e1.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 296, + 384, + 531, + 465 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 384, + 531, + 465 + ], + "spans": [ + { + "bbox": [ + 296, + 384, + 531, + 465 + ], + "type": "text", + "content": "Intuitively, optimizing the perturbative bound as a surrogate function effectively seeks out parameter solutions whose neighborhoods have uniformly low empirical training loss, thus achieving a flat minimum. Now, our key insight is that the bound is convex with respect to the radius " + }, + { + "bbox": [ + 296, + 384, + 531, + 465 + ], + "type": "inline_equation", + "content": "\\rho" + }, + { + "bbox": [ + 296, + 384, + 531, + 465 + ], + "type": "text", + "content": " of the flat minima, hence there exists an optimal " + }, + { + "bbox": [ + 296, + 384, + 531, + 465 + ], + "type": "inline_equation", + "content": "\\rho^{*}" + }, + { + "bbox": [ + 296, + 384, + 531, + 465 + ], + "type": "text", + "content": " which gives the tightest generalization bound:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 296, + 474, + 542, + 522 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 474, + 542, + 522 + ], + "spans": [ + { + "bbox": [ + 296, + 474, + 542, + 522 + ], + "type": "interline_equation", + "content": "\\rho^ {*} := \\arg \\min _ {\\rho} \\left[ \\max _ {| | \\boldsymbol {\\epsilon} | | _ {2} \\leq \\sqrt {k} \\rho} L _ {\\mathcal {S}} (\\boldsymbol {w} + \\boldsymbol {\\epsilon}) + \\sqrt {\\frac {\\frac {| | \\boldsymbol {w} | | _ {2} ^ {2}}{4 \\rho^ {2}} + \\log \\left(\\frac {n}{\\delta}\\right)}{n - 1}} \\right]. \\tag {1}", + "image_path": "5b75810689b9f9400372f0e1dd68dcbd5fe9385e5b9229aa21e4a96d216b9fc8.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 296, + 522, + 530, + 638 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 522, + 530, + 638 + ], + "spans": [ + { + "bbox": [ + 296, + 522, + 530, + 638 + ], + "type": "text", + "content": "We denote " + }, + { + "bbox": [ + 296, + 522, + 530, + 638 + ], + "type": "inline_equation", + "content": "\\rho^{*}" + }, + { + "bbox": [ + 296, + 522, + 530, + 638 + ], + "type": "text", + "content": " as the characteristic radius of the flat minima centered at " + }, + { + "bbox": [ + 296, + 522, + 530, + 638 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 296, + 522, + 530, + 638 + ], + "type": "text", + "content": ". It's worth mentioning that such optimal perturbation radius is overlooked by previous literature on sharpness-aware optimization, mainly due to arguments that the PAC-Bayesian bound is too loose to capture the real generalization error [35]. However, our Theorem 1 inspired from [8, 12] ensures a non-vacuous bound (i.e., the square root term falls below 1, see empirical evidence in Appendix B) in the over-parameterized \"deep learning\" regime, rendering the characteristic radius relevant for optimization." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 296, + 638, + 530, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 638, + 530, + 662 + ], + "spans": [ + { + "bbox": [ + 296, + 638, + 530, + 662 + ], + "type": "text", + "content": "Moreover, the perturbative bound can be approximated by a first-order Taylor expansion, yielding" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 298, + 680, + 535, + 730 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 680, + 535, + 730 + ], + "spans": [ + { + "bbox": [ + 298, + 680, + 535, + 730 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\max _ {\\| \\boldsymbol {\\epsilon} \\| _ {2} \\leq \\sqrt {k} \\rho} L _ {\\mathcal {S}} (\\boldsymbol {w} + \\boldsymbol {\\epsilon}) \\approx \\max _ {\\| \\boldsymbol {\\epsilon} \\| _ {2} \\leq \\sqrt {k} \\rho} \\left[ L _ {\\mathcal {S}} (\\boldsymbol {w}) + \\boldsymbol {\\epsilon} ^ {T} \\nabla_ {\\boldsymbol {w}} L _ {\\mathcal {S}} (\\boldsymbol {w}) \\right] \\\\ = L _ {\\mathcal {S}} (\\boldsymbol {w}) + \\sqrt {k} \\rho | | \\nabla_ {\\boldsymbol {w}} L _ {\\mathcal {S}} (\\boldsymbol {w}) | | _ {2} \\tag {2} \\\\ \\end{array}", + "image_path": "ffcf9548a88844d88037a252ad8361e2bad09f855faec4abec6ef3ef015f9511.jpg" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 279, + 746, + 298, + 756 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 279, + 746, + 298, + 756 + ], + "spans": [ + { + "bbox": [ + 279, + 746, + 298, + 756 + ], + "type": "text", + "content": "3501" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 595, + 842 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 46, + 103, + 301, + 214 + ], + "blocks": [ + { + "bbox": [ + 46, + 103, + 301, + 214 + ], + "lines": [ + { + "bbox": [ + 46, + 103, + 301, + 214 + ], + "spans": [ + { + "bbox": [ + 46, + 103, + 301, + 214 + ], + "type": "image", + "image_path": "16752340528c6eccc9ada79837d80a2dbbe5fcb032792dafd304ada06f4e15d3.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 45, + 216, + 300, + 236 + ], + "lines": [ + { + "bbox": [ + 45, + 216, + 300, + 236 + ], + "spans": [ + { + "bbox": [ + 45, + 216, + 300, + 236 + ], + "type": "text", + "content": "(a) Stage 1: Class-conditional sharpness-aware minimization. We take three classes and classifier perturbation as examples for illustration." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 314, + 118, + 532, + 215 + ], + "blocks": [ + { + "bbox": [ + 314, + 118, + 532, + 215 + ], + "lines": [ + { + "bbox": [ + 314, + 118, + 532, + 215 + ], + "spans": [ + { + "bbox": [ + 314, + 118, + 532, + 215 + ], + "type": "image", + "image_path": "11f0154026619240555f56b110bab0278f9a51eafa90e8353aa96b036bbf8ec5.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 314, + 216, + 530, + 236 + ], + "lines": [ + { + "bbox": [ + 314, + 216, + 530, + 236 + ], + "spans": [ + { + "bbox": [ + 314, + 216, + 530, + 236 + ], + "type": "text", + "content": "(b) Stage 2: Robust training of the classifier by progressively generating adversarial features." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 205, + 245, + 370, + 256 + ], + "lines": [ + { + "bbox": [ + 205, + 245, + 370, + 256 + ], + "spans": [ + { + "bbox": [ + 205, + 245, + 370, + 256 + ], + "type": "text", + "content": "Figure 3. The overall framework of CC-SAM." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 277, + 186, + 288 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 277, + 186, + 288 + ], + "spans": [ + { + "bbox": [ + 46, + 277, + 186, + 288 + ], + "type": "text", + "content": "and the optimal perturbation vector" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 70, + 304, + 279, + 356 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 304, + 279, + 356 + ], + "spans": [ + { + "bbox": [ + 70, + 304, + 279, + 356 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\hat{\\boldsymbol{\\epsilon}}^{*}(\\boldsymbol {w})\\approx \\operatorname *{arg max}_{||\\boldsymbol {\\epsilon}||_{2}\\leq \\sqrt{k}\\rho}\\left[L_{\\mathcal{S}}(\\boldsymbol {w}) + \\boldsymbol{\\epsilon}^{T}\\nabla_{\\boldsymbol{w}}L_{\\mathcal{S}}(\\boldsymbol {w})\\right] \\\\ = \\sqrt {k} \\rho \\frac {\\nabla_ {\\boldsymbol {w}} L _ {\\mathcal {S}} (\\boldsymbol {w})}{\\| \\nabla_ {\\boldsymbol {w}} L _ {\\mathcal {S}} (\\boldsymbol {w}) \\| _ {2}}. \\tag {3} \\\\ \\end{array}", + "image_path": "640e4e1d2c5b38a662df963b2b222968fd2a0fca4a7d9fc5cfab80ec81ff4293.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 361, + 279, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 361, + 279, + 384 + ], + "spans": [ + { + "bbox": [ + 46, + 361, + 279, + 384 + ], + "type": "text", + "content": "Substitute Eqn 2 into 1, and ignore " + }, + { + "bbox": [ + 46, + 361, + 279, + 384 + ], + "type": "inline_equation", + "content": "\\log \\left(\\frac{n}{\\delta}\\right)" + }, + { + "bbox": [ + 46, + 361, + 279, + 384 + ], + "type": "text", + "content": " by assuming " + }, + { + "bbox": [ + 46, + 361, + 279, + 384 + ], + "type": "inline_equation", + "content": "\\rho \\ll ||\\pmb{w}||_2" + }, + { + "bbox": [ + 46, + 361, + 279, + 384 + ], + "type": "text", + "content": " as we discussed in Sec 2.2.1, obtaining" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 75, + 399, + 279, + 428 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 399, + 279, + 428 + ], + "spans": [ + { + "bbox": [ + 75, + 399, + 279, + 428 + ], + "type": "interline_equation", + "content": "\\rho^ {*} \\approx \\left(\\frac {\\| \\boldsymbol {w} \\| _ {2}}{2 \\| \\nabla_ {\\boldsymbol {w}} L _ {\\mathcal {S}} (\\boldsymbol {w}) \\| _ {2}}\\right) ^ {\\frac {1}{2}} k ^ {- \\frac {1}{4}} (n - 1) ^ {- \\frac {1}{4}} \\tag {4}", + "image_path": "a758b9135dd3585d8b2d50f8999acaac5079d61953d890bbb93b144ba123cda9.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 433, + 253, + 444 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 433, + 253, + 444 + ], + "spans": [ + { + "bbox": [ + 46, + 433, + 253, + 444 + ], + "type": "text", + "content": "with the approximated optimal generalization bound" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 50, + 458, + 279, + 485 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 50, + 458, + 279, + 485 + ], + "spans": [ + { + "bbox": [ + 50, + 458, + 279, + 485 + ], + "type": "interline_equation", + "content": "\\widehat {L} _ {\\mathcal {T}} (\\boldsymbol {w}) \\approx \\max _ {| | \\boldsymbol {\\epsilon} | | _ {2} \\leq \\sqrt {k} \\rho^ {*}} L _ {\\mathcal {S}} (\\boldsymbol {w} + \\boldsymbol {\\epsilon}) + \\frac {1}{2 \\sqrt {n - 1}} \\cdot \\frac {| | \\boldsymbol {w} | | _ {2}}{\\rho^ {*}}. \\tag {5}", + "image_path": "d774041fe01a3557e68488e508f5279049243848ba25189f488dfc15b8f1d673.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 45, + 490, + 279, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 490, + 279, + 548 + ], + "spans": [ + { + "bbox": [ + 45, + 490, + 279, + 548 + ], + "type": "text", + "content": "The first term captures the sharpness of " + }, + { + "bbox": [ + 45, + 490, + 279, + 548 + ], + "type": "inline_equation", + "content": "L_{\\mathcal{S}}" + }, + { + "bbox": [ + 45, + 490, + 279, + 548 + ], + "type": "text", + "content": " at " + }, + { + "bbox": [ + 45, + 490, + 279, + 548 + ], + "type": "inline_equation", + "content": "\\pmb{w}" + }, + { + "bbox": [ + 45, + 490, + 279, + 548 + ], + "type": "text", + "content": ", and the second term serves as a regularizer on the magnitude of " + }, + { + "bbox": [ + 45, + 490, + 279, + 548 + ], + "type": "inline_equation", + "content": "\\pmb{w}" + }, + { + "bbox": [ + 45, + 490, + 279, + 548 + ], + "type": "text", + "content": " akin to the L2 regularization in SAM [13]. A key difference here is that our regularization term is well motivated by theoretic interpretation from the PAC-Bayesian bound." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 558, + 126, + 571 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 558, + 126, + 571 + ], + "spans": [ + { + "bbox": [ + 46, + 558, + 126, + 571 + ], + "type": "text", + "content": "3. Methodology" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 45, + 578, + 279, + 671 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 578, + 279, + 671 + ], + "spans": [ + { + "bbox": [ + 45, + 578, + 279, + 671 + ], + "type": "text", + "content": "In this section, we introduce our two-stage sharpness-aware optimization algorithm based on Decoupling [20] in detail. In the first stage of the decoupled training, both the feature extractor and classifier are trained under parameter perturbations at a class-conditioned scale. In the second stage, we generate adversarial features to further robustify the classifier while freezing the backbone. The complete pseudo-code is shown in Algorithm 1." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 46, + 676, + 279, + 699 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 676, + 279, + 699 + ], + "spans": [ + { + "bbox": [ + 46, + 676, + 279, + 699 + ], + "type": "text", + "content": "3.1. Stage 1: Class-Conditional Sharpness-Aware Minimization (CC-SAM)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 46, + 706, + 279, + 730 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 706, + 279, + 730 + ], + "spans": [ + { + "bbox": [ + 46, + 706, + 279, + 730 + ], + "type": "text", + "content": "Note that a key assumption we made in Theorem 1 when deriving the characteristic radius of flat minima is that the" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 296, + 277, + 530, + 382 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 277, + 530, + 382 + ], + "spans": [ + { + "bbox": [ + 296, + 277, + 530, + 382 + ], + "type": "text", + "content": "training data " + }, + { + "bbox": [ + 296, + 277, + 530, + 382 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 296, + 277, + 530, + 382 + ], + "type": "text", + "content": " are drawn i.i.d. from the target distribution " + }, + { + "bbox": [ + 296, + 277, + 530, + 382 + ], + "type": "inline_equation", + "content": "\\mathcal{T}" + }, + { + "bbox": [ + 296, + 277, + 530, + 382 + ], + "type": "text", + "content": ". However, as we explained in Sec 2.1, such assumption is invalid due to the severe label distribution shift between training and testing sets in DLTR. To resolve this fundamental conflict, we turn to the class conditional distribution " + }, + { + "bbox": [ + 296, + 277, + 530, + 382 + ], + "type": "inline_equation", + "content": "p_{s}(\\pmb{x}|\\pmb{y})" + }, + { + "bbox": [ + 296, + 277, + 530, + 382 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 296, + 277, + 530, + 382 + ], + "type": "inline_equation", + "content": "p_{t}(\\pmb{x}|\\pmb{y})" + }, + { + "bbox": [ + 296, + 277, + 530, + 382 + ], + "type": "text", + "content": " instead, for which the i.i.d. assumption reasonably holds. Accordingly, we decompose the total loss " + }, + { + "bbox": [ + 296, + 277, + 530, + 382 + ], + "type": "inline_equation", + "content": "L_{\\mathcal{T}}(\\pmb{w})" + }, + { + "bbox": [ + 296, + 277, + 530, + 382 + ], + "type": "text", + "content": " and derive a generalization bound for each class separately, obtaining" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 303, + 389, + 530, + 421 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 303, + 389, + 530, + 421 + ], + "spans": [ + { + "bbox": [ + 303, + 389, + 530, + 421 + ], + "type": "interline_equation", + "content": "L _ {\\mathcal {T}} (\\boldsymbol {w}) = \\sum_ {c = 1} ^ {k} L _ {\\mathcal {T}} ^ {c} (\\boldsymbol {w}) = \\sum_ {c = 1} ^ {k} \\mathbb {E} _ {\\boldsymbol {x} \\sim p _ {t} (\\boldsymbol {x} | c)} [ l (\\boldsymbol {w}, \\boldsymbol {x}, c) ], \\tag {6}", + "image_path": "022c1f0e60cbbcc425439ac91df5447de031bac19b56b45d713217de7b06a687.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 422, + 530, + 468 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 422, + 530, + 468 + ], + "spans": [ + { + "bbox": [ + 304, + 422, + 530, + 468 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\widehat {L} _ {\\mathcal {T}} ^ {c} (\\boldsymbol {w}) = \\max _ {| | \\boldsymbol {\\epsilon} | | _ {2} \\leq \\sqrt {k} \\rho_ {c} ^ {*}} L _ {\\mathcal {S}} ^ {c} (\\boldsymbol {w} + \\boldsymbol {\\epsilon}) + \\frac {1}{2 \\sqrt {n - 1}} \\cdot \\frac {| | \\boldsymbol {w} | | _ {2}}{\\rho_ {c} ^ {*}}, \\\\ \\approx \\left(2 \\| \\boldsymbol {w} \\| _ {2} \\| \\nabla_ {w} L _ {\\mathcal {S}} ^ {c} (\\boldsymbol {w}) \\| _ {2}\\right) ^ {\\frac {1}{2}} k ^ {\\frac {1}{4}} \\left(n _ {c} - 1\\right) ^ {- \\frac {1}{4}} \\tag {7} \\\\ \\end{array}", + "image_path": "da477f7917c5862b817183fda1e7d3c82426b29c65cd945736c40efa51705819.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 323, + 469, + 530, + 498 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 323, + 469, + 530, + 498 + ], + "spans": [ + { + "bbox": [ + 323, + 469, + 530, + 498 + ], + "type": "interline_equation", + "content": "\\rho_ {c} ^ {*} = \\left(\\frac {\\| \\boldsymbol {w} \\| _ {2}}{2 \\| \\nabla_ {w} L _ {\\mathcal {S}} ^ {c} (\\boldsymbol {w}) \\| _ {2}}\\right) ^ {\\frac {1}{2}} k ^ {- \\frac {1}{4}} \\left(n _ {c} - 1\\right) ^ {- \\frac {1}{4}}, \\tag {8}", + "image_path": "fb89f02bb8a8f501de59d4a7e9c1b3109178afc640237233071d88990050912a.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 500, + 530, + 526 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 500, + 530, + 526 + ], + "spans": [ + { + "bbox": [ + 307, + 500, + 530, + 526 + ], + "type": "interline_equation", + "content": "\\hat {\\boldsymbol {\\epsilon}} _ {c} ^ {*} (\\boldsymbol {w}) \\approx \\sqrt {k} p _ {c} ^ {*} \\frac {\\nabla_ {w} L _ {\\mathcal {S}} ^ {c} (\\boldsymbol {w})}{\\| \\nabla_ {w} L _ {\\mathcal {S}} ^ {c} (\\boldsymbol {w}) \\| _ {2}}. \\tag {9}", + "image_path": "9c8a482a558ef1ee47ee0d89669755f5e069de69a3f803f4aa9592fd9cfb65cf.jpg" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 296, + 548, + 530, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 548, + 530, + 688 + ], + "spans": [ + { + "bbox": [ + 296, + 548, + 530, + 688 + ], + "type": "text", + "content": "An immediate observation of Eqn 8 is that the characteristic radius " + }, + { + "bbox": [ + 296, + 548, + 530, + 688 + ], + "type": "inline_equation", + "content": "\\rho_c^*" + }, + { + "bbox": [ + 296, + 548, + 530, + 688 + ], + "type": "text", + "content": " is class-dependent, more specifically, negatively correlated with the label frequency " + }, + { + "bbox": [ + 296, + 548, + 530, + 688 + ], + "type": "inline_equation", + "content": "n_c" + }, + { + "bbox": [ + 296, + 548, + 530, + 688 + ], + "type": "text", + "content": ". One possible interpretation is that the model is more confident regarding the majority classes with higher " + }, + { + "bbox": [ + 296, + 548, + 530, + 688 + ], + "type": "inline_equation", + "content": "n_c" + }, + { + "bbox": [ + 296, + 548, + 530, + 688 + ], + "type": "text", + "content": ", due to non-asymptotic estimation error given limited samples, hence it requires a smaller perturbative region in the parameter space to ensure flat loss landscape with respect to that class. Additionally, the class-wise generalization bound " + }, + { + "bbox": [ + 296, + 548, + 530, + 688 + ], + "type": "inline_equation", + "content": "\\widehat{L}_{\\mathcal{T}}^{c}(\\boldsymbol{w})" + }, + { + "bbox": [ + 296, + 548, + 530, + 688 + ], + "type": "text", + "content": " in Eqn 7 is positively related to the class-wise gradient norm " + }, + { + "bbox": [ + 296, + 548, + 530, + 688 + ], + "type": "inline_equation", + "content": "\\|\\nabla_{\\boldsymbol{w}} L_{\\mathcal{S}}^{c}(\\boldsymbol{w})\\|_2" + }, + { + "bbox": [ + 296, + 548, + 530, + 688 + ], + "type": "text", + "content": ", which effectively enforces hard example mining." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 296, + 693, + 530, + 730 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 693, + 530, + 730 + ], + "spans": [ + { + "bbox": [ + 296, + 693, + 530, + 730 + ], + "type": "text", + "content": "It follows that optimizing the approximated generalization bound " + }, + { + "bbox": [ + 296, + 693, + 530, + 730 + ], + "type": "inline_equation", + "content": "\\widehat{L}_{\\mathcal{T}}(\\boldsymbol{w}) = \\sum_{c=1}^{k} \\widehat{L}_{\\mathcal{T}}^{c}(\\boldsymbol{w})" + }, + { + "bbox": [ + 296, + 693, + 530, + 730 + ], + "type": "text", + "content": " yields an effective algorithm in the first stage, namely Class-Conditional" + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 279, + 746, + 298, + 756 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 279, + 746, + 298, + 756 + ], + "spans": [ + { + "bbox": [ + 279, + 746, + 298, + 756 + ], + "type": "text", + "content": "3502" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 595, + 842 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 52, + 108, + 233, + 120 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 52, + 108, + 233, + 120 + ], + "spans": [ + { + "bbox": [ + 52, + 108, + 233, + 120 + ], + "type": "text", + "content": "Algorithm 1: Training Paradigm of CC-SAM" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 122, + 155, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 122, + 155, + 133 + ], + "spans": [ + { + "bbox": [ + 57, + 122, + 155, + 133 + ], + "type": "text", + "content": "Input: Training Dataset" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 86, + 134, + 210, + 146 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 86, + 134, + 210, + 146 + ], + "spans": [ + { + "bbox": [ + 86, + 134, + 210, + 146 + ], + "type": "interline_equation", + "content": "\\mathcal {S} \\sim p _ {s} (\\boldsymbol {x}, \\boldsymbol {y}) = p _ {s} (\\boldsymbol {x} | \\boldsymbol {y}) p _ {s} (\\boldsymbol {y})", + "image_path": "7a08b991f50867e86e4ed15cfbc52176c092a368a16c793dad8a6c9fc876f421.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 146, + 211, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 146, + 211, + 157 + ], + "spans": [ + { + "bbox": [ + 57, + 146, + 211, + 157 + ], + "type": "text", + "content": "Output: Model trained with CC-SAM" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 57, + 161, + 93, + 172 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 161, + 93, + 172 + ], + "spans": [ + { + "bbox": [ + 57, + 161, + 93, + 172 + ], + "type": "text", + "content": "Stage 1:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 57, + 173, + 185, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 173, + 185, + 185 + ], + "spans": [ + { + "bbox": [ + 57, + 173, + 185, + 185 + ], + "type": "text", + "content": "Initialize " + }, + { + "bbox": [ + 57, + 173, + 185, + 185 + ], + "type": "inline_equation", + "content": "\\pmb{w} = \\{\\varphi, \\theta\\}" + }, + { + "bbox": [ + 57, + 173, + 185, + 185 + ], + "type": "text", + "content": " randomly" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 57, + 185, + 153, + 196 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 185, + 153, + 196 + ], + "spans": [ + { + "bbox": [ + 57, + 185, + 153, + 196 + ], + "type": "text", + "content": "while not converged do" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 73, + 197, + 174, + 207 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 73, + 197, + 174, + 207 + ], + "spans": [ + { + "bbox": [ + 73, + 197, + 174, + 207 + ], + "type": "text", + "content": "foreach batch " + }, + { + "bbox": [ + 73, + 197, + 174, + 207 + ], + "type": "inline_equation", + "content": "\\mathcal{B}_i" + }, + { + "bbox": [ + 73, + 197, + 174, + 207 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 73, + 197, + 174, + 207 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 73, + 197, + 174, + 207 + ], + "type": "text", + "content": " do" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 88, + 207, + 228, + 218 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 88, + 207, + 228, + 218 + ], + "spans": [ + { + "bbox": [ + 88, + 207, + 228, + 218 + ], + "type": "text", + "content": "Compute empirical loss " + }, + { + "bbox": [ + 88, + 207, + 228, + 218 + ], + "type": "inline_equation", + "content": "L_{S}" + }, + { + "bbox": [ + 88, + 207, + 228, + 218 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 88, + 207, + 228, + 218 + ], + "type": "inline_equation", + "content": "\\mathcal{B}_i" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 88, + 218, + 239, + 230 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 88, + 218, + 239, + 230 + ], + "spans": [ + { + "bbox": [ + 88, + 218, + 239, + 230 + ], + "type": "text", + "content": "Estimate the class-specific gradient set" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 92, + 230, + 251, + 242 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 92, + 230, + 251, + 242 + ], + "spans": [ + { + "bbox": [ + 92, + 230, + 251, + 242 + ], + "type": "inline_equation", + "content": "G = \\{g_1,g_2,\\dots,g_k\\}" + }, + { + "bbox": [ + 92, + 230, + 251, + 242 + ], + "type": "text", + "content": " with respect to " + }, + { + "bbox": [ + 92, + 230, + 251, + 242 + ], + "type": "inline_equation", + "content": "L_{S}" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 89, + 242, + 252, + 265 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 242, + 252, + 265 + ], + "spans": [ + { + "bbox": [ + 89, + 242, + 252, + 265 + ], + "type": "text", + "content": "Perturb " + }, + { + "bbox": [ + 89, + 242, + 252, + 265 + ], + "type": "inline_equation", + "content": "\\pmb{w}" + }, + { + "bbox": [ + 89, + 242, + 252, + 265 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 89, + 242, + 252, + 265 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 89, + 242, + 252, + 265 + ], + "type": "text", + "content": " according to Eqn 8 and Eqn 9" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 89, + 265, + 220, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 265, + 220, + 277 + ], + "spans": [ + { + "bbox": [ + 89, + 265, + 220, + 277 + ], + "type": "text", + "content": "Update " + }, + { + "bbox": [ + 89, + 265, + 220, + 277 + ], + "type": "inline_equation", + "content": "\\pmb{w}" + }, + { + "bbox": [ + 89, + 265, + 220, + 277 + ], + "type": "text", + "content": " via Eqn 10 and Eqn 11" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 281, + 93, + 292 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 281, + 93, + 292 + ], + "spans": [ + { + "bbox": [ + 57, + 281, + 93, + 292 + ], + "type": "text", + "content": "Stage 2:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 293, + 96, + 303 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 293, + 96, + 303 + ], + "spans": [ + { + "bbox": [ + 57, + 293, + 96, + 303 + ], + "type": "text", + "content": "Freeze " + }, + { + "bbox": [ + 57, + 293, + 96, + 303 + ], + "type": "inline_equation", + "content": "\\varphi" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 57, + 304, + 152, + 315 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 304, + 152, + 315 + ], + "spans": [ + { + "bbox": [ + 57, + 304, + 152, + 315 + ], + "type": "text", + "content": "while not converged do" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 73, + 315, + 242, + 326 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 73, + 315, + 242, + 326 + ], + "spans": [ + { + "bbox": [ + 73, + 315, + 242, + 326 + ], + "type": "text", + "content": "Sample batches via class-balanced sampler" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 78, + 326, + 263, + 338 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 78, + 326, + 263, + 338 + ], + "spans": [ + { + "bbox": [ + 78, + 326, + 263, + 338 + ], + "type": "interline_equation", + "content": "\\mathcal {S} ^ {\\prime} = \\left\\{\\mathcal {B} _ {1} ^ {\\prime}, \\mathcal {B} _ {2} ^ {\\prime}, \\dots , \\mathcal {B} _ {m} ^ {\\prime} \\right\\} \\sim p _ {s} (\\boldsymbol {x} | \\boldsymbol {y}) \\cdot \\operatorname {U n i f o r m} (\\boldsymbol {y})", + "image_path": "2188e2b17bd4648087d76c9224aca81f454bd6507ea0241937ae43509e99b44d.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 73, + 338, + 177, + 348 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 73, + 338, + 177, + 348 + ], + "spans": [ + { + "bbox": [ + 73, + 338, + 177, + 348 + ], + "type": "text", + "content": "foreach batch " + }, + { + "bbox": [ + 73, + 338, + 177, + 348 + ], + "type": "inline_equation", + "content": "\\mathcal{B}_j^\\prime" + }, + { + "bbox": [ + 73, + 338, + 177, + 348 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 73, + 338, + 177, + 348 + ], + "type": "inline_equation", + "content": "S^{\\prime}" + }, + { + "bbox": [ + 73, + 338, + 177, + 348 + ], + "type": "text", + "content": " do" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 88, + 348, + 236, + 360 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 88, + 348, + 236, + 360 + ], + "spans": [ + { + "bbox": [ + 88, + 348, + 236, + 360 + ], + "type": "text", + "content": "Computing empirical loss " + }, + { + "bbox": [ + 88, + 348, + 236, + 360 + ], + "type": "inline_equation", + "content": "L_{\\mathcal{S}}" + }, + { + "bbox": [ + 88, + 348, + 236, + 360 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 88, + 348, + 236, + 360 + ], + "type": "inline_equation", + "content": "\\mathcal{B}_i^\\prime" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 88, + 360, + 249, + 371 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 88, + 360, + 249, + 371 + ], + "spans": [ + { + "bbox": [ + 88, + 360, + 249, + 371 + ], + "type": "text", + "content": "Obtain the adversarial feature via Eqn 12" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 88, + 372, + 251, + 394 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 88, + 372, + 251, + 394 + ], + "spans": [ + { + "bbox": [ + 88, + 372, + 251, + 394 + ], + "type": "text", + "content": "Evaluate overall loss according to Eqn 13 and Eqn 14" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 88, + 394, + 174, + 406 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 88, + 394, + 174, + 406 + ], + "spans": [ + { + "bbox": [ + 88, + 394, + 174, + 406 + ], + "type": "text", + "content": "Update " + }, + { + "bbox": [ + 88, + 394, + 174, + 406 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 88, + 394, + 174, + 406 + ], + "type": "text", + "content": " accordingly" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 45, + 442, + 221, + 454 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 442, + 221, + 454 + ], + "spans": [ + { + "bbox": [ + 45, + 442, + 221, + 454 + ], + "type": "text", + "content": "Sharpness-Aware Minimization (CC-SAM):" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 52, + 468, + 141, + 480 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 52, + 468, + 141, + 480 + ], + "spans": [ + { + "bbox": [ + 52, + 468, + 141, + 480 + ], + "type": "text", + "content": "Given learning rate " + }, + { + "bbox": [ + 52, + 468, + 141, + 480 + ], + "type": "inline_equation", + "content": "\\eta" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 104, + 482, + 279, + 497 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 104, + 482, + 279, + 497 + ], + "spans": [ + { + "bbox": [ + 104, + 482, + 279, + 497 + ], + "type": "interline_equation", + "content": "\\boldsymbol {w} \\leftarrow \\boldsymbol {w} - \\eta \\nabla_ {\\boldsymbol {w}} L _ {\\mathcal {S}} ^ {\\mathrm {C C - S A M}} (\\boldsymbol {w}) \\tag {10}", + "image_path": "6710399be27ad737340278ed67054fa1b5c681ec67b9d9f2d2055b10d2aa09ee.jpg" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 116, + 500, + 279, + 531 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 116, + 500, + 279, + 531 + ], + "spans": [ + { + "bbox": [ + 116, + 500, + 279, + 531 + ], + "type": "interline_equation", + "content": "\\approx \\boldsymbol {w} - \\eta \\sum_ {c = 1} ^ {k} \\nabla_ {\\boldsymbol {w}} \\widehat {L} _ {\\mathcal {T}} ^ {c} (\\boldsymbol {w}) | _ {\\boldsymbol {w} + \\hat {\\epsilon} _ {c} ^ {*} (\\boldsymbol {w})}, \\tag {11}", + "image_path": "33cca452e0a730163f350e157f55cdeaa81f880ec8e9aac0432c6cdf7337fc67.jpg" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 45, + 544, + 279, + 615 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 544, + 279, + 615 + ], + "spans": [ + { + "bbox": [ + 45, + 544, + 279, + 615 + ], + "type": "text", + "content": "which is computationally efficient since it only involves first-order gradients. Even when " + }, + { + "bbox": [ + 45, + 544, + 279, + 615 + ], + "type": "inline_equation", + "content": "\\boldsymbol{w}" + }, + { + "bbox": [ + 45, + 544, + 279, + 615 + ], + "type": "text", + "content": " is close to the optimal " + }, + { + "bbox": [ + 45, + 544, + 279, + 615 + ], + "type": "inline_equation", + "content": "\\boldsymbol{w}^*" + }, + { + "bbox": [ + 45, + 544, + 279, + 615 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 45, + 544, + 279, + 615 + ], + "type": "inline_equation", + "content": "\\nabla_{\\boldsymbol{w}}L_{S}(\\boldsymbol{w})|_{\\boldsymbol{w}^*} \\approx 0" + }, + { + "bbox": [ + 45, + 544, + 279, + 615 + ], + "type": "text", + "content": ", the gradient for each class " + }, + { + "bbox": [ + 45, + 544, + 279, + 615 + ], + "type": "inline_equation", + "content": "\\nabla_{\\boldsymbol{w}}L_{S}^{c}(\\boldsymbol{w})|_{\\boldsymbol{w}^*}" + }, + { + "bbox": [ + 45, + 544, + 279, + 615 + ], + "type": "text", + "content": " for estimating the optimal perturbation vector " + }, + { + "bbox": [ + 45, + 544, + 279, + 615 + ], + "type": "inline_equation", + "content": "\\hat{\\epsilon}_{c}^{*}(\\boldsymbol{w})" + }, + { + "bbox": [ + 45, + 544, + 279, + 615 + ], + "type": "text", + "content": " in Eqn 9 is most likely far from zero, circumventing the need for computing high-order terms." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 46, + 628, + 256, + 640 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 628, + 256, + 640 + ], + "spans": [ + { + "bbox": [ + 46, + 628, + 256, + 640 + ], + "type": "text", + "content": "3.2. Stage 2: Robust Training of the Classifier" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 45, + 648, + 279, + 730 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 648, + 279, + 730 + ], + "spans": [ + { + "bbox": [ + 45, + 648, + 279, + 730 + ], + "type": "text", + "content": "For the second stage, we freeze the backbone to maintain feature representation and concentrate on refining the decision boundary of the classifier. Following Decoupling, we adopt the class-balanced sampling strategy in this stage, to rectify the classifier with uniform label distribution. As depicted in the Figure 3(b), when taking original feature " + }, + { + "bbox": [ + 45, + 648, + 279, + 730 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 45, + 648, + 279, + 730 + ], + "type": "text", + "content": " as the input of classifier " + }, + { + "bbox": [ + 45, + 648, + 279, + 730 + ], + "type": "inline_equation", + "content": "h(z; \\theta)" + }, + { + "bbox": [ + 45, + 648, + 279, + 730 + ], + "type": "text", + "content": ", we can generate adversarial" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 297, + 106, + 437, + 117 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 106, + 437, + 117 + ], + "spans": [ + { + "bbox": [ + 297, + 106, + 437, + 117 + ], + "type": "text", + "content": "features in forward pass as follows:" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 351, + 127, + 530, + 153 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 351, + 127, + 530, + 153 + ], + "spans": [ + { + "bbox": [ + 351, + 127, + 530, + 153 + ], + "type": "interline_equation", + "content": "z _ {a d v} = z + \\lambda \\frac {\\nabla_ {z} L _ {\\mathcal {S}} (z ; \\theta)}{| | \\nabla_ {z} L _ {\\mathcal {S}} (z ; \\theta) | | _ {2}} \\tag {12}", + "image_path": "bb1c22180a86d26c5297f9df40976f61324b748cb6d970b81a0211bbfa1dda16.jpg" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 378, + 169, + 530, + 181 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 378, + 169, + 530, + 181 + ], + "spans": [ + { + "bbox": [ + 378, + 169, + 530, + 181 + ], + "type": "interline_equation", + "content": "L _ {\\mathcal {A}} = h \\left(\\boldsymbol {z} _ {\\mathbf {a d v}}; \\boldsymbol {\\theta}\\right) \\tag {13}", + "image_path": "8a4973362a35125a0835c4e76b93052efb517564ffabb6c66e450d0742ce68bc.jpg" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 296, + 188, + 530, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 188, + 530, + 224 + ], + "spans": [ + { + "bbox": [ + 296, + 188, + 530, + 224 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 296, + 188, + 530, + 224 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 296, + 188, + 530, + 224 + ], + "type": "text", + "content": " is the hyper-parameter to scale the adversarial gradient " + }, + { + "bbox": [ + 296, + 188, + 530, + 224 + ], + "type": "inline_equation", + "content": "\\nabla_{z}L_{\\mathcal{S}}(z;\\pmb {\\theta}) / ||\\nabla_{z}L_{\\mathcal{S}}(z;\\pmb {\\theta})||_{2}" + }, + { + "bbox": [ + 296, + 188, + 530, + 224 + ], + "type": "text", + "content": ". Moreover, we employ a progressive strategy to balance " + }, + { + "bbox": [ + 296, + 188, + 530, + 224 + ], + "type": "inline_equation", + "content": "L_{\\mathcal{S}}" + }, + { + "bbox": [ + 296, + 188, + 530, + 224 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 296, + 188, + 530, + 224 + ], + "type": "inline_equation", + "content": "L_{\\mathcal{A}}" + }, + { + "bbox": [ + 296, + 188, + 530, + 224 + ], + "type": "text", + "content": ". At epoch " + }, + { + "bbox": [ + 296, + 188, + 530, + 224 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 296, + 188, + 530, + 224 + ], + "type": "text", + "content": "," + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 362, + 232, + 530, + 255 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 362, + 232, + 530, + 255 + ], + "spans": [ + { + "bbox": [ + 362, + 232, + 530, + 255 + ], + "type": "interline_equation", + "content": "L = \\left(1 - \\frac {t}{T}\\right) L _ {\\mathcal {S}} + \\frac {t}{T} L _ {\\mathcal {A}} \\tag {14}", + "image_path": "1e8008d3caa5f05594db32ee34dd2902eac72647e7db349db798d2c686b17146.jpg" + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 296, + 262, + 530, + 286 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 262, + 530, + 286 + ], + "spans": [ + { + "bbox": [ + 296, + 262, + 530, + 286 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 296, + 262, + 530, + 286 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 296, + 262, + 530, + 286 + ], + "type": "text", + "content": " is the total number of epochs in this stage, and " + }, + { + "bbox": [ + 296, + 262, + 530, + 286 + ], + "type": "inline_equation", + "content": "L_{\\mathcal{A}}" + }, + { + "bbox": [ + 296, + 262, + 530, + 286 + ], + "type": "text", + "content": " dominates the training progressively." + } + ] + } + ], + "index": 35 + }, + { + "bbox": [ + 297, + 298, + 366, + 309 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 298, + 366, + 309 + ], + "spans": [ + { + "bbox": [ + 297, + 298, + 366, + 309 + ], + "type": "text", + "content": "4. Evaluation" + } + ] + } + ], + "index": 36 + }, + { + "bbox": [ + 296, + 318, + 530, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 318, + 530, + 411 + ], + "spans": [ + { + "bbox": [ + 296, + 318, + 530, + 411 + ], + "type": "text", + "content": "In this section, we evaluate our method on multiple mainstream DLTR datasets and compare it with popular baselines including the state-of-the-art (SOTA) methods. Moreover, we also report the result on open-set recognition [32, 47] tasks to demonstrate the excellent robustness of CC-SAM to out-of-distribution samples. In the end, an ablation study is presented to verify the effectiveness of our design choices." + } + ] + } + ], + "index": 37 + }, + { + "bbox": [ + 297, + 420, + 422, + 431 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 420, + 422, + 431 + ], + "spans": [ + { + "bbox": [ + 297, + 420, + 422, + 431 + ], + "type": "text", + "content": "4.1. Datasets and Baselines" + } + ] + } + ], + "index": 38 + }, + { + "bbox": [ + 296, + 438, + 530, + 484 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 438, + 530, + 484 + ], + "spans": [ + { + "bbox": [ + 296, + 438, + 530, + 484 + ], + "type": "text", + "content": "Following the mainstream evaluation protocol [7, 63], we conduct experiments on five major long-tailed datasets, CIFAR-10-LT, CIFAR-100-LT, Places-LT [32], ImageNet-LT [32], and iNaturalist 2018 [46]." + } + ] + } + ], + "index": 39 + }, + { + "bbox": [ + 296, + 484, + 530, + 554 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 484, + 530, + 554 + ], + "spans": [ + { + "bbox": [ + 296, + 484, + 530, + 554 + ], + "type": "text", + "content": "CIFAR-10-LT/CIFAR-100-LT: These two datasets are sampled from the original CIFAR with different imbalance ratios " + }, + { + "bbox": [ + 296, + 484, + 530, + 554 + ], + "type": "inline_equation", + "content": "\\beta = N_{max} / N_{min}" + }, + { + "bbox": [ + 296, + 484, + 530, + 554 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 296, + 484, + 530, + 554 + ], + "type": "inline_equation", + "content": "N_{max}" + }, + { + "bbox": [ + 296, + 484, + 530, + 554 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 296, + 484, + 530, + 554 + ], + "type": "inline_equation", + "content": "N_{min}" + }, + { + "bbox": [ + 296, + 484, + 530, + 554 + ], + "type": "text", + "content": " are the corresponding number of the most and least frequent classes. Following [26], we set the imbalance ratio as " + }, + { + "bbox": [ + 296, + 484, + 530, + 554 + ], + "type": "inline_equation", + "content": "\\{200, 100, 50\\}" + }, + { + "bbox": [ + 296, + 484, + 530, + 554 + ], + "type": "text", + "content": " for evaluation." + } + ] + } + ], + "index": 40 + }, + { + "bbox": [ + 296, + 554, + 530, + 612 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 554, + 530, + 612 + ], + "spans": [ + { + "bbox": [ + 296, + 554, + 530, + 612 + ], + "type": "text", + "content": "Places-LT & ImageNet-LT: Both datasets were first proposed by OLTR [32]. Places-LT contains 62.5K training images spanning 365 classes in total, with imbalance ratio 996. ImageNet-LT has 115.8K training images covering 1000 categories, with imbalance ratio being 256." + } + ] + } + ], + "index": 41 + }, + { + "bbox": [ + 296, + 612, + 530, + 659 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 612, + 530, + 659 + ], + "spans": [ + { + "bbox": [ + 296, + 612, + 530, + 659 + ], + "type": "text", + "content": "iNaturalist 2018: As a naturally long-tailed classification dataset, iNaturalist 2018 contains 437.5K trainin images from 8142 categories, and its imbalance ratio is 512. We follow the official spilt in our evaluations." + } + ] + } + ], + "index": 42 + }, + { + "bbox": [ + 296, + 660, + 530, + 730 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 660, + 530, + 730 + ], + "spans": [ + { + "bbox": [ + 296, + 660, + 530, + 730 + ], + "type": "text", + "content": "For fair comparison, we exclude ensemble or pretraining models in our experiments. The baselines and state-of-the-art methods evaluated include (1) class rebalancing: LDAM-DRW [7], LDAM-DRW + SAM [38], De-confound-TDE [45], Lifted Loss [36], Focal Loss [28], OpenMax [3], BBN [63], LADE [17], DisAlign [58]," + } + ] + } + ], + "index": 43 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 279, + 746, + 298, + 756 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 279, + 746, + 298, + 756 + ], + "spans": [ + { + "bbox": [ + 279, + 746, + 298, + 756 + ], + "type": "text", + "content": "3503" + } + ] + } + ], + "index": 44 + } + ], + "page_size": [ + 595, + 842 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 130, + 104, + 446, + 295 + ], + "blocks": [ + { + "bbox": [ + 130, + 104, + 446, + 295 + ], + "lines": [ + { + "bbox": [ + 130, + 104, + 446, + 295 + ], + "spans": [ + { + "bbox": [ + 130, + 104, + 446, + 295 + ], + "type": "table", + "html": "
CIFAR-10-LTCIFAR-100-LT
Imbalance Ratio2001005020010050
CE65.6870.7074.8134.8438.4343.90
CE + Mixup [57]65.8472.9679.4835.8440.0145.16
LDAM-DRW [7]73.5277.0381.0338.9142.0447.62
De-confound-TDE [45]-80.6083.60-44.1550.31
CE + Mixup + cRT [20]73.0679.1584.2141.7345.1250.86
BBN [63]73.4779.8281.1837.2142.5647.02
Contrastive Learning [51]-81.4085.36-46.7251.87
BGP [49]---41.2045.2050.50
MiSLAS [62]77.3182.0685.1642.3347.5052.62
VS + SAM [38]-82.40--46.60-
GCL [26]79.0382.6885.4644.8848.7153.55
CC-SAM80.9483.9286.2245.6650.8353.91
", + "image_path": "5612fbf3b20118985c67a1a642ae836d2e77e14a228a5ceb801c51710dc08f32.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 47, + 344, + 278, + 433 + ], + "blocks": [ + { + "bbox": [ + 45, + 303, + 531, + 327 + ], + "lines": [ + { + "bbox": [ + 45, + 303, + 531, + 327 + ], + "spans": [ + { + "bbox": [ + 45, + 303, + 531, + 327 + ], + "type": "text", + "content": "Table 1. Top-1 Accuracy on CIFAR-10-LT and CIFAR-100-LT. All methods take ResNet-32 as the backbone. All baseline results except BGP are directly adopted from [26]. \"--\" means the original paper didn't report the corresponding results." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 47, + 344, + 278, + 433 + ], + "lines": [ + { + "bbox": [ + 47, + 344, + 278, + 433 + ], + "spans": [ + { + "bbox": [ + 47, + 344, + 278, + 433 + ], + "type": "table", + "html": "
cRTStage 1 + dirStage 1 + magStage 2Acc
✓37.8
✓✓38.9
✓✓37.5
✓✓✓40.1
✓✓✓✓40.6
", + "image_path": "76060512f14fcbed006e9aa189bd555403581257e72d8a744a71b76af082d93e.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 45, + 442, + 279, + 496 + ], + "lines": [ + { + "bbox": [ + 45, + 442, + 279, + 496 + ], + "spans": [ + { + "bbox": [ + 45, + 442, + 279, + 496 + ], + "type": "text", + "content": "Table 2. Ablation studies on Places-LT. \"Stage " + }, + { + "bbox": [ + 45, + 442, + 279, + 496 + ], + "type": "inline_equation", + "content": "1 + \\mathrm{dir}" + }, + { + "bbox": [ + 45, + 442, + 279, + 496 + ], + "type": "text", + "content": " enforces parameter perturbation along the recommended direction " + }, + { + "bbox": [ + 45, + 442, + 279, + 496 + ], + "type": "inline_equation", + "content": "(\\nabla_w L_S^c (\\pmb {w}) / ||\\nabla_w L_S^c (\\pmb {w})||_2" + }, + { + "bbox": [ + 45, + 442, + 279, + 496 + ], + "type": "text", + "content": " in Eqn 9) with magnitude of 1, whereas \"Stage " + }, + { + "bbox": [ + 45, + 442, + 279, + 496 + ], + "type": "inline_equation", + "content": "1+" + }, + { + "bbox": [ + 45, + 442, + 279, + 496 + ], + "type": "text", + "content": " mag\" enforces perturbation with the recommended magnitude " + }, + { + "bbox": [ + 45, + 442, + 279, + 496 + ], + "type": "inline_equation", + "content": "(\\rho_c^*" + }, + { + "bbox": [ + 45, + 442, + 279, + 496 + ], + "type": "text", + "content": " in Eqn 9) in a random direction." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 45, + 518, + 278, + 576 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 518, + 278, + 576 + ], + "spans": [ + { + "bbox": [ + 45, + 518, + 278, + 576 + ], + "type": "text", + "content": "LUNA [5], BGP [49], VS+SAM [38] (2) information augmentation: RSG [50], (3) decoupled training: Decouple- " + }, + { + "bbox": [ + 45, + 518, + 278, + 576 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 45, + 518, + 278, + 576 + ], + "type": "text", + "content": " norm [20], cRT [20], MisLAS [62], GCL [26], (4) representation learning: Range Loss [59], OLTR [32], IEM [65], Contrastive Learning [51], ResLT [9]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 45, + 577, + 278, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 577, + 278, + 693 + ], + "spans": [ + { + "bbox": [ + 45, + 577, + 278, + 693 + ], + "type": "text", + "content": "Our code is implemented with Pytorch 1.4.0 and all experiments are carried out on Tesla V100 GPUs. We train each model with batch size of 64 (for CIFAR-10-LT and CIFAR-100-LT) / 128 (for Places-LT) / 256 (for ImageNet-LT) / 512 (for iNaturalist 2018), SGD optimizer with momentum of 0.9. We apply similar tricks in Balanced Softmax [40] as the mainstream methods have done. Although we implemented CC-SAM under a two-stage framework, it serves as a general technique to improve the existing DLTR methods without model perturbation (Appendix A.3)." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 45, + 694, + 278, + 730 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 694, + 278, + 730 + ], + "spans": [ + { + "bbox": [ + 45, + 694, + 278, + 730 + ], + "type": "text", + "content": "Intuitively, CC-SAM brings more computation overhead due to additional gradient descents. But we only perturbed the last several layers as an efficient version of CC-SAM in" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 296, + 346, + 530, + 369 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 346, + 530, + 369 + ], + "spans": [ + { + "bbox": [ + 296, + 346, + 530, + 369 + ], + "type": "text", + "content": "our evaluation. A simple training time comparison is presented in Appendix A.5." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 297, + 378, + 379, + 389 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 378, + 379, + 389 + ], + "spans": [ + { + "bbox": [ + 297, + 378, + 379, + 389 + ], + "type": "text", + "content": "4.2. Main Results" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 296, + 398, + 530, + 538 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 398, + 530, + 538 + ], + "spans": [ + { + "bbox": [ + 296, + 398, + 530, + 538 + ], + "type": "text", + "content": "We evaluate CC-SAM on the five mainstream public benchmarks mentioned above, and the corresponding results are reported in Table 1 and Table 3. The best method is bolded and the second best is underlined. According to the evaluations, we observe that CC-SAM unanimously attains the best ranking on CIFAR-LT datasets. On large scale datasets (Places-LT, ImageNet-LT, and iNaturalist 2018), CC-SAM shows competitive performance compared to other advanced methods, too. Specifically, it consistently brings better gains to the medium and tail classes, demonstrating the effectiveness of class-conditional flattening for deep long-tailed recognition." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 296, + 548, + 458, + 560 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 548, + 458, + 560 + ], + "spans": [ + { + "bbox": [ + 296, + 548, + 458, + 560 + ], + "type": "text", + "content": "4.3. Open Long-Tailed Recognition" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 296, + 566, + 530, + 730 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 566, + 530, + 730 + ], + "spans": [ + { + "bbox": [ + 296, + 566, + 530, + 730 + ], + "type": "text", + "content": "The ability to detect out-of-distribution samples from the open world, namely open-set recognition [47], provides a unique dimension for evaluating the robustness of deep learning models. Open-set long-tailed recognition, or OLTR [32], combines the open-set problems with deep long-tailed learning to offer even more challenging tasks for model generalization. Following the setting of OLTR [32], we enable CC-SAM to distinguish the close-set and open-set samples by applying a simple, non-learnable prototype-based metric. Results are presented in Table 4. We find that CC-SAM achieves the top F-measure performance and outperforms LUNA [5] on Places-LT, a SOTA with a sophisticated open-set detection method based on hierarchical metrics. The experiments demonstrate the remarkable gen" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 279, + 746, + 298, + 756 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 279, + 746, + 298, + 756 + ], + "spans": [ + { + "bbox": [ + 279, + 746, + 298, + 756 + ], + "type": "text", + "content": "3504" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 595, + 842 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 86, + 104, + 490, + 557 + ], + "blocks": [ + { + "bbox": [ + 86, + 104, + 490, + 557 + ], + "lines": [ + { + "bbox": [ + 86, + 104, + 490, + 557 + ], + "spans": [ + { + "bbox": [ + 86, + 104, + 490, + 557 + ], + "type": "table", + "html": "
DatasetMethodBackboneManyMediumFewOverall
ImageNet-LTCEResNeXt-5065.937.57.744.4
Decouple-τ-norm [20]ResNet-5056.644.227.446.7
Balanced Softmax [40]ResNeXt-5064.148.233.452.3
LADE [17]ResNeXt-5064.447.734.352.3
RSG [50]ResNeXt-5063.248.232.251.8
DisAlign [58]ResNet-5061.352.231.452.9
ResNeXt-5062.752.131.453.4
ResLT [9]ResNeXt-5063.053.335.552.9
BGP [49]ResNet-50---51.5
MiSLAS [62]ResNet-50---52.7
LDAM-DRW + SAM [38]ResNet-5062.052.134.853.1
GCL [26]ResNet-50---54.9
CC-SAMResNet-5061.449.537.152.4
ResNeXt-5063.153.441.155.4
Places-LTCEResNet-15245.727.38.230.2
Decouple-τ-norm [20]ResNet-15237.840.731.837.9
Balanced Softmax [40]ResNet-15242.039.330.538.6
LADE [17]ResNet-15242.839.031.238.8
RSG [50]ResNet-15241.941.432.039.3
DisAlign [58]ResNet-15240.442.430.139.3
ResLT [9]ResNet-15239.843.631.439.8
MiSLAS [62]ResNet-152---40.2
GCL [26]ResNet-152---40.6
CC-SAMResNet-15241.242.136.440.6
iNaturalist 2018CEResNet-5072.263.057.261.7
Decouple-τ-norm [20]ResNet-5065.665.365.965.6
Balanced Softmax [40]ResNet-50---70.6
LADE [17]ResNet-50---70.0
RSG [50]ResNet-50---70.3
DisAlign [58]ResNet-50---70.6
ResLT [9]ResNet-50---70.2
BGP [49]ResNet-5070.069.969.670.5
MiSLAS [62]ResNet-50---71.6
LDAM-DRW + SAM [38]ResNet-5064.170.571.270.1
GCL [26]ResNet-50---72.0
CC-SAMResNet-5065.470.972.270.9
", + "image_path": "78d973e9d5ee350abd3cc6f7f9c391092985b430903664dcc446c2817902631b.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 45, + 565, + 530, + 587 + ], + "lines": [ + { + "bbox": [ + 45, + 565, + 530, + 587 + ], + "spans": [ + { + "bbox": [ + 45, + 565, + 530, + 587 + ], + "type": "text", + "content": "Table 3. Top-1 Accuracy on Places-LT, ImageNet-LT and iNaturalist 2018. As for Places-LT, we take a pre-trained ResNet-152 as the backbone for a fair comparison." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 45, + 607, + 279, + 631 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 607, + 279, + 631 + ], + "spans": [ + { + "bbox": [ + 45, + 607, + 279, + 631 + ], + "type": "text", + "content": "eralization of CC-SAM due to its capacity to learn highly robust representations." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 45, + 640, + 137, + 654 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 640, + 137, + 654 + ], + "spans": [ + { + "bbox": [ + 45, + 640, + 137, + 654 + ], + "type": "text", + "content": "4.4. Ablation Study" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 45, + 660, + 279, + 730 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 660, + 279, + 730 + ], + "spans": [ + { + "bbox": [ + 45, + 660, + 279, + 730 + ], + "type": "text", + "content": "Since CC-SAM consists of multiple components, here we provide an ablation study to demonstrate their effectiveness. The result of experiments conducted on Places-LT are shown in Table 2. In particular, CC-SAM degrades to cRT without any of our proposed operations, which we take as a baseline. It is observed that CC-SAM benefits from the" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 296, + 607, + 530, + 632 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 607, + 530, + 632 + ], + "spans": [ + { + "bbox": [ + 296, + 607, + 530, + 632 + ], + "type": "text", + "content": "design choices of each individual stage, and the integration of both stages shows the best performance." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 296, + 636, + 531, + 730 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 636, + 531, + 730 + ], + "spans": [ + { + "bbox": [ + 296, + 636, + 531, + 730 + ], + "type": "text", + "content": "We also study the class-conditional perturbation by differentiating the impacts of choosing the right direction " + }, + { + "bbox": [ + 296, + 636, + 531, + 730 + ], + "type": "inline_equation", + "content": "\\nabla_{\\boldsymbol{w}}L_{S}^{c}(\\boldsymbol{w}) / ||\\nabla_{\\boldsymbol{w}}L_{S}^{c}(\\boldsymbol{w})||_{2}" + }, + { + "bbox": [ + 296, + 636, + 531, + 730 + ], + "type": "text", + "content": " and the right magnitude " + }, + { + "bbox": [ + 296, + 636, + 531, + 730 + ], + "type": "inline_equation", + "content": "\\rho_c^*" + }, + { + "bbox": [ + 296, + 636, + 531, + 730 + ], + "type": "text", + "content": " in Eqn 9, shown as \"stage " + }, + { + "bbox": [ + 296, + 636, + 531, + 730 + ], + "type": "inline_equation", + "content": "1 + \\mathrm{dir}" + }, + { + "bbox": [ + 296, + 636, + 531, + 730 + ], + "type": "text", + "content": "\" and \"stage " + }, + { + "bbox": [ + 296, + 636, + 531, + 730 + ], + "type": "inline_equation", + "content": "1 + \\mathrm{mag}" + }, + { + "bbox": [ + 296, + 636, + 531, + 730 + ], + "type": "text", + "content": "\" in Table 2 respectively. For the former, we choose a perturbation scale of 1. For the latter, we perturb the model parameters by a random direction sampled from a zero-mean Gaussian. It turns out the perturbation direction contributes more, and" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 279, + 746, + 298, + 756 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 279, + 746, + 298, + 756 + ], + "spans": [ + { + "bbox": [ + 279, + 746, + 298, + 756 + ], + "type": "text", + "content": "3505" + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 595, + 842 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 88, + 104, + 490, + 255 + ], + "blocks": [ + { + "bbox": [ + 88, + 104, + 490, + 255 + ], + "lines": [ + { + "bbox": [ + 88, + 104, + 490, + 255 + ], + "spans": [ + { + "bbox": [ + 88, + 104, + 490, + 255 + ], + "type": "table", + "html": "
ImageNet-LTPlaces-LT
MethodManyMediumFewF-measureManyMediumFewF-measure
CE40.110.40.40.29545.922.40.40.366
Lifted Loss [36]34.829.317.40.37441.035.223.80.459
Focal Loss [28]35.729.315.60.37141.034.822.30.453
Range Loss [59]34.729.417.20.37341.035.323.10.457
OpenMax [3]35.830.017.60.36841.135.423.20.458
OLTR [32]41.933.917.40.47444.636.825.20.464
IEM [65]46.142.320.10.52548.842.428.90.486
LUNA [5]48.244.723.60.57948.141.629.00.491
CC-SAM61.449.537.10.55241.241.836.40.510
", + "image_path": "27bf662d3d688c8baa8e5dbca283a5679d147a398a958d40f142734ad013089d.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 45, + 263, + 531, + 286 + ], + "lines": [ + { + "bbox": [ + 45, + 263, + 531, + 286 + ], + "spans": [ + { + "bbox": [ + 45, + 263, + 531, + 286 + ], + "type": "text", + "content": "Table 4. Open long-tail performance of top-1 accuracy on ImageNet-LT and Places-LT. F-measure is a balanced treatment of precision and recall. The backbone of CC-SAM is ResNet-50 for ImageNet-LT, while other compared methods are equipped with ResNet-10." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 45, + 305, + 280, + 330 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 305, + 280, + 330 + ], + "spans": [ + { + "bbox": [ + 45, + 305, + 280, + 330 + ], + "type": "text", + "content": "applying both attains the best improvement. For more ablation studies, please refer to Appendix A.1." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 45, + 340, + 207, + 352 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 340, + 207, + 352 + ], + "spans": [ + { + "bbox": [ + 45, + 340, + 207, + 352 + ], + "type": "text", + "content": "5. Related Work and Discussion" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 45, + 359, + 207, + 372 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 359, + 207, + 372 + ], + "spans": [ + { + "bbox": [ + 45, + 359, + 207, + 372 + ], + "type": "text", + "content": "5.1. Deep Long-Tailed Recognition" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 45, + 378, + 279, + 551 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 378, + 279, + 551 + ], + "spans": [ + { + "bbox": [ + 45, + 378, + 279, + 551 + ], + "type": "text", + "content": "A latest survey [60] systematically studies up-to-date DLTR algorithms published at the top-tier conferences and introduces a new taxonomy, dividing DLTR methodologies into three categories: class re-balancing, information augmentation, and module improvement. For class rebalancing, a branch of sampling strategies [20, 40, 52, 54] have been proposed to re-balance the training distribution. Meanwhile, some studies design dedicated loss functions [7, 10, 28, 45] to dynamically re-weight the gradients for optimization. As one representative of information augmentation, RSG [50] enriches the feature space of tail classes via dynamical prototypes and a maximized vector loss. Decoupled training [20, 58, 62] propose to learn effective representation and robust classifier according to different sampling strategies in two stages." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 45, + 552, + 279, + 609 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 552, + 279, + 609 + ], + "spans": [ + { + "bbox": [ + 45, + 552, + 279, + 609 + ], + "type": "text", + "content": "Our method follows the decoupled training paradigm. Moreover, to combat label distribution shift, our method seeks out flat minima by perturbing the model parameters at a class-conditional scale, which is unseen in existing DLTR literature." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 45, + 618, + 237, + 631 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 618, + 237, + 631 + ], + "spans": [ + { + "bbox": [ + 45, + 618, + 237, + 631 + ], + "type": "text", + "content": "5.2. Flattening for Deep Learning Models" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 45, + 636, + 279, + 730 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 636, + 279, + 730 + ], + "spans": [ + { + "bbox": [ + 45, + 636, + 279, + 730 + ], + "type": "text", + "content": "Sharpness as a generalization measure of deep learning models has been extensively explored in previous literature [1, 13, 16, 19]. Keskar et al. [21] empirically found that a large batch SGD optimizer might lead to sharp minima and poor generalization. Inspired by this, He et al. [16] implicitly averaged along the SGD trajectory to find the asymmetric valleys, namely asymmetric flat minima. Due to the power of flattening, it has also been applied to many" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 296, + 305, + 531, + 377 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 305, + 531, + 377 + ], + "spans": [ + { + "bbox": [ + 296, + 305, + 531, + 377 + ], + "type": "text", + "content": "learning tasks to achieve better generalization. Rangwani et al. [37] introduced a domain adversarial training framework by adding an auxiliary loss to smooth the local minima for domain adaptation. Shi et al. [44] demonstrated that a flattened model was resilient to catastrophic forgetting in class-incremental few-shot learning." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 296, + 389, + 531, + 622 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 389, + 531, + 622 + ], + "spans": [ + { + "bbox": [ + 296, + 389, + 531, + 622 + ], + "type": "text", + "content": "In this work, we investigate flattening in the context of DLTR. To our best knowledge, there are only a few works which succeed in applying flattening to improve generalization in DLTR/OLTR. GBP [49] is built upon the observation of high correlation between logits and gradient norm, and proposes a gradient penalty loss as regularization. However, it lacks rigorous theoretical interpretation by employing a very loose (vacuous) PAC-Bayesian bound and meanwhile under-performs compared to our approach (Table 1, 3). Liu et al. [30] develops a variant of SAM that re-weights the perturbation at instance level to combat data imbalance, without specifying an optimal perturbation or giving any theoretical justification; thus, we regard it as a simplistic and incomplete version of our method. Besides, Rangwani et al. [38] observes that re-weighting techniques in DLTR lead to saddle points for tail classes, and directly leverages SAM to alleviate this problem. In contrast, CC-SAM further advances [38] by providing a theoretically motivated class-conditional SAM design for better generalization and robustness to label distribution shift." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 296, + 636, + 531, + 730 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 636, + 531, + 730 + ], + "spans": [ + { + "bbox": [ + 296, + 636, + 531, + 730 + ], + "type": "text", + "content": "Acknowledgment We thank anonymous reviewers for their valuable comments. This work was supported by the NSFC under Grants 61932017 and 61971390, and the following grants from the Research Grants Council of the Hong Kong Special Administrative Region, China (Project Reference Number: 14201620 and 14201321). Lanqing Li would like to thank Guangyong Chen and Danruo Deng for their helpful comments on the PAC-Bayesian framework." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 279, + 746, + 299, + 756 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 279, + 746, + 299, + 756 + ], + "spans": [ + { + "bbox": [ + 279, + 746, + 299, + 756 + ], + "type": "text", + "content": "3506" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 595, + 842 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 105, + 102, + 117 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 105, + 102, + 117 + ], + "spans": [ + { + "bbox": [ + 47, + 105, + 102, + 117 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 125, + 279, + 729 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 51, + 125, + 279, + 168 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 51, + 125, + 279, + 168 + ], + "spans": [ + { + "bbox": [ + 51, + 125, + 279, + 168 + ], + "type": "text", + "content": "[1] Carlo Baldassi, Fabrizio Pittorino, and Riccardo Zecchina. Shaping the learning landscape in neural networks around wide flat minima. Proceedings of the National Academy of Sciences, 117(1):161-170, 2020. 2, 8" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 51, + 170, + 279, + 202 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 51, + 170, + 279, + 202 + ], + "spans": [ + { + "bbox": [ + 51, + 170, + 279, + 202 + ], + "type": "text", + "content": "[2] Albert-László Barabási and Réka Albert. Emergence of scaling in random networks. science, 286(5439):509-512, 1999. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 51, + 205, + 279, + 246 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 51, + 205, + 279, + 246 + ], + "spans": [ + { + "bbox": [ + 51, + 205, + 279, + 246 + ], + "type": "text", + "content": "[3] Abhijit Bendale and Terrance E Boult. Towards open set deep networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1563-1572, 2016. 5, 8" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 51, + 250, + 279, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 51, + 250, + 279, + 282 + ], + "spans": [ + { + "bbox": [ + 51, + 250, + 279, + 282 + ], + "type": "text", + "content": "[4] Andrew Brock, Jeff Donahue, and Karen Simonyan. Large scale gan training for high fidelity natural image synthesis. arXiv preprint arXiv:1809.11096, 2018. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 51, + 284, + 279, + 337 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 51, + 284, + 279, + 337 + ], + "spans": [ + { + "bbox": [ + 51, + 284, + 279, + 337 + ], + "type": "text", + "content": "[5] Jiarui Cai, Yizhou Wang, Hung-Min Hsu, Jenq-Neng Hwang, Kelsey Magrane, and Craig S Rose. Luna: Localizing unfamiliarity near acquaintance for open-set long-tailed recognition. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 36, pages 131-139, 2022. 6, 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 51, + 340, + 279, + 383 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 51, + 340, + 279, + 383 + ], + "spans": [ + { + "bbox": [ + 51, + 340, + 279, + 383 + ], + "type": "text", + "content": "[6] Jiarui Cai, Yizhou Wang, and Jenq-Neng Hwang. Ace: Ally complementary experts for solving long-tailed recognition in one-shot. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 112-121, 2021. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 51, + 385, + 279, + 427 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 51, + 385, + 279, + 427 + ], + "spans": [ + { + "bbox": [ + 51, + 385, + 279, + 427 + ], + "type": "text", + "content": "[7] Kaidi Cao, Colin Wei, Adrien Gaidon, Nikos Arechiga, and Tengyu Ma. Learning imbalanced datasets with label-distribution-aware margin loss. Advances in neural information processing systems, 32, 2019. 1, 2, 5, 6, 8" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 51, + 430, + 279, + 472 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 51, + 430, + 279, + 472 + ], + "spans": [ + { + "bbox": [ + 51, + 430, + 279, + 472 + ], + "type": "text", + "content": "[8] Niladri S. Chatterji, Behnam Neyshabur, and Hanie Sedghi. The intriguing role of module criticality in the generalization of deep networks. In 8th International Conference on Learning Representations, 2020. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 51, + 475, + 279, + 516 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 51, + 475, + 279, + 516 + ], + "spans": [ + { + "bbox": [ + 51, + 475, + 279, + 516 + ], + "type": "text", + "content": "[9] Jiequan Cui, Shu Liu, Zhuotao Tian, Zhisheng Zhong, and Jiaya Jia. Result: Residual learning for long-tailed recognition. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2022. 1, 6, 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 520, + 279, + 572 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 520, + 279, + 572 + ], + "spans": [ + { + "bbox": [ + 47, + 520, + 279, + 572 + ], + "type": "text", + "content": "[10] Yin Cui, Menglin Jia, Tsung-Yi Lin, Yang Song, and Serge Belongie. Class-balanced loss based on effective number of samples. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9268-9277, 2019. 1, 8" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 575, + 279, + 628 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 575, + 279, + 628 + ], + "spans": [ + { + "bbox": [ + 47, + 575, + 279, + 628 + ], + "type": "text", + "content": "[11] Danruo Deng, Guangyong Chen, Jianye Hao, Qiong Wang, and Pheng-Ann Heng. Flattening sharpness for dynamic gradient projection memory benefits continual learning. Advances in Neural Information Processing Systems, 34:18710-18721, 2021. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 47, + 631, + 279, + 684 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 631, + 279, + 684 + ], + "spans": [ + { + "bbox": [ + 47, + 631, + 279, + 684 + ], + "type": "text", + "content": "[12] Gintare Karolina Dziugaite and Daniel M. Roy. Computing nonvacuous generalization bounds for deep (stochastic) neural networks with many more parameters than training data. In Proceedings of the Thirty-Third Conference on Uncertainty in Artificial Intelligence, 2017. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 47, + 687, + 279, + 729 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 687, + 279, + 729 + ], + "spans": [ + { + "bbox": [ + 47, + 687, + 279, + 729 + ], + "type": "text", + "content": "[13] Pierre Foret, Ariel Kleiner, Hossein Mobahi, and Behnam Neyshabur. Sharpness-aware minimization for efficiently improving generalization. In 9th International Conference on Learning Representations, 2021. 1, 2, 3, 4, 8" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 299, + 107, + 530, + 729 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 299, + 107, + 530, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 107, + 530, + 139 + ], + "spans": [ + { + "bbox": [ + 299, + 107, + 530, + 139 + ], + "type": "text", + "content": "[14] Ishaan Gulrajani and David Lopez-Paz. In search of lost domain generalization. In 9th International Conference on Learning Representations, 2021. 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 299, + 140, + 530, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 140, + 530, + 172 + ], + "spans": [ + { + "bbox": [ + 299, + 140, + 530, + 172 + ], + "type": "text", + "content": "[15] Haibo He and Edwardo A Garcia. Learning from imbalanced data. IEEE Transactions on knowledge and data engineering, 21(9):1263-1284, 2009. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 299, + 174, + 530, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 174, + 530, + 205 + ], + "spans": [ + { + "bbox": [ + 299, + 174, + 530, + 205 + ], + "type": "text", + "content": "[16] Haowei He, Gao Huang, and Yang Yuan. Asymmetric valleys: Beyond sharp and flat local minima. Advances in neural information processing systems, 32, 2019. 2, 3, 8" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 299, + 207, + 530, + 260 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 207, + 530, + 260 + ], + "spans": [ + { + "bbox": [ + 299, + 207, + 530, + 260 + ], + "type": "text", + "content": "[17] Youngkyu Hong, Seungju Han, Kwanghee Choi, Seokjun Seo, Beomsu Kim, and Buru Chang. Disentangling label distribution for long-tailed visual recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6626-6636, 2021. 1, 2, 5, 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 299, + 261, + 530, + 325 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 261, + 530, + 325 + ], + "spans": [ + { + "bbox": [ + 299, + 261, + 530, + 325 + ], + "type": "text", + "content": "[18] Yuanfeng Ji, Lu Zhang, Jiaxiang Wu, Bingzhe Wu, Long-Kai Huang, Tingyang Xu, Yu Rong, Lanqing Li, Jie Ren, Ding Xue, et al. Drugood: Out-of-distribution (ood) dataset curator and benchmark for ai-aided drug discovery-a focus on affinity prediction problems with noise annotations. arXiv preprint arXiv:2201.09637, 2022. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 299, + 326, + 530, + 368 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 326, + 530, + 368 + ], + "spans": [ + { + "bbox": [ + 299, + 326, + 530, + 368 + ], + "type": "text", + "content": "[19] Yiding Jiang, Behnam Neyshabur, Hossein Mobahi, Dilip Krishnan, and Samy Bengio. Fantastic generalization measures and where to find them. In 8th International Conference on Learning Representations, 2020. 2, 3, 8" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 299, + 370, + 530, + 422 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 370, + 530, + 422 + ], + "spans": [ + { + "bbox": [ + 299, + 370, + 530, + 422 + ], + "type": "text", + "content": "[20] Bingyi Kang, Saining Xie, Marcus Rohrbach, Zhicheng Yan, Albert Gordo, Jiashi Feng, and Yannis Kalantidis. Decoupling representation and classifier for long-tailed recognition. arXiv preprint arXiv:1910.09217, 2019. 1, 2, 4, 6, 7, 8" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 299, + 425, + 530, + 478 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 425, + 530, + 478 + ], + "spans": [ + { + "bbox": [ + 299, + 425, + 530, + 478 + ], + "type": "text", + "content": "[21] Nitish Shirish Keskar, Dheevatsa Mudigere, Jorge Nocedal, Mikhail Smelyanskiy, and Ping Tak Peter Tang. On large-batch training for deep learning: Generalization gap and sharp minima. In 5th International Conference on Learning Representations, 2017. 1, 2, 8" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 299, + 479, + 530, + 522 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 479, + 530, + 522 + ], + "spans": [ + { + "bbox": [ + 299, + 479, + 530, + 522 + ], + "type": "text", + "content": "[22] Jaehyung Kim, Jongheon Jeong, and Jinwoo Shin. M2m: Imbalanced classification via major-to-minor translation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 13896-13905, 2020. 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 299, + 523, + 530, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 523, + 530, + 586 + ], + "spans": [ + { + "bbox": [ + 299, + 523, + 530, + 586 + ], + "type": "text", + "content": "[23] Pang Wei Koh, Shiori Sagawa, Henrik Marklund, Sang Michael Xie, Marvin Zhang, Akshay Balsubramani, Weihua Hu, Michihiro Yasunaga, Richard Lanas Phillips, Irena Gao, et al. Wilds: A benchmark of in-the-wild distribution shifts. In International Conference on Machine Learning, pages 5637-5664. PMLR, 2021. 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 299, + 589, + 530, + 641 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 589, + 530, + 641 + ], + "spans": [ + { + "bbox": [ + 299, + 589, + 530, + 641 + ], + "type": "text", + "content": "[24] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E. Hinton. Imagenet classification with deep convolutional neural networks. In Advances in Neural Information Processing Systems 25: 26th Annual Conference on Neural Information Processing Systems, pages 1106-1114, 2012. 1" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 299, + 643, + 530, + 695 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 643, + 530, + 695 + ], + "spans": [ + { + "bbox": [ + 299, + 643, + 530, + 695 + ], + "type": "text", + "content": "[25] Lanqing Li, Liang Zeng, Ziqi Gao, Shen Yuan, Yatao Bian, Bingzhe Wu, Hengtong Zhang, Chan Lu, Yang Yu, Wei Liu, et al. Imdrug: A benchmark for deep imbalanced learning in ai-aided drug discovery. arXiv preprint arXiv:2209.07921, 2022.1" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 299, + 698, + 530, + 729 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 698, + 530, + 729 + ], + "spans": [ + { + "bbox": [ + 299, + 698, + 530, + 729 + ], + "type": "text", + "content": "[26] Mengke Li, Yiu-ming Cheung, and Yang Lu. Long-tailed visual recognition via gaussian clouded logit adjustment. In Proceedings of the IEEE/CVF Conference on Computer Vi" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 279, + 746, + 298, + 756 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 279, + 746, + 298, + 756 + ], + "spans": [ + { + "bbox": [ + 279, + 746, + 298, + 756 + ], + "type": "text", + "content": "3507" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 595, + 842 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 106, + 279, + 729 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 65, + 106, + 279, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 106, + 279, + 127 + ], + "spans": [ + { + "bbox": [ + 65, + 106, + 279, + 127 + ], + "type": "text", + "content": "sion and Pattern Recognition, pages 6929-6938, 2022. 2, 5, 6, 7" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 128, + 279, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 128, + 279, + 182 + ], + "spans": [ + { + "bbox": [ + 48, + 128, + 279, + 182 + ], + "type": "text", + "content": "[27] Shuang Li, Kaixiong Gong, Chi Harold Liu, Yulin Wang, Feng Qiao, and Xinjing Cheng. Metasaug: Meta semantic augmentation for long-tailed visual recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5212-5221, 2021. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 183, + 279, + 224 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 183, + 279, + 224 + ], + "spans": [ + { + "bbox": [ + 48, + 183, + 279, + 224 + ], + "type": "text", + "content": "[28] Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dollar. Focal loss for dense object detection. In Proceedings of the IEEE international conference on computer vision, pages 2980-2988, 2017. 1, 5, 8" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 225, + 279, + 278 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 225, + 279, + 278 + ], + "spans": [ + { + "bbox": [ + 48, + 225, + 279, + 278 + ], + "type": "text", + "content": "[29] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In European conference on computer vision, pages 740-755. Springer, 2014. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 279, + 279, + 320 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 279, + 279, + 320 + ], + "spans": [ + { + "bbox": [ + 48, + 279, + 279, + 320 + ], + "type": "text", + "content": "[30] Hong Liu, Jeff Z HaoChen, Adrien Gaidon, and Tengyu Ma. Self-supervised learning is more robust to dataset imbalance. In International Conference on Learning Representations, 2021. 1, 8" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 322, + 279, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 322, + 279, + 374 + ], + "spans": [ + { + "bbox": [ + 48, + 322, + 279, + 374 + ], + "type": "text", + "content": "[31] Jialun Liu, Yifan Sun, Chuchu Han, Zhaopeng Dou, and Wenhui Li. Deep representation learning on long-tailed data: A learnable embedding augmentation perspective. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2970-2979, 2020. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 376, + 279, + 428 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 376, + 279, + 428 + ], + "spans": [ + { + "bbox": [ + 48, + 376, + 279, + 428 + ], + "type": "text", + "content": "[32] Ziwei Liu, Zhongqi Miao, Xiaohang Zhan, Jiayun Wang, Boqing Gong, and Stella X Yu. Large-scale long-tailed recognition in an open world. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2537-2546, 2019. 1, 2, 5, 6, 8" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 430, + 279, + 471 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 430, + 279, + 471 + ], + "spans": [ + { + "bbox": [ + 48, + 430, + 279, + 471 + ], + "type": "text", + "content": "[33] Jonathan Long, Evan Shelhamer, and Trevor Darrell. Fully convolutional networks for semantic segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3431-3440, 2015. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 473, + 279, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 473, + 279, + 503 + ], + "spans": [ + { + "bbox": [ + 48, + 473, + 279, + 503 + ], + "type": "text", + "content": "[34] David A McAllester. Pac-bayesian model averaging. In Proceedings of the twelfth annual conference on Computational learning theory, pages 164-170, 1999. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 505, + 279, + 546 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 505, + 279, + 546 + ], + "spans": [ + { + "bbox": [ + 48, + 505, + 279, + 546 + ], + "type": "text", + "content": "[35] Behnam Neyshabur, Srinadh Bhojanapalli, David McAllester, and Nati Srebro. Exploring generalization in deep learning. Advances in neural information processing systems, 30, 2017. 1, 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 548, + 279, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 548, + 279, + 599 + ], + "spans": [ + { + "bbox": [ + 48, + 548, + 279, + 599 + ], + "type": "text", + "content": "[36] Hyun Oh Song, Yu Xiang, Stefanie Jegelka, and Silvio Savarese. Deep metric learning via lifted structured feature embedding. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4004-4012, 2016. 5, 8" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 602, + 279, + 653 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 602, + 279, + 653 + ], + "spans": [ + { + "bbox": [ + 48, + 602, + 279, + 653 + ], + "type": "text", + "content": "[37] Harsh Rangwani, Sumukh K Aithal, Mayank Mishra, Arihant Jain, and Venkatesh Babu Radhakrishnan. A closer look at smoothness in domain adversarial training. In International Conference on Machine Learning, pages 18378-18399. PMLR, 2022. 8" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 655, + 279, + 706 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 655, + 279, + 706 + ], + "spans": [ + { + "bbox": [ + 48, + 655, + 279, + 706 + ], + "type": "text", + "content": "[38] Harsh Rangwani, Sumukh K Aithal, Mayank Mishra, and Venkatesh Babu Radhakrishnan. Escaping saddle points for effective generalization on class-imbalanced data. In Advances in Neural Information Processing Systems. 5, 6, 7, 8" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 708, + 279, + 729 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 708, + 279, + 729 + ], + "spans": [ + { + "bbox": [ + 48, + 708, + 279, + 729 + ], + "type": "text", + "content": "[39] William J Reed. The pareto, zipf and other power laws. Economics letters, 74(1):15-19, 2001. 1" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 299, + 107, + 531, + 729 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 299, + 107, + 530, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 107, + 530, + 149 + ], + "spans": [ + { + "bbox": [ + 299, + 107, + 530, + 149 + ], + "type": "text", + "content": "[40] Jiawei Ren, Cunjun Yu, Xiao Ma, Haiyu Zhao, Shuai Yi, et al. Balanced meta-softmax for long-tailed visual recognition. Advances in neural information processing systems, 33:4175-4186, 2020. 1, 6, 7, 8" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 299, + 150, + 530, + 191 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 150, + 530, + 191 + ], + "spans": [ + { + "bbox": [ + 299, + 150, + 530, + 191 + ], + "type": "text", + "content": "[41] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. Advances in neural information processing systems, 28, 2015. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 299, + 193, + 531, + 245 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 193, + 531, + 245 + ], + "spans": [ + { + "bbox": [ + 299, + 193, + 531, + 245 + ], + "type": "text", + "content": "[42] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, et al. Imagenet large scale visual recognition challenge. International journal of computer vision, 115(3):211-252, 2015. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 299, + 247, + 530, + 298 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 247, + 530, + 298 + ], + "spans": [ + { + "bbox": [ + 299, + 247, + 530, + 298 + ], + "type": "text", + "content": "[43] Ikuro Sato, Yamada Ryota, Masayuki Tanaka, Nakamasa Inoue, and Rei Kawakami. Pof: Post-training of feature extractor for improving generalization. In International Conference on Machine Learning, pages 19221-19230. PMLR, 2022. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 299, + 300, + 530, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 300, + 530, + 352 + ], + "spans": [ + { + "bbox": [ + 299, + 300, + 530, + 352 + ], + "type": "text", + "content": "[44] Guangyuan Shi, Jiaxin Chen, Wenlong Zhang, Li-Ming Zhan, and Xiao-Ming Wu. Overcoming catastrophic forgetting in incremental few-shot learning by finding flat minima. Advances in Neural Information Processing Systems, 34:6747-6761, 2021. 1, 3, 8" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 299, + 354, + 531, + 406 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 354, + 531, + 406 + ], + "spans": [ + { + "bbox": [ + 299, + 354, + 531, + 406 + ], + "type": "text", + "content": "[45] Jingru Tan, Changbao Wang, Buyu Li, Quanquan Li, Wanli Ouyang, Changqing Yin, and Junjie Yan. Equalization loss for long-tailed object recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11662-11671, 2020. 1, 5, 6, 8" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 299, + 408, + 530, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 408, + 530, + 470 + ], + "spans": [ + { + "bbox": [ + 299, + 408, + 530, + 470 + ], + "type": "text", + "content": "[46] Grant Van Horn, Oisin Mac Aodha, Yang Song, Yin Cui, Chen Sun, Alex Shepard, Hartwig Adam, Pietro Perona, and Serge Belongie. The inaturalist species classification and detection dataset. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 8769-8778, 2018. 5" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 299, + 472, + 530, + 504 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 472, + 530, + 504 + ], + "spans": [ + { + "bbox": [ + 299, + 472, + 530, + 504 + ], + "type": "text", + "content": "[47] Sagar Vaze, Kai Han, Andrea Vedaldi, and Andrew Zisserman. Open-set recognition: A good closed-set classifier is all you need. arXiv preprint arXiv:2110.06207, 2021. 5, 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 299, + 505, + 530, + 546 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 505, + 530, + 546 + ], + "spans": [ + { + "bbox": [ + 299, + 505, + 530, + 546 + ], + "type": "text", + "content": "[48] Athanasios Voulodimos, Nikolaos Doulamis, Anastasios Doulamis, and Eftychios Protopapadakis. Deep learning for computer vision: A brief review. Computational intelligence and neuroscience, 2018. 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 299, + 548, + 530, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 548, + 530, + 599 + ], + "spans": [ + { + "bbox": [ + 299, + 548, + 530, + 599 + ], + "type": "text", + "content": "[49] Dong Wang, Yicheng Liu, Liangji Fang, Fanhua Shang, Yuanyuan Liu, and Hongying Liu. Balanced gradient penalty improves deep long-tailed learning. In Proceedings of the 30th ACM International Conference on Multimedia, pages 5093-5101, 2022. 2, 6, 7, 8" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 299, + 601, + 531, + 653 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 601, + 531, + 653 + ], + "spans": [ + { + "bbox": [ + 299, + 601, + 531, + 653 + ], + "type": "text", + "content": "[50] Jianfeng Wang, Thomas Lukasiewicz, Xiaolin Hu, Jianfei Cai, and Zhenghua Xu. Rsg: A simple but effective module for learning imbalanced datasets. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3784-3793, 2021. 1, 6, 7, 8" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 299, + 655, + 530, + 708 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 655, + 530, + 708 + ], + "spans": [ + { + "bbox": [ + 299, + 655, + 530, + 708 + ], + "type": "text", + "content": "[51] Peng Wang, Kai Han, Xiu-Shen Wei, Lei Zhang, and Lei Wang. Contrastive learning based hybrid networks for long-tailed image classification. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 943-952, 2021. 1, 6" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 299, + 708, + 530, + 729 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 708, + 530, + 729 + ], + "spans": [ + { + "bbox": [ + 299, + 708, + 530, + 729 + ], + "type": "text", + "content": "[52] Tao Wang, Yu Li, Bingyi Kang, Junnan Li, Junhao Liew, Sheng Tang, Steven Hoi, and Jiashi Feng. The devil is" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 279, + 746, + 298, + 756 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 279, + 746, + 298, + 756 + ], + "spans": [ + { + "bbox": [ + 279, + 746, + 298, + 756 + ], + "type": "text", + "content": "3508" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 595, + 842 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 106, + 279, + 723 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 64, + 106, + 279, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 106, + 279, + 139 + ], + "spans": [ + { + "bbox": [ + 64, + 106, + 279, + 139 + ], + "type": "text", + "content": "in classification: A simple framework for long-tail instance segmentation. In European conference on computer vision, pages 728-744. Springer, 2020. 1, 8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 140, + 279, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 140, + 279, + 182 + ], + "spans": [ + { + "bbox": [ + 47, + 140, + 279, + 182 + ], + "type": "text", + "content": "[53] Xudong Wang, Long Lian, Zhongqi Miao, Ziwei Liu, and Stella Yu. Long-tailed recognition by routing diverse distribution-aware experts. In International Conference on Learning Representations, 2020. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 183, + 279, + 225 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 183, + 279, + 225 + ], + "spans": [ + { + "bbox": [ + 48, + 183, + 279, + 225 + ], + "type": "text", + "content": "[54] Yiru Wang, Weihao Gan, Jie Yang, Wei Wu, and Junjie Yan. Dynamic curriculum learning for imbalanced data classification. In Proceedings of the IEEE/CVF international conference on computer vision, pages 5017-5026, 2019. 1, 8" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 226, + 279, + 279 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 226, + 279, + 279 + ], + "spans": [ + { + "bbox": [ + 48, + 226, + 279, + 279 + ], + "type": "text", + "content": "[55] Xi Yin, Xiang Yu, Kihyuk Sohn, Xiaoming Liu, and Manmohan Chandraker. Feature transfer learning for face recognition with under-represented data. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5704-5713, 2019. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 281, + 279, + 322 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 281, + 279, + 322 + ], + "spans": [ + { + "bbox": [ + 48, + 281, + 279, + 322 + ], + "type": "text", + "content": "[56] Liang Zeng, Lanqing Li, Ziqi Gao, Peilin Zhao, and Jian Li. Imgcl: Revisiting graph contrastive learning on imbalanced node classification. arXiv preprint arXiv:2205.11332, 2022. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 324, + 279, + 356 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 324, + 279, + 356 + ], + "spans": [ + { + "bbox": [ + 48, + 324, + 279, + 356 + ], + "type": "text", + "content": "[57] Hongyi Zhang, Moustapha Cisse, Yann N Dauphin, and David Lopez-Paz. mixup: Beyond empirical risk minimization. arXiv preprint arXiv:1710.09412, 2017. 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 357, + 279, + 410 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 357, + 279, + 410 + ], + "spans": [ + { + "bbox": [ + 48, + 357, + 279, + 410 + ], + "type": "text", + "content": "[58] Songyang Zhang, Zeming Li, Shipeng Yan, Xuming He, and Jian Sun. Distribution alignment: A unified framework for long-tail visual recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2361-2370, 2021. 1, 5, 7, 8" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 410, + 279, + 463 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 410, + 279, + 463 + ], + "spans": [ + { + "bbox": [ + 48, + 410, + 279, + 463 + ], + "type": "text", + "content": "[59] Xiao Zhang, Zhiyuan Fang, Yandong Wen, Zhifeng Li, and Yu Qiao. Range loss for deep face recognition with long-tailed training data. In Proceedings of the IEEE International Conference on Computer Vision, pages 5409-5418, 2017. 1, 6, 8" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 464, + 279, + 496 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 464, + 279, + 496 + ], + "spans": [ + { + "bbox": [ + 48, + 464, + 279, + 496 + ], + "type": "text", + "content": "[60] Yifan Zhang, Bingyi Kang, Bryan Hooi, Shuicheng Yan, and Jiashi Feng. Deep long-tailed learning: A survey. arXiv preprint arXiv:2110.04596, 2021. 1, 8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 497, + 279, + 529 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 497, + 279, + 529 + ], + "spans": [ + { + "bbox": [ + 48, + 497, + 279, + 529 + ], + "type": "text", + "content": "[61] Yang Zhao, Hao Zhang, and Xiuyuan Hu. Penalizing gradient norm for efficiently improving generalization in deep learning. arXiv preprint arXiv:2202.03599, 2022. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 530, + 279, + 572 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 530, + 279, + 572 + ], + "spans": [ + { + "bbox": [ + 48, + 530, + 279, + 572 + ], + "type": "text", + "content": "[62] Zhisheng Zhong, Jiequan Cui, Shu Liu, and Jiaya Jia. Improving calibration for long-tailed recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16489-16498, 2021. 1, 2, 6, 7, 8" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 573, + 279, + 626 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 573, + 279, + 626 + ], + "spans": [ + { + "bbox": [ + 48, + 573, + 279, + 626 + ], + "type": "text", + "content": "[63] Boyan Zhou, Quan Cui, Xiu-Shen Wei, and Zhao-Min Chen. Bbn: Bilateral-branch network with cumulative learning for long-tailed visual recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9719–9728, 2020. 1, 5, 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 627, + 279, + 670 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 627, + 279, + 670 + ], + "spans": [ + { + "bbox": [ + 48, + 627, + 279, + 670 + ], + "type": "text", + "content": "[64] Bolei Zhou, Agata Lapedriza, Aditya Khosla, Aude Oliva, and Antonio Torralba. Places: A 10 million image database for scene recognition. IEEE transactions on pattern analysis and machine intelligence, 40(6):1452-1464, 2017. 1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 671, + 279, + 723 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 671, + 279, + 723 + ], + "spans": [ + { + "bbox": [ + 48, + 671, + 279, + 723 + ], + "type": "text", + "content": "[65] Linchao Zhu and Yi Yang. Inflated episodic memory with region self-attention for long-tailed visual recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4344–4353, 2020. 1, 6, 8" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 279, + 746, + 298, + 756 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 279, + 746, + 298, + 756 + ], + "spans": [ + { + "bbox": [ + 279, + 746, + 298, + 756 + ], + "type": "text", + "content": "3509" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 595, + 842 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Class-Incremental Exemplar Compression for Class-Incremental Learning/47061da4-a9f4-49a6-b036-8d247f22c620_content_list.json b/2023/Class-Incremental Exemplar Compression for Class-Incremental Learning/47061da4-a9f4-49a6-b036-8d247f22c620_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..40c9487f2f66329c7e9570677f0f3a5d6e7c7ec8 --- /dev/null +++ b/2023/Class-Incremental Exemplar Compression for Class-Incremental Learning/47061da4-a9f4-49a6-b036-8d247f22c620_content_list.json @@ -0,0 +1,1510 @@ +[ + { + "type": "text", + "text": "Class-Incremental Exemplar Compression for Class-Incremental Learning", + "text_level": 1, + "bbox": [ + 107, + 130, + 861, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Zilin Luo1 Yaoyao Liu2 Bernt Schiele2 Qianru Sun1", + "bbox": [ + 256, + 179, + 727, + 199 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Singapore Management University", + "bbox": [ + 349, + 215, + 635, + 234 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ Max Planck Institute for Informatics, Saarland Informatics Campus", + "bbox": [ + 222, + 234, + 761, + 252 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "zilin.luo.2021@phdcs.smu.edu.sg {yaoyao.liu, schiele}@mpi-inf.mpg.de qianrusun@smu.edu.sg", + "bbox": [ + 89, + 253, + 895, + 268 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 304, + 313, + 319 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Exemplar-based class-incremental learning (CIL) [36] finetunes the model with all samples of new classes but few-shot exemplars of old classes in each incremental phase, where the \"few-shot\" abides by the limited memory budget. In this paper, we break this \"few-shot\" limit based on a simple yet surprisingly effective idea: compressing exemplars by downsampling non-discriminative pixels and saving \"many-shot\" compressed exemplars in the memory. Without needing any manual annotation, we achieve this compression by generating 0-1 masks on discriminative pixels from class activation maps (CAM) [49]. We propose an adaptive mask generation model called class-incremental masking (CIM) to explicitly resolve two difficulties of using CAM: 1) transforming the heatmaps of CAM to 0-1 masks with an arbitrary threshold leads to a trade-off between the coverage on discriminative pixels and the quantity of exemplars, as the total memory is fixed; and 2) optimal thresholds vary for different object classes, which is particularly obvious in the dynamic environment of CIL. We optimize the CIM model alternatively with the conventional CIL model through a bilevel optimization problem [40]. We conduct extensive experiments on high-resolution CIL benchmarks including Food-101, ImageNet-100, and ImageNet-1000, and show that using the compressed exemplars by CIM can achieve a new state-of-the-art CIL accuracy, e.g., 4.8 percentage points higher than FOSTER [42] on 10-Phase ImageNet-1000. Our code is available at https://github.com/xffzl/CIM-CIL.", + "bbox": [ + 75, + 335, + 473, + 758 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 78, + 784, + 207, + 799 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Dynamic AI systems have a continual learning nature to learn new class data. They are expected to adapt to new classes while maintaining the knowledge of old classes, i.e., free from forgetting problems [31]. To evaluate this, the following protocol of class-incremental learning (CIL) was proposed by Rebuffi et al. [36]. The model training goes", + "bbox": [ + 75, + 810, + 468, + 902 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/7db32eb0839a7df78ee3689fff037f1a62dfd758d14530fd142bcae94fe50c05.jpg", + "image_caption": [ + "(a) iCaRL [baseline]" + ], + "image_footnote": [], + "bbox": [ + 501, + 325, + 890, + 393 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/abc4554419cc1d57e763cc862d6c8d5d7a4b70f5f5456fe93318f05708e3e8f9.jpg", + "image_caption": [ + "(c) MRDC [related]" + ], + "image_footnote": [], + "bbox": [ + 503, + 411, + 890, + 478 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/bf45685998f0ace0f9dd19d4778ae5140a66001ec4521222a428c128eda68dd4.jpg", + "image_caption": [ + "Figure 1. The phase-wise training data in different methods. (a) iCaRL [36] is the baseline method using full new class data and few-shot old class exemplars. (b) Mnemonics [27] distills all training samples into few-shot exemplars without increasing their quantity. (c) MRDC [43] compresses each exemplar uniformly into a low-resolution image using JPEG [41]. (d) Our approach based on the proposed class-incremental masking (CIM) down-samples only non-discriminative pixels in the image. The legend shows the symbols of special images generated by the methods." + ], + "image_footnote": [], + "bbox": [ + 535, + 494, + 857, + 527 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "through a number of phases. Each phase has new class data added and old class data discarded, and the resultant model is evaluated on the test data of all seen classes. A straightforward way to retain old class knowledge is keeping around a few old class exemplars in the memory and using them to re-train the model in subsequent phases. The number of exemplars is usually limited, e.g., $5 \\sim 20$ exemplars per class [12, 17, 25, 27, 36, 42, 44, 46, 48], as the total memory in CIL strictly budgeted, e.g., $2k$ exemplars.", + "bbox": [ + 496, + 672, + 890, + 808 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "This leads to a serious data imbalance between old and new classes, e.g., 20 per old class vs. $1.3k$ per new class (on ImageNet-1000 [9]), as illustrated in Figure 1a. The training is thus always dominated by new classes, and forgetting problems occur for old classes. Liu et al. [27] tried to mitigate this problem by parameterizing and distilling the exem", + "bbox": [ + 496, + 809, + 890, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "11371", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "plars, without increasing the number of them (Figure 1b). Wang et al. [43] traded off between the quality and quantity of exemplars by uniformly compressing exemplar images with JPEG [41] (Figure 1c). As shown in Figure 1d, our approach is also based on image compression. The idea is to downsample only non-discriminative pixels (e.g., background) and keep discriminative pixels (i.e., representative cues of foreground objects) as the original. In this way, we do not sacrifice the discriminativeness of exemplars when increasing their quantity. In particular, we aim for adaptive compression in dynamic environments of CIL, where the intuition is later phases need to be more conservative (i.e., less downsampling) as the model needs more visual cues to classify the increased number of classes.", + "bbox": [ + 75, + 90, + 472, + 301 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To achieve selective and adaptive compression, we need the location labels of discriminative pixels. Without extra labeling, we automatically generate the labels by utilizing the model's own \"attention\" on discriminative features, i.e., class activation maps (CAM) [49]. We take this method as a feasible baseline, and based on it, we propose an adaptive version called class-incremental masking (CIM). Specifically, for each input image (with its class label), we use its feature maps and classifier weights (corresponding to its class label) to compute a CAM by channel-wise multiplication, aggregation, and normalization. Then, we apply hard thresholding to generate a 0-1 mask. We notice that when generating the masks in the dynamic environments of CIL, the optimal hyperparameters (such as the value of hard threshold and the choice of activation functions) vary for different classes as well as in different incremental phases. Our adaptive version CIM tackles this by parameterizing a mask generation model and optimizing it in an end-to-end manner across all incremental phases. In each phase, the learned CIM model adaptively generates class-and phase-specific masks. We find that the compressed exemplars based on these masks have stronger representativeness, compared to using the conventional CAM.", + "bbox": [ + 75, + 301, + 472, + 648 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Technically, we have two models to optimize, i.e., the CIL model and the CIM model. These two cannot be optimized separately as they are dependent on computation: 1) the CIM model compresses exemplars to input into the CIL model; 2) the two models share network parameters. We exploit a global bilevel optimization problem (BOP) [7, 40] to alternate their training processes at two levels. This BOP goes through all incremental training phases. In particular, for each phase, we perform a local BOP with two steps to tune the parameters of the CIM model: 1) a temporary model is trained with the compressed exemplars as input; and 2) a validation loss on the uncom", + "bbox": [ + 75, + 648, + 472, + 832 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "pressed new data is computed and the gradients are backpropagated to optimize the parameters of CIM. To evaluate CIM, we conduct extensive experiments by plugging it in recent CIL methods, $^{3}$ LUCIR [17], DER [46], and FOSTER [42], on three high-resolution benchmarks, Food101 [3], ImageNet-100 [17], and ImageNet-1000 [9]. We find that using the compressed exemplars by CIM brings consistent and significant improvements, e.g., $4.2\\%$ and $4.8\\%$ higher than the SOTA method FOSTER [42], respectively, in the 5-phase and 10-phase settings of ImageNet1000, with a total memory budget for $5k$ exemplars.", + "bbox": [ + 500, + 90, + 893, + 258 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 268, + 640, + 284 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Class-Incremental Learning (CIL). There are three main lines of work to address the catastrophic forgetting problem [30, 31] in CIL. Regularization-based methods apply discrepancy (between old and new models) penalization terms in their objective functions, e.g., by comparing output logits [23, 36], intermediate features [12, 17, 24, 39], and prediction heatmaps [10]. Parameter-isolation-based methods increase the model parameters in each new incremental phase, to prevent knowledge forgetting caused by parameter overwritten. Some of them [18, 37, 42, 45, 46] proposed to progressively expand the size of the neural network to learn new coming data. Others [1, 21, 25, 47] froze a part of network parameters (to maintain the old class knowledge) to alleviate the problem of knowledge overwriting. Replay-based methods assume there is a clear memory budget allowing a handful of old-class exemplars in the memory. Exemplars can be used to re-train the model in each new phase [12, 17, 26, 27, 36, 43, 44]. This re-training usually contains two steps: one step of training the model on all new class data and old class exemplars, and one step of finetuning the model with a balanced subset (i.e., using an equal number of samples per class) [12, 17, 25, 26, 46].", + "bbox": [ + 496, + 292, + 893, + 626 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The replay-based methods focusing on memory optimization [27, 43] are closely related to our work. [27] proposed a bilevel optimization framework to distill the current new class data into exemplars before discarding them. It aims to improve the quality of exemplars without increasing the quantity. Another work [43] aimed to trade-off between the quality and quantity of exemplars by image compression using the JPEG algorithm, i.e., each exemplar is uniformly downsampled. Ours differs from these two works in three aspects. 1) Our CIM based image compression automatically segments the discriminative pixels in the exemplar and downsamples the non-discriminative pixels only. It barely weakens the representativeness of the exemplar. In contrast, the parameterization of image pixels in [27] hampers the model from capturing high-frequency (discriminative) features from the image, especially the high-resolution (e.g.,", + "bbox": [ + 496, + 627, + 895, + 869 + ], + "page_idx": 1 + }, + { + "type": "page_footnote", + "text": "3Using \"plug-in\" for evaluation is due to the fact that many baseline methods were originally evaluated in different CIL settings.", + "bbox": [ + 500, + 875, + 893, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_footnote", + "text": "1Note that we do not use mask labels to do image compression because storing them is expensive. Instead, we expand the mask to a bounding box, as elaborated in Section 4.", + "bbox": [ + 75, + 838, + 468, + 875 + ], + "page_idx": 1 + }, + { + "type": "page_footnote", + "text": "2Note that the CIM model is actually a plug-in branch in the CIL model, which is detailed in Section 4.2.", + "bbox": [ + 75, + 875, + 468, + 898 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "11372", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "$224 \\times 224)$ image [6]. 2) Our approach increases the diversity (i.e., quantity) of old class exemplars by reducing the memory consumption for each exemplar. In contrast, [27] keeps a fixed number of exemplars in the memory. 3) Our approach has an adaptive image compression strategy that fits well in the dynamic environments of CIL. In contrast, [43] uses uniform image compression (or uniformly increasing quality parameters from 100 to 1) without considering the properties of specific classes in each learning phase.", + "bbox": [ + 75, + 90, + 467, + 226 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Class Activation Map (CAM) [49] is a simple yet effective weakly-supervised object localization method. Its model is trained with only the image-level label and can generate pixel-level masks on foreground objects. Specifically, the masks are the results of hard-thresholding the heatmaps produced by feature maps and classifier weights. Advanced CAM variants include Grad-CAM [38], ReCAM [8], AdvCAM [22], etc. Our CIM is based on the vanilla CAM because it is computationally simple and efficient.", + "bbox": [ + 75, + 227, + 467, + 362 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Bilevel Optimization Problems (BOP) [7, 40] aims to solve a nested optimization problem, where the outer-level optimization is subjected to the result of the inner-level optimization. It has shown effectiveness in a wide range of machine learning areas, such as hyperparameter selection [29] and meta-learning [13]. For tackling CIL tasks, [27] leverages BOP to alternatively optimize the parameters of the CIL model and the parameterized exemplars. [25] applies BOP to learn the aggregation weights of the plastic and elastic branches in the CIL model. In our work, we use BOP to solve the optimization of the CIL model and the parameterized Class-Incremental Masking (CIM) model, where CIM is a plugin branch (in the CIL model), using few extra parameters. The process of BOP is quick and efficient.", + "bbox": [ + 75, + 363, + 467, + 574 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Preliminary", + "text_level": 1, + "bbox": [ + 76, + 588, + 202, + 604 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The following is the training pipeline of standard CIL with few-shot exemplars. Assume there are $N$ learning phases. In the 1-st phase, we load data $\\mathcal{D}_1$ containing all training samples of $c_1$ classes, and use $\\mathcal{D}_1$ to train the initial classification model $(\\theta_{1},\\omega_{1})$ , where $\\theta_{1}$ and $\\omega_{1}$ denote the parameters of the feature extractor and classifier, respectively. When the training is done, we evaluate the model performance on the test samples of $c_1$ classes. Before the 2nd phase, we discard most of the training samples due to the strict memory budget of CIL. In other words, we preserve only a handful of training samples $\\mathcal{E}_1$ (i.e., exemplars) in the memory, selected from $\\mathcal{D}_1$ . A common method for selecting exemplars is called feature herding [36] and has been used in many related works [27, 42, 43, 46]. We adopt it, too, in this work.", + "bbox": [ + 75, + 613, + 467, + 838 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In the $i$ -th phase $(i \\geq 2)$ , we load all exemplars $\\mathcal{E}_{1:i-1} = \\mathcal{E}_1 \\cup \\dots \\cup \\mathcal{E}_{i-1}$ from the memory and initialize the current model $(\\theta_i, \\omega_i)$ by the previous model $(\\theta_{i-1}, \\omega_{i-1})$ . We use $\\mathcal{E}_{1:i-1}$ and the new coming data $\\mathcal{D}_i$ (containing $c_i$", + "bbox": [ + 75, + 839, + 467, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "new classes) to train $(\\theta_{i},\\omega_{i})$ . Then, we evaluate the current model using a test set of all $\\sum_{j = 1}^{i}c_{j}$ classes seen so far. After that, we discard most of the training samples in $\\mathcal{D}_i$ and leave few-shot exemplars $\\varepsilon_{i}$ in the memory. It is clear that this discarding causes a strong data imbalance between old and new coming classes in the subsequent phase. In the following, we introduce our solution to this problem.", + "bbox": [ + 496, + 90, + 890, + 198 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4. Methodology", + "text_level": 1, + "bbox": [ + 498, + 210, + 635, + 227 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "As illustrated in Figure 1d, we alleviate the data imbalance problem by saving a larger number of compressed exemplars for old classes, where we leverage pixel-selective compression, i.e., downsample only non-discriminative pixels. To achieve this, the ideal case is that we have the pixel-level localization of foreground objects. However, the realistic case is that such localization labels are expensive, most CIL benchmarks do not have the labels, and it is not fair to compare with other CIL methods if using the labels. Without extra labeling, we introduce a CAM-based mask generation method, and based on it, we provide a baseline solution to pixel-selective compression in Section 4.1. The problem of mask generation in CIL is that the optimal generation hyperparameters such as hard thresholds are changing in the dynamic environment (with the increasing number of classes and phases). It is thus desirable to have an adaptive mask generation process. To this end, we propose class-incremental masking (CIM)—a learnable mask generation model, in Section 4.2.", + "bbox": [ + 496, + 234, + 890, + 521 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4.1. CAM-based Compression Pipeline", + "text_level": 1, + "bbox": [ + 498, + 530, + 800, + 546 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Generating pixel-level labels for large-scale datasets, e.g., ImageNet [9], is non-trivial. Using class activation maps (CAM) is a naive solution with little computation costs. Its key idea is to make use of the activation of the classification model itself: on the feature maps, activated pixels are more discriminative than non-activated ones for recognizing the object, where \"activated\" means of high activation values and with strong correlation with the classification weights of the object. After localizing activated pixels by CAM, we can generate a 0-1 mask on them, e.g., by hard thresholding their normalized values, and then up-sample the mask to the size of the input image [49].", + "bbox": [ + 496, + 553, + 890, + 733 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "From CAM to 0-1 Mask. We extract CAM in the following steps. Given an image $x$ from $\\mathcal{D}_i$ and its ground truth class label $y$ , let $F(x; \\theta_i)$ represent the feature block output by the feature extractor $\\theta_i$ , and $\\omega_{i,y}$ for the classification weights of class $y$ in the classifier $\\omega_i$ . The CAM of $x$ is:", + "bbox": [ + 496, + 734, + 890, + 809 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {M} ^ {\\mathrm {C A M}} = \\frac {A - \\operatorname* {m i n} (A)}{\\operatorname* {m a x} (A) - \\operatorname* {m i n} (A)}, A = \\omega_ {i, y} ^ {\\top} F (x; \\theta_ {i}), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 514, + 816, + 890, + 848 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\min (\\cdot)$ and $\\max (\\cdot)$ operations are used for normalization. Then, we upsample $\\mathcal{M}^{\\mathrm{CAM}}$ to the size of image $x$ and use the same notation.", + "bbox": [ + 496, + 854, + 890, + 898 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "11373", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/0749e512693154e7a16132e2e3713eb2930a05005bc40316fe6c880c3b65e52d.jpg", + "image_caption": [ + "Figure 2. The proposed compression pipeline. Given an image, we extract its CAM-based (baseline) or CIM-based (ours) mask, threshold it to be a 0-1 mask with a fixed threshold $\\tau$ , and generate a tight bounding box (bbox) that covers all masked pixels. Then, we downsample the pixels outside the bbox and sum the downsampled image up to the masked image to generate the final compressed image." + ], + "image_footnote": [], + "bbox": [ + 89, + 84, + 890, + 208 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Each value in $\\mathcal{M}^{\\mathrm{CAM}}$ denotes the activation strength of the model at a specific pixel location. Following the way of generating 0-1 masks in weakly-supervised semantic segmentation works [8, 11, 22], we apply a hard threshold $\\tau$ (between 0 and 1) over all values of $\\mathcal{M}^{\\mathrm{CAM}}$ , and get the 0-1 mask $\\mathcal{M}^{\\tau}$ : $\\mathcal{M}^{\\tau} = \\mathbb{I}(\\mathcal{M}^{\\mathrm{CAM}} > \\tau)$ , where $\\mathbb{I}(\\cdot)$ is the indicator function. In $\\mathcal{M}^{\\tau}$ , 1s indicate the locations of discriminative pixels, e.g., foreground pixels, based on which the model makes the prediction. While 0s indicate mostly background pixels or non-discrimination pixels that can be downsampled as they make little contribution to the prediction.", + "bbox": [ + 75, + 271, + 468, + 436 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "After generating 0-1 masks, it is ideal to keep them in the memory as meta information of compression. However, this is not efficient or feasible in CIL. There are two reasons. 1) The space for saving image-size masks is non-negligible. Each mask pixel is a one-bit boolean value, and one mask takes around $\\frac{1}{3 \\times 8} = \\frac{1}{24}$ of the memory of one RGB image. 2) The mask involves activated regions with irregular shapes. It is thus non-trivial to perform any standard down-sampling algorithm [34] on the remaining regions.", + "bbox": [ + 75, + 438, + 468, + 574 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "From 0-1 Masks to Bounding Boxes (BBox). A simple workaround is to generate a tight bounding box (bbox) to cover the positions of 1s in $\\mathcal{M}^{\\tau}$ , and use the_bbox for compression. Specifically, given $\\mathcal{M}^{\\tau}$ , we obtain the coordinate representation of the bounding box as:", + "bbox": [ + 75, + 574, + 468, + 650 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {B} = \\left[ \\min h, \\min w; \\max h, \\max w \\right] _ {(h, w): \\mathcal {M} ^ {\\tau} (h, w) = 1}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 88, + 659, + 468, + 676 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $h$ and $w$ denote the vertical and horizontal coordinates of the 1 on $\\mathcal{M}^{\\tau}$ , respectively. We highlight that $\\mathcal{B}$ consists of four integers only and takes negligible memory overhead compared to $\\mathcal{M}^{\\tau}$ . In addition, we \"reshape\" the irregular shape (of the activated region) in $\\mathcal{M}^{\\tau}$ into a rectangular $\\mathcal{B}$ , so our downsampling operation on the pixels outside the rectangular becomes easy.", + "bbox": [ + 75, + 684, + 468, + 789 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Compression with BBox. Given the image and its bbox on foreground, the compression is implemented by downsampling pixels outside the bbox. Specifically, as illustrated in Figure 2, we compress the image $x$ to $\\tilde{x}$ as follows,", + "bbox": [ + 75, + 789, + 468, + 849 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {x} = \\mathcal {M} ^ {\\mathcal {B}} \\odot x + (1 - \\mathcal {M} ^ {\\mathcal {B}}) \\odot x _ {\\eta}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 158, + 858, + 468, + 876 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathcal{M}^{\\mathcal{B}}$ is the binary mask according to $\\mathcal{B}$ , i.e., the val", + "bbox": [ + 76, + 883, + 468, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "ues of $\\mathcal{M}^{\\mathcal{B}}$ are 1 inside $\\mathcal{B}$ , and 0 otherwise. $x_{\\eta}$ is the fully downsampled version of $x$ with a downsampling ratio $\\eta$ ( $\\eta > 1$ ), $\\odot$ denotes the element-wise product, and $+$ denotes the element-wise addition. Both $\\odot$ and $+$ are applied independently on each RGB channel.", + "bbox": [ + 496, + 271, + 890, + 347 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The memory allocated for the compressed image $\\tilde{x}$ is as follows,", + "bbox": [ + 496, + 347, + 890, + 376 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} m _ {\\tilde {x}} = \\frac {H _ {\\mathcal {B}} W _ {\\mathcal {B}}}{H W} + \\frac {1}{\\eta} \\left(1 - \\frac {H _ {\\mathcal {B}} W _ {\\mathcal {B}}}{H W}\\right) \\tag {4} \\\\ = 1 - \\left(1 - \\frac {1}{\\eta}\\right) \\cdot \\left(1 - \\frac {H _ {\\mathcal {B}} W _ {\\mathcal {B}}}{H W}\\right), \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 566, + 386, + 890, + 450 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $H_{\\mathcal{B}}$ and $W_{\\mathcal{B}}$ are the height and width of $\\mathcal{B}$ , respectively. $H$ and $W$ are the height and width of the original image $x$ , respectively. $m_{\\tilde{x}}$ is always smaller than 1 where 1 denotes the memory unit of saving one original image $x$ . Therefore, we can save a larger number of compressed exemplars within the same memory budget. We denote the set of compressed exemplars in the $i$ -th phase as $\\tilde{\\mathcal{E}}_i$ .", + "bbox": [ + 496, + 459, + 890, + 566 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Compression Artifacts. The above compression introduces artifacts to the compressed images, i.e., there is a resolution mutation around bounding box edges. From the perspective of spectrum analysis [5,20], such mutation carries noisy and high-frequency components and impairs the model training in subsequent phases. We mitigate the effect of these artifacts by implementing the following data augmentation: in each training epoch, we transform a random subset of $\\mathcal{D}_i$ into compressed images with CAM-based bounding boxes using the same downsampling ratio. Using this augmentation enables the model to \"simulate\" the training with compressed images and learn to be invariant to compression artifacts.", + "bbox": [ + 496, + 566, + 890, + 762 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.2. Class-Incremental Masking (CIM)", + "text_level": 1, + "bbox": [ + 498, + 771, + 802, + 787 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Ideally, the mask generation process needs to adjust at different phases in CIL environments. The process involves two hyperparameters: the masking threshold and the choice of network activation functions. First, for the threshold, searching for its optimum (for all classes) is not trivial. Grid search is intuitive, but it is computationally expensive when the number of classes increases in CIL. Second, for the acti", + "bbox": [ + 496, + 795, + 890, + 898 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "11374", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/390c8a83693147661c7a7845ebf4c21b3bbe8dc5e18442187584b423f580a6f4.jpg", + "image_caption": [ + "Figure 3. Our CIM installs an \"extension\" on the backbone (e.g., ResNet-18 [16, 36] with four blocks) by adding a learnable activation function (e.g., PAU [32]) at the position of the original activation function (i.e., ReLU [33]). Such \"extension\" results in a new network branch (in green), whose weight layer parameters are directly copied from the original branch (in blue)." + ], + "image_footnote": [], + "bbox": [ + 78, + 71, + 472, + 215 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "vation function, the standard network of CIL methods uses ReLU [33, 36] and does not optimize it.", + "bbox": [ + 75, + 327, + 468, + 357 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We solve this issue by applying a learnable activation function in addition to the existing ReLU function in the CIL model. Physically, we have one neural network, while logically, we have two models to learn (in each incremental phase): the conventional CIL model with ReLU activations, and the adaptive mask generation model with learnable activations. We thus call our method class-incremental masking (CIM) based CIL. In the following, we elaborate on the network design and optimization pipeline.", + "bbox": [ + 75, + 359, + 468, + 494 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Network Design. Figure 3 demonstrates an example network architecture in our CIM-based CIL. The proposed CIM extends the network backbone by logically adding a network branch, where only the activation functions are learnable (e.g., Padé Activation Units $(\\mathrm{PAU}^4$ , [32])) and the parameters of weight layers are copied from the original branch. This design is motivated by the works of He et al. [15] and Bochkovskiy et al. [2], which indicate that layers with learnable activation functions can flexibly process object (localization) information at different network blocks. The difference is that we apply this flexibility to achieve adaptive mask generation for different CIL phases. We denote the CIM parameters (i.e., the parameters in learnable activation functions) in $i$ -th phase as $\\phi_i$ . We optimize the CIL model $(\\theta_i, \\omega_i)$ and the CIM model $\\phi_i$ via a global BOP, as elaborated below.", + "bbox": [ + 75, + 496, + 468, + 737 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Optimization Pipeline. We demonstrate the overall optimization flow in Algorithm 1, which consists of two levels of optimization: task-level and mask-level—the former one for CIL and the latter one for CIM. Note that to maintain an unified notation, we further define $\\tilde{\\mathcal{E}}_0 = \\varnothing$ .", + "bbox": [ + 75, + 739, + 468, + 815 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "1) Task-level Optimization. This level aims to optimize the CIL model $(\\theta_i, \\omega_i)$ to address the CIL task at hand. It", + "bbox": [ + 76, + 816, + 468, + 847 + ], + "page_idx": 4 + }, + { + "type": "code", + "sub_type": "algorithm", + "code_caption": [ + "Algorithm 1: CIM-based CIL (i-th phase, $i \\geq 1$ )" + ], + "code_body": "Input: New data $\\mathcal{D}_i$ ; old compressed exemplars $\\tilde{\\mathcal{E}}_{0:i - 1}(\\tilde{\\mathcal{E}}_0 = \\emptyset)$ ; last-phase CIL model $(\\theta_{i - 1},\\omega_{i - 1})$ and CIM model $\\phi_{i - 1}$ $(\\theta_0,\\omega_0$ are random parameters, $\\phi_0$ is set to ReLU). Output: New compressed exemplars $\\tilde{\\mathcal{E}}_i$ ; updated CIL model $(\\theta_i,\\omega_i)$ and CIM model $\\phi_{i}$ \n1 Initialize $(\\theta_{i},\\omega_{i})$ with $(\\theta_{i - 1},\\omega_{i - 1})$ . \n2 Initialize $\\phi_{i}$ with $\\phi_{i - 1}$ . \n3 for epochs do. \n4 Train $(\\theta_{i},\\omega_{i})$ using $\\tilde{\\mathcal{E}}_{0:i - 1}\\cup \\mathcal{D}_i$ by Eq. 5; \n5 Compress $\\mathcal{D}_i$ into $\\tilde{\\mathcal{D}}_i^\\phi$ using $\\phi_{i}$ . \n6 Temporarily update $\\theta_{i}$ to $\\theta_i^+$ using $\\tilde{\\mathcal{E}}_{0:i - 1}\\cup \\tilde{\\mathcal{D}}_i^\\phi$ by Eq.7; \n7 Learn $\\phi_{i}$ using $\\mathcal{D}_i$ by Eq.9; \n8 end \n9 Compress $\\mathcal{D}_i$ into $\\tilde{\\mathcal{D}}_i$ using the learned $\\phi_{i}$ by Eq. 3 \n0 Select exemplars $\\tilde{\\mathcal{E}}_i$ from $\\tilde{\\mathcal{D}}_i$ by, e.g., herding [36].", + "bbox": [ + 506, + 111, + 856, + 383 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "can be written as:", + "bbox": [ + 500, + 415, + 619, + 428 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\left(\\theta_ {i}, \\omega_ {i}\\right) \\leftarrow \\left(\\theta_ {i}, \\omega_ {i}\\right) - \\lambda \\nabla_ {\\left(\\theta , \\omega\\right)} \\mathcal {L} _ {\\mathrm {C I L}} \\left(\\tilde {\\mathcal {E}} _ {0: i - 1} \\cup \\mathcal {D} _ {i}; \\theta_ {i}, \\omega_ {i}\\right), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 514, + 440, + 890, + 470 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\lambda$ is the learning rate. We follow the implementation of CIL training loss $\\mathcal{L}_{\\mathrm{CIL}}$ in baseline methods [23, 42, 46, 50]. This means that we use different training losses when plugging CIM into different baseline methods.", + "bbox": [ + 496, + 470, + 890, + 531 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "2) Mask-level Optimization. This level aims to optimize the CIM model $\\phi_{i}$ to produce adaptive compression masks. It is formulated as a local BOP:", + "bbox": [ + 496, + 531, + 890, + 575 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {\\phi_ {i}} \\left[ \\mathcal {L} _ {\\text {v a l}} \\left(\\mathcal {D} _ {i}; \\theta_ {i} ^ {*}, \\omega_ {i}\\right) + \\mu \\mathcal {R} \\left(\\phi_ {i}\\right) \\right], \\tag {6a}\n$$\n", + "text_format": "latex", + "bbox": [ + 519, + 587, + 890, + 611 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\text {s . t .} \\theta_ {i} ^ {*} = \\underset {\\theta_ {i}} {\\arg \\min } \\mathcal {L} _ {\\operatorname {t r a i n}} \\left(\\tilde {\\mathcal {E}} _ {0: i - 1} \\cup \\tilde {\\mathcal {D}} _ {i} \\left(\\phi_ {i}\\right); \\theta_ {i}, \\omega_ {i}\\right). \\tag {6b}\n$$\n", + "text_format": "latex", + "bbox": [ + 521, + 614, + 890, + 641 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Eq. 6b denotes an inner-level optimization. It trains $\\theta_{i}$ with the data $\\tilde{\\mathcal{D}}_i(\\phi_i)$ compressed by using $\\phi_{i}$ , and converges as $\\theta_{i}^{*}$ . Eq. 6a denotes an outer-level optimization. It is based on the validation loss derived by $\\theta_{i}^{*}$ on the original data $\\mathcal{D}_i$ . $\\mathcal{R}(\\phi_i)$ is a constraint representing the memory limitation and $\\mu$ is its weight. In the following, we elaborate on the implementation details for the two levels.", + "bbox": [ + 496, + 650, + 890, + 755 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In the inner-level optimization, we train a temporary CIL model with compressed data. Specifically, we first compress new-class data $\\mathcal{D}_i$ into $\\tilde{\\mathcal{D}}_i(\\phi_i)$ using the generated masks by the CIM model $\\phi_i$ . Then, we implement the inner-level optimization as a one-step gradient descent (using the CIL training loss) as:", + "bbox": [ + 496, + 756, + 890, + 847 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\theta_ {i} ^ {+} \\leftarrow \\theta_ {i} - \\beta_ {1} \\nabla_ {\\theta} \\mathcal {L} _ {\\mathrm {C I L}} \\left(\\tilde {\\mathcal {E}} _ {0: i - 1} \\cup \\tilde {\\mathcal {D}} _ {i} \\left(\\phi_ {i}\\right); \\theta_ {i}, \\omega_ {i}\\right), \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 535, + 856, + 890, + 875 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\beta_{1}$ is the learning rate for $\\theta_{i}$ .", + "bbox": [ + 500, + 885, + 733, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_footnote", + "text": "4PAU uses a rational function of given degrees $m$ and $n$ , i.e., $(a_0 + a_1x + \\dots + a_mx^m) / (1 + b_1x + \\dots + b_nx^n)$ , and can be parameterized by $a_0, \\dots, a_m$ and $b_1, \\dots, b_n$ .", + "bbox": [ + 75, + 862, + 468, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "11375", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The aim of the outer-level optimization is to optimize $\\phi_{i}$ such that the temporary model $(\\theta_{i}^{+},\\omega_{i})$ (trained with compressed data) has a low validation loss on the original data $D_{i}$ . To achieve this, we back-propagate the loss on the original data $D_{i}$ to update $\\phi_{i}$ as:", + "bbox": [ + 76, + 90, + 468, + 167 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\phi_ {i} \\leftarrow \\phi_ {i} - \\beta_ {2} \\nabla_ {\\phi} \\left[ \\mathcal {L} _ {\\mathrm {C E}} \\left(\\mathcal {D} _ {i}; \\theta_ {i} ^ {+}, \\omega_ {i}\\right) + \\mu \\mathcal {R} \\left(\\phi_ {i}\\right) \\right], \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 102, + 178, + 468, + 196 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\mathcal{L}_{\\mathrm{CE}}$ denotes softmax cross-entropy loss and $\\beta_{2}$ is the learning rate for $\\phi_{i}$ . This trains $\\phi_{i}$ to capture the most discriminative features of new-class images. The constraint $\\mathcal{R}(\\phi_i)$ is implemented as a $\\ell_2$ -regularization term on the generated mask by $\\phi_{i}$ . The motivation for the regularization term is to make the mask coverage smaller, thus the compressed images take less memory.", + "bbox": [ + 75, + 207, + 468, + 313 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We empirically observe that by the above optimization flow, the output activation maps by $\\phi_{i}$ are easy to collapse, i.e., different images have the same map. To solve this issue, we add a cross-entropy loss term about $\\phi_{i}$ in Eq. 8 to regularize it to produce image-specific activation maps:", + "bbox": [ + 76, + 313, + 468, + 388 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\phi_ {i} \\leftarrow \\phi_ {i} - \\beta_ {2} \\nabla_ {\\phi} \\left[ \\mathcal {L} _ {\\mathrm {C E}} \\left(\\mathcal {D} _ {i}; \\theta_ {i} ^ {+}, \\omega_ {i}\\right) + \\mu \\mathcal {R} (\\phi_ {i}) \\right. \\tag {9} \\\\ + \\mu^ {\\prime} \\mathcal {L} _ {\\mathrm {C E}} \\left(\\tilde {\\mathcal {E}} _ {0: i - 1} \\cup \\mathcal {D} _ {i}; \\theta_ {i}, \\phi_ {i}, \\omega_ {i}\\right) ], \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 91, + 401, + 468, + 439 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\mu^{\\prime}$ is the weight.", + "bbox": [ + 76, + 450, + 230, + 465 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Limitations. Our CIM learns to generate adaptive masks for exemplar compression in CIL. It has three limitations that are left as future work. 1) It is not able to adjust any previous-phase exemplars, as the validation data (the original data of these exemplars) are not accessible anymore. 2) It introduces hundreds of activation parameters to the CIL model, although this is not a significant overhead compared to model parameters. Please check detailed overhead analyses in the supplementary materials. 3) Image compression is not that meaningful for low-resolution datasets (e.g., $32 \\times 32$ CIFAR-100). It is because the memory taken by the compression parameters (e.g., the parameters of CIM) and the RGB pixels of a low-resolution image are comparable. Using the memory to save more images is more meaningful.", + "bbox": [ + 75, + 465, + 468, + 676 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 76, + 690, + 207, + 707 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We incorporate CIM into two baseline CIL methods (i.e., DER [46] and FOSTER [42]) and boost their model performances consistently on three datasets. Below, we introduce datasets and experiment settings (Section 5.1), followed by results and analyses (Section 5.2)", + "bbox": [ + 75, + 715, + 468, + 792 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1. Experimental Settings", + "text_level": 1, + "bbox": [ + 76, + 801, + 284, + 816 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Datasets. We conduct experiments on three standard CIL benchmarks with high-resolution images. 1) Food-101 [3] consists of 101 food categories with 750 training and 250 test samples per category. All images have a maximum side length of 512 pixels. 2) ImageNet-1000 [9] is a large-scale", + "bbox": [ + 75, + 824, + 468, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "dataset with 1,000 classes and each class has around 1,300 training and 50 test samples. 3) ImageNet-100 is a 100-class subset randomly sampled from ImageNet-1000 with a fixed NumPy [14] random seed (1993), following [17]. We provide other details of these datasets, e.g., image sizes and pre-processing methods, in the supplementary materials.", + "bbox": [ + 496, + 90, + 890, + 181 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Protocols. We use two protocols: learning from scratch (LFS) and learning from half (LFH), following recent CIL works [42, 46]. In LFS, the model observes the same number of classes in all $N$ phases, where $N$ is optionally 5, 10, and 20. In LFH, the model is trained on half of the classes (e.g., 500 classes for ImageNet-1000) in the 1-th phase. Then, it learns the remaining classes evenly in the subsequent $N$ phases, where $N$ can be 5, 10, and 25. In both protocols, after the training of each phase we evaluate the resultant model on the test data of all seen classes. Our final report includes the average accuracy over all phases and the last-phase accuracy which indicates the degree of model forgetting. We run each experiment three times and report the average results.", + "bbox": [ + 496, + 181, + 892, + 393 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Memory Budget. There are two memory budget settings. 1) In the \"fixed\" setting, we remove some old-class exemplars when new exemplars from the current phase are added in the memory to maintain the \"fixed memory budget\". In this setting, we set the total memory to be 2,020 samples for Food-101 and 2,000 samples for ImageNet-100. For ImageNet-1000, we have two options—5,000 samples and 20,000 samples. 2) In the \"growing\" setting, a constant memory budget is allocated for each class across all phases and hence extra memory is appended when new classes come. In this setting, we set the budget to be 20 samples per class for all datasets. Following [42, 46], we apply the \"fixed\" setting in LFS experiments and the \"growing\" setting in LFH experiments.", + "bbox": [ + 496, + 393, + 892, + 606 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation Details. Our implementation is based on the standard deep learning library PyTorch [35] and image processing library OpenCV [4]. Following [42, 43, 46], we use an 18-layer ResNet [16] as the network backbone $\\theta$ and a fully-connected layer as the classifier $\\omega$ in all experiments. We use the same CIL training hyperparameters as in related works [17, 42, 46] for fair comparison: 1) there are 200 epochs in 1-st phase and 170 epochs in the subsequent phases; 2) the learning rate $\\lambda$ is initialized as 0.1 and decreases to zero with a cosine annealing scheduler [28]; 3) the SGD optimizer is deployed, with momentum factor set to 0.9 and weight decay set to 0.0005. For compression-related hyperparameters, we set the masking threshold $\\tau$ as 0.6 and the downsampling ratio $\\eta$ as 4.0. To build the CIM model, we apply PAUs with degrees $m = 5$ and $n = 4$ as learnable activation layers. For the optimization of the CIM", + "bbox": [ + 496, + 607, + 890, + 849 + ], + "page_idx": 5 + }, + { + "type": "page_footnote", + "text": "5Please note that we measure the memory by the number of original images. Each compressed exemplar in CIM takes less memory than the original image (Eq. 4), resulting in more exemplars in the same memory.", + "bbox": [ + 498, + 862, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "11376", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/622a8a85609712e4825ca61c0bfb9ea6637485efaf592c09b576824aaf19f70c.jpg", + "table_caption": [], + "table_footnote": [ + "† The paper of FOSTER [42] did not report the numerical results for $N = 5 / 10$ (LFS) and $N = 5$ (LFH) on ImageNet-100. We run these experiments using the public code (released by authors) and report the reproduced results.", + "* Our reproduced result (72.40) is significantly higher than the original result (69.34) reported in the paper of FOSTER." + ], + "table_body": "
MethodLearning from Scratch (LFS)Learning from Half (LFH)
Food-101ImageNet-100Food-101ImageNet-100
N=51020510205102551025
iCaRL [36]69.6662.1856.7073.9067.0662.3660.1353.4246.8762.5359.8852.97
WA [48]70.9463.6958.4574.6468.6263.2063.5557.6052.4865.7563.7158.34
PODNet [12]68.0361.2447.3872.1463.9653.6975.3770.0165.3275.5474.3368.33
AANets [25]69.4661.5948.8372.9865.7755.3676.0771.2266.9376.9675.5871.78
DER [46]73.8870.7664.3978.5076.1273.7978.1373.45-79.0877.73-
DER w/ ours75.6373.0969.1779.6377.5775.3679.2575.76-80.3079.05-
FOSTER [42]75.0372.7266.73\\( 79.93^{\\dagger} \\)\\( 76.55^{\\dagger} \\)74.4979.0875.0768.08\\( 80.07^{\\dagger} \\)77.54\\( 72.40^{\\times} \\)
FOSTER w/ ours76.4474.8570.2080.5877.9475.2379.7676.8670.5080.9378.6675.74
", + "bbox": [ + 86, + 88, + 883, + 277 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/b23490493f89c7f2690e20a2d27b79b921384b0beaa8d54c8eb9ec4f809b71aa.jpg", + "table_caption": [ + "Table 1. Average accuracies $(\\%)$ of two top-performing CIL methods [42, 46] with and without our CIM-CIL plugged-in, and other four baselines [12, 25, 36, 48], on two datasets (Food-101 and ImageNet-100) and using two protocols (learning from scratch (LFS) and learning from half (LFH)). Due to the space limits, we report the $95\\%$ confidence intervals for these results in the supplementary materials." + ], + "table_footnote": [], + "table_body": "
Memory BudgetMethodN=5N=10
Avg.LastAvg.Last
M=20kiCaRL [36]44.3627.7838.4022.70
WA [48]58.3750.6254.1045.66
DER [46]67.4959.7566.7358.62
FOSTER [42]69.2164.8868.3460.14
FOSTER w/ ours69.9366.0569.5362.07
M=5kFOSTER57.1949.4254.7244.96
FOSTER w/ ours61.3754.4659.4850.83
", + "bbox": [ + 78, + 378, + 468, + 532 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "model $\\phi$ (i.e., the mask-level optimization), we initially set $\\beta_{1}$ as 0.1 and $\\beta_{2}$ as 0.01 and reduce them to zero following the scheduler of $\\lambda$ . $\\mu$ and $\\mu^{\\prime}$ is set to 0.1 and 0.2, respectively. To smooth the training, we clip the gradient norm of $\\phi$ to be no more than 1. We report the result of hyperparameter sensitivity analysis in the supplementary materials.", + "bbox": [ + 75, + 608, + 468, + 700 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.2. Results and Analyses", + "text_level": 1, + "bbox": [ + 76, + 710, + 274, + 727 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Comparing with the State-of-the-art. In Table 1, we summarize the experimental results on two datasets (Food-101 and ImageNet-100) and in two CIL protocols (LFS and LFH). From the table, we have the following observations. 1) Our CIM-based CIL consistently improves the state-ofthe-art method FOSTER [42] with clear margins in all settings. E.g., our method surpasses it by an average of 1.4 percentage points on ImageNet-100, and 2.0 percentage points on Food-101. 2) Our CIM-based CIL achieves more significant improvements when $N$ becomes larger, e.g., on ImageNet-100 (LFH), our method improves FOSTER by", + "bbox": [ + 75, + 734, + 470, + 902 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/375693eb67a69177a651a4c1c7e2e5e3983b4e9d6f00954375df4f8f75ff0b64.jpg", + "table_caption": [ + "Table 2. Average and last accuracies $(\\%)$ on ImageNet-1000 of FOSTER with and without our method plugged-in, and other three baselines [36, 46, 48]. We show two memory budgets, $M = 20k$ (upper block) and $M = 5k$ (lower block), in the LFS setting." + ], + "table_footnote": [], + "table_body": "
Ablation MethodFood-101ImageNet-100
N=10201020
1 Baseline72.7266.7376.5572.37
2 Artifact Aug.71.3866.0375.6371.45
3 Full Comp.73.0367.3876.9273.26
4 Random Acti.73.1067.5476.8873.54
5 Center Acti.73.2967.8876.7873.82
6 Class Acti.73.7668.6577.2174.67
7 Phase-wise τ73.8369.1777.0674.78
8 Joint Train73.4469.0177.3474.59
9 BOP (ours)74.8570.2077.9475.23
10 LastBlock Only74.5569.8777.7274.86
11 Fg Compressed75.0270.1377.8775.46
", + "bbox": [ + 511, + 378, + 883, + 598 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 3. Average accuracies $(\\%)$ of different ablation methods. The experiments are conducted in the LFS setting.", + "bbox": [ + 498, + 602, + 890, + 632 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "0.9 and 3.3 percentage points when $N = 5$ and $N = 25$ , respectively. 3) Our CIM-based CIL achieves greater improvements consistently on Food-101 (than ImageNet-100). It improves baselines by 2.1 percentage points on Food-101, while the improvement is 1.4 on ImageNet-100 ( $N = 10$ , LFS). It shows that our method is particularly effective when the representative visual cues of a class are from some of its components, e.g., the \"cream\" of the class \"cake\".", + "bbox": [ + 496, + 643, + 890, + 763 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 2 shows the results on the large-scale dataset ImageNet-1000 in different memory settings ( $M = 20k$ and $M = 5k$ ). We can see that our CIM-based CIL improves FOSTER consistently. It is impressive that it achieves more improvements in the more strict memory setting ( $M = 5k$ ). Specifically, it boosts the average accuracy of FOSTER by 4.5 percentage points when $M = 5k$ , significantly higher than that of $M = 20k$ (1.0).", + "bbox": [ + 496, + 763, + 890, + 885 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Ablation Study. Table 3 shows the ablation results. First", + "bbox": [ + 500, + 885, + 890, + 901 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "11377", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/34b9068f8fd9872de0690fef02744f71377e2cad5de5f43ce08898974231e137.jpg", + "table_caption": [], + "table_footnote": [ + "Table 4. Comparing with Mnemonics [27] and MRDC [43]. We plug each of them in baseline LUCIR [17] for fair comparison." + ], + "table_body": "
MethodImageNet-100ImageNet-1000
N=6112661126
LUCIR baseline71.2269.6767.4565.2362.4359.88
w/ Mnemonics73.3072.1771.5066.1563.1263.08
w/ MRDC73.6272.8170.4467.6765.6062.74
w/ ours74.0573.7672.8468.0366.5463.77
", + "bbox": [ + 83, + 88, + 467, + 203 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "block: baselines. Row 1 is for the baseline FOSTER [42]. Row 2 shows the results of adding artifact augmentation (see Section 4.1). It shows directly apply this augmentation does not improve and even impair the model. Please note the models in below blocks all use this augmentation. Second block: activation methods. Rows 3-6 show the results of using different activation methods to compress exemplars. Row 3 is to downsample all pixels (i.e., no region is activated). Row 4 is to randomly select activation regions. Row 5 is to activate only the center region $(\\frac{1}{4}$ of the original image), while row 6 is to use naive CAM. Comparing them to row 1, we can see that using naively compressed exemplars can improve CIL models. Row 4 outperforms rows 3-5, validating that it is more reliable to use the model's activation to generate compressed exemplars. Third block: optimization methods. Rows 7-9 are on top of Row 6 and are the results of applying different optimization strategies. Row 7 is to manually select $\\tau$ using a held-out set ( $10\\%$ of the dataset). Row 8 is to jointly train CIL and CIM models (for each input batch). Row 9 is the proposed method of using a global BOP. Fourth block: two variants of CIM-based CIL. Rows 10-11 are two variants of row 9. In Row 10, only the activation layers in the last block of CIM are learnable, and previous blocks use ReLU. Compared to row 9, row 10 shows slightly worse performance. Row 11 shows the version of adding a weak downsampling $(\\eta' = 2.0)$ on discriminative regions, based on which more compressed exemplars are saved. It results in comparable performance to row 9 but increases costs.", + "bbox": [ + 76, + 248, + 472, + 686 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Comparing with Other Compression-based Methods. Table 4 shows our results comparing to two compression-based methods: Mnemonics [27] and MRDC [43]. We can see that our method consistently outperforms them in all settings. This is because our method does not sacrifice the discriminativeness of exemplars while improving the number (variance) of exemplars in a phase-adaptive manner. However, the two related methods either keep a fixed number of exemplars in the memory [27] or use uniform image compression without considering the properties of specific classes in different incremental phases [43].", + "bbox": [ + 75, + 686, + 468, + 853 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Visualizations (CAM vs. CIM). Figure 4 gives two visualization examples, \"Afghan hound\" and \"indigo bird\", each with the activation map as well as the bounding boxes.", + "bbox": [ + 75, + 854, + 470, + 901 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/21467937e61d1a11cdb13cd77972360696d8f03f49efae8b77e1b2e9adad253a.jpg", + "image_caption": [ + "Figure 4. Visualizations (CAM vs. CIM). The experimental setting is $N = 10$ (LFS) on ImageNet-100." + ], + "image_footnote": [], + "bbox": [ + 532, + 88, + 857, + 239 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/02ca84354d937caf9f9375f1866a70aeac36f0bd7b3478269512012eb3d29839.jpg", + "table_caption": [], + "table_footnote": [ + "Table 5. Results for small, middle and large objects in the setting of $N = 10$ (LFS) on ImageNet-100. \"Mean of #Exemplars\" denotes the average number of saved exemplars by our method. The baseline (FOSTER [42]) has this number as 20 for all classes." + ], + "table_body": "
MetricSmallMiddleLarge
Mean of #Exemplars39.4038.3034.77
Last Acc. (%, baseline)66.1368.4069.93
Last Acc. (%, ours)70.0071.1072.26
Improvement (%)+3.87+3.65+2.33
", + "bbox": [ + 500, + 279, + 885, + 371 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The first column shows their respective confusing classes appearing in earlier phases. CIM learns to focus on the discriminative (i.e., dissimilar to confusing classes) regions.", + "bbox": [ + 496, + 444, + 890, + 489 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Results of Different-Size Objects. Table 5 shows the results for small, middle, and large objects. These size categorization is according to ImageNet Object Localization Challenge [19]. We calculate the bbox coverage for each class and take top 30 classes with highest coverages as \"large\", near 30 classes with lowest coverages as \"small\" and the rest 40 classes as \"middle\". It is intriguing that our method achieves the highest improvement (over baseline) for small objects. Our explanation is that small objects benefit more from image compression (than large ones), as their images contain more background pixels to downsample.", + "bbox": [ + 496, + 489, + 892, + 656 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusions", + "text_level": 1, + "bbox": [ + 500, + 670, + 627, + 686 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We introduced a novel exemplar compression method for CIL, allowing us to save more representative exemplars but not increase memory budget. We achieved this compression by downsampling non-discriminative pixels of the image bounded by CAM masks. To generate adaptive masks, we proposed a novel method CIM that explicitly parameterizes a mask generation model and optimizes it in an end-to-end manner across incremental phases. Our method achieves consistent performance improvements over multiple baselines and can be taken as a flexible plug-and-play module.", + "bbox": [ + 496, + 696, + 890, + 849 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements. The author gratefully acknowledges the support by the Lee Kong Chian (LKC) Fellowship fund awarded by Singapore Management University.", + "bbox": [ + 498, + 854, + 892, + 901 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "11378", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 174, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Davide Abati, Jakub Tomczak, Tijmen Blankevoort, Simone Calderara, Rita Cucchiara, and Babak Ehteshami Bejnordi. Conditional channel gated networks for task-aware continual learning. In CVPR, 2020. 2", + "[2] Alexey Bochkovskiy, Chien-Yao Wang, and Hong-Yuan Mark Liao. Yolov4: Optimal speed and accuracy of object detection. arXiv, 2020. 5", + "[3] Lukas Bossard, Matthieu Guillaumin, and Luc Van Gool. Food-101 - mining discriminative components with random forests. In ECCV, 2014. 2, 6", + "[4] Gary Bradski. The opencv library. Dr. Dobb's Journal: Software Tools for the Professional Programmer, 2000. 6", + "[5] Kenneth R Castleman. Digital image processing. Prentice Hall Press, 1996. 4", + "[6] George Cazenavette, Tongzhou Wang, Antonio Torralba, Alexei A Efros, and Jun-Yan Zhu. Dataset distillation by matching training trajectories. In CVPR, 2022. 3", + "[7] Can Chen, Xi Chen, Chen Ma, Zixuan Liu, and Xue Liu. Gradient-based bi-level optimization for deep learning: A survey. arXiv, 2022. 2, 3", + "[8] Zhaozheng Chen, Tan Wang, Xiongwei Wu, Xian-Sheng Hua, Hanwang Zhang, and Qianru Sun. Class re-activation maps for weakly-supervised semantic segmentation. In CVPR, 2022. 3, 4", + "[9] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In CVPR, 2009. 1, 2, 3, 6", + "[10] Prithviraj Dhar, Rajat Vikram Singh, Kuan-Chuan Peng, Ziyan Wu, and Rama Chellappa. Learning without memorizing. In CVPR, 2019. 2", + "[11] Zhang Dong, Zhang Hanwang, Tang Jinhui, Hua Xiansheng, and Sun Qianru. Causal intervention for weakly supervised semantic segmentation. In NeurIPS, 2020. 4", + "[12] Arthur Douillard, Matthieu Cord, Charles Ollion, Thomas Robert, and Eduardo Valle. Podnet: Pooled outputs distillation for small-tasks incremental learning. In ECCV, 2020. 1, 2, 7", + "[13] Chelsea Finn, Pieter Abbeel, and Sergey Levine. Model-agnostic meta-learning for fast adaptation of deep networks. In ICML, 2017. 3", + "[14] Charles R Harris, K Jarrod Millman, Stefan J Van Der Walt, Ralf Gommers, Pauli Virtanen, David Cournapeau, Eric Wieser, Julian Taylor, Sebastian Berg, Nathaniel J Smith, et al. Array programming with numpy. Nature, 2020. 6", + "[15] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Delving deep into rectifiers: Surpassing human-level performance on imagenet classification. In ICCV, 2015. 5", + "[16] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, 2016. 5, 6", + "[17] Saihui Hou, Xinyu Pan, Chen Change Loy, Zilei Wang, and Dahua Lin. Learning a unified classifier incrementally via rebalancing. In CVPR, 2019. 1, 2, 6, 8", + "[18] Shenyang Huang, Vincent François-Lavet, and Guillaume Rabusseau. Neural architecture search for class-incremental learning. arXiv, 2019. 2" + ], + "bbox": [ + 78, + 114, + 470, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[19] Imagenet object localization challenge. https://www.kaggle.com/competitions/imagenet-object-localization-challenge/.8", + "[20] Anil K Jain. Fundamentals of digital image processing. Prentice-Hall, Inc., 1989. 4", + "[21] James Kirkpatrick, Razvan Pascanu, Neil Rabinowitz, Joel Veness, Guillaume Desjardins, Andrei A Rusu, Kieran Milan, John Quan, Tiago Ramalho, Agnieszka Grabska-Barwinska, et al. Overcoming catastrophic forgetting in neural networks. PNAS, 2017. 2", + "[22] Jungbeom Lee, Eunji Kim, and Sungroh Yoon. Antiadversarily manipulated attributions for weakly and semi-supervised semantic segmentation. In CVPR, 2021. 3, 4", + "[23] Zhizhong Li and Derek Hoiem. Learning without forgetting. PAMI, 2017. 2, 5", + "[24] Yaoyao Liu, Yingying Li, Bernt Schiele, and Qianru Sun. Online hyperparameter optimization for class-incremental learning. In AAAi, 2023. 2", + "[25] Yaoyao Liu, Bernt Schiele, and Qianru Sun. Adaptive aggregation networks for class-incremental learning. In CVPR, 2021. 1, 2, 3, 7", + "[26] Yaoyao Liu, Bernt Schiele, and Qianru Sun. Rmm: Reinforced memory management for class-incremental learning. In NeurIPS, 2021. 2", + "[27] Yaoyao Liu, Yuting Su, An-An Liu, Bernt Schiele, and Qianru Sun. Mnemonics training: Multi-class incremental learning without forgetting. In CVPR, 2020. 1, 2, 3, 8", + "[28] Ilya Loshchilov and Frank Hutter. Sgdr: Stochastic gradient descent with warm restarts. arXiv, 2016. 6", + "[29] Dougal Maclaurin, David Duvenaud, and Ryan Adams. Gradient-based hyperparameter optimization through reversible learning. In ICML, 2015. 3", + "[30] Michael McCloskey and Neal J Cohen. Catastrophic interference in connectionist networks: The sequential learning problem. In *Psychology of Learning and Motivation*. Elsevier, 1989. 2", + "[31] Ken McRae and Phil A Hetherington. Catastrophic interference is eliminated in pretrained networks. In Proceedings of the 15th Annual Conference of the Cognitive Science Society, 1993. 1, 2", + "[32] Alejandro Molina, Patrick Schramowski, and Kristian Kersting. Padé activation units: End-to-end learning of flexible activation functions in deep networks. In ICLR, 2019. 5", + "[33] Vinod Nair and Geoffrey E Hinton. Rectified linear units improve restricted boltzmann machines. In ICML, 2010. 5", + "[34] J Anthony Parker, Robert V Kenyon, and Donald E Troxel. Comparison of interpolating methods for image resampling. IEEE Transactions on Medical Imaging, 1983. 4", + "[35] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. In NeurIPS, 2019. 6", + "[36] Sylvestre-Alvise Rebuffi, Alexander Kolesnikov, Georg Sperl, and Christoph H Lampert. iCaRL: Incremental classifier and representation learning. In CVPR, 2017. 1, 2, 3, 5, 7" + ], + "bbox": [ + 501, + 92, + 890, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "11379", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[37] Andrei A Rusu, Neil C Rabinowitz, Guillaume Desjardins, Hubert Soyer, James Kirkpatrick, Koray Kavukcuoglu, Razvan Pascanu, and Raia Hadsell. Progressive neural networks. arXiv, 2016. 2", + "[38] Ramprasaath R Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra. Grad-cam: Visual explanations from deep networks via gradient-based localization. In ICCV, 2017. 3", + "[39] Christian Simon, Piotr Koniusz, and Mehrtash Harandi. On learning the geodesic path for incremental learning. In CVPR, 2021. 2", + "[40] Ankur Sinha, Pekka Malo, and Kalyanmoy Deb. A review on bilevel optimization: from classical to evolutionary approaches and applications. IEEE Transactions on Evolutionary Computation, 2017. 1, 2, 3", + "[41] Gregory K Wallace. The jpeg still picture compression standard. Communications of the ACM, 1991. 1, 2", + "[42] Fu-Yun Wang, Da-Wei Zhou, Han-Jia Ye, and De-Chuan Zhan. Foster: Feature boosting and compression for class incremental learning. arXiv, 2022. 1, 2, 3, 5, 6, 7, 8", + "[43] Liyuan Wang, Xingxing Zhang, Kuo Yang, Longhui Yu, Chongxuan Li, Lanqing Hong, Shifeng Zhang, Zhenguo Li, Yi Zhong, and Jun Zhu. Memory replay with data compression for continual learning. In ICLR, 2022. 1, 2, 3, 6, 8", + "[44] Yue Wu, Yinpeng Chen, Lijuan Wang, Yuancheng Ye, Zicheng Liu, Yandong Guo, and Yun Fu. Large scale incremental learning. In CVPR, 2019. 1, 2", + "[45] Ju Xu and Zhanxing Zhu. Reinforced continual learning. In NeurIPS, 2018. 2", + "[46] Shipeng Yan, Jiangwei Xie, and Xuming He. Der: Dynamically expandable representation for class incremental learning. In CVPR, 2021. 1, 2, 3, 5, 6, 7", + "[47] Friedemann Zenke, Ben Poole, and Surya Ganguli. Continual learning through synaptic intelligence. In ICML, 2017. 2", + "[48] Bowen Zhao, Xi Xiao, Guojun Gan, Bin Zhang, and Shu-Tao Xia. Maintaining discrimination and fairness in class incremental learning. In CVPR, 2020. 1, 7", + "[49] Bolei Zhou, Aditya Khosla, Agata Lapedriza, Aude Oliva, and Antonio Torralba. Learning deep features for discriminative localization. In CVPR, 2016. 1, 2, 3", + "[50] Da-Wei Zhou, Han-Jia Ye, and De-Chuan Zhan. Co-transport for class-incremental learning. In Proceedings of the 29th ACM International Conference on Multimedia, 2021. 5" + ], + "bbox": [ + 78, + 90, + 468, + 729 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "11380", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2023/Class-Incremental Exemplar Compression for Class-Incremental Learning/47061da4-a9f4-49a6-b036-8d247f22c620_model.json b/2023/Class-Incremental Exemplar Compression for Class-Incremental Learning/47061da4-a9f4-49a6-b036-8d247f22c620_model.json new file mode 100644 index 0000000000000000000000000000000000000000..0a87ccdd8a59c2c012624ce395103bac4fc810c7 --- /dev/null +++ b/2023/Class-Incremental Exemplar Compression for Class-Incremental Learning/47061da4-a9f4-49a6-b036-8d247f22c620_model.json @@ -0,0 +1,2112 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.109, + 0.131, + 0.862, + 0.154 + ], + "angle": 0, + "content": "Class-Incremental Exemplar Compression for Class-Incremental Learning" + }, + { + "type": "text", + "bbox": [ + 0.257, + 0.18, + 0.728, + 0.2 + ], + "angle": 0, + "content": "Zilin Luo1 Yaoyao Liu2 Bernt Schiele2 Qianru Sun1" + }, + { + "type": "text", + "bbox": [ + 0.35, + 0.216, + 0.636, + 0.235 + ], + "angle": 0, + "content": "\\(^{1}\\)Singapore Management University" + }, + { + "type": "text", + "bbox": [ + 0.223, + 0.235, + 0.763, + 0.253 + ], + "angle": 0, + "content": "\\(^{2}\\)Max Planck Institute for Informatics, Saarland Informatics Campus" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.255, + 0.897, + 0.27 + ], + "angle": 0, + "content": "zilin.luo.2021@phdcs.smu.edu.sg {yaoyao.liu, schiele}@mpi-inf.mpg.de qianrusun@smu.edu.sg" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.305, + 0.314, + 0.32 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.336, + 0.474, + 0.76 + ], + "angle": 0, + "content": "Exemplar-based class-incremental learning (CIL) [36] finetunes the model with all samples of new classes but few-shot exemplars of old classes in each incremental phase, where the \"few-shot\" abides by the limited memory budget. In this paper, we break this \"few-shot\" limit based on a simple yet surprisingly effective idea: compressing exemplars by downsampling non-discriminative pixels and saving \"many-shot\" compressed exemplars in the memory. Without needing any manual annotation, we achieve this compression by generating 0-1 masks on discriminative pixels from class activation maps (CAM) [49]. We propose an adaptive mask generation model called class-incremental masking (CIM) to explicitly resolve two difficulties of using CAM: 1) transforming the heatmaps of CAM to 0-1 masks with an arbitrary threshold leads to a trade-off between the coverage on discriminative pixels and the quantity of exemplars, as the total memory is fixed; and 2) optimal thresholds vary for different object classes, which is particularly obvious in the dynamic environment of CIL. We optimize the CIM model alternatively with the conventional CIL model through a bilevel optimization problem [40]. We conduct extensive experiments on high-resolution CIL benchmarks including Food-101, ImageNet-100, and ImageNet-1000, and show that using the compressed exemplars by CIM can achieve a new state-of-the-art CIL accuracy, e.g., 4.8 percentage points higher than FOSTER [42] on 10-Phase ImageNet-1000. Our code is available at https://github.com/xffzl/CIM-CIL." + }, + { + "type": "title", + "bbox": [ + 0.079, + 0.785, + 0.208, + 0.8 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.811, + 0.47, + 0.903 + ], + "angle": 0, + "content": "Dynamic AI systems have a continual learning nature to learn new class data. They are expected to adapt to new classes while maintaining the knowledge of old classes, i.e., free from forgetting problems [31]. To evaluate this, the following protocol of class-incremental learning (CIL) was proposed by Rebuffi et al. [36]. The model training goes" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.326, + 0.892, + 0.394 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.567, + 0.395, + 0.661, + 0.405 + ], + "angle": 0, + "content": "(a) iCaRL [baseline]" + }, + { + "type": "image", + "bbox": [ + 0.504, + 0.412, + 0.892, + 0.479 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.569, + 0.48, + 0.659, + 0.49 + ], + "angle": 0, + "content": "(c) MRDC [related]" + }, + { + "type": "image", + "bbox": [ + 0.536, + 0.495, + 0.858, + 0.528 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.541, + 0.892, + 0.666 + ], + "angle": 0, + "content": "Figure 1. The phase-wise training data in different methods. (a) iCaRL [36] is the baseline method using full new class data and few-shot old class exemplars. (b) Mnemonics [27] distills all training samples into few-shot exemplars without increasing their quantity. (c) MRDC [43] compresses each exemplar uniformly into a low-resolution image using JPEG [41]. (d) Our approach based on the proposed class-incremental masking (CIM) down-samples only non-discriminative pixels in the image. The legend shows the symbols of special images generated by the methods." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.673, + 0.892, + 0.809 + ], + "angle": 0, + "content": "through a number of phases. Each phase has new class data added and old class data discarded, and the resultant model is evaluated on the test data of all seen classes. A straightforward way to retain old class knowledge is keeping around a few old class exemplars in the memory and using them to re-train the model in subsequent phases. The number of exemplars is usually limited, e.g., \\(5 \\sim 20\\) exemplars per class [12, 17, 25, 27, 36, 42, 44, 46, 48], as the total memory in CIL strictly budgeted, e.g., \\(2k\\) exemplars." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.81, + 0.892, + 0.901 + ], + "angle": 0, + "content": "This leads to a serious data imbalance between old and new classes, e.g., 20 per old class vs. \\(1.3k\\) per new class (on ImageNet-1000 [9]), as illustrated in Figure 1a. The training is thus always dominated by new classes, and forgetting problems occur for old classes. Liu et al. [27] tried to mitigate this problem by parameterizing and distilling the exem" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "11371" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.302 + ], + "angle": 0, + "content": "plars, without increasing the number of them (Figure 1b). Wang et al. [43] traded off between the quality and quantity of exemplars by uniformly compressing exemplar images with JPEG [41] (Figure 1c). As shown in Figure 1d, our approach is also based on image compression. The idea is to downsample only non-discriminative pixels (e.g., background) and keep discriminative pixels (i.e., representative cues of foreground objects) as the original. In this way, we do not sacrifice the discriminativeness of exemplars when increasing their quantity. In particular, we aim for adaptive compression in dynamic environments of CIL, where the intuition is later phases need to be more conservative (i.e., less downsampling) as the model needs more visual cues to classify the increased number of classes." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.303, + 0.473, + 0.65 + ], + "angle": 0, + "content": "To achieve selective and adaptive compression, we need the location labels of discriminative pixels. Without extra labeling, we automatically generate the labels by utilizing the model's own \"attention\" on discriminative features, i.e., class activation maps (CAM) [49]. We take this method as a feasible baseline, and based on it, we propose an adaptive version called class-incremental masking (CIM). Specifically, for each input image (with its class label), we use its feature maps and classifier weights (corresponding to its class label) to compute a CAM by channel-wise multiplication, aggregation, and normalization. Then, we apply hard thresholding to generate a 0-1 mask. We notice that when generating the masks in the dynamic environments of CIL, the optimal hyperparameters (such as the value of hard threshold and the choice of activation functions) vary for different classes as well as in different incremental phases. Our adaptive version CIM tackles this by parameterizing a mask generation model and optimizing it in an end-to-end manner across all incremental phases. In each phase, the learned CIM model adaptively generates class-and phase-specific masks. We find that the compressed exemplars based on these masks have stronger representativeness, compared to using the conventional CAM." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.65, + 0.473, + 0.833 + ], + "angle": 0, + "content": "Technically, we have two models to optimize, i.e., the CIL model and the CIM model. These two cannot be optimized separately as they are dependent on computation: 1) the CIM model compresses exemplars to input into the CIL model; 2) the two models share network parameters. We exploit a global bilevel optimization problem (BOP) [7, 40] to alternate their training processes at two levels. This BOP goes through all incremental training phases. In particular, for each phase, we perform a local BOP with two steps to tune the parameters of the CIM model: 1) a temporary model is trained with the compressed exemplars as input; and 2) a validation loss on the uncom" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.092, + 0.895, + 0.26 + ], + "angle": 0, + "content": "pressed new data is computed and the gradients are backpropagated to optimize the parameters of CIM. To evaluate CIM, we conduct extensive experiments by plugging it in recent CIL methods,\\(^{3}\\) LUCIR [17], DER [46], and FOSTER [42], on three high-resolution benchmarks, Food101 [3], ImageNet-100 [17], and ImageNet-1000 [9]. We find that using the compressed exemplars by CIM brings consistent and significant improvements, e.g., \\(4.2\\%\\) and \\(4.8\\%\\) higher than the SOTA method FOSTER [42], respectively, in the 5-phase and 10-phase settings of ImageNet1000, with a total memory budget for \\(5k\\) exemplars." + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.269, + 0.642, + 0.285 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.294, + 0.895, + 0.627 + ], + "angle": 0, + "content": "Class-Incremental Learning (CIL). There are three main lines of work to address the catastrophic forgetting problem [30, 31] in CIL. Regularization-based methods apply discrepancy (between old and new models) penalization terms in their objective functions, e.g., by comparing output logits [23, 36], intermediate features [12, 17, 24, 39], and prediction heatmaps [10]. Parameter-isolation-based methods increase the model parameters in each new incremental phase, to prevent knowledge forgetting caused by parameter overwritten. Some of them [18, 37, 42, 45, 46] proposed to progressively expand the size of the neural network to learn new coming data. Others [1, 21, 25, 47] froze a part of network parameters (to maintain the old class knowledge) to alleviate the problem of knowledge overwriting. Replay-based methods assume there is a clear memory budget allowing a handful of old-class exemplars in the memory. Exemplars can be used to re-train the model in each new phase [12, 17, 26, 27, 36, 43, 44]. This re-training usually contains two steps: one step of training the model on all new class data and old class exemplars, and one step of finetuning the model with a balanced subset (i.e., using an equal number of samples per class) [12, 17, 25, 26, 46]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.628, + 0.896, + 0.87 + ], + "angle": 0, + "content": "The replay-based methods focusing on memory optimization [27, 43] are closely related to our work. [27] proposed a bilevel optimization framework to distill the current new class data into exemplars before discarding them. It aims to improve the quality of exemplars without increasing the quantity. Another work [43] aimed to trade-off between the quality and quantity of exemplars by image compression using the JPEG algorithm, i.e., each exemplar is uniformly downsampled. Ours differs from these two works in three aspects. 1) Our CIM based image compression automatically segments the discriminative pixels in the exemplar and downsamples the non-discriminative pixels only. It barely weakens the representativeness of the exemplar. In contrast, the parameterization of image pixels in [27] hampers the model from capturing high-frequency (discriminative) features from the image, especially the high-resolution (e.g.," + }, + { + "type": "page_footnote", + "bbox": [ + 0.5, + 0.875, + 0.894, + 0.901 + ], + "angle": 0, + "content": "3Using \"plug-in\" for evaluation is due to the fact that many baseline methods were originally evaluated in different CIL settings." + }, + { + "type": "page_footnote", + "bbox": [ + 0.076, + 0.839, + 0.47, + 0.875 + ], + "angle": 0, + "content": "1Note that we do not use mask labels to do image compression because storing them is expensive. Instead, we expand the mask to a bounding box, as elaborated in Section 4." + }, + { + "type": "page_footnote", + "bbox": [ + 0.076, + 0.875, + 0.469, + 0.9 + ], + "angle": 0, + "content": "2Note that the CIM model is actually a plug-in branch in the CIL model, which is detailed in Section 4.2." + }, + { + "type": "list", + "bbox": [ + 0.076, + 0.839, + 0.47, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "11372" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.468, + 0.227 + ], + "angle": 0, + "content": "\\(224 \\times 224)\\) image [6]. 2) Our approach increases the diversity (i.e., quantity) of old class exemplars by reducing the memory consumption for each exemplar. In contrast, [27] keeps a fixed number of exemplars in the memory. 3) Our approach has an adaptive image compression strategy that fits well in the dynamic environments of CIL. In contrast, [43] uses uniform image compression (or uniformly increasing quality parameters from 100 to 1) without considering the properties of specific classes in each learning phase." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.228, + 0.468, + 0.363 + ], + "angle": 0, + "content": "Class Activation Map (CAM) [49] is a simple yet effective weakly-supervised object localization method. Its model is trained with only the image-level label and can generate pixel-level masks on foreground objects. Specifically, the masks are the results of hard-thresholding the heatmaps produced by feature maps and classifier weights. Advanced CAM variants include Grad-CAM [38], ReCAM [8], AdvCAM [22], etc. Our CIM is based on the vanilla CAM because it is computationally simple and efficient." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.364, + 0.468, + 0.575 + ], + "angle": 0, + "content": "Bilevel Optimization Problems (BOP) [7, 40] aims to solve a nested optimization problem, where the outer-level optimization is subjected to the result of the inner-level optimization. It has shown effectiveness in a wide range of machine learning areas, such as hyperparameter selection [29] and meta-learning [13]. For tackling CIL tasks, [27] leverages BOP to alternatively optimize the parameters of the CIL model and the parameterized exemplars. [25] applies BOP to learn the aggregation weights of the plastic and elastic branches in the CIL model. In our work, we use BOP to solve the optimization of the CIL model and the parameterized Class-Incremental Masking (CIM) model, where CIM is a plugin branch (in the CIL model), using few extra parameters. The process of BOP is quick and efficient." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.589, + 0.203, + 0.606 + ], + "angle": 0, + "content": "3. Preliminary" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.614, + 0.468, + 0.839 + ], + "angle": 0, + "content": "The following is the training pipeline of standard CIL with few-shot exemplars. Assume there are \\(N\\) learning phases. In the 1-st phase, we load data \\(\\mathcal{D}_1\\) containing all training samples of \\(c_1\\) classes, and use \\(\\mathcal{D}_1\\) to train the initial classification model \\((\\theta_{1},\\omega_{1})\\), where \\(\\theta_{1}\\) and \\(\\omega_{1}\\) denote the parameters of the feature extractor and classifier, respectively. When the training is done, we evaluate the model performance on the test samples of \\(c_1\\) classes. Before the 2nd phase, we discard most of the training samples due to the strict memory budget of CIL. In other words, we preserve only a handful of training samples \\(\\mathcal{E}_1\\) (i.e., exemplars) in the memory, selected from \\(\\mathcal{D}_1\\). A common method for selecting exemplars is called feature herding [36] and has been used in many related works [27, 42, 43, 46]. We adopt it, too, in this work." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.468, + 0.902 + ], + "angle": 0, + "content": "In the \\(i\\)-th phase \\((i \\geq 2)\\), we load all exemplars \\(\\mathcal{E}_{1:i-1} = \\mathcal{E}_1 \\cup \\dots \\cup \\mathcal{E}_{i-1}\\) from the memory and initialize the current model \\((\\theta_i, \\omega_i)\\) by the previous model \\((\\theta_{i-1}, \\omega_{i-1})\\). We use \\(\\mathcal{E}_{1:i-1}\\) and the new coming data \\(\\mathcal{D}_i\\) (containing \\(c_i\\)" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.199 + ], + "angle": 0, + "content": "new classes) to train \\((\\theta_{i},\\omega_{i})\\). Then, we evaluate the current model using a test set of all \\(\\sum_{j = 1}^{i}c_{j}\\) classes seen so far. After that, we discard most of the training samples in \\(\\mathcal{D}_i\\) and leave few-shot exemplars \\(\\varepsilon_{i}\\) in the memory. It is clear that this discarding causes a strong data imbalance between old and new coming classes in the subsequent phase. In the following, we introduce our solution to this problem." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.211, + 0.636, + 0.228 + ], + "angle": 0, + "content": "4. Methodology" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.235, + 0.892, + 0.522 + ], + "angle": 0, + "content": "As illustrated in Figure 1d, we alleviate the data imbalance problem by saving a larger number of compressed exemplars for old classes, where we leverage pixel-selective compression, i.e., downsample only non-discriminative pixels. To achieve this, the ideal case is that we have the pixel-level localization of foreground objects. However, the realistic case is that such localization labels are expensive, most CIL benchmarks do not have the labels, and it is not fair to compare with other CIL methods if using the labels. Without extra labeling, we introduce a CAM-based mask generation method, and based on it, we provide a baseline solution to pixel-selective compression in Section 4.1. The problem of mask generation in CIL is that the optimal generation hyperparameters such as hard thresholds are changing in the dynamic environment (with the increasing number of classes and phases). It is thus desirable to have an adaptive mask generation process. To this end, we propose class-incremental masking (CIM)—a learnable mask generation model, in Section 4.2." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.531, + 0.802, + 0.547 + ], + "angle": 0, + "content": "4.1. CAM-based Compression Pipeline" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.554, + 0.892, + 0.734 + ], + "angle": 0, + "content": "Generating pixel-level labels for large-scale datasets, e.g., ImageNet [9], is non-trivial. Using class activation maps (CAM) is a naive solution with little computation costs. Its key idea is to make use of the activation of the classification model itself: on the feature maps, activated pixels are more discriminative than non-activated ones for recognizing the object, where \"activated\" means of high activation values and with strong correlation with the classification weights of the object. After localizing activated pixels by CAM, we can generate a 0-1 mask on them, e.g., by hard thresholding their normalized values, and then up-sample the mask to the size of the input image [49]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.735, + 0.892, + 0.81 + ], + "angle": 0, + "content": "From CAM to 0-1 Mask. We extract CAM in the following steps. Given an image \\( x \\) from \\( \\mathcal{D}_i \\) and its ground truth class label \\( y \\), let \\( F(x; \\theta_i) \\) represent the feature block output by the feature extractor \\( \\theta_i \\), and \\( \\omega_{i,y} \\) for the classification weights of class \\( y \\) in the classifier \\( \\omega_i \\). The CAM of \\( x \\) is:" + }, + { + "type": "equation", + "bbox": [ + 0.515, + 0.817, + 0.892, + 0.849 + ], + "angle": 0, + "content": "\\[\n\\mathcal {M} ^ {\\mathrm {C A M}} = \\frac {A - \\operatorname* {m i n} (A)}{\\operatorname* {m a x} (A) - \\operatorname* {m i n} (A)}, A = \\omega_ {i, y} ^ {\\top} F (x; \\theta_ {i}), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.892, + 0.9 + ], + "angle": 0, + "content": "where \\(\\min (\\cdot)\\) and \\(\\max (\\cdot)\\) operations are used for normalization. Then, we upsample \\(\\mathcal{M}^{\\mathrm{CAM}}\\) to the size of image \\(x\\) and use the same notation." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "11373" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.09, + 0.085, + 0.892, + 0.209 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.218, + 0.895, + 0.261 + ], + "angle": 0, + "content": "Figure 2. The proposed compression pipeline. Given an image, we extract its CAM-based (baseline) or CIM-based (ours) mask, threshold it to be a 0-1 mask with a fixed threshold \\(\\tau\\), and generate a tight bounding box (bbox) that covers all masked pixels. Then, we downsample the pixels outside the bbox and sum the downsampled image up to the masked image to generate the final compressed image." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.272, + 0.47, + 0.438 + ], + "angle": 0, + "content": "Each value in \\(\\mathcal{M}^{\\mathrm{CAM}}\\) denotes the activation strength of the model at a specific pixel location. Following the way of generating 0-1 masks in weakly-supervised semantic segmentation works [8, 11, 22], we apply a hard threshold \\(\\tau\\) (between 0 and 1) over all values of \\(\\mathcal{M}^{\\mathrm{CAM}}\\), and get the 0-1 mask \\(\\mathcal{M}^{\\tau}\\): \\(\\mathcal{M}^{\\tau} = \\mathbb{I}(\\mathcal{M}^{\\mathrm{CAM}} > \\tau)\\), where \\(\\mathbb{I}(\\cdot)\\) is the indicator function. In \\(\\mathcal{M}^{\\tau}\\), 1s indicate the locations of discriminative pixels, e.g., foreground pixels, based on which the model makes the prediction. While 0s indicate mostly background pixels or non-discrimination pixels that can be downsampled as they make little contribution to the prediction." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.439, + 0.47, + 0.575 + ], + "angle": 0, + "content": "After generating 0-1 masks, it is ideal to keep them in the memory as meta information of compression. However, this is not efficient or feasible in CIL. There are two reasons. 1) The space for saving image-size masks is non-negligible. Each mask pixel is a one-bit boolean value, and one mask takes around \\( \\frac{1}{3 \\times 8} = \\frac{1}{24} \\) of the memory of one RGB image. 2) The mask involves activated regions with irregular shapes. It is thus non-trivial to perform any standard down-sampling algorithm [34] on the remaining regions." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.575, + 0.47, + 0.651 + ], + "angle": 0, + "content": "From 0-1 Masks to Bounding Boxes (BBox). A simple workaround is to generate a tight bounding box (bbox) to cover the positions of 1s in \\(\\mathcal{M}^{\\tau}\\), and use the_bbox for compression. Specifically, given \\(\\mathcal{M}^{\\tau}\\), we obtain the coordinate representation of the bounding box as:" + }, + { + "type": "equation", + "bbox": [ + 0.089, + 0.66, + 0.469, + 0.677 + ], + "angle": 0, + "content": "\\[\n\\mathcal {B} = \\left[ \\min h, \\min w; \\max h, \\max w \\right] _ {(h, w): \\mathcal {M} ^ {\\tau} (h, w) = 1}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.685, + 0.469, + 0.79 + ], + "angle": 0, + "content": "where \\( h \\) and \\( w \\) denote the vertical and horizontal coordinates of the 1 on \\( \\mathcal{M}^{\\tau} \\), respectively. We highlight that \\( \\mathcal{B} \\) consists of four integers only and takes negligible memory overhead compared to \\( \\mathcal{M}^{\\tau} \\). In addition, we \"reshape\" the irregular shape (of the activated region) in \\( \\mathcal{M}^{\\tau} \\) into a rectangular \\( \\mathcal{B} \\), so our downsampling operation on the pixels outside the rectangular becomes easy." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.79, + 0.469, + 0.851 + ], + "angle": 0, + "content": "Compression with BBox. Given the image and its bbox on foreground, the compression is implemented by downsampling pixels outside the bbox. Specifically, as illustrated in Figure 2, we compress the image \\( x \\) to \\( \\tilde{x} \\) as follows," + }, + { + "type": "equation", + "bbox": [ + 0.159, + 0.859, + 0.469, + 0.877 + ], + "angle": 0, + "content": "\\[\n\\tilde {x} = \\mathcal {M} ^ {\\mathcal {B}} \\odot x + (1 - \\mathcal {M} ^ {\\mathcal {B}}) \\odot x _ {\\eta}, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.885, + 0.469, + 0.901 + ], + "angle": 0, + "content": "where \\(\\mathcal{M}^{\\mathcal{B}}\\) is the binary mask according to \\(\\mathcal{B}\\), i.e., the val" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.272, + 0.892, + 0.348 + ], + "angle": 0, + "content": "ues of \\(\\mathcal{M}^{\\mathcal{B}}\\) are 1 inside \\(\\mathcal{B}\\), and 0 otherwise. \\(x_{\\eta}\\) is the fully downsampled version of \\(x\\) with a downsampling ratio \\(\\eta\\) (\\(\\eta > 1\\)), \\(\\odot\\) denotes the element-wise product, and \\(+\\) denotes the element-wise addition. Both \\(\\odot\\) and \\(+\\) are applied independently on each RGB channel." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.348, + 0.892, + 0.377 + ], + "angle": 0, + "content": "The memory allocated for the compressed image \\(\\tilde{x}\\) is as follows," + }, + { + "type": "equation", + "bbox": [ + 0.568, + 0.387, + 0.892, + 0.452 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} m _ {\\tilde {x}} = \\frac {H _ {\\mathcal {B}} W _ {\\mathcal {B}}}{H W} + \\frac {1}{\\eta} \\left(1 - \\frac {H _ {\\mathcal {B}} W _ {\\mathcal {B}}}{H W}\\right) \\tag {4} \\\\ = 1 - \\left(1 - \\frac {1}{\\eta}\\right) \\cdot \\left(1 - \\frac {H _ {\\mathcal {B}} W _ {\\mathcal {B}}}{H W}\\right), \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.46, + 0.892, + 0.567 + ], + "angle": 0, + "content": "where \\( H_{\\mathcal{B}} \\) and \\( W_{\\mathcal{B}} \\) are the height and width of \\( \\mathcal{B} \\), respectively. \\( H \\) and \\( W \\) are the height and width of the original image \\( x \\), respectively. \\( m_{\\tilde{x}} \\) is always smaller than 1 where 1 denotes the memory unit of saving one original image \\( x \\). Therefore, we can save a larger number of compressed exemplars within the same memory budget. We denote the set of compressed exemplars in the \\( i \\)-th phase as \\( \\tilde{\\mathcal{E}}_i \\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.567, + 0.892, + 0.763 + ], + "angle": 0, + "content": "Compression Artifacts. The above compression introduces artifacts to the compressed images, i.e., there is a resolution mutation around bounding box edges. From the perspective of spectrum analysis [5,20], such mutation carries noisy and high-frequency components and impairs the model training in subsequent phases. We mitigate the effect of these artifacts by implementing the following data augmentation: in each training epoch, we transform a random subset of \\(\\mathcal{D}_i\\) into compressed images with CAM-based bounding boxes using the same downsampling ratio. Using this augmentation enables the model to \"simulate\" the training with compressed images and learn to be invariant to compression artifacts." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.772, + 0.803, + 0.789 + ], + "angle": 0, + "content": "4.2. Class-Incremental Masking (CIM)" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.796, + 0.892, + 0.9 + ], + "angle": 0, + "content": "Ideally, the mask generation process needs to adjust at different phases in CIL environments. The process involves two hyperparameters: the masking threshold and the choice of network activation functions. First, for the threshold, searching for its optimum (for all classes) is not trivial. Grid search is intuitive, but it is computationally expensive when the number of classes increases in CIL. Second, for the acti" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "11374" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.08, + 0.072, + 0.473, + 0.216 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.229, + 0.47, + 0.313 + ], + "angle": 0, + "content": "Figure 3. Our CIM installs an \"extension\" on the backbone (e.g., ResNet-18 [16, 36] with four blocks) by adding a learnable activation function (e.g., PAU [32]) at the position of the original activation function (i.e., ReLU [33]). Such \"extension\" results in a new network branch (in green), whose weight layer parameters are directly copied from the original branch (in blue)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.328, + 0.47, + 0.358 + ], + "angle": 0, + "content": "vation function, the standard network of CIL methods uses ReLU [33, 36] and does not optimize it." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.36, + 0.469, + 0.496 + ], + "angle": 0, + "content": "We solve this issue by applying a learnable activation function in addition to the existing ReLU function in the CIL model. Physically, we have one neural network, while logically, we have two models to learn (in each incremental phase): the conventional CIL model with ReLU activations, and the adaptive mask generation model with learnable activations. We thus call our method class-incremental masking (CIM) based CIL. In the following, we elaborate on the network design and optimization pipeline." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.497, + 0.47, + 0.738 + ], + "angle": 0, + "content": "Network Design. Figure 3 demonstrates an example network architecture in our CIM-based CIL. The proposed CIM extends the network backbone by logically adding a network branch, where only the activation functions are learnable (e.g., Padé Activation Units \\((\\mathrm{PAU}^4\\), [32])) and the parameters of weight layers are copied from the original branch. This design is motivated by the works of He et al. [15] and Bochkovskiy et al. [2], which indicate that layers with learnable activation functions can flexibly process object (localization) information at different network blocks. The difference is that we apply this flexibility to achieve adaptive mask generation for different CIL phases. We denote the CIM parameters (i.e., the parameters in learnable activation functions) in \\(i\\)-th phase as \\(\\phi_i\\). We optimize the CIL model \\((\\theta_i, \\omega_i)\\) and the CIM model \\(\\phi_i\\) via a global BOP, as elaborated below." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.74, + 0.469, + 0.816 + ], + "angle": 0, + "content": "Optimization Pipeline. We demonstrate the overall optimization flow in Algorithm 1, which consists of two levels of optimization: task-level and mask-level—the former one for CIL and the latter one for CIM. Note that to maintain an unified notation, we further define \\(\\tilde{\\mathcal{E}}_0 = \\varnothing\\)." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.817, + 0.47, + 0.848 + ], + "angle": 0, + "content": "1) Task-level Optimization. This level aims to optimize the CIL model \\((\\theta_i, \\omega_i)\\) to address the CIL task at hand. It" + }, + { + "type": "page_footnote", + "bbox": [ + 0.076, + 0.863, + 0.469, + 0.901 + ], + "angle": 0, + "content": "4PAU uses a rational function of given degrees \\(m\\) and \\(n\\), i.e., \\((a_0 + a_1x + \\dots + a_mx^m) / (1 + b_1x + \\dots + b_nx^n)\\), and can be parameterized by \\(a_0, \\dots, a_m\\) and \\(b_1, \\dots, b_n\\)." + }, + { + "type": "code_caption", + "bbox": [ + 0.51, + 0.095, + 0.832, + 0.11 + ], + "angle": 0, + "content": "Algorithm 1: CIM-based CIL (i-th phase, \\( i \\geq 1 \\))" + }, + { + "type": "algorithm", + "bbox": [ + 0.508, + 0.112, + 0.857, + 0.385 + ], + "angle": 0, + "content": "Input: New data \\(\\mathcal{D}_i\\) ; old compressed exemplars \\(\\tilde{\\mathcal{E}}_{0:i - 1}(\\tilde{\\mathcal{E}}_0 = \\emptyset)\\) ; last-phase CIL model \\((\\theta_{i - 1},\\omega_{i - 1})\\) and CIM model \\(\\phi_{i - 1}\\) \\((\\theta_0,\\omega_0\\) are random parameters, \\(\\phi_0\\) is set to ReLU). Output: New compressed exemplars \\(\\tilde{\\mathcal{E}}_i\\) ; updated CIL model \\((\\theta_i,\\omega_i)\\) and CIM model \\(\\phi_{i}\\) \n1 Initialize \\((\\theta_{i},\\omega_{i})\\) with \\((\\theta_{i - 1},\\omega_{i - 1})\\) . \n2 Initialize \\(\\phi_{i}\\) with \\(\\phi_{i - 1}\\) . \n3 for epochs do. \n4 Train \\((\\theta_{i},\\omega_{i})\\) using \\(\\tilde{\\mathcal{E}}_{0:i - 1}\\cup \\mathcal{D}_i\\) by Eq. 5; \n5 Compress \\(\\mathcal{D}_i\\) into \\(\\tilde{\\mathcal{D}}_i^\\phi\\) using \\(\\phi_{i}\\) . \n6 Temporarily update \\(\\theta_{i}\\) to \\(\\theta_i^+\\) using \\(\\tilde{\\mathcal{E}}_{0:i - 1}\\cup \\tilde{\\mathcal{D}}_i^\\phi\\) by Eq.7; \n7 Learn \\(\\phi_{i}\\) using \\(\\mathcal{D}_i\\) by Eq.9; \n8 end \n9 Compress \\(\\mathcal{D}_i\\) into \\(\\tilde{\\mathcal{D}}_i\\) using the learned \\(\\phi_{i}\\) by Eq. 3 \n0 Select exemplars \\(\\tilde{\\mathcal{E}}_i\\) from \\(\\tilde{\\mathcal{D}}_i\\) by, e.g., herding [36]." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.416, + 0.62, + 0.429 + ], + "angle": 0, + "content": "can be written as:" + }, + { + "type": "equation", + "bbox": [ + 0.515, + 0.441, + 0.892, + 0.472 + ], + "angle": 0, + "content": "\\[\n\\left(\\theta_ {i}, \\omega_ {i}\\right) \\leftarrow \\left(\\theta_ {i}, \\omega_ {i}\\right) - \\lambda \\nabla_ {\\left(\\theta , \\omega\\right)} \\mathcal {L} _ {\\mathrm {C I L}} \\left(\\tilde {\\mathcal {E}} _ {0: i - 1} \\cup \\mathcal {D} _ {i}; \\theta_ {i}, \\omega_ {i}\\right), \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.472, + 0.892, + 0.532 + ], + "angle": 0, + "content": "where \\(\\lambda\\) is the learning rate. We follow the implementation of CIL training loss \\(\\mathcal{L}_{\\mathrm{CIL}}\\) in baseline methods [23, 42, 46, 50]. This means that we use different training losses when plugging CIM into different baseline methods." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.532, + 0.892, + 0.577 + ], + "angle": 0, + "content": "2) Mask-level Optimization. This level aims to optimize the CIM model \\(\\phi_{i}\\) to produce adaptive compression masks. It is formulated as a local BOP:" + }, + { + "type": "equation", + "bbox": [ + 0.521, + 0.588, + 0.892, + 0.612 + ], + "angle": 0, + "content": "\\[\n\\min _ {\\phi_ {i}} \\left[ \\mathcal {L} _ {\\text {v a l}} \\left(\\mathcal {D} _ {i}; \\theta_ {i} ^ {*}, \\omega_ {i}\\right) + \\mu \\mathcal {R} \\left(\\phi_ {i}\\right) \\right], \\tag {6a}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.522, + 0.615, + 0.892, + 0.642 + ], + "angle": 0, + "content": "\\[\n\\text {s . t .} \\theta_ {i} ^ {*} = \\underset {\\theta_ {i}} {\\arg \\min } \\mathcal {L} _ {\\operatorname {t r a i n}} \\left(\\tilde {\\mathcal {E}} _ {0: i - 1} \\cup \\tilde {\\mathcal {D}} _ {i} \\left(\\phi_ {i}\\right); \\theta_ {i}, \\omega_ {i}\\right). \\tag {6b}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.651, + 0.892, + 0.756 + ], + "angle": 0, + "content": "Eq. 6b denotes an inner-level optimization. It trains \\(\\theta_{i}\\) with the data \\(\\tilde{\\mathcal{D}}_i(\\phi_i)\\) compressed by using \\(\\phi_{i}\\), and converges as \\(\\theta_{i}^{*}\\). Eq. 6a denotes an outer-level optimization. It is based on the validation loss derived by \\(\\theta_{i}^{*}\\) on the original data \\(\\mathcal{D}_i\\). \\(\\mathcal{R}(\\phi_i)\\) is a constraint representing the memory limitation and \\(\\mu\\) is its weight. In the following, we elaborate on the implementation details for the two levels." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.757, + 0.892, + 0.848 + ], + "angle": 0, + "content": "In the inner-level optimization, we train a temporary CIL model with compressed data. Specifically, we first compress new-class data \\(\\mathcal{D}_i\\) into \\(\\tilde{\\mathcal{D}}_i(\\phi_i)\\) using the generated masks by the CIM model \\(\\phi_i\\). Then, we implement the inner-level optimization as a one-step gradient descent (using the CIL training loss) as:" + }, + { + "type": "equation", + "bbox": [ + 0.537, + 0.857, + 0.892, + 0.876 + ], + "angle": 0, + "content": "\\[\n\\theta_ {i} ^ {+} \\leftarrow \\theta_ {i} - \\beta_ {1} \\nabla_ {\\theta} \\mathcal {L} _ {\\mathrm {C I L}} \\left(\\tilde {\\mathcal {E}} _ {0: i - 1} \\cup \\tilde {\\mathcal {D}} _ {i} \\left(\\phi_ {i}\\right); \\theta_ {i}, \\omega_ {i}\\right), \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.886, + 0.734, + 0.901 + ], + "angle": 0, + "content": "where \\(\\beta_{1}\\) is the learning rate for \\(\\theta_{i}\\)." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "11375" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.47, + 0.168 + ], + "angle": 0, + "content": "The aim of the outer-level optimization is to optimize \\(\\phi_{i}\\) such that the temporary model \\((\\theta_{i}^{+},\\omega_{i})\\) (trained with compressed data) has a low validation loss on the original data \\(D_{i}\\). To achieve this, we back-propagate the loss on the original data \\(D_{i}\\) to update \\(\\phi_{i}\\) as:" + }, + { + "type": "equation", + "bbox": [ + 0.103, + 0.179, + 0.47, + 0.197 + ], + "angle": 0, + "content": "\\[\n\\phi_ {i} \\leftarrow \\phi_ {i} - \\beta_ {2} \\nabla_ {\\phi} \\left[ \\mathcal {L} _ {\\mathrm {C E}} \\left(\\mathcal {D} _ {i}; \\theta_ {i} ^ {+}, \\omega_ {i}\\right) + \\mu \\mathcal {R} \\left(\\phi_ {i}\\right) \\right], \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.208, + 0.469, + 0.314 + ], + "angle": 0, + "content": "where \\(\\mathcal{L}_{\\mathrm{CE}}\\) denotes softmax cross-entropy loss and \\(\\beta_{2}\\) is the learning rate for \\(\\phi_{i}\\). This trains \\(\\phi_{i}\\) to capture the most discriminative features of new-class images. The constraint \\(\\mathcal{R}(\\phi_i)\\) is implemented as a \\(\\ell_2\\)-regularization term on the generated mask by \\(\\phi_{i}\\). The motivation for the regularization term is to make the mask coverage smaller, thus the compressed images take less memory." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.314, + 0.469, + 0.39 + ], + "angle": 0, + "content": "We empirically observe that by the above optimization flow, the output activation maps by \\(\\phi_{i}\\) are easy to collapse, i.e., different images have the same map. To solve this issue, we add a cross-entropy loss term about \\(\\phi_{i}\\) in Eq. 8 to regularize it to produce image-specific activation maps:" + }, + { + "type": "equation", + "bbox": [ + 0.093, + 0.402, + 0.47, + 0.44 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\phi_ {i} \\leftarrow \\phi_ {i} - \\beta_ {2} \\nabla_ {\\phi} \\left[ \\mathcal {L} _ {\\mathrm {C E}} \\left(\\mathcal {D} _ {i}; \\theta_ {i} ^ {+}, \\omega_ {i}\\right) + \\mu \\mathcal {R} (\\phi_ {i}) \\right. \\tag {9} \\\\ + \\mu^ {\\prime} \\mathcal {L} _ {\\mathrm {C E}} \\left(\\tilde {\\mathcal {E}} _ {0: i - 1} \\cup \\mathcal {D} _ {i}; \\theta_ {i}, \\phi_ {i}, \\omega_ {i}\\right) ], \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.451, + 0.232, + 0.466 + ], + "angle": 0, + "content": "where \\(\\mu^{\\prime}\\) is the weight." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.466, + 0.47, + 0.678 + ], + "angle": 0, + "content": "Limitations. Our CIM learns to generate adaptive masks for exemplar compression in CIL. It has three limitations that are left as future work. 1) It is not able to adjust any previous-phase exemplars, as the validation data (the original data of these exemplars) are not accessible anymore. 2) It introduces hundreds of activation parameters to the CIL model, although this is not a significant overhead compared to model parameters. Please check detailed overhead analyses in the supplementary materials. 3) Image compression is not that meaningful for low-resolution datasets (e.g., \\(32 \\times 32\\) CIFAR-100). It is because the memory taken by the compression parameters (e.g., the parameters of CIM) and the RGB pixels of a low-resolution image are comparable. Using the memory to save more images is more meaningful." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.691, + 0.209, + 0.708 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.717, + 0.469, + 0.793 + ], + "angle": 0, + "content": "We incorporate CIM into two baseline CIL methods (i.e., DER [46] and FOSTER [42]) and boost their model performances consistently on three datasets. Below, we introduce datasets and experiment settings (Section 5.1), followed by results and analyses (Section 5.2)" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.802, + 0.285, + 0.818 + ], + "angle": 0, + "content": "5.1. Experimental Settings" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Datasets. We conduct experiments on three standard CIL benchmarks with high-resolution images. 1) Food-101 [3] consists of 101 food categories with 750 training and 250 test samples per category. All images have a maximum side length of 512 pixels. 2) ImageNet-1000 [9] is a large-scale" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.182 + ], + "angle": 0, + "content": "dataset with 1,000 classes and each class has around 1,300 training and 50 test samples. 3) ImageNet-100 is a 100-class subset randomly sampled from ImageNet-1000 with a fixed NumPy [14] random seed (1993), following [17]. We provide other details of these datasets, e.g., image sizes and pre-processing methods, in the supplementary materials." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.183, + 0.893, + 0.394 + ], + "angle": 0, + "content": "Protocols. We use two protocols: learning from scratch (LFS) and learning from half (LFH), following recent CIL works [42, 46]. In LFS, the model observes the same number of classes in all \\( N \\) phases, where \\( N \\) is optionally 5, 10, and 20. In LFH, the model is trained on half of the classes (e.g., 500 classes for ImageNet-1000) in the 1-th phase. Then, it learns the remaining classes evenly in the subsequent \\( N \\) phases, where \\( N \\) can be 5, 10, and 25. In both protocols, after the training of each phase we evaluate the resultant model on the test data of all seen classes. Our final report includes the average accuracy over all phases and the last-phase accuracy which indicates the degree of model forgetting. We run each experiment three times and report the average results." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.395, + 0.893, + 0.607 + ], + "angle": 0, + "content": "Memory Budget. There are two memory budget settings. 1) In the \"fixed\" setting, we remove some old-class exemplars when new exemplars from the current phase are added in the memory to maintain the \"fixed memory budget\". In this setting, we set the total memory to be 2,020 samples for Food-101 and 2,000 samples for ImageNet-100. For ImageNet-1000, we have two options—5,000 samples and 20,000 samples. 2) In the \"growing\" setting, a constant memory budget is allocated for each class across all phases and hence extra memory is appended when new classes come. In this setting, we set the budget to be 20 samples per class for all datasets. Following [42, 46], we apply the \"fixed\" setting in LFS experiments and the \"growing\" setting in LFH experiments." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.608, + 0.892, + 0.85 + ], + "angle": 0, + "content": "Implementation Details. Our implementation is based on the standard deep learning library PyTorch [35] and image processing library OpenCV [4]. Following [42, 43, 46], we use an 18-layer ResNet [16] as the network backbone \\(\\theta\\) and a fully-connected layer as the classifier \\(\\omega\\) in all experiments. We use the same CIL training hyperparameters as in related works [17, 42, 46] for fair comparison: 1) there are 200 epochs in 1-st phase and 170 epochs in the subsequent phases; 2) the learning rate \\(\\lambda\\) is initialized as 0.1 and decreases to zero with a cosine annealing scheduler [28]; 3) the SGD optimizer is deployed, with momentum factor set to 0.9 and weight decay set to 0.0005. For compression-related hyperparameters, we set the masking threshold \\(\\tau\\) as 0.6 and the downsampling ratio \\(\\eta\\) as 4.0. To build the CIM model, we apply PAUs with degrees \\(m = 5\\) and \\(n = 4\\) as learnable activation layers. For the optimization of the CIM" + }, + { + "type": "page_footnote", + "bbox": [ + 0.499, + 0.863, + 0.892, + 0.901 + ], + "angle": 0, + "content": "5Please note that we measure the memory by the number of original images. Each compressed exemplar in CIM takes less memory than the original image (Eq. 4), resulting in more exemplars in the same memory." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "11376" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.088, + 0.089, + 0.885, + 0.278 + ], + "angle": 0, + "content": "
MethodLearning from Scratch (LFS)Learning from Half (LFH)
Food-101ImageNet-100Food-101ImageNet-100
N=51020510205102551025
iCaRL [36]69.6662.1856.7073.9067.0662.3660.1353.4246.8762.5359.8852.97
WA [48]70.9463.6958.4574.6468.6263.2063.5557.6052.4865.7563.7158.34
PODNet [12]68.0361.2447.3872.1463.9653.6975.3770.0165.3275.5474.3368.33
AANets [25]69.4661.5948.8372.9865.7755.3676.0771.2266.9376.9675.5871.78
DER [46]73.8870.7664.3978.5076.1273.7978.1373.45-79.0877.73-
DER w/ ours75.6373.0969.1779.6377.5775.3679.2575.76-80.3079.05-
FOSTER [42]75.0372.7266.73\\( 79.93^{\\dagger} \\)\\( 76.55^{\\dagger} \\)74.4979.0875.0768.08\\( 80.07^{\\dagger} \\)77.54\\( 72.40^{\\times} \\)
FOSTER w/ ours76.4474.8570.2080.5877.9475.2379.7676.8670.5080.9378.6675.74
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.099, + 0.279, + 0.882, + 0.308 + ], + "angle": 0, + "content": "† The paper of FOSTER [42] did not report the numerical results for \\( N = 5 / 10 \\) (LFS) and \\( N = 5 \\) (LFH) on ImageNet-100. We run these experiments using the public code (released by authors) and report the reproduced results." + }, + { + "type": "table_footnote", + "bbox": [ + 0.099, + 0.308, + 0.807, + 0.323 + ], + "angle": 0, + "content": "* Our reproduced result (72.40) is significantly higher than the original result (69.34) reported in the paper of FOSTER." + }, + { + "type": "list", + "bbox": [ + 0.099, + 0.279, + 0.882, + 0.323 + ], + "angle": 0, + "content": null + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.327, + 0.895, + 0.37 + ], + "angle": 0, + "content": "Table 1. Average accuracies \\((\\%)\\) of two top-performing CIL methods [42, 46] with and without our CIM-CIL plugged-in, and other four baselines [12, 25, 36, 48], on two datasets (Food-101 and ImageNet-100) and using two protocols (learning from scratch (LFS) and learning from half (LFH)). Due to the space limits, we report the \\(95\\%\\) confidence intervals for these results in the supplementary materials." + }, + { + "type": "table", + "bbox": [ + 0.08, + 0.379, + 0.47, + 0.533 + ], + "angle": 0, + "content": "
Memory BudgetMethodN=5N=10
Avg.LastAvg.Last
M=20kiCaRL [36]44.3627.7838.4022.70
WA [48]58.3750.6254.1045.66
DER [46]67.4959.7566.7358.62
FOSTER [42]69.2164.8868.3460.14
FOSTER w/ ours69.9366.0569.5362.07
M=5kFOSTER57.1949.4254.7244.96
FOSTER w/ ours61.3754.4659.4850.83
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.54, + 0.472, + 0.598 + ], + "angle": 0, + "content": "Table 2. Average and last accuracies \\((\\%)\\) on ImageNet-1000 of FOSTER with and without our method plugged-in, and other three baselines [36, 46, 48]. We show two memory budgets, \\(M = 20k\\) (upper block) and \\(M = 5k\\) (lower block), in the LFS setting." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.609, + 0.47, + 0.702 + ], + "angle": 0, + "content": "model \\(\\phi\\) (i.e., the mask-level optimization), we initially set \\(\\beta_{1}\\) as 0.1 and \\(\\beta_{2}\\) as 0.01 and reduce them to zero following the scheduler of \\(\\lambda\\). \\(\\mu\\) and \\(\\mu^{\\prime}\\) is set to 0.1 and 0.2, respectively. To smooth the training, we clip the gradient norm of \\(\\phi\\) to be no more than 1. We report the result of hyperparameter sensitivity analysis in the supplementary materials." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.711, + 0.276, + 0.728 + ], + "angle": 0, + "content": "5.2. Results and Analyses" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.735, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Comparing with the State-of-the-art. In Table 1, we summarize the experimental results on two datasets (Food-101 and ImageNet-100) and in two CIL protocols (LFS and LFH). From the table, we have the following observations. 1) Our CIM-based CIL consistently improves the state-ofthe-art method FOSTER [42] with clear margins in all settings. E.g., our method surpasses it by an average of 1.4 percentage points on ImageNet-100, and 2.0 percentage points on Food-101. 2) Our CIM-based CIL achieves more significant improvements when \\(N\\) becomes larger, e.g., on ImageNet-100 (LFH), our method improves FOSTER by" + }, + { + "type": "table", + "bbox": [ + 0.513, + 0.379, + 0.885, + 0.599 + ], + "angle": 0, + "content": "
Ablation MethodFood-101ImageNet-100
N=10201020
1 Baseline72.7266.7376.5572.37
2 Artifact Aug.71.3866.0375.6371.45
3 Full Comp.73.0367.3876.9273.26
4 Random Acti.73.1067.5476.8873.54
5 Center Acti.73.2967.8876.7873.82
6 Class Acti.73.7668.6577.2174.67
7 Phase-wise τ73.8369.1777.0674.78
8 Joint Train73.4469.0177.3474.59
9 BOP (ours)74.8570.2077.9475.23
10 LastBlock Only74.5569.8777.7274.86
11 Fg Compressed75.0270.1377.8775.46
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.603, + 0.892, + 0.633 + ], + "angle": 0, + "content": "Table 3. Average accuracies \\((\\%)\\) of different ablation methods. The experiments are conducted in the LFS setting." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.644, + 0.892, + 0.765 + ], + "angle": 0, + "content": "0.9 and 3.3 percentage points when \\( N = 5 \\) and \\( N = 25 \\), respectively. 3) Our CIM-based CIL achieves greater improvements consistently on Food-101 (than ImageNet-100). It improves baselines by 2.1 percentage points on Food-101, while the improvement is 1.4 on ImageNet-100 (\\( N = 10 \\), LFS). It shows that our method is particularly effective when the representative visual cues of a class are from some of its components, e.g., the \"cream\" of the class \"cake\"." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.892, + 0.886 + ], + "angle": 0, + "content": "Table 2 shows the results on the large-scale dataset ImageNet-1000 in different memory settings (\\(M = 20k\\) and \\(M = 5k\\)). We can see that our CIM-based CIL improves FOSTER consistently. It is impressive that it achieves more improvements in the more strict memory setting (\\(M = 5k\\)). Specifically, it boosts the average accuracy of FOSTER by 4.5 percentage points when \\(M = 5k\\), significantly higher than that of \\(M = 20k\\) (1.0)." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.886, + 0.892, + 0.902 + ], + "angle": 0, + "content": "Ablation Study. Table 3 shows the ablation results. First" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "11377" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.084, + 0.089, + 0.468, + 0.204 + ], + "angle": 0, + "content": "
MethodImageNet-100ImageNet-1000
N=6112661126
LUCIR baseline71.2269.6767.4565.2362.4359.88
w/ Mnemonics73.3072.1771.5066.1563.1263.08
w/ MRDC73.6272.8170.4467.6765.6062.74
w/ ours74.0573.7672.8468.0366.5463.77
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.077, + 0.207, + 0.47, + 0.237 + ], + "angle": 0, + "content": "Table 4. Comparing with Mnemonics [27] and MRDC [43]. We plug each of them in baseline LUCIR [17] for fair comparison." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.249, + 0.473, + 0.687 + ], + "angle": 0, + "content": "block: baselines. Row 1 is for the baseline FOSTER [42]. Row 2 shows the results of adding artifact augmentation (see Section 4.1). It shows directly apply this augmentation does not improve and even impair the model. Please note the models in below blocks all use this augmentation. Second block: activation methods. Rows 3-6 show the results of using different activation methods to compress exemplars. Row 3 is to downsample all pixels (i.e., no region is activated). Row 4 is to randomly select activation regions. Row 5 is to activate only the center region \\((\\frac{1}{4}\\) of the original image), while row 6 is to use naive CAM. Comparing them to row 1, we can see that using naively compressed exemplars can improve CIL models. Row 4 outperforms rows 3-5, validating that it is more reliable to use the model's activation to generate compressed exemplars. Third block: optimization methods. Rows 7-9 are on top of Row 6 and are the results of applying different optimization strategies. Row 7 is to manually select \\(\\tau\\) using a held-out set (\\(10\\%\\) of the dataset). Row 8 is to jointly train CIL and CIM models (for each input batch). Row 9 is the proposed method of using a global BOP. Fourth block: two variants of CIM-based CIL. Rows 10-11 are two variants of row 9. In Row 10, only the activation layers in the last block of CIM are learnable, and previous blocks use ReLU. Compared to row 9, row 10 shows slightly worse performance. Row 11 shows the version of adding a weak downsampling \\((\\eta' = 2.0)\\) on discriminative regions, based on which more compressed exemplars are saved. It results in comparable performance to row 9 but increases costs." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.688, + 0.47, + 0.854 + ], + "angle": 0, + "content": "Comparing with Other Compression-based Methods. Table 4 shows our results comparing to two compression-based methods: Mnemonics [27] and MRDC [43]. We can see that our method consistently outperforms them in all settings. This is because our method does not sacrifice the discriminativeness of exemplars while improving the number (variance) of exemplars in a phase-adaptive manner. However, the two related methods either keep a fixed number of exemplars in the memory [27] or use uniform image compression without considering the properties of specific classes in different incremental phases [43]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Visualizations (CAM vs. CIM). Figure 4 gives two visualization examples, \"Afghan hound\" and \"indigo bird\", each with the activation map as well as the bounding boxes." + }, + { + "type": "image", + "bbox": [ + 0.533, + 0.089, + 0.859, + 0.24 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.244, + 0.892, + 0.273 + ], + "angle": 0, + "content": "Figure 4. Visualizations (CAM vs. CIM). The experimental setting is \\( N = 10 \\) (LFS) on ImageNet-100." + }, + { + "type": "table", + "bbox": [ + 0.501, + 0.28, + 0.886, + 0.372 + ], + "angle": 0, + "content": "
MetricSmallMiddleLarge
Mean of #Exemplars39.4038.3034.77
Last Acc. (%, baseline)66.1368.4069.93
Last Acc. (%, ours)70.0071.1072.26
Improvement (%)+3.87+3.65+2.33
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.499, + 0.375, + 0.893, + 0.432 + ], + "angle": 0, + "content": "Table 5. Results for small, middle and large objects in the setting of \\( N = 10 \\) (LFS) on ImageNet-100. \"Mean of #Exemplars\" denotes the average number of saved exemplars by our method. The baseline (FOSTER [42]) has this number as 20 for all classes." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.445, + 0.892, + 0.49 + ], + "angle": 0, + "content": "The first column shows their respective confusing classes appearing in earlier phases. CIM learns to focus on the discriminative (i.e., dissimilar to confusing classes) regions." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.49, + 0.893, + 0.657 + ], + "angle": 0, + "content": "Results of Different-Size Objects. Table 5 shows the results for small, middle, and large objects. These size categorization is according to ImageNet Object Localization Challenge [19]. We calculate the bbox coverage for each class and take top 30 classes with highest coverages as \"large\", near 30 classes with lowest coverages as \"small\" and the rest 40 classes as \"middle\". It is intriguing that our method achieves the highest improvement (over baseline) for small objects. Our explanation is that small objects benefit more from image compression (than large ones), as their images contain more background pixels to downsample." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.671, + 0.628, + 0.687 + ], + "angle": 0, + "content": "6. Conclusions" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.697, + 0.892, + 0.85 + ], + "angle": 0, + "content": "We introduced a novel exemplar compression method for CIL, allowing us to save more representative exemplars but not increase memory budget. We achieved this compression by downsampling non-discriminative pixels of the image bounded by CAM masks. To generate adaptive masks, we proposed a novel method CIM that explicitly parameterizes a mask generation model and optimizes it in an end-to-end manner across incremental phases. Our method achieves consistent performance improvements over multiple baselines and can be taken as a flexible plug-and-play module." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.856, + 0.893, + 0.902 + ], + "angle": 0, + "content": "Acknowledgements. The author gratefully acknowledges the support by the Lee Kong Chian (LKC) Fellowship fund awarded by Singapore Management University." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "11378" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.175, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.17 + ], + "angle": 0, + "content": "[1] Davide Abati, Jakub Tomczak, Tijmen Blankevoort, Simone Calderara, Rita Cucchiara, and Babak Ehteshami Bejnordi. Conditional channel gated networks for task-aware continual learning. In CVPR, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.171, + 0.472, + 0.213 + ], + "angle": 0, + "content": "[2] Alexey Bochkovskiy, Chien-Yao Wang, and Hong-Yuan Mark Liao. Yolov4: Optimal speed and accuracy of object detection. arXiv, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.213, + 0.47, + 0.255 + ], + "angle": 0, + "content": "[3] Lukas Bossard, Matthieu Guillaumin, and Luc Van Gool. Food-101 - mining discriminative components with random forests. In ECCV, 2014. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.256, + 0.47, + 0.283 + ], + "angle": 0, + "content": "[4] Gary Bradski. The opencv library. Dr. Dobb's Journal: Software Tools for the Professional Programmer, 2000. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.283, + 0.47, + 0.311 + ], + "angle": 0, + "content": "[5] Kenneth R Castleman. Digital image processing. Prentice Hall Press, 1996. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.312, + 0.47, + 0.354 + ], + "angle": 0, + "content": "[6] George Cazenavette, Tongzhou Wang, Antonio Torralba, Alexei A Efros, and Jun-Yan Zhu. Dataset distillation by matching training trajectories. In CVPR, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.354, + 0.47, + 0.396 + ], + "angle": 0, + "content": "[7] Can Chen, Xi Chen, Chen Ma, Zixuan Liu, and Xue Liu. Gradient-based bi-level optimization for deep learning: A survey. arXiv, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.397, + 0.47, + 0.452 + ], + "angle": 0, + "content": "[8] Zhaozheng Chen, Tan Wang, Xiongwei Wu, Xian-Sheng Hua, Hanwang Zhang, and Qianru Sun. Class re-activation maps for weakly-supervised semantic segmentation. In CVPR, 2022. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.452, + 0.47, + 0.493 + ], + "angle": 0, + "content": "[9] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In CVPR, 2009. 1, 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.494, + 0.47, + 0.536 + ], + "angle": 0, + "content": "[10] Prithviraj Dhar, Rajat Vikram Singh, Kuan-Chuan Peng, Ziyan Wu, and Rama Chellappa. Learning without memorizing. In CVPR, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.536, + 0.47, + 0.579 + ], + "angle": 0, + "content": "[11] Zhang Dong, Zhang Hanwang, Tang Jinhui, Hua Xiansheng, and Sun Qianru. Causal intervention for weakly supervised semantic segmentation. In NeurIPS, 2020. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.579, + 0.47, + 0.633 + ], + "angle": 0, + "content": "[12] Arthur Douillard, Matthieu Cord, Charles Ollion, Thomas Robert, and Eduardo Valle. Podnet: Pooled outputs distillation for small-tasks incremental learning. In ECCV, 2020. 1, 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.634, + 0.47, + 0.676 + ], + "angle": 0, + "content": "[13] Chelsea Finn, Pieter Abbeel, and Sergey Levine. Model-agnostic meta-learning for fast adaptation of deep networks. In ICML, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.677, + 0.47, + 0.733 + ], + "angle": 0, + "content": "[14] Charles R Harris, K Jarrod Millman, Stefan J Van Der Walt, Ralf Gommers, Pauli Virtanen, David Cournapeau, Eric Wieser, Julian Taylor, Sebastian Berg, Nathaniel J Smith, et al. Array programming with numpy. Nature, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.733, + 0.47, + 0.774 + ], + "angle": 0, + "content": "[15] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Delving deep into rectifiers: Surpassing human-level performance on imagenet classification. In ICCV, 2015. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.774, + 0.47, + 0.816 + ], + "angle": 0, + "content": "[16] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, 2016. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.816, + 0.47, + 0.859 + ], + "angle": 0, + "content": "[17] Saihui Hou, Xinyu Pan, Chen Change Loy, Zilei Wang, and Dahua Lin. Learning a unified classifier incrementally via rebalancing. In CVPR, 2019. 1, 2, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.859, + 0.47, + 0.901 + ], + "angle": 0, + "content": "[18] Shenyang Huang, Vincent François-Lavet, and Guillaume Rabusseau. Neural architecture search for class-incremental learning. arXiv, 2019. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.134 + ], + "angle": 0, + "content": "[19] Imagenet object localization challenge. https://www.kaggle.com/competitions/imagenet-object-localization-challenge/.8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.136, + 0.892, + 0.163 + ], + "angle": 0, + "content": "[20] Anil K Jain. Fundamentals of digital image processing. Prentice-Hall, Inc., 1989. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.164, + 0.892, + 0.233 + ], + "angle": 0, + "content": "[21] James Kirkpatrick, Razvan Pascanu, Neil Rabinowitz, Joel Veness, Guillaume Desjardins, Andrei A Rusu, Kieran Milan, John Quan, Tiago Ramalho, Agnieszka Grabska-Barwinska, et al. Overcoming catastrophic forgetting in neural networks. PNAS, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.234, + 0.892, + 0.276 + ], + "angle": 0, + "content": "[22] Jungbeom Lee, Eunji Kim, and Sungroh Yoon. Antiadversarily manipulated attributions for weakly and semi-supervised semantic segmentation. In CVPR, 2021. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.277, + 0.892, + 0.304 + ], + "angle": 0, + "content": "[23] Zhizhong Li and Derek Hoiem. Learning without forgetting. PAMI, 2017. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.305, + 0.892, + 0.348 + ], + "angle": 0, + "content": "[24] Yaoyao Liu, Yingying Li, Bernt Schiele, and Qianru Sun. Online hyperparameter optimization for class-incremental learning. In AAAi, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.348, + 0.892, + 0.39 + ], + "angle": 0, + "content": "[25] Yaoyao Liu, Bernt Schiele, and Qianru Sun. Adaptive aggregation networks for class-incremental learning. In CVPR, 2021. 1, 2, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.391, + 0.892, + 0.432 + ], + "angle": 0, + "content": "[26] Yaoyao Liu, Bernt Schiele, and Qianru Sun. Rmm: Reinforced memory management for class-incremental learning. In NeurIPS, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.433, + 0.892, + 0.476 + ], + "angle": 0, + "content": "[27] Yaoyao Liu, Yuting Su, An-An Liu, Bernt Schiele, and Qianru Sun. Mnemonics training: Multi-class incremental learning without forgetting. In CVPR, 2020. 1, 2, 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.476, + 0.892, + 0.504 + ], + "angle": 0, + "content": "[28] Ilya Loshchilov and Frank Hutter. Sgdr: Stochastic gradient descent with warm restarts. arXiv, 2016. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.505, + 0.892, + 0.547 + ], + "angle": 0, + "content": "[29] Dougal Maclaurin, David Duvenaud, and Ryan Adams. Gradient-based hyperparameter optimization through reversible learning. In ICML, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.547, + 0.892, + 0.603 + ], + "angle": 0, + "content": "[30] Michael McCloskey and Neal J Cohen. Catastrophic interference in connectionist networks: The sequential learning problem. In *Psychology of Learning and Motivation*. Elsevier, 1989. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.604, + 0.892, + 0.659 + ], + "angle": 0, + "content": "[31] Ken McRae and Phil A Hetherington. Catastrophic interference is eliminated in pretrained networks. In Proceedings of the 15th Annual Conference of the Cognitive Science Society, 1993. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.66, + 0.892, + 0.703 + ], + "angle": 0, + "content": "[32] Alejandro Molina, Patrick Schramowski, and Kristian Kersting. Padé activation units: End-to-end learning of flexible activation functions in deep networks. In ICLR, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.703, + 0.892, + 0.732 + ], + "angle": 0, + "content": "[33] Vinod Nair and Geoffrey E Hinton. Rectified linear units improve restricted boltzmann machines. In ICML, 2010. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.732, + 0.892, + 0.774 + ], + "angle": 0, + "content": "[34] J Anthony Parker, Robert V Kenyon, and Donald E Troxel. Comparison of interpolating methods for image resampling. IEEE Transactions on Medical Imaging, 1983. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.774, + 0.892, + 0.844 + ], + "angle": 0, + "content": "[35] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. In NeurIPS, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.845, + 0.892, + 0.899 + ], + "angle": 0, + "content": "[36] Sylvestre-Alvise Rebuffi, Alexander Kolesnikov, Georg Sperl, and Christoph H Lampert. iCaRL: Incremental classifier and representation learning. In CVPR, 2017. 1, 2, 3, 5, 7" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "11379" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.147 + ], + "angle": 0, + "content": "[37] Andrei A Rusu, Neil C Rabinowitz, Guillaume Desjardins, Hubert Soyer, James Kirkpatrick, Koray Kavukcuoglu, Razvan Pascanu, and Raia Hadsell. Progressive neural networks. arXiv, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.149, + 0.468, + 0.205 + ], + "angle": 0, + "content": "[38] Ramprasaath R Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra. Grad-cam: Visual explanations from deep networks via gradient-based localization. In ICCV, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.207, + 0.468, + 0.247 + ], + "angle": 0, + "content": "[39] Christian Simon, Piotr Koniusz, and Mehrtash Harandi. On learning the geodesic path for incremental learning. In CVPR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.249, + 0.468, + 0.303 + ], + "angle": 0, + "content": "[40] Ankur Sinha, Pekka Malo, and Kalyanmoy Deb. A review on bilevel optimization: from classical to evolutionary approaches and applications. IEEE Transactions on Evolutionary Computation, 2017. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.305, + 0.468, + 0.333 + ], + "angle": 0, + "content": "[41] Gregory K Wallace. The jpeg still picture compression standard. Communications of the ACM, 1991. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.335, + 0.468, + 0.375 + ], + "angle": 0, + "content": "[42] Fu-Yun Wang, Da-Wei Zhou, Han-Jia Ye, and De-Chuan Zhan. Foster: Feature boosting and compression for class incremental learning. arXiv, 2022. 1, 2, 3, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.377, + 0.468, + 0.432 + ], + "angle": 0, + "content": "[43] Liyuan Wang, Xingxing Zhang, Kuo Yang, Longhui Yu, Chongxuan Li, Lanqing Hong, Shifeng Zhang, Zhenguo Li, Yi Zhong, and Jun Zhu. Memory replay with data compression for continual learning. In ICLR, 2022. 1, 2, 3, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.434, + 0.468, + 0.475 + ], + "angle": 0, + "content": "[44] Yue Wu, Yinpeng Chen, Lijuan Wang, Yuancheng Ye, Zicheng Liu, Yandong Guo, and Yun Fu. Large scale incremental learning. In CVPR, 2019. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.477, + 0.468, + 0.504 + ], + "angle": 0, + "content": "[45] Ju Xu and Zhanxing Zhu. Reinforced continual learning. In NeurIPS, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.505, + 0.468, + 0.546 + ], + "angle": 0, + "content": "[46] Shipeng Yan, Jiangwei Xie, and Xuming He. Der: Dynamically expandable representation for class incremental learning. In CVPR, 2021. 1, 2, 3, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.548, + 0.468, + 0.587 + ], + "angle": 0, + "content": "[47] Friedemann Zenke, Ben Poole, and Surya Ganguli. Continual learning through synaptic intelligence. In ICML, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.591, + 0.468, + 0.632 + ], + "angle": 0, + "content": "[48] Bowen Zhao, Xi Xiao, Guojun Gan, Bin Zhang, and Shu-Tao Xia. Maintaining discrimination and fairness in class incremental learning. In CVPR, 2020. 1, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.634, + 0.468, + 0.674 + ], + "angle": 0, + "content": "[49] Bolei Zhou, Aditya Khosla, Agata Lapedriza, Aude Oliva, and Antonio Torralba. Learning deep features for discriminative localization. In CVPR, 2016. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.676, + 0.468, + 0.73 + ], + "angle": 0, + "content": "[50] Da-Wei Zhou, Han-Jia Ye, and De-Chuan Zhan. Co-transport for class-incremental learning. In Proceedings of the 29th ACM International Conference on Multimedia, 2021. 5" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.73 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.956 + ], + "angle": 0, + "content": "11380" + } + ] +] \ No newline at end of file diff --git a/2023/Class-Incremental Exemplar Compression for Class-Incremental Learning/47061da4-a9f4-49a6-b036-8d247f22c620_origin.pdf b/2023/Class-Incremental Exemplar Compression for Class-Incremental Learning/47061da4-a9f4-49a6-b036-8d247f22c620_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..89f1367f86a9a6c50956f3a740bd47b860e393f5 --- /dev/null +++ b/2023/Class-Incremental Exemplar Compression for Class-Incremental Learning/47061da4-a9f4-49a6-b036-8d247f22c620_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:853119b2acd0af161d2be7b67ae7e59bc52773c1ec860b49695a94490b79da21 +size 1058336 diff --git a/2023/Class-Incremental Exemplar Compression for Class-Incremental Learning/full.md b/2023/Class-Incremental Exemplar Compression for Class-Incremental Learning/full.md new file mode 100644 index 0000000000000000000000000000000000000000..8ab9f2654ca9272a94f7fdf129e1ab6bcab73f74 --- /dev/null +++ b/2023/Class-Incremental Exemplar Compression for Class-Incremental Learning/full.md @@ -0,0 +1,310 @@ +# Class-Incremental Exemplar Compression for Class-Incremental Learning + +Zilin Luo1 Yaoyao Liu2 Bernt Schiele2 Qianru Sun1 + +$^{1}$ Singapore Management University + +$^{2}$ Max Planck Institute for Informatics, Saarland Informatics Campus + +zilin.luo.2021@phdcs.smu.edu.sg {yaoyao.liu, schiele}@mpi-inf.mpg.de qianrusun@smu.edu.sg + +# Abstract + +Exemplar-based class-incremental learning (CIL) [36] finetunes the model with all samples of new classes but few-shot exemplars of old classes in each incremental phase, where the "few-shot" abides by the limited memory budget. In this paper, we break this "few-shot" limit based on a simple yet surprisingly effective idea: compressing exemplars by downsampling non-discriminative pixels and saving "many-shot" compressed exemplars in the memory. Without needing any manual annotation, we achieve this compression by generating 0-1 masks on discriminative pixels from class activation maps (CAM) [49]. We propose an adaptive mask generation model called class-incremental masking (CIM) to explicitly resolve two difficulties of using CAM: 1) transforming the heatmaps of CAM to 0-1 masks with an arbitrary threshold leads to a trade-off between the coverage on discriminative pixels and the quantity of exemplars, as the total memory is fixed; and 2) optimal thresholds vary for different object classes, which is particularly obvious in the dynamic environment of CIL. We optimize the CIM model alternatively with the conventional CIL model through a bilevel optimization problem [40]. We conduct extensive experiments on high-resolution CIL benchmarks including Food-101, ImageNet-100, and ImageNet-1000, and show that using the compressed exemplars by CIM can achieve a new state-of-the-art CIL accuracy, e.g., 4.8 percentage points higher than FOSTER [42] on 10-Phase ImageNet-1000. Our code is available at https://github.com/xffzl/CIM-CIL. + +# 1. Introduction + +Dynamic AI systems have a continual learning nature to learn new class data. They are expected to adapt to new classes while maintaining the knowledge of old classes, i.e., free from forgetting problems [31]. To evaluate this, the following protocol of class-incremental learning (CIL) was proposed by Rebuffi et al. [36]. The model training goes + +![](images/7db32eb0839a7df78ee3689fff037f1a62dfd758d14530fd142bcae94fe50c05.jpg) +(a) iCaRL [baseline] + +![](images/abc4554419cc1d57e763cc862d6c8d5d7a4b70f5f5456fe93318f05708e3e8f9.jpg) +(c) MRDC [related] + +![](images/bf45685998f0ace0f9dd19d4778ae5140a66001ec4521222a428c128eda68dd4.jpg) +Figure 1. The phase-wise training data in different methods. (a) iCaRL [36] is the baseline method using full new class data and few-shot old class exemplars. (b) Mnemonics [27] distills all training samples into few-shot exemplars without increasing their quantity. (c) MRDC [43] compresses each exemplar uniformly into a low-resolution image using JPEG [41]. (d) Our approach based on the proposed class-incremental masking (CIM) down-samples only non-discriminative pixels in the image. The legend shows the symbols of special images generated by the methods. + +through a number of phases. Each phase has new class data added and old class data discarded, and the resultant model is evaluated on the test data of all seen classes. A straightforward way to retain old class knowledge is keeping around a few old class exemplars in the memory and using them to re-train the model in subsequent phases. The number of exemplars is usually limited, e.g., $5 \sim 20$ exemplars per class [12, 17, 25, 27, 36, 42, 44, 46, 48], as the total memory in CIL strictly budgeted, e.g., $2k$ exemplars. + +This leads to a serious data imbalance between old and new classes, e.g., 20 per old class vs. $1.3k$ per new class (on ImageNet-1000 [9]), as illustrated in Figure 1a. The training is thus always dominated by new classes, and forgetting problems occur for old classes. Liu et al. [27] tried to mitigate this problem by parameterizing and distilling the exem + +plars, without increasing the number of them (Figure 1b). Wang et al. [43] traded off between the quality and quantity of exemplars by uniformly compressing exemplar images with JPEG [41] (Figure 1c). As shown in Figure 1d, our approach is also based on image compression. The idea is to downsample only non-discriminative pixels (e.g., background) and keep discriminative pixels (i.e., representative cues of foreground objects) as the original. In this way, we do not sacrifice the discriminativeness of exemplars when increasing their quantity. In particular, we aim for adaptive compression in dynamic environments of CIL, where the intuition is later phases need to be more conservative (i.e., less downsampling) as the model needs more visual cues to classify the increased number of classes. + +To achieve selective and adaptive compression, we need the location labels of discriminative pixels. Without extra labeling, we automatically generate the labels by utilizing the model's own "attention" on discriminative features, i.e., class activation maps (CAM) [49]. We take this method as a feasible baseline, and based on it, we propose an adaptive version called class-incremental masking (CIM). Specifically, for each input image (with its class label), we use its feature maps and classifier weights (corresponding to its class label) to compute a CAM by channel-wise multiplication, aggregation, and normalization. Then, we apply hard thresholding to generate a 0-1 mask. We notice that when generating the masks in the dynamic environments of CIL, the optimal hyperparameters (such as the value of hard threshold and the choice of activation functions) vary for different classes as well as in different incremental phases. Our adaptive version CIM tackles this by parameterizing a mask generation model and optimizing it in an end-to-end manner across all incremental phases. In each phase, the learned CIM model adaptively generates class-and phase-specific masks. We find that the compressed exemplars based on these masks have stronger representativeness, compared to using the conventional CAM. + +Technically, we have two models to optimize, i.e., the CIL model and the CIM model. These two cannot be optimized separately as they are dependent on computation: 1) the CIM model compresses exemplars to input into the CIL model; 2) the two models share network parameters. We exploit a global bilevel optimization problem (BOP) [7, 40] to alternate their training processes at two levels. This BOP goes through all incremental training phases. In particular, for each phase, we perform a local BOP with two steps to tune the parameters of the CIM model: 1) a temporary model is trained with the compressed exemplars as input; and 2) a validation loss on the uncom + +pressed new data is computed and the gradients are backpropagated to optimize the parameters of CIM. To evaluate CIM, we conduct extensive experiments by plugging it in recent CIL methods, $^{3}$ LUCIR [17], DER [46], and FOSTER [42], on three high-resolution benchmarks, Food101 [3], ImageNet-100 [17], and ImageNet-1000 [9]. We find that using the compressed exemplars by CIM brings consistent and significant improvements, e.g., $4.2\%$ and $4.8\%$ higher than the SOTA method FOSTER [42], respectively, in the 5-phase and 10-phase settings of ImageNet1000, with a total memory budget for $5k$ exemplars. + +# 2. Related Work + +Class-Incremental Learning (CIL). There are three main lines of work to address the catastrophic forgetting problem [30, 31] in CIL. Regularization-based methods apply discrepancy (between old and new models) penalization terms in their objective functions, e.g., by comparing output logits [23, 36], intermediate features [12, 17, 24, 39], and prediction heatmaps [10]. Parameter-isolation-based methods increase the model parameters in each new incremental phase, to prevent knowledge forgetting caused by parameter overwritten. Some of them [18, 37, 42, 45, 46] proposed to progressively expand the size of the neural network to learn new coming data. Others [1, 21, 25, 47] froze a part of network parameters (to maintain the old class knowledge) to alleviate the problem of knowledge overwriting. Replay-based methods assume there is a clear memory budget allowing a handful of old-class exemplars in the memory. Exemplars can be used to re-train the model in each new phase [12, 17, 26, 27, 36, 43, 44]. This re-training usually contains two steps: one step of training the model on all new class data and old class exemplars, and one step of finetuning the model with a balanced subset (i.e., using an equal number of samples per class) [12, 17, 25, 26, 46]. + +The replay-based methods focusing on memory optimization [27, 43] are closely related to our work. [27] proposed a bilevel optimization framework to distill the current new class data into exemplars before discarding them. It aims to improve the quality of exemplars without increasing the quantity. Another work [43] aimed to trade-off between the quality and quantity of exemplars by image compression using the JPEG algorithm, i.e., each exemplar is uniformly downsampled. Ours differs from these two works in three aspects. 1) Our CIM based image compression automatically segments the discriminative pixels in the exemplar and downsamples the non-discriminative pixels only. It barely weakens the representativeness of the exemplar. In contrast, the parameterization of image pixels in [27] hampers the model from capturing high-frequency (discriminative) features from the image, especially the high-resolution (e.g., + +$224 \times 224)$ image [6]. 2) Our approach increases the diversity (i.e., quantity) of old class exemplars by reducing the memory consumption for each exemplar. In contrast, [27] keeps a fixed number of exemplars in the memory. 3) Our approach has an adaptive image compression strategy that fits well in the dynamic environments of CIL. In contrast, [43] uses uniform image compression (or uniformly increasing quality parameters from 100 to 1) without considering the properties of specific classes in each learning phase. + +Class Activation Map (CAM) [49] is a simple yet effective weakly-supervised object localization method. Its model is trained with only the image-level label and can generate pixel-level masks on foreground objects. Specifically, the masks are the results of hard-thresholding the heatmaps produced by feature maps and classifier weights. Advanced CAM variants include Grad-CAM [38], ReCAM [8], AdvCAM [22], etc. Our CIM is based on the vanilla CAM because it is computationally simple and efficient. + +Bilevel Optimization Problems (BOP) [7, 40] aims to solve a nested optimization problem, where the outer-level optimization is subjected to the result of the inner-level optimization. It has shown effectiveness in a wide range of machine learning areas, such as hyperparameter selection [29] and meta-learning [13]. For tackling CIL tasks, [27] leverages BOP to alternatively optimize the parameters of the CIL model and the parameterized exemplars. [25] applies BOP to learn the aggregation weights of the plastic and elastic branches in the CIL model. In our work, we use BOP to solve the optimization of the CIL model and the parameterized Class-Incremental Masking (CIM) model, where CIM is a plugin branch (in the CIL model), using few extra parameters. The process of BOP is quick and efficient. + +# 3. Preliminary + +The following is the training pipeline of standard CIL with few-shot exemplars. Assume there are $N$ learning phases. In the 1-st phase, we load data $\mathcal{D}_1$ containing all training samples of $c_1$ classes, and use $\mathcal{D}_1$ to train the initial classification model $(\theta_{1},\omega_{1})$ , where $\theta_{1}$ and $\omega_{1}$ denote the parameters of the feature extractor and classifier, respectively. When the training is done, we evaluate the model performance on the test samples of $c_1$ classes. Before the 2nd phase, we discard most of the training samples due to the strict memory budget of CIL. In other words, we preserve only a handful of training samples $\mathcal{E}_1$ (i.e., exemplars) in the memory, selected from $\mathcal{D}_1$ . A common method for selecting exemplars is called feature herding [36] and has been used in many related works [27, 42, 43, 46]. We adopt it, too, in this work. + +In the $i$ -th phase $(i \geq 2)$ , we load all exemplars $\mathcal{E}_{1:i-1} = \mathcal{E}_1 \cup \dots \cup \mathcal{E}_{i-1}$ from the memory and initialize the current model $(\theta_i, \omega_i)$ by the previous model $(\theta_{i-1}, \omega_{i-1})$ . We use $\mathcal{E}_{1:i-1}$ and the new coming data $\mathcal{D}_i$ (containing $c_i$ + +new classes) to train $(\theta_{i},\omega_{i})$ . Then, we evaluate the current model using a test set of all $\sum_{j = 1}^{i}c_{j}$ classes seen so far. After that, we discard most of the training samples in $\mathcal{D}_i$ and leave few-shot exemplars $\varepsilon_{i}$ in the memory. It is clear that this discarding causes a strong data imbalance between old and new coming classes in the subsequent phase. In the following, we introduce our solution to this problem. + +# 4. Methodology + +As illustrated in Figure 1d, we alleviate the data imbalance problem by saving a larger number of compressed exemplars for old classes, where we leverage pixel-selective compression, i.e., downsample only non-discriminative pixels. To achieve this, the ideal case is that we have the pixel-level localization of foreground objects. However, the realistic case is that such localization labels are expensive, most CIL benchmarks do not have the labels, and it is not fair to compare with other CIL methods if using the labels. Without extra labeling, we introduce a CAM-based mask generation method, and based on it, we provide a baseline solution to pixel-selective compression in Section 4.1. The problem of mask generation in CIL is that the optimal generation hyperparameters such as hard thresholds are changing in the dynamic environment (with the increasing number of classes and phases). It is thus desirable to have an adaptive mask generation process. To this end, we propose class-incremental masking (CIM)—a learnable mask generation model, in Section 4.2. + +# 4.1. CAM-based Compression Pipeline + +Generating pixel-level labels for large-scale datasets, e.g., ImageNet [9], is non-trivial. Using class activation maps (CAM) is a naive solution with little computation costs. Its key idea is to make use of the activation of the classification model itself: on the feature maps, activated pixels are more discriminative than non-activated ones for recognizing the object, where "activated" means of high activation values and with strong correlation with the classification weights of the object. After localizing activated pixels by CAM, we can generate a 0-1 mask on them, e.g., by hard thresholding their normalized values, and then up-sample the mask to the size of the input image [49]. + +From CAM to 0-1 Mask. We extract CAM in the following steps. Given an image $x$ from $\mathcal{D}_i$ and its ground truth class label $y$ , let $F(x; \theta_i)$ represent the feature block output by the feature extractor $\theta_i$ , and $\omega_{i,y}$ for the classification weights of class $y$ in the classifier $\omega_i$ . The CAM of $x$ is: + +$$ +\mathcal {M} ^ {\mathrm {C A M}} = \frac {A - \operatorname* {m i n} (A)}{\operatorname* {m a x} (A) - \operatorname* {m i n} (A)}, A = \omega_ {i, y} ^ {\top} F (x; \theta_ {i}), \tag {1} +$$ + +where $\min (\cdot)$ and $\max (\cdot)$ operations are used for normalization. Then, we upsample $\mathcal{M}^{\mathrm{CAM}}$ to the size of image $x$ and use the same notation. + +![](images/0749e512693154e7a16132e2e3713eb2930a05005bc40316fe6c880c3b65e52d.jpg) +Figure 2. The proposed compression pipeline. Given an image, we extract its CAM-based (baseline) or CIM-based (ours) mask, threshold it to be a 0-1 mask with a fixed threshold $\tau$ , and generate a tight bounding box (bbox) that covers all masked pixels. Then, we downsample the pixels outside the bbox and sum the downsampled image up to the masked image to generate the final compressed image. + +Each value in $\mathcal{M}^{\mathrm{CAM}}$ denotes the activation strength of the model at a specific pixel location. Following the way of generating 0-1 masks in weakly-supervised semantic segmentation works [8, 11, 22], we apply a hard threshold $\tau$ (between 0 and 1) over all values of $\mathcal{M}^{\mathrm{CAM}}$ , and get the 0-1 mask $\mathcal{M}^{\tau}$ : $\mathcal{M}^{\tau} = \mathbb{I}(\mathcal{M}^{\mathrm{CAM}} > \tau)$ , where $\mathbb{I}(\cdot)$ is the indicator function. In $\mathcal{M}^{\tau}$ , 1s indicate the locations of discriminative pixels, e.g., foreground pixels, based on which the model makes the prediction. While 0s indicate mostly background pixels or non-discrimination pixels that can be downsampled as they make little contribution to the prediction. + +After generating 0-1 masks, it is ideal to keep them in the memory as meta information of compression. However, this is not efficient or feasible in CIL. There are two reasons. 1) The space for saving image-size masks is non-negligible. Each mask pixel is a one-bit boolean value, and one mask takes around $\frac{1}{3 \times 8} = \frac{1}{24}$ of the memory of one RGB image. 2) The mask involves activated regions with irregular shapes. It is thus non-trivial to perform any standard down-sampling algorithm [34] on the remaining regions. + +From 0-1 Masks to Bounding Boxes (BBox). A simple workaround is to generate a tight bounding box (bbox) to cover the positions of 1s in $\mathcal{M}^{\tau}$ , and use the_bbox for compression. Specifically, given $\mathcal{M}^{\tau}$ , we obtain the coordinate representation of the bounding box as: + +$$ +\mathcal {B} = \left[ \min h, \min w; \max h, \max w \right] _ {(h, w): \mathcal {M} ^ {\tau} (h, w) = 1}, \tag {2} +$$ + +where $h$ and $w$ denote the vertical and horizontal coordinates of the 1 on $\mathcal{M}^{\tau}$ , respectively. We highlight that $\mathcal{B}$ consists of four integers only and takes negligible memory overhead compared to $\mathcal{M}^{\tau}$ . In addition, we "reshape" the irregular shape (of the activated region) in $\mathcal{M}^{\tau}$ into a rectangular $\mathcal{B}$ , so our downsampling operation on the pixels outside the rectangular becomes easy. + +Compression with BBox. Given the image and its bbox on foreground, the compression is implemented by downsampling pixels outside the bbox. Specifically, as illustrated in Figure 2, we compress the image $x$ to $\tilde{x}$ as follows, + +$$ +\tilde {x} = \mathcal {M} ^ {\mathcal {B}} \odot x + (1 - \mathcal {M} ^ {\mathcal {B}}) \odot x _ {\eta}, \tag {3} +$$ + +where $\mathcal{M}^{\mathcal{B}}$ is the binary mask according to $\mathcal{B}$ , i.e., the val + +ues of $\mathcal{M}^{\mathcal{B}}$ are 1 inside $\mathcal{B}$ , and 0 otherwise. $x_{\eta}$ is the fully downsampled version of $x$ with a downsampling ratio $\eta$ ( $\eta > 1$ ), $\odot$ denotes the element-wise product, and $+$ denotes the element-wise addition. Both $\odot$ and $+$ are applied independently on each RGB channel. + +The memory allocated for the compressed image $\tilde{x}$ is as follows, + +$$ +\begin{array}{l} m _ {\tilde {x}} = \frac {H _ {\mathcal {B}} W _ {\mathcal {B}}}{H W} + \frac {1}{\eta} \left(1 - \frac {H _ {\mathcal {B}} W _ {\mathcal {B}}}{H W}\right) \tag {4} \\ = 1 - \left(1 - \frac {1}{\eta}\right) \cdot \left(1 - \frac {H _ {\mathcal {B}} W _ {\mathcal {B}}}{H W}\right), \\ \end{array} +$$ + +where $H_{\mathcal{B}}$ and $W_{\mathcal{B}}$ are the height and width of $\mathcal{B}$ , respectively. $H$ and $W$ are the height and width of the original image $x$ , respectively. $m_{\tilde{x}}$ is always smaller than 1 where 1 denotes the memory unit of saving one original image $x$ . Therefore, we can save a larger number of compressed exemplars within the same memory budget. We denote the set of compressed exemplars in the $i$ -th phase as $\tilde{\mathcal{E}}_i$ . + +Compression Artifacts. The above compression introduces artifacts to the compressed images, i.e., there is a resolution mutation around bounding box edges. From the perspective of spectrum analysis [5,20], such mutation carries noisy and high-frequency components and impairs the model training in subsequent phases. We mitigate the effect of these artifacts by implementing the following data augmentation: in each training epoch, we transform a random subset of $\mathcal{D}_i$ into compressed images with CAM-based bounding boxes using the same downsampling ratio. Using this augmentation enables the model to "simulate" the training with compressed images and learn to be invariant to compression artifacts. + +# 4.2. Class-Incremental Masking (CIM) + +Ideally, the mask generation process needs to adjust at different phases in CIL environments. The process involves two hyperparameters: the masking threshold and the choice of network activation functions. First, for the threshold, searching for its optimum (for all classes) is not trivial. Grid search is intuitive, but it is computationally expensive when the number of classes increases in CIL. Second, for the acti + +![](images/390c8a83693147661c7a7845ebf4c21b3bbe8dc5e18442187584b423f580a6f4.jpg) +Figure 3. Our CIM installs an "extension" on the backbone (e.g., ResNet-18 [16, 36] with four blocks) by adding a learnable activation function (e.g., PAU [32]) at the position of the original activation function (i.e., ReLU [33]). Such "extension" results in a new network branch (in green), whose weight layer parameters are directly copied from the original branch (in blue). + +vation function, the standard network of CIL methods uses ReLU [33, 36] and does not optimize it. + +We solve this issue by applying a learnable activation function in addition to the existing ReLU function in the CIL model. Physically, we have one neural network, while logically, we have two models to learn (in each incremental phase): the conventional CIL model with ReLU activations, and the adaptive mask generation model with learnable activations. We thus call our method class-incremental masking (CIM) based CIL. In the following, we elaborate on the network design and optimization pipeline. + +Network Design. Figure 3 demonstrates an example network architecture in our CIM-based CIL. The proposed CIM extends the network backbone by logically adding a network branch, where only the activation functions are learnable (e.g., Padé Activation Units $(\mathrm{PAU}^4$ , [32])) and the parameters of weight layers are copied from the original branch. This design is motivated by the works of He et al. [15] and Bochkovskiy et al. [2], which indicate that layers with learnable activation functions can flexibly process object (localization) information at different network blocks. The difference is that we apply this flexibility to achieve adaptive mask generation for different CIL phases. We denote the CIM parameters (i.e., the parameters in learnable activation functions) in $i$ -th phase as $\phi_i$ . We optimize the CIL model $(\theta_i, \omega_i)$ and the CIM model $\phi_i$ via a global BOP, as elaborated below. + +Optimization Pipeline. We demonstrate the overall optimization flow in Algorithm 1, which consists of two levels of optimization: task-level and mask-level—the former one for CIL and the latter one for CIM. Note that to maintain an unified notation, we further define $\tilde{\mathcal{E}}_0 = \varnothing$ . + +1) Task-level Optimization. This level aims to optimize the CIL model $(\theta_i, \omega_i)$ to address the CIL task at hand. It + +Algorithm 1: CIM-based CIL (i-th phase, $i \geq 1$ ) +Input: New data $\mathcal{D}_i$ ; old compressed exemplars $\tilde{\mathcal{E}}_{0:i - 1}(\tilde{\mathcal{E}}_0 = \emptyset)$ ; last-phase CIL model $(\theta_{i - 1},\omega_{i - 1})$ and CIM model $\phi_{i - 1}$ $(\theta_0,\omega_0$ are random parameters, $\phi_0$ is set to ReLU). Output: New compressed exemplars $\tilde{\mathcal{E}}_i$ ; updated CIL model $(\theta_i,\omega_i)$ and CIM model $\phi_{i}$ +1 Initialize $(\theta_{i},\omega_{i})$ with $(\theta_{i - 1},\omega_{i - 1})$ . +2 Initialize $\phi_{i}$ with $\phi_{i - 1}$ . +3 for epochs do. +4 Train $(\theta_{i},\omega_{i})$ using $\tilde{\mathcal{E}}_{0:i - 1}\cup \mathcal{D}_i$ by Eq. 5; +5 Compress $\mathcal{D}_i$ into $\tilde{\mathcal{D}}_i^\phi$ using $\phi_{i}$ . +6 Temporarily update $\theta_{i}$ to $\theta_i^+$ using $\tilde{\mathcal{E}}_{0:i - 1}\cup \tilde{\mathcal{D}}_i^\phi$ by Eq.7; +7 Learn $\phi_{i}$ using $\mathcal{D}_i$ by Eq.9; +8 end +9 Compress $\mathcal{D}_i$ into $\tilde{\mathcal{D}}_i$ using the learned $\phi_{i}$ by Eq. 3 +0 Select exemplars $\tilde{\mathcal{E}}_i$ from $\tilde{\mathcal{D}}_i$ by, e.g., herding [36]. + +can be written as: + +$$ +\left(\theta_ {i}, \omega_ {i}\right) \leftarrow \left(\theta_ {i}, \omega_ {i}\right) - \lambda \nabla_ {\left(\theta , \omega\right)} \mathcal {L} _ {\mathrm {C I L}} \left(\tilde {\mathcal {E}} _ {0: i - 1} \cup \mathcal {D} _ {i}; \theta_ {i}, \omega_ {i}\right), \tag {5} +$$ + +where $\lambda$ is the learning rate. We follow the implementation of CIL training loss $\mathcal{L}_{\mathrm{CIL}}$ in baseline methods [23, 42, 46, 50]. This means that we use different training losses when plugging CIM into different baseline methods. + +2) Mask-level Optimization. This level aims to optimize the CIM model $\phi_{i}$ to produce adaptive compression masks. It is formulated as a local BOP: + +$$ +\min _ {\phi_ {i}} \left[ \mathcal {L} _ {\text {v a l}} \left(\mathcal {D} _ {i}; \theta_ {i} ^ {*}, \omega_ {i}\right) + \mu \mathcal {R} \left(\phi_ {i}\right) \right], \tag {6a} +$$ + +$$ +\text {s . t .} \theta_ {i} ^ {*} = \underset {\theta_ {i}} {\arg \min } \mathcal {L} _ {\operatorname {t r a i n}} \left(\tilde {\mathcal {E}} _ {0: i - 1} \cup \tilde {\mathcal {D}} _ {i} \left(\phi_ {i}\right); \theta_ {i}, \omega_ {i}\right). \tag {6b} +$$ + +Eq. 6b denotes an inner-level optimization. It trains $\theta_{i}$ with the data $\tilde{\mathcal{D}}_i(\phi_i)$ compressed by using $\phi_{i}$ , and converges as $\theta_{i}^{*}$ . Eq. 6a denotes an outer-level optimization. It is based on the validation loss derived by $\theta_{i}^{*}$ on the original data $\mathcal{D}_i$ . $\mathcal{R}(\phi_i)$ is a constraint representing the memory limitation and $\mu$ is its weight. In the following, we elaborate on the implementation details for the two levels. + +In the inner-level optimization, we train a temporary CIL model with compressed data. Specifically, we first compress new-class data $\mathcal{D}_i$ into $\tilde{\mathcal{D}}_i(\phi_i)$ using the generated masks by the CIM model $\phi_i$ . Then, we implement the inner-level optimization as a one-step gradient descent (using the CIL training loss) as: + +$$ +\theta_ {i} ^ {+} \leftarrow \theta_ {i} - \beta_ {1} \nabla_ {\theta} \mathcal {L} _ {\mathrm {C I L}} \left(\tilde {\mathcal {E}} _ {0: i - 1} \cup \tilde {\mathcal {D}} _ {i} \left(\phi_ {i}\right); \theta_ {i}, \omega_ {i}\right), \tag {7} +$$ + +where $\beta_{1}$ is the learning rate for $\theta_{i}$ . + +The aim of the outer-level optimization is to optimize $\phi_{i}$ such that the temporary model $(\theta_{i}^{+},\omega_{i})$ (trained with compressed data) has a low validation loss on the original data $D_{i}$ . To achieve this, we back-propagate the loss on the original data $D_{i}$ to update $\phi_{i}$ as: + +$$ +\phi_ {i} \leftarrow \phi_ {i} - \beta_ {2} \nabla_ {\phi} \left[ \mathcal {L} _ {\mathrm {C E}} \left(\mathcal {D} _ {i}; \theta_ {i} ^ {+}, \omega_ {i}\right) + \mu \mathcal {R} \left(\phi_ {i}\right) \right], \tag {8} +$$ + +where $\mathcal{L}_{\mathrm{CE}}$ denotes softmax cross-entropy loss and $\beta_{2}$ is the learning rate for $\phi_{i}$ . This trains $\phi_{i}$ to capture the most discriminative features of new-class images. The constraint $\mathcal{R}(\phi_i)$ is implemented as a $\ell_2$ -regularization term on the generated mask by $\phi_{i}$ . The motivation for the regularization term is to make the mask coverage smaller, thus the compressed images take less memory. + +We empirically observe that by the above optimization flow, the output activation maps by $\phi_{i}$ are easy to collapse, i.e., different images have the same map. To solve this issue, we add a cross-entropy loss term about $\phi_{i}$ in Eq. 8 to regularize it to produce image-specific activation maps: + +$$ +\begin{array}{l} \phi_ {i} \leftarrow \phi_ {i} - \beta_ {2} \nabla_ {\phi} \left[ \mathcal {L} _ {\mathrm {C E}} \left(\mathcal {D} _ {i}; \theta_ {i} ^ {+}, \omega_ {i}\right) + \mu \mathcal {R} (\phi_ {i}) \right. \tag {9} \\ + \mu^ {\prime} \mathcal {L} _ {\mathrm {C E}} \left(\tilde {\mathcal {E}} _ {0: i - 1} \cup \mathcal {D} _ {i}; \theta_ {i}, \phi_ {i}, \omega_ {i}\right) ], \\ \end{array} +$$ + +where $\mu^{\prime}$ is the weight. + +Limitations. Our CIM learns to generate adaptive masks for exemplar compression in CIL. It has three limitations that are left as future work. 1) It is not able to adjust any previous-phase exemplars, as the validation data (the original data of these exemplars) are not accessible anymore. 2) It introduces hundreds of activation parameters to the CIL model, although this is not a significant overhead compared to model parameters. Please check detailed overhead analyses in the supplementary materials. 3) Image compression is not that meaningful for low-resolution datasets (e.g., $32 \times 32$ CIFAR-100). It is because the memory taken by the compression parameters (e.g., the parameters of CIM) and the RGB pixels of a low-resolution image are comparable. Using the memory to save more images is more meaningful. + +# 5. Experiments + +We incorporate CIM into two baseline CIL methods (i.e., DER [46] and FOSTER [42]) and boost their model performances consistently on three datasets. Below, we introduce datasets and experiment settings (Section 5.1), followed by results and analyses (Section 5.2) + +# 5.1. Experimental Settings + +Datasets. We conduct experiments on three standard CIL benchmarks with high-resolution images. 1) Food-101 [3] consists of 101 food categories with 750 training and 250 test samples per category. All images have a maximum side length of 512 pixels. 2) ImageNet-1000 [9] is a large-scale + +dataset with 1,000 classes and each class has around 1,300 training and 50 test samples. 3) ImageNet-100 is a 100-class subset randomly sampled from ImageNet-1000 with a fixed NumPy [14] random seed (1993), following [17]. We provide other details of these datasets, e.g., image sizes and pre-processing methods, in the supplementary materials. + +Protocols. We use two protocols: learning from scratch (LFS) and learning from half (LFH), following recent CIL works [42, 46]. In LFS, the model observes the same number of classes in all $N$ phases, where $N$ is optionally 5, 10, and 20. In LFH, the model is trained on half of the classes (e.g., 500 classes for ImageNet-1000) in the 1-th phase. Then, it learns the remaining classes evenly in the subsequent $N$ phases, where $N$ can be 5, 10, and 25. In both protocols, after the training of each phase we evaluate the resultant model on the test data of all seen classes. Our final report includes the average accuracy over all phases and the last-phase accuracy which indicates the degree of model forgetting. We run each experiment three times and report the average results. + +Memory Budget. There are two memory budget settings. 1) In the "fixed" setting, we remove some old-class exemplars when new exemplars from the current phase are added in the memory to maintain the "fixed memory budget". In this setting, we set the total memory to be 2,020 samples for Food-101 and 2,000 samples for ImageNet-100. For ImageNet-1000, we have two options—5,000 samples and 20,000 samples. 2) In the "growing" setting, a constant memory budget is allocated for each class across all phases and hence extra memory is appended when new classes come. In this setting, we set the budget to be 20 samples per class for all datasets. Following [42, 46], we apply the "fixed" setting in LFS experiments and the "growing" setting in LFH experiments. + +Implementation Details. Our implementation is based on the standard deep learning library PyTorch [35] and image processing library OpenCV [4]. Following [42, 43, 46], we use an 18-layer ResNet [16] as the network backbone $\theta$ and a fully-connected layer as the classifier $\omega$ in all experiments. We use the same CIL training hyperparameters as in related works [17, 42, 46] for fair comparison: 1) there are 200 epochs in 1-st phase and 170 epochs in the subsequent phases; 2) the learning rate $\lambda$ is initialized as 0.1 and decreases to zero with a cosine annealing scheduler [28]; 3) the SGD optimizer is deployed, with momentum factor set to 0.9 and weight decay set to 0.0005. For compression-related hyperparameters, we set the masking threshold $\tau$ as 0.6 and the downsampling ratio $\eta$ as 4.0. To build the CIM model, we apply PAUs with degrees $m = 5$ and $n = 4$ as learnable activation layers. For the optimization of the CIM + +
MethodLearning from Scratch (LFS)Learning from Half (LFH)
Food-101ImageNet-100Food-101ImageNet-100
N=51020510205102551025
iCaRL [36]69.6662.1856.7073.9067.0662.3660.1353.4246.8762.5359.8852.97
WA [48]70.9463.6958.4574.6468.6263.2063.5557.6052.4865.7563.7158.34
PODNet [12]68.0361.2447.3872.1463.9653.6975.3770.0165.3275.5474.3368.33
AANets [25]69.4661.5948.8372.9865.7755.3676.0771.2266.9376.9675.5871.78
DER [46]73.8870.7664.3978.5076.1273.7978.1373.45-79.0877.73-
DER w/ ours75.6373.0969.1779.6377.5775.3679.2575.76-80.3079.05-
FOSTER [42]75.0372.7266.73\( 79.93^{\dagger} \)\( 76.55^{\dagger} \)74.4979.0875.0768.08\( 80.07^{\dagger} \)77.54\( 72.40^{\times} \)
FOSTER w/ ours76.4474.8570.2080.5877.9475.2379.7676.8670.5080.9378.6675.74
+ +† The paper of FOSTER [42] did not report the numerical results for $N = 5 / 10$ (LFS) and $N = 5$ (LFH) on ImageNet-100. We run these experiments using the public code (released by authors) and report the reproduced results. +* Our reproduced result (72.40) is significantly higher than the original result (69.34) reported in the paper of FOSTER. + +Table 1. Average accuracies $(\%)$ of two top-performing CIL methods [42, 46] with and without our CIM-CIL plugged-in, and other four baselines [12, 25, 36, 48], on two datasets (Food-101 and ImageNet-100) and using two protocols (learning from scratch (LFS) and learning from half (LFH)). Due to the space limits, we report the $95\%$ confidence intervals for these results in the supplementary materials. + +
Memory BudgetMethodN=5N=10
Avg.LastAvg.Last
M=20kiCaRL [36]44.3627.7838.4022.70
WA [48]58.3750.6254.1045.66
DER [46]67.4959.7566.7358.62
FOSTER [42]69.2164.8868.3460.14
FOSTER w/ ours69.9366.0569.5362.07
M=5kFOSTER57.1949.4254.7244.96
FOSTER w/ ours61.3754.4659.4850.83
+ +model $\phi$ (i.e., the mask-level optimization), we initially set $\beta_{1}$ as 0.1 and $\beta_{2}$ as 0.01 and reduce them to zero following the scheduler of $\lambda$ . $\mu$ and $\mu^{\prime}$ is set to 0.1 and 0.2, respectively. To smooth the training, we clip the gradient norm of $\phi$ to be no more than 1. We report the result of hyperparameter sensitivity analysis in the supplementary materials. + +# 5.2. Results and Analyses + +Comparing with the State-of-the-art. In Table 1, we summarize the experimental results on two datasets (Food-101 and ImageNet-100) and in two CIL protocols (LFS and LFH). From the table, we have the following observations. 1) Our CIM-based CIL consistently improves the state-ofthe-art method FOSTER [42] with clear margins in all settings. E.g., our method surpasses it by an average of 1.4 percentage points on ImageNet-100, and 2.0 percentage points on Food-101. 2) Our CIM-based CIL achieves more significant improvements when $N$ becomes larger, e.g., on ImageNet-100 (LFH), our method improves FOSTER by + +Table 2. Average and last accuracies $(\%)$ on ImageNet-1000 of FOSTER with and without our method plugged-in, and other three baselines [36, 46, 48]. We show two memory budgets, $M = 20k$ (upper block) and $M = 5k$ (lower block), in the LFS setting. + +
Ablation MethodFood-101ImageNet-100
N=10201020
1 Baseline72.7266.7376.5572.37
2 Artifact Aug.71.3866.0375.6371.45
3 Full Comp.73.0367.3876.9273.26
4 Random Acti.73.1067.5476.8873.54
5 Center Acti.73.2967.8876.7873.82
6 Class Acti.73.7668.6577.2174.67
7 Phase-wise τ73.8369.1777.0674.78
8 Joint Train73.4469.0177.3474.59
9 BOP (ours)74.8570.2077.9475.23
10 LastBlock Only74.5569.8777.7274.86
11 Fg Compressed75.0270.1377.8775.46
+ +Table 3. Average accuracies $(\%)$ of different ablation methods. The experiments are conducted in the LFS setting. + +0.9 and 3.3 percentage points when $N = 5$ and $N = 25$ , respectively. 3) Our CIM-based CIL achieves greater improvements consistently on Food-101 (than ImageNet-100). It improves baselines by 2.1 percentage points on Food-101, while the improvement is 1.4 on ImageNet-100 ( $N = 10$ , LFS). It shows that our method is particularly effective when the representative visual cues of a class are from some of its components, e.g., the "cream" of the class "cake". + +Table 2 shows the results on the large-scale dataset ImageNet-1000 in different memory settings ( $M = 20k$ and $M = 5k$ ). We can see that our CIM-based CIL improves FOSTER consistently. It is impressive that it achieves more improvements in the more strict memory setting ( $M = 5k$ ). Specifically, it boosts the average accuracy of FOSTER by 4.5 percentage points when $M = 5k$ , significantly higher than that of $M = 20k$ (1.0). + +Ablation Study. Table 3 shows the ablation results. First + +
MethodImageNet-100ImageNet-1000
N=6112661126
LUCIR baseline71.2269.6767.4565.2362.4359.88
w/ Mnemonics73.3072.1771.5066.1563.1263.08
w/ MRDC73.6272.8170.4467.6765.6062.74
w/ ours74.0573.7672.8468.0366.5463.77
+ +Table 4. Comparing with Mnemonics [27] and MRDC [43]. We plug each of them in baseline LUCIR [17] for fair comparison. + +block: baselines. Row 1 is for the baseline FOSTER [42]. Row 2 shows the results of adding artifact augmentation (see Section 4.1). It shows directly apply this augmentation does not improve and even impair the model. Please note the models in below blocks all use this augmentation. Second block: activation methods. Rows 3-6 show the results of using different activation methods to compress exemplars. Row 3 is to downsample all pixels (i.e., no region is activated). Row 4 is to randomly select activation regions. Row 5 is to activate only the center region $(\frac{1}{4}$ of the original image), while row 6 is to use naive CAM. Comparing them to row 1, we can see that using naively compressed exemplars can improve CIL models. Row 4 outperforms rows 3-5, validating that it is more reliable to use the model's activation to generate compressed exemplars. Third block: optimization methods. Rows 7-9 are on top of Row 6 and are the results of applying different optimization strategies. Row 7 is to manually select $\tau$ using a held-out set ( $10\%$ of the dataset). Row 8 is to jointly train CIL and CIM models (for each input batch). Row 9 is the proposed method of using a global BOP. Fourth block: two variants of CIM-based CIL. Rows 10-11 are two variants of row 9. In Row 10, only the activation layers in the last block of CIM are learnable, and previous blocks use ReLU. Compared to row 9, row 10 shows slightly worse performance. Row 11 shows the version of adding a weak downsampling $(\eta' = 2.0)$ on discriminative regions, based on which more compressed exemplars are saved. It results in comparable performance to row 9 but increases costs. + +Comparing with Other Compression-based Methods. Table 4 shows our results comparing to two compression-based methods: Mnemonics [27] and MRDC [43]. We can see that our method consistently outperforms them in all settings. This is because our method does not sacrifice the discriminativeness of exemplars while improving the number (variance) of exemplars in a phase-adaptive manner. However, the two related methods either keep a fixed number of exemplars in the memory [27] or use uniform image compression without considering the properties of specific classes in different incremental phases [43]. + +Visualizations (CAM vs. CIM). Figure 4 gives two visualization examples, "Afghan hound" and "indigo bird", each with the activation map as well as the bounding boxes. + +![](images/21467937e61d1a11cdb13cd77972360696d8f03f49efae8b77e1b2e9adad253a.jpg) +Figure 4. Visualizations (CAM vs. CIM). The experimental setting is $N = 10$ (LFS) on ImageNet-100. + +
MetricSmallMiddleLarge
Mean of #Exemplars39.4038.3034.77
Last Acc. (%, baseline)66.1368.4069.93
Last Acc. (%, ours)70.0071.1072.26
Improvement (%)+3.87+3.65+2.33
+ +Table 5. Results for small, middle and large objects in the setting of $N = 10$ (LFS) on ImageNet-100. "Mean of #Exemplars" denotes the average number of saved exemplars by our method. The baseline (FOSTER [42]) has this number as 20 for all classes. + +The first column shows their respective confusing classes appearing in earlier phases. CIM learns to focus on the discriminative (i.e., dissimilar to confusing classes) regions. + +Results of Different-Size Objects. Table 5 shows the results for small, middle, and large objects. These size categorization is according to ImageNet Object Localization Challenge [19]. We calculate the bbox coverage for each class and take top 30 classes with highest coverages as "large", near 30 classes with lowest coverages as "small" and the rest 40 classes as "middle". It is intriguing that our method achieves the highest improvement (over baseline) for small objects. Our explanation is that small objects benefit more from image compression (than large ones), as their images contain more background pixels to downsample. + +# 6. Conclusions + +We introduced a novel exemplar compression method for CIL, allowing us to save more representative exemplars but not increase memory budget. We achieved this compression by downsampling non-discriminative pixels of the image bounded by CAM masks. To generate adaptive masks, we proposed a novel method CIM that explicitly parameterizes a mask generation model and optimizes it in an end-to-end manner across incremental phases. Our method achieves consistent performance improvements over multiple baselines and can be taken as a flexible plug-and-play module. + +Acknowledgements. The author gratefully acknowledges the support by the Lee Kong Chian (LKC) Fellowship fund awarded by Singapore Management University. + +# References + +[1] Davide Abati, Jakub Tomczak, Tijmen Blankevoort, Simone Calderara, Rita Cucchiara, and Babak Ehteshami Bejnordi. Conditional channel gated networks for task-aware continual learning. In CVPR, 2020. 2 +[2] Alexey Bochkovskiy, Chien-Yao Wang, and Hong-Yuan Mark Liao. Yolov4: Optimal speed and accuracy of object detection. arXiv, 2020. 5 +[3] Lukas Bossard, Matthieu Guillaumin, and Luc Van Gool. Food-101 - mining discriminative components with random forests. In ECCV, 2014. 2, 6 +[4] Gary Bradski. The opencv library. Dr. Dobb's Journal: Software Tools for the Professional Programmer, 2000. 6 +[5] Kenneth R Castleman. Digital image processing. Prentice Hall Press, 1996. 4 +[6] George Cazenavette, Tongzhou Wang, Antonio Torralba, Alexei A Efros, and Jun-Yan Zhu. Dataset distillation by matching training trajectories. In CVPR, 2022. 3 +[7] Can Chen, Xi Chen, Chen Ma, Zixuan Liu, and Xue Liu. Gradient-based bi-level optimization for deep learning: A survey. arXiv, 2022. 2, 3 +[8] Zhaozheng Chen, Tan Wang, Xiongwei Wu, Xian-Sheng Hua, Hanwang Zhang, and Qianru Sun. Class re-activation maps for weakly-supervised semantic segmentation. In CVPR, 2022. 3, 4 +[9] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In CVPR, 2009. 1, 2, 3, 6 +[10] Prithviraj Dhar, Rajat Vikram Singh, Kuan-Chuan Peng, Ziyan Wu, and Rama Chellappa. Learning without memorizing. In CVPR, 2019. 2 +[11] Zhang Dong, Zhang Hanwang, Tang Jinhui, Hua Xiansheng, and Sun Qianru. Causal intervention for weakly supervised semantic segmentation. In NeurIPS, 2020. 4 +[12] Arthur Douillard, Matthieu Cord, Charles Ollion, Thomas Robert, and Eduardo Valle. Podnet: Pooled outputs distillation for small-tasks incremental learning. In ECCV, 2020. 1, 2, 7 +[13] Chelsea Finn, Pieter Abbeel, and Sergey Levine. Model-agnostic meta-learning for fast adaptation of deep networks. In ICML, 2017. 3 +[14] Charles R Harris, K Jarrod Millman, Stefan J Van Der Walt, Ralf Gommers, Pauli Virtanen, David Cournapeau, Eric Wieser, Julian Taylor, Sebastian Berg, Nathaniel J Smith, et al. Array programming with numpy. Nature, 2020. 6 +[15] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Delving deep into rectifiers: Surpassing human-level performance on imagenet classification. In ICCV, 2015. 5 +[16] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, 2016. 5, 6 +[17] Saihui Hou, Xinyu Pan, Chen Change Loy, Zilei Wang, and Dahua Lin. Learning a unified classifier incrementally via rebalancing. In CVPR, 2019. 1, 2, 6, 8 +[18] Shenyang Huang, Vincent François-Lavet, and Guillaume Rabusseau. Neural architecture search for class-incremental learning. arXiv, 2019. 2 + +[19] Imagenet object localization challenge. https://www.kaggle.com/competitions/imagenet-object-localization-challenge/.8 +[20] Anil K Jain. Fundamentals of digital image processing. Prentice-Hall, Inc., 1989. 4 +[21] James Kirkpatrick, Razvan Pascanu, Neil Rabinowitz, Joel Veness, Guillaume Desjardins, Andrei A Rusu, Kieran Milan, John Quan, Tiago Ramalho, Agnieszka Grabska-Barwinska, et al. Overcoming catastrophic forgetting in neural networks. PNAS, 2017. 2 +[22] Jungbeom Lee, Eunji Kim, and Sungroh Yoon. Antiadversarily manipulated attributions for weakly and semi-supervised semantic segmentation. In CVPR, 2021. 3, 4 +[23] Zhizhong Li and Derek Hoiem. Learning without forgetting. PAMI, 2017. 2, 5 +[24] Yaoyao Liu, Yingying Li, Bernt Schiele, and Qianru Sun. Online hyperparameter optimization for class-incremental learning. In AAAi, 2023. 2 +[25] Yaoyao Liu, Bernt Schiele, and Qianru Sun. Adaptive aggregation networks for class-incremental learning. In CVPR, 2021. 1, 2, 3, 7 +[26] Yaoyao Liu, Bernt Schiele, and Qianru Sun. Rmm: Reinforced memory management for class-incremental learning. In NeurIPS, 2021. 2 +[27] Yaoyao Liu, Yuting Su, An-An Liu, Bernt Schiele, and Qianru Sun. Mnemonics training: Multi-class incremental learning without forgetting. In CVPR, 2020. 1, 2, 3, 8 +[28] Ilya Loshchilov and Frank Hutter. Sgdr: Stochastic gradient descent with warm restarts. arXiv, 2016. 6 +[29] Dougal Maclaurin, David Duvenaud, and Ryan Adams. Gradient-based hyperparameter optimization through reversible learning. In ICML, 2015. 3 +[30] Michael McCloskey and Neal J Cohen. Catastrophic interference in connectionist networks: The sequential learning problem. In *Psychology of Learning and Motivation*. Elsevier, 1989. 2 +[31] Ken McRae and Phil A Hetherington. Catastrophic interference is eliminated in pretrained networks. In Proceedings of the 15th Annual Conference of the Cognitive Science Society, 1993. 1, 2 +[32] Alejandro Molina, Patrick Schramowski, and Kristian Kersting. Padé activation units: End-to-end learning of flexible activation functions in deep networks. In ICLR, 2019. 5 +[33] Vinod Nair and Geoffrey E Hinton. Rectified linear units improve restricted boltzmann machines. In ICML, 2010. 5 +[34] J Anthony Parker, Robert V Kenyon, and Donald E Troxel. Comparison of interpolating methods for image resampling. IEEE Transactions on Medical Imaging, 1983. 4 +[35] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. In NeurIPS, 2019. 6 +[36] Sylvestre-Alvise Rebuffi, Alexander Kolesnikov, Georg Sperl, and Christoph H Lampert. iCaRL: Incremental classifier and representation learning. In CVPR, 2017. 1, 2, 3, 5, 7 + +[37] Andrei A Rusu, Neil C Rabinowitz, Guillaume Desjardins, Hubert Soyer, James Kirkpatrick, Koray Kavukcuoglu, Razvan Pascanu, and Raia Hadsell. Progressive neural networks. arXiv, 2016. 2 +[38] Ramprasaath R Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra. Grad-cam: Visual explanations from deep networks via gradient-based localization. In ICCV, 2017. 3 +[39] Christian Simon, Piotr Koniusz, and Mehrtash Harandi. On learning the geodesic path for incremental learning. In CVPR, 2021. 2 +[40] Ankur Sinha, Pekka Malo, and Kalyanmoy Deb. A review on bilevel optimization: from classical to evolutionary approaches and applications. IEEE Transactions on Evolutionary Computation, 2017. 1, 2, 3 +[41] Gregory K Wallace. The jpeg still picture compression standard. Communications of the ACM, 1991. 1, 2 +[42] Fu-Yun Wang, Da-Wei Zhou, Han-Jia Ye, and De-Chuan Zhan. Foster: Feature boosting and compression for class incremental learning. arXiv, 2022. 1, 2, 3, 5, 6, 7, 8 +[43] Liyuan Wang, Xingxing Zhang, Kuo Yang, Longhui Yu, Chongxuan Li, Lanqing Hong, Shifeng Zhang, Zhenguo Li, Yi Zhong, and Jun Zhu. Memory replay with data compression for continual learning. In ICLR, 2022. 1, 2, 3, 6, 8 +[44] Yue Wu, Yinpeng Chen, Lijuan Wang, Yuancheng Ye, Zicheng Liu, Yandong Guo, and Yun Fu. Large scale incremental learning. In CVPR, 2019. 1, 2 +[45] Ju Xu and Zhanxing Zhu. Reinforced continual learning. In NeurIPS, 2018. 2 +[46] Shipeng Yan, Jiangwei Xie, and Xuming He. Der: Dynamically expandable representation for class incremental learning. In CVPR, 2021. 1, 2, 3, 5, 6, 7 +[47] Friedemann Zenke, Ben Poole, and Surya Ganguli. Continual learning through synaptic intelligence. In ICML, 2017. 2 +[48] Bowen Zhao, Xi Xiao, Guojun Gan, Bin Zhang, and Shu-Tao Xia. Maintaining discrimination and fairness in class incremental learning. In CVPR, 2020. 1, 7 +[49] Bolei Zhou, Aditya Khosla, Agata Lapedriza, Aude Oliva, and Antonio Torralba. Learning deep features for discriminative localization. In CVPR, 2016. 1, 2, 3 +[50] Da-Wei Zhou, Han-Jia Ye, and De-Chuan Zhan. Co-transport for class-incremental learning. In Proceedings of the 29th ACM International Conference on Multimedia, 2021. 5 \ No newline at end of file diff --git a/2023/Class-Incremental Exemplar Compression for Class-Incremental Learning/images.zip b/2023/Class-Incremental Exemplar Compression for Class-Incremental Learning/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..0d5e3ce581474a633a9954592968a9296170f181 --- /dev/null +++ b/2023/Class-Incremental Exemplar Compression for Class-Incremental Learning/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4267f516987f58fe7e67291b7c2e362ae9485aeb877478dc9c8f154517f415f4 +size 507191 diff --git a/2023/Class-Incremental Exemplar Compression for Class-Incremental Learning/layout.json b/2023/Class-Incremental Exemplar Compression for Class-Incremental Learning/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..155569f70845df3b369d9d673f8cf8ccf513b4b4 --- /dev/null +++ b/2023/Class-Incremental Exemplar Compression for Class-Incremental Learning/layout.json @@ -0,0 +1,10433 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 66, + 103, + 527, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 103, + 527, + 121 + ], + "spans": [ + { + "bbox": [ + 66, + 103, + 527, + 121 + ], + "type": "text", + "content": "Class-Incremental Exemplar Compression for Class-Incremental Learning" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 157, + 142, + 445, + 158 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 157, + 142, + 445, + 158 + ], + "spans": [ + { + "bbox": [ + 157, + 142, + 445, + 158 + ], + "type": "text", + "content": "Zilin Luo1 Yaoyao Liu2 Bernt Schiele2 Qianru Sun1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 214, + 171, + 389, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 214, + 171, + 389, + 186 + ], + "spans": [ + { + "bbox": [ + 214, + 171, + 389, + 186 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 214, + 171, + 389, + 186 + ], + "type": "text", + "content": "Singapore Management University" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 136, + 186, + 466, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 186, + 466, + 200 + ], + "spans": [ + { + "bbox": [ + 136, + 186, + 466, + 200 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 136, + 186, + 466, + 200 + ], + "type": "text", + "content": "Max Planck Institute for Informatics, Saarland Informatics Campus" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 201, + 548, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 201, + 548, + 213 + ], + "spans": [ + { + "bbox": [ + 55, + 201, + 548, + 213 + ], + "type": "text", + "content": "zilin.luo.2021@phdcs.smu.edu.sg {yaoyao.liu, schiele}@mpi-inf.mpg.de qianrusun@smu.edu.sg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 143, + 241, + 192, + 253 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 241, + 192, + 253 + ], + "spans": [ + { + "bbox": [ + 143, + 241, + 192, + 253 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 266, + 290, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 266, + 290, + 601 + ], + "spans": [ + { + "bbox": [ + 46, + 266, + 290, + 601 + ], + "type": "text", + "content": "Exemplar-based class-incremental learning (CIL) [36] finetunes the model with all samples of new classes but few-shot exemplars of old classes in each incremental phase, where the \"few-shot\" abides by the limited memory budget. In this paper, we break this \"few-shot\" limit based on a simple yet surprisingly effective idea: compressing exemplars by downsampling non-discriminative pixels and saving \"many-shot\" compressed exemplars in the memory. Without needing any manual annotation, we achieve this compression by generating 0-1 masks on discriminative pixels from class activation maps (CAM) [49]. We propose an adaptive mask generation model called class-incremental masking (CIM) to explicitly resolve two difficulties of using CAM: 1) transforming the heatmaps of CAM to 0-1 masks with an arbitrary threshold leads to a trade-off between the coverage on discriminative pixels and the quantity of exemplars, as the total memory is fixed; and 2) optimal thresholds vary for different object classes, which is particularly obvious in the dynamic environment of CIL. We optimize the CIM model alternatively with the conventional CIL model through a bilevel optimization problem [40]. We conduct extensive experiments on high-resolution CIL benchmarks including Food-101, ImageNet-100, and ImageNet-1000, and show that using the compressed exemplars by CIM can achieve a new state-of-the-art CIL accuracy, e.g., 4.8 percentage points higher than FOSTER [42] on 10-Phase ImageNet-1000. Our code is available at https://github.com/xffzl/CIM-CIL." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 621, + 127, + 633 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 621, + 127, + 633 + ], + "spans": [ + { + "bbox": [ + 48, + 621, + 127, + 633 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 642, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 642, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 642, + 287, + 715 + ], + "type": "text", + "content": "Dynamic AI systems have a continual learning nature to learn new class data. They are expected to adapt to new classes while maintaining the knowledge of old classes, i.e., free from forgetting problems [31]. To evaluate this, the following protocol of class-incremental learning (CIL) was proposed by Rebuffi et al. [36]. The model training goes" + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 307, + 258, + 545, + 312 + ], + "blocks": [ + { + "bbox": [ + 307, + 258, + 545, + 312 + ], + "lines": [ + { + "bbox": [ + 307, + 258, + 545, + 312 + ], + "spans": [ + { + "bbox": [ + 307, + 258, + 545, + 312 + ], + "type": "image", + "image_path": "7db32eb0839a7df78ee3689fff037f1a62dfd758d14530fd142bcae94fe50c05.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 347, + 312, + 404, + 320 + ], + "lines": [ + { + "bbox": [ + 347, + 312, + 404, + 320 + ], + "spans": [ + { + "bbox": [ + 347, + 312, + 404, + 320 + ], + "type": "text", + "content": "(a) iCaRL [baseline]" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 308, + 326, + 545, + 379 + ], + "blocks": [ + { + "bbox": [ + 308, + 326, + 545, + 379 + ], + "lines": [ + { + "bbox": [ + 308, + 326, + 545, + 379 + ], + "spans": [ + { + "bbox": [ + 308, + 326, + 545, + 379 + ], + "type": "image", + "image_path": "abc4554419cc1d57e763cc862d6c8d5d7a4b70f5f5456fe93318f05708e3e8f9.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 348, + 380, + 403, + 388 + ], + "lines": [ + { + "bbox": [ + 348, + 380, + 403, + 388 + ], + "spans": [ + { + "bbox": [ + 348, + 380, + 403, + 388 + ], + "type": "text", + "content": "(c) MRDC [related]" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 328, + 392, + 525, + 418 + ], + "blocks": [ + { + "bbox": [ + 328, + 392, + 525, + 418 + ], + "lines": [ + { + "bbox": [ + 328, + 392, + 525, + 418 + ], + "spans": [ + { + "bbox": [ + 328, + 392, + 525, + 418 + ], + "type": "image", + "image_path": "bf45685998f0ace0f9dd19d4778ae5140a66001ec4521222a428c128eda68dd4.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 428, + 545, + 527 + ], + "lines": [ + { + "bbox": [ + 304, + 428, + 545, + 527 + ], + "spans": [ + { + "bbox": [ + 304, + 428, + 545, + 527 + ], + "type": "text", + "content": "Figure 1. The phase-wise training data in different methods. (a) iCaRL [36] is the baseline method using full new class data and few-shot old class exemplars. (b) Mnemonics [27] distills all training samples into few-shot exemplars without increasing their quantity. (c) MRDC [43] compresses each exemplar uniformly into a low-resolution image using JPEG [41]. (d) Our approach based on the proposed class-incremental masking (CIM) down-samples only non-discriminative pixels in the image. The legend shows the symbols of special images generated by the methods." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 533, + 545, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 533, + 545, + 640 + ], + "spans": [ + { + "bbox": [ + 304, + 533, + 545, + 640 + ], + "type": "text", + "content": "through a number of phases. Each phase has new class data added and old class data discarded, and the resultant model is evaluated on the test data of all seen classes. A straightforward way to retain old class knowledge is keeping around a few old class exemplars in the memory and using them to re-train the model in subsequent phases. The number of exemplars is usually limited, e.g., " + }, + { + "bbox": [ + 304, + 533, + 545, + 640 + ], + "type": "inline_equation", + "content": "5 \\sim 20" + }, + { + "bbox": [ + 304, + 533, + 545, + 640 + ], + "type": "text", + "content": " exemplars per class [12, 17, 25, 27, 36, 42, 44, 46, 48], as the total memory in CIL strictly budgeted, e.g., " + }, + { + "bbox": [ + 304, + 533, + 545, + 640 + ], + "type": "inline_equation", + "content": "2k" + }, + { + "bbox": [ + 304, + 533, + 545, + 640 + ], + "type": "text", + "content": " exemplars." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 641, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 641, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 641, + 545, + 713 + ], + "type": "text", + "content": "This leads to a serious data imbalance between old and new classes, e.g., 20 per old class vs. " + }, + { + "bbox": [ + 304, + 641, + 545, + 713 + ], + "type": "inline_equation", + "content": "1.3k" + }, + { + "bbox": [ + 304, + 641, + 545, + 713 + ], + "type": "text", + "content": " per new class (on ImageNet-1000 [9]), as illustrated in Figure 1a. The training is thus always dominated by new classes, and forgetting problems occur for old classes. Liu et al. [27] tried to mitigate this problem by parameterizing and distilling the exem" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "11371" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 239 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 239 + ], + "type": "text", + "content": "plars, without increasing the number of them (Figure 1b). Wang et al. [43] traded off between the quality and quantity of exemplars by uniformly compressing exemplar images with JPEG [41] (Figure 1c). As shown in Figure 1d, our approach is also based on image compression. The idea is to downsample only non-discriminative pixels (e.g., background) and keep discriminative pixels (i.e., representative cues of foreground objects) as the original. In this way, we do not sacrifice the discriminativeness of exemplars when increasing their quantity. In particular, we aim for adaptive compression in dynamic environments of CIL, where the intuition is later phases need to be more conservative (i.e., less downsampling) as the model needs more visual cues to classify the increased number of classes." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 239, + 289, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 239, + 289, + 514 + ], + "spans": [ + { + "bbox": [ + 46, + 239, + 289, + 514 + ], + "type": "text", + "content": "To achieve selective and adaptive compression, we need the location labels of discriminative pixels. Without extra labeling, we automatically generate the labels by utilizing the model's own \"attention\" on discriminative features, i.e., class activation maps (CAM) [49]. We take this method as a feasible baseline, and based on it, we propose an adaptive version called class-incremental masking (CIM). Specifically, for each input image (with its class label), we use its feature maps and classifier weights (corresponding to its class label) to compute a CAM by channel-wise multiplication, aggregation, and normalization. Then, we apply hard thresholding to generate a 0-1 mask. We notice that when generating the masks in the dynamic environments of CIL, the optimal hyperparameters (such as the value of hard threshold and the choice of activation functions) vary for different classes as well as in different incremental phases. Our adaptive version CIM tackles this by parameterizing a mask generation model and optimizing it in an end-to-end manner across all incremental phases. In each phase, the learned CIM model adaptively generates class-and phase-specific masks. We find that the compressed exemplars based on these masks have stronger representativeness, compared to using the conventional CAM." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 514, + 289, + 659 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 514, + 289, + 659 + ], + "spans": [ + { + "bbox": [ + 46, + 514, + 289, + 659 + ], + "type": "text", + "content": "Technically, we have two models to optimize, i.e., the CIL model and the CIM model. These two cannot be optimized separately as they are dependent on computation: 1) the CIM model compresses exemplars to input into the CIL model; 2) the two models share network parameters. We exploit a global bilevel optimization problem (BOP) [7, 40] to alternate their training processes at two levels. This BOP goes through all incremental training phases. In particular, for each phase, we perform a local BOP with two steps to tune the parameters of the CIM model: 1) a temporary model is trained with the compressed exemplars as input; and 2) a validation loss on the uncom" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 306, + 72, + 547, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 547, + 205 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 547, + 205 + ], + "type": "text", + "content": "pressed new data is computed and the gradients are backpropagated to optimize the parameters of CIM. To evaluate CIM, we conduct extensive experiments by plugging it in recent CIL methods," + }, + { + "bbox": [ + 306, + 72, + 547, + 205 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 306, + 72, + 547, + 205 + ], + "type": "text", + "content": " LUCIR [17], DER [46], and FOSTER [42], on three high-resolution benchmarks, Food101 [3], ImageNet-100 [17], and ImageNet-1000 [9]. We find that using the compressed exemplars by CIM brings consistent and significant improvements, e.g., " + }, + { + "bbox": [ + 306, + 72, + 547, + 205 + ], + "type": "inline_equation", + "content": "4.2\\%" + }, + { + "bbox": [ + 306, + 72, + 547, + 205 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 306, + 72, + 547, + 205 + ], + "type": "inline_equation", + "content": "4.8\\%" + }, + { + "bbox": [ + 306, + 72, + 547, + 205 + ], + "type": "text", + "content": " higher than the SOTA method FOSTER [42], respectively, in the 5-phase and 10-phase settings of ImageNet1000, with a total memory budget for " + }, + { + "bbox": [ + 306, + 72, + 547, + 205 + ], + "type": "inline_equation", + "content": "5k" + }, + { + "bbox": [ + 306, + 72, + 547, + 205 + ], + "type": "text", + "content": " exemplars." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 306, + 213, + 392, + 225 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 213, + 392, + 225 + ], + "spans": [ + { + "bbox": [ + 306, + 213, + 392, + 225 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 232, + 547, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 232, + 547, + 496 + ], + "spans": [ + { + "bbox": [ + 304, + 232, + 547, + 496 + ], + "type": "text", + "content": "Class-Incremental Learning (CIL). There are three main lines of work to address the catastrophic forgetting problem [30, 31] in CIL. Regularization-based methods apply discrepancy (between old and new models) penalization terms in their objective functions, e.g., by comparing output logits [23, 36], intermediate features [12, 17, 24, 39], and prediction heatmaps [10]. Parameter-isolation-based methods increase the model parameters in each new incremental phase, to prevent knowledge forgetting caused by parameter overwritten. Some of them [18, 37, 42, 45, 46] proposed to progressively expand the size of the neural network to learn new coming data. Others [1, 21, 25, 47] froze a part of network parameters (to maintain the old class knowledge) to alleviate the problem of knowledge overwriting. Replay-based methods assume there is a clear memory budget allowing a handful of old-class exemplars in the memory. Exemplars can be used to re-train the model in each new phase [12, 17, 26, 27, 36, 43, 44]. This re-training usually contains two steps: one step of training the model on all new class data and old class exemplars, and one step of finetuning the model with a balanced subset (i.e., using an equal number of samples per class) [12, 17, 25, 26, 46]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 497, + 548, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 497, + 548, + 689 + ], + "spans": [ + { + "bbox": [ + 304, + 497, + 548, + 689 + ], + "type": "text", + "content": "The replay-based methods focusing on memory optimization [27, 43] are closely related to our work. [27] proposed a bilevel optimization framework to distill the current new class data into exemplars before discarding them. It aims to improve the quality of exemplars without increasing the quantity. Another work [43] aimed to trade-off between the quality and quantity of exemplars by image compression using the JPEG algorithm, i.e., each exemplar is uniformly downsampled. Ours differs from these two works in three aspects. 1) Our CIM based image compression automatically segments the discriminative pixels in the exemplar and downsamples the non-discriminative pixels only. It barely weakens the representativeness of the exemplar. In contrast, the parameterization of image pixels in [27] hampers the model from capturing high-frequency (discriminative) features from the image, especially the high-resolution (e.g.," + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 306, + 693, + 547, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 693, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 693, + 547, + 713 + ], + "type": "text", + "content": "3Using \"plug-in\" for evaluation is due to the fact that many baseline methods were originally evaluated in different CIL settings." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 664, + 287, + 693 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 664, + 287, + 693 + ], + "spans": [ + { + "bbox": [ + 46, + 664, + 287, + 693 + ], + "type": "text", + "content": "1Note that we do not use mask labels to do image compression because storing them is expensive. Instead, we expand the mask to a bounding box, as elaborated in Section 4." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 693, + 287, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 693, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 46, + 693, + 287, + 712 + ], + "type": "text", + "content": "2Note that the CIM model is actually a plug-in branch in the CIL model, which is detailed in Section 4.2." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "11372" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 286, + 179 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 286, + 179 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 286, + 179 + ], + "type": "inline_equation", + "content": "224 \\times 224)" + }, + { + "bbox": [ + 46, + 72, + 286, + 179 + ], + "type": "text", + "content": " image [6]. 2) Our approach increases the diversity (i.e., quantity) of old class exemplars by reducing the memory consumption for each exemplar. In contrast, [27] keeps a fixed number of exemplars in the memory. 3) Our approach has an adaptive image compression strategy that fits well in the dynamic environments of CIL. In contrast, [43] uses uniform image compression (or uniformly increasing quality parameters from 100 to 1) without considering the properties of specific classes in each learning phase." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 180, + 286, + 287 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 180, + 286, + 287 + ], + "spans": [ + { + "bbox": [ + 46, + 180, + 286, + 287 + ], + "type": "text", + "content": "Class Activation Map (CAM) [49] is a simple yet effective weakly-supervised object localization method. Its model is trained with only the image-level label and can generate pixel-level masks on foreground objects. Specifically, the masks are the results of hard-thresholding the heatmaps produced by feature maps and classifier weights. Advanced CAM variants include Grad-CAM [38], ReCAM [8], AdvCAM [22], etc. Our CIM is based on the vanilla CAM because it is computationally simple and efficient." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 288, + 286, + 455 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 288, + 286, + 455 + ], + "spans": [ + { + "bbox": [ + 46, + 288, + 286, + 455 + ], + "type": "text", + "content": "Bilevel Optimization Problems (BOP) [7, 40] aims to solve a nested optimization problem, where the outer-level optimization is subjected to the result of the inner-level optimization. It has shown effectiveness in a wide range of machine learning areas, such as hyperparameter selection [29] and meta-learning [13]. For tackling CIL tasks, [27] leverages BOP to alternatively optimize the parameters of the CIL model and the parameterized exemplars. [25] applies BOP to learn the aggregation weights of the plastic and elastic branches in the CIL model. In our work, we use BOP to solve the optimization of the CIL model and the parameterized Class-Incremental Masking (CIM) model, where CIM is a plugin branch (in the CIL model), using few extra parameters. The process of BOP is quick and efficient." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 466, + 124, + 479 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 466, + 124, + 479 + ], + "spans": [ + { + "bbox": [ + 47, + 466, + 124, + 479 + ], + "type": "text", + "content": "3. Preliminary" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 486, + 286, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 486, + 286, + 664 + ], + "spans": [ + { + "bbox": [ + 46, + 486, + 286, + 664 + ], + "type": "text", + "content": "The following is the training pipeline of standard CIL with few-shot exemplars. Assume there are " + }, + { + "bbox": [ + 46, + 486, + 286, + 664 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 486, + 286, + 664 + ], + "type": "text", + "content": " learning phases. In the 1-st phase, we load data " + }, + { + "bbox": [ + 46, + 486, + 286, + 664 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_1" + }, + { + "bbox": [ + 46, + 486, + 286, + 664 + ], + "type": "text", + "content": " containing all training samples of " + }, + { + "bbox": [ + 46, + 486, + 286, + 664 + ], + "type": "inline_equation", + "content": "c_1" + }, + { + "bbox": [ + 46, + 486, + 286, + 664 + ], + "type": "text", + "content": " classes, and use " + }, + { + "bbox": [ + 46, + 486, + 286, + 664 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_1" + }, + { + "bbox": [ + 46, + 486, + 286, + 664 + ], + "type": "text", + "content": " to train the initial classification model " + }, + { + "bbox": [ + 46, + 486, + 286, + 664 + ], + "type": "inline_equation", + "content": "(\\theta_{1},\\omega_{1})" + }, + { + "bbox": [ + 46, + 486, + 286, + 664 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 486, + 286, + 664 + ], + "type": "inline_equation", + "content": "\\theta_{1}" + }, + { + "bbox": [ + 46, + 486, + 286, + 664 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 486, + 286, + 664 + ], + "type": "inline_equation", + "content": "\\omega_{1}" + }, + { + "bbox": [ + 46, + 486, + 286, + 664 + ], + "type": "text", + "content": " denote the parameters of the feature extractor and classifier, respectively. When the training is done, we evaluate the model performance on the test samples of " + }, + { + "bbox": [ + 46, + 486, + 286, + 664 + ], + "type": "inline_equation", + "content": "c_1" + }, + { + "bbox": [ + 46, + 486, + 286, + 664 + ], + "type": "text", + "content": " classes. Before the 2nd phase, we discard most of the training samples due to the strict memory budget of CIL. In other words, we preserve only a handful of training samples " + }, + { + "bbox": [ + 46, + 486, + 286, + 664 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_1" + }, + { + "bbox": [ + 46, + 486, + 286, + 664 + ], + "type": "text", + "content": " (i.e., exemplars) in the memory, selected from " + }, + { + "bbox": [ + 46, + 486, + 286, + 664 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_1" + }, + { + "bbox": [ + 46, + 486, + 286, + 664 + ], + "type": "text", + "content": ". A common method for selecting exemplars is called feature herding [36] and has been used in many related works [27, 42, 43, 46]. We adopt it, too, in this work." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 665, + 286, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 286, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 286, + 714 + ], + "type": "text", + "content": "In the " + }, + { + "bbox": [ + 46, + 665, + 286, + 714 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 665, + 286, + 714 + ], + "type": "text", + "content": "-th phase " + }, + { + "bbox": [ + 46, + 665, + 286, + 714 + ], + "type": "inline_equation", + "content": "(i \\geq 2)" + }, + { + "bbox": [ + 46, + 665, + 286, + 714 + ], + "type": "text", + "content": ", we load all exemplars " + }, + { + "bbox": [ + 46, + 665, + 286, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_{1:i-1} = \\mathcal{E}_1 \\cup \\dots \\cup \\mathcal{E}_{i-1}" + }, + { + "bbox": [ + 46, + 665, + 286, + 714 + ], + "type": "text", + "content": " from the memory and initialize the current model " + }, + { + "bbox": [ + 46, + 665, + 286, + 714 + ], + "type": "inline_equation", + "content": "(\\theta_i, \\omega_i)" + }, + { + "bbox": [ + 46, + 665, + 286, + 714 + ], + "type": "text", + "content": " by the previous model " + }, + { + "bbox": [ + 46, + 665, + 286, + 714 + ], + "type": "inline_equation", + "content": "(\\theta_{i-1}, \\omega_{i-1})" + }, + { + "bbox": [ + 46, + 665, + 286, + 714 + ], + "type": "text", + "content": ". We use " + }, + { + "bbox": [ + 46, + 665, + 286, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_{1:i-1}" + }, + { + "bbox": [ + 46, + 665, + 286, + 714 + ], + "type": "text", + "content": " and the new coming data " + }, + { + "bbox": [ + 46, + 665, + 286, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_i" + }, + { + "bbox": [ + 46, + 665, + 286, + 714 + ], + "type": "text", + "content": " (containing " + }, + { + "bbox": [ + 46, + 665, + 286, + 714 + ], + "type": "inline_equation", + "content": "c_i" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 72, + 545, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 157 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 157 + ], + "type": "text", + "content": "new classes) to train " + }, + { + "bbox": [ + 304, + 72, + 545, + 157 + ], + "type": "inline_equation", + "content": "(\\theta_{i},\\omega_{i})" + }, + { + "bbox": [ + 304, + 72, + 545, + 157 + ], + "type": "text", + "content": ". Then, we evaluate the current model using a test set of all " + }, + { + "bbox": [ + 304, + 72, + 545, + 157 + ], + "type": "inline_equation", + "content": "\\sum_{j = 1}^{i}c_{j}" + }, + { + "bbox": [ + 304, + 72, + 545, + 157 + ], + "type": "text", + "content": " classes seen so far. After that, we discard most of the training samples in " + }, + { + "bbox": [ + 304, + 72, + 545, + 157 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_i" + }, + { + "bbox": [ + 304, + 72, + 545, + 157 + ], + "type": "text", + "content": " and leave few-shot exemplars " + }, + { + "bbox": [ + 304, + 72, + 545, + 157 + ], + "type": "inline_equation", + "content": "\\varepsilon_{i}" + }, + { + "bbox": [ + 304, + 72, + 545, + 157 + ], + "type": "text", + "content": " in the memory. It is clear that this discarding causes a strong data imbalance between old and new coming classes in the subsequent phase. In the following, we introduce our solution to this problem." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 167, + 389, + 180 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 167, + 389, + 180 + ], + "spans": [ + { + "bbox": [ + 305, + 167, + 389, + 180 + ], + "type": "text", + "content": "4. Methodology" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 186, + 545, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 186, + 545, + 413 + ], + "spans": [ + { + "bbox": [ + 304, + 186, + 545, + 413 + ], + "type": "text", + "content": "As illustrated in Figure 1d, we alleviate the data imbalance problem by saving a larger number of compressed exemplars for old classes, where we leverage pixel-selective compression, i.e., downsample only non-discriminative pixels. To achieve this, the ideal case is that we have the pixel-level localization of foreground objects. However, the realistic case is that such localization labels are expensive, most CIL benchmarks do not have the labels, and it is not fair to compare with other CIL methods if using the labels. Without extra labeling, we introduce a CAM-based mask generation method, and based on it, we provide a baseline solution to pixel-selective compression in Section 4.1. The problem of mask generation in CIL is that the optimal generation hyperparameters such as hard thresholds are changing in the dynamic environment (with the increasing number of classes and phases). It is thus desirable to have an adaptive mask generation process. To this end, we propose class-incremental masking (CIM)—a learnable mask generation model, in Section 4.2." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 420, + 490, + 433 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 420, + 490, + 433 + ], + "spans": [ + { + "bbox": [ + 305, + 420, + 490, + 433 + ], + "type": "text", + "content": "4.1. CAM-based Compression Pipeline" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 438, + 545, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 438, + 545, + 581 + ], + "spans": [ + { + "bbox": [ + 304, + 438, + 545, + 581 + ], + "type": "text", + "content": "Generating pixel-level labels for large-scale datasets, e.g., ImageNet [9], is non-trivial. Using class activation maps (CAM) is a naive solution with little computation costs. Its key idea is to make use of the activation of the classification model itself: on the feature maps, activated pixels are more discriminative than non-activated ones for recognizing the object, where \"activated\" means of high activation values and with strong correlation with the classification weights of the object. After localizing activated pixels by CAM, we can generate a 0-1 mask on them, e.g., by hard thresholding their normalized values, and then up-sample the mask to the size of the input image [49]." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 582, + 545, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 582, + 545, + 641 + ], + "spans": [ + { + "bbox": [ + 304, + 582, + 545, + 641 + ], + "type": "text", + "content": "From CAM to 0-1 Mask. We extract CAM in the following steps. Given an image " + }, + { + "bbox": [ + 304, + 582, + 545, + 641 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 304, + 582, + 545, + 641 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 304, + 582, + 545, + 641 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_i" + }, + { + "bbox": [ + 304, + 582, + 545, + 641 + ], + "type": "text", + "content": " and its ground truth class label " + }, + { + "bbox": [ + 304, + 582, + 545, + 641 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 304, + 582, + 545, + 641 + ], + "type": "text", + "content": ", let " + }, + { + "bbox": [ + 304, + 582, + 545, + 641 + ], + "type": "inline_equation", + "content": "F(x; \\theta_i)" + }, + { + "bbox": [ + 304, + 582, + 545, + 641 + ], + "type": "text", + "content": " represent the feature block output by the feature extractor " + }, + { + "bbox": [ + 304, + 582, + 545, + 641 + ], + "type": "inline_equation", + "content": "\\theta_i" + }, + { + "bbox": [ + 304, + 582, + 545, + 641 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 582, + 545, + 641 + ], + "type": "inline_equation", + "content": "\\omega_{i,y}" + }, + { + "bbox": [ + 304, + 582, + 545, + 641 + ], + "type": "text", + "content": " for the classification weights of class " + }, + { + "bbox": [ + 304, + 582, + 545, + 641 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 304, + 582, + 545, + 641 + ], + "type": "text", + "content": " in the classifier " + }, + { + "bbox": [ + 304, + 582, + 545, + 641 + ], + "type": "inline_equation", + "content": "\\omega_i" + }, + { + "bbox": [ + 304, + 582, + 545, + 641 + ], + "type": "text", + "content": ". The CAM of " + }, + { + "bbox": [ + 304, + 582, + 545, + 641 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 304, + 582, + 545, + 641 + ], + "type": "text", + "content": " is:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 315, + 647, + 545, + 672 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 647, + 545, + 672 + ], + "spans": [ + { + "bbox": [ + 315, + 647, + 545, + 672 + ], + "type": "interline_equation", + "content": "\\mathcal {M} ^ {\\mathrm {C A M}} = \\frac {A - \\operatorname* {m i n} (A)}{\\operatorname* {m a x} (A) - \\operatorname* {m i n} (A)}, A = \\omega_ {i, y} ^ {\\top} F (x; \\theta_ {i}), \\tag {1}", + "image_path": "cdf7dbedee6f439c57dde4fa35b8bb223772e3b0d4846f5e9a3daaafb5e3a563.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 677, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 545, + 712 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 677, + 545, + 712 + ], + "type": "inline_equation", + "content": "\\min (\\cdot)" + }, + { + "bbox": [ + 304, + 677, + 545, + 712 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 677, + 545, + 712 + ], + "type": "inline_equation", + "content": "\\max (\\cdot)" + }, + { + "bbox": [ + 304, + 677, + 545, + 712 + ], + "type": "text", + "content": " operations are used for normalization. Then, we upsample " + }, + { + "bbox": [ + 304, + 677, + 545, + 712 + ], + "type": "inline_equation", + "content": "\\mathcal{M}^{\\mathrm{CAM}}" + }, + { + "bbox": [ + 304, + 677, + 545, + 712 + ], + "type": "text", + "content": " to the size of image " + }, + { + "bbox": [ + 304, + 677, + 545, + 712 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 304, + 677, + 545, + 712 + ], + "type": "text", + "content": " and use the same notation." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "11373" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 55, + 67, + 545, + 165 + ], + "blocks": [ + { + "bbox": [ + 55, + 67, + 545, + 165 + ], + "lines": [ + { + "bbox": [ + 55, + 67, + 545, + 165 + ], + "spans": [ + { + "bbox": [ + 55, + 67, + 545, + 165 + ], + "type": "image", + "image_path": "0749e512693154e7a16132e2e3713eb2930a05005bc40316fe6c880c3b65e52d.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 172, + 547, + 206 + ], + "lines": [ + { + "bbox": [ + 46, + 172, + 547, + 206 + ], + "spans": [ + { + "bbox": [ + 46, + 172, + 547, + 206 + ], + "type": "text", + "content": "Figure 2. The proposed compression pipeline. Given an image, we extract its CAM-based (baseline) or CIM-based (ours) mask, threshold it to be a 0-1 mask with a fixed threshold " + }, + { + "bbox": [ + 46, + 172, + 547, + 206 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 46, + 172, + 547, + 206 + ], + "type": "text", + "content": ", and generate a tight bounding box (bbox) that covers all masked pixels. Then, we downsample the pixels outside the bbox and sum the downsampled image up to the masked image to generate the final compressed image." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 215, + 287, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 215, + 287, + 346 + ], + "spans": [ + { + "bbox": [ + 46, + 215, + 287, + 346 + ], + "type": "text", + "content": "Each value in " + }, + { + "bbox": [ + 46, + 215, + 287, + 346 + ], + "type": "inline_equation", + "content": "\\mathcal{M}^{\\mathrm{CAM}}" + }, + { + "bbox": [ + 46, + 215, + 287, + 346 + ], + "type": "text", + "content": " denotes the activation strength of the model at a specific pixel location. Following the way of generating 0-1 masks in weakly-supervised semantic segmentation works [8, 11, 22], we apply a hard threshold " + }, + { + "bbox": [ + 46, + 215, + 287, + 346 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 46, + 215, + 287, + 346 + ], + "type": "text", + "content": " (between 0 and 1) over all values of " + }, + { + "bbox": [ + 46, + 215, + 287, + 346 + ], + "type": "inline_equation", + "content": "\\mathcal{M}^{\\mathrm{CAM}}" + }, + { + "bbox": [ + 46, + 215, + 287, + 346 + ], + "type": "text", + "content": ", and get the 0-1 mask " + }, + { + "bbox": [ + 46, + 215, + 287, + 346 + ], + "type": "inline_equation", + "content": "\\mathcal{M}^{\\tau}" + }, + { + "bbox": [ + 46, + 215, + 287, + 346 + ], + "type": "text", + "content": ": " + }, + { + "bbox": [ + 46, + 215, + 287, + 346 + ], + "type": "inline_equation", + "content": "\\mathcal{M}^{\\tau} = \\mathbb{I}(\\mathcal{M}^{\\mathrm{CAM}} > \\tau)" + }, + { + "bbox": [ + 46, + 215, + 287, + 346 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 215, + 287, + 346 + ], + "type": "inline_equation", + "content": "\\mathbb{I}(\\cdot)" + }, + { + "bbox": [ + 46, + 215, + 287, + 346 + ], + "type": "text", + "content": " is the indicator function. In " + }, + { + "bbox": [ + 46, + 215, + 287, + 346 + ], + "type": "inline_equation", + "content": "\\mathcal{M}^{\\tau}" + }, + { + "bbox": [ + 46, + 215, + 287, + 346 + ], + "type": "text", + "content": ", 1s indicate the locations of discriminative pixels, e.g., foreground pixels, based on which the model makes the prediction. While 0s indicate mostly background pixels or non-discrimination pixels that can be downsampled as they make little contribution to the prediction." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 347, + 287, + 455 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 347, + 287, + 455 + ], + "spans": [ + { + "bbox": [ + 46, + 347, + 287, + 455 + ], + "type": "text", + "content": "After generating 0-1 masks, it is ideal to keep them in the memory as meta information of compression. However, this is not efficient or feasible in CIL. There are two reasons. 1) The space for saving image-size masks is non-negligible. Each mask pixel is a one-bit boolean value, and one mask takes around " + }, + { + "bbox": [ + 46, + 347, + 287, + 455 + ], + "type": "inline_equation", + "content": "\\frac{1}{3 \\times 8} = \\frac{1}{24}" + }, + { + "bbox": [ + 46, + 347, + 287, + 455 + ], + "type": "text", + "content": " of the memory of one RGB image. 2) The mask involves activated regions with irregular shapes. It is thus non-trivial to perform any standard down-sampling algorithm [34] on the remaining regions." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 455, + 287, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 455, + 287, + 515 + ], + "spans": [ + { + "bbox": [ + 46, + 455, + 287, + 515 + ], + "type": "text", + "content": "From 0-1 Masks to Bounding Boxes (BBox). A simple workaround is to generate a tight bounding box (bbox) to cover the positions of 1s in " + }, + { + "bbox": [ + 46, + 455, + 287, + 515 + ], + "type": "inline_equation", + "content": "\\mathcal{M}^{\\tau}" + }, + { + "bbox": [ + 46, + 455, + 287, + 515 + ], + "type": "text", + "content": ", and use the_bbox for compression. Specifically, given " + }, + { + "bbox": [ + 46, + 455, + 287, + 515 + ], + "type": "inline_equation", + "content": "\\mathcal{M}^{\\tau}" + }, + { + "bbox": [ + 46, + 455, + 287, + 515 + ], + "type": "text", + "content": ", we obtain the coordinate representation of the bounding box as:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 522, + 287, + 536 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 522, + 287, + 536 + ], + "spans": [ + { + "bbox": [ + 54, + 522, + 287, + 536 + ], + "type": "interline_equation", + "content": "\\mathcal {B} = \\left[ \\min h, \\min w; \\max h, \\max w \\right] _ {(h, w): \\mathcal {M} ^ {\\tau} (h, w) = 1}, \\tag {2}", + "image_path": "3cdae9d5971a566cf68104883a3663d9f1a64b8b02c4dd5f00f75af7fdc12f39.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 542, + 287, + 625 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 542, + 287, + 625 + ], + "spans": [ + { + "bbox": [ + 46, + 542, + 287, + 625 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 542, + 287, + 625 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 46, + 542, + 287, + 625 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 542, + 287, + 625 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 46, + 542, + 287, + 625 + ], + "type": "text", + "content": " denote the vertical and horizontal coordinates of the 1 on " + }, + { + "bbox": [ + 46, + 542, + 287, + 625 + ], + "type": "inline_equation", + "content": "\\mathcal{M}^{\\tau}" + }, + { + "bbox": [ + 46, + 542, + 287, + 625 + ], + "type": "text", + "content": ", respectively. We highlight that " + }, + { + "bbox": [ + 46, + 542, + 287, + 625 + ], + "type": "inline_equation", + "content": "\\mathcal{B}" + }, + { + "bbox": [ + 46, + 542, + 287, + 625 + ], + "type": "text", + "content": " consists of four integers only and takes negligible memory overhead compared to " + }, + { + "bbox": [ + 46, + 542, + 287, + 625 + ], + "type": "inline_equation", + "content": "\\mathcal{M}^{\\tau}" + }, + { + "bbox": [ + 46, + 542, + 287, + 625 + ], + "type": "text", + "content": ". In addition, we \"reshape\" the irregular shape (of the activated region) in " + }, + { + "bbox": [ + 46, + 542, + 287, + 625 + ], + "type": "inline_equation", + "content": "\\mathcal{M}^{\\tau}" + }, + { + "bbox": [ + 46, + 542, + 287, + 625 + ], + "type": "text", + "content": " into a rectangular " + }, + { + "bbox": [ + 46, + 542, + 287, + 625 + ], + "type": "inline_equation", + "content": "\\mathcal{B}" + }, + { + "bbox": [ + 46, + 542, + 287, + 625 + ], + "type": "text", + "content": ", so our downsampling operation on the pixels outside the rectangular becomes easy." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 625, + 287, + 673 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 625, + 287, + 673 + ], + "spans": [ + { + "bbox": [ + 46, + 625, + 287, + 673 + ], + "type": "text", + "content": "Compression with BBox. Given the image and its bbox on foreground, the compression is implemented by downsampling pixels outside the bbox. Specifically, as illustrated in Figure 2, we compress the image " + }, + { + "bbox": [ + 46, + 625, + 287, + 673 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 46, + 625, + 287, + 673 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 46, + 625, + 287, + 673 + ], + "type": "inline_equation", + "content": "\\tilde{x}" + }, + { + "bbox": [ + 46, + 625, + 287, + 673 + ], + "type": "text", + "content": " as follows," + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 97, + 680, + 287, + 694 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 97, + 680, + 287, + 694 + ], + "spans": [ + { + "bbox": [ + 97, + 680, + 287, + 694 + ], + "type": "interline_equation", + "content": "\\tilde {x} = \\mathcal {M} ^ {\\mathcal {B}} \\odot x + (1 - \\mathcal {M} ^ {\\mathcal {B}}) \\odot x _ {\\eta}, \\tag {3}", + "image_path": "30911f5f2cc933362346d97b428aea4dda812b287c07999a3e38c086a3ccadd6.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 700, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 700, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 700, + 287, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 700, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{M}^{\\mathcal{B}}" + }, + { + "bbox": [ + 47, + 700, + 287, + 713 + ], + "type": "text", + "content": " is the binary mask according to " + }, + { + "bbox": [ + 47, + 700, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{B}" + }, + { + "bbox": [ + 47, + 700, + 287, + 713 + ], + "type": "text", + "content": ", i.e., the val" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 215, + 545, + 275 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 215, + 545, + 275 + ], + "spans": [ + { + "bbox": [ + 304, + 215, + 545, + 275 + ], + "type": "text", + "content": "ues of " + }, + { + "bbox": [ + 304, + 215, + 545, + 275 + ], + "type": "inline_equation", + "content": "\\mathcal{M}^{\\mathcal{B}}" + }, + { + "bbox": [ + 304, + 215, + 545, + 275 + ], + "type": "text", + "content": " are 1 inside " + }, + { + "bbox": [ + 304, + 215, + 545, + 275 + ], + "type": "inline_equation", + "content": "\\mathcal{B}" + }, + { + "bbox": [ + 304, + 215, + 545, + 275 + ], + "type": "text", + "content": ", and 0 otherwise. " + }, + { + "bbox": [ + 304, + 215, + 545, + 275 + ], + "type": "inline_equation", + "content": "x_{\\eta}" + }, + { + "bbox": [ + 304, + 215, + 545, + 275 + ], + "type": "text", + "content": " is the fully downsampled version of " + }, + { + "bbox": [ + 304, + 215, + 545, + 275 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 304, + 215, + 545, + 275 + ], + "type": "text", + "content": " with a downsampling ratio " + }, + { + "bbox": [ + 304, + 215, + 545, + 275 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 304, + 215, + 545, + 275 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 304, + 215, + 545, + 275 + ], + "type": "inline_equation", + "content": "\\eta > 1" + }, + { + "bbox": [ + 304, + 215, + 545, + 275 + ], + "type": "text", + "content": "), " + }, + { + "bbox": [ + 304, + 215, + 545, + 275 + ], + "type": "inline_equation", + "content": "\\odot" + }, + { + "bbox": [ + 304, + 215, + 545, + 275 + ], + "type": "text", + "content": " denotes the element-wise product, and " + }, + { + "bbox": [ + 304, + 215, + 545, + 275 + ], + "type": "inline_equation", + "content": "+" + }, + { + "bbox": [ + 304, + 215, + 545, + 275 + ], + "type": "text", + "content": " denotes the element-wise addition. Both " + }, + { + "bbox": [ + 304, + 215, + 545, + 275 + ], + "type": "inline_equation", + "content": "\\odot" + }, + { + "bbox": [ + 304, + 215, + 545, + 275 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 215, + 545, + 275 + ], + "type": "inline_equation", + "content": "+" + }, + { + "bbox": [ + 304, + 215, + 545, + 275 + ], + "type": "text", + "content": " are applied independently on each RGB channel." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 275, + 545, + 298 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 275, + 545, + 298 + ], + "spans": [ + { + "bbox": [ + 304, + 275, + 545, + 298 + ], + "type": "text", + "content": "The memory allocated for the compressed image " + }, + { + "bbox": [ + 304, + 275, + 545, + 298 + ], + "type": "inline_equation", + "content": "\\tilde{x}" + }, + { + "bbox": [ + 304, + 275, + 545, + 298 + ], + "type": "text", + "content": " is as follows," + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 347, + 306, + 545, + 357 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 347, + 306, + 545, + 357 + ], + "spans": [ + { + "bbox": [ + 347, + 306, + 545, + 357 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} m _ {\\tilde {x}} = \\frac {H _ {\\mathcal {B}} W _ {\\mathcal {B}}}{H W} + \\frac {1}{\\eta} \\left(1 - \\frac {H _ {\\mathcal {B}} W _ {\\mathcal {B}}}{H W}\\right) \\tag {4} \\\\ = 1 - \\left(1 - \\frac {1}{\\eta}\\right) \\cdot \\left(1 - \\frac {H _ {\\mathcal {B}} W _ {\\mathcal {B}}}{H W}\\right), \\\\ \\end{array}", + "image_path": "f16dfffd74e2e71662448977c3ccb28dd23b68939325646a58c83e3277c6fe84.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 364, + 545, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 364, + 545, + 449 + ], + "spans": [ + { + "bbox": [ + 304, + 364, + 545, + 449 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 364, + 545, + 449 + ], + "type": "inline_equation", + "content": "H_{\\mathcal{B}}" + }, + { + "bbox": [ + 304, + 364, + 545, + 449 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 364, + 545, + 449 + ], + "type": "inline_equation", + "content": "W_{\\mathcal{B}}" + }, + { + "bbox": [ + 304, + 364, + 545, + 449 + ], + "type": "text", + "content": " are the height and width of " + }, + { + "bbox": [ + 304, + 364, + 545, + 449 + ], + "type": "inline_equation", + "content": "\\mathcal{B}" + }, + { + "bbox": [ + 304, + 364, + 545, + 449 + ], + "type": "text", + "content": ", respectively. " + }, + { + "bbox": [ + 304, + 364, + 545, + 449 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 304, + 364, + 545, + 449 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 364, + 545, + 449 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 304, + 364, + 545, + 449 + ], + "type": "text", + "content": " are the height and width of the original image " + }, + { + "bbox": [ + 304, + 364, + 545, + 449 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 304, + 364, + 545, + 449 + ], + "type": "text", + "content": ", respectively. " + }, + { + "bbox": [ + 304, + 364, + 545, + 449 + ], + "type": "inline_equation", + "content": "m_{\\tilde{x}}" + }, + { + "bbox": [ + 304, + 364, + 545, + 449 + ], + "type": "text", + "content": " is always smaller than 1 where 1 denotes the memory unit of saving one original image " + }, + { + "bbox": [ + 304, + 364, + 545, + 449 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 304, + 364, + 545, + 449 + ], + "type": "text", + "content": ". Therefore, we can save a larger number of compressed exemplars within the same memory budget. We denote the set of compressed exemplars in the " + }, + { + "bbox": [ + 304, + 364, + 545, + 449 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 364, + 545, + 449 + ], + "type": "text", + "content": "-th phase as " + }, + { + "bbox": [ + 304, + 364, + 545, + 449 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathcal{E}}_i" + }, + { + "bbox": [ + 304, + 364, + 545, + 449 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 449, + 545, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 449, + 545, + 604 + ], + "spans": [ + { + "bbox": [ + 304, + 449, + 545, + 604 + ], + "type": "text", + "content": "Compression Artifacts. The above compression introduces artifacts to the compressed images, i.e., there is a resolution mutation around bounding box edges. From the perspective of spectrum analysis [5,20], such mutation carries noisy and high-frequency components and impairs the model training in subsequent phases. We mitigate the effect of these artifacts by implementing the following data augmentation: in each training epoch, we transform a random subset of " + }, + { + "bbox": [ + 304, + 449, + 545, + 604 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_i" + }, + { + "bbox": [ + 304, + 449, + 545, + 604 + ], + "type": "text", + "content": " into compressed images with CAM-based bounding boxes using the same downsampling ratio. Using this augmentation enables the model to \"simulate\" the training with compressed images and learn to be invariant to compression artifacts." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 611, + 491, + 624 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 611, + 491, + 624 + ], + "spans": [ + { + "bbox": [ + 305, + 611, + 491, + 624 + ], + "type": "text", + "content": "4.2. Class-Incremental Masking (CIM)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 630, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 630, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 630, + 545, + 712 + ], + "type": "text", + "content": "Ideally, the mask generation process needs to adjust at different phases in CIL environments. The process involves two hyperparameters: the masking threshold and the choice of network activation functions. First, for the threshold, searching for its optimum (for all classes) is not trivial. Grid search is intuitive, but it is computationally expensive when the number of classes increases in CIL. Second, for the acti" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "11374" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 48, + 57, + 289, + 171 + ], + "blocks": [ + { + "bbox": [ + 48, + 57, + 289, + 171 + ], + "lines": [ + { + "bbox": [ + 48, + 57, + 289, + 171 + ], + "spans": [ + { + "bbox": [ + 48, + 57, + 289, + 171 + ], + "type": "image", + "image_path": "390c8a83693147661c7a7845ebf4c21b3bbe8dc5e18442187584b423f580a6f4.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 181, + 287, + 247 + ], + "lines": [ + { + "bbox": [ + 46, + 181, + 287, + 247 + ], + "spans": [ + { + "bbox": [ + 46, + 181, + 287, + 247 + ], + "type": "text", + "content": "Figure 3. Our CIM installs an \"extension\" on the backbone (e.g., ResNet-18 [16, 36] with four blocks) by adding a learnable activation function (e.g., PAU [32]) at the position of the original activation function (i.e., ReLU [33]). Such \"extension\" results in a new network branch (in green), whose weight layer parameters are directly copied from the original branch (in blue)." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 259, + 287, + 283 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 259, + 287, + 283 + ], + "spans": [ + { + "bbox": [ + 46, + 259, + 287, + 283 + ], + "type": "text", + "content": "vation function, the standard network of CIL methods uses ReLU [33, 36] and does not optimize it." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 285, + 287, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 285, + 287, + 392 + ], + "spans": [ + { + "bbox": [ + 46, + 285, + 287, + 392 + ], + "type": "text", + "content": "We solve this issue by applying a learnable activation function in addition to the existing ReLU function in the CIL model. Physically, we have one neural network, while logically, we have two models to learn (in each incremental phase): the conventional CIL model with ReLU activations, and the adaptive mask generation model with learnable activations. We thus call our method class-incremental masking (CIM) based CIL. In the following, we elaborate on the network design and optimization pipeline." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 393, + 287, + 584 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 393, + 287, + 584 + ], + "spans": [ + { + "bbox": [ + 46, + 393, + 287, + 584 + ], + "type": "text", + "content": "Network Design. Figure 3 demonstrates an example network architecture in our CIM-based CIL. The proposed CIM extends the network backbone by logically adding a network branch, where only the activation functions are learnable (e.g., Padé Activation Units " + }, + { + "bbox": [ + 46, + 393, + 287, + 584 + ], + "type": "inline_equation", + "content": "(\\mathrm{PAU}^4" + }, + { + "bbox": [ + 46, + 393, + 287, + 584 + ], + "type": "text", + "content": ", [32])) and the parameters of weight layers are copied from the original branch. This design is motivated by the works of He et al. [15] and Bochkovskiy et al. [2], which indicate that layers with learnable activation functions can flexibly process object (localization) information at different network blocks. The difference is that we apply this flexibility to achieve adaptive mask generation for different CIL phases. We denote the CIM parameters (i.e., the parameters in learnable activation functions) in " + }, + { + "bbox": [ + 46, + 393, + 287, + 584 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 393, + 287, + 584 + ], + "type": "text", + "content": "-th phase as " + }, + { + "bbox": [ + 46, + 393, + 287, + 584 + ], + "type": "inline_equation", + "content": "\\phi_i" + }, + { + "bbox": [ + 46, + 393, + 287, + 584 + ], + "type": "text", + "content": ". We optimize the CIL model " + }, + { + "bbox": [ + 46, + 393, + 287, + 584 + ], + "type": "inline_equation", + "content": "(\\theta_i, \\omega_i)" + }, + { + "bbox": [ + 46, + 393, + 287, + 584 + ], + "type": "text", + "content": " and the CIM model " + }, + { + "bbox": [ + 46, + 393, + 287, + 584 + ], + "type": "inline_equation", + "content": "\\phi_i" + }, + { + "bbox": [ + 46, + 393, + 287, + 584 + ], + "type": "text", + "content": " via a global BOP, as elaborated below." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 586, + 287, + 646 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 586, + 287, + 646 + ], + "spans": [ + { + "bbox": [ + 46, + 586, + 287, + 646 + ], + "type": "text", + "content": "Optimization Pipeline. We demonstrate the overall optimization flow in Algorithm 1, which consists of two levels of optimization: task-level and mask-level—the former one for CIL and the latter one for CIM. Note that to maintain an unified notation, we further define " + }, + { + "bbox": [ + 46, + 586, + 287, + 646 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathcal{E}}_0 = \\varnothing" + }, + { + "bbox": [ + 46, + 586, + 287, + 646 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 647, + 287, + 671 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 647, + 287, + 671 + ], + "spans": [ + { + "bbox": [ + 47, + 647, + 287, + 671 + ], + "type": "text", + "content": "1) Task-level Optimization. This level aims to optimize the CIL model " + }, + { + "bbox": [ + 47, + 647, + 287, + 671 + ], + "type": "inline_equation", + "content": "(\\theta_i, \\omega_i)" + }, + { + "bbox": [ + 47, + 647, + 287, + 671 + ], + "type": "text", + "content": " to address the CIL task at hand. It" + } + ] + } + ], + "index": 6 + }, + { + "type": "code", + "bbox": [ + 310, + 88, + 524, + 304 + ], + "blocks": [ + { + "bbox": [ + 312, + 75, + 509, + 87 + ], + "lines": [ + { + "bbox": [ + 312, + 75, + 509, + 87 + ], + "spans": [ + { + "bbox": [ + 312, + 75, + 509, + 87 + ], + "type": "text", + "content": "Algorithm 1: CIM-based CIL (i-th phase, " + }, + { + "bbox": [ + 312, + 75, + 509, + 87 + ], + "type": "inline_equation", + "content": "i \\geq 1" + }, + { + "bbox": [ + 312, + 75, + 509, + 87 + ], + "type": "text", + "content": ")" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "code_caption" + }, + { + "bbox": [ + 310, + 88, + 524, + 304 + ], + "lines": [ + { + "bbox": [ + 310, + 88, + 524, + 304 + ], + "spans": [ + { + "bbox": [ + 310, + 88, + 524, + 304 + ], + "type": "text", + "content": "Input: New data " + }, + { + "bbox": [ + 310, + 88, + 524, + 304 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_i" + }, + { + "bbox": [ + 310, + 88, + 524, + 304 + ], + "type": "text", + "content": " ; old compressed exemplars " + }, + { + "bbox": [ + 310, + 88, + 524, + 304 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathcal{E}}_{0:i - 1}(\\tilde{\\mathcal{E}}_0 = \\emptyset)" + }, + { + "bbox": [ + 310, + 88, + 524, + 304 + ], + "type": "text", + "content": " ; last-phase CIL model " + }, + { + "bbox": [ + 310, + 88, + 524, + 304 + ], + "type": "inline_equation", + "content": "(\\theta_{i - 1},\\omega_{i - 1})" + }, + { + "bbox": [ + 310, + 88, + 524, + 304 + ], + "type": "text", + "content": " and CIM model " + }, + { + "bbox": [ + 310, + 88, + 524, + 304 + ], + "type": "inline_equation", + "content": "\\phi_{i - 1}" + }, + { + "bbox": [ + 310, + 88, + 524, + 304 + ], + "type": "inline_equation", + "content": "(\\theta_0,\\omega_0" + }, + { + "bbox": [ + 310, + 88, + 524, + 304 + ], + "type": "text", + "content": " are random parameters, " + }, + { + "bbox": [ + 310, + 88, + 524, + 304 + ], + "type": "inline_equation", + "content": "\\phi_0" + }, + { + "bbox": [ + 310, + 88, + 524, + 304 + ], + "type": "text", + "content": " is set to ReLU). Output: New compressed exemplars " + }, + { + "bbox": [ + 310, + 88, + 524, + 304 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathcal{E}}_i" + }, + { + "bbox": [ + 310, + 88, + 524, + 304 + ], + "type": "text", + "content": " ; updated CIL model " + }, + { + "bbox": [ + 310, + 88, + 524, + 304 + ], + "type": "inline_equation", + "content": "(\\theta_i,\\omega_i)" + }, + { + "bbox": [ + 310, + 88, + 524, + 304 + ], + "type": "text", + "content": " and CIM model " + }, + { + "bbox": [ + 310, + 88, + 524, + 304 + ], + "type": "inline_equation", + "content": "\\phi_{i}" + }, + { + "bbox": [ + 310, + 88, + 524, + 304 + ], + "type": "text", + "content": " \n1 Initialize " + }, + { + "bbox": [ + 310, + 88, + 524, + 304 + ], + "type": "inline_equation", + "content": "(\\theta_{i},\\omega_{i})" + }, + { + "bbox": [ + 310, + 88, + 524, + 304 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 310, + 88, + 524, + 304 + ], + "type": "inline_equation", + "content": "(\\theta_{i - 1},\\omega_{i - 1})" + }, + { + "bbox": [ + 310, + 88, + 524, + 304 + ], + "type": "text", + "content": " . \n2 Initialize " + }, + { + "bbox": [ + 310, + 88, + 524, + 304 + ], + "type": "inline_equation", + "content": "\\phi_{i}" + }, + { + "bbox": [ + 310, + 88, + 524, + 304 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 310, + 88, + 524, + 304 + ], + "type": "inline_equation", + "content": "\\phi_{i - 1}" + }, + { + "bbox": [ + 310, + 88, + 524, + 304 + ], + "type": "text", + "content": " . \n3 for epochs do. \n4 Train " + }, + { + "bbox": [ + 310, + 88, + 524, + 304 + ], + "type": "inline_equation", + "content": "(\\theta_{i},\\omega_{i})" + }, + { + "bbox": [ + 310, + 88, + 524, + 304 + ], + "type": "text", + "content": " using " + }, + { + "bbox": [ + 310, + 88, + 524, + 304 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathcal{E}}_{0:i - 1}\\cup \\mathcal{D}_i" + }, + { + "bbox": [ + 310, + 88, + 524, + 304 + ], + "type": "text", + "content": " by Eq. 5; \n5 Compress " + }, + { + "bbox": [ + 310, + 88, + 524, + 304 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_i" + }, + { + "bbox": [ + 310, + 88, + 524, + 304 + ], + "type": "text", + "content": " into " + }, + { + "bbox": [ + 310, + 88, + 524, + 304 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathcal{D}}_i^\\phi" + }, + { + "bbox": [ + 310, + 88, + 524, + 304 + ], + "type": "text", + "content": " using " + }, + { + "bbox": [ + 310, + 88, + 524, + 304 + ], + "type": "inline_equation", + "content": "\\phi_{i}" + }, + { + "bbox": [ + 310, + 88, + 524, + 304 + ], + "type": "text", + "content": " . \n6 Temporarily update " + }, + { + "bbox": [ + 310, + 88, + 524, + 304 + ], + "type": "inline_equation", + "content": "\\theta_{i}" + }, + { + "bbox": [ + 310, + 88, + 524, + 304 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 310, + 88, + 524, + 304 + ], + "type": "inline_equation", + "content": "\\theta_i^+" + }, + { + "bbox": [ + 310, + 88, + 524, + 304 + ], + "type": "text", + "content": " using " + }, + { + "bbox": [ + 310, + 88, + 524, + 304 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathcal{E}}_{0:i - 1}\\cup \\tilde{\\mathcal{D}}_i^\\phi" + }, + { + "bbox": [ + 310, + 88, + 524, + 304 + ], + "type": "text", + "content": " by Eq.7; \n7 Learn " + }, + { + "bbox": [ + 310, + 88, + 524, + 304 + ], + "type": "inline_equation", + "content": "\\phi_{i}" + }, + { + "bbox": [ + 310, + 88, + 524, + 304 + ], + "type": "text", + "content": " using " + }, + { + "bbox": [ + 310, + 88, + 524, + 304 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_i" + }, + { + "bbox": [ + 310, + 88, + 524, + 304 + ], + "type": "text", + "content": " by Eq.9; \n8 end \n9 Compress " + }, + { + "bbox": [ + 310, + 88, + 524, + 304 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_i" + }, + { + "bbox": [ + 310, + 88, + 524, + 304 + ], + "type": "text", + "content": " into " + }, + { + "bbox": [ + 310, + 88, + 524, + 304 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathcal{D}}_i" + }, + { + "bbox": [ + 310, + 88, + 524, + 304 + ], + "type": "text", + "content": " using the learned " + }, + { + "bbox": [ + 310, + 88, + 524, + 304 + ], + "type": "inline_equation", + "content": "\\phi_{i}" + }, + { + "bbox": [ + 310, + 88, + 524, + 304 + ], + "type": "text", + "content": " by Eq. 3 \n0 Select exemplars " + }, + { + "bbox": [ + 310, + 88, + 524, + 304 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathcal{E}}_i" + }, + { + "bbox": [ + 310, + 88, + 524, + 304 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 310, + 88, + 524, + 304 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathcal{D}}_i" + }, + { + "bbox": [ + 310, + 88, + 524, + 304 + ], + "type": "text", + "content": " by, e.g., herding [36]." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "code_body" + } + ], + "index": 9, + "sub_type": "algorithm" + }, + { + "bbox": [ + 306, + 329, + 379, + 339 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 329, + 379, + 339 + ], + "spans": [ + { + "bbox": [ + 306, + 329, + 379, + 339 + ], + "type": "text", + "content": "can be written as:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 315, + 349, + 545, + 373 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 349, + 545, + 373 + ], + "spans": [ + { + "bbox": [ + 315, + 349, + 545, + 373 + ], + "type": "interline_equation", + "content": "\\left(\\theta_ {i}, \\omega_ {i}\\right) \\leftarrow \\left(\\theta_ {i}, \\omega_ {i}\\right) - \\lambda \\nabla_ {\\left(\\theta , \\omega\\right)} \\mathcal {L} _ {\\mathrm {C I L}} \\left(\\tilde {\\mathcal {E}} _ {0: i - 1} \\cup \\mathcal {D} _ {i}; \\theta_ {i}, \\omega_ {i}\\right), \\tag {5}", + "image_path": "eae6bf889f4bcbc5fdf91cb6cb5b8cf5bc7403a33b77dba0e6f6c186da8b2953.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 373, + 545, + 421 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 373, + 545, + 421 + ], + "spans": [ + { + "bbox": [ + 304, + 373, + 545, + 421 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 373, + 545, + 421 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 304, + 373, + 545, + 421 + ], + "type": "text", + "content": " is the learning rate. We follow the implementation of CIL training loss " + }, + { + "bbox": [ + 304, + 373, + 545, + 421 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{CIL}}" + }, + { + "bbox": [ + 304, + 373, + 545, + 421 + ], + "type": "text", + "content": " in baseline methods [23, 42, 46, 50]. This means that we use different training losses when plugging CIM into different baseline methods." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 421, + 545, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 421, + 545, + 456 + ], + "spans": [ + { + "bbox": [ + 304, + 421, + 545, + 456 + ], + "type": "text", + "content": "2) Mask-level Optimization. This level aims to optimize the CIM model " + }, + { + "bbox": [ + 304, + 421, + 545, + 456 + ], + "type": "inline_equation", + "content": "\\phi_{i}" + }, + { + "bbox": [ + 304, + 421, + 545, + 456 + ], + "type": "text", + "content": " to produce adaptive compression masks. It is formulated as a local BOP:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 318, + 465, + 545, + 484 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 465, + 545, + 484 + ], + "spans": [ + { + "bbox": [ + 318, + 465, + 545, + 484 + ], + "type": "interline_equation", + "content": "\\min _ {\\phi_ {i}} \\left[ \\mathcal {L} _ {\\text {v a l}} \\left(\\mathcal {D} _ {i}; \\theta_ {i} ^ {*}, \\omega_ {i}\\right) + \\mu \\mathcal {R} \\left(\\phi_ {i}\\right) \\right], \\tag {6a}", + "image_path": "5755cc81602b46f5fc396e7eddb0f974caea7fa9fe9640bc36f4f4f517f7b595.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 319, + 487, + 545, + 508 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 487, + 545, + 508 + ], + "spans": [ + { + "bbox": [ + 319, + 487, + 545, + 508 + ], + "type": "interline_equation", + "content": "\\text {s . t .} \\theta_ {i} ^ {*} = \\underset {\\theta_ {i}} {\\arg \\min } \\mathcal {L} _ {\\operatorname {t r a i n}} \\left(\\tilde {\\mathcal {E}} _ {0: i - 1} \\cup \\tilde {\\mathcal {D}} _ {i} \\left(\\phi_ {i}\\right); \\theta_ {i}, \\omega_ {i}\\right). \\tag {6b}", + "image_path": "f16ecdc8515a2d9e1f919da96aae6c3c2504d1f86f1f1584662cc6c41802fd0e.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 515, + 545, + 598 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 515, + 545, + 598 + ], + "spans": [ + { + "bbox": [ + 304, + 515, + 545, + 598 + ], + "type": "text", + "content": "Eq. 6b denotes an inner-level optimization. It trains " + }, + { + "bbox": [ + 304, + 515, + 545, + 598 + ], + "type": "inline_equation", + "content": "\\theta_{i}" + }, + { + "bbox": [ + 304, + 515, + 545, + 598 + ], + "type": "text", + "content": " with the data " + }, + { + "bbox": [ + 304, + 515, + 545, + 598 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathcal{D}}_i(\\phi_i)" + }, + { + "bbox": [ + 304, + 515, + 545, + 598 + ], + "type": "text", + "content": " compressed by using " + }, + { + "bbox": [ + 304, + 515, + 545, + 598 + ], + "type": "inline_equation", + "content": "\\phi_{i}" + }, + { + "bbox": [ + 304, + 515, + 545, + 598 + ], + "type": "text", + "content": ", and converges as " + }, + { + "bbox": [ + 304, + 515, + 545, + 598 + ], + "type": "inline_equation", + "content": "\\theta_{i}^{*}" + }, + { + "bbox": [ + 304, + 515, + 545, + 598 + ], + "type": "text", + "content": ". Eq. 6a denotes an outer-level optimization. It is based on the validation loss derived by " + }, + { + "bbox": [ + 304, + 515, + 545, + 598 + ], + "type": "inline_equation", + "content": "\\theta_{i}^{*}" + }, + { + "bbox": [ + 304, + 515, + 545, + 598 + ], + "type": "text", + "content": " on the original data " + }, + { + "bbox": [ + 304, + 515, + 545, + 598 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_i" + }, + { + "bbox": [ + 304, + 515, + 545, + 598 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 304, + 515, + 545, + 598 + ], + "type": "inline_equation", + "content": "\\mathcal{R}(\\phi_i)" + }, + { + "bbox": [ + 304, + 515, + 545, + 598 + ], + "type": "text", + "content": " is a constraint representing the memory limitation and " + }, + { + "bbox": [ + 304, + 515, + 545, + 598 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 304, + 515, + 545, + 598 + ], + "type": "text", + "content": " is its weight. In the following, we elaborate on the implementation details for the two levels." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 599, + 545, + 671 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 599, + 545, + 671 + ], + "spans": [ + { + "bbox": [ + 304, + 599, + 545, + 671 + ], + "type": "text", + "content": "In the inner-level optimization, we train a temporary CIL model with compressed data. Specifically, we first compress new-class data " + }, + { + "bbox": [ + 304, + 599, + 545, + 671 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_i" + }, + { + "bbox": [ + 304, + 599, + 545, + 671 + ], + "type": "text", + "content": " into " + }, + { + "bbox": [ + 304, + 599, + 545, + 671 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathcal{D}}_i(\\phi_i)" + }, + { + "bbox": [ + 304, + 599, + 545, + 671 + ], + "type": "text", + "content": " using the generated masks by the CIM model " + }, + { + "bbox": [ + 304, + 599, + 545, + 671 + ], + "type": "inline_equation", + "content": "\\phi_i" + }, + { + "bbox": [ + 304, + 599, + 545, + 671 + ], + "type": "text", + "content": ". Then, we implement the inner-level optimization as a one-step gradient descent (using the CIL training loss) as:" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 328, + 678, + 545, + 693 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 328, + 678, + 545, + 693 + ], + "spans": [ + { + "bbox": [ + 328, + 678, + 545, + 693 + ], + "type": "interline_equation", + "content": "\\theta_ {i} ^ {+} \\leftarrow \\theta_ {i} - \\beta_ {1} \\nabla_ {\\theta} \\mathcal {L} _ {\\mathrm {C I L}} \\left(\\tilde {\\mathcal {E}} _ {0: i - 1} \\cup \\tilde {\\mathcal {D}} _ {i} \\left(\\phi_ {i}\\right); \\theta_ {i}, \\omega_ {i}\\right), \\tag {7}", + "image_path": "3778394dce4c7059eaf5187f16881f427c2cd78fc28133c55129ab950323ad95.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 306, + 701, + 449, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 701, + 449, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 701, + 449, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 306, + 701, + 449, + 713 + ], + "type": "inline_equation", + "content": "\\beta_{1}" + }, + { + "bbox": [ + 306, + 701, + 449, + 713 + ], + "type": "text", + "content": " is the learning rate for " + }, + { + "bbox": [ + 306, + 701, + 449, + 713 + ], + "type": "inline_equation", + "content": "\\theta_{i}" + }, + { + "bbox": [ + 306, + 701, + 449, + 713 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 46, + 683, + 287, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 683, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 683, + 287, + 713 + ], + "type": "text", + "content": "4PAU uses a rational function of given degrees " + }, + { + "bbox": [ + 46, + 683, + 287, + 713 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 46, + 683, + 287, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 683, + 287, + 713 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 46, + 683, + 287, + 713 + ], + "type": "text", + "content": ", i.e., " + }, + { + "bbox": [ + 46, + 683, + 287, + 713 + ], + "type": "inline_equation", + "content": "(a_0 + a_1x + \\dots + a_mx^m) / (1 + b_1x + \\dots + b_nx^n)" + }, + { + "bbox": [ + 46, + 683, + 287, + 713 + ], + "type": "text", + "content": ", and can be parameterized by " + }, + { + "bbox": [ + 46, + 683, + 287, + 713 + ], + "type": "inline_equation", + "content": "a_0, \\dots, a_m" + }, + { + "bbox": [ + 46, + 683, + 287, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 683, + 287, + 713 + ], + "type": "inline_equation", + "content": "b_1, \\dots, b_n" + }, + { + "bbox": [ + 46, + 683, + 287, + 713 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "11375" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 287, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 287, + 133 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 287, + 133 + ], + "type": "text", + "content": "The aim of the outer-level optimization is to optimize " + }, + { + "bbox": [ + 47, + 72, + 287, + 133 + ], + "type": "inline_equation", + "content": "\\phi_{i}" + }, + { + "bbox": [ + 47, + 72, + 287, + 133 + ], + "type": "text", + "content": " such that the temporary model " + }, + { + "bbox": [ + 47, + 72, + 287, + 133 + ], + "type": "inline_equation", + "content": "(\\theta_{i}^{+},\\omega_{i})" + }, + { + "bbox": [ + 47, + 72, + 287, + 133 + ], + "type": "text", + "content": " (trained with compressed data) has a low validation loss on the original data " + }, + { + "bbox": [ + 47, + 72, + 287, + 133 + ], + "type": "inline_equation", + "content": "D_{i}" + }, + { + "bbox": [ + 47, + 72, + 287, + 133 + ], + "type": "text", + "content": ". To achieve this, we back-propagate the loss on the original data " + }, + { + "bbox": [ + 47, + 72, + 287, + 133 + ], + "type": "inline_equation", + "content": "D_{i}" + }, + { + "bbox": [ + 47, + 72, + 287, + 133 + ], + "type": "text", + "content": " to update " + }, + { + "bbox": [ + 47, + 72, + 287, + 133 + ], + "type": "inline_equation", + "content": "\\phi_{i}" + }, + { + "bbox": [ + 47, + 72, + 287, + 133 + ], + "type": "text", + "content": " as:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 63, + 141, + 287, + 156 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 141, + 287, + 156 + ], + "spans": [ + { + "bbox": [ + 63, + 141, + 287, + 156 + ], + "type": "interline_equation", + "content": "\\phi_ {i} \\leftarrow \\phi_ {i} - \\beta_ {2} \\nabla_ {\\phi} \\left[ \\mathcal {L} _ {\\mathrm {C E}} \\left(\\mathcal {D} _ {i}; \\theta_ {i} ^ {+}, \\omega_ {i}\\right) + \\mu \\mathcal {R} \\left(\\phi_ {i}\\right) \\right], \\tag {8}", + "image_path": "7737f09af3540c8411991938bd2f9344aa5dcf6527be6189aca209eacfb0f1c7.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 164, + 287, + 248 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 164, + 287, + 248 + ], + "spans": [ + { + "bbox": [ + 46, + 164, + 287, + 248 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 164, + 287, + 248 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{CE}}" + }, + { + "bbox": [ + 46, + 164, + 287, + 248 + ], + "type": "text", + "content": " denotes softmax cross-entropy loss and " + }, + { + "bbox": [ + 46, + 164, + 287, + 248 + ], + "type": "inline_equation", + "content": "\\beta_{2}" + }, + { + "bbox": [ + 46, + 164, + 287, + 248 + ], + "type": "text", + "content": " is the learning rate for " + }, + { + "bbox": [ + 46, + 164, + 287, + 248 + ], + "type": "inline_equation", + "content": "\\phi_{i}" + }, + { + "bbox": [ + 46, + 164, + 287, + 248 + ], + "type": "text", + "content": ". This trains " + }, + { + "bbox": [ + 46, + 164, + 287, + 248 + ], + "type": "inline_equation", + "content": "\\phi_{i}" + }, + { + "bbox": [ + 46, + 164, + 287, + 248 + ], + "type": "text", + "content": " to capture the most discriminative features of new-class images. The constraint " + }, + { + "bbox": [ + 46, + 164, + 287, + 248 + ], + "type": "inline_equation", + "content": "\\mathcal{R}(\\phi_i)" + }, + { + "bbox": [ + 46, + 164, + 287, + 248 + ], + "type": "text", + "content": " is implemented as a " + }, + { + "bbox": [ + 46, + 164, + 287, + 248 + ], + "type": "inline_equation", + "content": "\\ell_2" + }, + { + "bbox": [ + 46, + 164, + 287, + 248 + ], + "type": "text", + "content": "-regularization term on the generated mask by " + }, + { + "bbox": [ + 46, + 164, + 287, + 248 + ], + "type": "inline_equation", + "content": "\\phi_{i}" + }, + { + "bbox": [ + 46, + 164, + 287, + 248 + ], + "type": "text", + "content": ". The motivation for the regularization term is to make the mask coverage smaller, thus the compressed images take less memory." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 248, + 287, + 308 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 248, + 287, + 308 + ], + "spans": [ + { + "bbox": [ + 47, + 248, + 287, + 308 + ], + "type": "text", + "content": "We empirically observe that by the above optimization flow, the output activation maps by " + }, + { + "bbox": [ + 47, + 248, + 287, + 308 + ], + "type": "inline_equation", + "content": "\\phi_{i}" + }, + { + "bbox": [ + 47, + 248, + 287, + 308 + ], + "type": "text", + "content": " are easy to collapse, i.e., different images have the same map. To solve this issue, we add a cross-entropy loss term about " + }, + { + "bbox": [ + 47, + 248, + 287, + 308 + ], + "type": "inline_equation", + "content": "\\phi_{i}" + }, + { + "bbox": [ + 47, + 248, + 287, + 308 + ], + "type": "text", + "content": " in Eq. 8 to regularize it to produce image-specific activation maps:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 318, + 287, + 348 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 318, + 287, + 348 + ], + "spans": [ + { + "bbox": [ + 56, + 318, + 287, + 348 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\phi_ {i} \\leftarrow \\phi_ {i} - \\beta_ {2} \\nabla_ {\\phi} \\left[ \\mathcal {L} _ {\\mathrm {C E}} \\left(\\mathcal {D} _ {i}; \\theta_ {i} ^ {+}, \\omega_ {i}\\right) + \\mu \\mathcal {R} (\\phi_ {i}) \\right. \\tag {9} \\\\ + \\mu^ {\\prime} \\mathcal {L} _ {\\mathrm {C E}} \\left(\\tilde {\\mathcal {E}} _ {0: i - 1} \\cup \\mathcal {D} _ {i}; \\theta_ {i}, \\phi_ {i}, \\omega_ {i}\\right) ], \\\\ \\end{array}", + "image_path": "2dd6156b00668ceb508c0bd97c17082ea6cdd78cbb8f9ecca1336111905d48b7.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 357, + 141, + 369 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 357, + 141, + 369 + ], + "spans": [ + { + "bbox": [ + 47, + 357, + 141, + 369 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 357, + 141, + 369 + ], + "type": "inline_equation", + "content": "\\mu^{\\prime}" + }, + { + "bbox": [ + 47, + 357, + 141, + 369 + ], + "type": "text", + "content": " is the weight." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 369, + 287, + 536 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 369, + 287, + 536 + ], + "spans": [ + { + "bbox": [ + 46, + 369, + 287, + 536 + ], + "type": "text", + "content": "Limitations. Our CIM learns to generate adaptive masks for exemplar compression in CIL. It has three limitations that are left as future work. 1) It is not able to adjust any previous-phase exemplars, as the validation data (the original data of these exemplars) are not accessible anymore. 2) It introduces hundreds of activation parameters to the CIL model, although this is not a significant overhead compared to model parameters. Please check detailed overhead analyses in the supplementary materials. 3) Image compression is not that meaningful for low-resolution datasets (e.g., " + }, + { + "bbox": [ + 46, + 369, + 287, + 536 + ], + "type": "inline_equation", + "content": "32 \\times 32" + }, + { + "bbox": [ + 46, + 369, + 287, + 536 + ], + "type": "text", + "content": " CIFAR-100). It is because the memory taken by the compression parameters (e.g., the parameters of CIM) and the RGB pixels of a low-resolution image are comparable. Using the memory to save more images is more meaningful." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 547, + 127, + 560 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 547, + 127, + 560 + ], + "spans": [ + { + "bbox": [ + 47, + 547, + 127, + 560 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 567, + 287, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 567, + 287, + 628 + ], + "spans": [ + { + "bbox": [ + 46, + 567, + 287, + 628 + ], + "type": "text", + "content": "We incorporate CIM into two baseline CIL methods (i.e., DER [46] and FOSTER [42]) and boost their model performances consistently on three datasets. Below, we introduce datasets and experiment settings (Section 5.1), followed by results and analyses (Section 5.2)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 635, + 174, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 635, + 174, + 647 + ], + "spans": [ + { + "bbox": [ + 47, + 635, + 174, + 647 + ], + "type": "text", + "content": "5.1. Experimental Settings" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 653, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 287, + 713 + ], + "type": "text", + "content": "Datasets. We conduct experiments on three standard CIL benchmarks with high-resolution images. 1) Food-101 [3] consists of 101 food categories with 750 training and 250 test samples per category. All images have a maximum side length of 512 pixels. 2) ImageNet-1000 [9] is a large-scale" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "text", + "content": "dataset with 1,000 classes and each class has around 1,300 training and 50 test samples. 3) ImageNet-100 is a 100-class subset randomly sampled from ImageNet-1000 with a fixed NumPy [14] random seed (1993), following [17]. We provide other details of these datasets, e.g., image sizes and pre-processing methods, in the supplementary materials." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 144, + 546, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 144, + 546, + 312 + ], + "spans": [ + { + "bbox": [ + 304, + 144, + 546, + 312 + ], + "type": "text", + "content": "Protocols. We use two protocols: learning from scratch (LFS) and learning from half (LFH), following recent CIL works [42, 46]. In LFS, the model observes the same number of classes in all " + }, + { + "bbox": [ + 304, + 144, + 546, + 312 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 144, + 546, + 312 + ], + "type": "text", + "content": " phases, where " + }, + { + "bbox": [ + 304, + 144, + 546, + 312 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 144, + 546, + 312 + ], + "type": "text", + "content": " is optionally 5, 10, and 20. In LFH, the model is trained on half of the classes (e.g., 500 classes for ImageNet-1000) in the 1-th phase. Then, it learns the remaining classes evenly in the subsequent " + }, + { + "bbox": [ + 304, + 144, + 546, + 312 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 144, + 546, + 312 + ], + "type": "text", + "content": " phases, where " + }, + { + "bbox": [ + 304, + 144, + 546, + 312 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 144, + 546, + 312 + ], + "type": "text", + "content": " can be 5, 10, and 25. In both protocols, after the training of each phase we evaluate the resultant model on the test data of all seen classes. Our final report includes the average accuracy over all phases and the last-phase accuracy which indicates the degree of model forgetting. We run each experiment three times and report the average results." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 312, + 546, + 480 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 312, + 546, + 480 + ], + "spans": [ + { + "bbox": [ + 304, + 312, + 546, + 480 + ], + "type": "text", + "content": "Memory Budget. There are two memory budget settings. 1) In the \"fixed\" setting, we remove some old-class exemplars when new exemplars from the current phase are added in the memory to maintain the \"fixed memory budget\". In this setting, we set the total memory to be 2,020 samples for Food-101 and 2,000 samples for ImageNet-100. For ImageNet-1000, we have two options—5,000 samples and 20,000 samples. 2) In the \"growing\" setting, a constant memory budget is allocated for each class across all phases and hence extra memory is appended when new classes come. In this setting, we set the budget to be 20 samples per class for all datasets. Following [42, 46], we apply the \"fixed\" setting in LFS experiments and the \"growing\" setting in LFH experiments." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 481, + 545, + 673 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 481, + 545, + 673 + ], + "spans": [ + { + "bbox": [ + 304, + 481, + 545, + 673 + ], + "type": "text", + "content": "Implementation Details. Our implementation is based on the standard deep learning library PyTorch [35] and image processing library OpenCV [4]. Following [42, 43, 46], we use an 18-layer ResNet [16] as the network backbone " + }, + { + "bbox": [ + 304, + 481, + 545, + 673 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 304, + 481, + 545, + 673 + ], + "type": "text", + "content": " and a fully-connected layer as the classifier " + }, + { + "bbox": [ + 304, + 481, + 545, + 673 + ], + "type": "inline_equation", + "content": "\\omega" + }, + { + "bbox": [ + 304, + 481, + 545, + 673 + ], + "type": "text", + "content": " in all experiments. We use the same CIL training hyperparameters as in related works [17, 42, 46] for fair comparison: 1) there are 200 epochs in 1-st phase and 170 epochs in the subsequent phases; 2) the learning rate " + }, + { + "bbox": [ + 304, + 481, + 545, + 673 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 304, + 481, + 545, + 673 + ], + "type": "text", + "content": " is initialized as 0.1 and decreases to zero with a cosine annealing scheduler [28]; 3) the SGD optimizer is deployed, with momentum factor set to 0.9 and weight decay set to 0.0005. For compression-related hyperparameters, we set the masking threshold " + }, + { + "bbox": [ + 304, + 481, + 545, + 673 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 304, + 481, + 545, + 673 + ], + "type": "text", + "content": " as 0.6 and the downsampling ratio " + }, + { + "bbox": [ + 304, + 481, + 545, + 673 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 304, + 481, + 545, + 673 + ], + "type": "text", + "content": " as 4.0. To build the CIM model, we apply PAUs with degrees " + }, + { + "bbox": [ + 304, + 481, + 545, + 673 + ], + "type": "inline_equation", + "content": "m = 5" + }, + { + "bbox": [ + 304, + 481, + 545, + 673 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 481, + 545, + 673 + ], + "type": "inline_equation", + "content": "n = 4" + }, + { + "bbox": [ + 304, + 481, + 545, + 673 + ], + "type": "text", + "content": " as learnable activation layers. For the optimization of the CIM" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 305, + 683, + 545, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 683, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 683, + 545, + 713 + ], + "type": "text", + "content": "5Please note that we measure the memory by the number of original images. Each compressed exemplar in CIM takes less memory than the original image (Eq. 4), resulting in more exemplars in the same memory." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "11376" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 53, + 70, + 541, + 220 + ], + "blocks": [ + { + "bbox": [ + 53, + 70, + 541, + 220 + ], + "lines": [ + { + "bbox": [ + 53, + 70, + 541, + 220 + ], + "spans": [ + { + "bbox": [ + 53, + 70, + 541, + 220 + ], + "type": "table", + "html": "
MethodLearning from Scratch (LFS)Learning from Half (LFH)
Food-101ImageNet-100Food-101ImageNet-100
N=51020510205102551025
iCaRL [36]69.6662.1856.7073.9067.0662.3660.1353.4246.8762.5359.8852.97
WA [48]70.9463.6958.4574.6468.6263.2063.5557.6052.4865.7563.7158.34
PODNet [12]68.0361.2447.3872.1463.9653.6975.3770.0165.3275.5474.3368.33
AANets [25]69.4661.5948.8372.9865.7755.3676.0771.2266.9376.9675.5871.78
DER [46]73.8870.7664.3978.5076.1273.7978.1373.45-79.0877.73-
DER w/ ours75.6373.0969.1779.6377.5775.3679.2575.76-80.3079.05-
FOSTER [42]75.0372.7266.73\\( 79.93^{\\dagger} \\)\\( 76.55^{\\dagger} \\)74.4979.0875.0768.08\\( 80.07^{\\dagger} \\)77.54\\( 72.40^{\\times} \\)
FOSTER w/ ours76.4474.8570.2080.5877.9475.2379.7676.8670.5080.9378.6675.74
", + "image_path": "622a8a85609712e4825ca61c0bfb9ea6637485efaf592c09b576824aaf19f70c.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 60, + 220, + 539, + 243 + ], + "lines": [ + { + "bbox": [ + 60, + 220, + 539, + 243 + ], + "spans": [ + { + "bbox": [ + 60, + 220, + 539, + 243 + ], + "type": "text", + "content": "† The paper of FOSTER [42] did not report the numerical results for " + }, + { + "bbox": [ + 60, + 220, + 539, + 243 + ], + "type": "inline_equation", + "content": "N = 5 / 10" + }, + { + "bbox": [ + 60, + 220, + 539, + 243 + ], + "type": "text", + "content": " (LFS) and " + }, + { + "bbox": [ + 60, + 220, + 539, + 243 + ], + "type": "inline_equation", + "content": "N = 5" + }, + { + "bbox": [ + 60, + 220, + 539, + 243 + ], + "type": "text", + "content": " (LFH) on ImageNet-100. We run these experiments using the public code (released by authors) and report the reproduced results." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_footnote" + }, + { + "bbox": [ + 60, + 243, + 493, + 255 + ], + "lines": [ + { + "bbox": [ + 60, + 243, + 493, + 255 + ], + "spans": [ + { + "bbox": [ + 60, + 243, + 493, + 255 + ], + "type": "text", + "content": "* Our reproduced result (72.40) is significantly higher than the original result (69.34) reported in the paper of FOSTER." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 48, + 300, + 287, + 422 + ], + "blocks": [ + { + "bbox": [ + 46, + 258, + 547, + 293 + ], + "lines": [ + { + "bbox": [ + 46, + 258, + 547, + 293 + ], + "spans": [ + { + "bbox": [ + 46, + 258, + 547, + 293 + ], + "type": "text", + "content": "Table 1. Average accuracies " + }, + { + "bbox": [ + 46, + 258, + 547, + 293 + ], + "type": "inline_equation", + "content": "(\\%)" + }, + { + "bbox": [ + 46, + 258, + 547, + 293 + ], + "type": "text", + "content": " of two top-performing CIL methods [42, 46] with and without our CIM-CIL plugged-in, and other four baselines [12, 25, 36, 48], on two datasets (Food-101 and ImageNet-100) and using two protocols (learning from scratch (LFS) and learning from half (LFH)). Due to the space limits, we report the " + }, + { + "bbox": [ + 46, + 258, + 547, + 293 + ], + "type": "inline_equation", + "content": "95\\%" + }, + { + "bbox": [ + 46, + 258, + 547, + 293 + ], + "type": "text", + "content": " confidence intervals for these results in the supplementary materials." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 48, + 300, + 287, + 422 + ], + "lines": [ + { + "bbox": [ + 48, + 300, + 287, + 422 + ], + "spans": [ + { + "bbox": [ + 48, + 300, + 287, + 422 + ], + "type": "table", + "html": "
Memory BudgetMethodN=5N=10
Avg.LastAvg.Last
M=20kiCaRL [36]44.3627.7838.4022.70
WA [48]58.3750.6254.1045.66
DER [46]67.4959.7566.7358.62
FOSTER [42]69.2164.8868.3460.14
FOSTER w/ ours69.9366.0569.5362.07
M=5kFOSTER57.1949.4254.7244.96
FOSTER w/ ours61.3754.4659.4850.83
", + "image_path": "b23490493f89c7f2690e20a2d27b79b921384b0beaa8d54c8eb9ec4f809b71aa.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 482, + 287, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 482, + 287, + 555 + ], + "spans": [ + { + "bbox": [ + 46, + 482, + 287, + 555 + ], + "type": "text", + "content": "model " + }, + { + "bbox": [ + 46, + 482, + 287, + 555 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 46, + 482, + 287, + 555 + ], + "type": "text", + "content": " (i.e., the mask-level optimization), we initially set " + }, + { + "bbox": [ + 46, + 482, + 287, + 555 + ], + "type": "inline_equation", + "content": "\\beta_{1}" + }, + { + "bbox": [ + 46, + 482, + 287, + 555 + ], + "type": "text", + "content": " as 0.1 and " + }, + { + "bbox": [ + 46, + 482, + 287, + 555 + ], + "type": "inline_equation", + "content": "\\beta_{2}" + }, + { + "bbox": [ + 46, + 482, + 287, + 555 + ], + "type": "text", + "content": " as 0.01 and reduce them to zero following the scheduler of " + }, + { + "bbox": [ + 46, + 482, + 287, + 555 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 46, + 482, + 287, + 555 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 46, + 482, + 287, + 555 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 46, + 482, + 287, + 555 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 482, + 287, + 555 + ], + "type": "inline_equation", + "content": "\\mu^{\\prime}" + }, + { + "bbox": [ + 46, + 482, + 287, + 555 + ], + "type": "text", + "content": " is set to 0.1 and 0.2, respectively. To smooth the training, we clip the gradient norm of " + }, + { + "bbox": [ + 46, + 482, + 287, + 555 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 46, + 482, + 287, + 555 + ], + "type": "text", + "content": " to be no more than 1. We report the result of hyperparameter sensitivity analysis in the supplementary materials." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 563, + 168, + 576 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 563, + 168, + 576 + ], + "spans": [ + { + "bbox": [ + 47, + 563, + 168, + 576 + ], + "type": "text", + "content": "5.2. Results and Analyses" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 582, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 582, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 582, + 288, + 715 + ], + "type": "text", + "content": "Comparing with the State-of-the-art. In Table 1, we summarize the experimental results on two datasets (Food-101 and ImageNet-100) and in two CIL protocols (LFS and LFH). From the table, we have the following observations. 1) Our CIM-based CIL consistently improves the state-ofthe-art method FOSTER [42] with clear margins in all settings. E.g., our method surpasses it by an average of 1.4 percentage points on ImageNet-100, and 2.0 percentage points on Food-101. 2) Our CIM-based CIL achieves more significant improvements when " + }, + { + "bbox": [ + 46, + 582, + 288, + 715 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 582, + 288, + 715 + ], + "type": "text", + "content": " becomes larger, e.g., on ImageNet-100 (LFH), our method improves FOSTER by" + } + ] + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 313, + 300, + 541, + 474 + ], + "blocks": [ + { + "bbox": [ + 46, + 427, + 288, + 473 + ], + "lines": [ + { + "bbox": [ + 46, + 427, + 288, + 473 + ], + "spans": [ + { + "bbox": [ + 46, + 427, + 288, + 473 + ], + "type": "text", + "content": "Table 2. Average and last accuracies " + }, + { + "bbox": [ + 46, + 427, + 288, + 473 + ], + "type": "inline_equation", + "content": "(\\%)" + }, + { + "bbox": [ + 46, + 427, + 288, + 473 + ], + "type": "text", + "content": " on ImageNet-1000 of FOSTER with and without our method plugged-in, and other three baselines [36, 46, 48]. We show two memory budgets, " + }, + { + "bbox": [ + 46, + 427, + 288, + 473 + ], + "type": "inline_equation", + "content": "M = 20k" + }, + { + "bbox": [ + 46, + 427, + 288, + 473 + ], + "type": "text", + "content": " (upper block) and " + }, + { + "bbox": [ + 46, + 427, + 288, + 473 + ], + "type": "inline_equation", + "content": "M = 5k" + }, + { + "bbox": [ + 46, + 427, + 288, + 473 + ], + "type": "text", + "content": " (lower block), in the LFS setting." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 313, + 300, + 541, + 474 + ], + "lines": [ + { + "bbox": [ + 313, + 300, + 541, + 474 + ], + "spans": [ + { + "bbox": [ + 313, + 300, + 541, + 474 + ], + "type": "table", + "html": "
Ablation MethodFood-101ImageNet-100
N=10201020
1 Baseline72.7266.7376.5572.37
2 Artifact Aug.71.3866.0375.6371.45
3 Full Comp.73.0367.3876.9273.26
4 Random Acti.73.1067.5476.8873.54
5 Center Acti.73.2967.8876.7873.82
6 Class Acti.73.7668.6577.2174.67
7 Phase-wise τ73.8369.1777.0674.78
8 Joint Train73.4469.0177.3474.59
9 BOP (ours)74.8570.2077.9475.23
10 LastBlock Only74.5569.8777.7274.86
11 Fg Compressed75.0270.1377.8775.46
", + "image_path": "375693eb67a69177a651a4c1c7e2e5e3983b4e9d6f00954375df4f8f75ff0b64.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 477, + 545, + 501 + ], + "lines": [ + { + "bbox": [ + 305, + 477, + 545, + 501 + ], + "spans": [ + { + "bbox": [ + 305, + 477, + 545, + 501 + ], + "type": "text", + "content": "Table 3. Average accuracies " + }, + { + "bbox": [ + 305, + 477, + 545, + 501 + ], + "type": "inline_equation", + "content": "(\\%)" + }, + { + "bbox": [ + 305, + 477, + 545, + 501 + ], + "type": "text", + "content": " of different ablation methods. The experiments are conducted in the LFS setting." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 510, + 545, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 510, + 545, + 605 + ], + "spans": [ + { + "bbox": [ + 304, + 510, + 545, + 605 + ], + "type": "text", + "content": "0.9 and 3.3 percentage points when " + }, + { + "bbox": [ + 304, + 510, + 545, + 605 + ], + "type": "inline_equation", + "content": "N = 5" + }, + { + "bbox": [ + 304, + 510, + 545, + 605 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 510, + 545, + 605 + ], + "type": "inline_equation", + "content": "N = 25" + }, + { + "bbox": [ + 304, + 510, + 545, + 605 + ], + "type": "text", + "content": ", respectively. 3) Our CIM-based CIL achieves greater improvements consistently on Food-101 (than ImageNet-100). It improves baselines by 2.1 percentage points on Food-101, while the improvement is 1.4 on ImageNet-100 (" + }, + { + "bbox": [ + 304, + 510, + 545, + 605 + ], + "type": "inline_equation", + "content": "N = 10" + }, + { + "bbox": [ + 304, + 510, + 545, + 605 + ], + "type": "text", + "content": ", LFS). It shows that our method is particularly effective when the representative visual cues of a class are from some of its components, e.g., the \"cream\" of the class \"cake\"." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 605, + 545, + 701 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 545, + 701 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 545, + 701 + ], + "type": "text", + "content": "Table 2 shows the results on the large-scale dataset ImageNet-1000 in different memory settings (" + }, + { + "bbox": [ + 304, + 605, + 545, + 701 + ], + "type": "inline_equation", + "content": "M = 20k" + }, + { + "bbox": [ + 304, + 605, + 545, + 701 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 605, + 545, + 701 + ], + "type": "inline_equation", + "content": "M = 5k" + }, + { + "bbox": [ + 304, + 605, + 545, + 701 + ], + "type": "text", + "content": "). We can see that our CIM-based CIL improves FOSTER consistently. It is impressive that it achieves more improvements in the more strict memory setting (" + }, + { + "bbox": [ + 304, + 605, + 545, + 701 + ], + "type": "inline_equation", + "content": "M = 5k" + }, + { + "bbox": [ + 304, + 605, + 545, + 701 + ], + "type": "text", + "content": "). Specifically, it boosts the average accuracy of FOSTER by 4.5 percentage points when " + }, + { + "bbox": [ + 304, + 605, + 545, + 701 + ], + "type": "inline_equation", + "content": "M = 5k" + }, + { + "bbox": [ + 304, + 605, + 545, + 701 + ], + "type": "text", + "content": ", significantly higher than that of " + }, + { + "bbox": [ + 304, + 605, + 545, + 701 + ], + "type": "inline_equation", + "content": "M = 20k" + }, + { + "bbox": [ + 304, + 605, + 545, + 701 + ], + "type": "text", + "content": " (1.0)." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 701, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 701, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 306, + 701, + 545, + 714 + ], + "type": "text", + "content": "Ablation Study. Table 3 shows the ablation results. First" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "11377" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 51, + 70, + 286, + 161 + ], + "blocks": [ + { + "bbox": [ + 51, + 70, + 286, + 161 + ], + "lines": [ + { + "bbox": [ + 51, + 70, + 286, + 161 + ], + "spans": [ + { + "bbox": [ + 51, + 70, + 286, + 161 + ], + "type": "table", + "html": "
MethodImageNet-100ImageNet-1000
N=6112661126
LUCIR baseline71.2269.6767.4565.2362.4359.88
w/ Mnemonics73.3072.1771.5066.1563.1263.08
w/ MRDC73.6272.8170.4467.6765.6062.74
w/ ours74.0573.7672.8468.0366.5463.77
", + "image_path": "34b9068f8fd9872de0690fef02744f71377e2cad5de5f43ce08898974231e137.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 47, + 163, + 287, + 187 + ], + "lines": [ + { + "bbox": [ + 47, + 163, + 287, + 187 + ], + "spans": [ + { + "bbox": [ + 47, + 163, + 287, + 187 + ], + "type": "text", + "content": "Table 4. Comparing with Mnemonics [27] and MRDC [43]. We plug each of them in baseline LUCIR [17] for fair comparison." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 197, + 289, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 197, + 289, + 544 + ], + "spans": [ + { + "bbox": [ + 47, + 197, + 289, + 544 + ], + "type": "text", + "content": "block: baselines. Row 1 is for the baseline FOSTER [42]. Row 2 shows the results of adding artifact augmentation (see Section 4.1). It shows directly apply this augmentation does not improve and even impair the model. Please note the models in below blocks all use this augmentation. Second block: activation methods. Rows 3-6 show the results of using different activation methods to compress exemplars. Row 3 is to downsample all pixels (i.e., no region is activated). Row 4 is to randomly select activation regions. Row 5 is to activate only the center region " + }, + { + "bbox": [ + 47, + 197, + 289, + 544 + ], + "type": "inline_equation", + "content": "(\\frac{1}{4}" + }, + { + "bbox": [ + 47, + 197, + 289, + 544 + ], + "type": "text", + "content": " of the original image), while row 6 is to use naive CAM. Comparing them to row 1, we can see that using naively compressed exemplars can improve CIL models. Row 4 outperforms rows 3-5, validating that it is more reliable to use the model's activation to generate compressed exemplars. Third block: optimization methods. Rows 7-9 are on top of Row 6 and are the results of applying different optimization strategies. Row 7 is to manually select " + }, + { + "bbox": [ + 47, + 197, + 289, + 544 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 47, + 197, + 289, + 544 + ], + "type": "text", + "content": " using a held-out set (" + }, + { + "bbox": [ + 47, + 197, + 289, + 544 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 47, + 197, + 289, + 544 + ], + "type": "text", + "content": " of the dataset). Row 8 is to jointly train CIL and CIM models (for each input batch). Row 9 is the proposed method of using a global BOP. Fourth block: two variants of CIM-based CIL. Rows 10-11 are two variants of row 9. In Row 10, only the activation layers in the last block of CIM are learnable, and previous blocks use ReLU. Compared to row 9, row 10 shows slightly worse performance. Row 11 shows the version of adding a weak downsampling " + }, + { + "bbox": [ + 47, + 197, + 289, + 544 + ], + "type": "inline_equation", + "content": "(\\eta' = 2.0)" + }, + { + "bbox": [ + 47, + 197, + 289, + 544 + ], + "type": "text", + "content": " on discriminative regions, based on which more compressed exemplars are saved. It results in comparable performance to row 9 but increases costs." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 544, + 287, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 544, + 287, + 676 + ], + "spans": [ + { + "bbox": [ + 46, + 544, + 287, + 676 + ], + "type": "text", + "content": "Comparing with Other Compression-based Methods. Table 4 shows our results comparing to two compression-based methods: Mnemonics [27] and MRDC [43]. We can see that our method consistently outperforms them in all settings. This is because our method does not sacrifice the discriminativeness of exemplars while improving the number (variance) of exemplars in a phase-adaptive manner. However, the two related methods either keep a fixed number of exemplars in the memory [27] or use uniform image compression without considering the properties of specific classes in different incremental phases [43]." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 677, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 288, + 714 + ], + "type": "text", + "content": "Visualizations (CAM vs. CIM). Figure 4 gives two visualization examples, \"Afghan hound\" and \"indigo bird\", each with the activation map as well as the bounding boxes." + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 326, + 70, + 525, + 190 + ], + "blocks": [ + { + "bbox": [ + 326, + 70, + 525, + 190 + ], + "lines": [ + { + "bbox": [ + 326, + 70, + 525, + 190 + ], + "spans": [ + { + "bbox": [ + 326, + 70, + 525, + 190 + ], + "type": "image", + "image_path": "21467937e61d1a11cdb13cd77972360696d8f03f49efae8b77e1b2e9adad253a.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 193, + 545, + 216 + ], + "lines": [ + { + "bbox": [ + 305, + 193, + 545, + 216 + ], + "spans": [ + { + "bbox": [ + 305, + 193, + 545, + 216 + ], + "type": "text", + "content": "Figure 4. Visualizations (CAM vs. CIM). The experimental setting is " + }, + { + "bbox": [ + 305, + 193, + 545, + 216 + ], + "type": "inline_equation", + "content": "N = 10" + }, + { + "bbox": [ + 305, + 193, + 545, + 216 + ], + "type": "text", + "content": " (LFS) on ImageNet-100." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 306, + 221, + 542, + 294 + ], + "blocks": [ + { + "bbox": [ + 306, + 221, + 542, + 294 + ], + "lines": [ + { + "bbox": [ + 306, + 221, + 542, + 294 + ], + "spans": [ + { + "bbox": [ + 306, + 221, + 542, + 294 + ], + "type": "table", + "html": "
MetricSmallMiddleLarge
Mean of #Exemplars39.4038.3034.77
Last Acc. (%, baseline)66.1368.4069.93
Last Acc. (%, ours)70.0071.1072.26
Improvement (%)+3.87+3.65+2.33
", + "image_path": "02ca84354d937caf9f9375f1866a70aeac36f0bd7b3478269512012eb3d29839.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 305, + 297, + 546, + 342 + ], + "lines": [ + { + "bbox": [ + 305, + 297, + 546, + 342 + ], + "spans": [ + { + "bbox": [ + 305, + 297, + 546, + 342 + ], + "type": "text", + "content": "Table 5. Results for small, middle and large objects in the setting of " + }, + { + "bbox": [ + 305, + 297, + 546, + 342 + ], + "type": "inline_equation", + "content": "N = 10" + }, + { + "bbox": [ + 305, + 297, + 546, + 342 + ], + "type": "text", + "content": " (LFS) on ImageNet-100. \"Mean of #Exemplars\" denotes the average number of saved exemplars by our method. The baseline (FOSTER [42]) has this number as 20 for all classes." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 352, + 545, + 388 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 352, + 545, + 388 + ], + "spans": [ + { + "bbox": [ + 304, + 352, + 545, + 388 + ], + "type": "text", + "content": "The first column shows their respective confusing classes appearing in earlier phases. CIM learns to focus on the discriminative (i.e., dissimilar to confusing classes) regions." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 388, + 546, + 520 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 388, + 546, + 520 + ], + "spans": [ + { + "bbox": [ + 304, + 388, + 546, + 520 + ], + "type": "text", + "content": "Results of Different-Size Objects. Table 5 shows the results for small, middle, and large objects. These size categorization is according to ImageNet Object Localization Challenge [19]. We calculate the bbox coverage for each class and take top 30 classes with highest coverages as \"large\", near 30 classes with lowest coverages as \"small\" and the rest 40 classes as \"middle\". It is intriguing that our method achieves the highest improvement (over baseline) for small objects. Our explanation is that small objects benefit more from image compression (than large ones), as their images contain more background pixels to downsample." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 531, + 384, + 544 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 531, + 384, + 544 + ], + "spans": [ + { + "bbox": [ + 306, + 531, + 384, + 544 + ], + "type": "text", + "content": "6. Conclusions" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 552, + 545, + 673 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 552, + 545, + 673 + ], + "spans": [ + { + "bbox": [ + 304, + 552, + 545, + 673 + ], + "type": "text", + "content": "We introduced a novel exemplar compression method for CIL, allowing us to save more representative exemplars but not increase memory budget. We achieved this compression by downsampling non-discriminative pixels of the image bounded by CAM masks. To generate adaptive masks, we proposed a novel method CIM that explicitly parameterizes a mask generation model and optimizes it in an end-to-end manner across incremental phases. Our method achieves consistent performance improvements over multiple baselines and can be taken as a flexible plug-and-play module." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 677, + 546, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 677, + 546, + 714 + ], + "spans": [ + { + "bbox": [ + 305, + 677, + 546, + 714 + ], + "type": "text", + "content": "Acknowledgements. The author gratefully acknowledges the support by the Lee Kong Chian (LKC) Fellowship fund awarded by Singapore Management University." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "11378" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 134 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 134 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 134 + ], + "type": "text", + "content": "[1] Davide Abati, Jakub Tomczak, Tijmen Blankevoort, Simone Calderara, Rita Cucchiara, and Babak Ehteshami Bejnordi. Conditional channel gated networks for task-aware continual learning. In CVPR, 2020. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 135, + 288, + 168 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 135, + 288, + 168 + ], + "spans": [ + { + "bbox": [ + 53, + 135, + 288, + 168 + ], + "type": "text", + "content": "[2] Alexey Bochkovskiy, Chien-Yao Wang, and Hong-Yuan Mark Liao. Yolov4: Optimal speed and accuracy of object detection. arXiv, 2020. 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 168, + 287, + 201 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 168, + 287, + 201 + ], + "spans": [ + { + "bbox": [ + 53, + 168, + 287, + 201 + ], + "type": "text", + "content": "[3] Lukas Bossard, Matthieu Guillaumin, and Luc Van Gool. Food-101 - mining discriminative components with random forests. In ECCV, 2014. 2, 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 202, + 287, + 224 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 202, + 287, + 224 + ], + "spans": [ + { + "bbox": [ + 53, + 202, + 287, + 224 + ], + "type": "text", + "content": "[4] Gary Bradski. The opencv library. Dr. Dobb's Journal: Software Tools for the Professional Programmer, 2000. 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 224, + 287, + 246 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 224, + 287, + 246 + ], + "spans": [ + { + "bbox": [ + 53, + 224, + 287, + 246 + ], + "type": "text", + "content": "[5] Kenneth R Castleman. Digital image processing. Prentice Hall Press, 1996. 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 247, + 287, + 280 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 247, + 287, + 280 + ], + "spans": [ + { + "bbox": [ + 53, + 247, + 287, + 280 + ], + "type": "text", + "content": "[6] George Cazenavette, Tongzhou Wang, Antonio Torralba, Alexei A Efros, and Jun-Yan Zhu. Dataset distillation by matching training trajectories. In CVPR, 2022. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 280, + 287, + 313 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 280, + 287, + 313 + ], + "spans": [ + { + "bbox": [ + 53, + 280, + 287, + 313 + ], + "type": "text", + "content": "[7] Can Chen, Xi Chen, Chen Ma, Zixuan Liu, and Xue Liu. Gradient-based bi-level optimization for deep learning: A survey. arXiv, 2022. 2, 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 314, + 287, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 314, + 287, + 357 + ], + "spans": [ + { + "bbox": [ + 53, + 314, + 287, + 357 + ], + "type": "text", + "content": "[8] Zhaozheng Chen, Tan Wang, Xiongwei Wu, Xian-Sheng Hua, Hanwang Zhang, and Qianru Sun. Class re-activation maps for weakly-supervised semantic segmentation. In CVPR, 2022. 3, 4" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 357, + 287, + 390 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 357, + 287, + 390 + ], + "spans": [ + { + "bbox": [ + 53, + 357, + 287, + 390 + ], + "type": "text", + "content": "[9] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In CVPR, 2009. 1, 2, 3, 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 391, + 287, + 424 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 391, + 287, + 424 + ], + "spans": [ + { + "bbox": [ + 48, + 391, + 287, + 424 + ], + "type": "text", + "content": "[10] Prithviraj Dhar, Rajat Vikram Singh, Kuan-Chuan Peng, Ziyan Wu, and Rama Chellappa. Learning without memorizing. In CVPR, 2019. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 424, + 287, + 458 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 424, + 287, + 458 + ], + "spans": [ + { + "bbox": [ + 48, + 424, + 287, + 458 + ], + "type": "text", + "content": "[11] Zhang Dong, Zhang Hanwang, Tang Jinhui, Hua Xiansheng, and Sun Qianru. Causal intervention for weakly supervised semantic segmentation. In NeurIPS, 2020. 4" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 458, + 287, + 501 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 458, + 287, + 501 + ], + "spans": [ + { + "bbox": [ + 48, + 458, + 287, + 501 + ], + "type": "text", + "content": "[12] Arthur Douillard, Matthieu Cord, Charles Ollion, Thomas Robert, and Eduardo Valle. Podnet: Pooled outputs distillation for small-tasks incremental learning. In ECCV, 2020. 1, 2, 7" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 502, + 287, + 535 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 502, + 287, + 535 + ], + "spans": [ + { + "bbox": [ + 48, + 502, + 287, + 535 + ], + "type": "text", + "content": "[13] Chelsea Finn, Pieter Abbeel, and Sergey Levine. Model-agnostic meta-learning for fast adaptation of deep networks. In ICML, 2017. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 536, + 287, + 580 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 536, + 287, + 580 + ], + "spans": [ + { + "bbox": [ + 48, + 536, + 287, + 580 + ], + "type": "text", + "content": "[14] Charles R Harris, K Jarrod Millman, Stefan J Van Der Walt, Ralf Gommers, Pauli Virtanen, David Cournapeau, Eric Wieser, Julian Taylor, Sebastian Berg, Nathaniel J Smith, et al. Array programming with numpy. Nature, 2020. 6" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 580, + 287, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 580, + 287, + 613 + ], + "spans": [ + { + "bbox": [ + 48, + 580, + 287, + 613 + ], + "type": "text", + "content": "[15] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Delving deep into rectifiers: Surpassing human-level performance on imagenet classification. In ICCV, 2015. 5" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 48, + 613, + 287, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 613, + 287, + 646 + ], + "spans": [ + { + "bbox": [ + 48, + 613, + 287, + 646 + ], + "type": "text", + "content": "[16] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, 2016. 5, 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 48, + 646, + 287, + 680 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 646, + 287, + 680 + ], + "spans": [ + { + "bbox": [ + 48, + 646, + 287, + 680 + ], + "type": "text", + "content": "[17] Saihui Hou, Xinyu Pan, Chen Change Loy, Zilei Wang, and Dahua Lin. Learning a unified classifier incrementally via rebalancing. In CVPR, 2019. 1, 2, 6, 8" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "type": "text", + "content": "[18] Shenyang Huang, Vincent François-Lavet, and Guillaume Rabusseau. Neural architecture search for class-incremental learning. arXiv, 2019. 2" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 38, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "text", + "content": "[19] Imagenet object localization challenge. https://www.kaggle.com/competitions/imagenet-object-localization-challenge/.8" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 107, + 545, + 129 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 107, + 545, + 129 + ], + "spans": [ + { + "bbox": [ + 307, + 107, + 545, + 129 + ], + "type": "text", + "content": "[20] Anil K Jain. Fundamentals of digital image processing. Prentice-Hall, Inc., 1989. 4" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 129, + 545, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 129, + 545, + 184 + ], + "spans": [ + { + "bbox": [ + 307, + 129, + 545, + 184 + ], + "type": "text", + "content": "[21] James Kirkpatrick, Razvan Pascanu, Neil Rabinowitz, Joel Veness, Guillaume Desjardins, Andrei A Rusu, Kieran Milan, John Quan, Tiago Ramalho, Agnieszka Grabska-Barwinska, et al. Overcoming catastrophic forgetting in neural networks. PNAS, 2017. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 185, + 545, + 218 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 185, + 545, + 218 + ], + "spans": [ + { + "bbox": [ + 307, + 185, + 545, + 218 + ], + "type": "text", + "content": "[22] Jungbeom Lee, Eunji Kim, and Sungroh Yoon. Antiadversarily manipulated attributions for weakly and semi-supervised semantic segmentation. In CVPR, 2021. 3, 4" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 219, + 545, + 240 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 219, + 545, + 240 + ], + "spans": [ + { + "bbox": [ + 307, + 219, + 545, + 240 + ], + "type": "text", + "content": "[23] Zhizhong Li and Derek Hoiem. Learning without forgetting. PAMI, 2017. 2, 5" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 241, + 545, + 275 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 241, + 545, + 275 + ], + "spans": [ + { + "bbox": [ + 307, + 241, + 545, + 275 + ], + "type": "text", + "content": "[24] Yaoyao Liu, Yingying Li, Bernt Schiele, and Qianru Sun. Online hyperparameter optimization for class-incremental learning. In AAAi, 2023. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 275, + 545, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 275, + 545, + 308 + ], + "spans": [ + { + "bbox": [ + 307, + 275, + 545, + 308 + ], + "type": "text", + "content": "[25] Yaoyao Liu, Bernt Schiele, and Qianru Sun. Adaptive aggregation networks for class-incremental learning. In CVPR, 2021. 1, 2, 3, 7" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 309, + 545, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 309, + 545, + 342 + ], + "spans": [ + { + "bbox": [ + 307, + 309, + 545, + 342 + ], + "type": "text", + "content": "[26] Yaoyao Liu, Bernt Schiele, and Qianru Sun. Rmm: Reinforced memory management for class-incremental learning. In NeurIPS, 2021. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 342, + 545, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 342, + 545, + 376 + ], + "spans": [ + { + "bbox": [ + 307, + 342, + 545, + 376 + ], + "type": "text", + "content": "[27] Yaoyao Liu, Yuting Su, An-An Liu, Bernt Schiele, and Qianru Sun. Mnemonics training: Multi-class incremental learning without forgetting. In CVPR, 2020. 1, 2, 3, 8" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 376, + 545, + 399 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 376, + 545, + 399 + ], + "spans": [ + { + "bbox": [ + 307, + 376, + 545, + 399 + ], + "type": "text", + "content": "[28] Ilya Loshchilov and Frank Hutter. Sgdr: Stochastic gradient descent with warm restarts. arXiv, 2016. 6" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 307, + 399, + 545, + 433 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 399, + 545, + 433 + ], + "spans": [ + { + "bbox": [ + 307, + 399, + 545, + 433 + ], + "type": "text", + "content": "[29] Dougal Maclaurin, David Duvenaud, and Ryan Adams. Gradient-based hyperparameter optimization through reversible learning. In ICML, 2015. 3" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 307, + 433, + 545, + 477 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 433, + 545, + 477 + ], + "spans": [ + { + "bbox": [ + 307, + 433, + 545, + 477 + ], + "type": "text", + "content": "[30] Michael McCloskey and Neal J Cohen. Catastrophic interference in connectionist networks: The sequential learning problem. In *Psychology of Learning and Motivation*. Elsevier, 1989. 2" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 307, + 478, + 545, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 478, + 545, + 521 + ], + "spans": [ + { + "bbox": [ + 307, + 478, + 545, + 521 + ], + "type": "text", + "content": "[31] Ken McRae and Phil A Hetherington. Catastrophic interference is eliminated in pretrained networks. In Proceedings of the 15th Annual Conference of the Cognitive Science Society, 1993. 1, 2" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 307, + 522, + 545, + 556 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 522, + 545, + 556 + ], + "spans": [ + { + "bbox": [ + 307, + 522, + 545, + 556 + ], + "type": "text", + "content": "[32] Alejandro Molina, Patrick Schramowski, and Kristian Kersting. Padé activation units: End-to-end learning of flexible activation functions in deep networks. In ICLR, 2019. 5" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 307, + 556, + 545, + 579 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 556, + 545, + 579 + ], + "spans": [ + { + "bbox": [ + 307, + 556, + 545, + 579 + ], + "type": "text", + "content": "[33] Vinod Nair and Geoffrey E Hinton. Rectified linear units improve restricted boltzmann machines. In ICML, 2010. 5" + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 307, + 579, + 545, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 579, + 545, + 613 + ], + "spans": [ + { + "bbox": [ + 307, + 579, + 545, + 613 + ], + "type": "text", + "content": "[34] J Anthony Parker, Robert V Kenyon, and Donald E Troxel. Comparison of interpolating methods for image resampling. IEEE Transactions on Medical Imaging, 1983. 4" + } + ] + } + ], + "index": 35 + }, + { + "bbox": [ + 307, + 613, + 545, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 613, + 545, + 668 + ], + "spans": [ + { + "bbox": [ + 307, + 613, + 545, + 668 + ], + "type": "text", + "content": "[35] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. In NeurIPS, 2019. 6" + } + ] + } + ], + "index": 36 + }, + { + "bbox": [ + 307, + 669, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 669, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 307, + 669, + 545, + 712 + ], + "type": "text", + "content": "[36] Sylvestre-Alvise Rebuffi, Alexander Kolesnikov, Georg Sperl, and Christoph H Lampert. iCaRL: Incremental classifier and representation learning. In CVPR, 2017. 1, 2, 3, 5, 7" + } + ] + } + ], + "index": 37 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "11379" + } + ] + } + ], + "index": 39 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 578 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "text", + "content": "[37] Andrei A Rusu, Neil C Rabinowitz, Guillaume Desjardins, Hubert Soyer, James Kirkpatrick, Koray Kavukcuoglu, Razvan Pascanu, and Raia Hadsell. Progressive neural networks. arXiv, 2016. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 118, + 286, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 118, + 286, + 162 + ], + "spans": [ + { + "bbox": [ + 49, + 118, + 286, + 162 + ], + "type": "text", + "content": "[38] Ramprasaath R Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra. Grad-cam: Visual explanations from deep networks via gradient-based localization. In ICCV, 2017. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 163, + 286, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 163, + 286, + 195 + ], + "spans": [ + { + "bbox": [ + 49, + 163, + 286, + 195 + ], + "type": "text", + "content": "[39] Christian Simon, Piotr Koniusz, and Mehrtash Harandi. On learning the geodesic path for incremental learning. In CVPR, 2021. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 197, + 286, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 197, + 286, + 239 + ], + "spans": [ + { + "bbox": [ + 49, + 197, + 286, + 239 + ], + "type": "text", + "content": "[40] Ankur Sinha, Pekka Malo, and Kalyanmoy Deb. A review on bilevel optimization: from classical to evolutionary approaches and applications. IEEE Transactions on Evolutionary Computation, 2017. 1, 2, 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 49, + 241, + 286, + 263 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 241, + 286, + 263 + ], + "spans": [ + { + "bbox": [ + 49, + 241, + 286, + 263 + ], + "type": "text", + "content": "[41] Gregory K Wallace. The jpeg still picture compression standard. Communications of the ACM, 1991. 1, 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 49, + 265, + 286, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 265, + 286, + 297 + ], + "spans": [ + { + "bbox": [ + 49, + 265, + 286, + 297 + ], + "type": "text", + "content": "[42] Fu-Yun Wang, Da-Wei Zhou, Han-Jia Ye, and De-Chuan Zhan. Foster: Feature boosting and compression for class incremental learning. arXiv, 2022. 1, 2, 3, 5, 6, 7, 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 49, + 298, + 286, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 298, + 286, + 342 + ], + "spans": [ + { + "bbox": [ + 49, + 298, + 286, + 342 + ], + "type": "text", + "content": "[43] Liyuan Wang, Xingxing Zhang, Kuo Yang, Longhui Yu, Chongxuan Li, Lanqing Hong, Shifeng Zhang, Zhenguo Li, Yi Zhong, and Jun Zhu. Memory replay with data compression for continual learning. In ICLR, 2022. 1, 2, 3, 6, 8" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 49, + 343, + 286, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 343, + 286, + 376 + ], + "spans": [ + { + "bbox": [ + 49, + 343, + 286, + 376 + ], + "type": "text", + "content": "[44] Yue Wu, Yinpeng Chen, Lijuan Wang, Yuancheng Ye, Zicheng Liu, Yandong Guo, and Yun Fu. Large scale incremental learning. In CVPR, 2019. 1, 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 49, + 377, + 286, + 399 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 377, + 286, + 399 + ], + "spans": [ + { + "bbox": [ + 49, + 377, + 286, + 399 + ], + "type": "text", + "content": "[45] Ju Xu and Zhanxing Zhu. Reinforced continual learning. In NeurIPS, 2018. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 49, + 399, + 286, + 432 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 399, + 286, + 432 + ], + "spans": [ + { + "bbox": [ + 49, + 399, + 286, + 432 + ], + "type": "text", + "content": "[46] Shipeng Yan, Jiangwei Xie, and Xuming He. Der: Dynamically expandable representation for class incremental learning. In CVPR, 2021. 1, 2, 3, 5, 6, 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 49, + 434, + 286, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 434, + 286, + 464 + ], + "spans": [ + { + "bbox": [ + 49, + 434, + 286, + 464 + ], + "type": "text", + "content": "[47] Friedemann Zenke, Ben Poole, and Surya Ganguli. Continual learning through synaptic intelligence. In ICML, 2017. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 49, + 468, + 286, + 500 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 468, + 286, + 500 + ], + "spans": [ + { + "bbox": [ + 49, + 468, + 286, + 500 + ], + "type": "text", + "content": "[48] Bowen Zhao, Xi Xiao, Guojun Gan, Bin Zhang, and Shu-Tao Xia. Maintaining discrimination and fairness in class incremental learning. In CVPR, 2020. 1, 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 49, + 502, + 286, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 502, + 286, + 533 + ], + "spans": [ + { + "bbox": [ + 49, + 502, + 286, + 533 + ], + "type": "text", + "content": "[49] Bolei Zhou, Aditya Khosla, Agata Lapedriza, Aude Oliva, and Antonio Torralba. Learning deep features for discriminative localization. In CVPR, 2016. 1, 2, 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 49, + 535, + 286, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 535, + 286, + 578 + ], + "spans": [ + { + "bbox": [ + 49, + 535, + 286, + 578 + ], + "type": "text", + "content": "[50] Da-Wei Zhou, Han-Jia Ye, and De-Chuan Zhan. Co-transport for class-incremental learning. In Proceedings of the 29th ACM International Conference on Multimedia, 2021. 5" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "11380" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/CloSET_ Modeling Clothed Humans on Continuous Surface With Explicit Template Decomposition/4725afdd-f941-4b99-a289-05a751ed5c4e_content_list.json b/2023/CloSET_ Modeling Clothed Humans on Continuous Surface With Explicit Template Decomposition/4725afdd-f941-4b99-a289-05a751ed5c4e_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..bf0aa74602b602a96f80d18addce01d1fffdd41a --- /dev/null +++ b/2023/CloSET_ Modeling Clothed Humans on Continuous Surface With Explicit Template Decomposition/4725afdd-f941-4b99-a289-05a751ed5c4e_content_list.json @@ -0,0 +1,1345 @@ +[ + { + "type": "text", + "text": "CloSET: Modeling Clothed Humans on Continuous Surface with Explicit Template Decomposition", + "text_level": 1, + "bbox": [ + 158, + 128, + 816, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Hongwen Zhang $^{1}$ Siyou Lin $^{1}$ Ruizhi Shao $^{1}$ Yuxiang Zhang $^{1}$ Zerong Zheng $^{1}$ Han Huang $^{2}$ Yandong Guo $^{2}$ Yebin Liu $^{1}$ Tsinghua University ${}^{2}$ OPPO Research Institute", + "bbox": [ + 184, + 202, + 784, + 257 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 291, + 313, + 306 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Creating animatable avatars from static scans requires the modeling of clothing deformations in different poses. Existing learning-based methods typically add pose-dependent deformations upon a minimally-clothed mesh template or a learned implicit template, which have limitations in capturing details or hinder end-to-end learning. In this paper, we revisit point-based solutions and propose to decompose explicit garment-related templates and then add pose-dependent wrinkles to them. In this way, the clothing deformations are disentangled such that the pose-dependent wrinkles can be better learned and applied to unseen poses. Additionally, to tackle the seam artifact issues in recent state-of-the-art point-based methods, we propose to learn point features on a body surface, which establishes a continuous and compact feature space to capture the fine-grained and pose-dependent clothing geometry. To facilitate the research in this field, we also introduce a high-quality scan dataset of humans in real-world clothing. Our approach is validated on two existing datasets and our newly introduced dataset, showing better clothing deformation results in unseen poses. The project page with code and dataset can be found at https://www.liuyebin.com/closet.", + "bbox": [ + 75, + 325, + 473, + 657 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 691, + 209, + 708 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Animating 3D clothed humans requires the modeling of pose-dependent deformations in various poses. The diversity of clothing styles and body poses makes this task extremely challenging. Traditional methods are based on either simple rigging and skinning [4,20,33] or physics-based simulation [14, 23, 24, 49], which heavily rely on artist efforts or computational resources. Recent learning-based methods [10, 36, 39, 59] resort to modeling the clothing deformation directly from raw scans of clothed humans. Despite the promising progress, this task is still far from being solved due to the challenges in clothing representations, generalization to unseen poses, and data acquisition, etc.", + "bbox": [ + 75, + 719, + 470, + 902 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/9120cdfd2b4793367ccd1702c81f777d90bb4f9e6c091172060a3d4e1a02a13b.jpg", + "image_caption": [ + "Figure 1. Our method learns to decompose garment templates (top row) and add pose-dependent wrinkles upon them (bottom row)." + ], + "image_footnote": [], + "bbox": [ + 506, + 290, + 893, + 513 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "For the modeling of pose-dependent garment geometry, the representation of clothing plays a vital role in a learning-based scheme. As the relationship between body poses and clothing deformations is complex, an effective representation is desirable for neural networks to capture pose-dependent deformations. In the research of this line, meshes [2, 12, 38], implicit fields [10, 59], and point clouds [36, 39] have been adopted to represent clothing. In accordance with the chosen representation, the clothing deformation and geometry features are learned on top of a fixed-resolution template mesh [8, 38], a 3D implicit sampling space [10, 59], or an unfolded UV plane [2, 12, 36, 39]. Among these representations, the mesh is the most efficient one but is limited to a fixed topology due to its discretization scheme. The implicit fields naturally enable continuous feature learning in a resolution-free manner but are too flexible to satisfy the body structure prior, leading to geometry artifacts in unseen poses. The point clouds enjoy the compact nature and topology flexibility and have shown promising results in the recent state-of-the-art solutions [36, 39] to represent clothing, but the feature learning on UV planes still leads to discontinuity artifacts between body parts.", + "bbox": [ + 496, + 568, + 893, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 44 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "501", + "bbox": [ + 485, + 945, + 509, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To model the pose-dependent deformation of clothing, body templates such as SMPL [34] are typically leveraged to account for articulated motions. However, a body template alone is not ideal, since the body template only models the minimally-clothed humans and may hinder the learning of actual pose-dependent deformations, especially in cases of loose clothing. To overcome this issue, recent implicit approaches [59] make attempts to learn skinning weights in the 3D space to complement the imperfect body templates. However, their pose-dependent deformations are typically coarse due to the difficulty in learning implicit fields. For explicit solutions, the recent approach [32] suggests learning coarse templates implicitly at first and then the pose-dependent deformations explicitly. Despite its effectiveness, such a workaround requires a two-step modeling procedure and hinders end-to-end learning.", + "bbox": [ + 75, + 90, + 472, + 332 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this work, we propose CloSET, an end-to-end method to tackle the above issues by modeling Clothed humans on a continuous Surface with Explicit Template decomposition. We follow the spirit of recent state-of-the-art point-based approaches [32, 36, 39] as they show the efficiency and potential in modeling real-world garments. We take steps forward in the following aspects for better point-based modeling of clothed humans. First, we propose to decompose the clothing deformations into explicit garment templates and pose-dependent wrinkles. Specifically, our method learns a garment-related template and adds the pose-dependent displacement upon them, as shown in Fig. 1. Such a garment-related template preserves a shared topology for various poses and enables better learning of pose-dependent wrinkles. Different from the recent solution [32] that needs two-step procedures, our method can decompose the explicit templates in an end-to-end manner with more garment details. Second, we tackle the seam artifact issues that occurred in recent point-based methods [36, 39]. Instead of using unfolded UV planes, we propose to learn point features on a body surface, which supports a continuous and compact feature space. We achieve this by learning hierarchical point-based features on top of the body surface and then using barycentric interpolation to sample features continuously. Compared to feature learning in the UV space [39], on template meshes [8, 38], or in the 3D implicit space [57, 59], our body surface enables the network to capture not only fine-grained details but also long-range part correlations for pose-dependent geometry modeling. Third, we introduce a new scan dataset of humans in real-world clothing, which contains more than 2,000 high-quality scans of humans in diverse outfits, hoping to facilitate the research in this field. The main contributions of this work are summarized below:", + "bbox": [ + 75, + 330, + 472, + 845 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "- We propose a point-based clothed human modeling method by decomposing clothing deformations into explicit garment templates and pose-dependent wrin", + "bbox": [ + 94, + 854, + 470, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "kles in an end-to-end manner. These learnable templates provide a garment-aware canonical space so that pose-dependent deformations can be better learned and applied to unseen poses.", + "bbox": [ + 529, + 90, + 890, + 151 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose to learn point-based clothing features on a continuous body surface, which allows a continuous feature space for fine-grained detail modeling and helps to capture long-range part correlations for pose-dependent geometry modeling.", + "- We introduce a new high-quality scan dataset of clothed humans in real-world clothing to facilitate the research of clothed human modeling and animation from real-world scans." + ], + "bbox": [ + 517, + 160, + 893, + 304 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 316, + 640, + 333 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Representations for Modeling Clothed Humans. A key component in modeling clothed humans is the choice of representation, which mainly falls into two categories: implicit and explicit representations.", + "bbox": [ + 498, + 340, + 890, + 402 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Implicit Modeling. Implicit methods [3, 10-12, 15, 21, 29, 41, 42, 48, 59, 67, 70, 77] represent surfaces as the level set of an implicit neural scalar field. Recent state-of-the-art methods typically learn the clothing deformation field with a canonical space decomposition [9, 30, 45, 59, 65] or part-based modeling strategies [16, 25, 46, 55, 75]. Compared to mesh templates, implicit surfaces are not topologically constrained to specific templates [57, 58], and can model various clothes with complex topology. However, the learning space of an implicit surface is the whole 3D volume, which makes training and interpolation difficult, especially when the numbers of scan data are limited.", + "bbox": [ + 498, + 402, + 892, + 582 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Explicit Modeling. Mesh surfaces, the classic explicit representation, currently dominate the field of 3D modeling [6-8, 13, 23, 24, 26-28, 38, 43, 49, 60, 62, 64, 66, 69, 72] with their compactness and high efficiency in downstream tasks such as rendering, but they are mostly limited to a fixed topology and/or require scan data registered to a template. Thus, mesh-based representations forbid the learning of a universal model for topologically varying clothing types. Though some approaches have been proposed to allow varying mesh topology [44, 47, 63, 68, 79], they are still limited in their expressiveness. Point clouds enjoy both compactness and topological flexibility. Previous work generates sparse point clouds for 3D representation [1, 19, 31, 74]. However, the points need to be densely sampled over the surface to model surface geometry accurately. Due to the difficulty of generating a large point set, recent methods group points into patches [5, 17, 18, 22]. Each patch maps the 2D UV space to the 3D space, allowing arbitrarily dense sampling within this patch. SCALE [36] successfully applies this idea to modeling clothed humans, but produces notable discontinuity artifacts near patch boundaries.", + "bbox": [ + 498, + 583, + 893, + 901 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "502", + "bbox": [ + 485, + 945, + 509, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "POP [39] further utilizes a single fine-grained UV map for the whole body surface, leading to a more topologically flexible representation. However, the discontinuity of the UV map may lead to seam artifacts in POP. Very recently, FITE [32] suggests learning implicit coarse templates [76] at first and then explicit fine details. Despite the efficacy, it requires a two-step modeling procedure. Concurrently, SkiRT [37] proposes to improve the body template by learning the blend skinning weights with several data terms. In contrast, our method applies regularization to achieve template decomposition in an end-to-end manner and learns the point-based pose-dependent displacement more effectively.", + "bbox": [ + 75, + 90, + 472, + 273 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Pose-dependent Deformations for Animation. In the field of character animation, traditional methods utilize rigging and skinning techniques to repose characters [4, 20, 33, 34, 50]. but they fail to model realistic pose-dependent clothing deformations such as wrinkles and sliding motions between clothes and body. We conclude two key ingredients in modeling pose-dependent clothing: (i) pose-dependent feature learning; (ii) datasets with realistic clothing.", + "bbox": [ + 75, + 306, + 468, + 426 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Pose-dependent Feature Learning. Some traditional methods directly incorporate the entire pose parameters into the model [28, 38, 49, 71]. Such methods easily overfit on pose parameters and introduce spurious correlations, causing bad generalization to unseen poses. Recent work explores poses conditioning with local features, either with point clouds [36, 39] or implicit surfaces [59, 67], and shows superiority in improving geometry quality and in eliminating spurious correlations. Among them, the most relevant to ours is POP [39], which extracts local pose features by utilizing convolution on a UV position map. Despite its compelling performance, POP suffers from artifacts inherent to its UV-based representation, hence the convolution on the UV map produces discontinuity near UV islands' boundaries [39]. We address this issue by discarding the UV-based scheme and returning to the actual 3D body surface. We attach the features to a uniform set of points on a T-posed body template, and process them via a PointNet++ [53] structure for pose-dependent modeling. As validated by our experiments, our pose embedding method leads to both qualitative and quantitative improvements.", + "bbox": [ + 75, + 429, + 472, + 747 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Clothed Human Datasets. Another challenging issue for training an animatable avatar is the need for datasets of clothed humans in diverse poses. There are considerable efforts seeking to synthesize clothed datasets with physics-based simulation [14,23,24,39,49]. Although they are diverse in poses, there remains an observable domain gap between synthetic clothes and real data. Acquiring clothed scans with realistic details of clothing deformations [38, 51, 56, 73, 78] is crucial for the development of learning-based methods in this field.", + "bbox": [ + 75, + 750, + 472, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 498, + 89, + 591, + 106 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "As illustrated in Fig. 2, the proposed method CloSET learns garment-related and pose-dependent features on body surfaces (see Sec. 3.1), which can be sampled in a continuous manner and fed into two decoders for the generation of explicit garment templates and pose-dependent wrinkles (see Sec. 3.2).", + "bbox": [ + 496, + 114, + 893, + 205 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Continuous Surface Features", + "text_level": 1, + "bbox": [ + 498, + 212, + 761, + 228 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "As most parts of the clothing are deformed smoothly in different poses, a continuous feature space is desirable to model the garment details and pose-dependent garment geometry. To this end, our approach first learns features on top of a body template surface, i.e., a SMPL [34] or SMPL-X [50] model in a T-pose. Note that these features are not limited to those on template vertices as they can be continuously sampled from the body surface via barycentric interpolation. Hence, our feature space is more continuous than UV-based [2, 39] spaces, while being more compact than 3D implicit feature fields [59, 67].", + "bbox": [ + 496, + 236, + 893, + 402 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To model pose-dependent clothing deformations, the underlying unclothed body model is taken as input to the geometry feature encoder. For each scan, let $\\mathbf{V}^u = \\{\\pmb{v}_n^u\\}_{n=1}^N$ denote the posed vertex positions of the fitted unclothed body model, where $N = 6890$ for SMPL [34] and $N = 10475$ for SMPL-X [50]. These posed vertices act as the pose code and will be paired with the template vertices $\\mathbf{V}^t = \\{\\pmb{v}_n^t\\}_{n=1}^N$ of the body model in a T-pose, which shares the same mesh topology with $\\mathbf{V}^u$ . These point pairs are processed by the pose encoder $\\mathcal{F}_p$ to generate the pose-dependent geometry features $\\{\\phi_p(\\pmb{v}_n^t) \\in \\mathbb{R}^{C_p}\\}_{n=1}^N$ at vertices $\\mathbf{V}^t$ , i.e.,", + "bbox": [ + 496, + 402, + 893, + 584 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\left\\{\\phi_ {p} \\left(\\boldsymbol {v} _ {n} ^ {t}\\right) \\in \\mathbb {R} ^ {C _ {p}} \\right\\} _ {n = 1} ^ {N} = \\mathcal {F} _ {p} \\left(\\mathbf {V} ^ {t}, \\mathbf {V} ^ {u}\\right). \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 568, + 588, + 890, + 608 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To learn hierarchical features with different levels of receptive fields, we adopt PointNet++ [53] as the architecture of the pose encoder $\\mathcal{F}_p$ , where vertices $\\mathbf{V}^t$ are treated as the input point cloud in the PointNet++ network, while vertices $\\mathbf{V}^u$ act as the feature of $\\mathbf{V}^t$ . As the template vertices $\\mathbf{V}^t$ are constant, the encoder $\\mathcal{F}_p$ can focus on the feature learned from the posed vertices $\\mathbf{V}^u$ . Moreover, the PointNet++ based $\\mathcal{F}_p$ first abstracts features from the template vertices $\\mathbf{V}^t$ to sparser points $\\{\\mathbf{V}_l^t\\}_{l=1}^L$ at $L$ levels, where the number of $\\mathbf{V}_l^t$ decreases with $l$ increasing. Then, the features at $\\{\\mathbf{V}_l^t\\}_{l=1}^L$ are further propagated back to $\\mathbf{V}^t$ successively. In this way, the encoder can capture the long-range part correlations of the pose-dependent deformations.", + "bbox": [ + 496, + 613, + 893, + 810 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Similar to POP [39], our method can be trained under multi-outfit or outfit-specific settings. When trained with multiple outfits, the pose-dependent deformation should be aware of the outfit type, and hence requires the input of the garment features. Specifically, the garment-related features $\\{\\phi_g(\\pmb{v}_n^t) \\in \\mathbb{R}^{C_g}\\}_{n=1}^N$ are also defined on template", + "bbox": [ + 496, + 810, + 893, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "503", + "bbox": [ + 485, + 945, + 509, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/986709cebc27f14be506c0ee7a592d4621578e2cdc4a6679da0a796dbf422979.jpg", + "image_caption": [ + "Figure 2. Overview of the proposed method CloSET. Given an input body model, its pose code and garment code are processed hierarchically by point-based pose and garment encoders $\\mathcal{F}_p$ and $\\mathcal{F}_g$ for the learning of surface features $\\phi_p$ and $\\phi_g$ . For any point $\\pmb{p}_i^t$ lying on the template surface, its features $\\phi (\\pmb{p}_i^t)$ are sampled from surface features accordingly and fed into two decoders for the prediction of the explicit garment template and pose-dependent wrinkle displacements, which will be combined and transformed to the clothing point cloud." + ], + "image_footnote": [], + "bbox": [ + 140, + 88, + 823, + 287 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/3a4fb4fca7e8a7f309d99a898b4780b117f625d9bb392b329bd1dfe4fd193c44.jpg", + "image_caption": [ + "Figure 3. Comparison of the bilinear interpolation on the UV plane and the barycentric interpolation on the surface." + ], + "image_footnote": [], + "bbox": [ + 78, + 368, + 468, + 482 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "vertices $\\mathbf{V}^t$ , which are learned by feeding the garment code $\\{\\phi_{gc}(\\pmb{v}_n^t)\\}_{n=1}^N$ to a smaller PointNet++ encoder $\\mathcal{F}_g$ . Note that the garment-related features $\\phi_g$ are shared for each outfit across all poses and optimized during the training. Since both the pose-dependent and garment-related geometry features are aligned with each other, we denote them as the surface features $\\{\\phi(\\pmb{v}_n^t)\\}_{n=1}^N$ for simplicity. Note that the input of $\\phi_g(\\pmb{p}_i^t)$ has no side effect on the results when trained with only one outfit, as the garment features are invariant to the input poses.", + "bbox": [ + 75, + 532, + 468, + 683 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Continuous Feature Interpolation. In the implicit modeling solutions [42, 57, 59], features are learned in a spatially continuous manner, which contributes to the fine-grained modeling of clothing details. To sample continuous features in our scheme, we adopt barycentric interpolation on the surface features $\\phi$ . As illustrated in Fig. 3, for any point $p_i^t = \\mathbf{V}^t (\\pmb {b}_i)$ lying on the template surface, where $\\pmb {b}_i = [n_{i1},n_{i2},n_{i3},b_{i1},b_{i2},b_{i3}]$ denotes the corresponding vertex indices and barycentric coordinates in $\\mathbf{V}^t$ . Then, the corresponding surface features can be retrieved via barycentric interpolation, i.e.,", + "bbox": [ + 75, + 684, + 470, + 851 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\phi \\left(\\boldsymbol {p} _ {i} ^ {t}\\right) = \\sum_ {j = 1} ^ {3} \\left(b _ {i j} * \\phi \\left(\\boldsymbol {v} _ {n _ {i j}} ^ {t}\\right)\\right). \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 174, + 859, + 468, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this way, the point features are not limited to those learned on the template vertices and are continuously defined over the whole body surface without the seaming discontinuity issue in the UV plane.", + "bbox": [ + 498, + 369, + 890, + 431 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Point-based Clothing Deformation", + "text_level": 1, + "bbox": [ + 500, + 440, + 799, + 455 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Following previous work [36, 39], our approach represents the clothed body as a point cloud. For any point on the surface of the unclothed body model, the corresponding features are extracted from surface features to predict its displacement and normal vector.", + "bbox": [ + 498, + 463, + 890, + 537 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Explicit Template Decomposition. Instead of predicting the clothing deformation directly, our method decomposes the deformations into two components: garment-related template displacements and pose-dependent wrinkle displacements. To achieve this, the garment-related template is learned from the garment-related features and shared across all poses. Meanwhile, the learning of pose-dependent wrinkles are conditioned on both garment-related and pose-dependent features. Specifically, for the point $\\pmb{p}_i^u = \\mathbf{V}^u (\\pmb {b}_i)$ at the unclothed body mesh, it has the same vertex indices and barycentric coordinates $\\pmb{b}_{i}$ as the point $\\pmb{p}_i^t$ on the template surface. The pose-dependent and garment-related features of the point $\\pmb{p}_i^u$ are first sampled according to Eq. (2) based on $\\pmb{p}_i^t = \\mathbf{V}^t (\\pmb {b}_i)$ and then further fed into the garment decoder $\\mathcal{D}_g$ and the pose decoder $\\mathcal{D}_p$ for displacement predictions, i.e.,", + "bbox": [ + 496, + 539, + 892, + 781 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\boldsymbol {r} _ {i} ^ {g} = \\mathcal {D} _ {g} \\left(\\phi_ {g} \\left(\\boldsymbol {p} _ {i} ^ {t}\\right), \\boldsymbol {p} _ {i} ^ {t}\\right), \\tag {3} \\\\ \\boldsymbol {r} _ {i} ^ {p} = \\mathcal {D} _ {p} \\left(\\oplus \\left(\\phi_ {g} \\left(\\boldsymbol {p} _ {i} ^ {t}\\right), \\phi_ {p} \\left(\\boldsymbol {p} _ {i} ^ {t}\\right)\\right), \\boldsymbol {p} _ {i} ^ {t}\\right), \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 576, + 790, + 890, + 829 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\oplus$ denotes the concatenation operation, $r_i^g$ and $r_i^p$ are the displacements for garment templates and pose-dependent wrinkles, respectively. Finally, $r_i^g$ and $r_i^p$ will be added together as the clothing deformation $r_i = r_i^g + r_i^p$ .", + "bbox": [ + 498, + 839, + 890, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "504", + "bbox": [ + 485, + 945, + 509, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Local Transformation. Similar to [36, 39], the displacement $\\boldsymbol{r}_i$ is learned in a local coordinate system. It is further transformed to the world coordinate system by applying the following transformation, i.e., $\\boldsymbol{x}_i = \\mathcal{T}_i \\boldsymbol{r}_i + \\boldsymbol{p}_i^u$ , where $\\mathcal{T}_i$ denotes the local transformation calculated based on the unclothed body model. Following [36, 39], the transformation matrix $\\mathcal{T}_i$ is defined at the point $\\boldsymbol{p}_i^u$ on the unclothed body model, which naturally supports the barycentric interpolation. Similarly, the normal $\\boldsymbol{n}_i$ of each point is predicted together with $\\boldsymbol{r}_i$ from $\\mathcal{D}$ and transformed by $\\mathcal{T}_i$ .", + "bbox": [ + 76, + 90, + 472, + 243 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Loss Functions", + "text_level": 1, + "bbox": [ + 76, + 252, + 230, + 266 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Following previous work [36, 39], the point-based clothing deformation is learned with the summation of loss functions: $\\mathcal{L} = \\mathcal{L}_{data} + \\lambda_{rgl}\\mathcal{L}_{rgl}$ . where $\\mathcal{L}_{data}$ and $\\mathcal{L}_{rgl}$ denote the data and regularization terms respectively, and the weight $\\lambda_{rgl}$ balances the loss terms.", + "bbox": [ + 76, + 276, + 468, + 351 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Data Term. The data term $\\mathcal{L}_{data}$ is calculated on the final predicted points and normals, i.e., $\\mathcal{L}_{data} = \\lambda_p\\mathcal{L}_p + \\lambda_n\\mathcal{L}_n$ . Specifically, $\\mathcal{L}_p$ is the normalized Chamfer distance to minimize the bi-directional distances between the point sets of the prediction and the ground-truth scan:", + "bbox": [ + 76, + 352, + 468, + 426 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {p} = C h a m f e r \\left(\\{\\boldsymbol {x} _ {i} \\} _ {i = 1} ^ {M}, \\{\\hat {\\boldsymbol {x}} _ {j} \\} _ {j = 1} ^ {N _ {s}}\\right) =\n$$\n", + "text_format": "latex", + "bbox": [ + 76, + 428, + 352, + 450 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\frac {1}{M} \\sum_ {i = 1} ^ {M} \\min _ {j} \\| \\boldsymbol {x} _ {i} - \\hat {\\boldsymbol {x}} _ {j} \\| _ {2} ^ {2} + \\frac {1}{N _ {s}} \\sum_ {j = 1} ^ {N _ {s}} \\min _ {i} \\| \\boldsymbol {x} _ {i} - \\hat {\\boldsymbol {x}} _ {j} \\| _ {2} ^ {2}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 89, + 463, + 468, + 506 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\hat{\\pmb{x}}_j$ is the point sampled from the ground-truth surface, $M$ and $N_{s}$ denote the number of the predicted and ground-truth points, respectively.", + "bbox": [ + 76, + 517, + 468, + 563 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The normal loss $\\mathcal{L}_n$ is an averaged $L1$ distance between the normal of each predicted point and its nearest ground-truth counterpart:", + "bbox": [ + 76, + 564, + 468, + 609 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {n} = L 1 \\left(\\left\\{\\boldsymbol {n} _ {i} \\right\\} _ {i = 1} ^ {M}, \\left\\{\\hat {\\boldsymbol {n}} _ {i} \\right\\} _ {i = 1} ^ {M}\\right) = \\frac {1}{M} \\sum_ {i = 1} ^ {M} \\| \\boldsymbol {n} _ {i} - \\hat {\\boldsymbol {n}} _ {i} \\|, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 89, + 619, + 468, + 661 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\hat{\\pmb{n}}_i$ is the normal of its nearest point in the groundtruth point set.", + "bbox": [ + 76, + 672, + 468, + 702 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Note that we do not apply data terms on the garment templates, as we found such a strategy leads to noisy template learning in our experiments.", + "bbox": [ + 76, + 704, + 468, + 748 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Regularization Term. The regularization terms are added to prevent the predicted deformations from being extremely large and regularize the garment code. Moreover, following the previous implicit template learning solution [30, 76], we also add regularization on the pose-dependent displacement $\\boldsymbol{r}_i^p$ to encourage it to be as small as possible. As the pose-dependent displacement represents the clothing deformation in various poses, such a regularization implies that the pose-invariant deformation should be retained in the template displacement $\\boldsymbol{r}_i^g$ , which forms", + "bbox": [ + 75, + 750, + 468, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "the garment-related template shared by all poses. Overall, the regularization term can be written as follows:", + "bbox": [ + 498, + 90, + 890, + 121 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {r g l} = \\frac {1}{M} \\sum_ {i = 1} ^ {M} \\| \\mathbf {r} _ {i} \\| _ {2} ^ {2} + \\frac {\\lambda_ {p d}}{M} \\sum_ {i = 1} ^ {M} \\| \\mathbf {r} _ {i} ^ {p} \\| _ {2} ^ {2} + \\frac {\\lambda_ {g c}}{N} \\sum_ {n = 1} ^ {N} \\| \\phi_ {g c} (\\boldsymbol {v} _ {n} ^ {t}) \\| _ {2} ^ {2}. \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 500, + 127, + 903, + 181 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 500, + 193, + 632, + 210 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Network Architecture. For a fair comparison with POP [39], we modify the official PointNet++ [53] (PN++) architecture so that our encoders have comparable network parameters as POP [39]. The modified PointNet++ architecture has 6 layers for feature abstraction and 6 layers for feature propagation (i.e., $L = 6$ ). Since the input point cloud $\\mathbf{V}^t$ has constant coordinates, the farthest point sampling in PointNet++ is only performed at the first forward process, and the sampling indices are saved for the next run. In this way, the runtime is significantly reduced for both training and inference so that our pose and garment encoders can have similar network parameters and runtime speeds to POP. Note that the pose and garment encoders in our method can also be replaced with recent state-of-the-art point-based encoders such as PointMLP [40] and PointNeXt [54]. More details about the network architecture and implementation can be found in the Supp.Mat.", + "bbox": [ + 496, + 218, + 893, + 476 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Datasets. We use CAPE [38], ReSynth [39], and our newly introduced dataset THuman-CloSET for training and evaluation.", + "bbox": [ + 496, + 492, + 890, + 536 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "CAPE [38] is a captured human dataset consisting of multiple humans in various motions. The outfits in this dataset mainly include common clothing such as T-shirts. We follow SCALE [36] to choose blazerlong (with outfits of blazer jacket and long trousers) and shortlong (with outfits of short T-shirt and long trousers) from subject 03375 to validate the efficacy of our method.", + "bbox": [ + 496, + 537, + 890, + 643 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "ReSynth [39] is a synthetic dataset introduced in POP [39]. It is created by using physics simulation, and contains challenging outfits such as skirts and jackets. We use the official training and test split as [39].", + "bbox": [ + 496, + 643, + 890, + 703 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "THuman-CloSET is our newly introduced dataset, containing high-quality clothed human scans captured by a dense camera rig. We introduce THuman-CloSET for the reason that existing pose-dependent clothing datasets [38, 39] are with either relatively tight clothing or synthetic clothing via physics simulation. In THuman-CloSET, there are more than 2,000 scans of 15 outfitsits with a large variation in clothing style, including T-shirts, pants, skirts, dresses, jackets, and coats, to name a few. For each outfit, the subject is guided to perform different poses by imitating the poses in CAPE. Moreover, each subject has a scan with minimal clothing in A-pose. THuman-CloSET contains well-fitted body models in the form of SMPL-X [50]. Note that", + "bbox": [ + 496, + 704, + 893, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "505", + "bbox": [ + 485, + 945, + 509, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/20a840774c6c54741557ef6b34db82457544b7b5e90d5e298d8413945862b2f4.jpg", + "table_caption": [ + "Table 1. Quantitative comparison with previous point-based methods on ReSynth. † denotes the methods using 1/8 training data." + ], + "table_footnote": [], + "table_body": "
MethodChamfer-L2↓Normal diff. ↓
MeanMaxMeanMax
outfit-specific
SCALE [36]1.4918.4511.0411.321
POP [39]1.3567.3391.0131.289
multi-outfit
Baseline (POP [39]) †1.4907.8591.0501.326
Baseline w. PN++ †1.2905.9401.0281.330
CloSET (Ours) †1.2405.5431.0191.315
", + "bbox": [ + 89, + 121, + 455, + 276 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/a8dd29c01a9960cbfd6eb7c1360e50292b94d164a12f6902f7d72fa8f3afda22.jpg", + "table_caption": [ + "Table 2. Quantitative comparison of different methods on the proposed THuman-CloSET dataset in the outfit-specific setting." + ], + "table_footnote": [], + "table_body": "
Subject IDSCANImateSNARFPOPCloSET
CDNMLCDNMLCDNMLCDNML
sweater-0001.061.647.112.090.761.550.681.48
longshirt-0011.421.856.662.211.541.831.391.71
skirt-0051.931.749.392.311.661.431.491.36
", + "bbox": [ + 84, + 316, + 460, + 409 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "the loose clothing makes the fitting of the underlying body models quite challenging. For more accurate fitting of the body models, we first fit a SMPL-X model on the scan of the subject in minimal clothing and then adopt its shape parameters for fitting the outfit scans in different poses. More details can be found in the Supp.Mat. In our experiments, we use the outfit scans in 100 different poses for training and use the remaining poses for evaluation. We hope our new dataset can open a promising direction for clothed human modeling and animation from real-world scans.", + "bbox": [ + 75, + 434, + 467, + 585 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Metrics. Following previous work [39], we generate 50K points from our method and point-based baselines and adopt the Chamfer Distance (see Eq. (4)) and the $L1$ normal discrepancy (see Eq. (5)) for quantitative evaluation. By default, the Chamfer distance (CD) and normal discrepancy (NML) are reported in the unit of $\\times 10^{-4}m^2$ and $\\times 10^{-1}$ , respectively. To evaluate the implicit modeling methods, the points are sampled from the surface extracted using Marching Cubes [35].", + "bbox": [ + 75, + 604, + 467, + 741 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Comparison with the State-of-the-art Methods", + "text_level": 1, + "bbox": [ + 76, + 751, + 467, + 767 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We compare results with recent state-of-the-art methods, including point-based approaches SCALE [36], POP [39], and SkiRT [37], and implicit approaches SCANimate [59] and SNARF [10].", + "bbox": [ + 76, + 775, + 467, + 835 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "ReSynth. Tab. 1 reports the results of the pose-dependent clothing predictions on unseen motion sequences from the ReSynth [39] dataset, where all 12 outfits fits are used for eval", + "bbox": [ + 76, + 854, + 467, + 900 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/162f1615d267928e094ee28f0ff99127ee6b796a9c1c91830b2b52edb2f5b597.jpg", + "table_caption": [ + "Table 3. Quantitative comparison of different methods on the ReSynth dataset in the outfit-specific setting. The garment styles are non-skirt, skirt, and dress for carla-004, Christine-027, and felice-004, respectively." + ], + "table_footnote": [], + "table_body": "
Subject IDSCANImatePOPSkiRTCloSET
CDNMLCDNMLCDNMLCDNML
carla-0040.901.520.511.020.481.060.491.04
christine-0273.211.661.720.971.540.991.490.97
felice-00420.792.947.341.246.451.256.011.16
", + "bbox": [ + 504, + 148, + 887, + 241 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "uation. As can be seen, the proposed approach has the lowest mean and max errors, which outperforms all other approaches including POP [39]. Note that our approach needs fewer data for the pose-dependent deformation modeling. By using only 1/8 data, our approach achieves a performance comparable to or even better than other models trained with full data. Tab. 3 also reports outfit-specific performances on 3 selected subject-outfit types, including jackets, skirts, and dresses. In comparison with the recent state-of-the-art method SkiRT [37], our method achieves better results on challenging skirt/dress outfits and comparable results on non-skirt clothing.", + "bbox": [ + 496, + 258, + 890, + 441 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "THuman-CloSET. The effectiveness of our method is also validated on our real-world THuman-CloSET dataset. The sparse training poses and loose clothing make this dataset very challenging for clothed human modeling. Tab. 2 reports the quantitative comparisons of different methods on three representative outfits. Fig. 4 also shows example results of different methods, where we follow previous work [39] to obtain meshed results via Poisson surface reconstruction. We can see that our method generalizes better to unseen poses and produces more natural pose-dependent wrinkles than other methods. In our experiments, we found that SNARF [10] fails to learn correct skinning weights due to loose clothing and limited training poses. As discussed in FITE [32], there is an ill-posed issue of jointly optimizing the canonical shape and the skinning fields, which becomes more severe in our dataset.", + "bbox": [ + 496, + 459, + 890, + 700 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Ablation Study", + "text_level": 1, + "bbox": [ + 500, + 710, + 651, + 726 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation of Continuous Surface Features. The point features in our method are learned on the body surface, which provides a continuous and compact feature learning space. To validate this, Tab. 4 summarizes the feature learning space of different approaches and their performances on two representative outfits from CAPE [38]. Here, we only include the proposed Continuous Surface Features (CSF) in Tab. 4 by applying the continuous features on POP [39] for fair comparisons with SCALE [36] and POP [39]. As discussed previously, existing solutions learn features either in a discontinuous space (e.g., CAPE [38] on the fixed resolu", + "bbox": [ + 496, + 734, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "506", + "bbox": [ + 485, + 945, + 509, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/ef3d7d5c5c228210c963ddd9ab79b78c1e8b679a7e644349e3d7b564cbb524b6.jpg", + "image_caption": [ + "Figure 4. Comparison of different clothed human modeling methods on the proposed real-world scan dataset." + ], + "image_footnote": [], + "bbox": [ + 86, + 87, + 880, + 441 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b67fe7458d39ddb96fdec5ae53b785611ac7b2f4d219d300ec39e7cd889e0bff.jpg", + "image_caption": [ + "Figure 5. Comparison of the approach learned on UV planes (POP [39]) and the approach learned on continuous surfaces (Ours). Our solution alleviates the seam artifacts of POP." + ], + "image_footnote": [], + "bbox": [ + 78, + 470, + 486, + 694 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "tion mesh, POP [39] on the 2D UV plane) or in a space that is too flexible (e.g., NASA [16] in the implicit 3D space), while our approach learns features on continuous and compact surface space. Though SCALE [36] has also investigated using the point-based encoder (PointNet [52]) for pose-dependent feature extraction, it only uses the global features which lack fine-grained information. In contrast, we adopt PointNet++ [53] (PN++) to learn hierarchical surface features, so that the pose-dependent features can be learned more effectively. Fig. 5 shows the qualitative re", + "bbox": [ + 75, + 750, + 472, + 902 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/aeb9246b5368d70939e6a78f56ddd6d5a5f06bba89bb507dbf44f54273f728aa.jpg", + "table_caption": [ + "Table 4. Comparison of the modeling ability of different approaches and their feature learning space on the CAPE dataset." + ], + "table_footnote": [], + "table_body": "
MethodsFeaturesChamfer-L2↓Normal diff. ↓
blazerlongshortlongblazerlongshortlong
CAPE [38]Mesh1.961.371.281.15
NASA [16]3D space1.370.951.291.17
SCALE [36]Global1.461.031.341.16
SCALE [36]UV plane1.070.891.221.12
POP [39]UV plane0.780.571.291.24
CSFSurface0.710.541.151.09
", + "bbox": [ + 501, + 503, + 890, + 612 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/c8c8de97171999ad25851958bad5294fde4bb79c4f425bca5d0cbc5a88e9f864.jpg", + "table_caption": [ + "Table 5. Ablation study on the effectiveness of continuous surface features (CSF) and Explicit Template Decomposition (ETD) on a dress outfit (felice-004 from ReSynth [39])." + ], + "table_footnote": [], + "table_body": "
MethodPOPPOP + ETDCSFCSF + ETD
CD7.347.056.536.01
NML1.241.171.161.16
", + "bbox": [ + 508, + 662, + 879, + 724 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "sults of the ablation approaches learned on UV planes and continuous surfaces. We can see that our solution clearly alleviates the seam artifacts of POP.", + "bbox": [ + 498, + 731, + 890, + 776 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Evaluation of Explicit Template Decomposition. The decomposed templates help to capture more accurate pose-dependent deformations and produce more natural wrinkles in unseen poses, especially for outfits that differ largely from the body template. To validate the effectiveness of our decomposition strategy, Tab. 5 reports the ablation experiments on a dress outfit (felice-004) of the ReSynth [39]", + "bbox": [ + 496, + 794, + 893, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "507", + "bbox": [ + 485, + 944, + 511, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/982d99d42e975eb97fe7747d0b50555014dc1693aef34c30fab0bb6819d7a963.jpg", + "image_caption": [ + "Figure 6. Comparison of the clothing deformation in unseen poses. Explicit template decomposition (ETD) helps to capture more natural pose-dependent wrinkle details than POP [39]." + ], + "image_footnote": [], + "bbox": [ + 86, + 87, + 883, + 458 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/b27cfd4580ea5deef04c17561fb87501427671c1e9a0f35c176e6b4d01073783.jpg", + "image_caption": [ + "Figure 7. Comparison of the learned templates with FITE [32]." + ], + "image_footnote": [], + "bbox": [ + 81, + 496, + 486, + 638 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "dataset. We can see that the proposed Explicit Template Decomposition (ETD) brings clear performance gains over baseline methods. Fig. 6 shows the visual improvements of pose-dependent wrinkles when applying the explicit template decomposition. Note that the templates are decomposed in an end-to-end manner in our method. Compared with the implicit template learned in the recent approach FITE [32], the explicit templates in our method contain more details, as shown in Fig. 7.", + "bbox": [ + 75, + 665, + 472, + 801 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusions and Future Work", + "text_level": 1, + "bbox": [ + 76, + 814, + 354, + 830 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this work, we present CloSET, a point-based clothed human modeling method that is built upon a continuous surface and learns to decompose explicit garment templates for better learning of pose-dependent deformations. By learn-", + "bbox": [ + 75, + 839, + 470, + 902 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "ing features on a continuous surface, our solution gets rid of the seam artifacts in previous state-of-the-art point-based methods [36, 39]. Moreover, the explicit template decomposition helps to capture more accurate and natural pose-dependent wrinkles. To facilitate the research in this direction, we also introduce a high-quality real-world scan dataset with diverse outfit styles and accurate body model fitting.", + "bbox": [ + 496, + 500, + 890, + 619 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Limitations and Future Work. Due to the incorrect skinning weight used in our template, the issue of the nonuniform point distribution remains for the skirt and dress outfits. Combining our method with recent learnable skinning solutions [37, 59] could alleviate this issue and further improve the results. Currently, our method does not leverage information from adjacent poses. Enforcing temporal consistency and correspondences between adjacent frames would be interesting for future work. Moreover, incorporating physics-based losses into the learning process like SNUG [61] would also be a promising solution to address the artifacts like self-intersections.", + "bbox": [ + 496, + 621, + 892, + 801 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements. This work was supported by the National Key R&D Program of China (2022YFF0902200), the National Natural Science Foundation of China (No.62125107 and No.61827805), and the China Postdoctoral Science Foundation (No.2022M721844).", + "bbox": [ + 496, + 824, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "508", + "bbox": [ + 486, + 945, + 509, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Panos Achlioptas, Olga Diamanti, Ioannis Mitliagkas, and Leonidas Guibas. Learning representations and generative models for 3D point clouds. In ICML, pages 40-49, 2018. 2", + "[2] Thiemo Alldieck, Gerard Pons-Moll, Christian Theobalt, and Marcus Magnor. Tex2shape: Detailed full human body geometry from a single image. In ICCV, pages 2293-2303, 2019. 1, 3", + "[3] Ziqian Bai, Timur Bagautdinov, Javier Romero, Michael Zollhöfer, Ping Tan, and Shunsuke Saito. AutoAvatar: Autoregressive neural fields for dynamic avatar modeling. ECCV, 2022. 2", + "[4] Ilya Baran and Jovan Popovic. Automatic rigging and animation of 3D characters. ACM TOG, 26(3):72-es, 2007. 1, 3", + "[5] Jan Bednářík, Shaifali Parashar, Erhan Gundogdu, Mathieu Salzmann, and Pascal Fua. Shape reconstruction by learning differentiable surface representations. In CVPR, pages 4715-4724, 2020. 2", + "[6] Hugo Bertiche, Meysam Madadi, Emilio Tylson, and Sergio Escalera. DeePSD: Automatic deep skinning and pose space deformation for 3D garment animation. In ICCV, pages 5471-5480, 2021. 2", + "[7] Bharat Lal Bhatnagar, Garvita Tiwari, Christian Theobalt, and Gerard Pons-Moll. Multi-Garment Net: Learning to dress 3D people from images. In ICCV, pages 5420-5430, 2019. 2", + "[8] Andrei Burov, Matthias Nießner, and Justus Thies. Dynamic surface function networks for clothed human bodies. In ICCV, pages 10754-10764, October 2021. 1, 2", + "[9] Xu Chen, Tianjian Jiang, Jie Song, Jinlong Yang, Michael J Black, Andreas Geiger, and Otmar Hilliges. gDNA: Towards generative detailed neural avatars. In CVPR, pages 20427-20437, 2022. 2", + "[10] Xu Chen, Yufeng Zheng, Michael J. Black, Otmar Hilliges, and Andreas Geiger. SNARF: Differentiable forward skinning for animating non-rigid neural implicit shapes. In ICCV, pages 11594–11604, October 2021. 1, 2, 6, 7", + "[11] Julian Chibane, Aymen Mir, and Gerard Pons-Moll. Neural unsigned distance fields for implicit function learning. In NeurIPS, pages 21638-21652, 2020. 2", + "[12] Enric Corona, Albert Pumarola, Guillem Alenya, Gerard Pons-Moll, and Francesc Moreno-Noguer. SMPLicit: Topology-aware generative model for clothed people. In CVPR, pages 11875-11885, 2021. 1, 2", + "[13] Edilson De Aguiar, Leonid Sigal, Adrien Treuille, and Jessica K Hodgins. Stable spaces for real-time clothing. In ACM TOG, volume 29, page 106. ACM, 2010. 2", + "[14] Deform Dynamics. https://deformdynamics.com/.1,3", + "[15] Boyang Deng, JP Lewis, Timothy Jeruzalski, Gerard Pons-Moll, Geoffrey Hinton, Mohammad Norouzi, and Andrea Tagliasacchi. Neural articulated shape approximation. In ECCV, pages 612-628, 2020. 2", + "[16] Boyang Deng, John P Lewis, Timothy Jeruzalski, Gerard Pons-Moll, Geoffrey Hinton, Mohammad Norouzi, and An-" + ], + "bbox": [ + 78, + 114, + 468, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "drea Tagliasacchi. Nasa neural articulated shape approximation. In ECCV, pages 612-628. Springer, 2020. 2, 7", + "[17] Zhantao Deng, Jan Bednárík, Mathieu Salzmann, and Pascal Fua. Better patch stitching for parametric surface reconstruction. In 3DV, pages 593-602, 2020. 2", + "[18] Theo Deprelle, Thibault Groueix, Matthew Fisher, Vladimir Kim, Bryan Russell, and Mathieu Aubry. Learning elementary structures for 3D shape generation and matching. In NeurIPS, pages 7433-7443, 2019. 2", + "[19] Haoqiang Fan, Hao Su, and Leonidas J Guibas. A point set generation network for 3D object reconstruction from a single image. In CVPR, pages 2463-2471, 2017. 2", + "[20] Andrew Feng, Dan Casas, and Ari Shapiro. Avatar reshaping and automatic rigging using a deformable model. In Proceedings of the ACM SIGGRAPH Conference on Motion in Games, pages 57-64, 2015. 1, 3", + "[21] Amos Gropp, Lior Yariv, Niv Haim, Matan Atzmon, and Yaron Lipman. Implicit geometric regularization for learning shapes. In ICML, pages 3569-3579, 2020. 2", + "[22] Thibault Groueix, Matthew Fisher, Vladimir G Kim, Bryan C Russell, and Mathieu Aubry. 3D-CODED: 3D correspondences by deep deformation. In ECCV, pages 230-246, 2018. 2", + "[23] Peng Guan, Loretta Reiss, David A Hirshberg, Alexander Weiss, and Michael J Black. DRAPE: DRessing Any PErson. ACM TOG, 31(4):35-1, 2012. 1, 2, 3", + "[24] Erhan Gundogdu, Victor Constantin, Amrollah Seifoddini, Minh Dang, Mathieu Salzmann, and Pascal Fua. GarNet: A two-stream network for fast and accurate 3D cloth draping. In CVPR, pages 8739-8748, 2019. 1, 2, 3", + "[25] Boyan Jiang, Xinlin Ren, Mingsong Dou, Xiangyang Xue, Yanwei Fu, and Yinda Zhang. LoRD: Local 4d implicit representation for high-fidelity dynamic human modeling. In ECCV, pages 307-326. Springer, 2022. 2", + "[26] Boyi Jiang, Juyong Zhang, Yang Hong, Jinhao Luo, Ligang Liu, and Hujun Bao. BCNet: Learning body and cloth shape from a single image. In ECCV, pages 18-35. Springer, 2020. 2", + "[27] Hyomin Kim, Hyeonseo Nam, Junkon Kim, Jaesik Park, and Seungyong Lee. LaplacianFusion: Detailed 3D clothed-human body reconstruction. ACM TOG, 41(6):1-14, 2022. 2", + "[28] Zorah Lahner, Daniel Cremers, and Tony Tung. Deepwrinkles: Accurate and realistic clothing modeling. In ECCV, pages 667-684, 2018. 2, 3", + "[29] Ruilong Li, Julian Tanke, Minh Vo, Michael Zollhöfer, Jürgen Gall, Angjoo Kanazawa, and Christoph Lassner. TAVA: Template-free animatable volumetric actors. In ECCV, pages 419-436. Springer, 2022. 2", + "[30] Zhe Li, Zerong Zheng, Hongwen Zhang, Chaonan Ji, and Yebin Liu. Avatarcap: Animatable avatar conditioned monocular human volumetric capture. In ECCV, pages 322-341. Springer, 2022. 2, 5", + "[31] Chen-Hsuan Lin, Chen Kong, and Simon Lucey. Learning efficient point cloud generation for dense 3D object reconstruction. In AAAI, pages 7114-7121, 2018. 2" + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "509", + "bbox": [ + 486, + 945, + 509, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[32] Siyou Lin, Hongwen Zhang, Zerong Zheng, Ruizhi Shao, and Yebin Liu. Learning implicit templates for point-based clothed human modeling. ECCV, 2022. 2, 3, 6, 8", + "[33] Lijuan Liu, Youyi Zheng, Di Tang, Yi Yuan, Changjie Fan, and Kun Zhou. NeuroSkinning: Automatic skin binding for production characters with deep graph networks. ACM TOG, 38(4):1-12, 2019. 1, 3", + "[34] Matthew Loper, Naureen Mahmood, Javier Romero, Gerard Pons-Moll, and Michael J Black. SMPL: A skinned multiperson linear model. ACM TOG, 34(6):248, 2015. 2, 3", + "[35] William E Lorensen and Harvey E Cline. Marching cubes: A high resolution 3D surface construction algorithm. In SIGGRAPH, volume 21, pages 163-169, 1987. 6", + "[36] Qianli Ma, Shunsuke Saito, Jinlong Yang, Siyu Tang, and Michael J Black. SCALE: Modeling clothed humans with a surface codec of articulated local elements. In CVPR, pages 16082-16093, 2021. 1, 2, 3, 4, 5, 6, 7, 8", + "[37] Qianli Ma, Jinlong Yang, Michael J Black, and Siyu Tang. Neural point-based shape modeling of humans in challenging clothing. 3DV, 2022. 3, 6, 8", + "[38] Qianli Ma, Jinlong Yang, Anurag Ranjan, Sergi Pujades, Gerard Pons-Moll, Siyu Tang, and Michael J Black. Learning to dress 3D people in generative clothing. In CVPR, pages 6469-6478, 2020. 1, 2, 3, 5, 6, 7", + "[39] Qianli Ma, Jinlong Yang, Siyu Tang, and Michael J Black. The power of points for modeling humans in clothing. In ICCV, pages 10974-10984, 2021. 1, 2, 3, 4, 5, 6, 7, 8", + "[40] Xu Ma, Can Qin, Haoxuan You, Haoxi Ran, and Yun Fu. Rethinking network design and local geometry in point cloud: A simple residual mlp framework. In ICLR, 2022. 5", + "[41] Lars Mescheder, Michael Oechsle, Michael Niemeyer, Sebastian Nowozin, and Andreas Geiger. Occupancy networks: Learning 3D reconstruction in function space. In CVPR, pages 4460-4470, 2019. 2", + "[42] Marko Mihajlovic, Yan Zhang, Michael J Black, and Siyu Tang. LEAP: Learning articulated occupancy of people. In CVPR, pages 10461-10471, 2021. 2, 4", + "[43] Alexandros Neophytou and Adrian Hilton. A layered model of human body and garment deformation. In 3DV, pages 171-178, 2014. 2", + "[44] Hayato Onizuka, Zehra Hayirci, Diego Thomas, Akihiro Sugimoto, Hideaki Uchiyama, and Rin-ichiro Taniguchi. TetraTSDF: 3D human reconstruction from a single image with a tetrahedral outer shell. In CVPR, pages 6011–6020, 2020. 2", + "[45] Pablo Palafox, Aljaž Božić, Justus Thies, Matthias Nießner, and Angela Dai. NPMs: Neural parametric models for 3D deformable shapes. In ICCV, pages 12695-12705, October 2021. 2", + "[46] Pablo Palafox, Nikolaos Sarafianos, Tony Tung, and Angela Dai. SPAMs: Structured implicit parametric models. In CVPR, pages 12851-12860, 2022. 2", + "[47] Junyi Pan, Xiaoguang Han, Weikai Chen, Jiapeng Tang, and Kui Jia. Deep mesh reconstruction from single RGB images via topology modification networks. In ICCV, pages 9963-9972, 2019. 2" + ], + "bbox": [ + 78, + 90, + 468, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[48] Jeong Joon Park, Peter Florence, Julian Straub, Richard Newcombe, and Steven Lovegrove. DeepSDF: Learning continuous signed distance functions for shape representation. In CVPR, pages 165-174, 2019. 2", + "[49] Chaitanya Patel, Zhouyingcheng Liao, and Gerard Pons-Moll. TailorNet: Predicting clothing in 3D as a function of human pose, shape and garment style. In CVPR, pages 7363-7373, 2020. 1, 2, 3", + "[50] Georgios Pavlakos, Vasileios Choutas, Nima Ghorbani, Timo Bolkart, Ahmed AA Osman, Dimitrios Tzionas, and Michael J Black. Expressive body capture: 3D hands, face, and body from a single image. In CVPR, pages 10975-10985, 2019. 3, 5", + "[51] Gerard Pons-Moll, Sergi Pujades, Sonny Hu, and Michael Black. ClothCap: Seamless 4d clothing capture and retargeting. ACM TOG, 36(4), 2017. Two first authors contributed equally. 3", + "[52] Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. PointNet: Deep learning on point sets for 3D classification and segmentation. In CVPR, pages 652-660, 2017. 7", + "[53] Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J Guibas. Pointnet++: Deep hierarchical feature learning on point sets in a metric space. NeurIPS, 30, 2017. 3, 5, 7", + "[54] Guocheng Qian, Yuchen Li, Houwen Peng, Jinjie Mai, Hasan Hammoud, Mohamed Elhoseiny, and Bernard Ghanem. Pointnext: Revisiting pointnet++ with improved training and scaling strategies. NeurIPS, 35:23192-23204, 2022. 5", + "[55] Shenhan Qian, Jiale Xu, Ziwei Liu, Liqian Ma, and Shenghua Gao. UNIF: United neural implicit functions for clothed human reconstruction and animation. In ECCV, 2022. 2", + "[56] Renderpeople, 2020. https://renderpeople.com.3", + "[57] Shunsuke Saito, Zeng Huang, Ryota Natsume, Shigeo Morishima, Angjoo Kanazawa, and Hao Li. PIFu: Pixel-aligned implicit function for high-resolution clothed human digitization. In ICCV, pages 2304-2314, 2019. 2, 4", + "[58] Shunsuke Saito, Tomas Simon, Jason Saragih, and Hanbyul Joo. PIFuHD: Multi-level pixel-aligned implicit function for high-resolution 3D human digitization. In CVPR, pages 84–93, 2020. 2", + "[59] Shunsuke Saito, Jinlong Yang, Qianli Ma, and Michael J Black. SCANimate: Weakly supervised learning of skinned clothed avatar networks. In CVPR, pages 2886-2897, 2021. 1, 2, 3, 4, 6, 7, 8", + "[60] Igor Santesteban, Miguel A. Otaduy, and Dan Casas. Learning-Based Animation of Clothing for Virtual Try-On. Comput. Graph. Forum, 38(2):355-366, 2019. 2", + "[61] Igor Santesteban, Miguel A Otaduy, and Dan Casas. SNUG: Self-supervised neural dynamic garments. In CVPR, pages 8140-8150, 2022. 8", + "[62] Igor Santesteban, Nils Thuerey, Miguel A Otaduy, and Dan Casas. Self-supervised collision handling via generative 3D garment models for virtual try-on. In CVPR, pages 11763-11773, 2021. 2", + "[63] Yu Shen, Junbang Liang, and Ming C Lin. Gan-based garment generation using sewing pattern images. In ECCV, volume 1, page 3, 2020. 2" + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "510", + "bbox": [ + 486, + 945, + 511, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[64] Garvita Tiwari, Bharat Lal Bhatnagar, Tony Tung, and Gerard Pons-Moll. SIZER: A dataset and model for parsing 3D clothing and learning size sensitive 3D clothing. In ECCV, volume 12348, pages 1-18, 2020. 2", + "[65] Garvita Tiwari, Nikolaos Sarafianos, Tony Tung, and Gerard Pons-Moll. Neural-GIF: Neural generalized implicit functions for animating people in clothing. In ICCV, pages 11708-11718, 2021. 2", + "[66] Raquel Vidaurre, Igor Santesteban, Elena Garces, and Dan Casas. Fully convolutional graph neural networks for parametric virtual try-on. In Comput. Graph. Forum, volume 39, pages 145-156. Wiley Online Library, 2020. 2", + "[67] Shaofei Wang, Marko Mihajlovic, Qianli Ma, Andreas Geiger, and Siyu Tang. Meta­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­.\nclothed human models from few depth images. In NeurIPS, 2021. 2, 3", + "[68] Jane Wu, Zhenglin Geng, Hui Zhou, and Ronald Fedkiw. Skinning a parameterization of three-dimensional space for neural network cloth. arXiv preprint arXiv:2006.04874, 2020. 2", + "[69] Donglai Xiang, Fabian Prada, Timur Bagautdinov, Weipeng Xu, Yuan Dong, He Wen, Jessica Hodgins, and Chenglei Wu. Modeling clothing as a separate layer for an animatable human avatar. ACM TOG, dec 2021. 2", + "[70] Yuliang Xiu, Jinlong Yang, Dimitrios Tzionas, and Michael J Black. ICON: implicit clothed humans obtained from normals. In CVPR, pages 13286-13296. IEEE, 2022. 2", + "[71] Jinlong Yang, Jean-Sebastien Franco, Franck Hetroy-Wheeler, and Stefanie Wuhrer. Analyzing clothing layer deformation statistics of 3D human motions. In ECCV, September 2018. 3", + "[72] Shan Yang, Zherong Pan, Tanya Amert, Ke Wang, Licheng Yu, Tamara Berg, and Ming C Lin. Physics-inspired garment recovery from a single-view image. ACM TOG, 37(5):1-14, 2018. 2", + "[73] Tao Yu, Zerong Zheng, Kaiwen Guo, Pengpeng Liu, Qionghai Dai, and Yebin Liu. Function4D: Real-time human volumetric capture from very sparse consumer rgbd sensors. In CVPR, June 2021. 3", + "[74] Ilya Zakharkin, Kirill Mazur, Artur Grigorev, and Victor Lempitsky. Point-based modeling of human clothing. In ICCV, pages 14718-14727, October 2021. 2", + "[75] Zerong Zheng, Han Huang, Tao Yu, Hongwen Zhang, Yandong Guo, and Yebin Liu. Structured local radiance fields for human avatar modeling. In CVPR, pages 15893-15903, 2022. 2", + "[76] Zerong Zheng, Tao Yu, Qionghai Dai, and Yebin Liu. Deep implicit templates for 3D shape representation. In CVPR, pages 1429-1439, 2021. 3, 5", + "[77] Zerong Zheng, Tao Yu, Yebin Liu, and Qionghai Dai. PaMIR: Parametric model-conditioned implicit representation for image-based human reconstruction. IEEE TPAMI, 2021. 2", + "[78] Zerong Zheng, Tao Yu, Yixuan Wei, Qionghai Dai, and Yebin Liu. DeepHuman: 3D human reconstruction from a single image. In ICCV, pages 7739-7749, 2019. 3" + ], + "bbox": [ + 78, + 90, + 468, + 901 + ], + "page_idx": 10 + }, + { + "type": "ref_text", + "text": "[79] Heming Zhu, Yu Cao, Hang Jin, Weikai Chen, Dong Du, Zhangye Wang, Shuguang Cui, and Xiaoguang Han. Deep Fashion3D: A dataset and benchmark for 3D garment reconstruction from single images. In ECCV, volume 12346, pages 512-530, 2020. 2", + "bbox": [ + 501, + 90, + 892, + 161 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "511", + "bbox": [ + 486, + 945, + 509, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2023/CloSET_ Modeling Clothed Humans on Continuous Surface With Explicit Template Decomposition/4725afdd-f941-4b99-a289-05a751ed5c4e_model.json b/2023/CloSET_ Modeling Clothed Humans on Continuous Surface With Explicit Template Decomposition/4725afdd-f941-4b99-a289-05a751ed5c4e_model.json new file mode 100644 index 0000000000000000000000000000000000000000..0b0e3aca5d4c3c85f85e85aef732805672ae63a5 --- /dev/null +++ b/2023/CloSET_ Modeling Clothed Humans on Continuous Surface With Explicit Template Decomposition/4725afdd-f941-4b99-a289-05a751ed5c4e_model.json @@ -0,0 +1,2224 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.0, + 0.812, + 0.045 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.16, + 0.13, + 0.817, + 0.178 + ], + "angle": 0, + "content": "CloSET: Modeling Clothed Humans on Continuous Surface with Explicit Template Decomposition" + }, + { + "type": "text", + "bbox": [ + 0.185, + 0.203, + 0.785, + 0.258 + ], + "angle": 0, + "content": "Hongwen Zhang\\(^{1}\\) Siyou Lin\\(^{1}\\) Ruizhi Shao\\(^{1}\\) Yuxiang Zhang\\(^{1}\\) Zerong Zheng\\(^{1}\\) Han Huang\\(^{2}\\) Yandong Guo\\(^{2}\\) Yebin Liu\\(^{1}\\) Tsinghua University \\({}^{2}\\)OPPO Research Institute" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.292, + 0.314, + 0.308 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.326, + 0.474, + 0.659 + ], + "angle": 0, + "content": "Creating animatable avatars from static scans requires the modeling of clothing deformations in different poses. Existing learning-based methods typically add pose-dependent deformations upon a minimally-clothed mesh template or a learned implicit template, which have limitations in capturing details or hinder end-to-end learning. In this paper, we revisit point-based solutions and propose to decompose explicit garment-related templates and then add pose-dependent wrinkles to them. In this way, the clothing deformations are disentangled such that the pose-dependent wrinkles can be better learned and applied to unseen poses. Additionally, to tackle the seam artifact issues in recent state-of-the-art point-based methods, we propose to learn point features on a body surface, which establishes a continuous and compact feature space to capture the fine-grained and pose-dependent clothing geometry. To facilitate the research in this field, we also introduce a high-quality scan dataset of humans in real-world clothing. Our approach is validated on two existing datasets and our newly introduced dataset, showing better clothing deformation results in unseen poses. The project page with code and dataset can be found at https://www.liuyebin.com/closet." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.692, + 0.21, + 0.709 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.72, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Animating 3D clothed humans requires the modeling of pose-dependent deformations in various poses. The diversity of clothing styles and body poses makes this task extremely challenging. Traditional methods are based on either simple rigging and skinning [4,20,33] or physics-based simulation [14, 23, 24, 49], which heavily rely on artist efforts or computational resources. Recent learning-based methods [10, 36, 39, 59] resort to modeling the clothing deformation directly from raw scans of clothed humans. Despite the promising progress, this task is still far from being solved due to the challenges in clothing representations, generalization to unseen poses, and data acquisition, etc." + }, + { + "type": "image", + "bbox": [ + 0.508, + 0.291, + 0.895, + 0.515 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.525, + 0.895, + 0.554 + ], + "angle": 0, + "content": "Figure 1. Our method learns to decompose garment templates (top row) and add pose-dependent wrinkles upon them (bottom row)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.569, + 0.895, + 0.903 + ], + "angle": 0, + "content": "For the modeling of pose-dependent garment geometry, the representation of clothing plays a vital role in a learning-based scheme. As the relationship between body poses and clothing deformations is complex, an effective representation is desirable for neural networks to capture pose-dependent deformations. In the research of this line, meshes [2, 12, 38], implicit fields [10, 59], and point clouds [36, 39] have been adopted to represent clothing. In accordance with the chosen representation, the clothing deformation and geometry features are learned on top of a fixed-resolution template mesh [8, 38], a 3D implicit sampling space [10, 59], or an unfolded UV plane [2, 12, 36, 39]. Among these representations, the mesh is the most efficient one but is limited to a fixed topology due to its discretization scheme. The implicit fields naturally enable continuous feature learning in a resolution-free manner but are too flexible to satisfy the body structure prior, leading to geometry artifacts in unseen poses. The point clouds enjoy the compact nature and topology flexibility and have shown promising results in the recent state-of-the-art solutions [36, 39] to represent clothing, but the feature learning on UV planes still leads to discontinuity artifacts between body parts." + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.946, + 0.511, + 0.957 + ], + "angle": 0, + "content": "501" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.333 + ], + "angle": 0, + "content": "To model the pose-dependent deformation of clothing, body templates such as SMPL [34] are typically leveraged to account for articulated motions. However, a body template alone is not ideal, since the body template only models the minimally-clothed humans and may hinder the learning of actual pose-dependent deformations, especially in cases of loose clothing. To overcome this issue, recent implicit approaches [59] make attempts to learn skinning weights in the 3D space to complement the imperfect body templates. However, their pose-dependent deformations are typically coarse due to the difficulty in learning implicit fields. For explicit solutions, the recent approach [32] suggests learning coarse templates implicitly at first and then the pose-dependent deformations explicitly. Despite its effectiveness, such a workaround requires a two-step modeling procedure and hinders end-to-end learning." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.332, + 0.473, + 0.847 + ], + "angle": 0, + "content": "In this work, we propose CloSET, an end-to-end method to tackle the above issues by modeling Clothed humans on a continuous Surface with Explicit Template decomposition. We follow the spirit of recent state-of-the-art point-based approaches [32, 36, 39] as they show the efficiency and potential in modeling real-world garments. We take steps forward in the following aspects for better point-based modeling of clothed humans. First, we propose to decompose the clothing deformations into explicit garment templates and pose-dependent wrinkles. Specifically, our method learns a garment-related template and adds the pose-dependent displacement upon them, as shown in Fig. 1. Such a garment-related template preserves a shared topology for various poses and enables better learning of pose-dependent wrinkles. Different from the recent solution [32] that needs two-step procedures, our method can decompose the explicit templates in an end-to-end manner with more garment details. Second, we tackle the seam artifact issues that occurred in recent point-based methods [36, 39]. Instead of using unfolded UV planes, we propose to learn point features on a body surface, which supports a continuous and compact feature space. We achieve this by learning hierarchical point-based features on top of the body surface and then using barycentric interpolation to sample features continuously. Compared to feature learning in the UV space [39], on template meshes [8, 38], or in the 3D implicit space [57, 59], our body surface enables the network to capture not only fine-grained details but also long-range part correlations for pose-dependent geometry modeling. Third, we introduce a new scan dataset of humans in real-world clothing, which contains more than 2,000 high-quality scans of humans in diverse outfits, hoping to facilitate the research in this field. The main contributions of this work are summarized below:" + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.856, + 0.472, + 0.903 + ], + "angle": 0, + "content": "- We propose a point-based clothed human modeling method by decomposing clothing deformations into explicit garment templates and pose-dependent wrin" + }, + { + "type": "text", + "bbox": [ + 0.531, + 0.092, + 0.892, + 0.152 + ], + "angle": 0, + "content": "kles in an end-to-end manner. These learnable templates provide a garment-aware canonical space so that pose-dependent deformations can be better learned and applied to unseen poses." + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.161, + 0.892, + 0.237 + ], + "angle": 0, + "content": "- We propose to learn point-based clothing features on a continuous body surface, which allows a continuous feature space for fine-grained detail modeling and helps to capture long-range part correlations for pose-dependent geometry modeling." + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.245, + 0.895, + 0.305 + ], + "angle": 0, + "content": "- We introduce a new high-quality scan dataset of clothed humans in real-world clothing to facilitate the research of clothed human modeling and animation from real-world scans." + }, + { + "type": "list", + "bbox": [ + 0.519, + 0.161, + 0.895, + 0.305 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.317, + 0.642, + 0.334 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.342, + 0.892, + 0.403 + ], + "angle": 0, + "content": "Representations for Modeling Clothed Humans. A key component in modeling clothed humans is the choice of representation, which mainly falls into two categories: implicit and explicit representations." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.403, + 0.893, + 0.583 + ], + "angle": 0, + "content": "Implicit Modeling. Implicit methods [3, 10-12, 15, 21, 29, 41, 42, 48, 59, 67, 70, 77] represent surfaces as the level set of an implicit neural scalar field. Recent state-of-the-art methods typically learn the clothing deformation field with a canonical space decomposition [9, 30, 45, 59, 65] or part-based modeling strategies [16, 25, 46, 55, 75]. Compared to mesh templates, implicit surfaces are not topologically constrained to specific templates [57, 58], and can model various clothes with complex topology. However, the learning space of an implicit surface is the whole 3D volume, which makes training and interpolation difficult, especially when the numbers of scan data are limited." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.584, + 0.895, + 0.902 + ], + "angle": 0, + "content": "Explicit Modeling. Mesh surfaces, the classic explicit representation, currently dominate the field of 3D modeling [6-8, 13, 23, 24, 26-28, 38, 43, 49, 60, 62, 64, 66, 69, 72] with their compactness and high efficiency in downstream tasks such as rendering, but they are mostly limited to a fixed topology and/or require scan data registered to a template. Thus, mesh-based representations forbid the learning of a universal model for topologically varying clothing types. Though some approaches have been proposed to allow varying mesh topology [44, 47, 63, 68, 79], they are still limited in their expressiveness. Point clouds enjoy both compactness and topological flexibility. Previous work generates sparse point clouds for 3D representation [1, 19, 31, 74]. However, the points need to be densely sampled over the surface to model surface geometry accurately. Due to the difficulty of generating a large point set, recent methods group points into patches [5, 17, 18, 22]. Each patch maps the 2D UV space to the 3D space, allowing arbitrarily dense sampling within this patch. SCALE [36] successfully applies this idea to modeling clothed humans, but produces notable discontinuity artifacts near patch boundaries." + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.946, + 0.511, + 0.957 + ], + "angle": 0, + "content": "502" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.091, + 0.473, + 0.274 + ], + "angle": 0, + "content": "POP [39] further utilizes a single fine-grained UV map for the whole body surface, leading to a more topologically flexible representation. However, the discontinuity of the UV map may lead to seam artifacts in POP. Very recently, FITE [32] suggests learning implicit coarse templates [76] at first and then explicit fine details. Despite the efficacy, it requires a two-step modeling procedure. Concurrently, SkiRT [37] proposes to improve the body template by learning the blend skinning weights with several data terms. In contrast, our method applies regularization to achieve template decomposition in an end-to-end manner and learns the point-based pose-dependent displacement more effectively." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.307, + 0.47, + 0.428 + ], + "angle": 0, + "content": "Pose-dependent Deformations for Animation. In the field of character animation, traditional methods utilize rigging and skinning techniques to repose characters [4, 20, 33, 34, 50]. but they fail to model realistic pose-dependent clothing deformations such as wrinkles and sliding motions between clothes and body. We conclude two key ingredients in modeling pose-dependent clothing: (i) pose-dependent feature learning; (ii) datasets with realistic clothing." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.43, + 0.473, + 0.748 + ], + "angle": 0, + "content": "Pose-dependent Feature Learning. Some traditional methods directly incorporate the entire pose parameters into the model [28, 38, 49, 71]. Such methods easily overfit on pose parameters and introduce spurious correlations, causing bad generalization to unseen poses. Recent work explores poses conditioning with local features, either with point clouds [36, 39] or implicit surfaces [59, 67], and shows superiority in improving geometry quality and in eliminating spurious correlations. Among them, the most relevant to ours is POP [39], which extracts local pose features by utilizing convolution on a UV position map. Despite its compelling performance, POP suffers from artifacts inherent to its UV-based representation, hence the convolution on the UV map produces discontinuity near UV islands' boundaries [39]. We address this issue by discarding the UV-based scheme and returning to the actual 3D body surface. We attach the features to a uniform set of points on a T-posed body template, and process them via a PointNet++ [53] structure for pose-dependent modeling. As validated by our experiments, our pose embedding method leads to both qualitative and quantitative improvements." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.473, + 0.903 + ], + "angle": 0, + "content": "Clothed Human Datasets. Another challenging issue for training an animatable avatar is the need for datasets of clothed humans in diverse poses. There are considerable efforts seeking to synthesize clothed datasets with physics-based simulation [14,23,24,39,49]. Although they are diverse in poses, there remains an observable domain gap between synthetic clothes and real data. Acquiring clothed scans with realistic details of clothing deformations [38, 51, 56, 73, 78] is crucial for the development of learning-based methods in this field." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.09, + 0.593, + 0.107 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.115, + 0.895, + 0.207 + ], + "angle": 0, + "content": "As illustrated in Fig. 2, the proposed method CloSET learns garment-related and pose-dependent features on body surfaces (see Sec. 3.1), which can be sampled in a continuous manner and fed into two decoders for the generation of explicit garment templates and pose-dependent wrinkles (see Sec. 3.2)." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.213, + 0.762, + 0.229 + ], + "angle": 0, + "content": "3.1. Continuous Surface Features" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.237, + 0.895, + 0.403 + ], + "angle": 0, + "content": "As most parts of the clothing are deformed smoothly in different poses, a continuous feature space is desirable to model the garment details and pose-dependent garment geometry. To this end, our approach first learns features on top of a body template surface, i.e., a SMPL [34] or SMPL-X [50] model in a T-pose. Note that these features are not limited to those on template vertices as they can be continuously sampled from the body surface via barycentric interpolation. Hence, our feature space is more continuous than UV-based [2, 39] spaces, while being more compact than 3D implicit feature fields [59, 67]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.403, + 0.895, + 0.585 + ], + "angle": 0, + "content": "To model pose-dependent clothing deformations, the underlying unclothed body model is taken as input to the geometry feature encoder. For each scan, let \\(\\mathbf{V}^u = \\{\\pmb{v}_n^u\\}_{n=1}^N\\) denote the posed vertex positions of the fitted unclothed body model, where \\(N = 6890\\) for SMPL [34] and \\(N = 10475\\) for SMPL-X [50]. These posed vertices act as the pose code and will be paired with the template vertices \\(\\mathbf{V}^t = \\{\\pmb{v}_n^t\\}_{n=1}^N\\) of the body model in a T-pose, which shares the same mesh topology with \\(\\mathbf{V}^u\\). These point pairs are processed by the pose encoder \\(\\mathcal{F}_p\\) to generate the pose-dependent geometry features \\(\\{\\phi_p(\\pmb{v}_n^t) \\in \\mathbb{R}^{C_p}\\}_{n=1}^N\\) at vertices \\(\\mathbf{V}^t\\), i.e.," + }, + { + "type": "equation", + "bbox": [ + 0.57, + 0.589, + 0.892, + 0.609 + ], + "angle": 0, + "content": "\\[\n\\left\\{\\phi_ {p} \\left(\\boldsymbol {v} _ {n} ^ {t}\\right) \\in \\mathbb {R} ^ {C _ {p}} \\right\\} _ {n = 1} ^ {N} = \\mathcal {F} _ {p} \\left(\\mathbf {V} ^ {t}, \\mathbf {V} ^ {u}\\right). \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.614, + 0.895, + 0.811 + ], + "angle": 0, + "content": "To learn hierarchical features with different levels of receptive fields, we adopt PointNet++ [53] as the architecture of the pose encoder \\(\\mathcal{F}_p\\), where vertices \\(\\mathbf{V}^t\\) are treated as the input point cloud in the PointNet++ network, while vertices \\(\\mathbf{V}^u\\) act as the feature of \\(\\mathbf{V}^t\\). As the template vertices \\(\\mathbf{V}^t\\) are constant, the encoder \\(\\mathcal{F}_p\\) can focus on the feature learned from the posed vertices \\(\\mathbf{V}^u\\). Moreover, the PointNet++ based \\(\\mathcal{F}_p\\) first abstracts features from the template vertices \\(\\mathbf{V}^t\\) to sparser points \\(\\{\\mathbf{V}_l^t\\}_{l=1}^L\\) at \\(L\\) levels, where the number of \\(\\mathbf{V}_l^t\\) decreases with \\(l\\) increasing. Then, the features at \\(\\{\\mathbf{V}_l^t\\}_{l=1}^L\\) are further propagated back to \\(\\mathbf{V}^t\\) successively. In this way, the encoder can capture the long-range part correlations of the pose-dependent deformations." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.811, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Similar to POP [39], our method can be trained under multi-outfit or outfit-specific settings. When trained with multiple outfits, the pose-dependent deformation should be aware of the outfit type, and hence requires the input of the garment features. Specifically, the garment-related features \\(\\{\\phi_g(\\pmb{v}_n^t) \\in \\mathbb{R}^{C_g}\\}_{n=1}^N\\) are also defined on template" + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.946, + 0.511, + 0.957 + ], + "angle": 0, + "content": "503" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.142, + 0.089, + 0.825, + 0.288 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.293, + 0.895, + 0.352 + ], + "angle": 0, + "content": "Figure 2. Overview of the proposed method CloSET. Given an input body model, its pose code and garment code are processed hierarchically by point-based pose and garment encoders \\(\\mathcal{F}_p\\) and \\(\\mathcal{F}_g\\) for the learning of surface features \\(\\phi_p\\) and \\(\\phi_g\\). For any point \\(\\pmb{p}_i^t\\) lying on the template surface, its features \\(\\phi (\\pmb{p}_i^t)\\) are sampled from surface features accordingly and fed into two decoders for the prediction of the explicit garment template and pose-dependent wrinkle displacements, which will be combined and transformed to the clothing point cloud." + }, + { + "type": "image", + "bbox": [ + 0.08, + 0.369, + 0.47, + 0.483 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.489, + 0.471, + 0.517 + ], + "angle": 0, + "content": "Figure 3. Comparison of the bilinear interpolation on the UV plane and the barycentric interpolation on the surface." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.534, + 0.47, + 0.684 + ], + "angle": 0, + "content": "vertices \\(\\mathbf{V}^t\\), which are learned by feeding the garment code \\(\\{\\phi_{gc}(\\pmb{v}_n^t)\\}_{n=1}^N\\) to a smaller PointNet++ encoder \\(\\mathcal{F}_g\\). Note that the garment-related features \\(\\phi_g\\) are shared for each outfit across all poses and optimized during the training. Since both the pose-dependent and garment-related geometry features are aligned with each other, we denote them as the surface features \\(\\{\\phi(\\pmb{v}_n^t)\\}_{n=1}^N\\) for simplicity. Note that the input of \\(\\phi_g(\\pmb{p}_i^t)\\) has no side effect on the results when trained with only one outfit, as the garment features are invariant to the input poses." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.685, + 0.471, + 0.852 + ], + "angle": 0, + "content": "Continuous Feature Interpolation. In the implicit modeling solutions [42, 57, 59], features are learned in a spatially continuous manner, which contributes to the fine-grained modeling of clothing details. To sample continuous features in our scheme, we adopt barycentric interpolation on the surface features \\(\\phi\\). As illustrated in Fig. 3, for any point \\(p_i^t = \\mathbf{V}^t (\\pmb {b}_i)\\) lying on the template surface, where \\(\\pmb {b}_i = [n_{i1},n_{i2},n_{i3},b_{i1},b_{i2},b_{i3}]\\) denotes the corresponding vertex indices and barycentric coordinates in \\(\\mathbf{V}^t\\). Then, the corresponding surface features can be retrieved via barycentric interpolation, i.e.," + }, + { + "type": "equation", + "bbox": [ + 0.175, + 0.861, + 0.47, + 0.904 + ], + "angle": 0, + "content": "\\[\n\\phi \\left(\\boldsymbol {p} _ {i} ^ {t}\\right) = \\sum_ {j = 1} ^ {3} \\left(b _ {i j} * \\phi \\left(\\boldsymbol {v} _ {n _ {i j}} ^ {t}\\right)\\right). \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.371, + 0.892, + 0.432 + ], + "angle": 0, + "content": "In this way, the point features are not limited to those learned on the template vertices and are continuously defined over the whole body surface without the seaming discontinuity issue in the UV plane." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.441, + 0.8, + 0.457 + ], + "angle": 0, + "content": "3.2. Point-based Clothing Deformation" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.464, + 0.892, + 0.538 + ], + "angle": 0, + "content": "Following previous work [36, 39], our approach represents the clothed body as a point cloud. For any point on the surface of the unclothed body model, the corresponding features are extracted from surface features to predict its displacement and normal vector." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.54, + 0.893, + 0.782 + ], + "angle": 0, + "content": "Explicit Template Decomposition. Instead of predicting the clothing deformation directly, our method decomposes the deformations into two components: garment-related template displacements and pose-dependent wrinkle displacements. To achieve this, the garment-related template is learned from the garment-related features and shared across all poses. Meanwhile, the learning of pose-dependent wrinkles are conditioned on both garment-related and pose-dependent features. Specifically, for the point \\( \\pmb{p}_i^u = \\mathbf{V}^u (\\pmb {b}_i) \\) at the unclothed body mesh, it has the same vertex indices and barycentric coordinates \\( \\pmb{b}_{i} \\) as the point \\( \\pmb{p}_i^t \\) on the template surface. The pose-dependent and garment-related features of the point \\( \\pmb{p}_i^u \\) are first sampled according to Eq. (2) based on \\( \\pmb{p}_i^t = \\mathbf{V}^t (\\pmb {b}_i) \\) and then further fed into the garment decoder \\( \\mathcal{D}_g \\) and the pose decoder \\( \\mathcal{D}_p \\) for displacement predictions, i.e.," + }, + { + "type": "equation", + "bbox": [ + 0.578, + 0.791, + 0.891, + 0.83 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\boldsymbol {r} _ {i} ^ {g} = \\mathcal {D} _ {g} \\left(\\phi_ {g} \\left(\\boldsymbol {p} _ {i} ^ {t}\\right), \\boldsymbol {p} _ {i} ^ {t}\\right), \\tag {3} \\\\ \\boldsymbol {r} _ {i} ^ {p} = \\mathcal {D} _ {p} \\left(\\oplus \\left(\\phi_ {g} \\left(\\boldsymbol {p} _ {i} ^ {t}\\right), \\phi_ {p} \\left(\\boldsymbol {p} _ {i} ^ {t}\\right)\\right), \\boldsymbol {p} _ {i} ^ {t}\\right), \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.84, + 0.892, + 0.903 + ], + "angle": 0, + "content": "where \\(\\oplus\\) denotes the concatenation operation, \\(r_i^g\\) and \\(r_i^p\\) are the displacements for garment templates and pose-dependent wrinkles, respectively. Finally, \\(r_i^g\\) and \\(r_i^p\\) will be added together as the clothing deformation \\(r_i = r_i^g + r_i^p\\)." + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.946, + 0.511, + 0.957 + ], + "angle": 0, + "content": "504" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.473, + 0.244 + ], + "angle": 0, + "content": "Local Transformation. Similar to [36, 39], the displacement \\( \\boldsymbol{r}_i \\) is learned in a local coordinate system. It is further transformed to the world coordinate system by applying the following transformation, i.e., \\( \\boldsymbol{x}_i = \\mathcal{T}_i \\boldsymbol{r}_i + \\boldsymbol{p}_i^u \\), where \\( \\mathcal{T}_i \\) denotes the local transformation calculated based on the unclothed body model. Following [36, 39], the transformation matrix \\( \\mathcal{T}_i \\) is defined at the point \\( \\boldsymbol{p}_i^u \\) on the unclothed body model, which naturally supports the barycentric interpolation. Similarly, the normal \\( \\boldsymbol{n}_i \\) of each point is predicted together with \\( \\boldsymbol{r}_i \\) from \\( \\mathcal{D} \\) and transformed by \\( \\mathcal{T}_i \\)." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.253, + 0.231, + 0.267 + ], + "angle": 0, + "content": "3.3. Loss Functions" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.277, + 0.47, + 0.352 + ], + "angle": 0, + "content": "Following previous work [36, 39], the point-based clothing deformation is learned with the summation of loss functions: \\(\\mathcal{L} = \\mathcal{L}_{data} + \\lambda_{rgl}\\mathcal{L}_{rgl}\\). where \\(\\mathcal{L}_{data}\\) and \\(\\mathcal{L}_{rgl}\\) denote the data and regularization terms respectively, and the weight \\(\\lambda_{rgl}\\) balances the loss terms." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.353, + 0.47, + 0.427 + ], + "angle": 0, + "content": "Data Term. The data term \\(\\mathcal{L}_{data}\\) is calculated on the final predicted points and normals, i.e., \\(\\mathcal{L}_{data} = \\lambda_p\\mathcal{L}_p + \\lambda_n\\mathcal{L}_n\\). Specifically, \\(\\mathcal{L}_p\\) is the normalized Chamfer distance to minimize the bi-directional distances between the point sets of the prediction and the ground-truth scan:" + }, + { + "type": "equation", + "bbox": [ + 0.078, + 0.429, + 0.353, + 0.452 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {p} = C h a m f e r \\left(\\{\\boldsymbol {x} _ {i} \\} _ {i = 1} ^ {M}, \\{\\hat {\\boldsymbol {x}} _ {j} \\} _ {j = 1} ^ {N _ {s}}\\right) =\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.091, + 0.464, + 0.47, + 0.507 + ], + "angle": 0, + "content": "\\[\n\\frac {1}{M} \\sum_ {i = 1} ^ {M} \\min _ {j} \\| \\boldsymbol {x} _ {i} - \\hat {\\boldsymbol {x}} _ {j} \\| _ {2} ^ {2} + \\frac {1}{N _ {s}} \\sum_ {j = 1} ^ {N _ {s}} \\min _ {i} \\| \\boldsymbol {x} _ {i} - \\hat {\\boldsymbol {x}} _ {j} \\| _ {2} ^ {2}, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.518, + 0.47, + 0.564 + ], + "angle": 0, + "content": "where \\(\\hat{\\pmb{x}}_j\\) is the point sampled from the ground-truth surface, \\(M\\) and \\(N_{s}\\) denote the number of the predicted and ground-truth points, respectively." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.565, + 0.469, + 0.61 + ], + "angle": 0, + "content": "The normal loss \\(\\mathcal{L}_n\\) is an averaged \\(L1\\) distance between the normal of each predicted point and its nearest ground-truth counterpart:" + }, + { + "type": "equation", + "bbox": [ + 0.09, + 0.621, + 0.47, + 0.662 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {n} = L 1 \\left(\\left\\{\\boldsymbol {n} _ {i} \\right\\} _ {i = 1} ^ {M}, \\left\\{\\hat {\\boldsymbol {n}} _ {i} \\right\\} _ {i = 1} ^ {M}\\right) = \\frac {1}{M} \\sum_ {i = 1} ^ {M} \\| \\boldsymbol {n} _ {i} - \\hat {\\boldsymbol {n}} _ {i} \\|, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.674, + 0.469, + 0.703 + ], + "angle": 0, + "content": "where \\(\\hat{\\pmb{n}}_i\\) is the normal of its nearest point in the groundtruth point set." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.705, + 0.469, + 0.749 + ], + "angle": 0, + "content": "Note that we do not apply data terms on the garment templates, as we found such a strategy leads to noisy template learning in our experiments." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.47, + 0.902 + ], + "angle": 0, + "content": "Regularization Term. The regularization terms are added to prevent the predicted deformations from being extremely large and regularize the garment code. Moreover, following the previous implicit template learning solution [30, 76], we also add regularization on the pose-dependent displacement \\( \\boldsymbol{r}_i^p \\) to encourage it to be as small as possible. As the pose-dependent displacement represents the clothing deformation in various poses, such a regularization implies that the pose-invariant deformation should be retained in the template displacement \\( \\boldsymbol{r}_i^g \\), which forms" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.892, + 0.122 + ], + "angle": 0, + "content": "the garment-related template shared by all poses. Overall, the regularization term can be written as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.5, + 0.128, + 0.905, + 0.183 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {r g l} = \\frac {1}{M} \\sum_ {i = 1} ^ {M} \\| \\mathbf {r} _ {i} \\| _ {2} ^ {2} + \\frac {\\lambda_ {p d}}{M} \\sum_ {i = 1} ^ {M} \\| \\mathbf {r} _ {i} ^ {p} \\| _ {2} ^ {2} + \\frac {\\lambda_ {g c}}{N} \\sum_ {n = 1} ^ {N} \\| \\phi_ {g c} (\\boldsymbol {v} _ {n} ^ {t}) \\| _ {2} ^ {2}. \\tag {6}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.194, + 0.633, + 0.211 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.219, + 0.895, + 0.477 + ], + "angle": 0, + "content": "Network Architecture. For a fair comparison with POP [39], we modify the official PointNet++ [53] (PN++) architecture so that our encoders have comparable network parameters as POP [39]. The modified PointNet++ architecture has 6 layers for feature abstraction and 6 layers for feature propagation (i.e., \\( L = 6 \\)). Since the input point cloud \\(\\mathbf{V}^t\\) has constant coordinates, the farthest point sampling in PointNet++ is only performed at the first forward process, and the sampling indices are saved for the next run. In this way, the runtime is significantly reduced for both training and inference so that our pose and garment encoders can have similar network parameters and runtime speeds to POP. Note that the pose and garment encoders in our method can also be replaced with recent state-of-the-art point-based encoders such as PointMLP [40] and PointNeXt [54]. More details about the network architecture and implementation can be found in the Supp.Mat." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.493, + 0.892, + 0.537 + ], + "angle": 0, + "content": "Datasets. We use CAPE [38], ReSynth [39], and our newly introduced dataset THuman-CloSET for training and evaluation." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.539, + 0.892, + 0.644 + ], + "angle": 0, + "content": "CAPE [38] is a captured human dataset consisting of multiple humans in various motions. The outfits in this dataset mainly include common clothing such as T-shirts. We follow SCALE [36] to choose blazerlong (with outfits of blazer jacket and long trousers) and shortlong (with outfits of short T-shirt and long trousers) from subject 03375 to validate the efficacy of our method." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.645, + 0.892, + 0.704 + ], + "angle": 0, + "content": "ReSynth [39] is a synthetic dataset introduced in POP [39]. It is created by using physics simulation, and contains challenging outfits such as skirts and jackets. We use the official training and test split as [39]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.705, + 0.895, + 0.901 + ], + "angle": 0, + "content": "THuman-CloSET is our newly introduced dataset, containing high-quality clothed human scans captured by a dense camera rig. We introduce THuman-CloSET for the reason that existing pose-dependent clothing datasets [38, 39] are with either relatively tight clothing or synthetic clothing via physics simulation. In THuman-CloSET, there are more than 2,000 scans of 15 outfitsits with a large variation in clothing style, including T-shirts, pants, skirts, dresses, jackets, and coats, to name a few. For each outfit, the subject is guided to perform different poses by imitating the poses in CAPE. Moreover, each subject has a scan with minimal clothing in A-pose. THuman-CloSET contains well-fitted body models in the form of SMPL-X [50]. Note that" + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.946, + 0.511, + 0.957 + ], + "angle": 0, + "content": "505" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.09, + 0.468, + 0.118 + ], + "angle": 0, + "content": "Table 1. Quantitative comparison with previous point-based methods on ReSynth. † denotes the methods using 1/8 training data." + }, + { + "type": "table", + "bbox": [ + 0.09, + 0.122, + 0.457, + 0.277 + ], + "angle": 0, + "content": "
MethodChamfer-L2↓Normal diff. ↓
MeanMaxMeanMax
outfit-specific
SCALE [36]1.4918.4511.0411.321
POP [39]1.3567.3391.0131.289
multi-outfit
Baseline (POP [39]) †1.4907.8591.0501.326
Baseline w. PN++ †1.2905.9401.0281.330
CloSET (Ours) †1.2405.5431.0191.315
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.285, + 0.468, + 0.314 + ], + "angle": 0, + "content": "Table 2. Quantitative comparison of different methods on the proposed THuman-CloSET dataset in the outfit-specific setting." + }, + { + "type": "table", + "bbox": [ + 0.086, + 0.318, + 0.462, + 0.41 + ], + "angle": 0, + "content": "
Subject IDSCANImateSNARFPOPCloSET
CDNMLCDNMLCDNMLCDNML
sweater-0001.061.647.112.090.761.550.681.48
longshirt-0011.421.856.662.211.541.831.391.71
skirt-0051.931.749.392.311.661.431.491.36
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.435, + 0.468, + 0.587 + ], + "angle": 0, + "content": "the loose clothing makes the fitting of the underlying body models quite challenging. For more accurate fitting of the body models, we first fit a SMPL-X model on the scan of the subject in minimal clothing and then adopt its shape parameters for fitting the outfit scans in different poses. More details can be found in the Supp.Mat. In our experiments, we use the outfit scans in 100 different poses for training and use the remaining poses for evaluation. We hope our new dataset can open a promising direction for clothed human modeling and animation from real-world scans." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.606, + 0.468, + 0.742 + ], + "angle": 0, + "content": "Metrics. Following previous work [39], we generate 50K points from our method and point-based baselines and adopt the Chamfer Distance (see Eq. (4)) and the \\(L1\\) normal discrepancy (see Eq. (5)) for quantitative evaluation. By default, the Chamfer distance (CD) and normal discrepancy (NML) are reported in the unit of \\(\\times 10^{-4}m^2\\) and \\(\\times 10^{-1}\\), respectively. To evaluate the implicit modeling methods, the points are sampled from the surface extracted using Marching Cubes [35]." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.752, + 0.468, + 0.768 + ], + "angle": 0, + "content": "4.1. Comparison with the State-of-the-art Methods" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.776, + 0.468, + 0.837 + ], + "angle": 0, + "content": "We compare results with recent state-of-the-art methods, including point-based approaches SCALE [36], POP [39], and SkiRT [37], and implicit approaches SCANimate [59] and SNARF [10]." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.856, + 0.468, + 0.901 + ], + "angle": 0, + "content": "ReSynth. Tab. 1 reports the results of the pose-dependent clothing predictions on unseen motion sequences from the ReSynth [39] dataset, where all 12 outfits fits are used for eval" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.09, + 0.892, + 0.146 + ], + "angle": 0, + "content": "Table 3. Quantitative comparison of different methods on the ReSynth dataset in the outfit-specific setting. The garment styles are non-skirt, skirt, and dress for carla-004, Christine-027, and felice-004, respectively." + }, + { + "type": "table", + "bbox": [ + 0.505, + 0.15, + 0.888, + 0.242 + ], + "angle": 0, + "content": "
Subject IDSCANImatePOPSkiRTCloSET
CDNMLCDNMLCDNMLCDNML
carla-0040.901.520.511.020.481.060.491.04
christine-0273.211.661.720.971.540.991.490.97
felice-00420.792.947.341.246.451.256.011.16
" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.26, + 0.892, + 0.442 + ], + "angle": 0, + "content": "uation. As can be seen, the proposed approach has the lowest mean and max errors, which outperforms all other approaches including POP [39]. Note that our approach needs fewer data for the pose-dependent deformation modeling. By using only 1/8 data, our approach achieves a performance comparable to or even better than other models trained with full data. Tab. 3 also reports outfit-specific performances on 3 selected subject-outfit types, including jackets, skirts, and dresses. In comparison with the recent state-of-the-art method SkiRT [37], our method achieves better results on challenging skirt/dress outfits and comparable results on non-skirt clothing." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.46, + 0.892, + 0.702 + ], + "angle": 0, + "content": "THuman-CloSET. The effectiveness of our method is also validated on our real-world THuman-CloSET dataset. The sparse training poses and loose clothing make this dataset very challenging for clothed human modeling. Tab. 2 reports the quantitative comparisons of different methods on three representative outfits. Fig. 4 also shows example results of different methods, where we follow previous work [39] to obtain meshed results via Poisson surface reconstruction. We can see that our method generalizes better to unseen poses and produces more natural pose-dependent wrinkles than other methods. In our experiments, we found that SNARF [10] fails to learn correct skinning weights due to loose clothing and limited training poses. As discussed in FITE [32], there is an ill-posed issue of jointly optimizing the canonical shape and the skinning fields, which becomes more severe in our dataset." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.712, + 0.653, + 0.727 + ], + "angle": 0, + "content": "4.2. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.735, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Evaluation of Continuous Surface Features. The point features in our method are learned on the body surface, which provides a continuous and compact feature learning space. To validate this, Tab. 4 summarizes the feature learning space of different approaches and their performances on two representative outfits from CAPE [38]. Here, we only include the proposed Continuous Surface Features (CSF) in Tab. 4 by applying the continuous features on POP [39] for fair comparisons with SCALE [36] and POP [39]. As discussed previously, existing solutions learn features either in a discontinuous space (e.g., CAPE [38] on the fixed resolu" + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.946, + 0.511, + 0.957 + ], + "angle": 0, + "content": "506" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.088, + 0.088, + 0.881, + 0.442 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.16, + 0.444, + 0.807, + 0.459 + ], + "angle": 0, + "content": "Figure 4. Comparison of different clothed human modeling methods on the proposed real-world scan dataset." + }, + { + "type": "image", + "bbox": [ + 0.079, + 0.471, + 0.487, + 0.695 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.698, + 0.47, + 0.74 + ], + "angle": 0, + "content": "Figure 5. Comparison of the approach learned on UV planes (POP [39]) and the approach learned on continuous surfaces (Ours). Our solution alleviates the seam artifacts of POP." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.473, + 0.903 + ], + "angle": 0, + "content": "tion mesh, POP [39] on the 2D UV plane) or in a space that is too flexible (e.g., NASA [16] in the implicit 3D space), while our approach learns features on continuous and compact surface space. Though SCALE [36] has also investigated using the point-based encoder (PointNet [52]) for pose-dependent feature extraction, it only uses the global features which lack fine-grained information. In contrast, we adopt PointNet++ [53] (PN++) to learn hierarchical surface features, so that the pose-dependent features can be learned more effectively. Fig. 5 shows the qualitative re" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.471, + 0.892, + 0.5 + ], + "angle": 0, + "content": "Table 4. Comparison of the modeling ability of different approaches and their feature learning space on the CAPE dataset." + }, + { + "type": "table", + "bbox": [ + 0.502, + 0.504, + 0.892, + 0.613 + ], + "angle": 0, + "content": "
MethodsFeaturesChamfer-L2↓Normal diff. ↓
blazerlongshortlongblazerlongshortlong
CAPE [38]Mesh1.961.371.281.15
NASA [16]3D space1.370.951.291.17
SCALE [36]Global1.461.031.341.16
SCALE [36]UV plane1.070.891.221.12
POP [39]UV plane0.780.571.291.24
CSFSurface0.710.541.151.09
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.617, + 0.892, + 0.658 + ], + "angle": 0, + "content": "Table 5. Ablation study on the effectiveness of continuous surface features (CSF) and Explicit Template Decomposition (ETD) on a dress outfit (felice-004 from ReSynth [39])." + }, + { + "type": "table", + "bbox": [ + 0.509, + 0.663, + 0.88, + 0.726 + ], + "angle": 0, + "content": "
MethodPOPPOP + ETDCSFCSF + ETD
CD7.347.056.536.01
NML1.241.171.161.16
" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.732, + 0.892, + 0.777 + ], + "angle": 0, + "content": "sults of the ablation approaches learned on UV planes and continuous surfaces. We can see that our solution clearly alleviates the seam artifacts of POP." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.795, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Evaluation of Explicit Template Decomposition. The decomposed templates help to capture more accurate pose-dependent deformations and produce more natural wrinkles in unseen poses, especially for outfits that differ largely from the body template. To validate the effectiveness of our decomposition strategy, Tab. 5 reports the ablation experiments on a dress outfit (felice-004) of the ReSynth [39]" + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.945, + 0.512, + 0.957 + ], + "angle": 0, + "content": "507" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.087, + 0.088, + 0.885, + 0.459 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.463, + 0.893, + 0.491 + ], + "angle": 0, + "content": "Figure 6. Comparison of the clothing deformation in unseen poses. Explicit template decomposition (ETD) helps to capture more natural pose-dependent wrinkle details than POP [39]." + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.497, + 0.487, + 0.64 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.085, + 0.644, + 0.461, + 0.658 + ], + "angle": 0, + "content": "Figure 7. Comparison of the learned templates with FITE [32]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.666, + 0.473, + 0.803 + ], + "angle": 0, + "content": "dataset. We can see that the proposed Explicit Template Decomposition (ETD) brings clear performance gains over baseline methods. Fig. 6 shows the visual improvements of pose-dependent wrinkles when applying the explicit template decomposition. Note that the templates are decomposed in an end-to-end manner in our method. Compared with the implicit template learned in the recent approach FITE [32], the explicit templates in our method contain more details, as shown in Fig. 7." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.815, + 0.355, + 0.831 + ], + "angle": 0, + "content": "5. Conclusions and Future Work" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.471, + 0.903 + ], + "angle": 0, + "content": "In this work, we present CloSET, a point-based clothed human modeling method that is built upon a continuous surface and learns to decompose explicit garment templates for better learning of pose-dependent deformations. By learn-" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.5, + 0.892, + 0.621 + ], + "angle": 0, + "content": "ing features on a continuous surface, our solution gets rid of the seam artifacts in previous state-of-the-art point-based methods [36, 39]. Moreover, the explicit template decomposition helps to capture more accurate and natural pose-dependent wrinkles. To facilitate the research in this direction, we also introduce a high-quality real-world scan dataset with diverse outfit styles and accurate body model fitting." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.622, + 0.893, + 0.803 + ], + "angle": 0, + "content": "Limitations and Future Work. Due to the incorrect skinning weight used in our template, the issue of the nonuniform point distribution remains for the skirt and dress outfits. Combining our method with recent learnable skinning solutions [37, 59] could alleviate this issue and further improve the results. Currently, our method does not leverage information from adjacent poses. Enforcing temporal consistency and correspondences between adjacent frames would be interesting for future work. Moreover, incorporating physics-based losses into the learning process like SNUG [61] would also be a promising solution to address the artifacts like self-intersections." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.825, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Acknowledgements. This work was supported by the National Key R&D Program of China (2022YFF0902200), the National Natural Science Foundation of China (No.62125107 and No.61827805), and the China Postdoctoral Science Foundation (No.2022M721844)." + }, + { + "type": "page_number", + "bbox": [ + 0.487, + 0.946, + 0.511, + 0.957 + ], + "angle": 0, + "content": "508" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.116, + 0.47, + 0.158 + ], + "angle": 0, + "content": "[1] Panos Achlioptas, Olga Diamanti, Ioannis Mitliagkas, and Leonidas Guibas. Learning representations and generative models for 3D point clouds. In ICML, pages 40-49, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.159, + 0.47, + 0.213 + ], + "angle": 0, + "content": "[2] Thiemo Alldieck, Gerard Pons-Moll, Christian Theobalt, and Marcus Magnor. Tex2shape: Detailed full human body geometry from a single image. In ICCV, pages 2293-2303, 2019. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.215, + 0.47, + 0.271 + ], + "angle": 0, + "content": "[3] Ziqian Bai, Timur Bagautdinov, Javier Romero, Michael Zollhöfer, Ping Tan, and Shunsuke Saito. AutoAvatar: Autoregressive neural fields for dynamic avatar modeling. ECCV, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.273, + 0.47, + 0.313 + ], + "angle": 0, + "content": "[4] Ilya Baran and Jovan Popovic. Automatic rigging and animation of 3D characters. ACM TOG, 26(3):72-es, 2007. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.316, + 0.47, + 0.37 + ], + "angle": 0, + "content": "[5] Jan Bednářík, Shaifali Parashar, Erhan Gundogdu, Mathieu Salzmann, and Pascal Fua. Shape reconstruction by learning differentiable surface representations. In CVPR, pages 4715-4724, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.373, + 0.47, + 0.427 + ], + "angle": 0, + "content": "[6] Hugo Bertiche, Meysam Madadi, Emilio Tylson, and Sergio Escalera. DeePSD: Automatic deep skinning and pose space deformation for 3D garment animation. In ICCV, pages 5471-5480, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.429, + 0.47, + 0.484 + ], + "angle": 0, + "content": "[7] Bharat Lal Bhatnagar, Garvita Tiwari, Christian Theobalt, and Gerard Pons-Moll. Multi-Garment Net: Learning to dress 3D people from images. In ICCV, pages 5420-5430, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.487, + 0.47, + 0.528 + ], + "angle": 0, + "content": "[8] Andrei Burov, Matthias Nießner, and Justus Thies. Dynamic surface function networks for clothed human bodies. In ICCV, pages 10754-10764, October 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.53, + 0.47, + 0.584 + ], + "angle": 0, + "content": "[9] Xu Chen, Tianjian Jiang, Jie Song, Jinlong Yang, Michael J Black, Andreas Geiger, and Otmar Hilliges. gDNA: Towards generative detailed neural avatars. In CVPR, pages 20427-20437, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.587, + 0.47, + 0.642 + ], + "angle": 0, + "content": "[10] Xu Chen, Yufeng Zheng, Michael J. Black, Otmar Hilliges, and Andreas Geiger. SNARF: Differentiable forward skinning for animating non-rigid neural implicit shapes. In ICCV, pages 11594–11604, October 2021. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.644, + 0.47, + 0.685 + ], + "angle": 0, + "content": "[11] Julian Chibane, Aymen Mir, and Gerard Pons-Moll. Neural unsigned distance fields for implicit function learning. In NeurIPS, pages 21638-21652, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.687, + 0.47, + 0.742 + ], + "angle": 0, + "content": "[12] Enric Corona, Albert Pumarola, Guillem Alenya, Gerard Pons-Moll, and Francesc Moreno-Noguer. SMPLicit: Topology-aware generative model for clothed people. In CVPR, pages 11875-11885, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.744, + 0.47, + 0.785 + ], + "angle": 0, + "content": "[13] Edilson De Aguiar, Leonid Sigal, Adrien Treuille, and Jessica K Hodgins. Stable spaces for real-time clothing. In ACM TOG, volume 29, page 106. ACM, 2010. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.787, + 0.468, + 0.814 + ], + "angle": 0, + "content": "[14] Deform Dynamics. https://deformdynamics.com/.1,3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.817, + 0.47, + 0.871 + ], + "angle": 0, + "content": "[15] Boyang Deng, JP Lewis, Timothy Jeruzalski, Gerard Pons-Moll, Geoffrey Hinton, Mohammad Norouzi, and Andrea Tagliasacchi. Neural articulated shape approximation. In ECCV, pages 612-628, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.47, + 0.901 + ], + "angle": 0, + "content": "[16] Boyang Deng, John P Lewis, Timothy Jeruzalski, Gerard Pons-Moll, Geoffrey Hinton, Mohammad Norouzi, and An-" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.116, + 0.47, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.121 + ], + "angle": 0, + "content": "drea Tagliasacchi. Nasa neural articulated shape approximation. In ECCV, pages 612-628. Springer, 2020. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.123, + 0.892, + 0.164 + ], + "angle": 0, + "content": "[17] Zhantao Deng, Jan Bednárík, Mathieu Salzmann, and Pascal Fua. Better patch stitching for parametric surface reconstruction. In 3DV, pages 593-602, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.166, + 0.892, + 0.222 + ], + "angle": 0, + "content": "[18] Theo Deprelle, Thibault Groueix, Matthew Fisher, Vladimir Kim, Bryan Russell, and Mathieu Aubry. Learning elementary structures for 3D shape generation and matching. In NeurIPS, pages 7433-7443, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.224, + 0.892, + 0.266 + ], + "angle": 0, + "content": "[19] Haoqiang Fan, Hao Su, and Leonidas J Guibas. A point set generation network for 3D object reconstruction from a single image. In CVPR, pages 2463-2471, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.268, + 0.892, + 0.323 + ], + "angle": 0, + "content": "[20] Andrew Feng, Dan Casas, and Ari Shapiro. Avatar reshaping and automatic rigging using a deformable model. In Proceedings of the ACM SIGGRAPH Conference on Motion in Games, pages 57-64, 2015. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.325, + 0.892, + 0.367 + ], + "angle": 0, + "content": "[21] Amos Gropp, Lior Yariv, Niv Haim, Matan Atzmon, and Yaron Lipman. Implicit geometric regularization for learning shapes. In ICML, pages 3569-3579, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.369, + 0.892, + 0.424 + ], + "angle": 0, + "content": "[22] Thibault Groueix, Matthew Fisher, Vladimir G Kim, Bryan C Russell, and Mathieu Aubry. 3D-CODED: 3D correspondences by deep deformation. In ECCV, pages 230-246, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.426, + 0.892, + 0.468 + ], + "angle": 0, + "content": "[23] Peng Guan, Loretta Reiss, David A Hirshberg, Alexander Weiss, and Michael J Black. DRAPE: DRessing Any PErson. ACM TOG, 31(4):35-1, 2012. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.47, + 0.892, + 0.525 + ], + "angle": 0, + "content": "[24] Erhan Gundogdu, Victor Constantin, Amrollah Seifoddini, Minh Dang, Mathieu Salzmann, and Pascal Fua. GarNet: A two-stream network for fast and accurate 3D cloth draping. In CVPR, pages 8739-8748, 2019. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.528, + 0.892, + 0.583 + ], + "angle": 0, + "content": "[25] Boyan Jiang, Xinlin Ren, Mingsong Dou, Xiangyang Xue, Yanwei Fu, and Yinda Zhang. LoRD: Local 4d implicit representation for high-fidelity dynamic human modeling. In ECCV, pages 307-326. Springer, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.585, + 0.892, + 0.639 + ], + "angle": 0, + "content": "[26] Boyi Jiang, Juyong Zhang, Yang Hong, Jinhao Luo, Ligang Liu, and Hujun Bao. BCNet: Learning body and cloth shape from a single image. In ECCV, pages 18-35. Springer, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.643, + 0.892, + 0.696 + ], + "angle": 0, + "content": "[27] Hyomin Kim, Hyeonseo Nam, Junkon Kim, Jaesik Park, and Seungyong Lee. LaplacianFusion: Detailed 3D clothed-human body reconstruction. ACM TOG, 41(6):1-14, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.7, + 0.892, + 0.742 + ], + "angle": 0, + "content": "[28] Zorah Lahner, Daniel Cremers, and Tony Tung. Deepwrinkles: Accurate and realistic clothing modeling. In ECCV, pages 667-684, 2018. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.744, + 0.892, + 0.8 + ], + "angle": 0, + "content": "[29] Ruilong Li, Julian Tanke, Minh Vo, Michael Zollhöfer, Jürgen Gall, Angjoo Kanazawa, and Christoph Lassner. TAVA: Template-free animatable volumetric actors. In ECCV, pages 419-436. Springer, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.802, + 0.892, + 0.857 + ], + "angle": 0, + "content": "[30] Zhe Li, Zerong Zheng, Hongwen Zhang, Chaonan Ji, and Yebin Liu. Avatarcap: Animatable avatar conditioned monocular human volumetric capture. In ECCV, pages 322-341. Springer, 2022. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.859, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[31] Chen-Hsuan Lin, Chen Kong, and Simon Lucey. Learning efficient point cloud generation for dense 3D object reconstruction. In AAAI, pages 7114-7121, 2018. 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.487, + 0.946, + 0.511, + 0.957 + ], + "angle": 0, + "content": "509" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.469, + 0.134 + ], + "angle": 0, + "content": "[32] Siyou Lin, Hongwen Zhang, Zerong Zheng, Ruizhi Shao, and Yebin Liu. Learning implicit templates for point-based clothed human modeling. ECCV, 2022. 2, 3, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.136, + 0.469, + 0.191 + ], + "angle": 0, + "content": "[33] Lijuan Liu, Youyi Zheng, Di Tang, Yi Yuan, Changjie Fan, and Kun Zhou. NeuroSkinning: Automatic skin binding for production characters with deep graph networks. ACM TOG, 38(4):1-12, 2019. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.194, + 0.469, + 0.235 + ], + "angle": 0, + "content": "[34] Matthew Loper, Naureen Mahmood, Javier Romero, Gerard Pons-Moll, and Michael J Black. SMPL: A skinned multiperson linear model. ACM TOG, 34(6):248, 2015. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.238, + 0.469, + 0.279 + ], + "angle": 0, + "content": "[35] William E Lorensen and Harvey E Cline. Marching cubes: A high resolution 3D surface construction algorithm. In SIGGRAPH, volume 21, pages 163-169, 1987. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.282, + 0.469, + 0.336 + ], + "angle": 0, + "content": "[36] Qianli Ma, Shunsuke Saito, Jinlong Yang, Siyu Tang, and Michael J Black. SCALE: Modeling clothed humans with a surface codec of articulated local elements. In CVPR, pages 16082-16093, 2021. 1, 2, 3, 4, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.339, + 0.469, + 0.38 + ], + "angle": 0, + "content": "[37] Qianli Ma, Jinlong Yang, Michael J Black, and Siyu Tang. Neural point-based shape modeling of humans in challenging clothing. 3DV, 2022. 3, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.383, + 0.469, + 0.437 + ], + "angle": 0, + "content": "[38] Qianli Ma, Jinlong Yang, Anurag Ranjan, Sergi Pujades, Gerard Pons-Moll, Siyu Tang, and Michael J Black. Learning to dress 3D people in generative clothing. In CVPR, pages 6469-6478, 2020. 1, 2, 3, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.44, + 0.469, + 0.481 + ], + "angle": 0, + "content": "[39] Qianli Ma, Jinlong Yang, Siyu Tang, and Michael J Black. The power of points for modeling humans in clothing. In ICCV, pages 10974-10984, 2021. 1, 2, 3, 4, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.484, + 0.469, + 0.525 + ], + "angle": 0, + "content": "[40] Xu Ma, Can Qin, Haoxuan You, Haoxi Ran, and Yun Fu. Rethinking network design and local geometry in point cloud: A simple residual mlp framework. In ICLR, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.528, + 0.469, + 0.582 + ], + "angle": 0, + "content": "[41] Lars Mescheder, Michael Oechsle, Michael Niemeyer, Sebastian Nowozin, and Andreas Geiger. Occupancy networks: Learning 3D reconstruction in function space. In CVPR, pages 4460-4470, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.585, + 0.469, + 0.627 + ], + "angle": 0, + "content": "[42] Marko Mihajlovic, Yan Zhang, Michael J Black, and Siyu Tang. LEAP: Learning articulated occupancy of people. In CVPR, pages 10461-10471, 2021. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.63, + 0.469, + 0.669 + ], + "angle": 0, + "content": "[43] Alexandros Neophytou and Adrian Hilton. A layered model of human body and garment deformation. In 3DV, pages 171-178, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.673, + 0.469, + 0.74 + ], + "angle": 0, + "content": "[44] Hayato Onizuka, Zehra Hayirci, Diego Thomas, Akihiro Sugimoto, Hideaki Uchiyama, and Rin-ichiro Taniguchi. TetraTSDF: 3D human reconstruction from a single image with a tetrahedral outer shell. In CVPR, pages 6011–6020, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.744, + 0.469, + 0.798 + ], + "angle": 0, + "content": "[45] Pablo Palafox, Aljaž Božić, Justus Thies, Matthias Nießner, and Angela Dai. NPMs: Neural parametric models for 3D deformable shapes. In ICCV, pages 12695-12705, October 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.802, + 0.469, + 0.842 + ], + "angle": 0, + "content": "[46] Pablo Palafox, Nikolaos Sarafianos, Tony Tung, and Angela Dai. SPAMs: Structured implicit parametric models. In CVPR, pages 12851-12860, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.846, + 0.469, + 0.899 + ], + "angle": 0, + "content": "[47] Junyi Pan, Xiaoguang Han, Weikai Chen, Jiapeng Tang, and Kui Jia. Deep mesh reconstruction from single RGB images via topology modification networks. In ICCV, pages 9963-9972, 2019. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.469, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.147 + ], + "angle": 0, + "content": "[48] Jeong Joon Park, Peter Florence, Julian Straub, Richard Newcombe, and Steven Lovegrove. DeepSDF: Learning continuous signed distance functions for shape representation. In CVPR, pages 165-174, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.149, + 0.892, + 0.203 + ], + "angle": 0, + "content": "[49] Chaitanya Patel, Zhouyingcheng Liao, and Gerard Pons-Moll. TailorNet: Predicting clothing in 3D as a function of human pose, shape and garment style. In CVPR, pages 7363-7373, 2020. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.205, + 0.892, + 0.272 + ], + "angle": 0, + "content": "[50] Georgios Pavlakos, Vasileios Choutas, Nima Ghorbani, Timo Bolkart, Ahmed AA Osman, Dimitrios Tzionas, and Michael J Black. Expressive body capture: 3D hands, face, and body from a single image. In CVPR, pages 10975-10985, 2019. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.274, + 0.892, + 0.328 + ], + "angle": 0, + "content": "[51] Gerard Pons-Moll, Sergi Pujades, Sonny Hu, and Michael Black. ClothCap: Seamless 4d clothing capture and retargeting. ACM TOG, 36(4), 2017. Two first authors contributed equally. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.33, + 0.892, + 0.37 + ], + "angle": 0, + "content": "[52] Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. PointNet: Deep learning on point sets for 3D classification and segmentation. In CVPR, pages 652-660, 2017. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.372, + 0.892, + 0.412 + ], + "angle": 0, + "content": "[53] Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J Guibas. Pointnet++: Deep hierarchical feature learning on point sets in a metric space. NeurIPS, 30, 2017. 3, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.414, + 0.892, + 0.48 + ], + "angle": 0, + "content": "[54] Guocheng Qian, Yuchen Li, Houwen Peng, Jinjie Mai, Hasan Hammoud, Mohamed Elhoseiny, and Bernard Ghanem. Pointnext: Revisiting pointnet++ with improved training and scaling strategies. NeurIPS, 35:23192-23204, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.483, + 0.892, + 0.536 + ], + "angle": 0, + "content": "[55] Shenhan Qian, Jiale Xu, Ziwei Liu, Liqian Ma, and Shenghua Gao. UNIF: United neural implicit functions for clothed human reconstruction and animation. In ECCV, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.539, + 0.892, + 0.552 + ], + "angle": 0, + "content": "[56] Renderpeople, 2020. https://renderpeople.com.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.554, + 0.892, + 0.607 + ], + "angle": 0, + "content": "[57] Shunsuke Saito, Zeng Huang, Ryota Natsume, Shigeo Morishima, Angjoo Kanazawa, and Hao Li. PIFu: Pixel-aligned implicit function for high-resolution clothed human digitization. In ICCV, pages 2304-2314, 2019. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.609, + 0.892, + 0.661 + ], + "angle": 0, + "content": "[58] Shunsuke Saito, Tomas Simon, Jason Saragih, and Hanbyul Joo. PIFuHD: Multi-level pixel-aligned implicit function for high-resolution 3D human digitization. In CVPR, pages 84–93, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.664, + 0.892, + 0.718 + ], + "angle": 0, + "content": "[59] Shunsuke Saito, Jinlong Yang, Qianli Ma, and Michael J Black. SCANimate: Weakly supervised learning of skinned clothed avatar networks. In CVPR, pages 2886-2897, 2021. 1, 2, 3, 4, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.72, + 0.892, + 0.761 + ], + "angle": 0, + "content": "[60] Igor Santesteban, Miguel A. Otaduy, and Dan Casas. Learning-Based Animation of Clothing for Virtual Try-On. Comput. Graph. Forum, 38(2):355-366, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.763, + 0.892, + 0.801 + ], + "angle": 0, + "content": "[61] Igor Santesteban, Miguel A Otaduy, and Dan Casas. SNUG: Self-supervised neural dynamic garments. In CVPR, pages 8140-8150, 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.804, + 0.892, + 0.857 + ], + "angle": 0, + "content": "[62] Igor Santesteban, Nils Thuerey, Miguel A Otaduy, and Dan Casas. Self-supervised collision handling via generative 3D garment models for virtual try-on. In CVPR, pages 11763-11773, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.859, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[63] Yu Shen, Junbang Liang, and Ming C Lin. Gan-based garment generation using sewing pattern images. In ECCV, volume 1, page 3, 2020. 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.487, + 0.946, + 0.512, + 0.957 + ], + "angle": 0, + "content": "510" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.147 + ], + "angle": 0, + "content": "[64] Garvita Tiwari, Bharat Lal Bhatnagar, Tony Tung, and Gerard Pons-Moll. SIZER: A dataset and model for parsing 3D clothing and learning size sensitive 3D clothing. In ECCV, volume 12348, pages 1-18, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.15, + 0.47, + 0.205 + ], + "angle": 0, + "content": "[65] Garvita Tiwari, Nikolaos Sarafianos, Tony Tung, and Gerard Pons-Moll. Neural-GIF: Neural generalized implicit functions for animating people in clothing. In ICCV, pages 11708-11718, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.208, + 0.47, + 0.263 + ], + "angle": 0, + "content": "[66] Raquel Vidaurre, Igor Santesteban, Elena Garces, and Dan Casas. Fully convolutional graph neural networks for parametric virtual try-on. In Comput. Graph. Forum, volume 39, pages 145-156. Wiley Online Library, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.266, + 0.47, + 0.32 + ], + "angle": 0, + "content": "[67] Shaofei Wang, Marko Mihajlovic, Qianli Ma, Andreas Geiger, and Siyu Tang. Meta­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­.\nclothed human models from few depth images. In NeurIPS, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.324, + 0.47, + 0.377 + ], + "angle": 0, + "content": "[68] Jane Wu, Zhenglin Geng, Hui Zhou, and Ronald Fedkiw. Skinning a parameterization of three-dimensional space for neural network cloth. arXiv preprint arXiv:2006.04874, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.381, + 0.47, + 0.436 + ], + "angle": 0, + "content": "[69] Donglai Xiang, Fabian Prada, Timur Bagautdinov, Weipeng Xu, Yuan Dong, He Wen, Jessica Hodgins, and Chenglei Wu. Modeling clothing as a separate layer for an animatable human avatar. ACM TOG, dec 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.439, + 0.47, + 0.481 + ], + "angle": 0, + "content": "[70] Yuliang Xiu, Jinlong Yang, Dimitrios Tzionas, and Michael J Black. ICON: implicit clothed humans obtained from normals. In CVPR, pages 13286-13296. IEEE, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.483, + 0.47, + 0.538 + ], + "angle": 0, + "content": "[71] Jinlong Yang, Jean-Sebastien Franco, Franck Hetroy-Wheeler, and Stefanie Wuhrer. Analyzing clothing layer deformation statistics of 3D human motions. In ECCV, September 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.541, + 0.47, + 0.594 + ], + "angle": 0, + "content": "[72] Shan Yang, Zherong Pan, Tanya Amert, Ke Wang, Licheng Yu, Tamara Berg, and Ming C Lin. Physics-inspired garment recovery from a single-view image. ACM TOG, 37(5):1-14, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.599, + 0.47, + 0.653 + ], + "angle": 0, + "content": "[73] Tao Yu, Zerong Zheng, Kaiwen Guo, Pengpeng Liu, Qionghai Dai, and Yebin Liu. Function4D: Real-time human volumetric capture from very sparse consumer rgbd sensors. In CVPR, June 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.656, + 0.47, + 0.697 + ], + "angle": 0, + "content": "[74] Ilya Zakharkin, Kirill Mazur, Artur Grigorev, and Victor Lempitsky. Point-based modeling of human clothing. In ICCV, pages 14718-14727, October 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.7, + 0.47, + 0.754 + ], + "angle": 0, + "content": "[75] Zerong Zheng, Han Huang, Tao Yu, Hongwen Zhang, Yandong Guo, and Yebin Liu. Structured local radiance fields for human avatar modeling. In CVPR, pages 15893-15903, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.758, + 0.47, + 0.799 + ], + "angle": 0, + "content": "[76] Zerong Zheng, Tao Yu, Qionghai Dai, and Yebin Liu. Deep implicit templates for 3D shape representation. In CVPR, pages 1429-1439, 2021. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.802, + 0.47, + 0.856 + ], + "angle": 0, + "content": "[77] Zerong Zheng, Tao Yu, Yebin Liu, and Qionghai Dai. PaMIR: Parametric model-conditioned implicit representation for image-based human reconstruction. IEEE TPAMI, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.859, + 0.47, + 0.902 + ], + "angle": 0, + "content": "[78] Zerong Zheng, Tao Yu, Yixuan Wei, Qionghai Dai, and Yebin Liu. DeepHuman: 3D human reconstruction from a single image. In ICCV, pages 7739-7749, 2019. 3" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.092, + 0.893, + 0.162 + ], + "angle": 0, + "content": "[79] Heming Zhu, Yu Cao, Hang Jin, Weikai Chen, Dong Du, Zhangye Wang, Shuguang Cui, and Xiaoguang Han. Deep Fashion3D: A dataset and benchmark for 3D garment reconstruction from single images. In ECCV, volume 12346, pages 512-530, 2020. 2" + }, + { + "type": "page_number", + "bbox": [ + 0.487, + 0.946, + 0.51, + 0.957 + ], + "angle": 0, + "content": "511" + } + ] +] \ No newline at end of file diff --git a/2023/CloSET_ Modeling Clothed Humans on Continuous Surface With Explicit Template Decomposition/4725afdd-f941-4b99-a289-05a751ed5c4e_origin.pdf b/2023/CloSET_ Modeling Clothed Humans on Continuous Surface With Explicit Template Decomposition/4725afdd-f941-4b99-a289-05a751ed5c4e_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..da0e73ea4f60feb8f5ba2625bbed9a23e80d6759 --- /dev/null +++ b/2023/CloSET_ Modeling Clothed Humans on Continuous Surface With Explicit Template Decomposition/4725afdd-f941-4b99-a289-05a751ed5c4e_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:860c35e8d0210a15fe730e7327be046116efd1c44c7e137dcf0843491f97d6c4 +size 6947476 diff --git a/2023/CloSET_ Modeling Clothed Humans on Continuous Surface With Explicit Template Decomposition/full.md b/2023/CloSET_ Modeling Clothed Humans on Continuous Surface With Explicit Template Decomposition/full.md new file mode 100644 index 0000000000000000000000000000000000000000..baf40a3ed4ef6e0f38a768d546cf4e5acc86d483 --- /dev/null +++ b/2023/CloSET_ Modeling Clothed Humans on Continuous Surface With Explicit Template Decomposition/full.md @@ -0,0 +1,294 @@ +# CloSET: Modeling Clothed Humans on Continuous Surface with Explicit Template Decomposition + +Hongwen Zhang $^{1}$ Siyou Lin $^{1}$ Ruizhi Shao $^{1}$ Yuxiang Zhang $^{1}$ Zerong Zheng $^{1}$ Han Huang $^{2}$ Yandong Guo $^{2}$ Yebin Liu $^{1}$ Tsinghua University ${}^{2}$ OPPO Research Institute + +# Abstract + +Creating animatable avatars from static scans requires the modeling of clothing deformations in different poses. Existing learning-based methods typically add pose-dependent deformations upon a minimally-clothed mesh template or a learned implicit template, which have limitations in capturing details or hinder end-to-end learning. In this paper, we revisit point-based solutions and propose to decompose explicit garment-related templates and then add pose-dependent wrinkles to them. In this way, the clothing deformations are disentangled such that the pose-dependent wrinkles can be better learned and applied to unseen poses. Additionally, to tackle the seam artifact issues in recent state-of-the-art point-based methods, we propose to learn point features on a body surface, which establishes a continuous and compact feature space to capture the fine-grained and pose-dependent clothing geometry. To facilitate the research in this field, we also introduce a high-quality scan dataset of humans in real-world clothing. Our approach is validated on two existing datasets and our newly introduced dataset, showing better clothing deformation results in unseen poses. The project page with code and dataset can be found at https://www.liuyebin.com/closet. + +# 1. Introduction + +Animating 3D clothed humans requires the modeling of pose-dependent deformations in various poses. The diversity of clothing styles and body poses makes this task extremely challenging. Traditional methods are based on either simple rigging and skinning [4,20,33] or physics-based simulation [14, 23, 24, 49], which heavily rely on artist efforts or computational resources. Recent learning-based methods [10, 36, 39, 59] resort to modeling the clothing deformation directly from raw scans of clothed humans. Despite the promising progress, this task is still far from being solved due to the challenges in clothing representations, generalization to unseen poses, and data acquisition, etc. + +![](images/9120cdfd2b4793367ccd1702c81f777d90bb4f9e6c091172060a3d4e1a02a13b.jpg) +Figure 1. Our method learns to decompose garment templates (top row) and add pose-dependent wrinkles upon them (bottom row). + +For the modeling of pose-dependent garment geometry, the representation of clothing plays a vital role in a learning-based scheme. As the relationship between body poses and clothing deformations is complex, an effective representation is desirable for neural networks to capture pose-dependent deformations. In the research of this line, meshes [2, 12, 38], implicit fields [10, 59], and point clouds [36, 39] have been adopted to represent clothing. In accordance with the chosen representation, the clothing deformation and geometry features are learned on top of a fixed-resolution template mesh [8, 38], a 3D implicit sampling space [10, 59], or an unfolded UV plane [2, 12, 36, 39]. Among these representations, the mesh is the most efficient one but is limited to a fixed topology due to its discretization scheme. The implicit fields naturally enable continuous feature learning in a resolution-free manner but are too flexible to satisfy the body structure prior, leading to geometry artifacts in unseen poses. The point clouds enjoy the compact nature and topology flexibility and have shown promising results in the recent state-of-the-art solutions [36, 39] to represent clothing, but the feature learning on UV planes still leads to discontinuity artifacts between body parts. + +To model the pose-dependent deformation of clothing, body templates such as SMPL [34] are typically leveraged to account for articulated motions. However, a body template alone is not ideal, since the body template only models the minimally-clothed humans and may hinder the learning of actual pose-dependent deformations, especially in cases of loose clothing. To overcome this issue, recent implicit approaches [59] make attempts to learn skinning weights in the 3D space to complement the imperfect body templates. However, their pose-dependent deformations are typically coarse due to the difficulty in learning implicit fields. For explicit solutions, the recent approach [32] suggests learning coarse templates implicitly at first and then the pose-dependent deformations explicitly. Despite its effectiveness, such a workaround requires a two-step modeling procedure and hinders end-to-end learning. + +In this work, we propose CloSET, an end-to-end method to tackle the above issues by modeling Clothed humans on a continuous Surface with Explicit Template decomposition. We follow the spirit of recent state-of-the-art point-based approaches [32, 36, 39] as they show the efficiency and potential in modeling real-world garments. We take steps forward in the following aspects for better point-based modeling of clothed humans. First, we propose to decompose the clothing deformations into explicit garment templates and pose-dependent wrinkles. Specifically, our method learns a garment-related template and adds the pose-dependent displacement upon them, as shown in Fig. 1. Such a garment-related template preserves a shared topology for various poses and enables better learning of pose-dependent wrinkles. Different from the recent solution [32] that needs two-step procedures, our method can decompose the explicit templates in an end-to-end manner with more garment details. Second, we tackle the seam artifact issues that occurred in recent point-based methods [36, 39]. Instead of using unfolded UV planes, we propose to learn point features on a body surface, which supports a continuous and compact feature space. We achieve this by learning hierarchical point-based features on top of the body surface and then using barycentric interpolation to sample features continuously. Compared to feature learning in the UV space [39], on template meshes [8, 38], or in the 3D implicit space [57, 59], our body surface enables the network to capture not only fine-grained details but also long-range part correlations for pose-dependent geometry modeling. Third, we introduce a new scan dataset of humans in real-world clothing, which contains more than 2,000 high-quality scans of humans in diverse outfits, hoping to facilitate the research in this field. The main contributions of this work are summarized below: + +- We propose a point-based clothed human modeling method by decomposing clothing deformations into explicit garment templates and pose-dependent wrin + +kles in an end-to-end manner. These learnable templates provide a garment-aware canonical space so that pose-dependent deformations can be better learned and applied to unseen poses. + +- We propose to learn point-based clothing features on a continuous body surface, which allows a continuous feature space for fine-grained detail modeling and helps to capture long-range part correlations for pose-dependent geometry modeling. +- We introduce a new high-quality scan dataset of clothed humans in real-world clothing to facilitate the research of clothed human modeling and animation from real-world scans. + +# 2. Related Work + +Representations for Modeling Clothed Humans. A key component in modeling clothed humans is the choice of representation, which mainly falls into two categories: implicit and explicit representations. + +Implicit Modeling. Implicit methods [3, 10-12, 15, 21, 29, 41, 42, 48, 59, 67, 70, 77] represent surfaces as the level set of an implicit neural scalar field. Recent state-of-the-art methods typically learn the clothing deformation field with a canonical space decomposition [9, 30, 45, 59, 65] or part-based modeling strategies [16, 25, 46, 55, 75]. Compared to mesh templates, implicit surfaces are not topologically constrained to specific templates [57, 58], and can model various clothes with complex topology. However, the learning space of an implicit surface is the whole 3D volume, which makes training and interpolation difficult, especially when the numbers of scan data are limited. + +Explicit Modeling. Mesh surfaces, the classic explicit representation, currently dominate the field of 3D modeling [6-8, 13, 23, 24, 26-28, 38, 43, 49, 60, 62, 64, 66, 69, 72] with their compactness and high efficiency in downstream tasks such as rendering, but they are mostly limited to a fixed topology and/or require scan data registered to a template. Thus, mesh-based representations forbid the learning of a universal model for topologically varying clothing types. Though some approaches have been proposed to allow varying mesh topology [44, 47, 63, 68, 79], they are still limited in their expressiveness. Point clouds enjoy both compactness and topological flexibility. Previous work generates sparse point clouds for 3D representation [1, 19, 31, 74]. However, the points need to be densely sampled over the surface to model surface geometry accurately. Due to the difficulty of generating a large point set, recent methods group points into patches [5, 17, 18, 22]. Each patch maps the 2D UV space to the 3D space, allowing arbitrarily dense sampling within this patch. SCALE [36] successfully applies this idea to modeling clothed humans, but produces notable discontinuity artifacts near patch boundaries. + +POP [39] further utilizes a single fine-grained UV map for the whole body surface, leading to a more topologically flexible representation. However, the discontinuity of the UV map may lead to seam artifacts in POP. Very recently, FITE [32] suggests learning implicit coarse templates [76] at first and then explicit fine details. Despite the efficacy, it requires a two-step modeling procedure. Concurrently, SkiRT [37] proposes to improve the body template by learning the blend skinning weights with several data terms. In contrast, our method applies regularization to achieve template decomposition in an end-to-end manner and learns the point-based pose-dependent displacement more effectively. + +Pose-dependent Deformations for Animation. In the field of character animation, traditional methods utilize rigging and skinning techniques to repose characters [4, 20, 33, 34, 50]. but they fail to model realistic pose-dependent clothing deformations such as wrinkles and sliding motions between clothes and body. We conclude two key ingredients in modeling pose-dependent clothing: (i) pose-dependent feature learning; (ii) datasets with realistic clothing. + +Pose-dependent Feature Learning. Some traditional methods directly incorporate the entire pose parameters into the model [28, 38, 49, 71]. Such methods easily overfit on pose parameters and introduce spurious correlations, causing bad generalization to unseen poses. Recent work explores poses conditioning with local features, either with point clouds [36, 39] or implicit surfaces [59, 67], and shows superiority in improving geometry quality and in eliminating spurious correlations. Among them, the most relevant to ours is POP [39], which extracts local pose features by utilizing convolution on a UV position map. Despite its compelling performance, POP suffers from artifacts inherent to its UV-based representation, hence the convolution on the UV map produces discontinuity near UV islands' boundaries [39]. We address this issue by discarding the UV-based scheme and returning to the actual 3D body surface. We attach the features to a uniform set of points on a T-posed body template, and process them via a PointNet++ [53] structure for pose-dependent modeling. As validated by our experiments, our pose embedding method leads to both qualitative and quantitative improvements. + +Clothed Human Datasets. Another challenging issue for training an animatable avatar is the need for datasets of clothed humans in diverse poses. There are considerable efforts seeking to synthesize clothed datasets with physics-based simulation [14,23,24,39,49]. Although they are diverse in poses, there remains an observable domain gap between synthetic clothes and real data. Acquiring clothed scans with realistic details of clothing deformations [38, 51, 56, 73, 78] is crucial for the development of learning-based methods in this field. + +# 3. Method + +As illustrated in Fig. 2, the proposed method CloSET learns garment-related and pose-dependent features on body surfaces (see Sec. 3.1), which can be sampled in a continuous manner and fed into two decoders for the generation of explicit garment templates and pose-dependent wrinkles (see Sec. 3.2). + +# 3.1. Continuous Surface Features + +As most parts of the clothing are deformed smoothly in different poses, a continuous feature space is desirable to model the garment details and pose-dependent garment geometry. To this end, our approach first learns features on top of a body template surface, i.e., a SMPL [34] or SMPL-X [50] model in a T-pose. Note that these features are not limited to those on template vertices as they can be continuously sampled from the body surface via barycentric interpolation. Hence, our feature space is more continuous than UV-based [2, 39] spaces, while being more compact than 3D implicit feature fields [59, 67]. + +To model pose-dependent clothing deformations, the underlying unclothed body model is taken as input to the geometry feature encoder. For each scan, let $\mathbf{V}^u = \{\pmb{v}_n^u\}_{n=1}^N$ denote the posed vertex positions of the fitted unclothed body model, where $N = 6890$ for SMPL [34] and $N = 10475$ for SMPL-X [50]. These posed vertices act as the pose code and will be paired with the template vertices $\mathbf{V}^t = \{\pmb{v}_n^t\}_{n=1}^N$ of the body model in a T-pose, which shares the same mesh topology with $\mathbf{V}^u$ . These point pairs are processed by the pose encoder $\mathcal{F}_p$ to generate the pose-dependent geometry features $\{\phi_p(\pmb{v}_n^t) \in \mathbb{R}^{C_p}\}_{n=1}^N$ at vertices $\mathbf{V}^t$ , i.e., + +$$ +\left\{\phi_ {p} \left(\boldsymbol {v} _ {n} ^ {t}\right) \in \mathbb {R} ^ {C _ {p}} \right\} _ {n = 1} ^ {N} = \mathcal {F} _ {p} \left(\mathbf {V} ^ {t}, \mathbf {V} ^ {u}\right). \tag {1} +$$ + +To learn hierarchical features with different levels of receptive fields, we adopt PointNet++ [53] as the architecture of the pose encoder $\mathcal{F}_p$ , where vertices $\mathbf{V}^t$ are treated as the input point cloud in the PointNet++ network, while vertices $\mathbf{V}^u$ act as the feature of $\mathbf{V}^t$ . As the template vertices $\mathbf{V}^t$ are constant, the encoder $\mathcal{F}_p$ can focus on the feature learned from the posed vertices $\mathbf{V}^u$ . Moreover, the PointNet++ based $\mathcal{F}_p$ first abstracts features from the template vertices $\mathbf{V}^t$ to sparser points $\{\mathbf{V}_l^t\}_{l=1}^L$ at $L$ levels, where the number of $\mathbf{V}_l^t$ decreases with $l$ increasing. Then, the features at $\{\mathbf{V}_l^t\}_{l=1}^L$ are further propagated back to $\mathbf{V}^t$ successively. In this way, the encoder can capture the long-range part correlations of the pose-dependent deformations. + +Similar to POP [39], our method can be trained under multi-outfit or outfit-specific settings. When trained with multiple outfits, the pose-dependent deformation should be aware of the outfit type, and hence requires the input of the garment features. Specifically, the garment-related features $\{\phi_g(\pmb{v}_n^t) \in \mathbb{R}^{C_g}\}_{n=1}^N$ are also defined on template + +![](images/986709cebc27f14be506c0ee7a592d4621578e2cdc4a6679da0a796dbf422979.jpg) +Figure 2. Overview of the proposed method CloSET. Given an input body model, its pose code and garment code are processed hierarchically by point-based pose and garment encoders $\mathcal{F}_p$ and $\mathcal{F}_g$ for the learning of surface features $\phi_p$ and $\phi_g$ . For any point $\pmb{p}_i^t$ lying on the template surface, its features $\phi (\pmb{p}_i^t)$ are sampled from surface features accordingly and fed into two decoders for the prediction of the explicit garment template and pose-dependent wrinkle displacements, which will be combined and transformed to the clothing point cloud. + +![](images/3a4fb4fca7e8a7f309d99a898b4780b117f625d9bb392b329bd1dfe4fd193c44.jpg) +Figure 3. Comparison of the bilinear interpolation on the UV plane and the barycentric interpolation on the surface. + +vertices $\mathbf{V}^t$ , which are learned by feeding the garment code $\{\phi_{gc}(\pmb{v}_n^t)\}_{n=1}^N$ to a smaller PointNet++ encoder $\mathcal{F}_g$ . Note that the garment-related features $\phi_g$ are shared for each outfit across all poses and optimized during the training. Since both the pose-dependent and garment-related geometry features are aligned with each other, we denote them as the surface features $\{\phi(\pmb{v}_n^t)\}_{n=1}^N$ for simplicity. Note that the input of $\phi_g(\pmb{p}_i^t)$ has no side effect on the results when trained with only one outfit, as the garment features are invariant to the input poses. + +Continuous Feature Interpolation. In the implicit modeling solutions [42, 57, 59], features are learned in a spatially continuous manner, which contributes to the fine-grained modeling of clothing details. To sample continuous features in our scheme, we adopt barycentric interpolation on the surface features $\phi$ . As illustrated in Fig. 3, for any point $p_i^t = \mathbf{V}^t (\pmb {b}_i)$ lying on the template surface, where $\pmb {b}_i = [n_{i1},n_{i2},n_{i3},b_{i1},b_{i2},b_{i3}]$ denotes the corresponding vertex indices and barycentric coordinates in $\mathbf{V}^t$ . Then, the corresponding surface features can be retrieved via barycentric interpolation, i.e., + +$$ +\phi \left(\boldsymbol {p} _ {i} ^ {t}\right) = \sum_ {j = 1} ^ {3} \left(b _ {i j} * \phi \left(\boldsymbol {v} _ {n _ {i j}} ^ {t}\right)\right). \tag {2} +$$ + +In this way, the point features are not limited to those learned on the template vertices and are continuously defined over the whole body surface without the seaming discontinuity issue in the UV plane. + +# 3.2. Point-based Clothing Deformation + +Following previous work [36, 39], our approach represents the clothed body as a point cloud. For any point on the surface of the unclothed body model, the corresponding features are extracted from surface features to predict its displacement and normal vector. + +Explicit Template Decomposition. Instead of predicting the clothing deformation directly, our method decomposes the deformations into two components: garment-related template displacements and pose-dependent wrinkle displacements. To achieve this, the garment-related template is learned from the garment-related features and shared across all poses. Meanwhile, the learning of pose-dependent wrinkles are conditioned on both garment-related and pose-dependent features. Specifically, for the point $\pmb{p}_i^u = \mathbf{V}^u (\pmb {b}_i)$ at the unclothed body mesh, it has the same vertex indices and barycentric coordinates $\pmb{b}_{i}$ as the point $\pmb{p}_i^t$ on the template surface. The pose-dependent and garment-related features of the point $\pmb{p}_i^u$ are first sampled according to Eq. (2) based on $\pmb{p}_i^t = \mathbf{V}^t (\pmb {b}_i)$ and then further fed into the garment decoder $\mathcal{D}_g$ and the pose decoder $\mathcal{D}_p$ for displacement predictions, i.e., + +$$ +\begin{array}{l} \boldsymbol {r} _ {i} ^ {g} = \mathcal {D} _ {g} \left(\phi_ {g} \left(\boldsymbol {p} _ {i} ^ {t}\right), \boldsymbol {p} _ {i} ^ {t}\right), \tag {3} \\ \boldsymbol {r} _ {i} ^ {p} = \mathcal {D} _ {p} \left(\oplus \left(\phi_ {g} \left(\boldsymbol {p} _ {i} ^ {t}\right), \phi_ {p} \left(\boldsymbol {p} _ {i} ^ {t}\right)\right), \boldsymbol {p} _ {i} ^ {t}\right), \\ \end{array} +$$ + +where $\oplus$ denotes the concatenation operation, $r_i^g$ and $r_i^p$ are the displacements for garment templates and pose-dependent wrinkles, respectively. Finally, $r_i^g$ and $r_i^p$ will be added together as the clothing deformation $r_i = r_i^g + r_i^p$ . + +Local Transformation. Similar to [36, 39], the displacement $\boldsymbol{r}_i$ is learned in a local coordinate system. It is further transformed to the world coordinate system by applying the following transformation, i.e., $\boldsymbol{x}_i = \mathcal{T}_i \boldsymbol{r}_i + \boldsymbol{p}_i^u$ , where $\mathcal{T}_i$ denotes the local transformation calculated based on the unclothed body model. Following [36, 39], the transformation matrix $\mathcal{T}_i$ is defined at the point $\boldsymbol{p}_i^u$ on the unclothed body model, which naturally supports the barycentric interpolation. Similarly, the normal $\boldsymbol{n}_i$ of each point is predicted together with $\boldsymbol{r}_i$ from $\mathcal{D}$ and transformed by $\mathcal{T}_i$ . + +# 3.3. Loss Functions + +Following previous work [36, 39], the point-based clothing deformation is learned with the summation of loss functions: $\mathcal{L} = \mathcal{L}_{data} + \lambda_{rgl}\mathcal{L}_{rgl}$ . where $\mathcal{L}_{data}$ and $\mathcal{L}_{rgl}$ denote the data and regularization terms respectively, and the weight $\lambda_{rgl}$ balances the loss terms. + +Data Term. The data term $\mathcal{L}_{data}$ is calculated on the final predicted points and normals, i.e., $\mathcal{L}_{data} = \lambda_p\mathcal{L}_p + \lambda_n\mathcal{L}_n$ . Specifically, $\mathcal{L}_p$ is the normalized Chamfer distance to minimize the bi-directional distances between the point sets of the prediction and the ground-truth scan: + +$$ +\mathcal {L} _ {p} = C h a m f e r \left(\{\boldsymbol {x} _ {i} \} _ {i = 1} ^ {M}, \{\hat {\boldsymbol {x}} _ {j} \} _ {j = 1} ^ {N _ {s}}\right) = +$$ + +$$ +\frac {1}{M} \sum_ {i = 1} ^ {M} \min _ {j} \| \boldsymbol {x} _ {i} - \hat {\boldsymbol {x}} _ {j} \| _ {2} ^ {2} + \frac {1}{N _ {s}} \sum_ {j = 1} ^ {N _ {s}} \min _ {i} \| \boldsymbol {x} _ {i} - \hat {\boldsymbol {x}} _ {j} \| _ {2} ^ {2}, \tag {4} +$$ + +where $\hat{\pmb{x}}_j$ is the point sampled from the ground-truth surface, $M$ and $N_{s}$ denote the number of the predicted and ground-truth points, respectively. + +The normal loss $\mathcal{L}_n$ is an averaged $L1$ distance between the normal of each predicted point and its nearest ground-truth counterpart: + +$$ +\mathcal {L} _ {n} = L 1 \left(\left\{\boldsymbol {n} _ {i} \right\} _ {i = 1} ^ {M}, \left\{\hat {\boldsymbol {n}} _ {i} \right\} _ {i = 1} ^ {M}\right) = \frac {1}{M} \sum_ {i = 1} ^ {M} \| \boldsymbol {n} _ {i} - \hat {\boldsymbol {n}} _ {i} \|, \tag {5} +$$ + +where $\hat{\pmb{n}}_i$ is the normal of its nearest point in the groundtruth point set. + +Note that we do not apply data terms on the garment templates, as we found such a strategy leads to noisy template learning in our experiments. + +Regularization Term. The regularization terms are added to prevent the predicted deformations from being extremely large and regularize the garment code. Moreover, following the previous implicit template learning solution [30, 76], we also add regularization on the pose-dependent displacement $\boldsymbol{r}_i^p$ to encourage it to be as small as possible. As the pose-dependent displacement represents the clothing deformation in various poses, such a regularization implies that the pose-invariant deformation should be retained in the template displacement $\boldsymbol{r}_i^g$ , which forms + +the garment-related template shared by all poses. Overall, the regularization term can be written as follows: + +$$ +\mathcal {L} _ {r g l} = \frac {1}{M} \sum_ {i = 1} ^ {M} \| \mathbf {r} _ {i} \| _ {2} ^ {2} + \frac {\lambda_ {p d}}{M} \sum_ {i = 1} ^ {M} \| \mathbf {r} _ {i} ^ {p} \| _ {2} ^ {2} + \frac {\lambda_ {g c}}{N} \sum_ {n = 1} ^ {N} \| \phi_ {g c} (\boldsymbol {v} _ {n} ^ {t}) \| _ {2} ^ {2}. \tag {6} +$$ + +# 4. Experiments + +Network Architecture. For a fair comparison with POP [39], we modify the official PointNet++ [53] (PN++) architecture so that our encoders have comparable network parameters as POP [39]. The modified PointNet++ architecture has 6 layers for feature abstraction and 6 layers for feature propagation (i.e., $L = 6$ ). Since the input point cloud $\mathbf{V}^t$ has constant coordinates, the farthest point sampling in PointNet++ is only performed at the first forward process, and the sampling indices are saved for the next run. In this way, the runtime is significantly reduced for both training and inference so that our pose and garment encoders can have similar network parameters and runtime speeds to POP. Note that the pose and garment encoders in our method can also be replaced with recent state-of-the-art point-based encoders such as PointMLP [40] and PointNeXt [54]. More details about the network architecture and implementation can be found in the Supp.Mat. + +Datasets. We use CAPE [38], ReSynth [39], and our newly introduced dataset THuman-CloSET for training and evaluation. + +CAPE [38] is a captured human dataset consisting of multiple humans in various motions. The outfits in this dataset mainly include common clothing such as T-shirts. We follow SCALE [36] to choose blazerlong (with outfits of blazer jacket and long trousers) and shortlong (with outfits of short T-shirt and long trousers) from subject 03375 to validate the efficacy of our method. + +ReSynth [39] is a synthetic dataset introduced in POP [39]. It is created by using physics simulation, and contains challenging outfits such as skirts and jackets. We use the official training and test split as [39]. + +THuman-CloSET is our newly introduced dataset, containing high-quality clothed human scans captured by a dense camera rig. We introduce THuman-CloSET for the reason that existing pose-dependent clothing datasets [38, 39] are with either relatively tight clothing or synthetic clothing via physics simulation. In THuman-CloSET, there are more than 2,000 scans of 15 outfitsits with a large variation in clothing style, including T-shirts, pants, skirts, dresses, jackets, and coats, to name a few. For each outfit, the subject is guided to perform different poses by imitating the poses in CAPE. Moreover, each subject has a scan with minimal clothing in A-pose. THuman-CloSET contains well-fitted body models in the form of SMPL-X [50]. Note that + +Table 1. Quantitative comparison with previous point-based methods on ReSynth. † denotes the methods using 1/8 training data. + +
MethodChamfer-L2↓Normal diff. ↓
MeanMaxMeanMax
outfit-specific
SCALE [36]1.4918.4511.0411.321
POP [39]1.3567.3391.0131.289
multi-outfit
Baseline (POP [39]) †1.4907.8591.0501.326
Baseline w. PN++ †1.2905.9401.0281.330
CloSET (Ours) †1.2405.5431.0191.315
+ +Table 2. Quantitative comparison of different methods on the proposed THuman-CloSET dataset in the outfit-specific setting. + +
Subject IDSCANImateSNARFPOPCloSET
CDNMLCDNMLCDNMLCDNML
sweater-0001.061.647.112.090.761.550.681.48
longshirt-0011.421.856.662.211.541.831.391.71
skirt-0051.931.749.392.311.661.431.491.36
+ +the loose clothing makes the fitting of the underlying body models quite challenging. For more accurate fitting of the body models, we first fit a SMPL-X model on the scan of the subject in minimal clothing and then adopt its shape parameters for fitting the outfit scans in different poses. More details can be found in the Supp.Mat. In our experiments, we use the outfit scans in 100 different poses for training and use the remaining poses for evaluation. We hope our new dataset can open a promising direction for clothed human modeling and animation from real-world scans. + +Metrics. Following previous work [39], we generate 50K points from our method and point-based baselines and adopt the Chamfer Distance (see Eq. (4)) and the $L1$ normal discrepancy (see Eq. (5)) for quantitative evaluation. By default, the Chamfer distance (CD) and normal discrepancy (NML) are reported in the unit of $\times 10^{-4}m^2$ and $\times 10^{-1}$ , respectively. To evaluate the implicit modeling methods, the points are sampled from the surface extracted using Marching Cubes [35]. + +# 4.1. Comparison with the State-of-the-art Methods + +We compare results with recent state-of-the-art methods, including point-based approaches SCALE [36], POP [39], and SkiRT [37], and implicit approaches SCANimate [59] and SNARF [10]. + +ReSynth. Tab. 1 reports the results of the pose-dependent clothing predictions on unseen motion sequences from the ReSynth [39] dataset, where all 12 outfits fits are used for eval + +Table 3. Quantitative comparison of different methods on the ReSynth dataset in the outfit-specific setting. The garment styles are non-skirt, skirt, and dress for carla-004, Christine-027, and felice-004, respectively. + +
Subject IDSCANImatePOPSkiRTCloSET
CDNMLCDNMLCDNMLCDNML
carla-0040.901.520.511.020.481.060.491.04
christine-0273.211.661.720.971.540.991.490.97
felice-00420.792.947.341.246.451.256.011.16
+ +uation. As can be seen, the proposed approach has the lowest mean and max errors, which outperforms all other approaches including POP [39]. Note that our approach needs fewer data for the pose-dependent deformation modeling. By using only 1/8 data, our approach achieves a performance comparable to or even better than other models trained with full data. Tab. 3 also reports outfit-specific performances on 3 selected subject-outfit types, including jackets, skirts, and dresses. In comparison with the recent state-of-the-art method SkiRT [37], our method achieves better results on challenging skirt/dress outfits and comparable results on non-skirt clothing. + +THuman-CloSET. The effectiveness of our method is also validated on our real-world THuman-CloSET dataset. The sparse training poses and loose clothing make this dataset very challenging for clothed human modeling. Tab. 2 reports the quantitative comparisons of different methods on three representative outfits. Fig. 4 also shows example results of different methods, where we follow previous work [39] to obtain meshed results via Poisson surface reconstruction. We can see that our method generalizes better to unseen poses and produces more natural pose-dependent wrinkles than other methods. In our experiments, we found that SNARF [10] fails to learn correct skinning weights due to loose clothing and limited training poses. As discussed in FITE [32], there is an ill-posed issue of jointly optimizing the canonical shape and the skinning fields, which becomes more severe in our dataset. + +# 4.2. Ablation Study + +Evaluation of Continuous Surface Features. The point features in our method are learned on the body surface, which provides a continuous and compact feature learning space. To validate this, Tab. 4 summarizes the feature learning space of different approaches and their performances on two representative outfits from CAPE [38]. Here, we only include the proposed Continuous Surface Features (CSF) in Tab. 4 by applying the continuous features on POP [39] for fair comparisons with SCALE [36] and POP [39]. As discussed previously, existing solutions learn features either in a discontinuous space (e.g., CAPE [38] on the fixed resolu + +![](images/ef3d7d5c5c228210c963ddd9ab79b78c1e8b679a7e644349e3d7b564cbb524b6.jpg) +Figure 4. Comparison of different clothed human modeling methods on the proposed real-world scan dataset. + +![](images/b67fe7458d39ddb96fdec5ae53b785611ac7b2f4d219d300ec39e7cd889e0bff.jpg) +Figure 5. Comparison of the approach learned on UV planes (POP [39]) and the approach learned on continuous surfaces (Ours). Our solution alleviates the seam artifacts of POP. + +tion mesh, POP [39] on the 2D UV plane) or in a space that is too flexible (e.g., NASA [16] in the implicit 3D space), while our approach learns features on continuous and compact surface space. Though SCALE [36] has also investigated using the point-based encoder (PointNet [52]) for pose-dependent feature extraction, it only uses the global features which lack fine-grained information. In contrast, we adopt PointNet++ [53] (PN++) to learn hierarchical surface features, so that the pose-dependent features can be learned more effectively. Fig. 5 shows the qualitative re + +Table 4. Comparison of the modeling ability of different approaches and their feature learning space on the CAPE dataset. + +
MethodsFeaturesChamfer-L2↓Normal diff. ↓
blazerlongshortlongblazerlongshortlong
CAPE [38]Mesh1.961.371.281.15
NASA [16]3D space1.370.951.291.17
SCALE [36]Global1.461.031.341.16
SCALE [36]UV plane1.070.891.221.12
POP [39]UV plane0.780.571.291.24
CSFSurface0.710.541.151.09
+ +Table 5. Ablation study on the effectiveness of continuous surface features (CSF) and Explicit Template Decomposition (ETD) on a dress outfit (felice-004 from ReSynth [39]). + +
MethodPOPPOP + ETDCSFCSF + ETD
CD7.347.056.536.01
NML1.241.171.161.16
+ +sults of the ablation approaches learned on UV planes and continuous surfaces. We can see that our solution clearly alleviates the seam artifacts of POP. + +Evaluation of Explicit Template Decomposition. The decomposed templates help to capture more accurate pose-dependent deformations and produce more natural wrinkles in unseen poses, especially for outfits that differ largely from the body template. To validate the effectiveness of our decomposition strategy, Tab. 5 reports the ablation experiments on a dress outfit (felice-004) of the ReSynth [39] + +![](images/982d99d42e975eb97fe7747d0b50555014dc1693aef34c30fab0bb6819d7a963.jpg) +Figure 6. Comparison of the clothing deformation in unseen poses. Explicit template decomposition (ETD) helps to capture more natural pose-dependent wrinkle details than POP [39]. + +![](images/b27cfd4580ea5deef04c17561fb87501427671c1e9a0f35c176e6b4d01073783.jpg) +Figure 7. Comparison of the learned templates with FITE [32]. + +dataset. We can see that the proposed Explicit Template Decomposition (ETD) brings clear performance gains over baseline methods. Fig. 6 shows the visual improvements of pose-dependent wrinkles when applying the explicit template decomposition. Note that the templates are decomposed in an end-to-end manner in our method. Compared with the implicit template learned in the recent approach FITE [32], the explicit templates in our method contain more details, as shown in Fig. 7. + +# 5. Conclusions and Future Work + +In this work, we present CloSET, a point-based clothed human modeling method that is built upon a continuous surface and learns to decompose explicit garment templates for better learning of pose-dependent deformations. By learn- + +ing features on a continuous surface, our solution gets rid of the seam artifacts in previous state-of-the-art point-based methods [36, 39]. Moreover, the explicit template decomposition helps to capture more accurate and natural pose-dependent wrinkles. To facilitate the research in this direction, we also introduce a high-quality real-world scan dataset with diverse outfit styles and accurate body model fitting. + +Limitations and Future Work. Due to the incorrect skinning weight used in our template, the issue of the nonuniform point distribution remains for the skirt and dress outfits. Combining our method with recent learnable skinning solutions [37, 59] could alleviate this issue and further improve the results. Currently, our method does not leverage information from adjacent poses. Enforcing temporal consistency and correspondences between adjacent frames would be interesting for future work. Moreover, incorporating physics-based losses into the learning process like SNUG [61] would also be a promising solution to address the artifacts like self-intersections. + +Acknowledgements. This work was supported by the National Key R&D Program of China (2022YFF0902200), the National Natural Science Foundation of China (No.62125107 and No.61827805), and the China Postdoctoral Science Foundation (No.2022M721844). + +# References + +[1] Panos Achlioptas, Olga Diamanti, Ioannis Mitliagkas, and Leonidas Guibas. Learning representations and generative models for 3D point clouds. In ICML, pages 40-49, 2018. 2 +[2] Thiemo Alldieck, Gerard Pons-Moll, Christian Theobalt, and Marcus Magnor. Tex2shape: Detailed full human body geometry from a single image. In ICCV, pages 2293-2303, 2019. 1, 3 +[3] Ziqian Bai, Timur Bagautdinov, Javier Romero, Michael Zollhöfer, Ping Tan, and Shunsuke Saito. AutoAvatar: Autoregressive neural fields for dynamic avatar modeling. ECCV, 2022. 2 +[4] Ilya Baran and Jovan Popovic. Automatic rigging and animation of 3D characters. ACM TOG, 26(3):72-es, 2007. 1, 3 +[5] Jan Bednářík, Shaifali Parashar, Erhan Gundogdu, Mathieu Salzmann, and Pascal Fua. Shape reconstruction by learning differentiable surface representations. In CVPR, pages 4715-4724, 2020. 2 +[6] Hugo Bertiche, Meysam Madadi, Emilio Tylson, and Sergio Escalera. DeePSD: Automatic deep skinning and pose space deformation for 3D garment animation. In ICCV, pages 5471-5480, 2021. 2 +[7] Bharat Lal Bhatnagar, Garvita Tiwari, Christian Theobalt, and Gerard Pons-Moll. Multi-Garment Net: Learning to dress 3D people from images. In ICCV, pages 5420-5430, 2019. 2 +[8] Andrei Burov, Matthias Nießner, and Justus Thies. Dynamic surface function networks for clothed human bodies. In ICCV, pages 10754-10764, October 2021. 1, 2 +[9] Xu Chen, Tianjian Jiang, Jie Song, Jinlong Yang, Michael J Black, Andreas Geiger, and Otmar Hilliges. gDNA: Towards generative detailed neural avatars. In CVPR, pages 20427-20437, 2022. 2 +[10] Xu Chen, Yufeng Zheng, Michael J. Black, Otmar Hilliges, and Andreas Geiger. SNARF: Differentiable forward skinning for animating non-rigid neural implicit shapes. In ICCV, pages 11594–11604, October 2021. 1, 2, 6, 7 +[11] Julian Chibane, Aymen Mir, and Gerard Pons-Moll. Neural unsigned distance fields for implicit function learning. In NeurIPS, pages 21638-21652, 2020. 2 +[12] Enric Corona, Albert Pumarola, Guillem Alenya, Gerard Pons-Moll, and Francesc Moreno-Noguer. SMPLicit: Topology-aware generative model for clothed people. In CVPR, pages 11875-11885, 2021. 1, 2 +[13] Edilson De Aguiar, Leonid Sigal, Adrien Treuille, and Jessica K Hodgins. Stable spaces for real-time clothing. In ACM TOG, volume 29, page 106. ACM, 2010. 2 +[14] Deform Dynamics. https://deformdynamics.com/.1,3 +[15] Boyang Deng, JP Lewis, Timothy Jeruzalski, Gerard Pons-Moll, Geoffrey Hinton, Mohammad Norouzi, and Andrea Tagliasacchi. Neural articulated shape approximation. In ECCV, pages 612-628, 2020. 2 +[16] Boyang Deng, John P Lewis, Timothy Jeruzalski, Gerard Pons-Moll, Geoffrey Hinton, Mohammad Norouzi, and An- + +drea Tagliasacchi. Nasa neural articulated shape approximation. In ECCV, pages 612-628. Springer, 2020. 2, 7 +[17] Zhantao Deng, Jan Bednárík, Mathieu Salzmann, and Pascal Fua. Better patch stitching for parametric surface reconstruction. In 3DV, pages 593-602, 2020. 2 +[18] Theo Deprelle, Thibault Groueix, Matthew Fisher, Vladimir Kim, Bryan Russell, and Mathieu Aubry. Learning elementary structures for 3D shape generation and matching. In NeurIPS, pages 7433-7443, 2019. 2 +[19] Haoqiang Fan, Hao Su, and Leonidas J Guibas. A point set generation network for 3D object reconstruction from a single image. In CVPR, pages 2463-2471, 2017. 2 +[20] Andrew Feng, Dan Casas, and Ari Shapiro. Avatar reshaping and automatic rigging using a deformable model. In Proceedings of the ACM SIGGRAPH Conference on Motion in Games, pages 57-64, 2015. 1, 3 +[21] Amos Gropp, Lior Yariv, Niv Haim, Matan Atzmon, and Yaron Lipman. Implicit geometric regularization for learning shapes. In ICML, pages 3569-3579, 2020. 2 +[22] Thibault Groueix, Matthew Fisher, Vladimir G Kim, Bryan C Russell, and Mathieu Aubry. 3D-CODED: 3D correspondences by deep deformation. In ECCV, pages 230-246, 2018. 2 +[23] Peng Guan, Loretta Reiss, David A Hirshberg, Alexander Weiss, and Michael J Black. DRAPE: DRessing Any PErson. ACM TOG, 31(4):35-1, 2012. 1, 2, 3 +[24] Erhan Gundogdu, Victor Constantin, Amrollah Seifoddini, Minh Dang, Mathieu Salzmann, and Pascal Fua. GarNet: A two-stream network for fast and accurate 3D cloth draping. In CVPR, pages 8739-8748, 2019. 1, 2, 3 +[25] Boyan Jiang, Xinlin Ren, Mingsong Dou, Xiangyang Xue, Yanwei Fu, and Yinda Zhang. LoRD: Local 4d implicit representation for high-fidelity dynamic human modeling. In ECCV, pages 307-326. Springer, 2022. 2 +[26] Boyi Jiang, Juyong Zhang, Yang Hong, Jinhao Luo, Ligang Liu, and Hujun Bao. BCNet: Learning body and cloth shape from a single image. In ECCV, pages 18-35. Springer, 2020. 2 +[27] Hyomin Kim, Hyeonseo Nam, Junkon Kim, Jaesik Park, and Seungyong Lee. LaplacianFusion: Detailed 3D clothed-human body reconstruction. ACM TOG, 41(6):1-14, 2022. 2 +[28] Zorah Lahner, Daniel Cremers, and Tony Tung. Deepwrinkles: Accurate and realistic clothing modeling. In ECCV, pages 667-684, 2018. 2, 3 +[29] Ruilong Li, Julian Tanke, Minh Vo, Michael Zollhöfer, Jürgen Gall, Angjoo Kanazawa, and Christoph Lassner. TAVA: Template-free animatable volumetric actors. In ECCV, pages 419-436. Springer, 2022. 2 +[30] Zhe Li, Zerong Zheng, Hongwen Zhang, Chaonan Ji, and Yebin Liu. Avatarcap: Animatable avatar conditioned monocular human volumetric capture. In ECCV, pages 322-341. Springer, 2022. 2, 5 +[31] Chen-Hsuan Lin, Chen Kong, and Simon Lucey. Learning efficient point cloud generation for dense 3D object reconstruction. In AAAI, pages 7114-7121, 2018. 2 + +[32] Siyou Lin, Hongwen Zhang, Zerong Zheng, Ruizhi Shao, and Yebin Liu. Learning implicit templates for point-based clothed human modeling. ECCV, 2022. 2, 3, 6, 8 +[33] Lijuan Liu, Youyi Zheng, Di Tang, Yi Yuan, Changjie Fan, and Kun Zhou. NeuroSkinning: Automatic skin binding for production characters with deep graph networks. ACM TOG, 38(4):1-12, 2019. 1, 3 +[34] Matthew Loper, Naureen Mahmood, Javier Romero, Gerard Pons-Moll, and Michael J Black. SMPL: A skinned multiperson linear model. ACM TOG, 34(6):248, 2015. 2, 3 +[35] William E Lorensen and Harvey E Cline. Marching cubes: A high resolution 3D surface construction algorithm. In SIGGRAPH, volume 21, pages 163-169, 1987. 6 +[36] Qianli Ma, Shunsuke Saito, Jinlong Yang, Siyu Tang, and Michael J Black. SCALE: Modeling clothed humans with a surface codec of articulated local elements. In CVPR, pages 16082-16093, 2021. 1, 2, 3, 4, 5, 6, 7, 8 +[37] Qianli Ma, Jinlong Yang, Michael J Black, and Siyu Tang. Neural point-based shape modeling of humans in challenging clothing. 3DV, 2022. 3, 6, 8 +[38] Qianli Ma, Jinlong Yang, Anurag Ranjan, Sergi Pujades, Gerard Pons-Moll, Siyu Tang, and Michael J Black. Learning to dress 3D people in generative clothing. In CVPR, pages 6469-6478, 2020. 1, 2, 3, 5, 6, 7 +[39] Qianli Ma, Jinlong Yang, Siyu Tang, and Michael J Black. The power of points for modeling humans in clothing. In ICCV, pages 10974-10984, 2021. 1, 2, 3, 4, 5, 6, 7, 8 +[40] Xu Ma, Can Qin, Haoxuan You, Haoxi Ran, and Yun Fu. Rethinking network design and local geometry in point cloud: A simple residual mlp framework. In ICLR, 2022. 5 +[41] Lars Mescheder, Michael Oechsle, Michael Niemeyer, Sebastian Nowozin, and Andreas Geiger. Occupancy networks: Learning 3D reconstruction in function space. In CVPR, pages 4460-4470, 2019. 2 +[42] Marko Mihajlovic, Yan Zhang, Michael J Black, and Siyu Tang. LEAP: Learning articulated occupancy of people. In CVPR, pages 10461-10471, 2021. 2, 4 +[43] Alexandros Neophytou and Adrian Hilton. A layered model of human body and garment deformation. In 3DV, pages 171-178, 2014. 2 +[44] Hayato Onizuka, Zehra Hayirci, Diego Thomas, Akihiro Sugimoto, Hideaki Uchiyama, and Rin-ichiro Taniguchi. TetraTSDF: 3D human reconstruction from a single image with a tetrahedral outer shell. In CVPR, pages 6011–6020, 2020. 2 +[45] Pablo Palafox, Aljaž Božić, Justus Thies, Matthias Nießner, and Angela Dai. NPMs: Neural parametric models for 3D deformable shapes. In ICCV, pages 12695-12705, October 2021. 2 +[46] Pablo Palafox, Nikolaos Sarafianos, Tony Tung, and Angela Dai. SPAMs: Structured implicit parametric models. In CVPR, pages 12851-12860, 2022. 2 +[47] Junyi Pan, Xiaoguang Han, Weikai Chen, Jiapeng Tang, and Kui Jia. Deep mesh reconstruction from single RGB images via topology modification networks. In ICCV, pages 9963-9972, 2019. 2 + +[48] Jeong Joon Park, Peter Florence, Julian Straub, Richard Newcombe, and Steven Lovegrove. DeepSDF: Learning continuous signed distance functions for shape representation. In CVPR, pages 165-174, 2019. 2 +[49] Chaitanya Patel, Zhouyingcheng Liao, and Gerard Pons-Moll. TailorNet: Predicting clothing in 3D as a function of human pose, shape and garment style. In CVPR, pages 7363-7373, 2020. 1, 2, 3 +[50] Georgios Pavlakos, Vasileios Choutas, Nima Ghorbani, Timo Bolkart, Ahmed AA Osman, Dimitrios Tzionas, and Michael J Black. Expressive body capture: 3D hands, face, and body from a single image. In CVPR, pages 10975-10985, 2019. 3, 5 +[51] Gerard Pons-Moll, Sergi Pujades, Sonny Hu, and Michael Black. ClothCap: Seamless 4d clothing capture and retargeting. ACM TOG, 36(4), 2017. Two first authors contributed equally. 3 +[52] Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. PointNet: Deep learning on point sets for 3D classification and segmentation. In CVPR, pages 652-660, 2017. 7 +[53] Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J Guibas. Pointnet++: Deep hierarchical feature learning on point sets in a metric space. NeurIPS, 30, 2017. 3, 5, 7 +[54] Guocheng Qian, Yuchen Li, Houwen Peng, Jinjie Mai, Hasan Hammoud, Mohamed Elhoseiny, and Bernard Ghanem. Pointnext: Revisiting pointnet++ with improved training and scaling strategies. NeurIPS, 35:23192-23204, 2022. 5 +[55] Shenhan Qian, Jiale Xu, Ziwei Liu, Liqian Ma, and Shenghua Gao. UNIF: United neural implicit functions for clothed human reconstruction and animation. In ECCV, 2022. 2 +[56] Renderpeople, 2020. https://renderpeople.com.3 +[57] Shunsuke Saito, Zeng Huang, Ryota Natsume, Shigeo Morishima, Angjoo Kanazawa, and Hao Li. PIFu: Pixel-aligned implicit function for high-resolution clothed human digitization. In ICCV, pages 2304-2314, 2019. 2, 4 +[58] Shunsuke Saito, Tomas Simon, Jason Saragih, and Hanbyul Joo. PIFuHD: Multi-level pixel-aligned implicit function for high-resolution 3D human digitization. In CVPR, pages 84–93, 2020. 2 +[59] Shunsuke Saito, Jinlong Yang, Qianli Ma, and Michael J Black. SCANimate: Weakly supervised learning of skinned clothed avatar networks. In CVPR, pages 2886-2897, 2021. 1, 2, 3, 4, 6, 7, 8 +[60] Igor Santesteban, Miguel A. Otaduy, and Dan Casas. Learning-Based Animation of Clothing for Virtual Try-On. Comput. Graph. Forum, 38(2):355-366, 2019. 2 +[61] Igor Santesteban, Miguel A Otaduy, and Dan Casas. SNUG: Self-supervised neural dynamic garments. In CVPR, pages 8140-8150, 2022. 8 +[62] Igor Santesteban, Nils Thuerey, Miguel A Otaduy, and Dan Casas. Self-supervised collision handling via generative 3D garment models for virtual try-on. In CVPR, pages 11763-11773, 2021. 2 +[63] Yu Shen, Junbang Liang, and Ming C Lin. Gan-based garment generation using sewing pattern images. In ECCV, volume 1, page 3, 2020. 2 + +[64] Garvita Tiwari, Bharat Lal Bhatnagar, Tony Tung, and Gerard Pons-Moll. SIZER: A dataset and model for parsing 3D clothing and learning size sensitive 3D clothing. In ECCV, volume 12348, pages 1-18, 2020. 2 +[65] Garvita Tiwari, Nikolaos Sarafianos, Tony Tung, and Gerard Pons-Moll. Neural-GIF: Neural generalized implicit functions for animating people in clothing. In ICCV, pages 11708-11718, 2021. 2 +[66] Raquel Vidaurre, Igor Santesteban, Elena Garces, and Dan Casas. Fully convolutional graph neural networks for parametric virtual try-on. In Comput. Graph. Forum, volume 39, pages 145-156. Wiley Online Library, 2020. 2 +[67] Shaofei Wang, Marko Mihajlovic, Qianli Ma, Andreas Geiger, and Siyu Tang. Meta­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­. +clothed human models from few depth images. In NeurIPS, 2021. 2, 3 +[68] Jane Wu, Zhenglin Geng, Hui Zhou, and Ronald Fedkiw. Skinning a parameterization of three-dimensional space for neural network cloth. arXiv preprint arXiv:2006.04874, 2020. 2 +[69] Donglai Xiang, Fabian Prada, Timur Bagautdinov, Weipeng Xu, Yuan Dong, He Wen, Jessica Hodgins, and Chenglei Wu. Modeling clothing as a separate layer for an animatable human avatar. ACM TOG, dec 2021. 2 +[70] Yuliang Xiu, Jinlong Yang, Dimitrios Tzionas, and Michael J Black. ICON: implicit clothed humans obtained from normals. In CVPR, pages 13286-13296. IEEE, 2022. 2 +[71] Jinlong Yang, Jean-Sebastien Franco, Franck Hetroy-Wheeler, and Stefanie Wuhrer. Analyzing clothing layer deformation statistics of 3D human motions. In ECCV, September 2018. 3 +[72] Shan Yang, Zherong Pan, Tanya Amert, Ke Wang, Licheng Yu, Tamara Berg, and Ming C Lin. Physics-inspired garment recovery from a single-view image. ACM TOG, 37(5):1-14, 2018. 2 +[73] Tao Yu, Zerong Zheng, Kaiwen Guo, Pengpeng Liu, Qionghai Dai, and Yebin Liu. Function4D: Real-time human volumetric capture from very sparse consumer rgbd sensors. In CVPR, June 2021. 3 +[74] Ilya Zakharkin, Kirill Mazur, Artur Grigorev, and Victor Lempitsky. Point-based modeling of human clothing. In ICCV, pages 14718-14727, October 2021. 2 +[75] Zerong Zheng, Han Huang, Tao Yu, Hongwen Zhang, Yandong Guo, and Yebin Liu. Structured local radiance fields for human avatar modeling. In CVPR, pages 15893-15903, 2022. 2 +[76] Zerong Zheng, Tao Yu, Qionghai Dai, and Yebin Liu. Deep implicit templates for 3D shape representation. In CVPR, pages 1429-1439, 2021. 3, 5 +[77] Zerong Zheng, Tao Yu, Yebin Liu, and Qionghai Dai. PaMIR: Parametric model-conditioned implicit representation for image-based human reconstruction. IEEE TPAMI, 2021. 2 +[78] Zerong Zheng, Tao Yu, Yixuan Wei, Qionghai Dai, and Yebin Liu. DeepHuman: 3D human reconstruction from a single image. In ICCV, pages 7739-7749, 2019. 3 + +[79] Heming Zhu, Yu Cao, Hang Jin, Weikai Chen, Dong Du, Zhangye Wang, Shuguang Cui, and Xiaoguang Han. Deep Fashion3D: A dataset and benchmark for 3D garment reconstruction from single images. In ECCV, volume 12346, pages 512-530, 2020. 2 \ No newline at end of file diff --git a/2023/CloSET_ Modeling Clothed Humans on Continuous Surface With Explicit Template Decomposition/images.zip b/2023/CloSET_ Modeling Clothed Humans on Continuous Surface With Explicit Template Decomposition/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..d6f13a58beb8d4500d3973172de33f3096b3711e --- /dev/null +++ b/2023/CloSET_ Modeling Clothed Humans on Continuous Surface With Explicit Template Decomposition/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f1e396185da57ca340849f13e0096a4bcbf9c747a6662ad3b18bf7513fb51644 +size 686140 diff --git a/2023/CloSET_ Modeling Clothed Humans on Continuous Surface With Explicit Template Decomposition/layout.json b/2023/CloSET_ Modeling Clothed Humans on Continuous Surface With Explicit Template Decomposition/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..9cfbab2772d33cb86e86a856670118ef5f1eb99d --- /dev/null +++ b/2023/CloSET_ Modeling Clothed Humans on Continuous Surface With Explicit Template Decomposition/layout.json @@ -0,0 +1,8589 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 97, + 102, + 500, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 97, + 102, + 500, + 140 + ], + "spans": [ + { + "bbox": [ + 97, + 102, + 500, + 140 + ], + "type": "text", + "content": "CloSET: Modeling Clothed Humans on Continuous Surface with Explicit Template Decomposition" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 113, + 160, + 480, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 113, + 160, + 480, + 204 + ], + "spans": [ + { + "bbox": [ + 113, + 160, + 480, + 204 + ], + "type": "text", + "content": "Hongwen Zhang" + }, + { + "bbox": [ + 113, + 160, + 480, + 204 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 113, + 160, + 480, + 204 + ], + "type": "text", + "content": " Siyou Lin" + }, + { + "bbox": [ + 113, + 160, + 480, + 204 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 113, + 160, + 480, + 204 + ], + "type": "text", + "content": " Ruizhi Shao" + }, + { + "bbox": [ + 113, + 160, + 480, + 204 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 113, + 160, + 480, + 204 + ], + "type": "text", + "content": " Yuxiang Zhang" + }, + { + "bbox": [ + 113, + 160, + 480, + 204 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 113, + 160, + 480, + 204 + ], + "type": "text", + "content": " Zerong Zheng" + }, + { + "bbox": [ + 113, + 160, + 480, + 204 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 113, + 160, + 480, + 204 + ], + "type": "text", + "content": " Han Huang" + }, + { + "bbox": [ + 113, + 160, + 480, + 204 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 113, + 160, + 480, + 204 + ], + "type": "text", + "content": " Yandong Guo" + }, + { + "bbox": [ + 113, + 160, + 480, + 204 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 113, + 160, + 480, + 204 + ], + "type": "text", + "content": " Yebin Liu" + }, + { + "bbox": [ + 113, + 160, + 480, + 204 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 113, + 160, + 480, + 204 + ], + "type": "text", + "content": " Tsinghua University " + }, + { + "bbox": [ + 113, + 160, + 480, + 204 + ], + "type": "inline_equation", + "content": "{}^{2}" + }, + { + "bbox": [ + 113, + 160, + 480, + 204 + ], + "type": "text", + "content": "OPPO Research Institute" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 143, + 231, + 192, + 243 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 231, + 192, + 243 + ], + "spans": [ + { + "bbox": [ + 143, + 231, + 192, + 243 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 258, + 290, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 258, + 290, + 521 + ], + "spans": [ + { + "bbox": [ + 46, + 258, + 290, + 521 + ], + "type": "text", + "content": "Creating animatable avatars from static scans requires the modeling of clothing deformations in different poses. Existing learning-based methods typically add pose-dependent deformations upon a minimally-clothed mesh template or a learned implicit template, which have limitations in capturing details or hinder end-to-end learning. In this paper, we revisit point-based solutions and propose to decompose explicit garment-related templates and then add pose-dependent wrinkles to them. In this way, the clothing deformations are disentangled such that the pose-dependent wrinkles can be better learned and applied to unseen poses. Additionally, to tackle the seam artifact issues in recent state-of-the-art point-based methods, we propose to learn point features on a body surface, which establishes a continuous and compact feature space to capture the fine-grained and pose-dependent clothing geometry. To facilitate the research in this field, we also introduce a high-quality scan dataset of humans in real-world clothing. Our approach is validated on two existing datasets and our newly introduced dataset, showing better clothing deformation results in unseen poses. The project page with code and dataset can be found at https://www.liuyebin.com/closet." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 548, + 128, + 561 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 548, + 128, + 561 + ], + "spans": [ + { + "bbox": [ + 47, + 548, + 128, + 561 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 570, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 570, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 570, + 288, + 715 + ], + "type": "text", + "content": "Animating 3D clothed humans requires the modeling of pose-dependent deformations in various poses. The diversity of clothing styles and body poses makes this task extremely challenging. Traditional methods are based on either simple rigging and skinning [4,20,33] or physics-based simulation [14, 23, 24, 49], which heavily rely on artist efforts or computational resources. Recent learning-based methods [10, 36, 39, 59] resort to modeling the clothing deformation directly from raw scans of clothed humans. Despite the promising progress, this task is still far from being solved due to the challenges in clothing representations, generalization to unseen poses, and data acquisition, etc." + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 310, + 230, + 547, + 407 + ], + "blocks": [ + { + "bbox": [ + 310, + 230, + 547, + 407 + ], + "lines": [ + { + "bbox": [ + 310, + 230, + 547, + 407 + ], + "spans": [ + { + "bbox": [ + 310, + 230, + 547, + 407 + ], + "type": "image", + "image_path": "9120cdfd2b4793367ccd1702c81f777d90bb4f9e6c091172060a3d4e1a02a13b.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 415, + 547, + 438 + ], + "lines": [ + { + "bbox": [ + 305, + 415, + 547, + 438 + ], + "spans": [ + { + "bbox": [ + 305, + 415, + 547, + 438 + ], + "type": "text", + "content": "Figure 1. Our method learns to decompose garment templates (top row) and add pose-dependent wrinkles upon them (bottom row)." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 450, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 450, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 450, + 547, + 715 + ], + "type": "text", + "content": "For the modeling of pose-dependent garment geometry, the representation of clothing plays a vital role in a learning-based scheme. As the relationship between body poses and clothing deformations is complex, an effective representation is desirable for neural networks to capture pose-dependent deformations. In the research of this line, meshes [2, 12, 38], implicit fields [10, 59], and point clouds [36, 39] have been adopted to represent clothing. In accordance with the chosen representation, the clothing deformation and geometry features are learned on top of a fixed-resolution template mesh [8, 38], a 3D implicit sampling space [10, 59], or an unfolded UV plane [2, 12, 36, 39]. Among these representations, the mesh is the most efficient one but is limited to a fixed topology due to its discretization scheme. The implicit fields naturally enable continuous feature learning in a resolution-free manner but are too flexible to satisfy the body structure prior, leading to geometry artifacts in unseen poses. The point clouds enjoy the compact nature and topology flexibility and have shown promising results in the recent state-of-the-art solutions [36, 39] to represent clothing, but the feature learning on UV planes still leads to discontinuity artifacts between body parts." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "text", + "content": "501" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 263 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 263 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 263 + ], + "type": "text", + "content": "To model the pose-dependent deformation of clothing, body templates such as SMPL [34] are typically leveraged to account for articulated motions. However, a body template alone is not ideal, since the body template only models the minimally-clothed humans and may hinder the learning of actual pose-dependent deformations, especially in cases of loose clothing. To overcome this issue, recent implicit approaches [59] make attempts to learn skinning weights in the 3D space to complement the imperfect body templates. However, their pose-dependent deformations are typically coarse due to the difficulty in learning implicit fields. For explicit solutions, the recent approach [32] suggests learning coarse templates implicitly at first and then the pose-dependent deformations explicitly. Despite its effectiveness, such a workaround requires a two-step modeling procedure and hinders end-to-end learning." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 262, + 289, + 670 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 262, + 289, + 670 + ], + "spans": [ + { + "bbox": [ + 46, + 262, + 289, + 670 + ], + "type": "text", + "content": "In this work, we propose CloSET, an end-to-end method to tackle the above issues by modeling Clothed humans on a continuous Surface with Explicit Template decomposition. We follow the spirit of recent state-of-the-art point-based approaches [32, 36, 39] as they show the efficiency and potential in modeling real-world garments. We take steps forward in the following aspects for better point-based modeling of clothed humans. First, we propose to decompose the clothing deformations into explicit garment templates and pose-dependent wrinkles. Specifically, our method learns a garment-related template and adds the pose-dependent displacement upon them, as shown in Fig. 1. Such a garment-related template preserves a shared topology for various poses and enables better learning of pose-dependent wrinkles. Different from the recent solution [32] that needs two-step procedures, our method can decompose the explicit templates in an end-to-end manner with more garment details. Second, we tackle the seam artifact issues that occurred in recent point-based methods [36, 39]. Instead of using unfolded UV planes, we propose to learn point features on a body surface, which supports a continuous and compact feature space. We achieve this by learning hierarchical point-based features on top of the body surface and then using barycentric interpolation to sample features continuously. Compared to feature learning in the UV space [39], on template meshes [8, 38], or in the 3D implicit space [57, 59], our body surface enables the network to capture not only fine-grained details but also long-range part correlations for pose-dependent geometry modeling. Third, we introduce a new scan dataset of humans in real-world clothing, which contains more than 2,000 high-quality scans of humans in diverse outfits, hoping to facilitate the research in this field. The main contributions of this work are summarized below:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 677, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 677, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 58, + 677, + 288, + 715 + ], + "type": "text", + "content": "- We propose a point-based clothed human modeling method by decomposing clothing deformations into explicit garment templates and pose-dependent wrin" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 324, + 72, + 545, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 324, + 72, + 545, + 120 + ], + "spans": [ + { + "bbox": [ + 324, + 72, + 545, + 120 + ], + "type": "text", + "content": "kles in an end-to-end manner. These learnable templates provide a garment-aware canonical space so that pose-dependent deformations can be better learned and applied to unseen poses." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 317, + 127, + 547, + 241 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 317, + 127, + 545, + 187 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 127, + 545, + 187 + ], + "spans": [ + { + "bbox": [ + 317, + 127, + 545, + 187 + ], + "type": "text", + "content": "- We propose to learn point-based clothing features on a continuous body surface, which allows a continuous feature space for fine-grained detail modeling and helps to capture long-range part correlations for pose-dependent geometry modeling." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 317, + 194, + 547, + 241 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 194, + 547, + 241 + ], + "spans": [ + { + "bbox": [ + 317, + 194, + 547, + 241 + ], + "type": "text", + "content": "- We introduce a new high-quality scan dataset of clothed humans in real-world clothing to facilitate the research of clothed human modeling and animation from real-world scans." + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 251, + 392, + 264 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 251, + 392, + 264 + ], + "spans": [ + { + "bbox": [ + 306, + 251, + 392, + 264 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 270, + 545, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 270, + 545, + 319 + ], + "spans": [ + { + "bbox": [ + 305, + 270, + 545, + 319 + ], + "type": "text", + "content": "Representations for Modeling Clothed Humans. A key component in modeling clothed humans is the choice of representation, which mainly falls into two categories: implicit and explicit representations." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 319, + 546, + 461 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 319, + 546, + 461 + ], + "spans": [ + { + "bbox": [ + 305, + 319, + 546, + 461 + ], + "type": "text", + "content": "Implicit Modeling. Implicit methods [3, 10-12, 15, 21, 29, 41, 42, 48, 59, 67, 70, 77] represent surfaces as the level set of an implicit neural scalar field. Recent state-of-the-art methods typically learn the clothing deformation field with a canonical space decomposition [9, 30, 45, 59, 65] or part-based modeling strategies [16, 25, 46, 55, 75]. Compared to mesh templates, implicit surfaces are not topologically constrained to specific templates [57, 58], and can model various clothes with complex topology. However, the learning space of an implicit surface is the whole 3D volume, which makes training and interpolation difficult, especially when the numbers of scan data are limited." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 462, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 462, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 305, + 462, + 547, + 714 + ], + "type": "text", + "content": "Explicit Modeling. Mesh surfaces, the classic explicit representation, currently dominate the field of 3D modeling [6-8, 13, 23, 24, 26-28, 38, 43, 49, 60, 62, 64, 66, 69, 72] with their compactness and high efficiency in downstream tasks such as rendering, but they are mostly limited to a fixed topology and/or require scan data registered to a template. Thus, mesh-based representations forbid the learning of a universal model for topologically varying clothing types. Though some approaches have been proposed to allow varying mesh topology [44, 47, 63, 68, 79], they are still limited in their expressiveness. Point clouds enjoy both compactness and topological flexibility. Previous work generates sparse point clouds for 3D representation [1, 19, 31, 74]. However, the points need to be densely sampled over the surface to model surface geometry accurately. Due to the difficulty of generating a large point set, recent methods group points into patches [5, 17, 18, 22]. Each patch maps the 2D UV space to the 3D space, allowing arbitrarily dense sampling within this patch. SCALE [36] successfully applies this idea to modeling clothed humans, but produces notable discontinuity artifacts near patch boundaries." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "text", + "content": "502" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 217 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 217 + ], + "type": "text", + "content": "POP [39] further utilizes a single fine-grained UV map for the whole body surface, leading to a more topologically flexible representation. However, the discontinuity of the UV map may lead to seam artifacts in POP. Very recently, FITE [32] suggests learning implicit coarse templates [76] at first and then explicit fine details. Despite the efficacy, it requires a two-step modeling procedure. Concurrently, SkiRT [37] proposes to improve the body template by learning the blend skinning weights with several data terms. In contrast, our method applies regularization to achieve template decomposition in an end-to-end manner and learns the point-based pose-dependent displacement more effectively." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 243, + 287, + 338 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 243, + 287, + 338 + ], + "spans": [ + { + "bbox": [ + 46, + 243, + 287, + 338 + ], + "type": "text", + "content": "Pose-dependent Deformations for Animation. In the field of character animation, traditional methods utilize rigging and skinning techniques to repose characters [4, 20, 33, 34, 50]. but they fail to model realistic pose-dependent clothing deformations such as wrinkles and sliding motions between clothes and body. We conclude two key ingredients in modeling pose-dependent clothing: (i) pose-dependent feature learning; (ii) datasets with realistic clothing." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 340, + 289, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 340, + 289, + 592 + ], + "spans": [ + { + "bbox": [ + 46, + 340, + 289, + 592 + ], + "type": "text", + "content": "Pose-dependent Feature Learning. Some traditional methods directly incorporate the entire pose parameters into the model [28, 38, 49, 71]. Such methods easily overfit on pose parameters and introduce spurious correlations, causing bad generalization to unseen poses. Recent work explores poses conditioning with local features, either with point clouds [36, 39] or implicit surfaces [59, 67], and shows superiority in improving geometry quality and in eliminating spurious correlations. Among them, the most relevant to ours is POP [39], which extracts local pose features by utilizing convolution on a UV position map. Despite its compelling performance, POP suffers from artifacts inherent to its UV-based representation, hence the convolution on the UV map produces discontinuity near UV islands' boundaries [39]. We address this issue by discarding the UV-based scheme and returning to the actual 3D body surface. We attach the features to a uniform set of points on a T-posed body template, and process them via a PointNet++ [53] structure for pose-dependent modeling. As validated by our experiments, our pose embedding method leads to both qualitative and quantitative improvements." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 594, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 289, + 715 + ], + "type": "text", + "content": "Clothed Human Datasets. Another challenging issue for training an animatable avatar is the need for datasets of clothed humans in diverse poses. There are considerable efforts seeking to synthesize clothed datasets with physics-based simulation [14,23,24,39,49]. Although they are diverse in poses, there remains an observable domain gap between synthetic clothes and real data. Acquiring clothed scans with realistic details of clothing deformations [38, 51, 56, 73, 78] is crucial for the development of learning-based methods in this field." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 305, + 71, + 362, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 71, + 362, + 84 + ], + "spans": [ + { + "bbox": [ + 305, + 71, + 362, + 84 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 91, + 547, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 91, + 547, + 163 + ], + "spans": [ + { + "bbox": [ + 304, + 91, + 547, + 163 + ], + "type": "text", + "content": "As illustrated in Fig. 2, the proposed method CloSET learns garment-related and pose-dependent features on body surfaces (see Sec. 3.1), which can be sampled in a continuous manner and fed into two decoders for the generation of explicit garment templates and pose-dependent wrinkles (see Sec. 3.2)." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 305, + 168, + 466, + 181 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 168, + 466, + 181 + ], + "spans": [ + { + "bbox": [ + 305, + 168, + 466, + 181 + ], + "type": "text", + "content": "3.1. Continuous Surface Features" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 187, + 547, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 187, + 547, + 319 + ], + "spans": [ + { + "bbox": [ + 304, + 187, + 547, + 319 + ], + "type": "text", + "content": "As most parts of the clothing are deformed smoothly in different poses, a continuous feature space is desirable to model the garment details and pose-dependent garment geometry. To this end, our approach first learns features on top of a body template surface, i.e., a SMPL [34] or SMPL-X [50] model in a T-pose. Note that these features are not limited to those on template vertices as they can be continuously sampled from the body surface via barycentric interpolation. Hence, our feature space is more continuous than UV-based [2, 39] spaces, while being more compact than 3D implicit feature fields [59, 67]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 319, + 547, + 463 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 319, + 547, + 463 + ], + "spans": [ + { + "bbox": [ + 304, + 319, + 547, + 463 + ], + "type": "text", + "content": "To model pose-dependent clothing deformations, the underlying unclothed body model is taken as input to the geometry feature encoder. For each scan, let " + }, + { + "bbox": [ + 304, + 319, + 547, + 463 + ], + "type": "inline_equation", + "content": "\\mathbf{V}^u = \\{\\pmb{v}_n^u\\}_{n=1}^N" + }, + { + "bbox": [ + 304, + 319, + 547, + 463 + ], + "type": "text", + "content": " denote the posed vertex positions of the fitted unclothed body model, where " + }, + { + "bbox": [ + 304, + 319, + 547, + 463 + ], + "type": "inline_equation", + "content": "N = 6890" + }, + { + "bbox": [ + 304, + 319, + 547, + 463 + ], + "type": "text", + "content": " for SMPL [34] and " + }, + { + "bbox": [ + 304, + 319, + 547, + 463 + ], + "type": "inline_equation", + "content": "N = 10475" + }, + { + "bbox": [ + 304, + 319, + 547, + 463 + ], + "type": "text", + "content": " for SMPL-X [50]. These posed vertices act as the pose code and will be paired with the template vertices " + }, + { + "bbox": [ + 304, + 319, + 547, + 463 + ], + "type": "inline_equation", + "content": "\\mathbf{V}^t = \\{\\pmb{v}_n^t\\}_{n=1}^N" + }, + { + "bbox": [ + 304, + 319, + 547, + 463 + ], + "type": "text", + "content": " of the body model in a T-pose, which shares the same mesh topology with " + }, + { + "bbox": [ + 304, + 319, + 547, + 463 + ], + "type": "inline_equation", + "content": "\\mathbf{V}^u" + }, + { + "bbox": [ + 304, + 319, + 547, + 463 + ], + "type": "text", + "content": ". These point pairs are processed by the pose encoder " + }, + { + "bbox": [ + 304, + 319, + 547, + 463 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_p" + }, + { + "bbox": [ + 304, + 319, + 547, + 463 + ], + "type": "text", + "content": " to generate the pose-dependent geometry features " + }, + { + "bbox": [ + 304, + 319, + 547, + 463 + ], + "type": "inline_equation", + "content": "\\{\\phi_p(\\pmb{v}_n^t) \\in \\mathbb{R}^{C_p}\\}_{n=1}^N" + }, + { + "bbox": [ + 304, + 319, + 547, + 463 + ], + "type": "text", + "content": " at vertices " + }, + { + "bbox": [ + 304, + 319, + 547, + 463 + ], + "type": "inline_equation", + "content": "\\mathbf{V}^t" + }, + { + "bbox": [ + 304, + 319, + 547, + 463 + ], + "type": "text", + "content": ", i.e.," + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 348, + 466, + 545, + 482 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 348, + 466, + 545, + 482 + ], + "spans": [ + { + "bbox": [ + 348, + 466, + 545, + 482 + ], + "type": "interline_equation", + "content": "\\left\\{\\phi_ {p} \\left(\\boldsymbol {v} _ {n} ^ {t}\\right) \\in \\mathbb {R} ^ {C _ {p}} \\right\\} _ {n = 1} ^ {N} = \\mathcal {F} _ {p} \\left(\\mathbf {V} ^ {t}, \\mathbf {V} ^ {u}\\right). \\tag {1}", + "image_path": "be6c8b6ca3b1328e14ae3ccb2b0882b2db6ec6c84682499898a439857273163a.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 486, + 547, + 642 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 486, + 547, + 642 + ], + "spans": [ + { + "bbox": [ + 304, + 486, + 547, + 642 + ], + "type": "text", + "content": "To learn hierarchical features with different levels of receptive fields, we adopt PointNet++ [53] as the architecture of the pose encoder " + }, + { + "bbox": [ + 304, + 486, + 547, + 642 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_p" + }, + { + "bbox": [ + 304, + 486, + 547, + 642 + ], + "type": "text", + "content": ", where vertices " + }, + { + "bbox": [ + 304, + 486, + 547, + 642 + ], + "type": "inline_equation", + "content": "\\mathbf{V}^t" + }, + { + "bbox": [ + 304, + 486, + 547, + 642 + ], + "type": "text", + "content": " are treated as the input point cloud in the PointNet++ network, while vertices " + }, + { + "bbox": [ + 304, + 486, + 547, + 642 + ], + "type": "inline_equation", + "content": "\\mathbf{V}^u" + }, + { + "bbox": [ + 304, + 486, + 547, + 642 + ], + "type": "text", + "content": " act as the feature of " + }, + { + "bbox": [ + 304, + 486, + 547, + 642 + ], + "type": "inline_equation", + "content": "\\mathbf{V}^t" + }, + { + "bbox": [ + 304, + 486, + 547, + 642 + ], + "type": "text", + "content": ". As the template vertices " + }, + { + "bbox": [ + 304, + 486, + 547, + 642 + ], + "type": "inline_equation", + "content": "\\mathbf{V}^t" + }, + { + "bbox": [ + 304, + 486, + 547, + 642 + ], + "type": "text", + "content": " are constant, the encoder " + }, + { + "bbox": [ + 304, + 486, + 547, + 642 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_p" + }, + { + "bbox": [ + 304, + 486, + 547, + 642 + ], + "type": "text", + "content": " can focus on the feature learned from the posed vertices " + }, + { + "bbox": [ + 304, + 486, + 547, + 642 + ], + "type": "inline_equation", + "content": "\\mathbf{V}^u" + }, + { + "bbox": [ + 304, + 486, + 547, + 642 + ], + "type": "text", + "content": ". Moreover, the PointNet++ based " + }, + { + "bbox": [ + 304, + 486, + 547, + 642 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_p" + }, + { + "bbox": [ + 304, + 486, + 547, + 642 + ], + "type": "text", + "content": " first abstracts features from the template vertices " + }, + { + "bbox": [ + 304, + 486, + 547, + 642 + ], + "type": "inline_equation", + "content": "\\mathbf{V}^t" + }, + { + "bbox": [ + 304, + 486, + 547, + 642 + ], + "type": "text", + "content": " to sparser points " + }, + { + "bbox": [ + 304, + 486, + 547, + 642 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{V}_l^t\\}_{l=1}^L" + }, + { + "bbox": [ + 304, + 486, + 547, + 642 + ], + "type": "text", + "content": " at " + }, + { + "bbox": [ + 304, + 486, + 547, + 642 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 304, + 486, + 547, + 642 + ], + "type": "text", + "content": " levels, where the number of " + }, + { + "bbox": [ + 304, + 486, + 547, + 642 + ], + "type": "inline_equation", + "content": "\\mathbf{V}_l^t" + }, + { + "bbox": [ + 304, + 486, + 547, + 642 + ], + "type": "text", + "content": " decreases with " + }, + { + "bbox": [ + 304, + 486, + 547, + 642 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 304, + 486, + 547, + 642 + ], + "type": "text", + "content": " increasing. Then, the features at " + }, + { + "bbox": [ + 304, + 486, + 547, + 642 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{V}_l^t\\}_{l=1}^L" + }, + { + "bbox": [ + 304, + 486, + 547, + 642 + ], + "type": "text", + "content": " are further propagated back to " + }, + { + "bbox": [ + 304, + 486, + 547, + 642 + ], + "type": "inline_equation", + "content": "\\mathbf{V}^t" + }, + { + "bbox": [ + 304, + 486, + 547, + 642 + ], + "type": "text", + "content": " successively. In this way, the encoder can capture the long-range part correlations of the pose-dependent deformations." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 642, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 642, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 642, + 547, + 715 + ], + "type": "text", + "content": "Similar to POP [39], our method can be trained under multi-outfit or outfit-specific settings. When trained with multiple outfits, the pose-dependent deformation should be aware of the outfit type, and hence requires the input of the garment features. Specifically, the garment-related features " + }, + { + "bbox": [ + 304, + 642, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\{\\phi_g(\\pmb{v}_n^t) \\in \\mathbb{R}^{C_g}\\}_{n=1}^N" + }, + { + "bbox": [ + 304, + 642, + 547, + 715 + ], + "type": "text", + "content": " are also defined on template" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "text", + "content": "503" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 86, + 70, + 504, + 228 + ], + "blocks": [ + { + "bbox": [ + 86, + 70, + 504, + 228 + ], + "lines": [ + { + "bbox": [ + 86, + 70, + 504, + 228 + ], + "spans": [ + { + "bbox": [ + 86, + 70, + 504, + 228 + ], + "type": "image", + "image_path": "986709cebc27f14be506c0ee7a592d4621578e2cdc4a6679da0a796dbf422979.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 232, + 547, + 278 + ], + "lines": [ + { + "bbox": [ + 46, + 232, + 547, + 278 + ], + "spans": [ + { + "bbox": [ + 46, + 232, + 547, + 278 + ], + "type": "text", + "content": "Figure 2. Overview of the proposed method CloSET. Given an input body model, its pose code and garment code are processed hierarchically by point-based pose and garment encoders " + }, + { + "bbox": [ + 46, + 232, + 547, + 278 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_p" + }, + { + "bbox": [ + 46, + 232, + 547, + 278 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 232, + 547, + 278 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_g" + }, + { + "bbox": [ + 46, + 232, + 547, + 278 + ], + "type": "text", + "content": " for the learning of surface features " + }, + { + "bbox": [ + 46, + 232, + 547, + 278 + ], + "type": "inline_equation", + "content": "\\phi_p" + }, + { + "bbox": [ + 46, + 232, + 547, + 278 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 232, + 547, + 278 + ], + "type": "inline_equation", + "content": "\\phi_g" + }, + { + "bbox": [ + 46, + 232, + 547, + 278 + ], + "type": "text", + "content": ". For any point " + }, + { + "bbox": [ + 46, + 232, + 547, + 278 + ], + "type": "inline_equation", + "content": "\\pmb{p}_i^t" + }, + { + "bbox": [ + 46, + 232, + 547, + 278 + ], + "type": "text", + "content": " lying on the template surface, its features " + }, + { + "bbox": [ + 46, + 232, + 547, + 278 + ], + "type": "inline_equation", + "content": "\\phi (\\pmb{p}_i^t)" + }, + { + "bbox": [ + 46, + 232, + 547, + 278 + ], + "type": "text", + "content": " are sampled from surface features accordingly and fed into two decoders for the prediction of the explicit garment template and pose-dependent wrinkle displacements, which will be combined and transformed to the clothing point cloud." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 48, + 292, + 287, + 382 + ], + "blocks": [ + { + "bbox": [ + 48, + 292, + 287, + 382 + ], + "lines": [ + { + "bbox": [ + 48, + 292, + 287, + 382 + ], + "spans": [ + { + "bbox": [ + 48, + 292, + 287, + 382 + ], + "type": "image", + "image_path": "3a4fb4fca7e8a7f309d99a898b4780b117f625d9bb392b329bd1dfe4fd193c44.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 47, + 387, + 288, + 409 + ], + "lines": [ + { + "bbox": [ + 47, + 387, + 288, + 409 + ], + "spans": [ + { + "bbox": [ + 47, + 387, + 288, + 409 + ], + "type": "text", + "content": "Figure 3. Comparison of the bilinear interpolation on the UV plane and the barycentric interpolation on the surface." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 422, + 287, + 541 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 422, + 287, + 541 + ], + "spans": [ + { + "bbox": [ + 46, + 422, + 287, + 541 + ], + "type": "text", + "content": "vertices " + }, + { + "bbox": [ + 46, + 422, + 287, + 541 + ], + "type": "inline_equation", + "content": "\\mathbf{V}^t" + }, + { + "bbox": [ + 46, + 422, + 287, + 541 + ], + "type": "text", + "content": ", which are learned by feeding the garment code " + }, + { + "bbox": [ + 46, + 422, + 287, + 541 + ], + "type": "inline_equation", + "content": "\\{\\phi_{gc}(\\pmb{v}_n^t)\\}_{n=1}^N" + }, + { + "bbox": [ + 46, + 422, + 287, + 541 + ], + "type": "text", + "content": " to a smaller PointNet++ encoder " + }, + { + "bbox": [ + 46, + 422, + 287, + 541 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_g" + }, + { + "bbox": [ + 46, + 422, + 287, + 541 + ], + "type": "text", + "content": ". Note that the garment-related features " + }, + { + "bbox": [ + 46, + 422, + 287, + 541 + ], + "type": "inline_equation", + "content": "\\phi_g" + }, + { + "bbox": [ + 46, + 422, + 287, + 541 + ], + "type": "text", + "content": " are shared for each outfit across all poses and optimized during the training. Since both the pose-dependent and garment-related geometry features are aligned with each other, we denote them as the surface features " + }, + { + "bbox": [ + 46, + 422, + 287, + 541 + ], + "type": "inline_equation", + "content": "\\{\\phi(\\pmb{v}_n^t)\\}_{n=1}^N" + }, + { + "bbox": [ + 46, + 422, + 287, + 541 + ], + "type": "text", + "content": " for simplicity. Note that the input of " + }, + { + "bbox": [ + 46, + 422, + 287, + 541 + ], + "type": "inline_equation", + "content": "\\phi_g(\\pmb{p}_i^t)" + }, + { + "bbox": [ + 46, + 422, + 287, + 541 + ], + "type": "text", + "content": " has no side effect on the results when trained with only one outfit, as the garment features are invariant to the input poses." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 542, + 288, + 674 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 542, + 288, + 674 + ], + "spans": [ + { + "bbox": [ + 46, + 542, + 288, + 674 + ], + "type": "text", + "content": "Continuous Feature Interpolation. In the implicit modeling solutions [42, 57, 59], features are learned in a spatially continuous manner, which contributes to the fine-grained modeling of clothing details. To sample continuous features in our scheme, we adopt barycentric interpolation on the surface features " + }, + { + "bbox": [ + 46, + 542, + 288, + 674 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 46, + 542, + 288, + 674 + ], + "type": "text", + "content": ". As illustrated in Fig. 3, for any point " + }, + { + "bbox": [ + 46, + 542, + 288, + 674 + ], + "type": "inline_equation", + "content": "p_i^t = \\mathbf{V}^t (\\pmb {b}_i)" + }, + { + "bbox": [ + 46, + 542, + 288, + 674 + ], + "type": "text", + "content": " lying on the template surface, where " + }, + { + "bbox": [ + 46, + 542, + 288, + 674 + ], + "type": "inline_equation", + "content": "\\pmb {b}_i = [n_{i1},n_{i2},n_{i3},b_{i1},b_{i2},b_{i3}]" + }, + { + "bbox": [ + 46, + 542, + 288, + 674 + ], + "type": "text", + "content": " denotes the corresponding vertex indices and barycentric coordinates in " + }, + { + "bbox": [ + 46, + 542, + 288, + 674 + ], + "type": "inline_equation", + "content": "\\mathbf{V}^t" + }, + { + "bbox": [ + 46, + 542, + 288, + 674 + ], + "type": "text", + "content": ". Then, the corresponding surface features can be retrieved via barycentric interpolation, i.e.," + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 107, + 681, + 287, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 107, + 681, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 107, + 681, + 287, + 715 + ], + "type": "interline_equation", + "content": "\\phi \\left(\\boldsymbol {p} _ {i} ^ {t}\\right) = \\sum_ {j = 1} ^ {3} \\left(b _ {i j} * \\phi \\left(\\boldsymbol {v} _ {n _ {i j}} ^ {t}\\right)\\right). \\tag {2}", + "image_path": "c4e2f568700630842105fab72464a0de51f1e13b30fd0ef1fb07274b1f510e97.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 293, + 545, + 342 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 293, + 545, + 342 + ], + "spans": [ + { + "bbox": [ + 305, + 293, + 545, + 342 + ], + "type": "text", + "content": "In this way, the point features are not limited to those learned on the template vertices and are continuously defined over the whole body surface without the seaming discontinuity issue in the UV plane." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 349, + 489, + 361 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 349, + 489, + 361 + ], + "spans": [ + { + "bbox": [ + 306, + 349, + 489, + 361 + ], + "type": "text", + "content": "3.2. Point-based Clothing Deformation" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 367, + 545, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 367, + 545, + 426 + ], + "spans": [ + { + "bbox": [ + 305, + 367, + 545, + 426 + ], + "type": "text", + "content": "Following previous work [36, 39], our approach represents the clothed body as a point cloud. For any point on the surface of the unclothed body model, the corresponding features are extracted from surface features to predict its displacement and normal vector." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 427, + 546, + 619 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 427, + 546, + 619 + ], + "spans": [ + { + "bbox": [ + 304, + 427, + 546, + 619 + ], + "type": "text", + "content": "Explicit Template Decomposition. Instead of predicting the clothing deformation directly, our method decomposes the deformations into two components: garment-related template displacements and pose-dependent wrinkle displacements. To achieve this, the garment-related template is learned from the garment-related features and shared across all poses. Meanwhile, the learning of pose-dependent wrinkles are conditioned on both garment-related and pose-dependent features. Specifically, for the point " + }, + { + "bbox": [ + 304, + 427, + 546, + 619 + ], + "type": "inline_equation", + "content": "\\pmb{p}_i^u = \\mathbf{V}^u (\\pmb {b}_i)" + }, + { + "bbox": [ + 304, + 427, + 546, + 619 + ], + "type": "text", + "content": " at the unclothed body mesh, it has the same vertex indices and barycentric coordinates " + }, + { + "bbox": [ + 304, + 427, + 546, + 619 + ], + "type": "inline_equation", + "content": "\\pmb{b}_{i}" + }, + { + "bbox": [ + 304, + 427, + 546, + 619 + ], + "type": "text", + "content": " as the point " + }, + { + "bbox": [ + 304, + 427, + 546, + 619 + ], + "type": "inline_equation", + "content": "\\pmb{p}_i^t" + }, + { + "bbox": [ + 304, + 427, + 546, + 619 + ], + "type": "text", + "content": " on the template surface. The pose-dependent and garment-related features of the point " + }, + { + "bbox": [ + 304, + 427, + 546, + 619 + ], + "type": "inline_equation", + "content": "\\pmb{p}_i^u" + }, + { + "bbox": [ + 304, + 427, + 546, + 619 + ], + "type": "text", + "content": " are first sampled according to Eq. (2) based on " + }, + { + "bbox": [ + 304, + 427, + 546, + 619 + ], + "type": "inline_equation", + "content": "\\pmb{p}_i^t = \\mathbf{V}^t (\\pmb {b}_i)" + }, + { + "bbox": [ + 304, + 427, + 546, + 619 + ], + "type": "text", + "content": " and then further fed into the garment decoder " + }, + { + "bbox": [ + 304, + 427, + 546, + 619 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_g" + }, + { + "bbox": [ + 304, + 427, + 546, + 619 + ], + "type": "text", + "content": " and the pose decoder " + }, + { + "bbox": [ + 304, + 427, + 546, + 619 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_p" + }, + { + "bbox": [ + 304, + 427, + 546, + 619 + ], + "type": "text", + "content": " for displacement predictions, i.e.," + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 353, + 626, + 545, + 657 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 353, + 626, + 545, + 657 + ], + "spans": [ + { + "bbox": [ + 353, + 626, + 545, + 657 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\boldsymbol {r} _ {i} ^ {g} = \\mathcal {D} _ {g} \\left(\\phi_ {g} \\left(\\boldsymbol {p} _ {i} ^ {t}\\right), \\boldsymbol {p} _ {i} ^ {t}\\right), \\tag {3} \\\\ \\boldsymbol {r} _ {i} ^ {p} = \\mathcal {D} _ {p} \\left(\\oplus \\left(\\phi_ {g} \\left(\\boldsymbol {p} _ {i} ^ {t}\\right), \\phi_ {p} \\left(\\boldsymbol {p} _ {i} ^ {t}\\right)\\right), \\boldsymbol {p} _ {i} ^ {t}\\right), \\\\ \\end{array}", + "image_path": "b7e3f59cd502a3030e5ca1644cf3b03f9cfb09235818d54610bcd7f22848835e.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 665, + 545, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 665, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 305, + 665, + 545, + 715 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 665, + 545, + 715 + ], + "type": "inline_equation", + "content": "\\oplus" + }, + { + "bbox": [ + 305, + 665, + 545, + 715 + ], + "type": "text", + "content": " denotes the concatenation operation, " + }, + { + "bbox": [ + 305, + 665, + 545, + 715 + ], + "type": "inline_equation", + "content": "r_i^g" + }, + { + "bbox": [ + 305, + 665, + 545, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 665, + 545, + 715 + ], + "type": "inline_equation", + "content": "r_i^p" + }, + { + "bbox": [ + 305, + 665, + 545, + 715 + ], + "type": "text", + "content": " are the displacements for garment templates and pose-dependent wrinkles, respectively. Finally, " + }, + { + "bbox": [ + 305, + 665, + 545, + 715 + ], + "type": "inline_equation", + "content": "r_i^g" + }, + { + "bbox": [ + 305, + 665, + 545, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 665, + 545, + 715 + ], + "type": "inline_equation", + "content": "r_i^p" + }, + { + "bbox": [ + 305, + 665, + 545, + 715 + ], + "type": "text", + "content": " will be added together as the clothing deformation " + }, + { + "bbox": [ + 305, + 665, + 545, + 715 + ], + "type": "inline_equation", + "content": "r_i = r_i^g + r_i^p" + }, + { + "bbox": [ + 305, + 665, + 545, + 715 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "text", + "content": "504" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 289, + 193 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 289, + 193 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 289, + 193 + ], + "type": "text", + "content": "Local Transformation. Similar to [36, 39], the displacement " + }, + { + "bbox": [ + 47, + 72, + 289, + 193 + ], + "type": "inline_equation", + "content": "\\boldsymbol{r}_i" + }, + { + "bbox": [ + 47, + 72, + 289, + 193 + ], + "type": "text", + "content": " is learned in a local coordinate system. It is further transformed to the world coordinate system by applying the following transformation, i.e., " + }, + { + "bbox": [ + 47, + 72, + 289, + 193 + ], + "type": "inline_equation", + "content": "\\boldsymbol{x}_i = \\mathcal{T}_i \\boldsymbol{r}_i + \\boldsymbol{p}_i^u" + }, + { + "bbox": [ + 47, + 72, + 289, + 193 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 47, + 72, + 289, + 193 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_i" + }, + { + "bbox": [ + 47, + 72, + 289, + 193 + ], + "type": "text", + "content": " denotes the local transformation calculated based on the unclothed body model. Following [36, 39], the transformation matrix " + }, + { + "bbox": [ + 47, + 72, + 289, + 193 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_i" + }, + { + "bbox": [ + 47, + 72, + 289, + 193 + ], + "type": "text", + "content": " is defined at the point " + }, + { + "bbox": [ + 47, + 72, + 289, + 193 + ], + "type": "inline_equation", + "content": "\\boldsymbol{p}_i^u" + }, + { + "bbox": [ + 47, + 72, + 289, + 193 + ], + "type": "text", + "content": " on the unclothed body model, which naturally supports the barycentric interpolation. Similarly, the normal " + }, + { + "bbox": [ + 47, + 72, + 289, + 193 + ], + "type": "inline_equation", + "content": "\\boldsymbol{n}_i" + }, + { + "bbox": [ + 47, + 72, + 289, + 193 + ], + "type": "text", + "content": " of each point is predicted together with " + }, + { + "bbox": [ + 47, + 72, + 289, + 193 + ], + "type": "inline_equation", + "content": "\\boldsymbol{r}_i" + }, + { + "bbox": [ + 47, + 72, + 289, + 193 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 47, + 72, + 289, + 193 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 47, + 72, + 289, + 193 + ], + "type": "text", + "content": " and transformed by " + }, + { + "bbox": [ + 47, + 72, + 289, + 193 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_i" + }, + { + "bbox": [ + 47, + 72, + 289, + 193 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 200, + 141, + 211 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 200, + 141, + 211 + ], + "spans": [ + { + "bbox": [ + 47, + 200, + 141, + 211 + ], + "type": "text", + "content": "3.3. Loss Functions" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 219, + 287, + 278 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 219, + 287, + 278 + ], + "spans": [ + { + "bbox": [ + 47, + 219, + 287, + 278 + ], + "type": "text", + "content": "Following previous work [36, 39], the point-based clothing deformation is learned with the summation of loss functions: " + }, + { + "bbox": [ + 47, + 219, + 287, + 278 + ], + "type": "inline_equation", + "content": "\\mathcal{L} = \\mathcal{L}_{data} + \\lambda_{rgl}\\mathcal{L}_{rgl}" + }, + { + "bbox": [ + 47, + 219, + 287, + 278 + ], + "type": "text", + "content": ". where " + }, + { + "bbox": [ + 47, + 219, + 287, + 278 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{data}" + }, + { + "bbox": [ + 47, + 219, + 287, + 278 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 219, + 287, + 278 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{rgl}" + }, + { + "bbox": [ + 47, + 219, + 287, + 278 + ], + "type": "text", + "content": " denote the data and regularization terms respectively, and the weight " + }, + { + "bbox": [ + 47, + 219, + 287, + 278 + ], + "type": "inline_equation", + "content": "\\lambda_{rgl}" + }, + { + "bbox": [ + 47, + 219, + 287, + 278 + ], + "type": "text", + "content": " balances the loss terms." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 279, + 287, + 338 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 279, + 287, + 338 + ], + "spans": [ + { + "bbox": [ + 47, + 279, + 287, + 338 + ], + "type": "text", + "content": "Data Term. The data term " + }, + { + "bbox": [ + 47, + 279, + 287, + 338 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{data}" + }, + { + "bbox": [ + 47, + 279, + 287, + 338 + ], + "type": "text", + "content": " is calculated on the final predicted points and normals, i.e., " + }, + { + "bbox": [ + 47, + 279, + 287, + 338 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{data} = \\lambda_p\\mathcal{L}_p + \\lambda_n\\mathcal{L}_n" + }, + { + "bbox": [ + 47, + 279, + 287, + 338 + ], + "type": "text", + "content": ". Specifically, " + }, + { + "bbox": [ + 47, + 279, + 287, + 338 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_p" + }, + { + "bbox": [ + 47, + 279, + 287, + 338 + ], + "type": "text", + "content": " is the normalized Chamfer distance to minimize the bi-directional distances between the point sets of the prediction and the ground-truth scan:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 339, + 216, + 357 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 339, + 216, + 357 + ], + "spans": [ + { + "bbox": [ + 47, + 339, + 216, + 357 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {p} = C h a m f e r \\left(\\{\\boldsymbol {x} _ {i} \\} _ {i = 1} ^ {M}, \\{\\hat {\\boldsymbol {x}} _ {j} \\} _ {j = 1} ^ {N _ {s}}\\right) =", + "image_path": "435347b97d5a01670e7797849e228fc61e4e3d03349486c7668be2dc744e040f.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 367, + 287, + 401 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 367, + 287, + 401 + ], + "spans": [ + { + "bbox": [ + 55, + 367, + 287, + 401 + ], + "type": "interline_equation", + "content": "\\frac {1}{M} \\sum_ {i = 1} ^ {M} \\min _ {j} \\| \\boldsymbol {x} _ {i} - \\hat {\\boldsymbol {x}} _ {j} \\| _ {2} ^ {2} + \\frac {1}{N _ {s}} \\sum_ {j = 1} ^ {N _ {s}} \\min _ {i} \\| \\boldsymbol {x} _ {i} - \\hat {\\boldsymbol {x}} _ {j} \\| _ {2} ^ {2}, \\tag {4}", + "image_path": "70a23f08e8efb05856d5c9b3f6a3a9ee92b8939326646d80277d82ab5a54b52b.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 410, + 287, + 446 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 410, + 287, + 446 + ], + "spans": [ + { + "bbox": [ + 47, + 410, + 287, + 446 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 410, + 287, + 446 + ], + "type": "inline_equation", + "content": "\\hat{\\pmb{x}}_j" + }, + { + "bbox": [ + 47, + 410, + 287, + 446 + ], + "type": "text", + "content": " is the point sampled from the ground-truth surface, " + }, + { + "bbox": [ + 47, + 410, + 287, + 446 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 47, + 410, + 287, + 446 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 410, + 287, + 446 + ], + "type": "inline_equation", + "content": "N_{s}" + }, + { + "bbox": [ + 47, + 410, + 287, + 446 + ], + "type": "text", + "content": " denote the number of the predicted and ground-truth points, respectively." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 447, + 287, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 447, + 287, + 483 + ], + "spans": [ + { + "bbox": [ + 47, + 447, + 287, + 483 + ], + "type": "text", + "content": "The normal loss " + }, + { + "bbox": [ + 47, + 447, + 287, + 483 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_n" + }, + { + "bbox": [ + 47, + 447, + 287, + 483 + ], + "type": "text", + "content": " is an averaged " + }, + { + "bbox": [ + 47, + 447, + 287, + 483 + ], + "type": "inline_equation", + "content": "L1" + }, + { + "bbox": [ + 47, + 447, + 287, + 483 + ], + "type": "text", + "content": " distance between the normal of each predicted point and its nearest ground-truth counterpart:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 491, + 287, + 524 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 491, + 287, + 524 + ], + "spans": [ + { + "bbox": [ + 55, + 491, + 287, + 524 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {n} = L 1 \\left(\\left\\{\\boldsymbol {n} _ {i} \\right\\} _ {i = 1} ^ {M}, \\left\\{\\hat {\\boldsymbol {n}} _ {i} \\right\\} _ {i = 1} ^ {M}\\right) = \\frac {1}{M} \\sum_ {i = 1} ^ {M} \\| \\boldsymbol {n} _ {i} - \\hat {\\boldsymbol {n}} _ {i} \\|, \\tag {5}", + "image_path": "3a1801d6a3e8ace969001da7435bcfb32e9feba7440257e54e6a672ceee37f37.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 533, + 287, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 533, + 287, + 556 + ], + "spans": [ + { + "bbox": [ + 47, + 533, + 287, + 556 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 533, + 287, + 556 + ], + "type": "inline_equation", + "content": "\\hat{\\pmb{n}}_i" + }, + { + "bbox": [ + 47, + 533, + 287, + 556 + ], + "type": "text", + "content": " is the normal of its nearest point in the groundtruth point set." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 558, + 287, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 558, + 287, + 593 + ], + "spans": [ + { + "bbox": [ + 47, + 558, + 287, + 593 + ], + "type": "text", + "content": "Note that we do not apply data terms on the garment templates, as we found such a strategy leads to noisy template learning in our experiments." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 594, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 287, + 714 + ], + "type": "text", + "content": "Regularization Term. The regularization terms are added to prevent the predicted deformations from being extremely large and regularize the garment code. Moreover, following the previous implicit template learning solution [30, 76], we also add regularization on the pose-dependent displacement " + }, + { + "bbox": [ + 46, + 594, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\boldsymbol{r}_i^p" + }, + { + "bbox": [ + 46, + 594, + 287, + 714 + ], + "type": "text", + "content": " to encourage it to be as small as possible. As the pose-dependent displacement represents the clothing deformation in various poses, such a regularization implies that the pose-invariant deformation should be retained in the template displacement " + }, + { + "bbox": [ + 46, + 594, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\boldsymbol{r}_i^g" + }, + { + "bbox": [ + 46, + 594, + 287, + 714 + ], + "type": "text", + "content": ", which forms" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 72, + 545, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 545, + 96 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 545, + 96 + ], + "type": "text", + "content": "the garment-related template shared by all poses. Overall, the regularization term can be written as follows:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 101, + 553, + 144 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 101, + 553, + 144 + ], + "spans": [ + { + "bbox": [ + 306, + 101, + 553, + 144 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {r g l} = \\frac {1}{M} \\sum_ {i = 1} ^ {M} \\| \\mathbf {r} _ {i} \\| _ {2} ^ {2} + \\frac {\\lambda_ {p d}}{M} \\sum_ {i = 1} ^ {M} \\| \\mathbf {r} _ {i} ^ {p} \\| _ {2} ^ {2} + \\frac {\\lambda_ {g c}}{N} \\sum_ {n = 1} ^ {N} \\| \\phi_ {g c} (\\boldsymbol {v} _ {n} ^ {t}) \\| _ {2} ^ {2}. \\tag {6}", + "image_path": "a8b1aed63ef8a0ec172df9c2cd5da452a684f61587017cdbb305be4def676634.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 153, + 387, + 167 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 153, + 387, + 167 + ], + "spans": [ + { + "bbox": [ + 306, + 153, + 387, + 167 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 173, + 547, + 377 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 173, + 547, + 377 + ], + "spans": [ + { + "bbox": [ + 304, + 173, + 547, + 377 + ], + "type": "text", + "content": "Network Architecture. For a fair comparison with POP [39], we modify the official PointNet++ [53] (PN++) architecture so that our encoders have comparable network parameters as POP [39]. The modified PointNet++ architecture has 6 layers for feature abstraction and 6 layers for feature propagation (i.e., " + }, + { + "bbox": [ + 304, + 173, + 547, + 377 + ], + "type": "inline_equation", + "content": "L = 6" + }, + { + "bbox": [ + 304, + 173, + 547, + 377 + ], + "type": "text", + "content": "). Since the input point cloud " + }, + { + "bbox": [ + 304, + 173, + 547, + 377 + ], + "type": "inline_equation", + "content": "\\mathbf{V}^t" + }, + { + "bbox": [ + 304, + 173, + 547, + 377 + ], + "type": "text", + "content": " has constant coordinates, the farthest point sampling in PointNet++ is only performed at the first forward process, and the sampling indices are saved for the next run. In this way, the runtime is significantly reduced for both training and inference so that our pose and garment encoders can have similar network parameters and runtime speeds to POP. Note that the pose and garment encoders in our method can also be replaced with recent state-of-the-art point-based encoders such as PointMLP [40] and PointNeXt [54]. More details about the network architecture and implementation can be found in the Supp.Mat." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 390, + 545, + 425 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 390, + 545, + 425 + ], + "spans": [ + { + "bbox": [ + 304, + 390, + 545, + 425 + ], + "type": "text", + "content": "Datasets. We use CAPE [38], ReSynth [39], and our newly introduced dataset THuman-CloSET for training and evaluation." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 426, + 545, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 426, + 545, + 510 + ], + "spans": [ + { + "bbox": [ + 304, + 426, + 545, + 510 + ], + "type": "text", + "content": "CAPE [38] is a captured human dataset consisting of multiple humans in various motions. The outfits in this dataset mainly include common clothing such as T-shirts. We follow SCALE [36] to choose blazerlong (with outfits of blazer jacket and long trousers) and shortlong (with outfits of short T-shirt and long trousers) from subject 03375 to validate the efficacy of our method." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 510, + 545, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 510, + 545, + 557 + ], + "spans": [ + { + "bbox": [ + 304, + 510, + 545, + 557 + ], + "type": "text", + "content": "ReSynth [39] is a synthetic dataset introduced in POP [39]. It is created by using physics simulation, and contains challenging outfits such as skirts and jackets. We use the official training and test split as [39]." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 558, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 558, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 558, + 547, + 713 + ], + "type": "text", + "content": "THuman-CloSET is our newly introduced dataset, containing high-quality clothed human scans captured by a dense camera rig. We introduce THuman-CloSET for the reason that existing pose-dependent clothing datasets [38, 39] are with either relatively tight clothing or synthetic clothing via physics simulation. In THuman-CloSET, there are more than 2,000 scans of 15 outfitsits with a large variation in clothing style, including T-shirts, pants, skirts, dresses, jackets, and coats, to name a few. For each outfit, the subject is guided to perform different poses by imitating the poses in CAPE. Moreover, each subject has a scan with minimal clothing in A-pose. THuman-CloSET contains well-fitted body models in the form of SMPL-X [50]. Note that" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "text", + "content": "505" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 55, + 96, + 279, + 219 + ], + "blocks": [ + { + "bbox": [ + 47, + 71, + 286, + 93 + ], + "lines": [ + { + "bbox": [ + 47, + 71, + 286, + 93 + ], + "spans": [ + { + "bbox": [ + 47, + 71, + 286, + 93 + ], + "type": "text", + "content": "Table 1. Quantitative comparison with previous point-based methods on ReSynth. † denotes the methods using 1/8 training data." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 55, + 96, + 279, + 219 + ], + "lines": [ + { + "bbox": [ + 55, + 96, + 279, + 219 + ], + "spans": [ + { + "bbox": [ + 55, + 96, + 279, + 219 + ], + "type": "table", + "html": "
MethodChamfer-L2↓Normal diff. ↓
MeanMaxMeanMax
outfit-specific
SCALE [36]1.4918.4511.0411.321
POP [39]1.3567.3391.0131.289
multi-outfit
Baseline (POP [39]) †1.4907.8591.0501.326
Baseline w. PN++ †1.2905.9401.0281.330
CloSET (Ours) †1.2405.5431.0191.315
", + "image_path": "20a840774c6c54741557ef6b34db82457544b7b5e90d5e298d8413945862b2f4.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 52, + 251, + 282, + 324 + ], + "blocks": [ + { + "bbox": [ + 47, + 225, + 286, + 248 + ], + "lines": [ + { + "bbox": [ + 47, + 225, + 286, + 248 + ], + "spans": [ + { + "bbox": [ + 47, + 225, + 286, + 248 + ], + "type": "text", + "content": "Table 2. Quantitative comparison of different methods on the proposed THuman-CloSET dataset in the outfit-specific setting." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 52, + 251, + 282, + 324 + ], + "lines": [ + { + "bbox": [ + 52, + 251, + 282, + 324 + ], + "spans": [ + { + "bbox": [ + 52, + 251, + 282, + 324 + ], + "type": "table", + "html": "
Subject IDSCANImateSNARFPOPCloSET
CDNMLCDNMLCDNMLCDNML
sweater-0001.061.647.112.090.761.550.681.48
longshirt-0011.421.856.662.211.541.831.391.71
skirt-0051.931.749.392.311.661.431.491.36
", + "image_path": "a8dd29c01a9960cbfd6eb7c1360e50292b94d164a12f6902f7d72fa8f3afda22.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 344, + 286, + 464 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 344, + 286, + 464 + ], + "spans": [ + { + "bbox": [ + 46, + 344, + 286, + 464 + ], + "type": "text", + "content": "the loose clothing makes the fitting of the underlying body models quite challenging. For more accurate fitting of the body models, we first fit a SMPL-X model on the scan of the subject in minimal clothing and then adopt its shape parameters for fitting the outfit scans in different poses. More details can be found in the Supp.Mat. In our experiments, we use the outfit scans in 100 different poses for training and use the remaining poses for evaluation. We hope our new dataset can open a promising direction for clothed human modeling and animation from real-world scans." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 479, + 286, + 587 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 479, + 286, + 587 + ], + "spans": [ + { + "bbox": [ + 46, + 479, + 286, + 587 + ], + "type": "text", + "content": "Metrics. Following previous work [39], we generate 50K points from our method and point-based baselines and adopt the Chamfer Distance (see Eq. (4)) and the " + }, + { + "bbox": [ + 46, + 479, + 286, + 587 + ], + "type": "inline_equation", + "content": "L1" + }, + { + "bbox": [ + 46, + 479, + 286, + 587 + ], + "type": "text", + "content": " normal discrepancy (see Eq. (5)) for quantitative evaluation. By default, the Chamfer distance (CD) and normal discrepancy (NML) are reported in the unit of " + }, + { + "bbox": [ + 46, + 479, + 286, + 587 + ], + "type": "inline_equation", + "content": "\\times 10^{-4}m^2" + }, + { + "bbox": [ + 46, + 479, + 286, + 587 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 479, + 286, + 587 + ], + "type": "inline_equation", + "content": "\\times 10^{-1}" + }, + { + "bbox": [ + 46, + 479, + 286, + 587 + ], + "type": "text", + "content": ", respectively. To evaluate the implicit modeling methods, the points are sampled from the surface extracted using Marching Cubes [35]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 595, + 286, + 608 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 595, + 286, + 608 + ], + "spans": [ + { + "bbox": [ + 47, + 595, + 286, + 608 + ], + "type": "text", + "content": "4.1. Comparison with the State-of-the-art Methods" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 614, + 286, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 614, + 286, + 662 + ], + "spans": [ + { + "bbox": [ + 47, + 614, + 286, + 662 + ], + "type": "text", + "content": "We compare results with recent state-of-the-art methods, including point-based approaches SCALE [36], POP [39], and SkiRT [37], and implicit approaches SCANimate [59] and SNARF [10]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 677, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 677, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 677, + 286, + 713 + ], + "type": "text", + "content": "ReSynth. Tab. 1 reports the results of the pose-dependent clothing predictions on unseen motion sequences from the ReSynth [39] dataset, where all 12 outfits fits are used for eval" + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 309, + 118, + 543, + 191 + ], + "blocks": [ + { + "bbox": [ + 305, + 71, + 545, + 115 + ], + "lines": [ + { + "bbox": [ + 305, + 71, + 545, + 115 + ], + "spans": [ + { + "bbox": [ + 305, + 71, + 545, + 115 + ], + "type": "text", + "content": "Table 3. Quantitative comparison of different methods on the ReSynth dataset in the outfit-specific setting. The garment styles are non-skirt, skirt, and dress for carla-004, Christine-027, and felice-004, respectively." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 309, + 118, + 543, + 191 + ], + "lines": [ + { + "bbox": [ + 309, + 118, + 543, + 191 + ], + "spans": [ + { + "bbox": [ + 309, + 118, + 543, + 191 + ], + "type": "table", + "html": "
Subject IDSCANImatePOPSkiRTCloSET
CDNMLCDNMLCDNMLCDNML
carla-0040.901.520.511.020.481.060.491.04
christine-0273.211.661.720.971.540.991.490.97
felice-00420.792.947.341.246.451.256.011.16
", + "image_path": "162f1615d267928e094ee28f0ff99127ee6b796a9c1c91830b2b52edb2f5b597.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 205, + 545, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 205, + 545, + 350 + ], + "spans": [ + { + "bbox": [ + 304, + 205, + 545, + 350 + ], + "type": "text", + "content": "uation. As can be seen, the proposed approach has the lowest mean and max errors, which outperforms all other approaches including POP [39]. Note that our approach needs fewer data for the pose-dependent deformation modeling. By using only 1/8 data, our approach achieves a performance comparable to or even better than other models trained with full data. Tab. 3 also reports outfit-specific performances on 3 selected subject-outfit types, including jackets, skirts, and dresses. In comparison with the recent state-of-the-art method SkiRT [37], our method achieves better results on challenging skirt/dress outfits and comparable results on non-skirt clothing." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 364, + 545, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 364, + 545, + 555 + ], + "spans": [ + { + "bbox": [ + 304, + 364, + 545, + 555 + ], + "type": "text", + "content": "THuman-CloSET. The effectiveness of our method is also validated on our real-world THuman-CloSET dataset. The sparse training poses and loose clothing make this dataset very challenging for clothed human modeling. Tab. 2 reports the quantitative comparisons of different methods on three representative outfits. Fig. 4 also shows example results of different methods, where we follow previous work [39] to obtain meshed results via Poisson surface reconstruction. We can see that our method generalizes better to unseen poses and produces more natural pose-dependent wrinkles than other methods. In our experiments, we found that SNARF [10] fails to learn correct skinning weights due to loose clothing and limited training poses. As discussed in FITE [32], there is an ill-posed issue of jointly optimizing the canonical shape and the skinning fields, which becomes more severe in our dataset." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 563, + 399, + 575 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 563, + 399, + 575 + ], + "spans": [ + { + "bbox": [ + 306, + 563, + 399, + 575 + ], + "type": "text", + "content": "4.2. Ablation Study" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "content": "Evaluation of Continuous Surface Features. The point features in our method are learned on the body surface, which provides a continuous and compact feature learning space. To validate this, Tab. 4 summarizes the feature learning space of different approaches and their performances on two representative outfits from CAPE [38]. Here, we only include the proposed Continuous Surface Features (CSF) in Tab. 4 by applying the continuous features on POP [39] for fair comparisons with SCALE [36] and POP [39]. As discussed previously, existing solutions learn features either in a discontinuous space (e.g., CAPE [38] on the fixed resolu" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "text", + "content": "506" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 53, + 69, + 539, + 350 + ], + "blocks": [ + { + "bbox": [ + 53, + 69, + 539, + 350 + ], + "lines": [ + { + "bbox": [ + 53, + 69, + 539, + 350 + ], + "spans": [ + { + "bbox": [ + 53, + 69, + 539, + 350 + ], + "type": "image", + "image_path": "ef3d7d5c5c228210c963ddd9ab79b78c1e8b679a7e644349e3d7b564cbb524b6.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 97, + 351, + 493, + 363 + ], + "lines": [ + { + "bbox": [ + 97, + 351, + 493, + 363 + ], + "spans": [ + { + "bbox": [ + 97, + 351, + 493, + 363 + ], + "type": "text", + "content": "Figure 4. Comparison of different clothed human modeling methods on the proposed real-world scan dataset." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 48, + 373, + 298, + 550 + ], + "blocks": [ + { + "bbox": [ + 48, + 373, + 298, + 550 + ], + "lines": [ + { + "bbox": [ + 48, + 373, + 298, + 550 + ], + "spans": [ + { + "bbox": [ + 48, + 373, + 298, + 550 + ], + "type": "image", + "image_path": "b67fe7458d39ddb96fdec5ae53b785611ac7b2f4d219d300ec39e7cd889e0bff.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 552, + 287, + 586 + ], + "lines": [ + { + "bbox": [ + 46, + 552, + 287, + 586 + ], + "spans": [ + { + "bbox": [ + 46, + 552, + 287, + 586 + ], + "type": "text", + "content": "Figure 5. Comparison of the approach learned on UV planes (POP [39]) and the approach learned on continuous surfaces (Ours). Our solution alleviates the seam artifacts of POP." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 594, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 289, + 715 + ], + "type": "text", + "content": "tion mesh, POP [39] on the 2D UV plane) or in a space that is too flexible (e.g., NASA [16] in the implicit 3D space), while our approach learns features on continuous and compact surface space. Though SCALE [36] has also investigated using the point-based encoder (PointNet [52]) for pose-dependent feature extraction, it only uses the global features which lack fine-grained information. In contrast, we adopt PointNet++ [53] (PN++) to learn hierarchical surface features, so that the pose-dependent features can be learned more effectively. Fig. 5 shows the qualitative re" + } + ] + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 307, + 399, + 545, + 485 + ], + "blocks": [ + { + "bbox": [ + 305, + 373, + 545, + 396 + ], + "lines": [ + { + "bbox": [ + 305, + 373, + 545, + 396 + ], + "spans": [ + { + "bbox": [ + 305, + 373, + 545, + 396 + ], + "type": "text", + "content": "Table 4. Comparison of the modeling ability of different approaches and their feature learning space on the CAPE dataset." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 307, + 399, + 545, + 485 + ], + "lines": [ + { + "bbox": [ + 307, + 399, + 545, + 485 + ], + "spans": [ + { + "bbox": [ + 307, + 399, + 545, + 485 + ], + "type": "table", + "html": "
MethodsFeaturesChamfer-L2↓Normal diff. ↓
blazerlongshortlongblazerlongshortlong
CAPE [38]Mesh1.961.371.281.15
NASA [16]3D space1.370.951.291.17
SCALE [36]Global1.461.031.341.16
SCALE [36]UV plane1.070.891.221.12
POP [39]UV plane0.780.571.291.24
CSFSurface0.710.541.151.09
", + "image_path": "aeb9246b5368d70939e6a78f56ddd6d5a5f06bba89bb507dbf44f54273f728aa.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 311, + 525, + 538, + 574 + ], + "blocks": [ + { + "bbox": [ + 305, + 488, + 545, + 521 + ], + "lines": [ + { + "bbox": [ + 305, + 488, + 545, + 521 + ], + "spans": [ + { + "bbox": [ + 305, + 488, + 545, + 521 + ], + "type": "text", + "content": "Table 5. Ablation study on the effectiveness of continuous surface features (CSF) and Explicit Template Decomposition (ETD) on a dress outfit (felice-004 from ReSynth [39])." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 311, + 525, + 538, + 574 + ], + "lines": [ + { + "bbox": [ + 311, + 525, + 538, + 574 + ], + "spans": [ + { + "bbox": [ + 311, + 525, + 538, + 574 + ], + "type": "table", + "html": "
MethodPOPPOP + ETDCSFCSF + ETD
CD7.347.056.536.01
NML1.241.171.161.16
", + "image_path": "c8c8de97171999ad25851958bad5294fde4bb79c4f425bca5d0cbc5a88e9f864.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 579, + 545, + 615 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 579, + 545, + 615 + ], + "spans": [ + { + "bbox": [ + 305, + 579, + 545, + 615 + ], + "type": "text", + "content": "sults of the ablation approaches learned on UV planes and continuous surfaces. We can see that our solution clearly alleviates the seam artifacts of POP." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 629, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 629, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 629, + 547, + 715 + ], + "type": "text", + "content": "Evaluation of Explicit Template Decomposition. The decomposed templates help to capture more accurate pose-dependent deformations and produce more natural wrinkles in unseen poses, especially for outfits that differ largely from the body template. To validate the effectiveness of our decomposition strategy, Tab. 5 reports the ablation experiments on a dress outfit (felice-004) of the ReSynth [39]" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 748, + 313, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 748, + 313, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 748, + 313, + 757 + ], + "type": "text", + "content": "507" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 53, + 69, + 541, + 363 + ], + "blocks": [ + { + "bbox": [ + 53, + 69, + 541, + 363 + ], + "lines": [ + { + "bbox": [ + 53, + 69, + 541, + 363 + ], + "spans": [ + { + "bbox": [ + 53, + 69, + 541, + 363 + ], + "type": "image", + "image_path": "982d99d42e975eb97fe7747d0b50555014dc1693aef34c30fab0bb6819d7a963.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 366, + 546, + 388 + ], + "lines": [ + { + "bbox": [ + 46, + 366, + 546, + 388 + ], + "spans": [ + { + "bbox": [ + 46, + 366, + 546, + 388 + ], + "type": "text", + "content": "Figure 6. Comparison of the clothing deformation in unseen poses. Explicit template decomposition (ETD) helps to capture more natural pose-dependent wrinkle details than POP [39]." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 50, + 393, + 298, + 506 + ], + "blocks": [ + { + "bbox": [ + 50, + 393, + 298, + 506 + ], + "lines": [ + { + "bbox": [ + 50, + 393, + 298, + 506 + ], + "spans": [ + { + "bbox": [ + 50, + 393, + 298, + 506 + ], + "type": "image", + "image_path": "b27cfd4580ea5deef04c17561fb87501427671c1e9a0f35c176e6b4d01073783.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 52, + 510, + 282, + 521 + ], + "lines": [ + { + "bbox": [ + 52, + 510, + 282, + 521 + ], + "spans": [ + { + "bbox": [ + 52, + 510, + 282, + 521 + ], + "type": "text", + "content": "Figure 7. Comparison of the learned templates with FITE [32]." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 527, + 289, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 527, + 289, + 635 + ], + "spans": [ + { + "bbox": [ + 46, + 527, + 289, + 635 + ], + "type": "text", + "content": "dataset. We can see that the proposed Explicit Template Decomposition (ETD) brings clear performance gains over baseline methods. Fig. 6 shows the visual improvements of pose-dependent wrinkles when applying the explicit template decomposition. Note that the templates are decomposed in an end-to-end manner in our method. Compared with the implicit template learned in the recent approach FITE [32], the explicit templates in our method contain more details, as shown in Fig. 7." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 645, + 217, + 658 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 645, + 217, + 658 + ], + "spans": [ + { + "bbox": [ + 47, + 645, + 217, + 658 + ], + "type": "text", + "content": "5. Conclusions and Future Work" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "type": "text", + "content": "In this work, we present CloSET, a point-based clothed human modeling method that is built upon a continuous surface and learns to decompose explicit garment templates for better learning of pose-dependent deformations. By learn-" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 396, + 545, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 396, + 545, + 491 + ], + "spans": [ + { + "bbox": [ + 304, + 396, + 545, + 491 + ], + "type": "text", + "content": "ing features on a continuous surface, our solution gets rid of the seam artifacts in previous state-of-the-art point-based methods [36, 39]. Moreover, the explicit template decomposition helps to capture more accurate and natural pose-dependent wrinkles. To facilitate the research in this direction, we also introduce a high-quality real-world scan dataset with diverse outfit styles and accurate body model fitting." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 492, + 546, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 492, + 546, + 635 + ], + "spans": [ + { + "bbox": [ + 304, + 492, + 546, + 635 + ], + "type": "text", + "content": "Limitations and Future Work. Due to the incorrect skinning weight used in our template, the issue of the nonuniform point distribution remains for the skirt and dress outfits. Combining our method with recent learnable skinning solutions [37, 59] could alleviate this issue and further improve the results. Currently, our method does not leverage information from adjacent poses. Enforcing temporal consistency and correspondences between adjacent frames would be interesting for future work. Moreover, incorporating physics-based losses into the learning process like SNUG [61] would also be a promising solution to address the artifacts like self-intersections." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "type": "text", + "content": "Acknowledgements. This work was supported by the National Key R&D Program of China (2022YFF0902200), the National Natural Science Foundation of China (No.62125107 and No.61827805), and the China Postdoctoral Science Foundation (No.2022M721844)." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "text", + "content": "508" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 125 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 125 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 125 + ], + "type": "text", + "content": "[1] Panos Achlioptas, Olga Diamanti, Ioannis Mitliagkas, and Leonidas Guibas. Learning representations and generative models for 3D point clouds. In ICML, pages 40-49, 2018. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 125, + 287, + 168 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 125, + 287, + 168 + ], + "spans": [ + { + "bbox": [ + 53, + 125, + 287, + 168 + ], + "type": "text", + "content": "[2] Thiemo Alldieck, Gerard Pons-Moll, Christian Theobalt, and Marcus Magnor. Tex2shape: Detailed full human body geometry from a single image. In ICCV, pages 2293-2303, 2019. 1, 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 170, + 287, + 214 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 170, + 287, + 214 + ], + "spans": [ + { + "bbox": [ + 53, + 170, + 287, + 214 + ], + "type": "text", + "content": "[3] Ziqian Bai, Timur Bagautdinov, Javier Romero, Michael Zollhöfer, Ping Tan, and Shunsuke Saito. AutoAvatar: Autoregressive neural fields for dynamic avatar modeling. ECCV, 2022. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 216, + 287, + 247 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 216, + 287, + 247 + ], + "spans": [ + { + "bbox": [ + 53, + 216, + 287, + 247 + ], + "type": "text", + "content": "[4] Ilya Baran and Jovan Popovic. Automatic rigging and animation of 3D characters. ACM TOG, 26(3):72-es, 2007. 1, 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 250, + 287, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 250, + 287, + 293 + ], + "spans": [ + { + "bbox": [ + 53, + 250, + 287, + 293 + ], + "type": "text", + "content": "[5] Jan Bednářík, Shaifali Parashar, Erhan Gundogdu, Mathieu Salzmann, and Pascal Fua. Shape reconstruction by learning differentiable surface representations. In CVPR, pages 4715-4724, 2020. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 295, + 287, + 338 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 295, + 287, + 338 + ], + "spans": [ + { + "bbox": [ + 53, + 295, + 287, + 338 + ], + "type": "text", + "content": "[6] Hugo Bertiche, Meysam Madadi, Emilio Tylson, and Sergio Escalera. DeePSD: Automatic deep skinning and pose space deformation for 3D garment animation. In ICCV, pages 5471-5480, 2021. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 339, + 287, + 383 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 339, + 287, + 383 + ], + "spans": [ + { + "bbox": [ + 53, + 339, + 287, + 383 + ], + "type": "text", + "content": "[7] Bharat Lal Bhatnagar, Garvita Tiwari, Christian Theobalt, and Gerard Pons-Moll. Multi-Garment Net: Learning to dress 3D people from images. In ICCV, pages 5420-5430, 2019. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 385, + 287, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 385, + 287, + 418 + ], + "spans": [ + { + "bbox": [ + 53, + 385, + 287, + 418 + ], + "type": "text", + "content": "[8] Andrei Burov, Matthias Nießner, and Justus Thies. Dynamic surface function networks for clothed human bodies. In ICCV, pages 10754-10764, October 2021. 1, 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 419, + 287, + 462 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 419, + 287, + 462 + ], + "spans": [ + { + "bbox": [ + 53, + 419, + 287, + 462 + ], + "type": "text", + "content": "[9] Xu Chen, Tianjian Jiang, Jie Song, Jinlong Yang, Michael J Black, Andreas Geiger, and Otmar Hilliges. gDNA: Towards generative detailed neural avatars. In CVPR, pages 20427-20437, 2022. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 464, + 287, + 508 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 464, + 287, + 508 + ], + "spans": [ + { + "bbox": [ + 48, + 464, + 287, + 508 + ], + "type": "text", + "content": "[10] Xu Chen, Yufeng Zheng, Michael J. Black, Otmar Hilliges, and Andreas Geiger. SNARF: Differentiable forward skinning for animating non-rigid neural implicit shapes. In ICCV, pages 11594–11604, October 2021. 1, 2, 6, 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 510, + 287, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 510, + 287, + 542 + ], + "spans": [ + { + "bbox": [ + 48, + 510, + 287, + 542 + ], + "type": "text", + "content": "[11] Julian Chibane, Aymen Mir, and Gerard Pons-Moll. Neural unsigned distance fields for implicit function learning. In NeurIPS, pages 21638-21652, 2020. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 544, + 287, + 587 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 544, + 287, + 587 + ], + "spans": [ + { + "bbox": [ + 48, + 544, + 287, + 587 + ], + "type": "text", + "content": "[12] Enric Corona, Albert Pumarola, Guillem Alenya, Gerard Pons-Moll, and Francesc Moreno-Noguer. SMPLicit: Topology-aware generative model for clothed people. In CVPR, pages 11875-11885, 2021. 1, 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 589, + 287, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 589, + 287, + 621 + ], + "spans": [ + { + "bbox": [ + 48, + 589, + 287, + 621 + ], + "type": "text", + "content": "[13] Edilson De Aguiar, Leonid Sigal, Adrien Treuille, and Jessica K Hodgins. Stable spaces for real-time clothing. In ACM TOG, volume 29, page 106. ACM, 2010. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 623, + 286, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 623, + 286, + 644 + ], + "spans": [ + { + "bbox": [ + 48, + 623, + 286, + 644 + ], + "type": "text", + "content": "[14] Deform Dynamics. https://deformdynamics.com/.1,3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 647, + 287, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 647, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 647, + 287, + 689 + ], + "type": "text", + "content": "[15] Boyang Deng, JP Lewis, Timothy Jeruzalski, Gerard Pons-Moll, Geoffrey Hinton, Mohammad Norouzi, and Andrea Tagliasacchi. Neural articulated shape approximation. In ECCV, pages 612-628, 2020. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "type": "text", + "content": "[16] Boyang Deng, John P Lewis, Timothy Jeruzalski, Gerard Pons-Moll, Geoffrey Hinton, Mohammad Norouzi, and An-" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 34, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "text", + "content": "drea Tagliasacchi. Nasa neural articulated shape approximation. In ECCV, pages 612-628. Springer, 2020. 2, 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 97, + 545, + 129 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 97, + 545, + 129 + ], + "spans": [ + { + "bbox": [ + 308, + 97, + 545, + 129 + ], + "type": "text", + "content": "[17] Zhantao Deng, Jan Bednárík, Mathieu Salzmann, and Pascal Fua. Better patch stitching for parametric surface reconstruction. In 3DV, pages 593-602, 2020. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 131, + 545, + 175 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 131, + 545, + 175 + ], + "spans": [ + { + "bbox": [ + 307, + 131, + 545, + 175 + ], + "type": "text", + "content": "[18] Theo Deprelle, Thibault Groueix, Matthew Fisher, Vladimir Kim, Bryan Russell, and Mathieu Aubry. Learning elementary structures for 3D shape generation and matching. In NeurIPS, pages 7433-7443, 2019. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 177, + 545, + 210 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 177, + 545, + 210 + ], + "spans": [ + { + "bbox": [ + 308, + 177, + 545, + 210 + ], + "type": "text", + "content": "[19] Haoqiang Fan, Hao Su, and Leonidas J Guibas. A point set generation network for 3D object reconstruction from a single image. In CVPR, pages 2463-2471, 2017. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 212, + 545, + 255 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 212, + 545, + 255 + ], + "spans": [ + { + "bbox": [ + 308, + 212, + 545, + 255 + ], + "type": "text", + "content": "[20] Andrew Feng, Dan Casas, and Ari Shapiro. Avatar reshaping and automatic rigging using a deformable model. In Proceedings of the ACM SIGGRAPH Conference on Motion in Games, pages 57-64, 2015. 1, 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 257, + 545, + 290 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 257, + 545, + 290 + ], + "spans": [ + { + "bbox": [ + 308, + 257, + 545, + 290 + ], + "type": "text", + "content": "[21] Amos Gropp, Lior Yariv, Niv Haim, Matan Atzmon, and Yaron Lipman. Implicit geometric regularization for learning shapes. In ICML, pages 3569-3579, 2020. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 292, + 545, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 292, + 545, + 335 + ], + "spans": [ + { + "bbox": [ + 308, + 292, + 545, + 335 + ], + "type": "text", + "content": "[22] Thibault Groueix, Matthew Fisher, Vladimir G Kim, Bryan C Russell, and Mathieu Aubry. 3D-CODED: 3D correspondences by deep deformation. In ECCV, pages 230-246, 2018. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 337, + 545, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 337, + 545, + 370 + ], + "spans": [ + { + "bbox": [ + 308, + 337, + 545, + 370 + ], + "type": "text", + "content": "[23] Peng Guan, Loretta Reiss, David A Hirshberg, Alexander Weiss, and Michael J Black. DRAPE: DRessing Any PErson. ACM TOG, 31(4):35-1, 2012. 1, 2, 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 372, + 545, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 372, + 545, + 415 + ], + "spans": [ + { + "bbox": [ + 308, + 372, + 545, + 415 + ], + "type": "text", + "content": "[24] Erhan Gundogdu, Victor Constantin, Amrollah Seifoddini, Minh Dang, Mathieu Salzmann, and Pascal Fua. GarNet: A two-stream network for fast and accurate 3D cloth draping. In CVPR, pages 8739-8748, 2019. 1, 2, 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 418, + 545, + 461 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 418, + 545, + 461 + ], + "spans": [ + { + "bbox": [ + 308, + 418, + 545, + 461 + ], + "type": "text", + "content": "[25] Boyan Jiang, Xinlin Ren, Mingsong Dou, Xiangyang Xue, Yanwei Fu, and Yinda Zhang. LoRD: Local 4d implicit representation for high-fidelity dynamic human modeling. In ECCV, pages 307-326. Springer, 2022. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 463, + 545, + 506 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 463, + 545, + 506 + ], + "spans": [ + { + "bbox": [ + 308, + 463, + 545, + 506 + ], + "type": "text", + "content": "[26] Boyi Jiang, Juyong Zhang, Yang Hong, Jinhao Luo, Ligang Liu, and Hujun Bao. BCNet: Learning body and cloth shape from a single image. In ECCV, pages 18-35. Springer, 2020. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 308, + 509, + 545, + 551 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 509, + 545, + 551 + ], + "spans": [ + { + "bbox": [ + 308, + 509, + 545, + 551 + ], + "type": "text", + "content": "[27] Hyomin Kim, Hyeonseo Nam, Junkon Kim, Jaesik Park, and Seungyong Lee. LaplacianFusion: Detailed 3D clothed-human body reconstruction. ACM TOG, 41(6):1-14, 2022. 2" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 308, + 554, + 545, + 587 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 554, + 545, + 587 + ], + "spans": [ + { + "bbox": [ + 308, + 554, + 545, + 587 + ], + "type": "text", + "content": "[28] Zorah Lahner, Daniel Cremers, and Tony Tung. Deepwrinkles: Accurate and realistic clothing modeling. In ECCV, pages 667-684, 2018. 2, 3" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 308, + 589, + 545, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 589, + 545, + 633 + ], + "spans": [ + { + "bbox": [ + 308, + 589, + 545, + 633 + ], + "type": "text", + "content": "[29] Ruilong Li, Julian Tanke, Minh Vo, Michael Zollhöfer, Jürgen Gall, Angjoo Kanazawa, and Christoph Lassner. TAVA: Template-free animatable volumetric actors. In ECCV, pages 419-436. Springer, 2022. 2" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 308, + 635, + 545, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 635, + 545, + 678 + ], + "spans": [ + { + "bbox": [ + 308, + 635, + 545, + 678 + ], + "type": "text", + "content": "[30] Zhe Li, Zerong Zheng, Hongwen Zhang, Chaonan Ji, and Yebin Liu. Avatarcap: Animatable avatar conditioned monocular human volumetric capture. In ECCV, pages 322-341. Springer, 2022. 2, 5" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 308, + 680, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 680, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 680, + 545, + 713 + ], + "type": "text", + "content": "[31] Chen-Hsuan Lin, Chen Kong, and Simon Lucey. Learning efficient point cloud generation for dense 3D object reconstruction. In AAAI, pages 7114-7121, 2018. 2" + } + ] + } + ], + "index": 33 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "text", + "content": "509" + } + ] + } + ], + "index": 35 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "type": "text", + "content": "[32] Siyou Lin, Hongwen Zhang, Zerong Zheng, Ruizhi Shao, and Yebin Liu. Learning implicit templates for point-based clothed human modeling. ECCV, 2022. 2, 3, 6, 8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 107, + 287, + 151 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 107, + 287, + 151 + ], + "spans": [ + { + "bbox": [ + 48, + 107, + 287, + 151 + ], + "type": "text", + "content": "[33] Lijuan Liu, Youyi Zheng, Di Tang, Yi Yuan, Changjie Fan, and Kun Zhou. NeuroSkinning: Automatic skin binding for production characters with deep graph networks. ACM TOG, 38(4):1-12, 2019. 1, 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 153, + 287, + 186 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 153, + 287, + 186 + ], + "spans": [ + { + "bbox": [ + 48, + 153, + 287, + 186 + ], + "type": "text", + "content": "[34] Matthew Loper, Naureen Mahmood, Javier Romero, Gerard Pons-Moll, and Michael J Black. SMPL: A skinned multiperson linear model. ACM TOG, 34(6):248, 2015. 2, 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 188, + 287, + 220 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 188, + 287, + 220 + ], + "spans": [ + { + "bbox": [ + 48, + 188, + 287, + 220 + ], + "type": "text", + "content": "[35] William E Lorensen and Harvey E Cline. Marching cubes: A high resolution 3D surface construction algorithm. In SIGGRAPH, volume 21, pages 163-169, 1987. 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 223, + 287, + 266 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 223, + 287, + 266 + ], + "spans": [ + { + "bbox": [ + 48, + 223, + 287, + 266 + ], + "type": "text", + "content": "[36] Qianli Ma, Shunsuke Saito, Jinlong Yang, Siyu Tang, and Michael J Black. SCALE: Modeling clothed humans with a surface codec of articulated local elements. In CVPR, pages 16082-16093, 2021. 1, 2, 3, 4, 5, 6, 7, 8" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 268, + 287, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 268, + 287, + 300 + ], + "spans": [ + { + "bbox": [ + 48, + 268, + 287, + 300 + ], + "type": "text", + "content": "[37] Qianli Ma, Jinlong Yang, Michael J Black, and Siyu Tang. Neural point-based shape modeling of humans in challenging clothing. 3DV, 2022. 3, 6, 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 303, + 287, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 303, + 287, + 346 + ], + "spans": [ + { + "bbox": [ + 48, + 303, + 287, + 346 + ], + "type": "text", + "content": "[38] Qianli Ma, Jinlong Yang, Anurag Ranjan, Sergi Pujades, Gerard Pons-Moll, Siyu Tang, and Michael J Black. Learning to dress 3D people in generative clothing. In CVPR, pages 6469-6478, 2020. 1, 2, 3, 5, 6, 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 348, + 287, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 348, + 287, + 380 + ], + "spans": [ + { + "bbox": [ + 48, + 348, + 287, + 380 + ], + "type": "text", + "content": "[39] Qianli Ma, Jinlong Yang, Siyu Tang, and Michael J Black. The power of points for modeling humans in clothing. In ICCV, pages 10974-10984, 2021. 1, 2, 3, 4, 5, 6, 7, 8" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 383, + 287, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 383, + 287, + 415 + ], + "spans": [ + { + "bbox": [ + 48, + 383, + 287, + 415 + ], + "type": "text", + "content": "[40] Xu Ma, Can Qin, Haoxuan You, Haoxi Ran, and Yun Fu. Rethinking network design and local geometry in point cloud: A simple residual mlp framework. In ICLR, 2022. 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 418, + 287, + 460 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 418, + 287, + 460 + ], + "spans": [ + { + "bbox": [ + 48, + 418, + 287, + 460 + ], + "type": "text", + "content": "[41] Lars Mescheder, Michael Oechsle, Michael Niemeyer, Sebastian Nowozin, and Andreas Geiger. Occupancy networks: Learning 3D reconstruction in function space. In CVPR, pages 4460-4470, 2019. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 463, + 287, + 496 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 463, + 287, + 496 + ], + "spans": [ + { + "bbox": [ + 48, + 463, + 287, + 496 + ], + "type": "text", + "content": "[42] Marko Mihajlovic, Yan Zhang, Michael J Black, and Siyu Tang. LEAP: Learning articulated occupancy of people. In CVPR, pages 10461-10471, 2021. 2, 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 498, + 287, + 529 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 498, + 287, + 529 + ], + "spans": [ + { + "bbox": [ + 48, + 498, + 287, + 529 + ], + "type": "text", + "content": "[43] Alexandros Neophytou and Adrian Hilton. A layered model of human body and garment deformation. In 3DV, pages 171-178, 2014. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 533, + 287, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 533, + 287, + 586 + ], + "spans": [ + { + "bbox": [ + 48, + 533, + 287, + 586 + ], + "type": "text", + "content": "[44] Hayato Onizuka, Zehra Hayirci, Diego Thomas, Akihiro Sugimoto, Hideaki Uchiyama, and Rin-ichiro Taniguchi. TetraTSDF: 3D human reconstruction from a single image with a tetrahedral outer shell. In CVPR, pages 6011–6020, 2020. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 589, + 287, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 589, + 287, + 632 + ], + "spans": [ + { + "bbox": [ + 48, + 589, + 287, + 632 + ], + "type": "text", + "content": "[45] Pablo Palafox, Aljaž Božić, Justus Thies, Matthias Nießner, and Angela Dai. NPMs: Neural parametric models for 3D deformable shapes. In ICCV, pages 12695-12705, October 2021. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 635, + 287, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 635, + 287, + 666 + ], + "spans": [ + { + "bbox": [ + 48, + 635, + 287, + 666 + ], + "type": "text", + "content": "[46] Pablo Palafox, Nikolaos Sarafianos, Tony Tung, and Angela Dai. SPAMs: Structured implicit parametric models. In CVPR, pages 12851-12860, 2022. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "type": "text", + "content": "[47] Junyi Pan, Xiaoguang Han, Weikai Chen, Jiapeng Tang, and Kui Jia. Deep mesh reconstruction from single RGB images via topology modification networks. In ICCV, pages 9963-9972, 2019. 2" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 33, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "text", + "content": "[48] Jeong Joon Park, Peter Florence, Julian Straub, Richard Newcombe, and Steven Lovegrove. DeepSDF: Learning continuous signed distance functions for shape representation. In CVPR, pages 165-174, 2019. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 118, + 545, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 545, + 160 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 545, + 160 + ], + "type": "text", + "content": "[49] Chaitanya Patel, Zhouyingcheng Liao, and Gerard Pons-Moll. TailorNet: Predicting clothing in 3D as a function of human pose, shape and garment style. In CVPR, pages 7363-7373, 2020. 1, 2, 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 162, + 545, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 162, + 545, + 215 + ], + "spans": [ + { + "bbox": [ + 308, + 162, + 545, + 215 + ], + "type": "text", + "content": "[50] Georgios Pavlakos, Vasileios Choutas, Nima Ghorbani, Timo Bolkart, Ahmed AA Osman, Dimitrios Tzionas, and Michael J Black. Expressive body capture: 3D hands, face, and body from a single image. In CVPR, pages 10975-10985, 2019. 3, 5" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 217, + 545, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 217, + 545, + 259 + ], + "spans": [ + { + "bbox": [ + 308, + 217, + 545, + 259 + ], + "type": "text", + "content": "[51] Gerard Pons-Moll, Sergi Pujades, Sonny Hu, and Michael Black. ClothCap: Seamless 4d clothing capture and retargeting. ACM TOG, 36(4), 2017. Two first authors contributed equally. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 261, + 545, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 261, + 545, + 293 + ], + "spans": [ + { + "bbox": [ + 308, + 261, + 545, + 293 + ], + "type": "text", + "content": "[52] Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. PointNet: Deep learning on point sets for 3D classification and segmentation. In CVPR, pages 652-660, 2017. 7" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 294, + 545, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 294, + 545, + 326 + ], + "spans": [ + { + "bbox": [ + 308, + 294, + 545, + 326 + ], + "type": "text", + "content": "[53] Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J Guibas. Pointnet++: Deep hierarchical feature learning on point sets in a metric space. NeurIPS, 30, 2017. 3, 5, 7" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 327, + 545, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 327, + 545, + 380 + ], + "spans": [ + { + "bbox": [ + 308, + 327, + 545, + 380 + ], + "type": "text", + "content": "[54] Guocheng Qian, Yuchen Li, Houwen Peng, Jinjie Mai, Hasan Hammoud, Mohamed Elhoseiny, and Bernard Ghanem. Pointnext: Revisiting pointnet++ with improved training and scaling strategies. NeurIPS, 35:23192-23204, 2022. 5" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 382, + 545, + 424 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 382, + 545, + 424 + ], + "spans": [ + { + "bbox": [ + 308, + 382, + 545, + 424 + ], + "type": "text", + "content": "[55] Shenhan Qian, Jiale Xu, Ziwei Liu, Liqian Ma, and Shenghua Gao. UNIF: United neural implicit functions for clothed human reconstruction and animation. In ECCV, 2022. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 426, + 545, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 426, + 545, + 437 + ], + "spans": [ + { + "bbox": [ + 308, + 426, + 545, + 437 + ], + "type": "text", + "content": "[56] Renderpeople, 2020. https://renderpeople.com.3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 438, + 545, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 438, + 545, + 480 + ], + "spans": [ + { + "bbox": [ + 308, + 438, + 545, + 480 + ], + "type": "text", + "content": "[57] Shunsuke Saito, Zeng Huang, Ryota Natsume, Shigeo Morishima, Angjoo Kanazawa, and Hao Li. PIFu: Pixel-aligned implicit function for high-resolution clothed human digitization. In ICCV, pages 2304-2314, 2019. 2, 4" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 482, + 545, + 523 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 482, + 545, + 523 + ], + "spans": [ + { + "bbox": [ + 308, + 482, + 545, + 523 + ], + "type": "text", + "content": "[58] Shunsuke Saito, Tomas Simon, Jason Saragih, and Hanbyul Joo. PIFuHD: Multi-level pixel-aligned implicit function for high-resolution 3D human digitization. In CVPR, pages 84–93, 2020. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 525, + 545, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 525, + 545, + 568 + ], + "spans": [ + { + "bbox": [ + 308, + 525, + 545, + 568 + ], + "type": "text", + "content": "[59] Shunsuke Saito, Jinlong Yang, Qianli Ma, and Michael J Black. SCANimate: Weakly supervised learning of skinned clothed avatar networks. In CVPR, pages 2886-2897, 2021. 1, 2, 3, 4, 6, 7, 8" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 308, + 570, + 545, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 570, + 545, + 602 + ], + "spans": [ + { + "bbox": [ + 308, + 570, + 545, + 602 + ], + "type": "text", + "content": "[60] Igor Santesteban, Miguel A. Otaduy, and Dan Casas. Learning-Based Animation of Clothing for Virtual Try-On. Comput. Graph. Forum, 38(2):355-366, 2019. 2" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 308, + 604, + 545, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 604, + 545, + 634 + ], + "spans": [ + { + "bbox": [ + 308, + 604, + 545, + 634 + ], + "type": "text", + "content": "[61] Igor Santesteban, Miguel A Otaduy, and Dan Casas. SNUG: Self-supervised neural dynamic garments. In CVPR, pages 8140-8150, 2022. 8" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 308, + 636, + 545, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 636, + 545, + 678 + ], + "spans": [ + { + "bbox": [ + 308, + 636, + 545, + 678 + ], + "type": "text", + "content": "[62] Igor Santesteban, Nils Thuerey, Miguel A Otaduy, and Dan Casas. Self-supervised collision handling via generative 3D garment models for virtual try-on. In CVPR, pages 11763-11773, 2021. 2" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 308, + 680, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 680, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 680, + 545, + 713 + ], + "type": "text", + "content": "[63] Yu Shen, Junbang Liang, and Ming C Lin. Gan-based garment generation using sewing pattern images. In ECCV, volume 1, page 3, 2020. 2" + } + ] + } + ], + "index": 32 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 298, + 749, + 313, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 749, + 313, + 757 + ], + "spans": [ + { + "bbox": [ + 298, + 749, + 313, + 757 + ], + "type": "text", + "content": "510" + } + ] + } + ], + "index": 34 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 714 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "text", + "content": "[64] Garvita Tiwari, Bharat Lal Bhatnagar, Tony Tung, and Gerard Pons-Moll. SIZER: A dataset and model for parsing 3D clothing and learning size sensitive 3D clothing. In ECCV, volume 12348, pages 1-18, 2020. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 118, + 287, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 118, + 287, + 162 + ], + "spans": [ + { + "bbox": [ + 48, + 118, + 287, + 162 + ], + "type": "text", + "content": "[65] Garvita Tiwari, Nikolaos Sarafianos, Tony Tung, and Gerard Pons-Moll. Neural-GIF: Neural generalized implicit functions for animating people in clothing. In ICCV, pages 11708-11718, 2021. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 164, + 287, + 208 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 164, + 287, + 208 + ], + "spans": [ + { + "bbox": [ + 48, + 164, + 287, + 208 + ], + "type": "text", + "content": "[66] Raquel Vidaurre, Igor Santesteban, Elena Garces, and Dan Casas. Fully convolutional graph neural networks for parametric virtual try-on. In Comput. Graph. Forum, volume 39, pages 145-156. Wiley Online Library, 2020. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 210, + 287, + 253 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 210, + 287, + 253 + ], + "spans": [ + { + "bbox": [ + 48, + 210, + 287, + 253 + ], + "type": "text", + "content": "[67] Shaofei Wang, Marko Mihajlovic, Qianli Ma, Andreas Geiger, and Siyu Tang. Meta­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­.\nclothed human models from few depth images. In NeurIPS, 2021. 2, 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 256, + 287, + 298 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 256, + 287, + 298 + ], + "spans": [ + { + "bbox": [ + 48, + 256, + 287, + 298 + ], + "type": "text", + "content": "[68] Jane Wu, Zhenglin Geng, Hui Zhou, and Ronald Fedkiw. Skinning a parameterization of three-dimensional space for neural network cloth. arXiv preprint arXiv:2006.04874, 2020. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 301, + 287, + 345 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 301, + 287, + 345 + ], + "spans": [ + { + "bbox": [ + 48, + 301, + 287, + 345 + ], + "type": "text", + "content": "[69] Donglai Xiang, Fabian Prada, Timur Bagautdinov, Weipeng Xu, Yuan Dong, He Wen, Jessica Hodgins, and Chenglei Wu. Modeling clothing as a separate layer for an animatable human avatar. ACM TOG, dec 2021. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 347, + 287, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 347, + 287, + 380 + ], + "spans": [ + { + "bbox": [ + 48, + 347, + 287, + 380 + ], + "type": "text", + "content": "[70] Yuliang Xiu, Jinlong Yang, Dimitrios Tzionas, and Michael J Black. ICON: implicit clothed humans obtained from normals. In CVPR, pages 13286-13296. IEEE, 2022. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 382, + 287, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 382, + 287, + 426 + ], + "spans": [ + { + "bbox": [ + 48, + 382, + 287, + 426 + ], + "type": "text", + "content": "[71] Jinlong Yang, Jean-Sebastien Franco, Franck Hetroy-Wheeler, and Stefanie Wuhrer. Analyzing clothing layer deformation statistics of 3D human motions. In ECCV, September 2018. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 428, + 287, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 428, + 287, + 470 + ], + "spans": [ + { + "bbox": [ + 48, + 428, + 287, + 470 + ], + "type": "text", + "content": "[72] Shan Yang, Zherong Pan, Tanya Amert, Ke Wang, Licheng Yu, Tamara Berg, and Ming C Lin. Physics-inspired garment recovery from a single-view image. ACM TOG, 37(5):1-14, 2018. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 474, + 287, + 517 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 474, + 287, + 517 + ], + "spans": [ + { + "bbox": [ + 48, + 474, + 287, + 517 + ], + "type": "text", + "content": "[73] Tao Yu, Zerong Zheng, Kaiwen Guo, Pengpeng Liu, Qionghai Dai, and Yebin Liu. Function4D: Real-time human volumetric capture from very sparse consumer rgbd sensors. In CVPR, June 2021. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 519, + 287, + 552 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 519, + 287, + 552 + ], + "spans": [ + { + "bbox": [ + 48, + 519, + 287, + 552 + ], + "type": "text", + "content": "[74] Ilya Zakharkin, Kirill Mazur, Artur Grigorev, and Victor Lempitsky. Point-based modeling of human clothing. In ICCV, pages 14718-14727, October 2021. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 554, + 287, + 597 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 554, + 287, + 597 + ], + "spans": [ + { + "bbox": [ + 48, + 554, + 287, + 597 + ], + "type": "text", + "content": "[75] Zerong Zheng, Han Huang, Tao Yu, Hongwen Zhang, Yandong Guo, and Yebin Liu. Structured local radiance fields for human avatar modeling. In CVPR, pages 15893-15903, 2022. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 600, + 287, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 600, + 287, + 632 + ], + "spans": [ + { + "bbox": [ + 48, + 600, + 287, + 632 + ], + "type": "text", + "content": "[76] Zerong Zheng, Tao Yu, Qionghai Dai, and Yebin Liu. Deep implicit templates for 3D shape representation. In CVPR, pages 1429-1439, 2021. 3, 5" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 635, + 287, + 677 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 635, + 287, + 677 + ], + "spans": [ + { + "bbox": [ + 48, + 635, + 287, + 677 + ], + "type": "text", + "content": "[77] Zerong Zheng, Tao Yu, Yebin Liu, and Qionghai Dai. PaMIR: Parametric model-conditioned implicit representation for image-based human reconstruction. IEEE TPAMI, 2021. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 680, + 287, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 680, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 48, + 680, + 287, + 714 + ], + "type": "text", + "content": "[78] Zerong Zheng, Tao Yu, Yixuan Wei, Qionghai Dai, and Yebin Liu. DeepHuman: 3D human reconstruction from a single image. In ICCV, pages 7739-7749, 2019. 3" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 72, + 546, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 72, + 546, + 128 + ], + "spans": [ + { + "bbox": [ + 307, + 72, + 546, + 128 + ], + "type": "text", + "content": "[79] Heming Zhu, Yu Cao, Hang Jin, Weikai Chen, Dong Du, Zhangye Wang, Shuguang Cui, and Xiaoguang Han. Deep Fashion3D: A dataset and benchmark for 3D garment reconstruction from single images. In ECCV, volume 12346, pages 512-530, 2020. 2" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "text", + "content": "511" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Clothed Human Performance Capture With a Double-Layer Neural Radiance Fields/edeeca28-ac72-40e3-9181-9737d0c908f2_content_list.json b/2023/Clothed Human Performance Capture With a Double-Layer Neural Radiance Fields/edeeca28-ac72-40e3-9181-9737d0c908f2_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..e91da11aab6bef23b4adb416332ac2af0e24e435 --- /dev/null +++ b/2023/Clothed Human Performance Capture With a Double-Layer Neural Radiance Fields/edeeca28-ac72-40e3-9181-9737d0c908f2_content_list.json @@ -0,0 +1,1997 @@ +[ + { + "type": "text", + "text": "Clothed Human Performance Capture with a Double-layer Neural Radiance Fields", + "text_level": 1, + "bbox": [ + 100, + 161, + 868, + 200 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Kangkan Wang $^{1,2}$ , Guofeng Zhang $^{3}$ , Suxu Cong $^{1}$ , Jian Yang $^{1,2}$", + "bbox": [ + 231, + 228, + 734, + 247 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1Key Lab of Intelligent Perception and Systems for High-Dimensional Information of Ministry of Education, 2Jiangsu Key Lab of Image and Video Understanding for Social Security,", + "bbox": [ + 77, + 250, + 934, + 283 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "School of Computer Science and Engineering, Nanjing University of Science and Technology, China \n3State Key Laboratory of CAD&CG, Zhejiang University, China", + "bbox": [ + 107, + 283, + 904, + 317 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 340, + 312, + 355 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "This paper addresses the challenge of capturing performance for the clothed humans from sparse-view or monocular videos. Previous methods capture the performance of full humans with a personalized template or recover the garments from a single frame with static human poses. However, it is inconvenient to extract cloth semantics and capture clothing motion with one-piece template, while single frame-based methods may suffer from instable tracking across videos. To address these problems, we propose a novel method for human performance capture by tracking clothing and human body motion separately with a double-layer neural radiance fields (NeRFs). Specifically, we propose a double-layer NeRFs for the body and garments, and track the densely deforming template of the clothing and body by jointly optimizing the deformation fields and the canonical double-layer NeRFs. In the optimization, we introduce a physics-aware cloth simulation network which can help generate physically plausible cloth dynamics and body-cloth interactions. Compared with existing methods, our method is fully differentiable and can capture both the body and clothing motion robustly from dynamic videos. Also, our method represents the clothing with an independent NeRFs, allowing us to model implicit fields of general clothes feasibly. The experimental evaluations validate its effectiveness on real multi-view or monocular videos.", + "bbox": [ + 75, + 370, + 472, + 717 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 77, + 735, + 208, + 749 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Performance capture for clothed humans is one of the essential problems in the metaverse, and it not only captures the inner human body motion but also recovers the outer clothing motion which has many promising applications such as virtual try-on, video editing, and telepresence. From sparse-view or monocular videos of a moving human", + "bbox": [ + 75, + 758, + 470, + 843 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "in general clothes, its goal is to recover the dynamic 3D shape sequence of the human body and clothing simultaneously that are consistent with the observed frames in both human shape and motion. This is a very challenging problem since the dynamic human could be with arbitrary motions and with complex non-rigid cloth deformations, and the clothing in motion is difficult to maintain physically plausible interactions with the body.", + "bbox": [ + 499, + 342, + 892, + 454 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Previous systems [4, 21, 38, 39] reconstruct 3D clothed humans by using depth sensors or fitting a personalized template [9, 10, 37] to the image observations (e.g., body joints and silhouettes). Only recovering one-piece geometry which unifies the human body and clothes, these systems fail to track the motion of the clothing and achieve clothing editing on 3D humans, which are the prerequisites in many VR/AR applications like virtual dressing. On the contrary, cloth can be extracted and tracked from depth scans [26, 40] accurately by fitting pre-built cloth templates to the scans which have limited applications when 3D data are unavailable. Existing garment estimation methods [3, 12, 43, 44] from color images require the person facing the camera and in static poses. When the human is moving and the clothing is deforming, these methods may recover the 3D garments unreliable. Recent methods [16, 31] track body and clothing motion simultaneously from videos, but they need to re-build cloth template for a new performer and the running efficiency is very low due to online cloth simulation or computationally-exhaustive optimization, which prohibits them from being widely deployed for daily applications.", + "bbox": [ + 497, + 458, + 892, + 750 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recent works [20, 34, 41] adopt dynamic human NeRFs to capture human motion and obtain impressive tracking results. By capturing the temporally-varying appearance in the videos, dynamic NeRFs [34] can provide dense photometric constraints to track the deforming geometry of the performer. However, they represent the human with a single NeRFs without modeling cloth, and the clothing motion cannot be extracted. In this paper, we aim to track the cloth", + "bbox": [ + 497, + 755, + 892, + 866 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 129, + 49, + 200, + 85 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 252, + 48, + 791, + 87 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author: wangkangkan@njust.edu.cn", + "bbox": [ + 92, + 853, + 376, + 866 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "21098", + "bbox": [ + 463, + 885, + 505, + 897 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ing and body motion simultaneously with dynamic NeRFs. However, this problem is rather challenging due to two major questions we need to solve: how to represent dynamic clothing and human body with NeRFs, and how to capture clothing and human body motion with plausible body-cloth interactions based on the implicit representation.", + "bbox": [ + 75, + 125, + 468, + 207 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we propose a novel method for clothed human performance capture with a double-layer NeRFs. Specifically, a double-layer NeRFs is modeled for both the body and clothing in the template space, and transformed to the observation space with the corresponding deformation fields, and the rendered images are then synthesized by composing the two dynamic NeRFs. We first estimate the template shape in canonical frame and learn the geometry network supervised by the template geometry. In the rendering, we compose the double-layer NeRFs with the guidance of deformed body and clothing meshes. Then, by minimizing the difference between the rendered color and observed color, the deformation fields and the canonical N-NeRFs are optimized jointly. The deformation field is represented as the inverse deformation of the template mesh, thus the densely deforming geometry of the template can be recovered simultaneously. In addition, we adopt a physics-aware network learnt from simulation data between various cloth types and humans to constrain the dynamic clothing and preserve physically plausible body-cloth interactions, resulting in realistic cloth geometry tracking. Compared to previous methods, our method is fully differentiable and can recover the realistic motion of both the clothing and body from dynamic videos with arbitrary human poses and complex cloth deformations. The experimental qualitative and quantitative results on datasets of DynaCap [8] and DeepCap [10] prove that the proposed approach can robustly and accurately capture the motion for clothed humans. In summary, the primary contributions of our work include:", + "bbox": [ + 75, + 209, + 470, + 609 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose a double-layer NeRFs for dynamic humans in general clothing, allowing us to model implicit humans with a variety of clothes (e.g., loose dresses).", + "- To the best of our knowledge, we propose the first framework to capture clothing motion separately from the human body using the double-layer NeRFs, which provides dense appearance constraints on the geometry tracking and improves the robustness and accuracy.", + "- A differentiable physics-aware network is learnt for d-different common garments and used to preserve physically plausible cloth deformations in motion capture." + ], + "bbox": [ + 92, + 615, + 468, + 781 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 77, + 792, + 218, + 806 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Clothed human reconstruction. Depth-based methods [4,21,38,39] reconstruct 3D humans in realtime by fusing the depth maps in a volumetric way. However, depth-based methods cannot synthesize photo-realistic videos and", + "bbox": [ + 75, + 809, + 468, + 866 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "have constrained applications in daily life due to data capture limitations, e.g., high power consumption and sensitive to sunlight. Based on deep learning techniques, some approaches predict 3D humans from color images through implicit function [27, 28], coarse-to-fine mesh deformation [45], surface normal optimization [36, 42], or UV mapping [2]. Although these methods can reconstruct human details, the recovered shapes do not contain human semantics like human pose, shape and clothing. To recover the clothing, some methods [3, 19] add displacements on vertices of SMPL model [18] which only support tight clothing and recover coarse-level geometry. Template-based methods [9, 10, 37] recover 3D human motions by non-rigidly fitting a person-specific template mesh to 2D observed features (e.g., human joints or silhouettes). With a personalized template, these methods obtain impressive performance capture. However, the recovered meshes integrate the body and clothing in one piece, which hinders their applications like garment retargeting and virtual trying-on.", + "bbox": [ + 497, + 125, + 890, + 388 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Clothing motion capture. Traditional methods [26, 40] capture the cloth motion by precisely fitting a cloth template to 3D data from depth sensors, which cannot generalize to the video-based setting and are limited to applications in indoor scenes. Deep learning-based approaches [3, 12, 43, 44] estimate the 3D garment shape from color images. However, these methods require the human facing the camera with a static pose and do not try temporally tracking of dynamic garments. Some works [23, 29] predict the cloth shape conditioned on the human pose and shape but the shapes are not consistent with the observed data. Recent works [16, 31] capture both human body and clothing motion from dynamic videos. MulayCap [31] optimizes a simulated garment on SMPL model with semantic cloth segmentation, but their traditional simulation and optimization are both computationally expensive. DeepCloth [16] incorporates a physics-based simulator to supervise the cloth deformations estimated with multi-view joints and silhouettes as DeepCap [10]. However, their simulation layer is not differentiable and needs to adjust the cloth material types online, which is a tedious and time consuming step. Moreover, with the same limitation as DeepCap, they cannot recover the detailed cloth deformation based on the sparse features.", + "bbox": [ + 497, + 391, + 890, + 710 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Implicit Human Reconstruction. Clothed humans can be reconstructed through implicit representation-based methods such as voxel representation [33, 42], implicit function [27, 28], or neural radiance fields (NeRFs) [20, 24, 25]. Recent works combine explicit and implicit representations to produce high-quality avatar [11] or reconstruct high-fidelity garment mesh [44], which show the advantage of implicit representation on recovering geometry details. However, these methods are limited to self-rotating human motion or a single color image. Neuralbody [25] and AniNErf [24] extend NeRFs for a dynamic human with defor", + "bbox": [ + 497, + 713, + 890, + 866 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "21099", + "bbox": [ + 463, + 885, + 504, + 897 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/be976881f92fa9d91395654b0c248d79871f51f1ec80781719aa574f9cd75ff7.jpg", + "image_caption": [ + "Figure 1. Overview of the proposed approach. Our approach can capture both the clothing and human body motion, and achieve novel-view synthesis from sparse-view or monocular videos. We learn a double-layer neural radiance fields (NeRFs) for the clothing and human body in canonical frame, and jointly optimize the canonical double-layer NeRFs and the human deformation constrained with a physics-aware simulation network. Please refer to Sec. 3 for details of our method." + ], + "image_footnote": [], + "bbox": [ + 100, + 122, + 868, + 406 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "mation fields based on the linear blend skinning (LBS) [15]. Some human NeRFs methods [17, 41] model nonrigid human deformations with a point displacement on top of LBS. However, without using any clothing motion prior, these methods cannot effectively recover the motion of humans in loose clothes, and the geometry of clothes is not reconstructed reliably under nonrigid cloth deformations. The recent work, NerfCap [34] incorporates the embedded graph prior into the human NeRFs and represents nonlinear surface deformations for loose clothes successfully. However, their method does not model the clothing separately and thus fails to track the clothing motion.", + "bbox": [ + 75, + 472, + 472, + 638 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Proposed Approach", + "text_level": 1, + "bbox": [ + 75, + 647, + 268, + 663 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Given sparse-view or monocular videos of a clothed human, we track the motion of both the human body and clothing and synthesize novel-view videos of the human with a novel performance capture method based on a double-layer neural radiance fields (NeRFs). Figure 1 shows an overview of the proposed method. The double-layer NeRFs is learnt for the human body and clothing in canonical frame (Sec. 3.2). Then, the canonical NeRFs is dynamically transformed into the observed frame with inverse deformation fields of the deformed template for clothing and body respectively. From the dynamic double-layer NeRFs, we render the images using a composite rendering. By matching the synthesized images with the video frames and constraining the cloth deformations with a physics-aware simulation", + "bbox": [ + 75, + 672, + 472, + 866 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "network, we estimate the double-layer NeRFs and the human deformation jointly (Sec. 3.3).", + "bbox": [ + 497, + 472, + 890, + 501 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Canonical Human Shape", + "text_level": 1, + "bbox": [ + 497, + 509, + 729, + 526 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We decompose the clothed human into undressed body and garments (like T-shirt and pants), and represent them with SMPL model [18] and a parametric PCA model respectively. To include the human details (e.g., the head and hand), we obtain the canonical body model $M$ by deforming the SMPL model to a personalized template with non-rigid deformation [32]. We design different styles (size and sleeve length, et al.) for each garment category, and simulate these garments on the template SMPL model using Marvelous Designer [1]. Then, we obtain a linear model through PCA on the simulated meshes,", + "bbox": [ + 497, + 530, + 890, + 684 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {G} (\\alpha) = \\bar {\\boldsymbol {G}} + \\boldsymbol {B} _ {g} \\alpha , \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 623, + 695, + 890, + 712 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\bar{G}$ is the mean garment shape, $B_{g}$ denotes the linear basis matrix, and $\\alpha$ are parameters of the cloth shape $G(\\alpha)$ .", + "bbox": [ + 497, + 722, + 890, + 751 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Double-layer NeRFs for Dynamic Humans", + "text_level": 1, + "bbox": [ + 497, + 760, + 862, + 776 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Unlike a single NeRFs for clothed humans [17,24,25,34, 41], we represent the clothing with an independent NeRFs on the body, which forms a double-layer NeRFs. The implicit representation can be extended to multiple garments. Canonical NeRF. We use an occupancy network [22] to represent canonical geometry since the occupancy can mod", + "bbox": [ + 497, + 782, + 890, + 866 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "21100", + "bbox": [ + 463, + 885, + 505, + 897 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "el geometry more precisely than density [20] as,", + "bbox": [ + 75, + 125, + 394, + 138 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {o} (\\boldsymbol {x}) = \\boldsymbol {F} _ {o} \\left(\\gamma_ {\\boldsymbol {x}} (\\boldsymbol {x})\\right), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 203, + 149, + 468, + 165 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\gamma_{\\pmb{x}}$ is the positional encoding [20]. The color network $F_{c}$ in canonical frame is formulated as,", + "bbox": [ + 77, + 174, + 468, + 201 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {c} _ {i} (\\boldsymbol {x}) = \\boldsymbol {F} _ {c} \\left(\\gamma_ {\\boldsymbol {x}} (\\boldsymbol {x}), \\varphi_ {i}\\right), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 189, + 212, + 468, + 228 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\varphi_{i}$ is an appearance latent code [24, 25] for frame $i$ . We define occupancy network $\\{F_o^b,F_o^g\\}$ , and color network $\\{F_c^b,\\varphi^b,F_c^g,\\varphi^g\\}$ for the body and clothing, respectively.", + "bbox": [ + 77, + 237, + 468, + 277 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To facilitate the learning of occupancy network, we incorporate a signed distance fields (SDFs) network $S$ ,", + "bbox": [ + 75, + 279, + 468, + 306 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {S}: \\boldsymbol {p} \\in \\mathbb {R} ^ {3} \\rightarrow \\boldsymbol {s} \\in \\mathbb {R}. \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 196, + 315, + 468, + 331 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "which predicts the SDF value $s$ for point $p$ . We use the losses proposed in [30] to learn the SDFs of the shape,", + "bbox": [ + 75, + 340, + 468, + 369 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} L _ {s d f} = \\sum_ {\\boldsymbol {p} \\in \\Phi} | \\boldsymbol {S} (\\boldsymbol {p}) | + \\sum_ {\\boldsymbol {p} \\in \\Phi} (1 - \\langle \\nabla \\boldsymbol {S} (\\boldsymbol {p}), \\bar {\\boldsymbol {n}}) \\rangle \\tag {5} \\\\ + \\sum_ {\\boldsymbol {p} \\in \\Omega} | \\| \\nabla \\boldsymbol {S} (\\boldsymbol {p}) \\| _ {2} - 1 | + \\sum_ {\\boldsymbol {p} \\in \\Omega \\backslash \\Phi} \\exp (- \\delta \\cdot | \\boldsymbol {S} (\\boldsymbol {p}) |), \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 77, + 377, + 485, + 441 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\bar{n}$ is the surface normal, $\\nabla S(\\pmb{p})$ indicates the gradient of the SDFs, $\\Omega$ and $\\Phi$ denotes the 3D space and shape surface respectively. The first and second terms are used to estimate valid SDFs and normals on the surfaces, respectively. The third term constrains the norm of SDF gradients to be 1, and the last term penalizes SDF values of non-surface points near 0 with $\\delta \\gg 1$ . We define SDF network $\\{\\pmb{S}^b, \\pmb{S}^g\\}$ for both the body and clothing.", + "bbox": [ + 75, + 451, + 468, + 562 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Supervised by the predictions of SDF network, the occupancy network is learnt with the following loss,", + "bbox": [ + 75, + 562, + 468, + 590 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nL _ {o c c} = \\sum_ {\\boldsymbol {x} \\in \\Omega} L _ {C} \\left(\\boldsymbol {F} _ {\\boldsymbol {o}} \\left(\\gamma_ {\\boldsymbol {x}} (\\boldsymbol {x})\\right), \\boldsymbol {o}\\right), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 164, + 599, + 468, + 628 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\pmb{o}$ is the occupancy indicated by SDFs (if $S(\\pmb{x}) < = 0$ , $o = 1$ ; otherwise, $o = 0$ ), and $L_{C}$ denotes the cross-entropy loss. We adopt the compositional design of SDF and occupancy because it is better for volumetric learning with continuous occupancy fields [22] compared to discrete labels extracted from SDFs. The template is varying during learning geometry model, and the loss is formulated as,", + "bbox": [ + 75, + 637, + 468, + 733 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nL _ {g e o} = \\mu_ {1} L _ {s d f} + \\mu_ {2} L _ {o c c}. \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 184, + 745, + 468, + 760 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mu_{1}$ and $\\mu_{2}$ are the weight for each term. Through the optimization, the NeRF geometry is consistent with the estimated template so that densely deforming geometry of the template can be tracked by utilizing dense photometric constraints from NeRF color model.", + "bbox": [ + 75, + 769, + 468, + 836 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Deformation Model. Following [34], we transform sampled points in observed space to canonical space with the", + "bbox": [ + 75, + 838, + 468, + 866 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "inverse deformation of the nearest vertex on the deformed template. In this way, the human deformation model is disentangled into the non-rigid surface deformation (i.e., embedded graph deformation as in [34]) and the articulated skeletal motion. For the clothing and body, we define the deformation network $F_{d}^{g}$ with latent codes $\\omega^{g}$ and $F_{d}^{b}$ with latent codes $\\omega^{b}$ , respectively. Through nonrigid deformation, we obtain the clothing mesh $\\hat{G}$ and body mesh $\\hat{M}$ . Then, the linear blend skinning [15] is applied to obtain the final cloth mesh $\\tilde{G}$ and body mesh $\\tilde{M}$ in the observation space. We recover the SMPL model with [13] for input video frames and also estimate a canonical body shape.", + "bbox": [ + 497, + 125, + 890, + 305 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Composite Rendering. In [6], a blending weight is predicted for each sampled point to blend the static background and dynamic target in a scene. This blending method cannot separate the clothing and human body in the NeRFs effectively. We propose an efficient composite rendering strategy for the dynamic double-layer NeRF. Specifically, for a casted ray $r$ , we find the intersecting triangle with the deformed meshes. If the intersected triangle is from the body mesh, we set the mask $m_r^b$ of all sampled points on $r$ to 1; otherwise, $m_r^b$ is set to 0. Similarly, if the intersected triangle is from the cloth mesh, we set the mask $m_r^g$ to 1; otherwise, $m_r^g$ is set to 0. The complete human is rendered as,", + "bbox": [ + 497, + 307, + 890, + 472 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {T} _ {r} ^ {i} = \\prod_ {j = 1} ^ {i - 1} \\left(1 - \\boldsymbol {\\sigma} _ {g} ^ {j} \\left(1 - \\boldsymbol {m} _ {r} ^ {b}\\right)\\right) \\left(1 - \\boldsymbol {\\sigma} _ {b} ^ {j} \\boldsymbol {m} _ {r} ^ {b}\\right), \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 542, + 485, + 890, + 524 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\boldsymbol {C}} _ {r} = \\sum_ {i = 1} ^ {n} \\boldsymbol {T} _ {r} ^ {i} \\left(\\boldsymbol {o} _ {g} ^ {i} \\left(1 - \\boldsymbol {m} _ {r} ^ {b}\\right) \\boldsymbol {c} _ {g} ^ {i} + \\boldsymbol {o} _ {b} ^ {i} \\boldsymbol {m} _ {r} ^ {b} \\boldsymbol {c} _ {b} ^ {i}\\right),\n$$\n", + "text_format": "latex", + "bbox": [ + 542, + 527, + 847, + 562 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\{o_g,o_b\\}$ and $\\{c_{g},c_{b}\\}$ denote the point occupancy and color respectively, which are defined for both clothing and body. $T_r^i$ is the accumulated transparency along the ray. We also render the clothing color $\\hat{C}_r^g$ and body color $\\hat{C}_r^b$", + "bbox": [ + 499, + 574, + 890, + 631 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Clothed Human Performance Capture with Double-layer NeRFs", + "text_level": 1, + "bbox": [ + 499, + 642, + 889, + 669 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We estimate the double-layer NeRFs and template deformation of the body and clothing by minimizing the following loss between the synthesized color and observed color:", + "bbox": [ + 499, + 678, + 890, + 719 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} L _ {r g b} = \\sum_ {\\boldsymbol {r} \\in R} \\| \\hat {\\boldsymbol {C}} (\\boldsymbol {r}) - \\boldsymbol {C} (\\boldsymbol {r}) \\| _ {2} ^ {2} \\tag {9} \\\\ + \\sum_ {\\boldsymbol {r} \\in \\boldsymbol {R}} \\| (\\hat {\\boldsymbol {C}} (\\boldsymbol {r}) ^ {b} - \\boldsymbol {C} (\\boldsymbol {r})) (1 - \\boldsymbol {m} _ {r} ^ {g}) \\| _ {2} ^ {2} \\\\ + \\sum_ {\\boldsymbol {r} \\in \\boldsymbol {R}} \\| (\\hat {\\boldsymbol {C}} (\\boldsymbol {r}) ^ {g} - \\boldsymbol {C} (\\boldsymbol {r})) (1 - \\boldsymbol {m} _ {r} ^ {b}) \\| _ {2} ^ {2}, \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 536, + 732, + 890, + 826 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $R$ is the set of casted rays. While using the dense photometric constraint defined in Eq. 9, the model training", + "bbox": [ + 499, + 838, + 890, + 866 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "21101", + "bbox": [ + 463, + 885, + 502, + 897 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "remains ill-posed and body-cloth collisions are prone to occur. To solve these problems, we design several regularization losses to constrain the joint optimization.", + "bbox": [ + 75, + 125, + 468, + 167 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Physics-aware constraint. Cloth simulation is adopted in previous works [16, 40] to obtain realistic performance capture. Different from their traditional simulation method, we introduce a differential network that learns a mapping from human and cloth parameters to physics-aware cloth shapes. We simulate the data in Marvelous Designer using the method of [23]. For each garment category, we use 25 garment styles $\\gamma \\in \\mathbb{R}^4$ as [23]. To account for the effects of different materials, we simulate 8 different materials in Marvelous Designer, and represent the material type $\\pmb{\\tau}$ using one-hot labels. We train the simulation network $D_{\\phi}$ using a MLP to minimize the L1-loss between the simulated cloth meshes and the predicted ones $G_{s} = D_{\\phi}(\\theta ,\\beta ,\\gamma ,\\tau)$ with the input of human pose and shape, and garment and material styles. The physical simulation loss enforces the deformed cloth mesh $\\tilde{G}$ to be close to $G_{s}$ ,", + "bbox": [ + 75, + 169, + 472, + 391 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nL _ {s i m} = \\sum_ {t \\in \\boldsymbol {G}} \\rho \\left(\\left\\| \\tilde {\\boldsymbol {G}} ^ {t} - \\boldsymbol {G} _ {s} ^ {t} \\right\\| _ {2} ^ {2}\\right), \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 173, + 402, + 468, + 432 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $t$ denotes vertex index of $G$ , and $\\rho$ is the Geman-McClure robust function [5]. The simulation loss can penalize body-cloth collisions and preserve physically plausible deformations.", + "bbox": [ + 75, + 444, + 468, + 498 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As-rigid-as-possible Loss. To make sure the template deforms smoothly, we utilize an as-rigid-as-possible loss [32] on the embedded graph deformation of the human body and clothing, $L_{arap} = L_{arap}(\\hat{M}) + L_{arap}(\\hat{G})$ .", + "bbox": [ + 75, + 501, + 468, + 558 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Mask Loss. We render the mask image $R(\\mathbf{U})$ of the combined mesh $\\mathbf{U} = (M, G)$ by using a differentiable renderer [35] based on point cloud, and compute the mask loss with the input human mask image $\\bar{R}$ :", + "bbox": [ + 75, + 558, + 468, + 612 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nL _ {\\text {m a s k}} (\\boldsymbol {U}) = \\left\\| R (\\boldsymbol {U}) - \\bar {R} \\right\\| _ {2} ^ {2}. \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 173, + 625, + 468, + 644 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We utilize the mask loss on both the deformed mesh and the simulated mesh as, $L_{mask} = L_{mask}(\\tilde{M},\\tilde{G}) + L_{mask}(\\tilde{M},G_s)$ .", + "bbox": [ + 75, + 656, + 468, + 698 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Attachment Loss. During human motion, some cloth parts are always attached to the body (e.g., the dress strap is fixed on the shoulder). To ensure this constraint, we define an attachment loss:", + "bbox": [ + 75, + 700, + 468, + 754 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nL _ {\\text {a t t a c h}} = \\sum_ {i, j \\in A} \\rho \\left(\\| \\tilde {\\boldsymbol {G}} ^ {i} - \\tilde {\\boldsymbol {M}} ^ {j} \\| _ {2} ^ {2}\\right), \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 159, + 766, + 468, + 796 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $A$ is a set of selected vertices on the clothing mesh attached to corresponding points of the body mesh.", + "bbox": [ + 75, + 809, + 468, + 837 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Interpenetration Loss. Interpenetrations may occur between the predicted garments and the body. We use the", + "bbox": [ + 75, + 838, + 468, + 866 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "interpenetration term [7] to alleviate this problem:", + "bbox": [ + 499, + 125, + 830, + 138 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nL _ {i n t e r p} \\left(M _ {1}, M _ {2}\\right) = \\frac {1}{| C |} \\sum_ {i, j \\in C} R e L U \\left(\\left(M _ {1} ^ {i} - M _ {2} ^ {j}\\right) \\cdot N _ {1} ^ {i}\\right), \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 502, + 147, + 890, + 192 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $C$ is the set of corresponding points between the mesh $M_{1}$ and mesh $M_2$ , $|C|$ denotes the number of corresponding points, and $N_{1}$ is the normal of mesh $M_{1}$ . We use interpenetration loss in both canonical and posed space,", + "bbox": [ + 497, + 193, + 890, + 249 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nL _ {i n t e r p} = L _ {i n t e r p} (\\hat {\\boldsymbol {M}}, \\hat {\\boldsymbol {G}}) + L _ {i n t e r p} (\\tilde {\\boldsymbol {M}}, \\tilde {\\boldsymbol {G}}). \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 527, + 256, + 890, + 273 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3.1 Total Loss Function", + "text_level": 1, + "bbox": [ + 499, + 288, + 690, + 300 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We first learn the geometry network of both the clothing and human body in canonical frame. We use the first frame of multiple views or a video in self-rotating motion to estimate the canonical geometry by optimizing all the losses. Then, by fixing the clothing shape and the geometry network, we track the motion of each video frame and estimate the appearance of NeRFs using the following total loss:", + "bbox": [ + 497, + 309, + 890, + 406 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} L = \\lambda_ {1} L _ {r g b} + \\lambda_ {2} L _ {a r a p} + \\lambda_ {3} L _ {s i m} \\tag {15} \\\\ + \\lambda_ {4} L _ {\\text {m a s k}} + \\lambda_ {5} L _ {\\text {a t t a c h}} + \\lambda_ {6} L _ {\\text {i n t e r p}}, \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 546, + 415, + 890, + 448 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\{\\lambda_1\\ldots \\lambda_6\\}$ are the balancing weights. For a video with $N$ frames, the estimated parameter set $X$ is:", + "bbox": [ + 499, + 454, + 889, + 483 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nX = \\left\\{\\boldsymbol {F} _ {c} ^ {b}, \\boldsymbol {F} _ {c} ^ {g}, \\boldsymbol {F} _ {d} ^ {b}, \\boldsymbol {F} _ {d} ^ {g} \\right\\} \\cup \\left\\{\\varphi_ {i} ^ {b}, \\varphi_ {i} ^ {g}, \\omega_ {i} ^ {b}, \\omega_ {i} ^ {g} \\mid i \\in 1, \\dots , N \\right\\}, \\tag {16}\n$$\n", + "text_format": "latex", + "bbox": [ + 500, + 491, + 889, + 520 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "which includes the weights of the appearance and deformation networks and the corresponding conditional variables for each frame in the video.", + "bbox": [ + 497, + 520, + 889, + 561 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 499, + 574, + 631, + 589 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Implementation details", + "text_level": 1, + "bbox": [ + 499, + 596, + 712, + 611 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Datasets and metrics. We design common garments (e.g., dress, T-shirt, skirt, pants, et al.) and simulate the data in the Marvelous Designer [1] under various human shapes and poses and cloth materials. We test our method on publicly datasets, DynaCap [8], and DeepCap [10]. We use four cameras around the human for training and the remaining cameras for test. Figure 2 shows some recovered results using our approach on different clothed humans. Our method is also tested on a monocular video in outdoor scene. Please refer to the supplementary video for more experimental results. The geometry accuracy is evaluated with the intersection over union (IoU) between the ground truth human masks and 2D projections of the recovered human shapes. Following [10, 34], the IoU (\\%) is computed on all views (AMVIoU), all views except the input view (RVIoU), and the input view (SVIoU). The novel-view synthesis quality is evaluated with two metrics, i.e., peak signal-to-noise ratio (PSNR) and structural similarity index (SSIM) as [24, 34].", + "bbox": [ + 497, + 617, + 890, + 866 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "21102", + "bbox": [ + 463, + 885, + 504, + 897 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/8a0fc36af3de34eb7391906c13c0ef74a624df0294341d8ad22b62a2d3895e1c.jpg", + "image_caption": [ + "Figure 2. The results of our approach on different humans in various garments. In each example, we show the ground truth image in a novel view, the reconstructed geometry (overlayed on the image), and the novel-view synthesis image. From left to right: the S4 from DeepCap dataset [10], \"FranziRed\" from DynaCap dataset [8], and the S1 from DeepCap dataset [10]." + ], + "image_footnote": [], + "bbox": [ + 206, + 125, + 759, + 495 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Architecture and experimental settings. We adopt the S-DF network from [30], and the same occupancy and color network as [34]. In the non-rigid deformation network, there are 8 layers with a skip connection at the fourth layer and the hidden size is 1024. The deformation and appearance latent codes both have dimensions of 128. We set both weight $\\mu_{1}$ and $\\mu_{2}$ to 1. We set the loss weight $\\{\\lambda_1,\\dots,\\lambda_6\\}$ to $\\{1.0,0.1,0.02,50,0.1,100.0\\}$ empirically. The Adam optimizer [14] is used in the optimization, and the learning rate is $5e^{-4}$ at the start of training and exponentially decays to $5e^{-5}$ during the training. With a GPU of RTX 2080 Ti, our network training costs about 12 hours tested on a four-view video with 300 frames.", + "bbox": [ + 75, + 546, + 472, + 725 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Comparison to state-of-the-art methods", + "text_level": 1, + "bbox": [ + 75, + 732, + 418, + 748 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We compare our method with two state-of-the-art approaches for human performance capture, DeepCap [10] and NerfCap [34]. We also compare with BCNet [12], TailorNet [23], and ICON [36]. For DeepCap and NerfCap, the experimental results are from [34]. For other compared methods, we implement them using their released codes. We do not compare with DeepCloth [16] since their codes are not released and DeepCap [10] outperforms them in ge", + "bbox": [ + 75, + 755, + 468, + 866 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/1ff9c9dbaf1d11b40b0a8bfc59010e8887f868c9ab35e9a11252fef88c0e7adc.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodAMVIoURVIoUSVIoU
DeepCap [10]86.1785.7586.90
NerfCap [34]88.9688.3290.09
ICON [36]--93.56
BCNet [12]82.3381.6782.94
TailorNet [23]83.8783.5084.51
Our method89.5888.7591.03
", + "bbox": [ + 531, + 543, + 863, + 644 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/2db4dd8c42fc8b8e0a0b51f273787f61d3c9e050f3892f677c041b3f00f83b88.jpg", + "table_caption": [ + "Evaluated on the subject S4 from DeepCap dataset [10]." + ], + "table_footnote": [], + "table_body": "
MethodAMVIoURVIoUSVIoU
DeepCap [10]85.3885.3785.41
NerfCap [34]88.0888.0788.12
ICON [36]--95.21
BCNet [12]78.9178.7780.83
TailorNet [23]78.1578.0078.21
Our method88.7488.4089.68
", + "bbox": [ + 529, + 663, + 862, + 764 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluated on the 'FranziRed' from DynaCap dataset [8].", + "bbox": [ + 505, + 764, + 882, + 777 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1. Comparison to the state-of-the-arts in surface reconstruction accuracy. Note that, the accuracy of ICON [36] is high since they use the human mask in estimating the geometry, and their dress shapes are not recovered accurately as shown in Fig. 3.", + "bbox": [ + 497, + 781, + 890, + 833 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "ometry tracking as reported in [16]. We conduct the qualitative and quantitative experiments on S4 of DeepCap and", + "bbox": [ + 497, + 838, + 892, + 866 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "21103", + "bbox": [ + 463, + 885, + 504, + 897 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "\"FranziRed\" from DynaCap. The test dataset are the same as [34] which include 300 - 400 frames randomly sampled from the original dataset.", + "bbox": [ + 75, + 125, + 468, + 167 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/88b1a4cb9af775bb7d201b363d90d292d84745d0ef7b38652693afd627517315.jpg", + "image_caption": [ + "(a) (b)", + "#", + "(c)", + "(d)", + "(e)", + "(f)", + "(g)", + "Figure 3. Qualitative comparison with the state-of-the-arts. (a) Ground truth image. (b) DeepCap [10]. (c) NerfCap [34]. (d) ICON [36]. (e) BCNet [12]. (f) TailorNet [23]. (g) Our method." + ], + "image_footnote": [], + "bbox": [ + 87, + 172, + 458, + 456 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We show the qualitative comparisons in Figure 3 and list the quantitative results in Table 1. We only compute the error of the input view for ICON since they use their own camera system. Our method performs significantly better than the compared methods [12,23,36] without a cloth template. BCNet [12] also separates the modeling of the clothing and body, but as a supervised learning method, the performance may degrade dramatically because of data domain gap. When the human is moving, BCNet fails to recover the dresses accurately due to arbitrary cloth deformation. In contrast, the appearance loss of our double-layer NeRF-s poses dense constraints on the clothing and body motion capture in a weakly-supervised manner. As a result, our method can achieve a much higher accuracy of geometry tracking even without the ground truth. The predicted garments of TailorNet [23] are not consistent with the real images since their model is conditioned on human poses and shapes. ICON obtains higher IoU accuracy by using the human mask in recovering the geometry. However, there are obvious artifacts on the recovered dresses of ICON [36]. Based on one-piece personalized template, it is hard for both DeepCap [10] and NerfCap [34] to extract the clothing motion, which is important for downstream applications in AR/VR. In contrast, our method tracks the motion for both the clothing and body. Without a pre-scanned cloth template, our method still obtains a high tracking accuracy", + "bbox": [ + 75, + 507, + 472, + 866 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/54e80774fb5acbf141fa10a175ca80032e1281ddec166103e1cfbaeb2c0f9092.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodS4 of [10]‘FranziRed” of [8]
PSNRSSIMPSNRSSIM
NerfCap [34]24.5290.7923.1891.49
Our method23.0690.2821.7090.54
", + "bbox": [ + 512, + 123, + 880, + 182 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/cab623925684cc5840f29ec54590789ecc64a3609c95233775725ef32c2c547a.jpg", + "table_caption": [ + "Table 2. Comparison to NerfCap [34] in terms of novel-view synthesis on two datasets." + ], + "table_footnote": [], + "table_body": "
MethodAMVIoUPSNRSSIM
w/o simulation loss89.0322.6689.35
w/o mask loss88.7523.2189.49
w/o color loss89.36--
blending render [6]88.7523.6490.51
1 camera view88.4921.3487.12
2 camera views89.2321.8988.38
6 camera views90.1023.8090.60
ours(4 views)89.5823.0690.28
", + "bbox": [ + 521, + 210, + 870, + 338 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 3. Ablation studies on S4 from DeepCap dataset [10].", + "bbox": [ + 517, + 340, + 870, + 353 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "thanks to the implicit representation.", + "bbox": [ + 499, + 366, + 742, + 378 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We also compare the quality of free-viewpoint synthesis with NerfCap [34]. Figure 4 shows two compared examples and Table 2 reports the quantitative comparison. Without using any clothing template, our method still obtains close or comparable synthesis quality to NerfCap that uses a ground truth template. Our results are rendered by composing the NeRFs of the body and clothing, which is more complicated than a single NeRFs of NerfCap. For example, when the mutual motion occurs between the body and clothing (like cloth sliding), new body parts may appear, leading to the decrease of our synthesis quality.", + "bbox": [ + 497, + 380, + 890, + 533 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b31ed73df970870e0ac60c671fb255b4b56a6dec0a49dcb91bc50f4c4e39e3ad.jpg", + "image_caption": [ + "(a)", + "Figure 4. Comparison of novel-view synthesis with NerfCap [34]. (a,d) Ground truth. (b,e) NerfCap [34]. (c,f) Our method." + ], + "image_footnote": [], + "bbox": [ + 532, + 540, + 579, + 638 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/4b6c2bf4ed8e4d070e86dd39d566ddb6a26ddc5765438fc0fd4defb254a15aab.jpg", + "image_caption": [ + "(b)" + ], + "image_footnote": [], + "bbox": [ + 581, + 540, + 626, + 638 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a25fb15d5ad67a3b9a172f16580039cd498fa17f3bae6ab361029b724c0a5e22.jpg", + "image_caption": [ + "(c)" + ], + "image_footnote": [], + "bbox": [ + 628, + 541, + 673, + 638 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/9db08b2b3e6b76f75dd2bce570b5b16060b886a071774cef6c4b96c8e6b6dfb6.jpg", + "image_caption": [ + "(d)" + ], + "image_footnote": [], + "bbox": [ + 680, + 543, + 739, + 638 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/cd2c2cea695505c9a22d825c8c53bf9392a26ef91e14bbdb6e0c842d7fa094d2.jpg", + "image_caption": [ + "(e)" + ], + "image_footnote": [], + "bbox": [ + 741, + 543, + 798, + 638 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a4ea041c9b8c5221d0d53fa4a1e91d258adf299ff1937c3ff6548ee5a3325bed.jpg", + "image_caption": [ + "(f)" + ], + "image_footnote": [], + "bbox": [ + 800, + 543, + 855, + 638 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Ablation studies", + "text_level": 1, + "bbox": [ + 499, + 691, + 660, + 705 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The ablation experiments are performed to demonstrate the effectiveness of our framework on the S4 from DeepCap dataset [10]. The quantitative results are listed in Table 3.", + "bbox": [ + 497, + 713, + 890, + 755 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Effectiveness of different losses. We first evaluate the effectiveness of the main losses in our optimization, i.e., the simulation loss, the mask loss, and the color loss. We report the accuracy by removing each of them, and show qualitative comparison examples in Figure 5. Without the simulation loss, the body-cloth collision becomes severe, showing that the simulation loss can preserve physically-plausible body-cloth interactions. Using the input human mask can", + "bbox": [ + 497, + 755, + 890, + 866 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "21104", + "bbox": [ + 463, + 885, + 504, + 897 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "improve the alignment accuracy with the human silhouettes. The color loss of dynamic NeRF can further enhance the detailed deformation on the human surface by posing dense appearance constraints for the geometry tracking.", + "bbox": [ + 75, + 125, + 468, + 181 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/2cafdd527f3a2550598e5a1ade66938596b8b9596ec78c10737c0b37c78c0f11.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 122, + 188, + 181, + 277 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/4d736e527bb28a2679c32af1379f3c4f5499cb79e9066a309d145f9ee5ace126.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 181, + 188, + 240, + 277 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/1766e4d22506c87b72375e406ff51d7835b77fd9b23b1b9a2afa8852aba8d93a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 243, + 188, + 302, + 277 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/c79b7ed46ec3a8a1bafa4155948909b1c12191c4f583d9a179b50a9c545e5544.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 304, + 188, + 361, + 277 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/e7bed27739caa1a681482e887dc055a37127764a71a461e8ff4cab7f6f296a95.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 364, + 188, + 423, + 277 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/1724da4f1adf14bc96d2506cd00674659143a08ffed2de310f3e4e9c25ebeb36.jpg", + "image_caption": [ + "(a)" + ], + "image_footnote": [], + "bbox": [ + 124, + 279, + 181, + 366 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/aa1efa9baa10faff6110740345a60648ac0dccb7f31a35de16fc3dbd19552feb.jpg", + "image_caption": [ + "(b)" + ], + "image_footnote": [], + "bbox": [ + 181, + 279, + 240, + 366 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/d8231ca330ce52dd0abb612881dbc9a4ca8db648163bd6e74e2ab55f7807d1fb.jpg", + "image_caption": [ + "(c)" + ], + "image_footnote": [], + "bbox": [ + 243, + 279, + 302, + 366 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/18e9d9f5c403d40fe372bae16fc436608f3be6fa451955b4702ca51345da250f.jpg", + "image_caption": [ + "(d)" + ], + "image_footnote": [], + "bbox": [ + 304, + 279, + 361, + 366 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/7c308394e2760380e490274e723d9cac1c30133251f4fcce88de743707d497db.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 364, + 279, + 423, + 366 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/79ee651b5e22c62d990dd3556beff6d0154366dca669112e480546541eeba10d.jpg", + "image_caption": [ + "Figure 5. Ablation study for different losses. (a) Ground truth. (b) Using complete losses. (c) W/o simulation loss. (d) W/o mask loss. (e) W/o color loss. There are unnatural cloth deformations without the simulation loss. The rendered mask cannot align with the silhouette well without the mask loss. Compared to that without color loss, the surface detailed deformation can be further improved using complete losses (b).", + "(a)" + ], + "image_footnote": [], + "bbox": [ + 80, + 482, + 134, + 589 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/5e97de6abcc6dd2fd0494bb139acfce978e0a9275dc0fe761e051e706832b6f3.jpg", + "image_caption": [ + "(b)" + ], + "image_footnote": [], + "bbox": [ + 156, + 484, + 200, + 589 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/c152489fa55ea4a2a8c57e0c7574c55aef8536ad8288d2b2217715f51ab727e0.jpg", + "image_caption": [ + "Figure 6. Comparison of different rendering. (a) Ground truth. (b) Our composite rendering. (c) Blend rendering [6]. We show the synthesis of the clothing, body, and full human. There are much cloth appearance on the body caused by the blending [6], while our method separates the appearance of the clothing and body better." + ], + "image_footnote": [], + "bbox": [ + 201, + 484, + 253, + 587 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/0be68f33d69b1ca701cedbe9cb6c1297bddd553ac882b75fddd130432b06c198.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 255, + 484, + 304, + 587 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/64ffa03747a1eda5401011102a27b6b127ce8c51d2d269a3097e0e9bef886d47.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 322, + 497, + 366, + 554 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/52a19c0ba71493b4000011081db876501e483239da9460d86251f195d1db8658.jpg", + "image_caption": [ + "(c)" + ], + "image_footnote": [], + "bbox": [ + 369, + 484, + 421, + 587 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/5aadb97f34f34b404da349258615dfd2c46bc7e5b4b789249abd4ab597f21b12.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 423, + 484, + 473, + 587 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effectiveness of the composite rendering. We compare our composite rendering with the blending rendering in [6]. Through free composition of two NeRFs for each point, the method of [6] can obtain high fitting color. However, the blending may generate missing areas or artifacts in the novel-view synthesis. Moreover, it fails to recover the appearance of the body and clothing reliably as shown in Figure 6. Comparatively, we can recover the appearance of the body and clothing separately and compose them into a full human more accurately using the composite rendering.", + "bbox": [ + 75, + 672, + 467, + 809 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The number of camera views. We investigate the influence of different number of input camera views on the performance of our approach. We compare the results of our method under $\\{1,2,4,6\\}$ views, and show the comparison", + "bbox": [ + 75, + 811, + 467, + 866 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "example of novel-view synthesis in Figure 7. The results become slightly better using more camera views because the moving humans are captured more completely.", + "bbox": [ + 497, + 125, + 890, + 167 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/d34100012af1e0ae21a4d54372f1f017c0e088b6729939077a77d3d7391303ff.jpg", + "image_caption": [ + "Figure 7. Ablation study for different numbers of camera views. With more cameras, the novel-view synthesis quality becomes better (marked in cycle)." + ], + "image_footnote": [], + "bbox": [ + 547, + 176, + 608, + 286 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/d43466951f76608d0ce7b92136b54930a7814df294209549227d43b1c06c6b0c.jpg", + "image_caption": [ + "Ground Truth", + "1 view" + ], + "image_footnote": [], + "bbox": [ + 611, + 176, + 665, + 286 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/61a22e02dc1f226e377f5244f301979de0d86d0ee3925d60d2b0bfb408662ad4.jpg", + "image_caption": [ + "2 views" + ], + "image_footnote": [], + "bbox": [ + 668, + 176, + 724, + 286 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/f8cbdcbc3c5ae4e5a2719bb517038dfd695e299c91495ddf922d5cf52fb223af.jpg", + "image_caption": [ + "4 views" + ], + "image_footnote": [], + "bbox": [ + 727, + 176, + 784, + 286 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/9da0dea23bbf11da7f27bf2335f933d2b1f86d3f758d49e35bf1e02ff5ee8b20.jpg", + "image_caption": [ + "6 views" + ], + "image_footnote": [], + "bbox": [ + 786, + 176, + 845, + 286 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4. Applications", + "text_level": 1, + "bbox": [ + 499, + 356, + 631, + 371 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Benefiting from separate tracking results of human body and clothing, we can enable interesting applications like cloth retargeting as shown in Figure 8. Since the body mesh and cloth mesh are captured, we can exchange the garments between two person by simulating the garments on the other body. Moreover, free-viewpoint motion can be created by simulating the clothing on human body sequences.", + "bbox": [ + 497, + 377, + 890, + 475 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/d2a02633a71bc2f21ee5194b63edd462afd85e4c557463deea43581dba62d4be.jpg", + "image_caption": [ + "Figure 8. Application of cloth retargeting. From the captured results using our method, we can retarget the garments between the S4 of DeepCap and the 'FranziRed' of DynaCap." + ], + "image_footnote": [], + "bbox": [ + 522, + 483, + 608, + 574 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/04e8d1615ac71b0c66e619a82bed89d58395189f9065a4bb4a2345125442166e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 576, + 484, + 635, + 573 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/a783863e72fc48f9fb39e39b52c50ce68a20bea0a10d9452f9e4c4b6bacce24f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 645, + 484, + 684, + 573 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/d48f036eb9b4af9b10453d785a9649e0b97c099067737d9949a0adef63d2eea0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 694, + 484, + 747, + 573 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/1b1772b24e41c5a1d224b24850638f1b56d14c5ebe4c2dd63961fbe11dd3f3a0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 763, + 484, + 806, + 573 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/7c03eb5680cda026af5bc8f3493b915476c20c3aa1a02398bd4071445cad4686.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 820, + 484, + 870, + 573 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 499, + 634, + 618, + 648 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we proposed a novel performance capture method for clothed humans with a double-layer neural radiance fields (NeRF). Our double-layer NeRFs can effectively model moving humans in daily-life clothes. By matching the rendered color with the observed color, our method estimates the clothing and body motion and temporally-varying appearance robustly. In addition, we learn a differential physics-aware network for different garment types which supports physically-plausible body-cloth interactions and cloth dynamics. The experimental results on real datasets demonstrate the effectiveness of the proposed method.", + "bbox": [ + 497, + 657, + 890, + 809 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments This work was supported by the Fundamental Research Funds for the Central Universities (NJ2022028), and the Natural Science Foundation of China (No.U1713208), and Program for Changjiang Scholars.", + "bbox": [ + 497, + 811, + 890, + 866 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "21105", + "bbox": [ + 463, + 885, + 504, + 897 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 124, + 173, + 138 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] https://www.marvelousdesigner.com/.3,5", + "[2] Thiemo Alldieck, Gerard Pons-Moll, Christian Theobalt, and Marcus Magnor. Tex2Shape: Detailed full human body geometry from a single image. In ICCV, 2019. 2", + "[3] Bharat Lal Bhatnagar, Garvita Tiwari, Christian Theobalt, and Gerard Pons-Moll. Multi-Garment Net: Learning to dress 3D people from images. In ICCV, 2019. 1, 2", + "[4] Mingsong Dou, Sameh Khamis, Yury Degtyarev, Philip Davidson, Sean Fanello, Adarsh Kowdle, Sergio Orts Escolano, Christoph Rhemann, David Kim, Jonathan Taylor, Pushmeet Kohli, Vladimir Tankovich, and Shahram Izadi. Fusion4D: Real-time performance capture of challenging scenes. ACM TOG, 35:1-13, July 2016. 1, 2", + "[5] Stuart Ganan and D McClure. Bayesian image analysis: An application to single photon emission tomography. In Proceedings of the American Statistical Association, pages 12-18, 1985. 5", + "[6] Chen Gao, Ayush Saraf, Johannes Kopf, and Jia-Bin Huang. Dynamic view synthesis from dynamic monocular video. In ICCV, 2021. 4, 7, 8", + "[7] Erhan Gundogdu, Victor Constantin, Amrollah Seifoddini, Minh Dang, Mathieu Salzmann, and Pascal Fua. GarNet: A two-stream network for fast and accurate 3d cloth draping. In ICCV, 2019. 5", + "[8] Marc Habermann, Lingjie Liu, Weipeng Xu, Michael Zollhoefer, Gerard Pons-Moll, and Christian Theobalt. Real-time deep dynamic characters. ACM Transactions on Graphics, 40(4), August 2021. 2, 5, 6, 7", + "[9] Marc Habermann, Weipeng Xu, Michael Zollhoefer, Gerard Pons-Moll, and Christian Theobalt. LiveCap: Real-time human performance capture from monocular video. ACM TOG, 38(2):14:1-14:17, March 2019. 1, 2", + "[10] Marc Habermann, Weipeng Xu, Michael Zollhoefer, Gerard Pons-Moll, and Christian Theobalt. DeepCap: Monocular human performance capture using weak supervision. In CVPR, 2020. 1, 2, 5, 6, 7", + "[11] Boyi Jiang, Yang Hong, Hujun Bao, and Juyong Zhang. SelfRecon: Self reconstruction your digital avatar from monocular video. In CVPR, 2022. 2", + "[12] Boyi Jiang, Juyong Zhang, Yang Hong, Jinhao Luo, Ligang Liu, and Hujun Bao. BCNet: Learning body and cloth shape from a single image. In ECCV, 2020. 1, 2, 6, 7", + "[13] Hanbyul Joo, Tomas Simon, and Yaser Sheikh. Total capture: A 3D deformation model for tracking faces, hands, and bodies. In CVPR, 2018. 4", + "[14] Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In International Conference for Learning Representations, 2015. 6", + "[15] John P Lewis, Matt Cordner, and Nickson Fong. Pose space deformation: A unified approach to shape interpolation and skeleton-driven deformation. In SIGGRAPH, 2000. 3, 4", + "[16] Yue Li, Marc Habermann, Bernhard Thomaszewski, Stelian Coros, Thabo Beeler, and Christian Theobalt. Deep physics-aware inference of cloth deformation for monocular human performance capture. In International Conference on 3D Vision (3DV), 2021. 1, 2, 5, 6" + ], + "bbox": [ + 78, + 147, + 468, + 865 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[17] Lingjie Liu, Marc Habermann, Viktor Rudnev, Kripasindhu Sarkar, Jiatao Gu, and Christian Theobalt. Neural Actor: Neural free-view synthesis of human actors with pose control. ACM TOG, 40(6):1-16, 2021. 3", + "[18] Matthew Loper, Naureen Mahmood, Javier Romero, Gerard Pons-Moll, and Michael J. Black. SMPL: A skinned multiperson linear model. ACM TOG, 34(6):248, 2015. 2, 3", + "[19] Qianli Ma, Jinlong Yang, Anurag Ranjan, Sergi Pujades, Gerard Pons-Moll, Siyu Tang, and Michael J. Black. Learning to dress 3D people in generative clothing. In CVPR, 2020. 2", + "[20] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. N-eRF: Representing scenes as neural radiance fields for view synthesis. In ECCV, 2020. 1, 2, 4", + "[21] Richard Newcombe, Dieter Fox, and Steve Seitz. Dynamic-Fusion: Reconstruction and tracking of non-rigid scenes in real-time. In CVPR, 2015. 1, 2", + "[22] Michael Oechsle, Songyou Peng, and Andreas Geiger. U-NISURF: Unifying neural implicit surfaces and radiance fields for multi-view reconstruction. In ICCV, 2021. 3, 4", + "[23] Chaitanya Patel, Zhouyingcheng Liao, and Gerard Pons-Moll. TailorNet: Predicting clothing in 3D as a function of human pose, shape and garment style. In CVPR, 2020. 2, 5, 6, 7", + "[24] Sida Peng, Junting Dong, Qianqian Wang, Shangzhan Zhang, Qing Shuai, Xiaowei Zhou, and Hujun Bao. Animatable neural radiance fields for modeling dynamic human bodies. In ICCV, 2021. 2, 3, 4, 5", + "[25] Sida Peng, Yuanqing Zhang, Yinghao Xu, Qianqian Wang, Qing Shuai, Hujun Bao, and Xiaowei Zhou. Neural Body: Implicit neural representations with structured latent codes for novel view synthesis of dynamic humans. In CVPR, 2021. 2, 3, 4", + "[26] Gerard Pons-Moll, Sergi Pujades, Sonny Hu, and Michael J. Black. ClothCap: Seamless 4D clothing capture and retargeting. ACM TOG, 36(4), July 2017. 1, 2", + "[27] Shunsuke Saito, Zeng Huang, Ryota Natsume, Shigeo Morishima, Angjoo Kanazawa, and Hao Li. PIFu: Pixel-aligned implicit function for high-resolution clothed human digitization. In ICCV, 2019. 2", + "[28] Shunsuke Saito, Tomas Simon, Jason Saragih, and Hanbyul Joo. PIFuHD: Multi-level pixel-aligned implicit function for high-resolution 3D human digitization. In CVPR, 2020. 2", + "[29] Igor Santesteban, Miguel A. Otaduy, and Dan Casas. Learning-based animation of clothing for virtual try-on. Computer Graphics Forum (Proc. Eurographics), 2019. 2", + "[30] Vincent Sitzmann, Julien NP Martel, Alexander W Bergman, David B Lindell, and Gordon Wetzstein. Implicit neural representations with periodic activation functions. In Advances in Neural Information Processing Systems, 2020. 4, 6", + "[31] Zhaoqi Su, Weilin Wan, Tao Yu, Lingjie Liu, Lu Fang, Wenping Wang, and Yebin Liu. MulayCap: Multi-layer human performance capture using a monocular video camera. IEEE Transactions on Visualization and Computer Graphics, 28(4):1862-1879, 2022. 1, 2" + ], + "bbox": [ + 502, + 127, + 890, + 865 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "21106", + "bbox": [ + 463, + 887, + 504, + 897 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[32] Robert W. Sumner, Johannes Schmid, and Mark Pauly. Embedded deformation for shape manipulation. ACM TOG, 26(3), 2007. 3, 5", + "[33] Gül Varol, Duygu Ceylan, Bryan Russell, Jimei Yang, Ersin Yumer, Ivan Laptev, and Cordelia Schmid. BodyNet: Volumetric inference of 3D human body shapes. In ECCV, 2018. 2", + "[34] Kangkan Wang, Sida Peng, Xiaowei Zhou, Jian Yang, and Guofeng Zhang. NerfCap: Human performance capture with dynamic neural radiance fields. IEEE Transactions on Visualization and Computer Graphics, 2022. 1, 3, 4, 5, 6, 7", + "[35] Olivia Wiles, Georgia Gkioxari, Richard Szeliski, and Justin Johnson. Synsin: End-to-end view synthesis from a single image. In CVPR, pages 7467-7477, 2020. 5", + "[36] Yuliang Xiu, Jinlong Yang, Dimitrios Tzionas, and Michael J. Black. ICON: Implicit clothed humans obtained from normals. In CVPR, June 2022. 2, 6, 7", + "[37] Weipeng Xu, Avishek Chatterjee, Michael Zollhöfer, Helge Rhodin, Dushyant Mehta, Hans-Peter Seidel, and Christian Theobalt. MonoPerfCap: Human performance capture from monocular video. ACM TOG, 37(2):1-15, July 2018. 1, 2", + "[38] Tao Yu, Kaiwen Guo, Feng Xu, Yuan Dong, Zhaqi Su, Jianhui Zhao, Jianguo Li, Qionghai Dai, and Yebin Liu. Body-Fusion: Real-time capture of human motion and surface geometry using a single depth camera. In ICCV, October 2017. 1, 2", + "[39] Tao Yu, Zerong Zheng, Kaiwen Guo, Jianhui Zhao, Qionghai Dai, Hao Li, Gerard Pons-Moll, and Yebin Liu. DoubleFusion: Real-time capture of human performances with inner body shapes from a single depth sensor. In CVPR, June 2018. 1, 2", + "[40] Tao Yu, Zerong Zheng, Yuan Zhong, Jianhui Zhao, Qionghai Dai, Gerard Pons-Moll, and Yebin Liu. SimulCap: Single-view human performance capture with cloth simulation. In CVPR, 2019. 1, 2, 5", + "[41] Fuqiang Zhao, Wei Yang, Jiakai Zhang, Pei Lin, Yingliang Zhang, Jingyi Yu, and Lan Xu. HumanNeRF: Generalizable neural human radiance field from sparse inputs. In CVPR, 2022. 1, 3", + "[42] Zerong Zheng, Tao Yu, Yixuan Wei, Qionghai Dai, and Yebin Liu. DeepHuman: 3D human reconstruction from a single image. In ICCV, 2019. 2", + "[43] Heming Zhu, Yu Cao, Hang Jin, Weikai Chen, Dong Du, Zhangye Wang, Shuguang Cui, and Xiaoguang Han. Deep Fashion3D: A dataset and benchmark for 3d garment reconstruction from single images. In ECCV, pages 512-530, 2020. 1, 2", + "[44] Heming Zhu, Lingteng Qiu, Yuda Qiu, and Xiaoguang Han. Registering Explicit to Implicit: Towards high-fidelity garment mesh reconstruction from single images. In CVPR, pages 3845-3854, 2022. 1, 2", + "[45] Hao Zhu, Xinxin Zuo, Sen Wang, Xun Cao, and Ruigang Yang. Detailed human shape estimation from a single image by hierarchical mesh deformation. In CVPR, 2019. 2" + ], + "bbox": [ + 78, + 125, + 468, + 825 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "21107", + "bbox": [ + 465, + 885, + 504, + 897 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2023/Clothed Human Performance Capture With a Double-Layer Neural Radiance Fields/edeeca28-ac72-40e3-9181-9737d0c908f2_model.json b/2023/Clothed Human Performance Capture With a Double-Layer Neural Radiance Fields/edeeca28-ac72-40e3-9181-9737d0c908f2_model.json new file mode 100644 index 0000000000000000000000000000000000000000..dc22a1708c8257d28b8a8ff37a9805900d8c5220 --- /dev/null +++ b/2023/Clothed Human Performance Capture With a Double-Layer Neural Radiance Fields/edeeca28-ac72-40e3-9181-9737d0c908f2_model.json @@ -0,0 +1,2695 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.13, + 0.051, + 0.201, + 0.086 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.253, + 0.049, + 0.793, + 0.088 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.102, + 0.162, + 0.87, + 0.201 + ], + "angle": 0, + "content": "Clothed Human Performance Capture with a Double-layer Neural Radiance Fields" + }, + { + "type": "text", + "bbox": [ + 0.232, + 0.229, + 0.736, + 0.248 + ], + "angle": 0, + "content": "Kangkan Wang\\(^{1,2}\\), Guofeng Zhang\\(^{3}\\), Suxu Cong\\(^{1}\\), Jian Yang\\(^{1,2}\\)" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.251, + 0.935, + 0.284 + ], + "angle": 0, + "content": "1Key Lab of Intelligent Perception and Systems for High-Dimensional Information of Ministry of Education, 2Jiangsu Key Lab of Image and Video Understanding for Social Security," + }, + { + "type": "text", + "bbox": [ + 0.109, + 0.285, + 0.905, + 0.318 + ], + "angle": 0, + "content": "School of Computer Science and Engineering, Nanjing University of Science and Technology, China \n3State Key Laboratory of CAD&CG, Zhejiang University, China" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.342, + 0.314, + 0.356 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.371, + 0.473, + 0.718 + ], + "angle": 0, + "content": "This paper addresses the challenge of capturing performance for the clothed humans from sparse-view or monocular videos. Previous methods capture the performance of full humans with a personalized template or recover the garments from a single frame with static human poses. However, it is inconvenient to extract cloth semantics and capture clothing motion with one-piece template, while single frame-based methods may suffer from instable tracking across videos. To address these problems, we propose a novel method for human performance capture by tracking clothing and human body motion separately with a double-layer neural radiance fields (NeRFs). Specifically, we propose a double-layer NeRFs for the body and garments, and track the densely deforming template of the clothing and body by jointly optimizing the deformation fields and the canonical double-layer NeRFs. In the optimization, we introduce a physics-aware cloth simulation network which can help generate physically plausible cloth dynamics and body-cloth interactions. Compared with existing methods, our method is fully differentiable and can capture both the body and clothing motion robustly from dynamic videos. Also, our method represents the clothing with an independent NeRFs, allowing us to model implicit fields of general clothes feasibly. The experimental evaluations validate its effectiveness on real multi-view or monocular videos." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.736, + 0.21, + 0.75 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.759, + 0.471, + 0.844 + ], + "angle": 0, + "content": "Performance capture for clothed humans is one of the essential problems in the metaverse, and it not only captures the inner human body motion but also recovers the outer clothing motion which has many promising applications such as virtual try-on, video editing, and telepresence. From sparse-view or monocular videos of a moving human" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.343, + 0.893, + 0.455 + ], + "angle": 0, + "content": "in general clothes, its goal is to recover the dynamic 3D shape sequence of the human body and clothing simultaneously that are consistent with the observed frames in both human shape and motion. This is a very challenging problem since the dynamic human could be with arbitrary motions and with complex non-rigid cloth deformations, and the clothing in motion is difficult to maintain physically plausible interactions with the body." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.459, + 0.893, + 0.751 + ], + "angle": 0, + "content": "Previous systems [4, 21, 38, 39] reconstruct 3D clothed humans by using depth sensors or fitting a personalized template [9, 10, 37] to the image observations (e.g., body joints and silhouettes). Only recovering one-piece geometry which unifies the human body and clothes, these systems fail to track the motion of the clothing and achieve clothing editing on 3D humans, which are the prerequisites in many VR/AR applications like virtual dressing. On the contrary, cloth can be extracted and tracked from depth scans [26, 40] accurately by fitting pre-built cloth templates to the scans which have limited applications when 3D data are unavailable. Existing garment estimation methods [3, 12, 43, 44] from color images require the person facing the camera and in static poses. When the human is moving and the clothing is deforming, these methods may recover the 3D garments unreliable. Recent methods [16, 31] track body and clothing motion simultaneously from videos, but they need to re-build cloth template for a new performer and the running efficiency is very low due to online cloth simulation or computationally-exhaustive optimization, which prohibits them from being widely deployed for daily applications." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.756, + 0.894, + 0.868 + ], + "angle": 0, + "content": "Recent works [20, 34, 41] adopt dynamic human NeRFs to capture human motion and obtain impressive tracking results. By capturing the temporally-varying appearance in the videos, dynamic NeRFs [34] can provide dense photometric constraints to track the deforming geometry of the performer. However, they represent the human with a single NeRFs without modeling cloth, and the clothing motion cannot be extracted. In this paper, we aim to track the cloth" + }, + { + "type": "page_footnote", + "bbox": [ + 0.094, + 0.854, + 0.378, + 0.868 + ], + "angle": 0, + "content": "*Corresponding author: wangkangkan@njust.edu.cn" + }, + { + "type": "page_number", + "bbox": [ + 0.465, + 0.887, + 0.506, + 0.898 + ], + "angle": 0, + "content": "21098" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.126, + 0.469, + 0.209 + ], + "angle": 0, + "content": "ing and body motion simultaneously with dynamic NeRFs. However, this problem is rather challenging due to two major questions we need to solve: how to represent dynamic clothing and human body with NeRFs, and how to capture clothing and human body motion with plausible body-cloth interactions based on the implicit representation." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.21, + 0.471, + 0.61 + ], + "angle": 0, + "content": "In this paper, we propose a novel method for clothed human performance capture with a double-layer NeRFs. Specifically, a double-layer NeRFs is modeled for both the body and clothing in the template space, and transformed to the observation space with the corresponding deformation fields, and the rendered images are then synthesized by composing the two dynamic NeRFs. We first estimate the template shape in canonical frame and learn the geometry network supervised by the template geometry. In the rendering, we compose the double-layer NeRFs with the guidance of deformed body and clothing meshes. Then, by minimizing the difference between the rendered color and observed color, the deformation fields and the canonical N-NeRFs are optimized jointly. The deformation field is represented as the inverse deformation of the template mesh, thus the densely deforming geometry of the template can be recovered simultaneously. In addition, we adopt a physics-aware network learnt from simulation data between various cloth types and humans to constrain the dynamic clothing and preserve physically plausible body-cloth interactions, resulting in realistic cloth geometry tracking. Compared to previous methods, our method is fully differentiable and can recover the realistic motion of both the clothing and body from dynamic videos with arbitrary human poses and complex cloth deformations. The experimental qualitative and quantitative results on datasets of DynaCap [8] and DeepCap [10] prove that the proposed approach can robustly and accurately capture the motion for clothed humans. In summary, the primary contributions of our work include:" + }, + { + "type": "text", + "bbox": [ + 0.094, + 0.616, + 0.469, + 0.657 + ], + "angle": 0, + "content": "- We propose a double-layer NeRFs for dynamic humans in general clothing, allowing us to model implicit humans with a variety of clothes (e.g., loose dresses)." + }, + { + "type": "text", + "bbox": [ + 0.094, + 0.664, + 0.469, + 0.733 + ], + "angle": 0, + "content": "- To the best of our knowledge, we propose the first framework to capture clothing motion separately from the human body using the double-layer NeRFs, which provides dense appearance constraints on the geometry tracking and improves the robustness and accuracy." + }, + { + "type": "text", + "bbox": [ + 0.094, + 0.741, + 0.469, + 0.782 + ], + "angle": 0, + "content": "- A differentiable physics-aware network is learnt for d-different common garments and used to preserve physically plausible cloth deformations in motion capture." + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.616, + 0.469, + 0.782 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.793, + 0.219, + 0.807 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.811, + 0.469, + 0.867 + ], + "angle": 0, + "content": "Clothed human reconstruction. Depth-based methods [4,21,38,39] reconstruct 3D humans in realtime by fusing the depth maps in a volumetric way. However, depth-based methods cannot synthesize photo-realistic videos and" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.126, + 0.892, + 0.389 + ], + "angle": 0, + "content": "have constrained applications in daily life due to data capture limitations, e.g., high power consumption and sensitive to sunlight. Based on deep learning techniques, some approaches predict 3D humans from color images through implicit function [27, 28], coarse-to-fine mesh deformation [45], surface normal optimization [36, 42], or UV mapping [2]. Although these methods can reconstruct human details, the recovered shapes do not contain human semantics like human pose, shape and clothing. To recover the clothing, some methods [3, 19] add displacements on vertices of SMPL model [18] which only support tight clothing and recover coarse-level geometry. Template-based methods [9, 10, 37] recover 3D human motions by non-rigidly fitting a person-specific template mesh to 2D observed features (e.g., human joints or silhouettes). With a personalized template, these methods obtain impressive performance capture. However, the recovered meshes integrate the body and clothing in one piece, which hinders their applications like garment retargeting and virtual trying-on." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.393, + 0.892, + 0.711 + ], + "angle": 0, + "content": "Clothing motion capture. Traditional methods [26, 40] capture the cloth motion by precisely fitting a cloth template to 3D data from depth sensors, which cannot generalize to the video-based setting and are limited to applications in indoor scenes. Deep learning-based approaches [3, 12, 43, 44] estimate the 3D garment shape from color images. However, these methods require the human facing the camera with a static pose and do not try temporally tracking of dynamic garments. Some works [23, 29] predict the cloth shape conditioned on the human pose and shape but the shapes are not consistent with the observed data. Recent works [16, 31] capture both human body and clothing motion from dynamic videos. MulayCap [31] optimizes a simulated garment on SMPL model with semantic cloth segmentation, but their traditional simulation and optimization are both computationally expensive. DeepCloth [16] incorporates a physics-based simulator to supervise the cloth deformations estimated with multi-view joints and silhouettes as DeepCap [10]. However, their simulation layer is not differentiable and needs to adjust the cloth material types online, which is a tedious and time consuming step. Moreover, with the same limitation as DeepCap, they cannot recover the detailed cloth deformation based on the sparse features." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.714, + 0.892, + 0.867 + ], + "angle": 0, + "content": "Implicit Human Reconstruction. Clothed humans can be reconstructed through implicit representation-based methods such as voxel representation [33, 42], implicit function [27, 28], or neural radiance fields (NeRFs) [20, 24, 25]. Recent works combine explicit and implicit representations to produce high-quality avatar [11] or reconstruct high-fidelity garment mesh [44], which show the advantage of implicit representation on recovering geometry details. However, these methods are limited to self-rotating human motion or a single color image. Neuralbody [25] and AniNErf [24] extend NeRFs for a dynamic human with defor" + }, + { + "type": "page_number", + "bbox": [ + 0.465, + 0.887, + 0.505, + 0.898 + ], + "angle": 0, + "content": "21099" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.102, + 0.123, + 0.87, + 0.407 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.413, + 0.893, + 0.465 + ], + "angle": 0, + "content": "Figure 1. Overview of the proposed approach. Our approach can capture both the clothing and human body motion, and achieve novel-view synthesis from sparse-view or monocular videos. We learn a double-layer neural radiance fields (NeRFs) for the clothing and human body in canonical frame, and jointly optimize the canonical double-layer NeRFs and the human deformation constrained with a physics-aware simulation network. Please refer to Sec. 3 for details of our method." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.473, + 0.473, + 0.64 + ], + "angle": 0, + "content": "mation fields based on the linear blend skinning (LBS) [15]. Some human NeRFs methods [17, 41] model nonrigid human deformations with a point displacement on top of LBS. However, without using any clothing motion prior, these methods cannot effectively recover the motion of humans in loose clothes, and the geometry of clothes is not reconstructed reliably under nonrigid cloth deformations. The recent work, NerfCap [34] incorporates the embedded graph prior into the human NeRFs and represents nonlinear surface deformations for loose clothes successfully. However, their method does not model the clothing separately and thus fails to track the clothing motion." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.648, + 0.269, + 0.664 + ], + "angle": 0, + "content": "3. Proposed Approach" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.673, + 0.473, + 0.868 + ], + "angle": 0, + "content": "Given sparse-view or monocular videos of a clothed human, we track the motion of both the human body and clothing and synthesize novel-view videos of the human with a novel performance capture method based on a double-layer neural radiance fields (NeRFs). Figure 1 shows an overview of the proposed method. The double-layer NeRFs is learnt for the human body and clothing in canonical frame (Sec. 3.2). Then, the canonical NeRFs is dynamically transformed into the observed frame with inverse deformation fields of the deformed template for clothing and body respectively. From the dynamic double-layer NeRFs, we render the images using a composite rendering. By matching the synthesized images with the video frames and constraining the cloth deformations with a physics-aware simulation" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.473, + 0.892, + 0.502 + ], + "angle": 0, + "content": "network, we estimate the double-layer NeRFs and the human deformation jointly (Sec. 3.3)." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.51, + 0.73, + 0.527 + ], + "angle": 0, + "content": "3.1. Canonical Human Shape" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.532, + 0.892, + 0.685 + ], + "angle": 0, + "content": "We decompose the clothed human into undressed body and garments (like T-shirt and pants), and represent them with SMPL model [18] and a parametric PCA model respectively. To include the human details (e.g., the head and hand), we obtain the canonical body model \\(M\\) by deforming the SMPL model to a personalized template with non-rigid deformation [32]. We design different styles (size and sleeve length, et al.) for each garment category, and simulate these garments on the template SMPL model using Marvelous Designer [1]. Then, we obtain a linear model through PCA on the simulated meshes," + }, + { + "type": "equation", + "bbox": [ + 0.625, + 0.696, + 0.891, + 0.713 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {G} (\\alpha) = \\bar {\\boldsymbol {G}} + \\boldsymbol {B} _ {g} \\alpha , \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.723, + 0.892, + 0.752 + ], + "angle": 0, + "content": "where \\(\\bar{G}\\) is the mean garment shape, \\(B_{g}\\) denotes the linear basis matrix, and \\(\\alpha\\) are parameters of the cloth shape \\(G(\\alpha)\\)." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.761, + 0.863, + 0.777 + ], + "angle": 0, + "content": "3.2. Double-layer NeRFs for Dynamic Humans" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.783, + 0.892, + 0.868 + ], + "angle": 0, + "content": "Unlike a single NeRFs for clothed humans [17,24,25,34, 41], we represent the clothing with an independent NeRFs on the body, which forms a double-layer NeRFs. The implicit representation can be extended to multiple garments. Canonical NeRF. We use an occupancy network [22] to represent canonical geometry since the occupancy can mod" + }, + { + "type": "page_number", + "bbox": [ + 0.465, + 0.887, + 0.506, + 0.898 + ], + "angle": 0, + "content": "21100" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.127, + 0.396, + 0.14 + ], + "angle": 0, + "content": "el geometry more precisely than density [20] as," + }, + { + "type": "equation", + "bbox": [ + 0.205, + 0.15, + 0.469, + 0.166 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {o} (\\boldsymbol {x}) = \\boldsymbol {F} _ {o} \\left(\\gamma_ {\\boldsymbol {x}} (\\boldsymbol {x})\\right), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.175, + 0.469, + 0.203 + ], + "angle": 0, + "content": "where \\(\\gamma_{\\pmb{x}}\\) is the positional encoding [20]. The color network \\(F_{c}\\) in canonical frame is formulated as," + }, + { + "type": "equation", + "bbox": [ + 0.191, + 0.213, + 0.469, + 0.229 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {c} _ {i} (\\boldsymbol {x}) = \\boldsymbol {F} _ {c} \\left(\\gamma_ {\\boldsymbol {x}} (\\boldsymbol {x}), \\varphi_ {i}\\right), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.238, + 0.469, + 0.279 + ], + "angle": 0, + "content": "where \\(\\varphi_{i}\\) is an appearance latent code [24, 25] for frame \\(i\\). We define occupancy network \\(\\{F_o^b,F_o^g\\}\\), and color network \\(\\{F_c^b,\\varphi^b,F_c^g,\\varphi^g\\}\\) for the body and clothing, respectively." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.28, + 0.469, + 0.307 + ], + "angle": 0, + "content": "To facilitate the learning of occupancy network, we incorporate a signed distance fields (SDFs) network \\( S \\)," + }, + { + "type": "equation", + "bbox": [ + 0.198, + 0.316, + 0.469, + 0.332 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {S}: \\boldsymbol {p} \\in \\mathbb {R} ^ {3} \\rightarrow \\boldsymbol {s} \\in \\mathbb {R}. \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.341, + 0.469, + 0.37 + ], + "angle": 0, + "content": "which predicts the SDF value \\( s \\) for point \\( p \\). We use the losses proposed in [30] to learn the SDFs of the shape," + }, + { + "type": "equation", + "bbox": [ + 0.078, + 0.378, + 0.487, + 0.442 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} L _ {s d f} = \\sum_ {\\boldsymbol {p} \\in \\Phi} | \\boldsymbol {S} (\\boldsymbol {p}) | + \\sum_ {\\boldsymbol {p} \\in \\Phi} (1 - \\langle \\nabla \\boldsymbol {S} (\\boldsymbol {p}), \\bar {\\boldsymbol {n}}) \\rangle \\tag {5} \\\\ + \\sum_ {\\boldsymbol {p} \\in \\Omega} | \\| \\nabla \\boldsymbol {S} (\\boldsymbol {p}) \\| _ {2} - 1 | + \\sum_ {\\boldsymbol {p} \\in \\Omega \\backslash \\Phi} \\exp (- \\delta \\cdot | \\boldsymbol {S} (\\boldsymbol {p}) |), \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.452, + 0.469, + 0.563 + ], + "angle": 0, + "content": "where \\(\\bar{n}\\) is the surface normal, \\(\\nabla S(\\pmb{p})\\) indicates the gradient of the SDFs, \\(\\Omega\\) and \\(\\Phi\\) denotes the 3D space and shape surface respectively. The first and second terms are used to estimate valid SDFs and normals on the surfaces, respectively. The third term constrains the norm of SDF gradients to be 1, and the last term penalizes SDF values of non-surface points near 0 with \\(\\delta \\gg 1\\). We define SDF network \\(\\{\\pmb{S}^b, \\pmb{S}^g\\}\\) for both the body and clothing." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.564, + 0.469, + 0.591 + ], + "angle": 0, + "content": "Supervised by the predictions of SDF network, the occupancy network is learnt with the following loss," + }, + { + "type": "equation", + "bbox": [ + 0.166, + 0.6, + 0.469, + 0.629 + ], + "angle": 0, + "content": "\\[\nL _ {o c c} = \\sum_ {\\boldsymbol {x} \\in \\Omega} L _ {C} \\left(\\boldsymbol {F} _ {\\boldsymbol {o}} \\left(\\gamma_ {\\boldsymbol {x}} (\\boldsymbol {x})\\right), \\boldsymbol {o}\\right), \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.638, + 0.469, + 0.735 + ], + "angle": 0, + "content": "where \\( \\pmb{o} \\) is the occupancy indicated by SDFs (if \\( S(\\pmb{x}) < = 0 \\), \\( o = 1 \\); otherwise, \\( o = 0 \\)), and \\( L_{C} \\) denotes the cross-entropy loss. We adopt the compositional design of SDF and occupancy because it is better for volumetric learning with continuous occupancy fields [22] compared to discrete labels extracted from SDFs. The template is varying during learning geometry model, and the loss is formulated as," + }, + { + "type": "equation", + "bbox": [ + 0.185, + 0.746, + 0.469, + 0.761 + ], + "angle": 0, + "content": "\\[\nL _ {g e o} = \\mu_ {1} L _ {s d f} + \\mu_ {2} L _ {o c c}. \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.77, + 0.469, + 0.837 + ], + "angle": 0, + "content": "where \\(\\mu_{1}\\) and \\(\\mu_{2}\\) are the weight for each term. Through the optimization, the NeRF geometry is consistent with the estimated template so that densely deforming geometry of the template can be tracked by utilizing dense photometric constraints from NeRF color model." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.839, + 0.469, + 0.867 + ], + "angle": 0, + "content": "Deformation Model. Following [34], we transform sampled points in observed space to canonical space with the" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.127, + 0.892, + 0.306 + ], + "angle": 0, + "content": "inverse deformation of the nearest vertex on the deformed template. In this way, the human deformation model is disentangled into the non-rigid surface deformation (i.e., embedded graph deformation as in [34]) and the articulated skeletal motion. For the clothing and body, we define the deformation network \\( F_{d}^{g} \\) with latent codes \\( \\omega^{g} \\) and \\( F_{d}^{b} \\) with latent codes \\( \\omega^{b} \\), respectively. Through nonrigid deformation, we obtain the clothing mesh \\( \\hat{G} \\) and body mesh \\( \\hat{M} \\). Then, the linear blend skinning [15] is applied to obtain the final cloth mesh \\( \\tilde{G} \\) and body mesh \\( \\tilde{M} \\) in the observation space. We recover the SMPL model with [13] for input video frames and also estimate a canonical body shape." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.308, + 0.892, + 0.473 + ], + "angle": 0, + "content": "Composite Rendering. In [6], a blending weight is predicted for each sampled point to blend the static background and dynamic target in a scene. This blending method cannot separate the clothing and human body in the NeRFs effectively. We propose an efficient composite rendering strategy for the dynamic double-layer NeRF. Specifically, for a casted ray \\( r \\), we find the intersecting triangle with the deformed meshes. If the intersected triangle is from the body mesh, we set the mask \\( m_r^b \\) of all sampled points on \\( r \\) to 1; otherwise, \\( m_r^b \\) is set to 0. Similarly, if the intersected triangle is from the cloth mesh, we set the mask \\( m_r^g \\) to 1; otherwise, \\( m_r^g \\) is set to 0. The complete human is rendered as," + }, + { + "type": "equation", + "bbox": [ + 0.544, + 0.486, + 0.891, + 0.525 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {T} _ {r} ^ {i} = \\prod_ {j = 1} ^ {i - 1} \\left(1 - \\boldsymbol {\\sigma} _ {g} ^ {j} \\left(1 - \\boldsymbol {m} _ {r} ^ {b}\\right)\\right) \\left(1 - \\boldsymbol {\\sigma} _ {b} ^ {j} \\boldsymbol {m} _ {r} ^ {b}\\right), \\tag {8}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.543, + 0.528, + 0.848, + 0.563 + ], + "angle": 0, + "content": "\\[\n\\hat {\\boldsymbol {C}} _ {r} = \\sum_ {i = 1} ^ {n} \\boldsymbol {T} _ {r} ^ {i} \\left(\\boldsymbol {o} _ {g} ^ {i} \\left(1 - \\boldsymbol {m} _ {r} ^ {b}\\right) \\boldsymbol {c} _ {g} ^ {i} + \\boldsymbol {o} _ {b} ^ {i} \\boldsymbol {m} _ {r} ^ {b} \\boldsymbol {c} _ {b} ^ {i}\\right),\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.575, + 0.891, + 0.632 + ], + "angle": 0, + "content": "where \\(\\{o_g,o_b\\}\\) and \\(\\{c_{g},c_{b}\\}\\) denote the point occupancy and color respectively, which are defined for both clothing and body. \\(T_r^i\\) is the accumulated transparency along the ray. We also render the clothing color \\(\\hat{C}_r^g\\) and body color \\(\\hat{C}_r^b\\)" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.643, + 0.89, + 0.671 + ], + "angle": 0, + "content": "3.3. Clothed Human Performance Capture with Double-layer NeRFs" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.679, + 0.891, + 0.72 + ], + "angle": 0, + "content": "We estimate the double-layer NeRFs and template deformation of the body and clothing by minimizing the following loss between the synthesized color and observed color:" + }, + { + "type": "equation", + "bbox": [ + 0.537, + 0.733, + 0.891, + 0.827 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} L _ {r g b} = \\sum_ {\\boldsymbol {r} \\in R} \\| \\hat {\\boldsymbol {C}} (\\boldsymbol {r}) - \\boldsymbol {C} (\\boldsymbol {r}) \\| _ {2} ^ {2} \\tag {9} \\\\ + \\sum_ {\\boldsymbol {r} \\in \\boldsymbol {R}} \\| (\\hat {\\boldsymbol {C}} (\\boldsymbol {r}) ^ {b} - \\boldsymbol {C} (\\boldsymbol {r})) (1 - \\boldsymbol {m} _ {r} ^ {g}) \\| _ {2} ^ {2} \\\\ + \\sum_ {\\boldsymbol {r} \\in \\boldsymbol {R}} \\| (\\hat {\\boldsymbol {C}} (\\boldsymbol {r}) ^ {g} - \\boldsymbol {C} (\\boldsymbol {r})) (1 - \\boldsymbol {m} _ {r} ^ {b}) \\| _ {2} ^ {2}, \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.839, + 0.891, + 0.867 + ], + "angle": 0, + "content": "where \\(R\\) is the set of casted rays. While using the dense photometric constraint defined in Eq. 9, the model training" + }, + { + "type": "page_number", + "bbox": [ + 0.465, + 0.887, + 0.503, + 0.898 + ], + "angle": 0, + "content": "21101" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.127, + 0.47, + 0.168 + ], + "angle": 0, + "content": "remains ill-posed and body-cloth collisions are prone to occur. To solve these problems, we design several regularization losses to constrain the joint optimization." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.17, + 0.473, + 0.392 + ], + "angle": 0, + "content": "Physics-aware constraint. Cloth simulation is adopted in previous works [16, 40] to obtain realistic performance capture. Different from their traditional simulation method, we introduce a differential network that learns a mapping from human and cloth parameters to physics-aware cloth shapes. We simulate the data in Marvelous Designer using the method of [23]. For each garment category, we use 25 garment styles \\(\\gamma \\in \\mathbb{R}^4\\) as [23]. To account for the effects of different materials, we simulate 8 different materials in Marvelous Designer, and represent the material type \\(\\pmb{\\tau}\\) using one-hot labels. We train the simulation network \\(D_{\\phi}\\) using a MLP to minimize the L1-loss between the simulated cloth meshes and the predicted ones \\(G_{s} = D_{\\phi}(\\theta ,\\beta ,\\gamma ,\\tau)\\) with the input of human pose and shape, and garment and material styles. The physical simulation loss enforces the deformed cloth mesh \\(\\tilde{G}\\) to be close to \\(G_{s}\\)," + }, + { + "type": "equation", + "bbox": [ + 0.174, + 0.403, + 0.47, + 0.433 + ], + "angle": 0, + "content": "\\[\nL _ {s i m} = \\sum_ {t \\in \\boldsymbol {G}} \\rho \\left(\\left\\| \\tilde {\\boldsymbol {G}} ^ {t} - \\boldsymbol {G} _ {s} ^ {t} \\right\\| _ {2} ^ {2}\\right), \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.445, + 0.47, + 0.499 + ], + "angle": 0, + "content": "where \\( t \\) denotes vertex index of \\( G \\), and \\( \\rho \\) is the Geman-McClure robust function [5]. The simulation loss can penalize body-cloth collisions and preserve physically plausible deformations." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.502, + 0.469, + 0.559 + ], + "angle": 0, + "content": "As-rigid-as-possible Loss. To make sure the template deforms smoothly, we utilize an as-rigid-as-possible loss [32] on the embedded graph deformation of the human body and clothing, \\( L_{arap} = L_{arap}(\\hat{M}) + L_{arap}(\\hat{G}) \\)." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.559, + 0.469, + 0.614 + ], + "angle": 0, + "content": "Mask Loss. We render the mask image \\( R(\\mathbf{U}) \\) of the combined mesh \\( \\mathbf{U} = (M, G) \\) by using a differentiable renderer [35] based on point cloud, and compute the mask loss with the input human mask image \\( \\bar{R} \\):" + }, + { + "type": "equation", + "bbox": [ + 0.174, + 0.627, + 0.47, + 0.645 + ], + "angle": 0, + "content": "\\[\nL _ {\\text {m a s k}} (\\boldsymbol {U}) = \\left\\| R (\\boldsymbol {U}) - \\bar {R} \\right\\| _ {2} ^ {2}. \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.657, + 0.469, + 0.699 + ], + "angle": 0, + "content": "We utilize the mask loss on both the deformed mesh and the simulated mesh as, \\( L_{mask} = L_{mask}(\\tilde{M},\\tilde{G}) + L_{mask}(\\tilde{M},G_s) \\)." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.701, + 0.469, + 0.755 + ], + "angle": 0, + "content": "Attachment Loss. During human motion, some cloth parts are always attached to the body (e.g., the dress strap is fixed on the shoulder). To ensure this constraint, we define an attachment loss:" + }, + { + "type": "equation", + "bbox": [ + 0.16, + 0.767, + 0.469, + 0.797 + ], + "angle": 0, + "content": "\\[\nL _ {\\text {a t t a c h}} = \\sum_ {i, j \\in A} \\rho \\left(\\| \\tilde {\\boldsymbol {G}} ^ {i} - \\tilde {\\boldsymbol {M}} ^ {j} \\| _ {2} ^ {2}\\right), \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.81, + 0.469, + 0.838 + ], + "angle": 0, + "content": "where \\(A\\) is a set of selected vertices on the clothing mesh attached to corresponding points of the body mesh." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.839, + 0.469, + 0.867 + ], + "angle": 0, + "content": "Interpenetration Loss. Interpenetrations may occur between the predicted garments and the body. We use the" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.127, + 0.831, + 0.14 + ], + "angle": 0, + "content": "interpenetration term [7] to alleviate this problem:" + }, + { + "type": "equation", + "bbox": [ + 0.504, + 0.148, + 0.892, + 0.193 + ], + "angle": 0, + "content": "\\[\nL _ {i n t e r p} \\left(M _ {1}, M _ {2}\\right) = \\frac {1}{| C |} \\sum_ {i, j \\in C} R e L U \\left(\\left(M _ {1} ^ {i} - M _ {2} ^ {j}\\right) \\cdot N _ {1} ^ {i}\\right), \\tag {13}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.194, + 0.892, + 0.25 + ], + "angle": 0, + "content": "where \\(C\\) is the set of corresponding points between the mesh \\(M_{1}\\) and mesh \\(M_2\\), \\(|C|\\) denotes the number of corresponding points, and \\(N_{1}\\) is the normal of mesh \\(M_{1}\\). We use interpenetration loss in both canonical and posed space," + }, + { + "type": "equation", + "bbox": [ + 0.529, + 0.257, + 0.892, + 0.274 + ], + "angle": 0, + "content": "\\[\nL _ {i n t e r p} = L _ {i n t e r p} (\\hat {\\boldsymbol {M}}, \\hat {\\boldsymbol {G}}) + L _ {i n t e r p} (\\tilde {\\boldsymbol {M}}, \\tilde {\\boldsymbol {G}}). \\tag {14}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.289, + 0.692, + 0.301 + ], + "angle": 0, + "content": "3.3.1 Total Loss Function" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.31, + 0.892, + 0.407 + ], + "angle": 0, + "content": "We first learn the geometry network of both the clothing and human body in canonical frame. We use the first frame of multiple views or a video in self-rotating motion to estimate the canonical geometry by optimizing all the losses. Then, by fixing the clothing shape and the geometry network, we track the motion of each video frame and estimate the appearance of NeRFs using the following total loss:" + }, + { + "type": "equation", + "bbox": [ + 0.547, + 0.416, + 0.891, + 0.449 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} L = \\lambda_ {1} L _ {r g b} + \\lambda_ {2} L _ {a r a p} + \\lambda_ {3} L _ {s i m} \\tag {15} \\\\ + \\lambda_ {4} L _ {\\text {m a s k}} + \\lambda_ {5} L _ {\\text {a t t a c h}} + \\lambda_ {6} L _ {\\text {i n t e r p}}, \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.456, + 0.89, + 0.484 + ], + "angle": 0, + "content": "where \\(\\{\\lambda_1\\ldots \\lambda_6\\}\\) are the balancing weights. For a video with \\(N\\) frames, the estimated parameter set \\(X\\) is:" + }, + { + "type": "equation", + "bbox": [ + 0.501, + 0.492, + 0.89, + 0.521 + ], + "angle": 0, + "content": "\\[\nX = \\left\\{\\boldsymbol {F} _ {c} ^ {b}, \\boldsymbol {F} _ {c} ^ {g}, \\boldsymbol {F} _ {d} ^ {b}, \\boldsymbol {F} _ {d} ^ {g} \\right\\} \\cup \\left\\{\\varphi_ {i} ^ {b}, \\varphi_ {i} ^ {g}, \\omega_ {i} ^ {b}, \\omega_ {i} ^ {g} \\mid i \\in 1, \\dots , N \\right\\}, \\tag {16}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.521, + 0.89, + 0.562 + ], + "angle": 0, + "content": "which includes the weights of the appearance and deformation networks and the corresponding conditional variables for each frame in the video." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.575, + 0.632, + 0.59 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.597, + 0.714, + 0.612 + ], + "angle": 0, + "content": "4.1. Implementation details" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.618, + 0.892, + 0.868 + ], + "angle": 0, + "content": "Datasets and metrics. We design common garments (e.g., dress, T-shirt, skirt, pants, et al.) and simulate the data in the Marvelous Designer [1] under various human shapes and poses and cloth materials. We test our method on publicly datasets, DynaCap [8], and DeepCap [10]. We use four cameras around the human for training and the remaining cameras for test. Figure 2 shows some recovered results using our approach on different clothed humans. Our method is also tested on a monocular video in outdoor scene. Please refer to the supplementary video for more experimental results. The geometry accuracy is evaluated with the intersection over union (IoU) between the ground truth human masks and 2D projections of the recovered human shapes. Following [10, 34], the IoU (\\%) is computed on all views (AMVIoU), all views except the input view (RVIoU), and the input view (SVIoU). The novel-view synthesis quality is evaluated with two metrics, i.e., peak signal-to-noise ratio (PSNR) and structural similarity index (SSIM) as [24, 34]." + }, + { + "type": "page_number", + "bbox": [ + 0.465, + 0.887, + 0.505, + 0.898 + ], + "angle": 0, + "content": "21102" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.208, + 0.126, + 0.761, + 0.496 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.5, + 0.896, + 0.54 + ], + "angle": 0, + "content": "Figure 2. The results of our approach on different humans in various garments. In each example, we show the ground truth image in a novel view, the reconstructed geometry (overlayed on the image), and the novel-view synthesis image. From left to right: the S4 from DeepCap dataset [10], \"FranziRed\" from DynaCap dataset [8], and the S1 from DeepCap dataset [10]." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.547, + 0.473, + 0.726 + ], + "angle": 0, + "content": "Architecture and experimental settings. We adopt the S-DF network from [30], and the same occupancy and color network as [34]. In the non-rigid deformation network, there are 8 layers with a skip connection at the fourth layer and the hidden size is 1024. The deformation and appearance latent codes both have dimensions of 128. We set both weight \\(\\mu_{1}\\) and \\(\\mu_{2}\\) to 1. We set the loss weight \\(\\{\\lambda_1,\\dots,\\lambda_6\\}\\) to \\(\\{1.0,0.1,0.02,50,0.1,100.0\\}\\) empirically. The Adam optimizer [14] is used in the optimization, and the learning rate is \\(5e^{-4}\\) at the start of training and exponentially decays to \\(5e^{-5}\\) during the training. With a GPU of RTX 2080 Ti, our network training costs about 12 hours tested on a four-view video with 300 frames." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.733, + 0.42, + 0.749 + ], + "angle": 0, + "content": "4.2. Comparison to state-of-the-art methods" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.756, + 0.47, + 0.868 + ], + "angle": 0, + "content": "We compare our method with two state-of-the-art approaches for human performance capture, DeepCap [10] and NerfCap [34]. We also compare with BCNet [12], TailorNet [23], and ICON [36]. For DeepCap and NerfCap, the experimental results are from [34]. For other compared methods, we implement them using their released codes. We do not compare with DeepCloth [16] since their codes are not released and DeepCap [10] outperforms them in ge" + }, + { + "type": "table", + "bbox": [ + 0.532, + 0.544, + 0.864, + 0.646 + ], + "angle": 0, + "content": "
MethodAMVIoURVIoUSVIoU
DeepCap [10]86.1785.7586.90
NerfCap [34]88.9688.3290.09
ICON [36]--93.56
BCNet [12]82.3381.6782.94
TailorNet [23]83.8783.5084.51
Our method89.5888.7591.03
" + }, + { + "type": "table_caption", + "bbox": [ + 0.509, + 0.646, + 0.878, + 0.659 + ], + "angle": 0, + "content": "Evaluated on the subject S4 from DeepCap dataset [10]." + }, + { + "type": "table", + "bbox": [ + 0.531, + 0.664, + 0.863, + 0.766 + ], + "angle": 0, + "content": "
MethodAMVIoURVIoUSVIoU
DeepCap [10]85.3885.3785.41
NerfCap [34]88.0888.0788.12
ICON [36]--95.21
BCNet [12]78.9178.7780.83
TailorNet [23]78.1578.0078.21
Our method88.7488.4089.68
" + }, + { + "type": "table_caption", + "bbox": [ + 0.507, + 0.765, + 0.883, + 0.779 + ], + "angle": 0, + "content": "Evaluated on the 'FranziRed' from DynaCap dataset [8]." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.782, + 0.892, + 0.834 + ], + "angle": 0, + "content": "Table 1. Comparison to the state-of-the-arts in surface reconstruction accuracy. Note that, the accuracy of ICON [36] is high since they use the human mask in estimating the geometry, and their dress shapes are not recovered accurately as shown in Fig. 3." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.839, + 0.893, + 0.868 + ], + "angle": 0, + "content": "ometry tracking as reported in [16]. We conduct the qualitative and quantitative experiments on S4 of DeepCap and" + }, + { + "type": "page_number", + "bbox": [ + 0.465, + 0.887, + 0.505, + 0.898 + ], + "angle": 0, + "content": "21103" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.126, + 0.47, + 0.168 + ], + "angle": 0, + "content": "\"FranziRed\" from DynaCap. The test dataset are the same as [34] which include 300 - 400 frames randomly sampled from the original dataset." + }, + { + "type": "image", + "bbox": [ + 0.089, + 0.173, + 0.46, + 0.457 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.108, + 0.457, + 0.178, + 0.468 + ], + "angle": 0, + "content": "(a) (b)" + }, + { + "type": "image_caption", + "bbox": [ + 0.184, + 0.457, + 0.203, + 0.468 + ], + "angle": 0, + "content": "#" + }, + { + "type": "image_caption", + "bbox": [ + 0.205, + 0.457, + 0.256, + 0.468 + ], + "angle": 0, + "content": "(c)" + }, + { + "type": "image_caption", + "bbox": [ + 0.25, + 0.457, + 0.298, + 0.468 + ], + "angle": 0, + "content": "(d)" + }, + { + "type": "image_caption", + "bbox": [ + 0.301, + 0.457, + 0.35, + 0.468 + ], + "angle": 0, + "content": "(e)" + }, + { + "type": "image_caption", + "bbox": [ + 0.352, + 0.457, + 0.388, + 0.468 + ], + "angle": 0, + "content": "(f)" + }, + { + "type": "image_caption", + "bbox": [ + 0.391, + 0.457, + 0.441, + 0.468 + ], + "angle": 0, + "content": "(g)" + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.468, + 0.469, + 0.505 + ], + "angle": 0, + "content": "Figure 3. Qualitative comparison with the state-of-the-arts. (a) Ground truth image. (b) DeepCap [10]. (c) NerfCap [34]. (d) ICON [36]. (e) BCNet [12]. (f) TailorNet [23]. (g) Our method." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.508, + 0.473, + 0.868 + ], + "angle": 0, + "content": "We show the qualitative comparisons in Figure 3 and list the quantitative results in Table 1. We only compute the error of the input view for ICON since they use their own camera system. Our method performs significantly better than the compared methods [12,23,36] without a cloth template. BCNet [12] also separates the modeling of the clothing and body, but as a supervised learning method, the performance may degrade dramatically because of data domain gap. When the human is moving, BCNet fails to recover the dresses accurately due to arbitrary cloth deformation. In contrast, the appearance loss of our double-layer NeRF-s poses dense constraints on the clothing and body motion capture in a weakly-supervised manner. As a result, our method can achieve a much higher accuracy of geometry tracking even without the ground truth. The predicted garments of TailorNet [23] are not consistent with the real images since their model is conditioned on human poses and shapes. ICON obtains higher IoU accuracy by using the human mask in recovering the geometry. However, there are obvious artifacts on the recovered dresses of ICON [36]. Based on one-piece personalized template, it is hard for both DeepCap [10] and NerfCap [34] to extract the clothing motion, which is important for downstream applications in AR/VR. In contrast, our method tracks the motion for both the clothing and body. Without a pre-scanned cloth template, our method still obtains a high tracking accuracy" + }, + { + "type": "table", + "bbox": [ + 0.513, + 0.124, + 0.881, + 0.183 + ], + "angle": 0, + "content": "
MethodS4 of [10]‘FranziRed” of [8]
PSNRSSIMPSNRSSIM
NerfCap [34]24.5290.7923.1891.49
Our method23.0690.2821.7090.54
" + }, + { + "type": "table_caption", + "bbox": [ + 0.5, + 0.185, + 0.89, + 0.209 + ], + "angle": 0, + "content": "Table 2. Comparison to NerfCap [34] in terms of novel-view synthesis on two datasets." + }, + { + "type": "table", + "bbox": [ + 0.522, + 0.211, + 0.872, + 0.339 + ], + "angle": 0, + "content": "
MethodAMVIoUPSNRSSIM
w/o simulation loss89.0322.6689.35
w/o mask loss88.7523.2189.49
w/o color loss89.36--
blending render [6]88.7523.6490.51
1 camera view88.4921.3487.12
2 camera views89.2321.8988.38
6 camera views90.1023.8090.60
ours(4 views)89.5823.0690.28
" + }, + { + "type": "table_caption", + "bbox": [ + 0.518, + 0.342, + 0.872, + 0.355 + ], + "angle": 0, + "content": "Table 3. Ablation studies on S4 from DeepCap dataset [10]." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.367, + 0.744, + 0.38 + ], + "angle": 0, + "content": "thanks to the implicit representation." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.381, + 0.892, + 0.534 + ], + "angle": 0, + "content": "We also compare the quality of free-viewpoint synthesis with NerfCap [34]. Figure 4 shows two compared examples and Table 2 reports the quantitative comparison. Without using any clothing template, our method still obtains close or comparable synthesis quality to NerfCap that uses a ground truth template. Our results are rendered by composing the NeRFs of the body and clothing, which is more complicated than a single NeRFs of NerfCap. For example, when the mutual motion occurs between the body and clothing (like cloth sliding), new body parts may appear, leading to the decrease of our synthesis quality." + }, + { + "type": "image", + "bbox": [ + 0.534, + 0.541, + 0.58, + 0.64 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.55, + 0.64, + 0.567, + 0.651 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.582, + 0.541, + 0.627, + 0.64 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.595, + 0.64, + 0.612, + 0.651 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image", + "bbox": [ + 0.629, + 0.542, + 0.674, + 0.639 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.641, + 0.64, + 0.659, + 0.651 + ], + "angle": 0, + "content": "(c)" + }, + { + "type": "image", + "bbox": [ + 0.681, + 0.544, + 0.74, + 0.639 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.704, + 0.64, + 0.721, + 0.651 + ], + "angle": 0, + "content": "(d)" + }, + { + "type": "image", + "bbox": [ + 0.742, + 0.544, + 0.799, + 0.639 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.761, + 0.64, + 0.778, + 0.651 + ], + "angle": 0, + "content": "(e)" + }, + { + "type": "image", + "bbox": [ + 0.801, + 0.544, + 0.857, + 0.639 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.821, + 0.64, + 0.835, + 0.651 + ], + "angle": 0, + "content": "(f)" + }, + { + "type": "image_caption", + "bbox": [ + 0.5, + 0.653, + 0.89, + 0.679 + ], + "angle": 0, + "content": "Figure 4. Comparison of novel-view synthesis with NerfCap [34]. (a,d) Ground truth. (b,e) NerfCap [34]. (c,f) Our method." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.692, + 0.661, + 0.706 + ], + "angle": 0, + "content": "4.3. Ablation studies" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.714, + 0.892, + 0.756 + ], + "angle": 0, + "content": "The ablation experiments are performed to demonstrate the effectiveness of our framework on the S4 from DeepCap dataset [10]. The quantitative results are listed in Table 3." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.756, + 0.892, + 0.867 + ], + "angle": 0, + "content": "Effectiveness of different losses. We first evaluate the effectiveness of the main losses in our optimization, i.e., the simulation loss, the mask loss, and the color loss. We report the accuracy by removing each of them, and show qualitative comparison examples in Figure 5. Without the simulation loss, the body-cloth collision becomes severe, showing that the simulation loss can preserve physically-plausible body-cloth interactions. Using the input human mask can" + }, + { + "type": "page_number", + "bbox": [ + 0.465, + 0.887, + 0.505, + 0.898 + ], + "angle": 0, + "content": "21104" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.127, + 0.47, + 0.182 + ], + "angle": 0, + "content": "improve the alignment accuracy with the human silhouettes. The color loss of dynamic NeRF can further enhance the detailed deformation on the human surface by posing dense appearance constraints for the geometry tracking." + }, + { + "type": "image", + "bbox": [ + 0.124, + 0.189, + 0.182, + 0.279 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.183, + 0.189, + 0.242, + 0.279 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.244, + 0.189, + 0.304, + 0.279 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.305, + 0.189, + 0.363, + 0.279 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.365, + 0.189, + 0.424, + 0.279 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.125, + 0.28, + 0.182, + 0.368 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.183, + 0.28, + 0.242, + 0.368 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.244, + 0.28, + 0.304, + 0.368 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.305, + 0.28, + 0.363, + 0.368 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.365, + 0.28, + 0.424, + 0.368 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.145, + 0.369, + 0.163, + 0.381 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image_caption", + "bbox": [ + 0.203, + 0.369, + 0.221, + 0.38 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image_caption", + "bbox": [ + 0.266, + 0.369, + 0.282, + 0.38 + ], + "angle": 0, + "content": "(c)" + }, + { + "type": "image_caption", + "bbox": [ + 0.324, + 0.369, + 0.344, + 0.38 + ], + "angle": 0, + "content": "(d)" + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.383, + 0.469, + 0.472 + ], + "angle": 0, + "content": "Figure 5. Ablation study for different losses. (a) Ground truth. (b) Using complete losses. (c) W/o simulation loss. (d) W/o mask loss. (e) W/o color loss. There are unnatural cloth deformations without the simulation loss. The rendered mask cannot align with the silhouette well without the mask loss. Compared to that without color loss, the surface detailed deformation can be further improved using complete losses (b)." + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.483, + 0.135, + 0.59 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.099, + 0.593, + 0.117, + 0.604 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.157, + 0.485, + 0.2, + 0.59 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.202, + 0.485, + 0.254, + 0.589 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.218, + 0.593, + 0.236, + 0.604 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image", + "bbox": [ + 0.256, + 0.485, + 0.305, + 0.589 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.323, + 0.498, + 0.367, + 0.555 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.37, + 0.485, + 0.422, + 0.589 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.424, + 0.485, + 0.475, + 0.589 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.387, + 0.593, + 0.405, + 0.604 + ], + "angle": 0, + "content": "(c)" + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.607, + 0.469, + 0.67 + ], + "angle": 0, + "content": "Figure 6. Comparison of different rendering. (a) Ground truth. (b) Our composite rendering. (c) Blend rendering [6]. We show the synthesis of the clothing, body, and full human. There are much cloth appearance on the body caused by the blending [6], while our method separates the appearance of the clothing and body better." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.673, + 0.468, + 0.811 + ], + "angle": 0, + "content": "Effectiveness of the composite rendering. We compare our composite rendering with the blending rendering in [6]. Through free composition of two NeRFs for each point, the method of [6] can obtain high fitting color. However, the blending may generate missing areas or artifacts in the novel-view synthesis. Moreover, it fails to recover the appearance of the body and clothing reliably as shown in Figure 6. Comparatively, we can recover the appearance of the body and clothing separately and compose them into a full human more accurately using the composite rendering." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.812, + 0.468, + 0.867 + ], + "angle": 0, + "content": "The number of camera views. We investigate the influence of different number of input camera views on the performance of our approach. We compare the results of our method under \\(\\{1,2,4,6\\}\\) views, and show the comparison" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.127, + 0.892, + 0.168 + ], + "angle": 0, + "content": "example of novel-view synthesis in Figure 7. The results become slightly better using more camera views because the moving humans are captured more completely." + }, + { + "type": "image", + "bbox": [ + 0.548, + 0.177, + 0.61, + 0.287 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.546, + 0.29, + 0.61, + 0.299 + ], + "angle": 0, + "content": "Ground Truth" + }, + { + "type": "image", + "bbox": [ + 0.612, + 0.177, + 0.667, + 0.287 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.624, + 0.29, + 0.653, + 0.299 + ], + "angle": 0, + "content": "1 view" + }, + { + "type": "image", + "bbox": [ + 0.669, + 0.177, + 0.726, + 0.287 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.681, + 0.29, + 0.718, + 0.299 + ], + "angle": 0, + "content": "2 views" + }, + { + "type": "image", + "bbox": [ + 0.728, + 0.177, + 0.785, + 0.287 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.741, + 0.29, + 0.777, + 0.298 + ], + "angle": 0, + "content": "4 views" + }, + { + "type": "image", + "bbox": [ + 0.788, + 0.177, + 0.846, + 0.287 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.798, + 0.29, + 0.835, + 0.298 + ], + "angle": 0, + "content": "6 views" + }, + { + "type": "image_caption", + "bbox": [ + 0.5, + 0.3, + 0.891, + 0.338 + ], + "angle": 0, + "content": "Figure 7. Ablation study for different numbers of camera views. With more cameras, the novel-view synthesis quality becomes better (marked in cycle)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.357, + 0.633, + 0.372 + ], + "angle": 0, + "content": "4.4. Applications" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.378, + 0.892, + 0.476 + ], + "angle": 0, + "content": "Benefiting from separate tracking results of human body and clothing, we can enable interesting applications like cloth retargeting as shown in Figure 8. Since the body mesh and cloth mesh are captured, we can exchange the garments between two person by simulating the garments on the other body. Moreover, free-viewpoint motion can be created by simulating the clothing on human body sequences." + }, + { + "type": "image", + "bbox": [ + 0.524, + 0.484, + 0.609, + 0.576 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.577, + 0.485, + 0.636, + 0.574 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.646, + 0.485, + 0.685, + 0.574 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.695, + 0.485, + 0.749, + 0.574 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.764, + 0.485, + 0.807, + 0.574 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.821, + 0.485, + 0.871, + 0.574 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.579, + 0.891, + 0.618 + ], + "angle": 0, + "content": "Figure 8. Application of cloth retargeting. From the captured results using our method, we can retarget the garments between the S4 of DeepCap and the 'FranziRed' of DynaCap." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.635, + 0.619, + 0.649 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.658, + 0.892, + 0.81 + ], + "angle": 0, + "content": "In this paper, we proposed a novel performance capture method for clothed humans with a double-layer neural radiance fields (NeRF). Our double-layer NeRFs can effectively model moving humans in daily-life clothes. By matching the rendered color with the observed color, our method estimates the clothing and body motion and temporally-varying appearance robustly. In addition, we learn a differential physics-aware network for different garment types which supports physically-plausible body-cloth interactions and cloth dynamics. The experimental results on real datasets demonstrate the effectiveness of the proposed method." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.812, + 0.891, + 0.867 + ], + "angle": 0, + "content": "Acknowledgments This work was supported by the Fundamental Research Funds for the Central Universities (NJ2022028), and the Natural Science Foundation of China (No.U1713208), and Program for Changjiang Scholars." + }, + { + "type": "page_number", + "bbox": [ + 0.465, + 0.887, + 0.505, + 0.898 + ], + "angle": 0, + "content": "21105" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.125, + 0.174, + 0.139 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.148, + 0.445, + 0.161 + ], + "angle": 0, + "content": "[1] https://www.marvelousdesigner.com/.3,5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.162, + 0.469, + 0.2 + ], + "angle": 0, + "content": "[2] Thiemo Alldieck, Gerard Pons-Moll, Christian Theobalt, and Marcus Magnor. Tex2Shape: Detailed full human body geometry from a single image. In ICCV, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.2, + 0.469, + 0.238 + ], + "angle": 0, + "content": "[3] Bharat Lal Bhatnagar, Garvita Tiwari, Christian Theobalt, and Gerard Pons-Moll. Multi-Garment Net: Learning to dress 3D people from images. In ICCV, 2019. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.239, + 0.469, + 0.314 + ], + "angle": 0, + "content": "[4] Mingsong Dou, Sameh Khamis, Yury Degtyarev, Philip Davidson, Sean Fanello, Adarsh Kowdle, Sergio Orts Escolano, Christoph Rhemann, David Kim, Jonathan Taylor, Pushmeet Kohli, Vladimir Tankovich, and Shahram Izadi. Fusion4D: Real-time performance capture of challenging scenes. ACM TOG, 35:1-13, July 2016. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.315, + 0.469, + 0.365 + ], + "angle": 0, + "content": "[5] Stuart Ganan and D McClure. Bayesian image analysis: An application to single photon emission tomography. In Proceedings of the American Statistical Association, pages 12-18, 1985. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.367, + 0.469, + 0.404 + ], + "angle": 0, + "content": "[6] Chen Gao, Ayush Saraf, Johannes Kopf, and Jia-Bin Huang. Dynamic view synthesis from dynamic monocular video. In ICCV, 2021. 4, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.405, + 0.469, + 0.455 + ], + "angle": 0, + "content": "[7] Erhan Gundogdu, Victor Constantin, Amrollah Seifoddini, Minh Dang, Mathieu Salzmann, and Pascal Fua. GarNet: A two-stream network for fast and accurate 3d cloth draping. In ICCV, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.457, + 0.469, + 0.507 + ], + "angle": 0, + "content": "[8] Marc Habermann, Lingjie Liu, Weipeng Xu, Michael Zollhoefer, Gerard Pons-Moll, and Christian Theobalt. Real-time deep dynamic characters. ACM Transactions on Graphics, 40(4), August 2021. 2, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.508, + 0.469, + 0.558 + ], + "angle": 0, + "content": "[9] Marc Habermann, Weipeng Xu, Michael Zollhoefer, Gerard Pons-Moll, and Christian Theobalt. LiveCap: Real-time human performance capture from monocular video. ACM TOG, 38(2):14:1-14:17, March 2019. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.559, + 0.469, + 0.609 + ], + "angle": 0, + "content": "[10] Marc Habermann, Weipeng Xu, Michael Zollhoefer, Gerard Pons-Moll, and Christian Theobalt. DeepCap: Monocular human performance capture using weak supervision. In CVPR, 2020. 1, 2, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.61, + 0.469, + 0.647 + ], + "angle": 0, + "content": "[11] Boyi Jiang, Yang Hong, Hujun Bao, and Juyong Zhang. SelfRecon: Self reconstruction your digital avatar from monocular video. In CVPR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.649, + 0.469, + 0.687 + ], + "angle": 0, + "content": "[12] Boyi Jiang, Juyong Zhang, Yang Hong, Jinhao Luo, Ligang Liu, and Hujun Bao. BCNet: Learning body and cloth shape from a single image. In ECCV, 2020. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.688, + 0.469, + 0.724 + ], + "angle": 0, + "content": "[13] Hanbyul Joo, Tomas Simon, and Yaser Sheikh. Total capture: A 3D deformation model for tracking faces, hands, and bodies. In CVPR, 2018. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.726, + 0.469, + 0.763 + ], + "angle": 0, + "content": "[14] Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In International Conference for Learning Representations, 2015. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.765, + 0.469, + 0.803 + ], + "angle": 0, + "content": "[15] John P Lewis, Matt Cordner, and Nickson Fong. Pose space deformation: A unified approach to shape interpolation and skeleton-driven deformation. In SIGGRAPH, 2000. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.804, + 0.469, + 0.866 + ], + "angle": 0, + "content": "[16] Yue Li, Marc Habermann, Bernhard Thomaszewski, Stelian Coros, Thabo Beeler, and Christian Theobalt. Deep physics-aware inference of cloth deformation for monocular human performance capture. In International Conference on 3D Vision (3DV), 2021. 1, 2, 5, 6" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.148, + 0.469, + 0.866 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.128, + 0.892, + 0.178 + ], + "angle": 0, + "content": "[17] Lingjie Liu, Marc Habermann, Viktor Rudnev, Kripasindhu Sarkar, Jiatao Gu, and Christian Theobalt. Neural Actor: Neural free-view synthesis of human actors with pose control. ACM TOG, 40(6):1-16, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.181, + 0.892, + 0.218 + ], + "angle": 0, + "content": "[18] Matthew Loper, Naureen Mahmood, Javier Romero, Gerard Pons-Moll, and Michael J. Black. SMPL: A skinned multiperson linear model. ACM TOG, 34(6):248, 2015. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.22, + 0.892, + 0.27 + ], + "angle": 0, + "content": "[19] Qianli Ma, Jinlong Yang, Anurag Ranjan, Sergi Pujades, Gerard Pons-Moll, Siyu Tang, and Michael J. Black. Learning to dress 3D people in generative clothing. In CVPR, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.273, + 0.892, + 0.323 + ], + "angle": 0, + "content": "[20] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. N-eRF: Representing scenes as neural radiance fields for view synthesis. In ECCV, 2020. 1, 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.326, + 0.892, + 0.363 + ], + "angle": 0, + "content": "[21] Richard Newcombe, Dieter Fox, and Steve Seitz. Dynamic-Fusion: Reconstruction and tracking of non-rigid scenes in real-time. In CVPR, 2015. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.366, + 0.892, + 0.404 + ], + "angle": 0, + "content": "[22] Michael Oechsle, Songyou Peng, and Andreas Geiger. U-NISURF: Unifying neural implicit surfaces and radiance fields for multi-view reconstruction. In ICCV, 2021. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.406, + 0.892, + 0.455 + ], + "angle": 0, + "content": "[23] Chaitanya Patel, Zhouyingcheng Liao, and Gerard Pons-Moll. TailorNet: Predicting clothing in 3D as a function of human pose, shape and garment style. In CVPR, 2020. 2, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.459, + 0.892, + 0.509 + ], + "angle": 0, + "content": "[24] Sida Peng, Junting Dong, Qianqian Wang, Shangzhan Zhang, Qing Shuai, Xiaowei Zhou, and Hujun Bao. Animatable neural radiance fields for modeling dynamic human bodies. In ICCV, 2021. 2, 3, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.512, + 0.892, + 0.574 + ], + "angle": 0, + "content": "[25] Sida Peng, Yuanqing Zhang, Yinghao Xu, Qianqian Wang, Qing Shuai, Hujun Bao, and Xiaowei Zhou. Neural Body: Implicit neural representations with structured latent codes for novel view synthesis of dynamic humans. In CVPR, 2021. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.577, + 0.892, + 0.615 + ], + "angle": 0, + "content": "[26] Gerard Pons-Moll, Sergi Pujades, Sonny Hu, and Michael J. Black. ClothCap: Seamless 4D clothing capture and retargeting. ACM TOG, 36(4), July 2017. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.617, + 0.892, + 0.667 + ], + "angle": 0, + "content": "[27] Shunsuke Saito, Zeng Huang, Ryota Natsume, Shigeo Morishima, Angjoo Kanazawa, and Hao Li. PIFu: Pixel-aligned implicit function for high-resolution clothed human digitization. In ICCV, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.669, + 0.892, + 0.708 + ], + "angle": 0, + "content": "[28] Shunsuke Saito, Tomas Simon, Jason Saragih, and Hanbyul Joo. PIFuHD: Multi-level pixel-aligned implicit function for high-resolution 3D human digitization. In CVPR, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.71, + 0.892, + 0.748 + ], + "angle": 0, + "content": "[29] Igor Santesteban, Miguel A. Otaduy, and Dan Casas. Learning-based animation of clothing for virtual try-on. Computer Graphics Forum (Proc. Eurographics), 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.75, + 0.892, + 0.801 + ], + "angle": 0, + "content": "[30] Vincent Sitzmann, Julien NP Martel, Alexander W Bergman, David B Lindell, and Gordon Wetzstein. Implicit neural representations with periodic activation functions. In Advances in Neural Information Processing Systems, 2020. 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.803, + 0.892, + 0.866 + ], + "angle": 0, + "content": "[31] Zhaoqi Su, Weilin Wan, Tao Yu, Lingjie Liu, Lu Fang, Wenping Wang, and Yebin Liu. MulayCap: Multi-layer human performance capture using a monocular video camera. IEEE Transactions on Visualization and Computer Graphics, 28(4):1862-1879, 2022. 1, 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.128, + 0.892, + 0.866 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.465, + 0.888, + 0.505, + 0.898 + ], + "angle": 0, + "content": "21106" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.127, + 0.47, + 0.165 + ], + "angle": 0, + "content": "[32] Robert W. Sumner, Johannes Schmid, and Mark Pauly. Embedded deformation for shape manipulation. ACM TOG, 26(3), 2007. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.167, + 0.47, + 0.215 + ], + "angle": 0, + "content": "[33] Gül Varol, Duygu Ceylan, Bryan Russell, Jimei Yang, Ersin Yumer, Ivan Laptev, and Cordelia Schmid. BodyNet: Volumetric inference of 3D human body shapes. In ECCV, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.218, + 0.469, + 0.269 + ], + "angle": 0, + "content": "[34] Kangkan Wang, Sida Peng, Xiaowei Zhou, Jian Yang, and Guofeng Zhang. NerfCap: Human performance capture with dynamic neural radiance fields. IEEE Transactions on Visualization and Computer Graphics, 2022. 1, 3, 4, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.271, + 0.469, + 0.308 + ], + "angle": 0, + "content": "[35] Olivia Wiles, Georgia Gkioxari, Richard Szeliski, and Justin Johnson. Synsin: End-to-end view synthesis from a single image. In CVPR, pages 7467-7477, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.31, + 0.469, + 0.347 + ], + "angle": 0, + "content": "[36] Yuliang Xiu, Jinlong Yang, Dimitrios Tzionas, and Michael J. Black. ICON: Implicit clothed humans obtained from normals. In CVPR, June 2022. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.349, + 0.469, + 0.399 + ], + "angle": 0, + "content": "[37] Weipeng Xu, Avishek Chatterjee, Michael Zollhöfer, Helge Rhodin, Dushyant Mehta, Hans-Peter Seidel, and Christian Theobalt. MonoPerfCap: Human performance capture from monocular video. ACM TOG, 37(2):1-15, July 2018. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.401, + 0.469, + 0.462 + ], + "angle": 0, + "content": "[38] Tao Yu, Kaiwen Guo, Feng Xu, Yuan Dong, Zhaqi Su, Jianhui Zhao, Jianguo Li, Qionghai Dai, and Yebin Liu. Body-Fusion: Real-time capture of human motion and surface geometry using a single depth camera. In ICCV, October 2017. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.465, + 0.469, + 0.526 + ], + "angle": 0, + "content": "[39] Tao Yu, Zerong Zheng, Kaiwen Guo, Jianhui Zhao, Qionghai Dai, Hao Li, Gerard Pons-Moll, and Yebin Liu. DoubleFusion: Real-time capture of human performances with inner body shapes from a single depth sensor. In CVPR, June 2018. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.529, + 0.469, + 0.579 + ], + "angle": 0, + "content": "[40] Tao Yu, Zerong Zheng, Yuan Zhong, Jianhui Zhao, Qionghai Dai, Gerard Pons-Moll, and Yebin Liu. SimulCap: Single-view human performance capture with cloth simulation. In CVPR, 2019. 1, 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.581, + 0.469, + 0.63 + ], + "angle": 0, + "content": "[41] Fuqiang Zhao, Wei Yang, Jiakai Zhang, Pei Lin, Yingliang Zhang, Jingyi Yu, and Lan Xu. HumanNeRF: Generalizable neural human radiance field from sparse inputs. In CVPR, 2022. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.633, + 0.469, + 0.671 + ], + "angle": 0, + "content": "[42] Zerong Zheng, Tao Yu, Yixuan Wei, Qionghai Dai, and Yebin Liu. DeepHuman: 3D human reconstruction from a single image. In ICCV, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.672, + 0.469, + 0.734 + ], + "angle": 0, + "content": "[43] Heming Zhu, Yu Cao, Hang Jin, Weikai Chen, Dong Du, Zhangye Wang, Shuguang Cui, and Xiaoguang Han. Deep Fashion3D: A dataset and benchmark for 3d garment reconstruction from single images. In ECCV, pages 512-530, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.737, + 0.469, + 0.786 + ], + "angle": 0, + "content": "[44] Heming Zhu, Lingteng Qiu, Yuda Qiu, and Xiaoguang Han. Registering Explicit to Implicit: Towards high-fidelity garment mesh reconstruction from single images. In CVPR, pages 3845-3854, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.789, + 0.469, + 0.826 + ], + "angle": 0, + "content": "[45] Hao Zhu, Xinxin Zuo, Sen Wang, Xun Cao, and Ruigang Yang. Detailed human shape estimation from a single image by hierarchical mesh deformation. In CVPR, 2019. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.127, + 0.47, + 0.826 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.466, + 0.887, + 0.505, + 0.898 + ], + "angle": 0, + "content": "21107" + } + ] +] \ No newline at end of file diff --git a/2023/Clothed Human Performance Capture With a Double-Layer Neural Radiance Fields/edeeca28-ac72-40e3-9181-9737d0c908f2_origin.pdf b/2023/Clothed Human Performance Capture With a Double-Layer Neural Radiance Fields/edeeca28-ac72-40e3-9181-9737d0c908f2_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..1a6a01196b1e8a1b213ed2b83b129438c7aa5da3 --- /dev/null +++ b/2023/Clothed Human Performance Capture With a Double-Layer Neural Radiance Fields/edeeca28-ac72-40e3-9181-9737d0c908f2_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d2783f0c247cf4e0ca787cc4056cf3d3bbb844a1da9807fe693a607480936ff5 +size 1634672 diff --git a/2023/Clothed Human Performance Capture With a Double-Layer Neural Radiance Fields/full.md b/2023/Clothed Human Performance Capture With a Double-Layer Neural Radiance Fields/full.md new file mode 100644 index 0000000000000000000000000000000000000000..dd27ba09ca63a2f07279fba486b486c1c87eed96 --- /dev/null +++ b/2023/Clothed Human Performance Capture With a Double-Layer Neural Radiance Fields/full.md @@ -0,0 +1,409 @@ +# Clothed Human Performance Capture with a Double-layer Neural Radiance Fields + +Kangkan Wang $^{1,2}$ , Guofeng Zhang $^{3}$ , Suxu Cong $^{1}$ , Jian Yang $^{1,2}$ + +1Key Lab of Intelligent Perception and Systems for High-Dimensional Information of Ministry of Education, 2Jiangsu Key Lab of Image and Video Understanding for Social Security, + +School of Computer Science and Engineering, Nanjing University of Science and Technology, China +3State Key Laboratory of CAD&CG, Zhejiang University, China + +# Abstract + +This paper addresses the challenge of capturing performance for the clothed humans from sparse-view or monocular videos. Previous methods capture the performance of full humans with a personalized template or recover the garments from a single frame with static human poses. However, it is inconvenient to extract cloth semantics and capture clothing motion with one-piece template, while single frame-based methods may suffer from instable tracking across videos. To address these problems, we propose a novel method for human performance capture by tracking clothing and human body motion separately with a double-layer neural radiance fields (NeRFs). Specifically, we propose a double-layer NeRFs for the body and garments, and track the densely deforming template of the clothing and body by jointly optimizing the deformation fields and the canonical double-layer NeRFs. In the optimization, we introduce a physics-aware cloth simulation network which can help generate physically plausible cloth dynamics and body-cloth interactions. Compared with existing methods, our method is fully differentiable and can capture both the body and clothing motion robustly from dynamic videos. Also, our method represents the clothing with an independent NeRFs, allowing us to model implicit fields of general clothes feasibly. The experimental evaluations validate its effectiveness on real multi-view or monocular videos. + +# 1. Introduction + +Performance capture for clothed humans is one of the essential problems in the metaverse, and it not only captures the inner human body motion but also recovers the outer clothing motion which has many promising applications such as virtual try-on, video editing, and telepresence. From sparse-view or monocular videos of a moving human + +in general clothes, its goal is to recover the dynamic 3D shape sequence of the human body and clothing simultaneously that are consistent with the observed frames in both human shape and motion. This is a very challenging problem since the dynamic human could be with arbitrary motions and with complex non-rigid cloth deformations, and the clothing in motion is difficult to maintain physically plausible interactions with the body. + +Previous systems [4, 21, 38, 39] reconstruct 3D clothed humans by using depth sensors or fitting a personalized template [9, 10, 37] to the image observations (e.g., body joints and silhouettes). Only recovering one-piece geometry which unifies the human body and clothes, these systems fail to track the motion of the clothing and achieve clothing editing on 3D humans, which are the prerequisites in many VR/AR applications like virtual dressing. On the contrary, cloth can be extracted and tracked from depth scans [26, 40] accurately by fitting pre-built cloth templates to the scans which have limited applications when 3D data are unavailable. Existing garment estimation methods [3, 12, 43, 44] from color images require the person facing the camera and in static poses. When the human is moving and the clothing is deforming, these methods may recover the 3D garments unreliable. Recent methods [16, 31] track body and clothing motion simultaneously from videos, but they need to re-build cloth template for a new performer and the running efficiency is very low due to online cloth simulation or computationally-exhaustive optimization, which prohibits them from being widely deployed for daily applications. + +Recent works [20, 34, 41] adopt dynamic human NeRFs to capture human motion and obtain impressive tracking results. By capturing the temporally-varying appearance in the videos, dynamic NeRFs [34] can provide dense photometric constraints to track the deforming geometry of the performer. However, they represent the human with a single NeRFs without modeling cloth, and the clothing motion cannot be extracted. In this paper, we aim to track the cloth + +ing and body motion simultaneously with dynamic NeRFs. However, this problem is rather challenging due to two major questions we need to solve: how to represent dynamic clothing and human body with NeRFs, and how to capture clothing and human body motion with plausible body-cloth interactions based on the implicit representation. + +In this paper, we propose a novel method for clothed human performance capture with a double-layer NeRFs. Specifically, a double-layer NeRFs is modeled for both the body and clothing in the template space, and transformed to the observation space with the corresponding deformation fields, and the rendered images are then synthesized by composing the two dynamic NeRFs. We first estimate the template shape in canonical frame and learn the geometry network supervised by the template geometry. In the rendering, we compose the double-layer NeRFs with the guidance of deformed body and clothing meshes. Then, by minimizing the difference between the rendered color and observed color, the deformation fields and the canonical N-NeRFs are optimized jointly. The deformation field is represented as the inverse deformation of the template mesh, thus the densely deforming geometry of the template can be recovered simultaneously. In addition, we adopt a physics-aware network learnt from simulation data between various cloth types and humans to constrain the dynamic clothing and preserve physically plausible body-cloth interactions, resulting in realistic cloth geometry tracking. Compared to previous methods, our method is fully differentiable and can recover the realistic motion of both the clothing and body from dynamic videos with arbitrary human poses and complex cloth deformations. The experimental qualitative and quantitative results on datasets of DynaCap [8] and DeepCap [10] prove that the proposed approach can robustly and accurately capture the motion for clothed humans. In summary, the primary contributions of our work include: + +- We propose a double-layer NeRFs for dynamic humans in general clothing, allowing us to model implicit humans with a variety of clothes (e.g., loose dresses). +- To the best of our knowledge, we propose the first framework to capture clothing motion separately from the human body using the double-layer NeRFs, which provides dense appearance constraints on the geometry tracking and improves the robustness and accuracy. +- A differentiable physics-aware network is learnt for d-different common garments and used to preserve physically plausible cloth deformations in motion capture. + +# 2. Related Work + +Clothed human reconstruction. Depth-based methods [4,21,38,39] reconstruct 3D humans in realtime by fusing the depth maps in a volumetric way. However, depth-based methods cannot synthesize photo-realistic videos and + +have constrained applications in daily life due to data capture limitations, e.g., high power consumption and sensitive to sunlight. Based on deep learning techniques, some approaches predict 3D humans from color images through implicit function [27, 28], coarse-to-fine mesh deformation [45], surface normal optimization [36, 42], or UV mapping [2]. Although these methods can reconstruct human details, the recovered shapes do not contain human semantics like human pose, shape and clothing. To recover the clothing, some methods [3, 19] add displacements on vertices of SMPL model [18] which only support tight clothing and recover coarse-level geometry. Template-based methods [9, 10, 37] recover 3D human motions by non-rigidly fitting a person-specific template mesh to 2D observed features (e.g., human joints or silhouettes). With a personalized template, these methods obtain impressive performance capture. However, the recovered meshes integrate the body and clothing in one piece, which hinders their applications like garment retargeting and virtual trying-on. + +Clothing motion capture. Traditional methods [26, 40] capture the cloth motion by precisely fitting a cloth template to 3D data from depth sensors, which cannot generalize to the video-based setting and are limited to applications in indoor scenes. Deep learning-based approaches [3, 12, 43, 44] estimate the 3D garment shape from color images. However, these methods require the human facing the camera with a static pose and do not try temporally tracking of dynamic garments. Some works [23, 29] predict the cloth shape conditioned on the human pose and shape but the shapes are not consistent with the observed data. Recent works [16, 31] capture both human body and clothing motion from dynamic videos. MulayCap [31] optimizes a simulated garment on SMPL model with semantic cloth segmentation, but their traditional simulation and optimization are both computationally expensive. DeepCloth [16] incorporates a physics-based simulator to supervise the cloth deformations estimated with multi-view joints and silhouettes as DeepCap [10]. However, their simulation layer is not differentiable and needs to adjust the cloth material types online, which is a tedious and time consuming step. Moreover, with the same limitation as DeepCap, they cannot recover the detailed cloth deformation based on the sparse features. + +Implicit Human Reconstruction. Clothed humans can be reconstructed through implicit representation-based methods such as voxel representation [33, 42], implicit function [27, 28], or neural radiance fields (NeRFs) [20, 24, 25]. Recent works combine explicit and implicit representations to produce high-quality avatar [11] or reconstruct high-fidelity garment mesh [44], which show the advantage of implicit representation on recovering geometry details. However, these methods are limited to self-rotating human motion or a single color image. Neuralbody [25] and AniNErf [24] extend NeRFs for a dynamic human with defor + +![](images/be976881f92fa9d91395654b0c248d79871f51f1ec80781719aa574f9cd75ff7.jpg) +Figure 1. Overview of the proposed approach. Our approach can capture both the clothing and human body motion, and achieve novel-view synthesis from sparse-view or monocular videos. We learn a double-layer neural radiance fields (NeRFs) for the clothing and human body in canonical frame, and jointly optimize the canonical double-layer NeRFs and the human deformation constrained with a physics-aware simulation network. Please refer to Sec. 3 for details of our method. + +mation fields based on the linear blend skinning (LBS) [15]. Some human NeRFs methods [17, 41] model nonrigid human deformations with a point displacement on top of LBS. However, without using any clothing motion prior, these methods cannot effectively recover the motion of humans in loose clothes, and the geometry of clothes is not reconstructed reliably under nonrigid cloth deformations. The recent work, NerfCap [34] incorporates the embedded graph prior into the human NeRFs and represents nonlinear surface deformations for loose clothes successfully. However, their method does not model the clothing separately and thus fails to track the clothing motion. + +# 3. Proposed Approach + +Given sparse-view or monocular videos of a clothed human, we track the motion of both the human body and clothing and synthesize novel-view videos of the human with a novel performance capture method based on a double-layer neural radiance fields (NeRFs). Figure 1 shows an overview of the proposed method. The double-layer NeRFs is learnt for the human body and clothing in canonical frame (Sec. 3.2). Then, the canonical NeRFs is dynamically transformed into the observed frame with inverse deformation fields of the deformed template for clothing and body respectively. From the dynamic double-layer NeRFs, we render the images using a composite rendering. By matching the synthesized images with the video frames and constraining the cloth deformations with a physics-aware simulation + +network, we estimate the double-layer NeRFs and the human deformation jointly (Sec. 3.3). + +# 3.1. Canonical Human Shape + +We decompose the clothed human into undressed body and garments (like T-shirt and pants), and represent them with SMPL model [18] and a parametric PCA model respectively. To include the human details (e.g., the head and hand), we obtain the canonical body model $M$ by deforming the SMPL model to a personalized template with non-rigid deformation [32]. We design different styles (size and sleeve length, et al.) for each garment category, and simulate these garments on the template SMPL model using Marvelous Designer [1]. Then, we obtain a linear model through PCA on the simulated meshes, + +$$ +\boldsymbol {G} (\alpha) = \bar {\boldsymbol {G}} + \boldsymbol {B} _ {g} \alpha , \tag {1} +$$ + +where $\bar{G}$ is the mean garment shape, $B_{g}$ denotes the linear basis matrix, and $\alpha$ are parameters of the cloth shape $G(\alpha)$ . + +# 3.2. Double-layer NeRFs for Dynamic Humans + +Unlike a single NeRFs for clothed humans [17,24,25,34, 41], we represent the clothing with an independent NeRFs on the body, which forms a double-layer NeRFs. The implicit representation can be extended to multiple garments. Canonical NeRF. We use an occupancy network [22] to represent canonical geometry since the occupancy can mod + +el geometry more precisely than density [20] as, + +$$ +\boldsymbol {o} (\boldsymbol {x}) = \boldsymbol {F} _ {o} \left(\gamma_ {\boldsymbol {x}} (\boldsymbol {x})\right), \tag {2} +$$ + +where $\gamma_{\pmb{x}}$ is the positional encoding [20]. The color network $F_{c}$ in canonical frame is formulated as, + +$$ +\boldsymbol {c} _ {i} (\boldsymbol {x}) = \boldsymbol {F} _ {c} \left(\gamma_ {\boldsymbol {x}} (\boldsymbol {x}), \varphi_ {i}\right), \tag {3} +$$ + +where $\varphi_{i}$ is an appearance latent code [24, 25] for frame $i$ . We define occupancy network $\{F_o^b,F_o^g\}$ , and color network $\{F_c^b,\varphi^b,F_c^g,\varphi^g\}$ for the body and clothing, respectively. + +To facilitate the learning of occupancy network, we incorporate a signed distance fields (SDFs) network $S$ , + +$$ +\boldsymbol {S}: \boldsymbol {p} \in \mathbb {R} ^ {3} \rightarrow \boldsymbol {s} \in \mathbb {R}. \tag {4} +$$ + +which predicts the SDF value $s$ for point $p$ . We use the losses proposed in [30] to learn the SDFs of the shape, + +$$ +\begin{array}{l} L _ {s d f} = \sum_ {\boldsymbol {p} \in \Phi} | \boldsymbol {S} (\boldsymbol {p}) | + \sum_ {\boldsymbol {p} \in \Phi} (1 - \langle \nabla \boldsymbol {S} (\boldsymbol {p}), \bar {\boldsymbol {n}}) \rangle \tag {5} \\ + \sum_ {\boldsymbol {p} \in \Omega} | \| \nabla \boldsymbol {S} (\boldsymbol {p}) \| _ {2} - 1 | + \sum_ {\boldsymbol {p} \in \Omega \backslash \Phi} \exp (- \delta \cdot | \boldsymbol {S} (\boldsymbol {p}) |), \\ \end{array} +$$ + +where $\bar{n}$ is the surface normal, $\nabla S(\pmb{p})$ indicates the gradient of the SDFs, $\Omega$ and $\Phi$ denotes the 3D space and shape surface respectively. The first and second terms are used to estimate valid SDFs and normals on the surfaces, respectively. The third term constrains the norm of SDF gradients to be 1, and the last term penalizes SDF values of non-surface points near 0 with $\delta \gg 1$ . We define SDF network $\{\pmb{S}^b, \pmb{S}^g\}$ for both the body and clothing. + +Supervised by the predictions of SDF network, the occupancy network is learnt with the following loss, + +$$ +L _ {o c c} = \sum_ {\boldsymbol {x} \in \Omega} L _ {C} \left(\boldsymbol {F} _ {\boldsymbol {o}} \left(\gamma_ {\boldsymbol {x}} (\boldsymbol {x})\right), \boldsymbol {o}\right), \tag {6} +$$ + +where $\pmb{o}$ is the occupancy indicated by SDFs (if $S(\pmb{x}) < = 0$ , $o = 1$ ; otherwise, $o = 0$ ), and $L_{C}$ denotes the cross-entropy loss. We adopt the compositional design of SDF and occupancy because it is better for volumetric learning with continuous occupancy fields [22] compared to discrete labels extracted from SDFs. The template is varying during learning geometry model, and the loss is formulated as, + +$$ +L _ {g e o} = \mu_ {1} L _ {s d f} + \mu_ {2} L _ {o c c}. \tag {7} +$$ + +where $\mu_{1}$ and $\mu_{2}$ are the weight for each term. Through the optimization, the NeRF geometry is consistent with the estimated template so that densely deforming geometry of the template can be tracked by utilizing dense photometric constraints from NeRF color model. + +Deformation Model. Following [34], we transform sampled points in observed space to canonical space with the + +inverse deformation of the nearest vertex on the deformed template. In this way, the human deformation model is disentangled into the non-rigid surface deformation (i.e., embedded graph deformation as in [34]) and the articulated skeletal motion. For the clothing and body, we define the deformation network $F_{d}^{g}$ with latent codes $\omega^{g}$ and $F_{d}^{b}$ with latent codes $\omega^{b}$ , respectively. Through nonrigid deformation, we obtain the clothing mesh $\hat{G}$ and body mesh $\hat{M}$ . Then, the linear blend skinning [15] is applied to obtain the final cloth mesh $\tilde{G}$ and body mesh $\tilde{M}$ in the observation space. We recover the SMPL model with [13] for input video frames and also estimate a canonical body shape. + +Composite Rendering. In [6], a blending weight is predicted for each sampled point to blend the static background and dynamic target in a scene. This blending method cannot separate the clothing and human body in the NeRFs effectively. We propose an efficient composite rendering strategy for the dynamic double-layer NeRF. Specifically, for a casted ray $r$ , we find the intersecting triangle with the deformed meshes. If the intersected triangle is from the body mesh, we set the mask $m_r^b$ of all sampled points on $r$ to 1; otherwise, $m_r^b$ is set to 0. Similarly, if the intersected triangle is from the cloth mesh, we set the mask $m_r^g$ to 1; otherwise, $m_r^g$ is set to 0. The complete human is rendered as, + +$$ +\boldsymbol {T} _ {r} ^ {i} = \prod_ {j = 1} ^ {i - 1} \left(1 - \boldsymbol {\sigma} _ {g} ^ {j} \left(1 - \boldsymbol {m} _ {r} ^ {b}\right)\right) \left(1 - \boldsymbol {\sigma} _ {b} ^ {j} \boldsymbol {m} _ {r} ^ {b}\right), \tag {8} +$$ + +$$ +\hat {\boldsymbol {C}} _ {r} = \sum_ {i = 1} ^ {n} \boldsymbol {T} _ {r} ^ {i} \left(\boldsymbol {o} _ {g} ^ {i} \left(1 - \boldsymbol {m} _ {r} ^ {b}\right) \boldsymbol {c} _ {g} ^ {i} + \boldsymbol {o} _ {b} ^ {i} \boldsymbol {m} _ {r} ^ {b} \boldsymbol {c} _ {b} ^ {i}\right), +$$ + +where $\{o_g,o_b\}$ and $\{c_{g},c_{b}\}$ denote the point occupancy and color respectively, which are defined for both clothing and body. $T_r^i$ is the accumulated transparency along the ray. We also render the clothing color $\hat{C}_r^g$ and body color $\hat{C}_r^b$ + +# 3.3. Clothed Human Performance Capture with Double-layer NeRFs + +We estimate the double-layer NeRFs and template deformation of the body and clothing by minimizing the following loss between the synthesized color and observed color: + +$$ +\begin{array}{l} L _ {r g b} = \sum_ {\boldsymbol {r} \in R} \| \hat {\boldsymbol {C}} (\boldsymbol {r}) - \boldsymbol {C} (\boldsymbol {r}) \| _ {2} ^ {2} \tag {9} \\ + \sum_ {\boldsymbol {r} \in \boldsymbol {R}} \| (\hat {\boldsymbol {C}} (\boldsymbol {r}) ^ {b} - \boldsymbol {C} (\boldsymbol {r})) (1 - \boldsymbol {m} _ {r} ^ {g}) \| _ {2} ^ {2} \\ + \sum_ {\boldsymbol {r} \in \boldsymbol {R}} \| (\hat {\boldsymbol {C}} (\boldsymbol {r}) ^ {g} - \boldsymbol {C} (\boldsymbol {r})) (1 - \boldsymbol {m} _ {r} ^ {b}) \| _ {2} ^ {2}, \\ \end{array} +$$ + +where $R$ is the set of casted rays. While using the dense photometric constraint defined in Eq. 9, the model training + +remains ill-posed and body-cloth collisions are prone to occur. To solve these problems, we design several regularization losses to constrain the joint optimization. + +Physics-aware constraint. Cloth simulation is adopted in previous works [16, 40] to obtain realistic performance capture. Different from their traditional simulation method, we introduce a differential network that learns a mapping from human and cloth parameters to physics-aware cloth shapes. We simulate the data in Marvelous Designer using the method of [23]. For each garment category, we use 25 garment styles $\gamma \in \mathbb{R}^4$ as [23]. To account for the effects of different materials, we simulate 8 different materials in Marvelous Designer, and represent the material type $\pmb{\tau}$ using one-hot labels. We train the simulation network $D_{\phi}$ using a MLP to minimize the L1-loss between the simulated cloth meshes and the predicted ones $G_{s} = D_{\phi}(\theta ,\beta ,\gamma ,\tau)$ with the input of human pose and shape, and garment and material styles. The physical simulation loss enforces the deformed cloth mesh $\tilde{G}$ to be close to $G_{s}$ , + +$$ +L _ {s i m} = \sum_ {t \in \boldsymbol {G}} \rho \left(\left\| \tilde {\boldsymbol {G}} ^ {t} - \boldsymbol {G} _ {s} ^ {t} \right\| _ {2} ^ {2}\right), \tag {10} +$$ + +where $t$ denotes vertex index of $G$ , and $\rho$ is the Geman-McClure robust function [5]. The simulation loss can penalize body-cloth collisions and preserve physically plausible deformations. + +As-rigid-as-possible Loss. To make sure the template deforms smoothly, we utilize an as-rigid-as-possible loss [32] on the embedded graph deformation of the human body and clothing, $L_{arap} = L_{arap}(\hat{M}) + L_{arap}(\hat{G})$ . + +Mask Loss. We render the mask image $R(\mathbf{U})$ of the combined mesh $\mathbf{U} = (M, G)$ by using a differentiable renderer [35] based on point cloud, and compute the mask loss with the input human mask image $\bar{R}$ : + +$$ +L _ {\text {m a s k}} (\boldsymbol {U}) = \left\| R (\boldsymbol {U}) - \bar {R} \right\| _ {2} ^ {2}. \tag {11} +$$ + +We utilize the mask loss on both the deformed mesh and the simulated mesh as, $L_{mask} = L_{mask}(\tilde{M},\tilde{G}) + L_{mask}(\tilde{M},G_s)$ . + +Attachment Loss. During human motion, some cloth parts are always attached to the body (e.g., the dress strap is fixed on the shoulder). To ensure this constraint, we define an attachment loss: + +$$ +L _ {\text {a t t a c h}} = \sum_ {i, j \in A} \rho \left(\| \tilde {\boldsymbol {G}} ^ {i} - \tilde {\boldsymbol {M}} ^ {j} \| _ {2} ^ {2}\right), \tag {12} +$$ + +where $A$ is a set of selected vertices on the clothing mesh attached to corresponding points of the body mesh. + +Interpenetration Loss. Interpenetrations may occur between the predicted garments and the body. We use the + +interpenetration term [7] to alleviate this problem: + +$$ +L _ {i n t e r p} \left(M _ {1}, M _ {2}\right) = \frac {1}{| C |} \sum_ {i, j \in C} R e L U \left(\left(M _ {1} ^ {i} - M _ {2} ^ {j}\right) \cdot N _ {1} ^ {i}\right), \tag {13} +$$ + +where $C$ is the set of corresponding points between the mesh $M_{1}$ and mesh $M_2$ , $|C|$ denotes the number of corresponding points, and $N_{1}$ is the normal of mesh $M_{1}$ . We use interpenetration loss in both canonical and posed space, + +$$ +L _ {i n t e r p} = L _ {i n t e r p} (\hat {\boldsymbol {M}}, \hat {\boldsymbol {G}}) + L _ {i n t e r p} (\tilde {\boldsymbol {M}}, \tilde {\boldsymbol {G}}). \tag {14} +$$ + +# 3.3.1 Total Loss Function + +We first learn the geometry network of both the clothing and human body in canonical frame. We use the first frame of multiple views or a video in self-rotating motion to estimate the canonical geometry by optimizing all the losses. Then, by fixing the clothing shape and the geometry network, we track the motion of each video frame and estimate the appearance of NeRFs using the following total loss: + +$$ +\begin{array}{l} L = \lambda_ {1} L _ {r g b} + \lambda_ {2} L _ {a r a p} + \lambda_ {3} L _ {s i m} \tag {15} \\ + \lambda_ {4} L _ {\text {m a s k}} + \lambda_ {5} L _ {\text {a t t a c h}} + \lambda_ {6} L _ {\text {i n t e r p}}, \\ \end{array} +$$ + +where $\{\lambda_1\ldots \lambda_6\}$ are the balancing weights. For a video with $N$ frames, the estimated parameter set $X$ is: + +$$ +X = \left\{\boldsymbol {F} _ {c} ^ {b}, \boldsymbol {F} _ {c} ^ {g}, \boldsymbol {F} _ {d} ^ {b}, \boldsymbol {F} _ {d} ^ {g} \right\} \cup \left\{\varphi_ {i} ^ {b}, \varphi_ {i} ^ {g}, \omega_ {i} ^ {b}, \omega_ {i} ^ {g} \mid i \in 1, \dots , N \right\}, \tag {16} +$$ + +which includes the weights of the appearance and deformation networks and the corresponding conditional variables for each frame in the video. + +# 4. Experiments + +# 4.1. Implementation details + +Datasets and metrics. We design common garments (e.g., dress, T-shirt, skirt, pants, et al.) and simulate the data in the Marvelous Designer [1] under various human shapes and poses and cloth materials. We test our method on publicly datasets, DynaCap [8], and DeepCap [10]. We use four cameras around the human for training and the remaining cameras for test. Figure 2 shows some recovered results using our approach on different clothed humans. Our method is also tested on a monocular video in outdoor scene. Please refer to the supplementary video for more experimental results. The geometry accuracy is evaluated with the intersection over union (IoU) between the ground truth human masks and 2D projections of the recovered human shapes. Following [10, 34], the IoU (\%) is computed on all views (AMVIoU), all views except the input view (RVIoU), and the input view (SVIoU). The novel-view synthesis quality is evaluated with two metrics, i.e., peak signal-to-noise ratio (PSNR) and structural similarity index (SSIM) as [24, 34]. + +![](images/8a0fc36af3de34eb7391906c13c0ef74a624df0294341d8ad22b62a2d3895e1c.jpg) +Figure 2. The results of our approach on different humans in various garments. In each example, we show the ground truth image in a novel view, the reconstructed geometry (overlayed on the image), and the novel-view synthesis image. From left to right: the S4 from DeepCap dataset [10], "FranziRed" from DynaCap dataset [8], and the S1 from DeepCap dataset [10]. + +Architecture and experimental settings. We adopt the S-DF network from [30], and the same occupancy and color network as [34]. In the non-rigid deformation network, there are 8 layers with a skip connection at the fourth layer and the hidden size is 1024. The deformation and appearance latent codes both have dimensions of 128. We set both weight $\mu_{1}$ and $\mu_{2}$ to 1. We set the loss weight $\{\lambda_1,\dots,\lambda_6\}$ to $\{1.0,0.1,0.02,50,0.1,100.0\}$ empirically. The Adam optimizer [14] is used in the optimization, and the learning rate is $5e^{-4}$ at the start of training and exponentially decays to $5e^{-5}$ during the training. With a GPU of RTX 2080 Ti, our network training costs about 12 hours tested on a four-view video with 300 frames. + +# 4.2. Comparison to state-of-the-art methods + +We compare our method with two state-of-the-art approaches for human performance capture, DeepCap [10] and NerfCap [34]. We also compare with BCNet [12], TailorNet [23], and ICON [36]. For DeepCap and NerfCap, the experimental results are from [34]. For other compared methods, we implement them using their released codes. We do not compare with DeepCloth [16] since their codes are not released and DeepCap [10] outperforms them in ge + +
MethodAMVIoURVIoUSVIoU
DeepCap [10]86.1785.7586.90
NerfCap [34]88.9688.3290.09
ICON [36]--93.56
BCNet [12]82.3381.6782.94
TailorNet [23]83.8783.5084.51
Our method89.5888.7591.03
+ +Evaluated on the subject S4 from DeepCap dataset [10]. + +
MethodAMVIoURVIoUSVIoU
DeepCap [10]85.3885.3785.41
NerfCap [34]88.0888.0788.12
ICON [36]--95.21
BCNet [12]78.9178.7780.83
TailorNet [23]78.1578.0078.21
Our method88.7488.4089.68
+ +Evaluated on the 'FranziRed' from DynaCap dataset [8]. + +Table 1. Comparison to the state-of-the-arts in surface reconstruction accuracy. Note that, the accuracy of ICON [36] is high since they use the human mask in estimating the geometry, and their dress shapes are not recovered accurately as shown in Fig. 3. + +ometry tracking as reported in [16]. We conduct the qualitative and quantitative experiments on S4 of DeepCap and + +"FranziRed" from DynaCap. The test dataset are the same as [34] which include 300 - 400 frames randomly sampled from the original dataset. + +![](images/88b1a4cb9af775bb7d201b363d90d292d84745d0ef7b38652693afd627517315.jpg) +(a) (b) +# +(c) +(d) +(e) +(f) +(g) +Figure 3. Qualitative comparison with the state-of-the-arts. (a) Ground truth image. (b) DeepCap [10]. (c) NerfCap [34]. (d) ICON [36]. (e) BCNet [12]. (f) TailorNet [23]. (g) Our method. + +We show the qualitative comparisons in Figure 3 and list the quantitative results in Table 1. We only compute the error of the input view for ICON since they use their own camera system. Our method performs significantly better than the compared methods [12,23,36] without a cloth template. BCNet [12] also separates the modeling of the clothing and body, but as a supervised learning method, the performance may degrade dramatically because of data domain gap. When the human is moving, BCNet fails to recover the dresses accurately due to arbitrary cloth deformation. In contrast, the appearance loss of our double-layer NeRF-s poses dense constraints on the clothing and body motion capture in a weakly-supervised manner. As a result, our method can achieve a much higher accuracy of geometry tracking even without the ground truth. The predicted garments of TailorNet [23] are not consistent with the real images since their model is conditioned on human poses and shapes. ICON obtains higher IoU accuracy by using the human mask in recovering the geometry. However, there are obvious artifacts on the recovered dresses of ICON [36]. Based on one-piece personalized template, it is hard for both DeepCap [10] and NerfCap [34] to extract the clothing motion, which is important for downstream applications in AR/VR. In contrast, our method tracks the motion for both the clothing and body. Without a pre-scanned cloth template, our method still obtains a high tracking accuracy + +
MethodS4 of [10]‘FranziRed” of [8]
PSNRSSIMPSNRSSIM
NerfCap [34]24.5290.7923.1891.49
Our method23.0690.2821.7090.54
+ +Table 2. Comparison to NerfCap [34] in terms of novel-view synthesis on two datasets. + +
MethodAMVIoUPSNRSSIM
w/o simulation loss89.0322.6689.35
w/o mask loss88.7523.2189.49
w/o color loss89.36--
blending render [6]88.7523.6490.51
1 camera view88.4921.3487.12
2 camera views89.2321.8988.38
6 camera views90.1023.8090.60
ours(4 views)89.5823.0690.28
+ +Table 3. Ablation studies on S4 from DeepCap dataset [10]. + +thanks to the implicit representation. + +We also compare the quality of free-viewpoint synthesis with NerfCap [34]. Figure 4 shows two compared examples and Table 2 reports the quantitative comparison. Without using any clothing template, our method still obtains close or comparable synthesis quality to NerfCap that uses a ground truth template. Our results are rendered by composing the NeRFs of the body and clothing, which is more complicated than a single NeRFs of NerfCap. For example, when the mutual motion occurs between the body and clothing (like cloth sliding), new body parts may appear, leading to the decrease of our synthesis quality. + +![](images/b31ed73df970870e0ac60c671fb255b4b56a6dec0a49dcb91bc50f4c4e39e3ad.jpg) +(a) +Figure 4. Comparison of novel-view synthesis with NerfCap [34]. (a,d) Ground truth. (b,e) NerfCap [34]. (c,f) Our method. + +![](images/4b6c2bf4ed8e4d070e86dd39d566ddb6a26ddc5765438fc0fd4defb254a15aab.jpg) +(b) + +![](images/a25fb15d5ad67a3b9a172f16580039cd498fa17f3bae6ab361029b724c0a5e22.jpg) +(c) + +![](images/9db08b2b3e6b76f75dd2bce570b5b16060b886a071774cef6c4b96c8e6b6dfb6.jpg) +(d) + +![](images/cd2c2cea695505c9a22d825c8c53bf9392a26ef91e14bbdb6e0c842d7fa094d2.jpg) +(e) + +![](images/a4ea041c9b8c5221d0d53fa4a1e91d258adf299ff1937c3ff6548ee5a3325bed.jpg) +(f) + +# 4.3. Ablation studies + +The ablation experiments are performed to demonstrate the effectiveness of our framework on the S4 from DeepCap dataset [10]. The quantitative results are listed in Table 3. + +Effectiveness of different losses. We first evaluate the effectiveness of the main losses in our optimization, i.e., the simulation loss, the mask loss, and the color loss. We report the accuracy by removing each of them, and show qualitative comparison examples in Figure 5. Without the simulation loss, the body-cloth collision becomes severe, showing that the simulation loss can preserve physically-plausible body-cloth interactions. Using the input human mask can + +improve the alignment accuracy with the human silhouettes. The color loss of dynamic NeRF can further enhance the detailed deformation on the human surface by posing dense appearance constraints for the geometry tracking. + +![](images/2cafdd527f3a2550598e5a1ade66938596b8b9596ec78c10737c0b37c78c0f11.jpg) + +![](images/4d736e527bb28a2679c32af1379f3c4f5499cb79e9066a309d145f9ee5ace126.jpg) + +![](images/1766e4d22506c87b72375e406ff51d7835b77fd9b23b1b9a2afa8852aba8d93a.jpg) + +![](images/c79b7ed46ec3a8a1bafa4155948909b1c12191c4f583d9a179b50a9c545e5544.jpg) + +![](images/e7bed27739caa1a681482e887dc055a37127764a71a461e8ff4cab7f6f296a95.jpg) + +![](images/1724da4f1adf14bc96d2506cd00674659143a08ffed2de310f3e4e9c25ebeb36.jpg) +(a) + +![](images/aa1efa9baa10faff6110740345a60648ac0dccb7f31a35de16fc3dbd19552feb.jpg) +(b) + +![](images/d8231ca330ce52dd0abb612881dbc9a4ca8db648163bd6e74e2ab55f7807d1fb.jpg) +(c) + +![](images/18e9d9f5c403d40fe372bae16fc436608f3be6fa451955b4702ca51345da250f.jpg) +(d) + +![](images/7c308394e2760380e490274e723d9cac1c30133251f4fcce88de743707d497db.jpg) + +![](images/79ee651b5e22c62d990dd3556beff6d0154366dca669112e480546541eeba10d.jpg) +Figure 5. Ablation study for different losses. (a) Ground truth. (b) Using complete losses. (c) W/o simulation loss. (d) W/o mask loss. (e) W/o color loss. There are unnatural cloth deformations without the simulation loss. The rendered mask cannot align with the silhouette well without the mask loss. Compared to that without color loss, the surface detailed deformation can be further improved using complete losses (b). +(a) + +![](images/5e97de6abcc6dd2fd0494bb139acfce978e0a9275dc0fe761e051e706832b6f3.jpg) +(b) + +![](images/c152489fa55ea4a2a8c57e0c7574c55aef8536ad8288d2b2217715f51ab727e0.jpg) +Figure 6. Comparison of different rendering. (a) Ground truth. (b) Our composite rendering. (c) Blend rendering [6]. We show the synthesis of the clothing, body, and full human. There are much cloth appearance on the body caused by the blending [6], while our method separates the appearance of the clothing and body better. + +![](images/0be68f33d69b1ca701cedbe9cb6c1297bddd553ac882b75fddd130432b06c198.jpg) + +![](images/64ffa03747a1eda5401011102a27b6b127ce8c51d2d269a3097e0e9bef886d47.jpg) + +![](images/52a19c0ba71493b4000011081db876501e483239da9460d86251f195d1db8658.jpg) +(c) + +![](images/5aadb97f34f34b404da349258615dfd2c46bc7e5b4b789249abd4ab597f21b12.jpg) + +Effectiveness of the composite rendering. We compare our composite rendering with the blending rendering in [6]. Through free composition of two NeRFs for each point, the method of [6] can obtain high fitting color. However, the blending may generate missing areas or artifacts in the novel-view synthesis. Moreover, it fails to recover the appearance of the body and clothing reliably as shown in Figure 6. Comparatively, we can recover the appearance of the body and clothing separately and compose them into a full human more accurately using the composite rendering. + +The number of camera views. We investigate the influence of different number of input camera views on the performance of our approach. We compare the results of our method under $\{1,2,4,6\}$ views, and show the comparison + +example of novel-view synthesis in Figure 7. The results become slightly better using more camera views because the moving humans are captured more completely. + +![](images/d34100012af1e0ae21a4d54372f1f017c0e088b6729939077a77d3d7391303ff.jpg) +Figure 7. Ablation study for different numbers of camera views. With more cameras, the novel-view synthesis quality becomes better (marked in cycle). + +![](images/d43466951f76608d0ce7b92136b54930a7814df294209549227d43b1c06c6b0c.jpg) +Ground Truth +1 view + +![](images/61a22e02dc1f226e377f5244f301979de0d86d0ee3925d60d2b0bfb408662ad4.jpg) +2 views + +![](images/f8cbdcbc3c5ae4e5a2719bb517038dfd695e299c91495ddf922d5cf52fb223af.jpg) +4 views + +![](images/9da0dea23bbf11da7f27bf2335f933d2b1f86d3f758d49e35bf1e02ff5ee8b20.jpg) +6 views + +# 4.4. Applications + +Benefiting from separate tracking results of human body and clothing, we can enable interesting applications like cloth retargeting as shown in Figure 8. Since the body mesh and cloth mesh are captured, we can exchange the garments between two person by simulating the garments on the other body. Moreover, free-viewpoint motion can be created by simulating the clothing on human body sequences. + +![](images/d2a02633a71bc2f21ee5194b63edd462afd85e4c557463deea43581dba62d4be.jpg) +Figure 8. Application of cloth retargeting. From the captured results using our method, we can retarget the garments between the S4 of DeepCap and the 'FranziRed' of DynaCap. + +![](images/04e8d1615ac71b0c66e619a82bed89d58395189f9065a4bb4a2345125442166e.jpg) + +![](images/a783863e72fc48f9fb39e39b52c50ce68a20bea0a10d9452f9e4c4b6bacce24f.jpg) + +![](images/d48f036eb9b4af9b10453d785a9649e0b97c099067737d9949a0adef63d2eea0.jpg) + +![](images/1b1772b24e41c5a1d224b24850638f1b56d14c5ebe4c2dd63961fbe11dd3f3a0.jpg) + +![](images/7c03eb5680cda026af5bc8f3493b915476c20c3aa1a02398bd4071445cad4686.jpg) + +# 5. Conclusion + +In this paper, we proposed a novel performance capture method for clothed humans with a double-layer neural radiance fields (NeRF). Our double-layer NeRFs can effectively model moving humans in daily-life clothes. By matching the rendered color with the observed color, our method estimates the clothing and body motion and temporally-varying appearance robustly. In addition, we learn a differential physics-aware network for different garment types which supports physically-plausible body-cloth interactions and cloth dynamics. The experimental results on real datasets demonstrate the effectiveness of the proposed method. + +Acknowledgments This work was supported by the Fundamental Research Funds for the Central Universities (NJ2022028), and the Natural Science Foundation of China (No.U1713208), and Program for Changjiang Scholars. + +# References + +[1] https://www.marvelousdesigner.com/.3,5 +[2] Thiemo Alldieck, Gerard Pons-Moll, Christian Theobalt, and Marcus Magnor. Tex2Shape: Detailed full human body geometry from a single image. In ICCV, 2019. 2 +[3] Bharat Lal Bhatnagar, Garvita Tiwari, Christian Theobalt, and Gerard Pons-Moll. Multi-Garment Net: Learning to dress 3D people from images. In ICCV, 2019. 1, 2 +[4] Mingsong Dou, Sameh Khamis, Yury Degtyarev, Philip Davidson, Sean Fanello, Adarsh Kowdle, Sergio Orts Escolano, Christoph Rhemann, David Kim, Jonathan Taylor, Pushmeet Kohli, Vladimir Tankovich, and Shahram Izadi. Fusion4D: Real-time performance capture of challenging scenes. ACM TOG, 35:1-13, July 2016. 1, 2 +[5] Stuart Ganan and D McClure. Bayesian image analysis: An application to single photon emission tomography. In Proceedings of the American Statistical Association, pages 12-18, 1985. 5 +[6] Chen Gao, Ayush Saraf, Johannes Kopf, and Jia-Bin Huang. Dynamic view synthesis from dynamic monocular video. In ICCV, 2021. 4, 7, 8 +[7] Erhan Gundogdu, Victor Constantin, Amrollah Seifoddini, Minh Dang, Mathieu Salzmann, and Pascal Fua. GarNet: A two-stream network for fast and accurate 3d cloth draping. In ICCV, 2019. 5 +[8] Marc Habermann, Lingjie Liu, Weipeng Xu, Michael Zollhoefer, Gerard Pons-Moll, and Christian Theobalt. Real-time deep dynamic characters. ACM Transactions on Graphics, 40(4), August 2021. 2, 5, 6, 7 +[9] Marc Habermann, Weipeng Xu, Michael Zollhoefer, Gerard Pons-Moll, and Christian Theobalt. LiveCap: Real-time human performance capture from monocular video. ACM TOG, 38(2):14:1-14:17, March 2019. 1, 2 +[10] Marc Habermann, Weipeng Xu, Michael Zollhoefer, Gerard Pons-Moll, and Christian Theobalt. DeepCap: Monocular human performance capture using weak supervision. In CVPR, 2020. 1, 2, 5, 6, 7 +[11] Boyi Jiang, Yang Hong, Hujun Bao, and Juyong Zhang. SelfRecon: Self reconstruction your digital avatar from monocular video. In CVPR, 2022. 2 +[12] Boyi Jiang, Juyong Zhang, Yang Hong, Jinhao Luo, Ligang Liu, and Hujun Bao. BCNet: Learning body and cloth shape from a single image. In ECCV, 2020. 1, 2, 6, 7 +[13] Hanbyul Joo, Tomas Simon, and Yaser Sheikh. Total capture: A 3D deformation model for tracking faces, hands, and bodies. In CVPR, 2018. 4 +[14] Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In International Conference for Learning Representations, 2015. 6 +[15] John P Lewis, Matt Cordner, and Nickson Fong. Pose space deformation: A unified approach to shape interpolation and skeleton-driven deformation. In SIGGRAPH, 2000. 3, 4 +[16] Yue Li, Marc Habermann, Bernhard Thomaszewski, Stelian Coros, Thabo Beeler, and Christian Theobalt. Deep physics-aware inference of cloth deformation for monocular human performance capture. In International Conference on 3D Vision (3DV), 2021. 1, 2, 5, 6 + +[17] Lingjie Liu, Marc Habermann, Viktor Rudnev, Kripasindhu Sarkar, Jiatao Gu, and Christian Theobalt. Neural Actor: Neural free-view synthesis of human actors with pose control. ACM TOG, 40(6):1-16, 2021. 3 +[18] Matthew Loper, Naureen Mahmood, Javier Romero, Gerard Pons-Moll, and Michael J. Black. SMPL: A skinned multiperson linear model. ACM TOG, 34(6):248, 2015. 2, 3 +[19] Qianli Ma, Jinlong Yang, Anurag Ranjan, Sergi Pujades, Gerard Pons-Moll, Siyu Tang, and Michael J. Black. Learning to dress 3D people in generative clothing. In CVPR, 2020. 2 +[20] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. N-eRF: Representing scenes as neural radiance fields for view synthesis. In ECCV, 2020. 1, 2, 4 +[21] Richard Newcombe, Dieter Fox, and Steve Seitz. Dynamic-Fusion: Reconstruction and tracking of non-rigid scenes in real-time. In CVPR, 2015. 1, 2 +[22] Michael Oechsle, Songyou Peng, and Andreas Geiger. U-NISURF: Unifying neural implicit surfaces and radiance fields for multi-view reconstruction. In ICCV, 2021. 3, 4 +[23] Chaitanya Patel, Zhouyingcheng Liao, and Gerard Pons-Moll. TailorNet: Predicting clothing in 3D as a function of human pose, shape and garment style. In CVPR, 2020. 2, 5, 6, 7 +[24] Sida Peng, Junting Dong, Qianqian Wang, Shangzhan Zhang, Qing Shuai, Xiaowei Zhou, and Hujun Bao. Animatable neural radiance fields for modeling dynamic human bodies. In ICCV, 2021. 2, 3, 4, 5 +[25] Sida Peng, Yuanqing Zhang, Yinghao Xu, Qianqian Wang, Qing Shuai, Hujun Bao, and Xiaowei Zhou. Neural Body: Implicit neural representations with structured latent codes for novel view synthesis of dynamic humans. In CVPR, 2021. 2, 3, 4 +[26] Gerard Pons-Moll, Sergi Pujades, Sonny Hu, and Michael J. Black. ClothCap: Seamless 4D clothing capture and retargeting. ACM TOG, 36(4), July 2017. 1, 2 +[27] Shunsuke Saito, Zeng Huang, Ryota Natsume, Shigeo Morishima, Angjoo Kanazawa, and Hao Li. PIFu: Pixel-aligned implicit function for high-resolution clothed human digitization. In ICCV, 2019. 2 +[28] Shunsuke Saito, Tomas Simon, Jason Saragih, and Hanbyul Joo. PIFuHD: Multi-level pixel-aligned implicit function for high-resolution 3D human digitization. In CVPR, 2020. 2 +[29] Igor Santesteban, Miguel A. Otaduy, and Dan Casas. Learning-based animation of clothing for virtual try-on. Computer Graphics Forum (Proc. Eurographics), 2019. 2 +[30] Vincent Sitzmann, Julien NP Martel, Alexander W Bergman, David B Lindell, and Gordon Wetzstein. Implicit neural representations with periodic activation functions. In Advances in Neural Information Processing Systems, 2020. 4, 6 +[31] Zhaoqi Su, Weilin Wan, Tao Yu, Lingjie Liu, Lu Fang, Wenping Wang, and Yebin Liu. MulayCap: Multi-layer human performance capture using a monocular video camera. IEEE Transactions on Visualization and Computer Graphics, 28(4):1862-1879, 2022. 1, 2 + +[32] Robert W. Sumner, Johannes Schmid, and Mark Pauly. Embedded deformation for shape manipulation. ACM TOG, 26(3), 2007. 3, 5 +[33] Gül Varol, Duygu Ceylan, Bryan Russell, Jimei Yang, Ersin Yumer, Ivan Laptev, and Cordelia Schmid. BodyNet: Volumetric inference of 3D human body shapes. In ECCV, 2018. 2 +[34] Kangkan Wang, Sida Peng, Xiaowei Zhou, Jian Yang, and Guofeng Zhang. NerfCap: Human performance capture with dynamic neural radiance fields. IEEE Transactions on Visualization and Computer Graphics, 2022. 1, 3, 4, 5, 6, 7 +[35] Olivia Wiles, Georgia Gkioxari, Richard Szeliski, and Justin Johnson. Synsin: End-to-end view synthesis from a single image. In CVPR, pages 7467-7477, 2020. 5 +[36] Yuliang Xiu, Jinlong Yang, Dimitrios Tzionas, and Michael J. Black. ICON: Implicit clothed humans obtained from normals. In CVPR, June 2022. 2, 6, 7 +[37] Weipeng Xu, Avishek Chatterjee, Michael Zollhöfer, Helge Rhodin, Dushyant Mehta, Hans-Peter Seidel, and Christian Theobalt. MonoPerfCap: Human performance capture from monocular video. ACM TOG, 37(2):1-15, July 2018. 1, 2 +[38] Tao Yu, Kaiwen Guo, Feng Xu, Yuan Dong, Zhaqi Su, Jianhui Zhao, Jianguo Li, Qionghai Dai, and Yebin Liu. Body-Fusion: Real-time capture of human motion and surface geometry using a single depth camera. In ICCV, October 2017. 1, 2 +[39] Tao Yu, Zerong Zheng, Kaiwen Guo, Jianhui Zhao, Qionghai Dai, Hao Li, Gerard Pons-Moll, and Yebin Liu. DoubleFusion: Real-time capture of human performances with inner body shapes from a single depth sensor. In CVPR, June 2018. 1, 2 +[40] Tao Yu, Zerong Zheng, Yuan Zhong, Jianhui Zhao, Qionghai Dai, Gerard Pons-Moll, and Yebin Liu. SimulCap: Single-view human performance capture with cloth simulation. In CVPR, 2019. 1, 2, 5 +[41] Fuqiang Zhao, Wei Yang, Jiakai Zhang, Pei Lin, Yingliang Zhang, Jingyi Yu, and Lan Xu. HumanNeRF: Generalizable neural human radiance field from sparse inputs. In CVPR, 2022. 1, 3 +[42] Zerong Zheng, Tao Yu, Yixuan Wei, Qionghai Dai, and Yebin Liu. DeepHuman: 3D human reconstruction from a single image. In ICCV, 2019. 2 +[43] Heming Zhu, Yu Cao, Hang Jin, Weikai Chen, Dong Du, Zhangye Wang, Shuguang Cui, and Xiaoguang Han. Deep Fashion3D: A dataset and benchmark for 3d garment reconstruction from single images. In ECCV, pages 512-530, 2020. 1, 2 +[44] Heming Zhu, Lingteng Qiu, Yuda Qiu, and Xiaoguang Han. Registering Explicit to Implicit: Towards high-fidelity garment mesh reconstruction from single images. In CVPR, pages 3845-3854, 2022. 1, 2 +[45] Hao Zhu, Xinxin Zuo, Sen Wang, Xun Cao, and Ruigang Yang. Detailed human shape estimation from a single image by hierarchical mesh deformation. In CVPR, 2019. 2 \ No newline at end of file diff --git a/2023/Clothed Human Performance Capture With a Double-Layer Neural Radiance Fields/images.zip b/2023/Clothed Human Performance Capture With a Double-Layer Neural Radiance Fields/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..f34f20c49caa4f83f51d81bbb17ed2a337730366 --- /dev/null +++ b/2023/Clothed Human Performance Capture With a Double-Layer Neural Radiance Fields/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c5704a4145e80e9b892cd48d513379e2317b9d8d3d9b8ef01a9128f446fdf430 +size 638634 diff --git a/2023/Clothed Human Performance Capture With a Double-Layer Neural Radiance Fields/layout.json b/2023/Clothed Human Performance Capture With a Double-Layer Neural Radiance Fields/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..6e117c2b257b7ac78fdffd0fe9d0a45b01dea67d --- /dev/null +++ b/2023/Clothed Human Performance Capture With a Double-Layer Neural Radiance Fields/layout.json @@ -0,0 +1,10192 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 60, + 136, + 517, + 169 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 60, + 136, + 517, + 169 + ], + "spans": [ + { + "bbox": [ + 60, + 136, + 517, + 169 + ], + "type": "text", + "content": "Clothed Human Performance Capture with a Double-layer Neural Radiance Fields" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 192, + 437, + 208 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 192, + 437, + 208 + ], + "spans": [ + { + "bbox": [ + 138, + 192, + 437, + 208 + ], + "type": "text", + "content": "Kangkan Wang" + }, + { + "bbox": [ + 138, + 192, + 437, + 208 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 138, + 192, + 437, + 208 + ], + "type": "text", + "content": ", Guofeng Zhang" + }, + { + "bbox": [ + 138, + 192, + 437, + 208 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 138, + 192, + 437, + 208 + ], + "type": "text", + "content": ", Suxu Cong" + }, + { + "bbox": [ + 138, + 192, + 437, + 208 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 138, + 192, + 437, + 208 + ], + "type": "text", + "content": ", Jian Yang" + }, + { + "bbox": [ + 138, + 192, + 437, + 208 + ], + "type": "inline_equation", + "content": "^{1,2}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 211, + 556, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 211, + 556, + 239 + ], + "spans": [ + { + "bbox": [ + 46, + 211, + 556, + 239 + ], + "type": "text", + "content": "1Key Lab of Intelligent Perception and Systems for High-Dimensional Information of Ministry of Education, 2Jiangsu Key Lab of Image and Video Understanding for Social Security," + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 64, + 239, + 538, + 267 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 239, + 538, + 267 + ], + "spans": [ + { + "bbox": [ + 64, + 239, + 538, + 267 + ], + "type": "text", + "content": "School of Computer Science and Engineering, Nanjing University of Science and Technology, China \n3State Key Laboratory of CAD&CG, Zhejiang University, China" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 139, + 287, + 186, + 299 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 139, + 287, + 186, + 299 + ], + "spans": [ + { + "bbox": [ + 139, + 287, + 186, + 299 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 45, + 312, + 281, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 312, + 281, + 604 + ], + "spans": [ + { + "bbox": [ + 45, + 312, + 281, + 604 + ], + "type": "text", + "content": "This paper addresses the challenge of capturing performance for the clothed humans from sparse-view or monocular videos. Previous methods capture the performance of full humans with a personalized template or recover the garments from a single frame with static human poses. However, it is inconvenient to extract cloth semantics and capture clothing motion with one-piece template, while single frame-based methods may suffer from instable tracking across videos. To address these problems, we propose a novel method for human performance capture by tracking clothing and human body motion separately with a double-layer neural radiance fields (NeRFs). Specifically, we propose a double-layer NeRFs for the body and garments, and track the densely deforming template of the clothing and body by jointly optimizing the deformation fields and the canonical double-layer NeRFs. In the optimization, we introduce a physics-aware cloth simulation network which can help generate physically plausible cloth dynamics and body-cloth interactions. Compared with existing methods, our method is fully differentiable and can capture both the body and clothing motion robustly from dynamic videos. Also, our method represents the clothing with an independent NeRFs, allowing us to model implicit fields of general clothes feasibly. The experimental evaluations validate its effectiveness on real multi-view or monocular videos." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 619, + 124, + 631 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 619, + 124, + 631 + ], + "spans": [ + { + "bbox": [ + 46, + 619, + 124, + 631 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 45, + 639, + 280, + 710 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 639, + 280, + 710 + ], + "spans": [ + { + "bbox": [ + 45, + 639, + 280, + 710 + ], + "type": "text", + "content": "Performance capture for clothed humans is one of the essential problems in the metaverse, and it not only captures the inner human body motion but also recovers the outer clothing motion which has many promising applications such as virtual try-on, video editing, and telepresence. From sparse-view or monocular videos of a moving human" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 297, + 288, + 531, + 383 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 288, + 531, + 383 + ], + "spans": [ + { + "bbox": [ + 297, + 288, + 531, + 383 + ], + "type": "text", + "content": "in general clothes, its goal is to recover the dynamic 3D shape sequence of the human body and clothing simultaneously that are consistent with the observed frames in both human shape and motion. This is a very challenging problem since the dynamic human could be with arbitrary motions and with complex non-rigid cloth deformations, and the clothing in motion is difficult to maintain physically plausible interactions with the body." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 296, + 386, + 531, + 632 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 386, + 531, + 632 + ], + "spans": [ + { + "bbox": [ + 296, + 386, + 531, + 632 + ], + "type": "text", + "content": "Previous systems [4, 21, 38, 39] reconstruct 3D clothed humans by using depth sensors or fitting a personalized template [9, 10, 37] to the image observations (e.g., body joints and silhouettes). Only recovering one-piece geometry which unifies the human body and clothes, these systems fail to track the motion of the clothing and achieve clothing editing on 3D humans, which are the prerequisites in many VR/AR applications like virtual dressing. On the contrary, cloth can be extracted and tracked from depth scans [26, 40] accurately by fitting pre-built cloth templates to the scans which have limited applications when 3D data are unavailable. Existing garment estimation methods [3, 12, 43, 44] from color images require the person facing the camera and in static poses. When the human is moving and the clothing is deforming, these methods may recover the 3D garments unreliable. Recent methods [16, 31] track body and clothing motion simultaneously from videos, but they need to re-build cloth template for a new performer and the running efficiency is very low due to online cloth simulation or computationally-exhaustive optimization, which prohibits them from being widely deployed for daily applications." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 296, + 636, + 531, + 730 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 636, + 531, + 730 + ], + "spans": [ + { + "bbox": [ + 296, + 636, + 531, + 730 + ], + "type": "text", + "content": "Recent works [20, 34, 41] adopt dynamic human NeRFs to capture human motion and obtain impressive tracking results. By capturing the temporally-varying appearance in the videos, dynamic NeRFs [34] can provide dense photometric constraints to track the deforming geometry of the performer. However, they represent the human with a single NeRFs without modeling cloth, and the clothing motion cannot be extracted. In this paper, we aim to track the cloth" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 77, + 42, + 119, + 72 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 42, + 119, + 72 + ], + "spans": [ + { + "bbox": [ + 77, + 42, + 119, + 72 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 150, + 41, + 471, + 74 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 150, + 41, + 471, + 74 + ], + "spans": [ + { + "bbox": [ + 150, + 41, + 471, + 74 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 719, + 224, + 730 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 719, + 224, + 730 + ], + "spans": [ + { + "bbox": [ + 55, + 719, + 224, + 730 + ], + "type": "text", + "content": "*Corresponding author: wangkangkan@njust.edu.cn" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 276, + 746, + 301, + 756 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 276, + 746, + 301, + 756 + ], + "spans": [ + { + "bbox": [ + 276, + 746, + 301, + 756 + ], + "type": "text", + "content": "21098" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 595, + 842 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 45, + 106, + 279, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 106, + 279, + 175 + ], + "spans": [ + { + "bbox": [ + 45, + 106, + 279, + 175 + ], + "type": "text", + "content": "ing and body motion simultaneously with dynamic NeRFs. However, this problem is rather challenging due to two major questions we need to solve: how to represent dynamic clothing and human body with NeRFs, and how to capture clothing and human body motion with plausible body-cloth interactions based on the implicit representation." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 45, + 176, + 280, + 513 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 176, + 280, + 513 + ], + "spans": [ + { + "bbox": [ + 45, + 176, + 280, + 513 + ], + "type": "text", + "content": "In this paper, we propose a novel method for clothed human performance capture with a double-layer NeRFs. Specifically, a double-layer NeRFs is modeled for both the body and clothing in the template space, and transformed to the observation space with the corresponding deformation fields, and the rendered images are then synthesized by composing the two dynamic NeRFs. We first estimate the template shape in canonical frame and learn the geometry network supervised by the template geometry. In the rendering, we compose the double-layer NeRFs with the guidance of deformed body and clothing meshes. Then, by minimizing the difference between the rendered color and observed color, the deformation fields and the canonical N-NeRFs are optimized jointly. The deformation field is represented as the inverse deformation of the template mesh, thus the densely deforming geometry of the template can be recovered simultaneously. In addition, we adopt a physics-aware network learnt from simulation data between various cloth types and humans to constrain the dynamic clothing and preserve physically plausible body-cloth interactions, resulting in realistic cloth geometry tracking. Compared to previous methods, our method is fully differentiable and can recover the realistic motion of both the clothing and body from dynamic videos with arbitrary human poses and complex cloth deformations. The experimental qualitative and quantitative results on datasets of DynaCap [8] and DeepCap [10] prove that the proposed approach can robustly and accurately capture the motion for clothed humans. In summary, the primary contributions of our work include:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 518, + 279, + 658 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 55, + 518, + 279, + 553 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 518, + 279, + 553 + ], + "spans": [ + { + "bbox": [ + 55, + 518, + 279, + 553 + ], + "type": "text", + "content": "- We propose a double-layer NeRFs for dynamic humans in general clothing, allowing us to model implicit humans with a variety of clothes (e.g., loose dresses)." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 559, + 279, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 559, + 279, + 617 + ], + "spans": [ + { + "bbox": [ + 55, + 559, + 279, + 617 + ], + "type": "text", + "content": "- To the best of our knowledge, we propose the first framework to capture clothing motion separately from the human body using the double-layer NeRFs, which provides dense appearance constraints on the geometry tracking and improves the robustness and accuracy." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 623, + 279, + 658 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 623, + 279, + 658 + ], + "spans": [ + { + "bbox": [ + 55, + 623, + 279, + 658 + ], + "type": "text", + "content": "- A differentiable physics-aware network is learnt for d-different common garments and used to preserve physically plausible cloth deformations in motion capture." + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 46, + 667, + 130, + 679 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 667, + 130, + 679 + ], + "spans": [ + { + "bbox": [ + 46, + 667, + 130, + 679 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 45, + 682, + 279, + 730 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 682, + 279, + 730 + ], + "spans": [ + { + "bbox": [ + 45, + 682, + 279, + 730 + ], + "type": "text", + "content": "Clothed human reconstruction. Depth-based methods [4,21,38,39] reconstruct 3D humans in realtime by fusing the depth maps in a volumetric way. However, depth-based methods cannot synthesize photo-realistic videos and" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 296, + 106, + 530, + 327 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 106, + 530, + 327 + ], + "spans": [ + { + "bbox": [ + 296, + 106, + 530, + 327 + ], + "type": "text", + "content": "have constrained applications in daily life due to data capture limitations, e.g., high power consumption and sensitive to sunlight. Based on deep learning techniques, some approaches predict 3D humans from color images through implicit function [27, 28], coarse-to-fine mesh deformation [45], surface normal optimization [36, 42], or UV mapping [2]. Although these methods can reconstruct human details, the recovered shapes do not contain human semantics like human pose, shape and clothing. To recover the clothing, some methods [3, 19] add displacements on vertices of SMPL model [18] which only support tight clothing and recover coarse-level geometry. Template-based methods [9, 10, 37] recover 3D human motions by non-rigidly fitting a person-specific template mesh to 2D observed features (e.g., human joints or silhouettes). With a personalized template, these methods obtain impressive performance capture. However, the recovered meshes integrate the body and clothing in one piece, which hinders their applications like garment retargeting and virtual trying-on." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 296, + 330, + 530, + 598 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 330, + 530, + 598 + ], + "spans": [ + { + "bbox": [ + 296, + 330, + 530, + 598 + ], + "type": "text", + "content": "Clothing motion capture. Traditional methods [26, 40] capture the cloth motion by precisely fitting a cloth template to 3D data from depth sensors, which cannot generalize to the video-based setting and are limited to applications in indoor scenes. Deep learning-based approaches [3, 12, 43, 44] estimate the 3D garment shape from color images. However, these methods require the human facing the camera with a static pose and do not try temporally tracking of dynamic garments. Some works [23, 29] predict the cloth shape conditioned on the human pose and shape but the shapes are not consistent with the observed data. Recent works [16, 31] capture both human body and clothing motion from dynamic videos. MulayCap [31] optimizes a simulated garment on SMPL model with semantic cloth segmentation, but their traditional simulation and optimization are both computationally expensive. DeepCloth [16] incorporates a physics-based simulator to supervise the cloth deformations estimated with multi-view joints and silhouettes as DeepCap [10]. However, their simulation layer is not differentiable and needs to adjust the cloth material types online, which is a tedious and time consuming step. Moreover, with the same limitation as DeepCap, they cannot recover the detailed cloth deformation based on the sparse features." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 296, + 601, + 530, + 730 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 601, + 530, + 730 + ], + "spans": [ + { + "bbox": [ + 296, + 601, + 530, + 730 + ], + "type": "text", + "content": "Implicit Human Reconstruction. Clothed humans can be reconstructed through implicit representation-based methods such as voxel representation [33, 42], implicit function [27, 28], or neural radiance fields (NeRFs) [20, 24, 25]. Recent works combine explicit and implicit representations to produce high-quality avatar [11] or reconstruct high-fidelity garment mesh [44], which show the advantage of implicit representation on recovering geometry details. However, these methods are limited to self-rotating human motion or a single color image. Neuralbody [25] and AniNErf [24] extend NeRFs for a dynamic human with defor" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 276, + 746, + 300, + 756 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 276, + 746, + 300, + 756 + ], + "spans": [ + { + "bbox": [ + 276, + 746, + 300, + 756 + ], + "type": "text", + "content": "21099" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 595, + 842 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 60, + 103, + 517, + 342 + ], + "blocks": [ + { + "bbox": [ + 60, + 103, + 517, + 342 + ], + "lines": [ + { + "bbox": [ + 60, + 103, + 517, + 342 + ], + "spans": [ + { + "bbox": [ + 60, + 103, + 517, + 342 + ], + "type": "image", + "image_path": "be976881f92fa9d91395654b0c248d79871f51f1ec80781719aa574f9cd75ff7.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 45, + 347, + 531, + 391 + ], + "lines": [ + { + "bbox": [ + 45, + 347, + 531, + 391 + ], + "spans": [ + { + "bbox": [ + 45, + 347, + 531, + 391 + ], + "type": "text", + "content": "Figure 1. Overview of the proposed approach. Our approach can capture both the clothing and human body motion, and achieve novel-view synthesis from sparse-view or monocular videos. We learn a double-layer neural radiance fields (NeRFs) for the clothing and human body in canonical frame, and jointly optimize the canonical double-layer NeRFs and the human deformation constrained with a physics-aware simulation network. Please refer to Sec. 3 for details of our method." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 45, + 398, + 281, + 538 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 398, + 281, + 538 + ], + "spans": [ + { + "bbox": [ + 45, + 398, + 281, + 538 + ], + "type": "text", + "content": "mation fields based on the linear blend skinning (LBS) [15]. Some human NeRFs methods [17, 41] model nonrigid human deformations with a point displacement on top of LBS. However, without using any clothing motion prior, these methods cannot effectively recover the motion of humans in loose clothes, and the geometry of clothes is not reconstructed reliably under nonrigid cloth deformations. The recent work, NerfCap [34] incorporates the embedded graph prior into the human NeRFs and represents nonlinear surface deformations for loose clothes successfully. However, their method does not model the clothing separately and thus fails to track the clothing motion." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 45, + 545, + 160, + 559 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 545, + 160, + 559 + ], + "spans": [ + { + "bbox": [ + 45, + 545, + 160, + 559 + ], + "type": "text", + "content": "3. Proposed Approach" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 45, + 566, + 281, + 730 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 566, + 281, + 730 + ], + "spans": [ + { + "bbox": [ + 45, + 566, + 281, + 730 + ], + "type": "text", + "content": "Given sparse-view or monocular videos of a clothed human, we track the motion of both the human body and clothing and synthesize novel-view videos of the human with a novel performance capture method based on a double-layer neural radiance fields (NeRFs). Figure 1 shows an overview of the proposed method. The double-layer NeRFs is learnt for the human body and clothing in canonical frame (Sec. 3.2). Then, the canonical NeRFs is dynamically transformed into the observed frame with inverse deformation fields of the deformed template for clothing and body respectively. From the dynamic double-layer NeRFs, we render the images using a composite rendering. By matching the synthesized images with the video frames and constraining the cloth deformations with a physics-aware simulation" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 296, + 398, + 530, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 398, + 530, + 422 + ], + "spans": [ + { + "bbox": [ + 296, + 398, + 530, + 422 + ], + "type": "text", + "content": "network, we estimate the double-layer NeRFs and the human deformation jointly (Sec. 3.3)." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 296, + 429, + 434, + 443 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 429, + 434, + 443 + ], + "spans": [ + { + "bbox": [ + 296, + 429, + 434, + 443 + ], + "type": "text", + "content": "3.1. Canonical Human Shape" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 296, + 447, + 530, + 576 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 447, + 530, + 576 + ], + "spans": [ + { + "bbox": [ + 296, + 447, + 530, + 576 + ], + "type": "text", + "content": "We decompose the clothed human into undressed body and garments (like T-shirt and pants), and represent them with SMPL model [18] and a parametric PCA model respectively. To include the human details (e.g., the head and hand), we obtain the canonical body model " + }, + { + "bbox": [ + 296, + 447, + 530, + 576 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 296, + 447, + 530, + 576 + ], + "type": "text", + "content": " by deforming the SMPL model to a personalized template with non-rigid deformation [32]. We design different styles (size and sleeve length, et al.) for each garment category, and simulate these garments on the template SMPL model using Marvelous Designer [1]. Then, we obtain a linear model through PCA on the simulated meshes," + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 371, + 586, + 530, + 600 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 371, + 586, + 530, + 600 + ], + "spans": [ + { + "bbox": [ + 371, + 586, + 530, + 600 + ], + "type": "interline_equation", + "content": "\\boldsymbol {G} (\\alpha) = \\bar {\\boldsymbol {G}} + \\boldsymbol {B} _ {g} \\alpha , \\tag {1}", + "image_path": "d307b5176cfed35603383785e5e1e52d01f1696218556e6859a8db3ce469fd7f.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 296, + 608, + 530, + 633 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 608, + 530, + 633 + ], + "spans": [ + { + "bbox": [ + 296, + 608, + 530, + 633 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 296, + 608, + 530, + 633 + ], + "type": "inline_equation", + "content": "\\bar{G}" + }, + { + "bbox": [ + 296, + 608, + 530, + 633 + ], + "type": "text", + "content": " is the mean garment shape, " + }, + { + "bbox": [ + 296, + 608, + 530, + 633 + ], + "type": "inline_equation", + "content": "B_{g}" + }, + { + "bbox": [ + 296, + 608, + 530, + 633 + ], + "type": "text", + "content": " denotes the linear basis matrix, and " + }, + { + "bbox": [ + 296, + 608, + 530, + 633 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 296, + 608, + 530, + 633 + ], + "type": "text", + "content": " are parameters of the cloth shape " + }, + { + "bbox": [ + 296, + 608, + 530, + 633 + ], + "type": "inline_equation", + "content": "G(\\alpha)" + }, + { + "bbox": [ + 296, + 608, + 530, + 633 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 296, + 640, + 513, + 654 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 640, + 513, + 654 + ], + "spans": [ + { + "bbox": [ + 296, + 640, + 513, + 654 + ], + "type": "text", + "content": "3.2. Double-layer NeRFs for Dynamic Humans" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 296, + 659, + 530, + 730 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 659, + 530, + 730 + ], + "spans": [ + { + "bbox": [ + 296, + 659, + 530, + 730 + ], + "type": "text", + "content": "Unlike a single NeRFs for clothed humans [17,24,25,34, 41], we represent the clothing with an independent NeRFs on the body, which forms a double-layer NeRFs. The implicit representation can be extended to multiple garments. Canonical NeRF. We use an occupancy network [22] to represent canonical geometry since the occupancy can mod" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 276, + 746, + 301, + 756 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 276, + 746, + 301, + 756 + ], + "spans": [ + { + "bbox": [ + 276, + 746, + 301, + 756 + ], + "type": "text", + "content": "21100" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 595, + 842 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 45, + 106, + 235, + 117 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 106, + 235, + 117 + ], + "spans": [ + { + "bbox": [ + 45, + 106, + 235, + 117 + ], + "type": "text", + "content": "el geometry more precisely than density [20] as," + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 121, + 126, + 279, + 139 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 121, + 126, + 279, + 139 + ], + "spans": [ + { + "bbox": [ + 121, + 126, + 279, + 139 + ], + "type": "interline_equation", + "content": "\\boldsymbol {o} (\\boldsymbol {x}) = \\boldsymbol {F} _ {o} \\left(\\gamma_ {\\boldsymbol {x}} (\\boldsymbol {x})\\right), \\tag {2}", + "image_path": "46669aa09837321259dbb222b98a7899eb31d1a8d5aa8cec257aa3b5887e505b.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 147, + 279, + 170 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 147, + 279, + 170 + ], + "spans": [ + { + "bbox": [ + 46, + 147, + 279, + 170 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 147, + 279, + 170 + ], + "type": "inline_equation", + "content": "\\gamma_{\\pmb{x}}" + }, + { + "bbox": [ + 46, + 147, + 279, + 170 + ], + "type": "text", + "content": " is the positional encoding [20]. The color network " + }, + { + "bbox": [ + 46, + 147, + 279, + 170 + ], + "type": "inline_equation", + "content": "F_{c}" + }, + { + "bbox": [ + 46, + 147, + 279, + 170 + ], + "type": "text", + "content": " in canonical frame is formulated as," + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 113, + 179, + 279, + 192 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 113, + 179, + 279, + 192 + ], + "spans": [ + { + "bbox": [ + 113, + 179, + 279, + 192 + ], + "type": "interline_equation", + "content": "\\boldsymbol {c} _ {i} (\\boldsymbol {x}) = \\boldsymbol {F} _ {c} \\left(\\gamma_ {\\boldsymbol {x}} (\\boldsymbol {x}), \\varphi_ {i}\\right), \\tag {3}", + "image_path": "fa39e685b1b28f7a0d1b38d7a0d73c1bda97c1159c7164cfe84bcebd1d7da0d7.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 200, + 279, + 234 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 200, + 279, + 234 + ], + "spans": [ + { + "bbox": [ + 46, + 200, + 279, + 234 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 200, + 279, + 234 + ], + "type": "inline_equation", + "content": "\\varphi_{i}" + }, + { + "bbox": [ + 46, + 200, + 279, + 234 + ], + "type": "text", + "content": " is an appearance latent code [24, 25] for frame " + }, + { + "bbox": [ + 46, + 200, + 279, + 234 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 200, + 279, + 234 + ], + "type": "text", + "content": ". We define occupancy network " + }, + { + "bbox": [ + 46, + 200, + 279, + 234 + ], + "type": "inline_equation", + "content": "\\{F_o^b,F_o^g\\}" + }, + { + "bbox": [ + 46, + 200, + 279, + 234 + ], + "type": "text", + "content": ", and color network " + }, + { + "bbox": [ + 46, + 200, + 279, + 234 + ], + "type": "inline_equation", + "content": "\\{F_c^b,\\varphi^b,F_c^g,\\varphi^g\\}" + }, + { + "bbox": [ + 46, + 200, + 279, + 234 + ], + "type": "text", + "content": " for the body and clothing, respectively." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 45, + 235, + 279, + 258 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 235, + 279, + 258 + ], + "spans": [ + { + "bbox": [ + 45, + 235, + 279, + 258 + ], + "type": "text", + "content": "To facilitate the learning of occupancy network, we incorporate a signed distance fields (SDFs) network " + }, + { + "bbox": [ + 45, + 235, + 279, + 258 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 45, + 235, + 279, + 258 + ], + "type": "text", + "content": "," + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 117, + 266, + 279, + 279 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 117, + 266, + 279, + 279 + ], + "spans": [ + { + "bbox": [ + 117, + 266, + 279, + 279 + ], + "type": "interline_equation", + "content": "\\boldsymbol {S}: \\boldsymbol {p} \\in \\mathbb {R} ^ {3} \\rightarrow \\boldsymbol {s} \\in \\mathbb {R}. \\tag {4}", + "image_path": "85ba9263eaf66a58c4550b59f3555188966d230c7adbb5d7157c8de94f5c73de.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 45, + 287, + 279, + 311 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 287, + 279, + 311 + ], + "spans": [ + { + "bbox": [ + 45, + 287, + 279, + 311 + ], + "type": "text", + "content": "which predicts the SDF value " + }, + { + "bbox": [ + 45, + 287, + 279, + 311 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 45, + 287, + 279, + 311 + ], + "type": "text", + "content": " for point " + }, + { + "bbox": [ + 45, + 287, + 279, + 311 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 45, + 287, + 279, + 311 + ], + "type": "text", + "content": ". We use the losses proposed in [30] to learn the SDFs of the shape," + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 318, + 289, + 372 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 318, + 289, + 372 + ], + "spans": [ + { + "bbox": [ + 46, + 318, + 289, + 372 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} L _ {s d f} = \\sum_ {\\boldsymbol {p} \\in \\Phi} | \\boldsymbol {S} (\\boldsymbol {p}) | + \\sum_ {\\boldsymbol {p} \\in \\Phi} (1 - \\langle \\nabla \\boldsymbol {S} (\\boldsymbol {p}), \\bar {\\boldsymbol {n}}) \\rangle \\tag {5} \\\\ + \\sum_ {\\boldsymbol {p} \\in \\Omega} | \\| \\nabla \\boldsymbol {S} (\\boldsymbol {p}) \\| _ {2} - 1 | + \\sum_ {\\boldsymbol {p} \\in \\Omega \\backslash \\Phi} \\exp (- \\delta \\cdot | \\boldsymbol {S} (\\boldsymbol {p}) |), \\\\ \\end{array}", + "image_path": "9fbe37bf3f122171069e2f1d7e6e8d31135f054a3b79b5c43b932da540296d79.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 45, + 380, + 279, + 474 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 380, + 279, + 474 + ], + "spans": [ + { + "bbox": [ + 45, + 380, + 279, + 474 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 45, + 380, + 279, + 474 + ], + "type": "inline_equation", + "content": "\\bar{n}" + }, + { + "bbox": [ + 45, + 380, + 279, + 474 + ], + "type": "text", + "content": " is the surface normal, " + }, + { + "bbox": [ + 45, + 380, + 279, + 474 + ], + "type": "inline_equation", + "content": "\\nabla S(\\pmb{p})" + }, + { + "bbox": [ + 45, + 380, + 279, + 474 + ], + "type": "text", + "content": " indicates the gradient of the SDFs, " + }, + { + "bbox": [ + 45, + 380, + 279, + 474 + ], + "type": "inline_equation", + "content": "\\Omega" + }, + { + "bbox": [ + 45, + 380, + 279, + 474 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 45, + 380, + 279, + 474 + ], + "type": "inline_equation", + "content": "\\Phi" + }, + { + "bbox": [ + 45, + 380, + 279, + 474 + ], + "type": "text", + "content": " denotes the 3D space and shape surface respectively. The first and second terms are used to estimate valid SDFs and normals on the surfaces, respectively. The third term constrains the norm of SDF gradients to be 1, and the last term penalizes SDF values of non-surface points near 0 with " + }, + { + "bbox": [ + 45, + 380, + 279, + 474 + ], + "type": "inline_equation", + "content": "\\delta \\gg 1" + }, + { + "bbox": [ + 45, + 380, + 279, + 474 + ], + "type": "text", + "content": ". We define SDF network " + }, + { + "bbox": [ + 45, + 380, + 279, + 474 + ], + "type": "inline_equation", + "content": "\\{\\pmb{S}^b, \\pmb{S}^g\\}" + }, + { + "bbox": [ + 45, + 380, + 279, + 474 + ], + "type": "text", + "content": " for both the body and clothing." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 45, + 474, + 279, + 497 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 474, + 279, + 497 + ], + "spans": [ + { + "bbox": [ + 45, + 474, + 279, + 497 + ], + "type": "text", + "content": "Supervised by the predictions of SDF network, the occupancy network is learnt with the following loss," + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 98, + 505, + 279, + 529 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 98, + 505, + 279, + 529 + ], + "spans": [ + { + "bbox": [ + 98, + 505, + 279, + 529 + ], + "type": "interline_equation", + "content": "L _ {o c c} = \\sum_ {\\boldsymbol {x} \\in \\Omega} L _ {C} \\left(\\boldsymbol {F} _ {\\boldsymbol {o}} \\left(\\gamma_ {\\boldsymbol {x}} (\\boldsymbol {x})\\right), \\boldsymbol {o}\\right), \\tag {6}", + "image_path": "e50435f9963b899ef6638f2504df82e93898829319f48b3a82fbf99a5ee36aa3.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 45, + 537, + 279, + 618 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 537, + 279, + 618 + ], + "spans": [ + { + "bbox": [ + 45, + 537, + 279, + 618 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 45, + 537, + 279, + 618 + ], + "type": "inline_equation", + "content": "\\pmb{o}" + }, + { + "bbox": [ + 45, + 537, + 279, + 618 + ], + "type": "text", + "content": " is the occupancy indicated by SDFs (if " + }, + { + "bbox": [ + 45, + 537, + 279, + 618 + ], + "type": "inline_equation", + "content": "S(\\pmb{x}) < = 0" + }, + { + "bbox": [ + 45, + 537, + 279, + 618 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 45, + 537, + 279, + 618 + ], + "type": "inline_equation", + "content": "o = 1" + }, + { + "bbox": [ + 45, + 537, + 279, + 618 + ], + "type": "text", + "content": "; otherwise, " + }, + { + "bbox": [ + 45, + 537, + 279, + 618 + ], + "type": "inline_equation", + "content": "o = 0" + }, + { + "bbox": [ + 45, + 537, + 279, + 618 + ], + "type": "text", + "content": "), and " + }, + { + "bbox": [ + 45, + 537, + 279, + 618 + ], + "type": "inline_equation", + "content": "L_{C}" + }, + { + "bbox": [ + 45, + 537, + 279, + 618 + ], + "type": "text", + "content": " denotes the cross-entropy loss. We adopt the compositional design of SDF and occupancy because it is better for volumetric learning with continuous occupancy fields [22] compared to discrete labels extracted from SDFs. The template is varying during learning geometry model, and the loss is formulated as," + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 110, + 628, + 279, + 640 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 110, + 628, + 279, + 640 + ], + "spans": [ + { + "bbox": [ + 110, + 628, + 279, + 640 + ], + "type": "interline_equation", + "content": "L _ {g e o} = \\mu_ {1} L _ {s d f} + \\mu_ {2} L _ {o c c}. \\tag {7}", + "image_path": "db6ed71a04fba78528d4fe229d80311d84ddf6c61704021f8874be15e60c4a7d.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 45, + 648, + 279, + 704 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 648, + 279, + 704 + ], + "spans": [ + { + "bbox": [ + 45, + 648, + 279, + 704 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 45, + 648, + 279, + 704 + ], + "type": "inline_equation", + "content": "\\mu_{1}" + }, + { + "bbox": [ + 45, + 648, + 279, + 704 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 45, + 648, + 279, + 704 + ], + "type": "inline_equation", + "content": "\\mu_{2}" + }, + { + "bbox": [ + 45, + 648, + 279, + 704 + ], + "type": "text", + "content": " are the weight for each term. Through the optimization, the NeRF geometry is consistent with the estimated template so that densely deforming geometry of the template can be tracked by utilizing dense photometric constraints from NeRF color model." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 45, + 706, + 279, + 730 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 706, + 279, + 730 + ], + "spans": [ + { + "bbox": [ + 45, + 706, + 279, + 730 + ], + "type": "text", + "content": "Deformation Model. Following [34], we transform sampled points in observed space to canonical space with the" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 296, + 106, + 530, + 257 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 106, + 530, + 257 + ], + "spans": [ + { + "bbox": [ + 296, + 106, + 530, + 257 + ], + "type": "text", + "content": "inverse deformation of the nearest vertex on the deformed template. In this way, the human deformation model is disentangled into the non-rigid surface deformation (i.e., embedded graph deformation as in [34]) and the articulated skeletal motion. For the clothing and body, we define the deformation network " + }, + { + "bbox": [ + 296, + 106, + 530, + 257 + ], + "type": "inline_equation", + "content": "F_{d}^{g}" + }, + { + "bbox": [ + 296, + 106, + 530, + 257 + ], + "type": "text", + "content": " with latent codes " + }, + { + "bbox": [ + 296, + 106, + 530, + 257 + ], + "type": "inline_equation", + "content": "\\omega^{g}" + }, + { + "bbox": [ + 296, + 106, + 530, + 257 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 296, + 106, + 530, + 257 + ], + "type": "inline_equation", + "content": "F_{d}^{b}" + }, + { + "bbox": [ + 296, + 106, + 530, + 257 + ], + "type": "text", + "content": " with latent codes " + }, + { + "bbox": [ + 296, + 106, + 530, + 257 + ], + "type": "inline_equation", + "content": "\\omega^{b}" + }, + { + "bbox": [ + 296, + 106, + 530, + 257 + ], + "type": "text", + "content": ", respectively. Through nonrigid deformation, we obtain the clothing mesh " + }, + { + "bbox": [ + 296, + 106, + 530, + 257 + ], + "type": "inline_equation", + "content": "\\hat{G}" + }, + { + "bbox": [ + 296, + 106, + 530, + 257 + ], + "type": "text", + "content": " and body mesh " + }, + { + "bbox": [ + 296, + 106, + 530, + 257 + ], + "type": "inline_equation", + "content": "\\hat{M}" + }, + { + "bbox": [ + 296, + 106, + 530, + 257 + ], + "type": "text", + "content": ". Then, the linear blend skinning [15] is applied to obtain the final cloth mesh " + }, + { + "bbox": [ + 296, + 106, + 530, + 257 + ], + "type": "inline_equation", + "content": "\\tilde{G}" + }, + { + "bbox": [ + 296, + 106, + 530, + 257 + ], + "type": "text", + "content": " and body mesh " + }, + { + "bbox": [ + 296, + 106, + 530, + 257 + ], + "type": "inline_equation", + "content": "\\tilde{M}" + }, + { + "bbox": [ + 296, + 106, + 530, + 257 + ], + "type": "text", + "content": " in the observation space. We recover the SMPL model with [13] for input video frames and also estimate a canonical body shape." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 296, + 259, + 530, + 398 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 259, + 530, + 398 + ], + "spans": [ + { + "bbox": [ + 296, + 259, + 530, + 398 + ], + "type": "text", + "content": "Composite Rendering. In [6], a blending weight is predicted for each sampled point to blend the static background and dynamic target in a scene. This blending method cannot separate the clothing and human body in the NeRFs effectively. We propose an efficient composite rendering strategy for the dynamic double-layer NeRF. Specifically, for a casted ray " + }, + { + "bbox": [ + 296, + 259, + 530, + 398 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 296, + 259, + 530, + 398 + ], + "type": "text", + "content": ", we find the intersecting triangle with the deformed meshes. If the intersected triangle is from the body mesh, we set the mask " + }, + { + "bbox": [ + 296, + 259, + 530, + 398 + ], + "type": "inline_equation", + "content": "m_r^b" + }, + { + "bbox": [ + 296, + 259, + 530, + 398 + ], + "type": "text", + "content": " of all sampled points on " + }, + { + "bbox": [ + 296, + 259, + 530, + 398 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 296, + 259, + 530, + 398 + ], + "type": "text", + "content": " to 1; otherwise, " + }, + { + "bbox": [ + 296, + 259, + 530, + 398 + ], + "type": "inline_equation", + "content": "m_r^b" + }, + { + "bbox": [ + 296, + 259, + 530, + 398 + ], + "type": "text", + "content": " is set to 0. Similarly, if the intersected triangle is from the cloth mesh, we set the mask " + }, + { + "bbox": [ + 296, + 259, + 530, + 398 + ], + "type": "inline_equation", + "content": "m_r^g" + }, + { + "bbox": [ + 296, + 259, + 530, + 398 + ], + "type": "text", + "content": " to 1; otherwise, " + }, + { + "bbox": [ + 296, + 259, + 530, + 398 + ], + "type": "inline_equation", + "content": "m_r^g" + }, + { + "bbox": [ + 296, + 259, + 530, + 398 + ], + "type": "text", + "content": " is set to 0. The complete human is rendered as," + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 323, + 409, + 530, + 442 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 323, + 409, + 530, + 442 + ], + "spans": [ + { + "bbox": [ + 323, + 409, + 530, + 442 + ], + "type": "interline_equation", + "content": "\\boldsymbol {T} _ {r} ^ {i} = \\prod_ {j = 1} ^ {i - 1} \\left(1 - \\boldsymbol {\\sigma} _ {g} ^ {j} \\left(1 - \\boldsymbol {m} _ {r} ^ {b}\\right)\\right) \\left(1 - \\boldsymbol {\\sigma} _ {b} ^ {j} \\boldsymbol {m} _ {r} ^ {b}\\right), \\tag {8}", + "image_path": "4565d218a4ad8c0870a71551f855cd9991477e7250ead54c2e59b20159b803b8.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 323, + 444, + 504, + 474 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 323, + 444, + 504, + 474 + ], + "spans": [ + { + "bbox": [ + 323, + 444, + 504, + 474 + ], + "type": "interline_equation", + "content": "\\hat {\\boldsymbol {C}} _ {r} = \\sum_ {i = 1} ^ {n} \\boldsymbol {T} _ {r} ^ {i} \\left(\\boldsymbol {o} _ {g} ^ {i} \\left(1 - \\boldsymbol {m} _ {r} ^ {b}\\right) \\boldsymbol {c} _ {g} ^ {i} + \\boldsymbol {o} _ {b} ^ {i} \\boldsymbol {m} _ {r} ^ {b} \\boldsymbol {c} _ {b} ^ {i}\\right),", + "image_path": "03fd1d954d7fa8dfc8cf2fbcf370be049f1926579fd6a162b5b723357398b999.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 297, + 484, + 530, + 532 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 484, + 530, + 532 + ], + "spans": [ + { + "bbox": [ + 297, + 484, + 530, + 532 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 297, + 484, + 530, + 532 + ], + "type": "inline_equation", + "content": "\\{o_g,o_b\\}" + }, + { + "bbox": [ + 297, + 484, + 530, + 532 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 297, + 484, + 530, + 532 + ], + "type": "inline_equation", + "content": "\\{c_{g},c_{b}\\}" + }, + { + "bbox": [ + 297, + 484, + 530, + 532 + ], + "type": "text", + "content": " denote the point occupancy and color respectively, which are defined for both clothing and body. " + }, + { + "bbox": [ + 297, + 484, + 530, + 532 + ], + "type": "inline_equation", + "content": "T_r^i" + }, + { + "bbox": [ + 297, + 484, + 530, + 532 + ], + "type": "text", + "content": " is the accumulated transparency along the ray. We also render the clothing color " + }, + { + "bbox": [ + 297, + 484, + 530, + 532 + ], + "type": "inline_equation", + "content": "\\hat{C}_r^g" + }, + { + "bbox": [ + 297, + 484, + 530, + 532 + ], + "type": "text", + "content": " and body color " + }, + { + "bbox": [ + 297, + 484, + 530, + 532 + ], + "type": "inline_equation", + "content": "\\hat{C}_r^b" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 297, + 541, + 529, + 564 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 541, + 529, + 564 + ], + "spans": [ + { + "bbox": [ + 297, + 541, + 529, + 564 + ], + "type": "text", + "content": "3.3. Clothed Human Performance Capture with Double-layer NeRFs" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 297, + 571, + 530, + 606 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 571, + 530, + 606 + ], + "spans": [ + { + "bbox": [ + 297, + 571, + 530, + 606 + ], + "type": "text", + "content": "We estimate the double-layer NeRFs and template deformation of the body and clothing by minimizing the following loss between the synthesized color and observed color:" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 319, + 617, + 530, + 696 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 617, + 530, + 696 + ], + "spans": [ + { + "bbox": [ + 319, + 617, + 530, + 696 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} L _ {r g b} = \\sum_ {\\boldsymbol {r} \\in R} \\| \\hat {\\boldsymbol {C}} (\\boldsymbol {r}) - \\boldsymbol {C} (\\boldsymbol {r}) \\| _ {2} ^ {2} \\tag {9} \\\\ + \\sum_ {\\boldsymbol {r} \\in \\boldsymbol {R}} \\| (\\hat {\\boldsymbol {C}} (\\boldsymbol {r}) ^ {b} - \\boldsymbol {C} (\\boldsymbol {r})) (1 - \\boldsymbol {m} _ {r} ^ {g}) \\| _ {2} ^ {2} \\\\ + \\sum_ {\\boldsymbol {r} \\in \\boldsymbol {R}} \\| (\\hat {\\boldsymbol {C}} (\\boldsymbol {r}) ^ {g} - \\boldsymbol {C} (\\boldsymbol {r})) (1 - \\boldsymbol {m} _ {r} ^ {b}) \\| _ {2} ^ {2}, \\\\ \\end{array}", + "image_path": "65f0afcf64bfcea648065f464eee312d6fd690712196a197e7cbc2f07c3f45b9.jpg" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 297, + 706, + 530, + 730 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 706, + 530, + 730 + ], + "spans": [ + { + "bbox": [ + 297, + 706, + 530, + 730 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 297, + 706, + 530, + 730 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 297, + 706, + 530, + 730 + ], + "type": "text", + "content": " is the set of casted rays. While using the dense photometric constraint defined in Eq. 9, the model training" + } + ] + } + ], + "index": 24 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 276, + 746, + 299, + 756 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 276, + 746, + 299, + 756 + ], + "spans": [ + { + "bbox": [ + 276, + 746, + 299, + 756 + ], + "type": "text", + "content": "21101" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 595, + 842 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 45, + 106, + 279, + 141 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 106, + 279, + 141 + ], + "spans": [ + { + "bbox": [ + 45, + 106, + 279, + 141 + ], + "type": "text", + "content": "remains ill-posed and body-cloth collisions are prone to occur. To solve these problems, we design several regularization losses to constrain the joint optimization." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 45, + 143, + 281, + 330 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 143, + 281, + 330 + ], + "spans": [ + { + "bbox": [ + 45, + 143, + 281, + 330 + ], + "type": "text", + "content": "Physics-aware constraint. Cloth simulation is adopted in previous works [16, 40] to obtain realistic performance capture. Different from their traditional simulation method, we introduce a differential network that learns a mapping from human and cloth parameters to physics-aware cloth shapes. We simulate the data in Marvelous Designer using the method of [23]. For each garment category, we use 25 garment styles " + }, + { + "bbox": [ + 45, + 143, + 281, + 330 + ], + "type": "inline_equation", + "content": "\\gamma \\in \\mathbb{R}^4" + }, + { + "bbox": [ + 45, + 143, + 281, + 330 + ], + "type": "text", + "content": " as [23]. To account for the effects of different materials, we simulate 8 different materials in Marvelous Designer, and represent the material type " + }, + { + "bbox": [ + 45, + 143, + 281, + 330 + ], + "type": "inline_equation", + "content": "\\pmb{\\tau}" + }, + { + "bbox": [ + 45, + 143, + 281, + 330 + ], + "type": "text", + "content": " using one-hot labels. We train the simulation network " + }, + { + "bbox": [ + 45, + 143, + 281, + 330 + ], + "type": "inline_equation", + "content": "D_{\\phi}" + }, + { + "bbox": [ + 45, + 143, + 281, + 330 + ], + "type": "text", + "content": " using a MLP to minimize the L1-loss between the simulated cloth meshes and the predicted ones " + }, + { + "bbox": [ + 45, + 143, + 281, + 330 + ], + "type": "inline_equation", + "content": "G_{s} = D_{\\phi}(\\theta ,\\beta ,\\gamma ,\\tau)" + }, + { + "bbox": [ + 45, + 143, + 281, + 330 + ], + "type": "text", + "content": " with the input of human pose and shape, and garment and material styles. The physical simulation loss enforces the deformed cloth mesh " + }, + { + "bbox": [ + 45, + 143, + 281, + 330 + ], + "type": "inline_equation", + "content": "\\tilde{G}" + }, + { + "bbox": [ + 45, + 143, + 281, + 330 + ], + "type": "text", + "content": " to be close to " + }, + { + "bbox": [ + 45, + 143, + 281, + 330 + ], + "type": "inline_equation", + "content": "G_{s}" + }, + { + "bbox": [ + 45, + 143, + 281, + 330 + ], + "type": "text", + "content": "," + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 103, + 339, + 279, + 364 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 103, + 339, + 279, + 364 + ], + "spans": [ + { + "bbox": [ + 103, + 339, + 279, + 364 + ], + "type": "interline_equation", + "content": "L _ {s i m} = \\sum_ {t \\in \\boldsymbol {G}} \\rho \\left(\\left\\| \\tilde {\\boldsymbol {G}} ^ {t} - \\boldsymbol {G} _ {s} ^ {t} \\right\\| _ {2} ^ {2}\\right), \\tag {10}", + "image_path": "da98cfa9d8dc6f096a5f138996ecfc26a5120d1be0a057360b1481cc480483ab.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 45, + 374, + 279, + 420 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 374, + 279, + 420 + ], + "spans": [ + { + "bbox": [ + 45, + 374, + 279, + 420 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 45, + 374, + 279, + 420 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 45, + 374, + 279, + 420 + ], + "type": "text", + "content": " denotes vertex index of " + }, + { + "bbox": [ + 45, + 374, + 279, + 420 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 45, + 374, + 279, + 420 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 45, + 374, + 279, + 420 + ], + "type": "inline_equation", + "content": "\\rho" + }, + { + "bbox": [ + 45, + 374, + 279, + 420 + ], + "type": "text", + "content": " is the Geman-McClure robust function [5]. The simulation loss can penalize body-cloth collisions and preserve physically plausible deformations." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 45, + 422, + 279, + 470 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 422, + 279, + 470 + ], + "spans": [ + { + "bbox": [ + 45, + 422, + 279, + 470 + ], + "type": "text", + "content": "As-rigid-as-possible Loss. To make sure the template deforms smoothly, we utilize an as-rigid-as-possible loss [32] on the embedded graph deformation of the human body and clothing, " + }, + { + "bbox": [ + 45, + 422, + 279, + 470 + ], + "type": "inline_equation", + "content": "L_{arap} = L_{arap}(\\hat{M}) + L_{arap}(\\hat{G})" + }, + { + "bbox": [ + 45, + 422, + 279, + 470 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 45, + 470, + 279, + 516 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 470, + 279, + 516 + ], + "spans": [ + { + "bbox": [ + 45, + 470, + 279, + 516 + ], + "type": "text", + "content": "Mask Loss. We render the mask image " + }, + { + "bbox": [ + 45, + 470, + 279, + 516 + ], + "type": "inline_equation", + "content": "R(\\mathbf{U})" + }, + { + "bbox": [ + 45, + 470, + 279, + 516 + ], + "type": "text", + "content": " of the combined mesh " + }, + { + "bbox": [ + 45, + 470, + 279, + 516 + ], + "type": "inline_equation", + "content": "\\mathbf{U} = (M, G)" + }, + { + "bbox": [ + 45, + 470, + 279, + 516 + ], + "type": "text", + "content": " by using a differentiable renderer [35] based on point cloud, and compute the mask loss with the input human mask image " + }, + { + "bbox": [ + 45, + 470, + 279, + 516 + ], + "type": "inline_equation", + "content": "\\bar{R}" + }, + { + "bbox": [ + 45, + 470, + 279, + 516 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 103, + 527, + 279, + 543 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 103, + 527, + 279, + 543 + ], + "spans": [ + { + "bbox": [ + 103, + 527, + 279, + 543 + ], + "type": "interline_equation", + "content": "L _ {\\text {m a s k}} (\\boldsymbol {U}) = \\left\\| R (\\boldsymbol {U}) - \\bar {R} \\right\\| _ {2} ^ {2}. \\tag {11}", + "image_path": "c5dbf61ac2160cd9ed513d7bf74f21c014ef6e7511cdfeffb42906b6d1420a03.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 45, + 553, + 279, + 588 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 553, + 279, + 588 + ], + "spans": [ + { + "bbox": [ + 45, + 553, + 279, + 588 + ], + "type": "text", + "content": "We utilize the mask loss on both the deformed mesh and the simulated mesh as, " + }, + { + "bbox": [ + 45, + 553, + 279, + 588 + ], + "type": "inline_equation", + "content": "L_{mask} = L_{mask}(\\tilde{M},\\tilde{G}) + L_{mask}(\\tilde{M},G_s)" + }, + { + "bbox": [ + 45, + 553, + 279, + 588 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 45, + 590, + 279, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 590, + 279, + 635 + ], + "spans": [ + { + "bbox": [ + 45, + 590, + 279, + 635 + ], + "type": "text", + "content": "Attachment Loss. During human motion, some cloth parts are always attached to the body (e.g., the dress strap is fixed on the shoulder). To ensure this constraint, we define an attachment loss:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 95, + 645, + 279, + 671 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 95, + 645, + 279, + 671 + ], + "spans": [ + { + "bbox": [ + 95, + 645, + 279, + 671 + ], + "type": "interline_equation", + "content": "L _ {\\text {a t t a c h}} = \\sum_ {i, j \\in A} \\rho \\left(\\| \\tilde {\\boldsymbol {G}} ^ {i} - \\tilde {\\boldsymbol {M}} ^ {j} \\| _ {2} ^ {2}\\right), \\tag {12}", + "image_path": "da78a89c804c76cf078b6aab7ec1bf956f54c9ff5f4db447c4474de101af6e0a.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 45, + 682, + 279, + 705 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 682, + 279, + 705 + ], + "spans": [ + { + "bbox": [ + 45, + 682, + 279, + 705 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 45, + 682, + 279, + 705 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 45, + 682, + 279, + 705 + ], + "type": "text", + "content": " is a set of selected vertices on the clothing mesh attached to corresponding points of the body mesh." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 45, + 706, + 279, + 730 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 706, + 279, + 730 + ], + "spans": [ + { + "bbox": [ + 45, + 706, + 279, + 730 + ], + "type": "text", + "content": "Interpenetration Loss. Interpenetrations may occur between the predicted garments and the body. We use the" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 297, + 106, + 494, + 117 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 106, + 494, + 117 + ], + "spans": [ + { + "bbox": [ + 297, + 106, + 494, + 117 + ], + "type": "text", + "content": "interpenetration term [7] to alleviate this problem:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 299, + 124, + 530, + 162 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 124, + 530, + 162 + ], + "spans": [ + { + "bbox": [ + 299, + 124, + 530, + 162 + ], + "type": "interline_equation", + "content": "L _ {i n t e r p} \\left(M _ {1}, M _ {2}\\right) = \\frac {1}{| C |} \\sum_ {i, j \\in C} R e L U \\left(\\left(M _ {1} ^ {i} - M _ {2} ^ {j}\\right) \\cdot N _ {1} ^ {i}\\right), \\tag {13}", + "image_path": "42c08245d0afb2e26675afc65605c6421f506c7e8b1f75de930d8704b95f8aa1.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 296, + 163, + 530, + 210 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 163, + 530, + 210 + ], + "spans": [ + { + "bbox": [ + 296, + 163, + 530, + 210 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 296, + 163, + 530, + 210 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 296, + 163, + 530, + 210 + ], + "type": "text", + "content": " is the set of corresponding points between the mesh " + }, + { + "bbox": [ + 296, + 163, + 530, + 210 + ], + "type": "inline_equation", + "content": "M_{1}" + }, + { + "bbox": [ + 296, + 163, + 530, + 210 + ], + "type": "text", + "content": " and mesh " + }, + { + "bbox": [ + 296, + 163, + 530, + 210 + ], + "type": "inline_equation", + "content": "M_2" + }, + { + "bbox": [ + 296, + 163, + 530, + 210 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 296, + 163, + 530, + 210 + ], + "type": "inline_equation", + "content": "|C|" + }, + { + "bbox": [ + 296, + 163, + 530, + 210 + ], + "type": "text", + "content": " denotes the number of corresponding points, and " + }, + { + "bbox": [ + 296, + 163, + 530, + 210 + ], + "type": "inline_equation", + "content": "N_{1}" + }, + { + "bbox": [ + 296, + 163, + 530, + 210 + ], + "type": "text", + "content": " is the normal of mesh " + }, + { + "bbox": [ + 296, + 163, + 530, + 210 + ], + "type": "inline_equation", + "content": "M_{1}" + }, + { + "bbox": [ + 296, + 163, + 530, + 210 + ], + "type": "text", + "content": ". We use interpenetration loss in both canonical and posed space," + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 314, + 216, + 530, + 230 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 216, + 530, + 230 + ], + "spans": [ + { + "bbox": [ + 314, + 216, + 530, + 230 + ], + "type": "interline_equation", + "content": "L _ {i n t e r p} = L _ {i n t e r p} (\\hat {\\boldsymbol {M}}, \\hat {\\boldsymbol {G}}) + L _ {i n t e r p} (\\tilde {\\boldsymbol {M}}, \\tilde {\\boldsymbol {G}}). \\tag {14}", + "image_path": "2010dc5b8f8a4b400e4e0ab92ef8638c237be17f61f13b2f87a8c492d1c17f73.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 297, + 243, + 411, + 253 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 243, + 411, + 253 + ], + "spans": [ + { + "bbox": [ + 297, + 243, + 411, + 253 + ], + "type": "text", + "content": "3.3.1 Total Loss Function" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 296, + 261, + 530, + 342 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 261, + 530, + 342 + ], + "spans": [ + { + "bbox": [ + 296, + 261, + 530, + 342 + ], + "type": "text", + "content": "We first learn the geometry network of both the clothing and human body in canonical frame. We use the first frame of multiple views or a video in self-rotating motion to estimate the canonical geometry by optimizing all the losses. Then, by fixing the clothing shape and the geometry network, we track the motion of each video frame and estimate the appearance of NeRFs using the following total loss:" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 325, + 350, + 530, + 378 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 350, + 530, + 378 + ], + "spans": [ + { + "bbox": [ + 325, + 350, + 530, + 378 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} L = \\lambda_ {1} L _ {r g b} + \\lambda_ {2} L _ {a r a p} + \\lambda_ {3} L _ {s i m} \\tag {15} \\\\ + \\lambda_ {4} L _ {\\text {m a s k}} + \\lambda_ {5} L _ {\\text {a t t a c h}} + \\lambda_ {6} L _ {\\text {i n t e r p}}, \\\\ \\end{array}", + "image_path": "fbd2847583975e504a4f15d27b2ebf36e46f5a97e71d2a73e89d87432c4759fc.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 297, + 383, + 529, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 383, + 529, + 407 + ], + "spans": [ + { + "bbox": [ + 297, + 383, + 529, + 407 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 297, + 383, + 529, + 407 + ], + "type": "inline_equation", + "content": "\\{\\lambda_1\\ldots \\lambda_6\\}" + }, + { + "bbox": [ + 297, + 383, + 529, + 407 + ], + "type": "text", + "content": " are the balancing weights. For a video with " + }, + { + "bbox": [ + 297, + 383, + 529, + 407 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 297, + 383, + 529, + 407 + ], + "type": "text", + "content": " frames, the estimated parameter set " + }, + { + "bbox": [ + 297, + 383, + 529, + 407 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 297, + 383, + 529, + 407 + ], + "type": "text", + "content": " is:" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 298, + 414, + 529, + 438 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 414, + 529, + 438 + ], + "spans": [ + { + "bbox": [ + 298, + 414, + 529, + 438 + ], + "type": "interline_equation", + "content": "X = \\left\\{\\boldsymbol {F} _ {c} ^ {b}, \\boldsymbol {F} _ {c} ^ {g}, \\boldsymbol {F} _ {d} ^ {b}, \\boldsymbol {F} _ {d} ^ {g} \\right\\} \\cup \\left\\{\\varphi_ {i} ^ {b}, \\varphi_ {i} ^ {g}, \\omega_ {i} ^ {b}, \\omega_ {i} ^ {g} \\mid i \\in 1, \\dots , N \\right\\}, \\tag {16}", + "image_path": "63ea930e646636eb5e398e43149c8f1aa372b6cca850584b7b7b3e305792d2ed.jpg" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 296, + 438, + 529, + 473 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 438, + 529, + 473 + ], + "spans": [ + { + "bbox": [ + 296, + 438, + 529, + 473 + ], + "type": "text", + "content": "which includes the weights of the appearance and deformation networks and the corresponding conditional variables for each frame in the video." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 297, + 484, + 376, + 496 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 484, + 376, + 496 + ], + "spans": [ + { + "bbox": [ + 297, + 484, + 376, + 496 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 297, + 502, + 424, + 515 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 502, + 424, + 515 + ], + "spans": [ + { + "bbox": [ + 297, + 502, + 424, + 515 + ], + "type": "text", + "content": "4.1. Implementation details" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 296, + 520, + 530, + 730 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 520, + 530, + 730 + ], + "spans": [ + { + "bbox": [ + 296, + 520, + 530, + 730 + ], + "type": "text", + "content": "Datasets and metrics. We design common garments (e.g., dress, T-shirt, skirt, pants, et al.) and simulate the data in the Marvelous Designer [1] under various human shapes and poses and cloth materials. We test our method on publicly datasets, DynaCap [8], and DeepCap [10]. We use four cameras around the human for training and the remaining cameras for test. Figure 2 shows some recovered results using our approach on different clothed humans. Our method is also tested on a monocular video in outdoor scene. Please refer to the supplementary video for more experimental results. The geometry accuracy is evaluated with the intersection over union (IoU) between the ground truth human masks and 2D projections of the recovered human shapes. Following [10, 34], the IoU (\\%) is computed on all views (AMVIoU), all views except the input view (RVIoU), and the input view (SVIoU). The novel-view synthesis quality is evaluated with two metrics, i.e., peak signal-to-noise ratio (PSNR) and structural similarity index (SSIM) as [24, 34]." + } + ] + } + ], + "index": 24 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 276, + 746, + 300, + 756 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 276, + 746, + 300, + 756 + ], + "spans": [ + { + "bbox": [ + 276, + 746, + 300, + 756 + ], + "type": "text", + "content": "21102" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 595, + 842 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 123, + 106, + 452, + 417 + ], + "blocks": [ + { + "bbox": [ + 123, + 106, + 452, + 417 + ], + "lines": [ + { + "bbox": [ + 123, + 106, + 452, + 417 + ], + "spans": [ + { + "bbox": [ + 123, + 106, + 452, + 417 + ], + "type": "image", + "image_path": "8a0fc36af3de34eb7391906c13c0ef74a624df0294341d8ad22b62a2d3895e1c.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 45, + 421, + 533, + 454 + ], + "lines": [ + { + "bbox": [ + 45, + 421, + 533, + 454 + ], + "spans": [ + { + "bbox": [ + 45, + 421, + 533, + 454 + ], + "type": "text", + "content": "Figure 2. The results of our approach on different humans in various garments. In each example, we show the ground truth image in a novel view, the reconstructed geometry (overlayed on the image), and the novel-view synthesis image. From left to right: the S4 from DeepCap dataset [10], \"FranziRed\" from DynaCap dataset [8], and the S1 from DeepCap dataset [10]." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 45, + 460, + 281, + 611 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 460, + 281, + 611 + ], + "spans": [ + { + "bbox": [ + 45, + 460, + 281, + 611 + ], + "type": "text", + "content": "Architecture and experimental settings. We adopt the S-DF network from [30], and the same occupancy and color network as [34]. In the non-rigid deformation network, there are 8 layers with a skip connection at the fourth layer and the hidden size is 1024. The deformation and appearance latent codes both have dimensions of 128. We set both weight " + }, + { + "bbox": [ + 45, + 460, + 281, + 611 + ], + "type": "inline_equation", + "content": "\\mu_{1}" + }, + { + "bbox": [ + 45, + 460, + 281, + 611 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 45, + 460, + 281, + 611 + ], + "type": "inline_equation", + "content": "\\mu_{2}" + }, + { + "bbox": [ + 45, + 460, + 281, + 611 + ], + "type": "text", + "content": " to 1. We set the loss weight " + }, + { + "bbox": [ + 45, + 460, + 281, + 611 + ], + "type": "inline_equation", + "content": "\\{\\lambda_1,\\dots,\\lambda_6\\}" + }, + { + "bbox": [ + 45, + 460, + 281, + 611 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 45, + 460, + 281, + 611 + ], + "type": "inline_equation", + "content": "\\{1.0,0.1,0.02,50,0.1,100.0\\}" + }, + { + "bbox": [ + 45, + 460, + 281, + 611 + ], + "type": "text", + "content": " empirically. The Adam optimizer [14] is used in the optimization, and the learning rate is " + }, + { + "bbox": [ + 45, + 460, + 281, + 611 + ], + "type": "inline_equation", + "content": "5e^{-4}" + }, + { + "bbox": [ + 45, + 460, + 281, + 611 + ], + "type": "text", + "content": " at the start of training and exponentially decays to " + }, + { + "bbox": [ + 45, + 460, + 281, + 611 + ], + "type": "inline_equation", + "content": "5e^{-5}" + }, + { + "bbox": [ + 45, + 460, + 281, + 611 + ], + "type": "text", + "content": " during the training. With a GPU of RTX 2080 Ti, our network training costs about 12 hours tested on a four-view video with 300 frames." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 45, + 617, + 249, + 630 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 617, + 249, + 630 + ], + "spans": [ + { + "bbox": [ + 45, + 617, + 249, + 630 + ], + "type": "text", + "content": "4.2. Comparison to state-of-the-art methods" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 45, + 636, + 279, + 730 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 636, + 279, + 730 + ], + "spans": [ + { + "bbox": [ + 45, + 636, + 279, + 730 + ], + "type": "text", + "content": "We compare our method with two state-of-the-art approaches for human performance capture, DeepCap [10] and NerfCap [34]. We also compare with BCNet [12], TailorNet [23], and ICON [36]. For DeepCap and NerfCap, the experimental results are from [34]. For other compared methods, we implement them using their released codes. We do not compare with DeepCloth [16] since their codes are not released and DeepCap [10] outperforms them in ge" + } + ] + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 316, + 458, + 514, + 543 + ], + "blocks": [ + { + "bbox": [ + 316, + 458, + 514, + 543 + ], + "lines": [ + { + "bbox": [ + 316, + 458, + 514, + 543 + ], + "spans": [ + { + "bbox": [ + 316, + 458, + 514, + 543 + ], + "type": "table", + "html": "
MethodAMVIoURVIoUSVIoU
DeepCap [10]86.1785.7586.90
NerfCap [34]88.9688.3290.09
ICON [36]--93.56
BCNet [12]82.3381.6782.94
TailorNet [23]83.8783.5084.51
Our method89.5888.7591.03
", + "image_path": "1ff9c9dbaf1d11b40b0a8bfc59010e8887f868c9ab35e9a11252fef88c0e7adc.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 315, + 559, + 513, + 644 + ], + "blocks": [ + { + "bbox": [ + 302, + 543, + 522, + 554 + ], + "lines": [ + { + "bbox": [ + 302, + 543, + 522, + 554 + ], + "spans": [ + { + "bbox": [ + 302, + 543, + 522, + 554 + ], + "type": "text", + "content": "Evaluated on the subject S4 from DeepCap dataset [10]." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 315, + 559, + 513, + 644 + ], + "lines": [ + { + "bbox": [ + 315, + 559, + 513, + 644 + ], + "spans": [ + { + "bbox": [ + 315, + 559, + 513, + 644 + ], + "type": "table", + "html": "
MethodAMVIoURVIoUSVIoU
DeepCap [10]85.3885.3785.41
NerfCap [34]88.0888.0788.12
ICON [36]--95.21
BCNet [12]78.9178.7780.83
TailorNet [23]78.1578.0078.21
Our method88.7488.4089.68
", + "image_path": "2db4dd8c42fc8b8e0a0b51f273787f61d3c9e050f3892f677c041b3f00f83b88.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 301, + 644, + 525, + 655 + ], + "lines": [ + { + "bbox": [ + 301, + 644, + 525, + 655 + ], + "spans": [ + { + "bbox": [ + 301, + 644, + 525, + 655 + ], + "type": "text", + "content": "Evaluated on the 'FranziRed' from DynaCap dataset [8]." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 296, + 658, + 530, + 702 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 658, + 530, + 702 + ], + "spans": [ + { + "bbox": [ + 296, + 658, + 530, + 702 + ], + "type": "text", + "content": "Table 1. Comparison to the state-of-the-arts in surface reconstruction accuracy. Note that, the accuracy of ICON [36] is high since they use the human mask in estimating the geometry, and their dress shapes are not recovered accurately as shown in Fig. 3." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 296, + 706, + 531, + 730 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 706, + 531, + 730 + ], + "spans": [ + { + "bbox": [ + 296, + 706, + 531, + 730 + ], + "type": "text", + "content": "ometry tracking as reported in [16]. We conduct the qualitative and quantitative experiments on S4 of DeepCap and" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 276, + 746, + 300, + 756 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 276, + 746, + 300, + 756 + ], + "spans": [ + { + "bbox": [ + 276, + 746, + 300, + 756 + ], + "type": "text", + "content": "21103" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 595, + 842 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 45, + 106, + 279, + 141 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 106, + 279, + 141 + ], + "spans": [ + { + "bbox": [ + 45, + 106, + 279, + 141 + ], + "type": "text", + "content": "\"FranziRed\" from DynaCap. The test dataset are the same as [34] which include 300 - 400 frames randomly sampled from the original dataset." + } + ] + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 52, + 145, + 273, + 384 + ], + "blocks": [ + { + "bbox": [ + 52, + 145, + 273, + 384 + ], + "lines": [ + { + "bbox": [ + 52, + 145, + 273, + 384 + ], + "spans": [ + { + "bbox": [ + 52, + 145, + 273, + 384 + ], + "type": "image", + "image_path": "88b1a4cb9af775bb7d201b363d90d292d84745d0ef7b38652693afd627517315.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 64, + 384, + 105, + 394 + ], + "lines": [ + { + "bbox": [ + 64, + 384, + 105, + 394 + ], + "spans": [ + { + "bbox": [ + 64, + 384, + 105, + 394 + ], + "type": "text", + "content": "(a) (b)" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 109, + 384, + 120, + 394 + ], + "lines": [ + { + "bbox": [ + 109, + 384, + 120, + 394 + ], + "spans": [ + { + "bbox": [ + 109, + 384, + 120, + 394 + ], + "type": "text", + "content": "#" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 121, + 384, + 152, + 394 + ], + "lines": [ + { + "bbox": [ + 121, + 384, + 152, + 394 + ], + "spans": [ + { + "bbox": [ + 121, + 384, + 152, + 394 + ], + "type": "text", + "content": "(c)" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 148, + 384, + 177, + 394 + ], + "lines": [ + { + "bbox": [ + 148, + 384, + 177, + 394 + ], + "spans": [ + { + "bbox": [ + 148, + 384, + 177, + 394 + ], + "type": "text", + "content": "(d)" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 179, + 384, + 208, + 394 + ], + "lines": [ + { + "bbox": [ + 179, + 384, + 208, + 394 + ], + "spans": [ + { + "bbox": [ + 179, + 384, + 208, + 394 + ], + "type": "text", + "content": "(e)" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 209, + 384, + 230, + 394 + ], + "lines": [ + { + "bbox": [ + 209, + 384, + 230, + 394 + ], + "spans": [ + { + "bbox": [ + 209, + 384, + 230, + 394 + ], + "type": "text", + "content": "(f)" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 232, + 384, + 262, + 394 + ], + "lines": [ + { + "bbox": [ + 232, + 384, + 262, + 394 + ], + "spans": [ + { + "bbox": [ + 232, + 384, + 262, + 394 + ], + "type": "text", + "content": "(g)" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 45, + 394, + 279, + 425 + ], + "lines": [ + { + "bbox": [ + 45, + 394, + 279, + 425 + ], + "spans": [ + { + "bbox": [ + 45, + 394, + 279, + 425 + ], + "type": "text", + "content": "Figure 3. Qualitative comparison with the state-of-the-arts. (a) Ground truth image. (b) DeepCap [10]. (c) NerfCap [34]. (d) ICON [36]. (e) BCNet [12]. (f) TailorNet [23]. (g) Our method." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "bbox": [ + 45, + 427, + 281, + 730 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 427, + 281, + 730 + ], + "spans": [ + { + "bbox": [ + 45, + 427, + 281, + 730 + ], + "type": "text", + "content": "We show the qualitative comparisons in Figure 3 and list the quantitative results in Table 1. We only compute the error of the input view for ICON since they use their own camera system. Our method performs significantly better than the compared methods [12,23,36] without a cloth template. BCNet [12] also separates the modeling of the clothing and body, but as a supervised learning method, the performance may degrade dramatically because of data domain gap. When the human is moving, BCNet fails to recover the dresses accurately due to arbitrary cloth deformation. In contrast, the appearance loss of our double-layer NeRF-s poses dense constraints on the clothing and body motion capture in a weakly-supervised manner. As a result, our method can achieve a much higher accuracy of geometry tracking even without the ground truth. The predicted garments of TailorNet [23] are not consistent with the real images since their model is conditioned on human poses and shapes. ICON obtains higher IoU accuracy by using the human mask in recovering the geometry. However, there are obvious artifacts on the recovered dresses of ICON [36]. Based on one-piece personalized template, it is hard for both DeepCap [10] and NerfCap [34] to extract the clothing motion, which is important for downstream applications in AR/VR. In contrast, our method tracks the motion for both the clothing and body. Without a pre-scanned cloth template, our method still obtains a high tracking accuracy" + } + ] + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 305, + 104, + 524, + 154 + ], + "blocks": [ + { + "bbox": [ + 305, + 104, + 524, + 154 + ], + "lines": [ + { + "bbox": [ + 305, + 104, + 524, + 154 + ], + "spans": [ + { + "bbox": [ + 305, + 104, + 524, + 154 + ], + "type": "table", + "html": "
MethodS4 of [10]‘FranziRed” of [8]
PSNRSSIMPSNRSSIM
NerfCap [34]24.5290.7923.1891.49
Our method23.0690.2821.7090.54
", + "image_path": "54e80774fb5acbf141fa10a175ca80032e1281ddec166103e1cfbaeb2c0f9092.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "type": "table", + "bbox": [ + 310, + 177, + 518, + 285 + ], + "blocks": [ + { + "bbox": [ + 297, + 155, + 529, + 175 + ], + "lines": [ + { + "bbox": [ + 297, + 155, + 529, + 175 + ], + "spans": [ + { + "bbox": [ + 297, + 155, + 529, + 175 + ], + "type": "text", + "content": "Table 2. Comparison to NerfCap [34] in terms of novel-view synthesis on two datasets." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 310, + 177, + 518, + 285 + ], + "lines": [ + { + "bbox": [ + 310, + 177, + 518, + 285 + ], + "spans": [ + { + "bbox": [ + 310, + 177, + 518, + 285 + ], + "type": "table", + "html": "
MethodAMVIoUPSNRSSIM
w/o simulation loss89.0322.6689.35
w/o mask loss88.7523.2189.49
w/o color loss89.36--
blending render [6]88.7523.6490.51
1 camera view88.4921.3487.12
2 camera views89.2321.8988.38
6 camera views90.1023.8090.60
ours(4 views)89.5823.0690.28
", + "image_path": "cab623925684cc5840f29ec54590789ecc64a3609c95233775725ef32c2c547a.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_body" + } + ], + "index": 13 + }, + { + "bbox": [ + 308, + 287, + 518, + 298 + ], + "lines": [ + { + "bbox": [ + 308, + 287, + 518, + 298 + ], + "spans": [ + { + "bbox": [ + 308, + 287, + 518, + 298 + ], + "type": "text", + "content": "Table 3. Ablation studies on S4 from DeepCap dataset [10]." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 297, + 309, + 442, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 309, + 442, + 319 + ], + "spans": [ + { + "bbox": [ + 297, + 309, + 442, + 319 + ], + "type": "text", + "content": "thanks to the implicit representation." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 296, + 320, + 530, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 320, + 530, + 449 + ], + "spans": [ + { + "bbox": [ + 296, + 320, + 530, + 449 + ], + "type": "text", + "content": "We also compare the quality of free-viewpoint synthesis with NerfCap [34]. Figure 4 shows two compared examples and Table 2 reports the quantitative comparison. Without using any clothing template, our method still obtains close or comparable synthesis quality to NerfCap that uses a ground truth template. Our results are rendered by composing the NeRFs of the body and clothing, which is more complicated than a single NeRFs of NerfCap. For example, when the mutual motion occurs between the body and clothing (like cloth sliding), new body parts may appear, leading to the decrease of our synthesis quality." + } + ] + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 317, + 455, + 345, + 538 + ], + "blocks": [ + { + "bbox": [ + 317, + 455, + 345, + 538 + ], + "lines": [ + { + "bbox": [ + 317, + 455, + 345, + 538 + ], + "spans": [ + { + "bbox": [ + 317, + 455, + 345, + 538 + ], + "type": "image", + "image_path": "b31ed73df970870e0ac60c671fb255b4b56a6dec0a49dcb91bc50f4c4e39e3ad.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 327, + 538, + 337, + 548 + ], + "lines": [ + { + "bbox": [ + 327, + 538, + 337, + 548 + ], + "spans": [ + { + "bbox": [ + 327, + 538, + 337, + 548 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 297, + 549, + 529, + 571 + ], + "lines": [ + { + "bbox": [ + 297, + 549, + 529, + 571 + ], + "spans": [ + { + "bbox": [ + 297, + 549, + 529, + 571 + ], + "type": "text", + "content": "Figure 4. Comparison of novel-view synthesis with NerfCap [34]. (a,d) Ground truth. (b,e) NerfCap [34]. (c,f) Our method." + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 346, + 455, + 373, + 538 + ], + "blocks": [ + { + "bbox": [ + 346, + 455, + 373, + 538 + ], + "lines": [ + { + "bbox": [ + 346, + 455, + 373, + 538 + ], + "spans": [ + { + "bbox": [ + 346, + 455, + 373, + 538 + ], + "type": "image", + "image_path": "4b6c2bf4ed8e4d070e86dd39d566ddb6a26ddc5765438fc0fd4defb254a15aab.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 354, + 538, + 364, + 548 + ], + "lines": [ + { + "bbox": [ + 354, + 538, + 364, + 548 + ], + "spans": [ + { + "bbox": [ + 354, + 538, + 364, + 548 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 374, + 456, + 401, + 538 + ], + "blocks": [ + { + "bbox": [ + 374, + 456, + 401, + 538 + ], + "lines": [ + { + "bbox": [ + 374, + 456, + 401, + 538 + ], + "spans": [ + { + "bbox": [ + 374, + 456, + 401, + 538 + ], + "type": "image", + "image_path": "a25fb15d5ad67a3b9a172f16580039cd498fa17f3bae6ab361029b724c0a5e22.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 381, + 538, + 392, + 548 + ], + "lines": [ + { + "bbox": [ + 381, + 538, + 392, + 548 + ], + "spans": [ + { + "bbox": [ + 381, + 538, + 392, + 548 + ], + "type": "text", + "content": "(c)" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 405, + 458, + 440, + 538 + ], + "blocks": [ + { + "bbox": [ + 405, + 458, + 440, + 538 + ], + "lines": [ + { + "bbox": [ + 405, + 458, + 440, + 538 + ], + "spans": [ + { + "bbox": [ + 405, + 458, + 440, + 538 + ], + "type": "image", + "image_path": "9db08b2b3e6b76f75dd2bce570b5b16060b886a071774cef6c4b96c8e6b6dfb6.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 418, + 538, + 428, + 548 + ], + "lines": [ + { + "bbox": [ + 418, + 538, + 428, + 548 + ], + "spans": [ + { + "bbox": [ + 418, + 538, + 428, + 548 + ], + "type": "text", + "content": "(d)" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 441, + 458, + 475, + 538 + ], + "blocks": [ + { + "bbox": [ + 441, + 458, + 475, + 538 + ], + "lines": [ + { + "bbox": [ + 441, + 458, + 475, + 538 + ], + "spans": [ + { + "bbox": [ + 441, + 458, + 475, + 538 + ], + "type": "image", + "image_path": "cd2c2cea695505c9a22d825c8c53bf9392a26ef91e14bbdb6e0c842d7fa094d2.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 452, + 538, + 462, + 548 + ], + "lines": [ + { + "bbox": [ + 452, + 538, + 462, + 548 + ], + "spans": [ + { + "bbox": [ + 452, + 538, + 462, + 548 + ], + "type": "text", + "content": "(e)" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_caption" + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 476, + 458, + 509, + 538 + ], + "blocks": [ + { + "bbox": [ + 476, + 458, + 509, + 538 + ], + "lines": [ + { + "bbox": [ + 476, + 458, + 509, + 538 + ], + "spans": [ + { + "bbox": [ + 476, + 458, + 509, + 538 + ], + "type": "image", + "image_path": "a4ea041c9b8c5221d0d53fa4a1e91d258adf299ff1937c3ff6548ee5a3325bed.jpg" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 488, + 538, + 496, + 548 + ], + "lines": [ + { + "bbox": [ + 488, + 538, + 496, + 548 + ], + "spans": [ + { + "bbox": [ + 488, + 538, + 496, + 548 + ], + "type": "text", + "content": "(f)" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_caption" + } + ], + "index": 27 + }, + { + "bbox": [ + 297, + 582, + 393, + 594 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 582, + 393, + 594 + ], + "spans": [ + { + "bbox": [ + 297, + 582, + 393, + 594 + ], + "type": "text", + "content": "4.3. Ablation studies" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 296, + 601, + 530, + 636 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 601, + 530, + 636 + ], + "spans": [ + { + "bbox": [ + 296, + 601, + 530, + 636 + ], + "type": "text", + "content": "The ablation experiments are performed to demonstrate the effectiveness of our framework on the S4 from DeepCap dataset [10]. The quantitative results are listed in Table 3." + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 296, + 636, + 530, + 730 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 636, + 530, + 730 + ], + "spans": [ + { + "bbox": [ + 296, + 636, + 530, + 730 + ], + "type": "text", + "content": "Effectiveness of different losses. We first evaluate the effectiveness of the main losses in our optimization, i.e., the simulation loss, the mask loss, and the color loss. We report the accuracy by removing each of them, and show qualitative comparison examples in Figure 5. Without the simulation loss, the body-cloth collision becomes severe, showing that the simulation loss can preserve physically-plausible body-cloth interactions. Using the input human mask can" + } + ] + } + ], + "index": 32 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 276, + 746, + 300, + 756 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 276, + 746, + 300, + 756 + ], + "spans": [ + { + "bbox": [ + 276, + 746, + 300, + 756 + ], + "type": "text", + "content": "21104" + } + ] + } + ], + "index": 33 + } + ], + "page_size": [ + 595, + 842 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 45, + 106, + 279, + 153 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 106, + 279, + 153 + ], + "spans": [ + { + "bbox": [ + 45, + 106, + 279, + 153 + ], + "type": "text", + "content": "improve the alignment accuracy with the human silhouettes. The color loss of dynamic NeRF can further enhance the detailed deformation on the human surface by posing dense appearance constraints for the geometry tracking." + } + ] + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 73, + 159, + 108, + 234 + ], + "blocks": [ + { + "bbox": [ + 73, + 159, + 108, + 234 + ], + "lines": [ + { + "bbox": [ + 73, + 159, + 108, + 234 + ], + "spans": [ + { + "bbox": [ + 73, + 159, + 108, + 234 + ], + "type": "image", + "image_path": "2cafdd527f3a2550598e5a1ade66938596b8b9596ec78c10737c0b37c78c0f11.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 108, + 159, + 143, + 234 + ], + "blocks": [ + { + "bbox": [ + 108, + 159, + 143, + 234 + ], + "lines": [ + { + "bbox": [ + 108, + 159, + 143, + 234 + ], + "spans": [ + { + "bbox": [ + 108, + 159, + 143, + 234 + ], + "type": "image", + "image_path": "4d736e527bb28a2679c32af1379f3c4f5499cb79e9066a309d145f9ee5ace126.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 145, + 159, + 180, + 234 + ], + "blocks": [ + { + "bbox": [ + 145, + 159, + 180, + 234 + ], + "lines": [ + { + "bbox": [ + 145, + 159, + 180, + 234 + ], + "spans": [ + { + "bbox": [ + 145, + 159, + 180, + 234 + ], + "type": "image", + "image_path": "1766e4d22506c87b72375e406ff51d7835b77fd9b23b1b9a2afa8852aba8d93a.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 181, + 159, + 215, + 234 + ], + "blocks": [ + { + "bbox": [ + 181, + 159, + 215, + 234 + ], + "lines": [ + { + "bbox": [ + 181, + 159, + 215, + 234 + ], + "spans": [ + { + "bbox": [ + 181, + 159, + 215, + 234 + ], + "type": "image", + "image_path": "c79b7ed46ec3a8a1bafa4155948909b1c12191c4f583d9a179b50a9c545e5544.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 217, + 159, + 252, + 234 + ], + "blocks": [ + { + "bbox": [ + 217, + 159, + 252, + 234 + ], + "lines": [ + { + "bbox": [ + 217, + 159, + 252, + 234 + ], + "spans": [ + { + "bbox": [ + 217, + 159, + 252, + 234 + ], + "type": "image", + "image_path": "e7bed27739caa1a681482e887dc055a37127764a71a461e8ff4cab7f6f296a95.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 74, + 235, + 108, + 309 + ], + "blocks": [ + { + "bbox": [ + 74, + 235, + 108, + 309 + ], + "lines": [ + { + "bbox": [ + 74, + 235, + 108, + 309 + ], + "spans": [ + { + "bbox": [ + 74, + 235, + 108, + 309 + ], + "type": "image", + "image_path": "1724da4f1adf14bc96d2506cd00674659143a08ffed2de310f3e4e9c25ebeb36.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 86, + 310, + 96, + 320 + ], + "lines": [ + { + "bbox": [ + 86, + 310, + 96, + 320 + ], + "spans": [ + { + "bbox": [ + 86, + 310, + 96, + 320 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 108, + 235, + 143, + 309 + ], + "blocks": [ + { + "bbox": [ + 108, + 235, + 143, + 309 + ], + "lines": [ + { + "bbox": [ + 108, + 235, + 143, + 309 + ], + "spans": [ + { + "bbox": [ + 108, + 235, + 143, + 309 + ], + "type": "image", + "image_path": "aa1efa9baa10faff6110740345a60648ac0dccb7f31a35de16fc3dbd19552feb.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 120, + 310, + 131, + 319 + ], + "lines": [ + { + "bbox": [ + 120, + 310, + 131, + 319 + ], + "spans": [ + { + "bbox": [ + 120, + 310, + 131, + 319 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 145, + 235, + 180, + 309 + ], + "blocks": [ + { + "bbox": [ + 145, + 235, + 180, + 309 + ], + "lines": [ + { + "bbox": [ + 145, + 235, + 180, + 309 + ], + "spans": [ + { + "bbox": [ + 145, + 235, + 180, + 309 + ], + "type": "image", + "image_path": "d8231ca330ce52dd0abb612881dbc9a4ca8db648163bd6e74e2ab55f7807d1fb.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 158, + 310, + 167, + 319 + ], + "lines": [ + { + "bbox": [ + 158, + 310, + 167, + 319 + ], + "spans": [ + { + "bbox": [ + 158, + 310, + 167, + 319 + ], + "type": "text", + "content": "(c)" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 181, + 235, + 215, + 309 + ], + "blocks": [ + { + "bbox": [ + 181, + 235, + 215, + 309 + ], + "lines": [ + { + "bbox": [ + 181, + 235, + 215, + 309 + ], + "spans": [ + { + "bbox": [ + 181, + 235, + 215, + 309 + ], + "type": "image", + "image_path": "18e9d9f5c403d40fe372bae16fc436608f3be6fa451955b4702ca51345da250f.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 192, + 310, + 204, + 319 + ], + "lines": [ + { + "bbox": [ + 192, + 310, + 204, + 319 + ], + "spans": [ + { + "bbox": [ + 192, + 310, + 204, + 319 + ], + "type": "text", + "content": "(d)" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 217, + 235, + 252, + 309 + ], + "blocks": [ + { + "bbox": [ + 217, + 235, + 252, + 309 + ], + "lines": [ + { + "bbox": [ + 217, + 235, + 252, + 309 + ], + "spans": [ + { + "bbox": [ + 217, + 235, + 252, + 309 + ], + "type": "image", + "image_path": "7c308394e2760380e490274e723d9cac1c30133251f4fcce88de743707d497db.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 48, + 406, + 80, + 496 + ], + "blocks": [ + { + "bbox": [ + 45, + 322, + 279, + 397 + ], + "lines": [ + { + "bbox": [ + 45, + 322, + 279, + 397 + ], + "spans": [ + { + "bbox": [ + 45, + 322, + 279, + 397 + ], + "type": "text", + "content": "Figure 5. Ablation study for different losses. (a) Ground truth. (b) Using complete losses. (c) W/o simulation loss. (d) W/o mask loss. (e) W/o color loss. There are unnatural cloth deformations without the simulation loss. The rendered mask cannot align with the silhouette well without the mask loss. Compared to that without color loss, the surface detailed deformation can be further improved using complete losses (b)." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 48, + 406, + 80, + 496 + ], + "lines": [ + { + "bbox": [ + 48, + 406, + 80, + 496 + ], + "spans": [ + { + "bbox": [ + 48, + 406, + 80, + 496 + ], + "type": "image", + "image_path": "79ee651b5e22c62d990dd3556beff6d0154366dca669112e480546541eeba10d.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 58, + 499, + 69, + 508 + ], + "lines": [ + { + "bbox": [ + 58, + 499, + 69, + 508 + ], + "spans": [ + { + "bbox": [ + 58, + 499, + 69, + 508 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 93, + 408, + 119, + 496 + ], + "blocks": [ + { + "bbox": [ + 93, + 408, + 119, + 496 + ], + "lines": [ + { + "bbox": [ + 93, + 408, + 119, + 496 + ], + "spans": [ + { + "bbox": [ + 93, + 408, + 119, + 496 + ], + "type": "image", + "image_path": "5e97de6abcc6dd2fd0494bb139acfce978e0a9275dc0fe761e051e706832b6f3.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 129, + 499, + 140, + 508 + ], + "lines": [ + { + "bbox": [ + 129, + 499, + 140, + 508 + ], + "spans": [ + { + "bbox": [ + 129, + 499, + 140, + 508 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 120, + 408, + 151, + 495 + ], + "blocks": [ + { + "bbox": [ + 120, + 408, + 151, + 495 + ], + "lines": [ + { + "bbox": [ + 120, + 408, + 151, + 495 + ], + "spans": [ + { + "bbox": [ + 120, + 408, + 151, + 495 + ], + "type": "image", + "image_path": "c152489fa55ea4a2a8c57e0c7574c55aef8536ad8288d2b2217715f51ab727e0.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 45, + 511, + 279, + 564 + ], + "lines": [ + { + "bbox": [ + 45, + 511, + 279, + 564 + ], + "spans": [ + { + "bbox": [ + 45, + 511, + 279, + 564 + ], + "type": "text", + "content": "Figure 6. Comparison of different rendering. (a) Ground truth. (b) Our composite rendering. (c) Blend rendering [6]. We show the synthesis of the clothing, body, and full human. There are much cloth appearance on the body caused by the blending [6], while our method separates the appearance of the clothing and body better." + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 152, + 408, + 181, + 495 + ], + "blocks": [ + { + "bbox": [ + 152, + 408, + 181, + 495 + ], + "lines": [ + { + "bbox": [ + 152, + 408, + 181, + 495 + ], + "spans": [ + { + "bbox": [ + 152, + 408, + 181, + 495 + ], + "type": "image", + "image_path": "0be68f33d69b1ca701cedbe9cb6c1297bddd553ac882b75fddd130432b06c198.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 192, + 419, + 218, + 467 + ], + "blocks": [ + { + "bbox": [ + 192, + 419, + 218, + 467 + ], + "lines": [ + { + "bbox": [ + 192, + 419, + 218, + 467 + ], + "spans": [ + { + "bbox": [ + 192, + 419, + 218, + 467 + ], + "type": "image", + "image_path": "64ffa03747a1eda5401011102a27b6b127ce8c51d2d269a3097e0e9bef886d47.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 220, + 408, + 251, + 495 + ], + "blocks": [ + { + "bbox": [ + 220, + 408, + 251, + 495 + ], + "lines": [ + { + "bbox": [ + 220, + 408, + 251, + 495 + ], + "spans": [ + { + "bbox": [ + 220, + 408, + 251, + 495 + ], + "type": "image", + "image_path": "52a19c0ba71493b4000011081db876501e483239da9460d86251f195d1db8658.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 230, + 499, + 240, + 508 + ], + "lines": [ + { + "bbox": [ + 230, + 499, + 240, + 508 + ], + "spans": [ + { + "bbox": [ + 230, + 499, + 240, + 508 + ], + "type": "text", + "content": "(c)" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 252, + 408, + 282, + 495 + ], + "blocks": [ + { + "bbox": [ + 252, + 408, + 282, + 495 + ], + "lines": [ + { + "bbox": [ + 252, + 408, + 282, + 495 + ], + "spans": [ + { + "bbox": [ + 252, + 408, + 282, + 495 + ], + "type": "image", + "image_path": "5aadb97f34f34b404da349258615dfd2c46bc7e5b4b789249abd4ab597f21b12.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + } + ], + "index": 24 + }, + { + "bbox": [ + 45, + 566, + 278, + 682 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 566, + 278, + 682 + ], + "spans": [ + { + "bbox": [ + 45, + 566, + 278, + 682 + ], + "type": "text", + "content": "Effectiveness of the composite rendering. We compare our composite rendering with the blending rendering in [6]. Through free composition of two NeRFs for each point, the method of [6] can obtain high fitting color. However, the blending may generate missing areas or artifacts in the novel-view synthesis. Moreover, it fails to recover the appearance of the body and clothing reliably as shown in Figure 6. Comparatively, we can recover the appearance of the body and clothing separately and compose them into a full human more accurately using the composite rendering." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 45, + 683, + 278, + 730 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 683, + 278, + 730 + ], + "spans": [ + { + "bbox": [ + 45, + 683, + 278, + 730 + ], + "type": "text", + "content": "The number of camera views. We investigate the influence of different number of input camera views on the performance of our approach. We compare the results of our method under " + }, + { + "bbox": [ + 45, + 683, + 278, + 730 + ], + "type": "inline_equation", + "content": "\\{1,2,4,6\\}" + }, + { + "bbox": [ + 45, + 683, + 278, + 730 + ], + "type": "text", + "content": " views, and show the comparison" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 296, + 106, + 530, + 141 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 106, + 530, + 141 + ], + "spans": [ + { + "bbox": [ + 296, + 106, + 530, + 141 + ], + "type": "text", + "content": "example of novel-view synthesis in Figure 7. The results become slightly better using more camera views because the moving humans are captured more completely." + } + ] + } + ], + "index": 29 + }, + { + "type": "image", + "bbox": [ + 326, + 149, + 362, + 241 + ], + "blocks": [ + { + "bbox": [ + 326, + 149, + 362, + 241 + ], + "lines": [ + { + "bbox": [ + 326, + 149, + 362, + 241 + ], + "spans": [ + { + "bbox": [ + 326, + 149, + 362, + 241 + ], + "type": "image", + "image_path": "d34100012af1e0ae21a4d54372f1f017c0e088b6729939077a77d3d7391303ff.jpg" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 297, + 252, + 530, + 284 + ], + "lines": [ + { + "bbox": [ + 297, + 252, + 530, + 284 + ], + "spans": [ + { + "bbox": [ + 297, + 252, + 530, + 284 + ], + "type": "text", + "content": "Figure 7. Ablation study for different numbers of camera views. With more cameras, the novel-view synthesis quality becomes better (marked in cycle)." + } + ] + } + ], + "index": 40, + "angle": 0, + "type": "image_caption" + } + ], + "index": 30 + }, + { + "type": "image", + "bbox": [ + 364, + 149, + 396, + 241 + ], + "blocks": [ + { + "bbox": [ + 324, + 244, + 362, + 251 + ], + "lines": [ + { + "bbox": [ + 324, + 244, + 362, + 251 + ], + "spans": [ + { + "bbox": [ + 324, + 244, + 362, + 251 + ], + "type": "text", + "content": "Ground Truth" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 364, + 149, + 396, + 241 + ], + "lines": [ + { + "bbox": [ + 364, + 149, + 396, + 241 + ], + "spans": [ + { + "bbox": [ + 364, + 149, + 396, + 241 + ], + "type": "image", + "image_path": "d43466951f76608d0ce7b92136b54930a7814df294209549227d43b1c06c6b0c.jpg" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 371, + 244, + 388, + 251 + ], + "lines": [ + { + "bbox": [ + 371, + 244, + 388, + 251 + ], + "spans": [ + { + "bbox": [ + 371, + 244, + 388, + 251 + ], + "type": "text", + "content": "1 view" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_caption" + } + ], + "index": 32 + }, + { + "type": "image", + "bbox": [ + 398, + 149, + 431, + 241 + ], + "blocks": [ + { + "bbox": [ + 398, + 149, + 431, + 241 + ], + "lines": [ + { + "bbox": [ + 398, + 149, + 431, + 241 + ], + "spans": [ + { + "bbox": [ + 398, + 149, + 431, + 241 + ], + "type": "image", + "image_path": "61a22e02dc1f226e377f5244f301979de0d86d0ee3925d60d2b0bfb408662ad4.jpg" + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 405, + 244, + 427, + 251 + ], + "lines": [ + { + "bbox": [ + 405, + 244, + 427, + 251 + ], + "spans": [ + { + "bbox": [ + 405, + 244, + 427, + 251 + ], + "type": "text", + "content": "2 views" + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "image_caption" + } + ], + "index": 34 + }, + { + "type": "image", + "bbox": [ + 433, + 149, + 467, + 241 + ], + "blocks": [ + { + "bbox": [ + 433, + 149, + 467, + 241 + ], + "lines": [ + { + "bbox": [ + 433, + 149, + 467, + 241 + ], + "spans": [ + { + "bbox": [ + 433, + 149, + 467, + 241 + ], + "type": "image", + "image_path": "f8cbdcbc3c5ae4e5a2719bb517038dfd695e299c91495ddf922d5cf52fb223af.jpg" + } + ] + } + ], + "index": 36, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 440, + 244, + 462, + 250 + ], + "lines": [ + { + "bbox": [ + 440, + 244, + 462, + 250 + ], + "spans": [ + { + "bbox": [ + 440, + 244, + 462, + 250 + ], + "type": "text", + "content": "4 views" + } + ] + } + ], + "index": 37, + "angle": 0, + "type": "image_caption" + } + ], + "index": 36 + }, + { + "type": "image", + "bbox": [ + 468, + 149, + 503, + 241 + ], + "blocks": [ + { + "bbox": [ + 468, + 149, + 503, + 241 + ], + "lines": [ + { + "bbox": [ + 468, + 149, + 503, + 241 + ], + "spans": [ + { + "bbox": [ + 468, + 149, + 503, + 241 + ], + "type": "image", + "image_path": "9da0dea23bbf11da7f27bf2335f933d2b1f86d3f758d49e35bf1e02ff5ee8b20.jpg" + } + ] + } + ], + "index": 38, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 474, + 244, + 496, + 250 + ], + "lines": [ + { + "bbox": [ + 474, + 244, + 496, + 250 + ], + "spans": [ + { + "bbox": [ + 474, + 244, + 496, + 250 + ], + "type": "text", + "content": "6 views" + } + ] + } + ], + "index": 39, + "angle": 0, + "type": "image_caption" + } + ], + "index": 38 + }, + { + "bbox": [ + 297, + 300, + 376, + 313 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 300, + 376, + 313 + ], + "spans": [ + { + "bbox": [ + 297, + 300, + 376, + 313 + ], + "type": "text", + "content": "4.4. Applications" + } + ] + } + ], + "index": 41 + }, + { + "bbox": [ + 296, + 318, + 530, + 400 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 318, + 530, + 400 + ], + "spans": [ + { + "bbox": [ + 296, + 318, + 530, + 400 + ], + "type": "text", + "content": "Benefiting from separate tracking results of human body and clothing, we can enable interesting applications like cloth retargeting as shown in Figure 8. Since the body mesh and cloth mesh are captured, we can exchange the garments between two person by simulating the garments on the other body. Moreover, free-viewpoint motion can be created by simulating the clothing on human body sequences." + } + ] + } + ], + "index": 42 + }, + { + "type": "image", + "bbox": [ + 311, + 407, + 362, + 484 + ], + "blocks": [ + { + "bbox": [ + 311, + 407, + 362, + 484 + ], + "lines": [ + { + "bbox": [ + 311, + 407, + 362, + 484 + ], + "spans": [ + { + "bbox": [ + 311, + 407, + 362, + 484 + ], + "type": "image", + "image_path": "d2a02633a71bc2f21ee5194b63edd462afd85e4c557463deea43581dba62d4be.jpg" + } + ] + } + ], + "index": 43, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 296, + 487, + 530, + 520 + ], + "lines": [ + { + "bbox": [ + 296, + 487, + 530, + 520 + ], + "spans": [ + { + "bbox": [ + 296, + 487, + 530, + 520 + ], + "type": "text", + "content": "Figure 8. Application of cloth retargeting. From the captured results using our method, we can retarget the garments between the S4 of DeepCap and the 'FranziRed' of DynaCap." + } + ] + } + ], + "index": 49, + "angle": 0, + "type": "image_caption" + } + ], + "index": 43 + }, + { + "type": "image", + "bbox": [ + 343, + 408, + 378, + 483 + ], + "blocks": [ + { + "bbox": [ + 343, + 408, + 378, + 483 + ], + "lines": [ + { + "bbox": [ + 343, + 408, + 378, + 483 + ], + "spans": [ + { + "bbox": [ + 343, + 408, + 378, + 483 + ], + "type": "image", + "image_path": "04e8d1615ac71b0c66e619a82bed89d58395189f9065a4bb4a2345125442166e.jpg" + } + ] + } + ], + "index": 44, + "angle": 0, + "type": "image_body" + } + ], + "index": 44 + }, + { + "type": "image", + "bbox": [ + 384, + 408, + 407, + 483 + ], + "blocks": [ + { + "bbox": [ + 384, + 408, + 407, + 483 + ], + "lines": [ + { + "bbox": [ + 384, + 408, + 407, + 483 + ], + "spans": [ + { + "bbox": [ + 384, + 408, + 407, + 483 + ], + "type": "image", + "image_path": "a783863e72fc48f9fb39e39b52c50ce68a20bea0a10d9452f9e4c4b6bacce24f.jpg" + } + ] + } + ], + "index": 45, + "angle": 0, + "type": "image_body" + } + ], + "index": 45 + }, + { + "type": "image", + "bbox": [ + 413, + 408, + 445, + 483 + ], + "blocks": [ + { + "bbox": [ + 413, + 408, + 445, + 483 + ], + "lines": [ + { + "bbox": [ + 413, + 408, + 445, + 483 + ], + "spans": [ + { + "bbox": [ + 413, + 408, + 445, + 483 + ], + "type": "image", + "image_path": "d48f036eb9b4af9b10453d785a9649e0b97c099067737d9949a0adef63d2eea0.jpg" + } + ] + } + ], + "index": 46, + "angle": 0, + "type": "image_body" + } + ], + "index": 46 + }, + { + "type": "image", + "bbox": [ + 454, + 408, + 480, + 483 + ], + "blocks": [ + { + "bbox": [ + 454, + 408, + 480, + 483 + ], + "lines": [ + { + "bbox": [ + 454, + 408, + 480, + 483 + ], + "spans": [ + { + "bbox": [ + 454, + 408, + 480, + 483 + ], + "type": "image", + "image_path": "1b1772b24e41c5a1d224b24850638f1b56d14c5ebe4c2dd63961fbe11dd3f3a0.jpg" + } + ] + } + ], + "index": 47, + "angle": 0, + "type": "image_body" + } + ], + "index": 47 + }, + { + "type": "image", + "bbox": [ + 488, + 408, + 518, + 483 + ], + "blocks": [ + { + "bbox": [ + 488, + 408, + 518, + 483 + ], + "lines": [ + { + "bbox": [ + 488, + 408, + 518, + 483 + ], + "spans": [ + { + "bbox": [ + 488, + 408, + 518, + 483 + ], + "type": "image", + "image_path": "7c03eb5680cda026af5bc8f3493b915476c20c3aa1a02398bd4071445cad4686.jpg" + } + ] + } + ], + "index": 48, + "angle": 0, + "type": "image_body" + } + ], + "index": 48 + }, + { + "bbox": [ + 297, + 534, + 368, + 546 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 534, + 368, + 546 + ], + "spans": [ + { + "bbox": [ + 297, + 534, + 368, + 546 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 50 + }, + { + "bbox": [ + 296, + 554, + 530, + 682 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 554, + 530, + 682 + ], + "spans": [ + { + "bbox": [ + 296, + 554, + 530, + 682 + ], + "type": "text", + "content": "In this paper, we proposed a novel performance capture method for clothed humans with a double-layer neural radiance fields (NeRF). Our double-layer NeRFs can effectively model moving humans in daily-life clothes. By matching the rendered color with the observed color, our method estimates the clothing and body motion and temporally-varying appearance robustly. In addition, we learn a differential physics-aware network for different garment types which supports physically-plausible body-cloth interactions and cloth dynamics. The experimental results on real datasets demonstrate the effectiveness of the proposed method." + } + ] + } + ], + "index": 51 + }, + { + "bbox": [ + 296, + 683, + 530, + 730 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 683, + 530, + 730 + ], + "spans": [ + { + "bbox": [ + 296, + 683, + 530, + 730 + ], + "type": "text", + "content": "Acknowledgments This work was supported by the Fundamental Research Funds for the Central Universities (NJ2022028), and the Natural Science Foundation of China (No.U1713208), and Program for Changjiang Scholars." + } + ] + } + ], + "index": 52 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 276, + 746, + 300, + 756 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 276, + 746, + 300, + 756 + ], + "spans": [ + { + "bbox": [ + 276, + 746, + 300, + 756 + ], + "type": "text", + "content": "21105" + } + ] + } + ], + "index": 53 + } + ], + "page_size": [ + 595, + 842 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 105, + 103, + 117 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 105, + 103, + 117 + ], + "spans": [ + { + "bbox": [ + 47, + 105, + 103, + 117 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 124, + 279, + 729 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 51, + 124, + 264, + 135 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 51, + 124, + 264, + 135 + ], + "spans": [ + { + "bbox": [ + 51, + 124, + 264, + 135 + ], + "type": "text", + "content": "[1] https://www.marvelousdesigner.com/.3,5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 51, + 136, + 279, + 168 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 51, + 136, + 279, + 168 + ], + "spans": [ + { + "bbox": [ + 51, + 136, + 279, + 168 + ], + "type": "text", + "content": "[2] Thiemo Alldieck, Gerard Pons-Moll, Christian Theobalt, and Marcus Magnor. Tex2Shape: Detailed full human body geometry from a single image. In ICCV, 2019. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 51, + 168, + 279, + 200 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 51, + 168, + 279, + 200 + ], + "spans": [ + { + "bbox": [ + 51, + 168, + 279, + 200 + ], + "type": "text", + "content": "[3] Bharat Lal Bhatnagar, Garvita Tiwari, Christian Theobalt, and Gerard Pons-Moll. Multi-Garment Net: Learning to dress 3D people from images. In ICCV, 2019. 1, 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 51, + 201, + 279, + 264 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 51, + 201, + 279, + 264 + ], + "spans": [ + { + "bbox": [ + 51, + 201, + 279, + 264 + ], + "type": "text", + "content": "[4] Mingsong Dou, Sameh Khamis, Yury Degtyarev, Philip Davidson, Sean Fanello, Adarsh Kowdle, Sergio Orts Escolano, Christoph Rhemann, David Kim, Jonathan Taylor, Pushmeet Kohli, Vladimir Tankovich, and Shahram Izadi. Fusion4D: Real-time performance capture of challenging scenes. ACM TOG, 35:1-13, July 2016. 1, 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 51, + 265, + 279, + 307 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 51, + 265, + 279, + 307 + ], + "spans": [ + { + "bbox": [ + 51, + 265, + 279, + 307 + ], + "type": "text", + "content": "[5] Stuart Ganan and D McClure. Bayesian image analysis: An application to single photon emission tomography. In Proceedings of the American Statistical Association, pages 12-18, 1985. 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 51, + 309, + 279, + 340 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 51, + 309, + 279, + 340 + ], + "spans": [ + { + "bbox": [ + 51, + 309, + 279, + 340 + ], + "type": "text", + "content": "[6] Chen Gao, Ayush Saraf, Johannes Kopf, and Jia-Bin Huang. Dynamic view synthesis from dynamic monocular video. In ICCV, 2021. 4, 7, 8" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 51, + 341, + 279, + 383 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 51, + 341, + 279, + 383 + ], + "spans": [ + { + "bbox": [ + 51, + 341, + 279, + 383 + ], + "type": "text", + "content": "[7] Erhan Gundogdu, Victor Constantin, Amrollah Seifoddini, Minh Dang, Mathieu Salzmann, and Pascal Fua. GarNet: A two-stream network for fast and accurate 3d cloth draping. In ICCV, 2019. 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 51, + 384, + 279, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 51, + 384, + 279, + 426 + ], + "spans": [ + { + "bbox": [ + 51, + 384, + 279, + 426 + ], + "type": "text", + "content": "[8] Marc Habermann, Lingjie Liu, Weipeng Xu, Michael Zollhoefer, Gerard Pons-Moll, and Christian Theobalt. Real-time deep dynamic characters. ACM Transactions on Graphics, 40(4), August 2021. 2, 5, 6, 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 51, + 427, + 279, + 469 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 51, + 427, + 279, + 469 + ], + "spans": [ + { + "bbox": [ + 51, + 427, + 279, + 469 + ], + "type": "text", + "content": "[9] Marc Habermann, Weipeng Xu, Michael Zollhoefer, Gerard Pons-Moll, and Christian Theobalt. LiveCap: Real-time human performance capture from monocular video. ACM TOG, 38(2):14:1-14:17, March 2019. 1, 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 470, + 279, + 512 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 470, + 279, + 512 + ], + "spans": [ + { + "bbox": [ + 47, + 470, + 279, + 512 + ], + "type": "text", + "content": "[10] Marc Habermann, Weipeng Xu, Michael Zollhoefer, Gerard Pons-Moll, and Christian Theobalt. DeepCap: Monocular human performance capture using weak supervision. In CVPR, 2020. 1, 2, 5, 6, 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 513, + 279, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 513, + 279, + 544 + ], + "spans": [ + { + "bbox": [ + 47, + 513, + 279, + 544 + ], + "type": "text", + "content": "[11] Boyi Jiang, Yang Hong, Hujun Bao, and Juyong Zhang. SelfRecon: Self reconstruction your digital avatar from monocular video. In CVPR, 2022. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 47, + 546, + 279, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 546, + 279, + 578 + ], + "spans": [ + { + "bbox": [ + 47, + 546, + 279, + 578 + ], + "type": "text", + "content": "[12] Boyi Jiang, Juyong Zhang, Yang Hong, Jinhao Luo, Ligang Liu, and Hujun Bao. BCNet: Learning body and cloth shape from a single image. In ECCV, 2020. 1, 2, 6, 7" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 47, + 579, + 279, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 579, + 279, + 609 + ], + "spans": [ + { + "bbox": [ + 47, + 579, + 279, + 609 + ], + "type": "text", + "content": "[13] Hanbyul Joo, Tomas Simon, and Yaser Sheikh. Total capture: A 3D deformation model for tracking faces, hands, and bodies. In CVPR, 2018. 4" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 47, + 611, + 279, + 642 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 611, + 279, + 642 + ], + "spans": [ + { + "bbox": [ + 47, + 611, + 279, + 642 + ], + "type": "text", + "content": "[14] Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In International Conference for Learning Representations, 2015. 6" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 47, + 644, + 279, + 676 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 644, + 279, + 676 + ], + "spans": [ + { + "bbox": [ + 47, + 644, + 279, + 676 + ], + "type": "text", + "content": "[15] John P Lewis, Matt Cordner, and Nickson Fong. Pose space deformation: A unified approach to shape interpolation and skeleton-driven deformation. In SIGGRAPH, 2000. 3, 4" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 47, + 676, + 279, + 729 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 676, + 279, + 729 + ], + "spans": [ + { + "bbox": [ + 47, + 676, + 279, + 729 + ], + "type": "text", + "content": "[16] Yue Li, Marc Habermann, Bernhard Thomaszewski, Stelian Coros, Thabo Beeler, and Christian Theobalt. Deep physics-aware inference of cloth deformation for monocular human performance capture. In International Conference on 3D Vision (3DV), 2021. 1, 2, 5, 6" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 299, + 107, + 530, + 729 + ], + "type": "list", + "angle": 0, + "index": 33, + "blocks": [ + { + "bbox": [ + 299, + 107, + 530, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 107, + 530, + 149 + ], + "spans": [ + { + "bbox": [ + 299, + 107, + 530, + 149 + ], + "type": "text", + "content": "[17] Lingjie Liu, Marc Habermann, Viktor Rudnev, Kripasindhu Sarkar, Jiatao Gu, and Christian Theobalt. Neural Actor: Neural free-view synthesis of human actors with pose control. ACM TOG, 40(6):1-16, 2021. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 299, + 152, + 530, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 152, + 530, + 183 + ], + "spans": [ + { + "bbox": [ + 299, + 152, + 530, + 183 + ], + "type": "text", + "content": "[18] Matthew Loper, Naureen Mahmood, Javier Romero, Gerard Pons-Moll, and Michael J. Black. SMPL: A skinned multiperson linear model. ACM TOG, 34(6):248, 2015. 2, 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 299, + 185, + 530, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 185, + 530, + 227 + ], + "spans": [ + { + "bbox": [ + 299, + 185, + 530, + 227 + ], + "type": "text", + "content": "[19] Qianli Ma, Jinlong Yang, Anurag Ranjan, Sergi Pujades, Gerard Pons-Moll, Siyu Tang, and Michael J. Black. Learning to dress 3D people in generative clothing. In CVPR, 2020. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 299, + 229, + 530, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 229, + 530, + 271 + ], + "spans": [ + { + "bbox": [ + 299, + 229, + 530, + 271 + ], + "type": "text", + "content": "[20] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. N-eRF: Representing scenes as neural radiance fields for view synthesis. In ECCV, 2020. 1, 2, 4" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 299, + 274, + 530, + 305 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 274, + 530, + 305 + ], + "spans": [ + { + "bbox": [ + 299, + 274, + 530, + 305 + ], + "type": "text", + "content": "[21] Richard Newcombe, Dieter Fox, and Steve Seitz. Dynamic-Fusion: Reconstruction and tracking of non-rigid scenes in real-time. In CVPR, 2015. 1, 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 299, + 308, + 530, + 340 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 308, + 530, + 340 + ], + "spans": [ + { + "bbox": [ + 299, + 308, + 530, + 340 + ], + "type": "text", + "content": "[22] Michael Oechsle, Songyou Peng, and Andreas Geiger. U-NISURF: Unifying neural implicit surfaces and radiance fields for multi-view reconstruction. In ICCV, 2021. 3, 4" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 299, + 341, + 530, + 383 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 341, + 530, + 383 + ], + "spans": [ + { + "bbox": [ + 299, + 341, + 530, + 383 + ], + "type": "text", + "content": "[23] Chaitanya Patel, Zhouyingcheng Liao, and Gerard Pons-Moll. TailorNet: Predicting clothing in 3D as a function of human pose, shape and garment style. In CVPR, 2020. 2, 5, 6, 7" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 299, + 386, + 530, + 428 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 386, + 530, + 428 + ], + "spans": [ + { + "bbox": [ + 299, + 386, + 530, + 428 + ], + "type": "text", + "content": "[24] Sida Peng, Junting Dong, Qianqian Wang, Shangzhan Zhang, Qing Shuai, Xiaowei Zhou, and Hujun Bao. Animatable neural radiance fields for modeling dynamic human bodies. In ICCV, 2021. 2, 3, 4, 5" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 299, + 431, + 530, + 483 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 431, + 530, + 483 + ], + "spans": [ + { + "bbox": [ + 299, + 431, + 530, + 483 + ], + "type": "text", + "content": "[25] Sida Peng, Yuanqing Zhang, Yinghao Xu, Qianqian Wang, Qing Shuai, Hujun Bao, and Xiaowei Zhou. Neural Body: Implicit neural representations with structured latent codes for novel view synthesis of dynamic humans. In CVPR, 2021. 2, 3, 4" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 299, + 485, + 530, + 517 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 485, + 530, + 517 + ], + "spans": [ + { + "bbox": [ + 299, + 485, + 530, + 517 + ], + "type": "text", + "content": "[26] Gerard Pons-Moll, Sergi Pujades, Sonny Hu, and Michael J. Black. ClothCap: Seamless 4D clothing capture and retargeting. ACM TOG, 36(4), July 2017. 1, 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 299, + 519, + 530, + 561 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 519, + 530, + 561 + ], + "spans": [ + { + "bbox": [ + 299, + 519, + 530, + 561 + ], + "type": "text", + "content": "[27] Shunsuke Saito, Zeng Huang, Ryota Natsume, Shigeo Morishima, Angjoo Kanazawa, and Hao Li. PIFu: Pixel-aligned implicit function for high-resolution clothed human digitization. In ICCV, 2019. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 299, + 563, + 530, + 596 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 563, + 530, + 596 + ], + "spans": [ + { + "bbox": [ + 299, + 563, + 530, + 596 + ], + "type": "text", + "content": "[28] Shunsuke Saito, Tomas Simon, Jason Saragih, and Hanbyul Joo. PIFuHD: Multi-level pixel-aligned implicit function for high-resolution 3D human digitization. In CVPR, 2020. 2" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 299, + 597, + 530, + 629 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 597, + 530, + 629 + ], + "spans": [ + { + "bbox": [ + 299, + 597, + 530, + 629 + ], + "type": "text", + "content": "[29] Igor Santesteban, Miguel A. Otaduy, and Dan Casas. Learning-based animation of clothing for virtual try-on. Computer Graphics Forum (Proc. Eurographics), 2019. 2" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 299, + 631, + 530, + 674 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 631, + 530, + 674 + ], + "spans": [ + { + "bbox": [ + 299, + 631, + 530, + 674 + ], + "type": "text", + "content": "[30] Vincent Sitzmann, Julien NP Martel, Alexander W Bergman, David B Lindell, and Gordon Wetzstein. Implicit neural representations with periodic activation functions. In Advances in Neural Information Processing Systems, 2020. 4, 6" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 299, + 676, + 530, + 729 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 676, + 530, + 729 + ], + "spans": [ + { + "bbox": [ + 299, + 676, + 530, + 729 + ], + "type": "text", + "content": "[31] Zhaoqi Su, Weilin Wan, Tao Yu, Lingjie Liu, Lu Fang, Wenping Wang, and Yebin Liu. MulayCap: Multi-layer human performance capture using a monocular video camera. IEEE Transactions on Visualization and Computer Graphics, 28(4):1862-1879, 2022. 1, 2" + } + ] + } + ], + "index": 32 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 276, + 747, + 300, + 756 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 276, + 747, + 300, + 756 + ], + "spans": [ + { + "bbox": [ + 276, + 747, + 300, + 756 + ], + "type": "text", + "content": "21106" + } + ] + } + ], + "index": 34 + } + ], + "page_size": [ + 595, + 842 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 106, + 279, + 695 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 47, + 106, + 279, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 106, + 279, + 138 + ], + "spans": [ + { + "bbox": [ + 47, + 106, + 279, + 138 + ], + "type": "text", + "content": "[32] Robert W. Sumner, Johannes Schmid, and Mark Pauly. Embedded deformation for shape manipulation. ACM TOG, 26(3), 2007. 3, 5" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 140, + 279, + 181 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 140, + 279, + 181 + ], + "spans": [ + { + "bbox": [ + 47, + 140, + 279, + 181 + ], + "type": "text", + "content": "[33] Gül Varol, Duygu Ceylan, Bryan Russell, Jimei Yang, Ersin Yumer, Ivan Laptev, and Cordelia Schmid. BodyNet: Volumetric inference of 3D human body shapes. In ECCV, 2018. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 183, + 279, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 183, + 279, + 226 + ], + "spans": [ + { + "bbox": [ + 47, + 183, + 279, + 226 + ], + "type": "text", + "content": "[34] Kangkan Wang, Sida Peng, Xiaowei Zhou, Jian Yang, and Guofeng Zhang. NerfCap: Human performance capture with dynamic neural radiance fields. IEEE Transactions on Visualization and Computer Graphics, 2022. 1, 3, 4, 5, 6, 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 228, + 279, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 228, + 279, + 259 + ], + "spans": [ + { + "bbox": [ + 47, + 228, + 279, + 259 + ], + "type": "text", + "content": "[35] Olivia Wiles, Georgia Gkioxari, Richard Szeliski, and Justin Johnson. Synsin: End-to-end view synthesis from a single image. In CVPR, pages 7467-7477, 2020. 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 261, + 279, + 292 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 261, + 279, + 292 + ], + "spans": [ + { + "bbox": [ + 47, + 261, + 279, + 292 + ], + "type": "text", + "content": "[36] Yuliang Xiu, Jinlong Yang, Dimitrios Tzionas, and Michael J. Black. ICON: Implicit clothed humans obtained from normals. In CVPR, June 2022. 2, 6, 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 293, + 279, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 293, + 279, + 335 + ], + "spans": [ + { + "bbox": [ + 47, + 293, + 279, + 335 + ], + "type": "text", + "content": "[37] Weipeng Xu, Avishek Chatterjee, Michael Zollhöfer, Helge Rhodin, Dushyant Mehta, Hans-Peter Seidel, and Christian Theobalt. MonoPerfCap: Human performance capture from monocular video. ACM TOG, 37(2):1-15, July 2018. 1, 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 337, + 279, + 389 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 337, + 279, + 389 + ], + "spans": [ + { + "bbox": [ + 47, + 337, + 279, + 389 + ], + "type": "text", + "content": "[38] Tao Yu, Kaiwen Guo, Feng Xu, Yuan Dong, Zhaqi Su, Jianhui Zhao, Jianguo Li, Qionghai Dai, and Yebin Liu. Body-Fusion: Real-time capture of human motion and surface geometry using a single depth camera. In ICCV, October 2017. 1, 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 391, + 279, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 391, + 279, + 442 + ], + "spans": [ + { + "bbox": [ + 47, + 391, + 279, + 442 + ], + "type": "text", + "content": "[39] Tao Yu, Zerong Zheng, Kaiwen Guo, Jianhui Zhao, Qionghai Dai, Hao Li, Gerard Pons-Moll, and Yebin Liu. DoubleFusion: Real-time capture of human performances with inner body shapes from a single depth sensor. In CVPR, June 2018. 1, 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 445, + 279, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 445, + 279, + 487 + ], + "spans": [ + { + "bbox": [ + 47, + 445, + 279, + 487 + ], + "type": "text", + "content": "[40] Tao Yu, Zerong Zheng, Yuan Zhong, Jianhui Zhao, Qionghai Dai, Gerard Pons-Moll, and Yebin Liu. SimulCap: Single-view human performance capture with cloth simulation. In CVPR, 2019. 1, 2, 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 489, + 279, + 530 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 489, + 279, + 530 + ], + "spans": [ + { + "bbox": [ + 47, + 489, + 279, + 530 + ], + "type": "text", + "content": "[41] Fuqiang Zhao, Wei Yang, Jiakai Zhang, Pei Lin, Yingliang Zhang, Jingyi Yu, and Lan Xu. HumanNeRF: Generalizable neural human radiance field from sparse inputs. In CVPR, 2022. 1, 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 532, + 279, + 564 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 532, + 279, + 564 + ], + "spans": [ + { + "bbox": [ + 47, + 532, + 279, + 564 + ], + "type": "text", + "content": "[42] Zerong Zheng, Tao Yu, Yixuan Wei, Qionghai Dai, and Yebin Liu. DeepHuman: 3D human reconstruction from a single image. In ICCV, 2019. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 565, + 279, + 618 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 565, + 279, + 618 + ], + "spans": [ + { + "bbox": [ + 47, + 565, + 279, + 618 + ], + "type": "text", + "content": "[43] Heming Zhu, Yu Cao, Hang Jin, Weikai Chen, Dong Du, Zhangye Wang, Shuguang Cui, and Xiaoguang Han. Deep Fashion3D: A dataset and benchmark for 3d garment reconstruction from single images. In ECCV, pages 512-530, 2020. 1, 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 47, + 620, + 279, + 661 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 620, + 279, + 661 + ], + "spans": [ + { + "bbox": [ + 47, + 620, + 279, + 661 + ], + "type": "text", + "content": "[44] Heming Zhu, Lingteng Qiu, Yuda Qiu, and Xiaoguang Han. Registering Explicit to Implicit: Towards high-fidelity garment mesh reconstruction from single images. In CVPR, pages 3845-3854, 2022. 1, 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 47, + 664, + 279, + 695 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 664, + 279, + 695 + ], + "spans": [ + { + "bbox": [ + 47, + 664, + 279, + 695 + ], + "type": "text", + "content": "[45] Hao Zhu, Xinxin Zuo, Sen Wang, Xun Cao, and Ruigang Yang. Detailed human shape estimation from a single image by hierarchical mesh deformation. In CVPR, 2019. 2" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 277, + 746, + 300, + 756 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 277, + 746, + 300, + 756 + ], + "spans": [ + { + "bbox": [ + 277, + 746, + 300, + 756 + ], + "type": "text", + "content": "21107" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 595, + 842 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Cloud-Device Collaborative Adaptation to Continual Changing Environments in the Real-World/071f2817-bbaa-4424-acc0-27135344a577_content_list.json b/2023/Cloud-Device Collaborative Adaptation to Continual Changing Environments in the Real-World/071f2817-bbaa-4424-acc0-27135344a577_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..e4070ce75197cd47861247058ea147934ed2138e --- /dev/null +++ b/2023/Cloud-Device Collaborative Adaptation to Continual Changing Environments in the Real-World/071f2817-bbaa-4424-acc0-27135344a577_content_list.json @@ -0,0 +1,1580 @@ +[ + { + "type": "text", + "text": "Cloud-Device Collaborative Adaptation to Continual Changing Environments in the Real-world", + "text_level": 1, + "bbox": [ + 94, + 128, + 875, + 174 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yulu Gan $^{1*}$ , Mingjie Pan $^{1*}$ , Rongyu Zhang $^{2}$ , Zijian Ling $^{3}$ , Lingran Zhao $^{1}$ , Jiaming Liu $^{1}$ , Shanghang Zhang $^{1\\dagger}$", + "bbox": [ + 258, + 202, + 709, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ National Key Laboratory for Multimedia Information Processing, School of Computer Science, Peking University,", + "bbox": [ + 104, + 239, + 864, + 257 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ The Chinese University of Hong Kong, Shenzhen, $^{3}$ Imperial College London", + "bbox": [ + 225, + 256, + 743, + 273 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/204b99c60bcd114a0b5d4892094edba84b4b5ebf619a0e3955fa852653c7ad60.jpg", + "image_caption": [ + "(a) The problem and our main idea" + ], + "image_footnote": [], + "bbox": [ + 93, + 303, + 460, + 448 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/fe33a44805963d570b10019b9b4ddaace2b9e62a57530e3f15f36c16259e63d9.jpg", + "image_caption": [ + "(b) Comparison of our method with others on continuous domain shifts", + "Figure 1. (a) Models deployed on devices are preferably lightweight. However, device models will suffer from severe performance degradation when facing continual distribution shift data. Our main idea is to improve the continual domain adaptation capability of the device model by performing our proposed Cloud-Device Collaborative Adaptation paradigm. (b) We compare our method with previous works [16,34,35]. Our method surpasses the state-of-the-art approach and exhibits a solid ability when facing continual distribution shifts." + ], + "image_footnote": [], + "bbox": [ + 503, + 316, + 867, + 448 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 560, + 313, + 575 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "When facing changing environments in the real world, the lightweight model on client devices suffers from severe performance drops under distribution shifts. The main limitations of the existing device model lie in (1) unable to update due to the computation limit of the device, (2) the limited generalization ability of the lightweight model. Meanwhile, recent large models have shown strong generalization capability on the cloud while they can not be deployed on client devices due to poor computation constraints. To enable the device model to deal with changing environments, we propose a new learning paradigm of Cloud-Device Collaborative Continual Adaptation, which encourages collaboration between cloud and device and improves the generalization of the device model. Based on this paradigm, we further propose an Uncertainty-based Visual Prompt Adapted (U-VPA) teacher-student model to transfer the generalization capability of the large model on the cloud to the device model. Specifically, we first de", + "bbox": [ + 75, + 592, + 473, + 864 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "sign the Uncertainty Guided Sampling (UGS) to screen out challenging data continuously and transmit the most out-of-distribution samples from the device to the cloud. Then we propose a Visual Prompt Learning Strategy with Uncertainty guided updating (VPLU) to specifically deal with the selected samples with more distribution shifts. We transmit the visual prompts to the device and concatenate them with the incoming data to pull the device testing distribution closer to the cloud training distribution. We conduct extensive experiments on two object detection datasets with continually changing environments. Our proposed U-VPA teacher-student framework outperforms previous state-of-the-art test time adaptation and device-cloud collaboration methods. The code and datasets will be released.", + "bbox": [ + 500, + 561, + 892, + 772 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 500, + 800, + 630, + 814 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Real-world usually contains various environmental changes along with continual distribution shifts [35]. People usually deploy economically lightweight models on devices to boost the scalability and practicability. The lightweight model can suffer severe performance degrad", + "bbox": [ + 496, + 825, + 892, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Equal contribution", + "bbox": [ + 94, + 875, + 202, + 887 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "† Corresponding author", + "bbox": [ + 96, + 887, + 218, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "12157", + "bbox": [ + 478, + 924, + 517, + 936 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "tion under continual distribution shifts [25, 29, 33, 35]. The main challenges are: (1) The poor computational ability of devices. Due to the properties of device infrastructure, the deployed model can not be updated on time, thus lagging its performance for the real world under distribution shift. (2) The limited generalization ability of the lightweight model. Since lightweight models are of relatively small capacity, they can not handle continually changing environments. In contrast, recent large models that are trained on the cloud server show significant generalization ability [2, 39]. While in industry, these large models can not be directly applied due to the limited infrastructure.", + "bbox": [ + 75, + 90, + 472, + 271 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Therefore, we enable the device model to tackle real-world environmental changing by proposing a Cloud-Device Collaborative Continual Adaptation paradigm, as shown in Fig.1 (a). Previous Cloud-Device Collaborative methods only focus on improving model representation on variance in video frames but neglect the model generalization ability for continually changing data distribution. In our new paradigm, we fully exploit the sufficient knowledge of the large cloud model and transfer the continual generalization ability to the device lightweight model.", + "bbox": [ + 75, + 272, + 472, + 422 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In particular, we design an Uncertainty-based Visual Prompt Adapted (U-VPA) teacher-student model, which consists of an Uncertainty Guided Sampling (UGS) strategy and a Visual Prompt Learning Strategy with Uncertainty Guided Updating (VPLU). Due to the communication bandwidth constraint [11, 13, 19], different with some work targeting filtering the parameters [22, 26], we design the UGS to screen out the most environment-specific samples and decrease the required bandwidth compared with transmitting the whole sequence. To leverage the strong generalization ability of large models, we introduce the VPLU to align source-target domain distribution and transfer the representation of the large teacher model to the light student model. The light student model and visual prompts are then delivered to the device, thus alleviating the continuously changing scenarios in the real world.", + "bbox": [ + 75, + 422, + 472, + 664 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Experimental results show that our method outperforms the state-of-the-art methods on synthetic and real-world distribution shift datasets, as shown in Fig.1 (b). Besides, we can achieve the same performance as the entire data with fewer reflowed data (42% of total data). As another benefit, fewer reflowed data reduce the bandwidth pressure of the uplink. As for the downlink, we can deliver the visual prompts (0.43% of the model's parameters) with almost negligible bandwidth to the device and apply the visual prompt to the input data to improve the performance of the device model by 3.9% in mAP.", + "bbox": [ + 75, + 665, + 472, + 830 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our contributions can be summarized as follows:", + "bbox": [ + 96, + 830, + 423, + 845 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "- We make the first attempt to deal with continually changing scenarios by proposing a Cloud-Device Collaborative Continual Adaptation paradigm, which aims", + "bbox": [ + 94, + 854, + 468, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "to transfer the generalization ability from the large cloud model to the lightweight device model. Our method is a general paradigm that can apply to real-world systems.", + "bbox": [ + 529, + 90, + 890, + 151 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "- We design an Uncertainty-based Visual Prompt Adapted (U-VPA) teacher-student model, which consists of UGS and VPLU. We introduce UGS to screen out the most environment-specific samples and decrease the required bandwidth compared with transmitting the whole sequence. We propose a VPLU to align source-target data distribution and transfer the representation of the large teacher model to the lightweight student model.", + "bbox": [ + 517, + 161, + 890, + 297 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "- Experiments show that our proposed framework and method surpass other state-of-the-art methods and can effectively improve the continuous domain adaptation capability of the device model.", + "bbox": [ + 517, + 306, + 890, + 367 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related work", + "text_level": 1, + "bbox": [ + 500, + 392, + 637, + 407 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Test time adaptation. Test-time adaptation(TTA) aims to get better model generalization performance on target data without access to source data. Feature modulation is used in TENT [34] to optimize BN layers of the input pretrained model by entropy minimization. SHOT [20] takes entropy minimization and a diversity regularizer to achieve information maximization. [18, 24] both take a productive way to enhance model performance on the target domain.", + "bbox": [ + 496, + 417, + 890, + 537 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We differ from TTA in that we consider continuously changing domains. Besides, our model is deployed on the device. There will be restrictions on computing resources and uplink and downlink transmission bandwidth. Considering TTA can effectively use the test data to make updates, we leverage TTA to enable the cloud to improve generalization performance continuously.", + "bbox": [ + 496, + 537, + 890, + 643 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Prompt learning. Prompt is originally from the natural language process (NLP) and aims to adapt the large-scale pre-trained language model to downstream tasks. Recently, several methods [1, 9, 21, 40, 41] are proposed to prove the feasibility of using prompts to adapt the large pre-trained model to downstream tasks in computer vision or vision-language tasks. [1, 21, 37] take prompt as additional input containing a small number of learnable parameters and explore the design of prompts.", + "bbox": [ + 496, + 643, + 890, + 779 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Prompt learning aims at mining the pre-trained model's information. We introduce visual prompt tuning to improve the performance of the device model on the current domain, with a negligible computational overhead in our Cloud-Device Collaboration (DCC) framework.", + "bbox": [ + 496, + 780, + 890, + 854 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Cloud-device collaboration. Cloud-device collaboration aims to bring benefits to both the cloud and the device jointly. Several methods [4, 6, 7, 15] try to unload part or", + "bbox": [ + 496, + 854, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "12158", + "bbox": [ + 480, + 924, + 517, + 936 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/30c87bae3bde82773df112a90c27fb43c35c3e88475bd3911825c0580a8d2b34.jpg", + "image_caption": [ + "Figure 2. The whole framework. (a): Testing on the device. On the client device, we select data that need to be sent to cloud by Uncertainty Guided Sampling (UGS). These uncertainty samples are then used in the cloud to update the teacher-student models. The visual prompt trained on the cloud will be delivered to the device through the downlink and applied to the input data $x$ . Then the reformulated input data will be put into the device model for testing. (b) Training on the cloud. We add our proposed visual prompt to the uncertain samples at the pixel level to get the reformulated images. We then put the reformulated images into the teacher-student model. The teacher-student model and visual prompts are optimized by align loss and supervised loss. (c) Visual prompt learning strategy with uncertainty guided sampling (VPLU). We update the visual prompts by uncertainty-based EMA update. The update weight of the visual prompt will be larger when the uncertainty is higher. The student model is updated using the uncertainty samples and the pseudo labels." + ], + "image_footnote": [], + "bbox": [ + 93, + 103, + 883, + 444 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "all computation to the cloud to alleviate the computation deficiency on the device. But they do not leverage the collaboration between the cloud and the device. AMS [16] and DCCL [38] have considered the collaboration but they do not take the real-world continual distribution shifts into account. Compared with methods above, our proposed CloudDevice Collaborative Continual Adaptation paradigm can deal with the real-world environment changes.", + "bbox": [ + 75, + 580, + 468, + 700 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In our work, we use Uncertainty Guided Sampling (UGS) and visual prompt for more efficient bandwidth usage. Besides, our model can adapt to different domains during test time. Our teacher model on the cloud and our student model on the device can be improved simultaneously.", + "bbox": [ + 75, + 700, + 470, + 777 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 76, + 789, + 168, + 806 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Preliminary", + "text_level": 1, + "bbox": [ + 76, + 814, + 207, + 830 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Given a lightweight model $\\theta_{device}$ that is deployed on the device, our goal is to perform well on the continually changing target domains $D_{t1}, D_{t2}, \\ldots, D_{tT}$ with limited communication bandwidth, where $D_{ti} = \\{(x_i^T)\\}_{i=1}^{N_t}$ , and $N_t$ repre", + "bbox": [ + 75, + 837, + 470, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "sents the scale of the target domain. The distributions of the target domains are arbitrary since they can change or reoccur over time. We thus introduce a large teacher model $\\theta_{tea}$ and lightweight student model $\\theta_{stu}$ cloned from $\\theta_{device}$ , to boost the generalization ability of $\\theta_{device}$ .", + "bbox": [ + 496, + 580, + 890, + 657 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Overall framework", + "text_level": 1, + "bbox": [ + 500, + 666, + 684, + 680 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In our Cloud-Device Collaborative Continual Adaptation paradigm, the key idea is to improve the device model's performance continuously by receiving adapted parameters from the cloud server. So we propose Visual Prompt Adapted (U-VPA) teacher-student model. As shown in Fig. 2, the U-VPA teacher-student model contains Uncertainty Guided Sampling (UGS) and Visual Prompt Learning Strategy with Uncertainty Guided Updating (VPLU).", + "bbox": [ + 496, + 688, + 890, + 809 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "When perceiving scene changing or distribution shift, poor computational device intends to get help from a cloud server with powerful computing ability. However, due to the communication bandwidth constraint, we must consider the required bandwidth when concurring the distribution shifts. Therefore, we design a Uncertainty Guided Sam", + "bbox": [ + 496, + 810, + 892, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "12159", + "bbox": [ + 480, + 925, + 517, + 936 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "pling (UGS) to screen out the most environment-specific samples and decrease the required bandwidth compared with transmitting the whole sequence. After receiving the learned parameters from the cloud to deal with these samples of relatively large distribution shifts, it can adapt to the new scene when inference.", + "bbox": [ + 75, + 90, + 468, + 180 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "On the cloud server, due to its strengths in computing, we intend to leverage the strong generalization ability of large models to tackle the received environment-specific and challenging samples. We introduce Visual Prompt Learning Strategy with Uncertainty guided Updating (VPLU) to align source-target domain distribution and transfer the representation of the large teacher model to the light student model. The light student model and visual prompts is then delivered to the device, thus alleviating the continuously changing scenarios in the real-world.", + "bbox": [ + 75, + 181, + 468, + 333 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We will discuss the details of UGS and VPLU in Sec. 3.3.1 and Sec. 3.3.2.", + "bbox": [ + 76, + 334, + 467, + 363 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. U-VPA teacher-student model", + "text_level": 1, + "bbox": [ + 76, + 377, + 341, + 391 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3.1 Uncertainty Guided Sampling(UGS)", + "text_level": 1, + "bbox": [ + 76, + 401, + 387, + 416 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Due to bandwidth limitations, it is not practical for the device to return all data. Therefore, we select images that need to be transmitted based on uncertainty estimation. Inspired by the Dropout methods [8, 12], since we are doing an object detection task, we can obtain $n$ probabilities $p_i(y|x)$ of the model by dropout method, and we calculate the variance of these $n$ prediction $p_i(y|x)$ as uncertainty.", + "bbox": [ + 75, + 426, + 468, + 532 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nV _ {u n c} = \\left(\\frac {1}{n} \\sum_ {i} ^ {n} \\left(p _ {i} (y | x) - \\mu\\right) ^ {2}\\right) ^ {\\frac {1}{2}} \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 160, + 547, + 468, + 584 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $p(y|x)$ represents the predictions of the input image $x$ and $\\mu$ is the mean. $V_{unc}$ is the variance of $n$ probabilities of the model, which represents the uncertainty of the model for the data. The device model filters the input images $x$ whose uncertainty is larger than a threshold and sends back to the cloud server.", + "bbox": [ + 75, + 595, + 468, + 684 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3.2 Visual Prompt Learning Strategy with Uncertainty Guided Updating (VPLU).", + "text_level": 1, + "bbox": [ + 76, + 709, + 467, + 739 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We optimize model on the cloud server by using the data transmitted from the device. It then delivers the learned parameters and processed visual prompt to the device model for inference. We thus introduce VPLU with a large teacher model and light student model, to mitigate variant domain distribution and improve the generalization ability of light model. Note that, teacher model is initialized by the source model pre-trained on source domain. It takes source and target domain as input, while student model only trains on the target domain.", + "bbox": [ + 75, + 750, + 468, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Cross-domain alignment. In order to mitigate distribution shift, we align the features in the source and target domains. Inspired by [10], we apply feature alignment on teacher model to improve its performance on target domain. We use a domain discriminator to classify data in feature space and try to distinguish which domain the data comes from. Besides, we use a gradient reversal layer to achieve a uniform training process. Source domain data is used to supervise teacher model, avoiding performance degradation.", + "bbox": [ + 496, + 90, + 890, + 226 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Cross-model transfer. After improving teacher model's domain-invariant representation ability, we transfer its ability against distribution shifts and generalization to student model via visual prompts and knowledge transferring.", + "bbox": [ + 496, + 227, + 890, + 286 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Visual Prompt Adaptation(VPA) learning. Specifically, inspired by the prompt learning in NLP [23], which can effectively explores the knowledge of source domain, we propose visual prompts as hints to continually transfer teacher model's generalized representation to the student model.", + "bbox": [ + 496, + 287, + 890, + 362 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In particular, visual prompts $\\phi$ are learnable parameters. During the forward process of the model, we can obtain the reformulated image $x^{*}$ by adding visual prompts on the input images $x$ at the pixel level.", + "bbox": [ + 496, + 363, + 890, + 422 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nx ^ {*} = x + \\phi , \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 651, + 438, + 890, + 453 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In our framework, whether in training on server or inference on device, we all use the reformulated image $x^{*}$ as model's input, instead of origin images $x$ .", + "bbox": [ + 496, + 458, + 890, + 503 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Uncertainty-aware Prompt Updation(U-EMA). Visual prompt acts on the input data and participates in the forward process of both teacher and student models, so it can also benefit from their optimization during training. Moreover, to make the prompt update more stable and consider the current model's uncertainty to data, we update the visual prompt using by uncertainty-based exponential moving average(U-EMA). The weight of the exponential moving average is based on the uncertainty estimation as follows:", + "bbox": [ + 496, + 503, + 890, + 638 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\phi_ {t} \\leftarrow \\beta \\phi_ {t - 1} + (1 - \\beta) \\phi_ {t}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 604, + 654, + 890, + 670 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Note that $\\phi_t$ represents the parameters of the visual prompt that need to be updated, and $\\phi_{t - 1}$ indicates the visual prompt parameters of the last training step t. We follow [30] to set $\\alpha = 0.999$ . And $\\beta = \\alpha -V_{unc}$ , where $V_{unc}$ is defined in Eq. (1) as the uncertainty of current data. Higher uncertainty means a larger distribution shift in the current data. To save more domain-related information, the prompt update weight will be larger at this time. That is, the larger the uncertainty, the greater the weight of the prompt update.", + "bbox": [ + 496, + 674, + 890, + 810 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As another essential role of the visual prompt, the processed visual prompt will be sent to the device to improve the adaptability of the device model by applying the visual prompt on input images.", + "bbox": [ + 496, + 810, + 890, + 869 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Knowlegde transfer. Along with visual prompt adapted tuning, we align teacher student feature and adopt teacher", + "bbox": [ + 500, + 869, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "12160", + "bbox": [ + 480, + 924, + 517, + 935 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "model to generate pseudo label for student model. It is penalized by Knowledge transfer loss $\\mathcal{L}_k$ , to promote the consistency between teacher student model representation. Specifically, The teacher model will generate predictions for the reformulated images $x^{*}$ as pseudo labels. Teacher model has stronger generalization ability, so the generated pseudo labels are of high quality. To further improve the pseudo labels quality, we screen those predictions with confidence higher than the threshold as the input to the student model. Overall, we transfer the knowledge from the teacher to the student by feeding selected images $x^{*}$ and corresponding pseudo labels into the student model for training.", + "bbox": [ + 75, + 90, + 472, + 273 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Training objectives", + "text_level": 1, + "bbox": [ + 76, + 280, + 261, + 296 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The teacher and student models are jointly optimized on the cloud server, using uncertain samples $x_{t}^{i}\\in D_{t}$ returned from device. The source model is trained on source pairs of $\\{x_s^i,y_s^i\\}_{i = 1}^{n_s}\\in D_s$ where $x_{s}^{i}\\in X_{s}$ and $y_{s}^{i}\\in Y_{s}$ are images and labels, respectively.", + "bbox": [ + 76, + 303, + 468, + 378 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For the training process of the teacher model, given the reformulated taregt images $x_{t}^{*}$ , as the Eq. (4) shown, we use adversal alignment loss [31] and supervision loss to optimize the teacher model (including visual prompt $\\phi$ ).", + "bbox": [ + 76, + 378, + 468, + 439 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} L _ {t e a} = - \\lambda L _ {d} \\left(G _ {d} \\left(F \\left(x _ {s} ^ {*}; \\theta_ {t e a}\\right); F \\left(x _ {t} ^ {*}; \\theta_ {t e a}\\right)\\right), y _ {d}\\right) \\tag {4} \\\\ + L _ {D e t} \\left(X \\left(x _ {s} ^ {*}, \\theta_ {t e a}\\right), y _ {s}\\right), \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 99, + 458, + 468, + 494 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Note that the first part is adversal alignment loss [31], and another one is the loss function of Faster RCNN [27] (using labeled data in source domain $D_{s}$ ). $X$ represents the model which predict detection outputs. $L_{d}$ represent domain classifier. $F$ indicates the feature extractor and $G_{d}$ is the domain classifier. $\\lambda$ is the weight of alignment loss.", + "bbox": [ + 76, + 501, + 468, + 590 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For the optimization of the student model, we use pseudo labels $y_{t}^{p}$ generated by teacher model as supervision:", + "bbox": [ + 76, + 592, + 468, + 622 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nL _ {s t u} = L _ {D e t} \\left(X \\left(x _ {t} ^ {*}, \\theta_ {s t u}\\right), y _ {t} ^ {p}\\right), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 166, + 636, + 468, + 652 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The visual prompt benefited from both $L_{tea}$ and $L_{stu}$ . And we update the visual prompt by uncertainty-aware exponential moving average as Eq. (3) shown.", + "bbox": [ + 76, + 657, + 468, + 703 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 76, + 714, + 209, + 732 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In Sec.4.1, we give the details of the dataset and experiment setting. We evaluate the performance of CCA on continual changing scenarios in Sec.4.2. We give discussion on our motivations and ablation study in Sec.4.3, Sec.4.4.", + "bbox": [ + 76, + 739, + 468, + 800 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Experimental setup", + "text_level": 1, + "bbox": [ + 76, + 808, + 264, + 824 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1.1 Datasets and settings", + "text_level": 1, + "bbox": [ + 76, + 832, + 276, + 847 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We evaluate our proposed method on two test time adaptation benchmark tasks for object detection: Cityscapes-to-Cityscape-C and Cityscapes-to-ACDC-Detection dataset.", + "bbox": [ + 76, + 854, + 468, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Cityscape-C is originally created to benchmark robustness tasks [14]. We select the five most relevant corruptions to the autonomous driving scenario (including brightness, motion_blur, rain, fog and snow). Each corruption dataset contains five levels of severity. The corruptions are applied to images from the validation set of the clean Cityscapes dataset [5]. There are 2,500 images in five corruption datasets.", + "bbox": [ + 496, + 90, + 890, + 210 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For Cityscapes-to-Cityscapes-C, we follow [3] to train the pre-trained model on the source domain (Cityscapes). Specifically, we use the SGD optimizer with a learning rate of 1e-2 as the official implementation to train the source model.", + "bbox": [ + 496, + 212, + 890, + 286 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "ACDC-Detection is created from the Adverse Conditions Dataset (ACDC) [28]. The ACDC dataset shares the same classes with Cityscapes and is collected in various adverse visual conditions: Fog, Night, Rain and Snow. We convert the segmented labels into labels for object detection. We follow CoTTA [35] to set our task. We use 400 unlabeled images from each adverse condition for the adaptation. To mimic the scenario in real life where a similar environment might be revisited and to evaluate the continuous generalization capability of our methods, we repeat the same sequence group (of the three conditions) 10 times (i.e., in total 30: $\\mathrm{Fog} \\to \\mathrm{Rain} \\to \\mathrm{Snow} \\to \\mathrm{Fog}\\ldots$ ). This also provides an evaluation of the adaptation performance in the long term.", + "bbox": [ + 496, + 289, + 892, + 500 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For the Cityscapes-to-ACDC-Detection task, we use the same way to train the pre-trained model on the source domain (Cityscapes) as the Cityscapes-to-Cityscapes-C task.", + "bbox": [ + 496, + 501, + 890, + 547 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1.2 Implementation Details", + "text_level": 1, + "bbox": [ + 500, + 571, + 718, + 585 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this paper, all experiments are conducted with PyTorch. For the Cityscapes-to-Cityscapes-C task and Cityscapes-to-ACDC-Detection task, the batch size is set to 8. We follow [3] to train the pre-trained model on the source domain (Cityscapes). The number of samples used for uncertainty estimation is set to 10 in the experiment.", + "bbox": [ + 496, + 595, + 890, + 686 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We use FasterRCNN with the backbone of ResNet-101 as the large teacher model on the cloud. And we use FasterRCNN with the backbone of ResNet-18 as the lightweight student model (same as the device model). The above large and light models are initialized with corresponding pretrained models on the source domain (Cityscapes dataset). For the UGS module, we set the uncertainty threshold to 0.008, which filters out $42\\%$ samples in the Cityscape-C dataset. For the VPA module, We set the visual prompts with size $3 \\times 200 \\times 200$ , which can be added to the left-up space of the image. The student model's parameters and visual prompts (about $0.4\\%$ of the student model's parameters) will be transmitted to the device.", + "bbox": [ + 496, + 688, + 890, + 883 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We use $\\mathrm{mAP} @ 0.5(\\%)$ as the evaluation metric. The", + "bbox": [ + 519, + 885, + 890, + 898 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "12161", + "bbox": [ + 480, + 926, + 517, + 936 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/63c7d5a1c314a8231196355787e421c5aed56988cc3a9351ea48aec844f4caa6.jpg", + "table_caption": [ + "Table 1. Continual generalization capability on Cityscapes-to-Cityscapes-C. Object detection results (mAP@0.5 in %) on the Cityscapes-to-Cityscapes-C online continual test-time adaptation task. Gain(%) means the improvement of our method compared with Source-only. We evaluate the four test conditions continually for ten times to evaluate the long-term adaptation performance. All results are evaluated on the FasterRCNN architecture with the largest corruption severity level 5. Our approach surpasses the SOTA method and exhibits significant continual generalization and anti-forgetting abilities." + ], + "table_footnote": [], + "table_body": "
Timet
Round1510All
ConditionFogMotionRainSnowBrightnessFogMotionRainSnowBrightnessFogMotionRainSnowBrightnessMeanGain
Source-only [27]24.79.522.41.329.024.79.522.41.329.024.79.522.41.329.017.4/
TENT-continual [34]24.89.222.51.429.420.06.419.80.823.79.03.810.70.511.413.0-4.4
CoTTA [35]25.110.122.41.328.921.89.520.74.228.29.33.910.13.317.514.7-2.7
Pseudo-Label [17]27.317.224.76.933.730.120.126.98.036.531.020.927.77.238.224.1+6.7
AMS [16]28.417.726.07.634.731.820.428.27.738.034.221.428.97.039.425.1+7.7
Ours(proposed)28.017.826.810.036.537.524.235.514.244.044.027.340.915.150.131.0+13.6
", + "bbox": [ + 76, + 166, + 890, + 277 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/6e9418e45faf3427333cb1117d518f576bf87e9f1b9a172343e277f78a655b5c.jpg", + "table_caption": [ + "Table 2. Continuous generalization capability on Cityscapes-to-ACDC-Detection. Object detection results (mAP@0.5 in %) on the Cityscapes-to-ACDC-Detection online continual test-time adaptation task. Gain(%) means the improvement of our method compared with Source-only. We evaluate the four test conditions continually for ten times to evaluate the long-term adaptation performance. All results are evaluated on the FasterRCNN architecture. Our results outperform the SOTA method, thus verifying the effectiveness of our method on the real distribution shift." + ], + "table_footnote": [], + "table_body": "
Timet
Round14710All
ConditionFogNightRainSnowFogNightRainSnowFogNightRainSnowFogNightRainSnowMeanGain
Source-only [27]20.57.414.314.220.57.414.314.220.57.414.314.220.57.414.314.214.1/
TENT-continual [34]20.57.514.313.920.07.213.413.018.66.612.312.116.66.011.110.912.8-2.3
CoTTA [35]20.77.514.714.320.97.414.714.420.47.313.814.416.36.411.712.213.8-0.3
Pseudo-Label [17]20.98.715.517.623.410.316.918.924.310.917.819.525.111.517.520.017.6+3.5
AMS [16]21.69.016.017.223.210.717.219.124.711.417.219.926.111.717.820.517.8+3.7
Ours(proposed)21.19.315.717.224.711.817.820.226.312.018.520.626.112.318.020.818.5+4.4
", + "bbox": [ + 76, + 358, + 890, + 483 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "overall process of test-time adaptation strictly follows CoTTA [35]. The process first tests the coming batch of data then train with it. We do batch-level updates on the cloud model, and the performance on the device model is taken as the average of the results of all the data.", + "bbox": [ + 75, + 498, + 470, + 575 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Baselines. Our method is compared with multiple baselines, including those for test time adaptation (TTA), continual test time adaptation (CTTA) and cloud-end collaboration methods as follows: Source-only [27]. Test the model performance on the different target domains using the pre-trained source model (training on the source domain). Pseudo-label [17]. Use pseudo label generated by the teacher model to supervise the student model. Tent-continual [34]. Tent is an effective method for the test-time adaptation (TTA). Tent fine-tune the batch normalization layer during training to adapt the model to the target domain. CoTTA [35]. CoTTA is an attractive approach for the continual test-time adaptation (CTTA). CoTTA mitigate error accumulation and catastrophic forgetting with weighted averaging, data augmented averaging, and random recovery neurons. AMS [16]. AMS is an cloud engagement approach. To solve the resource limitation of the device model, AMS do distillation in the cloud, and then transmit the model's parameters to the device in the cloud, the teacher network will not update.", + "bbox": [ + 75, + 598, + 470, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Results and Analysis", + "text_level": 1, + "bbox": [ + 500, + 497, + 694, + 513 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Synthetic continual distribution shift. We evaluate our proposed method on the synthetic continual distribution shift dataset (Cityscapes $\\rightarrow$ Cityscapes-C task). As shown in Table 1, TENT-continual [34] has a decreasing performance when facing continuously changing domains. Because TENT does not consider continuous distribution shifts, it uses self-entropy to update the bn layer, which is prone to catastrophic forgetting when facing constant distribution shifts. CoTTA [35] is a method designed specifically for continuous test time adaptation. CoTTA applies weighted averaging, data-augmented averaging, and random recovery neurons to overcome catastrophic forgetting and error accumulation. However, when doing the object detection task on the continuously changing domains, the model's performance still decreases, indicating that solutions used by CoTTA are insufficient against catastrophic forgetting. We believe this is caused by the poor generalization abilities of the lightweight device model. Therefore, it is reasonable to introduce the cloud for the Cloud-Device collaborative continual adaptation paradigm.", + "bbox": [ + 496, + 521, + 890, + 824 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Both Pseudo-Label and AMS [16] use the Cloud-Device collaboration framework, which transfers the teacher's knowledge to the student on the cloud. The generalization of the large model provides continuous guidance for the student model, which avoids the problem of catastrophic for", + "bbox": [ + 496, + 824, + 893, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "12162", + "bbox": [ + 478, + 925, + 519, + 936 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/ee175a1cf1b4809dbd6315810cfed251a4429d8ad11b0bc8c8365c6fc6e9a4b0.jpg", + "image_caption": [ + "(a)" + ], + "image_footnote": [], + "bbox": [ + 99, + 98, + 269, + 208 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/921a84d4117dc510009e17c94cf4f4d6258c468eace56ec72089c3acc77942d2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 274, + 98, + 444, + 207 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/fd4b8e6533c44311937a2084116cea2f2d5006fd96df6bfd797526dabe8e023c.jpg", + "image_caption": [ + "Figure 3. (a)-(b): Statistical distribution of confidence and uncertainty. (a): distribution of confidence. (b): distribution of uncertainty. Comparing (a) and (b), the uncertainty value is more stable when facing continuous distribution shift, while the confidence value fluctuates significantly with continuous distribution shift. (c)-(e): T-SNE [32] visualization of samples in different domains. (c): all samples. (d): samples filtered by confidence. (e): samples filtered by UGS. Samples filtered by UGS are more environment-specific. They are more worthy of being transmitted to the cloud server for optimizing the teacher-student models." + ], + "image_footnote": [], + "bbox": [ + 98, + 222, + 444, + 321 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "getting and error accumulation. (Pseudo-Label and AMS are $6.7\\%$ and $7.7\\%$ higher than Source-only).", + "bbox": [ + 75, + 507, + 468, + 539 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "However, pseudo label and AMS do not consider updating the teacher model. Our U-VPA framework jointly optimizes the teacher and student model, which makes our method surpass Source-only by $13.6\\%$ , outperforming all the above methods. This validates that our Cloud-Device collaboration paradigm can effectively improve the continuous domain adaptation ability of the device.", + "bbox": [ + 75, + 545, + 468, + 651 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Real-world continual distribution shift. We also evaluate our proposed method on the real-world continual distribution shift dataset (Cityscapes $\\rightarrow$ ACDC-Detection task). As shown in Table 2, our method improves by $4.4\\%$ over the source-only method. By comparing the performance change of the same domain in 10 rounds, we can find that our method does not have the problem of catastrophic forgetting. For example, the TENT-continual method faces the problem of catastrophic forgetting. The $\\mathrm{mAP}(\\%)$ drops in the fourth, seventh, and tenth rounds compared to the first round when facing the night domain, specifically $-0.2\\%$ , $-0.5\\%$ , $-4.1\\%$ . In contrast, our method leads to a continuous performance improvement (e.g., our method's $\\mathrm{mAP}(\\%)$ increases in the fourth, seventh, and tenth rounds, respectively, compared to the first round when facing the Night domain, specifically, $+2.5\\%$ , $+0.2\\%$ , $+0.3\\%$ ).", + "bbox": [ + 75, + 659, + 470, + 900 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/136e8a9008aa01371940180382e53f496479df865cae23273717523acae80955.jpg", + "image_caption": [ + "Figure 4. Detection results (mAP@0.5) of both teacher and student model. Solid lines are results of models using VPA under the brightness domain. Teacher and student models have jointly improved in continuous domain adaptation. Dotted lines are results without using VPA. The teacher quickly reaches the bottleneck, and student's improvement is slow. The analysis above illustrates that the proposed U-VPA framework enables the teacher-student model jointly improve the continuous domain adaption capability." + ], + "image_footnote": [], + "bbox": [ + 547, + 104, + 823, + 250 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Discussion", + "text_level": 1, + "bbox": [ + 500, + 390, + 617, + 404 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Teacher-student joint optimization with VPA. Previous approaches (e.g., AMS) focus on improving the student model. However, the student model's performance is correlated with the teacher model. Our method enables the teacher and student models to improve performance together by using the visual prompt as a bridge and using the temporal domain information brought by the visual prompt.", + "bbox": [ + 496, + 414, + 890, + 520 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The optimization of the teacher model comes from two modules: (1) feature alignment, (2) visual prompt adaptation(VPA). Our experiments have proved that although both of the above can bring about gains, visual prompt adaptation is the main reason for the success of our framework.", + "bbox": [ + 496, + 520, + 890, + 595 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "As shown in Fig. 4, if only feature alignment is used, the performance of the teacher model reaches the bottleneck after the second round, and the student's optimization will be very slow (dotted lines in the figure).", + "bbox": [ + 496, + 595, + 890, + 656 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "While our method benefits from VPA, the teacher and student models are jointly improved, demonstrating that our proposed U-VPA teacher-student model can optimize the student-teacher model together.", + "bbox": [ + 496, + 657, + 890, + 717 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Uncertainty Guided Sampling (UGS) strategy. We demonstrate in Table 5 that using only a small amount of reflowed data can outperform the results using total data by the UGS strategy. Our model's performance using the UGS strategy to select part of the data to transmit into the cloud for training is $31\\%$ . Compared to the model performance of $28.3\\%$ obtained with the total data, there is a $2.7\\%$ improvement. The amount of data using the random picking strategy is the same as that with the UGS strategy. And the UGS strategy has a better performance than random picking. The above shows the effectiveness of the UGS strategy.", + "bbox": [ + 496, + 718, + 890, + 883 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We further reveal the advantages of uncertainty when the", + "bbox": [ + 519, + 885, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "12163", + "bbox": [ + 480, + 925, + 517, + 936 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/2f4ca34effd08c4dee527535a2dad924eb9ddf51240dd71cfd80ab131504f21a.jpg", + "table_caption": [ + "Table 3. Ablation study. We conduct ablation studies for proposed Pseudo-Label, VPA, UGS, and U-EMA. All the experiments are done on Cityscapes-to-Cityscapes-C, and the evaluation metric is mAP (\\%). The experimental result shows that all these proposed modules positively impact the performance improvement of the devices model when facing continuous domains." + ], + "table_footnote": [], + "table_body": "
Pseudo-LabelVPAUGSU-EMAFogMotionRainSnowBrightnessMeanGain
124.79.522.41.329.017.4/
2✓31.020.927.77.238.224.1+6.7
3✓✓37.122.232.711.243.127.4+10.0
4✓✓✓38.324.836.512.346.728.3+10.9
5✓✓✓✓44.027.340.915.150.131.0+13.6
", + "bbox": [ + 111, + 133, + 856, + 227 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/69d695cb754c6422ff3f9bb71fbd1b89dcced81df12e929cd337f9cfe91a6bfc.jpg", + "table_caption": [ + "Table 4. Comparing the difference between selecting samples with confidence and uncertainty. Experiments show that better performance is obtained by choosing the reflowed data model through uncertainty estimation." + ], + "table_footnote": [], + "table_body": "
FogMotionRainSnowBrightnessMean
Confidence39.624.937.214.345.127.4
Uncertainty44.027.340.915.150.131.0
", + "bbox": [ + 81, + 294, + 460, + 356 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/81d83f6566881418cf173a69e24758d7d43dd9fe9e0e6ccc666ae3b48c0963cc.jpg", + "table_caption": [ + "Table 5. Effect of uplink transmission of full and partial data on model's performance. Our method returns partial data ( $\\tilde{42\\%}$ ) for training but achieves better performance than using full data. This indicates that our UGS strategy reduces the data transmission communication cost and improves the model performance." + ], + "table_footnote": [], + "table_body": "
25%50%75%100%
Random23.526.127.928.3
UGS28.231.031.2/
Gain+4.7+4.9+3.3/
", + "bbox": [ + 81, + 438, + 460, + 503 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "distribution shifts exist. We find uncertainty is more reliable compared with the confidence score. Regarding Fig.3 (a)-(b), in the presence of distribution shifts, confidence distribution varies significantly, while uncertainty distribution is balanced. Therefore, the samples selected using confidence will face the problem of long-tail distribution, which is not conducive to cloud model training [36]. Uncertainty can more accurately reflect the difficulty of identifying samples by device without using the training of cloud model and discover the samples that are more worthy of attention.", + "bbox": [ + 75, + 515, + 468, + 666 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Besides, we explore the properties of the data filtered by different filter strategies. In Fig.3 (c)-(e), each point represents a sample filtered by the corresponding strategy. We find that samples selected by uncertainty are environmental-specific, so using UGS can filter out more valuable data.", + "bbox": [ + 75, + 667, + 468, + 744 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4. Ablation study", + "text_level": 1, + "bbox": [ + 76, + 753, + 227, + 771 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effect of the Uncertainty Guided Sampling (UGS). Our proposed UGS strategy not only reduces the communication cost of the uplink from the device to the cloud but also improves performance by $0.9\\%$ , as shown in Table 3. Furthermore, as shown in Table 5, the model using UGS outperforms baseline with only $42\\%$ data for train, which significantly improves the compression ratio of the video.", + "bbox": [ + 75, + 777, + 467, + 883 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effect of the Visual Prompt Adapt (VPA). Our pro", + "bbox": [ + 96, + 885, + 467, + 901 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "posed VPA teacher-student model uses visual prompts to make the student-teacher model jointly improve performance. As shown in Table 3, compared to the pure pseudo label method, the usage of VPA can increase performance by $3.3\\%$ . Furthermore, we discussed in Sec. 4.3 that VPA could significantly increase the upper limit of the teacher-student framework from challenging continuously changing distribution shift scenarios. The combination of VPA and other modules can further improve performance steadily.", + "bbox": [ + 496, + 236, + 890, + 372 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effect of the Uncertainty-aware prompt EMA (U-EMA) update mechanism. The U-EMA update mechanism we proposed in training the cloud model makes the cloud model pay more attention to the images worthy of attention by dynamically adjusting the prompt weights with uncertainty. As shown in Table 3, U-EMA can bring a $2.7\\%$ performance improvement.", + "bbox": [ + 496, + 373, + 890, + 481 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion & Limitations", + "text_level": 1, + "bbox": [ + 500, + 500, + 741, + 516 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We propose a Cloud-Device Collaborative Continual Adaptation paradigm to deal with continual changing environments on devices. We design an Uncertainty Guided Sampling (UGS) strategy to transmit the most out-of-distribution samples from device to cloud. Besides, we design a Visual Prompt Learning Strategy with Uncertainty Guided Updating (VPLU) to transfer the generalization capability of large model on the cloud to the device model. Experiments of our method on a wide range of continually changing environments show an improvement of $4.4 - 13.6\\%$ in mAP@0.5 on the object detection task. Our proposed method offers flexibility in its application to most existing cloud-device systems. However, it is worth noting that our experiments were conducted using simulated datasets rather than real systems. Moving forward, an interesting direction for future research would be to test our method on a real system with specific bandwidth limitations, a specific cloud, and a specific device. By taking into account the hardware characteristics of the system, there should have rooms to improve its performance further.", + "bbox": [ + 496, + 527, + 890, + 828 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgement. This research was partly supported by the foundation of the National Key R&D Program of China (2022ZD0116305). We thank Xiaoqi Li for the insightful discussions and for helping with writing.", + "bbox": [ + 496, + 829, + 890, + 901 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "12164", + "bbox": [ + 480, + 925, + 519, + 936 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Hyojin Bahng, Ali Jahanian, Swami Sankaranarayanan, and Phillip Isola. Exploring visual prompts for adapting large-scale models. 2022. 2", + "[2] Rishi Bommasani, Drew A. Hudson, Ehsan Adeli, Russ B. Altman, Simran Arora, Sydney von Arx, Michael S. Bernstein, Jeannette Bohg, Antoine Bosselut, Emma Brunskill, Erik Brynjolfsson, Shyamal Buch, Dallas Card, Rodrigo Castellon, Niladri S. Chatterji, Annie Chen, Kathleen Creel, Jared Davis, Dora Demszky, Chris Donahue, Moussa Doumbouya, Esin Durmus, Stefano Ermon, John Etchemendy, Kawin Ethayarajh, Li Fei-Fei, Chelsea Finn, Trevor Gale, Lauren Gillespie, Karan Goel, Noah D. Goodman, Shelby Grossman, Neel Guha, Tatsunori Hashimoto, Peter Henderson, John Hewitt, Daniel E. Ho, Jenny Hong, Kyle Hsu, Jing Huang, Thomas Icard, Sahil Jain, Dan Jurafsky, Pratyusha Kalluri, Siddharth Karamcheti, Geoff Keeling, Fereshte Khani, Omar Khattab, Pang Wei Koh, Mark Krass, Ranjay Krishna, Rohith Kuditipudi, Ananya Kumar, Faisal Ladhak, Mina Lee, Tony Lee, Jure Leskovec, Isabelle Levent, Xiang Lisa Li, Xuechen Li, Tengyu Ma, Ali Ahmad Malik, Christopher D. Manning, Suvir Mirchandani, Eric Mitchell, Zanele Munyikwa, Suraj Nair, Avanika Narayan, Deepak Narayanan, Ben Newman, Allen Nie, Juan Carlos Niebles, Hamed Nilforoshan, Julian Nyarko, Giray Ogut, Laurel Orr, Isabel Papadimitriou, Joon Sung Park, Chris Piech, Eva Portelance, Christopher Potts, Aditi Raghunathan, Rob Reich, Hongyu Ren, Frieda Rong, Yusuf H. Roohani, Camilo Ruiz, Jack Ryan, Christopher Re, Dorsa Sadigh, Shiori Sagawa, Keshav Santhanam, Andy Shih, Krishnan Srinivasan, Alex Tamkin, Rohan Taori, Armin W. Thomas, Florian Tramèr, Rose E. Wang, and William Yang Wang. On the opportunities and risks of foundation models. arXiv: Learning, 2021. 2", + "[3] Kai Chen, Jiaqi Wang, Jiangmiao Pang, Yuhang Cao, Yu Xiong, Xiaoxiao Li, Shuyang Sun, Wansen Feng, Ziwei Liu, Jiarui Xu, Zheng Zhang, Dazhi Cheng, Chenchen Zhu, Tianheng Cheng, Qijie Zhao, Buyu Li, Xin Lu, Rui Zhu, Yue Wu, Jifeng Dai, Jingdong Wang, Jianping Shi, Wanli Ouyang, Chen Change Loy, and Dahua Lin. MMDetection: Open mmlab detection toolbox and benchmark. arXiv preprint arXiv:1906.07155, 2019. 5", + "[4] Sandeep Chinchali, Apoorva Sharma, James Harrison, Amine Elhafsi, Daniel Kang, Evgenya Pergament, Eyal Cidon, Sachin Katti, and Marco Pavone. Network offloading policies for cloud robotics: a learning-based approach. Autonomous Robots, 2019. 2", + "[5] Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The cityscapes dataset for semantic urban scene understanding. In Proc. of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 5", + "[6] Daniel Crankshaw, Xin Wang, Giulio Zhou, Michael J. Franklin, Joseph E. Gonzalez, and Ion Stoica. Clipper: A low-latency online prediction serving system. networked systems design and implementation, 2016. 2" + ], + "bbox": [ + 86, + 114, + 470, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[7] Sadjad Fouladi, John Emmons, Emre Orbay, Catherine Wu, Riad S. Wahby, and Keith Winstein. Salsify: Low-latency network video through tighter integration between a video codec and a transport protocol. networked systems design and implementation, 2018. 2", + "[8] Yarin Gal and Zoubin Ghahramani. Dropout as a bayesian approximation: Representing model uncertainty in deep learning. international conference on machine learning, 2015. 4", + "[9] Yulu Gan, Yan Bai, Yihang Lou, Xianzheng Ma, Renrui Zhang, Nian Shi, and Lin Luo. Decorate the newcomers: Visual domain prompt for continual test time adaptation, 2023. 2", + "[10] Yaroslav Ganin and Victor Lempitsky. Unsupervised domain adaptation by backpropagation. international conference on machine learning, 2015. 4", + "[11] Amir Ghodrati, Babak Ehteshami Bejnordi, and Amirhossein Habibian. Frameexit: Conditional early exiting for efficient video recognition. arXiv: Computer Vision and Pattern Recognition, 2021. 2", + "[12] Chuan Guo, Geoff Pleiss, Yu Sun, and Kilian Q. Weinberger. On calibration of modern neural networks. international conference on machine learning, 2017. 4", + "[13] Amirhossein Habibian, Davide Abati, Taco S. Cohen, and Babak Ehteshami Bejnordi. Skip-convolutions for efficient video processing. 2021. 2", + "[14] Dan Hendrycks and Thomas Dietterich. Benchmarking neural network robustness to common corruptions and perturbations. Proceedings of the International Conference on Learning Representations, 2019. 5", + "[15] Yiping Kang, Johann Hauswald, Cao Gao, Austin Rovinski, Trevor Mudge, Jason Mars, and Lingjia Tang. Neurosurgeon: Collaborative intelligence between the cloud and mobile edge. architectural support for programming languages and operating systems, 2017. 2", + "[16] Mehrdad Khani, Pouya Hamadanian, Arash Nasr-Esfahany, and Mohammad Alizadeh. Real-time video inference on edge devices via adaptive model streaming. arXiv: Learning, 2020. 1, 3, 6", + "[17] Dong-Hyun Lee. Pseudo-label: The simple and efficient semi-supervised learning method for deep neural networks. 2022. 6", + "[18] Rui Li, Qianfen Jiao, Wenming Cao, Hau-San Wong, and Si Wu. Model adaptation: Unsupervised domain adaptation without source data. computer vision and pattern recognition, 2020. 2", + "[19] Xiaoqi Li, Jiaming Liu, Shizun Wang, Cheng Lyu, Ming Lu, Yurong Chen, Anbang Yao, Yandong Guo, and Shanghang Zhang. Efficient meta-tuning for content-aware neural video delivery, 2022. 2", + "[20] Jian Liang, Dapeng Hu, and Jiashi Feng. Do we really need to access the source data? source hypothesis transfer for unsupervised domain adaptation. international conference on machine learning, 2020. 2", + "[21] Ser-Nam Lim, Bharath Hariharan, Luming Tang, Menglin Jia, Bor-Chun Chen, Claire Cardie, and Serge Belongie. Visual prompt tuning. 2022. 2" + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "12165", + "bbox": [ + 480, + 926, + 517, + 936 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[22] Jiaming Liu, Ming Lu, Kaixin Chen, Xiaqi Li, Shizun Wang, Zhaoqing Wang, Enhua Wu, Yurong Chen, Chuang Zhang, and Ming Wu. Overfitting the data: Compact neural video delivery via content-aware feature modulation, 2021. 2", + "[23] Pengfei Liu, Weizhe Yuan, Jinlan Fu, Zhengbao Jiang, Hiroaki Hayashi, and Graham Neubig. Pre-train, prompt, and predict: A systematic survey of prompting methods in natural language processing. arXiv: Computation and Language, 2021. 4", + "[24] Yuang Liu, Wei Zhang, and Jun Wang. Source-free domain adaptation for semantic segmentation. computer vision and pattern recognition, 2021.", + "[25] Robert A. Marsden, Felix Wiewel, Mario Döbler, Yang Yang, and Bin Yang. Continual unsupervised domain adaptation for semantic segmentation using a class-specific transfer. 2022. 2", + "[26] Zhongnan Qu, Cong Liu, and Lothar Thiele. Deep partial updating: Towards communication efficient updating for ondevice inference. 2020. 2", + "[27] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2015. 5, 6", + "[28] Christos Sakaridis, Dengxin Dai, and Luc Van Gool. ACDC: The adverse conditions dataset with correspondences for semantic driving scene understanding. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), October 2021. 5", + "[29] Christian Simon, Masoud Faraki, Yi-Hsuan Tsai, Xiang Yu, Samuel Schulter, Yumin Suh, Mehrtash Harandi, and Manmohan Chandraker. On generalizing beyond domains in cross-domain continual learning. 2022. 2", + "[30] Antti Tarvainen and Harri Valpola. Mean teachers are better role models: Weight-averaged consistency targets improve semi-supervised deep learning results. Learning, 2017. 4", + "[31] Eric Tzeng, Judy Hoffman, Kate Saenko, and Trevor Darrell. Adversarial discriminative domain adaptation. computer vision and pattern recognition, 2017. 5", + "[32] Laurens Van Der Maaten. Barnes-hut-sne. arXiv preprint arXiv:1301.3342, 2013. 7", + "[33] Riccardo Volpi, Diane Larlus, and Grégory Rogez. Continual adaptation of visual representations via domain randomization and meta-learning. computer vision and pattern recognition, 2020. 2", + "[34] Dequan Wang, Evan Shelhamer, Shaoteng Liu, Bruno A. Olshausen, and Trevor Darrell. Tent: Fully test-time adaptation by entropy minimization. Learning, 2021. 1, 2, 6", + "[35] Qin Wang, Olga Fink, Luc Van Gool, and Dengxin Dai. Continual test-time domain adaptation. In Proceedings of Conference on Computer Vision and Pattern Recognition, 2022. 1, 2, 5, 6", + "[36] Tong Wang, Yousong Zhu, Chaoyang Zhao, Wei Zeng, Jinqiao Wang, and Ming Tang. Adaptive class suppression loss for long-tail object detection. computer vision and pattern recognition, 2021. 8" + ], + "bbox": [ + 78, + 90, + 470, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[37] Zifeng Wang, Zizhao Zhang, Chen-Yu Lee, Han Zhang, Ruoxi Sun, Xiaoqi Ren, Guolong Su, Vincent Perot, Jennifer Dy, and Tomas Pfister. Learning to prompt for continual learning. 2022. 2", + "[38] Jiangchao Yao, Feng Wang, Kunyang Jia, Bo Han, Jingren Zhou, and Hongxia Yang. Device-cloud collaborative learning for recommendation. knowledge discovery and data mining, 2021. 3", + "[39] Chongzhi Zhang, Mingyuan Zhang, Shanghang Zhang, Daisheng Jin, Qiang Zhou, Zhongang Cai, Haiyu Zhao, Shuai Yi, Xianglong Liu, and Ziwei Liu. Delving deep into the generalization of vision transformers under distribution shifts. arXiv: Computer Vision and Pattern Recognition, 2021. 2", + "[40] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Zwei Liu. Conditional prompt learning for vision-language models. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2", + "[41] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Learning to prompt for vision-language models. International Journal of Computer Vision (IJCV), 2022. 2" + ], + "bbox": [ + 501, + 90, + 890, + 387 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "12166", + "bbox": [ + 480, + 926, + 519, + 936 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2023/Cloud-Device Collaborative Adaptation to Continual Changing Environments in the Real-World/071f2817-bbaa-4424-acc0-27135344a577_model.json b/2023/Cloud-Device Collaborative Adaptation to Continual Changing Environments in the Real-World/071f2817-bbaa-4424-acc0-27135344a577_model.json new file mode 100644 index 0000000000000000000000000000000000000000..7a11fcbc432ff1a45b6f98984137284185338c21 --- /dev/null +++ b/2023/Cloud-Device Collaborative Adaptation to Continual Changing Environments in the Real-World/071f2817-bbaa-4424-acc0-27135344a577_model.json @@ -0,0 +1,2068 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.095, + 0.13, + 0.877, + 0.175 + ], + "angle": 0, + "content": "Cloud-Device Collaborative Adaptation to Continual Changing Environments in the Real-world" + }, + { + "type": "text", + "bbox": [ + 0.259, + 0.203, + 0.71, + 0.24 + ], + "angle": 0, + "content": "Yulu Gan\\(^{1*}\\), Mingjie Pan\\(^{1*}\\), Rongyu Zhang\\(^{2}\\), Zijian Ling\\(^{3}\\), Lingran Zhao\\(^{1}\\), Jiaming Liu\\(^{1}\\), Shanghang Zhang\\(^{1\\dagger}\\)" + }, + { + "type": "text", + "bbox": [ + 0.106, + 0.24, + 0.865, + 0.258 + ], + "angle": 0, + "content": "\\(^{1}\\)National Key Laboratory for Multimedia Information Processing, School of Computer Science, Peking University," + }, + { + "type": "text", + "bbox": [ + 0.227, + 0.257, + 0.744, + 0.275 + ], + "angle": 0, + "content": "\\(^{2}\\)The Chinese University of Hong Kong, Shenzhen, \\(^{3}\\)Imperial College London" + }, + { + "type": "image", + "bbox": [ + 0.094, + 0.304, + 0.462, + 0.449 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.18, + 0.452, + 0.359, + 0.464 + ], + "angle": 0, + "content": "(a) The problem and our main idea" + }, + { + "type": "image", + "bbox": [ + 0.504, + 0.318, + 0.868, + 0.449 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.507, + 0.452, + 0.867, + 0.464 + ], + "angle": 0, + "content": "(b) Comparison of our method with others on continuous domain shifts" + }, + { + "type": "image_caption", + "bbox": [ + 0.075, + 0.489, + 0.893, + 0.545 + ], + "angle": 0, + "content": "Figure 1. (a) Models deployed on devices are preferably lightweight. However, device models will suffer from severe performance degradation when facing continual distribution shift data. Our main idea is to improve the continual domain adaptation capability of the device model by performing our proposed Cloud-Device Collaborative Adaptation paradigm. (b) We compare our method with previous works [16,34,35]. Our method surpasses the state-of-the-art approach and exhibits a solid ability when facing continual distribution shifts." + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.561, + 0.314, + 0.577 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.593, + 0.474, + 0.866 + ], + "angle": 0, + "content": "When facing changing environments in the real world, the lightweight model on client devices suffers from severe performance drops under distribution shifts. The main limitations of the existing device model lie in (1) unable to update due to the computation limit of the device, (2) the limited generalization ability of the lightweight model. Meanwhile, recent large models have shown strong generalization capability on the cloud while they can not be deployed on client devices due to poor computation constraints. To enable the device model to deal with changing environments, we propose a new learning paradigm of Cloud-Device Collaborative Continual Adaptation, which encourages collaboration between cloud and device and improves the generalization of the device model. Based on this paradigm, we further propose an Uncertainty-based Visual Prompt Adapted (U-VPA) teacher-student model to transfer the generalization capability of the large model on the cloud to the device model. Specifically, we first de" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.562, + 0.893, + 0.773 + ], + "angle": 0, + "content": "sign the Uncertainty Guided Sampling (UGS) to screen out challenging data continuously and transmit the most out-of-distribution samples from the device to the cloud. Then we propose a Visual Prompt Learning Strategy with Uncertainty guided updating (VPLU) to specifically deal with the selected samples with more distribution shifts. We transmit the visual prompts to the device and concatenate them with the incoming data to pull the device testing distribution closer to the cloud training distribution. We conduct extensive experiments on two object detection datasets with continually changing environments. Our proposed U-VPA teacher-student framework outperforms previous state-of-the-art test time adaptation and device-cloud collaboration methods. The code and datasets will be released." + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.801, + 0.631, + 0.815 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.826, + 0.893, + 0.903 + ], + "angle": 0, + "content": "Real-world usually contains various environmental changes along with continual distribution shifts [35]. People usually deploy economically lightweight models on devices to boost the scalability and practicability. The lightweight model can suffer severe performance degrad" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.875, + 0.204, + 0.888 + ], + "angle": 0, + "content": "*Equal contribution" + }, + { + "type": "page_footnote", + "bbox": [ + 0.098, + 0.888, + 0.22, + 0.9 + ], + "angle": 0, + "content": "† Corresponding author" + }, + { + "type": "list", + "bbox": [ + 0.096, + 0.875, + 0.22, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.925, + 0.519, + 0.937 + ], + "angle": 0, + "content": "12157" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.272 + ], + "angle": 0, + "content": "tion under continual distribution shifts [25, 29, 33, 35]. The main challenges are: (1) The poor computational ability of devices. Due to the properties of device infrastructure, the deployed model can not be updated on time, thus lagging its performance for the real world under distribution shift. (2) The limited generalization ability of the lightweight model. Since lightweight models are of relatively small capacity, they can not handle continually changing environments. In contrast, recent large models that are trained on the cloud server show significant generalization ability [2, 39]. While in industry, these large models can not be directly applied due to the limited infrastructure." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.273, + 0.473, + 0.424 + ], + "angle": 0, + "content": "Therefore, we enable the device model to tackle real-world environmental changing by proposing a Cloud-Device Collaborative Continual Adaptation paradigm, as shown in Fig.1 (a). Previous Cloud-Device Collaborative methods only focus on improving model representation on variance in video frames but neglect the model generalization ability for continually changing data distribution. In our new paradigm, we fully exploit the sufficient knowledge of the large cloud model and transfer the continual generalization ability to the device lightweight model." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.424, + 0.473, + 0.665 + ], + "angle": 0, + "content": "In particular, we design an Uncertainty-based Visual Prompt Adapted (U-VPA) teacher-student model, which consists of an Uncertainty Guided Sampling (UGS) strategy and a Visual Prompt Learning Strategy with Uncertainty Guided Updating (VPLU). Due to the communication bandwidth constraint [11, 13, 19], different with some work targeting filtering the parameters [22, 26], we design the UGS to screen out the most environment-specific samples and decrease the required bandwidth compared with transmitting the whole sequence. To leverage the strong generalization ability of large models, we introduce the VPLU to align source-target domain distribution and transfer the representation of the large teacher model to the light student model. The light student model and visual prompts are then delivered to the device, thus alleviating the continuously changing scenarios in the real world." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.666, + 0.473, + 0.831 + ], + "angle": 0, + "content": "Experimental results show that our method outperforms the state-of-the-art methods on synthetic and real-world distribution shift datasets, as shown in Fig.1 (b). Besides, we can achieve the same performance as the entire data with fewer reflowed data (42% of total data). As another benefit, fewer reflowed data reduce the bandwidth pressure of the uplink. As for the downlink, we can deliver the visual prompts (0.43% of the model's parameters) with almost negligible bandwidth to the device and apply the visual prompt to the input data to improve the performance of the device model by 3.9% in mAP." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.832, + 0.424, + 0.846 + ], + "angle": 0, + "content": "Our contributions can be summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.856, + 0.47, + 0.901 + ], + "angle": 0, + "content": "- We make the first attempt to deal with continually changing scenarios by proposing a Cloud-Device Collaborative Continual Adaptation paradigm, which aims" + }, + { + "type": "text", + "bbox": [ + 0.531, + 0.092, + 0.892, + 0.152 + ], + "angle": 0, + "content": "to transfer the generalization ability from the large cloud model to the lightweight device model. Our method is a general paradigm that can apply to real-world systems." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.162, + 0.892, + 0.298 + ], + "angle": 0, + "content": "- We design an Uncertainty-based Visual Prompt Adapted (U-VPA) teacher-student model, which consists of UGS and VPLU. We introduce UGS to screen out the most environment-specific samples and decrease the required bandwidth compared with transmitting the whole sequence. We propose a VPLU to align source-target data distribution and transfer the representation of the large teacher model to the lightweight student model." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.307, + 0.892, + 0.368 + ], + "angle": 0, + "content": "- Experiments show that our proposed framework and method surpass other state-of-the-art methods and can effectively improve the continuous domain adaptation capability of the device model." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.393, + 0.638, + 0.408 + ], + "angle": 0, + "content": "2. Related work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.418, + 0.892, + 0.538 + ], + "angle": 0, + "content": "Test time adaptation. Test-time adaptation(TTA) aims to get better model generalization performance on target data without access to source data. Feature modulation is used in TENT [34] to optimize BN layers of the input pretrained model by entropy minimization. SHOT [20] takes entropy minimization and a diversity regularizer to achieve information maximization. [18, 24] both take a productive way to enhance model performance on the target domain." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.539, + 0.892, + 0.644 + ], + "angle": 0, + "content": "We differ from TTA in that we consider continuously changing domains. Besides, our model is deployed on the device. There will be restrictions on computing resources and uplink and downlink transmission bandwidth. Considering TTA can effectively use the test data to make updates, we leverage TTA to enable the cloud to improve generalization performance continuously." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.645, + 0.892, + 0.78 + ], + "angle": 0, + "content": "Prompt learning. Prompt is originally from the natural language process (NLP) and aims to adapt the large-scale pre-trained language model to downstream tasks. Recently, several methods [1, 9, 21, 40, 41] are proposed to prove the feasibility of using prompts to adapt the large pre-trained model to downstream tasks in computer vision or vision-language tasks. [1, 21, 37] take prompt as additional input containing a small number of learnable parameters and explore the design of prompts." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.781, + 0.892, + 0.855 + ], + "angle": 0, + "content": "Prompt learning aims at mining the pre-trained model's information. We introduce visual prompt tuning to improve the performance of the device model on the current domain, with a negligible computational overhead in our Cloud-Device Collaboration (DCC) framework." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Cloud-device collaboration. Cloud-device collaboration aims to bring benefits to both the cloud and the device jointly. Several methods [4, 6, 7, 15] try to unload part or" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.925, + 0.519, + 0.937 + ], + "angle": 0, + "content": "12158" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.094, + 0.104, + 0.885, + 0.445 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.461, + 0.896, + 0.574 + ], + "angle": 0, + "content": "Figure 2. The whole framework. (a): Testing on the device. On the client device, we select data that need to be sent to cloud by Uncertainty Guided Sampling (UGS). These uncertainty samples are then used in the cloud to update the teacher-student models. The visual prompt trained on the cloud will be delivered to the device through the downlink and applied to the input data \\( x \\). Then the reformulated input data will be put into the device model for testing. (b) Training on the cloud. We add our proposed visual prompt to the uncertain samples at the pixel level to get the reformulated images. We then put the reformulated images into the teacher-student model. The teacher-student model and visual prompts are optimized by align loss and supervised loss. (c) Visual prompt learning strategy with uncertainty guided sampling (VPLU). We update the visual prompts by uncertainty-based EMA update. The update weight of the visual prompt will be larger when the uncertainty is higher. The student model is updated using the uncertainty samples and the pseudo labels." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.581, + 0.47, + 0.701 + ], + "angle": 0, + "content": "all computation to the cloud to alleviate the computation deficiency on the device. But they do not leverage the collaboration between the cloud and the device. AMS [16] and DCCL [38] have considered the collaboration but they do not take the real-world continual distribution shifts into account. Compared with methods above, our proposed CloudDevice Collaborative Continual Adaptation paradigm can deal with the real-world environment changes." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.702, + 0.471, + 0.779 + ], + "angle": 0, + "content": "In our work, we use Uncertainty Guided Sampling (UGS) and visual prompt for more efficient bandwidth usage. Besides, our model can adapt to different domains during test time. Our teacher model on the cloud and our student model on the device can be improved simultaneously." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.79, + 0.169, + 0.807 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.815, + 0.208, + 0.831 + ], + "angle": 0, + "content": "3.1. Preliminary" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.838, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Given a lightweight model \\(\\theta_{device}\\) that is deployed on the device, our goal is to perform well on the continually changing target domains \\(D_{t1}, D_{t2}, \\ldots, D_{tT}\\) with limited communication bandwidth, where \\(D_{ti} = \\{(x_i^T)\\}_{i=1}^{N_t}\\), and \\(N_t\\) repre" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.581, + 0.892, + 0.658 + ], + "angle": 0, + "content": "sents the scale of the target domain. The distributions of the target domains are arbitrary since they can change or reoccur over time. We thus introduce a large teacher model \\(\\theta_{tea}\\) and lightweight student model \\(\\theta_{stu}\\) cloned from \\(\\theta_{device}\\), to boost the generalization ability of \\(\\theta_{device}\\)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.667, + 0.686, + 0.681 + ], + "angle": 0, + "content": "3.2. Overall framework" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.689, + 0.892, + 0.81 + ], + "angle": 0, + "content": "In our Cloud-Device Collaborative Continual Adaptation paradigm, the key idea is to improve the device model's performance continuously by receiving adapted parameters from the cloud server. So we propose Visual Prompt Adapted (U-VPA) teacher-student model. As shown in Fig. 2, the U-VPA teacher-student model contains Uncertainty Guided Sampling (UGS) and Visual Prompt Learning Strategy with Uncertainty Guided Updating (VPLU)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.811, + 0.893, + 0.901 + ], + "angle": 0, + "content": "When perceiving scene changing or distribution shift, poor computational device intends to get help from a cloud server with powerful computing ability. However, due to the communication bandwidth constraint, we must consider the required bandwidth when concurring the distribution shifts. Therefore, we design a Uncertainty Guided Sam" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.926, + 0.519, + 0.938 + ], + "angle": 0, + "content": "12159" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.181 + ], + "angle": 0, + "content": "pling (UGS) to screen out the most environment-specific samples and decrease the required bandwidth compared with transmitting the whole sequence. After receiving the learned parameters from the cloud to deal with these samples of relatively large distribution shifts, it can adapt to the new scene when inference." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.183, + 0.469, + 0.334 + ], + "angle": 0, + "content": "On the cloud server, due to its strengths in computing, we intend to leverage the strong generalization ability of large models to tackle the received environment-specific and challenging samples. We introduce Visual Prompt Learning Strategy with Uncertainty guided Updating (VPLU) to align source-target domain distribution and transfer the representation of the large teacher model to the light student model. The light student model and visual prompts is then delivered to the device, thus alleviating the continuously changing scenarios in the real-world." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.335, + 0.468, + 0.364 + ], + "angle": 0, + "content": "We will discuss the details of UGS and VPLU in Sec. 3.3.1 and Sec. 3.3.2." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.378, + 0.343, + 0.392 + ], + "angle": 0, + "content": "3.3. U-VPA teacher-student model" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.402, + 0.388, + 0.417 + ], + "angle": 0, + "content": "3.3.1 Uncertainty Guided Sampling(UGS)" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.427, + 0.469, + 0.533 + ], + "angle": 0, + "content": "Due to bandwidth limitations, it is not practical for the device to return all data. Therefore, we select images that need to be transmitted based on uncertainty estimation. Inspired by the Dropout methods [8, 12], since we are doing an object detection task, we can obtain \\( n \\) probabilities \\( p_i(y|x) \\) of the model by dropout method, and we calculate the variance of these \\( n \\) prediction \\( p_i(y|x) \\) as uncertainty." + }, + { + "type": "equation", + "bbox": [ + 0.161, + 0.548, + 0.469, + 0.585 + ], + "angle": 0, + "content": "\\[\nV _ {u n c} = \\left(\\frac {1}{n} \\sum_ {i} ^ {n} \\left(p _ {i} (y | x) - \\mu\\right) ^ {2}\\right) ^ {\\frac {1}{2}} \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.596, + 0.469, + 0.685 + ], + "angle": 0, + "content": "where \\( p(y|x) \\) represents the predictions of the input image \\( x \\) and \\( \\mu \\) is the mean. \\( V_{unc} \\) is the variance of \\( n \\) probabilities of the model, which represents the uncertainty of the model for the data. The device model filters the input images \\( x \\) whose uncertainty is larger than a threshold and sends back to the cloud server." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.71, + 0.468, + 0.74 + ], + "angle": 0, + "content": "3.3.2 Visual Prompt Learning Strategy with Uncertainty Guided Updating (VPLU)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.469, + 0.901 + ], + "angle": 0, + "content": "We optimize model on the cloud server by using the data transmitted from the device. It then delivers the learned parameters and processed visual prompt to the device model for inference. We thus introduce VPLU with a large teacher model and light student model, to mitigate variant domain distribution and improve the generalization ability of light model. Note that, teacher model is initialized by the source model pre-trained on source domain. It takes source and target domain as input, while student model only trains on the target domain." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.227 + ], + "angle": 0, + "content": "Cross-domain alignment. In order to mitigate distribution shift, we align the features in the source and target domains. Inspired by [10], we apply feature alignment on teacher model to improve its performance on target domain. We use a domain discriminator to classify data in feature space and try to distinguish which domain the data comes from. Besides, we use a gradient reversal layer to achieve a uniform training process. Source domain data is used to supervise teacher model, avoiding performance degradation." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.228, + 0.892, + 0.287 + ], + "angle": 0, + "content": "Cross-model transfer. After improving teacher model's domain-invariant representation ability, we transfer its ability against distribution shifts and generalization to student model via visual prompts and knowledge transferring." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.288, + 0.892, + 0.363 + ], + "angle": 0, + "content": "Visual Prompt Adaptation(VPA) learning. Specifically, inspired by the prompt learning in NLP [23], which can effectively explores the knowledge of source domain, we propose visual prompts as hints to continually transfer teacher model's generalized representation to the student model." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.364, + 0.892, + 0.424 + ], + "angle": 0, + "content": "In particular, visual prompts \\(\\phi\\) are learnable parameters. During the forward process of the model, we can obtain the reformulated image \\(x^{*}\\) by adding visual prompts on the input images \\(x\\) at the pixel level." + }, + { + "type": "equation", + "bbox": [ + 0.652, + 0.439, + 0.891, + 0.454 + ], + "angle": 0, + "content": "\\[\nx ^ {*} = x + \\phi , \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.459, + 0.891, + 0.504 + ], + "angle": 0, + "content": "In our framework, whether in training on server or inference on device, we all use the reformulated image \\( x^{*} \\) as model's input, instead of origin images \\( x \\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.505, + 0.892, + 0.64 + ], + "angle": 0, + "content": "Uncertainty-aware Prompt Updation(U-EMA). Visual prompt acts on the input data and participates in the forward process of both teacher and student models, so it can also benefit from their optimization during training. Moreover, to make the prompt update more stable and consider the current model's uncertainty to data, we update the visual prompt using by uncertainty-based exponential moving average(U-EMA). The weight of the exponential moving average is based on the uncertainty estimation as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.606, + 0.655, + 0.891, + 0.671 + ], + "angle": 0, + "content": "\\[\n\\phi_ {t} \\leftarrow \\beta \\phi_ {t - 1} + (1 - \\beta) \\phi_ {t}, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.675, + 0.892, + 0.811 + ], + "angle": 0, + "content": "Note that \\(\\phi_t\\) represents the parameters of the visual prompt that need to be updated, and \\(\\phi_{t - 1}\\) indicates the visual prompt parameters of the last training step t. We follow [30] to set \\(\\alpha = 0.999\\). And \\(\\beta = \\alpha -V_{unc}\\), where \\(V_{unc}\\) is defined in Eq. (1) as the uncertainty of current data. Higher uncertainty means a larger distribution shift in the current data. To save more domain-related information, the prompt update weight will be larger at this time. That is, the larger the uncertainty, the greater the weight of the prompt update." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.811, + 0.892, + 0.871 + ], + "angle": 0, + "content": "As another essential role of the visual prompt, the processed visual prompt will be sent to the device to improve the adaptability of the device model by applying the visual prompt on input images." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Knowlegde transfer. Along with visual prompt adapted tuning, we align teacher student feature and adopt teacher" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.925, + 0.519, + 0.936 + ], + "angle": 0, + "content": "12160" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.274 + ], + "angle": 0, + "content": "model to generate pseudo label for student model. It is penalized by Knowledge transfer loss \\(\\mathcal{L}_k\\), to promote the consistency between teacher student model representation. Specifically, The teacher model will generate predictions for the reformulated images \\(x^{*}\\) as pseudo labels. Teacher model has stronger generalization ability, so the generated pseudo labels are of high quality. To further improve the pseudo labels quality, we screen those predictions with confidence higher than the threshold as the input to the student model. Overall, we transfer the knowledge from the teacher to the student by feeding selected images \\(x^{*}\\) and corresponding pseudo labels into the student model for training." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.281, + 0.262, + 0.297 + ], + "angle": 0, + "content": "3.4. Training objectives" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.304, + 0.47, + 0.379 + ], + "angle": 0, + "content": "The teacher and student models are jointly optimized on the cloud server, using uncertain samples \\( x_{t}^{i}\\in D_{t} \\) returned from device. The source model is trained on source pairs of \\( \\{x_s^i,y_s^i\\}_{i = 1}^{n_s}\\in D_s \\) where \\( x_{s}^{i}\\in X_{s} \\) and \\( y_{s}^{i}\\in Y_{s} \\) are images and labels, respectively." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.38, + 0.47, + 0.44 + ], + "angle": 0, + "content": "For the training process of the teacher model, given the reformulated taregt images \\( x_{t}^{*} \\), as the Eq. (4) shown, we use adversal alignment loss [31] and supervision loss to optimize the teacher model (including visual prompt \\( \\phi \\))." + }, + { + "type": "equation", + "bbox": [ + 0.1, + 0.459, + 0.47, + 0.495 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} L _ {t e a} = - \\lambda L _ {d} \\left(G _ {d} \\left(F \\left(x _ {s} ^ {*}; \\theta_ {t e a}\\right); F \\left(x _ {t} ^ {*}; \\theta_ {t e a}\\right)\\right), y _ {d}\\right) \\tag {4} \\\\ + L _ {D e t} \\left(X \\left(x _ {s} ^ {*}, \\theta_ {t e a}\\right), y _ {s}\\right), \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.502, + 0.47, + 0.592 + ], + "angle": 0, + "content": "Note that the first part is adversal alignment loss [31], and another one is the loss function of Faster RCNN [27] (using labeled data in source domain \\(D_{s}\\)). \\(X\\) represents the model which predict detection outputs. \\(L_{d}\\) represent domain classifier. \\(F\\) indicates the feature extractor and \\(G_{d}\\) is the domain classifier. \\(\\lambda\\) is the weight of alignment loss." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.593, + 0.469, + 0.623 + ], + "angle": 0, + "content": "For the optimization of the student model, we use pseudo labels \\( y_{t}^{p} \\) generated by teacher model as supervision:" + }, + { + "type": "equation", + "bbox": [ + 0.168, + 0.637, + 0.469, + 0.654 + ], + "angle": 0, + "content": "\\[\nL _ {s t u} = L _ {D e t} \\left(X \\left(x _ {t} ^ {*}, \\theta_ {s t u}\\right), y _ {t} ^ {p}\\right), \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.659, + 0.469, + 0.704 + ], + "angle": 0, + "content": "The visual prompt benefited from both \\( L_{tea} \\) and \\( L_{stu} \\). And we update the visual prompt by uncertainty-aware exponential moving average as Eq. (3) shown." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.715, + 0.21, + 0.733 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.741, + 0.469, + 0.801 + ], + "angle": 0, + "content": "In Sec.4.1, we give the details of the dataset and experiment setting. We evaluate the performance of CCA on continual changing scenarios in Sec.4.2. We give discussion on our motivations and ablation study in Sec.4.3, Sec.4.4." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.809, + 0.265, + 0.825 + ], + "angle": 0, + "content": "4.1. Experimental setup" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.833, + 0.277, + 0.848 + ], + "angle": 0, + "content": "4.1.1 Datasets and settings" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.856, + 0.469, + 0.901 + ], + "angle": 0, + "content": "We evaluate our proposed method on two test time adaptation benchmark tasks for object detection: Cityscapes-to-Cityscape-C and Cityscapes-to-ACDC-Detection dataset." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.212 + ], + "angle": 0, + "content": "Cityscape-C is originally created to benchmark robustness tasks [14]. We select the five most relevant corruptions to the autonomous driving scenario (including brightness, motion_blur, rain, fog and snow). Each corruption dataset contains five levels of severity. The corruptions are applied to images from the validation set of the clean Cityscapes dataset [5]. There are 2,500 images in five corruption datasets." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.213, + 0.892, + 0.287 + ], + "angle": 0, + "content": "For Cityscapes-to-Cityscapes-C, we follow [3] to train the pre-trained model on the source domain (Cityscapes). Specifically, we use the SGD optimizer with a learning rate of 1e-2 as the official implementation to train the source model." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.29, + 0.893, + 0.501 + ], + "angle": 0, + "content": "ACDC-Detection is created from the Adverse Conditions Dataset (ACDC) [28]. The ACDC dataset shares the same classes with Cityscapes and is collected in various adverse visual conditions: Fog, Night, Rain and Snow. We convert the segmented labels into labels for object detection. We follow CoTTA [35] to set our task. We use 400 unlabeled images from each adverse condition for the adaptation. To mimic the scenario in real life where a similar environment might be revisited and to evaluate the continuous generalization capability of our methods, we repeat the same sequence group (of the three conditions) 10 times (i.e., in total 30: \\(\\mathrm{Fog} \\to \\mathrm{Rain} \\to \\mathrm{Snow} \\to \\mathrm{Fog}\\ldots\\)). This also provides an evaluation of the adaptation performance in the long term." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.502, + 0.892, + 0.548 + ], + "angle": 0, + "content": "For the Cityscapes-to-ACDC-Detection task, we use the same way to train the pre-trained model on the source domain (Cityscapes) as the Cityscapes-to-Cityscapes-C task." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.572, + 0.719, + 0.587 + ], + "angle": 0, + "content": "4.1.2 Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.597, + 0.892, + 0.687 + ], + "angle": 0, + "content": "In this paper, all experiments are conducted with PyTorch. For the Cityscapes-to-Cityscapes-C task and Cityscapes-to-ACDC-Detection task, the batch size is set to 8. We follow [3] to train the pre-trained model on the source domain (Cityscapes). The number of samples used for uncertainty estimation is set to 10 in the experiment." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.689, + 0.892, + 0.884 + ], + "angle": 0, + "content": "We use FasterRCNN with the backbone of ResNet-101 as the large teacher model on the cloud. And we use FasterRCNN with the backbone of ResNet-18 as the lightweight student model (same as the device model). The above large and light models are initialized with corresponding pretrained models on the source domain (Cityscapes dataset). For the UGS module, we set the uncertainty threshold to 0.008, which filters out \\(42\\%\\) samples in the Cityscape-C dataset. For the VPA module, We set the visual prompts with size \\(3 \\times 200 \\times 200\\), which can be added to the left-up space of the image. The student model's parameters and visual prompts (about \\(0.4\\%\\) of the student model's parameters) will be transmitted to the device." + }, + { + "type": "text", + "bbox": [ + 0.52, + 0.886, + 0.892, + 0.9 + ], + "angle": 0, + "content": "We use \\(\\mathrm{mAP} @ 0.5(\\%)\\) as the evaluation metric. The" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.927, + 0.518, + 0.938 + ], + "angle": 0, + "content": "12161" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.089, + 0.895, + 0.161 + ], + "angle": 0, + "content": "Table 1. Continual generalization capability on Cityscapes-to-Cityscapes-C. Object detection results (mAP@0.5 in %) on the Cityscapes-to-Cityscapes-C online continual test-time adaptation task. Gain(%) means the improvement of our method compared with Source-only. We evaluate the four test conditions continually for ten times to evaluate the long-term adaptation performance. All results are evaluated on the FasterRCNN architecture with the largest corruption severity level 5. Our approach surpasses the SOTA method and exhibits significant continual generalization and anti-forgetting abilities." + }, + { + "type": "table", + "bbox": [ + 0.078, + 0.167, + 0.892, + 0.279 + ], + "angle": 0, + "content": "
Timet
Round1510All
ConditionFogMotionRainSnowBrightnessFogMotionRainSnowBrightnessFogMotionRainSnowBrightnessMeanGain
Source-only [27]24.79.522.41.329.024.79.522.41.329.024.79.522.41.329.017.4/
TENT-continual [34]24.89.222.51.429.420.06.419.80.823.79.03.810.70.511.413.0-4.4
CoTTA [35]25.110.122.41.328.921.89.520.74.228.29.33.910.13.317.514.7-2.7
Pseudo-Label [17]27.317.224.76.933.730.120.126.98.036.531.020.927.77.238.224.1+6.7
AMS [16]28.417.726.07.634.731.820.428.27.738.034.221.428.97.039.425.1+7.7
Ours(proposed)28.017.826.810.036.537.524.235.514.244.044.027.340.915.150.131.0+13.6
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.282, + 0.895, + 0.352 + ], + "angle": 0, + "content": "Table 2. Continuous generalization capability on Cityscapes-to-ACDC-Detection. Object detection results (mAP@0.5 in %) on the Cityscapes-to-ACDC-Detection online continual test-time adaptation task. Gain(%) means the improvement of our method compared with Source-only. We evaluate the four test conditions continually for ten times to evaluate the long-term adaptation performance. All results are evaluated on the FasterRCNN architecture. Our results outperform the SOTA method, thus verifying the effectiveness of our method on the real distribution shift." + }, + { + "type": "table", + "bbox": [ + 0.078, + 0.359, + 0.892, + 0.484 + ], + "angle": 0, + "content": "
Timet
Round14710All
ConditionFogNightRainSnowFogNightRainSnowFogNightRainSnowFogNightRainSnowMeanGain
Source-only [27]20.57.414.314.220.57.414.314.220.57.414.314.220.57.414.314.214.1/
TENT-continual [34]20.57.514.313.920.07.213.413.018.66.612.312.116.66.011.110.912.8-2.3
CoTTA [35]20.77.514.714.320.97.414.714.420.47.313.814.416.36.411.712.213.8-0.3
Pseudo-Label [17]20.98.715.517.623.410.316.918.924.310.917.819.525.111.517.520.017.6+3.5
AMS [16]21.69.016.017.223.210.717.219.124.711.417.219.926.111.717.820.517.8+3.7
Ours(proposed)21.19.315.717.224.711.817.820.226.312.018.520.626.112.318.020.818.5+4.4
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.499, + 0.471, + 0.576 + ], + "angle": 0, + "content": "overall process of test-time adaptation strictly follows CoTTA [35]. The process first tests the coming batch of data then train with it. We do batch-level updates on the cloud model, and the performance on the device model is taken as the average of the results of all the data." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.599, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Baselines. Our method is compared with multiple baselines, including those for test time adaptation (TTA), continual test time adaptation (CTTA) and cloud-end collaboration methods as follows: Source-only [27]. Test the model performance on the different target domains using the pre-trained source model (training on the source domain). Pseudo-label [17]. Use pseudo label generated by the teacher model to supervise the student model. Tent-continual [34]. Tent is an effective method for the test-time adaptation (TTA). Tent fine-tune the batch normalization layer during training to adapt the model to the target domain. CoTTA [35]. CoTTA is an attractive approach for the continual test-time adaptation (CTTA). CoTTA mitigate error accumulation and catastrophic forgetting with weighted averaging, data augmented averaging, and random recovery neurons. AMS [16]. AMS is an cloud engagement approach. To solve the resource limitation of the device model, AMS do distillation in the cloud, and then transmit the model's parameters to the device in the cloud, the teacher network will not update." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.498, + 0.696, + 0.515 + ], + "angle": 0, + "content": "4.2. Results and Analysis" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.522, + 0.892, + 0.825 + ], + "angle": 0, + "content": "Synthetic continual distribution shift. We evaluate our proposed method on the synthetic continual distribution shift dataset (Cityscapes \\(\\rightarrow\\) Cityscapes-C task). As shown in Table 1, TENT-continual [34] has a decreasing performance when facing continuously changing domains. Because TENT does not consider continuous distribution shifts, it uses self-entropy to update the bn layer, which is prone to catastrophic forgetting when facing constant distribution shifts. CoTTA [35] is a method designed specifically for continuous test time adaptation. CoTTA applies weighted averaging, data-augmented averaging, and random recovery neurons to overcome catastrophic forgetting and error accumulation. However, when doing the object detection task on the continuously changing domains, the model's performance still decreases, indicating that solutions used by CoTTA are insufficient against catastrophic forgetting. We believe this is caused by the poor generalization abilities of the lightweight device model. Therefore, it is reasonable to introduce the cloud for the Cloud-Device collaborative continual adaptation paradigm." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.825, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Both Pseudo-Label and AMS [16] use the Cloud-Device collaboration framework, which transfers the teacher's knowledge to the student on the cloud. The generalization of the large model provides continuous guidance for the student model, which avoids the problem of catastrophic for" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.926, + 0.521, + 0.938 + ], + "angle": 0, + "content": "12162" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.1, + 0.099, + 0.27, + 0.209 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.276, + 0.099, + 0.446, + 0.208 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.177, + 0.212, + 0.188, + 0.221 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.099, + 0.223, + 0.446, + 0.322 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.345, + 0.473, + 0.485 + ], + "angle": 0, + "content": "Figure 3. (a)-(b): Statistical distribution of confidence and uncertainty. (a): distribution of confidence. (b): distribution of uncertainty. Comparing (a) and (b), the uncertainty value is more stable when facing continuous distribution shift, while the confidence value fluctuates significantly with continuous distribution shift. (c)-(e): T-SNE [32] visualization of samples in different domains. (c): all samples. (d): samples filtered by confidence. (e): samples filtered by UGS. Samples filtered by UGS are more environment-specific. They are more worthy of being transmitted to the cloud server for optimizing the teacher-student models." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.508, + 0.47, + 0.54 + ], + "angle": 0, + "content": "getting and error accumulation. (Pseudo-Label and AMS are \\(6.7\\%\\) and \\(7.7\\%\\) higher than Source-only)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.546, + 0.47, + 0.652 + ], + "angle": 0, + "content": "However, pseudo label and AMS do not consider updating the teacher model. Our U-VPA framework jointly optimizes the teacher and student model, which makes our method surpass Source-only by \\(13.6\\%\\), outperforming all the above methods. This validates that our Cloud-Device collaboration paradigm can effectively improve the continuous domain adaptation ability of the device." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.66, + 0.471, + 0.901 + ], + "angle": 0, + "content": "Real-world continual distribution shift. We also evaluate our proposed method on the real-world continual distribution shift dataset (Cityscapes \\(\\rightarrow\\) ACDC-Detection task). As shown in Table 2, our method improves by \\(4.4\\%\\) over the source-only method. By comparing the performance change of the same domain in 10 rounds, we can find that our method does not have the problem of catastrophic forgetting. For example, the TENT-continual method faces the problem of catastrophic forgetting. The \\(\\mathrm{mAP}(\\%)\\) drops in the fourth, seventh, and tenth rounds compared to the first round when facing the night domain, specifically \\(-0.2\\%\\), \\(-0.5\\%\\), \\(-4.1\\%\\). In contrast, our method leads to a continuous performance improvement (e.g., our method's \\(\\mathrm{mAP}(\\%)\\) increases in the fourth, seventh, and tenth rounds, respectively, compared to the first round when facing the Night domain, specifically, \\(+2.5\\%\\), \\(+0.2\\%\\), \\(+0.3\\%\\))." + }, + { + "type": "image", + "bbox": [ + 0.549, + 0.105, + 0.825, + 0.25 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.262, + 0.895, + 0.376 + ], + "angle": 0, + "content": "Figure 4. Detection results (mAP@0.5) of both teacher and student model. Solid lines are results of models using VPA under the brightness domain. Teacher and student models have jointly improved in continuous domain adaptation. Dotted lines are results without using VPA. The teacher quickly reaches the bottleneck, and student's improvement is slow. The analysis above illustrates that the proposed U-VPA framework enables the teacher-student model jointly improve the continuous domain adaption capability." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.391, + 0.618, + 0.405 + ], + "angle": 0, + "content": "4.3. Discussion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.415, + 0.892, + 0.521 + ], + "angle": 0, + "content": "Teacher-student joint optimization with VPA. Previous approaches (e.g., AMS) focus on improving the student model. However, the student model's performance is correlated with the teacher model. Our method enables the teacher and student models to improve performance together by using the visual prompt as a bridge and using the temporal domain information brought by the visual prompt." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.521, + 0.892, + 0.596 + ], + "angle": 0, + "content": "The optimization of the teacher model comes from two modules: (1) feature alignment, (2) visual prompt adaptation(VPA). Our experiments have proved that although both of the above can bring about gains, visual prompt adaptation is the main reason for the success of our framework." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.597, + 0.892, + 0.657 + ], + "angle": 0, + "content": "As shown in Fig. 4, if only feature alignment is used, the performance of the teacher model reaches the bottleneck after the second round, and the student's optimization will be very slow (dotted lines in the figure)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.658, + 0.892, + 0.718 + ], + "angle": 0, + "content": "While our method benefits from VPA, the teacher and student models are jointly improved, demonstrating that our proposed U-VPA teacher-student model can optimize the student-teacher model together." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.719, + 0.892, + 0.885 + ], + "angle": 0, + "content": "Uncertainty Guided Sampling (UGS) strategy. We demonstrate in Table 5 that using only a small amount of reflowed data can outperform the results using total data by the UGS strategy. Our model's performance using the UGS strategy to select part of the data to transmit into the cloud for training is \\(31\\%\\). Compared to the model performance of \\(28.3\\%\\) obtained with the total data, there is a \\(2.7\\%\\) improvement. The amount of data using the random picking strategy is the same as that with the UGS strategy. And the UGS strategy has a better performance than random picking. The above shows the effectiveness of the UGS strategy." + }, + { + "type": "text", + "bbox": [ + 0.52, + 0.886, + 0.892, + 0.901 + ], + "angle": 0, + "content": "We further reveal the advantages of uncertainty when the" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.926, + 0.519, + 0.938 + ], + "angle": 0, + "content": "12163" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.082, + 0.895, + 0.127 + ], + "angle": 0, + "content": "Table 3. Ablation study. We conduct ablation studies for proposed Pseudo-Label, VPA, UGS, and U-EMA. All the experiments are done on Cityscapes-to-Cityscapes-C, and the evaluation metric is mAP (\\%). The experimental result shows that all these proposed modules positively impact the performance improvement of the devices model when facing continuous domains." + }, + { + "type": "table", + "bbox": [ + 0.112, + 0.135, + 0.857, + 0.228 + ], + "angle": 0, + "content": "
Pseudo-LabelVPAUGSU-EMAFogMotionRainSnowBrightnessMeanGain
124.79.522.41.329.017.4/
2✓31.020.927.77.238.224.1+6.7
3✓✓37.122.232.711.243.127.4+10.0
4✓✓✓38.324.836.512.346.728.3+10.9
5✓✓✓✓44.027.340.915.150.131.0+13.6
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.235, + 0.472, + 0.292 + ], + "angle": 0, + "content": "Table 4. Comparing the difference between selecting samples with confidence and uncertainty. Experiments show that better performance is obtained by choosing the reflowed data model through uncertainty estimation." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.295, + 0.462, + 0.357 + ], + "angle": 0, + "content": "
FogMotionRainSnowBrightnessMean
Confidence39.624.937.214.345.127.4
Uncertainty44.027.340.915.150.131.0
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.364, + 0.47, + 0.435 + ], + "angle": 0, + "content": "Table 5. Effect of uplink transmission of full and partial data on model's performance. Our method returns partial data (\\( \\tilde{42\\%} \\)) for training but achieves better performance than using full data. This indicates that our UGS strategy reduces the data transmission communication cost and improves the model performance." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.439, + 0.462, + 0.504 + ], + "angle": 0, + "content": "
25%50%75%100%
Random23.526.127.928.3
UGS28.231.031.2/
Gain+4.7+4.9+3.3/
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.516, + 0.47, + 0.667 + ], + "angle": 0, + "content": "distribution shifts exist. We find uncertainty is more reliable compared with the confidence score. Regarding Fig.3 (a)-(b), in the presence of distribution shifts, confidence distribution varies significantly, while uncertainty distribution is balanced. Therefore, the samples selected using confidence will face the problem of long-tail distribution, which is not conducive to cloud model training [36]. Uncertainty can more accurately reflect the difficulty of identifying samples by device without using the training of cloud model and discover the samples that are more worthy of attention." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.668, + 0.47, + 0.745 + ], + "angle": 0, + "content": "Besides, we explore the properties of the data filtered by different filter strategies. In Fig.3 (c)-(e), each point represents a sample filtered by the corresponding strategy. We find that samples selected by uncertainty are environmental-specific, so using UGS can filter out more valuable data." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.755, + 0.228, + 0.772 + ], + "angle": 0, + "content": "4.4. Ablation study" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.779, + 0.468, + 0.884 + ], + "angle": 0, + "content": "Effect of the Uncertainty Guided Sampling (UGS). Our proposed UGS strategy not only reduces the communication cost of the uplink from the device to the cloud but also improves performance by \\(0.9\\%\\), as shown in Table 3. Furthermore, as shown in Table 5, the model using UGS outperforms baseline with only \\(42\\%\\) data for train, which significantly improves the compression ratio of the video." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.886, + 0.468, + 0.902 + ], + "angle": 0, + "content": "Effect of the Visual Prompt Adapt (VPA). Our pro" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.237, + 0.892, + 0.373 + ], + "angle": 0, + "content": "posed VPA teacher-student model uses visual prompts to make the student-teacher model jointly improve performance. As shown in Table 3, compared to the pure pseudo label method, the usage of VPA can increase performance by \\(3.3\\%\\). Furthermore, we discussed in Sec. 4.3 that VPA could significantly increase the upper limit of the teacher-student framework from challenging continuously changing distribution shift scenarios. The combination of VPA and other modules can further improve performance steadily." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.374, + 0.892, + 0.482 + ], + "angle": 0, + "content": "Effect of the Uncertainty-aware prompt EMA (U-EMA) update mechanism. The U-EMA update mechanism we proposed in training the cloud model makes the cloud model pay more attention to the images worthy of attention by dynamically adjusting the prompt weights with uncertainty. As shown in Table 3, U-EMA can bring a \\(2.7\\%\\) performance improvement." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.5, + 0.742, + 0.517 + ], + "angle": 0, + "content": "5. Conclusion & Limitations" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.528, + 0.892, + 0.829 + ], + "angle": 0, + "content": "We propose a Cloud-Device Collaborative Continual Adaptation paradigm to deal with continual changing environments on devices. We design an Uncertainty Guided Sampling (UGS) strategy to transmit the most out-of-distribution samples from device to cloud. Besides, we design a Visual Prompt Learning Strategy with Uncertainty Guided Updating (VPLU) to transfer the generalization capability of large model on the cloud to the device model. Experiments of our method on a wide range of continually changing environments show an improvement of \\(4.4 - 13.6\\%\\) in mAP@0.5 on the object detection task. Our proposed method offers flexibility in its application to most existing cloud-device systems. However, it is worth noting that our experiments were conducted using simulated datasets rather than real systems. Moving forward, an interesting direction for future research would be to test our method on a real system with specific bandwidth limitations, a specific cloud, and a specific device. By taking into account the hardware characteristics of the system, there should have rooms to improve its performance further." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.83, + 0.892, + 0.902 + ], + "angle": 0, + "content": "Acknowledgement. This research was partly supported by the foundation of the National Key R&D Program of China (2022ZD0116305). We thank Xiaoqi Li for the insightful discussions and for helping with writing." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.926, + 0.521, + 0.938 + ], + "angle": 0, + "content": "12164" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.116, + 0.47, + 0.157 + ], + "angle": 0, + "content": "[1] Hyojin Bahng, Ali Jahanian, Swami Sankaranarayanan, and Phillip Isola. Exploring visual prompts for adapting large-scale models. 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.16, + 0.471, + 0.575 + ], + "angle": 0, + "content": "[2] Rishi Bommasani, Drew A. Hudson, Ehsan Adeli, Russ B. Altman, Simran Arora, Sydney von Arx, Michael S. Bernstein, Jeannette Bohg, Antoine Bosselut, Emma Brunskill, Erik Brynjolfsson, Shyamal Buch, Dallas Card, Rodrigo Castellon, Niladri S. Chatterji, Annie Chen, Kathleen Creel, Jared Davis, Dora Demszky, Chris Donahue, Moussa Doumbouya, Esin Durmus, Stefano Ermon, John Etchemendy, Kawin Ethayarajh, Li Fei-Fei, Chelsea Finn, Trevor Gale, Lauren Gillespie, Karan Goel, Noah D. Goodman, Shelby Grossman, Neel Guha, Tatsunori Hashimoto, Peter Henderson, John Hewitt, Daniel E. Ho, Jenny Hong, Kyle Hsu, Jing Huang, Thomas Icard, Sahil Jain, Dan Jurafsky, Pratyusha Kalluri, Siddharth Karamcheti, Geoff Keeling, Fereshte Khani, Omar Khattab, Pang Wei Koh, Mark Krass, Ranjay Krishna, Rohith Kuditipudi, Ananya Kumar, Faisal Ladhak, Mina Lee, Tony Lee, Jure Leskovec, Isabelle Levent, Xiang Lisa Li, Xuechen Li, Tengyu Ma, Ali Ahmad Malik, Christopher D. Manning, Suvir Mirchandani, Eric Mitchell, Zanele Munyikwa, Suraj Nair, Avanika Narayan, Deepak Narayanan, Ben Newman, Allen Nie, Juan Carlos Niebles, Hamed Nilforoshan, Julian Nyarko, Giray Ogut, Laurel Orr, Isabel Papadimitriou, Joon Sung Park, Chris Piech, Eva Portelance, Christopher Potts, Aditi Raghunathan, Rob Reich, Hongyu Ren, Frieda Rong, Yusuf H. Roohani, Camilo Ruiz, Jack Ryan, Christopher Re, Dorsa Sadigh, Shiori Sagawa, Keshav Santhanam, Andy Shih, Krishnan Srinivasan, Alex Tamkin, Rohan Taori, Armin W. Thomas, Florian Tramèr, Rose E. Wang, and William Yang Wang. On the opportunities and risks of foundation models. arXiv: Learning, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.577, + 0.47, + 0.687 + ], + "angle": 0, + "content": "[3] Kai Chen, Jiaqi Wang, Jiangmiao Pang, Yuhang Cao, Yu Xiong, Xiaoxiao Li, Shuyang Sun, Wansen Feng, Ziwei Liu, Jiarui Xu, Zheng Zhang, Dazhi Cheng, Chenchen Zhu, Tianheng Cheng, Qijie Zhao, Buyu Li, Xin Lu, Rui Zhu, Yue Wu, Jifeng Dai, Jingdong Wang, Jianping Shi, Wanli Ouyang, Chen Change Loy, and Dahua Lin. MMDetection: Open mmlab detection toolbox and benchmark. arXiv preprint arXiv:1906.07155, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.689, + 0.47, + 0.758 + ], + "angle": 0, + "content": "[4] Sandeep Chinchali, Apoorva Sharma, James Harrison, Amine Elhafsi, Daniel Kang, Evgenya Pergament, Eyal Cidon, Sachin Katti, and Marco Pavone. Network offloading policies for cloud robotics: a learning-based approach. Autonomous Robots, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.761, + 0.47, + 0.843 + ], + "angle": 0, + "content": "[5] Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The cityscapes dataset for semantic urban scene understanding. In Proc. of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.846, + 0.47, + 0.901 + ], + "angle": 0, + "content": "[6] Daniel Crankshaw, Xin Wang, Giulio Zhou, Michael J. Franklin, Joseph E. Gonzalez, and Ion Stoica. Clipper: A low-latency online prediction serving system. networked systems design and implementation, 2016. 2" + }, + { + "type": "list", + "bbox": [ + 0.087, + 0.116, + 0.471, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.093, + 0.892, + 0.162 + ], + "angle": 0, + "content": "[7] Sadjad Fouladi, John Emmons, Emre Orbay, Catherine Wu, Riad S. Wahby, and Keith Winstein. Salsify: Low-latency network video through tighter integration between a video codec and a transport protocol. networked systems design and implementation, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.164, + 0.892, + 0.218 + ], + "angle": 0, + "content": "[8] Yarin Gal and Zoubin Ghahramani. Dropout as a bayesian approximation: Representing model uncertainty in deep learning. international conference on machine learning, 2015. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.221, + 0.892, + 0.274 + ], + "angle": 0, + "content": "[9] Yulu Gan, Yan Bai, Yihang Lou, Xianzheng Ma, Renrui Zhang, Nian Shi, and Lin Luo. Decorate the newcomers: Visual domain prompt for continual test time adaptation, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.277, + 0.892, + 0.318 + ], + "angle": 0, + "content": "[10] Yaroslav Ganin and Victor Lempitsky. Unsupervised domain adaptation by backpropagation. international conference on machine learning, 2015. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.32, + 0.892, + 0.375 + ], + "angle": 0, + "content": "[11] Amir Ghodrati, Babak Ehteshami Bejnordi, and Amirhossein Habibian. Frameexit: Conditional early exiting for efficient video recognition. arXiv: Computer Vision and Pattern Recognition, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.377, + 0.892, + 0.418 + ], + "angle": 0, + "content": "[12] Chuan Guo, Geoff Pleiss, Yu Sun, and Kilian Q. Weinberger. On calibration of modern neural networks. international conference on machine learning, 2017. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.42, + 0.892, + 0.461 + ], + "angle": 0, + "content": "[13] Amirhossein Habibian, Davide Abati, Taco S. Cohen, and Babak Ehteshami Bejnordi. Skip-convolutions for efficient video processing. 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.463, + 0.892, + 0.517 + ], + "angle": 0, + "content": "[14] Dan Hendrycks and Thomas Dietterich. Benchmarking neural network robustness to common corruptions and perturbations. Proceedings of the International Conference on Learning Representations, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.519, + 0.892, + 0.588 + ], + "angle": 0, + "content": "[15] Yiping Kang, Johann Hauswald, Cao Gao, Austin Rovinski, Trevor Mudge, Jason Mars, and Lingjia Tang. Neurosurgeon: Collaborative intelligence between the cloud and mobile edge. architectural support for programming languages and operating systems, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.59, + 0.892, + 0.644 + ], + "angle": 0, + "content": "[16] Mehrdad Khani, Pouya Hamadanian, Arash Nasr-Esfahany, and Mohammad Alizadeh. Real-time video inference on edge devices via adaptive model streaming. arXiv: Learning, 2020. 1, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.647, + 0.892, + 0.686 + ], + "angle": 0, + "content": "[17] Dong-Hyun Lee. Pseudo-label: The simple and efficient semi-supervised learning method for deep neural networks. 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.689, + 0.892, + 0.744 + ], + "angle": 0, + "content": "[18] Rui Li, Qianfen Jiao, Wenming Cao, Hau-San Wong, and Si Wu. Model adaptation: Unsupervised domain adaptation without source data. computer vision and pattern recognition, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.746, + 0.892, + 0.801 + ], + "angle": 0, + "content": "[19] Xiaoqi Li, Jiaming Liu, Shizun Wang, Cheng Lyu, Ming Lu, Yurong Chen, Anbang Yao, Yandong Guo, and Shanghang Zhang. Efficient meta-tuning for content-aware neural video delivery, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.803, + 0.892, + 0.858 + ], + "angle": 0, + "content": "[20] Jian Liang, Dapeng Hu, and Jiashi Feng. Do we really need to access the source data? source hypothesis transfer for unsupervised domain adaptation. international conference on machine learning, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.86, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[21] Ser-Nam Lim, Bharath Hariharan, Luming Tang, Menglin Jia, Bor-Chun Chen, Claire Cardie, and Serge Belongie. Visual prompt tuning. 2022. 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.927, + 0.519, + 0.938 + ], + "angle": 0, + "content": "12165" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.16 + ], + "angle": 0, + "content": "[22] Jiaming Liu, Ming Lu, Kaixin Chen, Xiaqi Li, Shizun Wang, Zhaoqing Wang, Enhua Wu, Yurong Chen, Chuang Zhang, and Ming Wu. Overfitting the data: Compact neural video delivery via content-aware feature modulation, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.165, + 0.47, + 0.233 + ], + "angle": 0, + "content": "[23] Pengfei Liu, Weizhe Yuan, Jinlan Fu, Zhengbao Jiang, Hiroaki Hayashi, and Graham Neubig. Pre-train, prompt, and predict: A systematic survey of prompting methods in natural language processing. arXiv: Computation and Language, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.236, + 0.471, + 0.277 + ], + "angle": 0, + "content": "[24] Yuang Liu, Wei Zhang, and Jun Wang. Source-free domain adaptation for semantic segmentation. computer vision and pattern recognition, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.28, + 0.469, + 0.333 + ], + "angle": 0, + "content": "[25] Robert A. Marsden, Felix Wiewel, Mario Döbler, Yang Yang, and Bin Yang. Continual unsupervised domain adaptation for semantic segmentation using a class-specific transfer. 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.337, + 0.469, + 0.378 + ], + "angle": 0, + "content": "[26] Zhongnan Qu, Cong Liu, and Lothar Thiele. Deep partial updating: Towards communication efficient updating for ondevice inference. 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.381, + 0.469, + 0.436 + ], + "angle": 0, + "content": "[27] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2015. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.439, + 0.469, + 0.508 + ], + "angle": 0, + "content": "[28] Christos Sakaridis, Dengxin Dai, and Luc Van Gool. ACDC: The adverse conditions dataset with correspondences for semantic driving scene understanding. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), October 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.511, + 0.469, + 0.566 + ], + "angle": 0, + "content": "[29] Christian Simon, Masoud Faraki, Yi-Hsuan Tsai, Xiang Yu, Samuel Schulter, Yumin Suh, Mehrtash Harandi, and Manmohan Chandraker. On generalizing beyond domains in cross-domain continual learning. 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.569, + 0.469, + 0.61 + ], + "angle": 0, + "content": "[30] Antti Tarvainen and Harri Valpola. Mean teachers are better role models: Weight-averaged consistency targets improve semi-supervised deep learning results. Learning, 2017. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.613, + 0.469, + 0.653 + ], + "angle": 0, + "content": "[31] Eric Tzeng, Judy Hoffman, Kate Saenko, and Trevor Darrell. Adversarial discriminative domain adaptation. computer vision and pattern recognition, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.656, + 0.469, + 0.683 + ], + "angle": 0, + "content": "[32] Laurens Van Der Maaten. Barnes-hut-sne. arXiv preprint arXiv:1301.3342, 2013. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.686, + 0.469, + 0.74 + ], + "angle": 0, + "content": "[33] Riccardo Volpi, Diane Larlus, and Grégory Rogez. Continual adaptation of visual representations via domain randomization and meta-learning. computer vision and pattern recognition, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.744, + 0.469, + 0.785 + ], + "angle": 0, + "content": "[34] Dequan Wang, Evan Shelhamer, Shaoteng Liu, Bruno A. Olshausen, and Trevor Darrell. Tent: Fully test-time adaptation by entropy minimization. Learning, 2021. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.788, + 0.469, + 0.842 + ], + "angle": 0, + "content": "[35] Qin Wang, Olga Fink, Luc Van Gool, and Dengxin Dai. Continual test-time domain adaptation. In Proceedings of Conference on Computer Vision and Pattern Recognition, 2022. 1, 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.846, + 0.469, + 0.9 + ], + "angle": 0, + "content": "[36] Tong Wang, Yousong Zhu, Chaoyang Zhao, Wei Zeng, Jinqiao Wang, and Ming Tang. Adaptive class suppression loss for long-tail object detection. computer vision and pattern recognition, 2021. 8" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.471, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.092, + 0.892, + 0.147 + ], + "angle": 0, + "content": "[37] Zifeng Wang, Zizhao Zhang, Chen-Yu Lee, Han Zhang, Ruoxi Sun, Xiaoqi Ren, Guolong Su, Vincent Perot, Jennifer Dy, and Tomas Pfister. Learning to prompt for continual learning. 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.149, + 0.892, + 0.205 + ], + "angle": 0, + "content": "[38] Jiangchao Yao, Feng Wang, Kunyang Jia, Bo Han, Jingren Zhou, and Hongxia Yang. Device-cloud collaborative learning for recommendation. knowledge discovery and data mining, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.207, + 0.892, + 0.287 + ], + "angle": 0, + "content": "[39] Chongzhi Zhang, Mingyuan Zhang, Shanghang Zhang, Daisheng Jin, Qiang Zhou, Zhongang Cai, Haiyu Zhao, Shuai Yi, Xianglong Liu, and Ziwei Liu. Delving deep into the generalization of vision transformers under distribution shifts. arXiv: Computer Vision and Pattern Recognition, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.29, + 0.892, + 0.345 + ], + "angle": 0, + "content": "[40] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Zwei Liu. Conditional prompt learning for vision-language models. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.347, + 0.892, + 0.388 + ], + "angle": 0, + "content": "[41] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Learning to prompt for vision-language models. International Journal of Computer Vision (IJCV), 2022. 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.092, + 0.892, + 0.388 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.927, + 0.52, + 0.938 + ], + "angle": 0, + "content": "12166" + } + ] +] \ No newline at end of file diff --git a/2023/Cloud-Device Collaborative Adaptation to Continual Changing Environments in the Real-World/071f2817-bbaa-4424-acc0-27135344a577_origin.pdf b/2023/Cloud-Device Collaborative Adaptation to Continual Changing Environments in the Real-World/071f2817-bbaa-4424-acc0-27135344a577_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..a8f6a881fa31eeb39ba5d5154370b7b29937f27f --- /dev/null +++ b/2023/Cloud-Device Collaborative Adaptation to Continual Changing Environments in the Real-World/071f2817-bbaa-4424-acc0-27135344a577_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:03515356d4690554d1cc6b4bb8f5c09b87b09e6178dc2e0987e8c739115b8e86 +size 4954524 diff --git a/2023/Cloud-Device Collaborative Adaptation to Continual Changing Environments in the Real-World/full.md b/2023/Cloud-Device Collaborative Adaptation to Continual Changing Environments in the Real-World/full.md new file mode 100644 index 0000000000000000000000000000000000000000..27a4bdb05e17f08ccdb913d23e61af33252ff679 --- /dev/null +++ b/2023/Cloud-Device Collaborative Adaptation to Continual Changing Environments in the Real-World/full.md @@ -0,0 +1,299 @@ +# Cloud-Device Collaborative Adaptation to Continual Changing Environments in the Real-world + +Yulu Gan $^{1*}$ , Mingjie Pan $^{1*}$ , Rongyu Zhang $^{2}$ , Zijian Ling $^{3}$ , Lingran Zhao $^{1}$ , Jiaming Liu $^{1}$ , Shanghang Zhang $^{1\dagger}$ + +$^{1}$ National Key Laboratory for Multimedia Information Processing, School of Computer Science, Peking University, + +$^{2}$ The Chinese University of Hong Kong, Shenzhen, $^{3}$ Imperial College London + +![](images/204b99c60bcd114a0b5d4892094edba84b4b5ebf619a0e3955fa852653c7ad60.jpg) +(a) The problem and our main idea + +![](images/fe33a44805963d570b10019b9b4ddaace2b9e62a57530e3f15f36c16259e63d9.jpg) +(b) Comparison of our method with others on continuous domain shifts +Figure 1. (a) Models deployed on devices are preferably lightweight. However, device models will suffer from severe performance degradation when facing continual distribution shift data. Our main idea is to improve the continual domain adaptation capability of the device model by performing our proposed Cloud-Device Collaborative Adaptation paradigm. (b) We compare our method with previous works [16,34,35]. Our method surpasses the state-of-the-art approach and exhibits a solid ability when facing continual distribution shifts. + +# Abstract + +When facing changing environments in the real world, the lightweight model on client devices suffers from severe performance drops under distribution shifts. The main limitations of the existing device model lie in (1) unable to update due to the computation limit of the device, (2) the limited generalization ability of the lightweight model. Meanwhile, recent large models have shown strong generalization capability on the cloud while they can not be deployed on client devices due to poor computation constraints. To enable the device model to deal with changing environments, we propose a new learning paradigm of Cloud-Device Collaborative Continual Adaptation, which encourages collaboration between cloud and device and improves the generalization of the device model. Based on this paradigm, we further propose an Uncertainty-based Visual Prompt Adapted (U-VPA) teacher-student model to transfer the generalization capability of the large model on the cloud to the device model. Specifically, we first de + +sign the Uncertainty Guided Sampling (UGS) to screen out challenging data continuously and transmit the most out-of-distribution samples from the device to the cloud. Then we propose a Visual Prompt Learning Strategy with Uncertainty guided updating (VPLU) to specifically deal with the selected samples with more distribution shifts. We transmit the visual prompts to the device and concatenate them with the incoming data to pull the device testing distribution closer to the cloud training distribution. We conduct extensive experiments on two object detection datasets with continually changing environments. Our proposed U-VPA teacher-student framework outperforms previous state-of-the-art test time adaptation and device-cloud collaboration methods. The code and datasets will be released. + +# 1. Introduction + +Real-world usually contains various environmental changes along with continual distribution shifts [35]. People usually deploy economically lightweight models on devices to boost the scalability and practicability. The lightweight model can suffer severe performance degrad + +tion under continual distribution shifts [25, 29, 33, 35]. The main challenges are: (1) The poor computational ability of devices. Due to the properties of device infrastructure, the deployed model can not be updated on time, thus lagging its performance for the real world under distribution shift. (2) The limited generalization ability of the lightweight model. Since lightweight models are of relatively small capacity, they can not handle continually changing environments. In contrast, recent large models that are trained on the cloud server show significant generalization ability [2, 39]. While in industry, these large models can not be directly applied due to the limited infrastructure. + +Therefore, we enable the device model to tackle real-world environmental changing by proposing a Cloud-Device Collaborative Continual Adaptation paradigm, as shown in Fig.1 (a). Previous Cloud-Device Collaborative methods only focus on improving model representation on variance in video frames but neglect the model generalization ability for continually changing data distribution. In our new paradigm, we fully exploit the sufficient knowledge of the large cloud model and transfer the continual generalization ability to the device lightweight model. + +In particular, we design an Uncertainty-based Visual Prompt Adapted (U-VPA) teacher-student model, which consists of an Uncertainty Guided Sampling (UGS) strategy and a Visual Prompt Learning Strategy with Uncertainty Guided Updating (VPLU). Due to the communication bandwidth constraint [11, 13, 19], different with some work targeting filtering the parameters [22, 26], we design the UGS to screen out the most environment-specific samples and decrease the required bandwidth compared with transmitting the whole sequence. To leverage the strong generalization ability of large models, we introduce the VPLU to align source-target domain distribution and transfer the representation of the large teacher model to the light student model. The light student model and visual prompts are then delivered to the device, thus alleviating the continuously changing scenarios in the real world. + +Experimental results show that our method outperforms the state-of-the-art methods on synthetic and real-world distribution shift datasets, as shown in Fig.1 (b). Besides, we can achieve the same performance as the entire data with fewer reflowed data (42% of total data). As another benefit, fewer reflowed data reduce the bandwidth pressure of the uplink. As for the downlink, we can deliver the visual prompts (0.43% of the model's parameters) with almost negligible bandwidth to the device and apply the visual prompt to the input data to improve the performance of the device model by 3.9% in mAP. + +Our contributions can be summarized as follows: + +- We make the first attempt to deal with continually changing scenarios by proposing a Cloud-Device Collaborative Continual Adaptation paradigm, which aims + +to transfer the generalization ability from the large cloud model to the lightweight device model. Our method is a general paradigm that can apply to real-world systems. + +- We design an Uncertainty-based Visual Prompt Adapted (U-VPA) teacher-student model, which consists of UGS and VPLU. We introduce UGS to screen out the most environment-specific samples and decrease the required bandwidth compared with transmitting the whole sequence. We propose a VPLU to align source-target data distribution and transfer the representation of the large teacher model to the lightweight student model. + +- Experiments show that our proposed framework and method surpass other state-of-the-art methods and can effectively improve the continuous domain adaptation capability of the device model. + +# 2. Related work + +Test time adaptation. Test-time adaptation(TTA) aims to get better model generalization performance on target data without access to source data. Feature modulation is used in TENT [34] to optimize BN layers of the input pretrained model by entropy minimization. SHOT [20] takes entropy minimization and a diversity regularizer to achieve information maximization. [18, 24] both take a productive way to enhance model performance on the target domain. + +We differ from TTA in that we consider continuously changing domains. Besides, our model is deployed on the device. There will be restrictions on computing resources and uplink and downlink transmission bandwidth. Considering TTA can effectively use the test data to make updates, we leverage TTA to enable the cloud to improve generalization performance continuously. + +Prompt learning. Prompt is originally from the natural language process (NLP) and aims to adapt the large-scale pre-trained language model to downstream tasks. Recently, several methods [1, 9, 21, 40, 41] are proposed to prove the feasibility of using prompts to adapt the large pre-trained model to downstream tasks in computer vision or vision-language tasks. [1, 21, 37] take prompt as additional input containing a small number of learnable parameters and explore the design of prompts. + +Prompt learning aims at mining the pre-trained model's information. We introduce visual prompt tuning to improve the performance of the device model on the current domain, with a negligible computational overhead in our Cloud-Device Collaboration (DCC) framework. + +Cloud-device collaboration. Cloud-device collaboration aims to bring benefits to both the cloud and the device jointly. Several methods [4, 6, 7, 15] try to unload part or + +![](images/30c87bae3bde82773df112a90c27fb43c35c3e88475bd3911825c0580a8d2b34.jpg) +Figure 2. The whole framework. (a): Testing on the device. On the client device, we select data that need to be sent to cloud by Uncertainty Guided Sampling (UGS). These uncertainty samples are then used in the cloud to update the teacher-student models. The visual prompt trained on the cloud will be delivered to the device through the downlink and applied to the input data $x$ . Then the reformulated input data will be put into the device model for testing. (b) Training on the cloud. We add our proposed visual prompt to the uncertain samples at the pixel level to get the reformulated images. We then put the reformulated images into the teacher-student model. The teacher-student model and visual prompts are optimized by align loss and supervised loss. (c) Visual prompt learning strategy with uncertainty guided sampling (VPLU). We update the visual prompts by uncertainty-based EMA update. The update weight of the visual prompt will be larger when the uncertainty is higher. The student model is updated using the uncertainty samples and the pseudo labels. + +all computation to the cloud to alleviate the computation deficiency on the device. But they do not leverage the collaboration between the cloud and the device. AMS [16] and DCCL [38] have considered the collaboration but they do not take the real-world continual distribution shifts into account. Compared with methods above, our proposed CloudDevice Collaborative Continual Adaptation paradigm can deal with the real-world environment changes. + +In our work, we use Uncertainty Guided Sampling (UGS) and visual prompt for more efficient bandwidth usage. Besides, our model can adapt to different domains during test time. Our teacher model on the cloud and our student model on the device can be improved simultaneously. + +# 3. Method + +# 3.1. Preliminary + +Given a lightweight model $\theta_{device}$ that is deployed on the device, our goal is to perform well on the continually changing target domains $D_{t1}, D_{t2}, \ldots, D_{tT}$ with limited communication bandwidth, where $D_{ti} = \{(x_i^T)\}_{i=1}^{N_t}$ , and $N_t$ repre + +sents the scale of the target domain. The distributions of the target domains are arbitrary since they can change or reoccur over time. We thus introduce a large teacher model $\theta_{tea}$ and lightweight student model $\theta_{stu}$ cloned from $\theta_{device}$ , to boost the generalization ability of $\theta_{device}$ . + +# 3.2. Overall framework + +In our Cloud-Device Collaborative Continual Adaptation paradigm, the key idea is to improve the device model's performance continuously by receiving adapted parameters from the cloud server. So we propose Visual Prompt Adapted (U-VPA) teacher-student model. As shown in Fig. 2, the U-VPA teacher-student model contains Uncertainty Guided Sampling (UGS) and Visual Prompt Learning Strategy with Uncertainty Guided Updating (VPLU). + +When perceiving scene changing or distribution shift, poor computational device intends to get help from a cloud server with powerful computing ability. However, due to the communication bandwidth constraint, we must consider the required bandwidth when concurring the distribution shifts. Therefore, we design a Uncertainty Guided Sam + +pling (UGS) to screen out the most environment-specific samples and decrease the required bandwidth compared with transmitting the whole sequence. After receiving the learned parameters from the cloud to deal with these samples of relatively large distribution shifts, it can adapt to the new scene when inference. + +On the cloud server, due to its strengths in computing, we intend to leverage the strong generalization ability of large models to tackle the received environment-specific and challenging samples. We introduce Visual Prompt Learning Strategy with Uncertainty guided Updating (VPLU) to align source-target domain distribution and transfer the representation of the large teacher model to the light student model. The light student model and visual prompts is then delivered to the device, thus alleviating the continuously changing scenarios in the real-world. + +We will discuss the details of UGS and VPLU in Sec. 3.3.1 and Sec. 3.3.2. + +# 3.3. U-VPA teacher-student model + +# 3.3.1 Uncertainty Guided Sampling(UGS) + +Due to bandwidth limitations, it is not practical for the device to return all data. Therefore, we select images that need to be transmitted based on uncertainty estimation. Inspired by the Dropout methods [8, 12], since we are doing an object detection task, we can obtain $n$ probabilities $p_i(y|x)$ of the model by dropout method, and we calculate the variance of these $n$ prediction $p_i(y|x)$ as uncertainty. + +$$ +V _ {u n c} = \left(\frac {1}{n} \sum_ {i} ^ {n} \left(p _ {i} (y | x) - \mu\right) ^ {2}\right) ^ {\frac {1}{2}} \tag {1} +$$ + +where $p(y|x)$ represents the predictions of the input image $x$ and $\mu$ is the mean. $V_{unc}$ is the variance of $n$ probabilities of the model, which represents the uncertainty of the model for the data. The device model filters the input images $x$ whose uncertainty is larger than a threshold and sends back to the cloud server. + +# 3.3.2 Visual Prompt Learning Strategy with Uncertainty Guided Updating (VPLU). + +We optimize model on the cloud server by using the data transmitted from the device. It then delivers the learned parameters and processed visual prompt to the device model for inference. We thus introduce VPLU with a large teacher model and light student model, to mitigate variant domain distribution and improve the generalization ability of light model. Note that, teacher model is initialized by the source model pre-trained on source domain. It takes source and target domain as input, while student model only trains on the target domain. + +Cross-domain alignment. In order to mitigate distribution shift, we align the features in the source and target domains. Inspired by [10], we apply feature alignment on teacher model to improve its performance on target domain. We use a domain discriminator to classify data in feature space and try to distinguish which domain the data comes from. Besides, we use a gradient reversal layer to achieve a uniform training process. Source domain data is used to supervise teacher model, avoiding performance degradation. + +Cross-model transfer. After improving teacher model's domain-invariant representation ability, we transfer its ability against distribution shifts and generalization to student model via visual prompts and knowledge transferring. + +Visual Prompt Adaptation(VPA) learning. Specifically, inspired by the prompt learning in NLP [23], which can effectively explores the knowledge of source domain, we propose visual prompts as hints to continually transfer teacher model's generalized representation to the student model. + +In particular, visual prompts $\phi$ are learnable parameters. During the forward process of the model, we can obtain the reformulated image $x^{*}$ by adding visual prompts on the input images $x$ at the pixel level. + +$$ +x ^ {*} = x + \phi , \tag {2} +$$ + +In our framework, whether in training on server or inference on device, we all use the reformulated image $x^{*}$ as model's input, instead of origin images $x$ . + +Uncertainty-aware Prompt Updation(U-EMA). Visual prompt acts on the input data and participates in the forward process of both teacher and student models, so it can also benefit from their optimization during training. Moreover, to make the prompt update more stable and consider the current model's uncertainty to data, we update the visual prompt using by uncertainty-based exponential moving average(U-EMA). The weight of the exponential moving average is based on the uncertainty estimation as follows: + +$$ +\phi_ {t} \leftarrow \beta \phi_ {t - 1} + (1 - \beta) \phi_ {t}, \tag {3} +$$ + +Note that $\phi_t$ represents the parameters of the visual prompt that need to be updated, and $\phi_{t - 1}$ indicates the visual prompt parameters of the last training step t. We follow [30] to set $\alpha = 0.999$ . And $\beta = \alpha -V_{unc}$ , where $V_{unc}$ is defined in Eq. (1) as the uncertainty of current data. Higher uncertainty means a larger distribution shift in the current data. To save more domain-related information, the prompt update weight will be larger at this time. That is, the larger the uncertainty, the greater the weight of the prompt update. + +As another essential role of the visual prompt, the processed visual prompt will be sent to the device to improve the adaptability of the device model by applying the visual prompt on input images. + +Knowlegde transfer. Along with visual prompt adapted tuning, we align teacher student feature and adopt teacher + +model to generate pseudo label for student model. It is penalized by Knowledge transfer loss $\mathcal{L}_k$ , to promote the consistency between teacher student model representation. Specifically, The teacher model will generate predictions for the reformulated images $x^{*}$ as pseudo labels. Teacher model has stronger generalization ability, so the generated pseudo labels are of high quality. To further improve the pseudo labels quality, we screen those predictions with confidence higher than the threshold as the input to the student model. Overall, we transfer the knowledge from the teacher to the student by feeding selected images $x^{*}$ and corresponding pseudo labels into the student model for training. + +# 3.4. Training objectives + +The teacher and student models are jointly optimized on the cloud server, using uncertain samples $x_{t}^{i}\in D_{t}$ returned from device. The source model is trained on source pairs of $\{x_s^i,y_s^i\}_{i = 1}^{n_s}\in D_s$ where $x_{s}^{i}\in X_{s}$ and $y_{s}^{i}\in Y_{s}$ are images and labels, respectively. + +For the training process of the teacher model, given the reformulated taregt images $x_{t}^{*}$ , as the Eq. (4) shown, we use adversal alignment loss [31] and supervision loss to optimize the teacher model (including visual prompt $\phi$ ). + +$$ +\begin{array}{l} L _ {t e a} = - \lambda L _ {d} \left(G _ {d} \left(F \left(x _ {s} ^ {*}; \theta_ {t e a}\right); F \left(x _ {t} ^ {*}; \theta_ {t e a}\right)\right), y _ {d}\right) \tag {4} \\ + L _ {D e t} \left(X \left(x _ {s} ^ {*}, \theta_ {t e a}\right), y _ {s}\right), \\ \end{array} +$$ + +Note that the first part is adversal alignment loss [31], and another one is the loss function of Faster RCNN [27] (using labeled data in source domain $D_{s}$ ). $X$ represents the model which predict detection outputs. $L_{d}$ represent domain classifier. $F$ indicates the feature extractor and $G_{d}$ is the domain classifier. $\lambda$ is the weight of alignment loss. + +For the optimization of the student model, we use pseudo labels $y_{t}^{p}$ generated by teacher model as supervision: + +$$ +L _ {s t u} = L _ {D e t} \left(X \left(x _ {t} ^ {*}, \theta_ {s t u}\right), y _ {t} ^ {p}\right), \tag {5} +$$ + +The visual prompt benefited from both $L_{tea}$ and $L_{stu}$ . And we update the visual prompt by uncertainty-aware exponential moving average as Eq. (3) shown. + +# 4. Experiments + +In Sec.4.1, we give the details of the dataset and experiment setting. We evaluate the performance of CCA on continual changing scenarios in Sec.4.2. We give discussion on our motivations and ablation study in Sec.4.3, Sec.4.4. + +# 4.1. Experimental setup + +# 4.1.1 Datasets and settings + +We evaluate our proposed method on two test time adaptation benchmark tasks for object detection: Cityscapes-to-Cityscape-C and Cityscapes-to-ACDC-Detection dataset. + +Cityscape-C is originally created to benchmark robustness tasks [14]. We select the five most relevant corruptions to the autonomous driving scenario (including brightness, motion_blur, rain, fog and snow). Each corruption dataset contains five levels of severity. The corruptions are applied to images from the validation set of the clean Cityscapes dataset [5]. There are 2,500 images in five corruption datasets. + +For Cityscapes-to-Cityscapes-C, we follow [3] to train the pre-trained model on the source domain (Cityscapes). Specifically, we use the SGD optimizer with a learning rate of 1e-2 as the official implementation to train the source model. + +ACDC-Detection is created from the Adverse Conditions Dataset (ACDC) [28]. The ACDC dataset shares the same classes with Cityscapes and is collected in various adverse visual conditions: Fog, Night, Rain and Snow. We convert the segmented labels into labels for object detection. We follow CoTTA [35] to set our task. We use 400 unlabeled images from each adverse condition for the adaptation. To mimic the scenario in real life where a similar environment might be revisited and to evaluate the continuous generalization capability of our methods, we repeat the same sequence group (of the three conditions) 10 times (i.e., in total 30: $\mathrm{Fog} \to \mathrm{Rain} \to \mathrm{Snow} \to \mathrm{Fog}\ldots$ ). This also provides an evaluation of the adaptation performance in the long term. + +For the Cityscapes-to-ACDC-Detection task, we use the same way to train the pre-trained model on the source domain (Cityscapes) as the Cityscapes-to-Cityscapes-C task. + +# 4.1.2 Implementation Details + +In this paper, all experiments are conducted with PyTorch. For the Cityscapes-to-Cityscapes-C task and Cityscapes-to-ACDC-Detection task, the batch size is set to 8. We follow [3] to train the pre-trained model on the source domain (Cityscapes). The number of samples used for uncertainty estimation is set to 10 in the experiment. + +We use FasterRCNN with the backbone of ResNet-101 as the large teacher model on the cloud. And we use FasterRCNN with the backbone of ResNet-18 as the lightweight student model (same as the device model). The above large and light models are initialized with corresponding pretrained models on the source domain (Cityscapes dataset). For the UGS module, we set the uncertainty threshold to 0.008, which filters out $42\%$ samples in the Cityscape-C dataset. For the VPA module, We set the visual prompts with size $3 \times 200 \times 200$ , which can be added to the left-up space of the image. The student model's parameters and visual prompts (about $0.4\%$ of the student model's parameters) will be transmitted to the device. + +We use $\mathrm{mAP} @ 0.5(\%)$ as the evaluation metric. The + +Table 1. Continual generalization capability on Cityscapes-to-Cityscapes-C. Object detection results (mAP@0.5 in %) on the Cityscapes-to-Cityscapes-C online continual test-time adaptation task. Gain(%) means the improvement of our method compared with Source-only. We evaluate the four test conditions continually for ten times to evaluate the long-term adaptation performance. All results are evaluated on the FasterRCNN architecture with the largest corruption severity level 5. Our approach surpasses the SOTA method and exhibits significant continual generalization and anti-forgetting abilities. + +
Timet
Round1510All
ConditionFogMotionRainSnowBrightnessFogMotionRainSnowBrightnessFogMotionRainSnowBrightnessMeanGain
Source-only [27]24.79.522.41.329.024.79.522.41.329.024.79.522.41.329.017.4/
TENT-continual [34]24.89.222.51.429.420.06.419.80.823.79.03.810.70.511.413.0-4.4
CoTTA [35]25.110.122.41.328.921.89.520.74.228.29.33.910.13.317.514.7-2.7
Pseudo-Label [17]27.317.224.76.933.730.120.126.98.036.531.020.927.77.238.224.1+6.7
AMS [16]28.417.726.07.634.731.820.428.27.738.034.221.428.97.039.425.1+7.7
Ours(proposed)28.017.826.810.036.537.524.235.514.244.044.027.340.915.150.131.0+13.6
+ +Table 2. Continuous generalization capability on Cityscapes-to-ACDC-Detection. Object detection results (mAP@0.5 in %) on the Cityscapes-to-ACDC-Detection online continual test-time adaptation task. Gain(%) means the improvement of our method compared with Source-only. We evaluate the four test conditions continually for ten times to evaluate the long-term adaptation performance. All results are evaluated on the FasterRCNN architecture. Our results outperform the SOTA method, thus verifying the effectiveness of our method on the real distribution shift. + +
Timet
Round14710All
ConditionFogNightRainSnowFogNightRainSnowFogNightRainSnowFogNightRainSnowMeanGain
Source-only [27]20.57.414.314.220.57.414.314.220.57.414.314.220.57.414.314.214.1/
TENT-continual [34]20.57.514.313.920.07.213.413.018.66.612.312.116.66.011.110.912.8-2.3
CoTTA [35]20.77.514.714.320.97.414.714.420.47.313.814.416.36.411.712.213.8-0.3
Pseudo-Label [17]20.98.715.517.623.410.316.918.924.310.917.819.525.111.517.520.017.6+3.5
AMS [16]21.69.016.017.223.210.717.219.124.711.417.219.926.111.717.820.517.8+3.7
Ours(proposed)21.19.315.717.224.711.817.820.226.312.018.520.626.112.318.020.818.5+4.4
+ +overall process of test-time adaptation strictly follows CoTTA [35]. The process first tests the coming batch of data then train with it. We do batch-level updates on the cloud model, and the performance on the device model is taken as the average of the results of all the data. + +Baselines. Our method is compared with multiple baselines, including those for test time adaptation (TTA), continual test time adaptation (CTTA) and cloud-end collaboration methods as follows: Source-only [27]. Test the model performance on the different target domains using the pre-trained source model (training on the source domain). Pseudo-label [17]. Use pseudo label generated by the teacher model to supervise the student model. Tent-continual [34]. Tent is an effective method for the test-time adaptation (TTA). Tent fine-tune the batch normalization layer during training to adapt the model to the target domain. CoTTA [35]. CoTTA is an attractive approach for the continual test-time adaptation (CTTA). CoTTA mitigate error accumulation and catastrophic forgetting with weighted averaging, data augmented averaging, and random recovery neurons. AMS [16]. AMS is an cloud engagement approach. To solve the resource limitation of the device model, AMS do distillation in the cloud, and then transmit the model's parameters to the device in the cloud, the teacher network will not update. + +# 4.2. Results and Analysis + +Synthetic continual distribution shift. We evaluate our proposed method on the synthetic continual distribution shift dataset (Cityscapes $\rightarrow$ Cityscapes-C task). As shown in Table 1, TENT-continual [34] has a decreasing performance when facing continuously changing domains. Because TENT does not consider continuous distribution shifts, it uses self-entropy to update the bn layer, which is prone to catastrophic forgetting when facing constant distribution shifts. CoTTA [35] is a method designed specifically for continuous test time adaptation. CoTTA applies weighted averaging, data-augmented averaging, and random recovery neurons to overcome catastrophic forgetting and error accumulation. However, when doing the object detection task on the continuously changing domains, the model's performance still decreases, indicating that solutions used by CoTTA are insufficient against catastrophic forgetting. We believe this is caused by the poor generalization abilities of the lightweight device model. Therefore, it is reasonable to introduce the cloud for the Cloud-Device collaborative continual adaptation paradigm. + +Both Pseudo-Label and AMS [16] use the Cloud-Device collaboration framework, which transfers the teacher's knowledge to the student on the cloud. The generalization of the large model provides continuous guidance for the student model, which avoids the problem of catastrophic for + +![](images/ee175a1cf1b4809dbd6315810cfed251a4429d8ad11b0bc8c8365c6fc6e9a4b0.jpg) +(a) + +![](images/921a84d4117dc510009e17c94cf4f4d6258c468eace56ec72089c3acc77942d2.jpg) + +![](images/fd4b8e6533c44311937a2084116cea2f2d5006fd96df6bfd797526dabe8e023c.jpg) +Figure 3. (a)-(b): Statistical distribution of confidence and uncertainty. (a): distribution of confidence. (b): distribution of uncertainty. Comparing (a) and (b), the uncertainty value is more stable when facing continuous distribution shift, while the confidence value fluctuates significantly with continuous distribution shift. (c)-(e): T-SNE [32] visualization of samples in different domains. (c): all samples. (d): samples filtered by confidence. (e): samples filtered by UGS. Samples filtered by UGS are more environment-specific. They are more worthy of being transmitted to the cloud server for optimizing the teacher-student models. + +getting and error accumulation. (Pseudo-Label and AMS are $6.7\%$ and $7.7\%$ higher than Source-only). + +However, pseudo label and AMS do not consider updating the teacher model. Our U-VPA framework jointly optimizes the teacher and student model, which makes our method surpass Source-only by $13.6\%$ , outperforming all the above methods. This validates that our Cloud-Device collaboration paradigm can effectively improve the continuous domain adaptation ability of the device. + +Real-world continual distribution shift. We also evaluate our proposed method on the real-world continual distribution shift dataset (Cityscapes $\rightarrow$ ACDC-Detection task). As shown in Table 2, our method improves by $4.4\%$ over the source-only method. By comparing the performance change of the same domain in 10 rounds, we can find that our method does not have the problem of catastrophic forgetting. For example, the TENT-continual method faces the problem of catastrophic forgetting. The $\mathrm{mAP}(\%)$ drops in the fourth, seventh, and tenth rounds compared to the first round when facing the night domain, specifically $-0.2\%$ , $-0.5\%$ , $-4.1\%$ . In contrast, our method leads to a continuous performance improvement (e.g., our method's $\mathrm{mAP}(\%)$ increases in the fourth, seventh, and tenth rounds, respectively, compared to the first round when facing the Night domain, specifically, $+2.5\%$ , $+0.2\%$ , $+0.3\%$ ). + +![](images/136e8a9008aa01371940180382e53f496479df865cae23273717523acae80955.jpg) +Figure 4. Detection results (mAP@0.5) of both teacher and student model. Solid lines are results of models using VPA under the brightness domain. Teacher and student models have jointly improved in continuous domain adaptation. Dotted lines are results without using VPA. The teacher quickly reaches the bottleneck, and student's improvement is slow. The analysis above illustrates that the proposed U-VPA framework enables the teacher-student model jointly improve the continuous domain adaption capability. + +# 4.3. Discussion + +Teacher-student joint optimization with VPA. Previous approaches (e.g., AMS) focus on improving the student model. However, the student model's performance is correlated with the teacher model. Our method enables the teacher and student models to improve performance together by using the visual prompt as a bridge and using the temporal domain information brought by the visual prompt. + +The optimization of the teacher model comes from two modules: (1) feature alignment, (2) visual prompt adaptation(VPA). Our experiments have proved that although both of the above can bring about gains, visual prompt adaptation is the main reason for the success of our framework. + +As shown in Fig. 4, if only feature alignment is used, the performance of the teacher model reaches the bottleneck after the second round, and the student's optimization will be very slow (dotted lines in the figure). + +While our method benefits from VPA, the teacher and student models are jointly improved, demonstrating that our proposed U-VPA teacher-student model can optimize the student-teacher model together. + +Uncertainty Guided Sampling (UGS) strategy. We demonstrate in Table 5 that using only a small amount of reflowed data can outperform the results using total data by the UGS strategy. Our model's performance using the UGS strategy to select part of the data to transmit into the cloud for training is $31\%$ . Compared to the model performance of $28.3\%$ obtained with the total data, there is a $2.7\%$ improvement. The amount of data using the random picking strategy is the same as that with the UGS strategy. And the UGS strategy has a better performance than random picking. The above shows the effectiveness of the UGS strategy. + +We further reveal the advantages of uncertainty when the + +Table 3. Ablation study. We conduct ablation studies for proposed Pseudo-Label, VPA, UGS, and U-EMA. All the experiments are done on Cityscapes-to-Cityscapes-C, and the evaluation metric is mAP (\%). The experimental result shows that all these proposed modules positively impact the performance improvement of the devices model when facing continuous domains. + +
Pseudo-LabelVPAUGSU-EMAFogMotionRainSnowBrightnessMeanGain
124.79.522.41.329.017.4/
2✓31.020.927.77.238.224.1+6.7
3✓✓37.122.232.711.243.127.4+10.0
4✓✓✓38.324.836.512.346.728.3+10.9
5✓✓✓✓44.027.340.915.150.131.0+13.6
+ +Table 4. Comparing the difference between selecting samples with confidence and uncertainty. Experiments show that better performance is obtained by choosing the reflowed data model through uncertainty estimation. + +
FogMotionRainSnowBrightnessMean
Confidence39.624.937.214.345.127.4
Uncertainty44.027.340.915.150.131.0
+ +Table 5. Effect of uplink transmission of full and partial data on model's performance. Our method returns partial data ( $\tilde{42\%}$ ) for training but achieves better performance than using full data. This indicates that our UGS strategy reduces the data transmission communication cost and improves the model performance. + +
25%50%75%100%
Random23.526.127.928.3
UGS28.231.031.2/
Gain+4.7+4.9+3.3/
+ +distribution shifts exist. We find uncertainty is more reliable compared with the confidence score. Regarding Fig.3 (a)-(b), in the presence of distribution shifts, confidence distribution varies significantly, while uncertainty distribution is balanced. Therefore, the samples selected using confidence will face the problem of long-tail distribution, which is not conducive to cloud model training [36]. Uncertainty can more accurately reflect the difficulty of identifying samples by device without using the training of cloud model and discover the samples that are more worthy of attention. + +Besides, we explore the properties of the data filtered by different filter strategies. In Fig.3 (c)-(e), each point represents a sample filtered by the corresponding strategy. We find that samples selected by uncertainty are environmental-specific, so using UGS can filter out more valuable data. + +# 4.4. Ablation study + +Effect of the Uncertainty Guided Sampling (UGS). Our proposed UGS strategy not only reduces the communication cost of the uplink from the device to the cloud but also improves performance by $0.9\%$ , as shown in Table 3. Furthermore, as shown in Table 5, the model using UGS outperforms baseline with only $42\%$ data for train, which significantly improves the compression ratio of the video. + +Effect of the Visual Prompt Adapt (VPA). Our pro + +posed VPA teacher-student model uses visual prompts to make the student-teacher model jointly improve performance. As shown in Table 3, compared to the pure pseudo label method, the usage of VPA can increase performance by $3.3\%$ . Furthermore, we discussed in Sec. 4.3 that VPA could significantly increase the upper limit of the teacher-student framework from challenging continuously changing distribution shift scenarios. The combination of VPA and other modules can further improve performance steadily. + +Effect of the Uncertainty-aware prompt EMA (U-EMA) update mechanism. The U-EMA update mechanism we proposed in training the cloud model makes the cloud model pay more attention to the images worthy of attention by dynamically adjusting the prompt weights with uncertainty. As shown in Table 3, U-EMA can bring a $2.7\%$ performance improvement. + +# 5. Conclusion & Limitations + +We propose a Cloud-Device Collaborative Continual Adaptation paradigm to deal with continual changing environments on devices. We design an Uncertainty Guided Sampling (UGS) strategy to transmit the most out-of-distribution samples from device to cloud. Besides, we design a Visual Prompt Learning Strategy with Uncertainty Guided Updating (VPLU) to transfer the generalization capability of large model on the cloud to the device model. Experiments of our method on a wide range of continually changing environments show an improvement of $4.4 - 13.6\%$ in mAP@0.5 on the object detection task. Our proposed method offers flexibility in its application to most existing cloud-device systems. However, it is worth noting that our experiments were conducted using simulated datasets rather than real systems. Moving forward, an interesting direction for future research would be to test our method on a real system with specific bandwidth limitations, a specific cloud, and a specific device. By taking into account the hardware characteristics of the system, there should have rooms to improve its performance further. + +Acknowledgement. This research was partly supported by the foundation of the National Key R&D Program of China (2022ZD0116305). We thank Xiaoqi Li for the insightful discussions and for helping with writing. + +# References + +[1] Hyojin Bahng, Ali Jahanian, Swami Sankaranarayanan, and Phillip Isola. Exploring visual prompts for adapting large-scale models. 2022. 2 +[2] Rishi Bommasani, Drew A. Hudson, Ehsan Adeli, Russ B. Altman, Simran Arora, Sydney von Arx, Michael S. Bernstein, Jeannette Bohg, Antoine Bosselut, Emma Brunskill, Erik Brynjolfsson, Shyamal Buch, Dallas Card, Rodrigo Castellon, Niladri S. Chatterji, Annie Chen, Kathleen Creel, Jared Davis, Dora Demszky, Chris Donahue, Moussa Doumbouya, Esin Durmus, Stefano Ermon, John Etchemendy, Kawin Ethayarajh, Li Fei-Fei, Chelsea Finn, Trevor Gale, Lauren Gillespie, Karan Goel, Noah D. Goodman, Shelby Grossman, Neel Guha, Tatsunori Hashimoto, Peter Henderson, John Hewitt, Daniel E. Ho, Jenny Hong, Kyle Hsu, Jing Huang, Thomas Icard, Sahil Jain, Dan Jurafsky, Pratyusha Kalluri, Siddharth Karamcheti, Geoff Keeling, Fereshte Khani, Omar Khattab, Pang Wei Koh, Mark Krass, Ranjay Krishna, Rohith Kuditipudi, Ananya Kumar, Faisal Ladhak, Mina Lee, Tony Lee, Jure Leskovec, Isabelle Levent, Xiang Lisa Li, Xuechen Li, Tengyu Ma, Ali Ahmad Malik, Christopher D. Manning, Suvir Mirchandani, Eric Mitchell, Zanele Munyikwa, Suraj Nair, Avanika Narayan, Deepak Narayanan, Ben Newman, Allen Nie, Juan Carlos Niebles, Hamed Nilforoshan, Julian Nyarko, Giray Ogut, Laurel Orr, Isabel Papadimitriou, Joon Sung Park, Chris Piech, Eva Portelance, Christopher Potts, Aditi Raghunathan, Rob Reich, Hongyu Ren, Frieda Rong, Yusuf H. Roohani, Camilo Ruiz, Jack Ryan, Christopher Re, Dorsa Sadigh, Shiori Sagawa, Keshav Santhanam, Andy Shih, Krishnan Srinivasan, Alex Tamkin, Rohan Taori, Armin W. Thomas, Florian Tramèr, Rose E. Wang, and William Yang Wang. On the opportunities and risks of foundation models. arXiv: Learning, 2021. 2 +[3] Kai Chen, Jiaqi Wang, Jiangmiao Pang, Yuhang Cao, Yu Xiong, Xiaoxiao Li, Shuyang Sun, Wansen Feng, Ziwei Liu, Jiarui Xu, Zheng Zhang, Dazhi Cheng, Chenchen Zhu, Tianheng Cheng, Qijie Zhao, Buyu Li, Xin Lu, Rui Zhu, Yue Wu, Jifeng Dai, Jingdong Wang, Jianping Shi, Wanli Ouyang, Chen Change Loy, and Dahua Lin. MMDetection: Open mmlab detection toolbox and benchmark. arXiv preprint arXiv:1906.07155, 2019. 5 +[4] Sandeep Chinchali, Apoorva Sharma, James Harrison, Amine Elhafsi, Daniel Kang, Evgenya Pergament, Eyal Cidon, Sachin Katti, and Marco Pavone. Network offloading policies for cloud robotics: a learning-based approach. Autonomous Robots, 2019. 2 +[5] Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The cityscapes dataset for semantic urban scene understanding. In Proc. of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 5 +[6] Daniel Crankshaw, Xin Wang, Giulio Zhou, Michael J. Franklin, Joseph E. Gonzalez, and Ion Stoica. Clipper: A low-latency online prediction serving system. networked systems design and implementation, 2016. 2 + +[7] Sadjad Fouladi, John Emmons, Emre Orbay, Catherine Wu, Riad S. Wahby, and Keith Winstein. Salsify: Low-latency network video through tighter integration between a video codec and a transport protocol. networked systems design and implementation, 2018. 2 +[8] Yarin Gal and Zoubin Ghahramani. Dropout as a bayesian approximation: Representing model uncertainty in deep learning. international conference on machine learning, 2015. 4 +[9] Yulu Gan, Yan Bai, Yihang Lou, Xianzheng Ma, Renrui Zhang, Nian Shi, and Lin Luo. Decorate the newcomers: Visual domain prompt for continual test time adaptation, 2023. 2 +[10] Yaroslav Ganin and Victor Lempitsky. Unsupervised domain adaptation by backpropagation. international conference on machine learning, 2015. 4 +[11] Amir Ghodrati, Babak Ehteshami Bejnordi, and Amirhossein Habibian. Frameexit: Conditional early exiting for efficient video recognition. arXiv: Computer Vision and Pattern Recognition, 2021. 2 +[12] Chuan Guo, Geoff Pleiss, Yu Sun, and Kilian Q. Weinberger. On calibration of modern neural networks. international conference on machine learning, 2017. 4 +[13] Amirhossein Habibian, Davide Abati, Taco S. Cohen, and Babak Ehteshami Bejnordi. Skip-convolutions for efficient video processing. 2021. 2 +[14] Dan Hendrycks and Thomas Dietterich. Benchmarking neural network robustness to common corruptions and perturbations. Proceedings of the International Conference on Learning Representations, 2019. 5 +[15] Yiping Kang, Johann Hauswald, Cao Gao, Austin Rovinski, Trevor Mudge, Jason Mars, and Lingjia Tang. Neurosurgeon: Collaborative intelligence between the cloud and mobile edge. architectural support for programming languages and operating systems, 2017. 2 +[16] Mehrdad Khani, Pouya Hamadanian, Arash Nasr-Esfahany, and Mohammad Alizadeh. Real-time video inference on edge devices via adaptive model streaming. arXiv: Learning, 2020. 1, 3, 6 +[17] Dong-Hyun Lee. Pseudo-label: The simple and efficient semi-supervised learning method for deep neural networks. 2022. 6 +[18] Rui Li, Qianfen Jiao, Wenming Cao, Hau-San Wong, and Si Wu. Model adaptation: Unsupervised domain adaptation without source data. computer vision and pattern recognition, 2020. 2 +[19] Xiaoqi Li, Jiaming Liu, Shizun Wang, Cheng Lyu, Ming Lu, Yurong Chen, Anbang Yao, Yandong Guo, and Shanghang Zhang. Efficient meta-tuning for content-aware neural video delivery, 2022. 2 +[20] Jian Liang, Dapeng Hu, and Jiashi Feng. Do we really need to access the source data? source hypothesis transfer for unsupervised domain adaptation. international conference on machine learning, 2020. 2 +[21] Ser-Nam Lim, Bharath Hariharan, Luming Tang, Menglin Jia, Bor-Chun Chen, Claire Cardie, and Serge Belongie. Visual prompt tuning. 2022. 2 + +[22] Jiaming Liu, Ming Lu, Kaixin Chen, Xiaqi Li, Shizun Wang, Zhaoqing Wang, Enhua Wu, Yurong Chen, Chuang Zhang, and Ming Wu. Overfitting the data: Compact neural video delivery via content-aware feature modulation, 2021. 2 +[23] Pengfei Liu, Weizhe Yuan, Jinlan Fu, Zhengbao Jiang, Hiroaki Hayashi, and Graham Neubig. Pre-train, prompt, and predict: A systematic survey of prompting methods in natural language processing. arXiv: Computation and Language, 2021. 4 +[24] Yuang Liu, Wei Zhang, and Jun Wang. Source-free domain adaptation for semantic segmentation. computer vision and pattern recognition, 2021. +[25] Robert A. Marsden, Felix Wiewel, Mario Döbler, Yang Yang, and Bin Yang. Continual unsupervised domain adaptation for semantic segmentation using a class-specific transfer. 2022. 2 +[26] Zhongnan Qu, Cong Liu, and Lothar Thiele. Deep partial updating: Towards communication efficient updating for ondevice inference. 2020. 2 +[27] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2015. 5, 6 +[28] Christos Sakaridis, Dengxin Dai, and Luc Van Gool. ACDC: The adverse conditions dataset with correspondences for semantic driving scene understanding. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), October 2021. 5 +[29] Christian Simon, Masoud Faraki, Yi-Hsuan Tsai, Xiang Yu, Samuel Schulter, Yumin Suh, Mehrtash Harandi, and Manmohan Chandraker. On generalizing beyond domains in cross-domain continual learning. 2022. 2 +[30] Antti Tarvainen and Harri Valpola. Mean teachers are better role models: Weight-averaged consistency targets improve semi-supervised deep learning results. Learning, 2017. 4 +[31] Eric Tzeng, Judy Hoffman, Kate Saenko, and Trevor Darrell. Adversarial discriminative domain adaptation. computer vision and pattern recognition, 2017. 5 +[32] Laurens Van Der Maaten. Barnes-hut-sne. arXiv preprint arXiv:1301.3342, 2013. 7 +[33] Riccardo Volpi, Diane Larlus, and Grégory Rogez. Continual adaptation of visual representations via domain randomization and meta-learning. computer vision and pattern recognition, 2020. 2 +[34] Dequan Wang, Evan Shelhamer, Shaoteng Liu, Bruno A. Olshausen, and Trevor Darrell. Tent: Fully test-time adaptation by entropy minimization. Learning, 2021. 1, 2, 6 +[35] Qin Wang, Olga Fink, Luc Van Gool, and Dengxin Dai. Continual test-time domain adaptation. In Proceedings of Conference on Computer Vision and Pattern Recognition, 2022. 1, 2, 5, 6 +[36] Tong Wang, Yousong Zhu, Chaoyang Zhao, Wei Zeng, Jinqiao Wang, and Ming Tang. Adaptive class suppression loss for long-tail object detection. computer vision and pattern recognition, 2021. 8 + +[37] Zifeng Wang, Zizhao Zhang, Chen-Yu Lee, Han Zhang, Ruoxi Sun, Xiaoqi Ren, Guolong Su, Vincent Perot, Jennifer Dy, and Tomas Pfister. Learning to prompt for continual learning. 2022. 2 +[38] Jiangchao Yao, Feng Wang, Kunyang Jia, Bo Han, Jingren Zhou, and Hongxia Yang. Device-cloud collaborative learning for recommendation. knowledge discovery and data mining, 2021. 3 +[39] Chongzhi Zhang, Mingyuan Zhang, Shanghang Zhang, Daisheng Jin, Qiang Zhou, Zhongang Cai, Haiyu Zhao, Shuai Yi, Xianglong Liu, and Ziwei Liu. Delving deep into the generalization of vision transformers under distribution shifts. arXiv: Computer Vision and Pattern Recognition, 2021. 2 +[40] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Zwei Liu. Conditional prompt learning for vision-language models. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2 +[41] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Learning to prompt for vision-language models. International Journal of Computer Vision (IJCV), 2022. 2 \ No newline at end of file diff --git a/2023/Cloud-Device Collaborative Adaptation to Continual Changing Environments in the Real-World/images.zip b/2023/Cloud-Device Collaborative Adaptation to Continual Changing Environments in the Real-World/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..6e9ba805e38926c172b85c6d6afeae85140067ad --- /dev/null +++ b/2023/Cloud-Device Collaborative Adaptation to Continual Changing Environments in the Real-World/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:66be567566ceaa3174cf70d4a022714492d6c37736e66aa382a4d1772aacc531 +size 443888 diff --git a/2023/Cloud-Device Collaborative Adaptation to Continual Changing Environments in the Real-World/layout.json b/2023/Cloud-Device Collaborative Adaptation to Continual Changing Environments in the Real-World/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..9a7a327e44b7e54f1e6025f24757486c9f7257d4 --- /dev/null +++ b/2023/Cloud-Device Collaborative Adaptation to Continual Changing Environments in the Real-World/layout.json @@ -0,0 +1,7969 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 58, + 102, + 536, + 138 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 102, + 536, + 138 + ], + "spans": [ + { + "bbox": [ + 58, + 102, + 536, + 138 + ], + "type": "text", + "content": "Cloud-Device Collaborative Adaptation to Continual Changing Environments in the Real-world" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 158, + 160, + 434, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 158, + 160, + 434, + 190 + ], + "spans": [ + { + "bbox": [ + 158, + 160, + 434, + 190 + ], + "type": "text", + "content": "Yulu Gan" + }, + { + "bbox": [ + 158, + 160, + 434, + 190 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 158, + 160, + 434, + 190 + ], + "type": "text", + "content": ", Mingjie Pan" + }, + { + "bbox": [ + 158, + 160, + 434, + 190 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 158, + 160, + 434, + 190 + ], + "type": "text", + "content": ", Rongyu Zhang" + }, + { + "bbox": [ + 158, + 160, + 434, + 190 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 158, + 160, + 434, + 190 + ], + "type": "text", + "content": ", Zijian Ling" + }, + { + "bbox": [ + 158, + 160, + 434, + 190 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 158, + 160, + 434, + 190 + ], + "type": "text", + "content": ", Lingran Zhao" + }, + { + "bbox": [ + 158, + 160, + 434, + 190 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 158, + 160, + 434, + 190 + ], + "type": "text", + "content": ", Jiaming Liu" + }, + { + "bbox": [ + 158, + 160, + 434, + 190 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 158, + 160, + 434, + 190 + ], + "type": "text", + "content": ", Shanghang Zhang" + }, + { + "bbox": [ + 158, + 160, + 434, + 190 + ], + "type": "inline_equation", + "content": "^{1\\dagger}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 64, + 190, + 529, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 190, + 529, + 204 + ], + "spans": [ + { + "bbox": [ + 64, + 190, + 529, + 204 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 64, + 190, + 529, + 204 + ], + "type": "text", + "content": "National Key Laboratory for Multimedia Information Processing, School of Computer Science, Peking University," + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 203, + 455, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 203, + 455, + 217 + ], + "spans": [ + { + "bbox": [ + 138, + 203, + 455, + 217 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 138, + 203, + 455, + 217 + ], + "type": "text", + "content": "The Chinese University of Hong Kong, Shenzhen, " + }, + { + "bbox": [ + 138, + 203, + 455, + 217 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 138, + 203, + 455, + 217 + ], + "type": "text", + "content": "Imperial College London" + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 57, + 240, + 282, + 355 + ], + "blocks": [ + { + "bbox": [ + 57, + 240, + 282, + 355 + ], + "lines": [ + { + "bbox": [ + 57, + 240, + 282, + 355 + ], + "spans": [ + { + "bbox": [ + 57, + 240, + 282, + 355 + ], + "type": "image", + "image_path": "204b99c60bcd114a0b5d4892094edba84b4b5ebf619a0e3955fa852653c7ad60.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 110, + 357, + 219, + 367 + ], + "lines": [ + { + "bbox": [ + 110, + 357, + 219, + 367 + ], + "spans": [ + { + "bbox": [ + 110, + 357, + 219, + 367 + ], + "type": "text", + "content": "(a) The problem and our main idea" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 308, + 251, + 531, + 355 + ], + "blocks": [ + { + "bbox": [ + 308, + 251, + 531, + 355 + ], + "lines": [ + { + "bbox": [ + 308, + 251, + 531, + 355 + ], + "spans": [ + { + "bbox": [ + 308, + 251, + 531, + 355 + ], + "type": "image", + "image_path": "fe33a44805963d570b10019b9b4ddaace2b9e62a57530e3f15f36c16259e63d9.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 310, + 357, + 530, + 367 + ], + "lines": [ + { + "bbox": [ + 310, + 357, + 530, + 367 + ], + "spans": [ + { + "bbox": [ + 310, + 357, + 530, + 367 + ], + "type": "text", + "content": "(b) Comparison of our method with others on continuous domain shifts" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 45, + 387, + 546, + 431 + ], + "lines": [ + { + "bbox": [ + 45, + 387, + 546, + 431 + ], + "spans": [ + { + "bbox": [ + 45, + 387, + 546, + 431 + ], + "type": "text", + "content": "Figure 1. (a) Models deployed on devices are preferably lightweight. However, device models will suffer from severe performance degradation when facing continual distribution shift data. Our main idea is to improve the continual domain adaptation capability of the device model by performing our proposed Cloud-Device Collaborative Adaptation paradigm. (b) We compare our method with previous works [16,34,35]. Our method surpasses the state-of-the-art approach and exhibits a solid ability when facing continual distribution shifts." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 143, + 444, + 192, + 456 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 444, + 192, + 456 + ], + "spans": [ + { + "bbox": [ + 143, + 444, + 192, + 456 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 469, + 290, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 469, + 290, + 685 + ], + "spans": [ + { + "bbox": [ + 46, + 469, + 290, + 685 + ], + "type": "text", + "content": "When facing changing environments in the real world, the lightweight model on client devices suffers from severe performance drops under distribution shifts. The main limitations of the existing device model lie in (1) unable to update due to the computation limit of the device, (2) the limited generalization ability of the lightweight model. Meanwhile, recent large models have shown strong generalization capability on the cloud while they can not be deployed on client devices due to poor computation constraints. To enable the device model to deal with changing environments, we propose a new learning paradigm of Cloud-Device Collaborative Continual Adaptation, which encourages collaboration between cloud and device and improves the generalization of the device model. Based on this paradigm, we further propose an Uncertainty-based Visual Prompt Adapted (U-VPA) teacher-student model to transfer the generalization capability of the large model on the cloud to the device model. Specifically, we first de" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 445, + 546, + 612 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 445, + 546, + 612 + ], + "spans": [ + { + "bbox": [ + 306, + 445, + 546, + 612 + ], + "type": "text", + "content": "sign the Uncertainty Guided Sampling (UGS) to screen out challenging data continuously and transmit the most out-of-distribution samples from the device to the cloud. Then we propose a Visual Prompt Learning Strategy with Uncertainty guided updating (VPLU) to specifically deal with the selected samples with more distribution shifts. We transmit the visual prompts to the device and concatenate them with the incoming data to pull the device testing distribution closer to the cloud training distribution. We conduct extensive experiments on two object detection datasets with continually changing environments. Our proposed U-VPA teacher-student framework outperforms previous state-of-the-art test time adaptation and device-cloud collaboration methods. The code and datasets will be released." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 634, + 386, + 645 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 634, + 386, + 645 + ], + "spans": [ + { + "bbox": [ + 306, + 634, + 386, + 645 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 654, + 546, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 546, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 546, + 715 + ], + "type": "text", + "content": "Real-world usually contains various environmental changes along with continual distribution shifts [35]. People usually deploy economically lightweight models on devices to boost the scalability and practicability. The lightweight model can suffer severe performance degrad" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 693, + 124, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 693, + 124, + 703 + ], + "spans": [ + { + "bbox": [ + 58, + 693, + 124, + 703 + ], + "type": "text", + "content": "*Equal contribution" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 59, + 703, + 134, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 703, + 134, + 712 + ], + "spans": [ + { + "bbox": [ + 59, + 703, + 134, + 712 + ], + "type": "text", + "content": "† Corresponding author" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 293, + 732, + 317, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 732, + 317, + 742 + ], + "spans": [ + { + "bbox": [ + 293, + 732, + 317, + 742 + ], + "type": "text", + "content": "12157" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 215 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 215 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 215 + ], + "type": "text", + "content": "tion under continual distribution shifts [25, 29, 33, 35]. The main challenges are: (1) The poor computational ability of devices. Due to the properties of device infrastructure, the deployed model can not be updated on time, thus lagging its performance for the real world under distribution shift. (2) The limited generalization ability of the lightweight model. Since lightweight models are of relatively small capacity, they can not handle continually changing environments. In contrast, recent large models that are trained on the cloud server show significant generalization ability [2, 39]. While in industry, these large models can not be directly applied due to the limited infrastructure." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 216, + 289, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 216, + 289, + 335 + ], + "spans": [ + { + "bbox": [ + 46, + 216, + 289, + 335 + ], + "type": "text", + "content": "Therefore, we enable the device model to tackle real-world environmental changing by proposing a Cloud-Device Collaborative Continual Adaptation paradigm, as shown in Fig.1 (a). Previous Cloud-Device Collaborative methods only focus on improving model representation on variance in video frames but neglect the model generalization ability for continually changing data distribution. In our new paradigm, we fully exploit the sufficient knowledge of the large cloud model and transfer the continual generalization ability to the device lightweight model." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 335, + 289, + 526 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 335, + 289, + 526 + ], + "spans": [ + { + "bbox": [ + 46, + 335, + 289, + 526 + ], + "type": "text", + "content": "In particular, we design an Uncertainty-based Visual Prompt Adapted (U-VPA) teacher-student model, which consists of an Uncertainty Guided Sampling (UGS) strategy and a Visual Prompt Learning Strategy with Uncertainty Guided Updating (VPLU). Due to the communication bandwidth constraint [11, 13, 19], different with some work targeting filtering the parameters [22, 26], we design the UGS to screen out the most environment-specific samples and decrease the required bandwidth compared with transmitting the whole sequence. To leverage the strong generalization ability of large models, we introduce the VPLU to align source-target domain distribution and transfer the representation of the large teacher model to the light student model. The light student model and visual prompts are then delivered to the device, thus alleviating the continuously changing scenarios in the real world." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 527, + 289, + 658 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 527, + 289, + 658 + ], + "spans": [ + { + "bbox": [ + 46, + 527, + 289, + 658 + ], + "type": "text", + "content": "Experimental results show that our method outperforms the state-of-the-art methods on synthetic and real-world distribution shift datasets, as shown in Fig.1 (b). Besides, we can achieve the same performance as the entire data with fewer reflowed data (42% of total data). As another benefit, fewer reflowed data reduce the bandwidth pressure of the uplink. As for the downlink, we can deliver the visual prompts (0.43% of the model's parameters) with almost negligible bandwidth to the device and apply the visual prompt to the input data to improve the performance of the device model by 3.9% in mAP." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 59, + 658, + 259, + 670 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 658, + 259, + 670 + ], + "spans": [ + { + "bbox": [ + 59, + 658, + 259, + 670 + ], + "type": "text", + "content": "Our contributions can be summarized as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 58, + 677, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 677, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 58, + 677, + 287, + 713 + ], + "type": "text", + "content": "- We make the first attempt to deal with continually changing scenarios by proposing a Cloud-Device Collaborative Continual Adaptation paradigm, which aims" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 324, + 72, + 545, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 324, + 72, + 545, + 120 + ], + "spans": [ + { + "bbox": [ + 324, + 72, + 545, + 120 + ], + "type": "text", + "content": "to transfer the generalization ability from the large cloud model to the lightweight device model. Our method is a general paradigm that can apply to real-world systems." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 317, + 128, + 545, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 128, + 545, + 236 + ], + "spans": [ + { + "bbox": [ + 317, + 128, + 545, + 236 + ], + "type": "text", + "content": "- We design an Uncertainty-based Visual Prompt Adapted (U-VPA) teacher-student model, which consists of UGS and VPLU. We introduce UGS to screen out the most environment-specific samples and decrease the required bandwidth compared with transmitting the whole sequence. We propose a VPLU to align source-target data distribution and transfer the representation of the large teacher model to the lightweight student model." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 317, + 243, + 545, + 291 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 243, + 545, + 291 + ], + "spans": [ + { + "bbox": [ + 317, + 243, + 545, + 291 + ], + "type": "text", + "content": "- Experiments show that our proposed framework and method surpass other state-of-the-art methods and can effectively improve the continuous domain adaptation capability of the device model." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 311, + 390, + 323 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 311, + 390, + 323 + ], + "spans": [ + { + "bbox": [ + 306, + 311, + 390, + 323 + ], + "type": "text", + "content": "2. Related work" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 331, + 545, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 331, + 545, + 426 + ], + "spans": [ + { + "bbox": [ + 304, + 331, + 545, + 426 + ], + "type": "text", + "content": "Test time adaptation. Test-time adaptation(TTA) aims to get better model generalization performance on target data without access to source data. Feature modulation is used in TENT [34] to optimize BN layers of the input pretrained model by entropy minimization. SHOT [20] takes entropy minimization and a diversity regularizer to achieve information maximization. [18, 24] both take a productive way to enhance model performance on the target domain." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 426, + 545, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 426, + 545, + 510 + ], + "spans": [ + { + "bbox": [ + 304, + 426, + 545, + 510 + ], + "type": "text", + "content": "We differ from TTA in that we consider continuously changing domains. Besides, our model is deployed on the device. There will be restrictions on computing resources and uplink and downlink transmission bandwidth. Considering TTA can effectively use the test data to make updates, we leverage TTA to enable the cloud to improve generalization performance continuously." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 510, + 545, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 510, + 545, + 617 + ], + "spans": [ + { + "bbox": [ + 304, + 510, + 545, + 617 + ], + "type": "text", + "content": "Prompt learning. Prompt is originally from the natural language process (NLP) and aims to adapt the large-scale pre-trained language model to downstream tasks. Recently, several methods [1, 9, 21, 40, 41] are proposed to prove the feasibility of using prompts to adapt the large pre-trained model to downstream tasks in computer vision or vision-language tasks. [1, 21, 37] take prompt as additional input containing a small number of learnable parameters and explore the design of prompts." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 618, + 545, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 618, + 545, + 677 + ], + "spans": [ + { + "bbox": [ + 304, + 618, + 545, + 677 + ], + "type": "text", + "content": "Prompt learning aims at mining the pre-trained model's information. We introduce visual prompt tuning to improve the performance of the device model on the current domain, with a negligible computational overhead in our Cloud-Device Collaboration (DCC) framework." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "content": "Cloud-device collaboration. Cloud-device collaboration aims to bring benefits to both the cloud and the device jointly. Several methods [4, 6, 7, 15] try to unload part or" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 732, + 317, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 732, + 317, + 742 + ], + "spans": [ + { + "bbox": [ + 294, + 732, + 317, + 742 + ], + "type": "text", + "content": "12158" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 82, + 541, + 352 + ], + "blocks": [ + { + "bbox": [ + 57, + 82, + 541, + 352 + ], + "lines": [ + { + "bbox": [ + 57, + 82, + 541, + 352 + ], + "spans": [ + { + "bbox": [ + 57, + 82, + 541, + 352 + ], + "type": "image", + "image_path": "30c87bae3bde82773df112a90c27fb43c35c3e88475bd3911825c0580a8d2b34.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 365, + 548, + 454 + ], + "lines": [ + { + "bbox": [ + 46, + 365, + 548, + 454 + ], + "spans": [ + { + "bbox": [ + 46, + 365, + 548, + 454 + ], + "type": "text", + "content": "Figure 2. The whole framework. (a): Testing on the device. On the client device, we select data that need to be sent to cloud by Uncertainty Guided Sampling (UGS). These uncertainty samples are then used in the cloud to update the teacher-student models. The visual prompt trained on the cloud will be delivered to the device through the downlink and applied to the input data " + }, + { + "bbox": [ + 46, + 365, + 548, + 454 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 46, + 365, + 548, + 454 + ], + "type": "text", + "content": ". Then the reformulated input data will be put into the device model for testing. (b) Training on the cloud. We add our proposed visual prompt to the uncertain samples at the pixel level to get the reformulated images. We then put the reformulated images into the teacher-student model. The teacher-student model and visual prompts are optimized by align loss and supervised loss. (c) Visual prompt learning strategy with uncertainty guided sampling (VPLU). We update the visual prompts by uncertainty-based EMA update. The update weight of the visual prompt will be larger when the uncertainty is higher. The student model is updated using the uncertainty samples and the pseudo labels." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 460, + 287, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 460, + 287, + 555 + ], + "spans": [ + { + "bbox": [ + 46, + 460, + 287, + 555 + ], + "type": "text", + "content": "all computation to the cloud to alleviate the computation deficiency on the device. But they do not leverage the collaboration between the cloud and the device. AMS [16] and DCCL [38] have considered the collaboration but they do not take the real-world continual distribution shifts into account. Compared with methods above, our proposed CloudDevice Collaborative Continual Adaptation paradigm can deal with the real-world environment changes." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 555, + 288, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 555, + 288, + 616 + ], + "spans": [ + { + "bbox": [ + 46, + 555, + 288, + 616 + ], + "type": "text", + "content": "In our work, we use Uncertainty Guided Sampling (UGS) and visual prompt for more efficient bandwidth usage. Besides, our model can adapt to different domains during test time. Our teacher model on the cloud and our student model on the device can be improved simultaneously." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 625, + 103, + 639 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 625, + 103, + 639 + ], + "spans": [ + { + "bbox": [ + 47, + 625, + 103, + 639 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 645, + 127, + 658 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 645, + 127, + 658 + ], + "spans": [ + { + "bbox": [ + 47, + 645, + 127, + 658 + ], + "type": "text", + "content": "3.1. Preliminary" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 663, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 663, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 663, + 288, + 715 + ], + "type": "text", + "content": "Given a lightweight model " + }, + { + "bbox": [ + 46, + 663, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\theta_{device}" + }, + { + "bbox": [ + 46, + 663, + 288, + 715 + ], + "type": "text", + "content": " that is deployed on the device, our goal is to perform well on the continually changing target domains " + }, + { + "bbox": [ + 46, + 663, + 288, + 715 + ], + "type": "inline_equation", + "content": "D_{t1}, D_{t2}, \\ldots, D_{tT}" + }, + { + "bbox": [ + 46, + 663, + 288, + 715 + ], + "type": "text", + "content": " with limited communication bandwidth, where " + }, + { + "bbox": [ + 46, + 663, + 288, + 715 + ], + "type": "inline_equation", + "content": "D_{ti} = \\{(x_i^T)\\}_{i=1}^{N_t}" + }, + { + "bbox": [ + 46, + 663, + 288, + 715 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 663, + 288, + 715 + ], + "type": "inline_equation", + "content": "N_t" + }, + { + "bbox": [ + 46, + 663, + 288, + 715 + ], + "type": "text", + "content": " repre" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 460, + 545, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 460, + 545, + 521 + ], + "spans": [ + { + "bbox": [ + 304, + 460, + 545, + 521 + ], + "type": "text", + "content": "sents the scale of the target domain. The distributions of the target domains are arbitrary since they can change or reoccur over time. We thus introduce a large teacher model " + }, + { + "bbox": [ + 304, + 460, + 545, + 521 + ], + "type": "inline_equation", + "content": "\\theta_{tea}" + }, + { + "bbox": [ + 304, + 460, + 545, + 521 + ], + "type": "text", + "content": " and lightweight student model " + }, + { + "bbox": [ + 304, + 460, + 545, + 521 + ], + "type": "inline_equation", + "content": "\\theta_{stu}" + }, + { + "bbox": [ + 304, + 460, + 545, + 521 + ], + "type": "text", + "content": " cloned from " + }, + { + "bbox": [ + 304, + 460, + 545, + 521 + ], + "type": "inline_equation", + "content": "\\theta_{device}" + }, + { + "bbox": [ + 304, + 460, + 545, + 521 + ], + "type": "text", + "content": ", to boost the generalization ability of " + }, + { + "bbox": [ + 304, + 460, + 545, + 521 + ], + "type": "inline_equation", + "content": "\\theta_{device}" + }, + { + "bbox": [ + 304, + 460, + 545, + 521 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 528, + 419, + 539 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 528, + 419, + 539 + ], + "spans": [ + { + "bbox": [ + 306, + 528, + 419, + 539 + ], + "type": "text", + "content": "3.2. Overall framework" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 545, + 545, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 545, + 545, + 641 + ], + "spans": [ + { + "bbox": [ + 304, + 545, + 545, + 641 + ], + "type": "text", + "content": "In our Cloud-Device Collaborative Continual Adaptation paradigm, the key idea is to improve the device model's performance continuously by receiving adapted parameters from the cloud server. So we propose Visual Prompt Adapted (U-VPA) teacher-student model. As shown in Fig. 2, the U-VPA teacher-student model contains Uncertainty Guided Sampling (UGS) and Visual Prompt Learning Strategy with Uncertainty Guided Updating (VPLU)." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 642, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 642, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 642, + 546, + 713 + ], + "type": "text", + "content": "When perceiving scene changing or distribution shift, poor computational device intends to get help from a cloud server with powerful computing ability. However, due to the communication bandwidth constraint, we must consider the required bandwidth when concurring the distribution shifts. Therefore, we design a Uncertainty Guided Sam" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 733, + 317, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 733, + 317, + 742 + ], + "spans": [ + { + "bbox": [ + 294, + 733, + 317, + 742 + ], + "type": "text", + "content": "12159" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 143 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 143 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 143 + ], + "type": "text", + "content": "pling (UGS) to screen out the most environment-specific samples and decrease the required bandwidth compared with transmitting the whole sequence. After receiving the learned parameters from the cloud to deal with these samples of relatively large distribution shifts, it can adapt to the new scene when inference." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 144, + 287, + 264 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 144, + 287, + 264 + ], + "spans": [ + { + "bbox": [ + 46, + 144, + 287, + 264 + ], + "type": "text", + "content": "On the cloud server, due to its strengths in computing, we intend to leverage the strong generalization ability of large models to tackle the received environment-specific and challenging samples. We introduce Visual Prompt Learning Strategy with Uncertainty guided Updating (VPLU) to align source-target domain distribution and transfer the representation of the large teacher model to the light student model. The light student model and visual prompts is then delivered to the device, thus alleviating the continuously changing scenarios in the real-world." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 265, + 286, + 288 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 265, + 286, + 288 + ], + "spans": [ + { + "bbox": [ + 47, + 265, + 286, + 288 + ], + "type": "text", + "content": "We will discuss the details of UGS and VPLU in Sec. 3.3.1 and Sec. 3.3.2." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 299, + 209, + 310 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 299, + 209, + 310 + ], + "spans": [ + { + "bbox": [ + 47, + 299, + 209, + 310 + ], + "type": "text", + "content": "3.3. U-VPA teacher-student model" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 318, + 237, + 330 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 318, + 237, + 330 + ], + "spans": [ + { + "bbox": [ + 47, + 318, + 237, + 330 + ], + "type": "text", + "content": "3.3.1 Uncertainty Guided Sampling(UGS)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 338, + 287, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 338, + 287, + 422 + ], + "spans": [ + { + "bbox": [ + 46, + 338, + 287, + 422 + ], + "type": "text", + "content": "Due to bandwidth limitations, it is not practical for the device to return all data. Therefore, we select images that need to be transmitted based on uncertainty estimation. Inspired by the Dropout methods [8, 12], since we are doing an object detection task, we can obtain " + }, + { + "bbox": [ + 46, + 338, + 287, + 422 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 46, + 338, + 287, + 422 + ], + "type": "text", + "content": " probabilities " + }, + { + "bbox": [ + 46, + 338, + 287, + 422 + ], + "type": "inline_equation", + "content": "p_i(y|x)" + }, + { + "bbox": [ + 46, + 338, + 287, + 422 + ], + "type": "text", + "content": " of the model by dropout method, and we calculate the variance of these " + }, + { + "bbox": [ + 46, + 338, + 287, + 422 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 46, + 338, + 287, + 422 + ], + "type": "text", + "content": " prediction " + }, + { + "bbox": [ + 46, + 338, + 287, + 422 + ], + "type": "inline_equation", + "content": "p_i(y|x)" + }, + { + "bbox": [ + 46, + 338, + 287, + 422 + ], + "type": "text", + "content": " as uncertainty." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 98, + 434, + 287, + 463 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 98, + 434, + 287, + 463 + ], + "spans": [ + { + "bbox": [ + 98, + 434, + 287, + 463 + ], + "type": "interline_equation", + "content": "V _ {u n c} = \\left(\\frac {1}{n} \\sum_ {i} ^ {n} \\left(p _ {i} (y | x) - \\mu\\right) ^ {2}\\right) ^ {\\frac {1}{2}} \\tag {1}", + "image_path": "14f749d35c361a101c23a32d81bf1d3f6c2fde31e237ee84513927fcbd71a6b0.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 472, + 287, + 542 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 472, + 287, + 542 + ], + "spans": [ + { + "bbox": [ + 46, + 472, + 287, + 542 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 472, + 287, + 542 + ], + "type": "inline_equation", + "content": "p(y|x)" + }, + { + "bbox": [ + 46, + 472, + 287, + 542 + ], + "type": "text", + "content": " represents the predictions of the input image " + }, + { + "bbox": [ + 46, + 472, + 287, + 542 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 46, + 472, + 287, + 542 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 472, + 287, + 542 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 46, + 472, + 287, + 542 + ], + "type": "text", + "content": " is the mean. " + }, + { + "bbox": [ + 46, + 472, + 287, + 542 + ], + "type": "inline_equation", + "content": "V_{unc}" + }, + { + "bbox": [ + 46, + 472, + 287, + 542 + ], + "type": "text", + "content": " is the variance of " + }, + { + "bbox": [ + 46, + 472, + 287, + 542 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 46, + 472, + 287, + 542 + ], + "type": "text", + "content": " probabilities of the model, which represents the uncertainty of the model for the data. The device model filters the input images " + }, + { + "bbox": [ + 46, + 472, + 287, + 542 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 46, + 472, + 287, + 542 + ], + "type": "text", + "content": " whose uncertainty is larger than a threshold and sends back to the cloud server." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 562, + 286, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 562, + 286, + 586 + ], + "spans": [ + { + "bbox": [ + 47, + 562, + 286, + 586 + ], + "type": "text", + "content": "3.3.2 Visual Prompt Learning Strategy with Uncertainty Guided Updating (VPLU)." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 594, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 287, + 713 + ], + "type": "text", + "content": "We optimize model on the cloud server by using the data transmitted from the device. It then delivers the learned parameters and processed visual prompt to the device model for inference. We thus introduce VPLU with a large teacher model and light student model, to mitigate variant domain distribution and improve the generalization ability of light model. Note that, teacher model is initialized by the source model pre-trained on source domain. It takes source and target domain as input, while student model only trains on the target domain." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 72, + 545, + 179 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 179 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 179 + ], + "type": "text", + "content": "Cross-domain alignment. In order to mitigate distribution shift, we align the features in the source and target domains. Inspired by [10], we apply feature alignment on teacher model to improve its performance on target domain. We use a domain discriminator to classify data in feature space and try to distinguish which domain the data comes from. Besides, we use a gradient reversal layer to achieve a uniform training process. Source domain data is used to supervise teacher model, avoiding performance degradation." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 180, + 545, + 227 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 180, + 545, + 227 + ], + "spans": [ + { + "bbox": [ + 304, + 180, + 545, + 227 + ], + "type": "text", + "content": "Cross-model transfer. After improving teacher model's domain-invariant representation ability, we transfer its ability against distribution shifts and generalization to student model via visual prompts and knowledge transferring." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 228, + 545, + 287 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 228, + 545, + 287 + ], + "spans": [ + { + "bbox": [ + 304, + 228, + 545, + 287 + ], + "type": "text", + "content": "Visual Prompt Adaptation(VPA) learning. Specifically, inspired by the prompt learning in NLP [23], which can effectively explores the knowledge of source domain, we propose visual prompts as hints to continually transfer teacher model's generalized representation to the student model." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 288, + 545, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 288, + 545, + 335 + ], + "spans": [ + { + "bbox": [ + 304, + 288, + 545, + 335 + ], + "type": "text", + "content": "In particular, visual prompts " + }, + { + "bbox": [ + 304, + 288, + 545, + 335 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 304, + 288, + 545, + 335 + ], + "type": "text", + "content": " are learnable parameters. During the forward process of the model, we can obtain the reformulated image " + }, + { + "bbox": [ + 304, + 288, + 545, + 335 + ], + "type": "inline_equation", + "content": "x^{*}" + }, + { + "bbox": [ + 304, + 288, + 545, + 335 + ], + "type": "text", + "content": " by adding visual prompts on the input images " + }, + { + "bbox": [ + 304, + 288, + 545, + 335 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 304, + 288, + 545, + 335 + ], + "type": "text", + "content": " at the pixel level." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 399, + 347, + 545, + 359 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 399, + 347, + 545, + 359 + ], + "spans": [ + { + "bbox": [ + 399, + 347, + 545, + 359 + ], + "type": "interline_equation", + "content": "x ^ {*} = x + \\phi , \\tag {2}", + "image_path": "8d3b6d5b93a9f3c1d6f0d9f602d02f205c52716aae8af4de41c9a26a61283d6f.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 363, + 545, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 363, + 545, + 399 + ], + "spans": [ + { + "bbox": [ + 304, + 363, + 545, + 399 + ], + "type": "text", + "content": "In our framework, whether in training on server or inference on device, we all use the reformulated image " + }, + { + "bbox": [ + 304, + 363, + 545, + 399 + ], + "type": "inline_equation", + "content": "x^{*}" + }, + { + "bbox": [ + 304, + 363, + 545, + 399 + ], + "type": "text", + "content": " as model's input, instead of origin images " + }, + { + "bbox": [ + 304, + 363, + 545, + 399 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 304, + 363, + 545, + 399 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 399, + 545, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 399, + 545, + 506 + ], + "spans": [ + { + "bbox": [ + 304, + 399, + 545, + 506 + ], + "type": "text", + "content": "Uncertainty-aware Prompt Updation(U-EMA). Visual prompt acts on the input data and participates in the forward process of both teacher and student models, so it can also benefit from their optimization during training. Moreover, to make the prompt update more stable and consider the current model's uncertainty to data, we update the visual prompt using by uncertainty-based exponential moving average(U-EMA). The weight of the exponential moving average is based on the uncertainty estimation as follows:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 370, + 518, + 545, + 531 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 370, + 518, + 545, + 531 + ], + "spans": [ + { + "bbox": [ + 370, + 518, + 545, + 531 + ], + "type": "interline_equation", + "content": "\\phi_ {t} \\leftarrow \\beta \\phi_ {t - 1} + (1 - \\beta) \\phi_ {t}, \\tag {3}", + "image_path": "da6e276fc13f1e4b8f7764accb8286722b8f40a7b5d435d2f132d5e5a416d25b.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 534, + 545, + 642 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 534, + 545, + 642 + ], + "spans": [ + { + "bbox": [ + 304, + 534, + 545, + 642 + ], + "type": "text", + "content": "Note that " + }, + { + "bbox": [ + 304, + 534, + 545, + 642 + ], + "type": "inline_equation", + "content": "\\phi_t" + }, + { + "bbox": [ + 304, + 534, + 545, + 642 + ], + "type": "text", + "content": " represents the parameters of the visual prompt that need to be updated, and " + }, + { + "bbox": [ + 304, + 534, + 545, + 642 + ], + "type": "inline_equation", + "content": "\\phi_{t - 1}" + }, + { + "bbox": [ + 304, + 534, + 545, + 642 + ], + "type": "text", + "content": " indicates the visual prompt parameters of the last training step t. We follow [30] to set " + }, + { + "bbox": [ + 304, + 534, + 545, + 642 + ], + "type": "inline_equation", + "content": "\\alpha = 0.999" + }, + { + "bbox": [ + 304, + 534, + 545, + 642 + ], + "type": "text", + "content": ". And " + }, + { + "bbox": [ + 304, + 534, + 545, + 642 + ], + "type": "inline_equation", + "content": "\\beta = \\alpha -V_{unc}" + }, + { + "bbox": [ + 304, + 534, + 545, + 642 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 534, + 545, + 642 + ], + "type": "inline_equation", + "content": "V_{unc}" + }, + { + "bbox": [ + 304, + 534, + 545, + 642 + ], + "type": "text", + "content": " is defined in Eq. (1) as the uncertainty of current data. Higher uncertainty means a larger distribution shift in the current data. To save more domain-related information, the prompt update weight will be larger at this time. That is, the larger the uncertainty, the greater the weight of the prompt update." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 642, + 545, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 642, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 304, + 642, + 545, + 689 + ], + "type": "text", + "content": "As another essential role of the visual prompt, the processed visual prompt will be sent to the device to improve the adaptability of the device model by applying the visual prompt on input images." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "content": "Knowlegde transfer. Along with visual prompt adapted tuning, we align teacher student feature and adopt teacher" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 732, + 317, + 741 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 732, + 317, + 741 + ], + "spans": [ + { + "bbox": [ + 294, + 732, + 317, + 741 + ], + "type": "text", + "content": "12160" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 217 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 217 + ], + "type": "text", + "content": "model to generate pseudo label for student model. It is penalized by Knowledge transfer loss " + }, + { + "bbox": [ + 46, + 72, + 289, + 217 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_k" + }, + { + "bbox": [ + 46, + 72, + 289, + 217 + ], + "type": "text", + "content": ", to promote the consistency between teacher student model representation. Specifically, The teacher model will generate predictions for the reformulated images " + }, + { + "bbox": [ + 46, + 72, + 289, + 217 + ], + "type": "inline_equation", + "content": "x^{*}" + }, + { + "bbox": [ + 46, + 72, + 289, + 217 + ], + "type": "text", + "content": " as pseudo labels. Teacher model has stronger generalization ability, so the generated pseudo labels are of high quality. To further improve the pseudo labels quality, we screen those predictions with confidence higher than the threshold as the input to the student model. Overall, we transfer the knowledge from the teacher to the student by feeding selected images " + }, + { + "bbox": [ + 46, + 72, + 289, + 217 + ], + "type": "inline_equation", + "content": "x^{*}" + }, + { + "bbox": [ + 46, + 72, + 289, + 217 + ], + "type": "text", + "content": " and corresponding pseudo labels into the student model for training." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 222, + 160, + 235 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 222, + 160, + 235 + ], + "spans": [ + { + "bbox": [ + 47, + 222, + 160, + 235 + ], + "type": "text", + "content": "3.4. Training objectives" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 240, + 287, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 240, + 287, + 300 + ], + "spans": [ + { + "bbox": [ + 47, + 240, + 287, + 300 + ], + "type": "text", + "content": "The teacher and student models are jointly optimized on the cloud server, using uncertain samples " + }, + { + "bbox": [ + 47, + 240, + 287, + 300 + ], + "type": "inline_equation", + "content": "x_{t}^{i}\\in D_{t}" + }, + { + "bbox": [ + 47, + 240, + 287, + 300 + ], + "type": "text", + "content": " returned from device. The source model is trained on source pairs of " + }, + { + "bbox": [ + 47, + 240, + 287, + 300 + ], + "type": "inline_equation", + "content": "\\{x_s^i,y_s^i\\}_{i = 1}^{n_s}\\in D_s" + }, + { + "bbox": [ + 47, + 240, + 287, + 300 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 47, + 240, + 287, + 300 + ], + "type": "inline_equation", + "content": "x_{s}^{i}\\in X_{s}" + }, + { + "bbox": [ + 47, + 240, + 287, + 300 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 240, + 287, + 300 + ], + "type": "inline_equation", + "content": "y_{s}^{i}\\in Y_{s}" + }, + { + "bbox": [ + 47, + 240, + 287, + 300 + ], + "type": "text", + "content": " are images and labels, respectively." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 300, + 287, + 348 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 300, + 287, + 348 + ], + "spans": [ + { + "bbox": [ + 47, + 300, + 287, + 348 + ], + "type": "text", + "content": "For the training process of the teacher model, given the reformulated taregt images " + }, + { + "bbox": [ + 47, + 300, + 287, + 348 + ], + "type": "inline_equation", + "content": "x_{t}^{*}" + }, + { + "bbox": [ + 47, + 300, + 287, + 348 + ], + "type": "text", + "content": ", as the Eq. (4) shown, we use adversal alignment loss [31] and supervision loss to optimize the teacher model (including visual prompt " + }, + { + "bbox": [ + 47, + 300, + 287, + 348 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 47, + 300, + 287, + 348 + ], + "type": "text", + "content": ")." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 363, + 287, + 392 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 363, + 287, + 392 + ], + "spans": [ + { + "bbox": [ + 61, + 363, + 287, + 392 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} L _ {t e a} = - \\lambda L _ {d} \\left(G _ {d} \\left(F \\left(x _ {s} ^ {*}; \\theta_ {t e a}\\right); F \\left(x _ {t} ^ {*}; \\theta_ {t e a}\\right)\\right), y _ {d}\\right) \\tag {4} \\\\ + L _ {D e t} \\left(X \\left(x _ {s} ^ {*}, \\theta_ {t e a}\\right), y _ {s}\\right), \\\\ \\end{array}", + "image_path": "46b59461466ce57006609b9d59acb036869484aff8fe186f2c1758854e1b1c64.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 397, + 287, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 397, + 287, + 468 + ], + "spans": [ + { + "bbox": [ + 47, + 397, + 287, + 468 + ], + "type": "text", + "content": "Note that the first part is adversal alignment loss [31], and another one is the loss function of Faster RCNN [27] (using labeled data in source domain " + }, + { + "bbox": [ + 47, + 397, + 287, + 468 + ], + "type": "inline_equation", + "content": "D_{s}" + }, + { + "bbox": [ + 47, + 397, + 287, + 468 + ], + "type": "text", + "content": "). " + }, + { + "bbox": [ + 47, + 397, + 287, + 468 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 47, + 397, + 287, + 468 + ], + "type": "text", + "content": " represents the model which predict detection outputs. " + }, + { + "bbox": [ + 47, + 397, + 287, + 468 + ], + "type": "inline_equation", + "content": "L_{d}" + }, + { + "bbox": [ + 47, + 397, + 287, + 468 + ], + "type": "text", + "content": " represent domain classifier. " + }, + { + "bbox": [ + 47, + 397, + 287, + 468 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 47, + 397, + 287, + 468 + ], + "type": "text", + "content": " indicates the feature extractor and " + }, + { + "bbox": [ + 47, + 397, + 287, + 468 + ], + "type": "inline_equation", + "content": "G_{d}" + }, + { + "bbox": [ + 47, + 397, + 287, + 468 + ], + "type": "text", + "content": " is the domain classifier. " + }, + { + "bbox": [ + 47, + 397, + 287, + 468 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 47, + 397, + 287, + 468 + ], + "type": "text", + "content": " is the weight of alignment loss." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 469, + 287, + 493 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 469, + 287, + 493 + ], + "spans": [ + { + "bbox": [ + 47, + 469, + 287, + 493 + ], + "type": "text", + "content": "For the optimization of the student model, we use pseudo labels " + }, + { + "bbox": [ + 47, + 469, + 287, + 493 + ], + "type": "inline_equation", + "content": "y_{t}^{p}" + }, + { + "bbox": [ + 47, + 469, + 287, + 493 + ], + "type": "text", + "content": " generated by teacher model as supervision:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 102, + 504, + 287, + 517 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 102, + 504, + 287, + 517 + ], + "spans": [ + { + "bbox": [ + 102, + 504, + 287, + 517 + ], + "type": "interline_equation", + "content": "L _ {s t u} = L _ {D e t} \\left(X \\left(x _ {t} ^ {*}, \\theta_ {s t u}\\right), y _ {t} ^ {p}\\right), \\tag {5}", + "image_path": "3c17c9fb553a384ff7cb5e6539d5cbfa5a56d819eb10cbbb6cb635ab15a94310.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 521, + 287, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 521, + 287, + 557 + ], + "spans": [ + { + "bbox": [ + 47, + 521, + 287, + 557 + ], + "type": "text", + "content": "The visual prompt benefited from both " + }, + { + "bbox": [ + 47, + 521, + 287, + 557 + ], + "type": "inline_equation", + "content": "L_{tea}" + }, + { + "bbox": [ + 47, + 521, + 287, + 557 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 521, + 287, + 557 + ], + "type": "inline_equation", + "content": "L_{stu}" + }, + { + "bbox": [ + 47, + 521, + 287, + 557 + ], + "type": "text", + "content": ". And we update the visual prompt by uncertainty-aware exponential moving average as Eq. (3) shown." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 566, + 128, + 580 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 566, + 128, + 580 + ], + "spans": [ + { + "bbox": [ + 47, + 566, + 128, + 580 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 586, + 287, + 634 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 586, + 287, + 634 + ], + "spans": [ + { + "bbox": [ + 47, + 586, + 287, + 634 + ], + "type": "text", + "content": "In Sec.4.1, we give the details of the dataset and experiment setting. We evaluate the performance of CCA on continual changing scenarios in Sec.4.2. We give discussion on our motivations and ablation study in Sec.4.3, Sec.4.4." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 640, + 162, + 653 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 640, + 162, + 653 + ], + "spans": [ + { + "bbox": [ + 47, + 640, + 162, + 653 + ], + "type": "text", + "content": "4.1. Experimental setup" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 47, + 659, + 169, + 671 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 659, + 169, + 671 + ], + "spans": [ + { + "bbox": [ + 47, + 659, + 169, + 671 + ], + "type": "text", + "content": "4.1.1 Datasets and settings" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 47, + 677, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 677, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 677, + 287, + 713 + ], + "type": "text", + "content": "We evaluate our proposed method on two test time adaptation benchmark tasks for object detection: Cityscapes-to-Cityscape-C and Cityscapes-to-ACDC-Detection dataset." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 72, + 545, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 167 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 167 + ], + "type": "text", + "content": "Cityscape-C is originally created to benchmark robustness tasks [14]. We select the five most relevant corruptions to the autonomous driving scenario (including brightness, motion_blur, rain, fog and snow). Each corruption dataset contains five levels of severity. The corruptions are applied to images from the validation set of the clean Cityscapes dataset [5]. There are 2,500 images in five corruption datasets." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 168, + 545, + 227 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 168, + 545, + 227 + ], + "spans": [ + { + "bbox": [ + 304, + 168, + 545, + 227 + ], + "type": "text", + "content": "For Cityscapes-to-Cityscapes-C, we follow [3] to train the pre-trained model on the source domain (Cityscapes). Specifically, we use the SGD optimizer with a learning rate of 1e-2 as the official implementation to train the source model." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 229, + 546, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 229, + 546, + 396 + ], + "spans": [ + { + "bbox": [ + 304, + 229, + 546, + 396 + ], + "type": "text", + "content": "ACDC-Detection is created from the Adverse Conditions Dataset (ACDC) [28]. The ACDC dataset shares the same classes with Cityscapes and is collected in various adverse visual conditions: Fog, Night, Rain and Snow. We convert the segmented labels into labels for object detection. We follow CoTTA [35] to set our task. We use 400 unlabeled images from each adverse condition for the adaptation. To mimic the scenario in real life where a similar environment might be revisited and to evaluate the continuous generalization capability of our methods, we repeat the same sequence group (of the three conditions) 10 times (i.e., in total 30: " + }, + { + "bbox": [ + 304, + 229, + 546, + 396 + ], + "type": "inline_equation", + "content": "\\mathrm{Fog} \\to \\mathrm{Rain} \\to \\mathrm{Snow} \\to \\mathrm{Fog}\\ldots" + }, + { + "bbox": [ + 304, + 229, + 546, + 396 + ], + "type": "text", + "content": "). This also provides an evaluation of the adaptation performance in the long term." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 397, + 545, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 397, + 545, + 434 + ], + "spans": [ + { + "bbox": [ + 304, + 397, + 545, + 434 + ], + "type": "text", + "content": "For the Cityscapes-to-ACDC-Detection task, we use the same way to train the pre-trained model on the source domain (Cityscapes) as the Cityscapes-to-Cityscapes-C task." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 306, + 453, + 440, + 464 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 453, + 440, + 464 + ], + "spans": [ + { + "bbox": [ + 306, + 453, + 440, + 464 + ], + "type": "text", + "content": "4.1.2 Implementation Details" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 472, + 545, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 472, + 545, + 544 + ], + "spans": [ + { + "bbox": [ + 304, + 472, + 545, + 544 + ], + "type": "text", + "content": "In this paper, all experiments are conducted with PyTorch. For the Cityscapes-to-Cityscapes-C task and Cityscapes-to-ACDC-Detection task, the batch size is set to 8. We follow [3] to train the pre-trained model on the source domain (Cityscapes). The number of samples used for uncertainty estimation is set to 10 in the experiment." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 545, + 545, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 545, + 545, + 700 + ], + "spans": [ + { + "bbox": [ + 304, + 545, + 545, + 700 + ], + "type": "text", + "content": "We use FasterRCNN with the backbone of ResNet-101 as the large teacher model on the cloud. And we use FasterRCNN with the backbone of ResNet-18 as the lightweight student model (same as the device model). The above large and light models are initialized with corresponding pretrained models on the source domain (Cityscapes dataset). For the UGS module, we set the uncertainty threshold to 0.008, which filters out " + }, + { + "bbox": [ + 304, + 545, + 545, + 700 + ], + "type": "inline_equation", + "content": "42\\%" + }, + { + "bbox": [ + 304, + 545, + 545, + 700 + ], + "type": "text", + "content": " samples in the Cityscape-C dataset. For the VPA module, We set the visual prompts with size " + }, + { + "bbox": [ + 304, + 545, + 545, + 700 + ], + "type": "inline_equation", + "content": "3 \\times 200 \\times 200" + }, + { + "bbox": [ + 304, + 545, + 545, + 700 + ], + "type": "text", + "content": ", which can be added to the left-up space of the image. The student model's parameters and visual prompts (about " + }, + { + "bbox": [ + 304, + 545, + 545, + 700 + ], + "type": "inline_equation", + "content": "0.4\\%" + }, + { + "bbox": [ + 304, + 545, + 545, + 700 + ], + "type": "text", + "content": " of the student model's parameters) will be transmitted to the device." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 318, + 701, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 701, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 318, + 701, + 545, + 712 + ], + "type": "text", + "content": "We use " + }, + { + "bbox": [ + 318, + 701, + 545, + 712 + ], + "type": "inline_equation", + "content": "\\mathrm{mAP} @ 0.5(\\%)" + }, + { + "bbox": [ + 318, + 701, + 545, + 712 + ], + "type": "text", + "content": " as the evaluation metric. The" + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 734, + 317, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 734, + 317, + 742 + ], + "spans": [ + { + "bbox": [ + 294, + 734, + 317, + 742 + ], + "type": "text", + "content": "12161" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 47, + 132, + 545, + 220 + ], + "blocks": [ + { + "bbox": [ + 46, + 70, + 547, + 127 + ], + "lines": [ + { + "bbox": [ + 46, + 70, + 547, + 127 + ], + "spans": [ + { + "bbox": [ + 46, + 70, + 547, + 127 + ], + "type": "text", + "content": "Table 1. Continual generalization capability on Cityscapes-to-Cityscapes-C. Object detection results (mAP@0.5 in %) on the Cityscapes-to-Cityscapes-C online continual test-time adaptation task. Gain(%) means the improvement of our method compared with Source-only. We evaluate the four test conditions continually for ten times to evaluate the long-term adaptation performance. All results are evaluated on the FasterRCNN architecture with the largest corruption severity level 5. Our approach surpasses the SOTA method and exhibits significant continual generalization and anti-forgetting abilities." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 47, + 132, + 545, + 220 + ], + "lines": [ + { + "bbox": [ + 47, + 132, + 545, + 220 + ], + "spans": [ + { + "bbox": [ + 47, + 132, + 545, + 220 + ], + "type": "table", + "html": "
Timet
Round1510All
ConditionFogMotionRainSnowBrightnessFogMotionRainSnowBrightnessFogMotionRainSnowBrightnessMeanGain
Source-only [27]24.79.522.41.329.024.79.522.41.329.024.79.522.41.329.017.4/
TENT-continual [34]24.89.222.51.429.420.06.419.80.823.79.03.810.70.511.413.0-4.4
CoTTA [35]25.110.122.41.328.921.89.520.74.228.29.33.910.13.317.514.7-2.7
Pseudo-Label [17]27.317.224.76.933.730.120.126.98.036.531.020.927.77.238.224.1+6.7
AMS [16]28.417.726.07.634.731.820.428.27.738.034.221.428.97.039.425.1+7.7
Ours(proposed)28.017.826.810.036.537.524.235.514.244.044.027.340.915.150.131.0+13.6
", + "image_path": "63c7d5a1c314a8231196355787e421c5aed56988cc3a9351ea48aec844f4caa6.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 47, + 284, + 545, + 383 + ], + "blocks": [ + { + "bbox": [ + 46, + 223, + 547, + 278 + ], + "lines": [ + { + "bbox": [ + 46, + 223, + 547, + 278 + ], + "spans": [ + { + "bbox": [ + 46, + 223, + 547, + 278 + ], + "type": "text", + "content": "Table 2. Continuous generalization capability on Cityscapes-to-ACDC-Detection. Object detection results (mAP@0.5 in %) on the Cityscapes-to-ACDC-Detection online continual test-time adaptation task. Gain(%) means the improvement of our method compared with Source-only. We evaluate the four test conditions continually for ten times to evaluate the long-term adaptation performance. All results are evaluated on the FasterRCNN architecture. Our results outperform the SOTA method, thus verifying the effectiveness of our method on the real distribution shift." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 47, + 284, + 545, + 383 + ], + "lines": [ + { + "bbox": [ + 47, + 284, + 545, + 383 + ], + "spans": [ + { + "bbox": [ + 47, + 284, + 545, + 383 + ], + "type": "table", + "html": "
Timet
Round14710All
ConditionFogNightRainSnowFogNightRainSnowFogNightRainSnowFogNightRainSnowMeanGain
Source-only [27]20.57.414.314.220.57.414.314.220.57.414.314.220.57.414.314.214.1/
TENT-continual [34]20.57.514.313.920.07.213.413.018.66.612.312.116.66.011.110.912.8-2.3
CoTTA [35]20.77.514.714.320.97.414.714.420.47.313.814.416.36.411.712.213.8-0.3
Pseudo-Label [17]20.98.715.517.623.410.316.918.924.310.917.819.525.111.517.520.017.6+3.5
AMS [16]21.69.016.017.223.210.717.219.124.711.417.219.926.111.717.820.517.8+3.7
Ours(proposed)21.19.315.717.224.711.817.820.226.312.018.520.626.112.318.020.818.5+4.4
", + "image_path": "6e9418e45faf3427333cb1117d518f576bf87e9f1b9a172343e277f78a655b5c.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 395, + 288, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 395, + 288, + 456 + ], + "spans": [ + { + "bbox": [ + 46, + 395, + 288, + 456 + ], + "type": "text", + "content": "overall process of test-time adaptation strictly follows CoTTA [35]. The process first tests the coming batch of data then train with it. We do batch-level updates on the cloud model, and the performance on the device model is taken as the average of the results of all the data." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 474, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 474, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 474, + 288, + 715 + ], + "type": "text", + "content": "Baselines. Our method is compared with multiple baselines, including those for test time adaptation (TTA), continual test time adaptation (CTTA) and cloud-end collaboration methods as follows: Source-only [27]. Test the model performance on the different target domains using the pre-trained source model (training on the source domain). Pseudo-label [17]. Use pseudo label generated by the teacher model to supervise the student model. Tent-continual [34]. Tent is an effective method for the test-time adaptation (TTA). Tent fine-tune the batch normalization layer during training to adapt the model to the target domain. CoTTA [35]. CoTTA is an attractive approach for the continual test-time adaptation (CTTA). CoTTA mitigate error accumulation and catastrophic forgetting with weighted averaging, data augmented averaging, and random recovery neurons. AMS [16]. AMS is an cloud engagement approach. To solve the resource limitation of the device model, AMS do distillation in the cloud, and then transmit the model's parameters to the device in the cloud, the teacher network will not update." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 394, + 425, + 407 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 394, + 425, + 407 + ], + "spans": [ + { + "bbox": [ + 306, + 394, + 425, + 407 + ], + "type": "text", + "content": "4.2. Results and Analysis" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 413, + 545, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 413, + 545, + 653 + ], + "spans": [ + { + "bbox": [ + 304, + 413, + 545, + 653 + ], + "type": "text", + "content": "Synthetic continual distribution shift. We evaluate our proposed method on the synthetic continual distribution shift dataset (Cityscapes " + }, + { + "bbox": [ + 304, + 413, + 545, + 653 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 304, + 413, + 545, + 653 + ], + "type": "text", + "content": " Cityscapes-C task). As shown in Table 1, TENT-continual [34] has a decreasing performance when facing continuously changing domains. Because TENT does not consider continuous distribution shifts, it uses self-entropy to update the bn layer, which is prone to catastrophic forgetting when facing constant distribution shifts. CoTTA [35] is a method designed specifically for continuous test time adaptation. CoTTA applies weighted averaging, data-augmented averaging, and random recovery neurons to overcome catastrophic forgetting and error accumulation. However, when doing the object detection task on the continuously changing domains, the model's performance still decreases, indicating that solutions used by CoTTA are insufficient against catastrophic forgetting. We believe this is caused by the poor generalization abilities of the lightweight device model. Therefore, it is reasonable to introduce the cloud for the Cloud-Device collaborative continual adaptation paradigm." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 653, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 653, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 653, + 547, + 715 + ], + "type": "text", + "content": "Both Pseudo-Label and AMS [16] use the Cloud-Device collaboration framework, which transfers the teacher's knowledge to the student on the cloud. The generalization of the large model provides continuous guidance for the student model, which avoids the problem of catastrophic for" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 733, + 318, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 733, + 318, + 742 + ], + "spans": [ + { + "bbox": [ + 293, + 733, + 318, + 742 + ], + "type": "text", + "content": "12162" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 61, + 78, + 165, + 165 + ], + "blocks": [ + { + "bbox": [ + 61, + 78, + 165, + 165 + ], + "lines": [ + { + "bbox": [ + 61, + 78, + 165, + 165 + ], + "spans": [ + { + "bbox": [ + 61, + 78, + 165, + 165 + ], + "type": "image", + "image_path": "ee175a1cf1b4809dbd6315810cfed251a4429d8ad11b0bc8c8365c6fc6e9a4b0.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 108, + 167, + 115, + 175 + ], + "lines": [ + { + "bbox": [ + 108, + 167, + 115, + 175 + ], + "spans": [ + { + "bbox": [ + 108, + 167, + 115, + 175 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 168, + 78, + 272, + 164 + ], + "blocks": [ + { + "bbox": [ + 168, + 78, + 272, + 164 + ], + "lines": [ + { + "bbox": [ + 168, + 78, + 272, + 164 + ], + "spans": [ + { + "bbox": [ + 168, + 78, + 272, + 164 + ], + "type": "image", + "image_path": "921a84d4117dc510009e17c94cf4f4d6258c468eace56ec72089c3acc77942d2.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 60, + 176, + 272, + 255 + ], + "blocks": [ + { + "bbox": [ + 60, + 176, + 272, + 255 + ], + "lines": [ + { + "bbox": [ + 60, + 176, + 272, + 255 + ], + "spans": [ + { + "bbox": [ + 60, + 176, + 272, + 255 + ], + "type": "image", + "image_path": "fd4b8e6533c44311937a2084116cea2f2d5006fd96df6bfd797526dabe8e023c.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 273, + 289, + 384 + ], + "lines": [ + { + "bbox": [ + 46, + 273, + 289, + 384 + ], + "spans": [ + { + "bbox": [ + 46, + 273, + 289, + 384 + ], + "type": "text", + "content": "Figure 3. (a)-(b): Statistical distribution of confidence and uncertainty. (a): distribution of confidence. (b): distribution of uncertainty. Comparing (a) and (b), the uncertainty value is more stable when facing continuous distribution shift, while the confidence value fluctuates significantly with continuous distribution shift. (c)-(e): T-SNE [32] visualization of samples in different domains. (c): all samples. (d): samples filtered by confidence. (e): samples filtered by UGS. Samples filtered by UGS are more environment-specific. They are more worthy of being transmitted to the cloud server for optimizing the teacher-student models." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 402, + 287, + 427 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 402, + 287, + 427 + ], + "spans": [ + { + "bbox": [ + 46, + 402, + 287, + 427 + ], + "type": "text", + "content": "getting and error accumulation. (Pseudo-Label and AMS are " + }, + { + "bbox": [ + 46, + 402, + 287, + 427 + ], + "type": "inline_equation", + "content": "6.7\\%" + }, + { + "bbox": [ + 46, + 402, + 287, + 427 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 402, + 287, + 427 + ], + "type": "inline_equation", + "content": "7.7\\%" + }, + { + "bbox": [ + 46, + 402, + 287, + 427 + ], + "type": "text", + "content": " higher than Source-only)." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 432, + 287, + 516 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 432, + 287, + 516 + ], + "spans": [ + { + "bbox": [ + 46, + 432, + 287, + 516 + ], + "type": "text", + "content": "However, pseudo label and AMS do not consider updating the teacher model. Our U-VPA framework jointly optimizes the teacher and student model, which makes our method surpass Source-only by " + }, + { + "bbox": [ + 46, + 432, + 287, + 516 + ], + "type": "inline_equation", + "content": "13.6\\%" + }, + { + "bbox": [ + 46, + 432, + 287, + 516 + ], + "type": "text", + "content": ", outperforming all the above methods. This validates that our Cloud-Device collaboration paradigm can effectively improve the continuous domain adaptation ability of the device." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 522, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 522, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 522, + 288, + 713 + ], + "type": "text", + "content": "Real-world continual distribution shift. We also evaluate our proposed method on the real-world continual distribution shift dataset (Cityscapes " + }, + { + "bbox": [ + 46, + 522, + 288, + 713 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 46, + 522, + 288, + 713 + ], + "type": "text", + "content": " ACDC-Detection task). As shown in Table 2, our method improves by " + }, + { + "bbox": [ + 46, + 522, + 288, + 713 + ], + "type": "inline_equation", + "content": "4.4\\%" + }, + { + "bbox": [ + 46, + 522, + 288, + 713 + ], + "type": "text", + "content": " over the source-only method. By comparing the performance change of the same domain in 10 rounds, we can find that our method does not have the problem of catastrophic forgetting. For example, the TENT-continual method faces the problem of catastrophic forgetting. The " + }, + { + "bbox": [ + 46, + 522, + 288, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{mAP}(\\%)" + }, + { + "bbox": [ + 46, + 522, + 288, + 713 + ], + "type": "text", + "content": " drops in the fourth, seventh, and tenth rounds compared to the first round when facing the night domain, specifically " + }, + { + "bbox": [ + 46, + 522, + 288, + 713 + ], + "type": "inline_equation", + "content": "-0.2\\%" + }, + { + "bbox": [ + 46, + 522, + 288, + 713 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 522, + 288, + 713 + ], + "type": "inline_equation", + "content": "-0.5\\%" + }, + { + "bbox": [ + 46, + 522, + 288, + 713 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 522, + 288, + 713 + ], + "type": "inline_equation", + "content": "-4.1\\%" + }, + { + "bbox": [ + 46, + 522, + 288, + 713 + ], + "type": "text", + "content": ". In contrast, our method leads to a continuous performance improvement (e.g., our method's " + }, + { + "bbox": [ + 46, + 522, + 288, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{mAP}(\\%)" + }, + { + "bbox": [ + 46, + 522, + 288, + 713 + ], + "type": "text", + "content": " increases in the fourth, seventh, and tenth rounds, respectively, compared to the first round when facing the Night domain, specifically, " + }, + { + "bbox": [ + 46, + 522, + 288, + 713 + ], + "type": "inline_equation", + "content": "+2.5\\%" + }, + { + "bbox": [ + 46, + 522, + 288, + 713 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 522, + 288, + 713 + ], + "type": "inline_equation", + "content": "+0.2\\%" + }, + { + "bbox": [ + 46, + 522, + 288, + 713 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 522, + 288, + 713 + ], + "type": "inline_equation", + "content": "+0.3\\%" + }, + { + "bbox": [ + 46, + 522, + 288, + 713 + ], + "type": "text", + "content": ")." + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 335, + 83, + 504, + 198 + ], + "blocks": [ + { + "bbox": [ + 335, + 83, + 504, + 198 + ], + "lines": [ + { + "bbox": [ + 335, + 83, + 504, + 198 + ], + "spans": [ + { + "bbox": [ + 335, + 83, + 504, + 198 + ], + "type": "image", + "image_path": "136e8a9008aa01371940180382e53f496479df865cae23273717523acae80955.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 207, + 547, + 297 + ], + "lines": [ + { + "bbox": [ + 305, + 207, + 547, + 297 + ], + "spans": [ + { + "bbox": [ + 305, + 207, + 547, + 297 + ], + "type": "text", + "content": "Figure 4. Detection results (mAP@0.5) of both teacher and student model. Solid lines are results of models using VPA under the brightness domain. Teacher and student models have jointly improved in continuous domain adaptation. Dotted lines are results without using VPA. The teacher quickly reaches the bottleneck, and student's improvement is slow. The analysis above illustrates that the proposed U-VPA framework enables the teacher-student model jointly improve the continuous domain adaption capability." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 309, + 378, + 320 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 309, + 378, + 320 + ], + "spans": [ + { + "bbox": [ + 306, + 309, + 378, + 320 + ], + "type": "text", + "content": "4.3. Discussion" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 328, + 545, + 412 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 328, + 545, + 412 + ], + "spans": [ + { + "bbox": [ + 304, + 328, + 545, + 412 + ], + "type": "text", + "content": "Teacher-student joint optimization with VPA. Previous approaches (e.g., AMS) focus on improving the student model. However, the student model's performance is correlated with the teacher model. Our method enables the teacher and student models to improve performance together by using the visual prompt as a bridge and using the temporal domain information brought by the visual prompt." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 412, + 545, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 412, + 545, + 472 + ], + "spans": [ + { + "bbox": [ + 304, + 412, + 545, + 472 + ], + "type": "text", + "content": "The optimization of the teacher model comes from two modules: (1) feature alignment, (2) visual prompt adaptation(VPA). Our experiments have proved that although both of the above can bring about gains, visual prompt adaptation is the main reason for the success of our framework." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 472, + 545, + 520 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 472, + 545, + 520 + ], + "spans": [ + { + "bbox": [ + 304, + 472, + 545, + 520 + ], + "type": "text", + "content": "As shown in Fig. 4, if only feature alignment is used, the performance of the teacher model reaches the bottleneck after the second round, and the student's optimization will be very slow (dotted lines in the figure)." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 521, + 545, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 521, + 545, + 568 + ], + "spans": [ + { + "bbox": [ + 304, + 521, + 545, + 568 + ], + "type": "text", + "content": "While our method benefits from VPA, the teacher and student models are jointly improved, demonstrating that our proposed U-VPA teacher-student model can optimize the student-teacher model together." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 569, + 545, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 569, + 545, + 700 + ], + "spans": [ + { + "bbox": [ + 304, + 569, + 545, + 700 + ], + "type": "text", + "content": "Uncertainty Guided Sampling (UGS) strategy. We demonstrate in Table 5 that using only a small amount of reflowed data can outperform the results using total data by the UGS strategy. Our model's performance using the UGS strategy to select part of the data to transmit into the cloud for training is " + }, + { + "bbox": [ + 304, + 569, + 545, + 700 + ], + "type": "inline_equation", + "content": "31\\%" + }, + { + "bbox": [ + 304, + 569, + 545, + 700 + ], + "type": "text", + "content": ". Compared to the model performance of " + }, + { + "bbox": [ + 304, + 569, + 545, + 700 + ], + "type": "inline_equation", + "content": "28.3\\%" + }, + { + "bbox": [ + 304, + 569, + 545, + 700 + ], + "type": "text", + "content": " obtained with the total data, there is a " + }, + { + "bbox": [ + 304, + 569, + 545, + 700 + ], + "type": "inline_equation", + "content": "2.7\\%" + }, + { + "bbox": [ + 304, + 569, + 545, + 700 + ], + "type": "text", + "content": " improvement. The amount of data using the random picking strategy is the same as that with the UGS strategy. And the UGS strategy has a better performance than random picking. The above shows the effectiveness of the UGS strategy." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 318, + 701, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 701, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 318, + 701, + 545, + 713 + ], + "type": "text", + "content": "We further reveal the advantages of uncertainty when the" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 733, + 317, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 733, + 317, + 742 + ], + "spans": [ + { + "bbox": [ + 294, + 733, + 317, + 742 + ], + "type": "text", + "content": "12163" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 68, + 106, + 524, + 180 + ], + "blocks": [ + { + "bbox": [ + 46, + 64, + 547, + 100 + ], + "lines": [ + { + "bbox": [ + 46, + 64, + 547, + 100 + ], + "spans": [ + { + "bbox": [ + 46, + 64, + 547, + 100 + ], + "type": "text", + "content": "Table 3. Ablation study. We conduct ablation studies for proposed Pseudo-Label, VPA, UGS, and U-EMA. All the experiments are done on Cityscapes-to-Cityscapes-C, and the evaluation metric is mAP (\\%). The experimental result shows that all these proposed modules positively impact the performance improvement of the devices model when facing continuous domains." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 68, + 106, + 524, + 180 + ], + "lines": [ + { + "bbox": [ + 68, + 106, + 524, + 180 + ], + "spans": [ + { + "bbox": [ + 68, + 106, + 524, + 180 + ], + "type": "table", + "html": "
Pseudo-LabelVPAUGSU-EMAFogMotionRainSnowBrightnessMeanGain
124.79.522.41.329.017.4/
2✓31.020.927.77.238.224.1+6.7
3✓✓37.122.232.711.243.127.4+10.0
4✓✓✓38.324.836.512.346.728.3+10.9
5✓✓✓✓44.027.340.915.150.131.0+13.6
", + "image_path": "2f4ca34effd08c4dee527535a2dad924eb9ddf51240dd71cfd80ab131504f21a.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 50, + 233, + 282, + 282 + ], + "blocks": [ + { + "bbox": [ + 46, + 186, + 288, + 231 + ], + "lines": [ + { + "bbox": [ + 46, + 186, + 288, + 231 + ], + "spans": [ + { + "bbox": [ + 46, + 186, + 288, + 231 + ], + "type": "text", + "content": "Table 4. Comparing the difference between selecting samples with confidence and uncertainty. Experiments show that better performance is obtained by choosing the reflowed data model through uncertainty estimation." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 50, + 233, + 282, + 282 + ], + "lines": [ + { + "bbox": [ + 50, + 233, + 282, + 282 + ], + "spans": [ + { + "bbox": [ + 50, + 233, + 282, + 282 + ], + "type": "table", + "html": "
FogMotionRainSnowBrightnessMean
Confidence39.624.937.214.345.127.4
Uncertainty44.027.340.915.150.131.0
", + "image_path": "69d695cb754c6422ff3f9bb71fbd1b89dcced81df12e929cd337f9cfe91a6bfc.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 50, + 347, + 282, + 399 + ], + "blocks": [ + { + "bbox": [ + 46, + 288, + 287, + 344 + ], + "lines": [ + { + "bbox": [ + 46, + 288, + 287, + 344 + ], + "spans": [ + { + "bbox": [ + 46, + 288, + 287, + 344 + ], + "type": "text", + "content": "Table 5. Effect of uplink transmission of full and partial data on model's performance. Our method returns partial data (" + }, + { + "bbox": [ + 46, + 288, + 287, + 344 + ], + "type": "inline_equation", + "content": "\\tilde{42\\%}" + }, + { + "bbox": [ + 46, + 288, + 287, + 344 + ], + "type": "text", + "content": ") for training but achieves better performance than using full data. This indicates that our UGS strategy reduces the data transmission communication cost and improves the model performance." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 50, + 347, + 282, + 399 + ], + "lines": [ + { + "bbox": [ + 50, + 347, + 282, + 399 + ], + "spans": [ + { + "bbox": [ + 50, + 347, + 282, + 399 + ], + "type": "table", + "html": "
25%50%75%100%
Random23.526.127.928.3
UGS28.231.031.2/
Gain+4.7+4.9+3.3/
", + "image_path": "81d83f6566881418cf173a69e24758d7d43dd9fe9e0e6ccc666ae3b48c0963cc.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 408, + 287, + 528 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 408, + 287, + 528 + ], + "spans": [ + { + "bbox": [ + 46, + 408, + 287, + 528 + ], + "type": "text", + "content": "distribution shifts exist. We find uncertainty is more reliable compared with the confidence score. Regarding Fig.3 (a)-(b), in the presence of distribution shifts, confidence distribution varies significantly, while uncertainty distribution is balanced. Therefore, the samples selected using confidence will face the problem of long-tail distribution, which is not conducive to cloud model training [36]. Uncertainty can more accurately reflect the difficulty of identifying samples by device without using the training of cloud model and discover the samples that are more worthy of attention." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 529, + 287, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 529, + 287, + 590 + ], + "spans": [ + { + "bbox": [ + 46, + 529, + 287, + 590 + ], + "type": "text", + "content": "Besides, we explore the properties of the data filtered by different filter strategies. In Fig.3 (c)-(e), each point represents a sample filtered by the corresponding strategy. We find that samples selected by uncertainty are environmental-specific, so using UGS can filter out more valuable data." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 597, + 139, + 611 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 597, + 139, + 611 + ], + "spans": [ + { + "bbox": [ + 47, + 597, + 139, + 611 + ], + "type": "text", + "content": "4.4. Ablation study" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 616, + 286, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 616, + 286, + 700 + ], + "spans": [ + { + "bbox": [ + 46, + 616, + 286, + 700 + ], + "type": "text", + "content": "Effect of the Uncertainty Guided Sampling (UGS). Our proposed UGS strategy not only reduces the communication cost of the uplink from the device to the cloud but also improves performance by " + }, + { + "bbox": [ + 46, + 616, + 286, + 700 + ], + "type": "inline_equation", + "content": "0.9\\%" + }, + { + "bbox": [ + 46, + 616, + 286, + 700 + ], + "type": "text", + "content": ", as shown in Table 3. Furthermore, as shown in Table 5, the model using UGS outperforms baseline with only " + }, + { + "bbox": [ + 46, + 616, + 286, + 700 + ], + "type": "inline_equation", + "content": "42\\%" + }, + { + "bbox": [ + 46, + 616, + 286, + 700 + ], + "type": "text", + "content": " data for train, which significantly improves the compression ratio of the video." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 59, + 701, + 286, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 701, + 286, + 714 + ], + "spans": [ + { + "bbox": [ + 59, + 701, + 286, + 714 + ], + "type": "text", + "content": "Effect of the Visual Prompt Adapt (VPA). Our pro" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 187, + 545, + 295 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 187, + 545, + 295 + ], + "spans": [ + { + "bbox": [ + 304, + 187, + 545, + 295 + ], + "type": "text", + "content": "posed VPA teacher-student model uses visual prompts to make the student-teacher model jointly improve performance. As shown in Table 3, compared to the pure pseudo label method, the usage of VPA can increase performance by " + }, + { + "bbox": [ + 304, + 187, + 545, + 295 + ], + "type": "inline_equation", + "content": "3.3\\%" + }, + { + "bbox": [ + 304, + 187, + 545, + 295 + ], + "type": "text", + "content": ". Furthermore, we discussed in Sec. 4.3 that VPA could significantly increase the upper limit of the teacher-student framework from challenging continuously changing distribution shift scenarios. The combination of VPA and other modules can further improve performance steadily." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 296, + 545, + 381 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 296, + 545, + 381 + ], + "spans": [ + { + "bbox": [ + 304, + 296, + 545, + 381 + ], + "type": "text", + "content": "Effect of the Uncertainty-aware prompt EMA (U-EMA) update mechanism. The U-EMA update mechanism we proposed in training the cloud model makes the cloud model pay more attention to the images worthy of attention by dynamically adjusting the prompt weights with uncertainty. As shown in Table 3, U-EMA can bring a " + }, + { + "bbox": [ + 304, + 296, + 545, + 381 + ], + "type": "inline_equation", + "content": "2.7\\%" + }, + { + "bbox": [ + 304, + 296, + 545, + 381 + ], + "type": "text", + "content": " performance improvement." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 396, + 454, + 409 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 396, + 454, + 409 + ], + "spans": [ + { + "bbox": [ + 306, + 396, + 454, + 409 + ], + "type": "text", + "content": "5. Conclusion & Limitations" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 418, + 545, + 656 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 418, + 545, + 656 + ], + "spans": [ + { + "bbox": [ + 304, + 418, + 545, + 656 + ], + "type": "text", + "content": "We propose a Cloud-Device Collaborative Continual Adaptation paradigm to deal with continual changing environments on devices. We design an Uncertainty Guided Sampling (UGS) strategy to transmit the most out-of-distribution samples from device to cloud. Besides, we design a Visual Prompt Learning Strategy with Uncertainty Guided Updating (VPLU) to transfer the generalization capability of large model on the cloud to the device model. Experiments of our method on a wide range of continually changing environments show an improvement of " + }, + { + "bbox": [ + 304, + 418, + 545, + 656 + ], + "type": "inline_equation", + "content": "4.4 - 13.6\\%" + }, + { + "bbox": [ + 304, + 418, + 545, + 656 + ], + "type": "text", + "content": " in mAP@0.5 on the object detection task. Our proposed method offers flexibility in its application to most existing cloud-device systems. However, it is worth noting that our experiments were conducted using simulated datasets rather than real systems. Moving forward, an interesting direction for future research would be to test our method on a real system with specific bandwidth limitations, a specific cloud, and a specific device. By taking into account the hardware characteristics of the system, there should have rooms to improve its performance further." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 657, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 657, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 657, + 545, + 714 + ], + "type": "text", + "content": "Acknowledgement. This research was partly supported by the foundation of the National Key R&D Program of China (2022ZD0116305). We thank Xiaoqi Li for the insightful discussions and for helping with writing." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 733, + 318, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 733, + 318, + 742 + ], + "spans": [ + { + "bbox": [ + 294, + 733, + 318, + 742 + ], + "type": "text", + "content": "12164" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 53, + 91, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 124 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 124 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 124 + ], + "type": "text", + "content": "[1] Hyojin Bahng, Ali Jahanian, Swami Sankaranarayanan, and Phillip Isola. Exploring visual prompts for adapting large-scale models. 2022. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 126, + 288, + 455 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 126, + 288, + 455 + ], + "spans": [ + { + "bbox": [ + 53, + 126, + 288, + 455 + ], + "type": "text", + "content": "[2] Rishi Bommasani, Drew A. Hudson, Ehsan Adeli, Russ B. Altman, Simran Arora, Sydney von Arx, Michael S. Bernstein, Jeannette Bohg, Antoine Bosselut, Emma Brunskill, Erik Brynjolfsson, Shyamal Buch, Dallas Card, Rodrigo Castellon, Niladri S. Chatterji, Annie Chen, Kathleen Creel, Jared Davis, Dora Demszky, Chris Donahue, Moussa Doumbouya, Esin Durmus, Stefano Ermon, John Etchemendy, Kawin Ethayarajh, Li Fei-Fei, Chelsea Finn, Trevor Gale, Lauren Gillespie, Karan Goel, Noah D. Goodman, Shelby Grossman, Neel Guha, Tatsunori Hashimoto, Peter Henderson, John Hewitt, Daniel E. Ho, Jenny Hong, Kyle Hsu, Jing Huang, Thomas Icard, Sahil Jain, Dan Jurafsky, Pratyusha Kalluri, Siddharth Karamcheti, Geoff Keeling, Fereshte Khani, Omar Khattab, Pang Wei Koh, Mark Krass, Ranjay Krishna, Rohith Kuditipudi, Ananya Kumar, Faisal Ladhak, Mina Lee, Tony Lee, Jure Leskovec, Isabelle Levent, Xiang Lisa Li, Xuechen Li, Tengyu Ma, Ali Ahmad Malik, Christopher D. Manning, Suvir Mirchandani, Eric Mitchell, Zanele Munyikwa, Suraj Nair, Avanika Narayan, Deepak Narayanan, Ben Newman, Allen Nie, Juan Carlos Niebles, Hamed Nilforoshan, Julian Nyarko, Giray Ogut, Laurel Orr, Isabel Papadimitriou, Joon Sung Park, Chris Piech, Eva Portelance, Christopher Potts, Aditi Raghunathan, Rob Reich, Hongyu Ren, Frieda Rong, Yusuf H. Roohani, Camilo Ruiz, Jack Ryan, Christopher Re, Dorsa Sadigh, Shiori Sagawa, Keshav Santhanam, Andy Shih, Krishnan Srinivasan, Alex Tamkin, Rohan Taori, Armin W. Thomas, Florian Tramèr, Rose E. Wang, and William Yang Wang. On the opportunities and risks of foundation models. arXiv: Learning, 2021. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 456, + 287, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 456, + 287, + 544 + ], + "spans": [ + { + "bbox": [ + 53, + 456, + 287, + 544 + ], + "type": "text", + "content": "[3] Kai Chen, Jiaqi Wang, Jiangmiao Pang, Yuhang Cao, Yu Xiong, Xiaoxiao Li, Shuyang Sun, Wansen Feng, Ziwei Liu, Jiarui Xu, Zheng Zhang, Dazhi Cheng, Chenchen Zhu, Tianheng Cheng, Qijie Zhao, Buyu Li, Xin Lu, Rui Zhu, Yue Wu, Jifeng Dai, Jingdong Wang, Jianping Shi, Wanli Ouyang, Chen Change Loy, and Dahua Lin. MMDetection: Open mmlab detection toolbox and benchmark. arXiv preprint arXiv:1906.07155, 2019. 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 545, + 287, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 545, + 287, + 600 + ], + "spans": [ + { + "bbox": [ + 53, + 545, + 287, + 600 + ], + "type": "text", + "content": "[4] Sandeep Chinchali, Apoorva Sharma, James Harrison, Amine Elhafsi, Daniel Kang, Evgenya Pergament, Eyal Cidon, Sachin Katti, and Marco Pavone. Network offloading policies for cloud robotics: a learning-based approach. Autonomous Robots, 2019. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 602, + 287, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 602, + 287, + 667 + ], + "spans": [ + { + "bbox": [ + 53, + 602, + 287, + 667 + ], + "type": "text", + "content": "[5] Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The cityscapes dataset for semantic urban scene understanding. In Proc. of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 670, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 670, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 53, + 670, + 287, + 713 + ], + "type": "text", + "content": "[6] Daniel Crankshaw, Xin Wang, Giulio Zhou, Michael J. Franklin, Joseph E. Gonzalez, and Ion Stoica. Clipper: A low-latency online prediction serving system. networked systems design and implementation, 2016. 2" + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 23, + "blocks": [ + { + "bbox": [ + 312, + 73, + 545, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 73, + 545, + 128 + ], + "spans": [ + { + "bbox": [ + 312, + 73, + 545, + 128 + ], + "type": "text", + "content": "[7] Sadjad Fouladi, John Emmons, Emre Orbay, Catherine Wu, Riad S. Wahby, and Keith Winstein. Salsify: Low-latency network video through tighter integration between a video codec and a transport protocol. networked systems design and implementation, 2018. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 312, + 129, + 545, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 129, + 545, + 172 + ], + "spans": [ + { + "bbox": [ + 312, + 129, + 545, + 172 + ], + "type": "text", + "content": "[8] Yarin Gal and Zoubin Ghahramani. Dropout as a bayesian approximation: Representing model uncertainty in deep learning. international conference on machine learning, 2015. 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 312, + 175, + 545, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 175, + 545, + 217 + ], + "spans": [ + { + "bbox": [ + 312, + 175, + 545, + 217 + ], + "type": "text", + "content": "[9] Yulu Gan, Yan Bai, Yihang Lou, Xianzheng Ma, Renrui Zhang, Nian Shi, and Lin Luo. Decorate the newcomers: Visual domain prompt for continual test time adaptation, 2023. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 307, + 219, + 545, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 219, + 545, + 251 + ], + "spans": [ + { + "bbox": [ + 307, + 219, + 545, + 251 + ], + "type": "text", + "content": "[10] Yaroslav Ganin and Victor Lempitsky. Unsupervised domain adaptation by backpropagation. international conference on machine learning, 2015. 4" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 308, + 253, + 545, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 253, + 545, + 297 + ], + "spans": [ + { + "bbox": [ + 308, + 253, + 545, + 297 + ], + "type": "text", + "content": "[11] Amir Ghodrati, Babak Ehteshami Bejnordi, and Amirhossein Habibian. Frameexit: Conditional early exiting for efficient video recognition. arXiv: Computer Vision and Pattern Recognition, 2021. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 308, + 298, + 545, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 298, + 545, + 331 + ], + "spans": [ + { + "bbox": [ + 308, + 298, + 545, + 331 + ], + "type": "text", + "content": "[12] Chuan Guo, Geoff Pleiss, Yu Sun, and Kilian Q. Weinberger. On calibration of modern neural networks. international conference on machine learning, 2017. 4" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 308, + 332, + 545, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 332, + 545, + 365 + ], + "spans": [ + { + "bbox": [ + 308, + 332, + 545, + 365 + ], + "type": "text", + "content": "[13] Amirhossein Habibian, Davide Abati, Taco S. Cohen, and Babak Ehteshami Bejnordi. Skip-convolutions for efficient video processing. 2021. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 308, + 366, + 545, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 366, + 545, + 409 + ], + "spans": [ + { + "bbox": [ + 308, + 366, + 545, + 409 + ], + "type": "text", + "content": "[14] Dan Hendrycks and Thomas Dietterich. Benchmarking neural network robustness to common corruptions and perturbations. Proceedings of the International Conference on Learning Representations, 2019. 5" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 411, + 545, + 465 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 411, + 545, + 465 + ], + "spans": [ + { + "bbox": [ + 308, + 411, + 545, + 465 + ], + "type": "text", + "content": "[15] Yiping Kang, Johann Hauswald, Cao Gao, Austin Rovinski, Trevor Mudge, Jason Mars, and Lingjia Tang. Neurosurgeon: Collaborative intelligence between the cloud and mobile edge. architectural support for programming languages and operating systems, 2017. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 467, + 545, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 467, + 545, + 510 + ], + "spans": [ + { + "bbox": [ + 308, + 467, + 545, + 510 + ], + "type": "text", + "content": "[16] Mehrdad Khani, Pouya Hamadanian, Arash Nasr-Esfahany, and Mohammad Alizadeh. Real-time video inference on edge devices via adaptive model streaming. arXiv: Learning, 2020. 1, 3, 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 512, + 545, + 543 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 512, + 545, + 543 + ], + "spans": [ + { + "bbox": [ + 308, + 512, + 545, + 543 + ], + "type": "text", + "content": "[17] Dong-Hyun Lee. Pseudo-label: The simple and efficient semi-supervised learning method for deep neural networks. 2022. 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 545, + 545, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 545, + 545, + 589 + ], + "spans": [ + { + "bbox": [ + 308, + 545, + 545, + 589 + ], + "type": "text", + "content": "[18] Rui Li, Qianfen Jiao, Wenming Cao, Hau-San Wong, and Si Wu. Model adaptation: Unsupervised domain adaptation without source data. computer vision and pattern recognition, 2020. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 590, + 545, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 590, + 545, + 634 + ], + "spans": [ + { + "bbox": [ + 308, + 590, + 545, + 634 + ], + "type": "text", + "content": "[19] Xiaoqi Li, Jiaming Liu, Shizun Wang, Cheng Lyu, Ming Lu, Yurong Chen, Anbang Yao, Yandong Guo, and Shanghang Zhang. Efficient meta-tuning for content-aware neural video delivery, 2022. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 635, + 545, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 635, + 545, + 679 + ], + "spans": [ + { + "bbox": [ + 308, + 635, + 545, + 679 + ], + "type": "text", + "content": "[20] Jian Liang, Dapeng Hu, and Jiashi Feng. Do we really need to access the source data? source hypothesis transfer for unsupervised domain adaptation. international conference on machine learning, 2020. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 681, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 681, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 681, + 545, + 713 + ], + "type": "text", + "content": "[21] Ser-Nam Lim, Bharath Hariharan, Luming Tang, Menglin Jia, Bor-Chun Chen, Claire Cardie, and Serge Belongie. Visual prompt tuning. 2022. 2" + } + ] + } + ], + "index": 22 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 734, + 317, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 734, + 317, + 742 + ], + "spans": [ + { + "bbox": [ + 294, + 734, + 317, + 742 + ], + "type": "text", + "content": "12165" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 126 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 126 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 126 + ], + "type": "text", + "content": "[22] Jiaming Liu, Ming Lu, Kaixin Chen, Xiaqi Li, Shizun Wang, Zhaoqing Wang, Enhua Wu, Yurong Chen, Chuang Zhang, and Ming Wu. Overfitting the data: Compact neural video delivery via content-aware feature modulation, 2021. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 130, + 287, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 130, + 287, + 184 + ], + "spans": [ + { + "bbox": [ + 48, + 130, + 287, + 184 + ], + "type": "text", + "content": "[23] Pengfei Liu, Weizhe Yuan, Jinlan Fu, Zhengbao Jiang, Hiroaki Hayashi, and Graham Neubig. Pre-train, prompt, and predict: A systematic survey of prompting methods in natural language processing. arXiv: Computation and Language, 2021. 4" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 186, + 288, + 219 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 186, + 288, + 219 + ], + "spans": [ + { + "bbox": [ + 48, + 186, + 288, + 219 + ], + "type": "text", + "content": "[24] Yuang Liu, Wei Zhang, and Jun Wang. Source-free domain adaptation for semantic segmentation. computer vision and pattern recognition, 2021." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 221, + 287, + 263 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 221, + 287, + 263 + ], + "spans": [ + { + "bbox": [ + 48, + 221, + 287, + 263 + ], + "type": "text", + "content": "[25] Robert A. Marsden, Felix Wiewel, Mario Döbler, Yang Yang, and Bin Yang. Continual unsupervised domain adaptation for semantic segmentation using a class-specific transfer. 2022. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 266, + 287, + 299 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 266, + 287, + 299 + ], + "spans": [ + { + "bbox": [ + 48, + 266, + 287, + 299 + ], + "type": "text", + "content": "[26] Zhongnan Qu, Cong Liu, and Lothar Thiele. Deep partial updating: Towards communication efficient updating for ondevice inference. 2020. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 301, + 287, + 345 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 301, + 287, + 345 + ], + "spans": [ + { + "bbox": [ + 48, + 301, + 287, + 345 + ], + "type": "text", + "content": "[27] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2015. 5, 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 347, + 287, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 347, + 287, + 402 + ], + "spans": [ + { + "bbox": [ + 48, + 347, + 287, + 402 + ], + "type": "text", + "content": "[28] Christos Sakaridis, Dengxin Dai, and Luc Van Gool. ACDC: The adverse conditions dataset with correspondences for semantic driving scene understanding. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), October 2021. 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 404, + 287, + 448 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 404, + 287, + 448 + ], + "spans": [ + { + "bbox": [ + 48, + 404, + 287, + 448 + ], + "type": "text", + "content": "[29] Christian Simon, Masoud Faraki, Yi-Hsuan Tsai, Xiang Yu, Samuel Schulter, Yumin Suh, Mehrtash Harandi, and Manmohan Chandraker. On generalizing beyond domains in cross-domain continual learning. 2022. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 450, + 287, + 483 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 450, + 287, + 483 + ], + "spans": [ + { + "bbox": [ + 48, + 450, + 287, + 483 + ], + "type": "text", + "content": "[30] Antti Tarvainen and Harri Valpola. Mean teachers are better role models: Weight-averaged consistency targets improve semi-supervised deep learning results. Learning, 2017. 4" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 485, + 287, + 517 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 485, + 287, + 517 + ], + "spans": [ + { + "bbox": [ + 48, + 485, + 287, + 517 + ], + "type": "text", + "content": "[31] Eric Tzeng, Judy Hoffman, Kate Saenko, and Trevor Darrell. Adversarial discriminative domain adaptation. computer vision and pattern recognition, 2017. 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 519, + 287, + 540 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 519, + 287, + 540 + ], + "spans": [ + { + "bbox": [ + 48, + 519, + 287, + 540 + ], + "type": "text", + "content": "[32] Laurens Van Der Maaten. Barnes-hut-sne. arXiv preprint arXiv:1301.3342, 2013. 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 543, + 287, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 543, + 287, + 586 + ], + "spans": [ + { + "bbox": [ + 48, + 543, + 287, + 586 + ], + "type": "text", + "content": "[33] Riccardo Volpi, Diane Larlus, and Grégory Rogez. Continual adaptation of visual representations via domain randomization and meta-learning. computer vision and pattern recognition, 2020. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 589, + 287, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 589, + 287, + 621 + ], + "spans": [ + { + "bbox": [ + 48, + 589, + 287, + 621 + ], + "type": "text", + "content": "[34] Dequan Wang, Evan Shelhamer, Shaoteng Liu, Bruno A. Olshausen, and Trevor Darrell. Tent: Fully test-time adaptation by entropy minimization. Learning, 2021. 1, 2, 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 624, + 287, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 624, + 287, + 666 + ], + "spans": [ + { + "bbox": [ + 48, + 624, + 287, + 666 + ], + "type": "text", + "content": "[35] Qin Wang, Olga Fink, Luc Van Gool, and Dengxin Dai. Continual test-time domain adaptation. In Proceedings of Conference on Computer Vision and Pattern Recognition, 2022. 1, 2, 5, 6" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "type": "text", + "content": "[36] Tong Wang, Yousong Zhu, Chaoyang Zhao, Wei Zeng, Jinqiao Wang, and Ming Tang. Adaptive class suppression loss for long-tail object detection. computer vision and pattern recognition, 2021. 8" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 72, + 545, + 307 + ], + "type": "list", + "angle": 0, + "index": 21, + "blocks": [ + { + "bbox": [ + 307, + 72, + 545, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 72, + 545, + 116 + ], + "spans": [ + { + "bbox": [ + 307, + 72, + 545, + 116 + ], + "type": "text", + "content": "[37] Zifeng Wang, Zizhao Zhang, Chen-Yu Lee, Han Zhang, Ruoxi Sun, Xiaoqi Ren, Guolong Su, Vincent Perot, Jennifer Dy, and Tomas Pfister. Learning to prompt for continual learning. 2022. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 118, + 545, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 545, + 162 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 545, + 162 + ], + "type": "text", + "content": "[38] Jiangchao Yao, Feng Wang, Kunyang Jia, Bo Han, Jingren Zhou, and Hongxia Yang. Device-cloud collaborative learning for recommendation. knowledge discovery and data mining, 2021. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 163, + 545, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 163, + 545, + 227 + ], + "spans": [ + { + "bbox": [ + 307, + 163, + 545, + 227 + ], + "type": "text", + "content": "[39] Chongzhi Zhang, Mingyuan Zhang, Shanghang Zhang, Daisheng Jin, Qiang Zhou, Zhongang Cai, Haiyu Zhao, Shuai Yi, Xianglong Liu, and Ziwei Liu. Delving deep into the generalization of vision transformers under distribution shifts. arXiv: Computer Vision and Pattern Recognition, 2021. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 229, + 545, + 273 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 229, + 545, + 273 + ], + "spans": [ + { + "bbox": [ + 307, + 229, + 545, + 273 + ], + "type": "text", + "content": "[40] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Zwei Liu. Conditional prompt learning for vision-language models. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 274, + 545, + 307 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 274, + 545, + 307 + ], + "spans": [ + { + "bbox": [ + 307, + 274, + 545, + 307 + ], + "type": "text", + "content": "[41] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Learning to prompt for vision-language models. International Journal of Computer Vision (IJCV), 2022. 2" + } + ] + } + ], + "index": 20 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 734, + 318, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 734, + 318, + 742 + ], + "spans": [ + { + "bbox": [ + 294, + 734, + 318, + 742 + ], + "type": "text", + "content": "12166" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Clover_ Towards a Unified Video-Language Alignment and Fusion Model/eddf9a27-0fae-4139-acd3-5b79001d8460_content_list.json b/2023/Clover_ Towards a Unified Video-Language Alignment and Fusion Model/eddf9a27-0fae-4139-acd3-5b79001d8460_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..c235cfe667c111d5d6527f688004c6bf055d1475 --- /dev/null +++ b/2023/Clover_ Towards a Unified Video-Language Alignment and Fusion Model/eddf9a27-0fae-4139-acd3-5b79001d8460_content_list.json @@ -0,0 +1,1693 @@ +[ + { + "type": "text", + "text": "Clover: Towards A Unified Video-Language Alignment and Fusion Model", + "text_level": 1, + "bbox": [ + 99, + 128, + 870, + 151 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Jingjia Huang\\*", + "bbox": [ + 116, + 178, + 250, + 200 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yinan Li\\*", + "bbox": [ + 282, + 178, + 380, + 198 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Jiashi Feng", + "bbox": [ + 413, + 179, + 517, + 199 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Xinglong Wu", + "bbox": [ + 558, + 179, + 679, + 199 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Xiaoshuai Sun", + "bbox": [ + 720, + 179, + 857, + 196 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Rongrong Ji", + "bbox": [ + 428, + 204, + 540, + 224 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{\\text{心}}$ Key Laboratory of Multimedia Trusted Perception and Efficient Computing, Ministry of Education of China, Xiamen University, 361005, China", + "bbox": [ + 178, + 232, + 790, + 268 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ByteDance Inc, 100043, China", + "bbox": [ + 357, + 268, + 612, + 287 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "yinanlee@stu.xmu.edu.cn, {xssun, rrji} @ xmu.edu.cn", + "bbox": [ + 272, + 287, + 696, + 305 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{huangjingjia, jshfeng, wuxinglong} @bytedance.com", + "bbox": [ + 271, + 305, + 694, + 323 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 357, + 312, + 372 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Building a universal Video-Language model for solving various video understanding tasks (e.g., text-video retrieval, video question answering) is an open challenge to the machine learning field. Towards this goal, most recent works build the model by stacking uni-modal and cross-modal feature encoders and train it with pair-wise contrastive pre-text tasks. Though offering attractive generality, the resulted models have to compromise between efficiency and performance. They mostly adopt different architectures to deal with different downstream tasks. We find this is because the pair-wise training cannot well align and fuse features from different modalities. We then introduce Clover—a Correlated Video-Language pre-training method—towards a universal Video-Language model for solving multiple video understanding tasks with neither performance nor efficiency compromise. It improves cross-modal feature alignment and fusion via a novel tri-modal alignment pre-training task. Additionally, we propose to enhance the tri-modal alignment via incorporating learning from semantic masked samples and a new pair-wise ranking loss. Clover establishes new state-of-the-arts on multiple downstream tasks, including three retrieval tasks for both zero-shot and fine-tuning settings, and eight video question answering tasks. Codes and pre-trained models will be released at https://github.com/LeeYN-43/Clover.", + "bbox": [ + 76, + 388, + 473, + 765 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 796, + 209, + 811 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Video-Language pre-training (VidL) aims to learn generalizable multi-modal models from large-scale video-text", + "bbox": [ + 76, + 821, + 468, + 852 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "samples so as to better solve various challenging Video-Language understanding tasks, such as text-video retrieval [1,4,38,55] and video question answering [16,47,52]. Recent studies [9,11,23,24,49,56,58,61] have shown that VidL leads to significant performance improvement and achieves state-of-the-art results on various downstream text-video retrieval and video question answering (VQA) benchmarks.", + "bbox": [ + 500, + 358, + 890, + 464 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Though achieving encouraging performance, existing VidL models mostly adopt different architectures to deal with different downstream tasks. For the text-video retrieval tasks, they [2, 10, 11, 13, 31, 39] typically use two individual uni-modal encoders for processing video and text data separately, for the sake of retrieval efficiency. While for video question answering tasks, the models usually adopt the multi-modal joint encoder design to learn the association and interaction of different modalities.", + "bbox": [ + 496, + 470, + 892, + 606 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Building a unified model capable of solving various Video-Language tasks is a long-standing challenge for machine learning research. A few recent works [9,24] attempt to learn a unified VidL model for both tasks, which uses the multi-modal encoder to conduct text-video retrieval. However, the model requires an exhaustive pair-wise comparison between the query texts and gallery videos. Given $N$ text queries and $M$ category videos, the computation complexity of the multi-modal encoder model would be $O(NM)$ , which makes it infeasible for large-scale videotext retrieval applications. Another straightforward solution is to simply combine the uni-modal and multi-modal encoders (Fig. 1 (a)), and perform the retrieval and VQA tasks through the uni-modal and multi-modal encoders respectively. Its computation complexity for retrieval tasks is only $O(N + M)$ . However, the experiment results in Fig.1 (c) show that, without a carefully designed correlating mechanism between these two types of encoders, the simple combination i.e., COMB yields a compromised per", + "bbox": [ + 496, + 613, + 893, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Equal Contribution in alphabetical order.", + "bbox": [ + 93, + 862, + 318, + 875 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "$\\dagger$ Work done when interning at ByteDance Inc.", + "bbox": [ + 96, + 875, + 341, + 887 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "$\\ddagger$ Corresponding Author.", + "bbox": [ + 96, + 887, + 225, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "14856", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "formance compared with the models i.e. IND individually designed for retrieval and VQA.", + "bbox": [ + 76, + 90, + 468, + 119 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this work, we aim to address the above issues and build a unified pre-training model that attains high efficiency and performance simultaneously. We observe: (i) well aligning the features of the video and text from the same data pair is important for text-video matching; (ii) effectively fusing video and text features into unified representations is critical for video-text understanding. However, existing pre-training strategies that rely on either simple supervised or contrastive pre-text tasks hardly achieve promising feature alignment and fusion capability simultaneously. Motivated by these observations, we develop a new VidL method from these two aspects. Specifically, we propose Correlated Video-Language pre-training (Clover), a VidL method that not only unifies Video-Language alignment and fusion, but also makes them mutually boosted. The fused multi-modal representation contains richer context information than the uni-modal representations [11,35]. As an intermediate modality between video and text, the multi-modal representations are good anchors for cross-modality alignment. Meanwhile, keeping the fused representation closer to the uni-modal representation containing consistent semantic information and away from the inconsistent one will enhance the learning of semantic information in the fused modality. Therefore, we propose the Tri-Modal Alignment (TMA) to get Video-Language alignment and fusion mutually boosted, which takes the alignment between the multimodal representation and text/video representations as an auxiliary objective. We note that since the tri-modal alignment is well compatible with the classical pre-training tasks [3,7,35] e.g., Masked Language Modeling, its computation overhead is negligible. To help the model maintain fine-grained discriminative capability while improving its generalizability, we further introduce a pair-wise ranking loss that urges the model to be aware of the concept missing in masked samples compared to original samples.", + "bbox": [ + 76, + 123, + 472, + 654 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Extensive experiments are conducted on multiple downstream tasks, including three retrieval tasks with different experimental setups (i.e. zero-shot and fine-tune) and eight video question answering tasks. The results demonstrate that Clover is able to get the cross-modal fusion and alignment capability mutually improved, and consistently outperforms current SOTAs on various downstream tasks. It achieves an average performance improvement of $4.9\\%$ and $8.7\\%$ Recall@10 score on the zero-shot and fine-tune settings of the three downstream retrieval datasets, while the average accuracy improvement over current SOTAs is $2.3\\%$ on the eight video question answering datasets.", + "bbox": [ + 76, + 656, + 468, + 837 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In summary, we make the following contributions: (1) we introduce Clover, a pre-training method achieving unified Video-Language alignment and fusion model that can be easily transferred to various downstream video under", + "bbox": [ + 76, + 839, + 468, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "standing tasks while attaining both high efficiency and performance; (2) we propose a novel tri-modal alignment pretraining task, which correlates the uni-modal encoder and multi-modal encoder to get them mutually boosted.", + "bbox": [ + 498, + 90, + 890, + 151 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 165, + 640, + 181 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Video-Language pre-training for text-video retrieval. Existing pre-training methods for text-video retrieval can be mainly divided into two categories. The first kind of method employs two individual encoders to embed video and text, and project them into a common latent space [2,10,11,13,31,39,42,50,54,59]. Then, they leverage a contrastive objective [48] to align the paired representations. Thanks to their high efficiency, this category of methods is widely used in real-world applications. The other kind of method utilizes a joint multi-modal encoder to learn comprehensive representation for a given video-text pair, and predict whether the video and text are matched or not via a binary classifier [9,23,25,45,53,63]. Despite the good performance they achieved, they need to exhaustively pair all the videos and texts, and feed them into the model during inference, making them highly inefficient.", + "bbox": [ + 498, + 190, + 890, + 431 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Video-Language pre-training for video question answering. Video question answering aims to automatically answer natural language questions given a context video, which requires a joint understanding of both the video and language. Previous works mainly focus on designing better spatio-temporal attention networks [8,16,26,29,52,62] and question-video relation networks [15, 19, 22, 41] to extract more accurate multi-modal representations. Recently, many works leverage transformer-based models to build an additional multi-modal encoder upon the uni-modal encoders [9, 24, 43] or input image patches and word tokens [58, 63] directly to fuse cross-modality features. The multi-modal encoder is typically driven by the pre-training tasks e.g. Masked Language Modeling [7] and Visual-Text Matching [46], to perform token-level cross-modal fusion.", + "bbox": [ + 498, + 433, + 890, + 657 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Video-Language pre-training for multiple downstream tasks. There are a few recent works trying to construct a unified pre-trained model for multiple downstream tasks. ActBert [63], clipBert [23], VIOLET [9] and ALPRO [24] utilize a cross-modal encoder for both retrieval and VQA tasks. They take video-text retrieval as a binary classification task and construct the classifier upon a visual-text matching head, which brings about extreme computation overload. While All-in-One [49] and HD-VILA [56] proposes a transformer-based unified backbone architecture for joint-modal representation learning, they require a large amount of data for pre-training. HERO [25] and VLM [53] design a task-agnostic model that can be finetuned on multiple downstream tasks. However, they are not end-to-end trainable and need pre-extract video features, which limits the performance of the model.", + "bbox": [ + 498, + 659, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "14857", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/e10e97ed4a81136d4a17e33218dc1c503bf507b171e29ed863dbd8c1937f6ff1.jpg", + "image_caption": [ + "Figure 1. Different attempts on achieving a universal VidL model. (a) Naive combination by stacking the cross-modal encoder upon the uni-modal encoders with weak association (COMB). (b) Our proposed Clover pre-trained model that better correlates the cross-modal encoder and the uni-modal encoders via Tri-Modal Alignment. (c) Comparisons of the transfer performance on MSRVTT [60] (zero-shot) and TGIF-FrameQA [16] datasets. Benefiting from the proposed tri-modal alignment, our proposed universal model outperforms both the individually designed models (IND) and COMB." + ], + "image_footnote": [], + "bbox": [ + 133, + 44, + 330, + 224 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/ec9fc763dad1a29fcec5d1414d161e7394bd05e4dca6bd2e9ea9edf152f08099.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 346, + 44, + 539, + 222 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/9ce533bff35816aad9d5ac584995945adccd7dc2128ac5f4e5f208dc4b488567.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 540, + 65, + 836, + 222 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Unlike the aforementioned methods, our Clover is a fully end-to-end VidL model that can be easily transferred to various downstream tasks while attaining both high efficiency and performance.", + "bbox": [ + 75, + 319, + 468, + 378 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Note that there are some methods focus on transferring the pre-trained CLIP model to downstream tasks and achieve significant performance [6, 14, 17, 21, 30, 37, 51, 57]. The CLIP model is pretrained with 400M image-text data pairs. In contrast, our Clover focuses on presenting a better video-text pretrained model, and only leverages 5M data pairs for pretraining.", + "bbox": [ + 75, + 380, + 468, + 484 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 76, + 500, + 169, + 516 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Motivation and Overview", + "text_level": 1, + "bbox": [ + 76, + 525, + 308, + 540 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Recent years have witnessed a significant progress in Video-Language pre-training. Nevertheless, building a universal Video-Language model for solving various video understanding tasks (e.g., text-video retrieval, video question answering) is still a challenging problem. The key to the problem lies in how to unify the representation learning for cross-modal alignment and fusion efficiently.", + "bbox": [ + 75, + 549, + 468, + 655 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Cross-modal alignment aims to learn projection functions $f(\\cdot)$ and $g(\\cdot)$ that project videos and texts into a common embedding space where the similarity between the video and text with consistent semantic information is maximized and otherwise minimized. Taking a video $V$ as the anchor, the learning objective of cross-modal alignment is formulated as:", + "bbox": [ + 75, + 655, + 468, + 758 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\arg \\max _ {f, g} [ \\mathrm {s} (f (V), g (T ^ {+})) - \\mathrm {s} (f (V), g (T ^ {-})) ], \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 99, + 771, + 468, + 797 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $T^{+}, T^{-}$ represent texts that are semantically consistent and inconsistent with $V$ , respectively. The function $\\mathrm{s}(\\cdot, \\cdot)$ measures the dot-product similarity between two embeddings. A model with strong cross-modal alignment capability is desired in the video-retrieval task, which requires matching the semantically aligned videos and texts.", + "bbox": [ + 75, + 809, + 468, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Cross-modal fusion aims to integrate the correlation and interaction carried by the video and text modalities into a unified multi-modal embedding. Specifically, it can be formulated as learning a function $Fusion(\\cdot, \\cdot)$ that takes inputs of different modals and outputs a unified representation $M = Fusion(V, T)$ , which is then used to solve downstream tasks like VQA.", + "bbox": [ + 498, + 319, + 893, + 426 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Existing pre-training strategies rely on either simple supervised or contrastive pre-text tasks that hardly achieve feature alignment and fusion simultaneously. To better correlate the cross-modal alignment and fusion, we propose the Correlated Video-Language pre-training method (Clover) with three key pre-training innovations (Sec. 3.2): tri-modal alignment; pair-wise ranking loss; and semantic enhanced masked language modeling. The classical masked language modeling task is also incorporated into our method, which is able to boost the generalizability of the model as well as the interaction between visual and language.", + "bbox": [ + 496, + 444, + 893, + 611 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Architecture. We briefly explain the architecture of Clover here, which is shown in Fig. 2 as well. It consists of three components: a video encoder, a text encoder and a multi-modal encoder. Following [9], we use VideoSwin Transformer [32] as our video encoder. Given a video $V$ , it outputs a sequence of video embeddings: $V_{e} = \\{v_{1},\\dots,v_{K}\\} \\subset \\mathbb{R}^{D}$ , where $K$ is the number of flattened patches. The text encoder of Clover is a 12-layer bidirectional transformer encoder model [7]. Given an input text sentence $T$ , the encoder outputs an embedding sequence $T_{e} = \\{t_{\\mathrm{cls}},t_{1},\\dots,t_{L - 1}\\} \\subset \\mathbb{R}^{D}$ , where $t_\\mathrm{cls}$ indicates the embedding of the [CLS] token. We employ a 3-layer bidirectional Transformer encoder to learn the fused cross-modal representation of a video-text pair. The multi-modal encoder takes the concatenation of video and text embeddings as input, and outputs the fused multi-modal embeddings $M_{e} = \\{m_{v_{1}},\\dots,m_{v_{K}},m_{\\mathrm{CLS}},m_{t_{1}},\\dots,m_{t_{L - 1}}\\}$ , where $m \\in \\mathbb{R}^D$ .", + "bbox": [ + 496, + 628, + 893, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "14858", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/e65fcad67801b10efb9b33c1cd0706b0c71eb46ae63cba5b6967603291a25045.jpg", + "image_caption": [ + "(a) Model Architecture" + ], + "image_footnote": [], + "bbox": [ + 138, + 37, + 529, + 270 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/f5a4f991b0ee9d61c942d9c19a6013a766a7fcf5625ed33e2ea302d77e7375b4.jpg", + "image_caption": [ + "Tri-modal Alignment", + "Pair-wise Ranking" + ], + "image_footnote": [], + "bbox": [ + 550, + 54, + 810, + 169 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/b4d05872044a0c430888f7696323550bfbb071c6f1f9ad4618af462785ddb845.jpg", + "image_caption": [ + "(b) Pre-training Tasks", + "Figure 2. Overview of Clover (a) Model architecture. (b) Tri-modal alignment pre-training task with masked samples and the pair-wise ranking. “×” indicates the embedding of the token [MASK]. For more detail on model architecture and pre-training tasks refers to Sec. 3." + ], + "image_footnote": [], + "bbox": [ + 555, + 199, + 656, + 273 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/78b295b1c86e26c840f9fa5f51b179df781aba52a90ad1ed12dd6dab20efa22a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 689, + 199, + 792, + 273 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Tri-Modal Alignment", + "text_level": 1, + "bbox": [ + 76, + 349, + 277, + 366 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Existing VidL pre-training methods typically align the embeddings of data from different modalities through pairwise contrastive learning. Differently, we propose a TriModal Alignment (TMA) learning task that not only enforces the video and text modalities to be aligned but also encourages them to well align with the third modality, i.e., their fusion modality. In TMA, as an intermediate modality between video and text, the fused multi-modal representations act as anchors for cross-modality alignment, which reduces the difficulty of the direct alignment between video and text. Meanwhile, it keeps the fused representation closer to the uni-modal representation containing consistent semantic information and away from the others, so as to enhance the learning of semantic information in the fused representation.", + "bbox": [ + 75, + 376, + 468, + 602 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As shown in Fig. 2 (a), given an input video-text pair $\\langle V, T \\rangle$ and their embeddings $\\langle V_e, T_e \\rangle$ from the uni-modal encoder, we mask some regions in $V$ to make a masked video with embedding $V_m$ , and similarly mask some tokens in $T$ to make a masked text embedding $T_m$ . Then, we pair the incomplete samples with the complete samples as $\\langle V_m, T_e \\rangle$ and $\\langle V_e, T_m \\rangle$ . We adopt multi-modal encoder to get fused multi-modal embedding sequence $M_{V_m}$ and $M_{T_m}$ of $\\langle V_m, T_e \\rangle$ and $\\langle V_e, T_m \\rangle$ , respectively. We denote the embedding of [CLS] token in $M_{V_m}$ and $M_{T_m}$ as $M_{V_m f}$ and $M_{T_m f}$ .", + "bbox": [ + 76, + 606, + 468, + 773 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For the convenience of description, we use superscripts $i$ and $j$ to index the data sample, and introduce TMA from the perspective of video and language, respectively. Firstly, for a video representation $V_{e}^{i}$ , besides its associated text embedding $T_{e}^{i}$ and $T_{m}^{i}$ , we also consider its associated fusion embedding $M_{V_{mf}}^{i}$ as its positive pair—that should be pushed closer within the embedding space. To align these tri-modal representations, we propose a novel exclusivity loss computed within a batch of $B$ samples as follows:", + "bbox": [ + 75, + 773, + 470, + 901 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} L _ {v} = - \\sum_ {i = 1} ^ {B} \\left[ \\log \\frac {e ^ {s (V _ {e} ^ {i} , T _ {e} ^ {i}) / \\tau}}{e ^ {s (V _ {e} ^ {i} , T _ {e} ^ {i}) / \\tau} + Z} + \\log \\frac {e ^ {s (V _ {e} ^ {i} , T _ {m} ^ {i}) / \\tau}}{e ^ {s (V _ {e} ^ {i} , T _ {m} ^ {i}) / \\tau} + Z} \\right. \\\\ \\left. + \\log \\frac {e ^ {s (V _ {e} ^ {i} , M _ {V _ {m} f} ^ {i}) / \\tau}}{e ^ {s (V _ {e} ^ {i} , M _ {V _ {m} f} ^ {i}) / \\tau} + Z} \\right], \\mathrm {w h e r e} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 514, + 368, + 874, + 446 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nZ = \\sum_ {j \\neq i} ^ {B} \\left[ e ^ {s \\left(V _ {e} ^ {i}, T _ {e} ^ {j}\\right) / \\tau} + e ^ {s \\left(V _ {e} ^ {i}, T _ {m} ^ {j}\\right) / \\tau} + e ^ {s \\left(V _ {e} ^ {i}, M _ {V _ {m f}} ^ {j}\\right) / \\tau} \\right]. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 519, + 448, + 890, + 497 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Here $\\tau$ is a temperature scalar, and $s(\\cdot ,\\cdot)$ is the dot-product similarity function to measure the degree of alignment between different modalities. In each term of Eq. (2), we exclude the other positive pairs when performing contrastive learning on one positive pair. Thus the intra-suppression among positive pairs is effectively avoided, and the video-/text-/fusion-modal representations are better aligned. For example, in the first term of Eq. (2), $e^{s(V_e^i,M_{Vmf}^i) / \\tau}$ and $e^{s(V_e^i,T_m^i) / \\tau}$ are excluded from the denominator, which prevents the model from yielding the sub-optimal solution that gets $T_{e}^{i}$ closer to $V_{e}^{i}$ while pushes $T_{m}^{i}$ and $M_{Vmf}^{i}$ away from $V_{e}^{i}$ . Meanwhile, to align the video to the text and fusion modalities, we have the following objective:", + "bbox": [ + 496, + 498, + 892, + 688 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} L _ {v ^ {\\prime}} = - \\sum_ {i = 1} ^ {B} \\left[ \\log \\frac {e ^ {s (T _ {e} ^ {i} , V _ {e} ^ {i}) / \\tau}}{\\sum_ {j = 1} ^ {B} e ^ {s (T _ {e} ^ {i} , V _ {e} ^ {j}) / \\tau}} + \\log \\frac {e ^ {s (T _ {m} ^ {i} , V _ {e} ^ {i}) / \\tau}}{\\sum_ {j = 1} ^ {B} e ^ {s (T _ {m} ^ {i} , V _ {e} ^ {j}) / \\tau}} \\right] \\\\ \\left. + \\log \\frac {e ^ {s \\left(M _ {V _ {m f}} ^ {i} , V _ {e} ^ {i}\\right) / \\tau}}{\\sum_ {j = 1} ^ {B} e ^ {s \\left(M _ {V _ {m f}} ^ {i} , V _ {e} ^ {j}\\right) / \\tau}} \\right]. \\tag {3} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 500, + 693, + 890, + 787 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The above two losses (2) and (3) are summed up to give the tri-modal alignment objective w.r.t. the visual modality: $L_{V} = L_{v} + L_{v^{\\prime}}$ . Similarly, from the perspective of text modality, we define the tri-modal alignment objective $L_{T} = L_{t} + L_{t^{\\prime}}$ w.r.t. the text modality. Then, we obtain the final loss of tri-modal alignment $L_{TmA} = L_{V} + L_{T}$ .", + "bbox": [ + 496, + 789, + 893, + 880 + ], + "page_idx": 3 + }, + { + "type": "page_footnote", + "text": "For the detail of $L_{T}$ , please refer to the supplementary material.", + "bbox": [ + 516, + 886, + 861, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "14859", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To get the masked samples for TMA, we conduct masking over the videos and texts to form masked samples using the following two strategies.", + "bbox": [ + 76, + 90, + 468, + 136 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Video-block masking strategy. Objects in a video often appear at similar spatial locations across consecutive frames, which compose a tube area in the spatial-temporal domain. In order to construct incomplete video samples with less information leakage, we extend the block-wise mask [3] to video via performing a block-wise mask at the same position in all video frames. In this paper, we randomly replace $20\\%$ patches with a learnable mask token to obtain $V_{m}$ .", + "bbox": [ + 75, + 136, + 467, + 271 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Semantic text masking strategy. The detailed semantic information in the text basically lies in the verb, noun and adjective words. To facilitate learning their representations, we construct incomplete text samples by randomly masking some verbs, nouns and adjectives in the sentence. Specifically, given a sentence, we use a part-of-speech tagger [33] to tag each word. Then we pick the verb phrases and nouns, and replace $30\\%$ of them with a special [MASK] token. Note that to avoid drastic semantic changes, we do not mask the auxiliary verb like have, should, will, would, etc. With the masking strategy, when some key elements are masked from the texts, it would form an expression carrying partial information rather than completely changing the information expressed.", + "bbox": [ + 75, + 272, + 467, + 483 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Training Objective", + "text_level": 1, + "bbox": [ + 76, + 493, + 256, + 508 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Pair-wise ranking. In TMA, we take the masked pairs $\\langle V_e, T_m \\rangle$ and $\\langle V_m, T_e \\rangle$ as positive pairs, considering that masked samples would still carry partial meaningful information though some others are missed. For instance, as shown in Fig. 2 (a), given the text \"A black swan swimming in a calm lake\", \"swimming\" and \"lake\" are masked with [MASK] tokens, generating the masked pair $\\langle V_e, T_m \\rangle$ . Although some information is removed, the masked sentence still contains other concepts that make it partially matched with the video, such as the \"black swan\". Therefore, in contrastive learning, $V_m$ and $T_m$ can be considered pseudo positive candidates for $T$ and $V$ . However, compared to the original pair $\\langle V_e, T_e \\rangle$ , we consider that the semantic consistency in the masked pairs should be weaker. Based on this prior, we further propose a pair-wise ranking loss:", + "bbox": [ + 75, + 515, + 467, + 724 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} L _ {r a n k} = \\max \\left(0, - \\left(\\operatorname {s i m} \\left(V _ {e}, T _ {e}\\right) / \\tau - \\operatorname {s i m} \\left(V _ {e}, T _ {m}\\right) / \\tau\\right) + \\lambda\\right) \\\\ + \\max \\left(0, - \\left(\\operatorname {s i m} \\left(V _ {e}, T _ {e}\\right) / \\tau - \\operatorname {s i m} \\left(V _ {m}, T _ {e}\\right) / \\tau\\right) + \\lambda\\right), \\tag {4} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 84, + 734, + 467, + 779 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\lambda > 0$ is a margin hyper-parameter. Eq. (4) urges the model to be aware of the gap of semantic consistency between $\\langle V_e, T_m \\rangle / \\langle V_m, T_e \\rangle$ and $\\langle V_e, T_e \\rangle$ brought about by concepts missing in the masked pairs. With the pairwise ranking objective, our model is able to maintain fine-grained perceptual capability while improving its generalizability.", + "bbox": [ + 75, + 779, + 467, + 883 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Semantic enhanced masked language modeling. Masked", + "bbox": [ + 76, + 885, + 467, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "language modeling (MLM) is a classical pre-training task in VidL, which promotes the interaction between different modalities in the cross-modal encoder. Combining the classical MLM with our semantic text masking strategy, we present the Semantic Enhanced Masked Language Modeling task, which facilitates the representation learning on the key concepts, i.e., verb, noun and adjective words. Besides, considering the class-imbalance between different words, we use focal loss [28] instead of the traditional cross-entropy loss to improve the MLM loss. We apply the MLM loss to the reconstruction of the masked text tokens i.e., $m_{t_j} \\in M_{T_m}$ , where $t_j = [Mask]$ . The MLM loss is defined as:", + "bbox": [ + 496, + 90, + 890, + 286 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nL _ {m l m} = - \\frac {1}{B} \\sum_ {i = 1} ^ {B} \\sum_ {m _ {t _ {j}} \\in M _ {T _ {m}} ^ {i}} [ (1 - p _ {t _ {j}} ^ {i}) ^ {\\gamma} p _ {t _ {j}} ^ {i} ] \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 544, + 296, + 890, + 344 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $p_{t_j}^i$ denote the predicted probability distribution of the masked token $t_j$ in $i$ th sentence in the batch, and $\\gamma$ is a hyper-parameter. Finally, the overall pre-training objective of Clover is:", + "bbox": [ + 496, + 356, + 890, + 416 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nL = L _ {T m A} + L _ {r a n k} + L _ {m l m}. \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 593, + 431, + 890, + 446 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 500, + 473, + 632, + 489 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Experiment Setup", + "text_level": 1, + "bbox": [ + 500, + 498, + 674, + 515 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Pre-training datasets. Following recent work [2, 11, 24], we jointly pre-train our Clover on a video dataset WebVid2M [2] with 2.5M video-text pairs and an image dataset Google Conceptual Captions (CC3M) [44] with 3.3M image-text pairs (we only obtain 2.8M image-text pairs in CC3M due to image url broken). During pretraining, we treat image data as one frame video data.", + "bbox": [ + 496, + 522, + 890, + 627 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Downstream tasks. We evaluate our proposed models on the following downstream tasks. (a) Text-to-Video Retrieval on MSRVTT [55], LSMDC [38] and DiDeMo [1]. For DiDeMo, we follow [23, 31] and evaluate Clover on the paragraph-to-video retrieval, where text sentences for each video are concatenated together as one text query. We do not use the ground-truth proposal for fair comparison with previous works; (b) Multiple-choice QA on TGIF-Action [16], TGIF-Transition [16], MSRVTT-MC [60] and LSMDC-MC [47]; (c) Open-Ended QA on TGIF-Frame [16], MSRVTT-QA [52], MSVD-QA [52] and LSMDC-FiB [38]. For retrieval tasks, we only use the two uni-modal encoders of Clover for fine-tuning and inference. We adopt the two encoders to get the video and text embeddings, and calculate their cosine similarity for retrieval. For QA tasks, we use all the modules for fine-tuning and inference. More details on these datasets and their evaluation usage are provided in the supplementary material.", + "bbox": [ + 496, + 628, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "14860", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/32b0f0c8294d9d1377aca281733e777483cc0b9ad5df83854884d35996023f79.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodPre-training datasetDiDeMoMSRVTTLSMDC
R@1R@5R@10MedRR@1R@5R@10MedRR@1R@5R@10MedR
Fine-tune
clipBert [23]COCO [5], VG [20]20.448.060.8622.046.859.96----
Frozen [2]W2M+C3M31.059.872.4331.059.570.5315.030.839.820
VIOLET [9]W2M+C3M+Y180M32.662.874.7-34.563.073.4-16.136.641.2-
HD-VILA [56]HDV100M28.857.469.1435.665.378.0317.434.144.115
ALPRO [24]W2M+C3M35.967.578.8333.960.773.23----
TMVM [27]W2M+C3M36.564.975.4336.264.275.7317.837.145.913.5
All-in-1 [49]W2M+H100M32.761.473.5-37.968.177.1-----
OA-Trans [50]W2M+C3M34.864.475.1335.863.476.5318.234.343.718.5
MILES [12]W2M+C3M36.663.974.0337.763.673.8317.835.644.115.5
MCQ [11]W2M+C3M37.062.273.9337.664.875.1317.935.444.515
Clover (ours)W2M+C3M50.176.785.6140.569.879.4224.844.054.58
Zero-shot
MIL-NCE [39]H100M----9.924.032.429.6----
Frozen [2]W2M+C3M21.146.056.2718.739.651.6109.322.030.151.0
VIOLET [9]W2M+C3M+Y180M23.549.859.8-25.949.559.7-----
ALPRO [24]W2M+C3M23.847.357.9624.144.755.48----
OA-Trans [50]W2M+C3M23.550.459.8623.447.555.68----
MILES [12]W2M+C3M27.250.363.6526.147.256.9711.124.730.650.7
MCQ [11]W2M+C3M25.650.661.1526.046.456.4712.225.932.242
Clover (ours)W2M+C3M29.555.266.3426.449.560.0614.729.238.224
", + "bbox": [ + 86, + 87, + 885, + 414 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/b09de1abe8278fb0d246f116ea7bdda2ddf0af280e62d894b23ed1f4072cff7a.jpg", + "table_caption": [ + "Table 1. Text-to-video retrieval performance comparison under fine-tune and zero-shot setups. Here higher R@k (Recall K) and lower MedR (Median Recall) indicate better performance. W2M, C3M, H100M, HDV100M, Y180M are short for WebVid2M [2], CC3M [44], HowTo100M [40], HD-VILA-100M [56], YT-Temporal-180M [61], respectively." + ], + "table_footnote": [], + "table_body": "
MethodPre-training datasetTGIFMSRVTTLSMDCMSVD
ActionTransitionFrameMCQAMCFiBQA
clipBert [23]COCO, VG82.887.860.388.237.4---
JuskAsk [58]HTVQA69M [58]----41.5--46.3
ALPRO [24]W2M+C3M----42.1--45.9
All-in-1 [49]W2M+H100M92.794.364.292.042.983.1-47.9
VIOLET [9]W2M+C3M+Y180M92.595.768.991.943.982.853.747.9
MERLOT [61]Y180M94.096.269.590.943.181.752.9-
Clover (ours)W2M+C3M95.098.271.695.244.183.754.152.4
", + "bbox": [ + 161, + 465, + 807, + 601 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 2. Performance comparison on transferring to downstream video question answering tasks.", + "bbox": [ + 196, + 604, + 769, + 619 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation details. Following [9], we initialize the video encoder with a VideoSwin-Base [32] model pretrained on Kinetics-400 [18]. The text encoder is initialized from pre-trained Bert-Base [7]. The multi-modal encoder is initialized from the first three layers of the pre-trained Bert-Base model. We train Clover end-to-end during both pretraining and fine-tuning. We pre-train Clover for 40 epochs, using a batch size of 1024 on 64 NVIDIA A100 GPUs. We use AdamW [34] optimizer with a weight decay 0.005 and betas (0.9, 0.98). The learning rate is first warmed-up by 4 epochs to 5e-5 and then decays following a cosine annealing decay schedule. We resize all the video frames to $224 \\times 224$ and split each frame into patches with a size of $32 \\times 32$ . We choose hyper-parameter $\\tau = 0.05$ , $\\lambda = 5$ and $\\gamma = 2$ . We set dimension $D = 768$ in our model. For each video, we randomly sample 8 frames while preserving their order in-between. For fine-tuning on retrieval tasks, we only fine", + "bbox": [ + 75, + 631, + 472, + 888 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "tune the uni-modal encoders of Clover with InfoNCE [48] loss. For fine-tuning on video QA task, we add a simple MLPs that takes the multi-modal [CLS] embedding as input for classification, and optimize the whole Clover model with cross-entropy loss. During fine-tuning, we follow the conventional set-up in [9, 36], and all the fine-tuning experiments are performed on 8 NVIDIA A100 GPUs.", + "bbox": [ + 496, + 631, + 890, + 737 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Comparing to State-of-the-art", + "text_level": 1, + "bbox": [ + 498, + 753, + 766, + 770 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Text-to-video retrieval. Tab. 1 shows the retrieval results on the DiDeMo [1], MSRVTT [55] and LSMDC [38] datasets with both zero-shot and fine-tuning settings. Our Clover surpasses previous works on all the datasets by a large margin. The significant performance gain under zero-shot evaluation demonstrates its stronger generalization ability. Specifically, Clover brings $4.9\\%$ gain in R@10 over MCQ [11], the current SOTA method on re", + "bbox": [ + 496, + 779, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "14861", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/c151788e4250461e372350677b7b20477bedcb6c5ecf8c73c413ad11346ea0a2.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodDiDeMoMSRVTTLSMDC-MCTGIF-Frame
R@1R@5R@10MedRR@1R@5R@10MedRAccAcc
Baseline22.647.958.2719.841.751.11078.868.9
+ TMA24.749.760.0622.742.252.2980.069.3
+ TMA+SM25.349.560.5622.542.752.0980.569.4
+ TMA+SM+RankL26.451.161.3523.443.352.4980.769.7
", + "bbox": [ + 122, + 88, + 851, + 188 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/d29dd64ea3adbd4f6e0d052e97885b173ebc857608f71f872b430560f5aa986d.jpg", + "table_caption": [ + "Table 3. Effects of our pre-training tasks for Clover. We report zero-shot text-video retrieval performance on DiDeMo and MSRVTT, and fine-tune video QA performance on LSMDC-MC and TGIF-Frame. TMA, SM and RankL: tri-modal alignment, semantic masking strategy and pair-wise ranking loss." + ], + "table_footnote": [], + "table_body": "
MethodAveraged similarity scores of positive pairsAveraged similarity margin between positive and negative samples
w/o TMA0.560.30
w TMA0.650.35
", + "bbox": [ + 84, + 247, + 465, + 301 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 4. Affect of TMA on the video and text representations in learned embedding space. The experiment is conducted on MSRVTT-1kA test set under zero-shot setting.", + "bbox": [ + 75, + 311, + 468, + 354 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "trieval tasks with the same pre-train data, averaged on the three datasets. Moreover, Clover outperforms VIOLET [9], though VIOLET adopts a considerably larger pre-training dataset (YTT180M [61] of 180M visual-text pairs vs. WebVid2M+CC3M of only 5.5M samples). Clover also presents superior performance when fine-tuned on the downstream datasets, compared with the SOTAs. On all three datasets, Clover surpasses previous works by a large margin across all the metrics and obtains an average improvement of $8.7\\%$ on R@10. Moreover, different from some previous work [9,23,24] that adopts a joint multi-modal encoder and requires exhaustively pairing every video and text during retrieval, Clover directly deploys the video encoder and text encoder for retrieval tasks, making it much more efficient.", + "bbox": [ + 73, + 364, + 468, + 575 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Video question answering. Tab. 2 reports the results of Clover and current SOTAs on four video QA datasets. Though using much less pre-training data, Clover outperforms another unified VidL model i.e., VIOLET [9] significantly. Moreover, compared to the current SOTA method MERLOT [61], Clover brings improvements of $1.0\\%$ on TGIF-Action, $2.0\\%$ on TGIF-Transition, $2.1\\%$ on TGIF-Frame, $4.3\\%$ on MSRVTT-MC, $1.0\\%$ on MSRVTT-QA, $2.0\\%$ on LSMDC-MC and $1.2\\%$ on LSMDC-FiB. Note that JustAsk [58] and MERLOT [61] are specifically designed for video QA and trained on orders of magnitude larger datasets (e.g., HTVQA69M [58], YTT180M [61]). The experiment results again confirm the superiority of Clover.", + "bbox": [ + 75, + 575, + 470, + 773 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Analysis", + "text_level": 1, + "bbox": [ + 76, + 785, + 181, + 801 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We conduct ablation experiments on two retrieval datasets (DiDeMo and MSRvTT) and two Video QA datasets (TGIF-Frame and LSMDC-MC). Due to the computational resource limit, we randomly sample 1 million video-text pairs from WebVid2M [2] to build the WebVid1M subset for model pre-training under all the ablation stud", + "bbox": [ + 75, + 810, + 470, + 902 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "ies. We keep the model architecture unchanged but pretrain the model with only MLM and InfoNCE losses as a baseline, which is similar to the baseline adopted in VIOLET [9]. The only difference is that we replace the ITM applied to the cross-modal encoder outputs in VIOLET [9] with InfoNCE applied to the outputs of uni-modal encoders and separate the text encoder from the cross-modal encoder. All the ablation experiments are conducted with batch size 1024 on 32 NVIDIA A100 GPUs.", + "bbox": [ + 496, + 250, + 890, + 383 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Effect of tri-modal alignment. Tri-modal alignment (TMA) aims to better correlate cross-modal alignment and fusion. To evaluate the contribution of TMA, we remove the pair-wise ranking objective and semantic masking strategy, and employ the classical MLM task as in the baseline method. Compared with the baseline, TMA explicitly associates the outputs of uni-modal encoders with the outputs of the multi-modal encoder. As shown in Tab. 3, the model trained with TMA outperforms the baseline on both retrieval and video QA, demonstrating its effectiveness. For a better understanding of the effect of TMA, we further report averaged cosine similarity of videos and texts in MSRVTT dataset (in zero-shot setting) calculated by models trained with/without TMA, respectively. For a fair comparison, we select all video queries (197 in total) that were successfully responded to by both models. We recall 100 results for each query, and take the ground-truth pairs as positive and others as negative. As shown in Tab. 4, the model armed with tri-modal alignment assigns higher similarity scores to positive pairs compared to the model without it. Meanwhile, with tri-modal alignment, the margin between the positive and negative samples is also increased. It reveals that with the help of tri-modal alignment, the distance between video and text that contain consistent semantics is reduced, while the margin between the misaligned samples is increased. The results demonstrate that with fused multi-modal representation as anchors, tri-modal alignment helps the video and language modalities to be better aligned.", + "bbox": [ + 496, + 386, + 892, + 808 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Effect of semantic masking strategy. We incorporate a semantic masking strategy to form the masked pairs in trimodal alignment task. As shown in Tab. 3, on more diverse evaluation benchmarks, e.g. DiDemo, where each video is paired with 5 different ground-truth texts, using masked samples brings more gains. The result reveals that the pro", + "bbox": [ + 496, + 810, + 893, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "14862", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/87698ca4b49657409b4a4907dccfda3dd3ad1fd1d27a70ef4b6cc1a9ec47fc23.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodModel ArchitectureTraining ObjectivesMSRVTTDiDeMoTGIF-FrameLSMDC-MC
Two uni-modal encoderCross-modal encoderInfoNCEMLMTMA + RanklossR@1R@1AccAcc
IND-A✓✓20.722.9--
IND-F✓✓✓--69.079.0
COMB✓✓✓✓19.822.668.978.8
Clover✓✓✓✓23.426.469.780.7
", + "bbox": [ + 122, + 70, + 851, + 181 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 5. Comparisons with task-independently trained models (IND) and naive combination of the uni- and cross- modal encoder (COMB)", + "bbox": [ + 76, + 194, + 890, + 208 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/2e58ff4bf58f40c41af7f8dbfc21d7d6b1964f84c19ba8bcd3d377f29760a0ef.jpg", + "image_caption": [ + "Baseline: a man is showing the interior of a car \nClover: inside of a car with large space and safety" + ], + "image_footnote": [], + "bbox": [ + 81, + 222, + 181, + 277 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/2bfdfe5b1d7a73efa803cb07f85464665ab823cc62ecb196edb800aa697daf4f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 184, + 222, + 274, + 277 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/202c7cf70b69e855b2960ae0660acce90836cfb7db406a4be60349be6a47a035.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 281, + 222, + 370, + 277 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/5e30e897642881d85a6c1c33a1789b9097d291c544fb3e51b74158bca0a9611e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 377, + 222, + 467, + 277 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/99598dfc66d0e3f4188dab9891b69b52d491f464e26d08e0eeb7ac8b96890ffe.jpg", + "image_caption": [ + "Baseline: a man is folding pieces of paper \nClover: a person folds a paper airplane" + ], + "image_footnote": [], + "bbox": [ + 84, + 313, + 179, + 367 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/462c84091090bc713fa3dc7e0b3c4abbf7db80a09ffc99fb3af2f2fb488d8d6f.jpg", + "image_caption": [ + "Figure 3. Qualitative results of zero-shot video to text retrieval results on MSRVTT [55]." + ], + "image_footnote": [], + "bbox": [ + 187, + 313, + 274, + 367 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/6668d484ba85d5bf6d7f8da3c2497dce53ecc2b883b832b200ad5ec0cb702042.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 282, + 313, + 370, + 367 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/6fa58e8590c243eba2efd5b6862e74bb238afb64237eb90ff1c090b5a8c50102.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 380, + 313, + 465, + 367 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "posed semantic masking strategy facilitates the model to capture key semantic information in more complex scenes and makes it achieve better results.", + "bbox": [ + 75, + 446, + 468, + 489 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effect of pair-wise ranking. To make the model capture the semantic information difference between the masked pairs and complete pairs, we adopt the pair-wise rank loss. As shown in Tab. 3, with pair-wise ranking loss, the performance of the model is further improved. We also show the visualization results below to further illustrate the effect of pair-wise ranking.", + "bbox": [ + 75, + 491, + 468, + 597 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Clover makes cross-modal alignment and fusion mutually improving. We compare Clover with the ones that use the same architecture but employ different pre-training tasks in Tab. 5. $IND-A$ represents the model only trained with InfoNCE loss for cross-modal alignment. $IND-F$ indicates the model trained with MLM loss for cross-modal fusion. We also report results achieved by the model (i.e., COMB) that simply combines the uni-modal encoders and multi-modal encoder as well as the training objectives. We can see that the simple combination hurts the performance, while our Clover achieves superior performance than the task-independently trained $IND-A$ and $IND-F$ . It reveals that Clover is able to get the model's cross-modal alignment and fusion capability mutually enhanced.", + "bbox": [ + 75, + 598, + 468, + 809 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Qualitative analysis. In Fig. 3 and Fig. 4, we show the qualitative results of Clover and the baseline method on zero-shot video-text retrieval and VQA tasks. Specifically, we present the query videos with the matched texts, and the video-question pair with the answer. Clover empowers the model with stronger video-text understanding capabilities.", + "bbox": [ + 75, + 810, + 468, + 900 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/f1316f9f5675cb4a33cdd88e7d78c7d679a44a60f81fc45521fde877815a75d7.jpg", + "image_caption": [ + "Question: What is a guy hugging? \nBaseline: Animal Clover: Lion" + ], + "image_footnote": [], + "bbox": [ + 506, + 220, + 598, + 275 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/62686c8452653c9a3206b7e51ae02eeba485c0ad74d9cb941cf6c3a314fc6875.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 604, + 220, + 696, + 275 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/43a8c699237e35b1055cbe51ad2b450317e3bbd0048a74aa4c068f9ab63af8bd.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 700, + 220, + 792, + 275 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/21ddc29e4c885e022189e3ffe838d9ce27bf821c830a1a38f0f0e3e595ed5455.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 797, + 220, + 890, + 275 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/8c74ef8541c5fb818a88aa3822a1d8d064b92fa65ca94854e52825b0f879095b.jpg", + "image_caption": [ + "Question: What does a person wash with a brush? Baseline: Face Clover: Toy" + ], + "image_footnote": [], + "bbox": [ + 506, + 309, + 596, + 363 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/46e79f251ce445f111bc523f64810d47c0dc1708866bb8d86c25c782de0551f0.jpg", + "image_caption": [ + "Figure 4. Qualitative results of video question answering results on MSRVTT-QA [52]." + ], + "image_footnote": [], + "bbox": [ + 604, + 309, + 694, + 363 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/f939e05940ef2722fd4631faab0281c3a85afc150b665c2e58d77303aeaa9cb1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 700, + 309, + 792, + 363 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/dfa832d6663c0c667b2f931a3dafbe07f37026a77bf46b809ea3dc63e748e4e7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 797, + 309, + 890, + 363 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "ity. For example, for the second video in Fig. 3, Clover returns the text with \"paper airplane\", which is a more accurate description of the video content; for the first video in Fig. 4, Clover generates a more accurate answer \"Lion\" than the baseline answer \"Animal\". The results demonstrate the superiority of Clover.", + "bbox": [ + 498, + 439, + 890, + 529 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 542, + 617, + 556 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we present Clover, a new end-to-end Video-Language pre-training method for both high-efficiency video-text retrieval and video question answering. Clover introduces a novel Tri-modal Alignment task to better align the representations from visual, text and fused modalities, which explicitly correlate the uni-modal encoder and multi-modal encoder. It also introduces semantic masking strategy and pair-wise ranking loss to further improve the cross-modality modal training. Extensive experiments conducted on the three retrieval datasets and eight video QA datasets clearly demonstrated, as a general video-text model, its consistent superiority for video-text understanding.", + "bbox": [ + 496, + 569, + 890, + 763 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements. This work was supported by National Key R&D Program of China (No.2022ZD0118201), the National Science Fund for Distinguished Young Scholars (No.62025603), the National Natural Science Foundation of China (No. U21B2037, No. U22B2051, No. 62176222, No. 62176223, No. 62176226, No. 62072386, No. 62072387, No. 62072389, No. 62002305 and No. 62272401), and the Natural Science Foundation of Fujian Province of China (No.2021J01002, No.2022J06001).", + "bbox": [ + 496, + 765, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "14863", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 174, + 106 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Lisa Anne Hendricks, Oliver Wang, Eli Shechtman, Josef Sivic, Trevor Darrell, and Bryan Russell. Localizing moments in video with natural language. In Proceedings of the IEEE international conference on computer vision, pages 5803-5812, 2017. 1, 5, 6", + "[2] Max Bain, Arsha Nagrani, Gül Varol, and Andrew Zisserman. Frozen in time: A joint video and image encoder for end-to-end retrieval. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1728-1738, 2021. 1, 2, 5, 6, 7", + "[3] Hangbo Bao, Li Dong, Songhao Piao, and Furu Wei. Beit: Bert pre-training of image transformers. In International Conference on Learning Representations, 2021. 2, 5", + "[4] David Chen and William B Dolan. Collecting highly parallel data for paraphrase evaluation. In Proceedings of the 49th annual meeting of the association for computational linguistics: human language technologies, pages 190–200, 2011. 1", + "[5] Xinlei Chen, Hao Fang, Tsung-Yi Lin, Ramakrishna Vedantam, Saurabh Gupta, Piotr Dólár, and C Lawrence Zitnick. Microsoft coco captions: Data collection and evaluation server. arXiv preprint arXiv:1504.00325, 2015. 6", + "[6] Xing Cheng, Hezheng Lin, Xiangyu Wu, Fan Yang, and Dong Shen. Improving video-text retrieval by multi-stream corpus alignment and dual softmax loss. arXiv preprint arXiv:2109.04290, 2021. 3", + "[7] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 4171–4186, 2019. 2, 3, 6", + "[8] Chenyou Fan, Xiaofan Zhang, Shu Zhang, Wensheng Wang, Chi Zhang, and Heng Huang. Heterogeneous memory enhanced multimodal attention model for video question answering. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1999-2007, 2019. 2", + "[9] Tsu-Jui Fu, Linjie Li, Zhe Gan, Kevin Lin, William Yang Wang, Lijuan Wang, and Zicheng Liu. Violet: End-to-end video-language transformers with masked visual-token modeling. arXiv preprint arXiv:2111.12681, 2021. 1, 2, 3, 6, 7", + "[10] Valentin Gabeur, Chen Sun, Karteek Alahari, and Cordelia Schmid. Multi-modal transformer for video retrieval. In European Conference on Computer Vision, pages 214-229. Springer, 2020. 1, 2", + "[11] Yuying Ge, Yixiao Ge, Xihui Liu, Dian Li, Ying Shan, Xiaohu Qie, and Ping Luo. Bridging video-text retrieval with multiple choice questions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16167-16176, 2022. 1, 2, 5, 6", + "[12] Yuying Ge, Yixiao Ge, Xihui Liu, Alex Jinpeng Wang, Jianping Wu, Ying Shan, Xiaohu Qie, and Ping Luo. Miles: Visual bert pre-training with injected language semantics for" + ], + "bbox": [ + 78, + 114, + 472, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "video-text retrieval. Proceedings of the European Conference on Computer Vision (ECCV), 2022. 6", + "[13] Simon Ging, Mohammadreza Zolfaghari, Hamed Pirsiavash, and Thomas Brox. Coot: Cooperative hierarchical transformer for video-text representation learning. Advances in neural information processing systems, 33:22605-22618, 2020. 1, 2", + "[14] Satya Krishna Gorti, Noel Vouitsis, Junwei Ma, Keyvan Golestan, Maksims Volkovs, Animesh Garg, and Guangwei Yu. X-pool: Cross-modal language-video attention for text-video retrieval. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5006-5015, 2022. 3", + "[15] Deng Huang, Peihao Chen, Runhao Zeng, Qing Du, Mingkui Tan, and Chuang Gan. Location-aware graph convolutional networks for video question answering. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 34, pages 11021-11028, 2020. 2", + "[16] Yunseok Jang, Yale Song, Youngjae Yu, Youngjin Kim, and Gunhee Kim. Tgif-qa: Toward spatio-temporal reasoning in visual question answering. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2758–2766, 2017. 1, 2, 3, 5", + "[17] Jie Jiang, Shaobo Min, Weijie Kong, Hongfa Wang, Zhifeng Li, and Wei Liu. Tencent text-video retrieval: Hierarchical cross-modal interactions with multi-level representations. IEEE Access, 2022. 3", + "[18] Will Kay, Joao Carreira, Karen Simonyan, Brian Zhang, Chloe Hillier, Sudheendra Vijayanarasimhan, Fabio Viola, Tim Green, Trevor Back, Paul Natsev, et al. The kinetics human action video dataset. arXiv preprint arXiv:1705.06950, 2017. 6", + "[19] Junyeong Kim, Minuk Ma, Kyungsu Kim, Sungjin Kim, and Chang D Yoo. Progressive attention memory network for movie story question answering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8337-8346, 2019. 2", + "[20] Ranjay Krishna, Yuke Zhu, Oliver Groth, Justin Johnson, Kenji Hata, Joshua Kravitz, Stephanie Chen, Yannis Kalantidis, Li-Jia Li, David A Shamma, et al. Visual genome: Connecting language and vision using crowdsourced dense image annotations. International journal of computer vision, 123(1):32-73, 2017. 6", + "[21] Alexander Kunitsyn, Maksim Kalashnikov, Maksim Dzabraev, and Andrei Ivaniuta. Mdmt-2: Multidomain multimodal transformer for video retrieval, one more step towards generalization. arXiv preprint arXiv:2203.07086, 2022. 3", + "[22] Thao Minh Le, Vuong Le, Svetha Venkatesh, and Truyen Tran. Hierarchical conditional relation networks for video question answering. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9972-9981, 2020. 2", + "[23] Jie Lei, Linjie Li, Luowei Zhou, Zhe Gan, Tamara L Berg, Mohit Bansal, and Jingjing Liu. Less is more: Clipbert for video-and-language learning via sparse sampling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7331-7341, 2021. 1, 2, 5, 6, 7" + ], + "bbox": [ + 501, + 92, + 893, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "14864", + "bbox": [ + 480, + 945, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[24] Dongxu Li, Junnan Li, Hongdong Li, Juan Carlos Niebles, and Steven CH Hoi. Align and prompt: Video-and-language pre-training with entity prompts. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4953-4963, 2022. 1, 2, 5, 6, 7", + "[25] Linjie Li, Yen-Chun Chen, Yu Cheng, Zhe Gan, Licheng Yu, and Jingjing Liu. Hero: Hierarchical encoder for video+ language omni-representation pre-training. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 2046-2065, 2020. 2", + "[26] Xiangpeng Li, Jingkuan Song, Lianli Gao, Xianglong Liu, Wenbing Huang, Xiangnan He, and Chuang Gan. Beyond rnns: Positional self-attention with co-attention for video question answering. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 33, pages 8658–8665, 2019. 2", + "[27] Chengzhi Lin, Ancong Wu, Junwei Liang, Jun Zhang, Wenhang Ge, Wei-Shi Zheng, and Chunhua Shen. Text-adaptive multiple visual prototype matching for video-text retrieval. Advances in Neural Information Processing Systems, 2022. 6", + "[28] Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dólár. Focal loss for dense object detection. In Proceedings of the IEEE international conference on computer vision, pages 2980-2988, 2017. 5", + "[29] Fei Liu, Jing Liu, Weining Wang, and Hanqing Lu. Hair: Hierarchical visual-semantic relational reasoning for video question answering. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1698–1707, 2021. 2", + "[30] Ruyang Liu, Jingjia Huang, Ge Li, Jiashi Feng, Xinglong Wu, and Thomas H Li. Revisiting temporal modeling for clip-based image-to-video knowledge transferring. arXiv preprint arXiv:2301.11116, 2023. 3", + "[31] Yang Liu, Samuel Albanie, Arsha Nagrani, and Andrew Zisserman. Use what you have: Video retrieval using representations from collaborative experts. In BMVC, 2019. 1, 2, 5", + "[32] Ze Liu, Jia Ning, Yue Cao, Yixuan Wei, Zheng Zhang, Stephen Lin, and Han Hu. Video swim transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3202-3211, 2022. 3, 6", + "[33] Edward Loper and Steven Bird. Nltk: The natural language toolkit. arXiv preprint cs/0205028, 2002. 5", + "[34] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In International Conference on Learning Representations, 2018. 6", + "[35] Huaishao Luo, Lei Ji, Botian Shi, Haoyang Huang, Nan Duan, Tianrui Li, Jason Li, Taroon Bharti, and Ming Zhou. Univl: A unified video and language pre-training model for multimodal understanding and generation. arXiv preprint arXiv:2002.06353, 2020. 2", + "[36] Huaishao Luo, Lei Ji, Ming Zhong, Yang Chen, Wen Lei, Nan Duan, and Tianrui Li. Clip4clip: An empirical study of clip for end to end video clip retrieval. arXiv preprint arXiv:2104.08860, 2021. 6" + ], + "bbox": [ + 78, + 90, + 470, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[37] Yiwei Ma, Guohai Xu, Xiaoshuai Sun, Ming Yan, Ji Zhang, and Rongrong Ji. X-clip: End-to-end multi-grained contrastive learning for video-text retrieval. In Proceedings of the 30th ACM International Conference on Multimedia, pages 638–647, 2022. 3", + "[38] Tegan Maharaj, Nicolas Ballas, Anna Rohrbach, Aaron Courville, and Christopher Pal. A dataset and exploration of models for understanding video data through fill-in-the-blank question-answering. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 6884–6893, 2017. 1, 5, 6", + "[39] Antoine Miech, Jean-Baptiste Alayrac, Lucas Smaira, Ivan Laptev, Josef Sivic, and Andrew Zisserman. End-to-end learning of visual representations from uncurated instructional videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9879-9889, 2020. 1, 2, 6", + "[40] Antoine Miech, Dimitri Zhukov, Jean-Baptiste Alayrac, Makarand Tapaswi, Ivan Laptev, and Josef Sivic. Howto100m: Learning a text-video embedding by watching hundred million narrated video clips. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2630-2640, 2019. 6", + "[41] Jungin Park, Jiyoung Lee, and Kwanghoon Sohn. Bridge to answer: Structure-aware graph interaction network for video question answering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15526-15535, 2021. 2", + "[42] Andrew Rouditchenko, Angie Boggust, David Harwath, Brian Chen, Dhiraj Joshi, Samuel Thomas, Kartik Audhkhasi, Hilde Kuehne, Rameswar Panda, Rogerio Feris, et al. Avlnet: Learning audio-visual language representations from instructional videos. In Annual Conference of the International Speech Communication Association. International Speech Communication Association, 2021. 2", + "[43] Paul Hongsuck Seo, Arsha Nagrani, and Cordelia Schmid. Look before you speak: Visually contextualized utterances. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16877-16887, 2021. 2", + "[44] Piyush Sharma, Nan Ding, Sebastian Goodman, and Radu Soricut. Conceptual captions: A cleaned, hypernymed, image alt-text dataset for automatic image captioning. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 2556-2565, 2018. 5, 6", + "[45] Chen Sun, Austin Myers, Carl Vondrick, Kevin Murphy, and Cordelia Schmid. Videobert: A joint model for video and language representation learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7464-7473, 2019. 2", + "[46] Hao Tan and Mohit Bansal. Lxmert: Learning cross-modality encoder representations from transformers. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages 5100-5111, 2019. 2" + ], + "bbox": [ + 501, + 90, + 890, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "14865", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[47] Atousa Torabi, Niket Tandon, and Leonid Sigal. Learning language-visual embedding for movie understanding with natural-language. arXiv preprint arXiv:1609.08124, 2016. 1, 5", + "[48] Aaron Van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding. arXiv e-prints, pages arXiv-1807, 2018. 2, 6", + "[49] Alex Jinpeng Wang, Yixiao Ge, Rui Yan, Yuying Ge, Xudong Lin, Guanyu Cai, Jianping Wu, Ying Shan, Xiaohu Qie, and Mike Zheng Shou. All in one: Exploring unified video-language pre-training. arXiv preprint arXiv:2203.07303, 2022. 1, 2, 6", + "[50] Jinpeng Wang, Yixiao Ge, Guanyu Cai, Rui Yan, Xudong Lin, Ying Shan, Xiaohu Qie, and Mike Zheng Shou. Object-aware video-language pre-training for retrieval. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3313-3322, 2022. 2, 6", + "[51] Yi Wang, Kunchang Li, Yizhuo Li, Yinan He, Bingkun Huang, Zhiyu Zhao, Hongjie Zhang, Jilan Xu, Yi Liu, Zun Wang, et al. Internvideo: General video foundation models via generative and discriminative learning. arXiv preprint arXiv:2212.03191, 2022. 3", + "[52] Dejing Xu, Zhou Zhao, Jun Xiao, Fei Wu, Hanwang Zhang, Xiangnan He, and Yueting Zhuang. Video question answering via gradually refined attention over appearance and motion. In Proceedings of the 25th ACM international conference on Multimedia, pages 1645-1653, 2017. 1, 2, 5, 8", + "[53] Hu Xu, Gargi Ghosh, Po-Yao Huang, Prahal Arora, Masoumeh Aminzadeh, Christoph Feichtenhofer, Florian Metze, and Luke Zettlemoyer. Vlm: Task-agnostic videolanguage model pre-training for video understanding. In Findings of the Association for Computational Linguistics: ACL-IJCNLP 2021, pages 4227–4239, 2021. 2", + "[54] Hu Xu, Gargi Ghosh, Po-Yao Huang, Dmytro Okhonko, Armen Aghajanyan, Florian Metze, Luke Zettlemoyer, and Christoph Feichtenhofer. Videoclip: Contrastive pre-training for zero-shot video-text understanding. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 6787-6800, 2021. 2", + "[55] Jun Xu, Tao Mei, Ting Yao, and Yong Rui. Msr-vtt: A large video description dataset for bridging video and language. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5288-5296, 2016. 1, 5, 6, 8", + "[56] Hongwei Xue, Tiankai Hang, Yanhong Zeng, Yuchong Sun, Bei Liu, Huan Yang, Jianlong Fu, and Baining Guo. Advancing high-resolution video-language representation with large-scale video transcriptions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5036-5045, June 2022. 1, 2, 6", + "[57] Hongwei Xue, Yuchong Sun, Bei Liu, Jianlong Fu, Ruihua Song, Houqiang Li, and Jiebo Luo. Clip-vip: Adapting pretrained image-text model to video-language representation alignment. arXiv preprint arXiv:2209.06430, 2022. 3", + "[58] Antoine Yang, Antoine Miech, Josef Sivic, Ivan Laptev, and Cordelia Schmid. Just ask: Learning to answer questions from millions of narrated videos. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1686-1697, 2021. 1, 2, 6, 7" + ], + "bbox": [ + 78, + 90, + 468, + 900 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[59] Jianwei Yang, Yonatan Bisk, and Jianfeng Gao. Taco: Token-aware cascade contrastive learning for video-text alignment. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11562-11572, 2021. 2", + "[60] Youngjae Yu, Jongseok Kim, and Gunhee Kim. A joint sequence fusion model for video question answering and retrieval. In Proceedings of the European Conference on Computer Vision (ECCV), pages 471-487, 2018. 3, 5", + "[61] Rowan Zellers, Ximing Lu, Jack Hessel, Youngjae Yu, Jae Sung Park, Jize Cao, Ali Farhadi, and Yejin Choi. Merlot: Multimodal neural script knowledge models. Advances in Neural Information Processing Systems, 34, 2021. 1, 6, 7", + "[62] Zheng-Jun Zha, Jiawei Liu, Tianhao Yang, and Yongdong Zhang. Spatiotemporal-textual co-attention network for video question answering. ACM Transactions on Multimedia Computing, Communications, and Applications (TOMM), 15(2s):1-18, 2019. 2", + "[63] Linchao Zhu and Yi Yang. Actbert: Learning global-local video-text representations. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8746-8755, 2020. 2" + ], + "bbox": [ + 501, + 92, + 892, + 401 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "14866", + "bbox": [ + 480, + 945, + 519, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2023/Clover_ Towards a Unified Video-Language Alignment and Fusion Model/eddf9a27-0fae-4139-acd3-5b79001d8460_model.json b/2023/Clover_ Towards a Unified Video-Language Alignment and Fusion Model/eddf9a27-0fae-4139-acd3-5b79001d8460_model.json new file mode 100644 index 0000000000000000000000000000000000000000..1bf76f842a1f4f3f1ce926c333e0c0cfa051045c --- /dev/null +++ b/2023/Clover_ Towards a Unified Video-Language Alignment and Fusion Model/eddf9a27-0fae-4139-acd3-5b79001d8460_model.json @@ -0,0 +1,2400 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.101, + 0.13, + 0.872, + 0.152 + ], + "angle": 0, + "content": "Clover: Towards A Unified Video-Language Alignment and Fusion Model" + }, + { + "type": "text", + "bbox": [ + 0.117, + 0.179, + 0.251, + 0.201 + ], + "angle": 0, + "content": "Jingjia Huang\\*" + }, + { + "type": "text", + "bbox": [ + 0.284, + 0.179, + 0.382, + 0.199 + ], + "angle": 0, + "content": "Yinan Li\\*" + }, + { + "type": "text", + "bbox": [ + 0.415, + 0.18, + 0.518, + 0.2 + ], + "angle": 0, + "content": "Jiashi Feng" + }, + { + "type": "text", + "bbox": [ + 0.56, + 0.18, + 0.681, + 0.2 + ], + "angle": 0, + "content": "Xinglong Wu" + }, + { + "type": "text", + "bbox": [ + 0.721, + 0.18, + 0.859, + 0.198 + ], + "angle": 0, + "content": "Xiaoshuai Sun" + }, + { + "type": "text", + "bbox": [ + 0.429, + 0.205, + 0.541, + 0.225 + ], + "angle": 0, + "content": "Rongrong Ji" + }, + { + "type": "text", + "bbox": [ + 0.179, + 0.233, + 0.792, + 0.27 + ], + "angle": 0, + "content": "\\(^{\\text{心}}\\)Key Laboratory of Multimedia Trusted Perception and Efficient Computing, Ministry of Education of China, Xiamen University, 361005, China" + }, + { + "type": "text", + "bbox": [ + 0.358, + 0.27, + 0.613, + 0.288 + ], + "angle": 0, + "content": "ByteDance Inc, 100043, China" + }, + { + "type": "text", + "bbox": [ + 0.273, + 0.288, + 0.697, + 0.306 + ], + "angle": 0, + "content": "yinanlee@stu.xmu.edu.cn, {xssun, rrji} @ xmu.edu.cn" + }, + { + "type": "text", + "bbox": [ + 0.272, + 0.306, + 0.696, + 0.324 + ], + "angle": 0, + "content": "{huangjingjia, jshfeng, wuxinglong} @bytedance.com" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.358, + 0.313, + 0.373 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.389, + 0.474, + 0.766 + ], + "angle": 0, + "content": "Building a universal Video-Language model for solving various video understanding tasks (e.g., text-video retrieval, video question answering) is an open challenge to the machine learning field. Towards this goal, most recent works build the model by stacking uni-modal and cross-modal feature encoders and train it with pair-wise contrastive pre-text tasks. Though offering attractive generality, the resulted models have to compromise between efficiency and performance. They mostly adopt different architectures to deal with different downstream tasks. We find this is because the pair-wise training cannot well align and fuse features from different modalities. We then introduce Clover—a Correlated Video-Language pre-training method—towards a universal Video-Language model for solving multiple video understanding tasks with neither performance nor efficiency compromise. It improves cross-modal feature alignment and fusion via a novel tri-modal alignment pre-training task. Additionally, we propose to enhance the tri-modal alignment via incorporating learning from semantic masked samples and a new pair-wise ranking loss. Clover establishes new state-of-the-arts on multiple downstream tasks, including three retrieval tasks for both zero-shot and fine-tuning settings, and eight video question answering tasks. Codes and pre-trained models will be released at https://github.com/LeeYN-43/Clover." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.797, + 0.21, + 0.812 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.823, + 0.47, + 0.853 + ], + "angle": 0, + "content": "Video-Language pre-training (VidL) aims to learn generalizable multi-modal models from large-scale video-text" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.359, + 0.892, + 0.465 + ], + "angle": 0, + "content": "samples so as to better solve various challenging Video-Language understanding tasks, such as text-video retrieval [1,4,38,55] and video question answering [16,47,52]. Recent studies [9,11,23,24,49,56,58,61] have shown that VidL leads to significant performance improvement and achieves state-of-the-art results on various downstream text-video retrieval and video question answering (VQA) benchmarks." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.471, + 0.893, + 0.607 + ], + "angle": 0, + "content": "Though achieving encouraging performance, existing VidL models mostly adopt different architectures to deal with different downstream tasks. For the text-video retrieval tasks, they [2, 10, 11, 13, 31, 39] typically use two individual uni-modal encoders for processing video and text data separately, for the sake of retrieval efficiency. While for video question answering tasks, the models usually adopt the multi-modal joint encoder design to learn the association and interaction of different modalities." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.614, + 0.895, + 0.902 + ], + "angle": 0, + "content": "Building a unified model capable of solving various Video-Language tasks is a long-standing challenge for machine learning research. A few recent works [9,24] attempt to learn a unified VidL model for both tasks, which uses the multi-modal encoder to conduct text-video retrieval. However, the model requires an exhaustive pair-wise comparison between the query texts and gallery videos. Given \\(N\\) text queries and \\(M\\) category videos, the computation complexity of the multi-modal encoder model would be \\(O(NM)\\), which makes it infeasible for large-scale videotext retrieval applications. Another straightforward solution is to simply combine the uni-modal and multi-modal encoders (Fig. 1 (a)), and perform the retrieval and VQA tasks through the uni-modal and multi-modal encoders respectively. Its computation complexity for retrieval tasks is only \\(O(N + M)\\). However, the experiment results in Fig.1 (c) show that, without a carefully designed correlating mechanism between these two types of encoders, the simple combination i.e., COMB yields a compromised per" + }, + { + "type": "page_footnote", + "bbox": [ + 0.094, + 0.863, + 0.32, + 0.876 + ], + "angle": 0, + "content": "*Equal Contribution in alphabetical order." + }, + { + "type": "page_footnote", + "bbox": [ + 0.097, + 0.876, + 0.342, + 0.888 + ], + "angle": 0, + "content": "\\(\\dagger\\) Work done when interning at ByteDance Inc." + }, + { + "type": "page_footnote", + "bbox": [ + 0.097, + 0.888, + 0.226, + 0.9 + ], + "angle": 0, + "content": "\\(\\ddagger\\) Corresponding Author." + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.863, + 0.342, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "14856" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.47, + 0.121 + ], + "angle": 0, + "content": "formance compared with the models i.e. IND individually designed for retrieval and VQA." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.124, + 0.473, + 0.655 + ], + "angle": 0, + "content": "In this work, we aim to address the above issues and build a unified pre-training model that attains high efficiency and performance simultaneously. We observe: (i) well aligning the features of the video and text from the same data pair is important for text-video matching; (ii) effectively fusing video and text features into unified representations is critical for video-text understanding. However, existing pre-training strategies that rely on either simple supervised or contrastive pre-text tasks hardly achieve promising feature alignment and fusion capability simultaneously. Motivated by these observations, we develop a new VidL method from these two aspects. Specifically, we propose Correlated Video-Language pre-training (Clover), a VidL method that not only unifies Video-Language alignment and fusion, but also makes them mutually boosted. The fused multi-modal representation contains richer context information than the uni-modal representations [11,35]. As an intermediate modality between video and text, the multi-modal representations are good anchors for cross-modality alignment. Meanwhile, keeping the fused representation closer to the uni-modal representation containing consistent semantic information and away from the inconsistent one will enhance the learning of semantic information in the fused modality. Therefore, we propose the Tri-Modal Alignment (TMA) to get Video-Language alignment and fusion mutually boosted, which takes the alignment between the multimodal representation and text/video representations as an auxiliary objective. We note that since the tri-modal alignment is well compatible with the classical pre-training tasks [3,7,35] e.g., Masked Language Modeling, its computation overhead is negligible. To help the model maintain fine-grained discriminative capability while improving its generalizability, we further introduce a pair-wise ranking loss that urges the model to be aware of the concept missing in masked samples compared to original samples." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.657, + 0.47, + 0.838 + ], + "angle": 0, + "content": "Extensive experiments are conducted on multiple downstream tasks, including three retrieval tasks with different experimental setups (i.e. zero-shot and fine-tune) and eight video question answering tasks. The results demonstrate that Clover is able to get the cross-modal fusion and alignment capability mutually improved, and consistently outperforms current SOTAs on various downstream tasks. It achieves an average performance improvement of \\(4.9\\%\\) and \\(8.7\\%\\) Recall@10 score on the zero-shot and fine-tune settings of the three downstream retrieval datasets, while the average accuracy improvement over current SOTAs is \\(2.3\\%\\) on the eight video question answering datasets." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.84, + 0.47, + 0.901 + ], + "angle": 0, + "content": "In summary, we make the following contributions: (1) we introduce Clover, a pre-training method achieving unified Video-Language alignment and fusion model that can be easily transferred to various downstream video under" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.892, + 0.152 + ], + "angle": 0, + "content": "standing tasks while attaining both high efficiency and performance; (2) we propose a novel tri-modal alignment pretraining task, which correlates the uni-modal encoder and multi-modal encoder to get them mutually boosted." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.166, + 0.642, + 0.182 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.191, + 0.892, + 0.432 + ], + "angle": 0, + "content": "Video-Language pre-training for text-video retrieval. Existing pre-training methods for text-video retrieval can be mainly divided into two categories. The first kind of method employs two individual encoders to embed video and text, and project them into a common latent space [2,10,11,13,31,39,42,50,54,59]. Then, they leverage a contrastive objective [48] to align the paired representations. Thanks to their high efficiency, this category of methods is widely used in real-world applications. The other kind of method utilizes a joint multi-modal encoder to learn comprehensive representation for a given video-text pair, and predict whether the video and text are matched or not via a binary classifier [9,23,25,45,53,63]. Despite the good performance they achieved, they need to exhaustively pair all the videos and texts, and feed them into the model during inference, making them highly inefficient." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.434, + 0.892, + 0.658 + ], + "angle": 0, + "content": "Video-Language pre-training for video question answering. Video question answering aims to automatically answer natural language questions given a context video, which requires a joint understanding of both the video and language. Previous works mainly focus on designing better spatio-temporal attention networks [8,16,26,29,52,62] and question-video relation networks [15, 19, 22, 41] to extract more accurate multi-modal representations. Recently, many works leverage transformer-based models to build an additional multi-modal encoder upon the uni-modal encoders [9, 24, 43] or input image patches and word tokens [58, 63] directly to fuse cross-modality features. The multi-modal encoder is typically driven by the pre-training tasks e.g. Masked Language Modeling [7] and Visual-Text Matching [46], to perform token-level cross-modal fusion." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.66, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Video-Language pre-training for multiple downstream tasks. There are a few recent works trying to construct a unified pre-trained model for multiple downstream tasks. ActBert [63], clipBert [23], VIOLET [9] and ALPRO [24] utilize a cross-modal encoder for both retrieval and VQA tasks. They take video-text retrieval as a binary classification task and construct the classifier upon a visual-text matching head, which brings about extreme computation overload. While All-in-One [49] and HD-VILA [56] proposes a transformer-based unified backbone architecture for joint-modal representation learning, they require a large amount of data for pre-training. HERO [25] and VLM [53] design a task-agnostic model that can be finetuned on multiple downstream tasks. However, they are not end-to-end trainable and need pre-extract video features, which limits the performance of the model." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "14857" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.134, + 0.045, + 0.331, + 0.225 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.347, + 0.045, + 0.54, + 0.223 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.542, + 0.066, + 0.837, + 0.223 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.239, + 0.895, + 0.31 + ], + "angle": 0, + "content": "Figure 1. Different attempts on achieving a universal VidL model. (a) Naive combination by stacking the cross-modal encoder upon the uni-modal encoders with weak association (COMB). (b) Our proposed Clover pre-trained model that better correlates the cross-modal encoder and the uni-modal encoders via Tri-Modal Alignment. (c) Comparisons of the transfer performance on MSRVTT [60] (zero-shot) and TGIF-FrameQA [16] datasets. Benefiting from the proposed tri-modal alignment, our proposed universal model outperforms both the individually designed models (IND) and COMB." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.32, + 0.47, + 0.38 + ], + "angle": 0, + "content": "Unlike the aforementioned methods, our Clover is a fully end-to-end VidL model that can be easily transferred to various downstream tasks while attaining both high efficiency and performance." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.381, + 0.47, + 0.486 + ], + "angle": 0, + "content": "Note that there are some methods focus on transferring the pre-trained CLIP model to downstream tasks and achieve significant performance [6, 14, 17, 21, 30, 37, 51, 57]. The CLIP model is pretrained with 400M image-text data pairs. In contrast, our Clover focuses on presenting a better video-text pretrained model, and only leverages 5M data pairs for pretraining." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.501, + 0.17, + 0.517 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.526, + 0.31, + 0.541 + ], + "angle": 0, + "content": "3.1. Motivation and Overview" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.55, + 0.47, + 0.656 + ], + "angle": 0, + "content": "Recent years have witnessed a significant progress in Video-Language pre-training. Nevertheless, building a universal Video-Language model for solving various video understanding tasks (e.g., text-video retrieval, video question answering) is still a challenging problem. The key to the problem lies in how to unify the representation learning for cross-modal alignment and fusion efficiently." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.656, + 0.47, + 0.76 + ], + "angle": 0, + "content": "Cross-modal alignment aims to learn projection functions \\( f(\\cdot) \\) and \\( g(\\cdot) \\) that project videos and texts into a common embedding space where the similarity between the video and text with consistent semantic information is maximized and otherwise minimized. Taking a video \\( V \\) as the anchor, the learning objective of cross-modal alignment is formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.1, + 0.772, + 0.47, + 0.799 + ], + "angle": 0, + "content": "\\[\n\\arg \\max _ {f, g} [ \\mathrm {s} (f (V), g (T ^ {+})) - \\mathrm {s} (f (V), g (T ^ {-})) ], \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.81, + 0.47, + 0.902 + ], + "angle": 0, + "content": "where \\( T^{+}, T^{-} \\) represent texts that are semantically consistent and inconsistent with \\( V \\), respectively. The function \\( \\mathrm{s}(\\cdot, \\cdot) \\) measures the dot-product similarity between two embeddings. A model with strong cross-modal alignment capability is desired in the video-retrieval task, which requires matching the semantically aligned videos and texts." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.32, + 0.895, + 0.427 + ], + "angle": 0, + "content": "Cross-modal fusion aims to integrate the correlation and interaction carried by the video and text modalities into a unified multi-modal embedding. Specifically, it can be formulated as learning a function \\( Fusion(\\cdot, \\cdot) \\) that takes inputs of different modals and outputs a unified representation \\( M = Fusion(V, T) \\), which is then used to solve downstream tasks like VQA." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.445, + 0.895, + 0.612 + ], + "angle": 0, + "content": "Existing pre-training strategies rely on either simple supervised or contrastive pre-text tasks that hardly achieve feature alignment and fusion simultaneously. To better correlate the cross-modal alignment and fusion, we propose the Correlated Video-Language pre-training method (Clover) with three key pre-training innovations (Sec. 3.2): tri-modal alignment; pair-wise ranking loss; and semantic enhanced masked language modeling. The classical masked language modeling task is also incorporated into our method, which is able to boost the generalizability of the model as well as the interaction between visual and language." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.629, + 0.895, + 0.901 + ], + "angle": 0, + "content": "Architecture. We briefly explain the architecture of Clover here, which is shown in Fig. 2 as well. It consists of three components: a video encoder, a text encoder and a multi-modal encoder. Following [9], we use VideoSwin Transformer [32] as our video encoder. Given a video \\(V\\), it outputs a sequence of video embeddings: \\(V_{e} = \\{v_{1},\\dots,v_{K}\\} \\subset \\mathbb{R}^{D}\\), where \\(K\\) is the number of flattened patches. The text encoder of Clover is a 12-layer bidirectional transformer encoder model [7]. Given an input text sentence \\(T\\), the encoder outputs an embedding sequence \\(T_{e} = \\{t_{\\mathrm{cls}},t_{1},\\dots,t_{L - 1}\\} \\subset \\mathbb{R}^{D}\\), where \\(t_\\mathrm{cls}\\) indicates the embedding of the [CLS] token. We employ a 3-layer bidirectional Transformer encoder to learn the fused cross-modal representation of a video-text pair. The multi-modal encoder takes the concatenation of video and text embeddings as input, and outputs the fused multi-modal embeddings \\(M_{e} = \\{m_{v_{1}},\\dots,m_{v_{K}},m_{\\mathrm{CLS}},m_{t_{1}},\\dots,m_{t_{L - 1}}\\}\\), where \\(m \\in \\mathbb{R}^D\\)." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "14858" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.139, + 0.039, + 0.53, + 0.271 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.272, + 0.281, + 0.39, + 0.294 + ], + "angle": 0, + "content": "(a) Model Architecture" + }, + { + "type": "image_caption", + "bbox": [ + 0.552, + 0.04, + 0.66, + 0.052 + ], + "angle": 0, + "content": "Tri-modal Alignment" + }, + { + "type": "image", + "bbox": [ + 0.552, + 0.055, + 0.811, + 0.17 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.559, + 0.174, + 0.656, + 0.187 + ], + "angle": 0, + "content": "Pair-wise Ranking" + }, + { + "type": "image", + "bbox": [ + 0.557, + 0.2, + 0.657, + 0.274 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.691, + 0.2, + 0.793, + 0.274 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.624, + 0.283, + 0.737, + 0.296 + ], + "angle": 0, + "content": "(b) Pre-training Tasks" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.311, + 0.894, + 0.34 + ], + "angle": 0, + "content": "Figure 2. Overview of Clover (a) Model architecture. (b) Tri-modal alignment pre-training task with masked samples and the pair-wise ranking. “×” indicates the embedding of the token [MASK]. For more detail on model architecture and pre-training tasks refers to Sec. 3." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.351, + 0.279, + 0.367 + ], + "angle": 0, + "content": "3.2. Tri-Modal Alignment" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.377, + 0.47, + 0.603 + ], + "angle": 0, + "content": "Existing VidL pre-training methods typically align the embeddings of data from different modalities through pairwise contrastive learning. Differently, we propose a TriModal Alignment (TMA) learning task that not only enforces the video and text modalities to be aligned but also encourages them to well align with the third modality, i.e., their fusion modality. In TMA, as an intermediate modality between video and text, the fused multi-modal representations act as anchors for cross-modality alignment, which reduces the difficulty of the direct alignment between video and text. Meanwhile, it keeps the fused representation closer to the uni-modal representation containing consistent semantic information and away from the others, so as to enhance the learning of semantic information in the fused representation." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.607, + 0.47, + 0.775 + ], + "angle": 0, + "content": "As shown in Fig. 2 (a), given an input video-text pair \\(\\langle V, T \\rangle\\) and their embeddings \\(\\langle V_e, T_e \\rangle\\) from the uni-modal encoder, we mask some regions in \\(V\\) to make a masked video with embedding \\(V_m\\), and similarly mask some tokens in \\(T\\) to make a masked text embedding \\(T_m\\). Then, we pair the incomplete samples with the complete samples as \\(\\langle V_m, T_e \\rangle\\) and \\(\\langle V_e, T_m \\rangle\\). We adopt multi-modal encoder to get fused multi-modal embedding sequence \\(M_{V_m}\\) and \\(M_{T_m}\\) of \\(\\langle V_m, T_e \\rangle\\) and \\(\\langle V_e, T_m \\rangle\\), respectively. We denote the embedding of [CLS] token in \\(M_{V_m}\\) and \\(M_{T_m}\\) as \\(M_{V_m f}\\) and \\(M_{T_m f}\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.775, + 0.471, + 0.902 + ], + "angle": 0, + "content": "For the convenience of description, we use superscripts \\(i\\) and \\(j\\) to index the data sample, and introduce TMA from the perspective of video and language, respectively. Firstly, for a video representation \\(V_{e}^{i}\\), besides its associated text embedding \\(T_{e}^{i}\\) and \\(T_{m}^{i}\\), we also consider its associated fusion embedding \\(M_{V_{mf}}^{i}\\) as its positive pair—that should be pushed closer within the embedding space. To align these tri-modal representations, we propose a novel exclusivity loss computed within a batch of \\(B\\) samples as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.515, + 0.369, + 0.875, + 0.447 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} L _ {v} = - \\sum_ {i = 1} ^ {B} \\left[ \\log \\frac {e ^ {s (V _ {e} ^ {i} , T _ {e} ^ {i}) / \\tau}}{e ^ {s (V _ {e} ^ {i} , T _ {e} ^ {i}) / \\tau} + Z} + \\log \\frac {e ^ {s (V _ {e} ^ {i} , T _ {m} ^ {i}) / \\tau}}{e ^ {s (V _ {e} ^ {i} , T _ {m} ^ {i}) / \\tau} + Z} \\right. \\\\ \\left. + \\log \\frac {e ^ {s (V _ {e} ^ {i} , M _ {V _ {m} f} ^ {i}) / \\tau}}{e ^ {s (V _ {e} ^ {i} , M _ {V _ {m} f} ^ {i}) / \\tau} + Z} \\right], \\mathrm {w h e r e} \\\\ \\end{array}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.521, + 0.449, + 0.892, + 0.498 + ], + "angle": 0, + "content": "\\[\nZ = \\sum_ {j \\neq i} ^ {B} \\left[ e ^ {s \\left(V _ {e} ^ {i}, T _ {e} ^ {j}\\right) / \\tau} + e ^ {s \\left(V _ {e} ^ {i}, T _ {m} ^ {j}\\right) / \\tau} + e ^ {s \\left(V _ {e} ^ {i}, M _ {V _ {m f}} ^ {j}\\right) / \\tau} \\right]. \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.499, + 0.893, + 0.689 + ], + "angle": 0, + "content": "Here \\(\\tau\\) is a temperature scalar, and \\(s(\\cdot ,\\cdot)\\) is the dot-product similarity function to measure the degree of alignment between different modalities. In each term of Eq. (2), we exclude the other positive pairs when performing contrastive learning on one positive pair. Thus the intra-suppression among positive pairs is effectively avoided, and the video-/text-/fusion-modal representations are better aligned. For example, in the first term of Eq. (2), \\(e^{s(V_e^i,M_{Vmf}^i) / \\tau}\\) and \\(e^{s(V_e^i,T_m^i) / \\tau}\\) are excluded from the denominator, which prevents the model from yielding the sub-optimal solution that gets \\(T_{e}^{i}\\) closer to \\(V_{e}^{i}\\) while pushes \\(T_{m}^{i}\\) and \\(M_{Vmf}^{i}\\) away from \\(V_{e}^{i}\\). Meanwhile, to align the video to the text and fusion modalities, we have the following objective:" + }, + { + "type": "equation", + "bbox": [ + 0.501, + 0.694, + 0.892, + 0.789 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} L _ {v ^ {\\prime}} = - \\sum_ {i = 1} ^ {B} \\left[ \\log \\frac {e ^ {s (T _ {e} ^ {i} , V _ {e} ^ {i}) / \\tau}}{\\sum_ {j = 1} ^ {B} e ^ {s (T _ {e} ^ {i} , V _ {e} ^ {j}) / \\tau}} + \\log \\frac {e ^ {s (T _ {m} ^ {i} , V _ {e} ^ {i}) / \\tau}}{\\sum_ {j = 1} ^ {B} e ^ {s (T _ {m} ^ {i} , V _ {e} ^ {j}) / \\tau}} \\right] \\\\ \\left. + \\log \\frac {e ^ {s \\left(M _ {V _ {m f}} ^ {i} , V _ {e} ^ {i}\\right) / \\tau}}{\\sum_ {j = 1} ^ {B} e ^ {s \\left(M _ {V _ {m f}} ^ {i} , V _ {e} ^ {j}\\right) / \\tau}} \\right]. \\tag {3} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.79, + 0.894, + 0.881 + ], + "angle": 0, + "content": "The above two losses (2) and (3) are summed up to give the tri-modal alignment objective w.r.t. the visual modality: \\( L_{V} = L_{v} + L_{v^{\\prime}} \\). Similarly, from the perspective of text modality, we define the tri-modal alignment objective \\( L_{T} = L_{t} + L_{t^{\\prime}} \\) w.r.t. the text modality. Then, we obtain the final loss of tri-modal alignment \\( L_{TmA} = L_{V} + L_{T} \\)." + }, + { + "type": "page_footnote", + "bbox": [ + 0.517, + 0.887, + 0.862, + 0.901 + ], + "angle": 0, + "content": "For the detail of \\(L_{T}\\), please refer to the supplementary material." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "14859" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.47, + 0.137 + ], + "angle": 0, + "content": "To get the masked samples for TMA, we conduct masking over the videos and texts to form masked samples using the following two strategies." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.137, + 0.468, + 0.272 + ], + "angle": 0, + "content": "Video-block masking strategy. Objects in a video often appear at similar spatial locations across consecutive frames, which compose a tube area in the spatial-temporal domain. In order to construct incomplete video samples with less information leakage, we extend the block-wise mask [3] to video via performing a block-wise mask at the same position in all video frames. In this paper, we randomly replace \\(20\\%\\) patches with a learnable mask token to obtain \\(V_{m}\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.273, + 0.468, + 0.484 + ], + "angle": 0, + "content": "Semantic text masking strategy. The detailed semantic information in the text basically lies in the verb, noun and adjective words. To facilitate learning their representations, we construct incomplete text samples by randomly masking some verbs, nouns and adjectives in the sentence. Specifically, given a sentence, we use a part-of-speech tagger [33] to tag each word. Then we pick the verb phrases and nouns, and replace \\(30\\%\\) of them with a special [MASK] token. Note that to avoid drastic semantic changes, we do not mask the auxiliary verb like have, should, will, would, etc. With the masking strategy, when some key elements are masked from the texts, it would form an expression carrying partial information rather than completely changing the information expressed." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.494, + 0.258, + 0.51 + ], + "angle": 0, + "content": "3.3. Training Objective" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.516, + 0.468, + 0.725 + ], + "angle": 0, + "content": "Pair-wise ranking. In TMA, we take the masked pairs \\(\\langle V_e, T_m \\rangle\\) and \\(\\langle V_m, T_e \\rangle\\) as positive pairs, considering that masked samples would still carry partial meaningful information though some others are missed. For instance, as shown in Fig. 2 (a), given the text \"A black swan swimming in a calm lake\", \"swimming\" and \"lake\" are masked with [MASK] tokens, generating the masked pair \\(\\langle V_e, T_m \\rangle\\). Although some information is removed, the masked sentence still contains other concepts that make it partially matched with the video, such as the \"black swan\". Therefore, in contrastive learning, \\(V_m\\) and \\(T_m\\) can be considered pseudo positive candidates for \\(T\\) and \\(V\\). However, compared to the original pair \\(\\langle V_e, T_e \\rangle\\), we consider that the semantic consistency in the masked pairs should be weaker. Based on this prior, we further propose a pair-wise ranking loss:" + }, + { + "type": "equation", + "bbox": [ + 0.086, + 0.735, + 0.468, + 0.78 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} L _ {r a n k} = \\max \\left(0, - \\left(\\operatorname {s i m} \\left(V _ {e}, T _ {e}\\right) / \\tau - \\operatorname {s i m} \\left(V _ {e}, T _ {m}\\right) / \\tau\\right) + \\lambda\\right) \\\\ + \\max \\left(0, - \\left(\\operatorname {s i m} \\left(V _ {e}, T _ {e}\\right) / \\tau - \\operatorname {s i m} \\left(V _ {m}, T _ {e}\\right) / \\tau\\right) + \\lambda\\right), \\tag {4} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.78, + 0.468, + 0.885 + ], + "angle": 0, + "content": "where \\(\\lambda > 0\\) is a margin hyper-parameter. Eq. (4) urges the model to be aware of the gap of semantic consistency between \\(\\langle V_e, T_m \\rangle / \\langle V_m, T_e \\rangle\\) and \\(\\langle V_e, T_e \\rangle\\) brought about by concepts missing in the masked pairs. With the pairwise ranking objective, our model is able to maintain fine-grained perceptual capability while improving its generalizability." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.886, + 0.468, + 0.901 + ], + "angle": 0, + "content": "Semantic enhanced masked language modeling. Masked" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.287 + ], + "angle": 0, + "content": "language modeling (MLM) is a classical pre-training task in VidL, which promotes the interaction between different modalities in the cross-modal encoder. Combining the classical MLM with our semantic text masking strategy, we present the Semantic Enhanced Masked Language Modeling task, which facilitates the representation learning on the key concepts, i.e., verb, noun and adjective words. Besides, considering the class-imbalance between different words, we use focal loss [28] instead of the traditional cross-entropy loss to improve the MLM loss. We apply the MLM loss to the reconstruction of the masked text tokens i.e., \\( m_{t_j} \\in M_{T_m} \\), where \\( t_j = [Mask] \\). The MLM loss is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.545, + 0.297, + 0.892, + 0.345 + ], + "angle": 0, + "content": "\\[\nL _ {m l m} = - \\frac {1}{B} \\sum_ {i = 1} ^ {B} \\sum_ {m _ {t _ {j}} \\in M _ {T _ {m}} ^ {i}} [ (1 - p _ {t _ {j}} ^ {i}) ^ {\\gamma} p _ {t _ {j}} ^ {i} ] \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.357, + 0.892, + 0.417 + ], + "angle": 0, + "content": "where \\( p_{t_j}^i \\) denote the predicted probability distribution of the masked token \\( t_j \\) in \\( i \\)th sentence in the batch, and \\( \\gamma \\) is a hyper-parameter. Finally, the overall pre-training objective of Clover is:" + }, + { + "type": "equation", + "bbox": [ + 0.594, + 0.432, + 0.892, + 0.447 + ], + "angle": 0, + "content": "\\[\nL = L _ {T m A} + L _ {r a n k} + L _ {m l m}. \\tag {6}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.474, + 0.633, + 0.491 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.499, + 0.676, + 0.516 + ], + "angle": 0, + "content": "4.1. Experiment Setup" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.523, + 0.892, + 0.628 + ], + "angle": 0, + "content": "Pre-training datasets. Following recent work [2, 11, 24], we jointly pre-train our Clover on a video dataset WebVid2M [2] with 2.5M video-text pairs and an image dataset Google Conceptual Captions (CC3M) [44] with 3.3M image-text pairs (we only obtain 2.8M image-text pairs in CC3M due to image url broken). During pretraining, we treat image data as one frame video data." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.63, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Downstream tasks. We evaluate our proposed models on the following downstream tasks. (a) Text-to-Video Retrieval on MSRVTT [55], LSMDC [38] and DiDeMo [1]. For DiDeMo, we follow [23, 31] and evaluate Clover on the paragraph-to-video retrieval, where text sentences for each video are concatenated together as one text query. We do not use the ground-truth proposal for fair comparison with previous works; (b) Multiple-choice QA on TGIF-Action [16], TGIF-Transition [16], MSRVTT-MC [60] and LSMDC-MC [47]; (c) Open-Ended QA on TGIF-Frame [16], MSRVTT-QA [52], MSVD-QA [52] and LSMDC-FiB [38]. For retrieval tasks, we only use the two uni-modal encoders of Clover for fine-tuning and inference. We adopt the two encoders to get the video and text embeddings, and calculate their cosine similarity for retrieval. For QA tasks, we use all the modules for fine-tuning and inference. More details on these datasets and their evaluation usage are provided in the supplementary material." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "14860" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.087, + 0.088, + 0.887, + 0.415 + ], + "angle": 0, + "content": "
MethodPre-training datasetDiDeMoMSRVTTLSMDC
R@1R@5R@10MedRR@1R@5R@10MedRR@1R@5R@10MedR
Fine-tune
clipBert [23]COCO [5], VG [20]20.448.060.8622.046.859.96----
Frozen [2]W2M+C3M31.059.872.4331.059.570.5315.030.839.820
VIOLET [9]W2M+C3M+Y180M32.662.874.7-34.563.073.4-16.136.641.2-
HD-VILA [56]HDV100M28.857.469.1435.665.378.0317.434.144.115
ALPRO [24]W2M+C3M35.967.578.8333.960.773.23----
TMVM [27]W2M+C3M36.564.975.4336.264.275.7317.837.145.913.5
All-in-1 [49]W2M+H100M32.761.473.5-37.968.177.1-----
OA-Trans [50]W2M+C3M34.864.475.1335.863.476.5318.234.343.718.5
MILES [12]W2M+C3M36.663.974.0337.763.673.8317.835.644.115.5
MCQ [11]W2M+C3M37.062.273.9337.664.875.1317.935.444.515
Clover (ours)W2M+C3M50.176.785.6140.569.879.4224.844.054.58
Zero-shot
MIL-NCE [39]H100M----9.924.032.429.6----
Frozen [2]W2M+C3M21.146.056.2718.739.651.6109.322.030.151.0
VIOLET [9]W2M+C3M+Y180M23.549.859.8-25.949.559.7-----
ALPRO [24]W2M+C3M23.847.357.9624.144.755.48----
OA-Trans [50]W2M+C3M23.550.459.8623.447.555.68----
MILES [12]W2M+C3M27.250.363.6526.147.256.9711.124.730.650.7
MCQ [11]W2M+C3M25.650.661.1526.046.456.4712.225.932.242
Clover (ours)W2M+C3M29.555.266.3426.449.560.0614.729.238.224
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.424, + 0.892, + 0.465 + ], + "angle": 0, + "content": "Table 1. Text-to-video retrieval performance comparison under fine-tune and zero-shot setups. Here higher R@k (Recall K) and lower MedR (Median Recall) indicate better performance. W2M, C3M, H100M, HDV100M, Y180M are short for WebVid2M [2], CC3M [44], HowTo100M [40], HD-VILA-100M [56], YT-Temporal-180M [61], respectively." + }, + { + "type": "table", + "bbox": [ + 0.162, + 0.466, + 0.808, + 0.602 + ], + "angle": 0, + "content": "
MethodPre-training datasetTGIFMSRVTTLSMDCMSVD
ActionTransitionFrameMCQAMCFiBQA
clipBert [23]COCO, VG82.887.860.388.237.4---
JuskAsk [58]HTVQA69M [58]----41.5--46.3
ALPRO [24]W2M+C3M----42.1--45.9
All-in-1 [49]W2M+H100M92.794.364.292.042.983.1-47.9
VIOLET [9]W2M+C3M+Y180M92.595.768.991.943.982.853.747.9
MERLOT [61]Y180M94.096.269.590.943.181.752.9-
Clover (ours)W2M+C3M95.098.271.695.244.183.754.152.4
" + }, + { + "type": "table_caption", + "bbox": [ + 0.197, + 0.606, + 0.771, + 0.621 + ], + "angle": 0, + "content": "Table 2. Performance comparison on transferring to downstream video question answering tasks." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.632, + 0.473, + 0.89 + ], + "angle": 0, + "content": "Implementation details. Following [9], we initialize the video encoder with a VideoSwin-Base [32] model pretrained on Kinetics-400 [18]. The text encoder is initialized from pre-trained Bert-Base [7]. The multi-modal encoder is initialized from the first three layers of the pre-trained Bert-Base model. We train Clover end-to-end during both pretraining and fine-tuning. We pre-train Clover for 40 epochs, using a batch size of 1024 on 64 NVIDIA A100 GPUs. We use AdamW [34] optimizer with a weight decay 0.005 and betas (0.9, 0.98). The learning rate is first warmed-up by 4 epochs to 5e-5 and then decays following a cosine annealing decay schedule. We resize all the video frames to \\(224 \\times 224\\) and split each frame into patches with a size of \\(32 \\times 32\\). We choose hyper-parameter \\(\\tau = 0.05\\), \\(\\lambda = 5\\) and \\(\\gamma = 2\\). We set dimension \\(D = 768\\) in our model. For each video, we randomly sample 8 frames while preserving their order in-between. For fine-tuning on retrieval tasks, we only fine" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.632, + 0.892, + 0.738 + ], + "angle": 0, + "content": "tune the uni-modal encoders of Clover with InfoNCE [48] loss. For fine-tuning on video QA task, we add a simple MLPs that takes the multi-modal [CLS] embedding as input for classification, and optimize the whole Clover model with cross-entropy loss. During fine-tuning, we follow the conventional set-up in [9, 36], and all the fine-tuning experiments are performed on 8 NVIDIA A100 GPUs." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.755, + 0.767, + 0.771 + ], + "angle": 0, + "content": "4.2. Comparing to State-of-the-art" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.78, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Text-to-video retrieval. Tab. 1 shows the retrieval results on the DiDeMo [1], MSRVTT [55] and LSMDC [38] datasets with both zero-shot and fine-tuning settings. Our Clover surpasses previous works on all the datasets by a large margin. The significant performance gain under zero-shot evaluation demonstrates its stronger generalization ability. Specifically, Clover brings \\(4.9\\%\\) gain in R@10 over MCQ [11], the current SOTA method on re" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.518, + 0.957 + ], + "angle": 0, + "content": "14861" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.123, + 0.089, + 0.852, + 0.189 + ], + "angle": 0, + "content": "
MethodDiDeMoMSRVTTLSMDC-MCTGIF-Frame
R@1R@5R@10MedRR@1R@5R@10MedRAccAcc
Baseline22.647.958.2719.841.751.11078.868.9
+ TMA24.749.760.0622.742.252.2980.069.3
+ TMA+SM25.349.560.5622.542.752.0980.569.4
+ TMA+SM+RankL26.451.161.3523.443.352.4980.769.7
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.199, + 0.895, + 0.242 + ], + "angle": 0, + "content": "Table 3. Effects of our pre-training tasks for Clover. We report zero-shot text-video retrieval performance on DiDeMo and MSRVTT, and fine-tune video QA performance on LSMDC-MC and TGIF-Frame. TMA, SM and RankL: tri-modal alignment, semantic masking strategy and pair-wise ranking loss." + }, + { + "type": "table", + "bbox": [ + 0.085, + 0.248, + 0.466, + 0.302 + ], + "angle": 0, + "content": "
MethodAveraged similarity scores of positive pairsAveraged similarity margin between positive and negative samples
w/o TMA0.560.30
w TMA0.650.35
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.312, + 0.47, + 0.355 + ], + "angle": 0, + "content": "Table 4. Affect of TMA on the video and text representations in learned embedding space. The experiment is conducted on MSRVTT-1kA test set under zero-shot setting." + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.365, + 0.47, + 0.576 + ], + "angle": 0, + "content": "trieval tasks with the same pre-train data, averaged on the three datasets. Moreover, Clover outperforms VIOLET [9], though VIOLET adopts a considerably larger pre-training dataset (YTT180M [61] of 180M visual-text pairs vs. WebVid2M+CC3M of only 5.5M samples). Clover also presents superior performance when fine-tuned on the downstream datasets, compared with the SOTAs. On all three datasets, Clover surpasses previous works by a large margin across all the metrics and obtains an average improvement of \\(8.7\\%\\) on R@10. Moreover, different from some previous work [9,23,24] that adopts a joint multi-modal encoder and requires exhaustively pairing every video and text during retrieval, Clover directly deploys the video encoder and text encoder for retrieval tasks, making it much more efficient." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.577, + 0.471, + 0.775 + ], + "angle": 0, + "content": "Video question answering. Tab. 2 reports the results of Clover and current SOTAs on four video QA datasets. Though using much less pre-training data, Clover outperforms another unified VidL model i.e., VIOLET [9] significantly. Moreover, compared to the current SOTA method MERLOT [61], Clover brings improvements of \\(1.0\\%\\) on TGIF-Action, \\(2.0\\%\\) on TGIF-Transition, \\(2.1\\%\\) on TGIF-Frame, \\(4.3\\%\\) on MSRVTT-MC, \\(1.0\\%\\) on MSRVTT-QA, \\(2.0\\%\\) on LSMDC-MC and \\(1.2\\%\\) on LSMDC-FiB. Note that JustAsk [58] and MERLOT [61] are specifically designed for video QA and trained on orders of magnitude larger datasets (e.g., HTVQA69M [58], YTT180M [61]). The experiment results again confirm the superiority of Clover." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.786, + 0.182, + 0.803 + ], + "angle": 0, + "content": "4.3. Analysis" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.811, + 0.471, + 0.903 + ], + "angle": 0, + "content": "We conduct ablation experiments on two retrieval datasets (DiDeMo and MSRvTT) and two Video QA datasets (TGIF-Frame and LSMDC-MC). Due to the computational resource limit, we randomly sample 1 million video-text pairs from WebVid2M [2] to build the WebVid1M subset for model pre-training under all the ablation stud" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.25, + 0.892, + 0.385 + ], + "angle": 0, + "content": "ies. We keep the model architecture unchanged but pretrain the model with only MLM and InfoNCE losses as a baseline, which is similar to the baseline adopted in VIOLET [9]. The only difference is that we replace the ITM applied to the cross-modal encoder outputs in VIOLET [9] with InfoNCE applied to the outputs of uni-modal encoders and separate the text encoder from the cross-modal encoder. All the ablation experiments are conducted with batch size 1024 on 32 NVIDIA A100 GPUs." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.387, + 0.893, + 0.809 + ], + "angle": 0, + "content": "Effect of tri-modal alignment. Tri-modal alignment (TMA) aims to better correlate cross-modal alignment and fusion. To evaluate the contribution of TMA, we remove the pair-wise ranking objective and semantic masking strategy, and employ the classical MLM task as in the baseline method. Compared with the baseline, TMA explicitly associates the outputs of uni-modal encoders with the outputs of the multi-modal encoder. As shown in Tab. 3, the model trained with TMA outperforms the baseline on both retrieval and video QA, demonstrating its effectiveness. For a better understanding of the effect of TMA, we further report averaged cosine similarity of videos and texts in MSRVTT dataset (in zero-shot setting) calculated by models trained with/without TMA, respectively. For a fair comparison, we select all video queries (197 in total) that were successfully responded to by both models. We recall 100 results for each query, and take the ground-truth pairs as positive and others as negative. As shown in Tab. 4, the model armed with tri-modal alignment assigns higher similarity scores to positive pairs compared to the model without it. Meanwhile, with tri-modal alignment, the margin between the positive and negative samples is also increased. It reveals that with the help of tri-modal alignment, the distance between video and text that contain consistent semantics is reduced, while the margin between the misaligned samples is increased. The results demonstrate that with fused multi-modal representation as anchors, tri-modal alignment helps the video and language modalities to be better aligned." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.811, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Effect of semantic masking strategy. We incorporate a semantic masking strategy to form the masked pairs in trimodal alignment task. As shown in Tab. 3, on more diverse evaluation benchmarks, e.g. DiDemo, where each video is paired with 5 different ground-truth texts, using masked samples brings more gains. The result reveals that the pro" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "14862" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.123, + 0.071, + 0.852, + 0.183 + ], + "angle": 0, + "content": "
MethodModel ArchitectureTraining ObjectivesMSRVTTDiDeMoTGIF-FrameLSMDC-MC
Two uni-modal encoderCross-modal encoderInfoNCEMLMTMA + RanklossR@1R@1AccAcc
IND-A✓✓20.722.9--
IND-F✓✓✓--69.079.0
COMB✓✓✓✓19.822.668.978.8
Clover✓✓✓✓23.426.469.780.7
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.195, + 0.892, + 0.209 + ], + "angle": 0, + "content": "Table 5. Comparisons with task-independently trained models (IND) and naive combination of the uni- and cross- modal encoder (COMB)" + }, + { + "type": "image", + "bbox": [ + 0.083, + 0.223, + 0.182, + 0.279 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.186, + 0.223, + 0.276, + 0.279 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.282, + 0.223, + 0.371, + 0.279 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.378, + 0.223, + 0.468, + 0.279 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.086, + 0.28, + 0.375, + 0.306 + ], + "angle": 0, + "content": "Baseline: a man is showing the interior of a car \nClover: inside of a car with large space and safety" + }, + { + "type": "image", + "bbox": [ + 0.086, + 0.314, + 0.18, + 0.368 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.188, + 0.314, + 0.275, + 0.368 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.283, + 0.314, + 0.371, + 0.368 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.381, + 0.314, + 0.466, + 0.368 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.086, + 0.371, + 0.328, + 0.396 + ], + "angle": 0, + "content": "Baseline: a man is folding pieces of paper \nClover: a person folds a paper airplane" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.413, + 0.47, + 0.441 + ], + "angle": 0, + "content": "Figure 3. Qualitative results of zero-shot video to text retrieval results on MSRVTT [55]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.447, + 0.469, + 0.491 + ], + "angle": 0, + "content": "posed semantic masking strategy facilitates the model to capture key semantic information in more complex scenes and makes it achieve better results." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.492, + 0.469, + 0.598 + ], + "angle": 0, + "content": "Effect of pair-wise ranking. To make the model capture the semantic information difference between the masked pairs and complete pairs, we adopt the pair-wise rank loss. As shown in Tab. 3, with pair-wise ranking loss, the performance of the model is further improved. We also show the visualization results below to further illustrate the effect of pair-wise ranking." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.599, + 0.469, + 0.81 + ], + "angle": 0, + "content": "Clover makes cross-modal alignment and fusion mutually improving. We compare Clover with the ones that use the same architecture but employ different pre-training tasks in Tab. 5. \\( IND-A \\) represents the model only trained with InfoNCE loss for cross-modal alignment. \\( IND-F \\) indicates the model trained with MLM loss for cross-modal fusion. We also report results achieved by the model (i.e., COMB) that simply combines the uni-modal encoders and multi-modal encoder as well as the training objectives. We can see that the simple combination hurts the performance, while our Clover achieves superior performance than the task-independently trained \\( IND-A \\) and \\( IND-F \\). It reveals that Clover is able to get the model's cross-modal alignment and fusion capability mutually enhanced." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.811, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Qualitative analysis. In Fig. 3 and Fig. 4, we show the qualitative results of Clover and the baseline method on zero-shot video-text retrieval and VQA tasks. Specifically, we present the query videos with the matched texts, and the video-question pair with the answer. Clover empowers the model with stronger video-text understanding capabilities." + }, + { + "type": "image", + "bbox": [ + 0.507, + 0.222, + 0.599, + 0.276 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.605, + 0.222, + 0.697, + 0.276 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.701, + 0.222, + 0.794, + 0.276 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.799, + 0.222, + 0.892, + 0.276 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.509, + 0.278, + 0.708, + 0.3 + ], + "angle": 0, + "content": "Question: What is a guy hugging? \nBaseline: Animal Clover: Lion" + }, + { + "type": "image", + "bbox": [ + 0.507, + 0.31, + 0.598, + 0.364 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.605, + 0.31, + 0.696, + 0.364 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.701, + 0.31, + 0.794, + 0.364 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.799, + 0.31, + 0.892, + 0.364 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.509, + 0.368, + 0.781, + 0.391 + ], + "angle": 0, + "content": "Question: What does a person wash with a brush? Baseline: Face Clover: Toy" + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.41, + 0.892, + 0.437 + ], + "angle": 0, + "content": "Figure 4. Qualitative results of video question answering results on MSRVTT-QA [52]." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.44, + 0.892, + 0.53 + ], + "angle": 0, + "content": "ity. For example, for the second video in Fig. 3, Clover returns the text with \"paper airplane\", which is a more accurate description of the video content; for the first video in Fig. 4, Clover generates a more accurate answer \"Lion\" than the baseline answer \"Animal\". The results demonstrate the superiority of Clover." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.544, + 0.619, + 0.558 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.57, + 0.892, + 0.765 + ], + "angle": 0, + "content": "In this paper, we present Clover, a new end-to-end Video-Language pre-training method for both high-efficiency video-text retrieval and video question answering. Clover introduces a novel Tri-modal Alignment task to better align the representations from visual, text and fused modalities, which explicitly correlate the uni-modal encoder and multi-modal encoder. It also introduces semantic masking strategy and pair-wise ranking loss to further improve the cross-modality modal training. Extensive experiments conducted on the three retrieval datasets and eight video QA datasets clearly demonstrated, as a general video-text model, its consistent superiority for video-text understanding." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.766, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Acknowledgements. This work was supported by National Key R&D Program of China (No.2022ZD0118201), the National Science Fund for Distinguished Young Scholars (No.62025603), the National Natural Science Foundation of China (No. U21B2037, No. U22B2051, No. 62176222, No. 62176223, No. 62176226, No. 62072386, No. 62072387, No. 62072389, No. 62002305 and No. 62272401), and the Natural Science Foundation of Fujian Province of China (No.2021J01002, No.2022J06001)." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "14863" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.079, + 0.09, + 0.176, + 0.107 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.116, + 0.473, + 0.185 + ], + "angle": 0, + "content": "[1] Lisa Anne Hendricks, Oliver Wang, Eli Shechtman, Josef Sivic, Trevor Darrell, and Bryan Russell. Localizing moments in video with natural language. In Proceedings of the IEEE international conference on computer vision, pages 5803-5812, 2017. 1, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.187, + 0.472, + 0.257 + ], + "angle": 0, + "content": "[2] Max Bain, Arsha Nagrani, Gül Varol, and Andrew Zisserman. Frozen in time: A joint video and image encoder for end-to-end retrieval. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1728-1738, 2021. 1, 2, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.259, + 0.472, + 0.3 + ], + "angle": 0, + "content": "[3] Hangbo Bao, Li Dong, Songhao Piao, and Furu Wei. Beit: Bert pre-training of image transformers. In International Conference on Learning Representations, 2021. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.302, + 0.472, + 0.37 + ], + "angle": 0, + "content": "[4] David Chen and William B Dolan. Collecting highly parallel data for paraphrase evaluation. In Proceedings of the 49th annual meeting of the association for computational linguistics: human language technologies, pages 190–200, 2011. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.373, + 0.472, + 0.43 + ], + "angle": 0, + "content": "[5] Xinlei Chen, Hao Fang, Tsung-Yi Lin, Ramakrishna Vedantam, Saurabh Gupta, Piotr Dólár, and C Lawrence Zitnick. Microsoft coco captions: Data collection and evaluation server. arXiv preprint arXiv:1504.00325, 2015. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.432, + 0.472, + 0.487 + ], + "angle": 0, + "content": "[6] Xing Cheng, Hezheng Lin, Xiangyu Wu, Fan Yang, and Dong Shen. Improving video-text retrieval by multi-stream corpus alignment and dual softmax loss. arXiv preprint arXiv:2109.04290, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.489, + 0.472, + 0.586 + ], + "angle": 0, + "content": "[7] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 4171–4186, 2019. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.588, + 0.472, + 0.669 + ], + "angle": 0, + "content": "[8] Chenyou Fan, Xiaofan Zhang, Shu Zhang, Wensheng Wang, Chi Zhang, and Heng Huang. Heterogeneous memory enhanced multimodal attention model for video question answering. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1999-2007, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.673, + 0.472, + 0.729 + ], + "angle": 0, + "content": "[9] Tsu-Jui Fu, Linjie Li, Zhe Gan, Kevin Lin, William Yang Wang, Lijuan Wang, and Zicheng Liu. Violet: End-to-end video-language transformers with masked visual-token modeling. arXiv preprint arXiv:2111.12681, 2021. 1, 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.731, + 0.472, + 0.786 + ], + "angle": 0, + "content": "[10] Valentin Gabeur, Chen Sun, Karteek Alahari, and Cordelia Schmid. Multi-modal transformer for video retrieval. In European Conference on Computer Vision, pages 214-229. Springer, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.788, + 0.472, + 0.857 + ], + "angle": 0, + "content": "[11] Yuying Ge, Yixiao Ge, Xihui Liu, Dian Li, Ying Shan, Xiaohu Qie, and Ping Luo. Bridging video-text retrieval with multiple choice questions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16167-16176, 2022. 1, 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.859, + 0.472, + 0.902 + ], + "angle": 0, + "content": "[12] Yuying Ge, Yixiao Ge, Xihui Liu, Alex Jinpeng Wang, Jianping Wu, Ying Shan, Xiaohu Qie, and Ping Luo. Miles: Visual bert pre-training with injected language semantics for" + }, + { + "type": "list", + "bbox": [ + 0.079, + 0.116, + 0.473, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "video-text retrieval. Proceedings of the European Conference on Computer Vision (ECCV), 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.122, + 0.894, + 0.188 + ], + "angle": 0, + "content": "[13] Simon Ging, Mohammadreza Zolfaghari, Hamed Pirsiavash, and Thomas Brox. Coot: Cooperative hierarchical transformer for video-text representation learning. Advances in neural information processing systems, 33:22605-22618, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.19, + 0.894, + 0.272 + ], + "angle": 0, + "content": "[14] Satya Krishna Gorti, Noel Vouitsis, Junwei Ma, Keyvan Golestan, Maksims Volkovs, Animesh Garg, and Guangwei Yu. X-pool: Cross-modal language-video attention for text-video retrieval. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5006-5015, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.274, + 0.895, + 0.343 + ], + "angle": 0, + "content": "[15] Deng Huang, Peihao Chen, Runhao Zeng, Qing Du, Mingkui Tan, and Chuang Gan. Location-aware graph convolutional networks for video question answering. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 34, pages 11021-11028, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.344, + 0.894, + 0.412 + ], + "angle": 0, + "content": "[16] Yunseok Jang, Yale Song, Youngjae Yu, Youngjin Kim, and Gunhee Kim. Tgif-qa: Toward spatio-temporal reasoning in visual question answering. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2758–2766, 2017. 1, 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.414, + 0.892, + 0.468 + ], + "angle": 0, + "content": "[17] Jie Jiang, Shaobo Min, Weijie Kong, Hongfa Wang, Zhifeng Li, and Wei Liu. Tencent text-video retrieval: Hierarchical cross-modal interactions with multi-level representations. IEEE Access, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.47, + 0.892, + 0.537 + ], + "angle": 0, + "content": "[18] Will Kay, Joao Carreira, Karen Simonyan, Brian Zhang, Chloe Hillier, Sudheendra Vijayanarasimhan, Fabio Viola, Tim Green, Trevor Back, Paul Natsev, et al. The kinetics human action video dataset. arXiv preprint arXiv:1705.06950, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.539, + 0.895, + 0.607 + ], + "angle": 0, + "content": "[19] Junyeong Kim, Minuk Ma, Kyungsu Kim, Sungjin Kim, and Chang D Yoo. Progressive attention memory network for movie story question answering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8337-8346, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.609, + 0.892, + 0.691 + ], + "angle": 0, + "content": "[20] Ranjay Krishna, Yuke Zhu, Oliver Groth, Justin Johnson, Kenji Hata, Joshua Kravitz, Stephanie Chen, Yannis Kalantidis, Li-Jia Li, David A Shamma, et al. Visual genome: Connecting language and vision using crowdsourced dense image annotations. International journal of computer vision, 123(1):32-73, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.692, + 0.892, + 0.748 + ], + "angle": 0, + "content": "[21] Alexander Kunitsyn, Maksim Kalashnikov, Maksim Dzabraev, and Andrei Ivaniuta. Mdmt-2: Multidomain multimodal transformer for video retrieval, one more step towards generalization. arXiv preprint arXiv:2203.07086, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.749, + 0.892, + 0.816 + ], + "angle": 0, + "content": "[22] Thao Minh Le, Vuong Le, Svetha Venkatesh, and Truyen Tran. Hierarchical conditional relation networks for video question answering. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9972-9981, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.818, + 0.892, + 0.899 + ], + "angle": 0, + "content": "[23] Jie Lei, Linjie Li, Luowei Zhou, Zhe Gan, Tamara L Berg, Mohit Bansal, and Jingjing Liu. Less is more: Clipbert for video-and-language learning via sparse sampling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7331-7341, 2021. 1, 2, 5, 6, 7" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.895, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "14864" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.162 + ], + "angle": 0, + "content": "[24] Dongxu Li, Junnan Li, Hongdong Li, Juan Carlos Niebles, and Steven CH Hoi. Align and prompt: Video-and-language pre-training with entity prompts. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4953-4963, 2022. 1, 2, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.165, + 0.472, + 0.234 + ], + "angle": 0, + "content": "[25] Linjie Li, Yen-Chun Chen, Yu Cheng, Zhe Gan, Licheng Yu, and Jingjing Liu. Hero: Hierarchical encoder for video+ language omni-representation pre-training. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 2046-2065, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.236, + 0.471, + 0.319 + ], + "angle": 0, + "content": "[26] Xiangpeng Li, Jingkuan Song, Lianli Gao, Xianglong Liu, Wenbing Huang, Xiangnan He, and Chuang Gan. Beyond rnns: Positional self-attention with co-attention for video question answering. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 33, pages 8658–8665, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.322, + 0.47, + 0.39 + ], + "angle": 0, + "content": "[27] Chengzhi Lin, Ancong Wu, Junwei Liang, Jun Zhang, Wenhang Ge, Wei-Shi Zheng, and Chunhua Shen. Text-adaptive multiple visual prototype matching for video-text retrieval. Advances in Neural Information Processing Systems, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.394, + 0.47, + 0.45 + ], + "angle": 0, + "content": "[28] Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dólár. Focal loss for dense object detection. In Proceedings of the IEEE international conference on computer vision, pages 2980-2988, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.452, + 0.469, + 0.521 + ], + "angle": 0, + "content": "[29] Fei Liu, Jing Liu, Weining Wang, and Hanqing Lu. Hair: Hierarchical visual-semantic relational reasoning for video question answering. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1698–1707, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.524, + 0.469, + 0.58 + ], + "angle": 0, + "content": "[30] Ruyang Liu, Jingjia Huang, Ge Li, Jiashi Feng, Xinglong Wu, and Thomas H Li. Revisiting temporal modeling for clip-based image-to-video knowledge transferring. arXiv preprint arXiv:2301.11116, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.582, + 0.469, + 0.637 + ], + "angle": 0, + "content": "[31] Yang Liu, Samuel Albanie, Arsha Nagrani, and Andrew Zisserman. Use what you have: Video retrieval using representations from collaborative experts. In BMVC, 2019. 1, 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.64, + 0.469, + 0.696 + ], + "angle": 0, + "content": "[32] Ze Liu, Jia Ning, Yue Cao, Yixuan Wei, Zheng Zhang, Stephen Lin, and Han Hu. Video swim transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3202-3211, 2022. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.698, + 0.469, + 0.726 + ], + "angle": 0, + "content": "[33] Edward Loper and Steven Bird. Nltk: The natural language toolkit. arXiv preprint cs/0205028, 2002. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.729, + 0.469, + 0.77 + ], + "angle": 0, + "content": "[34] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In International Conference on Learning Representations, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.773, + 0.469, + 0.842 + ], + "angle": 0, + "content": "[35] Huaishao Luo, Lei Ji, Botian Shi, Haoyang Huang, Nan Duan, Tianrui Li, Jason Li, Taroon Bharti, and Ming Zhou. Univl: A unified video and language pre-training model for multimodal understanding and generation. arXiv preprint arXiv:2002.06353, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.845, + 0.469, + 0.9 + ], + "angle": 0, + "content": "[36] Huaishao Luo, Lei Ji, Ming Zhong, Yang Chen, Wen Lei, Nan Duan, and Tianrui Li. Clip4clip: An empirical study of clip for end to end video clip retrieval. arXiv preprint arXiv:2104.08860, 2021. 6" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.092, + 0.892, + 0.162 + ], + "angle": 0, + "content": "[37] Yiwei Ma, Guohai Xu, Xiaoshuai Sun, Ming Yan, Ji Zhang, and Rongrong Ji. X-clip: End-to-end multi-grained contrastive learning for video-text retrieval. In Proceedings of the 30th ACM International Conference on Multimedia, pages 638–647, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.165, + 0.892, + 0.245 + ], + "angle": 0, + "content": "[38] Tegan Maharaj, Nicolas Ballas, Anna Rohrbach, Aaron Courville, and Christopher Pal. A dataset and exploration of models for understanding video data through fill-in-the-blank question-answering. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 6884–6893, 2017. 1, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.249, + 0.892, + 0.331 + ], + "angle": 0, + "content": "[39] Antoine Miech, Jean-Baptiste Alayrac, Lucas Smaira, Ivan Laptev, Josef Sivic, and Andrew Zisserman. End-to-end learning of visual representations from uncurated instructional videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9879-9889, 2020. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.334, + 0.892, + 0.417 + ], + "angle": 0, + "content": "[40] Antoine Miech, Dimitri Zhukov, Jean-Baptiste Alayrac, Makarand Tapaswi, Ivan Laptev, and Josef Sivic. Howto100m: Learning a text-video embedding by watching hundred million narrated video clips. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2630-2640, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.42, + 0.892, + 0.488 + ], + "angle": 0, + "content": "[41] Jungin Park, Jiyoung Lee, and Kwanghoon Sohn. Bridge to answer: Structure-aware graph interaction network for video question answering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15526-15535, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.491, + 0.892, + 0.587 + ], + "angle": 0, + "content": "[42] Andrew Rouditchenko, Angie Boggust, David Harwath, Brian Chen, Dhiraj Joshi, Samuel Thomas, Kartik Audhkhasi, Hilde Kuehne, Rameswar Panda, Rogerio Feris, et al. Avlnet: Learning audio-visual language representations from instructional videos. In Annual Conference of the International Speech Communication Association. International Speech Communication Association, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.59, + 0.892, + 0.657 + ], + "angle": 0, + "content": "[43] Paul Hongsuck Seo, Arsha Nagrani, and Cordelia Schmid. Look before you speak: Visually contextualized utterances. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16877-16887, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.661, + 0.892, + 0.744 + ], + "angle": 0, + "content": "[44] Piyush Sharma, Nan Ding, Sebastian Goodman, and Radu Soricut. Conceptual captions: A cleaned, hypernymed, image alt-text dataset for automatic image captioning. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 2556-2565, 2018. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.747, + 0.892, + 0.815 + ], + "angle": 0, + "content": "[45] Chen Sun, Austin Myers, Carl Vondrick, Kevin Murphy, and Cordelia Schmid. Videobert: A joint model for video and language representation learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7464-7473, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.818, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[46] Hao Tan and Mohit Bansal. Lxmert: Learning cross-modality encoder representations from transformers. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages 5100-5111, 2019. 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.092, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "14865" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.146 + ], + "angle": 0, + "content": "[47] Atousa Torabi, Niket Tandon, and Leonid Sigal. Learning language-visual embedding for movie understanding with natural-language. arXiv preprint arXiv:1609.08124, 2016. 1, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.149, + 0.47, + 0.189 + ], + "angle": 0, + "content": "[48] Aaron Van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding. arXiv e-prints, pages arXiv-1807, 2018. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.191, + 0.47, + 0.259 + ], + "angle": 0, + "content": "[49] Alex Jinpeng Wang, Yixiao Ge, Rui Yan, Yuying Ge, Xudong Lin, Guanyu Cai, Jianping Wu, Ying Shan, Xiaohu Qie, and Mike Zheng Shou. All in one: Exploring unified video-language pre-training. arXiv preprint arXiv:2203.07303, 2022. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.261, + 0.47, + 0.329 + ], + "angle": 0, + "content": "[50] Jinpeng Wang, Yixiao Ge, Guanyu Cai, Rui Yan, Xudong Lin, Ying Shan, Xiaohu Qie, and Mike Zheng Shou. Object-aware video-language pre-training for retrieval. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3313-3322, 2022. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.331, + 0.47, + 0.398 + ], + "angle": 0, + "content": "[51] Yi Wang, Kunchang Li, Yizhuo Li, Yinan He, Bingkun Huang, Zhiyu Zhao, Hongjie Zhang, Jilan Xu, Yi Liu, Zun Wang, et al. Internvideo: General video foundation models via generative and discriminative learning. arXiv preprint arXiv:2212.03191, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.401, + 0.47, + 0.469 + ], + "angle": 0, + "content": "[52] Dejing Xu, Zhou Zhao, Jun Xiao, Fei Wu, Hanwang Zhang, Xiangnan He, and Yueting Zhuang. Video question answering via gradually refined attention over appearance and motion. In Proceedings of the 25th ACM international conference on Multimedia, pages 1645-1653, 2017. 1, 2, 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.47, + 0.47, + 0.552 + ], + "angle": 0, + "content": "[53] Hu Xu, Gargi Ghosh, Po-Yao Huang, Prahal Arora, Masoumeh Aminzadeh, Christoph Feichtenhofer, Florian Metze, and Luke Zettlemoyer. Vlm: Task-agnostic videolanguage model pre-training for video understanding. In Findings of the Association for Computational Linguistics: ACL-IJCNLP 2021, pages 4227–4239, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.554, + 0.47, + 0.636 + ], + "angle": 0, + "content": "[54] Hu Xu, Gargi Ghosh, Po-Yao Huang, Dmytro Okhonko, Armen Aghajanyan, Florian Metze, Luke Zettlemoyer, and Christoph Feichtenhofer. Videoclip: Contrastive pre-training for zero-shot video-text understanding. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 6787-6800, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.638, + 0.47, + 0.692 + ], + "angle": 0, + "content": "[55] Jun Xu, Tao Mei, Ting Yao, and Yong Rui. Msr-vtt: A large video description dataset for bridging video and language. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5288-5296, 2016. 1, 5, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.693, + 0.47, + 0.775 + ], + "angle": 0, + "content": "[56] Hongwei Xue, Tiankai Hang, Yanhong Zeng, Yuchong Sun, Bei Liu, Huan Yang, Jianlong Fu, and Baining Guo. Advancing high-resolution video-language representation with large-scale video transcriptions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5036-5045, June 2022. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.777, + 0.47, + 0.831 + ], + "angle": 0, + "content": "[57] Hongwei Xue, Yuchong Sun, Bei Liu, Jianlong Fu, Ruihua Song, Houqiang Li, and Jiebo Luo. Clip-vip: Adapting pretrained image-text model to video-language representation alignment. arXiv preprint arXiv:2209.06430, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.833, + 0.47, + 0.901 + ], + "angle": 0, + "content": "[58] Antoine Yang, Antoine Miech, Josef Sivic, Ivan Laptev, and Cordelia Schmid. Just ask: Learning to answer questions from millions of narrated videos. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1686-1697, 2021. 1, 2, 6, 7" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.893, + 0.161 + ], + "angle": 0, + "content": "[59] Jianwei Yang, Yonatan Bisk, and Jianfeng Gao. Taco: Token-aware cascade contrastive learning for video-text alignment. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11562-11572, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.164, + 0.893, + 0.219 + ], + "angle": 0, + "content": "[60] Youngjae Yu, Jongseok Kim, and Gunhee Kim. A joint sequence fusion model for video question answering and retrieval. In Proceedings of the European Conference on Computer Vision (ECCV), pages 471-487, 2018. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.221, + 0.893, + 0.275 + ], + "angle": 0, + "content": "[61] Rowan Zellers, Ximing Lu, Jack Hessel, Youngjae Yu, Jae Sung Park, Jize Cao, Ali Farhadi, and Yejin Choi. Merlot: Multimodal neural script knowledge models. Advances in Neural Information Processing Systems, 34, 2021. 1, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.277, + 0.893, + 0.345 + ], + "angle": 0, + "content": "[62] Zheng-Jun Zha, Jiawei Liu, Tianhao Yang, and Yongdong Zhang. Spatiotemporal-textual co-attention network for video question answering. ACM Transactions on Multimedia Computing, Communications, and Applications (TOMM), 15(2s):1-18, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.348, + 0.893, + 0.402 + ], + "angle": 0, + "content": "[63] Linchao Zhu and Yi Yang. Actbert: Learning global-local video-text representations. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8746-8755, 2020. 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.893, + 0.402 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "14866" + } + ] +] \ No newline at end of file diff --git a/2023/Clover_ Towards a Unified Video-Language Alignment and Fusion Model/eddf9a27-0fae-4139-acd3-5b79001d8460_origin.pdf b/2023/Clover_ Towards a Unified Video-Language Alignment and Fusion Model/eddf9a27-0fae-4139-acd3-5b79001d8460_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..1f1b681aa31d164c5ebed11e7822d5e4d2333117 --- /dev/null +++ b/2023/Clover_ Towards a Unified Video-Language Alignment and Fusion Model/eddf9a27-0fae-4139-acd3-5b79001d8460_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:259ffbfdfd76b99b262ea75c378405b6d2457365a939181cf90bce5ad84234ca +size 1912148 diff --git a/2023/Clover_ Towards a Unified Video-Language Alignment and Fusion Model/full.md b/2023/Clover_ Towards a Unified Video-Language Alignment and Fusion Model/full.md new file mode 100644 index 0000000000000000000000000000000000000000..bf339ee6ac403245d830aeac6dbf03d92afd8415 --- /dev/null +++ b/2023/Clover_ Towards a Unified Video-Language Alignment and Fusion Model/full.md @@ -0,0 +1,330 @@ +# Clover: Towards A Unified Video-Language Alignment and Fusion Model + +Jingjia Huang\* + +Yinan Li\* + +Jiashi Feng + +Xinglong Wu + +Xiaoshuai Sun + +Rongrong Ji + +$^{\text{心}}$ Key Laboratory of Multimedia Trusted Perception and Efficient Computing, Ministry of Education of China, Xiamen University, 361005, China + +ByteDance Inc, 100043, China + +yinanlee@stu.xmu.edu.cn, {xssun, rrji} @ xmu.edu.cn + +{huangjingjia, jshfeng, wuxinglong} @bytedance.com + +# Abstract + +Building a universal Video-Language model for solving various video understanding tasks (e.g., text-video retrieval, video question answering) is an open challenge to the machine learning field. Towards this goal, most recent works build the model by stacking uni-modal and cross-modal feature encoders and train it with pair-wise contrastive pre-text tasks. Though offering attractive generality, the resulted models have to compromise between efficiency and performance. They mostly adopt different architectures to deal with different downstream tasks. We find this is because the pair-wise training cannot well align and fuse features from different modalities. We then introduce Clover—a Correlated Video-Language pre-training method—towards a universal Video-Language model for solving multiple video understanding tasks with neither performance nor efficiency compromise. It improves cross-modal feature alignment and fusion via a novel tri-modal alignment pre-training task. Additionally, we propose to enhance the tri-modal alignment via incorporating learning from semantic masked samples and a new pair-wise ranking loss. Clover establishes new state-of-the-arts on multiple downstream tasks, including three retrieval tasks for both zero-shot and fine-tuning settings, and eight video question answering tasks. Codes and pre-trained models will be released at https://github.com/LeeYN-43/Clover. + +# 1. Introduction + +Video-Language pre-training (VidL) aims to learn generalizable multi-modal models from large-scale video-text + +samples so as to better solve various challenging Video-Language understanding tasks, such as text-video retrieval [1,4,38,55] and video question answering [16,47,52]. Recent studies [9,11,23,24,49,56,58,61] have shown that VidL leads to significant performance improvement and achieves state-of-the-art results on various downstream text-video retrieval and video question answering (VQA) benchmarks. + +Though achieving encouraging performance, existing VidL models mostly adopt different architectures to deal with different downstream tasks. For the text-video retrieval tasks, they [2, 10, 11, 13, 31, 39] typically use two individual uni-modal encoders for processing video and text data separately, for the sake of retrieval efficiency. While for video question answering tasks, the models usually adopt the multi-modal joint encoder design to learn the association and interaction of different modalities. + +Building a unified model capable of solving various Video-Language tasks is a long-standing challenge for machine learning research. A few recent works [9,24] attempt to learn a unified VidL model for both tasks, which uses the multi-modal encoder to conduct text-video retrieval. However, the model requires an exhaustive pair-wise comparison between the query texts and gallery videos. Given $N$ text queries and $M$ category videos, the computation complexity of the multi-modal encoder model would be $O(NM)$ , which makes it infeasible for large-scale videotext retrieval applications. Another straightforward solution is to simply combine the uni-modal and multi-modal encoders (Fig. 1 (a)), and perform the retrieval and VQA tasks through the uni-modal and multi-modal encoders respectively. Its computation complexity for retrieval tasks is only $O(N + M)$ . However, the experiment results in Fig.1 (c) show that, without a carefully designed correlating mechanism between these two types of encoders, the simple combination i.e., COMB yields a compromised per + +formance compared with the models i.e. IND individually designed for retrieval and VQA. + +In this work, we aim to address the above issues and build a unified pre-training model that attains high efficiency and performance simultaneously. We observe: (i) well aligning the features of the video and text from the same data pair is important for text-video matching; (ii) effectively fusing video and text features into unified representations is critical for video-text understanding. However, existing pre-training strategies that rely on either simple supervised or contrastive pre-text tasks hardly achieve promising feature alignment and fusion capability simultaneously. Motivated by these observations, we develop a new VidL method from these two aspects. Specifically, we propose Correlated Video-Language pre-training (Clover), a VidL method that not only unifies Video-Language alignment and fusion, but also makes them mutually boosted. The fused multi-modal representation contains richer context information than the uni-modal representations [11,35]. As an intermediate modality between video and text, the multi-modal representations are good anchors for cross-modality alignment. Meanwhile, keeping the fused representation closer to the uni-modal representation containing consistent semantic information and away from the inconsistent one will enhance the learning of semantic information in the fused modality. Therefore, we propose the Tri-Modal Alignment (TMA) to get Video-Language alignment and fusion mutually boosted, which takes the alignment between the multimodal representation and text/video representations as an auxiliary objective. We note that since the tri-modal alignment is well compatible with the classical pre-training tasks [3,7,35] e.g., Masked Language Modeling, its computation overhead is negligible. To help the model maintain fine-grained discriminative capability while improving its generalizability, we further introduce a pair-wise ranking loss that urges the model to be aware of the concept missing in masked samples compared to original samples. + +Extensive experiments are conducted on multiple downstream tasks, including three retrieval tasks with different experimental setups (i.e. zero-shot and fine-tune) and eight video question answering tasks. The results demonstrate that Clover is able to get the cross-modal fusion and alignment capability mutually improved, and consistently outperforms current SOTAs on various downstream tasks. It achieves an average performance improvement of $4.9\%$ and $8.7\%$ Recall@10 score on the zero-shot and fine-tune settings of the three downstream retrieval datasets, while the average accuracy improvement over current SOTAs is $2.3\%$ on the eight video question answering datasets. + +In summary, we make the following contributions: (1) we introduce Clover, a pre-training method achieving unified Video-Language alignment and fusion model that can be easily transferred to various downstream video under + +standing tasks while attaining both high efficiency and performance; (2) we propose a novel tri-modal alignment pretraining task, which correlates the uni-modal encoder and multi-modal encoder to get them mutually boosted. + +# 2. Related Work + +Video-Language pre-training for text-video retrieval. Existing pre-training methods for text-video retrieval can be mainly divided into two categories. The first kind of method employs two individual encoders to embed video and text, and project them into a common latent space [2,10,11,13,31,39,42,50,54,59]. Then, they leverage a contrastive objective [48] to align the paired representations. Thanks to their high efficiency, this category of methods is widely used in real-world applications. The other kind of method utilizes a joint multi-modal encoder to learn comprehensive representation for a given video-text pair, and predict whether the video and text are matched or not via a binary classifier [9,23,25,45,53,63]. Despite the good performance they achieved, they need to exhaustively pair all the videos and texts, and feed them into the model during inference, making them highly inefficient. + +Video-Language pre-training for video question answering. Video question answering aims to automatically answer natural language questions given a context video, which requires a joint understanding of both the video and language. Previous works mainly focus on designing better spatio-temporal attention networks [8,16,26,29,52,62] and question-video relation networks [15, 19, 22, 41] to extract more accurate multi-modal representations. Recently, many works leverage transformer-based models to build an additional multi-modal encoder upon the uni-modal encoders [9, 24, 43] or input image patches and word tokens [58, 63] directly to fuse cross-modality features. The multi-modal encoder is typically driven by the pre-training tasks e.g. Masked Language Modeling [7] and Visual-Text Matching [46], to perform token-level cross-modal fusion. + +Video-Language pre-training for multiple downstream tasks. There are a few recent works trying to construct a unified pre-trained model for multiple downstream tasks. ActBert [63], clipBert [23], VIOLET [9] and ALPRO [24] utilize a cross-modal encoder for both retrieval and VQA tasks. They take video-text retrieval as a binary classification task and construct the classifier upon a visual-text matching head, which brings about extreme computation overload. While All-in-One [49] and HD-VILA [56] proposes a transformer-based unified backbone architecture for joint-modal representation learning, they require a large amount of data for pre-training. HERO [25] and VLM [53] design a task-agnostic model that can be finetuned on multiple downstream tasks. However, they are not end-to-end trainable and need pre-extract video features, which limits the performance of the model. + +![](images/e10e97ed4a81136d4a17e33218dc1c503bf507b171e29ed863dbd8c1937f6ff1.jpg) +Figure 1. Different attempts on achieving a universal VidL model. (a) Naive combination by stacking the cross-modal encoder upon the uni-modal encoders with weak association (COMB). (b) Our proposed Clover pre-trained model that better correlates the cross-modal encoder and the uni-modal encoders via Tri-Modal Alignment. (c) Comparisons of the transfer performance on MSRVTT [60] (zero-shot) and TGIF-FrameQA [16] datasets. Benefiting from the proposed tri-modal alignment, our proposed universal model outperforms both the individually designed models (IND) and COMB. + +![](images/ec9fc763dad1a29fcec5d1414d161e7394bd05e4dca6bd2e9ea9edf152f08099.jpg) + +![](images/9ce533bff35816aad9d5ac584995945adccd7dc2128ac5f4e5f208dc4b488567.jpg) + +Unlike the aforementioned methods, our Clover is a fully end-to-end VidL model that can be easily transferred to various downstream tasks while attaining both high efficiency and performance. + +Note that there are some methods focus on transferring the pre-trained CLIP model to downstream tasks and achieve significant performance [6, 14, 17, 21, 30, 37, 51, 57]. The CLIP model is pretrained with 400M image-text data pairs. In contrast, our Clover focuses on presenting a better video-text pretrained model, and only leverages 5M data pairs for pretraining. + +# 3. Method + +# 3.1. Motivation and Overview + +Recent years have witnessed a significant progress in Video-Language pre-training. Nevertheless, building a universal Video-Language model for solving various video understanding tasks (e.g., text-video retrieval, video question answering) is still a challenging problem. The key to the problem lies in how to unify the representation learning for cross-modal alignment and fusion efficiently. + +Cross-modal alignment aims to learn projection functions $f(\cdot)$ and $g(\cdot)$ that project videos and texts into a common embedding space where the similarity between the video and text with consistent semantic information is maximized and otherwise minimized. Taking a video $V$ as the anchor, the learning objective of cross-modal alignment is formulated as: + +$$ +\arg \max _ {f, g} [ \mathrm {s} (f (V), g (T ^ {+})) - \mathrm {s} (f (V), g (T ^ {-})) ], \tag {1} +$$ + +where $T^{+}, T^{-}$ represent texts that are semantically consistent and inconsistent with $V$ , respectively. The function $\mathrm{s}(\cdot, \cdot)$ measures the dot-product similarity between two embeddings. A model with strong cross-modal alignment capability is desired in the video-retrieval task, which requires matching the semantically aligned videos and texts. + +Cross-modal fusion aims to integrate the correlation and interaction carried by the video and text modalities into a unified multi-modal embedding. Specifically, it can be formulated as learning a function $Fusion(\cdot, \cdot)$ that takes inputs of different modals and outputs a unified representation $M = Fusion(V, T)$ , which is then used to solve downstream tasks like VQA. + +Existing pre-training strategies rely on either simple supervised or contrastive pre-text tasks that hardly achieve feature alignment and fusion simultaneously. To better correlate the cross-modal alignment and fusion, we propose the Correlated Video-Language pre-training method (Clover) with three key pre-training innovations (Sec. 3.2): tri-modal alignment; pair-wise ranking loss; and semantic enhanced masked language modeling. The classical masked language modeling task is also incorporated into our method, which is able to boost the generalizability of the model as well as the interaction between visual and language. + +Architecture. We briefly explain the architecture of Clover here, which is shown in Fig. 2 as well. It consists of three components: a video encoder, a text encoder and a multi-modal encoder. Following [9], we use VideoSwin Transformer [32] as our video encoder. Given a video $V$ , it outputs a sequence of video embeddings: $V_{e} = \{v_{1},\dots,v_{K}\} \subset \mathbb{R}^{D}$ , where $K$ is the number of flattened patches. The text encoder of Clover is a 12-layer bidirectional transformer encoder model [7]. Given an input text sentence $T$ , the encoder outputs an embedding sequence $T_{e} = \{t_{\mathrm{cls}},t_{1},\dots,t_{L - 1}\} \subset \mathbb{R}^{D}$ , where $t_\mathrm{cls}$ indicates the embedding of the [CLS] token. We employ a 3-layer bidirectional Transformer encoder to learn the fused cross-modal representation of a video-text pair. The multi-modal encoder takes the concatenation of video and text embeddings as input, and outputs the fused multi-modal embeddings $M_{e} = \{m_{v_{1}},\dots,m_{v_{K}},m_{\mathrm{CLS}},m_{t_{1}},\dots,m_{t_{L - 1}}\}$ , where $m \in \mathbb{R}^D$ . + +![](images/e65fcad67801b10efb9b33c1cd0706b0c71eb46ae63cba5b6967603291a25045.jpg) +(a) Model Architecture + +![](images/f5a4f991b0ee9d61c942d9c19a6013a766a7fcf5625ed33e2ea302d77e7375b4.jpg) +Tri-modal Alignment +Pair-wise Ranking + +![](images/b4d05872044a0c430888f7696323550bfbb071c6f1f9ad4618af462785ddb845.jpg) +(b) Pre-training Tasks +Figure 2. Overview of Clover (a) Model architecture. (b) Tri-modal alignment pre-training task with masked samples and the pair-wise ranking. “×” indicates the embedding of the token [MASK]. For more detail on model architecture and pre-training tasks refers to Sec. 3. + +![](images/78b295b1c86e26c840f9fa5f51b179df781aba52a90ad1ed12dd6dab20efa22a.jpg) + +# 3.2. Tri-Modal Alignment + +Existing VidL pre-training methods typically align the embeddings of data from different modalities through pairwise contrastive learning. Differently, we propose a TriModal Alignment (TMA) learning task that not only enforces the video and text modalities to be aligned but also encourages them to well align with the third modality, i.e., their fusion modality. In TMA, as an intermediate modality between video and text, the fused multi-modal representations act as anchors for cross-modality alignment, which reduces the difficulty of the direct alignment between video and text. Meanwhile, it keeps the fused representation closer to the uni-modal representation containing consistent semantic information and away from the others, so as to enhance the learning of semantic information in the fused representation. + +As shown in Fig. 2 (a), given an input video-text pair $\langle V, T \rangle$ and their embeddings $\langle V_e, T_e \rangle$ from the uni-modal encoder, we mask some regions in $V$ to make a masked video with embedding $V_m$ , and similarly mask some tokens in $T$ to make a masked text embedding $T_m$ . Then, we pair the incomplete samples with the complete samples as $\langle V_m, T_e \rangle$ and $\langle V_e, T_m \rangle$ . We adopt multi-modal encoder to get fused multi-modal embedding sequence $M_{V_m}$ and $M_{T_m}$ of $\langle V_m, T_e \rangle$ and $\langle V_e, T_m \rangle$ , respectively. We denote the embedding of [CLS] token in $M_{V_m}$ and $M_{T_m}$ as $M_{V_m f}$ and $M_{T_m f}$ . + +For the convenience of description, we use superscripts $i$ and $j$ to index the data sample, and introduce TMA from the perspective of video and language, respectively. Firstly, for a video representation $V_{e}^{i}$ , besides its associated text embedding $T_{e}^{i}$ and $T_{m}^{i}$ , we also consider its associated fusion embedding $M_{V_{mf}}^{i}$ as its positive pair—that should be pushed closer within the embedding space. To align these tri-modal representations, we propose a novel exclusivity loss computed within a batch of $B$ samples as follows: + +$$ +\begin{array}{l} L _ {v} = - \sum_ {i = 1} ^ {B} \left[ \log \frac {e ^ {s (V _ {e} ^ {i} , T _ {e} ^ {i}) / \tau}}{e ^ {s (V _ {e} ^ {i} , T _ {e} ^ {i}) / \tau} + Z} + \log \frac {e ^ {s (V _ {e} ^ {i} , T _ {m} ^ {i}) / \tau}}{e ^ {s (V _ {e} ^ {i} , T _ {m} ^ {i}) / \tau} + Z} \right. \\ \left. + \log \frac {e ^ {s (V _ {e} ^ {i} , M _ {V _ {m} f} ^ {i}) / \tau}}{e ^ {s (V _ {e} ^ {i} , M _ {V _ {m} f} ^ {i}) / \tau} + Z} \right], \mathrm {w h e r e} \\ \end{array} +$$ + +$$ +Z = \sum_ {j \neq i} ^ {B} \left[ e ^ {s \left(V _ {e} ^ {i}, T _ {e} ^ {j}\right) / \tau} + e ^ {s \left(V _ {e} ^ {i}, T _ {m} ^ {j}\right) / \tau} + e ^ {s \left(V _ {e} ^ {i}, M _ {V _ {m f}} ^ {j}\right) / \tau} \right]. \tag {2} +$$ + +Here $\tau$ is a temperature scalar, and $s(\cdot ,\cdot)$ is the dot-product similarity function to measure the degree of alignment between different modalities. In each term of Eq. (2), we exclude the other positive pairs when performing contrastive learning on one positive pair. Thus the intra-suppression among positive pairs is effectively avoided, and the video-/text-/fusion-modal representations are better aligned. For example, in the first term of Eq. (2), $e^{s(V_e^i,M_{Vmf}^i) / \tau}$ and $e^{s(V_e^i,T_m^i) / \tau}$ are excluded from the denominator, which prevents the model from yielding the sub-optimal solution that gets $T_{e}^{i}$ closer to $V_{e}^{i}$ while pushes $T_{m}^{i}$ and $M_{Vmf}^{i}$ away from $V_{e}^{i}$ . Meanwhile, to align the video to the text and fusion modalities, we have the following objective: + +$$ +\begin{array}{l} L _ {v ^ {\prime}} = - \sum_ {i = 1} ^ {B} \left[ \log \frac {e ^ {s (T _ {e} ^ {i} , V _ {e} ^ {i}) / \tau}}{\sum_ {j = 1} ^ {B} e ^ {s (T _ {e} ^ {i} , V _ {e} ^ {j}) / \tau}} + \log \frac {e ^ {s (T _ {m} ^ {i} , V _ {e} ^ {i}) / \tau}}{\sum_ {j = 1} ^ {B} e ^ {s (T _ {m} ^ {i} , V _ {e} ^ {j}) / \tau}} \right] \\ \left. + \log \frac {e ^ {s \left(M _ {V _ {m f}} ^ {i} , V _ {e} ^ {i}\right) / \tau}}{\sum_ {j = 1} ^ {B} e ^ {s \left(M _ {V _ {m f}} ^ {i} , V _ {e} ^ {j}\right) / \tau}} \right]. \tag {3} \\ \end{array} +$$ + +The above two losses (2) and (3) are summed up to give the tri-modal alignment objective w.r.t. the visual modality: $L_{V} = L_{v} + L_{v^{\prime}}$ . Similarly, from the perspective of text modality, we define the tri-modal alignment objective $L_{T} = L_{t} + L_{t^{\prime}}$ w.r.t. the text modality. Then, we obtain the final loss of tri-modal alignment $L_{TmA} = L_{V} + L_{T}$ . + +To get the masked samples for TMA, we conduct masking over the videos and texts to form masked samples using the following two strategies. + +Video-block masking strategy. Objects in a video often appear at similar spatial locations across consecutive frames, which compose a tube area in the spatial-temporal domain. In order to construct incomplete video samples with less information leakage, we extend the block-wise mask [3] to video via performing a block-wise mask at the same position in all video frames. In this paper, we randomly replace $20\%$ patches with a learnable mask token to obtain $V_{m}$ . + +Semantic text masking strategy. The detailed semantic information in the text basically lies in the verb, noun and adjective words. To facilitate learning their representations, we construct incomplete text samples by randomly masking some verbs, nouns and adjectives in the sentence. Specifically, given a sentence, we use a part-of-speech tagger [33] to tag each word. Then we pick the verb phrases and nouns, and replace $30\%$ of them with a special [MASK] token. Note that to avoid drastic semantic changes, we do not mask the auxiliary verb like have, should, will, would, etc. With the masking strategy, when some key elements are masked from the texts, it would form an expression carrying partial information rather than completely changing the information expressed. + +# 3.3. Training Objective + +Pair-wise ranking. In TMA, we take the masked pairs $\langle V_e, T_m \rangle$ and $\langle V_m, T_e \rangle$ as positive pairs, considering that masked samples would still carry partial meaningful information though some others are missed. For instance, as shown in Fig. 2 (a), given the text "A black swan swimming in a calm lake", "swimming" and "lake" are masked with [MASK] tokens, generating the masked pair $\langle V_e, T_m \rangle$ . Although some information is removed, the masked sentence still contains other concepts that make it partially matched with the video, such as the "black swan". Therefore, in contrastive learning, $V_m$ and $T_m$ can be considered pseudo positive candidates for $T$ and $V$ . However, compared to the original pair $\langle V_e, T_e \rangle$ , we consider that the semantic consistency in the masked pairs should be weaker. Based on this prior, we further propose a pair-wise ranking loss: + +$$ +\begin{array}{l} L _ {r a n k} = \max \left(0, - \left(\operatorname {s i m} \left(V _ {e}, T _ {e}\right) / \tau - \operatorname {s i m} \left(V _ {e}, T _ {m}\right) / \tau\right) + \lambda\right) \\ + \max \left(0, - \left(\operatorname {s i m} \left(V _ {e}, T _ {e}\right) / \tau - \operatorname {s i m} \left(V _ {m}, T _ {e}\right) / \tau\right) + \lambda\right), \tag {4} \\ \end{array} +$$ + +where $\lambda > 0$ is a margin hyper-parameter. Eq. (4) urges the model to be aware of the gap of semantic consistency between $\langle V_e, T_m \rangle / \langle V_m, T_e \rangle$ and $\langle V_e, T_e \rangle$ brought about by concepts missing in the masked pairs. With the pairwise ranking objective, our model is able to maintain fine-grained perceptual capability while improving its generalizability. + +Semantic enhanced masked language modeling. Masked + +language modeling (MLM) is a classical pre-training task in VidL, which promotes the interaction between different modalities in the cross-modal encoder. Combining the classical MLM with our semantic text masking strategy, we present the Semantic Enhanced Masked Language Modeling task, which facilitates the representation learning on the key concepts, i.e., verb, noun and adjective words. Besides, considering the class-imbalance between different words, we use focal loss [28] instead of the traditional cross-entropy loss to improve the MLM loss. We apply the MLM loss to the reconstruction of the masked text tokens i.e., $m_{t_j} \in M_{T_m}$ , where $t_j = [Mask]$ . The MLM loss is defined as: + +$$ +L _ {m l m} = - \frac {1}{B} \sum_ {i = 1} ^ {B} \sum_ {m _ {t _ {j}} \in M _ {T _ {m}} ^ {i}} [ (1 - p _ {t _ {j}} ^ {i}) ^ {\gamma} p _ {t _ {j}} ^ {i} ] \tag {5} +$$ + +where $p_{t_j}^i$ denote the predicted probability distribution of the masked token $t_j$ in $i$ th sentence in the batch, and $\gamma$ is a hyper-parameter. Finally, the overall pre-training objective of Clover is: + +$$ +L = L _ {T m A} + L _ {r a n k} + L _ {m l m}. \tag {6} +$$ + +# 4. Experiments + +# 4.1. Experiment Setup + +Pre-training datasets. Following recent work [2, 11, 24], we jointly pre-train our Clover on a video dataset WebVid2M [2] with 2.5M video-text pairs and an image dataset Google Conceptual Captions (CC3M) [44] with 3.3M image-text pairs (we only obtain 2.8M image-text pairs in CC3M due to image url broken). During pretraining, we treat image data as one frame video data. + +Downstream tasks. We evaluate our proposed models on the following downstream tasks. (a) Text-to-Video Retrieval on MSRVTT [55], LSMDC [38] and DiDeMo [1]. For DiDeMo, we follow [23, 31] and evaluate Clover on the paragraph-to-video retrieval, where text sentences for each video are concatenated together as one text query. We do not use the ground-truth proposal for fair comparison with previous works; (b) Multiple-choice QA on TGIF-Action [16], TGIF-Transition [16], MSRVTT-MC [60] and LSMDC-MC [47]; (c) Open-Ended QA on TGIF-Frame [16], MSRVTT-QA [52], MSVD-QA [52] and LSMDC-FiB [38]. For retrieval tasks, we only use the two uni-modal encoders of Clover for fine-tuning and inference. We adopt the two encoders to get the video and text embeddings, and calculate their cosine similarity for retrieval. For QA tasks, we use all the modules for fine-tuning and inference. More details on these datasets and their evaluation usage are provided in the supplementary material. + +
MethodPre-training datasetDiDeMoMSRVTTLSMDC
R@1R@5R@10MedRR@1R@5R@10MedRR@1R@5R@10MedR
Fine-tune
clipBert [23]COCO [5], VG [20]20.448.060.8622.046.859.96----
Frozen [2]W2M+C3M31.059.872.4331.059.570.5315.030.839.820
VIOLET [9]W2M+C3M+Y180M32.662.874.7-34.563.073.4-16.136.641.2-
HD-VILA [56]HDV100M28.857.469.1435.665.378.0317.434.144.115
ALPRO [24]W2M+C3M35.967.578.8333.960.773.23----
TMVM [27]W2M+C3M36.564.975.4336.264.275.7317.837.145.913.5
All-in-1 [49]W2M+H100M32.761.473.5-37.968.177.1-----
OA-Trans [50]W2M+C3M34.864.475.1335.863.476.5318.234.343.718.5
MILES [12]W2M+C3M36.663.974.0337.763.673.8317.835.644.115.5
MCQ [11]W2M+C3M37.062.273.9337.664.875.1317.935.444.515
Clover (ours)W2M+C3M50.176.785.6140.569.879.4224.844.054.58
Zero-shot
MIL-NCE [39]H100M----9.924.032.429.6----
Frozen [2]W2M+C3M21.146.056.2718.739.651.6109.322.030.151.0
VIOLET [9]W2M+C3M+Y180M23.549.859.8-25.949.559.7-----
ALPRO [24]W2M+C3M23.847.357.9624.144.755.48----
OA-Trans [50]W2M+C3M23.550.459.8623.447.555.68----
MILES [12]W2M+C3M27.250.363.6526.147.256.9711.124.730.650.7
MCQ [11]W2M+C3M25.650.661.1526.046.456.4712.225.932.242
Clover (ours)W2M+C3M29.555.266.3426.449.560.0614.729.238.224
+ +Table 1. Text-to-video retrieval performance comparison under fine-tune and zero-shot setups. Here higher R@k (Recall K) and lower MedR (Median Recall) indicate better performance. W2M, C3M, H100M, HDV100M, Y180M are short for WebVid2M [2], CC3M [44], HowTo100M [40], HD-VILA-100M [56], YT-Temporal-180M [61], respectively. + +
MethodPre-training datasetTGIFMSRVTTLSMDCMSVD
ActionTransitionFrameMCQAMCFiBQA
clipBert [23]COCO, VG82.887.860.388.237.4---
JuskAsk [58]HTVQA69M [58]----41.5--46.3
ALPRO [24]W2M+C3M----42.1--45.9
All-in-1 [49]W2M+H100M92.794.364.292.042.983.1-47.9
VIOLET [9]W2M+C3M+Y180M92.595.768.991.943.982.853.747.9
MERLOT [61]Y180M94.096.269.590.943.181.752.9-
Clover (ours)W2M+C3M95.098.271.695.244.183.754.152.4
+ +Table 2. Performance comparison on transferring to downstream video question answering tasks. + +Implementation details. Following [9], we initialize the video encoder with a VideoSwin-Base [32] model pretrained on Kinetics-400 [18]. The text encoder is initialized from pre-trained Bert-Base [7]. The multi-modal encoder is initialized from the first three layers of the pre-trained Bert-Base model. We train Clover end-to-end during both pretraining and fine-tuning. We pre-train Clover for 40 epochs, using a batch size of 1024 on 64 NVIDIA A100 GPUs. We use AdamW [34] optimizer with a weight decay 0.005 and betas (0.9, 0.98). The learning rate is first warmed-up by 4 epochs to 5e-5 and then decays following a cosine annealing decay schedule. We resize all the video frames to $224 \times 224$ and split each frame into patches with a size of $32 \times 32$ . We choose hyper-parameter $\tau = 0.05$ , $\lambda = 5$ and $\gamma = 2$ . We set dimension $D = 768$ in our model. For each video, we randomly sample 8 frames while preserving their order in-between. For fine-tuning on retrieval tasks, we only fine + +tune the uni-modal encoders of Clover with InfoNCE [48] loss. For fine-tuning on video QA task, we add a simple MLPs that takes the multi-modal [CLS] embedding as input for classification, and optimize the whole Clover model with cross-entropy loss. During fine-tuning, we follow the conventional set-up in [9, 36], and all the fine-tuning experiments are performed on 8 NVIDIA A100 GPUs. + +# 4.2. Comparing to State-of-the-art + +Text-to-video retrieval. Tab. 1 shows the retrieval results on the DiDeMo [1], MSRVTT [55] and LSMDC [38] datasets with both zero-shot and fine-tuning settings. Our Clover surpasses previous works on all the datasets by a large margin. The significant performance gain under zero-shot evaluation demonstrates its stronger generalization ability. Specifically, Clover brings $4.9\%$ gain in R@10 over MCQ [11], the current SOTA method on re + +
MethodDiDeMoMSRVTTLSMDC-MCTGIF-Frame
R@1R@5R@10MedRR@1R@5R@10MedRAccAcc
Baseline22.647.958.2719.841.751.11078.868.9
+ TMA24.749.760.0622.742.252.2980.069.3
+ TMA+SM25.349.560.5622.542.752.0980.569.4
+ TMA+SM+RankL26.451.161.3523.443.352.4980.769.7
+ +Table 3. Effects of our pre-training tasks for Clover. We report zero-shot text-video retrieval performance on DiDeMo and MSRVTT, and fine-tune video QA performance on LSMDC-MC and TGIF-Frame. TMA, SM and RankL: tri-modal alignment, semantic masking strategy and pair-wise ranking loss. + +
MethodAveraged similarity scores of positive pairsAveraged similarity margin between positive and negative samples
w/o TMA0.560.30
w TMA0.650.35
+ +Table 4. Affect of TMA on the video and text representations in learned embedding space. The experiment is conducted on MSRVTT-1kA test set under zero-shot setting. + +trieval tasks with the same pre-train data, averaged on the three datasets. Moreover, Clover outperforms VIOLET [9], though VIOLET adopts a considerably larger pre-training dataset (YTT180M [61] of 180M visual-text pairs vs. WebVid2M+CC3M of only 5.5M samples). Clover also presents superior performance when fine-tuned on the downstream datasets, compared with the SOTAs. On all three datasets, Clover surpasses previous works by a large margin across all the metrics and obtains an average improvement of $8.7\%$ on R@10. Moreover, different from some previous work [9,23,24] that adopts a joint multi-modal encoder and requires exhaustively pairing every video and text during retrieval, Clover directly deploys the video encoder and text encoder for retrieval tasks, making it much more efficient. + +Video question answering. Tab. 2 reports the results of Clover and current SOTAs on four video QA datasets. Though using much less pre-training data, Clover outperforms another unified VidL model i.e., VIOLET [9] significantly. Moreover, compared to the current SOTA method MERLOT [61], Clover brings improvements of $1.0\%$ on TGIF-Action, $2.0\%$ on TGIF-Transition, $2.1\%$ on TGIF-Frame, $4.3\%$ on MSRVTT-MC, $1.0\%$ on MSRVTT-QA, $2.0\%$ on LSMDC-MC and $1.2\%$ on LSMDC-FiB. Note that JustAsk [58] and MERLOT [61] are specifically designed for video QA and trained on orders of magnitude larger datasets (e.g., HTVQA69M [58], YTT180M [61]). The experiment results again confirm the superiority of Clover. + +# 4.3. Analysis + +We conduct ablation experiments on two retrieval datasets (DiDeMo and MSRvTT) and two Video QA datasets (TGIF-Frame and LSMDC-MC). Due to the computational resource limit, we randomly sample 1 million video-text pairs from WebVid2M [2] to build the WebVid1M subset for model pre-training under all the ablation stud + +ies. We keep the model architecture unchanged but pretrain the model with only MLM and InfoNCE losses as a baseline, which is similar to the baseline adopted in VIOLET [9]. The only difference is that we replace the ITM applied to the cross-modal encoder outputs in VIOLET [9] with InfoNCE applied to the outputs of uni-modal encoders and separate the text encoder from the cross-modal encoder. All the ablation experiments are conducted with batch size 1024 on 32 NVIDIA A100 GPUs. + +Effect of tri-modal alignment. Tri-modal alignment (TMA) aims to better correlate cross-modal alignment and fusion. To evaluate the contribution of TMA, we remove the pair-wise ranking objective and semantic masking strategy, and employ the classical MLM task as in the baseline method. Compared with the baseline, TMA explicitly associates the outputs of uni-modal encoders with the outputs of the multi-modal encoder. As shown in Tab. 3, the model trained with TMA outperforms the baseline on both retrieval and video QA, demonstrating its effectiveness. For a better understanding of the effect of TMA, we further report averaged cosine similarity of videos and texts in MSRVTT dataset (in zero-shot setting) calculated by models trained with/without TMA, respectively. For a fair comparison, we select all video queries (197 in total) that were successfully responded to by both models. We recall 100 results for each query, and take the ground-truth pairs as positive and others as negative. As shown in Tab. 4, the model armed with tri-modal alignment assigns higher similarity scores to positive pairs compared to the model without it. Meanwhile, with tri-modal alignment, the margin between the positive and negative samples is also increased. It reveals that with the help of tri-modal alignment, the distance between video and text that contain consistent semantics is reduced, while the margin between the misaligned samples is increased. The results demonstrate that with fused multi-modal representation as anchors, tri-modal alignment helps the video and language modalities to be better aligned. + +Effect of semantic masking strategy. We incorporate a semantic masking strategy to form the masked pairs in trimodal alignment task. As shown in Tab. 3, on more diverse evaluation benchmarks, e.g. DiDemo, where each video is paired with 5 different ground-truth texts, using masked samples brings more gains. The result reveals that the pro + +
MethodModel ArchitectureTraining ObjectivesMSRVTTDiDeMoTGIF-FrameLSMDC-MC
Two uni-modal encoderCross-modal encoderInfoNCEMLMTMA + RanklossR@1R@1AccAcc
IND-A✓✓20.722.9--
IND-F✓✓✓--69.079.0
COMB✓✓✓✓19.822.668.978.8
Clover✓✓✓✓23.426.469.780.7
+ +Table 5. Comparisons with task-independently trained models (IND) and naive combination of the uni- and cross- modal encoder (COMB) + +![](images/2e58ff4bf58f40c41af7f8dbfc21d7d6b1964f84c19ba8bcd3d377f29760a0ef.jpg) +Baseline: a man is showing the interior of a car +Clover: inside of a car with large space and safety + +![](images/2bfdfe5b1d7a73efa803cb07f85464665ab823cc62ecb196edb800aa697daf4f.jpg) + +![](images/202c7cf70b69e855b2960ae0660acce90836cfb7db406a4be60349be6a47a035.jpg) + +![](images/5e30e897642881d85a6c1c33a1789b9097d291c544fb3e51b74158bca0a9611e.jpg) + +![](images/99598dfc66d0e3f4188dab9891b69b52d491f464e26d08e0eeb7ac8b96890ffe.jpg) +Baseline: a man is folding pieces of paper +Clover: a person folds a paper airplane + +![](images/462c84091090bc713fa3dc7e0b3c4abbf7db80a09ffc99fb3af2f2fb488d8d6f.jpg) +Figure 3. Qualitative results of zero-shot video to text retrieval results on MSRVTT [55]. + +![](images/6668d484ba85d5bf6d7f8da3c2497dce53ecc2b883b832b200ad5ec0cb702042.jpg) + +![](images/6fa58e8590c243eba2efd5b6862e74bb238afb64237eb90ff1c090b5a8c50102.jpg) + +posed semantic masking strategy facilitates the model to capture key semantic information in more complex scenes and makes it achieve better results. + +Effect of pair-wise ranking. To make the model capture the semantic information difference between the masked pairs and complete pairs, we adopt the pair-wise rank loss. As shown in Tab. 3, with pair-wise ranking loss, the performance of the model is further improved. We also show the visualization results below to further illustrate the effect of pair-wise ranking. + +Clover makes cross-modal alignment and fusion mutually improving. We compare Clover with the ones that use the same architecture but employ different pre-training tasks in Tab. 5. $IND-A$ represents the model only trained with InfoNCE loss for cross-modal alignment. $IND-F$ indicates the model trained with MLM loss for cross-modal fusion. We also report results achieved by the model (i.e., COMB) that simply combines the uni-modal encoders and multi-modal encoder as well as the training objectives. We can see that the simple combination hurts the performance, while our Clover achieves superior performance than the task-independently trained $IND-A$ and $IND-F$ . It reveals that Clover is able to get the model's cross-modal alignment and fusion capability mutually enhanced. + +Qualitative analysis. In Fig. 3 and Fig. 4, we show the qualitative results of Clover and the baseline method on zero-shot video-text retrieval and VQA tasks. Specifically, we present the query videos with the matched texts, and the video-question pair with the answer. Clover empowers the model with stronger video-text understanding capabilities. + +![](images/f1316f9f5675cb4a33cdd88e7d78c7d679a44a60f81fc45521fde877815a75d7.jpg) +Question: What is a guy hugging? +Baseline: Animal Clover: Lion + +![](images/62686c8452653c9a3206b7e51ae02eeba485c0ad74d9cb941cf6c3a314fc6875.jpg) + +![](images/43a8c699237e35b1055cbe51ad2b450317e3bbd0048a74aa4c068f9ab63af8bd.jpg) + +![](images/21ddc29e4c885e022189e3ffe838d9ce27bf821c830a1a38f0f0e3e595ed5455.jpg) + +![](images/8c74ef8541c5fb818a88aa3822a1d8d064b92fa65ca94854e52825b0f879095b.jpg) +Question: What does a person wash with a brush? Baseline: Face Clover: Toy + +![](images/46e79f251ce445f111bc523f64810d47c0dc1708866bb8d86c25c782de0551f0.jpg) +Figure 4. Qualitative results of video question answering results on MSRVTT-QA [52]. + +![](images/f939e05940ef2722fd4631faab0281c3a85afc150b665c2e58d77303aeaa9cb1.jpg) + +![](images/dfa832d6663c0c667b2f931a3dafbe07f37026a77bf46b809ea3dc63e748e4e7.jpg) + +ity. For example, for the second video in Fig. 3, Clover returns the text with "paper airplane", which is a more accurate description of the video content; for the first video in Fig. 4, Clover generates a more accurate answer "Lion" than the baseline answer "Animal". The results demonstrate the superiority of Clover. + +# 5. Conclusion + +In this paper, we present Clover, a new end-to-end Video-Language pre-training method for both high-efficiency video-text retrieval and video question answering. Clover introduces a novel Tri-modal Alignment task to better align the representations from visual, text and fused modalities, which explicitly correlate the uni-modal encoder and multi-modal encoder. It also introduces semantic masking strategy and pair-wise ranking loss to further improve the cross-modality modal training. Extensive experiments conducted on the three retrieval datasets and eight video QA datasets clearly demonstrated, as a general video-text model, its consistent superiority for video-text understanding. + +Acknowledgements. This work was supported by National Key R&D Program of China (No.2022ZD0118201), the National Science Fund for Distinguished Young Scholars (No.62025603), the National Natural Science Foundation of China (No. U21B2037, No. U22B2051, No. 62176222, No. 62176223, No. 62176226, No. 62072386, No. 62072387, No. 62072389, No. 62002305 and No. 62272401), and the Natural Science Foundation of Fujian Province of China (No.2021J01002, No.2022J06001). + +# References + +[1] Lisa Anne Hendricks, Oliver Wang, Eli Shechtman, Josef Sivic, Trevor Darrell, and Bryan Russell. Localizing moments in video with natural language. In Proceedings of the IEEE international conference on computer vision, pages 5803-5812, 2017. 1, 5, 6 +[2] Max Bain, Arsha Nagrani, Gül Varol, and Andrew Zisserman. Frozen in time: A joint video and image encoder for end-to-end retrieval. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1728-1738, 2021. 1, 2, 5, 6, 7 +[3] Hangbo Bao, Li Dong, Songhao Piao, and Furu Wei. Beit: Bert pre-training of image transformers. In International Conference on Learning Representations, 2021. 2, 5 +[4] David Chen and William B Dolan. Collecting highly parallel data for paraphrase evaluation. In Proceedings of the 49th annual meeting of the association for computational linguistics: human language technologies, pages 190–200, 2011. 1 +[5] Xinlei Chen, Hao Fang, Tsung-Yi Lin, Ramakrishna Vedantam, Saurabh Gupta, Piotr Dólár, and C Lawrence Zitnick. Microsoft coco captions: Data collection and evaluation server. arXiv preprint arXiv:1504.00325, 2015. 6 +[6] Xing Cheng, Hezheng Lin, Xiangyu Wu, Fan Yang, and Dong Shen. Improving video-text retrieval by multi-stream corpus alignment and dual softmax loss. arXiv preprint arXiv:2109.04290, 2021. 3 +[7] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 4171–4186, 2019. 2, 3, 6 +[8] Chenyou Fan, Xiaofan Zhang, Shu Zhang, Wensheng Wang, Chi Zhang, and Heng Huang. Heterogeneous memory enhanced multimodal attention model for video question answering. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1999-2007, 2019. 2 +[9] Tsu-Jui Fu, Linjie Li, Zhe Gan, Kevin Lin, William Yang Wang, Lijuan Wang, and Zicheng Liu. Violet: End-to-end video-language transformers with masked visual-token modeling. arXiv preprint arXiv:2111.12681, 2021. 1, 2, 3, 6, 7 +[10] Valentin Gabeur, Chen Sun, Karteek Alahari, and Cordelia Schmid. Multi-modal transformer for video retrieval. In European Conference on Computer Vision, pages 214-229. Springer, 2020. 1, 2 +[11] Yuying Ge, Yixiao Ge, Xihui Liu, Dian Li, Ying Shan, Xiaohu Qie, and Ping Luo. Bridging video-text retrieval with multiple choice questions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16167-16176, 2022. 1, 2, 5, 6 +[12] Yuying Ge, Yixiao Ge, Xihui Liu, Alex Jinpeng Wang, Jianping Wu, Ying Shan, Xiaohu Qie, and Ping Luo. Miles: Visual bert pre-training with injected language semantics for + +video-text retrieval. Proceedings of the European Conference on Computer Vision (ECCV), 2022. 6 +[13] Simon Ging, Mohammadreza Zolfaghari, Hamed Pirsiavash, and Thomas Brox. Coot: Cooperative hierarchical transformer for video-text representation learning. Advances in neural information processing systems, 33:22605-22618, 2020. 1, 2 +[14] Satya Krishna Gorti, Noel Vouitsis, Junwei Ma, Keyvan Golestan, Maksims Volkovs, Animesh Garg, and Guangwei Yu. X-pool: Cross-modal language-video attention for text-video retrieval. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5006-5015, 2022. 3 +[15] Deng Huang, Peihao Chen, Runhao Zeng, Qing Du, Mingkui Tan, and Chuang Gan. Location-aware graph convolutional networks for video question answering. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 34, pages 11021-11028, 2020. 2 +[16] Yunseok Jang, Yale Song, Youngjae Yu, Youngjin Kim, and Gunhee Kim. Tgif-qa: Toward spatio-temporal reasoning in visual question answering. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2758–2766, 2017. 1, 2, 3, 5 +[17] Jie Jiang, Shaobo Min, Weijie Kong, Hongfa Wang, Zhifeng Li, and Wei Liu. Tencent text-video retrieval: Hierarchical cross-modal interactions with multi-level representations. IEEE Access, 2022. 3 +[18] Will Kay, Joao Carreira, Karen Simonyan, Brian Zhang, Chloe Hillier, Sudheendra Vijayanarasimhan, Fabio Viola, Tim Green, Trevor Back, Paul Natsev, et al. The kinetics human action video dataset. arXiv preprint arXiv:1705.06950, 2017. 6 +[19] Junyeong Kim, Minuk Ma, Kyungsu Kim, Sungjin Kim, and Chang D Yoo. Progressive attention memory network for movie story question answering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8337-8346, 2019. 2 +[20] Ranjay Krishna, Yuke Zhu, Oliver Groth, Justin Johnson, Kenji Hata, Joshua Kravitz, Stephanie Chen, Yannis Kalantidis, Li-Jia Li, David A Shamma, et al. Visual genome: Connecting language and vision using crowdsourced dense image annotations. International journal of computer vision, 123(1):32-73, 2017. 6 +[21] Alexander Kunitsyn, Maksim Kalashnikov, Maksim Dzabraev, and Andrei Ivaniuta. Mdmt-2: Multidomain multimodal transformer for video retrieval, one more step towards generalization. arXiv preprint arXiv:2203.07086, 2022. 3 +[22] Thao Minh Le, Vuong Le, Svetha Venkatesh, and Truyen Tran. Hierarchical conditional relation networks for video question answering. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9972-9981, 2020. 2 +[23] Jie Lei, Linjie Li, Luowei Zhou, Zhe Gan, Tamara L Berg, Mohit Bansal, and Jingjing Liu. Less is more: Clipbert for video-and-language learning via sparse sampling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7331-7341, 2021. 1, 2, 5, 6, 7 + +[24] Dongxu Li, Junnan Li, Hongdong Li, Juan Carlos Niebles, and Steven CH Hoi. Align and prompt: Video-and-language pre-training with entity prompts. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4953-4963, 2022. 1, 2, 5, 6, 7 +[25] Linjie Li, Yen-Chun Chen, Yu Cheng, Zhe Gan, Licheng Yu, and Jingjing Liu. Hero: Hierarchical encoder for video+ language omni-representation pre-training. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 2046-2065, 2020. 2 +[26] Xiangpeng Li, Jingkuan Song, Lianli Gao, Xianglong Liu, Wenbing Huang, Xiangnan He, and Chuang Gan. Beyond rnns: Positional self-attention with co-attention for video question answering. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 33, pages 8658–8665, 2019. 2 +[27] Chengzhi Lin, Ancong Wu, Junwei Liang, Jun Zhang, Wenhang Ge, Wei-Shi Zheng, and Chunhua Shen. Text-adaptive multiple visual prototype matching for video-text retrieval. Advances in Neural Information Processing Systems, 2022. 6 +[28] Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dólár. Focal loss for dense object detection. In Proceedings of the IEEE international conference on computer vision, pages 2980-2988, 2017. 5 +[29] Fei Liu, Jing Liu, Weining Wang, and Hanqing Lu. Hair: Hierarchical visual-semantic relational reasoning for video question answering. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1698–1707, 2021. 2 +[30] Ruyang Liu, Jingjia Huang, Ge Li, Jiashi Feng, Xinglong Wu, and Thomas H Li. Revisiting temporal modeling for clip-based image-to-video knowledge transferring. arXiv preprint arXiv:2301.11116, 2023. 3 +[31] Yang Liu, Samuel Albanie, Arsha Nagrani, and Andrew Zisserman. Use what you have: Video retrieval using representations from collaborative experts. In BMVC, 2019. 1, 2, 5 +[32] Ze Liu, Jia Ning, Yue Cao, Yixuan Wei, Zheng Zhang, Stephen Lin, and Han Hu. Video swim transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3202-3211, 2022. 3, 6 +[33] Edward Loper and Steven Bird. Nltk: The natural language toolkit. arXiv preprint cs/0205028, 2002. 5 +[34] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In International Conference on Learning Representations, 2018. 6 +[35] Huaishao Luo, Lei Ji, Botian Shi, Haoyang Huang, Nan Duan, Tianrui Li, Jason Li, Taroon Bharti, and Ming Zhou. Univl: A unified video and language pre-training model for multimodal understanding and generation. arXiv preprint arXiv:2002.06353, 2020. 2 +[36] Huaishao Luo, Lei Ji, Ming Zhong, Yang Chen, Wen Lei, Nan Duan, and Tianrui Li. Clip4clip: An empirical study of clip for end to end video clip retrieval. arXiv preprint arXiv:2104.08860, 2021. 6 + +[37] Yiwei Ma, Guohai Xu, Xiaoshuai Sun, Ming Yan, Ji Zhang, and Rongrong Ji. X-clip: End-to-end multi-grained contrastive learning for video-text retrieval. In Proceedings of the 30th ACM International Conference on Multimedia, pages 638–647, 2022. 3 +[38] Tegan Maharaj, Nicolas Ballas, Anna Rohrbach, Aaron Courville, and Christopher Pal. A dataset and exploration of models for understanding video data through fill-in-the-blank question-answering. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 6884–6893, 2017. 1, 5, 6 +[39] Antoine Miech, Jean-Baptiste Alayrac, Lucas Smaira, Ivan Laptev, Josef Sivic, and Andrew Zisserman. End-to-end learning of visual representations from uncurated instructional videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9879-9889, 2020. 1, 2, 6 +[40] Antoine Miech, Dimitri Zhukov, Jean-Baptiste Alayrac, Makarand Tapaswi, Ivan Laptev, and Josef Sivic. Howto100m: Learning a text-video embedding by watching hundred million narrated video clips. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2630-2640, 2019. 6 +[41] Jungin Park, Jiyoung Lee, and Kwanghoon Sohn. Bridge to answer: Structure-aware graph interaction network for video question answering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15526-15535, 2021. 2 +[42] Andrew Rouditchenko, Angie Boggust, David Harwath, Brian Chen, Dhiraj Joshi, Samuel Thomas, Kartik Audhkhasi, Hilde Kuehne, Rameswar Panda, Rogerio Feris, et al. Avlnet: Learning audio-visual language representations from instructional videos. In Annual Conference of the International Speech Communication Association. International Speech Communication Association, 2021. 2 +[43] Paul Hongsuck Seo, Arsha Nagrani, and Cordelia Schmid. Look before you speak: Visually contextualized utterances. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16877-16887, 2021. 2 +[44] Piyush Sharma, Nan Ding, Sebastian Goodman, and Radu Soricut. Conceptual captions: A cleaned, hypernymed, image alt-text dataset for automatic image captioning. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 2556-2565, 2018. 5, 6 +[45] Chen Sun, Austin Myers, Carl Vondrick, Kevin Murphy, and Cordelia Schmid. Videobert: A joint model for video and language representation learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7464-7473, 2019. 2 +[46] Hao Tan and Mohit Bansal. Lxmert: Learning cross-modality encoder representations from transformers. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages 5100-5111, 2019. 2 + +[47] Atousa Torabi, Niket Tandon, and Leonid Sigal. Learning language-visual embedding for movie understanding with natural-language. arXiv preprint arXiv:1609.08124, 2016. 1, 5 +[48] Aaron Van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding. arXiv e-prints, pages arXiv-1807, 2018. 2, 6 +[49] Alex Jinpeng Wang, Yixiao Ge, Rui Yan, Yuying Ge, Xudong Lin, Guanyu Cai, Jianping Wu, Ying Shan, Xiaohu Qie, and Mike Zheng Shou. All in one: Exploring unified video-language pre-training. arXiv preprint arXiv:2203.07303, 2022. 1, 2, 6 +[50] Jinpeng Wang, Yixiao Ge, Guanyu Cai, Rui Yan, Xudong Lin, Ying Shan, Xiaohu Qie, and Mike Zheng Shou. Object-aware video-language pre-training for retrieval. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3313-3322, 2022. 2, 6 +[51] Yi Wang, Kunchang Li, Yizhuo Li, Yinan He, Bingkun Huang, Zhiyu Zhao, Hongjie Zhang, Jilan Xu, Yi Liu, Zun Wang, et al. Internvideo: General video foundation models via generative and discriminative learning. arXiv preprint arXiv:2212.03191, 2022. 3 +[52] Dejing Xu, Zhou Zhao, Jun Xiao, Fei Wu, Hanwang Zhang, Xiangnan He, and Yueting Zhuang. Video question answering via gradually refined attention over appearance and motion. In Proceedings of the 25th ACM international conference on Multimedia, pages 1645-1653, 2017. 1, 2, 5, 8 +[53] Hu Xu, Gargi Ghosh, Po-Yao Huang, Prahal Arora, Masoumeh Aminzadeh, Christoph Feichtenhofer, Florian Metze, and Luke Zettlemoyer. Vlm: Task-agnostic videolanguage model pre-training for video understanding. In Findings of the Association for Computational Linguistics: ACL-IJCNLP 2021, pages 4227–4239, 2021. 2 +[54] Hu Xu, Gargi Ghosh, Po-Yao Huang, Dmytro Okhonko, Armen Aghajanyan, Florian Metze, Luke Zettlemoyer, and Christoph Feichtenhofer. Videoclip: Contrastive pre-training for zero-shot video-text understanding. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 6787-6800, 2021. 2 +[55] Jun Xu, Tao Mei, Ting Yao, and Yong Rui. Msr-vtt: A large video description dataset for bridging video and language. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5288-5296, 2016. 1, 5, 6, 8 +[56] Hongwei Xue, Tiankai Hang, Yanhong Zeng, Yuchong Sun, Bei Liu, Huan Yang, Jianlong Fu, and Baining Guo. Advancing high-resolution video-language representation with large-scale video transcriptions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5036-5045, June 2022. 1, 2, 6 +[57] Hongwei Xue, Yuchong Sun, Bei Liu, Jianlong Fu, Ruihua Song, Houqiang Li, and Jiebo Luo. Clip-vip: Adapting pretrained image-text model to video-language representation alignment. arXiv preprint arXiv:2209.06430, 2022. 3 +[58] Antoine Yang, Antoine Miech, Josef Sivic, Ivan Laptev, and Cordelia Schmid. Just ask: Learning to answer questions from millions of narrated videos. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1686-1697, 2021. 1, 2, 6, 7 + +[59] Jianwei Yang, Yonatan Bisk, and Jianfeng Gao. Taco: Token-aware cascade contrastive learning for video-text alignment. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11562-11572, 2021. 2 +[60] Youngjae Yu, Jongseok Kim, and Gunhee Kim. A joint sequence fusion model for video question answering and retrieval. In Proceedings of the European Conference on Computer Vision (ECCV), pages 471-487, 2018. 3, 5 +[61] Rowan Zellers, Ximing Lu, Jack Hessel, Youngjae Yu, Jae Sung Park, Jize Cao, Ali Farhadi, and Yejin Choi. Merlot: Multimodal neural script knowledge models. Advances in Neural Information Processing Systems, 34, 2021. 1, 6, 7 +[62] Zheng-Jun Zha, Jiawei Liu, Tianhao Yang, and Yongdong Zhang. Spatiotemporal-textual co-attention network for video question answering. ACM Transactions on Multimedia Computing, Communications, and Applications (TOMM), 15(2s):1-18, 2019. 2 +[63] Linchao Zhu and Yi Yang. Actbert: Learning global-local video-text representations. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8746-8755, 2020. 2 \ No newline at end of file diff --git a/2023/Clover_ Towards a Unified Video-Language Alignment and Fusion Model/images.zip b/2023/Clover_ Towards a Unified Video-Language Alignment and Fusion Model/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..d4043189f310730f05192778288f29368c21450f --- /dev/null +++ b/2023/Clover_ Towards a Unified Video-Language Alignment and Fusion Model/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:111e009f4c3d2a352a021a077c6fee5033bc02f2e5d0b81a7341c4a0450a9d16 +size 506199 diff --git a/2023/Clover_ Towards a Unified Video-Language Alignment and Fusion Model/layout.json b/2023/Clover_ Towards a Unified Video-Language Alignment and Fusion Model/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..51f51973f2c232766e5992ca56145de7d37a590a --- /dev/null +++ b/2023/Clover_ Towards a Unified Video-Language Alignment and Fusion Model/layout.json @@ -0,0 +1,9446 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 61, + 102, + 533, + 120 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 102, + 533, + 120 + ], + "spans": [ + { + "bbox": [ + 61, + 102, + 533, + 120 + ], + "type": "text", + "content": "Clover: Towards A Unified Video-Language Alignment and Fusion Model" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 71, + 141, + 153, + 159 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 71, + 141, + 153, + 159 + ], + "spans": [ + { + "bbox": [ + 71, + 141, + 153, + 159 + ], + "type": "text", + "content": "Jingjia Huang\\*" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 173, + 141, + 233, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 173, + 141, + 233, + 157 + ], + "spans": [ + { + "bbox": [ + 173, + 141, + 233, + 157 + ], + "type": "text", + "content": "Yinan Li\\*" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 253, + 142, + 317, + 158 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 253, + 142, + 317, + 158 + ], + "spans": [ + { + "bbox": [ + 253, + 142, + 317, + 158 + ], + "type": "text", + "content": "Jiashi Feng" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 342, + 142, + 416, + 158 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 342, + 142, + 416, + 158 + ], + "spans": [ + { + "bbox": [ + 342, + 142, + 416, + 158 + ], + "type": "text", + "content": "Xinglong Wu" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 441, + 142, + 525, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 441, + 142, + 525, + 156 + ], + "spans": [ + { + "bbox": [ + 441, + 142, + 525, + 156 + ], + "type": "text", + "content": "Xiaoshuai Sun" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 262, + 162, + 331, + 178 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 262, + 162, + 331, + 178 + ], + "spans": [ + { + "bbox": [ + 262, + 162, + 331, + 178 + ], + "type": "text", + "content": "Rongrong Ji" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 109, + 184, + 484, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 109, + 184, + 484, + 213 + ], + "spans": [ + { + "bbox": [ + 109, + 184, + 484, + 213 + ], + "type": "inline_equation", + "content": "^{\\text{心}}" + }, + { + "bbox": [ + 109, + 184, + 484, + 213 + ], + "type": "text", + "content": "Key Laboratory of Multimedia Trusted Perception and Efficient Computing, Ministry of Education of China, Xiamen University, 361005, China" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 219, + 213, + 375, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 219, + 213, + 375, + 228 + ], + "spans": [ + { + "bbox": [ + 219, + 213, + 375, + 228 + ], + "type": "text", + "content": "ByteDance Inc, 100043, China" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 167, + 228, + 426, + 242 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 167, + 228, + 426, + 242 + ], + "spans": [ + { + "bbox": [ + 167, + 228, + 426, + 242 + ], + "type": "text", + "content": "yinanlee@stu.xmu.edu.cn, {xssun, rrji} @ xmu.edu.cn" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 166, + 242, + 425, + 256 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 242, + 425, + 256 + ], + "spans": [ + { + "bbox": [ + 166, + 242, + 425, + 256 + ], + "type": "text", + "content": "{huangjingjia, jshfeng, wuxinglong} @bytedance.com" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 143, + 283, + 191, + 295 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 283, + 191, + 295 + ], + "spans": [ + { + "bbox": [ + 143, + 283, + 191, + 295 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 47, + 308, + 290, + 606 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 308, + 290, + 606 + ], + "spans": [ + { + "bbox": [ + 47, + 308, + 290, + 606 + ], + "type": "text", + "content": "Building a universal Video-Language model for solving various video understanding tasks (e.g., text-video retrieval, video question answering) is an open challenge to the machine learning field. Towards this goal, most recent works build the model by stacking uni-modal and cross-modal feature encoders and train it with pair-wise contrastive pre-text tasks. Though offering attractive generality, the resulted models have to compromise between efficiency and performance. They mostly adopt different architectures to deal with different downstream tasks. We find this is because the pair-wise training cannot well align and fuse features from different modalities. We then introduce Clover—a Correlated Video-Language pre-training method—towards a universal Video-Language model for solving multiple video understanding tasks with neither performance nor efficiency compromise. It improves cross-modal feature alignment and fusion via a novel tri-modal alignment pre-training task. Additionally, we propose to enhance the tri-modal alignment via incorporating learning from semantic masked samples and a new pair-wise ranking loss. Clover establishes new state-of-the-arts on multiple downstream tasks, including three retrieval tasks for both zero-shot and fine-tuning settings, and eight video question answering tasks. Codes and pre-trained models will be released at https://github.com/LeeYN-43/Clover." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 47, + 631, + 128, + 643 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 631, + 128, + 643 + ], + "spans": [ + { + "bbox": [ + 47, + 631, + 128, + 643 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 47, + 651, + 287, + 675 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 651, + 287, + 675 + ], + "spans": [ + { + "bbox": [ + 47, + 651, + 287, + 675 + ], + "type": "text", + "content": "Video-Language pre-training (VidL) aims to learn generalizable multi-modal models from large-scale video-text" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 284, + 545, + 368 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 284, + 545, + 368 + ], + "spans": [ + { + "bbox": [ + 306, + 284, + 545, + 368 + ], + "type": "text", + "content": "samples so as to better solve various challenging Video-Language understanding tasks, such as text-video retrieval [1,4,38,55] and video question answering [16,47,52]. Recent studies [9,11,23,24,49,56,58,61] have shown that VidL leads to significant performance improvement and achieves state-of-the-art results on various downstream text-video retrieval and video question answering (VQA) benchmarks." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 373, + 546, + 480 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 373, + 546, + 480 + ], + "spans": [ + { + "bbox": [ + 304, + 373, + 546, + 480 + ], + "type": "text", + "content": "Though achieving encouraging performance, existing VidL models mostly adopt different architectures to deal with different downstream tasks. For the text-video retrieval tasks, they [2, 10, 11, 13, 31, 39] typically use two individual uni-modal encoders for processing video and text data separately, for the sake of retrieval efficiency. While for video question answering tasks, the models usually adopt the multi-modal joint encoder design to learn the association and interaction of different modalities." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 486, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 486, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 486, + 547, + 714 + ], + "type": "text", + "content": "Building a unified model capable of solving various Video-Language tasks is a long-standing challenge for machine learning research. A few recent works [9,24] attempt to learn a unified VidL model for both tasks, which uses the multi-modal encoder to conduct text-video retrieval. However, the model requires an exhaustive pair-wise comparison between the query texts and gallery videos. Given " + }, + { + "bbox": [ + 304, + 486, + 547, + 714 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 486, + 547, + 714 + ], + "type": "text", + "content": " text queries and " + }, + { + "bbox": [ + 304, + 486, + 547, + 714 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 304, + 486, + 547, + 714 + ], + "type": "text", + "content": " category videos, the computation complexity of the multi-modal encoder model would be " + }, + { + "bbox": [ + 304, + 486, + 547, + 714 + ], + "type": "inline_equation", + "content": "O(NM)" + }, + { + "bbox": [ + 304, + 486, + 547, + 714 + ], + "type": "text", + "content": ", which makes it infeasible for large-scale videotext retrieval applications. Another straightforward solution is to simply combine the uni-modal and multi-modal encoders (Fig. 1 (a)), and perform the retrieval and VQA tasks through the uni-modal and multi-modal encoders respectively. Its computation complexity for retrieval tasks is only " + }, + { + "bbox": [ + 304, + 486, + 547, + 714 + ], + "type": "inline_equation", + "content": "O(N + M)" + }, + { + "bbox": [ + 304, + 486, + 547, + 714 + ], + "type": "text", + "content": ". However, the experiment results in Fig.1 (c) show that, without a carefully designed correlating mechanism between these two types of encoders, the simple combination i.e., COMB yields a compromised per" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 57, + 683, + 195, + 693 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 683, + 195, + 693 + ], + "spans": [ + { + "bbox": [ + 57, + 683, + 195, + 693 + ], + "type": "text", + "content": "*Equal Contribution in alphabetical order." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 59, + 693, + 209, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 693, + 209, + 703 + ], + "spans": [ + { + "bbox": [ + 59, + 693, + 209, + 703 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 59, + 693, + 209, + 703 + ], + "type": "text", + "content": " Work done when interning at ByteDance Inc." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 59, + 703, + 138, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 703, + 138, + 712 + ], + "spans": [ + { + "bbox": [ + 59, + 703, + 138, + 712 + ], + "type": "inline_equation", + "content": "\\ddagger" + }, + { + "bbox": [ + 59, + 703, + 138, + 712 + ], + "type": "text", + "content": " Corresponding Author." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "14856" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 287, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 287, + 95 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 287, + 95 + ], + "type": "text", + "content": "formance compared with the models i.e. IND individually designed for retrieval and VQA." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 98, + 289, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 98, + 289, + 518 + ], + "spans": [ + { + "bbox": [ + 47, + 98, + 289, + 518 + ], + "type": "text", + "content": "In this work, we aim to address the above issues and build a unified pre-training model that attains high efficiency and performance simultaneously. We observe: (i) well aligning the features of the video and text from the same data pair is important for text-video matching; (ii) effectively fusing video and text features into unified representations is critical for video-text understanding. However, existing pre-training strategies that rely on either simple supervised or contrastive pre-text tasks hardly achieve promising feature alignment and fusion capability simultaneously. Motivated by these observations, we develop a new VidL method from these two aspects. Specifically, we propose Correlated Video-Language pre-training (Clover), a VidL method that not only unifies Video-Language alignment and fusion, but also makes them mutually boosted. The fused multi-modal representation contains richer context information than the uni-modal representations [11,35]. As an intermediate modality between video and text, the multi-modal representations are good anchors for cross-modality alignment. Meanwhile, keeping the fused representation closer to the uni-modal representation containing consistent semantic information and away from the inconsistent one will enhance the learning of semantic information in the fused modality. Therefore, we propose the Tri-Modal Alignment (TMA) to get Video-Language alignment and fusion mutually boosted, which takes the alignment between the multimodal representation and text/video representations as an auxiliary objective. We note that since the tri-modal alignment is well compatible with the classical pre-training tasks [3,7,35] e.g., Masked Language Modeling, its computation overhead is negligible. To help the model maintain fine-grained discriminative capability while improving its generalizability, we further introduce a pair-wise ranking loss that urges the model to be aware of the concept missing in masked samples compared to original samples." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 520, + 287, + 663 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 520, + 287, + 663 + ], + "spans": [ + { + "bbox": [ + 47, + 520, + 287, + 663 + ], + "type": "text", + "content": "Extensive experiments are conducted on multiple downstream tasks, including three retrieval tasks with different experimental setups (i.e. zero-shot and fine-tune) and eight video question answering tasks. The results demonstrate that Clover is able to get the cross-modal fusion and alignment capability mutually improved, and consistently outperforms current SOTAs on various downstream tasks. It achieves an average performance improvement of " + }, + { + "bbox": [ + 47, + 520, + 287, + 663 + ], + "type": "inline_equation", + "content": "4.9\\%" + }, + { + "bbox": [ + 47, + 520, + 287, + 663 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 520, + 287, + 663 + ], + "type": "inline_equation", + "content": "8.7\\%" + }, + { + "bbox": [ + 47, + 520, + 287, + 663 + ], + "type": "text", + "content": " Recall@10 score on the zero-shot and fine-tune settings of the three downstream retrieval datasets, while the average accuracy improvement over current SOTAs is " + }, + { + "bbox": [ + 47, + 520, + 287, + 663 + ], + "type": "inline_equation", + "content": "2.3\\%" + }, + { + "bbox": [ + 47, + 520, + 287, + 663 + ], + "type": "text", + "content": " on the eight video question answering datasets." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 665, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 665, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 665, + 287, + 713 + ], + "type": "text", + "content": "In summary, we make the following contributions: (1) we introduce Clover, a pre-training method achieving unified Video-Language alignment and fusion model that can be easily transferred to various downstream video under" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 305, + 72, + 545, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 545, + 120 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 545, + 120 + ], + "type": "text", + "content": "standing tasks while attaining both high efficiency and performance; (2) we propose a novel tri-modal alignment pretraining task, which correlates the uni-modal encoder and multi-modal encoder to get them mutually boosted." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 306, + 131, + 392, + 144 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 131, + 392, + 144 + ], + "spans": [ + { + "bbox": [ + 306, + 131, + 392, + 144 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 305, + 151, + 545, + 342 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 151, + 545, + 342 + ], + "spans": [ + { + "bbox": [ + 305, + 151, + 545, + 342 + ], + "type": "text", + "content": "Video-Language pre-training for text-video retrieval. Existing pre-training methods for text-video retrieval can be mainly divided into two categories. The first kind of method employs two individual encoders to embed video and text, and project them into a common latent space [2,10,11,13,31,39,42,50,54,59]. Then, they leverage a contrastive objective [48] to align the paired representations. Thanks to their high efficiency, this category of methods is widely used in real-world applications. The other kind of method utilizes a joint multi-modal encoder to learn comprehensive representation for a given video-text pair, and predict whether the video and text are matched or not via a binary classifier [9,23,25,45,53,63]. Despite the good performance they achieved, they need to exhaustively pair all the videos and texts, and feed them into the model during inference, making them highly inefficient." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 343, + 545, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 343, + 545, + 521 + ], + "spans": [ + { + "bbox": [ + 305, + 343, + 545, + 521 + ], + "type": "text", + "content": "Video-Language pre-training for video question answering. Video question answering aims to automatically answer natural language questions given a context video, which requires a joint understanding of both the video and language. Previous works mainly focus on designing better spatio-temporal attention networks [8,16,26,29,52,62] and question-video relation networks [15, 19, 22, 41] to extract more accurate multi-modal representations. Recently, many works leverage transformer-based models to build an additional multi-modal encoder upon the uni-modal encoders [9, 24, 43] or input image patches and word tokens [58, 63] directly to fuse cross-modality features. The multi-modal encoder is typically driven by the pre-training tasks e.g. Masked Language Modeling [7] and Visual-Text Matching [46], to perform token-level cross-modal fusion." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 522, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 522, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 522, + 545, + 713 + ], + "type": "text", + "content": "Video-Language pre-training for multiple downstream tasks. There are a few recent works trying to construct a unified pre-trained model for multiple downstream tasks. ActBert [63], clipBert [23], VIOLET [9] and ALPRO [24] utilize a cross-modal encoder for both retrieval and VQA tasks. They take video-text retrieval as a binary classification task and construct the classifier upon a visual-text matching head, which brings about extreme computation overload. While All-in-One [49] and HD-VILA [56] proposes a transformer-based unified backbone architecture for joint-modal representation learning, they require a large amount of data for pre-training. HERO [25] and VLM [53] design a task-agnostic model that can be finetuned on multiple downstream tasks. However, they are not end-to-end trainable and need pre-extract video features, which limits the performance of the model." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14857" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 82, + 35, + 202, + 178 + ], + "blocks": [ + { + "bbox": [ + 82, + 35, + 202, + 178 + ], + "lines": [ + { + "bbox": [ + 82, + 35, + 202, + 178 + ], + "spans": [ + { + "bbox": [ + 82, + 35, + 202, + 178 + ], + "type": "image", + "image_path": "e10e97ed4a81136d4a17e33218dc1c503bf507b171e29ed863dbd8c1937f6ff1.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 189, + 547, + 245 + ], + "lines": [ + { + "bbox": [ + 46, + 189, + 547, + 245 + ], + "spans": [ + { + "bbox": [ + 46, + 189, + 547, + 245 + ], + "type": "text", + "content": "Figure 1. Different attempts on achieving a universal VidL model. (a) Naive combination by stacking the cross-modal encoder upon the uni-modal encoders with weak association (COMB). (b) Our proposed Clover pre-trained model that better correlates the cross-modal encoder and the uni-modal encoders via Tri-Modal Alignment. (c) Comparisons of the transfer performance on MSRVTT [60] (zero-shot) and TGIF-FrameQA [16] datasets. Benefiting from the proposed tri-modal alignment, our proposed universal model outperforms both the individually designed models (IND) and COMB." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 212, + 35, + 330, + 176 + ], + "blocks": [ + { + "bbox": [ + 212, + 35, + 330, + 176 + ], + "lines": [ + { + "bbox": [ + 212, + 35, + 330, + 176 + ], + "spans": [ + { + "bbox": [ + 212, + 35, + 330, + 176 + ], + "type": "image", + "image_path": "ec9fc763dad1a29fcec5d1414d161e7394bd05e4dca6bd2e9ea9edf152f08099.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 331, + 52, + 512, + 176 + ], + "blocks": [ + { + "bbox": [ + 331, + 52, + 512, + 176 + ], + "lines": [ + { + "bbox": [ + 331, + 52, + 512, + 176 + ], + "spans": [ + { + "bbox": [ + 331, + 52, + 512, + 176 + ], + "type": "image", + "image_path": "9ce533bff35816aad9d5ac584995945adccd7dc2128ac5f4e5f208dc4b488567.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 253, + 287, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 253, + 287, + 300 + ], + "spans": [ + { + "bbox": [ + 46, + 253, + 287, + 300 + ], + "type": "text", + "content": "Unlike the aforementioned methods, our Clover is a fully end-to-end VidL model that can be easily transferred to various downstream tasks while attaining both high efficiency and performance." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 301, + 287, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 301, + 287, + 384 + ], + "spans": [ + { + "bbox": [ + 46, + 301, + 287, + 384 + ], + "type": "text", + "content": "Note that there are some methods focus on transferring the pre-trained CLIP model to downstream tasks and achieve significant performance [6, 14, 17, 21, 30, 37, 51, 57]. The CLIP model is pretrained with 400M image-text data pairs. In contrast, our Clover focuses on presenting a better video-text pretrained model, and only leverages 5M data pairs for pretraining." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 396, + 104, + 409 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 396, + 104, + 409 + ], + "spans": [ + { + "bbox": [ + 47, + 396, + 104, + 409 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 416, + 189, + 428 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 416, + 189, + 428 + ], + "spans": [ + { + "bbox": [ + 47, + 416, + 189, + 428 + ], + "type": "text", + "content": "3.1. Motivation and Overview" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 435, + 287, + 519 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 435, + 287, + 519 + ], + "spans": [ + { + "bbox": [ + 46, + 435, + 287, + 519 + ], + "type": "text", + "content": "Recent years have witnessed a significant progress in Video-Language pre-training. Nevertheless, building a universal Video-Language model for solving various video understanding tasks (e.g., text-video retrieval, video question answering) is still a challenging problem. The key to the problem lies in how to unify the representation learning for cross-modal alignment and fusion efficiently." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 519, + 287, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 519, + 287, + 601 + ], + "spans": [ + { + "bbox": [ + 46, + 519, + 287, + 601 + ], + "type": "text", + "content": "Cross-modal alignment aims to learn projection functions " + }, + { + "bbox": [ + 46, + 519, + 287, + 601 + ], + "type": "inline_equation", + "content": "f(\\cdot)" + }, + { + "bbox": [ + 46, + 519, + 287, + 601 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 519, + 287, + 601 + ], + "type": "inline_equation", + "content": "g(\\cdot)" + }, + { + "bbox": [ + 46, + 519, + 287, + 601 + ], + "type": "text", + "content": " that project videos and texts into a common embedding space where the similarity between the video and text with consistent semantic information is maximized and otherwise minimized. Taking a video " + }, + { + "bbox": [ + 46, + 519, + 287, + 601 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 46, + 519, + 287, + 601 + ], + "type": "text", + "content": " as the anchor, the learning objective of cross-modal alignment is formulated as:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 61, + 611, + 287, + 632 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 611, + 287, + 632 + ], + "spans": [ + { + "bbox": [ + 61, + 611, + 287, + 632 + ], + "type": "interline_equation", + "content": "\\arg \\max _ {f, g} [ \\mathrm {s} (f (V), g (T ^ {+})) - \\mathrm {s} (f (V), g (T ^ {-})) ], \\tag {1}", + "image_path": "12a28f313dd26bdfd4949df9be66d465c4a29b8a96a6eb14732dceed5de3e0d3.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 641, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 641, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 641, + 287, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 641, + 287, + 714 + ], + "type": "inline_equation", + "content": "T^{+}, T^{-}" + }, + { + "bbox": [ + 46, + 641, + 287, + 714 + ], + "type": "text", + "content": " represent texts that are semantically consistent and inconsistent with " + }, + { + "bbox": [ + 46, + 641, + 287, + 714 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 46, + 641, + 287, + 714 + ], + "type": "text", + "content": ", respectively. The function " + }, + { + "bbox": [ + 46, + 641, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\mathrm{s}(\\cdot, \\cdot)" + }, + { + "bbox": [ + 46, + 641, + 287, + 714 + ], + "type": "text", + "content": " measures the dot-product similarity between two embeddings. A model with strong cross-modal alignment capability is desired in the video-retrieval task, which requires matching the semantically aligned videos and texts." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 253, + 547, + 338 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 253, + 547, + 338 + ], + "spans": [ + { + "bbox": [ + 305, + 253, + 547, + 338 + ], + "type": "text", + "content": "Cross-modal fusion aims to integrate the correlation and interaction carried by the video and text modalities into a unified multi-modal embedding. Specifically, it can be formulated as learning a function " + }, + { + "bbox": [ + 305, + 253, + 547, + 338 + ], + "type": "inline_equation", + "content": "Fusion(\\cdot, \\cdot)" + }, + { + "bbox": [ + 305, + 253, + 547, + 338 + ], + "type": "text", + "content": " that takes inputs of different modals and outputs a unified representation " + }, + { + "bbox": [ + 305, + 253, + 547, + 338 + ], + "type": "inline_equation", + "content": "M = Fusion(V, T)" + }, + { + "bbox": [ + 305, + 253, + 547, + 338 + ], + "type": "text", + "content": ", which is then used to solve downstream tasks like VQA." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 352, + 547, + 484 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 352, + 547, + 484 + ], + "spans": [ + { + "bbox": [ + 304, + 352, + 547, + 484 + ], + "type": "text", + "content": "Existing pre-training strategies rely on either simple supervised or contrastive pre-text tasks that hardly achieve feature alignment and fusion simultaneously. To better correlate the cross-modal alignment and fusion, we propose the Correlated Video-Language pre-training method (Clover) with three key pre-training innovations (Sec. 3.2): tri-modal alignment; pair-wise ranking loss; and semantic enhanced masked language modeling. The classical masked language modeling task is also incorporated into our method, which is able to boost the generalizability of the model as well as the interaction between visual and language." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 498, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 498, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 498, + 547, + 713 + ], + "type": "text", + "content": "Architecture. We briefly explain the architecture of Clover here, which is shown in Fig. 2 as well. It consists of three components: a video encoder, a text encoder and a multi-modal encoder. Following [9], we use VideoSwin Transformer [32] as our video encoder. Given a video " + }, + { + "bbox": [ + 304, + 498, + 547, + 713 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 304, + 498, + 547, + 713 + ], + "type": "text", + "content": ", it outputs a sequence of video embeddings: " + }, + { + "bbox": [ + 304, + 498, + 547, + 713 + ], + "type": "inline_equation", + "content": "V_{e} = \\{v_{1},\\dots,v_{K}\\} \\subset \\mathbb{R}^{D}" + }, + { + "bbox": [ + 304, + 498, + 547, + 713 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 498, + 547, + 713 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 304, + 498, + 547, + 713 + ], + "type": "text", + "content": " is the number of flattened patches. The text encoder of Clover is a 12-layer bidirectional transformer encoder model [7]. Given an input text sentence " + }, + { + "bbox": [ + 304, + 498, + 547, + 713 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 304, + 498, + 547, + 713 + ], + "type": "text", + "content": ", the encoder outputs an embedding sequence " + }, + { + "bbox": [ + 304, + 498, + 547, + 713 + ], + "type": "inline_equation", + "content": "T_{e} = \\{t_{\\mathrm{cls}},t_{1},\\dots,t_{L - 1}\\} \\subset \\mathbb{R}^{D}" + }, + { + "bbox": [ + 304, + 498, + 547, + 713 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 498, + 547, + 713 + ], + "type": "inline_equation", + "content": "t_\\mathrm{cls}" + }, + { + "bbox": [ + 304, + 498, + 547, + 713 + ], + "type": "text", + "content": " indicates the embedding of the [CLS] token. We employ a 3-layer bidirectional Transformer encoder to learn the fused cross-modal representation of a video-text pair. The multi-modal encoder takes the concatenation of video and text embeddings as input, and outputs the fused multi-modal embeddings " + }, + { + "bbox": [ + 304, + 498, + 547, + 713 + ], + "type": "inline_equation", + "content": "M_{e} = \\{m_{v_{1}},\\dots,m_{v_{K}},m_{\\mathrm{CLS}},m_{t_{1}},\\dots,m_{t_{L - 1}}\\}" + }, + { + "bbox": [ + 304, + 498, + 547, + 713 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 498, + 547, + 713 + ], + "type": "inline_equation", + "content": "m \\in \\mathbb{R}^D" + }, + { + "bbox": [ + 304, + 498, + 547, + 713 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "14858" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 85, + 30, + 324, + 214 + ], + "blocks": [ + { + "bbox": [ + 85, + 30, + 324, + 214 + ], + "lines": [ + { + "bbox": [ + 85, + 30, + 324, + 214 + ], + "spans": [ + { + "bbox": [ + 85, + 30, + 324, + 214 + ], + "type": "image", + "image_path": "e65fcad67801b10efb9b33c1cd0706b0c71eb46ae63cba5b6967603291a25045.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 166, + 222, + 238, + 232 + ], + "lines": [ + { + "bbox": [ + 166, + 222, + 238, + 232 + ], + "spans": [ + { + "bbox": [ + 166, + 222, + 238, + 232 + ], + "type": "text", + "content": "(a) Model Architecture" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 337, + 43, + 496, + 134 + ], + "blocks": [ + { + "bbox": [ + 337, + 31, + 403, + 41 + ], + "lines": [ + { + "bbox": [ + 337, + 31, + 403, + 41 + ], + "spans": [ + { + "bbox": [ + 337, + 31, + 403, + 41 + ], + "type": "text", + "content": "Tri-modal Alignment" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 337, + 43, + 496, + 134 + ], + "lines": [ + { + "bbox": [ + 337, + 43, + 496, + 134 + ], + "spans": [ + { + "bbox": [ + 337, + 43, + 496, + 134 + ], + "type": "image", + "image_path": "f5a4f991b0ee9d61c942d9c19a6013a766a7fcf5625ed33e2ea302d77e7375b4.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 342, + 137, + 401, + 148 + ], + "lines": [ + { + "bbox": [ + 342, + 137, + 401, + 148 + ], + "spans": [ + { + "bbox": [ + 342, + 137, + 401, + 148 + ], + "type": "text", + "content": "Pair-wise Ranking" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 340, + 158, + 402, + 217 + ], + "blocks": [ + { + "bbox": [ + 340, + 158, + 402, + 217 + ], + "lines": [ + { + "bbox": [ + 340, + 158, + 402, + 217 + ], + "spans": [ + { + "bbox": [ + 340, + 158, + 402, + 217 + ], + "type": "image", + "image_path": "b4d05872044a0c430888f7696323550bfbb071c6f1f9ad4618af462785ddb845.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 381, + 224, + 451, + 234 + ], + "lines": [ + { + "bbox": [ + 381, + 224, + 451, + 234 + ], + "spans": [ + { + "bbox": [ + 381, + 224, + 451, + 234 + ], + "type": "text", + "content": "(b) Pre-training Tasks" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 246, + 547, + 269 + ], + "lines": [ + { + "bbox": [ + 46, + 246, + 547, + 269 + ], + "spans": [ + { + "bbox": [ + 46, + 246, + 547, + 269 + ], + "type": "text", + "content": "Figure 2. Overview of Clover (a) Model architecture. (b) Tri-modal alignment pre-training task with masked samples and the pair-wise ranking. “×” indicates the embedding of the token [MASK]. For more detail on model architecture and pre-training tasks refers to Sec. 3." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 422, + 158, + 485, + 217 + ], + "blocks": [ + { + "bbox": [ + 422, + 158, + 485, + 217 + ], + "lines": [ + { + "bbox": [ + 422, + 158, + 485, + 217 + ], + "spans": [ + { + "bbox": [ + 422, + 158, + 485, + 217 + ], + "type": "image", + "image_path": "78b295b1c86e26c840f9fa5f51b179df781aba52a90ad1ed12dd6dab20efa22a.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 277, + 170, + 290 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 277, + 170, + 290 + ], + "spans": [ + { + "bbox": [ + 47, + 277, + 170, + 290 + ], + "type": "text", + "content": "3.2. Tri-Modal Alignment" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 298, + 287, + 477 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 298, + 287, + 477 + ], + "spans": [ + { + "bbox": [ + 46, + 298, + 287, + 477 + ], + "type": "text", + "content": "Existing VidL pre-training methods typically align the embeddings of data from different modalities through pairwise contrastive learning. Differently, we propose a TriModal Alignment (TMA) learning task that not only enforces the video and text modalities to be aligned but also encourages them to well align with the third modality, i.e., their fusion modality. In TMA, as an intermediate modality between video and text, the fused multi-modal representations act as anchors for cross-modality alignment, which reduces the difficulty of the direct alignment between video and text. Meanwhile, it keeps the fused representation closer to the uni-modal representation containing consistent semantic information and away from the others, so as to enhance the learning of semantic information in the fused representation." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 480, + 287, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 480, + 287, + 613 + ], + "spans": [ + { + "bbox": [ + 47, + 480, + 287, + 613 + ], + "type": "text", + "content": "As shown in Fig. 2 (a), given an input video-text pair " + }, + { + "bbox": [ + 47, + 480, + 287, + 613 + ], + "type": "inline_equation", + "content": "\\langle V, T \\rangle" + }, + { + "bbox": [ + 47, + 480, + 287, + 613 + ], + "type": "text", + "content": " and their embeddings " + }, + { + "bbox": [ + 47, + 480, + 287, + 613 + ], + "type": "inline_equation", + "content": "\\langle V_e, T_e \\rangle" + }, + { + "bbox": [ + 47, + 480, + 287, + 613 + ], + "type": "text", + "content": " from the uni-modal encoder, we mask some regions in " + }, + { + "bbox": [ + 47, + 480, + 287, + 613 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 47, + 480, + 287, + 613 + ], + "type": "text", + "content": " to make a masked video with embedding " + }, + { + "bbox": [ + 47, + 480, + 287, + 613 + ], + "type": "inline_equation", + "content": "V_m" + }, + { + "bbox": [ + 47, + 480, + 287, + 613 + ], + "type": "text", + "content": ", and similarly mask some tokens in " + }, + { + "bbox": [ + 47, + 480, + 287, + 613 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 47, + 480, + 287, + 613 + ], + "type": "text", + "content": " to make a masked text embedding " + }, + { + "bbox": [ + 47, + 480, + 287, + 613 + ], + "type": "inline_equation", + "content": "T_m" + }, + { + "bbox": [ + 47, + 480, + 287, + 613 + ], + "type": "text", + "content": ". Then, we pair the incomplete samples with the complete samples as " + }, + { + "bbox": [ + 47, + 480, + 287, + 613 + ], + "type": "inline_equation", + "content": "\\langle V_m, T_e \\rangle" + }, + { + "bbox": [ + 47, + 480, + 287, + 613 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 480, + 287, + 613 + ], + "type": "inline_equation", + "content": "\\langle V_e, T_m \\rangle" + }, + { + "bbox": [ + 47, + 480, + 287, + 613 + ], + "type": "text", + "content": ". We adopt multi-modal encoder to get fused multi-modal embedding sequence " + }, + { + "bbox": [ + 47, + 480, + 287, + 613 + ], + "type": "inline_equation", + "content": "M_{V_m}" + }, + { + "bbox": [ + 47, + 480, + 287, + 613 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 480, + 287, + 613 + ], + "type": "inline_equation", + "content": "M_{T_m}" + }, + { + "bbox": [ + 47, + 480, + 287, + 613 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 47, + 480, + 287, + 613 + ], + "type": "inline_equation", + "content": "\\langle V_m, T_e \\rangle" + }, + { + "bbox": [ + 47, + 480, + 287, + 613 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 480, + 287, + 613 + ], + "type": "inline_equation", + "content": "\\langle V_e, T_m \\rangle" + }, + { + "bbox": [ + 47, + 480, + 287, + 613 + ], + "type": "text", + "content": ", respectively. We denote the embedding of [CLS] token in " + }, + { + "bbox": [ + 47, + 480, + 287, + 613 + ], + "type": "inline_equation", + "content": "M_{V_m}" + }, + { + "bbox": [ + 47, + 480, + 287, + 613 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 480, + 287, + 613 + ], + "type": "inline_equation", + "content": "M_{T_m}" + }, + { + "bbox": [ + 47, + 480, + 287, + 613 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 47, + 480, + 287, + 613 + ], + "type": "inline_equation", + "content": "M_{V_m f}" + }, + { + "bbox": [ + 47, + 480, + 287, + 613 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 480, + 287, + 613 + ], + "type": "inline_equation", + "content": "M_{T_m f}" + }, + { + "bbox": [ + 47, + 480, + 287, + 613 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 613, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 613, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 613, + 288, + 714 + ], + "type": "text", + "content": "For the convenience of description, we use superscripts " + }, + { + "bbox": [ + 46, + 613, + 288, + 714 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 613, + 288, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 613, + 288, + 714 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 46, + 613, + 288, + 714 + ], + "type": "text", + "content": " to index the data sample, and introduce TMA from the perspective of video and language, respectively. Firstly, for a video representation " + }, + { + "bbox": [ + 46, + 613, + 288, + 714 + ], + "type": "inline_equation", + "content": "V_{e}^{i}" + }, + { + "bbox": [ + 46, + 613, + 288, + 714 + ], + "type": "text", + "content": ", besides its associated text embedding " + }, + { + "bbox": [ + 46, + 613, + 288, + 714 + ], + "type": "inline_equation", + "content": "T_{e}^{i}" + }, + { + "bbox": [ + 46, + 613, + 288, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 613, + 288, + 714 + ], + "type": "inline_equation", + "content": "T_{m}^{i}" + }, + { + "bbox": [ + 46, + 613, + 288, + 714 + ], + "type": "text", + "content": ", we also consider its associated fusion embedding " + }, + { + "bbox": [ + 46, + 613, + 288, + 714 + ], + "type": "inline_equation", + "content": "M_{V_{mf}}^{i}" + }, + { + "bbox": [ + 46, + 613, + 288, + 714 + ], + "type": "text", + "content": " as its positive pair—that should be pushed closer within the embedding space. To align these tri-modal representations, we propose a novel exclusivity loss computed within a batch of " + }, + { + "bbox": [ + 46, + 613, + 288, + 714 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 46, + 613, + 288, + 714 + ], + "type": "text", + "content": " samples as follows:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 315, + 292, + 535, + 354 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 292, + 535, + 354 + ], + "spans": [ + { + "bbox": [ + 315, + 292, + 535, + 354 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} L _ {v} = - \\sum_ {i = 1} ^ {B} \\left[ \\log \\frac {e ^ {s (V _ {e} ^ {i} , T _ {e} ^ {i}) / \\tau}}{e ^ {s (V _ {e} ^ {i} , T _ {e} ^ {i}) / \\tau} + Z} + \\log \\frac {e ^ {s (V _ {e} ^ {i} , T _ {m} ^ {i}) / \\tau}}{e ^ {s (V _ {e} ^ {i} , T _ {m} ^ {i}) / \\tau} + Z} \\right. \\\\ \\left. + \\log \\frac {e ^ {s (V _ {e} ^ {i} , M _ {V _ {m} f} ^ {i}) / \\tau}}{e ^ {s (V _ {e} ^ {i} , M _ {V _ {m} f} ^ {i}) / \\tau} + Z} \\right], \\mathrm {w h e r e} \\\\ \\end{array}", + "image_path": "2c0ac14bcdfe12f27e21f735cd9e90f7214c41c47591071e05c2376b81953e41.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 318, + 355, + 545, + 394 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 355, + 545, + 394 + ], + "spans": [ + { + "bbox": [ + 318, + 355, + 545, + 394 + ], + "type": "interline_equation", + "content": "Z = \\sum_ {j \\neq i} ^ {B} \\left[ e ^ {s \\left(V _ {e} ^ {i}, T _ {e} ^ {j}\\right) / \\tau} + e ^ {s \\left(V _ {e} ^ {i}, T _ {m} ^ {j}\\right) / \\tau} + e ^ {s \\left(V _ {e} ^ {i}, M _ {V _ {m f}} ^ {j}\\right) / \\tau} \\right]. \\tag {2}", + "image_path": "2018e6cca81fef4e1a41af971e550fdfd18158c04250e93aa74119fbbfdd29ac.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 395, + 546, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 395, + 546, + 545 + ], + "spans": [ + { + "bbox": [ + 304, + 395, + 546, + 545 + ], + "type": "text", + "content": "Here " + }, + { + "bbox": [ + 304, + 395, + 546, + 545 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 304, + 395, + 546, + 545 + ], + "type": "text", + "content": " is a temperature scalar, and " + }, + { + "bbox": [ + 304, + 395, + 546, + 545 + ], + "type": "inline_equation", + "content": "s(\\cdot ,\\cdot)" + }, + { + "bbox": [ + 304, + 395, + 546, + 545 + ], + "type": "text", + "content": " is the dot-product similarity function to measure the degree of alignment between different modalities. In each term of Eq. (2), we exclude the other positive pairs when performing contrastive learning on one positive pair. Thus the intra-suppression among positive pairs is effectively avoided, and the video-/text-/fusion-modal representations are better aligned. For example, in the first term of Eq. (2), " + }, + { + "bbox": [ + 304, + 395, + 546, + 545 + ], + "type": "inline_equation", + "content": "e^{s(V_e^i,M_{Vmf}^i) / \\tau}" + }, + { + "bbox": [ + 304, + 395, + 546, + 545 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 395, + 546, + 545 + ], + "type": "inline_equation", + "content": "e^{s(V_e^i,T_m^i) / \\tau}" + }, + { + "bbox": [ + 304, + 395, + 546, + 545 + ], + "type": "text", + "content": " are excluded from the denominator, which prevents the model from yielding the sub-optimal solution that gets " + }, + { + "bbox": [ + 304, + 395, + 546, + 545 + ], + "type": "inline_equation", + "content": "T_{e}^{i}" + }, + { + "bbox": [ + 304, + 395, + 546, + 545 + ], + "type": "text", + "content": " closer to " + }, + { + "bbox": [ + 304, + 395, + 546, + 545 + ], + "type": "inline_equation", + "content": "V_{e}^{i}" + }, + { + "bbox": [ + 304, + 395, + 546, + 545 + ], + "type": "text", + "content": " while pushes " + }, + { + "bbox": [ + 304, + 395, + 546, + 545 + ], + "type": "inline_equation", + "content": "T_{m}^{i}" + }, + { + "bbox": [ + 304, + 395, + 546, + 545 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 395, + 546, + 545 + ], + "type": "inline_equation", + "content": "M_{Vmf}^{i}" + }, + { + "bbox": [ + 304, + 395, + 546, + 545 + ], + "type": "text", + "content": " away from " + }, + { + "bbox": [ + 304, + 395, + 546, + 545 + ], + "type": "inline_equation", + "content": "V_{e}^{i}" + }, + { + "bbox": [ + 304, + 395, + 546, + 545 + ], + "type": "text", + "content": ". Meanwhile, to align the video to the text and fusion modalities, we have the following objective:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 549, + 545, + 624 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 549, + 545, + 624 + ], + "spans": [ + { + "bbox": [ + 306, + 549, + 545, + 624 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} L _ {v ^ {\\prime}} = - \\sum_ {i = 1} ^ {B} \\left[ \\log \\frac {e ^ {s (T _ {e} ^ {i} , V _ {e} ^ {i}) / \\tau}}{\\sum_ {j = 1} ^ {B} e ^ {s (T _ {e} ^ {i} , V _ {e} ^ {j}) / \\tau}} + \\log \\frac {e ^ {s (T _ {m} ^ {i} , V _ {e} ^ {i}) / \\tau}}{\\sum_ {j = 1} ^ {B} e ^ {s (T _ {m} ^ {i} , V _ {e} ^ {j}) / \\tau}} \\right] \\\\ \\left. + \\log \\frac {e ^ {s \\left(M _ {V _ {m f}} ^ {i} , V _ {e} ^ {i}\\right) / \\tau}}{\\sum_ {j = 1} ^ {B} e ^ {s \\left(M _ {V _ {m f}} ^ {i} , V _ {e} ^ {j}\\right) / \\tau}} \\right]. \\tag {3} \\\\ \\end{array}", + "image_path": "0412bff660057fe044fa52f0cf27a36bc77b591018eb1211a452f2cee3a136e7.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 625, + 547, + 697 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 625, + 547, + 697 + ], + "spans": [ + { + "bbox": [ + 304, + 625, + 547, + 697 + ], + "type": "text", + "content": "The above two losses (2) and (3) are summed up to give the tri-modal alignment objective w.r.t. the visual modality: " + }, + { + "bbox": [ + 304, + 625, + 547, + 697 + ], + "type": "inline_equation", + "content": "L_{V} = L_{v} + L_{v^{\\prime}}" + }, + { + "bbox": [ + 304, + 625, + 547, + 697 + ], + "type": "text", + "content": ". Similarly, from the perspective of text modality, we define the tri-modal alignment objective " + }, + { + "bbox": [ + 304, + 625, + 547, + 697 + ], + "type": "inline_equation", + "content": "L_{T} = L_{t} + L_{t^{\\prime}}" + }, + { + "bbox": [ + 304, + 625, + 547, + 697 + ], + "type": "text", + "content": " w.r.t. the text modality. Then, we obtain the final loss of tri-modal alignment " + }, + { + "bbox": [ + 304, + 625, + 547, + 697 + ], + "type": "inline_equation", + "content": "L_{TmA} = L_{V} + L_{T}" + }, + { + "bbox": [ + 304, + 625, + 547, + 697 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 316, + 702, + 527, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 702, + 527, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 702, + 527, + 713 + ], + "type": "text", + "content": "For the detail of " + }, + { + "bbox": [ + 316, + 702, + 527, + 713 + ], + "type": "inline_equation", + "content": "L_{T}" + }, + { + "bbox": [ + 316, + 702, + 527, + 713 + ], + "type": "text", + "content": ", please refer to the supplementary material." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14859" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 287, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 287, + 108 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 287, + 108 + ], + "type": "text", + "content": "To get the masked samples for TMA, we conduct masking over the videos and texts to form masked samples using the following two strategies." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 108, + 286, + 215 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 108, + 286, + 215 + ], + "spans": [ + { + "bbox": [ + 46, + 108, + 286, + 215 + ], + "type": "text", + "content": "Video-block masking strategy. Objects in a video often appear at similar spatial locations across consecutive frames, which compose a tube area in the spatial-temporal domain. In order to construct incomplete video samples with less information leakage, we extend the block-wise mask [3] to video via performing a block-wise mask at the same position in all video frames. In this paper, we randomly replace " + }, + { + "bbox": [ + 46, + 108, + 286, + 215 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 46, + 108, + 286, + 215 + ], + "type": "text", + "content": " patches with a learnable mask token to obtain " + }, + { + "bbox": [ + 46, + 108, + 286, + 215 + ], + "type": "inline_equation", + "content": "V_{m}" + }, + { + "bbox": [ + 46, + 108, + 286, + 215 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 216, + 286, + 383 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 216, + 286, + 383 + ], + "spans": [ + { + "bbox": [ + 46, + 216, + 286, + 383 + ], + "type": "text", + "content": "Semantic text masking strategy. The detailed semantic information in the text basically lies in the verb, noun and adjective words. To facilitate learning their representations, we construct incomplete text samples by randomly masking some verbs, nouns and adjectives in the sentence. Specifically, given a sentence, we use a part-of-speech tagger [33] to tag each word. Then we pick the verb phrases and nouns, and replace " + }, + { + "bbox": [ + 46, + 216, + 286, + 383 + ], + "type": "inline_equation", + "content": "30\\%" + }, + { + "bbox": [ + 46, + 216, + 286, + 383 + ], + "type": "text", + "content": " of them with a special [MASK] token. Note that to avoid drastic semantic changes, we do not mask the auxiliary verb like have, should, will, would, etc. With the masking strategy, when some key elements are masked from the texts, it would form an expression carrying partial information rather than completely changing the information expressed." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 391, + 157, + 403 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 391, + 157, + 403 + ], + "spans": [ + { + "bbox": [ + 47, + 391, + 157, + 403 + ], + "type": "text", + "content": "3.3. Training Objective" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 408, + 286, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 408, + 286, + 574 + ], + "spans": [ + { + "bbox": [ + 46, + 408, + 286, + 574 + ], + "type": "text", + "content": "Pair-wise ranking. In TMA, we take the masked pairs " + }, + { + "bbox": [ + 46, + 408, + 286, + 574 + ], + "type": "inline_equation", + "content": "\\langle V_e, T_m \\rangle" + }, + { + "bbox": [ + 46, + 408, + 286, + 574 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 408, + 286, + 574 + ], + "type": "inline_equation", + "content": "\\langle V_m, T_e \\rangle" + }, + { + "bbox": [ + 46, + 408, + 286, + 574 + ], + "type": "text", + "content": " as positive pairs, considering that masked samples would still carry partial meaningful information though some others are missed. For instance, as shown in Fig. 2 (a), given the text \"A black swan swimming in a calm lake\", \"swimming\" and \"lake\" are masked with [MASK] tokens, generating the masked pair " + }, + { + "bbox": [ + 46, + 408, + 286, + 574 + ], + "type": "inline_equation", + "content": "\\langle V_e, T_m \\rangle" + }, + { + "bbox": [ + 46, + 408, + 286, + 574 + ], + "type": "text", + "content": ". Although some information is removed, the masked sentence still contains other concepts that make it partially matched with the video, such as the \"black swan\". Therefore, in contrastive learning, " + }, + { + "bbox": [ + 46, + 408, + 286, + 574 + ], + "type": "inline_equation", + "content": "V_m" + }, + { + "bbox": [ + 46, + 408, + 286, + 574 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 408, + 286, + 574 + ], + "type": "inline_equation", + "content": "T_m" + }, + { + "bbox": [ + 46, + 408, + 286, + 574 + ], + "type": "text", + "content": " can be considered pseudo positive candidates for " + }, + { + "bbox": [ + 46, + 408, + 286, + 574 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 46, + 408, + 286, + 574 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 408, + 286, + 574 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 46, + 408, + 286, + 574 + ], + "type": "text", + "content": ". However, compared to the original pair " + }, + { + "bbox": [ + 46, + 408, + 286, + 574 + ], + "type": "inline_equation", + "content": "\\langle V_e, T_e \\rangle" + }, + { + "bbox": [ + 46, + 408, + 286, + 574 + ], + "type": "text", + "content": ", we consider that the semantic consistency in the masked pairs should be weaker. Based on this prior, we further propose a pair-wise ranking loss:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 52, + 582, + 286, + 617 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 52, + 582, + 286, + 617 + ], + "spans": [ + { + "bbox": [ + 52, + 582, + 286, + 617 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} L _ {r a n k} = \\max \\left(0, - \\left(\\operatorname {s i m} \\left(V _ {e}, T _ {e}\\right) / \\tau - \\operatorname {s i m} \\left(V _ {e}, T _ {m}\\right) / \\tau\\right) + \\lambda\\right) \\\\ + \\max \\left(0, - \\left(\\operatorname {s i m} \\left(V _ {e}, T _ {e}\\right) / \\tau - \\operatorname {s i m} \\left(V _ {m}, T _ {e}\\right) / \\tau\\right) + \\lambda\\right), \\tag {4} \\\\ \\end{array}", + "image_path": "542aec5c364e6ab11fa9d4f927777c2b25bbea01f4c5030a5b385ccf4dcb5d5c.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 617, + 286, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 617, + 286, + 700 + ], + "spans": [ + { + "bbox": [ + 46, + 617, + 286, + 700 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 617, + 286, + 700 + ], + "type": "inline_equation", + "content": "\\lambda > 0" + }, + { + "bbox": [ + 46, + 617, + 286, + 700 + ], + "type": "text", + "content": " is a margin hyper-parameter. Eq. (4) urges the model to be aware of the gap of semantic consistency between " + }, + { + "bbox": [ + 46, + 617, + 286, + 700 + ], + "type": "inline_equation", + "content": "\\langle V_e, T_m \\rangle / \\langle V_m, T_e \\rangle" + }, + { + "bbox": [ + 46, + 617, + 286, + 700 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 617, + 286, + 700 + ], + "type": "inline_equation", + "content": "\\langle V_e, T_e \\rangle" + }, + { + "bbox": [ + 46, + 617, + 286, + 700 + ], + "type": "text", + "content": " brought about by concepts missing in the masked pairs. With the pairwise ranking objective, our model is able to maintain fine-grained perceptual capability while improving its generalizability." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 701, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 701, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 701, + 286, + 713 + ], + "type": "text", + "content": "Semantic enhanced masked language modeling. Masked" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 72, + 545, + 227 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 227 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 227 + ], + "type": "text", + "content": "language modeling (MLM) is a classical pre-training task in VidL, which promotes the interaction between different modalities in the cross-modal encoder. Combining the classical MLM with our semantic text masking strategy, we present the Semantic Enhanced Masked Language Modeling task, which facilitates the representation learning on the key concepts, i.e., verb, noun and adjective words. Besides, considering the class-imbalance between different words, we use focal loss [28] instead of the traditional cross-entropy loss to improve the MLM loss. We apply the MLM loss to the reconstruction of the masked text tokens i.e., " + }, + { + "bbox": [ + 304, + 72, + 545, + 227 + ], + "type": "inline_equation", + "content": "m_{t_j} \\in M_{T_m}" + }, + { + "bbox": [ + 304, + 72, + 545, + 227 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 72, + 545, + 227 + ], + "type": "inline_equation", + "content": "t_j = [Mask]" + }, + { + "bbox": [ + 304, + 72, + 545, + 227 + ], + "type": "text", + "content": ". The MLM loss is defined as:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 333, + 235, + 545, + 273 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 235, + 545, + 273 + ], + "spans": [ + { + "bbox": [ + 333, + 235, + 545, + 273 + ], + "type": "interline_equation", + "content": "L _ {m l m} = - \\frac {1}{B} \\sum_ {i = 1} ^ {B} \\sum_ {m _ {t _ {j}} \\in M _ {T _ {m}} ^ {i}} [ (1 - p _ {t _ {j}} ^ {i}) ^ {\\gamma} p _ {t _ {j}} ^ {i} ] \\tag {5}", + "image_path": "c2e54a0f7dcef8a2a36c6873894b9ea7e761f735ec613d079a13e9bd4eeb0cb5.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 282, + 545, + 330 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 282, + 545, + 330 + ], + "spans": [ + { + "bbox": [ + 304, + 282, + 545, + 330 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 282, + 545, + 330 + ], + "type": "inline_equation", + "content": "p_{t_j}^i" + }, + { + "bbox": [ + 304, + 282, + 545, + 330 + ], + "type": "text", + "content": " denote the predicted probability distribution of the masked token " + }, + { + "bbox": [ + 304, + 282, + 545, + 330 + ], + "type": "inline_equation", + "content": "t_j" + }, + { + "bbox": [ + 304, + 282, + 545, + 330 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 304, + 282, + 545, + 330 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 282, + 545, + 330 + ], + "type": "text", + "content": "th sentence in the batch, and " + }, + { + "bbox": [ + 304, + 282, + 545, + 330 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 304, + 282, + 545, + 330 + ], + "type": "text", + "content": " is a hyper-parameter. Finally, the overall pre-training objective of Clover is:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 363, + 342, + 545, + 354 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 363, + 342, + 545, + 354 + ], + "spans": [ + { + "bbox": [ + 363, + 342, + 545, + 354 + ], + "type": "interline_equation", + "content": "L = L _ {T m A} + L _ {r a n k} + L _ {m l m}. \\tag {6}", + "image_path": "68551386bb24cba559c062910c0bc04cdfd903501419b8b6187c2abf0fb92b48.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 375, + 387, + 388 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 375, + 387, + 388 + ], + "spans": [ + { + "bbox": [ + 306, + 375, + 387, + 388 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 395, + 413, + 408 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 395, + 413, + 408 + ], + "spans": [ + { + "bbox": [ + 306, + 395, + 413, + 408 + ], + "type": "text", + "content": "4.1. Experiment Setup" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 414, + 545, + 497 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 414, + 545, + 497 + ], + "spans": [ + { + "bbox": [ + 304, + 414, + 545, + 497 + ], + "type": "text", + "content": "Pre-training datasets. Following recent work [2, 11, 24], we jointly pre-train our Clover on a video dataset WebVid2M [2] with 2.5M video-text pairs and an image dataset Google Conceptual Captions (CC3M) [44] with 3.3M image-text pairs (we only obtain 2.8M image-text pairs in CC3M due to image url broken). During pretraining, we treat image data as one frame video data." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 498, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 498, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 498, + 545, + 713 + ], + "type": "text", + "content": "Downstream tasks. We evaluate our proposed models on the following downstream tasks. (a) Text-to-Video Retrieval on MSRVTT [55], LSMDC [38] and DiDeMo [1]. For DiDeMo, we follow [23, 31] and evaluate Clover on the paragraph-to-video retrieval, where text sentences for each video are concatenated together as one text query. We do not use the ground-truth proposal for fair comparison with previous works; (b) Multiple-choice QA on TGIF-Action [16], TGIF-Transition [16], MSRVTT-MC [60] and LSMDC-MC [47]; (c) Open-Ended QA on TGIF-Frame [16], MSRVTT-QA [52], MSVD-QA [52] and LSMDC-FiB [38]. For retrieval tasks, we only use the two uni-modal encoders of Clover for fine-tuning and inference. We adopt the two encoders to get the video and text embeddings, and calculate their cosine similarity for retrieval. For QA tasks, we use all the modules for fine-tuning and inference. More details on these datasets and their evaluation usage are provided in the supplementary material." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14860" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 53, + 69, + 542, + 328 + ], + "blocks": [ + { + "bbox": [ + 53, + 69, + 542, + 328 + ], + "lines": [ + { + "bbox": [ + 53, + 69, + 542, + 328 + ], + "spans": [ + { + "bbox": [ + 53, + 69, + 542, + 328 + ], + "type": "table", + "html": "
MethodPre-training datasetDiDeMoMSRVTTLSMDC
R@1R@5R@10MedRR@1R@5R@10MedRR@1R@5R@10MedR
Fine-tune
clipBert [23]COCO [5], VG [20]20.448.060.8622.046.859.96----
Frozen [2]W2M+C3M31.059.872.4331.059.570.5315.030.839.820
VIOLET [9]W2M+C3M+Y180M32.662.874.7-34.563.073.4-16.136.641.2-
HD-VILA [56]HDV100M28.857.469.1435.665.378.0317.434.144.115
ALPRO [24]W2M+C3M35.967.578.8333.960.773.23----
TMVM [27]W2M+C3M36.564.975.4336.264.275.7317.837.145.913.5
All-in-1 [49]W2M+H100M32.761.473.5-37.968.177.1-----
OA-Trans [50]W2M+C3M34.864.475.1335.863.476.5318.234.343.718.5
MILES [12]W2M+C3M36.663.974.0337.763.673.8317.835.644.115.5
MCQ [11]W2M+C3M37.062.273.9337.664.875.1317.935.444.515
Clover (ours)W2M+C3M50.176.785.6140.569.879.4224.844.054.58
Zero-shot
MIL-NCE [39]H100M----9.924.032.429.6----
Frozen [2]W2M+C3M21.146.056.2718.739.651.6109.322.030.151.0
VIOLET [9]W2M+C3M+Y180M23.549.859.8-25.949.559.7-----
ALPRO [24]W2M+C3M23.847.357.9624.144.755.48----
OA-Trans [50]W2M+C3M23.550.459.8623.447.555.68----
MILES [12]W2M+C3M27.250.363.6526.147.256.9711.124.730.650.7
MCQ [11]W2M+C3M25.650.661.1526.046.456.4712.225.932.242
Clover (ours)W2M+C3M29.555.266.3426.449.560.0614.729.238.224
", + "image_path": "32b0f0c8294d9d1377aca281733e777483cc0b9ad5df83854884d35996023f79.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 99, + 369, + 494, + 476 + ], + "blocks": [ + { + "bbox": [ + 47, + 335, + 545, + 368 + ], + "lines": [ + { + "bbox": [ + 47, + 335, + 545, + 368 + ], + "spans": [ + { + "bbox": [ + 47, + 335, + 545, + 368 + ], + "type": "text", + "content": "Table 1. Text-to-video retrieval performance comparison under fine-tune and zero-shot setups. Here higher R@k (Recall K) and lower MedR (Median Recall) indicate better performance. W2M, C3M, H100M, HDV100M, Y180M are short for WebVid2M [2], CC3M [44], HowTo100M [40], HD-VILA-100M [56], YT-Temporal-180M [61], respectively." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 99, + 369, + 494, + 476 + ], + "lines": [ + { + "bbox": [ + 99, + 369, + 494, + 476 + ], + "spans": [ + { + "bbox": [ + 99, + 369, + 494, + 476 + ], + "type": "table", + "html": "
MethodPre-training datasetTGIFMSRVTTLSMDCMSVD
ActionTransitionFrameMCQAMCFiBQA
clipBert [23]COCO, VG82.887.860.388.237.4---
JuskAsk [58]HTVQA69M [58]----41.5--46.3
ALPRO [24]W2M+C3M----42.1--45.9
All-in-1 [49]W2M+H100M92.794.364.292.042.983.1-47.9
VIOLET [9]W2M+C3M+Y180M92.595.768.991.943.982.853.747.9
MERLOT [61]Y180M94.096.269.590.943.181.752.9-
Clover (ours)W2M+C3M95.098.271.695.244.183.754.152.4
", + "image_path": "b09de1abe8278fb0d246f116ea7bdda2ddf0af280e62d894b23ed1f4072cff7a.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 120, + 479, + 471, + 491 + ], + "lines": [ + { + "bbox": [ + 120, + 479, + 471, + 491 + ], + "spans": [ + { + "bbox": [ + 120, + 479, + 471, + 491 + ], + "type": "text", + "content": "Table 2. Performance comparison on transferring to downstream video question answering tasks." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 500, + 289, + 704 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 500, + 289, + 704 + ], + "spans": [ + { + "bbox": [ + 46, + 500, + 289, + 704 + ], + "type": "text", + "content": "Implementation details. Following [9], we initialize the video encoder with a VideoSwin-Base [32] model pretrained on Kinetics-400 [18]. The text encoder is initialized from pre-trained Bert-Base [7]. The multi-modal encoder is initialized from the first three layers of the pre-trained Bert-Base model. We train Clover end-to-end during both pretraining and fine-tuning. We pre-train Clover for 40 epochs, using a batch size of 1024 on 64 NVIDIA A100 GPUs. We use AdamW [34] optimizer with a weight decay 0.005 and betas (0.9, 0.98). The learning rate is first warmed-up by 4 epochs to 5e-5 and then decays following a cosine annealing decay schedule. We resize all the video frames to " + }, + { + "bbox": [ + 46, + 500, + 289, + 704 + ], + "type": "inline_equation", + "content": "224 \\times 224" + }, + { + "bbox": [ + 46, + 500, + 289, + 704 + ], + "type": "text", + "content": " and split each frame into patches with a size of " + }, + { + "bbox": [ + 46, + 500, + 289, + 704 + ], + "type": "inline_equation", + "content": "32 \\times 32" + }, + { + "bbox": [ + 46, + 500, + 289, + 704 + ], + "type": "text", + "content": ". We choose hyper-parameter " + }, + { + "bbox": [ + 46, + 500, + 289, + 704 + ], + "type": "inline_equation", + "content": "\\tau = 0.05" + }, + { + "bbox": [ + 46, + 500, + 289, + 704 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 500, + 289, + 704 + ], + "type": "inline_equation", + "content": "\\lambda = 5" + }, + { + "bbox": [ + 46, + 500, + 289, + 704 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 500, + 289, + 704 + ], + "type": "inline_equation", + "content": "\\gamma = 2" + }, + { + "bbox": [ + 46, + 500, + 289, + 704 + ], + "type": "text", + "content": ". We set dimension " + }, + { + "bbox": [ + 46, + 500, + 289, + 704 + ], + "type": "inline_equation", + "content": "D = 768" + }, + { + "bbox": [ + 46, + 500, + 289, + 704 + ], + "type": "text", + "content": " in our model. For each video, we randomly sample 8 frames while preserving their order in-between. For fine-tuning on retrieval tasks, we only fine" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 500, + 545, + 584 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 500, + 545, + 584 + ], + "spans": [ + { + "bbox": [ + 304, + 500, + 545, + 584 + ], + "type": "text", + "content": "tune the uni-modal encoders of Clover with InfoNCE [48] loss. For fine-tuning on video QA task, we add a simple MLPs that takes the multi-modal [CLS] embedding as input for classification, and optimize the whole Clover model with cross-entropy loss. During fine-tuning, we follow the conventional set-up in [9, 36], and all the fine-tuning experiments are performed on 8 NVIDIA A100 GPUs." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 305, + 597, + 469, + 610 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 597, + 469, + 610 + ], + "spans": [ + { + "bbox": [ + 305, + 597, + 469, + 610 + ], + "type": "text", + "content": "4.2. Comparing to State-of-the-art" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "content": "Text-to-video retrieval. Tab. 1 shows the retrieval results on the DiDeMo [1], MSRVTT [55] and LSMDC [38] datasets with both zero-shot and fine-tuning settings. Our Clover surpasses previous works on all the datasets by a large margin. The significant performance gain under zero-shot evaluation demonstrates its stronger generalization ability. Specifically, Clover brings " + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "inline_equation", + "content": "4.9\\%" + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "content": " gain in R@10 over MCQ [11], the current SOTA method on re" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "14861" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 75, + 70, + 521, + 149 + ], + "blocks": [ + { + "bbox": [ + 75, + 70, + 521, + 149 + ], + "lines": [ + { + "bbox": [ + 75, + 70, + 521, + 149 + ], + "spans": [ + { + "bbox": [ + 75, + 70, + 521, + 149 + ], + "type": "table", + "html": "
MethodDiDeMoMSRVTTLSMDC-MCTGIF-Frame
R@1R@5R@10MedRR@1R@5R@10MedRAccAcc
Baseline22.647.958.2719.841.751.11078.868.9
+ TMA24.749.760.0622.742.252.2980.069.3
+ TMA+SM25.349.560.5622.542.752.0980.569.4
+ TMA+SM+RankL26.451.161.3523.443.352.4980.769.7
", + "image_path": "c151788e4250461e372350677b7b20477bedcb6c5ecf8c73c413ad11346ea0a2.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 52, + 196, + 285, + 239 + ], + "blocks": [ + { + "bbox": [ + 46, + 157, + 547, + 191 + ], + "lines": [ + { + "bbox": [ + 46, + 157, + 547, + 191 + ], + "spans": [ + { + "bbox": [ + 46, + 157, + 547, + 191 + ], + "type": "text", + "content": "Table 3. Effects of our pre-training tasks for Clover. We report zero-shot text-video retrieval performance on DiDeMo and MSRVTT, and fine-tune video QA performance on LSMDC-MC and TGIF-Frame. TMA, SM and RankL: tri-modal alignment, semantic masking strategy and pair-wise ranking loss." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 52, + 196, + 285, + 239 + ], + "lines": [ + { + "bbox": [ + 52, + 196, + 285, + 239 + ], + "spans": [ + { + "bbox": [ + 52, + 196, + 285, + 239 + ], + "type": "table", + "html": "
MethodAveraged similarity scores of positive pairsAveraged similarity margin between positive and negative samples
w/o TMA0.560.30
w TMA0.650.35
", + "image_path": "d29dd64ea3adbd4f6e0d052e97885b173ebc857608f71f872b430560f5aa986d.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 247, + 287, + 281 + ], + "lines": [ + { + "bbox": [ + 46, + 247, + 287, + 281 + ], + "spans": [ + { + "bbox": [ + 46, + 247, + 287, + 281 + ], + "type": "text", + "content": "Table 4. Affect of TMA on the video and text representations in learned embedding space. The experiment is conducted on MSRVTT-1kA test set under zero-shot setting." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 45, + 289, + 287, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 289, + 287, + 456 + ], + "spans": [ + { + "bbox": [ + 45, + 289, + 287, + 456 + ], + "type": "text", + "content": "trieval tasks with the same pre-train data, averaged on the three datasets. Moreover, Clover outperforms VIOLET [9], though VIOLET adopts a considerably larger pre-training dataset (YTT180M [61] of 180M visual-text pairs vs. WebVid2M+CC3M of only 5.5M samples). Clover also presents superior performance when fine-tuned on the downstream datasets, compared with the SOTAs. On all three datasets, Clover surpasses previous works by a large margin across all the metrics and obtains an average improvement of " + }, + { + "bbox": [ + 45, + 289, + 287, + 456 + ], + "type": "inline_equation", + "content": "8.7\\%" + }, + { + "bbox": [ + 45, + 289, + 287, + 456 + ], + "type": "text", + "content": " on R@10. Moreover, different from some previous work [9,23,24] that adopts a joint multi-modal encoder and requires exhaustively pairing every video and text during retrieval, Clover directly deploys the video encoder and text encoder for retrieval tasks, making it much more efficient." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 456, + 288, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 456, + 288, + 613 + ], + "spans": [ + { + "bbox": [ + 46, + 456, + 288, + 613 + ], + "type": "text", + "content": "Video question answering. Tab. 2 reports the results of Clover and current SOTAs on four video QA datasets. Though using much less pre-training data, Clover outperforms another unified VidL model i.e., VIOLET [9] significantly. Moreover, compared to the current SOTA method MERLOT [61], Clover brings improvements of " + }, + { + "bbox": [ + 46, + 456, + 288, + 613 + ], + "type": "inline_equation", + "content": "1.0\\%" + }, + { + "bbox": [ + 46, + 456, + 288, + 613 + ], + "type": "text", + "content": " on TGIF-Action, " + }, + { + "bbox": [ + 46, + 456, + 288, + 613 + ], + "type": "inline_equation", + "content": "2.0\\%" + }, + { + "bbox": [ + 46, + 456, + 288, + 613 + ], + "type": "text", + "content": " on TGIF-Transition, " + }, + { + "bbox": [ + 46, + 456, + 288, + 613 + ], + "type": "inline_equation", + "content": "2.1\\%" + }, + { + "bbox": [ + 46, + 456, + 288, + 613 + ], + "type": "text", + "content": " on TGIF-Frame, " + }, + { + "bbox": [ + 46, + 456, + 288, + 613 + ], + "type": "inline_equation", + "content": "4.3\\%" + }, + { + "bbox": [ + 46, + 456, + 288, + 613 + ], + "type": "text", + "content": " on MSRVTT-MC, " + }, + { + "bbox": [ + 46, + 456, + 288, + 613 + ], + "type": "inline_equation", + "content": "1.0\\%" + }, + { + "bbox": [ + 46, + 456, + 288, + 613 + ], + "type": "text", + "content": " on MSRVTT-QA, " + }, + { + "bbox": [ + 46, + 456, + 288, + 613 + ], + "type": "inline_equation", + "content": "2.0\\%" + }, + { + "bbox": [ + 46, + 456, + 288, + 613 + ], + "type": "text", + "content": " on LSMDC-MC and " + }, + { + "bbox": [ + 46, + 456, + 288, + 613 + ], + "type": "inline_equation", + "content": "1.2\\%" + }, + { + "bbox": [ + 46, + 456, + 288, + 613 + ], + "type": "text", + "content": " on LSMDC-FiB. Note that JustAsk [58] and MERLOT [61] are specifically designed for video QA and trained on orders of magnitude larger datasets (e.g., HTVQA69M [58], YTT180M [61]). The experiment results again confirm the superiority of Clover." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 622, + 111, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 622, + 111, + 635 + ], + "spans": [ + { + "bbox": [ + 47, + 622, + 111, + 635 + ], + "type": "text", + "content": "4.3. Analysis" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 642, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 642, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 642, + 288, + 715 + ], + "type": "text", + "content": "We conduct ablation experiments on two retrieval datasets (DiDeMo and MSRvTT) and two Video QA datasets (TGIF-Frame and LSMDC-MC). Due to the computational resource limit, we randomly sample 1 million video-text pairs from WebVid2M [2] to build the WebVid1M subset for model pre-training under all the ablation stud" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 198, + 545, + 304 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 198, + 545, + 304 + ], + "spans": [ + { + "bbox": [ + 304, + 198, + 545, + 304 + ], + "type": "text", + "content": "ies. We keep the model architecture unchanged but pretrain the model with only MLM and InfoNCE losses as a baseline, which is similar to the baseline adopted in VIOLET [9]. The only difference is that we replace the ITM applied to the cross-modal encoder outputs in VIOLET [9] with InfoNCE applied to the outputs of uni-modal encoders and separate the text encoder from the cross-modal encoder. All the ablation experiments are conducted with batch size 1024 on 32 NVIDIA A100 GPUs." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 306, + 546, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 306, + 546, + 640 + ], + "spans": [ + { + "bbox": [ + 304, + 306, + 546, + 640 + ], + "type": "text", + "content": "Effect of tri-modal alignment. Tri-modal alignment (TMA) aims to better correlate cross-modal alignment and fusion. To evaluate the contribution of TMA, we remove the pair-wise ranking objective and semantic masking strategy, and employ the classical MLM task as in the baseline method. Compared with the baseline, TMA explicitly associates the outputs of uni-modal encoders with the outputs of the multi-modal encoder. As shown in Tab. 3, the model trained with TMA outperforms the baseline on both retrieval and video QA, demonstrating its effectiveness. For a better understanding of the effect of TMA, we further report averaged cosine similarity of videos and texts in MSRVTT dataset (in zero-shot setting) calculated by models trained with/without TMA, respectively. For a fair comparison, we select all video queries (197 in total) that were successfully responded to by both models. We recall 100 results for each query, and take the ground-truth pairs as positive and others as negative. As shown in Tab. 4, the model armed with tri-modal alignment assigns higher similarity scores to positive pairs compared to the model without it. Meanwhile, with tri-modal alignment, the margin between the positive and negative samples is also increased. It reveals that with the help of tri-modal alignment, the distance between video and text that contain consistent semantics is reduced, while the margin between the misaligned samples is increased. The results demonstrate that with fused multi-modal representation as anchors, tri-modal alignment helps the video and language modalities to be better aligned." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 642, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 642, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 642, + 547, + 715 + ], + "type": "text", + "content": "Effect of semantic masking strategy. We incorporate a semantic masking strategy to form the masked pairs in trimodal alignment task. As shown in Tab. 3, on more diverse evaluation benchmarks, e.g. DiDemo, where each video is paired with 5 different ground-truth texts, using masked samples brings more gains. The result reveals that the pro" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "14862" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 75, + 56, + 521, + 144 + ], + "blocks": [ + { + "bbox": [ + 75, + 56, + 521, + 144 + ], + "lines": [ + { + "bbox": [ + 75, + 56, + 521, + 144 + ], + "spans": [ + { + "bbox": [ + 75, + 56, + 521, + 144 + ], + "type": "table", + "html": "
MethodModel ArchitectureTraining ObjectivesMSRVTTDiDeMoTGIF-FrameLSMDC-MC
Two uni-modal encoderCross-modal encoderInfoNCEMLMTMA + RanklossR@1R@1AccAcc
IND-A✓✓20.722.9--
IND-F✓✓✓--69.079.0
COMB✓✓✓✓19.822.668.978.8
Clover✓✓✓✓23.426.469.780.7
", + "image_path": "87698ca4b49657409b4a4907dccfda3dd3ad1fd1d27a70ef4b6cc1a9ec47fc23.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 154, + 545, + 165 + ], + "lines": [ + { + "bbox": [ + 47, + 154, + 545, + 165 + ], + "spans": [ + { + "bbox": [ + 47, + 154, + 545, + 165 + ], + "type": "text", + "content": "Table 5. Comparisons with task-independently trained models (IND) and naive combination of the uni- and cross- modal encoder (COMB)" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 50, + 176, + 111, + 220 + ], + "blocks": [ + { + "bbox": [ + 50, + 176, + 111, + 220 + ], + "lines": [ + { + "bbox": [ + 50, + 176, + 111, + 220 + ], + "spans": [ + { + "bbox": [ + 50, + 176, + 111, + 220 + ], + "type": "image", + "image_path": "2e58ff4bf58f40c41af7f8dbfc21d7d6b1964f84c19ba8bcd3d377f29760a0ef.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 52, + 221, + 229, + 242 + ], + "lines": [ + { + "bbox": [ + 52, + 221, + 229, + 242 + ], + "spans": [ + { + "bbox": [ + 52, + 221, + 229, + 242 + ], + "type": "text", + "content": "Baseline: a man is showing the interior of a car \nClover: inside of a car with large space and safety" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 113, + 176, + 168, + 220 + ], + "blocks": [ + { + "bbox": [ + 113, + 176, + 168, + 220 + ], + "lines": [ + { + "bbox": [ + 113, + 176, + 168, + 220 + ], + "spans": [ + { + "bbox": [ + 113, + 176, + 168, + 220 + ], + "type": "image", + "image_path": "2bfdfe5b1d7a73efa803cb07f85464665ab823cc62ecb196edb800aa697daf4f.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 172, + 176, + 227, + 220 + ], + "blocks": [ + { + "bbox": [ + 172, + 176, + 227, + 220 + ], + "lines": [ + { + "bbox": [ + 172, + 176, + 227, + 220 + ], + "spans": [ + { + "bbox": [ + 172, + 176, + 227, + 220 + ], + "type": "image", + "image_path": "202c7cf70b69e855b2960ae0660acce90836cfb7db406a4be60349be6a47a035.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 231, + 176, + 286, + 220 + ], + "blocks": [ + { + "bbox": [ + 231, + 176, + 286, + 220 + ], + "lines": [ + { + "bbox": [ + 231, + 176, + 286, + 220 + ], + "spans": [ + { + "bbox": [ + 231, + 176, + 286, + 220 + ], + "type": "image", + "image_path": "5e30e897642881d85a6c1c33a1789b9097d291c544fb3e51b74158bca0a9611e.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 52, + 248, + 110, + 291 + ], + "blocks": [ + { + "bbox": [ + 52, + 248, + 110, + 291 + ], + "lines": [ + { + "bbox": [ + 52, + 248, + 110, + 291 + ], + "spans": [ + { + "bbox": [ + 52, + 248, + 110, + 291 + ], + "type": "image", + "image_path": "99598dfc66d0e3f4188dab9891b69b52d491f464e26d08e0eeb7ac8b96890ffe.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 52, + 293, + 200, + 313 + ], + "lines": [ + { + "bbox": [ + 52, + 293, + 200, + 313 + ], + "spans": [ + { + "bbox": [ + 52, + 293, + 200, + 313 + ], + "type": "text", + "content": "Baseline: a man is folding pieces of paper \nClover: a person folds a paper airplane" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 115, + 248, + 168, + 291 + ], + "blocks": [ + { + "bbox": [ + 115, + 248, + 168, + 291 + ], + "lines": [ + { + "bbox": [ + 115, + 248, + 168, + 291 + ], + "spans": [ + { + "bbox": [ + 115, + 248, + 168, + 291 + ], + "type": "image", + "image_path": "462c84091090bc713fa3dc7e0b3c4abbf7db80a09ffc99fb3af2f2fb488d8d6f.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 327, + 287, + 349 + ], + "lines": [ + { + "bbox": [ + 46, + 327, + 287, + 349 + ], + "spans": [ + { + "bbox": [ + 46, + 327, + 287, + 349 + ], + "type": "text", + "content": "Figure 3. Qualitative results of zero-shot video to text retrieval results on MSRVTT [55]." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 173, + 248, + 227, + 291 + ], + "blocks": [ + { + "bbox": [ + 173, + 248, + 227, + 291 + ], + "lines": [ + { + "bbox": [ + 173, + 248, + 227, + 291 + ], + "spans": [ + { + "bbox": [ + 173, + 248, + 227, + 291 + ], + "type": "image", + "image_path": "6668d484ba85d5bf6d7f8da3c2497dce53ecc2b883b832b200ad5ec0cb702042.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 233, + 248, + 285, + 291 + ], + "blocks": [ + { + "bbox": [ + 233, + 248, + 285, + 291 + ], + "lines": [ + { + "bbox": [ + 233, + 248, + 285, + 291 + ], + "spans": [ + { + "bbox": [ + 233, + 248, + 285, + 291 + ], + "type": "image", + "image_path": "6fa58e8590c243eba2efd5b6862e74bb238afb64237eb90ff1c090b5a8c50102.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 354, + 287, + 388 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 354, + 287, + 388 + ], + "spans": [ + { + "bbox": [ + 46, + 354, + 287, + 388 + ], + "type": "text", + "content": "posed semantic masking strategy facilitates the model to capture key semantic information in more complex scenes and makes it achieve better results." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 46, + 389, + 287, + 473 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 389, + 287, + 473 + ], + "spans": [ + { + "bbox": [ + 46, + 389, + 287, + 473 + ], + "type": "text", + "content": "Effect of pair-wise ranking. To make the model capture the semantic information difference between the masked pairs and complete pairs, we adopt the pair-wise rank loss. As shown in Tab. 3, with pair-wise ranking loss, the performance of the model is further improved. We also show the visualization results below to further illustrate the effect of pair-wise ranking." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 46, + 474, + 287, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 474, + 287, + 641 + ], + "spans": [ + { + "bbox": [ + 46, + 474, + 287, + 641 + ], + "type": "text", + "content": "Clover makes cross-modal alignment and fusion mutually improving. We compare Clover with the ones that use the same architecture but employ different pre-training tasks in Tab. 5. " + }, + { + "bbox": [ + 46, + 474, + 287, + 641 + ], + "type": "inline_equation", + "content": "IND-A" + }, + { + "bbox": [ + 46, + 474, + 287, + 641 + ], + "type": "text", + "content": " represents the model only trained with InfoNCE loss for cross-modal alignment. " + }, + { + "bbox": [ + 46, + 474, + 287, + 641 + ], + "type": "inline_equation", + "content": "IND-F" + }, + { + "bbox": [ + 46, + 474, + 287, + 641 + ], + "type": "text", + "content": " indicates the model trained with MLM loss for cross-modal fusion. We also report results achieved by the model (i.e., COMB) that simply combines the uni-modal encoders and multi-modal encoder as well as the training objectives. We can see that the simple combination hurts the performance, while our Clover achieves superior performance than the task-independently trained " + }, + { + "bbox": [ + 46, + 474, + 287, + 641 + ], + "type": "inline_equation", + "content": "IND-A" + }, + { + "bbox": [ + 46, + 474, + 287, + 641 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 474, + 287, + 641 + ], + "type": "inline_equation", + "content": "IND-F" + }, + { + "bbox": [ + 46, + 474, + 287, + 641 + ], + "type": "text", + "content": ". It reveals that Clover is able to get the model's cross-modal alignment and fusion capability mutually enhanced." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 46, + 642, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 642, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 642, + 287, + 713 + ], + "type": "text", + "content": "Qualitative analysis. In Fig. 3 and Fig. 4, we show the qualitative results of Clover and the baseline method on zero-shot video-text retrieval and VQA tasks. Specifically, we present the query videos with the matched texts, and the video-question pair with the answer. Clover empowers the model with stronger video-text understanding capabilities." + } + ] + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 310, + 175, + 366, + 218 + ], + "blocks": [ + { + "bbox": [ + 310, + 175, + 366, + 218 + ], + "lines": [ + { + "bbox": [ + 310, + 175, + 366, + 218 + ], + "spans": [ + { + "bbox": [ + 310, + 175, + 366, + 218 + ], + "type": "image", + "image_path": "f1316f9f5675cb4a33cdd88e7d78c7d679a44a60f81fc45521fde877815a75d7.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 311, + 220, + 433, + 237 + ], + "lines": [ + { + "bbox": [ + 311, + 220, + 433, + 237 + ], + "spans": [ + { + "bbox": [ + 311, + 220, + 433, + 237 + ], + "type": "text", + "content": "Question: What is a guy hugging? \nBaseline: Animal Clover: Lion" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 370, + 175, + 426, + 218 + ], + "blocks": [ + { + "bbox": [ + 370, + 175, + 426, + 218 + ], + "lines": [ + { + "bbox": [ + 370, + 175, + 426, + 218 + ], + "spans": [ + { + "bbox": [ + 370, + 175, + 426, + 218 + ], + "type": "image", + "image_path": "62686c8452653c9a3206b7e51ae02eeba485c0ad74d9cb941cf6c3a314fc6875.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 429, + 175, + 485, + 218 + ], + "blocks": [ + { + "bbox": [ + 429, + 175, + 485, + 218 + ], + "lines": [ + { + "bbox": [ + 429, + 175, + 485, + 218 + ], + "spans": [ + { + "bbox": [ + 429, + 175, + 485, + 218 + ], + "type": "image", + "image_path": "43a8c699237e35b1055cbe51ad2b450317e3bbd0048a74aa4c068f9ab63af8bd.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 488, + 175, + 545, + 218 + ], + "blocks": [ + { + "bbox": [ + 488, + 175, + 545, + 218 + ], + "lines": [ + { + "bbox": [ + 488, + 175, + 545, + 218 + ], + "spans": [ + { + "bbox": [ + 488, + 175, + 545, + 218 + ], + "type": "image", + "image_path": "21ddc29e4c885e022189e3ffe838d9ce27bf821c830a1a38f0f0e3e595ed5455.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 310, + 245, + 365, + 288 + ], + "blocks": [ + { + "bbox": [ + 310, + 245, + 365, + 288 + ], + "lines": [ + { + "bbox": [ + 310, + 245, + 365, + 288 + ], + "spans": [ + { + "bbox": [ + 310, + 245, + 365, + 288 + ], + "type": "image", + "image_path": "8c74ef8541c5fb818a88aa3822a1d8d064b92fa65ca94854e52825b0f879095b.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 311, + 291, + 477, + 309 + ], + "lines": [ + { + "bbox": [ + 311, + 291, + 477, + 309 + ], + "spans": [ + { + "bbox": [ + 311, + 291, + 477, + 309 + ], + "type": "text", + "content": "Question: What does a person wash with a brush? Baseline: Face Clover: Toy" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_caption" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 370, + 245, + 425, + 288 + ], + "blocks": [ + { + "bbox": [ + 370, + 245, + 425, + 288 + ], + "lines": [ + { + "bbox": [ + 370, + 245, + 425, + 288 + ], + "spans": [ + { + "bbox": [ + 370, + 245, + 425, + 288 + ], + "type": "image", + "image_path": "46e79f251ce445f111bc523f64810d47c0dc1708866bb8d86c25c782de0551f0.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 324, + 545, + 346 + ], + "lines": [ + { + "bbox": [ + 305, + 324, + 545, + 346 + ], + "spans": [ + { + "bbox": [ + 305, + 324, + 545, + 346 + ], + "type": "text", + "content": "Figure 4. Qualitative results of video question answering results on MSRVTT-QA [52]." + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_caption" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 429, + 245, + 485, + 288 + ], + "blocks": [ + { + "bbox": [ + 429, + 245, + 485, + 288 + ], + "lines": [ + { + "bbox": [ + 429, + 245, + 485, + 288 + ], + "spans": [ + { + "bbox": [ + 429, + 245, + 485, + 288 + ], + "type": "image", + "image_path": "f939e05940ef2722fd4631faab0281c3a85afc150b665c2e58d77303aeaa9cb1.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 488, + 245, + 545, + 288 + ], + "blocks": [ + { + "bbox": [ + 488, + 245, + 545, + 288 + ], + "lines": [ + { + "bbox": [ + 488, + 245, + 545, + 288 + ], + "spans": [ + { + "bbox": [ + 488, + 245, + 545, + 288 + ], + "type": "image", + "image_path": "dfa832d6663c0c667b2f931a3dafbe07f37026a77bf46b809ea3dc63e748e4e7.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + } + ], + "index": 25 + }, + { + "bbox": [ + 305, + 348, + 545, + 419 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 348, + 545, + 419 + ], + "spans": [ + { + "bbox": [ + 305, + 348, + 545, + 419 + ], + "type": "text", + "content": "ity. For example, for the second video in Fig. 3, Clover returns the text with \"paper airplane\", which is a more accurate description of the video content; for the first video in Fig. 4, Clover generates a more accurate answer \"Lion\" than the baseline answer \"Animal\". The results demonstrate the superiority of Clover." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 306, + 430, + 378, + 441 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 430, + 378, + 441 + ], + "spans": [ + { + "bbox": [ + 306, + 430, + 378, + 441 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 304, + 451, + 545, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 451, + 545, + 605 + ], + "spans": [ + { + "bbox": [ + 304, + 451, + 545, + 605 + ], + "type": "text", + "content": "In this paper, we present Clover, a new end-to-end Video-Language pre-training method for both high-efficiency video-text retrieval and video question answering. Clover introduces a novel Tri-modal Alignment task to better align the representations from visual, text and fused modalities, which explicitly correlate the uni-modal encoder and multi-modal encoder. It also introduces semantic masking strategy and pair-wise ranking loss to further improve the cross-modality modal training. Extensive experiments conducted on the three retrieval datasets and eight video QA datasets clearly demonstrated, as a general video-text model, its consistent superiority for video-text understanding." + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 304, + 606, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 606, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 606, + 545, + 713 + ], + "type": "text", + "content": "Acknowledgements. This work was supported by National Key R&D Program of China (No.2022ZD0118201), the National Science Fund for Distinguished Young Scholars (No.62025603), the National Natural Science Foundation of China (No. U21B2037, No. U22B2051, No. 62176222, No. 62176223, No. 62176226, No. 62072386, No. 62072387, No. 62072389, No. 62002305 and No. 62272401), and the Natural Science Foundation of Fujian Province of China (No.2021J01002, No.2022J06001)." + } + ] + } + ], + "index": 31 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14863" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 289, + 714 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 53, + 91, + 289, + 146 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 289, + 146 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 289, + 146 + ], + "type": "text", + "content": "[1] Lisa Anne Hendricks, Oliver Wang, Eli Shechtman, Josef Sivic, Trevor Darrell, and Bryan Russell. Localizing moments in video with natural language. In Proceedings of the IEEE international conference on computer vision, pages 5803-5812, 2017. 1, 5, 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 148, + 288, + 203 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 148, + 288, + 203 + ], + "spans": [ + { + "bbox": [ + 53, + 148, + 288, + 203 + ], + "type": "text", + "content": "[2] Max Bain, Arsha Nagrani, Gül Varol, and Andrew Zisserman. Frozen in time: A joint video and image encoder for end-to-end retrieval. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1728-1738, 2021. 1, 2, 5, 6, 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 205, + 288, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 205, + 288, + 237 + ], + "spans": [ + { + "bbox": [ + 53, + 205, + 288, + 237 + ], + "type": "text", + "content": "[3] Hangbo Bao, Li Dong, Songhao Piao, and Furu Wei. Beit: Bert pre-training of image transformers. In International Conference on Learning Representations, 2021. 2, 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 239, + 288, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 239, + 288, + 293 + ], + "spans": [ + { + "bbox": [ + 53, + 239, + 288, + 293 + ], + "type": "text", + "content": "[4] David Chen and William B Dolan. Collecting highly parallel data for paraphrase evaluation. In Proceedings of the 49th annual meeting of the association for computational linguistics: human language technologies, pages 190–200, 2011. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 295, + 288, + 340 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 295, + 288, + 340 + ], + "spans": [ + { + "bbox": [ + 53, + 295, + 288, + 340 + ], + "type": "text", + "content": "[5] Xinlei Chen, Hao Fang, Tsung-Yi Lin, Ramakrishna Vedantam, Saurabh Gupta, Piotr Dólár, and C Lawrence Zitnick. Microsoft coco captions: Data collection and evaluation server. arXiv preprint arXiv:1504.00325, 2015. 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 342, + 288, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 342, + 288, + 385 + ], + "spans": [ + { + "bbox": [ + 53, + 342, + 288, + 385 + ], + "type": "text", + "content": "[6] Xing Cheng, Hezheng Lin, Xiangyu Wu, Fan Yang, and Dong Shen. Improving video-text retrieval by multi-stream corpus alignment and dual softmax loss. arXiv preprint arXiv:2109.04290, 2021. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 387, + 288, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 387, + 288, + 464 + ], + "spans": [ + { + "bbox": [ + 53, + 387, + 288, + 464 + ], + "type": "text", + "content": "[7] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 4171–4186, 2019. 2, 3, 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 465, + 288, + 529 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 465, + 288, + 529 + ], + "spans": [ + { + "bbox": [ + 53, + 465, + 288, + 529 + ], + "type": "text", + "content": "[8] Chenyou Fan, Xiaofan Zhang, Shu Zhang, Wensheng Wang, Chi Zhang, and Heng Huang. Heterogeneous memory enhanced multimodal attention model for video question answering. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1999-2007, 2019. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 533, + 288, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 533, + 288, + 577 + ], + "spans": [ + { + "bbox": [ + 53, + 533, + 288, + 577 + ], + "type": "text", + "content": "[9] Tsu-Jui Fu, Linjie Li, Zhe Gan, Kevin Lin, William Yang Wang, Lijuan Wang, and Zicheng Liu. Violet: End-to-end video-language transformers with masked visual-token modeling. arXiv preprint arXiv:2111.12681, 2021. 1, 2, 3, 6, 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 578, + 288, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 578, + 288, + 622 + ], + "spans": [ + { + "bbox": [ + 48, + 578, + 288, + 622 + ], + "type": "text", + "content": "[10] Valentin Gabeur, Chen Sun, Karteek Alahari, and Cordelia Schmid. Multi-modal transformer for video retrieval. In European Conference on Computer Vision, pages 214-229. Springer, 2020. 1, 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 624, + 288, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 624, + 288, + 678 + ], + "spans": [ + { + "bbox": [ + 48, + 624, + 288, + 678 + ], + "type": "text", + "content": "[11] Yuying Ge, Yixiao Ge, Xihui Liu, Dian Li, Ying Shan, Xiaohu Qie, and Ping Luo. Bridging video-text retrieval with multiple choice questions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16167-16176, 2022. 1, 2, 5, 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 680, + 288, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 680, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 48, + 680, + 288, + 714 + ], + "type": "text", + "content": "[12] Yuying Ge, Yixiao Ge, Xihui Liu, Alex Jinpeng Wang, Jianping Wu, Ying Shan, Xiaohu Qie, and Ping Luo. Miles: Visual bert pre-training with injected language semantics for" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 712 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "text", + "content": "video-text retrieval. Proceedings of the European Conference on Computer Vision (ECCV), 2022. 6" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 96, + 547, + 148 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 96, + 547, + 148 + ], + "spans": [ + { + "bbox": [ + 307, + 96, + 547, + 148 + ], + "type": "text", + "content": "[13] Simon Ging, Mohammadreza Zolfaghari, Hamed Pirsiavash, and Thomas Brox. Coot: Cooperative hierarchical transformer for video-text representation learning. Advances in neural information processing systems, 33:22605-22618, 2020. 1, 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 150, + 547, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 150, + 547, + 215 + ], + "spans": [ + { + "bbox": [ + 307, + 150, + 547, + 215 + ], + "type": "text", + "content": "[14] Satya Krishna Gorti, Noel Vouitsis, Junwei Ma, Keyvan Golestan, Maksims Volkovs, Animesh Garg, and Guangwei Yu. X-pool: Cross-modal language-video attention for text-video retrieval. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5006-5015, 2022. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 217, + 547, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 217, + 547, + 271 + ], + "spans": [ + { + "bbox": [ + 307, + 217, + 547, + 271 + ], + "type": "text", + "content": "[15] Deng Huang, Peihao Chen, Runhao Zeng, Qing Du, Mingkui Tan, and Chuang Gan. Location-aware graph convolutional networks for video question answering. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 34, pages 11021-11028, 2020. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 272, + 547, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 272, + 547, + 326 + ], + "spans": [ + { + "bbox": [ + 307, + 272, + 547, + 326 + ], + "type": "text", + "content": "[16] Yunseok Jang, Yale Song, Youngjae Yu, Youngjin Kim, and Gunhee Kim. Tgif-qa: Toward spatio-temporal reasoning in visual question answering. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2758–2766, 2017. 1, 2, 3, 5" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 327, + 545, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 327, + 545, + 370 + ], + "spans": [ + { + "bbox": [ + 307, + 327, + 545, + 370 + ], + "type": "text", + "content": "[17] Jie Jiang, Shaobo Min, Weijie Kong, Hongfa Wang, Zhifeng Li, and Wei Liu. Tencent text-video retrieval: Hierarchical cross-modal interactions with multi-level representations. IEEE Access, 2022. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 372, + 545, + 425 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 372, + 545, + 425 + ], + "spans": [ + { + "bbox": [ + 307, + 372, + 545, + 425 + ], + "type": "text", + "content": "[18] Will Kay, Joao Carreira, Karen Simonyan, Brian Zhang, Chloe Hillier, Sudheendra Vijayanarasimhan, Fabio Viola, Tim Green, Trevor Back, Paul Natsev, et al. The kinetics human action video dataset. arXiv preprint arXiv:1705.06950, 2017. 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 426, + 547, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 426, + 547, + 480 + ], + "spans": [ + { + "bbox": [ + 307, + 426, + 547, + 480 + ], + "type": "text", + "content": "[19] Junyeong Kim, Minuk Ma, Kyungsu Kim, Sungjin Kim, and Chang D Yoo. Progressive attention memory network for movie story question answering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8337-8346, 2019. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 482, + 545, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 482, + 545, + 547 + ], + "spans": [ + { + "bbox": [ + 307, + 482, + 545, + 547 + ], + "type": "text", + "content": "[20] Ranjay Krishna, Yuke Zhu, Oliver Groth, Justin Johnson, Kenji Hata, Joshua Kravitz, Stephanie Chen, Yannis Kalantidis, Li-Jia Li, David A Shamma, et al. Visual genome: Connecting language and vision using crowdsourced dense image annotations. International journal of computer vision, 123(1):32-73, 2017. 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 548, + 545, + 592 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 548, + 545, + 592 + ], + "spans": [ + { + "bbox": [ + 307, + 548, + 545, + 592 + ], + "type": "text", + "content": "[21] Alexander Kunitsyn, Maksim Kalashnikov, Maksim Dzabraev, and Andrei Ivaniuta. Mdmt-2: Multidomain multimodal transformer for video retrieval, one more step towards generalization. arXiv preprint arXiv:2203.07086, 2022. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 593, + 545, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 593, + 545, + 646 + ], + "spans": [ + { + "bbox": [ + 307, + 593, + 545, + 646 + ], + "type": "text", + "content": "[22] Thao Minh Le, Vuong Le, Svetha Venkatesh, and Truyen Tran. Hierarchical conditional relation networks for video question answering. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9972-9981, 2020. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 647, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 647, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 307, + 647, + 545, + 712 + ], + "type": "text", + "content": "[23] Jie Lei, Linjie Li, Luowei Zhou, Zhe Gan, Tamara L Berg, Mohit Bansal, and Jingjing Liu. Less is more: Clipbert for video-and-language learning via sparse sampling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7331-7341, 2021. 1, 2, 5, 6, 7" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "text", + "content": "14864" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 128 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 128 + ], + "type": "text", + "content": "[24] Dongxu Li, Junnan Li, Hongdong Li, Juan Carlos Niebles, and Steven CH Hoi. Align and prompt: Video-and-language pre-training with entity prompts. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4953-4963, 2022. 1, 2, 5, 6, 7" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 130, + 288, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 130, + 288, + 185 + ], + "spans": [ + { + "bbox": [ + 48, + 130, + 288, + 185 + ], + "type": "text", + "content": "[25] Linjie Li, Yen-Chun Chen, Yu Cheng, Zhe Gan, Licheng Yu, and Jingjing Liu. Hero: Hierarchical encoder for video+ language omni-representation pre-training. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 2046-2065, 2020. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 186, + 288, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 186, + 288, + 252 + ], + "spans": [ + { + "bbox": [ + 48, + 186, + 288, + 252 + ], + "type": "text", + "content": "[26] Xiangpeng Li, Jingkuan Song, Lianli Gao, Xianglong Liu, Wenbing Huang, Xiangnan He, and Chuang Gan. Beyond rnns: Positional self-attention with co-attention for video question answering. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 33, pages 8658–8665, 2019. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 255, + 287, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 255, + 287, + 308 + ], + "spans": [ + { + "bbox": [ + 48, + 255, + 287, + 308 + ], + "type": "text", + "content": "[27] Chengzhi Lin, Ancong Wu, Junwei Liang, Jun Zhang, Wenhang Ge, Wei-Shi Zheng, and Chunhua Shen. Text-adaptive multiple visual prototype matching for video-text retrieval. Advances in Neural Information Processing Systems, 2022. 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 312, + 287, + 356 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 312, + 287, + 356 + ], + "spans": [ + { + "bbox": [ + 48, + 312, + 287, + 356 + ], + "type": "text", + "content": "[28] Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dólár. Focal loss for dense object detection. In Proceedings of the IEEE international conference on computer vision, pages 2980-2988, 2017. 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 357, + 287, + 412 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 357, + 287, + 412 + ], + "spans": [ + { + "bbox": [ + 48, + 357, + 287, + 412 + ], + "type": "text", + "content": "[29] Fei Liu, Jing Liu, Weining Wang, and Hanqing Lu. Hair: Hierarchical visual-semantic relational reasoning for video question answering. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1698–1707, 2021. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 415, + 287, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 415, + 287, + 459 + ], + "spans": [ + { + "bbox": [ + 48, + 415, + 287, + 459 + ], + "type": "text", + "content": "[30] Ruyang Liu, Jingjia Huang, Ge Li, Jiashi Feng, Xinglong Wu, and Thomas H Li. Revisiting temporal modeling for clip-based image-to-video knowledge transferring. arXiv preprint arXiv:2301.11116, 2023. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 460, + 287, + 504 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 460, + 287, + 504 + ], + "spans": [ + { + "bbox": [ + 48, + 460, + 287, + 504 + ], + "type": "text", + "content": "[31] Yang Liu, Samuel Albanie, Arsha Nagrani, and Andrew Zisserman. Use what you have: Video retrieval using representations from collaborative experts. In BMVC, 2019. 1, 2, 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 506, + 287, + 551 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 506, + 287, + 551 + ], + "spans": [ + { + "bbox": [ + 48, + 506, + 287, + 551 + ], + "type": "text", + "content": "[32] Ze Liu, Jia Ning, Yue Cao, Yixuan Wei, Zheng Zhang, Stephen Lin, and Han Hu. Video swim transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3202-3211, 2022. 3, 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 552, + 287, + 574 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 552, + 287, + 574 + ], + "spans": [ + { + "bbox": [ + 48, + 552, + 287, + 574 + ], + "type": "text", + "content": "[33] Edward Loper and Steven Bird. Nltk: The natural language toolkit. arXiv preprint cs/0205028, 2002. 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 577, + 287, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 577, + 287, + 609 + ], + "spans": [ + { + "bbox": [ + 48, + 577, + 287, + 609 + ], + "type": "text", + "content": "[34] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In International Conference on Learning Representations, 2018. 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 612, + 287, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 612, + 287, + 666 + ], + "spans": [ + { + "bbox": [ + 48, + 612, + 287, + 666 + ], + "type": "text", + "content": "[35] Huaishao Luo, Lei Ji, Botian Shi, Haoyang Huang, Nan Duan, Tianrui Li, Jason Li, Taroon Bharti, and Ming Zhou. Univl: A unified video and language pre-training model for multimodal understanding and generation. arXiv preprint arXiv:2002.06353, 2020. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 669, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 669, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 669, + 287, + 712 + ], + "type": "text", + "content": "[36] Huaishao Luo, Lei Ji, Ming Zhong, Yang Chen, Wen Lei, Nan Duan, and Tianrui Li. Clip4clip: An empirical study of clip for end to end video clip retrieval. arXiv preprint arXiv:2104.08860, 2021. 6" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 72, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 24, + "blocks": [ + { + "bbox": [ + 307, + 72, + 545, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 72, + 545, + 128 + ], + "spans": [ + { + "bbox": [ + 307, + 72, + 545, + 128 + ], + "type": "text", + "content": "[37] Yiwei Ma, Guohai Xu, Xiaoshuai Sun, Ming Yan, Ji Zhang, and Rongrong Ji. X-clip: End-to-end multi-grained contrastive learning for video-text retrieval. In Proceedings of the 30th ACM International Conference on Multimedia, pages 638–647, 2022. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 130, + 545, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 130, + 545, + 194 + ], + "spans": [ + { + "bbox": [ + 307, + 130, + 545, + 194 + ], + "type": "text", + "content": "[38] Tegan Maharaj, Nicolas Ballas, Anna Rohrbach, Aaron Courville, and Christopher Pal. A dataset and exploration of models for understanding video data through fill-in-the-blank question-answering. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 6884–6893, 2017. 1, 5, 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 197, + 545, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 197, + 545, + 262 + ], + "spans": [ + { + "bbox": [ + 307, + 197, + 545, + 262 + ], + "type": "text", + "content": "[39] Antoine Miech, Jean-Baptiste Alayrac, Lucas Smaira, Ivan Laptev, Josef Sivic, and Andrew Zisserman. End-to-end learning of visual representations from uncurated instructional videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9879-9889, 2020. 1, 2, 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 264, + 545, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 264, + 545, + 330 + ], + "spans": [ + { + "bbox": [ + 307, + 264, + 545, + 330 + ], + "type": "text", + "content": "[40] Antoine Miech, Dimitri Zhukov, Jean-Baptiste Alayrac, Makarand Tapaswi, Ivan Laptev, and Josef Sivic. Howto100m: Learning a text-video embedding by watching hundred million narrated video clips. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2630-2640, 2019. 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 332, + 545, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 332, + 545, + 386 + ], + "spans": [ + { + "bbox": [ + 307, + 332, + 545, + 386 + ], + "type": "text", + "content": "[41] Jungin Park, Jiyoung Lee, and Kwanghoon Sohn. Bridge to answer: Structure-aware graph interaction network for video question answering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15526-15535, 2021. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 388, + 545, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 388, + 545, + 464 + ], + "spans": [ + { + "bbox": [ + 307, + 388, + 545, + 464 + ], + "type": "text", + "content": "[42] Andrew Rouditchenko, Angie Boggust, David Harwath, Brian Chen, Dhiraj Joshi, Samuel Thomas, Kartik Audhkhasi, Hilde Kuehne, Rameswar Panda, Rogerio Feris, et al. Avlnet: Learning audio-visual language representations from instructional videos. In Annual Conference of the International Speech Communication Association. International Speech Communication Association, 2021. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 467, + 545, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 467, + 545, + 520 + ], + "spans": [ + { + "bbox": [ + 307, + 467, + 545, + 520 + ], + "type": "text", + "content": "[43] Paul Hongsuck Seo, Arsha Nagrani, and Cordelia Schmid. Look before you speak: Visually contextualized utterances. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16877-16887, 2021. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 523, + 545, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 523, + 545, + 589 + ], + "spans": [ + { + "bbox": [ + 307, + 523, + 545, + 589 + ], + "type": "text", + "content": "[44] Piyush Sharma, Nan Ding, Sebastian Goodman, and Radu Soricut. Conceptual captions: A cleaned, hypernymed, image alt-text dataset for automatic image captioning. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 2556-2565, 2018. 5, 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 591, + 545, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 591, + 545, + 645 + ], + "spans": [ + { + "bbox": [ + 307, + 591, + 545, + 645 + ], + "type": "text", + "content": "[45] Chen Sun, Austin Myers, Carl Vondrick, Kevin Murphy, and Cordelia Schmid. Videobert: A joint model for video and language representation learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7464-7473, 2019. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 647, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 647, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 307, + 647, + 545, + 712 + ], + "type": "text", + "content": "[46] Hao Tan and Mohit Bansal. Lxmert: Learning cross-modality encoder representations from transformers. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages 5100-5111, 2019. 2" + } + ] + } + ], + "index": 23 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "14865" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 115 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 115 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 115 + ], + "type": "text", + "content": "[47] Atousa Torabi, Niket Tandon, and Leonid Sigal. Learning language-visual embedding for movie understanding with natural-language. arXiv preprint arXiv:1609.08124, 2016. 1, 5" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 118, + 287, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 118, + 287, + 149 + ], + "spans": [ + { + "bbox": [ + 48, + 118, + 287, + 149 + ], + "type": "text", + "content": "[48] Aaron Van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding. arXiv e-prints, pages arXiv-1807, 2018. 2, 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 151, + 287, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 151, + 287, + 205 + ], + "spans": [ + { + "bbox": [ + 48, + 151, + 287, + 205 + ], + "type": "text", + "content": "[49] Alex Jinpeng Wang, Yixiao Ge, Rui Yan, Yuying Ge, Xudong Lin, Guanyu Cai, Jianping Wu, Ying Shan, Xiaohu Qie, and Mike Zheng Shou. All in one: Exploring unified video-language pre-training. arXiv preprint arXiv:2203.07303, 2022. 1, 2, 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 206, + 287, + 260 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 206, + 287, + 260 + ], + "spans": [ + { + "bbox": [ + 48, + 206, + 287, + 260 + ], + "type": "text", + "content": "[50] Jinpeng Wang, Yixiao Ge, Guanyu Cai, Rui Yan, Xudong Lin, Ying Shan, Xiaohu Qie, and Mike Zheng Shou. Object-aware video-language pre-training for retrieval. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3313-3322, 2022. 2, 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 262, + 287, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 262, + 287, + 315 + ], + "spans": [ + { + "bbox": [ + 48, + 262, + 287, + 315 + ], + "type": "text", + "content": "[51] Yi Wang, Kunchang Li, Yizhuo Li, Yinan He, Bingkun Huang, Zhiyu Zhao, Hongjie Zhang, Jilan Xu, Yi Liu, Zun Wang, et al. Internvideo: General video foundation models via generative and discriminative learning. arXiv preprint arXiv:2212.03191, 2022. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 317, + 287, + 371 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 317, + 287, + 371 + ], + "spans": [ + { + "bbox": [ + 48, + 317, + 287, + 371 + ], + "type": "text", + "content": "[52] Dejing Xu, Zhou Zhao, Jun Xiao, Fei Wu, Hanwang Zhang, Xiangnan He, and Yueting Zhuang. Video question answering via gradually refined attention over appearance and motion. In Proceedings of the 25th ACM international conference on Multimedia, pages 1645-1653, 2017. 1, 2, 5, 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 372, + 287, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 372, + 287, + 437 + ], + "spans": [ + { + "bbox": [ + 48, + 372, + 287, + 437 + ], + "type": "text", + "content": "[53] Hu Xu, Gargi Ghosh, Po-Yao Huang, Prahal Arora, Masoumeh Aminzadeh, Christoph Feichtenhofer, Florian Metze, and Luke Zettlemoyer. Vlm: Task-agnostic videolanguage model pre-training for video understanding. In Findings of the Association for Computational Linguistics: ACL-IJCNLP 2021, pages 4227–4239, 2021. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 438, + 287, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 438, + 287, + 503 + ], + "spans": [ + { + "bbox": [ + 48, + 438, + 287, + 503 + ], + "type": "text", + "content": "[54] Hu Xu, Gargi Ghosh, Po-Yao Huang, Dmytro Okhonko, Armen Aghajanyan, Florian Metze, Luke Zettlemoyer, and Christoph Feichtenhofer. Videoclip: Contrastive pre-training for zero-shot video-text understanding. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 6787-6800, 2021. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 505, + 287, + 548 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 505, + 287, + 548 + ], + "spans": [ + { + "bbox": [ + 48, + 505, + 287, + 548 + ], + "type": "text", + "content": "[55] Jun Xu, Tao Mei, Ting Yao, and Yong Rui. Msr-vtt: A large video description dataset for bridging video and language. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5288-5296, 2016. 1, 5, 6, 8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 548, + 287, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 548, + 287, + 613 + ], + "spans": [ + { + "bbox": [ + 48, + 548, + 287, + 613 + ], + "type": "text", + "content": "[56] Hongwei Xue, Tiankai Hang, Yanhong Zeng, Yuchong Sun, Bei Liu, Huan Yang, Jianlong Fu, and Baining Guo. Advancing high-resolution video-language representation with large-scale video transcriptions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5036-5045, June 2022. 1, 2, 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 615, + 287, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 615, + 287, + 658 + ], + "spans": [ + { + "bbox": [ + 48, + 615, + 287, + 658 + ], + "type": "text", + "content": "[57] Hongwei Xue, Yuchong Sun, Bei Liu, Jianlong Fu, Ruihua Song, Houqiang Li, and Jiebo Luo. Clip-vip: Adapting pretrained image-text model to video-language representation alignment. arXiv preprint arXiv:2209.06430, 2022. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 659, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 659, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 659, + 287, + 713 + ], + "type": "text", + "content": "[58] Antoine Yang, Antoine Miech, Josef Sivic, Ivan Laptev, and Cordelia Schmid. Just ask: Learning to answer questions from millions of narrated videos. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1686-1697, 2021. 1, 2, 6, 7" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 546, + 318 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 307, + 73, + 546, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 546, + 127 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 546, + 127 + ], + "type": "text", + "content": "[59] Jianwei Yang, Yonatan Bisk, and Jianfeng Gao. Taco: Token-aware cascade contrastive learning for video-text alignment. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11562-11572, 2021. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 307, + 129, + 546, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 129, + 546, + 173 + ], + "spans": [ + { + "bbox": [ + 307, + 129, + 546, + 173 + ], + "type": "text", + "content": "[60] Youngjae Yu, Jongseok Kim, and Gunhee Kim. A joint sequence fusion model for video question answering and retrieval. In Proceedings of the European Conference on Computer Vision (ECCV), pages 471-487, 2018. 3, 5" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 175, + 546, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 175, + 546, + 217 + ], + "spans": [ + { + "bbox": [ + 307, + 175, + 546, + 217 + ], + "type": "text", + "content": "[61] Rowan Zellers, Ximing Lu, Jack Hessel, Youngjae Yu, Jae Sung Park, Jize Cao, Ali Farhadi, and Yejin Choi. Merlot: Multimodal neural script knowledge models. Advances in Neural Information Processing Systems, 34, 2021. 1, 6, 7" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 219, + 546, + 273 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 219, + 546, + 273 + ], + "spans": [ + { + "bbox": [ + 307, + 219, + 546, + 273 + ], + "type": "text", + "content": "[62] Zheng-Jun Zha, Jiawei Liu, Tianhao Yang, and Yongdong Zhang. Spatiotemporal-textual co-attention network for video question answering. ACM Transactions on Multimedia Computing, Communications, and Applications (TOMM), 15(2s):1-18, 2019. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 275, + 546, + 318 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 275, + 546, + 318 + ], + "spans": [ + { + "bbox": [ + 308, + 275, + 546, + 318 + ], + "type": "text", + "content": "[63] Linchao Zhu and Yi Yang. Actbert: Learning global-local video-text representations. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8746-8755, 2020. 2" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "text", + "content": "14866" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Co-SLAM_ Joint Coordinate and Sparse Parametric Encodings for Neural Real-Time SLAM/98d936fc-c445-4925-9f16-8d4e0f8fc517_content_list.json b/2023/Co-SLAM_ Joint Coordinate and Sparse Parametric Encodings for Neural Real-Time SLAM/98d936fc-c445-4925-9f16-8d4e0f8fc517_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..864c69ba8edb1b1cdb87212ee6ce736e59ec367d --- /dev/null +++ b/2023/Co-SLAM_ Joint Coordinate and Sparse Parametric Encodings for Neural Real-Time SLAM/98d936fc-c445-4925-9f16-8d4e0f8fc517_content_list.json @@ -0,0 +1,1499 @@ +[ + { + "type": "text", + "text": "Co-SLAM: Joint Coordinate and Sparse Parametric Encodings for Neural Real-Time SLAM", + "text_level": 1, + "bbox": [ + 148, + 130, + 821, + 174 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Hengyi Wang* Jingwen Wang* Lourdes Agapito Department of Computer Science, University College London", + "bbox": [ + 240, + 191, + 728, + 227 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{hengyi.wang.21, jingwen.wang.17, 1.agapito}@ucl.ac.uk", + "bbox": [ + 243, + 229, + 720, + 244 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/8111988349f3fd4a7c740f7c31b3fb88b8c910219b7fa1c4c1bb67a412bb2096.jpg", + "image_caption": [ + "Figure 1. We present Co-SLAM, a neural RGB-D SLAM method that performs online tracking and mapping in real time. We propose a new hybrid representation based on a joint coordinate and sparse-parametric encoding with global bundle adjustment. Our method shows fast, high-fidelity scene reconstruction with efficient memory use and plausible hole-filling." + ], + "image_footnote": [], + "bbox": [ + 106, + 263, + 870, + 433 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 501, + 313, + 516 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We present Co-SLAM, a neural RGB-D SLAM system based on a hybrid representation, that performs robust camera tracking and high-fidelity surface reconstruction in real time. Co-SLAM represents the scene as a multi-resolution hash-grid to exploit its high convergence speed and ability to represent high-frequency local features. In addition, Co-SLAM incorporates one-blob encoding, to encourage surface coherence and completion in unobserved areas. This joint parametric-coordinate encoding enables real-time and robust performance by bringing the best of both worlds: fast convergence and surface hole filling. Moreover, our ray sampling strategy allows Co-SLAM to perform global bundle adjustment over all keyframes instead of requiring keyframe selection to maintain a small number of active keyframes as competing neural SLAM approaches do. Experimental results show that Co-SLAM runs at $10 - 17\\mathrm{Hz}$ and achieves state-of-the-art scene reconstruction results, and competitive tracking performance in various datasets and benchmarks (ScanNet, TUM, Replica, Synthetic RGBD). Project page: https://hengyiwang.github.io/projects/CoSLAM", + "bbox": [ + 75, + 535, + 473, + 853 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 501, + 501, + 630, + 516 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Real-time joint camera tracking and dense surface reconstruction from RGB-D sensors has been a core problem in computer vision and robotics for decades. Traditional SLAM solutions exist that can robustly track the position of the camera while fusing depth and/or color measurements into a single high-fidelity map. However, they rely on handcrafted loss terms and do not exploit data-driven priors.", + "bbox": [ + 496, + 530, + 892, + 636 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recent attention has turned to learning-based models that can exploit the ability of neural network architectures to learn smoothness and coherence priors directly from data. Coordinate-based networks have probably become the most popular representation, since they can be trained to predict the geometric and appearance properties of any point in the scene in a self-supervised way, directly from images. The most notable example, Neural Radiance Fields (NeRF) [14], encodes scene density and color in the weights of a neural network. In combination with volume rendering, NeRF is trained to re-synthesize the input images and has a remarkable ability to generalize to nearby unseen views.", + "bbox": [ + 496, + 638, + 893, + 820 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Coordinate-based networks embed input point coordinates into a high dimensional space, using sinusoidal or other frequency embeddings, allowing them to capture high-frequency details that are essential for high-fidelity geometry reconstruction [1]. Combined with the smooth-", + "bbox": [ + 496, + 824, + 893, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 44 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "$\\star$ Indicates equal contribution.", + "bbox": [ + 101, + 887, + 264, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "13293", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/2ba0bddad9beabc2c6bae49e057106ca11f3115b09d5938ba98b1af9e30ddece.jpg", + "image_caption": [ + "COORDINATE" + ], + "image_footnote": [], + "bbox": [ + 135, + 101, + 305, + 200 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/a9c1a9797c2c88598b833a18c8e4d45d6f2cd2627a7b9dca26ffd6a61b4d388a.jpg", + "image_caption": [ + "PARAMETRIC" + ], + "image_footnote": [], + "bbox": [ + 308, + 101, + 480, + 200 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/0c45c4169fc7c1832ce4cae4d3488c083aa0c7db5c645185826166ae26e71a7a.jpg", + "image_caption": [ + "JOINT", + "DenseGrid+OneBlob" + ], + "image_footnote": [], + "bbox": [ + 483, + 101, + 656, + 200 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/f26fe639dca6d1f67e9f783296c256cbcb248cdaa268a2738a784e729607d2bc.jpg", + "image_caption": [ + "REFERENCE" + ], + "image_footnote": [], + "bbox": [ + 658, + 101, + 831, + 200 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/11ad094643c600a866f02ffbc44b7151dac17206dd1ef0be1c412f1d17995a72.jpg", + "image_caption": [ + "Frequency", + "OneBlob", + "Figure 2. Illustration of the effect of different encodings on completion. COORDINATE based encodings achieve hole filling but require long training times. PARAMETRIC encodings allow fast training, but fail to complete unobserved regions. JOINT coordinate and parametric encoding (Ours) allows smooth scene completion and fast training. NICE-SLAM [42] uses a dense parametric encoding." + ], + "image_footnote": [], + "bbox": [ + 135, + 215, + 307, + 315 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/b1f72d2dd45e1ccd0da4c20c3931a3f2b9f639752551be2345924165f2f077f0.jpg", + "image_caption": [ + "DenseGrid", + "Grid" + ], + "image_footnote": [], + "bbox": [ + 310, + 215, + 480, + 315 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/432fdebd8001049624c51a4fdd2c3e0b09b5e01ad677c0e5c50cafe7bb9c1449.jpg", + "image_caption": [ + "HashGrid+OneBlob (Ours)" + ], + "image_footnote": [], + "bbox": [ + 483, + 215, + 656, + 315 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/70d111e0030d4375e1d7c7cc86a9a191b1865b3d1ff7c4b4a4b1cb05793338f4.jpg", + "image_caption": [ + "NICE-SLAM [42]", + "GT" + ], + "image_footnote": [], + "bbox": [ + 658, + 215, + 831, + 315 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "ness and coherence priors inherently encoded in the MLP weights, they constitute a good choice for sequential tracking and mapping [26]. However, the weakness of MLP-based approaches is the long training times required (sometimes hours) to learn a single scene. For that reason, recent real-time capable SLAM systems built on coordinate networks with frequency embeddings such as iMAP [26] need to resort to strategies to sparsify ray sampling and reduce tracking iterations to maintain interactive operation. This comes at the cost of loss of detail in the reconstructions which are oversmoothed (Fig. 5) and potential errors in camera tracking.", + "bbox": [ + 75, + 393, + 468, + 575 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Optimizable feature grids, also known as parametric embeddings, have recently become a powerful alternative scene representation to monolithic MLPs, given their ability to represent high-fidelity local features and their extremely fast convergence (orders of magnitude faster) [7, 10, 15, 32, 40]. Recent efforts focus on sparse alternatives to these parametric embeddings such as octrees [28], tri-plane [2], hash-grid [15] or sparse voxel grid [12, 13] to improve the memory efficiency of dense grids. While these representations can be fast to train and are therefore well suited to real-time operation, they fundamentally lack the smoothness, and coherence priors inherent to MLPs and struggle with hole-filling in areas without observation. NICE-SLAM [42] is a recent example of a multi-resolution feature grid-based SLAM method. Although it does not suffer from over-smoothness and captures local detail (as shown in Fig. 2) it cannot perform hole-filling which might in turn lead to drift in camera pose estimation.", + "bbox": [ + 75, + 579, + 468, + 851 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our first contribution is to design a joint coordinate and sparse grid encoding for input points that brings together the benefits of both worlds to the real-time SLAM frame-", + "bbox": [ + 75, + 854, + 468, + 898 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "work. On the one hand, the smoothness and coherence priors provided by coordinate encodings (we use one-blob [16] encoding), and on the other hand the optimization speed and local details of sparse feature encodings (we use hash grid [15]), resulting in more robust camera tracking and high-fidelity maps with better completion and hole filling.", + "bbox": [ + 496, + 393, + 890, + 484 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our second contribution relates to the bundle adjustment (BA) step in the joint optimization of the map and camera poses. So far, all neural SLAM systems [26, 42] perform BA using rays sampled from a very small subset of selected keyframes. Restricting the optimization to a very small number of viewpoints results in decreased robustness in camera tracking and increased computation due to the need for a keyframe-selection strategy. Instead, Co-SLAM performs global BA, sampling rays from all past keyframes, which results in an important boost in robustness and performance in pose estimation. In addition, we show that our BA optimization requires a fraction of the iterations of NICE-SLAM [42] to achieve similar errors. In practice, Co-SLAM achieves SOTA performance in camera tracking and 3D reconstruction while maintaining real time performance.", + "bbox": [ + 496, + 484, + 892, + 710 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Co-SLAM runs at 15-17Hz on Replica and Synthetic RGB-D datasets [1], and 12-13Hz on ScanNet [5] and TUM [25] scenes — faster than NICE-SLAM (0.1-1Hz) [42] and iMAP [26]. We perform extensive evaluations on various datasets (Replica [24], Synthetic RGBD [1], ScanNet [6], TUM [25]) where we outperform NICE-SLAM [42] and iMAP [26] in reconstruction and achieve better or at least on-par tracking accuracy.", + "bbox": [ + 496, + 710, + 890, + 832 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 844, + 640, + 859 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Dense Visual SLAM. Taking advantage of commodity depth sensors, KinectFusion [18] performs frame-to-model", + "bbox": [ + 500, + 869, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "13294", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "camera tracking via projective iterative-closest-point (ICP), and incrementally updates the scene geometry via TSDF-Fusion. Subsequent works focused on addressing the scalability issue by adopting more efficient data structures, such as surface elements (surfels) [35, 36], VoxelHashing [3, 9, 19] or Octrees [31, 41]. While most works focus more heavily on scene reconstruction and only track per-frame poses, BAD-SLAM [21] proposes full direct bundle adjustment (BA) to jointly optimize keyframe (KF) poses and the dense 3D structure. Several recent works [11, 29, 30, 38] leverage deep learning to improve the accuracy and robustness of traditional SLAM, and even achieve dense reconstruction with monocular SLAM. While these approaches introduced some learned components, the scene representation and overall pipeline still follow traditional SLAM methods.", + "bbox": [ + 75, + 90, + 472, + 318 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Neural Implicit Representations. Recently neural implicit representations [14] that encode 3D geometry and the appearance of a scene within the weights of a neural network have gained popularity due to their expressiveness and compactness. Among these works, NeRF [14] and its variants adopt coordinate encoding [16, 20] with MLPs and show impressive scene reconstruction using differentiable rendering. Given that coordinate encoding-based methods require lengthy training, many follow-up works [7, 10, 27] propose parametric encodings that increase parameter size but speed up the training. To improve the memory efficiency of parametric encoding-based methods, sparse parametric encodings, such as Octree [28], Tri-plane [2], or sparse voxel grid [12, 13, 15], have been proposed. While these methods focus on novel view synthesis, others instead focus on surface reconstruction from RGB images, with implicit surface representations and differentiable renderers [8, 33, 39, 40]. Other methods [1, 26, 32, 37, 42] use depth measurements as additional supervision for surface reconstruction.", + "bbox": [ + 75, + 329, + 472, + 617 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Neural Implicit SLAM. iMAP [26] adopts an MLP representation to perform joint tracking and mapping in quasi-real time. Smooth, plausible filling of unobserved regions is achieved thanks to the coherence priors inherent to the MLP. iMAP introduces elaborate keyframe selection and information-guided pixel sampling for speed, resulting in $10\\mathrm{Hz}$ tracking and $2\\mathrm{Hz}$ mapping. To reduce the computational overhead and improve scalability, NICE-SLAM [42] adopts a multi-level feature grid for scene representation. However, as feature grids only perform local updates, they fail to achieve plausible hole-filling. With Co-SLAM we aim to address both issues. To achieve real-time performance and memory-efficiency, while maintaining high-fidelity surface reconstruction and plausible hole filling, we propose to combine the use of coordinate and sparse parametric encodings for scene representation, and perform dense global bundle adjustment using rays sampled from all keyframes.", + "bbox": [ + 75, + 628, + 472, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 500, + 89, + 591, + 106 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Fig. 3 shows an overview of Co-SLAM. Given an input RGB-D stream $\\{I_t\\}_{t=1}^N \\{D_t\\}_{t=1}^N$ with known camera intrinsics $K \\in \\mathbb{R}^{3 \\times 3}$ , we perform dense mapping and tracking by jointly optimizing camera poses $\\{\\xi_t\\}_{t=1}^N$ and a neural scene representation $f_\\theta$ . Specifically, our implicit representation maps world coordinates $\\mathbf{x}$ into color $\\mathbf{c}$ and truncated signed distance (TSDF) $s$ values:", + "bbox": [ + 496, + 114, + 893, + 220 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nf _ {\\theta} (\\mathbf {x}) \\mapsto (\\mathbf {c}, s). \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 640, + 229, + 890, + 247 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Similar to most SLAM systems, the process is split into tracking and mapping. Initialization is performed by running a few training iterations on the first frame. For each subsequent frame, the camera pose is optimized first, initialized with a simple constant-speed motion model. A small fraction of pixels/rays are then sampled and copied to the global pixel-set. At each mapping iteration, global bundle adjustment is performed over a set of pixels randomly sampled from the global pixel-set, to jointly optimize the scene representation $\\theta$ and all camera poses $\\{\\xi_t\\}$ .", + "bbox": [ + 496, + 255, + 893, + 407 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Joint Coordinate and Parametric Encoding", + "text_level": 1, + "bbox": [ + 498, + 414, + 870, + 431 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Thanks to the coherence and smoothness priors inherent to MLPs, coordinate-based representations achieve high-fidelity scene reconstruction. However, these methods often suffer from slow convergence and catastrophic forgetting, when optimized in a sequential setting. Instead, parametric encoding based methods improve the computational efficiency, but they fall short of hole filling and smoothness. Since both properties of speed and coherence are crucial for a real-world SLAM system, we propose a joint coordinate and parametric encoding that combines the best of both worlds: we adopt coordinate encoding for scene representation while using sparse parametric encoding to speed up training. Specifically, we use One-blob [16] encoding $\\gamma (\\mathbf{x})$ instead of embedding spatial coordinates into multiple frequency bands. As scene representation we adopt a multiresolution hash-based feature grid [15] $\\mathcal{V}_{\\alpha} = \\{\\mathcal{V}_{\\alpha}^{l}\\}_{l = 1}^{L}$ . The spatial resolution of each level is set between the coarsest $R_{min}$ and the finest resolution $R_{max}$ in a progressive manner. Feature vectors $\\mathcal{V}_{\\alpha}(\\mathbf{x})$ at each sampled point $\\mathbf{x}$ are queried via trilinear interpolation. The geometry decoder outputs the predicted SDF value $s$ and a feature vector $\\mathbf{h}$ :", + "bbox": [ + 496, + 438, + 893, + 755 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nf _ {\\tau} (\\gamma (\\mathbf {x}), \\mathcal {V} _ {\\alpha} (\\mathbf {x})) \\mapsto (\\mathbf {h}, s). \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 602, + 763, + 890, + 781 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Finally, the color MLP predicts the RGB value:", + "bbox": [ + 500, + 789, + 815, + 805 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nf _ {\\phi} (\\gamma (\\mathbf {x}), \\mathbf {h}) \\mapsto \\mathbf {c}. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 633, + 814, + 890, + 830 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Here $\\theta = \\{\\alpha, \\phi, \\tau\\}$ are learnable parameters. Injecting the One-blob encoding in the hash-based multi-resolution feature grid representation, results in fast convergence, efficient memory use, and hole filling needed for online SLAM.", + "bbox": [ + 496, + 839, + 893, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "13295", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/a415353fd7bfe788cf3dc708a6db1a6c53309e215917109d95c500cd4bebdb26.jpg", + "image_caption": [ + "Figure 3. Overview of Co-SLAM. 1) Scene representation: using our new joint coordinate+parametric encoding, input coordinates are mapped to RGB and SDF values via two shallow MLPs. 2) Tracking: optimize per-frame camera poses $\\xi_{t}$ by minimizing losses. 3) Mapping: global bundle adjustment to jointly optimize the scene representation and camera poses taking rays sampled from all keyframes." + ], + "image_footnote": [], + "bbox": [ + 98, + 88, + 872, + 335 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Depth and Color Rendering", + "text_level": 1, + "bbox": [ + 76, + 395, + 328, + 412 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Following [26, 42], we render depth and color by integrating the predicted values along the sampled rays. Specifically, given the camera origin $\\mathbf{o}$ and ray direction $\\mathbf{r}$ , we uniformly sample $M$ points $\\mathbf{x}_i = \\mathbf{o} + d_i\\mathbf{r}$ , $i \\in \\{1, \\dots, M\\}$ with depth values $\\{t_1, \\dots, t_M\\}$ and predicted colors $\\{\\mathbf{c}_1, \\dots, \\mathbf{c}_M\\}$ and render color and depth as", + "bbox": [ + 76, + 417, + 470, + 511 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathbf {c}} = \\frac {1}{\\sum_ {i = 1} ^ {M} w _ {i}} \\sum_ {i = 1} ^ {M} w _ {i} \\mathbf {c} _ {i}, \\quad \\hat {d} = \\frac {1}{\\sum_ {i = 1} ^ {M} w _ {i}} \\sum_ {i = 1} ^ {M} w _ {i} d _ {i}, \\quad (4)\n$$\n", + "text_format": "latex", + "bbox": [ + 91, + 518, + 470, + 561 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\{w_i\\}$ are the computed weights along the ray. A conversion function is needed to convert predicted SDF $s_i$ to weight $w_i$ . Instead of adopting the rendering equations proposed in Neus [33, 39], we follow the simple bell-shaped model of [1] and compute weights $w_i$ directly by multiplying the two Sigmoid functions $\\sigma(\\cdot)$", + "bbox": [ + 75, + 569, + 468, + 660 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nw _ {i} = \\sigma \\left(\\frac {s _ {i}}{t r}\\right) \\sigma \\left(- \\frac {s _ {i}}{t r}\\right), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 189, + 667, + 468, + 696 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $tr$ is the truncation distance.", + "bbox": [ + 76, + 703, + 308, + 715 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Depth-guided Sampling. For sampling along each ray, we observe that importance sampling does not show significant improvement while slowing down our tracking and mapping. Instead, we use depth-guided sampling: In addition to $M_{c}$ points uniformly sampled between near and far bound, for rays with a valid depth measurement, we further uniformly sample $M_{f}$ near-surface points within the range $[d - d_{s}, d + d_{s}]$ , where $d_{s}$ is a small offset.", + "bbox": [ + 75, + 717, + 468, + 839 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Tracking and Bundle Adjustment", + "text_level": 1, + "bbox": [ + 76, + 847, + 372, + 864 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Objective Functions. Our tracking and bundle adjustment are performed via minimizing our objective functions with", + "bbox": [ + 75, + 869, + 470, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "respect to learnable parameters $\\theta$ and camera parameters $\\xi_{t}$ . The color and depth rendering losses are $\\ell_2$ errors between the rendered results and observations:", + "bbox": [ + 498, + 396, + 892, + 441 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {r g b} = \\frac {1}{N} \\sum_ {n = 1} ^ {N} \\left(\\hat {\\mathbf {c}} _ {n} - \\mathbf {c} _ {n}\\right) ^ {2}, \\mathcal {L} _ {d} = \\frac {1}{| R _ {d} |} \\sum_ {r \\in R _ {d}} \\left(\\hat {d} _ {r} - D [ u, v ]\\right) ^ {2}. \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 500, + 450, + 890, + 506 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $R_{d}$ is the set of rays that have a valid depth measurement, and $u, v$ is the corresponding pixel on the image plane. To achieve accurate, smooth reconstructions with detailed geometry, we also apply approximate SDF and feature smoothness losses. Specifically, for samples within the truncation region, i.e. points where $|D[u,v] - d| \\leq tr$ , we use the distance between the sampled point and its observed depth value as an approximation of the ground-truth SDF value for supervision:", + "bbox": [ + 498, + 507, + 892, + 643 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {s d f} = \\frac {1}{| R _ {d} |} \\sum_ {r \\in R _ {d}} \\frac {1}{| S _ {r} ^ {t r} |} \\sum_ {p \\in S _ {r} ^ {t r}} \\left(s _ {p} - (D [ u, v ] - d)\\right) ^ {2}. \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 506, + 654, + 890, + 691 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For points that are far from the surface $((D[u,v] - d|) > tr)$ , we apply a free-space loss which forces the SDF prediction to be the truncated distance $tr$ :", + "bbox": [ + 498, + 704, + 890, + 750 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {f s} = \\frac {1}{| R _ {d} |} \\sum_ {r \\in R _ {d}} \\frac {1}{| S _ {r} ^ {f s} |} \\sum_ {p \\in S _ {r} ^ {f s}} (s _ {p} - t r) ^ {2}. \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 557, + 758, + 890, + 799 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To prevent the noisy reconstructions caused by hash collisions in unobserved free-space regions we perform additional regularization on the interpolated features $\\mathcal{V}_{\\alpha}(\\mathbf{x})$ ..", + "bbox": [ + 498, + 810, + 890, + 856 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {s m o o t h}} = \\frac {1}{| \\mathcal {G} |} \\sum_ {\\mathbf {x} \\in \\mathcal {G}} \\Delta_ {x} ^ {2} + \\Delta_ {y} ^ {2} + \\Delta_ {z} ^ {2}, \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 571, + 866, + 890, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "13296", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\Delta_{x,y,z} = \\mathcal{V}_{\\alpha}(\\mathbf{x} + \\epsilon_{x,y,z}) - \\mathcal{V}_{\\alpha}(\\mathbf{x})$ denotes the feature-metric difference between adjacent sampled vertices on the hash-grid along the three dimensions. Since performing regularization on the entire feature grid is computationally infeasible for real-time mapping, we only perform it in a small random region in each iteration.", + "bbox": [ + 75, + 90, + 468, + 181 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Camera Tracking. We track the camera-to-world transformation matrix $\\mathbf{T}_{wc} = \\exp (\\xi_t^{\\wedge})\\in \\mathbb{S}\\mathbb{E}(3)$ at each frame $t$ When a new frame comes in, we first initialize the pose of the current frame $i$ using constant speed assumption:", + "bbox": [ + 76, + 181, + 470, + 242 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {T} _ {t} = \\mathbf {T} _ {t - 1} \\mathbf {T} _ {t - 2} ^ {- 1} \\mathbf {T} _ {t - 1} \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 196, + 252, + 468, + 270 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Then, we select $N_{t}$ pixels within the current frame and iteratively optimize the pose by minimizing our objective function with respect to the camera parameters $\\xi_{t}$ .", + "bbox": [ + 75, + 280, + 468, + 325 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Bundle Adjustment. In neural SLAM, bundle adjustment usually consists of keyframe selection and joint optimization of camera poses and scene representation. Classic dense visual SLAM methods require saving keyframe (KF) images as the loss is formulated densely over all pixels. In contrast, the advantage of neural SLAM, as first shown by iMAP [26], is that BA can work with a sparse set of sampled rays. This is because the scene is represented as an implicit field using a neural network. However, iMAP and NICE-SLAM do not take full advantage of this - they still store full keyframe images following the classic SLAM paradigm and rely on keyframe selection (e.g. information gain, visual overlapping) to perform joint optimization on a small fraction of keyframes (usually less than 10).", + "bbox": [ + 75, + 325, + 468, + 536 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In Co-SLAM, we go further and drop the need for storing full keyframe images or keyframe selection. Instead, we only store a subset of pixels (around $5\\%$ ) to represent each keyframe. This allows us to insert new keyframes more frequently and maintain a much larger keyframe database. For joint optimization, we randomly sample a total number of $N_{g}$ rays from our global keyframe list to optimize our scene representation as well as camera poses. The joint optimization is performed in an alternating fashion. Specifically, we firstly optimize the scene representation $\\theta$ for $k_{m}$ steps and update camera poses using the accumulated gradient on camera parameters $\\{\\xi_t\\}$ . Since each camera pose uses only 6 parameters, this approach can improve the robustness of camera pose optimization with negligible extra computational cost on gradient accumulation.", + "bbox": [ + 75, + 537, + 468, + 763 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 76, + 777, + 209, + 792 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Experimental Setup", + "text_level": 1, + "bbox": [ + 76, + 801, + 266, + 816 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Datasets. We evaluate Co-SLAM on a variety of scenes from four different datasets. Following iMAP and NICE-SLAM, we quantitatively evaluate the reconstruction quality on 8 synthetic scenes from Replica [24]. We also evaluate on 7 synthetic scenes from NeuralRGBD [1], which", + "bbox": [ + 75, + 824, + 468, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "simulates noisy depth maps. For pose estimation, we evaluate the results on 6 scenes from ScanNet [5] with their ground truth pose obtained with BundleFusion [6], and 3 scenes from TUM RGB-D dataset [25] with their ground truth pose provided by a motion capture system.", + "bbox": [ + 496, + 90, + 890, + 167 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Metrics. We evaluate the reconstruction quality using Depth L1 (cm), Accuracy (cm), Completion (cm), and Completion ratio (\\%) with a threshold of 5cm. Following NICE-SLAM [42], we remove the unobserved regions that are outside of any camera frustum. In addition, we also perform an extra mesh culling that removes the noisy points within the camera frustum but outside the target scene. We observe that all methods experience a performance gain with our mesh culling strategy. Please refer to our supplementary material for more details. For evaluation of camera tracking, we adopt ATE RMSE [25] (cm).", + "bbox": [ + 496, + 167, + 890, + 333 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Baselines. We consider iMAP [26] and NICE-SLAM [42] as our main baselines for reconstruction quality and camera tracking. For a fair comparison, we evaluate iMAP and NICE-SLAM with the same mesh culling strategy as Co-SLAM. Note that iMAP* denotes the iMAP reimplementation released by the NICE-SLAM [42] authors, which is much slower than the original implementation. To investigate the trade-off between accuracy and frame rate on real-world datasets, we report results of two versions of our method: Ours refers to our proposed approach (which achieves real-time operation) while Ours† indicates our method ran with twice as many tracking iterations.", + "bbox": [ + 496, + 334, + 890, + 515 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Implementation Details. We run Co-SLAM on a desktop PC with a 3.60GHz Intel Core i7-12700K CPU and NVIDIA RTX 3090ti GPU. For experiments with default settings (Ours), which runs at 17 FPS on the Replica dataset, we use $N_{t} = 1024$ pixels with 10 iterations for tracking and $5\\%$ of pixels from every $5^{th}$ frame for global bundle adjustment. We sample $M_{r} = 32$ regular points and $M_{d} = 11$ depth points along each camera ray, with $tr = 10cm$ . Please refer to supplementary materials for more specific settings on all datasets.", + "bbox": [ + 496, + 516, + 890, + 667 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Evaluation of Tracking and Reconstruction", + "text_level": 1, + "bbox": [ + 498, + 679, + 867, + 695 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Replica dataset [24]. We evaluate on the same simulated RGB-D sequences as iMAP [26]. As shown in Tab. 1, our method achieves higher reconstruction accuracy and faster speed. Fig. 5 shows the qualitative results, from which we can observe that iMAP achieves plausible completion in unobserved areas but results are over-smoothed, while NICE-SLAM maintains more reconstruction details, but results contain some artifacts (e.g. the floors beside the bed, the back of the chairs). Co-SLAM successfully retains the advantages of both methods achieving consistent completion as well as high-fidelity reconstruction results.", + "bbox": [ + 496, + 703, + 890, + 869 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Synthetic dataset [1]. We perform further experiments on the synthetic dataset from NeuralRGBD [1]. Unlike the", + "bbox": [ + 496, + 869, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "13297", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/4c383fab653da07a7a1877b6d63d8a64fa124846ec18c79790ac03eb8f76e64b.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
DatasetMethodDepth L1 (cm)↓Acc. (cm)↓Comp. (cm)↓Comp. Ratio↑Tracking (ms) ↓Mapping (ms) ↓FPS ↑#param. ↓
Replica [24]TSDF-Fusion [4]6.361.623.9483.93N/AN/AN/A16.8 M
iMAP [23]4.643.624.9380.51101 (200, 6)448 (1000, 10)9.90.26 M
NICE-SLAM [42]1.902.372.6491.1378 (200, 10)5470 (1000, 60)0.9117.4 M
Ours1.512.102.0893.4458 (1024, 10)98 (2048, 10)17.40.26 M
Synthetic RGBD [1]TSDF-Fusion [4]10.871.625.1681.52N/AN/AN/A16.8 M
iMAP* [23]43.9118.3026.4120.731550 (5000, 50)14730 (5000, 300)0.340.22 M
NICE-SLAM [42]6.325.965.3077.46123 (1024, 10)3792 (1000, 60)1.313.11 M
Ours3.022.952.9686.8864 (1024, 10)104 (2048, 10)15.60.26 M
", + "bbox": [ + 86, + 89, + 879, + 203 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/2df974ef22e9a7cb52d511cd8076970ff2f54f8da091b3c3658303844252d056.jpg", + "table_caption": [ + "Table 1. Reconstruction quality and run-time memory comparison on Replica [24] and Synthetic-RGBD [1] with respective settings. TSDF-Fusion is reconstructed with poses estimated by Co-SLAM. Run-time is reported in time (#pixel, #iter) for a comprehensive comparison. The model size is averaged across all scenes." + ], + "table_footnote": [], + "table_body": "
MethodTrack. (ms) ↓Map. (ms) ↓FPS ↑#param. ↓
ScanNetiMAP*30.4×5044.9×3000.370.2 M
NICE-SLAM12.3×50125.3×600.6810.3 M
Ours†7.8×2020.2×106.40.8 M
Ours7.8×1020.2×1012.80.8 M
TUMiMAP*29.6×20044.3×3000.070.2 M
NICE-SLAM47.1×200189.2×600.08101.6 M
Ours†7.5×2019.0×206.71.6 M
Ours7.5×1019.0×2013.31.6 M
", + "bbox": [ + 81, + 262, + 465, + 383 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/d10b20d8bc3bfc2b66c23767fb1bfc894ff3eef77b3176aba9e349984c780846.jpg", + "image_caption": [ + "Figure 4. Reconstruction results on Synthetic RGB-D dataset [1]. Our method can recover thin structures and achieve plausible scene completion given noisy depth measurement." + ], + "image_footnote": [], + "bbox": [ + 81, + 488, + 460, + 779 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Replica dataset, it contains many thin structures and simulates the noise present in real depth sensor measurements. Quantitatively, our method significantly outperforms baseline methods (see Tab. 1) while still operating in real time", + "bbox": [ + 75, + 839, + 470, + 902 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/bb5a4ed5c52591299ca664fb283800007ce89195df09b5d234fd97421ea0fe15.jpg", + "table_caption": [ + "Table 2. Run-time and memory comparison on ScanNet [5] and TUM-RGBD [25] with respective settings. Run-time is reported in ms/iter $\\times$ #iter for a detailed comparison. NICE-SLAM and iMAP\\* run mapping at every frame on TUM-RGBD. Mapping happens every 5 frames in all other cases. The model size is averaged across all scenes." + ], + "table_footnote": [], + "table_body": "
Scene ID000000590106016901810207Avg.
iMAP* [26]55.9532.0617.5070.5132.1011.9136.67
NICE-SLAM [42]8.6412.258.0910.2812.935.599.63
Ours†7.1311.149.365.9011.817.148.75
Ours7.1812.299.576.6213.437.139.37
", + "bbox": [ + 503, + 262, + 888, + 327 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/09b7cbcc0fd91bb16ffe58eddfa79942f4d15faf70ab09aa38d4976189ff737e.jpg", + "table_caption": [ + "Table 3. ATE RMSE (cm) results of an average of 5 runs on ScanNet. Co-SLAM achieves better or on-par performance compared to NICE-SLAM [42] with significantly faster optimization speed." + ], + "table_footnote": [], + "table_body": "
fr1/desk (cm)fr2/xyz (cm)fr3/office (cm)
iMAP [26]4.92.05.8
iMAP* [26]7.22.19.0
NICE-SLAM [42]2.71.83.0
Ours†2.41.72.4
Ours2.71.92.6
BAD-SLAM [22]1.71.11.7
Kintinous [34]3.72.93.0
ORB-SLAM2 [17]1.60.41.0
", + "bbox": [ + 503, + 393, + 888, + 523 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 4. ATE RSME (cm) results on TUM RGB-D dataset. Our method achieves the best tracking performance among neural SLAM methods and maintains high-fidelity reconstruction.", + "bbox": [ + 498, + 527, + 890, + 570 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "(15 FPS). Fig. 4 shows some example qualitative results. Overall, Co-SLAM can capture fine details (e.g. wine bottles, chair legs, etc.) and produces complete and smooth reconstructions. NICE-SLAM yields less detailed and noisier reconstructions and cannot perform hole filling, while iMAP* lost track on some occasions.", + "bbox": [ + 496, + 582, + 890, + 672 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "ScanNet dataset [5]. We evaluate the camera tracking accuracy of Co-SLAM on 6 real-world sequences from ScanNet. The absolute trajectory error (ATE) is obtained by comparing predicted and ground-truth (generated by BundleFusion [6]) trajectories. Tab. 3 shows that quantitatively, our method achieves better tracking results in comparison to NICE-SLAM [42] with fewer tracking and mapping iterations while running at $6 - 12\\mathrm{Hz}$ (see Tab. 2). Fig. 6 also shows Co-SLAM achieves better reconstruction quality with smoother results and finer details (e.g. bike).", + "bbox": [ + 496, + 672, + 892, + 824 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "TUM dataset [25]. We further evaluate the tracking accuracy on the TUM dataset [25]. As shown in Tab. 4, our method achieves competitive tracking performance at 13 FPS. By increasing the number of tracking iterations $(\\mathrm{Ours}^{\\dagger})$ , our method achieves the best tracking performance", + "bbox": [ + 496, + 824, + 893, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "13298", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/356ee828facc0cdf57961b5b966cf106268208b2b266af6892c988897ac39be9.jpg", + "image_caption": [ + "Figure 5. Reconstruction results on Replica [24] dataset. In comparison to our baselines, our methods achieve accurate and high-quality scene reconstruction and completion on various scenes." + ], + "image_footnote": [], + "bbox": [ + 86, + 87, + 880, + 391 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/bf6bb8784c679cd7e1a4bd0e64ec1975812ed32587c1b88493dba3a66a0c64a9.jpg", + "image_caption": [ + "Figure 6. Reconstruction results on ScanNet [5]. In comparison to the previous methods, our reconstructions are smoother and contain more details thanks to our proposed joint encoding and global bundle adjustment strategy." + ], + "image_footnote": [], + "bbox": [ + 84, + 430, + 883, + 765 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "among neural SLAM methods, though at the expense of the FPS dropping to 6.7 (see Tab. 2). Although Co-SLAM still cannot outperform classic SLAM methods, it reduces the tracking performance gap between neural and classic methods, while improving the fidelity and completeness of the reconstructions.", + "bbox": [ + 75, + 806, + 470, + 898 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Performance Analysis", + "text_level": 1, + "bbox": [ + 500, + 806, + 705, + 821 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Run time and memory analysis. In our default setting (Ours), Co-SLAM can operate above $15\\mathrm{Hz}$ on a desktop PC with a 3.60GHz Intel Core i7-12700K CPU and NVIDIA RTX 3090ti GPU. For more challenging scenarios such as", + "bbox": [ + 498, + 839, + 893, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "13299", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/982816ef5dc39cf889ae88d3955a899477f5192a2f1eedae883c0a73844f4364.jpg", + "image_caption": [ + "Figure 7. Completion ratio vs. model size for Co-SLAM w/ and w/o coordinate encoding. Each model corresponds to a different hash-table size. iMAP [26] and NICE-SLAM [42] shown for reference." + ], + "image_footnote": [], + "bbox": [ + 96, + 88, + 446, + 251 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/09a456c7b3b11123174424c3b75c1046d442912cebf880e998381dd1cbcb3a10.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
w/o hash enc.w/o one-blob enc.Full model
Acc. (cm.) ↓3.692.132.10
Comp. (cm.) ↓3.432.132.08
Comp. Ratio ↑82.4993.1793.44
", + "bbox": [ + 89, + 321, + 454, + 383 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "the ScanNet and TUM datasets, Co-SLAM still achieves $5 - 13\\mathrm{Hz}$ runtime. Fig. 7 shows reconstruction quality with respect to memory use. Thanks to the sparse parametric encoding, our method requires significantly less memory than NICE-SLAM [42] while operating in real time and achieving accurate reconstruction results. Surprisingly, we found that further compressing the memory footprint (increasing the chances of hash collisions), Co-SLAM still outperforms iMAP [26], suggesting that our joint encoding improves the representation power of single encoding. Note that this figure is for illustration purposes, so we use the same spatial resolution throughout our hash encoding. Ideally, one can reduce the spatial resolution further to minimize hash collisions and achieve a better reconstruction quality.", + "bbox": [ + 75, + 441, + 467, + 652 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Scene completion. Fig. 2 shows an illustration of hole filling using different encoding strategies on a small scene. Coordinate encoding-based methods achieve plausible completion at the cost of lengthy training times, while parametric encoding-based methods fail at hole filling due to their local nature. By applying our new joint encoding, we observe that smooth hole filling can be achieved and fine structures are preserved by Co-SLAM.", + "bbox": [ + 75, + 654, + 467, + 773 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4. Ablations", + "text_level": 1, + "bbox": [ + 76, + 786, + 187, + 800 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effect of joint coordinate and parametric encoding. Tab. 5 illustrates a quantitative evaluation using different encodings. Our full model leads to higher accuracy and better completion than using single encodings (only one-blob or only hash-encoding). In addition, Fig. 7 illustrates that when compressing the size of the hash lookup table, our", + "bbox": [ + 75, + 809, + 467, + 900 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/96ca123e53ddd63153f2e86c70cfff72741afc257549d9f3424a183adcd12faf.jpg", + "table_caption": [ + "Table 5. Ablation study on different encodings. Defaultetable size is 13. Our full model with joint encoding achieves better completion and more accurate reconstructions. See also Fig. 2." + ], + "table_footnote": [], + "table_body": "
NameKF selection#KFPose optim.ATE (cm)Std. (cm)
LocalGlobal010All
w/o BA✓16.811.69
LBA✓✓✓9.691.38
GBA-10✓✓✓9.540.67
GBA✓✓✓8.750.33
", + "bbox": [ + 504, + 88, + 890, + 170 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 6. Ablation of BA strategies on Co-SLAM: (LBA) BA with rays from 10 local keyframes; (GBA-10) BA with rays from 10 keyframes randomly selected from all keyframes; (GBA) BA with rays from all keyframes (our full method). All methods sample the same number of total rays per iteration (2048).", + "bbox": [ + 498, + 175, + 890, + 246 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "full model with joint coordinate and parametric encoding is more robust in comparison to using a hash-based feature grid without the coordinate encoding.", + "bbox": [ + 498, + 263, + 890, + 308 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effect of global bundle adjustment. Tab. 6 shows the average ATE of our SLAM method on the 6 ScanNet scenes using different BA strategies: (w/o BA) pure tracking; (LBA) BA with rays from 10 local keyframes, a similar strategy to NICE-SLAM; (GBA-10) BA using rays from only 10 keyframes randomly selected from all past keyframes; (GBA) denotes the global BA strategy of Co-SLAM. We observe that using rays from a small (10) number of keyframes (LBA and GBA-10) leads to higher ATE errors. However, when keyframes are chosen from the entire sequence (GBA-10), instead of locally (LBA) the standard deviation is greatly reduced. Sampling rays from all keyframes (GBA) is the best overall strategy, even when all methods sample the same number of total rays (2048).", + "bbox": [ + 498, + 310, + 890, + 521 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 539, + 617, + 555 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We presented Co-SLAM, a dense real-time neural RGB-D SLAM system. We show that using a joint coordinate and parametric encoding with tiny MLPs as scene representation and training it with global bundle adjustment, achieves high-fidelity mapping and accurate tracking with plausible hole filling and efficient memory use.", + "bbox": [ + 498, + 566, + 890, + 656 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Limitations. Co-SLAM relies on inputs from an RGB-D sensor and is therefore sensitive to illumination changes and inaccurate depth measurements. Instead of sampling keyframe pixels randomly, an information-guided pixel sampling strategy could be helpful to further reduce the number of pixels and improve the tracking speed. Incorporating loop closure could lead to further improvements.", + "bbox": [ + 498, + 657, + 890, + 763 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 500, + 782, + 666, + 800 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The research presented here has been supported by a sponsored research award from Cisco Research and the UCL Centre for Doctoral Training in Foundational AI under UKRI grant number EP/S021566/1. This project made use of time on Tier 2 HPC facility JADE2, funded by EPSRC (EP/T022205/1).", + "bbox": [ + 498, + 809, + 890, + 898 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "13300", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Dejan Azinović, Ricardo Martin-Brualla, Dan B. Goldman, Matthias Nießner, and Justus Thies. Neural RGB-D surface reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6290-6301, 2022. 1, 2, 3, 4, 5, 6", + "[2] Eric R. Chan, Connor Z. Lin, Matthew A. Chan, Koki Nagano, Boxiao Pan, Shalini De Mello, Orazio Gallo, Leonidas J. Guibas, Jonathan Tremblay, and Sameh Khamis. Efficient geometry-aware 3D generative adversarial networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16123-16133, 2022. 2, 3", + "[3] Jiawen Chen, Dennis Bautembach, and Shahram Izadi. Scalable real-time volumetric surface reconstruction. ACM Transactions on Graphics (ToG), 32(4):1-16, 2013. 3", + "[4] Brian Curless and Marc Levoy. A volumetric method for building complex models from range images. In Proceedings of the 23rd Annual Conference on Computer Graphics and Interactive Techniques, pages 303-312, 1996. 6", + "[5] Angela Dai, Angel X. Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 5828–5839, 2017. 2, 5, 6, 7", + "[6] Angela Dai, Matthias Nießner, Michael Zollhöfer, Shahram Izadi, and Christian Theobalt. Bundlefusion: Real-time globally consistent 3d reconstruction using on-the-fly surface reintegration. ACM Transactions on Graphics (ToG), 36(4):1, 2017. 2, 5, 6", + "[7] Sara Fridovich-Keil, Alex Yu, Matthew Tancik, Qinhong Chen, Benjamin Recht, and Angjoo Kanazawa. Plenoxels: Radiance Fields Without Neural Networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5501-5510, 2022. 2, 3", + "[8] Haoyu Guo, Sida Peng, Haotong Lin, Qianqian Wang, Guofeng Zhang, Hujun Bao, and Xiaowei Zhou. Neural 3D Scene Reconstruction with the Manhattan-world Assumption. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5511-5520, 2022. 3", + "[9] Olaf Kähler, Victor Adrian Prisacariu, Carl Yuheng Ren, Xin Sun, Philip H. S. Torr, and David William Murray. Very high frame rate volumetric integration of depth images on mobile devices. IEEE Trans. Vis. Comput. Graph., 21(11):1241-1250, 2015. 3", + "[10] Animesh Karnewar, Tobias Ritschel, Oliver Wang, and Niloy Mitra. Relu fields: The little non-linearity that could. In ACM SIGGRAPH 2022 Conference Proceedings, pages 1-9, 2022. 2, 3", + "[11] L Koestler, N Yang, N Zeller, and D Cremers. Tandem: Tracking and dense mapping in real-time using deep multiview stereo. In Conference on Robot Learning (CoRL), 2021. 3", + "[12] Hai Li, Xingrui Yang, Hongjia Zhai, Yuqian Liu, Hujun Bao, and Guofeng Zhang. Vox-Surf: Voxel-based Implicit Surface" + ], + "bbox": [ + 78, + 114, + 470, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Representation. arXiv preprint arXiv:2208.10925, 2022. 2, 3", + "[13] Lingjie Liu, Jiatao Gu, Kyaw Zaw Lin, Tat-Seng Chua, and Christian Theobalt. Neural sparse voxel fields. Advances in Neural Information Processing Systems, 33:15651-15663, 2020. 2, 3", + "[14] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In European Conference on Computer Vision, pages 405-421. Springer, 2020. 1, 3", + "[15] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM Transactions on Graphics, 41(4):102:1-102:15, July 2022. 2, 3", + "[16] Thomas Müller, Brian McWilliams, Fabrice Rousselle, Markus Gross, and Jan Novák. Neural importance sampling. ACM Transactions on Graphics (TOG), 38(5):1-19, 2019. 2, 3", + "[17] Raul Mur-Artal and Juan D. Tardós. Orb-slam2: An opensource slam system for monocular, stereo, and rgb-d cameras. IEEE transactions on robotics, 33(5):1255-1262, 2017. 6", + "[18] Richard A Newcombe, Shahram Izadi, Otmar Hilliges, David Molyneaux, David Kim, Andrew J Davison, Pushmeet Kohli, Jamie Shotton, Steve Hodges, and Andrew Fitzgibbon. KinectFusion: Real-time dense surface mapping and tracking. In IEEE International Symposium on Mixed and Augmented Reality, pages 127-136, 2011. 2", + "[19] Matthias Nießner, Michael Zollhöfer, Shahram Izadi, and Marc Stamminger. Real-time 3D reconstruction at scale using voxel hashing. ACM Trans. Graph., 32(6):169:1-169:11, 2013. 3", + "[20] Nasim Rahaman, Aristide Baratin, Devansh Arpit, Felix Draxler, Min Lin, Fred Hamprecht, Yoshua Bengio, and Aaron Courville. On the spectral bias of neural networks. In International Conference on Machine Learning, pages 5301-5310. PMLR, 2019. 3", + "[21] Thomas Schops, Torsten Sattler, and Marc Pollefeys. BAD SLAM: Bundle adjusted direct RGB-D SLAM. In CVPR, 2019. 3", + "[22] Thomas Schops, Torsten Sattler, and Marc Pollefeys. Bad slam: Bundle adjusted direct rgb-d slam. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 134-144, 2019. 6", + "[23] Vincent Sitzmann, Julien Martel, Alexander Bergman, David Lindell, and Gordon Wetzstein. Implicit neural representations with periodic activation functions. Advances in Neural Information Processing Systems, 33:7462-7473, 2020. 6", + "[24] Julian Straub, Thomas Whelan, Lingni Ma, Yufan Chen, Erik Wijmans, Simon Green, Jakob J. Engel, Raul Mur-Artal, Carl Ren, and Shobhit Verma. The Replica dataset: A digital replica of indoor spaces. arXiv preprint arXiv:1906.05797, 2019. 2, 5, 6, 7", + "[25] Jürgen Sturm, Nikolas Engelhard, Felix Endres, Wolfram Burgard, and Daniel Cremers. A benchmark for the evaluation of RGB-D SLAM systems. In 2012 IEEE/RSJ Interna" + ], + "bbox": [ + 501, + 92, + 893, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "13301", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "tional Conference on Intelligent Robots and Systems, pages 573-580.IEEE,2012.2,5,6", + "[26] Edgar Sucar, Shikun Liu, Joseph Ortiz, and Andrew J. Davison. iMAP: Implicit mapping and positioning in real-time. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6229-6238, 2021. 2, 3, 4, 5, 6, 7, 8", + "[27] Cheng Sun, Min Sun, and Hwann-Tzong Chen. Direct voxel grid optimization: Super-fast convergence for radiance fields reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5459-5469, 2022. 3", + "[28] Towaki Takikawa, Joey Litalien, Kangxue Yin, Karsten Kreis, Charles Loop, Derek Nowrouzezahrai, Alec Jacobson, Morgan McGuire, and Sanja Fidler. Neural geometric level of detail: Real-time rendering with implicit 3D shapes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11358-11367, 2021. 2, 3", + "[29] K. Tateno, F. Tombari, I. Laina, and N. Navab. CNN-SLAM: Real-time dense monocular slam with learned depth prediction. In CVPR, 2017. 3", + "[30] Zachary Teed and Jia Deng. DROID-SLAM: Deep Visual SLAM for Monocular, Stereo, and RGB-D Cameras. Advances in neural information processing systems, 2021. 3", + "[31] Emanuele Vespa, Nikolay Nikolov, Marius Grimm, Luigi Nardi, Paul HJ Kelly, and Stefan Leutenegger. Efficient octree-based volumetric SLAM supporting signed-distance and occupancy mapping. RAL, 2018. 3", + "[32] Jingwen Wang, Tymoteusz Bleja, and Lourdes Agapito. Go-surf: Neural feature grid optimization for fast, high-fidelity rgb-d surface reconstruction. arXiv preprint arXiv:2206.14735, 2022. 2, 3", + "[33] Peng Wang, Lingjie Liu, Yuan Liu, Christian Theobalt, Taku Komura, and Wenping Wang. Neus: Learning neural implicit surfaces by volume rendering for multi-view reconstruction. Advances in Neural Information Processing Systems, 2021. 3, 4", + "[34] Thomas Whelan, Michael Kaess, Maurice Fallon, Hordur Johannsson, John Leonard, and John McDonald. Kintinous: Spatially extended kinectfusion. 2012. 6", + "[35] Thomas Whelan, Stefan Leutenegger, Renato F. Salas-Moreno, Ben Glocker, and Andrew J. Davison. ElasticFusion: Dense SLAM without a pose graph. In Robotics: Science and Systems, volume 11, 2015. 3", + "[36] Thomas Whelan, Renato F Salas-Moreno, Ben Glocker, Andrew J Davison, and Stefan Leutenegger. ElasticFusion: Real-time dense SLAM and light source estimation. In International Journal of Robotics Research, volume 35, pages 1697-1716, 2016. 3", + "[37] Tong Wu, Jiaqi Wang, Xingang Pan, Xudong Xu, Christian Theobalt, Ziwei Liu, and Dahua Lin. Voxurf: Voxel-based Efficient and Accurate Neural Surface Reconstruction. arXiv preprint arXiv:2208.12697, 2022. 3", + "[38] Nan Yang, Lukas von Stumberg, Rui Wang, and Daniel Cremers. D3vo: Deep depth, deep pose and deep uncertainty for monocular visual odometry. In Proceedings of" + ], + "bbox": [ + 78, + 92, + 470, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1281-1292, 2020. 3", + "[39] Lior Yariv, Jiatao Gu, Yoni Kasten, and Yaron Lipman. Volume rendering of neural implicit surfaces. Advances in Neural Information Processing Systems, 34:4805-4815, 2021. 3, 4", + "[40] Zehao Yu, Songyou Peng, Michael Niemeyer, Torsten Sattler, and Andreas Geiger. Monosdf: Exploring monocular geometric cues for neural implicit surface reconstruction. Advances in Neural Information Processing Systems, 2022. 2, 3", + "[41] Ming Zeng, Fukai Zhao, Jiaxiang Zheng, and Xinguo Liu. Octree-based fusion for realtime 3d reconstruction. Graphical Models, 75(3):126-136, 2013. 3", + "[42] Zihan Zhu, Songyou Peng, Viktor Larsson, Weiwei Xu, Hujun Bao, Zhaopeng Cui, Martin R. Oswald, and Marc Pollefeys. Nice-slam: Neural implicit scalable encoding for slam. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12786-12796, 2022. 2, 3, 4, 5, 6, 7, 8" + ], + "bbox": [ + 501, + 92, + 890, + 372 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "13302", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2023/Co-SLAM_ Joint Coordinate and Sparse Parametric Encodings for Neural Real-Time SLAM/98d936fc-c445-4925-9f16-8d4e0f8fc517_model.json b/2023/Co-SLAM_ Joint Coordinate and Sparse Parametric Encodings for Neural Real-Time SLAM/98d936fc-c445-4925-9f16-8d4e0f8fc517_model.json new file mode 100644 index 0000000000000000000000000000000000000000..d6b39b92bd5c069997e60758a90ed8dbf0ba8959 --- /dev/null +++ b/2023/Co-SLAM_ Joint Coordinate and Sparse Parametric Encodings for Neural Real-Time SLAM/98d936fc-c445-4925-9f16-8d4e0f8fc517_model.json @@ -0,0 +1,2101 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.0, + 0.812, + 0.045 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.149, + 0.131, + 0.823, + 0.175 + ], + "angle": 0, + "content": "Co-SLAM: Joint Coordinate and Sparse Parametric Encodings for Neural Real-Time SLAM" + }, + { + "type": "text", + "bbox": [ + 0.241, + 0.193, + 0.73, + 0.228 + ], + "angle": 0, + "content": "Hengyi Wang* Jingwen Wang* Lourdes Agapito Department of Computer Science, University College London" + }, + { + "type": "text", + "bbox": [ + 0.245, + 0.231, + 0.721, + 0.246 + ], + "angle": 0, + "content": "{hengyi.wang.21, jingwen.wang.17, 1.agapito}@ucl.ac.uk" + }, + { + "type": "image", + "bbox": [ + 0.107, + 0.264, + 0.871, + 0.434 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.445, + 0.893, + 0.488 + ], + "angle": 0, + "content": "Figure 1. We present Co-SLAM, a neural RGB-D SLAM method that performs online tracking and mapping in real time. We propose a new hybrid representation based on a joint coordinate and sparse-parametric encoding with global bundle adjustment. Our method shows fast, high-fidelity scene reconstruction with efficient memory use and plausible hole-filling." + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.502, + 0.314, + 0.517 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.536, + 0.474, + 0.854 + ], + "angle": 0, + "content": "We present Co-SLAM, a neural RGB-D SLAM system based on a hybrid representation, that performs robust camera tracking and high-fidelity surface reconstruction in real time. Co-SLAM represents the scene as a multi-resolution hash-grid to exploit its high convergence speed and ability to represent high-frequency local features. In addition, Co-SLAM incorporates one-blob encoding, to encourage surface coherence and completion in unobserved areas. This joint parametric-coordinate encoding enables real-time and robust performance by bringing the best of both worlds: fast convergence and surface hole filling. Moreover, our ray sampling strategy allows Co-SLAM to perform global bundle adjustment over all keyframes instead of requiring keyframe selection to maintain a small number of active keyframes as competing neural SLAM approaches do. Experimental results show that Co-SLAM runs at \\(10 - 17\\mathrm{Hz}\\) and achieves state-of-the-art scene reconstruction results, and competitive tracking performance in various datasets and benchmarks (ScanNet, TUM, Replica, Synthetic RGBD). Project page: https://hengyiwang.github.io/projects/CoSLAM" + }, + { + "type": "title", + "bbox": [ + 0.502, + 0.502, + 0.631, + 0.517 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.531, + 0.893, + 0.637 + ], + "angle": 0, + "content": "Real-time joint camera tracking and dense surface reconstruction from RGB-D sensors has been a core problem in computer vision and robotics for decades. Traditional SLAM solutions exist that can robustly track the position of the camera while fusing depth and/or color measurements into a single high-fidelity map. However, they rely on handcrafted loss terms and do not exploit data-driven priors." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.64, + 0.895, + 0.821 + ], + "angle": 0, + "content": "Recent attention has turned to learning-based models that can exploit the ability of neural network architectures to learn smoothness and coherence priors directly from data. Coordinate-based networks have probably become the most popular representation, since they can be trained to predict the geometric and appearance properties of any point in the scene in a self-supervised way, directly from images. The most notable example, Neural Radiance Fields (NeRF) [14], encodes scene density and color in the weights of a neural network. In combination with volume rendering, NeRF is trained to re-synthesize the input images and has a remarkable ability to generalize to nearby unseen views." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.825, + 0.895, + 0.901 + ], + "angle": 0, + "content": "Coordinate-based networks embed input point coordinates into a high dimensional space, using sinusoidal or other frequency embeddings, allowing them to capture high-frequency details that are essential for high-fidelity geometry reconstruction [1]. Combined with the smooth-" + }, + { + "type": "page_footnote", + "bbox": [ + 0.102, + 0.888, + 0.265, + 0.9 + ], + "angle": 0, + "content": "\\(\\star\\) Indicates equal contribution." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13293" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.178, + 0.089, + 0.265, + 0.1 + ], + "angle": 0, + "content": "COORDINATE" + }, + { + "type": "image", + "bbox": [ + 0.136, + 0.102, + 0.307, + 0.202 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.195, + 0.205, + 0.247, + 0.216 + ], + "angle": 0, + "content": "Frequency" + }, + { + "type": "image_caption", + "bbox": [ + 0.354, + 0.089, + 0.439, + 0.1 + ], + "angle": 0, + "content": "PARAMETRIC" + }, + { + "type": "image", + "bbox": [ + 0.31, + 0.102, + 0.482, + 0.202 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.37, + 0.204, + 0.423, + 0.215 + ], + "angle": 0, + "content": "DenseGrid" + }, + { + "type": "image_caption", + "bbox": [ + 0.552, + 0.089, + 0.591, + 0.1 + ], + "angle": 0, + "content": "JOINT" + }, + { + "type": "image", + "bbox": [ + 0.485, + 0.102, + 0.657, + 0.202 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.521, + 0.204, + 0.621, + 0.214 + ], + "angle": 0, + "content": "DenseGrid+OneBlob" + }, + { + "type": "image_caption", + "bbox": [ + 0.707, + 0.089, + 0.786, + 0.1 + ], + "angle": 0, + "content": "REFERENCE" + }, + { + "type": "image", + "bbox": [ + 0.66, + 0.102, + 0.832, + 0.202 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.702, + 0.204, + 0.79, + 0.215 + ], + "angle": 0, + "content": "NICE-SLAM [42]" + }, + { + "type": "image", + "bbox": [ + 0.136, + 0.216, + 0.308, + 0.316 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.199, + 0.318, + 0.244, + 0.328 + ], + "angle": 0, + "content": "OneBlob" + }, + { + "type": "image", + "bbox": [ + 0.312, + 0.216, + 0.482, + 0.316 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.373, + 0.318, + 0.421, + 0.328 + ], + "angle": 0, + "content": "Grid" + }, + { + "type": "image", + "bbox": [ + 0.485, + 0.216, + 0.657, + 0.316 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.507, + 0.318, + 0.635, + 0.329 + ], + "angle": 0, + "content": "HashGrid+OneBlob (Ours)" + }, + { + "type": "image", + "bbox": [ + 0.66, + 0.216, + 0.832, + 0.316 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.737, + 0.319, + 0.755, + 0.328 + ], + "angle": 0, + "content": "GT" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.336, + 0.892, + 0.379 + ], + "angle": 0, + "content": "Figure 2. Illustration of the effect of different encodings on completion. COORDINATE based encodings achieve hole filling but require long training times. PARAMETRIC encodings allow fast training, but fail to complete unobserved regions. JOINT coordinate and parametric encoding (Ours) allows smooth scene completion and fast training. NICE-SLAM [42] uses a dense parametric encoding." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.395, + 0.47, + 0.576 + ], + "angle": 0, + "content": "ness and coherence priors inherently encoded in the MLP weights, they constitute a good choice for sequential tracking and mapping [26]. However, the weakness of MLP-based approaches is the long training times required (sometimes hours) to learn a single scene. For that reason, recent real-time capable SLAM systems built on coordinate networks with frequency embeddings such as iMAP [26] need to resort to strategies to sparsify ray sampling and reduce tracking iterations to maintain interactive operation. This comes at the cost of loss of detail in the reconstructions which are oversmoothed (Fig. 5) and potential errors in camera tracking." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.58, + 0.469, + 0.852 + ], + "angle": 0, + "content": "Optimizable feature grids, also known as parametric embeddings, have recently become a powerful alternative scene representation to monolithic MLPs, given their ability to represent high-fidelity local features and their extremely fast convergence (orders of magnitude faster) [7, 10, 15, 32, 40]. Recent efforts focus on sparse alternatives to these parametric embeddings such as octrees [28], tri-plane [2], hash-grid [15] or sparse voxel grid [12, 13] to improve the memory efficiency of dense grids. While these representations can be fast to train and are therefore well suited to real-time operation, they fundamentally lack the smoothness, and coherence priors inherent to MLPs and struggle with hole-filling in areas without observation. NICE-SLAM [42] is a recent example of a multi-resolution feature grid-based SLAM method. Although it does not suffer from over-smoothness and captures local detail (as shown in Fig. 2) it cannot perform hole-filling which might in turn lead to drift in camera pose estimation." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.47, + 0.9 + ], + "angle": 0, + "content": "Our first contribution is to design a joint coordinate and sparse grid encoding for input points that brings together the benefits of both worlds to the real-time SLAM frame-" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.395, + 0.892, + 0.485 + ], + "angle": 0, + "content": "work. On the one hand, the smoothness and coherence priors provided by coordinate encodings (we use one-blob [16] encoding), and on the other hand the optimization speed and local details of sparse feature encodings (we use hash grid [15]), resulting in more robust camera tracking and high-fidelity maps with better completion and hole filling." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.486, + 0.893, + 0.711 + ], + "angle": 0, + "content": "Our second contribution relates to the bundle adjustment (BA) step in the joint optimization of the map and camera poses. So far, all neural SLAM systems [26, 42] perform BA using rays sampled from a very small subset of selected keyframes. Restricting the optimization to a very small number of viewpoints results in decreased robustness in camera tracking and increased computation due to the need for a keyframe-selection strategy. Instead, Co-SLAM performs global BA, sampling rays from all past keyframes, which results in an important boost in robustness and performance in pose estimation. In addition, we show that our BA optimization requires a fraction of the iterations of NICE-SLAM [42] to achieve similar errors. In practice, Co-SLAM achieves SOTA performance in camera tracking and 3D reconstruction while maintaining real time performance." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.712, + 0.892, + 0.833 + ], + "angle": 0, + "content": "Co-SLAM runs at 15-17Hz on Replica and Synthetic RGB-D datasets [1], and 12-13Hz on ScanNet [5] and TUM [25] scenes — faster than NICE-SLAM (0.1-1Hz) [42] and iMAP [26]. We perform extensive evaluations on various datasets (Replica [24], Synthetic RGBD [1], ScanNet [6], TUM [25]) where we outperform NICE-SLAM [42] and iMAP [26] in reconstruction and achieve better or at least on-par tracking accuracy." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.845, + 0.642, + 0.861 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Dense Visual SLAM. Taking advantage of commodity depth sensors, KinectFusion [18] performs frame-to-model" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13294" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.319 + ], + "angle": 0, + "content": "camera tracking via projective iterative-closest-point (ICP), and incrementally updates the scene geometry via TSDF-Fusion. Subsequent works focused on addressing the scalability issue by adopting more efficient data structures, such as surface elements (surfels) [35, 36], VoxelHashing [3, 9, 19] or Octrees [31, 41]. While most works focus more heavily on scene reconstruction and only track per-frame poses, BAD-SLAM [21] proposes full direct bundle adjustment (BA) to jointly optimize keyframe (KF) poses and the dense 3D structure. Several recent works [11, 29, 30, 38] leverage deep learning to improve the accuracy and robustness of traditional SLAM, and even achieve dense reconstruction with monocular SLAM. While these approaches introduced some learned components, the scene representation and overall pipeline still follow traditional SLAM methods." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.33, + 0.473, + 0.618 + ], + "angle": 0, + "content": "Neural Implicit Representations. Recently neural implicit representations [14] that encode 3D geometry and the appearance of a scene within the weights of a neural network have gained popularity due to their expressiveness and compactness. Among these works, NeRF [14] and its variants adopt coordinate encoding [16, 20] with MLPs and show impressive scene reconstruction using differentiable rendering. Given that coordinate encoding-based methods require lengthy training, many follow-up works [7, 10, 27] propose parametric encodings that increase parameter size but speed up the training. To improve the memory efficiency of parametric encoding-based methods, sparse parametric encodings, such as Octree [28], Tri-plane [2], or sparse voxel grid [12, 13, 15], have been proposed. While these methods focus on novel view synthesis, others instead focus on surface reconstruction from RGB images, with implicit surface representations and differentiable renderers [8, 33, 39, 40]. Other methods [1, 26, 32, 37, 42] use depth measurements as additional supervision for surface reconstruction." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.629, + 0.473, + 0.903 + ], + "angle": 0, + "content": "Neural Implicit SLAM. iMAP [26] adopts an MLP representation to perform joint tracking and mapping in quasi-real time. Smooth, plausible filling of unobserved regions is achieved thanks to the coherence priors inherent to the MLP. iMAP introduces elaborate keyframe selection and information-guided pixel sampling for speed, resulting in \\(10\\mathrm{Hz}\\) tracking and \\(2\\mathrm{Hz}\\) mapping. To reduce the computational overhead and improve scalability, NICE-SLAM [42] adopts a multi-level feature grid for scene representation. However, as feature grids only perform local updates, they fail to achieve plausible hole-filling. With Co-SLAM we aim to address both issues. To achieve real-time performance and memory-efficiency, while maintaining high-fidelity surface reconstruction and plausible hole filling, we propose to combine the use of coordinate and sparse parametric encodings for scene representation, and perform dense global bundle adjustment using rays sampled from all keyframes." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.09, + 0.593, + 0.107 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.115, + 0.895, + 0.222 + ], + "angle": 0, + "content": "Fig. 3 shows an overview of Co-SLAM. Given an input RGB-D stream \\(\\{I_t\\}_{t=1}^N \\{D_t\\}_{t=1}^N\\) with known camera intrinsics \\(K \\in \\mathbb{R}^{3 \\times 3}\\), we perform dense mapping and tracking by jointly optimizing camera poses \\(\\{\\xi_t\\}_{t=1}^N\\) and a neural scene representation \\(f_\\theta\\). Specifically, our implicit representation maps world coordinates \\(\\mathbf{x}\\) into color \\(\\mathbf{c}\\) and truncated signed distance (TSDF) \\(s\\) values:" + }, + { + "type": "equation", + "bbox": [ + 0.642, + 0.231, + 0.892, + 0.248 + ], + "angle": 0, + "content": "\\[\nf _ {\\theta} (\\mathbf {x}) \\mapsto (\\mathbf {c}, s). \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.256, + 0.895, + 0.409 + ], + "angle": 0, + "content": "Similar to most SLAM systems, the process is split into tracking and mapping. Initialization is performed by running a few training iterations on the first frame. For each subsequent frame, the camera pose is optimized first, initialized with a simple constant-speed motion model. A small fraction of pixels/rays are then sampled and copied to the global pixel-set. At each mapping iteration, global bundle adjustment is performed over a set of pixels randomly sampled from the global pixel-set, to jointly optimize the scene representation \\(\\theta\\) and all camera poses \\(\\{\\xi_t\\}\\)." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.415, + 0.871, + 0.432 + ], + "angle": 0, + "content": "3.1. Joint Coordinate and Parametric Encoding" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.439, + 0.895, + 0.756 + ], + "angle": 0, + "content": "Thanks to the coherence and smoothness priors inherent to MLPs, coordinate-based representations achieve high-fidelity scene reconstruction. However, these methods often suffer from slow convergence and catastrophic forgetting, when optimized in a sequential setting. Instead, parametric encoding based methods improve the computational efficiency, but they fall short of hole filling and smoothness. Since both properties of speed and coherence are crucial for a real-world SLAM system, we propose a joint coordinate and parametric encoding that combines the best of both worlds: we adopt coordinate encoding for scene representation while using sparse parametric encoding to speed up training. Specifically, we use One-blob [16] encoding \\(\\gamma (\\mathbf{x})\\) instead of embedding spatial coordinates into multiple frequency bands. As scene representation we adopt a multiresolution hash-based feature grid [15] \\(\\mathcal{V}_{\\alpha} = \\{\\mathcal{V}_{\\alpha}^{l}\\}_{l = 1}^{L}\\). The spatial resolution of each level is set between the coarsest \\(R_{min}\\) and the finest resolution \\(R_{max}\\) in a progressive manner. Feature vectors \\(\\mathcal{V}_{\\alpha}(\\mathbf{x})\\) at each sampled point \\(\\mathbf{x}\\) are queried via trilinear interpolation. The geometry decoder outputs the predicted SDF value \\(s\\) and a feature vector \\(\\mathbf{h}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.604, + 0.765, + 0.892, + 0.782 + ], + "angle": 0, + "content": "\\[\nf _ {\\tau} (\\gamma (\\mathbf {x}), \\mathcal {V} _ {\\alpha} (\\mathbf {x})) \\mapsto (\\mathbf {h}, s). \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.79, + 0.816, + 0.806 + ], + "angle": 0, + "content": "Finally, the color MLP predicts the RGB value:" + }, + { + "type": "equation", + "bbox": [ + 0.635, + 0.815, + 0.892, + 0.832 + ], + "angle": 0, + "content": "\\[\nf _ {\\phi} (\\gamma (\\mathbf {x}), \\mathbf {h}) \\mapsto \\mathbf {c}. \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.895, + 0.902 + ], + "angle": 0, + "content": "Here \\(\\theta = \\{\\alpha, \\phi, \\tau\\}\\) are learnable parameters. Injecting the One-blob encoding in the hash-based multi-resolution feature grid representation, results in fast convergence, efficient memory use, and hole filling needed for online SLAM." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "13295" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.099, + 0.089, + 0.873, + 0.337 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.341, + 0.895, + 0.385 + ], + "angle": 0, + "content": "Figure 3. Overview of Co-SLAM. 1) Scene representation: using our new joint coordinate+parametric encoding, input coordinates are mapped to RGB and SDF values via two shallow MLPs. 2) Tracking: optimize per-frame camera poses \\(\\xi_{t}\\) by minimizing losses. 3) Mapping: global bundle adjustment to jointly optimize the scene representation and camera poses taking rays sampled from all keyframes." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.396, + 0.33, + 0.414 + ], + "angle": 0, + "content": "3.2. Depth and Color Rendering" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.419, + 0.471, + 0.512 + ], + "angle": 0, + "content": "Following [26, 42], we render depth and color by integrating the predicted values along the sampled rays. Specifically, given the camera origin \\(\\mathbf{o}\\) and ray direction \\(\\mathbf{r}\\), we uniformly sample \\(M\\) points \\(\\mathbf{x}_i = \\mathbf{o} + d_i\\mathbf{r}\\), \\(i \\in \\{1, \\dots, M\\}\\) with depth values \\(\\{t_1, \\dots, t_M\\}\\) and predicted colors \\(\\{\\mathbf{c}_1, \\dots, \\mathbf{c}_M\\}\\) and render color and depth as" + }, + { + "type": "equation", + "bbox": [ + 0.093, + 0.52, + 0.472, + 0.562 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathbf {c}} = \\frac {1}{\\sum_ {i = 1} ^ {M} w _ {i}} \\sum_ {i = 1} ^ {M} w _ {i} \\mathbf {c} _ {i}, \\quad \\hat {d} = \\frac {1}{\\sum_ {i = 1} ^ {M} w _ {i}} \\sum_ {i = 1} ^ {M} w _ {i} d _ {i}, \\quad (4)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.57, + 0.47, + 0.661 + ], + "angle": 0, + "content": "where \\(\\{w_i\\}\\) are the computed weights along the ray. A conversion function is needed to convert predicted SDF \\(s_i\\) to weight \\(w_i\\). Instead of adopting the rendering equations proposed in Neus [33, 39], we follow the simple bell-shaped model of [1] and compute weights \\(w_i\\) directly by multiplying the two Sigmoid functions \\(\\sigma(\\cdot)\\)" + }, + { + "type": "equation", + "bbox": [ + 0.19, + 0.668, + 0.47, + 0.697 + ], + "angle": 0, + "content": "\\[\nw _ {i} = \\sigma \\left(\\frac {s _ {i}}{t r}\\right) \\sigma \\left(- \\frac {s _ {i}}{t r}\\right), \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.704, + 0.31, + 0.717 + ], + "angle": 0, + "content": "where \\( tr \\) is the truncation distance." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.718, + 0.47, + 0.84 + ], + "angle": 0, + "content": "Depth-guided Sampling. For sampling along each ray, we observe that importance sampling does not show significant improvement while slowing down our tracking and mapping. Instead, we use depth-guided sampling: In addition to \\( M_{c} \\) points uniformly sampled between near and far bound, for rays with a valid depth measurement, we further uniformly sample \\( M_{f} \\) near-surface points within the range \\( [d - d_{s}, d + d_{s}] \\), where \\( d_{s} \\) is a small offset." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.848, + 0.373, + 0.865 + ], + "angle": 0, + "content": "3.3. Tracking and Bundle Adjustment" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.871, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Objective Functions. Our tracking and bundle adjustment are performed via minimizing our objective functions with" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.397, + 0.893, + 0.442 + ], + "angle": 0, + "content": "respect to learnable parameters \\(\\theta\\) and camera parameters \\(\\xi_{t}\\). The color and depth rendering losses are \\(\\ell_2\\) errors between the rendered results and observations:" + }, + { + "type": "equation", + "bbox": [ + 0.5, + 0.452, + 0.892, + 0.507 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {r g b} = \\frac {1}{N} \\sum_ {n = 1} ^ {N} \\left(\\hat {\\mathbf {c}} _ {n} - \\mathbf {c} _ {n}\\right) ^ {2}, \\mathcal {L} _ {d} = \\frac {1}{| R _ {d} |} \\sum_ {r \\in R _ {d}} \\left(\\hat {d} _ {r} - D [ u, v ]\\right) ^ {2}. \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.508, + 0.893, + 0.644 + ], + "angle": 0, + "content": "where \\( R_{d} \\) is the set of rays that have a valid depth measurement, and \\( u, v \\) is the corresponding pixel on the image plane. To achieve accurate, smooth reconstructions with detailed geometry, we also apply approximate SDF and feature smoothness losses. Specifically, for samples within the truncation region, i.e. points where \\( |D[u,v] - d| \\leq tr \\), we use the distance between the sampled point and its observed depth value as an approximation of the ground-truth SDF value for supervision:" + }, + { + "type": "equation", + "bbox": [ + 0.508, + 0.655, + 0.892, + 0.693 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {s d f} = \\frac {1}{| R _ {d} |} \\sum_ {r \\in R _ {d}} \\frac {1}{| S _ {r} ^ {t r} |} \\sum_ {p \\in S _ {r} ^ {t r}} \\left(s _ {p} - (D [ u, v ] - d)\\right) ^ {2}. \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.705, + 0.892, + 0.75 + ], + "angle": 0, + "content": "For points that are far from the surface \\(((D[u,v] - d|) > tr)\\), we apply a free-space loss which forces the SDF prediction to be the truncated distance \\(tr\\):" + }, + { + "type": "equation", + "bbox": [ + 0.558, + 0.759, + 0.892, + 0.8 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {f s} = \\frac {1}{| R _ {d} |} \\sum_ {r \\in R _ {d}} \\frac {1}{| S _ {r} ^ {f s} |} \\sum_ {p \\in S _ {r} ^ {f s}} (s _ {p} - t r) ^ {2}. \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.811, + 0.892, + 0.857 + ], + "angle": 0, + "content": "To prevent the noisy reconstructions caused by hash collisions in unobserved free-space regions we perform additional regularization on the interpolated features \\(\\mathcal{V}_{\\alpha}(\\mathbf{x})\\) .." + }, + { + "type": "equation", + "bbox": [ + 0.573, + 0.867, + 0.892, + 0.904 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {s m o o t h}} = \\frac {1}{| \\mathcal {G} |} \\sum_ {\\mathbf {x} \\in \\mathcal {G}} \\Delta_ {x} ^ {2} + \\Delta_ {y} ^ {2} + \\Delta_ {z} ^ {2}, \\tag {9}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "13296" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.182 + ], + "angle": 0, + "content": "where \\(\\Delta_{x,y,z} = \\mathcal{V}_{\\alpha}(\\mathbf{x} + \\epsilon_{x,y,z}) - \\mathcal{V}_{\\alpha}(\\mathbf{x})\\) denotes the feature-metric difference between adjacent sampled vertices on the hash-grid along the three dimensions. Since performing regularization on the entire feature grid is computationally infeasible for real-time mapping, we only perform it in a small random region in each iteration." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.183, + 0.472, + 0.243 + ], + "angle": 0, + "content": "Camera Tracking. We track the camera-to-world transformation matrix \\(\\mathbf{T}_{wc} = \\exp (\\xi_t^{\\wedge})\\in \\mathbb{S}\\mathbb{E}(3)\\) at each frame \\(t\\) When a new frame comes in, we first initialize the pose of the current frame \\(i\\) using constant speed assumption:" + }, + { + "type": "equation", + "bbox": [ + 0.197, + 0.253, + 0.469, + 0.271 + ], + "angle": 0, + "content": "\\[\n\\mathbf {T} _ {t} = \\mathbf {T} _ {t - 1} \\mathbf {T} _ {t - 2} ^ {- 1} \\mathbf {T} _ {t - 1} \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.281, + 0.469, + 0.326 + ], + "angle": 0, + "content": "Then, we select \\(N_{t}\\) pixels within the current frame and iteratively optimize the pose by minimizing our objective function with respect to the camera parameters \\(\\xi_{t}\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.327, + 0.469, + 0.537 + ], + "angle": 0, + "content": "Bundle Adjustment. In neural SLAM, bundle adjustment usually consists of keyframe selection and joint optimization of camera poses and scene representation. Classic dense visual SLAM methods require saving keyframe (KF) images as the loss is formulated densely over all pixels. In contrast, the advantage of neural SLAM, as first shown by iMAP [26], is that BA can work with a sparse set of sampled rays. This is because the scene is represented as an implicit field using a neural network. However, iMAP and NICE-SLAM do not take full advantage of this - they still store full keyframe images following the classic SLAM paradigm and rely on keyframe selection (e.g. information gain, visual overlapping) to perform joint optimization on a small fraction of keyframes (usually less than 10)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.538, + 0.47, + 0.765 + ], + "angle": 0, + "content": "In Co-SLAM, we go further and drop the need for storing full keyframe images or keyframe selection. Instead, we only store a subset of pixels (around \\(5\\%\\)) to represent each keyframe. This allows us to insert new keyframes more frequently and maintain a much larger keyframe database. For joint optimization, we randomly sample a total number of \\(N_{g}\\) rays from our global keyframe list to optimize our scene representation as well as camera poses. The joint optimization is performed in an alternating fashion. Specifically, we firstly optimize the scene representation \\(\\theta\\) for \\(k_{m}\\) steps and update camera poses using the accumulated gradient on camera parameters \\(\\{\\xi_t\\}\\). Since each camera pose uses only 6 parameters, this approach can improve the robustness of camera pose optimization with negligible extra computational cost on gradient accumulation." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.778, + 0.21, + 0.794 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.802, + 0.267, + 0.818 + ], + "angle": 0, + "content": "4.1. Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Datasets. We evaluate Co-SLAM on a variety of scenes from four different datasets. Following iMAP and NICE-SLAM, we quantitatively evaluate the reconstruction quality on 8 synthetic scenes from Replica [24]. We also evaluate on 7 synthetic scenes from NeuralRGBD [1], which" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.168 + ], + "angle": 0, + "content": "simulates noisy depth maps. For pose estimation, we evaluate the results on 6 scenes from ScanNet [5] with their ground truth pose obtained with BundleFusion [6], and 3 scenes from TUM RGB-D dataset [25] with their ground truth pose provided by a motion capture system." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.168, + 0.892, + 0.334 + ], + "angle": 0, + "content": "Metrics. We evaluate the reconstruction quality using Depth L1 (cm), Accuracy (cm), Completion (cm), and Completion ratio (\\%) with a threshold of 5cm. Following NICE-SLAM [42], we remove the unobserved regions that are outside of any camera frustum. In addition, we also perform an extra mesh culling that removes the noisy points within the camera frustum but outside the target scene. We observe that all methods experience a performance gain with our mesh culling strategy. Please refer to our supplementary material for more details. For evaluation of camera tracking, we adopt ATE RMSE [25] (cm)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.335, + 0.892, + 0.516 + ], + "angle": 0, + "content": "Baselines. We consider iMAP [26] and NICE-SLAM [42] as our main baselines for reconstruction quality and camera tracking. For a fair comparison, we evaluate iMAP and NICE-SLAM with the same mesh culling strategy as Co-SLAM. Note that iMAP* denotes the iMAP reimplementation released by the NICE-SLAM [42] authors, which is much slower than the original implementation. To investigate the trade-off between accuracy and frame rate on real-world datasets, we report results of two versions of our method: Ours refers to our proposed approach (which achieves real-time operation) while Ours† indicates our method ran with twice as many tracking iterations." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.517, + 0.892, + 0.668 + ], + "angle": 0, + "content": "Implementation Details. We run Co-SLAM on a desktop PC with a 3.60GHz Intel Core i7-12700K CPU and NVIDIA RTX 3090ti GPU. For experiments with default settings (Ours), which runs at 17 FPS on the Replica dataset, we use \\( N_{t} = 1024 \\) pixels with 10 iterations for tracking and \\( 5\\% \\) of pixels from every \\( 5^{th} \\) frame for global bundle adjustment. We sample \\( M_{r} = 32 \\) regular points and \\( M_{d} = 11 \\) depth points along each camera ray, with \\( tr = 10cm \\). Please refer to supplementary materials for more specific settings on all datasets." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.68, + 0.868, + 0.696 + ], + "angle": 0, + "content": "4.2. Evaluation of Tracking and Reconstruction" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.704, + 0.892, + 0.87 + ], + "angle": 0, + "content": "Replica dataset [24]. We evaluate on the same simulated RGB-D sequences as iMAP [26]. As shown in Tab. 1, our method achieves higher reconstruction accuracy and faster speed. Fig. 5 shows the qualitative results, from which we can observe that iMAP achieves plausible completion in unobserved areas but results are over-smoothed, while NICE-SLAM maintains more reconstruction details, but results contain some artifacts (e.g. the floors beside the bed, the back of the chairs). Co-SLAM successfully retains the advantages of both methods achieving consistent completion as well as high-fidelity reconstruction results." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.871, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Synthetic dataset [1]. We perform further experiments on the synthetic dataset from NeuralRGBD [1]. Unlike the" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13297" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.087, + 0.09, + 0.88, + 0.204 + ], + "angle": 0, + "content": "
DatasetMethodDepth L1 (cm)↓Acc. (cm)↓Comp. (cm)↓Comp. Ratio↑Tracking (ms) ↓Mapping (ms) ↓FPS ↑#param. ↓
Replica [24]TSDF-Fusion [4]6.361.623.9483.93N/AN/AN/A16.8 M
iMAP [23]4.643.624.9380.51101 (200, 6)448 (1000, 10)9.90.26 M
NICE-SLAM [42]1.902.372.6491.1378 (200, 10)5470 (1000, 60)0.9117.4 M
Ours1.512.102.0893.4458 (1024, 10)98 (2048, 10)17.40.26 M
Synthetic RGBD [1]TSDF-Fusion [4]10.871.625.1681.52N/AN/AN/A16.8 M
iMAP* [23]43.9118.3026.4120.731550 (5000, 50)14730 (5000, 300)0.340.22 M
NICE-SLAM [42]6.325.965.3077.46123 (1024, 10)3792 (1000, 60)1.313.11 M
Ours3.022.952.9686.8864 (1024, 10)104 (2048, 10)15.60.26 M
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.211, + 0.895, + 0.255 + ], + "angle": 0, + "content": "Table 1. Reconstruction quality and run-time memory comparison on Replica [24] and Synthetic-RGBD [1] with respective settings. TSDF-Fusion is reconstructed with poses estimated by Co-SLAM. Run-time is reported in time (#pixel, #iter) for a comprehensive comparison. The model size is averaged across all scenes." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.263, + 0.466, + 0.385 + ], + "angle": 0, + "content": "
MethodTrack. (ms) ↓Map. (ms) ↓FPS ↑#param. ↓
ScanNetiMAP*30.4×5044.9×3000.370.2 M
NICE-SLAM12.3×50125.3×600.6810.3 M
Ours†7.8×2020.2×106.40.8 M
Ours7.8×1020.2×1012.80.8 M
TUMiMAP*29.6×20044.3×3000.070.2 M
NICE-SLAM47.1×200189.2×600.08101.6 M
Ours†7.5×2019.0×206.71.6 M
Ours7.5×1019.0×2013.31.6 M
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.394, + 0.472, + 0.477 + ], + "angle": 0, + "content": "Table 2. Run-time and memory comparison on ScanNet [5] and TUM-RGBD [25] with respective settings. Run-time is reported in ms/iter \\(\\times\\) #iter for a detailed comparison. NICE-SLAM and iMAP\\* run mapping at every frame on TUM-RGBD. Mapping happens every 5 frames in all other cases. The model size is averaged across all scenes." + }, + { + "type": "image", + "bbox": [ + 0.083, + 0.489, + 0.462, + 0.78 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.782, + 0.47, + 0.826 + ], + "angle": 0, + "content": "Figure 4. Reconstruction results on Synthetic RGB-D dataset [1]. Our method can recover thin structures and achieve plausible scene completion given noisy depth measurement." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Replica dataset, it contains many thin structures and simulates the noise present in real depth sensor measurements. Quantitatively, our method significantly outperforms baseline methods (see Tab. 1) while still operating in real time" + }, + { + "type": "table", + "bbox": [ + 0.504, + 0.263, + 0.89, + 0.328 + ], + "angle": 0, + "content": "
Scene ID000000590106016901810207Avg.
iMAP* [26]55.9532.0617.5070.5132.1011.9136.67
NICE-SLAM [42]8.6412.258.0910.2812.935.599.63
Ours†7.1311.149.365.9011.817.148.75
Ours7.1812.299.576.6213.437.139.37
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.337, + 0.893, + 0.38 + ], + "angle": 0, + "content": "Table 3. ATE RMSE (cm) results of an average of 5 runs on ScanNet. Co-SLAM achieves better or on-par performance compared to NICE-SLAM [42] with significantly faster optimization speed." + }, + { + "type": "table", + "bbox": [ + 0.504, + 0.394, + 0.889, + 0.524 + ], + "angle": 0, + "content": "
fr1/desk (cm)fr2/xyz (cm)fr3/office (cm)
iMAP [26]4.92.05.8
iMAP* [26]7.22.19.0
NICE-SLAM [42]2.71.83.0
Ours†2.41.72.4
Ours2.71.92.6
BAD-SLAM [22]1.71.11.7
Kintinous [34]3.72.93.0
ORB-SLAM2 [17]1.60.41.0
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.528, + 0.892, + 0.571 + ], + "angle": 0, + "content": "Table 4. ATE RSME (cm) results on TUM RGB-D dataset. Our method achieves the best tracking performance among neural SLAM methods and maintains high-fidelity reconstruction." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.583, + 0.892, + 0.673 + ], + "angle": 0, + "content": "(15 FPS). Fig. 4 shows some example qualitative results. Overall, Co-SLAM can capture fine details (e.g. wine bottles, chair legs, etc.) and produces complete and smooth reconstructions. NICE-SLAM yields less detailed and noisier reconstructions and cannot perform hole filling, while iMAP* lost track on some occasions." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.674, + 0.893, + 0.825 + ], + "angle": 0, + "content": "ScanNet dataset [5]. We evaluate the camera tracking accuracy of Co-SLAM on 6 real-world sequences from ScanNet. The absolute trajectory error (ATE) is obtained by comparing predicted and ground-truth (generated by BundleFusion [6]) trajectories. Tab. 3 shows that quantitatively, our method achieves better tracking results in comparison to NICE-SLAM [42] with fewer tracking and mapping iterations while running at \\(6 - 12\\mathrm{Hz}\\) (see Tab. 2). Fig. 6 also shows Co-SLAM achieves better reconstruction quality with smoother results and finer details (e.g. bike)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.825, + 0.894, + 0.903 + ], + "angle": 0, + "content": "TUM dataset [25]. We further evaluate the tracking accuracy on the TUM dataset [25]. As shown in Tab. 4, our method achieves competitive tracking performance at 13 FPS. By increasing the number of tracking iterations \\((\\mathrm{Ours}^{\\dagger})\\), our method achieves the best tracking performance" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "13298" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.088, + 0.088, + 0.881, + 0.392 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.396, + 0.893, + 0.424 + ], + "angle": 0, + "content": "Figure 5. Reconstruction results on Replica [24] dataset. In comparison to our baselines, our methods achieve accurate and high-quality scene reconstruction and completion on various scenes." + }, + { + "type": "image", + "bbox": [ + 0.085, + 0.431, + 0.885, + 0.766 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.771, + 0.893, + 0.8 + ], + "angle": 0, + "content": "Figure 6. Reconstruction results on ScanNet [5]. In comparison to the previous methods, our reconstructions are smoother and contain more details thanks to our proposed joint encoding and global bundle adjustment strategy." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.808, + 0.471, + 0.899 + ], + "angle": 0, + "content": "among neural SLAM methods, though at the expense of the FPS dropping to 6.7 (see Tab. 2). Although Co-SLAM still cannot outperform classic SLAM methods, it reduces the tracking performance gap between neural and classic methods, while improving the fidelity and completeness of the reconstructions." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.807, + 0.707, + 0.823 + ], + "angle": 0, + "content": "4.3. Performance Analysis" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.84, + 0.894, + 0.903 + ], + "angle": 0, + "content": "Run time and memory analysis. In our default setting (Ours), Co-SLAM can operate above \\(15\\mathrm{Hz}\\) on a desktop PC with a 3.60GHz Intel Core i7-12700K CPU and NVIDIA RTX 3090ti GPU. For more challenging scenarios such as" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "13299" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.098, + 0.089, + 0.447, + 0.252 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.257, + 0.47, + 0.312 + ], + "angle": 0, + "content": "Figure 7. Completion ratio vs. model size for Co-SLAM w/ and w/o coordinate encoding. Each model corresponds to a different hash-table size. iMAP [26] and NICE-SLAM [42] shown for reference." + }, + { + "type": "table", + "bbox": [ + 0.09, + 0.322, + 0.455, + 0.384 + ], + "angle": 0, + "content": "
w/o hash enc.w/o one-blob enc.Full model
Acc. (cm.) ↓3.692.132.10
Comp. (cm.) ↓3.432.132.08
Comp. Ratio ↑82.4993.1793.44
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.388, + 0.47, + 0.43 + ], + "angle": 0, + "content": "Table 5. Ablation study on different encodings. Defaultetable size is 13. Our full model with joint encoding achieves better completion and more accurate reconstructions. See also Fig. 2." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.443, + 0.468, + 0.653 + ], + "angle": 0, + "content": "the ScanNet and TUM datasets, Co-SLAM still achieves \\(5 - 13\\mathrm{Hz}\\) runtime. Fig. 7 shows reconstruction quality with respect to memory use. Thanks to the sparse parametric encoding, our method requires significantly less memory than NICE-SLAM [42] while operating in real time and achieving accurate reconstruction results. Surprisingly, we found that further compressing the memory footprint (increasing the chances of hash collisions), Co-SLAM still outperforms iMAP [26], suggesting that our joint encoding improves the representation power of single encoding. Note that this figure is for illustration purposes, so we use the same spatial resolution throughout our hash encoding. Ideally, one can reduce the spatial resolution further to minimize hash collisions and achieve a better reconstruction quality." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.655, + 0.468, + 0.775 + ], + "angle": 0, + "content": "Scene completion. Fig. 2 shows an illustration of hole filling using different encoding strategies on a small scene. Coordinate encoding-based methods achieve plausible completion at the cost of lengthy training times, while parametric encoding-based methods fail at hole filling due to their local nature. By applying our new joint encoding, we observe that smooth hole filling can be achieved and fine structures are preserved by Co-SLAM." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.787, + 0.188, + 0.801 + ], + "angle": 0, + "content": "4.4. Ablations" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.81, + 0.468, + 0.901 + ], + "angle": 0, + "content": "Effect of joint coordinate and parametric encoding. Tab. 5 illustrates a quantitative evaluation using different encodings. Our full model leads to higher accuracy and better completion than using single encodings (only one-blob or only hash-encoding). In addition, Fig. 7 illustrates that when compressing the size of the hash lookup table, our" + }, + { + "type": "table", + "bbox": [ + 0.505, + 0.089, + 0.892, + 0.171 + ], + "angle": 0, + "content": "
NameKF selection#KFPose optim.ATE (cm)Std. (cm)
LocalGlobal010All
w/o BA✓16.811.69
LBA✓✓✓9.691.38
GBA-10✓✓✓9.540.67
GBA✓✓✓8.750.33
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.176, + 0.892, + 0.247 + ], + "angle": 0, + "content": "Table 6. Ablation of BA strategies on Co-SLAM: (LBA) BA with rays from 10 local keyframes; (GBA-10) BA with rays from 10 keyframes randomly selected from all keyframes; (GBA) BA with rays from all keyframes (our full method). All methods sample the same number of total rays per iteration (2048)." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.264, + 0.892, + 0.309 + ], + "angle": 0, + "content": "full model with joint coordinate and parametric encoding is more robust in comparison to using a hash-based feature grid without the coordinate encoding." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.311, + 0.892, + 0.522 + ], + "angle": 0, + "content": "Effect of global bundle adjustment. Tab. 6 shows the average ATE of our SLAM method on the 6 ScanNet scenes using different BA strategies: (w/o BA) pure tracking; (LBA) BA with rays from 10 local keyframes, a similar strategy to NICE-SLAM; (GBA-10) BA using rays from only 10 keyframes randomly selected from all past keyframes; (GBA) denotes the global BA strategy of Co-SLAM. We observe that using rays from a small (10) number of keyframes (LBA and GBA-10) leads to higher ATE errors. However, when keyframes are chosen from the entire sequence (GBA-10), instead of locally (LBA) the standard deviation is greatly reduced. Sampling rays from all keyframes (GBA) is the best overall strategy, even when all methods sample the same number of total rays (2048)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.54, + 0.618, + 0.556 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.567, + 0.892, + 0.657 + ], + "angle": 0, + "content": "We presented Co-SLAM, a dense real-time neural RGB-D SLAM system. We show that using a joint coordinate and parametric encoding with tiny MLPs as scene representation and training it with global bundle adjustment, achieves high-fidelity mapping and accurate tracking with plausible hole filling and efficient memory use." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.659, + 0.892, + 0.765 + ], + "angle": 0, + "content": "Limitations. Co-SLAM relies on inputs from an RGB-D sensor and is therefore sensitive to illumination changes and inaccurate depth measurements. Instead of sampling keyframe pixels randomly, an information-guided pixel sampling strategy could be helpful to further reduce the number of pixels and improve the tracking speed. Incorporating loop closure could lead to further improvements." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.784, + 0.667, + 0.801 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.81, + 0.892, + 0.9 + ], + "angle": 0, + "content": "The research presented here has been supported by a sponsored research award from Cisco Research and the UCL Centre for Doctoral Training in Foundational AI under UKRI grant number EP/S021566/1. This project made use of time on Tier 2 HPC facility JADE2, funded by EPSRC (EP/T022205/1)." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13300" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.105 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.184 + ], + "angle": 0, + "content": "[1] Dejan Azinović, Ricardo Martin-Brualla, Dan B. Goldman, Matthias Nießner, and Justus Thies. Neural RGB-D surface reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6290-6301, 2022. 1, 2, 3, 4, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.186, + 0.472, + 0.282 + ], + "angle": 0, + "content": "[2] Eric R. Chan, Connor Z. Lin, Matthew A. Chan, Koki Nagano, Boxiao Pan, Shalini De Mello, Orazio Gallo, Leonidas J. Guibas, Jonathan Tremblay, and Sameh Khamis. Efficient geometry-aware 3D generative adversarial networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16123-16133, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.284, + 0.471, + 0.325 + ], + "angle": 0, + "content": "[3] Jiawen Chen, Dennis Bautembach, and Shahram Izadi. Scalable real-time volumetric surface reconstruction. ACM Transactions on Graphics (ToG), 32(4):1-16, 2013. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.326, + 0.471, + 0.381 + ], + "angle": 0, + "content": "[4] Brian Curless and Marc Levoy. A volumetric method for building complex models from range images. In Proceedings of the 23rd Annual Conference on Computer Graphics and Interactive Techniques, pages 303-312, 1996. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.382, + 0.47, + 0.464 + ], + "angle": 0, + "content": "[5] Angela Dai, Angel X. Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 5828–5839, 2017. 2, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.466, + 0.47, + 0.535 + ], + "angle": 0, + "content": "[6] Angela Dai, Matthias Nießner, Michael Zollhöfer, Shahram Izadi, and Christian Theobalt. Bundlefusion: Real-time globally consistent 3d reconstruction using on-the-fly surface reintegration. ACM Transactions on Graphics (ToG), 36(4):1, 2017. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.536, + 0.47, + 0.606 + ], + "angle": 0, + "content": "[7] Sara Fridovich-Keil, Alex Yu, Matthew Tancik, Qinhong Chen, Benjamin Recht, and Angjoo Kanazawa. Plenoxels: Radiance Fields Without Neural Networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5501-5510, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.607, + 0.47, + 0.688 + ], + "angle": 0, + "content": "[8] Haoyu Guo, Sida Peng, Haotong Lin, Qianqian Wang, Guofeng Zhang, Hujun Bao, and Xiaowei Zhou. Neural 3D Scene Reconstruction with the Manhattan-world Assumption. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5511-5520, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.69, + 0.47, + 0.759 + ], + "angle": 0, + "content": "[9] Olaf Kähler, Victor Adrian Prisacariu, Carl Yuheng Ren, Xin Sun, Philip H. S. Torr, and David William Murray. Very high frame rate volumetric integration of depth images on mobile devices. IEEE Trans. Vis. Comput. Graph., 21(11):1241-1250, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.761, + 0.469, + 0.815 + ], + "angle": 0, + "content": "[10] Animesh Karnewar, Tobias Ritschel, Oliver Wang, and Niloy Mitra. Relu fields: The little non-linearity that could. In ACM SIGGRAPH 2022 Conference Proceedings, pages 1-9, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.817, + 0.469, + 0.871 + ], + "angle": 0, + "content": "[11] L Koestler, N Yang, N Zeller, and D Cremers. Tandem: Tracking and dense mapping in real-time using deep multiview stereo. In Conference on Robot Learning (CoRL), 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[12] Hai Li, Xingrui Yang, Hongjia Zhai, Yuqian Liu, Hujun Bao, and Guofeng Zhang. Vox-Surf: Voxel-based Implicit Surface" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.119 + ], + "angle": 0, + "content": "Representation. arXiv preprint arXiv:2208.10925, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.122, + 0.892, + 0.176 + ], + "angle": 0, + "content": "[13] Lingjie Liu, Jiatao Gu, Kyaw Zaw Lin, Tat-Seng Chua, and Christian Theobalt. Neural sparse voxel fields. Advances in Neural Information Processing Systems, 33:15651-15663, 2020. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.179, + 0.892, + 0.248 + ], + "angle": 0, + "content": "[14] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In European Conference on Computer Vision, pages 405-421. Springer, 2020. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.25, + 0.892, + 0.305 + ], + "angle": 0, + "content": "[15] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM Transactions on Graphics, 41(4):102:1-102:15, July 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.307, + 0.892, + 0.359 + ], + "angle": 0, + "content": "[16] Thomas Müller, Brian McWilliams, Fabrice Rousselle, Markus Gross, and Jan Novák. Neural importance sampling. ACM Transactions on Graphics (TOG), 38(5):1-19, 2019. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.363, + 0.892, + 0.417 + ], + "angle": 0, + "content": "[17] Raul Mur-Artal and Juan D. Tardós. Orb-slam2: An opensource slam system for monocular, stereo, and rgb-d cameras. IEEE transactions on robotics, 33(5):1255-1262, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.42, + 0.892, + 0.503 + ], + "angle": 0, + "content": "[18] Richard A Newcombe, Shahram Izadi, Otmar Hilliges, David Molyneaux, David Kim, Andrew J Davison, Pushmeet Kohli, Jamie Shotton, Steve Hodges, and Andrew Fitzgibbon. KinectFusion: Real-time dense surface mapping and tracking. In IEEE International Symposium on Mixed and Augmented Reality, pages 127-136, 2011. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.504, + 0.892, + 0.559 + ], + "angle": 0, + "content": "[19] Matthias Nießner, Michael Zollhöfer, Shahram Izadi, and Marc Stamminger. Real-time 3D reconstruction at scale using voxel hashing. ACM Trans. Graph., 32(6):169:1-169:11, 2013. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.562, + 0.892, + 0.629 + ], + "angle": 0, + "content": "[20] Nasim Rahaman, Aristide Baratin, Devansh Arpit, Felix Draxler, Min Lin, Fred Hamprecht, Yoshua Bengio, and Aaron Courville. On the spectral bias of neural networks. In International Conference on Machine Learning, pages 5301-5310. PMLR, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.632, + 0.892, + 0.672 + ], + "angle": 0, + "content": "[21] Thomas Schops, Torsten Sattler, and Marc Pollefeys. BAD SLAM: Bundle adjusted direct RGB-D SLAM. In CVPR, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.675, + 0.894, + 0.731 + ], + "angle": 0, + "content": "[22] Thomas Schops, Torsten Sattler, and Marc Pollefeys. Bad slam: Bundle adjusted direct rgb-d slam. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 134-144, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.732, + 0.892, + 0.788 + ], + "angle": 0, + "content": "[23] Vincent Sitzmann, Julien Martel, Alexander Bergman, David Lindell, and Gordon Wetzstein. Implicit neural representations with periodic activation functions. Advances in Neural Information Processing Systems, 33:7462-7473, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.789, + 0.892, + 0.857 + ], + "angle": 0, + "content": "[24] Julian Straub, Thomas Whelan, Lingni Ma, Yufan Chen, Erik Wijmans, Simon Green, Jakob J. Engel, Raul Mur-Artal, Carl Ren, and Shobhit Verma. The Replica dataset: A digital replica of indoor spaces. arXiv preprint arXiv:1906.05797, 2019. 2, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.859, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[25] Jürgen Sturm, Nikolas Engelhard, Felix Endres, Wolfram Burgard, and Daniel Cremers. A benchmark for the evaluation of RGB-D SLAM systems. In 2012 IEEE/RSJ Interna" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "13301" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.11, + 0.093, + 0.469, + 0.12 + ], + "angle": 0, + "content": "tional Conference on Intelligent Robots and Systems, pages 573-580.IEEE,2012.2,5,6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.122, + 0.47, + 0.19 + ], + "angle": 0, + "content": "[26] Edgar Sucar, Shikun Liu, Joseph Ortiz, and Andrew J. Davison. iMAP: Implicit mapping and positioning in real-time. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6229-6238, 2021. 2, 3, 4, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.192, + 0.471, + 0.261 + ], + "angle": 0, + "content": "[27] Cheng Sun, Min Sun, and Hwann-Tzong Chen. Direct voxel grid optimization: Super-fast convergence for radiance fields reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5459-5469, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.263, + 0.471, + 0.358 + ], + "angle": 0, + "content": "[28] Towaki Takikawa, Joey Litalien, Kangxue Yin, Karsten Kreis, Charles Loop, Derek Nowrouzezahrai, Alec Jacobson, Morgan McGuire, and Sanja Fidler. Neural geometric level of detail: Real-time rendering with implicit 3D shapes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11358-11367, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.361, + 0.469, + 0.403 + ], + "angle": 0, + "content": "[29] K. Tateno, F. Tombari, I. Laina, and N. Navab. CNN-SLAM: Real-time dense monocular slam with learned depth prediction. In CVPR, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.405, + 0.469, + 0.447 + ], + "angle": 0, + "content": "[30] Zachary Teed and Jia Deng. DROID-SLAM: Deep Visual SLAM for Monocular, Stereo, and RGB-D Cameras. Advances in neural information processing systems, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.448, + 0.469, + 0.503 + ], + "angle": 0, + "content": "[31] Emanuele Vespa, Nikolay Nikolov, Marius Grimm, Luigi Nardi, Paul HJ Kelly, and Stefan Leutenegger. Efficient octree-based volumetric SLAM supporting signed-distance and occupancy mapping. RAL, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.505, + 0.469, + 0.559 + ], + "angle": 0, + "content": "[32] Jingwen Wang, Tymoteusz Bleja, and Lourdes Agapito. Go-surf: Neural feature grid optimization for fast, high-fidelity rgb-d surface reconstruction. arXiv preprint arXiv:2206.14735, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.562, + 0.469, + 0.629 + ], + "angle": 0, + "content": "[33] Peng Wang, Lingjie Liu, Yuan Liu, Christian Theobalt, Taku Komura, and Wenping Wang. Neus: Learning neural implicit surfaces by volume rendering for multi-view reconstruction. Advances in Neural Information Processing Systems, 2021. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.632, + 0.469, + 0.673 + ], + "angle": 0, + "content": "[34] Thomas Whelan, Michael Kaess, Maurice Fallon, Hordur Johannsson, John Leonard, and John McDonald. Kintinous: Spatially extended kinectfusion. 2012. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.676, + 0.469, + 0.731 + ], + "angle": 0, + "content": "[35] Thomas Whelan, Stefan Leutenegger, Renato F. Salas-Moreno, Ben Glocker, and Andrew J. Davison. ElasticFusion: Dense SLAM without a pose graph. In Robotics: Science and Systems, volume 11, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.733, + 0.469, + 0.8 + ], + "angle": 0, + "content": "[36] Thomas Whelan, Renato F Salas-Moreno, Ben Glocker, Andrew J Davison, and Stefan Leutenegger. ElasticFusion: Real-time dense SLAM and light source estimation. In International Journal of Robotics Research, volume 35, pages 1697-1716, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.803, + 0.469, + 0.857 + ], + "angle": 0, + "content": "[37] Tong Wu, Jiaqi Wang, Xingang Pan, Xudong Xu, Christian Theobalt, Ziwei Liu, and Dahua Lin. Voxurf: Voxel-based Efficient and Accurate Neural Surface Reconstruction. arXiv preprint arXiv:2208.12697, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.86, + 0.47, + 0.901 + ], + "angle": 0, + "content": "[38] Nan Yang, Lukas von Stumberg, Rui Wang, and Daniel Cremers. D3vo: Deep depth, deep pose and deep uncertainty for monocular visual odometry. In Proceedings of" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.093, + 0.471, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1281-1292, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.122, + 0.892, + 0.176 + ], + "angle": 0, + "content": "[39] Lior Yariv, Jiatao Gu, Yoni Kasten, and Yaron Lipman. Volume rendering of neural implicit surfaces. Advances in Neural Information Processing Systems, 34:4805-4815, 2021. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.179, + 0.892, + 0.246 + ], + "angle": 0, + "content": "[40] Zehao Yu, Songyou Peng, Michael Niemeyer, Torsten Sattler, and Andreas Geiger. Monosdf: Exploring monocular geometric cues for neural implicit surface reconstruction. Advances in Neural Information Processing Systems, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.249, + 0.892, + 0.29 + ], + "angle": 0, + "content": "[41] Ming Zeng, Fukai Zhao, Jiaxiang Zheng, and Xinguo Liu. Octree-based fusion for realtime 3d reconstruction. Graphical Models, 75(3):126-136, 2013. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.292, + 0.892, + 0.373 + ], + "angle": 0, + "content": "[42] Zihan Zhu, Songyou Peng, Viktor Larsson, Weiwei Xu, Hujun Bao, Zhaopeng Cui, Martin R. Oswald, and Marc Pollefeys. Nice-slam: Neural implicit scalable encoding for slam. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12786-12796, 2022. 2, 3, 4, 5, 6, 7, 8" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.373 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13302" + } + ] +] \ No newline at end of file diff --git a/2023/Co-SLAM_ Joint Coordinate and Sparse Parametric Encodings for Neural Real-Time SLAM/98d936fc-c445-4925-9f16-8d4e0f8fc517_origin.pdf b/2023/Co-SLAM_ Joint Coordinate and Sparse Parametric Encodings for Neural Real-Time SLAM/98d936fc-c445-4925-9f16-8d4e0f8fc517_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..a82eba72dd28e48f8291a39f21f27daabffeea85 --- /dev/null +++ b/2023/Co-SLAM_ Joint Coordinate and Sparse Parametric Encodings for Neural Real-Time SLAM/98d936fc-c445-4925-9f16-8d4e0f8fc517_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2e97de8547a8aac4852262877dedc0dfc9f811a091051bee590b03dcba1d20f9 +size 9676439 diff --git a/2023/Co-SLAM_ Joint Coordinate and Sparse Parametric Encodings for Neural Real-Time SLAM/full.md b/2023/Co-SLAM_ Joint Coordinate and Sparse Parametric Encodings for Neural Real-Time SLAM/full.md new file mode 100644 index 0000000000000000000000000000000000000000..f3cce546fe320306f612d43535a528d01a12118a --- /dev/null +++ b/2023/Co-SLAM_ Joint Coordinate and Sparse Parametric Encodings for Neural Real-Time SLAM/full.md @@ -0,0 +1,303 @@ +# Co-SLAM: Joint Coordinate and Sparse Parametric Encodings for Neural Real-Time SLAM + +Hengyi Wang* Jingwen Wang* Lourdes Agapito Department of Computer Science, University College London + +{hengyi.wang.21, jingwen.wang.17, 1.agapito}@ucl.ac.uk + +![](images/8111988349f3fd4a7c740f7c31b3fb88b8c910219b7fa1c4c1bb67a412bb2096.jpg) +Figure 1. We present Co-SLAM, a neural RGB-D SLAM method that performs online tracking and mapping in real time. We propose a new hybrid representation based on a joint coordinate and sparse-parametric encoding with global bundle adjustment. Our method shows fast, high-fidelity scene reconstruction with efficient memory use and plausible hole-filling. + +# Abstract + +We present Co-SLAM, a neural RGB-D SLAM system based on a hybrid representation, that performs robust camera tracking and high-fidelity surface reconstruction in real time. Co-SLAM represents the scene as a multi-resolution hash-grid to exploit its high convergence speed and ability to represent high-frequency local features. In addition, Co-SLAM incorporates one-blob encoding, to encourage surface coherence and completion in unobserved areas. This joint parametric-coordinate encoding enables real-time and robust performance by bringing the best of both worlds: fast convergence and surface hole filling. Moreover, our ray sampling strategy allows Co-SLAM to perform global bundle adjustment over all keyframes instead of requiring keyframe selection to maintain a small number of active keyframes as competing neural SLAM approaches do. Experimental results show that Co-SLAM runs at $10 - 17\mathrm{Hz}$ and achieves state-of-the-art scene reconstruction results, and competitive tracking performance in various datasets and benchmarks (ScanNet, TUM, Replica, Synthetic RGBD). Project page: https://hengyiwang.github.io/projects/CoSLAM + +# 1. Introduction + +Real-time joint camera tracking and dense surface reconstruction from RGB-D sensors has been a core problem in computer vision and robotics for decades. Traditional SLAM solutions exist that can robustly track the position of the camera while fusing depth and/or color measurements into a single high-fidelity map. However, they rely on handcrafted loss terms and do not exploit data-driven priors. + +Recent attention has turned to learning-based models that can exploit the ability of neural network architectures to learn smoothness and coherence priors directly from data. Coordinate-based networks have probably become the most popular representation, since they can be trained to predict the geometric and appearance properties of any point in the scene in a self-supervised way, directly from images. The most notable example, Neural Radiance Fields (NeRF) [14], encodes scene density and color in the weights of a neural network. In combination with volume rendering, NeRF is trained to re-synthesize the input images and has a remarkable ability to generalize to nearby unseen views. + +Coordinate-based networks embed input point coordinates into a high dimensional space, using sinusoidal or other frequency embeddings, allowing them to capture high-frequency details that are essential for high-fidelity geometry reconstruction [1]. Combined with the smooth- + +![](images/2ba0bddad9beabc2c6bae49e057106ca11f3115b09d5938ba98b1af9e30ddece.jpg) +COORDINATE + +![](images/a9c1a9797c2c88598b833a18c8e4d45d6f2cd2627a7b9dca26ffd6a61b4d388a.jpg) +PARAMETRIC + +![](images/0c45c4169fc7c1832ce4cae4d3488c083aa0c7db5c645185826166ae26e71a7a.jpg) +JOINT +DenseGrid+OneBlob + +![](images/f26fe639dca6d1f67e9f783296c256cbcb248cdaa268a2738a784e729607d2bc.jpg) +REFERENCE + +![](images/11ad094643c600a866f02ffbc44b7151dac17206dd1ef0be1c412f1d17995a72.jpg) +Frequency +OneBlob +Figure 2. Illustration of the effect of different encodings on completion. COORDINATE based encodings achieve hole filling but require long training times. PARAMETRIC encodings allow fast training, but fail to complete unobserved regions. JOINT coordinate and parametric encoding (Ours) allows smooth scene completion and fast training. NICE-SLAM [42] uses a dense parametric encoding. + +![](images/b1f72d2dd45e1ccd0da4c20c3931a3f2b9f639752551be2345924165f2f077f0.jpg) +DenseGrid +Grid + +![](images/432fdebd8001049624c51a4fdd2c3e0b09b5e01ad677c0e5c50cafe7bb9c1449.jpg) +HashGrid+OneBlob (Ours) + +![](images/70d111e0030d4375e1d7c7cc86a9a191b1865b3d1ff7c4b4a4b1cb05793338f4.jpg) +NICE-SLAM [42] +GT + +ness and coherence priors inherently encoded in the MLP weights, they constitute a good choice for sequential tracking and mapping [26]. However, the weakness of MLP-based approaches is the long training times required (sometimes hours) to learn a single scene. For that reason, recent real-time capable SLAM systems built on coordinate networks with frequency embeddings such as iMAP [26] need to resort to strategies to sparsify ray sampling and reduce tracking iterations to maintain interactive operation. This comes at the cost of loss of detail in the reconstructions which are oversmoothed (Fig. 5) and potential errors in camera tracking. + +Optimizable feature grids, also known as parametric embeddings, have recently become a powerful alternative scene representation to monolithic MLPs, given their ability to represent high-fidelity local features and their extremely fast convergence (orders of magnitude faster) [7, 10, 15, 32, 40]. Recent efforts focus on sparse alternatives to these parametric embeddings such as octrees [28], tri-plane [2], hash-grid [15] or sparse voxel grid [12, 13] to improve the memory efficiency of dense grids. While these representations can be fast to train and are therefore well suited to real-time operation, they fundamentally lack the smoothness, and coherence priors inherent to MLPs and struggle with hole-filling in areas without observation. NICE-SLAM [42] is a recent example of a multi-resolution feature grid-based SLAM method. Although it does not suffer from over-smoothness and captures local detail (as shown in Fig. 2) it cannot perform hole-filling which might in turn lead to drift in camera pose estimation. + +Our first contribution is to design a joint coordinate and sparse grid encoding for input points that brings together the benefits of both worlds to the real-time SLAM frame- + +work. On the one hand, the smoothness and coherence priors provided by coordinate encodings (we use one-blob [16] encoding), and on the other hand the optimization speed and local details of sparse feature encodings (we use hash grid [15]), resulting in more robust camera tracking and high-fidelity maps with better completion and hole filling. + +Our second contribution relates to the bundle adjustment (BA) step in the joint optimization of the map and camera poses. So far, all neural SLAM systems [26, 42] perform BA using rays sampled from a very small subset of selected keyframes. Restricting the optimization to a very small number of viewpoints results in decreased robustness in camera tracking and increased computation due to the need for a keyframe-selection strategy. Instead, Co-SLAM performs global BA, sampling rays from all past keyframes, which results in an important boost in robustness and performance in pose estimation. In addition, we show that our BA optimization requires a fraction of the iterations of NICE-SLAM [42] to achieve similar errors. In practice, Co-SLAM achieves SOTA performance in camera tracking and 3D reconstruction while maintaining real time performance. + +Co-SLAM runs at 15-17Hz on Replica and Synthetic RGB-D datasets [1], and 12-13Hz on ScanNet [5] and TUM [25] scenes — faster than NICE-SLAM (0.1-1Hz) [42] and iMAP [26]. We perform extensive evaluations on various datasets (Replica [24], Synthetic RGBD [1], ScanNet [6], TUM [25]) where we outperform NICE-SLAM [42] and iMAP [26] in reconstruction and achieve better or at least on-par tracking accuracy. + +# 2. Related Work + +Dense Visual SLAM. Taking advantage of commodity depth sensors, KinectFusion [18] performs frame-to-model + +camera tracking via projective iterative-closest-point (ICP), and incrementally updates the scene geometry via TSDF-Fusion. Subsequent works focused on addressing the scalability issue by adopting more efficient data structures, such as surface elements (surfels) [35, 36], VoxelHashing [3, 9, 19] or Octrees [31, 41]. While most works focus more heavily on scene reconstruction and only track per-frame poses, BAD-SLAM [21] proposes full direct bundle adjustment (BA) to jointly optimize keyframe (KF) poses and the dense 3D structure. Several recent works [11, 29, 30, 38] leverage deep learning to improve the accuracy and robustness of traditional SLAM, and even achieve dense reconstruction with monocular SLAM. While these approaches introduced some learned components, the scene representation and overall pipeline still follow traditional SLAM methods. + +Neural Implicit Representations. Recently neural implicit representations [14] that encode 3D geometry and the appearance of a scene within the weights of a neural network have gained popularity due to their expressiveness and compactness. Among these works, NeRF [14] and its variants adopt coordinate encoding [16, 20] with MLPs and show impressive scene reconstruction using differentiable rendering. Given that coordinate encoding-based methods require lengthy training, many follow-up works [7, 10, 27] propose parametric encodings that increase parameter size but speed up the training. To improve the memory efficiency of parametric encoding-based methods, sparse parametric encodings, such as Octree [28], Tri-plane [2], or sparse voxel grid [12, 13, 15], have been proposed. While these methods focus on novel view synthesis, others instead focus on surface reconstruction from RGB images, with implicit surface representations and differentiable renderers [8, 33, 39, 40]. Other methods [1, 26, 32, 37, 42] use depth measurements as additional supervision for surface reconstruction. + +Neural Implicit SLAM. iMAP [26] adopts an MLP representation to perform joint tracking and mapping in quasi-real time. Smooth, plausible filling of unobserved regions is achieved thanks to the coherence priors inherent to the MLP. iMAP introduces elaborate keyframe selection and information-guided pixel sampling for speed, resulting in $10\mathrm{Hz}$ tracking and $2\mathrm{Hz}$ mapping. To reduce the computational overhead and improve scalability, NICE-SLAM [42] adopts a multi-level feature grid for scene representation. However, as feature grids only perform local updates, they fail to achieve plausible hole-filling. With Co-SLAM we aim to address both issues. To achieve real-time performance and memory-efficiency, while maintaining high-fidelity surface reconstruction and plausible hole filling, we propose to combine the use of coordinate and sparse parametric encodings for scene representation, and perform dense global bundle adjustment using rays sampled from all keyframes. + +# 3. Method + +Fig. 3 shows an overview of Co-SLAM. Given an input RGB-D stream $\{I_t\}_{t=1}^N \{D_t\}_{t=1}^N$ with known camera intrinsics $K \in \mathbb{R}^{3 \times 3}$ , we perform dense mapping and tracking by jointly optimizing camera poses $\{\xi_t\}_{t=1}^N$ and a neural scene representation $f_\theta$ . Specifically, our implicit representation maps world coordinates $\mathbf{x}$ into color $\mathbf{c}$ and truncated signed distance (TSDF) $s$ values: + +$$ +f _ {\theta} (\mathbf {x}) \mapsto (\mathbf {c}, s). \tag {1} +$$ + +Similar to most SLAM systems, the process is split into tracking and mapping. Initialization is performed by running a few training iterations on the first frame. For each subsequent frame, the camera pose is optimized first, initialized with a simple constant-speed motion model. A small fraction of pixels/rays are then sampled and copied to the global pixel-set. At each mapping iteration, global bundle adjustment is performed over a set of pixels randomly sampled from the global pixel-set, to jointly optimize the scene representation $\theta$ and all camera poses $\{\xi_t\}$ . + +# 3.1. Joint Coordinate and Parametric Encoding + +Thanks to the coherence and smoothness priors inherent to MLPs, coordinate-based representations achieve high-fidelity scene reconstruction. However, these methods often suffer from slow convergence and catastrophic forgetting, when optimized in a sequential setting. Instead, parametric encoding based methods improve the computational efficiency, but they fall short of hole filling and smoothness. Since both properties of speed and coherence are crucial for a real-world SLAM system, we propose a joint coordinate and parametric encoding that combines the best of both worlds: we adopt coordinate encoding for scene representation while using sparse parametric encoding to speed up training. Specifically, we use One-blob [16] encoding $\gamma (\mathbf{x})$ instead of embedding spatial coordinates into multiple frequency bands. As scene representation we adopt a multiresolution hash-based feature grid [15] $\mathcal{V}_{\alpha} = \{\mathcal{V}_{\alpha}^{l}\}_{l = 1}^{L}$ . The spatial resolution of each level is set between the coarsest $R_{min}$ and the finest resolution $R_{max}$ in a progressive manner. Feature vectors $\mathcal{V}_{\alpha}(\mathbf{x})$ at each sampled point $\mathbf{x}$ are queried via trilinear interpolation. The geometry decoder outputs the predicted SDF value $s$ and a feature vector $\mathbf{h}$ : + +$$ +f _ {\tau} (\gamma (\mathbf {x}), \mathcal {V} _ {\alpha} (\mathbf {x})) \mapsto (\mathbf {h}, s). \tag {2} +$$ + +Finally, the color MLP predicts the RGB value: + +$$ +f _ {\phi} (\gamma (\mathbf {x}), \mathbf {h}) \mapsto \mathbf {c}. \tag {3} +$$ + +Here $\theta = \{\alpha, \phi, \tau\}$ are learnable parameters. Injecting the One-blob encoding in the hash-based multi-resolution feature grid representation, results in fast convergence, efficient memory use, and hole filling needed for online SLAM. + +![](images/a415353fd7bfe788cf3dc708a6db1a6c53309e215917109d95c500cd4bebdb26.jpg) +Figure 3. Overview of Co-SLAM. 1) Scene representation: using our new joint coordinate+parametric encoding, input coordinates are mapped to RGB and SDF values via two shallow MLPs. 2) Tracking: optimize per-frame camera poses $\xi_{t}$ by minimizing losses. 3) Mapping: global bundle adjustment to jointly optimize the scene representation and camera poses taking rays sampled from all keyframes. + +# 3.2. Depth and Color Rendering + +Following [26, 42], we render depth and color by integrating the predicted values along the sampled rays. Specifically, given the camera origin $\mathbf{o}$ and ray direction $\mathbf{r}$ , we uniformly sample $M$ points $\mathbf{x}_i = \mathbf{o} + d_i\mathbf{r}$ , $i \in \{1, \dots, M\}$ with depth values $\{t_1, \dots, t_M\}$ and predicted colors $\{\mathbf{c}_1, \dots, \mathbf{c}_M\}$ and render color and depth as + +$$ +\hat {\mathbf {c}} = \frac {1}{\sum_ {i = 1} ^ {M} w _ {i}} \sum_ {i = 1} ^ {M} w _ {i} \mathbf {c} _ {i}, \quad \hat {d} = \frac {1}{\sum_ {i = 1} ^ {M} w _ {i}} \sum_ {i = 1} ^ {M} w _ {i} d _ {i}, \quad (4) +$$ + +where $\{w_i\}$ are the computed weights along the ray. A conversion function is needed to convert predicted SDF $s_i$ to weight $w_i$ . Instead of adopting the rendering equations proposed in Neus [33, 39], we follow the simple bell-shaped model of [1] and compute weights $w_i$ directly by multiplying the two Sigmoid functions $\sigma(\cdot)$ + +$$ +w _ {i} = \sigma \left(\frac {s _ {i}}{t r}\right) \sigma \left(- \frac {s _ {i}}{t r}\right), \tag {5} +$$ + +where $tr$ is the truncation distance. + +Depth-guided Sampling. For sampling along each ray, we observe that importance sampling does not show significant improvement while slowing down our tracking and mapping. Instead, we use depth-guided sampling: In addition to $M_{c}$ points uniformly sampled between near and far bound, for rays with a valid depth measurement, we further uniformly sample $M_{f}$ near-surface points within the range $[d - d_{s}, d + d_{s}]$ , where $d_{s}$ is a small offset. + +# 3.3. Tracking and Bundle Adjustment + +Objective Functions. Our tracking and bundle adjustment are performed via minimizing our objective functions with + +respect to learnable parameters $\theta$ and camera parameters $\xi_{t}$ . The color and depth rendering losses are $\ell_2$ errors between the rendered results and observations: + +$$ +\mathcal {L} _ {r g b} = \frac {1}{N} \sum_ {n = 1} ^ {N} \left(\hat {\mathbf {c}} _ {n} - \mathbf {c} _ {n}\right) ^ {2}, \mathcal {L} _ {d} = \frac {1}{| R _ {d} |} \sum_ {r \in R _ {d}} \left(\hat {d} _ {r} - D [ u, v ]\right) ^ {2}. \tag {6} +$$ + +where $R_{d}$ is the set of rays that have a valid depth measurement, and $u, v$ is the corresponding pixel on the image plane. To achieve accurate, smooth reconstructions with detailed geometry, we also apply approximate SDF and feature smoothness losses. Specifically, for samples within the truncation region, i.e. points where $|D[u,v] - d| \leq tr$ , we use the distance between the sampled point and its observed depth value as an approximation of the ground-truth SDF value for supervision: + +$$ +\mathcal {L} _ {s d f} = \frac {1}{| R _ {d} |} \sum_ {r \in R _ {d}} \frac {1}{| S _ {r} ^ {t r} |} \sum_ {p \in S _ {r} ^ {t r}} \left(s _ {p} - (D [ u, v ] - d)\right) ^ {2}. \tag {7} +$$ + +For points that are far from the surface $((D[u,v] - d|) > tr)$ , we apply a free-space loss which forces the SDF prediction to be the truncated distance $tr$ : + +$$ +\mathcal {L} _ {f s} = \frac {1}{| R _ {d} |} \sum_ {r \in R _ {d}} \frac {1}{| S _ {r} ^ {f s} |} \sum_ {p \in S _ {r} ^ {f s}} (s _ {p} - t r) ^ {2}. \tag {8} +$$ + +To prevent the noisy reconstructions caused by hash collisions in unobserved free-space regions we perform additional regularization on the interpolated features $\mathcal{V}_{\alpha}(\mathbf{x})$ .. + +$$ +\mathcal {L} _ {\text {s m o o t h}} = \frac {1}{| \mathcal {G} |} \sum_ {\mathbf {x} \in \mathcal {G}} \Delta_ {x} ^ {2} + \Delta_ {y} ^ {2} + \Delta_ {z} ^ {2}, \tag {9} +$$ + +where $\Delta_{x,y,z} = \mathcal{V}_{\alpha}(\mathbf{x} + \epsilon_{x,y,z}) - \mathcal{V}_{\alpha}(\mathbf{x})$ denotes the feature-metric difference between adjacent sampled vertices on the hash-grid along the three dimensions. Since performing regularization on the entire feature grid is computationally infeasible for real-time mapping, we only perform it in a small random region in each iteration. + +Camera Tracking. We track the camera-to-world transformation matrix $\mathbf{T}_{wc} = \exp (\xi_t^{\wedge})\in \mathbb{S}\mathbb{E}(3)$ at each frame $t$ When a new frame comes in, we first initialize the pose of the current frame $i$ using constant speed assumption: + +$$ +\mathbf {T} _ {t} = \mathbf {T} _ {t - 1} \mathbf {T} _ {t - 2} ^ {- 1} \mathbf {T} _ {t - 1} \tag {10} +$$ + +Then, we select $N_{t}$ pixels within the current frame and iteratively optimize the pose by minimizing our objective function with respect to the camera parameters $\xi_{t}$ . + +Bundle Adjustment. In neural SLAM, bundle adjustment usually consists of keyframe selection and joint optimization of camera poses and scene representation. Classic dense visual SLAM methods require saving keyframe (KF) images as the loss is formulated densely over all pixels. In contrast, the advantage of neural SLAM, as first shown by iMAP [26], is that BA can work with a sparse set of sampled rays. This is because the scene is represented as an implicit field using a neural network. However, iMAP and NICE-SLAM do not take full advantage of this - they still store full keyframe images following the classic SLAM paradigm and rely on keyframe selection (e.g. information gain, visual overlapping) to perform joint optimization on a small fraction of keyframes (usually less than 10). + +In Co-SLAM, we go further and drop the need for storing full keyframe images or keyframe selection. Instead, we only store a subset of pixels (around $5\%$ ) to represent each keyframe. This allows us to insert new keyframes more frequently and maintain a much larger keyframe database. For joint optimization, we randomly sample a total number of $N_{g}$ rays from our global keyframe list to optimize our scene representation as well as camera poses. The joint optimization is performed in an alternating fashion. Specifically, we firstly optimize the scene representation $\theta$ for $k_{m}$ steps and update camera poses using the accumulated gradient on camera parameters $\{\xi_t\}$ . Since each camera pose uses only 6 parameters, this approach can improve the robustness of camera pose optimization with negligible extra computational cost on gradient accumulation. + +# 4. Experiments + +# 4.1. Experimental Setup + +Datasets. We evaluate Co-SLAM on a variety of scenes from four different datasets. Following iMAP and NICE-SLAM, we quantitatively evaluate the reconstruction quality on 8 synthetic scenes from Replica [24]. We also evaluate on 7 synthetic scenes from NeuralRGBD [1], which + +simulates noisy depth maps. For pose estimation, we evaluate the results on 6 scenes from ScanNet [5] with their ground truth pose obtained with BundleFusion [6], and 3 scenes from TUM RGB-D dataset [25] with their ground truth pose provided by a motion capture system. + +Metrics. We evaluate the reconstruction quality using Depth L1 (cm), Accuracy (cm), Completion (cm), and Completion ratio (\%) with a threshold of 5cm. Following NICE-SLAM [42], we remove the unobserved regions that are outside of any camera frustum. In addition, we also perform an extra mesh culling that removes the noisy points within the camera frustum but outside the target scene. We observe that all methods experience a performance gain with our mesh culling strategy. Please refer to our supplementary material for more details. For evaluation of camera tracking, we adopt ATE RMSE [25] (cm). + +Baselines. We consider iMAP [26] and NICE-SLAM [42] as our main baselines for reconstruction quality and camera tracking. For a fair comparison, we evaluate iMAP and NICE-SLAM with the same mesh culling strategy as Co-SLAM. Note that iMAP* denotes the iMAP reimplementation released by the NICE-SLAM [42] authors, which is much slower than the original implementation. To investigate the trade-off between accuracy and frame rate on real-world datasets, we report results of two versions of our method: Ours refers to our proposed approach (which achieves real-time operation) while Ours† indicates our method ran with twice as many tracking iterations. + +Implementation Details. We run Co-SLAM on a desktop PC with a 3.60GHz Intel Core i7-12700K CPU and NVIDIA RTX 3090ti GPU. For experiments with default settings (Ours), which runs at 17 FPS on the Replica dataset, we use $N_{t} = 1024$ pixels with 10 iterations for tracking and $5\%$ of pixels from every $5^{th}$ frame for global bundle adjustment. We sample $M_{r} = 32$ regular points and $M_{d} = 11$ depth points along each camera ray, with $tr = 10cm$ . Please refer to supplementary materials for more specific settings on all datasets. + +# 4.2. Evaluation of Tracking and Reconstruction + +Replica dataset [24]. We evaluate on the same simulated RGB-D sequences as iMAP [26]. As shown in Tab. 1, our method achieves higher reconstruction accuracy and faster speed. Fig. 5 shows the qualitative results, from which we can observe that iMAP achieves plausible completion in unobserved areas but results are over-smoothed, while NICE-SLAM maintains more reconstruction details, but results contain some artifacts (e.g. the floors beside the bed, the back of the chairs). Co-SLAM successfully retains the advantages of both methods achieving consistent completion as well as high-fidelity reconstruction results. + +Synthetic dataset [1]. We perform further experiments on the synthetic dataset from NeuralRGBD [1]. Unlike the + +
DatasetMethodDepth L1 (cm)↓Acc. (cm)↓Comp. (cm)↓Comp. Ratio↑Tracking (ms) ↓Mapping (ms) ↓FPS ↑#param. ↓
Replica [24]TSDF-Fusion [4]6.361.623.9483.93N/AN/AN/A16.8 M
iMAP [23]4.643.624.9380.51101 (200, 6)448 (1000, 10)9.90.26 M
NICE-SLAM [42]1.902.372.6491.1378 (200, 10)5470 (1000, 60)0.9117.4 M
Ours1.512.102.0893.4458 (1024, 10)98 (2048, 10)17.40.26 M
Synthetic RGBD [1]TSDF-Fusion [4]10.871.625.1681.52N/AN/AN/A16.8 M
iMAP* [23]43.9118.3026.4120.731550 (5000, 50)14730 (5000, 300)0.340.22 M
NICE-SLAM [42]6.325.965.3077.46123 (1024, 10)3792 (1000, 60)1.313.11 M
Ours3.022.952.9686.8864 (1024, 10)104 (2048, 10)15.60.26 M
+ +Table 1. Reconstruction quality and run-time memory comparison on Replica [24] and Synthetic-RGBD [1] with respective settings. TSDF-Fusion is reconstructed with poses estimated by Co-SLAM. Run-time is reported in time (#pixel, #iter) for a comprehensive comparison. The model size is averaged across all scenes. + +
MethodTrack. (ms) ↓Map. (ms) ↓FPS ↑#param. ↓
ScanNetiMAP*30.4×5044.9×3000.370.2 M
NICE-SLAM12.3×50125.3×600.6810.3 M
Ours†7.8×2020.2×106.40.8 M
Ours7.8×1020.2×1012.80.8 M
TUMiMAP*29.6×20044.3×3000.070.2 M
NICE-SLAM47.1×200189.2×600.08101.6 M
Ours†7.5×2019.0×206.71.6 M
Ours7.5×1019.0×2013.31.6 M
+ +![](images/d10b20d8bc3bfc2b66c23767fb1bfc894ff3eef77b3176aba9e349984c780846.jpg) +Figure 4. Reconstruction results on Synthetic RGB-D dataset [1]. Our method can recover thin structures and achieve plausible scene completion given noisy depth measurement. + +Replica dataset, it contains many thin structures and simulates the noise present in real depth sensor measurements. Quantitatively, our method significantly outperforms baseline methods (see Tab. 1) while still operating in real time + +Table 2. Run-time and memory comparison on ScanNet [5] and TUM-RGBD [25] with respective settings. Run-time is reported in ms/iter $\times$ #iter for a detailed comparison. NICE-SLAM and iMAP\* run mapping at every frame on TUM-RGBD. Mapping happens every 5 frames in all other cases. The model size is averaged across all scenes. + +
Scene ID000000590106016901810207Avg.
iMAP* [26]55.9532.0617.5070.5132.1011.9136.67
NICE-SLAM [42]8.6412.258.0910.2812.935.599.63
Ours†7.1311.149.365.9011.817.148.75
Ours7.1812.299.576.6213.437.139.37
+ +Table 3. ATE RMSE (cm) results of an average of 5 runs on ScanNet. Co-SLAM achieves better or on-par performance compared to NICE-SLAM [42] with significantly faster optimization speed. + +
fr1/desk (cm)fr2/xyz (cm)fr3/office (cm)
iMAP [26]4.92.05.8
iMAP* [26]7.22.19.0
NICE-SLAM [42]2.71.83.0
Ours†2.41.72.4
Ours2.71.92.6
BAD-SLAM [22]1.71.11.7
Kintinous [34]3.72.93.0
ORB-SLAM2 [17]1.60.41.0
+ +Table 4. ATE RSME (cm) results on TUM RGB-D dataset. Our method achieves the best tracking performance among neural SLAM methods and maintains high-fidelity reconstruction. + +(15 FPS). Fig. 4 shows some example qualitative results. Overall, Co-SLAM can capture fine details (e.g. wine bottles, chair legs, etc.) and produces complete and smooth reconstructions. NICE-SLAM yields less detailed and noisier reconstructions and cannot perform hole filling, while iMAP* lost track on some occasions. + +ScanNet dataset [5]. We evaluate the camera tracking accuracy of Co-SLAM on 6 real-world sequences from ScanNet. The absolute trajectory error (ATE) is obtained by comparing predicted and ground-truth (generated by BundleFusion [6]) trajectories. Tab. 3 shows that quantitatively, our method achieves better tracking results in comparison to NICE-SLAM [42] with fewer tracking and mapping iterations while running at $6 - 12\mathrm{Hz}$ (see Tab. 2). Fig. 6 also shows Co-SLAM achieves better reconstruction quality with smoother results and finer details (e.g. bike). + +TUM dataset [25]. We further evaluate the tracking accuracy on the TUM dataset [25]. As shown in Tab. 4, our method achieves competitive tracking performance at 13 FPS. By increasing the number of tracking iterations $(\mathrm{Ours}^{\dagger})$ , our method achieves the best tracking performance + +![](images/356ee828facc0cdf57961b5b966cf106268208b2b266af6892c988897ac39be9.jpg) +Figure 5. Reconstruction results on Replica [24] dataset. In comparison to our baselines, our methods achieve accurate and high-quality scene reconstruction and completion on various scenes. + +![](images/bf6bb8784c679cd7e1a4bd0e64ec1975812ed32587c1b88493dba3a66a0c64a9.jpg) +Figure 6. Reconstruction results on ScanNet [5]. In comparison to the previous methods, our reconstructions are smoother and contain more details thanks to our proposed joint encoding and global bundle adjustment strategy. + +among neural SLAM methods, though at the expense of the FPS dropping to 6.7 (see Tab. 2). Although Co-SLAM still cannot outperform classic SLAM methods, it reduces the tracking performance gap between neural and classic methods, while improving the fidelity and completeness of the reconstructions. + +# 4.3. Performance Analysis + +Run time and memory analysis. In our default setting (Ours), Co-SLAM can operate above $15\mathrm{Hz}$ on a desktop PC with a 3.60GHz Intel Core i7-12700K CPU and NVIDIA RTX 3090ti GPU. For more challenging scenarios such as + +![](images/982816ef5dc39cf889ae88d3955a899477f5192a2f1eedae883c0a73844f4364.jpg) +Figure 7. Completion ratio vs. model size for Co-SLAM w/ and w/o coordinate encoding. Each model corresponds to a different hash-table size. iMAP [26] and NICE-SLAM [42] shown for reference. + +
w/o hash enc.w/o one-blob enc.Full model
Acc. (cm.) ↓3.692.132.10
Comp. (cm.) ↓3.432.132.08
Comp. Ratio ↑82.4993.1793.44
+ +the ScanNet and TUM datasets, Co-SLAM still achieves $5 - 13\mathrm{Hz}$ runtime. Fig. 7 shows reconstruction quality with respect to memory use. Thanks to the sparse parametric encoding, our method requires significantly less memory than NICE-SLAM [42] while operating in real time and achieving accurate reconstruction results. Surprisingly, we found that further compressing the memory footprint (increasing the chances of hash collisions), Co-SLAM still outperforms iMAP [26], suggesting that our joint encoding improves the representation power of single encoding. Note that this figure is for illustration purposes, so we use the same spatial resolution throughout our hash encoding. Ideally, one can reduce the spatial resolution further to minimize hash collisions and achieve a better reconstruction quality. + +Scene completion. Fig. 2 shows an illustration of hole filling using different encoding strategies on a small scene. Coordinate encoding-based methods achieve plausible completion at the cost of lengthy training times, while parametric encoding-based methods fail at hole filling due to their local nature. By applying our new joint encoding, we observe that smooth hole filling can be achieved and fine structures are preserved by Co-SLAM. + +# 4.4. Ablations + +Effect of joint coordinate and parametric encoding. Tab. 5 illustrates a quantitative evaluation using different encodings. Our full model leads to higher accuracy and better completion than using single encodings (only one-blob or only hash-encoding). In addition, Fig. 7 illustrates that when compressing the size of the hash lookup table, our + +Table 5. Ablation study on different encodings. Defaultetable size is 13. Our full model with joint encoding achieves better completion and more accurate reconstructions. See also Fig. 2. + +
NameKF selection#KFPose optim.ATE (cm)Std. (cm)
LocalGlobal010All
w/o BA✓16.811.69
LBA✓✓✓9.691.38
GBA-10✓✓✓9.540.67
GBA✓✓✓8.750.33
+ +Table 6. Ablation of BA strategies on Co-SLAM: (LBA) BA with rays from 10 local keyframes; (GBA-10) BA with rays from 10 keyframes randomly selected from all keyframes; (GBA) BA with rays from all keyframes (our full method). All methods sample the same number of total rays per iteration (2048). + +full model with joint coordinate and parametric encoding is more robust in comparison to using a hash-based feature grid without the coordinate encoding. + +Effect of global bundle adjustment. Tab. 6 shows the average ATE of our SLAM method on the 6 ScanNet scenes using different BA strategies: (w/o BA) pure tracking; (LBA) BA with rays from 10 local keyframes, a similar strategy to NICE-SLAM; (GBA-10) BA using rays from only 10 keyframes randomly selected from all past keyframes; (GBA) denotes the global BA strategy of Co-SLAM. We observe that using rays from a small (10) number of keyframes (LBA and GBA-10) leads to higher ATE errors. However, when keyframes are chosen from the entire sequence (GBA-10), instead of locally (LBA) the standard deviation is greatly reduced. Sampling rays from all keyframes (GBA) is the best overall strategy, even when all methods sample the same number of total rays (2048). + +# 5. Conclusion + +We presented Co-SLAM, a dense real-time neural RGB-D SLAM system. We show that using a joint coordinate and parametric encoding with tiny MLPs as scene representation and training it with global bundle adjustment, achieves high-fidelity mapping and accurate tracking with plausible hole filling and efficient memory use. + +Limitations. Co-SLAM relies on inputs from an RGB-D sensor and is therefore sensitive to illumination changes and inaccurate depth measurements. Instead of sampling keyframe pixels randomly, an information-guided pixel sampling strategy could be helpful to further reduce the number of pixels and improve the tracking speed. Incorporating loop closure could lead to further improvements. + +# Acknowledgements + +The research presented here has been supported by a sponsored research award from Cisco Research and the UCL Centre for Doctoral Training in Foundational AI under UKRI grant number EP/S021566/1. This project made use of time on Tier 2 HPC facility JADE2, funded by EPSRC (EP/T022205/1). + +# References + +[1] Dejan Azinović, Ricardo Martin-Brualla, Dan B. Goldman, Matthias Nießner, and Justus Thies. Neural RGB-D surface reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6290-6301, 2022. 1, 2, 3, 4, 5, 6 +[2] Eric R. Chan, Connor Z. Lin, Matthew A. Chan, Koki Nagano, Boxiao Pan, Shalini De Mello, Orazio Gallo, Leonidas J. Guibas, Jonathan Tremblay, and Sameh Khamis. Efficient geometry-aware 3D generative adversarial networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16123-16133, 2022. 2, 3 +[3] Jiawen Chen, Dennis Bautembach, and Shahram Izadi. Scalable real-time volumetric surface reconstruction. ACM Transactions on Graphics (ToG), 32(4):1-16, 2013. 3 +[4] Brian Curless and Marc Levoy. A volumetric method for building complex models from range images. In Proceedings of the 23rd Annual Conference on Computer Graphics and Interactive Techniques, pages 303-312, 1996. 6 +[5] Angela Dai, Angel X. Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 5828–5839, 2017. 2, 5, 6, 7 +[6] Angela Dai, Matthias Nießner, Michael Zollhöfer, Shahram Izadi, and Christian Theobalt. Bundlefusion: Real-time globally consistent 3d reconstruction using on-the-fly surface reintegration. ACM Transactions on Graphics (ToG), 36(4):1, 2017. 2, 5, 6 +[7] Sara Fridovich-Keil, Alex Yu, Matthew Tancik, Qinhong Chen, Benjamin Recht, and Angjoo Kanazawa. Plenoxels: Radiance Fields Without Neural Networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5501-5510, 2022. 2, 3 +[8] Haoyu Guo, Sida Peng, Haotong Lin, Qianqian Wang, Guofeng Zhang, Hujun Bao, and Xiaowei Zhou. Neural 3D Scene Reconstruction with the Manhattan-world Assumption. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5511-5520, 2022. 3 +[9] Olaf Kähler, Victor Adrian Prisacariu, Carl Yuheng Ren, Xin Sun, Philip H. S. Torr, and David William Murray. Very high frame rate volumetric integration of depth images on mobile devices. IEEE Trans. Vis. Comput. Graph., 21(11):1241-1250, 2015. 3 +[10] Animesh Karnewar, Tobias Ritschel, Oliver Wang, and Niloy Mitra. Relu fields: The little non-linearity that could. In ACM SIGGRAPH 2022 Conference Proceedings, pages 1-9, 2022. 2, 3 +[11] L Koestler, N Yang, N Zeller, and D Cremers. Tandem: Tracking and dense mapping in real-time using deep multiview stereo. In Conference on Robot Learning (CoRL), 2021. 3 +[12] Hai Li, Xingrui Yang, Hongjia Zhai, Yuqian Liu, Hujun Bao, and Guofeng Zhang. Vox-Surf: Voxel-based Implicit Surface + +Representation. arXiv preprint arXiv:2208.10925, 2022. 2, 3 +[13] Lingjie Liu, Jiatao Gu, Kyaw Zaw Lin, Tat-Seng Chua, and Christian Theobalt. Neural sparse voxel fields. Advances in Neural Information Processing Systems, 33:15651-15663, 2020. 2, 3 +[14] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In European Conference on Computer Vision, pages 405-421. Springer, 2020. 1, 3 +[15] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM Transactions on Graphics, 41(4):102:1-102:15, July 2022. 2, 3 +[16] Thomas Müller, Brian McWilliams, Fabrice Rousselle, Markus Gross, and Jan Novák. Neural importance sampling. ACM Transactions on Graphics (TOG), 38(5):1-19, 2019. 2, 3 +[17] Raul Mur-Artal and Juan D. Tardós. Orb-slam2: An opensource slam system for monocular, stereo, and rgb-d cameras. IEEE transactions on robotics, 33(5):1255-1262, 2017. 6 +[18] Richard A Newcombe, Shahram Izadi, Otmar Hilliges, David Molyneaux, David Kim, Andrew J Davison, Pushmeet Kohli, Jamie Shotton, Steve Hodges, and Andrew Fitzgibbon. KinectFusion: Real-time dense surface mapping and tracking. In IEEE International Symposium on Mixed and Augmented Reality, pages 127-136, 2011. 2 +[19] Matthias Nießner, Michael Zollhöfer, Shahram Izadi, and Marc Stamminger. Real-time 3D reconstruction at scale using voxel hashing. ACM Trans. Graph., 32(6):169:1-169:11, 2013. 3 +[20] Nasim Rahaman, Aristide Baratin, Devansh Arpit, Felix Draxler, Min Lin, Fred Hamprecht, Yoshua Bengio, and Aaron Courville. On the spectral bias of neural networks. In International Conference on Machine Learning, pages 5301-5310. PMLR, 2019. 3 +[21] Thomas Schops, Torsten Sattler, and Marc Pollefeys. BAD SLAM: Bundle adjusted direct RGB-D SLAM. In CVPR, 2019. 3 +[22] Thomas Schops, Torsten Sattler, and Marc Pollefeys. Bad slam: Bundle adjusted direct rgb-d slam. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 134-144, 2019. 6 +[23] Vincent Sitzmann, Julien Martel, Alexander Bergman, David Lindell, and Gordon Wetzstein. Implicit neural representations with periodic activation functions. Advances in Neural Information Processing Systems, 33:7462-7473, 2020. 6 +[24] Julian Straub, Thomas Whelan, Lingni Ma, Yufan Chen, Erik Wijmans, Simon Green, Jakob J. Engel, Raul Mur-Artal, Carl Ren, and Shobhit Verma. The Replica dataset: A digital replica of indoor spaces. arXiv preprint arXiv:1906.05797, 2019. 2, 5, 6, 7 +[25] Jürgen Sturm, Nikolas Engelhard, Felix Endres, Wolfram Burgard, and Daniel Cremers. A benchmark for the evaluation of RGB-D SLAM systems. In 2012 IEEE/RSJ Interna + +tional Conference on Intelligent Robots and Systems, pages 573-580.IEEE,2012.2,5,6 +[26] Edgar Sucar, Shikun Liu, Joseph Ortiz, and Andrew J. Davison. iMAP: Implicit mapping and positioning in real-time. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6229-6238, 2021. 2, 3, 4, 5, 6, 7, 8 +[27] Cheng Sun, Min Sun, and Hwann-Tzong Chen. Direct voxel grid optimization: Super-fast convergence for radiance fields reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5459-5469, 2022. 3 +[28] Towaki Takikawa, Joey Litalien, Kangxue Yin, Karsten Kreis, Charles Loop, Derek Nowrouzezahrai, Alec Jacobson, Morgan McGuire, and Sanja Fidler. Neural geometric level of detail: Real-time rendering with implicit 3D shapes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11358-11367, 2021. 2, 3 +[29] K. Tateno, F. Tombari, I. Laina, and N. Navab. CNN-SLAM: Real-time dense monocular slam with learned depth prediction. In CVPR, 2017. 3 +[30] Zachary Teed and Jia Deng. DROID-SLAM: Deep Visual SLAM for Monocular, Stereo, and RGB-D Cameras. Advances in neural information processing systems, 2021. 3 +[31] Emanuele Vespa, Nikolay Nikolov, Marius Grimm, Luigi Nardi, Paul HJ Kelly, and Stefan Leutenegger. Efficient octree-based volumetric SLAM supporting signed-distance and occupancy mapping. RAL, 2018. 3 +[32] Jingwen Wang, Tymoteusz Bleja, and Lourdes Agapito. Go-surf: Neural feature grid optimization for fast, high-fidelity rgb-d surface reconstruction. arXiv preprint arXiv:2206.14735, 2022. 2, 3 +[33] Peng Wang, Lingjie Liu, Yuan Liu, Christian Theobalt, Taku Komura, and Wenping Wang. Neus: Learning neural implicit surfaces by volume rendering for multi-view reconstruction. Advances in Neural Information Processing Systems, 2021. 3, 4 +[34] Thomas Whelan, Michael Kaess, Maurice Fallon, Hordur Johannsson, John Leonard, and John McDonald. Kintinous: Spatially extended kinectfusion. 2012. 6 +[35] Thomas Whelan, Stefan Leutenegger, Renato F. Salas-Moreno, Ben Glocker, and Andrew J. Davison. ElasticFusion: Dense SLAM without a pose graph. In Robotics: Science and Systems, volume 11, 2015. 3 +[36] Thomas Whelan, Renato F Salas-Moreno, Ben Glocker, Andrew J Davison, and Stefan Leutenegger. ElasticFusion: Real-time dense SLAM and light source estimation. In International Journal of Robotics Research, volume 35, pages 1697-1716, 2016. 3 +[37] Tong Wu, Jiaqi Wang, Xingang Pan, Xudong Xu, Christian Theobalt, Ziwei Liu, and Dahua Lin. Voxurf: Voxel-based Efficient and Accurate Neural Surface Reconstruction. arXiv preprint arXiv:2208.12697, 2022. 3 +[38] Nan Yang, Lukas von Stumberg, Rui Wang, and Daniel Cremers. D3vo: Deep depth, deep pose and deep uncertainty for monocular visual odometry. In Proceedings of + +the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1281-1292, 2020. 3 +[39] Lior Yariv, Jiatao Gu, Yoni Kasten, and Yaron Lipman. Volume rendering of neural implicit surfaces. Advances in Neural Information Processing Systems, 34:4805-4815, 2021. 3, 4 +[40] Zehao Yu, Songyou Peng, Michael Niemeyer, Torsten Sattler, and Andreas Geiger. Monosdf: Exploring monocular geometric cues for neural implicit surface reconstruction. Advances in Neural Information Processing Systems, 2022. 2, 3 +[41] Ming Zeng, Fukai Zhao, Jiaxiang Zheng, and Xinguo Liu. Octree-based fusion for realtime 3d reconstruction. Graphical Models, 75(3):126-136, 2013. 3 +[42] Zihan Zhu, Songyou Peng, Viktor Larsson, Weiwei Xu, Hujun Bao, Zhaopeng Cui, Martin R. Oswald, and Marc Pollefeys. Nice-slam: Neural implicit scalable encoding for slam. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12786-12796, 2022. 2, 3, 4, 5, 6, 7, 8 \ No newline at end of file diff --git a/2023/Co-SLAM_ Joint Coordinate and Sparse Parametric Encodings for Neural Real-Time SLAM/images.zip b/2023/Co-SLAM_ Joint Coordinate and Sparse Parametric Encodings for Neural Real-Time SLAM/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..e7dec0ceee66815ec64a96d504f4afdd7bd0df55 --- /dev/null +++ b/2023/Co-SLAM_ Joint Coordinate and Sparse Parametric Encodings for Neural Real-Time SLAM/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:20510bd1293aa55a5172fb9502c0e386091d4dfb1fc91810cf91b15d84b82678 +size 846671 diff --git a/2023/Co-SLAM_ Joint Coordinate and Sparse Parametric Encodings for Neural Real-Time SLAM/layout.json b/2023/Co-SLAM_ Joint Coordinate and Sparse Parametric Encodings for Neural Real-Time SLAM/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..6a9487669efeea18bc4a8257354dfbec7bc1d555 --- /dev/null +++ b/2023/Co-SLAM_ Joint Coordinate and Sparse Parametric Encodings for Neural Real-Time SLAM/layout.json @@ -0,0 +1,7933 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 91, + 103, + 503, + 138 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 103, + 503, + 138 + ], + "spans": [ + { + "bbox": [ + 91, + 103, + 503, + 138 + ], + "type": "text", + "content": "Co-SLAM: Joint Coordinate and Sparse Parametric Encodings for Neural Real-Time SLAM" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 147, + 152, + 446, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 147, + 152, + 446, + 180 + ], + "spans": [ + { + "bbox": [ + 147, + 152, + 446, + 180 + ], + "type": "text", + "content": "Hengyi Wang* Jingwen Wang* Lourdes Agapito Department of Computer Science, University College London" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 149, + 182, + 441, + 194 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 149, + 182, + 441, + 194 + ], + "spans": [ + { + "bbox": [ + 149, + 182, + 441, + 194 + ], + "type": "text", + "content": "{hengyi.wang.21, jingwen.wang.17, 1.agapito}@ucl.ac.uk" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 65, + 209, + 533, + 343 + ], + "blocks": [ + { + "bbox": [ + 65, + 209, + 533, + 343 + ], + "lines": [ + { + "bbox": [ + 65, + 209, + 533, + 343 + ], + "spans": [ + { + "bbox": [ + 65, + 209, + 533, + 343 + ], + "type": "image", + "image_path": "8111988349f3fd4a7c740f7c31b3fb88b8c910219b7fa1c4c1bb67a412bb2096.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 352, + 546, + 386 + ], + "lines": [ + { + "bbox": [ + 46, + 352, + 546, + 386 + ], + "spans": [ + { + "bbox": [ + 46, + 352, + 546, + 386 + ], + "type": "text", + "content": "Figure 1. We present Co-SLAM, a neural RGB-D SLAM method that performs online tracking and mapping in real time. We propose a new hybrid representation based on a joint coordinate and sparse-parametric encoding with global bundle adjustment. Our method shows fast, high-fidelity scene reconstruction with efficient memory use and plausible hole-filling." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 143, + 397, + 192, + 409 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 397, + 192, + 409 + ], + "spans": [ + { + "bbox": [ + 143, + 397, + 192, + 409 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 424, + 290, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 424, + 290, + 676 + ], + "spans": [ + { + "bbox": [ + 46, + 424, + 290, + 676 + ], + "type": "text", + "content": "We present Co-SLAM, a neural RGB-D SLAM system based on a hybrid representation, that performs robust camera tracking and high-fidelity surface reconstruction in real time. Co-SLAM represents the scene as a multi-resolution hash-grid to exploit its high convergence speed and ability to represent high-frequency local features. In addition, Co-SLAM incorporates one-blob encoding, to encourage surface coherence and completion in unobserved areas. This joint parametric-coordinate encoding enables real-time and robust performance by bringing the best of both worlds: fast convergence and surface hole filling. Moreover, our ray sampling strategy allows Co-SLAM to perform global bundle adjustment over all keyframes instead of requiring keyframe selection to maintain a small number of active keyframes as competing neural SLAM approaches do. Experimental results show that Co-SLAM runs at " + }, + { + "bbox": [ + 46, + 424, + 290, + 676 + ], + "type": "inline_equation", + "content": "10 - 17\\mathrm{Hz}" + }, + { + "bbox": [ + 46, + 424, + 290, + 676 + ], + "type": "text", + "content": " and achieves state-of-the-art scene reconstruction results, and competitive tracking performance in various datasets and benchmarks (ScanNet, TUM, Replica, Synthetic RGBD). Project page: https://hengyiwang.github.io/projects/CoSLAM" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 307, + 397, + 386, + 409 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 397, + 386, + 409 + ], + "spans": [ + { + "bbox": [ + 307, + 397, + 386, + 409 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 420, + 546, + 504 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 420, + 546, + 504 + ], + "spans": [ + { + "bbox": [ + 304, + 420, + 546, + 504 + ], + "type": "text", + "content": "Real-time joint camera tracking and dense surface reconstruction from RGB-D sensors has been a core problem in computer vision and robotics for decades. Traditional SLAM solutions exist that can robustly track the position of the camera while fusing depth and/or color measurements into a single high-fidelity map. However, they rely on handcrafted loss terms and do not exploit data-driven priors." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 506, + 547, + 650 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 506, + 547, + 650 + ], + "spans": [ + { + "bbox": [ + 304, + 506, + 547, + 650 + ], + "type": "text", + "content": "Recent attention has turned to learning-based models that can exploit the ability of neural network architectures to learn smoothness and coherence priors directly from data. Coordinate-based networks have probably become the most popular representation, since they can be trained to predict the geometric and appearance properties of any point in the scene in a self-supervised way, directly from images. The most notable example, Neural Radiance Fields (NeRF) [14], encodes scene density and color in the weights of a neural network. In combination with volume rendering, NeRF is trained to re-synthesize the input images and has a remarkable ability to generalize to nearby unseen views." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 653, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 653, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 653, + 547, + 713 + ], + "type": "text", + "content": "Coordinate-based networks embed input point coordinates into a high dimensional space, using sinusoidal or other frequency embeddings, allowing them to capture high-frequency details that are essential for high-fidelity geometry reconstruction [1]. Combined with the smooth-" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 62, + 703, + 162, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 703, + 162, + 712 + ], + "spans": [ + { + "bbox": [ + 62, + 703, + 162, + 712 + ], + "type": "inline_equation", + "content": "\\star" + }, + { + "bbox": [ + 62, + 703, + 162, + 712 + ], + "type": "text", + "content": " Indicates equal contribution." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13293" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 83, + 80, + 187, + 159 + ], + "blocks": [ + { + "bbox": [ + 108, + 70, + 162, + 79 + ], + "lines": [ + { + "bbox": [ + 108, + 70, + 162, + 79 + ], + "spans": [ + { + "bbox": [ + 108, + 70, + 162, + 79 + ], + "type": "text", + "content": "COORDINATE" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 83, + 80, + 187, + 159 + ], + "lines": [ + { + "bbox": [ + 83, + 80, + 187, + 159 + ], + "spans": [ + { + "bbox": [ + 83, + 80, + 187, + 159 + ], + "type": "image", + "image_path": "2ba0bddad9beabc2c6bae49e057106ca11f3115b09d5938ba98b1af9e30ddece.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 189, + 80, + 294, + 159 + ], + "blocks": [ + { + "bbox": [ + 216, + 70, + 268, + 79 + ], + "lines": [ + { + "bbox": [ + 216, + 70, + 268, + 79 + ], + "spans": [ + { + "bbox": [ + 216, + 70, + 268, + 79 + ], + "type": "text", + "content": "PARAMETRIC" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 189, + 80, + 294, + 159 + ], + "lines": [ + { + "bbox": [ + 189, + 80, + 294, + 159 + ], + "spans": [ + { + "bbox": [ + 189, + 80, + 294, + 159 + ], + "type": "image", + "image_path": "a9c1a9797c2c88598b833a18c8e4d45d6f2cd2627a7b9dca26ffd6a61b4d388a.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 296, + 80, + 402, + 159 + ], + "blocks": [ + { + "bbox": [ + 337, + 70, + 361, + 79 + ], + "lines": [ + { + "bbox": [ + 337, + 70, + 361, + 79 + ], + "spans": [ + { + "bbox": [ + 337, + 70, + 361, + 79 + ], + "type": "text", + "content": "JOINT" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 296, + 80, + 402, + 159 + ], + "lines": [ + { + "bbox": [ + 296, + 80, + 402, + 159 + ], + "spans": [ + { + "bbox": [ + 296, + 80, + 402, + 159 + ], + "type": "image", + "image_path": "0c45c4169fc7c1832ce4cae4d3488c083aa0c7db5c645185826166ae26e71a7a.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 318, + 161, + 380, + 169 + ], + "lines": [ + { + "bbox": [ + 318, + 161, + 380, + 169 + ], + "spans": [ + { + "bbox": [ + 318, + 161, + 380, + 169 + ], + "type": "text", + "content": "DenseGrid+OneBlob" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 403, + 80, + 509, + 159 + ], + "blocks": [ + { + "bbox": [ + 432, + 70, + 481, + 79 + ], + "lines": [ + { + "bbox": [ + 432, + 70, + 481, + 79 + ], + "spans": [ + { + "bbox": [ + 432, + 70, + 481, + 79 + ], + "type": "text", + "content": "REFERENCE" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 403, + 80, + 509, + 159 + ], + "lines": [ + { + "bbox": [ + 403, + 80, + 509, + 159 + ], + "spans": [ + { + "bbox": [ + 403, + 80, + 509, + 159 + ], + "type": "image", + "image_path": "f26fe639dca6d1f67e9f783296c256cbcb248cdaa268a2738a784e729607d2bc.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 83, + 171, + 188, + 250 + ], + "blocks": [ + { + "bbox": [ + 119, + 162, + 151, + 171 + ], + "lines": [ + { + "bbox": [ + 119, + 162, + 151, + 171 + ], + "spans": [ + { + "bbox": [ + 119, + 162, + 151, + 171 + ], + "type": "text", + "content": "Frequency" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 83, + 171, + 188, + 250 + ], + "lines": [ + { + "bbox": [ + 83, + 171, + 188, + 250 + ], + "spans": [ + { + "bbox": [ + 83, + 171, + 188, + 250 + ], + "type": "image", + "image_path": "11ad094643c600a866f02ffbc44b7151dac17206dd1ef0be1c412f1d17995a72.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 121, + 251, + 149, + 259 + ], + "lines": [ + { + "bbox": [ + 121, + 251, + 149, + 259 + ], + "spans": [ + { + "bbox": [ + 121, + 251, + 149, + 259 + ], + "type": "text", + "content": "OneBlob" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 266, + 545, + 300 + ], + "lines": [ + { + "bbox": [ + 46, + 266, + 545, + 300 + ], + "spans": [ + { + "bbox": [ + 46, + 266, + 545, + 300 + ], + "type": "text", + "content": "Figure 2. Illustration of the effect of different encodings on completion. COORDINATE based encodings achieve hole filling but require long training times. PARAMETRIC encodings allow fast training, but fail to complete unobserved regions. JOINT coordinate and parametric encoding (Ours) allows smooth scene completion and fast training. NICE-SLAM [42] uses a dense parametric encoding." + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 190, + 171, + 294, + 250 + ], + "blocks": [ + { + "bbox": [ + 226, + 161, + 258, + 170 + ], + "lines": [ + { + "bbox": [ + 226, + 161, + 258, + 170 + ], + "spans": [ + { + "bbox": [ + 226, + 161, + 258, + 170 + ], + "type": "text", + "content": "DenseGrid" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 190, + 171, + 294, + 250 + ], + "lines": [ + { + "bbox": [ + 190, + 171, + 294, + 250 + ], + "spans": [ + { + "bbox": [ + 190, + 171, + 294, + 250 + ], + "type": "image", + "image_path": "b1f72d2dd45e1ccd0da4c20c3931a3f2b9f639752551be2345924165f2f077f0.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 228, + 251, + 257, + 259 + ], + "lines": [ + { + "bbox": [ + 228, + 251, + 257, + 259 + ], + "spans": [ + { + "bbox": [ + 228, + 251, + 257, + 259 + ], + "type": "text", + "content": "Grid" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 296, + 171, + 402, + 250 + ], + "blocks": [ + { + "bbox": [ + 296, + 171, + 402, + 250 + ], + "lines": [ + { + "bbox": [ + 296, + 171, + 402, + 250 + ], + "spans": [ + { + "bbox": [ + 296, + 171, + 402, + 250 + ], + "type": "image", + "image_path": "432fdebd8001049624c51a4fdd2c3e0b09b5e01ad677c0e5c50cafe7bb9c1449.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 310, + 251, + 388, + 260 + ], + "lines": [ + { + "bbox": [ + 310, + 251, + 388, + 260 + ], + "spans": [ + { + "bbox": [ + 310, + 251, + 388, + 260 + ], + "type": "text", + "content": "HashGrid+OneBlob (Ours)" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 403, + 171, + 509, + 250 + ], + "blocks": [ + { + "bbox": [ + 429, + 161, + 483, + 170 + ], + "lines": [ + { + "bbox": [ + 429, + 161, + 483, + 170 + ], + "spans": [ + { + "bbox": [ + 429, + 161, + 483, + 170 + ], + "type": "text", + "content": "NICE-SLAM [42]" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 403, + 171, + 509, + 250 + ], + "lines": [ + { + "bbox": [ + 403, + 171, + 509, + 250 + ], + "spans": [ + { + "bbox": [ + 403, + 171, + 509, + 250 + ], + "type": "image", + "image_path": "70d111e0030d4375e1d7c7cc86a9a191b1865b3d1ff7c4b4a4b1cb05793338f4.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 451, + 252, + 462, + 259 + ], + "lines": [ + { + "bbox": [ + 451, + 252, + 462, + 259 + ], + "spans": [ + { + "bbox": [ + 451, + 252, + 462, + 259 + ], + "type": "text", + "content": "GT" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "bbox": [ + 46, + 312, + 287, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 312, + 287, + 456 + ], + "spans": [ + { + "bbox": [ + 46, + 312, + 287, + 456 + ], + "type": "text", + "content": "ness and coherence priors inherently encoded in the MLP weights, they constitute a good choice for sequential tracking and mapping [26]. However, the weakness of MLP-based approaches is the long training times required (sometimes hours) to learn a single scene. For that reason, recent real-time capable SLAM systems built on coordinate networks with frequency embeddings such as iMAP [26] need to resort to strategies to sparsify ray sampling and reduce tracking iterations to maintain interactive operation. This comes at the cost of loss of detail in the reconstructions which are oversmoothed (Fig. 5) and potential errors in camera tracking." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 46, + 459, + 287, + 674 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 459, + 287, + 674 + ], + "spans": [ + { + "bbox": [ + 46, + 459, + 287, + 674 + ], + "type": "text", + "content": "Optimizable feature grids, also known as parametric embeddings, have recently become a powerful alternative scene representation to monolithic MLPs, given their ability to represent high-fidelity local features and their extremely fast convergence (orders of magnitude faster) [7, 10, 15, 32, 40]. Recent efforts focus on sparse alternatives to these parametric embeddings such as octrees [28], tri-plane [2], hash-grid [15] or sparse voxel grid [12, 13] to improve the memory efficiency of dense grids. While these representations can be fast to train and are therefore well suited to real-time operation, they fundamentally lack the smoothness, and coherence priors inherent to MLPs and struggle with hole-filling in areas without observation. NICE-SLAM [42] is a recent example of a multi-resolution feature grid-based SLAM method. Although it does not suffer from over-smoothness and captures local detail (as shown in Fig. 2) it cannot perform hole-filling which might in turn lead to drift in camera pose estimation." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 46, + 677, + 287, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 287, + 712 + ], + "type": "text", + "content": "Our first contribution is to design a joint coordinate and sparse grid encoding for input points that brings together the benefits of both worlds to the real-time SLAM frame-" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 304, + 312, + 545, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 312, + 545, + 384 + ], + "spans": [ + { + "bbox": [ + 304, + 312, + 545, + 384 + ], + "type": "text", + "content": "work. On the one hand, the smoothness and coherence priors provided by coordinate encodings (we use one-blob [16] encoding), and on the other hand the optimization speed and local details of sparse feature encodings (we use hash grid [15]), resulting in more robust camera tracking and high-fidelity maps with better completion and hole filling." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 304, + 384, + 546, + 563 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 384, + 546, + 563 + ], + "spans": [ + { + "bbox": [ + 304, + 384, + 546, + 563 + ], + "type": "text", + "content": "Our second contribution relates to the bundle adjustment (BA) step in the joint optimization of the map and camera poses. So far, all neural SLAM systems [26, 42] perform BA using rays sampled from a very small subset of selected keyframes. Restricting the optimization to a very small number of viewpoints results in decreased robustness in camera tracking and increased computation due to the need for a keyframe-selection strategy. Instead, Co-SLAM performs global BA, sampling rays from all past keyframes, which results in an important boost in robustness and performance in pose estimation. In addition, we show that our BA optimization requires a fraction of the iterations of NICE-SLAM [42] to achieve similar errors. In practice, Co-SLAM achieves SOTA performance in camera tracking and 3D reconstruction while maintaining real time performance." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 304, + 563, + 545, + 659 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 563, + 545, + 659 + ], + "spans": [ + { + "bbox": [ + 304, + 563, + 545, + 659 + ], + "type": "text", + "content": "Co-SLAM runs at 15-17Hz on Replica and Synthetic RGB-D datasets [1], and 12-13Hz on ScanNet [5] and TUM [25] scenes — faster than NICE-SLAM (0.1-1Hz) [42] and iMAP [26]. We perform extensive evaluations on various datasets (Replica [24], Synthetic RGBD [1], ScanNet [6], TUM [25]) where we outperform NICE-SLAM [42] and iMAP [26] in reconstruction and achieve better or at least on-par tracking accuracy." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 306, + 669, + 392, + 681 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 669, + 392, + 681 + ], + "spans": [ + { + "bbox": [ + 306, + 669, + 392, + 681 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "content": "Dense Visual SLAM. Taking advantage of commodity depth sensors, KinectFusion [18] performs frame-to-model" + } + ] + } + ], + "index": 28 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13294" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 252 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 252 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 252 + ], + "type": "text", + "content": "camera tracking via projective iterative-closest-point (ICP), and incrementally updates the scene geometry via TSDF-Fusion. Subsequent works focused on addressing the scalability issue by adopting more efficient data structures, such as surface elements (surfels) [35, 36], VoxelHashing [3, 9, 19] or Octrees [31, 41]. While most works focus more heavily on scene reconstruction and only track per-frame poses, BAD-SLAM [21] proposes full direct bundle adjustment (BA) to jointly optimize keyframe (KF) poses and the dense 3D structure. Several recent works [11, 29, 30, 38] leverage deep learning to improve the accuracy and robustness of traditional SLAM, and even achieve dense reconstruction with monocular SLAM. While these approaches introduced some learned components, the scene representation and overall pipeline still follow traditional SLAM methods." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 261, + 289, + 489 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 261, + 289, + 489 + ], + "spans": [ + { + "bbox": [ + 46, + 261, + 289, + 489 + ], + "type": "text", + "content": "Neural Implicit Representations. Recently neural implicit representations [14] that encode 3D geometry and the appearance of a scene within the weights of a neural network have gained popularity due to their expressiveness and compactness. Among these works, NeRF [14] and its variants adopt coordinate encoding [16, 20] with MLPs and show impressive scene reconstruction using differentiable rendering. Given that coordinate encoding-based methods require lengthy training, many follow-up works [7, 10, 27] propose parametric encodings that increase parameter size but speed up the training. To improve the memory efficiency of parametric encoding-based methods, sparse parametric encodings, such as Octree [28], Tri-plane [2], or sparse voxel grid [12, 13, 15], have been proposed. While these methods focus on novel view synthesis, others instead focus on surface reconstruction from RGB images, with implicit surface representations and differentiable renderers [8, 33, 39, 40]. Other methods [1, 26, 32, 37, 42] use depth measurements as additional supervision for surface reconstruction." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 498, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 498, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 498, + 289, + 715 + ], + "type": "text", + "content": "Neural Implicit SLAM. iMAP [26] adopts an MLP representation to perform joint tracking and mapping in quasi-real time. Smooth, plausible filling of unobserved regions is achieved thanks to the coherence priors inherent to the MLP. iMAP introduces elaborate keyframe selection and information-guided pixel sampling for speed, resulting in " + }, + { + "bbox": [ + 46, + 498, + 289, + 715 + ], + "type": "inline_equation", + "content": "10\\mathrm{Hz}" + }, + { + "bbox": [ + 46, + 498, + 289, + 715 + ], + "type": "text", + "content": " tracking and " + }, + { + "bbox": [ + 46, + 498, + 289, + 715 + ], + "type": "inline_equation", + "content": "2\\mathrm{Hz}" + }, + { + "bbox": [ + 46, + 498, + 289, + 715 + ], + "type": "text", + "content": " mapping. To reduce the computational overhead and improve scalability, NICE-SLAM [42] adopts a multi-level feature grid for scene representation. However, as feature grids only perform local updates, they fail to achieve plausible hole-filling. With Co-SLAM we aim to address both issues. To achieve real-time performance and memory-efficiency, while maintaining high-fidelity surface reconstruction and plausible hole filling, we propose to combine the use of coordinate and sparse parametric encodings for scene representation, and perform dense global bundle adjustment using rays sampled from all keyframes." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 306, + 71, + 362, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 71, + 362, + 84 + ], + "spans": [ + { + "bbox": [ + 306, + 71, + 362, + 84 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 304, + 91, + 547, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 91, + 547, + 175 + ], + "spans": [ + { + "bbox": [ + 304, + 91, + 547, + 175 + ], + "type": "text", + "content": "Fig. 3 shows an overview of Co-SLAM. Given an input RGB-D stream " + }, + { + "bbox": [ + 304, + 91, + 547, + 175 + ], + "type": "inline_equation", + "content": "\\{I_t\\}_{t=1}^N \\{D_t\\}_{t=1}^N" + }, + { + "bbox": [ + 304, + 91, + 547, + 175 + ], + "type": "text", + "content": " with known camera intrinsics " + }, + { + "bbox": [ + 304, + 91, + 547, + 175 + ], + "type": "inline_equation", + "content": "K \\in \\mathbb{R}^{3 \\times 3}" + }, + { + "bbox": [ + 304, + 91, + 547, + 175 + ], + "type": "text", + "content": ", we perform dense mapping and tracking by jointly optimizing camera poses " + }, + { + "bbox": [ + 304, + 91, + 547, + 175 + ], + "type": "inline_equation", + "content": "\\{\\xi_t\\}_{t=1}^N" + }, + { + "bbox": [ + 304, + 91, + 547, + 175 + ], + "type": "text", + "content": " and a neural scene representation " + }, + { + "bbox": [ + 304, + 91, + 547, + 175 + ], + "type": "inline_equation", + "content": "f_\\theta" + }, + { + "bbox": [ + 304, + 91, + 547, + 175 + ], + "type": "text", + "content": ". Specifically, our implicit representation maps world coordinates " + }, + { + "bbox": [ + 304, + 91, + 547, + 175 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 304, + 91, + 547, + 175 + ], + "type": "text", + "content": " into color " + }, + { + "bbox": [ + 304, + 91, + 547, + 175 + ], + "type": "inline_equation", + "content": "\\mathbf{c}" + }, + { + "bbox": [ + 304, + 91, + 547, + 175 + ], + "type": "text", + "content": " and truncated signed distance (TSDF) " + }, + { + "bbox": [ + 304, + 91, + 547, + 175 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 304, + 91, + 547, + 175 + ], + "type": "text", + "content": " values:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 392, + 182, + 545, + 196 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 392, + 182, + 545, + 196 + ], + "spans": [ + { + "bbox": [ + 392, + 182, + 545, + 196 + ], + "type": "interline_equation", + "content": "f _ {\\theta} (\\mathbf {x}) \\mapsto (\\mathbf {c}, s). \\tag {1}", + "image_path": "fa3814e4b3c7981d9d089c9f015d78354c191804ac4f9eb1c2e2883b3e948cd1.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 202, + 547, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 202, + 547, + 323 + ], + "spans": [ + { + "bbox": [ + 304, + 202, + 547, + 323 + ], + "type": "text", + "content": "Similar to most SLAM systems, the process is split into tracking and mapping. Initialization is performed by running a few training iterations on the first frame. For each subsequent frame, the camera pose is optimized first, initialized with a simple constant-speed motion model. A small fraction of pixels/rays are then sampled and copied to the global pixel-set. At each mapping iteration, global bundle adjustment is performed over a set of pixels randomly sampled from the global pixel-set, to jointly optimize the scene representation " + }, + { + "bbox": [ + 304, + 202, + 547, + 323 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 304, + 202, + 547, + 323 + ], + "type": "text", + "content": " and all camera poses " + }, + { + "bbox": [ + 304, + 202, + 547, + 323 + ], + "type": "inline_equation", + "content": "\\{\\xi_t\\}" + }, + { + "bbox": [ + 304, + 202, + 547, + 323 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 328, + 533, + 342 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 328, + 533, + 342 + ], + "spans": [ + { + "bbox": [ + 305, + 328, + 533, + 342 + ], + "type": "text", + "content": "3.1. Joint Coordinate and Parametric Encoding" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 347, + 547, + 598 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 347, + 547, + 598 + ], + "spans": [ + { + "bbox": [ + 304, + 347, + 547, + 598 + ], + "type": "text", + "content": "Thanks to the coherence and smoothness priors inherent to MLPs, coordinate-based representations achieve high-fidelity scene reconstruction. However, these methods often suffer from slow convergence and catastrophic forgetting, when optimized in a sequential setting. Instead, parametric encoding based methods improve the computational efficiency, but they fall short of hole filling and smoothness. Since both properties of speed and coherence are crucial for a real-world SLAM system, we propose a joint coordinate and parametric encoding that combines the best of both worlds: we adopt coordinate encoding for scene representation while using sparse parametric encoding to speed up training. Specifically, we use One-blob [16] encoding " + }, + { + "bbox": [ + 304, + 347, + 547, + 598 + ], + "type": "inline_equation", + "content": "\\gamma (\\mathbf{x})" + }, + { + "bbox": [ + 304, + 347, + 547, + 598 + ], + "type": "text", + "content": " instead of embedding spatial coordinates into multiple frequency bands. As scene representation we adopt a multiresolution hash-based feature grid [15] " + }, + { + "bbox": [ + 304, + 347, + 547, + 598 + ], + "type": "inline_equation", + "content": "\\mathcal{V}_{\\alpha} = \\{\\mathcal{V}_{\\alpha}^{l}\\}_{l = 1}^{L}" + }, + { + "bbox": [ + 304, + 347, + 547, + 598 + ], + "type": "text", + "content": ". The spatial resolution of each level is set between the coarsest " + }, + { + "bbox": [ + 304, + 347, + 547, + 598 + ], + "type": "inline_equation", + "content": "R_{min}" + }, + { + "bbox": [ + 304, + 347, + 547, + 598 + ], + "type": "text", + "content": " and the finest resolution " + }, + { + "bbox": [ + 304, + 347, + 547, + 598 + ], + "type": "inline_equation", + "content": "R_{max}" + }, + { + "bbox": [ + 304, + 347, + 547, + 598 + ], + "type": "text", + "content": " in a progressive manner. Feature vectors " + }, + { + "bbox": [ + 304, + 347, + 547, + 598 + ], + "type": "inline_equation", + "content": "\\mathcal{V}_{\\alpha}(\\mathbf{x})" + }, + { + "bbox": [ + 304, + 347, + 547, + 598 + ], + "type": "text", + "content": " at each sampled point " + }, + { + "bbox": [ + 304, + 347, + 547, + 598 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 304, + 347, + 547, + 598 + ], + "type": "text", + "content": " are queried via trilinear interpolation. The geometry decoder outputs the predicted SDF value " + }, + { + "bbox": [ + 304, + 347, + 547, + 598 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 304, + 347, + 547, + 598 + ], + "type": "text", + "content": " and a feature vector " + }, + { + "bbox": [ + 304, + 347, + 547, + 598 + ], + "type": "inline_equation", + "content": "\\mathbf{h}" + }, + { + "bbox": [ + 304, + 347, + 547, + 598 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 369, + 605, + 545, + 619 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 369, + 605, + 545, + 619 + ], + "spans": [ + { + "bbox": [ + 369, + 605, + 545, + 619 + ], + "type": "interline_equation", + "content": "f _ {\\tau} (\\gamma (\\mathbf {x}), \\mathcal {V} _ {\\alpha} (\\mathbf {x})) \\mapsto (\\mathbf {h}, s). \\tag {2}", + "image_path": "2e1446adf58ba7c9421722db74ec3d165131af34948d3d94b20122e647e793c5.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 625, + 499, + 638 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 625, + 499, + 638 + ], + "spans": [ + { + "bbox": [ + 306, + 625, + 499, + 638 + ], + "type": "text", + "content": "Finally, the color MLP predicts the RGB value:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 388, + 645, + 545, + 658 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 388, + 645, + 545, + 658 + ], + "spans": [ + { + "bbox": [ + 388, + 645, + 545, + 658 + ], + "type": "interline_equation", + "content": "f _ {\\phi} (\\gamma (\\mathbf {x}), \\mathbf {h}) \\mapsto \\mathbf {c}. \\tag {3}", + "image_path": "ee77b681f441c140ccc01299630dad3cde3b1661c3b6a1ab14d1c20b7f6bb22c.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 665, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 547, + 714 + ], + "type": "text", + "content": "Here " + }, + { + "bbox": [ + 304, + 665, + 547, + 714 + ], + "type": "inline_equation", + "content": "\\theta = \\{\\alpha, \\phi, \\tau\\}" + }, + { + "bbox": [ + 304, + 665, + 547, + 714 + ], + "type": "text", + "content": " are learnable parameters. Injecting the One-blob encoding in the hash-based multi-resolution feature grid representation, results in fast convergence, efficient memory use, and hole filling needed for online SLAM." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "13295" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 60, + 70, + 534, + 266 + ], + "blocks": [ + { + "bbox": [ + 60, + 70, + 534, + 266 + ], + "lines": [ + { + "bbox": [ + 60, + 70, + 534, + 266 + ], + "spans": [ + { + "bbox": [ + 60, + 70, + 534, + 266 + ], + "type": "image", + "image_path": "a415353fd7bfe788cf3dc708a6db1a6c53309e215917109d95c500cd4bebdb26.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 270, + 547, + 304 + ], + "lines": [ + { + "bbox": [ + 46, + 270, + 547, + 304 + ], + "spans": [ + { + "bbox": [ + 46, + 270, + 547, + 304 + ], + "type": "text", + "content": "Figure 3. Overview of Co-SLAM. 1) Scene representation: using our new joint coordinate+parametric encoding, input coordinates are mapped to RGB and SDF values via two shallow MLPs. 2) Tracking: optimize per-frame camera poses " + }, + { + "bbox": [ + 46, + 270, + 547, + 304 + ], + "type": "inline_equation", + "content": "\\xi_{t}" + }, + { + "bbox": [ + 46, + 270, + 547, + 304 + ], + "type": "text", + "content": " by minimizing losses. 3) Mapping: global bundle adjustment to jointly optimize the scene representation and camera poses taking rays sampled from all keyframes." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 313, + 201, + 327 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 313, + 201, + 327 + ], + "spans": [ + { + "bbox": [ + 47, + 313, + 201, + 327 + ], + "type": "text", + "content": "3.2. Depth and Color Rendering" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 331, + 288, + 405 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 331, + 288, + 405 + ], + "spans": [ + { + "bbox": [ + 47, + 331, + 288, + 405 + ], + "type": "text", + "content": "Following [26, 42], we render depth and color by integrating the predicted values along the sampled rays. Specifically, given the camera origin " + }, + { + "bbox": [ + 47, + 331, + 288, + 405 + ], + "type": "inline_equation", + "content": "\\mathbf{o}" + }, + { + "bbox": [ + 47, + 331, + 288, + 405 + ], + "type": "text", + "content": " and ray direction " + }, + { + "bbox": [ + 47, + 331, + 288, + 405 + ], + "type": "inline_equation", + "content": "\\mathbf{r}" + }, + { + "bbox": [ + 47, + 331, + 288, + 405 + ], + "type": "text", + "content": ", we uniformly sample " + }, + { + "bbox": [ + 47, + 331, + 288, + 405 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 47, + 331, + 288, + 405 + ], + "type": "text", + "content": " points " + }, + { + "bbox": [ + 47, + 331, + 288, + 405 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_i = \\mathbf{o} + d_i\\mathbf{r}" + }, + { + "bbox": [ + 47, + 331, + 288, + 405 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 47, + 331, + 288, + 405 + ], + "type": "inline_equation", + "content": "i \\in \\{1, \\dots, M\\}" + }, + { + "bbox": [ + 47, + 331, + 288, + 405 + ], + "type": "text", + "content": " with depth values " + }, + { + "bbox": [ + 47, + 331, + 288, + 405 + ], + "type": "inline_equation", + "content": "\\{t_1, \\dots, t_M\\}" + }, + { + "bbox": [ + 47, + 331, + 288, + 405 + ], + "type": "text", + "content": " and predicted colors " + }, + { + "bbox": [ + 47, + 331, + 288, + 405 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{c}_1, \\dots, \\mathbf{c}_M\\}" + }, + { + "bbox": [ + 47, + 331, + 288, + 405 + ], + "type": "text", + "content": " and render color and depth as" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 411, + 288, + 445 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 411, + 288, + 445 + ], + "spans": [ + { + "bbox": [ + 56, + 411, + 288, + 445 + ], + "type": "interline_equation", + "content": "\\hat {\\mathbf {c}} = \\frac {1}{\\sum_ {i = 1} ^ {M} w _ {i}} \\sum_ {i = 1} ^ {M} w _ {i} \\mathbf {c} _ {i}, \\quad \\hat {d} = \\frac {1}{\\sum_ {i = 1} ^ {M} w _ {i}} \\sum_ {i = 1} ^ {M} w _ {i} d _ {i}, \\quad (4)", + "image_path": "e1739e99db55d7867effb76208b81d42f1e3db54e60edef93efd445b50f413d9.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 451, + 287, + 523 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 451, + 287, + 523 + ], + "spans": [ + { + "bbox": [ + 46, + 451, + 287, + 523 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 451, + 287, + 523 + ], + "type": "inline_equation", + "content": "\\{w_i\\}" + }, + { + "bbox": [ + 46, + 451, + 287, + 523 + ], + "type": "text", + "content": " are the computed weights along the ray. A conversion function is needed to convert predicted SDF " + }, + { + "bbox": [ + 46, + 451, + 287, + 523 + ], + "type": "inline_equation", + "content": "s_i" + }, + { + "bbox": [ + 46, + 451, + 287, + 523 + ], + "type": "text", + "content": " to weight " + }, + { + "bbox": [ + 46, + 451, + 287, + 523 + ], + "type": "inline_equation", + "content": "w_i" + }, + { + "bbox": [ + 46, + 451, + 287, + 523 + ], + "type": "text", + "content": ". Instead of adopting the rendering equations proposed in Neus [33, 39], we follow the simple bell-shaped model of [1] and compute weights " + }, + { + "bbox": [ + 46, + 451, + 287, + 523 + ], + "type": "inline_equation", + "content": "w_i" + }, + { + "bbox": [ + 46, + 451, + 287, + 523 + ], + "type": "text", + "content": " directly by multiplying the two Sigmoid functions " + }, + { + "bbox": [ + 46, + 451, + 287, + 523 + ], + "type": "inline_equation", + "content": "\\sigma(\\cdot)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 116, + 529, + 287, + 552 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 116, + 529, + 287, + 552 + ], + "spans": [ + { + "bbox": [ + 116, + 529, + 287, + 552 + ], + "type": "interline_equation", + "content": "w _ {i} = \\sigma \\left(\\frac {s _ {i}}{t r}\\right) \\sigma \\left(- \\frac {s _ {i}}{t r}\\right), \\tag {5}", + "image_path": "ceca597fd17284dc7a30bde0b5179becc7a491af8f791a82317d9fe0f202009e.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 557, + 189, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 557, + 189, + 567 + ], + "spans": [ + { + "bbox": [ + 47, + 557, + 189, + 567 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 557, + 189, + 567 + ], + "type": "inline_equation", + "content": "tr" + }, + { + "bbox": [ + 47, + 557, + 189, + 567 + ], + "type": "text", + "content": " is the truncation distance." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 568, + 287, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 568, + 287, + 665 + ], + "spans": [ + { + "bbox": [ + 46, + 568, + 287, + 665 + ], + "type": "text", + "content": "Depth-guided Sampling. For sampling along each ray, we observe that importance sampling does not show significant improvement while slowing down our tracking and mapping. Instead, we use depth-guided sampling: In addition to " + }, + { + "bbox": [ + 46, + 568, + 287, + 665 + ], + "type": "inline_equation", + "content": "M_{c}" + }, + { + "bbox": [ + 46, + 568, + 287, + 665 + ], + "type": "text", + "content": " points uniformly sampled between near and far bound, for rays with a valid depth measurement, we further uniformly sample " + }, + { + "bbox": [ + 46, + 568, + 287, + 665 + ], + "type": "inline_equation", + "content": "M_{f}" + }, + { + "bbox": [ + 46, + 568, + 287, + 665 + ], + "type": "text", + "content": " near-surface points within the range " + }, + { + "bbox": [ + 46, + 568, + 287, + 665 + ], + "type": "inline_equation", + "content": "[d - d_{s}, d + d_{s}]" + }, + { + "bbox": [ + 46, + 568, + 287, + 665 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 568, + 287, + 665 + ], + "type": "inline_equation", + "content": "d_{s}" + }, + { + "bbox": [ + 46, + 568, + 287, + 665 + ], + "type": "text", + "content": " is a small offset." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 671, + 228, + 685 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 671, + 228, + 685 + ], + "spans": [ + { + "bbox": [ + 47, + 671, + 228, + 685 + ], + "type": "text", + "content": "3.3. Tracking and Bundle Adjustment" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 689, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 689, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 689, + 288, + 715 + ], + "type": "text", + "content": "Objective Functions. Our tracking and bundle adjustment are performed via minimizing our objective functions with" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 314, + 546, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 314, + 546, + 350 + ], + "spans": [ + { + "bbox": [ + 305, + 314, + 546, + 350 + ], + "type": "text", + "content": "respect to learnable parameters " + }, + { + "bbox": [ + 305, + 314, + 546, + 350 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 305, + 314, + 546, + 350 + ], + "type": "text", + "content": " and camera parameters " + }, + { + "bbox": [ + 305, + 314, + 546, + 350 + ], + "type": "inline_equation", + "content": "\\xi_{t}" + }, + { + "bbox": [ + 305, + 314, + 546, + 350 + ], + "type": "text", + "content": ". The color and depth rendering losses are " + }, + { + "bbox": [ + 305, + 314, + 546, + 350 + ], + "type": "inline_equation", + "content": "\\ell_2" + }, + { + "bbox": [ + 305, + 314, + 546, + 350 + ], + "type": "text", + "content": " errors between the rendered results and observations:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 357, + 545, + 401 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 357, + 545, + 401 + ], + "spans": [ + { + "bbox": [ + 306, + 357, + 545, + 401 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {r g b} = \\frac {1}{N} \\sum_ {n = 1} ^ {N} \\left(\\hat {\\mathbf {c}} _ {n} - \\mathbf {c} _ {n}\\right) ^ {2}, \\mathcal {L} _ {d} = \\frac {1}{| R _ {d} |} \\sum_ {r \\in R _ {d}} \\left(\\hat {d} _ {r} - D [ u, v ]\\right) ^ {2}. \\tag {6}", + "image_path": "a12c41bf29a0f3a779ab3ac22ac76dd922940eb080c1845fe2958e31c3933489.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 402, + 546, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 402, + 546, + 510 + ], + "spans": [ + { + "bbox": [ + 305, + 402, + 546, + 510 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 402, + 546, + 510 + ], + "type": "inline_equation", + "content": "R_{d}" + }, + { + "bbox": [ + 305, + 402, + 546, + 510 + ], + "type": "text", + "content": " is the set of rays that have a valid depth measurement, and " + }, + { + "bbox": [ + 305, + 402, + 546, + 510 + ], + "type": "inline_equation", + "content": "u, v" + }, + { + "bbox": [ + 305, + 402, + 546, + 510 + ], + "type": "text", + "content": " is the corresponding pixel on the image plane. To achieve accurate, smooth reconstructions with detailed geometry, we also apply approximate SDF and feature smoothness losses. Specifically, for samples within the truncation region, i.e. points where " + }, + { + "bbox": [ + 305, + 402, + 546, + 510 + ], + "type": "inline_equation", + "content": "|D[u,v] - d| \\leq tr" + }, + { + "bbox": [ + 305, + 402, + 546, + 510 + ], + "type": "text", + "content": ", we use the distance between the sampled point and its observed depth value as an approximation of the ground-truth SDF value for supervision:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 310, + 518, + 545, + 548 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 310, + 518, + 545, + 548 + ], + "spans": [ + { + "bbox": [ + 310, + 518, + 545, + 548 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {s d f} = \\frac {1}{| R _ {d} |} \\sum_ {r \\in R _ {d}} \\frac {1}{| S _ {r} ^ {t r} |} \\sum_ {p \\in S _ {r} ^ {t r}} \\left(s _ {p} - (D [ u, v ] - d)\\right) ^ {2}. \\tag {7}", + "image_path": "9dd25c141d38e08b4ac7e112a213acf232d00cec07b620297466c7bd375dfc06.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 558, + 545, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 558, + 545, + 594 + ], + "spans": [ + { + "bbox": [ + 305, + 558, + 545, + 594 + ], + "type": "text", + "content": "For points that are far from the surface " + }, + { + "bbox": [ + 305, + 558, + 545, + 594 + ], + "type": "inline_equation", + "content": "((D[u,v] - d|) > tr)" + }, + { + "bbox": [ + 305, + 558, + 545, + 594 + ], + "type": "text", + "content": ", we apply a free-space loss which forces the SDF prediction to be the truncated distance " + }, + { + "bbox": [ + 305, + 558, + 545, + 594 + ], + "type": "inline_equation", + "content": "tr" + }, + { + "bbox": [ + 305, + 558, + 545, + 594 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 341, + 601, + 545, + 633 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 341, + 601, + 545, + 633 + ], + "spans": [ + { + "bbox": [ + 341, + 601, + 545, + 633 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {f s} = \\frac {1}{| R _ {d} |} \\sum_ {r \\in R _ {d}} \\frac {1}{| S _ {r} ^ {f s} |} \\sum_ {p \\in S _ {r} ^ {f s}} (s _ {p} - t r) ^ {2}. \\tag {8}", + "image_path": "3b5407b43ee4e00d8ff2bb085c8def6a021f8c3ee2bd8dc021e95cdb45c16ee2.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 305, + 642, + 545, + 678 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 642, + 545, + 678 + ], + "spans": [ + { + "bbox": [ + 305, + 642, + 545, + 678 + ], + "type": "text", + "content": "To prevent the noisy reconstructions caused by hash collisions in unobserved free-space regions we perform additional regularization on the interpolated features " + }, + { + "bbox": [ + 305, + 642, + 545, + 678 + ], + "type": "inline_equation", + "content": "\\mathcal{V}_{\\alpha}(\\mathbf{x})" + }, + { + "bbox": [ + 305, + 642, + 545, + 678 + ], + "type": "text", + "content": " .." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 350, + 686, + 545, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 350, + 686, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 350, + 686, + 545, + 715 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {s m o o t h}} = \\frac {1}{| \\mathcal {G} |} \\sum_ {\\mathbf {x} \\in \\mathcal {G}} \\Delta_ {x} ^ {2} + \\Delta_ {y} ^ {2} + \\Delta_ {z} ^ {2}, \\tag {9}", + "image_path": "065feed42e19216f66d67806e520cf6a9a033a67a4d7ca9cef4a537229f85639.jpg" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "13296" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 144 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 144 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 72, + 287, + 144 + ], + "type": "inline_equation", + "content": "\\Delta_{x,y,z} = \\mathcal{V}_{\\alpha}(\\mathbf{x} + \\epsilon_{x,y,z}) - \\mathcal{V}_{\\alpha}(\\mathbf{x})" + }, + { + "bbox": [ + 46, + 72, + 287, + 144 + ], + "type": "text", + "content": " denotes the feature-metric difference between adjacent sampled vertices on the hash-grid along the three dimensions. Since performing regularization on the entire feature grid is computationally infeasible for real-time mapping, we only perform it in a small random region in each iteration." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 144, + 288, + 192 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 144, + 288, + 192 + ], + "spans": [ + { + "bbox": [ + 47, + 144, + 288, + 192 + ], + "type": "text", + "content": "Camera Tracking. We track the camera-to-world transformation matrix " + }, + { + "bbox": [ + 47, + 144, + 288, + 192 + ], + "type": "inline_equation", + "content": "\\mathbf{T}_{wc} = \\exp (\\xi_t^{\\wedge})\\in \\mathbb{S}\\mathbb{E}(3)" + }, + { + "bbox": [ + 47, + 144, + 288, + 192 + ], + "type": "text", + "content": " at each frame " + }, + { + "bbox": [ + 47, + 144, + 288, + 192 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 47, + 144, + 288, + 192 + ], + "type": "text", + "content": " When a new frame comes in, we first initialize the pose of the current frame " + }, + { + "bbox": [ + 47, + 144, + 288, + 192 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 47, + 144, + 288, + 192 + ], + "type": "text", + "content": " using constant speed assumption:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 120, + 200, + 287, + 214 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 120, + 200, + 287, + 214 + ], + "spans": [ + { + "bbox": [ + 120, + 200, + 287, + 214 + ], + "type": "interline_equation", + "content": "\\mathbf {T} _ {t} = \\mathbf {T} _ {t - 1} \\mathbf {T} _ {t - 2} ^ {- 1} \\mathbf {T} _ {t - 1} \\tag {10}", + "image_path": "ffda182565e2453a1f0844b66c968e12b5fb4fe551f79d7bfa6585f0c6d1c453.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 222, + 287, + 258 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 222, + 287, + 258 + ], + "spans": [ + { + "bbox": [ + 46, + 222, + 287, + 258 + ], + "type": "text", + "content": "Then, we select " + }, + { + "bbox": [ + 46, + 222, + 287, + 258 + ], + "type": "inline_equation", + "content": "N_{t}" + }, + { + "bbox": [ + 46, + 222, + 287, + 258 + ], + "type": "text", + "content": " pixels within the current frame and iteratively optimize the pose by minimizing our objective function with respect to the camera parameters " + }, + { + "bbox": [ + 46, + 222, + 287, + 258 + ], + "type": "inline_equation", + "content": "\\xi_{t}" + }, + { + "bbox": [ + 46, + 222, + 287, + 258 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 258, + 287, + 425 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 258, + 287, + 425 + ], + "spans": [ + { + "bbox": [ + 46, + 258, + 287, + 425 + ], + "type": "text", + "content": "Bundle Adjustment. In neural SLAM, bundle adjustment usually consists of keyframe selection and joint optimization of camera poses and scene representation. Classic dense visual SLAM methods require saving keyframe (KF) images as the loss is formulated densely over all pixels. In contrast, the advantage of neural SLAM, as first shown by iMAP [26], is that BA can work with a sparse set of sampled rays. This is because the scene is represented as an implicit field using a neural network. However, iMAP and NICE-SLAM do not take full advantage of this - they still store full keyframe images following the classic SLAM paradigm and rely on keyframe selection (e.g. information gain, visual overlapping) to perform joint optimization on a small fraction of keyframes (usually less than 10)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 426, + 287, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 426, + 287, + 605 + ], + "spans": [ + { + "bbox": [ + 46, + 426, + 287, + 605 + ], + "type": "text", + "content": "In Co-SLAM, we go further and drop the need for storing full keyframe images or keyframe selection. Instead, we only store a subset of pixels (around " + }, + { + "bbox": [ + 46, + 426, + 287, + 605 + ], + "type": "inline_equation", + "content": "5\\%" + }, + { + "bbox": [ + 46, + 426, + 287, + 605 + ], + "type": "text", + "content": ") to represent each keyframe. This allows us to insert new keyframes more frequently and maintain a much larger keyframe database. For joint optimization, we randomly sample a total number of " + }, + { + "bbox": [ + 46, + 426, + 287, + 605 + ], + "type": "inline_equation", + "content": "N_{g}" + }, + { + "bbox": [ + 46, + 426, + 287, + 605 + ], + "type": "text", + "content": " rays from our global keyframe list to optimize our scene representation as well as camera poses. The joint optimization is performed in an alternating fashion. Specifically, we firstly optimize the scene representation " + }, + { + "bbox": [ + 46, + 426, + 287, + 605 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 46, + 426, + 287, + 605 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 46, + 426, + 287, + 605 + ], + "type": "inline_equation", + "content": "k_{m}" + }, + { + "bbox": [ + 46, + 426, + 287, + 605 + ], + "type": "text", + "content": " steps and update camera poses using the accumulated gradient on camera parameters " + }, + { + "bbox": [ + 46, + 426, + 287, + 605 + ], + "type": "inline_equation", + "content": "\\{\\xi_t\\}" + }, + { + "bbox": [ + 46, + 426, + 287, + 605 + ], + "type": "text", + "content": ". Since each camera pose uses only 6 parameters, this approach can improve the robustness of camera pose optimization with negligible extra computational cost on gradient accumulation." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 616, + 128, + 628 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 616, + 128, + 628 + ], + "spans": [ + { + "bbox": [ + 47, + 616, + 128, + 628 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 635, + 163, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 635, + 163, + 647 + ], + "spans": [ + { + "bbox": [ + 47, + 635, + 163, + 647 + ], + "type": "text", + "content": "4.1. Experimental Setup" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 653, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 287, + 713 + ], + "type": "text", + "content": "Datasets. We evaluate Co-SLAM on a variety of scenes from four different datasets. Following iMAP and NICE-SLAM, we quantitatively evaluate the reconstruction quality on 8 synthetic scenes from Replica [24]. We also evaluate on 7 synthetic scenes from NeuralRGBD [1], which" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "text", + "content": "simulates noisy depth maps. For pose estimation, we evaluate the results on 6 scenes from ScanNet [5] with their ground truth pose obtained with BundleFusion [6], and 3 scenes from TUM RGB-D dataset [25] with their ground truth pose provided by a motion capture system." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 133, + 545, + 264 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 133, + 545, + 264 + ], + "spans": [ + { + "bbox": [ + 304, + 133, + 545, + 264 + ], + "type": "text", + "content": "Metrics. We evaluate the reconstruction quality using Depth L1 (cm), Accuracy (cm), Completion (cm), and Completion ratio (\\%) with a threshold of 5cm. Following NICE-SLAM [42], we remove the unobserved regions that are outside of any camera frustum. In addition, we also perform an extra mesh culling that removes the noisy points within the camera frustum but outside the target scene. We observe that all methods experience a performance gain with our mesh culling strategy. Please refer to our supplementary material for more details. For evaluation of camera tracking, we adopt ATE RMSE [25] (cm)." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 265, + 545, + 408 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 265, + 545, + 408 + ], + "spans": [ + { + "bbox": [ + 304, + 265, + 545, + 408 + ], + "type": "text", + "content": "Baselines. We consider iMAP [26] and NICE-SLAM [42] as our main baselines for reconstruction quality and camera tracking. For a fair comparison, we evaluate iMAP and NICE-SLAM with the same mesh culling strategy as Co-SLAM. Note that iMAP* denotes the iMAP reimplementation released by the NICE-SLAM [42] authors, which is much slower than the original implementation. To investigate the trade-off between accuracy and frame rate on real-world datasets, we report results of two versions of our method: Ours refers to our proposed approach (which achieves real-time operation) while Ours† indicates our method ran with twice as many tracking iterations." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 409, + 545, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 409, + 545, + 529 + ], + "spans": [ + { + "bbox": [ + 304, + 409, + 545, + 529 + ], + "type": "text", + "content": "Implementation Details. We run Co-SLAM on a desktop PC with a 3.60GHz Intel Core i7-12700K CPU and NVIDIA RTX 3090ti GPU. For experiments with default settings (Ours), which runs at 17 FPS on the Replica dataset, we use " + }, + { + "bbox": [ + 304, + 409, + 545, + 529 + ], + "type": "inline_equation", + "content": "N_{t} = 1024" + }, + { + "bbox": [ + 304, + 409, + 545, + 529 + ], + "type": "text", + "content": " pixels with 10 iterations for tracking and " + }, + { + "bbox": [ + 304, + 409, + 545, + 529 + ], + "type": "inline_equation", + "content": "5\\%" + }, + { + "bbox": [ + 304, + 409, + 545, + 529 + ], + "type": "text", + "content": " of pixels from every " + }, + { + "bbox": [ + 304, + 409, + 545, + 529 + ], + "type": "inline_equation", + "content": "5^{th}" + }, + { + "bbox": [ + 304, + 409, + 545, + 529 + ], + "type": "text", + "content": " frame for global bundle adjustment. We sample " + }, + { + "bbox": [ + 304, + 409, + 545, + 529 + ], + "type": "inline_equation", + "content": "M_{r} = 32" + }, + { + "bbox": [ + 304, + 409, + 545, + 529 + ], + "type": "text", + "content": " regular points and " + }, + { + "bbox": [ + 304, + 409, + 545, + 529 + ], + "type": "inline_equation", + "content": "M_{d} = 11" + }, + { + "bbox": [ + 304, + 409, + 545, + 529 + ], + "type": "text", + "content": " depth points along each camera ray, with " + }, + { + "bbox": [ + 304, + 409, + 545, + 529 + ], + "type": "inline_equation", + "content": "tr = 10cm" + }, + { + "bbox": [ + 304, + 409, + 545, + 529 + ], + "type": "text", + "content": ". Please refer to supplementary materials for more specific settings on all datasets." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 538, + 531, + 551 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 538, + 531, + 551 + ], + "spans": [ + { + "bbox": [ + 305, + 538, + 531, + 551 + ], + "type": "text", + "content": "4.2. Evaluation of Tracking and Reconstruction" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 557, + 545, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 557, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 304, + 557, + 545, + 689 + ], + "type": "text", + "content": "Replica dataset [24]. We evaluate on the same simulated RGB-D sequences as iMAP [26]. As shown in Tab. 1, our method achieves higher reconstruction accuracy and faster speed. Fig. 5 shows the qualitative results, from which we can observe that iMAP achieves plausible completion in unobserved areas but results are over-smoothed, while NICE-SLAM maintains more reconstruction details, but results contain some artifacts (e.g. the floors beside the bed, the back of the chairs). Co-SLAM successfully retains the advantages of both methods achieving consistent completion as well as high-fidelity reconstruction results." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 689, + 545, + 713 + ], + "type": "text", + "content": "Synthetic dataset [1]. We perform further experiments on the synthetic dataset from NeuralRGBD [1]. Unlike the" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13297" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 53, + 71, + 538, + 161 + ], + "blocks": [ + { + "bbox": [ + 53, + 71, + 538, + 161 + ], + "lines": [ + { + "bbox": [ + 53, + 71, + 538, + 161 + ], + "spans": [ + { + "bbox": [ + 53, + 71, + 538, + 161 + ], + "type": "table", + "html": "
DatasetMethodDepth L1 (cm)↓Acc. (cm)↓Comp. (cm)↓Comp. Ratio↑Tracking (ms) ↓Mapping (ms) ↓FPS ↑#param. ↓
Replica [24]TSDF-Fusion [4]6.361.623.9483.93N/AN/AN/A16.8 M
iMAP [23]4.643.624.9380.51101 (200, 6)448 (1000, 10)9.90.26 M
NICE-SLAM [42]1.902.372.6491.1378 (200, 10)5470 (1000, 60)0.9117.4 M
Ours1.512.102.0893.4458 (1024, 10)98 (2048, 10)17.40.26 M
Synthetic RGBD [1]TSDF-Fusion [4]10.871.625.1681.52N/AN/AN/A16.8 M
iMAP* [23]43.9118.3026.4120.731550 (5000, 50)14730 (5000, 300)0.340.22 M
NICE-SLAM [42]6.325.965.3077.46123 (1024, 10)3792 (1000, 60)1.313.11 M
Ours3.022.952.9686.8864 (1024, 10)104 (2048, 10)15.60.26 M
", + "image_path": "4c383fab653da07a7a1877b6d63d8a64fa124846ec18c79790ac03eb8f76e64b.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 50, + 208, + 285, + 304 + ], + "blocks": [ + { + "bbox": [ + 46, + 167, + 547, + 201 + ], + "lines": [ + { + "bbox": [ + 46, + 167, + 547, + 201 + ], + "spans": [ + { + "bbox": [ + 46, + 167, + 547, + 201 + ], + "type": "text", + "content": "Table 1. Reconstruction quality and run-time memory comparison on Replica [24] and Synthetic-RGBD [1] with respective settings. TSDF-Fusion is reconstructed with poses estimated by Co-SLAM. Run-time is reported in time (#pixel, #iter) for a comprehensive comparison. The model size is averaged across all scenes." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 50, + 208, + 285, + 304 + ], + "lines": [ + { + "bbox": [ + 50, + 208, + 285, + 304 + ], + "spans": [ + { + "bbox": [ + 50, + 208, + 285, + 304 + ], + "type": "table", + "html": "
MethodTrack. (ms) ↓Map. (ms) ↓FPS ↑#param. ↓
ScanNetiMAP*30.4×5044.9×3000.370.2 M
NICE-SLAM12.3×50125.3×600.6810.3 M
Ours†7.8×2020.2×106.40.8 M
Ours7.8×1020.2×1012.80.8 M
TUMiMAP*29.6×20044.3×3000.070.2 M
NICE-SLAM47.1×200189.2×600.08101.6 M
Ours†7.5×2019.0×206.71.6 M
Ours7.5×1019.0×2013.31.6 M
", + "image_path": "2df974ef22e9a7cb52d511cd8076970ff2f54f8da091b3c3658303844252d056.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 50, + 387, + 282, + 617 + ], + "blocks": [ + { + "bbox": [ + 50, + 387, + 282, + 617 + ], + "lines": [ + { + "bbox": [ + 50, + 387, + 282, + 617 + ], + "spans": [ + { + "bbox": [ + 50, + 387, + 282, + 617 + ], + "type": "image", + "image_path": "d10b20d8bc3bfc2b66c23767fb1bfc894ff3eef77b3176aba9e349984c780846.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 619, + 287, + 654 + ], + "lines": [ + { + "bbox": [ + 46, + 619, + 287, + 654 + ], + "spans": [ + { + "bbox": [ + 46, + 619, + 287, + 654 + ], + "type": "text", + "content": "Figure 4. Reconstruction results on Synthetic RGB-D dataset [1]. Our method can recover thin structures and achieve plausible scene completion given noisy depth measurement." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "type": "text", + "content": "Replica dataset, it contains many thin structures and simulates the noise present in real depth sensor measurements. Quantitatively, our method significantly outperforms baseline methods (see Tab. 1) while still operating in real time" + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 308, + 208, + 544, + 259 + ], + "blocks": [ + { + "bbox": [ + 46, + 312, + 288, + 377 + ], + "lines": [ + { + "bbox": [ + 46, + 312, + 288, + 377 + ], + "spans": [ + { + "bbox": [ + 46, + 312, + 288, + 377 + ], + "type": "text", + "content": "Table 2. Run-time and memory comparison on ScanNet [5] and TUM-RGBD [25] with respective settings. Run-time is reported in ms/iter " + }, + { + "bbox": [ + 46, + 312, + 288, + 377 + ], + "type": "inline_equation", + "content": "\\times" + }, + { + "bbox": [ + 46, + 312, + 288, + 377 + ], + "type": "text", + "content": " #iter for a detailed comparison. NICE-SLAM and iMAP\\* run mapping at every frame on TUM-RGBD. Mapping happens every 5 frames in all other cases. The model size is averaged across all scenes." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 308, + 208, + 544, + 259 + ], + "lines": [ + { + "bbox": [ + 308, + 208, + 544, + 259 + ], + "spans": [ + { + "bbox": [ + 308, + 208, + 544, + 259 + ], + "type": "table", + "html": "
Scene ID000000590106016901810207Avg.
iMAP* [26]55.9532.0617.5070.5132.1011.9136.67
NICE-SLAM [42]8.6412.258.0910.2812.935.599.63
Ours†7.1311.149.365.9011.817.148.75
Ours7.1812.299.576.6213.437.139.37
", + "image_path": "bb5a4ed5c52591299ca664fb283800007ce89195df09b5d234fd97421ea0fe15.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 308, + 312, + 544, + 415 + ], + "blocks": [ + { + "bbox": [ + 305, + 266, + 546, + 300 + ], + "lines": [ + { + "bbox": [ + 305, + 266, + 546, + 300 + ], + "spans": [ + { + "bbox": [ + 305, + 266, + 546, + 300 + ], + "type": "text", + "content": "Table 3. ATE RMSE (cm) results of an average of 5 runs on ScanNet. Co-SLAM achieves better or on-par performance compared to NICE-SLAM [42] with significantly faster optimization speed." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 308, + 312, + 544, + 415 + ], + "lines": [ + { + "bbox": [ + 308, + 312, + 544, + 415 + ], + "spans": [ + { + "bbox": [ + 308, + 312, + 544, + 415 + ], + "type": "table", + "html": "
fr1/desk (cm)fr2/xyz (cm)fr3/office (cm)
iMAP [26]4.92.05.8
iMAP* [26]7.22.19.0
NICE-SLAM [42]2.71.83.0
Ours†2.41.72.4
Ours2.71.92.6
BAD-SLAM [22]1.71.11.7
Kintinous [34]3.72.93.0
ORB-SLAM2 [17]1.60.41.0
", + "image_path": "09b7cbcc0fd91bb16ffe58eddfa79942f4d15faf70ab09aa38d4976189ff737e.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 418, + 545, + 452 + ], + "lines": [ + { + "bbox": [ + 305, + 418, + 545, + 452 + ], + "spans": [ + { + "bbox": [ + 305, + 418, + 545, + 452 + ], + "type": "text", + "content": "Table 4. ATE RSME (cm) results on TUM RGB-D dataset. Our method achieves the best tracking performance among neural SLAM methods and maintains high-fidelity reconstruction." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 461, + 545, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 461, + 545, + 533 + ], + "spans": [ + { + "bbox": [ + 304, + 461, + 545, + 533 + ], + "type": "text", + "content": "(15 FPS). Fig. 4 shows some example qualitative results. Overall, Co-SLAM can capture fine details (e.g. wine bottles, chair legs, etc.) and produces complete and smooth reconstructions. NICE-SLAM yields less detailed and noisier reconstructions and cannot perform hole filling, while iMAP* lost track on some occasions." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 533, + 546, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 533, + 546, + 653 + ], + "spans": [ + { + "bbox": [ + 304, + 533, + 546, + 653 + ], + "type": "text", + "content": "ScanNet dataset [5]. We evaluate the camera tracking accuracy of Co-SLAM on 6 real-world sequences from ScanNet. The absolute trajectory error (ATE) is obtained by comparing predicted and ground-truth (generated by BundleFusion [6]) trajectories. Tab. 3 shows that quantitatively, our method achieves better tracking results in comparison to NICE-SLAM [42] with fewer tracking and mapping iterations while running at " + }, + { + "bbox": [ + 304, + 533, + 546, + 653 + ], + "type": "inline_equation", + "content": "6 - 12\\mathrm{Hz}" + }, + { + "bbox": [ + 304, + 533, + 546, + 653 + ], + "type": "text", + "content": " (see Tab. 2). Fig. 6 also shows Co-SLAM achieves better reconstruction quality with smoother results and finer details (e.g. bike)." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 653, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 653, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 653, + 547, + 715 + ], + "type": "text", + "content": "TUM dataset [25]. We further evaluate the tracking accuracy on the TUM dataset [25]. As shown in Tab. 4, our method achieves competitive tracking performance at 13 FPS. By increasing the number of tracking iterations " + }, + { + "bbox": [ + 304, + 653, + 547, + 715 + ], + "type": "inline_equation", + "content": "(\\mathrm{Ours}^{\\dagger})" + }, + { + "bbox": [ + 304, + 653, + 547, + 715 + ], + "type": "text", + "content": ", our method achieves the best tracking performance" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "13298" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 53, + 69, + 539, + 310 + ], + "blocks": [ + { + "bbox": [ + 53, + 69, + 539, + 310 + ], + "lines": [ + { + "bbox": [ + 53, + 69, + 539, + 310 + ], + "spans": [ + { + "bbox": [ + 53, + 69, + 539, + 310 + ], + "type": "image", + "image_path": "356ee828facc0cdf57961b5b966cf106268208b2b266af6892c988897ac39be9.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 47, + 313, + 546, + 335 + ], + "lines": [ + { + "bbox": [ + 47, + 313, + 546, + 335 + ], + "spans": [ + { + "bbox": [ + 47, + 313, + 546, + 335 + ], + "type": "text", + "content": "Figure 5. Reconstruction results on Replica [24] dataset. In comparison to our baselines, our methods achieve accurate and high-quality scene reconstruction and completion on various scenes." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 52, + 341, + 541, + 606 + ], + "blocks": [ + { + "bbox": [ + 52, + 341, + 541, + 606 + ], + "lines": [ + { + "bbox": [ + 52, + 341, + 541, + 606 + ], + "spans": [ + { + "bbox": [ + 52, + 341, + 541, + 606 + ], + "type": "image", + "image_path": "bf6bb8784c679cd7e1a4bd0e64ec1975812ed32587c1b88493dba3a66a0c64a9.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 610, + 546, + 633 + ], + "lines": [ + { + "bbox": [ + 46, + 610, + 546, + 633 + ], + "spans": [ + { + "bbox": [ + 46, + 610, + 546, + 633 + ], + "type": "text", + "content": "Figure 6. Reconstruction results on ScanNet [5]. In comparison to the previous methods, our reconstructions are smoother and contain more details thanks to our proposed joint encoding and global bundle adjustment strategy." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 639, + 288, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 639, + 288, + 712 + ], + "spans": [ + { + "bbox": [ + 46, + 639, + 288, + 712 + ], + "type": "text", + "content": "among neural SLAM methods, though at the expense of the FPS dropping to 6.7 (see Tab. 2). Although Co-SLAM still cannot outperform classic SLAM methods, it reduces the tracking performance gap between neural and classic methods, while improving the fidelity and completeness of the reconstructions." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 306, + 639, + 432, + 651 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 639, + 432, + 651 + ], + "spans": [ + { + "bbox": [ + 306, + 639, + 432, + 651 + ], + "type": "text", + "content": "4.3. Performance Analysis" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 305, + 665, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 665, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 305, + 665, + 547, + 715 + ], + "type": "text", + "content": "Run time and memory analysis. In our default setting (Ours), Co-SLAM can operate above " + }, + { + "bbox": [ + 305, + 665, + 547, + 715 + ], + "type": "inline_equation", + "content": "15\\mathrm{Hz}" + }, + { + "bbox": [ + 305, + 665, + 547, + 715 + ], + "type": "text", + "content": " on a desktop PC with a 3.60GHz Intel Core i7-12700K CPU and NVIDIA RTX 3090ti GPU. For more challenging scenarios such as" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "13299" + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 70, + 273, + 199 + ], + "blocks": [ + { + "bbox": [ + 59, + 70, + 273, + 199 + ], + "lines": [ + { + "bbox": [ + 59, + 70, + 273, + 199 + ], + "spans": [ + { + "bbox": [ + 59, + 70, + 273, + 199 + ], + "type": "image", + "image_path": "982816ef5dc39cf889ae88d3955a899477f5192a2f1eedae883c0a73844f4364.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 203, + 287, + 247 + ], + "lines": [ + { + "bbox": [ + 46, + 203, + 287, + 247 + ], + "spans": [ + { + "bbox": [ + 46, + 203, + 287, + 247 + ], + "type": "text", + "content": "Figure 7. Completion ratio vs. model size for Co-SLAM w/ and w/o coordinate encoding. Each model corresponds to a different hash-table size. iMAP [26] and NICE-SLAM [42] shown for reference." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 55, + 255, + 278, + 304 + ], + "blocks": [ + { + "bbox": [ + 55, + 255, + 278, + 304 + ], + "lines": [ + { + "bbox": [ + 55, + 255, + 278, + 304 + ], + "spans": [ + { + "bbox": [ + 55, + 255, + 278, + 304 + ], + "type": "table", + "html": "
w/o hash enc.w/o one-blob enc.Full model
Acc. (cm.) ↓3.692.132.10
Comp. (cm.) ↓3.432.132.08
Comp. Ratio ↑82.4993.1793.44
", + "image_path": "09a456c7b3b11123174424c3b75c1046d442912cebf880e998381dd1cbcb3a10.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 350, + 286, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 350, + 286, + 517 + ], + "spans": [ + { + "bbox": [ + 46, + 350, + 286, + 517 + ], + "type": "text", + "content": "the ScanNet and TUM datasets, Co-SLAM still achieves " + }, + { + "bbox": [ + 46, + 350, + 286, + 517 + ], + "type": "inline_equation", + "content": "5 - 13\\mathrm{Hz}" + }, + { + "bbox": [ + 46, + 350, + 286, + 517 + ], + "type": "text", + "content": " runtime. Fig. 7 shows reconstruction quality with respect to memory use. Thanks to the sparse parametric encoding, our method requires significantly less memory than NICE-SLAM [42] while operating in real time and achieving accurate reconstruction results. Surprisingly, we found that further compressing the memory footprint (increasing the chances of hash collisions), Co-SLAM still outperforms iMAP [26], suggesting that our joint encoding improves the representation power of single encoding. Note that this figure is for illustration purposes, so we use the same spatial resolution throughout our hash encoding. Ideally, one can reduce the spatial resolution further to minimize hash collisions and achieve a better reconstruction quality." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 518, + 286, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 518, + 286, + 613 + ], + "spans": [ + { + "bbox": [ + 46, + 518, + 286, + 613 + ], + "type": "text", + "content": "Scene completion. Fig. 2 shows an illustration of hole filling using different encoding strategies on a small scene. Coordinate encoding-based methods achieve plausible completion at the cost of lengthy training times, while parametric encoding-based methods fail at hole filling due to their local nature. By applying our new joint encoding, we observe that smooth hole filling can be achieved and fine structures are preserved by Co-SLAM." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 623, + 115, + 634 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 623, + 115, + 634 + ], + "spans": [ + { + "bbox": [ + 47, + 623, + 115, + 634 + ], + "type": "text", + "content": "4.4. Ablations" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 641, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 641, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 641, + 286, + 713 + ], + "type": "text", + "content": "Effect of joint coordinate and parametric encoding. Tab. 5 illustrates a quantitative evaluation using different encodings. Our full model leads to higher accuracy and better completion than using single encodings (only one-blob or only hash-encoding). In addition, Fig. 7 illustrates that when compressing the size of the hash lookup table, our" + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 309, + 70, + 545, + 135 + ], + "blocks": [ + { + "bbox": [ + 46, + 307, + 287, + 340 + ], + "lines": [ + { + "bbox": [ + 46, + 307, + 287, + 340 + ], + "spans": [ + { + "bbox": [ + 46, + 307, + 287, + 340 + ], + "type": "text", + "content": "Table 5. Ablation study on different encodings. Defaultetable size is 13. Our full model with joint encoding achieves better completion and more accurate reconstructions. See also Fig. 2." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 309, + 70, + 545, + 135 + ], + "lines": [ + { + "bbox": [ + 309, + 70, + 545, + 135 + ], + "spans": [ + { + "bbox": [ + 309, + 70, + 545, + 135 + ], + "type": "table", + "html": "
NameKF selection#KFPose optim.ATE (cm)Std. (cm)
LocalGlobal010All
w/o BA✓16.811.69
LBA✓✓✓9.691.38
GBA-10✓✓✓9.540.67
GBA✓✓✓8.750.33
", + "image_path": "96ca123e53ddd63153f2e86c70cfff72741afc257549d9f3424a183adcd12faf.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 139, + 545, + 195 + ], + "lines": [ + { + "bbox": [ + 305, + 139, + 545, + 195 + ], + "spans": [ + { + "bbox": [ + 305, + 139, + 545, + 195 + ], + "type": "text", + "content": "Table 6. Ablation of BA strategies on Co-SLAM: (LBA) BA with rays from 10 local keyframes; (GBA-10) BA with rays from 10 keyframes randomly selected from all keyframes; (GBA) BA with rays from all keyframes (our full method). All methods sample the same number of total rays per iteration (2048)." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 305, + 209, + 545, + 244 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 209, + 545, + 244 + ], + "spans": [ + { + "bbox": [ + 305, + 209, + 545, + 244 + ], + "type": "text", + "content": "full model with joint coordinate and parametric encoding is more robust in comparison to using a hash-based feature grid without the coordinate encoding." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 246, + 545, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 246, + 545, + 413 + ], + "spans": [ + { + "bbox": [ + 305, + 246, + 545, + 413 + ], + "type": "text", + "content": "Effect of global bundle adjustment. Tab. 6 shows the average ATE of our SLAM method on the 6 ScanNet scenes using different BA strategies: (w/o BA) pure tracking; (LBA) BA with rays from 10 local keyframes, a similar strategy to NICE-SLAM; (GBA-10) BA using rays from only 10 keyframes randomly selected from all past keyframes; (GBA) denotes the global BA strategy of Co-SLAM. We observe that using rays from a small (10) number of keyframes (LBA and GBA-10) leads to higher ATE errors. However, when keyframes are chosen from the entire sequence (GBA-10), instead of locally (LBA) the standard deviation is greatly reduced. Sampling rays from all keyframes (GBA) is the best overall strategy, even when all methods sample the same number of total rays (2048)." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 427, + 378, + 440 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 427, + 378, + 440 + ], + "spans": [ + { + "bbox": [ + 306, + 427, + 378, + 440 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 449, + 545, + 520 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 449, + 545, + 520 + ], + "spans": [ + { + "bbox": [ + 305, + 449, + 545, + 520 + ], + "type": "text", + "content": "We presented Co-SLAM, a dense real-time neural RGB-D SLAM system. We show that using a joint coordinate and parametric encoding with tiny MLPs as scene representation and training it with global bundle adjustment, achieves high-fidelity mapping and accurate tracking with plausible hole filling and efficient memory use." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 521, + 545, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 521, + 545, + 605 + ], + "spans": [ + { + "bbox": [ + 305, + 521, + 545, + 605 + ], + "type": "text", + "content": "Limitations. Co-SLAM relies on inputs from an RGB-D sensor and is therefore sensitive to illumination changes and inaccurate depth measurements. Instead of sampling keyframe pixels randomly, an information-guided pixel sampling strategy could be helpful to further reduce the number of pixels and improve the tracking speed. Incorporating loop closure could lead to further improvements." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 620, + 408, + 634 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 620, + 408, + 634 + ], + "spans": [ + { + "bbox": [ + 306, + 620, + 408, + 634 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 305, + 641, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 641, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 305, + 641, + 545, + 712 + ], + "type": "text", + "content": "The research presented here has been supported by a sponsored research award from Cisco Research and the UCL Centre for Doctoral Training in Foundational AI under UKRI grant number EP/S021566/1. This project made use of time on Tier 2 HPC facility JADE2, funded by EPSRC (EP/T022205/1)." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13300" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "type": "text", + "content": "[1] Dejan Azinović, Ricardo Martin-Brualla, Dan B. Goldman, Matthias Nießner, and Justus Thies. Neural RGB-D surface reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6290-6301, 2022. 1, 2, 3, 4, 5, 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 147, + 288, + 223 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 147, + 288, + 223 + ], + "spans": [ + { + "bbox": [ + 53, + 147, + 288, + 223 + ], + "type": "text", + "content": "[2] Eric R. Chan, Connor Z. Lin, Matthew A. Chan, Koki Nagano, Boxiao Pan, Shalini De Mello, Orazio Gallo, Leonidas J. Guibas, Jonathan Tremblay, and Sameh Khamis. Efficient geometry-aware 3D generative adversarial networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16123-16133, 2022. 2, 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 224, + 288, + 257 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 224, + 288, + 257 + ], + "spans": [ + { + "bbox": [ + 53, + 224, + 288, + 257 + ], + "type": "text", + "content": "[3] Jiawen Chen, Dennis Bautembach, and Shahram Izadi. Scalable real-time volumetric surface reconstruction. ACM Transactions on Graphics (ToG), 32(4):1-16, 2013. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 258, + 288, + 301 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 258, + 288, + 301 + ], + "spans": [ + { + "bbox": [ + 53, + 258, + 288, + 301 + ], + "type": "text", + "content": "[4] Brian Curless and Marc Levoy. A volumetric method for building complex models from range images. In Proceedings of the 23rd Annual Conference on Computer Graphics and Interactive Techniques, pages 303-312, 1996. 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 302, + 287, + 367 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 302, + 287, + 367 + ], + "spans": [ + { + "bbox": [ + 53, + 302, + 287, + 367 + ], + "type": "text", + "content": "[5] Angela Dai, Angel X. Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 5828–5839, 2017. 2, 5, 6, 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 369, + 287, + 423 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 369, + 287, + 423 + ], + "spans": [ + { + "bbox": [ + 53, + 369, + 287, + 423 + ], + "type": "text", + "content": "[6] Angela Dai, Matthias Nießner, Michael Zollhöfer, Shahram Izadi, and Christian Theobalt. Bundlefusion: Real-time globally consistent 3d reconstruction using on-the-fly surface reintegration. ACM Transactions on Graphics (ToG), 36(4):1, 2017. 2, 5, 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 424, + 287, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 424, + 287, + 479 + ], + "spans": [ + { + "bbox": [ + 53, + 424, + 287, + 479 + ], + "type": "text", + "content": "[7] Sara Fridovich-Keil, Alex Yu, Matthew Tancik, Qinhong Chen, Benjamin Recht, and Angjoo Kanazawa. Plenoxels: Radiance Fields Without Neural Networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5501-5510, 2022. 2, 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 480, + 287, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 480, + 287, + 544 + ], + "spans": [ + { + "bbox": [ + 53, + 480, + 287, + 544 + ], + "type": "text", + "content": "[8] Haoyu Guo, Sida Peng, Haotong Lin, Qianqian Wang, Guofeng Zhang, Hujun Bao, and Xiaowei Zhou. Neural 3D Scene Reconstruction with the Manhattan-world Assumption. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5511-5520, 2022. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 546, + 287, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 546, + 287, + 601 + ], + "spans": [ + { + "bbox": [ + 53, + 546, + 287, + 601 + ], + "type": "text", + "content": "[9] Olaf Kähler, Victor Adrian Prisacariu, Carl Yuheng Ren, Xin Sun, Philip H. S. Torr, and David William Murray. Very high frame rate volumetric integration of depth images on mobile devices. IEEE Trans. Vis. Comput. Graph., 21(11):1241-1250, 2015. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 602, + 287, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 602, + 287, + 645 + ], + "spans": [ + { + "bbox": [ + 48, + 602, + 287, + 645 + ], + "type": "text", + "content": "[10] Animesh Karnewar, Tobias Ritschel, Oliver Wang, and Niloy Mitra. Relu fields: The little non-linearity that could. In ACM SIGGRAPH 2022 Conference Proceedings, pages 1-9, 2022. 2, 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 647, + 287, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 647, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 647, + 287, + 689 + ], + "type": "text", + "content": "[11] L Koestler, N Yang, N Zeller, and D Cremers. Tandem: Tracking and dense mapping in real-time using deep multiview stereo. In Conference on Robot Learning (CoRL), 2021. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "type": "text", + "content": "[12] Hai Li, Xingrui Yang, Hongjia Zhai, Yuqian Liu, Hujun Bao, and Guofeng Zhang. Vox-Surf: Voxel-based Implicit Surface" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 713 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 94 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 94 + ], + "type": "text", + "content": "Representation. arXiv preprint arXiv:2208.10925, 2022. 2, 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 308, + 96, + 545, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 96, + 545, + 139 + ], + "spans": [ + { + "bbox": [ + 308, + 96, + 545, + 139 + ], + "type": "text", + "content": "[13] Lingjie Liu, Jiatao Gu, Kyaw Zaw Lin, Tat-Seng Chua, and Christian Theobalt. Neural sparse voxel fields. Advances in Neural Information Processing Systems, 33:15651-15663, 2020. 2, 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 141, + 545, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 141, + 545, + 196 + ], + "spans": [ + { + "bbox": [ + 307, + 141, + 545, + 196 + ], + "type": "text", + "content": "[14] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In European Conference on Computer Vision, pages 405-421. Springer, 2020. 1, 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 198, + 545, + 241 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 198, + 545, + 241 + ], + "spans": [ + { + "bbox": [ + 308, + 198, + 545, + 241 + ], + "type": "text", + "content": "[15] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM Transactions on Graphics, 41(4):102:1-102:15, July 2022. 2, 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 243, + 545, + 284 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 243, + 545, + 284 + ], + "spans": [ + { + "bbox": [ + 308, + 243, + 545, + 284 + ], + "type": "text", + "content": "[16] Thomas Müller, Brian McWilliams, Fabrice Rousselle, Markus Gross, and Jan Novák. Neural importance sampling. ACM Transactions on Graphics (TOG), 38(5):1-19, 2019. 2, 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 287, + 545, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 287, + 545, + 330 + ], + "spans": [ + { + "bbox": [ + 308, + 287, + 545, + 330 + ], + "type": "text", + "content": "[17] Raul Mur-Artal and Juan D. Tardós. Orb-slam2: An opensource slam system for monocular, stereo, and rgb-d cameras. IEEE transactions on robotics, 33(5):1255-1262, 2017. 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 332, + 545, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 332, + 545, + 398 + ], + "spans": [ + { + "bbox": [ + 308, + 332, + 545, + 398 + ], + "type": "text", + "content": "[18] Richard A Newcombe, Shahram Izadi, Otmar Hilliges, David Molyneaux, David Kim, Andrew J Davison, Pushmeet Kohli, Jamie Shotton, Steve Hodges, and Andrew Fitzgibbon. KinectFusion: Real-time dense surface mapping and tracking. In IEEE International Symposium on Mixed and Augmented Reality, pages 127-136, 2011. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 399, + 545, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 399, + 545, + 442 + ], + "spans": [ + { + "bbox": [ + 308, + 399, + 545, + 442 + ], + "type": "text", + "content": "[19] Matthias Nießner, Michael Zollhöfer, Shahram Izadi, and Marc Stamminger. Real-time 3D reconstruction at scale using voxel hashing. ACM Trans. Graph., 32(6):169:1-169:11, 2013. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 445, + 545, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 445, + 545, + 498 + ], + "spans": [ + { + "bbox": [ + 308, + 445, + 545, + 498 + ], + "type": "text", + "content": "[20] Nasim Rahaman, Aristide Baratin, Devansh Arpit, Felix Draxler, Min Lin, Fred Hamprecht, Yoshua Bengio, and Aaron Courville. On the spectral bias of neural networks. In International Conference on Machine Learning, pages 5301-5310. PMLR, 2019. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 500, + 545, + 532 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 500, + 545, + 532 + ], + "spans": [ + { + "bbox": [ + 308, + 500, + 545, + 532 + ], + "type": "text", + "content": "[21] Thomas Schops, Torsten Sattler, and Marc Pollefeys. BAD SLAM: Bundle adjusted direct RGB-D SLAM. In CVPR, 2019. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 534, + 547, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 534, + 547, + 578 + ], + "spans": [ + { + "bbox": [ + 308, + 534, + 547, + 578 + ], + "type": "text", + "content": "[22] Thomas Schops, Torsten Sattler, and Marc Pollefeys. Bad slam: Bundle adjusted direct rgb-d slam. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 134-144, 2019. 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 579, + 545, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 579, + 545, + 624 + ], + "spans": [ + { + "bbox": [ + 308, + 579, + 545, + 624 + ], + "type": "text", + "content": "[23] Vincent Sitzmann, Julien Martel, Alexander Bergman, David Lindell, and Gordon Wetzstein. Implicit neural representations with periodic activation functions. Advances in Neural Information Processing Systems, 33:7462-7473, 2020. 6" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 624, + 545, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 624, + 545, + 678 + ], + "spans": [ + { + "bbox": [ + 308, + 624, + 545, + 678 + ], + "type": "text", + "content": "[24] Julian Straub, Thomas Whelan, Lingni Ma, Yufan Chen, Erik Wijmans, Simon Green, Jakob J. Engel, Raul Mur-Artal, Carl Ren, and Shobhit Verma. The Replica dataset: A digital replica of indoor spaces. arXiv preprint arXiv:1906.05797, 2019. 2, 5, 6, 7" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 680, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 680, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 680, + 545, + 713 + ], + "type": "text", + "content": "[25] Jürgen Sturm, Nikolas Engelhard, Felix Endres, Wolfram Burgard, and Daniel Cremers. A benchmark for the evaluation of RGB-D SLAM systems. In 2012 IEEE/RSJ Interna" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13301" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 73, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 67, + 73, + 287, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 73, + 287, + 95 + ], + "spans": [ + { + "bbox": [ + 67, + 73, + 287, + 95 + ], + "type": "text", + "content": "tional Conference on Intelligent Robots and Systems, pages 573-580.IEEE,2012.2,5,6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 96, + 287, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 96, + 287, + 150 + ], + "spans": [ + { + "bbox": [ + 48, + 96, + 287, + 150 + ], + "type": "text", + "content": "[26] Edgar Sucar, Shikun Liu, Joseph Ortiz, and Andrew J. Davison. iMAP: Implicit mapping and positioning in real-time. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6229-6238, 2021. 2, 3, 4, 5, 6, 7, 8" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 152, + 288, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 152, + 288, + 206 + ], + "spans": [ + { + "bbox": [ + 48, + 152, + 288, + 206 + ], + "type": "text", + "content": "[27] Cheng Sun, Min Sun, and Hwann-Tzong Chen. Direct voxel grid optimization: Super-fast convergence for radiance fields reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5459-5469, 2022. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 208, + 288, + 283 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 208, + 288, + 283 + ], + "spans": [ + { + "bbox": [ + 48, + 208, + 288, + 283 + ], + "type": "text", + "content": "[28] Towaki Takikawa, Joey Litalien, Kangxue Yin, Karsten Kreis, Charles Loop, Derek Nowrouzezahrai, Alec Jacobson, Morgan McGuire, and Sanja Fidler. Neural geometric level of detail: Real-time rendering with implicit 3D shapes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11358-11367, 2021. 2, 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 285, + 287, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 285, + 287, + 319 + ], + "spans": [ + { + "bbox": [ + 48, + 285, + 287, + 319 + ], + "type": "text", + "content": "[29] K. Tateno, F. Tombari, I. Laina, and N. Navab. CNN-SLAM: Real-time dense monocular slam with learned depth prediction. In CVPR, 2017. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 320, + 287, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 320, + 287, + 354 + ], + "spans": [ + { + "bbox": [ + 48, + 320, + 287, + 354 + ], + "type": "text", + "content": "[30] Zachary Teed and Jia Deng. DROID-SLAM: Deep Visual SLAM for Monocular, Stereo, and RGB-D Cameras. Advances in neural information processing systems, 2021. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 354, + 287, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 354, + 287, + 398 + ], + "spans": [ + { + "bbox": [ + 48, + 354, + 287, + 398 + ], + "type": "text", + "content": "[31] Emanuele Vespa, Nikolay Nikolov, Marius Grimm, Luigi Nardi, Paul HJ Kelly, and Stefan Leutenegger. Efficient octree-based volumetric SLAM supporting signed-distance and occupancy mapping. RAL, 2018. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 399, + 287, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 399, + 287, + 442 + ], + "spans": [ + { + "bbox": [ + 48, + 399, + 287, + 442 + ], + "type": "text", + "content": "[32] Jingwen Wang, Tymoteusz Bleja, and Lourdes Agapito. Go-surf: Neural feature grid optimization for fast, high-fidelity rgb-d surface reconstruction. arXiv preprint arXiv:2206.14735, 2022. 2, 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 445, + 287, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 445, + 287, + 498 + ], + "spans": [ + { + "bbox": [ + 48, + 445, + 287, + 498 + ], + "type": "text", + "content": "[33] Peng Wang, Lingjie Liu, Yuan Liu, Christian Theobalt, Taku Komura, and Wenping Wang. Neus: Learning neural implicit surfaces by volume rendering for multi-view reconstruction. Advances in Neural Information Processing Systems, 2021. 3, 4" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 500, + 287, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 500, + 287, + 533 + ], + "spans": [ + { + "bbox": [ + 48, + 500, + 287, + 533 + ], + "type": "text", + "content": "[34] Thomas Whelan, Michael Kaess, Maurice Fallon, Hordur Johannsson, John Leonard, and John McDonald. Kintinous: Spatially extended kinectfusion. 2012. 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 535, + 287, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 535, + 287, + 578 + ], + "spans": [ + { + "bbox": [ + 48, + 535, + 287, + 578 + ], + "type": "text", + "content": "[35] Thomas Whelan, Stefan Leutenegger, Renato F. Salas-Moreno, Ben Glocker, and Andrew J. Davison. ElasticFusion: Dense SLAM without a pose graph. In Robotics: Science and Systems, volume 11, 2015. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 580, + 287, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 580, + 287, + 633 + ], + "spans": [ + { + "bbox": [ + 48, + 580, + 287, + 633 + ], + "type": "text", + "content": "[36] Thomas Whelan, Renato F Salas-Moreno, Ben Glocker, Andrew J Davison, and Stefan Leutenegger. ElasticFusion: Real-time dense SLAM and light source estimation. In International Journal of Robotics Research, volume 35, pages 1697-1716, 2016. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 635, + 287, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 635, + 287, + 678 + ], + "spans": [ + { + "bbox": [ + 48, + 635, + 287, + 678 + ], + "type": "text", + "content": "[37] Tong Wu, Jiaqi Wang, Xingang Pan, Xudong Xu, Christian Theobalt, Ziwei Liu, and Dahua Lin. Voxurf: Voxel-based Efficient and Accurate Neural Surface Reconstruction. arXiv preprint arXiv:2208.12697, 2022. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 681, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 681, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 681, + 287, + 713 + ], + "type": "text", + "content": "[38] Nan Yang, Lukas von Stumberg, Rui Wang, and Daniel Cremers. D3vo: Deep depth, deep pose and deep uncertainty for monocular visual odometry. In Proceedings of" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 295 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "type": "text", + "content": "the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1281-1292, 2020. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 96, + 545, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 96, + 545, + 139 + ], + "spans": [ + { + "bbox": [ + 307, + 96, + 545, + 139 + ], + "type": "text", + "content": "[39] Lior Yariv, Jiatao Gu, Yoni Kasten, and Yaron Lipman. Volume rendering of neural implicit surfaces. Advances in Neural Information Processing Systems, 34:4805-4815, 2021. 3, 4" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 141, + 545, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 141, + 545, + 194 + ], + "spans": [ + { + "bbox": [ + 307, + 141, + 545, + 194 + ], + "type": "text", + "content": "[40] Zehao Yu, Songyou Peng, Michael Niemeyer, Torsten Sattler, and Andreas Geiger. Monosdf: Exploring monocular geometric cues for neural implicit surface reconstruction. Advances in Neural Information Processing Systems, 2022. 2, 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 197, + 545, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 197, + 545, + 229 + ], + "spans": [ + { + "bbox": [ + 307, + 197, + 545, + 229 + ], + "type": "text", + "content": "[41] Ming Zeng, Fukai Zhao, Jiaxiang Zheng, and Xinguo Liu. Octree-based fusion for realtime 3d reconstruction. Graphical Models, 75(3):126-136, 2013. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 231, + 545, + 295 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 231, + 545, + 295 + ], + "spans": [ + { + "bbox": [ + 308, + 231, + 545, + 295 + ], + "type": "text", + "content": "[42] Zihan Zhu, Songyou Peng, Viktor Larsson, Weiwei Xu, Hujun Bao, Zhaopeng Cui, Martin R. Oswald, and Marc Pollefeys. Nice-slam: Neural implicit scalable encoding for slam. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12786-12796, 2022. 2, 3, 4, 5, 6, 7, 8" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "13302" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Co-Salient Object Detection With Uncertainty-Aware Group Exchange-Masking/9cd15a5b-7ef1-4ba9-a24c-48575755f531_content_list.json b/2023/Co-Salient Object Detection With Uncertainty-Aware Group Exchange-Masking/9cd15a5b-7ef1-4ba9-a24c-48575755f531_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..d499eddc2a7a325a8e94db052730020e17f2b806 --- /dev/null +++ b/2023/Co-Salient Object Detection With Uncertainty-Aware Group Exchange-Masking/9cd15a5b-7ef1-4ba9-a24c-48575755f531_content_list.json @@ -0,0 +1,1690 @@ +[ + { + "type": "text", + "text": "Co-Salient Object Detection with Uncertainty-aware Group Exchange-Masking", + "text_level": 1, + "bbox": [ + 84, + 130, + 883, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yang Wu $^{1}$ Huihui Song $^{1*}$ Bo Liu $^{2}$ Kaihua Zhang $^{1}$ Dong Liu $^{3}$", + "bbox": [ + 202, + 179, + 764, + 198 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ B-DAT and CICAET, Nanjing University of Information Science and Technology, Nanjing, China", + "bbox": [ + 91, + 198, + 879, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ Walmart Global Tech, Sunnyvale, CA, 94086, USA", + "bbox": [ + 277, + 215, + 691, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3Netflix Inc, Los Gatos, CA, 95032, USA", + "bbox": [ + 320, + 233, + 648, + 250 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "songhuihui@nuist.edu.cn", + "bbox": [ + 375, + 253, + 581, + 268 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 304, + 312, + 319 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The traditional definition of co-salient object detection (CoSOD) task is to segment the common salient objects in a group of relevant images. Existing CoSOD models by default adopt the group consensus assumption. This brings about model robustness defect under the condition of irrelevant images in the testing image group, which hinders the use of CoSOD models in real-world applications. To address this issue, this paper presents a group exchange-masking (GEM) strategy for robust CoSOD model learning. With two groups of image containing different types of salient object as input, the GEM first selects a set of images from each group by the proposed learning based strategy, then these images are exchanged. The proposed feature extraction module considers both the uncertainty caused by the irrelevant images and group consensus in the remaining relevant images. We design a latent variable generator branch which is made of conditional variational autoencoder to generate uncertainly-based global stochastic features. A CoSOD transformer branch is devised to capture the correlation-based local features that contain the group consistency information. At last, the output of two branches are concatenated and fed into a transformer-based decoder, producing robust co-saliency prediction. Extensive evaluations on co-saliency detection with and without irrelevant images demonstrate the superiority of our method over a variety of state-of-the-art methods.", + "bbox": [ + 75, + 335, + 473, + 729 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 756, + 207, + 771 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Co-salient object detection (CoSOD) is to segment the common salient objects in a group of relevant images. By detecting the co-salient object in a group of images, the images' background and redundant content are re", + "bbox": [ + 75, + 781, + 468, + 842 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/94854fdc8c47034ac3e62cc651aca8f0721e02daba5b07d7d0b6cc6c9a217f61.jpg", + "image_caption": [ + "Figure 1. (a) When there exists an irrelevant image in the test group, the state-of-the-art CoSOD models such as DCFM [39] and GCoNet [9] tend to generate false positive predictions for the noisy image, yet our method can achieve an accurate prediction due to the use of group exchange-masking strategy in (b)." + ], + "image_footnote": [], + "bbox": [ + 504, + 304, + 651, + 448 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/a70bec057487f1f85338cccfaa7d71bf4ebaa5ea13ba93fe6c3e2531cd1797c5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 653, + 304, + 885, + 448 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "moved, which helps the downstream tasks such as object tracking [38], co-segmentation [48] and video colocalization [14], to name a few.", + "bbox": [ + 496, + 564, + 890, + 609 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Group consensus assumption is widely used by existing CoSOD models, i.e., these models presume that all images in the same group contain the common salient targets. The widely-used CoSOD benchmark datasets, such as COCO-SEG [33], CoCA [50], and CoSOD3k [8] organize the training and testing images that contain the same object as a group. Many existing CoSOD models consider the group consensus characteristic in modeling. For example, early works such as [2, 10, 13] extract hand-crafted features for inter-image co-object correspondence discovery. The deep learning models proposed in [15, 20, 36, 39, 44, 46, 48] use one group of relevant images as training data input for consensus representation learning. Among them, a variety of novel model design techniques have been developed to make full use of the group consistency characteristic, such as the low-rank feature learning [46], co-attention model [20] and intra-saliency correlation learning [15]. The issues of the group consensus assumption are partially studied in recent literature [9], which further models the inter", + "bbox": [ + 496, + 613, + 892, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author. This work is supported in part by National Key Research and Development Program of China under Grant No. 2018AAA0100400, in part by the NSFC under Grant Nos. 62276141, 61872189.", + "bbox": [ + 75, + 849, + 468, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "19639", + "bbox": [ + 480, + 924, + 517, + 936 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "group separability for more discriminative feature learning by a group collaborating module.", + "bbox": [ + 76, + 90, + 468, + 119 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we find that the group consensus assumption also restricts the CoSOD model's robustness against the images without common object. As illustrated by Figure 1(a), state-of-the-art CoSOD models tend to output false positive predictions for the irrelevant image. This issue hinders the use of CoSOD models in real-world applications where the testing inputs are likely to contain irrelevant images. To enhance the model's robustness, we propose a learning framework called group exchange-masking (GEM). The GEM is illustrated by Figure 1(b). Given two image groups that contain different types of co-salient objects, we exchange several images between one group and the other. Those exchanged images are called noisy images. The number of noisy images is chosen to be less than the number of remaining relevant images in the group, so that the co-salient object in the noisy images forms a negative object but not the dominant co-salient object. The \"masking\" strategy refers to the label regeneration of the noisy images. Because there is no co-salient object, in the regenerated label, the original ground-truth object is masked. The learning objective is to correctly predict both the co-salient objects in the original relevant images and the added noisy images.", + "bbox": [ + 75, + 122, + 470, + 467 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Adding noisy images to the training image group brings about uncertainty to the CoSOD model learning since there is some probability of no expected common object in each image. We design a dual-path image feature extraction module to model the group uncertainty in addition to the group consensus property. Specifically, we design a latent variable generator branch (LVGB) to extract the uncertainty-based global image features. The LVGB module is motivated by the conditional variational autoencoder (CVAE) [32] that is widely used to address the uncertainty in vision applications including image background modeling [19], RGB-D saliency detection [43] and image reconstruction [41]. In parallel with LVGB, we feed the image group into a CoSOD Transformer Branch (CoSOD-TB). The CoSOD-TB partitions each image group into local patches, and the attention mechanism in the transformer enables this branch to model patch-wise correlation-based local features. As a result, the group consistency information can be captured by this branch. The outputs of the two branches are concatenated and fed into a transformer-based decoder for co-saliency prediction. The proposed model has the following technical contributions.", + "bbox": [ + 75, + 468, + 470, + 801 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "- A robust CoSOD model learning mechanism, called group exchange-masking is proposed. By exchanging images between two groups, we augment the training data containing irrelevant images as noise to enhance model's robustness. This is different from the traditional CoSOD model learning frameworks that use", + "bbox": [ + 94, + 809, + 470, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "groups of relevant images as training data.", + "bbox": [ + 532, + 90, + 812, + 107 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose a dual-path feature extraction module composed of the LVGB and the CoSOD-TB. The LVGB is designed to model the uncertainty of cosalient object existence. The CoSOD-TB is for the consensus feature extraction of the salient object in the relevant images.", + "- Extensive evaluations on three benchmark datasets, including CoSal2015 [42], CoCA [15], and CoSOD3k [7] show that the superiority of the proposed model to the state-of-the-art methods in terms of all evaluation metrics. Besides, the proposed model demonstrates good robustness for dealing with noisy data without co-salient objects." + ], + "bbox": [ + 517, + 116, + 893, + 325 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 349, + 640, + 366 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Co-salient Object Detection", + "text_level": 1, + "bbox": [ + 500, + 375, + 748, + 391 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In the past, CoSOD methods used to extract handcrafted features such as Gabor and SIFT features from images and then detect co-saliency by utilizing the consistency of low-level features between the images being tested [2]. A series of studies attempt to capture the intra-image constraints by employing a manifold ranking scheme to produce saliency maps [21], or using a global association constraint with clustering [10], or translational alignment [13]. More recently, there is a surge of deep learning-based CoSOD models that learn feature representation and saliency predictor in an end-to-end manner [9, 12, 36, 46, 47, 50]. Wei et al. [36] designs a collaborative learning framework for CoSOD that discovers the collaborative and interactive relationships between intra-image and single-image feature representations in a group-wise manner. Hsu et al. [12] present an unsupervised CNN-based model for CoSOD. In [46], a hierarchical framework is proposed for CoSOD where the initial CoSOD results generated by the CNN model is refined by label smoothing. Zhang et al. [50] propose a gradient-induced model for CoSOD that utilizes the image gradient information to induce more attentions to the discriminative co-salient feature learning. In [47], a deep graph neural network model is proposed to characterize the intra-image and inter-image region correspondence for CoSOD. In [9], a group collaborative learning strategy is proposed to explore inter-group relations for discriminating feature learning.", + "bbox": [ + 496, + 398, + 890, + 792 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Robust Model and Feature Learning", + "text_level": 1, + "bbox": [ + 500, + 801, + 816, + 816 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In past years, with the development of deep learning, robust model and feature learning have received more and more attention as a possible solution to overcoming the bottleneck [28]. Various methods have been proposed to improve the robustness of the model. In [3] and [30], ways to", + "bbox": [ + 496, + 825, + 893, + 902 + ], + "page_idx": 1 + }, + { + "type": "footer", + "text": "19640", + "bbox": [ + 480, + 924, + 519, + 936 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/f3256742c23bdfa633fb2e92436604e49167e6b70d2961de41df887a448d6a2d.jpg", + "image_caption": [ + "Figure 2. Pipeline of the proposed method. First, two groups of images $\\{\\mathcal{I}_1,\\mathcal{I}_2\\}$ with their masks $\\{\\mathcal{V}_1,\\mathcal{V}_2\\}$ are processed by group exchange-masking, yielding the exchanged images $\\{\\mathcal{I}_{e1},\\mathcal{I}_{e2}\\}$ and their corresponding masked labels $\\{\\mathcal{V}_{e1},\\mathcal{V}_{e2}\\}$ . Then, $\\{\\mathcal{I}_{e1},\\mathcal{I}_{e2}\\}$ and $\\{\\mathcal{V}_{e1},\\mathcal{V}_{e2}\\}$ are passed through the latent variable generator branch, extracting the uncertainty-based global image features $\\{\\mathcal{V}_1,\\mathcal{V}_2\\}$ to eliminate the bias caused by noisy images. Meanwhile, $\\{\\mathcal{I}_{e1},\\mathcal{I}_{e2}\\}$ are passed through the weight-shared CoSOD transformer branch, producing two groups of feature sequences $\\{\\mathcal{F}_1,\\mathcal{F}_2\\}$ with group consistency information and long-range independent information. Finally, $\\mathcal{F}$ and $\\mathcal{V}$ are concatenated and fed into the transformer decoder, yielding the predicted co-saliency maps $\\hat{\\mathcal{O}}$ ." + ], + "image_footnote": [], + "bbox": [ + 81, + 88, + 893, + 348 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "transfer the policy learned in a simulator to the real world are proposed to make features more robust. Rajeswaran et al. [29] learn a robust policy by sampling the worst case trajectories from a class of parametrized models, to learn a robust policy. Xie et al. [37] enhance the robustness of the model through randomly resizing and padding the training images. Several works leverage randomness to improve the robustness of models. Liu et al. [24] propose a noise layer that introduces randomness to both the input and the hidden layer output. Liu et al. [25] introduce a new min-max formulation that combines adversarial training with Bayesian Neural Networks, which achieves promising results.", + "bbox": [ + 75, + 468, + 472, + 650 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Proposed Method", + "text_level": 1, + "bbox": [ + 76, + 662, + 250, + 680 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Figure 2 illustrates the pipeline of our framework. Given two groups of $2N$ images $\\mathcal{I}_i = \\{I_i^n\\in \\mathbb{R}^{H\\times W\\times 3}\\}_{n = 1}^N,i = 1,2$ with height $H$ and width $W$ and manually-labeled binary masks $\\mathcal{V}_i = \\{\\mathbf{Y}_i^n\\in \\mathbb{R}^{H\\times W}\\}_{n = 1}^N$ with different categories as input, we employ the GEM model to reorganize them, yielding $\\mathcal{I}_{ei}$ and $\\mathcal{V}_{ei}$ . Each pair of $\\{\\mathcal{I}_{ei},\\mathcal{V}_{ei}\\}$ are fed to the LVGB. In LVGB, we draw $\\mathbf{z}_i\\in \\mathbb{R}^K$ from the Gaussian distribution $P_{\\theta}(\\mathbf{z}_i|\\mathcal{I}_{ei})$ with the learning parameters $\\pmb{\\theta}$ and the posterior of $\\mathbf{z}_i$ is formulated as $Q_{\\phi}(\\mathbf{z}_i|\\mathcal{I}_{ei},\\mathcal{V}_{ei})$ with the learning parameters $\\phi$ . We use the KL-Divergence metric to reduce the distance between $P_{\\theta}(\\mathbf{z}_i|\\mathcal{I}_{ei})$ and $Q_{\\phi}(\\mathbf{z}_i|\\mathcal{I}_{ei},\\mathcal{V}_{ei})$ . We further re-parameterize and expand $\\mathbf{z}_i$ to yield a latent sequence $\\mathcal{V}_i$ which has the same dimension as the feature sequences produced by trans", + "bbox": [ + 75, + 688, + 472, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Algorithm 1 Group exchange-masking", + "text_level": 1, + "bbox": [ + 501, + 467, + 761, + 483 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Input: $\\mathcal{I}_1 = \\{I_1^n\\}_{n=1}^N$ , $\\mathcal{Y}_1 = \\{\\mathbf{Y}_1^n\\}_{n=1}^N$ ,", + "bbox": [ + 506, + 484, + 771, + 502 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {I} _ {2} = \\left\\{I _ {2} ^ {n} \\right\\} _ {n = 1} ^ {N}, \\mathcal {Y} _ {2} = \\left\\{\\mathbf {Y} _ {2} ^ {n} \\right\\} _ {n = 1} ^ {N}.\n$$\n", + "text_format": "latex", + "bbox": [ + 557, + 502, + 771, + 517 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Output: $\\mathcal{I}_{e1},\\mathcal{Y}_{e1},\\mathcal{I}_{e2},\\mathcal{Y}_{e2}$", + "bbox": [ + 506, + 517, + 694, + 532 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "1: $\\mathcal{I}_{e1},\\mathcal{Y}_{e1},\\mathcal{I}_{e2},\\mathcal{Y}_{e2}\\gets$ Rearranging $\\mathcal{I}_1,\\mathcal{Y}_1,\\mathcal{I}_2,\\mathcal{Y}_2$ via solving (1).", + "bbox": [ + 508, + 534, + 890, + 564 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "2: for $n = 1, 2, \\ldots, k$ do", + "3: $I_{e1}^{n} \\gets I_{2}^{n}, I_{e2}^{n} \\gets I_{1}^{n}$ ,", + "4: $\\mathbf{Y}_{e1}^{n}\\gets \\mathbf{0}\\in \\mathbb{R}^{H\\times W\\times 1},$", + "5: $\\mathbf{Y}_{e2}^{\\vec{n}}\\gets \\mathbf{0}\\in \\mathbb{R}^{H\\times W\\times 1}$", + "6: end for" + ], + "bbox": [ + 508, + 565, + 694, + 637 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "formers. Meanwhile, the image groups are passed through the CoSOD-TB with shared weights, producing the token sequences $\\mathcal{F}_i$ which capture group consensus and long-range dependency information. At last, $\\mathcal{V}_i$ and $\\mathcal{F}_i$ are concatenated in channel dimension and passed through several subsequent processes as shown in Figure 2 followed by an up-sampling layer, yielding the corresponding predicted cosaliency maps $\\hat{\\mathcal{O}} = \\{\\hat{\\mathbf{O}}^n\\in \\mathbb{R}^{H\\times W}\\}_{n = 1}^{2N}$ .", + "bbox": [ + 496, + 667, + 890, + 792 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Group Exchange-Masking", + "text_level": 1, + "bbox": [ + 500, + 801, + 740, + 816 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The widely-used group consensus assumption limits the robustness of the CoSOD model, especially when there are test images without common objects in the group. To this end, we reorganize the two input groups with the GEM tailored to CoSOD. The GEM is summarised in Algorithm 1:", + "bbox": [ + 496, + 824, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "19641", + "bbox": [ + 478, + 926, + 517, + 936 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/7c477db33a779d751b18b96233bdfb2f1ffbecf1bd6510b8e32065d3b31e6119.jpg", + "image_caption": [ + "Figure 3. Architecture of the LVGB." + ], + "image_footnote": [], + "bbox": [ + 78, + 90, + 472, + 224 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "taking $\\mathcal{I}_i$ and $\\mathcal{Y}_i$ as input, the goal of GEM is to obtain $\\mathcal{I}_{ei}$ and $\\mathcal{Y}_{ei}$ which contain noisy images in each group and corresponding full masked ground truths. Specifically, we rearrange the two groups of images and their binary ground-truth masks in a decreasing order by solving (1). After this, we sample the top- $k$ images of each group as noisy images and exchange them in each other's group, ensuring the maximum destructiveness of noisy images and the effectiveness of training.", + "bbox": [ + 75, + 277, + 468, + 412 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "This training process is essentially a min-max optimization process [26] as", + "bbox": [ + 76, + 412, + 468, + 444 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {\\varphi} \\sum_ {n = 1} ^ {k} \\sum_ {i = 1} ^ {2} \\max _ {\\mathcal {I} _ {i} (n,:)} \\mathcal {L} \\left(f _ {\\varphi} \\left(\\mathcal {I} _ {i} (n,:)\\right), \\hat {\\mathcal {O}} _ {i} (n,:)\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 122, + 454, + 468, + 496 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where the loss $\\mathcal{L}$ is defined by (7), and $f$ denotes our whole model with learnable parameters $\\varphi$ . Solving (1) finds out the most noisy images which can maximize the training loss, and our goal is to minimize the loss function w.r.t the intra-group noisy images to improve the model's robustness. At last, we mask the corresponding labels of the noisy images with the all-zeros maps.", + "bbox": [ + 75, + 505, + 468, + 612 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Latent Variable Generator", + "text_level": 1, + "bbox": [ + 76, + 619, + 318, + 636 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "By training the model with noisy images, the GEM is able to increase the model's robustness which can identify whether there exist co-objects or not. However, the model will focus on non-co-salient regions as a result of the uncertainty introduced by adding noisy images to the training image groups. This means that the model will become overconfident in the background regions in the early stage of training, leading to inaccurate prediction.", + "bbox": [ + 75, + 643, + 468, + 763 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To address this issue, we further propose the LVGB that generates a low-dimensional latent space to represent the most common patterns in image groups [17]. The latent variables sampled from the feature space characterize uncertainties of group consensus and can be used to modulate the intermediate features from other branches to highlight the co-objects.", + "bbox": [ + 75, + 763, + 468, + 869 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Figure 3 shows the details of our LVGB. The generator takes $\\mathcal{I}_{ei}$ and latent variable $\\mathbf{z}_i$ as input, and uses an encoder", + "bbox": [ + 76, + 869, + 468, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "to produce stochastic prediction $P_{\\omega}(\\mathcal{Y}_{ei}|\\mathcal{I}_{ei},\\mathbf{z}_i)$ with learnable parameters $\\omega$ , where $\\mathcal{I}_{ei}$ is conditional variable, and $\\mathcal{Y}_{ei}$ is output variable. $\\mathbf{z}_i$ is drawn from the Gaussian distribution $P_{\\theta}(\\mathbf{z}_i|\\mathcal{I}_{ei})$ , and the posterior of $\\mathbf{z}_i$ is formulated as $Q_{\\phi}(\\mathbf{z}_i|\\mathcal{I}_{ei},\\mathcal{Y}_{ei})$ via an encoder that maps the input variable to the latent space. The loss of the LVG is defined as [32,43]", + "bbox": [ + 498, + 90, + 890, + 183 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} _ {L V G} = \\sum_ {i = 1} ^ {2} E _ {\\mathbf {z} \\sim Q _ {\\phi} (\\mathbf {z} _ {i} | \\mathcal {I} _ {e i}, \\mathcal {Y} _ {e i})} [ - \\log P _ {\\boldsymbol {\\omega}} (\\mathcal {Y} _ {e i} | \\mathcal {I} _ {e i}, \\mathbf {z} _ {i}) ] \\\\ + \\mathrm {K L} \\left(Q _ {\\phi} \\left(\\mathbf {z} _ {i} \\mid \\mathcal {I} _ {e i}, \\mathcal {Y} _ {e i}\\right) \\mid \\mid P _ {\\boldsymbol {\\theta}} \\left(\\mathbf {z} _ {i} \\mid \\mathcal {I} _ {e i}\\right)\\right), \\tag {2} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 514, + 191, + 890, + 266 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where the Kullback-Leibler Divergence metric KL is defined as", + "bbox": [ + 498, + 267, + 890, + 295 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {K L} (Q | | P) = \\sum Q (x) \\log \\frac {Q (x)}{P (x)}. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 580, + 304, + 890, + 338 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In (2), the prior net $P_{\\pmb{\\theta}}(\\mathbf{z}_i|\\mathcal{I}_{ei})$ is defined as a Gaussian distribution that maps the image group $\\mathcal{I}_{ei}$ to a low-dimensional latent space that encodes the most common patterns in each group. We randomly sample prior latent variable $\\mathbf{z}_i\\sim \\mathcal{N}(\\pmb{\\mu}_{1i},\\mathrm{diag}(\\pmb{\\sigma}_{1i}^2))$ , where $\\pmb{\\mu}_{1i},\\pmb{\\sigma}_{1i}^{2}\\in \\mathbb{R}^{K}$ represent the mean and standard deviation vectors and $\\pmb{\\theta}$ is the learnable parameter of the mapping function which consists five conv layers [43] as shown in Figure 3. In the posterior net, we use the similar encoder as the prior net to model $Q_{\\phi}(\\mathbf{z}_i|\\mathcal{I}_{ei},\\mathcal{Y}_{ei})$ that maps the concatenation of $\\mathcal{I}_{ei}$ and $\\mathcal{Y}_{ei}$ to the posterior latent variable $\\mathbf{z}_i\\sim \\mathcal{N}(\\pmb{\\mu}_{2i},\\mathrm{diag}(\\pmb{\\sigma}_{2i}^2))$ where $\\pmb{\\mu}_{2i},\\pmb{\\sigma}_{2i}^{2}\\in \\mathbb{R}^{K}$ . The KL in (2) is used to measure the difference between probability distributions prior $P_{\\pmb{\\theta}}(\\mathbf{z}_i|\\mathcal{I}_{ei})$ and posterior $Q_{\\phi}(\\mathbf{z}_i|\\mathcal{I}_{ei},\\mathcal{Y}_{ei})$ . It represents the information loss that occurs when using $Q_{\\phi}(\\mathbf{z}_i|\\mathcal{I}_{ei},\\mathcal{Y}_{ei})$ to approximate $P_{\\pmb{\\theta}}(\\mathbf{z}_i|\\mathcal{I}_{ei})$ , and a smaller value indicates greater similarity between the two distributions.", + "bbox": [ + 498, + 347, + 890, + 603 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Each position in latent space symbolizes potential labeling changes or other potential factors that could lead to a variety of co-saliency predictions [17, 43]. The proposed GEM can meet the needs for diverse ground truths as we learn to select the most difficult samples to mask their ground truths to all-zeros maps. We hope diverse annotations in the posterior net $Q_{\\phi}(\\mathbf{z}_i|\\mathcal{I}_{ei},\\mathcal{Y}_{ei})$ during training can compel the prior net $P_{\\theta}(\\mathbf{z}_i|\\mathcal{I}_{ei})$ to encode labeling variants of the supplied inputs $\\mathcal{I}_{ei}$ . The statistics $\\mathbf{z}_i$ is further processed by feature expanding [43]. Given a pair of $(\\boldsymbol{\\mu}_k,\\boldsymbol{\\sigma}_k)$ in each position of $K$ dimensional vector, we parameterize them to obtain the latent vector", + "bbox": [ + 496, + 604, + 890, + 785 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {z} _ {k} = \\boldsymbol {\\sigma} _ {k} \\odot \\boldsymbol {\\epsilon} + \\boldsymbol {\\mu} _ {k}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 625, + 797, + 890, + 814 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\epsilon$ follows the standard normal distribution $\\epsilon \\sim \\mathcal{N}(0,1)$ . To fuse with the features $\\mathcal{F}_i$ from the CoSOD-TB, as shown in Figure 3, we expand $\\mathbf{z}_k$ to the feature sequence which has the same size as $\\mathcal{F}_i$ , yielding the stochastic feature $\\mathcal{V}_i$ . The visualization effect of $\\mathcal{V}_i$ is partly shown in Fig-", + "bbox": [ + 498, + 825, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "19642", + "bbox": [ + 480, + 924, + 517, + 936 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/669de5bedfbe1b9bbc55320728ec41581aa49c7de768d2a3bb2dd9bb401c1643.jpg", + "image_caption": [ + "Figure 4. Architecture of the CoSOD-TB." + ], + "image_footnote": [], + "bbox": [ + 78, + 88, + 472, + 198 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "ure 5, as we can see, these features encode rich group consistency information that can well highlight the co-salient regions.", + "bbox": [ + 75, + 251, + 468, + 296 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. CoSOD Transformer", + "text_level": 1, + "bbox": [ + 76, + 308, + 274, + 324 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "After getting the stochastic features $\\nu_{i}$ from the LVGB, we leverage these features to guide the learning of generating the general features $\\mathcal{F}_i$ from the CoSOD-TB to focus on the co-salient regions (see the bottom row of Figure 5).", + "bbox": [ + 75, + 332, + 468, + 393 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Figure 4 shows the architecture of the proposed CoSOD-TB. Each input image $I_{i}^{n} \\in \\mathcal{I}_{ei}$ is cropped into $d$ patches with size $H / 4 \\times W / 4$ , and then all the image patches in each group construct the token sequences $\\mathcal{X}_i^0 \\in \\mathbb{R}^{N \\times \\frac{H}{4} \\times \\frac{W}{4} \\times 3d}$ . Then, $\\mathcal{X}_i^0$ is sent into the backbone of the transformer with a T2T architecture [40], obtaining the token sequences $\\mathcal{X}_i^1 \\in \\mathbb{R}^{N \\times \\frac{H}{16} \\times \\frac{W}{16} \\times c}$ which encode both local and global information. After this, we design the group token $\\mathcal{X}^G \\in \\mathbb{R}^{N \\times 1 \\times c}$ to capture group-wise common information that is essential to extract co-objects, and the saliency token $\\mathcal{X}^S \\in \\mathbb{R}^{N \\times 1 \\times c}$ to capture the specific information that encodes more co-object structure details. Afterwards, $\\mathcal{X}^G$ and $\\mathcal{X}^1$ are concatenated in the channel dimension and fed into the Transformer Layer as the design in ViT [4] to get the new token sequence. We further use an multi-layer perceptron (MLP) layer to integrate global information into $\\mathcal{X}^G$ . We fuse the $\\mathcal{X}^S$ to integrate saliency information into sequences and split the $\\mathcal{X}^G$ and $\\mathcal{X}^S$ to up-sample the token sequence and get the token sequence $\\mathcal{X}^2$ . After undergoing some processing as shown in Figure 4, $\\mathcal{F}_i$ is then concatenated with the stochastic features $\\mathcal{V}_i$ followed by a decoder that includes two Transformer Layers and an MLP layer. During decoding, due to the use of stochastic features, the model is more possible to overcome the prejudices formed in the early training. As shown in the third row of Figure 5, the CoSOD-TB can integrate $\\mathcal{V}_i$ and $\\mathcal{F}_i$ well, helping the model focus on more reliable co-salient object regions. Finally, the output sequences are reshaped to produce the predicted co-saliency maps $\\hat{\\mathcal{O}}$ .", + "bbox": [ + 75, + 393, + 468, + 834 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The loss function of the CoSOD-TB is defined as", + "bbox": [ + 96, + 835, + 423, + 849 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {T R A N S} = \\frac {1}{2 N} \\sum_ {n = 1} ^ {N} \\sum_ {i = 1} ^ {2} \\ell_ {B C E} \\left(\\mathcal {Y} _ {e i} (n,:), \\hat {\\mathcal {O}} _ {i} (n,:)\\right), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 89, + 862, + 468, + 904 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/d6c019f4ac60ca461c8d3cfa98d78d068e0b4a53949b07943107ed80b1d12da8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 504, + 90, + 591, + 137 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/d4d641740597727ff39488d3d9ad453f69f56ef2e2955afa833acc4bfe5bdbb6.jpg", + "image_caption": [ + "V" + ], + "image_footnote": [], + "bbox": [ + 506, + 138, + 591, + 195 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/cb8608e47752fd79e1e9882da09ef7f81477508803ef2786099b76293ac9c958.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 596, + 90, + 684, + 137 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/124f6d34d2e8ab093b7373f739fab125f59dca7bc4c5a77ff4cbc0264c1b2ee1.jpg", + "image_caption": [ + "Watermelon_2" + ], + "image_footnote": [], + "bbox": [ + 689, + 90, + 777, + 137 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/b7147fc66c9abfc49b9649e6c16b391411ce5b53c35cbdbe07fab06424ba15c6.jpg", + "image_caption": [ + "F" + ], + "image_footnote": [], + "bbox": [ + 782, + 90, + 870, + 137 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/b1e0fa65a5d649d1604930a3813f64fe7d0dd798010061490e716100294cb3d3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 504, + 150, + 591, + 195 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/d1b1a274ba14bd665d087b38eccbfb94a9239c7786bf1eb7e9145ac4567a7e8c.jpg", + "image_caption": [ + "F", + "V" + ], + "image_footnote": [], + "bbox": [ + 596, + 150, + 684, + 195 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/8beadddbfde6357610eea1203d2dff9b9710a533484077f218e47a01485f18b6.jpg", + "image_caption": [ + "V2-1" + ], + "image_footnote": [], + "bbox": [ + 689, + 150, + 777, + 195 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/ab7f5f382d81f5218b6d9e49207827580a92de728e545fa866f39e53f75f7b83.jpg", + "image_caption": [ + "V2-2" + ], + "image_footnote": [], + "bbox": [ + 782, + 150, + 870, + 195 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/a2d31cdfb33e17e2f8ea6334113ea29fe6aed812f8759f471aef85f46460a783.jpg", + "image_caption": [ + "$\\mathcal{F}_i^1$ & $\\mathcal{V}_i^{1 - 1}$" + ], + "image_footnote": [], + "bbox": [ + 504, + 208, + 591, + 252 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/1a7138bfd942a6f1341a68be263627baf31060264d3cdb9b4659f72ad21e651c.jpg", + "image_caption": [ + "$\\mathcal{F}_i^1$ & $\\nu_{i}^{1 - 2}$" + ], + "image_footnote": [], + "bbox": [ + 596, + 208, + 684, + 252 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/e21aa753c738f59337856cd23dbc34a6595e6bc1e805dd0695f85f19bf622bae.jpg", + "image_caption": [ + "$\\mathcal{F}_i^2\\& \\mathcal{V}_i^{2 - 1}$" + ], + "image_footnote": [], + "bbox": [ + 689, + 208, + 777, + 252 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/e4c4e261fb78a898daf990052e6f879f3ea7537a35314cf71b96742f2f7bbaf7.jpg", + "image_caption": [ + "$\\mathcal{F}_i^2$ & $\\mathcal{V}_i^{2 - 2}$" + ], + "image_footnote": [], + "bbox": [ + 782, + 208, + 870, + 252 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/66c2176887f164ea6c566eab77593af273de301961cdd4dde98720b6dce642b3.jpg", + "image_caption": [ + "Figure 5. Comparison of effects of the \"watermelon\" group before and after fusing stochastic features. The first row: visualizations of the images and features obtained by the CoSOD-TB; The second row: visualizations of the features generated by the LVGB; The third row: visualizations of the integrated features." + ], + "image_footnote": [], + "bbox": [ + 879, + 90, + 908, + 255 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\ell_{BCE}$ is a binary cross-entropy (BCE) loss [27] defined as", + "bbox": [ + 496, + 375, + 890, + 402 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\ell_ {B C E} (\\mathcal {Y} (n,:), \\hat {\\mathcal {O}} (n,:)) = - (\\mathcal {Y} (n,:) ^ {\\top} \\log (\\hat {\\mathcal {O}} (n,:)) \\tag {6} \\\\ - \\left(1 - \\mathcal {Y} (n,:)\\right) ^ {\\top} \\log \\left(1 - \\tilde {\\mathcal {O}} (n,:)\\right)). \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 514, + 412, + 888, + 452 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Loss Function", + "text_level": 1, + "bbox": [ + 500, + 477, + 643, + 491 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The LVG and the CoSOD-TB are jointly trained in an end-to-end manner by optimizing the following multi-task loss", + "bbox": [ + 496, + 500, + 890, + 542 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\lambda_ {1} \\mathcal {L} _ {L V G} + \\lambda_ {2} \\mathcal {L} _ {T R A N S}, \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 596, + 546, + 890, + 561 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\lambda_{1},\\lambda_{2}$ are the hyperparameters to balance each loss.", + "bbox": [ + 498, + 566, + 885, + 583 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experimental Results", + "text_level": 1, + "bbox": [ + 498, + 595, + 704, + 612 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Implementation Details", + "text_level": 1, + "bbox": [ + 498, + 619, + 715, + 636 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our model is implemented under the PyTorch1.9.0 framework [27]. Our computing platform's acceleration is provided by a GeForce GTX 2080Ti GPU. The transformer backbone is the pre-trained T2T-ViT $t - 14$ [40] model since it has a similar computational complexity as CNNs-based ResNet50 [11] which is smaller than the VGG-16 [31] widely used in CoSOD. The training dataset are COCOSEG released by [33] and DUTS released by [34]. They totally include about 208,250 images from 369 categories, and the relevant binary ground-truth masks are provided.", + "bbox": [ + 496, + 643, + 890, + 794 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In each training episode, we randomly select two groups, which are processed by the proposed GEM with $k = 1$ . Each group contains $N = 5$ images. The hyperparameters in (7) are set to $\\lambda_1 = 0.25$ and $\\lambda_2 = 1$ . We resize the images to $224 \\times 224 \\times 3$ and use them as input. We train the network over 60,000 steps totally and the training process takes about 11 hours. During the training process, we use", + "bbox": [ + 496, + 795, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "19643", + "bbox": [ + 480, + 926, + 517, + 936 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/5adf55c23f5906c200b6b7e773e66a611ef4a37658406c17f8f6e2c3bb2cad26.jpg", + "image_caption": [ + "Figure 6. The qualitative comparisons with the recent state-of-the-art methods, including DCFM [39], CADC [49], DeepACG [45] and GCoNet [9]." + ], + "image_footnote": [], + "bbox": [ + 119, + 87, + 849, + 303 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "the Adam algorithm [16] to optimize the whole network. The initial learning rate is set to $10e - 4$ and decay to $10e - 5$ at the 20,000-th epoch. The inference time is 31 fps in average, meeting the real-time application requirements.", + "bbox": [ + 75, + 357, + 468, + 417 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Datasets and Evaluation Metrics", + "text_level": 1, + "bbox": [ + 76, + 440, + 362, + 455 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The test sets include CoSal2015 [42], CoCA [50] and CoSOD3k [8]. Among them, CoSal2015 contains 2,015 images from 50 categories, and the objects therein have different sizes and shapes, which makes CoSal2015 challenging; CoCA is the newest dataset, including 1,297 images from 80 categories. These images have extremely different styles and very complicated backgrounds, making it a very difficult data set; CoSOD3k contains 3,316 images form 160 categories, which is by far the largest test set. It contains more disturbing targets, making it very challenging.", + "bbox": [ + 75, + 468, + 468, + 619 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Four evaluation metrics are used for comparison, including MAE [33], $E_{\\phi}^{max}$ [6], $S_{\\alpha}$ [5] and $F_{\\beta}^{max}$ [1], which are to assess the average pixel-wise absolute difference, local and global similarity, structural similarity between the predictions and the ground truths, and the weighted harmonic mean of precision and recall, respectively.", + "bbox": [ + 75, + 623, + 468, + 715 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Comparisons with State-of-the-art Methods", + "text_level": 1, + "bbox": [ + 76, + 736, + 447, + 753 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "On the basis of unified evaluation codes [7] for fair comparison, we compare our method with several state-of-the-art methods published in recent three years, including RCAN [20], CSMG [46], SSNM [44], GCAGC [47], GICD [50], ICNet [15], GCoNet [9], DeepACG [45], CoEG-Net [7], HrSSMN [48], CADC [49], DCFM [39]. Among them, CADC [49] and DCFM [39] are the latest cutting-edge methods. More experimental results can be found in the supplementary materials.", + "bbox": [ + 75, + 763, + 468, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Qualitative Results. Figure 6 shows some visual comparison results with four latest state-of-the-art methods, including DCFM, CADC, DeepACG and GCoNet. The selected four groups are very challenging and suffer from highly confusing interference co-objects, extremely complex backgrounds, and drastic scale changes. Specifically, in the group \"Baseball\", the co-salient object is very small and the interference objects have similar shapes. Because all other methods are trapped in overconfidence, obsessed with the wrong results, and unable to extricate themselves. They cannot accurately detect the tiny objects and the background is wrongly segmented. In the group \"Backpack\", benefiting from training with noisy samples, our model is more robust that can accurately locate co-salient objects and segment target details, while other methods DCFM, CADC and GCoNet misjudge the co-salient objects. Camouflage co-objects in complex backgrounds are generally acknowledged as difficult samples [18], and the co-salient items in the \"Gift box\" group are perfectly in place. Thanks to CoSOD-TB's ability to capture long-range dependencies, our model learns strong feature representation that is able to accurately segment the camouflage co-salient objects, while CADC and GCoNet cannot finely segment the objects, and the other methods even fail in detecting the targets. The group \"teddy bear\" is also very challenging, and the co-salient objects in it suffer from interference objects and backgrounds with similar colors. Besides, the objects also have very different sizes. In this case, our model still achieves promising results, demonstrating its strong robustness to a variety of challenging factors. However, other methods yield a large number of inaccurate segmentation regions.", + "bbox": [ + 496, + 357, + 893, + 843 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Quantitative Results. Figure 7 shows the PR and F-measure curves of our method and other state-of-the-art", + "bbox": [ + 498, + 869, + 893, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "19644", + "bbox": [ + 480, + 926, + 519, + 936 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/a6e745f2961ddb5612fb833de3595d57838f3f4118b6c08d3a166e04ceb4eba5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 94, + 87, + 354, + 251 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/99edc3e7150f9a04bdb9dc5e06abba64ec881c3e45bb05246a32c4dc9e1b4999.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 354, + 85, + 614, + 251 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/520c16fb7a0153f745e34bf8b8931bd9da222d2d9f43802bfe3db202108cd973.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 616, + 85, + 875, + 251 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/7bedf32491bcaba9436116a59b84f70b9f46a5d1acdce0420a4888440240dd33.jpg", + "image_caption": [ + "Figure 7. Comparisons with state-of-the-art methods since the year of 2019 in terms of PR and F-measure curves on three benchmark datasets." + ], + "image_footnote": [], + "bbox": [ + 93, + 255, + 354, + 406 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/01981976420ccf7d7042e0193aa7e54e7cd345ad809be4748e1232ac25b5ff34.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 354, + 255, + 614, + 406 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/4051291ff930bb6c7daa9699aedd6930d9c5b19efd289e33baaa41fdb329785e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 617, + 256, + 875, + 406 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/136123f8483aa28270822f20d726b808ff61c099e28dfc303496138d062b9fb5.jpg", + "table_caption": [ + "Table 1. Statistic comparisons of our model with the other state-of-the-arts since the year of 2019." + ], + "table_footnote": [], + "table_body": "
MethodsCoSal2015CoSOD3kCoCA
MAE↓Sα↑Eφmax↑Fβmax↑MAE↓Sα↑Eφmax↑Fβmax↑MAE↓Sα↑Eφmax↑Fβmax↑
RCAN(IJCAI2019)0.1260.7790.8420.7640.1300.7440.8080.6880.1600.6160.7020.422
CSMG(CVPR2019)0.1300.7740.8180.7770.1570.7110.7230.6450.1240.6320.7340.503
SSNM(AAAI2020)0.1020.7880.8430.7940.1200.7260.7560.6750.1160.6280.7410.482
GCAGC(CVPR2020)0.0850.8170.8660.8130.1000.7850.8160.7400.1180.6690.7540.523
GICD(ECCV2020)0.0720.8420.8840.8340.0890.7940.8310.7430.1250.6580.7010.504
ICNet(NIPS2020)0.0580.8570.9000.8580.0890.7940.8450.7620.1470.6540.7050.514
CoEGNet(TPAMI2021)0.0770.8360.8820.8320.0920.7620.8250.7360.1060.6120.7170.493
GCoNet(CVPR2021)0.0690.8450.8870.8470.0710.8020.8600.7500.1050.6730.7600.524
DeepACG(CVPR2021)0.0660.8530.8930.8470.0790.8110.8590.7790.1040.6850.7590.564
CADC(ICCV2021)0.0640.8660.9060.8620.0960.8010.8400.7590.1320.6810.7440.548
HrSSMN(TMM2022)0.0620.8450.8950.8410.0870.7880.8420.7530.1060.6710.7390.532
DCFM(CVPR2022)0.0670.8380.8920.8560.0670.8100.8740.8050.0850.7100.7830.598
OURS0.0530.8850.9330.8820.0610.8530.9110.8290.0950.7260.8080.599
", + "bbox": [ + 81, + 472, + 890, + 657 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "methods. As we can see, the PR curves produced by our method wrap around the curves generated by other methods. Also, all the F-measure curves of our method are above those generated by other methods.", + "bbox": [ + 75, + 684, + 468, + 744 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Meanwhile, we list the statistic comparison results of all compared methods in Table 1. It is clear that our approach achieves the most competitive result compared to its counterparts. Specifically, on the CoSal2015 dataset, our method achieves the best scores of 0.053, 0.885, 0.933, and 0.882 in terms of all metrics, with a significant gain of $0.5\\%$ , $1.9\\%$ , $2.7\\%$ and $2.0\\%$ , respectively, compared to the second-best preforming ICNet [15] and CADC [49]. On the most recent and difficult datasets, CoSOD3k and CoCA, our method also achieves the best performance. Especially", + "bbox": [ + 75, + 750, + 470, + 902 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "on CoSOD3k, our method reaches the best scores of 0.061, 0.853, 0.911 and 0.829 in terms of $MAE$ , $E_{\\phi}^{max}$ , $S_{\\alpha}$ and $F_{\\beta}^{max}$ , respectively, with a great gain of $0.6\\%$ , $4.3\\%$ , $3.7\\%$ and $2.4\\%$ compared to the second-best performing DCFM [39]. The effectiveness of our method has been fully validated on the large-scale and challenging datasets.", + "bbox": [ + 498, + 684, + 890, + 776 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Ablation Study", + "text_level": 1, + "bbox": [ + 500, + 786, + 653, + 801 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To verify the effectiveness of our designs, we conduct the ablation study of our method on three datasets. We have made a lot of changes to VST [23] and regard the modified version as the baseline. Table 2 lists the results, and we can observe that every key design in our model makes a significant contribution to the overall performance. Taking", + "bbox": [ + 496, + 809, + 892, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "19645", + "bbox": [ + 480, + 926, + 517, + 936 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/c23b576430664ccda45ace2cfbd4fd8b35dc78da64f5414f260df75c29bec77f.jpg", + "table_caption": [ + "Table 2. Ablations of our method on the effectiveness of the GEM , LVGB, and CoSOD-TB." + ], + "table_footnote": [], + "table_body": "
StrategiesCoSal2015CoSOD3kCoCA
GEMLVGBCoSOD-TBMAE ↓\\( S_{\\alpha} \\uparrow \\)\\( E_{\\phi}^{max} \\uparrow \\)\\( F_{\\beta}^{max} \\uparrow \\)MAE ↓\\( S_{\\alpha} \\uparrow \\)\\( E_{\\phi}^{max} \\uparrow \\)\\( F_{\\beta}^{max} \\uparrow \\)MAE ↓\\( S_{\\alpha} \\uparrow \\)\\( E_{\\phi}^{max} \\uparrow \\)\\( F_{\\beta}^{max} \\uparrow \\)
0.0600.8540.8870.8600.0750.7870.8630.7780.1050.7100.7850.564
✓0.0610.8830.9280.8770.0650.8340.8800.8170.1090.7100.7880.571
✓0.0580.8770.9260.8720.0630.8400.8710.8050.1000.7190.7890.574
✓0.0530.8740.8930.8800.0620.8420.9000.8140.0980.7160.7980.583
✓✓0.0550.8690.9300.8760.0630.8470.9060.8190.1040.7240.8020.597
✓✓0.0600.8860.9250.8720.0690.8500.8950.8230.1000.7180.7920.587
✓✓0.0540.8800.9190.8780.0610.8490.8890.8290.0960.7200.8050.595
✓✓✓0.0530.8850.9330.8820.0610.8530.9110.8290.0950.7260.8080.599
", + "bbox": [ + 81, + 99, + 893, + 220 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/37e499642ec966a5aedf7adc854d385a3fcd23e6f6b1c4ece0e712283928f47f.jpg", + "table_caption": [ + "Table 3. Complexity analyses. The number of inputs is set to 5 to maintain unity." + ], + "table_footnote": [], + "table_body": "
methodsFLOPs(G)↓param.(M)↓runtime(fps)↑Fβmax↑
GICD(ECCV20)364.7278.040.80.504
GCoNet(CVPR21)259.9142.0116.20.524
CADC(ICCV21)330.0392.818.00.548
DCFM(CVPR22)251.9142.384.40.598
OURS211.852.331.00.599
", + "bbox": [ + 81, + 256, + 467, + 325 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/1c635b45307c1fbe73aa2cbccc55df733e7ad7676be5a92ef6eb938f076349a2.jpg", + "image_caption": [ + "Figure 8. Comparison with the current most competitive method DCFM [39] in simulated real scenes." + ], + "image_footnote": [], + "bbox": [ + 81, + 342, + 455, + 474 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "the results in CoSal2015 as examples, without GEM, the performance of our model gets worse in terms of all metrics especially for the $S_{\\alpha}$ that drops from 0.885 to 0.869 by $1.6\\%$ . With the GEM alone, the performance of our model improves greatly, and especially $S_{\\alpha}$ and $E_{\\phi}^{max}$ have a gain of $2.9\\%$ and $4.1\\%$ , respectively. Also, LVGB achieves the same excellent effect as GEM, because the uncertainty features provided by the LVGB can help the model consider more possibilities to enhance the robustness and overcome the over-confidence to non-salient regions. The CoSOD-TB makes important contributions to fine segmentation of images. With the CoSOD-TB, the MAE gets $0.7\\%$ better from 0.060 to 0.053. The results on CoSOD3k and CoCA show the same trend, proving the effectiveness of each design of our method.", + "bbox": [ + 75, + 530, + 467, + 757 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.5. Practical Application", + "text_level": 1, + "bbox": [ + 76, + 768, + 276, + 787 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "For the practicability of the algorithm, we conduct complexity comparisons with the same settings as DCFM [39]. The results can be seen in Table 3. Our method achieves both smallest FLOPs, parameters and best performance in terms of $F_{\\beta}^{max}$ . Meanwhile, the runtime of our method is 31 fps, meeting the real-time requirements. This shows that our method is more suitable for deployment to various ap", + "bbox": [ + 75, + 794, + 467, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "plication devices.", + "bbox": [ + 500, + 233, + 617, + 247 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We further test the performance of the model in the real scenes. In the real-life scenarios, there may not always exist co-salient objects within the group of images [22]. Figure 8 shows the images in the simulated real scenes that may not always exist co-objects. In these cases, the model should segment the co-objects existing in most images, while outputting all-zero masks on images without co-objects. Our method is better than the most competitive method DCFM [39] in dealing with real situations. When dealing with the images without co-objects, our model produces all-zero maps as much as possible. It shows that our method is more practical than other methods. More experimental results can be found in the supplementary materials.", + "bbox": [ + 496, + 250, + 890, + 446 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 465, + 617, + 479 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we have discovered that the paradigm of group consensus assumption has reduced the model's robustness and practical application value when confronted with irrelevant images in groups. A CoSOD model learning framework that is distinct from the classic CoSOD model learning framework has been developed. First, the group exchange-masking strategy has been devised, which is capable of automatically selecting the most informative noisy images from two groups and exchanging them in each other's groups to assist the model in learning more robust representations encoded with rich group consensus information. Second, the latent variable generator branch has been created to provide uncertainly-based global stochastic features that can regulate intermediate features from other branches to focus on co-objects. Third, the CoSOD transformer branch has been created to capture correlation-based global characteristics that carry information about group consistency. These branches' features are concatenated and put into a transformer-based decoder, yielding high-quality co-saliency maps. Extensive evaluations with and without irrelevant images have demonstrated the superiority of our method over a variety of state-of-the-art methods. In the future, we plan to build a new dataset for real scenarios that can provide a platform to evaluate model robustness and practical value in a more reliable way. Further more, we will consider the potential application of our method in defending adversarial attacks [35].", + "bbox": [ + 496, + 493, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "19646", + "bbox": [ + 480, + 926, + 517, + 936 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 174, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Radhakrishna Achanta, Sheila Hemami, Francisco Estrada, and Sabine Susstrunk. Frequency-tuned salient region detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1597-1604. IEEE, 2009. 6", + "[2] Kai-Yueh Chang, Tyng-Luh Liu, and Shang-Hong Lai. From co-saliency to co-segmentation: An efficient and fully unsupervised energy minimization model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2129-2136. IEEE, 2011. 1, 2", + "[3] Paul Christiano, Zain Shah, Igor Mordatch, Jonas Schneider, Trevor Blackwell, Joshua Tobin, Pieter Abbeel, and Wojciech Zaremba. Transfer from simulation to real world through learning deep inverse dynamics model. arXiv preprint arXiv:1610.03518, 2016. 2", + "[4] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 5", + "[5] Deng-Ping Fan, Ming-Ming Cheng, Yun Liu, Tao Li, and Ali Borji. Structure-measure: A new way to evaluate foreground maps. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4548-4557, 2017. 6", + "[6] Deng-Ping Fan, Cheng Gong, Yang Cao, Bo Ren, MingMing Cheng, and Ali Borji. Enhanced-alignment measure for binary foreground map evaluation. arXiv preprint arXiv:1805.10421, 2018. 6", + "[7] Deng-Ping Fan, Tengpeng Li, Zheng Lin, Ge-Peng Ji, Dingwen Zhang, Ming-Ming Cheng, Huazhu Fu, and Jianbing Shen. Re-thinking co-salient object detection. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2021. 2, 6", + "[8] Deng-Ping Fan, Zheng Lin, Ge-Peng Ji, Dingwen Zhang, Huazhu Fu, and Ming-Ming Cheng. Taking a deeper look at co-salient object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2919-2929, 2020. 1, 6", + "[9] Qi Fan, Deng-Ping Fan, Huazhu Fu, Chi-Keung Tang, Ling Shao, and Yu-Wing Tai. Group collaborative learning for cosalient object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12288–12298, 2021. 1, 2, 6", + "[10] Huazhu Fu, Xiaochun Cao, and Zhuowen Tu. Cluster-based co-saliency detection. IEEE Transactions on Image Processing, 22(10):3766-3778, 2013. 1, 2", + "[11] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 770-778, 2016. 5", + "[12] Kuang-Jui Hsu, Yen-Yu Lin, and Yung-Yu Chuang. Coattention cnns for unsupervised object co-segmentation. In Proceedings of the International Joint Conference on Artificial Intelligence, pages 748-756, 2018. 2" + ], + "bbox": [ + 78, + 114, + 470, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[13] David E Jacobs, Dan B Goldman, and Eli Shechtman. Cosaliency: Where people look when comparing images. In Proceedings of the 23nd Annual ACM Symposium on User Interface Software and Technology, pages 219-228, 2010. 1, 2", + "[14] Koteswar Rao Jerripothula, Jianfei Cai, and Junsong Yuan. Cats: Co-saliency activated tracklet selection for video colocalization. In Proceedings of the European Conference on Computer Vision, pages 187-202. Springer, 2016. 1", + "[15] Wen-Da Jin, Jun Xu, Ming-Ming Cheng, Yi Zhang, and Wei Guo. Icnet: Intra-saliency correlation network for co-saliency detection. Advances in Neural Information Processing Systems, 33:18749-18759, 2020. 1, 2, 6, 7", + "[16] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6", + "[17] Simon Kohl, Bernardino Romera-Paredes, Clemens Meyer, Jeffrey De Fauw, Joseph R Ledsam, Klaus Maier-Hein, SM Eslami, Danilo Jimenez Rezende, and Olaf Ronneberger. A probabilistic u-net for segmentation of ambiguous images. Advances in Neural Information Processing Systems, 31, 2018. 4", + "[18] Aixuan Li, Jing Zhang, Yunqiu Lv, Bowen Liu, Tong Zhang, and Yuchao Dai. Uncertainty-aware joint salient object and camouflaged object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10071-10081, 2021. 6", + "[19] Bo Li, Zhengxing Sun, and Yuqi Guo. Supervae: Superpixelwise variational autoencoder for salient object detection. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 33, pages 8569-8576, 2019. 2", + "[20] Bo Li, Zhengxing Sun, Lv Tang, Yunhan Sun, and Jinlong Shi. Detecting robust co-saliency with recurrent coattention neural network. In Proceedings of the International Joint Conference on Artificial Intelligence, volume 2, page 6, 2019. 1, 6", + "[21] Yijun Li, Keren Fu, Zhi Liu, and Jie Yang. Efficient saliency-model-guided visual co-saliency detection. IEEE Signal Processing Letters, 22(5):588-592, 2014. 2", + "[22] Jiawei Liu, Jing Zhang, Kaihao Zhang, and Nick Barnes. Generalised co-salient object detection. arXiv preprint arXiv:2208.09668, 2022. 8", + "[23] Nian Liu, Ni Zhang, Kaiyuan Wan, Ling Shao, and Junwei Han. Visual saliency transformer. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2021. 7", + "[24] Xuanqing Liu, Minhao Cheng, Huan Zhang, and Cho-Jui Hsieh. Towards robust neural networks via random self-ensemble. In Proceedings of the European Conference on Computer Vision, pages 369–385, 2018. 3", + "[25] Xuanqing Liu, Yao Li, Chongruo Wu, and Cho-Jui Hsieh. Adv-bnn: Improved adversarial defense through robust bayesian neural network. arXiv preprint arXiv:1810.01279, 2018.3", + "[26] Aleksander Madry, Aleksandar Makelov, Ludwig Schmidt, Dimitris Tsipras, and Adrian Vladu. Towards deep learning models resistant to adversarial attacks. In International Conference on Learning Representations, 2018. 4" + ], + "bbox": [ + 501, + 92, + 893, + 901 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "19647", + "bbox": [ + 480, + 926, + 519, + 936 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[27] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. Advances in Neural Information Processing Systems, 32:8026-8037, 2019. 5", + "[28] Lerrel Pinto, James Davidson, Rahul Sukthankar, and Abhi nav Gupta. Robust adversarial reinforcement learning. In International Conference on Machine Learning, pages 2817-2826. PMLR, 2017. 2", + "[29] Aravind Rajeswaran, Sarvjeet Ghotra, Balaraman Ravindran, and Sergey Levine. Epopt: Learning robust neural network policies using model ensembles. arXiv preprint arXiv:1610.01283, 2016. 3", + "[30] Andrei A Rusu, Matej Večerík, Thomas Rothörl, Nicolas Heess, Razvan Pascanu, and Raia Hadsell. Sim-to-real robot learning from pixels with progressive nets. In Conference on robot learning, pages 262-270. PMLR, 2017. 2", + "[31] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556, 2014. 5", + "[32] Kihyuk Sohn, Honglak Lee, and Xinchen Yan. Learning structured output representation using deep conditional generative models. Advances in Neural Information Processing Systems, 28, 2015. 2, 4", + "[33] Chong Wang, Zheng-Jun Zha, Dong Liu, and Hongtao Xie. Robust deep co-saliency detection with group semantic. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 33, pages 8917-8924, 2019. 1, 5, 6", + "[34] Lijun Wang, Huchuan Lu, Yifan Wang, Mengyang Feng, Dong Wang, Baocai Yin, and Xiang Ruan. Learning to detect salient objects with image-level supervision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 136-145, 2017. 5", + "[35] Yisen Wang, Difan Zou, Jinfeng Yi, James Bailey, Xingjun Ma, and Quanquan Gu. Improving adversarial robustness requires revisiting misclassified examples. In International Conference on Learning Representations, 2019. 8", + "[36] Lina Wei, Shanshan Zhao, Omar El Farouk Bourahla, Xi Li, and Fei Wu. Group-wise deep co-saliency detection. arXiv preprint arXiv:1707.07381, 2017. 1, 2", + "[37] Cihang Xie, Jianyu Wang, Zhishuai Zhang, Zhou Ren, and Alan Yuille. Mitigating adversarial effects through randomization. arXiv preprint arXiv:1711.01991, 2017. 3", + "[38] Xi Yang, Shaoyi Li, Jun Ma, Jun-yan Yang, and Jie Yan. Co-saliency-regularized correlation filter for object tracking. Signal Processing: Image Communication, 103:116655, 2022. 1", + "[39] Siyue Yu, Jimin Xiao, Bingfeng Zhang, and Eng Gee Lim. Democracy does matter: Comprehensive feature mining for co-salient object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 979–988, 2022. 1, 6, 7, 8", + "[40] Li Yuan, Yunpeng Chen, Tao Wang, Weihao Yu, Yujun Shi, Zi-Hang Jiang, Francis EH Tay, Jiashi Feng, and Shuicheng Yan. Tokens-to-token vit: Training vision transformers from scratch onImagenet. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2021. 5" + ], + "bbox": [ + 78, + 90, + 468, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[41] Chen Zhang, Riccardo Barbano, and Bangti Jin. Conditional variational autoencoder for learned image reconstruction. Computation, 9(11):114, 2021. 2", + "[42] Dingwen Zhang, Junwei Han, Chao Li, and Jingdong Wang. Co-saliency detection via looking deep and wide. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2994-3002, 2015. 2, 6", + "[43] Jing Zhang, Deng-Ping Fan, Yuchao Dai, Saeed Anwar, Fatemeh Saleh, Sadegh Aliakbarian, and Nick Barnes. Uncertainty inspired rgb-d saliency detection. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2021. 2, 4", + "[44] Kaihua Zhang, Jin Chen, Bo Liu, and Qingshan Liu. Deep object co-segmentation via spatial-semantic network modulation. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 34, pages 12813–12820, 2020. 1, 6", + "[45] Kaihua Zhang, Mingliang Dong, Bo Liu, Xiao-Tong Yuan, and Qingshan Liu. Deepacg: Co-saliency detection via semantic-aware contrast gromov-wasserstein distance. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13703-13712, 2021. 6", + "[46] Kaihua Zhang, Tengpeng Li, Bo Liu, and Qingshan Liu. Cosaliency detection via mask-guided fully convolutional networks with multi-scale label smoothing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3095-3104, 2019. 1, 2, 6", + "[47] Kaihua Zhang, Tengpeng Li, Shiwen Shen, Bo Liu, Jin Chen, and Qingshan Liu. Adaptive graph convolutional network with attention graph clustering for co-saliency detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9050-9059, 2020. 2, 6", + "[48] Kaihua Zhang, Yang Wu, Mingliang Dong, Bo Liu, Dong Liu, and Qingshan Liu. Deep object co-segmentation and c-osaliency detection via high-order spatial-semantic network modulation. IEEE Transactions on Multimedia, 2022. 1, 6", + "[49] Ni Zhang, Junwei Han, Nian Liu, and Ling Shao. Summarize and search: Learning consensus-aware dynamic convolution for co-saliency detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4167-4176, 2021. 6, 7", + "[50] Zhao Zhang, Wenda Jin, Jun Xu, and Ming-Ming Cheng. Gradient-induced co-saliency detection. In Proceedings of the European Conference on Computer Vision, pages 455-472. Springer, 2020. 1, 2, 6" + ], + "bbox": [ + 501, + 92, + 893, + 726 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "19648", + "bbox": [ + 480, + 926, + 519, + 936 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2023/Co-Salient Object Detection With Uncertainty-Aware Group Exchange-Masking/9cd15a5b-7ef1-4ba9-a24c-48575755f531_model.json b/2023/Co-Salient Object Detection With Uncertainty-Aware Group Exchange-Masking/9cd15a5b-7ef1-4ba9-a24c-48575755f531_model.json new file mode 100644 index 0000000000000000000000000000000000000000..dcf1105ffe59375eaeffc816ae43f8cc93c5fbc6 --- /dev/null +++ b/2023/Co-Salient Object Detection With Uncertainty-Aware Group Exchange-Masking/9cd15a5b-7ef1-4ba9-a24c-48575755f531_model.json @@ -0,0 +1,2376 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.086, + 0.131, + 0.885, + 0.154 + ], + "angle": 0, + "content": "Co-Salient Object Detection with Uncertainty-aware Group Exchange-Masking" + }, + { + "type": "text", + "bbox": [ + 0.204, + 0.18, + 0.766, + 0.199 + ], + "angle": 0, + "content": "Yang Wu\\(^{1}\\) Huihui Song\\(^{1*}\\) Bo Liu\\(^{2}\\) Kaihua Zhang\\(^{1}\\) Dong Liu\\(^{3}\\)" + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.199, + 0.88, + 0.217 + ], + "angle": 0, + "content": "\\(^{1}\\)B-DAT and CICAET, Nanjing University of Information Science and Technology, Nanjing, China" + }, + { + "type": "text", + "bbox": [ + 0.278, + 0.217, + 0.692, + 0.234 + ], + "angle": 0, + "content": "\\(^{2}\\)Walmart Global Tech, Sunnyvale, CA, 94086, USA" + }, + { + "type": "text", + "bbox": [ + 0.321, + 0.234, + 0.65, + 0.251 + ], + "angle": 0, + "content": "3Netflix Inc, Los Gatos, CA, 95032, USA" + }, + { + "type": "text", + "bbox": [ + 0.377, + 0.255, + 0.583, + 0.269 + ], + "angle": 0, + "content": "songhuihui@nuist.edu.cn" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.305, + 0.313, + 0.32 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.336, + 0.474, + 0.73 + ], + "angle": 0, + "content": "The traditional definition of co-salient object detection (CoSOD) task is to segment the common salient objects in a group of relevant images. Existing CoSOD models by default adopt the group consensus assumption. This brings about model robustness defect under the condition of irrelevant images in the testing image group, which hinders the use of CoSOD models in real-world applications. To address this issue, this paper presents a group exchange-masking (GEM) strategy for robust CoSOD model learning. With two groups of image containing different types of salient object as input, the GEM first selects a set of images from each group by the proposed learning based strategy, then these images are exchanged. The proposed feature extraction module considers both the uncertainty caused by the irrelevant images and group consensus in the remaining relevant images. We design a latent variable generator branch which is made of conditional variational autoencoder to generate uncertainly-based global stochastic features. A CoSOD transformer branch is devised to capture the correlation-based local features that contain the group consistency information. At last, the output of two branches are concatenated and fed into a transformer-based decoder, producing robust co-saliency prediction. Extensive evaluations on co-saliency detection with and without irrelevant images demonstrate the superiority of our method over a variety of state-of-the-art methods." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.757, + 0.208, + 0.772 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.782, + 0.47, + 0.843 + ], + "angle": 0, + "content": "Co-salient object detection (CoSOD) is to segment the common salient objects in a group of relevant images. By detecting the co-salient object in a group of images, the images' background and redundant content are re" + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.305, + 0.653, + 0.449 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.654, + 0.305, + 0.887, + 0.449 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.464, + 0.893, + 0.534 + ], + "angle": 0, + "content": "Figure 1. (a) When there exists an irrelevant image in the test group, the state-of-the-art CoSOD models such as DCFM [39] and GCoNet [9] tend to generate false positive predictions for the noisy image, yet our method can achieve an accurate prediction due to the use of group exchange-masking strategy in (b)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.565, + 0.892, + 0.611 + ], + "angle": 0, + "content": "moved, which helps the downstream tasks such as object tracking [38], co-segmentation [48] and video colocalization [14], to name a few." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.614, + 0.893, + 0.901 + ], + "angle": 0, + "content": "Group consensus assumption is widely used by existing CoSOD models, i.e., these models presume that all images in the same group contain the common salient targets. The widely-used CoSOD benchmark datasets, such as COCO-SEG [33], CoCA [50], and CoSOD3k [8] organize the training and testing images that contain the same object as a group. Many existing CoSOD models consider the group consensus characteristic in modeling. For example, early works such as [2, 10, 13] extract hand-crafted features for inter-image co-object correspondence discovery. The deep learning models proposed in [15, 20, 36, 39, 44, 46, 48] use one group of relevant images as training data input for consensus representation learning. Among them, a variety of novel model design techniques have been developed to make full use of the group consistency characteristic, such as the low-rank feature learning [46], co-attention model [20] and intra-saliency correlation learning [15]. The issues of the group consensus assumption are partially studied in recent literature [9], which further models the inter" + }, + { + "type": "page_footnote", + "bbox": [ + 0.076, + 0.851, + 0.47, + 0.9 + ], + "angle": 0, + "content": "*Corresponding author. This work is supported in part by National Key Research and Development Program of China under Grant No. 2018AAA0100400, in part by the NSFC under Grant Nos. 62276141, 61872189." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.925, + 0.519, + 0.937 + ], + "angle": 0, + "content": "19639" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.47, + 0.121 + ], + "angle": 0, + "content": "group separability for more discriminative feature learning by a group collaborating module." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.123, + 0.471, + 0.468 + ], + "angle": 0, + "content": "In this paper, we find that the group consensus assumption also restricts the CoSOD model's robustness against the images without common object. As illustrated by Figure 1(a), state-of-the-art CoSOD models tend to output false positive predictions for the irrelevant image. This issue hinders the use of CoSOD models in real-world applications where the testing inputs are likely to contain irrelevant images. To enhance the model's robustness, we propose a learning framework called group exchange-masking (GEM). The GEM is illustrated by Figure 1(b). Given two image groups that contain different types of co-salient objects, we exchange several images between one group and the other. Those exchanged images are called noisy images. The number of noisy images is chosen to be less than the number of remaining relevant images in the group, so that the co-salient object in the noisy images forms a negative object but not the dominant co-salient object. The \"masking\" strategy refers to the label regeneration of the noisy images. Because there is no co-salient object, in the regenerated label, the original ground-truth object is masked. The learning objective is to correctly predict both the co-salient objects in the original relevant images and the added noisy images." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.469, + 0.471, + 0.802 + ], + "angle": 0, + "content": "Adding noisy images to the training image group brings about uncertainty to the CoSOD model learning since there is some probability of no expected common object in each image. We design a dual-path image feature extraction module to model the group uncertainty in addition to the group consensus property. Specifically, we design a latent variable generator branch (LVGB) to extract the uncertainty-based global image features. The LVGB module is motivated by the conditional variational autoencoder (CVAE) [32] that is widely used to address the uncertainty in vision applications including image background modeling [19], RGB-D saliency detection [43] and image reconstruction [41]. In parallel with LVGB, we feed the image group into a CoSOD Transformer Branch (CoSOD-TB). The CoSOD-TB partitions each image group into local patches, and the attention mechanism in the transformer enables this branch to model patch-wise correlation-based local features. As a result, the group consistency information can be captured by this branch. The outputs of the two branches are concatenated and fed into a transformer-based decoder for co-saliency prediction. The proposed model has the following technical contributions." + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.81, + 0.472, + 0.902 + ], + "angle": 0, + "content": "- A robust CoSOD model learning mechanism, called group exchange-masking is proposed. By exchanging images between two groups, we augment the training data containing irrelevant images as noise to enhance model's robustness. This is different from the traditional CoSOD model learning frameworks that use" + }, + { + "type": "text", + "bbox": [ + 0.533, + 0.092, + 0.813, + 0.108 + ], + "angle": 0, + "content": "groups of relevant images as training data." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.117, + 0.892, + 0.208 + ], + "angle": 0, + "content": "- We propose a dual-path feature extraction module composed of the LVGB and the CoSOD-TB. The LVGB is designed to model the uncertainty of cosalient object existence. The CoSOD-TB is for the consensus feature extraction of the salient object in the relevant images." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.219, + 0.895, + 0.326 + ], + "angle": 0, + "content": "- Extensive evaluations on three benchmark datasets, including CoSal2015 [42], CoCA [15], and CoSOD3k [7] show that the superiority of the proposed model to the state-of-the-art methods in terms of all evaluation metrics. Besides, the proposed model demonstrates good robustness for dealing with noisy data without co-salient objects." + }, + { + "type": "list", + "bbox": [ + 0.518, + 0.117, + 0.895, + 0.326 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.351, + 0.642, + 0.367 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.376, + 0.749, + 0.392 + ], + "angle": 0, + "content": "2.1. Co-salient Object Detection" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.399, + 0.892, + 0.793 + ], + "angle": 0, + "content": "In the past, CoSOD methods used to extract handcrafted features such as Gabor and SIFT features from images and then detect co-saliency by utilizing the consistency of low-level features between the images being tested [2]. A series of studies attempt to capture the intra-image constraints by employing a manifold ranking scheme to produce saliency maps [21], or using a global association constraint with clustering [10], or translational alignment [13]. More recently, there is a surge of deep learning-based CoSOD models that learn feature representation and saliency predictor in an end-to-end manner [9, 12, 36, 46, 47, 50]. Wei et al. [36] designs a collaborative learning framework for CoSOD that discovers the collaborative and interactive relationships between intra-image and single-image feature representations in a group-wise manner. Hsu et al. [12] present an unsupervised CNN-based model for CoSOD. In [46], a hierarchical framework is proposed for CoSOD where the initial CoSOD results generated by the CNN model is refined by label smoothing. Zhang et al. [50] propose a gradient-induced model for CoSOD that utilizes the image gradient information to induce more attentions to the discriminative co-salient feature learning. In [47], a deep graph neural network model is proposed to characterize the intra-image and inter-image region correspondence for CoSOD. In [9], a group collaborative learning strategy is proposed to explore inter-group relations for discriminating feature learning." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.802, + 0.818, + 0.818 + ], + "angle": 0, + "content": "2.2. Robust Model and Feature Learning" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.826, + 0.894, + 0.903 + ], + "angle": 0, + "content": "In past years, with the development of deep learning, robust model and feature learning have received more and more attention as a possible solution to overcoming the bottleneck [28]. Various methods have been proposed to improve the robustness of the model. In [3] and [30], ways to" + }, + { + "type": "footer", + "bbox": [ + 0.481, + 0.925, + 0.52, + 0.937 + ], + "angle": 0, + "content": "19640" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.089, + 0.895, + 0.349 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.36, + 0.895, + 0.445 + ], + "angle": 0, + "content": "Figure 2. Pipeline of the proposed method. First, two groups of images \\(\\{\\mathcal{I}_1,\\mathcal{I}_2\\}\\) with their masks \\(\\{\\mathcal{V}_1,\\mathcal{V}_2\\}\\) are processed by group exchange-masking, yielding the exchanged images \\(\\{\\mathcal{I}_{e1},\\mathcal{I}_{e2}\\}\\) and their corresponding masked labels \\(\\{\\mathcal{V}_{e1},\\mathcal{V}_{e2}\\}\\). Then, \\(\\{\\mathcal{I}_{e1},\\mathcal{I}_{e2}\\}\\) and \\(\\{\\mathcal{V}_{e1},\\mathcal{V}_{e2}\\}\\) are passed through the latent variable generator branch, extracting the uncertainty-based global image features \\(\\{\\mathcal{V}_1,\\mathcal{V}_2\\}\\) to eliminate the bias caused by noisy images. Meanwhile, \\(\\{\\mathcal{I}_{e1},\\mathcal{I}_{e2}\\}\\) are passed through the weight-shared CoSOD transformer branch, producing two groups of feature sequences \\(\\{\\mathcal{F}_1,\\mathcal{F}_2\\}\\) with group consistency information and long-range independent information. Finally, \\(\\mathcal{F}\\) and \\(\\mathcal{V}\\) are concatenated and fed into the transformer decoder, yielding the predicted co-saliency maps \\(\\hat{\\mathcal{O}}\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.469, + 0.473, + 0.651 + ], + "angle": 0, + "content": "transfer the policy learned in a simulator to the real world are proposed to make features more robust. Rajeswaran et al. [29] learn a robust policy by sampling the worst case trajectories from a class of parametrized models, to learn a robust policy. Xie et al. [37] enhance the robustness of the model through randomly resizing and padding the training images. Several works leverage randomness to improve the robustness of models. Liu et al. [24] propose a noise layer that introduces randomness to both the input and the hidden layer output. Liu et al. [25] introduce a new min-max formulation that combines adversarial training with Bayesian Neural Networks, which achieves promising results." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.664, + 0.251, + 0.681 + ], + "angle": 0, + "content": "3. Proposed Method" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.689, + 0.473, + 0.901 + ], + "angle": 0, + "content": "Figure 2 illustrates the pipeline of our framework. Given two groups of \\(2N\\) images \\(\\mathcal{I}_i = \\{I_i^n\\in \\mathbb{R}^{H\\times W\\times 3}\\}_{n = 1}^N,i = 1,2\\) with height \\(H\\) and width \\(W\\) and manually-labeled binary masks \\(\\mathcal{V}_i = \\{\\mathbf{Y}_i^n\\in \\mathbb{R}^{H\\times W}\\}_{n = 1}^N\\) with different categories as input, we employ the GEM model to reorganize them, yielding \\(\\mathcal{I}_{ei}\\) and \\(\\mathcal{V}_{ei}\\). Each pair of \\(\\{\\mathcal{I}_{ei},\\mathcal{V}_{ei}\\}\\) are fed to the LVGB. In LVGB, we draw \\(\\mathbf{z}_i\\in \\mathbb{R}^K\\) from the Gaussian distribution \\(P_{\\theta}(\\mathbf{z}_i|\\mathcal{I}_{ei})\\) with the learning parameters \\(\\pmb{\\theta}\\) and the posterior of \\(\\mathbf{z}_i\\) is formulated as \\(Q_{\\phi}(\\mathbf{z}_i|\\mathcal{I}_{ei},\\mathcal{V}_{ei})\\) with the learning parameters \\(\\phi\\). We use the KL-Divergence metric to reduce the distance between \\(P_{\\theta}(\\mathbf{z}_i|\\mathcal{I}_{ei})\\) and \\(Q_{\\phi}(\\mathbf{z}_i|\\mathcal{I}_{ei},\\mathcal{V}_{ei})\\). We further re-parameterize and expand \\(\\mathbf{z}_i\\) to yield a latent sequence \\(\\mathcal{V}_i\\) which has the same dimension as the feature sequences produced by trans" + }, + { + "type": "title", + "bbox": [ + 0.502, + 0.468, + 0.762, + 0.484 + ], + "angle": 0, + "content": "Algorithm 1 Group exchange-masking" + }, + { + "type": "text", + "bbox": [ + 0.507, + 0.486, + 0.772, + 0.503 + ], + "angle": 0, + "content": "Input: \\(\\mathcal{I}_1 = \\{I_1^n\\}_{n=1}^N\\), \\(\\mathcal{Y}_1 = \\{\\mathbf{Y}_1^n\\}_{n=1}^N\\)," + }, + { + "type": "equation", + "bbox": [ + 0.558, + 0.503, + 0.772, + 0.518 + ], + "angle": 0, + "content": "\\[\n\\mathcal {I} _ {2} = \\left\\{I _ {2} ^ {n} \\right\\} _ {n = 1} ^ {N}, \\mathcal {Y} _ {2} = \\left\\{\\mathbf {Y} _ {2} ^ {n} \\right\\} _ {n = 1} ^ {N}.\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.518, + 0.695, + 0.533 + ], + "angle": 0, + "content": "Output: \\(\\mathcal{I}_{e1},\\mathcal{Y}_{e1},\\mathcal{I}_{e2},\\mathcal{Y}_{e2}\\)" + }, + { + "type": "text", + "bbox": [ + 0.509, + 0.535, + 0.892, + 0.565 + ], + "angle": 0, + "content": "1: \\(\\mathcal{I}_{e1},\\mathcal{Y}_{e1},\\mathcal{I}_{e2},\\mathcal{Y}_{e2}\\gets\\) Rearranging \\(\\mathcal{I}_1,\\mathcal{Y}_1,\\mathcal{I}_2,\\mathcal{Y}_2\\) via solving (1)." + }, + { + "type": "text", + "bbox": [ + 0.51, + 0.566, + 0.681, + 0.579 + ], + "angle": 0, + "content": "2: for \\( n = 1, 2, \\ldots, k \\) do" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.58, + 0.687, + 0.595 + ], + "angle": 0, + "content": "3: \\( I_{e1}^{n} \\gets I_{2}^{n}, I_{e2}^{n} \\gets I_{1}^{n} \\)," + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.595, + 0.696, + 0.61 + ], + "angle": 0, + "content": "4: \\(\\mathbf{Y}_{e1}^{n}\\gets \\mathbf{0}\\in \\mathbb{R}^{H\\times W\\times 1},\\)" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.61, + 0.696, + 0.626 + ], + "angle": 0, + "content": "5: \\(\\mathbf{Y}_{e2}^{\\vec{n}}\\gets \\mathbf{0}\\in \\mathbb{R}^{H\\times W\\times 1}\\)" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.628, + 0.583, + 0.638 + ], + "angle": 0, + "content": "6: end for" + }, + { + "type": "list", + "bbox": [ + 0.509, + 0.566, + 0.696, + 0.638 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.669, + 0.892, + 0.794 + ], + "angle": 0, + "content": "formers. Meanwhile, the image groups are passed through the CoSOD-TB with shared weights, producing the token sequences \\(\\mathcal{F}_i\\) which capture group consensus and long-range dependency information. At last, \\(\\mathcal{V}_i\\) and \\(\\mathcal{F}_i\\) are concatenated in channel dimension and passed through several subsequent processes as shown in Figure 2 followed by an up-sampling layer, yielding the corresponding predicted cosaliency maps \\(\\hat{\\mathcal{O}} = \\{\\hat{\\mathbf{O}}^n\\in \\mathbb{R}^{H\\times W}\\}_{n = 1}^{2N}\\)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.802, + 0.741, + 0.818 + ], + "angle": 0, + "content": "3.1. Group Exchange-Masking" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.825, + 0.892, + 0.901 + ], + "angle": 0, + "content": "The widely-used group consensus assumption limits the robustness of the CoSOD model, especially when there are test images without common objects in the group. To this end, we reorganize the two input groups with the GEM tailored to CoSOD. The GEM is summarised in Algorithm 1:" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.927, + 0.518, + 0.938 + ], + "angle": 0, + "content": "19641" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.08, + 0.091, + 0.473, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.164, + 0.239, + 0.382, + 0.253 + ], + "angle": 0, + "content": "Figure 3. Architecture of the LVGB." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.278, + 0.47, + 0.413 + ], + "angle": 0, + "content": "taking \\(\\mathcal{I}_i\\) and \\(\\mathcal{Y}_i\\) as input, the goal of GEM is to obtain \\(\\mathcal{I}_{ei}\\) and \\(\\mathcal{Y}_{ei}\\) which contain noisy images in each group and corresponding full masked ground truths. Specifically, we rearrange the two groups of images and their binary ground-truth masks in a decreasing order by solving (1). After this, we sample the top-\\(k\\) images of each group as noisy images and exchange them in each other's group, ensuring the maximum destructiveness of noisy images and the effectiveness of training." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.414, + 0.47, + 0.445 + ], + "angle": 0, + "content": "This training process is essentially a min-max optimization process [26] as" + }, + { + "type": "equation", + "bbox": [ + 0.124, + 0.455, + 0.47, + 0.497 + ], + "angle": 0, + "content": "\\[\n\\min _ {\\varphi} \\sum_ {n = 1} ^ {k} \\sum_ {i = 1} ^ {2} \\max _ {\\mathcal {I} _ {i} (n,:)} \\mathcal {L} \\left(f _ {\\varphi} \\left(\\mathcal {I} _ {i} (n,:)\\right), \\hat {\\mathcal {O}} _ {i} (n,:)\\right), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.506, + 0.47, + 0.613 + ], + "angle": 0, + "content": "where the loss \\(\\mathcal{L}\\) is defined by (7), and \\(f\\) denotes our whole model with learnable parameters \\(\\varphi\\). Solving (1) finds out the most noisy images which can maximize the training loss, and our goal is to minimize the loss function w.r.t the intra-group noisy images to improve the model's robustness. At last, we mask the corresponding labels of the noisy images with the all-zeros maps." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.621, + 0.319, + 0.637 + ], + "angle": 0, + "content": "3.2. Latent Variable Generator" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.645, + 0.469, + 0.765 + ], + "angle": 0, + "content": "By training the model with noisy images, the GEM is able to increase the model's robustness which can identify whether there exist co-objects or not. However, the model will focus on non-co-salient regions as a result of the uncertainty introduced by adding noisy images to the training image groups. This means that the model will become overconfident in the background regions in the early stage of training, leading to inaccurate prediction." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.469, + 0.87 + ], + "angle": 0, + "content": "To address this issue, we further propose the LVGB that generates a low-dimensional latent space to represent the most common patterns in image groups [17]. The latent variables sampled from the feature space characterize uncertainties of group consensus and can be used to modulate the intermediate features from other branches to highlight the co-objects." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.47, + 0.901 + ], + "angle": 0, + "content": "Figure 3 shows the details of our LVGB. The generator takes \\(\\mathcal{I}_{ei}\\) and latent variable \\(\\mathbf{z}_i\\) as input, and uses an encoder" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.892, + 0.184 + ], + "angle": 0, + "content": "to produce stochastic prediction \\(P_{\\omega}(\\mathcal{Y}_{ei}|\\mathcal{I}_{ei},\\mathbf{z}_i)\\) with learnable parameters \\(\\omega\\), where \\(\\mathcal{I}_{ei}\\) is conditional variable, and \\(\\mathcal{Y}_{ei}\\) is output variable. \\(\\mathbf{z}_i\\) is drawn from the Gaussian distribution \\(P_{\\theta}(\\mathbf{z}_i|\\mathcal{I}_{ei})\\), and the posterior of \\(\\mathbf{z}_i\\) is formulated as \\(Q_{\\phi}(\\mathbf{z}_i|\\mathcal{I}_{ei},\\mathcal{Y}_{ei})\\) via an encoder that maps the input variable to the latent space. The loss of the LVG is defined as [32,43]" + }, + { + "type": "equation", + "bbox": [ + 0.516, + 0.193, + 0.891, + 0.267 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} _ {L V G} = \\sum_ {i = 1} ^ {2} E _ {\\mathbf {z} \\sim Q _ {\\phi} (\\mathbf {z} _ {i} | \\mathcal {I} _ {e i}, \\mathcal {Y} _ {e i})} [ - \\log P _ {\\boldsymbol {\\omega}} (\\mathcal {Y} _ {e i} | \\mathcal {I} _ {e i}, \\mathbf {z} _ {i}) ] \\\\ + \\mathrm {K L} \\left(Q _ {\\phi} \\left(\\mathbf {z} _ {i} \\mid \\mathcal {I} _ {e i}, \\mathcal {Y} _ {e i}\\right) \\mid \\mid P _ {\\boldsymbol {\\theta}} \\left(\\mathbf {z} _ {i} \\mid \\mathcal {I} _ {e i}\\right)\\right), \\tag {2} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.268, + 0.892, + 0.296 + ], + "angle": 0, + "content": "where the Kullback-Leibler Divergence metric KL is defined as" + }, + { + "type": "equation", + "bbox": [ + 0.581, + 0.305, + 0.892, + 0.339 + ], + "angle": 0, + "content": "\\[\n\\operatorname {K L} (Q | | P) = \\sum Q (x) \\log \\frac {Q (x)}{P (x)}. \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.348, + 0.892, + 0.604 + ], + "angle": 0, + "content": "In (2), the prior net \\(P_{\\pmb{\\theta}}(\\mathbf{z}_i|\\mathcal{I}_{ei})\\) is defined as a Gaussian distribution that maps the image group \\(\\mathcal{I}_{ei}\\) to a low-dimensional latent space that encodes the most common patterns in each group. We randomly sample prior latent variable \\(\\mathbf{z}_i\\sim \\mathcal{N}(\\pmb{\\mu}_{1i},\\mathrm{diag}(\\pmb{\\sigma}_{1i}^2))\\), where \\(\\pmb{\\mu}_{1i},\\pmb{\\sigma}_{1i}^{2}\\in \\mathbb{R}^{K}\\) represent the mean and standard deviation vectors and \\(\\pmb{\\theta}\\) is the learnable parameter of the mapping function which consists five conv layers [43] as shown in Figure 3. In the posterior net, we use the similar encoder as the prior net to model \\(Q_{\\phi}(\\mathbf{z}_i|\\mathcal{I}_{ei},\\mathcal{Y}_{ei})\\) that maps the concatenation of \\(\\mathcal{I}_{ei}\\) and \\(\\mathcal{Y}_{ei}\\) to the posterior latent variable \\(\\mathbf{z}_i\\sim \\mathcal{N}(\\pmb{\\mu}_{2i},\\mathrm{diag}(\\pmb{\\sigma}_{2i}^2))\\) where \\(\\pmb{\\mu}_{2i},\\pmb{\\sigma}_{2i}^{2}\\in \\mathbb{R}^{K}\\). The KL in (2) is used to measure the difference between probability distributions prior \\(P_{\\pmb{\\theta}}(\\mathbf{z}_i|\\mathcal{I}_{ei})\\) and posterior \\(Q_{\\phi}(\\mathbf{z}_i|\\mathcal{I}_{ei},\\mathcal{Y}_{ei})\\). It represents the information loss that occurs when using \\(Q_{\\phi}(\\mathbf{z}_i|\\mathcal{I}_{ei},\\mathcal{Y}_{ei})\\) to approximate \\(P_{\\pmb{\\theta}}(\\mathbf{z}_i|\\mathcal{I}_{ei})\\), and a smaller value indicates greater similarity between the two distributions." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.605, + 0.892, + 0.786 + ], + "angle": 0, + "content": "Each position in latent space symbolizes potential labeling changes or other potential factors that could lead to a variety of co-saliency predictions [17, 43]. The proposed GEM can meet the needs for diverse ground truths as we learn to select the most difficult samples to mask their ground truths to all-zeros maps. We hope diverse annotations in the posterior net \\( Q_{\\phi}(\\mathbf{z}_i|\\mathcal{I}_{ei},\\mathcal{Y}_{ei}) \\) during training can compel the prior net \\( P_{\\theta}(\\mathbf{z}_i|\\mathcal{I}_{ei}) \\) to encode labeling variants of the supplied inputs \\( \\mathcal{I}_{ei} \\). The statistics \\( \\mathbf{z}_i \\) is further processed by feature expanding [43]. Given a pair of \\( (\\boldsymbol{\\mu}_k,\\boldsymbol{\\sigma}_k) \\) in each position of \\( K \\) dimensional vector, we parameterize them to obtain the latent vector" + }, + { + "type": "equation", + "bbox": [ + 0.627, + 0.799, + 0.891, + 0.815 + ], + "angle": 0, + "content": "\\[\n\\mathbf {z} _ {k} = \\boldsymbol {\\sigma} _ {k} \\odot \\boldsymbol {\\epsilon} + \\boldsymbol {\\mu} _ {k}, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.826, + 0.892, + 0.901 + ], + "angle": 0, + "content": "where \\(\\epsilon\\) follows the standard normal distribution \\(\\epsilon \\sim \\mathcal{N}(0,1)\\). To fuse with the features \\(\\mathcal{F}_i\\) from the CoSOD-TB, as shown in Figure 3, we expand \\(\\mathbf{z}_k\\) to the feature sequence which has the same size as \\(\\mathcal{F}_i\\), yielding the stochastic feature \\(\\mathcal{V}_i\\). The visualization effect of \\(\\mathcal{V}_i\\) is partly shown in Fig-" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.925, + 0.519, + 0.937 + ], + "angle": 0, + "content": "19642" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.08, + 0.089, + 0.473, + 0.199 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.148, + 0.21, + 0.398, + 0.224 + ], + "angle": 0, + "content": "Figure 4. Architecture of the CoSOD-TB." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.252, + 0.47, + 0.297 + ], + "angle": 0, + "content": "ure 5, as we can see, these features encode rich group consistency information that can well highlight the co-salient regions." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.309, + 0.276, + 0.325 + ], + "angle": 0, + "content": "3.3. CoSOD Transformer" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.333, + 0.469, + 0.394 + ], + "angle": 0, + "content": "After getting the stochastic features \\(\\nu_{i}\\) from the LVGB, we leverage these features to guide the learning of generating the general features \\(\\mathcal{F}_i\\) from the CoSOD-TB to focus on the co-salient regions (see the bottom row of Figure 5)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.395, + 0.47, + 0.835 + ], + "angle": 0, + "content": "Figure 4 shows the architecture of the proposed CoSOD-TB. Each input image \\( I_{i}^{n} \\in \\mathcal{I}_{ei} \\) is cropped into \\( d \\) patches with size \\( H / 4 \\times W / 4 \\), and then all the image patches in each group construct the token sequences \\( \\mathcal{X}_i^0 \\in \\mathbb{R}^{N \\times \\frac{H}{4} \\times \\frac{W}{4} \\times 3d} \\). Then, \\( \\mathcal{X}_i^0 \\) is sent into the backbone of the transformer with a T2T architecture [40], obtaining the token sequences \\( \\mathcal{X}_i^1 \\in \\mathbb{R}^{N \\times \\frac{H}{16} \\times \\frac{W}{16} \\times c} \\) which encode both local and global information. After this, we design the group token \\( \\mathcal{X}^G \\in \\mathbb{R}^{N \\times 1 \\times c} \\) to capture group-wise common information that is essential to extract co-objects, and the saliency token \\( \\mathcal{X}^S \\in \\mathbb{R}^{N \\times 1 \\times c} \\) to capture the specific information that encodes more co-object structure details. Afterwards, \\( \\mathcal{X}^G \\) and \\( \\mathcal{X}^1 \\) are concatenated in the channel dimension and fed into the Transformer Layer as the design in ViT [4] to get the new token sequence. We further use an multi-layer perceptron (MLP) layer to integrate global information into \\( \\mathcal{X}^G \\). We fuse the \\( \\mathcal{X}^S \\) to integrate saliency information into sequences and split the \\( \\mathcal{X}^G \\) and \\( \\mathcal{X}^S \\) to up-sample the token sequence and get the token sequence \\( \\mathcal{X}^2 \\). After undergoing some processing as shown in Figure 4, \\( \\mathcal{F}_i \\) is then concatenated with the stochastic features \\( \\mathcal{V}_i \\) followed by a decoder that includes two Transformer Layers and an MLP layer. During decoding, due to the use of stochastic features, the model is more possible to overcome the prejudices formed in the early training. As shown in the third row of Figure 5, the CoSOD-TB can integrate \\( \\mathcal{V}_i \\) and \\( \\mathcal{F}_i \\) well, helping the model focus on more reliable co-salient object regions. Finally, the output sequences are reshaped to produce the predicted co-saliency maps \\( \\hat{\\mathcal{O}} \\)." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.837, + 0.424, + 0.851 + ], + "angle": 0, + "content": "The loss function of the CoSOD-TB is defined as" + }, + { + "type": "equation", + "bbox": [ + 0.091, + 0.863, + 0.47, + 0.905 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {T R A N S} = \\frac {1}{2 N} \\sum_ {n = 1} ^ {N} \\sum_ {i = 1} ^ {2} \\ell_ {B C E} \\left(\\mathcal {Y} _ {e i} (n,:), \\hat {\\mathcal {O}} _ {i} (n,:)\\right), \\tag {5}\n\\]" + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.092, + 0.593, + 0.138 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.507, + 0.139, + 0.593, + 0.196 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.598, + 0.092, + 0.686, + 0.138 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.6, + 0.139, + 0.648, + 0.151 + ], + "angle": 0, + "content": "F" + }, + { + "type": "image", + "bbox": [ + 0.691, + 0.092, + 0.779, + 0.138 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.695, + 0.139, + 0.778, + 0.15 + ], + "angle": 0, + "content": "Watermelon_2" + }, + { + "type": "image", + "bbox": [ + 0.784, + 0.092, + 0.871, + 0.138 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.808, + 0.139, + 0.844, + 0.15 + ], + "angle": 0, + "content": "F" + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.151, + 0.592, + 0.196 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.536, + 0.196, + 0.559, + 0.208 + ], + "angle": 0, + "content": "V" + }, + { + "type": "image", + "bbox": [ + 0.598, + 0.151, + 0.685, + 0.196 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.629, + 0.196, + 0.648, + 0.208 + ], + "angle": 0, + "content": "V" + }, + { + "type": "image", + "bbox": [ + 0.691, + 0.151, + 0.778, + 0.196 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.7, + 0.196, + 0.749, + 0.208 + ], + "angle": 0, + "content": "V2-1" + }, + { + "type": "image", + "bbox": [ + 0.784, + 0.151, + 0.871, + 0.196 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.817, + 0.196, + 0.845, + 0.208 + ], + "angle": 0, + "content": "V2-2" + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.209, + 0.592, + 0.253 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.524, + 0.254, + 0.58, + 0.267 + ], + "angle": 0, + "content": "\\(\\mathcal{F}_i^1\\) & \\(\\mathcal{V}_i^{1 - 1}\\)" + }, + { + "type": "image", + "bbox": [ + 0.598, + 0.209, + 0.685, + 0.253 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.619, + 0.254, + 0.674, + 0.267 + ], + "angle": 0, + "content": "\\(\\mathcal{F}_i^1\\) & \\(\\nu_{i}^{1 - 2}\\)" + }, + { + "type": "image", + "bbox": [ + 0.691, + 0.209, + 0.778, + 0.253 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.71, + 0.254, + 0.767, + 0.267 + ], + "angle": 0, + "content": "\\(\\mathcal{F}_i^2\\& \\mathcal{V}_i^{2 - 1}\\)" + }, + { + "type": "image", + "bbox": [ + 0.784, + 0.209, + 0.871, + 0.253 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.798, + 0.254, + 0.856, + 0.267 + ], + "angle": 0, + "content": "\\(\\mathcal{F}_i^2\\) & \\(\\mathcal{V}_i^{2 - 2}\\)" + }, + { + "type": "image", + "bbox": [ + 0.88, + 0.092, + 0.909, + 0.256 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.283, + 0.893, + 0.352 + ], + "angle": 0, + "content": "Figure 5. Comparison of effects of the \"watermelon\" group before and after fusing stochastic features. The first row: visualizations of the images and features obtained by the CoSOD-TB; The second row: visualizations of the features generated by the LVGB; The third row: visualizations of the integrated features." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.376, + 0.891, + 0.404 + ], + "angle": 0, + "content": "where \\(\\ell_{BCE}\\) is a binary cross-entropy (BCE) loss [27] defined as" + }, + { + "type": "equation", + "bbox": [ + 0.515, + 0.414, + 0.89, + 0.453 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\ell_ {B C E} (\\mathcal {Y} (n,:), \\hat {\\mathcal {O}} (n,:)) = - (\\mathcal {Y} (n,:) ^ {\\top} \\log (\\hat {\\mathcal {O}} (n,:)) \\tag {6} \\\\ - \\left(1 - \\mathcal {Y} (n,:)\\right) ^ {\\top} \\log \\left(1 - \\tilde {\\mathcal {O}} (n,:)\\right)). \\\\ \\end{array}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.478, + 0.645, + 0.492 + ], + "angle": 0, + "content": "3.4. Loss Function" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.501, + 0.891, + 0.544 + ], + "angle": 0, + "content": "The LVG and the CoSOD-TB are jointly trained in an end-to-end manner by optimizing the following multi-task loss" + }, + { + "type": "equation", + "bbox": [ + 0.597, + 0.547, + 0.891, + 0.563 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\lambda_ {1} \\mathcal {L} _ {L V G} + \\lambda_ {2} \\mathcal {L} _ {T R A N S}, \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.568, + 0.886, + 0.584 + ], + "angle": 0, + "content": "where \\(\\lambda_{1},\\lambda_{2}\\) are the hyperparameters to balance each loss." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.596, + 0.705, + 0.613 + ], + "angle": 0, + "content": "4. Experimental Results" + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.621, + 0.716, + 0.637 + ], + "angle": 0, + "content": "4.1. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.644, + 0.892, + 0.795 + ], + "angle": 0, + "content": "Our model is implemented under the PyTorch1.9.0 framework [27]. Our computing platform's acceleration is provided by a GeForce GTX 2080Ti GPU. The transformer backbone is the pre-trained T2T-ViT \\(t - 14\\) [40] model since it has a similar computational complexity as CNNs-based ResNet50 [11] which is smaller than the VGG-16 [31] widely used in CoSOD. The training dataset are COCOSEG released by [33] and DUTS released by [34]. They totally include about 208,250 images from 369 categories, and the relevant binary ground-truth masks are provided." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.796, + 0.892, + 0.901 + ], + "angle": 0, + "content": "In each training episode, we randomly select two groups, which are processed by the proposed GEM with \\( k = 1 \\). Each group contains \\( N = 5 \\) images. The hyperparameters in (7) are set to \\( \\lambda_1 = 0.25 \\) and \\( \\lambda_2 = 1 \\). We resize the images to \\( 224 \\times 224 \\times 3 \\) and use them as input. We train the network over 60,000 steps totally and the training process takes about 11 hours. During the training process, we use" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.927, + 0.519, + 0.938 + ], + "angle": 0, + "content": "19643" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.12, + 0.088, + 0.85, + 0.304 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.32, + 0.895, + 0.348 + ], + "angle": 0, + "content": "Figure 6. The qualitative comparisons with the recent state-of-the-art methods, including DCFM [39], CADC [49], DeepACG [45] and GCoNet [9]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.358, + 0.47, + 0.419 + ], + "angle": 0, + "content": "the Adam algorithm [16] to optimize the whole network. The initial learning rate is set to \\( 10e - 4 \\) and decay to \\( 10e - 5 \\) at the 20,000-th epoch. The inference time is 31 fps in average, meeting the real-time application requirements." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.441, + 0.364, + 0.456 + ], + "angle": 0, + "content": "4.2. Datasets and Evaluation Metrics" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.469, + 0.47, + 0.621 + ], + "angle": 0, + "content": "The test sets include CoSal2015 [42], CoCA [50] and CoSOD3k [8]. Among them, CoSal2015 contains 2,015 images from 50 categories, and the objects therein have different sizes and shapes, which makes CoSal2015 challenging; CoCA is the newest dataset, including 1,297 images from 80 categories. These images have extremely different styles and very complicated backgrounds, making it a very difficult data set; CoSOD3k contains 3,316 images form 160 categories, which is by far the largest test set. It contains more disturbing targets, making it very challenging." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.624, + 0.47, + 0.716 + ], + "angle": 0, + "content": "Four evaluation metrics are used for comparison, including MAE [33], \\( E_{\\phi}^{max} \\) [6], \\( S_{\\alpha} \\) [5] and \\( F_{\\beta}^{max} \\) [1], which are to assess the average pixel-wise absolute difference, local and global similarity, structural similarity between the predictions and the ground truths, and the weighted harmonic mean of precision and recall, respectively." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.737, + 0.449, + 0.754 + ], + "angle": 0, + "content": "4.3. Comparisons with State-of-the-art Methods" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.47, + 0.903 + ], + "angle": 0, + "content": "On the basis of unified evaluation codes [7] for fair comparison, we compare our method with several state-of-the-art methods published in recent three years, including RCAN [20], CSMG [46], SSNM [44], GCAGC [47], GICD [50], ICNet [15], GCoNet [9], DeepACG [45], CoEG-Net [7], HrSSMN [48], CADC [49], DCFM [39]. Among them, CADC [49] and DCFM [39] are the latest cutting-edge methods. More experimental results can be found in the supplementary materials." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.358, + 0.895, + 0.844 + ], + "angle": 0, + "content": "Qualitative Results. Figure 6 shows some visual comparison results with four latest state-of-the-art methods, including DCFM, CADC, DeepACG and GCoNet. The selected four groups are very challenging and suffer from highly confusing interference co-objects, extremely complex backgrounds, and drastic scale changes. Specifically, in the group \"Baseball\", the co-salient object is very small and the interference objects have similar shapes. Because all other methods are trapped in overconfidence, obsessed with the wrong results, and unable to extricate themselves. They cannot accurately detect the tiny objects and the background is wrongly segmented. In the group \"Backpack\", benefiting from training with noisy samples, our model is more robust that can accurately locate co-salient objects and segment target details, while other methods DCFM, CADC and GCoNet misjudge the co-salient objects. Camouflage co-objects in complex backgrounds are generally acknowledged as difficult samples [18], and the co-salient items in the \"Gift box\" group are perfectly in place. Thanks to CoSOD-TB's ability to capture long-range dependencies, our model learns strong feature representation that is able to accurately segment the camouflage co-salient objects, while CADC and GCoNet cannot finely segment the objects, and the other methods even fail in detecting the targets. The group \"teddy bear\" is also very challenging, and the co-salient objects in it suffer from interference objects and backgrounds with similar colors. Besides, the objects also have very different sizes. In this case, our model still achieves promising results, demonstrating its strong robustness to a variety of challenging factors. However, other methods yield a large number of inaccurate segmentation regions." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.871, + 0.895, + 0.901 + ], + "angle": 0, + "content": "Quantitative Results. Figure 7 shows the PR and F-measure curves of our method and other state-of-the-art" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.927, + 0.52, + 0.938 + ], + "angle": 0, + "content": "19644" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.088, + 0.355, + 0.252 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.356, + 0.087, + 0.615, + 0.252 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.617, + 0.087, + 0.876, + 0.252 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.094, + 0.256, + 0.355, + 0.407 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.356, + 0.256, + 0.615, + 0.407 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.618, + 0.257, + 0.876, + 0.407 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.422, + 0.893, + 0.449 + ], + "angle": 0, + "content": "Figure 7. Comparisons with state-of-the-art methods since the year of 2019 in terms of PR and F-measure curves on three benchmark datasets." + }, + { + "type": "table_caption", + "bbox": [ + 0.195, + 0.459, + 0.773, + 0.472 + ], + "angle": 0, + "content": "Table 1. Statistic comparisons of our model with the other state-of-the-arts since the year of 2019." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.473, + 0.892, + 0.658 + ], + "angle": 0, + "content": "
MethodsCoSal2015CoSOD3kCoCA
MAE↓Sα↑Eφmax↑Fβmax↑MAE↓Sα↑Eφmax↑Fβmax↑MAE↓Sα↑Eφmax↑Fβmax↑
RCAN(IJCAI2019)0.1260.7790.8420.7640.1300.7440.8080.6880.1600.6160.7020.422
CSMG(CVPR2019)0.1300.7740.8180.7770.1570.7110.7230.6450.1240.6320.7340.503
SSNM(AAAI2020)0.1020.7880.8430.7940.1200.7260.7560.6750.1160.6280.7410.482
GCAGC(CVPR2020)0.0850.8170.8660.8130.1000.7850.8160.7400.1180.6690.7540.523
GICD(ECCV2020)0.0720.8420.8840.8340.0890.7940.8310.7430.1250.6580.7010.504
ICNet(NIPS2020)0.0580.8570.9000.8580.0890.7940.8450.7620.1470.6540.7050.514
CoEGNet(TPAMI2021)0.0770.8360.8820.8320.0920.7620.8250.7360.1060.6120.7170.493
GCoNet(CVPR2021)0.0690.8450.8870.8470.0710.8020.8600.7500.1050.6730.7600.524
DeepACG(CVPR2021)0.0660.8530.8930.8470.0790.8110.8590.7790.1040.6850.7590.564
CADC(ICCV2021)0.0640.8660.9060.8620.0960.8010.8400.7590.1320.6810.7440.548
HrSSMN(TMM2022)0.0620.8450.8950.8410.0870.7880.8420.7530.1060.6710.7390.532
DCFM(CVPR2022)0.0670.8380.8920.8560.0670.8100.8740.8050.0850.7100.7830.598
OURS0.0530.8850.9330.8820.0610.8530.9110.8290.0950.7260.8080.599
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.685, + 0.47, + 0.746 + ], + "angle": 0, + "content": "methods. As we can see, the PR curves produced by our method wrap around the curves generated by other methods. Also, all the F-measure curves of our method are above those generated by other methods." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Meanwhile, we list the statistic comparison results of all compared methods in Table 1. It is clear that our approach achieves the most competitive result compared to its counterparts. Specifically, on the CoSal2015 dataset, our method achieves the best scores of 0.053, 0.885, 0.933, and 0.882 in terms of all metrics, with a significant gain of \\(0.5\\%\\), \\(1.9\\%\\), \\(2.7\\%\\) and \\(2.0\\%\\), respectively, compared to the second-best preforming ICNet [15] and CADC [49]. On the most recent and difficult datasets, CoSOD3k and CoCA, our method also achieves the best performance. Especially" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.685, + 0.892, + 0.777 + ], + "angle": 0, + "content": "on CoSOD3k, our method reaches the best scores of 0.061, 0.853, 0.911 and 0.829 in terms of \\( MAE \\), \\( E_{\\phi}^{max} \\), \\( S_{\\alpha} \\) and \\( F_{\\beta}^{max} \\), respectively, with a great gain of \\( 0.6\\% \\), \\( 4.3\\% \\), \\( 3.7\\% \\) and \\( 2.4\\% \\) compared to the second-best performing DCFM [39]. The effectiveness of our method has been fully validated on the large-scale and challenging datasets." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.787, + 0.655, + 0.803 + ], + "angle": 0, + "content": "4.4. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.81, + 0.893, + 0.903 + ], + "angle": 0, + "content": "To verify the effectiveness of our designs, we conduct the ablation study of our method on three datasets. We have made a lot of changes to VST [23] and regard the modified version as the baseline. Table 2 lists the results, and we can observe that every key design in our model makes a significant contribution to the overall performance. Taking" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.927, + 0.519, + 0.938 + ], + "angle": 0, + "content": "19645" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.21, + 0.088, + 0.757, + 0.1 + ], + "angle": 0, + "content": "Table 2. Ablations of our method on the effectiveness of the GEM , LVGB, and CoSOD-TB." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.1, + 0.895, + 0.222 + ], + "angle": 0, + "content": "
StrategiesCoSal2015CoSOD3kCoCA
GEMLVGBCoSOD-TBMAE ↓\\( S_{\\alpha} \\uparrow \\)\\( E_{\\phi}^{max} \\uparrow \\)\\( F_{\\beta}^{max} \\uparrow \\)MAE ↓\\( S_{\\alpha} \\uparrow \\)\\( E_{\\phi}^{max} \\uparrow \\)\\( F_{\\beta}^{max} \\uparrow \\)MAE ↓\\( S_{\\alpha} \\uparrow \\)\\( E_{\\phi}^{max} \\uparrow \\)\\( F_{\\beta}^{max} \\uparrow \\)
0.0600.8540.8870.8600.0750.7870.8630.7780.1050.7100.7850.564
✓0.0610.8830.9280.8770.0650.8340.8800.8170.1090.7100.7880.571
✓0.0580.8770.9260.8720.0630.8400.8710.8050.1000.7190.7890.574
✓0.0530.8740.8930.8800.0620.8420.9000.8140.0980.7160.7980.583
✓✓0.0550.8690.9300.8760.0630.8470.9060.8190.1040.7240.8020.597
✓✓0.0600.8860.9250.8720.0690.8500.8950.8230.1000.7180.7920.587
✓✓0.0540.8800.9190.8780.0610.8490.8890.8290.0960.7200.8050.595
✓✓✓0.0530.8850.9330.8820.0610.8530.9110.8290.0950.7260.8080.599
" + }, + { + "type": "table_caption", + "bbox": [ + 0.078, + 0.23, + 0.468, + 0.256 + ], + "angle": 0, + "content": "Table 3. Complexity analyses. The number of inputs is set to 5 to maintain unity." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.257, + 0.468, + 0.327 + ], + "angle": 0, + "content": "
methodsFLOPs(G)↓param.(M)↓runtime(fps)↑Fβmax↑
GICD(ECCV20)364.7278.040.80.504
GCoNet(CVPR21)259.9142.0116.20.524
CADC(ICCV21)330.0392.818.00.548
DCFM(CVPR22)251.9142.384.40.598
OURS211.852.331.00.599
" + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.343, + 0.457, + 0.476 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.491, + 0.468, + 0.518 + ], + "angle": 0, + "content": "Figure 8. Comparison with the current most competitive method DCFM [39] in simulated real scenes." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.531, + 0.468, + 0.758 + ], + "angle": 0, + "content": "the results in CoSal2015 as examples, without GEM, the performance of our model gets worse in terms of all metrics especially for the \\( S_{\\alpha} \\) that drops from 0.885 to 0.869 by \\( 1.6\\% \\). With the GEM alone, the performance of our model improves greatly, and especially \\( S_{\\alpha} \\) and \\( E_{\\phi}^{max} \\) have a gain of \\( 2.9\\% \\) and \\( 4.1\\% \\), respectively. Also, LVGB achieves the same excellent effect as GEM, because the uncertainty features provided by the LVGB can help the model consider more possibilities to enhance the robustness and overcome the over-confidence to non-salient regions. The CoSOD-TB makes important contributions to fine segmentation of images. With the CoSOD-TB, the MAE gets \\( 0.7\\% \\) better from 0.060 to 0.053. The results on CoSOD3k and CoCA show the same trend, proving the effectiveness of each design of our method." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.77, + 0.277, + 0.788 + ], + "angle": 0, + "content": "4.5. Practical Application" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.795, + 0.468, + 0.901 + ], + "angle": 0, + "content": "For the practicability of the algorithm, we conduct complexity comparisons with the same settings as DCFM [39]. The results can be seen in Table 3. Our method achieves both smallest FLOPs, parameters and best performance in terms of \\(F_{\\beta}^{max}\\). Meanwhile, the runtime of our method is 31 fps, meeting the real-time requirements. This shows that our method is more suitable for deployment to various ap" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.234, + 0.618, + 0.248 + ], + "angle": 0, + "content": "plication devices." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.25, + 0.892, + 0.447 + ], + "angle": 0, + "content": "We further test the performance of the model in the real scenes. In the real-life scenarios, there may not always exist co-salient objects within the group of images [22]. Figure 8 shows the images in the simulated real scenes that may not always exist co-objects. In these cases, the model should segment the co-objects existing in most images, while outputting all-zero masks on images without co-objects. Our method is better than the most competitive method DCFM [39] in dealing with real situations. When dealing with the images without co-objects, our model produces all-zero maps as much as possible. It shows that our method is more practical than other methods. More experimental results can be found in the supplementary materials." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.466, + 0.619, + 0.481 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.494, + 0.892, + 0.901 + ], + "angle": 0, + "content": "In this paper, we have discovered that the paradigm of group consensus assumption has reduced the model's robustness and practical application value when confronted with irrelevant images in groups. A CoSOD model learning framework that is distinct from the classic CoSOD model learning framework has been developed. First, the group exchange-masking strategy has been devised, which is capable of automatically selecting the most informative noisy images from two groups and exchanging them in each other's groups to assist the model in learning more robust representations encoded with rich group consensus information. Second, the latent variable generator branch has been created to provide uncertainly-based global stochastic features that can regulate intermediate features from other branches to focus on co-objects. Third, the CoSOD transformer branch has been created to capture correlation-based global characteristics that carry information about group consistency. These branches' features are concatenated and put into a transformer-based decoder, yielding high-quality co-saliency maps. Extensive evaluations with and without irrelevant images have demonstrated the superiority of our method over a variety of state-of-the-art methods. In the future, we plan to build a new dataset for real scenarios that can provide a platform to evaluate model robustness and practical value in a more reliable way. Further more, we will consider the potential application of our method in defending adversarial attacks [35]." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.927, + 0.519, + 0.938 + ], + "angle": 0, + "content": "19646" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.175, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.116, + 0.47, + 0.185 + ], + "angle": 0, + "content": "[1] Radhakrishna Achanta, Sheila Hemami, Francisco Estrada, and Sabine Susstrunk. Frequency-tuned salient region detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1597-1604. IEEE, 2009. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.187, + 0.472, + 0.257 + ], + "angle": 0, + "content": "[2] Kai-Yueh Chang, Tyng-Luh Liu, and Shang-Hong Lai. From co-saliency to co-segmentation: An efficient and fully unsupervised energy minimization model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2129-2136. IEEE, 2011. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.259, + 0.471, + 0.328 + ], + "angle": 0, + "content": "[3] Paul Christiano, Zain Shah, Igor Mordatch, Jonas Schneider, Trevor Blackwell, Joshua Tobin, Pieter Abbeel, and Wojciech Zaremba. Transfer from simulation to real world through learning deep inverse dynamics model. arXiv preprint arXiv:1610.03518, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.33, + 0.47, + 0.413 + ], + "angle": 0, + "content": "[4] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.415, + 0.47, + 0.471 + ], + "angle": 0, + "content": "[5] Deng-Ping Fan, Ming-Ming Cheng, Yun Liu, Tao Li, and Ali Borji. Structure-measure: A new way to evaluate foreground maps. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4548-4557, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.473, + 0.47, + 0.528 + ], + "angle": 0, + "content": "[6] Deng-Ping Fan, Cheng Gong, Yang Cao, Bo Ren, MingMing Cheng, and Ali Borji. Enhanced-alignment measure for binary foreground map evaluation. arXiv preprint arXiv:1805.10421, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.53, + 0.47, + 0.599 + ], + "angle": 0, + "content": "[7] Deng-Ping Fan, Tengpeng Li, Zheng Lin, Ge-Peng Ji, Dingwen Zhang, Ming-Ming Cheng, Huazhu Fu, and Jianbing Shen. Re-thinking co-salient object detection. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2021. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.601, + 0.47, + 0.671 + ], + "angle": 0, + "content": "[8] Deng-Ping Fan, Zheng Lin, Ge-Peng Ji, Dingwen Zhang, Huazhu Fu, and Ming-Ming Cheng. Taking a deeper look at co-salient object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2919-2929, 2020. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.673, + 0.47, + 0.743 + ], + "angle": 0, + "content": "[9] Qi Fan, Deng-Ping Fan, Huazhu Fu, Chi-Keung Tang, Ling Shao, and Yu-Wing Tai. Group collaborative learning for cosalient object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12288–12298, 2021. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.745, + 0.47, + 0.786 + ], + "angle": 0, + "content": "[10] Huazhu Fu, Xiaochun Cao, and Zhuowen Tu. Cluster-based co-saliency detection. IEEE Transactions on Image Processing, 22(10):3766-3778, 2013. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.788, + 0.47, + 0.844 + ], + "angle": 0, + "content": "[11] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 770-778, 2016. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.845, + 0.47, + 0.902 + ], + "angle": 0, + "content": "[12] Kuang-Jui Hsu, Yen-Yu Lin, and Yung-Yu Chuang. Coattention cnns for unsupervised object co-segmentation. In Proceedings of the International Joint Conference on Artificial Intelligence, pages 748-756, 2018. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.116, + 0.472, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.161 + ], + "angle": 0, + "content": "[13] David E Jacobs, Dan B Goldman, and Eli Shechtman. Cosaliency: Where people look when comparing images. In Proceedings of the 23nd Annual ACM Symposium on User Interface Software and Technology, pages 219-228, 2010. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.163, + 0.894, + 0.218 + ], + "angle": 0, + "content": "[14] Koteswar Rao Jerripothula, Jianfei Cai, and Junsong Yuan. Cats: Co-saliency activated tracklet selection for video colocalization. In Proceedings of the European Conference on Computer Vision, pages 187-202. Springer, 2016. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.219, + 0.894, + 0.273 + ], + "angle": 0, + "content": "[15] Wen-Da Jin, Jun Xu, Ming-Ming Cheng, Yi Zhang, and Wei Guo. Icnet: Intra-saliency correlation network for co-saliency detection. Advances in Neural Information Processing Systems, 33:18749-18759, 2020. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.274, + 0.894, + 0.314 + ], + "angle": 0, + "content": "[16] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.315, + 0.894, + 0.398 + ], + "angle": 0, + "content": "[17] Simon Kohl, Bernardino Romera-Paredes, Clemens Meyer, Jeffrey De Fauw, Joseph R Ledsam, Klaus Maier-Hein, SM Eslami, Danilo Jimenez Rezende, and Olaf Ronneberger. A probabilistic u-net for segmentation of ambiguous images. Advances in Neural Information Processing Systems, 31, 2018. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.4, + 0.894, + 0.469 + ], + "angle": 0, + "content": "[18] Aixuan Li, Jing Zhang, Yunqiu Lv, Bowen Liu, Tong Zhang, and Yuchao Dai. Uncertainty-aware joint salient object and camouflaged object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10071-10081, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.47, + 0.894, + 0.525 + ], + "angle": 0, + "content": "[19] Bo Li, Zhengxing Sun, and Yuqi Guo. Supervae: Superpixelwise variational autoencoder for salient object detection. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 33, pages 8569-8576, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.526, + 0.894, + 0.593 + ], + "angle": 0, + "content": "[20] Bo Li, Zhengxing Sun, Lv Tang, Yunhan Sun, and Jinlong Shi. Detecting robust co-saliency with recurrent coattention neural network. In Proceedings of the International Joint Conference on Artificial Intelligence, volume 2, page 6, 2019. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.595, + 0.894, + 0.637 + ], + "angle": 0, + "content": "[21] Yijun Li, Keren Fu, Zhi Liu, and Jie Yang. Efficient saliency-model-guided visual co-saliency detection. IEEE Signal Processing Letters, 22(5):588-592, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.637, + 0.894, + 0.677 + ], + "angle": 0, + "content": "[22] Jiawei Liu, Jing Zhang, Kaihao Zhang, and Nick Barnes. Generalised co-salient object detection. arXiv preprint arXiv:2208.09668, 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.679, + 0.894, + 0.733 + ], + "angle": 0, + "content": "[23] Nian Liu, Ni Zhang, Kaiyuan Wan, Ling Shao, and Junwei Han. Visual saliency transformer. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.734, + 0.894, + 0.789 + ], + "angle": 0, + "content": "[24] Xuanqing Liu, Minhao Cheng, Huan Zhang, and Cho-Jui Hsieh. Towards robust neural networks via random self-ensemble. In Proceedings of the European Conference on Computer Vision, pages 369–385, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.79, + 0.894, + 0.844 + ], + "angle": 0, + "content": "[25] Xuanqing Liu, Yao Li, Chongruo Wu, and Cho-Jui Hsieh. Adv-bnn: Improved adversarial defense through robust bayesian neural network. arXiv preprint arXiv:1810.01279, 2018.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.845, + 0.894, + 0.902 + ], + "angle": 0, + "content": "[26] Aleksander Madry, Aleksandar Makelov, Ludwig Schmidt, Dimitris Tsipras, and Adrian Vladu. Towards deep learning models resistant to adversarial attacks. In International Conference on Learning Representations, 2018. 4" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.927, + 0.52, + 0.938 + ], + "angle": 0, + "content": "19647" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.174 + ], + "angle": 0, + "content": "[27] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. Advances in Neural Information Processing Systems, 32:8026-8037, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.176, + 0.47, + 0.232 + ], + "angle": 0, + "content": "[28] Lerrel Pinto, James Davidson, Rahul Sukthankar, and Abhi nav Gupta. Robust adversarial reinforcement learning. In International Conference on Machine Learning, pages 2817-2826. PMLR, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.233, + 0.47, + 0.287 + ], + "angle": 0, + "content": "[29] Aravind Rajeswaran, Sarvjeet Ghotra, Balaraman Ravindran, and Sergey Levine. Epopt: Learning robust neural network policies using model ensembles. arXiv preprint arXiv:1610.01283, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.288, + 0.47, + 0.343 + ], + "angle": 0, + "content": "[30] Andrei A Rusu, Matej Večerík, Thomas Rothörl, Nicolas Heess, Razvan Pascanu, and Raia Hadsell. Sim-to-real robot learning from pixels with progressive nets. In Conference on robot learning, pages 262-270. PMLR, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.344, + 0.47, + 0.385 + ], + "angle": 0, + "content": "[31] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556, 2014. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.386, + 0.47, + 0.44 + ], + "angle": 0, + "content": "[32] Kihyuk Sohn, Honglak Lee, and Xinchen Yan. Learning structured output representation using deep conditional generative models. Advances in Neural Information Processing Systems, 28, 2015. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.442, + 0.47, + 0.496 + ], + "angle": 0, + "content": "[33] Chong Wang, Zheng-Jun Zha, Dong Liu, and Hongtao Xie. Robust deep co-saliency detection with group semantic. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 33, pages 8917-8924, 2019. 1, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.497, + 0.47, + 0.566 + ], + "angle": 0, + "content": "[34] Lijun Wang, Huchuan Lu, Yifan Wang, Mengyang Feng, Dong Wang, Baocai Yin, and Xiang Ruan. Learning to detect salient objects with image-level supervision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 136-145, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.567, + 0.47, + 0.622 + ], + "angle": 0, + "content": "[35] Yisen Wang, Difan Zou, Jinfeng Yi, James Bailey, Xingjun Ma, and Quanquan Gu. Improving adversarial robustness requires revisiting misclassified examples. In International Conference on Learning Representations, 2019. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.623, + 0.47, + 0.664 + ], + "angle": 0, + "content": "[36] Lina Wei, Shanshan Zhao, Omar El Farouk Bourahla, Xi Li, and Fei Wu. Group-wise deep co-saliency detection. arXiv preprint arXiv:1707.07381, 2017. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.665, + 0.47, + 0.706 + ], + "angle": 0, + "content": "[37] Cihang Xie, Jianyu Wang, Zhishuai Zhang, Zhou Ren, and Alan Yuille. Mitigating adversarial effects through randomization. arXiv preprint arXiv:1711.01991, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.707, + 0.47, + 0.761 + ], + "angle": 0, + "content": "[38] Xi Yang, Shaoyi Li, Jun Ma, Jun-yan Yang, and Jie Yan. Co-saliency-regularized correlation filter for object tracking. Signal Processing: Image Communication, 103:116655, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.762, + 0.47, + 0.831 + ], + "angle": 0, + "content": "[39] Siyue Yu, Jimin Xiao, Bingfeng Zhang, and Eng Gee Lim. Democracy does matter: Comprehensive feature mining for co-salient object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 979–988, 2022. 1, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.832, + 0.47, + 0.901 + ], + "angle": 0, + "content": "[40] Li Yuan, Yunpeng Chen, Tao Wang, Weihao Yu, Yujun Shi, Zi-Hang Jiang, Francis EH Tay, Jiashi Feng, and Shuicheng Yan. Tokens-to-token vit: Training vision transformers from scratch onImagenet. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2021. 5" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.134 + ], + "angle": 0, + "content": "[41] Chen Zhang, Riccardo Barbano, and Bangti Jin. Conditional variational autoencoder for learned image reconstruction. Computation, 9(11):114, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.136, + 0.892, + 0.19 + ], + "angle": 0, + "content": "[42] Dingwen Zhang, Junwei Han, Chao Li, and Jingdong Wang. Co-saliency detection via looking deep and wide. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2994-3002, 2015. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.192, + 0.892, + 0.259 + ], + "angle": 0, + "content": "[43] Jing Zhang, Deng-Ping Fan, Yuchao Dai, Saeed Anwar, Fatemeh Saleh, Sadegh Aliakbarian, and Nick Barnes. Uncertainty inspired rgb-d saliency detection. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2021. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.262, + 0.892, + 0.318 + ], + "angle": 0, + "content": "[44] Kaihua Zhang, Jin Chen, Bo Liu, and Qingshan Liu. Deep object co-segmentation via spatial-semantic network modulation. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 34, pages 12813–12820, 2020. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.32, + 0.892, + 0.388 + ], + "angle": 0, + "content": "[45] Kaihua Zhang, Mingliang Dong, Bo Liu, Xiao-Tong Yuan, and Qingshan Liu. Deepacg: Co-saliency detection via semantic-aware contrast gromov-wasserstein distance. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13703-13712, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.39, + 0.892, + 0.459 + ], + "angle": 0, + "content": "[46] Kaihua Zhang, Tengpeng Li, Bo Liu, and Qingshan Liu. Cosaliency detection via mask-guided fully convolutional networks with multi-scale label smoothing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3095-3104, 2019. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.461, + 0.892, + 0.543 + ], + "angle": 0, + "content": "[47] Kaihua Zhang, Tengpeng Li, Shiwen Shen, Bo Liu, Jin Chen, and Qingshan Liu. Adaptive graph convolutional network with attention graph clustering for co-saliency detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9050-9059, 2020. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.545, + 0.892, + 0.599 + ], + "angle": 0, + "content": "[48] Kaihua Zhang, Yang Wu, Mingliang Dong, Bo Liu, Dong Liu, and Qingshan Liu. Deep object co-segmentation and c-osaliency detection via high-order spatial-semantic network modulation. IEEE Transactions on Multimedia, 2022. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.601, + 0.892, + 0.669 + ], + "angle": 0, + "content": "[49] Ni Zhang, Junwei Han, Nian Liu, and Ling Shao. Summarize and search: Learning consensus-aware dynamic convolution for co-saliency detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4167-4176, 2021. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.671, + 0.894, + 0.727 + ], + "angle": 0, + "content": "[50] Zhao Zhang, Wenda Jin, Jun Xu, and Ming-Ming Cheng. Gradient-induced co-saliency detection. In Proceedings of the European Conference on Computer Vision, pages 455-472. Springer, 2020. 1, 2, 6" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.727 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.927, + 0.52, + 0.938 + ], + "angle": 0, + "content": "19648" + } + ] +] \ No newline at end of file diff --git a/2023/Co-Salient Object Detection With Uncertainty-Aware Group Exchange-Masking/9cd15a5b-7ef1-4ba9-a24c-48575755f531_origin.pdf b/2023/Co-Salient Object Detection With Uncertainty-Aware Group Exchange-Masking/9cd15a5b-7ef1-4ba9-a24c-48575755f531_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..8d23ded354b610d92629905477ddfbc5e57d451b --- /dev/null +++ b/2023/Co-Salient Object Detection With Uncertainty-Aware Group Exchange-Masking/9cd15a5b-7ef1-4ba9-a24c-48575755f531_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f4027daf67917af9bf4024325bda76a557fbd57cbfa2cde8a89ad5268a72bc29 +size 3589680 diff --git a/2023/Co-Salient Object Detection With Uncertainty-Aware Group Exchange-Masking/full.md b/2023/Co-Salient Object Detection With Uncertainty-Aware Group Exchange-Masking/full.md new file mode 100644 index 0000000000000000000000000000000000000000..c00629a465c52a2066b3e8395ecb7cd0fcd49b50 --- /dev/null +++ b/2023/Co-Salient Object Detection With Uncertainty-Aware Group Exchange-Masking/full.md @@ -0,0 +1,335 @@ +# Co-Salient Object Detection with Uncertainty-aware Group Exchange-Masking + +Yang Wu $^{1}$ Huihui Song $^{1*}$ Bo Liu $^{2}$ Kaihua Zhang $^{1}$ Dong Liu $^{3}$ + +$^{1}$ B-DAT and CICAET, Nanjing University of Information Science and Technology, Nanjing, China + +$^{2}$ Walmart Global Tech, Sunnyvale, CA, 94086, USA + +3Netflix Inc, Los Gatos, CA, 95032, USA + +songhuihui@nuist.edu.cn + +# Abstract + +The traditional definition of co-salient object detection (CoSOD) task is to segment the common salient objects in a group of relevant images. Existing CoSOD models by default adopt the group consensus assumption. This brings about model robustness defect under the condition of irrelevant images in the testing image group, which hinders the use of CoSOD models in real-world applications. To address this issue, this paper presents a group exchange-masking (GEM) strategy for robust CoSOD model learning. With two groups of image containing different types of salient object as input, the GEM first selects a set of images from each group by the proposed learning based strategy, then these images are exchanged. The proposed feature extraction module considers both the uncertainty caused by the irrelevant images and group consensus in the remaining relevant images. We design a latent variable generator branch which is made of conditional variational autoencoder to generate uncertainly-based global stochastic features. A CoSOD transformer branch is devised to capture the correlation-based local features that contain the group consistency information. At last, the output of two branches are concatenated and fed into a transformer-based decoder, producing robust co-saliency prediction. Extensive evaluations on co-saliency detection with and without irrelevant images demonstrate the superiority of our method over a variety of state-of-the-art methods. + +# 1. Introduction + +Co-salient object detection (CoSOD) is to segment the common salient objects in a group of relevant images. By detecting the co-salient object in a group of images, the images' background and redundant content are re + +![](images/94854fdc8c47034ac3e62cc651aca8f0721e02daba5b07d7d0b6cc6c9a217f61.jpg) +Figure 1. (a) When there exists an irrelevant image in the test group, the state-of-the-art CoSOD models such as DCFM [39] and GCoNet [9] tend to generate false positive predictions for the noisy image, yet our method can achieve an accurate prediction due to the use of group exchange-masking strategy in (b). + +![](images/a70bec057487f1f85338cccfaa7d71bf4ebaa5ea13ba93fe6c3e2531cd1797c5.jpg) + +moved, which helps the downstream tasks such as object tracking [38], co-segmentation [48] and video colocalization [14], to name a few. + +Group consensus assumption is widely used by existing CoSOD models, i.e., these models presume that all images in the same group contain the common salient targets. The widely-used CoSOD benchmark datasets, such as COCO-SEG [33], CoCA [50], and CoSOD3k [8] organize the training and testing images that contain the same object as a group. Many existing CoSOD models consider the group consensus characteristic in modeling. For example, early works such as [2, 10, 13] extract hand-crafted features for inter-image co-object correspondence discovery. The deep learning models proposed in [15, 20, 36, 39, 44, 46, 48] use one group of relevant images as training data input for consensus representation learning. Among them, a variety of novel model design techniques have been developed to make full use of the group consistency characteristic, such as the low-rank feature learning [46], co-attention model [20] and intra-saliency correlation learning [15]. The issues of the group consensus assumption are partially studied in recent literature [9], which further models the inter + +group separability for more discriminative feature learning by a group collaborating module. + +In this paper, we find that the group consensus assumption also restricts the CoSOD model's robustness against the images without common object. As illustrated by Figure 1(a), state-of-the-art CoSOD models tend to output false positive predictions for the irrelevant image. This issue hinders the use of CoSOD models in real-world applications where the testing inputs are likely to contain irrelevant images. To enhance the model's robustness, we propose a learning framework called group exchange-masking (GEM). The GEM is illustrated by Figure 1(b). Given two image groups that contain different types of co-salient objects, we exchange several images between one group and the other. Those exchanged images are called noisy images. The number of noisy images is chosen to be less than the number of remaining relevant images in the group, so that the co-salient object in the noisy images forms a negative object but not the dominant co-salient object. The "masking" strategy refers to the label regeneration of the noisy images. Because there is no co-salient object, in the regenerated label, the original ground-truth object is masked. The learning objective is to correctly predict both the co-salient objects in the original relevant images and the added noisy images. + +Adding noisy images to the training image group brings about uncertainty to the CoSOD model learning since there is some probability of no expected common object in each image. We design a dual-path image feature extraction module to model the group uncertainty in addition to the group consensus property. Specifically, we design a latent variable generator branch (LVGB) to extract the uncertainty-based global image features. The LVGB module is motivated by the conditional variational autoencoder (CVAE) [32] that is widely used to address the uncertainty in vision applications including image background modeling [19], RGB-D saliency detection [43] and image reconstruction [41]. In parallel with LVGB, we feed the image group into a CoSOD Transformer Branch (CoSOD-TB). The CoSOD-TB partitions each image group into local patches, and the attention mechanism in the transformer enables this branch to model patch-wise correlation-based local features. As a result, the group consistency information can be captured by this branch. The outputs of the two branches are concatenated and fed into a transformer-based decoder for co-saliency prediction. The proposed model has the following technical contributions. + +- A robust CoSOD model learning mechanism, called group exchange-masking is proposed. By exchanging images between two groups, we augment the training data containing irrelevant images as noise to enhance model's robustness. This is different from the traditional CoSOD model learning frameworks that use + +groups of relevant images as training data. + +- We propose a dual-path feature extraction module composed of the LVGB and the CoSOD-TB. The LVGB is designed to model the uncertainty of cosalient object existence. The CoSOD-TB is for the consensus feature extraction of the salient object in the relevant images. +- Extensive evaluations on three benchmark datasets, including CoSal2015 [42], CoCA [15], and CoSOD3k [7] show that the superiority of the proposed model to the state-of-the-art methods in terms of all evaluation metrics. Besides, the proposed model demonstrates good robustness for dealing with noisy data without co-salient objects. + +# 2. Related Work + +# 2.1. Co-salient Object Detection + +In the past, CoSOD methods used to extract handcrafted features such as Gabor and SIFT features from images and then detect co-saliency by utilizing the consistency of low-level features between the images being tested [2]. A series of studies attempt to capture the intra-image constraints by employing a manifold ranking scheme to produce saliency maps [21], or using a global association constraint with clustering [10], or translational alignment [13]. More recently, there is a surge of deep learning-based CoSOD models that learn feature representation and saliency predictor in an end-to-end manner [9, 12, 36, 46, 47, 50]. Wei et al. [36] designs a collaborative learning framework for CoSOD that discovers the collaborative and interactive relationships between intra-image and single-image feature representations in a group-wise manner. Hsu et al. [12] present an unsupervised CNN-based model for CoSOD. In [46], a hierarchical framework is proposed for CoSOD where the initial CoSOD results generated by the CNN model is refined by label smoothing. Zhang et al. [50] propose a gradient-induced model for CoSOD that utilizes the image gradient information to induce more attentions to the discriminative co-salient feature learning. In [47], a deep graph neural network model is proposed to characterize the intra-image and inter-image region correspondence for CoSOD. In [9], a group collaborative learning strategy is proposed to explore inter-group relations for discriminating feature learning. + +# 2.2. Robust Model and Feature Learning + +In past years, with the development of deep learning, robust model and feature learning have received more and more attention as a possible solution to overcoming the bottleneck [28]. Various methods have been proposed to improve the robustness of the model. In [3] and [30], ways to + +![](images/f3256742c23bdfa633fb2e92436604e49167e6b70d2961de41df887a448d6a2d.jpg) +Figure 2. Pipeline of the proposed method. First, two groups of images $\{\mathcal{I}_1,\mathcal{I}_2\}$ with their masks $\{\mathcal{V}_1,\mathcal{V}_2\}$ are processed by group exchange-masking, yielding the exchanged images $\{\mathcal{I}_{e1},\mathcal{I}_{e2}\}$ and their corresponding masked labels $\{\mathcal{V}_{e1},\mathcal{V}_{e2}\}$ . Then, $\{\mathcal{I}_{e1},\mathcal{I}_{e2}\}$ and $\{\mathcal{V}_{e1},\mathcal{V}_{e2}\}$ are passed through the latent variable generator branch, extracting the uncertainty-based global image features $\{\mathcal{V}_1,\mathcal{V}_2\}$ to eliminate the bias caused by noisy images. Meanwhile, $\{\mathcal{I}_{e1},\mathcal{I}_{e2}\}$ are passed through the weight-shared CoSOD transformer branch, producing two groups of feature sequences $\{\mathcal{F}_1,\mathcal{F}_2\}$ with group consistency information and long-range independent information. Finally, $\mathcal{F}$ and $\mathcal{V}$ are concatenated and fed into the transformer decoder, yielding the predicted co-saliency maps $\hat{\mathcal{O}}$ . + +transfer the policy learned in a simulator to the real world are proposed to make features more robust. Rajeswaran et al. [29] learn a robust policy by sampling the worst case trajectories from a class of parametrized models, to learn a robust policy. Xie et al. [37] enhance the robustness of the model through randomly resizing and padding the training images. Several works leverage randomness to improve the robustness of models. Liu et al. [24] propose a noise layer that introduces randomness to both the input and the hidden layer output. Liu et al. [25] introduce a new min-max formulation that combines adversarial training with Bayesian Neural Networks, which achieves promising results. + +# 3. Proposed Method + +Figure 2 illustrates the pipeline of our framework. Given two groups of $2N$ images $\mathcal{I}_i = \{I_i^n\in \mathbb{R}^{H\times W\times 3}\}_{n = 1}^N,i = 1,2$ with height $H$ and width $W$ and manually-labeled binary masks $\mathcal{V}_i = \{\mathbf{Y}_i^n\in \mathbb{R}^{H\times W}\}_{n = 1}^N$ with different categories as input, we employ the GEM model to reorganize them, yielding $\mathcal{I}_{ei}$ and $\mathcal{V}_{ei}$ . Each pair of $\{\mathcal{I}_{ei},\mathcal{V}_{ei}\}$ are fed to the LVGB. In LVGB, we draw $\mathbf{z}_i\in \mathbb{R}^K$ from the Gaussian distribution $P_{\theta}(\mathbf{z}_i|\mathcal{I}_{ei})$ with the learning parameters $\pmb{\theta}$ and the posterior of $\mathbf{z}_i$ is formulated as $Q_{\phi}(\mathbf{z}_i|\mathcal{I}_{ei},\mathcal{V}_{ei})$ with the learning parameters $\phi$ . We use the KL-Divergence metric to reduce the distance between $P_{\theta}(\mathbf{z}_i|\mathcal{I}_{ei})$ and $Q_{\phi}(\mathbf{z}_i|\mathcal{I}_{ei},\mathcal{V}_{ei})$ . We further re-parameterize and expand $\mathbf{z}_i$ to yield a latent sequence $\mathcal{V}_i$ which has the same dimension as the feature sequences produced by trans + +# Algorithm 1 Group exchange-masking + +Input: $\mathcal{I}_1 = \{I_1^n\}_{n=1}^N$ , $\mathcal{Y}_1 = \{\mathbf{Y}_1^n\}_{n=1}^N$ , + +$$ +\mathcal {I} _ {2} = \left\{I _ {2} ^ {n} \right\} _ {n = 1} ^ {N}, \mathcal {Y} _ {2} = \left\{\mathbf {Y} _ {2} ^ {n} \right\} _ {n = 1} ^ {N}. +$$ + +Output: $\mathcal{I}_{e1},\mathcal{Y}_{e1},\mathcal{I}_{e2},\mathcal{Y}_{e2}$ + +1: $\mathcal{I}_{e1},\mathcal{Y}_{e1},\mathcal{I}_{e2},\mathcal{Y}_{e2}\gets$ Rearranging $\mathcal{I}_1,\mathcal{Y}_1,\mathcal{I}_2,\mathcal{Y}_2$ via solving (1). + +2: for $n = 1, 2, \ldots, k$ do +3: $I_{e1}^{n} \gets I_{2}^{n}, I_{e2}^{n} \gets I_{1}^{n}$ , +4: $\mathbf{Y}_{e1}^{n}\gets \mathbf{0}\in \mathbb{R}^{H\times W\times 1},$ +5: $\mathbf{Y}_{e2}^{\vec{n}}\gets \mathbf{0}\in \mathbb{R}^{H\times W\times 1}$ +6: end for + +formers. Meanwhile, the image groups are passed through the CoSOD-TB with shared weights, producing the token sequences $\mathcal{F}_i$ which capture group consensus and long-range dependency information. At last, $\mathcal{V}_i$ and $\mathcal{F}_i$ are concatenated in channel dimension and passed through several subsequent processes as shown in Figure 2 followed by an up-sampling layer, yielding the corresponding predicted cosaliency maps $\hat{\mathcal{O}} = \{\hat{\mathbf{O}}^n\in \mathbb{R}^{H\times W}\}_{n = 1}^{2N}$ . + +# 3.1. Group Exchange-Masking + +The widely-used group consensus assumption limits the robustness of the CoSOD model, especially when there are test images without common objects in the group. To this end, we reorganize the two input groups with the GEM tailored to CoSOD. The GEM is summarised in Algorithm 1: + +![](images/7c477db33a779d751b18b96233bdfb2f1ffbecf1bd6510b8e32065d3b31e6119.jpg) +Figure 3. Architecture of the LVGB. + +taking $\mathcal{I}_i$ and $\mathcal{Y}_i$ as input, the goal of GEM is to obtain $\mathcal{I}_{ei}$ and $\mathcal{Y}_{ei}$ which contain noisy images in each group and corresponding full masked ground truths. Specifically, we rearrange the two groups of images and their binary ground-truth masks in a decreasing order by solving (1). After this, we sample the top- $k$ images of each group as noisy images and exchange them in each other's group, ensuring the maximum destructiveness of noisy images and the effectiveness of training. + +This training process is essentially a min-max optimization process [26] as + +$$ +\min _ {\varphi} \sum_ {n = 1} ^ {k} \sum_ {i = 1} ^ {2} \max _ {\mathcal {I} _ {i} (n,:)} \mathcal {L} \left(f _ {\varphi} \left(\mathcal {I} _ {i} (n,:)\right), \hat {\mathcal {O}} _ {i} (n,:)\right), \tag {1} +$$ + +where the loss $\mathcal{L}$ is defined by (7), and $f$ denotes our whole model with learnable parameters $\varphi$ . Solving (1) finds out the most noisy images which can maximize the training loss, and our goal is to minimize the loss function w.r.t the intra-group noisy images to improve the model's robustness. At last, we mask the corresponding labels of the noisy images with the all-zeros maps. + +# 3.2. Latent Variable Generator + +By training the model with noisy images, the GEM is able to increase the model's robustness which can identify whether there exist co-objects or not. However, the model will focus on non-co-salient regions as a result of the uncertainty introduced by adding noisy images to the training image groups. This means that the model will become overconfident in the background regions in the early stage of training, leading to inaccurate prediction. + +To address this issue, we further propose the LVGB that generates a low-dimensional latent space to represent the most common patterns in image groups [17]. The latent variables sampled from the feature space characterize uncertainties of group consensus and can be used to modulate the intermediate features from other branches to highlight the co-objects. + +Figure 3 shows the details of our LVGB. The generator takes $\mathcal{I}_{ei}$ and latent variable $\mathbf{z}_i$ as input, and uses an encoder + +to produce stochastic prediction $P_{\omega}(\mathcal{Y}_{ei}|\mathcal{I}_{ei},\mathbf{z}_i)$ with learnable parameters $\omega$ , where $\mathcal{I}_{ei}$ is conditional variable, and $\mathcal{Y}_{ei}$ is output variable. $\mathbf{z}_i$ is drawn from the Gaussian distribution $P_{\theta}(\mathbf{z}_i|\mathcal{I}_{ei})$ , and the posterior of $\mathbf{z}_i$ is formulated as $Q_{\phi}(\mathbf{z}_i|\mathcal{I}_{ei},\mathcal{Y}_{ei})$ via an encoder that maps the input variable to the latent space. The loss of the LVG is defined as [32,43] + +$$ +\begin{array}{l} \mathcal {L} _ {L V G} = \sum_ {i = 1} ^ {2} E _ {\mathbf {z} \sim Q _ {\phi} (\mathbf {z} _ {i} | \mathcal {I} _ {e i}, \mathcal {Y} _ {e i})} [ - \log P _ {\boldsymbol {\omega}} (\mathcal {Y} _ {e i} | \mathcal {I} _ {e i}, \mathbf {z} _ {i}) ] \\ + \mathrm {K L} \left(Q _ {\phi} \left(\mathbf {z} _ {i} \mid \mathcal {I} _ {e i}, \mathcal {Y} _ {e i}\right) \mid \mid P _ {\boldsymbol {\theta}} \left(\mathbf {z} _ {i} \mid \mathcal {I} _ {e i}\right)\right), \tag {2} \\ \end{array} +$$ + +where the Kullback-Leibler Divergence metric KL is defined as + +$$ +\operatorname {K L} (Q | | P) = \sum Q (x) \log \frac {Q (x)}{P (x)}. \tag {3} +$$ + +In (2), the prior net $P_{\pmb{\theta}}(\mathbf{z}_i|\mathcal{I}_{ei})$ is defined as a Gaussian distribution that maps the image group $\mathcal{I}_{ei}$ to a low-dimensional latent space that encodes the most common patterns in each group. We randomly sample prior latent variable $\mathbf{z}_i\sim \mathcal{N}(\pmb{\mu}_{1i},\mathrm{diag}(\pmb{\sigma}_{1i}^2))$ , where $\pmb{\mu}_{1i},\pmb{\sigma}_{1i}^{2}\in \mathbb{R}^{K}$ represent the mean and standard deviation vectors and $\pmb{\theta}$ is the learnable parameter of the mapping function which consists five conv layers [43] as shown in Figure 3. In the posterior net, we use the similar encoder as the prior net to model $Q_{\phi}(\mathbf{z}_i|\mathcal{I}_{ei},\mathcal{Y}_{ei})$ that maps the concatenation of $\mathcal{I}_{ei}$ and $\mathcal{Y}_{ei}$ to the posterior latent variable $\mathbf{z}_i\sim \mathcal{N}(\pmb{\mu}_{2i},\mathrm{diag}(\pmb{\sigma}_{2i}^2))$ where $\pmb{\mu}_{2i},\pmb{\sigma}_{2i}^{2}\in \mathbb{R}^{K}$ . The KL in (2) is used to measure the difference between probability distributions prior $P_{\pmb{\theta}}(\mathbf{z}_i|\mathcal{I}_{ei})$ and posterior $Q_{\phi}(\mathbf{z}_i|\mathcal{I}_{ei},\mathcal{Y}_{ei})$ . It represents the information loss that occurs when using $Q_{\phi}(\mathbf{z}_i|\mathcal{I}_{ei},\mathcal{Y}_{ei})$ to approximate $P_{\pmb{\theta}}(\mathbf{z}_i|\mathcal{I}_{ei})$ , and a smaller value indicates greater similarity between the two distributions. + +Each position in latent space symbolizes potential labeling changes or other potential factors that could lead to a variety of co-saliency predictions [17, 43]. The proposed GEM can meet the needs for diverse ground truths as we learn to select the most difficult samples to mask their ground truths to all-zeros maps. We hope diverse annotations in the posterior net $Q_{\phi}(\mathbf{z}_i|\mathcal{I}_{ei},\mathcal{Y}_{ei})$ during training can compel the prior net $P_{\theta}(\mathbf{z}_i|\mathcal{I}_{ei})$ to encode labeling variants of the supplied inputs $\mathcal{I}_{ei}$ . The statistics $\mathbf{z}_i$ is further processed by feature expanding [43]. Given a pair of $(\boldsymbol{\mu}_k,\boldsymbol{\sigma}_k)$ in each position of $K$ dimensional vector, we parameterize them to obtain the latent vector + +$$ +\mathbf {z} _ {k} = \boldsymbol {\sigma} _ {k} \odot \boldsymbol {\epsilon} + \boldsymbol {\mu} _ {k}, \tag {4} +$$ + +where $\epsilon$ follows the standard normal distribution $\epsilon \sim \mathcal{N}(0,1)$ . To fuse with the features $\mathcal{F}_i$ from the CoSOD-TB, as shown in Figure 3, we expand $\mathbf{z}_k$ to the feature sequence which has the same size as $\mathcal{F}_i$ , yielding the stochastic feature $\mathcal{V}_i$ . The visualization effect of $\mathcal{V}_i$ is partly shown in Fig- + +![](images/669de5bedfbe1b9bbc55320728ec41581aa49c7de768d2a3bb2dd9bb401c1643.jpg) +Figure 4. Architecture of the CoSOD-TB. + +ure 5, as we can see, these features encode rich group consistency information that can well highlight the co-salient regions. + +# 3.3. CoSOD Transformer + +After getting the stochastic features $\nu_{i}$ from the LVGB, we leverage these features to guide the learning of generating the general features $\mathcal{F}_i$ from the CoSOD-TB to focus on the co-salient regions (see the bottom row of Figure 5). + +Figure 4 shows the architecture of the proposed CoSOD-TB. Each input image $I_{i}^{n} \in \mathcal{I}_{ei}$ is cropped into $d$ patches with size $H / 4 \times W / 4$ , and then all the image patches in each group construct the token sequences $\mathcal{X}_i^0 \in \mathbb{R}^{N \times \frac{H}{4} \times \frac{W}{4} \times 3d}$ . Then, $\mathcal{X}_i^0$ is sent into the backbone of the transformer with a T2T architecture [40], obtaining the token sequences $\mathcal{X}_i^1 \in \mathbb{R}^{N \times \frac{H}{16} \times \frac{W}{16} \times c}$ which encode both local and global information. After this, we design the group token $\mathcal{X}^G \in \mathbb{R}^{N \times 1 \times c}$ to capture group-wise common information that is essential to extract co-objects, and the saliency token $\mathcal{X}^S \in \mathbb{R}^{N \times 1 \times c}$ to capture the specific information that encodes more co-object structure details. Afterwards, $\mathcal{X}^G$ and $\mathcal{X}^1$ are concatenated in the channel dimension and fed into the Transformer Layer as the design in ViT [4] to get the new token sequence. We further use an multi-layer perceptron (MLP) layer to integrate global information into $\mathcal{X}^G$ . We fuse the $\mathcal{X}^S$ to integrate saliency information into sequences and split the $\mathcal{X}^G$ and $\mathcal{X}^S$ to up-sample the token sequence and get the token sequence $\mathcal{X}^2$ . After undergoing some processing as shown in Figure 4, $\mathcal{F}_i$ is then concatenated with the stochastic features $\mathcal{V}_i$ followed by a decoder that includes two Transformer Layers and an MLP layer. During decoding, due to the use of stochastic features, the model is more possible to overcome the prejudices formed in the early training. As shown in the third row of Figure 5, the CoSOD-TB can integrate $\mathcal{V}_i$ and $\mathcal{F}_i$ well, helping the model focus on more reliable co-salient object regions. Finally, the output sequences are reshaped to produce the predicted co-saliency maps $\hat{\mathcal{O}}$ . + +The loss function of the CoSOD-TB is defined as + +$$ +\mathcal {L} _ {T R A N S} = \frac {1}{2 N} \sum_ {n = 1} ^ {N} \sum_ {i = 1} ^ {2} \ell_ {B C E} \left(\mathcal {Y} _ {e i} (n,:), \hat {\mathcal {O}} _ {i} (n,:)\right), \tag {5} +$$ + +![](images/d6c019f4ac60ca461c8d3cfa98d78d068e0b4a53949b07943107ed80b1d12da8.jpg) + +![](images/d4d641740597727ff39488d3d9ad453f69f56ef2e2955afa833acc4bfe5bdbb6.jpg) +V + +![](images/cb8608e47752fd79e1e9882da09ef7f81477508803ef2786099b76293ac9c958.jpg) + +![](images/124f6d34d2e8ab093b7373f739fab125f59dca7bc4c5a77ff4cbc0264c1b2ee1.jpg) +Watermelon_2 + +![](images/b7147fc66c9abfc49b9649e6c16b391411ce5b53c35cbdbe07fab06424ba15c6.jpg) +F + +![](images/b1e0fa65a5d649d1604930a3813f64fe7d0dd798010061490e716100294cb3d3.jpg) + +![](images/d1b1a274ba14bd665d087b38eccbfb94a9239c7786bf1eb7e9145ac4567a7e8c.jpg) +F +V + +![](images/8beadddbfde6357610eea1203d2dff9b9710a533484077f218e47a01485f18b6.jpg) +V2-1 + +![](images/ab7f5f382d81f5218b6d9e49207827580a92de728e545fa866f39e53f75f7b83.jpg) +V2-2 + +![](images/a2d31cdfb33e17e2f8ea6334113ea29fe6aed812f8759f471aef85f46460a783.jpg) +$\mathcal{F}_i^1$ & $\mathcal{V}_i^{1 - 1}$ + +![](images/1a7138bfd942a6f1341a68be263627baf31060264d3cdb9b4659f72ad21e651c.jpg) +$\mathcal{F}_i^1$ & $\nu_{i}^{1 - 2}$ + +![](images/e21aa753c738f59337856cd23dbc34a6595e6bc1e805dd0695f85f19bf622bae.jpg) +$\mathcal{F}_i^2\& \mathcal{V}_i^{2 - 1}$ + +![](images/e4c4e261fb78a898daf990052e6f879f3ea7537a35314cf71b96742f2f7bbaf7.jpg) +$\mathcal{F}_i^2$ & $\mathcal{V}_i^{2 - 2}$ + +![](images/66c2176887f164ea6c566eab77593af273de301961cdd4dde98720b6dce642b3.jpg) +Figure 5. Comparison of effects of the "watermelon" group before and after fusing stochastic features. The first row: visualizations of the images and features obtained by the CoSOD-TB; The second row: visualizations of the features generated by the LVGB; The third row: visualizations of the integrated features. + +where $\ell_{BCE}$ is a binary cross-entropy (BCE) loss [27] defined as + +$$ +\begin{array}{l} \ell_ {B C E} (\mathcal {Y} (n,:), \hat {\mathcal {O}} (n,:)) = - (\mathcal {Y} (n,:) ^ {\top} \log (\hat {\mathcal {O}} (n,:)) \tag {6} \\ - \left(1 - \mathcal {Y} (n,:)\right) ^ {\top} \log \left(1 - \tilde {\mathcal {O}} (n,:)\right)). \\ \end{array} +$$ + +# 3.4. Loss Function + +The LVG and the CoSOD-TB are jointly trained in an end-to-end manner by optimizing the following multi-task loss + +$$ +\mathcal {L} = \lambda_ {1} \mathcal {L} _ {L V G} + \lambda_ {2} \mathcal {L} _ {T R A N S}, \tag {7} +$$ + +where $\lambda_{1},\lambda_{2}$ are the hyperparameters to balance each loss. + +# 4. Experimental Results + +# 4.1. Implementation Details + +Our model is implemented under the PyTorch1.9.0 framework [27]. Our computing platform's acceleration is provided by a GeForce GTX 2080Ti GPU. The transformer backbone is the pre-trained T2T-ViT $t - 14$ [40] model since it has a similar computational complexity as CNNs-based ResNet50 [11] which is smaller than the VGG-16 [31] widely used in CoSOD. The training dataset are COCOSEG released by [33] and DUTS released by [34]. They totally include about 208,250 images from 369 categories, and the relevant binary ground-truth masks are provided. + +In each training episode, we randomly select two groups, which are processed by the proposed GEM with $k = 1$ . Each group contains $N = 5$ images. The hyperparameters in (7) are set to $\lambda_1 = 0.25$ and $\lambda_2 = 1$ . We resize the images to $224 \times 224 \times 3$ and use them as input. We train the network over 60,000 steps totally and the training process takes about 11 hours. During the training process, we use + +![](images/5adf55c23f5906c200b6b7e773e66a611ef4a37658406c17f8f6e2c3bb2cad26.jpg) +Figure 6. The qualitative comparisons with the recent state-of-the-art methods, including DCFM [39], CADC [49], DeepACG [45] and GCoNet [9]. + +the Adam algorithm [16] to optimize the whole network. The initial learning rate is set to $10e - 4$ and decay to $10e - 5$ at the 20,000-th epoch. The inference time is 31 fps in average, meeting the real-time application requirements. + +# 4.2. Datasets and Evaluation Metrics + +The test sets include CoSal2015 [42], CoCA [50] and CoSOD3k [8]. Among them, CoSal2015 contains 2,015 images from 50 categories, and the objects therein have different sizes and shapes, which makes CoSal2015 challenging; CoCA is the newest dataset, including 1,297 images from 80 categories. These images have extremely different styles and very complicated backgrounds, making it a very difficult data set; CoSOD3k contains 3,316 images form 160 categories, which is by far the largest test set. It contains more disturbing targets, making it very challenging. + +Four evaluation metrics are used for comparison, including MAE [33], $E_{\phi}^{max}$ [6], $S_{\alpha}$ [5] and $F_{\beta}^{max}$ [1], which are to assess the average pixel-wise absolute difference, local and global similarity, structural similarity between the predictions and the ground truths, and the weighted harmonic mean of precision and recall, respectively. + +# 4.3. Comparisons with State-of-the-art Methods + +On the basis of unified evaluation codes [7] for fair comparison, we compare our method with several state-of-the-art methods published in recent three years, including RCAN [20], CSMG [46], SSNM [44], GCAGC [47], GICD [50], ICNet [15], GCoNet [9], DeepACG [45], CoEG-Net [7], HrSSMN [48], CADC [49], DCFM [39]. Among them, CADC [49] and DCFM [39] are the latest cutting-edge methods. More experimental results can be found in the supplementary materials. + +Qualitative Results. Figure 6 shows some visual comparison results with four latest state-of-the-art methods, including DCFM, CADC, DeepACG and GCoNet. The selected four groups are very challenging and suffer from highly confusing interference co-objects, extremely complex backgrounds, and drastic scale changes. Specifically, in the group "Baseball", the co-salient object is very small and the interference objects have similar shapes. Because all other methods are trapped in overconfidence, obsessed with the wrong results, and unable to extricate themselves. They cannot accurately detect the tiny objects and the background is wrongly segmented. In the group "Backpack", benefiting from training with noisy samples, our model is more robust that can accurately locate co-salient objects and segment target details, while other methods DCFM, CADC and GCoNet misjudge the co-salient objects. Camouflage co-objects in complex backgrounds are generally acknowledged as difficult samples [18], and the co-salient items in the "Gift box" group are perfectly in place. Thanks to CoSOD-TB's ability to capture long-range dependencies, our model learns strong feature representation that is able to accurately segment the camouflage co-salient objects, while CADC and GCoNet cannot finely segment the objects, and the other methods even fail in detecting the targets. The group "teddy bear" is also very challenging, and the co-salient objects in it suffer from interference objects and backgrounds with similar colors. Besides, the objects also have very different sizes. In this case, our model still achieves promising results, demonstrating its strong robustness to a variety of challenging factors. However, other methods yield a large number of inaccurate segmentation regions. + +Quantitative Results. Figure 7 shows the PR and F-measure curves of our method and other state-of-the-art + +![](images/a6e745f2961ddb5612fb833de3595d57838f3f4118b6c08d3a166e04ceb4eba5.jpg) + +![](images/99edc3e7150f9a04bdb9dc5e06abba64ec881c3e45bb05246a32c4dc9e1b4999.jpg) + +![](images/520c16fb7a0153f745e34bf8b8931bd9da222d2d9f43802bfe3db202108cd973.jpg) + +![](images/7bedf32491bcaba9436116a59b84f70b9f46a5d1acdce0420a4888440240dd33.jpg) +Figure 7. Comparisons with state-of-the-art methods since the year of 2019 in terms of PR and F-measure curves on three benchmark datasets. + +![](images/01981976420ccf7d7042e0193aa7e54e7cd345ad809be4748e1232ac25b5ff34.jpg) + +![](images/4051291ff930bb6c7daa9699aedd6930d9c5b19efd289e33baaa41fdb329785e.jpg) + +Table 1. Statistic comparisons of our model with the other state-of-the-arts since the year of 2019. + +
MethodsCoSal2015CoSOD3kCoCA
MAE↓Sα↑Eφmax↑Fβmax↑MAE↓Sα↑Eφmax↑Fβmax↑MAE↓Sα↑Eφmax↑Fβmax↑
RCAN(IJCAI2019)0.1260.7790.8420.7640.1300.7440.8080.6880.1600.6160.7020.422
CSMG(CVPR2019)0.1300.7740.8180.7770.1570.7110.7230.6450.1240.6320.7340.503
SSNM(AAAI2020)0.1020.7880.8430.7940.1200.7260.7560.6750.1160.6280.7410.482
GCAGC(CVPR2020)0.0850.8170.8660.8130.1000.7850.8160.7400.1180.6690.7540.523
GICD(ECCV2020)0.0720.8420.8840.8340.0890.7940.8310.7430.1250.6580.7010.504
ICNet(NIPS2020)0.0580.8570.9000.8580.0890.7940.8450.7620.1470.6540.7050.514
CoEGNet(TPAMI2021)0.0770.8360.8820.8320.0920.7620.8250.7360.1060.6120.7170.493
GCoNet(CVPR2021)0.0690.8450.8870.8470.0710.8020.8600.7500.1050.6730.7600.524
DeepACG(CVPR2021)0.0660.8530.8930.8470.0790.8110.8590.7790.1040.6850.7590.564
CADC(ICCV2021)0.0640.8660.9060.8620.0960.8010.8400.7590.1320.6810.7440.548
HrSSMN(TMM2022)0.0620.8450.8950.8410.0870.7880.8420.7530.1060.6710.7390.532
DCFM(CVPR2022)0.0670.8380.8920.8560.0670.8100.8740.8050.0850.7100.7830.598
OURS0.0530.8850.9330.8820.0610.8530.9110.8290.0950.7260.8080.599
+ +methods. As we can see, the PR curves produced by our method wrap around the curves generated by other methods. Also, all the F-measure curves of our method are above those generated by other methods. + +Meanwhile, we list the statistic comparison results of all compared methods in Table 1. It is clear that our approach achieves the most competitive result compared to its counterparts. Specifically, on the CoSal2015 dataset, our method achieves the best scores of 0.053, 0.885, 0.933, and 0.882 in terms of all metrics, with a significant gain of $0.5\%$ , $1.9\%$ , $2.7\%$ and $2.0\%$ , respectively, compared to the second-best preforming ICNet [15] and CADC [49]. On the most recent and difficult datasets, CoSOD3k and CoCA, our method also achieves the best performance. Especially + +on CoSOD3k, our method reaches the best scores of 0.061, 0.853, 0.911 and 0.829 in terms of $MAE$ , $E_{\phi}^{max}$ , $S_{\alpha}$ and $F_{\beta}^{max}$ , respectively, with a great gain of $0.6\%$ , $4.3\%$ , $3.7\%$ and $2.4\%$ compared to the second-best performing DCFM [39]. The effectiveness of our method has been fully validated on the large-scale and challenging datasets. + +# 4.4. Ablation Study + +To verify the effectiveness of our designs, we conduct the ablation study of our method on three datasets. We have made a lot of changes to VST [23] and regard the modified version as the baseline. Table 2 lists the results, and we can observe that every key design in our model makes a significant contribution to the overall performance. Taking + +Table 2. Ablations of our method on the effectiveness of the GEM , LVGB, and CoSOD-TB. + +
StrategiesCoSal2015CoSOD3kCoCA
GEMLVGBCoSOD-TBMAE ↓\( S_{\alpha} \uparrow \)\( E_{\phi}^{max} \uparrow \)\( F_{\beta}^{max} \uparrow \)MAE ↓\( S_{\alpha} \uparrow \)\( E_{\phi}^{max} \uparrow \)\( F_{\beta}^{max} \uparrow \)MAE ↓\( S_{\alpha} \uparrow \)\( E_{\phi}^{max} \uparrow \)\( F_{\beta}^{max} \uparrow \)
0.0600.8540.8870.8600.0750.7870.8630.7780.1050.7100.7850.564
✓0.0610.8830.9280.8770.0650.8340.8800.8170.1090.7100.7880.571
✓0.0580.8770.9260.8720.0630.8400.8710.8050.1000.7190.7890.574
✓0.0530.8740.8930.8800.0620.8420.9000.8140.0980.7160.7980.583
✓✓0.0550.8690.9300.8760.0630.8470.9060.8190.1040.7240.8020.597
✓✓0.0600.8860.9250.8720.0690.8500.8950.8230.1000.7180.7920.587
✓✓0.0540.8800.9190.8780.0610.8490.8890.8290.0960.7200.8050.595
✓✓✓0.0530.8850.9330.8820.0610.8530.9110.8290.0950.7260.8080.599
+ +Table 3. Complexity analyses. The number of inputs is set to 5 to maintain unity. + +
methodsFLOPs(G)↓param.(M)↓runtime(fps)↑Fβmax↑
GICD(ECCV20)364.7278.040.80.504
GCoNet(CVPR21)259.9142.0116.20.524
CADC(ICCV21)330.0392.818.00.548
DCFM(CVPR22)251.9142.384.40.598
OURS211.852.331.00.599
+ +![](images/1c635b45307c1fbe73aa2cbccc55df733e7ad7676be5a92ef6eb938f076349a2.jpg) +Figure 8. Comparison with the current most competitive method DCFM [39] in simulated real scenes. + +the results in CoSal2015 as examples, without GEM, the performance of our model gets worse in terms of all metrics especially for the $S_{\alpha}$ that drops from 0.885 to 0.869 by $1.6\%$ . With the GEM alone, the performance of our model improves greatly, and especially $S_{\alpha}$ and $E_{\phi}^{max}$ have a gain of $2.9\%$ and $4.1\%$ , respectively. Also, LVGB achieves the same excellent effect as GEM, because the uncertainty features provided by the LVGB can help the model consider more possibilities to enhance the robustness and overcome the over-confidence to non-salient regions. The CoSOD-TB makes important contributions to fine segmentation of images. With the CoSOD-TB, the MAE gets $0.7\%$ better from 0.060 to 0.053. The results on CoSOD3k and CoCA show the same trend, proving the effectiveness of each design of our method. + +# 4.5. Practical Application + +For the practicability of the algorithm, we conduct complexity comparisons with the same settings as DCFM [39]. The results can be seen in Table 3. Our method achieves both smallest FLOPs, parameters and best performance in terms of $F_{\beta}^{max}$ . Meanwhile, the runtime of our method is 31 fps, meeting the real-time requirements. This shows that our method is more suitable for deployment to various ap + +plication devices. + +We further test the performance of the model in the real scenes. In the real-life scenarios, there may not always exist co-salient objects within the group of images [22]. Figure 8 shows the images in the simulated real scenes that may not always exist co-objects. In these cases, the model should segment the co-objects existing in most images, while outputting all-zero masks on images without co-objects. Our method is better than the most competitive method DCFM [39] in dealing with real situations. When dealing with the images without co-objects, our model produces all-zero maps as much as possible. It shows that our method is more practical than other methods. More experimental results can be found in the supplementary materials. + +# 5. Conclusion + +In this paper, we have discovered that the paradigm of group consensus assumption has reduced the model's robustness and practical application value when confronted with irrelevant images in groups. A CoSOD model learning framework that is distinct from the classic CoSOD model learning framework has been developed. First, the group exchange-masking strategy has been devised, which is capable of automatically selecting the most informative noisy images from two groups and exchanging them in each other's groups to assist the model in learning more robust representations encoded with rich group consensus information. Second, the latent variable generator branch has been created to provide uncertainly-based global stochastic features that can regulate intermediate features from other branches to focus on co-objects. Third, the CoSOD transformer branch has been created to capture correlation-based global characteristics that carry information about group consistency. These branches' features are concatenated and put into a transformer-based decoder, yielding high-quality co-saliency maps. Extensive evaluations with and without irrelevant images have demonstrated the superiority of our method over a variety of state-of-the-art methods. In the future, we plan to build a new dataset for real scenarios that can provide a platform to evaluate model robustness and practical value in a more reliable way. Further more, we will consider the potential application of our method in defending adversarial attacks [35]. + +# References + +[1] Radhakrishna Achanta, Sheila Hemami, Francisco Estrada, and Sabine Susstrunk. Frequency-tuned salient region detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1597-1604. IEEE, 2009. 6 +[2] Kai-Yueh Chang, Tyng-Luh Liu, and Shang-Hong Lai. From co-saliency to co-segmentation: An efficient and fully unsupervised energy minimization model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2129-2136. IEEE, 2011. 1, 2 +[3] Paul Christiano, Zain Shah, Igor Mordatch, Jonas Schneider, Trevor Blackwell, Joshua Tobin, Pieter Abbeel, and Wojciech Zaremba. Transfer from simulation to real world through learning deep inverse dynamics model. arXiv preprint arXiv:1610.03518, 2016. 2 +[4] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 5 +[5] Deng-Ping Fan, Ming-Ming Cheng, Yun Liu, Tao Li, and Ali Borji. Structure-measure: A new way to evaluate foreground maps. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4548-4557, 2017. 6 +[6] Deng-Ping Fan, Cheng Gong, Yang Cao, Bo Ren, MingMing Cheng, and Ali Borji. Enhanced-alignment measure for binary foreground map evaluation. arXiv preprint arXiv:1805.10421, 2018. 6 +[7] Deng-Ping Fan, Tengpeng Li, Zheng Lin, Ge-Peng Ji, Dingwen Zhang, Ming-Ming Cheng, Huazhu Fu, and Jianbing Shen. Re-thinking co-salient object detection. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2021. 2, 6 +[8] Deng-Ping Fan, Zheng Lin, Ge-Peng Ji, Dingwen Zhang, Huazhu Fu, and Ming-Ming Cheng. Taking a deeper look at co-salient object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2919-2929, 2020. 1, 6 +[9] Qi Fan, Deng-Ping Fan, Huazhu Fu, Chi-Keung Tang, Ling Shao, and Yu-Wing Tai. Group collaborative learning for cosalient object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12288–12298, 2021. 1, 2, 6 +[10] Huazhu Fu, Xiaochun Cao, and Zhuowen Tu. Cluster-based co-saliency detection. IEEE Transactions on Image Processing, 22(10):3766-3778, 2013. 1, 2 +[11] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 770-778, 2016. 5 +[12] Kuang-Jui Hsu, Yen-Yu Lin, and Yung-Yu Chuang. Coattention cnns for unsupervised object co-segmentation. In Proceedings of the International Joint Conference on Artificial Intelligence, pages 748-756, 2018. 2 + +[13] David E Jacobs, Dan B Goldman, and Eli Shechtman. Cosaliency: Where people look when comparing images. In Proceedings of the 23nd Annual ACM Symposium on User Interface Software and Technology, pages 219-228, 2010. 1, 2 +[14] Koteswar Rao Jerripothula, Jianfei Cai, and Junsong Yuan. Cats: Co-saliency activated tracklet selection for video colocalization. In Proceedings of the European Conference on Computer Vision, pages 187-202. Springer, 2016. 1 +[15] Wen-Da Jin, Jun Xu, Ming-Ming Cheng, Yi Zhang, and Wei Guo. Icnet: Intra-saliency correlation network for co-saliency detection. Advances in Neural Information Processing Systems, 33:18749-18759, 2020. 1, 2, 6, 7 +[16] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6 +[17] Simon Kohl, Bernardino Romera-Paredes, Clemens Meyer, Jeffrey De Fauw, Joseph R Ledsam, Klaus Maier-Hein, SM Eslami, Danilo Jimenez Rezende, and Olaf Ronneberger. A probabilistic u-net for segmentation of ambiguous images. Advances in Neural Information Processing Systems, 31, 2018. 4 +[18] Aixuan Li, Jing Zhang, Yunqiu Lv, Bowen Liu, Tong Zhang, and Yuchao Dai. Uncertainty-aware joint salient object and camouflaged object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10071-10081, 2021. 6 +[19] Bo Li, Zhengxing Sun, and Yuqi Guo. Supervae: Superpixelwise variational autoencoder for salient object detection. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 33, pages 8569-8576, 2019. 2 +[20] Bo Li, Zhengxing Sun, Lv Tang, Yunhan Sun, and Jinlong Shi. Detecting robust co-saliency with recurrent coattention neural network. In Proceedings of the International Joint Conference on Artificial Intelligence, volume 2, page 6, 2019. 1, 6 +[21] Yijun Li, Keren Fu, Zhi Liu, and Jie Yang. Efficient saliency-model-guided visual co-saliency detection. IEEE Signal Processing Letters, 22(5):588-592, 2014. 2 +[22] Jiawei Liu, Jing Zhang, Kaihao Zhang, and Nick Barnes. Generalised co-salient object detection. arXiv preprint arXiv:2208.09668, 2022. 8 +[23] Nian Liu, Ni Zhang, Kaiyuan Wan, Ling Shao, and Junwei Han. Visual saliency transformer. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2021. 7 +[24] Xuanqing Liu, Minhao Cheng, Huan Zhang, and Cho-Jui Hsieh. Towards robust neural networks via random self-ensemble. In Proceedings of the European Conference on Computer Vision, pages 369–385, 2018. 3 +[25] Xuanqing Liu, Yao Li, Chongruo Wu, and Cho-Jui Hsieh. Adv-bnn: Improved adversarial defense through robust bayesian neural network. arXiv preprint arXiv:1810.01279, 2018.3 +[26] Aleksander Madry, Aleksandar Makelov, Ludwig Schmidt, Dimitris Tsipras, and Adrian Vladu. Towards deep learning models resistant to adversarial attacks. In International Conference on Learning Representations, 2018. 4 + +[27] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. Advances in Neural Information Processing Systems, 32:8026-8037, 2019. 5 +[28] Lerrel Pinto, James Davidson, Rahul Sukthankar, and Abhi nav Gupta. Robust adversarial reinforcement learning. In International Conference on Machine Learning, pages 2817-2826. PMLR, 2017. 2 +[29] Aravind Rajeswaran, Sarvjeet Ghotra, Balaraman Ravindran, and Sergey Levine. Epopt: Learning robust neural network policies using model ensembles. arXiv preprint arXiv:1610.01283, 2016. 3 +[30] Andrei A Rusu, Matej Večerík, Thomas Rothörl, Nicolas Heess, Razvan Pascanu, and Raia Hadsell. Sim-to-real robot learning from pixels with progressive nets. In Conference on robot learning, pages 262-270. PMLR, 2017. 2 +[31] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556, 2014. 5 +[32] Kihyuk Sohn, Honglak Lee, and Xinchen Yan. Learning structured output representation using deep conditional generative models. Advances in Neural Information Processing Systems, 28, 2015. 2, 4 +[33] Chong Wang, Zheng-Jun Zha, Dong Liu, and Hongtao Xie. Robust deep co-saliency detection with group semantic. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 33, pages 8917-8924, 2019. 1, 5, 6 +[34] Lijun Wang, Huchuan Lu, Yifan Wang, Mengyang Feng, Dong Wang, Baocai Yin, and Xiang Ruan. Learning to detect salient objects with image-level supervision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 136-145, 2017. 5 +[35] Yisen Wang, Difan Zou, Jinfeng Yi, James Bailey, Xingjun Ma, and Quanquan Gu. Improving adversarial robustness requires revisiting misclassified examples. In International Conference on Learning Representations, 2019. 8 +[36] Lina Wei, Shanshan Zhao, Omar El Farouk Bourahla, Xi Li, and Fei Wu. Group-wise deep co-saliency detection. arXiv preprint arXiv:1707.07381, 2017. 1, 2 +[37] Cihang Xie, Jianyu Wang, Zhishuai Zhang, Zhou Ren, and Alan Yuille. Mitigating adversarial effects through randomization. arXiv preprint arXiv:1711.01991, 2017. 3 +[38] Xi Yang, Shaoyi Li, Jun Ma, Jun-yan Yang, and Jie Yan. Co-saliency-regularized correlation filter for object tracking. Signal Processing: Image Communication, 103:116655, 2022. 1 +[39] Siyue Yu, Jimin Xiao, Bingfeng Zhang, and Eng Gee Lim. Democracy does matter: Comprehensive feature mining for co-salient object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 979–988, 2022. 1, 6, 7, 8 +[40] Li Yuan, Yunpeng Chen, Tao Wang, Weihao Yu, Yujun Shi, Zi-Hang Jiang, Francis EH Tay, Jiashi Feng, and Shuicheng Yan. Tokens-to-token vit: Training vision transformers from scratch onImagenet. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2021. 5 + +[41] Chen Zhang, Riccardo Barbano, and Bangti Jin. Conditional variational autoencoder for learned image reconstruction. Computation, 9(11):114, 2021. 2 +[42] Dingwen Zhang, Junwei Han, Chao Li, and Jingdong Wang. Co-saliency detection via looking deep and wide. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2994-3002, 2015. 2, 6 +[43] Jing Zhang, Deng-Ping Fan, Yuchao Dai, Saeed Anwar, Fatemeh Saleh, Sadegh Aliakbarian, and Nick Barnes. Uncertainty inspired rgb-d saliency detection. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2021. 2, 4 +[44] Kaihua Zhang, Jin Chen, Bo Liu, and Qingshan Liu. Deep object co-segmentation via spatial-semantic network modulation. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 34, pages 12813–12820, 2020. 1, 6 +[45] Kaihua Zhang, Mingliang Dong, Bo Liu, Xiao-Tong Yuan, and Qingshan Liu. Deepacg: Co-saliency detection via semantic-aware contrast gromov-wasserstein distance. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13703-13712, 2021. 6 +[46] Kaihua Zhang, Tengpeng Li, Bo Liu, and Qingshan Liu. Cosaliency detection via mask-guided fully convolutional networks with multi-scale label smoothing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3095-3104, 2019. 1, 2, 6 +[47] Kaihua Zhang, Tengpeng Li, Shiwen Shen, Bo Liu, Jin Chen, and Qingshan Liu. Adaptive graph convolutional network with attention graph clustering for co-saliency detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9050-9059, 2020. 2, 6 +[48] Kaihua Zhang, Yang Wu, Mingliang Dong, Bo Liu, Dong Liu, and Qingshan Liu. Deep object co-segmentation and c-osaliency detection via high-order spatial-semantic network modulation. IEEE Transactions on Multimedia, 2022. 1, 6 +[49] Ni Zhang, Junwei Han, Nian Liu, and Ling Shao. Summarize and search: Learning consensus-aware dynamic convolution for co-saliency detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4167-4176, 2021. 6, 7 +[50] Zhao Zhang, Wenda Jin, Jun Xu, and Ming-Ming Cheng. Gradient-induced co-saliency detection. In Proceedings of the European Conference on Computer Vision, pages 455-472. Springer, 2020. 1, 2, 6 \ No newline at end of file diff --git a/2023/Co-Salient Object Detection With Uncertainty-Aware Group Exchange-Masking/images.zip b/2023/Co-Salient Object Detection With Uncertainty-Aware Group Exchange-Masking/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..dc3823a92ebace1f11d7d3ae88b9a1eb6d37e526 --- /dev/null +++ b/2023/Co-Salient Object Detection With Uncertainty-Aware Group Exchange-Masking/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d25ed2e305fb4fcb101522fb907d6fcb1e851e261ec85f6c60bd71c136134e8e +size 782392 diff --git a/2023/Co-Salient Object Detection With Uncertainty-Aware Group Exchange-Masking/layout.json b/2023/Co-Salient Object Detection With Uncertainty-Aware Group Exchange-Masking/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..d3e7cf5aaf1f34d3ab9002b8b4a266d1f0ffd399 --- /dev/null +++ b/2023/Co-Salient Object Detection With Uncertainty-Aware Group Exchange-Masking/layout.json @@ -0,0 +1,10435 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 52, + 103, + 541, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 52, + 103, + 541, + 121 + ], + "spans": [ + { + "bbox": [ + 52, + 103, + 541, + 121 + ], + "type": "text", + "content": "Co-Salient Object Detection with Uncertainty-aware Group Exchange-Masking" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 124, + 142, + 468, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 124, + 142, + 468, + 157 + ], + "spans": [ + { + "bbox": [ + 124, + 142, + 468, + 157 + ], + "type": "text", + "content": "Yang Wu" + }, + { + "bbox": [ + 124, + 142, + 468, + 157 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 124, + 142, + 468, + 157 + ], + "type": "text", + "content": " Huihui Song" + }, + { + "bbox": [ + 124, + 142, + 468, + 157 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 124, + 142, + 468, + 157 + ], + "type": "text", + "content": " Bo Liu" + }, + { + "bbox": [ + 124, + 142, + 468, + 157 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 124, + 142, + 468, + 157 + ], + "type": "text", + "content": " Kaihua Zhang" + }, + { + "bbox": [ + 124, + 142, + 468, + 157 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 124, + 142, + 468, + 157 + ], + "type": "text", + "content": " Dong Liu" + }, + { + "bbox": [ + 124, + 142, + 468, + 157 + ], + "type": "inline_equation", + "content": "^{3}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 157, + 538, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 157, + 538, + 171 + ], + "spans": [ + { + "bbox": [ + 56, + 157, + 538, + 171 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 56, + 157, + 538, + 171 + ], + "type": "text", + "content": "B-DAT and CICAET, Nanjing University of Information Science and Technology, Nanjing, China" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 170, + 171, + 423, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 170, + 171, + 423, + 185 + ], + "spans": [ + { + "bbox": [ + 170, + 171, + 423, + 185 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 170, + 171, + 423, + 185 + ], + "type": "text", + "content": "Walmart Global Tech, Sunnyvale, CA, 94086, USA" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 196, + 185, + 397, + 198 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 196, + 185, + 397, + 198 + ], + "spans": [ + { + "bbox": [ + 196, + 185, + 397, + 198 + ], + "type": "text", + "content": "3Netflix Inc, Los Gatos, CA, 95032, USA" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 230, + 201, + 356, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 230, + 201, + 356, + 213 + ], + "spans": [ + { + "bbox": [ + 230, + 201, + 356, + 213 + ], + "type": "text", + "content": "songhuihui@nuist.edu.cn" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 143, + 241, + 191, + 253 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 241, + 191, + 253 + ], + "spans": [ + { + "bbox": [ + 143, + 241, + 191, + 253 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 266, + 290, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 266, + 290, + 578 + ], + "spans": [ + { + "bbox": [ + 46, + 266, + 290, + 578 + ], + "type": "text", + "content": "The traditional definition of co-salient object detection (CoSOD) task is to segment the common salient objects in a group of relevant images. Existing CoSOD models by default adopt the group consensus assumption. This brings about model robustness defect under the condition of irrelevant images in the testing image group, which hinders the use of CoSOD models in real-world applications. To address this issue, this paper presents a group exchange-masking (GEM) strategy for robust CoSOD model learning. With two groups of image containing different types of salient object as input, the GEM first selects a set of images from each group by the proposed learning based strategy, then these images are exchanged. The proposed feature extraction module considers both the uncertainty caused by the irrelevant images and group consensus in the remaining relevant images. We design a latent variable generator branch which is made of conditional variational autoencoder to generate uncertainly-based global stochastic features. A CoSOD transformer branch is devised to capture the correlation-based local features that contain the group consistency information. At last, the output of two branches are concatenated and fed into a transformer-based decoder, producing robust co-saliency prediction. Extensive evaluations on co-saliency detection with and without irrelevant images demonstrate the superiority of our method over a variety of state-of-the-art methods." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 599, + 127, + 611 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 599, + 127, + 611 + ], + "spans": [ + { + "bbox": [ + 47, + 599, + 127, + 611 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 619, + 287, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 619, + 287, + 667 + ], + "spans": [ + { + "bbox": [ + 46, + 619, + 287, + 667 + ], + "type": "text", + "content": "Co-salient object detection (CoSOD) is to segment the common salient objects in a group of relevant images. By detecting the co-salient object in a group of images, the images' background and redundant content are re" + } + ] + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 309, + 241, + 399, + 355 + ], + "blocks": [ + { + "bbox": [ + 309, + 241, + 399, + 355 + ], + "lines": [ + { + "bbox": [ + 309, + 241, + 399, + 355 + ], + "spans": [ + { + "bbox": [ + 309, + 241, + 399, + 355 + ], + "type": "image", + "image_path": "94854fdc8c47034ac3e62cc651aca8f0721e02daba5b07d7d0b6cc6c9a217f61.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 367, + 546, + 422 + ], + "lines": [ + { + "bbox": [ + 305, + 367, + 546, + 422 + ], + "spans": [ + { + "bbox": [ + 305, + 367, + 546, + 422 + ], + "type": "text", + "content": "Figure 1. (a) When there exists an irrelevant image in the test group, the state-of-the-art CoSOD models such as DCFM [39] and GCoNet [9] tend to generate false positive predictions for the noisy image, yet our method can achieve an accurate prediction due to the use of group exchange-masking strategy in (b)." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 400, + 241, + 542, + 355 + ], + "blocks": [ + { + "bbox": [ + 400, + 241, + 542, + 355 + ], + "lines": [ + { + "bbox": [ + 400, + 241, + 542, + 355 + ], + "spans": [ + { + "bbox": [ + 400, + 241, + 542, + 355 + ], + "type": "image", + "image_path": "a70bec057487f1f85338cccfaa7d71bf4ebaa5ea13ba93fe6c3e2531cd1797c5.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 447, + 545, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 447, + 545, + 483 + ], + "spans": [ + { + "bbox": [ + 304, + 447, + 545, + 483 + ], + "type": "text", + "content": "moved, which helps the downstream tasks such as object tracking [38], co-segmentation [48] and video colocalization [14], to name a few." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 486, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 486, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 486, + 546, + 713 + ], + "type": "text", + "content": "Group consensus assumption is widely used by existing CoSOD models, i.e., these models presume that all images in the same group contain the common salient targets. The widely-used CoSOD benchmark datasets, such as COCO-SEG [33], CoCA [50], and CoSOD3k [8] organize the training and testing images that contain the same object as a group. Many existing CoSOD models consider the group consensus characteristic in modeling. For example, early works such as [2, 10, 13] extract hand-crafted features for inter-image co-object correspondence discovery. The deep learning models proposed in [15, 20, 36, 39, 44, 46, 48] use one group of relevant images as training data input for consensus representation learning. Among them, a variety of novel model design techniques have been developed to make full use of the group consistency characteristic, such as the low-rank feature learning [46], co-attention model [20] and intra-saliency correlation learning [15]. The issues of the group consensus assumption are partially studied in recent literature [9], which further models the inter" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 673, + 287, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 673, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 46, + 673, + 287, + 712 + ], + "type": "text", + "content": "*Corresponding author. This work is supported in part by National Key Research and Development Program of China under Grant No. 2018AAA0100400, in part by the NSFC under Grant Nos. 62276141, 61872189." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 294, + 732, + 317, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 732, + 317, + 742 + ], + "spans": [ + { + "bbox": [ + 294, + 732, + 317, + 742 + ], + "type": "text", + "content": "19639" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 287, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 287, + 95 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 287, + 95 + ], + "type": "text", + "content": "group separability for more discriminative feature learning by a group collaborating module." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 97, + 288, + 370 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 97, + 288, + 370 + ], + "spans": [ + { + "bbox": [ + 46, + 97, + 288, + 370 + ], + "type": "text", + "content": "In this paper, we find that the group consensus assumption also restricts the CoSOD model's robustness against the images without common object. As illustrated by Figure 1(a), state-of-the-art CoSOD models tend to output false positive predictions for the irrelevant image. This issue hinders the use of CoSOD models in real-world applications where the testing inputs are likely to contain irrelevant images. To enhance the model's robustness, we propose a learning framework called group exchange-masking (GEM). The GEM is illustrated by Figure 1(b). Given two image groups that contain different types of co-salient objects, we exchange several images between one group and the other. Those exchanged images are called noisy images. The number of noisy images is chosen to be less than the number of remaining relevant images in the group, so that the co-salient object in the noisy images forms a negative object but not the dominant co-salient object. The \"masking\" strategy refers to the label regeneration of the noisy images. Because there is no co-salient object, in the regenerated label, the original ground-truth object is masked. The learning objective is to correctly predict both the co-salient objects in the original relevant images and the added noisy images." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 371, + 288, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 371, + 288, + 635 + ], + "spans": [ + { + "bbox": [ + 46, + 371, + 288, + 635 + ], + "type": "text", + "content": "Adding noisy images to the training image group brings about uncertainty to the CoSOD model learning since there is some probability of no expected common object in each image. We design a dual-path image feature extraction module to model the group uncertainty in addition to the group consensus property. Specifically, we design a latent variable generator branch (LVGB) to extract the uncertainty-based global image features. The LVGB module is motivated by the conditional variational autoencoder (CVAE) [32] that is widely used to address the uncertainty in vision applications including image background modeling [19], RGB-D saliency detection [43] and image reconstruction [41]. In parallel with LVGB, we feed the image group into a CoSOD Transformer Branch (CoSOD-TB). The CoSOD-TB partitions each image group into local patches, and the attention mechanism in the transformer enables this branch to model patch-wise correlation-based local features. As a result, the group consistency information can be captured by this branch. The outputs of the two branches are concatenated and fed into a transformer-based decoder for co-saliency prediction. The proposed model has the following technical contributions." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 58, + 641, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 641, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 58, + 641, + 288, + 714 + ], + "type": "text", + "content": "- A robust CoSOD model learning mechanism, called group exchange-masking is proposed. By exchanging images between two groups, we augment the training data containing irrelevant images as noise to enhance model's robustness. This is different from the traditional CoSOD model learning frameworks that use" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 326, + 72, + 497, + 85 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 72, + 497, + 85 + ], + "spans": [ + { + "bbox": [ + 326, + 72, + 497, + 85 + ], + "type": "text", + "content": "groups of relevant images as training data." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 317, + 92, + 547, + 258 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 317, + 92, + 545, + 164 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 92, + 545, + 164 + ], + "spans": [ + { + "bbox": [ + 317, + 92, + 545, + 164 + ], + "type": "text", + "content": "- We propose a dual-path feature extraction module composed of the LVGB and the CoSOD-TB. The LVGB is designed to model the uncertainty of cosalient object existence. The CoSOD-TB is for the consensus feature extraction of the salient object in the relevant images." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 317, + 173, + 547, + 258 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 173, + 547, + 258 + ], + "spans": [ + { + "bbox": [ + 317, + 173, + 547, + 258 + ], + "type": "text", + "content": "- Extensive evaluations on three benchmark datasets, including CoSal2015 [42], CoCA [15], and CoSOD3k [7] show that the superiority of the proposed model to the state-of-the-art methods in terms of all evaluation metrics. Besides, the proposed model demonstrates good robustness for dealing with noisy data without co-salient objects." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 277, + 392, + 290 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 277, + 392, + 290 + ], + "spans": [ + { + "bbox": [ + 306, + 277, + 392, + 290 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 297, + 458, + 310 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 297, + 458, + 310 + ], + "spans": [ + { + "bbox": [ + 306, + 297, + 458, + 310 + ], + "type": "text", + "content": "2.1. Co-salient Object Detection" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 316, + 545, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 316, + 545, + 628 + ], + "spans": [ + { + "bbox": [ + 304, + 316, + 545, + 628 + ], + "type": "text", + "content": "In the past, CoSOD methods used to extract handcrafted features such as Gabor and SIFT features from images and then detect co-saliency by utilizing the consistency of low-level features between the images being tested [2]. A series of studies attempt to capture the intra-image constraints by employing a manifold ranking scheme to produce saliency maps [21], or using a global association constraint with clustering [10], or translational alignment [13]. More recently, there is a surge of deep learning-based CoSOD models that learn feature representation and saliency predictor in an end-to-end manner [9, 12, 36, 46, 47, 50]. Wei et al. [36] designs a collaborative learning framework for CoSOD that discovers the collaborative and interactive relationships between intra-image and single-image feature representations in a group-wise manner. Hsu et al. [12] present an unsupervised CNN-based model for CoSOD. In [46], a hierarchical framework is proposed for CoSOD where the initial CoSOD results generated by the CNN model is refined by label smoothing. Zhang et al. [50] propose a gradient-induced model for CoSOD that utilizes the image gradient information to induce more attentions to the discriminative co-salient feature learning. In [47], a deep graph neural network model is proposed to characterize the intra-image and inter-image region correspondence for CoSOD. In [9], a group collaborative learning strategy is proposed to explore inter-group relations for discriminating feature learning." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 635, + 500, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 635, + 500, + 647 + ], + "spans": [ + { + "bbox": [ + 306, + 635, + 500, + 647 + ], + "type": "text", + "content": "2.2. Robust Model and Feature Learning" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 654, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 547, + 715 + ], + "type": "text", + "content": "In past years, with the development of deep learning, robust model and feature learning have received more and more attention as a possible solution to overcoming the bottleneck [28]. Various methods have been proposed to improve the robustness of the model. In [3] and [30], ways to" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 732, + 318, + 742 + ], + "type": "footer", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 732, + 318, + 742 + ], + "spans": [ + { + "bbox": [ + 294, + 732, + 318, + 742 + ], + "type": "text", + "content": "19640" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 70, + 547, + 276 + ], + "blocks": [ + { + "bbox": [ + 50, + 70, + 547, + 276 + ], + "lines": [ + { + "bbox": [ + 50, + 70, + 547, + 276 + ], + "spans": [ + { + "bbox": [ + 50, + 70, + 547, + 276 + ], + "type": "image", + "image_path": "f3256742c23bdfa633fb2e92436604e49167e6b70d2961de41df887a448d6a2d.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 285, + 547, + 352 + ], + "lines": [ + { + "bbox": [ + 46, + 285, + 547, + 352 + ], + "spans": [ + { + "bbox": [ + 46, + 285, + 547, + 352 + ], + "type": "text", + "content": "Figure 2. Pipeline of the proposed method. First, two groups of images " + }, + { + "bbox": [ + 46, + 285, + 547, + 352 + ], + "type": "inline_equation", + "content": "\\{\\mathcal{I}_1,\\mathcal{I}_2\\}" + }, + { + "bbox": [ + 46, + 285, + 547, + 352 + ], + "type": "text", + "content": " with their masks " + }, + { + "bbox": [ + 46, + 285, + 547, + 352 + ], + "type": "inline_equation", + "content": "\\{\\mathcal{V}_1,\\mathcal{V}_2\\}" + }, + { + "bbox": [ + 46, + 285, + 547, + 352 + ], + "type": "text", + "content": " are processed by group exchange-masking, yielding the exchanged images " + }, + { + "bbox": [ + 46, + 285, + 547, + 352 + ], + "type": "inline_equation", + "content": "\\{\\mathcal{I}_{e1},\\mathcal{I}_{e2}\\}" + }, + { + "bbox": [ + 46, + 285, + 547, + 352 + ], + "type": "text", + "content": " and their corresponding masked labels " + }, + { + "bbox": [ + 46, + 285, + 547, + 352 + ], + "type": "inline_equation", + "content": "\\{\\mathcal{V}_{e1},\\mathcal{V}_{e2}\\}" + }, + { + "bbox": [ + 46, + 285, + 547, + 352 + ], + "type": "text", + "content": ". Then, " + }, + { + "bbox": [ + 46, + 285, + 547, + 352 + ], + "type": "inline_equation", + "content": "\\{\\mathcal{I}_{e1},\\mathcal{I}_{e2}\\}" + }, + { + "bbox": [ + 46, + 285, + 547, + 352 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 285, + 547, + 352 + ], + "type": "inline_equation", + "content": "\\{\\mathcal{V}_{e1},\\mathcal{V}_{e2}\\}" + }, + { + "bbox": [ + 46, + 285, + 547, + 352 + ], + "type": "text", + "content": " are passed through the latent variable generator branch, extracting the uncertainty-based global image features " + }, + { + "bbox": [ + 46, + 285, + 547, + 352 + ], + "type": "inline_equation", + "content": "\\{\\mathcal{V}_1,\\mathcal{V}_2\\}" + }, + { + "bbox": [ + 46, + 285, + 547, + 352 + ], + "type": "text", + "content": " to eliminate the bias caused by noisy images. Meanwhile, " + }, + { + "bbox": [ + 46, + 285, + 547, + 352 + ], + "type": "inline_equation", + "content": "\\{\\mathcal{I}_{e1},\\mathcal{I}_{e2}\\}" + }, + { + "bbox": [ + 46, + 285, + 547, + 352 + ], + "type": "text", + "content": " are passed through the weight-shared CoSOD transformer branch, producing two groups of feature sequences " + }, + { + "bbox": [ + 46, + 285, + 547, + 352 + ], + "type": "inline_equation", + "content": "\\{\\mathcal{F}_1,\\mathcal{F}_2\\}" + }, + { + "bbox": [ + 46, + 285, + 547, + 352 + ], + "type": "text", + "content": " with group consistency information and long-range independent information. Finally, " + }, + { + "bbox": [ + 46, + 285, + 547, + 352 + ], + "type": "inline_equation", + "content": "\\mathcal{F}" + }, + { + "bbox": [ + 46, + 285, + 547, + 352 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 285, + 547, + 352 + ], + "type": "inline_equation", + "content": "\\mathcal{V}" + }, + { + "bbox": [ + 46, + 285, + 547, + 352 + ], + "type": "text", + "content": " are concatenated and fed into the transformer decoder, yielding the predicted co-saliency maps " + }, + { + "bbox": [ + 46, + 285, + 547, + 352 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{O}}" + }, + { + "bbox": [ + 46, + 285, + 547, + 352 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 371, + 289, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 371, + 289, + 515 + ], + "spans": [ + { + "bbox": [ + 46, + 371, + 289, + 515 + ], + "type": "text", + "content": "transfer the policy learned in a simulator to the real world are proposed to make features more robust. Rajeswaran et al. [29] learn a robust policy by sampling the worst case trajectories from a class of parametrized models, to learn a robust policy. Xie et al. [37] enhance the robustness of the model through randomly resizing and padding the training images. Several works leverage randomness to improve the robustness of models. Liu et al. [24] propose a noise layer that introduces randomness to both the input and the hidden layer output. Liu et al. [25] introduce a new min-max formulation that combines adversarial training with Bayesian Neural Networks, which achieves promising results." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 525, + 153, + 539 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 525, + 153, + 539 + ], + "spans": [ + { + "bbox": [ + 47, + 525, + 153, + 539 + ], + "type": "text", + "content": "3. Proposed Method" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 545, + 289, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 545, + 289, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 545, + 289, + 713 + ], + "type": "text", + "content": "Figure 2 illustrates the pipeline of our framework. Given two groups of " + }, + { + "bbox": [ + 46, + 545, + 289, + 713 + ], + "type": "inline_equation", + "content": "2N" + }, + { + "bbox": [ + 46, + 545, + 289, + 713 + ], + "type": "text", + "content": " images " + }, + { + "bbox": [ + 46, + 545, + 289, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{I}_i = \\{I_i^n\\in \\mathbb{R}^{H\\times W\\times 3}\\}_{n = 1}^N,i = 1,2" + }, + { + "bbox": [ + 46, + 545, + 289, + 713 + ], + "type": "text", + "content": " with height " + }, + { + "bbox": [ + 46, + 545, + 289, + 713 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 46, + 545, + 289, + 713 + ], + "type": "text", + "content": " and width " + }, + { + "bbox": [ + 46, + 545, + 289, + 713 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 46, + 545, + 289, + 713 + ], + "type": "text", + "content": " and manually-labeled binary masks " + }, + { + "bbox": [ + 46, + 545, + 289, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{V}_i = \\{\\mathbf{Y}_i^n\\in \\mathbb{R}^{H\\times W}\\}_{n = 1}^N" + }, + { + "bbox": [ + 46, + 545, + 289, + 713 + ], + "type": "text", + "content": " with different categories as input, we employ the GEM model to reorganize them, yielding " + }, + { + "bbox": [ + 46, + 545, + 289, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{I}_{ei}" + }, + { + "bbox": [ + 46, + 545, + 289, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 545, + 289, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{V}_{ei}" + }, + { + "bbox": [ + 46, + 545, + 289, + 713 + ], + "type": "text", + "content": ". Each pair of " + }, + { + "bbox": [ + 46, + 545, + 289, + 713 + ], + "type": "inline_equation", + "content": "\\{\\mathcal{I}_{ei},\\mathcal{V}_{ei}\\}" + }, + { + "bbox": [ + 46, + 545, + 289, + 713 + ], + "type": "text", + "content": " are fed to the LVGB. In LVGB, we draw " + }, + { + "bbox": [ + 46, + 545, + 289, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_i\\in \\mathbb{R}^K" + }, + { + "bbox": [ + 46, + 545, + 289, + 713 + ], + "type": "text", + "content": " from the Gaussian distribution " + }, + { + "bbox": [ + 46, + 545, + 289, + 713 + ], + "type": "inline_equation", + "content": "P_{\\theta}(\\mathbf{z}_i|\\mathcal{I}_{ei})" + }, + { + "bbox": [ + 46, + 545, + 289, + 713 + ], + "type": "text", + "content": " with the learning parameters " + }, + { + "bbox": [ + 46, + 545, + 289, + 713 + ], + "type": "inline_equation", + "content": "\\pmb{\\theta}" + }, + { + "bbox": [ + 46, + 545, + 289, + 713 + ], + "type": "text", + "content": " and the posterior of " + }, + { + "bbox": [ + 46, + 545, + 289, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_i" + }, + { + "bbox": [ + 46, + 545, + 289, + 713 + ], + "type": "text", + "content": " is formulated as " + }, + { + "bbox": [ + 46, + 545, + 289, + 713 + ], + "type": "inline_equation", + "content": "Q_{\\phi}(\\mathbf{z}_i|\\mathcal{I}_{ei},\\mathcal{V}_{ei})" + }, + { + "bbox": [ + 46, + 545, + 289, + 713 + ], + "type": "text", + "content": " with the learning parameters " + }, + { + "bbox": [ + 46, + 545, + 289, + 713 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 46, + 545, + 289, + 713 + ], + "type": "text", + "content": ". We use the KL-Divergence metric to reduce the distance between " + }, + { + "bbox": [ + 46, + 545, + 289, + 713 + ], + "type": "inline_equation", + "content": "P_{\\theta}(\\mathbf{z}_i|\\mathcal{I}_{ei})" + }, + { + "bbox": [ + 46, + 545, + 289, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 545, + 289, + 713 + ], + "type": "inline_equation", + "content": "Q_{\\phi}(\\mathbf{z}_i|\\mathcal{I}_{ei},\\mathcal{V}_{ei})" + }, + { + "bbox": [ + 46, + 545, + 289, + 713 + ], + "type": "text", + "content": ". We further re-parameterize and expand " + }, + { + "bbox": [ + 46, + 545, + 289, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_i" + }, + { + "bbox": [ + 46, + 545, + 289, + 713 + ], + "type": "text", + "content": " to yield a latent sequence " + }, + { + "bbox": [ + 46, + 545, + 289, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{V}_i" + }, + { + "bbox": [ + 46, + 545, + 289, + 713 + ], + "type": "text", + "content": " which has the same dimension as the feature sequences produced by trans" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 307, + 370, + 466, + 383 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 370, + 466, + 383 + ], + "spans": [ + { + "bbox": [ + 307, + 370, + 466, + 383 + ], + "type": "text", + "content": "Algorithm 1 Group exchange-masking" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 310, + 384, + 472, + 398 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 310, + 384, + 472, + 398 + ], + "spans": [ + { + "bbox": [ + 310, + 384, + 472, + 398 + ], + "type": "text", + "content": "Input: " + }, + { + "bbox": [ + 310, + 384, + 472, + 398 + ], + "type": "inline_equation", + "content": "\\mathcal{I}_1 = \\{I_1^n\\}_{n=1}^N" + }, + { + "bbox": [ + 310, + 384, + 472, + 398 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 310, + 384, + 472, + 398 + ], + "type": "inline_equation", + "content": "\\mathcal{Y}_1 = \\{\\mathbf{Y}_1^n\\}_{n=1}^N" + }, + { + "bbox": [ + 310, + 384, + 472, + 398 + ], + "type": "text", + "content": "," + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 341, + 398, + 472, + 410 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 341, + 398, + 472, + 410 + ], + "spans": [ + { + "bbox": [ + 341, + 398, + 472, + 410 + ], + "type": "interline_equation", + "content": "\\mathcal {I} _ {2} = \\left\\{I _ {2} ^ {n} \\right\\} _ {n = 1} ^ {N}, \\mathcal {Y} _ {2} = \\left\\{\\mathbf {Y} _ {2} ^ {n} \\right\\} _ {n = 1} ^ {N}.", + "image_path": "02250f6efb24344ff5a368b2b683d244b70bbfe6697c05c114df828416b386ac.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 310, + 410, + 425, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 310, + 410, + 425, + 422 + ], + "spans": [ + { + "bbox": [ + 310, + 410, + 425, + 422 + ], + "type": "text", + "content": "Output: " + }, + { + "bbox": [ + 310, + 410, + 425, + 422 + ], + "type": "inline_equation", + "content": "\\mathcal{I}_{e1},\\mathcal{Y}_{e1},\\mathcal{I}_{e2},\\mathcal{Y}_{e2}" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 311, + 423, + 545, + 447 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 423, + 545, + 447 + ], + "spans": [ + { + "bbox": [ + 311, + 423, + 545, + 447 + ], + "type": "text", + "content": "1: " + }, + { + "bbox": [ + 311, + 423, + 545, + 447 + ], + "type": "inline_equation", + "content": "\\mathcal{I}_{e1},\\mathcal{Y}_{e1},\\mathcal{I}_{e2},\\mathcal{Y}_{e2}\\gets" + }, + { + "bbox": [ + 311, + 423, + 545, + 447 + ], + "type": "text", + "content": " Rearranging " + }, + { + "bbox": [ + 311, + 423, + 545, + 447 + ], + "type": "inline_equation", + "content": "\\mathcal{I}_1,\\mathcal{Y}_1,\\mathcal{I}_2,\\mathcal{Y}_2" + }, + { + "bbox": [ + 311, + 423, + 545, + 447 + ], + "type": "text", + "content": " via solving (1)." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 311, + 448, + 425, + 505 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 312, + 448, + 416, + 458 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 448, + 416, + 458 + ], + "spans": [ + { + "bbox": [ + 312, + 448, + 416, + 458 + ], + "type": "text", + "content": "2: for " + }, + { + "bbox": [ + 312, + 448, + 416, + 458 + ], + "type": "inline_equation", + "content": "n = 1, 2, \\ldots, k" + }, + { + "bbox": [ + 312, + 448, + 416, + 458 + ], + "type": "text", + "content": " do" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 312, + 459, + 420, + 471 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 459, + 420, + 471 + ], + "spans": [ + { + "bbox": [ + 312, + 459, + 420, + 471 + ], + "type": "text", + "content": "3: " + }, + { + "bbox": [ + 312, + 459, + 420, + 471 + ], + "type": "inline_equation", + "content": "I_{e1}^{n} \\gets I_{2}^{n}, I_{e2}^{n} \\gets I_{1}^{n}" + }, + { + "bbox": [ + 312, + 459, + 420, + 471 + ], + "type": "text", + "content": "," + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 312, + 471, + 425, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 471, + 425, + 483 + ], + "spans": [ + { + "bbox": [ + 312, + 471, + 425, + 483 + ], + "type": "text", + "content": "4: " + }, + { + "bbox": [ + 312, + 471, + 425, + 483 + ], + "type": "inline_equation", + "content": "\\mathbf{Y}_{e1}^{n}\\gets \\mathbf{0}\\in \\mathbb{R}^{H\\times W\\times 1}," + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 312, + 483, + 425, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 483, + 425, + 495 + ], + "spans": [ + { + "bbox": [ + 312, + 483, + 425, + 495 + ], + "type": "text", + "content": "5: " + }, + { + "bbox": [ + 312, + 483, + 425, + 495 + ], + "type": "inline_equation", + "content": "\\mathbf{Y}_{e2}^{\\vec{n}}\\gets \\mathbf{0}\\in \\mathbb{R}^{H\\times W\\times 1}" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 312, + 497, + 356, + 505 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 497, + 356, + 505 + ], + "spans": [ + { + "bbox": [ + 312, + 497, + 356, + 505 + ], + "type": "text", + "content": "6: end for" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 304, + 529, + 545, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 529, + 545, + 628 + ], + "spans": [ + { + "bbox": [ + 304, + 529, + 545, + 628 + ], + "type": "text", + "content": "formers. Meanwhile, the image groups are passed through the CoSOD-TB with shared weights, producing the token sequences " + }, + { + "bbox": [ + 304, + 529, + 545, + 628 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_i" + }, + { + "bbox": [ + 304, + 529, + 545, + 628 + ], + "type": "text", + "content": " which capture group consensus and long-range dependency information. At last, " + }, + { + "bbox": [ + 304, + 529, + 545, + 628 + ], + "type": "inline_equation", + "content": "\\mathcal{V}_i" + }, + { + "bbox": [ + 304, + 529, + 545, + 628 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 529, + 545, + 628 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_i" + }, + { + "bbox": [ + 304, + 529, + 545, + 628 + ], + "type": "text", + "content": " are concatenated in channel dimension and passed through several subsequent processes as shown in Figure 2 followed by an up-sampling layer, yielding the corresponding predicted cosaliency maps " + }, + { + "bbox": [ + 304, + 529, + 545, + 628 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{O}} = \\{\\hat{\\mathbf{O}}^n\\in \\mathbb{R}^{H\\times W}\\}_{n = 1}^{2N}" + }, + { + "bbox": [ + 304, + 529, + 545, + 628 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 635, + 453, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 635, + 453, + 647 + ], + "spans": [ + { + "bbox": [ + 306, + 635, + 453, + 647 + ], + "type": "text", + "content": "3.1. Group Exchange-Masking" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "type": "text", + "content": "The widely-used group consensus assumption limits the robustness of the CoSOD model, especially when there are test images without common objects in the group. To this end, we reorganize the two input groups with the GEM tailored to CoSOD. The GEM is summarised in Algorithm 1:" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 734, + 317, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 734, + 317, + 742 + ], + "spans": [ + { + "bbox": [ + 293, + 734, + 317, + 742 + ], + "type": "text", + "content": "19641" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 48, + 72, + 289, + 178 + ], + "blocks": [ + { + "bbox": [ + 48, + 72, + 289, + 178 + ], + "lines": [ + { + "bbox": [ + 48, + 72, + 289, + 178 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 289, + 178 + ], + "type": "image", + "image_path": "7c477db33a779d751b18b96233bdfb2f1ffbecf1bd6510b8e32065d3b31e6119.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 100, + 189, + 233, + 200 + ], + "lines": [ + { + "bbox": [ + 100, + 189, + 233, + 200 + ], + "spans": [ + { + "bbox": [ + 100, + 189, + 233, + 200 + ], + "type": "text", + "content": "Figure 3. Architecture of the LVGB." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 220, + 287, + 327 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 220, + 287, + 327 + ], + "spans": [ + { + "bbox": [ + 46, + 220, + 287, + 327 + ], + "type": "text", + "content": "taking " + }, + { + "bbox": [ + 46, + 220, + 287, + 327 + ], + "type": "inline_equation", + "content": "\\mathcal{I}_i" + }, + { + "bbox": [ + 46, + 220, + 287, + 327 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 220, + 287, + 327 + ], + "type": "inline_equation", + "content": "\\mathcal{Y}_i" + }, + { + "bbox": [ + 46, + 220, + 287, + 327 + ], + "type": "text", + "content": " as input, the goal of GEM is to obtain " + }, + { + "bbox": [ + 46, + 220, + 287, + 327 + ], + "type": "inline_equation", + "content": "\\mathcal{I}_{ei}" + }, + { + "bbox": [ + 46, + 220, + 287, + 327 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 220, + 287, + 327 + ], + "type": "inline_equation", + "content": "\\mathcal{Y}_{ei}" + }, + { + "bbox": [ + 46, + 220, + 287, + 327 + ], + "type": "text", + "content": " which contain noisy images in each group and corresponding full masked ground truths. Specifically, we rearrange the two groups of images and their binary ground-truth masks in a decreasing order by solving (1). After this, we sample the top-" + }, + { + "bbox": [ + 46, + 220, + 287, + 327 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 220, + 287, + 327 + ], + "type": "text", + "content": " images of each group as noisy images and exchange them in each other's group, ensuring the maximum destructiveness of noisy images and the effectiveness of training." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 327, + 287, + 352 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 327, + 287, + 352 + ], + "spans": [ + { + "bbox": [ + 47, + 327, + 287, + 352 + ], + "type": "text", + "content": "This training process is essentially a min-max optimization process [26] as" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 75, + 360, + 287, + 393 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 360, + 287, + 393 + ], + "spans": [ + { + "bbox": [ + 75, + 360, + 287, + 393 + ], + "type": "interline_equation", + "content": "\\min _ {\\varphi} \\sum_ {n = 1} ^ {k} \\sum_ {i = 1} ^ {2} \\max _ {\\mathcal {I} _ {i} (n,:)} \\mathcal {L} \\left(f _ {\\varphi} \\left(\\mathcal {I} _ {i} (n,:)\\right), \\hat {\\mathcal {O}} _ {i} (n,:)\\right), \\tag {1}", + "image_path": "dec3e1e6785677c0e769c8b8e40c1b7156f541aa28459f443ab08c3109639cc7.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 400, + 287, + 485 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 400, + 287, + 485 + ], + "spans": [ + { + "bbox": [ + 46, + 400, + 287, + 485 + ], + "type": "text", + "content": "where the loss " + }, + { + "bbox": [ + 46, + 400, + 287, + 485 + ], + "type": "inline_equation", + "content": "\\mathcal{L}" + }, + { + "bbox": [ + 46, + 400, + 287, + 485 + ], + "type": "text", + "content": " is defined by (7), and " + }, + { + "bbox": [ + 46, + 400, + 287, + 485 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 46, + 400, + 287, + 485 + ], + "type": "text", + "content": " denotes our whole model with learnable parameters " + }, + { + "bbox": [ + 46, + 400, + 287, + 485 + ], + "type": "inline_equation", + "content": "\\varphi" + }, + { + "bbox": [ + 46, + 400, + 287, + 485 + ], + "type": "text", + "content": ". Solving (1) finds out the most noisy images which can maximize the training loss, and our goal is to minimize the loss function w.r.t the intra-group noisy images to improve the model's robustness. At last, we mask the corresponding labels of the noisy images with the all-zeros maps." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 491, + 195, + 504 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 491, + 195, + 504 + ], + "spans": [ + { + "bbox": [ + 47, + 491, + 195, + 504 + ], + "type": "text", + "content": "3.2. Latent Variable Generator" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 510, + 287, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 510, + 287, + 605 + ], + "spans": [ + { + "bbox": [ + 46, + 510, + 287, + 605 + ], + "type": "text", + "content": "By training the model with noisy images, the GEM is able to increase the model's robustness which can identify whether there exist co-objects or not. However, the model will focus on non-co-salient regions as a result of the uncertainty introduced by adding noisy images to the training image groups. This means that the model will become overconfident in the background regions in the early stage of training, leading to inaccurate prediction." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 605, + 287, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 287, + 689 + ], + "type": "text", + "content": "To address this issue, we further propose the LVGB that generates a low-dimensional latent space to represent the most common patterns in image groups [17]. The latent variables sampled from the feature space characterize uncertainties of group consensus and can be used to modulate the intermediate features from other branches to highlight the co-objects." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "content": "Figure 3 shows the details of our LVGB. The generator takes " + }, + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{I}_{ei}" + }, + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "content": " and latent variable " + }, + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_i" + }, + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "content": " as input, and uses an encoder" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 72, + 545, + 145 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 545, + 145 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 545, + 145 + ], + "type": "text", + "content": "to produce stochastic prediction " + }, + { + "bbox": [ + 305, + 72, + 545, + 145 + ], + "type": "inline_equation", + "content": "P_{\\omega}(\\mathcal{Y}_{ei}|\\mathcal{I}_{ei},\\mathbf{z}_i)" + }, + { + "bbox": [ + 305, + 72, + 545, + 145 + ], + "type": "text", + "content": " with learnable parameters " + }, + { + "bbox": [ + 305, + 72, + 545, + 145 + ], + "type": "inline_equation", + "content": "\\omega" + }, + { + "bbox": [ + 305, + 72, + 545, + 145 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 305, + 72, + 545, + 145 + ], + "type": "inline_equation", + "content": "\\mathcal{I}_{ei}" + }, + { + "bbox": [ + 305, + 72, + 545, + 145 + ], + "type": "text", + "content": " is conditional variable, and " + }, + { + "bbox": [ + 305, + 72, + 545, + 145 + ], + "type": "inline_equation", + "content": "\\mathcal{Y}_{ei}" + }, + { + "bbox": [ + 305, + 72, + 545, + 145 + ], + "type": "text", + "content": " is output variable. " + }, + { + "bbox": [ + 305, + 72, + 545, + 145 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_i" + }, + { + "bbox": [ + 305, + 72, + 545, + 145 + ], + "type": "text", + "content": " is drawn from the Gaussian distribution " + }, + { + "bbox": [ + 305, + 72, + 545, + 145 + ], + "type": "inline_equation", + "content": "P_{\\theta}(\\mathbf{z}_i|\\mathcal{I}_{ei})" + }, + { + "bbox": [ + 305, + 72, + 545, + 145 + ], + "type": "text", + "content": ", and the posterior of " + }, + { + "bbox": [ + 305, + 72, + 545, + 145 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_i" + }, + { + "bbox": [ + 305, + 72, + 545, + 145 + ], + "type": "text", + "content": " is formulated as " + }, + { + "bbox": [ + 305, + 72, + 545, + 145 + ], + "type": "inline_equation", + "content": "Q_{\\phi}(\\mathbf{z}_i|\\mathcal{I}_{ei},\\mathcal{Y}_{ei})" + }, + { + "bbox": [ + 305, + 72, + 545, + 145 + ], + "type": "text", + "content": " via an encoder that maps the input variable to the latent space. The loss of the LVG is defined as [32,43]" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 315, + 152, + 545, + 211 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 152, + 545, + 211 + ], + "spans": [ + { + "bbox": [ + 315, + 152, + 545, + 211 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} _ {L V G} = \\sum_ {i = 1} ^ {2} E _ {\\mathbf {z} \\sim Q _ {\\phi} (\\mathbf {z} _ {i} | \\mathcal {I} _ {e i}, \\mathcal {Y} _ {e i})} [ - \\log P _ {\\boldsymbol {\\omega}} (\\mathcal {Y} _ {e i} | \\mathcal {I} _ {e i}, \\mathbf {z} _ {i}) ] \\\\ + \\mathrm {K L} \\left(Q _ {\\phi} \\left(\\mathbf {z} _ {i} \\mid \\mathcal {I} _ {e i}, \\mathcal {Y} _ {e i}\\right) \\mid \\mid P _ {\\boldsymbol {\\theta}} \\left(\\mathbf {z} _ {i} \\mid \\mathcal {I} _ {e i}\\right)\\right), \\tag {2} \\\\ \\end{array}", + "image_path": "52d3151ba7800f4f72643353f42e663bfece80509551b97efb2b66818c7e62f5.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 212, + 545, + 234 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 212, + 545, + 234 + ], + "spans": [ + { + "bbox": [ + 305, + 212, + 545, + 234 + ], + "type": "text", + "content": "where the Kullback-Leibler Divergence metric KL is defined as" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 355, + 241, + 545, + 268 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 355, + 241, + 545, + 268 + ], + "spans": [ + { + "bbox": [ + 355, + 241, + 545, + 268 + ], + "type": "interline_equation", + "content": "\\operatorname {K L} (Q | | P) = \\sum Q (x) \\log \\frac {Q (x)}{P (x)}. \\tag {3}", + "image_path": "ce5cd9eab0886942f61639dc6961045d77d5dd92de2f48a7e67c982c2c319904.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 275, + 545, + 478 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 275, + 545, + 478 + ], + "spans": [ + { + "bbox": [ + 305, + 275, + 545, + 478 + ], + "type": "text", + "content": "In (2), the prior net " + }, + { + "bbox": [ + 305, + 275, + 545, + 478 + ], + "type": "inline_equation", + "content": "P_{\\pmb{\\theta}}(\\mathbf{z}_i|\\mathcal{I}_{ei})" + }, + { + "bbox": [ + 305, + 275, + 545, + 478 + ], + "type": "text", + "content": " is defined as a Gaussian distribution that maps the image group " + }, + { + "bbox": [ + 305, + 275, + 545, + 478 + ], + "type": "inline_equation", + "content": "\\mathcal{I}_{ei}" + }, + { + "bbox": [ + 305, + 275, + 545, + 478 + ], + "type": "text", + "content": " to a low-dimensional latent space that encodes the most common patterns in each group. We randomly sample prior latent variable " + }, + { + "bbox": [ + 305, + 275, + 545, + 478 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_i\\sim \\mathcal{N}(\\pmb{\\mu}_{1i},\\mathrm{diag}(\\pmb{\\sigma}_{1i}^2))" + }, + { + "bbox": [ + 305, + 275, + 545, + 478 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 305, + 275, + 545, + 478 + ], + "type": "inline_equation", + "content": "\\pmb{\\mu}_{1i},\\pmb{\\sigma}_{1i}^{2}\\in \\mathbb{R}^{K}" + }, + { + "bbox": [ + 305, + 275, + 545, + 478 + ], + "type": "text", + "content": " represent the mean and standard deviation vectors and " + }, + { + "bbox": [ + 305, + 275, + 545, + 478 + ], + "type": "inline_equation", + "content": "\\pmb{\\theta}" + }, + { + "bbox": [ + 305, + 275, + 545, + 478 + ], + "type": "text", + "content": " is the learnable parameter of the mapping function which consists five conv layers [43] as shown in Figure 3. In the posterior net, we use the similar encoder as the prior net to model " + }, + { + "bbox": [ + 305, + 275, + 545, + 478 + ], + "type": "inline_equation", + "content": "Q_{\\phi}(\\mathbf{z}_i|\\mathcal{I}_{ei},\\mathcal{Y}_{ei})" + }, + { + "bbox": [ + 305, + 275, + 545, + 478 + ], + "type": "text", + "content": " that maps the concatenation of " + }, + { + "bbox": [ + 305, + 275, + 545, + 478 + ], + "type": "inline_equation", + "content": "\\mathcal{I}_{ei}" + }, + { + "bbox": [ + 305, + 275, + 545, + 478 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 275, + 545, + 478 + ], + "type": "inline_equation", + "content": "\\mathcal{Y}_{ei}" + }, + { + "bbox": [ + 305, + 275, + 545, + 478 + ], + "type": "text", + "content": " to the posterior latent variable " + }, + { + "bbox": [ + 305, + 275, + 545, + 478 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_i\\sim \\mathcal{N}(\\pmb{\\mu}_{2i},\\mathrm{diag}(\\pmb{\\sigma}_{2i}^2))" + }, + { + "bbox": [ + 305, + 275, + 545, + 478 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 305, + 275, + 545, + 478 + ], + "type": "inline_equation", + "content": "\\pmb{\\mu}_{2i},\\pmb{\\sigma}_{2i}^{2}\\in \\mathbb{R}^{K}" + }, + { + "bbox": [ + 305, + 275, + 545, + 478 + ], + "type": "text", + "content": ". The KL in (2) is used to measure the difference between probability distributions prior " + }, + { + "bbox": [ + 305, + 275, + 545, + 478 + ], + "type": "inline_equation", + "content": "P_{\\pmb{\\theta}}(\\mathbf{z}_i|\\mathcal{I}_{ei})" + }, + { + "bbox": [ + 305, + 275, + 545, + 478 + ], + "type": "text", + "content": " and posterior " + }, + { + "bbox": [ + 305, + 275, + 545, + 478 + ], + "type": "inline_equation", + "content": "Q_{\\phi}(\\mathbf{z}_i|\\mathcal{I}_{ei},\\mathcal{Y}_{ei})" + }, + { + "bbox": [ + 305, + 275, + 545, + 478 + ], + "type": "text", + "content": ". It represents the information loss that occurs when using " + }, + { + "bbox": [ + 305, + 275, + 545, + 478 + ], + "type": "inline_equation", + "content": "Q_{\\phi}(\\mathbf{z}_i|\\mathcal{I}_{ei},\\mathcal{Y}_{ei})" + }, + { + "bbox": [ + 305, + 275, + 545, + 478 + ], + "type": "text", + "content": " to approximate " + }, + { + "bbox": [ + 305, + 275, + 545, + 478 + ], + "type": "inline_equation", + "content": "P_{\\pmb{\\theta}}(\\mathbf{z}_i|\\mathcal{I}_{ei})" + }, + { + "bbox": [ + 305, + 275, + 545, + 478 + ], + "type": "text", + "content": ", and a smaller value indicates greater similarity between the two distributions." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 479, + 545, + 622 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 479, + 545, + 622 + ], + "spans": [ + { + "bbox": [ + 304, + 479, + 545, + 622 + ], + "type": "text", + "content": "Each position in latent space symbolizes potential labeling changes or other potential factors that could lead to a variety of co-saliency predictions [17, 43]. The proposed GEM can meet the needs for diverse ground truths as we learn to select the most difficult samples to mask their ground truths to all-zeros maps. We hope diverse annotations in the posterior net " + }, + { + "bbox": [ + 304, + 479, + 545, + 622 + ], + "type": "inline_equation", + "content": "Q_{\\phi}(\\mathbf{z}_i|\\mathcal{I}_{ei},\\mathcal{Y}_{ei})" + }, + { + "bbox": [ + 304, + 479, + 545, + 622 + ], + "type": "text", + "content": " during training can compel the prior net " + }, + { + "bbox": [ + 304, + 479, + 545, + 622 + ], + "type": "inline_equation", + "content": "P_{\\theta}(\\mathbf{z}_i|\\mathcal{I}_{ei})" + }, + { + "bbox": [ + 304, + 479, + 545, + 622 + ], + "type": "text", + "content": " to encode labeling variants of the supplied inputs " + }, + { + "bbox": [ + 304, + 479, + 545, + 622 + ], + "type": "inline_equation", + "content": "\\mathcal{I}_{ei}" + }, + { + "bbox": [ + 304, + 479, + 545, + 622 + ], + "type": "text", + "content": ". The statistics " + }, + { + "bbox": [ + 304, + 479, + 545, + 622 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_i" + }, + { + "bbox": [ + 304, + 479, + 545, + 622 + ], + "type": "text", + "content": " is further processed by feature expanding [43]. Given a pair of " + }, + { + "bbox": [ + 304, + 479, + 545, + 622 + ], + "type": "inline_equation", + "content": "(\\boldsymbol{\\mu}_k,\\boldsymbol{\\sigma}_k)" + }, + { + "bbox": [ + 304, + 479, + 545, + 622 + ], + "type": "text", + "content": " in each position of " + }, + { + "bbox": [ + 304, + 479, + 545, + 622 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 304, + 479, + 545, + 622 + ], + "type": "text", + "content": " dimensional vector, we parameterize them to obtain the latent vector" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 383, + 632, + 545, + 645 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 383, + 632, + 545, + 645 + ], + "spans": [ + { + "bbox": [ + 383, + 632, + 545, + 645 + ], + "type": "interline_equation", + "content": "\\mathbf {z} _ {k} = \\boldsymbol {\\sigma} _ {k} \\odot \\boldsymbol {\\epsilon} + \\boldsymbol {\\mu} _ {k}, \\tag {4}", + "image_path": "1b3e26363a279c278ac680466b36781712aa3de5ab4dbdd8c0a8e6346b90df9f.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 305, + 654, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 654, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 654, + 545, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 654, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 305, + 654, + 545, + 713 + ], + "type": "text", + "content": " follows the standard normal distribution " + }, + { + "bbox": [ + 305, + 654, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\epsilon \\sim \\mathcal{N}(0,1)" + }, + { + "bbox": [ + 305, + 654, + 545, + 713 + ], + "type": "text", + "content": ". To fuse with the features " + }, + { + "bbox": [ + 305, + 654, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_i" + }, + { + "bbox": [ + 305, + 654, + 545, + 713 + ], + "type": "text", + "content": " from the CoSOD-TB, as shown in Figure 3, we expand " + }, + { + "bbox": [ + 305, + 654, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_k" + }, + { + "bbox": [ + 305, + 654, + 545, + 713 + ], + "type": "text", + "content": " to the feature sequence which has the same size as " + }, + { + "bbox": [ + 305, + 654, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_i" + }, + { + "bbox": [ + 305, + 654, + 545, + 713 + ], + "type": "text", + "content": ", yielding the stochastic feature " + }, + { + "bbox": [ + 305, + 654, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{V}_i" + }, + { + "bbox": [ + 305, + 654, + 545, + 713 + ], + "type": "text", + "content": ". The visualization effect of " + }, + { + "bbox": [ + 305, + 654, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{V}_i" + }, + { + "bbox": [ + 305, + 654, + 545, + 713 + ], + "type": "text", + "content": " is partly shown in Fig-" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 732, + 317, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 732, + 317, + 742 + ], + "spans": [ + { + "bbox": [ + 294, + 732, + 317, + 742 + ], + "type": "text", + "content": "19642" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 48, + 70, + 289, + 157 + ], + "blocks": [ + { + "bbox": [ + 48, + 70, + 289, + 157 + ], + "lines": [ + { + "bbox": [ + 48, + 70, + 289, + 157 + ], + "spans": [ + { + "bbox": [ + 48, + 70, + 289, + 157 + ], + "type": "image", + "image_path": "669de5bedfbe1b9bbc55320728ec41581aa49c7de768d2a3bb2dd9bb401c1643.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 90, + 166, + 243, + 177 + ], + "lines": [ + { + "bbox": [ + 90, + 166, + 243, + 177 + ], + "spans": [ + { + "bbox": [ + 90, + 166, + 243, + 177 + ], + "type": "text", + "content": "Figure 4. Architecture of the CoSOD-TB." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 199, + 287, + 235 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 199, + 287, + 235 + ], + "spans": [ + { + "bbox": [ + 46, + 199, + 287, + 235 + ], + "type": "text", + "content": "ure 5, as we can see, these features encode rich group consistency information that can well highlight the co-salient regions." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 244, + 168, + 257 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 244, + 168, + 257 + ], + "spans": [ + { + "bbox": [ + 47, + 244, + 168, + 257 + ], + "type": "text", + "content": "3.3. CoSOD Transformer" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 263, + 287, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 263, + 287, + 312 + ], + "spans": [ + { + "bbox": [ + 46, + 263, + 287, + 312 + ], + "type": "text", + "content": "After getting the stochastic features " + }, + { + "bbox": [ + 46, + 263, + 287, + 312 + ], + "type": "inline_equation", + "content": "\\nu_{i}" + }, + { + "bbox": [ + 46, + 263, + 287, + 312 + ], + "type": "text", + "content": " from the LVGB, we leverage these features to guide the learning of generating the general features " + }, + { + "bbox": [ + 46, + 263, + 287, + 312 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_i" + }, + { + "bbox": [ + 46, + 263, + 287, + 312 + ], + "type": "text", + "content": " from the CoSOD-TB to focus on the co-salient regions (see the bottom row of Figure 5)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 312, + 287, + 661 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 312, + 287, + 661 + ], + "spans": [ + { + "bbox": [ + 46, + 312, + 287, + 661 + ], + "type": "text", + "content": "Figure 4 shows the architecture of the proposed CoSOD-TB. Each input image " + }, + { + "bbox": [ + 46, + 312, + 287, + 661 + ], + "type": "inline_equation", + "content": "I_{i}^{n} \\in \\mathcal{I}_{ei}" + }, + { + "bbox": [ + 46, + 312, + 287, + 661 + ], + "type": "text", + "content": " is cropped into " + }, + { + "bbox": [ + 46, + 312, + 287, + 661 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 46, + 312, + 287, + 661 + ], + "type": "text", + "content": " patches with size " + }, + { + "bbox": [ + 46, + 312, + 287, + 661 + ], + "type": "inline_equation", + "content": "H / 4 \\times W / 4" + }, + { + "bbox": [ + 46, + 312, + 287, + 661 + ], + "type": "text", + "content": ", and then all the image patches in each group construct the token sequences " + }, + { + "bbox": [ + 46, + 312, + 287, + 661 + ], + "type": "inline_equation", + "content": "\\mathcal{X}_i^0 \\in \\mathbb{R}^{N \\times \\frac{H}{4} \\times \\frac{W}{4} \\times 3d}" + }, + { + "bbox": [ + 46, + 312, + 287, + 661 + ], + "type": "text", + "content": ". Then, " + }, + { + "bbox": [ + 46, + 312, + 287, + 661 + ], + "type": "inline_equation", + "content": "\\mathcal{X}_i^0" + }, + { + "bbox": [ + 46, + 312, + 287, + 661 + ], + "type": "text", + "content": " is sent into the backbone of the transformer with a T2T architecture [40], obtaining the token sequences " + }, + { + "bbox": [ + 46, + 312, + 287, + 661 + ], + "type": "inline_equation", + "content": "\\mathcal{X}_i^1 \\in \\mathbb{R}^{N \\times \\frac{H}{16} \\times \\frac{W}{16} \\times c}" + }, + { + "bbox": [ + 46, + 312, + 287, + 661 + ], + "type": "text", + "content": " which encode both local and global information. After this, we design the group token " + }, + { + "bbox": [ + 46, + 312, + 287, + 661 + ], + "type": "inline_equation", + "content": "\\mathcal{X}^G \\in \\mathbb{R}^{N \\times 1 \\times c}" + }, + { + "bbox": [ + 46, + 312, + 287, + 661 + ], + "type": "text", + "content": " to capture group-wise common information that is essential to extract co-objects, and the saliency token " + }, + { + "bbox": [ + 46, + 312, + 287, + 661 + ], + "type": "inline_equation", + "content": "\\mathcal{X}^S \\in \\mathbb{R}^{N \\times 1 \\times c}" + }, + { + "bbox": [ + 46, + 312, + 287, + 661 + ], + "type": "text", + "content": " to capture the specific information that encodes more co-object structure details. Afterwards, " + }, + { + "bbox": [ + 46, + 312, + 287, + 661 + ], + "type": "inline_equation", + "content": "\\mathcal{X}^G" + }, + { + "bbox": [ + 46, + 312, + 287, + 661 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 312, + 287, + 661 + ], + "type": "inline_equation", + "content": "\\mathcal{X}^1" + }, + { + "bbox": [ + 46, + 312, + 287, + 661 + ], + "type": "text", + "content": " are concatenated in the channel dimension and fed into the Transformer Layer as the design in ViT [4] to get the new token sequence. We further use an multi-layer perceptron (MLP) layer to integrate global information into " + }, + { + "bbox": [ + 46, + 312, + 287, + 661 + ], + "type": "inline_equation", + "content": "\\mathcal{X}^G" + }, + { + "bbox": [ + 46, + 312, + 287, + 661 + ], + "type": "text", + "content": ". We fuse the " + }, + { + "bbox": [ + 46, + 312, + 287, + 661 + ], + "type": "inline_equation", + "content": "\\mathcal{X}^S" + }, + { + "bbox": [ + 46, + 312, + 287, + 661 + ], + "type": "text", + "content": " to integrate saliency information into sequences and split the " + }, + { + "bbox": [ + 46, + 312, + 287, + 661 + ], + "type": "inline_equation", + "content": "\\mathcal{X}^G" + }, + { + "bbox": [ + 46, + 312, + 287, + 661 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 312, + 287, + 661 + ], + "type": "inline_equation", + "content": "\\mathcal{X}^S" + }, + { + "bbox": [ + 46, + 312, + 287, + 661 + ], + "type": "text", + "content": " to up-sample the token sequence and get the token sequence " + }, + { + "bbox": [ + 46, + 312, + 287, + 661 + ], + "type": "inline_equation", + "content": "\\mathcal{X}^2" + }, + { + "bbox": [ + 46, + 312, + 287, + 661 + ], + "type": "text", + "content": ". After undergoing some processing as shown in Figure 4, " + }, + { + "bbox": [ + 46, + 312, + 287, + 661 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_i" + }, + { + "bbox": [ + 46, + 312, + 287, + 661 + ], + "type": "text", + "content": " is then concatenated with the stochastic features " + }, + { + "bbox": [ + 46, + 312, + 287, + 661 + ], + "type": "inline_equation", + "content": "\\mathcal{V}_i" + }, + { + "bbox": [ + 46, + 312, + 287, + 661 + ], + "type": "text", + "content": " followed by a decoder that includes two Transformer Layers and an MLP layer. During decoding, due to the use of stochastic features, the model is more possible to overcome the prejudices formed in the early training. As shown in the third row of Figure 5, the CoSOD-TB can integrate " + }, + { + "bbox": [ + 46, + 312, + 287, + 661 + ], + "type": "inline_equation", + "content": "\\mathcal{V}_i" + }, + { + "bbox": [ + 46, + 312, + 287, + 661 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 312, + 287, + 661 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_i" + }, + { + "bbox": [ + 46, + 312, + 287, + 661 + ], + "type": "text", + "content": " well, helping the model focus on more reliable co-salient object regions. Finally, the output sequences are reshaped to produce the predicted co-saliency maps " + }, + { + "bbox": [ + 46, + 312, + 287, + 661 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{O}}" + }, + { + "bbox": [ + 46, + 312, + 287, + 661 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 59, + 662, + 259, + 673 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 662, + 259, + 673 + ], + "spans": [ + { + "bbox": [ + 59, + 662, + 259, + 673 + ], + "type": "text", + "content": "The loss function of the CoSOD-TB is defined as" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 683, + 287, + 716 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 683, + 287, + 716 + ], + "spans": [ + { + "bbox": [ + 55, + 683, + 287, + 716 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {T R A N S} = \\frac {1}{2 N} \\sum_ {n = 1} ^ {N} \\sum_ {i = 1} ^ {2} \\ell_ {B C E} \\left(\\mathcal {Y} _ {e i} (n,:), \\hat {\\mathcal {O}} _ {i} (n,:)\\right), \\tag {5}", + "image_path": "cebcf822c47e798fda839673d8fc2f5bc471881de4dbbeb96e1635b511e3df5d.jpg" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 309, + 72, + 362, + 109 + ], + "blocks": [ + { + "bbox": [ + 309, + 72, + 362, + 109 + ], + "lines": [ + { + "bbox": [ + 309, + 72, + 362, + 109 + ], + "spans": [ + { + "bbox": [ + 309, + 72, + 362, + 109 + ], + "type": "image", + "image_path": "d6c019f4ac60ca461c8d3cfa98d78d068e0b4a53949b07943107ed80b1d12da8.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 310, + 110, + 362, + 155 + ], + "blocks": [ + { + "bbox": [ + 310, + 110, + 362, + 155 + ], + "lines": [ + { + "bbox": [ + 310, + 110, + 362, + 155 + ], + "spans": [ + { + "bbox": [ + 310, + 110, + 362, + 155 + ], + "type": "image", + "image_path": "d4d641740597727ff39488d3d9ad453f69f56ef2e2955afa833acc4bfe5bdbb6.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 328, + 155, + 342, + 164 + ], + "lines": [ + { + "bbox": [ + 328, + 155, + 342, + 164 + ], + "spans": [ + { + "bbox": [ + 328, + 155, + 342, + 164 + ], + "type": "text", + "content": "V" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 365, + 72, + 419, + 109 + ], + "blocks": [ + { + "bbox": [ + 365, + 72, + 419, + 109 + ], + "lines": [ + { + "bbox": [ + 365, + 72, + 419, + 109 + ], + "spans": [ + { + "bbox": [ + 365, + 72, + 419, + 109 + ], + "type": "image", + "image_path": "cb8608e47752fd79e1e9882da09ef7f81477508803ef2786099b76293ac9c958.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 422, + 72, + 476, + 109 + ], + "blocks": [ + { + "bbox": [ + 422, + 72, + 476, + 109 + ], + "lines": [ + { + "bbox": [ + 422, + 72, + 476, + 109 + ], + "spans": [ + { + "bbox": [ + 422, + 72, + 476, + 109 + ], + "type": "image", + "image_path": "124f6d34d2e8ab093b7373f739fab125f59dca7bc4c5a77ff4cbc0264c1b2ee1.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 425, + 110, + 476, + 118 + ], + "lines": [ + { + "bbox": [ + 425, + 110, + 476, + 118 + ], + "spans": [ + { + "bbox": [ + 425, + 110, + 476, + 118 + ], + "type": "text", + "content": "Watermelon_2" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 479, + 72, + 533, + 109 + ], + "blocks": [ + { + "bbox": [ + 479, + 72, + 533, + 109 + ], + "lines": [ + { + "bbox": [ + 479, + 72, + 533, + 109 + ], + "spans": [ + { + "bbox": [ + 479, + 72, + 533, + 109 + ], + "type": "image", + "image_path": "b7147fc66c9abfc49b9649e6c16b391411ce5b53c35cbdbe07fab06424ba15c6.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 494, + 110, + 516, + 118 + ], + "lines": [ + { + "bbox": [ + 494, + 110, + 516, + 118 + ], + "spans": [ + { + "bbox": [ + 494, + 110, + 516, + 118 + ], + "type": "text", + "content": "F" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 309, + 119, + 362, + 155 + ], + "blocks": [ + { + "bbox": [ + 309, + 119, + 362, + 155 + ], + "lines": [ + { + "bbox": [ + 309, + 119, + 362, + 155 + ], + "spans": [ + { + "bbox": [ + 309, + 119, + 362, + 155 + ], + "type": "image", + "image_path": "b1e0fa65a5d649d1604930a3813f64fe7d0dd798010061490e716100294cb3d3.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 365, + 119, + 419, + 155 + ], + "blocks": [ + { + "bbox": [ + 367, + 110, + 396, + 119 + ], + "lines": [ + { + "bbox": [ + 367, + 110, + 396, + 119 + ], + "spans": [ + { + "bbox": [ + 367, + 110, + 396, + 119 + ], + "type": "text", + "content": "F" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 365, + 119, + 419, + 155 + ], + "lines": [ + { + "bbox": [ + 365, + 119, + 419, + 155 + ], + "spans": [ + { + "bbox": [ + 365, + 119, + 419, + 155 + ], + "type": "image", + "image_path": "d1b1a274ba14bd665d087b38eccbfb94a9239c7786bf1eb7e9145ac4567a7e8c.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 384, + 155, + 396, + 164 + ], + "lines": [ + { + "bbox": [ + 384, + 155, + 396, + 164 + ], + "spans": [ + { + "bbox": [ + 384, + 155, + 396, + 164 + ], + "type": "text", + "content": "V" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 422, + 119, + 476, + 155 + ], + "blocks": [ + { + "bbox": [ + 422, + 119, + 476, + 155 + ], + "lines": [ + { + "bbox": [ + 422, + 119, + 476, + 155 + ], + "spans": [ + { + "bbox": [ + 422, + 119, + 476, + 155 + ], + "type": "image", + "image_path": "8beadddbfde6357610eea1203d2dff9b9710a533484077f218e47a01485f18b6.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 428, + 155, + 458, + 164 + ], + "lines": [ + { + "bbox": [ + 428, + 155, + 458, + 164 + ], + "spans": [ + { + "bbox": [ + 428, + 155, + 458, + 164 + ], + "type": "text", + "content": "V2-1" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 479, + 119, + 533, + 155 + ], + "blocks": [ + { + "bbox": [ + 479, + 119, + 533, + 155 + ], + "lines": [ + { + "bbox": [ + 479, + 119, + 533, + 155 + ], + "spans": [ + { + "bbox": [ + 479, + 119, + 533, + 155 + ], + "type": "image", + "image_path": "ab7f5f382d81f5218b6d9e49207827580a92de728e545fa866f39e53f75f7b83.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 500, + 155, + 517, + 164 + ], + "lines": [ + { + "bbox": [ + 500, + 155, + 517, + 164 + ], + "spans": [ + { + "bbox": [ + 500, + 155, + 517, + 164 + ], + "type": "text", + "content": "V2-2" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 309, + 165, + 362, + 200 + ], + "blocks": [ + { + "bbox": [ + 309, + 165, + 362, + 200 + ], + "lines": [ + { + "bbox": [ + 309, + 165, + 362, + 200 + ], + "spans": [ + { + "bbox": [ + 309, + 165, + 362, + 200 + ], + "type": "image", + "image_path": "a2d31cdfb33e17e2f8ea6334113ea29fe6aed812f8759f471aef85f46460a783.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 320, + 201, + 354, + 211 + ], + "lines": [ + { + "bbox": [ + 320, + 201, + 354, + 211 + ], + "spans": [ + { + "bbox": [ + 320, + 201, + 354, + 211 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_i^1" + }, + { + "bbox": [ + 320, + 201, + 354, + 211 + ], + "type": "text", + "content": " & " + }, + { + "bbox": [ + 320, + 201, + 354, + 211 + ], + "type": "inline_equation", + "content": "\\mathcal{V}_i^{1 - 1}" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 365, + 165, + 419, + 200 + ], + "blocks": [ + { + "bbox": [ + 365, + 165, + 419, + 200 + ], + "lines": [ + { + "bbox": [ + 365, + 165, + 419, + 200 + ], + "spans": [ + { + "bbox": [ + 365, + 165, + 419, + 200 + ], + "type": "image", + "image_path": "1a7138bfd942a6f1341a68be263627baf31060264d3cdb9b4659f72ad21e651c.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 378, + 201, + 412, + 211 + ], + "lines": [ + { + "bbox": [ + 378, + 201, + 412, + 211 + ], + "spans": [ + { + "bbox": [ + 378, + 201, + 412, + 211 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_i^1" + }, + { + "bbox": [ + 378, + 201, + 412, + 211 + ], + "type": "text", + "content": " & " + }, + { + "bbox": [ + 378, + 201, + 412, + 211 + ], + "type": "inline_equation", + "content": "\\nu_{i}^{1 - 2}" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_caption" + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 422, + 165, + 476, + 200 + ], + "blocks": [ + { + "bbox": [ + 422, + 165, + 476, + 200 + ], + "lines": [ + { + "bbox": [ + 422, + 165, + 476, + 200 + ], + "spans": [ + { + "bbox": [ + 422, + 165, + 476, + 200 + ], + "type": "image", + "image_path": "e21aa753c738f59337856cd23dbc34a6595e6bc1e805dd0695f85f19bf622bae.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 434, + 201, + 469, + 211 + ], + "lines": [ + { + "bbox": [ + 434, + 201, + 469, + 211 + ], + "spans": [ + { + "bbox": [ + 434, + 201, + 469, + 211 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_i^2\\& \\mathcal{V}_i^{2 - 1}" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_caption" + } + ], + "index": 28 + }, + { + "type": "image", + "bbox": [ + 479, + 165, + 533, + 200 + ], + "blocks": [ + { + "bbox": [ + 479, + 165, + 533, + 200 + ], + "lines": [ + { + "bbox": [ + 479, + 165, + 533, + 200 + ], + "spans": [ + { + "bbox": [ + 479, + 165, + 533, + 200 + ], + "type": "image", + "image_path": "e4c4e261fb78a898daf990052e6f879f3ea7537a35314cf71b96742f2f7bbaf7.jpg" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 488, + 201, + 523, + 211 + ], + "lines": [ + { + "bbox": [ + 488, + 201, + 523, + 211 + ], + "spans": [ + { + "bbox": [ + 488, + 201, + 523, + 211 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_i^2" + }, + { + "bbox": [ + 488, + 201, + 523, + 211 + ], + "type": "text", + "content": " & " + }, + { + "bbox": [ + 488, + 201, + 523, + 211 + ], + "type": "inline_equation", + "content": "\\mathcal{V}_i^{2 - 2}" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_caption" + } + ], + "index": 30 + }, + { + "type": "image", + "bbox": [ + 538, + 72, + 556, + 202 + ], + "blocks": [ + { + "bbox": [ + 538, + 72, + 556, + 202 + ], + "lines": [ + { + "bbox": [ + 538, + 72, + 556, + 202 + ], + "spans": [ + { + "bbox": [ + 538, + 72, + 556, + 202 + ], + "type": "image", + "image_path": "66c2176887f164ea6c566eab77593af273de301961cdd4dde98720b6dce642b3.jpg" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 224, + 546, + 278 + ], + "lines": [ + { + "bbox": [ + 304, + 224, + 546, + 278 + ], + "spans": [ + { + "bbox": [ + 304, + 224, + 546, + 278 + ], + "type": "text", + "content": "Figure 5. Comparison of effects of the \"watermelon\" group before and after fusing stochastic features. The first row: visualizations of the images and features obtained by the CoSOD-TB; The second row: visualizations of the features generated by the LVGB; The third row: visualizations of the integrated features." + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_caption" + } + ], + "index": 32 + }, + { + "bbox": [ + 304, + 297, + 545, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 297, + 545, + 319 + ], + "spans": [ + { + "bbox": [ + 304, + 297, + 545, + 319 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 297, + 545, + 319 + ], + "type": "inline_equation", + "content": "\\ell_{BCE}" + }, + { + "bbox": [ + 304, + 297, + 545, + 319 + ], + "type": "text", + "content": " is a binary cross-entropy (BCE) loss [27] defined as" + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 315, + 327, + 544, + 358 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 327, + 544, + 358 + ], + "spans": [ + { + "bbox": [ + 315, + 327, + 544, + 358 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\ell_ {B C E} (\\mathcal {Y} (n,:), \\hat {\\mathcal {O}} (n,:)) = - (\\mathcal {Y} (n,:) ^ {\\top} \\log (\\hat {\\mathcal {O}} (n,:)) \\tag {6} \\\\ - \\left(1 - \\mathcal {Y} (n,:)\\right) ^ {\\top} \\log \\left(1 - \\tilde {\\mathcal {O}} (n,:)\\right)). \\\\ \\end{array}", + "image_path": "ce21f3a47a096b0601f7e5eedfce1163aadba06b794848e66977db967b690c23.jpg" + } + ] + } + ], + "index": 35 + }, + { + "bbox": [ + 306, + 378, + 394, + 389 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 378, + 394, + 389 + ], + "spans": [ + { + "bbox": [ + 306, + 378, + 394, + 389 + ], + "type": "text", + "content": "3.4. Loss Function" + } + ] + } + ], + "index": 36 + }, + { + "bbox": [ + 304, + 396, + 545, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 396, + 545, + 430 + ], + "spans": [ + { + "bbox": [ + 304, + 396, + 545, + 430 + ], + "type": "text", + "content": "The LVG and the CoSOD-TB are jointly trained in an end-to-end manner by optimizing the following multi-task loss" + } + ] + } + ], + "index": 37 + }, + { + "bbox": [ + 365, + 433, + 545, + 445 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 365, + 433, + 545, + 445 + ], + "spans": [ + { + "bbox": [ + 365, + 433, + 545, + 445 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\lambda_ {1} \\mathcal {L} _ {L V G} + \\lambda_ {2} \\mathcal {L} _ {T R A N S}, \\tag {7}", + "image_path": "c99793cb1a3113207b9a28dea3e0dbaf0b2eb0a9f3777409c5802da25d785d4b.jpg" + } + ] + } + ], + "index": 38 + }, + { + "bbox": [ + 305, + 449, + 542, + 462 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 449, + 542, + 462 + ], + "spans": [ + { + "bbox": [ + 305, + 449, + 542, + 462 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 449, + 542, + 462 + ], + "type": "inline_equation", + "content": "\\lambda_{1},\\lambda_{2}" + }, + { + "bbox": [ + 305, + 449, + 542, + 462 + ], + "type": "text", + "content": " are the hyperparameters to balance each loss." + } + ] + } + ], + "index": 39 + }, + { + "bbox": [ + 305, + 472, + 431, + 485 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 472, + 431, + 485 + ], + "spans": [ + { + "bbox": [ + 305, + 472, + 431, + 485 + ], + "type": "text", + "content": "4. Experimental Results" + } + ] + } + ], + "index": 40 + }, + { + "bbox": [ + 305, + 491, + 438, + 504 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 491, + 438, + 504 + ], + "spans": [ + { + "bbox": [ + 305, + 491, + 438, + 504 + ], + "type": "text", + "content": "4.1. Implementation Details" + } + ] + } + ], + "index": 41 + }, + { + "bbox": [ + 304, + 510, + 545, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 510, + 545, + 629 + ], + "spans": [ + { + "bbox": [ + 304, + 510, + 545, + 629 + ], + "type": "text", + "content": "Our model is implemented under the PyTorch1.9.0 framework [27]. Our computing platform's acceleration is provided by a GeForce GTX 2080Ti GPU. The transformer backbone is the pre-trained T2T-ViT " + }, + { + "bbox": [ + 304, + 510, + 545, + 629 + ], + "type": "inline_equation", + "content": "t - 14" + }, + { + "bbox": [ + 304, + 510, + 545, + 629 + ], + "type": "text", + "content": " [40] model since it has a similar computational complexity as CNNs-based ResNet50 [11] which is smaller than the VGG-16 [31] widely used in CoSOD. The training dataset are COCOSEG released by [33] and DUTS released by [34]. They totally include about 208,250 images from 369 categories, and the relevant binary ground-truth masks are provided." + } + ] + } + ], + "index": 42 + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "content": "In each training episode, we randomly select two groups, which are processed by the proposed GEM with " + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "inline_equation", + "content": "k = 1" + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "content": ". Each group contains " + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "inline_equation", + "content": "N = 5" + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "content": " images. The hyperparameters in (7) are set to " + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\lambda_1 = 0.25" + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\lambda_2 = 1" + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "content": ". We resize the images to " + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "inline_equation", + "content": "224 \\times 224 \\times 3" + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "content": " and use them as input. We train the network over 60,000 steps totally and the training process takes about 11 hours. During the training process, we use" + } + ] + } + ], + "index": 43 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 734, + 317, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 734, + 317, + 742 + ], + "spans": [ + { + "bbox": [ + 294, + 734, + 317, + 742 + ], + "type": "text", + "content": "19643" + } + ] + } + ], + "index": 44 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 73, + 69, + 520, + 240 + ], + "blocks": [ + { + "bbox": [ + 73, + 69, + 520, + 240 + ], + "lines": [ + { + "bbox": [ + 73, + 69, + 520, + 240 + ], + "spans": [ + { + "bbox": [ + 73, + 69, + 520, + 240 + ], + "type": "image", + "image_path": "5adf55c23f5906c200b6b7e773e66a611ef4a37658406c17f8f6e2c3bb2cad26.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 253, + 547, + 275 + ], + "lines": [ + { + "bbox": [ + 46, + 253, + 547, + 275 + ], + "spans": [ + { + "bbox": [ + 46, + 253, + 547, + 275 + ], + "type": "text", + "content": "Figure 6. The qualitative comparisons with the recent state-of-the-art methods, including DCFM [39], CADC [49], DeepACG [45] and GCoNet [9]." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 283, + 287, + 331 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 283, + 287, + 331 + ], + "spans": [ + { + "bbox": [ + 46, + 283, + 287, + 331 + ], + "type": "text", + "content": "the Adam algorithm [16] to optimize the whole network. The initial learning rate is set to " + }, + { + "bbox": [ + 46, + 283, + 287, + 331 + ], + "type": "inline_equation", + "content": "10e - 4" + }, + { + "bbox": [ + 46, + 283, + 287, + 331 + ], + "type": "text", + "content": " and decay to " + }, + { + "bbox": [ + 46, + 283, + 287, + 331 + ], + "type": "inline_equation", + "content": "10e - 5" + }, + { + "bbox": [ + 46, + 283, + 287, + 331 + ], + "type": "text", + "content": " at the 20,000-th epoch. The inference time is 31 fps in average, meeting the real-time application requirements." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 349, + 222, + 361 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 349, + 222, + 361 + ], + "spans": [ + { + "bbox": [ + 47, + 349, + 222, + 361 + ], + "type": "text", + "content": "4.2. Datasets and Evaluation Metrics" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 371, + 287, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 371, + 287, + 491 + ], + "spans": [ + { + "bbox": [ + 46, + 371, + 287, + 491 + ], + "type": "text", + "content": "The test sets include CoSal2015 [42], CoCA [50] and CoSOD3k [8]. Among them, CoSal2015 contains 2,015 images from 50 categories, and the objects therein have different sizes and shapes, which makes CoSal2015 challenging; CoCA is the newest dataset, including 1,297 images from 80 categories. These images have extremely different styles and very complicated backgrounds, making it a very difficult data set; CoSOD3k contains 3,316 images form 160 categories, which is by far the largest test set. It contains more disturbing targets, making it very challenging." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 494, + 287, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 494, + 287, + 567 + ], + "spans": [ + { + "bbox": [ + 46, + 494, + 287, + 567 + ], + "type": "text", + "content": "Four evaluation metrics are used for comparison, including MAE [33], " + }, + { + "bbox": [ + 46, + 494, + 287, + 567 + ], + "type": "inline_equation", + "content": "E_{\\phi}^{max}" + }, + { + "bbox": [ + 46, + 494, + 287, + 567 + ], + "type": "text", + "content": " [6], " + }, + { + "bbox": [ + 46, + 494, + 287, + 567 + ], + "type": "inline_equation", + "content": "S_{\\alpha}" + }, + { + "bbox": [ + 46, + 494, + 287, + 567 + ], + "type": "text", + "content": " [5] and " + }, + { + "bbox": [ + 46, + 494, + 287, + 567 + ], + "type": "inline_equation", + "content": "F_{\\beta}^{max}" + }, + { + "bbox": [ + 46, + 494, + 287, + 567 + ], + "type": "text", + "content": " [1], which are to assess the average pixel-wise absolute difference, local and global similarity, structural similarity between the predictions and the ground truths, and the weighted harmonic mean of precision and recall, respectively." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 583, + 274, + 597 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 583, + 274, + 597 + ], + "spans": [ + { + "bbox": [ + 47, + 583, + 274, + 597 + ], + "type": "text", + "content": "4.3. Comparisons with State-of-the-art Methods" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 605, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 287, + 715 + ], + "type": "text", + "content": "On the basis of unified evaluation codes [7] for fair comparison, we compare our method with several state-of-the-art methods published in recent three years, including RCAN [20], CSMG [46], SSNM [44], GCAGC [47], GICD [50], ICNet [15], GCoNet [9], DeepACG [45], CoEG-Net [7], HrSSMN [48], CADC [49], DCFM [39]. Among them, CADC [49] and DCFM [39] are the latest cutting-edge methods. More experimental results can be found in the supplementary materials." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 283, + 547, + 668 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 283, + 547, + 668 + ], + "spans": [ + { + "bbox": [ + 304, + 283, + 547, + 668 + ], + "type": "text", + "content": "Qualitative Results. Figure 6 shows some visual comparison results with four latest state-of-the-art methods, including DCFM, CADC, DeepACG and GCoNet. The selected four groups are very challenging and suffer from highly confusing interference co-objects, extremely complex backgrounds, and drastic scale changes. Specifically, in the group \"Baseball\", the co-salient object is very small and the interference objects have similar shapes. Because all other methods are trapped in overconfidence, obsessed with the wrong results, and unable to extricate themselves. They cannot accurately detect the tiny objects and the background is wrongly segmented. In the group \"Backpack\", benefiting from training with noisy samples, our model is more robust that can accurately locate co-salient objects and segment target details, while other methods DCFM, CADC and GCoNet misjudge the co-salient objects. Camouflage co-objects in complex backgrounds are generally acknowledged as difficult samples [18], and the co-salient items in the \"Gift box\" group are perfectly in place. Thanks to CoSOD-TB's ability to capture long-range dependencies, our model learns strong feature representation that is able to accurately segment the camouflage co-salient objects, while CADC and GCoNet cannot finely segment the objects, and the other methods even fail in detecting the targets. The group \"teddy bear\" is also very challenging, and the co-salient objects in it suffer from interference objects and backgrounds with similar colors. Besides, the objects also have very different sizes. In this case, our model still achieves promising results, demonstrating its strong robustness to a variety of challenging factors. However, other methods yield a large number of inaccurate segmentation regions." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 689, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 689, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 689, + 547, + 713 + ], + "type": "text", + "content": "Quantitative Results. Figure 7 shows the PR and F-measure curves of our method and other state-of-the-art" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 734, + 318, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 734, + 318, + 742 + ], + "spans": [ + { + "bbox": [ + 294, + 734, + 318, + 742 + ], + "type": "text", + "content": "19644" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 69, + 217, + 199 + ], + "blocks": [ + { + "bbox": [ + 58, + 69, + 217, + 199 + ], + "lines": [ + { + "bbox": [ + 58, + 69, + 217, + 199 + ], + "spans": [ + { + "bbox": [ + 58, + 69, + 217, + 199 + ], + "type": "image", + "image_path": "a6e745f2961ddb5612fb833de3595d57838f3f4118b6c08d3a166e04ceb4eba5.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 217, + 68, + 376, + 199 + ], + "blocks": [ + { + "bbox": [ + 217, + 68, + 376, + 199 + ], + "lines": [ + { + "bbox": [ + 217, + 68, + 376, + 199 + ], + "spans": [ + { + "bbox": [ + 217, + 68, + 376, + 199 + ], + "type": "image", + "image_path": "99edc3e7150f9a04bdb9dc5e06abba64ec881c3e45bb05246a32c4dc9e1b4999.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 377, + 68, + 536, + 199 + ], + "blocks": [ + { + "bbox": [ + 377, + 68, + 536, + 199 + ], + "lines": [ + { + "bbox": [ + 377, + 68, + 536, + 199 + ], + "spans": [ + { + "bbox": [ + 377, + 68, + 536, + 199 + ], + "type": "image", + "image_path": "520c16fb7a0153f745e34bf8b8931bd9da222d2d9f43802bfe3db202108cd973.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 57, + 202, + 217, + 322 + ], + "blocks": [ + { + "bbox": [ + 57, + 202, + 217, + 322 + ], + "lines": [ + { + "bbox": [ + 57, + 202, + 217, + 322 + ], + "spans": [ + { + "bbox": [ + 57, + 202, + 217, + 322 + ], + "type": "image", + "image_path": "7bedf32491bcaba9436116a59b84f70b9f46a5d1acdce0420a4888440240dd33.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 334, + 546, + 355 + ], + "lines": [ + { + "bbox": [ + 46, + 334, + 546, + 355 + ], + "spans": [ + { + "bbox": [ + 46, + 334, + 546, + 355 + ], + "type": "text", + "content": "Figure 7. Comparisons with state-of-the-art methods since the year of 2019 in terms of PR and F-measure curves on three benchmark datasets." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 217, + 202, + 376, + 322 + ], + "blocks": [ + { + "bbox": [ + 217, + 202, + 376, + 322 + ], + "lines": [ + { + "bbox": [ + 217, + 202, + 376, + 322 + ], + "spans": [ + { + "bbox": [ + 217, + 202, + 376, + 322 + ], + "type": "image", + "image_path": "01981976420ccf7d7042e0193aa7e54e7cd345ad809be4748e1232ac25b5ff34.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 378, + 203, + 536, + 322 + ], + "blocks": [ + { + "bbox": [ + 378, + 203, + 536, + 322 + ], + "lines": [ + { + "bbox": [ + 378, + 203, + 536, + 322 + ], + "spans": [ + { + "bbox": [ + 378, + 203, + 536, + 322 + ], + "type": "image", + "image_path": "4051291ff930bb6c7daa9699aedd6930d9c5b19efd289e33baaa41fdb329785e.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 50, + 374, + 545, + 521 + ], + "blocks": [ + { + "bbox": [ + 119, + 363, + 473, + 373 + ], + "lines": [ + { + "bbox": [ + 119, + 363, + 473, + 373 + ], + "spans": [ + { + "bbox": [ + 119, + 363, + 473, + 373 + ], + "type": "text", + "content": "Table 1. Statistic comparisons of our model with the other state-of-the-arts since the year of 2019." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 50, + 374, + 545, + 521 + ], + "lines": [ + { + "bbox": [ + 50, + 374, + 545, + 521 + ], + "spans": [ + { + "bbox": [ + 50, + 374, + 545, + 521 + ], + "type": "table", + "html": "
MethodsCoSal2015CoSOD3kCoCA
MAE↓Sα↑Eφmax↑Fβmax↑MAE↓Sα↑Eφmax↑Fβmax↑MAE↓Sα↑Eφmax↑Fβmax↑
RCAN(IJCAI2019)0.1260.7790.8420.7640.1300.7440.8080.6880.1600.6160.7020.422
CSMG(CVPR2019)0.1300.7740.8180.7770.1570.7110.7230.6450.1240.6320.7340.503
SSNM(AAAI2020)0.1020.7880.8430.7940.1200.7260.7560.6750.1160.6280.7410.482
GCAGC(CVPR2020)0.0850.8170.8660.8130.1000.7850.8160.7400.1180.6690.7540.523
GICD(ECCV2020)0.0720.8420.8840.8340.0890.7940.8310.7430.1250.6580.7010.504
ICNet(NIPS2020)0.0580.8570.9000.8580.0890.7940.8450.7620.1470.6540.7050.514
CoEGNet(TPAMI2021)0.0770.8360.8820.8320.0920.7620.8250.7360.1060.6120.7170.493
GCoNet(CVPR2021)0.0690.8450.8870.8470.0710.8020.8600.7500.1050.6730.7600.524
DeepACG(CVPR2021)0.0660.8530.8930.8470.0790.8110.8590.7790.1040.6850.7590.564
CADC(ICCV2021)0.0640.8660.9060.8620.0960.8010.8400.7590.1320.6810.7440.548
HrSSMN(TMM2022)0.0620.8450.8950.8410.0870.7880.8420.7530.1060.6710.7390.532
DCFM(CVPR2022)0.0670.8380.8920.8560.0670.8100.8740.8050.0850.7100.7830.598
OURS0.0530.8850.9330.8820.0610.8530.9110.8290.0950.7260.8080.599
", + "image_path": "136123f8483aa28270822f20d726b808ff61c099e28dfc303496138d062b9fb5.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 542, + 287, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 542, + 287, + 590 + ], + "spans": [ + { + "bbox": [ + 46, + 542, + 287, + 590 + ], + "type": "text", + "content": "methods. As we can see, the PR curves produced by our method wrap around the curves generated by other methods. Also, all the F-measure curves of our method are above those generated by other methods." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 594, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 288, + 715 + ], + "type": "text", + "content": "Meanwhile, we list the statistic comparison results of all compared methods in Table 1. It is clear that our approach achieves the most competitive result compared to its counterparts. Specifically, on the CoSal2015 dataset, our method achieves the best scores of 0.053, 0.885, 0.933, and 0.882 in terms of all metrics, with a significant gain of " + }, + { + "bbox": [ + 46, + 594, + 288, + 715 + ], + "type": "inline_equation", + "content": "0.5\\%" + }, + { + "bbox": [ + 46, + 594, + 288, + 715 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 594, + 288, + 715 + ], + "type": "inline_equation", + "content": "1.9\\%" + }, + { + "bbox": [ + 46, + 594, + 288, + 715 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 594, + 288, + 715 + ], + "type": "inline_equation", + "content": "2.7\\%" + }, + { + "bbox": [ + 46, + 594, + 288, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 594, + 288, + 715 + ], + "type": "inline_equation", + "content": "2.0\\%" + }, + { + "bbox": [ + 46, + 594, + 288, + 715 + ], + "type": "text", + "content": ", respectively, compared to the second-best preforming ICNet [15] and CADC [49]. On the most recent and difficult datasets, CoSOD3k and CoCA, our method also achieves the best performance. Especially" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 542, + 545, + 615 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 542, + 545, + 615 + ], + "spans": [ + { + "bbox": [ + 305, + 542, + 545, + 615 + ], + "type": "text", + "content": "on CoSOD3k, our method reaches the best scores of 0.061, 0.853, 0.911 and 0.829 in terms of " + }, + { + "bbox": [ + 305, + 542, + 545, + 615 + ], + "type": "inline_equation", + "content": "MAE" + }, + { + "bbox": [ + 305, + 542, + 545, + 615 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 305, + 542, + 545, + 615 + ], + "type": "inline_equation", + "content": "E_{\\phi}^{max}" + }, + { + "bbox": [ + 305, + 542, + 545, + 615 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 305, + 542, + 545, + 615 + ], + "type": "inline_equation", + "content": "S_{\\alpha}" + }, + { + "bbox": [ + 305, + 542, + 545, + 615 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 542, + 545, + 615 + ], + "type": "inline_equation", + "content": "F_{\\beta}^{max}" + }, + { + "bbox": [ + 305, + 542, + 545, + 615 + ], + "type": "text", + "content": ", respectively, with a great gain of " + }, + { + "bbox": [ + 305, + 542, + 545, + 615 + ], + "type": "inline_equation", + "content": "0.6\\%" + }, + { + "bbox": [ + 305, + 542, + 545, + 615 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 305, + 542, + 545, + 615 + ], + "type": "inline_equation", + "content": "4.3\\%" + }, + { + "bbox": [ + 305, + 542, + 545, + 615 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 305, + 542, + 545, + 615 + ], + "type": "inline_equation", + "content": "3.7\\%" + }, + { + "bbox": [ + 305, + 542, + 545, + 615 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 542, + 545, + 615 + ], + "type": "inline_equation", + "content": "2.4\\%" + }, + { + "bbox": [ + 305, + 542, + 545, + 615 + ], + "type": "text", + "content": " compared to the second-best performing DCFM [39]. The effectiveness of our method has been fully validated on the large-scale and challenging datasets." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 623, + 400, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 623, + 400, + 635 + ], + "spans": [ + { + "bbox": [ + 306, + 623, + 400, + 635 + ], + "type": "text", + "content": "4.4. Ablation Study" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 641, + 546, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 641, + 546, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 641, + 546, + 715 + ], + "type": "text", + "content": "To verify the effectiveness of our designs, we conduct the ablation study of our method on three datasets. We have made a lot of changes to VST [23] and regard the modified version as the baseline. Table 2 lists the results, and we can observe that every key design in our model makes a significant contribution to the overall performance. Taking" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 734, + 317, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 734, + 317, + 742 + ], + "spans": [ + { + "bbox": [ + 294, + 734, + 317, + 742 + ], + "type": "text", + "content": "19645" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 50, + 79, + 547, + 175 + ], + "blocks": [ + { + "bbox": [ + 128, + 69, + 463, + 79 + ], + "lines": [ + { + "bbox": [ + 128, + 69, + 463, + 79 + ], + "spans": [ + { + "bbox": [ + 128, + 69, + 463, + 79 + ], + "type": "text", + "content": "Table 2. Ablations of our method on the effectiveness of the GEM , LVGB, and CoSOD-TB." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 50, + 79, + 547, + 175 + ], + "lines": [ + { + "bbox": [ + 50, + 79, + 547, + 175 + ], + "spans": [ + { + "bbox": [ + 50, + 79, + 547, + 175 + ], + "type": "table", + "html": "
StrategiesCoSal2015CoSOD3kCoCA
GEMLVGBCoSOD-TBMAE ↓\\( S_{\\alpha} \\uparrow \\)\\( E_{\\phi}^{max} \\uparrow \\)\\( F_{\\beta}^{max} \\uparrow \\)MAE ↓\\( S_{\\alpha} \\uparrow \\)\\( E_{\\phi}^{max} \\uparrow \\)\\( F_{\\beta}^{max} \\uparrow \\)MAE ↓\\( S_{\\alpha} \\uparrow \\)\\( E_{\\phi}^{max} \\uparrow \\)\\( F_{\\beta}^{max} \\uparrow \\)
0.0600.8540.8870.8600.0750.7870.8630.7780.1050.7100.7850.564
✓0.0610.8830.9280.8770.0650.8340.8800.8170.1090.7100.7880.571
✓0.0580.8770.9260.8720.0630.8400.8710.8050.1000.7190.7890.574
✓0.0530.8740.8930.8800.0620.8420.9000.8140.0980.7160.7980.583
✓✓0.0550.8690.9300.8760.0630.8470.9060.8190.1040.7240.8020.597
✓✓0.0600.8860.9250.8720.0690.8500.8950.8230.1000.7180.7920.587
✓✓0.0540.8800.9190.8780.0610.8490.8890.8290.0960.7200.8050.595
✓✓✓0.0530.8850.9330.8820.0610.8530.9110.8290.0950.7260.8080.599
", + "image_path": "c23b576430664ccda45ace2cfbd4fd8b35dc78da64f5414f260df75c29bec77f.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 50, + 203, + 286, + 258 + ], + "blocks": [ + { + "bbox": [ + 47, + 182, + 286, + 202 + ], + "lines": [ + { + "bbox": [ + 47, + 182, + 286, + 202 + ], + "spans": [ + { + "bbox": [ + 47, + 182, + 286, + 202 + ], + "type": "text", + "content": "Table 3. Complexity analyses. The number of inputs is set to 5 to maintain unity." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 50, + 203, + 286, + 258 + ], + "lines": [ + { + "bbox": [ + 50, + 203, + 286, + 258 + ], + "spans": [ + { + "bbox": [ + 50, + 203, + 286, + 258 + ], + "type": "table", + "html": "
methodsFLOPs(G)↓param.(M)↓runtime(fps)↑Fβmax↑
GICD(ECCV20)364.7278.040.80.504
GCoNet(CVPR21)259.9142.0116.20.524
CADC(ICCV21)330.0392.818.00.548
DCFM(CVPR22)251.9142.384.40.598
OURS211.852.331.00.599
", + "image_path": "37e499642ec966a5aedf7adc854d385a3fcd23e6f6b1c4ece0e712283928f47f.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 50, + 271, + 279, + 376 + ], + "blocks": [ + { + "bbox": [ + 50, + 271, + 279, + 376 + ], + "lines": [ + { + "bbox": [ + 50, + 271, + 279, + 376 + ], + "spans": [ + { + "bbox": [ + 50, + 271, + 279, + 376 + ], + "type": "image", + "image_path": "1c635b45307c1fbe73aa2cbccc55df733e7ad7676be5a92ef6eb938f076349a2.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 388, + 286, + 410 + ], + "lines": [ + { + "bbox": [ + 46, + 388, + 286, + 410 + ], + "spans": [ + { + "bbox": [ + 46, + 388, + 286, + 410 + ], + "type": "text", + "content": "Figure 8. Comparison with the current most competitive method DCFM [39] in simulated real scenes." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 420, + 286, + 600 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 420, + 286, + 600 + ], + "spans": [ + { + "bbox": [ + 46, + 420, + 286, + 600 + ], + "type": "text", + "content": "the results in CoSal2015 as examples, without GEM, the performance of our model gets worse in terms of all metrics especially for the " + }, + { + "bbox": [ + 46, + 420, + 286, + 600 + ], + "type": "inline_equation", + "content": "S_{\\alpha}" + }, + { + "bbox": [ + 46, + 420, + 286, + 600 + ], + "type": "text", + "content": " that drops from 0.885 to 0.869 by " + }, + { + "bbox": [ + 46, + 420, + 286, + 600 + ], + "type": "inline_equation", + "content": "1.6\\%" + }, + { + "bbox": [ + 46, + 420, + 286, + 600 + ], + "type": "text", + "content": ". With the GEM alone, the performance of our model improves greatly, and especially " + }, + { + "bbox": [ + 46, + 420, + 286, + 600 + ], + "type": "inline_equation", + "content": "S_{\\alpha}" + }, + { + "bbox": [ + 46, + 420, + 286, + 600 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 420, + 286, + 600 + ], + "type": "inline_equation", + "content": "E_{\\phi}^{max}" + }, + { + "bbox": [ + 46, + 420, + 286, + 600 + ], + "type": "text", + "content": " have a gain of " + }, + { + "bbox": [ + 46, + 420, + 286, + 600 + ], + "type": "inline_equation", + "content": "2.9\\%" + }, + { + "bbox": [ + 46, + 420, + 286, + 600 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 420, + 286, + 600 + ], + "type": "inline_equation", + "content": "4.1\\%" + }, + { + "bbox": [ + 46, + 420, + 286, + 600 + ], + "type": "text", + "content": ", respectively. Also, LVGB achieves the same excellent effect as GEM, because the uncertainty features provided by the LVGB can help the model consider more possibilities to enhance the robustness and overcome the over-confidence to non-salient regions. The CoSOD-TB makes important contributions to fine segmentation of images. With the CoSOD-TB, the MAE gets " + }, + { + "bbox": [ + 46, + 420, + 286, + 600 + ], + "type": "inline_equation", + "content": "0.7\\%" + }, + { + "bbox": [ + 46, + 420, + 286, + 600 + ], + "type": "text", + "content": " better from 0.060 to 0.053. The results on CoSOD3k and CoCA show the same trend, proving the effectiveness of each design of our method." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 609, + 169, + 624 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 609, + 169, + 624 + ], + "spans": [ + { + "bbox": [ + 47, + 609, + 169, + 624 + ], + "type": "text", + "content": "4.5. Practical Application" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 629, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 629, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 629, + 286, + 713 + ], + "type": "text", + "content": "For the practicability of the algorithm, we conduct complexity comparisons with the same settings as DCFM [39]. The results can be seen in Table 3. Our method achieves both smallest FLOPs, parameters and best performance in terms of " + }, + { + "bbox": [ + 46, + 629, + 286, + 713 + ], + "type": "inline_equation", + "content": "F_{\\beta}^{max}" + }, + { + "bbox": [ + 46, + 629, + 286, + 713 + ], + "type": "text", + "content": ". Meanwhile, the runtime of our method is 31 fps, meeting the real-time requirements. This shows that our method is more suitable for deployment to various ap" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 185, + 378, + 196 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 185, + 378, + 196 + ], + "spans": [ + { + "bbox": [ + 306, + 185, + 378, + 196 + ], + "type": "text", + "content": "plication devices." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 198, + 545, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 198, + 545, + 354 + ], + "spans": [ + { + "bbox": [ + 304, + 198, + 545, + 354 + ], + "type": "text", + "content": "We further test the performance of the model in the real scenes. In the real-life scenarios, there may not always exist co-salient objects within the group of images [22]. Figure 8 shows the images in the simulated real scenes that may not always exist co-objects. In these cases, the model should segment the co-objects existing in most images, while outputting all-zero masks on images without co-objects. Our method is better than the most competitive method DCFM [39] in dealing with real situations. When dealing with the images without co-objects, our model produces all-zero maps as much as possible. It shows that our method is more practical than other methods. More experimental results can be found in the supplementary materials." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 369, + 378, + 380 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 369, + 378, + 380 + ], + "spans": [ + { + "bbox": [ + 306, + 369, + 378, + 380 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 391, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 391, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 391, + 545, + 713 + ], + "type": "text", + "content": "In this paper, we have discovered that the paradigm of group consensus assumption has reduced the model's robustness and practical application value when confronted with irrelevant images in groups. A CoSOD model learning framework that is distinct from the classic CoSOD model learning framework has been developed. First, the group exchange-masking strategy has been devised, which is capable of automatically selecting the most informative noisy images from two groups and exchanging them in each other's groups to assist the model in learning more robust representations encoded with rich group consensus information. Second, the latent variable generator branch has been created to provide uncertainly-based global stochastic features that can regulate intermediate features from other branches to focus on co-objects. Third, the CoSOD transformer branch has been created to capture correlation-based global characteristics that carry information about group consistency. These branches' features are concatenated and put into a transformer-based decoder, yielding high-quality co-saliency maps. Extensive evaluations with and without irrelevant images have demonstrated the superiority of our method over a variety of state-of-the-art methods. In the future, we plan to build a new dataset for real scenarios that can provide a platform to evaluate model robustness and practical value in a more reliable way. Further more, we will consider the potential application of our method in defending adversarial attacks [35]." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 734, + 317, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 734, + 317, + 742 + ], + "spans": [ + { + "bbox": [ + 294, + 734, + 317, + 742 + ], + "type": "text", + "content": "19646" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 714 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 146 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 146 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 146 + ], + "type": "text", + "content": "[1] Radhakrishna Achanta, Sheila Hemami, Francisco Estrada, and Sabine Susstrunk. Frequency-tuned salient region detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1597-1604. IEEE, 2009. 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 148, + 288, + 203 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 148, + 288, + 203 + ], + "spans": [ + { + "bbox": [ + 53, + 148, + 288, + 203 + ], + "type": "text", + "content": "[2] Kai-Yueh Chang, Tyng-Luh Liu, and Shang-Hong Lai. From co-saliency to co-segmentation: An efficient and fully unsupervised energy minimization model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2129-2136. IEEE, 2011. 1, 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 205, + 288, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 205, + 288, + 259 + ], + "spans": [ + { + "bbox": [ + 53, + 205, + 288, + 259 + ], + "type": "text", + "content": "[3] Paul Christiano, Zain Shah, Igor Mordatch, Jonas Schneider, Trevor Blackwell, Joshua Tobin, Pieter Abbeel, and Wojciech Zaremba. Transfer from simulation to real world through learning deep inverse dynamics model. arXiv preprint arXiv:1610.03518, 2016. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 261, + 287, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 261, + 287, + 327 + ], + "spans": [ + { + "bbox": [ + 53, + 261, + 287, + 327 + ], + "type": "text", + "content": "[4] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 328, + 287, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 328, + 287, + 373 + ], + "spans": [ + { + "bbox": [ + 53, + 328, + 287, + 373 + ], + "type": "text", + "content": "[5] Deng-Ping Fan, Ming-Ming Cheng, Yun Liu, Tao Li, and Ali Borji. Structure-measure: A new way to evaluate foreground maps. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4548-4557, 2017. 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 374, + 287, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 374, + 287, + 418 + ], + "spans": [ + { + "bbox": [ + 53, + 374, + 287, + 418 + ], + "type": "text", + "content": "[6] Deng-Ping Fan, Cheng Gong, Yang Cao, Bo Ren, MingMing Cheng, and Ali Borji. Enhanced-alignment measure for binary foreground map evaluation. arXiv preprint arXiv:1805.10421, 2018. 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 419, + 287, + 474 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 419, + 287, + 474 + ], + "spans": [ + { + "bbox": [ + 53, + 419, + 287, + 474 + ], + "type": "text", + "content": "[7] Deng-Ping Fan, Tengpeng Li, Zheng Lin, Ge-Peng Ji, Dingwen Zhang, Ming-Ming Cheng, Huazhu Fu, and Jianbing Shen. Re-thinking co-salient object detection. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2021. 2, 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 475, + 287, + 531 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 475, + 287, + 531 + ], + "spans": [ + { + "bbox": [ + 53, + 475, + 287, + 531 + ], + "type": "text", + "content": "[8] Deng-Ping Fan, Zheng Lin, Ge-Peng Ji, Dingwen Zhang, Huazhu Fu, and Ming-Ming Cheng. Taking a deeper look at co-salient object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2919-2929, 2020. 1, 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 533, + 287, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 533, + 287, + 588 + ], + "spans": [ + { + "bbox": [ + 53, + 533, + 287, + 588 + ], + "type": "text", + "content": "[9] Qi Fan, Deng-Ping Fan, Huazhu Fu, Chi-Keung Tang, Ling Shao, and Yu-Wing Tai. Group collaborative learning for cosalient object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12288–12298, 2021. 1, 2, 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 590, + 287, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 590, + 287, + 622 + ], + "spans": [ + { + "bbox": [ + 48, + 590, + 287, + 622 + ], + "type": "text", + "content": "[10] Huazhu Fu, Xiaochun Cao, and Zhuowen Tu. Cluster-based co-saliency detection. IEEE Transactions on Image Processing, 22(10):3766-3778, 2013. 1, 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 624, + 287, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 624, + 287, + 668 + ], + "spans": [ + { + "bbox": [ + 48, + 624, + 287, + 668 + ], + "type": "text", + "content": "[11] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 770-778, 2016. 5" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 669, + 287, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 669, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 48, + 669, + 287, + 714 + ], + "type": "text", + "content": "[12] Kuang-Jui Hsu, Yen-Yu Lin, and Yung-Yu Chuang. Coattention cnns for unsupervised object co-segmentation. In Proceedings of the International Joint Conference on Artificial Intelligence, pages 748-756, 2018. 2" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 714 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 307, + 73, + 547, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 547, + 127 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 547, + 127 + ], + "type": "text", + "content": "[13] David E Jacobs, Dan B Goldman, and Eli Shechtman. Cosaliency: Where people look when comparing images. In Proceedings of the 23nd Annual ACM Symposium on User Interface Software and Technology, pages 219-228, 2010. 1, 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 129, + 547, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 129, + 547, + 172 + ], + "spans": [ + { + "bbox": [ + 307, + 129, + 547, + 172 + ], + "type": "text", + "content": "[14] Koteswar Rao Jerripothula, Jianfei Cai, and Junsong Yuan. Cats: Co-saliency activated tracklet selection for video colocalization. In Proceedings of the European Conference on Computer Vision, pages 187-202. Springer, 2016. 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 173, + 547, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 173, + 547, + 216 + ], + "spans": [ + { + "bbox": [ + 307, + 173, + 547, + 216 + ], + "type": "text", + "content": "[15] Wen-Da Jin, Jun Xu, Ming-Ming Cheng, Yi Zhang, and Wei Guo. Icnet: Intra-saliency correlation network for co-saliency detection. Advances in Neural Information Processing Systems, 33:18749-18759, 2020. 1, 2, 6, 7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 217, + 547, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 217, + 547, + 248 + ], + "spans": [ + { + "bbox": [ + 307, + 217, + 547, + 248 + ], + "type": "text", + "content": "[16] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 249, + 547, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 249, + 547, + 315 + ], + "spans": [ + { + "bbox": [ + 307, + 249, + 547, + 315 + ], + "type": "text", + "content": "[17] Simon Kohl, Bernardino Romera-Paredes, Clemens Meyer, Jeffrey De Fauw, Joseph R Ledsam, Klaus Maier-Hein, SM Eslami, Danilo Jimenez Rezende, and Olaf Ronneberger. A probabilistic u-net for segmentation of ambiguous images. Advances in Neural Information Processing Systems, 31, 2018. 4" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 316, + 547, + 371 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 316, + 547, + 371 + ], + "spans": [ + { + "bbox": [ + 307, + 316, + 547, + 371 + ], + "type": "text", + "content": "[18] Aixuan Li, Jing Zhang, Yunqiu Lv, Bowen Liu, Tong Zhang, and Yuchao Dai. Uncertainty-aware joint salient object and camouflaged object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10071-10081, 2021. 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 372, + 547, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 372, + 547, + 415 + ], + "spans": [ + { + "bbox": [ + 307, + 372, + 547, + 415 + ], + "type": "text", + "content": "[19] Bo Li, Zhengxing Sun, and Yuqi Guo. Supervae: Superpixelwise variational autoencoder for salient object detection. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 33, pages 8569-8576, 2019. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 416, + 547, + 469 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 416, + 547, + 469 + ], + "spans": [ + { + "bbox": [ + 307, + 416, + 547, + 469 + ], + "type": "text", + "content": "[20] Bo Li, Zhengxing Sun, Lv Tang, Yunhan Sun, and Jinlong Shi. Detecting robust co-saliency with recurrent coattention neural network. In Proceedings of the International Joint Conference on Artificial Intelligence, volume 2, page 6, 2019. 1, 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 471, + 547, + 504 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 471, + 547, + 504 + ], + "spans": [ + { + "bbox": [ + 307, + 471, + 547, + 504 + ], + "type": "text", + "content": "[21] Yijun Li, Keren Fu, Zhi Liu, and Jie Yang. Efficient saliency-model-guided visual co-saliency detection. IEEE Signal Processing Letters, 22(5):588-592, 2014. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 504, + 547, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 504, + 547, + 536 + ], + "spans": [ + { + "bbox": [ + 307, + 504, + 547, + 536 + ], + "type": "text", + "content": "[22] Jiawei Liu, Jing Zhang, Kaihao Zhang, and Nick Barnes. Generalised co-salient object detection. arXiv preprint arXiv:2208.09668, 2022. 8" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 537, + 547, + 580 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 537, + 547, + 580 + ], + "spans": [ + { + "bbox": [ + 307, + 537, + 547, + 580 + ], + "type": "text", + "content": "[23] Nian Liu, Ni Zhang, Kaiyuan Wan, Ling Shao, and Junwei Han. Visual saliency transformer. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2021. 7" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 581, + 547, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 581, + 547, + 624 + ], + "spans": [ + { + "bbox": [ + 307, + 581, + 547, + 624 + ], + "type": "text", + "content": "[24] Xuanqing Liu, Minhao Cheng, Huan Zhang, and Cho-Jui Hsieh. Towards robust neural networks via random self-ensemble. In Proceedings of the European Conference on Computer Vision, pages 369–385, 2018. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 625, + 547, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 625, + 547, + 668 + ], + "spans": [ + { + "bbox": [ + 307, + 625, + 547, + 668 + ], + "type": "text", + "content": "[25] Xuanqing Liu, Yao Li, Chongruo Wu, and Cho-Jui Hsieh. Adv-bnn: Improved adversarial defense through robust bayesian neural network. arXiv preprint arXiv:1810.01279, 2018.3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 669, + 547, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 669, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 307, + 669, + 547, + 714 + ], + "type": "text", + "content": "[26] Aleksander Madry, Aleksandar Makelov, Ludwig Schmidt, Dimitris Tsipras, and Adrian Vladu. Towards deep learning models resistant to adversarial attacks. In International Conference on Learning Representations, 2018. 4" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 734, + 318, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 734, + 318, + 742 + ], + "spans": [ + { + "bbox": [ + 294, + 734, + 318, + 742 + ], + "type": "text", + "content": "19647" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 137 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 137 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 137 + ], + "type": "text", + "content": "[27] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. Advances in Neural Information Processing Systems, 32:8026-8037, 2019. 5" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 139, + 287, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 139, + 287, + 183 + ], + "spans": [ + { + "bbox": [ + 48, + 139, + 287, + 183 + ], + "type": "text", + "content": "[28] Lerrel Pinto, James Davidson, Rahul Sukthankar, and Abhi nav Gupta. Robust adversarial reinforcement learning. In International Conference on Machine Learning, pages 2817-2826. PMLR, 2017. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 184, + 287, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 184, + 287, + 227 + ], + "spans": [ + { + "bbox": [ + 48, + 184, + 287, + 227 + ], + "type": "text", + "content": "[29] Aravind Rajeswaran, Sarvjeet Ghotra, Balaraman Ravindran, and Sergey Levine. Epopt: Learning robust neural network policies using model ensembles. arXiv preprint arXiv:1610.01283, 2016. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 228, + 287, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 228, + 287, + 271 + ], + "spans": [ + { + "bbox": [ + 48, + 228, + 287, + 271 + ], + "type": "text", + "content": "[30] Andrei A Rusu, Matej Večerík, Thomas Rothörl, Nicolas Heess, Razvan Pascanu, and Raia Hadsell. Sim-to-real robot learning from pixels with progressive nets. In Conference on robot learning, pages 262-270. PMLR, 2017. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 272, + 287, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 272, + 287, + 304 + ], + "spans": [ + { + "bbox": [ + 48, + 272, + 287, + 304 + ], + "type": "text", + "content": "[31] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556, 2014. 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 305, + 287, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 305, + 287, + 348 + ], + "spans": [ + { + "bbox": [ + 48, + 305, + 287, + 348 + ], + "type": "text", + "content": "[32] Kihyuk Sohn, Honglak Lee, and Xinchen Yan. Learning structured output representation using deep conditional generative models. Advances in Neural Information Processing Systems, 28, 2015. 2, 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 350, + 287, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 350, + 287, + 392 + ], + "spans": [ + { + "bbox": [ + 48, + 350, + 287, + 392 + ], + "type": "text", + "content": "[33] Chong Wang, Zheng-Jun Zha, Dong Liu, and Hongtao Xie. Robust deep co-saliency detection with group semantic. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 33, pages 8917-8924, 2019. 1, 5, 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 393, + 287, + 448 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 393, + 287, + 448 + ], + "spans": [ + { + "bbox": [ + 48, + 393, + 287, + 448 + ], + "type": "text", + "content": "[34] Lijun Wang, Huchuan Lu, Yifan Wang, Mengyang Feng, Dong Wang, Baocai Yin, and Xiang Ruan. Learning to detect salient objects with image-level supervision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 136-145, 2017. 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 449, + 287, + 492 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 449, + 287, + 492 + ], + "spans": [ + { + "bbox": [ + 48, + 449, + 287, + 492 + ], + "type": "text", + "content": "[35] Yisen Wang, Difan Zou, Jinfeng Yi, James Bailey, Xingjun Ma, and Quanquan Gu. Improving adversarial robustness requires revisiting misclassified examples. In International Conference on Learning Representations, 2019. 8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 493, + 287, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 493, + 287, + 525 + ], + "spans": [ + { + "bbox": [ + 48, + 493, + 287, + 525 + ], + "type": "text", + "content": "[36] Lina Wei, Shanshan Zhao, Omar El Farouk Bourahla, Xi Li, and Fei Wu. Group-wise deep co-saliency detection. arXiv preprint arXiv:1707.07381, 2017. 1, 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 526, + 287, + 559 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 526, + 287, + 559 + ], + "spans": [ + { + "bbox": [ + 48, + 526, + 287, + 559 + ], + "type": "text", + "content": "[37] Cihang Xie, Jianyu Wang, Zhishuai Zhang, Zhou Ren, and Alan Yuille. Mitigating adversarial effects through randomization. arXiv preprint arXiv:1711.01991, 2017. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 559, + 287, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 559, + 287, + 602 + ], + "spans": [ + { + "bbox": [ + 48, + 559, + 287, + 602 + ], + "type": "text", + "content": "[38] Xi Yang, Shaoyi Li, Jun Ma, Jun-yan Yang, and Jie Yan. Co-saliency-regularized correlation filter for object tracking. Signal Processing: Image Communication, 103:116655, 2022. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 603, + 287, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 603, + 287, + 658 + ], + "spans": [ + { + "bbox": [ + 48, + 603, + 287, + 658 + ], + "type": "text", + "content": "[39] Siyue Yu, Jimin Xiao, Bingfeng Zhang, and Eng Gee Lim. Democracy does matter: Comprehensive feature mining for co-salient object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 979–988, 2022. 1, 6, 7, 8" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 658, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 287, + 713 + ], + "type": "text", + "content": "[40] Li Yuan, Yunpeng Chen, Tao Wang, Weihao Yu, Yujun Shi, Zi-Hang Jiang, Francis EH Tay, Jiashi Feng, and Shuicheng Yan. Tokens-to-token vit: Training vision transformers from scratch onImagenet. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2021. 5" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 575 + ], + "type": "list", + "angle": 0, + "index": 25, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "text", + "content": "[41] Chen Zhang, Riccardo Barbano, and Bangti Jin. Conditional variational autoencoder for learned image reconstruction. Computation, 9(11):114, 2021. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 107, + 545, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 107, + 545, + 150 + ], + "spans": [ + { + "bbox": [ + 307, + 107, + 545, + 150 + ], + "type": "text", + "content": "[42] Dingwen Zhang, Junwei Han, Chao Li, and Jingdong Wang. Co-saliency detection via looking deep and wide. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2994-3002, 2015. 2, 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 152, + 545, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 152, + 545, + 205 + ], + "spans": [ + { + "bbox": [ + 307, + 152, + 545, + 205 + ], + "type": "text", + "content": "[43] Jing Zhang, Deng-Ping Fan, Yuchao Dai, Saeed Anwar, Fatemeh Saleh, Sadegh Aliakbarian, and Nick Barnes. Uncertainty inspired rgb-d saliency detection. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2021. 2, 4" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 207, + 545, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 207, + 545, + 251 + ], + "spans": [ + { + "bbox": [ + 307, + 207, + 545, + 251 + ], + "type": "text", + "content": "[44] Kaihua Zhang, Jin Chen, Bo Liu, and Qingshan Liu. Deep object co-segmentation via spatial-semantic network modulation. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 34, pages 12813–12820, 2020. 1, 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 253, + 545, + 307 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 253, + 545, + 307 + ], + "spans": [ + { + "bbox": [ + 307, + 253, + 545, + 307 + ], + "type": "text", + "content": "[45] Kaihua Zhang, Mingliang Dong, Bo Liu, Xiao-Tong Yuan, and Qingshan Liu. Deepacg: Co-saliency detection via semantic-aware contrast gromov-wasserstein distance. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13703-13712, 2021. 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 308, + 545, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 308, + 545, + 363 + ], + "spans": [ + { + "bbox": [ + 307, + 308, + 545, + 363 + ], + "type": "text", + "content": "[46] Kaihua Zhang, Tengpeng Li, Bo Liu, and Qingshan Liu. Cosaliency detection via mask-guided fully convolutional networks with multi-scale label smoothing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3095-3104, 2019. 1, 2, 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 365, + 545, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 365, + 545, + 430 + ], + "spans": [ + { + "bbox": [ + 307, + 365, + 545, + 430 + ], + "type": "text", + "content": "[47] Kaihua Zhang, Tengpeng Li, Shiwen Shen, Bo Liu, Jin Chen, and Qingshan Liu. Adaptive graph convolutional network with attention graph clustering for co-saliency detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9050-9059, 2020. 2, 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 431, + 545, + 474 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 431, + 545, + 474 + ], + "spans": [ + { + "bbox": [ + 307, + 431, + 545, + 474 + ], + "type": "text", + "content": "[48] Kaihua Zhang, Yang Wu, Mingliang Dong, Bo Liu, Dong Liu, and Qingshan Liu. Deep object co-segmentation and c-osaliency detection via high-order spatial-semantic network modulation. IEEE Transactions on Multimedia, 2022. 1, 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 475, + 545, + 529 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 475, + 545, + 529 + ], + "spans": [ + { + "bbox": [ + 307, + 475, + 545, + 529 + ], + "type": "text", + "content": "[49] Ni Zhang, Junwei Han, Nian Liu, and Ling Shao. Summarize and search: Learning consensus-aware dynamic convolution for co-saliency detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4167-4176, 2021. 6, 7" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 531, + 547, + 575 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 531, + 547, + 575 + ], + "spans": [ + { + "bbox": [ + 307, + 531, + 547, + 575 + ], + "type": "text", + "content": "[50] Zhao Zhang, Wenda Jin, Jun Xu, and Ming-Ming Cheng. Gradient-induced co-saliency detection. In Proceedings of the European Conference on Computer Vision, pages 455-472. Springer, 2020. 1, 2, 6" + } + ] + } + ], + "index": 24 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 734, + 318, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 734, + 318, + 742 + ], + "spans": [ + { + "bbox": [ + 294, + 734, + 318, + 742 + ], + "type": "text", + "content": "19648" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Co-Speech Gesture Synthesis by Reinforcement Learning With Contrastive Pre-Trained Rewards/6b1eace7-6844-4cc2-b39b-76192850eed9_content_list.json b/2023/Co-Speech Gesture Synthesis by Reinforcement Learning With Contrastive Pre-Trained Rewards/6b1eace7-6844-4cc2-b39b-76192850eed9_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..fb344bfc12b82862b17791ebb1f15ec99fe9163c --- /dev/null +++ b/2023/Co-Speech Gesture Synthesis by Reinforcement Learning With Contrastive Pre-Trained Rewards/6b1eace7-6844-4cc2-b39b-76192850eed9_content_list.json @@ -0,0 +1,1416 @@ +[ + { + "type": "text", + "text": "Co-speech Gesture Synthesis by Reinforcement Learning with Contrastive Pre-trained Rewards", + "text_level": 1, + "bbox": [ + 111, + 128, + 859, + 174 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Mingyang Sun $^{1*}$ , Mengchen Zhao $^{2*}$ , Yaqing Hou $^{1\\dagger}$ , Minglei Li $^{3}$ , Huang Xu $^{3}$ , Songcen Xu $^{2}$ , Jianye Hao $^{2,4}$ , Dalian University of Technology, China ${ }^{2}$ Noah's Ark Lab, Huawei, China ${ }^{3}$ Huawei Cloud Computing Technologies Co., Ltd, China ${ }^{4}$ Tianjin University, China", + "bbox": [ + 86, + 200, + 898, + 257 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "mysun@mail.dlut.edu.cn, houyq@dlut.edu.cn", + "bbox": [ + 308, + 258, + 676, + 273 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{zhaomengchen,liminglei29,xuhuang2,xusongcen,haojianye}@huawei.com", + "bbox": [ + 200, + 276, + 784, + 292 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 325, + 313, + 343 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "There is a growing demand of automatically synthesizing co-speech gestures for virtual characters. However, it remains a challenge due to the complex relationship between input speeches and target gestures. Most existing works focus on predicting the next gesture that fits the data best, however, such methods are myopic and lack the ability to plan for future gestures. In this paper, we propose a novel reinforcement learning (RL) framework called RACER to generate sequences of gestures that maximize the overall satisfactory. RACER employs a vector quantized variational autoencoder to learn compact representations of gestures and a GPT-based policy architecture to generate coherent sequence of gestures autoregressively. In particular, we propose a contrastive pre-training approach to calculate the rewards, which integrates contextual information into action evaluation and successfully captures the complex relationships between multi-modal speech-gesture data. Experimental results show that our method significantly outperforms existing baselines in terms of both objective metrics and subjective human judgements. Demos can be found at https://github.com/RLracer/RACER.git.", + "bbox": [ + 75, + 358, + 472, + 676 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 704, + 209, + 720 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Gesturing is important for human speakers to improve their expressiveness. It conveys the necessary non-verbal information to the audience, giving the speech a more emotional touch so that to enhance persuasiveness and credibility. Similarly, in virtual world, high-quality 3D gesture animations can make a talking character more vividly. For example, in human-computer interaction scenarios, vivid gestures performed by virtual characters can help listeners to concentrate and improve the intimacy between humans and", + "bbox": [ + 75, + 729, + 468, + 867 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "characters [36]. Attracted by these merits, there has been a growing demand of automatically synthesizing high-quality co-speech gestures in computer animation.", + "bbox": [ + 500, + 328, + 890, + 373 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "However, automatically synthesizing co-speech gestures remains a challenge due to the complicated relationship between speech audios and gestures. On one hand, a speaker may play different gestures when speaking the same words due to different mental and physical states. On the other hand, a speaker may also play similar gestures when speaking different words. Therefore, co-speech gesture synthesizing is inherently a \"many-to-many\" problem [21]. Moreover, in order to ensure the overall fluency and consistency, we must take into consideration both the contextual information and the subsequent effect upon playing a gesture [23, 38]. Therefore, gesture synthesizing is rather a sequential decision making problem than a simple matching between speeches and gestures.", + "bbox": [ + 496, + 378, + 892, + 589 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Compared with traditional rule-based approaches [35], data-driven gesture synthesis approaches [10,39] has shown many advantages, including the low development cost and the ability to generalize. Most existing data-driven approaches consider gesture synthesis as a classification task [6, 26] or a regression task [17] in a deterministic way i.e. the same speech or text input always maps to the same gesture output. However, these models rely on the assumption that there exists a unique ground-truth label for each input sequence of speech, which contradicts to the \"many-to-many\" nature of the problem. As a consequence, they sacrifice diversity and semantics of gestures and tend to learn some averaged gestures for any input speeches. Some other works adopt adversarial learning framework, where a discriminator is trained to distinguish between generated gestures and the recorded gestures in dataset [10,38]. Although they improve the generalizability of the gesture generator to some extent, they still fail to explore the essential relationship between speeches and gestures.", + "bbox": [ + 496, + 594, + 893, + 880 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In order to address the above challenges, we propose", + "bbox": [ + 517, + 885, + 890, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Equal contribution.", + "bbox": [ + 94, + 875, + 207, + 887 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "† Corresponding author.", + "bbox": [ + 96, + 887, + 220, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "2331", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "a novel Reinforcement leArning framework with Contrasitive prE-trained Rewards (RACER) for generating high-quality co-speech gestures. RACER is trained in an offline manner but can be used to generate the next gesture for a speaking character in real time. RACER consists of the following three components. Firstly, in order to extract meaningful gestures from the infinite action space, RACER adopts a vector quantized variational autoencoder (VQ-VAE) [32] to learn compact gesture representations, which significantly reduces the action space. Secondly, we construct the Q-value network using a GPT-based [28] model, which has natural advantages of generating coherence sequence of gestures. Thirdly, inspired by the contrastive language-image pre-training (CLIP) [27] and recent advances of contrastive learning [30], we propose a contrastive speech-gesture pre-training method to compute the rewards, which guide the RL agent to explore sophisticated relations between speeches and gestures. Note that these rewards evaluate the quality of gestures as a sequence. By contrast, in conventional supervised learning frameworks, the focus is on predicting only the next gesture, ignoring the quality of generated sequence as a whole. To sum up, the core contributions of this paper are as follows:", + "bbox": [ + 76, + 90, + 472, + 439 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "(1) We formally model the co-speech gesture synthesis problem as a Markov decision process and propose a novel RL based approach called RACER to learn the optimal gesture synthesis policy. RACER can be trained in an offline manner and used to synthesis co-speech gestures in real time.", + "(2) We introduce VQ-VAE to encode and quantize the motion segments to a codebook, which significantly reduces the action space and facilitates the RL phase.", + "(3) We propose a contrastive speech-gesture pre-training method to compute the rewards, which guide the RL agent to discover deeper relations from multi-modal speech-gesture data.", + "(4) Extensive experimental results show that RACER outperforms the existing baselines in terms of both objective metrics and subjective human judgements. This demonstrates the superiority and the potential of RL in co-speech gesture synthesis tasks." + ], + "bbox": [ + 93, + 449, + 468, + 750 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 76, + 762, + 218, + 777 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Data-driven Motion Synthesis", + "text_level": 1, + "bbox": [ + 76, + 786, + 341, + 801 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Human motion synthesis has been studied for a long time [21, 37]. Traditional approaches such as motion graph [16] cut the motion capture data and synthesize new motions according to crafted transition rules. With the development of deep learning, using neural networks to synthesize motions in an end-to-end manner becomes popular. Some of", + "bbox": [ + 75, + 810, + 470, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "works focus on motion prediction [13, 15], which could predict current motion with a high accuracy, but they often fail to predict longer motion sequences. Recent works focus more on the motion synthesis, which aims at generating long-term natural motions such as walking, jumping, waving and dancing. Many advanced techniques have been applied to long-term motion synthesis, including conditional GANs [33], VAEs [21], auto-regressive models [7] and Transformers [5, 22]. A most recent work called Bailando explores the potential of RL in synthesizing dances for given music [29]. However, in Bailando, RL is used only for fine tuning so that the exact contributions of RL remain unclear. By contrast, our proposed method RACER is trained solely by RL and its effectiveness is demonstrated by extensive experiments.", + "bbox": [ + 496, + 90, + 893, + 318 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Co-speech Gesture Synthesis", + "text_level": 1, + "bbox": [ + 500, + 325, + 756, + 342 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Co-speech gesture synthesis is challenging due to the complicated relations between speeches and gestures that are hidden in the multi-modal dataset. One line of conventional methods is rule-based generation [35], where the core idea is to pre-define a set of gesture units and design rules to connect speech words and the gesture units. However, it requires too much human effort in designing the rules, which is costly and inefficient in practice. Benefit from the recent advances in deep learning, many data-driven approaches have been proposed to learn the matching rules from speech-gesture data. Existing works focus on exploring the effectiveness of different network architectures, including the multi-layer perceptron (MLP) [18], convolutional neural networks (CNNs) [11], recurrent neural networks (RNNs) [4] and Transformers [18, 29]. Some recent works incorporate adversarial loss to enhance gesture fidelity [1,9,10,38]. However, they simply construct a target gesture as the unique label of the current segment of speech, ignoring the \"many-to-many\" nature of the problem. To model the complicated relation between speeches and gestures, Alexanderson et al. [2] propose a probabilistic model MoGlow based on normalizing flows [12] which maps input speeches to Gaussian distributions of gestures. However, the randomly sampled gestures contain non-negligible noise and therefore suffer from low interpretability.", + "bbox": [ + 496, + 349, + 893, + 728 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3. Preliminaries", + "text_level": 1, + "bbox": [ + 500, + 739, + 638, + 753 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The underlying environment of reinforcement learning is typically modeled by a Markov decision process (MDP), which is defined by a tuple $(S, A, P, R, \\gamma)$ . $S$ represents the state space and $A$ represents the action space. The transition function $P(\\cdot | s, a)$ gives the probabilities of transitioning to the next state after taking action $a$ at state $s$ . $R(\\cdot | s, a)$ is the associated reward function and $\\gamma \\in [0, 1)$ represents the discount factor. A policy $\\pi(\\cdot | s)$ maps each state $s \\in S$ to a distribution over actions. For an agent following the policy", + "bbox": [ + 496, + 763, + 893, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2332", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/024374574dbb552d472bc6a0aa14c844fa433a13bf08e57e07ca6aa9589c87b1.jpg", + "image_caption": [ + "Figure 1. The overall architecture of RACER. Given a piece of speech audio and the initial gesture code $a_0$ , the Q-network represented by transformer layers autoregressively calculates the Q-values and selects a sequence of actions $(a_1, \\dots, a_T)$ . The action sequence will then be transformed to quantitative features by querying the codebook and finally be decoded to motion sequences by the decoder of VQ-VAE." + ], + "image_footnote": [], + "bbox": [ + 178, + 102, + 795, + 378 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "$\\pi$ , the action value function, denoted by $Q^{\\pi}(s,a)$ , is defined as the expectation of cumulative discounted future rewards:", + "bbox": [ + 76, + 460, + 470, + 491 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nQ ^ {\\pi} (s, a) := \\mathbb {E} \\left[ \\sum_ {t = 0} \\gamma^ {t} R \\left(s _ {t}, a _ {t}\\right) \\right], \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 91, + 511, + 468, + 553 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\ns _ {0} = s, a _ {0} = a, s _ {t} \\sim P (\\cdot | s _ {t - 1}, a _ {t - 1}), a _ {t} \\sim \\pi (\\cdot | s _ {t}).\n$$\n", + "text_format": "latex", + "bbox": [ + 93, + 555, + 434, + 571 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The goal in RL is to learn an optimal policy $\\pi^{*}$ that maximizes the expected cumulative reward. We denote by $Q^{*}(s,a)$ the optimal action value function under the optimal policy $\\pi^{*}$ . The Bellman equations for the optimal Q-values can be represented by:", + "bbox": [ + 76, + 580, + 468, + 657 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nQ ^ {*} (s, a) := R (s, a) + \\gamma \\mathbb {E} _ {s ^ {\\prime} \\sim P} \\max _ {a ^ {\\prime} \\in \\mathcal {A}} Q ^ {*} \\left(s ^ {\\prime}, a ^ {\\prime}\\right). \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 96, + 679, + 468, + 702 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "A standard approach for computing the optimal policy is Q-learning, which iteratively improves the estimation of $Q^{*}$ by applying the Bellman operator derived from Eq. (2). While traditional Q-learning requires interactions with the environment, recent advances in offline RL allow us to learn the optimal policy directly from a fixed dataset, which we will elaborate in Sec. 4.3.", + "bbox": [ + 75, + 710, + 468, + 815 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4. Our Approach", + "text_level": 1, + "bbox": [ + 76, + 830, + 225, + 847 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The overview of our proposed framework RACER is shown in Fig. 1. At each time step $t$ , the state $s$ consists of the generated action tokens $(a_{1},\\ldots ,a_{t - 1})$ and input audio.", + "bbox": [ + 76, + 854, + 470, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Unlike existing methods which directly learn a mapping from audio features to the continuous high-dimensional motion space, we encode and quantize the motion into a finite codebook $\\mathcal{Z} = \\{z_i\\}_{i=1}^N$ by VQ-VAE, where $N$ is the size of codebook and each code $z_i$ represents a gesture lexeme feature. The details of action design are introduced in Sec. 4.1. We use a GPT-like unidirectional model as the Q-network that autoregressively outputs action tokens following a greedy strategy. An action token $a$ will be mapped to a gesture lexeme feature $z$ and then be decoded to a specific gesture motion. Moreover, we propose a contrastive speech-gesture pre-training model to compute the immediate rewards for the actions, which will be elaborated in Sec. 4.2. In addition, we will introduce how to train the Q-network in a fully offline manner in Sec. 4.3.", + "bbox": [ + 496, + 460, + 890, + 686 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4.1. Action Design", + "text_level": 1, + "bbox": [ + 500, + 696, + 643, + 712 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Instead of designing handcrafted gestures with expensive manual efforts, our goal is to summarize gesture lexemes into a representative codebook from a large gesture dataset in an unsupervised manner. To collect distinctive gesture representations and reconstruct them back to real gestures, we employ a VQ-VAE model which has successfully demonstrated its ability to learn effective latent representations from temporal data. The structure of VQ-VAE is shown in Fig. 2. VQ-VAE extends the regular autoencoder by adding a discrete codebook $\\mathcal{Z}$ to the network, which consists of a list of vectors with their corresponding indexes. A piece of motion in raw dataset can be represented by a ma", + "bbox": [ + 496, + 719, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "2333", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/62534f56cfd55598bcfb00c688e358e6566c967fad156f743763c34360e4349d.jpg", + "image_caption": [ + "Figure 2. Structure of motion VQ-VAE." + ], + "image_footnote": [], + "bbox": [ + 83, + 103, + 460, + 253 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "trix $\\mathcal{M} \\in \\mathbb{R}^{T \\times J}$ , where $T$ is the number of frames and $J$ is the dimension of joint features. $\\mathcal{M}$ can then be transformed to latent features $e \\in \\mathbb{R}^{T' \\times C}$ by a 1-D convolutional network, where $T'$ is the length of feature sequence after down-sampling, and $C$ is the channel dimension. Each row of $e$ , represented by $e_i$ , will be substituted by its closest vector $z_j$ in the codebook:", + "bbox": [ + 75, + 318, + 468, + 425 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {e} _ {i} = \\arg \\min _ {\\boldsymbol {z} _ {j} \\in \\mathcal {Z}} | | \\boldsymbol {e} _ {i} - \\boldsymbol {z} _ {j} | |. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 184, + 441, + 468, + 465 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Then, the quantized features $\\hat{e}_i$ will be decoded to $\\hat{\\mathcal{M}}$ via a decoder so that to reconstruct $\\mathcal{M}$ . Following [32], the encoder and decoder are simultaneously learned by minimizing the following loss function:", + "bbox": [ + 76, + 474, + 468, + 536 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} _ {V Q} = \\mathcal {L} _ {m} (\\mathcal {M}, \\hat {\\mathcal {M}}) \\tag {4} \\\\ + \\left\\| \\hat {\\boldsymbol {e}} - \\operatorname {s g} (\\boldsymbol {e}) \\right\\| _ {2} ^ {2} + \\beta \\left\\| \\operatorname {s g} (\\hat {\\boldsymbol {e}}) - \\boldsymbol {e} \\right\\| _ {2} ^ {2}, \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 127, + 560, + 468, + 599 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathfrak{sg}[\\cdot ]$ stands for the stop gradient operator that prevents the gradient from backpropagating through its operand. The first term $\\mathcal{L}_m(\\mathcal{M},\\hat{\\mathcal{M}})$ represents the reconstruction error, which is expanded as:", + "bbox": [ + 76, + 611, + 468, + 672 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} _ {m} (\\mathcal {M}, \\hat {\\mathcal {M}}) = | | \\mathcal {M}, \\hat {\\mathcal {M}} | | _ {1} \\tag {5} \\\\ + \\alpha_ {1} | | \\mathcal {M} ^ {\\prime}, \\hat {\\mathcal {M}} ^ {\\prime} | | _ {1} + \\alpha_ {2} | | \\mathcal {M} ^ {\\prime \\prime}, \\hat {\\mathcal {M}} ^ {\\prime \\prime} | | _ {1}, \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 84, + 696, + 468, + 736 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathcal{M}'$ and $\\mathcal{M}''$ are the $1^{st}$ order and $2^{nd}$ order of partial derivatives of motion sequence $\\mathcal{M}$ , representing the velocity and the acceleration of the motion, respectively. $\\alpha_{1}$ and $\\alpha_{2}$ are hyper-parameters used to trade-off between the two losses. We can see from Eq. (5) that VQ-VAE reconstructs not only the original position of each joints of the character, but also the velocities and the accelerations of its motions. Notably, since $\\mathcal{Z}$ is discrete, we will simply copy the gradients from the reconstruction error and pass them to the encoder during training. The second term in Equation (4) aims", + "bbox": [ + 75, + 750, + 468, + 900 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/cbc062a5b975db7d728e2d923065c654787eb6db9c9f177cb17cb0c69ce0daa1.jpg", + "image_caption": [ + "Figure 3. Summary of the reward model. The model jointly trains a motion encoder and an audio encoder to predict the correct pairings of a batch of (motion, audio) training examples." + ], + "image_footnote": [], + "bbox": [ + 509, + 116, + 846, + 284 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "to optimize the codebook by pushing the $\\hat{e}$ and its associated $z$ close to $e$ , which is the output of the encoder. The last term in Equation (4) aims to optimize the encoder by pushing $e$ close to its nearest latent vector in the codebook. The VQ-VAE mode will be trained separately, which means that its model parameters as well as the codebook remain unchanged during reward computation and reinforcement learning. Note that in the reinforcement learning phase, the indexes of the codebook will be treated as the action tokens, as introduced in Sec. 4.3.", + "bbox": [ + 496, + 366, + 890, + 517 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.2. Reward Design", + "text_level": 1, + "bbox": [ + 500, + 529, + 653, + 545 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Reward function is key to reinforcement learning because it contains most of the knowledge that the agent can learn. Regarding to the task of co-speech gesture synthesis, the original dataset explicitly indicates the ground-truth gestures for given speech audios. However, due to the \"many-to-many\" nature of the problem, a ground-truth gesture may not be optimal under different contexts. This motivates us to learn a reward model that is able to discover and utilize deeper relations between speeches and gestures. There are two advantages of using the reward model. First, if there are multiple gestures correspond to one speech slice, the reward model is able to distinguish between these gestures so that to enrich the supervision signals. Second, as the goal of RL is to maximize the accumulated rewards of action sequences, the agent will learn to generate sequences of gestures that maximize the overall satisfaction.", + "bbox": [ + 496, + 551, + 890, + 792 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To get a proper reward function, we train a model to evaluate the degree of correspondence between the sequences of gestures and the sequences of speeches. Since the reward model tells the agent the degree of correspondence of the speech-gesture pairs, it should be able to distinguish between matched pairs and unmatched pairs. Moreover, the reward model should also generalize to speech-gesture pairs", + "bbox": [ + 496, + 795, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "2334", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "that are not covered by the dataset. This coincides with the idea of contrastive learning. Multimodal contrastive learning techniques have been developed recently, such as CLIP [27], a well-known visual-text contrastive learning model, designed to match paired image and caption embeddings.", + "bbox": [ + 75, + 90, + 468, + 167 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Similar to CLIP, given a batch of $n$ (motion, audio) pairs, the reward model is trained to predict which of the $n \\times n$ possible (motion, audio) pairs across a batch actually occurred. To do this, the model converts each input motion $\\mathcal{M}$ and audio $\\mathcal{U}$ into vector representations $\\boldsymbol{m}$ and $\\boldsymbol{u}$ by a motion encoder $\\mathcal{E}_m$ and an audio encoder $\\mathcal{E}_a$ . Two encoders are jointly trained to maximize the cosine similarity of each real pair $(m_i$ and $u_i)$ :", + "bbox": [ + 75, + 167, + 470, + 287 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {R} = \\ell_ {i} ^ {\\boldsymbol {m} \\rightarrow \\boldsymbol {u}} + \\ell_ {i} ^ {\\boldsymbol {u} \\rightarrow \\boldsymbol {m}}, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 192, + 297, + 468, + 316 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where", + "bbox": [ + 76, + 327, + 122, + 340 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\ell_ {i} ^ {\\boldsymbol {m} \\rightarrow \\boldsymbol {u}} = - \\log \\frac {e x p (m _ {i} \\cdot u _ {i} / \\tau)}{\\sum_ {k = 1} ^ {n} e x p (m _ {i} \\cdot u _ {k} / \\tau)},\n$$\n", + "text_format": "latex", + "bbox": [ + 138, + 349, + 406, + 388 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\ell_ {i} ^ {\\boldsymbol {u} \\rightarrow \\boldsymbol {m}} = - \\log \\frac {\\exp (u _ {i} \\cdot m _ {i} / \\tau)}{\\sum_ {k = 1} ^ {n} \\exp (u _ {i} \\cdot m _ {k} / \\tau)}.\n$$\n", + "text_format": "latex", + "bbox": [ + 142, + 386, + 403, + 419 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "$\\tau \\in \\mathbb{R}^{+}$ represents a temperature parameter. In this work, both encoders are structured as 1-D temporal convolutional networks and can input sequences of different lengths.", + "bbox": [ + 75, + 429, + 468, + 474 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As with VQ-VAE, the training of the reward model is in a separate phase. During reinforcement learning, two encoders of the learned reward model respectively encode the generated motion $\\hat{\\mathcal{M}}$ and input speech $\\mathcal{U}$ into two vectors, with their inner product as the reward:", + "bbox": [ + 75, + 474, + 468, + 551 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nR (\\hat {\\mathcal {M}}, \\mathcal {U}) = \\mathcal {E} _ {m} (\\hat {\\mathcal {M}}) \\cdot \\mathcal {E} _ {u} (\\mathcal {U}). \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 171, + 563, + 468, + 582 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.3. Offline Reinforcement Learning", + "text_level": 1, + "bbox": [ + 76, + 590, + 361, + 606 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Recall that we have defined several essential factors of the MDP. State $s$ includes the generated action tokens $\\pmb{a}_{0:t-1}$ and the audio sequence $\\pmb{u}_{0:t}$ up to the current time step $t$ . Each action $a$ corresponds to an index of the codebook learned by VQ-VAE. The reward $r$ is outputted by the reward model to evaluate how well the predicted gesture matches the input audio. Now we are able to augment the original speech-gesture dataset to a trajectory dataset $\\mathcal{D} = \\{s_i, a_i, r_i, s_i' | i = 1, \\dots, N\\}$ . Note that these trajectories can be collected before the start of the reinforcement learning phase. We denote by $\\pi_{\\beta}$ the behavior policy that generates these trajectories.", + "bbox": [ + 75, + 613, + 468, + 794 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Offline datasets usually do not provide complete state-action coverage. That is, the set $\\{(s,a)|(s,a,r,s')\\in \\mathcal{D}\\}$ is typically a small subset of the full space $S\\times \\mathcal{A}$ . Standard online reinforcement learning methods such as DDPG [24] and SAC [31] are not appropriate due to issues with bootstrapping from out-of-distribution actions. A \"safe\" strategy to mitigate the distributional shift problem is to be con", + "bbox": [ + 75, + 795, + 468, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "servative: if we explicitly estimate the value of unseen outcomes conservatively (i.e. assign them a low value), then the estimated value or performance of the policy that executes unseen behaviors is guaranteed to be small. Using such conservative estimates for policy optimization will prevent the policy from executing unseen actions and it will perform reliably. To this end, we adopt a simple but effective offline reinforcement learning algorithm called conservative Q-learning (CQL) [20] to learn the optimal policy from the augmented offline dataset.", + "bbox": [ + 496, + 90, + 890, + 241 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Specifically, for the conservative off-policy evaluation, the Q-function, $\\hat{Q}^{\\pi} := \\lim_{k \\to \\infty} \\hat{Q}_k$ is trained via an iterative update:", + "bbox": [ + 496, + 242, + 890, + 287 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\hat {Q} ^ {k + 1} \\leftarrow \\arg \\min _ {Q} \\frac {1}{2} \\underset {s, a, s ^ {\\prime} \\sim \\mathcal {D}} {\\mathbb {E}} \\left[ (Q (s, a) - \\hat {\\mathcal {B}} ^ {\\pi} \\hat {Q} ^ {k} (s, a)) ^ {2} \\right] \\\\ + \\alpha \\left(\\underset {a \\sim \\mu (a | s)} {\\mathbb {E}} [ Q (s, a) ] - \\underset {a \\sim \\hat {\\pi} _ {\\beta} (a | s)} {\\mathbb {E}} [ Q (s, a) ]\\right), \\tag {9} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 506, + 308, + 890, + 402 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where", + "bbox": [ + 500, + 404, + 545, + 416 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathcal {B}} ^ {\\pi} \\hat {Q} ^ {k} (s, a) = \\mathbb {E} _ {s ^ {\\prime} \\sim P \\left(s ^ {\\prime} \\mid s, a\\right)} [ R (s, a) + \\gamma \\mathbb {E} _ {a ^ {\\prime} \\sim \\hat {\\pi} ^ {k} (s, a)} \\hat {Q} ^ {k} \\left(s ^ {\\prime}, a ^ {\\prime}\\right) ] \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 500, + 426, + 908, + 458 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "is the Bellman operator. This equation consists of the normal bootstrapping error and the regularization term with a trade-off factor $\\alpha$ . Inside the regularization term, the first term always pushes the Q-value down on the $(s,a)$ pairs sampled from the learning policy $\\mu$ whereas the second term pushes Q-value up on the $(s,a)$ pairs sampled from the offline data set. In practice, we selected the binding variant of CQL and DQN as our model, which can be optimized by minimizing the following loss function:", + "bbox": [ + 496, + 460, + 890, + 595 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\underset {Q} {\\operatorname {m i n}} \\alpha_ {s \\sim \\mathcal {D}} \\left[ \\log \\sum_ {a} \\exp (Q (s, a)) - \\underset {a \\sim \\hat {\\pi} _ {\\beta}} {\\mathbb {E}} [ Q (s, a) ] \\right] \\\\ + \\frac {1}{2} _ {s, a, s ^ {\\prime} \\sim \\mathcal {D}} \\mathbb {E} \\left[ (Q (s, a) - \\hat {\\mathcal {B}} ^ {\\pi} \\hat {Q} ^ {k} (s, a)) ^ {2} \\right]. \\tag {11} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 517, + 616, + 890, + 703 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Since we want to use the Q-value function to autoregressively generate the action sequences, we construct a powerful GPT-based backbone model to estimate the Q-values, as is shown in Fig. 1. To construct the input, each motion code sequence is embedded to learnable features and concatenated with the audio features in the temporal dimension. We add a learned positional encoding to this concatenated $(2\\times T^{\\prime}\\times C)$ -dimensional tensor and feed it to Transformer layers, whose structure is shown in the left of Fig. 1. Last, we employ a linear layer to map the output of Transformer layers to the Q-value $q\\in \\mathbb{R}^{T^{\\prime}\\times N}$ , where $N$ is the size of learned codebook. In the Transformer layer, the attention is the core component that determines the computational", + "bbox": [ + 496, + 704, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "2335", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "dependency among sequential elements of inputs, which is computed as:", + "bbox": [ + 75, + 90, + 468, + 121 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {A t t e n t i o n} (\\boldsymbol {Q}, \\boldsymbol {K}, \\boldsymbol {V}, \\boldsymbol {M}) = \\operatorname {s o f t m a x} \\left(\\frac {\\boldsymbol {Q} \\boldsymbol {K} ^ {T}}{\\sqrt {C}} + \\boldsymbol {M}\\right) \\boldsymbol {V}, \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 86, + 141, + 468, + 186 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $Q, K, V$ denote the query, key and value from input, and $V$ is the mask. In the original GPT model, \"causal attention\" realizes the intercommunication of the current and previous data to compute the state for the time of interest. However, our model needs to infer actions from the speech at each time step, so it needs to process two input sequences simultaneously. To comply the causality conditioned among features of the audio and the action, $M$ is designed to be a $2 \\times 2$ repeated block matrix with lower triangular matrix of size $T'$ .", + "bbox": [ + 75, + 188, + 468, + 338 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "During inference, given the audio of arbitrary length and the initial motion code $a_0$ , the Q-network can output Q-values of all actions. We will select the action with the maximum Q-value and then add it to the input, so that we can autoregressively generate the entire action sequence.", + "bbox": [ + 75, + 339, + 468, + 415 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 76, + 428, + 209, + 444 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1. Dataset", + "text_level": 1, + "bbox": [ + 76, + 452, + 171, + 467 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this paper, we use two speech-gesture datasets: the Trinity dataset and a Chinese dataset collected by us. Trinity Gesture dataset is a large database of speech and gestures jointly collected by [8]. This dataset consists of 242 minutes of motion capture and audio of one male actor talking on different topics. The actor's motion was captured with a 20-camera Vicon system and solved onto a skeleton with 69 joints. In this paper, we use the official release version of The GENEA Challenge 2020 Dataset [19], where 221 minutes are used as training data, and the remaining 21 minutes are kept for testing. Additionally, we collected an 1-hour high-quality Chinese dataset using motion capture equipment. This dataset contains 3D full-body gestures and the aligned Chinese speech audio. And we retargeted the skeletons to be consistent with the Trinity dataset. In this dataset, the speaker described various speech scenes in Chinese and made high-standard common speech gestures, ensuring a high degree of matching between gestures and speech.", + "bbox": [ + 75, + 476, + 468, + 748 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.2. Implementation Details", + "text_level": 1, + "bbox": [ + 76, + 756, + 294, + 772 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "All the motion data are downsampled to 20 frames per second on the Trinity and Chinese datasets. We focus on upper-body gestures in this work. For visualization, we employ a character model consist of 16 upper-body joints, including a rotational root. The audio features are extracted by the public audio processing toolbox librosa, including mel-scaled spectrogram and onset. In VQ-VAE, the codebook size $N$ and the channel dimension $C$ are both set to", + "bbox": [ + 75, + 780, + 468, + 898 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "512, and the temporal downsampling rate of encoders is set to 8. During the training of VQ-VAE, motion data are cropped to uniform length of $T = 120$ (6 seconds) and sampled in batch size of 64. The trade-off $\\beta$ in $\\mathcal{L}_{VQ}$ is set to 0.1. $\\alpha_{1}$ and $\\alpha_{2}$ in $\\mathcal{L}_m$ are both 1. We adopt Adam optimizer with $\\beta_{1} = 0.9$ and $\\beta_{2} = 0.999$ for 500 epochs on Trinity but 400 epochs on the Chinese dataset with learning rate 0.00003. During the training of reward model, both audio and motion data are clipped to 6s. The optimizer settings are consistent with those of the VQ-VAE. The structure of the Q network is basically the same as that of the GPT network, where the embedding dimension is 768 and the attention layer is implemented in 12 heads with dropout probability 0.1. During the off-line reinforcement learning, the motion sequences are encoded to motion codes and sampled to length of 15 and sampled in batch size of 256. The Q-network is optimized using Adam optimizer with $\\beta_{1} = 0.5$ and $\\beta_{2} = 0.999$ for 1500 epochs on Trinity but 1200 epochs on Chinese dataset, where the learning rate is initialized as 0.0001. The discount factor $\\gamma$ is set to 0.99. At runtime, a Gaussian filter with a kernel size of $K = 5$ is used to smooth the denormalized gesture sequence along the time dimension. The entire framework takes 36 hours on Trinity and 30 hours on the Chinese dataset to complete training on one NVIDIA Quadro RTX 5000 GPU.", + "bbox": [ + 496, + 90, + 890, + 468 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.3. Evaluation Metrics", + "text_level": 1, + "bbox": [ + 500, + 474, + 683, + 489 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We adopt four commonly used evaluation metrics (MAJE, MAD, FGD and PMB) to compare various methods quantitatively. Mean absolute joint error (MAJE) measures the mean of the absolute errors between the generated joint positions and ground truth over all the time steps, which indicates how closely the generated joint positions follow the ground truth. Mean acceleration difference (MAD) measures the mean of the $l2$ norm differences between the generated joint accelerations and ground truth over all the time steps, which indicates how closely the ground truth and the generated joint movements match. Fréchet Gesture Distance (FGD) was proposed by which measures the difference between the distributions of the latent features of the generated gestures and ground truth [38]. The latent features are extracted from an auto-encoder trained on the Human 3.6M dataset*. Specifically, we use $\\hat{H}$ and $H$ represent the latent features of generated gestures and ground truth. The Fréchet distance is formulated as", + "bbox": [ + 496, + 498, + 890, + 770 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {F G D} (\\hat {H}, H) = \\left| \\left| \\hat {\\mu} - \\mu \\right| \\right| _ {2} ^ {2} + \\operatorname {T r} \\left(\\hat {\\Sigma} + \\Sigma - 2 \\sqrt {\\hat {\\Sigma} \\Sigma}\\right), \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 519, + 787, + 890, + 825 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\mu$ and $\\Sigma$ are the first and second moments of the latent feature distribution of the ground truth $H$ , and $\\hat{\\mu}$ and $\\hat{\\Sigma}$ are", + "bbox": [ + 496, + 825, + 890, + 856 + ], + "page_idx": 5 + }, + { + "type": "page_footnote", + "text": "*We obtain the codes and models from: https://github.com/ai4r/Gesture-Generation-from-Trimodal-Context.git", + "bbox": [ + 500, + 862, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "2336", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/4743a398604ebafc3d5f9f060bb87a01a0e2df00753d5f6f064cdc726ebc942e.jpg", + "table_caption": [ + "Table 1. Comparison of RACER to Style Gesture, Gesticulator, S2AG and Bailando on Trinity datasets. $\\downarrow$ means the lower is better and $\\uparrow$ means the higher is better." + ], + "table_footnote": [], + "table_body": "
DatasetMethodMAJE(mm) ↓MAD(mm/s2) ↓FGD↓PMB(%) ↑
TrinityGround Truth00095.74
Style Gesture97.294.2636.9854.54
Gesticulator82.413.6231.0471.00
S2AG54.931.4920.3679.53
Bailando61.891.7417.2984.21
Ours50.331.2113.4489.58
Chinese DatasetStyle Gesture62.953.4630.1153.56
Bailando45.451.2925.0768.53
Ours43.251.199.2174.29
", + "bbox": [ + 163, + 128, + 805, + 303 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "the first and second moments of the latent feature distribution of the generated gestures $\\hat{H}$ . FGD makes a reasonable assessment of the perceived plausibility of the synthesized gestures. The Percentage of Matched Beats PMB is a new metric to calculating the matching rate of audio and motion beats for evaluating rhythm performance [3]. In PMB, a motion beat is considered an matched if the distance to a nearby audio beat is within the range of $\\delta$ :", + "bbox": [ + 75, + 332, + 472, + 454 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {P M B} \\left(B ^ {m}, B ^ {a}\\right) = \\frac {1}{N _ {m}} \\sum_ {i = 1} ^ {N _ {m}} \\sum_ {j = b _ {*} ^ {a} + 1} ^ {N _ {a}} \\mathbb {1} \\left[ \\left| \\left| b _ {j} ^ {m} - b _ {j} ^ {a} \\right| \\right| _ {1} < \\delta \\right], \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 86, + 463, + 468, + 518 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $B^{m}$ and $B^{a}$ represent the motion beats and audio beats, and $N_{m}$ and $N_{a}$ represent the number of beats. Motion beat can be obtained by identifying the local minima of joints deceleration [14]. $b_{*}^{a}$ indicates the last matched audio beat. The $\\delta$ is set to 0.2s, consistent with the original paper.", + "bbox": [ + 75, + 520, + 470, + 598 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.4. Comparison to Existing Methods", + "text_level": 1, + "bbox": [ + 76, + 604, + 367, + 622 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In this section, we compare RACER with several state-of-the-art methods to demonstrate the advances made by our method. On the Trinity dataset, we compare with the methods of Style Gesture [2], Gesticulator [18], Speed to Affective Gesture (S2AG) [4] and Bailando [29]. On the Chinese dataset, we compared with Style Gesture and Bailando. Style Gesture generates gestures based on only the speech audio features by a normalizing flow probabilistic model. Gesticulator leverages the audio and text of the speech to generate semantically consistent gestures. Compared with these methods, S2AG adds more modality, including speaker identity and seed gesture poses. Different from other methods, Bailando is a novel music-to-dance framework with a choreographic memory and actor-critic GPT model. Since Bailando's model architecture is similar to ours and both incorporate reinforcement learning into the training process, we apply it as a key contrast method to the gesture synthesis task. The hyperparameter setting is", + "bbox": [ + 75, + 628, + 472, + 902 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "consistent with the original paper and the open source code, and the feature processing of the training data is consistent with ours. For a fair comparison, we keep the same skeleton and motion frame rate as the comparison systems.", + "bbox": [ + 496, + 332, + 890, + 393 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Quantitative experimental results are listed in Tab. 1. According to the comparison, our proposed model consistently performs favorably against all the other methods on all evaluations. Specially, on Trinity, our method improves $4.60(8.37\\%)$ and $0.28(18.79\\%)$ than the best compared baseline model S2AG on MAJE and MAD, respectively. In terms of FGD and PMB, Bailando, with its advanced network structure and reinforcement learning fine-tuning process, also obtained better evaluation results (17.29 and $84.21\\%$ ) than S2AC (20.36 and $79.53\\%$ ). Our approach further achieves significantly higher performance (13.44 and $89.58\\%$ ) in both items. Similarly, our method has also achieved excellent performance in Chinese dataset compared with Style Gesture and Bailando, especially on FGD (9.21) and PMB $(74.29\\%)$ . The superiority of our method on all evaluation metrics reveals that RACER can rely on offline RL not only to synthesize more real-like motion than compared baseline methods, but also but also to achieve outstanding performance in motion quality and rhythm.", + "bbox": [ + 496, + 397, + 892, + 685 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "User Study. We conducted a user study to further assess the real visual performance of our method in a pairwise manner. We choose Style Gesture and Bailando to compare. We randomly sampled 10 30-second long slices generated from the test set of Trinity dataset. Each motion slice is bound to the Mixamo [25] model and rendered as a video by Blender. The experiment is conducted with 12 participants. The participants are asked to rate the slices from the following three aspects respectively: (1) realism, (2) speech-to-gesture matching, (3) speech-to-gesture rhythm matching. The scores assigned to each rating are in the range of 1-5, corresponding from worst to best. Rating statistics results are shown in Fig. 4. Notably, our method significantly surpasses the compared other methods.", + "bbox": [ + 496, + 688, + 893, + 901 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "2337", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/ed8dd396ec7bcff39edea51b308332d2b6815811120a0dd70f55761258459fa1.jpg", + "table_caption": [ + "Table 2. Ablation study results on the Trinity dataset." + ], + "table_footnote": [], + "table_body": "
MethodMAJE(mm) ↓MAD(mm/s2) ↓FGD↓PMB(%) ↑Reward↑
RACER(Supervised Learning)56.221.8116.6978.54-
RACER(Distance Reward)58.962.9423.5572.09-
RACER(DQN)59.441.8115.4676.634.23
RACER(Actor-Critic)60.181.7617.6978.023.93
Ours50.331.2113.4489.589.34
", + "bbox": [ + 133, + 114, + 838, + 229 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/d9b4d68c071bd6920dd2606bb27e3381c26fdcbde6835d65ef8145604b6d3c84.jpg", + "image_caption": [ + "Figure 4. User study results comparing RACER against Bailando and Style Gesture. Bars with different colors represent rank counts. In total, 120 comparisons have been rated. The total scores are listed on the right." + ], + "image_footnote": [], + "bbox": [ + 83, + 268, + 467, + 400 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.5. Ablation Studies", + "text_level": 1, + "bbox": [ + 76, + 496, + 240, + 511 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To gain more insights into the proposed components of our method, we test some variants of RACER on Trinity dataset: (1) \"Supervised Learning\" indicates that we directly treat gesture code sequences as labels to train GPT-like models through supervised learning instead of reinforcement learning. (2) \"Distance reward\" indicates that our reward function is replaced by the Euclidean distance calculation between the predicted poses and ground truth, which is equivalent to taking the supervision signal in the traditional regression problem directly as the reward. (3) \"DQN\" means that we use the online RL algorithm DQN [34] instead of offline RL to learn the policy. The Q-value network of RACER(DQN) keeps consistent with ours. (4) \"Actor-Critic\" means that we use the online RL algorithm Actor-Critic instead of offline RL to learn the policy. For RACER(Actor-Critic), we regard the last 3 Transformer layers of GPT model with a additional linear-softmax layer as the branch of actor network. Besides, we add 3 Transformer layers as the branch of critic network to estimate the V-values.", + "bbox": [ + 73, + 521, + 468, + 821 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The quantitative scores are shown in Tab. 2. Firstly, we explore the effectiveness of offline reinforcement learning itself. Compared to RACER, the evaluation metrics of RACER(Supervised Learning) sharply drops $5.89(11\\%)$ , $0.60(50\\%)$ , $3.25(24\\%)$ and $11.04\\%$ , respectively. This", + "bbox": [ + 75, + 824, + 468, + 902 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "shows that offline RL with proposed reward function can help the model learn to generate sequences of discrete tokens better than supervised learning. Secondly, to explore the advantages of the contrastive multi-modal reward, we evaluate the performance of RACER(Distance Reward). Our method has achieved better performance, especially on the MAD and PMB matrices. This confirms that the contrastive multi-modal reward function we propose can handle the feature relationship between multiple modes and improve the matching degree of motion and speech. Moreover, to verify the advantages of offline RL over online RL, we introduce two classic online RL algorithms, DQN and Actor-Critic, to determine the variations. Under the same number of training steps, DQN and AC algorithms can only get the highest rewards of 4.23 and 3.93 respectively, which is a significant gap compared to CQL. Naturally, other matrices are also significantly worse than CQL.", + "bbox": [ + 496, + 258, + 893, + 515 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Discussion and Conclusion", + "text_level": 1, + "bbox": [ + 500, + 539, + 748, + 556 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we propose a novel co-speech gesture synthesis framework called RACER. RACER regards the motion generated procedure as a sequential decision making process and employs reinforcement learning to learn the autoregressive gesture synthesis policy. There are mainly three sub-components of RACER. First, a vector quantized variational autoencoder (VQ-VAE) is employed to learn compact representations of gestures from massive real-world gesture data. Second, RACER uses a GPT-based architecture as policy network, which could autoregressively generate the next gestures given a sequence of speech audio. Thirdly, we propose a contrastive pre-training approach to calculate the rewards, which integrates contextual information into action evaluation and successfully captures the complex relationships between multi-modal speech-gesture data. We conduct extensive experiments to evaluate the performance of RACER. The experimental results show that RACER significantly outperforms the existing baselines in various objective and subjective metrics. Moreover, RACER shows a great generalization ability with respect to different contexts, which indicates the potentials of RACER in more complex motion synthesis tasks.", + "bbox": [ + 496, + 568, + 893, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "2338", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Chaitanya Ahuja, Dong Won Lee, Ryo Ishii, and Louis-Philippe Morency. No gestures left behind: Learning relationships between spoken language and freeform gestures. In Findings of the Association for Computational Linguistics: EMNLP 2020, pages 1884–1895, Nov. 2020. 2", + "[2] Simon Alexanderson, Gustav Eje Henter, Taras Kucherenko, and Jonas Beskow. Style-controllable speech-driven gesture synthesis using normalising flows. Computer Graphics Forum, 39, 2020. 2, 7", + "[3] Tenglong Ao, Qingzhe Gao, Yuke Lou, Baoquan Chen, and Libin Liu. Rhythmic gesticulator: Rhythm-aware cospeech gesture synthesis with hierarchical neural embeddings. ArXiv, abs/2210.01448, 2022. 7", + "[4] Uttaran Bhattacharya, Elizabeth Childs, Nicholas Rewkowski, and Dinesh Manocha. Speech2affectivegestures: Synthesizing co-speech gestures with generative adversarial affective expression learning. In Proceedings of the 29th ACM International Conference on Multimedia, page 2027-2036, 2021. 2, 7", + "[5] Uttaran Bhattacharya, Nicholas Rewkowski, Abhishek Banerjee, Pooja Guhan, Aniket Bera, and Dinesh Manocha. Text2gestures: A transformer-based network for generating emotive body gestures for virtual agents. In 2021 IEEE Virtual Reality and 3D User Interfaces (VR), pages 1-10, 2021. 2", + "[6] Chung-Cheng Chiu, Louis-Philippe Morency, and Stacy Marsella. Predicting co-verbal gestures: A deep and temporal modeling approach. In Intelligent Virtual Agents, pages 152–166, 2015. 1", + "[7] Hsu-Kuang Chiu, Ehsan Adeli, Borui Wang, De-An Huang, and Juan Carlos Niebles. Action-agnostic human pose forecasting. In 2019 IEEE Winter Conference on Applications of Computer Vision (WACV), pages 1423–1432, 2019. 2", + "[8] Ylva Ferstl and Rachel McDonnell. Investigating the use of recurrent motion modelling for speech gesture generation. In Proceedings of the 18th International Conference on Intelligent Virtual Agents, page 93–98, 2018. 6", + "[9] Ylva Ferstl, Michael Neff, and Rachel McDonnell. Multi-objective adversarial gesture generation. In Proceedings of the 12th ACM SIGGRAPH Conference on Motion, Interaction and Games, 2019. 2", + "[10] Shiry Ginosar, Amir Bar, Gefen Kohavi, Caroline Chan, Andrew Owens, and Jitendra Malik. Learning individual styles of conversational gesture. In 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3492-3501, 2019. 1, 2", + "[11] Ikhsanul Habibie, Weipeng Xu, Dushyant Mehta, Lingjie Liu, Hans-Peter Seidel, Gerard Pons-Moll, Mohamed Elgharib, and Christian Theobalt. Learning speech-driven 3d conversational gestures from video. In Proceedings of the 21st ACM International Conference on Intelligent Virtual Agents, page 101-108, 2021. 2", + "[12] Gustav Eje Henter, Simon Alexanderson, and Jonas Beskow. Moglow: Probabilistic and controllable motion synthesis using normalising flows. ACM Trans. Graph., 2020. 2" + ], + "bbox": [ + 78, + 114, + 470, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[13] Alejandro Hernandez, Jurgen Gall, and Francesc Moreno-Noguer. Human motion prediction via spatio-temporal inpainting. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), October 2019. 2", + "[14] Chieh Ho, Wei-Tze Tsai, Keng-Sheng Lin, and Homer H. Chen. Extraction and alignment evaluation of motion beats for street dance. In 2013 IEEE International Conference on Acoustics, Speech and Signal Processing, pages 2429–2433, 2013. 7", + "[15] Yu Hua, Fan Xuanzhe, Hou Yaqing, Liu Yi, Kang Cai, Zhou Dongsheng, and Zhang Qiang. Towards efficient 3d human motion prediction using deformable transformer-based adversarial network. In 2022 International Conference on Robotics and Automation (ICRA), pages 861-867, 2022. 2", + "[16] Lucas Kovar, Michael Gleicher, and Frédéric Pighin. Motion graphs. In ACM SIGGRAPH 2008 Classes, SIGGRAPH '08, 2008. 2", + "[17] Taras Kucherenko, Dai Hasegawa, Gustav Eje Henter, Naoshi Kaneko, and Hedvig Kjellström. Analyzing input and output representations for speech-driven gesture generation. In Proceedings of the 19th ACM International Conference on Intelligent Virtual Agents, page 97-104, 2019. 1", + "[18] Taras Kucherenko, Patrik Jonell, Sanne van Waveren, Gustav Eje Henter, Simon Alexandersson, Iolanda Leite, and Hedvig Kjellström. Gesticulator: A framework for semantically-aware speech-driven gesture generation. In Proceedings of the 2020 International Conference on Multimodal Interaction, page 242-250, 2020. 2, 7", + "[19] Taras Kucherenko, Patrik Jonell, Youngwoo Yoon, Pieter Wolfert, and Gustav Eje Henter. A large, crowdsourced evaluation of gesture generation systems on common data: The genea challenge 2020. In 26th International Conference on Intelligent User Interfaces, page 11-21, 2021. 6", + "[20] Aviral Kumar, Aurick Zhou, George Tucker, and Sergey Levine. Conservative q-learning for offline reinforcement learning. In H. Larochelle, M. Ranzato, R. Hadsell, M.F. Balcan, and H. Lin, editors, Advances in Neural Information Processing Systems, volume 33, pages 1179-1191. Curran Associates, Inc., 2020. 5", + "[21] Jing Li, Di Kang, Wenjie Pei, Xuefei Zhe, Ying Zhang, Zhenyu He, and Linchao Bao. Audio2gestures: Generating diverse gestures from speech audio with conditional variational autoencoders. In 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 11273-11282, 2021. 1, 2", + "[22] Ruilong Li, Shan Yang, David A. Ross, and Angjoo Kanazawa. Ai choreographer: Music conditioned 3d dance generation with aist++. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 13401-13412, October 2021. 2", + "[23] Yuanzhi Liang, Qianyu Feng, Linchao Zhu, Li Hu, Pan Pan, and Yi Yang. Seeg: Semantic energized co-speech gesture generation. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10463-10472, 2022. 1", + "[24] Timothy P. Lillicrap, Jonathan J. Hunt, Alexander Pritzel, Nicolas Heess, Tom Erez, Yuval Tassa, David Silver, and" + ], + "bbox": [ + 501, + 92, + 890, + 901 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "2339", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Daan Wierstra. Continuous control with deep reinforcement learning. In 4th International Conference on Learning Representations, ICLR 2016, San Juan, Puerto Rico, May 2-4, 2016, Conference Track Proceedings, 2016. 5", + "[25] Mixamo. https://www MIXamo.com/. 7", + "[26] Shenhan Qian, Zhi Tu, Yihao Zhi, Wen Liu, and Shenghua Gao. Speech drives templates: Co-speech gesture synthesis with learned templates. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 11077-11086, October 2021. 1", + "[27] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision, 2021. 2, 5", + "[28] Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, and Ilya Sutskever. Language Models are Unsupervised Multitask Learners. 2019. 2", + "[29] Li Siyao, Weijiang Yu, Tianpei Gu, Chunze Lin, Quan Wang, Chen Qian, Chen Change Loy, and Ziwei Liu. Bailando: 3d dance generation by actor-critic gpt with choreographic memory. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11050-11059, June 2022. 2, 7", + "[30] Yonglong Tian, Chen Sun, Ben Poole, Dilip Krishnan, Cordelia Schmid, and Phillip Isola. What makes for good views for contrastive learning? In H. Larochelle, M. Ranzato, R. Hadsell, M.F. Balcan, and H. Lin, editors, Advances in Neural Information Processing Systems, volume 33, pages 6827-6839. Curran Associates, Inc., 2020. 2", + "[31] Pieter Abbeel Sergey Levine Tuomas Haarnoja, Aurick Zhou. Soft actor-critic: Off-policy maximum entropy deep reinforcement learning with a stochastic actor. ArXiv, abs/1801.01290, 2018. 5", + "[32] Aaron van den Oord, Oriol Vinyals, and Koray Kavukcuoglu. Neural discrete representation learning. In Proceedings of the 31st International Conference on Neural Information Processing Systems, NIPS'17, page 6309-6318, 2017. 2, 4", + "[33] Ruben Villegas, Jimei Yang, Duygu Ceylan, and Honglak Lee. Neural kinematic networks for unsupervised motion retargeting. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), June 2018. 2", + "[34] David Silver Alex Graves Ioannis Antonoglou Daan Wierstra Martin Riedmiller Volodymyr Mnih, Koray Kavukcuoglu. Playing atari with deep reinforcement learning. ArXiv, abs/1312.5602, 2013. 8", + "[35] Petra Wagner, Zofia Malisz, and Stefan Kopp. Gesture and speech in interaction: An overview. Speech Communication, 57:209-232, 2014. 1, 2", + "[36] Jason R. Wilson, Nah Young Lee, Annie Saechao, Sharon Hershenson, Matthias Scheutz, and Linda Tickle-Degnen. Hand gestures and verbal acknowledgments improve humanrobot rapport. In Social Robotics, pages 334-344, 2017. 1", + "[37] Xinchen Yan, Akash Rastogi, Ruben Villegas, Kalyan Sunkavalli, Eli Shechtman, Sunil Hadap, Ersin Yumer, and Honglak Lee. Mt-vae: Learning motion transformations to" + ], + "bbox": [ + 78, + 90, + 468, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "generate multimodal human dynamics. In Proceedings of the European Conference on Computer Vision (ECCV), September 2018. 2", + "[38] Youngwoo Yoon, Bok Cha, Joo-Haeng Lee, Minsu Jang, Jaeyeon Lee, Jaehong Kim, and Geehyuk Lee. Speech gesture generation from the trimodal context of text, audio, and speaker identity. ACM Trans. Graph., 39(6), nov 2020. 1, 2, 6", + "[39] Youngwoo Yoon, Woo-Ri Ko, Minsu Jang, Jaeyeon Lee, Jae-hong Kim, and Geehyuk Lee. Robots learn social skills: End-to-end learning of co-speech gesture generation for humanoid robots. In 2019 International Conference on Robotics and Automation (ICRA), pages 4303-4309, 2019. 1" + ], + "bbox": [ + 501, + 92, + 890, + 286 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "2340", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2023/Co-Speech Gesture Synthesis by Reinforcement Learning With Contrastive Pre-Trained Rewards/6b1eace7-6844-4cc2-b39b-76192850eed9_model.json b/2023/Co-Speech Gesture Synthesis by Reinforcement Learning With Contrastive Pre-Trained Rewards/6b1eace7-6844-4cc2-b39b-76192850eed9_model.json new file mode 100644 index 0000000000000000000000000000000000000000..b8d2d7a5a96aec3e1da164f38e15b901e8caf3cc --- /dev/null +++ b/2023/Co-Speech Gesture Synthesis by Reinforcement Learning With Contrastive Pre-Trained Rewards/6b1eace7-6844-4cc2-b39b-76192850eed9_model.json @@ -0,0 +1,1892 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.112, + 0.13, + 0.861, + 0.175 + ], + "angle": 0, + "content": "Co-speech Gesture Synthesis by Reinforcement Learning with Contrastive Pre-trained Rewards" + }, + { + "type": "text", + "bbox": [ + 0.088, + 0.202, + 0.9, + 0.258 + ], + "angle": 0, + "content": "Mingyang Sun\\(^{1*}\\), Mengchen Zhao\\(^{2*}\\), Yaqing Hou\\(^{1\\dagger}\\), Minglei Li\\(^{3}\\), Huang Xu\\(^{3}\\), Songcen Xu\\(^{2}\\), Jianye Hao\\(^{2,4}\\), Dalian University of Technology, China \\({ }^{2}\\)Noah's Ark Lab, Huawei, China \\({ }^{3}\\)Huawei Cloud Computing Technologies Co., Ltd, China \\({ }^{4}\\)Tianjin University, China" + }, + { + "type": "text", + "bbox": [ + 0.31, + 0.26, + 0.678, + 0.275 + ], + "angle": 0, + "content": "mysun@mail.dlut.edu.cn, houyq@dlut.edu.cn" + }, + { + "type": "text", + "bbox": [ + 0.202, + 0.277, + 0.785, + 0.293 + ], + "angle": 0, + "content": "{zhaomengchen,liminglei29,xuhuang2,xusongcen,haojianye}@huawei.com" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.327, + 0.314, + 0.344 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.359, + 0.473, + 0.678 + ], + "angle": 0, + "content": "There is a growing demand of automatically synthesizing co-speech gestures for virtual characters. However, it remains a challenge due to the complex relationship between input speeches and target gestures. Most existing works focus on predicting the next gesture that fits the data best, however, such methods are myopic and lack the ability to plan for future gestures. In this paper, we propose a novel reinforcement learning (RL) framework called RACER to generate sequences of gestures that maximize the overall satisfactory. RACER employs a vector quantized variational autoencoder to learn compact representations of gestures and a GPT-based policy architecture to generate coherent sequence of gestures autoregressively. In particular, we propose a contrastive pre-training approach to calculate the rewards, which integrates contextual information into action evaluation and successfully captures the complex relationships between multi-modal speech-gesture data. Experimental results show that our method significantly outperforms existing baselines in terms of both objective metrics and subjective human judgements. Demos can be found at https://github.com/RLracer/RACER.git." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.705, + 0.21, + 0.721 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.73, + 0.47, + 0.868 + ], + "angle": 0, + "content": "Gesturing is important for human speakers to improve their expressiveness. It conveys the necessary non-verbal information to the audience, giving the speech a more emotional touch so that to enhance persuasiveness and credibility. Similarly, in virtual world, high-quality 3D gesture animations can make a talking character more vividly. For example, in human-computer interaction scenarios, vivid gestures performed by virtual characters can help listeners to concentrate and improve the intimacy between humans and" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.329, + 0.892, + 0.374 + ], + "angle": 0, + "content": "characters [36]. Attracted by these merits, there has been a growing demand of automatically synthesizing high-quality co-speech gestures in computer animation." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.379, + 0.893, + 0.59 + ], + "angle": 0, + "content": "However, automatically synthesizing co-speech gestures remains a challenge due to the complicated relationship between speech audios and gestures. On one hand, a speaker may play different gestures when speaking the same words due to different mental and physical states. On the other hand, a speaker may also play similar gestures when speaking different words. Therefore, co-speech gesture synthesizing is inherently a \"many-to-many\" problem [21]. Moreover, in order to ensure the overall fluency and consistency, we must take into consideration both the contextual information and the subsequent effect upon playing a gesture [23, 38]. Therefore, gesture synthesizing is rather a sequential decision making problem than a simple matching between speeches and gestures." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.595, + 0.895, + 0.881 + ], + "angle": 0, + "content": "Compared with traditional rule-based approaches [35], data-driven gesture synthesis approaches [10,39] has shown many advantages, including the low development cost and the ability to generalize. Most existing data-driven approaches consider gesture synthesis as a classification task [6, 26] or a regression task [17] in a deterministic way i.e. the same speech or text input always maps to the same gesture output. However, these models rely on the assumption that there exists a unique ground-truth label for each input sequence of speech, which contradicts to the \"many-to-many\" nature of the problem. As a consequence, they sacrifice diversity and semantics of gestures and tend to learn some averaged gestures for any input speeches. Some other works adopt adversarial learning framework, where a discriminator is trained to distinguish between generated gestures and the recorded gestures in dataset [10,38]. Although they improve the generalizability of the gesture generator to some extent, they still fail to explore the essential relationship between speeches and gestures." + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.886, + 0.892, + 0.902 + ], + "angle": 0, + "content": "In order to address the above challenges, we propose" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.875, + 0.208, + 0.888 + ], + "angle": 0, + "content": "*Equal contribution." + }, + { + "type": "page_footnote", + "bbox": [ + 0.097, + 0.888, + 0.222, + 0.9 + ], + "angle": 0, + "content": "† Corresponding author." + }, + { + "type": "list", + "bbox": [ + 0.096, + 0.875, + 0.222, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "2331" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.473, + 0.44 + ], + "angle": 0, + "content": "a novel Reinforcement leArning framework with Contrasitive prE-trained Rewards (RACER) for generating high-quality co-speech gestures. RACER is trained in an offline manner but can be used to generate the next gesture for a speaking character in real time. RACER consists of the following three components. Firstly, in order to extract meaningful gestures from the infinite action space, RACER adopts a vector quantized variational autoencoder (VQ-VAE) [32] to learn compact gesture representations, which significantly reduces the action space. Secondly, we construct the Q-value network using a GPT-based [28] model, which has natural advantages of generating coherence sequence of gestures. Thirdly, inspired by the contrastive language-image pre-training (CLIP) [27] and recent advances of contrastive learning [30], we propose a contrastive speech-gesture pre-training method to compute the rewards, which guide the RL agent to explore sophisticated relations between speeches and gestures. Note that these rewards evaluate the quality of gestures as a sequence. By contrast, in conventional supervised learning frameworks, the focus is on predicting only the next gesture, ignoring the quality of generated sequence as a whole. To sum up, the core contributions of this paper are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.094, + 0.45, + 0.47, + 0.542 + ], + "angle": 0, + "content": "(1) We formally model the co-speech gesture synthesis problem as a Markov decision process and propose a novel RL based approach called RACER to learn the optimal gesture synthesis policy. RACER can be trained in an offline manner and used to synthesis co-speech gestures in real time." + }, + { + "type": "text", + "bbox": [ + 0.094, + 0.55, + 0.47, + 0.597 + ], + "angle": 0, + "content": "(2) We introduce VQ-VAE to encode and quantize the motion segments to a codebook, which significantly reduces the action space and facilitates the RL phase." + }, + { + "type": "text", + "bbox": [ + 0.094, + 0.605, + 0.47, + 0.667 + ], + "angle": 0, + "content": "(3) We propose a contrastive speech-gesture pre-training method to compute the rewards, which guide the RL agent to discover deeper relations from multi-modal speech-gesture data." + }, + { + "type": "text", + "bbox": [ + 0.094, + 0.675, + 0.47, + 0.751 + ], + "angle": 0, + "content": "(4) Extensive experimental results show that RACER outperforms the existing baselines in terms of both objective metrics and subjective human judgements. This demonstrates the superiority and the potential of RL in co-speech gesture synthesis tasks." + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.45, + 0.47, + 0.751 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.763, + 0.22, + 0.779 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.787, + 0.343, + 0.803 + ], + "angle": 0, + "content": "2.1. Data-driven Motion Synthesis" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.811, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Human motion synthesis has been studied for a long time [21, 37]. Traditional approaches such as motion graph [16] cut the motion capture data and synthesize new motions according to crafted transition rules. With the development of deep learning, using neural networks to synthesize motions in an end-to-end manner becomes popular. Some of" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.895, + 0.319 + ], + "angle": 0, + "content": "works focus on motion prediction [13, 15], which could predict current motion with a high accuracy, but they often fail to predict longer motion sequences. Recent works focus more on the motion synthesis, which aims at generating long-term natural motions such as walking, jumping, waving and dancing. Many advanced techniques have been applied to long-term motion synthesis, including conditional GANs [33], VAEs [21], auto-regressive models [7] and Transformers [5, 22]. A most recent work called Bailando explores the potential of RL in synthesizing dances for given music [29]. However, in Bailando, RL is used only for fine tuning so that the exact contributions of RL remain unclear. By contrast, our proposed method RACER is trained solely by RL and its effectiveness is demonstrated by extensive experiments." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.327, + 0.757, + 0.343 + ], + "angle": 0, + "content": "2.2. Co-speech Gesture Synthesis" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.351, + 0.895, + 0.729 + ], + "angle": 0, + "content": "Co-speech gesture synthesis is challenging due to the complicated relations between speeches and gestures that are hidden in the multi-modal dataset. One line of conventional methods is rule-based generation [35], where the core idea is to pre-define a set of gesture units and design rules to connect speech words and the gesture units. However, it requires too much human effort in designing the rules, which is costly and inefficient in practice. Benefit from the recent advances in deep learning, many data-driven approaches have been proposed to learn the matching rules from speech-gesture data. Existing works focus on exploring the effectiveness of different network architectures, including the multi-layer perceptron (MLP) [18], convolutional neural networks (CNNs) [11], recurrent neural networks (RNNs) [4] and Transformers [18, 29]. Some recent works incorporate adversarial loss to enhance gesture fidelity [1,9,10,38]. However, they simply construct a target gesture as the unique label of the current segment of speech, ignoring the \"many-to-many\" nature of the problem. To model the complicated relation between speeches and gestures, Alexanderson et al. [2] propose a probabilistic model MoGlow based on normalizing flows [12] which maps input speeches to Gaussian distributions of gestures. However, the randomly sampled gestures contain non-negligible noise and therefore suffer from low interpretability." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.74, + 0.64, + 0.755 + ], + "angle": 0, + "content": "3. Preliminaries" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.895, + 0.903 + ], + "angle": 0, + "content": "The underlying environment of reinforcement learning is typically modeled by a Markov decision process (MDP), which is defined by a tuple \\((S, A, P, R, \\gamma)\\). \\( S \\) represents the state space and \\( A \\) represents the action space. The transition function \\( P(\\cdot | s, a) \\) gives the probabilities of transitioning to the next state after taking action \\( a \\) at state \\( s \\). \\( R(\\cdot | s, a) \\) is the associated reward function and \\( \\gamma \\in [0, 1) \\) represents the discount factor. A policy \\( \\pi(\\cdot | s) \\) maps each state \\( s \\in S \\) to a distribution over actions. For an agent following the policy" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2332" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.179, + 0.103, + 0.796, + 0.38 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.392, + 0.895, + 0.436 + ], + "angle": 0, + "content": "Figure 1. The overall architecture of RACER. Given a piece of speech audio and the initial gesture code \\(a_0\\), the Q-network represented by transformer layers autoregressively calculates the Q-values and selects a sequence of actions \\((a_1, \\dots, a_T)\\). The action sequence will then be transformed to quantitative features by querying the codebook and finally be decoded to motion sequences by the decoder of VQ-VAE." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.461, + 0.471, + 0.492 + ], + "angle": 0, + "content": "\\(\\pi\\), the action value function, denoted by \\(Q^{\\pi}(s,a)\\), is defined as the expectation of cumulative discounted future rewards:" + }, + { + "type": "equation", + "bbox": [ + 0.093, + 0.512, + 0.47, + 0.554 + ], + "angle": 0, + "content": "\\[\nQ ^ {\\pi} (s, a) := \\mathbb {E} \\left[ \\sum_ {t = 0} \\gamma^ {t} R \\left(s _ {t}, a _ {t}\\right) \\right], \\tag {1}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.094, + 0.556, + 0.435, + 0.573 + ], + "angle": 0, + "content": "\\[\ns _ {0} = s, a _ {0} = a, s _ {t} \\sim P (\\cdot | s _ {t - 1}, a _ {t - 1}), a _ {t} \\sim \\pi (\\cdot | s _ {t}).\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.582, + 0.47, + 0.658 + ], + "angle": 0, + "content": "The goal in RL is to learn an optimal policy \\(\\pi^{*}\\) that maximizes the expected cumulative reward. We denote by \\(Q^{*}(s,a)\\) the optimal action value function under the optimal policy \\(\\pi^{*}\\). The Bellman equations for the optimal Q-values can be represented by:" + }, + { + "type": "equation", + "bbox": [ + 0.098, + 0.68, + 0.47, + 0.703 + ], + "angle": 0, + "content": "\\[\nQ ^ {*} (s, a) := R (s, a) + \\gamma \\mathbb {E} _ {s ^ {\\prime} \\sim P} \\max _ {a ^ {\\prime} \\in \\mathcal {A}} Q ^ {*} \\left(s ^ {\\prime}, a ^ {\\prime}\\right). \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.712, + 0.469, + 0.816 + ], + "angle": 0, + "content": "A standard approach for computing the optimal policy is Q-learning, which iteratively improves the estimation of \\(Q^{*}\\) by applying the Bellman operator derived from Eq. (2). While traditional Q-learning requires interactions with the environment, recent advances in offline RL allow us to learn the optimal policy directly from a fixed dataset, which we will elaborate in Sec. 4.3." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.831, + 0.226, + 0.848 + ], + "angle": 0, + "content": "4. Our Approach" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.856, + 0.471, + 0.902 + ], + "angle": 0, + "content": "The overview of our proposed framework RACER is shown in Fig. 1. At each time step \\( t \\), the state \\( s \\) consists of the generated action tokens \\( (a_{1},\\ldots ,a_{t - 1}) \\) and input audio." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.461, + 0.892, + 0.688 + ], + "angle": 0, + "content": "Unlike existing methods which directly learn a mapping from audio features to the continuous high-dimensional motion space, we encode and quantize the motion into a finite codebook \\(\\mathcal{Z} = \\{z_i\\}_{i=1}^N\\) by VQ-VAE, where \\(N\\) is the size of codebook and each code \\(z_i\\) represents a gesture lexeme feature. The details of action design are introduced in Sec. 4.1. We use a GPT-like unidirectional model as the Q-network that autoregressively outputs action tokens following a greedy strategy. An action token \\(a\\) will be mapped to a gesture lexeme feature \\(z\\) and then be decoded to a specific gesture motion. Moreover, we propose a contrastive speech-gesture pre-training model to compute the immediate rewards for the actions, which will be elaborated in Sec. 4.2. In addition, we will introduce how to train the Q-network in a fully offline manner in Sec. 4.3." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.697, + 0.644, + 0.713 + ], + "angle": 0, + "content": "4.1. Action Design" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Instead of designing handcrafted gestures with expensive manual efforts, our goal is to summarize gesture lexemes into a representative codebook from a large gesture dataset in an unsupervised manner. To collect distinctive gesture representations and reconstruct them back to real gestures, we employ a VQ-VAE model which has successfully demonstrated its ability to learn effective latent representations from temporal data. The structure of VQ-VAE is shown in Fig. 2. VQ-VAE extends the regular autoencoder by adding a discrete codebook \\(\\mathcal{Z}\\) to the network, which consists of a list of vectors with their corresponding indexes. A piece of motion in raw dataset can be represented by a ma" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2333" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.084, + 0.104, + 0.462, + 0.254 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.154, + 0.279, + 0.393, + 0.293 + ], + "angle": 0, + "content": "Figure 2. Structure of motion VQ-VAE." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.319, + 0.47, + 0.426 + ], + "angle": 0, + "content": "trix \\(\\mathcal{M} \\in \\mathbb{R}^{T \\times J}\\), where \\(T\\) is the number of frames and \\(J\\) is the dimension of joint features. \\(\\mathcal{M}\\) can then be transformed to latent features \\(e \\in \\mathbb{R}^{T' \\times C}\\) by a 1-D convolutional network, where \\(T'\\) is the length of feature sequence after down-sampling, and \\(C\\) is the channel dimension. Each row of \\(e\\), represented by \\(e_i\\), will be substituted by its closest vector \\(z_j\\) in the codebook:" + }, + { + "type": "equation", + "bbox": [ + 0.186, + 0.442, + 0.469, + 0.466 + ], + "angle": 0, + "content": "\\[\n\\hat {e} _ {i} = \\arg \\min _ {\\boldsymbol {z} _ {j} \\in \\mathcal {Z}} | | \\boldsymbol {e} _ {i} - \\boldsymbol {z} _ {j} | |. \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.476, + 0.47, + 0.537 + ], + "angle": 0, + "content": "Then, the quantized features \\(\\hat{e}_i\\) will be decoded to \\(\\hat{\\mathcal{M}}\\) via a decoder so that to reconstruct \\(\\mathcal{M}\\). Following [32], the encoder and decoder are simultaneously learned by minimizing the following loss function:" + }, + { + "type": "equation", + "bbox": [ + 0.128, + 0.561, + 0.469, + 0.601 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} _ {V Q} = \\mathcal {L} _ {m} (\\mathcal {M}, \\hat {\\mathcal {M}}) \\tag {4} \\\\ + \\left\\| \\hat {\\boldsymbol {e}} - \\operatorname {s g} (\\boldsymbol {e}) \\right\\| _ {2} ^ {2} + \\beta \\left\\| \\operatorname {s g} (\\hat {\\boldsymbol {e}}) - \\boldsymbol {e} \\right\\| _ {2} ^ {2}, \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.612, + 0.47, + 0.673 + ], + "angle": 0, + "content": "where \\(\\mathfrak{sg}[\\cdot ]\\) stands for the stop gradient operator that prevents the gradient from backpropagating through its operand. The first term \\(\\mathcal{L}_m(\\mathcal{M},\\hat{\\mathcal{M}})\\) represents the reconstruction error, which is expanded as:" + }, + { + "type": "equation", + "bbox": [ + 0.086, + 0.697, + 0.469, + 0.737 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} _ {m} (\\mathcal {M}, \\hat {\\mathcal {M}}) = | | \\mathcal {M}, \\hat {\\mathcal {M}} | | _ {1} \\tag {5} \\\\ + \\alpha_ {1} | | \\mathcal {M} ^ {\\prime}, \\hat {\\mathcal {M}} ^ {\\prime} | | _ {1} + \\alpha_ {2} | | \\mathcal {M} ^ {\\prime \\prime}, \\hat {\\mathcal {M}} ^ {\\prime \\prime} | | _ {1}, \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.47, + 0.901 + ], + "angle": 0, + "content": "where \\(\\mathcal{M}'\\) and \\(\\mathcal{M}''\\) are the \\(1^{st}\\) order and \\(2^{nd}\\) order of partial derivatives of motion sequence \\(\\mathcal{M}\\), representing the velocity and the acceleration of the motion, respectively. \\(\\alpha_{1}\\) and \\(\\alpha_{2}\\) are hyper-parameters used to trade-off between the two losses. We can see from Eq. (5) that VQ-VAE reconstructs not only the original position of each joints of the character, but also the velocities and the accelerations of its motions. Notably, since \\(\\mathcal{Z}\\) is discrete, we will simply copy the gradients from the reconstruction error and pass them to the encoder during training. The second term in Equation (4) aims" + }, + { + "type": "image", + "bbox": [ + 0.51, + 0.117, + 0.847, + 0.285 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.298, + 0.892, + 0.341 + ], + "angle": 0, + "content": "Figure 3. Summary of the reward model. The model jointly trains a motion encoder and an audio encoder to predict the correct pairings of a batch of (motion, audio) training examples." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.367, + 0.892, + 0.518 + ], + "angle": 0, + "content": "to optimize the codebook by pushing the \\(\\hat{e}\\) and its associated \\(z\\) close to \\(e\\), which is the output of the encoder. The last term in Equation (4) aims to optimize the encoder by pushing \\(e\\) close to its nearest latent vector in the codebook. The VQ-VAE mode will be trained separately, which means that its model parameters as well as the codebook remain unchanged during reward computation and reinforcement learning. Note that in the reinforcement learning phase, the indexes of the codebook will be treated as the action tokens, as introduced in Sec. 4.3." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.53, + 0.655, + 0.546 + ], + "angle": 0, + "content": "4.2. Reward Design" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.553, + 0.892, + 0.794 + ], + "angle": 0, + "content": "Reward function is key to reinforcement learning because it contains most of the knowledge that the agent can learn. Regarding to the task of co-speech gesture synthesis, the original dataset explicitly indicates the ground-truth gestures for given speech audios. However, due to the \"many-to-many\" nature of the problem, a ground-truth gesture may not be optimal under different contexts. This motivates us to learn a reward model that is able to discover and utilize deeper relations between speeches and gestures. There are two advantages of using the reward model. First, if there are multiple gestures correspond to one speech slice, the reward model is able to distinguish between these gestures so that to enrich the supervision signals. Second, as the goal of RL is to maximize the accumulated rewards of action sequences, the agent will learn to generate sequences of gestures that maximize the overall satisfaction." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.796, + 0.892, + 0.901 + ], + "angle": 0, + "content": "To get a proper reward function, we train a model to evaluate the degree of correspondence between the sequences of gestures and the sequences of speeches. Since the reward model tells the agent the degree of correspondence of the speech-gesture pairs, it should be able to distinguish between matched pairs and unmatched pairs. Moreover, the reward model should also generalize to speech-gesture pairs" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2334" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.168 + ], + "angle": 0, + "content": "that are not covered by the dataset. This coincides with the idea of contrastive learning. Multimodal contrastive learning techniques have been developed recently, such as CLIP [27], a well-known visual-text contrastive learning model, designed to match paired image and caption embeddings." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.168, + 0.471, + 0.288 + ], + "angle": 0, + "content": "Similar to CLIP, given a batch of \\( n \\) (motion, audio) pairs, the reward model is trained to predict which of the \\( n \\times n \\) possible (motion, audio) pairs across a batch actually occurred. To do this, the model converts each input motion \\( \\mathcal{M} \\) and audio \\( \\mathcal{U} \\) into vector representations \\( \\boldsymbol{m} \\) and \\( \\boldsymbol{u} \\) by a motion encoder \\( \\mathcal{E}_m \\) and an audio encoder \\( \\mathcal{E}_a \\). Two encoders are jointly trained to maximize the cosine similarity of each real pair \\( (m_i \\) and \\( u_i) \\):" + }, + { + "type": "equation", + "bbox": [ + 0.194, + 0.299, + 0.47, + 0.318 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {R} = \\ell_ {i} ^ {\\boldsymbol {m} \\rightarrow \\boldsymbol {u}} + \\ell_ {i} ^ {\\boldsymbol {u} \\rightarrow \\boldsymbol {m}}, \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.328, + 0.124, + 0.342 + ], + "angle": 0, + "content": "where" + }, + { + "type": "equation", + "bbox": [ + 0.14, + 0.35, + 0.408, + 0.39 + ], + "angle": 0, + "content": "\\[\n\\ell_ {i} ^ {\\boldsymbol {m} \\rightarrow \\boldsymbol {u}} = - \\log \\frac {e x p (m _ {i} \\cdot u _ {i} / \\tau)}{\\sum_ {k = 1} ^ {n} e x p (m _ {i} \\cdot u _ {k} / \\tau)},\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.143, + 0.387, + 0.405, + 0.42 + ], + "angle": 0, + "content": "\\[\n\\ell_ {i} ^ {\\boldsymbol {u} \\rightarrow \\boldsymbol {m}} = - \\log \\frac {\\exp (u _ {i} \\cdot m _ {i} / \\tau)}{\\sum_ {k = 1} ^ {n} \\exp (u _ {i} \\cdot m _ {k} / \\tau)}.\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.43, + 0.469, + 0.476 + ], + "angle": 0, + "content": "\\(\\tau \\in \\mathbb{R}^{+}\\) represents a temperature parameter. In this work, both encoders are structured as 1-D temporal convolutional networks and can input sequences of different lengths." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.476, + 0.47, + 0.552 + ], + "angle": 0, + "content": "As with VQ-VAE, the training of the reward model is in a separate phase. During reinforcement learning, two encoders of the learned reward model respectively encode the generated motion \\(\\hat{\\mathcal{M}}\\) and input speech \\(\\mathcal{U}\\) into two vectors, with their inner product as the reward:" + }, + { + "type": "equation", + "bbox": [ + 0.173, + 0.564, + 0.469, + 0.583 + ], + "angle": 0, + "content": "\\[\nR (\\hat {\\mathcal {M}}, \\mathcal {U}) = \\mathcal {E} _ {m} (\\hat {\\mathcal {M}}) \\cdot \\mathcal {E} _ {u} (\\mathcal {U}). \\tag {8}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.591, + 0.362, + 0.607 + ], + "angle": 0, + "content": "4.3. Offline Reinforcement Learning" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.614, + 0.47, + 0.795 + ], + "angle": 0, + "content": "Recall that we have defined several essential factors of the MDP. State \\(s\\) includes the generated action tokens \\(\\pmb{a}_{0:t-1}\\) and the audio sequence \\(\\pmb{u}_{0:t}\\) up to the current time step \\(t\\). Each action \\(a\\) corresponds to an index of the codebook learned by VQ-VAE. The reward \\(r\\) is outputted by the reward model to evaluate how well the predicted gesture matches the input audio. Now we are able to augment the original speech-gesture dataset to a trajectory dataset \\(\\mathcal{D} = \\{s_i, a_i, r_i, s_i' | i = 1, \\dots, N\\}\\). Note that these trajectories can be collected before the start of the reinforcement learning phase. We denote by \\(\\pi_{\\beta}\\) the behavior policy that generates these trajectories." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.796, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Offline datasets usually do not provide complete state-action coverage. That is, the set \\(\\{(s,a)|(s,a,r,s')\\in \\mathcal{D}\\}\\) is typically a small subset of the full space \\(S\\times \\mathcal{A}\\). Standard online reinforcement learning methods such as DDPG [24] and SAC [31] are not appropriate due to issues with bootstrapping from out-of-distribution actions. A \"safe\" strategy to mitigate the distributional shift problem is to be con" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.242 + ], + "angle": 0, + "content": "servative: if we explicitly estimate the value of unseen outcomes conservatively (i.e. assign them a low value), then the estimated value or performance of the policy that executes unseen behaviors is guaranteed to be small. Using such conservative estimates for policy optimization will prevent the policy from executing unseen actions and it will perform reliably. To this end, we adopt a simple but effective offline reinforcement learning algorithm called conservative Q-learning (CQL) [20] to learn the optimal policy from the augmented offline dataset." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.243, + 0.892, + 0.289 + ], + "angle": 0, + "content": "Specifically, for the conservative off-policy evaluation, the Q-function, \\(\\hat{Q}^{\\pi} := \\lim_{k \\to \\infty} \\hat{Q}_k\\) is trained via an iterative update:" + }, + { + "type": "equation", + "bbox": [ + 0.508, + 0.309, + 0.892, + 0.404 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\hat {Q} ^ {k + 1} \\leftarrow \\arg \\min _ {Q} \\frac {1}{2} \\underset {s, a, s ^ {\\prime} \\sim \\mathcal {D}} {\\mathbb {E}} \\left[ (Q (s, a) - \\hat {\\mathcal {B}} ^ {\\pi} \\hat {Q} ^ {k} (s, a)) ^ {2} \\right] \\\\ + \\alpha \\left(\\underset {a \\sim \\mu (a | s)} {\\mathbb {E}} [ Q (s, a) ] - \\underset {a \\sim \\hat {\\pi} _ {\\beta} (a | s)} {\\mathbb {E}} [ Q (s, a) ]\\right), \\tag {9} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.501, + 0.405, + 0.546, + 0.417 + ], + "angle": 0, + "content": "where" + }, + { + "type": "equation", + "bbox": [ + 0.5, + 0.427, + 0.91, + 0.459 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathcal {B}} ^ {\\pi} \\hat {Q} ^ {k} (s, a) = \\mathbb {E} _ {s ^ {\\prime} \\sim P \\left(s ^ {\\prime} \\mid s, a\\right)} [ R (s, a) + \\gamma \\mathbb {E} _ {a ^ {\\prime} \\sim \\hat {\\pi} ^ {k} (s, a)} \\hat {Q} ^ {k} \\left(s ^ {\\prime}, a ^ {\\prime}\\right) ] \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.461, + 0.892, + 0.596 + ], + "angle": 0, + "content": "is the Bellman operator. This equation consists of the normal bootstrapping error and the regularization term with a trade-off factor \\(\\alpha\\). Inside the regularization term, the first term always pushes the Q-value down on the \\((s,a)\\) pairs sampled from the learning policy \\(\\mu\\) whereas the second term pushes Q-value up on the \\((s,a)\\) pairs sampled from the offline data set. In practice, we selected the binding variant of CQL and DQN as our model, which can be optimized by minimizing the following loss function:" + }, + { + "type": "equation", + "bbox": [ + 0.518, + 0.617, + 0.891, + 0.704 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\underset {Q} {\\operatorname {m i n}} \\alpha_ {s \\sim \\mathcal {D}} \\left[ \\log \\sum_ {a} \\exp (Q (s, a)) - \\underset {a \\sim \\hat {\\pi} _ {\\beta}} {\\mathbb {E}} [ Q (s, a) ] \\right] \\\\ + \\frac {1}{2} _ {s, a, s ^ {\\prime} \\sim \\mathcal {D}} \\mathbb {E} \\left[ (Q (s, a) - \\hat {\\mathcal {B}} ^ {\\pi} \\hat {Q} ^ {k} (s, a)) ^ {2} \\right]. \\tag {11} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.705, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Since we want to use the Q-value function to autoregressively generate the action sequences, we construct a powerful GPT-based backbone model to estimate the Q-values, as is shown in Fig. 1. To construct the input, each motion code sequence is embedded to learnable features and concatenated with the audio features in the temporal dimension. We add a learned positional encoding to this concatenated \\((2\\times T^{\\prime}\\times C)\\)-dimensional tensor and feed it to Transformer layers, whose structure is shown in the left of Fig. 1. Last, we employ a linear layer to map the output of Transformer layers to the Q-value \\(q\\in \\mathbb{R}^{T^{\\prime}\\times N}\\), where \\(N\\) is the size of learned codebook. In the Transformer layer, the attention is the core component that determines the computational" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2335" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.122 + ], + "angle": 0, + "content": "dependency among sequential elements of inputs, which is computed as:" + }, + { + "type": "equation", + "bbox": [ + 0.088, + 0.142, + 0.469, + 0.188 + ], + "angle": 0, + "content": "\\[\n\\operatorname {A t t e n t i o n} (\\boldsymbol {Q}, \\boldsymbol {K}, \\boldsymbol {V}, \\boldsymbol {M}) = \\operatorname {s o f t m a x} \\left(\\frac {\\boldsymbol {Q} \\boldsymbol {K} ^ {T}}{\\sqrt {C}} + \\boldsymbol {M}\\right) \\boldsymbol {V}, \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.189, + 0.469, + 0.339 + ], + "angle": 0, + "content": "where \\(Q, K, V\\) denote the query, key and value from input, and \\(V\\) is the mask. In the original GPT model, \"causal attention\" realizes the intercommunication of the current and previous data to compute the state for the time of interest. However, our model needs to infer actions from the speech at each time step, so it needs to process two input sequences simultaneously. To comply the causality conditioned among features of the audio and the action, \\(M\\) is designed to be a \\(2 \\times 2\\) repeated block matrix with lower triangular matrix of size \\(T'\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.34, + 0.469, + 0.416 + ], + "angle": 0, + "content": "During inference, given the audio of arbitrary length and the initial motion code \\( a_0 \\), the Q-network can output Q-values of all actions. We will select the action with the maximum Q-value and then add it to the input, so that we can autoregressively generate the entire action sequence." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.429, + 0.21, + 0.445 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.453, + 0.173, + 0.468 + ], + "angle": 0, + "content": "5.1. Dataset" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.477, + 0.47, + 0.749 + ], + "angle": 0, + "content": "In this paper, we use two speech-gesture datasets: the Trinity dataset and a Chinese dataset collected by us. Trinity Gesture dataset is a large database of speech and gestures jointly collected by [8]. This dataset consists of 242 minutes of motion capture and audio of one male actor talking on different topics. The actor's motion was captured with a 20-camera Vicon system and solved onto a skeleton with 69 joints. In this paper, we use the official release version of The GENEA Challenge 2020 Dataset [19], where 221 minutes are used as training data, and the remaining 21 minutes are kept for testing. Additionally, we collected an 1-hour high-quality Chinese dataset using motion capture equipment. This dataset contains 3D full-body gestures and the aligned Chinese speech audio. And we retargeted the skeletons to be consistent with the Trinity dataset. In this dataset, the speaker described various speech scenes in Chinese and made high-standard common speech gestures, ensuring a high degree of matching between gestures and speech." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.757, + 0.295, + 0.773 + ], + "angle": 0, + "content": "5.2. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.781, + 0.469, + 0.9 + ], + "angle": 0, + "content": "All the motion data are downsampled to 20 frames per second on the Trinity and Chinese datasets. We focus on upper-body gestures in this work. For visualization, we employ a character model consist of 16 upper-body joints, including a rotational root. The audio features are extracted by the public audio processing toolbox librosa, including mel-scaled spectrogram and onset. In VQ-VAE, the codebook size \\(N\\) and the channel dimension \\(C\\) are both set to" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.469 + ], + "angle": 0, + "content": "512, and the temporal downsampling rate of encoders is set to 8. During the training of VQ-VAE, motion data are cropped to uniform length of \\( T = 120 \\) (6 seconds) and sampled in batch size of 64. The trade-off \\( \\beta \\) in \\( \\mathcal{L}_{VQ} \\) is set to 0.1. \\( \\alpha_{1} \\) and \\( \\alpha_{2} \\) in \\( \\mathcal{L}_m \\) are both 1. We adopt Adam optimizer with \\( \\beta_{1} = 0.9 \\) and \\( \\beta_{2} = 0.999 \\) for 500 epochs on Trinity but 400 epochs on the Chinese dataset with learning rate 0.00003. During the training of reward model, both audio and motion data are clipped to 6s. The optimizer settings are consistent with those of the VQ-VAE. The structure of the Q network is basically the same as that of the GPT network, where the embedding dimension is 768 and the attention layer is implemented in 12 heads with dropout probability 0.1. During the off-line reinforcement learning, the motion sequences are encoded to motion codes and sampled to length of 15 and sampled in batch size of 256. The Q-network is optimized using Adam optimizer with \\( \\beta_{1} = 0.5 \\) and \\( \\beta_{2} = 0.999 \\) for 1500 epochs on Trinity but 1200 epochs on Chinese dataset, where the learning rate is initialized as 0.0001. The discount factor \\( \\gamma \\) is set to 0.99. At runtime, a Gaussian filter with a kernel size of \\( K = 5 \\) is used to smooth the denormalized gesture sequence along the time dimension. The entire framework takes 36 hours on Trinity and 30 hours on the Chinese dataset to complete training on one NVIDIA Quadro RTX 5000 GPU." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.476, + 0.684, + 0.491 + ], + "angle": 0, + "content": "5.3. Evaluation Metrics" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.499, + 0.892, + 0.771 + ], + "angle": 0, + "content": "We adopt four commonly used evaluation metrics (MAJE, MAD, FGD and PMB) to compare various methods quantitatively. Mean absolute joint error (MAJE) measures the mean of the absolute errors between the generated joint positions and ground truth over all the time steps, which indicates how closely the generated joint positions follow the ground truth. Mean acceleration difference (MAD) measures the mean of the \\( l2 \\) norm differences between the generated joint accelerations and ground truth over all the time steps, which indicates how closely the ground truth and the generated joint movements match. Fréchet Gesture Distance (FGD) was proposed by which measures the difference between the distributions of the latent features of the generated gestures and ground truth [38]. The latent features are extracted from an auto-encoder trained on the Human 3.6M dataset*. Specifically, we use \\( \\hat{H} \\) and \\( H \\) represent the latent features of generated gestures and ground truth. The Fréchet distance is formulated as" + }, + { + "type": "equation", + "bbox": [ + 0.52, + 0.788, + 0.891, + 0.827 + ], + "angle": 0, + "content": "\\[\n\\operatorname {F G D} (\\hat {H}, H) = \\left| \\left| \\hat {\\mu} - \\mu \\right| \\right| _ {2} ^ {2} + \\operatorname {T r} \\left(\\hat {\\Sigma} + \\Sigma - 2 \\sqrt {\\hat {\\Sigma} \\Sigma}\\right), \\tag {13}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.827, + 0.892, + 0.857 + ], + "angle": 0, + "content": "where \\(\\mu\\) and \\(\\Sigma\\) are the first and second moments of the latent feature distribution of the ground truth \\(H\\), and \\(\\hat{\\mu}\\) and \\(\\hat{\\Sigma}\\) are" + }, + { + "type": "page_footnote", + "bbox": [ + 0.5, + 0.863, + 0.892, + 0.901 + ], + "angle": 0, + "content": "*We obtain the codes and models from: https://github.com/ai4r/Gesture-Generation-from-Trimodal-Context.git" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2336" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.089, + 0.892, + 0.119 + ], + "angle": 0, + "content": "Table 1. Comparison of RACER to Style Gesture, Gesticulator, S2AG and Bailando on Trinity datasets. \\( \\downarrow \\) means the lower is better and \\( \\uparrow \\) means the higher is better." + }, + { + "type": "table", + "bbox": [ + 0.164, + 0.13, + 0.807, + 0.304 + ], + "angle": 0, + "content": "
DatasetMethodMAJE(mm) ↓MAD(mm/s2) ↓FGD↓PMB(%) ↑
TrinityGround Truth00095.74
Style Gesture97.294.2636.9854.54
Gesticulator82.413.6231.0471.00
S2AG54.931.4920.3679.53
Bailando61.891.7417.2984.21
Ours50.331.2113.4489.58
Chinese DatasetStyle Gesture62.953.4630.1153.56
Bailando45.451.2925.0768.53
Ours43.251.199.2174.29
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.333, + 0.473, + 0.455 + ], + "angle": 0, + "content": "the first and second moments of the latent feature distribution of the generated gestures \\(\\hat{H}\\). FGD makes a reasonable assessment of the perceived plausibility of the synthesized gestures. The Percentage of Matched Beats PMB is a new metric to calculating the matching rate of audio and motion beats for evaluating rhythm performance [3]. In PMB, a motion beat is considered an matched if the distance to a nearby audio beat is within the range of \\(\\delta\\):" + }, + { + "type": "equation", + "bbox": [ + 0.088, + 0.464, + 0.469, + 0.52 + ], + "angle": 0, + "content": "\\[\n\\operatorname {P M B} \\left(B ^ {m}, B ^ {a}\\right) = \\frac {1}{N _ {m}} \\sum_ {i = 1} ^ {N _ {m}} \\sum_ {j = b _ {*} ^ {a} + 1} ^ {N _ {a}} \\mathbb {1} \\left[ \\left| \\left| b _ {j} ^ {m} - b _ {j} ^ {a} \\right| \\right| _ {1} < \\delta \\right], \\tag {14}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.521, + 0.472, + 0.599 + ], + "angle": 0, + "content": "where \\( B^{m} \\) and \\( B^{a} \\) represent the motion beats and audio beats, and \\( N_{m} \\) and \\( N_{a} \\) represent the number of beats. Motion beat can be obtained by identifying the local minima of joints deceleration [14]. \\( b_{*}^{a} \\) indicates the last matched audio beat. The \\( \\delta \\) is set to 0.2s, consistent with the original paper." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.606, + 0.369, + 0.623 + ], + "angle": 0, + "content": "5.4. Comparison to Existing Methods" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.629, + 0.473, + 0.903 + ], + "angle": 0, + "content": "In this section, we compare RACER with several state-of-the-art methods to demonstrate the advances made by our method. On the Trinity dataset, we compare with the methods of Style Gesture [2], Gesticulator [18], Speed to Affective Gesture (S2AG) [4] and Bailando [29]. On the Chinese dataset, we compared with Style Gesture and Bailando. Style Gesture generates gestures based on only the speech audio features by a normalizing flow probabilistic model. Gesticulator leverages the audio and text of the speech to generate semantically consistent gestures. Compared with these methods, S2AG adds more modality, including speaker identity and seed gesture poses. Different from other methods, Bailando is a novel music-to-dance framework with a choreographic memory and actor-critic GPT model. Since Bailando's model architecture is similar to ours and both incorporate reinforcement learning into the training process, we apply it as a key contrast method to the gesture synthesis task. The hyperparameter setting is" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.333, + 0.892, + 0.395 + ], + "angle": 0, + "content": "consistent with the original paper and the open source code, and the feature processing of the training data is consistent with ours. For a fair comparison, we keep the same skeleton and motion frame rate as the comparison systems." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.398, + 0.893, + 0.686 + ], + "angle": 0, + "content": "Quantitative experimental results are listed in Tab. 1. According to the comparison, our proposed model consistently performs favorably against all the other methods on all evaluations. Specially, on Trinity, our method improves \\(4.60(8.37\\%)\\) and \\(0.28(18.79\\%)\\) than the best compared baseline model S2AG on MAJE and MAD, respectively. In terms of FGD and PMB, Bailando, with its advanced network structure and reinforcement learning fine-tuning process, also obtained better evaluation results (17.29 and \\(84.21\\%\\) ) than S2AC (20.36 and \\(79.53\\%\\)). Our approach further achieves significantly higher performance (13.44 and \\(89.58\\%\\)) in both items. Similarly, our method has also achieved excellent performance in Chinese dataset compared with Style Gesture and Bailando, especially on FGD (9.21) and PMB \\((74.29\\%)\\). The superiority of our method on all evaluation metrics reveals that RACER can rely on offline RL not only to synthesize more real-like motion than compared baseline methods, but also but also to achieve outstanding performance in motion quality and rhythm." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.689, + 0.895, + 0.902 + ], + "angle": 0, + "content": "User Study. We conducted a user study to further assess the real visual performance of our method in a pairwise manner. We choose Style Gesture and Bailando to compare. We randomly sampled 10 30-second long slices generated from the test set of Trinity dataset. Each motion slice is bound to the Mixamo [25] model and rendered as a video by Blender. The experiment is conducted with 12 participants. The participants are asked to rate the slices from the following three aspects respectively: (1) realism, (2) speech-to-gesture matching, (3) speech-to-gesture rhythm matching. The scores assigned to each rating are in the range of 1-5, corresponding from worst to best. Rating statistics results are shown in Fig. 4. Notably, our method significantly surpasses the compared other methods." + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2337" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.327, + 0.09, + 0.644, + 0.104 + ], + "angle": 0, + "content": "Table 2. Ablation study results on the Trinity dataset." + }, + { + "type": "table", + "bbox": [ + 0.134, + 0.115, + 0.839, + 0.23 + ], + "angle": 0, + "content": "
MethodMAJE(mm) ↓MAD(mm/s2) ↓FGD↓PMB(%) ↑Reward↑
RACER(Supervised Learning)56.221.8116.6978.54-
RACER(Distance Reward)58.962.9423.5572.09-
RACER(DQN)59.441.8115.4676.634.23
RACER(Actor-Critic)60.181.7617.6978.023.93
Ours50.331.2113.4489.589.34
" + }, + { + "type": "image", + "bbox": [ + 0.084, + 0.269, + 0.468, + 0.401 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.413, + 0.471, + 0.468 + ], + "angle": 0, + "content": "Figure 4. User study results comparing RACER against Bailando and Style Gesture. Bars with different colors represent rank counts. In total, 120 comparisons have been rated. The total scores are listed on the right." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.497, + 0.241, + 0.512 + ], + "angle": 0, + "content": "5.5. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.522, + 0.47, + 0.822 + ], + "angle": 0, + "content": "To gain more insights into the proposed components of our method, we test some variants of RACER on Trinity dataset: (1) \"Supervised Learning\" indicates that we directly treat gesture code sequences as labels to train GPT-like models through supervised learning instead of reinforcement learning. (2) \"Distance reward\" indicates that our reward function is replaced by the Euclidean distance calculation between the predicted poses and ground truth, which is equivalent to taking the supervision signal in the traditional regression problem directly as the reward. (3) \"DQN\" means that we use the online RL algorithm DQN [34] instead of offline RL to learn the policy. The Q-value network of RACER(DQN) keeps consistent with ours. (4) \"Actor-Critic\" means that we use the online RL algorithm Actor-Critic instead of offline RL to learn the policy. For RACER(Actor-Critic), we regard the last 3 Transformer layers of GPT model with a additional linear-softmax layer as the branch of actor network. Besides, we add 3 Transformer layers as the branch of critic network to estimate the V-values." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.47, + 0.903 + ], + "angle": 0, + "content": "The quantitative scores are shown in Tab. 2. Firstly, we explore the effectiveness of offline reinforcement learning itself. Compared to RACER, the evaluation metrics of RACER(Supervised Learning) sharply drops \\(5.89(11\\%)\\), \\(0.60(50\\%)\\), \\(3.25(24\\%)\\) and \\(11.04\\%\\), respectively. This" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.259, + 0.895, + 0.516 + ], + "angle": 0, + "content": "shows that offline RL with proposed reward function can help the model learn to generate sequences of discrete tokens better than supervised learning. Secondly, to explore the advantages of the contrastive multi-modal reward, we evaluate the performance of RACER(Distance Reward). Our method has achieved better performance, especially on the MAD and PMB matrices. This confirms that the contrastive multi-modal reward function we propose can handle the feature relationship between multiple modes and improve the matching degree of motion and speech. Moreover, to verify the advantages of offline RL over online RL, we introduce two classic online RL algorithms, DQN and Actor-Critic, to determine the variations. Under the same number of training steps, DQN and AC algorithms can only get the highest rewards of 4.23 and 3.93 respectively, which is a significant gap compared to CQL. Naturally, other matrices are also significantly worse than CQL." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.54, + 0.749, + 0.557 + ], + "angle": 0, + "content": "6. Discussion and Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.569, + 0.895, + 0.903 + ], + "angle": 0, + "content": "In this paper, we propose a novel co-speech gesture synthesis framework called RACER. RACER regards the motion generated procedure as a sequential decision making process and employs reinforcement learning to learn the autoregressive gesture synthesis policy. There are mainly three sub-components of RACER. First, a vector quantized variational autoencoder (VQ-VAE) is employed to learn compact representations of gestures from massive real-world gesture data. Second, RACER uses a GPT-based architecture as policy network, which could autoregressively generate the next gestures given a sequence of speech audio. Thirdly, we propose a contrastive pre-training approach to calculate the rewards, which integrates contextual information into action evaluation and successfully captures the complex relationships between multi-modal speech-gesture data. We conduct extensive experiments to evaluate the performance of RACER. The experimental results show that RACER significantly outperforms the existing baselines in various objective and subjective metrics. Moreover, RACER shows a great generalization ability with respect to different contexts, which indicates the potentials of RACER in more complex motion synthesis tasks." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2338" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.116, + 0.47, + 0.186 + ], + "angle": 0, + "content": "[1] Chaitanya Ahuja, Dong Won Lee, Ryo Ishii, and Louis-Philippe Morency. No gestures left behind: Learning relationships between spoken language and freeform gestures. In Findings of the Association for Computational Linguistics: EMNLP 2020, pages 1884–1895, Nov. 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.187, + 0.472, + 0.243 + ], + "angle": 0, + "content": "[2] Simon Alexanderson, Gustav Eje Henter, Taras Kucherenko, and Jonas Beskow. Style-controllable speech-driven gesture synthesis using normalising flows. Computer Graphics Forum, 39, 2020. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.245, + 0.472, + 0.3 + ], + "angle": 0, + "content": "[3] Tenglong Ao, Qingzhe Gao, Yuke Lou, Baoquan Chen, and Libin Liu. Rhythmic gesticulator: Rhythm-aware cospeech gesture synthesis with hierarchical neural embeddings. ArXiv, abs/2210.01448, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.302, + 0.472, + 0.386 + ], + "angle": 0, + "content": "[4] Uttaran Bhattacharya, Elizabeth Childs, Nicholas Rewkowski, and Dinesh Manocha. Speech2affectivegestures: Synthesizing co-speech gestures with generative adversarial affective expression learning. In Proceedings of the 29th ACM International Conference on Multimedia, page 2027-2036, 2021. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.388, + 0.472, + 0.47 + ], + "angle": 0, + "content": "[5] Uttaran Bhattacharya, Nicholas Rewkowski, Abhishek Banerjee, Pooja Guhan, Aniket Bera, and Dinesh Manocha. Text2gestures: A transformer-based network for generating emotive body gestures for virtual agents. In 2021 IEEE Virtual Reality and 3D User Interfaces (VR), pages 1-10, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.472, + 0.472, + 0.528 + ], + "angle": 0, + "content": "[6] Chung-Cheng Chiu, Louis-Philippe Morency, and Stacy Marsella. Predicting co-verbal gestures: A deep and temporal modeling approach. In Intelligent Virtual Agents, pages 152–166, 2015. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.53, + 0.472, + 0.586 + ], + "angle": 0, + "content": "[7] Hsu-Kuang Chiu, Ehsan Adeli, Borui Wang, De-An Huang, and Juan Carlos Niebles. Action-agnostic human pose forecasting. In 2019 IEEE Winter Conference on Applications of Computer Vision (WACV), pages 1423–1432, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.588, + 0.472, + 0.643 + ], + "angle": 0, + "content": "[8] Ylva Ferstl and Rachel McDonnell. Investigating the use of recurrent motion modelling for speech gesture generation. In Proceedings of the 18th International Conference on Intelligent Virtual Agents, page 93–98, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.645, + 0.472, + 0.7 + ], + "angle": 0, + "content": "[9] Ylva Ferstl, Michael Neff, and Rachel McDonnell. Multi-objective adversarial gesture generation. In Proceedings of the 12th ACM SIGGRAPH Conference on Motion, Interaction and Games, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.702, + 0.47, + 0.771 + ], + "angle": 0, + "content": "[10] Shiry Ginosar, Amir Bar, Gefen Kohavi, Caroline Chan, Andrew Owens, and Jitendra Malik. Learning individual styles of conversational gesture. In 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3492-3501, 2019. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.774, + 0.47, + 0.857 + ], + "angle": 0, + "content": "[11] Ikhsanul Habibie, Weipeng Xu, Dushyant Mehta, Lingjie Liu, Hans-Peter Seidel, Gerard Pons-Moll, Mohamed Elgharib, and Christian Theobalt. Learning speech-driven 3d conversational gestures from video. In Proceedings of the 21st ACM International Conference on Intelligent Virtual Agents, page 101-108, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.859, + 0.47, + 0.902 + ], + "angle": 0, + "content": "[12] Gustav Eje Henter, Simon Alexanderson, and Jonas Beskow. Moglow: Probabilistic and controllable motion synthesis using normalising flows. ACM Trans. Graph., 2020. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.116, + 0.472, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.148 + ], + "angle": 0, + "content": "[13] Alejandro Hernandez, Jurgen Gall, and Francesc Moreno-Noguer. Human motion prediction via spatio-temporal inpainting. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), October 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.15, + 0.892, + 0.219 + ], + "angle": 0, + "content": "[14] Chieh Ho, Wei-Tze Tsai, Keng-Sheng Lin, and Homer H. Chen. Extraction and alignment evaluation of motion beats for street dance. In 2013 IEEE International Conference on Acoustics, Speech and Signal Processing, pages 2429–2433, 2013. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.221, + 0.892, + 0.29 + ], + "angle": 0, + "content": "[15] Yu Hua, Fan Xuanzhe, Hou Yaqing, Liu Yi, Kang Cai, Zhou Dongsheng, and Zhang Qiang. Towards efficient 3d human motion prediction using deformable transformer-based adversarial network. In 2022 International Conference on Robotics and Automation (ICRA), pages 861-867, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.292, + 0.892, + 0.331 + ], + "angle": 0, + "content": "[16] Lucas Kovar, Michael Gleicher, and Frédéric Pighin. Motion graphs. In ACM SIGGRAPH 2008 Classes, SIGGRAPH '08, 2008. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.334, + 0.892, + 0.404 + ], + "angle": 0, + "content": "[17] Taras Kucherenko, Dai Hasegawa, Gustav Eje Henter, Naoshi Kaneko, and Hedvig Kjellström. Analyzing input and output representations for speech-driven gesture generation. In Proceedings of the 19th ACM International Conference on Intelligent Virtual Agents, page 97-104, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.406, + 0.892, + 0.489 + ], + "angle": 0, + "content": "[18] Taras Kucherenko, Patrik Jonell, Sanne van Waveren, Gustav Eje Henter, Simon Alexandersson, Iolanda Leite, and Hedvig Kjellström. Gesticulator: A framework for semantically-aware speech-driven gesture generation. In Proceedings of the 2020 International Conference on Multimodal Interaction, page 242-250, 2020. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.491, + 0.892, + 0.56 + ], + "angle": 0, + "content": "[19] Taras Kucherenko, Patrik Jonell, Youngwoo Yoon, Pieter Wolfert, and Gustav Eje Henter. A large, crowdsourced evaluation of gesture generation systems on common data: The genea challenge 2020. In 26th International Conference on Intelligent User Interfaces, page 11-21, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.562, + 0.892, + 0.644 + ], + "angle": 0, + "content": "[20] Aviral Kumar, Aurick Zhou, George Tucker, and Sergey Levine. Conservative q-learning for offline reinforcement learning. In H. Larochelle, M. Ranzato, R. Hadsell, M.F. Balcan, and H. Lin, editors, Advances in Neural Information Processing Systems, volume 33, pages 1179-1191. Curran Associates, Inc., 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.646, + 0.892, + 0.729 + ], + "angle": 0, + "content": "[21] Jing Li, Di Kang, Wenjie Pei, Xuefei Zhe, Ying Zhang, Zhenyu He, and Linchao Bao. Audio2gestures: Generating diverse gestures from speech audio with conditional variational autoencoders. In 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 11273-11282, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.731, + 0.892, + 0.8 + ], + "angle": 0, + "content": "[22] Ruilong Li, Shan Yang, David A. Ross, and Angjoo Kanazawa. Ai choreographer: Music conditioned 3d dance generation with aist++. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 13401-13412, October 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.802, + 0.892, + 0.87 + ], + "angle": 0, + "content": "[23] Yuanzhi Liang, Qianyu Feng, Linchao Zhu, Li Hu, Pan Pan, and Yi Yang. Seeg: Semantic energized co-speech gesture generation. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10463-10472, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.873, + 0.892, + 0.902 + ], + "angle": 0, + "content": "[24] Timothy P. Lillicrap, Jonathan J. Hunt, Alexander Pritzel, Nicolas Heess, Tom Erez, Yuval Tassa, David Silver, and" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2339" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.11, + 0.092, + 0.469, + 0.148 + ], + "angle": 0, + "content": "Daan Wierstra. Continuous control with deep reinforcement learning. In 4th International Conference on Learning Representations, ICLR 2016, San Juan, Puerto Rico, May 2-4, 2016, Conference Track Proceedings, 2016. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.15, + 0.346, + 0.163 + ], + "angle": 0, + "content": "[25] Mixamo. https://www MIXamo.com/. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.165, + 0.469, + 0.233 + ], + "angle": 0, + "content": "[26] Shenhan Qian, Zhi Tu, Yihao Zhi, Wen Liu, and Shenghua Gao. Speech drives templates: Co-speech gesture synthesis with learned templates. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 11077-11086, October 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.236, + 0.469, + 0.305 + ], + "angle": 0, + "content": "[27] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision, 2021. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.307, + 0.468, + 0.347 + ], + "angle": 0, + "content": "[28] Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, and Ilya Sutskever. Language Models are Unsupervised Multitask Learners. 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.35, + 0.469, + 0.431 + ], + "angle": 0, + "content": "[29] Li Siyao, Weijiang Yu, Tianpei Gu, Chunze Lin, Quan Wang, Chen Qian, Chen Change Loy, and Ziwei Liu. Bailando: 3d dance generation by actor-critic gpt with choreographic memory. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11050-11059, June 2022. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.434, + 0.469, + 0.516 + ], + "angle": 0, + "content": "[30] Yonglong Tian, Chen Sun, Ben Poole, Dilip Krishnan, Cordelia Schmid, and Phillip Isola. What makes for good views for contrastive learning? In H. Larochelle, M. Ranzato, R. Hadsell, M.F. Balcan, and H. Lin, editors, Advances in Neural Information Processing Systems, volume 33, pages 6827-6839. Curran Associates, Inc., 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.519, + 0.469, + 0.572 + ], + "angle": 0, + "content": "[31] Pieter Abbeel Sergey Levine Tuomas Haarnoja, Aurick Zhou. Soft actor-critic: Off-policy maximum entropy deep reinforcement learning with a stochastic actor. ArXiv, abs/1801.01290, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.576, + 0.469, + 0.643 + ], + "angle": 0, + "content": "[32] Aaron van den Oord, Oriol Vinyals, and Koray Kavukcuoglu. Neural discrete representation learning. In Proceedings of the 31st International Conference on Neural Information Processing Systems, NIPS'17, page 6309-6318, 2017. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.646, + 0.469, + 0.701 + ], + "angle": 0, + "content": "[33] Ruben Villegas, Jimei Yang, Duygu Ceylan, and Honglak Lee. Neural kinematic networks for unsupervised motion retargeting. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), June 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.703, + 0.469, + 0.757 + ], + "angle": 0, + "content": "[34] David Silver Alex Graves Ioannis Antonoglou Daan Wierstra Martin Riedmiller Volodymyr Mnih, Koray Kavukcuoglu. Playing atari with deep reinforcement learning. ArXiv, abs/1312.5602, 2013. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.76, + 0.469, + 0.8 + ], + "angle": 0, + "content": "[35] Petra Wagner, Zofia Malisz, and Stefan Kopp. Gesture and speech in interaction: An overview. Speech Communication, 57:209-232, 2014. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.803, + 0.469, + 0.857 + ], + "angle": 0, + "content": "[36] Jason R. Wilson, Nah Young Lee, Annie Saechao, Sharon Hershenson, Matthias Scheutz, and Linda Tickle-Degnen. Hand gestures and verbal acknowledgments improve humanrobot rapport. In Social Robotics, pages 334-344, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.86, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[37] Xinchen Yan, Akash Rastogi, Ruben Villegas, Kalyan Sunkavalli, Eli Shechtman, Sunil Hadap, Ersin Yumer, and Honglak Lee. Mt-vae: Learning motion transformations to" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.469, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.892, + 0.133 + ], + "angle": 0, + "content": "generate multimodal human dynamics. In Proceedings of the European Conference on Computer Vision (ECCV), September 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.136, + 0.892, + 0.204 + ], + "angle": 0, + "content": "[38] Youngwoo Yoon, Bok Cha, Joo-Haeng Lee, Minsu Jang, Jaeyeon Lee, Jaehong Kim, and Geehyuk Lee. Speech gesture generation from the trimodal context of text, audio, and speaker identity. ACM Trans. Graph., 39(6), nov 2020. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.207, + 0.892, + 0.287 + ], + "angle": 0, + "content": "[39] Youngwoo Yoon, Woo-Ri Ko, Minsu Jang, Jaeyeon Lee, Jae-hong Kim, and Geehyuk Lee. Robots learn social skills: End-to-end learning of co-speech gesture generation for humanoid robots. In 2019 International Conference on Robotics and Automation (ICRA), pages 4303-4309, 2019. 1" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.287 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.956 + ], + "angle": 0, + "content": "2340" + } + ] +] \ No newline at end of file diff --git a/2023/Co-Speech Gesture Synthesis by Reinforcement Learning With Contrastive Pre-Trained Rewards/6b1eace7-6844-4cc2-b39b-76192850eed9_origin.pdf b/2023/Co-Speech Gesture Synthesis by Reinforcement Learning With Contrastive Pre-Trained Rewards/6b1eace7-6844-4cc2-b39b-76192850eed9_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..82306c09bb47bb57c5b1cf858788f2a838a6af0a --- /dev/null +++ b/2023/Co-Speech Gesture Synthesis by Reinforcement Learning With Contrastive Pre-Trained Rewards/6b1eace7-6844-4cc2-b39b-76192850eed9_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:618434bf512d70feac793f185a368c13b5003f9538d115c0649c3d99c5128503 +size 1088789 diff --git a/2023/Co-Speech Gesture Synthesis by Reinforcement Learning With Contrastive Pre-Trained Rewards/full.md b/2023/Co-Speech Gesture Synthesis by Reinforcement Learning With Contrastive Pre-Trained Rewards/full.md new file mode 100644 index 0000000000000000000000000000000000000000..57160a500cb69d7c1674cf1a3661b802e880e088 --- /dev/null +++ b/2023/Co-Speech Gesture Synthesis by Reinforcement Learning With Contrastive Pre-Trained Rewards/full.md @@ -0,0 +1,285 @@ +# Co-speech Gesture Synthesis by Reinforcement Learning with Contrastive Pre-trained Rewards + +Mingyang Sun $^{1*}$ , Mengchen Zhao $^{2*}$ , Yaqing Hou $^{1\dagger}$ , Minglei Li $^{3}$ , Huang Xu $^{3}$ , Songcen Xu $^{2}$ , Jianye Hao $^{2,4}$ , Dalian University of Technology, China ${ }^{2}$ Noah's Ark Lab, Huawei, China ${ }^{3}$ Huawei Cloud Computing Technologies Co., Ltd, China ${ }^{4}$ Tianjin University, China + +mysun@mail.dlut.edu.cn, houyq@dlut.edu.cn + +{zhaomengchen,liminglei29,xuhuang2,xusongcen,haojianye}@huawei.com + +# Abstract + +There is a growing demand of automatically synthesizing co-speech gestures for virtual characters. However, it remains a challenge due to the complex relationship between input speeches and target gestures. Most existing works focus on predicting the next gesture that fits the data best, however, such methods are myopic and lack the ability to plan for future gestures. In this paper, we propose a novel reinforcement learning (RL) framework called RACER to generate sequences of gestures that maximize the overall satisfactory. RACER employs a vector quantized variational autoencoder to learn compact representations of gestures and a GPT-based policy architecture to generate coherent sequence of gestures autoregressively. In particular, we propose a contrastive pre-training approach to calculate the rewards, which integrates contextual information into action evaluation and successfully captures the complex relationships between multi-modal speech-gesture data. Experimental results show that our method significantly outperforms existing baselines in terms of both objective metrics and subjective human judgements. Demos can be found at https://github.com/RLracer/RACER.git. + +# 1. Introduction + +Gesturing is important for human speakers to improve their expressiveness. It conveys the necessary non-verbal information to the audience, giving the speech a more emotional touch so that to enhance persuasiveness and credibility. Similarly, in virtual world, high-quality 3D gesture animations can make a talking character more vividly. For example, in human-computer interaction scenarios, vivid gestures performed by virtual characters can help listeners to concentrate and improve the intimacy between humans and + +characters [36]. Attracted by these merits, there has been a growing demand of automatically synthesizing high-quality co-speech gestures in computer animation. + +However, automatically synthesizing co-speech gestures remains a challenge due to the complicated relationship between speech audios and gestures. On one hand, a speaker may play different gestures when speaking the same words due to different mental and physical states. On the other hand, a speaker may also play similar gestures when speaking different words. Therefore, co-speech gesture synthesizing is inherently a "many-to-many" problem [21]. Moreover, in order to ensure the overall fluency and consistency, we must take into consideration both the contextual information and the subsequent effect upon playing a gesture [23, 38]. Therefore, gesture synthesizing is rather a sequential decision making problem than a simple matching between speeches and gestures. + +Compared with traditional rule-based approaches [35], data-driven gesture synthesis approaches [10,39] has shown many advantages, including the low development cost and the ability to generalize. Most existing data-driven approaches consider gesture synthesis as a classification task [6, 26] or a regression task [17] in a deterministic way i.e. the same speech or text input always maps to the same gesture output. However, these models rely on the assumption that there exists a unique ground-truth label for each input sequence of speech, which contradicts to the "many-to-many" nature of the problem. As a consequence, they sacrifice diversity and semantics of gestures and tend to learn some averaged gestures for any input speeches. Some other works adopt adversarial learning framework, where a discriminator is trained to distinguish between generated gestures and the recorded gestures in dataset [10,38]. Although they improve the generalizability of the gesture generator to some extent, they still fail to explore the essential relationship between speeches and gestures. + +In order to address the above challenges, we propose + +a novel Reinforcement leArning framework with Contrasitive prE-trained Rewards (RACER) for generating high-quality co-speech gestures. RACER is trained in an offline manner but can be used to generate the next gesture for a speaking character in real time. RACER consists of the following three components. Firstly, in order to extract meaningful gestures from the infinite action space, RACER adopts a vector quantized variational autoencoder (VQ-VAE) [32] to learn compact gesture representations, which significantly reduces the action space. Secondly, we construct the Q-value network using a GPT-based [28] model, which has natural advantages of generating coherence sequence of gestures. Thirdly, inspired by the contrastive language-image pre-training (CLIP) [27] and recent advances of contrastive learning [30], we propose a contrastive speech-gesture pre-training method to compute the rewards, which guide the RL agent to explore sophisticated relations between speeches and gestures. Note that these rewards evaluate the quality of gestures as a sequence. By contrast, in conventional supervised learning frameworks, the focus is on predicting only the next gesture, ignoring the quality of generated sequence as a whole. To sum up, the core contributions of this paper are as follows: + +(1) We formally model the co-speech gesture synthesis problem as a Markov decision process and propose a novel RL based approach called RACER to learn the optimal gesture synthesis policy. RACER can be trained in an offline manner and used to synthesis co-speech gestures in real time. +(2) We introduce VQ-VAE to encode and quantize the motion segments to a codebook, which significantly reduces the action space and facilitates the RL phase. +(3) We propose a contrastive speech-gesture pre-training method to compute the rewards, which guide the RL agent to discover deeper relations from multi-modal speech-gesture data. +(4) Extensive experimental results show that RACER outperforms the existing baselines in terms of both objective metrics and subjective human judgements. This demonstrates the superiority and the potential of RL in co-speech gesture synthesis tasks. + +# 2. Related Work + +# 2.1. Data-driven Motion Synthesis + +Human motion synthesis has been studied for a long time [21, 37]. Traditional approaches such as motion graph [16] cut the motion capture data and synthesize new motions according to crafted transition rules. With the development of deep learning, using neural networks to synthesize motions in an end-to-end manner becomes popular. Some of + +works focus on motion prediction [13, 15], which could predict current motion with a high accuracy, but they often fail to predict longer motion sequences. Recent works focus more on the motion synthesis, which aims at generating long-term natural motions such as walking, jumping, waving and dancing. Many advanced techniques have been applied to long-term motion synthesis, including conditional GANs [33], VAEs [21], auto-regressive models [7] and Transformers [5, 22]. A most recent work called Bailando explores the potential of RL in synthesizing dances for given music [29]. However, in Bailando, RL is used only for fine tuning so that the exact contributions of RL remain unclear. By contrast, our proposed method RACER is trained solely by RL and its effectiveness is demonstrated by extensive experiments. + +# 2.2. Co-speech Gesture Synthesis + +Co-speech gesture synthesis is challenging due to the complicated relations between speeches and gestures that are hidden in the multi-modal dataset. One line of conventional methods is rule-based generation [35], where the core idea is to pre-define a set of gesture units and design rules to connect speech words and the gesture units. However, it requires too much human effort in designing the rules, which is costly and inefficient in practice. Benefit from the recent advances in deep learning, many data-driven approaches have been proposed to learn the matching rules from speech-gesture data. Existing works focus on exploring the effectiveness of different network architectures, including the multi-layer perceptron (MLP) [18], convolutional neural networks (CNNs) [11], recurrent neural networks (RNNs) [4] and Transformers [18, 29]. Some recent works incorporate adversarial loss to enhance gesture fidelity [1,9,10,38]. However, they simply construct a target gesture as the unique label of the current segment of speech, ignoring the "many-to-many" nature of the problem. To model the complicated relation between speeches and gestures, Alexanderson et al. [2] propose a probabilistic model MoGlow based on normalizing flows [12] which maps input speeches to Gaussian distributions of gestures. However, the randomly sampled gestures contain non-negligible noise and therefore suffer from low interpretability. + +# 3. Preliminaries + +The underlying environment of reinforcement learning is typically modeled by a Markov decision process (MDP), which is defined by a tuple $(S, A, P, R, \gamma)$ . $S$ represents the state space and $A$ represents the action space. The transition function $P(\cdot | s, a)$ gives the probabilities of transitioning to the next state after taking action $a$ at state $s$ . $R(\cdot | s, a)$ is the associated reward function and $\gamma \in [0, 1)$ represents the discount factor. A policy $\pi(\cdot | s)$ maps each state $s \in S$ to a distribution over actions. For an agent following the policy + +![](images/024374574dbb552d472bc6a0aa14c844fa433a13bf08e57e07ca6aa9589c87b1.jpg) +Figure 1. The overall architecture of RACER. Given a piece of speech audio and the initial gesture code $a_0$ , the Q-network represented by transformer layers autoregressively calculates the Q-values and selects a sequence of actions $(a_1, \dots, a_T)$ . The action sequence will then be transformed to quantitative features by querying the codebook and finally be decoded to motion sequences by the decoder of VQ-VAE. + +$\pi$ , the action value function, denoted by $Q^{\pi}(s,a)$ , is defined as the expectation of cumulative discounted future rewards: + +$$ +Q ^ {\pi} (s, a) := \mathbb {E} \left[ \sum_ {t = 0} \gamma^ {t} R \left(s _ {t}, a _ {t}\right) \right], \tag {1} +$$ + +$$ +s _ {0} = s, a _ {0} = a, s _ {t} \sim P (\cdot | s _ {t - 1}, a _ {t - 1}), a _ {t} \sim \pi (\cdot | s _ {t}). +$$ + +The goal in RL is to learn an optimal policy $\pi^{*}$ that maximizes the expected cumulative reward. We denote by $Q^{*}(s,a)$ the optimal action value function under the optimal policy $\pi^{*}$ . The Bellman equations for the optimal Q-values can be represented by: + +$$ +Q ^ {*} (s, a) := R (s, a) + \gamma \mathbb {E} _ {s ^ {\prime} \sim P} \max _ {a ^ {\prime} \in \mathcal {A}} Q ^ {*} \left(s ^ {\prime}, a ^ {\prime}\right). \tag {2} +$$ + +A standard approach for computing the optimal policy is Q-learning, which iteratively improves the estimation of $Q^{*}$ by applying the Bellman operator derived from Eq. (2). While traditional Q-learning requires interactions with the environment, recent advances in offline RL allow us to learn the optimal policy directly from a fixed dataset, which we will elaborate in Sec. 4.3. + +# 4. Our Approach + +The overview of our proposed framework RACER is shown in Fig. 1. At each time step $t$ , the state $s$ consists of the generated action tokens $(a_{1},\ldots ,a_{t - 1})$ and input audio. + +Unlike existing methods which directly learn a mapping from audio features to the continuous high-dimensional motion space, we encode and quantize the motion into a finite codebook $\mathcal{Z} = \{z_i\}_{i=1}^N$ by VQ-VAE, where $N$ is the size of codebook and each code $z_i$ represents a gesture lexeme feature. The details of action design are introduced in Sec. 4.1. We use a GPT-like unidirectional model as the Q-network that autoregressively outputs action tokens following a greedy strategy. An action token $a$ will be mapped to a gesture lexeme feature $z$ and then be decoded to a specific gesture motion. Moreover, we propose a contrastive speech-gesture pre-training model to compute the immediate rewards for the actions, which will be elaborated in Sec. 4.2. In addition, we will introduce how to train the Q-network in a fully offline manner in Sec. 4.3. + +# 4.1. Action Design + +Instead of designing handcrafted gestures with expensive manual efforts, our goal is to summarize gesture lexemes into a representative codebook from a large gesture dataset in an unsupervised manner. To collect distinctive gesture representations and reconstruct them back to real gestures, we employ a VQ-VAE model which has successfully demonstrated its ability to learn effective latent representations from temporal data. The structure of VQ-VAE is shown in Fig. 2. VQ-VAE extends the regular autoencoder by adding a discrete codebook $\mathcal{Z}$ to the network, which consists of a list of vectors with their corresponding indexes. A piece of motion in raw dataset can be represented by a ma + +![](images/62534f56cfd55598bcfb00c688e358e6566c967fad156f743763c34360e4349d.jpg) +Figure 2. Structure of motion VQ-VAE. + +trix $\mathcal{M} \in \mathbb{R}^{T \times J}$ , where $T$ is the number of frames and $J$ is the dimension of joint features. $\mathcal{M}$ can then be transformed to latent features $e \in \mathbb{R}^{T' \times C}$ by a 1-D convolutional network, where $T'$ is the length of feature sequence after down-sampling, and $C$ is the channel dimension. Each row of $e$ , represented by $e_i$ , will be substituted by its closest vector $z_j$ in the codebook: + +$$ +\hat {e} _ {i} = \arg \min _ {\boldsymbol {z} _ {j} \in \mathcal {Z}} | | \boldsymbol {e} _ {i} - \boldsymbol {z} _ {j} | |. \tag {3} +$$ + +Then, the quantized features $\hat{e}_i$ will be decoded to $\hat{\mathcal{M}}$ via a decoder so that to reconstruct $\mathcal{M}$ . Following [32], the encoder and decoder are simultaneously learned by minimizing the following loss function: + +$$ +\begin{array}{l} \mathcal {L} _ {V Q} = \mathcal {L} _ {m} (\mathcal {M}, \hat {\mathcal {M}}) \tag {4} \\ + \left\| \hat {\boldsymbol {e}} - \operatorname {s g} (\boldsymbol {e}) \right\| _ {2} ^ {2} + \beta \left\| \operatorname {s g} (\hat {\boldsymbol {e}}) - \boldsymbol {e} \right\| _ {2} ^ {2}, \\ \end{array} +$$ + +where $\mathfrak{sg}[\cdot ]$ stands for the stop gradient operator that prevents the gradient from backpropagating through its operand. The first term $\mathcal{L}_m(\mathcal{M},\hat{\mathcal{M}})$ represents the reconstruction error, which is expanded as: + +$$ +\begin{array}{l} \mathcal {L} _ {m} (\mathcal {M}, \hat {\mathcal {M}}) = | | \mathcal {M}, \hat {\mathcal {M}} | | _ {1} \tag {5} \\ + \alpha_ {1} | | \mathcal {M} ^ {\prime}, \hat {\mathcal {M}} ^ {\prime} | | _ {1} + \alpha_ {2} | | \mathcal {M} ^ {\prime \prime}, \hat {\mathcal {M}} ^ {\prime \prime} | | _ {1}, \\ \end{array} +$$ + +where $\mathcal{M}'$ and $\mathcal{M}''$ are the $1^{st}$ order and $2^{nd}$ order of partial derivatives of motion sequence $\mathcal{M}$ , representing the velocity and the acceleration of the motion, respectively. $\alpha_{1}$ and $\alpha_{2}$ are hyper-parameters used to trade-off between the two losses. We can see from Eq. (5) that VQ-VAE reconstructs not only the original position of each joints of the character, but also the velocities and the accelerations of its motions. Notably, since $\mathcal{Z}$ is discrete, we will simply copy the gradients from the reconstruction error and pass them to the encoder during training. The second term in Equation (4) aims + +![](images/cbc062a5b975db7d728e2d923065c654787eb6db9c9f177cb17cb0c69ce0daa1.jpg) +Figure 3. Summary of the reward model. The model jointly trains a motion encoder and an audio encoder to predict the correct pairings of a batch of (motion, audio) training examples. + +to optimize the codebook by pushing the $\hat{e}$ and its associated $z$ close to $e$ , which is the output of the encoder. The last term in Equation (4) aims to optimize the encoder by pushing $e$ close to its nearest latent vector in the codebook. The VQ-VAE mode will be trained separately, which means that its model parameters as well as the codebook remain unchanged during reward computation and reinforcement learning. Note that in the reinforcement learning phase, the indexes of the codebook will be treated as the action tokens, as introduced in Sec. 4.3. + +# 4.2. Reward Design + +Reward function is key to reinforcement learning because it contains most of the knowledge that the agent can learn. Regarding to the task of co-speech gesture synthesis, the original dataset explicitly indicates the ground-truth gestures for given speech audios. However, due to the "many-to-many" nature of the problem, a ground-truth gesture may not be optimal under different contexts. This motivates us to learn a reward model that is able to discover and utilize deeper relations between speeches and gestures. There are two advantages of using the reward model. First, if there are multiple gestures correspond to one speech slice, the reward model is able to distinguish between these gestures so that to enrich the supervision signals. Second, as the goal of RL is to maximize the accumulated rewards of action sequences, the agent will learn to generate sequences of gestures that maximize the overall satisfaction. + +To get a proper reward function, we train a model to evaluate the degree of correspondence between the sequences of gestures and the sequences of speeches. Since the reward model tells the agent the degree of correspondence of the speech-gesture pairs, it should be able to distinguish between matched pairs and unmatched pairs. Moreover, the reward model should also generalize to speech-gesture pairs + +that are not covered by the dataset. This coincides with the idea of contrastive learning. Multimodal contrastive learning techniques have been developed recently, such as CLIP [27], a well-known visual-text contrastive learning model, designed to match paired image and caption embeddings. + +Similar to CLIP, given a batch of $n$ (motion, audio) pairs, the reward model is trained to predict which of the $n \times n$ possible (motion, audio) pairs across a batch actually occurred. To do this, the model converts each input motion $\mathcal{M}$ and audio $\mathcal{U}$ into vector representations $\boldsymbol{m}$ and $\boldsymbol{u}$ by a motion encoder $\mathcal{E}_m$ and an audio encoder $\mathcal{E}_a$ . Two encoders are jointly trained to maximize the cosine similarity of each real pair $(m_i$ and $u_i)$ : + +$$ +\mathcal {L} _ {R} = \ell_ {i} ^ {\boldsymbol {m} \rightarrow \boldsymbol {u}} + \ell_ {i} ^ {\boldsymbol {u} \rightarrow \boldsymbol {m}}, \tag {6} +$$ + +where + +$$ +\ell_ {i} ^ {\boldsymbol {m} \rightarrow \boldsymbol {u}} = - \log \frac {e x p (m _ {i} \cdot u _ {i} / \tau)}{\sum_ {k = 1} ^ {n} e x p (m _ {i} \cdot u _ {k} / \tau)}, +$$ + +$$ +\ell_ {i} ^ {\boldsymbol {u} \rightarrow \boldsymbol {m}} = - \log \frac {\exp (u _ {i} \cdot m _ {i} / \tau)}{\sum_ {k = 1} ^ {n} \exp (u _ {i} \cdot m _ {k} / \tau)}. +$$ + +$\tau \in \mathbb{R}^{+}$ represents a temperature parameter. In this work, both encoders are structured as 1-D temporal convolutional networks and can input sequences of different lengths. + +As with VQ-VAE, the training of the reward model is in a separate phase. During reinforcement learning, two encoders of the learned reward model respectively encode the generated motion $\hat{\mathcal{M}}$ and input speech $\mathcal{U}$ into two vectors, with their inner product as the reward: + +$$ +R (\hat {\mathcal {M}}, \mathcal {U}) = \mathcal {E} _ {m} (\hat {\mathcal {M}}) \cdot \mathcal {E} _ {u} (\mathcal {U}). \tag {8} +$$ + +# 4.3. Offline Reinforcement Learning + +Recall that we have defined several essential factors of the MDP. State $s$ includes the generated action tokens $\pmb{a}_{0:t-1}$ and the audio sequence $\pmb{u}_{0:t}$ up to the current time step $t$ . Each action $a$ corresponds to an index of the codebook learned by VQ-VAE. The reward $r$ is outputted by the reward model to evaluate how well the predicted gesture matches the input audio. Now we are able to augment the original speech-gesture dataset to a trajectory dataset $\mathcal{D} = \{s_i, a_i, r_i, s_i' | i = 1, \dots, N\}$ . Note that these trajectories can be collected before the start of the reinforcement learning phase. We denote by $\pi_{\beta}$ the behavior policy that generates these trajectories. + +Offline datasets usually do not provide complete state-action coverage. That is, the set $\{(s,a)|(s,a,r,s')\in \mathcal{D}\}$ is typically a small subset of the full space $S\times \mathcal{A}$ . Standard online reinforcement learning methods such as DDPG [24] and SAC [31] are not appropriate due to issues with bootstrapping from out-of-distribution actions. A "safe" strategy to mitigate the distributional shift problem is to be con + +servative: if we explicitly estimate the value of unseen outcomes conservatively (i.e. assign them a low value), then the estimated value or performance of the policy that executes unseen behaviors is guaranteed to be small. Using such conservative estimates for policy optimization will prevent the policy from executing unseen actions and it will perform reliably. To this end, we adopt a simple but effective offline reinforcement learning algorithm called conservative Q-learning (CQL) [20] to learn the optimal policy from the augmented offline dataset. + +Specifically, for the conservative off-policy evaluation, the Q-function, $\hat{Q}^{\pi} := \lim_{k \to \infty} \hat{Q}_k$ is trained via an iterative update: + +$$ +\begin{array}{l} \hat {Q} ^ {k + 1} \leftarrow \arg \min _ {Q} \frac {1}{2} \underset {s, a, s ^ {\prime} \sim \mathcal {D}} {\mathbb {E}} \left[ (Q (s, a) - \hat {\mathcal {B}} ^ {\pi} \hat {Q} ^ {k} (s, a)) ^ {2} \right] \\ + \alpha \left(\underset {a \sim \mu (a | s)} {\mathbb {E}} [ Q (s, a) ] - \underset {a \sim \hat {\pi} _ {\beta} (a | s)} {\mathbb {E}} [ Q (s, a) ]\right), \tag {9} \\ \end{array} +$$ + +where + +$$ +\hat {\mathcal {B}} ^ {\pi} \hat {Q} ^ {k} (s, a) = \mathbb {E} _ {s ^ {\prime} \sim P \left(s ^ {\prime} \mid s, a\right)} [ R (s, a) + \gamma \mathbb {E} _ {a ^ {\prime} \sim \hat {\pi} ^ {k} (s, a)} \hat {Q} ^ {k} \left(s ^ {\prime}, a ^ {\prime}\right) ] \tag {10} +$$ + +is the Bellman operator. This equation consists of the normal bootstrapping error and the regularization term with a trade-off factor $\alpha$ . Inside the regularization term, the first term always pushes the Q-value down on the $(s,a)$ pairs sampled from the learning policy $\mu$ whereas the second term pushes Q-value up on the $(s,a)$ pairs sampled from the offline data set. In practice, we selected the binding variant of CQL and DQN as our model, which can be optimized by minimizing the following loss function: + +$$ +\begin{array}{l} \underset {Q} {\operatorname {m i n}} \alpha_ {s \sim \mathcal {D}} \left[ \log \sum_ {a} \exp (Q (s, a)) - \underset {a \sim \hat {\pi} _ {\beta}} {\mathbb {E}} [ Q (s, a) ] \right] \\ + \frac {1}{2} _ {s, a, s ^ {\prime} \sim \mathcal {D}} \mathbb {E} \left[ (Q (s, a) - \hat {\mathcal {B}} ^ {\pi} \hat {Q} ^ {k} (s, a)) ^ {2} \right]. \tag {11} \\ \end{array} +$$ + +Since we want to use the Q-value function to autoregressively generate the action sequences, we construct a powerful GPT-based backbone model to estimate the Q-values, as is shown in Fig. 1. To construct the input, each motion code sequence is embedded to learnable features and concatenated with the audio features in the temporal dimension. We add a learned positional encoding to this concatenated $(2\times T^{\prime}\times C)$ -dimensional tensor and feed it to Transformer layers, whose structure is shown in the left of Fig. 1. Last, we employ a linear layer to map the output of Transformer layers to the Q-value $q\in \mathbb{R}^{T^{\prime}\times N}$ , where $N$ is the size of learned codebook. In the Transformer layer, the attention is the core component that determines the computational + +dependency among sequential elements of inputs, which is computed as: + +$$ +\operatorname {A t t e n t i o n} (\boldsymbol {Q}, \boldsymbol {K}, \boldsymbol {V}, \boldsymbol {M}) = \operatorname {s o f t m a x} \left(\frac {\boldsymbol {Q} \boldsymbol {K} ^ {T}}{\sqrt {C}} + \boldsymbol {M}\right) \boldsymbol {V}, \tag {12} +$$ + +where $Q, K, V$ denote the query, key and value from input, and $V$ is the mask. In the original GPT model, "causal attention" realizes the intercommunication of the current and previous data to compute the state for the time of interest. However, our model needs to infer actions from the speech at each time step, so it needs to process two input sequences simultaneously. To comply the causality conditioned among features of the audio and the action, $M$ is designed to be a $2 \times 2$ repeated block matrix with lower triangular matrix of size $T'$ . + +During inference, given the audio of arbitrary length and the initial motion code $a_0$ , the Q-network can output Q-values of all actions. We will select the action with the maximum Q-value and then add it to the input, so that we can autoregressively generate the entire action sequence. + +# 5. Experiments + +# 5.1. Dataset + +In this paper, we use two speech-gesture datasets: the Trinity dataset and a Chinese dataset collected by us. Trinity Gesture dataset is a large database of speech and gestures jointly collected by [8]. This dataset consists of 242 minutes of motion capture and audio of one male actor talking on different topics. The actor's motion was captured with a 20-camera Vicon system and solved onto a skeleton with 69 joints. In this paper, we use the official release version of The GENEA Challenge 2020 Dataset [19], where 221 minutes are used as training data, and the remaining 21 minutes are kept for testing. Additionally, we collected an 1-hour high-quality Chinese dataset using motion capture equipment. This dataset contains 3D full-body gestures and the aligned Chinese speech audio. And we retargeted the skeletons to be consistent with the Trinity dataset. In this dataset, the speaker described various speech scenes in Chinese and made high-standard common speech gestures, ensuring a high degree of matching between gestures and speech. + +# 5.2. Implementation Details + +All the motion data are downsampled to 20 frames per second on the Trinity and Chinese datasets. We focus on upper-body gestures in this work. For visualization, we employ a character model consist of 16 upper-body joints, including a rotational root. The audio features are extracted by the public audio processing toolbox librosa, including mel-scaled spectrogram and onset. In VQ-VAE, the codebook size $N$ and the channel dimension $C$ are both set to + +512, and the temporal downsampling rate of encoders is set to 8. During the training of VQ-VAE, motion data are cropped to uniform length of $T = 120$ (6 seconds) and sampled in batch size of 64. The trade-off $\beta$ in $\mathcal{L}_{VQ}$ is set to 0.1. $\alpha_{1}$ and $\alpha_{2}$ in $\mathcal{L}_m$ are both 1. We adopt Adam optimizer with $\beta_{1} = 0.9$ and $\beta_{2} = 0.999$ for 500 epochs on Trinity but 400 epochs on the Chinese dataset with learning rate 0.00003. During the training of reward model, both audio and motion data are clipped to 6s. The optimizer settings are consistent with those of the VQ-VAE. The structure of the Q network is basically the same as that of the GPT network, where the embedding dimension is 768 and the attention layer is implemented in 12 heads with dropout probability 0.1. During the off-line reinforcement learning, the motion sequences are encoded to motion codes and sampled to length of 15 and sampled in batch size of 256. The Q-network is optimized using Adam optimizer with $\beta_{1} = 0.5$ and $\beta_{2} = 0.999$ for 1500 epochs on Trinity but 1200 epochs on Chinese dataset, where the learning rate is initialized as 0.0001. The discount factor $\gamma$ is set to 0.99. At runtime, a Gaussian filter with a kernel size of $K = 5$ is used to smooth the denormalized gesture sequence along the time dimension. The entire framework takes 36 hours on Trinity and 30 hours on the Chinese dataset to complete training on one NVIDIA Quadro RTX 5000 GPU. + +# 5.3. Evaluation Metrics + +We adopt four commonly used evaluation metrics (MAJE, MAD, FGD and PMB) to compare various methods quantitatively. Mean absolute joint error (MAJE) measures the mean of the absolute errors between the generated joint positions and ground truth over all the time steps, which indicates how closely the generated joint positions follow the ground truth. Mean acceleration difference (MAD) measures the mean of the $l2$ norm differences between the generated joint accelerations and ground truth over all the time steps, which indicates how closely the ground truth and the generated joint movements match. Fréchet Gesture Distance (FGD) was proposed by which measures the difference between the distributions of the latent features of the generated gestures and ground truth [38]. The latent features are extracted from an auto-encoder trained on the Human 3.6M dataset*. Specifically, we use $\hat{H}$ and $H$ represent the latent features of generated gestures and ground truth. The Fréchet distance is formulated as + +$$ +\operatorname {F G D} (\hat {H}, H) = \left| \left| \hat {\mu} - \mu \right| \right| _ {2} ^ {2} + \operatorname {T r} \left(\hat {\Sigma} + \Sigma - 2 \sqrt {\hat {\Sigma} \Sigma}\right), \tag {13} +$$ + +where $\mu$ and $\Sigma$ are the first and second moments of the latent feature distribution of the ground truth $H$ , and $\hat{\mu}$ and $\hat{\Sigma}$ are + +Table 1. Comparison of RACER to Style Gesture, Gesticulator, S2AG and Bailando on Trinity datasets. $\downarrow$ means the lower is better and $\uparrow$ means the higher is better. + +
DatasetMethodMAJE(mm) ↓MAD(mm/s2) ↓FGD↓PMB(%) ↑
TrinityGround Truth00095.74
Style Gesture97.294.2636.9854.54
Gesticulator82.413.6231.0471.00
S2AG54.931.4920.3679.53
Bailando61.891.7417.2984.21
Ours50.331.2113.4489.58
Chinese DatasetStyle Gesture62.953.4630.1153.56
Bailando45.451.2925.0768.53
Ours43.251.199.2174.29
+ +the first and second moments of the latent feature distribution of the generated gestures $\hat{H}$ . FGD makes a reasonable assessment of the perceived plausibility of the synthesized gestures. The Percentage of Matched Beats PMB is a new metric to calculating the matching rate of audio and motion beats for evaluating rhythm performance [3]. In PMB, a motion beat is considered an matched if the distance to a nearby audio beat is within the range of $\delta$ : + +$$ +\operatorname {P M B} \left(B ^ {m}, B ^ {a}\right) = \frac {1}{N _ {m}} \sum_ {i = 1} ^ {N _ {m}} \sum_ {j = b _ {*} ^ {a} + 1} ^ {N _ {a}} \mathbb {1} \left[ \left| \left| b _ {j} ^ {m} - b _ {j} ^ {a} \right| \right| _ {1} < \delta \right], \tag {14} +$$ + +where $B^{m}$ and $B^{a}$ represent the motion beats and audio beats, and $N_{m}$ and $N_{a}$ represent the number of beats. Motion beat can be obtained by identifying the local minima of joints deceleration [14]. $b_{*}^{a}$ indicates the last matched audio beat. The $\delta$ is set to 0.2s, consistent with the original paper. + +# 5.4. Comparison to Existing Methods + +In this section, we compare RACER with several state-of-the-art methods to demonstrate the advances made by our method. On the Trinity dataset, we compare with the methods of Style Gesture [2], Gesticulator [18], Speed to Affective Gesture (S2AG) [4] and Bailando [29]. On the Chinese dataset, we compared with Style Gesture and Bailando. Style Gesture generates gestures based on only the speech audio features by a normalizing flow probabilistic model. Gesticulator leverages the audio and text of the speech to generate semantically consistent gestures. Compared with these methods, S2AG adds more modality, including speaker identity and seed gesture poses. Different from other methods, Bailando is a novel music-to-dance framework with a choreographic memory and actor-critic GPT model. Since Bailando's model architecture is similar to ours and both incorporate reinforcement learning into the training process, we apply it as a key contrast method to the gesture synthesis task. The hyperparameter setting is + +consistent with the original paper and the open source code, and the feature processing of the training data is consistent with ours. For a fair comparison, we keep the same skeleton and motion frame rate as the comparison systems. + +Quantitative experimental results are listed in Tab. 1. According to the comparison, our proposed model consistently performs favorably against all the other methods on all evaluations. Specially, on Trinity, our method improves $4.60(8.37\%)$ and $0.28(18.79\%)$ than the best compared baseline model S2AG on MAJE and MAD, respectively. In terms of FGD and PMB, Bailando, with its advanced network structure and reinforcement learning fine-tuning process, also obtained better evaluation results (17.29 and $84.21\%$ ) than S2AC (20.36 and $79.53\%$ ). Our approach further achieves significantly higher performance (13.44 and $89.58\%$ ) in both items. Similarly, our method has also achieved excellent performance in Chinese dataset compared with Style Gesture and Bailando, especially on FGD (9.21) and PMB $(74.29\%)$ . The superiority of our method on all evaluation metrics reveals that RACER can rely on offline RL not only to synthesize more real-like motion than compared baseline methods, but also but also to achieve outstanding performance in motion quality and rhythm. + +User Study. We conducted a user study to further assess the real visual performance of our method in a pairwise manner. We choose Style Gesture and Bailando to compare. We randomly sampled 10 30-second long slices generated from the test set of Trinity dataset. Each motion slice is bound to the Mixamo [25] model and rendered as a video by Blender. The experiment is conducted with 12 participants. The participants are asked to rate the slices from the following three aspects respectively: (1) realism, (2) speech-to-gesture matching, (3) speech-to-gesture rhythm matching. The scores assigned to each rating are in the range of 1-5, corresponding from worst to best. Rating statistics results are shown in Fig. 4. Notably, our method significantly surpasses the compared other methods. + +Table 2. Ablation study results on the Trinity dataset. + +
MethodMAJE(mm) ↓MAD(mm/s2) ↓FGD↓PMB(%) ↑Reward↑
RACER(Supervised Learning)56.221.8116.6978.54-
RACER(Distance Reward)58.962.9423.5572.09-
RACER(DQN)59.441.8115.4676.634.23
RACER(Actor-Critic)60.181.7617.6978.023.93
Ours50.331.2113.4489.589.34
+ +![](images/d9b4d68c071bd6920dd2606bb27e3381c26fdcbde6835d65ef8145604b6d3c84.jpg) +Figure 4. User study results comparing RACER against Bailando and Style Gesture. Bars with different colors represent rank counts. In total, 120 comparisons have been rated. The total scores are listed on the right. + +# 5.5. Ablation Studies + +To gain more insights into the proposed components of our method, we test some variants of RACER on Trinity dataset: (1) "Supervised Learning" indicates that we directly treat gesture code sequences as labels to train GPT-like models through supervised learning instead of reinforcement learning. (2) "Distance reward" indicates that our reward function is replaced by the Euclidean distance calculation between the predicted poses and ground truth, which is equivalent to taking the supervision signal in the traditional regression problem directly as the reward. (3) "DQN" means that we use the online RL algorithm DQN [34] instead of offline RL to learn the policy. The Q-value network of RACER(DQN) keeps consistent with ours. (4) "Actor-Critic" means that we use the online RL algorithm Actor-Critic instead of offline RL to learn the policy. For RACER(Actor-Critic), we regard the last 3 Transformer layers of GPT model with a additional linear-softmax layer as the branch of actor network. Besides, we add 3 Transformer layers as the branch of critic network to estimate the V-values. + +The quantitative scores are shown in Tab. 2. Firstly, we explore the effectiveness of offline reinforcement learning itself. Compared to RACER, the evaluation metrics of RACER(Supervised Learning) sharply drops $5.89(11\%)$ , $0.60(50\%)$ , $3.25(24\%)$ and $11.04\%$ , respectively. This + +shows that offline RL with proposed reward function can help the model learn to generate sequences of discrete tokens better than supervised learning. Secondly, to explore the advantages of the contrastive multi-modal reward, we evaluate the performance of RACER(Distance Reward). Our method has achieved better performance, especially on the MAD and PMB matrices. This confirms that the contrastive multi-modal reward function we propose can handle the feature relationship between multiple modes and improve the matching degree of motion and speech. Moreover, to verify the advantages of offline RL over online RL, we introduce two classic online RL algorithms, DQN and Actor-Critic, to determine the variations. Under the same number of training steps, DQN and AC algorithms can only get the highest rewards of 4.23 and 3.93 respectively, which is a significant gap compared to CQL. Naturally, other matrices are also significantly worse than CQL. + +# 6. Discussion and Conclusion + +In this paper, we propose a novel co-speech gesture synthesis framework called RACER. RACER regards the motion generated procedure as a sequential decision making process and employs reinforcement learning to learn the autoregressive gesture synthesis policy. There are mainly three sub-components of RACER. First, a vector quantized variational autoencoder (VQ-VAE) is employed to learn compact representations of gestures from massive real-world gesture data. Second, RACER uses a GPT-based architecture as policy network, which could autoregressively generate the next gestures given a sequence of speech audio. Thirdly, we propose a contrastive pre-training approach to calculate the rewards, which integrates contextual information into action evaluation and successfully captures the complex relationships between multi-modal speech-gesture data. We conduct extensive experiments to evaluate the performance of RACER. The experimental results show that RACER significantly outperforms the existing baselines in various objective and subjective metrics. Moreover, RACER shows a great generalization ability with respect to different contexts, which indicates the potentials of RACER in more complex motion synthesis tasks. + +# References + +[1] Chaitanya Ahuja, Dong Won Lee, Ryo Ishii, and Louis-Philippe Morency. No gestures left behind: Learning relationships between spoken language and freeform gestures. In Findings of the Association for Computational Linguistics: EMNLP 2020, pages 1884–1895, Nov. 2020. 2 +[2] Simon Alexanderson, Gustav Eje Henter, Taras Kucherenko, and Jonas Beskow. Style-controllable speech-driven gesture synthesis using normalising flows. Computer Graphics Forum, 39, 2020. 2, 7 +[3] Tenglong Ao, Qingzhe Gao, Yuke Lou, Baoquan Chen, and Libin Liu. Rhythmic gesticulator: Rhythm-aware cospeech gesture synthesis with hierarchical neural embeddings. ArXiv, abs/2210.01448, 2022. 7 +[4] Uttaran Bhattacharya, Elizabeth Childs, Nicholas Rewkowski, and Dinesh Manocha. Speech2affectivegestures: Synthesizing co-speech gestures with generative adversarial affective expression learning. In Proceedings of the 29th ACM International Conference on Multimedia, page 2027-2036, 2021. 2, 7 +[5] Uttaran Bhattacharya, Nicholas Rewkowski, Abhishek Banerjee, Pooja Guhan, Aniket Bera, and Dinesh Manocha. Text2gestures: A transformer-based network for generating emotive body gestures for virtual agents. In 2021 IEEE Virtual Reality and 3D User Interfaces (VR), pages 1-10, 2021. 2 +[6] Chung-Cheng Chiu, Louis-Philippe Morency, and Stacy Marsella. Predicting co-verbal gestures: A deep and temporal modeling approach. In Intelligent Virtual Agents, pages 152–166, 2015. 1 +[7] Hsu-Kuang Chiu, Ehsan Adeli, Borui Wang, De-An Huang, and Juan Carlos Niebles. Action-agnostic human pose forecasting. In 2019 IEEE Winter Conference on Applications of Computer Vision (WACV), pages 1423–1432, 2019. 2 +[8] Ylva Ferstl and Rachel McDonnell. Investigating the use of recurrent motion modelling for speech gesture generation. In Proceedings of the 18th International Conference on Intelligent Virtual Agents, page 93–98, 2018. 6 +[9] Ylva Ferstl, Michael Neff, and Rachel McDonnell. Multi-objective adversarial gesture generation. In Proceedings of the 12th ACM SIGGRAPH Conference on Motion, Interaction and Games, 2019. 2 +[10] Shiry Ginosar, Amir Bar, Gefen Kohavi, Caroline Chan, Andrew Owens, and Jitendra Malik. Learning individual styles of conversational gesture. In 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3492-3501, 2019. 1, 2 +[11] Ikhsanul Habibie, Weipeng Xu, Dushyant Mehta, Lingjie Liu, Hans-Peter Seidel, Gerard Pons-Moll, Mohamed Elgharib, and Christian Theobalt. Learning speech-driven 3d conversational gestures from video. In Proceedings of the 21st ACM International Conference on Intelligent Virtual Agents, page 101-108, 2021. 2 +[12] Gustav Eje Henter, Simon Alexanderson, and Jonas Beskow. Moglow: Probabilistic and controllable motion synthesis using normalising flows. ACM Trans. Graph., 2020. 2 + +[13] Alejandro Hernandez, Jurgen Gall, and Francesc Moreno-Noguer. Human motion prediction via spatio-temporal inpainting. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), October 2019. 2 +[14] Chieh Ho, Wei-Tze Tsai, Keng-Sheng Lin, and Homer H. Chen. Extraction and alignment evaluation of motion beats for street dance. In 2013 IEEE International Conference on Acoustics, Speech and Signal Processing, pages 2429–2433, 2013. 7 +[15] Yu Hua, Fan Xuanzhe, Hou Yaqing, Liu Yi, Kang Cai, Zhou Dongsheng, and Zhang Qiang. Towards efficient 3d human motion prediction using deformable transformer-based adversarial network. In 2022 International Conference on Robotics and Automation (ICRA), pages 861-867, 2022. 2 +[16] Lucas Kovar, Michael Gleicher, and Frédéric Pighin. Motion graphs. In ACM SIGGRAPH 2008 Classes, SIGGRAPH '08, 2008. 2 +[17] Taras Kucherenko, Dai Hasegawa, Gustav Eje Henter, Naoshi Kaneko, and Hedvig Kjellström. Analyzing input and output representations for speech-driven gesture generation. In Proceedings of the 19th ACM International Conference on Intelligent Virtual Agents, page 97-104, 2019. 1 +[18] Taras Kucherenko, Patrik Jonell, Sanne van Waveren, Gustav Eje Henter, Simon Alexandersson, Iolanda Leite, and Hedvig Kjellström. Gesticulator: A framework for semantically-aware speech-driven gesture generation. In Proceedings of the 2020 International Conference on Multimodal Interaction, page 242-250, 2020. 2, 7 +[19] Taras Kucherenko, Patrik Jonell, Youngwoo Yoon, Pieter Wolfert, and Gustav Eje Henter. A large, crowdsourced evaluation of gesture generation systems on common data: The genea challenge 2020. In 26th International Conference on Intelligent User Interfaces, page 11-21, 2021. 6 +[20] Aviral Kumar, Aurick Zhou, George Tucker, and Sergey Levine. Conservative q-learning for offline reinforcement learning. In H. Larochelle, M. Ranzato, R. Hadsell, M.F. Balcan, and H. Lin, editors, Advances in Neural Information Processing Systems, volume 33, pages 1179-1191. Curran Associates, Inc., 2020. 5 +[21] Jing Li, Di Kang, Wenjie Pei, Xuefei Zhe, Ying Zhang, Zhenyu He, and Linchao Bao. Audio2gestures: Generating diverse gestures from speech audio with conditional variational autoencoders. In 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 11273-11282, 2021. 1, 2 +[22] Ruilong Li, Shan Yang, David A. Ross, and Angjoo Kanazawa. Ai choreographer: Music conditioned 3d dance generation with aist++. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 13401-13412, October 2021. 2 +[23] Yuanzhi Liang, Qianyu Feng, Linchao Zhu, Li Hu, Pan Pan, and Yi Yang. Seeg: Semantic energized co-speech gesture generation. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10463-10472, 2022. 1 +[24] Timothy P. Lillicrap, Jonathan J. Hunt, Alexander Pritzel, Nicolas Heess, Tom Erez, Yuval Tassa, David Silver, and + +Daan Wierstra. Continuous control with deep reinforcement learning. In 4th International Conference on Learning Representations, ICLR 2016, San Juan, Puerto Rico, May 2-4, 2016, Conference Track Proceedings, 2016. 5 +[25] Mixamo. https://www MIXamo.com/. 7 +[26] Shenhan Qian, Zhi Tu, Yihao Zhi, Wen Liu, and Shenghua Gao. Speech drives templates: Co-speech gesture synthesis with learned templates. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 11077-11086, October 2021. 1 +[27] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision, 2021. 2, 5 +[28] Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, and Ilya Sutskever. Language Models are Unsupervised Multitask Learners. 2019. 2 +[29] Li Siyao, Weijiang Yu, Tianpei Gu, Chunze Lin, Quan Wang, Chen Qian, Chen Change Loy, and Ziwei Liu. Bailando: 3d dance generation by actor-critic gpt with choreographic memory. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11050-11059, June 2022. 2, 7 +[30] Yonglong Tian, Chen Sun, Ben Poole, Dilip Krishnan, Cordelia Schmid, and Phillip Isola. What makes for good views for contrastive learning? In H. Larochelle, M. Ranzato, R. Hadsell, M.F. Balcan, and H. Lin, editors, Advances in Neural Information Processing Systems, volume 33, pages 6827-6839. Curran Associates, Inc., 2020. 2 +[31] Pieter Abbeel Sergey Levine Tuomas Haarnoja, Aurick Zhou. Soft actor-critic: Off-policy maximum entropy deep reinforcement learning with a stochastic actor. ArXiv, abs/1801.01290, 2018. 5 +[32] Aaron van den Oord, Oriol Vinyals, and Koray Kavukcuoglu. Neural discrete representation learning. In Proceedings of the 31st International Conference on Neural Information Processing Systems, NIPS'17, page 6309-6318, 2017. 2, 4 +[33] Ruben Villegas, Jimei Yang, Duygu Ceylan, and Honglak Lee. Neural kinematic networks for unsupervised motion retargeting. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), June 2018. 2 +[34] David Silver Alex Graves Ioannis Antonoglou Daan Wierstra Martin Riedmiller Volodymyr Mnih, Koray Kavukcuoglu. Playing atari with deep reinforcement learning. ArXiv, abs/1312.5602, 2013. 8 +[35] Petra Wagner, Zofia Malisz, and Stefan Kopp. Gesture and speech in interaction: An overview. Speech Communication, 57:209-232, 2014. 1, 2 +[36] Jason R. Wilson, Nah Young Lee, Annie Saechao, Sharon Hershenson, Matthias Scheutz, and Linda Tickle-Degnen. Hand gestures and verbal acknowledgments improve humanrobot rapport. In Social Robotics, pages 334-344, 2017. 1 +[37] Xinchen Yan, Akash Rastogi, Ruben Villegas, Kalyan Sunkavalli, Eli Shechtman, Sunil Hadap, Ersin Yumer, and Honglak Lee. Mt-vae: Learning motion transformations to + +generate multimodal human dynamics. In Proceedings of the European Conference on Computer Vision (ECCV), September 2018. 2 +[38] Youngwoo Yoon, Bok Cha, Joo-Haeng Lee, Minsu Jang, Jaeyeon Lee, Jaehong Kim, and Geehyuk Lee. Speech gesture generation from the trimodal context of text, audio, and speaker identity. ACM Trans. Graph., 39(6), nov 2020. 1, 2, 6 +[39] Youngwoo Yoon, Woo-Ri Ko, Minsu Jang, Jaeyeon Lee, Jae-hong Kim, and Geehyuk Lee. Robots learn social skills: End-to-end learning of co-speech gesture generation for humanoid robots. In 2019 International Conference on Robotics and Automation (ICRA), pages 4303-4309, 2019. 1 \ No newline at end of file diff --git a/2023/Co-Speech Gesture Synthesis by Reinforcement Learning With Contrastive Pre-Trained Rewards/images.zip b/2023/Co-Speech Gesture Synthesis by Reinforcement Learning With Contrastive Pre-Trained Rewards/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..574c7aecb9a046a24b8a417cd0534f1864173847 --- /dev/null +++ b/2023/Co-Speech Gesture Synthesis by Reinforcement Learning With Contrastive Pre-Trained Rewards/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9f1849727d9c173c285c65fa4689d0dade818fc0767b95502b58f37ce5b12c92 +size 335063 diff --git a/2023/Co-Speech Gesture Synthesis by Reinforcement Learning With Contrastive Pre-Trained Rewards/layout.json b/2023/Co-Speech Gesture Synthesis by Reinforcement Learning With Contrastive Pre-Trained Rewards/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..11ec75839d543b24a8377085992a4c9d8ae1e06a --- /dev/null +++ b/2023/Co-Speech Gesture Synthesis by Reinforcement Learning With Contrastive Pre-Trained Rewards/layout.json @@ -0,0 +1,8528 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 68, + 102, + 526, + 138 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 102, + 526, + 138 + ], + "spans": [ + { + "bbox": [ + 68, + 102, + 526, + 138 + ], + "type": "text", + "content": "Co-speech Gesture Synthesis by Reinforcement Learning with Contrastive Pre-trained Rewards" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 159, + 550, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 159, + 550, + 204 + ], + "spans": [ + { + "bbox": [ + 53, + 159, + 550, + 204 + ], + "type": "text", + "content": "Mingyang Sun" + }, + { + "bbox": [ + 53, + 159, + 550, + 204 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 53, + 159, + 550, + 204 + ], + "type": "text", + "content": ", Mengchen Zhao" + }, + { + "bbox": [ + 53, + 159, + 550, + 204 + ], + "type": "inline_equation", + "content": "^{2*}" + }, + { + "bbox": [ + 53, + 159, + 550, + 204 + ], + "type": "text", + "content": ", Yaqing Hou" + }, + { + "bbox": [ + 53, + 159, + 550, + 204 + ], + "type": "inline_equation", + "content": "^{1\\dagger}" + }, + { + "bbox": [ + 53, + 159, + 550, + 204 + ], + "type": "text", + "content": ", Minglei Li" + }, + { + "bbox": [ + 53, + 159, + 550, + 204 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 53, + 159, + 550, + 204 + ], + "type": "text", + "content": ", Huang Xu" + }, + { + "bbox": [ + 53, + 159, + 550, + 204 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 53, + 159, + 550, + 204 + ], + "type": "text", + "content": ", Songcen Xu" + }, + { + "bbox": [ + 53, + 159, + 550, + 204 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 53, + 159, + 550, + 204 + ], + "type": "text", + "content": ", Jianye Hao" + }, + { + "bbox": [ + 53, + 159, + 550, + 204 + ], + "type": "inline_equation", + "content": "^{2,4}" + }, + { + "bbox": [ + 53, + 159, + 550, + 204 + ], + "type": "text", + "content": ", Dalian University of Technology, China " + }, + { + "bbox": [ + 53, + 159, + 550, + 204 + ], + "type": "inline_equation", + "content": "{ }^{2}" + }, + { + "bbox": [ + 53, + 159, + 550, + 204 + ], + "type": "text", + "content": "Noah's Ark Lab, Huawei, China " + }, + { + "bbox": [ + 53, + 159, + 550, + 204 + ], + "type": "inline_equation", + "content": "{ }^{3}" + }, + { + "bbox": [ + 53, + 159, + 550, + 204 + ], + "type": "text", + "content": "Huawei Cloud Computing Technologies Co., Ltd, China " + }, + { + "bbox": [ + 53, + 159, + 550, + 204 + ], + "type": "inline_equation", + "content": "{ }^{4}" + }, + { + "bbox": [ + 53, + 159, + 550, + 204 + ], + "type": "text", + "content": "Tianjin University, China" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 189, + 205, + 414, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 189, + 205, + 414, + 217 + ], + "spans": [ + { + "bbox": [ + 189, + 205, + 414, + 217 + ], + "type": "text", + "content": "mysun@mail.dlut.edu.cn, houyq@dlut.edu.cn" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 123, + 219, + 480, + 232 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 123, + 219, + 480, + 232 + ], + "spans": [ + { + "bbox": [ + 123, + 219, + 480, + 232 + ], + "type": "text", + "content": "{zhaomengchen,liminglei29,xuhuang2,xusongcen,haojianye}@huawei.com" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 143, + 258, + 192, + 272 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 258, + 192, + 272 + ], + "spans": [ + { + "bbox": [ + 143, + 258, + 192, + 272 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 284, + 289, + 536 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 284, + 289, + 536 + ], + "spans": [ + { + "bbox": [ + 46, + 284, + 289, + 536 + ], + "type": "text", + "content": "There is a growing demand of automatically synthesizing co-speech gestures for virtual characters. However, it remains a challenge due to the complex relationship between input speeches and target gestures. Most existing works focus on predicting the next gesture that fits the data best, however, such methods are myopic and lack the ability to plan for future gestures. In this paper, we propose a novel reinforcement learning (RL) framework called RACER to generate sequences of gestures that maximize the overall satisfactory. RACER employs a vector quantized variational autoencoder to learn compact representations of gestures and a GPT-based policy architecture to generate coherent sequence of gestures autoregressively. In particular, we propose a contrastive pre-training approach to calculate the rewards, which integrates contextual information into action evaluation and successfully captures the complex relationships between multi-modal speech-gesture data. Experimental results show that our method significantly outperforms existing baselines in terms of both objective metrics and subjective human judgements. Demos can be found at https://github.com/RLracer/RACER.git." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 558, + 128, + 571 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 558, + 128, + 571 + ], + "spans": [ + { + "bbox": [ + 47, + 558, + 128, + 571 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 578, + 287, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 578, + 287, + 687 + ], + "spans": [ + { + "bbox": [ + 46, + 578, + 287, + 687 + ], + "type": "text", + "content": "Gesturing is important for human speakers to improve their expressiveness. It conveys the necessary non-verbal information to the audience, giving the speech a more emotional touch so that to enhance persuasiveness and credibility. Similarly, in virtual world, high-quality 3D gesture animations can make a talking character more vividly. For example, in human-computer interaction scenarios, vivid gestures performed by virtual characters can help listeners to concentrate and improve the intimacy between humans and" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 260, + 545, + 296 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 260, + 545, + 296 + ], + "spans": [ + { + "bbox": [ + 306, + 260, + 545, + 296 + ], + "type": "text", + "content": "characters [36]. Attracted by these merits, there has been a growing demand of automatically synthesizing high-quality co-speech gestures in computer animation." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 300, + 546, + 467 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 300, + 546, + 467 + ], + "spans": [ + { + "bbox": [ + 304, + 300, + 546, + 467 + ], + "type": "text", + "content": "However, automatically synthesizing co-speech gestures remains a challenge due to the complicated relationship between speech audios and gestures. On one hand, a speaker may play different gestures when speaking the same words due to different mental and physical states. On the other hand, a speaker may also play similar gestures when speaking different words. Therefore, co-speech gesture synthesizing is inherently a \"many-to-many\" problem [21]. Moreover, in order to ensure the overall fluency and consistency, we must take into consideration both the contextual information and the subsequent effect upon playing a gesture [23, 38]. Therefore, gesture synthesizing is rather a sequential decision making problem than a simple matching between speeches and gestures." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 471, + 547, + 697 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 471, + 547, + 697 + ], + "spans": [ + { + "bbox": [ + 304, + 471, + 547, + 697 + ], + "type": "text", + "content": "Compared with traditional rule-based approaches [35], data-driven gesture synthesis approaches [10,39] has shown many advantages, including the low development cost and the ability to generalize. Most existing data-driven approaches consider gesture synthesis as a classification task [6, 26] or a regression task [17] in a deterministic way i.e. the same speech or text input always maps to the same gesture output. However, these models rely on the assumption that there exists a unique ground-truth label for each input sequence of speech, which contradicts to the \"many-to-many\" nature of the problem. As a consequence, they sacrifice diversity and semantics of gestures and tend to learn some averaged gestures for any input speeches. Some other works adopt adversarial learning framework, where a discriminator is trained to distinguish between generated gestures and the recorded gestures in dataset [10,38]. Although they improve the generalizability of the gesture generator to some extent, they still fail to explore the essential relationship between speeches and gestures." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 317, + 701, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 701, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 317, + 701, + 545, + 714 + ], + "type": "text", + "content": "In order to address the above challenges, we propose" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 693, + 127, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 693, + 127, + 703 + ], + "spans": [ + { + "bbox": [ + 58, + 693, + 127, + 703 + ], + "type": "text", + "content": "*Equal contribution." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 59, + 703, + 135, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 703, + 135, + 712 + ], + "spans": [ + { + "bbox": [ + 59, + 703, + 135, + 712 + ], + "type": "text", + "content": "† Corresponding author." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "2331" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 289, + 348 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 289, + 348 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 289, + 348 + ], + "type": "text", + "content": "a novel Reinforcement leArning framework with Contrasitive prE-trained Rewards (RACER) for generating high-quality co-speech gestures. RACER is trained in an offline manner but can be used to generate the next gesture for a speaking character in real time. RACER consists of the following three components. Firstly, in order to extract meaningful gestures from the infinite action space, RACER adopts a vector quantized variational autoencoder (VQ-VAE) [32] to learn compact gesture representations, which significantly reduces the action space. Secondly, we construct the Q-value network using a GPT-based [28] model, which has natural advantages of generating coherence sequence of gestures. Thirdly, inspired by the contrastive language-image pre-training (CLIP) [27] and recent advances of contrastive learning [30], we propose a contrastive speech-gesture pre-training method to compute the rewards, which guide the RL agent to explore sophisticated relations between speeches and gestures. Note that these rewards evaluate the quality of gestures as a sequence. By contrast, in conventional supervised learning frameworks, the focus is on predicting only the next gesture, ignoring the quality of generated sequence as a whole. To sum up, the core contributions of this paper are as follows:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 356, + 287, + 594 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 57, + 356, + 287, + 429 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 356, + 287, + 429 + ], + "spans": [ + { + "bbox": [ + 57, + 356, + 287, + 429 + ], + "type": "text", + "content": "(1) We formally model the co-speech gesture synthesis problem as a Markov decision process and propose a novel RL based approach called RACER to learn the optimal gesture synthesis policy. RACER can be trained in an offline manner and used to synthesis co-speech gestures in real time." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 57, + 435, + 287, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 435, + 287, + 472 + ], + "spans": [ + { + "bbox": [ + 57, + 435, + 287, + 472 + ], + "type": "text", + "content": "(2) We introduce VQ-VAE to encode and quantize the motion segments to a codebook, which significantly reduces the action space and facilitates the RL phase." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 479, + 287, + 528 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 479, + 287, + 528 + ], + "spans": [ + { + "bbox": [ + 57, + 479, + 287, + 528 + ], + "type": "text", + "content": "(3) We propose a contrastive speech-gesture pre-training method to compute the rewards, which guide the RL agent to discover deeper relations from multi-modal speech-gesture data." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 57, + 534, + 287, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 534, + 287, + 594 + ], + "spans": [ + { + "bbox": [ + 57, + 534, + 287, + 594 + ], + "type": "text", + "content": "(4) Extensive experimental results show that RACER outperforms the existing baselines in terms of both objective metrics and subjective human judgements. This demonstrates the superiority and the potential of RL in co-speech gesture synthesis tasks." + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 47, + 604, + 134, + 616 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 604, + 134, + 616 + ], + "spans": [ + { + "bbox": [ + 47, + 604, + 134, + 616 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 623, + 209, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 623, + 209, + 635 + ], + "spans": [ + { + "bbox": [ + 47, + 623, + 209, + 635 + ], + "type": "text", + "content": "2.1. Data-driven Motion Synthesis" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 642, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 642, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 642, + 288, + 714 + ], + "type": "text", + "content": "Human motion synthesis has been studied for a long time [21, 37]. Traditional approaches such as motion graph [16] cut the motion capture data and synthesize new motions according to crafted transition rules. With the development of deep learning, using neural networks to synthesize motions in an end-to-end manner becomes popular. Some of" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 72, + 547, + 252 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 252 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 252 + ], + "type": "text", + "content": "works focus on motion prediction [13, 15], which could predict current motion with a high accuracy, but they often fail to predict longer motion sequences. Recent works focus more on the motion synthesis, which aims at generating long-term natural motions such as walking, jumping, waving and dancing. Many advanced techniques have been applied to long-term motion synthesis, including conditional GANs [33], VAEs [21], auto-regressive models [7] and Transformers [5, 22]. A most recent work called Bailando explores the potential of RL in synthesizing dances for given music [29]. However, in Bailando, RL is used only for fine tuning so that the exact contributions of RL remain unclear. By contrast, our proposed method RACER is trained solely by RL and its effectiveness is demonstrated by extensive experiments." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 258, + 463, + 271 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 258, + 463, + 271 + ], + "spans": [ + { + "bbox": [ + 306, + 258, + 463, + 271 + ], + "type": "text", + "content": "2.2. Co-speech Gesture Synthesis" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 277, + 547, + 577 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 277, + 547, + 577 + ], + "spans": [ + { + "bbox": [ + 304, + 277, + 547, + 577 + ], + "type": "text", + "content": "Co-speech gesture synthesis is challenging due to the complicated relations between speeches and gestures that are hidden in the multi-modal dataset. One line of conventional methods is rule-based generation [35], where the core idea is to pre-define a set of gesture units and design rules to connect speech words and the gesture units. However, it requires too much human effort in designing the rules, which is costly and inefficient in practice. Benefit from the recent advances in deep learning, many data-driven approaches have been proposed to learn the matching rules from speech-gesture data. Existing works focus on exploring the effectiveness of different network architectures, including the multi-layer perceptron (MLP) [18], convolutional neural networks (CNNs) [11], recurrent neural networks (RNNs) [4] and Transformers [18, 29]. Some recent works incorporate adversarial loss to enhance gesture fidelity [1,9,10,38]. However, they simply construct a target gesture as the unique label of the current segment of speech, ignoring the \"many-to-many\" nature of the problem. To model the complicated relation between speeches and gestures, Alexanderson et al. [2] propose a probabilistic model MoGlow based on normalizing flows [12] which maps input speeches to Gaussian distributions of gestures. However, the randomly sampled gestures contain non-negligible noise and therefore suffer from low interpretability." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 586, + 391, + 597 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 586, + 391, + 597 + ], + "spans": [ + { + "bbox": [ + 306, + 586, + 391, + 597 + ], + "type": "text", + "content": "3. Preliminaries" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "type": "text", + "content": "The underlying environment of reinforcement learning is typically modeled by a Markov decision process (MDP), which is defined by a tuple " + }, + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "type": "inline_equation", + "content": "(S, A, P, R, \\gamma)" + }, + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "type": "text", + "content": " represents the state space and " + }, + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "type": "text", + "content": " represents the action space. The transition function " + }, + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "type": "inline_equation", + "content": "P(\\cdot | s, a)" + }, + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "type": "text", + "content": " gives the probabilities of transitioning to the next state after taking action " + }, + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "type": "inline_equation", + "content": "a" + }, + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "type": "text", + "content": " at state " + }, + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "type": "inline_equation", + "content": "R(\\cdot | s, a)" + }, + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "type": "text", + "content": " is the associated reward function and " + }, + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\gamma \\in [0, 1)" + }, + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "type": "text", + "content": " represents the discount factor. A policy " + }, + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\pi(\\cdot | s)" + }, + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "type": "text", + "content": " maps each state " + }, + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "type": "inline_equation", + "content": "s \\in S" + }, + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "type": "text", + "content": " to a distribution over actions. For an agent following the policy" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2332" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 109, + 81, + 487, + 300 + ], + "blocks": [ + { + "bbox": [ + 109, + 81, + 487, + 300 + ], + "lines": [ + { + "bbox": [ + 109, + 81, + 487, + 300 + ], + "spans": [ + { + "bbox": [ + 109, + 81, + 487, + 300 + ], + "type": "image", + "image_path": "024374574dbb552d472bc6a0aa14c844fa433a13bf08e57e07ca6aa9589c87b1.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 310, + 547, + 345 + ], + "lines": [ + { + "bbox": [ + 46, + 310, + 547, + 345 + ], + "spans": [ + { + "bbox": [ + 46, + 310, + 547, + 345 + ], + "type": "text", + "content": "Figure 1. The overall architecture of RACER. Given a piece of speech audio and the initial gesture code " + }, + { + "bbox": [ + 46, + 310, + 547, + 345 + ], + "type": "inline_equation", + "content": "a_0" + }, + { + "bbox": [ + 46, + 310, + 547, + 345 + ], + "type": "text", + "content": ", the Q-network represented by transformer layers autoregressively calculates the Q-values and selects a sequence of actions " + }, + { + "bbox": [ + 46, + 310, + 547, + 345 + ], + "type": "inline_equation", + "content": "(a_1, \\dots, a_T)" + }, + { + "bbox": [ + 46, + 310, + 547, + 345 + ], + "type": "text", + "content": ". The action sequence will then be transformed to quantitative features by querying the codebook and finally be decoded to motion sequences by the decoder of VQ-VAE." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 365, + 288, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 365, + 288, + 389 + ], + "spans": [ + { + "bbox": [ + 47, + 365, + 288, + 389 + ], + "type": "inline_equation", + "content": "\\pi" + }, + { + "bbox": [ + 47, + 365, + 288, + 389 + ], + "type": "text", + "content": ", the action value function, denoted by " + }, + { + "bbox": [ + 47, + 365, + 288, + 389 + ], + "type": "inline_equation", + "content": "Q^{\\pi}(s,a)" + }, + { + "bbox": [ + 47, + 365, + 288, + 389 + ], + "type": "text", + "content": ", is defined as the expectation of cumulative discounted future rewards:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 405, + 287, + 438 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 405, + 287, + 438 + ], + "spans": [ + { + "bbox": [ + 56, + 405, + 287, + 438 + ], + "type": "interline_equation", + "content": "Q ^ {\\pi} (s, a) := \\mathbb {E} \\left[ \\sum_ {t = 0} \\gamma^ {t} R \\left(s _ {t}, a _ {t}\\right) \\right], \\tag {1}", + "image_path": "bfe817845951af3a4d4731cce2c70e1136018396dc0cb37ef6ae3618bd6622c4.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 57, + 440, + 266, + 453 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 440, + 266, + 453 + ], + "spans": [ + { + "bbox": [ + 57, + 440, + 266, + 453 + ], + "type": "interline_equation", + "content": "s _ {0} = s, a _ {0} = a, s _ {t} \\sim P (\\cdot | s _ {t - 1}, a _ {t - 1}), a _ {t} \\sim \\pi (\\cdot | s _ {t}).", + "image_path": "0e7bc483cbca68428c804feb355d052e8472ff1bbc74b44d2bd8eae1f15981c9.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 460, + 287, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 460, + 287, + 521 + ], + "spans": [ + { + "bbox": [ + 47, + 460, + 287, + 521 + ], + "type": "text", + "content": "The goal in RL is to learn an optimal policy " + }, + { + "bbox": [ + 47, + 460, + 287, + 521 + ], + "type": "inline_equation", + "content": "\\pi^{*}" + }, + { + "bbox": [ + 47, + 460, + 287, + 521 + ], + "type": "text", + "content": " that maximizes the expected cumulative reward. We denote by " + }, + { + "bbox": [ + 47, + 460, + 287, + 521 + ], + "type": "inline_equation", + "content": "Q^{*}(s,a)" + }, + { + "bbox": [ + 47, + 460, + 287, + 521 + ], + "type": "text", + "content": " the optimal action value function under the optimal policy " + }, + { + "bbox": [ + 47, + 460, + 287, + 521 + ], + "type": "inline_equation", + "content": "\\pi^{*}" + }, + { + "bbox": [ + 47, + 460, + 287, + 521 + ], + "type": "text", + "content": ". The Bellman equations for the optimal Q-values can be represented by:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 59, + 538, + 287, + 556 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 538, + 287, + 556 + ], + "spans": [ + { + "bbox": [ + 59, + 538, + 287, + 556 + ], + "type": "interline_equation", + "content": "Q ^ {*} (s, a) := R (s, a) + \\gamma \\mathbb {E} _ {s ^ {\\prime} \\sim P} \\max _ {a ^ {\\prime} \\in \\mathcal {A}} Q ^ {*} \\left(s ^ {\\prime}, a ^ {\\prime}\\right). \\tag {2}", + "image_path": "41bfc2d967b95b8e13bbbae68941dfc6e0431d2358572f66116416a5049d8a59.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 563, + 287, + 646 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 563, + 287, + 646 + ], + "spans": [ + { + "bbox": [ + 46, + 563, + 287, + 646 + ], + "type": "text", + "content": "A standard approach for computing the optimal policy is Q-learning, which iteratively improves the estimation of " + }, + { + "bbox": [ + 46, + 563, + 287, + 646 + ], + "type": "inline_equation", + "content": "Q^{*}" + }, + { + "bbox": [ + 46, + 563, + 287, + 646 + ], + "type": "text", + "content": " by applying the Bellman operator derived from Eq. (2). While traditional Q-learning requires interactions with the environment, recent advances in offline RL allow us to learn the optimal policy directly from a fixed dataset, which we will elaborate in Sec. 4.3." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 658, + 138, + 671 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 658, + 138, + 671 + ], + "spans": [ + { + "bbox": [ + 47, + 658, + 138, + 671 + ], + "type": "text", + "content": "4. Our Approach" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 677, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 677, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 677, + 288, + 714 + ], + "type": "text", + "content": "The overview of our proposed framework RACER is shown in Fig. 1. At each time step " + }, + { + "bbox": [ + 47, + 677, + 288, + 714 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 47, + 677, + 288, + 714 + ], + "type": "text", + "content": ", the state " + }, + { + "bbox": [ + 47, + 677, + 288, + 714 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 47, + 677, + 288, + 714 + ], + "type": "text", + "content": " consists of the generated action tokens " + }, + { + "bbox": [ + 47, + 677, + 288, + 714 + ], + "type": "inline_equation", + "content": "(a_{1},\\ldots ,a_{t - 1})" + }, + { + "bbox": [ + 47, + 677, + 288, + 714 + ], + "type": "text", + "content": " and input audio." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 365, + 545, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 365, + 545, + 544 + ], + "spans": [ + { + "bbox": [ + 304, + 365, + 545, + 544 + ], + "type": "text", + "content": "Unlike existing methods which directly learn a mapping from audio features to the continuous high-dimensional motion space, we encode and quantize the motion into a finite codebook " + }, + { + "bbox": [ + 304, + 365, + 545, + 544 + ], + "type": "inline_equation", + "content": "\\mathcal{Z} = \\{z_i\\}_{i=1}^N" + }, + { + "bbox": [ + 304, + 365, + 545, + 544 + ], + "type": "text", + "content": " by VQ-VAE, where " + }, + { + "bbox": [ + 304, + 365, + 545, + 544 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 365, + 545, + 544 + ], + "type": "text", + "content": " is the size of codebook and each code " + }, + { + "bbox": [ + 304, + 365, + 545, + 544 + ], + "type": "inline_equation", + "content": "z_i" + }, + { + "bbox": [ + 304, + 365, + 545, + 544 + ], + "type": "text", + "content": " represents a gesture lexeme feature. The details of action design are introduced in Sec. 4.1. We use a GPT-like unidirectional model as the Q-network that autoregressively outputs action tokens following a greedy strategy. An action token " + }, + { + "bbox": [ + 304, + 365, + 545, + 544 + ], + "type": "inline_equation", + "content": "a" + }, + { + "bbox": [ + 304, + 365, + 545, + 544 + ], + "type": "text", + "content": " will be mapped to a gesture lexeme feature " + }, + { + "bbox": [ + 304, + 365, + 545, + 544 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 304, + 365, + 545, + 544 + ], + "type": "text", + "content": " and then be decoded to a specific gesture motion. Moreover, we propose a contrastive speech-gesture pre-training model to compute the immediate rewards for the actions, which will be elaborated in Sec. 4.2. In addition, we will introduce how to train the Q-network in a fully offline manner in Sec. 4.3." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 552, + 394, + 564 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 552, + 394, + 564 + ], + "spans": [ + { + "bbox": [ + 306, + 552, + 394, + 564 + ], + "type": "text", + "content": "4.1. Action Design" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "text", + "content": "Instead of designing handcrafted gestures with expensive manual efforts, our goal is to summarize gesture lexemes into a representative codebook from a large gesture dataset in an unsupervised manner. To collect distinctive gesture representations and reconstruct them back to real gestures, we employ a VQ-VAE model which has successfully demonstrated its ability to learn effective latent representations from temporal data. The structure of VQ-VAE is shown in Fig. 2. VQ-VAE extends the regular autoencoder by adding a discrete codebook " + }, + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{Z}" + }, + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "text", + "content": " to the network, which consists of a list of vectors with their corresponding indexes. A piece of motion in raw dataset can be represented by a ma" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2333" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 51, + 82, + 282, + 201 + ], + "blocks": [ + { + "bbox": [ + 51, + 82, + 282, + 201 + ], + "lines": [ + { + "bbox": [ + 51, + 82, + 282, + 201 + ], + "spans": [ + { + "bbox": [ + 51, + 82, + 282, + 201 + ], + "type": "image", + "image_path": "62534f56cfd55598bcfb00c688e358e6566c967fad156f743763c34360e4349d.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 94, + 220, + 240, + 232 + ], + "lines": [ + { + "bbox": [ + 94, + 220, + 240, + 232 + ], + "spans": [ + { + "bbox": [ + 94, + 220, + 240, + 232 + ], + "type": "text", + "content": "Figure 2. Structure of motion VQ-VAE." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 252, + 287, + 337 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 252, + 287, + 337 + ], + "spans": [ + { + "bbox": [ + 46, + 252, + 287, + 337 + ], + "type": "text", + "content": "trix " + }, + { + "bbox": [ + 46, + 252, + 287, + 337 + ], + "type": "inline_equation", + "content": "\\mathcal{M} \\in \\mathbb{R}^{T \\times J}" + }, + { + "bbox": [ + 46, + 252, + 287, + 337 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 252, + 287, + 337 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 46, + 252, + 287, + 337 + ], + "type": "text", + "content": " is the number of frames and " + }, + { + "bbox": [ + 46, + 252, + 287, + 337 + ], + "type": "inline_equation", + "content": "J" + }, + { + "bbox": [ + 46, + 252, + 287, + 337 + ], + "type": "text", + "content": " is the dimension of joint features. " + }, + { + "bbox": [ + 46, + 252, + 287, + 337 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 46, + 252, + 287, + 337 + ], + "type": "text", + "content": " can then be transformed to latent features " + }, + { + "bbox": [ + 46, + 252, + 287, + 337 + ], + "type": "inline_equation", + "content": "e \\in \\mathbb{R}^{T' \\times C}" + }, + { + "bbox": [ + 46, + 252, + 287, + 337 + ], + "type": "text", + "content": " by a 1-D convolutional network, where " + }, + { + "bbox": [ + 46, + 252, + 287, + 337 + ], + "type": "inline_equation", + "content": "T'" + }, + { + "bbox": [ + 46, + 252, + 287, + 337 + ], + "type": "text", + "content": " is the length of feature sequence after down-sampling, and " + }, + { + "bbox": [ + 46, + 252, + 287, + 337 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 46, + 252, + 287, + 337 + ], + "type": "text", + "content": " is the channel dimension. Each row of " + }, + { + "bbox": [ + 46, + 252, + 287, + 337 + ], + "type": "inline_equation", + "content": "e" + }, + { + "bbox": [ + 46, + 252, + 287, + 337 + ], + "type": "text", + "content": ", represented by " + }, + { + "bbox": [ + 46, + 252, + 287, + 337 + ], + "type": "inline_equation", + "content": "e_i" + }, + { + "bbox": [ + 46, + 252, + 287, + 337 + ], + "type": "text", + "content": ", will be substituted by its closest vector " + }, + { + "bbox": [ + 46, + 252, + 287, + 337 + ], + "type": "inline_equation", + "content": "z_j" + }, + { + "bbox": [ + 46, + 252, + 287, + 337 + ], + "type": "text", + "content": " in the codebook:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 113, + 350, + 287, + 369 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 113, + 350, + 287, + 369 + ], + "spans": [ + { + "bbox": [ + 113, + 350, + 287, + 369 + ], + "type": "interline_equation", + "content": "\\hat {e} _ {i} = \\arg \\min _ {\\boldsymbol {z} _ {j} \\in \\mathcal {Z}} | | \\boldsymbol {e} _ {i} - \\boldsymbol {z} _ {j} | |. \\tag {3}", + "image_path": "ceee3d988762b13ea5325a9faf2fda5ab68b4a19b0f1bf9a408a9298211bb5f2.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 376, + 287, + 425 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 376, + 287, + 425 + ], + "spans": [ + { + "bbox": [ + 47, + 376, + 287, + 425 + ], + "type": "text", + "content": "Then, the quantized features " + }, + { + "bbox": [ + 47, + 376, + 287, + 425 + ], + "type": "inline_equation", + "content": "\\hat{e}_i" + }, + { + "bbox": [ + 47, + 376, + 287, + 425 + ], + "type": "text", + "content": " will be decoded to " + }, + { + "bbox": [ + 47, + 376, + 287, + 425 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{M}}" + }, + { + "bbox": [ + 47, + 376, + 287, + 425 + ], + "type": "text", + "content": " via a decoder so that to reconstruct " + }, + { + "bbox": [ + 47, + 376, + 287, + 425 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 47, + 376, + 287, + 425 + ], + "type": "text", + "content": ". Following [32], the encoder and decoder are simultaneously learned by minimizing the following loss function:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 78, + 444, + 287, + 475 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 78, + 444, + 287, + 475 + ], + "spans": [ + { + "bbox": [ + 78, + 444, + 287, + 475 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} _ {V Q} = \\mathcal {L} _ {m} (\\mathcal {M}, \\hat {\\mathcal {M}}) \\tag {4} \\\\ + \\left\\| \\hat {\\boldsymbol {e}} - \\operatorname {s g} (\\boldsymbol {e}) \\right\\| _ {2} ^ {2} + \\beta \\left\\| \\operatorname {s g} (\\hat {\\boldsymbol {e}}) - \\boldsymbol {e} \\right\\| _ {2} ^ {2}, \\\\ \\end{array}", + "image_path": "532f9299ce401be5cfd099341c4269f51c19c85b8b48fa13fdd3a7162995db0e.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 484, + 287, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 484, + 287, + 533 + ], + "spans": [ + { + "bbox": [ + 47, + 484, + 287, + 533 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 484, + 287, + 533 + ], + "type": "inline_equation", + "content": "\\mathfrak{sg}[\\cdot ]" + }, + { + "bbox": [ + 47, + 484, + 287, + 533 + ], + "type": "text", + "content": " stands for the stop gradient operator that prevents the gradient from backpropagating through its operand. The first term " + }, + { + "bbox": [ + 47, + 484, + 287, + 533 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_m(\\mathcal{M},\\hat{\\mathcal{M}})" + }, + { + "bbox": [ + 47, + 484, + 287, + 533 + ], + "type": "text", + "content": " represents the reconstruction error, which is expanded as:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 52, + 552, + 287, + 583 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 52, + 552, + 287, + 583 + ], + "spans": [ + { + "bbox": [ + 52, + 552, + 287, + 583 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} _ {m} (\\mathcal {M}, \\hat {\\mathcal {M}}) = | | \\mathcal {M}, \\hat {\\mathcal {M}} | | _ {1} \\tag {5} \\\\ + \\alpha_ {1} | | \\mathcal {M} ^ {\\prime}, \\hat {\\mathcal {M}} ^ {\\prime} | | _ {1} + \\alpha_ {2} | | \\mathcal {M} ^ {\\prime \\prime}, \\hat {\\mathcal {M}} ^ {\\prime \\prime} | | _ {1}, \\\\ \\end{array}", + "image_path": "427ccdcd5e66ef5d92c594ff8369cb702bd61b46cb8fadc4b16ccc2d75751251.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 594, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 287, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 594, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{M}'" + }, + { + "bbox": [ + 46, + 594, + 287, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 594, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{M}''" + }, + { + "bbox": [ + 46, + 594, + 287, + 713 + ], + "type": "text", + "content": " are the " + }, + { + "bbox": [ + 46, + 594, + 287, + 713 + ], + "type": "inline_equation", + "content": "1^{st}" + }, + { + "bbox": [ + 46, + 594, + 287, + 713 + ], + "type": "text", + "content": " order and " + }, + { + "bbox": [ + 46, + 594, + 287, + 713 + ], + "type": "inline_equation", + "content": "2^{nd}" + }, + { + "bbox": [ + 46, + 594, + 287, + 713 + ], + "type": "text", + "content": " order of partial derivatives of motion sequence " + }, + { + "bbox": [ + 46, + 594, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 46, + 594, + 287, + 713 + ], + "type": "text", + "content": ", representing the velocity and the acceleration of the motion, respectively. " + }, + { + "bbox": [ + 46, + 594, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\alpha_{1}" + }, + { + "bbox": [ + 46, + 594, + 287, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 594, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\alpha_{2}" + }, + { + "bbox": [ + 46, + 594, + 287, + 713 + ], + "type": "text", + "content": " are hyper-parameters used to trade-off between the two losses. We can see from Eq. (5) that VQ-VAE reconstructs not only the original position of each joints of the character, but also the velocities and the accelerations of its motions. Notably, since " + }, + { + "bbox": [ + 46, + 594, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{Z}" + }, + { + "bbox": [ + 46, + 594, + 287, + 713 + ], + "type": "text", + "content": " is discrete, we will simply copy the gradients from the reconstruction error and pass them to the encoder during training. The second term in Equation (4) aims" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 312, + 92, + 518, + 225 + ], + "blocks": [ + { + "bbox": [ + 312, + 92, + 518, + 225 + ], + "lines": [ + { + "bbox": [ + 312, + 92, + 518, + 225 + ], + "spans": [ + { + "bbox": [ + 312, + 92, + 518, + 225 + ], + "type": "image", + "image_path": "cbc062a5b975db7d728e2d923065c654787eb6db9c9f177cb17cb0c69ce0daa1.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 236, + 545, + 270 + ], + "lines": [ + { + "bbox": [ + 305, + 236, + 545, + 270 + ], + "spans": [ + { + "bbox": [ + 305, + 236, + 545, + 270 + ], + "type": "text", + "content": "Figure 3. Summary of the reward model. The model jointly trains a motion encoder and an audio encoder to predict the correct pairings of a batch of (motion, audio) training examples." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 290, + 545, + 410 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 290, + 545, + 410 + ], + "spans": [ + { + "bbox": [ + 304, + 290, + 545, + 410 + ], + "type": "text", + "content": "to optimize the codebook by pushing the " + }, + { + "bbox": [ + 304, + 290, + 545, + 410 + ], + "type": "inline_equation", + "content": "\\hat{e}" + }, + { + "bbox": [ + 304, + 290, + 545, + 410 + ], + "type": "text", + "content": " and its associated " + }, + { + "bbox": [ + 304, + 290, + 545, + 410 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 304, + 290, + 545, + 410 + ], + "type": "text", + "content": " close to " + }, + { + "bbox": [ + 304, + 290, + 545, + 410 + ], + "type": "inline_equation", + "content": "e" + }, + { + "bbox": [ + 304, + 290, + 545, + 410 + ], + "type": "text", + "content": ", which is the output of the encoder. The last term in Equation (4) aims to optimize the encoder by pushing " + }, + { + "bbox": [ + 304, + 290, + 545, + 410 + ], + "type": "inline_equation", + "content": "e" + }, + { + "bbox": [ + 304, + 290, + 545, + 410 + ], + "type": "text", + "content": " close to its nearest latent vector in the codebook. The VQ-VAE mode will be trained separately, which means that its model parameters as well as the codebook remain unchanged during reward computation and reinforcement learning. Note that in the reinforcement learning phase, the indexes of the codebook will be treated as the action tokens, as introduced in Sec. 4.3." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 419, + 400, + 432 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 419, + 400, + 432 + ], + "spans": [ + { + "bbox": [ + 306, + 419, + 400, + 432 + ], + "type": "text", + "content": "4.2. Reward Design" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 437, + 545, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 437, + 545, + 628 + ], + "spans": [ + { + "bbox": [ + 304, + 437, + 545, + 628 + ], + "type": "text", + "content": "Reward function is key to reinforcement learning because it contains most of the knowledge that the agent can learn. Regarding to the task of co-speech gesture synthesis, the original dataset explicitly indicates the ground-truth gestures for given speech audios. However, due to the \"many-to-many\" nature of the problem, a ground-truth gesture may not be optimal under different contexts. This motivates us to learn a reward model that is able to discover and utilize deeper relations between speeches and gestures. There are two advantages of using the reward model. First, if there are multiple gestures correspond to one speech slice, the reward model is able to distinguish between these gestures so that to enrich the supervision signals. Second, as the goal of RL is to maximize the accumulated rewards of action sequences, the agent will learn to generate sequences of gestures that maximize the overall satisfaction." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "content": "To get a proper reward function, we train a model to evaluate the degree of correspondence between the sequences of gestures and the sequences of speeches. Since the reward model tells the agent the degree of correspondence of the speech-gesture pairs, it should be able to distinguish between matched pairs and unmatched pairs. Moreover, the reward model should also generalize to speech-gesture pairs" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2334" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 133 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 133 + ], + "type": "text", + "content": "that are not covered by the dataset. This coincides with the idea of contrastive learning. Multimodal contrastive learning techniques have been developed recently, such as CLIP [27], a well-known visual-text contrastive learning model, designed to match paired image and caption embeddings." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 133, + 288, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 133, + 288, + 228 + ], + "spans": [ + { + "bbox": [ + 46, + 133, + 288, + 228 + ], + "type": "text", + "content": "Similar to CLIP, given a batch of " + }, + { + "bbox": [ + 46, + 133, + 288, + 228 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 46, + 133, + 288, + 228 + ], + "type": "text", + "content": " (motion, audio) pairs, the reward model is trained to predict which of the " + }, + { + "bbox": [ + 46, + 133, + 288, + 228 + ], + "type": "inline_equation", + "content": "n \\times n" + }, + { + "bbox": [ + 46, + 133, + 288, + 228 + ], + "type": "text", + "content": " possible (motion, audio) pairs across a batch actually occurred. To do this, the model converts each input motion " + }, + { + "bbox": [ + 46, + 133, + 288, + 228 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 46, + 133, + 288, + 228 + ], + "type": "text", + "content": " and audio " + }, + { + "bbox": [ + 46, + 133, + 288, + 228 + ], + "type": "inline_equation", + "content": "\\mathcal{U}" + }, + { + "bbox": [ + 46, + 133, + 288, + 228 + ], + "type": "text", + "content": " into vector representations " + }, + { + "bbox": [ + 46, + 133, + 288, + 228 + ], + "type": "inline_equation", + "content": "\\boldsymbol{m}" + }, + { + "bbox": [ + 46, + 133, + 288, + 228 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 133, + 288, + 228 + ], + "type": "inline_equation", + "content": "\\boldsymbol{u}" + }, + { + "bbox": [ + 46, + 133, + 288, + 228 + ], + "type": "text", + "content": " by a motion encoder " + }, + { + "bbox": [ + 46, + 133, + 288, + 228 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_m" + }, + { + "bbox": [ + 46, + 133, + 288, + 228 + ], + "type": "text", + "content": " and an audio encoder " + }, + { + "bbox": [ + 46, + 133, + 288, + 228 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_a" + }, + { + "bbox": [ + 46, + 133, + 288, + 228 + ], + "type": "text", + "content": ". Two encoders are jointly trained to maximize the cosine similarity of each real pair " + }, + { + "bbox": [ + 46, + 133, + 288, + 228 + ], + "type": "inline_equation", + "content": "(m_i" + }, + { + "bbox": [ + 46, + 133, + 288, + 228 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 133, + 288, + 228 + ], + "type": "inline_equation", + "content": "u_i)" + }, + { + "bbox": [ + 46, + 133, + 288, + 228 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 118, + 236, + 287, + 251 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 118, + 236, + 287, + 251 + ], + "spans": [ + { + "bbox": [ + 118, + 236, + 287, + 251 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {R} = \\ell_ {i} ^ {\\boldsymbol {m} \\rightarrow \\boldsymbol {u}} + \\ell_ {i} ^ {\\boldsymbol {u} \\rightarrow \\boldsymbol {m}}, \\tag {6}", + "image_path": "1d7e9d5829059329b91dfabfaa881b7325b15b7e6ccf90ab54014159fc711d6a.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 259, + 75, + 270 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 259, + 75, + 270 + ], + "spans": [ + { + "bbox": [ + 47, + 259, + 75, + 270 + ], + "type": "text", + "content": "where" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 85, + 277, + 249, + 308 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 85, + 277, + 249, + 308 + ], + "spans": [ + { + "bbox": [ + 85, + 277, + 249, + 308 + ], + "type": "interline_equation", + "content": "\\ell_ {i} ^ {\\boldsymbol {m} \\rightarrow \\boldsymbol {u}} = - \\log \\frac {e x p (m _ {i} \\cdot u _ {i} / \\tau)}{\\sum_ {k = 1} ^ {n} e x p (m _ {i} \\cdot u _ {k} / \\tau)},", + "image_path": "93dcd473c247575b9a098c7c9b3ee4934c71f0db09630828e6c8e80b9e43b98a.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 87, + 306, + 247, + 332 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 87, + 306, + 247, + 332 + ], + "spans": [ + { + "bbox": [ + 87, + 306, + 247, + 332 + ], + "type": "interline_equation", + "content": "\\ell_ {i} ^ {\\boldsymbol {u} \\rightarrow \\boldsymbol {m}} = - \\log \\frac {\\exp (u _ {i} \\cdot m _ {i} / \\tau)}{\\sum_ {k = 1} ^ {n} \\exp (u _ {i} \\cdot m _ {k} / \\tau)}.", + "image_path": "9fec33ace80c6310e14530549edaaf9de74baa8d0371a455a1ec68e3b1f40b26.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 340, + 287, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 340, + 287, + 376 + ], + "spans": [ + { + "bbox": [ + 46, + 340, + 287, + 376 + ], + "type": "inline_equation", + "content": "\\tau \\in \\mathbb{R}^{+}" + }, + { + "bbox": [ + 46, + 340, + 287, + 376 + ], + "type": "text", + "content": " represents a temperature parameter. In this work, both encoders are structured as 1-D temporal convolutional networks and can input sequences of different lengths." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 376, + 287, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 376, + 287, + 437 + ], + "spans": [ + { + "bbox": [ + 46, + 376, + 287, + 437 + ], + "type": "text", + "content": "As with VQ-VAE, the training of the reward model is in a separate phase. During reinforcement learning, two encoders of the learned reward model respectively encode the generated motion " + }, + { + "bbox": [ + 46, + 376, + 287, + 437 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{M}}" + }, + { + "bbox": [ + 46, + 376, + 287, + 437 + ], + "type": "text", + "content": " and input speech " + }, + { + "bbox": [ + 46, + 376, + 287, + 437 + ], + "type": "inline_equation", + "content": "\\mathcal{U}" + }, + { + "bbox": [ + 46, + 376, + 287, + 437 + ], + "type": "text", + "content": " into two vectors, with their inner product as the reward:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 446, + 287, + 461 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 105, + 446, + 287, + 461 + ], + "spans": [ + { + "bbox": [ + 105, + 446, + 287, + 461 + ], + "type": "interline_equation", + "content": "R (\\hat {\\mathcal {M}}, \\mathcal {U}) = \\mathcal {E} _ {m} (\\hat {\\mathcal {M}}) \\cdot \\mathcal {E} _ {u} (\\mathcal {U}). \\tag {8}", + "image_path": "2c3862a0cc62a75812a9d3b09c962b295dbcb175ea76d6fb6def1fabf092890d.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 468, + 221, + 480 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 468, + 221, + 480 + ], + "spans": [ + { + "bbox": [ + 47, + 468, + 221, + 480 + ], + "type": "text", + "content": "4.3. Offline Reinforcement Learning" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 486, + 287, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 486, + 287, + 629 + ], + "spans": [ + { + "bbox": [ + 46, + 486, + 287, + 629 + ], + "type": "text", + "content": "Recall that we have defined several essential factors of the MDP. State " + }, + { + "bbox": [ + 46, + 486, + 287, + 629 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 46, + 486, + 287, + 629 + ], + "type": "text", + "content": " includes the generated action tokens " + }, + { + "bbox": [ + 46, + 486, + 287, + 629 + ], + "type": "inline_equation", + "content": "\\pmb{a}_{0:t-1}" + }, + { + "bbox": [ + 46, + 486, + 287, + 629 + ], + "type": "text", + "content": " and the audio sequence " + }, + { + "bbox": [ + 46, + 486, + 287, + 629 + ], + "type": "inline_equation", + "content": "\\pmb{u}_{0:t}" + }, + { + "bbox": [ + 46, + 486, + 287, + 629 + ], + "type": "text", + "content": " up to the current time step " + }, + { + "bbox": [ + 46, + 486, + 287, + 629 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 486, + 287, + 629 + ], + "type": "text", + "content": ". Each action " + }, + { + "bbox": [ + 46, + 486, + 287, + 629 + ], + "type": "inline_equation", + "content": "a" + }, + { + "bbox": [ + 46, + 486, + 287, + 629 + ], + "type": "text", + "content": " corresponds to an index of the codebook learned by VQ-VAE. The reward " + }, + { + "bbox": [ + 46, + 486, + 287, + 629 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 46, + 486, + 287, + 629 + ], + "type": "text", + "content": " is outputted by the reward model to evaluate how well the predicted gesture matches the input audio. Now we are able to augment the original speech-gesture dataset to a trajectory dataset " + }, + { + "bbox": [ + 46, + 486, + 287, + 629 + ], + "type": "inline_equation", + "content": "\\mathcal{D} = \\{s_i, a_i, r_i, s_i' | i = 1, \\dots, N\\}" + }, + { + "bbox": [ + 46, + 486, + 287, + 629 + ], + "type": "text", + "content": ". Note that these trajectories can be collected before the start of the reinforcement learning phase. We denote by " + }, + { + "bbox": [ + 46, + 486, + 287, + 629 + ], + "type": "inline_equation", + "content": "\\pi_{\\beta}" + }, + { + "bbox": [ + 46, + 486, + 287, + 629 + ], + "type": "text", + "content": " the behavior policy that generates these trajectories." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 630, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 630, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 630, + 287, + 713 + ], + "type": "text", + "content": "Offline datasets usually do not provide complete state-action coverage. That is, the set " + }, + { + "bbox": [ + 46, + 630, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\{(s,a)|(s,a,r,s')\\in \\mathcal{D}\\}" + }, + { + "bbox": [ + 46, + 630, + 287, + 713 + ], + "type": "text", + "content": " is typically a small subset of the full space " + }, + { + "bbox": [ + 46, + 630, + 287, + 713 + ], + "type": "inline_equation", + "content": "S\\times \\mathcal{A}" + }, + { + "bbox": [ + 46, + 630, + 287, + 713 + ], + "type": "text", + "content": ". Standard online reinforcement learning methods such as DDPG [24] and SAC [31] are not appropriate due to issues with bootstrapping from out-of-distribution actions. A \"safe\" strategy to mitigate the distributional shift problem is to be con" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 72, + 545, + 191 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 191 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 191 + ], + "type": "text", + "content": "servative: if we explicitly estimate the value of unseen outcomes conservatively (i.e. assign them a low value), then the estimated value or performance of the policy that executes unseen behaviors is guaranteed to be small. Using such conservative estimates for policy optimization will prevent the policy from executing unseen actions and it will perform reliably. To this end, we adopt a simple but effective offline reinforcement learning algorithm called conservative Q-learning (CQL) [20] to learn the optimal policy from the augmented offline dataset." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 192, + 545, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 192, + 545, + 228 + ], + "spans": [ + { + "bbox": [ + 304, + 192, + 545, + 228 + ], + "type": "text", + "content": "Specifically, for the conservative off-policy evaluation, the Q-function, " + }, + { + "bbox": [ + 304, + 192, + 545, + 228 + ], + "type": "inline_equation", + "content": "\\hat{Q}^{\\pi} := \\lim_{k \\to \\infty} \\hat{Q}_k" + }, + { + "bbox": [ + 304, + 192, + 545, + 228 + ], + "type": "text", + "content": " is trained via an iterative update:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 310, + 244, + 545, + 319 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 310, + 244, + 545, + 319 + ], + "spans": [ + { + "bbox": [ + 310, + 244, + 545, + 319 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\hat {Q} ^ {k + 1} \\leftarrow \\arg \\min _ {Q} \\frac {1}{2} \\underset {s, a, s ^ {\\prime} \\sim \\mathcal {D}} {\\mathbb {E}} \\left[ (Q (s, a) - \\hat {\\mathcal {B}} ^ {\\pi} \\hat {Q} ^ {k} (s, a)) ^ {2} \\right] \\\\ + \\alpha \\left(\\underset {a \\sim \\mu (a | s)} {\\mathbb {E}} [ Q (s, a) ] - \\underset {a \\sim \\hat {\\pi} _ {\\beta} (a | s)} {\\mathbb {E}} [ Q (s, a) ]\\right), \\tag {9} \\\\ \\end{array}", + "image_path": "ee8f95f17d8abcc4606e6f53143868b1b4dc0a7188ebd1605ea02b78b6f1a5f3.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 320, + 334, + 330 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 320, + 334, + 330 + ], + "spans": [ + { + "bbox": [ + 306, + 320, + 334, + 330 + ], + "type": "text", + "content": "where" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 338, + 556, + 363 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 338, + 556, + 363 + ], + "spans": [ + { + "bbox": [ + 306, + 338, + 556, + 363 + ], + "type": "interline_equation", + "content": "\\hat {\\mathcal {B}} ^ {\\pi} \\hat {Q} ^ {k} (s, a) = \\mathbb {E} _ {s ^ {\\prime} \\sim P \\left(s ^ {\\prime} \\mid s, a\\right)} [ R (s, a) + \\gamma \\mathbb {E} _ {a ^ {\\prime} \\sim \\hat {\\pi} ^ {k} (s, a)} \\hat {Q} ^ {k} \\left(s ^ {\\prime}, a ^ {\\prime}\\right) ] \\tag {10}", + "image_path": "deded8a2b57fa37bb2fa0110acc968a0d8c1d37509d62e965e52c0c74e20243a.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 365, + 545, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 365, + 545, + 472 + ], + "spans": [ + { + "bbox": [ + 304, + 365, + 545, + 472 + ], + "type": "text", + "content": "is the Bellman operator. This equation consists of the normal bootstrapping error and the regularization term with a trade-off factor " + }, + { + "bbox": [ + 304, + 365, + 545, + 472 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 304, + 365, + 545, + 472 + ], + "type": "text", + "content": ". Inside the regularization term, the first term always pushes the Q-value down on the " + }, + { + "bbox": [ + 304, + 365, + 545, + 472 + ], + "type": "inline_equation", + "content": "(s,a)" + }, + { + "bbox": [ + 304, + 365, + 545, + 472 + ], + "type": "text", + "content": " pairs sampled from the learning policy " + }, + { + "bbox": [ + 304, + 365, + 545, + 472 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 304, + 365, + 545, + 472 + ], + "type": "text", + "content": " whereas the second term pushes Q-value up on the " + }, + { + "bbox": [ + 304, + 365, + 545, + 472 + ], + "type": "inline_equation", + "content": "(s,a)" + }, + { + "bbox": [ + 304, + 365, + 545, + 472 + ], + "type": "text", + "content": " pairs sampled from the offline data set. In practice, we selected the binding variant of CQL and DQN as our model, which can be optimized by minimizing the following loss function:" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 488, + 545, + 557 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 488, + 545, + 557 + ], + "spans": [ + { + "bbox": [ + 317, + 488, + 545, + 557 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\underset {Q} {\\operatorname {m i n}} \\alpha_ {s \\sim \\mathcal {D}} \\left[ \\log \\sum_ {a} \\exp (Q (s, a)) - \\underset {a \\sim \\hat {\\pi} _ {\\beta}} {\\mathbb {E}} [ Q (s, a) ] \\right] \\\\ + \\frac {1}{2} _ {s, a, s ^ {\\prime} \\sim \\mathcal {D}} \\mathbb {E} \\left[ (Q (s, a) - \\hat {\\mathcal {B}} ^ {\\pi} \\hat {Q} ^ {k} (s, a)) ^ {2} \\right]. \\tag {11} \\\\ \\end{array}", + "image_path": "420ec536c6e1440308feea9cd4c2883f96b87bf5029d2cf01e70a1734c5ba518.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 558, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 558, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 558, + 545, + 713 + ], + "type": "text", + "content": "Since we want to use the Q-value function to autoregressively generate the action sequences, we construct a powerful GPT-based backbone model to estimate the Q-values, as is shown in Fig. 1. To construct the input, each motion code sequence is embedded to learnable features and concatenated with the audio features in the temporal dimension. We add a learned positional encoding to this concatenated " + }, + { + "bbox": [ + 304, + 558, + 545, + 713 + ], + "type": "inline_equation", + "content": "(2\\times T^{\\prime}\\times C)" + }, + { + "bbox": [ + 304, + 558, + 545, + 713 + ], + "type": "text", + "content": "-dimensional tensor and feed it to Transformer layers, whose structure is shown in the left of Fig. 1. Last, we employ a linear layer to map the output of Transformer layers to the Q-value " + }, + { + "bbox": [ + 304, + 558, + 545, + 713 + ], + "type": "inline_equation", + "content": "q\\in \\mathbb{R}^{T^{\\prime}\\times N}" + }, + { + "bbox": [ + 304, + 558, + 545, + 713 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 558, + 545, + 713 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 558, + 545, + 713 + ], + "type": "text", + "content": " is the size of learned codebook. In the Transformer layer, the attention is the core component that determines the computational" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2335" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 96 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 96 + ], + "type": "text", + "content": "dependency among sequential elements of inputs, which is computed as:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 53, + 112, + 287, + 148 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 112, + 287, + 148 + ], + "spans": [ + { + "bbox": [ + 53, + 112, + 287, + 148 + ], + "type": "interline_equation", + "content": "\\operatorname {A t t e n t i o n} (\\boldsymbol {Q}, \\boldsymbol {K}, \\boldsymbol {V}, \\boldsymbol {M}) = \\operatorname {s o f t m a x} \\left(\\frac {\\boldsymbol {Q} \\boldsymbol {K} ^ {T}}{\\sqrt {C}} + \\boldsymbol {M}\\right) \\boldsymbol {V}, \\tag {12}", + "image_path": "c229aa776eb9a8988b6ed4ede4be76bdb57136eafb11a7ec894f4cc3059c6002.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 149, + 287, + 268 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 149, + 287, + 268 + ], + "spans": [ + { + "bbox": [ + 46, + 149, + 287, + 268 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 149, + 287, + 268 + ], + "type": "inline_equation", + "content": "Q, K, V" + }, + { + "bbox": [ + 46, + 149, + 287, + 268 + ], + "type": "text", + "content": " denote the query, key and value from input, and " + }, + { + "bbox": [ + 46, + 149, + 287, + 268 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 46, + 149, + 287, + 268 + ], + "type": "text", + "content": " is the mask. In the original GPT model, \"causal attention\" realizes the intercommunication of the current and previous data to compute the state for the time of interest. However, our model needs to infer actions from the speech at each time step, so it needs to process two input sequences simultaneously. To comply the causality conditioned among features of the audio and the action, " + }, + { + "bbox": [ + 46, + 149, + 287, + 268 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 46, + 149, + 287, + 268 + ], + "type": "text", + "content": " is designed to be a " + }, + { + "bbox": [ + 46, + 149, + 287, + 268 + ], + "type": "inline_equation", + "content": "2 \\times 2" + }, + { + "bbox": [ + 46, + 149, + 287, + 268 + ], + "type": "text", + "content": " repeated block matrix with lower triangular matrix of size " + }, + { + "bbox": [ + 46, + 149, + 287, + 268 + ], + "type": "inline_equation", + "content": "T'" + }, + { + "bbox": [ + 46, + 149, + 287, + 268 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 269, + 287, + 329 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 269, + 287, + 329 + ], + "spans": [ + { + "bbox": [ + 46, + 269, + 287, + 329 + ], + "type": "text", + "content": "During inference, given the audio of arbitrary length and the initial motion code " + }, + { + "bbox": [ + 46, + 269, + 287, + 329 + ], + "type": "inline_equation", + "content": "a_0" + }, + { + "bbox": [ + 46, + 269, + 287, + 329 + ], + "type": "text", + "content": ", the Q-network can output Q-values of all actions. We will select the action with the maximum Q-value and then add it to the input, so that we can autoregressively generate the entire action sequence." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 339, + 128, + 352 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 339, + 128, + 352 + ], + "spans": [ + { + "bbox": [ + 47, + 339, + 128, + 352 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 358, + 105, + 370 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 358, + 105, + 370 + ], + "spans": [ + { + "bbox": [ + 47, + 358, + 105, + 370 + ], + "type": "text", + "content": "5.1. Dataset" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 377, + 287, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 377, + 287, + 593 + ], + "spans": [ + { + "bbox": [ + 46, + 377, + 287, + 593 + ], + "type": "text", + "content": "In this paper, we use two speech-gesture datasets: the Trinity dataset and a Chinese dataset collected by us. Trinity Gesture dataset is a large database of speech and gestures jointly collected by [8]. This dataset consists of 242 minutes of motion capture and audio of one male actor talking on different topics. The actor's motion was captured with a 20-camera Vicon system and solved onto a skeleton with 69 joints. In this paper, we use the official release version of The GENEA Challenge 2020 Dataset [19], where 221 minutes are used as training data, and the remaining 21 minutes are kept for testing. Additionally, we collected an 1-hour high-quality Chinese dataset using motion capture equipment. This dataset contains 3D full-body gestures and the aligned Chinese speech audio. And we retargeted the skeletons to be consistent with the Trinity dataset. In this dataset, the speaker described various speech scenes in Chinese and made high-standard common speech gestures, ensuring a high degree of matching between gestures and speech." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 599, + 180, + 612 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 599, + 180, + 612 + ], + "spans": [ + { + "bbox": [ + 47, + 599, + 180, + 612 + ], + "type": "text", + "content": "5.2. Implementation Details" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 618, + 287, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 618, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 46, + 618, + 287, + 712 + ], + "type": "text", + "content": "All the motion data are downsampled to 20 frames per second on the Trinity and Chinese datasets. We focus on upper-body gestures in this work. For visualization, we employ a character model consist of 16 upper-body joints, including a rotational root. The audio features are extracted by the public audio processing toolbox librosa, including mel-scaled spectrogram and onset. In VQ-VAE, the codebook size " + }, + { + "bbox": [ + 46, + 618, + 287, + 712 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 618, + 287, + 712 + ], + "type": "text", + "content": " and the channel dimension " + }, + { + "bbox": [ + 46, + 618, + 287, + 712 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 46, + 618, + 287, + 712 + ], + "type": "text", + "content": " are both set to" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 72, + 545, + 371 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 371 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 371 + ], + "type": "text", + "content": "512, and the temporal downsampling rate of encoders is set to 8. During the training of VQ-VAE, motion data are cropped to uniform length of " + }, + { + "bbox": [ + 304, + 72, + 545, + 371 + ], + "type": "inline_equation", + "content": "T = 120" + }, + { + "bbox": [ + 304, + 72, + 545, + 371 + ], + "type": "text", + "content": " (6 seconds) and sampled in batch size of 64. The trade-off " + }, + { + "bbox": [ + 304, + 72, + 545, + 371 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 304, + 72, + 545, + 371 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 304, + 72, + 545, + 371 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{VQ}" + }, + { + "bbox": [ + 304, + 72, + 545, + 371 + ], + "type": "text", + "content": " is set to 0.1. " + }, + { + "bbox": [ + 304, + 72, + 545, + 371 + ], + "type": "inline_equation", + "content": "\\alpha_{1}" + }, + { + "bbox": [ + 304, + 72, + 545, + 371 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 72, + 545, + 371 + ], + "type": "inline_equation", + "content": "\\alpha_{2}" + }, + { + "bbox": [ + 304, + 72, + 545, + 371 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 304, + 72, + 545, + 371 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_m" + }, + { + "bbox": [ + 304, + 72, + 545, + 371 + ], + "type": "text", + "content": " are both 1. We adopt Adam optimizer with " + }, + { + "bbox": [ + 304, + 72, + 545, + 371 + ], + "type": "inline_equation", + "content": "\\beta_{1} = 0.9" + }, + { + "bbox": [ + 304, + 72, + 545, + 371 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 72, + 545, + 371 + ], + "type": "inline_equation", + "content": "\\beta_{2} = 0.999" + }, + { + "bbox": [ + 304, + 72, + 545, + 371 + ], + "type": "text", + "content": " for 500 epochs on Trinity but 400 epochs on the Chinese dataset with learning rate 0.00003. During the training of reward model, both audio and motion data are clipped to 6s. The optimizer settings are consistent with those of the VQ-VAE. The structure of the Q network is basically the same as that of the GPT network, where the embedding dimension is 768 and the attention layer is implemented in 12 heads with dropout probability 0.1. During the off-line reinforcement learning, the motion sequences are encoded to motion codes and sampled to length of 15 and sampled in batch size of 256. The Q-network is optimized using Adam optimizer with " + }, + { + "bbox": [ + 304, + 72, + 545, + 371 + ], + "type": "inline_equation", + "content": "\\beta_{1} = 0.5" + }, + { + "bbox": [ + 304, + 72, + 545, + 371 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 72, + 545, + 371 + ], + "type": "inline_equation", + "content": "\\beta_{2} = 0.999" + }, + { + "bbox": [ + 304, + 72, + 545, + 371 + ], + "type": "text", + "content": " for 1500 epochs on Trinity but 1200 epochs on Chinese dataset, where the learning rate is initialized as 0.0001. The discount factor " + }, + { + "bbox": [ + 304, + 72, + 545, + 371 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 304, + 72, + 545, + 371 + ], + "type": "text", + "content": " is set to 0.99. At runtime, a Gaussian filter with a kernel size of " + }, + { + "bbox": [ + 304, + 72, + 545, + 371 + ], + "type": "inline_equation", + "content": "K = 5" + }, + { + "bbox": [ + 304, + 72, + 545, + 371 + ], + "type": "text", + "content": " is used to smooth the denormalized gesture sequence along the time dimension. The entire framework takes 36 hours on Trinity and 30 hours on the Chinese dataset to complete training on one NVIDIA Quadro RTX 5000 GPU." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 376, + 418, + 388 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 376, + 418, + 388 + ], + "spans": [ + { + "bbox": [ + 306, + 376, + 418, + 388 + ], + "type": "text", + "content": "5.3. Evaluation Metrics" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 395, + 545, + 610 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 395, + 545, + 610 + ], + "spans": [ + { + "bbox": [ + 304, + 395, + 545, + 610 + ], + "type": "text", + "content": "We adopt four commonly used evaluation metrics (MAJE, MAD, FGD and PMB) to compare various methods quantitatively. Mean absolute joint error (MAJE) measures the mean of the absolute errors between the generated joint positions and ground truth over all the time steps, which indicates how closely the generated joint positions follow the ground truth. Mean acceleration difference (MAD) measures the mean of the " + }, + { + "bbox": [ + 304, + 395, + 545, + 610 + ], + "type": "inline_equation", + "content": "l2" + }, + { + "bbox": [ + 304, + 395, + 545, + 610 + ], + "type": "text", + "content": " norm differences between the generated joint accelerations and ground truth over all the time steps, which indicates how closely the ground truth and the generated joint movements match. Fréchet Gesture Distance (FGD) was proposed by which measures the difference between the distributions of the latent features of the generated gestures and ground truth [38]. The latent features are extracted from an auto-encoder trained on the Human 3.6M dataset*. Specifically, we use " + }, + { + "bbox": [ + 304, + 395, + 545, + 610 + ], + "type": "inline_equation", + "content": "\\hat{H}" + }, + { + "bbox": [ + 304, + 395, + 545, + 610 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 395, + 545, + 610 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 304, + 395, + 545, + 610 + ], + "type": "text", + "content": " represent the latent features of generated gestures and ground truth. The Fréchet distance is formulated as" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 318, + 624, + 545, + 654 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 624, + 545, + 654 + ], + "spans": [ + { + "bbox": [ + 318, + 624, + 545, + 654 + ], + "type": "interline_equation", + "content": "\\operatorname {F G D} (\\hat {H}, H) = \\left| \\left| \\hat {\\mu} - \\mu \\right| \\right| _ {2} ^ {2} + \\operatorname {T r} \\left(\\hat {\\Sigma} + \\Sigma - 2 \\sqrt {\\hat {\\Sigma} \\Sigma}\\right), \\tag {13}", + "image_path": "d5d8a16420b8137286fb0d817e2f423c7d465f0f7dc7437c6f296a5b8b14dfbd.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 654, + 545, + 678 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 545, + 678 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 545, + 678 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 654, + 545, + 678 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 304, + 654, + 545, + 678 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 654, + 545, + 678 + ], + "type": "inline_equation", + "content": "\\Sigma" + }, + { + "bbox": [ + 304, + 654, + 545, + 678 + ], + "type": "text", + "content": " are the first and second moments of the latent feature distribution of the ground truth " + }, + { + "bbox": [ + 304, + 654, + 545, + 678 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 304, + 654, + 545, + 678 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 654, + 545, + 678 + ], + "type": "inline_equation", + "content": "\\hat{\\mu}" + }, + { + "bbox": [ + 304, + 654, + 545, + 678 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 654, + 545, + 678 + ], + "type": "inline_equation", + "content": "\\hat{\\Sigma}" + }, + { + "bbox": [ + 304, + 654, + 545, + 678 + ], + "type": "text", + "content": " are" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 306, + 683, + 545, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 683, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 683, + 545, + 713 + ], + "type": "text", + "content": "*We obtain the codes and models from: https://github.com/ai4r/Gesture-Generation-from-Trimodal-Context.git" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2336" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 100, + 102, + 493, + 240 + ], + "blocks": [ + { + "bbox": [ + 46, + 70, + 545, + 94 + ], + "lines": [ + { + "bbox": [ + 46, + 70, + 545, + 94 + ], + "spans": [ + { + "bbox": [ + 46, + 70, + 545, + 94 + ], + "type": "text", + "content": "Table 1. Comparison of RACER to Style Gesture, Gesticulator, S2AG and Bailando on Trinity datasets. " + }, + { + "bbox": [ + 46, + 70, + 545, + 94 + ], + "type": "inline_equation", + "content": "\\downarrow" + }, + { + "bbox": [ + 46, + 70, + 545, + 94 + ], + "type": "text", + "content": " means the lower is better and " + }, + { + "bbox": [ + 46, + 70, + 545, + 94 + ], + "type": "inline_equation", + "content": "\\uparrow" + }, + { + "bbox": [ + 46, + 70, + 545, + 94 + ], + "type": "text", + "content": " means the higher is better." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 100, + 102, + 493, + 240 + ], + "lines": [ + { + "bbox": [ + 100, + 102, + 493, + 240 + ], + "spans": [ + { + "bbox": [ + 100, + 102, + 493, + 240 + ], + "type": "table", + "html": "
DatasetMethodMAJE(mm) ↓MAD(mm/s2) ↓FGD↓PMB(%) ↑
TrinityGround Truth00095.74
Style Gesture97.294.2636.9854.54
Gesticulator82.413.6231.0471.00
S2AG54.931.4920.3679.53
Bailando61.891.7417.2984.21
Ours50.331.2113.4489.58
Chinese DatasetStyle Gesture62.953.4630.1153.56
Bailando45.451.2925.0768.53
Ours43.251.199.2174.29
", + "image_path": "4743a398604ebafc3d5f9f060bb87a01a0e2df00753d5f6f064cdc726ebc942e.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 263, + 289, + 360 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 263, + 289, + 360 + ], + "spans": [ + { + "bbox": [ + 46, + 263, + 289, + 360 + ], + "type": "text", + "content": "the first and second moments of the latent feature distribution of the generated gestures " + }, + { + "bbox": [ + 46, + 263, + 289, + 360 + ], + "type": "inline_equation", + "content": "\\hat{H}" + }, + { + "bbox": [ + 46, + 263, + 289, + 360 + ], + "type": "text", + "content": ". FGD makes a reasonable assessment of the perceived plausibility of the synthesized gestures. The Percentage of Matched Beats PMB is a new metric to calculating the matching rate of audio and motion beats for evaluating rhythm performance [3]. In PMB, a motion beat is considered an matched if the distance to a nearby audio beat is within the range of " + }, + { + "bbox": [ + 46, + 263, + 289, + 360 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 46, + 263, + 289, + 360 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 367, + 287, + 411 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 367, + 287, + 411 + ], + "spans": [ + { + "bbox": [ + 53, + 367, + 287, + 411 + ], + "type": "interline_equation", + "content": "\\operatorname {P M B} \\left(B ^ {m}, B ^ {a}\\right) = \\frac {1}{N _ {m}} \\sum_ {i = 1} ^ {N _ {m}} \\sum_ {j = b _ {*} ^ {a} + 1} ^ {N _ {a}} \\mathbb {1} \\left[ \\left| \\left| b _ {j} ^ {m} - b _ {j} ^ {a} \\right| \\right| _ {1} < \\delta \\right], \\tag {14}", + "image_path": "4b177d88285b3cce2f1a701a9a5ce9f2e77a22cae0083dc12a90800458719c95.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 412, + 288, + 474 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 412, + 288, + 474 + ], + "spans": [ + { + "bbox": [ + 46, + 412, + 288, + 474 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 412, + 288, + 474 + ], + "type": "inline_equation", + "content": "B^{m}" + }, + { + "bbox": [ + 46, + 412, + 288, + 474 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 412, + 288, + 474 + ], + "type": "inline_equation", + "content": "B^{a}" + }, + { + "bbox": [ + 46, + 412, + 288, + 474 + ], + "type": "text", + "content": " represent the motion beats and audio beats, and " + }, + { + "bbox": [ + 46, + 412, + 288, + 474 + ], + "type": "inline_equation", + "content": "N_{m}" + }, + { + "bbox": [ + 46, + 412, + 288, + 474 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 412, + 288, + 474 + ], + "type": "inline_equation", + "content": "N_{a}" + }, + { + "bbox": [ + 46, + 412, + 288, + 474 + ], + "type": "text", + "content": " represent the number of beats. Motion beat can be obtained by identifying the local minima of joints deceleration [14]. " + }, + { + "bbox": [ + 46, + 412, + 288, + 474 + ], + "type": "inline_equation", + "content": "b_{*}^{a}" + }, + { + "bbox": [ + 46, + 412, + 288, + 474 + ], + "type": "text", + "content": " indicates the last matched audio beat. The " + }, + { + "bbox": [ + 46, + 412, + 288, + 474 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 46, + 412, + 288, + 474 + ], + "type": "text", + "content": " is set to 0.2s, consistent with the original paper." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 479, + 225, + 493 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 479, + 225, + 493 + ], + "spans": [ + { + "bbox": [ + 47, + 479, + 225, + 493 + ], + "type": "text", + "content": "5.4. Comparison to Existing Methods" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 498, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 498, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 498, + 289, + 715 + ], + "type": "text", + "content": "In this section, we compare RACER with several state-of-the-art methods to demonstrate the advances made by our method. On the Trinity dataset, we compare with the methods of Style Gesture [2], Gesticulator [18], Speed to Affective Gesture (S2AG) [4] and Bailando [29]. On the Chinese dataset, we compared with Style Gesture and Bailando. Style Gesture generates gestures based on only the speech audio features by a normalizing flow probabilistic model. Gesticulator leverages the audio and text of the speech to generate semantically consistent gestures. Compared with these methods, S2AG adds more modality, including speaker identity and seed gesture poses. Different from other methods, Bailando is a novel music-to-dance framework with a choreographic memory and actor-critic GPT model. Since Bailando's model architecture is similar to ours and both incorporate reinforcement learning into the training process, we apply it as a key contrast method to the gesture synthesis task. The hyperparameter setting is" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 263, + 545, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 263, + 545, + 312 + ], + "spans": [ + { + "bbox": [ + 304, + 263, + 545, + 312 + ], + "type": "text", + "content": "consistent with the original paper and the open source code, and the feature processing of the training data is consistent with ours. For a fair comparison, we keep the same skeleton and motion frame rate as the comparison systems." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 315, + 546, + 543 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 315, + 546, + 543 + ], + "spans": [ + { + "bbox": [ + 304, + 315, + 546, + 543 + ], + "type": "text", + "content": "Quantitative experimental results are listed in Tab. 1. According to the comparison, our proposed model consistently performs favorably against all the other methods on all evaluations. Specially, on Trinity, our method improves " + }, + { + "bbox": [ + 304, + 315, + 546, + 543 + ], + "type": "inline_equation", + "content": "4.60(8.37\\%)" + }, + { + "bbox": [ + 304, + 315, + 546, + 543 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 315, + 546, + 543 + ], + "type": "inline_equation", + "content": "0.28(18.79\\%)" + }, + { + "bbox": [ + 304, + 315, + 546, + 543 + ], + "type": "text", + "content": " than the best compared baseline model S2AG on MAJE and MAD, respectively. In terms of FGD and PMB, Bailando, with its advanced network structure and reinforcement learning fine-tuning process, also obtained better evaluation results (17.29 and " + }, + { + "bbox": [ + 304, + 315, + 546, + 543 + ], + "type": "inline_equation", + "content": "84.21\\%" + }, + { + "bbox": [ + 304, + 315, + 546, + 543 + ], + "type": "text", + "content": " ) than S2AC (20.36 and " + }, + { + "bbox": [ + 304, + 315, + 546, + 543 + ], + "type": "inline_equation", + "content": "79.53\\%" + }, + { + "bbox": [ + 304, + 315, + 546, + 543 + ], + "type": "text", + "content": "). Our approach further achieves significantly higher performance (13.44 and " + }, + { + "bbox": [ + 304, + 315, + 546, + 543 + ], + "type": "inline_equation", + "content": "89.58\\%" + }, + { + "bbox": [ + 304, + 315, + 546, + 543 + ], + "type": "text", + "content": ") in both items. Similarly, our method has also achieved excellent performance in Chinese dataset compared with Style Gesture and Bailando, especially on FGD (9.21) and PMB " + }, + { + "bbox": [ + 304, + 315, + 546, + 543 + ], + "type": "inline_equation", + "content": "(74.29\\%)" + }, + { + "bbox": [ + 304, + 315, + 546, + 543 + ], + "type": "text", + "content": ". The superiority of our method on all evaluation metrics reveals that RACER can rely on offline RL not only to synthesize more real-like motion than compared baseline methods, but also but also to achieve outstanding performance in motion quality and rhythm." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 545, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 545, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 545, + 547, + 714 + ], + "type": "text", + "content": "User Study. We conducted a user study to further assess the real visual performance of our method in a pairwise manner. We choose Style Gesture and Bailando to compare. We randomly sampled 10 30-second long slices generated from the test set of Trinity dataset. Each motion slice is bound to the Mixamo [25] model and rendered as a video by Blender. The experiment is conducted with 12 participants. The participants are asked to rate the slices from the following three aspects respectively: (1) realism, (2) speech-to-gesture matching, (3) speech-to-gesture rhythm matching. The scores assigned to each rating are in the range of 1-5, corresponding from worst to best. Rating statistics results are shown in Fig. 4. Notably, our method significantly surpasses the compared other methods." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "2337" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 82, + 91, + 513, + 182 + ], + "blocks": [ + { + "bbox": [ + 200, + 71, + 394, + 82 + ], + "lines": [ + { + "bbox": [ + 200, + 71, + 394, + 82 + ], + "spans": [ + { + "bbox": [ + 200, + 71, + 394, + 82 + ], + "type": "text", + "content": "Table 2. Ablation study results on the Trinity dataset." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 82, + 91, + 513, + 182 + ], + "lines": [ + { + "bbox": [ + 82, + 91, + 513, + 182 + ], + "spans": [ + { + "bbox": [ + 82, + 91, + 513, + 182 + ], + "type": "table", + "html": "
MethodMAJE(mm) ↓MAD(mm/s2) ↓FGD↓PMB(%) ↑Reward↑
RACER(Supervised Learning)56.221.8116.6978.54-
RACER(Distance Reward)58.962.9423.5572.09-
RACER(DQN)59.441.8115.4676.634.23
RACER(Actor-Critic)60.181.7617.6978.023.93
Ours50.331.2113.4489.589.34
", + "image_path": "ed8dd396ec7bcff39edea51b308332d2b6815811120a0dd70f55761258459fa1.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 51, + 213, + 286, + 317 + ], + "blocks": [ + { + "bbox": [ + 51, + 213, + 286, + 317 + ], + "lines": [ + { + "bbox": [ + 51, + 213, + 286, + 317 + ], + "spans": [ + { + "bbox": [ + 51, + 213, + 286, + 317 + ], + "type": "image", + "image_path": "d9b4d68c071bd6920dd2606bb27e3381c26fdcbde6835d65ef8145604b6d3c84.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 327, + 288, + 370 + ], + "lines": [ + { + "bbox": [ + 46, + 327, + 288, + 370 + ], + "spans": [ + { + "bbox": [ + 46, + 327, + 288, + 370 + ], + "type": "text", + "content": "Figure 4. User study results comparing RACER against Bailando and Style Gesture. Bars with different colors represent rank counts. In total, 120 comparisons have been rated. The total scores are listed on the right." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 393, + 147, + 405 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 393, + 147, + 405 + ], + "spans": [ + { + "bbox": [ + 47, + 393, + 147, + 405 + ], + "type": "text", + "content": "5.5. Ablation Studies" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 45, + 413, + 287, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 413, + 287, + 651 + ], + "spans": [ + { + "bbox": [ + 45, + 413, + 287, + 651 + ], + "type": "text", + "content": "To gain more insights into the proposed components of our method, we test some variants of RACER on Trinity dataset: (1) \"Supervised Learning\" indicates that we directly treat gesture code sequences as labels to train GPT-like models through supervised learning instead of reinforcement learning. (2) \"Distance reward\" indicates that our reward function is replaced by the Euclidean distance calculation between the predicted poses and ground truth, which is equivalent to taking the supervision signal in the traditional regression problem directly as the reward. (3) \"DQN\" means that we use the online RL algorithm DQN [34] instead of offline RL to learn the policy. The Q-value network of RACER(DQN) keeps consistent with ours. (4) \"Actor-Critic\" means that we use the online RL algorithm Actor-Critic instead of offline RL to learn the policy. For RACER(Actor-Critic), we regard the last 3 Transformer layers of GPT model with a additional linear-softmax layer as the branch of actor network. Besides, we add 3 Transformer layers as the branch of critic network to estimate the V-values." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 653, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 287, + 715 + ], + "type": "text", + "content": "The quantitative scores are shown in Tab. 2. Firstly, we explore the effectiveness of offline reinforcement learning itself. Compared to RACER, the evaluation metrics of RACER(Supervised Learning) sharply drops " + }, + { + "bbox": [ + 46, + 653, + 287, + 715 + ], + "type": "inline_equation", + "content": "5.89(11\\%)" + }, + { + "bbox": [ + 46, + 653, + 287, + 715 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 653, + 287, + 715 + ], + "type": "inline_equation", + "content": "0.60(50\\%)" + }, + { + "bbox": [ + 46, + 653, + 287, + 715 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 653, + 287, + 715 + ], + "type": "inline_equation", + "content": "3.25(24\\%)" + }, + { + "bbox": [ + 46, + 653, + 287, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 653, + 287, + 715 + ], + "type": "inline_equation", + "content": "11.04\\%" + }, + { + "bbox": [ + 46, + 653, + 287, + 715 + ], + "type": "text", + "content": ", respectively. This" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 205, + 547, + 408 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 205, + 547, + 408 + ], + "spans": [ + { + "bbox": [ + 304, + 205, + 547, + 408 + ], + "type": "text", + "content": "shows that offline RL with proposed reward function can help the model learn to generate sequences of discrete tokens better than supervised learning. Secondly, to explore the advantages of the contrastive multi-modal reward, we evaluate the performance of RACER(Distance Reward). Our method has achieved better performance, especially on the MAD and PMB matrices. This confirms that the contrastive multi-modal reward function we propose can handle the feature relationship between multiple modes and improve the matching degree of motion and speech. Moreover, to verify the advantages of offline RL over online RL, we introduce two classic online RL algorithms, DQN and Actor-Critic, to determine the variations. Under the same number of training steps, DQN and AC algorithms can only get the highest rewards of 4.23 and 3.93 respectively, which is a significant gap compared to CQL. Naturally, other matrices are also significantly worse than CQL." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 427, + 458, + 441 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 427, + 458, + 441 + ], + "spans": [ + { + "bbox": [ + 306, + 427, + 458, + 441 + ], + "type": "text", + "content": "6. Discussion and Conclusion" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 450, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 450, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 450, + 547, + 715 + ], + "type": "text", + "content": "In this paper, we propose a novel co-speech gesture synthesis framework called RACER. RACER regards the motion generated procedure as a sequential decision making process and employs reinforcement learning to learn the autoregressive gesture synthesis policy. There are mainly three sub-components of RACER. First, a vector quantized variational autoencoder (VQ-VAE) is employed to learn compact representations of gestures from massive real-world gesture data. Second, RACER uses a GPT-based architecture as policy network, which could autoregressively generate the next gestures given a sequence of speech audio. Thirdly, we propose a contrastive pre-training approach to calculate the rewards, which integrates contextual information into action evaluation and successfully captures the complex relationships between multi-modal speech-gesture data. We conduct extensive experiments to evaluate the performance of RACER. The experimental results show that RACER significantly outperforms the existing baselines in various objective and subjective metrics. Moreover, RACER shows a great generalization ability with respect to different contexts, which indicates the potentials of RACER in more complex motion synthesis tasks." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2338" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 714 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 147 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 147 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 147 + ], + "type": "text", + "content": "[1] Chaitanya Ahuja, Dong Won Lee, Ryo Ishii, and Louis-Philippe Morency. No gestures left behind: Learning relationships between spoken language and freeform gestures. In Findings of the Association for Computational Linguistics: EMNLP 2020, pages 1884–1895, Nov. 2020. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 148, + 288, + 192 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 148, + 288, + 192 + ], + "spans": [ + { + "bbox": [ + 53, + 148, + 288, + 192 + ], + "type": "text", + "content": "[2] Simon Alexanderson, Gustav Eje Henter, Taras Kucherenko, and Jonas Beskow. Style-controllable speech-driven gesture synthesis using normalising flows. Computer Graphics Forum, 39, 2020. 2, 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 194, + 288, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 194, + 288, + 237 + ], + "spans": [ + { + "bbox": [ + 53, + 194, + 288, + 237 + ], + "type": "text", + "content": "[3] Tenglong Ao, Qingzhe Gao, Yuke Lou, Baoquan Chen, and Libin Liu. Rhythmic gesticulator: Rhythm-aware cospeech gesture synthesis with hierarchical neural embeddings. ArXiv, abs/2210.01448, 2022. 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 239, + 288, + 305 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 239, + 288, + 305 + ], + "spans": [ + { + "bbox": [ + 53, + 239, + 288, + 305 + ], + "type": "text", + "content": "[4] Uttaran Bhattacharya, Elizabeth Childs, Nicholas Rewkowski, and Dinesh Manocha. Speech2affectivegestures: Synthesizing co-speech gestures with generative adversarial affective expression learning. In Proceedings of the 29th ACM International Conference on Multimedia, page 2027-2036, 2021. 2, 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 307, + 288, + 372 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 307, + 288, + 372 + ], + "spans": [ + { + "bbox": [ + 53, + 307, + 288, + 372 + ], + "type": "text", + "content": "[5] Uttaran Bhattacharya, Nicholas Rewkowski, Abhishek Banerjee, Pooja Guhan, Aniket Bera, and Dinesh Manocha. Text2gestures: A transformer-based network for generating emotive body gestures for virtual agents. In 2021 IEEE Virtual Reality and 3D User Interfaces (VR), pages 1-10, 2021. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 373, + 288, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 373, + 288, + 418 + ], + "spans": [ + { + "bbox": [ + 53, + 373, + 288, + 418 + ], + "type": "text", + "content": "[6] Chung-Cheng Chiu, Louis-Philippe Morency, and Stacy Marsella. Predicting co-verbal gestures: A deep and temporal modeling approach. In Intelligent Virtual Agents, pages 152–166, 2015. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 419, + 288, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 419, + 288, + 464 + ], + "spans": [ + { + "bbox": [ + 53, + 419, + 288, + 464 + ], + "type": "text", + "content": "[7] Hsu-Kuang Chiu, Ehsan Adeli, Borui Wang, De-An Huang, and Juan Carlos Niebles. Action-agnostic human pose forecasting. In 2019 IEEE Winter Conference on Applications of Computer Vision (WACV), pages 1423–1432, 2019. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 465, + 288, + 509 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 465, + 288, + 509 + ], + "spans": [ + { + "bbox": [ + 53, + 465, + 288, + 509 + ], + "type": "text", + "content": "[8] Ylva Ferstl and Rachel McDonnell. Investigating the use of recurrent motion modelling for speech gesture generation. In Proceedings of the 18th International Conference on Intelligent Virtual Agents, page 93–98, 2018. 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 510, + 288, + 554 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 510, + 288, + 554 + ], + "spans": [ + { + "bbox": [ + 53, + 510, + 288, + 554 + ], + "type": "text", + "content": "[9] Ylva Ferstl, Michael Neff, and Rachel McDonnell. Multi-objective adversarial gesture generation. In Proceedings of the 12th ACM SIGGRAPH Conference on Motion, Interaction and Games, 2019. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 555, + 287, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 555, + 287, + 610 + ], + "spans": [ + { + "bbox": [ + 48, + 555, + 287, + 610 + ], + "type": "text", + "content": "[10] Shiry Ginosar, Amir Bar, Gefen Kohavi, Caroline Chan, Andrew Owens, and Jitendra Malik. Learning individual styles of conversational gesture. In 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3492-3501, 2019. 1, 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 613, + 287, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 613, + 287, + 678 + ], + "spans": [ + { + "bbox": [ + 48, + 613, + 287, + 678 + ], + "type": "text", + "content": "[11] Ikhsanul Habibie, Weipeng Xu, Dushyant Mehta, Lingjie Liu, Hans-Peter Seidel, Gerard Pons-Moll, Mohamed Elgharib, and Christian Theobalt. Learning speech-driven 3d conversational gestures from video. In Proceedings of the 21st ACM International Conference on Intelligent Virtual Agents, page 101-108, 2021. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 680, + 287, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 680, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 48, + 680, + 287, + 714 + ], + "type": "text", + "content": "[12] Gustav Eje Henter, Simon Alexanderson, and Jonas Beskow. Moglow: Probabilistic and controllable motion synthesis using normalising flows. ACM Trans. Graph., 2020. 2" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 714 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 117 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 117 + ], + "type": "text", + "content": "[13] Alejandro Hernandez, Jurgen Gall, and Francesc Moreno-Noguer. Human motion prediction via spatio-temporal inpainting. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), October 2019. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 118, + 545, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 545, + 173 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 545, + 173 + ], + "type": "text", + "content": "[14] Chieh Ho, Wei-Tze Tsai, Keng-Sheng Lin, and Homer H. Chen. Extraction and alignment evaluation of motion beats for street dance. In 2013 IEEE International Conference on Acoustics, Speech and Signal Processing, pages 2429–2433, 2013. 7" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 175, + 545, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 175, + 545, + 229 + ], + "spans": [ + { + "bbox": [ + 307, + 175, + 545, + 229 + ], + "type": "text", + "content": "[15] Yu Hua, Fan Xuanzhe, Hou Yaqing, Liu Yi, Kang Cai, Zhou Dongsheng, and Zhang Qiang. Towards efficient 3d human motion prediction using deformable transformer-based adversarial network. In 2022 International Conference on Robotics and Automation (ICRA), pages 861-867, 2022. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 231, + 545, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 231, + 545, + 262 + ], + "spans": [ + { + "bbox": [ + 307, + 231, + 545, + 262 + ], + "type": "text", + "content": "[16] Lucas Kovar, Michael Gleicher, and Frédéric Pighin. Motion graphs. In ACM SIGGRAPH 2008 Classes, SIGGRAPH '08, 2008. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 264, + 545, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 264, + 545, + 319 + ], + "spans": [ + { + "bbox": [ + 307, + 264, + 545, + 319 + ], + "type": "text", + "content": "[17] Taras Kucherenko, Dai Hasegawa, Gustav Eje Henter, Naoshi Kaneko, and Hedvig Kjellström. Analyzing input and output representations for speech-driven gesture generation. In Proceedings of the 19th ACM International Conference on Intelligent Virtual Agents, page 97-104, 2019. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 321, + 545, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 321, + 545, + 387 + ], + "spans": [ + { + "bbox": [ + 307, + 321, + 545, + 387 + ], + "type": "text", + "content": "[18] Taras Kucherenko, Patrik Jonell, Sanne van Waveren, Gustav Eje Henter, Simon Alexandersson, Iolanda Leite, and Hedvig Kjellström. Gesticulator: A framework for semantically-aware speech-driven gesture generation. In Proceedings of the 2020 International Conference on Multimodal Interaction, page 242-250, 2020. 2, 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 388, + 545, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 388, + 545, + 443 + ], + "spans": [ + { + "bbox": [ + 307, + 388, + 545, + 443 + ], + "type": "text", + "content": "[19] Taras Kucherenko, Patrik Jonell, Youngwoo Yoon, Pieter Wolfert, and Gustav Eje Henter. A large, crowdsourced evaluation of gesture generation systems on common data: The genea challenge 2020. In 26th International Conference on Intelligent User Interfaces, page 11-21, 2021. 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 445, + 545, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 445, + 545, + 510 + ], + "spans": [ + { + "bbox": [ + 307, + 445, + 545, + 510 + ], + "type": "text", + "content": "[20] Aviral Kumar, Aurick Zhou, George Tucker, and Sergey Levine. Conservative q-learning for offline reinforcement learning. In H. Larochelle, M. Ranzato, R. Hadsell, M.F. Balcan, and H. Lin, editors, Advances in Neural Information Processing Systems, volume 33, pages 1179-1191. Curran Associates, Inc., 2020. 5" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 511, + 545, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 511, + 545, + 577 + ], + "spans": [ + { + "bbox": [ + 307, + 511, + 545, + 577 + ], + "type": "text", + "content": "[21] Jing Li, Di Kang, Wenjie Pei, Xuefei Zhe, Ying Zhang, Zhenyu He, and Linchao Bao. Audio2gestures: Generating diverse gestures from speech audio with conditional variational autoencoders. In 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 11273-11282, 2021. 1, 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 578, + 545, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 578, + 545, + 633 + ], + "spans": [ + { + "bbox": [ + 307, + 578, + 545, + 633 + ], + "type": "text", + "content": "[22] Ruilong Li, Shan Yang, David A. Ross, and Angjoo Kanazawa. Ai choreographer: Music conditioned 3d dance generation with aist++. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 13401-13412, October 2021. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 635, + 545, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 635, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 307, + 635, + 545, + 689 + ], + "type": "text", + "content": "[23] Yuanzhi Liang, Qianyu Feng, Linchao Zhu, Li Hu, Pan Pan, and Yi Yang. Seeg: Semantic energized co-speech gesture generation. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10463-10472, 2022. 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 691, + 545, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 691, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 307, + 691, + 545, + 714 + ], + "type": "text", + "content": "[24] Timothy P. Lillicrap, Jonathan J. Hunt, Alexander Pritzel, Nicolas Heess, Tom Erez, Yuval Tassa, David Silver, and" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2339" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 67, + 72, + 287, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 72, + 287, + 117 + ], + "spans": [ + { + "bbox": [ + 67, + 72, + 287, + 117 + ], + "type": "text", + "content": "Daan Wierstra. Continuous control with deep reinforcement learning. In 4th International Conference on Learning Representations, ICLR 2016, San Juan, Puerto Rico, May 2-4, 2016, Conference Track Proceedings, 2016. 5" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 118, + 211, + 129 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 118, + 211, + 129 + ], + "spans": [ + { + "bbox": [ + 48, + 118, + 211, + 129 + ], + "type": "text", + "content": "[25] Mixamo. https://www MIXamo.com/. 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 130, + 287, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 130, + 287, + 184 + ], + "spans": [ + { + "bbox": [ + 48, + 130, + 287, + 184 + ], + "type": "text", + "content": "[26] Shenhan Qian, Zhi Tu, Yihao Zhi, Wen Liu, and Shenghua Gao. Speech drives templates: Co-speech gesture synthesis with learned templates. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 11077-11086, October 2021. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 186, + 287, + 241 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 186, + 287, + 241 + ], + "spans": [ + { + "bbox": [ + 48, + 186, + 287, + 241 + ], + "type": "text", + "content": "[27] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision, 2021. 2, 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 243, + 286, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 243, + 286, + 274 + ], + "spans": [ + { + "bbox": [ + 48, + 243, + 286, + 274 + ], + "type": "text", + "content": "[28] Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, and Ilya Sutskever. Language Models are Unsupervised Multitask Learners. 2019. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 277, + 287, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 277, + 287, + 341 + ], + "spans": [ + { + "bbox": [ + 48, + 277, + 287, + 341 + ], + "type": "text", + "content": "[29] Li Siyao, Weijiang Yu, Tianpei Gu, Chunze Lin, Quan Wang, Chen Qian, Chen Change Loy, and Ziwei Liu. Bailando: 3d dance generation by actor-critic gpt with choreographic memory. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11050-11059, June 2022. 2, 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 343, + 287, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 343, + 287, + 408 + ], + "spans": [ + { + "bbox": [ + 48, + 343, + 287, + 408 + ], + "type": "text", + "content": "[30] Yonglong Tian, Chen Sun, Ben Poole, Dilip Krishnan, Cordelia Schmid, and Phillip Isola. What makes for good views for contrastive learning? In H. Larochelle, M. Ranzato, R. Hadsell, M.F. Balcan, and H. Lin, editors, Advances in Neural Information Processing Systems, volume 33, pages 6827-6839. Curran Associates, Inc., 2020. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 411, + 287, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 411, + 287, + 453 + ], + "spans": [ + { + "bbox": [ + 48, + 411, + 287, + 453 + ], + "type": "text", + "content": "[31] Pieter Abbeel Sergey Levine Tuomas Haarnoja, Aurick Zhou. Soft actor-critic: Off-policy maximum entropy deep reinforcement learning with a stochastic actor. ArXiv, abs/1801.01290, 2018. 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 456, + 287, + 509 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 456, + 287, + 509 + ], + "spans": [ + { + "bbox": [ + 48, + 456, + 287, + 509 + ], + "type": "text", + "content": "[32] Aaron van den Oord, Oriol Vinyals, and Koray Kavukcuoglu. Neural discrete representation learning. In Proceedings of the 31st International Conference on Neural Information Processing Systems, NIPS'17, page 6309-6318, 2017. 2, 4" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 511, + 287, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 511, + 287, + 555 + ], + "spans": [ + { + "bbox": [ + 48, + 511, + 287, + 555 + ], + "type": "text", + "content": "[33] Ruben Villegas, Jimei Yang, Duygu Ceylan, and Honglak Lee. Neural kinematic networks for unsupervised motion retargeting. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), June 2018. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 556, + 287, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 556, + 287, + 599 + ], + "spans": [ + { + "bbox": [ + 48, + 556, + 287, + 599 + ], + "type": "text", + "content": "[34] David Silver Alex Graves Ioannis Antonoglou Daan Wierstra Martin Riedmiller Volodymyr Mnih, Koray Kavukcuoglu. Playing atari with deep reinforcement learning. ArXiv, abs/1312.5602, 2013. 8" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 601, + 287, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 601, + 287, + 633 + ], + "spans": [ + { + "bbox": [ + 48, + 601, + 287, + 633 + ], + "type": "text", + "content": "[35] Petra Wagner, Zofia Malisz, and Stefan Kopp. Gesture and speech in interaction: An overview. Speech Communication, 57:209-232, 2014. 1, 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 635, + 287, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 635, + 287, + 678 + ], + "spans": [ + { + "bbox": [ + 48, + 635, + 287, + 678 + ], + "type": "text", + "content": "[36] Jason R. Wilson, Nah Young Lee, Annie Saechao, Sharon Hershenson, Matthias Scheutz, and Linda Tickle-Degnen. Hand gestures and verbal acknowledgments improve humanrobot rapport. In Social Robotics, pages 334-344, 2017. 1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 681, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 681, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 681, + 287, + 713 + ], + "type": "text", + "content": "[37] Xinchen Yan, Akash Rastogi, Ruben Villegas, Kalyan Sunkavalli, Eli Shechtman, Sunil Hadap, Ersin Yumer, and Honglak Lee. Mt-vae: Learning motion transformations to" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 227 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 327, + 73, + 545, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 545, + 105 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 545, + 105 + ], + "type": "text", + "content": "generate multimodal human dynamics. In Proceedings of the European Conference on Computer Vision (ECCV), September 2018. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 107, + 545, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 107, + 545, + 161 + ], + "spans": [ + { + "bbox": [ + 307, + 107, + 545, + 161 + ], + "type": "text", + "content": "[38] Youngwoo Yoon, Bok Cha, Joo-Haeng Lee, Minsu Jang, Jaeyeon Lee, Jaehong Kim, and Geehyuk Lee. Speech gesture generation from the trimodal context of text, audio, and speaker identity. ACM Trans. Graph., 39(6), nov 2020. 1, 2, 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 163, + 545, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 163, + 545, + 227 + ], + "spans": [ + { + "bbox": [ + 307, + 163, + 545, + 227 + ], + "type": "text", + "content": "[39] Youngwoo Yoon, Woo-Ri Ko, Minsu Jang, Jaeyeon Lee, Jae-hong Kim, and Geehyuk Lee. Robots learn social skills: End-to-end learning of co-speech gesture generation for humanoid robots. In 2019 International Conference on Robotics and Automation (ICRA), pages 4303-4309, 2019. 1" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "2340" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Co-Training 2L Submodels for Visual Recognition/51b2445a-3487-4e2d-824b-ee9b871d1274_content_list.json b/2023/Co-Training 2L Submodels for Visual Recognition/51b2445a-3487-4e2d-824b-ee9b871d1274_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..950d435715554da9a93d4cddf38cf80ef4c71f45 --- /dev/null +++ b/2023/Co-Training 2L Submodels for Visual Recognition/51b2445a-3487-4e2d-824b-ee9b871d1274_content_list.json @@ -0,0 +1,1576 @@ +[ + { + "type": "text", + "text": "Co-training $2^{L}$ Submodels for Visual Recognition", + "text_level": 1, + "bbox": [ + 236, + 128, + 733, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Hugo Touvron\\*,†", + "bbox": [ + 212, + 186, + 356, + 204 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Piotr Bojanowski*", + "bbox": [ + 212, + 205, + 359, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Matthieu Cord†", + "bbox": [ + 431, + 186, + 557, + 203 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Jakob Verbeek*", + "bbox": [ + 431, + 205, + 557, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Maxime Oquab\\*", + "bbox": [ + 627, + 186, + 759, + 204 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Hervé Jégou*", + "bbox": [ + 638, + 205, + 750, + 223 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "*Meta AI / FAIR Paris", + "bbox": [ + 276, + 224, + 455, + 242 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "†Sorbonne University", + "bbox": [ + 524, + 224, + 699, + 244 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 268, + 313, + 284 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We introduce submodel co-training, a regularization method related to co-training, self-distillation and stochastic depth. Given a neural network to be trained, for each sample we implicitly instantiate two altered networks, \"submodels\", with stochastic depth: we activate only a subset of the layers. Each network serves as a soft teacher to the other, by providing a loss that complements the regular loss provided by the one-hot label. Our approach, dubbed \"cosub\", uses a single set of weights, and does not involve a pre-trained external model or temporal averaging.", + "bbox": [ + 75, + 300, + 470, + 450 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Experimentally, we show that submodel co-training is effective to train backbones for recognition tasks such as image classification and semantic segmentation. Our approach is compatible with multiple architectures, including RegNet, ViT, PiT, XCiT, Swin and ConvNext. Our training strategy improves their results in comparable settings. For instance, a ViT-B pretrained with cosub on ImageNet-21k obtains $87.4\\%$ top-1 acc. @448 on ImageNet-val.", + "bbox": [ + 75, + 450, + 468, + 571 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 588, + 209, + 603 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Although the fundamental ideas of deep trainable neural networks have been around for decades, only recently have barriers been removed to allow breakthroughs in successfully training deep neural architectures in practice. Many of these barriers are related to non-convex optimization in one way or another, which is central to the success of modern neural networks. The optimization challenges have been addressed from multiple angles in the literature. First, modern architectures are designed to facilitate the optimization of very deep networks. An exceptionally successful design principle is using residual connections [24, 25]. Although this does not change the expressiveness of the functions that the network can implement, the improved gradient flow alleviates, to some extent, the difficulties of optimizing very deep networks. Another key element to the optimization is the importance of data, revealed by the step-change in visual recognition performance resulting from the ImageNet dataset [11], and the popularization of transfer learning with pre-training on large datasets [39, 58].", + "bbox": [ + 75, + 613, + 470, + 902 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/6605072b50b4cf03d4e74edaf5c0b4bb06c32e1a462ecca593c4342f23471d3c.jpg", + "image_caption": [ + "Figure 1. Co-training of submodels (cosub): for each image, two sub-models are sampled by randomly dropping layers of the full model. The training signal for each submodel mixes the cross-entropy loss from the image label with a self-distillation loss obtained from the other submodel." + ], + "image_footnote": [], + "bbox": [ + 588, + 263, + 807, + 529 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "However, even when (pre-)trained with millions of images, recent deep networks with millions if not billions of parameters, are still heavily overparameterized. Traditional regularization like weight decay, dropout [46], or label smoothing [47] are limited in their ability to address this issue. Data-augmentation strategies, including those mixing different images like Mixup [61] and CutMix [60], have proven to provide a complementary data-driven form of regularization. More recently, multiple works propose to resort to self-supervised pre-training. These approaches rely on a proxy objective that generally provides more supervision signal than the one available from labels, like in recent (masked) auto-encoders [5,16,22], which were popular in the early deep learning literature [7,19,27]. Similarly, contrastive approaches [23] or self-distillation [9] provide a richer supervision less prone to supervision collapse [12]. Overall, self-supervised learning makes it possible to learn larger models with less data, possibly reducing the need of a pre-training stage [15].", + "bbox": [ + 496, + 613, + 893, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "11701", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Distillation is a complementary approach to improve optimization. Distillation techniques were originally developed to transfer knowledge from a teacher model to a student model [4, 28], allowing the student to improve over learning from the data directly. In contrast to traditional distillation, co-distillation does not require pre-training a (strong) teacher. Instead, a pool of models supervise each other. Practically, it faces several limitations, including the difficulty of jointly training more than two students for complexity reasons, as it involves duplicating the weights.", + "bbox": [ + 75, + 90, + 472, + 243 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we propose a practical way to enable cotraining for a very large number of students. We consider a single target model to be trained, and we instantiate two submodels on-the-fly, simply by layerwise dropout [20,31]. This gives us two neural networks through which we can backpropagate to the shared parameters of the target model. In addition to the regular training loss, each submodel serves as a teacher to the other, which provides an additional supervision signal ensuring the consistency across the submodels. Our approach is illustrated in Figure 1: the parameter $\\lambda$ controls the importance of the co-training loss compared to the label loss, and our experiments show that it significantly increases the final model accuracy.", + "bbox": [ + 75, + 248, + 472, + 445 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "This co-training across different submodels, which we refer to as cosub, can be regarded as a massive co-training between $2^{L}$ models that share a common set of parameters, where $L$ is the number of layers in the target architecture. The target model can be interpreted as the expectation of all models. With a layer drop-rate set to 0.5, for instance for a ViT-H model, all submodels are equiprobable, and then it amounts to averaging the weights of $2^{2 \\times 32}$ models.", + "bbox": [ + 75, + 449, + 472, + 570 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our contributions can be summarized as follows:", + "bbox": [ + 76, + 575, + 403, + 589 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We introduce a novel training approach for deep neural networks: we co-train submodels. This significantly improves the training of most models, establishing the new state of the art in multiple cases. For instance, after pre-training ViT-B onImagenet-21k and fine-tuning it at resolution 448, we obtain $87.4\\%$ top-1 accuracy onImagenet-val.", + "- We provide an efficient implementation to subsample models on the fly. It is a simple yet effective variation of stochastic depth [31] to drop residual blocks.", + "- We provide multiple analyses and ablations. Noticeably, we show that our submodels are effective models by themselves even with significant trimming, similar to LayerDrop [20] in natural language processing.", + "- We validate our approach on multiple architectures (like ViT, ResNet, RegNet, PiT, XCiT, Swin, ConvNext), both for image classification -trained from scratch or with transfer-, and semantic segmentation.", + "- We will share models/code for reproducibility in the DeiT repository." + ], + "bbox": [ + 94, + 598, + 467, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related work", + "text_level": 1, + "bbox": [ + 500, + 89, + 637, + 104 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Knowledge distillation. Originally, distillation was introduced as a way to train a model such that it reproduces the performance of another model [4, 28]. The typical use-case is to improve the quality of a relatively small model by leveraging a strong teacher, whose complexity may be prohibitive for a practical deployment. The teacher's soft labels have a similar effect as label smoothing [59]. As shown by Wei et al. [53], the teacher's supervision takes into account the effects of the data augmentation, which sometimes causes a misalignment between the real label and the image. Knowledge distillation can transfer inductive biases [1] in a soft way in a student model when using a teacher model these biases are enforced in a hard way. Touvron et al. [50] proposed a variant of distillation adapted to Vision Transformer (ViT), showing the effectiveness of using a ConvNet teacher for a ViT student.", + "bbox": [ + 496, + 119, + 890, + 362 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Mixture models. Ensembling has a long history that we can trace back to the origins of statistics [37] and the possibility of improving the precision of measurements with multiple observations. In machine learning, bagging [8] combines multiple weak classifiers to produce a strong one. This idea is naturally extended to neural networks, where it can offer improved stability or accuracy, or other properties like anytime inference [44]. A mixture model can also be seen as a larger model with an internal parallel structure.", + "bbox": [ + 496, + 368, + 890, + 503 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Co-distillation. At the interface of mixture models and distillation, the concept of co-distillation [62, 64] does not require a prior teacher: the mixture serves as the teacher to all the networks in the mixture. The models are jointly optimized, which leads to improved accuracy of the individual models [2]. This can be regarded as a form of collaborative learning between the different elements of the mixture [45]. Compared to training only one model, co-distillation involves training two or more models, each requiring storage for weights and activation, and computing backward and forward passes.", + "bbox": [ + 496, + 510, + 890, + 676 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Exponential moving average teacher. A special form of teacher is a model obtained from the student model itself. Tarvainen and Valpila [49] show that such a model obtained by temporal averaging of intermediate checkpoints during the training is an effective teacher, that can be obtained with an exponential moving average (EMA). This idea has also been adopted in self-supervised training with learning schemes like DINO [9]. This method involves storing a copy of the weights corresponding to the temporal averaging carried out by the EMA model.", + "bbox": [ + 496, + 683, + 890, + 834 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Dropout and model populations. Dropout [46] is an effective way to regularize models. With residual architectures, one effective way to train deeper architecture is stochastic depth [31]. It reduces the size of a network at", + "bbox": [ + 496, + 839, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "11702", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/827e24b5c3a9e8f3960ee8f180d6c08bfd939a86fcf599f96570ddf5a723151d.jpg", + "image_caption": [ + "(a) supervised" + ], + "image_footnote": [], + "bbox": [ + 127, + 99, + 205, + 194 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/0577d9e58a051c00e478e494123d045453ed591c00b97aa9463d301198da72c3.jpg", + "image_caption": [ + "(b) knowledge distillation" + ], + "image_footnote": [], + "bbox": [ + 253, + 99, + 367, + 194 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/52415c6721f74905e1f9e2aa415e28ad94c2ee870d6e251fa66f3018e5d87f32.jpg", + "image_caption": [ + "(c) self-distillation" + ], + "image_footnote": [], + "bbox": [ + 411, + 99, + 529, + 195 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/760887e7f56f7476c1a24653682b1296bc185273378ec01af654b59712b33505.jpg", + "image_caption": [ + "(d) co-training", + "Figure 2. Brief summary of related works and of submodel co-training (cosub). (a) supervised baseline: the supervision is provided as a one-hot label. (b) Knowledge distillation [28]: a teacher provides a (soft) target. A simple yet effective variant [50] combines the predicted label of the teacher with the image label. KD requires a pre-existing teacher that must be stored for inference. (c) In self-distillation, the teacher is obtained from the model itself, typically by model averaging. (d) Co-training involves two distinct models that serves as teacher to each other. (e) Submodel co-training generates on-the-fly two distinct models (amongst $2^{L}$ ) for each sample, which serves a teacher to each other. We only need to store a single set of weights for the model and optimizer because the random selections $\\theta_{1} = \\mathcal{T}(\\theta)$ (resp., $\\theta_{2}$ ) of submodels allow us to back-propagate on $\\theta$ ." + ], + "image_footnote": [], + "bbox": [ + 557, + 95, + 694, + 196 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/d966bd13264237f7aecd65717121e981cfc89e5e8bfa8b9d0122c9a9f8cebd80.jpg", + "image_caption": [ + "(e) submodel co-training" + ], + "image_footnote": [], + "bbox": [ + 702, + 94, + 887, + 202 + ], + "page_idx": 2 + }, + { + "type": "table", + "img_path": "images/8fda16f12e1ae55dc6aaba2940166446304a7a36da955f09397b95af7e9cd754.jpg", + "table_caption": [], + "table_footnote": [ + "\\*backward pass assumed $2\\times$ the complexity of forward $\\dagger$ requires a trained teacher model" + ], + "table_body": "
methodmodel paramscompute*
weightsoptimizer
(a) supervised×1×1×1
(b) KD†×2×1×1.33
(c) mean teacher×2×1×1.33
(d) co-training×2×2×2
(e) Cosub (ours)×1×1×2
", + "bbox": [ + 596, + 233, + 890, + 325 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "training time by dropping residual blocks during training. The initial goal of this approach was to improve the training of deep network. However, there are other implications: in natural language processing, Fan et al. show that the counterpart of stochastic depth, namely LayerDrop [20], is effective to reduce the transformer depth on demand. Interestingly, submodels extracted from a model trained with LayerDrop are stronger than those trained with the full target depth with a proper selection strategy for layers.", + "bbox": [ + 75, + 369, + 468, + 505 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In our paper, we regard stochastic depth as a regularization technique, but at the same time we adopt the point of view of LayerDrop, i.e., an effective way to train a population of submodels that share parameters, or entire layers in our case. From that viewpoint, stochastic depth amounts to training $2^{L}$ distinct models. This is different from population training as involved in network space design [41], whose goal is related to network architecture search [18], which aims at optimizing the architecture itself.", + "bbox": [ + 75, + 507, + 468, + 641 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "More recently, model soups [56] are models obtained by averaging the weights of multiple models finetuned with different hyperparameter configurations. The authors show that it often improves accuracy and robustness. We point out that stochastic depth can be regarded as a special form of model soup over the entire population of submodels that we can instantiate with stochastic depth. We further discuss this relationship in the next section.", + "bbox": [ + 75, + 643, + 468, + 763 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Submodel co-training", + "text_level": 1, + "bbox": [ + 76, + 777, + 284, + 795 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we present our cosub approach, for constraining submodels. We consider a neural network $f_{\\theta}$ parameterized by learnable parameters $\\theta$ .", + "bbox": [ + 75, + 804, + 468, + 849 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Submodel instantiation: model augmentation operator. We first define a model augmentation operator $\\mathcal{T}$ . For a given neural network, it provides a set of parameters $\\theta^{\\prime} =$", + "bbox": [ + 75, + 854, + 468, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "$\\mathcal{T}(\\theta, R)$ that allow us to define variations $f_{\\theta'}$ of the function $f_{\\theta}$ by drawing a random variable $R$ . The model augmentation is such that $f_{\\theta}$ and $f_{\\theta'}$ share parameters, hence any update on $\\theta'$ modifies $\\theta$ and therefore $f_{\\theta}$ accordingly. A simple way to define $\\mathcal{T}$ is to replace some parameters by zeros, which corresponds to dropout [46]. In this paper, we focus on stochastic depth, which has interesting connections with model averaging [56].", + "bbox": [ + 496, + 369, + 890, + 491 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Overview. For a given training sample $x$ within a batch, the training is as follows (see also Figure 1):", + "bbox": [ + 496, + 496, + 890, + 526 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. We first data-augment the image, producing $\\hat{x}$ .", + "2. The image is duplicated with the batch, effectively doubling the batch size, which hence contains two identical copies of each augmented image1.", + "3. We determine the stochastic depth pattern for each sample according to the target drop-rate $\\tau$ , which amounts to producing two functions $f_{\\theta_1}$ and $f_{\\theta_2}$ . See Section 4 for the details of this procedure.", + "4. The forward pass proceeds as usual: we compute the soft output labels $y_{1} = f_{\\theta_{1}}(\\hat{x})$ and $y_{2} = f_{\\theta_{2}}(\\hat{x})$ .", + "5. We compute the losses and the backwards pass accordingly. Note that the gradients on $\\theta_{1}$ and $\\theta_{2}$ are directly used to update $\\theta$ , as they are just subsets of $\\theta$ ." + ], + "bbox": [ + 511, + 529, + 890, + 724 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Loss. Each submodel is trained using a weighted average of (i) the standard binary cross-entropy loss obtained from the image label $y$ , and (ii) a binary cross-entropy loss w.r.t. the soft-labels computed for the the same image by the other submodel. The respective weight of the standard binary cross-entropy loss $\\mathcal{L}_{\\mathrm{label}}$ versus the cosub loss $\\mathcal{L}_{\\mathrm{cosub}}$ is controlled by the hyper-parameter $\\lambda$ , as", + "bbox": [ + 496, + 732, + 890, + 838 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {t o t}} = \\lambda \\mathcal {L} _ {\\text {l a b e l}} + (1 - \\lambda) \\mathcal {L} _ {\\text {c o s u b}}. \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 580, + 849, + 890, + 864 + ], + "page_idx": 2 + }, + { + "type": "page_footnote", + "text": "Different data-augmentations did not provide any advantage in our preliminary experiments. We simply duplicate the batch for simplicity.", + "bbox": [ + 500, + 875, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "11703", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In details, the loss writes as", + "bbox": [ + 76, + 90, + 263, + 104 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} _ {\\mathrm {t o t}} = \\lambda \\left(\\frac {\\mathcal {L} (y _ {1} , y) + \\mathcal {L} (y _ {2} , y)}{2}\\right) \\\\ + (1 - \\lambda) \\left(\\frac {\\mathcal {L} \\left(y _ {1} , \\mathrm {s g} \\left(y _ {2}\\right)\\right) + \\mathcal {L} \\left(y _ {2} , \\mathrm {s g} \\left(y _ {1}\\right)\\right)}{2}\\right), \\quad (2) \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 93, + 113, + 468, + 185 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathcal{L}(y, y')$ is either the binary cross-entropy (BCE) or a cross-entropy (CE) loss. Importantly, when applying this loss, we do not back-propagate on the second term $y'$ . This is indicated by the stop-gradient operator $\\mathrm{sg}(\\cdot)$ .", + "bbox": [ + 76, + 193, + 468, + 255 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Discussion. The submodel instantiation provides a model that is, by itself, a valid neural network model. Co-training submodels is a way to enforce that all such submodels produce a consistent output. Amongst these submodels, a very special case is when we retain all residual blocks, which is the model that we primarily intend to use at inference time. Note that Fan et al. [20] show, in an NLP context, that submodels extracted from a deeper model are superior, for a given target depth. With proper rescaling of the residual branches, this specific submodel with all blocks activated can be seen as the expectation over all models. This can be shown as follows: let us consider an extra scalar parameter $s_l$ associated with a given residual block $r_l(x)$ , which we use as a multiplicative factor on output of each residual: $s_l = 1$ if the residual block $r_l$ is included in the submodel, $s_l = 0$ otherwise. Each submodel is fully parameterized by a binary vector $s = (s_1, \\dots, s_L)$ , where $L$ is the total number of layers. Therefore all submodels have the same parameters $\\theta$ , and only differ by $s$ indicating the zeroed residual blocks. The weight expectation $\\bar{\\theta}$ is hence", + "bbox": [ + 76, + 258, + 472, + 561 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\bar {\\theta} = \\mathbb {E} _ {s \\in [ 0, 1 ] ^ {L}} [ \\theta ] = \\theta , \\\\ \\bar {s} = \\mathbb {E} _ {s \\in [ 0, 1 ] ^ {L}} [ s ] = [ 1 - \\tau , \\dots , 1 - \\tau ] ^ {\\top}, \\tag {3} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 138, + 571, + 468, + 611 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\bar{s}$ is the scaling factor used in stochastic depth. Under a uniform distribution of submodels that is obtained when the stochastic drop rate $\\tau = 0.5$ , the inference-time model is exactly the average of all $2^{L}$ models. While stochastic depth is often regarded as a regularization technique, this averaging interpretation relates it to the recent model soup [56].", + "bbox": [ + 76, + 619, + 468, + 712 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Efficient stochastic depth", + "text_level": 1, + "bbox": [ + 76, + 724, + 313, + 742 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We revisit the original stochastic depth [31] formulation in order enable an efficient implementation, which we will release for PyTorch [40]. Instantiating a submodel for a sample amounts to selecting a subset of residual layers of the model and performing training on these. This can be implemented using the stochastic depth (SD) approach, whose objective was initially to improve the training of very deep networks. In stochastic depth, for each sample and each layers of the network, we select whether the layers will be dropped or not with a certain probability $\\tau$ .", + "bbox": [ + 76, + 750, + 470, + 902 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/9707da9db87eb1046491989ddb46f98c5c976595da4accbd49ee490813dd094b.jpg", + "image_caption": [ + "Figure 3. Efficient implementation of stochastic depth, using the permute-select approach. In this example, we drop the residuals for samples 3 and 6, corresponding to a drop rate $\\tau = 0.25$ . For a drop rate higher than 0.1, the overhead of the approach is negligible in terms of memory and compute." + ], + "image_footnote": [], + "bbox": [ + 498, + 90, + 893, + 202 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In practice, e.g. in the timm library [54], stochastic depth is implemented by masking with zeros the residuals added for a given sample of a batch. This is not efficient: this naive approach performs the computations for a residual then throws it away, wasting computation.", + "bbox": [ + 496, + 276, + 890, + 351 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our efficient stochastic depth (ESD) approach addresses this issue, saving both memory and compute, and is illustrated in Figure 3. For each layer, given a batch size $B$ and a drop rate $\\tau$ , we apply the layer to $B_{\\mathrm{keep}} = \\mathrm{round}(B \\times (1 - \\tau))$ samples in the batch. In contrast to the original version of SD, our efficient version drops a fixed number of samples at each layer, where $d$ is adjusted such that $B_{\\mathrm{keep}}$ is an integer. In our experiments, this did not have an effect on the final accuracy of the models. Our efficient implementation proceeds as follows: (i) we apply a random permutation of the $B$ samples, then (ii) we select the first $B_{\\mathrm{keep}}$ samples. We then compute the residual function for the selected subset, then add the result onto the full batch using the built-in index_add function and scaling the result by $\\frac{1}{1 - \\tau}$ to have the correct scaling as in the original SD formulation.", + "bbox": [ + 496, + 352, + 892, + 579 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Discussion: progressive vs uniform rate. In the original paper [31], stochastic depth drops a layer with a probability that is linearly increasing during the forward pass: layers closer to the output has a higher chance to be dropped. However this strategy is limited with high drop rates, as later layers are excessively dropped. Touvron et al. [52] adopt a uniform rate drop per layer. It is as effective as the progressive rate with vision transformers, but it makes it possible to target a drop-rate greater than 0.5 on average. For this reason we adopt an uniform drop rate everywhere.", + "bbox": [ + 496, + 583, + 890, + 733 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our efficient stochastic depth implementation variant works both with uniform and progressive drop-rate. In the case of progressive, the effective batch size is decreasing during to the forward pass. One limitation is that our technique implies a quantization of the drop-rate, which can be problematic for small batch sizes. Therefore one must take care to properly verify that the drop-rate is not too severely quantized, in particular in the progressive case for which the drop-rate can be very high in the last layers. In the supplemental material (Appendix C), we discuss with more details the drop-rate quantization effect.", + "bbox": [ + 496, + 734, + 893, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "11704", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/f52e69c2b34f9bd6846bc6b0ef914df1a935f850c1a0c262ab10a20143d0e470.jpg", + "image_caption": [ + "Figure 4. Cosub as population training: each submodel extracted by a transformation $\\mathcal{T}$ is a valid neural network. Our cosub strategy can hence be regarded as co-training a large number of subnetworks. We plot the accuracy of the submodels as a function of the number of layers that we preserve. We drop layers with probability $\\tau = 0.2$ for ViT-B and $\\tau = 0.45$ for ViT-L. On average our method provides a significant boost in performance for the whole population of submodels extracted from the main model." + ], + "image_footnote": [], + "bbox": [ + 76, + 85, + 467, + 244 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/3f58449cc66ba136867a1ad66050621265394342767c650048bb84a81d54237a.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelViT-SViT-MViT-BViT-LViT-H
τ(0.05)(0.1)(0.2)(0.45)(0.6)
baseline81.482.583.784.584.9
baseline + cosub81.582.883.984.985.5
", + "bbox": [ + 84, + 352, + 467, + 409 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 76, + 510, + 209, + 527 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We evaluate our approach on residual architectures [24, 25], since they are readily compatible with the stochastic dropout. We take as our main reference the vanilla Vision Transformer introduced by Dosovitskiy et al. [14]. However, as shown in our experiments, our approach is effective with all the residual architectures that we have considered.", + "bbox": [ + 75, + 537, + 468, + 628 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.1. Baseline and training settings", + "text_level": 1, + "bbox": [ + 76, + 645, + 339, + 662 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We adopt the state-of-the-art training procedure proposed in the DeiT III paper [51] as baseline for transformers architectures and the training procedure from Wightman et al. [55] for the convnets. All hyper-parameters are identical except onImagenet-21k, where the hyper-parameter $\\tau$ are adjusted depending on the training setting. We recapitulate the training hyper-parameters in Appendix A.", + "bbox": [ + 75, + 671, + 468, + 777 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We additionally adopt and measure the impact of LayerDecay for the fine-tuning when transferring from Imagenet-21k to Imagenet-1k. This method slightly boosts the performance, as discussed later in this section, and in Table 10. This method was adopted in multiple recent papers and in particular for fine-tuning of self-supervised approaches [5, 22], yet the contribution of this fine-tuning ingredient was not quantitatively measured.", + "bbox": [ + 75, + 779, + 470, + 902 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/2761f414f2b6cb09fe404b33e487ee0287e7becad429b48a223bebc0a2cdbf21.jpg", + "table_caption": [ + "Table 1. Comparison of top-1 accuracy with ViT architecture trained with/without cosub at resolution 224 (800 epochs) onImagenet-1k. The improvement is not significant for the smaller architecture ViT-S. Cosub is gradually more effective when we increase the model size and the SD rate." + ], + "table_footnote": [], + "table_body": "
ModelτOriginalBaseline+cosubΔ
Transformers
ViT-L [14]0.4576.584.584.9+0.4
CaiT-L24 [52]0.45-83.884.4+0.6
PiT-B [26]0.2582.083.884.1+0.3
XCiT-L12 [17]0.20-82.683.0+0.4
Swin-B [34]0.2083.582.983.3+0.4
Swin-L [34]0.45-80.884.0+3.2
Convnets
ResNet-50 [24]0.1076.280.280.3+0.1
ResNet-101 [24]0.2077.481.882.1+0.4
ResNet-152 [24]0.3078.382.483.1+0.7
RegNet-16GF [42]0.3080.482.983.8+0.9
", + "bbox": [ + 504, + 88, + 890, + 263 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/7361f96e908cc41e563cafcb4f06fe4616c70c7dc8a08eff7bc9ea1de1078513.jpg", + "table_caption": [ + "Table 2. Benefit of cosub for different architectures trained from scratch on Imagenet-1k at resolution 224. We report top-1 acc. for the supervised baseline and cosub, as well as results reported in the corresponding papers when available (trained with different settings). We have adjusted the stochastic-depth drop-rate (SD) hyper-parameter for each architecture." + ], + "table_footnote": [], + "table_body": "
LossBCE-softBCE-hardCE-hard
Imagenet-val top-1 accuracy83.583.581.9
", + "bbox": [ + 503, + 345, + 890, + 383 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Table 3. Ablation on the loss for cosub with ViT-H trained at resolution $126\\times 126$ on Imagenet-1k during 800 epochs. The training is inherited from DeiT-III, which also uses BCE when training with Imagenet-1k only.", + "bbox": [ + 500, + 385, + 890, + 422 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.2. Empirical analysis of cosub", + "text_level": 1, + "bbox": [ + 500, + 452, + 746, + 468 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We perform various ablations on Imagenet-1k to analyse the impact of our training method on the learned networks.", + "bbox": [ + 498, + 477, + 890, + 508 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Performance for different model sizes and architectures. Table 1 provides the results obtained by the baseline and cosub when we vary the model size of vision transformers. The stochastic depth coefficient was optimized for the baseline and we keep it unchanged with cosub. As to be expected, our method is almost neutral for small models like ViT-S: $+0.1\\%$ top-1 accuracy, which is about the standard deviation of measurements. The improvement is increasingly important for larger models, up to a significant improvement of $+0.6\\%$ top-1 accuracy for ViT-H models.", + "bbox": [ + 496, + 513, + 890, + 665 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In Table 2, we show that our approach is beneficial with all architectures that we have tried. We report the results of the original paper, evaluate the performance with our baseline training, and measure the improvement brought by cosub. For almost all architectures, we observe a significant boost in performance. The exception is the ResNet-50, for which cosub improves the top-1 accuracy by only $+0.1\\%$ , similar to our observation with ViT-S. In Table B.1 in the appendix we present improved results obtained for multiple architectures pre-trained with cosub onImagenet-21k.", + "bbox": [ + 496, + 667, + 892, + 819 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Analysis of submodel performance. With cosub, we sample different subnetworks during training to perform the constraining. We analyse the impact of cosub on the accuracy of the sub networks themselves. In Figure 4 we consider the accuracy of submodels of different size of ViT-B and ViT-L", + "bbox": [ + 496, + 825, + 893, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "11705", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/59e659e7c74bc8c7438b99554efc213012e7414fb8b093ab2df8a5cd590f87a3.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodViT-LViT-H
(a)supervised baseline [51]84.584.9
(b)KD85.385.3
(c)mean teacher84.483.4
(d)co-training82.683.1
(e)cosub84.985.5
(b) + (e)KD + cosub85.385.7
", + "bbox": [ + 83, + 88, + 460, + 186 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/5df5b7a5ef28a32c10f0dbb652a1ec1cf3969c003ad89652f743207b373ea06a.jpg", + "table_caption": [ + "Table 4. Training strategies with distillation. We compare on Imagenet-1k at resolution $224 \\times 224$ different approaches involving co- or self-training with distillation. KD, mean teacher and co-training use the same $\\lambda = 0.5$ and same hyper-parameters as in Deit-III. Unlike cosub, the mean teacher (c) requires other hyper-parameters for EMA. We perform a small grid-search to adjust this parameters. Note (last row): our approach is complementary with KD, assuming a pre-trained teacher is available beforehand." + ], + "table_footnote": [], + "table_body": "
λ0.10.30.50.70.91.0
top1 accuracy79.0583.2783.5583.2083.0482.91
", + "bbox": [ + 81, + 291, + 468, + 323 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "models. Cosub improves the accuracy of the whole population of sub-networks and, in particular, the target network.", + "bbox": [ + 76, + 407, + 467, + 438 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Loss formulation. In Table 3 we experiment with different losses for cosub. With Imagenet1k, DeiT-III training uses BCE instead of CE for the main loss. With cosub, BCE is more compatible with the loss of the baseline and, as to be expected, we also observe a better performance with BCE. We have done ablations using hard and soft targets for the cosub loss. The results are similar, therefore by default we keep soft-targets for the cosub loss.", + "bbox": [ + 75, + 445, + 467, + 565 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Alternative teacher/student. In Table 4 we report the results obtained with the different distillation or co-training approaches depicted in Figure 2. Other approaches are not effective off-the-shelf, except KD that requires a pre-trained teacher. Our approach is on par with KD (lower for ViT-L, better or ViT-H), and in the last row we show that it even provides a slight boost to combine KD with cosub.", + "bbox": [ + 75, + 574, + 467, + 679 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Impact of the cosub loss. The hyper-parameter $\\lambda$ controls the tradeoff between the co-distillation loss and the cross-entropy classification loss. Setting $\\lambda = 1$ means that we have a regular supervised training setting, except that (i) we double the batch size by duplicating the image after data augmentation, and (ii) stochastic depth selects different layers for each image copy.", + "bbox": [ + 75, + 686, + 467, + 792 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In Table 5, we measure the impact of $\\lambda$ , with all other hyper-parameters being fixed, for ViT-H trained at resolution $126 \\times 126$ onImagenet-1k. We observe that the best ratio is to use an equal weighting of the cosub loss and the classic training loss. Using the cosub loss increases the performance by $0.6\\%$ Top-1 accuracy onImagenet-val, which is the typical improvement that we observe for large models.", + "bbox": [ + 75, + 794, + 467, + 900 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/058c06a8f9ff92a83df772fbadcc0db615c759b10e3f72538502427652e5ff61.jpg", + "table_caption": [ + "Table 5. Ablation of the parameter $\\lambda$ controlling the weight of the co-distillation loss across submodels (Imagenet1k-val, top1-acc). Model ViT-H trained at resolution ${126} \\times {126}$ onImagenet-1k during 800 epochs. $\\lambda =$ 1.0 corresponds a supervised baseline and $\\lambda = {0.5}$ corresponds to cosub." + ], + "table_footnote": [], + "table_body": "
ModelτepochsBaseline+cosub
valv2valv2
CaiT-M120.2040083.272.983.773.5
80082.972.683.673.1
PiT-B0.2540083.873.684.174.1
80082.471.983.172.8
ViT-B0.2040083.172.683.273.1
80083.773.183.973.5
120083.372.8--
160083.373.4--
ViT-H0.6040084.875.385.075.8
80084.975.685.576.3
", + "bbox": [ + 509, + 88, + 883, + 258 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/01139790c47b71d89479342505e12a79e81b3b229187ace569cf5aece4364efb.jpg", + "table_caption": [ + "Table 6. We compare ViT models trained with and without cosub on ImageNet-1k only with different number of epochs at resolution $224 \\times 224$ . One can see that cosub is more effective for larger models yielding higher values of the SD hyper-parameter $\\tau$ . It avoids the early saturation or overfitting of the performance that we typically observe with the baseline when we increase the training time without re-adjusting hyper-parameters. See also Table 5 for a direct comparison with and without the co-distillation loss, and Table 7 for the corresponding training times per epoch." + ], + "table_footnote": [], + "table_body": "
Training methodmodelGPUs usedMemory peak (GB)Time (min) by epoch
DeiT-IIIViT-L3221.48
ViT-H6427.612
DeiT-III + ESDViT-L3215.19
ViT-H6415.211
cosub (with ESD)ViT-L3226.916
ViT-H6425.017
", + "bbox": [ + 504, + 375, + 887, + 491 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 7. Training times of different models trained at resolution $224\\times 224$ with batch size 2048 on Imagenet-1k with DeiT-III and our approach. cosub uses our efficient stochastic depth (ESD), which amortizes the extra memory needed by cosub, especially for the largest models with high stochastic depth values (0.45 for ViT-L, and 0.6 for ViT-H). Timings are indicative and not representative of an optimized selection of the batch size, see Appendix C for measurements with adjusted parameters.", + "bbox": [ + 498, + 493, + 890, + 578 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Number of training epochs. In Table 6 we compare results on Imagenet-1k-val and Imagenet-v2 different for architectures trained with and without cosub on Imagenet-1k only at resolution $224 \\times 224$ with different number of epochs. We observe less overfitting with cosub and longer training schedule. In particular, with bigger architecture like ViT-H, we observe continuous improvement with a longer schedule where the baseline saturates.", + "bbox": [ + 498, + 599, + 890, + 720 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Training time. In Table 7 we compare the training costs of cosub and DeiT-III. Thanks to our efficient stochastic depth formulation we maintain a similar memory peak during training. For bigger architectures the gap in training speed between cosub and the baseline is decreasing.", + "bbox": [ + 498, + 728, + 890, + 803 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Resolution. In Table 8 we compare different ViT architectures trained with and without cosub at different resolutions onImagenet-1k. We fine-tune during 20 epochs at resolution $224 \\times 224$ before evaluation at this resolution. We observe that cosub gives significant improvements across the different resolutions and models.", + "bbox": [ + 498, + 809, + 890, + 898 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "11706", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/c99963d3bf244e97afce1eedc22e7bb744f63378ca0f4cf4b61c276a1164e907.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelResolutionDeit-III Baseline+cosub
valv2valv2
ViT-B128 × 12883.573.483.874.0
192 × 19283.873.684.174.0
224 × 22483.773.183.973.5
ViT-L128 × 12884.574.785.175.5
192 × 19284.975.185.275.7
224 × 22484.575.084.975.6
ViT-H126 × 12685.175.685.676.4
182 × 18285.175.985.776.6
224 × 22484.975.685.576.3
", + "bbox": [ + 84, + 88, + 460, + 244 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/7617323a07ba4969a56535b7dde3e90c81beb0748e68d020de1363082af727b0.jpg", + "table_caption": [ + "Table 8. Imagenet-1k val and v2 top-1 accuracy of ViT models trained with and without cosub for 800 epochs on Imagenet-1k at different resolutions, followed by finetuning for 20 epochs at resolution $224 \\times 224$ ." + ], + "table_footnote": [], + "table_body": "
resol. →112224336448112224336448
↓ modelImagenet-valImagenet-v2
ViT-S78.083.184.685.266.673.775.176.3
ViT-M80.685.086.086.369.676.076.877.2
ViT-B82.886.386.987.472.177.077.978.3
ViT-L85.487.588.188.375.779.179.880.0
ViT-H86.288.0--76.979.6--
ViT-g86.5---77.3---
", + "bbox": [ + 81, + 294, + 468, + 409 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/63ad22a6f9a7ee99ef5b228e82c68526660b1606c0f15405a60186dcd628864e.jpg", + "table_caption": [ + "Table 9. Performance of models at different resolutions. We report the results obtained on Imagenet-val by models of different sizes pre-trained with cosub on Imagenet-21k and fine-tuned on Imagenet-1k. Training schedule: 270 epochs except ViT-g (90 epochs). Except for ViT-S, the results at resolution 336 and 448 were pre-trained on Imagenet-21k at resolution 224 for efficiency reasons. We have not fine-tuned ViT-H and ViT-g at large resolutions since these models are computationally expensive." + ], + "table_footnote": [], + "table_body": "
MethodLong TrainingLayer DecayViT-SModel ViT-BViT-L
baselineXX82.685.286.8
XX82.585.887.4
cosubX✓82.786.087.5
✓X82.886.087.4
✓✓83.186.387.5
", + "bbox": [ + 83, + 512, + 460, + 613 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Imagenet-21k impact of layer-decay. In Table 10 we compare different number of epochs for the pre-training on Imagenet-21k and the finetuning on Imagenet-1k with and without layer-decay. We observe that both layer-decay and long training bring improvements with cosub.", + "bbox": [ + 75, + 681, + 468, + 758 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.3. Comparisons with the state of the art", + "text_level": 1, + "bbox": [ + 76, + 766, + 398, + 782 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Imagenet-1k. In Table 11 we compare architectures trained with cosub with state-of-the-art results from the literature for this architecture. We observe that architectures trained with cosub are very competitive. For instance, for ResNet-152, RegNetY-16GF, PiT-B we improve the best results reported in the literature by more than $0.9\\%$ top-1 accuracy on Imagenet-1k-val.", + "bbox": [ + 75, + 794, + 468, + 902 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/92643485a22938f5636cebd22bd797d810946cab5eebddb96c0d33038b8ad2f8.jpg", + "table_caption": [ + "Table 10. We measure the impact of layer-decay during the finetuning on Imagenet-1k for models pre-trained on Imagenet-21k with cosub during 90 epochs (default) and 270 epochs (long training)." + ], + "table_footnote": [], + "table_body": "
Modelnb params (×106)throughput (im/s)FLOPs (×109)Peak Mem (MB)Top-1 Acc.v2 Acc.
“Traditional” ConvNets
ResNet-50 [24, 55]25.625874.1218280.468.7
ResNet-101 [24, 55]44.515867.9226981.570.3
ResNet-101 - cosub44.515867.9226982.170.8
ResNet-152 [24, 55]60.2112211.6235982.070.6
ResNet-152 - cosub60.2112211.6235983.172.1
RegNetY-8GF [42, 55]39.211588.0393982.271.1
RegNetY-16GF [42, 50]83.671416.0520482.972.4
RegNetY-16GF - cosub83.671416.0520483.872.8
Vision Transformers derivatives
Swin-T [34]28.311094.5334581.369.5
Swin-B [34]87.853215.4469583.5-
PiT-S [26]23.518092.9329380.9-
PiT-S - cosub23.518092.9329381.369.7
PiT-B [26]73.861512.5756482.0-
PiT-B - cosub73.861512.5756484.174.1
Vanilla Vision Transformers
ViT-S [50]22.018914.698779.868.1
ViT-S - Deit III [51]22.018914.698781.470.5
ViT-S - cosub22.018914.698781.570.8
ViT-B [14]86.683117.5207877.9-
ViT-B - DeiT [50]86.683117.5207881.871.5
ViT-B - DeiT/distilled86.683117.5207883.473.2
ViT-B - DeiT III [51]86.683117.5207883.873.6
ViT-B - cosub86.683117.5207884.274.2
ViT-L - Deit III [51]304.427761.6378984.975.1
ViT-L - cosub304.427761.6378985.375.5
ViT-H - Deit III [51]632.1112167.4698485.275.9
ViT-H - cosub632.1112167.4698485.776.6
", + "bbox": [ + 504, + 88, + 893, + 431 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 11. Classification with Imagenet1k training. We compare with models trained on Imagenet-1k only at resolution $224 \\times 224$ without self-supervised pre-training (see supp. material for a comparison). We report Top-1 accuracy on Imagenet-1k-val and Imagenet-v2 with different measures of complexity: throughput, FLOPs, number of parameters and peak memory usage. The throughput and peak memory are measured on a single V100-32GB GPU with batch size fixed to 256 and mixed precision.", + "bbox": [ + 496, + 436, + 893, + 522 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Imagenet-21k. In Table 12 we compare ViT models pretrained with cosub on Imagenet-21k and finetuned with cosub on Imagenet-1k with other architectures and our baseline DeiT-III. Our results with vanilla ViT outperform the baseline and are competitive with recent architectures.", + "bbox": [ + 496, + 549, + 890, + 626 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Overfitting analysis. As recommended in prior works [51, 55] we perform an overfitting analysis. We evaluate our models trained with codist onImagenet-1k val andImagenet-v2 [43]. The results are reported in Figure 5. For ViT, we observe that cosub does not overfit more than the DeiT-III baseline [51]. Our results with other architectures in Table 12 concur with that observation: our results are comparatively stronger onImagenet-v2 than those reported in the literature for the exact same models.", + "bbox": [ + 496, + 631, + 892, + 767 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.4. Downstream tasks", + "text_level": 1, + "bbox": [ + 500, + 780, + 678, + 795 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Semantic segmentation. First we evaluate our ViT models pre-trained onImagenet with cosub for semantic segmentation on the ADE20k dataset [63]. ADE20k consists of 20k training and 5k validation images with labels over 150 categories. We adopt the training schedule from Swin: 160k iterations with UperNet [57]. At test time we evaluate with", + "bbox": [ + 496, + 809, + 893, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "11707", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/96d5499c49a9c7430027c1cc2f4ccb811fe22e565246ff59894e3860fd9dc429.jpg", + "table_caption": [], + "table_footnote": [ + "Table 12. Classification with ImageNet-21k pretraining. We report top-1 accuracy on the validation set of Imagenet1k and Imagenet-V2 with different measures of complexity. The peak memory usage is measured on a single V100-32GB GPU with batch size fixed to 256 and mixed precision. For Swin-L the memory peak is an estimation since we decreased the batch size to 128 to fit in memory. All models are evaluated at resolution 224 except EfficientNetV2 that use resolution 480. ViT are pretrained with a $\\times 3$ schedule, comparable to that used in the best DeiT-III baseline (270 vs. 240 epochs). All other cosub models are pretrained during 90 epochs on Imagenet21k, with 50 epochs of fine-tuning. The $\\tau$ hyper-parameter is set per model based on prior choices or best guess based on model size." + ], + "table_body": "
Architecturenb params (×106)throughput (im/s)FLOPs (×109)peak mem (MB)top1 valacc. v2
Convnets
EfficientNetV2-M [48]54.131225.0712786.275.9
EfficientNetV2-L [48]118.517953.0954086.876.9
ResNet-152 [24, 55]60.2112211.6235982.070.6
ResNet-152 - cosub60.2112211.6235983.173.1
RegnetY16GF - cosub83.671416.0520484.274.7
ConvNeXt-S- cosub50.27838.7221885.276.0
ConvNeXt-B [35]88.656315.4302985.875.6
ConvNeXt-B - cosub88.656315.4302985.876.9
ConvNeXt-L [35]197.834434.4486586.676.6
ConvNeXt-XL [35]350.224160.9695187.077.0
Transformers variations
Swin-B [34]87.853215.4469585.274.6
Swin-B - cosub87.853215.4469586.277.2
Swin-L [34]196.533734.5735086.376.3
Swin-L - cosub196.533734.5735087.178.1
PiT-B - cosub [26]73.861512.5756485.876.8
XCiT-S12 - cosub [17]26.213734.9133084.274.9
XCiT-M24 - cosub [17]84.455316.2201086.578.0
XCiT-L24 - cosub [17]189.033436.1331587.277.8
Vanilla Vision Transformers [14, 51]
ViT-S - cosub22.018914.698783.173.7
ViT-M - cosub39.013078.0132285.076.0
ViT-B - DeiT-III86.683117.6207885.776.5
ViT-B - cosub86.683117.6207886.377.0
ViT-L - DeiT-III304.427761.6378987.078.6
ViT-L - cosub304.427761.6378987.579.1
ViT-H - DeiT-III632.1112167.4698487.279.2
ViT-H - cosub632.1112167.4698488.080.0
", + "bbox": [ + 83, + 88, + 460, + 435 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/f8c82359cfd5310e3d42ed277bb8b03a5174a9a81497154610bcc05e2cced25e.jpg", + "image_caption": [ + "Figure 5. Overfitting measurement: top-1 accuracy on Imagenet-val vs. Imagenet-v2 for models in Tables 11 and 12 pre-trained on Imagenet-1k and Imagenet-21k, respectively. Our cosub ViTs (plain lines and points) do not overfit more than DeiT-III [51] overall. Our Imagenet-21k results for Swin and Convnext generalize much better on v2 than the original models." + ], + "image_footnote": [], + "bbox": [ + 76, + 588, + 264, + 738 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/8d84030110f5e627046f800c771d9c5379935fe7bc51afca709a02acd1ff4a8c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 276, + 588, + 450, + 738 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "a single scale and multi-scale. Our UperNet implementation is the same as in DeiT III [51]. Our results are reported in Table 13. We observe that vanilla ViTs trained with cos-sub outperform our baseline DeiT-III but also have better FLOPs-accuracy trade-offs than recent architectures.", + "bbox": [ + 75, + 824, + 470, + 902 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/f26f64938d3d25c0ecd5f347f517fdbae0e44618d7f62d08e13c86c1957f0d59.jpg", + "table_caption": [], + "table_footnote": [ + "Table 13. ADE20k semantic segmentation performance using UperNet [57], in comparable setting as prior works [13, 17, 34]. All models are pretrained on Imagenet-1k, except bottom models identified with $\\dagger$ , which are pre-trained on Imagenet-21k. By default the finetuning resolution on ADE20k is $512 \\times 512$ except when mentioned otherwise (for Swin)." + ], + "table_body": "
Backbone#params (×106)FLOPs (×109)Single-scale mIoUMulti-scale mIoU
DeiT-S52.01099-44.0
Swin-T59.994544.546.1
ViT-S – DeiT-III41.758845.646.8
ViT-S – cosub41.758847.048.0
PatchConvNet-B60140.6125848.148.6
PatchConvNet-B120229.8155049.450.3
MAE ViT-B127.7128348.1-
Swin-B121.0118848.149.7
ViT-B – DeiT-III127.7128349.350.2
ViT-B – cosub127.7128349.349.9
ViT-L – DeiT-III353.6223151.552.0
ViT-L – cosub353.6223152.553.1
PatchConvNet-B60†140.6125850.551.1
Swin-B† (640 × 640)121.0184150.051.6
ViT-B – DeiT-III†127.7128353.454.1
ViT-B – cosub†127.7128353.754.7
PatchConvNet-L120†383.7208652.252.9
Swin-L† (640 × 640)234.03230-53.5
ViT-L – DeiT-III†353.6223154.655.6
ViT-L – cosub†353.6223155.756.3
", + "bbox": [ + 506, + 88, + 887, + 351 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/dfc70bd0a6d8ace522a46704770d7ce48a4af37ac0b5cea2084ea240587dbb2f.jpg", + "table_caption": [], + "table_footnote": [ + "Table 14. ViT models pre-trained with cosub or DeiT-III on Imagenet-1k and finetuned on six different target datasets. We note that for small datasets (CIFAR, Flowers, and Cars) our approach is useful for small models but neutral for larger models. The gains are more significant when transferring to the larger iNaturalist-18 and iNaturalist-19 datasets." + ], + "table_body": "
ModelCifar-10Cifar-100FlowersCarsiNat-18iNat-19
ViT-S - DeiT-III98.990.696.489.967.172.7
ViT-B - DeiT-III99.392.598.693.473.678.0
ViT-L - DeiT-III99.393.498.994.575.679.3
ViT-S - cosub99.191.797.493.070.175.6
ViT-B - cosub99.192.698.493.574.178.1
ViT-L - cosub99.493.598.894.576.280.1
", + "bbox": [ + 506, + 433, + 883, + 523 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Transfer learning. We now measure how the performance improvements observed with cosub translate to other classification problems. For this purpose, we performed transfer learning on the six different datasets used in DeiT-III. Our results are reported Table 14. Our pre-trained and fine-tuned models with cosub generally improve the baseline. The gains are overall more significant on the more challenging datasets like iNaturalist 2018 and iNaturalist 2019.", + "bbox": [ + 498, + 604, + 890, + 726 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 739, + 617, + 753 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Co-training submodels (cosub) is an effective way to improve existing deep residual networks. It is straightforward to implement, just involving a few lines of code. It does not need a pre-trained teacher, and it only maintains a single set of weights for the model. Extensive experimental results on image classification, transfer learning and semantic segmentation show that cosub is overall extremely effective. It works off-the-shelf and improves the state of the art for various network architectures, including convnets like Regnet.", + "bbox": [ + 496, + 763, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "11708", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Samira Abnar, Mostafa Dehghani, and Willem Zuidema. Transferring inductive biases through knowledge distillation. arXiv preprint arXiv:2006.00555, 2020. 2", + "[2] Rohan Anil, Gabriel Pereyra, Alexandre Passos, Robert Ormandi, George E Dahl, and Geoffrey E Hinton. Large scale distributed neural network training through online distillation. arXiv preprint arXiv:1804.03235, 2018. 2", + "[3] Mahmoud Assran, Mathilde Caron, Ishan Misra, Piotr Bojanowski, Florian Bordes, Pascal Vincent, Armand Joulin, Mike Rabbat, and Nicolas Ballas. Masked siamese networks for label-efficient learning. In European Conference on Computer Vision, pages 456-473. Springer, 2022. I", + "[4] L. Ba and R. Caruana. Do deep nets really need to be deep? In NeurIPS, 2014. 2", + "[5] Hangbo Bao, Li Dong, and Furu Wei. Beit: Bert pre-training of image transformers. arXiv preprint arXiv:2106.08254, 2021. 1, 5, I, II", + "[6] Irwan Bello, W. Fedus, Xianzhi Du, E. D. Cubuk, A. Srinivas, Tsung-Yi Lin, Jonathon Shlens, and Barret Zoph. Revisiting ResNets: Improved training and scaling strategies. arXiv preprint arXiv:2103.07579, 2021. II", + "[7] H. Bourlard and Y. Kamp. Auto-association by multilayer perceptrons and singular value decomposition. Biological Cybernetics, 59:291-294, 1988. 1", + "[8] Leo Breiman. Bagging predictors. Machine learning, 24(2):123-140, 1996. 2", + "[9] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. arXiv preprint arXiv:2104.14294, 2021. 1, 2, II", + "[10] Kevin Clark, Minh-Thang Luong, Quoc V Le, and Christopher D Manning. Electra: Pre-training text encoders as discriminators rather than generators. arXiv preprint arXiv:2003.10555, 2020. I", + "[11] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In Conference on Computer Vision and Pattern Recognition, pages 248–255, 2009. 1", + "[12] Carl Doersch, Ankush Gupta, and Andrew Zisserman. Crosstransformers: spatially-aware few-shot transfer. arXiv preprint arXiv:2007.11498, 2020. 1", + "[13] Xiaoyi Dong, Jianmin Bao, Dongdong Chen, Weiming Zhang, Nenghai Yu, Lu Yuan, Dong Chen, and Baining Guo. Cswin transformer: A general vision transformer backbone with cross-shaped windows. arXiv preprint arXiv:2107.00652, 2021. 8", + "[14] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations, 2021. 5, 7, 8", + "[15] Alaaeldin El-Nouby, Gautier Izacard, Hugo Touvron, Ivan Laptev, Hervé Jegou, and Edouard Grave. Are large-scale datasets necessary for self-supervised pre-training? arXiv preprint arXiv:2112.10740, 2021. 1" + ], + "bbox": [ + 78, + 114, + 468, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[16] Alaaeldin El-Nouby, Natalia Neverova, Ivan Laptev, and Hervé Jégou. Training vision transformers for image retrieval. arXiv preprint arXiv:2102.05644, 2021. 1", + "[17] Alaaeldin El-Nouby, Hugo Touvron, Mathilde Caron, Piotr Bojanowski, Matthijs Douze, Armand Joulin, Ivan Laptev, Natalia Neverova, Gabriel Synnaeve, Jakob Verbeek, et al. Xcit: Cross-covariance image transformers. arXiv preprint arXiv:2106.09681, 2021. 5, 8, II", + "[18] Thomas Elsken, Jan Hendrik Metzen, and Frank Hutter. Neural architecture search: A survey. Journal of Machine Learning Research, 20(1):1997-2017, 2019. 3", + "[19] Dumitru Erhan, Aaron Courville, Yoshua Bengio, and Pascal Vincent. Why does unsupervised pre-training help deep learning? In AISTATS, 2010. 1", + "[20] Angela Fan, Edouard Grave, and Armand Joulin. Reducing transformer depth on demand with structured dropout. arXiv preprint arXiv:1909.11556, 2019. ICLR 2020. 2, 3, 4, IV", + "[21] Rohit Girdhar, Alaaeldin El-Nouby, Mannat Singh, Kalyan Vasudev Alwala, Armand Joulin, and Ishan Misra. Omnimea: Single model masked pretraining on images and videos. arXiv preprint arXiv:2206.08356, 2022. I", + "[22] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Doll'ar, and Ross B. Girshick. Masked autoencoders are scalable vision learners. arXiv preprint arXiv:2111.06377, 2021. 1, 5, I, II", + "[23] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. arXiv preprint arXiv:1911.05722, 2019. 1", + "[24] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Conference on Computer Vision and Pattern Recognition, 2016. 1, 5, 7, 8, II", + "[25] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Identity mappings in deep residual networks. arXiv preprint arXiv:1603.05027, 2016. 1, 5", + "[26] Byeongho Heo, Sangdoo Yun, Dongyoon Han, Sanghyuk Chun, Junsuk Choe, and Seong Joon Oh. Rethinking spatial dimensions of vision transformers. arXiv preprint arXiv:2103.16302, 2021. 5, 7, 8, II", + "[27] G. Hinton and R. Zemel. Autoencoders, minimum description length and Helmholtz free energy. In NeurIPS, 1993. 1", + "[28] Geoffrey E. Hinton, Oriol Vinyals, and J. Dean. Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531, 2015. 2, 3", + "[29] Grant Van Horn, Oisin Mac Aodha, Yang Song, Alexander Shepard, Hartwig Adam, Pietro Perona, and Serge J. Belongie. The iNaturalist species classification and detection dataset. arXiv preprint arXiv:1707.06642, 2017. II", + "[30] Grant Van Horn, Oisin Mac Aodha, Yang Song, Alexander Shepard, Hartwig Adam, Pietro Perona, and Serge J. Belongie. The inaturalist challenge 2018 dataset. arXiv preprint arXiv:1707.06642, 2018. II", + "[31] Gao Huang, Yu Sun, Zhuang Liu, Daniel Sedra, and Kilian Q. Weinberger. Deep networks with stochastic depth. In European Conference on Computer Vision, 2016. 2, 4, I" + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "11709", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[32] Jonathan Krause, Michael Stark, Jia Deng, and Li Fei-Fei. 3d object representations for fine-grained categorization. In IEEE Workshop on 3D Representation and Recognition, 2013. II", + "[33] Alex Krizhevsky. Learning multiple layers of features from tiny images. Technical report, CIFAR, 2009. II", + "[34] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. arXiv preprint arXiv:2103.14030, 2021. 5, 7, 8, II", + "[35] Zhuang Liu, Hanzi Mao, Chao-Yuan Wu, Christoph Feichtenhofer, Trevor Darrell, and Saining Xie. A convnet for the 2020s. arXiv preprint arXiv:2201.03545, 2022. 8, II", + "[36] I. Loshchilov and F. Hutter. Fixing weight decay regularization in adam. arXiv preprint arXiv:1711.05101, 2017. I", + "[37] G. J. McLachlan and D. Peel. Finite Mixture Models. Wiley Series in Probability and Statistics. John Wiley & Sons, 2000. 2", + "[38] M-E. Nilsback and A. Zisserman. Automated flower classification over a large number of classes. In Proceedings of the Indian Conference on Computer Vision, Graphics and Image Processing, 2008. II", + "[39] Maxime Oquab, Leon Bottou, Ivan Laptev, and Josef Sivic. Learning and transferring mid-level image representations using convolutional neural networks. In Conference on Computer Vision and Pattern Recognition, 2014. 1", + "[40] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. In Advances in neural information processing systems, pages 8026-8037, 2019. 4", + "[41] Ilija Radosavovic, Raj Prateek Kosaraju, Ross Girshick, Kaiming He, and Piotr Dólar. Designing network design spaces. In Conference on Computer Vision and Pattern Recognition, 2020. 3", + "[42] Ilija Radosavovic, Raj Prateek Kosaraju, Ross B. Girshick, Kaiming He, and Piotr Dollar. Designing network design spaces. Conference on Computer Vision and Pattern Recognition, 2020. 5, 7, II", + "[43] B. Recht, Rebecca Roelofs, L. Schmidt, and V. Shankar. Do ImageNet classifiers generalize to ImageNet? In International Conference on Machine Learning, 2019. 7", + "[44] Adria Ruiz and Jakob Verbeek. Anytime inference with distilled hierarchical neural ensembles. In AAAI, 2021. 2", + "[45] Guocong Song and Wei Chai. Collaborative learning for deep neural networks. NeurIPS, 31, 2018. 2", + "[46] Nitish Srivastava, Geoffrey Hinton, Alex Krizhevsky, Ilya Sutskever, and Ruslan Salakhutdinov. Dropout: A simple way to prevent neural networks from overfitting. In Journal of Machine Learning Research, 2014. 1, 2, 3", + "[47] Christian Szegedy, Vincent Vanhoucke, Sergey Ioffe, Jonathon Shlens, and Zbigniew Wojna. Rethinking the inception architecture for computer vision. Conference on Computer Vision and Pattern Recognition, 2016. 1", + "[48] Mingxing Tan and Quoc V. Le. Efficientnetv2: Smaller models and faster training. In International Conference on Machine Learning, 2021. 8" + ], + "bbox": [ + 78, + 90, + 468, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[49] Antti Tarvainen and Harri Valpola. Mean teachers are better role models: Weight-averaged consistency targets improve semi-supervised deep learning results. Neurips, 30, 2017. 2", + "[50] Hugo Touvron, M. Cord, M. Douze, F. Massa, Alexandre Sablayrolles, and H. Jégou. Training data-efficient image transformers & distillation through attention. International Conference on Machine Learning, 2021. 2, 3, 7", + "[51] Hugo Touvron, Matthieu Cord, and Hervé Jégou. Deit III: Revenge of the vit. arXiv preprint arXiv:2204.07118, 2022. 5, 6, 7, 8, I, II", + "[52] Hugo Touvron, Matthieu Cord, Alexandre Sablayrolles, Gabriel Synnaeve, and Hervé Jégou. Going deeper with image transformers. International Conference on Computer Vision, 2021. 4, 5", + "[53] Longhui Wei, An Xiao, Lingxi Xie, X. Chen, Xiaopeng Zhang, and Q. Tian. Circumventing outliers of autoaugment with knowledge distillation. In European Conference on Computer Vision, 2020. 2", + "[54] Ross Wightman. Pytorch image models. https://github.com/rwrightman/pytorch-image-models, 2019.4", + "[55] Ross Wightman, Hugo Touvron, and Hervé Jégou. Resnet strikes back: An improved training procedure in timm. arXiv preprint arXiv:2110.00476, 2021. 5, 7, 8, II", + "[56] Mitchell Wortsman, Gabriel Ilharco, Samir Ya Gadre, Rebecca Roelofs, Raphael Gontijo-Lopes, Ari S Morcos, Hongseok Namkoong, Ali Farhadi, Yair Carmon, Simon Kornblith, et al. Model soups: averaging weights of multiple fine-tuned models improves accuracy without increasing inference time. In International Conference on Machine Learning, 2022. 3, 4", + "[57] Tete Xiao, Yingcheng Liu, Bolei Zhou, Yuning Jiang, and Jian Sun. Unified perceptual parsing for scene understanding. In European Conference on Computer Vision, 2018. 7, 8", + "[58] Jason Yosinski, Jeff Clune, Yoshua Bengio, and Hod Lipson. How transferable are features in deep neural networks? arXiv preprint arXiv:1411.1792, 2014. 1", + "[59] L. Yuan, F. Tay, G. Li, T. Wang, and Jiashi Feng. Revisit knowledge distillation: a teacher-free framework. Conference on Computer Vision and Pattern Recognition, 2020. 2", + "[60] Sangdoo Yun, Dongyoon Han, Seong Joon Oh, Sanghyuk Chun, Junsuk Choe, and Youngjoon Yoo. CutMix: Regularization strategy to train strong classifiers with localizable features. arXiv preprint arXiv:1905.04899, 2019. 1", + "[61] Hongyi Zhang, Moustapha Cisse, Yann N. Dauphin, and David Lopez-Paz. mixup: Beyond empirical risk minimization. arXiv preprint arXiv:1710.09412, 2017. 1", + "[62] Ying Zhang, Tao Xiang, Timothy M Hospedales, and Huchuan Lu. Deep mutual learning. In Conference on Computer Vision and Pattern Recognition, pages 4320-4328, 2018. 2", + "[63] Bolei Zhou, Hang Zhao, Xavier Puig, Sanja Fidler, Adela Barriuso, and Antonio Torralba. Scene parsing through ade20k dataset. Conference on Computer Vision and Pattern Recognition, 2017. 7", + "[64] Xiatian Zhu, Shaogang Gong, et al. Knowledge distillation by on-the-fly native ensemble. Neurips, 31, 2018. 2" + ], + "bbox": [ + 501, + 92, + 890, + 895 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "11710", + "bbox": [ + 480, + 945, + 519, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2023/Co-Training 2L Submodels for Visual Recognition/51b2445a-3487-4e2d-824b-ee9b871d1274_model.json b/2023/Co-Training 2L Submodels for Visual Recognition/51b2445a-3487-4e2d-824b-ee9b871d1274_model.json new file mode 100644 index 0000000000000000000000000000000000000000..fbd909881f5640b7a77dd499f6a5f9f539194198 --- /dev/null +++ b/2023/Co-Training 2L Submodels for Visual Recognition/51b2445a-3487-4e2d-824b-ee9b871d1274_model.json @@ -0,0 +1,2442 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.237, + 0.13, + 0.735, + 0.154 + ], + "angle": 0, + "content": "Co-training \\(2^{L}\\) Submodels for Visual Recognition" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.187, + 0.357, + 0.205 + ], + "angle": 0, + "content": "Hugo Touvron\\*,†" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.206, + 0.361, + 0.222 + ], + "angle": 0, + "content": "Piotr Bojanowski*" + }, + { + "type": "text", + "bbox": [ + 0.432, + 0.187, + 0.558, + 0.204 + ], + "angle": 0, + "content": "Matthieu Cord†" + }, + { + "type": "text", + "bbox": [ + 0.432, + 0.206, + 0.558, + 0.222 + ], + "angle": 0, + "content": "Jakob Verbeek*" + }, + { + "type": "text", + "bbox": [ + 0.628, + 0.187, + 0.761, + 0.205 + ], + "angle": 0, + "content": "Maxime Oquab\\*" + }, + { + "type": "text", + "bbox": [ + 0.64, + 0.206, + 0.75, + 0.224 + ], + "angle": 0, + "content": "Hervé Jégou*" + }, + { + "type": "text", + "bbox": [ + 0.277, + 0.226, + 0.457, + 0.243 + ], + "angle": 0, + "content": "*Meta AI / FAIR Paris" + }, + { + "type": "text", + "bbox": [ + 0.526, + 0.226, + 0.7, + 0.245 + ], + "angle": 0, + "content": "†Sorbonne University" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.269, + 0.314, + 0.285 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.301, + 0.471, + 0.451 + ], + "angle": 0, + "content": "We introduce submodel co-training, a regularization method related to co-training, self-distillation and stochastic depth. Given a neural network to be trained, for each sample we implicitly instantiate two altered networks, \"submodels\", with stochastic depth: we activate only a subset of the layers. Each network serves as a soft teacher to the other, by providing a loss that complements the regular loss provided by the one-hot label. Our approach, dubbed \"cosub\", uses a single set of weights, and does not involve a pre-trained external model or temporal averaging." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.452, + 0.47, + 0.573 + ], + "angle": 0, + "content": "Experimentally, we show that submodel co-training is effective to train backbones for recognition tasks such as image classification and semantic segmentation. Our approach is compatible with multiple architectures, including RegNet, ViT, PiT, XCiT, Swin and ConvNext. Our training strategy improves their results in comparable settings. For instance, a ViT-B pretrained with cosub on ImageNet-21k obtains \\(87.4\\%\\) top-1 acc. @448 on ImageNet-val." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.589, + 0.21, + 0.604 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.614, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Although the fundamental ideas of deep trainable neural networks have been around for decades, only recently have barriers been removed to allow breakthroughs in successfully training deep neural architectures in practice. Many of these barriers are related to non-convex optimization in one way or another, which is central to the success of modern neural networks. The optimization challenges have been addressed from multiple angles in the literature. First, modern architectures are designed to facilitate the optimization of very deep networks. An exceptionally successful design principle is using residual connections [24, 25]. Although this does not change the expressiveness of the functions that the network can implement, the improved gradient flow alleviates, to some extent, the difficulties of optimizing very deep networks. Another key element to the optimization is the importance of data, revealed by the step-change in visual recognition performance resulting from the ImageNet dataset [11], and the popularization of transfer learning with pre-training on large datasets [39, 58]." + }, + { + "type": "image", + "bbox": [ + 0.589, + 0.264, + 0.808, + 0.53 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.539, + 0.894, + 0.589 + ], + "angle": 0, + "content": "Figure 1. Co-training of submodels (cosub): for each image, two sub-models are sampled by randomly dropping layers of the full model. The training signal for each submodel mixes the cross-entropy loss from the image label with a self-distillation loss obtained from the other submodel." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.614, + 0.895, + 0.903 + ], + "angle": 0, + "content": "However, even when (pre-)trained with millions of images, recent deep networks with millions if not billions of parameters, are still heavily overparameterized. Traditional regularization like weight decay, dropout [46], or label smoothing [47] are limited in their ability to address this issue. Data-augmentation strategies, including those mixing different images like Mixup [61] and CutMix [60], have proven to provide a complementary data-driven form of regularization. More recently, multiple works propose to resort to self-supervised pre-training. These approaches rely on a proxy objective that generally provides more supervision signal than the one available from labels, like in recent (masked) auto-encoders [5,16,22], which were popular in the early deep learning literature [7,19,27]. Similarly, contrastive approaches [23] or self-distillation [9] provide a richer supervision less prone to supervision collapse [12]. Overall, self-supervised learning makes it possible to learn larger models with less data, possibly reducing the need of a pre-training stage [15]." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "11701" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.244 + ], + "angle": 0, + "content": "Distillation is a complementary approach to improve optimization. Distillation techniques were originally developed to transfer knowledge from a teacher model to a student model [4, 28], allowing the student to improve over learning from the data directly. In contrast to traditional distillation, co-distillation does not require pre-training a (strong) teacher. Instead, a pool of models supervise each other. Practically, it faces several limitations, including the difficulty of jointly training more than two students for complexity reasons, as it involves duplicating the weights." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.249, + 0.473, + 0.446 + ], + "angle": 0, + "content": "In this paper, we propose a practical way to enable cotraining for a very large number of students. We consider a single target model to be trained, and we instantiate two submodels on-the-fly, simply by layerwise dropout [20,31]. This gives us two neural networks through which we can backpropagate to the shared parameters of the target model. In addition to the regular training loss, each submodel serves as a teacher to the other, which provides an additional supervision signal ensuring the consistency across the submodels. Our approach is illustrated in Figure 1: the parameter \\(\\lambda\\) controls the importance of the co-training loss compared to the label loss, and our experiments show that it significantly increases the final model accuracy." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.45, + 0.473, + 0.571 + ], + "angle": 0, + "content": "This co-training across different submodels, which we refer to as cosub, can be regarded as a massive co-training between \\(2^{L}\\) models that share a common set of parameters, where \\(L\\) is the number of layers in the target architecture. The target model can be interpreted as the expectation of all models. With a layer drop-rate set to 0.5, for instance for a ViT-H model, all submodels are equiprobable, and then it amounts to averaging the weights of \\(2^{2 \\times 32}\\) models." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.576, + 0.405, + 0.59 + ], + "angle": 0, + "content": "Our contributions can be summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.599, + 0.468, + 0.704 + ], + "angle": 0, + "content": "- We introduce a novel training approach for deep neural networks: we co-train submodels. This significantly improves the training of most models, establishing the new state of the art in multiple cases. For instance, after pre-training ViT-B onImagenet-21k and fine-tuning it at resolution 448, we obtain \\(87.4\\%\\) top-1 accuracy onImagenet-val." + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.705, + 0.468, + 0.749 + ], + "angle": 0, + "content": "- We provide an efficient implementation to subsample models on the fly. It is a simple yet effective variation of stochastic depth [31] to drop residual blocks." + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.751, + 0.468, + 0.811 + ], + "angle": 0, + "content": "- We provide multiple analyses and ablations. Noticeably, we show that our submodels are effective models by themselves even with significant trimming, similar to LayerDrop [20] in natural language processing." + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.811, + 0.468, + 0.87 + ], + "angle": 0, + "content": "- We validate our approach on multiple architectures (like ViT, ResNet, RegNet, PiT, XCiT, Swin, ConvNext), both for image classification -trained from scratch or with transfer-, and semantic segmentation." + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.871, + 0.468, + 0.901 + ], + "angle": 0, + "content": "- We will share models/code for reproducibility in the DeiT repository." + }, + { + "type": "list", + "bbox": [ + 0.096, + 0.599, + 0.468, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.09, + 0.638, + 0.106 + ], + "angle": 0, + "content": "2. Related work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.121, + 0.892, + 0.363 + ], + "angle": 0, + "content": "Knowledge distillation. Originally, distillation was introduced as a way to train a model such that it reproduces the performance of another model [4, 28]. The typical use-case is to improve the quality of a relatively small model by leveraging a strong teacher, whose complexity may be prohibitive for a practical deployment. The teacher's soft labels have a similar effect as label smoothing [59]. As shown by Wei et al. [53], the teacher's supervision takes into account the effects of the data augmentation, which sometimes causes a misalignment between the real label and the image. Knowledge distillation can transfer inductive biases [1] in a soft way in a student model when using a teacher model these biases are enforced in a hard way. Touvron et al. [50] proposed a variant of distillation adapted to Vision Transformer (ViT), showing the effectiveness of using a ConvNet teacher for a ViT student." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.369, + 0.892, + 0.505 + ], + "angle": 0, + "content": "Mixture models. Ensembling has a long history that we can trace back to the origins of statistics [37] and the possibility of improving the precision of measurements with multiple observations. In machine learning, bagging [8] combines multiple weak classifiers to produce a strong one. This idea is naturally extended to neural networks, where it can offer improved stability or accuracy, or other properties like anytime inference [44]. A mixture model can also be seen as a larger model with an internal parallel structure." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.511, + 0.892, + 0.678 + ], + "angle": 0, + "content": "Co-distillation. At the interface of mixture models and distillation, the concept of co-distillation [62, 64] does not require a prior teacher: the mixture serves as the teacher to all the networks in the mixture. The models are jointly optimized, which leads to improved accuracy of the individual models [2]. This can be regarded as a form of collaborative learning between the different elements of the mixture [45]. Compared to training only one model, co-distillation involves training two or more models, each requiring storage for weights and activation, and computing backward and forward passes." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.684, + 0.892, + 0.835 + ], + "angle": 0, + "content": "Exponential moving average teacher. A special form of teacher is a model obtained from the student model itself. Tarvainen and Valpila [49] show that such a model obtained by temporal averaging of intermediate checkpoints during the training is an effective teacher, that can be obtained with an exponential moving average (EMA). This idea has also been adopted in self-supervised training with learning schemes like DINO [9]. This method involves storing a copy of the weights corresponding to the temporal averaging carried out by the EMA model." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Dropout and model populations. Dropout [46] is an effective way to regularize models. With residual architectures, one effective way to train deeper architecture is stochastic depth [31]. It reduces the size of a network at" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "11702" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.129, + 0.1, + 0.207, + 0.195 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.125, + 0.209, + 0.204, + 0.222 + ], + "angle": 0, + "content": "(a) supervised" + }, + { + "type": "image", + "bbox": [ + 0.254, + 0.1, + 0.368, + 0.195 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.267, + 0.208, + 0.348, + 0.232 + ], + "angle": 0, + "content": "(b) knowledge distillation" + }, + { + "type": "image", + "bbox": [ + 0.412, + 0.1, + 0.53, + 0.196 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.42, + 0.208, + 0.523, + 0.221 + ], + "angle": 0, + "content": "(c) self-distillation" + }, + { + "type": "image", + "bbox": [ + 0.558, + 0.096, + 0.695, + 0.198 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.592, + 0.208, + 0.672, + 0.221 + ], + "angle": 0, + "content": "(d) co-training" + }, + { + "type": "image", + "bbox": [ + 0.704, + 0.095, + 0.888, + 0.203 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.739, + 0.207, + 0.875, + 0.221 + ], + "angle": 0, + "content": "(e) submodel co-training" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.243, + 0.593, + 0.351 + ], + "angle": 0, + "content": "Figure 2. Brief summary of related works and of submodel co-training (cosub). (a) supervised baseline: the supervision is provided as a one-hot label. (b) Knowledge distillation [28]: a teacher provides a (soft) target. A simple yet effective variant [50] combines the predicted label of the teacher with the image label. KD requires a pre-existing teacher that must be stored for inference. (c) In self-distillation, the teacher is obtained from the model itself, typically by model averaging. (d) Co-training involves two distinct models that serves as teacher to each other. (e) Submodel co-training generates on-the-fly two distinct models (amongst \\(2^{L}\\)) for each sample, which serves a teacher to each other. We only need to store a single set of weights for the model and optimizer because the random selections \\(\\theta_{1} = \\mathcal{T}(\\theta)\\) (resp., \\(\\theta_{2}\\)) of submodels allow us to back-propagate on \\(\\theta\\)." + }, + { + "type": "table", + "bbox": [ + 0.598, + 0.234, + 0.892, + 0.326 + ], + "angle": 0, + "content": "
methodmodel paramscompute*
weightsoptimizer
(a) supervised×1×1×1
(b) KD†×2×1×1.33
(c) mean teacher×2×1×1.33
(d) co-training×2×2×2
(e) Cosub (ours)×1×1×2
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.598, + 0.327, + 0.892, + 0.351 + ], + "angle": 0, + "content": "\\*backward pass assumed \\(2\\times\\) the complexity of forward \\(\\dagger\\) requires a trained teacher model" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.371, + 0.469, + 0.506 + ], + "angle": 0, + "content": "training time by dropping residual blocks during training. The initial goal of this approach was to improve the training of deep network. However, there are other implications: in natural language processing, Fan et al. show that the counterpart of stochastic depth, namely LayerDrop [20], is effective to reduce the transformer depth on demand. Interestingly, submodels extracted from a model trained with LayerDrop are stronger than those trained with the full target depth with a proper selection strategy for layers." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.508, + 0.469, + 0.642 + ], + "angle": 0, + "content": "In our paper, we regard stochastic depth as a regularization technique, but at the same time we adopt the point of view of LayerDrop, i.e., an effective way to train a population of submodels that share parameters, or entire layers in our case. From that viewpoint, stochastic depth amounts to training \\(2^{L}\\) distinct models. This is different from population training as involved in network space design [41], whose goal is related to network architecture search [18], which aims at optimizing the architecture itself." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.644, + 0.469, + 0.764 + ], + "angle": 0, + "content": "More recently, model soups [56] are models obtained by averaging the weights of multiple models finetuned with different hyperparameter configurations. The authors show that it often improves accuracy and robustness. We point out that stochastic depth can be regarded as a special form of model soup over the entire population of submodels that we can instantiate with stochastic depth. We further discuss this relationship in the next section." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.779, + 0.285, + 0.796 + ], + "angle": 0, + "content": "3. Submodel co-training" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.805, + 0.469, + 0.85 + ], + "angle": 0, + "content": "In this section, we present our cosub approach, for constraining submodels. We consider a neural network \\( f_{\\theta} \\) parameterized by learnable parameters \\( \\theta \\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Submodel instantiation: model augmentation operator. We first define a model augmentation operator \\(\\mathcal{T}\\). For a given neural network, it provides a set of parameters \\(\\theta^{\\prime} =\\)" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.371, + 0.892, + 0.492 + ], + "angle": 0, + "content": "\\(\\mathcal{T}(\\theta, R)\\) that allow us to define variations \\(f_{\\theta'}\\) of the function \\(f_{\\theta}\\) by drawing a random variable \\(R\\). The model augmentation is such that \\(f_{\\theta}\\) and \\(f_{\\theta'}\\) share parameters, hence any update on \\(\\theta'\\) modifies \\(\\theta\\) and therefore \\(f_{\\theta}\\) accordingly. A simple way to define \\(\\mathcal{T}\\) is to replace some parameters by zeros, which corresponds to dropout [46]. In this paper, we focus on stochastic depth, which has interesting connections with model averaging [56]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.497, + 0.891, + 0.527 + ], + "angle": 0, + "content": "Overview. For a given training sample \\( x \\) within a batch, the training is as follows (see also Figure 1):" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.53, + 0.84, + 0.544 + ], + "angle": 0, + "content": "1. We first data-augment the image, producing \\(\\hat{x}\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.545, + 0.892, + 0.589 + ], + "angle": 0, + "content": "2. The image is duplicated with the batch, effectively doubling the batch size, which hence contains two identical copies of each augmented image1." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.59, + 0.892, + 0.649 + ], + "angle": 0, + "content": "3. We determine the stochastic depth pattern for each sample according to the target drop-rate \\(\\tau\\), which amounts to producing two functions \\(f_{\\theta_1}\\) and \\(f_{\\theta_2}\\). See Section 4 for the details of this procedure." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.65, + 0.892, + 0.681 + ], + "angle": 0, + "content": "4. The forward pass proceeds as usual: we compute the soft output labels \\( y_{1} = f_{\\theta_{1}}(\\hat{x}) \\) and \\( y_{2} = f_{\\theta_{2}}(\\hat{x}) \\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.681, + 0.892, + 0.725 + ], + "angle": 0, + "content": "5. We compute the losses and the backwards pass accordingly. Note that the gradients on \\(\\theta_{1}\\) and \\(\\theta_{2}\\) are directly used to update \\(\\theta\\), as they are just subsets of \\(\\theta\\)." + }, + { + "type": "list", + "bbox": [ + 0.512, + 0.53, + 0.892, + 0.725 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.733, + 0.892, + 0.839 + ], + "angle": 0, + "content": "Loss. Each submodel is trained using a weighted average of (i) the standard binary cross-entropy loss obtained from the image label \\(y\\), and (ii) a binary cross-entropy loss w.r.t. the soft-labels computed for the the same image by the other submodel. The respective weight of the standard binary cross-entropy loss \\(\\mathcal{L}_{\\mathrm{label}}\\) versus the cosub loss \\(\\mathcal{L}_{\\mathrm{cosub}}\\) is controlled by the hyper-parameter \\(\\lambda\\), as" + }, + { + "type": "equation", + "bbox": [ + 0.581, + 0.85, + 0.892, + 0.866 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {t o t}} = \\lambda \\mathcal {L} _ {\\text {l a b e l}} + (1 - \\lambda) \\mathcal {L} _ {\\text {c o s u b}}. \\tag {1}\n\\]" + }, + { + "type": "page_footnote", + "bbox": [ + 0.5, + 0.875, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Different data-augmentations did not provide any advantage in our preliminary experiments. We simply duplicate the batch for simplicity." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "11703" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.078, + 0.092, + 0.264, + 0.106 + ], + "angle": 0, + "content": "In details, the loss writes as" + }, + { + "type": "equation", + "bbox": [ + 0.094, + 0.114, + 0.47, + 0.186 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} _ {\\mathrm {t o t}} = \\lambda \\left(\\frac {\\mathcal {L} (y _ {1} , y) + \\mathcal {L} (y _ {2} , y)}{2}\\right) \\\\ + (1 - \\lambda) \\left(\\frac {\\mathcal {L} \\left(y _ {1} , \\mathrm {s g} \\left(y _ {2}\\right)\\right) + \\mathcal {L} \\left(y _ {2} , \\mathrm {s g} \\left(y _ {1}\\right)\\right)}{2}\\right), \\quad (2) \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.194, + 0.47, + 0.256 + ], + "angle": 0, + "content": "where \\(\\mathcal{L}(y, y')\\) is either the binary cross-entropy (BCE) or a cross-entropy (CE) loss. Importantly, when applying this loss, we do not back-propagate on the second term \\(y'\\). This is indicated by the stop-gradient operator \\(\\mathrm{sg}(\\cdot)\\)." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.259, + 0.473, + 0.562 + ], + "angle": 0, + "content": "Discussion. The submodel instantiation provides a model that is, by itself, a valid neural network model. Co-training submodels is a way to enforce that all such submodels produce a consistent output. Amongst these submodels, a very special case is when we retain all residual blocks, which is the model that we primarily intend to use at inference time. Note that Fan et al. [20] show, in an NLP context, that submodels extracted from a deeper model are superior, for a given target depth. With proper rescaling of the residual branches, this specific submodel with all blocks activated can be seen as the expectation over all models. This can be shown as follows: let us consider an extra scalar parameter \\( s_l \\) associated with a given residual block \\( r_l(x) \\), which we use as a multiplicative factor on output of each residual: \\( s_l = 1 \\) if the residual block \\( r_l \\) is included in the submodel, \\( s_l = 0 \\) otherwise. Each submodel is fully parameterized by a binary vector \\( s = (s_1, \\dots, s_L) \\), where \\( L \\) is the total number of layers. Therefore all submodels have the same parameters \\( \\theta \\), and only differ by \\( s \\) indicating the zeroed residual blocks. The weight expectation \\( \\bar{\\theta} \\) is hence" + }, + { + "type": "equation", + "bbox": [ + 0.139, + 0.572, + 0.469, + 0.612 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\bar {\\theta} = \\mathbb {E} _ {s \\in [ 0, 1 ] ^ {L}} [ \\theta ] = \\theta , \\\\ \\bar {s} = \\mathbb {E} _ {s \\in [ 0, 1 ] ^ {L}} [ s ] = [ 1 - \\tau , \\dots , 1 - \\tau ] ^ {\\top}, \\tag {3} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.621, + 0.47, + 0.713 + ], + "angle": 0, + "content": "where \\(\\bar{s}\\) is the scaling factor used in stochastic depth. Under a uniform distribution of submodels that is obtained when the stochastic drop rate \\(\\tau = 0.5\\), the inference-time model is exactly the average of all \\(2^{L}\\) models. While stochastic depth is often regarded as a regularization technique, this averaging interpretation relates it to the recent model soup [56]." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.725, + 0.314, + 0.743 + ], + "angle": 0, + "content": "4. Efficient stochastic depth" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.75, + 0.471, + 0.903 + ], + "angle": 0, + "content": "We revisit the original stochastic depth [31] formulation in order enable an efficient implementation, which we will release for PyTorch [40]. Instantiating a submodel for a sample amounts to selecting a subset of residual layers of the model and performing training on these. This can be implemented using the stochastic depth (SD) approach, whose objective was initially to improve the training of very deep networks. In stochastic depth, for each sample and each layers of the network, we select whether the layers will be dropped or not with a certain probability \\(\\tau\\)." + }, + { + "type": "image", + "bbox": [ + 0.499, + 0.092, + 0.895, + 0.203 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.21, + 0.894, + 0.26 + ], + "angle": 0, + "content": "Figure 3. Efficient implementation of stochastic depth, using the permute-select approach. In this example, we drop the residuals for samples 3 and 6, corresponding to a drop rate \\(\\tau = 0.25\\). For a drop rate higher than 0.1, the overhead of the approach is negligible in terms of memory and compute." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.277, + 0.892, + 0.352 + ], + "angle": 0, + "content": "In practice, e.g. in the timm library [54], stochastic depth is implemented by masking with zeros the residuals added for a given sample of a batch. This is not efficient: this naive approach performs the computations for a residual then throws it away, wasting computation." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.353, + 0.893, + 0.58 + ], + "angle": 0, + "content": "Our efficient stochastic depth (ESD) approach addresses this issue, saving both memory and compute, and is illustrated in Figure 3. For each layer, given a batch size \\( B \\) and a drop rate \\( \\tau \\), we apply the layer to \\( B_{\\mathrm{keep}} = \\mathrm{round}(B \\times (1 - \\tau)) \\) samples in the batch. In contrast to the original version of SD, our efficient version drops a fixed number of samples at each layer, where \\( d \\) is adjusted such that \\( B_{\\mathrm{keep}} \\) is an integer. In our experiments, this did not have an effect on the final accuracy of the models. Our efficient implementation proceeds as follows: (i) we apply a random permutation of the \\( B \\) samples, then (ii) we select the first \\( B_{\\mathrm{keep}} \\) samples. We then compute the residual function for the selected subset, then add the result onto the full batch using the built-in index_add function and scaling the result by \\( \\frac{1}{1 - \\tau} \\) to have the correct scaling as in the original SD formulation." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.584, + 0.892, + 0.734 + ], + "angle": 0, + "content": "Discussion: progressive vs uniform rate. In the original paper [31], stochastic depth drops a layer with a probability that is linearly increasing during the forward pass: layers closer to the output has a higher chance to be dropped. However this strategy is limited with high drop rates, as later layers are excessively dropped. Touvron et al. [52] adopt a uniform rate drop per layer. It is as effective as the progressive rate with vision transformers, but it makes it possible to target a drop-rate greater than 0.5 on average. For this reason we adopt an uniform drop rate everywhere." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.735, + 0.895, + 0.902 + ], + "angle": 0, + "content": "Our efficient stochastic depth implementation variant works both with uniform and progressive drop-rate. In the case of progressive, the effective batch size is decreasing during to the forward pass. One limitation is that our technique implies a quantization of the drop-rate, which can be problematic for small batch sizes. Therefore one must take care to properly verify that the drop-rate is not too severely quantized, in particular in the progressive case for which the drop-rate can be very high in the last layers. In the supplemental material (Appendix C), we discuss with more details the drop-rate quantization effect." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "11704" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.078, + 0.087, + 0.468, + 0.246 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.25, + 0.473, + 0.336 + ], + "angle": 0, + "content": "Figure 4. Cosub as population training: each submodel extracted by a transformation \\(\\mathcal{T}\\) is a valid neural network. Our cosub strategy can hence be regarded as co-training a large number of subnetworks. We plot the accuracy of the submodels as a function of the number of layers that we preserve. We drop layers with probability \\(\\tau = 0.2\\) for ViT-B and \\(\\tau = 0.45\\) for ViT-L. On average our method provides a significant boost in performance for the whole population of submodels extracted from the main model." + }, + { + "type": "table", + "bbox": [ + 0.085, + 0.353, + 0.468, + 0.41 + ], + "angle": 0, + "content": "
ModelViT-SViT-MViT-BViT-LViT-H
τ(0.05)(0.1)(0.2)(0.45)(0.6)
baseline81.482.583.784.584.9
baseline + cosub81.582.883.984.985.5
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.42, + 0.472, + 0.471 + ], + "angle": 0, + "content": "Table 1. Comparison of top-1 accuracy with ViT architecture trained with/without cosub at resolution 224 (800 epochs) onImagenet-1k. The improvement is not significant for the smaller architecture ViT-S. Cosub is gradually more effective when we increase the model size and the SD rate." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.511, + 0.21, + 0.529 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.539, + 0.47, + 0.63 + ], + "angle": 0, + "content": "We evaluate our approach on residual architectures [24, 25], since they are readily compatible with the stochastic dropout. We take as our main reference the vanilla Vision Transformer introduced by Dosovitskiy et al. [14]. However, as shown in our experiments, our approach is effective with all the residual architectures that we have considered." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.646, + 0.341, + 0.663 + ], + "angle": 0, + "content": "5.1. Baseline and training settings" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.672, + 0.47, + 0.779 + ], + "angle": 0, + "content": "We adopt the state-of-the-art training procedure proposed in the DeiT III paper [51] as baseline for transformers architectures and the training procedure from Wightman et al. [55] for the convnets. All hyper-parameters are identical except onImagenet-21k, where the hyper-parameter \\(\\tau\\) are adjusted depending on the training setting. We recapitulate the training hyper-parameters in Appendix A." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.78, + 0.471, + 0.903 + ], + "angle": 0, + "content": "We additionally adopt and measure the impact of LayerDecay for the fine-tuning when transferring from Imagenet-21k to Imagenet-1k. This method slightly boosts the performance, as discussed later in this section, and in Table 10. This method was adopted in multiple recent papers and in particular for fine-tuning of self-supervised approaches [5, 22], yet the contribution of this fine-tuning ingredient was not quantitatively measured." + }, + { + "type": "table", + "bbox": [ + 0.505, + 0.089, + 0.892, + 0.265 + ], + "angle": 0, + "content": "
ModelτOriginalBaseline+cosubΔ
Transformers
ViT-L [14]0.4576.584.584.9+0.4
CaiT-L24 [52]0.45-83.884.4+0.6
PiT-B [26]0.2582.083.884.1+0.3
XCiT-L12 [17]0.20-82.683.0+0.4
Swin-B [34]0.2083.582.983.3+0.4
Swin-L [34]0.45-80.884.0+3.2
Convnets
ResNet-50 [24]0.1076.280.280.3+0.1
ResNet-101 [24]0.2077.481.882.1+0.4
ResNet-152 [24]0.3078.382.483.1+0.7
RegNet-16GF [42]0.3080.482.983.8+0.9
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.267, + 0.894, + 0.33 + ], + "angle": 0, + "content": "Table 2. Benefit of cosub for different architectures trained from scratch on Imagenet-1k at resolution 224. We report top-1 acc. for the supervised baseline and cosub, as well as results reported in the corresponding papers when available (trained with different settings). We have adjusted the stochastic-depth drop-rate (SD) hyper-parameter for each architecture." + }, + { + "type": "table", + "bbox": [ + 0.504, + 0.346, + 0.891, + 0.384 + ], + "angle": 0, + "content": "
LossBCE-softBCE-hardCE-hard
Imagenet-val top-1 accuracy83.583.581.9
" + }, + { + "type": "table_caption", + "bbox": [ + 0.5, + 0.386, + 0.892, + 0.424 + ], + "angle": 0, + "content": "Table 3. Ablation on the loss for cosub with ViT-H trained at resolution \\(126\\times 126\\) on Imagenet-1k during 800 epochs. The training is inherited from DeiT-III, which also uses BCE when training with Imagenet-1k only." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.453, + 0.747, + 0.469 + ], + "angle": 0, + "content": "5.2. Empirical analysis of cosub" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.478, + 0.891, + 0.509 + ], + "angle": 0, + "content": "We perform various ablations on Imagenet-1k to analyse the impact of our training method on the learned networks." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.515, + 0.892, + 0.666 + ], + "angle": 0, + "content": "Performance for different model sizes and architectures. Table 1 provides the results obtained by the baseline and cosub when we vary the model size of vision transformers. The stochastic depth coefficient was optimized for the baseline and we keep it unchanged with cosub. As to be expected, our method is almost neutral for small models like ViT-S: \\(+0.1\\%\\) top-1 accuracy, which is about the standard deviation of measurements. The improvement is increasingly important for larger models, up to a significant improvement of \\(+0.6\\%\\) top-1 accuracy for ViT-H models." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.668, + 0.893, + 0.82 + ], + "angle": 0, + "content": "In Table 2, we show that our approach is beneficial with all architectures that we have tried. We report the results of the original paper, evaluate the performance with our baseline training, and measure the improvement brought by cosub. For almost all architectures, we observe a significant boost in performance. The exception is the ResNet-50, for which cosub improves the top-1 accuracy by only \\(+0.1\\%\\), similar to our observation with ViT-S. In Table B.1 in the appendix we present improved results obtained for multiple architectures pre-trained with cosub onImagenet-21k." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.826, + 0.894, + 0.901 + ], + "angle": 0, + "content": "Analysis of submodel performance. With cosub, we sample different subnetworks during training to perform the constraining. We analyse the impact of cosub on the accuracy of the sub networks themselves. In Figure 4 we consider the accuracy of submodels of different size of ViT-B and ViT-L" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "11705" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.084, + 0.089, + 0.462, + 0.187 + ], + "angle": 0, + "content": "
MethodViT-LViT-H
(a)supervised baseline [51]84.584.9
(b)KD85.385.3
(c)mean teacher84.483.4
(d)co-training82.683.1
(e)cosub84.985.5
(b) + (e)KD + cosub85.385.7
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.188, + 0.47, + 0.273 + ], + "angle": 0, + "content": "Table 4. Training strategies with distillation. We compare on Imagenet-1k at resolution \\(224 \\times 224\\) different approaches involving co- or self-training with distillation. KD, mean teacher and co-training use the same \\(\\lambda = 0.5\\) and same hyper-parameters as in Deit-III. Unlike cosub, the mean teacher (c) requires other hyper-parameters for EMA. We perform a small grid-search to adjust this parameters. Note (last row): our approach is complementary with KD, assuming a pre-trained teacher is available beforehand." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.292, + 0.47, + 0.324 + ], + "angle": 0, + "content": "
λ0.10.30.50.70.91.0
top1 accuracy79.0583.2783.5583.2083.0482.91
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.328, + 0.468, + 0.377 + ], + "angle": 0, + "content": "Table 5. Ablation of the parameter \\( \\lambda \\) controlling the weight of the co-distillation loss across submodels (Imagenet1k-val, top1-acc). Model ViT-H trained at resolution \\( {126} \\times {126} \\) onImagenet-1k during 800 epochs. \\( \\lambda = \\) 1.0 corresponds a supervised baseline and \\( \\lambda = {0.5} \\) corresponds to cosub." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.409, + 0.468, + 0.439 + ], + "angle": 0, + "content": "models. Cosub improves the accuracy of the whole population of sub-networks and, in particular, the target network." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.446, + 0.468, + 0.566 + ], + "angle": 0, + "content": "Loss formulation. In Table 3 we experiment with different losses for cosub. With Imagenet1k, DeiT-III training uses BCE instead of CE for the main loss. With cosub, BCE is more compatible with the loss of the baseline and, as to be expected, we also observe a better performance with BCE. We have done ablations using hard and soft targets for the cosub loss. The results are similar, therefore by default we keep soft-targets for the cosub loss." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.575, + 0.468, + 0.68 + ], + "angle": 0, + "content": "Alternative teacher/student. In Table 4 we report the results obtained with the different distillation or co-training approaches depicted in Figure 2. Other approaches are not effective off-the-shelf, except KD that requires a pre-trained teacher. Our approach is on par with KD (lower for ViT-L, better or ViT-H), and in the last row we show that it even provides a slight boost to combine KD with cosub." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.687, + 0.468, + 0.793 + ], + "angle": 0, + "content": "Impact of the cosub loss. The hyper-parameter \\(\\lambda\\) controls the tradeoff between the co-distillation loss and the cross-entropy classification loss. Setting \\(\\lambda = 1\\) means that we have a regular supervised training setting, except that (i) we double the batch size by duplicating the image after data augmentation, and (ii) stochastic depth selects different layers for each image copy." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.795, + 0.468, + 0.901 + ], + "angle": 0, + "content": "In Table 5, we measure the impact of \\(\\lambda\\), with all other hyper-parameters being fixed, for ViT-H trained at resolution \\(126 \\times 126\\) onImagenet-1k. We observe that the best ratio is to use an equal weighting of the cosub loss and the classic training loss. Using the cosub loss increases the performance by \\(0.6\\%\\) Top-1 accuracy onImagenet-val, which is the typical improvement that we observe for large models." + }, + { + "type": "table", + "bbox": [ + 0.51, + 0.089, + 0.885, + 0.259 + ], + "angle": 0, + "content": "
ModelτepochsBaseline+cosub
valv2valv2
CaiT-M120.2040083.272.983.773.5
80082.972.683.673.1
PiT-B0.2540083.873.684.174.1
80082.471.983.172.8
ViT-B0.2040083.172.683.273.1
80083.773.183.973.5
120083.372.8--
160083.373.4--
ViT-H0.6040084.875.385.075.8
80084.975.685.576.3
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.263, + 0.892, + 0.36 + ], + "angle": 0, + "content": "Table 6. We compare ViT models trained with and without cosub on ImageNet-1k only with different number of epochs at resolution \\(224 \\times 224\\). One can see that cosub is more effective for larger models yielding higher values of the SD hyper-parameter \\(\\tau\\). It avoids the early saturation or overfitting of the performance that we typically observe with the baseline when we increase the training time without re-adjusting hyper-parameters. See also Table 5 for a direct comparison with and without the co-distillation loss, and Table 7 for the corresponding training times per epoch." + }, + { + "type": "table", + "bbox": [ + 0.505, + 0.376, + 0.888, + 0.492 + ], + "angle": 0, + "content": "
Training methodmodelGPUs usedMemory peak (GB)Time (min) by epoch
DeiT-IIIViT-L3221.48
ViT-H6427.612
DeiT-III + ESDViT-L3215.19
ViT-H6415.211
cosub (with ESD)ViT-L3226.916
ViT-H6425.017
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.494, + 0.892, + 0.579 + ], + "angle": 0, + "content": "Table 7. Training times of different models trained at resolution \\(224\\times 224\\) with batch size 2048 on Imagenet-1k with DeiT-III and our approach. cosub uses our efficient stochastic depth (ESD), which amortizes the extra memory needed by cosub, especially for the largest models with high stochastic depth values (0.45 for ViT-L, and 0.6 for ViT-H). Timings are indicative and not representative of an optimized selection of the batch size, see Appendix C for measurements with adjusted parameters." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.601, + 0.892, + 0.721 + ], + "angle": 0, + "content": "Number of training epochs. In Table 6 we compare results on Imagenet-1k-val and Imagenet-v2 different for architectures trained with and without cosub on Imagenet-1k only at resolution \\(224 \\times 224\\) with different number of epochs. We observe less overfitting with cosub and longer training schedule. In particular, with bigger architecture like ViT-H, we observe continuous improvement with a longer schedule where the baseline saturates." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.729, + 0.892, + 0.804 + ], + "angle": 0, + "content": "Training time. In Table 7 we compare the training costs of cosub and DeiT-III. Thanks to our efficient stochastic depth formulation we maintain a similar memory peak during training. For bigger architectures the gap in training speed between cosub and the baseline is decreasing." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.81, + 0.892, + 0.899 + ], + "angle": 0, + "content": "Resolution. In Table 8 we compare different ViT architectures trained with and without cosub at different resolutions onImagenet-1k. We fine-tune during 20 epochs at resolution \\(224 \\times 224\\) before evaluation at this resolution. We observe that cosub gives significant improvements across the different resolutions and models." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.956 + ], + "angle": 0, + "content": "11706" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.086, + 0.089, + 0.462, + 0.246 + ], + "angle": 0, + "content": "
ModelResolutionDeit-III Baseline+cosub
valv2valv2
ViT-B128 × 12883.573.483.874.0
192 × 19283.873.684.174.0
224 × 22483.773.183.973.5
ViT-L128 × 12884.574.785.175.5
192 × 19284.975.185.275.7
224 × 22484.575.084.975.6
ViT-H126 × 12685.175.685.676.4
182 × 18285.175.985.776.6
224 × 22484.975.685.576.3
" + }, + { + "type": "table_caption", + "bbox": [ + 0.078, + 0.25, + 0.47, + 0.288 + ], + "angle": 0, + "content": "Table 8. Imagenet-1k val and v2 top-1 accuracy of ViT models trained with and without cosub for 800 epochs on Imagenet-1k at different resolutions, followed by finetuning for 20 epochs at resolution \\(224 \\times 224\\)." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.295, + 0.469, + 0.41 + ], + "angle": 0, + "content": "
resol. →112224336448112224336448
↓ modelImagenet-valImagenet-v2
ViT-S78.083.184.685.266.673.775.176.3
ViT-M80.685.086.086.369.676.076.877.2
ViT-B82.886.386.987.472.177.077.978.3
ViT-L85.487.588.188.375.779.179.880.0
ViT-H86.288.0--76.979.6--
ViT-g86.5---77.3---
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.414, + 0.47, + 0.501 + ], + "angle": 0, + "content": "Table 9. Performance of models at different resolutions. We report the results obtained on Imagenet-val by models of different sizes pre-trained with cosub on Imagenet-21k and fine-tuned on Imagenet-1k. Training schedule: 270 epochs except ViT-g (90 epochs). Except for ViT-S, the results at resolution 336 and 448 were pre-trained on Imagenet-21k at resolution 224 for efficiency reasons. We have not fine-tuned ViT-H and ViT-g at large resolutions since these models are computationally expensive." + }, + { + "type": "table", + "bbox": [ + 0.084, + 0.513, + 0.462, + 0.614 + ], + "angle": 0, + "content": "
MethodLong TrainingLayer DecayViT-SModel ViT-BViT-L
baselineXX82.685.286.8
XX82.585.887.4
cosubX✓82.786.087.5
✓X82.886.087.4
✓✓83.186.387.5
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.618, + 0.47, + 0.656 + ], + "angle": 0, + "content": "Table 10. We measure the impact of layer-decay during the finetuning on Imagenet-1k for models pre-trained on Imagenet-21k with cosub during 90 epochs (default) and 270 epochs (long training)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.682, + 0.47, + 0.759 + ], + "angle": 0, + "content": "Imagenet-21k impact of layer-decay. In Table 10 we compare different number of epochs for the pre-training on Imagenet-21k and the finetuning on Imagenet-1k with and without layer-decay. We observe that both layer-decay and long training bring improvements with cosub." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.767, + 0.399, + 0.784 + ], + "angle": 0, + "content": "5.3. Comparisons with the state of the art" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.795, + 0.469, + 0.903 + ], + "angle": 0, + "content": "Imagenet-1k. In Table 11 we compare architectures trained with cosub with state-of-the-art results from the literature for this architecture. We observe that architectures trained with cosub are very competitive. For instance, for ResNet-152, RegNetY-16GF, PiT-B we improve the best results reported in the literature by more than \\(0.9\\%\\) top-1 accuracy on Imagenet-1k-val." + }, + { + "type": "table", + "bbox": [ + 0.505, + 0.089, + 0.895, + 0.432 + ], + "angle": 0, + "content": "
Modelnb params (×106)throughput (im/s)FLOPs (×109)Peak Mem (MB)Top-1 Acc.v2 Acc.
“Traditional” ConvNets
ResNet-50 [24, 55]25.625874.1218280.468.7
ResNet-101 [24, 55]44.515867.9226981.570.3
ResNet-101 - cosub44.515867.9226982.170.8
ResNet-152 [24, 55]60.2112211.6235982.070.6
ResNet-152 - cosub60.2112211.6235983.172.1
RegNetY-8GF [42, 55]39.211588.0393982.271.1
RegNetY-16GF [42, 50]83.671416.0520482.972.4
RegNetY-16GF - cosub83.671416.0520483.872.8
Vision Transformers derivatives
Swin-T [34]28.311094.5334581.369.5
Swin-B [34]87.853215.4469583.5-
PiT-S [26]23.518092.9329380.9-
PiT-S - cosub23.518092.9329381.369.7
PiT-B [26]73.861512.5756482.0-
PiT-B - cosub73.861512.5756484.174.1
Vanilla Vision Transformers
ViT-S [50]22.018914.698779.868.1
ViT-S - Deit III [51]22.018914.698781.470.5
ViT-S - cosub22.018914.698781.570.8
ViT-B [14]86.683117.5207877.9-
ViT-B - DeiT [50]86.683117.5207881.871.5
ViT-B - DeiT/distilled86.683117.5207883.473.2
ViT-B - DeiT III [51]86.683117.5207883.873.6
ViT-B - cosub86.683117.5207884.274.2
ViT-L - Deit III [51]304.427761.6378984.975.1
ViT-L - cosub304.427761.6378985.375.5
ViT-H - Deit III [51]632.1112167.4698485.275.9
ViT-H - cosub632.1112167.4698485.776.6
" + }, + { + "type": "table_caption", + "bbox": [ + 0.498, + 0.437, + 0.894, + 0.523 + ], + "angle": 0, + "content": "Table 11. Classification with Imagenet1k training. We compare with models trained on Imagenet-1k only at resolution \\(224 \\times 224\\) without self-supervised pre-training (see supp. material for a comparison). We report Top-1 accuracy on Imagenet-1k-val and Imagenet-v2 with different measures of complexity: throughput, FLOPs, number of parameters and peak memory usage. The throughput and peak memory are measured on a single V100-32GB GPU with batch size fixed to 256 and mixed precision." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.55, + 0.892, + 0.627 + ], + "angle": 0, + "content": "Imagenet-21k. In Table 12 we compare ViT models pretrained with cosub on Imagenet-21k and finetuned with cosub on Imagenet-1k with other architectures and our baseline DeiT-III. Our results with vanilla ViT outperform the baseline and are competitive with recent architectures." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.632, + 0.893, + 0.768 + ], + "angle": 0, + "content": "Overfitting analysis. As recommended in prior works [51, 55] we perform an overfitting analysis. We evaluate our models trained with codist onImagenet-1k val andImagenet-v2 [43]. The results are reported in Figure 5. For ViT, we observe that cosub does not overfit more than the DeiT-III baseline [51]. Our results with other architectures in Table 12 concur with that observation: our results are comparatively stronger onImagenet-v2 than those reported in the literature for the exact same models." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.781, + 0.679, + 0.796 + ], + "angle": 0, + "content": "5.4. Downstream tasks" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.81, + 0.894, + 0.903 + ], + "angle": 0, + "content": "Semantic segmentation. First we evaluate our ViT models pre-trained onImagenet with cosub for semantic segmentation on the ADE20k dataset [63]. ADE20k consists of 20k training and 5k validation images with labels over 150 categories. We adopt the training schedule from Swin: 160k iterations with UperNet [57]. At test time we evaluate with" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "11707" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.084, + 0.089, + 0.462, + 0.436 + ], + "angle": 0, + "content": "
Architecturenb params (×106)throughput (im/s)FLOPs (×109)peak mem (MB)top1 valacc. v2
Convnets
EfficientNetV2-M [48]54.131225.0712786.275.9
EfficientNetV2-L [48]118.517953.0954086.876.9
ResNet-152 [24, 55]60.2112211.6235982.070.6
ResNet-152 - cosub60.2112211.6235983.173.1
RegnetY16GF - cosub83.671416.0520484.274.7
ConvNeXt-S- cosub50.27838.7221885.276.0
ConvNeXt-B [35]88.656315.4302985.875.6
ConvNeXt-B - cosub88.656315.4302985.876.9
ConvNeXt-L [35]197.834434.4486586.676.6
ConvNeXt-XL [35]350.224160.9695187.077.0
Transformers variations
Swin-B [34]87.853215.4469585.274.6
Swin-B - cosub87.853215.4469586.277.2
Swin-L [34]196.533734.5735086.376.3
Swin-L - cosub196.533734.5735087.178.1
PiT-B - cosub [26]73.861512.5756485.876.8
XCiT-S12 - cosub [17]26.213734.9133084.274.9
XCiT-M24 - cosub [17]84.455316.2201086.578.0
XCiT-L24 - cosub [17]189.033436.1331587.277.8
Vanilla Vision Transformers [14, 51]
ViT-S - cosub22.018914.698783.173.7
ViT-M - cosub39.013078.0132285.076.0
ViT-B - DeiT-III86.683117.6207885.776.5
ViT-B - cosub86.683117.6207886.377.0
ViT-L - DeiT-III304.427761.6378987.078.6
ViT-L - cosub304.427761.6378987.579.1
ViT-H - DeiT-III632.1112167.4698487.279.2
ViT-H - cosub632.1112167.4698488.080.0
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.077, + 0.44, + 0.47, + 0.573 + ], + "angle": 0, + "content": "Table 12. Classification with ImageNet-21k pretraining. We report top-1 accuracy on the validation set of Imagenet1k and Imagenet-V2 with different measures of complexity. The peak memory usage is measured on a single V100-32GB GPU with batch size fixed to 256 and mixed precision. For Swin-L the memory peak is an estimation since we decreased the batch size to 128 to fit in memory. All models are evaluated at resolution 224 except EfficientNetV2 that use resolution 480. ViT are pretrained with a \\(\\times 3\\) schedule, comparable to that used in the best DeiT-III baseline (270 vs. 240 epochs). All other cosub models are pretrained during 90 epochs on Imagenet21k, with 50 epochs of fine-tuning. The \\(\\tau\\) hyper-parameter is set per model based on prior choices or best guess based on model size." + }, + { + "type": "image", + "bbox": [ + 0.078, + 0.589, + 0.266, + 0.739 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.277, + 0.589, + 0.451, + 0.739 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.743, + 0.47, + 0.804 + ], + "angle": 0, + "content": "Figure 5. Overfitting measurement: top-1 accuracy on Imagenet-val vs. Imagenet-v2 for models in Tables 11 and 12 pre-trained on Imagenet-1k and Imagenet-21k, respectively. Our cosub ViTs (plain lines and points) do not overfit more than DeiT-III [51] overall. Our Imagenet-21k results for Swin and Convnext generalize much better on v2 than the original models." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.471, + 0.903 + ], + "angle": 0, + "content": "a single scale and multi-scale. Our UperNet implementation is the same as in DeiT III [51]. Our results are reported in Table 13. We observe that vanilla ViTs trained with cos-sub outperform our baseline DeiT-III but also have better FLOPs-accuracy trade-offs than recent architectures." + }, + { + "type": "table", + "bbox": [ + 0.507, + 0.089, + 0.888, + 0.352 + ], + "angle": 0, + "content": "
Backbone#params (×106)FLOPs (×109)Single-scale mIoUMulti-scale mIoU
DeiT-S52.01099-44.0
Swin-T59.994544.546.1
ViT-S – DeiT-III41.758845.646.8
ViT-S – cosub41.758847.048.0
PatchConvNet-B60140.6125848.148.6
PatchConvNet-B120229.8155049.450.3
MAE ViT-B127.7128348.1-
Swin-B121.0118848.149.7
ViT-B – DeiT-III127.7128349.350.2
ViT-B – cosub127.7128349.349.9
ViT-L – DeiT-III353.6223151.552.0
ViT-L – cosub353.6223152.553.1
PatchConvNet-B60†140.6125850.551.1
Swin-B† (640 × 640)121.0184150.051.6
ViT-B – DeiT-III†127.7128353.454.1
ViT-B – cosub†127.7128353.754.7
PatchConvNet-L120†383.7208652.252.9
Swin-L† (640 × 640)234.03230-53.5
ViT-L – DeiT-III†353.6223154.655.6
ViT-L – cosub†353.6223155.756.3
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.5, + 0.359, + 0.892, + 0.42 + ], + "angle": 0, + "content": "Table 13. ADE20k semantic segmentation performance using UperNet [57], in comparable setting as prior works [13, 17, 34]. All models are pretrained on Imagenet-1k, except bottom models identified with \\(\\dagger\\), which are pre-trained on Imagenet-21k. By default the finetuning resolution on ADE20k is \\(512 \\times 512\\) except when mentioned otherwise (for Swin)." + }, + { + "type": "table", + "bbox": [ + 0.508, + 0.434, + 0.885, + 0.524 + ], + "angle": 0, + "content": "
ModelCifar-10Cifar-100FlowersCarsiNat-18iNat-19
ViT-S - DeiT-III98.990.696.489.967.172.7
ViT-B - DeiT-III99.392.598.693.473.678.0
ViT-L - DeiT-III99.393.498.994.575.679.3
ViT-S - cosub99.191.797.493.070.175.6
ViT-B - cosub99.192.698.493.574.178.1
ViT-L - cosub99.493.598.894.576.280.1
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.499, + 0.526, + 0.892, + 0.587 + ], + "angle": 0, + "content": "Table 14. ViT models pre-trained with cosub or DeiT-III on Imagenet-1k and finetuned on six different target datasets. We note that for small datasets (CIFAR, Flowers, and Cars) our approach is useful for small models but neutral for larger models. The gains are more significant when transferring to the larger iNaturalist-18 and iNaturalist-19 datasets." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.606, + 0.892, + 0.727 + ], + "angle": 0, + "content": "Transfer learning. We now measure how the performance improvements observed with cosub translate to other classification problems. For this purpose, we performed transfer learning on the six different datasets used in DeiT-III. Our results are reported Table 14. Our pre-trained and fine-tuned models with cosub generally improve the baseline. The gains are overall more significant on the more challenging datasets like iNaturalist 2018 and iNaturalist 2019." + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.74, + 0.619, + 0.755 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Co-training submodels (cosub) is an effective way to improve existing deep residual networks. It is straightforward to implement, just involving a few lines of code. It does not need a pre-trained teacher, and it only maintains a single set of weights for the model. Extensive experimental results on image classification, transfer learning and semantic segmentation show that cosub is overall extremely effective. It works off-the-shelf and improves the state of the art for various network architectures, including convnets like Regnet." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "11708" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.468, + 0.156 + ], + "angle": 0, + "content": "[1] Samira Abnar, Mostafa Dehghani, and Willem Zuidema. Transferring inductive biases through knowledge distillation. arXiv preprint arXiv:2006.00555, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.157, + 0.469, + 0.212 + ], + "angle": 0, + "content": "[2] Rohan Anil, Gabriel Pereyra, Alexandre Passos, Robert Ormandi, George E Dahl, and Geoffrey E Hinton. Large scale distributed neural network training through online distillation. arXiv preprint arXiv:1804.03235, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.213, + 0.469, + 0.284 + ], + "angle": 0, + "content": "[3] Mahmoud Assran, Mathilde Caron, Ishan Misra, Piotr Bojanowski, Florian Bordes, Pascal Vincent, Armand Joulin, Mike Rabbat, and Nicolas Ballas. Masked siamese networks for label-efficient learning. In European Conference on Computer Vision, pages 456-473. Springer, 2022. I" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.284, + 0.469, + 0.31 + ], + "angle": 0, + "content": "[4] L. Ba and R. Caruana. Do deep nets really need to be deep? In NeurIPS, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.312, + 0.469, + 0.352 + ], + "angle": 0, + "content": "[5] Hangbo Bao, Li Dong, and Furu Wei. Beit: Bert pre-training of image transformers. arXiv preprint arXiv:2106.08254, 2021. 1, 5, I, II" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.354, + 0.469, + 0.41 + ], + "angle": 0, + "content": "[6] Irwan Bello, W. Fedus, Xianzhi Du, E. D. Cubuk, A. Srinivas, Tsung-Yi Lin, Jonathon Shlens, and Barret Zoph. Revisiting ResNets: Improved training and scaling strategies. arXiv preprint arXiv:2103.07579, 2021. II" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.411, + 0.469, + 0.451 + ], + "angle": 0, + "content": "[7] H. Bourlard and Y. Kamp. Auto-association by multilayer perceptrons and singular value decomposition. Biological Cybernetics, 59:291-294, 1988. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.453, + 0.468, + 0.479 + ], + "angle": 0, + "content": "[8] Leo Breiman. Bagging predictors. Machine learning, 24(2):123-140, 1996. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.481, + 0.469, + 0.536 + ], + "angle": 0, + "content": "[9] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. arXiv preprint arXiv:2104.14294, 2021. 1, 2, II" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.538, + 0.469, + 0.592 + ], + "angle": 0, + "content": "[10] Kevin Clark, Minh-Thang Luong, Quoc V Le, and Christopher D Manning. Electra: Pre-training text encoders as discriminators rather than generators. arXiv preprint arXiv:2003.10555, 2020. I" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.594, + 0.469, + 0.648 + ], + "angle": 0, + "content": "[11] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In Conference on Computer Vision and Pattern Recognition, pages 248–255, 2009. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.65, + 0.469, + 0.69 + ], + "angle": 0, + "content": "[12] Carl Doersch, Ankush Gupta, and Andrew Zisserman. Crosstransformers: spatially-aware few-shot transfer. arXiv preprint arXiv:2007.11498, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.692, + 0.469, + 0.759 + ], + "angle": 0, + "content": "[13] Xiaoyi Dong, Jianmin Bao, Dongdong Chen, Weiming Zhang, Nenghai Yu, Lu Yuan, Dong Chen, and Baining Guo. Cswin transformer: A general vision transformer backbone with cross-shaped windows. arXiv preprint arXiv:2107.00652, 2021. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.762, + 0.469, + 0.844 + ], + "angle": 0, + "content": "[14] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations, 2021. 5, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.846, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[15] Alaaeldin El-Nouby, Gautier Izacard, Hugo Touvron, Ivan Laptev, Hervé Jegou, and Edouard Grave. Are large-scale datasets necessary for self-supervised pre-training? arXiv preprint arXiv:2112.10740, 2021. 1" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.469, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.134 + ], + "angle": 0, + "content": "[16] Alaaeldin El-Nouby, Natalia Neverova, Ivan Laptev, and Hervé Jégou. Training vision transformers for image retrieval. arXiv preprint arXiv:2102.05644, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.136, + 0.892, + 0.204 + ], + "angle": 0, + "content": "[17] Alaaeldin El-Nouby, Hugo Touvron, Mathilde Caron, Piotr Bojanowski, Matthijs Douze, Armand Joulin, Ivan Laptev, Natalia Neverova, Gabriel Synnaeve, Jakob Verbeek, et al. Xcit: Cross-covariance image transformers. arXiv preprint arXiv:2106.09681, 2021. 5, 8, II" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.206, + 0.892, + 0.247 + ], + "angle": 0, + "content": "[18] Thomas Elsken, Jan Hendrik Metzen, and Frank Hutter. Neural architecture search: A survey. Journal of Machine Learning Research, 20(1):1997-2017, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.249, + 0.892, + 0.29 + ], + "angle": 0, + "content": "[19] Dumitru Erhan, Aaron Courville, Yoshua Bengio, and Pascal Vincent. Why does unsupervised pre-training help deep learning? In AISTATS, 2010. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.292, + 0.892, + 0.333 + ], + "angle": 0, + "content": "[20] Angela Fan, Edouard Grave, and Armand Joulin. Reducing transformer depth on demand with structured dropout. arXiv preprint arXiv:1909.11556, 2019. ICLR 2020. 2, 3, 4, IV" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.335, + 0.892, + 0.389 + ], + "angle": 0, + "content": "[21] Rohit Girdhar, Alaaeldin El-Nouby, Mannat Singh, Kalyan Vasudev Alwala, Armand Joulin, and Ishan Misra. Omnimea: Single model masked pretraining on images and videos. arXiv preprint arXiv:2206.08356, 2022. I" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.391, + 0.892, + 0.445 + ], + "angle": 0, + "content": "[22] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Doll'ar, and Ross B. Girshick. Masked autoencoders are scalable vision learners. arXiv preprint arXiv:2111.06377, 2021. 1, 5, I, II" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.448, + 0.892, + 0.502 + ], + "angle": 0, + "content": "[23] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. arXiv preprint arXiv:1911.05722, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.504, + 0.892, + 0.559 + ], + "angle": 0, + "content": "[24] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Conference on Computer Vision and Pattern Recognition, 2016. 1, 5, 7, 8, II" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.561, + 0.892, + 0.602 + ], + "angle": 0, + "content": "[25] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Identity mappings in deep residual networks. arXiv preprint arXiv:1603.05027, 2016. 1, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.604, + 0.892, + 0.658 + ], + "angle": 0, + "content": "[26] Byeongho Heo, Sangdoo Yun, Dongyoon Han, Sanghyuk Chun, Junsuk Choe, and Seong Joon Oh. Rethinking spatial dimensions of vision transformers. arXiv preprint arXiv:2103.16302, 2021. 5, 7, 8, II" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.66, + 0.892, + 0.7 + ], + "angle": 0, + "content": "[27] G. Hinton and R. Zemel. Autoencoders, minimum description length and Helmholtz free energy. In NeurIPS, 1993. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.703, + 0.892, + 0.744 + ], + "angle": 0, + "content": "[28] Geoffrey E. Hinton, Oriol Vinyals, and J. Dean. Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531, 2015. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.746, + 0.892, + 0.801 + ], + "angle": 0, + "content": "[29] Grant Van Horn, Oisin Mac Aodha, Yang Song, Alexander Shepard, Hartwig Adam, Pietro Perona, and Serge J. Belongie. The iNaturalist species classification and detection dataset. arXiv preprint arXiv:1707.06642, 2017. II" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.803, + 0.892, + 0.857 + ], + "angle": 0, + "content": "[30] Grant Van Horn, Oisin Mac Aodha, Yang Song, Alexander Shepard, Hartwig Adam, Pietro Perona, and Serge J. Belongie. The inaturalist challenge 2018 dataset. arXiv preprint arXiv:1707.06642, 2018. II" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.859, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[31] Gao Huang, Yu Sun, Zhuang Liu, Daniel Sedra, and Kilian Q. Weinberger. Deep networks with stochastic depth. In European Conference on Computer Vision, 2016. 2, 4, I" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "11709" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.147 + ], + "angle": 0, + "content": "[32] Jonathan Krause, Michael Stark, Jia Deng, and Li Fei-Fei. 3d object representations for fine-grained categorization. In IEEE Workshop on 3D Representation and Recognition, 2013. II" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.149, + 0.469, + 0.175 + ], + "angle": 0, + "content": "[33] Alex Krizhevsky. Learning multiple layers of features from tiny images. Technical report, CIFAR, 2009. II" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.177, + 0.469, + 0.232 + ], + "angle": 0, + "content": "[34] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. arXiv preprint arXiv:2103.14030, 2021. 5, 7, 8, II" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.233, + 0.469, + 0.273 + ], + "angle": 0, + "content": "[35] Zhuang Liu, Hanzi Mao, Chao-Yuan Wu, Christoph Feichtenhofer, Trevor Darrell, and Saining Xie. A convnet for the 2020s. arXiv preprint arXiv:2201.03545, 2022. 8, II" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.274, + 0.469, + 0.301 + ], + "angle": 0, + "content": "[36] I. Loshchilov and F. Hutter. Fixing weight decay regularization in adam. arXiv preprint arXiv:1711.05101, 2017. I" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.302, + 0.469, + 0.342 + ], + "angle": 0, + "content": "[37] G. J. McLachlan and D. Peel. Finite Mixture Models. Wiley Series in Probability and Statistics. John Wiley & Sons, 2000. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.343, + 0.469, + 0.399 + ], + "angle": 0, + "content": "[38] M-E. Nilsback and A. Zisserman. Automated flower classification over a large number of classes. In Proceedings of the Indian Conference on Computer Vision, Graphics and Image Processing, 2008. II" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.401, + 0.469, + 0.455 + ], + "angle": 0, + "content": "[39] Maxime Oquab, Leon Bottou, Ivan Laptev, and Josef Sivic. Learning and transferring mid-level image representations using convolutional neural networks. In Conference on Computer Vision and Pattern Recognition, 2014. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.457, + 0.469, + 0.537 + ], + "angle": 0, + "content": "[40] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. In Advances in neural information processing systems, pages 8026-8037, 2019. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.539, + 0.469, + 0.593 + ], + "angle": 0, + "content": "[41] Ilija Radosavovic, Raj Prateek Kosaraju, Ross Girshick, Kaiming He, and Piotr Dólar. Designing network design spaces. In Conference on Computer Vision and Pattern Recognition, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.595, + 0.469, + 0.649 + ], + "angle": 0, + "content": "[42] Ilija Radosavovic, Raj Prateek Kosaraju, Ross B. Girshick, Kaiming He, and Piotr Dollar. Designing network design spaces. Conference on Computer Vision and Pattern Recognition, 2020. 5, 7, II" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.651, + 0.469, + 0.691 + ], + "angle": 0, + "content": "[43] B. Recht, Rebecca Roelofs, L. Schmidt, and V. Shankar. Do ImageNet classifiers generalize to ImageNet? In International Conference on Machine Learning, 2019. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.692, + 0.469, + 0.718 + ], + "angle": 0, + "content": "[44] Adria Ruiz and Jakob Verbeek. Anytime inference with distilled hierarchical neural ensembles. In AAAI, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.72, + 0.469, + 0.747 + ], + "angle": 0, + "content": "[45] Guocong Song and Wei Chai. Collaborative learning for deep neural networks. NeurIPS, 31, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.749, + 0.469, + 0.803 + ], + "angle": 0, + "content": "[46] Nitish Srivastava, Geoffrey Hinton, Alex Krizhevsky, Ilya Sutskever, and Ruslan Salakhutdinov. Dropout: A simple way to prevent neural networks from overfitting. In Journal of Machine Learning Research, 2014. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.805, + 0.469, + 0.858 + ], + "angle": 0, + "content": "[47] Christian Szegedy, Vincent Vanhoucke, Sergey Ioffe, Jonathon Shlens, and Zbigniew Wojna. Rethinking the inception architecture for computer vision. Conference on Computer Vision and Pattern Recognition, 2016. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.86, + 0.469, + 0.9 + ], + "angle": 0, + "content": "[48] Mingxing Tan and Quoc V. Le. Efficientnetv2: Smaller models and faster training. In International Conference on Machine Learning, 2021. 8" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.134 + ], + "angle": 0, + "content": "[49] Antti Tarvainen and Harri Valpola. Mean teachers are better role models: Weight-averaged consistency targets improve semi-supervised deep learning results. Neurips, 30, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.136, + 0.892, + 0.189 + ], + "angle": 0, + "content": "[50] Hugo Touvron, M. Cord, M. Douze, F. Massa, Alexandre Sablayrolles, and H. Jégou. Training data-efficient image transformers & distillation through attention. International Conference on Machine Learning, 2021. 2, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.191, + 0.892, + 0.23 + ], + "angle": 0, + "content": "[51] Hugo Touvron, Matthieu Cord, and Hervé Jégou. Deit III: Revenge of the vit. arXiv preprint arXiv:2204.07118, 2022. 5, 6, 7, 8, I, II" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.232, + 0.892, + 0.285 + ], + "angle": 0, + "content": "[52] Hugo Touvron, Matthieu Cord, Alexandre Sablayrolles, Gabriel Synnaeve, and Hervé Jégou. Going deeper with image transformers. International Conference on Computer Vision, 2021. 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.287, + 0.892, + 0.342 + ], + "angle": 0, + "content": "[53] Longhui Wei, An Xiao, Lingxi Xie, X. Chen, Xiaopeng Zhang, and Q. Tian. Circumventing outliers of autoaugment with knowledge distillation. In European Conference on Computer Vision, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.343, + 0.892, + 0.383 + ], + "angle": 0, + "content": "[54] Ross Wightman. Pytorch image models. https://github.com/rwrightman/pytorch-image-models, 2019.4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.385, + 0.892, + 0.424 + ], + "angle": 0, + "content": "[55] Ross Wightman, Hugo Touvron, and Hervé Jégou. Resnet strikes back: An improved training procedure in timm. arXiv preprint arXiv:2110.00476, 2021. 5, 7, 8, II" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.426, + 0.892, + 0.521 + ], + "angle": 0, + "content": "[56] Mitchell Wortsman, Gabriel Ilharco, Samir Ya Gadre, Rebecca Roelofs, Raphael Gontijo-Lopes, Ari S Morcos, Hongseok Namkoong, Ali Farhadi, Yair Carmon, Simon Kornblith, et al. Model soups: averaging weights of multiple fine-tuned models improves accuracy without increasing inference time. In International Conference on Machine Learning, 2022. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.523, + 0.892, + 0.576 + ], + "angle": 0, + "content": "[57] Tete Xiao, Yingcheng Liu, Bolei Zhou, Yuning Jiang, and Jian Sun. Unified perceptual parsing for scene understanding. In European Conference on Computer Vision, 2018. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.578, + 0.892, + 0.619 + ], + "angle": 0, + "content": "[58] Jason Yosinski, Jeff Clune, Yoshua Bengio, and Hod Lipson. How transferable are features in deep neural networks? arXiv preprint arXiv:1411.1792, 2014. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.621, + 0.892, + 0.66 + ], + "angle": 0, + "content": "[59] L. Yuan, F. Tay, G. Li, T. Wang, and Jiashi Feng. Revisit knowledge distillation: a teacher-free framework. Conference on Computer Vision and Pattern Recognition, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.661, + 0.892, + 0.716 + ], + "angle": 0, + "content": "[60] Sangdoo Yun, Dongyoon Han, Seong Joon Oh, Sanghyuk Chun, Junsuk Choe, and Youngjoon Yoo. CutMix: Regularization strategy to train strong classifiers with localizable features. arXiv preprint arXiv:1905.04899, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.717, + 0.892, + 0.757 + ], + "angle": 0, + "content": "[61] Hongyi Zhang, Moustapha Cisse, Yann N. Dauphin, and David Lopez-Paz. mixup: Beyond empirical risk minimization. arXiv preprint arXiv:1710.09412, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.759, + 0.892, + 0.812 + ], + "angle": 0, + "content": "[62] Ying Zhang, Tao Xiang, Timothy M Hospedales, and Huchuan Lu. Deep mutual learning. In Conference on Computer Vision and Pattern Recognition, pages 4320-4328, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.814, + 0.892, + 0.868 + ], + "angle": 0, + "content": "[63] Bolei Zhou, Hang Zhao, Xavier Puig, Sanja Fidler, Adela Barriuso, and Antonio Torralba. Scene parsing through ade20k dataset. Conference on Computer Vision and Pattern Recognition, 2017. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.87, + 0.892, + 0.896 + ], + "angle": 0, + "content": "[64] Xiatian Zhu, Shaogang Gong, et al. Knowledge distillation by on-the-fly native ensemble. Neurips, 31, 2018. 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.896 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "11710" + } + ] +] \ No newline at end of file diff --git a/2023/Co-Training 2L Submodels for Visual Recognition/51b2445a-3487-4e2d-824b-ee9b871d1274_origin.pdf b/2023/Co-Training 2L Submodels for Visual Recognition/51b2445a-3487-4e2d-824b-ee9b871d1274_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..ddac947c7142810998d27aa6fa2364c8f1ba53ad --- /dev/null +++ b/2023/Co-Training 2L Submodels for Visual Recognition/51b2445a-3487-4e2d-824b-ee9b871d1274_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:67315637cdc1d057855c834526548d2ada8a351f2fa3440b96da6faee0dd062f +size 823832 diff --git a/2023/Co-Training 2L Submodels for Visual Recognition/full.md b/2023/Co-Training 2L Submodels for Visual Recognition/full.md new file mode 100644 index 0000000000000000000000000000000000000000..35ab4dc8cb2aa17952ea678e40d7c64d80301061 --- /dev/null +++ b/2023/Co-Training 2L Submodels for Visual Recognition/full.md @@ -0,0 +1,329 @@ +# Co-training $2^{L}$ Submodels for Visual Recognition + +Hugo Touvron\*,† + +Piotr Bojanowski* + +Matthieu Cord† + +Jakob Verbeek* + +Maxime Oquab\* + +Hervé Jégou* + +*Meta AI / FAIR Paris + +†Sorbonne University + +# Abstract + +We introduce submodel co-training, a regularization method related to co-training, self-distillation and stochastic depth. Given a neural network to be trained, for each sample we implicitly instantiate two altered networks, "submodels", with stochastic depth: we activate only a subset of the layers. Each network serves as a soft teacher to the other, by providing a loss that complements the regular loss provided by the one-hot label. Our approach, dubbed "cosub", uses a single set of weights, and does not involve a pre-trained external model or temporal averaging. + +Experimentally, we show that submodel co-training is effective to train backbones for recognition tasks such as image classification and semantic segmentation. Our approach is compatible with multiple architectures, including RegNet, ViT, PiT, XCiT, Swin and ConvNext. Our training strategy improves their results in comparable settings. For instance, a ViT-B pretrained with cosub on ImageNet-21k obtains $87.4\%$ top-1 acc. @448 on ImageNet-val. + +# 1. Introduction + +Although the fundamental ideas of deep trainable neural networks have been around for decades, only recently have barriers been removed to allow breakthroughs in successfully training deep neural architectures in practice. Many of these barriers are related to non-convex optimization in one way or another, which is central to the success of modern neural networks. The optimization challenges have been addressed from multiple angles in the literature. First, modern architectures are designed to facilitate the optimization of very deep networks. An exceptionally successful design principle is using residual connections [24, 25]. Although this does not change the expressiveness of the functions that the network can implement, the improved gradient flow alleviates, to some extent, the difficulties of optimizing very deep networks. Another key element to the optimization is the importance of data, revealed by the step-change in visual recognition performance resulting from the ImageNet dataset [11], and the popularization of transfer learning with pre-training on large datasets [39, 58]. + +![](images/6605072b50b4cf03d4e74edaf5c0b4bb06c32e1a462ecca593c4342f23471d3c.jpg) +Figure 1. Co-training of submodels (cosub): for each image, two sub-models are sampled by randomly dropping layers of the full model. The training signal for each submodel mixes the cross-entropy loss from the image label with a self-distillation loss obtained from the other submodel. + +However, even when (pre-)trained with millions of images, recent deep networks with millions if not billions of parameters, are still heavily overparameterized. Traditional regularization like weight decay, dropout [46], or label smoothing [47] are limited in their ability to address this issue. Data-augmentation strategies, including those mixing different images like Mixup [61] and CutMix [60], have proven to provide a complementary data-driven form of regularization. More recently, multiple works propose to resort to self-supervised pre-training. These approaches rely on a proxy objective that generally provides more supervision signal than the one available from labels, like in recent (masked) auto-encoders [5,16,22], which were popular in the early deep learning literature [7,19,27]. Similarly, contrastive approaches [23] or self-distillation [9] provide a richer supervision less prone to supervision collapse [12]. Overall, self-supervised learning makes it possible to learn larger models with less data, possibly reducing the need of a pre-training stage [15]. + +Distillation is a complementary approach to improve optimization. Distillation techniques were originally developed to transfer knowledge from a teacher model to a student model [4, 28], allowing the student to improve over learning from the data directly. In contrast to traditional distillation, co-distillation does not require pre-training a (strong) teacher. Instead, a pool of models supervise each other. Practically, it faces several limitations, including the difficulty of jointly training more than two students for complexity reasons, as it involves duplicating the weights. + +In this paper, we propose a practical way to enable cotraining for a very large number of students. We consider a single target model to be trained, and we instantiate two submodels on-the-fly, simply by layerwise dropout [20,31]. This gives us two neural networks through which we can backpropagate to the shared parameters of the target model. In addition to the regular training loss, each submodel serves as a teacher to the other, which provides an additional supervision signal ensuring the consistency across the submodels. Our approach is illustrated in Figure 1: the parameter $\lambda$ controls the importance of the co-training loss compared to the label loss, and our experiments show that it significantly increases the final model accuracy. + +This co-training across different submodels, which we refer to as cosub, can be regarded as a massive co-training between $2^{L}$ models that share a common set of parameters, where $L$ is the number of layers in the target architecture. The target model can be interpreted as the expectation of all models. With a layer drop-rate set to 0.5, for instance for a ViT-H model, all submodels are equiprobable, and then it amounts to averaging the weights of $2^{2 \times 32}$ models. + +Our contributions can be summarized as follows: + +- We introduce a novel training approach for deep neural networks: we co-train submodels. This significantly improves the training of most models, establishing the new state of the art in multiple cases. For instance, after pre-training ViT-B onImagenet-21k and fine-tuning it at resolution 448, we obtain $87.4\%$ top-1 accuracy onImagenet-val. +- We provide an efficient implementation to subsample models on the fly. It is a simple yet effective variation of stochastic depth [31] to drop residual blocks. +- We provide multiple analyses and ablations. Noticeably, we show that our submodels are effective models by themselves even with significant trimming, similar to LayerDrop [20] in natural language processing. +- We validate our approach on multiple architectures (like ViT, ResNet, RegNet, PiT, XCiT, Swin, ConvNext), both for image classification -trained from scratch or with transfer-, and semantic segmentation. +- We will share models/code for reproducibility in the DeiT repository. + +# 2. Related work + +Knowledge distillation. Originally, distillation was introduced as a way to train a model such that it reproduces the performance of another model [4, 28]. The typical use-case is to improve the quality of a relatively small model by leveraging a strong teacher, whose complexity may be prohibitive for a practical deployment. The teacher's soft labels have a similar effect as label smoothing [59]. As shown by Wei et al. [53], the teacher's supervision takes into account the effects of the data augmentation, which sometimes causes a misalignment between the real label and the image. Knowledge distillation can transfer inductive biases [1] in a soft way in a student model when using a teacher model these biases are enforced in a hard way. Touvron et al. [50] proposed a variant of distillation adapted to Vision Transformer (ViT), showing the effectiveness of using a ConvNet teacher for a ViT student. + +Mixture models. Ensembling has a long history that we can trace back to the origins of statistics [37] and the possibility of improving the precision of measurements with multiple observations. In machine learning, bagging [8] combines multiple weak classifiers to produce a strong one. This idea is naturally extended to neural networks, where it can offer improved stability or accuracy, or other properties like anytime inference [44]. A mixture model can also be seen as a larger model with an internal parallel structure. + +Co-distillation. At the interface of mixture models and distillation, the concept of co-distillation [62, 64] does not require a prior teacher: the mixture serves as the teacher to all the networks in the mixture. The models are jointly optimized, which leads to improved accuracy of the individual models [2]. This can be regarded as a form of collaborative learning between the different elements of the mixture [45]. Compared to training only one model, co-distillation involves training two or more models, each requiring storage for weights and activation, and computing backward and forward passes. + +Exponential moving average teacher. A special form of teacher is a model obtained from the student model itself. Tarvainen and Valpila [49] show that such a model obtained by temporal averaging of intermediate checkpoints during the training is an effective teacher, that can be obtained with an exponential moving average (EMA). This idea has also been adopted in self-supervised training with learning schemes like DINO [9]. This method involves storing a copy of the weights corresponding to the temporal averaging carried out by the EMA model. + +Dropout and model populations. Dropout [46] is an effective way to regularize models. With residual architectures, one effective way to train deeper architecture is stochastic depth [31]. It reduces the size of a network at + +![](images/827e24b5c3a9e8f3960ee8f180d6c08bfd939a86fcf599f96570ddf5a723151d.jpg) +(a) supervised + +![](images/0577d9e58a051c00e478e494123d045453ed591c00b97aa9463d301198da72c3.jpg) +(b) knowledge distillation + +![](images/52415c6721f74905e1f9e2aa415e28ad94c2ee870d6e251fa66f3018e5d87f32.jpg) +(c) self-distillation + +![](images/760887e7f56f7476c1a24653682b1296bc185273378ec01af654b59712b33505.jpg) +(d) co-training +Figure 2. Brief summary of related works and of submodel co-training (cosub). (a) supervised baseline: the supervision is provided as a one-hot label. (b) Knowledge distillation [28]: a teacher provides a (soft) target. A simple yet effective variant [50] combines the predicted label of the teacher with the image label. KD requires a pre-existing teacher that must be stored for inference. (c) In self-distillation, the teacher is obtained from the model itself, typically by model averaging. (d) Co-training involves two distinct models that serves as teacher to each other. (e) Submodel co-training generates on-the-fly two distinct models (amongst $2^{L}$ ) for each sample, which serves a teacher to each other. We only need to store a single set of weights for the model and optimizer because the random selections $\theta_{1} = \mathcal{T}(\theta)$ (resp., $\theta_{2}$ ) of submodels allow us to back-propagate on $\theta$ . + +![](images/d966bd13264237f7aecd65717121e981cfc89e5e8bfa8b9d0122c9a9f8cebd80.jpg) +(e) submodel co-training + +
methodmodel paramscompute*
weightsoptimizer
(a) supervised×1×1×1
(b) KD†×2×1×1.33
(c) mean teacher×2×1×1.33
(d) co-training×2×2×2
(e) Cosub (ours)×1×1×2
+ +\*backward pass assumed $2\times$ the complexity of forward $\dagger$ requires a trained teacher model + +training time by dropping residual blocks during training. The initial goal of this approach was to improve the training of deep network. However, there are other implications: in natural language processing, Fan et al. show that the counterpart of stochastic depth, namely LayerDrop [20], is effective to reduce the transformer depth on demand. Interestingly, submodels extracted from a model trained with LayerDrop are stronger than those trained with the full target depth with a proper selection strategy for layers. + +In our paper, we regard stochastic depth as a regularization technique, but at the same time we adopt the point of view of LayerDrop, i.e., an effective way to train a population of submodels that share parameters, or entire layers in our case. From that viewpoint, stochastic depth amounts to training $2^{L}$ distinct models. This is different from population training as involved in network space design [41], whose goal is related to network architecture search [18], which aims at optimizing the architecture itself. + +More recently, model soups [56] are models obtained by averaging the weights of multiple models finetuned with different hyperparameter configurations. The authors show that it often improves accuracy and robustness. We point out that stochastic depth can be regarded as a special form of model soup over the entire population of submodels that we can instantiate with stochastic depth. We further discuss this relationship in the next section. + +# 3. Submodel co-training + +In this section, we present our cosub approach, for constraining submodels. We consider a neural network $f_{\theta}$ parameterized by learnable parameters $\theta$ . + +Submodel instantiation: model augmentation operator. We first define a model augmentation operator $\mathcal{T}$ . For a given neural network, it provides a set of parameters $\theta^{\prime} =$ + +$\mathcal{T}(\theta, R)$ that allow us to define variations $f_{\theta'}$ of the function $f_{\theta}$ by drawing a random variable $R$ . The model augmentation is such that $f_{\theta}$ and $f_{\theta'}$ share parameters, hence any update on $\theta'$ modifies $\theta$ and therefore $f_{\theta}$ accordingly. A simple way to define $\mathcal{T}$ is to replace some parameters by zeros, which corresponds to dropout [46]. In this paper, we focus on stochastic depth, which has interesting connections with model averaging [56]. + +Overview. For a given training sample $x$ within a batch, the training is as follows (see also Figure 1): + +1. We first data-augment the image, producing $\hat{x}$ . +2. The image is duplicated with the batch, effectively doubling the batch size, which hence contains two identical copies of each augmented image1. +3. We determine the stochastic depth pattern for each sample according to the target drop-rate $\tau$ , which amounts to producing two functions $f_{\theta_1}$ and $f_{\theta_2}$ . See Section 4 for the details of this procedure. +4. The forward pass proceeds as usual: we compute the soft output labels $y_{1} = f_{\theta_{1}}(\hat{x})$ and $y_{2} = f_{\theta_{2}}(\hat{x})$ . +5. We compute the losses and the backwards pass accordingly. Note that the gradients on $\theta_{1}$ and $\theta_{2}$ are directly used to update $\theta$ , as they are just subsets of $\theta$ . + +Loss. Each submodel is trained using a weighted average of (i) the standard binary cross-entropy loss obtained from the image label $y$ , and (ii) a binary cross-entropy loss w.r.t. the soft-labels computed for the the same image by the other submodel. The respective weight of the standard binary cross-entropy loss $\mathcal{L}_{\mathrm{label}}$ versus the cosub loss $\mathcal{L}_{\mathrm{cosub}}$ is controlled by the hyper-parameter $\lambda$ , as + +$$ +\mathcal {L} _ {\text {t o t}} = \lambda \mathcal {L} _ {\text {l a b e l}} + (1 - \lambda) \mathcal {L} _ {\text {c o s u b}}. \tag {1} +$$ + +In details, the loss writes as + +$$ +\begin{array}{l} \mathcal {L} _ {\mathrm {t o t}} = \lambda \left(\frac {\mathcal {L} (y _ {1} , y) + \mathcal {L} (y _ {2} , y)}{2}\right) \\ + (1 - \lambda) \left(\frac {\mathcal {L} \left(y _ {1} , \mathrm {s g} \left(y _ {2}\right)\right) + \mathcal {L} \left(y _ {2} , \mathrm {s g} \left(y _ {1}\right)\right)}{2}\right), \quad (2) \\ \end{array} +$$ + +where $\mathcal{L}(y, y')$ is either the binary cross-entropy (BCE) or a cross-entropy (CE) loss. Importantly, when applying this loss, we do not back-propagate on the second term $y'$ . This is indicated by the stop-gradient operator $\mathrm{sg}(\cdot)$ . + +Discussion. The submodel instantiation provides a model that is, by itself, a valid neural network model. Co-training submodels is a way to enforce that all such submodels produce a consistent output. Amongst these submodels, a very special case is when we retain all residual blocks, which is the model that we primarily intend to use at inference time. Note that Fan et al. [20] show, in an NLP context, that submodels extracted from a deeper model are superior, for a given target depth. With proper rescaling of the residual branches, this specific submodel with all blocks activated can be seen as the expectation over all models. This can be shown as follows: let us consider an extra scalar parameter $s_l$ associated with a given residual block $r_l(x)$ , which we use as a multiplicative factor on output of each residual: $s_l = 1$ if the residual block $r_l$ is included in the submodel, $s_l = 0$ otherwise. Each submodel is fully parameterized by a binary vector $s = (s_1, \dots, s_L)$ , where $L$ is the total number of layers. Therefore all submodels have the same parameters $\theta$ , and only differ by $s$ indicating the zeroed residual blocks. The weight expectation $\bar{\theta}$ is hence + +$$ +\begin{array}{l} \bar {\theta} = \mathbb {E} _ {s \in [ 0, 1 ] ^ {L}} [ \theta ] = \theta , \\ \bar {s} = \mathbb {E} _ {s \in [ 0, 1 ] ^ {L}} [ s ] = [ 1 - \tau , \dots , 1 - \tau ] ^ {\top}, \tag {3} \\ \end{array} +$$ + +where $\bar{s}$ is the scaling factor used in stochastic depth. Under a uniform distribution of submodels that is obtained when the stochastic drop rate $\tau = 0.5$ , the inference-time model is exactly the average of all $2^{L}$ models. While stochastic depth is often regarded as a regularization technique, this averaging interpretation relates it to the recent model soup [56]. + +# 4. Efficient stochastic depth + +We revisit the original stochastic depth [31] formulation in order enable an efficient implementation, which we will release for PyTorch [40]. Instantiating a submodel for a sample amounts to selecting a subset of residual layers of the model and performing training on these. This can be implemented using the stochastic depth (SD) approach, whose objective was initially to improve the training of very deep networks. In stochastic depth, for each sample and each layers of the network, we select whether the layers will be dropped or not with a certain probability $\tau$ . + +![](images/9707da9db87eb1046491989ddb46f98c5c976595da4accbd49ee490813dd094b.jpg) +Figure 3. Efficient implementation of stochastic depth, using the permute-select approach. In this example, we drop the residuals for samples 3 and 6, corresponding to a drop rate $\tau = 0.25$ . For a drop rate higher than 0.1, the overhead of the approach is negligible in terms of memory and compute. + +In practice, e.g. in the timm library [54], stochastic depth is implemented by masking with zeros the residuals added for a given sample of a batch. This is not efficient: this naive approach performs the computations for a residual then throws it away, wasting computation. + +Our efficient stochastic depth (ESD) approach addresses this issue, saving both memory and compute, and is illustrated in Figure 3. For each layer, given a batch size $B$ and a drop rate $\tau$ , we apply the layer to $B_{\mathrm{keep}} = \mathrm{round}(B \times (1 - \tau))$ samples in the batch. In contrast to the original version of SD, our efficient version drops a fixed number of samples at each layer, where $d$ is adjusted such that $B_{\mathrm{keep}}$ is an integer. In our experiments, this did not have an effect on the final accuracy of the models. Our efficient implementation proceeds as follows: (i) we apply a random permutation of the $B$ samples, then (ii) we select the first $B_{\mathrm{keep}}$ samples. We then compute the residual function for the selected subset, then add the result onto the full batch using the built-in index_add function and scaling the result by $\frac{1}{1 - \tau}$ to have the correct scaling as in the original SD formulation. + +Discussion: progressive vs uniform rate. In the original paper [31], stochastic depth drops a layer with a probability that is linearly increasing during the forward pass: layers closer to the output has a higher chance to be dropped. However this strategy is limited with high drop rates, as later layers are excessively dropped. Touvron et al. [52] adopt a uniform rate drop per layer. It is as effective as the progressive rate with vision transformers, but it makes it possible to target a drop-rate greater than 0.5 on average. For this reason we adopt an uniform drop rate everywhere. + +Our efficient stochastic depth implementation variant works both with uniform and progressive drop-rate. In the case of progressive, the effective batch size is decreasing during to the forward pass. One limitation is that our technique implies a quantization of the drop-rate, which can be problematic for small batch sizes. Therefore one must take care to properly verify that the drop-rate is not too severely quantized, in particular in the progressive case for which the drop-rate can be very high in the last layers. In the supplemental material (Appendix C), we discuss with more details the drop-rate quantization effect. + +![](images/f52e69c2b34f9bd6846bc6b0ef914df1a935f850c1a0c262ab10a20143d0e470.jpg) +Figure 4. Cosub as population training: each submodel extracted by a transformation $\mathcal{T}$ is a valid neural network. Our cosub strategy can hence be regarded as co-training a large number of subnetworks. We plot the accuracy of the submodels as a function of the number of layers that we preserve. We drop layers with probability $\tau = 0.2$ for ViT-B and $\tau = 0.45$ for ViT-L. On average our method provides a significant boost in performance for the whole population of submodels extracted from the main model. + +
ModelViT-SViT-MViT-BViT-LViT-H
τ(0.05)(0.1)(0.2)(0.45)(0.6)
baseline81.482.583.784.584.9
baseline + cosub81.582.883.984.985.5
+ +# 5. Experiments + +We evaluate our approach on residual architectures [24, 25], since they are readily compatible with the stochastic dropout. We take as our main reference the vanilla Vision Transformer introduced by Dosovitskiy et al. [14]. However, as shown in our experiments, our approach is effective with all the residual architectures that we have considered. + +# 5.1. Baseline and training settings + +We adopt the state-of-the-art training procedure proposed in the DeiT III paper [51] as baseline for transformers architectures and the training procedure from Wightman et al. [55] for the convnets. All hyper-parameters are identical except onImagenet-21k, where the hyper-parameter $\tau$ are adjusted depending on the training setting. We recapitulate the training hyper-parameters in Appendix A. + +We additionally adopt and measure the impact of LayerDecay for the fine-tuning when transferring from Imagenet-21k to Imagenet-1k. This method slightly boosts the performance, as discussed later in this section, and in Table 10. This method was adopted in multiple recent papers and in particular for fine-tuning of self-supervised approaches [5, 22], yet the contribution of this fine-tuning ingredient was not quantitatively measured. + +Table 1. Comparison of top-1 accuracy with ViT architecture trained with/without cosub at resolution 224 (800 epochs) onImagenet-1k. The improvement is not significant for the smaller architecture ViT-S. Cosub is gradually more effective when we increase the model size and the SD rate. + +
ModelτOriginalBaseline+cosubΔ
Transformers
ViT-L [14]0.4576.584.584.9+0.4
CaiT-L24 [52]0.45-83.884.4+0.6
PiT-B [26]0.2582.083.884.1+0.3
XCiT-L12 [17]0.20-82.683.0+0.4
Swin-B [34]0.2083.582.983.3+0.4
Swin-L [34]0.45-80.884.0+3.2
Convnets
ResNet-50 [24]0.1076.280.280.3+0.1
ResNet-101 [24]0.2077.481.882.1+0.4
ResNet-152 [24]0.3078.382.483.1+0.7
RegNet-16GF [42]0.3080.482.983.8+0.9
+ +Table 2. Benefit of cosub for different architectures trained from scratch on Imagenet-1k at resolution 224. We report top-1 acc. for the supervised baseline and cosub, as well as results reported in the corresponding papers when available (trained with different settings). We have adjusted the stochastic-depth drop-rate (SD) hyper-parameter for each architecture. + +
LossBCE-softBCE-hardCE-hard
Imagenet-val top-1 accuracy83.583.581.9
+ +Table 3. Ablation on the loss for cosub with ViT-H trained at resolution $126\times 126$ on Imagenet-1k during 800 epochs. The training is inherited from DeiT-III, which also uses BCE when training with Imagenet-1k only. + +# 5.2. Empirical analysis of cosub + +We perform various ablations on Imagenet-1k to analyse the impact of our training method on the learned networks. + +Performance for different model sizes and architectures. Table 1 provides the results obtained by the baseline and cosub when we vary the model size of vision transformers. The stochastic depth coefficient was optimized for the baseline and we keep it unchanged with cosub. As to be expected, our method is almost neutral for small models like ViT-S: $+0.1\%$ top-1 accuracy, which is about the standard deviation of measurements. The improvement is increasingly important for larger models, up to a significant improvement of $+0.6\%$ top-1 accuracy for ViT-H models. + +In Table 2, we show that our approach is beneficial with all architectures that we have tried. We report the results of the original paper, evaluate the performance with our baseline training, and measure the improvement brought by cosub. For almost all architectures, we observe a significant boost in performance. The exception is the ResNet-50, for which cosub improves the top-1 accuracy by only $+0.1\%$ , similar to our observation with ViT-S. In Table B.1 in the appendix we present improved results obtained for multiple architectures pre-trained with cosub onImagenet-21k. + +Analysis of submodel performance. With cosub, we sample different subnetworks during training to perform the constraining. We analyse the impact of cosub on the accuracy of the sub networks themselves. In Figure 4 we consider the accuracy of submodels of different size of ViT-B and ViT-L + +
MethodViT-LViT-H
(a)supervised baseline [51]84.584.9
(b)KD85.385.3
(c)mean teacher84.483.4
(d)co-training82.683.1
(e)cosub84.985.5
(b) + (e)KD + cosub85.385.7
+ +Table 4. Training strategies with distillation. We compare on Imagenet-1k at resolution $224 \times 224$ different approaches involving co- or self-training with distillation. KD, mean teacher and co-training use the same $\lambda = 0.5$ and same hyper-parameters as in Deit-III. Unlike cosub, the mean teacher (c) requires other hyper-parameters for EMA. We perform a small grid-search to adjust this parameters. Note (last row): our approach is complementary with KD, assuming a pre-trained teacher is available beforehand. + +
λ0.10.30.50.70.91.0
top1 accuracy79.0583.2783.5583.2083.0482.91
+ +models. Cosub improves the accuracy of the whole population of sub-networks and, in particular, the target network. + +Loss formulation. In Table 3 we experiment with different losses for cosub. With Imagenet1k, DeiT-III training uses BCE instead of CE for the main loss. With cosub, BCE is more compatible with the loss of the baseline and, as to be expected, we also observe a better performance with BCE. We have done ablations using hard and soft targets for the cosub loss. The results are similar, therefore by default we keep soft-targets for the cosub loss. + +Alternative teacher/student. In Table 4 we report the results obtained with the different distillation or co-training approaches depicted in Figure 2. Other approaches are not effective off-the-shelf, except KD that requires a pre-trained teacher. Our approach is on par with KD (lower for ViT-L, better or ViT-H), and in the last row we show that it even provides a slight boost to combine KD with cosub. + +Impact of the cosub loss. The hyper-parameter $\lambda$ controls the tradeoff between the co-distillation loss and the cross-entropy classification loss. Setting $\lambda = 1$ means that we have a regular supervised training setting, except that (i) we double the batch size by duplicating the image after data augmentation, and (ii) stochastic depth selects different layers for each image copy. + +In Table 5, we measure the impact of $\lambda$ , with all other hyper-parameters being fixed, for ViT-H trained at resolution $126 \times 126$ onImagenet-1k. We observe that the best ratio is to use an equal weighting of the cosub loss and the classic training loss. Using the cosub loss increases the performance by $0.6\%$ Top-1 accuracy onImagenet-val, which is the typical improvement that we observe for large models. + +Table 5. Ablation of the parameter $\lambda$ controlling the weight of the co-distillation loss across submodels (Imagenet1k-val, top1-acc). Model ViT-H trained at resolution ${126} \times {126}$ onImagenet-1k during 800 epochs. $\lambda =$ 1.0 corresponds a supervised baseline and $\lambda = {0.5}$ corresponds to cosub. + +
ModelτepochsBaseline+cosub
valv2valv2
CaiT-M120.2040083.272.983.773.5
80082.972.683.673.1
PiT-B0.2540083.873.684.174.1
80082.471.983.172.8
ViT-B0.2040083.172.683.273.1
80083.773.183.973.5
120083.372.8--
160083.373.4--
ViT-H0.6040084.875.385.075.8
80084.975.685.576.3
+ +Table 6. We compare ViT models trained with and without cosub on ImageNet-1k only with different number of epochs at resolution $224 \times 224$ . One can see that cosub is more effective for larger models yielding higher values of the SD hyper-parameter $\tau$ . It avoids the early saturation or overfitting of the performance that we typically observe with the baseline when we increase the training time without re-adjusting hyper-parameters. See also Table 5 for a direct comparison with and without the co-distillation loss, and Table 7 for the corresponding training times per epoch. + +
Training methodmodelGPUs usedMemory peak (GB)Time (min) by epoch
DeiT-IIIViT-L3221.48
ViT-H6427.612
DeiT-III + ESDViT-L3215.19
ViT-H6415.211
cosub (with ESD)ViT-L3226.916
ViT-H6425.017
+ +Table 7. Training times of different models trained at resolution $224\times 224$ with batch size 2048 on Imagenet-1k with DeiT-III and our approach. cosub uses our efficient stochastic depth (ESD), which amortizes the extra memory needed by cosub, especially for the largest models with high stochastic depth values (0.45 for ViT-L, and 0.6 for ViT-H). Timings are indicative and not representative of an optimized selection of the batch size, see Appendix C for measurements with adjusted parameters. + +Number of training epochs. In Table 6 we compare results on Imagenet-1k-val and Imagenet-v2 different for architectures trained with and without cosub on Imagenet-1k only at resolution $224 \times 224$ with different number of epochs. We observe less overfitting with cosub and longer training schedule. In particular, with bigger architecture like ViT-H, we observe continuous improvement with a longer schedule where the baseline saturates. + +Training time. In Table 7 we compare the training costs of cosub and DeiT-III. Thanks to our efficient stochastic depth formulation we maintain a similar memory peak during training. For bigger architectures the gap in training speed between cosub and the baseline is decreasing. + +Resolution. In Table 8 we compare different ViT architectures trained with and without cosub at different resolutions onImagenet-1k. We fine-tune during 20 epochs at resolution $224 \times 224$ before evaluation at this resolution. We observe that cosub gives significant improvements across the different resolutions and models. + +
ModelResolutionDeit-III Baseline+cosub
valv2valv2
ViT-B128 × 12883.573.483.874.0
192 × 19283.873.684.174.0
224 × 22483.773.183.973.5
ViT-L128 × 12884.574.785.175.5
192 × 19284.975.185.275.7
224 × 22484.575.084.975.6
ViT-H126 × 12685.175.685.676.4
182 × 18285.175.985.776.6
224 × 22484.975.685.576.3
+ +Table 8. Imagenet-1k val and v2 top-1 accuracy of ViT models trained with and without cosub for 800 epochs on Imagenet-1k at different resolutions, followed by finetuning for 20 epochs at resolution $224 \times 224$ . + +
resol. →112224336448112224336448
↓ modelImagenet-valImagenet-v2
ViT-S78.083.184.685.266.673.775.176.3
ViT-M80.685.086.086.369.676.076.877.2
ViT-B82.886.386.987.472.177.077.978.3
ViT-L85.487.588.188.375.779.179.880.0
ViT-H86.288.0--76.979.6--
ViT-g86.5---77.3---
+ +Table 9. Performance of models at different resolutions. We report the results obtained on Imagenet-val by models of different sizes pre-trained with cosub on Imagenet-21k and fine-tuned on Imagenet-1k. Training schedule: 270 epochs except ViT-g (90 epochs). Except for ViT-S, the results at resolution 336 and 448 were pre-trained on Imagenet-21k at resolution 224 for efficiency reasons. We have not fine-tuned ViT-H and ViT-g at large resolutions since these models are computationally expensive. + +
MethodLong TrainingLayer DecayViT-SModel ViT-BViT-L
baselineXX82.685.286.8
XX82.585.887.4
cosubX✓82.786.087.5
✓X82.886.087.4
✓✓83.186.387.5
+ +Imagenet-21k impact of layer-decay. In Table 10 we compare different number of epochs for the pre-training on Imagenet-21k and the finetuning on Imagenet-1k with and without layer-decay. We observe that both layer-decay and long training bring improvements with cosub. + +# 5.3. Comparisons with the state of the art + +Imagenet-1k. In Table 11 we compare architectures trained with cosub with state-of-the-art results from the literature for this architecture. We observe that architectures trained with cosub are very competitive. For instance, for ResNet-152, RegNetY-16GF, PiT-B we improve the best results reported in the literature by more than $0.9\%$ top-1 accuracy on Imagenet-1k-val. + +Table 10. We measure the impact of layer-decay during the finetuning on Imagenet-1k for models pre-trained on Imagenet-21k with cosub during 90 epochs (default) and 270 epochs (long training). + +
Modelnb params (×106)throughput (im/s)FLOPs (×109)Peak Mem (MB)Top-1 Acc.v2 Acc.
“Traditional” ConvNets
ResNet-50 [24, 55]25.625874.1218280.468.7
ResNet-101 [24, 55]44.515867.9226981.570.3
ResNet-101 - cosub44.515867.9226982.170.8
ResNet-152 [24, 55]60.2112211.6235982.070.6
ResNet-152 - cosub60.2112211.6235983.172.1
RegNetY-8GF [42, 55]39.211588.0393982.271.1
RegNetY-16GF [42, 50]83.671416.0520482.972.4
RegNetY-16GF - cosub83.671416.0520483.872.8
Vision Transformers derivatives
Swin-T [34]28.311094.5334581.369.5
Swin-B [34]87.853215.4469583.5-
PiT-S [26]23.518092.9329380.9-
PiT-S - cosub23.518092.9329381.369.7
PiT-B [26]73.861512.5756482.0-
PiT-B - cosub73.861512.5756484.174.1
Vanilla Vision Transformers
ViT-S [50]22.018914.698779.868.1
ViT-S - Deit III [51]22.018914.698781.470.5
ViT-S - cosub22.018914.698781.570.8
ViT-B [14]86.683117.5207877.9-
ViT-B - DeiT [50]86.683117.5207881.871.5
ViT-B - DeiT/distilled86.683117.5207883.473.2
ViT-B - DeiT III [51]86.683117.5207883.873.6
ViT-B - cosub86.683117.5207884.274.2
ViT-L - Deit III [51]304.427761.6378984.975.1
ViT-L - cosub304.427761.6378985.375.5
ViT-H - Deit III [51]632.1112167.4698485.275.9
ViT-H - cosub632.1112167.4698485.776.6
+ +Table 11. Classification with Imagenet1k training. We compare with models trained on Imagenet-1k only at resolution $224 \times 224$ without self-supervised pre-training (see supp. material for a comparison). We report Top-1 accuracy on Imagenet-1k-val and Imagenet-v2 with different measures of complexity: throughput, FLOPs, number of parameters and peak memory usage. The throughput and peak memory are measured on a single V100-32GB GPU with batch size fixed to 256 and mixed precision. + +Imagenet-21k. In Table 12 we compare ViT models pretrained with cosub on Imagenet-21k and finetuned with cosub on Imagenet-1k with other architectures and our baseline DeiT-III. Our results with vanilla ViT outperform the baseline and are competitive with recent architectures. + +Overfitting analysis. As recommended in prior works [51, 55] we perform an overfitting analysis. We evaluate our models trained with codist onImagenet-1k val andImagenet-v2 [43]. The results are reported in Figure 5. For ViT, we observe that cosub does not overfit more than the DeiT-III baseline [51]. Our results with other architectures in Table 12 concur with that observation: our results are comparatively stronger onImagenet-v2 than those reported in the literature for the exact same models. + +# 5.4. Downstream tasks + +Semantic segmentation. First we evaluate our ViT models pre-trained onImagenet with cosub for semantic segmentation on the ADE20k dataset [63]. ADE20k consists of 20k training and 5k validation images with labels over 150 categories. We adopt the training schedule from Swin: 160k iterations with UperNet [57]. At test time we evaluate with + +
Architecturenb params (×106)throughput (im/s)FLOPs (×109)peak mem (MB)top1 valacc. v2
Convnets
EfficientNetV2-M [48]54.131225.0712786.275.9
EfficientNetV2-L [48]118.517953.0954086.876.9
ResNet-152 [24, 55]60.2112211.6235982.070.6
ResNet-152 - cosub60.2112211.6235983.173.1
RegnetY16GF - cosub83.671416.0520484.274.7
ConvNeXt-S- cosub50.27838.7221885.276.0
ConvNeXt-B [35]88.656315.4302985.875.6
ConvNeXt-B - cosub88.656315.4302985.876.9
ConvNeXt-L [35]197.834434.4486586.676.6
ConvNeXt-XL [35]350.224160.9695187.077.0
Transformers variations
Swin-B [34]87.853215.4469585.274.6
Swin-B - cosub87.853215.4469586.277.2
Swin-L [34]196.533734.5735086.376.3
Swin-L - cosub196.533734.5735087.178.1
PiT-B - cosub [26]73.861512.5756485.876.8
XCiT-S12 - cosub [17]26.213734.9133084.274.9
XCiT-M24 - cosub [17]84.455316.2201086.578.0
XCiT-L24 - cosub [17]189.033436.1331587.277.8
Vanilla Vision Transformers [14, 51]
ViT-S - cosub22.018914.698783.173.7
ViT-M - cosub39.013078.0132285.076.0
ViT-B - DeiT-III86.683117.6207885.776.5
ViT-B - cosub86.683117.6207886.377.0
ViT-L - DeiT-III304.427761.6378987.078.6
ViT-L - cosub304.427761.6378987.579.1
ViT-H - DeiT-III632.1112167.4698487.279.2
ViT-H - cosub632.1112167.4698488.080.0
+ +Table 12. Classification with ImageNet-21k pretraining. We report top-1 accuracy on the validation set of Imagenet1k and Imagenet-V2 with different measures of complexity. The peak memory usage is measured on a single V100-32GB GPU with batch size fixed to 256 and mixed precision. For Swin-L the memory peak is an estimation since we decreased the batch size to 128 to fit in memory. All models are evaluated at resolution 224 except EfficientNetV2 that use resolution 480. ViT are pretrained with a $\times 3$ schedule, comparable to that used in the best DeiT-III baseline (270 vs. 240 epochs). All other cosub models are pretrained during 90 epochs on Imagenet21k, with 50 epochs of fine-tuning. The $\tau$ hyper-parameter is set per model based on prior choices or best guess based on model size. + +![](images/f8c82359cfd5310e3d42ed277bb8b03a5174a9a81497154610bcc05e2cced25e.jpg) +Figure 5. Overfitting measurement: top-1 accuracy on Imagenet-val vs. Imagenet-v2 for models in Tables 11 and 12 pre-trained on Imagenet-1k and Imagenet-21k, respectively. Our cosub ViTs (plain lines and points) do not overfit more than DeiT-III [51] overall. Our Imagenet-21k results for Swin and Convnext generalize much better on v2 than the original models. + +![](images/8d84030110f5e627046f800c771d9c5379935fe7bc51afca709a02acd1ff4a8c.jpg) + +a single scale and multi-scale. Our UperNet implementation is the same as in DeiT III [51]. Our results are reported in Table 13. We observe that vanilla ViTs trained with cos-sub outperform our baseline DeiT-III but also have better FLOPs-accuracy trade-offs than recent architectures. + +
Backbone#params (×106)FLOPs (×109)Single-scale mIoUMulti-scale mIoU
DeiT-S52.01099-44.0
Swin-T59.994544.546.1
ViT-S – DeiT-III41.758845.646.8
ViT-S – cosub41.758847.048.0
PatchConvNet-B60140.6125848.148.6
PatchConvNet-B120229.8155049.450.3
MAE ViT-B127.7128348.1-
Swin-B121.0118848.149.7
ViT-B – DeiT-III127.7128349.350.2
ViT-B – cosub127.7128349.349.9
ViT-L – DeiT-III353.6223151.552.0
ViT-L – cosub353.6223152.553.1
PatchConvNet-B60†140.6125850.551.1
Swin-B† (640 × 640)121.0184150.051.6
ViT-B – DeiT-III†127.7128353.454.1
ViT-B – cosub†127.7128353.754.7
PatchConvNet-L120†383.7208652.252.9
Swin-L† (640 × 640)234.03230-53.5
ViT-L – DeiT-III†353.6223154.655.6
ViT-L – cosub†353.6223155.756.3
+ +Table 13. ADE20k semantic segmentation performance using UperNet [57], in comparable setting as prior works [13, 17, 34]. All models are pretrained on Imagenet-1k, except bottom models identified with $\dagger$ , which are pre-trained on Imagenet-21k. By default the finetuning resolution on ADE20k is $512 \times 512$ except when mentioned otherwise (for Swin). + +
ModelCifar-10Cifar-100FlowersCarsiNat-18iNat-19
ViT-S - DeiT-III98.990.696.489.967.172.7
ViT-B - DeiT-III99.392.598.693.473.678.0
ViT-L - DeiT-III99.393.498.994.575.679.3
ViT-S - cosub99.191.797.493.070.175.6
ViT-B - cosub99.192.698.493.574.178.1
ViT-L - cosub99.493.598.894.576.280.1
+ +Table 14. ViT models pre-trained with cosub or DeiT-III on Imagenet-1k and finetuned on six different target datasets. We note that for small datasets (CIFAR, Flowers, and Cars) our approach is useful for small models but neutral for larger models. The gains are more significant when transferring to the larger iNaturalist-18 and iNaturalist-19 datasets. + +Transfer learning. We now measure how the performance improvements observed with cosub translate to other classification problems. For this purpose, we performed transfer learning on the six different datasets used in DeiT-III. Our results are reported Table 14. Our pre-trained and fine-tuned models with cosub generally improve the baseline. The gains are overall more significant on the more challenging datasets like iNaturalist 2018 and iNaturalist 2019. + +# 6. Conclusion + +Co-training submodels (cosub) is an effective way to improve existing deep residual networks. It is straightforward to implement, just involving a few lines of code. It does not need a pre-trained teacher, and it only maintains a single set of weights for the model. Extensive experimental results on image classification, transfer learning and semantic segmentation show that cosub is overall extremely effective. It works off-the-shelf and improves the state of the art for various network architectures, including convnets like Regnet. + +# References + +[1] Samira Abnar, Mostafa Dehghani, and Willem Zuidema. Transferring inductive biases through knowledge distillation. arXiv preprint arXiv:2006.00555, 2020. 2 +[2] Rohan Anil, Gabriel Pereyra, Alexandre Passos, Robert Ormandi, George E Dahl, and Geoffrey E Hinton. Large scale distributed neural network training through online distillation. arXiv preprint arXiv:1804.03235, 2018. 2 +[3] Mahmoud Assran, Mathilde Caron, Ishan Misra, Piotr Bojanowski, Florian Bordes, Pascal Vincent, Armand Joulin, Mike Rabbat, and Nicolas Ballas. Masked siamese networks for label-efficient learning. In European Conference on Computer Vision, pages 456-473. Springer, 2022. I +[4] L. Ba and R. Caruana. Do deep nets really need to be deep? In NeurIPS, 2014. 2 +[5] Hangbo Bao, Li Dong, and Furu Wei. Beit: Bert pre-training of image transformers. arXiv preprint arXiv:2106.08254, 2021. 1, 5, I, II +[6] Irwan Bello, W. Fedus, Xianzhi Du, E. D. Cubuk, A. Srinivas, Tsung-Yi Lin, Jonathon Shlens, and Barret Zoph. Revisiting ResNets: Improved training and scaling strategies. arXiv preprint arXiv:2103.07579, 2021. II +[7] H. Bourlard and Y. Kamp. Auto-association by multilayer perceptrons and singular value decomposition. Biological Cybernetics, 59:291-294, 1988. 1 +[8] Leo Breiman. Bagging predictors. Machine learning, 24(2):123-140, 1996. 2 +[9] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. arXiv preprint arXiv:2104.14294, 2021. 1, 2, II +[10] Kevin Clark, Minh-Thang Luong, Quoc V Le, and Christopher D Manning. Electra: Pre-training text encoders as discriminators rather than generators. arXiv preprint arXiv:2003.10555, 2020. I +[11] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In Conference on Computer Vision and Pattern Recognition, pages 248–255, 2009. 1 +[12] Carl Doersch, Ankush Gupta, and Andrew Zisserman. Crosstransformers: spatially-aware few-shot transfer. arXiv preprint arXiv:2007.11498, 2020. 1 +[13] Xiaoyi Dong, Jianmin Bao, Dongdong Chen, Weiming Zhang, Nenghai Yu, Lu Yuan, Dong Chen, and Baining Guo. Cswin transformer: A general vision transformer backbone with cross-shaped windows. arXiv preprint arXiv:2107.00652, 2021. 8 +[14] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations, 2021. 5, 7, 8 +[15] Alaaeldin El-Nouby, Gautier Izacard, Hugo Touvron, Ivan Laptev, Hervé Jegou, and Edouard Grave. Are large-scale datasets necessary for self-supervised pre-training? arXiv preprint arXiv:2112.10740, 2021. 1 + +[16] Alaaeldin El-Nouby, Natalia Neverova, Ivan Laptev, and Hervé Jégou. Training vision transformers for image retrieval. arXiv preprint arXiv:2102.05644, 2021. 1 +[17] Alaaeldin El-Nouby, Hugo Touvron, Mathilde Caron, Piotr Bojanowski, Matthijs Douze, Armand Joulin, Ivan Laptev, Natalia Neverova, Gabriel Synnaeve, Jakob Verbeek, et al. Xcit: Cross-covariance image transformers. arXiv preprint arXiv:2106.09681, 2021. 5, 8, II +[18] Thomas Elsken, Jan Hendrik Metzen, and Frank Hutter. Neural architecture search: A survey. Journal of Machine Learning Research, 20(1):1997-2017, 2019. 3 +[19] Dumitru Erhan, Aaron Courville, Yoshua Bengio, and Pascal Vincent. Why does unsupervised pre-training help deep learning? In AISTATS, 2010. 1 +[20] Angela Fan, Edouard Grave, and Armand Joulin. Reducing transformer depth on demand with structured dropout. arXiv preprint arXiv:1909.11556, 2019. ICLR 2020. 2, 3, 4, IV +[21] Rohit Girdhar, Alaaeldin El-Nouby, Mannat Singh, Kalyan Vasudev Alwala, Armand Joulin, and Ishan Misra. Omnimea: Single model masked pretraining on images and videos. arXiv preprint arXiv:2206.08356, 2022. I +[22] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Doll'ar, and Ross B. Girshick. Masked autoencoders are scalable vision learners. arXiv preprint arXiv:2111.06377, 2021. 1, 5, I, II +[23] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. arXiv preprint arXiv:1911.05722, 2019. 1 +[24] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Conference on Computer Vision and Pattern Recognition, 2016. 1, 5, 7, 8, II +[25] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Identity mappings in deep residual networks. arXiv preprint arXiv:1603.05027, 2016. 1, 5 +[26] Byeongho Heo, Sangdoo Yun, Dongyoon Han, Sanghyuk Chun, Junsuk Choe, and Seong Joon Oh. Rethinking spatial dimensions of vision transformers. arXiv preprint arXiv:2103.16302, 2021. 5, 7, 8, II +[27] G. Hinton and R. Zemel. Autoencoders, minimum description length and Helmholtz free energy. In NeurIPS, 1993. 1 +[28] Geoffrey E. Hinton, Oriol Vinyals, and J. Dean. Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531, 2015. 2, 3 +[29] Grant Van Horn, Oisin Mac Aodha, Yang Song, Alexander Shepard, Hartwig Adam, Pietro Perona, and Serge J. Belongie. The iNaturalist species classification and detection dataset. arXiv preprint arXiv:1707.06642, 2017. II +[30] Grant Van Horn, Oisin Mac Aodha, Yang Song, Alexander Shepard, Hartwig Adam, Pietro Perona, and Serge J. Belongie. The inaturalist challenge 2018 dataset. arXiv preprint arXiv:1707.06642, 2018. II +[31] Gao Huang, Yu Sun, Zhuang Liu, Daniel Sedra, and Kilian Q. Weinberger. Deep networks with stochastic depth. In European Conference on Computer Vision, 2016. 2, 4, I + +[32] Jonathan Krause, Michael Stark, Jia Deng, and Li Fei-Fei. 3d object representations for fine-grained categorization. In IEEE Workshop on 3D Representation and Recognition, 2013. II +[33] Alex Krizhevsky. Learning multiple layers of features from tiny images. Technical report, CIFAR, 2009. II +[34] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. arXiv preprint arXiv:2103.14030, 2021. 5, 7, 8, II +[35] Zhuang Liu, Hanzi Mao, Chao-Yuan Wu, Christoph Feichtenhofer, Trevor Darrell, and Saining Xie. A convnet for the 2020s. arXiv preprint arXiv:2201.03545, 2022. 8, II +[36] I. Loshchilov and F. Hutter. Fixing weight decay regularization in adam. arXiv preprint arXiv:1711.05101, 2017. I +[37] G. J. McLachlan and D. Peel. Finite Mixture Models. Wiley Series in Probability and Statistics. John Wiley & Sons, 2000. 2 +[38] M-E. Nilsback and A. Zisserman. Automated flower classification over a large number of classes. In Proceedings of the Indian Conference on Computer Vision, Graphics and Image Processing, 2008. II +[39] Maxime Oquab, Leon Bottou, Ivan Laptev, and Josef Sivic. Learning and transferring mid-level image representations using convolutional neural networks. In Conference on Computer Vision and Pattern Recognition, 2014. 1 +[40] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. In Advances in neural information processing systems, pages 8026-8037, 2019. 4 +[41] Ilija Radosavovic, Raj Prateek Kosaraju, Ross Girshick, Kaiming He, and Piotr Dólar. Designing network design spaces. In Conference on Computer Vision and Pattern Recognition, 2020. 3 +[42] Ilija Radosavovic, Raj Prateek Kosaraju, Ross B. Girshick, Kaiming He, and Piotr Dollar. Designing network design spaces. Conference on Computer Vision and Pattern Recognition, 2020. 5, 7, II +[43] B. Recht, Rebecca Roelofs, L. Schmidt, and V. Shankar. Do ImageNet classifiers generalize to ImageNet? In International Conference on Machine Learning, 2019. 7 +[44] Adria Ruiz and Jakob Verbeek. Anytime inference with distilled hierarchical neural ensembles. In AAAI, 2021. 2 +[45] Guocong Song and Wei Chai. Collaborative learning for deep neural networks. NeurIPS, 31, 2018. 2 +[46] Nitish Srivastava, Geoffrey Hinton, Alex Krizhevsky, Ilya Sutskever, and Ruslan Salakhutdinov. Dropout: A simple way to prevent neural networks from overfitting. In Journal of Machine Learning Research, 2014. 1, 2, 3 +[47] Christian Szegedy, Vincent Vanhoucke, Sergey Ioffe, Jonathon Shlens, and Zbigniew Wojna. Rethinking the inception architecture for computer vision. Conference on Computer Vision and Pattern Recognition, 2016. 1 +[48] Mingxing Tan and Quoc V. Le. Efficientnetv2: Smaller models and faster training. In International Conference on Machine Learning, 2021. 8 + +[49] Antti Tarvainen and Harri Valpola. Mean teachers are better role models: Weight-averaged consistency targets improve semi-supervised deep learning results. Neurips, 30, 2017. 2 +[50] Hugo Touvron, M. Cord, M. Douze, F. Massa, Alexandre Sablayrolles, and H. Jégou. Training data-efficient image transformers & distillation through attention. International Conference on Machine Learning, 2021. 2, 3, 7 +[51] Hugo Touvron, Matthieu Cord, and Hervé Jégou. Deit III: Revenge of the vit. arXiv preprint arXiv:2204.07118, 2022. 5, 6, 7, 8, I, II +[52] Hugo Touvron, Matthieu Cord, Alexandre Sablayrolles, Gabriel Synnaeve, and Hervé Jégou. Going deeper with image transformers. International Conference on Computer Vision, 2021. 4, 5 +[53] Longhui Wei, An Xiao, Lingxi Xie, X. Chen, Xiaopeng Zhang, and Q. Tian. Circumventing outliers of autoaugment with knowledge distillation. In European Conference on Computer Vision, 2020. 2 +[54] Ross Wightman. Pytorch image models. https://github.com/rwrightman/pytorch-image-models, 2019.4 +[55] Ross Wightman, Hugo Touvron, and Hervé Jégou. Resnet strikes back: An improved training procedure in timm. arXiv preprint arXiv:2110.00476, 2021. 5, 7, 8, II +[56] Mitchell Wortsman, Gabriel Ilharco, Samir Ya Gadre, Rebecca Roelofs, Raphael Gontijo-Lopes, Ari S Morcos, Hongseok Namkoong, Ali Farhadi, Yair Carmon, Simon Kornblith, et al. Model soups: averaging weights of multiple fine-tuned models improves accuracy without increasing inference time. In International Conference on Machine Learning, 2022. 3, 4 +[57] Tete Xiao, Yingcheng Liu, Bolei Zhou, Yuning Jiang, and Jian Sun. Unified perceptual parsing for scene understanding. In European Conference on Computer Vision, 2018. 7, 8 +[58] Jason Yosinski, Jeff Clune, Yoshua Bengio, and Hod Lipson. How transferable are features in deep neural networks? arXiv preprint arXiv:1411.1792, 2014. 1 +[59] L. Yuan, F. Tay, G. Li, T. Wang, and Jiashi Feng. Revisit knowledge distillation: a teacher-free framework. Conference on Computer Vision and Pattern Recognition, 2020. 2 +[60] Sangdoo Yun, Dongyoon Han, Seong Joon Oh, Sanghyuk Chun, Junsuk Choe, and Youngjoon Yoo. CutMix: Regularization strategy to train strong classifiers with localizable features. arXiv preprint arXiv:1905.04899, 2019. 1 +[61] Hongyi Zhang, Moustapha Cisse, Yann N. Dauphin, and David Lopez-Paz. mixup: Beyond empirical risk minimization. arXiv preprint arXiv:1710.09412, 2017. 1 +[62] Ying Zhang, Tao Xiang, Timothy M Hospedales, and Huchuan Lu. Deep mutual learning. In Conference on Computer Vision and Pattern Recognition, pages 4320-4328, 2018. 2 +[63] Bolei Zhou, Hang Zhao, Xavier Puig, Sanja Fidler, Adela Barriuso, and Antonio Torralba. Scene parsing through ade20k dataset. Conference on Computer Vision and Pattern Recognition, 2017. 7 +[64] Xiatian Zhu, Shaogang Gong, et al. Knowledge distillation by on-the-fly native ensemble. Neurips, 31, 2018. 2 \ No newline at end of file diff --git a/2023/Co-Training 2L Submodels for Visual Recognition/images.zip b/2023/Co-Training 2L Submodels for Visual Recognition/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..c1d322baef2b39ab744750853eece271eca861cb --- /dev/null +++ b/2023/Co-Training 2L Submodels for Visual Recognition/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:52c12bc3efa342a7a727a7646f5427f5faed54b22644e2d211bcf3622767a361 +size 700509 diff --git a/2023/Co-Training 2L Submodels for Visual Recognition/layout.json b/2023/Co-Training 2L Submodels for Visual Recognition/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..58c694a892b46214ac30acb4e53bdb8d84013c78 --- /dev/null +++ b/2023/Co-Training 2L Submodels for Visual Recognition/layout.json @@ -0,0 +1,9560 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 145, + 102, + 449, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 102, + 449, + 121 + ], + "spans": [ + { + "bbox": [ + 145, + 102, + 449, + 121 + ], + "type": "text", + "content": "Co-training " + }, + { + "bbox": [ + 145, + 102, + 449, + 121 + ], + "type": "inline_equation", + "content": "2^{L}" + }, + { + "bbox": [ + 145, + 102, + 449, + 121 + ], + "type": "text", + "content": " Submodels for Visual Recognition" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 148, + 218, + 162 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 148, + 218, + 162 + ], + "spans": [ + { + "bbox": [ + 130, + 148, + 218, + 162 + ], + "type": "text", + "content": "Hugo Touvron\\*,†" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 163, + 220, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 163, + 220, + 175 + ], + "spans": [ + { + "bbox": [ + 130, + 163, + 220, + 175 + ], + "type": "text", + "content": "Piotr Bojanowski*" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 264, + 148, + 341, + 161 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 264, + 148, + 341, + 161 + ], + "spans": [ + { + "bbox": [ + 264, + 148, + 341, + 161 + ], + "type": "text", + "content": "Matthieu Cord†" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 264, + 163, + 341, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 264, + 163, + 341, + 175 + ], + "spans": [ + { + "bbox": [ + 264, + 163, + 341, + 175 + ], + "type": "text", + "content": "Jakob Verbeek*" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 384, + 148, + 465, + 162 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 384, + 148, + 465, + 162 + ], + "spans": [ + { + "bbox": [ + 384, + 148, + 465, + 162 + ], + "type": "text", + "content": "Maxime Oquab\\*" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 391, + 163, + 459, + 177 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 391, + 163, + 459, + 177 + ], + "spans": [ + { + "bbox": [ + 391, + 163, + 459, + 177 + ], + "type": "text", + "content": "Hervé Jégou*" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 169, + 178, + 279, + 192 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 169, + 178, + 279, + 192 + ], + "spans": [ + { + "bbox": [ + 169, + 178, + 279, + 192 + ], + "type": "text", + "content": "*Meta AI / FAIR Paris" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 321, + 178, + 428, + 194 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 178, + 428, + 194 + ], + "spans": [ + { + "bbox": [ + 321, + 178, + 428, + 194 + ], + "type": "text", + "content": "†Sorbonne University" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "spans": [ + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 238, + 288, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 238, + 288, + 357 + ], + "spans": [ + { + "bbox": [ + 46, + 238, + 288, + 357 + ], + "type": "text", + "content": "We introduce submodel co-training, a regularization method related to co-training, self-distillation and stochastic depth. Given a neural network to be trained, for each sample we implicitly instantiate two altered networks, \"submodels\", with stochastic depth: we activate only a subset of the layers. Each network serves as a soft teacher to the other, by providing a loss that complements the regular loss provided by the one-hot label. Our approach, dubbed \"cosub\", uses a single set of weights, and does not involve a pre-trained external model or temporal averaging." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 46, + 357, + 287, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 357, + 287, + 453 + ], + "spans": [ + { + "bbox": [ + 46, + 357, + 287, + 453 + ], + "type": "text", + "content": "Experimentally, we show that submodel co-training is effective to train backbones for recognition tasks such as image classification and semantic segmentation. Our approach is compatible with multiple architectures, including RegNet, ViT, PiT, XCiT, Swin and ConvNext. Our training strategy improves their results in comparable settings. For instance, a ViT-B pretrained with cosub on ImageNet-21k obtains " + }, + { + "bbox": [ + 46, + 357, + 287, + 453 + ], + "type": "inline_equation", + "content": "87.4\\%" + }, + { + "bbox": [ + 46, + 357, + 287, + 453 + ], + "type": "text", + "content": " top-1 acc. @448 on ImageNet-val." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 47, + 466, + 128, + 478 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 466, + 128, + 478 + ], + "spans": [ + { + "bbox": [ + 47, + 466, + 128, + 478 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 46, + 486, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 486, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 486, + 288, + 715 + ], + "type": "text", + "content": "Although the fundamental ideas of deep trainable neural networks have been around for decades, only recently have barriers been removed to allow breakthroughs in successfully training deep neural architectures in practice. Many of these barriers are related to non-convex optimization in one way or another, which is central to the success of modern neural networks. The optimization challenges have been addressed from multiple angles in the literature. First, modern architectures are designed to facilitate the optimization of very deep networks. An exceptionally successful design principle is using residual connections [24, 25]. Although this does not change the expressiveness of the functions that the network can implement, the improved gradient flow alleviates, to some extent, the difficulties of optimizing very deep networks. Another key element to the optimization is the importance of data, revealed by the step-change in visual recognition performance resulting from the ImageNet dataset [11], and the popularization of transfer learning with pre-training on large datasets [39, 58]." + } + ] + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 360, + 209, + 494, + 419 + ], + "blocks": [ + { + "bbox": [ + 360, + 209, + 494, + 419 + ], + "lines": [ + { + "bbox": [ + 360, + 209, + 494, + 419 + ], + "spans": [ + { + "bbox": [ + 360, + 209, + 494, + 419 + ], + "type": "image", + "image_path": "6605072b50b4cf03d4e74edaf5c0b4bb06c32e1a462ecca593c4342f23471d3c.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 426, + 547, + 466 + ], + "lines": [ + { + "bbox": [ + 305, + 426, + 547, + 466 + ], + "spans": [ + { + "bbox": [ + 305, + 426, + 547, + 466 + ], + "type": "text", + "content": "Figure 1. Co-training of submodels (cosub): for each image, two sub-models are sampled by randomly dropping layers of the full model. The training signal for each submodel mixes the cross-entropy loss from the image label with a self-distillation loss obtained from the other submodel." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 486, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 486, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 486, + 547, + 715 + ], + "type": "text", + "content": "However, even when (pre-)trained with millions of images, recent deep networks with millions if not billions of parameters, are still heavily overparameterized. Traditional regularization like weight decay, dropout [46], or label smoothing [47] are limited in their ability to address this issue. Data-augmentation strategies, including those mixing different images like Mixup [61] and CutMix [60], have proven to provide a complementary data-driven form of regularization. More recently, multiple works propose to resort to self-supervised pre-training. These approaches rely on a proxy objective that generally provides more supervision signal than the one available from labels, like in recent (masked) auto-encoders [5,16,22], which were popular in the early deep learning literature [7,19,27]. Similarly, contrastive approaches [23] or self-distillation [9] provide a richer supervision less prone to supervision collapse [12]. Overall, self-supervised learning makes it possible to learn larger models with less data, possibly reducing the need of a pre-training stage [15]." + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "11701" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 193 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 193 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 193 + ], + "type": "text", + "content": "Distillation is a complementary approach to improve optimization. Distillation techniques were originally developed to transfer knowledge from a teacher model to a student model [4, 28], allowing the student to improve over learning from the data directly. In contrast to traditional distillation, co-distillation does not require pre-training a (strong) teacher. Instead, a pool of models supervise each other. Practically, it faces several limitations, including the difficulty of jointly training more than two students for complexity reasons, as it involves duplicating the weights." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 197, + 289, + 353 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 197, + 289, + 353 + ], + "spans": [ + { + "bbox": [ + 46, + 197, + 289, + 353 + ], + "type": "text", + "content": "In this paper, we propose a practical way to enable cotraining for a very large number of students. We consider a single target model to be trained, and we instantiate two submodels on-the-fly, simply by layerwise dropout [20,31]. This gives us two neural networks through which we can backpropagate to the shared parameters of the target model. In addition to the regular training loss, each submodel serves as a teacher to the other, which provides an additional supervision signal ensuring the consistency across the submodels. Our approach is illustrated in Figure 1: the parameter " + }, + { + "bbox": [ + 46, + 197, + 289, + 353 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 46, + 197, + 289, + 353 + ], + "type": "text", + "content": " controls the importance of the co-training loss compared to the label loss, and our experiments show that it significantly increases the final model accuracy." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 356, + 289, + 452 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 356, + 289, + 452 + ], + "spans": [ + { + "bbox": [ + 46, + 356, + 289, + 452 + ], + "type": "text", + "content": "This co-training across different submodels, which we refer to as cosub, can be regarded as a massive co-training between " + }, + { + "bbox": [ + 46, + 356, + 289, + 452 + ], + "type": "inline_equation", + "content": "2^{L}" + }, + { + "bbox": [ + 46, + 356, + 289, + 452 + ], + "type": "text", + "content": " models that share a common set of parameters, where " + }, + { + "bbox": [ + 46, + 356, + 289, + 452 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 46, + 356, + 289, + 452 + ], + "type": "text", + "content": " is the number of layers in the target architecture. The target model can be interpreted as the expectation of all models. With a layer drop-rate set to 0.5, for instance for a ViT-H model, all submodels are equiprobable, and then it amounts to averaging the weights of " + }, + { + "bbox": [ + 46, + 356, + 289, + 452 + ], + "type": "inline_equation", + "content": "2^{2 \\times 32}" + }, + { + "bbox": [ + 46, + 356, + 289, + 452 + ], + "type": "text", + "content": " models." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 456, + 247, + 467 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 456, + 247, + 467 + ], + "spans": [ + { + "bbox": [ + 47, + 456, + 247, + 467 + ], + "type": "text", + "content": "Our contributions can be summarized as follows:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 58, + 474, + 286, + 713 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 58, + 474, + 286, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 474, + 286, + 557 + ], + "spans": [ + { + "bbox": [ + 58, + 474, + 286, + 557 + ], + "type": "text", + "content": "- We introduce a novel training approach for deep neural networks: we co-train submodels. This significantly improves the training of most models, establishing the new state of the art in multiple cases. For instance, after pre-training ViT-B onImagenet-21k and fine-tuning it at resolution 448, we obtain " + }, + { + "bbox": [ + 58, + 474, + 286, + 557 + ], + "type": "inline_equation", + "content": "87.4\\%" + }, + { + "bbox": [ + 58, + 474, + 286, + 557 + ], + "type": "text", + "content": " top-1 accuracy onImagenet-val." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 58, + 558, + 286, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 558, + 286, + 593 + ], + "spans": [ + { + "bbox": [ + 58, + 558, + 286, + 593 + ], + "type": "text", + "content": "- We provide an efficient implementation to subsample models on the fly. It is a simple yet effective variation of stochastic depth [31] to drop residual blocks." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 58, + 594, + 286, + 642 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 594, + 286, + 642 + ], + "spans": [ + { + "bbox": [ + 58, + 594, + 286, + 642 + ], + "type": "text", + "content": "- We provide multiple analyses and ablations. Noticeably, we show that our submodels are effective models by themselves even with significant trimming, similar to LayerDrop [20] in natural language processing." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 58, + 642, + 286, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 642, + 286, + 689 + ], + "spans": [ + { + "bbox": [ + 58, + 642, + 286, + 689 + ], + "type": "text", + "content": "- We validate our approach on multiple architectures (like ViT, ResNet, RegNet, PiT, XCiT, Swin, ConvNext), both for image classification -trained from scratch or with transfer-, and semantic segmentation." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 58, + 689, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 689, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 58, + 689, + 286, + 713 + ], + "type": "text", + "content": "- We will share models/code for reproducibility in the DeiT repository." + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 71, + 390, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 71, + 390, + 83 + ], + "spans": [ + { + "bbox": [ + 306, + 71, + 390, + 83 + ], + "type": "text", + "content": "2. Related work" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 95, + 545, + 287 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 95, + 545, + 287 + ], + "spans": [ + { + "bbox": [ + 304, + 95, + 545, + 287 + ], + "type": "text", + "content": "Knowledge distillation. Originally, distillation was introduced as a way to train a model such that it reproduces the performance of another model [4, 28]. The typical use-case is to improve the quality of a relatively small model by leveraging a strong teacher, whose complexity may be prohibitive for a practical deployment. The teacher's soft labels have a similar effect as label smoothing [59]. As shown by Wei et al. [53], the teacher's supervision takes into account the effects of the data augmentation, which sometimes causes a misalignment between the real label and the image. Knowledge distillation can transfer inductive biases [1] in a soft way in a student model when using a teacher model these biases are enforced in a hard way. Touvron et al. [50] proposed a variant of distillation adapted to Vision Transformer (ViT), showing the effectiveness of using a ConvNet teacher for a ViT student." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 292, + 545, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 292, + 545, + 399 + ], + "spans": [ + { + "bbox": [ + 304, + 292, + 545, + 399 + ], + "type": "text", + "content": "Mixture models. Ensembling has a long history that we can trace back to the origins of statistics [37] and the possibility of improving the precision of measurements with multiple observations. In machine learning, bagging [8] combines multiple weak classifiers to produce a strong one. This idea is naturally extended to neural networks, where it can offer improved stability or accuracy, or other properties like anytime inference [44]. A mixture model can also be seen as a larger model with an internal parallel structure." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 404, + 545, + 536 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 404, + 545, + 536 + ], + "spans": [ + { + "bbox": [ + 304, + 404, + 545, + 536 + ], + "type": "text", + "content": "Co-distillation. At the interface of mixture models and distillation, the concept of co-distillation [62, 64] does not require a prior teacher: the mixture serves as the teacher to all the networks in the mixture. The models are jointly optimized, which leads to improved accuracy of the individual models [2]. This can be regarded as a form of collaborative learning between the different elements of the mixture [45]. Compared to training only one model, co-distillation involves training two or more models, each requiring storage for weights and activation, and computing backward and forward passes." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 541, + 545, + 661 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 541, + 545, + 661 + ], + "spans": [ + { + "bbox": [ + 304, + 541, + 545, + 661 + ], + "type": "text", + "content": "Exponential moving average teacher. A special form of teacher is a model obtained from the student model itself. Tarvainen and Valpila [49] show that such a model obtained by temporal averaging of intermediate checkpoints during the training is an effective teacher, that can be obtained with an exponential moving average (EMA). This idea has also been adopted in self-supervised training with learning schemes like DINO [9]. This method involves storing a copy of the weights corresponding to the temporal averaging carried out by the EMA model." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "content": "Dropout and model populations. Dropout [46] is an effective way to regularize models. With residual architectures, one effective way to train deeper architecture is stochastic depth [31]. It reduces the size of a network at" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "11702" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 78, + 79, + 126, + 154 + ], + "blocks": [ + { + "bbox": [ + 78, + 79, + 126, + 154 + ], + "lines": [ + { + "bbox": [ + 78, + 79, + 126, + 154 + ], + "spans": [ + { + "bbox": [ + 78, + 79, + 126, + 154 + ], + "type": "image", + "image_path": "827e24b5c3a9e8f3960ee8f180d6c08bfd939a86fcf599f96570ddf5a723151d.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 76, + 165, + 124, + 175 + ], + "lines": [ + { + "bbox": [ + 76, + 165, + 124, + 175 + ], + "spans": [ + { + "bbox": [ + 76, + 165, + 124, + 175 + ], + "type": "text", + "content": "(a) supervised" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 155, + 79, + 225, + 154 + ], + "blocks": [ + { + "bbox": [ + 155, + 79, + 225, + 154 + ], + "lines": [ + { + "bbox": [ + 155, + 79, + 225, + 154 + ], + "spans": [ + { + "bbox": [ + 155, + 79, + 225, + 154 + ], + "type": "image", + "image_path": "0577d9e58a051c00e478e494123d045453ed591c00b97aa9463d301198da72c3.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 163, + 164, + 212, + 183 + ], + "lines": [ + { + "bbox": [ + 163, + 164, + 212, + 183 + ], + "spans": [ + { + "bbox": [ + 163, + 164, + 212, + 183 + ], + "type": "text", + "content": "(b) knowledge distillation" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 252, + 79, + 324, + 155 + ], + "blocks": [ + { + "bbox": [ + 252, + 79, + 324, + 155 + ], + "lines": [ + { + "bbox": [ + 252, + 79, + 324, + 155 + ], + "spans": [ + { + "bbox": [ + 252, + 79, + 324, + 155 + ], + "type": "image", + "image_path": "52415c6721f74905e1f9e2aa415e28ad94c2ee870d6e251fa66f3018e5d87f32.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 257, + 164, + 320, + 175 + ], + "lines": [ + { + "bbox": [ + 257, + 164, + 320, + 175 + ], + "spans": [ + { + "bbox": [ + 257, + 164, + 320, + 175 + ], + "type": "text", + "content": "(c) self-distillation" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 341, + 76, + 425, + 156 + ], + "blocks": [ + { + "bbox": [ + 341, + 76, + 425, + 156 + ], + "lines": [ + { + "bbox": [ + 341, + 76, + 425, + 156 + ], + "spans": [ + { + "bbox": [ + 341, + 76, + 425, + 156 + ], + "type": "image", + "image_path": "760887e7f56f7476c1a24653682b1296bc185273378ec01af654b59712b33505.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 362, + 164, + 411, + 175 + ], + "lines": [ + { + "bbox": [ + 362, + 164, + 411, + 175 + ], + "spans": [ + { + "bbox": [ + 362, + 164, + 411, + 175 + ], + "type": "text", + "content": "(d) co-training" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 192, + 362, + 277 + ], + "lines": [ + { + "bbox": [ + 46, + 192, + 362, + 277 + ], + "spans": [ + { + "bbox": [ + 46, + 192, + 362, + 277 + ], + "type": "text", + "content": "Figure 2. Brief summary of related works and of submodel co-training (cosub). (a) supervised baseline: the supervision is provided as a one-hot label. (b) Knowledge distillation [28]: a teacher provides a (soft) target. A simple yet effective variant [50] combines the predicted label of the teacher with the image label. KD requires a pre-existing teacher that must be stored for inference. (c) In self-distillation, the teacher is obtained from the model itself, typically by model averaging. (d) Co-training involves two distinct models that serves as teacher to each other. (e) Submodel co-training generates on-the-fly two distinct models (amongst " + }, + { + "bbox": [ + 46, + 192, + 362, + 277 + ], + "type": "inline_equation", + "content": "2^{L}" + }, + { + "bbox": [ + 46, + 192, + 362, + 277 + ], + "type": "text", + "content": ") for each sample, which serves a teacher to each other. We only need to store a single set of weights for the model and optimizer because the random selections " + }, + { + "bbox": [ + 46, + 192, + 362, + 277 + ], + "type": "inline_equation", + "content": "\\theta_{1} = \\mathcal{T}(\\theta)" + }, + { + "bbox": [ + 46, + 192, + 362, + 277 + ], + "type": "text", + "content": " (resp., " + }, + { + "bbox": [ + 46, + 192, + 362, + 277 + ], + "type": "inline_equation", + "content": "\\theta_{2}" + }, + { + "bbox": [ + 46, + 192, + 362, + 277 + ], + "type": "text", + "content": ") of submodels allow us to back-propagate on " + }, + { + "bbox": [ + 46, + 192, + 362, + 277 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 46, + 192, + 362, + 277 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 430, + 75, + 543, + 160 + ], + "blocks": [ + { + "bbox": [ + 430, + 75, + 543, + 160 + ], + "lines": [ + { + "bbox": [ + 430, + 75, + 543, + 160 + ], + "spans": [ + { + "bbox": [ + 430, + 75, + 543, + 160 + ], + "type": "image", + "image_path": "d966bd13264237f7aecd65717121e981cfc89e5e8bfa8b9d0122c9a9f8cebd80.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 452, + 163, + 535, + 175 + ], + "lines": [ + { + "bbox": [ + 452, + 163, + 535, + 175 + ], + "spans": [ + { + "bbox": [ + 452, + 163, + 535, + 175 + ], + "type": "text", + "content": "(e) submodel co-training" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 365, + 185, + 545, + 258 + ], + "blocks": [ + { + "bbox": [ + 365, + 185, + 545, + 258 + ], + "lines": [ + { + "bbox": [ + 365, + 185, + 545, + 258 + ], + "spans": [ + { + "bbox": [ + 365, + 185, + 545, + 258 + ], + "type": "table", + "html": "
methodmodel paramscompute*
weightsoptimizer
(a) supervised×1×1×1
(b) KD†×2×1×1.33
(c) mean teacher×2×1×1.33
(d) co-training×2×2×2
(e) Cosub (ours)×1×1×2
", + "image_path": "8fda16f12e1ae55dc6aaba2940166446304a7a36da955f09397b95af7e9cd754.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 365, + 258, + 545, + 277 + ], + "lines": [ + { + "bbox": [ + 365, + 258, + 545, + 277 + ], + "spans": [ + { + "bbox": [ + 365, + 258, + 545, + 277 + ], + "type": "text", + "content": "\\*backward pass assumed " + }, + { + "bbox": [ + 365, + 258, + 545, + 277 + ], + "type": "inline_equation", + "content": "2\\times" + }, + { + "bbox": [ + 365, + 258, + 545, + 277 + ], + "type": "text", + "content": " the complexity of forward " + }, + { + "bbox": [ + 365, + 258, + 545, + 277 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 365, + 258, + 545, + 277 + ], + "type": "text", + "content": " requires a trained teacher model" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 293, + 287, + 400 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 293, + 287, + 400 + ], + "spans": [ + { + "bbox": [ + 46, + 293, + 287, + 400 + ], + "type": "text", + "content": "training time by dropping residual blocks during training. The initial goal of this approach was to improve the training of deep network. However, there are other implications: in natural language processing, Fan et al. show that the counterpart of stochastic depth, namely LayerDrop [20], is effective to reduce the transformer depth on demand. Interestingly, submodels extracted from a model trained with LayerDrop are stronger than those trained with the full target depth with a proper selection strategy for layers." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 46, + 402, + 287, + 508 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 402, + 287, + 508 + ], + "spans": [ + { + "bbox": [ + 46, + 402, + 287, + 508 + ], + "type": "text", + "content": "In our paper, we regard stochastic depth as a regularization technique, but at the same time we adopt the point of view of LayerDrop, i.e., an effective way to train a population of submodels that share parameters, or entire layers in our case. From that viewpoint, stochastic depth amounts to training " + }, + { + "bbox": [ + 46, + 402, + 287, + 508 + ], + "type": "inline_equation", + "content": "2^{L}" + }, + { + "bbox": [ + 46, + 402, + 287, + 508 + ], + "type": "text", + "content": " distinct models. This is different from population training as involved in network space design [41], whose goal is related to network architecture search [18], which aims at optimizing the architecture itself." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 46, + 510, + 287, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 510, + 287, + 605 + ], + "spans": [ + { + "bbox": [ + 46, + 510, + 287, + 605 + ], + "type": "text", + "content": "More recently, model soups [56] are models obtained by averaging the weights of multiple models finetuned with different hyperparameter configurations. The authors show that it often improves accuracy and robustness. We point out that stochastic depth can be regarded as a special form of model soup over the entire population of submodels that we can instantiate with stochastic depth. We further discuss this relationship in the next section." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 47, + 616, + 174, + 630 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 616, + 174, + 630 + ], + "spans": [ + { + "bbox": [ + 47, + 616, + 174, + 630 + ], + "type": "text", + "content": "3. Submodel co-training" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 46, + 637, + 287, + 673 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 637, + 287, + 673 + ], + "spans": [ + { + "bbox": [ + 46, + 637, + 287, + 673 + ], + "type": "text", + "content": "In this section, we present our cosub approach, for constraining submodels. We consider a neural network " + }, + { + "bbox": [ + 46, + 637, + 287, + 673 + ], + "type": "inline_equation", + "content": "f_{\\theta}" + }, + { + "bbox": [ + 46, + 637, + 287, + 673 + ], + "type": "text", + "content": " parameterized by learnable parameters " + }, + { + "bbox": [ + 46, + 637, + 287, + 673 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 46, + 637, + 287, + 673 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "text", + "content": "Submodel instantiation: model augmentation operator. We first define a model augmentation operator " + }, + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{T}" + }, + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "text", + "content": ". For a given neural network, it provides a set of parameters " + }, + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\theta^{\\prime} =" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 293, + 545, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 293, + 545, + 389 + ], + "spans": [ + { + "bbox": [ + 304, + 293, + 545, + 389 + ], + "type": "inline_equation", + "content": "\\mathcal{T}(\\theta, R)" + }, + { + "bbox": [ + 304, + 293, + 545, + 389 + ], + "type": "text", + "content": " that allow us to define variations " + }, + { + "bbox": [ + 304, + 293, + 545, + 389 + ], + "type": "inline_equation", + "content": "f_{\\theta'}" + }, + { + "bbox": [ + 304, + 293, + 545, + 389 + ], + "type": "text", + "content": " of the function " + }, + { + "bbox": [ + 304, + 293, + 545, + 389 + ], + "type": "inline_equation", + "content": "f_{\\theta}" + }, + { + "bbox": [ + 304, + 293, + 545, + 389 + ], + "type": "text", + "content": " by drawing a random variable " + }, + { + "bbox": [ + 304, + 293, + 545, + 389 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 304, + 293, + 545, + 389 + ], + "type": "text", + "content": ". The model augmentation is such that " + }, + { + "bbox": [ + 304, + 293, + 545, + 389 + ], + "type": "inline_equation", + "content": "f_{\\theta}" + }, + { + "bbox": [ + 304, + 293, + 545, + 389 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 293, + 545, + 389 + ], + "type": "inline_equation", + "content": "f_{\\theta'}" + }, + { + "bbox": [ + 304, + 293, + 545, + 389 + ], + "type": "text", + "content": " share parameters, hence any update on " + }, + { + "bbox": [ + 304, + 293, + 545, + 389 + ], + "type": "inline_equation", + "content": "\\theta'" + }, + { + "bbox": [ + 304, + 293, + 545, + 389 + ], + "type": "text", + "content": " modifies " + }, + { + "bbox": [ + 304, + 293, + 545, + 389 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 304, + 293, + 545, + 389 + ], + "type": "text", + "content": " and therefore " + }, + { + "bbox": [ + 304, + 293, + 545, + 389 + ], + "type": "inline_equation", + "content": "f_{\\theta}" + }, + { + "bbox": [ + 304, + 293, + 545, + 389 + ], + "type": "text", + "content": " accordingly. A simple way to define " + }, + { + "bbox": [ + 304, + 293, + 545, + 389 + ], + "type": "inline_equation", + "content": "\\mathcal{T}" + }, + { + "bbox": [ + 304, + 293, + 545, + 389 + ], + "type": "text", + "content": " is to replace some parameters by zeros, which corresponds to dropout [46]. In this paper, we focus on stochastic depth, which has interesting connections with model averaging [56]." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 393, + 545, + 417 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 393, + 545, + 417 + ], + "spans": [ + { + "bbox": [ + 304, + 393, + 545, + 417 + ], + "type": "text", + "content": "Overview. For a given training sample " + }, + { + "bbox": [ + 304, + 393, + 545, + 417 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 304, + 393, + 545, + 417 + ], + "type": "text", + "content": " within a batch, the training is as follows (see also Figure 1):" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 419, + 545, + 574 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 313, + 419, + 514, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 419, + 514, + 430 + ], + "spans": [ + { + "bbox": [ + 313, + 419, + 514, + 430 + ], + "type": "text", + "content": "1. We first data-augment the image, producing " + }, + { + "bbox": [ + 313, + 419, + 514, + 430 + ], + "type": "inline_equation", + "content": "\\hat{x}" + }, + { + "bbox": [ + 313, + 419, + 514, + 430 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 431, + 545, + 466 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 431, + 545, + 466 + ], + "spans": [ + { + "bbox": [ + 313, + 431, + 545, + 466 + ], + "type": "text", + "content": "2. The image is duplicated with the batch, effectively doubling the batch size, which hence contains two identical copies of each augmented image1." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 467, + 545, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 467, + 545, + 514 + ], + "spans": [ + { + "bbox": [ + 313, + 467, + 545, + 514 + ], + "type": "text", + "content": "3. We determine the stochastic depth pattern for each sample according to the target drop-rate " + }, + { + "bbox": [ + 313, + 467, + 545, + 514 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 313, + 467, + 545, + 514 + ], + "type": "text", + "content": ", which amounts to producing two functions " + }, + { + "bbox": [ + 313, + 467, + 545, + 514 + ], + "type": "inline_equation", + "content": "f_{\\theta_1}" + }, + { + "bbox": [ + 313, + 467, + 545, + 514 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 467, + 545, + 514 + ], + "type": "inline_equation", + "content": "f_{\\theta_2}" + }, + { + "bbox": [ + 313, + 467, + 545, + 514 + ], + "type": "text", + "content": ". See Section 4 for the details of this procedure." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 514, + 545, + 539 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 514, + 545, + 539 + ], + "spans": [ + { + "bbox": [ + 313, + 514, + 545, + 539 + ], + "type": "text", + "content": "4. The forward pass proceeds as usual: we compute the soft output labels " + }, + { + "bbox": [ + 313, + 514, + 545, + 539 + ], + "type": "inline_equation", + "content": "y_{1} = f_{\\theta_{1}}(\\hat{x})" + }, + { + "bbox": [ + 313, + 514, + 545, + 539 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 514, + 545, + 539 + ], + "type": "inline_equation", + "content": "y_{2} = f_{\\theta_{2}}(\\hat{x})" + }, + { + "bbox": [ + 313, + 514, + 545, + 539 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 313, + 539, + 545, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 539, + 545, + 574 + ], + "spans": [ + { + "bbox": [ + 313, + 539, + 545, + 574 + ], + "type": "text", + "content": "5. We compute the losses and the backwards pass accordingly. Note that the gradients on " + }, + { + "bbox": [ + 313, + 539, + 545, + 574 + ], + "type": "inline_equation", + "content": "\\theta_{1}" + }, + { + "bbox": [ + 313, + 539, + 545, + 574 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 539, + 545, + 574 + ], + "type": "inline_equation", + "content": "\\theta_{2}" + }, + { + "bbox": [ + 313, + 539, + 545, + 574 + ], + "type": "text", + "content": " are directly used to update " + }, + { + "bbox": [ + 313, + 539, + 545, + 574 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 313, + 539, + 545, + 574 + ], + "type": "text", + "content": ", as they are just subsets of " + }, + { + "bbox": [ + 313, + 539, + 545, + 574 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 313, + 539, + 545, + 574 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 304, + 580, + 545, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 580, + 545, + 664 + ], + "spans": [ + { + "bbox": [ + 304, + 580, + 545, + 664 + ], + "type": "text", + "content": "Loss. Each submodel is trained using a weighted average of (i) the standard binary cross-entropy loss obtained from the image label " + }, + { + "bbox": [ + 304, + 580, + 545, + 664 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 304, + 580, + 545, + 664 + ], + "type": "text", + "content": ", and (ii) a binary cross-entropy loss w.r.t. the soft-labels computed for the the same image by the other submodel. The respective weight of the standard binary cross-entropy loss " + }, + { + "bbox": [ + 304, + 580, + 545, + 664 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{label}}" + }, + { + "bbox": [ + 304, + 580, + 545, + 664 + ], + "type": "text", + "content": " versus the cosub loss " + }, + { + "bbox": [ + 304, + 580, + 545, + 664 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{cosub}}" + }, + { + "bbox": [ + 304, + 580, + 545, + 664 + ], + "type": "text", + "content": " is controlled by the hyper-parameter " + }, + { + "bbox": [ + 304, + 580, + 545, + 664 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 304, + 580, + 545, + 664 + ], + "type": "text", + "content": ", as" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 355, + 673, + 545, + 685 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 355, + 673, + 545, + 685 + ], + "spans": [ + { + "bbox": [ + 355, + 673, + 545, + 685 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {t o t}} = \\lambda \\mathcal {L} _ {\\text {l a b e l}} + (1 - \\lambda) \\mathcal {L} _ {\\text {c o s u b}}. \\tag {1}", + "image_path": "6c3cf832c50f194a5f439b210eda3398a15a7059210dcbf31c04e17100470d7e.jpg" + } + ] + } + ], + "index": 28 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 306, + 693, + 545, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 693, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 693, + 545, + 713 + ], + "type": "text", + "content": "Different data-augmentations did not provide any advantage in our preliminary experiments. We simply duplicate the batch for simplicity." + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "11703" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 161, + 83 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 161, + 83 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 161, + 83 + ], + "type": "text", + "content": "In details, the loss writes as" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 90, + 287, + 147 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 90, + 287, + 147 + ], + "spans": [ + { + "bbox": [ + 57, + 90, + 287, + 147 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} _ {\\mathrm {t o t}} = \\lambda \\left(\\frac {\\mathcal {L} (y _ {1} , y) + \\mathcal {L} (y _ {2} , y)}{2}\\right) \\\\ + (1 - \\lambda) \\left(\\frac {\\mathcal {L} \\left(y _ {1} , \\mathrm {s g} \\left(y _ {2}\\right)\\right) + \\mathcal {L} \\left(y _ {2} , \\mathrm {s g} \\left(y _ {1}\\right)\\right)}{2}\\right), \\quad (2) \\\\ \\end{array}", + "image_path": "8c37eca49fedac5e72932c5983796d28be304018ca9f495a8f2a7b27691efea6.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 153, + 287, + 202 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 153, + 287, + 202 + ], + "spans": [ + { + "bbox": [ + 47, + 153, + 287, + 202 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 153, + 287, + 202 + ], + "type": "inline_equation", + "content": "\\mathcal{L}(y, y')" + }, + { + "bbox": [ + 47, + 153, + 287, + 202 + ], + "type": "text", + "content": " is either the binary cross-entropy (BCE) or a cross-entropy (CE) loss. Importantly, when applying this loss, we do not back-propagate on the second term " + }, + { + "bbox": [ + 47, + 153, + 287, + 202 + ], + "type": "inline_equation", + "content": "y'" + }, + { + "bbox": [ + 47, + 153, + 287, + 202 + ], + "type": "text", + "content": ". This is indicated by the stop-gradient operator " + }, + { + "bbox": [ + 47, + 153, + 287, + 202 + ], + "type": "inline_equation", + "content": "\\mathrm{sg}(\\cdot)" + }, + { + "bbox": [ + 47, + 153, + 287, + 202 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 205, + 289, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 205, + 289, + 445 + ], + "spans": [ + { + "bbox": [ + 47, + 205, + 289, + 445 + ], + "type": "text", + "content": "Discussion. The submodel instantiation provides a model that is, by itself, a valid neural network model. Co-training submodels is a way to enforce that all such submodels produce a consistent output. Amongst these submodels, a very special case is when we retain all residual blocks, which is the model that we primarily intend to use at inference time. Note that Fan et al. [20] show, in an NLP context, that submodels extracted from a deeper model are superior, for a given target depth. With proper rescaling of the residual branches, this specific submodel with all blocks activated can be seen as the expectation over all models. This can be shown as follows: let us consider an extra scalar parameter " + }, + { + "bbox": [ + 47, + 205, + 289, + 445 + ], + "type": "inline_equation", + "content": "s_l" + }, + { + "bbox": [ + 47, + 205, + 289, + 445 + ], + "type": "text", + "content": " associated with a given residual block " + }, + { + "bbox": [ + 47, + 205, + 289, + 445 + ], + "type": "inline_equation", + "content": "r_l(x)" + }, + { + "bbox": [ + 47, + 205, + 289, + 445 + ], + "type": "text", + "content": ", which we use as a multiplicative factor on output of each residual: " + }, + { + "bbox": [ + 47, + 205, + 289, + 445 + ], + "type": "inline_equation", + "content": "s_l = 1" + }, + { + "bbox": [ + 47, + 205, + 289, + 445 + ], + "type": "text", + "content": " if the residual block " + }, + { + "bbox": [ + 47, + 205, + 289, + 445 + ], + "type": "inline_equation", + "content": "r_l" + }, + { + "bbox": [ + 47, + 205, + 289, + 445 + ], + "type": "text", + "content": " is included in the submodel, " + }, + { + "bbox": [ + 47, + 205, + 289, + 445 + ], + "type": "inline_equation", + "content": "s_l = 0" + }, + { + "bbox": [ + 47, + 205, + 289, + 445 + ], + "type": "text", + "content": " otherwise. Each submodel is fully parameterized by a binary vector " + }, + { + "bbox": [ + 47, + 205, + 289, + 445 + ], + "type": "inline_equation", + "content": "s = (s_1, \\dots, s_L)" + }, + { + "bbox": [ + 47, + 205, + 289, + 445 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 47, + 205, + 289, + 445 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 47, + 205, + 289, + 445 + ], + "type": "text", + "content": " is the total number of layers. Therefore all submodels have the same parameters " + }, + { + "bbox": [ + 47, + 205, + 289, + 445 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 47, + 205, + 289, + 445 + ], + "type": "text", + "content": ", and only differ by " + }, + { + "bbox": [ + 47, + 205, + 289, + 445 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 47, + 205, + 289, + 445 + ], + "type": "text", + "content": " indicating the zeroed residual blocks. The weight expectation " + }, + { + "bbox": [ + 47, + 205, + 289, + 445 + ], + "type": "inline_equation", + "content": "\\bar{\\theta}" + }, + { + "bbox": [ + 47, + 205, + 289, + 445 + ], + "type": "text", + "content": " is hence" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 85, + 453, + 287, + 484 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 85, + 453, + 287, + 484 + ], + "spans": [ + { + "bbox": [ + 85, + 453, + 287, + 484 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\bar {\\theta} = \\mathbb {E} _ {s \\in [ 0, 1 ] ^ {L}} [ \\theta ] = \\theta , \\\\ \\bar {s} = \\mathbb {E} _ {s \\in [ 0, 1 ] ^ {L}} [ s ] = [ 1 - \\tau , \\dots , 1 - \\tau ] ^ {\\top}, \\tag {3} \\\\ \\end{array}", + "image_path": "9f26b0903dc8aa5242e4481cc41663996a957840a31d21b06bbfa3eb6ef234a5.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 491, + 287, + 564 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 491, + 287, + 564 + ], + "spans": [ + { + "bbox": [ + 47, + 491, + 287, + 564 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 491, + 287, + 564 + ], + "type": "inline_equation", + "content": "\\bar{s}" + }, + { + "bbox": [ + 47, + 491, + 287, + 564 + ], + "type": "text", + "content": " is the scaling factor used in stochastic depth. Under a uniform distribution of submodels that is obtained when the stochastic drop rate " + }, + { + "bbox": [ + 47, + 491, + 287, + 564 + ], + "type": "inline_equation", + "content": "\\tau = 0.5" + }, + { + "bbox": [ + 47, + 491, + 287, + 564 + ], + "type": "text", + "content": ", the inference-time model is exactly the average of all " + }, + { + "bbox": [ + 47, + 491, + 287, + 564 + ], + "type": "inline_equation", + "content": "2^{L}" + }, + { + "bbox": [ + 47, + 491, + 287, + 564 + ], + "type": "text", + "content": " models. While stochastic depth is often regarded as a regularization technique, this averaging interpretation relates it to the recent model soup [56]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 574, + 192, + 588 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 574, + 192, + 588 + ], + "spans": [ + { + "bbox": [ + 47, + 574, + 192, + 588 + ], + "type": "text", + "content": "4. Efficient stochastic depth" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 594, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 594, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 47, + 594, + 288, + 715 + ], + "type": "text", + "content": "We revisit the original stochastic depth [31] formulation in order enable an efficient implementation, which we will release for PyTorch [40]. Instantiating a submodel for a sample amounts to selecting a subset of residual layers of the model and performing training on these. This can be implemented using the stochastic depth (SD) approach, whose objective was initially to improve the training of very deep networks. In stochastic depth, for each sample and each layers of the network, we select whether the layers will be dropped or not with a certain probability " + }, + { + "bbox": [ + 47, + 594, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 47, + 594, + 288, + 715 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 305, + 72, + 547, + 160 + ], + "blocks": [ + { + "bbox": [ + 305, + 72, + 547, + 160 + ], + "lines": [ + { + "bbox": [ + 305, + 72, + 547, + 160 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 547, + 160 + ], + "type": "image", + "image_path": "9707da9db87eb1046491989ddb46f98c5c976595da4accbd49ee490813dd094b.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 166, + 547, + 205 + ], + "lines": [ + { + "bbox": [ + 305, + 166, + 547, + 205 + ], + "spans": [ + { + "bbox": [ + 305, + 166, + 547, + 205 + ], + "type": "text", + "content": "Figure 3. Efficient implementation of stochastic depth, using the permute-select approach. In this example, we drop the residuals for samples 3 and 6, corresponding to a drop rate " + }, + { + "bbox": [ + 305, + 166, + 547, + 205 + ], + "type": "inline_equation", + "content": "\\tau = 0.25" + }, + { + "bbox": [ + 305, + 166, + 547, + 205 + ], + "type": "text", + "content": ". For a drop rate higher than 0.1, the overhead of the approach is negligible in terms of memory and compute." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 219, + 545, + 278 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 219, + 545, + 278 + ], + "spans": [ + { + "bbox": [ + 304, + 219, + 545, + 278 + ], + "type": "text", + "content": "In practice, e.g. in the timm library [54], stochastic depth is implemented by masking with zeros the residuals added for a given sample of a batch. This is not efficient: this naive approach performs the computations for a residual then throws it away, wasting computation." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 279, + 546, + 459 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 279, + 546, + 459 + ], + "spans": [ + { + "bbox": [ + 304, + 279, + 546, + 459 + ], + "type": "text", + "content": "Our efficient stochastic depth (ESD) approach addresses this issue, saving both memory and compute, and is illustrated in Figure 3. For each layer, given a batch size " + }, + { + "bbox": [ + 304, + 279, + 546, + 459 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 304, + 279, + 546, + 459 + ], + "type": "text", + "content": " and a drop rate " + }, + { + "bbox": [ + 304, + 279, + 546, + 459 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 304, + 279, + 546, + 459 + ], + "type": "text", + "content": ", we apply the layer to " + }, + { + "bbox": [ + 304, + 279, + 546, + 459 + ], + "type": "inline_equation", + "content": "B_{\\mathrm{keep}} = \\mathrm{round}(B \\times (1 - \\tau))" + }, + { + "bbox": [ + 304, + 279, + 546, + 459 + ], + "type": "text", + "content": " samples in the batch. In contrast to the original version of SD, our efficient version drops a fixed number of samples at each layer, where " + }, + { + "bbox": [ + 304, + 279, + 546, + 459 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 304, + 279, + 546, + 459 + ], + "type": "text", + "content": " is adjusted such that " + }, + { + "bbox": [ + 304, + 279, + 546, + 459 + ], + "type": "inline_equation", + "content": "B_{\\mathrm{keep}}" + }, + { + "bbox": [ + 304, + 279, + 546, + 459 + ], + "type": "text", + "content": " is an integer. In our experiments, this did not have an effect on the final accuracy of the models. Our efficient implementation proceeds as follows: (i) we apply a random permutation of the " + }, + { + "bbox": [ + 304, + 279, + 546, + 459 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 304, + 279, + 546, + 459 + ], + "type": "text", + "content": " samples, then (ii) we select the first " + }, + { + "bbox": [ + 304, + 279, + 546, + 459 + ], + "type": "inline_equation", + "content": "B_{\\mathrm{keep}}" + }, + { + "bbox": [ + 304, + 279, + 546, + 459 + ], + "type": "text", + "content": " samples. We then compute the residual function for the selected subset, then add the result onto the full batch using the built-in index_add function and scaling the result by " + }, + { + "bbox": [ + 304, + 279, + 546, + 459 + ], + "type": "inline_equation", + "content": "\\frac{1}{1 - \\tau}" + }, + { + "bbox": [ + 304, + 279, + 546, + 459 + ], + "type": "text", + "content": " to have the correct scaling as in the original SD formulation." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 462, + 545, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 462, + 545, + 581 + ], + "spans": [ + { + "bbox": [ + 304, + 462, + 545, + 581 + ], + "type": "text", + "content": "Discussion: progressive vs uniform rate. In the original paper [31], stochastic depth drops a layer with a probability that is linearly increasing during the forward pass: layers closer to the output has a higher chance to be dropped. However this strategy is limited with high drop rates, as later layers are excessively dropped. Touvron et al. [52] adopt a uniform rate drop per layer. It is as effective as the progressive rate with vision transformers, but it makes it possible to target a drop-rate greater than 0.5 on average. For this reason we adopt an uniform drop rate everywhere." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 582, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 582, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 582, + 547, + 714 + ], + "type": "text", + "content": "Our efficient stochastic depth implementation variant works both with uniform and progressive drop-rate. In the case of progressive, the effective batch size is decreasing during to the forward pass. One limitation is that our technique implies a quantization of the drop-rate, which can be problematic for small batch sizes. Therefore one must take care to properly verify that the drop-rate is not too severely quantized, in particular in the progressive case for which the drop-rate can be very high in the last layers. In the supplemental material (Appendix C), we discuss with more details the drop-rate quantization effect." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "11704" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 47, + 68, + 286, + 194 + ], + "blocks": [ + { + "bbox": [ + 47, + 68, + 286, + 194 + ], + "lines": [ + { + "bbox": [ + 47, + 68, + 286, + 194 + ], + "spans": [ + { + "bbox": [ + 47, + 68, + 286, + 194 + ], + "type": "image", + "image_path": "f52e69c2b34f9bd6846bc6b0ef914df1a935f850c1a0c262ab10a20143d0e470.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 198, + 289, + 266 + ], + "lines": [ + { + "bbox": [ + 46, + 198, + 289, + 266 + ], + "spans": [ + { + "bbox": [ + 46, + 198, + 289, + 266 + ], + "type": "text", + "content": "Figure 4. Cosub as population training: each submodel extracted by a transformation " + }, + { + "bbox": [ + 46, + 198, + 289, + 266 + ], + "type": "inline_equation", + "content": "\\mathcal{T}" + }, + { + "bbox": [ + 46, + 198, + 289, + 266 + ], + "type": "text", + "content": " is a valid neural network. Our cosub strategy can hence be regarded as co-training a large number of subnetworks. We plot the accuracy of the submodels as a function of the number of layers that we preserve. We drop layers with probability " + }, + { + "bbox": [ + 46, + 198, + 289, + 266 + ], + "type": "inline_equation", + "content": "\\tau = 0.2" + }, + { + "bbox": [ + 46, + 198, + 289, + 266 + ], + "type": "text", + "content": " for ViT-B and " + }, + { + "bbox": [ + 46, + 198, + 289, + 266 + ], + "type": "inline_equation", + "content": "\\tau = 0.45" + }, + { + "bbox": [ + 46, + 198, + 289, + 266 + ], + "type": "text", + "content": " for ViT-L. On average our method provides a significant boost in performance for the whole population of submodels extracted from the main model." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 52, + 279, + 286, + 324 + ], + "blocks": [ + { + "bbox": [ + 52, + 279, + 286, + 324 + ], + "lines": [ + { + "bbox": [ + 52, + 279, + 286, + 324 + ], + "spans": [ + { + "bbox": [ + 52, + 279, + 286, + 324 + ], + "type": "table", + "html": "
ModelViT-SViT-MViT-BViT-LViT-H
τ(0.05)(0.1)(0.2)(0.45)(0.6)
baseline81.482.583.784.584.9
baseline + cosub81.582.883.984.985.5
", + "image_path": "3f58449cc66ba136867a1ad66050621265394342767c650048bb84a81d54237a.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 404, + 128, + 418 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 404, + 128, + 418 + ], + "spans": [ + { + "bbox": [ + 47, + 404, + 128, + 418 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 426, + 287, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 426, + 287, + 498 + ], + "spans": [ + { + "bbox": [ + 46, + 426, + 287, + 498 + ], + "type": "text", + "content": "We evaluate our approach on residual architectures [24, 25], since they are readily compatible with the stochastic dropout. We take as our main reference the vanilla Vision Transformer introduced by Dosovitskiy et al. [14]. However, as shown in our experiments, our approach is effective with all the residual architectures that we have considered." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 511, + 208, + 525 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 511, + 208, + 525 + ], + "spans": [ + { + "bbox": [ + 47, + 511, + 208, + 525 + ], + "type": "text", + "content": "5.1. Baseline and training settings" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 532, + 287, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 532, + 287, + 616 + ], + "spans": [ + { + "bbox": [ + 46, + 532, + 287, + 616 + ], + "type": "text", + "content": "We adopt the state-of-the-art training procedure proposed in the DeiT III paper [51] as baseline for transformers architectures and the training procedure from Wightman et al. [55] for the convnets. All hyper-parameters are identical except onImagenet-21k, where the hyper-parameter " + }, + { + "bbox": [ + 46, + 532, + 287, + 616 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 46, + 532, + 287, + 616 + ], + "type": "text", + "content": " are adjusted depending on the training setting. We recapitulate the training hyper-parameters in Appendix A." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 617, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 617, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 617, + 288, + 715 + ], + "type": "text", + "content": "We additionally adopt and measure the impact of LayerDecay for the fine-tuning when transferring from Imagenet-21k to Imagenet-1k. This method slightly boosts the performance, as discussed later in this section, and in Table 10. This method was adopted in multiple recent papers and in particular for fine-tuning of self-supervised approaches [5, 22], yet the contribution of this fine-tuning ingredient was not quantitatively measured." + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 309, + 70, + 545, + 209 + ], + "blocks": [ + { + "bbox": [ + 46, + 332, + 288, + 373 + ], + "lines": [ + { + "bbox": [ + 46, + 332, + 288, + 373 + ], + "spans": [ + { + "bbox": [ + 46, + 332, + 288, + 373 + ], + "type": "text", + "content": "Table 1. Comparison of top-1 accuracy with ViT architecture trained with/without cosub at resolution 224 (800 epochs) onImagenet-1k. The improvement is not significant for the smaller architecture ViT-S. Cosub is gradually more effective when we increase the model size and the SD rate." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 309, + 70, + 545, + 209 + ], + "lines": [ + { + "bbox": [ + 309, + 70, + 545, + 209 + ], + "spans": [ + { + "bbox": [ + 309, + 70, + 545, + 209 + ], + "type": "table", + "html": "
ModelτOriginalBaseline+cosubΔ
Transformers
ViT-L [14]0.4576.584.584.9+0.4
CaiT-L24 [52]0.45-83.884.4+0.6
PiT-B [26]0.2582.083.884.1+0.3
XCiT-L12 [17]0.20-82.683.0+0.4
Swin-B [34]0.2083.582.983.3+0.4
Swin-L [34]0.45-80.884.0+3.2
Convnets
ResNet-50 [24]0.1076.280.280.3+0.1
ResNet-101 [24]0.2077.481.882.1+0.4
ResNet-152 [24]0.3078.382.483.1+0.7
RegNet-16GF [42]0.3080.482.983.8+0.9
", + "image_path": "2761f414f2b6cb09fe404b33e487ee0287e7becad429b48a223bebc0a2cdbf21.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 308, + 274, + 545, + 304 + ], + "blocks": [ + { + "bbox": [ + 305, + 211, + 547, + 261 + ], + "lines": [ + { + "bbox": [ + 305, + 211, + 547, + 261 + ], + "spans": [ + { + "bbox": [ + 305, + 211, + 547, + 261 + ], + "type": "text", + "content": "Table 2. Benefit of cosub for different architectures trained from scratch on Imagenet-1k at resolution 224. We report top-1 acc. for the supervised baseline and cosub, as well as results reported in the corresponding papers when available (trained with different settings). We have adjusted the stochastic-depth drop-rate (SD) hyper-parameter for each architecture." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 308, + 274, + 545, + 304 + ], + "lines": [ + { + "bbox": [ + 308, + 274, + 545, + 304 + ], + "spans": [ + { + "bbox": [ + 308, + 274, + 545, + 304 + ], + "type": "table", + "html": "
LossBCE-softBCE-hardCE-hard
Imagenet-val top-1 accuracy83.583.581.9
", + "image_path": "7361f96e908cc41e563cafcb4f06fe4616c70c7dc8a08eff7bc9ea1de1078513.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 305, + 545, + 335 + ], + "lines": [ + { + "bbox": [ + 306, + 305, + 545, + 335 + ], + "spans": [ + { + "bbox": [ + 306, + 305, + 545, + 335 + ], + "type": "text", + "content": "Table 3. Ablation on the loss for cosub with ViT-H trained at resolution " + }, + { + "bbox": [ + 306, + 305, + 545, + 335 + ], + "type": "inline_equation", + "content": "126\\times 126" + }, + { + "bbox": [ + 306, + 305, + 545, + 335 + ], + "type": "text", + "content": " on Imagenet-1k during 800 epochs. The training is inherited from DeiT-III, which also uses BCE when training with Imagenet-1k only." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 306, + 358, + 457, + 371 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 358, + 457, + 371 + ], + "spans": [ + { + "bbox": [ + 306, + 358, + 457, + 371 + ], + "type": "text", + "content": "5.2. Empirical analysis of cosub" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 378, + 545, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 378, + 545, + 403 + ], + "spans": [ + { + "bbox": [ + 305, + 378, + 545, + 403 + ], + "type": "text", + "content": "We perform various ablations on Imagenet-1k to analyse the impact of our training method on the learned networks." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 407, + 545, + 527 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 407, + 545, + 527 + ], + "spans": [ + { + "bbox": [ + 304, + 407, + 545, + 527 + ], + "type": "text", + "content": "Performance for different model sizes and architectures. Table 1 provides the results obtained by the baseline and cosub when we vary the model size of vision transformers. The stochastic depth coefficient was optimized for the baseline and we keep it unchanged with cosub. As to be expected, our method is almost neutral for small models like ViT-S: " + }, + { + "bbox": [ + 304, + 407, + 545, + 527 + ], + "type": "inline_equation", + "content": "+0.1\\%" + }, + { + "bbox": [ + 304, + 407, + 545, + 527 + ], + "type": "text", + "content": " top-1 accuracy, which is about the standard deviation of measurements. The improvement is increasingly important for larger models, up to a significant improvement of " + }, + { + "bbox": [ + 304, + 407, + 545, + 527 + ], + "type": "inline_equation", + "content": "+0.6\\%" + }, + { + "bbox": [ + 304, + 407, + 545, + 527 + ], + "type": "text", + "content": " top-1 accuracy for ViT-H models." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 529, + 546, + 649 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 529, + 546, + 649 + ], + "spans": [ + { + "bbox": [ + 304, + 529, + 546, + 649 + ], + "type": "text", + "content": "In Table 2, we show that our approach is beneficial with all architectures that we have tried. We report the results of the original paper, evaluate the performance with our baseline training, and measure the improvement brought by cosub. For almost all architectures, we observe a significant boost in performance. The exception is the ResNet-50, for which cosub improves the top-1 accuracy by only " + }, + { + "bbox": [ + 304, + 529, + 546, + 649 + ], + "type": "inline_equation", + "content": "+0.1\\%" + }, + { + "bbox": [ + 304, + 529, + 546, + 649 + ], + "type": "text", + "content": ", similar to our observation with ViT-S. In Table B.1 in the appendix we present improved results obtained for multiple architectures pre-trained with cosub onImagenet-21k." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 654, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 547, + 713 + ], + "type": "text", + "content": "Analysis of submodel performance. With cosub, we sample different subnetworks during training to perform the constraining. We analyse the impact of cosub on the accuracy of the sub networks themselves. In Figure 4 we consider the accuracy of submodels of different size of ViT-B and ViT-L" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "11705" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 51, + 70, + 282, + 148 + ], + "blocks": [ + { + "bbox": [ + 51, + 70, + 282, + 148 + ], + "lines": [ + { + "bbox": [ + 51, + 70, + 282, + 148 + ], + "spans": [ + { + "bbox": [ + 51, + 70, + 282, + 148 + ], + "type": "table", + "html": "
MethodViT-LViT-H
(a)supervised baseline [51]84.584.9
(b)KD85.385.3
(c)mean teacher84.483.4
(d)co-training82.683.1
(e)cosub84.985.5
(b) + (e)KD + cosub85.385.7
", + "image_path": "59e659e7c74bc8c7438b99554efc213012e7414fb8b093ab2df8a5cd590f87a3.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 50, + 231, + 287, + 256 + ], + "blocks": [ + { + "bbox": [ + 47, + 148, + 287, + 216 + ], + "lines": [ + { + "bbox": [ + 47, + 148, + 287, + 216 + ], + "spans": [ + { + "bbox": [ + 47, + 148, + 287, + 216 + ], + "type": "text", + "content": "Table 4. Training strategies with distillation. We compare on Imagenet-1k at resolution " + }, + { + "bbox": [ + 47, + 148, + 287, + 216 + ], + "type": "inline_equation", + "content": "224 \\times 224" + }, + { + "bbox": [ + 47, + 148, + 287, + 216 + ], + "type": "text", + "content": " different approaches involving co- or self-training with distillation. KD, mean teacher and co-training use the same " + }, + { + "bbox": [ + 47, + 148, + 287, + 216 + ], + "type": "inline_equation", + "content": "\\lambda = 0.5" + }, + { + "bbox": [ + 47, + 148, + 287, + 216 + ], + "type": "text", + "content": " and same hyper-parameters as in Deit-III. Unlike cosub, the mean teacher (c) requires other hyper-parameters for EMA. We perform a small grid-search to adjust this parameters. Note (last row): our approach is complementary with KD, assuming a pre-trained teacher is available beforehand." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 50, + 231, + 287, + 256 + ], + "lines": [ + { + "bbox": [ + 50, + 231, + 287, + 256 + ], + "spans": [ + { + "bbox": [ + 50, + 231, + 287, + 256 + ], + "type": "table", + "html": "
λ0.10.30.50.70.91.0
top1 accuracy79.0583.2783.5583.2083.0482.91
", + "image_path": "5df5b7a5ef28a32c10f0dbb652a1ec1cf3969c003ad89652f743207b373ea06a.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 323, + 286, + 347 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 323, + 286, + 347 + ], + "spans": [ + { + "bbox": [ + 47, + 323, + 286, + 347 + ], + "type": "text", + "content": "models. Cosub improves the accuracy of the whole population of sub-networks and, in particular, the target network." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 353, + 286, + 448 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 353, + 286, + 448 + ], + "spans": [ + { + "bbox": [ + 46, + 353, + 286, + 448 + ], + "type": "text", + "content": "Loss formulation. In Table 3 we experiment with different losses for cosub. With Imagenet1k, DeiT-III training uses BCE instead of CE for the main loss. With cosub, BCE is more compatible with the loss of the baseline and, as to be expected, we also observe a better performance with BCE. We have done ablations using hard and soft targets for the cosub loss. The results are similar, therefore by default we keep soft-targets for the cosub loss." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 455, + 286, + 538 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 455, + 286, + 538 + ], + "spans": [ + { + "bbox": [ + 46, + 455, + 286, + 538 + ], + "type": "text", + "content": "Alternative teacher/student. In Table 4 we report the results obtained with the different distillation or co-training approaches depicted in Figure 2. Other approaches are not effective off-the-shelf, except KD that requires a pre-trained teacher. Our approach is on par with KD (lower for ViT-L, better or ViT-H), and in the last row we show that it even provides a slight boost to combine KD with cosub." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 544, + 286, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 544, + 286, + 628 + ], + "spans": [ + { + "bbox": [ + 46, + 544, + 286, + 628 + ], + "type": "text", + "content": "Impact of the cosub loss. The hyper-parameter " + }, + { + "bbox": [ + 46, + 544, + 286, + 628 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 46, + 544, + 286, + 628 + ], + "type": "text", + "content": " controls the tradeoff between the co-distillation loss and the cross-entropy classification loss. Setting " + }, + { + "bbox": [ + 46, + 544, + 286, + 628 + ], + "type": "inline_equation", + "content": "\\lambda = 1" + }, + { + "bbox": [ + 46, + 544, + 286, + 628 + ], + "type": "text", + "content": " means that we have a regular supervised training setting, except that (i) we double the batch size by duplicating the image after data augmentation, and (ii) stochastic depth selects different layers for each image copy." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 629, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 629, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 629, + 286, + 713 + ], + "type": "text", + "content": "In Table 5, we measure the impact of " + }, + { + "bbox": [ + 46, + 629, + 286, + 713 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 46, + 629, + 286, + 713 + ], + "type": "text", + "content": ", with all other hyper-parameters being fixed, for ViT-H trained at resolution " + }, + { + "bbox": [ + 46, + 629, + 286, + 713 + ], + "type": "inline_equation", + "content": "126 \\times 126" + }, + { + "bbox": [ + 46, + 629, + 286, + 713 + ], + "type": "text", + "content": " onImagenet-1k. We observe that the best ratio is to use an equal weighting of the cosub loss and the classic training loss. Using the cosub loss increases the performance by " + }, + { + "bbox": [ + 46, + 629, + 286, + 713 + ], + "type": "inline_equation", + "content": "0.6\\%" + }, + { + "bbox": [ + 46, + 629, + 286, + 713 + ], + "type": "text", + "content": " Top-1 accuracy onImagenet-val, which is the typical improvement that we observe for large models." + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 312, + 70, + 541, + 205 + ], + "blocks": [ + { + "bbox": [ + 47, + 259, + 286, + 298 + ], + "lines": [ + { + "bbox": [ + 47, + 259, + 286, + 298 + ], + "spans": [ + { + "bbox": [ + 47, + 259, + 286, + 298 + ], + "type": "text", + "content": "Table 5. Ablation of the parameter " + }, + { + "bbox": [ + 47, + 259, + 286, + 298 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 47, + 259, + 286, + 298 + ], + "type": "text", + "content": " controlling the weight of the co-distillation loss across submodels (Imagenet1k-val, top1-acc). Model ViT-H trained at resolution " + }, + { + "bbox": [ + 47, + 259, + 286, + 298 + ], + "type": "inline_equation", + "content": "{126} \\times {126}" + }, + { + "bbox": [ + 47, + 259, + 286, + 298 + ], + "type": "text", + "content": " onImagenet-1k during 800 epochs. " + }, + { + "bbox": [ + 47, + 259, + 286, + 298 + ], + "type": "inline_equation", + "content": "\\lambda =" + }, + { + "bbox": [ + 47, + 259, + 286, + 298 + ], + "type": "text", + "content": " 1.0 corresponds a supervised baseline and " + }, + { + "bbox": [ + 47, + 259, + 286, + 298 + ], + "type": "inline_equation", + "content": "\\lambda = {0.5}" + }, + { + "bbox": [ + 47, + 259, + 286, + 298 + ], + "type": "text", + "content": " corresponds to cosub." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 312, + 70, + 541, + 205 + ], + "lines": [ + { + "bbox": [ + 312, + 70, + 541, + 205 + ], + "spans": [ + { + "bbox": [ + 312, + 70, + 541, + 205 + ], + "type": "table", + "html": "
ModelτepochsBaseline+cosub
valv2valv2
CaiT-M120.2040083.272.983.773.5
80082.972.683.673.1
PiT-B0.2540083.873.684.174.1
80082.471.983.172.8
ViT-B0.2040083.172.683.273.1
80083.773.183.973.5
120083.372.8--
160083.373.4--
ViT-H0.6040084.875.385.075.8
80084.975.685.576.3
", + "image_path": "058c06a8f9ff92a83df772fbadcc0db615c759b10e3f72538502427652e5ff61.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 309, + 297, + 543, + 389 + ], + "blocks": [ + { + "bbox": [ + 305, + 208, + 545, + 285 + ], + "lines": [ + { + "bbox": [ + 305, + 208, + 545, + 285 + ], + "spans": [ + { + "bbox": [ + 305, + 208, + 545, + 285 + ], + "type": "text", + "content": "Table 6. We compare ViT models trained with and without cosub on ImageNet-1k only with different number of epochs at resolution " + }, + { + "bbox": [ + 305, + 208, + 545, + 285 + ], + "type": "inline_equation", + "content": "224 \\times 224" + }, + { + "bbox": [ + 305, + 208, + 545, + 285 + ], + "type": "text", + "content": ". One can see that cosub is more effective for larger models yielding higher values of the SD hyper-parameter " + }, + { + "bbox": [ + 305, + 208, + 545, + 285 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 305, + 208, + 545, + 285 + ], + "type": "text", + "content": ". It avoids the early saturation or overfitting of the performance that we typically observe with the baseline when we increase the training time without re-adjusting hyper-parameters. See also Table 5 for a direct comparison with and without the co-distillation loss, and Table 7 for the corresponding training times per epoch." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 309, + 297, + 543, + 389 + ], + "lines": [ + { + "bbox": [ + 309, + 297, + 543, + 389 + ], + "spans": [ + { + "bbox": [ + 309, + 297, + 543, + 389 + ], + "type": "table", + "html": "
Training methodmodelGPUs usedMemory peak (GB)Time (min) by epoch
DeiT-IIIViT-L3221.48
ViT-H6427.612
DeiT-III + ESDViT-L3215.19
ViT-H6415.211
cosub (with ESD)ViT-L3226.916
ViT-H6425.017
", + "image_path": "01139790c47b71d89479342505e12a79e81b3b229187ace569cf5aece4364efb.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 391, + 545, + 458 + ], + "lines": [ + { + "bbox": [ + 305, + 391, + 545, + 458 + ], + "spans": [ + { + "bbox": [ + 305, + 391, + 545, + 458 + ], + "type": "text", + "content": "Table 7. Training times of different models trained at resolution " + }, + { + "bbox": [ + 305, + 391, + 545, + 458 + ], + "type": "inline_equation", + "content": "224\\times 224" + }, + { + "bbox": [ + 305, + 391, + 545, + 458 + ], + "type": "text", + "content": " with batch size 2048 on Imagenet-1k with DeiT-III and our approach. cosub uses our efficient stochastic depth (ESD), which amortizes the extra memory needed by cosub, especially for the largest models with high stochastic depth values (0.45 for ViT-L, and 0.6 for ViT-H). Timings are indicative and not representative of an optimized selection of the batch size, see Appendix C for measurements with adjusted parameters." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 305, + 475, + 545, + 571 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 475, + 545, + 571 + ], + "spans": [ + { + "bbox": [ + 305, + 475, + 545, + 571 + ], + "type": "text", + "content": "Number of training epochs. In Table 6 we compare results on Imagenet-1k-val and Imagenet-v2 different for architectures trained with and without cosub on Imagenet-1k only at resolution " + }, + { + "bbox": [ + 305, + 475, + 545, + 571 + ], + "type": "inline_equation", + "content": "224 \\times 224" + }, + { + "bbox": [ + 305, + 475, + 545, + 571 + ], + "type": "text", + "content": " with different number of epochs. We observe less overfitting with cosub and longer training schedule. In particular, with bigger architecture like ViT-H, we observe continuous improvement with a longer schedule where the baseline saturates." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 577, + 545, + 636 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 577, + 545, + 636 + ], + "spans": [ + { + "bbox": [ + 305, + 577, + 545, + 636 + ], + "type": "text", + "content": "Training time. In Table 7 we compare the training costs of cosub and DeiT-III. Thanks to our efficient stochastic depth formulation we maintain a similar memory peak during training. For bigger architectures the gap in training speed between cosub and the baseline is decreasing." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 641, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 641, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 305, + 641, + 545, + 712 + ], + "type": "text", + "content": "Resolution. In Table 8 we compare different ViT architectures trained with and without cosub at different resolutions onImagenet-1k. We fine-tune during 20 epochs at resolution " + }, + { + "bbox": [ + 305, + 641, + 545, + 712 + ], + "type": "inline_equation", + "content": "224 \\times 224" + }, + { + "bbox": [ + 305, + 641, + 545, + 712 + ], + "type": "text", + "content": " before evaluation at this resolution. We observe that cosub gives significant improvements across the different resolutions and models." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "11706" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 52, + 70, + 282, + 194 + ], + "blocks": [ + { + "bbox": [ + 52, + 70, + 282, + 194 + ], + "lines": [ + { + "bbox": [ + 52, + 70, + 282, + 194 + ], + "spans": [ + { + "bbox": [ + 52, + 70, + 282, + 194 + ], + "type": "table", + "html": "
ModelResolutionDeit-III Baseline+cosub
valv2valv2
ViT-B128 × 12883.573.483.874.0
192 × 19283.873.684.174.0
224 × 22483.773.183.973.5
ViT-L128 × 12884.574.785.175.5
192 × 19284.975.185.275.7
224 × 22484.575.084.975.6
ViT-H126 × 12685.175.685.676.4
182 × 18285.175.985.776.6
224 × 22484.975.685.576.3
", + "image_path": "c99963d3bf244e97afce1eedc22e7bb744f63378ca0f4cf4b61c276a1164e907.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 50, + 233, + 287, + 324 + ], + "blocks": [ + { + "bbox": [ + 47, + 198, + 287, + 228 + ], + "lines": [ + { + "bbox": [ + 47, + 198, + 287, + 228 + ], + "spans": [ + { + "bbox": [ + 47, + 198, + 287, + 228 + ], + "type": "text", + "content": "Table 8. Imagenet-1k val and v2 top-1 accuracy of ViT models trained with and without cosub for 800 epochs on Imagenet-1k at different resolutions, followed by finetuning for 20 epochs at resolution " + }, + { + "bbox": [ + 47, + 198, + 287, + 228 + ], + "type": "inline_equation", + "content": "224 \\times 224" + }, + { + "bbox": [ + 47, + 198, + 287, + 228 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 50, + 233, + 287, + 324 + ], + "lines": [ + { + "bbox": [ + 50, + 233, + 287, + 324 + ], + "spans": [ + { + "bbox": [ + 50, + 233, + 287, + 324 + ], + "type": "table", + "html": "
resol. →112224336448112224336448
↓ modelImagenet-valImagenet-v2
ViT-S78.083.184.685.266.673.775.176.3
ViT-M80.685.086.086.369.676.076.877.2
ViT-B82.886.386.987.472.177.077.978.3
ViT-L85.487.588.188.375.779.179.880.0
ViT-H86.288.0--76.979.6--
ViT-g86.5---77.3---
", + "image_path": "7617323a07ba4969a56535b7dde3e90c81beb0748e68d020de1363082af727b0.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 51, + 406, + 282, + 486 + ], + "blocks": [ + { + "bbox": [ + 46, + 327, + 287, + 396 + ], + "lines": [ + { + "bbox": [ + 46, + 327, + 287, + 396 + ], + "spans": [ + { + "bbox": [ + 46, + 327, + 287, + 396 + ], + "type": "text", + "content": "Table 9. Performance of models at different resolutions. We report the results obtained on Imagenet-val by models of different sizes pre-trained with cosub on Imagenet-21k and fine-tuned on Imagenet-1k. Training schedule: 270 epochs except ViT-g (90 epochs). Except for ViT-S, the results at resolution 336 and 448 were pre-trained on Imagenet-21k at resolution 224 for efficiency reasons. We have not fine-tuned ViT-H and ViT-g at large resolutions since these models are computationally expensive." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 51, + 406, + 282, + 486 + ], + "lines": [ + { + "bbox": [ + 51, + 406, + 282, + 486 + ], + "spans": [ + { + "bbox": [ + 51, + 406, + 282, + 486 + ], + "type": "table", + "html": "
MethodLong TrainingLayer DecayViT-SModel ViT-BViT-L
baselineXX82.685.286.8
XX82.585.887.4
cosubX✓82.786.087.5
✓X82.886.087.4
✓✓83.186.387.5
", + "image_path": "63ad22a6f9a7ee99ef5b228e82c68526660b1606c0f15405a60186dcd628864e.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 540, + 287, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 540, + 287, + 601 + ], + "spans": [ + { + "bbox": [ + 46, + 540, + 287, + 601 + ], + "type": "text", + "content": "Imagenet-21k impact of layer-decay. In Table 10 we compare different number of epochs for the pre-training on Imagenet-21k and the finetuning on Imagenet-1k with and without layer-decay. We observe that both layer-decay and long training bring improvements with cosub." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 607, + 244, + 620 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 607, + 244, + 620 + ], + "spans": [ + { + "bbox": [ + 47, + 607, + 244, + 620 + ], + "type": "text", + "content": "5.3. Comparisons with the state of the art" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 629, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 629, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 629, + 287, + 715 + ], + "type": "text", + "content": "Imagenet-1k. In Table 11 we compare architectures trained with cosub with state-of-the-art results from the literature for this architecture. We observe that architectures trained with cosub are very competitive. For instance, for ResNet-152, RegNetY-16GF, PiT-B we improve the best results reported in the literature by more than " + }, + { + "bbox": [ + 46, + 629, + 287, + 715 + ], + "type": "inline_equation", + "content": "0.9\\%" + }, + { + "bbox": [ + 46, + 629, + 287, + 715 + ], + "type": "text", + "content": " top-1 accuracy on Imagenet-1k-val." + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 309, + 70, + 547, + 342 + ], + "blocks": [ + { + "bbox": [ + 47, + 489, + 287, + 519 + ], + "lines": [ + { + "bbox": [ + 47, + 489, + 287, + 519 + ], + "spans": [ + { + "bbox": [ + 47, + 489, + 287, + 519 + ], + "type": "text", + "content": "Table 10. We measure the impact of layer-decay during the finetuning on Imagenet-1k for models pre-trained on Imagenet-21k with cosub during 90 epochs (default) and 270 epochs (long training)." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 309, + 70, + 547, + 342 + ], + "lines": [ + { + "bbox": [ + 309, + 70, + 547, + 342 + ], + "spans": [ + { + "bbox": [ + 309, + 70, + 547, + 342 + ], + "type": "table", + "html": "
Modelnb params (×106)throughput (im/s)FLOPs (×109)Peak Mem (MB)Top-1 Acc.v2 Acc.
“Traditional” ConvNets
ResNet-50 [24, 55]25.625874.1218280.468.7
ResNet-101 [24, 55]44.515867.9226981.570.3
ResNet-101 - cosub44.515867.9226982.170.8
ResNet-152 [24, 55]60.2112211.6235982.070.6
ResNet-152 - cosub60.2112211.6235983.172.1
RegNetY-8GF [42, 55]39.211588.0393982.271.1
RegNetY-16GF [42, 50]83.671416.0520482.972.4
RegNetY-16GF - cosub83.671416.0520483.872.8
Vision Transformers derivatives
Swin-T [34]28.311094.5334581.369.5
Swin-B [34]87.853215.4469583.5-
PiT-S [26]23.518092.9329380.9-
PiT-S - cosub23.518092.9329381.369.7
PiT-B [26]73.861512.5756482.0-
PiT-B - cosub73.861512.5756484.174.1
Vanilla Vision Transformers
ViT-S [50]22.018914.698779.868.1
ViT-S - Deit III [51]22.018914.698781.470.5
ViT-S - cosub22.018914.698781.570.8
ViT-B [14]86.683117.5207877.9-
ViT-B - DeiT [50]86.683117.5207881.871.5
ViT-B - DeiT/distilled86.683117.5207883.473.2
ViT-B - DeiT III [51]86.683117.5207883.873.6
ViT-B - cosub86.683117.5207884.274.2
ViT-L - Deit III [51]304.427761.6378984.975.1
ViT-L - cosub304.427761.6378985.375.5
ViT-H - Deit III [51]632.1112167.4698485.275.9
ViT-H - cosub632.1112167.4698485.776.6
", + "image_path": "92643485a22938f5636cebd22bd797d810946cab5eebddb96c0d33038b8ad2f8.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 346, + 547, + 414 + ], + "lines": [ + { + "bbox": [ + 304, + 346, + 547, + 414 + ], + "spans": [ + { + "bbox": [ + 304, + 346, + 547, + 414 + ], + "type": "text", + "content": "Table 11. Classification with Imagenet1k training. We compare with models trained on Imagenet-1k only at resolution " + }, + { + "bbox": [ + 304, + 346, + 547, + 414 + ], + "type": "inline_equation", + "content": "224 \\times 224" + }, + { + "bbox": [ + 304, + 346, + 547, + 414 + ], + "type": "text", + "content": " without self-supervised pre-training (see supp. material for a comparison). We report Top-1 accuracy on Imagenet-1k-val and Imagenet-v2 with different measures of complexity: throughput, FLOPs, number of parameters and peak memory usage. The throughput and peak memory are measured on a single V100-32GB GPU with batch size fixed to 256 and mixed precision." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 435, + 545, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 435, + 545, + 496 + ], + "spans": [ + { + "bbox": [ + 304, + 435, + 545, + 496 + ], + "type": "text", + "content": "Imagenet-21k. In Table 12 we compare ViT models pretrained with cosub on Imagenet-21k and finetuned with cosub on Imagenet-1k with other architectures and our baseline DeiT-III. Our results with vanilla ViT outperform the baseline and are competitive with recent architectures." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 500, + 546, + 608 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 500, + 546, + 608 + ], + "spans": [ + { + "bbox": [ + 304, + 500, + 546, + 608 + ], + "type": "text", + "content": "Overfitting analysis. As recommended in prior works [51, 55] we perform an overfitting analysis. We evaluate our models trained with codist onImagenet-1k val andImagenet-v2 [43]. The results are reported in Figure 5. For ViT, we observe that cosub does not overfit more than the DeiT-III baseline [51]. Our results with other architectures in Table 12 concur with that observation: our results are comparatively stronger onImagenet-v2 than those reported in the literature for the exact same models." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 618, + 415, + 630 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 618, + 415, + 630 + ], + "spans": [ + { + "bbox": [ + 306, + 618, + 415, + 630 + ], + "type": "text", + "content": "5.4. Downstream tasks" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 641, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 641, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 641, + 547, + 715 + ], + "type": "text", + "content": "Semantic segmentation. First we evaluate our ViT models pre-trained onImagenet with cosub for semantic segmentation on the ADE20k dataset [63]. ADE20k consists of 20k training and 5k validation images with labels over 150 categories. We adopt the training schedule from Swin: 160k iterations with UperNet [57]. At test time we evaluate with" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "11707" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 51, + 70, + 282, + 345 + ], + "blocks": [ + { + "bbox": [ + 51, + 70, + 282, + 345 + ], + "lines": [ + { + "bbox": [ + 51, + 70, + 282, + 345 + ], + "spans": [ + { + "bbox": [ + 51, + 70, + 282, + 345 + ], + "type": "table", + "html": "
Architecturenb params (×106)throughput (im/s)FLOPs (×109)peak mem (MB)top1 valacc. v2
Convnets
EfficientNetV2-M [48]54.131225.0712786.275.9
EfficientNetV2-L [48]118.517953.0954086.876.9
ResNet-152 [24, 55]60.2112211.6235982.070.6
ResNet-152 - cosub60.2112211.6235983.173.1
RegnetY16GF - cosub83.671416.0520484.274.7
ConvNeXt-S- cosub50.27838.7221885.276.0
ConvNeXt-B [35]88.656315.4302985.875.6
ConvNeXt-B - cosub88.656315.4302985.876.9
ConvNeXt-L [35]197.834434.4486586.676.6
ConvNeXt-XL [35]350.224160.9695187.077.0
Transformers variations
Swin-B [34]87.853215.4469585.274.6
Swin-B - cosub87.853215.4469586.277.2
Swin-L [34]196.533734.5735086.376.3
Swin-L - cosub196.533734.5735087.178.1
PiT-B - cosub [26]73.861512.5756485.876.8
XCiT-S12 - cosub [17]26.213734.9133084.274.9
XCiT-M24 - cosub [17]84.455316.2201086.578.0
XCiT-L24 - cosub [17]189.033436.1331587.277.8
Vanilla Vision Transformers [14, 51]
ViT-S - cosub22.018914.698783.173.7
ViT-M - cosub39.013078.0132285.076.0
ViT-B - DeiT-III86.683117.6207885.776.5
ViT-B - cosub86.683117.6207886.377.0
ViT-L - DeiT-III304.427761.6378987.078.6
ViT-L - cosub304.427761.6378987.579.1
ViT-H - DeiT-III632.1112167.4698487.279.2
ViT-H - cosub632.1112167.4698488.080.0
", + "image_path": "96d5499c49a9c7430027c1cc2f4ccb811fe22e565246ff59894e3860fd9dc429.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 47, + 348, + 287, + 453 + ], + "lines": [ + { + "bbox": [ + 47, + 348, + 287, + 453 + ], + "spans": [ + { + "bbox": [ + 47, + 348, + 287, + 453 + ], + "type": "text", + "content": "Table 12. Classification with ImageNet-21k pretraining. We report top-1 accuracy on the validation set of Imagenet1k and Imagenet-V2 with different measures of complexity. The peak memory usage is measured on a single V100-32GB GPU with batch size fixed to 256 and mixed precision. For Swin-L the memory peak is an estimation since we decreased the batch size to 128 to fit in memory. All models are evaluated at resolution 224 except EfficientNetV2 that use resolution 480. ViT are pretrained with a " + }, + { + "bbox": [ + 47, + 348, + 287, + 453 + ], + "type": "inline_equation", + "content": "\\times 3" + }, + { + "bbox": [ + 47, + 348, + 287, + 453 + ], + "type": "text", + "content": " schedule, comparable to that used in the best DeiT-III baseline (270 vs. 240 epochs). All other cosub models are pretrained during 90 epochs on Imagenet21k, with 50 epochs of fine-tuning. The " + }, + { + "bbox": [ + 47, + 348, + 287, + 453 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 47, + 348, + 287, + 453 + ], + "type": "text", + "content": " hyper-parameter is set per model based on prior choices or best guess based on model size." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 47, + 466, + 162, + 585 + ], + "blocks": [ + { + "bbox": [ + 47, + 466, + 162, + 585 + ], + "lines": [ + { + "bbox": [ + 47, + 466, + 162, + 585 + ], + "spans": [ + { + "bbox": [ + 47, + 466, + 162, + 585 + ], + "type": "image", + "image_path": "f8c82359cfd5310e3d42ed277bb8b03a5174a9a81497154610bcc05e2cced25e.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 588, + 287, + 636 + ], + "lines": [ + { + "bbox": [ + 46, + 588, + 287, + 636 + ], + "spans": [ + { + "bbox": [ + 46, + 588, + 287, + 636 + ], + "type": "text", + "content": "Figure 5. Overfitting measurement: top-1 accuracy on Imagenet-val vs. Imagenet-v2 for models in Tables 11 and 12 pre-trained on Imagenet-1k and Imagenet-21k, respectively. Our cosub ViTs (plain lines and points) do not overfit more than DeiT-III [51] overall. Our Imagenet-21k results for Swin and Convnext generalize much better on v2 than the original models." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 169, + 466, + 276, + 585 + ], + "blocks": [ + { + "bbox": [ + 169, + 466, + 276, + 585 + ], + "lines": [ + { + "bbox": [ + 169, + 466, + 276, + 585 + ], + "spans": [ + { + "bbox": [ + 169, + 466, + 276, + 585 + ], + "type": "image", + "image_path": "8d84030110f5e627046f800c771d9c5379935fe7bc51afca709a02acd1ff4a8c.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "type": "text", + "content": "a single scale and multi-scale. Our UperNet implementation is the same as in DeiT III [51]. Our results are reported in Table 13. We observe that vanilla ViTs trained with cos-sub outperform our baseline DeiT-III but also have better FLOPs-accuracy trade-offs than recent architectures." + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 310, + 70, + 543, + 278 + ], + "blocks": [ + { + "bbox": [ + 310, + 70, + 543, + 278 + ], + "lines": [ + { + "bbox": [ + 310, + 70, + 543, + 278 + ], + "spans": [ + { + "bbox": [ + 310, + 70, + 543, + 278 + ], + "type": "table", + "html": "
Backbone#params (×106)FLOPs (×109)Single-scale mIoUMulti-scale mIoU
DeiT-S52.01099-44.0
Swin-T59.994544.546.1
ViT-S – DeiT-III41.758845.646.8
ViT-S – cosub41.758847.048.0
PatchConvNet-B60140.6125848.148.6
PatchConvNet-B120229.8155049.450.3
MAE ViT-B127.7128348.1-
Swin-B121.0118848.149.7
ViT-B – DeiT-III127.7128349.350.2
ViT-B – cosub127.7128349.349.9
ViT-L – DeiT-III353.6223151.552.0
ViT-L – cosub353.6223152.553.1
PatchConvNet-B60†140.6125850.551.1
Swin-B† (640 × 640)121.0184150.051.6
ViT-B – DeiT-III†127.7128353.454.1
ViT-B – cosub†127.7128353.754.7
PatchConvNet-L120†383.7208652.252.9
Swin-L† (640 × 640)234.03230-53.5
ViT-L – DeiT-III†353.6223154.655.6
ViT-L – cosub†353.6223155.756.3
", + "image_path": "f26f64938d3d25c0ecd5f347f517fdbae0e44618d7f62d08e13c86c1957f0d59.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 306, + 284, + 545, + 332 + ], + "lines": [ + { + "bbox": [ + 306, + 284, + 545, + 332 + ], + "spans": [ + { + "bbox": [ + 306, + 284, + 545, + 332 + ], + "type": "text", + "content": "Table 13. ADE20k semantic segmentation performance using UperNet [57], in comparable setting as prior works [13, 17, 34]. All models are pretrained on Imagenet-1k, except bottom models identified with " + }, + { + "bbox": [ + 306, + 284, + 545, + 332 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 306, + 284, + 545, + 332 + ], + "type": "text", + "content": ", which are pre-trained on Imagenet-21k. By default the finetuning resolution on ADE20k is " + }, + { + "bbox": [ + 306, + 284, + 545, + 332 + ], + "type": "inline_equation", + "content": "512 \\times 512" + }, + { + "bbox": [ + 306, + 284, + 545, + 332 + ], + "type": "text", + "content": " except when mentioned otherwise (for Swin)." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 310, + 343, + 541, + 415 + ], + "blocks": [ + { + "bbox": [ + 310, + 343, + 541, + 415 + ], + "lines": [ + { + "bbox": [ + 310, + 343, + 541, + 415 + ], + "spans": [ + { + "bbox": [ + 310, + 343, + 541, + 415 + ], + "type": "table", + "html": "
ModelCifar-10Cifar-100FlowersCarsiNat-18iNat-19
ViT-S - DeiT-III98.990.696.489.967.172.7
ViT-B - DeiT-III99.392.598.693.473.678.0
ViT-L - DeiT-III99.393.498.994.575.679.3
ViT-S - cosub99.191.797.493.070.175.6
ViT-B - cosub99.192.698.493.574.178.1
ViT-L - cosub99.493.598.894.576.280.1
", + "image_path": "dfc70bd0a6d8ace522a46704770d7ce48a4af37ac0b5cea2084ea240587dbb2f.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 305, + 416, + 545, + 464 + ], + "lines": [ + { + "bbox": [ + 305, + 416, + 545, + 464 + ], + "spans": [ + { + "bbox": [ + 305, + 416, + 545, + 464 + ], + "type": "text", + "content": "Table 14. ViT models pre-trained with cosub or DeiT-III on Imagenet-1k and finetuned on six different target datasets. We note that for small datasets (CIFAR, Flowers, and Cars) our approach is useful for small models but neutral for larger models. The gains are more significant when transferring to the larger iNaturalist-18 and iNaturalist-19 datasets." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 479, + 545, + 575 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 479, + 545, + 575 + ], + "spans": [ + { + "bbox": [ + 305, + 479, + 545, + 575 + ], + "type": "text", + "content": "Transfer learning. We now measure how the performance improvements observed with cosub translate to other classification problems. For this purpose, we performed transfer learning on the six different datasets used in DeiT-III. Our results are reported Table 14. Our pre-trained and fine-tuned models with cosub generally improve the baseline. The gains are overall more significant on the more challenging datasets like iNaturalist 2018 and iNaturalist 2019." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 586, + 378, + 597 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 586, + 378, + 597 + ], + "spans": [ + { + "bbox": [ + 306, + 586, + 378, + 597 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": "Co-training submodels (cosub) is an effective way to improve existing deep residual networks. It is straightforward to implement, just involving a few lines of code. It does not need a pre-trained teacher, and it only maintains a single set of weights for the model. Extensive experimental results on image classification, transfer learning and semantic segmentation show that cosub is overall extremely effective. It works off-the-shelf and improves the state of the art for various network architectures, including convnets like Regnet." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "11708" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 53, + 91, + 286, + 123 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 286, + 123 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 286, + 123 + ], + "type": "text", + "content": "[1] Samira Abnar, Mostafa Dehghani, and Willem Zuidema. Transferring inductive biases through knowledge distillation. arXiv preprint arXiv:2006.00555, 2020. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 124, + 287, + 167 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 124, + 287, + 167 + ], + "spans": [ + { + "bbox": [ + 53, + 124, + 287, + 167 + ], + "type": "text", + "content": "[2] Rohan Anil, Gabriel Pereyra, Alexandre Passos, Robert Ormandi, George E Dahl, and Geoffrey E Hinton. Large scale distributed neural network training through online distillation. arXiv preprint arXiv:1804.03235, 2018. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 168, + 287, + 224 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 168, + 287, + 224 + ], + "spans": [ + { + "bbox": [ + 53, + 168, + 287, + 224 + ], + "type": "text", + "content": "[3] Mahmoud Assran, Mathilde Caron, Ishan Misra, Piotr Bojanowski, Florian Bordes, Pascal Vincent, Armand Joulin, Mike Rabbat, and Nicolas Ballas. Masked siamese networks for label-efficient learning. In European Conference on Computer Vision, pages 456-473. Springer, 2022. I" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 224, + 287, + 245 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 224, + 287, + 245 + ], + "spans": [ + { + "bbox": [ + 53, + 224, + 287, + 245 + ], + "type": "text", + "content": "[4] L. Ba and R. Caruana. Do deep nets really need to be deep? In NeurIPS, 2014. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 247, + 287, + 278 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 247, + 287, + 278 + ], + "spans": [ + { + "bbox": [ + 53, + 247, + 287, + 278 + ], + "type": "text", + "content": "[5] Hangbo Bao, Li Dong, and Furu Wei. Beit: Bert pre-training of image transformers. arXiv preprint arXiv:2106.08254, 2021. 1, 5, I, II" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 280, + 287, + 324 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 280, + 287, + 324 + ], + "spans": [ + { + "bbox": [ + 53, + 280, + 287, + 324 + ], + "type": "text", + "content": "[6] Irwan Bello, W. Fedus, Xianzhi Du, E. D. Cubuk, A. Srinivas, Tsung-Yi Lin, Jonathon Shlens, and Barret Zoph. Revisiting ResNets: Improved training and scaling strategies. arXiv preprint arXiv:2103.07579, 2021. II" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 325, + 287, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 325, + 287, + 357 + ], + "spans": [ + { + "bbox": [ + 53, + 325, + 287, + 357 + ], + "type": "text", + "content": "[7] H. Bourlard and Y. Kamp. Auto-association by multilayer perceptrons and singular value decomposition. Biological Cybernetics, 59:291-294, 1988. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 358, + 286, + 379 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 358, + 286, + 379 + ], + "spans": [ + { + "bbox": [ + 53, + 358, + 286, + 379 + ], + "type": "text", + "content": "[8] Leo Breiman. Bagging predictors. Machine learning, 24(2):123-140, 1996. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 380, + 287, + 424 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 380, + 287, + 424 + ], + "spans": [ + { + "bbox": [ + 53, + 380, + 287, + 424 + ], + "type": "text", + "content": "[9] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. arXiv preprint arXiv:2104.14294, 2021. 1, 2, II" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 426, + 287, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 426, + 287, + 468 + ], + "spans": [ + { + "bbox": [ + 48, + 426, + 287, + 468 + ], + "type": "text", + "content": "[10] Kevin Clark, Minh-Thang Luong, Quoc V Le, and Christopher D Manning. Electra: Pre-training text encoders as discriminators rather than generators. arXiv preprint arXiv:2003.10555, 2020. I" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 470, + 287, + 513 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 470, + 287, + 513 + ], + "spans": [ + { + "bbox": [ + 48, + 470, + 287, + 513 + ], + "type": "text", + "content": "[11] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In Conference on Computer Vision and Pattern Recognition, pages 248–255, 2009. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 514, + 287, + 546 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 514, + 287, + 546 + ], + "spans": [ + { + "bbox": [ + 48, + 514, + 287, + 546 + ], + "type": "text", + "content": "[12] Carl Doersch, Ankush Gupta, and Andrew Zisserman. Crosstransformers: spatially-aware few-shot transfer. arXiv preprint arXiv:2007.11498, 2020. 1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 548, + 287, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 548, + 287, + 601 + ], + "spans": [ + { + "bbox": [ + 48, + 548, + 287, + 601 + ], + "type": "text", + "content": "[13] Xiaoyi Dong, Jianmin Bao, Dongdong Chen, Weiming Zhang, Nenghai Yu, Lu Yuan, Dong Chen, and Baining Guo. Cswin transformer: A general vision transformer backbone with cross-shaped windows. arXiv preprint arXiv:2107.00652, 2021. 8" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 603, + 287, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 603, + 287, + 668 + ], + "spans": [ + { + "bbox": [ + 48, + 603, + 287, + 668 + ], + "type": "text", + "content": "[14] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations, 2021. 5, 7, 8" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 670, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 670, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 670, + 287, + 713 + ], + "type": "text", + "content": "[15] Alaaeldin El-Nouby, Gautier Izacard, Hugo Touvron, Ivan Laptev, Hervé Jegou, and Edouard Grave. Are large-scale datasets necessary for self-supervised pre-training? arXiv preprint arXiv:2112.10740, 2021. 1" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 33, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "text", + "content": "[16] Alaaeldin El-Nouby, Natalia Neverova, Ivan Laptev, and Hervé Jégou. Training vision transformers for image retrieval. arXiv preprint arXiv:2102.05644, 2021. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 107, + 545, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 107, + 545, + 161 + ], + "spans": [ + { + "bbox": [ + 307, + 107, + 545, + 161 + ], + "type": "text", + "content": "[17] Alaaeldin El-Nouby, Hugo Touvron, Mathilde Caron, Piotr Bojanowski, Matthijs Douze, Armand Joulin, Ivan Laptev, Natalia Neverova, Gabriel Synnaeve, Jakob Verbeek, et al. Xcit: Cross-covariance image transformers. arXiv preprint arXiv:2106.09681, 2021. 5, 8, II" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 163, + 545, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 163, + 545, + 195 + ], + "spans": [ + { + "bbox": [ + 307, + 163, + 545, + 195 + ], + "type": "text", + "content": "[18] Thomas Elsken, Jan Hendrik Metzen, and Frank Hutter. Neural architecture search: A survey. Journal of Machine Learning Research, 20(1):1997-2017, 2019. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 197, + 545, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 197, + 545, + 229 + ], + "spans": [ + { + "bbox": [ + 307, + 197, + 545, + 229 + ], + "type": "text", + "content": "[19] Dumitru Erhan, Aaron Courville, Yoshua Bengio, and Pascal Vincent. Why does unsupervised pre-training help deep learning? In AISTATS, 2010. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 231, + 545, + 263 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 231, + 545, + 263 + ], + "spans": [ + { + "bbox": [ + 307, + 231, + 545, + 263 + ], + "type": "text", + "content": "[20] Angela Fan, Edouard Grave, and Armand Joulin. Reducing transformer depth on demand with structured dropout. arXiv preprint arXiv:1909.11556, 2019. ICLR 2020. 2, 3, 4, IV" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 265, + 545, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 265, + 545, + 308 + ], + "spans": [ + { + "bbox": [ + 307, + 265, + 545, + 308 + ], + "type": "text", + "content": "[21] Rohit Girdhar, Alaaeldin El-Nouby, Mannat Singh, Kalyan Vasudev Alwala, Armand Joulin, and Ishan Misra. Omnimea: Single model masked pretraining on images and videos. arXiv preprint arXiv:2206.08356, 2022. I" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 309, + 545, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 309, + 545, + 352 + ], + "spans": [ + { + "bbox": [ + 307, + 309, + 545, + 352 + ], + "type": "text", + "content": "[22] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Doll'ar, and Ross B. Girshick. Masked autoencoders are scalable vision learners. arXiv preprint arXiv:2111.06377, 2021. 1, 5, I, II" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 354, + 545, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 354, + 545, + 397 + ], + "spans": [ + { + "bbox": [ + 307, + 354, + 545, + 397 + ], + "type": "text", + "content": "[23] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. arXiv preprint arXiv:1911.05722, 2019. 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 399, + 545, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 399, + 545, + 442 + ], + "spans": [ + { + "bbox": [ + 307, + 399, + 545, + 442 + ], + "type": "text", + "content": "[24] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Conference on Computer Vision and Pattern Recognition, 2016. 1, 5, 7, 8, II" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 444, + 545, + 476 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 444, + 545, + 476 + ], + "spans": [ + { + "bbox": [ + 307, + 444, + 545, + 476 + ], + "type": "text", + "content": "[25] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Identity mappings in deep residual networks. arXiv preprint arXiv:1603.05027, 2016. 1, 5" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 478, + 545, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 478, + 545, + 521 + ], + "spans": [ + { + "bbox": [ + 307, + 478, + 545, + 521 + ], + "type": "text", + "content": "[26] Byeongho Heo, Sangdoo Yun, Dongyoon Han, Sanghyuk Chun, Junsuk Choe, and Seong Joon Oh. Rethinking spatial dimensions of vision transformers. arXiv preprint arXiv:2103.16302, 2021. 5, 7, 8, II" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 522, + 545, + 554 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 522, + 545, + 554 + ], + "spans": [ + { + "bbox": [ + 307, + 522, + 545, + 554 + ], + "type": "text", + "content": "[27] G. Hinton and R. Zemel. Autoencoders, minimum description length and Helmholtz free energy. In NeurIPS, 1993. 1" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 556, + 545, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 556, + 545, + 589 + ], + "spans": [ + { + "bbox": [ + 307, + 556, + 545, + 589 + ], + "type": "text", + "content": "[28] Geoffrey E. Hinton, Oriol Vinyals, and J. Dean. Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531, 2015. 2, 3" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 307, + 590, + 545, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 590, + 545, + 634 + ], + "spans": [ + { + "bbox": [ + 307, + 590, + 545, + 634 + ], + "type": "text", + "content": "[29] Grant Van Horn, Oisin Mac Aodha, Yang Song, Alexander Shepard, Hartwig Adam, Pietro Perona, and Serge J. Belongie. The iNaturalist species classification and detection dataset. arXiv preprint arXiv:1707.06642, 2017. II" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 307, + 635, + 545, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 635, + 545, + 678 + ], + "spans": [ + { + "bbox": [ + 307, + 635, + 545, + 678 + ], + "type": "text", + "content": "[30] Grant Van Horn, Oisin Mac Aodha, Yang Song, Alexander Shepard, Hartwig Adam, Pietro Perona, and Serge J. Belongie. The inaturalist challenge 2018 dataset. arXiv preprint arXiv:1707.06642, 2018. II" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 307, + 680, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 680, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 680, + 545, + 713 + ], + "type": "text", + "content": "[31] Gao Huang, Yu Sun, Zhuang Liu, Daniel Sedra, and Kilian Q. Weinberger. Deep networks with stochastic depth. In European Conference on Computer Vision, 2016. 2, 4, I" + } + ] + } + ], + "index": 32 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "11709" + } + ] + } + ], + "index": 34 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "text", + "content": "[32] Jonathan Krause, Michael Stark, Jia Deng, and Li Fei-Fei. 3d object representations for fine-grained categorization. In IEEE Workshop on 3D Representation and Recognition, 2013. II" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 118, + 287, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 118, + 287, + 138 + ], + "spans": [ + { + "bbox": [ + 49, + 118, + 287, + 138 + ], + "type": "text", + "content": "[33] Alex Krizhevsky. Learning multiple layers of features from tiny images. Technical report, CIFAR, 2009. II" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 140, + 287, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 140, + 287, + 183 + ], + "spans": [ + { + "bbox": [ + 49, + 140, + 287, + 183 + ], + "type": "text", + "content": "[34] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. arXiv preprint arXiv:2103.14030, 2021. 5, 7, 8, II" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 184, + 287, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 184, + 287, + 216 + ], + "spans": [ + { + "bbox": [ + 49, + 184, + 287, + 216 + ], + "type": "text", + "content": "[35] Zhuang Liu, Hanzi Mao, Chao-Yuan Wu, Christoph Feichtenhofer, Trevor Darrell, and Saining Xie. A convnet for the 2020s. arXiv preprint arXiv:2201.03545, 2022. 8, II" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 49, + 217, + 287, + 238 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 217, + 287, + 238 + ], + "spans": [ + { + "bbox": [ + 49, + 217, + 287, + 238 + ], + "type": "text", + "content": "[36] I. Loshchilov and F. Hutter. Fixing weight decay regularization in adam. arXiv preprint arXiv:1711.05101, 2017. I" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 49, + 239, + 287, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 239, + 287, + 270 + ], + "spans": [ + { + "bbox": [ + 49, + 239, + 287, + 270 + ], + "type": "text", + "content": "[37] G. J. McLachlan and D. Peel. Finite Mixture Models. Wiley Series in Probability and Statistics. John Wiley & Sons, 2000. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 49, + 271, + 287, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 271, + 287, + 316 + ], + "spans": [ + { + "bbox": [ + 49, + 271, + 287, + 316 + ], + "type": "text", + "content": "[38] M-E. Nilsback and A. Zisserman. Automated flower classification over a large number of classes. In Proceedings of the Indian Conference on Computer Vision, Graphics and Image Processing, 2008. II" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 49, + 317, + 287, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 317, + 287, + 360 + ], + "spans": [ + { + "bbox": [ + 49, + 317, + 287, + 360 + ], + "type": "text", + "content": "[39] Maxime Oquab, Leon Bottou, Ivan Laptev, and Josef Sivic. Learning and transferring mid-level image representations using convolutional neural networks. In Conference on Computer Vision and Pattern Recognition, 2014. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 49, + 361, + 287, + 425 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 361, + 287, + 425 + ], + "spans": [ + { + "bbox": [ + 49, + 361, + 287, + 425 + ], + "type": "text", + "content": "[40] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. In Advances in neural information processing systems, pages 8026-8037, 2019. 4" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 49, + 426, + 287, + 469 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 426, + 287, + 469 + ], + "spans": [ + { + "bbox": [ + 49, + 426, + 287, + 469 + ], + "type": "text", + "content": "[41] Ilija Radosavovic, Raj Prateek Kosaraju, Ross Girshick, Kaiming He, and Piotr Dólar. Designing network design spaces. In Conference on Computer Vision and Pattern Recognition, 2020. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 49, + 471, + 287, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 471, + 287, + 514 + ], + "spans": [ + { + "bbox": [ + 49, + 471, + 287, + 514 + ], + "type": "text", + "content": "[42] Ilija Radosavovic, Raj Prateek Kosaraju, Ross B. Girshick, Kaiming He, and Piotr Dollar. Designing network design spaces. Conference on Computer Vision and Pattern Recognition, 2020. 5, 7, II" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 49, + 515, + 287, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 515, + 287, + 547 + ], + "spans": [ + { + "bbox": [ + 49, + 515, + 287, + 547 + ], + "type": "text", + "content": "[43] B. Recht, Rebecca Roelofs, L. Schmidt, and V. Shankar. Do ImageNet classifiers generalize to ImageNet? In International Conference on Machine Learning, 2019. 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 49, + 548, + 287, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 548, + 287, + 568 + ], + "spans": [ + { + "bbox": [ + 49, + 548, + 287, + 568 + ], + "type": "text", + "content": "[44] Adria Ruiz and Jakob Verbeek. Anytime inference with distilled hierarchical neural ensembles. In AAAI, 2021. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 49, + 570, + 287, + 591 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 570, + 287, + 591 + ], + "spans": [ + { + "bbox": [ + 49, + 570, + 287, + 591 + ], + "type": "text", + "content": "[45] Guocong Song and Wei Chai. Collaborative learning for deep neural networks. NeurIPS, 31, 2018. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 49, + 593, + 287, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 593, + 287, + 635 + ], + "spans": [ + { + "bbox": [ + 49, + 593, + 287, + 635 + ], + "type": "text", + "content": "[46] Nitish Srivastava, Geoffrey Hinton, Alex Krizhevsky, Ilya Sutskever, and Ruslan Salakhutdinov. Dropout: A simple way to prevent neural networks from overfitting. In Journal of Machine Learning Research, 2014. 1, 2, 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 49, + 637, + 287, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 637, + 287, + 679 + ], + "spans": [ + { + "bbox": [ + 49, + 637, + 287, + 679 + ], + "type": "text", + "content": "[47] Christian Szegedy, Vincent Vanhoucke, Sergey Ioffe, Jonathon Shlens, and Zbigniew Wojna. Rethinking the inception architecture for computer vision. Conference on Computer Vision and Pattern Recognition, 2016. 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 49, + 681, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 681, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 49, + 681, + 287, + 712 + ], + "type": "text", + "content": "[48] Mingxing Tan and Quoc V. Le. Efficientnetv2: Smaller models and faster training. In International Conference on Machine Learning, 2021. 8" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 709 + ], + "type": "list", + "angle": 0, + "index": 34, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "text", + "content": "[49] Antti Tarvainen and Harri Valpola. Mean teachers are better role models: Weight-averaged consistency targets improve semi-supervised deep learning results. Neurips, 30, 2017. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 107, + 545, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 107, + 545, + 149 + ], + "spans": [ + { + "bbox": [ + 308, + 107, + 545, + 149 + ], + "type": "text", + "content": "[50] Hugo Touvron, M. Cord, M. Douze, F. Massa, Alexandre Sablayrolles, and H. Jégou. Training data-efficient image transformers & distillation through attention. International Conference on Machine Learning, 2021. 2, 3, 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 151, + 545, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 151, + 545, + 182 + ], + "spans": [ + { + "bbox": [ + 308, + 151, + 545, + 182 + ], + "type": "text", + "content": "[51] Hugo Touvron, Matthieu Cord, and Hervé Jégou. Deit III: Revenge of the vit. arXiv preprint arXiv:2204.07118, 2022. 5, 6, 7, 8, I, II" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 183, + 545, + 225 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 183, + 545, + 225 + ], + "spans": [ + { + "bbox": [ + 308, + 183, + 545, + 225 + ], + "type": "text", + "content": "[52] Hugo Touvron, Matthieu Cord, Alexandre Sablayrolles, Gabriel Synnaeve, and Hervé Jégou. Going deeper with image transformers. International Conference on Computer Vision, 2021. 4, 5" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 227, + 545, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 227, + 545, + 270 + ], + "spans": [ + { + "bbox": [ + 308, + 227, + 545, + 270 + ], + "type": "text", + "content": "[53] Longhui Wei, An Xiao, Lingxi Xie, X. Chen, Xiaopeng Zhang, and Q. Tian. Circumventing outliers of autoaugment with knowledge distillation. In European Conference on Computer Vision, 2020. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 271, + 545, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 271, + 545, + 303 + ], + "spans": [ + { + "bbox": [ + 308, + 271, + 545, + 303 + ], + "type": "text", + "content": "[54] Ross Wightman. Pytorch image models. https://github.com/rwrightman/pytorch-image-models, 2019.4" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 304, + 545, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 304, + 545, + 335 + ], + "spans": [ + { + "bbox": [ + 308, + 304, + 545, + 335 + ], + "type": "text", + "content": "[55] Ross Wightman, Hugo Touvron, and Hervé Jégou. Resnet strikes back: An improved training procedure in timm. arXiv preprint arXiv:2110.00476, 2021. 5, 7, 8, II" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 337, + 545, + 412 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 337, + 545, + 412 + ], + "spans": [ + { + "bbox": [ + 308, + 337, + 545, + 412 + ], + "type": "text", + "content": "[56] Mitchell Wortsman, Gabriel Ilharco, Samir Ya Gadre, Rebecca Roelofs, Raphael Gontijo-Lopes, Ari S Morcos, Hongseok Namkoong, Ali Farhadi, Yair Carmon, Simon Kornblith, et al. Model soups: averaging weights of multiple fine-tuned models improves accuracy without increasing inference time. In International Conference on Machine Learning, 2022. 3, 4" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 414, + 545, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 414, + 545, + 456 + ], + "spans": [ + { + "bbox": [ + 308, + 414, + 545, + 456 + ], + "type": "text", + "content": "[57] Tete Xiao, Yingcheng Liu, Bolei Zhou, Yuning Jiang, and Jian Sun. Unified perceptual parsing for scene understanding. In European Conference on Computer Vision, 2018. 7, 8" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 457, + 545, + 490 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 457, + 545, + 490 + ], + "spans": [ + { + "bbox": [ + 308, + 457, + 545, + 490 + ], + "type": "text", + "content": "[58] Jason Yosinski, Jeff Clune, Yoshua Bengio, and Hod Lipson. How transferable are features in deep neural networks? arXiv preprint arXiv:1411.1792, 2014. 1" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 491, + 545, + 522 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 491, + 545, + 522 + ], + "spans": [ + { + "bbox": [ + 308, + 491, + 545, + 522 + ], + "type": "text", + "content": "[59] L. Yuan, F. Tay, G. Li, T. Wang, and Jiashi Feng. Revisit knowledge distillation: a teacher-free framework. Conference on Computer Vision and Pattern Recognition, 2020. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 308, + 523, + 545, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 523, + 545, + 567 + ], + "spans": [ + { + "bbox": [ + 308, + 523, + 545, + 567 + ], + "type": "text", + "content": "[60] Sangdoo Yun, Dongyoon Han, Seong Joon Oh, Sanghyuk Chun, Junsuk Choe, and Youngjoon Yoo. CutMix: Regularization strategy to train strong classifiers with localizable features. arXiv preprint arXiv:1905.04899, 2019. 1" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 308, + 567, + 545, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 567, + 545, + 599 + ], + "spans": [ + { + "bbox": [ + 308, + 567, + 545, + 599 + ], + "type": "text", + "content": "[61] Hongyi Zhang, Moustapha Cisse, Yann N. Dauphin, and David Lopez-Paz. mixup: Beyond empirical risk minimization. arXiv preprint arXiv:1710.09412, 2017. 1" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 308, + 601, + 545, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 601, + 545, + 643 + ], + "spans": [ + { + "bbox": [ + 308, + 601, + 545, + 643 + ], + "type": "text", + "content": "[62] Ying Zhang, Tao Xiang, Timothy M Hospedales, and Huchuan Lu. Deep mutual learning. In Conference on Computer Vision and Pattern Recognition, pages 4320-4328, 2018. 2" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 308, + 644, + 545, + 687 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 644, + 545, + 687 + ], + "spans": [ + { + "bbox": [ + 308, + 644, + 545, + 687 + ], + "type": "text", + "content": "[63] Bolei Zhou, Hang Zhao, Xavier Puig, Sanja Fidler, Adela Barriuso, and Antonio Torralba. Scene parsing through ade20k dataset. Conference on Computer Vision and Pattern Recognition, 2017. 7" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 308, + 689, + 545, + 709 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 689, + 545, + 709 + ], + "spans": [ + { + "bbox": [ + 308, + 689, + 545, + 709 + ], + "type": "text", + "content": "[64] Xiatian Zhu, Shaogang Gong, et al. Knowledge distillation by on-the-fly native ensemble. Neurips, 31, 2018. 2" + } + ] + } + ], + "index": 33 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "text", + "content": "11710" + } + ] + } + ], + "index": 35 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/CoMFormer_ Continual Learning in Semantic and Panoptic Segmentation/9493268b-7f35-48c3-8aa1-39be491edc53_content_list.json b/2023/CoMFormer_ Continual Learning in Semantic and Panoptic Segmentation/9493268b-7f35-48c3-8aa1-39be491edc53_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..71da2a36cc286df2297ca38335499d4360a2ffd4 --- /dev/null +++ b/2023/CoMFormer_ Continual Learning in Semantic and Panoptic Segmentation/9493268b-7f35-48c3-8aa1-39be491edc53_content_list.json @@ -0,0 +1,1539 @@ +[ + { + "type": "text", + "text": "CoMFormer: Continual Learning in Semantic and Panoptic Segmentation", + "text_level": 1, + "bbox": [ + 109, + 130, + 859, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Fabio Cermelli $^{1,2*}$ , Matthieu Cord $^{3,4}$ , Arthur Douillard $^{3\\dagger}$", + "bbox": [ + 258, + 179, + 712, + 196 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Politecnico di Torino, $^{2}$ Italian Institute of Technology, $^{3}$ Sorbonne Université, $^{4}$ valeo.ai", + "bbox": [ + 135, + 198, + 831, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "fabio.cermelli@polito.it, matthieu.cord@sorbonne-universite.fr, ar.douillard@gmail.com", + "bbox": [ + 99, + 219, + 859, + 232 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 268, + 313, + 284 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Continual learning for segmentation has recently seen increasing interest. However, all previous works focus on narrow semantic segmentation and disregard panoptic segmentation, an important task with real-world impacts. In this paper, we present the first continual learning model capable of operating on both semantic and panoptic segmentation. Inspired by recent transformer approaches that consider segmentation as a mask-classification problem, we design CoMFormer. Our method carefully exploits the properties of transformer architectures to learn new classes over time. Specifically, we propose a novel adaptive distillation loss along with a mask-based pseudo-labeling technique to effectively prevent forgetting. To evaluate our approach, we introduce a novel continual panoptic segmentation benchmark on the challenging ADE20K dataset. Our CoMFormer outperforms all the existing baselines by forgetting less old classes but also learning more effectively new classes. In addition, we also report an extensive evaluation in the large-scale continual semantic segmentation scenario showing that CoMFormer also significantly outperforms state-of-the-art methods.", + "bbox": [ + 75, + 300, + 473, + 618 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 646, + 209, + 662 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Image segmentation is a fundamental computer vision problem that enables machines to assign an image's pixels to discrete segments. Multiple segmentation tasks have been defined depending on the segments definitions. Semantic segmentation clusters pixels by classes, merging in a single segment pixels belonging to instances of the same class. Panoptic segmentation assigns to every pixel a semantic class while separating different instances into different segments. This latter kind of segmentation has real-world impacts in autonomous robots and vehicles [7, 41].", + "bbox": [ + 75, + 671, + 468, + 821 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Despite tremendous progress in image segmentation, the current approaches are trained on a static dataset with a pre", + "bbox": [ + 76, + 823, + 468, + 853 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/b8d3347a7632053e053c6c4995d36b1c1c99a69545bc93bac49c7dac066977d0.jpg", + "image_caption": [ + "Step t-1 - New Class: Car", + "Image", + "CoMFormer (t-1)" + ], + "image_footnote": [], + "bbox": [ + 504, + 277, + 684, + 373 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/9967a26e5908b0367f68e04b1f5dce02f2397f78defc3567ae5a1a9360a2e5f0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 686, + 279, + 815, + 357 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/0dff6aec1f26369fcf162fe0e066f1246e89d7c6f19a07a4cf9e31a8f5eba346.jpg", + "image_caption": [ + "Predictions" + ], + "image_footnote": [], + "bbox": [ + 821, + 277, + 890, + 375 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/249792d01ea386ca4a040a0c03d29c59e249569a2a81740d16f2510f1f498045.jpg", + "image_caption": [ + "Image", + "Annotations" + ], + "image_footnote": [], + "bbox": [ + 503, + 416, + 609, + 486 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/4b01263540676814d6bda30c505c2ef988424ff64e4e22e40641f9f3af7b13c2.jpg", + "image_caption": [ + "Step t - New Class: Person", + "CoMFormer (t)" + ], + "image_footnote": [], + "bbox": [ + 612, + 410, + 815, + 483 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/756eccefecc29ee7a7d99a8af6be340e0df1e397300f4c2b43762d9d604c6c58.jpg", + "image_caption": [ + "Predictions", + "Figure 1. Illustration of our model, CoMFormer, operating in continual segmentation. Relying on the mask classification paradigm, it is able to cope with both continual semantic and panoptic segmentation without any modification by predicting masks for both old (e.g. car in red) and new (e.g. person in green) classes. The figure reports two classes and no \"stuff\" (e.g. road, building) only for illustration purposes." + ], + "image_footnote": [], + "bbox": [ + 821, + 391, + 885, + 491 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "defined set of classes. Whenever an update of the model is required to fit new classes, the common solution is to train a model from scratch on the union of the old and new class data. A computationally more efficient solution would be to fine-tune the existing model solely on the new class data. Unfortunately, this approach would cause a catastrophic forgetting [21] of the old classes on which the model performance would be extremely degraded.", + "bbox": [ + 496, + 626, + 890, + 747 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The problem of updating the knowledge of the model over time is typically referred as continual learning. It has been traditionally studied in the context of image classification [17, 19, 29, 33, 43, 45] and is gaining attention on the segmentation task [3, 4, 15, 39, 59] due to the more realistic applications and the additional challenges it introduces, such as the background shift [4]. However, current state-of-the-art methods mainly focus on semantic segmentation and are not designed to work in other segmentation tasks, strongly limiting their application in the real world.", + "bbox": [ + 496, + 750, + 892, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 807, + 44 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Work done during the visiting period at Sorbonne Université.", + "bbox": [ + 93, + 862, + 423, + 875 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "$^{\\dagger}$ Work done at Sorbonne Université, currently affiliated to DeepMind.", + "bbox": [ + 98, + 875, + 462, + 887 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "", + "bbox": [ + 98, + 887, + 377, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "3010", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this paper, we design the first method operating in both continual semantic and panoptic segmentation, as illustrated in Fig. 1. Our method, CoMFormer (Continual MaskFormer), takes inspiration from recent transformer architectures [11, 12], approaching segmentation as a mask classification problem. Instead of predicting a class probability for each pixel, as in previous semantic segmentation works [9, 37], it predicts a set of binary masks, each associated with a single class prediction, effectively addressing both segmentation tasks without any modification in the training architecture and procedure. Differently from previous works [11, 12], however, CoMFormer forces the output binary masks to be mutually exclusive to one another: a pixel can only be predicted by a single binary mask to prevent having several masks classifying the same pixel with different classes. This behavior is crucial in continual learning to reduce the interference among old and new classes.", + "bbox": [ + 75, + 90, + 472, + 347 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Furthermore, CoMFormer introduces a novel adaptive distillation loss to alleviate forgetting. It enforces consistency of the model's classification predictions across learning steps only when it is useful to remember old classes, ensuring a better tradeoff between rigidity (not forgetting old classes) and plasticity (learning efficiently new classes). Finally, since at each training iteration the dataset reports annotations only for the current classes, we design a mask-based pseudo-labeling technique to generate annotations for the old classes, effectively alleviating forgetting. To reduce the noise, we consider the prediction confidence and we avoid interference with ground-truth annotations.", + "bbox": [ + 75, + 347, + 472, + 527 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We validate CoMFormer on both continual segmentation tasks. For panoptic segmentation, we define a new benchmark relying on the challenging ADE20K where we demonstrate that CoMFormer largely outperforms all previous baselines. On semantic segmentation, we show that CoMFormer outperforms the existing state-of-the-art methods on every setting of the large-scale ADE20K benchmark.", + "bbox": [ + 75, + 529, + 472, + 635 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To sum up, the contributions of this paper are as follows:", + "bbox": [ + 96, + 636, + 467, + 651 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We introduce continual panoptic segmentation which has real-world impacts in addition to being significantly more challenging than previous benchmarks.", + "- We propose CoMFormer to tackle both continual panoptic and semantic segmentation. To avoid forgetting, we design a novel adaptive distillation and an efficient mask-based pseudo-labeling strategy.", + "- Through extensive quantitative and qualitative benchmarks, we showcase the state-of-the-art performance of our model on both continual segmentation tasks." + ], + "bbox": [ + 94, + 652, + 467, + 801 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 76, + 811, + 225, + 828 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Semantic and Panoptic Segmentation. The two tasks have been traditionally treated separately, with specialized architectures proposed for either one or the other task, without interoperability. Semantic segmentation has been tradi", + "bbox": [ + 75, + 839, + 470, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "tionally addressed as a per-pixel classification task. Fully convolutional network [37] dominated the field by aggregating long-range dependencies in the features map [8, 9, 60] and exploiting contextual information [22, 24, 57, 58, 61]. Recently, transformers [26, 46, 54] are replacing convolutions by integrating long-range dependencies at every layer. Panoptic segmentation [28] has been proposed to unify semantic and instance segmentations. Initially, methods proposed to combine each task-specific architectures [10, 27, 32, 42] or defined new specialized architectures and objective functions for the panoptic task [1, 50, 51], drifting from a general solution for all the segmentation tasks. Recently, methods addressing segmentation as a mask classification problem [11, 12, 50, 56] have been proposed, introducing a transformer architecture able to solve multiple tasks at once. MaskFormer [12] was the first to propose a single architecture to address both panoptic and semantic segmentation. Mask2Former [11] improves it by adopting multi-scale features, masked attention and optimization tricks. Concurrently, kMaX-DeepLab [56] extended [50] proposing to reformulate the cross-attention as a clustering process. Despite their effectiveness in standard training setting, these methods suffer from catastrophic forgetting. In this work, we aim to extend their capability and we propose CoMFormer to learn to segment new classes over time.", + "bbox": [ + 496, + 90, + 893, + 470 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Continual Segmentation. Continual learning is a longstanding field addressing the problem of learning new knowledge over time while avoiding catastrophic forgetting [21, 44, 47]. Traditionally, it has been studied in the context of image classification [6, 17, 19, 29, 33, 43, 45, 55] but there is a growing interest in its applications in segmentation [2-5, 15, 16, 38-40, 59]. Continual semantic segmentation introduces additional challenges, as pointed out by [4]. In particular, considering the task as a per-pixel classification problem, catastrophic forgetting is exacerbated by the background shift, where old classes pixels are treated as background in following training steps. [4] proposed a revision of the standard knowledge distillation framework to deal with it. [17, 39] introduced a distillation loss preserving the representation at feature level while [38] proposed an approach using old classes' examples to alleviate forgetting. Recently, [59] presented a technique that decouples learning the representation of the new classes while freezing the representation for the old ones. Differently from previous work, we extend the benchmarks to solve both semantic and the more challenging panoptic segmentation task.", + "bbox": [ + 496, + 479, + 893, + 799 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Continual Learning with Transformers. Recent progresses made using transformers for computer vision [14, 36, 48, 49] attracted the attention of the continual learning community [19, 52, 53]. In particular, DyTox [19] proposed to specialize the architecture on each task using a different task-specific token. Learning-to-Prompt [53] stores a", + "bbox": [ + 496, + 809, + 893, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "3011", + "bbox": [ + 482, + 944, + 513, + 957 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/2c32f8ab18b9358039da99d591fe4e93e53421fcb9015a9bd11c35250c11d980.jpg", + "image_caption": [ + "Figure 2. Overview of CoMFormer. At learning step t, both the current (bottom) and old (top) models output, for each input query, a pair made of a class probability distribution and a binary mask. While learning the new class (person), we regularize the current model to prevent forgetting the old classes (car) by: (i) generating pseudo-labels for the old classes present in the image and (ii) by applying an adaptive distillation loss that reweights the contribution of each output based on the information it carries on the old classes. $\\psi$ indicates the dot product between the mask embeddings $\\mathcal{E}_{mask}$ and the pixel embeddings $\\mathcal{E}_{pixel}$ followed by the softmax activation." + ], + "image_footnote": [], + "bbox": [ + 106, + 88, + 867, + 292 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "pool of prompts that are employed to condition the whole forward execution of the patch tokens. To the best of our knowledge, transformers have been evaluated only in continual learning for image classification [19, 31] and we are the first to use them in the continual segmentation task.", + "bbox": [ + 75, + 385, + 470, + 460 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. CoMFormer", + "text_level": 1, + "bbox": [ + 76, + 472, + 209, + 488 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Problem Definition", + "text_level": 1, + "bbox": [ + 76, + 497, + 259, + 512 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The goal of image segmentation is to learn a model able to (i) partition the image into a set of N regions represented by binary masks and (ii) produce a class-probability distribution associated with each region. The differences among segmentation tasks rely on the semantics of the masks: semantic segmentation groups all the pixels of a class while panoptic segmentation distinguishes different object instances. In the following, we provide a general formulation for the two tasks since they only differ in the construction of the binary masks in the dataset.", + "bbox": [ + 75, + 521, + 468, + 671 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Continual segmentation aims to train the model in multiple learning steps $t = 1, \\ldots, T$ , introducing at every step a new set of classes. Formally, during the learning step $t$ a dataset $D_{t}$ consisting of a collection of image and label pairs is provided. The label of each image takes the form of a set of $N^{gt}$ ground truth segments $z^{gt}\\{(c_i^{gt}, m_i^{gt}) | c_i^{gt} \\in \\mathcal{K}^t, m_i^{gt} \\in \\{0,1\\}^{H \\times W}\\}_{i=1}^{N^{gt}}$ where $c_i^{gt}$ is the ground-truth class and $m_i^{gt}$ is binary mask $m^{gt}$ . $\\mathcal{K}^t$ is the set of classes introduced at step $t$ , and $H, W$ are the height and width of the images. The goal of training step $t$ is to learn a model able to predict segments for all the seen classes $\\mathcal{K}^{1:t}$ .", + "bbox": [ + 75, + 672, + 470, + 839 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We note that the dataset only contains annotations for the new classes $\\mathcal{K}^t$ while not reporting ground-truth segments for old $\\mathcal{K}^{1:t - 1}$ or future $\\mathcal{K}^{t + 1:T}$ classes. Moreover, differently from previous works [4], old and future class segment", + "bbox": [ + 75, + 839, + 470, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "annotations are simply absent during training rather than being collapsed into an artificial background class.", + "bbox": [ + 498, + 385, + 890, + 417 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. CoMFormer Architecture", + "text_level": 1, + "bbox": [ + 498, + 430, + 736, + 445 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To solve semantic and panoptic segmentation tasks within a single method, we take inspiration from Mask-Former architectures [11, 12] considering segmentation a mask classification problem. It consists in predicting for each image a set of $N$ pairs made of a class prediction and a binary mask. Formally, the output of a CoMFormer is $z = \\{(p_i, m_i)\\}_{i=1}^N$ , where $p_i$ represents a class probability distribution over the seen classes, i.e. $p_i \\in \\Delta^{K^t + 1}$ with $K^t = |\\mathcal{K}^{1:t}|$ , and $m_i$ is a binary mask such that $m_i \\in [0, 1]^{H \\times W}$ . We note that $p_i$ contains an additional \"no object\" class (denoted $\\varnothing$ ) to indicate that the mask does not correspond to any of the known $\\mathcal{K}^{1:t}$ categories.", + "bbox": [ + 496, + 455, + 890, + 638 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To obtain such output, the CoMFormer architecture, showcased in Fig. 2, is made of three components: (i) a backbone that extracts feature embeddings $\\mathcal{E}_{\\text {feat }}$ , (ii) a transformer decoder that takes as input N learnable queries and $\\mathcal{E}_{\\text {feat }}$ to output $N$ per-segment embeddings $Q$ , (iii) a pixel decoder that takes $\\mathcal{E}_{\\text {feat }}$ as input and extracts per-pixel embeddings $\\mathcal{E}_{\\text {pixel }}$ . The output class probabilities $\\{p_i \\in \\Delta^{K^t + 1}\\}_{i=1}^N$ are obtained by applying a linear classifier on $Q$ , which is enlarged as new classes are presented. We obtain the mask predictions $m_i$ combining $Q$ and $\\mathcal{E}_{\\text {pixel }}$ : first $Q$ is passed into a 2-layer MLP to obtain N mask embeddings $\\mathcal{E}_{\\text {mask }}$ and we then compute the dot product between the $i$ -th mask embedding and the pixel embeddings, followed by the softmax activation: $m_i[h, w] = \\operatorname{softmax}(\\mathcal{E}_{\\text {mask}}[:, i]^T \\cdot \\mathcal{E}_{\\text {pixel}}[:, h, w])$ .", + "bbox": [ + 496, + 640, + 892, + 869 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We point out that employing the softmax contrasts sharply with MaskFormers [11, 12], which use sigmoid ac", + "bbox": [ + 498, + 869, + 890, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3012", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/a565ed99ec382a9209e368e4e7d0ebbf02cf6e43a6821034004cc1717fc002b1.jpg", + "image_caption": [ + "Figure 3. Illustration of the mask predictions between sigmoid and softmax. MaskFormer has overlap across masks (e.g. in Query 1 and 2) predicted either as old or new classes, while CoMFormer avoids the issue using the softmax activation." + ], + "image_footnote": [], + "bbox": [ + 78, + 85, + 444, + 258 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "tivation to get the mask predictions. Using the sigmoid activation would result in the overlap of many masks on a single pixel, each potentially belonging to a distinct class, as seen in Fig. 3. This is particularly important in the context of continual learning, when the model may predict two similar masks associated with two distinct classes (e.g. cat and dog classes in Fig. 3) thereby generating interference among old and new classes and degrading the model's performance.", + "bbox": [ + 75, + 330, + 470, + 452 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Learning without Forgetting", + "text_level": 1, + "bbox": [ + 76, + 460, + 333, + 474 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To learn new classes, a simple solution is to fine-tune the model on the new dataset. However, because it lacks annotations for old classes, this operation results in catastrophic forgetting [21]. To alleviate the issue, we introduce two techniques for avoiding forgetting: an adaptive distillation loss and a mask-based pseudo-labeling strategy.", + "bbox": [ + 75, + 483, + 468, + 574 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Adaptive Distillation Loss. A common strategy to mitigate forgetting adopted by continual semantic segmentation approaches [4, 15, 39] is to distill the knowledge coming from the old model in the new one, either in form of features [15, 39] or as classification probabilities [4]. We take inspiration from the latter solution and we design a novel distillation loss tailored to the CoMFormer architecture. In particular, analyzing the two components constituting the output space, we find the mask prediction being robust to forgetting while the classification part is heavily affected (see Sec. 4.6). For this reason, we focus on the output class probabilities to design our distillation loss.", + "bbox": [ + 75, + 580, + 468, + 762 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "An adaptation of standard knowledge distillation works [4] would force the new model to mimic the old model probabilities over all the N outputs, independently from their content. Formally, given the set of N probability distributions $\\{p_i^o\\}_{i = 1}^N$ coming from the old model, a standard distillation loss is computed as", + "bbox": [ + 75, + 762, + 468, + 852 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {K D} = - \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\sum_ {k = 1} ^ {K ^ {t - 1} + 1} p _ {i} ^ {o} (k) \\log \\frac {\\tilde {p} _ {i} (k)}{p _ {i} ^ {o} (k)}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 130, + 859, + 468, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $K^{t - 1} + 1$ is the $\\varnothing$ class and $\\tilde{p}_i(k)$ represents the unbiased probability distribution of $i$ -th output for class $k$ :", + "bbox": [ + 498, + 90, + 890, + 122 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {p} _ {i} (k) = \\left\\{ \\begin{array}{l l} \\sum_ {j \\in \\mathcal {K} ^ {t}} p _ {i} (j) + p _ {i} (\\varnothing) & \\text {i f} k = \\varnothing , \\\\ p _ {i} (k) & \\text {o t h e r w i s e .} \\end{array} \\right. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 540, + 133, + 890, + 172 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Note that, similarly to [4], it forces the sum of the model new class $(\\mathcal{K}^t)$ and the \"no object\" class $(\\varnothing)$ probabilities to be similar to the old model $\\varnothing$ probability.", + "bbox": [ + 498, + 184, + 890, + 229 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "However, not every output probability brings relevant information: most of them are predicted as \"no object\" $(\\varnothing)$ and thus they do not carry enough details to remember the old classes while reducing the relative importance of the other outputs probabilities. For this reason, we re-weight their contribution based on the probability of not being $\\varnothing$ . Formally, we define our adaptive distillation loss as", + "bbox": [ + 498, + 231, + 890, + 335 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {A D} (p, p ^ {o}) = - \\sum_ {i = 1} ^ {N} \\frac {\\omega_ {i}}{\\sum_ {i = 1} ^ {N} \\omega_ {i}} \\sum_ {k = 1} ^ {K ^ {t - 1} + 1} p _ {i} ^ {o} (k) \\log \\frac {\\tilde {p} _ {i} (k)}{p _ {i} ^ {o} (k)}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 509, + 349, + 890, + 405 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where the weighting coefficient is $\\omega_{i} = (1 - p_{i}^{o}(\\varnothing))^{2}$ . Note that, with this formulation, we effectively reduce the contributions of outputs that bring small information about the old classes $(p_i^o (\\varnothing)\\approx 1)$ , while we increase the contribution of important ones (where $p_i^o (\\varnothing)\\approx 0$ ).", + "bbox": [ + 498, + 405, + 890, + 481 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Mask-based Pseudo-labeling Strategy. Forgetting is effectively reduced by the adaptive distillation loss but we are not, however, taking advantage of the presence of the old classes that are unlabeled in the dataset. We propose using the old model predictions to recognize old class segments and generate pseudo-labels to make efficient use of this information and improve the knowledge of past classes.", + "bbox": [ + 498, + 488, + 890, + 594 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "A simple strategy consists in considering the N output pairs coming from the old model independently and using all the pairs for which the predicted class is not $\\varnothing$ as pseudolabels. This strategy, however, does not consider two aspects: there may be pairs where the class is different from $\\varnothing$ but (i) the mask substantially overlaps with a ground-truth segment, or (ii) the mask is noisy and has low confidence.", + "bbox": [ + 498, + 594, + 890, + 700 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To overcome these issues, we propose a mask-based pseudo-labeling strategy that jointly considers the mask and class prediction confidences to avoid noisy labels and overlaps with the existing annotations. We define the model confidence as the multiplication among the class and mask probabilities. Formally, we compute the confidence for the $i$ -th output as $q_{i} = p_{i}^{max} \\cdot m_{i}^{o} \\in [0,1]^{H \\times W}$ , where $p_{i}^{max} = \\max_{k=1}^{K^{t-1}} p_{i}^{o}(k)$ and $m_{i}^{o}$ is the binary mask predicted by the old model. Note that we do not consider the $\\varnothing$ class in the max operation. We denote the pseudo-class $c_{i}^{ps} = \\operatorname{argmax}_{k=1}^{K^{t-1}} p_{i}^{o}(k)$ and we generate the pseudomask $m_{i}^{ps}$ considering two criteria: (i) there should be no overlap between the pseudo-mask and the ground-truth seg", + "bbox": [ + 498, + 700, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "3013", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "ments and (ii) to be included in a pseudo-mask, the confidence on a pixel should be maximum over all outputs. Formally, denoting the binarization of the predicted mask as $m_{i}^{bin} = m_{i} > 0.5$ and the union of the ground-truth segments $m_{all}^{gt} = \\sum_{i=1}^{N^{ps}} m_{i}^{gt}$ , we generate the pseudo-mask as", + "bbox": [ + 76, + 90, + 468, + 172 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nm _ {i} ^ {p s} [ h, w ] = \\left\\{ \\begin{array}{l l} m _ {i} ^ {b i n} [ h, w ] & \\text {i f} i = \\operatorname {a r g m a x} _ {j = 1} ^ {N} q _ {j} [ h, w ] \\\\ & \\text {a n d} m _ {a l l} ^ {g t} [ h, w ] = 0, \\\\ 0 & \\text {o t h e r w i s e .} \\end{array} \\right. \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 91, + 188, + 468, + 241 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Indeed, not all pseudo-mask should be included in the set of pseudo-labels. There may be pseudo-masks where no pixel is active, i.e. it is all zeros, or where the pseudomask contains only a small fraction of the original mask. For these reasons, we construct the set of pseudo-labels $z^{ps} = \\{(c_i^{ps},m_i^{ps})\\}_{i = 1}^{N^{ps}}$ such that we include in $z^{ps}$ only the pseudo-masks that retain at least half of the pixels w.r.t. the binary mask $m_{i}^{bin}$ and that have at least one active pixel.", + "bbox": [ + 76, + 247, + 468, + 367 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We denote the final annotation set $\\bar{z} = \\{\\bar{c}_i, \\bar{m}_i\\}_{i=1}^N$ as the union of the ground-truth labels $z^{gt}$ and the pseudo-labels $z^{ps}$ , i.e. $\\bar{z} = z^{gt} \\cup z^{ps}$ and $\\bar{N} = N^{gt} + N^{ps}$ .", + "bbox": [ + 76, + 368, + 468, + 415 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Overall training loss. To train CoMFormer, a one-to-one matching $\\sigma$ between the model predictions $z$ and the annotation set $\\bar{z}$ is required. Following standard practices [1, 11], we employ the Hungarian matching algorithm [30] and we search for the matching $\\sigma$ that minimize the assignment cost computed as $-p_i(\\bar{c}_j)\\cdot Dice(m_i,\\bar{m}_j)$ , where Dice indicates the Dice coefficient [13]. We assume the size of the prediction set to be larger than the annotations set (i.e. $N > \\bar{N}$ ) and to obtain a one-to-one matching we pad the annotations with \"no object\" tokens $\\varnothing$ .", + "bbox": [ + 75, + 421, + 468, + 571 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Once the best matching $\\sigma$ has been found, we define the overall training loss as:", + "bbox": [ + 76, + 571, + 468, + 602 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\mathcal {L} _ {\\text {S e g}} (z, \\bar {z}) + \\lambda_ {D} \\mathcal {L} _ {A D} (p, p ^ {o}), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 156, + 612, + 468, + 628 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\lambda_{D}$ is a trade-off hyper-parameter and", + "bbox": [ + 76, + 638, + 374, + 652 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} _ {\\text {S e g}} (z, \\bar {z}) = \\sum_ {j = 1} ^ {N} [ - \\alpha (1 - p _ {\\sigma (j)} (\\bar {c} _ {j})) ^ {\\gamma} \\log p _ {\\sigma (j)} (\\bar {c} _ {j}) + \\tag {6} \\\\ \\left. + \\lambda_ {m a s k} \\mathbb {1} _ {\\bar {c} _ {j} \\neq \\emptyset} \\mathcal {L} _ {m a s k} \\left(m _ {\\sigma (j)}, \\bar {m} _ {j}\\right) \\right]. \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 84, + 662, + 468, + 727 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We note that the first term in $\\mathcal{L}_{Seg}$ indicates the focal loss [34], where $\\alpha$ and $\\gamma$ are hyper-parameters. Finally, $\\mathcal{L}_{mask}$ is the sum of dice and cross-entropy losses, and $\\lambda_{mask}$ is a hyper-parameter. We note that $\\mathcal{L}_{mask}$ is computed only for valid segments, i.e. where $\\bar{c}_j \\neq \\emptyset$ .", + "bbox": [ + 76, + 733, + 468, + 810 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 76, + 821, + 207, + 838 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Dataset and Settings", + "text_level": 1, + "bbox": [ + 76, + 847, + 269, + 863 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We start from the widely adopted continual semantic segmentation benchmark defined in [4] and we extend it to", + "bbox": [ + 76, + 869, + 468, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "the newly proposed continual panoptic segmentation using a modified Continuum library [18]. We compare our novel model on the large-scale ADE20K [62] dataset since it supports both semantic and panoptic segmentation tasks. This challenging dataset contains 150 classes, divided into 100 \"things\" and 50 \"stuff\" categories. This dataset represents a wide variety of scenes, both interior and exterior, with images featuring an average of 9.9 classes, while other datasets like COCO [35] only have an average of 3.5 classes.", + "bbox": [ + 496, + 90, + 890, + 227 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Continual Learning Protocols. Previous continual semantic segmentation works [4, 15] describe a training protocol to assess the performance on multiple continual learning steps. In particular, [4] proposed three protocols for ADE20K with different numbers of tasks: (i) the 100-50 consists of two tasks, the first of 100 and the second of 50 classes; (ii) the 50-50 consists of three tasks of 50 classes; (iii) the 100-10 consists on 6 tasks, the first on 100 classes followed by 5 tasks of 10 classes. In addition, we use the 100-5 introduced in [15], consisting of 11 tasks, the first of 100 classes followed by 10 tasks of 5 classes. To divide the images into different tasks, we follow the splits provided by [4] for 100-50, 100-10, 50-50 for both semantic and panoptic segmentation. In particular, they ensured that each image appears only on a unique task. Differently, for the 100-5 we use the split proposed by [15], where the same image may appear on multiple tasks. We report the results of the 50-50 setting for both semantic and panoptic segmentation in the supplementary material.", + "bbox": [ + 496, + 234, + 890, + 522 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Metrics. We compare the different models using the mean Intersection over Union (mIoU) on semantic segmentation [20] and Panoptic Quality (PQ) on panoptic segmentation [28]. PQ is the product of two components: Segmentation Quality (SQ) which considers the IoU between the correctly classified segments, and Recognition Quality (RQ) which only considers the classification accuracy. For both tasks, we report the metric after the last step $T$ on the first classes $(C^1)$ , for the added classes $(C^{2:T})$ , and for all the classes (all). We also report the average of the final performance on all seen classes after each step (avg) following [15].", + "bbox": [ + 496, + 530, + 890, + 698 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Baselines", + "text_level": 1, + "bbox": [ + 500, + 709, + 607, + 723 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We benchmark our model against the state-of-the-art methods: MiB [4], PLOP [15], and RCIL [59]. In semantic segmentation, we report them with the original DeepLab-v3 segmentation architecture [9]. In addition, for a fair comparison, we report MiB and PLOP using the CoMFormer architecture. In particular, we implement the methods in two versions: (i) we replace the CoMFormer losses with a per-pixel cross-entropy loss on the dot product between mask and class logits and we apply the continual learning methods on that baseline (we refer to this as Per-Pixel), and (ii) by applying the distillation losses directly on the", + "bbox": [ + 496, + 734, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "3014", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/37dc5e3bba87f7e317eb9177039f1ee33715a70eda57e250b8b25100964ef948.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Method100-50 (2 tasks)100-10 (6 tasks)100-5 (11 tasks)
1-100101-150allavg1-100101-150allavg1-100101-150allavg
FT0.025.88.624.80.02.91.07.90.01.30.44.6
MiB35.119.329.835.427.110.021.429.124.06.518.125.6
PLOP41.026.636.238.630.517.526.132.928.115.724.030.5
CoMFormer41.127.736.738.836.017.129.735.334.415.928.234.0
Joint43.232.139.5—43.232.139.5—43.232.139.5—
", + "bbox": [ + 112, + 87, + 857, + 189 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/62191e1965c0a981a7970352023c96b01ab7eb0e8f1fc8a4e48f5179535b2283.jpg", + "table_caption": [ + "Table 1. Continual Panoptic Segmentation results on ADE20K dataset in PQ considering multiple training protocols." + ], + "table_footnote": [], + "table_body": "
BackboneMethod100-50 (2 tasks)100-10 (6 tasks)100-5 (11 tasks)
1-100101-150allavg1-100101-150allavg1-100101-150allavg
DeepLab-v3 [9]MiB [4]40.517.232.837.338.311.329.235.136.05.726.032.7
PLOP [15]41.914.932.937.440.514.131.636.639.17.828.835.3
RCIL [59]42.318.834.5—39.317.632.1—38.511.529.6—
Per-PixelMiB40.324.034.837.535.114.028.134.033.315.227.333.1
PLOP40.220.233.536.932.613.726.332.433.39.425.432.8
Mask-basedFT0.026.78.926.40.02.30.88.50.01.10.34.2
MiB37.024.132.638.323.510.626.629.621.06.116.127.7
PLOP44.226.238.241.134.815.928.535.233.614.127.133.6
CoMFormer44.726.238.441.240.615.632.337.439.513.630.936.5
Joint46.935.643.1—46.935.643.1—46.935.643.1—
", + "bbox": [ + 102, + 228, + 867, + 400 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 2. Continual Semantic Segmentation results on the ADE20K dataset in mIoU considering multiple training protocols. We report the numbers published by the original papers on DeepLab-v3 [9]. The remaining results are based on our re-implementation (see Sec. 4.2).", + "bbox": [ + 75, + 411, + 893, + 440 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "classification head of the CoMFormer architecture (Mask-based). We note that the first version cannot be applied to panoptic segmentation, since the task does not support the use of per-pixel losses. For PLOP, we apply its distillation loss local-POD on the intermediate features of the segmentation backbone, as in the original paper, and we use the pseudo-labeling strategy proposed in Sec. 3. RCIL was not re-implemented since it is not possible to extend their approach to non-convolutional architectures like ours. We include a naive finetuning (FT) without any capabilities for continual learning. Finally, we also report an upper bound (Joint) that is trained in the traditional segmentation setting. We made our best effort to find the most suitable hyperparameters for all the baselines.", + "bbox": [ + 75, + 454, + 472, + 667 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Implementation Details", + "text_level": 1, + "bbox": [ + 76, + 678, + 294, + 694 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Architecture. Following the previous benchmarks [4, 15] we use as backbone a ResNet101 [23] on semantic segmentation and we propose to use a ResNet50 [23] for panoptic segmentation. For both tasks, we use the pixel decoder and the transformer decoder proposed in [11] with 9 layers in total and $N = 100$ queries. In addition, we adopt all the improvements introduced in [11]: masked attention, multiscale features, and their optimization improvements. We report the performance of single-scale inference.", + "bbox": [ + 75, + 696, + 468, + 833 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Training parameters. We follow the Mask2Former [11] hyper-parameters and we use AdamW [25] optimizer with an initial learning rate of 0.0001 for the first step ( $t = 0$ ) and 0.00005 in the following ( $t > 0$ ). A learning rate multiplier", + "bbox": [ + 75, + 839, + 470, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "of 0.1 is applied to the backbone and we follow a polynomial learning rate schedule. We use weight decay of 0.05. We train the model for 160K iterations in the first step and 400 iterations per class in the following (e.g. learning 50 classes we train for $50 \\times 400 = 20000$ iterations). For semantic segmentation, we use a crop size of 512 and a batch size of 16. For panoptic segmentation, we use a crop size of 640 and a batch size of 8. For both, we use the standard random scale jittering between 0.5 and 2.0, random horizontal flipping, random cropping, as well as random color jittering as data augmentation [12]. We set $\\alpha = 20$ , $\\gamma = 2$ , and $\\lambda_{mask} = 5.0$ . $\\lambda_{D}$ is set to 1 for 100-50 and to 10 in the 100-10 and 100-5. Finally, following the standard protocol of continual learning [4, 15] we do not store any image of previous steps and we do not use rehearsal learning.", + "bbox": [ + 496, + 454, + 893, + 681 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.4. Continual Panoptic Segmentation", + "text_level": 1, + "bbox": [ + 498, + 694, + 795, + 710 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Tab. 1 reports the results on the new continual panoptic segmentation benchmark. CoMFormer exceeds all the baselines by a significant margin. Considering the 100-50, we can see that CoMFormer outperforms MiB on both old (+6 PQ) and new (+8.4 PQ) classes and PLOP on new classes (+1.1 PQ). Considering a longer sequence of tasks (100-10 and 100-5), we can see that the gap between CoMFormer and the other methods becomes more significant, especially in the old classes. In 100-10, it surpasses MiB and PLOP respectively by 8.9 and 5.5 PQ in old classes, while in the new classes it exceeds MiB by 7.1 and is comparable to PLOP. In the 100-5, it outperforms MiB and PLOP", + "bbox": [ + 496, + 719, + 893, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "3015", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/b87103b4b638a59f18c508ea4fd257ba27f4faa0c031eb21538dd618859f39c2.jpg", + "image_caption": [ + "Figure 4. Qualitative results of CoMformer v.s. MiB and PLOP on the 100-10 continual panoptic segmentation setting on ADE20K." + ], + "image_footnote": [], + "bbox": [ + 99, + 88, + 874, + 494 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "on both old (+10.4 and +6.3 PQ) and new classes (+9.4 and +0.2 PQ). Finally, we note that CoMFormer obtains performance close to the Joint baseline in the 100-50 setting, suffering a drop of only 2.1 PQ on the old classes and 4.4 PQ on the new ones. In the other settings, the gap is still relevant, indicating that learning from multiple training steps is still a very challenging task.", + "bbox": [ + 75, + 534, + 468, + 638 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In Fig. 4, we report a qualitative comparison with the baselines in the 100-10 setting using randomly chosen images. We can see that our method is able to maintain better the previous knowledge by precisely segmenting all the classes in the image; for example, w.r.t. to PLOP, our method correctly segments the chair in the second row and the boat in the third row. We can note a common error: models correctly segment but misclassify objects in the image, like the plane classified as car in the first row or the press classified as a door in the third row.", + "bbox": [ + 75, + 640, + 470, + 791 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.5. Continual Semantic Segmentation", + "text_level": 1, + "bbox": [ + 76, + 801, + 375, + 816 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We report the comparison with state-of-the-art methods in continual semantic segmentation in Tab. 2. CoMFormer achieves the best results on every setting, surpassing both methods based on DeepLab-v3 [9] and baselines implemented upon the same architecture. In particular, CoM-", + "bbox": [ + 75, + 825, + 468, + 902 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Former outperforms the best state-of-the-art method (RCIL [59] by $3.9\\%$ on the 100-50, by $0.2\\%$ on 100-10, and by $1.3\\%$ on the 100-5. Moreover, it always improves the performance on the old classes and the avg across different steps. In addition, CoMFormer exceeds all the baselines. Comparing it with PLOP (Mask-based), we can see that it obtains a substantial improvement, especially considering scenarios with longer training protocols: CoMformers achieves $30.9\\%$ $(32.3\\%)$ on the all metrics, while PLOP only $27.1\\%$ $(28.5\\%)$ on the 100-5 $(100 - 10)$ setting. Furthermore, the results confirm that improvements are not led by the newer architecture. Per-Pixel baselines achieve only similar or even lower performance than their counterparts. MiB (Per-Pixel), for example, slightly improves the one based on DeepLab-v3 on the 100-50 $(+2.0\\%$ on all) and in the 100-5 $(+1.3\\%)$ but its performance is slightly worse in the 100-10 $(-1.1\\%)$ . In addition, Per-Pixel baselines demonstrate a large performance drop in the old classes, showing that, without employing a distillation strategy tailored to the architecture, it is difficult to maintain satisfactory results.", + "bbox": [ + 496, + 532, + 890, + 835 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.6. Ablation Studies", + "text_level": 1, + "bbox": [ + 500, + 849, + 663, + 864 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Mask2Former vs CoMFormer. In Tab. 3, we compare the proposed CoMFormer against Mask2Former [11]. In", + "bbox": [ + 498, + 869, + 892, + 901 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "3016", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/3f374d9a7cdc3dd4c92500cfe5efc61f93e5dc131ffea2747179ed09a8ccda0b.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodSemantic - mIoUPanoptic - PQ
1-100101-150all1-100101-150all
100-50 (2 tasks)
Mask2Former44.326.038.137.924.533.4
CoMFormer44.726.238.441.127.736.7
100-10 (6 tasks)
Mask2Former38.715.731.030.210.823.7
CoMFormer40.615.632.336.017.129.7
100-5 (11 tasks)
Mask2Former32.613.726.325.95.619.1
CoMFormer39.513.630.934.415.928.2
", + "bbox": [ + 86, + 88, + 460, + 247 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/47a81d61d84a9eb2834060a4385d4baaa2030b778d385f655f506063f7c97902.jpg", + "table_caption": [ + "Table 3. Comparison between CoMFormer and Mask2Former [11] in continual semantic and panoptic segmentation. We apply to Mask2Former our mask-based pseudo-labeling strategy and the adaptive distillation loss for a fair comparison." + ], + "table_footnote": [], + "table_body": "
100-50 (2 tasks)100-10 (6 tasks)
PSKDAD1-100101-150all1-100101-150all
---0.025.88.60.02.91.0
--✓28.415.324.023.09.118.4
✓--40.726.335.930.117.225.8
✓✓-40.528.136.430.217.626.0
✓-✓41.127.736.736.017.129.7
", + "bbox": [ + 86, + 328, + 460, + 431 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "particular, to fairly compare the two methods, we add our adaptive distillation loss and the pseudo-labeling strategy to Mask2Former. Considering semantic segmentation, CoMFormer achieves better results on the old classes (on average $3\\%$ mIoU), while the two methods obtain comparable results on the new classes. On panoptic segmentation, CoMFormer exceeds Mask2Former in all the classes with a considerable gap (6 PQ on both old and new classes on average). We conclude that a vanilla MaskFormer architecture is not suited to continual learning because it suffers from forgetting, while CoMFormer outperforms it in any setting.", + "bbox": [ + 75, + 513, + 468, + 681 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Method components. We investigate the benefit of each method component in Tab. 4, reporting the results in PQ on the 100-50 and 100-10 settings in panoptic segmentation. Without applying any regularization technique, the performance on old classes drops to 0, denoting that it is important to deal with forgetting. Both the adaptive distillation loss (AD) and the pseudo-labeling strategy (PS) lead to improvement in the old classes when applied independently. However, we obtain the best results when combining them, especially considering the 100-10 scenario, where the old classes are improved by 5.8 PQ and the final performances of 3.9 PQ. Finally, the table shows that weighting the queries based on their semantic content is essential to improve performance: comparing a standard distillation", + "bbox": [ + 75, + 688, + 470, + 901 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/e72086cdd253abe8cb5f101eb96615153796bacf738f2cd0ee264e875ad74b44.jpg", + "table_caption": [ + "Table 4. Ablation of the different key components of our CoM-Former architecture. We report the PQ on the 100-50 and 100-10 panoptic settings. PS: Pseudo-labeling; AD: adaptive distillation Eq. (3); KD: standard distillation Eq. (1)" + ], + "table_footnote": [], + "table_body": "
Method100-50 (2 tasks)100-10 (6 tasks)
\\( RQ \\)\\( SQ \\)\\( PQ \\)\\( RQ \\)\\( SQ \\)\\( PQ \\)
MiB36.873.029.826.455.521.4
PLOP43.879.136.231.761.126.1
CoMFormer44.479.636.736.273.429.7
Joint47.679.039.547.679.039.5
", + "bbox": [ + 506, + 88, + 883, + 176 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 5. Panoptic metrics analysis. We report Recognition Quality (RQ), Segmentation Quality (SQ), and Panoptic Quality (PQ) over all the classes after the last step.", + "bbox": [ + 498, + 186, + 890, + 229 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "(KD Eq. (1)) with our adaptive distillation (AD Eq. (3)), we largely improve the performance on both settings: $+0.3$ PQ in 100-50 and $+3.7$ PQ in 100-10 considering all classes.", + "bbox": [ + 498, + 243, + 890, + 289 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Forgetting in CoMFormer. To investigate the reason for the forgetting, we considered separately the Recognition Quality (RQ) and the Segmentation Quality (SQ) in panoptic segmentation. RQ measures how well the model is able to classify segments, while SQ considers the precision of the segmentation. Note that, as reported in [28], SQ and RQ are not independent since SQ is computed only on segments correctly classified. The results on the 100-50 and 100-10 are reported in Tab. 5. The table clearly shows that, as hypothesized in Sec. 3.3, forgetting happens heavily on the classification ability while it is less marked on the segmentation ability of the model. CoMFormer, in fact, without providing regularization losses on the mask prediction, is able to match the segmentation ability of the Joint upper bound on the 100-50 and it gets close results on the 100-10 (-5.6). In addition, it is evident that classification ability is heavily affected by forgetting. CoMFormer has a gap with Joint of 3.2 and 11.4 RQ respectively in the 100-50 and 100-10. We note that, however, other baselines are more affected by forgetting, showcasing even larger gaps.", + "bbox": [ + 496, + 296, + 892, + 599 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 614, + 617, + 631 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we extend the task of continual learning beyond semantic segmentation by introducing the challenging and realistic continual panoptic segmentation setting. We propose CoMFormer, a strategy to address segmentation as a mask classification problem exploiting the properties of transformers to tackle both semantic and panoptic segmentation tasks. The extensive quantitative results highlight that CoMFormer outperforms all previous methods, both in semantic and panoptic segmentation, under a unified paradigm. In future work, we aim to extend the benchmark to instance segmentation to show that CoMFormer can be employed to address any continual segmentation task.", + "bbox": [ + 496, + 642, + 890, + 824 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments.", + "text_level": 1, + "bbox": [ + 500, + 830, + 635, + 843 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We gratefully acknowledge the HPC infrastructure and Team at IIT. Fabio Cermelli acknowledges travel support from ELISE (GA no 951847). This work was partly supported by ANR grant VISA DEEP (ANR-20-CHIA-0022).", + "bbox": [ + 496, + 844, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "3017", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 121, + 173, + 136 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In ECCV, pages 213-229. Springer, 2020. (pages 2, 5).", + "[2] Fabio Cermelli, Dario Fontanel, Antonio Tavera, Marco Ciccone, and Barbara Caputo. Incremental learning in semantic segmentation from image labels. In CVPR, pages 4371-4381, 2022. (page 2).", + "[3] Fabio Cermelli, Antonino Geraci, Dario Fontanel, and Barbara Caputo. Modeling missing annotations for incremental learning in object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3700-3710, 2022. (pages 1, 2).", + "[4] Fabio Cermelli, Massimiliano Mancini, Samuel Rota Bulò, Elisa Ricci, and Barbara Caputo. Modeling the background for incremental learning in semantic segmentation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2020. (pages 1, 2, 3, 4, 5, 6).", + "[5] Fabio Cermelli, Massimiliano Mancini, Yongqin Xian, Zeynep Akata, and Barbara Caputo. Prototype-based incremental few-shot segmentation. In Proceedings of the British Machine Vision Conference (BMVC), 2021. (page 2).", + "[6] Arslan Chaudhry, Marc'Aurelio Ranzato, Marcus Rohrbach, and Mohamed Elhoseiny. Efficient lifelong learning with a-gem. In Proceedings of the International Conference on Learning Representations (ICLR), 2019. (page 2).", + "[7] Hongwei Chen, Laihui Ding, Fengqin Yao, Pengfei Ren, and Shengke Wang. Panoptic segmentation of UAV images with deformable convolution network and mask scoring. In Twelfth International Conference on Graphics and Image Processing (ICGIP 2020), volume 11720, pages 312-321. SPIE, 2021. (page 1).", + "[8] Liang-Chieh Chen, George Papandreou, Iasonas Kokkinos, Kevin Murphy, and Alan L. Yuille. Deeplab: Semantic image segmentation with deep convolutional nets, atrous convolution, and fully connected crfs. In IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 2018. (page 2).", + "[9] Liang-Chieh Chen, George Papandreou, Florian Schroff, and Hartwig Adam. Rethinking atrous convolution for semantic image segmentation. In arXiv preprint library, 2017. (pages 2, 5, 6, 7).", + "[10] Bowen Cheng, Maxwell D Collins, Yukun Zhu, Ting Liu, Thomas S Huang, Hartwig Adam, and Liang-Chieh Chen. Panoptic-deeplab: A simple, strong, and fast baseline for bottom-up panoptic segmentation. In CVPR, pages 12475-12485, 2020. (page 2).", + "[11] Bowen Cheng, Ishan Misra, Alexander G Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. In CVPR, 2022. (pages 2, 3, 5, 6, 7, 8).", + "[12] Bowen Cheng, Alexander G. Schwing, and Alexander Kirillov. Per-pixel classification is not all you need for semantic" + ], + "bbox": [ + 78, + 147, + 470, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "segmentation. In Adv. Neural Inform. Process. Syst., 2021. (pages 2, 3, 6).", + "[13] Lee R. Dice. Measures of the amount of ecologic association between species. Ecology, 1945. (page 5).", + "[14] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In Proceedings of the International Conference on Learning Representations (ICLR), 2021. (page 2).", + "[15] Arthur Douillard, Yifu Chen, Arnaud Dapogny, and Matthieu Cord. Plop: Learning without forgetting for continual semantic segmentation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2021. (pages 1, 2, 4, 5, 6).", + "[16] Arthur Douillard, Yifu Chen, Arnaud Dapogny, and Matthieu Cord. Tackling catastrophic forgetting and background shift in continual semantic segmentation. In arXiv preprint library, 2021. (page 2).", + "[17] Arthur Douillard, Matthieu Cord, Charles Ollion, Thomas Robert, and Eduardo Valle. Podnet: Pooled outputs distillation for small-tasks incremental learning. In Proceedings of the IEEE European Conference on Computer Vision (ECCV), 2020. (pages 1, 2).", + "[18] Arthur Douillard and Timothee Lesort. Continuum: Simple management of complex continual learning scenarios. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR) Workshop, 2021. (page 5).", + "[19] Arthur Douillard, Alexandre Rame, Guillaume Couairon, and Matthieu Cord. Dytox: Transformers for continual learning with dynamic token expansion. In CVPR, 2022. (pages 1, 2, 3).", + "[20] Mark Everingham, S. M. Ali Eslami, Luc Van Gool, Christopher K. I. Williams, John M. Winn, and Andrew Zisserman. The pascal visual object classes challenge: A retrospective. In International Journal of Computer Vision (IJCV), 2015. (page 5).", + "[21] Robert French. Catastrophic forgetting in connectionist networks. Trends in cognitive sciences, 1999. (pages 1, 2, 4).", + "[22] Jun Fu, Jing Liu, Hajjie Tian, Yong Li, Yongjun Bao, Zhiwei Fang, and Hanqing Lu. Dual attention network for scene segmentation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2019. (page 2).", + "[23] K. He, X. Zhang, S. Ren, and J. Sun. Deep residual learning for image recognition. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016. (page 6).", + "[24] Zilong Huang, Xinggang Wang, Yunchao Wei, Lichao Huang, Humphrey Shi, Wenyu Liu, and Thomas S. Huang. Cynet: Criss-cross attention for semantic segmentation. In IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 2020. (page 2).", + "[25] Loshchilov Ilya, Hutter Frank, et al. Decoupled weight decay regularization. Proceedings of ICLR, 2019. (page 6)." + ], + "bbox": [ + 503, + 93, + 890, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "3018", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[26] Jitesh Jain, Anukriti Singh, Nikita Orlov, Zilong Huang, Jiachen Li, Steven Walton, and Humphrey Shi. Semask: Semantically masked transformers for semantic segmentation. arXiv preprint arXiv:2112.12782, 2021. (page 2).", + "[27] Alexander Kirillov, Ross Girshick, Kaiming He, and Piotr Dólar. Panoptic feature pyramid networks. In CVPR, pages 6399-6408, 2019. (page 2).", + "[28] Alexander Kirillov, Kaiming He, Ross Girshick, Carsten Rother, and Piotr Dólár. Panoptic segmentation. In CVPR, pages 9404-9413, 2019. (pages 2, 5, 8).", + "[29] James Kirkpatrick, Razvan Pascanu, Neil Rabinowitz, Joel Veness, Guillaume Desjardins, Andrei A. Rusu, Kieran Milan, John Quan, Tiago Ramalho, Agnieszka Grabska-Barwinska, Demis Hassabis, Claudia Clopath, Dharshan Kumaran, and Raia Hadsell. Overcoming catastrophic forgetting in neural networks. Proceedings of the National Academy of Sciences, 2017. (pages 1, 2).", + "[30] Harold W. Kuhn. The hungarian method for the assignment problem. Naval Research Logistics Quarterly, 1955. (page 5).", + "[31] Duo Li, Guimei Cao, Yunlu Xu, Zhanzhan Cheng, and Yi Niu. Technical report for iccv 2021 challenge sslad-track3b: Transformers are better continual learners. In ICCV 2021 Challenge SSLAD-Track3B, 2021. (page 3).", + "[32] Yanwei Li, Hengshuang Zhao, Xiaojuan Qi, Yukang Chen, Lu Qi, Liwei Wang, Zeming Li, Jian Sun, and Jiaya Jia. Fully convolutional networks for panoptic segmentation with point-based supervision. IEEE TPAMI, 2022. (page 2).", + "[33] Z. Li and D. Hoiem. Learning without forgetting. Proceedings of the IEEE European Conference on Computer Vision (ECCV), 2016. (pages 1, 2).", + "[34] Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dólár. Focal loss for dense object detection. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), 2017. (page 5).", + "[35] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and Lawrence Zitnick. Microsoft coco: Common objects in context. In Proceedings of the IEEE European Conference on Computer Vision (ECCV), 2014. (page 5).", + "[36] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), 2021. (page 2).", + "[37] J. Long, E. Shelhamer, and T. Darrell. Fully convolutional networks for semantic segmentation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2015. (page 2).", + "[38] Andrea Maracani, Umberto Michieli, Marco Toldo, and Pietro Zanuttigh. Recall: Replay-based continual learning in semantic segmentation. In ICCV, pages 7026-7035, 2021. (page 2).", + "[39] Umberto Michieli and Pietro Zanuttigh. Incremental learning techniques for semantic segmentation. In Proceedings of the IEEE International Conference on Computer Vision (ICCV) Workshop, 2019. (pages 1, 2, 4)." + ], + "bbox": [ + 78, + 90, + 468, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[40] Umberto Michieli and Pietro Zanuttigh. Continual semantic segmentation via repulsion-attraction of sparse and disentangled latent representations. In CVPR, pages 1114-1124, 2021. (page 2).", + "[41] Gerhard Neuhold, Tobias Ollmann, Samuel Rota Bulò, and Peter Kontschieder. The mapillary vistas dataset for semantic understanding of street scenes. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), 2017. (page 1).", + "[42] Lorenzo Porzi, Samuel Rota Bulò, Aleksander Colovic, and Peter Kontschieder. Seamless scene segmentation. In CVPR, June 2019. (page 2).", + "[43] Sylvestre-Alvise Rebuffi, Alexander Kolesnikov, Georg Sperl, and Christoph H. Lampert. icarl: Incremental classifier and representation learning. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2017. (pages 1, 2).", + "[44] Anthony Robins. Catastrophic forgetting, rehearsal and pseudorehearsal. Connection Science, 1995. (page 2).", + "[45] Andrei A Rusu, Neil C Rabinowitz, Guillaume Desjardins, Hubert Soyer, James Kirkpatrick, Koray Kavukcuoglu, Razvan Pascanu, and Raia Hadsell. Progressive neural networks. arXiv preprint library, 2016. (pages 1, 2).", + "[46] Robin Strudel, Ricardo Garcia, Ivan Laptev, and Cordelia Schmid. Segmenter: Transformer for semantic segmentation. In ICCV, pages 7262-7272, 2021. (page 2).", + "[47] Sebastian Thrun. Lifelong learning algorithms. In Springer Learning to Learn, 1998. (page 2).", + "[48] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Hervé Jégou. Training data-efficient image transformers and distillation through attention. In International Conference on Machine Learning (ICML), 2021. (page 2).", + "[49] Hugo Touvron, Matthieu Cord, Alexandre Sablayrolles, Gabriel Synnaeve, and Hervé Jégou. Going deeper with image transformers. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), 2021. (page 2).", + "[50] Huiyu Wang, Yukun Zhu, Hartwig Adam, Alan Yuille, and Liang-Chieh Chen. Max-deeplab: End-to-end panoptic segmentation with mask transformers. In CVPR, pages 5463-5474, 2021. (page 2).", + "[51] Huiyu Wang, Yukun Zhu, Bradley Green, Hartwig Adam, Alan Yuille, and Liang-Chieh Chen. Axial-deeplab: Standalone axial-attention for panoptic segmentation. In ECCV, pages 108–126. Springer, 2020. (page 2).", + "[52] Zifeng Wang, Zizhao Zhang, Sayna Ebrahimi, Ruoxi Sun, Han Zhang, Chen-Yu Lee, Xiaoqi Ren, Guolong Su, Vincent Perot, Jennifer Dy, et al. Dualprompt: Complementary prompting for rehearsal-free continual learning. ECCV, 2022. (page 2).", + "[53] Zifeng Wang, Zizhao Zhang, Chen-Yu Lee, Han Zhang, Ruoxi Sun, Xiaoqi Ren, Guolong Su, Vincent Perot, Jennifer Dy, and Tomas Pfister. Learning to prompt for continual learning. In CVPR, pages 139-149, 2022. (page 2).", + "[54] Enze Xie, Wenhai Wang, Zhiding Yu, Anima Anandkumar, Jose M Alvarez, and Ping Luo. Segformer: Simple and efficient design for semantic segmentation with transformers. Adv. Neural Inform. Process. Syst., 34, 2021. (page 2)." + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "3019", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[55] Shipeng Yan, Jiangwei Xie, and Xuming He. Der: Dynamically expandable representation for class incremental learning. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2021. (page 2).", + "[56] Qihang Yu, Huiyu Wang, Siyuan Qiao, Maxwell Collins, Yukun Zhu, Hartwig Adam, Alan Yuille, and Liang-Chieh Chen. k-means mask transformer. In European Conference on Computer Vision, pages 288-307. Springer, 2022. (page 2).", + "[57] Yuhui Yuan, Xilin Chen, and Jingdong Wang. Object-contextual representations for semantic segmentation. In Proceedings of the IEEE European Conference on Computer Vision (ECCV), 2020. (page 2).", + "[58] Yuhui Yuan and Jingdong Wang. Ocnet: Object context network for scene parsing. In arXiv preprint library, 2018. (page 2).", + "[59] Chang-Bin Zhang, Jia-Wen Xiao, Xialei Liu, Ying-Cong Chen, and Ming-Ming Cheng. Representation compensation networks for continual semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7053-7064, 2022. (pages 1, 2, 5, 6, 7).", + "[60] H. Zhao, J. Shi, X. Qi, X. Wang, and J. Jia. Pyramid scene parsing network. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2017. (page 2).", + "[61] Hengshuang Zhao, Yi Zhang, Shu Liu, Jianping Shi, Chen Change Loy, Dahua Lin, and Jiaya Jia. Pisanet: Pointwise spatial attention network for scene parsing. In Proceedings of the IEEE European Conference on Computer Vision (ECCV), 2018. (page 2).", + "[62] Bolei Zhou, Hang Zhao, Xavier Puig, Sanja Fidler, Adela Barriuso, and Antonio Torralba. Scene parsing through ade20k dataset. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2017. (page 5)." + ], + "bbox": [ + 78, + 90, + 470, + 599 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "3020", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2023/CoMFormer_ Continual Learning in Semantic and Panoptic Segmentation/9493268b-7f35-48c3-8aa1-39be491edc53_model.json b/2023/CoMFormer_ Continual Learning in Semantic and Panoptic Segmentation/9493268b-7f35-48c3-8aa1-39be491edc53_model.json new file mode 100644 index 0000000000000000000000000000000000000000..bcb9cf71a914c7208873fc45608dfa053ea3834a --- /dev/null +++ b/2023/CoMFormer_ Continual Learning in Semantic and Panoptic Segmentation/9493268b-7f35-48c3-8aa1-39be491edc53_model.json @@ -0,0 +1,2312 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.808, + 0.045 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.11, + 0.131, + 0.861, + 0.154 + ], + "angle": 0, + "content": "CoMFormer: Continual Learning in Semantic and Panoptic Segmentation" + }, + { + "type": "text", + "bbox": [ + 0.259, + 0.18, + 0.714, + 0.198 + ], + "angle": 0, + "content": "Fabio Cermelli \\(^{1,2*}\\), Matthieu Cord \\(^{3,4}\\), Arthur Douillard \\(^{3\\dagger}\\)" + }, + { + "type": "text", + "bbox": [ + 0.136, + 0.199, + 0.832, + 0.217 + ], + "angle": 0, + "content": "\\(^{1}\\)Politecnico di Torino, \\(^{2}\\)Italian Institute of Technology, \\(^{3}\\) Sorbonne Université, \\(^{4}\\) valeo.ai" + }, + { + "type": "text", + "bbox": [ + 0.1, + 0.22, + 0.86, + 0.233 + ], + "angle": 0, + "content": "fabio.cermelli@polito.it, matthieu.cord@sorbonne-universite.fr, ar.douillard@gmail.com" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.269, + 0.314, + 0.285 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.301, + 0.474, + 0.619 + ], + "angle": 0, + "content": "Continual learning for segmentation has recently seen increasing interest. However, all previous works focus on narrow semantic segmentation and disregard panoptic segmentation, an important task with real-world impacts. In this paper, we present the first continual learning model capable of operating on both semantic and panoptic segmentation. Inspired by recent transformer approaches that consider segmentation as a mask-classification problem, we design CoMFormer. Our method carefully exploits the properties of transformer architectures to learn new classes over time. Specifically, we propose a novel adaptive distillation loss along with a mask-based pseudo-labeling technique to effectively prevent forgetting. To evaluate our approach, we introduce a novel continual panoptic segmentation benchmark on the challenging ADE20K dataset. Our CoMFormer outperforms all the existing baselines by forgetting less old classes but also learning more effectively new classes. In addition, we also report an extensive evaluation in the large-scale continual semantic segmentation scenario showing that CoMFormer also significantly outperforms state-of-the-art methods." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.647, + 0.21, + 0.663 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.672, + 0.47, + 0.822 + ], + "angle": 0, + "content": "Image segmentation is a fundamental computer vision problem that enables machines to assign an image's pixels to discrete segments. Multiple segmentation tasks have been defined depending on the segments definitions. Semantic segmentation clusters pixels by classes, merging in a single segment pixels belonging to instances of the same class. Panoptic segmentation assigns to every pixel a semantic class while separating different instances into different segments. This latter kind of segmentation has real-world impacts in autonomous robots and vehicles [7, 41]." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.824, + 0.47, + 0.854 + ], + "angle": 0, + "content": "Despite tremendous progress in image segmentation, the current approaches are trained on a static dataset with a pre" + }, + { + "type": "image_caption", + "bbox": [ + 0.502, + 0.267, + 0.642, + 0.277 + ], + "angle": 0, + "content": "Step t-1 - New Class: Car" + }, + { + "type": "image", + "bbox": [ + 0.506, + 0.278, + 0.685, + 0.374 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.538, + 0.376, + 0.574, + 0.385 + ], + "angle": 0, + "content": "Image" + }, + { + "type": "image", + "bbox": [ + 0.687, + 0.28, + 0.816, + 0.358 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.699, + 0.375, + 0.794, + 0.385 + ], + "angle": 0, + "content": "CoMFormer (t-1)" + }, + { + "type": "image", + "bbox": [ + 0.822, + 0.279, + 0.892, + 0.375 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.825, + 0.376, + 0.887, + 0.385 + ], + "angle": 0, + "content": "Predictions" + }, + { + "type": "image_caption", + "bbox": [ + 0.503, + 0.401, + 0.65, + 0.411 + ], + "angle": 0, + "content": "Step t - New Class: Person" + }, + { + "type": "image", + "bbox": [ + 0.504, + 0.417, + 0.61, + 0.487 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.538, + 0.49, + 0.574, + 0.5 + ], + "angle": 0, + "content": "Image" + }, + { + "type": "image", + "bbox": [ + 0.614, + 0.411, + 0.816, + 0.484 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.613, + 0.49, + 0.68, + 0.5 + ], + "angle": 0, + "content": "Annotations" + }, + { + "type": "image_caption", + "bbox": [ + 0.709, + 0.49, + 0.788, + 0.501 + ], + "angle": 0, + "content": "CoMFormer (t)" + }, + { + "type": "image", + "bbox": [ + 0.822, + 0.392, + 0.887, + 0.492 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.825, + 0.493, + 0.887, + 0.502 + ], + "angle": 0, + "content": "Predictions" + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.513, + 0.895, + 0.61 + ], + "angle": 0, + "content": "Figure 1. Illustration of our model, CoMFormer, operating in continual segmentation. Relying on the mask classification paradigm, it is able to cope with both continual semantic and panoptic segmentation without any modification by predicting masks for both old (e.g. car in red) and new (e.g. person in green) classes. The figure reports two classes and no \"stuff\" (e.g. road, building) only for illustration purposes." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.627, + 0.892, + 0.748 + ], + "angle": 0, + "content": "defined set of classes. Whenever an update of the model is required to fit new classes, the common solution is to train a model from scratch on the union of the old and new class data. A computationally more efficient solution would be to fine-tune the existing model solely on the new class data. Unfortunately, this approach would cause a catastrophic forgetting [21] of the old classes on which the model performance would be extremely degraded." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.893, + 0.902 + ], + "angle": 0, + "content": "The problem of updating the knowledge of the model over time is typically referred as continual learning. It has been traditionally studied in the context of image classification [17, 19, 29, 33, 43, 45] and is gaining attention on the segmentation task [3, 4, 15, 39, 59] due to the more realistic applications and the additional challenges it introduces, such as the background shift [4]. However, current state-of-the-art methods mainly focus on semantic segmentation and are not designed to work in other segmentation tasks, strongly limiting their application in the real world." + }, + { + "type": "page_footnote", + "bbox": [ + 0.094, + 0.863, + 0.424, + 0.876 + ], + "angle": 0, + "content": "*Work done during the visiting period at Sorbonne Université." + }, + { + "type": "page_footnote", + "bbox": [ + 0.099, + 0.876, + 0.464, + 0.888 + ], + "angle": 0, + "content": "\\(^{\\dagger}\\)Work done at Sorbonne Université, currently affiliated to DeepMind." + }, + { + "type": "page_footnote", + "bbox": [ + 0.099, + 0.888, + 0.379, + 0.9 + ], + "angle": 0, + "content": "" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.863, + 0.464, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3010" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.348 + ], + "angle": 0, + "content": "In this paper, we design the first method operating in both continual semantic and panoptic segmentation, as illustrated in Fig. 1. Our method, CoMFormer (Continual MaskFormer), takes inspiration from recent transformer architectures [11, 12], approaching segmentation as a mask classification problem. Instead of predicting a class probability for each pixel, as in previous semantic segmentation works [9, 37], it predicts a set of binary masks, each associated with a single class prediction, effectively addressing both segmentation tasks without any modification in the training architecture and procedure. Differently from previous works [11, 12], however, CoMFormer forces the output binary masks to be mutually exclusive to one another: a pixel can only be predicted by a single binary mask to prevent having several masks classifying the same pixel with different classes. This behavior is crucial in continual learning to reduce the interference among old and new classes." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.348, + 0.473, + 0.529 + ], + "angle": 0, + "content": "Furthermore, CoMFormer introduces a novel adaptive distillation loss to alleviate forgetting. It enforces consistency of the model's classification predictions across learning steps only when it is useful to remember old classes, ensuring a better tradeoff between rigidity (not forgetting old classes) and plasticity (learning efficiently new classes). Finally, since at each training iteration the dataset reports annotations only for the current classes, we design a mask-based pseudo-labeling technique to generate annotations for the old classes, effectively alleviating forgetting. To reduce the noise, we consider the prediction confidence and we avoid interference with ground-truth annotations." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.53, + 0.473, + 0.636 + ], + "angle": 0, + "content": "We validate CoMFormer on both continual segmentation tasks. For panoptic segmentation, we define a new benchmark relying on the challenging ADE20K where we demonstrate that CoMFormer largely outperforms all previous baselines. On semantic segmentation, we show that CoMFormer outperforms the existing state-of-the-art methods on every setting of the large-scale ADE20K benchmark." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.637, + 0.468, + 0.652 + ], + "angle": 0, + "content": "To sum up, the contributions of this paper are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.653, + 0.468, + 0.697 + ], + "angle": 0, + "content": "- We introduce continual panoptic segmentation which has real-world impacts in addition to being significantly more challenging than previous benchmarks." + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.698, + 0.468, + 0.758 + ], + "angle": 0, + "content": "- We propose CoMFormer to tackle both continual panoptic and semantic segmentation. To avoid forgetting, we design a novel adaptive distillation and an efficient mask-based pseudo-labeling strategy." + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.759, + 0.468, + 0.803 + ], + "angle": 0, + "content": "- Through extensive quantitative and qualitative benchmarks, we showcase the state-of-the-art performance of our model on both continual segmentation tasks." + }, + { + "type": "list", + "bbox": [ + 0.096, + 0.653, + 0.468, + 0.803 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.812, + 0.227, + 0.829 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.472, + 0.903 + ], + "angle": 0, + "content": "Semantic and Panoptic Segmentation. The two tasks have been traditionally treated separately, with specialized architectures proposed for either one or the other task, without interoperability. Semantic segmentation has been tradi" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.895, + 0.471 + ], + "angle": 0, + "content": "tionally addressed as a per-pixel classification task. Fully convolutional network [37] dominated the field by aggregating long-range dependencies in the features map [8, 9, 60] and exploiting contextual information [22, 24, 57, 58, 61]. Recently, transformers [26, 46, 54] are replacing convolutions by integrating long-range dependencies at every layer. Panoptic segmentation [28] has been proposed to unify semantic and instance segmentations. Initially, methods proposed to combine each task-specific architectures [10, 27, 32, 42] or defined new specialized architectures and objective functions for the panoptic task [1, 50, 51], drifting from a general solution for all the segmentation tasks. Recently, methods addressing segmentation as a mask classification problem [11, 12, 50, 56] have been proposed, introducing a transformer architecture able to solve multiple tasks at once. MaskFormer [12] was the first to propose a single architecture to address both panoptic and semantic segmentation. Mask2Former [11] improves it by adopting multi-scale features, masked attention and optimization tricks. Concurrently, kMaX-DeepLab [56] extended [50] proposing to reformulate the cross-attention as a clustering process. Despite their effectiveness in standard training setting, these methods suffer from catastrophic forgetting. In this work, we aim to extend their capability and we propose CoMFormer to learn to segment new classes over time." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.481, + 0.895, + 0.8 + ], + "angle": 0, + "content": "Continual Segmentation. Continual learning is a longstanding field addressing the problem of learning new knowledge over time while avoiding catastrophic forgetting [21, 44, 47]. Traditionally, it has been studied in the context of image classification [6, 17, 19, 29, 33, 43, 45, 55] but there is a growing interest in its applications in segmentation [2-5, 15, 16, 38-40, 59]. Continual semantic segmentation introduces additional challenges, as pointed out by [4]. In particular, considering the task as a per-pixel classification problem, catastrophic forgetting is exacerbated by the background shift, where old classes pixels are treated as background in following training steps. [4] proposed a revision of the standard knowledge distillation framework to deal with it. [17, 39] introduced a distillation loss preserving the representation at feature level while [38] proposed an approach using old classes' examples to alleviate forgetting. Recently, [59] presented a technique that decouples learning the representation of the new classes while freezing the representation for the old ones. Differently from previous work, we extend the benchmarks to solve both semantic and the more challenging panoptic segmentation task." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.81, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Continual Learning with Transformers. Recent progresses made using transformers for computer vision [14, 36, 48, 49] attracted the attention of the continual learning community [19, 52, 53]. In particular, DyTox [19] proposed to specialize the architecture on each task using a different task-specific token. Learning-to-Prompt [53] stores a" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.958 + ], + "angle": 0, + "content": "3011" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.107, + 0.089, + 0.868, + 0.293 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.302, + 0.895, + 0.375 + ], + "angle": 0, + "content": "Figure 2. Overview of CoMFormer. At learning step t, both the current (bottom) and old (top) models output, for each input query, a pair made of a class probability distribution and a binary mask. While learning the new class (person), we regularize the current model to prevent forgetting the old classes (car) by: (i) generating pseudo-labels for the old classes present in the image and (ii) by applying an adaptive distillation loss that reweights the contribution of each output based on the information it carries on the old classes. \\(\\psi\\) indicates the dot product between the mask embeddings \\(\\mathcal{E}_{mask}\\) and the pixel embeddings \\(\\mathcal{E}_{pixel}\\) followed by the softmax activation." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.386, + 0.471, + 0.462 + ], + "angle": 0, + "content": "pool of prompts that are employed to condition the whole forward execution of the patch tokens. To the best of our knowledge, transformers have been evaluated only in continual learning for image classification [19, 31] and we are the first to use them in the continual segmentation task." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.473, + 0.21, + 0.489 + ], + "angle": 0, + "content": "3. CoMFormer" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.498, + 0.26, + 0.513 + ], + "angle": 0, + "content": "3.1. Problem Definition" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.522, + 0.47, + 0.672 + ], + "angle": 0, + "content": "The goal of image segmentation is to learn a model able to (i) partition the image into a set of N regions represented by binary masks and (ii) produce a class-probability distribution associated with each region. The differences among segmentation tasks rely on the semantics of the masks: semantic segmentation groups all the pixels of a class while panoptic segmentation distinguishes different object instances. In the following, we provide a general formulation for the two tasks since they only differ in the construction of the binary masks in the dataset." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.673, + 0.471, + 0.84 + ], + "angle": 0, + "content": "Continual segmentation aims to train the model in multiple learning steps \\( t = 1, \\ldots, T \\), introducing at every step a new set of classes. Formally, during the learning step \\( t \\) a dataset \\( D_{t} \\) consisting of a collection of image and label pairs is provided. The label of each image takes the form of a set of \\( N^{gt} \\) ground truth segments \\( z^{gt}\\{(c_i^{gt}, m_i^{gt}) | c_i^{gt} \\in \\mathcal{K}^t, m_i^{gt} \\in \\{0,1\\}^{H \\times W}\\}_{i=1}^{N^{gt}} \\) where \\( c_i^{gt} \\) is the ground-truth class and \\( m_i^{gt} \\) is binary mask \\( m^{gt} \\). \\( \\mathcal{K}^t \\) is the set of classes introduced at step \\( t \\), and \\( H, W \\) are the height and width of the images. The goal of training step \\( t \\) is to learn a model able to predict segments for all the seen classes \\( \\mathcal{K}^{1:t} \\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.471, + 0.902 + ], + "angle": 0, + "content": "We note that the dataset only contains annotations for the new classes \\(\\mathcal{K}^t\\) while not reporting ground-truth segments for old \\(\\mathcal{K}^{1:t - 1}\\) or future \\(\\mathcal{K}^{t + 1:T}\\) classes. Moreover, differently from previous works [4], old and future class segment" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.386, + 0.892, + 0.418 + ], + "angle": 0, + "content": "annotations are simply absent during training rather than being collapsed into an artificial background class." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.431, + 0.738, + 0.446 + ], + "angle": 0, + "content": "3.2. CoMFormer Architecture" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.456, + 0.892, + 0.639 + ], + "angle": 0, + "content": "To solve semantic and panoptic segmentation tasks within a single method, we take inspiration from Mask-Former architectures [11, 12] considering segmentation a mask classification problem. It consists in predicting for each image a set of \\(N\\) pairs made of a class prediction and a binary mask. Formally, the output of a CoMFormer is \\(z = \\{(p_i, m_i)\\}_{i=1}^N\\), where \\(p_i\\) represents a class probability distribution over the seen classes, i.e. \\(p_i \\in \\Delta^{K^t + 1}\\) with \\(K^t = |\\mathcal{K}^{1:t}|\\), and \\(m_i\\) is a binary mask such that \\(m_i \\in [0, 1]^{H \\times W}\\). We note that \\(p_i\\) contains an additional \"no object\" class (denoted \\(\\varnothing\\)) to indicate that the mask does not correspond to any of the known \\(\\mathcal{K}^{1:t}\\) categories." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.641, + 0.893, + 0.87 + ], + "angle": 0, + "content": "To obtain such output, the CoMFormer architecture, showcased in Fig. 2, is made of three components: (i) a backbone that extracts feature embeddings \\(\\mathcal{E}_{\\text {feat }}\\), (ii) a transformer decoder that takes as input N learnable queries and \\(\\mathcal{E}_{\\text {feat }}\\) to output \\(N\\) per-segment embeddings \\(Q\\), (iii) a pixel decoder that takes \\(\\mathcal{E}_{\\text {feat }}\\) as input and extracts per-pixel embeddings \\(\\mathcal{E}_{\\text {pixel }}\\). The output class probabilities \\(\\{p_i \\in \\Delta^{K^t + 1}\\}_{i=1}^N\\) are obtained by applying a linear classifier on \\(Q\\), which is enlarged as new classes are presented. We obtain the mask predictions \\(m_i\\) combining \\(Q\\) and \\(\\mathcal{E}_{\\text {pixel }}\\): first \\(Q\\) is passed into a 2-layer MLP to obtain N mask embeddings \\(\\mathcal{E}_{\\text {mask }}\\) and we then compute the dot product between the \\(i\\)-th mask embedding and the pixel embeddings, followed by the softmax activation: \\(m_i[h, w] = \\operatorname{softmax}(\\mathcal{E}_{\\text {mask}}[:, i]^T \\cdot \\mathcal{E}_{\\text {pixel}}[:, h, w])\\)." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.871, + 0.892, + 0.902 + ], + "angle": 0, + "content": "We point out that employing the softmax contrasts sharply with MaskFormers [11, 12], which use sigmoid ac" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3012" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.08, + 0.087, + 0.445, + 0.259 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.269, + 0.47, + 0.325 + ], + "angle": 0, + "content": "Figure 3. Illustration of the mask predictions between sigmoid and softmax. MaskFormer has overlap across masks (e.g. in Query 1 and 2) predicted either as old or new classes, while CoMFormer avoids the issue using the softmax activation." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.332, + 0.471, + 0.453 + ], + "angle": 0, + "content": "tivation to get the mask predictions. Using the sigmoid activation would result in the overlap of many masks on a single pixel, each potentially belonging to a distinct class, as seen in Fig. 3. This is particularly important in the context of continual learning, when the model may predict two similar masks associated with two distinct classes (e.g. cat and dog classes in Fig. 3) thereby generating interference among old and new classes and degrading the model's performance." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.461, + 0.334, + 0.476 + ], + "angle": 0, + "content": "3.3. Learning without Forgetting" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.484, + 0.47, + 0.575 + ], + "angle": 0, + "content": "To learn new classes, a simple solution is to fine-tune the model on the new dataset. However, because it lacks annotations for old classes, this operation results in catastrophic forgetting [21]. To alleviate the issue, we introduce two techniques for avoiding forgetting: an adaptive distillation loss and a mask-based pseudo-labeling strategy." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.581, + 0.47, + 0.763 + ], + "angle": 0, + "content": "Adaptive Distillation Loss. A common strategy to mitigate forgetting adopted by continual semantic segmentation approaches [4, 15, 39] is to distill the knowledge coming from the old model in the new one, either in form of features [15, 39] or as classification probabilities [4]. We take inspiration from the latter solution and we design a novel distillation loss tailored to the CoMFormer architecture. In particular, analyzing the two components constituting the output space, we find the mask prediction being robust to forgetting while the classification part is heavily affected (see Sec. 4.6). For this reason, we focus on the output class probabilities to design our distillation loss." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.763, + 0.47, + 0.853 + ], + "angle": 0, + "content": "An adaptation of standard knowledge distillation works [4] would force the new model to mimic the old model probabilities over all the N outputs, independently from their content. Formally, given the set of N probability distributions \\(\\{p_i^o\\}_{i = 1}^N\\) coming from the old model, a standard distillation loss is computed as" + }, + { + "type": "equation", + "bbox": [ + 0.132, + 0.861, + 0.469, + 0.904 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {K D} = - \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\sum_ {k = 1} ^ {K ^ {t - 1} + 1} p _ {i} ^ {o} (k) \\log \\frac {\\tilde {p} _ {i} (k)}{p _ {i} ^ {o} (k)}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.091, + 0.892, + 0.123 + ], + "angle": 0, + "content": "where \\( K^{t - 1} + 1 \\) is the \\( \\varnothing \\) class and \\( \\tilde{p}_i(k) \\) represents the unbiased probability distribution of \\( i \\)-th output for class \\( k \\):" + }, + { + "type": "equation", + "bbox": [ + 0.542, + 0.134, + 0.892, + 0.174 + ], + "angle": 0, + "content": "\\[\n\\tilde {p} _ {i} (k) = \\left\\{ \\begin{array}{l l} \\sum_ {j \\in \\mathcal {K} ^ {t}} p _ {i} (j) + p _ {i} (\\varnothing) & \\text {i f} k = \\varnothing , \\\\ p _ {i} (k) & \\text {o t h e r w i s e .} \\end{array} \\right. \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.185, + 0.892, + 0.231 + ], + "angle": 0, + "content": "Note that, similarly to [4], it forces the sum of the model new class \\((\\mathcal{K}^t)\\) and the \"no object\" class \\((\\varnothing)\\) probabilities to be similar to the old model \\(\\varnothing\\) probability." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.232, + 0.892, + 0.337 + ], + "angle": 0, + "content": "However, not every output probability brings relevant information: most of them are predicted as \"no object\" \\((\\varnothing)\\) and thus they do not carry enough details to remember the old classes while reducing the relative importance of the other outputs probabilities. For this reason, we re-weight their contribution based on the probability of not being \\(\\varnothing\\). Formally, we define our adaptive distillation loss as" + }, + { + "type": "equation", + "bbox": [ + 0.511, + 0.35, + 0.891, + 0.406 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {A D} (p, p ^ {o}) = - \\sum_ {i = 1} ^ {N} \\frac {\\omega_ {i}}{\\sum_ {i = 1} ^ {N} \\omega_ {i}} \\sum_ {k = 1} ^ {K ^ {t - 1} + 1} p _ {i} ^ {o} (k) \\log \\frac {\\tilde {p} _ {i} (k)}{p _ {i} ^ {o} (k)}, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.406, + 0.892, + 0.482 + ], + "angle": 0, + "content": "where the weighting coefficient is \\(\\omega_{i} = (1 - p_{i}^{o}(\\varnothing))^{2}\\). Note that, with this formulation, we effectively reduce the contributions of outputs that bring small information about the old classes \\((p_i^o (\\varnothing)\\approx 1)\\), while we increase the contribution of important ones (where \\(p_i^o (\\varnothing)\\approx 0\\))." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.489, + 0.892, + 0.595 + ], + "angle": 0, + "content": "Mask-based Pseudo-labeling Strategy. Forgetting is effectively reduced by the adaptive distillation loss but we are not, however, taking advantage of the presence of the old classes that are unlabeled in the dataset. We propose using the old model predictions to recognize old class segments and generate pseudo-labels to make efficient use of this information and improve the knowledge of past classes." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.595, + 0.892, + 0.701 + ], + "angle": 0, + "content": "A simple strategy consists in considering the N output pairs coming from the old model independently and using all the pairs for which the predicted class is not \\(\\varnothing\\) as pseudolabels. This strategy, however, does not consider two aspects: there may be pairs where the class is different from \\(\\varnothing\\) but (i) the mask substantially overlaps with a ground-truth segment, or (ii) the mask is noisy and has low confidence." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.701, + 0.892, + 0.901 + ], + "angle": 0, + "content": "To overcome these issues, we propose a mask-based pseudo-labeling strategy that jointly considers the mask and class prediction confidences to avoid noisy labels and overlaps with the existing annotations. We define the model confidence as the multiplication among the class and mask probabilities. Formally, we compute the confidence for the \\(i\\)-th output as \\(q_{i} = p_{i}^{max} \\cdot m_{i}^{o} \\in [0,1]^{H \\times W}\\), where \\(p_{i}^{max} = \\max_{k=1}^{K^{t-1}} p_{i}^{o}(k)\\) and \\(m_{i}^{o}\\) is the binary mask predicted by the old model. Note that we do not consider the \\(\\varnothing\\) class in the max operation. We denote the pseudo-class \\(c_{i}^{ps} = \\operatorname{argmax}_{k=1}^{K^{t-1}} p_{i}^{o}(k)\\) and we generate the pseudomask \\(m_{i}^{ps}\\) considering two criteria: (i) there should be no overlap between the pseudo-mask and the ground-truth seg" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3013" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.47, + 0.173 + ], + "angle": 0, + "content": "ments and (ii) to be included in a pseudo-mask, the confidence on a pixel should be maximum over all outputs. Formally, denoting the binarization of the predicted mask as \\( m_{i}^{bin} = m_{i} > 0.5 \\) and the union of the ground-truth segments \\( m_{all}^{gt} = \\sum_{i=1}^{N^{ps}} m_{i}^{gt} \\), we generate the pseudo-mask as" + }, + { + "type": "equation", + "bbox": [ + 0.093, + 0.189, + 0.47, + 0.242 + ], + "angle": 0, + "content": "\\[\nm _ {i} ^ {p s} [ h, w ] = \\left\\{ \\begin{array}{l l} m _ {i} ^ {b i n} [ h, w ] & \\text {i f} i = \\operatorname {a r g m a x} _ {j = 1} ^ {N} q _ {j} [ h, w ] \\\\ & \\text {a n d} m _ {a l l} ^ {g t} [ h, w ] = 0, \\\\ 0 & \\text {o t h e r w i s e .} \\end{array} \\right. \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.248, + 0.469, + 0.368 + ], + "angle": 0, + "content": "Indeed, not all pseudo-mask should be included in the set of pseudo-labels. There may be pseudo-masks where no pixel is active, i.e. it is all zeros, or where the pseudomask contains only a small fraction of the original mask. For these reasons, we construct the set of pseudo-labels \\( z^{ps} = \\{(c_i^{ps},m_i^{ps})\\}_{i = 1}^{N^{ps}} \\) such that we include in \\( z^{ps} \\) only the pseudo-masks that retain at least half of the pixels w.r.t. the binary mask \\( m_{i}^{bin} \\) and that have at least one active pixel." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.369, + 0.469, + 0.416 + ], + "angle": 0, + "content": "We denote the final annotation set \\(\\bar{z} = \\{\\bar{c}_i, \\bar{m}_i\\}_{i=1}^N\\) as the union of the ground-truth labels \\(z^{gt}\\) and the pseudo-labels \\(z^{ps}\\), i.e. \\(\\bar{z} = z^{gt} \\cup z^{ps}\\) and \\(\\bar{N} = N^{gt} + N^{ps}\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.422, + 0.469, + 0.572 + ], + "angle": 0, + "content": "Overall training loss. To train CoMFormer, a one-to-one matching \\(\\sigma\\) between the model predictions \\(z\\) and the annotation set \\(\\bar{z}\\) is required. Following standard practices [1, 11], we employ the Hungarian matching algorithm [30] and we search for the matching \\(\\sigma\\) that minimize the assignment cost computed as \\(-p_i(\\bar{c}_j)\\cdot Dice(m_i,\\bar{m}_j)\\), where Dice indicates the Dice coefficient [13]. We assume the size of the prediction set to be larger than the annotations set (i.e. \\(N > \\bar{N}\\)) and to obtain a one-to-one matching we pad the annotations with \"no object\" tokens \\(\\varnothing\\)." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.573, + 0.469, + 0.603 + ], + "angle": 0, + "content": "Once the best matching \\(\\sigma\\) has been found, we define the overall training loss as:" + }, + { + "type": "equation", + "bbox": [ + 0.158, + 0.613, + 0.469, + 0.63 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\mathcal {L} _ {\\text {S e g}} (z, \\bar {z}) + \\lambda_ {D} \\mathcal {L} _ {A D} (p, p ^ {o}), \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.639, + 0.375, + 0.654 + ], + "angle": 0, + "content": "where \\(\\lambda_{D}\\) is a trade-off hyper-parameter and" + }, + { + "type": "equation", + "bbox": [ + 0.085, + 0.663, + 0.469, + 0.728 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} _ {\\text {S e g}} (z, \\bar {z}) = \\sum_ {j = 1} ^ {N} [ - \\alpha (1 - p _ {\\sigma (j)} (\\bar {c} _ {j})) ^ {\\gamma} \\log p _ {\\sigma (j)} (\\bar {c} _ {j}) + \\tag {6} \\\\ \\left. + \\lambda_ {m a s k} \\mathbb {1} _ {\\bar {c} _ {j} \\neq \\emptyset} \\mathcal {L} _ {m a s k} \\left(m _ {\\sigma (j)}, \\bar {m} _ {j}\\right) \\right]. \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.734, + 0.469, + 0.811 + ], + "angle": 0, + "content": "We note that the first term in \\(\\mathcal{L}_{Seg}\\) indicates the focal loss [34], where \\(\\alpha\\) and \\(\\gamma\\) are hyper-parameters. Finally, \\(\\mathcal{L}_{mask}\\) is the sum of dice and cross-entropy losses, and \\(\\lambda_{mask}\\) is a hyper-parameter. We note that \\(\\mathcal{L}_{mask}\\) is computed only for valid segments, i.e. where \\(\\bar{c}_j \\neq \\emptyset\\)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.822, + 0.209, + 0.839 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.848, + 0.27, + 0.864 + ], + "angle": 0, + "content": "4.1. Dataset and Settings" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.469, + 0.901 + ], + "angle": 0, + "content": "We start from the widely adopted continual semantic segmentation benchmark defined in [4] and we extend it to" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.228 + ], + "angle": 0, + "content": "the newly proposed continual panoptic segmentation using a modified Continuum library [18]. We compare our novel model on the large-scale ADE20K [62] dataset since it supports both semantic and panoptic segmentation tasks. This challenging dataset contains 150 classes, divided into 100 \"things\" and 50 \"stuff\" categories. This dataset represents a wide variety of scenes, both interior and exterior, with images featuring an average of 9.9 classes, while other datasets like COCO [35] only have an average of 3.5 classes." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.236, + 0.892, + 0.523 + ], + "angle": 0, + "content": "Continual Learning Protocols. Previous continual semantic segmentation works [4, 15] describe a training protocol to assess the performance on multiple continual learning steps. In particular, [4] proposed three protocols for ADE20K with different numbers of tasks: (i) the 100-50 consists of two tasks, the first of 100 and the second of 50 classes; (ii) the 50-50 consists of three tasks of 50 classes; (iii) the 100-10 consists on 6 tasks, the first on 100 classes followed by 5 tasks of 10 classes. In addition, we use the 100-5 introduced in [15], consisting of 11 tasks, the first of 100 classes followed by 10 tasks of 5 classes. To divide the images into different tasks, we follow the splits provided by [4] for 100-50, 100-10, 50-50 for both semantic and panoptic segmentation. In particular, they ensured that each image appears only on a unique task. Differently, for the 100-5 we use the split proposed by [15], where the same image may appear on multiple tasks. We report the results of the 50-50 setting for both semantic and panoptic segmentation in the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.531, + 0.892, + 0.699 + ], + "angle": 0, + "content": "Metrics. We compare the different models using the mean Intersection over Union (mIoU) on semantic segmentation [20] and Panoptic Quality (PQ) on panoptic segmentation [28]. PQ is the product of two components: Segmentation Quality (SQ) which considers the IoU between the correctly classified segments, and Recognition Quality (RQ) which only considers the classification accuracy. For both tasks, we report the metric after the last step \\( T \\) on the first classes \\( (C^1) \\), for the added classes \\( (C^{2:T}) \\), and for all the classes (all). We also report the average of the final performance on all seen classes after each step (avg) following [15]." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.71, + 0.608, + 0.724 + ], + "angle": 0, + "content": "4.2. Baselines" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.735, + 0.892, + 0.901 + ], + "angle": 0, + "content": "We benchmark our model against the state-of-the-art methods: MiB [4], PLOP [15], and RCIL [59]. In semantic segmentation, we report them with the original DeepLab-v3 segmentation architecture [9]. In addition, for a fair comparison, we report MiB and PLOP using the CoMFormer architecture. In particular, we implement the methods in two versions: (i) we replace the CoMFormer losses with a per-pixel cross-entropy loss on the dot product between mask and class logits and we apply the continual learning methods on that baseline (we refer to this as Per-Pixel), and (ii) by applying the distillation losses directly on the" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3014" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.114, + 0.088, + 0.858, + 0.19 + ], + "angle": 0, + "content": "
Method100-50 (2 tasks)100-10 (6 tasks)100-5 (11 tasks)
1-100101-150allavg1-100101-150allavg1-100101-150allavg
FT0.025.88.624.80.02.91.07.90.01.30.44.6
MiB35.119.329.835.427.110.021.429.124.06.518.125.6
PLOP41.026.636.238.630.517.526.132.928.115.724.030.5
CoMFormer41.127.736.738.836.017.129.735.334.415.928.234.0
Joint43.232.139.5—43.232.139.5—43.232.139.5—
" + }, + { + "type": "table_caption", + "bbox": [ + 0.131, + 0.201, + 0.836, + 0.216 + ], + "angle": 0, + "content": "Table 1. Continual Panoptic Segmentation results on ADE20K dataset in PQ considering multiple training protocols." + }, + { + "type": "table", + "bbox": [ + 0.104, + 0.229, + 0.869, + 0.401 + ], + "angle": 0, + "content": "
BackboneMethod100-50 (2 tasks)100-10 (6 tasks)100-5 (11 tasks)
1-100101-150allavg1-100101-150allavg1-100101-150allavg
DeepLab-v3 [9]MiB [4]40.517.232.837.338.311.329.235.136.05.726.032.7
PLOP [15]41.914.932.937.440.514.131.636.639.17.828.835.3
RCIL [59]42.318.834.5—39.317.632.1—38.511.529.6—
Per-PixelMiB40.324.034.837.535.114.028.134.033.315.227.333.1
PLOP40.220.233.536.932.613.726.332.433.39.425.432.8
Mask-basedFT0.026.78.926.40.02.30.88.50.01.10.34.2
MiB37.024.132.638.323.510.626.629.621.06.116.127.7
PLOP44.226.238.241.134.815.928.535.233.614.127.133.6
CoMFormer44.726.238.441.240.615.632.337.439.513.630.936.5
Joint46.935.643.1—46.935.643.1—46.935.643.1—
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.412, + 0.895, + 0.441 + ], + "angle": 0, + "content": "Table 2. Continual Semantic Segmentation results on the ADE20K dataset in mIoU considering multiple training protocols. We report the numbers published by the original papers on DeepLab-v3 [9]. The remaining results are based on our re-implementation (see Sec. 4.2)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.455, + 0.473, + 0.668 + ], + "angle": 0, + "content": "classification head of the CoMFormer architecture (Mask-based). We note that the first version cannot be applied to panoptic segmentation, since the task does not support the use of per-pixel losses. For PLOP, we apply its distillation loss local-POD on the intermediate features of the segmentation backbone, as in the original paper, and we use the pseudo-labeling strategy proposed in Sec. 3. RCIL was not re-implemented since it is not possible to extend their approach to non-convolutional architectures like ours. We include a naive finetuning (FT) without any capabilities for continual learning. Finally, we also report an upper bound (Joint) that is trained in the traditional segmentation setting. We made our best effort to find the most suitable hyperparameters for all the baselines." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.679, + 0.295, + 0.695 + ], + "angle": 0, + "content": "4.3. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.697, + 0.47, + 0.834 + ], + "angle": 0, + "content": "Architecture. Following the previous benchmarks [4, 15] we use as backbone a ResNet101 [23] on semantic segmentation and we propose to use a ResNet50 [23] for panoptic segmentation. For both tasks, we use the pixel decoder and the transformer decoder proposed in [11] with 9 layers in total and \\( N = 100 \\) queries. In addition, we adopt all the improvements introduced in [11]: masked attention, multiscale features, and their optimization improvements. We report the performance of single-scale inference." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Training parameters. We follow the Mask2Former [11] hyper-parameters and we use AdamW [25] optimizer with an initial learning rate of 0.0001 for the first step (\\(t = 0\\)) and 0.00005 in the following (\\(t > 0\\)). A learning rate multiplier" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.455, + 0.895, + 0.683 + ], + "angle": 0, + "content": "of 0.1 is applied to the backbone and we follow a polynomial learning rate schedule. We use weight decay of 0.05. We train the model for 160K iterations in the first step and 400 iterations per class in the following (e.g. learning 50 classes we train for \\(50 \\times 400 = 20000\\) iterations). For semantic segmentation, we use a crop size of 512 and a batch size of 16. For panoptic segmentation, we use a crop size of 640 and a batch size of 8. For both, we use the standard random scale jittering between 0.5 and 2.0, random horizontal flipping, random cropping, as well as random color jittering as data augmentation [12]. We set \\(\\alpha = 20\\), \\(\\gamma = 2\\), and \\(\\lambda_{mask} = 5.0\\). \\(\\lambda_{D}\\) is set to 1 for 100-50 and to 10 in the 100-10 and 100-5. Finally, following the standard protocol of continual learning [4, 15] we do not store any image of previous steps and we do not use rehearsal learning." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.695, + 0.796, + 0.712 + ], + "angle": 0, + "content": "4.4. Continual Panoptic Segmentation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.895, + 0.901 + ], + "angle": 0, + "content": "Tab. 1 reports the results on the new continual panoptic segmentation benchmark. CoMFormer exceeds all the baselines by a significant margin. Considering the 100-50, we can see that CoMFormer outperforms MiB on both old (+6 PQ) and new (+8.4 PQ) classes and PLOP on new classes (+1.1 PQ). Considering a longer sequence of tasks (100-10 and 100-5), we can see that the gap between CoMFormer and the other methods becomes more significant, especially in the old classes. In 100-10, it surpasses MiB and PLOP respectively by 8.9 and 5.5 PQ in old classes, while in the new classes it exceeds MiB by 7.1 and is comparable to PLOP. In the 100-5, it outperforms MiB and PLOP" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3015" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.101, + 0.089, + 0.875, + 0.495 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.088, + 0.506, + 0.88, + 0.521 + ], + "angle": 0, + "content": "Figure 4. Qualitative results of CoMformer v.s. MiB and PLOP on the 100-10 continual panoptic segmentation setting on ADE20K." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.535, + 0.47, + 0.64 + ], + "angle": 0, + "content": "on both old (+10.4 and +6.3 PQ) and new classes (+9.4 and +0.2 PQ). Finally, we note that CoMFormer obtains performance close to the Joint baseline in the 100-50 setting, suffering a drop of only 2.1 PQ on the old classes and 4.4 PQ on the new ones. In the other settings, the gap is still relevant, indicating that learning from multiple training steps is still a very challenging task." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.641, + 0.471, + 0.792 + ], + "angle": 0, + "content": "In Fig. 4, we report a qualitative comparison with the baselines in the 100-10 setting using randomly chosen images. We can see that our method is able to maintain better the previous knowledge by precisely segmenting all the classes in the image; for example, w.r.t. to PLOP, our method correctly segments the chair in the second row and the boat in the third row. We can note a common error: models correctly segment but misclassify objects in the image, like the plane classified as car in the first row or the press classified as a door in the third row." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.802, + 0.377, + 0.818 + ], + "angle": 0, + "content": "4.5. Continual Semantic Segmentation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.826, + 0.47, + 0.903 + ], + "angle": 0, + "content": "We report the comparison with state-of-the-art methods in continual semantic segmentation in Tab. 2. CoMFormer achieves the best results on every setting, surpassing both methods based on DeepLab-v3 [9] and baselines implemented upon the same architecture. In particular, CoM-" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.534, + 0.892, + 0.837 + ], + "angle": 0, + "content": "Former outperforms the best state-of-the-art method (RCIL [59] by \\(3.9\\%\\) on the 100-50, by \\(0.2\\%\\) on 100-10, and by \\(1.3\\%\\) on the 100-5. Moreover, it always improves the performance on the old classes and the avg across different steps. In addition, CoMFormer exceeds all the baselines. Comparing it with PLOP (Mask-based), we can see that it obtains a substantial improvement, especially considering scenarios with longer training protocols: CoMformers achieves \\(30.9\\%\\) \\((32.3\\%)\\) on the all metrics, while PLOP only \\(27.1\\%\\) \\((28.5\\%)\\) on the 100-5 \\((100 - 10)\\) setting. Furthermore, the results confirm that improvements are not led by the newer architecture. Per-Pixel baselines achieve only similar or even lower performance than their counterparts. MiB (Per-Pixel), for example, slightly improves the one based on DeepLab-v3 on the 100-50 \\((+2.0\\%\\) on all) and in the 100-5 \\((+1.3\\%)\\) but its performance is slightly worse in the 100-10 \\((-1.1\\%)\\). In addition, Per-Pixel baselines demonstrate a large performance drop in the old classes, showing that, without employing a distillation strategy tailored to the architecture, it is difficult to maintain satisfactory results." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.851, + 0.665, + 0.866 + ], + "angle": 0, + "content": "4.6. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.871, + 0.893, + 0.902 + ], + "angle": 0, + "content": "Mask2Former vs CoMFormer. In Tab. 3, we compare the proposed CoMFormer against Mask2Former [11]. In" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3016" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.087, + 0.089, + 0.462, + 0.248 + ], + "angle": 0, + "content": "
MethodSemantic - mIoUPanoptic - PQ
1-100101-150all1-100101-150all
100-50 (2 tasks)
Mask2Former44.326.038.137.924.533.4
CoMFormer44.726.238.441.127.736.7
100-10 (6 tasks)
Mask2Former38.715.731.030.210.823.7
CoMFormer40.615.632.336.017.129.7
100-5 (11 tasks)
Mask2Former32.613.726.325.95.619.1
CoMFormer39.513.630.934.415.928.2
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.259, + 0.47, + 0.316 + ], + "angle": 0, + "content": "Table 3. Comparison between CoMFormer and Mask2Former [11] in continual semantic and panoptic segmentation. We apply to Mask2Former our mask-based pseudo-labeling strategy and the adaptive distillation loss for a fair comparison." + }, + { + "type": "table", + "bbox": [ + 0.088, + 0.329, + 0.462, + 0.432 + ], + "angle": 0, + "content": "
100-50 (2 tasks)100-10 (6 tasks)
PSKDAD1-100101-150all1-100101-150all
---0.025.88.60.02.91.0
--✓28.415.324.023.09.118.4
✓--40.726.335.930.117.225.8
✓✓-40.528.136.430.217.626.0
✓-✓41.127.736.736.017.129.7
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.442, + 0.47, + 0.498 + ], + "angle": 0, + "content": "Table 4. Ablation of the different key components of our CoM-Former architecture. We report the PQ on the 100-50 and 100-10 panoptic settings. PS: Pseudo-labeling; AD: adaptive distillation Eq. (3); KD: standard distillation Eq. (1)" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.515, + 0.47, + 0.683 + ], + "angle": 0, + "content": "particular, to fairly compare the two methods, we add our adaptive distillation loss and the pseudo-labeling strategy to Mask2Former. Considering semantic segmentation, CoMFormer achieves better results on the old classes (on average \\(3\\%\\) mIoU), while the two methods obtain comparable results on the new classes. On panoptic segmentation, CoMFormer exceeds Mask2Former in all the classes with a considerable gap (6 PQ on both old and new classes on average). We conclude that a vanilla MaskFormer architecture is not suited to continual learning because it suffers from forgetting, while CoMFormer outperforms it in any setting." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.689, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Method components. We investigate the benefit of each method component in Tab. 4, reporting the results in PQ on the 100-50 and 100-10 settings in panoptic segmentation. Without applying any regularization technique, the performance on old classes drops to 0, denoting that it is important to deal with forgetting. Both the adaptive distillation loss (AD) and the pseudo-labeling strategy (PS) lead to improvement in the old classes when applied independently. However, we obtain the best results when combining them, especially considering the 100-10 scenario, where the old classes are improved by 5.8 PQ and the final performances of 3.9 PQ. Finally, the table shows that weighting the queries based on their semantic content is essential to improve performance: comparing a standard distillation" + }, + { + "type": "table", + "bbox": [ + 0.508, + 0.089, + 0.885, + 0.178 + ], + "angle": 0, + "content": "
Method100-50 (2 tasks)100-10 (6 tasks)
\\( RQ \\)\\( SQ \\)\\( PQ \\)\\( RQ \\)\\( SQ \\)\\( PQ \\)
MiB36.873.029.826.455.521.4
PLOP43.879.136.231.761.126.1
CoMFormer44.479.636.736.273.429.7
Joint47.679.039.547.679.039.5
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.187, + 0.892, + 0.23 + ], + "angle": 0, + "content": "Table 5. Panoptic metrics analysis. We report Recognition Quality (RQ), Segmentation Quality (SQ), and Panoptic Quality (PQ) over all the classes after the last step." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.244, + 0.892, + 0.29 + ], + "angle": 0, + "content": "(KD Eq. (1)) with our adaptive distillation (AD Eq. (3)), we largely improve the performance on both settings: \\(+0.3\\) PQ in 100-50 and \\(+3.7\\) PQ in 100-10 considering all classes." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.297, + 0.893, + 0.6 + ], + "angle": 0, + "content": "Forgetting in CoMFormer. To investigate the reason for the forgetting, we considered separately the Recognition Quality (RQ) and the Segmentation Quality (SQ) in panoptic segmentation. RQ measures how well the model is able to classify segments, while SQ considers the precision of the segmentation. Note that, as reported in [28], SQ and RQ are not independent since SQ is computed only on segments correctly classified. The results on the 100-50 and 100-10 are reported in Tab. 5. The table clearly shows that, as hypothesized in Sec. 3.3, forgetting happens heavily on the classification ability while it is less marked on the segmentation ability of the model. CoMFormer, in fact, without providing regularization losses on the mask prediction, is able to match the segmentation ability of the Joint upper bound on the 100-50 and it gets close results on the 100-10 (-5.6). In addition, it is evident that classification ability is heavily affected by forgetting. CoMFormer has a gap with Joint of 3.2 and 11.4 RQ respectively in the 100-50 and 100-10. We note that, however, other baselines are more affected by forgetting, showcasing even larger gaps." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.616, + 0.619, + 0.632 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.643, + 0.892, + 0.825 + ], + "angle": 0, + "content": "In this paper, we extend the task of continual learning beyond semantic segmentation by introducing the challenging and realistic continual panoptic segmentation setting. We propose CoMFormer, a strategy to address segmentation as a mask classification problem exploiting the properties of transformers to tackle both semantic and panoptic segmentation tasks. The extensive quantitative results highlight that CoMFormer outperforms all previous methods, both in semantic and panoptic segmentation, under a unified paradigm. In future work, we aim to extend the benchmark to instance segmentation to show that CoMFormer can be employed to address any continual segmentation task." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.831, + 0.636, + 0.844 + ], + "angle": 0, + "content": "Acknowledgments." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.845, + 0.892, + 0.901 + ], + "angle": 0, + "content": "We gratefully acknowledge the HPC infrastructure and Team at IIT. Fabio Cermelli acknowledges travel support from ELISE (GA no 951847). This work was partly supported by ANR grant VISA DEEP (ANR-20-CHIA-0022)." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3017" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.122, + 0.174, + 0.137 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.148, + 0.47, + 0.203 + ], + "angle": 0, + "content": "[1] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In ECCV, pages 213-229. Springer, 2020. (pages 2, 5)." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.205, + 0.472, + 0.261 + ], + "angle": 0, + "content": "[2] Fabio Cermelli, Dario Fontanel, Antonio Tavera, Marco Ciccone, and Barbara Caputo. Incremental learning in semantic segmentation from image labels. In CVPR, pages 4371-4381, 2022. (page 2)." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.263, + 0.472, + 0.332 + ], + "angle": 0, + "content": "[3] Fabio Cermelli, Antonino Geraci, Dario Fontanel, and Barbara Caputo. Modeling missing annotations for incremental learning in object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3700-3710, 2022. (pages 1, 2)." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.334, + 0.472, + 0.403 + ], + "angle": 0, + "content": "[4] Fabio Cermelli, Massimiliano Mancini, Samuel Rota Bulò, Elisa Ricci, and Barbara Caputo. Modeling the background for incremental learning in semantic segmentation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2020. (pages 1, 2, 3, 4, 5, 6)." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.405, + 0.47, + 0.461 + ], + "angle": 0, + "content": "[5] Fabio Cermelli, Massimiliano Mancini, Yongqin Xian, Zeynep Akata, and Barbara Caputo. Prototype-based incremental few-shot segmentation. In Proceedings of the British Machine Vision Conference (BMVC), 2021. (page 2)." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.462, + 0.47, + 0.517 + ], + "angle": 0, + "content": "[6] Arslan Chaudhry, Marc'Aurelio Ranzato, Marcus Rohrbach, and Mohamed Elhoseiny. Efficient lifelong learning with a-gem. In Proceedings of the International Conference on Learning Representations (ICLR), 2019. (page 2)." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.519, + 0.47, + 0.602 + ], + "angle": 0, + "content": "[7] Hongwei Chen, Laihui Ding, Fengqin Yao, Pengfei Ren, and Shengke Wang. Panoptic segmentation of UAV images with deformable convolution network and mask scoring. In Twelfth International Conference on Graphics and Image Processing (ICGIP 2020), volume 11720, pages 312-321. SPIE, 2021. (page 1)." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.604, + 0.47, + 0.687 + ], + "angle": 0, + "content": "[8] Liang-Chieh Chen, George Papandreou, Iasonas Kokkinos, Kevin Murphy, and Alan L. Yuille. Deeplab: Semantic image segmentation with deep convolutional nets, atrous convolution, and fully connected crfs. In IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 2018. (page 2)." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.689, + 0.47, + 0.743 + ], + "angle": 0, + "content": "[9] Liang-Chieh Chen, George Papandreou, Florian Schroff, and Hartwig Adam. Rethinking atrous convolution for semantic image segmentation. In arXiv preprint library, 2017. (pages 2, 5, 6, 7)." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.746, + 0.47, + 0.815 + ], + "angle": 0, + "content": "[10] Bowen Cheng, Maxwell D Collins, Yukun Zhu, Ting Liu, Thomas S Huang, Hartwig Adam, and Liang-Chieh Chen. Panoptic-deeplab: A simple, strong, and fast baseline for bottom-up panoptic segmentation. In CVPR, pages 12475-12485, 2020. (page 2)." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.817, + 0.47, + 0.871 + ], + "angle": 0, + "content": "[11] Bowen Cheng, Ishan Misra, Alexander G Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. In CVPR, 2022. (pages 2, 3, 5, 6, 7, 8)." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.47, + 0.901 + ], + "angle": 0, + "content": "[12] Bowen Cheng, Alexander G. Schwing, and Alexander Kirillov. Per-pixel classification is not all you need for semantic" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.148, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.094, + 0.891, + 0.121 + ], + "angle": 0, + "content": "segmentation. In Adv. Neural Inform. Process. Syst., 2021. (pages 2, 3, 6)." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.124, + 0.892, + 0.152 + ], + "angle": 0, + "content": "[13] Lee R. Dice. Measures of the amount of ecologic association between species. Ecology, 1945. (page 5)." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.154, + 0.892, + 0.25 + ], + "angle": 0, + "content": "[14] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In Proceedings of the International Conference on Learning Representations (ICLR), 2021. (page 2)." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.253, + 0.892, + 0.322 + ], + "angle": 0, + "content": "[15] Arthur Douillard, Yifu Chen, Arnaud Dapogny, and Matthieu Cord. Plop: Learning without forgetting for continual semantic segmentation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2021. (pages 1, 2, 4, 5, 6)." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.325, + 0.892, + 0.38 + ], + "angle": 0, + "content": "[16] Arthur Douillard, Yifu Chen, Arnaud Dapogny, and Matthieu Cord. Tackling catastrophic forgetting and background shift in continual semantic segmentation. In arXiv preprint library, 2021. (page 2)." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.383, + 0.892, + 0.452 + ], + "angle": 0, + "content": "[17] Arthur Douillard, Matthieu Cord, Charles Ollion, Thomas Robert, and Eduardo Valle. Podnet: Pooled outputs distillation for small-tasks incremental learning. In Proceedings of the IEEE European Conference on Computer Vision (ECCV), 2020. (pages 1, 2)." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.454, + 0.892, + 0.51 + ], + "angle": 0, + "content": "[18] Arthur Douillard and Timothee Lesort. Continuum: Simple management of complex continual learning scenarios. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR) Workshop, 2021. (page 5)." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.513, + 0.892, + 0.567 + ], + "angle": 0, + "content": "[19] Arthur Douillard, Alexandre Rame, Guillaume Couairon, and Matthieu Cord. Dytox: Transformers for continual learning with dynamic token expansion. In CVPR, 2022. (pages 1, 2, 3)." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.57, + 0.892, + 0.639 + ], + "angle": 0, + "content": "[20] Mark Everingham, S. M. Ali Eslami, Luc Van Gool, Christopher K. I. Williams, John M. Winn, and Andrew Zisserman. The pascal visual object classes challenge: A retrospective. In International Journal of Computer Vision (IJCV), 2015. (page 5)." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.642, + 0.892, + 0.669 + ], + "angle": 0, + "content": "[21] Robert French. Catastrophic forgetting in connectionist networks. Trends in cognitive sciences, 1999. (pages 1, 2, 4)." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.672, + 0.892, + 0.74 + ], + "angle": 0, + "content": "[22] Jun Fu, Jing Liu, Hajjie Tian, Yong Li, Yongjun Bao, Zhiwei Fang, and Hanqing Lu. Dual attention network for scene segmentation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2019. (page 2)." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.744, + 0.892, + 0.799 + ], + "angle": 0, + "content": "[23] K. He, X. Zhang, S. Ren, and J. Sun. Deep residual learning for image recognition. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016. (page 6)." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.802, + 0.892, + 0.871 + ], + "angle": 0, + "content": "[24] Zilong Huang, Xinggang Wang, Yunchao Wei, Lichao Huang, Humphrey Shi, Wenyu Liu, and Thomas S. Huang. Cynet: Criss-cross attention for semantic segmentation. In IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 2020. (page 2)." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.873, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[25] Loshchilov Ilya, Hutter Frank, et al. Decoupled weight decay regularization. Proceedings of ICLR, 2019. (page 6)." + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.094, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "3018" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.148 + ], + "angle": 0, + "content": "[26] Jitesh Jain, Anukriti Singh, Nikita Orlov, Zilong Huang, Jiachen Li, Steven Walton, and Humphrey Shi. Semask: Semantically masked transformers for semantic segmentation. arXiv preprint arXiv:2112.12782, 2021. (page 2)." + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.149, + 0.47, + 0.192 + ], + "angle": 0, + "content": "[27] Alexander Kirillov, Ross Girshick, Kaiming He, and Piotr Dólar. Panoptic feature pyramid networks. In CVPR, pages 6399-6408, 2019. (page 2)." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.192, + 0.469, + 0.234 + ], + "angle": 0, + "content": "[28] Alexander Kirillov, Kaiming He, Ross Girshick, Carsten Rother, and Piotr Dólár. Panoptic segmentation. In CVPR, pages 9404-9413, 2019. (pages 2, 5, 8)." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.236, + 0.47, + 0.332 + ], + "angle": 0, + "content": "[29] James Kirkpatrick, Razvan Pascanu, Neil Rabinowitz, Joel Veness, Guillaume Desjardins, Andrei A. Rusu, Kieran Milan, John Quan, Tiago Ramalho, Agnieszka Grabska-Barwinska, Demis Hassabis, Claudia Clopath, Dharshan Kumaran, and Raia Hadsell. Overcoming catastrophic forgetting in neural networks. Proceedings of the National Academy of Sciences, 2017. (pages 1, 2)." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.334, + 0.469, + 0.375 + ], + "angle": 0, + "content": "[30] Harold W. Kuhn. The hungarian method for the assignment problem. Naval Research Logistics Quarterly, 1955. (page 5)." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.377, + 0.469, + 0.433 + ], + "angle": 0, + "content": "[31] Duo Li, Guimei Cao, Yunlu Xu, Zhanzhan Cheng, and Yi Niu. Technical report for iccv 2021 challenge sslad-track3b: Transformers are better continual learners. In ICCV 2021 Challenge SSLAD-Track3B, 2021. (page 3)." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.434, + 0.469, + 0.489 + ], + "angle": 0, + "content": "[32] Yanwei Li, Hengshuang Zhao, Xiaojuan Qi, Yukang Chen, Lu Qi, Liwei Wang, Zeming Li, Jian Sun, and Jiaya Jia. Fully convolutional networks for panoptic segmentation with point-based supervision. IEEE TPAMI, 2022. (page 2)." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.49, + 0.469, + 0.533 + ], + "angle": 0, + "content": "[33] Z. Li and D. Hoiem. Learning without forgetting. Proceedings of the IEEE European Conference on Computer Vision (ECCV), 2016. (pages 1, 2)." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.534, + 0.469, + 0.589 + ], + "angle": 0, + "content": "[34] Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dólár. Focal loss for dense object detection. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), 2017. (page 5)." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.59, + 0.469, + 0.66 + ], + "angle": 0, + "content": "[35] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and Lawrence Zitnick. Microsoft coco: Common objects in context. In Proceedings of the IEEE European Conference on Computer Vision (ECCV), 2014. (page 5)." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.661, + 0.469, + 0.731 + ], + "angle": 0, + "content": "[36] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), 2021. (page 2)." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.732, + 0.469, + 0.788 + ], + "angle": 0, + "content": "[37] J. Long, E. Shelhamer, and T. Darrell. Fully convolutional networks for semantic segmentation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2015. (page 2)." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.789, + 0.469, + 0.845 + ], + "angle": 0, + "content": "[38] Andrea Maracani, Umberto Michieli, Marco Toldo, and Pietro Zanuttigh. Recall: Replay-based continual learning in semantic segmentation. In ICCV, pages 7026-7035, 2021. (page 2)." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.846, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[39] Umberto Michieli and Pietro Zanuttigh. Incremental learning techniques for semantic segmentation. In Proceedings of the IEEE International Conference on Computer Vision (ICCV) Workshop, 2019. (pages 1, 2, 4)." + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.148 + ], + "angle": 0, + "content": "[40] Umberto Michieli and Pietro Zanuttigh. Continual semantic segmentation via repulsion-attraction of sparse and disentangled latent representations. In CVPR, pages 1114-1124, 2021. (page 2)." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.149, + 0.892, + 0.218 + ], + "angle": 0, + "content": "[41] Gerhard Neuhold, Tobias Ollmann, Samuel Rota Bulò, and Peter Kontschieder. The mapillary vistas dataset for semantic understanding of street scenes. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), 2017. (page 1)." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.218, + 0.892, + 0.26 + ], + "angle": 0, + "content": "[42] Lorenzo Porzi, Samuel Rota Bulò, Aleksander Colovic, and Peter Kontschieder. Seamless scene segmentation. In CVPR, June 2019. (page 2)." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.26, + 0.892, + 0.329 + ], + "angle": 0, + "content": "[43] Sylvestre-Alvise Rebuffi, Alexander Kolesnikov, Georg Sperl, and Christoph H. Lampert. icarl: Incremental classifier and representation learning. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2017. (pages 1, 2)." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.33, + 0.892, + 0.357 + ], + "angle": 0, + "content": "[44] Anthony Robins. Catastrophic forgetting, rehearsal and pseudorehearsal. Connection Science, 1995. (page 2)." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.358, + 0.892, + 0.413 + ], + "angle": 0, + "content": "[45] Andrei A Rusu, Neil C Rabinowitz, Guillaume Desjardins, Hubert Soyer, James Kirkpatrick, Koray Kavukcuoglu, Razvan Pascanu, and Raia Hadsell. Progressive neural networks. arXiv preprint library, 2016. (pages 1, 2)." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.414, + 0.892, + 0.455 + ], + "angle": 0, + "content": "[46] Robin Strudel, Ricardo Garcia, Ivan Laptev, and Cordelia Schmid. Segmenter: Transformer for semantic segmentation. In ICCV, pages 7262-7272, 2021. (page 2)." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.456, + 0.892, + 0.483 + ], + "angle": 0, + "content": "[47] Sebastian Thrun. Lifelong learning algorithms. In Springer Learning to Learn, 1998. (page 2)." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.484, + 0.892, + 0.553 + ], + "angle": 0, + "content": "[48] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Hervé Jégou. Training data-efficient image transformers and distillation through attention. In International Conference on Machine Learning (ICML), 2021. (page 2)." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.554, + 0.892, + 0.609 + ], + "angle": 0, + "content": "[49] Hugo Touvron, Matthieu Cord, Alexandre Sablayrolles, Gabriel Synnaeve, and Hervé Jégou. Going deeper with image transformers. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), 2021. (page 2)." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.61, + 0.892, + 0.664 + ], + "angle": 0, + "content": "[50] Huiyu Wang, Yukun Zhu, Hartwig Adam, Alan Yuille, and Liang-Chieh Chen. Max-deeplab: End-to-end panoptic segmentation with mask transformers. In CVPR, pages 5463-5474, 2021. (page 2)." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.665, + 0.892, + 0.72 + ], + "angle": 0, + "content": "[51] Huiyu Wang, Yukun Zhu, Bradley Green, Hartwig Adam, Alan Yuille, and Liang-Chieh Chen. Axial-deeplab: Standalone axial-attention for panoptic segmentation. In ECCV, pages 108–126. Springer, 2020. (page 2)." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.721, + 0.892, + 0.789 + ], + "angle": 0, + "content": "[52] Zifeng Wang, Zizhao Zhang, Sayna Ebrahimi, Ruoxi Sun, Han Zhang, Chen-Yu Lee, Xiaoqi Ren, Guolong Su, Vincent Perot, Jennifer Dy, et al. Dualprompt: Complementary prompting for rehearsal-free continual learning. ECCV, 2022. (page 2)." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.79, + 0.892, + 0.845 + ], + "angle": 0, + "content": "[53] Zifeng Wang, Zizhao Zhang, Chen-Yu Lee, Han Zhang, Ruoxi Sun, Xiaoqi Ren, Guolong Su, Vincent Perot, Jennifer Dy, and Tomas Pfister. Learning to prompt for continual learning. In CVPR, pages 139-149, 2022. (page 2)." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.846, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[54] Enze Xie, Wenhai Wang, Zhiding Yu, Anima Anandkumar, Jose M Alvarez, and Ping Luo. Segformer: Simple and efficient design for semantic segmentation with transformers. Adv. Neural Inform. Process. Syst., 34, 2021. (page 2)." + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3019" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.148 + ], + "angle": 0, + "content": "[55] Shipeng Yan, Jiangwei Xie, and Xuming He. Der: Dynamically expandable representation for class incremental learning. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2021. (page 2)." + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.149, + 0.47, + 0.217 + ], + "angle": 0, + "content": "[56] Qihang Yu, Huiyu Wang, Siyuan Qiao, Maxwell Collins, Yukun Zhu, Hartwig Adam, Alan Yuille, and Liang-Chieh Chen. k-means mask transformer. In European Conference on Computer Vision, pages 288-307. Springer, 2022. (page 2)." + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.219, + 0.471, + 0.275 + ], + "angle": 0, + "content": "[57] Yuhui Yuan, Xilin Chen, and Jingdong Wang. Object-contextual representations for semantic segmentation. In Proceedings of the IEEE European Conference on Computer Vision (ECCV), 2020. (page 2)." + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.277, + 0.47, + 0.318 + ], + "angle": 0, + "content": "[58] Yuhui Yuan and Jingdong Wang. Ocnet: Object context network for scene parsing. In arXiv preprint library, 2018. (page 2)." + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.32, + 0.47, + 0.401 + ], + "angle": 0, + "content": "[59] Chang-Bin Zhang, Jia-Wen Xiao, Xialei Liu, Ying-Cong Chen, and Ming-Ming Cheng. Representation compensation networks for continual semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7053-7064, 2022. (pages 1, 2, 5, 6, 7)." + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.403, + 0.47, + 0.459 + ], + "angle": 0, + "content": "[60] H. Zhao, J. Shi, X. Qi, X. Wang, and J. Jia. Pyramid scene parsing network. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2017. (page 2)." + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.46, + 0.469, + 0.529 + ], + "angle": 0, + "content": "[61] Hengshuang Zhao, Yi Zhang, Shu Liu, Jianping Shi, Chen Change Loy, Dahua Lin, and Jiaya Jia. Pisanet: Pointwise spatial attention network for scene parsing. In Proceedings of the IEEE European Conference on Computer Vision (ECCV), 2018. (page 2)." + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.531, + 0.469, + 0.6 + ], + "angle": 0, + "content": "[62] Bolei Zhou, Hang Zhao, Xavier Puig, Sanja Fidler, Adela Barriuso, and Antonio Torralba. Scene parsing through ade20k dataset. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2017. (page 5)." + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.471, + 0.6 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.956 + ], + "angle": 0, + "content": "3020" + } + ] +] \ No newline at end of file diff --git a/2023/CoMFormer_ Continual Learning in Semantic and Panoptic Segmentation/9493268b-7f35-48c3-8aa1-39be491edc53_origin.pdf b/2023/CoMFormer_ Continual Learning in Semantic and Panoptic Segmentation/9493268b-7f35-48c3-8aa1-39be491edc53_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..3f3c84bc97dde2e0402f612a2473ff36b8119036 --- /dev/null +++ b/2023/CoMFormer_ Continual Learning in Semantic and Panoptic Segmentation/9493268b-7f35-48c3-8aa1-39be491edc53_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d96016899a52660f829dcc9660cde4c0bf78b8637687e0bffcbacfbad184cfa6 +size 1288275 diff --git a/2023/CoMFormer_ Continual Learning in Semantic and Panoptic Segmentation/full.md b/2023/CoMFormer_ Continual Learning in Semantic and Panoptic Segmentation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..cb6d2d79e3d62f9f3901f2419dc6efc40229df1b --- /dev/null +++ b/2023/CoMFormer_ Continual Learning in Semantic and Panoptic Segmentation/full.md @@ -0,0 +1,308 @@ +# CoMFormer: Continual Learning in Semantic and Panoptic Segmentation + +Fabio Cermelli $^{1,2*}$ , Matthieu Cord $^{3,4}$ , Arthur Douillard $^{3\dagger}$ + +$^{1}$ Politecnico di Torino, $^{2}$ Italian Institute of Technology, $^{3}$ Sorbonne Université, $^{4}$ valeo.ai + +fabio.cermelli@polito.it, matthieu.cord@sorbonne-universite.fr, ar.douillard@gmail.com + +# Abstract + +Continual learning for segmentation has recently seen increasing interest. However, all previous works focus on narrow semantic segmentation and disregard panoptic segmentation, an important task with real-world impacts. In this paper, we present the first continual learning model capable of operating on both semantic and panoptic segmentation. Inspired by recent transformer approaches that consider segmentation as a mask-classification problem, we design CoMFormer. Our method carefully exploits the properties of transformer architectures to learn new classes over time. Specifically, we propose a novel adaptive distillation loss along with a mask-based pseudo-labeling technique to effectively prevent forgetting. To evaluate our approach, we introduce a novel continual panoptic segmentation benchmark on the challenging ADE20K dataset. Our CoMFormer outperforms all the existing baselines by forgetting less old classes but also learning more effectively new classes. In addition, we also report an extensive evaluation in the large-scale continual semantic segmentation scenario showing that CoMFormer also significantly outperforms state-of-the-art methods. + +# 1. Introduction + +Image segmentation is a fundamental computer vision problem that enables machines to assign an image's pixels to discrete segments. Multiple segmentation tasks have been defined depending on the segments definitions. Semantic segmentation clusters pixels by classes, merging in a single segment pixels belonging to instances of the same class. Panoptic segmentation assigns to every pixel a semantic class while separating different instances into different segments. This latter kind of segmentation has real-world impacts in autonomous robots and vehicles [7, 41]. + +Despite tremendous progress in image segmentation, the current approaches are trained on a static dataset with a pre + +![](images/b8d3347a7632053e053c6c4995d36b1c1c99a69545bc93bac49c7dac066977d0.jpg) +Step t-1 - New Class: Car +Image +CoMFormer (t-1) + +![](images/9967a26e5908b0367f68e04b1f5dce02f2397f78defc3567ae5a1a9360a2e5f0.jpg) + +![](images/0dff6aec1f26369fcf162fe0e066f1246e89d7c6f19a07a4cf9e31a8f5eba346.jpg) +Predictions + +![](images/249792d01ea386ca4a040a0c03d29c59e249569a2a81740d16f2510f1f498045.jpg) +Image +Annotations + +![](images/4b01263540676814d6bda30c505c2ef988424ff64e4e22e40641f9f3af7b13c2.jpg) +Step t - New Class: Person +CoMFormer (t) + +![](images/756eccefecc29ee7a7d99a8af6be340e0df1e397300f4c2b43762d9d604c6c58.jpg) +Predictions +Figure 1. Illustration of our model, CoMFormer, operating in continual segmentation. Relying on the mask classification paradigm, it is able to cope with both continual semantic and panoptic segmentation without any modification by predicting masks for both old (e.g. car in red) and new (e.g. person in green) classes. The figure reports two classes and no "stuff" (e.g. road, building) only for illustration purposes. + +defined set of classes. Whenever an update of the model is required to fit new classes, the common solution is to train a model from scratch on the union of the old and new class data. A computationally more efficient solution would be to fine-tune the existing model solely on the new class data. Unfortunately, this approach would cause a catastrophic forgetting [21] of the old classes on which the model performance would be extremely degraded. + +The problem of updating the knowledge of the model over time is typically referred as continual learning. It has been traditionally studied in the context of image classification [17, 19, 29, 33, 43, 45] and is gaining attention on the segmentation task [3, 4, 15, 39, 59] due to the more realistic applications and the additional challenges it introduces, such as the background shift [4]. However, current state-of-the-art methods mainly focus on semantic segmentation and are not designed to work in other segmentation tasks, strongly limiting their application in the real world. + +In this paper, we design the first method operating in both continual semantic and panoptic segmentation, as illustrated in Fig. 1. Our method, CoMFormer (Continual MaskFormer), takes inspiration from recent transformer architectures [11, 12], approaching segmentation as a mask classification problem. Instead of predicting a class probability for each pixel, as in previous semantic segmentation works [9, 37], it predicts a set of binary masks, each associated with a single class prediction, effectively addressing both segmentation tasks without any modification in the training architecture and procedure. Differently from previous works [11, 12], however, CoMFormer forces the output binary masks to be mutually exclusive to one another: a pixel can only be predicted by a single binary mask to prevent having several masks classifying the same pixel with different classes. This behavior is crucial in continual learning to reduce the interference among old and new classes. + +Furthermore, CoMFormer introduces a novel adaptive distillation loss to alleviate forgetting. It enforces consistency of the model's classification predictions across learning steps only when it is useful to remember old classes, ensuring a better tradeoff between rigidity (not forgetting old classes) and plasticity (learning efficiently new classes). Finally, since at each training iteration the dataset reports annotations only for the current classes, we design a mask-based pseudo-labeling technique to generate annotations for the old classes, effectively alleviating forgetting. To reduce the noise, we consider the prediction confidence and we avoid interference with ground-truth annotations. + +We validate CoMFormer on both continual segmentation tasks. For panoptic segmentation, we define a new benchmark relying on the challenging ADE20K where we demonstrate that CoMFormer largely outperforms all previous baselines. On semantic segmentation, we show that CoMFormer outperforms the existing state-of-the-art methods on every setting of the large-scale ADE20K benchmark. + +To sum up, the contributions of this paper are as follows: + +- We introduce continual panoptic segmentation which has real-world impacts in addition to being significantly more challenging than previous benchmarks. +- We propose CoMFormer to tackle both continual panoptic and semantic segmentation. To avoid forgetting, we design a novel adaptive distillation and an efficient mask-based pseudo-labeling strategy. +- Through extensive quantitative and qualitative benchmarks, we showcase the state-of-the-art performance of our model on both continual segmentation tasks. + +# 2. Related Works + +Semantic and Panoptic Segmentation. The two tasks have been traditionally treated separately, with specialized architectures proposed for either one or the other task, without interoperability. Semantic segmentation has been tradi + +tionally addressed as a per-pixel classification task. Fully convolutional network [37] dominated the field by aggregating long-range dependencies in the features map [8, 9, 60] and exploiting contextual information [22, 24, 57, 58, 61]. Recently, transformers [26, 46, 54] are replacing convolutions by integrating long-range dependencies at every layer. Panoptic segmentation [28] has been proposed to unify semantic and instance segmentations. Initially, methods proposed to combine each task-specific architectures [10, 27, 32, 42] or defined new specialized architectures and objective functions for the panoptic task [1, 50, 51], drifting from a general solution for all the segmentation tasks. Recently, methods addressing segmentation as a mask classification problem [11, 12, 50, 56] have been proposed, introducing a transformer architecture able to solve multiple tasks at once. MaskFormer [12] was the first to propose a single architecture to address both panoptic and semantic segmentation. Mask2Former [11] improves it by adopting multi-scale features, masked attention and optimization tricks. Concurrently, kMaX-DeepLab [56] extended [50] proposing to reformulate the cross-attention as a clustering process. Despite their effectiveness in standard training setting, these methods suffer from catastrophic forgetting. In this work, we aim to extend their capability and we propose CoMFormer to learn to segment new classes over time. + +Continual Segmentation. Continual learning is a longstanding field addressing the problem of learning new knowledge over time while avoiding catastrophic forgetting [21, 44, 47]. Traditionally, it has been studied in the context of image classification [6, 17, 19, 29, 33, 43, 45, 55] but there is a growing interest in its applications in segmentation [2-5, 15, 16, 38-40, 59]. Continual semantic segmentation introduces additional challenges, as pointed out by [4]. In particular, considering the task as a per-pixel classification problem, catastrophic forgetting is exacerbated by the background shift, where old classes pixels are treated as background in following training steps. [4] proposed a revision of the standard knowledge distillation framework to deal with it. [17, 39] introduced a distillation loss preserving the representation at feature level while [38] proposed an approach using old classes' examples to alleviate forgetting. Recently, [59] presented a technique that decouples learning the representation of the new classes while freezing the representation for the old ones. Differently from previous work, we extend the benchmarks to solve both semantic and the more challenging panoptic segmentation task. + +Continual Learning with Transformers. Recent progresses made using transformers for computer vision [14, 36, 48, 49] attracted the attention of the continual learning community [19, 52, 53]. In particular, DyTox [19] proposed to specialize the architecture on each task using a different task-specific token. Learning-to-Prompt [53] stores a + +![](images/2c32f8ab18b9358039da99d591fe4e93e53421fcb9015a9bd11c35250c11d980.jpg) +Figure 2. Overview of CoMFormer. At learning step t, both the current (bottom) and old (top) models output, for each input query, a pair made of a class probability distribution and a binary mask. While learning the new class (person), we regularize the current model to prevent forgetting the old classes (car) by: (i) generating pseudo-labels for the old classes present in the image and (ii) by applying an adaptive distillation loss that reweights the contribution of each output based on the information it carries on the old classes. $\psi$ indicates the dot product between the mask embeddings $\mathcal{E}_{mask}$ and the pixel embeddings $\mathcal{E}_{pixel}$ followed by the softmax activation. + +pool of prompts that are employed to condition the whole forward execution of the patch tokens. To the best of our knowledge, transformers have been evaluated only in continual learning for image classification [19, 31] and we are the first to use them in the continual segmentation task. + +# 3. CoMFormer + +# 3.1. Problem Definition + +The goal of image segmentation is to learn a model able to (i) partition the image into a set of N regions represented by binary masks and (ii) produce a class-probability distribution associated with each region. The differences among segmentation tasks rely on the semantics of the masks: semantic segmentation groups all the pixels of a class while panoptic segmentation distinguishes different object instances. In the following, we provide a general formulation for the two tasks since they only differ in the construction of the binary masks in the dataset. + +Continual segmentation aims to train the model in multiple learning steps $t = 1, \ldots, T$ , introducing at every step a new set of classes. Formally, during the learning step $t$ a dataset $D_{t}$ consisting of a collection of image and label pairs is provided. The label of each image takes the form of a set of $N^{gt}$ ground truth segments $z^{gt}\{(c_i^{gt}, m_i^{gt}) | c_i^{gt} \in \mathcal{K}^t, m_i^{gt} \in \{0,1\}^{H \times W}\}_{i=1}^{N^{gt}}$ where $c_i^{gt}$ is the ground-truth class and $m_i^{gt}$ is binary mask $m^{gt}$ . $\mathcal{K}^t$ is the set of classes introduced at step $t$ , and $H, W$ are the height and width of the images. The goal of training step $t$ is to learn a model able to predict segments for all the seen classes $\mathcal{K}^{1:t}$ . + +We note that the dataset only contains annotations for the new classes $\mathcal{K}^t$ while not reporting ground-truth segments for old $\mathcal{K}^{1:t - 1}$ or future $\mathcal{K}^{t + 1:T}$ classes. Moreover, differently from previous works [4], old and future class segment + +annotations are simply absent during training rather than being collapsed into an artificial background class. + +# 3.2. CoMFormer Architecture + +To solve semantic and panoptic segmentation tasks within a single method, we take inspiration from Mask-Former architectures [11, 12] considering segmentation a mask classification problem. It consists in predicting for each image a set of $N$ pairs made of a class prediction and a binary mask. Formally, the output of a CoMFormer is $z = \{(p_i, m_i)\}_{i=1}^N$ , where $p_i$ represents a class probability distribution over the seen classes, i.e. $p_i \in \Delta^{K^t + 1}$ with $K^t = |\mathcal{K}^{1:t}|$ , and $m_i$ is a binary mask such that $m_i \in [0, 1]^{H \times W}$ . We note that $p_i$ contains an additional "no object" class (denoted $\varnothing$ ) to indicate that the mask does not correspond to any of the known $\mathcal{K}^{1:t}$ categories. + +To obtain such output, the CoMFormer architecture, showcased in Fig. 2, is made of three components: (i) a backbone that extracts feature embeddings $\mathcal{E}_{\text {feat }}$ , (ii) a transformer decoder that takes as input N learnable queries and $\mathcal{E}_{\text {feat }}$ to output $N$ per-segment embeddings $Q$ , (iii) a pixel decoder that takes $\mathcal{E}_{\text {feat }}$ as input and extracts per-pixel embeddings $\mathcal{E}_{\text {pixel }}$ . The output class probabilities $\{p_i \in \Delta^{K^t + 1}\}_{i=1}^N$ are obtained by applying a linear classifier on $Q$ , which is enlarged as new classes are presented. We obtain the mask predictions $m_i$ combining $Q$ and $\mathcal{E}_{\text {pixel }}$ : first $Q$ is passed into a 2-layer MLP to obtain N mask embeddings $\mathcal{E}_{\text {mask }}$ and we then compute the dot product between the $i$ -th mask embedding and the pixel embeddings, followed by the softmax activation: $m_i[h, w] = \operatorname{softmax}(\mathcal{E}_{\text {mask}}[:, i]^T \cdot \mathcal{E}_{\text {pixel}}[:, h, w])$ . + +We point out that employing the softmax contrasts sharply with MaskFormers [11, 12], which use sigmoid ac + +![](images/a565ed99ec382a9209e368e4e7d0ebbf02cf6e43a6821034004cc1717fc002b1.jpg) +Figure 3. Illustration of the mask predictions between sigmoid and softmax. MaskFormer has overlap across masks (e.g. in Query 1 and 2) predicted either as old or new classes, while CoMFormer avoids the issue using the softmax activation. + +tivation to get the mask predictions. Using the sigmoid activation would result in the overlap of many masks on a single pixel, each potentially belonging to a distinct class, as seen in Fig. 3. This is particularly important in the context of continual learning, when the model may predict two similar masks associated with two distinct classes (e.g. cat and dog classes in Fig. 3) thereby generating interference among old and new classes and degrading the model's performance. + +# 3.3. Learning without Forgetting + +To learn new classes, a simple solution is to fine-tune the model on the new dataset. However, because it lacks annotations for old classes, this operation results in catastrophic forgetting [21]. To alleviate the issue, we introduce two techniques for avoiding forgetting: an adaptive distillation loss and a mask-based pseudo-labeling strategy. + +Adaptive Distillation Loss. A common strategy to mitigate forgetting adopted by continual semantic segmentation approaches [4, 15, 39] is to distill the knowledge coming from the old model in the new one, either in form of features [15, 39] or as classification probabilities [4]. We take inspiration from the latter solution and we design a novel distillation loss tailored to the CoMFormer architecture. In particular, analyzing the two components constituting the output space, we find the mask prediction being robust to forgetting while the classification part is heavily affected (see Sec. 4.6). For this reason, we focus on the output class probabilities to design our distillation loss. + +An adaptation of standard knowledge distillation works [4] would force the new model to mimic the old model probabilities over all the N outputs, independently from their content. Formally, given the set of N probability distributions $\{p_i^o\}_{i = 1}^N$ coming from the old model, a standard distillation loss is computed as + +$$ +\mathcal {L} _ {K D} = - \frac {1}{N} \sum_ {i = 1} ^ {N} \sum_ {k = 1} ^ {K ^ {t - 1} + 1} p _ {i} ^ {o} (k) \log \frac {\tilde {p} _ {i} (k)}{p _ {i} ^ {o} (k)}, \tag {1} +$$ + +where $K^{t - 1} + 1$ is the $\varnothing$ class and $\tilde{p}_i(k)$ represents the unbiased probability distribution of $i$ -th output for class $k$ : + +$$ +\tilde {p} _ {i} (k) = \left\{ \begin{array}{l l} \sum_ {j \in \mathcal {K} ^ {t}} p _ {i} (j) + p _ {i} (\varnothing) & \text {i f} k = \varnothing , \\ p _ {i} (k) & \text {o t h e r w i s e .} \end{array} \right. \tag {2} +$$ + +Note that, similarly to [4], it forces the sum of the model new class $(\mathcal{K}^t)$ and the "no object" class $(\varnothing)$ probabilities to be similar to the old model $\varnothing$ probability. + +However, not every output probability brings relevant information: most of them are predicted as "no object" $(\varnothing)$ and thus they do not carry enough details to remember the old classes while reducing the relative importance of the other outputs probabilities. For this reason, we re-weight their contribution based on the probability of not being $\varnothing$ . Formally, we define our adaptive distillation loss as + +$$ +\mathcal {L} _ {A D} (p, p ^ {o}) = - \sum_ {i = 1} ^ {N} \frac {\omega_ {i}}{\sum_ {i = 1} ^ {N} \omega_ {i}} \sum_ {k = 1} ^ {K ^ {t - 1} + 1} p _ {i} ^ {o} (k) \log \frac {\tilde {p} _ {i} (k)}{p _ {i} ^ {o} (k)}, \tag {3} +$$ + +where the weighting coefficient is $\omega_{i} = (1 - p_{i}^{o}(\varnothing))^{2}$ . Note that, with this formulation, we effectively reduce the contributions of outputs that bring small information about the old classes $(p_i^o (\varnothing)\approx 1)$ , while we increase the contribution of important ones (where $p_i^o (\varnothing)\approx 0$ ). + +Mask-based Pseudo-labeling Strategy. Forgetting is effectively reduced by the adaptive distillation loss but we are not, however, taking advantage of the presence of the old classes that are unlabeled in the dataset. We propose using the old model predictions to recognize old class segments and generate pseudo-labels to make efficient use of this information and improve the knowledge of past classes. + +A simple strategy consists in considering the N output pairs coming from the old model independently and using all the pairs for which the predicted class is not $\varnothing$ as pseudolabels. This strategy, however, does not consider two aspects: there may be pairs where the class is different from $\varnothing$ but (i) the mask substantially overlaps with a ground-truth segment, or (ii) the mask is noisy and has low confidence. + +To overcome these issues, we propose a mask-based pseudo-labeling strategy that jointly considers the mask and class prediction confidences to avoid noisy labels and overlaps with the existing annotations. We define the model confidence as the multiplication among the class and mask probabilities. Formally, we compute the confidence for the $i$ -th output as $q_{i} = p_{i}^{max} \cdot m_{i}^{o} \in [0,1]^{H \times W}$ , where $p_{i}^{max} = \max_{k=1}^{K^{t-1}} p_{i}^{o}(k)$ and $m_{i}^{o}$ is the binary mask predicted by the old model. Note that we do not consider the $\varnothing$ class in the max operation. We denote the pseudo-class $c_{i}^{ps} = \operatorname{argmax}_{k=1}^{K^{t-1}} p_{i}^{o}(k)$ and we generate the pseudomask $m_{i}^{ps}$ considering two criteria: (i) there should be no overlap between the pseudo-mask and the ground-truth seg + +ments and (ii) to be included in a pseudo-mask, the confidence on a pixel should be maximum over all outputs. Formally, denoting the binarization of the predicted mask as $m_{i}^{bin} = m_{i} > 0.5$ and the union of the ground-truth segments $m_{all}^{gt} = \sum_{i=1}^{N^{ps}} m_{i}^{gt}$ , we generate the pseudo-mask as + +$$ +m _ {i} ^ {p s} [ h, w ] = \left\{ \begin{array}{l l} m _ {i} ^ {b i n} [ h, w ] & \text {i f} i = \operatorname {a r g m a x} _ {j = 1} ^ {N} q _ {j} [ h, w ] \\ & \text {a n d} m _ {a l l} ^ {g t} [ h, w ] = 0, \\ 0 & \text {o t h e r w i s e .} \end{array} \right. \tag {4} +$$ + +Indeed, not all pseudo-mask should be included in the set of pseudo-labels. There may be pseudo-masks where no pixel is active, i.e. it is all zeros, or where the pseudomask contains only a small fraction of the original mask. For these reasons, we construct the set of pseudo-labels $z^{ps} = \{(c_i^{ps},m_i^{ps})\}_{i = 1}^{N^{ps}}$ such that we include in $z^{ps}$ only the pseudo-masks that retain at least half of the pixels w.r.t. the binary mask $m_{i}^{bin}$ and that have at least one active pixel. + +We denote the final annotation set $\bar{z} = \{\bar{c}_i, \bar{m}_i\}_{i=1}^N$ as the union of the ground-truth labels $z^{gt}$ and the pseudo-labels $z^{ps}$ , i.e. $\bar{z} = z^{gt} \cup z^{ps}$ and $\bar{N} = N^{gt} + N^{ps}$ . + +Overall training loss. To train CoMFormer, a one-to-one matching $\sigma$ between the model predictions $z$ and the annotation set $\bar{z}$ is required. Following standard practices [1, 11], we employ the Hungarian matching algorithm [30] and we search for the matching $\sigma$ that minimize the assignment cost computed as $-p_i(\bar{c}_j)\cdot Dice(m_i,\bar{m}_j)$ , where Dice indicates the Dice coefficient [13]. We assume the size of the prediction set to be larger than the annotations set (i.e. $N > \bar{N}$ ) and to obtain a one-to-one matching we pad the annotations with "no object" tokens $\varnothing$ . + +Once the best matching $\sigma$ has been found, we define the overall training loss as: + +$$ +\mathcal {L} = \mathcal {L} _ {\text {S e g}} (z, \bar {z}) + \lambda_ {D} \mathcal {L} _ {A D} (p, p ^ {o}), \tag {5} +$$ + +where $\lambda_{D}$ is a trade-off hyper-parameter and + +$$ +\begin{array}{l} \mathcal {L} _ {\text {S e g}} (z, \bar {z}) = \sum_ {j = 1} ^ {N} [ - \alpha (1 - p _ {\sigma (j)} (\bar {c} _ {j})) ^ {\gamma} \log p _ {\sigma (j)} (\bar {c} _ {j}) + \tag {6} \\ \left. + \lambda_ {m a s k} \mathbb {1} _ {\bar {c} _ {j} \neq \emptyset} \mathcal {L} _ {m a s k} \left(m _ {\sigma (j)}, \bar {m} _ {j}\right) \right]. \\ \end{array} +$$ + +We note that the first term in $\mathcal{L}_{Seg}$ indicates the focal loss [34], where $\alpha$ and $\gamma$ are hyper-parameters. Finally, $\mathcal{L}_{mask}$ is the sum of dice and cross-entropy losses, and $\lambda_{mask}$ is a hyper-parameter. We note that $\mathcal{L}_{mask}$ is computed only for valid segments, i.e. where $\bar{c}_j \neq \emptyset$ . + +# 4. Experiments + +# 4.1. Dataset and Settings + +We start from the widely adopted continual semantic segmentation benchmark defined in [4] and we extend it to + +the newly proposed continual panoptic segmentation using a modified Continuum library [18]. We compare our novel model on the large-scale ADE20K [62] dataset since it supports both semantic and panoptic segmentation tasks. This challenging dataset contains 150 classes, divided into 100 "things" and 50 "stuff" categories. This dataset represents a wide variety of scenes, both interior and exterior, with images featuring an average of 9.9 classes, while other datasets like COCO [35] only have an average of 3.5 classes. + +Continual Learning Protocols. Previous continual semantic segmentation works [4, 15] describe a training protocol to assess the performance on multiple continual learning steps. In particular, [4] proposed three protocols for ADE20K with different numbers of tasks: (i) the 100-50 consists of two tasks, the first of 100 and the second of 50 classes; (ii) the 50-50 consists of three tasks of 50 classes; (iii) the 100-10 consists on 6 tasks, the first on 100 classes followed by 5 tasks of 10 classes. In addition, we use the 100-5 introduced in [15], consisting of 11 tasks, the first of 100 classes followed by 10 tasks of 5 classes. To divide the images into different tasks, we follow the splits provided by [4] for 100-50, 100-10, 50-50 for both semantic and panoptic segmentation. In particular, they ensured that each image appears only on a unique task. Differently, for the 100-5 we use the split proposed by [15], where the same image may appear on multiple tasks. We report the results of the 50-50 setting for both semantic and panoptic segmentation in the supplementary material. + +Metrics. We compare the different models using the mean Intersection over Union (mIoU) on semantic segmentation [20] and Panoptic Quality (PQ) on panoptic segmentation [28]. PQ is the product of two components: Segmentation Quality (SQ) which considers the IoU between the correctly classified segments, and Recognition Quality (RQ) which only considers the classification accuracy. For both tasks, we report the metric after the last step $T$ on the first classes $(C^1)$ , for the added classes $(C^{2:T})$ , and for all the classes (all). We also report the average of the final performance on all seen classes after each step (avg) following [15]. + +# 4.2. Baselines + +We benchmark our model against the state-of-the-art methods: MiB [4], PLOP [15], and RCIL [59]. In semantic segmentation, we report them with the original DeepLab-v3 segmentation architecture [9]. In addition, for a fair comparison, we report MiB and PLOP using the CoMFormer architecture. In particular, we implement the methods in two versions: (i) we replace the CoMFormer losses with a per-pixel cross-entropy loss on the dot product between mask and class logits and we apply the continual learning methods on that baseline (we refer to this as Per-Pixel), and (ii) by applying the distillation losses directly on the + +
Method100-50 (2 tasks)100-10 (6 tasks)100-5 (11 tasks)
1-100101-150allavg1-100101-150allavg1-100101-150allavg
FT0.025.88.624.80.02.91.07.90.01.30.44.6
MiB35.119.329.835.427.110.021.429.124.06.518.125.6
PLOP41.026.636.238.630.517.526.132.928.115.724.030.5
CoMFormer41.127.736.738.836.017.129.735.334.415.928.234.0
Joint43.232.139.5—43.232.139.5—43.232.139.5—
+ +Table 1. Continual Panoptic Segmentation results on ADE20K dataset in PQ considering multiple training protocols. + +
BackboneMethod100-50 (2 tasks)100-10 (6 tasks)100-5 (11 tasks)
1-100101-150allavg1-100101-150allavg1-100101-150allavg
DeepLab-v3 [9]MiB [4]40.517.232.837.338.311.329.235.136.05.726.032.7
PLOP [15]41.914.932.937.440.514.131.636.639.17.828.835.3
RCIL [59]42.318.834.5—39.317.632.1—38.511.529.6—
Per-PixelMiB40.324.034.837.535.114.028.134.033.315.227.333.1
PLOP40.220.233.536.932.613.726.332.433.39.425.432.8
Mask-basedFT0.026.78.926.40.02.30.88.50.01.10.34.2
MiB37.024.132.638.323.510.626.629.621.06.116.127.7
PLOP44.226.238.241.134.815.928.535.233.614.127.133.6
CoMFormer44.726.238.441.240.615.632.337.439.513.630.936.5
Joint46.935.643.1—46.935.643.1—46.935.643.1—
+ +Table 2. Continual Semantic Segmentation results on the ADE20K dataset in mIoU considering multiple training protocols. We report the numbers published by the original papers on DeepLab-v3 [9]. The remaining results are based on our re-implementation (see Sec. 4.2). + +classification head of the CoMFormer architecture (Mask-based). We note that the first version cannot be applied to panoptic segmentation, since the task does not support the use of per-pixel losses. For PLOP, we apply its distillation loss local-POD on the intermediate features of the segmentation backbone, as in the original paper, and we use the pseudo-labeling strategy proposed in Sec. 3. RCIL was not re-implemented since it is not possible to extend their approach to non-convolutional architectures like ours. We include a naive finetuning (FT) without any capabilities for continual learning. Finally, we also report an upper bound (Joint) that is trained in the traditional segmentation setting. We made our best effort to find the most suitable hyperparameters for all the baselines. + +# 4.3. Implementation Details + +Architecture. Following the previous benchmarks [4, 15] we use as backbone a ResNet101 [23] on semantic segmentation and we propose to use a ResNet50 [23] for panoptic segmentation. For both tasks, we use the pixel decoder and the transformer decoder proposed in [11] with 9 layers in total and $N = 100$ queries. In addition, we adopt all the improvements introduced in [11]: masked attention, multiscale features, and their optimization improvements. We report the performance of single-scale inference. + +Training parameters. We follow the Mask2Former [11] hyper-parameters and we use AdamW [25] optimizer with an initial learning rate of 0.0001 for the first step ( $t = 0$ ) and 0.00005 in the following ( $t > 0$ ). A learning rate multiplier + +of 0.1 is applied to the backbone and we follow a polynomial learning rate schedule. We use weight decay of 0.05. We train the model for 160K iterations in the first step and 400 iterations per class in the following (e.g. learning 50 classes we train for $50 \times 400 = 20000$ iterations). For semantic segmentation, we use a crop size of 512 and a batch size of 16. For panoptic segmentation, we use a crop size of 640 and a batch size of 8. For both, we use the standard random scale jittering between 0.5 and 2.0, random horizontal flipping, random cropping, as well as random color jittering as data augmentation [12]. We set $\alpha = 20$ , $\gamma = 2$ , and $\lambda_{mask} = 5.0$ . $\lambda_{D}$ is set to 1 for 100-50 and to 10 in the 100-10 and 100-5. Finally, following the standard protocol of continual learning [4, 15] we do not store any image of previous steps and we do not use rehearsal learning. + +# 4.4. Continual Panoptic Segmentation + +Tab. 1 reports the results on the new continual panoptic segmentation benchmark. CoMFormer exceeds all the baselines by a significant margin. Considering the 100-50, we can see that CoMFormer outperforms MiB on both old (+6 PQ) and new (+8.4 PQ) classes and PLOP on new classes (+1.1 PQ). Considering a longer sequence of tasks (100-10 and 100-5), we can see that the gap between CoMFormer and the other methods becomes more significant, especially in the old classes. In 100-10, it surpasses MiB and PLOP respectively by 8.9 and 5.5 PQ in old classes, while in the new classes it exceeds MiB by 7.1 and is comparable to PLOP. In the 100-5, it outperforms MiB and PLOP + +![](images/b87103b4b638a59f18c508ea4fd257ba27f4faa0c031eb21538dd618859f39c2.jpg) +Figure 4. Qualitative results of CoMformer v.s. MiB and PLOP on the 100-10 continual panoptic segmentation setting on ADE20K. + +on both old (+10.4 and +6.3 PQ) and new classes (+9.4 and +0.2 PQ). Finally, we note that CoMFormer obtains performance close to the Joint baseline in the 100-50 setting, suffering a drop of only 2.1 PQ on the old classes and 4.4 PQ on the new ones. In the other settings, the gap is still relevant, indicating that learning from multiple training steps is still a very challenging task. + +In Fig. 4, we report a qualitative comparison with the baselines in the 100-10 setting using randomly chosen images. We can see that our method is able to maintain better the previous knowledge by precisely segmenting all the classes in the image; for example, w.r.t. to PLOP, our method correctly segments the chair in the second row and the boat in the third row. We can note a common error: models correctly segment but misclassify objects in the image, like the plane classified as car in the first row or the press classified as a door in the third row. + +# 4.5. Continual Semantic Segmentation + +We report the comparison with state-of-the-art methods in continual semantic segmentation in Tab. 2. CoMFormer achieves the best results on every setting, surpassing both methods based on DeepLab-v3 [9] and baselines implemented upon the same architecture. In particular, CoM- + +Former outperforms the best state-of-the-art method (RCIL [59] by $3.9\%$ on the 100-50, by $0.2\%$ on 100-10, and by $1.3\%$ on the 100-5. Moreover, it always improves the performance on the old classes and the avg across different steps. In addition, CoMFormer exceeds all the baselines. Comparing it with PLOP (Mask-based), we can see that it obtains a substantial improvement, especially considering scenarios with longer training protocols: CoMformers achieves $30.9\%$ $(32.3\%)$ on the all metrics, while PLOP only $27.1\%$ $(28.5\%)$ on the 100-5 $(100 - 10)$ setting. Furthermore, the results confirm that improvements are not led by the newer architecture. Per-Pixel baselines achieve only similar or even lower performance than their counterparts. MiB (Per-Pixel), for example, slightly improves the one based on DeepLab-v3 on the 100-50 $(+2.0\%$ on all) and in the 100-5 $(+1.3\%)$ but its performance is slightly worse in the 100-10 $(-1.1\%)$ . In addition, Per-Pixel baselines demonstrate a large performance drop in the old classes, showing that, without employing a distillation strategy tailored to the architecture, it is difficult to maintain satisfactory results. + +# 4.6. Ablation Studies + +Mask2Former vs CoMFormer. In Tab. 3, we compare the proposed CoMFormer against Mask2Former [11]. In + +
MethodSemantic - mIoUPanoptic - PQ
1-100101-150all1-100101-150all
100-50 (2 tasks)
Mask2Former44.326.038.137.924.533.4
CoMFormer44.726.238.441.127.736.7
100-10 (6 tasks)
Mask2Former38.715.731.030.210.823.7
CoMFormer40.615.632.336.017.129.7
100-5 (11 tasks)
Mask2Former32.613.726.325.95.619.1
CoMFormer39.513.630.934.415.928.2
+ +Table 3. Comparison between CoMFormer and Mask2Former [11] in continual semantic and panoptic segmentation. We apply to Mask2Former our mask-based pseudo-labeling strategy and the adaptive distillation loss for a fair comparison. + +
100-50 (2 tasks)100-10 (6 tasks)
PSKDAD1-100101-150all1-100101-150all
---0.025.88.60.02.91.0
--✓28.415.324.023.09.118.4
✓--40.726.335.930.117.225.8
✓✓-40.528.136.430.217.626.0
✓-✓41.127.736.736.017.129.7
+ +particular, to fairly compare the two methods, we add our adaptive distillation loss and the pseudo-labeling strategy to Mask2Former. Considering semantic segmentation, CoMFormer achieves better results on the old classes (on average $3\%$ mIoU), while the two methods obtain comparable results on the new classes. On panoptic segmentation, CoMFormer exceeds Mask2Former in all the classes with a considerable gap (6 PQ on both old and new classes on average). We conclude that a vanilla MaskFormer architecture is not suited to continual learning because it suffers from forgetting, while CoMFormer outperforms it in any setting. + +Method components. We investigate the benefit of each method component in Tab. 4, reporting the results in PQ on the 100-50 and 100-10 settings in panoptic segmentation. Without applying any regularization technique, the performance on old classes drops to 0, denoting that it is important to deal with forgetting. Both the adaptive distillation loss (AD) and the pseudo-labeling strategy (PS) lead to improvement in the old classes when applied independently. However, we obtain the best results when combining them, especially considering the 100-10 scenario, where the old classes are improved by 5.8 PQ and the final performances of 3.9 PQ. Finally, the table shows that weighting the queries based on their semantic content is essential to improve performance: comparing a standard distillation + +Table 4. Ablation of the different key components of our CoM-Former architecture. We report the PQ on the 100-50 and 100-10 panoptic settings. PS: Pseudo-labeling; AD: adaptive distillation Eq. (3); KD: standard distillation Eq. (1) + +
Method100-50 (2 tasks)100-10 (6 tasks)
\( RQ \)\( SQ \)\( PQ \)\( RQ \)\( SQ \)\( PQ \)
MiB36.873.029.826.455.521.4
PLOP43.879.136.231.761.126.1
CoMFormer44.479.636.736.273.429.7
Joint47.679.039.547.679.039.5
+ +Table 5. Panoptic metrics analysis. We report Recognition Quality (RQ), Segmentation Quality (SQ), and Panoptic Quality (PQ) over all the classes after the last step. + +(KD Eq. (1)) with our adaptive distillation (AD Eq. (3)), we largely improve the performance on both settings: $+0.3$ PQ in 100-50 and $+3.7$ PQ in 100-10 considering all classes. + +Forgetting in CoMFormer. To investigate the reason for the forgetting, we considered separately the Recognition Quality (RQ) and the Segmentation Quality (SQ) in panoptic segmentation. RQ measures how well the model is able to classify segments, while SQ considers the precision of the segmentation. Note that, as reported in [28], SQ and RQ are not independent since SQ is computed only on segments correctly classified. The results on the 100-50 and 100-10 are reported in Tab. 5. The table clearly shows that, as hypothesized in Sec. 3.3, forgetting happens heavily on the classification ability while it is less marked on the segmentation ability of the model. CoMFormer, in fact, without providing regularization losses on the mask prediction, is able to match the segmentation ability of the Joint upper bound on the 100-50 and it gets close results on the 100-10 (-5.6). In addition, it is evident that classification ability is heavily affected by forgetting. CoMFormer has a gap with Joint of 3.2 and 11.4 RQ respectively in the 100-50 and 100-10. We note that, however, other baselines are more affected by forgetting, showcasing even larger gaps. + +# 5. Conclusion + +In this paper, we extend the task of continual learning beyond semantic segmentation by introducing the challenging and realistic continual panoptic segmentation setting. We propose CoMFormer, a strategy to address segmentation as a mask classification problem exploiting the properties of transformers to tackle both semantic and panoptic segmentation tasks. The extensive quantitative results highlight that CoMFormer outperforms all previous methods, both in semantic and panoptic segmentation, under a unified paradigm. In future work, we aim to extend the benchmark to instance segmentation to show that CoMFormer can be employed to address any continual segmentation task. + +# Acknowledgments. + +We gratefully acknowledge the HPC infrastructure and Team at IIT. Fabio Cermelli acknowledges travel support from ELISE (GA no 951847). This work was partly supported by ANR grant VISA DEEP (ANR-20-CHIA-0022). + +# References + +[1] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In ECCV, pages 213-229. Springer, 2020. (pages 2, 5). +[2] Fabio Cermelli, Dario Fontanel, Antonio Tavera, Marco Ciccone, and Barbara Caputo. Incremental learning in semantic segmentation from image labels. In CVPR, pages 4371-4381, 2022. (page 2). +[3] Fabio Cermelli, Antonino Geraci, Dario Fontanel, and Barbara Caputo. Modeling missing annotations for incremental learning in object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3700-3710, 2022. (pages 1, 2). +[4] Fabio Cermelli, Massimiliano Mancini, Samuel Rota Bulò, Elisa Ricci, and Barbara Caputo. Modeling the background for incremental learning in semantic segmentation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2020. (pages 1, 2, 3, 4, 5, 6). +[5] Fabio Cermelli, Massimiliano Mancini, Yongqin Xian, Zeynep Akata, and Barbara Caputo. Prototype-based incremental few-shot segmentation. In Proceedings of the British Machine Vision Conference (BMVC), 2021. (page 2). +[6] Arslan Chaudhry, Marc'Aurelio Ranzato, Marcus Rohrbach, and Mohamed Elhoseiny. Efficient lifelong learning with a-gem. In Proceedings of the International Conference on Learning Representations (ICLR), 2019. (page 2). +[7] Hongwei Chen, Laihui Ding, Fengqin Yao, Pengfei Ren, and Shengke Wang. Panoptic segmentation of UAV images with deformable convolution network and mask scoring. In Twelfth International Conference on Graphics and Image Processing (ICGIP 2020), volume 11720, pages 312-321. SPIE, 2021. (page 1). +[8] Liang-Chieh Chen, George Papandreou, Iasonas Kokkinos, Kevin Murphy, and Alan L. Yuille. Deeplab: Semantic image segmentation with deep convolutional nets, atrous convolution, and fully connected crfs. In IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 2018. (page 2). +[9] Liang-Chieh Chen, George Papandreou, Florian Schroff, and Hartwig Adam. Rethinking atrous convolution for semantic image segmentation. In arXiv preprint library, 2017. (pages 2, 5, 6, 7). +[10] Bowen Cheng, Maxwell D Collins, Yukun Zhu, Ting Liu, Thomas S Huang, Hartwig Adam, and Liang-Chieh Chen. Panoptic-deeplab: A simple, strong, and fast baseline for bottom-up panoptic segmentation. In CVPR, pages 12475-12485, 2020. (page 2). +[11] Bowen Cheng, Ishan Misra, Alexander G Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. In CVPR, 2022. (pages 2, 3, 5, 6, 7, 8). +[12] Bowen Cheng, Alexander G. Schwing, and Alexander Kirillov. Per-pixel classification is not all you need for semantic + +segmentation. In Adv. Neural Inform. Process. Syst., 2021. (pages 2, 3, 6). +[13] Lee R. Dice. Measures of the amount of ecologic association between species. Ecology, 1945. (page 5). +[14] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In Proceedings of the International Conference on Learning Representations (ICLR), 2021. (page 2). +[15] Arthur Douillard, Yifu Chen, Arnaud Dapogny, and Matthieu Cord. Plop: Learning without forgetting for continual semantic segmentation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2021. (pages 1, 2, 4, 5, 6). +[16] Arthur Douillard, Yifu Chen, Arnaud Dapogny, and Matthieu Cord. Tackling catastrophic forgetting and background shift in continual semantic segmentation. In arXiv preprint library, 2021. (page 2). +[17] Arthur Douillard, Matthieu Cord, Charles Ollion, Thomas Robert, and Eduardo Valle. Podnet: Pooled outputs distillation for small-tasks incremental learning. In Proceedings of the IEEE European Conference on Computer Vision (ECCV), 2020. (pages 1, 2). +[18] Arthur Douillard and Timothee Lesort. Continuum: Simple management of complex continual learning scenarios. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR) Workshop, 2021. (page 5). +[19] Arthur Douillard, Alexandre Rame, Guillaume Couairon, and Matthieu Cord. Dytox: Transformers for continual learning with dynamic token expansion. In CVPR, 2022. (pages 1, 2, 3). +[20] Mark Everingham, S. M. Ali Eslami, Luc Van Gool, Christopher K. I. Williams, John M. Winn, and Andrew Zisserman. The pascal visual object classes challenge: A retrospective. In International Journal of Computer Vision (IJCV), 2015. (page 5). +[21] Robert French. Catastrophic forgetting in connectionist networks. Trends in cognitive sciences, 1999. (pages 1, 2, 4). +[22] Jun Fu, Jing Liu, Hajjie Tian, Yong Li, Yongjun Bao, Zhiwei Fang, and Hanqing Lu. Dual attention network for scene segmentation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2019. (page 2). +[23] K. He, X. Zhang, S. Ren, and J. Sun. Deep residual learning for image recognition. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016. (page 6). +[24] Zilong Huang, Xinggang Wang, Yunchao Wei, Lichao Huang, Humphrey Shi, Wenyu Liu, and Thomas S. Huang. Cynet: Criss-cross attention for semantic segmentation. In IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 2020. (page 2). +[25] Loshchilov Ilya, Hutter Frank, et al. Decoupled weight decay regularization. Proceedings of ICLR, 2019. (page 6). + +[26] Jitesh Jain, Anukriti Singh, Nikita Orlov, Zilong Huang, Jiachen Li, Steven Walton, and Humphrey Shi. Semask: Semantically masked transformers for semantic segmentation. arXiv preprint arXiv:2112.12782, 2021. (page 2). +[27] Alexander Kirillov, Ross Girshick, Kaiming He, and Piotr Dólar. Panoptic feature pyramid networks. In CVPR, pages 6399-6408, 2019. (page 2). +[28] Alexander Kirillov, Kaiming He, Ross Girshick, Carsten Rother, and Piotr Dólár. Panoptic segmentation. In CVPR, pages 9404-9413, 2019. (pages 2, 5, 8). +[29] James Kirkpatrick, Razvan Pascanu, Neil Rabinowitz, Joel Veness, Guillaume Desjardins, Andrei A. Rusu, Kieran Milan, John Quan, Tiago Ramalho, Agnieszka Grabska-Barwinska, Demis Hassabis, Claudia Clopath, Dharshan Kumaran, and Raia Hadsell. Overcoming catastrophic forgetting in neural networks. Proceedings of the National Academy of Sciences, 2017. (pages 1, 2). +[30] Harold W. Kuhn. The hungarian method for the assignment problem. Naval Research Logistics Quarterly, 1955. (page 5). +[31] Duo Li, Guimei Cao, Yunlu Xu, Zhanzhan Cheng, and Yi Niu. Technical report for iccv 2021 challenge sslad-track3b: Transformers are better continual learners. In ICCV 2021 Challenge SSLAD-Track3B, 2021. (page 3). +[32] Yanwei Li, Hengshuang Zhao, Xiaojuan Qi, Yukang Chen, Lu Qi, Liwei Wang, Zeming Li, Jian Sun, and Jiaya Jia. Fully convolutional networks for panoptic segmentation with point-based supervision. IEEE TPAMI, 2022. (page 2). +[33] Z. Li and D. Hoiem. Learning without forgetting. Proceedings of the IEEE European Conference on Computer Vision (ECCV), 2016. (pages 1, 2). +[34] Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dólár. Focal loss for dense object detection. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), 2017. (page 5). +[35] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and Lawrence Zitnick. Microsoft coco: Common objects in context. In Proceedings of the IEEE European Conference on Computer Vision (ECCV), 2014. (page 5). +[36] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), 2021. (page 2). +[37] J. Long, E. Shelhamer, and T. Darrell. Fully convolutional networks for semantic segmentation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2015. (page 2). +[38] Andrea Maracani, Umberto Michieli, Marco Toldo, and Pietro Zanuttigh. Recall: Replay-based continual learning in semantic segmentation. In ICCV, pages 7026-7035, 2021. (page 2). +[39] Umberto Michieli and Pietro Zanuttigh. Incremental learning techniques for semantic segmentation. In Proceedings of the IEEE International Conference on Computer Vision (ICCV) Workshop, 2019. (pages 1, 2, 4). + +[40] Umberto Michieli and Pietro Zanuttigh. Continual semantic segmentation via repulsion-attraction of sparse and disentangled latent representations. In CVPR, pages 1114-1124, 2021. (page 2). +[41] Gerhard Neuhold, Tobias Ollmann, Samuel Rota Bulò, and Peter Kontschieder. The mapillary vistas dataset for semantic understanding of street scenes. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), 2017. (page 1). +[42] Lorenzo Porzi, Samuel Rota Bulò, Aleksander Colovic, and Peter Kontschieder. Seamless scene segmentation. In CVPR, June 2019. (page 2). +[43] Sylvestre-Alvise Rebuffi, Alexander Kolesnikov, Georg Sperl, and Christoph H. Lampert. icarl: Incremental classifier and representation learning. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2017. (pages 1, 2). +[44] Anthony Robins. Catastrophic forgetting, rehearsal and pseudorehearsal. Connection Science, 1995. (page 2). +[45] Andrei A Rusu, Neil C Rabinowitz, Guillaume Desjardins, Hubert Soyer, James Kirkpatrick, Koray Kavukcuoglu, Razvan Pascanu, and Raia Hadsell. Progressive neural networks. arXiv preprint library, 2016. (pages 1, 2). +[46] Robin Strudel, Ricardo Garcia, Ivan Laptev, and Cordelia Schmid. Segmenter: Transformer for semantic segmentation. In ICCV, pages 7262-7272, 2021. (page 2). +[47] Sebastian Thrun. Lifelong learning algorithms. In Springer Learning to Learn, 1998. (page 2). +[48] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Hervé Jégou. Training data-efficient image transformers and distillation through attention. In International Conference on Machine Learning (ICML), 2021. (page 2). +[49] Hugo Touvron, Matthieu Cord, Alexandre Sablayrolles, Gabriel Synnaeve, and Hervé Jégou. Going deeper with image transformers. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), 2021. (page 2). +[50] Huiyu Wang, Yukun Zhu, Hartwig Adam, Alan Yuille, and Liang-Chieh Chen. Max-deeplab: End-to-end panoptic segmentation with mask transformers. In CVPR, pages 5463-5474, 2021. (page 2). +[51] Huiyu Wang, Yukun Zhu, Bradley Green, Hartwig Adam, Alan Yuille, and Liang-Chieh Chen. Axial-deeplab: Standalone axial-attention for panoptic segmentation. In ECCV, pages 108–126. Springer, 2020. (page 2). +[52] Zifeng Wang, Zizhao Zhang, Sayna Ebrahimi, Ruoxi Sun, Han Zhang, Chen-Yu Lee, Xiaoqi Ren, Guolong Su, Vincent Perot, Jennifer Dy, et al. Dualprompt: Complementary prompting for rehearsal-free continual learning. ECCV, 2022. (page 2). +[53] Zifeng Wang, Zizhao Zhang, Chen-Yu Lee, Han Zhang, Ruoxi Sun, Xiaoqi Ren, Guolong Su, Vincent Perot, Jennifer Dy, and Tomas Pfister. Learning to prompt for continual learning. In CVPR, pages 139-149, 2022. (page 2). +[54] Enze Xie, Wenhai Wang, Zhiding Yu, Anima Anandkumar, Jose M Alvarez, and Ping Luo. Segformer: Simple and efficient design for semantic segmentation with transformers. Adv. Neural Inform. Process. Syst., 34, 2021. (page 2). + +[55] Shipeng Yan, Jiangwei Xie, and Xuming He. Der: Dynamically expandable representation for class incremental learning. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2021. (page 2). +[56] Qihang Yu, Huiyu Wang, Siyuan Qiao, Maxwell Collins, Yukun Zhu, Hartwig Adam, Alan Yuille, and Liang-Chieh Chen. k-means mask transformer. In European Conference on Computer Vision, pages 288-307. Springer, 2022. (page 2). +[57] Yuhui Yuan, Xilin Chen, and Jingdong Wang. Object-contextual representations for semantic segmentation. In Proceedings of the IEEE European Conference on Computer Vision (ECCV), 2020. (page 2). +[58] Yuhui Yuan and Jingdong Wang. Ocnet: Object context network for scene parsing. In arXiv preprint library, 2018. (page 2). +[59] Chang-Bin Zhang, Jia-Wen Xiao, Xialei Liu, Ying-Cong Chen, and Ming-Ming Cheng. Representation compensation networks for continual semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7053-7064, 2022. (pages 1, 2, 5, 6, 7). +[60] H. Zhao, J. Shi, X. Qi, X. Wang, and J. Jia. Pyramid scene parsing network. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2017. (page 2). +[61] Hengshuang Zhao, Yi Zhang, Shu Liu, Jianping Shi, Chen Change Loy, Dahua Lin, and Jiaya Jia. Pisanet: Pointwise spatial attention network for scene parsing. In Proceedings of the IEEE European Conference on Computer Vision (ECCV), 2018. (page 2). +[62] Bolei Zhou, Hang Zhao, Xavier Puig, Sanja Fidler, Adela Barriuso, and Antonio Torralba. Scene parsing through ade20k dataset. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2017. (page 5). \ No newline at end of file diff --git a/2023/CoMFormer_ Continual Learning in Semantic and Panoptic Segmentation/images.zip b/2023/CoMFormer_ Continual Learning in Semantic and Panoptic Segmentation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..84d6ac8ca3e9645659de552f7c3b480ce0e6e4f4 --- /dev/null +++ b/2023/CoMFormer_ Continual Learning in Semantic and Panoptic Segmentation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1aaa4306e3d41e81504caa2f529096bf7473cd77a9534787d09d79e14852a838 +size 592703 diff --git a/2023/CoMFormer_ Continual Learning in Semantic and Panoptic Segmentation/layout.json b/2023/CoMFormer_ Continual Learning in Semantic and Panoptic Segmentation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..79ce9430a80258ae7f5281112b5f4c77130fe414 --- /dev/null +++ b/2023/CoMFormer_ Continual Learning in Semantic and Panoptic Segmentation/layout.json @@ -0,0 +1,9493 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 67, + 103, + 526, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 103, + 526, + 121 + ], + "spans": [ + { + "bbox": [ + 67, + 103, + 526, + 121 + ], + "type": "text", + "content": "CoMFormer: Continual Learning in Semantic and Panoptic Segmentation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 158, + 142, + 436, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 158, + 142, + 436, + 156 + ], + "spans": [ + { + "bbox": [ + 158, + 142, + 436, + 156 + ], + "type": "text", + "content": "Fabio Cermelli " + }, + { + "bbox": [ + 158, + 142, + 436, + 156 + ], + "type": "inline_equation", + "content": "^{1,2*}" + }, + { + "bbox": [ + 158, + 142, + 436, + 156 + ], + "type": "text", + "content": ", Matthieu Cord " + }, + { + "bbox": [ + 158, + 142, + 436, + 156 + ], + "type": "inline_equation", + "content": "^{3,4}" + }, + { + "bbox": [ + 158, + 142, + 436, + 156 + ], + "type": "text", + "content": ", Arthur Douillard " + }, + { + "bbox": [ + 158, + 142, + 436, + 156 + ], + "type": "inline_equation", + "content": "^{3\\dagger}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 83, + 157, + 509, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 83, + 157, + 509, + 171 + ], + "spans": [ + { + "bbox": [ + 83, + 157, + 509, + 171 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 83, + 157, + 509, + 171 + ], + "type": "text", + "content": "Politecnico di Torino, " + }, + { + "bbox": [ + 83, + 157, + 509, + 171 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 83, + 157, + 509, + 171 + ], + "type": "text", + "content": "Italian Institute of Technology, " + }, + { + "bbox": [ + 83, + 157, + 509, + 171 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 83, + 157, + 509, + 171 + ], + "type": "text", + "content": " Sorbonne Université, " + }, + { + "bbox": [ + 83, + 157, + 509, + 171 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 83, + 157, + 509, + 171 + ], + "type": "text", + "content": " valeo.ai" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 61, + 174, + 526, + 184 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 174, + 526, + 184 + ], + "spans": [ + { + "bbox": [ + 61, + 174, + 526, + 184 + ], + "type": "text", + "content": "fabio.cermelli@polito.it, matthieu.cord@sorbonne-universite.fr, ar.douillard@gmail.com" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "spans": [ + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 238, + 290, + 490 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 238, + 290, + 490 + ], + "spans": [ + { + "bbox": [ + 46, + 238, + 290, + 490 + ], + "type": "text", + "content": "Continual learning for segmentation has recently seen increasing interest. However, all previous works focus on narrow semantic segmentation and disregard panoptic segmentation, an important task with real-world impacts. In this paper, we present the first continual learning model capable of operating on both semantic and panoptic segmentation. Inspired by recent transformer approaches that consider segmentation as a mask-classification problem, we design CoMFormer. Our method carefully exploits the properties of transformer architectures to learn new classes over time. Specifically, we propose a novel adaptive distillation loss along with a mask-based pseudo-labeling technique to effectively prevent forgetting. To evaluate our approach, we introduce a novel continual panoptic segmentation benchmark on the challenging ADE20K dataset. Our CoMFormer outperforms all the existing baselines by forgetting less old classes but also learning more effectively new classes. In addition, we also report an extensive evaluation in the large-scale continual semantic segmentation scenario showing that CoMFormer also significantly outperforms state-of-the-art methods." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 512, + 128, + 525 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 512, + 128, + 525 + ], + "spans": [ + { + "bbox": [ + 47, + 512, + 128, + 525 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 532, + 287, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 532, + 287, + 651 + ], + "spans": [ + { + "bbox": [ + 46, + 532, + 287, + 651 + ], + "type": "text", + "content": "Image segmentation is a fundamental computer vision problem that enables machines to assign an image's pixels to discrete segments. Multiple segmentation tasks have been defined depending on the segments definitions. Semantic segmentation clusters pixels by classes, merging in a single segment pixels belonging to instances of the same class. Panoptic segmentation assigns to every pixel a semantic class while separating different instances into different segments. This latter kind of segmentation has real-world impacts in autonomous robots and vehicles [7, 41]." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 652, + 287, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 652, + 287, + 676 + ], + "spans": [ + { + "bbox": [ + 47, + 652, + 287, + 676 + ], + "type": "text", + "content": "Despite tremendous progress in image segmentation, the current approaches are trained on a static dataset with a pre" + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 309, + 220, + 419, + 296 + ], + "blocks": [ + { + "bbox": [ + 307, + 211, + 392, + 219 + ], + "lines": [ + { + "bbox": [ + 307, + 211, + 392, + 219 + ], + "spans": [ + { + "bbox": [ + 307, + 211, + 392, + 219 + ], + "type": "text", + "content": "Step t-1 - New Class: Car" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 309, + 220, + 419, + 296 + ], + "lines": [ + { + "bbox": [ + 309, + 220, + 419, + 296 + ], + "spans": [ + { + "bbox": [ + 309, + 220, + 419, + 296 + ], + "type": "image", + "image_path": "b8d3347a7632053e053c6c4995d36b1c1c99a69545bc93bac49c7dac066977d0.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 329, + 297, + 351, + 304 + ], + "lines": [ + { + "bbox": [ + 329, + 297, + 351, + 304 + ], + "spans": [ + { + "bbox": [ + 329, + 297, + 351, + 304 + ], + "type": "text", + "content": "Image" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 427, + 297, + 485, + 304 + ], + "lines": [ + { + "bbox": [ + 427, + 297, + 485, + 304 + ], + "spans": [ + { + "bbox": [ + 427, + 297, + 485, + 304 + ], + "type": "text", + "content": "CoMFormer (t-1)" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 420, + 221, + 499, + 283 + ], + "blocks": [ + { + "bbox": [ + 420, + 221, + 499, + 283 + ], + "lines": [ + { + "bbox": [ + 420, + 221, + 499, + 283 + ], + "spans": [ + { + "bbox": [ + 420, + 221, + 499, + 283 + ], + "type": "image", + "image_path": "9967a26e5908b0367f68e04b1f5dce02f2397f78defc3567ae5a1a9360a2e5f0.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 503, + 220, + 545, + 297 + ], + "blocks": [ + { + "bbox": [ + 503, + 220, + 545, + 297 + ], + "lines": [ + { + "bbox": [ + 503, + 220, + 545, + 297 + ], + "spans": [ + { + "bbox": [ + 503, + 220, + 545, + 297 + ], + "type": "image", + "image_path": "0dff6aec1f26369fcf162fe0e066f1246e89d7c6f19a07a4cf9e31a8f5eba346.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 504, + 297, + 542, + 304 + ], + "lines": [ + { + "bbox": [ + 504, + 297, + 542, + 304 + ], + "spans": [ + { + "bbox": [ + 504, + 297, + 542, + 304 + ], + "type": "text", + "content": "Predictions" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 308, + 330, + 373, + 385 + ], + "blocks": [ + { + "bbox": [ + 308, + 330, + 373, + 385 + ], + "lines": [ + { + "bbox": [ + 308, + 330, + 373, + 385 + ], + "spans": [ + { + "bbox": [ + 308, + 330, + 373, + 385 + ], + "type": "image", + "image_path": "249792d01ea386ca4a040a0c03d29c59e249569a2a81740d16f2510f1f498045.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 329, + 388, + 351, + 396 + ], + "lines": [ + { + "bbox": [ + 329, + 388, + 351, + 396 + ], + "spans": [ + { + "bbox": [ + 329, + 388, + 351, + 396 + ], + "type": "text", + "content": "Image" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 375, + 388, + 416, + 396 + ], + "lines": [ + { + "bbox": [ + 375, + 388, + 416, + 396 + ], + "spans": [ + { + "bbox": [ + 375, + 388, + 416, + 396 + ], + "type": "text", + "content": "Annotations" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 375, + 325, + 499, + 383 + ], + "blocks": [ + { + "bbox": [ + 307, + 317, + 397, + 325 + ], + "lines": [ + { + "bbox": [ + 307, + 317, + 397, + 325 + ], + "spans": [ + { + "bbox": [ + 307, + 317, + 397, + 325 + ], + "type": "text", + "content": "Step t - New Class: Person" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 375, + 325, + 499, + 383 + ], + "lines": [ + { + "bbox": [ + 375, + 325, + 499, + 383 + ], + "spans": [ + { + "bbox": [ + 375, + 325, + 499, + 383 + ], + "type": "image", + "image_path": "4b01263540676814d6bda30c505c2ef988424ff64e4e22e40641f9f3af7b13c2.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 433, + 388, + 482, + 396 + ], + "lines": [ + { + "bbox": [ + 433, + 388, + 482, + 396 + ], + "spans": [ + { + "bbox": [ + 433, + 388, + 482, + 396 + ], + "type": "text", + "content": "CoMFormer (t)" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 503, + 310, + 542, + 389 + ], + "blocks": [ + { + "bbox": [ + 503, + 310, + 542, + 389 + ], + "lines": [ + { + "bbox": [ + 503, + 310, + 542, + 389 + ], + "spans": [ + { + "bbox": [ + 503, + 310, + 542, + 389 + ], + "type": "image", + "image_path": "756eccefecc29ee7a7d99a8af6be340e0df1e397300f4c2b43762d9d604c6c58.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 504, + 390, + 542, + 397 + ], + "lines": [ + { + "bbox": [ + 504, + 390, + 542, + 397 + ], + "spans": [ + { + "bbox": [ + 504, + 390, + 542, + 397 + ], + "type": "text", + "content": "Predictions" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 304, + 406, + 547, + 483 + ], + "lines": [ + { + "bbox": [ + 304, + 406, + 547, + 483 + ], + "spans": [ + { + "bbox": [ + 304, + 406, + 547, + 483 + ], + "type": "text", + "content": "Figure 1. Illustration of our model, CoMFormer, operating in continual segmentation. Relying on the mask classification paradigm, it is able to cope with both continual semantic and panoptic segmentation without any modification by predicting masks for both old (e.g. car in red) and new (e.g. person in green) classes. The figure reports two classes and no \"stuff\" (e.g. road, building) only for illustration purposes." + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_caption" + } + ], + "index": 24 + }, + { + "bbox": [ + 304, + 496, + 545, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 496, + 545, + 592 + ], + "spans": [ + { + "bbox": [ + 304, + 496, + 545, + 592 + ], + "type": "text", + "content": "defined set of classes. Whenever an update of the model is required to fit new classes, the common solution is to train a model from scratch on the union of the old and new class data. A computationally more efficient solution would be to fine-tune the existing model solely on the new class data. Unfortunately, this approach would cause a catastrophic forgetting [21] of the old classes on which the model performance would be extremely degraded." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 304, + 594, + 546, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 546, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 546, + 714 + ], + "type": "text", + "content": "The problem of updating the knowledge of the model over time is typically referred as continual learning. It has been traditionally studied in the context of image classification [17, 19, 29, 33, 43, 45] and is gaining attention on the segmentation task [3, 4, 15, 39, 59] due to the more realistic applications and the additional challenges it introduces, such as the background shift [4]. However, current state-of-the-art methods mainly focus on semantic segmentation and are not designed to work in other segmentation tasks, strongly limiting their application in the real world." + } + ] + } + ], + "index": 28 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 494, + 35 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 494, + 35 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 494, + 35 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 57, + 683, + 259, + 693 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 683, + 259, + 693 + ], + "spans": [ + { + "bbox": [ + 57, + 683, + 259, + 693 + ], + "type": "text", + "content": "*Work done during the visiting period at Sorbonne Université." + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 60, + 693, + 283, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 60, + 693, + 283, + 703 + ], + "spans": [ + { + "bbox": [ + 60, + 693, + 283, + 703 + ], + "type": "inline_equation", + "content": "^{\\dagger}" + }, + { + "bbox": [ + 60, + 693, + 283, + 703 + ], + "type": "text", + "content": "Work done at Sorbonne Université, currently affiliated to DeepMind." + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 60, + 703, + 231, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 60, + 703, + 231, + 712 + ], + "spans": [ + { + "bbox": [ + 60, + 703, + 231, + 712 + ], + "type": "text", + "content": "" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3010" + } + ] + } + ], + "index": 33 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 275 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 275 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 275 + ], + "type": "text", + "content": "In this paper, we design the first method operating in both continual semantic and panoptic segmentation, as illustrated in Fig. 1. Our method, CoMFormer (Continual MaskFormer), takes inspiration from recent transformer architectures [11, 12], approaching segmentation as a mask classification problem. Instead of predicting a class probability for each pixel, as in previous semantic segmentation works [9, 37], it predicts a set of binary masks, each associated with a single class prediction, effectively addressing both segmentation tasks without any modification in the training architecture and procedure. Differently from previous works [11, 12], however, CoMFormer forces the output binary masks to be mutually exclusive to one another: a pixel can only be predicted by a single binary mask to prevent having several masks classifying the same pixel with different classes. This behavior is crucial in continual learning to reduce the interference among old and new classes." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 275, + 289, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 275, + 289, + 418 + ], + "spans": [ + { + "bbox": [ + 46, + 275, + 289, + 418 + ], + "type": "text", + "content": "Furthermore, CoMFormer introduces a novel adaptive distillation loss to alleviate forgetting. It enforces consistency of the model's classification predictions across learning steps only when it is useful to remember old classes, ensuring a better tradeoff between rigidity (not forgetting old classes) and plasticity (learning efficiently new classes). Finally, since at each training iteration the dataset reports annotations only for the current classes, we design a mask-based pseudo-labeling technique to generate annotations for the old classes, effectively alleviating forgetting. To reduce the noise, we consider the prediction confidence and we avoid interference with ground-truth annotations." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 419, + 289, + 503 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 419, + 289, + 503 + ], + "spans": [ + { + "bbox": [ + 46, + 419, + 289, + 503 + ], + "type": "text", + "content": "We validate CoMFormer on both continual segmentation tasks. For panoptic segmentation, we define a new benchmark relying on the challenging ADE20K where we demonstrate that CoMFormer largely outperforms all previous baselines. On semantic segmentation, we show that CoMFormer outperforms the existing state-of-the-art methods on every setting of the large-scale ADE20K benchmark." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 59, + 504, + 286, + 516 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 504, + 286, + 516 + ], + "spans": [ + { + "bbox": [ + 59, + 504, + 286, + 516 + ], + "type": "text", + "content": "To sum up, the contributions of this paper are as follows:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 58, + 517, + 286, + 635 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 59, + 517, + 286, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 517, + 286, + 552 + ], + "spans": [ + { + "bbox": [ + 59, + 517, + 286, + 552 + ], + "type": "text", + "content": "- We introduce continual panoptic segmentation which has real-world impacts in addition to being significantly more challenging than previous benchmarks." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 58, + 552, + 286, + 600 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 552, + 286, + 600 + ], + "spans": [ + { + "bbox": [ + 58, + 552, + 286, + 600 + ], + "type": "text", + "content": "- We propose CoMFormer to tackle both continual panoptic and semantic segmentation. To avoid forgetting, we design a novel adaptive distillation and an efficient mask-based pseudo-labeling strategy." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 58, + 601, + 286, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 601, + 286, + 635 + ], + "spans": [ + { + "bbox": [ + 58, + 601, + 286, + 635 + ], + "type": "text", + "content": "- Through extensive quantitative and qualitative benchmarks, we showcase the state-of-the-art performance of our model on both continual segmentation tasks." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 47, + 643, + 138, + 656 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 643, + 138, + 656 + ], + "spans": [ + { + "bbox": [ + 47, + 643, + 138, + 656 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "type": "text", + "content": "Semantic and Panoptic Segmentation. The two tasks have been traditionally treated separately, with specialized architectures proposed for either one or the other task, without interoperability. Semantic segmentation has been tradi" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 72, + 547, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 373 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 373 + ], + "type": "text", + "content": "tionally addressed as a per-pixel classification task. Fully convolutional network [37] dominated the field by aggregating long-range dependencies in the features map [8, 9, 60] and exploiting contextual information [22, 24, 57, 58, 61]. Recently, transformers [26, 46, 54] are replacing convolutions by integrating long-range dependencies at every layer. Panoptic segmentation [28] has been proposed to unify semantic and instance segmentations. Initially, methods proposed to combine each task-specific architectures [10, 27, 32, 42] or defined new specialized architectures and objective functions for the panoptic task [1, 50, 51], drifting from a general solution for all the segmentation tasks. Recently, methods addressing segmentation as a mask classification problem [11, 12, 50, 56] have been proposed, introducing a transformer architecture able to solve multiple tasks at once. MaskFormer [12] was the first to propose a single architecture to address both panoptic and semantic segmentation. Mask2Former [11] improves it by adopting multi-scale features, masked attention and optimization tricks. Concurrently, kMaX-DeepLab [56] extended [50] proposing to reformulate the cross-attention as a clustering process. Despite their effectiveness in standard training setting, these methods suffer from catastrophic forgetting. In this work, we aim to extend their capability and we propose CoMFormer to learn to segment new classes over time." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 380, + 547, + 633 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 380, + 547, + 633 + ], + "spans": [ + { + "bbox": [ + 304, + 380, + 547, + 633 + ], + "type": "text", + "content": "Continual Segmentation. Continual learning is a longstanding field addressing the problem of learning new knowledge over time while avoiding catastrophic forgetting [21, 44, 47]. Traditionally, it has been studied in the context of image classification [6, 17, 19, 29, 33, 43, 45, 55] but there is a growing interest in its applications in segmentation [2-5, 15, 16, 38-40, 59]. Continual semantic segmentation introduces additional challenges, as pointed out by [4]. In particular, considering the task as a per-pixel classification problem, catastrophic forgetting is exacerbated by the background shift, where old classes pixels are treated as background in following training steps. [4] proposed a revision of the standard knowledge distillation framework to deal with it. [17, 39] introduced a distillation loss preserving the representation at feature level while [38] proposed an approach using old classes' examples to alleviate forgetting. Recently, [59] presented a technique that decouples learning the representation of the new classes while freezing the representation for the old ones. Differently from previous work, we extend the benchmarks to solve both semantic and the more challenging panoptic segmentation task." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 641, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 641, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 641, + 547, + 715 + ], + "type": "text", + "content": "Continual Learning with Transformers. Recent progresses made using transformers for computer vision [14, 36, 48, 49] attracted the attention of the continual learning community [19, 52, 53]. In particular, DyTox [19] proposed to specialize the architecture on each task using a different task-specific token. Learning-to-Prompt [53] stores a" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 758 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 758 + ], + "type": "text", + "content": "3011" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 65, + 70, + 531, + 232 + ], + "blocks": [ + { + "bbox": [ + 65, + 70, + 531, + 232 + ], + "lines": [ + { + "bbox": [ + 65, + 70, + 531, + 232 + ], + "spans": [ + { + "bbox": [ + 65, + 70, + 531, + 232 + ], + "type": "image", + "image_path": "2c32f8ab18b9358039da99d591fe4e93e53421fcb9015a9bd11c35250c11d980.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 239, + 547, + 297 + ], + "lines": [ + { + "bbox": [ + 46, + 239, + 547, + 297 + ], + "spans": [ + { + "bbox": [ + 46, + 239, + 547, + 297 + ], + "type": "text", + "content": "Figure 2. Overview of CoMFormer. At learning step t, both the current (bottom) and old (top) models output, for each input query, a pair made of a class probability distribution and a binary mask. While learning the new class (person), we regularize the current model to prevent forgetting the old classes (car) by: (i) generating pseudo-labels for the old classes present in the image and (ii) by applying an adaptive distillation loss that reweights the contribution of each output based on the information it carries on the old classes. " + }, + { + "bbox": [ + 46, + 239, + 547, + 297 + ], + "type": "inline_equation", + "content": "\\psi" + }, + { + "bbox": [ + 46, + 239, + 547, + 297 + ], + "type": "text", + "content": " indicates the dot product between the mask embeddings " + }, + { + "bbox": [ + 46, + 239, + 547, + 297 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_{mask}" + }, + { + "bbox": [ + 46, + 239, + 547, + 297 + ], + "type": "text", + "content": " and the pixel embeddings " + }, + { + "bbox": [ + 46, + 239, + 547, + 297 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_{pixel}" + }, + { + "bbox": [ + 46, + 239, + 547, + 297 + ], + "type": "text", + "content": " followed by the softmax activation." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 305, + 288, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 305, + 288, + 365 + ], + "spans": [ + { + "bbox": [ + 46, + 305, + 288, + 365 + ], + "type": "text", + "content": "pool of prompts that are employed to condition the whole forward execution of the patch tokens. To the best of our knowledge, transformers have been evaluated only in continual learning for image classification [19, 31] and we are the first to use them in the continual segmentation task." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 374, + 128, + 387 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 374, + 128, + 387 + ], + "spans": [ + { + "bbox": [ + 47, + 374, + 128, + 387 + ], + "type": "text", + "content": "3. CoMFormer" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 394, + 159, + 406 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 394, + 159, + 406 + ], + "spans": [ + { + "bbox": [ + 47, + 394, + 159, + 406 + ], + "type": "text", + "content": "3.1. Problem Definition" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 413, + 287, + 532 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 413, + 287, + 532 + ], + "spans": [ + { + "bbox": [ + 46, + 413, + 287, + 532 + ], + "type": "text", + "content": "The goal of image segmentation is to learn a model able to (i) partition the image into a set of N regions represented by binary masks and (ii) produce a class-probability distribution associated with each region. The differences among segmentation tasks rely on the semantics of the masks: semantic segmentation groups all the pixels of a class while panoptic segmentation distinguishes different object instances. In the following, we provide a general formulation for the two tasks since they only differ in the construction of the binary masks in the dataset." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 533, + 288, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 533, + 288, + 665 + ], + "spans": [ + { + "bbox": [ + 46, + 533, + 288, + 665 + ], + "type": "text", + "content": "Continual segmentation aims to train the model in multiple learning steps " + }, + { + "bbox": [ + 46, + 533, + 288, + 665 + ], + "type": "inline_equation", + "content": "t = 1, \\ldots, T" + }, + { + "bbox": [ + 46, + 533, + 288, + 665 + ], + "type": "text", + "content": ", introducing at every step a new set of classes. Formally, during the learning step " + }, + { + "bbox": [ + 46, + 533, + 288, + 665 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 533, + 288, + 665 + ], + "type": "text", + "content": " a dataset " + }, + { + "bbox": [ + 46, + 533, + 288, + 665 + ], + "type": "inline_equation", + "content": "D_{t}" + }, + { + "bbox": [ + 46, + 533, + 288, + 665 + ], + "type": "text", + "content": " consisting of a collection of image and label pairs is provided. The label of each image takes the form of a set of " + }, + { + "bbox": [ + 46, + 533, + 288, + 665 + ], + "type": "inline_equation", + "content": "N^{gt}" + }, + { + "bbox": [ + 46, + 533, + 288, + 665 + ], + "type": "text", + "content": " ground truth segments " + }, + { + "bbox": [ + 46, + 533, + 288, + 665 + ], + "type": "inline_equation", + "content": "z^{gt}\\{(c_i^{gt}, m_i^{gt}) | c_i^{gt} \\in \\mathcal{K}^t, m_i^{gt} \\in \\{0,1\\}^{H \\times W}\\}_{i=1}^{N^{gt}}" + }, + { + "bbox": [ + 46, + 533, + 288, + 665 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 46, + 533, + 288, + 665 + ], + "type": "inline_equation", + "content": "c_i^{gt}" + }, + { + "bbox": [ + 46, + 533, + 288, + 665 + ], + "type": "text", + "content": " is the ground-truth class and " + }, + { + "bbox": [ + 46, + 533, + 288, + 665 + ], + "type": "inline_equation", + "content": "m_i^{gt}" + }, + { + "bbox": [ + 46, + 533, + 288, + 665 + ], + "type": "text", + "content": " is binary mask " + }, + { + "bbox": [ + 46, + 533, + 288, + 665 + ], + "type": "inline_equation", + "content": "m^{gt}" + }, + { + "bbox": [ + 46, + 533, + 288, + 665 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 46, + 533, + 288, + 665 + ], + "type": "inline_equation", + "content": "\\mathcal{K}^t" + }, + { + "bbox": [ + 46, + 533, + 288, + 665 + ], + "type": "text", + "content": " is the set of classes introduced at step " + }, + { + "bbox": [ + 46, + 533, + 288, + 665 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 533, + 288, + 665 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 533, + 288, + 665 + ], + "type": "inline_equation", + "content": "H, W" + }, + { + "bbox": [ + 46, + 533, + 288, + 665 + ], + "type": "text", + "content": " are the height and width of the images. The goal of training step " + }, + { + "bbox": [ + 46, + 533, + 288, + 665 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 533, + 288, + 665 + ], + "type": "text", + "content": " is to learn a model able to predict segments for all the seen classes " + }, + { + "bbox": [ + 46, + 533, + 288, + 665 + ], + "type": "inline_equation", + "content": "\\mathcal{K}^{1:t}" + }, + { + "bbox": [ + 46, + 533, + 288, + 665 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 665, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 288, + 714 + ], + "type": "text", + "content": "We note that the dataset only contains annotations for the new classes " + }, + { + "bbox": [ + 46, + 665, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{K}^t" + }, + { + "bbox": [ + 46, + 665, + 288, + 714 + ], + "type": "text", + "content": " while not reporting ground-truth segments for old " + }, + { + "bbox": [ + 46, + 665, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{K}^{1:t - 1}" + }, + { + "bbox": [ + 46, + 665, + 288, + 714 + ], + "type": "text", + "content": " or future " + }, + { + "bbox": [ + 46, + 665, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{K}^{t + 1:T}" + }, + { + "bbox": [ + 46, + 665, + 288, + 714 + ], + "type": "text", + "content": " classes. Moreover, differently from previous works [4], old and future class segment" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 305, + 545, + 331 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 305, + 545, + 331 + ], + "spans": [ + { + "bbox": [ + 305, + 305, + 545, + 331 + ], + "type": "text", + "content": "annotations are simply absent during training rather than being collapsed into an artificial background class." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 341, + 451, + 353 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 341, + 451, + 353 + ], + "spans": [ + { + "bbox": [ + 305, + 341, + 451, + 353 + ], + "type": "text", + "content": "3.2. CoMFormer Architecture" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 361, + 545, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 361, + 545, + 506 + ], + "spans": [ + { + "bbox": [ + 304, + 361, + 545, + 506 + ], + "type": "text", + "content": "To solve semantic and panoptic segmentation tasks within a single method, we take inspiration from Mask-Former architectures [11, 12] considering segmentation a mask classification problem. It consists in predicting for each image a set of " + }, + { + "bbox": [ + 304, + 361, + 545, + 506 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 361, + 545, + 506 + ], + "type": "text", + "content": " pairs made of a class prediction and a binary mask. Formally, the output of a CoMFormer is " + }, + { + "bbox": [ + 304, + 361, + 545, + 506 + ], + "type": "inline_equation", + "content": "z = \\{(p_i, m_i)\\}_{i=1}^N" + }, + { + "bbox": [ + 304, + 361, + 545, + 506 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 361, + 545, + 506 + ], + "type": "inline_equation", + "content": "p_i" + }, + { + "bbox": [ + 304, + 361, + 545, + 506 + ], + "type": "text", + "content": " represents a class probability distribution over the seen classes, i.e. " + }, + { + "bbox": [ + 304, + 361, + 545, + 506 + ], + "type": "inline_equation", + "content": "p_i \\in \\Delta^{K^t + 1}" + }, + { + "bbox": [ + 304, + 361, + 545, + 506 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 304, + 361, + 545, + 506 + ], + "type": "inline_equation", + "content": "K^t = |\\mathcal{K}^{1:t}|" + }, + { + "bbox": [ + 304, + 361, + 545, + 506 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 361, + 545, + 506 + ], + "type": "inline_equation", + "content": "m_i" + }, + { + "bbox": [ + 304, + 361, + 545, + 506 + ], + "type": "text", + "content": " is a binary mask such that " + }, + { + "bbox": [ + 304, + 361, + 545, + 506 + ], + "type": "inline_equation", + "content": "m_i \\in [0, 1]^{H \\times W}" + }, + { + "bbox": [ + 304, + 361, + 545, + 506 + ], + "type": "text", + "content": ". We note that " + }, + { + "bbox": [ + 304, + 361, + 545, + 506 + ], + "type": "inline_equation", + "content": "p_i" + }, + { + "bbox": [ + 304, + 361, + 545, + 506 + ], + "type": "text", + "content": " contains an additional \"no object\" class (denoted " + }, + { + "bbox": [ + 304, + 361, + 545, + 506 + ], + "type": "inline_equation", + "content": "\\varnothing" + }, + { + "bbox": [ + 304, + 361, + 545, + 506 + ], + "type": "text", + "content": ") to indicate that the mask does not correspond to any of the known " + }, + { + "bbox": [ + 304, + 361, + 545, + 506 + ], + "type": "inline_equation", + "content": "\\mathcal{K}^{1:t}" + }, + { + "bbox": [ + 304, + 361, + 545, + 506 + ], + "type": "text", + "content": " categories." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 507, + 546, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 507, + 546, + 689 + ], + "spans": [ + { + "bbox": [ + 304, + 507, + 546, + 689 + ], + "type": "text", + "content": "To obtain such output, the CoMFormer architecture, showcased in Fig. 2, is made of three components: (i) a backbone that extracts feature embeddings " + }, + { + "bbox": [ + 304, + 507, + 546, + 689 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_{\\text {feat }}" + }, + { + "bbox": [ + 304, + 507, + 546, + 689 + ], + "type": "text", + "content": ", (ii) a transformer decoder that takes as input N learnable queries and " + }, + { + "bbox": [ + 304, + 507, + 546, + 689 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_{\\text {feat }}" + }, + { + "bbox": [ + 304, + 507, + 546, + 689 + ], + "type": "text", + "content": " to output " + }, + { + "bbox": [ + 304, + 507, + 546, + 689 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 507, + 546, + 689 + ], + "type": "text", + "content": " per-segment embeddings " + }, + { + "bbox": [ + 304, + 507, + 546, + 689 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 304, + 507, + 546, + 689 + ], + "type": "text", + "content": ", (iii) a pixel decoder that takes " + }, + { + "bbox": [ + 304, + 507, + 546, + 689 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_{\\text {feat }}" + }, + { + "bbox": [ + 304, + 507, + 546, + 689 + ], + "type": "text", + "content": " as input and extracts per-pixel embeddings " + }, + { + "bbox": [ + 304, + 507, + 546, + 689 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_{\\text {pixel }}" + }, + { + "bbox": [ + 304, + 507, + 546, + 689 + ], + "type": "text", + "content": ". The output class probabilities " + }, + { + "bbox": [ + 304, + 507, + 546, + 689 + ], + "type": "inline_equation", + "content": "\\{p_i \\in \\Delta^{K^t + 1}\\}_{i=1}^N" + }, + { + "bbox": [ + 304, + 507, + 546, + 689 + ], + "type": "text", + "content": " are obtained by applying a linear classifier on " + }, + { + "bbox": [ + 304, + 507, + 546, + 689 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 304, + 507, + 546, + 689 + ], + "type": "text", + "content": ", which is enlarged as new classes are presented. We obtain the mask predictions " + }, + { + "bbox": [ + 304, + 507, + 546, + 689 + ], + "type": "inline_equation", + "content": "m_i" + }, + { + "bbox": [ + 304, + 507, + 546, + 689 + ], + "type": "text", + "content": " combining " + }, + { + "bbox": [ + 304, + 507, + 546, + 689 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 304, + 507, + 546, + 689 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 507, + 546, + 689 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_{\\text {pixel }}" + }, + { + "bbox": [ + 304, + 507, + 546, + 689 + ], + "type": "text", + "content": ": first " + }, + { + "bbox": [ + 304, + 507, + 546, + 689 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 304, + 507, + 546, + 689 + ], + "type": "text", + "content": " is passed into a 2-layer MLP to obtain N mask embeddings " + }, + { + "bbox": [ + 304, + 507, + 546, + 689 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_{\\text {mask }}" + }, + { + "bbox": [ + 304, + 507, + 546, + 689 + ], + "type": "text", + "content": " and we then compute the dot product between the " + }, + { + "bbox": [ + 304, + 507, + 546, + 689 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 507, + 546, + 689 + ], + "type": "text", + "content": "-th mask embedding and the pixel embeddings, followed by the softmax activation: " + }, + { + "bbox": [ + 304, + 507, + 546, + 689 + ], + "type": "inline_equation", + "content": "m_i[h, w] = \\operatorname{softmax}(\\mathcal{E}_{\\text {mask}}[:, i]^T \\cdot \\mathcal{E}_{\\text {pixel}}[:, h, w])" + }, + { + "bbox": [ + 304, + 507, + 546, + 689 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 689, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 689, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 305, + 689, + 545, + 714 + ], + "type": "text", + "content": "We point out that employing the softmax contrasts sharply with MaskFormers [11, 12], which use sigmoid ac" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3012" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 48, + 68, + 272, + 205 + ], + "blocks": [ + { + "bbox": [ + 48, + 68, + 272, + 205 + ], + "lines": [ + { + "bbox": [ + 48, + 68, + 272, + 205 + ], + "spans": [ + { + "bbox": [ + 48, + 68, + 272, + 205 + ], + "type": "image", + "image_path": "a565ed99ec382a9209e368e4e7d0ebbf02cf6e43a6821034004cc1717fc002b1.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 213, + 287, + 257 + ], + "lines": [ + { + "bbox": [ + 46, + 213, + 287, + 257 + ], + "spans": [ + { + "bbox": [ + 46, + 213, + 287, + 257 + ], + "type": "text", + "content": "Figure 3. Illustration of the mask predictions between sigmoid and softmax. MaskFormer has overlap across masks (e.g. in Query 1 and 2) predicted either as old or new classes, while CoMFormer avoids the issue using the softmax activation." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 262, + 288, + 358 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 262, + 288, + 358 + ], + "spans": [ + { + "bbox": [ + 46, + 262, + 288, + 358 + ], + "type": "text", + "content": "tivation to get the mask predictions. Using the sigmoid activation would result in the overlap of many masks on a single pixel, each potentially belonging to a distinct class, as seen in Fig. 3. This is particularly important in the context of continual learning, when the model may predict two similar masks associated with two distinct classes (e.g. cat and dog classes in Fig. 3) thereby generating interference among old and new classes and degrading the model's performance." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 365, + 204, + 376 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 365, + 204, + 376 + ], + "spans": [ + { + "bbox": [ + 47, + 365, + 204, + 376 + ], + "type": "text", + "content": "3.3. Learning without Forgetting" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 383, + 287, + 455 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 383, + 287, + 455 + ], + "spans": [ + { + "bbox": [ + 46, + 383, + 287, + 455 + ], + "type": "text", + "content": "To learn new classes, a simple solution is to fine-tune the model on the new dataset. However, because it lacks annotations for old classes, this operation results in catastrophic forgetting [21]. To alleviate the issue, we introduce two techniques for avoiding forgetting: an adaptive distillation loss and a mask-based pseudo-labeling strategy." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 460, + 287, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 460, + 287, + 604 + ], + "spans": [ + { + "bbox": [ + 46, + 460, + 287, + 604 + ], + "type": "text", + "content": "Adaptive Distillation Loss. A common strategy to mitigate forgetting adopted by continual semantic segmentation approaches [4, 15, 39] is to distill the knowledge coming from the old model in the new one, either in form of features [15, 39] or as classification probabilities [4]. We take inspiration from the latter solution and we design a novel distillation loss tailored to the CoMFormer architecture. In particular, analyzing the two components constituting the output space, we find the mask prediction being robust to forgetting while the classification part is heavily affected (see Sec. 4.6). For this reason, we focus on the output class probabilities to design our distillation loss." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 604, + 287, + 675 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 604, + 287, + 675 + ], + "spans": [ + { + "bbox": [ + 46, + 604, + 287, + 675 + ], + "type": "text", + "content": "An adaptation of standard knowledge distillation works [4] would force the new model to mimic the old model probabilities over all the N outputs, independently from their content. Formally, given the set of N probability distributions " + }, + { + "bbox": [ + 46, + 604, + 287, + 675 + ], + "type": "inline_equation", + "content": "\\{p_i^o\\}_{i = 1}^N" + }, + { + "bbox": [ + 46, + 604, + 287, + 675 + ], + "type": "text", + "content": " coming from the old model, a standard distillation loss is computed as" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 80, + 681, + 287, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 80, + 681, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 80, + 681, + 287, + 715 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {K D} = - \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\sum_ {k = 1} ^ {K ^ {t - 1} + 1} p _ {i} ^ {o} (k) \\log \\frac {\\tilde {p} _ {i} (k)}{p _ {i} ^ {o} (k)}, \\tag {1}", + "image_path": "ba00b0f162b61ffa2263556b8cb29e26bb58ea447b271ffaf152ca69ee7805c0.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 72, + 545, + 97 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 545, + 97 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 545, + 97 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 72, + 545, + 97 + ], + "type": "inline_equation", + "content": "K^{t - 1} + 1" + }, + { + "bbox": [ + 305, + 72, + 545, + 97 + ], + "type": "text", + "content": " is the " + }, + { + "bbox": [ + 305, + 72, + 545, + 97 + ], + "type": "inline_equation", + "content": "\\varnothing" + }, + { + "bbox": [ + 305, + 72, + 545, + 97 + ], + "type": "text", + "content": " class and " + }, + { + "bbox": [ + 305, + 72, + 545, + 97 + ], + "type": "inline_equation", + "content": "\\tilde{p}_i(k)" + }, + { + "bbox": [ + 305, + 72, + 545, + 97 + ], + "type": "text", + "content": " represents the unbiased probability distribution of " + }, + { + "bbox": [ + 305, + 72, + 545, + 97 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 305, + 72, + 545, + 97 + ], + "type": "text", + "content": "-th output for class " + }, + { + "bbox": [ + 305, + 72, + 545, + 97 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 305, + 72, + 545, + 97 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 331, + 106, + 545, + 137 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 331, + 106, + 545, + 137 + ], + "spans": [ + { + "bbox": [ + 331, + 106, + 545, + 137 + ], + "type": "interline_equation", + "content": "\\tilde {p} _ {i} (k) = \\left\\{ \\begin{array}{l l} \\sum_ {j \\in \\mathcal {K} ^ {t}} p _ {i} (j) + p _ {i} (\\varnothing) & \\text {i f} k = \\varnothing , \\\\ p _ {i} (k) & \\text {o t h e r w i s e .} \\end{array} \\right. \\tag {2}", + "image_path": "1a13bc103dc318e4ef276378a437b28ff2ec670c580550acadbf0ab752b4be68.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 146, + 545, + 182 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 146, + 545, + 182 + ], + "spans": [ + { + "bbox": [ + 305, + 146, + 545, + 182 + ], + "type": "text", + "content": "Note that, similarly to [4], it forces the sum of the model new class " + }, + { + "bbox": [ + 305, + 146, + 545, + 182 + ], + "type": "inline_equation", + "content": "(\\mathcal{K}^t)" + }, + { + "bbox": [ + 305, + 146, + 545, + 182 + ], + "type": "text", + "content": " and the \"no object\" class " + }, + { + "bbox": [ + 305, + 146, + 545, + 182 + ], + "type": "inline_equation", + "content": "(\\varnothing)" + }, + { + "bbox": [ + 305, + 146, + 545, + 182 + ], + "type": "text", + "content": " probabilities to be similar to the old model " + }, + { + "bbox": [ + 305, + 146, + 545, + 182 + ], + "type": "inline_equation", + "content": "\\varnothing" + }, + { + "bbox": [ + 305, + 146, + 545, + 182 + ], + "type": "text", + "content": " probability." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 183, + 545, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 183, + 545, + 266 + ], + "spans": [ + { + "bbox": [ + 305, + 183, + 545, + 266 + ], + "type": "text", + "content": "However, not every output probability brings relevant information: most of them are predicted as \"no object\" " + }, + { + "bbox": [ + 305, + 183, + 545, + 266 + ], + "type": "inline_equation", + "content": "(\\varnothing)" + }, + { + "bbox": [ + 305, + 183, + 545, + 266 + ], + "type": "text", + "content": " and thus they do not carry enough details to remember the old classes while reducing the relative importance of the other outputs probabilities. For this reason, we re-weight their contribution based on the probability of not being " + }, + { + "bbox": [ + 305, + 183, + 545, + 266 + ], + "type": "inline_equation", + "content": "\\varnothing" + }, + { + "bbox": [ + 305, + 183, + 545, + 266 + ], + "type": "text", + "content": ". Formally, we define our adaptive distillation loss as" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 312, + 277, + 545, + 321 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 277, + 545, + 321 + ], + "spans": [ + { + "bbox": [ + 312, + 277, + 545, + 321 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {A D} (p, p ^ {o}) = - \\sum_ {i = 1} ^ {N} \\frac {\\omega_ {i}}{\\sum_ {i = 1} ^ {N} \\omega_ {i}} \\sum_ {k = 1} ^ {K ^ {t - 1} + 1} p _ {i} ^ {o} (k) \\log \\frac {\\tilde {p} _ {i} (k)}{p _ {i} ^ {o} (k)}, \\tag {3}", + "image_path": "45915353f481dab71eb157a72c53bc5e396634dc8a70acca0b0a6d1e095e920b.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 321, + 545, + 381 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 321, + 545, + 381 + ], + "spans": [ + { + "bbox": [ + 305, + 321, + 545, + 381 + ], + "type": "text", + "content": "where the weighting coefficient is " + }, + { + "bbox": [ + 305, + 321, + 545, + 381 + ], + "type": "inline_equation", + "content": "\\omega_{i} = (1 - p_{i}^{o}(\\varnothing))^{2}" + }, + { + "bbox": [ + 305, + 321, + 545, + 381 + ], + "type": "text", + "content": ". Note that, with this formulation, we effectively reduce the contributions of outputs that bring small information about the old classes " + }, + { + "bbox": [ + 305, + 321, + 545, + 381 + ], + "type": "inline_equation", + "content": "(p_i^o (\\varnothing)\\approx 1)" + }, + { + "bbox": [ + 305, + 321, + 545, + 381 + ], + "type": "text", + "content": ", while we increase the contribution of important ones (where " + }, + { + "bbox": [ + 305, + 321, + 545, + 381 + ], + "type": "inline_equation", + "content": "p_i^o (\\varnothing)\\approx 0" + }, + { + "bbox": [ + 305, + 321, + 545, + 381 + ], + "type": "text", + "content": ")." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 387, + 545, + 471 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 387, + 545, + 471 + ], + "spans": [ + { + "bbox": [ + 305, + 387, + 545, + 471 + ], + "type": "text", + "content": "Mask-based Pseudo-labeling Strategy. Forgetting is effectively reduced by the adaptive distillation loss but we are not, however, taking advantage of the presence of the old classes that are unlabeled in the dataset. We propose using the old model predictions to recognize old class segments and generate pseudo-labels to make efficient use of this information and improve the knowledge of past classes." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 471, + 545, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 471, + 545, + 555 + ], + "spans": [ + { + "bbox": [ + 305, + 471, + 545, + 555 + ], + "type": "text", + "content": "A simple strategy consists in considering the N output pairs coming from the old model independently and using all the pairs for which the predicted class is not " + }, + { + "bbox": [ + 305, + 471, + 545, + 555 + ], + "type": "inline_equation", + "content": "\\varnothing" + }, + { + "bbox": [ + 305, + 471, + 545, + 555 + ], + "type": "text", + "content": " as pseudolabels. This strategy, however, does not consider two aspects: there may be pairs where the class is different from " + }, + { + "bbox": [ + 305, + 471, + 545, + 555 + ], + "type": "inline_equation", + "content": "\\varnothing" + }, + { + "bbox": [ + 305, + 471, + 545, + 555 + ], + "type": "text", + "content": " but (i) the mask substantially overlaps with a ground-truth segment, or (ii) the mask is noisy and has low confidence." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 305, + 555, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 555, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 555, + 545, + 713 + ], + "type": "text", + "content": "To overcome these issues, we propose a mask-based pseudo-labeling strategy that jointly considers the mask and class prediction confidences to avoid noisy labels and overlaps with the existing annotations. We define the model confidence as the multiplication among the class and mask probabilities. Formally, we compute the confidence for the " + }, + { + "bbox": [ + 305, + 555, + 545, + 713 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 305, + 555, + 545, + 713 + ], + "type": "text", + "content": "-th output as " + }, + { + "bbox": [ + 305, + 555, + 545, + 713 + ], + "type": "inline_equation", + "content": "q_{i} = p_{i}^{max} \\cdot m_{i}^{o} \\in [0,1]^{H \\times W}" + }, + { + "bbox": [ + 305, + 555, + 545, + 713 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 305, + 555, + 545, + 713 + ], + "type": "inline_equation", + "content": "p_{i}^{max} = \\max_{k=1}^{K^{t-1}} p_{i}^{o}(k)" + }, + { + "bbox": [ + 305, + 555, + 545, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 555, + 545, + 713 + ], + "type": "inline_equation", + "content": "m_{i}^{o}" + }, + { + "bbox": [ + 305, + 555, + 545, + 713 + ], + "type": "text", + "content": " is the binary mask predicted by the old model. Note that we do not consider the " + }, + { + "bbox": [ + 305, + 555, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\varnothing" + }, + { + "bbox": [ + 305, + 555, + 545, + 713 + ], + "type": "text", + "content": " class in the max operation. We denote the pseudo-class " + }, + { + "bbox": [ + 305, + 555, + 545, + 713 + ], + "type": "inline_equation", + "content": "c_{i}^{ps} = \\operatorname{argmax}_{k=1}^{K^{t-1}} p_{i}^{o}(k)" + }, + { + "bbox": [ + 305, + 555, + 545, + 713 + ], + "type": "text", + "content": " and we generate the pseudomask " + }, + { + "bbox": [ + 305, + 555, + 545, + 713 + ], + "type": "inline_equation", + "content": "m_{i}^{ps}" + }, + { + "bbox": [ + 305, + 555, + 545, + 713 + ], + "type": "text", + "content": " considering two criteria: (i) there should be no overlap between the pseudo-mask and the ground-truth seg" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3013" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 287, + 137 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 287, + 137 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 287, + 137 + ], + "type": "text", + "content": "ments and (ii) to be included in a pseudo-mask, the confidence on a pixel should be maximum over all outputs. Formally, denoting the binarization of the predicted mask as " + }, + { + "bbox": [ + 47, + 72, + 287, + 137 + ], + "type": "inline_equation", + "content": "m_{i}^{bin} = m_{i} > 0.5" + }, + { + "bbox": [ + 47, + 72, + 287, + 137 + ], + "type": "text", + "content": " and the union of the ground-truth segments " + }, + { + "bbox": [ + 47, + 72, + 287, + 137 + ], + "type": "inline_equation", + "content": "m_{all}^{gt} = \\sum_{i=1}^{N^{ps}} m_{i}^{gt}" + }, + { + "bbox": [ + 47, + 72, + 287, + 137 + ], + "type": "text", + "content": ", we generate the pseudo-mask as" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 149, + 287, + 191 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 149, + 287, + 191 + ], + "spans": [ + { + "bbox": [ + 56, + 149, + 287, + 191 + ], + "type": "interline_equation", + "content": "m _ {i} ^ {p s} [ h, w ] = \\left\\{ \\begin{array}{l l} m _ {i} ^ {b i n} [ h, w ] & \\text {i f} i = \\operatorname {a r g m a x} _ {j = 1} ^ {N} q _ {j} [ h, w ] \\\\ & \\text {a n d} m _ {a l l} ^ {g t} [ h, w ] = 0, \\\\ 0 & \\text {o t h e r w i s e .} \\end{array} \\right. \\tag {4}", + "image_path": "4963cce52da75d33e8c7e13c7270f3a70aea950a2cbc79b82ebf49f7c0363e69.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 196, + 287, + 291 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 196, + 287, + 291 + ], + "spans": [ + { + "bbox": [ + 47, + 196, + 287, + 291 + ], + "type": "text", + "content": "Indeed, not all pseudo-mask should be included in the set of pseudo-labels. There may be pseudo-masks where no pixel is active, i.e. it is all zeros, or where the pseudomask contains only a small fraction of the original mask. For these reasons, we construct the set of pseudo-labels " + }, + { + "bbox": [ + 47, + 196, + 287, + 291 + ], + "type": "inline_equation", + "content": "z^{ps} = \\{(c_i^{ps},m_i^{ps})\\}_{i = 1}^{N^{ps}}" + }, + { + "bbox": [ + 47, + 196, + 287, + 291 + ], + "type": "text", + "content": " such that we include in " + }, + { + "bbox": [ + 47, + 196, + 287, + 291 + ], + "type": "inline_equation", + "content": "z^{ps}" + }, + { + "bbox": [ + 47, + 196, + 287, + 291 + ], + "type": "text", + "content": " only the pseudo-masks that retain at least half of the pixels w.r.t. the binary mask " + }, + { + "bbox": [ + 47, + 196, + 287, + 291 + ], + "type": "inline_equation", + "content": "m_{i}^{bin}" + }, + { + "bbox": [ + 47, + 196, + 287, + 291 + ], + "type": "text", + "content": " and that have at least one active pixel." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 292, + 287, + 329 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 292, + 287, + 329 + ], + "spans": [ + { + "bbox": [ + 47, + 292, + 287, + 329 + ], + "type": "text", + "content": "We denote the final annotation set " + }, + { + "bbox": [ + 47, + 292, + 287, + 329 + ], + "type": "inline_equation", + "content": "\\bar{z} = \\{\\bar{c}_i, \\bar{m}_i\\}_{i=1}^N" + }, + { + "bbox": [ + 47, + 292, + 287, + 329 + ], + "type": "text", + "content": " as the union of the ground-truth labels " + }, + { + "bbox": [ + 47, + 292, + 287, + 329 + ], + "type": "inline_equation", + "content": "z^{gt}" + }, + { + "bbox": [ + 47, + 292, + 287, + 329 + ], + "type": "text", + "content": " and the pseudo-labels " + }, + { + "bbox": [ + 47, + 292, + 287, + 329 + ], + "type": "inline_equation", + "content": "z^{ps}" + }, + { + "bbox": [ + 47, + 292, + 287, + 329 + ], + "type": "text", + "content": ", i.e. " + }, + { + "bbox": [ + 47, + 292, + 287, + 329 + ], + "type": "inline_equation", + "content": "\\bar{z} = z^{gt} \\cup z^{ps}" + }, + { + "bbox": [ + 47, + 292, + 287, + 329 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 292, + 287, + 329 + ], + "type": "inline_equation", + "content": "\\bar{N} = N^{gt} + N^{ps}" + }, + { + "bbox": [ + 47, + 292, + 287, + 329 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 334, + 287, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 334, + 287, + 453 + ], + "spans": [ + { + "bbox": [ + 46, + 334, + 287, + 453 + ], + "type": "text", + "content": "Overall training loss. To train CoMFormer, a one-to-one matching " + }, + { + "bbox": [ + 46, + 334, + 287, + 453 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 46, + 334, + 287, + 453 + ], + "type": "text", + "content": " between the model predictions " + }, + { + "bbox": [ + 46, + 334, + 287, + 453 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 46, + 334, + 287, + 453 + ], + "type": "text", + "content": " and the annotation set " + }, + { + "bbox": [ + 46, + 334, + 287, + 453 + ], + "type": "inline_equation", + "content": "\\bar{z}" + }, + { + "bbox": [ + 46, + 334, + 287, + 453 + ], + "type": "text", + "content": " is required. Following standard practices [1, 11], we employ the Hungarian matching algorithm [30] and we search for the matching " + }, + { + "bbox": [ + 46, + 334, + 287, + 453 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 46, + 334, + 287, + 453 + ], + "type": "text", + "content": " that minimize the assignment cost computed as " + }, + { + "bbox": [ + 46, + 334, + 287, + 453 + ], + "type": "inline_equation", + "content": "-p_i(\\bar{c}_j)\\cdot Dice(m_i,\\bar{m}_j)" + }, + { + "bbox": [ + 46, + 334, + 287, + 453 + ], + "type": "text", + "content": ", where Dice indicates the Dice coefficient [13]. We assume the size of the prediction set to be larger than the annotations set (i.e. " + }, + { + "bbox": [ + 46, + 334, + 287, + 453 + ], + "type": "inline_equation", + "content": "N > \\bar{N}" + }, + { + "bbox": [ + 46, + 334, + 287, + 453 + ], + "type": "text", + "content": ") and to obtain a one-to-one matching we pad the annotations with \"no object\" tokens " + }, + { + "bbox": [ + 46, + 334, + 287, + 453 + ], + "type": "inline_equation", + "content": "\\varnothing" + }, + { + "bbox": [ + 46, + 334, + 287, + 453 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 453, + 287, + 477 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 453, + 287, + 477 + ], + "spans": [ + { + "bbox": [ + 47, + 453, + 287, + 477 + ], + "type": "text", + "content": "Once the best matching " + }, + { + "bbox": [ + 47, + 453, + 287, + 477 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 47, + 453, + 287, + 477 + ], + "type": "text", + "content": " has been found, we define the overall training loss as:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 96, + 485, + 287, + 498 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 96, + 485, + 287, + 498 + ], + "spans": [ + { + "bbox": [ + 96, + 485, + 287, + 498 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\mathcal {L} _ {\\text {S e g}} (z, \\bar {z}) + \\lambda_ {D} \\mathcal {L} _ {A D} (p, p ^ {o}), \\tag {5}", + "image_path": "d6dcf6df9d71f878b3c6f6e81be358d480824c665ca077c34cd7b611dcd57478.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 506, + 229, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 506, + 229, + 517 + ], + "spans": [ + { + "bbox": [ + 47, + 506, + 229, + 517 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 506, + 229, + 517 + ], + "type": "inline_equation", + "content": "\\lambda_{D}" + }, + { + "bbox": [ + 47, + 506, + 229, + 517 + ], + "type": "text", + "content": " is a trade-off hyper-parameter and" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 52, + 525, + 287, + 576 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 52, + 525, + 287, + 576 + ], + "spans": [ + { + "bbox": [ + 52, + 525, + 287, + 576 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} _ {\\text {S e g}} (z, \\bar {z}) = \\sum_ {j = 1} ^ {N} [ - \\alpha (1 - p _ {\\sigma (j)} (\\bar {c} _ {j})) ^ {\\gamma} \\log p _ {\\sigma (j)} (\\bar {c} _ {j}) + \\tag {6} \\\\ \\left. + \\lambda_ {m a s k} \\mathbb {1} _ {\\bar {c} _ {j} \\neq \\emptyset} \\mathcal {L} _ {m a s k} \\left(m _ {\\sigma (j)}, \\bar {m} _ {j}\\right) \\right]. \\\\ \\end{array}", + "image_path": "0f3936c57c556b7c82638d4b2cb777b7762a84453cd8a298ae515fcbb22c6470.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 581, + 287, + 642 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 581, + 287, + 642 + ], + "spans": [ + { + "bbox": [ + 47, + 581, + 287, + 642 + ], + "type": "text", + "content": "We note that the first term in " + }, + { + "bbox": [ + 47, + 581, + 287, + 642 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{Seg}" + }, + { + "bbox": [ + 47, + 581, + 287, + 642 + ], + "type": "text", + "content": " indicates the focal loss [34], where " + }, + { + "bbox": [ + 47, + 581, + 287, + 642 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 47, + 581, + 287, + 642 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 581, + 287, + 642 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 47, + 581, + 287, + 642 + ], + "type": "text", + "content": " are hyper-parameters. Finally, " + }, + { + "bbox": [ + 47, + 581, + 287, + 642 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{mask}" + }, + { + "bbox": [ + 47, + 581, + 287, + 642 + ], + "type": "text", + "content": " is the sum of dice and cross-entropy losses, and " + }, + { + "bbox": [ + 47, + 581, + 287, + 642 + ], + "type": "inline_equation", + "content": "\\lambda_{mask}" + }, + { + "bbox": [ + 47, + 581, + 287, + 642 + ], + "type": "text", + "content": " is a hyper-parameter. We note that " + }, + { + "bbox": [ + 47, + 581, + 287, + 642 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{mask}" + }, + { + "bbox": [ + 47, + 581, + 287, + 642 + ], + "type": "text", + "content": " is computed only for valid segments, i.e. where " + }, + { + "bbox": [ + 47, + 581, + 287, + 642 + ], + "type": "inline_equation", + "content": "\\bar{c}_j \\neq \\emptyset" + }, + { + "bbox": [ + 47, + 581, + 287, + 642 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 651, + 127, + 664 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 651, + 127, + 664 + ], + "spans": [ + { + "bbox": [ + 47, + 651, + 127, + 664 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 671, + 165, + 684 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 671, + 165, + 684 + ], + "spans": [ + { + "bbox": [ + 47, + 671, + 165, + 684 + ], + "type": "text", + "content": "4.1. Dataset and Settings" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "content": "We start from the widely adopted continual semantic segmentation benchmark defined in [4] and we extend it to" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 72, + 545, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 180 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 180 + ], + "type": "text", + "content": "the newly proposed continual panoptic segmentation using a modified Continuum library [18]. We compare our novel model on the large-scale ADE20K [62] dataset since it supports both semantic and panoptic segmentation tasks. This challenging dataset contains 150 classes, divided into 100 \"things\" and 50 \"stuff\" categories. This dataset represents a wide variety of scenes, both interior and exterior, with images featuring an average of 9.9 classes, while other datasets like COCO [35] only have an average of 3.5 classes." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 186, + 545, + 414 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 186, + 545, + 414 + ], + "spans": [ + { + "bbox": [ + 304, + 186, + 545, + 414 + ], + "type": "text", + "content": "Continual Learning Protocols. Previous continual semantic segmentation works [4, 15] describe a training protocol to assess the performance on multiple continual learning steps. In particular, [4] proposed three protocols for ADE20K with different numbers of tasks: (i) the 100-50 consists of two tasks, the first of 100 and the second of 50 classes; (ii) the 50-50 consists of three tasks of 50 classes; (iii) the 100-10 consists on 6 tasks, the first on 100 classes followed by 5 tasks of 10 classes. In addition, we use the 100-5 introduced in [15], consisting of 11 tasks, the first of 100 classes followed by 10 tasks of 5 classes. To divide the images into different tasks, we follow the splits provided by [4] for 100-50, 100-10, 50-50 for both semantic and panoptic segmentation. In particular, they ensured that each image appears only on a unique task. Differently, for the 100-5 we use the split proposed by [15], where the same image may appear on multiple tasks. We report the results of the 50-50 setting for both semantic and panoptic segmentation in the supplementary material." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 420, + 545, + 553 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 420, + 545, + 553 + ], + "spans": [ + { + "bbox": [ + 304, + 420, + 545, + 553 + ], + "type": "text", + "content": "Metrics. We compare the different models using the mean Intersection over Union (mIoU) on semantic segmentation [20] and Panoptic Quality (PQ) on panoptic segmentation [28]. PQ is the product of two components: Segmentation Quality (SQ) which considers the IoU between the correctly classified segments, and Recognition Quality (RQ) which only considers the classification accuracy. For both tasks, we report the metric after the last step " + }, + { + "bbox": [ + 304, + 420, + 545, + 553 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 304, + 420, + 545, + 553 + ], + "type": "text", + "content": " on the first classes " + }, + { + "bbox": [ + 304, + 420, + 545, + 553 + ], + "type": "inline_equation", + "content": "(C^1)" + }, + { + "bbox": [ + 304, + 420, + 545, + 553 + ], + "type": "text", + "content": ", for the added classes " + }, + { + "bbox": [ + 304, + 420, + 545, + 553 + ], + "type": "inline_equation", + "content": "(C^{2:T})" + }, + { + "bbox": [ + 304, + 420, + 545, + 553 + ], + "type": "text", + "content": ", and for all the classes (all). We also report the average of the final performance on all seen classes after each step (avg) following [15]." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 562, + 372, + 573 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 562, + 372, + 573 + ], + "spans": [ + { + "bbox": [ + 306, + 562, + 372, + 573 + ], + "type": "text", + "content": "4.2. Baselines" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "content": "We benchmark our model against the state-of-the-art methods: MiB [4], PLOP [15], and RCIL [59]. In semantic segmentation, we report them with the original DeepLab-v3 segmentation architecture [9]. In addition, for a fair comparison, we report MiB and PLOP using the CoMFormer architecture. In particular, we implement the methods in two versions: (i) we replace the CoMFormer losses with a per-pixel cross-entropy loss on the dot product between mask and class logits and we apply the continual learning methods on that baseline (we refer to this as Per-Pixel), and (ii) by applying the distillation losses directly on the" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3014" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 69, + 69, + 525, + 150 + ], + "blocks": [ + { + "bbox": [ + 69, + 69, + 525, + 150 + ], + "lines": [ + { + "bbox": [ + 69, + 69, + 525, + 150 + ], + "spans": [ + { + "bbox": [ + 69, + 69, + 525, + 150 + ], + "type": "table", + "html": "
Method100-50 (2 tasks)100-10 (6 tasks)100-5 (11 tasks)
1-100101-150allavg1-100101-150allavg1-100101-150allavg
FT0.025.88.624.80.02.91.07.90.01.30.44.6
MiB35.119.329.835.427.110.021.429.124.06.518.125.6
PLOP41.026.636.238.630.517.526.132.928.115.724.030.5
CoMFormer41.127.736.738.836.017.129.735.334.415.928.234.0
Joint43.232.139.5—43.232.139.5—43.232.139.5—
", + "image_path": "37dc5e3bba87f7e317eb9177039f1ee33715a70eda57e250b8b25100964ef948.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 63, + 181, + 531, + 317 + ], + "blocks": [ + { + "bbox": [ + 80, + 159, + 511, + 171 + ], + "lines": [ + { + "bbox": [ + 80, + 159, + 511, + 171 + ], + "spans": [ + { + "bbox": [ + 80, + 159, + 511, + 171 + ], + "type": "text", + "content": "Table 1. Continual Panoptic Segmentation results on ADE20K dataset in PQ considering multiple training protocols." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 63, + 181, + 531, + 317 + ], + "lines": [ + { + "bbox": [ + 63, + 181, + 531, + 317 + ], + "spans": [ + { + "bbox": [ + 63, + 181, + 531, + 317 + ], + "type": "table", + "html": "
BackboneMethod100-50 (2 tasks)100-10 (6 tasks)100-5 (11 tasks)
1-100101-150allavg1-100101-150allavg1-100101-150allavg
DeepLab-v3 [9]MiB [4]40.517.232.837.338.311.329.235.136.05.726.032.7
PLOP [15]41.914.932.937.440.514.131.636.639.17.828.835.3
RCIL [59]42.318.834.5—39.317.632.1—38.511.529.6—
Per-PixelMiB40.324.034.837.535.114.028.134.033.315.227.333.1
PLOP40.220.233.536.932.613.726.332.433.39.425.432.8
Mask-basedFT0.026.78.926.40.02.30.88.50.01.10.34.2
MiB37.024.132.638.323.510.626.629.621.06.116.127.7
PLOP44.226.238.241.134.815.928.535.233.614.127.133.6
CoMFormer44.726.238.441.240.615.632.337.439.513.630.936.5
Joint46.935.643.1—46.935.643.1—46.935.643.1—
", + "image_path": "62191e1965c0a981a7970352023c96b01ab7eb0e8f1fc8a4e48f5179535b2283.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 326, + 547, + 349 + ], + "lines": [ + { + "bbox": [ + 46, + 326, + 547, + 349 + ], + "spans": [ + { + "bbox": [ + 46, + 326, + 547, + 349 + ], + "type": "text", + "content": "Table 2. Continual Semantic Segmentation results on the ADE20K dataset in mIoU considering multiple training protocols. We report the numbers published by the original papers on DeepLab-v3 [9]. The remaining results are based on our re-implementation (see Sec. 4.2)." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 360, + 289, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 360, + 289, + 529 + ], + "spans": [ + { + "bbox": [ + 46, + 360, + 289, + 529 + ], + "type": "text", + "content": "classification head of the CoMFormer architecture (Mask-based). We note that the first version cannot be applied to panoptic segmentation, since the task does not support the use of per-pixel losses. For PLOP, we apply its distillation loss local-POD on the intermediate features of the segmentation backbone, as in the original paper, and we use the pseudo-labeling strategy proposed in Sec. 3. RCIL was not re-implemented since it is not possible to extend their approach to non-convolutional architectures like ours. We include a naive finetuning (FT) without any capabilities for continual learning. Finally, we also report an upper bound (Joint) that is trained in the traditional segmentation setting. We made our best effort to find the most suitable hyperparameters for all the baselines." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 537, + 180, + 550 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 537, + 180, + 550 + ], + "spans": [ + { + "bbox": [ + 47, + 537, + 180, + 550 + ], + "type": "text", + "content": "4.3. Implementation Details" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 552, + 287, + 660 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 552, + 287, + 660 + ], + "spans": [ + { + "bbox": [ + 46, + 552, + 287, + 660 + ], + "type": "text", + "content": "Architecture. Following the previous benchmarks [4, 15] we use as backbone a ResNet101 [23] on semantic segmentation and we propose to use a ResNet50 [23] for panoptic segmentation. For both tasks, we use the pixel decoder and the transformer decoder proposed in [11] with 9 layers in total and " + }, + { + "bbox": [ + 46, + 552, + 287, + 660 + ], + "type": "inline_equation", + "content": "N = 100" + }, + { + "bbox": [ + 46, + 552, + 287, + 660 + ], + "type": "text", + "content": " queries. In addition, we adopt all the improvements introduced in [11]: masked attention, multiscale features, and their optimization improvements. We report the performance of single-scale inference." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "type": "text", + "content": "Training parameters. We follow the Mask2Former [11] hyper-parameters and we use AdamW [25] optimizer with an initial learning rate of 0.0001 for the first step (" + }, + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "type": "inline_equation", + "content": "t = 0" + }, + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "type": "text", + "content": ") and 0.00005 in the following (" + }, + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "type": "inline_equation", + "content": "t > 0" + }, + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "type": "text", + "content": "). A learning rate multiplier" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 360, + 547, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 360, + 547, + 540 + ], + "spans": [ + { + "bbox": [ + 304, + 360, + 547, + 540 + ], + "type": "text", + "content": "of 0.1 is applied to the backbone and we follow a polynomial learning rate schedule. We use weight decay of 0.05. We train the model for 160K iterations in the first step and 400 iterations per class in the following (e.g. learning 50 classes we train for " + }, + { + "bbox": [ + 304, + 360, + 547, + 540 + ], + "type": "inline_equation", + "content": "50 \\times 400 = 20000" + }, + { + "bbox": [ + 304, + 360, + 547, + 540 + ], + "type": "text", + "content": " iterations). For semantic segmentation, we use a crop size of 512 and a batch size of 16. For panoptic segmentation, we use a crop size of 640 and a batch size of 8. For both, we use the standard random scale jittering between 0.5 and 2.0, random horizontal flipping, random cropping, as well as random color jittering as data augmentation [12]. We set " + }, + { + "bbox": [ + 304, + 360, + 547, + 540 + ], + "type": "inline_equation", + "content": "\\alpha = 20" + }, + { + "bbox": [ + 304, + 360, + 547, + 540 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 360, + 547, + 540 + ], + "type": "inline_equation", + "content": "\\gamma = 2" + }, + { + "bbox": [ + 304, + 360, + 547, + 540 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 360, + 547, + 540 + ], + "type": "inline_equation", + "content": "\\lambda_{mask} = 5.0" + }, + { + "bbox": [ + 304, + 360, + 547, + 540 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 304, + 360, + 547, + 540 + ], + "type": "inline_equation", + "content": "\\lambda_{D}" + }, + { + "bbox": [ + 304, + 360, + 547, + 540 + ], + "type": "text", + "content": " is set to 1 for 100-50 and to 10 in the 100-10 and 100-5. Finally, following the standard protocol of continual learning [4, 15] we do not store any image of previous steps and we do not use rehearsal learning." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 550, + 487, + 563 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 550, + 487, + 563 + ], + "spans": [ + { + "bbox": [ + 305, + 550, + 487, + 563 + ], + "type": "text", + "content": "4.4. Continual Panoptic Segmentation" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 570, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 547, + 713 + ], + "type": "text", + "content": "Tab. 1 reports the results on the new continual panoptic segmentation benchmark. CoMFormer exceeds all the baselines by a significant margin. Considering the 100-50, we can see that CoMFormer outperforms MiB on both old (+6 PQ) and new (+8.4 PQ) classes and PLOP on new classes (+1.1 PQ). Considering a longer sequence of tasks (100-10 and 100-5), we can see that the gap between CoMFormer and the other methods becomes more significant, especially in the old classes. In 100-10, it surpasses MiB and PLOP respectively by 8.9 and 5.5 PQ in old classes, while in the new classes it exceeds MiB by 7.1 and is comparable to PLOP. In the 100-5, it outperforms MiB and PLOP" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3015" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 61, + 70, + 535, + 392 + ], + "blocks": [ + { + "bbox": [ + 61, + 70, + 535, + 392 + ], + "lines": [ + { + "bbox": [ + 61, + 70, + 535, + 392 + ], + "spans": [ + { + "bbox": [ + 61, + 70, + 535, + 392 + ], + "type": "image", + "image_path": "b87103b4b638a59f18c508ea4fd257ba27f4faa0c031eb21538dd618859f39c2.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 53, + 400, + 538, + 412 + ], + "lines": [ + { + "bbox": [ + 53, + 400, + 538, + 412 + ], + "spans": [ + { + "bbox": [ + 53, + 400, + 538, + 412 + ], + "type": "text", + "content": "Figure 4. Qualitative results of CoMformer v.s. MiB and PLOP on the 100-10 continual panoptic segmentation setting on ADE20K." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 423, + 287, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 423, + 287, + 506 + ], + "spans": [ + { + "bbox": [ + 46, + 423, + 287, + 506 + ], + "type": "text", + "content": "on both old (+10.4 and +6.3 PQ) and new classes (+9.4 and +0.2 PQ). Finally, we note that CoMFormer obtains performance close to the Joint baseline in the 100-50 setting, suffering a drop of only 2.1 PQ on the old classes and 4.4 PQ on the new ones. In the other settings, the gap is still relevant, indicating that learning from multiple training steps is still a very challenging task." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 507, + 288, + 627 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 507, + 288, + 627 + ], + "spans": [ + { + "bbox": [ + 46, + 507, + 288, + 627 + ], + "type": "text", + "content": "In Fig. 4, we report a qualitative comparison with the baselines in the 100-10 setting using randomly chosen images. We can see that our method is able to maintain better the previous knowledge by precisely segmenting all the classes in the image; for example, w.r.t. to PLOP, our method correctly segments the chair in the second row and the boat in the third row. We can note a common error: models correctly segment but misclassify objects in the image, like the plane classified as car in the first row or the press classified as a door in the third row." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 635, + 230, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 635, + 230, + 647 + ], + "spans": [ + { + "bbox": [ + 47, + 635, + 230, + 647 + ], + "type": "text", + "content": "4.5. Continual Semantic Segmentation" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 654, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 654, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 654, + 287, + 715 + ], + "type": "text", + "content": "We report the comparison with state-of-the-art methods in continual semantic segmentation in Tab. 2. CoMFormer achieves the best results on every setting, surpassing both methods based on DeepLab-v3 [9] and baselines implemented upon the same architecture. In particular, CoM-" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 422, + 545, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 422, + 545, + 662 + ], + "spans": [ + { + "bbox": [ + 304, + 422, + 545, + 662 + ], + "type": "text", + "content": "Former outperforms the best state-of-the-art method (RCIL [59] by " + }, + { + "bbox": [ + 304, + 422, + 545, + 662 + ], + "type": "inline_equation", + "content": "3.9\\%" + }, + { + "bbox": [ + 304, + 422, + 545, + 662 + ], + "type": "text", + "content": " on the 100-50, by " + }, + { + "bbox": [ + 304, + 422, + 545, + 662 + ], + "type": "inline_equation", + "content": "0.2\\%" + }, + { + "bbox": [ + 304, + 422, + 545, + 662 + ], + "type": "text", + "content": " on 100-10, and by " + }, + { + "bbox": [ + 304, + 422, + 545, + 662 + ], + "type": "inline_equation", + "content": "1.3\\%" + }, + { + "bbox": [ + 304, + 422, + 545, + 662 + ], + "type": "text", + "content": " on the 100-5. Moreover, it always improves the performance on the old classes and the avg across different steps. In addition, CoMFormer exceeds all the baselines. Comparing it with PLOP (Mask-based), we can see that it obtains a substantial improvement, especially considering scenarios with longer training protocols: CoMformers achieves " + }, + { + "bbox": [ + 304, + 422, + 545, + 662 + ], + "type": "inline_equation", + "content": "30.9\\%" + }, + { + "bbox": [ + 304, + 422, + 545, + 662 + ], + "type": "inline_equation", + "content": "(32.3\\%)" + }, + { + "bbox": [ + 304, + 422, + 545, + 662 + ], + "type": "text", + "content": " on the all metrics, while PLOP only " + }, + { + "bbox": [ + 304, + 422, + 545, + 662 + ], + "type": "inline_equation", + "content": "27.1\\%" + }, + { + "bbox": [ + 304, + 422, + 545, + 662 + ], + "type": "inline_equation", + "content": "(28.5\\%)" + }, + { + "bbox": [ + 304, + 422, + 545, + 662 + ], + "type": "text", + "content": " on the 100-5 " + }, + { + "bbox": [ + 304, + 422, + 545, + 662 + ], + "type": "inline_equation", + "content": "(100 - 10)" + }, + { + "bbox": [ + 304, + 422, + 545, + 662 + ], + "type": "text", + "content": " setting. Furthermore, the results confirm that improvements are not led by the newer architecture. Per-Pixel baselines achieve only similar or even lower performance than their counterparts. MiB (Per-Pixel), for example, slightly improves the one based on DeepLab-v3 on the 100-50 " + }, + { + "bbox": [ + 304, + 422, + 545, + 662 + ], + "type": "inline_equation", + "content": "(+2.0\\%" + }, + { + "bbox": [ + 304, + 422, + 545, + 662 + ], + "type": "text", + "content": " on all) and in the 100-5 " + }, + { + "bbox": [ + 304, + 422, + 545, + 662 + ], + "type": "inline_equation", + "content": "(+1.3\\%)" + }, + { + "bbox": [ + 304, + 422, + 545, + 662 + ], + "type": "text", + "content": " but its performance is slightly worse in the 100-10 " + }, + { + "bbox": [ + 304, + 422, + 545, + 662 + ], + "type": "inline_equation", + "content": "(-1.1\\%)" + }, + { + "bbox": [ + 304, + 422, + 545, + 662 + ], + "type": "text", + "content": ". In addition, Per-Pixel baselines demonstrate a large performance drop in the old classes, showing that, without employing a distillation strategy tailored to the architecture, it is difficult to maintain satisfactory results." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 673, + 406, + 685 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 673, + 406, + 685 + ], + "spans": [ + { + "bbox": [ + 306, + 673, + 406, + 685 + ], + "type": "text", + "content": "4.6. Ablation Studies" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 689, + 546, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 689, + 546, + 714 + ], + "spans": [ + { + "bbox": [ + 305, + 689, + 546, + 714 + ], + "type": "text", + "content": "Mask2Former vs CoMFormer. In Tab. 3, we compare the proposed CoMFormer against Mask2Former [11]. In" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3016" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 53, + 70, + 282, + 196 + ], + "blocks": [ + { + "bbox": [ + 53, + 70, + 282, + 196 + ], + "lines": [ + { + "bbox": [ + 53, + 70, + 282, + 196 + ], + "spans": [ + { + "bbox": [ + 53, + 70, + 282, + 196 + ], + "type": "table", + "html": "
MethodSemantic - mIoUPanoptic - PQ
1-100101-150all1-100101-150all
100-50 (2 tasks)
Mask2Former44.326.038.137.924.533.4
CoMFormer44.726.238.441.127.736.7
100-10 (6 tasks)
Mask2Former38.715.731.030.210.823.7
CoMFormer40.615.632.336.017.129.7
100-5 (11 tasks)
Mask2Former32.613.726.325.95.619.1
CoMFormer39.513.630.934.415.928.2
", + "image_path": "3f374d9a7cdc3dd4c92500cfe5efc61f93e5dc131ffea2747179ed09a8ccda0b.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 53, + 260, + 282, + 342 + ], + "blocks": [ + { + "bbox": [ + 47, + 205, + 287, + 250 + ], + "lines": [ + { + "bbox": [ + 47, + 205, + 287, + 250 + ], + "spans": [ + { + "bbox": [ + 47, + 205, + 287, + 250 + ], + "type": "text", + "content": "Table 3. Comparison between CoMFormer and Mask2Former [11] in continual semantic and panoptic segmentation. We apply to Mask2Former our mask-based pseudo-labeling strategy and the adaptive distillation loss for a fair comparison." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 53, + 260, + 282, + 342 + ], + "lines": [ + { + "bbox": [ + 53, + 260, + 282, + 342 + ], + "spans": [ + { + "bbox": [ + 53, + 260, + 282, + 342 + ], + "type": "table", + "html": "
100-50 (2 tasks)100-10 (6 tasks)
PSKDAD1-100101-150all1-100101-150all
---0.025.88.60.02.91.0
--✓28.415.324.023.09.118.4
✓--40.726.335.930.117.225.8
✓✓-40.528.136.430.217.626.0
✓-✓41.127.736.736.017.129.7
", + "image_path": "47a81d61d84a9eb2834060a4385d4baaa2030b778d385f655f506063f7c97902.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 407, + 287, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 407, + 287, + 540 + ], + "spans": [ + { + "bbox": [ + 46, + 407, + 287, + 540 + ], + "type": "text", + "content": "particular, to fairly compare the two methods, we add our adaptive distillation loss and the pseudo-labeling strategy to Mask2Former. Considering semantic segmentation, CoMFormer achieves better results on the old classes (on average " + }, + { + "bbox": [ + 46, + 407, + 287, + 540 + ], + "type": "inline_equation", + "content": "3\\%" + }, + { + "bbox": [ + 46, + 407, + 287, + 540 + ], + "type": "text", + "content": " mIoU), while the two methods obtain comparable results on the new classes. On panoptic segmentation, CoMFormer exceeds Mask2Former in all the classes with a considerable gap (6 PQ on both old and new classes on average). We conclude that a vanilla MaskFormer architecture is not suited to continual learning because it suffers from forgetting, while CoMFormer outperforms it in any setting." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 545, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 545, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 545, + 288, + 714 + ], + "type": "text", + "content": "Method components. We investigate the benefit of each method component in Tab. 4, reporting the results in PQ on the 100-50 and 100-10 settings in panoptic segmentation. Without applying any regularization technique, the performance on old classes drops to 0, denoting that it is important to deal with forgetting. Both the adaptive distillation loss (AD) and the pseudo-labeling strategy (PS) lead to improvement in the old classes when applied independently. However, we obtain the best results when combining them, especially considering the 100-10 scenario, where the old classes are improved by 5.8 PQ and the final performances of 3.9 PQ. Finally, the table shows that weighting the queries based on their semantic content is essential to improve performance: comparing a standard distillation" + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 310, + 70, + 541, + 140 + ], + "blocks": [ + { + "bbox": [ + 47, + 350, + 287, + 394 + ], + "lines": [ + { + "bbox": [ + 47, + 350, + 287, + 394 + ], + "spans": [ + { + "bbox": [ + 47, + 350, + 287, + 394 + ], + "type": "text", + "content": "Table 4. Ablation of the different key components of our CoM-Former architecture. We report the PQ on the 100-50 and 100-10 panoptic settings. PS: Pseudo-labeling; AD: adaptive distillation Eq. (3); KD: standard distillation Eq. (1)" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 310, + 70, + 541, + 140 + ], + "lines": [ + { + "bbox": [ + 310, + 70, + 541, + 140 + ], + "spans": [ + { + "bbox": [ + 310, + 70, + 541, + 140 + ], + "type": "table", + "html": "
Method100-50 (2 tasks)100-10 (6 tasks)
\\( RQ \\)\\( SQ \\)\\( PQ \\)\\( RQ \\)\\( SQ \\)\\( PQ \\)
MiB36.873.029.826.455.521.4
PLOP43.879.136.231.761.126.1
CoMFormer44.479.636.736.273.429.7
Joint47.679.039.547.679.039.5
", + "image_path": "e72086cdd253abe8cb5f101eb96615153796bacf738f2cd0ee264e875ad74b44.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 148, + 545, + 182 + ], + "lines": [ + { + "bbox": [ + 305, + 148, + 545, + 182 + ], + "spans": [ + { + "bbox": [ + 305, + 148, + 545, + 182 + ], + "type": "text", + "content": "Table 5. Panoptic metrics analysis. We report Recognition Quality (RQ), Segmentation Quality (SQ), and Panoptic Quality (PQ) over all the classes after the last step." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 305, + 193, + 545, + 229 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 193, + 545, + 229 + ], + "spans": [ + { + "bbox": [ + 305, + 193, + 545, + 229 + ], + "type": "text", + "content": "(KD Eq. (1)) with our adaptive distillation (AD Eq. (3)), we largely improve the performance on both settings: " + }, + { + "bbox": [ + 305, + 193, + 545, + 229 + ], + "type": "inline_equation", + "content": "+0.3" + }, + { + "bbox": [ + 305, + 193, + 545, + 229 + ], + "type": "text", + "content": " PQ in 100-50 and " + }, + { + "bbox": [ + 305, + 193, + 545, + 229 + ], + "type": "inline_equation", + "content": "+3.7" + }, + { + "bbox": [ + 305, + 193, + 545, + 229 + ], + "type": "text", + "content": " PQ in 100-10 considering all classes." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 235, + 546, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 235, + 546, + 475 + ], + "spans": [ + { + "bbox": [ + 304, + 235, + 546, + 475 + ], + "type": "text", + "content": "Forgetting in CoMFormer. To investigate the reason for the forgetting, we considered separately the Recognition Quality (RQ) and the Segmentation Quality (SQ) in panoptic segmentation. RQ measures how well the model is able to classify segments, while SQ considers the precision of the segmentation. Note that, as reported in [28], SQ and RQ are not independent since SQ is computed only on segments correctly classified. The results on the 100-50 and 100-10 are reported in Tab. 5. The table clearly shows that, as hypothesized in Sec. 3.3, forgetting happens heavily on the classification ability while it is less marked on the segmentation ability of the model. CoMFormer, in fact, without providing regularization losses on the mask prediction, is able to match the segmentation ability of the Joint upper bound on the 100-50 and it gets close results on the 100-10 (-5.6). In addition, it is evident that classification ability is heavily affected by forgetting. CoMFormer has a gap with Joint of 3.2 and 11.4 RQ respectively in the 100-50 and 100-10. We note that, however, other baselines are more affected by forgetting, showcasing even larger gaps." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 487, + 378, + 500 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 487, + 378, + 500 + ], + "spans": [ + { + "bbox": [ + 306, + 487, + 378, + 500 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 509, + 545, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 509, + 545, + 653 + ], + "spans": [ + { + "bbox": [ + 304, + 509, + 545, + 653 + ], + "type": "text", + "content": "In this paper, we extend the task of continual learning beyond semantic segmentation by introducing the challenging and realistic continual panoptic segmentation setting. We propose CoMFormer, a strategy to address segmentation as a mask classification problem exploiting the properties of transformers to tackle both semantic and panoptic segmentation tasks. The extensive quantitative results highlight that CoMFormer outperforms all previous methods, both in semantic and panoptic segmentation, under a unified paradigm. In future work, we aim to extend the benchmark to instance segmentation to show that CoMFormer can be employed to address any continual segmentation task." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 658, + 389, + 668 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 658, + 389, + 668 + ], + "spans": [ + { + "bbox": [ + 306, + 658, + 389, + 668 + ], + "type": "text", + "content": "Acknowledgments." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 669, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 669, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 669, + 545, + 713 + ], + "type": "text", + "content": "We gratefully acknowledge the HPC infrastructure and Team at IIT. Fabio Cermelli acknowledges travel support from ELISE (GA no 951847). This work was partly supported by ANR grant VISA DEEP (ANR-20-CHIA-0022)." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3017" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 96, + 106, + 108 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 96, + 106, + 108 + ], + "spans": [ + { + "bbox": [ + 48, + 96, + 106, + 108 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 117, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 53, + 117, + 287, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 117, + 287, + 160 + ], + "spans": [ + { + "bbox": [ + 53, + 117, + 287, + 160 + ], + "type": "text", + "content": "[1] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In ECCV, pages 213-229. Springer, 2020. (pages 2, 5)." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 162, + 288, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 162, + 288, + 206 + ], + "spans": [ + { + "bbox": [ + 53, + 162, + 288, + 206 + ], + "type": "text", + "content": "[2] Fabio Cermelli, Dario Fontanel, Antonio Tavera, Marco Ciccone, and Barbara Caputo. Incremental learning in semantic segmentation from image labels. In CVPR, pages 4371-4381, 2022. (page 2)." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 208, + 288, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 208, + 288, + 262 + ], + "spans": [ + { + "bbox": [ + 53, + 208, + 288, + 262 + ], + "type": "text", + "content": "[3] Fabio Cermelli, Antonino Geraci, Dario Fontanel, and Barbara Caputo. Modeling missing annotations for incremental learning in object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3700-3710, 2022. (pages 1, 2)." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 264, + 288, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 264, + 288, + 319 + ], + "spans": [ + { + "bbox": [ + 53, + 264, + 288, + 319 + ], + "type": "text", + "content": "[4] Fabio Cermelli, Massimiliano Mancini, Samuel Rota Bulò, Elisa Ricci, and Barbara Caputo. Modeling the background for incremental learning in semantic segmentation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2020. (pages 1, 2, 3, 4, 5, 6)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 320, + 287, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 320, + 287, + 365 + ], + "spans": [ + { + "bbox": [ + 53, + 320, + 287, + 365 + ], + "type": "text", + "content": "[5] Fabio Cermelli, Massimiliano Mancini, Yongqin Xian, Zeynep Akata, and Barbara Caputo. Prototype-based incremental few-shot segmentation. In Proceedings of the British Machine Vision Conference (BMVC), 2021. (page 2)." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 365, + 287, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 365, + 287, + 409 + ], + "spans": [ + { + "bbox": [ + 53, + 365, + 287, + 409 + ], + "type": "text", + "content": "[6] Arslan Chaudhry, Marc'Aurelio Ranzato, Marcus Rohrbach, and Mohamed Elhoseiny. Efficient lifelong learning with a-gem. In Proceedings of the International Conference on Learning Representations (ICLR), 2019. (page 2)." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 411, + 287, + 476 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 411, + 287, + 476 + ], + "spans": [ + { + "bbox": [ + 53, + 411, + 287, + 476 + ], + "type": "text", + "content": "[7] Hongwei Chen, Laihui Ding, Fengqin Yao, Pengfei Ren, and Shengke Wang. Panoptic segmentation of UAV images with deformable convolution network and mask scoring. In Twelfth International Conference on Graphics and Image Processing (ICGIP 2020), volume 11720, pages 312-321. SPIE, 2021. (page 1)." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 478, + 287, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 478, + 287, + 544 + ], + "spans": [ + { + "bbox": [ + 53, + 478, + 287, + 544 + ], + "type": "text", + "content": "[8] Liang-Chieh Chen, George Papandreou, Iasonas Kokkinos, Kevin Murphy, and Alan L. Yuille. Deeplab: Semantic image segmentation with deep convolutional nets, atrous convolution, and fully connected crfs. In IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 2018. (page 2)." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 545, + 287, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 545, + 287, + 588 + ], + "spans": [ + { + "bbox": [ + 53, + 545, + 287, + 588 + ], + "type": "text", + "content": "[9] Liang-Chieh Chen, George Papandreou, Florian Schroff, and Hartwig Adam. Rethinking atrous convolution for semantic image segmentation. In arXiv preprint library, 2017. (pages 2, 5, 6, 7)." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 590, + 287, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 590, + 287, + 645 + ], + "spans": [ + { + "bbox": [ + 48, + 590, + 287, + 645 + ], + "type": "text", + "content": "[10] Bowen Cheng, Maxwell D Collins, Yukun Zhu, Ting Liu, Thomas S Huang, Hartwig Adam, and Liang-Chieh Chen. Panoptic-deeplab: A simple, strong, and fast baseline for bottom-up panoptic segmentation. In CVPR, pages 12475-12485, 2020. (page 2)." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 647, + 287, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 647, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 647, + 287, + 689 + ], + "type": "text", + "content": "[11] Bowen Cheng, Ishan Misra, Alexander G Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. In CVPR, 2022. (pages 2, 3, 5, 6, 7, 8)." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "type": "text", + "content": "[12] Bowen Cheng, Alexander G. Schwing, and Alexander Kirillov. Per-pixel classification is not all you need for semantic" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 74, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 327, + 74, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 74, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 327, + 74, + 545, + 95 + ], + "type": "text", + "content": "segmentation. In Adv. Neural Inform. Process. Syst., 2021. (pages 2, 3, 6)." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 308, + 98, + 545, + 120 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 98, + 545, + 120 + ], + "spans": [ + { + "bbox": [ + 308, + 98, + 545, + 120 + ], + "type": "text", + "content": "[13] Lee R. Dice. Measures of the amount of ecologic association between species. Ecology, 1945. (page 5)." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 121, + 545, + 198 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 121, + 545, + 198 + ], + "spans": [ + { + "bbox": [ + 308, + 121, + 545, + 198 + ], + "type": "text", + "content": "[14] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In Proceedings of the International Conference on Learning Representations (ICLR), 2021. (page 2)." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 200, + 545, + 255 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 200, + 545, + 255 + ], + "spans": [ + { + "bbox": [ + 308, + 200, + 545, + 255 + ], + "type": "text", + "content": "[15] Arthur Douillard, Yifu Chen, Arnaud Dapogny, and Matthieu Cord. Plop: Learning without forgetting for continual semantic segmentation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2021. (pages 1, 2, 4, 5, 6)." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 257, + 545, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 257, + 545, + 300 + ], + "spans": [ + { + "bbox": [ + 308, + 257, + 545, + 300 + ], + "type": "text", + "content": "[16] Arthur Douillard, Yifu Chen, Arnaud Dapogny, and Matthieu Cord. Tackling catastrophic forgetting and background shift in continual semantic segmentation. In arXiv preprint library, 2021. (page 2)." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 303, + 545, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 303, + 545, + 357 + ], + "spans": [ + { + "bbox": [ + 308, + 303, + 545, + 357 + ], + "type": "text", + "content": "[17] Arthur Douillard, Matthieu Cord, Charles Ollion, Thomas Robert, and Eduardo Valle. Podnet: Pooled outputs distillation for small-tasks incremental learning. In Proceedings of the IEEE European Conference on Computer Vision (ECCV), 2020. (pages 1, 2)." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 359, + 545, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 359, + 545, + 403 + ], + "spans": [ + { + "bbox": [ + 308, + 359, + 545, + 403 + ], + "type": "text", + "content": "[18] Arthur Douillard and Timothee Lesort. Continuum: Simple management of complex continual learning scenarios. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR) Workshop, 2021. (page 5)." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 406, + 545, + 449 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 406, + 545, + 449 + ], + "spans": [ + { + "bbox": [ + 308, + 406, + 545, + 449 + ], + "type": "text", + "content": "[19] Arthur Douillard, Alexandre Rame, Guillaume Couairon, and Matthieu Cord. Dytox: Transformers for continual learning with dynamic token expansion. In CVPR, 2022. (pages 1, 2, 3)." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 451, + 545, + 506 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 451, + 545, + 506 + ], + "spans": [ + { + "bbox": [ + 308, + 451, + 545, + 506 + ], + "type": "text", + "content": "[20] Mark Everingham, S. M. Ali Eslami, Luc Van Gool, Christopher K. I. Williams, John M. Winn, and Andrew Zisserman. The pascal visual object classes challenge: A retrospective. In International Journal of Computer Vision (IJCV), 2015. (page 5)." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 508, + 545, + 529 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 508, + 545, + 529 + ], + "spans": [ + { + "bbox": [ + 308, + 508, + 545, + 529 + ], + "type": "text", + "content": "[21] Robert French. Catastrophic forgetting in connectionist networks. Trends in cognitive sciences, 1999. (pages 1, 2, 4)." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 532, + 545, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 532, + 545, + 586 + ], + "spans": [ + { + "bbox": [ + 308, + 532, + 545, + 586 + ], + "type": "text", + "content": "[22] Jun Fu, Jing Liu, Hajjie Tian, Yong Li, Yongjun Bao, Zhiwei Fang, and Hanqing Lu. Dual attention network for scene segmentation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2019. (page 2)." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 589, + 545, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 589, + 545, + 632 + ], + "spans": [ + { + "bbox": [ + 308, + 589, + 545, + 632 + ], + "type": "text", + "content": "[23] K. He, X. Zhang, S. Ren, and J. Sun. Deep residual learning for image recognition. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016. (page 6)." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 635, + 545, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 635, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 308, + 635, + 545, + 689 + ], + "type": "text", + "content": "[24] Zilong Huang, Xinggang Wang, Yunchao Wei, Lichao Huang, Humphrey Shi, Wenyu Liu, and Thomas S. Huang. Cynet: Criss-cross attention for semantic segmentation. In IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 2020. (page 2)." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 691, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 691, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 691, + 545, + 713 + ], + "type": "text", + "content": "[25] Loshchilov Ilya, Hutter Frank, et al. Decoupled weight decay regularization. Proceedings of ICLR, 2019. (page 6)." + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3018" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 117 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 117 + ], + "type": "text", + "content": "[26] Jitesh Jain, Anukriti Singh, Nikita Orlov, Zilong Huang, Jiachen Li, Steven Walton, and Humphrey Shi. Semask: Semantically masked transformers for semantic segmentation. arXiv preprint arXiv:2112.12782, 2021. (page 2)." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 118, + 287, + 152 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 118, + 287, + 152 + ], + "spans": [ + { + "bbox": [ + 49, + 118, + 287, + 152 + ], + "type": "text", + "content": "[27] Alexander Kirillov, Ross Girshick, Kaiming He, and Piotr Dólar. Panoptic feature pyramid networks. In CVPR, pages 6399-6408, 2019. (page 2)." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 152, + 287, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 152, + 287, + 185 + ], + "spans": [ + { + "bbox": [ + 48, + 152, + 287, + 185 + ], + "type": "text", + "content": "[28] Alexander Kirillov, Kaiming He, Ross Girshick, Carsten Rother, and Piotr Dólár. Panoptic segmentation. In CVPR, pages 9404-9413, 2019. (pages 2, 5, 8)." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 186, + 287, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 186, + 287, + 262 + ], + "spans": [ + { + "bbox": [ + 48, + 186, + 287, + 262 + ], + "type": "text", + "content": "[29] James Kirkpatrick, Razvan Pascanu, Neil Rabinowitz, Joel Veness, Guillaume Desjardins, Andrei A. Rusu, Kieran Milan, John Quan, Tiago Ramalho, Agnieszka Grabska-Barwinska, Demis Hassabis, Claudia Clopath, Dharshan Kumaran, and Raia Hadsell. Overcoming catastrophic forgetting in neural networks. Proceedings of the National Academy of Sciences, 2017. (pages 1, 2)." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 264, + 287, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 264, + 287, + 297 + ], + "spans": [ + { + "bbox": [ + 48, + 264, + 287, + 297 + ], + "type": "text", + "content": "[30] Harold W. Kuhn. The hungarian method for the assignment problem. Naval Research Logistics Quarterly, 1955. (page 5)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 298, + 287, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 298, + 287, + 342 + ], + "spans": [ + { + "bbox": [ + 48, + 298, + 287, + 342 + ], + "type": "text", + "content": "[31] Duo Li, Guimei Cao, Yunlu Xu, Zhanzhan Cheng, and Yi Niu. Technical report for iccv 2021 challenge sslad-track3b: Transformers are better continual learners. In ICCV 2021 Challenge SSLAD-Track3B, 2021. (page 3)." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 343, + 287, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 343, + 287, + 387 + ], + "spans": [ + { + "bbox": [ + 48, + 343, + 287, + 387 + ], + "type": "text", + "content": "[32] Yanwei Li, Hengshuang Zhao, Xiaojuan Qi, Yukang Chen, Lu Qi, Liwei Wang, Zeming Li, Jian Sun, and Jiaya Jia. Fully convolutional networks for panoptic segmentation with point-based supervision. IEEE TPAMI, 2022. (page 2)." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 388, + 287, + 422 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 388, + 287, + 422 + ], + "spans": [ + { + "bbox": [ + 48, + 388, + 287, + 422 + ], + "type": "text", + "content": "[33] Z. Li and D. Hoiem. Learning without forgetting. Proceedings of the IEEE European Conference on Computer Vision (ECCV), 2016. (pages 1, 2)." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 422, + 287, + 466 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 422, + 287, + 466 + ], + "spans": [ + { + "bbox": [ + 48, + 422, + 287, + 466 + ], + "type": "text", + "content": "[34] Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dólár. Focal loss for dense object detection. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), 2017. (page 5)." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 467, + 287, + 522 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 467, + 287, + 522 + ], + "spans": [ + { + "bbox": [ + 48, + 467, + 287, + 522 + ], + "type": "text", + "content": "[35] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and Lawrence Zitnick. Microsoft coco: Common objects in context. In Proceedings of the IEEE European Conference on Computer Vision (ECCV), 2014. (page 5)." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 523, + 287, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 523, + 287, + 578 + ], + "spans": [ + { + "bbox": [ + 48, + 523, + 287, + 578 + ], + "type": "text", + "content": "[36] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), 2021. (page 2)." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 579, + 287, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 579, + 287, + 624 + ], + "spans": [ + { + "bbox": [ + 48, + 579, + 287, + 624 + ], + "type": "text", + "content": "[37] J. Long, E. Shelhamer, and T. Darrell. Fully convolutional networks for semantic segmentation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2015. (page 2)." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 624, + 287, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 624, + 287, + 669 + ], + "spans": [ + { + "bbox": [ + 48, + 624, + 287, + 669 + ], + "type": "text", + "content": "[38] Andrea Maracani, Umberto Michieli, Marco Toldo, and Pietro Zanuttigh. Recall: Replay-based continual learning in semantic segmentation. In ICCV, pages 7026-7035, 2021. (page 2)." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 670, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 670, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 670, + 287, + 713 + ], + "type": "text", + "content": "[39] Umberto Michieli and Pietro Zanuttigh. Incremental learning techniques for semantic segmentation. In Proceedings of the IEEE International Conference on Computer Vision (ICCV) Workshop, 2019. (pages 1, 2, 4)." + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 117 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 117 + ], + "type": "text", + "content": "[40] Umberto Michieli and Pietro Zanuttigh. Continual semantic segmentation via repulsion-attraction of sparse and disentangled latent representations. In CVPR, pages 1114-1124, 2021. (page 2)." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 118, + 545, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 545, + 172 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 545, + 172 + ], + "type": "text", + "content": "[41] Gerhard Neuhold, Tobias Ollmann, Samuel Rota Bulò, and Peter Kontschieder. The mapillary vistas dataset for semantic understanding of street scenes. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), 2017. (page 1)." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 172, + 545, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 172, + 545, + 205 + ], + "spans": [ + { + "bbox": [ + 307, + 172, + 545, + 205 + ], + "type": "text", + "content": "[42] Lorenzo Porzi, Samuel Rota Bulò, Aleksander Colovic, and Peter Kontschieder. Seamless scene segmentation. In CVPR, June 2019. (page 2)." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 205, + 545, + 260 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 205, + 545, + 260 + ], + "spans": [ + { + "bbox": [ + 308, + 205, + 545, + 260 + ], + "type": "text", + "content": "[43] Sylvestre-Alvise Rebuffi, Alexander Kolesnikov, Georg Sperl, and Christoph H. Lampert. icarl: Incremental classifier and representation learning. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2017. (pages 1, 2)." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 261, + 545, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 261, + 545, + 282 + ], + "spans": [ + { + "bbox": [ + 308, + 261, + 545, + 282 + ], + "type": "text", + "content": "[44] Anthony Robins. Catastrophic forgetting, rehearsal and pseudorehearsal. Connection Science, 1995. (page 2)." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 283, + 545, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 283, + 545, + 327 + ], + "spans": [ + { + "bbox": [ + 308, + 283, + 545, + 327 + ], + "type": "text", + "content": "[45] Andrei A Rusu, Neil C Rabinowitz, Guillaume Desjardins, Hubert Soyer, James Kirkpatrick, Koray Kavukcuoglu, Razvan Pascanu, and Raia Hadsell. Progressive neural networks. arXiv preprint library, 2016. (pages 1, 2)." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 327, + 545, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 327, + 545, + 360 + ], + "spans": [ + { + "bbox": [ + 308, + 327, + 545, + 360 + ], + "type": "text", + "content": "[46] Robin Strudel, Ricardo Garcia, Ivan Laptev, and Cordelia Schmid. Segmenter: Transformer for semantic segmentation. In ICCV, pages 7262-7272, 2021. (page 2)." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 361, + 545, + 382 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 361, + 545, + 382 + ], + "spans": [ + { + "bbox": [ + 308, + 361, + 545, + 382 + ], + "type": "text", + "content": "[47] Sebastian Thrun. Lifelong learning algorithms. In Springer Learning to Learn, 1998. (page 2)." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 383, + 545, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 383, + 545, + 437 + ], + "spans": [ + { + "bbox": [ + 308, + 383, + 545, + 437 + ], + "type": "text", + "content": "[48] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Hervé Jégou. Training data-efficient image transformers and distillation through attention. In International Conference on Machine Learning (ICML), 2021. (page 2)." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 438, + 545, + 482 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 438, + 545, + 482 + ], + "spans": [ + { + "bbox": [ + 308, + 438, + 545, + 482 + ], + "type": "text", + "content": "[49] Hugo Touvron, Matthieu Cord, Alexandre Sablayrolles, Gabriel Synnaeve, and Hervé Jégou. Going deeper with image transformers. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), 2021. (page 2)." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 483, + 545, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 483, + 545, + 525 + ], + "spans": [ + { + "bbox": [ + 308, + 483, + 545, + 525 + ], + "type": "text", + "content": "[50] Huiyu Wang, Yukun Zhu, Hartwig Adam, Alan Yuille, and Liang-Chieh Chen. Max-deeplab: End-to-end panoptic segmentation with mask transformers. In CVPR, pages 5463-5474, 2021. (page 2)." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 526, + 545, + 570 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 526, + 545, + 570 + ], + "spans": [ + { + "bbox": [ + 308, + 526, + 545, + 570 + ], + "type": "text", + "content": "[51] Huiyu Wang, Yukun Zhu, Bradley Green, Hartwig Adam, Alan Yuille, and Liang-Chieh Chen. Axial-deeplab: Standalone axial-attention for panoptic segmentation. In ECCV, pages 108–126. Springer, 2020. (page 2)." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 571, + 545, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 571, + 545, + 624 + ], + "spans": [ + { + "bbox": [ + 308, + 571, + 545, + 624 + ], + "type": "text", + "content": "[52] Zifeng Wang, Zizhao Zhang, Sayna Ebrahimi, Ruoxi Sun, Han Zhang, Chen-Yu Lee, Xiaoqi Ren, Guolong Su, Vincent Perot, Jennifer Dy, et al. Dualprompt: Complementary prompting for rehearsal-free continual learning. ECCV, 2022. (page 2)." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 625, + 545, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 625, + 545, + 669 + ], + "spans": [ + { + "bbox": [ + 308, + 625, + 545, + 669 + ], + "type": "text", + "content": "[53] Zifeng Wang, Zizhao Zhang, Chen-Yu Lee, Han Zhang, Ruoxi Sun, Xiaoqi Ren, Guolong Su, Vincent Perot, Jennifer Dy, and Tomas Pfister. Learning to prompt for continual learning. In CVPR, pages 139-149, 2022. (page 2)." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 308, + 670, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 670, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 670, + 545, + 713 + ], + "type": "text", + "content": "[54] Enze Xie, Wenhai Wang, Zhiding Yu, Anima Anandkumar, Jose M Alvarez, and Ping Luo. Segformer: Simple and efficient design for semantic segmentation with transformers. Adv. Neural Inform. Process. Syst., 34, 2021. (page 2)." + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3019" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 475 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 117 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 117 + ], + "type": "text", + "content": "[55] Shipeng Yan, Jiangwei Xie, and Xuming He. Der: Dynamically expandable representation for class incremental learning. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2021. (page 2)." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 118, + 287, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 118, + 287, + 171 + ], + "spans": [ + { + "bbox": [ + 49, + 118, + 287, + 171 + ], + "type": "text", + "content": "[56] Qihang Yu, Huiyu Wang, Siyuan Qiao, Maxwell Collins, Yukun Zhu, Hartwig Adam, Alan Yuille, and Liang-Chieh Chen. k-means mask transformer. In European Conference on Computer Vision, pages 288-307. Springer, 2022. (page 2)." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 173, + 288, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 173, + 288, + 217 + ], + "spans": [ + { + "bbox": [ + 49, + 173, + 288, + 217 + ], + "type": "text", + "content": "[57] Yuhui Yuan, Xilin Chen, and Jingdong Wang. Object-contextual representations for semantic segmentation. In Proceedings of the IEEE European Conference on Computer Vision (ECCV), 2020. (page 2)." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 219, + 287, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 219, + 287, + 251 + ], + "spans": [ + { + "bbox": [ + 49, + 219, + 287, + 251 + ], + "type": "text", + "content": "[58] Yuhui Yuan and Jingdong Wang. Ocnet: Object context network for scene parsing. In arXiv preprint library, 2018. (page 2)." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 49, + 253, + 287, + 317 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 253, + 287, + 317 + ], + "spans": [ + { + "bbox": [ + 49, + 253, + 287, + 317 + ], + "type": "text", + "content": "[59] Chang-Bin Zhang, Jia-Wen Xiao, Xialei Liu, Ying-Cong Chen, and Ming-Ming Cheng. Representation compensation networks for continual semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7053-7064, 2022. (pages 1, 2, 5, 6, 7)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 49, + 319, + 287, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 319, + 287, + 363 + ], + "spans": [ + { + "bbox": [ + 49, + 319, + 287, + 363 + ], + "type": "text", + "content": "[60] H. Zhao, J. Shi, X. Qi, X. Wang, and J. Jia. Pyramid scene parsing network. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2017. (page 2)." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 49, + 364, + 287, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 364, + 287, + 418 + ], + "spans": [ + { + "bbox": [ + 49, + 364, + 287, + 418 + ], + "type": "text", + "content": "[61] Hengshuang Zhao, Yi Zhang, Shu Liu, Jianping Shi, Chen Change Loy, Dahua Lin, and Jiaya Jia. Pisanet: Pointwise spatial attention network for scene parsing. In Proceedings of the IEEE European Conference on Computer Vision (ECCV), 2018. (page 2)." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 49, + 420, + 287, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 420, + 287, + 475 + ], + "spans": [ + { + "bbox": [ + 49, + 420, + 287, + 475 + ], + "type": "text", + "content": "[62] Bolei Zhou, Hang Zhao, Xavier Puig, Sanja Fidler, Adela Barriuso, and Antonio Torralba. Scene parsing through ade20k dataset. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2017. (page 5)." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "3020" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/CoWs on Pasture_ Baselines and Benchmarks for Language-Driven Zero-Shot Object Navigation/a31a35f2-4957-4777-92dc-31eae2e5deb5_content_list.json b/2023/CoWs on Pasture_ Baselines and Benchmarks for Language-Driven Zero-Shot Object Navigation/a31a35f2-4957-4777-92dc-31eae2e5deb5_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..1b260168665bbe623b383e414211086739e01f06 --- /dev/null +++ b/2023/CoWs on Pasture_ Baselines and Benchmarks for Language-Driven Zero-Shot Object Navigation/a31a35f2-4957-4777-92dc-31eae2e5deb5_content_list.json @@ -0,0 +1,1722 @@ +[ + { + "type": "text", + "text": "CoWs on PASTURE: Baselines and Benchmarks for Language-Driven Zero-Shot Object Navigation", + "text_level": 1, + "bbox": [ + 228, + 130, + 743, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Samir Yitzhak Gadre $^{\\diamond}$ Mitchell Wortsman $^{\\dagger}$ Gabriel Ilharco $^{\\dagger}$ Ludwig Schmidt $^{\\dagger}$ Shuran Song $^{\\diamond}$", + "bbox": [ + 88, + 202, + 875, + 223 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 255, + 313, + 272 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "For robots to be generally useful, they must be able to find arbitrary objects described by people (i.e., be language-driven) even without expensive navigation training on in-domain data (i.e., perform zero-shot inference). We explore these capabilities in a unified setting: language-driven zero-shot object navigation (L-ZSON). Inspired by the recent success of open-vocabulary models for image classification, we investigate a straightforward framework, CLIP on Wheels (CoW), to adapt open-vocabulary models to this task without fine-tuning. To better evaluate L-ZSON, we introduce the PASTURE benchmark, which considers finding uncommon objects, objects described by spatial and appearance attributes, and hidden objects described relative to visible objects. We conduct an in-depth empirical study by directly deploying 22 CoW baselines across HABITAT, ROBOTHOR, and PASTURE. In total, we evaluate over 90k navigation episodes and find that (1) CoW baselines often struggle to leverage language descriptions but are proficient at finding uncommon objects. (2) A simple CoW, with CLIP-based object localization and classical exploration—and no additional training—matches the navigation efficiency of a state-of-the-art ZSON method trained for 500M steps on HABITAT MP3D data. This same CoW provides a 15.6 percentage point improvement in success over a state-of-the-art ROBOTHOR ZSON model.", + "bbox": [ + 75, + 287, + 473, + 666 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 691, + 209, + 709 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To be more widely applicable, robots should be language-driven: able to deduce goals based on arbitrary text input instead of being constrained to a fixed set of object categories. While existing image classification, semantic segmentation, and object navigation benchmarks like ImageNet-1k [65], ImageNet-21k [22], MS-COCO [45], LVIS [28], HABITAT [67], and ROBOTHOR [18] include a vast array of everyday items, they do not capture all objects that matter to people. For instance, a lost \"toy airplane\" may", + "bbox": [ + 75, + 718, + 468, + 854 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Sample tasks Top-down visualization Egocentric Observations", + "text_level": 1, + "bbox": [ + 506, + 253, + 859, + 266 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "(a) Finding uncommon objects", + "bbox": [ + 503, + 268, + 630, + 279 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/16652846a9fca43c2db6cb634fa5e8d0256f61a82637d3a954319e5ebb771b2f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 504, + 280, + 888, + 330 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "(b) Finding objects based on attributes in the presence of distractors", + "bbox": [ + 503, + 333, + 782, + 344 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/31f0ab0fe5e9a07a5095a87d64338440d54a55360dd693f2b489171f040d34fa.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 504, + 345, + 888, + 395 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "(c) Finding hidden objects in the presence of distractors", + "bbox": [ + 504, + 397, + 735, + 407 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/866fb55efa25e9630e6e03edbb445623423d7bcd17120fc94c04cd76ad751fdc.jpg", + "image_caption": [ + "Figure 1. The PASTURE benchmark for L-ZSON. Text specifies navigation goal objects. Agents do not train on these tasks, making the evaluation protocol zero-shot. (a) Uncommon object goals like \"llama wicker basket\", not found in existing navigation benchmarks. (b) Appearance, spatial descriptions, which are necessary to find the correct object. (c) Hidden object descriptions, which localize objects that are not visible." + ], + "image_footnote": [], + "bbox": [ + 504, + 407, + 888, + 459 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "become relevant in a kindergarten classroom, but this object is not annotated in any of the aforementioned datasets.", + "bbox": [ + 496, + 579, + 893, + 609 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this paper, we study Language-driven zero-shot object navigation (L-ZSON), a more challenging but also more applicable version of object navigation [5, 18, 67, 79, 89] and ZSON [38, 46] tasks. In L-ZSON, an agent must find an object based on a description, which may contain different levels of granularity (e.g., \"toy airplane\", \"toy airplane under the bed\", or \"wooden toy airplane\"). L-ZSON encompasses ZSON, which specifies only the target category [38, 46]. Since L-ZSON is \"zero-shot\", we consider agents without access to navigation training on the target objects or domains. This reflects realistic application scenarios, where the environment and object set may not be known a priori.", + "bbox": [ + 496, + 611, + 893, + 792 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Performing L-ZSON in any environment with unstructured language input is challenging; however, recent advances in open-vocabulary models for image classification [35, 58, 61], object detection [4, 21, 27, 36, 43, 47, 49, 62, 88], and semantic segmentation [3, 6, 15, 33, 36, 37, 86] present a promising foundation. These models provide an interface where one specifies—in text—the arbitrary ob", + "bbox": [ + 496, + 795, + 893, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 44 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "$^{\\text{©}}$ Columbia University, $^{\\dagger}$ University of Washington. Correspondence to sy@cs.columbia.edu.", + "bbox": [ + 76, + 862, + 468, + 886 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "1For code, data, and videos, see cow.cs.columbia.edu/", + "bbox": [ + 96, + 887, + 424, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "23171", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "jects they wish to classify, detect, or segment. For example, CLIP [61] open-vocabulary classifiers compute similarity scores between an input image and a set of user-specified captions (e.g., \"a photo of a toy airplane.\", ..., selecting the caption with the highest score to determine the image classification label. Given the flexibility of these models, we would like to understand their capability to execute embodied tasks even without additional training.", + "bbox": [ + 75, + 90, + 468, + 210 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To this end, we present baselines and benchmarks for L-ZSON. More specifically:", + "bbox": [ + 76, + 212, + 468, + 244 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- A collection of baseline algorithms, CLIP on Wheels (CoW), which adapt open-vocabulary models to the task of L-ZSON. CoW takes inspiration from the semantic mapping line of work [11, 41, 53], and decomposes the navigation task into exploration when the language target is not confidently localized, and target-driven planning otherwise. CoW retains the textual user interface of the original open-vocabulary model and does not require any navigation training. We evaluate 22 CoWs, ablating over many open-vocabulary models, exploration policies, backbones, prompting strategies, and post-processing strategies.", + "- A new benchmark, PASTURE, to evaluate CoW and future methods on L-ZSON. We design PASTURE, visualized in Fig. 1, to study capabilities that traditional object navigation agents, which are trained on a fixed set of categories, do not possess. We consider the ability to find (1) uncommon objects (e.g., \"tie-dye surfboard\"), (2) objects by their spatial and appearance attributes in the presence of distractor objects (e.g., \"green apple\" vs. \"red apple\"), and (3) objects that cannot be visually observed (e.g., \"mug under the bed\")." + ], + "bbox": [ + 76, + 250, + 472, + 588 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Together the CoW baselines and PASTURE benchmark allow us to conduct extensive studies on the capabilities of open-vocabulary models in the context of L-ZSON embodied tasks. Our experiments demonstrate CoW's potential on uncommon objects and limitations in taking full advantage of language descriptions—thereby providing empirical motivation for future studies. To contextualize CoW relative to prior zero-shot methods, we additionally evaluate on the HABITAT MP3D dataset. We find that our best CoW achieves navigation efficiency (SPL) that matches a state-of-the-art ZSON method [46] that trains on MP3D training data for 500M steps. On a ROBOTHOR object subset, considered in prior work, the same CoW beats the leading method [38] by 15.6 percentage points in task success.", + "bbox": [ + 75, + 597, + 470, + 809 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 76, + 825, + 218, + 840 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Mapping and exploration. Exploring effectively with a mobile robot is a long-standing problem in vision and robotics. Classical methods often decompose the task", + "bbox": [ + 75, + 854, + 468, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "into map reconstruction [30, 32, 51, 52, 72], agent localization [17, 20, 54], and planning [41, 78]. Recent work investigates learned alternatives for exploration [7, 14, 56, 57, 63]. Here, agents are often trained end-to-end with self-supervised rewards (e.g., curiosity [57]) or supervised rewards (e.g., state visitation counts [25, 73, 75]). Learning-based methods typically need less hand-tuning, but require millions of training steps and reward engineering. We test both classical and learnable exploration strategies in the context of CoW to study their applicability to L-ZSON.", + "bbox": [ + 496, + 90, + 890, + 242 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Goal-conditioned navigation. Apart from open-ended exploration, many navigation tasks are goal-conditioned, where the agent needs to navigate to a specified position (i.e., point goal [11, 12, 26, 31, 66, 77, 81, 83]), view of the environment (i.e., image goal [48, 64, 89]), or object category (i.e., object goal [1, 9, 10, 12, 19, 44, 74, 79, 84]). We consider an object goal navigation task.", + "bbox": [ + 496, + 246, + 890, + 351 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Vision-Language Navigation. Prior work investigates language-based navigation, where language provides step-by-step instructions for the task [2, 34, 39, 40, 71]. This line of work demonstrates the benefits of additional language input for robot navigation, especially for long-horizon tasks (e.g., room-to-room navigation [40]). However, providing detailed step-by-step instructions (e.g., move 3 meters south [34]) could be challenging and time-consuming. In our L-ZSON task, an algorithm gets natural language as the goal description instead of low-level instructions. Prior work also investigates navigation with target descriptions in supervised settings [42, 60, 87]. In contrast, we explore a zero-shot evaluation protocol and consider finding hidden objects (e.g., \"mug under bed\") and uncommon objects (e.g., \"tie-dye surfboard\").", + "bbox": [ + 496, + 356, + 890, + 583 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Zero-shot object navigation (ZSON). Recent work studies object navigation in zero-shot settings, where agents are evaluated on object categories that they are not explicitly trained on [38, 46]. Our task encompasses ZSON; however it also considers cases where more information—object attributes or hidden objects descriptions—is specified. Khandelwal et al. [38] train on a subset of ROBOTHOR categories and evaluate on a held-out set. In concurrent work, Majumdar et al. [46] train on an image goal navigation task and evaluate on object navigation downstream by leveraging CLIP multi-modal embeddings. Both algorithms necessitate navigation training for millions of steps and train separate models for each simulation domain. In contrast, CoW baselines do not necessitate any simulation training and can be deployed in multiple environments.", + "bbox": [ + 496, + 587, + 890, + 813 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3. The L-ZSON Task", + "text_level": 1, + "bbox": [ + 500, + 828, + 679, + 844 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Language-driven zero-shot object navigation (L-ZSON) involves navigating to goal objects, specified in language, without explicit training to do so. Let $\\mathcal{O}$ denote a set of", + "bbox": [ + 496, + 854, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "23172", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/432331edfae158dc49e0cd33234c1db9de98c336e4e79390c16442d512f8dfb2.jpg", + "image_caption": [ + "Figure 2. CLIP on Wheels (CoW) overview. A CoW uses a policy to explore and an object localizer (e.g., an open-vocabulary object detector) to determine if an object goal is in view.", + "Figure 3. Mapping. Top-down map created from egocentric depth observations as a CoW roams a space. (a) Frontier Based Exploration [82] showing a planned path exploration path to the next frontier. (b) Back-projected object relevance scores provide object goal targets when a CoW has found an object." + ], + "image_footnote": [], + "bbox": [ + 80, + 88, + 504, + 214 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "natural language descriptions of target objects with potentially many attributes (e.g., \"plant\", \"snake plant\", \"plant under the bed\", etc.). This contrasts with definitions studied in prior object navigation [5, 18] and ZSON [38, 46] tasks, which focus on high-level categories like \"plant\". Let $S$ denote the set of navigation scenes. Let $p_0$ describe the initial pose of an agent. A navigation episode $\\tau \\in \\mathcal{T}$ is written as a tuple $\\tau = (s, o, p_0)$ , $s \\in S$ , $o \\in \\mathcal{O}$ . Each $\\tau$ is a zero-shot task as tuples of this form are not seen during training. Starting at $p_0$ , an embodied agent's goal is to find $o$ . The agent receives observations and sensor readings $I_t$ (e.g., RGB-D images). At each timestep $t$ , the agent executes a navigation action $a \\in \\mathcal{A}$ . A special action $\\text{STOP} \\in \\mathcal{A}$ terminates the episode. If the agent is within $c$ units of $o$ and $o$ meets a visibility criteria, the episode is successful.", + "bbox": [ + 75, + 272, + 473, + 500 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4. CLIP on Wheels (CoW) Baselines", + "text_level": 1, + "bbox": [ + 76, + 511, + 382, + 527 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To address L-ZSON, we investigate a simple baseline approach, CoW, which adapts open-vocabulary models like CLIP to make them suitable for the task. A CoW takes as input an egocentric RGB-D image and an object goal specified in language. As a CoW moves, it updates a top-down map of the world created using RGB-D observations and pose estimates (Sec. 4.1). Each CoW gets an exploration policy and a zero-shot object localization module as seen in Fig. 2. To observe diverse views of the scene, a CoW explores using a policy (Sec. 4.2). As the CoW roams, it keeps track of its confidence about the target object's location using an object localization module (Sec. 4.3) and its top-down map. When a CoW's confidence exceeds a threshold, it plans to the location of the goal and issues the STOP action. We now describe the ingredients used to make the CoWs evaluated in our experiments (Sec. 6).", + "bbox": [ + 75, + 536, + 470, + 779 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4.1. Depth-based Mapping", + "text_level": 1, + "bbox": [ + 76, + 786, + 284, + 803 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "As a CoW moves, it constructs a top-down map using input depth, pose estimates, and known agent height. The map is initialized using known camera intrinsics and the first depth image. Since a CoW knows the intended consequences of its actions (e.g., MOVEFORWARD should result in a $0.25\\mathrm{m}$ translation), each action is represented as a", + "bbox": [ + 75, + 810, + 468, + 900 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/744a07bb4f5244059a6ff11e7c1a5757595ad2d06b97fe38d12f86a9ccb77226.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 504, + 90, + 694, + 183 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/b530d1ea053a38abe377f49c828f5ba002be7ddfec13acf8b51e25e253fbd124.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 697, + 90, + 887, + 181 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "pose delta transform to approximate a transition. To deal with noise associated with actuation or depth, a CoW maintains a map at $0.125\\mathrm{m}$ resolution. To improve map accuracy, a CoW checks for failed actions by comparing successive depth frames for movements (see Appx. A for details). Using known agent height $(0.9\\mathrm{m})$ , map cells are projected to the ground plane to maintain a top-down representation of the world, which suffices for most navigation applications. Cells close to the floor are considered free space (white points in Fig. 3 (a)), while other cells are considered occupied (blue points in Fig. 3 (a)).", + "bbox": [ + 496, + 271, + 892, + 436 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4.2. Exploration", + "text_level": 1, + "bbox": [ + 500, + 446, + 627, + 463 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Exploration generates diverse egocentric views so a CoW is more likely to view the language-specified target object. We consider two exploration methods, frontier-based and learning-based.", + "bbox": [ + 496, + 470, + 890, + 531 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Frontier based exploration (FBE) [82]. Using the top-down map discussed in Sec. 4.1, a CoW can navigate using a simple exploration heuristic: move to the frontier between free and unknown space to discover new regions. Once the navigator reaches a frontier (visualized as purple points in Fig. 3 (a)), it moves greedily to the next closest frontier. Since the map is updated at every timestep, a noisy pose estimate can contribute to inaccuracies. For example, narrow passages may collapse in the map due to pose drift. To circumvent such problems, we reinitialize the map when no paths exist to any frontiers in the map.", + "bbox": [ + 496, + 534, + 892, + 700 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Learnable exploration. In addition to FBE, we consider learnable alternatives, which may explore more intelligently but incur substantial training costs. We investigate an architecture and reward structure similar to prior work in embodied AI (e.g., [25, 38, 75]). Specifically, we adopt a frozen CLIP backbone with a trainable GRU [16] and linear heads for the actor and critic networks. We train agents independently in HABITAT [67] and ROBOTHOR [18] simulation environments for 60M steps each, using DD-PPO [69, 77] in the AllenAct [76] framework. We employ a simple count-based reward [73]. All training scenes are disjoint from downstream navigation test scenes. For details on reward, hyperparameters, and training, see Appx. B.", + "bbox": [ + 496, + 704, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "23173", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4.3. Object Localization", + "text_level": 1, + "bbox": [ + 76, + 90, + 266, + 107 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Successful navigation depends on object localization: the ability to tell if and where an object is in an image. Regions of high object relevance, extracted from 2D images, are projected to the depth-based map (Fig. 3 (b)) where they serve as natural navigation targets. To determine if and when a target is in an image, we consider the following object localization modules, used in our experiments (Sec. 6). For more details see Appx. C.", + "bbox": [ + 75, + 114, + 468, + 236 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Adapting open-vocabulary classifiers. We experiment with three strategies to turn CLIP [61] models into object localizers. First, we utilize the CLIP text encoder to embed $k$ referring expressions, which specify regions where the target object may appear in the image. For example, \"a plant in the top left of the image.\" We then match the language embeddings against a CLIP visual embedding for the current observation. We compute similarity between the image and text features to determine relevance scores over the regions. Second, we discretize the image into $k$ smaller patches and obtain CLIP patch embeddings. We then convolve each patch embedding with a CLIP text embedding for the target object. If the object is in a patch, the relevance score for that patch should be high. Third, we modify an interpretability method [13, 70] designed to extract object relevancy from vision transformers (ViTs) [24]. Using a target CLIP text embedding and gradient information accumulated through the CLIP vision encoder, we construct a relevance map over input pixels, which qualitatively segments the target when it is in view.", + "bbox": [ + 75, + 239, + 472, + 541 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Adapting open-vocabulary detectors and segmentors. In addition to CLIP-based methods, we consider two additional open-vocabulary models for object localization. First, the MDETR segmentation model [36], which extends the DETR detector [8] to take arbitrary text and images as input and output box detections. The base model is fine-tuned on PhraseCut [80], a dataset of paired masks and attribute descriptions, to support segmentation. Second, we consider the OWL-ViT detector [49], which uses a set prediction fine-tuning recipe to turn CLIP-like models into object detectors. We use this MDETR and OWL-ViT models to directly query images for targets.", + "bbox": [ + 75, + 545, + 468, + 727 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Post-processing. The aforementioned models give continuous valued predictions. However, we are interested in binary masks giving if and where objects are in images. Hence, we threshold predictions for each model (see Appx. C for details). We further investigate two strategies for using the masks downstream: (1) using the entire mask or (2) using the center pixel. The second strategy is reasonable because only part of an object needs to be detected for successful navigation.", + "bbox": [ + 75, + 729, + 468, + 867 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Target driven planning. Recall, CoWs have depth sensors. We back-project object relevance from 2D images into the", + "bbox": [ + 76, + 869, + 468, + 902 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/028d123bae99555101968f4ed45ab80b17727dbff4968acaa1239af1a3cf8653.jpg", + "image_caption": [ + "Figure 4. Uncommon objects in PASTURE." + ], + "image_footnote": [], + "bbox": [ + 503, + 90, + 890, + 213 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "depth-based map (Sec. 4.1). We keep only the max relevance for each map cell (Fig. 3 (b)). CoWs can then plan to high relevance areas in the map. See Appx. D for additional method visualization.", + "bbox": [ + 496, + 244, + 892, + 305 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Incorporating object priors. Since CoW does not train or fine-tune on navigation datasets, we investigate alternative approaches to inject object-level priors into the model. For each target object, we prompt GPT-3.5 [55] to generate rooms where the target objects are likely to be found. For example, GPT-3.5 states that apples are likely to be found in \"kitchen\" or \"dining room\" scenes. Following this prior, a GPT-3.5 enabled CoW first uses its object localization module to localize a kitchen or a dining room, and then looks for an apple. This straightforward extension, demonstrates how outside information can be incorporated into CoW.", + "bbox": [ + 496, + 309, + 892, + 474 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "5. The PASTURE Benchmark", + "text_level": 1, + "bbox": [ + 498, + 488, + 746, + 503 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To evaluate CoW baselines and future methods on L-ZSON, we introduce the PASTURE evaluation benchmark. PASTURE builds on ROBOTHER validation scenes, which have parallel environments in the real-world. We target ROBOTHER to facilitate future real-world benchmarking. PASTURE probes for seven capabilities explained in Sec. 5.1. We provide dataset statistics in Sec. 5.2.", + "bbox": [ + 496, + 513, + 890, + 619 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "5.1. PASTURE Tasks", + "text_level": 1, + "bbox": [ + 500, + 628, + 653, + 643 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "PASTURE evaluates seven core L-ZSON capabilities.", + "bbox": [ + 517, + 651, + 870, + 667 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Uncommon objects. Traditional benchmarks (e.g., ROBOTHOR and HABITAT MP3D) evaluate agents on common object categories like TVs; however, given the rich diversity of objects in homes, we would like to understand navigation performance on uncommon objects. Hence we add 12 new objects to each room. We use names shown in Fig. 4 as instance labels, which are minimal descriptions to identify each object. Some identifiers refer to text in images (e.g., \"whiteboard saying CVPR\") or to appearance attributes (e.g., \"wooden toy airplane\"). Other objects are less common in North America, like \"mate\", which is a popular Argentinian drink.", + "bbox": [ + 496, + 670, + 890, + 851 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Appearance descriptions. To evaluate if baselines can take advantage of visual attributes, we introduce descriptions of the form: “ $\\{size\\}$ , $\\{color\\}$ , $\\{material\\}$ $\\{object\\}$ . For", + "bbox": [ + 496, + 854, + 892, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "23174", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "example: \"small, red apple\", \"orange basketball\", \"small, black, metallic alarm clock\". Objects are considered small if their 3D bounding box diagonal is below a threshold. We determine color and materials by inspection.", + "bbox": [ + 75, + 90, + 468, + 151 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Spatial descriptions. To test if agents can leverage spatial information in navigation, we introduce descriptions: “{object} on top of $\\{x\\}$ , near $\\{y\\}$ , $\\{z\\}$ , ...”. For example, “house plant on a dresser near a spray bottle”. To determine [on top of] relations, we use THOR metadata and to determine [nearness] we use a distance threshold between pairs of objects. We inspect all descriptions for correctness.", + "bbox": [ + 75, + 155, + 467, + 261 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Appearance descriptions with distractors. To probe if appearance attributes better facilitate finding objects in the presence of distractors, we reuse the appearance captions from before, but evaluate on an modified environment with two visually distinct instances of each ROBOTHOR object category. For example, for the task of finding a \"red apple\", we have both a red apple and a green apple in the room. A navigator must leverage appearance information—and not just class information—to successfully complete the task. Distractor objects are sufficiently far from the target objects so that finding a distractor cannot count as success.", + "bbox": [ + 75, + 265, + 467, + 431 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Spatial descriptions with distractors. This capability is similar to the one above; however, we evaluate with spatial descriptions in the presence of distractor objects.", + "bbox": [ + 75, + 435, + 467, + 479 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Hidden object descriptions. An ideal object navigator should find objects, even when they are hidden. Hence, we introduce descriptions: “{object} under/in {x}”. For example, “basketball in the dresser drawers” or “vase under the sofa”. We sample large objects (e.g., beds, sofas, dressers) in each scene to determine [under/in] relations. Additionally we remove visible instances of {object} from the room.", + "bbox": [ + 75, + 484, + 467, + 590 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Hidden object descriptions with distractors. We use the hidden object descriptions from before, but reintroduce visible instances of $\\{object\\}$ to serve as distractors. Consider finding a \"mug under the bed\". A distractor mug will also appear in the scene making the task more challenging.", + "bbox": [ + 75, + 594, + 467, + 669 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.2. Dataset Creation and Statistics", + "text_level": 1, + "bbox": [ + 76, + 680, + 349, + 695 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "PASTURE contains three variations for each of the original 15 validation ROBOTHOR rooms: uncommon objects added, additional object instances added, and target objects removed. For each of the seven settings presented above, we evaluate over 12 object instances in 15 rooms with two initial agent starting locations. Hence PASTURE consists of 2,520 tasks, which is a similar order of magnitude to ROBOTHOR (1,800) and HABITAT MP3D (2,195) validation sets. For appearance attributes, $47\\%$ of the objects are considered \"small\". Each object gets an average of 1.2 color descriptors out of 22 possible choices, and 0.6 material descriptors out of 5 possible choices. Similarly, for spatial attributes, each object gets one object it is on top of or in (e.g.,", + "bbox": [ + 75, + 704, + 467, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "“vase in a shelving unit”) and an average of 1.9 objects it is near. For a sample of appearance and spatial attributes see Fig. 1. For more dataset details and statistics see Appx. E.", + "bbox": [ + 498, + 90, + 890, + 137 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "6. Experiments", + "text_level": 1, + "bbox": [ + 500, + 148, + 632, + 166 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We first present our experimental setup, including the datasets, metrics, embodiment, and baselines considered in our study (Sec. 6.1). Then we present results on PASTURE, thereby elucidating the strengths and weakness of CoW baselines for L-ZSON (Sec. 6.2). Finally, we compare to prior ZSON art in ROBOTHER and HABITAT (MP3D) environments (Sec. 6.3).", + "bbox": [ + 496, + 175, + 890, + 280 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "6.1. Experimental setup", + "text_level": 1, + "bbox": [ + 500, + 290, + 686, + 306 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "**Environments.** We consider PASTURE (Sec. 5), ROBOTHER [18], and HABITAT (MP3D) [67] validation sets as our test sets. We utilize validation sets for testing because official test set ground-truth is not publicly available. Domains are setup with noise that is faithful to their original challenge settings. For ROBOTHER—and by extension PASTURE—this means actuation noise but no depth noise. For HABITAT this means considerable depth noise and reconstruction artifacts, but no actuation noise.", + "bbox": [ + 496, + 316, + 890, + 452 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Navigation Metrics. We adopt standard object navigation metrics to measure performance:", + "bbox": [ + 498, + 455, + 890, + 486 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- SUCCESS (SR): the fraction of episodes where the agent executes STOP within $1.0\\mathrm{m}$ of the target object.", + "- Success weighted by inverse path length (SPL): Success weighted by the oracle shortest path length and normalized by the actual path length [5]. This metric points to the success efficiency of the agent." + ], + "bbox": [ + 500, + 489, + 890, + 580 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In ROBOTHOR and PASTURE, the target must additionally be visible for the episode to be a success, which this is not the case in HABITAT—as specified in their 2021 challenge.", + "bbox": [ + 496, + 588, + 890, + 633 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Embodiment. The agent is a LoCoBot [29]. All agents have discrete actions: {MOVEFORWARD, ROTATERIGHT, ROTATELEFT, STOP}. The move action advances the agent by $0.25\\mathrm{m}$ , while rotation actions pivot the camera by $30^{\\circ}$ .", + "bbox": [ + 496, + 637, + 890, + 696 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "CoW Baselines. For exploration we consider policies presented in Sec. 4.2: FBE heuristic exploration, learnable exploration optimized on HABITAT (MP3D) train scenes, and learned exploration optimized on ROBOTHOR train scenes. Learned exploration requires training in simulation—which is counter to our zero-shot goals; nonetheless, we ablate these explorers to contextualize their performance within the CoW framework. FBE is the default CoW exploration strategy.", + "bbox": [ + 496, + 700, + 890, + 835 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For object localization, we consider:", + "bbox": [ + 517, + 837, + 761, + 851 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- CLIP with $k = 9$ referring expressions (CLIP-Ref.)", + "- CLIP with $k = 9$ patches (CLIP-Patch)", + "- CLIP with gradient relevance (CLIP-Grad.)" + ], + "bbox": [ + 500, + 854, + 852, + 904 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "23175", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/09f95fb87e8e3586cd849b725aebe4948ce6be28c2b80c8886917a367e39b204.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
IDCoW breedsPASTUREROBOTHOR
Uncom.Appear.SpaceAppear. distract SRSpace distract SRHid.Hid. distract SRAvg.
LocalizerArch.SRSRSRSRSRSRSRSPLSRSPLSR
▲CLIP-Ref.B/323.60.61.70.61.72.22.50.91.81.01.8
■CLIP-Ref.B/161.42.82.83.13.31.71.91.72.42.12.7
▲CLIP-PatchB/3218.113.313.38.610.817.517.89.014.210.620.3
■CLIP-PatchB/1610.611.47.810.88.116.415.67.711.59.715.7
▲CLIP-Grad.B/3216.111.911.79.710.314.416.19.212.99.715.2
■CLIP-Grad.B/168.110.88.68.66.711.111.46.79.38.611.6
◇MDETRB33.17.25.06.94.78.18.95.46.38.49.9
▲OWLB/3232.826.419.419.416.119.214.412.621.116.926.7
■OWLB/1631.926.918.919.414.718.115.812.620.817.227.5
ProcTHOR fine-tune (supervised) [19]n/an/an/an/an/an/an/an/an/a27.466.4
", + "bbox": [ + 84, + 88, + 883, + 330 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1. Benchmarking CoWs on PASTURE for L-ZSON. On PASTURE we identify several key takeaways. (1) Average success on PASTURE is lower than on ROBOTHOR; however, CoWs are surprisingly good at finding uncommon objects (Uncom.), often finding them at higher rates than more common ROBOTHOR objects. (2) Comparing square (■) vs. triangle (▲) IDs, we see that architectures (Arch.) using more compute (i.e., ViT-B/16) often perform comparably or worse than their competitors (i.e., ViT-B/32). This is especially true for CLIP [61] models (indicated in pink, orange, and purple). (3) Blue OWL-ViT [49] models perform best. (4) PASTURE tasks with distractor objects (distract) hurt performance and natural language specification is not sufficient to mitigate against the added difficulties in these tasks. (5) A supervised baseline shown in gray significantly outperforms CoWs on ROBOTHOR; however, it is unable to support PASTURE tasks out-of-the-box.", + "bbox": [ + 75, + 337, + 893, + 446 + ], + "page_idx": 5 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- MDETR segmentation model (MDETR)", + "- OWL-ViT detector (OWL)" + ], + "bbox": [ + 76, + 455, + 359, + 491 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Descriptions of these models are in Sec. 4.3 and additional details are in Appx. C. All models are open-vocabulary. No models are fine-tuned on navigation, and hence we consider their inference zero-shot on our tasks.2 We also consider various backbone architectures:", + "bbox": [ + 75, + 494, + 468, + 570 + ], + "page_idx": 5 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- A vision transformer [24], ViT-B/32 ( $\\triangle$ B/32)", + "- ViT-B/16 (■ B/16), which uses a smaller patch size of 16x16 and hence more compute.", + "- EfficientNet B3 ( $\\spadesuit$ B3), which is convolutional and similar in compute requirements to a ViT/B32." + ], + "bbox": [ + 76, + 575, + 468, + 659 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For every model we evaluate with post-processing as the default setting, where only the center pixel of detections is registered in the top-down CoW map. Recall, this is a sensible strategy as only some part of the object needs to be found for an episode to be successful. We find that this decision improves performance on our best performing models from 0.1 to 6.0 percentage points on ROBOTHOR SUCCESS. For a full comparison between models with and without post-processing see Appx. F. For details on hyperparameters, learned agents, object localization threshold tuning, and CLIP prompt-tuning, see Appx. C, G.", + "bbox": [ + 75, + 664, + 468, + 830 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "End-to-end learnable baselines. We also compare against methods that are trained in simulation for millions of steps:", + "bbox": [ + 76, + 834, + 468, + 864 + ], + "page_idx": 5 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- EmbCLIP-ZSON [38] trains on eight ROBOTHR categories, using CLIP language embeddings to specify the goal objects. At test time, the model is evaluated on four held-out object categories, which are also specified CLIP language embeddings for the target category names.", + "- SemanticNav-ZSON [46] trains models separately, one for each dataset, for image goal navigation. Image goals are specified with CLIP visual embeddings. At test time, image goals are swapped for CLIP language embeddings for the object goals. We compare to the MP3D model." + ], + "bbox": [ + 500, + 455, + 890, + 614 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Both EmbCLIP-ZSON and SemanticNav-ZSON leverage multi-modal CLIP visual and language embeddings in learnable frameworks that require simulation training.", + "bbox": [ + 498, + 618, + 890, + 664 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "6.2. CoWs on PASTURE", + "text_level": 1, + "bbox": [ + 500, + 676, + 678, + 690 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Tab. 1 shows our main results of CoWs evaluated on PASTURE. For category-level results see Appx. H. We now discuss several salient questions.", + "bbox": [ + 498, + 699, + 890, + 744 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "How well can CoWs find common objects vs. uncommon objects? Comparing ROBOTHOR and uncommon (Uncom.) PASTURE success rate (SR) in Tab. 1—first and last columns—we notice that CoWs often find uncommon objects at higher rates than common ROBOTHOR objects (e.g., by $\\sim 6$ percentage points SUCCESS for the OWL ViT-B/32 CoW ( $\\triangle$ )). We hypothesize that though uncommon objects are less prevalent in daily life, they are still represented in open-vocabulary datasets and hence recognizable for the object localization modules. We further ex", + "bbox": [ + 496, + 750, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_footnote", + "text": "2The claim is not that these models have never seen any synthetic data in their large-scale training sets, only that they are not trained to navigate.", + "bbox": [ + 76, + 875, + 468, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "23176", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/c08efe085810ec7489087ad49b99ca9d5562e6dc03738f53c5d2af944044b086.jpg", + "image_caption": [ + "(a) Attribute object navigation" + ], + "image_footnote": [], + "bbox": [ + 91, + 109, + 284, + 253 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e5c82a109c048ad9062ef6ee47c8ab41ed2c0d4d029e09f19abae3c746c5c1fe.jpg", + "image_caption": [ + "(b) Attribute object navigation with distractors", + "Figure 5. PASTURE object navigation with descriptions. In general object navigation with descriptions is more challenging than the ROBOTOR object navigation task, as indicated by trend lines lying below the $y = x$ line. (a) Appearance descriptions are more helpful than spatial descriptions. (b) Performance further drops when distractor objects are introduced to the environment. However, CoWs are still able to make better use of appearance description than spatial descriptions. (c) Models in the lower success regime (<15% ROBOTOR SUCCESS) perform comparably on finding hidden objects. However, this trend plateaus for higher success models. (d) Trends are similar when distractor objects are introduced for hidden object navigation." + ], + "image_footnote": [], + "bbox": [ + 290, + 112, + 478, + 252 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e8ea830b0c8da8f02a872487755d76c6447c60a7110e3b42e1f8f0862616823c.jpg", + "image_caption": [ + "(c) Hidden object navigation" + ], + "image_footnote": [], + "bbox": [ + 483, + 112, + 674, + 252 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/cb2d52d7231379da66db7359b05d1718824a1723620903f966b7f16b8b2de868.jpg", + "image_caption": [ + "(d) Hidden object navigation with distractors" + ], + "image_footnote": [], + "bbox": [ + 681, + 112, + 870, + 252 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "plore this hypothesis in Appx. E by visualizing CLIP retrieval results on LAION-5B [68] for the uncommon object categories. The relatively high performance on uncommon objects speaks to the flexibility of CoW baselines and their ability to inherit desirable properties from the open-vocabulary models that they are constructed from.", + "bbox": [ + 75, + 357, + 467, + 446 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Can CoWs utilize appearance and spatial descriptions? Looking at Fig. 5 (a) we see that neither appearance nor spatial descriptions improve CoW performance compared to their ROBOTHOR baseline performance (i.e., most points lie under the $y = x$ line). However, CoW is able to take better advantage of appearance descriptions than spatial descriptions. These results motivate future investigation on open-vocabulary object localization with a greater focus on textual object attributes.", + "bbox": [ + 75, + 450, + 467, + 585 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Can CoWs find visible objects in the presence of distractors? In Fig. 5 (b) we see that CoWs experience performance degradation when compared to Fig. 5 (a). We conclude that appearance and spatial attributes added as language input are not sufficient to deal with the complexity of distractors given current open-vocabulary models.", + "bbox": [ + 75, + 590, + 467, + 681 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Can CoWs find hidden objects? Looking at Fig. 5 (c) we notice that models in the lower success regime (less than $15\\%$ SUCCESS on ROBOTHOR) are able to find hidden objects at about the same rate as ROBOTHOR objects (i.e., they lie on the $y = x$ line). OWL models in the higher success regime $(>15\\%)$ do not continue this trend; however, they do achieve higher absolute accuracy as seen in Tab. 1. Dealing with occlusion is a longstanding problem in computer vision, and these results provide a foundation upon which future hidden object navigation work can improve.", + "bbox": [ + 75, + 685, + 467, + 835 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Can CoWs find hidden objects in the presence of distractors? Comparing Figs. 5 (c) and (d), we notice similar trends lines, with the best models performing worse with distractors. This suggests that distractors are less of a con", + "bbox": [ + 75, + 839, + 467, + 900 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/27e9702207cfe9fe9ea99b438d70b63e83f7f51d807f8aeb889883027058be5f.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
IDLoc.CoW breedsPASTURE (Avg.)ROBOTHOR
Arch.Exp. StrategySPLSRSPLSR
▲OWLB/32ROBOTHOR learn.10.217.313.120.9
▲OWLB/32HABITAT learn.8.619.49.820.4
▲OWLB/32FBE12.621.116.926.7
", + "bbox": [ + 504, + 354, + 883, + 429 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/21c0200d8363d30e2cc9b0663730b24c573ff41844f76c01d42cfc9936bd4b24.jpg", + "table_caption": [ + "Table 2. Exploration ablation. For a fixed object localizer (OWL-ViT B/32 with post processing), we ablate over different choices of exploration policy: the FBE heuristic, agents trained in ROBOTHOR, and HABITAT (MP3D). We find that FBE outperforms learnable alternatives on both PASTURE and ROBOTHOR. HABITAT learnable model perform worst, but are not trained on any PASTURE or ROBOTHOR data." + ], + "table_footnote": [], + "table_body": "
IDCoW breedsObj. PriorPASTURE Uncom.ROBOTHOR
Loc.Arch.SPLSRSPLSR
▲OWLB/32None20.532.816.826.7
▲OWLB/32GPT-3.522.236.917.027.5
", + "bbox": [ + 506, + 539, + 883, + 602 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 3. CoW with GPT-3.5 priors. Leveraging GPT-3.5, we generate priors for objects (e.g., apples are likely to be in dining room scenes). Instead of directly searching for the target object, CoW first searches for the scenes, which boosts performance.", + "bbox": [ + 498, + 608, + 890, + 665 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "cern in the case of hidden objects than for visible object targets. In light of the fact that detection methods generally work better on larger objects, we hypothesize this effect is because distractor objects are smaller (e.g., apples, vases, basketballs) than objects used to conceal target categories (e.g., beds, sofas, etc.).", + "bbox": [ + 496, + 681, + 890, + 773 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "What exploration method performs best? We ablate the decision to use FBE for most experiments by fixing an object localizer (OWL, B/32 $(\\triangle))$ and comparing against ROBOTHR learnable exploration and HABITAT learnable exploration in Tab. 2. We notice that FBE performs best in all cases; however, learnable exploration still performs well suggesting that these models do learn useful strategies for the downstream tasks.", + "bbox": [ + 496, + 779, + 890, + 898 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "23177", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/5beab85ea8a0266fdfe3a1171921383455df55bd1b781209469696e873de385b.jpg", + "image_caption": [ + "Figure 6. Failure analysis for OWL, B/32 (▲). Exploration and object localization errors occur at similar ratios, with increased localization failures in the presence of distractors." + ], + "image_footnote": [], + "bbox": [ + 124, + 95, + 421, + 261 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Can CoW incorporate object priors? Examining Tab. 3, we see that incorporating GPT-3.5 object-level priors improves performance on both PASTURE uncommon objects and ROBOTOR. These initial results suggest positive trends for incorporating outside knowledge into CoW. Future work may consider more sophisticated methods for injecting priors to steer navigation.", + "bbox": [ + 75, + 329, + 468, + 436 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "How do CoWs fail? We identify three high-level failure modes. (1) Exploration fail: the target is never seen. (2) Object localization fail: the target is seen but the localizer never fires. (3) Planning fail: the target is seen and the localizer fires, but planning fails due to inaccuracy in the map representation (Sec. 4.2). Looking at Fig. 6, we notice a large fraction of failures are due to exploration and object localization. This suggests CoWs may continue to improve as research in these fields progress. In Fig. 6 we also see that in cases where distractors are present a higher fraction of object localization failures occur, further suggesting that open-vocabulary models currently struggle to make full use of attribute prompts. See Appx. I for more failure analysis.", + "bbox": [ + 75, + 439, + 470, + 638 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6.3. Comparison to Prior Art", + "text_level": 1, + "bbox": [ + 76, + 648, + 305, + 666 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We primarily evaluate CoWs in general L-ZSON settings; however, we further evaluate CoWs on ZSON benchmarks to establish them as a strong baseline for these tasks. Recall, ZSON can be seen as a case of L-ZSON where only object goals are specified (no attributes).", + "bbox": [ + 75, + 672, + 468, + 748 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In Tab. 4, we see there exists a CoW that outperforms the end-to-end baselines in all cases except SUCCESS on HABITAT (MP3D). For instance, the CLIP-Grad., B/32 ( $\\triangleleft$ ) matches the SemanticNav-ZSON model on HABITAT (MP3D) SPL: 4.9 for CoW v.s. 4.8 for the competitor, while improving over EmbCLIP-ZSON ROBOTHOR by 15.6 percentage points. To contextualize this result, CoWs train for $O$ navigation steps, while SemanticNav-ZSON and EmbCLIP-ZSON train in the target evaluation simulators for 500M and 60M steps respectively.", + "bbox": [ + 75, + 750, + 470, + 901 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/9900d387144252b9e5c8d9ea3b0beb45ffe14a4ce3364e08d002845713818384.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
IDCoW breedsHABITAT (MP3D)ROBOTHON (subset)ROBOTHON (full)Nav. training steps
Loc.Arch.SPLSRSPLSRSPLSR
▲CLIP-Grad.B/324.99.215.023.79.715.20
▲OWLB/323.77.420.832.516.926.70
EmbCLIP-ZSON [38]---8.1-14.0*60M
SemanticNav-ZSON [46]4.815.3----500M
", + "bbox": [ + 514, + 89, + 879, + 195 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 4. Comparison to prior art on existing ZSON benchmarks. CoWs are able to match or out-compete existing methods that leverage millions of steps of navigation training in the evaluation simulator. *indicates a result from prior work that includes, non-zero-shot evaluation. Specifically, only 1/4 of the evaluations are zero-shot on ROBOTOR (subset) and the remaining 3/4 on categories seen during training.", + "bbox": [ + 496, + 204, + 890, + 303 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The superior performance of SemanticNav-ZSON in terms of MP3D SUCCESS indicates that there can be benefits to in-domain learning. Future work may consider unifying the benefits of CoW-like models and fine-tuned models.", + "bbox": [ + 496, + 316, + 890, + 378 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "7. Limitations and Conclusion", + "text_level": 1, + "bbox": [ + 500, + 392, + 756, + 407 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Limitations. While our evaluation of CoWs on HABITAT, ROBOTHER, and PASTURE is a step towards assessing their performance in different domains, ultimately, real-world performance matters most. Hence, the biggest limitation of our work is the lack of large-scale, real-world benchmarking—which is also missing in much of the related literature. Additionally, CoW inherits the metalimitations of the object localization and exploration methods considered. For example, object localizers require tuning a confidence threshold to balance precision and recall. Finally, we do not consider different agent embodiment or continuous action spaces. This is a pertinent investigation given recent findings of Pratt et al. [59] that agent morphology can be a big determinant of downstream performance.", + "bbox": [ + 496, + 420, + 890, + 632 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Conclusion. This paper introduces the PASTURE benchmark for language-driven zero-shot object navigation and several CLIP on Wheels baselines, translating the successes of existing zero-shot models to an embodied task. We view CoW as an instance of using open-vocabulary models, with text-based interfaces, to tackle robotics tasks in more flexible settings. We hope that the baselines and the proposed benchmark will spur the field to explore broader and more powerful forms of zero-shot embodied AI.", + "bbox": [ + 496, + 635, + 890, + 772 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgement. We would like to thank Jessie Chapman, Cheng Chi, Huy Ha, Zeyi Liu, Sachit Menon, and Sarah Pratt for valuable feedback. This work was supported in part by NSF CMMI-2037101, NSF IIS-2132519, and an Amazon Research Award. SYG is supported by a NSF Graduate Research Fellowship. The views and conclusions contained herein are those of the authors and should not be interpreted as necessarily representing the official policies, either expressed or implied, of the sponsors.", + "bbox": [ + 496, + 776, + 890, + 896 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "23178", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 174, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Ziad Al-Halah, Santhosh K. Ramakrishnan, and Kristen Grauman. Zero experience required: Plug & play modular transfer learning for semantic visual navigation. arXiv, 2022. 2", + "[2] Peter Anderson, Qi Wu, Damien Teney, Jake Bruce, Mark Johnson, Niko Sunderhauf, Ian Reid, Stephen Gould, and Anton Van Den Hengel. Vision-and-language navigation: Interpreting visually-grounded navigation instructions in real environments. CVPR, 2018. 2", + "[3] Donghyeon Baek, Youngmin Oh, and Bumsub Ham. Exploiting a joint embedding space for generalized zero-shot semantic segmentation. ICCV, 2021. 1", + "[4] Ankan Bansal, Karan Sikka, Gaurav Sharma, Rama Chellappa, and Ajay Divakaran. Zero-shot object detection. ECCV, 2018. 1", + "[5] Dhruv Batra, Aaron Gokaslan, Aniruddha Kembhavi, Oleksandr Maksymets, Roozbeh Mottaghi, Manolis Savva, Alexander Toshev, and Erik Wijmans. Objectnav revisited: On evaluation of embodied agents navigating to objects. arXiv, 2020. 1, 3, 5", + "[6] Maxime Bucher, Tuan-Hung VU, Matthieu Cord, and Patrick Pérez. Zero-shot semantic segmentation. NeurIPS, 2019. 1", + "[7] Yuri Burda, Harrison Edwards, Amos J. Storkey, and Oleg Klimov. Exploration by random network distillation. *ICLR*, 2018. 2", + "[8] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. ECCV, 2020. 4, 14", + "[9] Matthew Chang, Arjun Gupta, and Saurabh Gupta. Semantic visual navigation by watching youtube videos. NeurIPS, 2020. 2", + "[10] Devendra Singh Chaplot, Dhiraj Gandhi, Abhinav Gupta, and Ruslan Salakhutdinov. Object goal navigation using goal-oriented semantic exploration. NeurIPS, 2020. 2", + "[11] Devendra Singh Chaplot, Dhiraj Gandhi, Saurabh Gupta, Abhinav Kumar Gupta, and Ruslan Salakhutdinov. Learning to explore using active neural slam. ICLR, 2020. 2", + "[12] Prithvijit Chattopadhyay, Judy Hoffman, Roozbeh Mottaghi, and Aniruddha Kembhavi. Robustnav: Towards benchmarking robustness in embodied navigation. ICCV, 2021. 2", + "[13] Hila Chefer, Shir Gur, and Lior Wolf. Transformer interpretability beyond attention visualization. CVPR, 2021. 4, 13, 14", + "[14] Tao Chen, Saurabh Gupta, and Abhinav Gupta. Learning exploration policies for navigation. *ICLR*, 2019. 2", + "[15] Jiaxin Cheng, Soumyaroop Nandi, Prem Natarajan, and Wael Abd-Almageed. Sign: Spatial-information incorporated generative network for generalized zero-shot semantic segmentation. ICCV, 2021. 1", + "[16] Kyunghyun Cho, Bart van Merrienboer, Dzmitry Bahdanau, and Yoshua Bengio. On the properties of neural machine translation: Encoder-decoder approaches. In SSST@EMNLP, 2014. 3", + "[17] Andrew J Davison and David W Murray. Mobile robot localisation using active vision. ECCV, 1998. 2" + ], + "bbox": [ + 78, + 114, + 470, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[18] Matt Deitke, Winson Han, Alvaro Herrasti, Aniruddha Kembhavi, Eric Kolve, Roozbeh Mottaghi, Jordi Salvador, Dustin Schwenk, Eli VanderBilt, Matthew Wallingford, Luca Weihs, Mark Yatskar, and Ali Farhadi. Robothor: An open simulation-to-real embodied ai platform. CVPR, 2020. 1, 3, 5", + "[19] Matt Deitke, Eli VanderBilt, Alvaro Herrasti, Luca Weihs, Jordi Salvador, Kiana Ehsani, Winson Han, Eric Kolve, Ali Farhadi, Aniruddha Kembhavi, and Roozbeh Mottaghi. Proctor: Large-scale embodied ai using procedural generation. NeurIPS, 2022. 2, 6", + "[20] Frank Dellaert, Dieter Fox, Wolfram Burgard, and Sebastian Thrun. Monte carlo localization for mobile robots. ICRA, 1999. 2", + "[21] Berkan Demirel, Ramazan Gokberk Cinbis, and Nazli Ikizler-Cinbis. Zero-shot object detection by hybrid region embedding. BMVC, 2018. 1", + "[22] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, K. Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. CVPR, 2009. 1", + "[23] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv, 2020. 13", + "[24] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. ICLR, 2020. 4, 6", + "[25] Samir Yitzhak Gadre, Kiana Ehsani, Shuran Song, and Roozbeh Mottaghi. Continuous scene representations for embodied ai. CVPR, 2022. 2, 3", + "[26] Daniel Gordon, Abhishek Kadian, Devi Parikh, Judy Hoffman, and Dhruv Batra. Splitnet: Sim2sim and task2task transfer for embodied visual navigation. CVPR, 2019. 2", + "[27] Xiuye Gu, Tsung-Yi Lin, Weicheng Kuo, and Yin Cui. Zeroshot detection via vision and language knowledge distillation. arXiv, 2021. 1", + "[28] Agrim Gupta, Piotr Dólar, and Ross B. Girshick. Lvis: A dataset for large vocabulary instance segmentation. CVPR, 2019. 1", + "[29] Abhinav Kumar Gupta, Adithyavairavan Murali, Dhiraj Gandhi, and Terrel Pinto. Robot learning in homes: Improving generalization and reducing dataset bias. NeurIPS, 2018. 5", + "[30] Saurabh Gupta, James Davidson, Sergey Levine, Rahul Sukthankar, and Jitendra Malik. Cognitive mapping and planning for visual navigation. CVPR, pages 2616-2625, 2017. 2", + "[31] Meera Hahn, Devendra Singh Chaplot, and Shubham Tulsiani. No rl, no simulation: Learning to navigate without navigating. NeurIPS, 2021. 2", + "[32] Peter Henry, Michael Krainin, Evan Herbst, Xiaofeng Ren, and Dieter Fox. RGB-d mapping: Using depth cameras for dense 3d modeling of indoor environments. Experimental robotics, 2014. 2" + ], + "bbox": [ + 501, + 92, + 890, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "23179", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[33] Ping Hu, Stan Sclaroff, and Kate Saenko. Uncertainty-aware learning for zero-shot semantic segmentation. NeurIPS, 2020. 1", + "[34] Chenguang Huang, Oier Mees, Andy Zeng, and Wolfram Burgard. Visual language maps for robot navigation. arXiv preprint arXiv:2210.05714, 2022. 2", + "[35] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc V Le, Yunhsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. ICML, 2021. 1, 13", + "[36] Aishwarya Kamath, Mannat Singh, Yann LeCun, Ishan Misra, Gabriel Synnaeve, and Nicolas Carion. Mdetr - modulated detection for end-to-end multi-modal understanding. ICCV, 2021. 1, 4, 13, 14", + "[37] Naoki Kato, Toshihiko Yamasaki, and Kiyoharu Aizawa. Zero-shot semantic segmentation via variational mapping. ICCV-W, 2019. 1", + "[38] Apoorv Khandelwal, Luca Weihs, Roozbeh Mottaghi, and Aniruddha Kembhavi. Simple but effective: Clip embeddings for embodied ai. arXiv, 2021. 1, 2, 3, 6, 8", + "[39] Jacob Krantz, Erik Wijmans, Arjun Majumdar, Dhruv Batra, and Stefan Lee. Beyond the nav-graph: Vision-and-language navigation in continuous environments. In European Conference on Computer Vision, pages 104-120. Springer, 2020. 2", + "[40] Alexander Ku, Peter Anderson, Roma Patel, Eugene Ie, and Jason Baldridge. Room-across-room: Multilingual vision-and-language navigation with dense spatiotemporal grounding. arXiv preprint arXiv:2010.07954, 2020. 2", + "[41] Benjamin Kuipers and Yung-Tai Byun. A robot exploration and mapping strategy based on a semantic hierarchy of spatial representations. Robotics and autonomous systems, 1991. 2", + "[42] Xinghang Li, Di Guo, Huaping Liu, and Fuchun Sun. Revece: Remote embodied visual referring expression in continuous environment. RA-L, 2022. 2", + "[43] Zhihui Li, Lina Yao, Xiaqin Zhang, Xianzhi Wang, Salil S. Kanhere, and Huaxiang Zhang. Zero-shot object detection with textual descriptions. AAAI, 2019. 1", + "[44] Yiqing Liang, Boyuan Chen, and Shuran Song. Sscnav: Confidence-aware semantic scene completion for visual semantic navigation. ICRA, 2021. 2", + "[45] Tsung-Yi Lin, Michael Maire, Serge Belongie, Lubomir Bourdev, Ross Girshick, James Hays, Pietro Perona, Deva Ramanan, C. Lawrence Zitnick, and Piotr Dólar. Microsoft coco: Common objects in context. ECCV, 2014. 1", + "[46] Arjun Majumdar, Gunjan Aggarwal, Bhavika Devnani, Judy Hoffman, and Dhruv Batra. Zson: Zero-shot object-goal navigation using multimodal goal embeddings. arXiv preprint arXiv:2206.12403, 2022. 1, 2, 3, 6, 8", + "[47] Qiaomei Mao, Chong Wang, Sheng Yu, Ye Zheng, and Yuqi Li. Zero-shot object detection with attributes-based category similarity. IEEE Transactions on Circuits and Systems II: Express Briefs, 2020. 1", + "[48] Lina Mezghani, Sainbayar Sukhbaatar, Thibaut Lavril, Oleksandr Maksymets, Dhruv Batra, Piotr Bojanowski, and Alahari Karteek. Memory-augmented reinforcement learning for image-goal navigation. arXiv, 2021. 2" + ], + "bbox": [ + 78, + 90, + 468, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[49] Matthias Minderer, Alexey A. Gritsenko, Austin Stone, Maxim Neumann, Dirk Weissenborn, Alexey Dosovitskiy, Aravindh Mahendran, Anurag Arnab, Mostafa Dehghani, Zhuoran Shen, Xiao Wang, Xiaohua Zhai, Thomas Kipf, and Neil Houlsby. Simple open-vocabulary object detection with vision transformers. ECCV, 2022. 1, 4, 6, 13, 14", + "[50] Grégoire Montavon, Sebastian Lapuschkin, Alexander Binder, Wojciech Samek, and Klaus-Robert Müller. Explaining nonlinear classification decisions with deep taylor decomposition. Pattern recognition, 65:211-222, 2017. 13", + "[51] Hans Moravec and Alberto Elfes. High resolution maps from wide angle sonar. ICRA, 1985. 2", + "[52] Richard A Newcombe, Shahram Izadi, Otmar Hilliges, David Molyneaux, David Kim, Andrew J Davison, Pushmeet Kohi, Jamie Shotton, Steve Hodges, and Andrew Fitzgibbon. Kinectfusion: Real-time dense surface mapping and tracking. ISMAR, 2011. 2", + "[53] Lachlan Nicholson, Michael Milford, and Niko Sunderhauf. Quadricslam: Dual quadrics from object detections as landmarks in object-oriented slam. RA-L, 2019. 2", + "[54] Clark F Olson and Larry H Matthies. Maximum likelihood rover localization by matching range maps. ICRA, 1998. 2", + "[55] Long Ouyang, Jeff Wu, Xu Jiang, Diogo Almeida, Carroll L Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, et al. Training language models to follow instructions with human feedback. NeurIPS, 2022. 4, 15", + "[56] Simone Parisi, Victoria Dean, Deepak Pathak, and Abhinav Kumar Gupta. Interesting object, curious agent: Learning task-agnostic exploration. NeurIPS, 2021. 2", + "[57] Deepak Pathak, Pulkit Agrawal, Alexei A. Efros, and Trevor Darrell. Curiosity-driven exploration by self-supervised prediction. ICML, 2017. 2", + "[58] Hieu Pham, Zihang Dai, Golnaz Ghiasi, Hanxiao Liu, Adams Wei Yu, Minh-Thang Luong, Mingxing Tan, and Quoc V. Le. Combined scaling for zero-shot transfer learning. arXiv, 2021. 1, 13", + "[59] Sarah Pratt, Luca Weihs, and Ali Farhadi. The introspective agent: Interdependence of strategy, physiology, and sensing for embodied agents. arXiv, 2022. 8", + "[60] Yuankai Qi, Qi Wu, Peter Anderson, Xin Wang, William Yang Wang, Chunhua Shen, and Anton van den Hengel. Reverie: Remote embodied visual referring expression in real indoor environments. CVPR, 2020. 2", + "[61] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. ICML, 2021. 1, 2, 4, 6, 13", + "[62] Shafin Rahman, Salman Hameed Khan, and Fatih Murat Porkikli. Zero-shot object detection: Learning to simultaneously recognize and localize novel concepts. ACCV, 2018. 1", + "[63] Roberta Raileanu and Tim Rocktäschel. Ride: Rewarding impact-driven exploration for procedurally-generated environments. *ICLR*, 2020. 2" + ], + "bbox": [ + 501, + 92, + 890, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "23180", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[64] Santhosh K. Ramakrishnan, Devendra Singh Chaplot, Ziad Al-Halah, Jitendra Malik, and Kristen Grauman. Poni: Potential functions for objectgoal navigation with interaction-free learning. CVPR, 2022. 2", + "[65] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, Alexander C. Berg, and Li Fei-Fei. Imagenet large scale visual recognition challenge. IJCV, 2015. 1", + "[66] Manolis Savva, Angel X Chang, Alexey Dosovitskiy, Thomas Funkhouser, and Vladlen Koltun. Minos: Multimodal indoor simulator for navigation in complex environments. arXiv, 2017. 2", + "[67] Manolis Savva, Abhishek Kadian, Oleksandr Maksymets, Yili Zhao, Erik Wijmans, Bhavana Jain, Julian Straub, Jia Liu, Vladlen Koltun, Jitendra Malik, Devi Parikh, and Dhruv Batra. Habitat: A Platform for Embodied AI Research. ICCV, 2019. 1, 3, 5", + "[68] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. NeurIPS, 2022. 7, 16, 18", + "[69] John Schulman, Filip Wolski, Prafulla Dhariwal, Alec Radford, and Oleg Klimov. Proximal policy optimization algorithms. arXiv, 2017. 3", + "[70] Ramprasaath R. Selvaraju, Abhishek Das, Ramakrishna Vedantam, Michael Cogswell, Devi Parikh, and Dhruv Batra. Grad-cam: Visual explanations from deep networks via gradient-based localization. IJCV, 2019. 4, 13, 14", + "[71] Mohit Shridhar, Jesse Thomason, Daniel Gordon, Yonatan Bisk, Winson Han, Roozbeh Mottaghi, Luke Zettlemoyer, and Dieter Fox. Alfred: A benchmark for interpreting grounded instructions for everyday tasks. CVPR, 2020. 2", + "[72] Shuran Song, Linguang Zhang, and Jianxiong Xiao. Robot in a room: Toward perfect object recognition in closed environments. arXiv, 2015. 2", + "[73] Alexander L. Strehl and Michael L. Littman. An analysis of model-based interval estimation for markov decision processes. J. Comput. Syst. Sci., 2008, 2, 3", + "[74] Saim Wani, Shivansh Patel, Unnat Jain, Angel X. Chang, and Manolis Savva. Mution: Benchmarking semantic map memory using multi-object navigation. NeurIPS, 2020. 2", + "[75] Luca Weihs, Matt Deitke, Aniruddha Kembhavi, and Roozbeh Mottaghi. Visual room rearrangement. CVPR, 2021. 2, 3", + "[76] Luca Weihs, Jordi Salvador, Klemen Kotar, Unnat Jain, KuoHao Zeng, Roozbeh Mottaghi, and Aniruddha Kembhavi. Allenact: A framework for embodied ai research. arXiv, 2020. 3, 12", + "[77] Erik Wijmans, Abhishek Kadian, Ari S. Morcos, Stefan Lee, Irfan Essa, D. Parikh, Manolis Savva, and Dhruv Batra. Dppo: Learning near-perfect pointgoal navigators from 2.5 billion frames. *ICLR*, 2019. 2, 3", + "[78] Brian H Wilcox. Robotic vehicles for planetary exploration. Applied Intelligence, 1992. 2" + ], + "bbox": [ + 78, + 90, + 468, + 900 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[79] Mitchell Wortsman, Kiana Ehsani, Mohammad Rastegari, Ali Farhadi, and Roozbeh Mottaghi. Learning to learn how to learn: Self-adaptive visual navigation using meta-learning. CVPR, 2019. 1, 2", + "[80] Chenyun Wu, Zhe Lin, Scott D. Cohen, Trung Bui, and Subhransu Maji. Phrasecut: Language-based image segmentation in the wild. CVPR, 2020. 4, 14", + "[81] Fei Xia, Amir R Zamir, Zhiyang He, Alexander Sax, Jitendra Malik, and Silvio Savarese. Gibson env: Real-world perception for embodied agents. CVPR, 2018. 2", + "[82] Brian Yamauchi. A frontier-based approach for autonomous exploration. CIRA, 1997. 3", + "[83] Claudia Yan, Dipendra Misra, Andrew Bennnett, Aaron Walsman, Yonatan Bisk, and Yoav Artzi. Chalet: Cornell house agent learning environment. arXiv, 2018. 2", + "[84] Wei Yang, Xiaolong Wang, Ali Farhadi, Abhinav Gupta, and Roozbeh Mottaghi. Visual semantic navigation using scene priors. arXiv, 2018. 2", + "[85] Xiaohua Zhai, Xiao Wang, Basil Mustafa, Andreas Steiner, Daniel Keysers, Alexander Kolesnikov, and Lucas Beyer. Lit: Zero-shot transfer with locked-image text tuning. CVPR, 2022. 14", + "[86] Chong Zhou, Chen Change Loy, and Bo Dai. Denseclip: Extract free dense labels from clip. arXiv, 2021. 1", + "[87] Fengda Zhu, Xiwen Liang, Yi Zhu, Xiaojun Chang, and Xiaodan Liang. Soon: Scenario oriented object navigation with graph-based exploration. CVPR, 2021. 2", + "[88] Pengkai Zhu, Hanxiao Wang, and Venkatesh Saligrama. Zero shot detection. IEEE Transactions on Circuits and Systems for Video Technology, 2020. 1", + "[89] Yuke Zhu, Roozbeh Mottaghi, Eric Kolve, Joseph J Lim, Abhinav Gupta, Li Fei-Fei, and Ali Farhadi. Target-driven visual navigation in indoor scenes using deep reinforcement learning. ICRA, 2017. 1, 2" + ], + "bbox": [ + 501, + 92, + 890, + 574 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "23181", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2023/CoWs on Pasture_ Baselines and Benchmarks for Language-Driven Zero-Shot Object Navigation/a31a35f2-4957-4777-92dc-31eae2e5deb5_model.json b/2023/CoWs on Pasture_ Baselines and Benchmarks for Language-Driven Zero-Shot Object Navigation/a31a35f2-4957-4777-92dc-31eae2e5deb5_model.json new file mode 100644 index 0000000000000000000000000000000000000000..2e6aa6779a69c0ba20510beb2a5bd5ec084e8c9b --- /dev/null +++ b/2023/CoWs on Pasture_ Baselines and Benchmarks for Language-Driven Zero-Shot Object Navigation/a31a35f2-4957-4777-92dc-31eae2e5deb5_model.json @@ -0,0 +1,2807 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.0, + 0.812, + 0.045 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.229, + 0.131, + 0.744, + 0.177 + ], + "angle": 0, + "content": "CoWs on PASTURE: Baselines and Benchmarks for Language-Driven Zero-Shot Object Navigation" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.203, + 0.877, + 0.224 + ], + "angle": 0, + "content": "Samir Yitzhak Gadre\\(^{\\diamond}\\) Mitchell Wortsman\\(^{\\dagger}\\) Gabriel Ilharco\\(^{\\dagger}\\) Ludwig Schmidt\\(^{\\dagger}\\) Shuran Song\\(^{\\diamond}\\)" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.256, + 0.314, + 0.273 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.288, + 0.474, + 0.667 + ], + "angle": 0, + "content": "For robots to be generally useful, they must be able to find arbitrary objects described by people (i.e., be language-driven) even without expensive navigation training on in-domain data (i.e., perform zero-shot inference). We explore these capabilities in a unified setting: language-driven zero-shot object navigation (L-ZSON). Inspired by the recent success of open-vocabulary models for image classification, we investigate a straightforward framework, CLIP on Wheels (CoW), to adapt open-vocabulary models to this task without fine-tuning. To better evaluate L-ZSON, we introduce the PASTURE benchmark, which considers finding uncommon objects, objects described by spatial and appearance attributes, and hidden objects described relative to visible objects. We conduct an in-depth empirical study by directly deploying 22 CoW baselines across HABITAT, ROBOTHOR, and PASTURE. In total, we evaluate over 90k navigation episodes and find that (1) CoW baselines often struggle to leverage language descriptions but are proficient at finding uncommon objects. (2) A simple CoW, with CLIP-based object localization and classical exploration—and no additional training—matches the navigation efficiency of a state-of-the-art ZSON method trained for 500M steps on HABITAT MP3D data. This same CoW provides a 15.6 percentage point improvement in success over a state-of-the-art ROBOTHOR ZSON model." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.693, + 0.21, + 0.71 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.719, + 0.47, + 0.856 + ], + "angle": 0, + "content": "To be more widely applicable, robots should be language-driven: able to deduce goals based on arbitrary text input instead of being constrained to a fixed set of object categories. While existing image classification, semantic segmentation, and object navigation benchmarks like ImageNet-1k [65], ImageNet-21k [22], MS-COCO [45], LVIS [28], HABITAT [67], and ROBOTHOR [18] include a vast array of everyday items, they do not capture all objects that matter to people. For instance, a lost \"toy airplane\" may" + }, + { + "type": "title", + "bbox": [ + 0.508, + 0.255, + 0.86, + 0.267 + ], + "angle": 0, + "content": "Sample tasks Top-down visualization Egocentric Observations" + }, + { + "type": "text", + "bbox": [ + 0.504, + 0.269, + 0.632, + 0.28 + ], + "angle": 0, + "content": "(a) Finding uncommon objects" + }, + { + "type": "image", + "bbox": [ + 0.506, + 0.281, + 0.889, + 0.331 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.504, + 0.334, + 0.783, + 0.345 + ], + "angle": 0, + "content": "(b) Finding objects based on attributes in the presence of distractors" + }, + { + "type": "image", + "bbox": [ + 0.506, + 0.346, + 0.89, + 0.396 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.506, + 0.398, + 0.736, + 0.408 + ], + "angle": 0, + "content": "(c) Finding hidden objects in the presence of distractors" + }, + { + "type": "image", + "bbox": [ + 0.506, + 0.409, + 0.89, + 0.46 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.47, + 0.895, + 0.568 + ], + "angle": 0, + "content": "Figure 1. The PASTURE benchmark for L-ZSON. Text specifies navigation goal objects. Agents do not train on these tasks, making the evaluation protocol zero-shot. (a) Uncommon object goals like \"llama wicker basket\", not found in existing navigation benchmarks. (b) Appearance, spatial descriptions, which are necessary to find the correct object. (c) Hidden object descriptions, which localize objects that are not visible." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.58, + 0.895, + 0.61 + ], + "angle": 0, + "content": "become relevant in a kindergarten classroom, but this object is not annotated in any of the aforementioned datasets." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.612, + 0.895, + 0.793 + ], + "angle": 0, + "content": "In this paper, we study Language-driven zero-shot object navigation (L-ZSON), a more challenging but also more applicable version of object navigation [5, 18, 67, 79, 89] and ZSON [38, 46] tasks. In L-ZSON, an agent must find an object based on a description, which may contain different levels of granularity (e.g., \"toy airplane\", \"toy airplane under the bed\", or \"wooden toy airplane\"). L-ZSON encompasses ZSON, which specifies only the target category [38, 46]. Since L-ZSON is \"zero-shot\", we consider agents without access to navigation training on the target objects or domains. This reflects realistic application scenarios, where the environment and object set may not be known a priori." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.796, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Performing L-ZSON in any environment with unstructured language input is challenging; however, recent advances in open-vocabulary models for image classification [35, 58, 61], object detection [4, 21, 27, 36, 43, 47, 49, 62, 88], and semantic segmentation [3, 6, 15, 33, 36, 37, 86] present a promising foundation. These models provide an interface where one specifies—in text—the arbitrary ob" + }, + { + "type": "page_footnote", + "bbox": [ + 0.078, + 0.863, + 0.47, + 0.887 + ], + "angle": 0, + "content": "\\(^{\\text{©}}\\) Columbia University, \\(^{\\dagger}\\)University of Washington. Correspondence to sy@cs.columbia.edu." + }, + { + "type": "page_footnote", + "bbox": [ + 0.098, + 0.888, + 0.425, + 0.9 + ], + "angle": 0, + "content": "1For code, data, and videos, see cow.cs.columbia.edu/" + }, + { + "type": "list", + "bbox": [ + 0.078, + 0.863, + 0.47, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "23171" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.212 + ], + "angle": 0, + "content": "jects they wish to classify, detect, or segment. For example, CLIP [61] open-vocabulary classifiers compute similarity scores between an input image and a set of user-specified captions (e.g., \"a photo of a toy airplane.\", ..., selecting the caption with the highest score to determine the image classification label. Given the flexibility of these models, we would like to understand their capability to execute embodied tasks even without additional training." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.213, + 0.47, + 0.245 + ], + "angle": 0, + "content": "To this end, we present baselines and benchmarks for L-ZSON. More specifically:" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.25, + 0.473, + 0.432 + ], + "angle": 0, + "content": "- A collection of baseline algorithms, CLIP on Wheels (CoW), which adapt open-vocabulary models to the task of L-ZSON. CoW takes inspiration from the semantic mapping line of work [11, 41, 53], and decomposes the navigation task into exploration when the language target is not confidently localized, and target-driven planning otherwise. CoW retains the textual user interface of the original open-vocabulary model and does not require any navigation training. We evaluate 22 CoWs, ablating over many open-vocabulary models, exploration policies, backbones, prompting strategies, and post-processing strategies." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.438, + 0.473, + 0.589 + ], + "angle": 0, + "content": "- A new benchmark, PASTURE, to evaluate CoW and future methods on L-ZSON. We design PASTURE, visualized in Fig. 1, to study capabilities that traditional object navigation agents, which are trained on a fixed set of categories, do not possess. We consider the ability to find (1) uncommon objects (e.g., \"tie-dye surfboard\"), (2) objects by their spatial and appearance attributes in the presence of distractor objects (e.g., \"green apple\" vs. \"red apple\"), and (3) objects that cannot be visually observed (e.g., \"mug under the bed\")." + }, + { + "type": "list", + "bbox": [ + 0.078, + 0.25, + 0.473, + 0.589 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.598, + 0.471, + 0.81 + ], + "angle": 0, + "content": "Together the CoW baselines and PASTURE benchmark allow us to conduct extensive studies on the capabilities of open-vocabulary models in the context of L-ZSON embodied tasks. Our experiments demonstrate CoW's potential on uncommon objects and limitations in taking full advantage of language descriptions—thereby providing empirical motivation for future studies. To contextualize CoW relative to prior zero-shot methods, we additionally evaluate on the HABITAT MP3D dataset. We find that our best CoW achieves navigation efficiency (SPL) that matches a state-of-the-art ZSON method [46] that trains on MP3D training data for 500M steps. On a ROBOTHOR object subset, considered in prior work, the same CoW beats the leading method [38] by 15.6 percentage points in task success." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.826, + 0.22, + 0.842 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.47, + 0.901 + ], + "angle": 0, + "content": "Mapping and exploration. Exploring effectively with a mobile robot is a long-standing problem in vision and robotics. Classical methods often decompose the task" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.243 + ], + "angle": 0, + "content": "into map reconstruction [30, 32, 51, 52, 72], agent localization [17, 20, 54], and planning [41, 78]. Recent work investigates learned alternatives for exploration [7, 14, 56, 57, 63]. Here, agents are often trained end-to-end with self-supervised rewards (e.g., curiosity [57]) or supervised rewards (e.g., state visitation counts [25, 73, 75]). Learning-based methods typically need less hand-tuning, but require millions of training steps and reward engineering. We test both classical and learnable exploration strategies in the context of CoW to study their applicability to L-ZSON." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.247, + 0.892, + 0.352 + ], + "angle": 0, + "content": "Goal-conditioned navigation. Apart from open-ended exploration, many navigation tasks are goal-conditioned, where the agent needs to navigate to a specified position (i.e., point goal [11, 12, 26, 31, 66, 77, 81, 83]), view of the environment (i.e., image goal [48, 64, 89]), or object category (i.e., object goal [1, 9, 10, 12, 19, 44, 74, 79, 84]). We consider an object goal navigation task." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.357, + 0.892, + 0.584 + ], + "angle": 0, + "content": "Vision-Language Navigation. Prior work investigates language-based navigation, where language provides step-by-step instructions for the task [2, 34, 39, 40, 71]. This line of work demonstrates the benefits of additional language input for robot navigation, especially for long-horizon tasks (e.g., room-to-room navigation [40]). However, providing detailed step-by-step instructions (e.g., move 3 meters south [34]) could be challenging and time-consuming. In our L-ZSON task, an algorithm gets natural language as the goal description instead of low-level instructions. Prior work also investigates navigation with target descriptions in supervised settings [42, 60, 87]. In contrast, we explore a zero-shot evaluation protocol and consider finding hidden objects (e.g., \"mug under bed\") and uncommon objects (e.g., \"tie-dye surfboard\")." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.588, + 0.892, + 0.814 + ], + "angle": 0, + "content": "Zero-shot object navigation (ZSON). Recent work studies object navigation in zero-shot settings, where agents are evaluated on object categories that they are not explicitly trained on [38, 46]. Our task encompasses ZSON; however it also considers cases where more information—object attributes or hidden objects descriptions—is specified. Khandelwal et al. [38] train on a subset of ROBOTHOR categories and evaluate on a held-out set. In concurrent work, Majumdar et al. [46] train on an image goal navigation task and evaluate on object navigation downstream by leveraging CLIP multi-modal embeddings. Both algorithms necessitate navigation training for millions of steps and train separate models for each simulation domain. In contrast, CoW baselines do not necessitate any simulation training and can be deployed in multiple environments." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.829, + 0.681, + 0.845 + ], + "angle": 0, + "content": "3. The L-ZSON Task" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Language-driven zero-shot object navigation (L-ZSON) involves navigating to goal objects, specified in language, without explicit training to do so. Let \\(\\mathcal{O}\\) denote a set of" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "23172" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.081, + 0.089, + 0.506, + 0.215 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.224, + 0.47, + 0.267 + ], + "angle": 0, + "content": "Figure 2. CLIP on Wheels (CoW) overview. A CoW uses a policy to explore and an object localizer (e.g., an open-vocabulary object detector) to determine if an object goal is in view." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.273, + 0.474, + 0.501 + ], + "angle": 0, + "content": "natural language descriptions of target objects with potentially many attributes (e.g., \"plant\", \"snake plant\", \"plant under the bed\", etc.). This contrasts with definitions studied in prior object navigation [5, 18] and ZSON [38, 46] tasks, which focus on high-level categories like \"plant\". Let \\( S \\) denote the set of navigation scenes. Let \\( p_0 \\) describe the initial pose of an agent. A navigation episode \\( \\tau \\in \\mathcal{T} \\) is written as a tuple \\( \\tau = (s, o, p_0) \\), \\( s \\in S \\), \\( o \\in \\mathcal{O} \\). Each \\( \\tau \\) is a zero-shot task as tuples of this form are not seen during training. Starting at \\( p_0 \\), an embodied agent's goal is to find \\( o \\). The agent receives observations and sensor readings \\( I_t \\) (e.g., RGB-D images). At each timestep \\( t \\), the agent executes a navigation action \\( a \\in \\mathcal{A} \\). A special action \\( \\text{STOP} \\in \\mathcal{A} \\) terminates the episode. If the agent is within \\( c \\) units of \\( o \\) and \\( o \\) meets a visibility criteria, the episode is successful." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.512, + 0.383, + 0.529 + ], + "angle": 0, + "content": "4. CLIP on Wheels (CoW) Baselines" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.537, + 0.471, + 0.78 + ], + "angle": 0, + "content": "To address L-ZSON, we investigate a simple baseline approach, CoW, which adapts open-vocabulary models like CLIP to make them suitable for the task. A CoW takes as input an egocentric RGB-D image and an object goal specified in language. As a CoW moves, it updates a top-down map of the world created using RGB-D observations and pose estimates (Sec. 4.1). Each CoW gets an exploration policy and a zero-shot object localization module as seen in Fig. 2. To observe diverse views of the scene, a CoW explores using a policy (Sec. 4.2). As the CoW roams, it keeps track of its confidence about the target object's location using an object localization module (Sec. 4.3) and its top-down map. When a CoW's confidence exceeds a threshold, it plans to the location of the goal and issues the STOP action. We now describe the ingredients used to make the CoWs evaluated in our experiments (Sec. 6)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.787, + 0.285, + 0.804 + ], + "angle": 0, + "content": "4.1. Depth-based Mapping" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.811, + 0.47, + 0.901 + ], + "angle": 0, + "content": "As a CoW moves, it constructs a top-down map using input depth, pose estimates, and known agent height. The map is initialized using known camera intrinsics and the first depth image. Since a CoW knows the intended consequences of its actions (e.g., MOVEFORWARD should result in a \\(0.25\\mathrm{m}\\) translation), each action is represented as a" + }, + { + "type": "image", + "bbox": [ + 0.506, + 0.091, + 0.695, + 0.184 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.699, + 0.092, + 0.888, + 0.182 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.193, + 0.892, + 0.263 + ], + "angle": 0, + "content": "Figure 3. Mapping. Top-down map created from egocentric depth observations as a CoW roams a space. (a) Frontier Based Exploration [82] showing a planned path exploration path to the next frontier. (b) Back-projected object relevance scores provide object goal targets when a CoW has found an object." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.272, + 0.893, + 0.438 + ], + "angle": 0, + "content": "pose delta transform to approximate a transition. To deal with noise associated with actuation or depth, a CoW maintains a map at \\(0.125\\mathrm{m}\\) resolution. To improve map accuracy, a CoW checks for failed actions by comparing successive depth frames for movements (see Appx. A for details). Using known agent height \\((0.9\\mathrm{m})\\), map cells are projected to the ground plane to maintain a top-down representation of the world, which suffices for most navigation applications. Cells close to the floor are considered free space (white points in Fig. 3 (a)), while other cells are considered occupied (blue points in Fig. 3 (a))." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.448, + 0.629, + 0.464 + ], + "angle": 0, + "content": "4.2. Exploration" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.471, + 0.892, + 0.532 + ], + "angle": 0, + "content": "Exploration generates diverse egocentric views so a CoW is more likely to view the language-specified target object. We consider two exploration methods, frontier-based and learning-based." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.535, + 0.893, + 0.701 + ], + "angle": 0, + "content": "Frontier based exploration (FBE) [82]. Using the top-down map discussed in Sec. 4.1, a CoW can navigate using a simple exploration heuristic: move to the frontier between free and unknown space to discover new regions. Once the navigator reaches a frontier (visualized as purple points in Fig. 3 (a)), it moves greedily to the next closest frontier. Since the map is updated at every timestep, a noisy pose estimate can contribute to inaccuracies. For example, narrow passages may collapse in the map due to pose drift. To circumvent such problems, we reinitialize the map when no paths exist to any frontiers in the map." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.705, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Learnable exploration. In addition to FBE, we consider learnable alternatives, which may explore more intelligently but incur substantial training costs. We investigate an architecture and reward structure similar to prior work in embodied AI (e.g., [25, 38, 75]). Specifically, we adopt a frozen CLIP backbone with a trainable GRU [16] and linear heads for the actor and critic networks. We train agents independently in HABITAT [67] and ROBOTHOR [18] simulation environments for 60M steps each, using DD-PPO [69, 77] in the AllenAct [76] framework. We employ a simple count-based reward [73]. All training scenes are disjoint from downstream navigation test scenes. For details on reward, hyperparameters, and training, see Appx. B." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "23173" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.077, + 0.091, + 0.267, + 0.108 + ], + "angle": 0, + "content": "4.3. Object Localization" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.115, + 0.47, + 0.237 + ], + "angle": 0, + "content": "Successful navigation depends on object localization: the ability to tell if and where an object is in an image. Regions of high object relevance, extracted from 2D images, are projected to the depth-based map (Fig. 3 (b)) where they serve as natural navigation targets. To determine if and when a target is in an image, we consider the following object localization modules, used in our experiments (Sec. 6). For more details see Appx. C." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.24, + 0.473, + 0.542 + ], + "angle": 0, + "content": "Adapting open-vocabulary classifiers. We experiment with three strategies to turn CLIP [61] models into object localizers. First, we utilize the CLIP text encoder to embed \\( k \\) referring expressions, which specify regions where the target object may appear in the image. For example, \"a plant in the top left of the image.\" We then match the language embeddings against a CLIP visual embedding for the current observation. We compute similarity between the image and text features to determine relevance scores over the regions. Second, we discretize the image into \\( k \\) smaller patches and obtain CLIP patch embeddings. We then convolve each patch embedding with a CLIP text embedding for the target object. If the object is in a patch, the relevance score for that patch should be high. Third, we modify an interpretability method [13, 70] designed to extract object relevancy from vision transformers (ViTs) [24]. Using a target CLIP text embedding and gradient information accumulated through the CLIP vision encoder, we construct a relevance map over input pixels, which qualitatively segments the target when it is in view." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.546, + 0.47, + 0.728 + ], + "angle": 0, + "content": "Adapting open-vocabulary detectors and segmentors. In addition to CLIP-based methods, we consider two additional open-vocabulary models for object localization. First, the MDETR segmentation model [36], which extends the DETR detector [8] to take arbitrary text and images as input and output box detections. The base model is fine-tuned on PhraseCut [80], a dataset of paired masks and attribute descriptions, to support segmentation. Second, we consider the OWL-ViT detector [49], which uses a set prediction fine-tuning recipe to turn CLIP-like models into object detectors. We use this MDETR and OWL-ViT models to directly query images for targets." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.731, + 0.47, + 0.868 + ], + "angle": 0, + "content": "Post-processing. The aforementioned models give continuous valued predictions. However, we are interested in binary masks giving if and where objects are in images. Hence, we threshold predictions for each model (see Appx. C for details). We further investigate two strategies for using the masks downstream: (1) using the entire mask or (2) using the center pixel. The second strategy is reasonable because only part of an object needs to be detected for successful navigation." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.47, + 0.903 + ], + "angle": 0, + "content": "Target driven planning. Recall, CoWs have depth sensors. We back-project object relevance from 2D images into the" + }, + { + "type": "image", + "bbox": [ + 0.504, + 0.091, + 0.891, + 0.214 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.564, + 0.223, + 0.828, + 0.238 + ], + "angle": 0, + "content": "Figure 4. Uncommon objects in PASTURE." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.245, + 0.893, + 0.306 + ], + "angle": 0, + "content": "depth-based map (Sec. 4.1). We keep only the max relevance for each map cell (Fig. 3 (b)). CoWs can then plan to high relevance areas in the map. See Appx. D for additional method visualization." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.31, + 0.893, + 0.476 + ], + "angle": 0, + "content": "Incorporating object priors. Since CoW does not train or fine-tune on navigation datasets, we investigate alternative approaches to inject object-level priors into the model. For each target object, we prompt GPT-3.5 [55] to generate rooms where the target objects are likely to be found. For example, GPT-3.5 states that apples are likely to be found in \"kitchen\" or \"dining room\" scenes. Following this prior, a GPT-3.5 enabled CoW first uses its object localization module to localize a kitchen or a dining room, and then looks for an apple. This straightforward extension, demonstrates how outside information can be incorporated into CoW." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.489, + 0.747, + 0.505 + ], + "angle": 0, + "content": "5. The PASTURE Benchmark" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.514, + 0.892, + 0.621 + ], + "angle": 0, + "content": "To evaluate CoW baselines and future methods on L-ZSON, we introduce the PASTURE evaluation benchmark. PASTURE builds on ROBOTHER validation scenes, which have parallel environments in the real-world. We target ROBOTHER to facilitate future real-world benchmarking. PASTURE probes for seven capabilities explained in Sec. 5.1. We provide dataset statistics in Sec. 5.2." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.629, + 0.654, + 0.644 + ], + "angle": 0, + "content": "5.1. PASTURE Tasks" + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.652, + 0.871, + 0.668 + ], + "angle": 0, + "content": "PASTURE evaluates seven core L-ZSON capabilities." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.671, + 0.892, + 0.852 + ], + "angle": 0, + "content": "Uncommon objects. Traditional benchmarks (e.g., ROBOTHOR and HABITAT MP3D) evaluate agents on common object categories like TVs; however, given the rich diversity of objects in homes, we would like to understand navigation performance on uncommon objects. Hence we add 12 new objects to each room. We use names shown in Fig. 4 as instance labels, which are minimal descriptions to identify each object. Some identifiers refer to text in images (e.g., \"whiteboard saying CVPR\") or to appearance attributes (e.g., \"wooden toy airplane\"). Other objects are less common in North America, like \"mate\", which is a popular Argentinian drink." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.893, + 0.903 + ], + "angle": 0, + "content": "Appearance descriptions. To evaluate if baselines can take advantage of visual attributes, we introduce descriptions of the form: “\\(\\{size\\}\\), \\(\\{color\\}\\), \\(\\{material\\}\\) \\(\\{object\\}\\). For" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "23174" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.152 + ], + "angle": 0, + "content": "example: \"small, red apple\", \"orange basketball\", \"small, black, metallic alarm clock\". Objects are considered small if their 3D bounding box diagonal is below a threshold. We determine color and materials by inspection." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.156, + 0.468, + 0.262 + ], + "angle": 0, + "content": "Spatial descriptions. To test if agents can leverage spatial information in navigation, we introduce descriptions: “{object} on top of \\(\\{x\\}\\), near \\(\\{y\\}\\), \\(\\{z\\}\\), ...”. For example, “house plant on a dresser near a spray bottle”. To determine [on top of] relations, we use THOR metadata and to determine [nearness] we use a distance threshold between pairs of objects. We inspect all descriptions for correctness." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.266, + 0.468, + 0.432 + ], + "angle": 0, + "content": "Appearance descriptions with distractors. To probe if appearance attributes better facilitate finding objects in the presence of distractors, we reuse the appearance captions from before, but evaluate on an modified environment with two visually distinct instances of each ROBOTHOR object category. For example, for the task of finding a \"red apple\", we have both a red apple and a green apple in the room. A navigator must leverage appearance information—and not just class information—to successfully complete the task. Distractor objects are sufficiently far from the target objects so that finding a distractor cannot count as success." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.436, + 0.468, + 0.481 + ], + "angle": 0, + "content": "Spatial descriptions with distractors. This capability is similar to the one above; however, we evaluate with spatial descriptions in the presence of distractor objects." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.485, + 0.468, + 0.591 + ], + "angle": 0, + "content": "Hidden object descriptions. An ideal object navigator should find objects, even when they are hidden. Hence, we introduce descriptions: “{object} under/in {x}”. For example, “basketball in the dresser drawers” or “vase under the sofa”. We sample large objects (e.g., beds, sofas, dressers) in each scene to determine [under/in] relations. Additionally we remove visible instances of {object} from the room." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.595, + 0.468, + 0.67 + ], + "angle": 0, + "content": "Hidden object descriptions with distractors. We use the hidden object descriptions from before, but reintroduce visible instances of \\(\\{object\\}\\) to serve as distractors. Consider finding a \"mug under the bed\". A distractor mug will also appear in the scene making the task more challenging." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.681, + 0.35, + 0.696 + ], + "angle": 0, + "content": "5.2. Dataset Creation and Statistics" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.705, + 0.468, + 0.901 + ], + "angle": 0, + "content": "PASTURE contains three variations for each of the original 15 validation ROBOTHOR rooms: uncommon objects added, additional object instances added, and target objects removed. For each of the seven settings presented above, we evaluate over 12 object instances in 15 rooms with two initial agent starting locations. Hence PASTURE consists of 2,520 tasks, which is a similar order of magnitude to ROBOTHOR (1,800) and HABITAT MP3D (2,195) validation sets. For appearance attributes, \\(47\\%\\) of the objects are considered \"small\". Each object gets an average of 1.2 color descriptors out of 22 possible choices, and 0.6 material descriptors out of 5 possible choices. Similarly, for spatial attributes, each object gets one object it is on top of or in (e.g.," + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.892, + 0.138 + ], + "angle": 0, + "content": "“vase in a shelving unit”) and an average of 1.9 objects it is near. For a sample of appearance and spatial attributes see Fig. 1. For more dataset details and statistics see Appx. E." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.15, + 0.633, + 0.167 + ], + "angle": 0, + "content": "6. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.176, + 0.892, + 0.281 + ], + "angle": 0, + "content": "We first present our experimental setup, including the datasets, metrics, embodiment, and baselines considered in our study (Sec. 6.1). Then we present results on PASTURE, thereby elucidating the strengths and weakness of CoW baselines for L-ZSON (Sec. 6.2). Finally, we compare to prior ZSON art in ROBOTHER and HABITAT (MP3D) environments (Sec. 6.3)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.291, + 0.687, + 0.307 + ], + "angle": 0, + "content": "6.1. Experimental setup" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.317, + 0.892, + 0.453 + ], + "angle": 0, + "content": "**Environments.** We consider PASTURE (Sec. 5), ROBOTHER [18], and HABITAT (MP3D) [67] validation sets as our test sets. We utilize validation sets for testing because official test set ground-truth is not publicly available. Domains are setup with noise that is faithful to their original challenge settings. For ROBOTHER—and by extension PASTURE—this means actuation noise but no depth noise. For HABITAT this means considerable depth noise and reconstruction artifacts, but no actuation noise." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.457, + 0.892, + 0.487 + ], + "angle": 0, + "content": "Navigation Metrics. We adopt standard object navigation metrics to measure performance:" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.49, + 0.891, + 0.519 + ], + "angle": 0, + "content": "- SUCCESS (SR): the fraction of episodes where the agent executes STOP within \\(1.0\\mathrm{m}\\) of the target object." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.522, + 0.891, + 0.581 + ], + "angle": 0, + "content": "- Success weighted by inverse path length (SPL): Success weighted by the oracle shortest path length and normalized by the actual path length [5]. This metric points to the success efficiency of the agent." + }, + { + "type": "list", + "bbox": [ + 0.5, + 0.49, + 0.891, + 0.581 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.589, + 0.891, + 0.634 + ], + "angle": 0, + "content": "In ROBOTHOR and PASTURE, the target must additionally be visible for the episode to be a success, which this is not the case in HABITAT—as specified in their 2021 challenge." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.638, + 0.891, + 0.698 + ], + "angle": 0, + "content": "Embodiment. The agent is a LoCoBot [29]. All agents have discrete actions: {MOVEFORWARD, ROTATERIGHT, ROTATELEFT, STOP}. The move action advances the agent by \\(0.25\\mathrm{m}\\), while rotation actions pivot the camera by \\(30^{\\circ}\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.701, + 0.892, + 0.837 + ], + "angle": 0, + "content": "CoW Baselines. For exploration we consider policies presented in Sec. 4.2: FBE heuristic exploration, learnable exploration optimized on HABITAT (MP3D) train scenes, and learned exploration optimized on ROBOTHOR train scenes. Learned exploration requires training in simulation—which is counter to our zero-shot goals; nonetheless, we ablate these explorers to contextualize their performance within the CoW framework. FBE is the default CoW exploration strategy." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.838, + 0.762, + 0.852 + ], + "angle": 0, + "content": "For object localization, we consider:" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.855, + 0.854, + 0.87 + ], + "angle": 0, + "content": "- CLIP with \\( k = 9 \\) referring expressions (CLIP-Ref.)" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.873, + 0.775, + 0.888 + ], + "angle": 0, + "content": "- CLIP with \\( k = 9 \\) patches (CLIP-Patch)" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.891, + 0.802, + 0.905 + ], + "angle": 0, + "content": "- CLIP with gradient relevance (CLIP-Grad.)" + }, + { + "type": "list", + "bbox": [ + 0.5, + 0.855, + 0.854, + 0.905 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "23175" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.086, + 0.089, + 0.885, + 0.332 + ], + "angle": 0, + "content": "
IDCoW breedsPASTUREROBOTHOR
Uncom.Appear.SpaceAppear. distract SRSpace distract SRHid.Hid. distract SRAvg.
LocalizerArch.SRSRSRSRSRSRSRSPLSRSPLSR
▲CLIP-Ref.B/323.60.61.70.61.72.22.50.91.81.01.8
■CLIP-Ref.B/161.42.82.83.13.31.71.91.72.42.12.7
▲CLIP-PatchB/3218.113.313.38.610.817.517.89.014.210.620.3
■CLIP-PatchB/1610.611.47.810.88.116.415.67.711.59.715.7
▲CLIP-Grad.B/3216.111.911.79.710.314.416.19.212.99.715.2
■CLIP-Grad.B/168.110.88.68.66.711.111.46.79.38.611.6
◇MDETRB33.17.25.06.94.78.18.95.46.38.49.9
▲OWLB/3232.826.419.419.416.119.214.412.621.116.926.7
■OWLB/1631.926.918.919.414.718.115.812.620.817.227.5
ProcTHOR fine-tune (supervised) [19]n/an/an/an/an/an/an/an/an/a27.466.4
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.338, + 0.895, + 0.448 + ], + "angle": 0, + "content": "Table 1. Benchmarking CoWs on PASTURE for L-ZSON. On PASTURE we identify several key takeaways. (1) Average success on PASTURE is lower than on ROBOTHOR; however, CoWs are surprisingly good at finding uncommon objects (Uncom.), often finding them at higher rates than more common ROBOTHOR objects. (2) Comparing square (■) vs. triangle (▲) IDs, we see that architectures (Arch.) using more compute (i.e., ViT-B/16) often perform comparably or worse than their competitors (i.e., ViT-B/32). This is especially true for CLIP [61] models (indicated in pink, orange, and purple). (3) Blue OWL-ViT [49] models perform best. (4) PASTURE tasks with distractor objects (distract) hurt performance and natural language specification is not sufficient to mitigate against the added difficulties in these tasks. (5) A supervised baseline shown in gray significantly outperforms CoWs on ROBOTHOR; however, it is unable to support PASTURE tasks out-of-the-box." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.457, + 0.361, + 0.473 + ], + "angle": 0, + "content": "- MDETR segmentation model (MDETR)" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.478, + 0.273, + 0.492 + ], + "angle": 0, + "content": "- OWL-ViT detector (OWL)" + }, + { + "type": "list", + "bbox": [ + 0.078, + 0.457, + 0.361, + 0.492 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.496, + 0.47, + 0.571 + ], + "angle": 0, + "content": "Descriptions of these models are in Sec. 4.3 and additional details are in Appx. C. All models are open-vocabulary. No models are fine-tuned on navigation, and hence we consider their inference zero-shot on our tasks.2 We also consider various backbone architectures:" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.576, + 0.395, + 0.591 + ], + "angle": 0, + "content": "- A vision transformer [24], ViT-B/32 (\\(\\triangle\\) B/32)" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.595, + 0.47, + 0.626 + ], + "angle": 0, + "content": "- ViT-B/16 (■ B/16), which uses a smaller patch size of 16x16 and hence more compute." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.63, + 0.469, + 0.66 + ], + "angle": 0, + "content": "- EfficientNet B3 (\\(\\spadesuit\\)B3), which is convolutional and similar in compute requirements to a ViT/B32." + }, + { + "type": "list", + "bbox": [ + 0.078, + 0.576, + 0.47, + 0.66 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.665, + 0.469, + 0.831 + ], + "angle": 0, + "content": "For every model we evaluate with post-processing as the default setting, where only the center pixel of detections is registered in the top-down CoW map. Recall, this is a sensible strategy as only some part of the object needs to be found for an episode to be successful. We find that this decision improves performance on our best performing models from 0.1 to 6.0 percentage points on ROBOTHOR SUCCESS. For a full comparison between models with and without post-processing see Appx. F. For details on hyperparameters, learned agents, object localization threshold tuning, and CLIP prompt-tuning, see Appx. C, G." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.835, + 0.469, + 0.865 + ], + "angle": 0, + "content": "End-to-end learnable baselines. We also compare against methods that are trained in simulation for millions of steps:" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.457, + 0.892, + 0.533 + ], + "angle": 0, + "content": "- EmbCLIP-ZSON [38] trains on eight ROBOTHR categories, using CLIP language embeddings to specify the goal objects. At test time, the model is evaluated on four held-out object categories, which are also specified CLIP language embeddings for the target category names." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.539, + 0.892, + 0.615 + ], + "angle": 0, + "content": "- SemanticNav-ZSON [46] trains models separately, one for each dataset, for image goal navigation. Image goals are specified with CLIP visual embeddings. At test time, image goals are swapped for CLIP language embeddings for the object goals. We compare to the MP3D model." + }, + { + "type": "list", + "bbox": [ + 0.5, + 0.457, + 0.892, + 0.615 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.619, + 0.892, + 0.665 + ], + "angle": 0, + "content": "Both EmbCLIP-ZSON and SemanticNav-ZSON leverage multi-modal CLIP visual and language embeddings in learnable frameworks that require simulation training." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.677, + 0.679, + 0.691 + ], + "angle": 0, + "content": "6.2. CoWs on PASTURE" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.7, + 0.892, + 0.746 + ], + "angle": 0, + "content": "Tab. 1 shows our main results of CoWs evaluated on PASTURE. For category-level results see Appx. H. We now discuss several salient questions." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.892, + 0.901 + ], + "angle": 0, + "content": "How well can CoWs find common objects vs. uncommon objects? Comparing ROBOTHOR and uncommon (Uncom.) PASTURE success rate (SR) in Tab. 1—first and last columns—we notice that CoWs often find uncommon objects at higher rates than common ROBOTHOR objects (e.g., by \\(\\sim 6\\) percentage points SUCCESS for the OWL ViT-B/32 CoW (\\(\\triangle\\))). We hypothesize that though uncommon objects are less prevalent in daily life, they are still represented in open-vocabulary datasets and hence recognizable for the object localization modules. We further ex" + }, + { + "type": "page_footnote", + "bbox": [ + 0.077, + 0.875, + 0.469, + 0.901 + ], + "angle": 0, + "content": "2The claim is not that these models have never seen any synthetic data in their large-scale training sets, only that they are not trained to navigate." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "23176" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.132, + 0.097, + 0.266, + 0.107 + ], + "angle": 0, + "content": "(a) Attribute object navigation" + }, + { + "type": "image", + "bbox": [ + 0.093, + 0.111, + 0.285, + 0.254 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.328, + 0.097, + 0.462, + 0.115 + ], + "angle": 0, + "content": "(b) Attribute object navigation with distractors" + }, + { + "type": "image", + "bbox": [ + 0.291, + 0.113, + 0.48, + 0.253 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.529, + 0.097, + 0.654, + 0.108 + ], + "angle": 0, + "content": "(c) Hidden object navigation" + }, + { + "type": "image", + "bbox": [ + 0.485, + 0.113, + 0.676, + 0.253 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.724, + 0.097, + 0.85, + 0.115 + ], + "angle": 0, + "content": "(d) Hidden object navigation with distractors" + }, + { + "type": "image", + "bbox": [ + 0.682, + 0.113, + 0.871, + 0.253 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.258, + 0.892, + 0.342 + ], + "angle": 0, + "content": "Figure 5. PASTURE object navigation with descriptions. In general object navigation with descriptions is more challenging than the ROBOTOR object navigation task, as indicated by trend lines lying below the \\( y = x \\) line. (a) Appearance descriptions are more helpful than spatial descriptions. (b) Performance further drops when distractor objects are introduced to the environment. However, CoWs are still able to make better use of appearance description than spatial descriptions. (c) Models in the lower success regime (<15% ROBOTOR SUCCESS) perform comparably on finding hidden objects. However, this trend plateaus for higher success models. (d) Trends are similar when distractor objects are introduced for hidden object navigation." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.358, + 0.468, + 0.448 + ], + "angle": 0, + "content": "plore this hypothesis in Appx. E by visualizing CLIP retrieval results on LAION-5B [68] for the uncommon object categories. The relatively high performance on uncommon objects speaks to the flexibility of CoW baselines and their ability to inherit desirable properties from the open-vocabulary models that they are constructed from." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.452, + 0.468, + 0.587 + ], + "angle": 0, + "content": "Can CoWs utilize appearance and spatial descriptions? Looking at Fig. 5 (a) we see that neither appearance nor spatial descriptions improve CoW performance compared to their ROBOTHOR baseline performance (i.e., most points lie under the \\( y = x \\) line). However, CoW is able to take better advantage of appearance descriptions than spatial descriptions. These results motivate future investigation on open-vocabulary object localization with a greater focus on textual object attributes." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.591, + 0.468, + 0.682 + ], + "angle": 0, + "content": "Can CoWs find visible objects in the presence of distractors? In Fig. 5 (b) we see that CoWs experience performance degradation when compared to Fig. 5 (a). We conclude that appearance and spatial attributes added as language input are not sufficient to deal with the complexity of distractors given current open-vocabulary models." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.686, + 0.468, + 0.837 + ], + "angle": 0, + "content": "Can CoWs find hidden objects? Looking at Fig. 5 (c) we notice that models in the lower success regime (less than \\(15\\%\\) SUCCESS on ROBOTHOR) are able to find hidden objects at about the same rate as ROBOTHOR objects (i.e., they lie on the \\(y = x\\) line). OWL models in the higher success regime \\((>15\\%)\\) do not continue this trend; however, they do achieve higher absolute accuracy as seen in Tab. 1. Dealing with occlusion is a longstanding problem in computer vision, and these results provide a foundation upon which future hidden object navigation work can improve." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.468, + 0.901 + ], + "angle": 0, + "content": "Can CoWs find hidden objects in the presence of distractors? Comparing Figs. 5 (c) and (d), we notice similar trends lines, with the best models performing worse with distractors. This suggests that distractors are less of a con" + }, + { + "type": "table", + "bbox": [ + 0.505, + 0.355, + 0.885, + 0.43 + ], + "angle": 0, + "content": "
IDLoc.CoW breedsPASTURE (Avg.)ROBOTHOR
Arch.Exp. StrategySPLSRSPLSR
▲OWLB/32ROBOTHOR learn.10.217.313.120.9
▲OWLB/32HABITAT learn.8.619.49.820.4
▲OWLB/32FBE12.621.116.926.7
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.438, + 0.892, + 0.534 + ], + "angle": 0, + "content": "Table 2. Exploration ablation. For a fixed object localizer (OWL-ViT B/32 with post processing), we ablate over different choices of exploration policy: the FBE heuristic, agents trained in ROBOTHOR, and HABITAT (MP3D). We find that FBE outperforms learnable alternatives on both PASTURE and ROBOTHOR. HABITAT learnable model perform worst, but are not trained on any PASTURE or ROBOTHOR data." + }, + { + "type": "table", + "bbox": [ + 0.507, + 0.54, + 0.885, + 0.603 + ], + "angle": 0, + "content": "
IDCoW breedsObj. PriorPASTURE Uncom.ROBOTHOR
Loc.Arch.SPLSRSPLSR
▲OWLB/32None20.532.816.826.7
▲OWLB/32GPT-3.522.236.917.027.5
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.609, + 0.892, + 0.666 + ], + "angle": 0, + "content": "Table 3. CoW with GPT-3.5 priors. Leveraging GPT-3.5, we generate priors for objects (e.g., apples are likely to be in dining room scenes). Instead of directly searching for the target object, CoW first searches for the scenes, which boosts performance." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.683, + 0.892, + 0.774 + ], + "angle": 0, + "content": "cern in the case of hidden objects than for visible object targets. In light of the fact that detection methods generally work better on larger objects, we hypothesize this effect is because distractor objects are smaller (e.g., apples, vases, basketballs) than objects used to conceal target categories (e.g., beds, sofas, etc.)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.78, + 0.892, + 0.9 + ], + "angle": 0, + "content": "What exploration method performs best? We ablate the decision to use FBE for most experiments by fixing an object localizer (OWL, B/32 \\((\\triangle))\\) and comparing against ROBOTHR learnable exploration and HABITAT learnable exploration in Tab. 2. We notice that FBE performs best in all cases; however, learnable exploration still performs well suggesting that these models do learn useful strategies for the downstream tasks." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "23177" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.125, + 0.096, + 0.422, + 0.262 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.274, + 0.473, + 0.318 + ], + "angle": 0, + "content": "Figure 6. Failure analysis for OWL, B/32 (▲). Exploration and object localization errors occur at similar ratios, with increased localization failures in the presence of distractors." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.33, + 0.47, + 0.437 + ], + "angle": 0, + "content": "Can CoW incorporate object priors? Examining Tab. 3, we see that incorporating GPT-3.5 object-level priors improves performance on both PASTURE uncommon objects and ROBOTOR. These initial results suggest positive trends for incorporating outside knowledge into CoW. Future work may consider more sophisticated methods for injecting priors to steer navigation." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.44, + 0.471, + 0.639 + ], + "angle": 0, + "content": "How do CoWs fail? We identify three high-level failure modes. (1) Exploration fail: the target is never seen. (2) Object localization fail: the target is seen but the localizer never fires. (3) Planning fail: the target is seen and the localizer fires, but planning fails due to inaccuracy in the map representation (Sec. 4.2). Looking at Fig. 6, we notice a large fraction of failures are due to exploration and object localization. This suggests CoWs may continue to improve as research in these fields progress. In Fig. 6 we also see that in cases where distractors are present a higher fraction of object localization failures occur, further suggesting that open-vocabulary models currently struggle to make full use of attribute prompts. See Appx. I for more failure analysis." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.649, + 0.307, + 0.667 + ], + "angle": 0, + "content": "6.3. Comparison to Prior Art" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.673, + 0.469, + 0.749 + ], + "angle": 0, + "content": "We primarily evaluate CoWs in general L-ZSON settings; however, we further evaluate CoWs on ZSON benchmarks to establish them as a strong baseline for these tasks. Recall, ZSON can be seen as a case of L-ZSON where only object goals are specified (no attributes)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.471, + 0.902 + ], + "angle": 0, + "content": "In Tab. 4, we see there exists a CoW that outperforms the end-to-end baselines in all cases except SUCCESS on HABITAT (MP3D). For instance, the CLIP-Grad., B/32 (\\(\\triangleleft\\)) matches the SemanticNav-ZSON model on HABITAT (MP3D) SPL: 4.9 for CoW v.s. 4.8 for the competitor, while improving over EmbCLIP-ZSON ROBOTHOR by 15.6 percentage points. To contextualize this result, CoWs train for \\(O\\) navigation steps, while SemanticNav-ZSON and EmbCLIP-ZSON train in the target evaluation simulators for 500M and 60M steps respectively." + }, + { + "type": "table", + "bbox": [ + 0.516, + 0.09, + 0.88, + 0.196 + ], + "angle": 0, + "content": "
IDCoW breedsHABITAT (MP3D)ROBOTHON (subset)ROBOTHON (full)Nav. training steps
Loc.Arch.SPLSRSPLSRSPLSR
▲CLIP-Grad.B/324.99.215.023.79.715.20
▲OWLB/323.77.420.832.516.926.70
EmbCLIP-ZSON [38]---8.1-14.0*60M
SemanticNav-ZSON [46]4.815.3----500M
" + }, + { + "type": "table_caption", + "bbox": [ + 0.498, + 0.205, + 0.892, + 0.304 + ], + "angle": 0, + "content": "Table 4. Comparison to prior art on existing ZSON benchmarks. CoWs are able to match or out-compete existing methods that leverage millions of steps of navigation training in the evaluation simulator. *indicates a result from prior work that includes, non-zero-shot evaluation. Specifically, only 1/4 of the evaluations are zero-shot on ROBOTOR (subset) and the remaining 3/4 on categories seen during training." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.318, + 0.892, + 0.38 + ], + "angle": 0, + "content": "The superior performance of SemanticNav-ZSON in terms of MP3D SUCCESS indicates that there can be benefits to in-domain learning. Future work may consider unifying the benefits of CoW-like models and fine-tuned models." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.393, + 0.757, + 0.409 + ], + "angle": 0, + "content": "7. Limitations and Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.421, + 0.892, + 0.633 + ], + "angle": 0, + "content": "Limitations. While our evaluation of CoWs on HABITAT, ROBOTHER, and PASTURE is a step towards assessing their performance in different domains, ultimately, real-world performance matters most. Hence, the biggest limitation of our work is the lack of large-scale, real-world benchmarking—which is also missing in much of the related literature. Additionally, CoW inherits the metalimitations of the object localization and exploration methods considered. For example, object localizers require tuning a confidence threshold to balance precision and recall. Finally, we do not consider different agent embodiment or continuous action spaces. This is a pertinent investigation given recent findings of Pratt et al. [59] that agent morphology can be a big determinant of downstream performance." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.636, + 0.892, + 0.773 + ], + "angle": 0, + "content": "Conclusion. This paper introduces the PASTURE benchmark for language-driven zero-shot object navigation and several CLIP on Wheels baselines, translating the successes of existing zero-shot models to an embodied task. We view CoW as an instance of using open-vocabulary models, with text-based interfaces, to tackle robotics tasks in more flexible settings. We hope that the baselines and the proposed benchmark will spur the field to explore broader and more powerful forms of zero-shot embodied AI." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.777, + 0.892, + 0.897 + ], + "angle": 0, + "content": "Acknowledgement. We would like to thank Jessie Chapman, Cheng Chi, Huy Ha, Zeyi Liu, Sachit Menon, and Sarah Pratt for valuable feedback. This work was supported in part by NSF CMMI-2037101, NSF IIS-2132519, and an Amazon Research Award. SYG is supported by a NSF Graduate Research Fellowship. The views and conclusions contained herein are those of the authors and should not be interpreted as necessarily representing the official policies, either expressed or implied, of the sponsors." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "23178" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.175, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.169 + ], + "angle": 0, + "content": "[1] Ziad Al-Halah, Santhosh K. Ramakrishnan, and Kristen Grauman. Zero experience required: Plug & play modular transfer learning for semantic visual navigation. arXiv, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.172, + 0.472, + 0.24 + ], + "angle": 0, + "content": "[2] Peter Anderson, Qi Wu, Damien Teney, Jake Bruce, Mark Johnson, Niko Sunderhauf, Ian Reid, Stephen Gould, and Anton Van Den Hengel. Vision-and-language navigation: Interpreting visually-grounded navigation instructions in real environments. CVPR, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.242, + 0.471, + 0.282 + ], + "angle": 0, + "content": "[3] Donghyeon Baek, Youngmin Oh, and Bumsub Ham. Exploiting a joint embedding space for generalized zero-shot semantic segmentation. ICCV, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.284, + 0.471, + 0.324 + ], + "angle": 0, + "content": "[4] Ankan Bansal, Karan Sikka, Gaurav Sharma, Rama Chellappa, and Ajay Divakaran. Zero-shot object detection. ECCV, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.326, + 0.47, + 0.394 + ], + "angle": 0, + "content": "[5] Dhruv Batra, Aaron Gokaslan, Aniruddha Kembhavi, Oleksandr Maksymets, Roozbeh Mottaghi, Manolis Savva, Alexander Toshev, and Erik Wijmans. Objectnav revisited: On evaluation of embodied agents navigating to objects. arXiv, 2020. 1, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.396, + 0.471, + 0.436 + ], + "angle": 0, + "content": "[6] Maxime Bucher, Tuan-Hung VU, Matthieu Cord, and Patrick Pérez. Zero-shot semantic segmentation. NeurIPS, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.438, + 0.47, + 0.478 + ], + "angle": 0, + "content": "[7] Yuri Burda, Harrison Edwards, Amos J. Storkey, and Oleg Klimov. Exploration by random network distillation. *ICLR*, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.48, + 0.471, + 0.521 + ], + "angle": 0, + "content": "[8] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. ECCV, 2020. 4, 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.523, + 0.471, + 0.562 + ], + "angle": 0, + "content": "[9] Matthew Chang, Arjun Gupta, and Saurabh Gupta. Semantic visual navigation by watching youtube videos. NeurIPS, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.565, + 0.47, + 0.606 + ], + "angle": 0, + "content": "[10] Devendra Singh Chaplot, Dhiraj Gandhi, Abhinav Gupta, and Ruslan Salakhutdinov. Object goal navigation using goal-oriented semantic exploration. NeurIPS, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.607, + 0.47, + 0.647 + ], + "angle": 0, + "content": "[11] Devendra Singh Chaplot, Dhiraj Gandhi, Saurabh Gupta, Abhinav Kumar Gupta, and Ruslan Salakhutdinov. Learning to explore using active neural slam. ICLR, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.649, + 0.469, + 0.69 + ], + "angle": 0, + "content": "[12] Prithvijit Chattopadhyay, Judy Hoffman, Roozbeh Mottaghi, and Aniruddha Kembhavi. Robustnav: Towards benchmarking robustness in embodied navigation. ICCV, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.691, + 0.47, + 0.732 + ], + "angle": 0, + "content": "[13] Hila Chefer, Shir Gur, and Lior Wolf. Transformer interpretability beyond attention visualization. CVPR, 2021. 4, 13, 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.734, + 0.47, + 0.761 + ], + "angle": 0, + "content": "[14] Tao Chen, Saurabh Gupta, and Abhinav Gupta. Learning exploration policies for navigation. *ICLR*, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.762, + 0.47, + 0.815 + ], + "angle": 0, + "content": "[15] Jiaxin Cheng, Soumyaroop Nandi, Prem Natarajan, and Wael Abd-Almageed. Sign: Spatial-information incorporated generative network for generalized zero-shot semantic segmentation. ICCV, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.818, + 0.47, + 0.871 + ], + "angle": 0, + "content": "[16] Kyunghyun Cho, Bart van Merrienboer, Dzmitry Bahdanau, and Yoshua Bengio. On the properties of neural machine translation: Encoder-decoder approaches. In SSST@EMNLP, 2014. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.47, + 0.901 + ], + "angle": 0, + "content": "[17] Andrew J Davison and David W Murray. Mobile robot localisation using active vision. ECCV, 1998. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.174 + ], + "angle": 0, + "content": "[18] Matt Deitke, Winson Han, Alvaro Herrasti, Aniruddha Kembhavi, Eric Kolve, Roozbeh Mottaghi, Jordi Salvador, Dustin Schwenk, Eli VanderBilt, Matthew Wallingford, Luca Weihs, Mark Yatskar, and Ali Farhadi. Robothor: An open simulation-to-real embodied ai platform. CVPR, 2020. 1, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.176, + 0.892, + 0.244 + ], + "angle": 0, + "content": "[19] Matt Deitke, Eli VanderBilt, Alvaro Herrasti, Luca Weihs, Jordi Salvador, Kiana Ehsani, Winson Han, Eric Kolve, Ali Farhadi, Aniruddha Kembhavi, and Roozbeh Mottaghi. Proctor: Large-scale embodied ai using procedural generation. NeurIPS, 2022. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.246, + 0.892, + 0.285 + ], + "angle": 0, + "content": "[20] Frank Dellaert, Dieter Fox, Wolfram Burgard, and Sebastian Thrun. Monte carlo localization for mobile robots. ICRA, 1999. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.288, + 0.892, + 0.328 + ], + "angle": 0, + "content": "[21] Berkan Demirel, Ramazan Gokberk Cinbis, and Nazli Ikizler-Cinbis. Zero-shot object detection by hybrid region embedding. BMVC, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.33, + 0.892, + 0.37 + ], + "angle": 0, + "content": "[22] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, K. Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. CVPR, 2009. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.372, + 0.892, + 0.44 + ], + "angle": 0, + "content": "[23] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv, 2020. 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.442, + 0.892, + 0.523 + ], + "angle": 0, + "content": "[24] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. ICLR, 2020. 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.525, + 0.892, + 0.565 + ], + "angle": 0, + "content": "[25] Samir Yitzhak Gadre, Kiana Ehsani, Shuran Song, and Roozbeh Mottaghi. Continuous scene representations for embodied ai. CVPR, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.567, + 0.892, + 0.607 + ], + "angle": 0, + "content": "[26] Daniel Gordon, Abhishek Kadian, Devi Parikh, Judy Hoffman, and Dhruv Batra. Splitnet: Sim2sim and task2task transfer for embodied visual navigation. CVPR, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.609, + 0.892, + 0.649 + ], + "angle": 0, + "content": "[27] Xiuye Gu, Tsung-Yi Lin, Weicheng Kuo, and Yin Cui. Zeroshot detection via vision and language knowledge distillation. arXiv, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.651, + 0.892, + 0.69 + ], + "angle": 0, + "content": "[28] Agrim Gupta, Piotr Dólar, and Ross B. Girshick. Lvis: A dataset for large vocabulary instance segmentation. CVPR, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.692, + 0.892, + 0.747 + ], + "angle": 0, + "content": "[29] Abhinav Kumar Gupta, Adithyavairavan Murali, Dhiraj Gandhi, and Terrel Pinto. Robot learning in homes: Improving generalization and reducing dataset bias. NeurIPS, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.749, + 0.892, + 0.801 + ], + "angle": 0, + "content": "[30] Saurabh Gupta, James Davidson, Sergey Levine, Rahul Sukthankar, and Jitendra Malik. Cognitive mapping and planning for visual navigation. CVPR, pages 2616-2625, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.804, + 0.892, + 0.845 + ], + "angle": 0, + "content": "[31] Meera Hahn, Devendra Singh Chaplot, and Shubham Tulsiani. No rl, no simulation: Learning to navigate without navigating. NeurIPS, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.847, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[32] Peter Henry, Michael Krainin, Evan Herbst, Xiaofeng Ren, and Dieter Fox. RGB-d mapping: Using depth cameras for dense 3d modeling of indoor environments. Experimental robotics, 2014. 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "23179" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.469, + 0.133 + ], + "angle": 0, + "content": "[33] Ping Hu, Stan Sclaroff, and Kate Saenko. Uncertainty-aware learning for zero-shot semantic segmentation. NeurIPS, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.136, + 0.469, + 0.177 + ], + "angle": 0, + "content": "[34] Chenguang Huang, Oier Mees, Andy Zeng, and Wolfram Burgard. Visual language maps for robot navigation. arXiv preprint arXiv:2210.05714, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.179, + 0.469, + 0.234 + ], + "angle": 0, + "content": "[35] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc V Le, Yunhsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. ICML, 2021. 1, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.236, + 0.469, + 0.289 + ], + "angle": 0, + "content": "[36] Aishwarya Kamath, Mannat Singh, Yann LeCun, Ishan Misra, Gabriel Synnaeve, and Nicolas Carion. Mdetr - modulated detection for end-to-end multi-modal understanding. ICCV, 2021. 1, 4, 13, 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.292, + 0.468, + 0.332 + ], + "angle": 0, + "content": "[37] Naoki Kato, Toshihiko Yamasaki, and Kiyoharu Aizawa. Zero-shot semantic segmentation via variational mapping. ICCV-W, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.334, + 0.469, + 0.376 + ], + "angle": 0, + "content": "[38] Apoorv Khandelwal, Luca Weihs, Roozbeh Mottaghi, and Aniruddha Kembhavi. Simple but effective: Clip embeddings for embodied ai. arXiv, 2021. 1, 2, 3, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.378, + 0.469, + 0.433 + ], + "angle": 0, + "content": "[39] Jacob Krantz, Erik Wijmans, Arjun Majumdar, Dhruv Batra, and Stefan Lee. Beyond the nav-graph: Vision-and-language navigation in continuous environments. In European Conference on Computer Vision, pages 104-120. Springer, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.435, + 0.469, + 0.489 + ], + "angle": 0, + "content": "[40] Alexander Ku, Peter Anderson, Roma Patel, Eugene Ie, and Jason Baldridge. Room-across-room: Multilingual vision-and-language navigation with dense spatiotemporal grounding. arXiv preprint arXiv:2010.07954, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.491, + 0.469, + 0.544 + ], + "angle": 0, + "content": "[41] Benjamin Kuipers and Yung-Tai Byun. A robot exploration and mapping strategy based on a semantic hierarchy of spatial representations. Robotics and autonomous systems, 1991. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.547, + 0.469, + 0.587 + ], + "angle": 0, + "content": "[42] Xinghang Li, Di Guo, Huaping Liu, and Fuchun Sun. Revece: Remote embodied visual referring expression in continuous environment. RA-L, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.59, + 0.469, + 0.631 + ], + "angle": 0, + "content": "[43] Zhihui Li, Lina Yao, Xiaqin Zhang, Xianzhi Wang, Salil S. Kanhere, and Huaxiang Zhang. Zero-shot object detection with textual descriptions. AAAI, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.633, + 0.469, + 0.674 + ], + "angle": 0, + "content": "[44] Yiqing Liang, Boyuan Chen, and Shuran Song. Sscnav: Confidence-aware semantic scene completion for visual semantic navigation. ICRA, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.676, + 0.469, + 0.731 + ], + "angle": 0, + "content": "[45] Tsung-Yi Lin, Michael Maire, Serge Belongie, Lubomir Bourdev, Ross Girshick, James Hays, Pietro Perona, Deva Ramanan, C. Lawrence Zitnick, and Piotr Dólar. Microsoft coco: Common objects in context. ECCV, 2014. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.733, + 0.469, + 0.788 + ], + "angle": 0, + "content": "[46] Arjun Majumdar, Gunjan Aggarwal, Bhavika Devnani, Judy Hoffman, and Dhruv Batra. Zson: Zero-shot object-goal navigation using multimodal goal embeddings. arXiv preprint arXiv:2206.12403, 2022. 1, 2, 3, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.789, + 0.469, + 0.844 + ], + "angle": 0, + "content": "[47] Qiaomei Mao, Chong Wang, Sheng Yu, Ye Zheng, and Yuqi Li. Zero-shot object detection with attributes-based category similarity. IEEE Transactions on Circuits and Systems II: Express Briefs, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.846, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[48] Lina Mezghani, Sainbayar Sukhbaatar, Thibaut Lavril, Oleksandr Maksymets, Dhruv Batra, Piotr Bojanowski, and Alahari Karteek. Memory-augmented reinforcement learning for image-goal navigation. arXiv, 2021. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.469, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.175 + ], + "angle": 0, + "content": "[49] Matthias Minderer, Alexey A. Gritsenko, Austin Stone, Maxim Neumann, Dirk Weissenborn, Alexey Dosovitskiy, Aravindh Mahendran, Anurag Arnab, Mostafa Dehghani, Zhuoran Shen, Xiao Wang, Xiaohua Zhai, Thomas Kipf, and Neil Houlsby. Simple open-vocabulary object detection with vision transformers. ECCV, 2022. 1, 4, 6, 13, 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.178, + 0.892, + 0.233 + ], + "angle": 0, + "content": "[50] Grégoire Montavon, Sebastian Lapuschkin, Alexander Binder, Wojciech Samek, and Klaus-Robert Müller. Explaining nonlinear classification decisions with deep taylor decomposition. Pattern recognition, 65:211-222, 2017. 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.235, + 0.892, + 0.261 + ], + "angle": 0, + "content": "[51] Hans Moravec and Alberto Elfes. High resolution maps from wide angle sonar. ICRA, 1985. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.263, + 0.892, + 0.331 + ], + "angle": 0, + "content": "[52] Richard A Newcombe, Shahram Izadi, Otmar Hilliges, David Molyneaux, David Kim, Andrew J Davison, Pushmeet Kohi, Jamie Shotton, Steve Hodges, and Andrew Fitzgibbon. Kinectfusion: Real-time dense surface mapping and tracking. ISMAR, 2011. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.334, + 0.892, + 0.374 + ], + "angle": 0, + "content": "[53] Lachlan Nicholson, Michael Milford, and Niko Sunderhauf. Quadricslam: Dual quadrics from object detections as landmarks in object-oriented slam. RA-L, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.377, + 0.892, + 0.404 + ], + "angle": 0, + "content": "[54] Clark F Olson and Larry H Matthies. Maximum likelihood rover localization by matching range maps. ICRA, 1998. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.406, + 0.892, + 0.473 + ], + "angle": 0, + "content": "[55] Long Ouyang, Jeff Wu, Xu Jiang, Diogo Almeida, Carroll L Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, et al. Training language models to follow instructions with human feedback. NeurIPS, 2022. 4, 15" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.476, + 0.892, + 0.517 + ], + "angle": 0, + "content": "[56] Simone Parisi, Victoria Dean, Deepak Pathak, and Abhinav Kumar Gupta. Interesting object, curious agent: Learning task-agnostic exploration. NeurIPS, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.519, + 0.892, + 0.559 + ], + "angle": 0, + "content": "[57] Deepak Pathak, Pulkit Agrawal, Alexei A. Efros, and Trevor Darrell. Curiosity-driven exploration by self-supervised prediction. ICML, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.562, + 0.892, + 0.617 + ], + "angle": 0, + "content": "[58] Hieu Pham, Zihang Dai, Golnaz Ghiasi, Hanxiao Liu, Adams Wei Yu, Minh-Thang Luong, Mingxing Tan, and Quoc V. Le. Combined scaling for zero-shot transfer learning. arXiv, 2021. 1, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.619, + 0.892, + 0.659 + ], + "angle": 0, + "content": "[59] Sarah Pratt, Luca Weihs, and Ali Farhadi. The introspective agent: Interdependence of strategy, physiology, and sensing for embodied agents. arXiv, 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.661, + 0.892, + 0.716 + ], + "angle": 0, + "content": "[60] Yuankai Qi, Qi Wu, Peter Anderson, Xin Wang, William Yang Wang, Chunhua Shen, and Anton van den Hengel. Reverie: Remote embodied visual referring expression in real indoor environments. CVPR, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.718, + 0.892, + 0.8 + ], + "angle": 0, + "content": "[61] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. ICML, 2021. 1, 2, 4, 6, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.802, + 0.892, + 0.856 + ], + "angle": 0, + "content": "[62] Shafin Rahman, Salman Hameed Khan, and Fatih Murat Porkikli. Zero-shot object detection: Learning to simultaneously recognize and localize novel concepts. ACCV, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.859, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[63] Roberta Raileanu and Tim Rocktäschel. Ride: Rewarding impact-driven exploration for procedurally-generated environments. *ICLR*, 2020. 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "23180" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.147 + ], + "angle": 0, + "content": "[64] Santhosh K. Ramakrishnan, Devendra Singh Chaplot, Ziad Al-Halah, Jitendra Malik, and Kristen Grauman. Poni: Potential functions for objectgoal navigation with interaction-free learning. CVPR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.151, + 0.47, + 0.219 + ], + "angle": 0, + "content": "[65] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, Alexander C. Berg, and Li Fei-Fei. Imagenet large scale visual recognition challenge. IJCV, 2015. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.222, + 0.469, + 0.276 + ], + "angle": 0, + "content": "[66] Manolis Savva, Angel X Chang, Alexey Dosovitskiy, Thomas Funkhouser, and Vladlen Koltun. Minos: Multimodal indoor simulator for navigation in complex environments. arXiv, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.28, + 0.469, + 0.348 + ], + "angle": 0, + "content": "[67] Manolis Savva, Abhishek Kadian, Oleksandr Maksymets, Yili Zhao, Erik Wijmans, Bhavana Jain, Julian Straub, Jia Liu, Vladlen Koltun, Jitendra Malik, Devi Parikh, and Dhruv Batra. Habitat: A Platform for Embodied AI Research. ICCV, 2019. 1, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.351, + 0.469, + 0.42 + ], + "angle": 0, + "content": "[68] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. NeurIPS, 2022. 7, 16, 18" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.423, + 0.469, + 0.464 + ], + "angle": 0, + "content": "[69] John Schulman, Filip Wolski, Prafulla Dhariwal, Alec Radford, and Oleg Klimov. Proximal policy optimization algorithms. arXiv, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.467, + 0.469, + 0.522 + ], + "angle": 0, + "content": "[70] Ramprasaath R. Selvaraju, Abhishek Das, Ramakrishna Vedantam, Michael Cogswell, Devi Parikh, and Dhruv Batra. Grad-cam: Visual explanations from deep networks via gradient-based localization. IJCV, 2019. 4, 13, 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.525, + 0.469, + 0.58 + ], + "angle": 0, + "content": "[71] Mohit Shridhar, Jesse Thomason, Daniel Gordon, Yonatan Bisk, Winson Han, Roozbeh Mottaghi, Luke Zettlemoyer, and Dieter Fox. Alfred: A benchmark for interpreting grounded instructions for everyday tasks. CVPR, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.582, + 0.469, + 0.623 + ], + "angle": 0, + "content": "[72] Shuran Song, Linguang Zhang, and Jianxiong Xiao. Robot in a room: Toward perfect object recognition in closed environments. arXiv, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.626, + 0.469, + 0.667 + ], + "angle": 0, + "content": "[73] Alexander L. Strehl and Michael L. Littman. An analysis of model-based interval estimation for markov decision processes. J. Comput. Syst. Sci., 2008, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.67, + 0.469, + 0.711 + ], + "angle": 0, + "content": "[74] Saim Wani, Shivansh Patel, Unnat Jain, Angel X. Chang, and Manolis Savva. Mution: Benchmarking semantic map memory using multi-object navigation. NeurIPS, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.714, + 0.469, + 0.755 + ], + "angle": 0, + "content": "[75] Luca Weihs, Matt Deitke, Aniruddha Kembhavi, and Roozbeh Mottaghi. Visual room rearrangement. CVPR, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.758, + 0.469, + 0.812 + ], + "angle": 0, + "content": "[76] Luca Weihs, Jordi Salvador, Klemen Kotar, Unnat Jain, KuoHao Zeng, Roozbeh Mottaghi, and Aniruddha Kembhavi. Allenact: A framework for embodied ai research. arXiv, 2020. 3, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.816, + 0.469, + 0.87 + ], + "angle": 0, + "content": "[77] Erik Wijmans, Abhishek Kadian, Ari S. Morcos, Stefan Lee, Irfan Essa, D. Parikh, Manolis Savva, and Dhruv Batra. Dppo: Learning near-perfect pointgoal navigators from 2.5 billion frames. *ICLR*, 2019. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.468, + 0.901 + ], + "angle": 0, + "content": "[78] Brian H Wilcox. Robotic vehicles for planetary exploration. Applied Intelligence, 1992. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.147 + ], + "angle": 0, + "content": "[79] Mitchell Wortsman, Kiana Ehsani, Mohammad Rastegari, Ali Farhadi, and Roozbeh Mottaghi. Learning to learn how to learn: Self-adaptive visual navigation using meta-learning. CVPR, 2019. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.149, + 0.892, + 0.19 + ], + "angle": 0, + "content": "[80] Chenyun Wu, Zhe Lin, Scott D. Cohen, Trung Bui, and Subhransu Maji. Phrasecut: Language-based image segmentation in the wild. CVPR, 2020. 4, 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.193, + 0.892, + 0.233 + ], + "angle": 0, + "content": "[81] Fei Xia, Amir R Zamir, Zhiyang He, Alexander Sax, Jitendra Malik, and Silvio Savarese. Gibson env: Real-world perception for embodied agents. CVPR, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.235, + 0.892, + 0.262 + ], + "angle": 0, + "content": "[82] Brian Yamauchi. A frontier-based approach for autonomous exploration. CIRA, 1997. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.264, + 0.892, + 0.305 + ], + "angle": 0, + "content": "[83] Claudia Yan, Dipendra Misra, Andrew Bennnett, Aaron Walsman, Yonatan Bisk, and Yoav Artzi. Chalet: Cornell house agent learning environment. arXiv, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.307, + 0.892, + 0.348 + ], + "angle": 0, + "content": "[84] Wei Yang, Xiaolong Wang, Ali Farhadi, Abhinav Gupta, and Roozbeh Mottaghi. Visual semantic navigation using scene priors. arXiv, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.35, + 0.892, + 0.403 + ], + "angle": 0, + "content": "[85] Xiaohua Zhai, Xiao Wang, Basil Mustafa, Andreas Steiner, Daniel Keysers, Alexander Kolesnikov, and Lucas Beyer. Lit: Zero-shot transfer with locked-image text tuning. CVPR, 2022. 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.406, + 0.892, + 0.433 + ], + "angle": 0, + "content": "[86] Chong Zhou, Chen Change Loy, and Bo Dai. Denseclip: Extract free dense labels from clip. arXiv, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.435, + 0.892, + 0.476 + ], + "angle": 0, + "content": "[87] Fengda Zhu, Xiwen Liang, Yi Zhu, Xiaojun Chang, and Xiaodan Liang. Soon: Scenario oriented object navigation with graph-based exploration. CVPR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.478, + 0.892, + 0.519 + ], + "angle": 0, + "content": "[88] Pengkai Zhu, Hanxiao Wang, and Venkatesh Saligrama. Zero shot detection. IEEE Transactions on Circuits and Systems for Video Technology, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.521, + 0.892, + 0.575 + ], + "angle": 0, + "content": "[89] Yuke Zhu, Roozbeh Mottaghi, Eric Kolve, Joseph J Lim, Abhinav Gupta, Li Fei-Fei, and Ali Farhadi. Target-driven visual navigation in indoor scenes using deep reinforcement learning. ICRA, 2017. 1, 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.575 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "23181" + } + ] +] \ No newline at end of file diff --git a/2023/CoWs on Pasture_ Baselines and Benchmarks for Language-Driven Zero-Shot Object Navigation/a31a35f2-4957-4777-92dc-31eae2e5deb5_origin.pdf b/2023/CoWs on Pasture_ Baselines and Benchmarks for Language-Driven Zero-Shot Object Navigation/a31a35f2-4957-4777-92dc-31eae2e5deb5_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..1e0400118eaa3bc46afc2e771de06a24c535c322 --- /dev/null +++ b/2023/CoWs on Pasture_ Baselines and Benchmarks for Language-Driven Zero-Shot Object Navigation/a31a35f2-4957-4777-92dc-31eae2e5deb5_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6a8160f5d1f73ca49d0e01391d6092c7393fecb6ee58abf874e700bf1daa903e +size 7174877 diff --git a/2023/CoWs on Pasture_ Baselines and Benchmarks for Language-Driven Zero-Shot Object Navigation/full.md b/2023/CoWs on Pasture_ Baselines and Benchmarks for Language-Driven Zero-Shot Object Navigation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..a836e0739bbea5944cd05e5c2bb9130d4d7b0107 --- /dev/null +++ b/2023/CoWs on Pasture_ Baselines and Benchmarks for Language-Driven Zero-Shot Object Navigation/full.md @@ -0,0 +1,346 @@ +# CoWs on PASTURE: Baselines and Benchmarks for Language-Driven Zero-Shot Object Navigation + +Samir Yitzhak Gadre $^{\diamond}$ Mitchell Wortsman $^{\dagger}$ Gabriel Ilharco $^{\dagger}$ Ludwig Schmidt $^{\dagger}$ Shuran Song $^{\diamond}$ + +# Abstract + +For robots to be generally useful, they must be able to find arbitrary objects described by people (i.e., be language-driven) even without expensive navigation training on in-domain data (i.e., perform zero-shot inference). We explore these capabilities in a unified setting: language-driven zero-shot object navigation (L-ZSON). Inspired by the recent success of open-vocabulary models for image classification, we investigate a straightforward framework, CLIP on Wheels (CoW), to adapt open-vocabulary models to this task without fine-tuning. To better evaluate L-ZSON, we introduce the PASTURE benchmark, which considers finding uncommon objects, objects described by spatial and appearance attributes, and hidden objects described relative to visible objects. We conduct an in-depth empirical study by directly deploying 22 CoW baselines across HABITAT, ROBOTHOR, and PASTURE. In total, we evaluate over 90k navigation episodes and find that (1) CoW baselines often struggle to leverage language descriptions but are proficient at finding uncommon objects. (2) A simple CoW, with CLIP-based object localization and classical exploration—and no additional training—matches the navigation efficiency of a state-of-the-art ZSON method trained for 500M steps on HABITAT MP3D data. This same CoW provides a 15.6 percentage point improvement in success over a state-of-the-art ROBOTHOR ZSON model. + +# 1. Introduction + +To be more widely applicable, robots should be language-driven: able to deduce goals based on arbitrary text input instead of being constrained to a fixed set of object categories. While existing image classification, semantic segmentation, and object navigation benchmarks like ImageNet-1k [65], ImageNet-21k [22], MS-COCO [45], LVIS [28], HABITAT [67], and ROBOTHOR [18] include a vast array of everyday items, they do not capture all objects that matter to people. For instance, a lost "toy airplane" may + +# Sample tasks Top-down visualization Egocentric Observations + +(a) Finding uncommon objects + +![](images/16652846a9fca43c2db6cb634fa5e8d0256f61a82637d3a954319e5ebb771b2f.jpg) + +(b) Finding objects based on attributes in the presence of distractors + +![](images/31f0ab0fe5e9a07a5095a87d64338440d54a55360dd693f2b489171f040d34fa.jpg) + +(c) Finding hidden objects in the presence of distractors + +![](images/866fb55efa25e9630e6e03edbb445623423d7bcd17120fc94c04cd76ad751fdc.jpg) +Figure 1. The PASTURE benchmark for L-ZSON. Text specifies navigation goal objects. Agents do not train on these tasks, making the evaluation protocol zero-shot. (a) Uncommon object goals like "llama wicker basket", not found in existing navigation benchmarks. (b) Appearance, spatial descriptions, which are necessary to find the correct object. (c) Hidden object descriptions, which localize objects that are not visible. + +become relevant in a kindergarten classroom, but this object is not annotated in any of the aforementioned datasets. + +In this paper, we study Language-driven zero-shot object navigation (L-ZSON), a more challenging but also more applicable version of object navigation [5, 18, 67, 79, 89] and ZSON [38, 46] tasks. In L-ZSON, an agent must find an object based on a description, which may contain different levels of granularity (e.g., "toy airplane", "toy airplane under the bed", or "wooden toy airplane"). L-ZSON encompasses ZSON, which specifies only the target category [38, 46]. Since L-ZSON is "zero-shot", we consider agents without access to navigation training on the target objects or domains. This reflects realistic application scenarios, where the environment and object set may not be known a priori. + +Performing L-ZSON in any environment with unstructured language input is challenging; however, recent advances in open-vocabulary models for image classification [35, 58, 61], object detection [4, 21, 27, 36, 43, 47, 49, 62, 88], and semantic segmentation [3, 6, 15, 33, 36, 37, 86] present a promising foundation. These models provide an interface where one specifies—in text—the arbitrary ob + +jects they wish to classify, detect, or segment. For example, CLIP [61] open-vocabulary classifiers compute similarity scores between an input image and a set of user-specified captions (e.g., "a photo of a toy airplane.", ..., selecting the caption with the highest score to determine the image classification label. Given the flexibility of these models, we would like to understand their capability to execute embodied tasks even without additional training. + +To this end, we present baselines and benchmarks for L-ZSON. More specifically: + +- A collection of baseline algorithms, CLIP on Wheels (CoW), which adapt open-vocabulary models to the task of L-ZSON. CoW takes inspiration from the semantic mapping line of work [11, 41, 53], and decomposes the navigation task into exploration when the language target is not confidently localized, and target-driven planning otherwise. CoW retains the textual user interface of the original open-vocabulary model and does not require any navigation training. We evaluate 22 CoWs, ablating over many open-vocabulary models, exploration policies, backbones, prompting strategies, and post-processing strategies. +- A new benchmark, PASTURE, to evaluate CoW and future methods on L-ZSON. We design PASTURE, visualized in Fig. 1, to study capabilities that traditional object navigation agents, which are trained on a fixed set of categories, do not possess. We consider the ability to find (1) uncommon objects (e.g., "tie-dye surfboard"), (2) objects by their spatial and appearance attributes in the presence of distractor objects (e.g., "green apple" vs. "red apple"), and (3) objects that cannot be visually observed (e.g., "mug under the bed"). + +Together the CoW baselines and PASTURE benchmark allow us to conduct extensive studies on the capabilities of open-vocabulary models in the context of L-ZSON embodied tasks. Our experiments demonstrate CoW's potential on uncommon objects and limitations in taking full advantage of language descriptions—thereby providing empirical motivation for future studies. To contextualize CoW relative to prior zero-shot methods, we additionally evaluate on the HABITAT MP3D dataset. We find that our best CoW achieves navigation efficiency (SPL) that matches a state-of-the-art ZSON method [46] that trains on MP3D training data for 500M steps. On a ROBOTHOR object subset, considered in prior work, the same CoW beats the leading method [38] by 15.6 percentage points in task success. + +# 2. Related Work + +Mapping and exploration. Exploring effectively with a mobile robot is a long-standing problem in vision and robotics. Classical methods often decompose the task + +into map reconstruction [30, 32, 51, 52, 72], agent localization [17, 20, 54], and planning [41, 78]. Recent work investigates learned alternatives for exploration [7, 14, 56, 57, 63]. Here, agents are often trained end-to-end with self-supervised rewards (e.g., curiosity [57]) or supervised rewards (e.g., state visitation counts [25, 73, 75]). Learning-based methods typically need less hand-tuning, but require millions of training steps and reward engineering. We test both classical and learnable exploration strategies in the context of CoW to study their applicability to L-ZSON. + +Goal-conditioned navigation. Apart from open-ended exploration, many navigation tasks are goal-conditioned, where the agent needs to navigate to a specified position (i.e., point goal [11, 12, 26, 31, 66, 77, 81, 83]), view of the environment (i.e., image goal [48, 64, 89]), or object category (i.e., object goal [1, 9, 10, 12, 19, 44, 74, 79, 84]). We consider an object goal navigation task. + +Vision-Language Navigation. Prior work investigates language-based navigation, where language provides step-by-step instructions for the task [2, 34, 39, 40, 71]. This line of work demonstrates the benefits of additional language input for robot navigation, especially for long-horizon tasks (e.g., room-to-room navigation [40]). However, providing detailed step-by-step instructions (e.g., move 3 meters south [34]) could be challenging and time-consuming. In our L-ZSON task, an algorithm gets natural language as the goal description instead of low-level instructions. Prior work also investigates navigation with target descriptions in supervised settings [42, 60, 87]. In contrast, we explore a zero-shot evaluation protocol and consider finding hidden objects (e.g., "mug under bed") and uncommon objects (e.g., "tie-dye surfboard"). + +Zero-shot object navigation (ZSON). Recent work studies object navigation in zero-shot settings, where agents are evaluated on object categories that they are not explicitly trained on [38, 46]. Our task encompasses ZSON; however it also considers cases where more information—object attributes or hidden objects descriptions—is specified. Khandelwal et al. [38] train on a subset of ROBOTHOR categories and evaluate on a held-out set. In concurrent work, Majumdar et al. [46] train on an image goal navigation task and evaluate on object navigation downstream by leveraging CLIP multi-modal embeddings. Both algorithms necessitate navigation training for millions of steps and train separate models for each simulation domain. In contrast, CoW baselines do not necessitate any simulation training and can be deployed in multiple environments. + +# 3. The L-ZSON Task + +Language-driven zero-shot object navigation (L-ZSON) involves navigating to goal objects, specified in language, without explicit training to do so. Let $\mathcal{O}$ denote a set of + +![](images/432331edfae158dc49e0cd33234c1db9de98c336e4e79390c16442d512f8dfb2.jpg) +Figure 2. CLIP on Wheels (CoW) overview. A CoW uses a policy to explore and an object localizer (e.g., an open-vocabulary object detector) to determine if an object goal is in view. +Figure 3. Mapping. Top-down map created from egocentric depth observations as a CoW roams a space. (a) Frontier Based Exploration [82] showing a planned path exploration path to the next frontier. (b) Back-projected object relevance scores provide object goal targets when a CoW has found an object. + +natural language descriptions of target objects with potentially many attributes (e.g., "plant", "snake plant", "plant under the bed", etc.). This contrasts with definitions studied in prior object navigation [5, 18] and ZSON [38, 46] tasks, which focus on high-level categories like "plant". Let $S$ denote the set of navigation scenes. Let $p_0$ describe the initial pose of an agent. A navigation episode $\tau \in \mathcal{T}$ is written as a tuple $\tau = (s, o, p_0)$ , $s \in S$ , $o \in \mathcal{O}$ . Each $\tau$ is a zero-shot task as tuples of this form are not seen during training. Starting at $p_0$ , an embodied agent's goal is to find $o$ . The agent receives observations and sensor readings $I_t$ (e.g., RGB-D images). At each timestep $t$ , the agent executes a navigation action $a \in \mathcal{A}$ . A special action $\text{STOP} \in \mathcal{A}$ terminates the episode. If the agent is within $c$ units of $o$ and $o$ meets a visibility criteria, the episode is successful. + +# 4. CLIP on Wheels (CoW) Baselines + +To address L-ZSON, we investigate a simple baseline approach, CoW, which adapts open-vocabulary models like CLIP to make them suitable for the task. A CoW takes as input an egocentric RGB-D image and an object goal specified in language. As a CoW moves, it updates a top-down map of the world created using RGB-D observations and pose estimates (Sec. 4.1). Each CoW gets an exploration policy and a zero-shot object localization module as seen in Fig. 2. To observe diverse views of the scene, a CoW explores using a policy (Sec. 4.2). As the CoW roams, it keeps track of its confidence about the target object's location using an object localization module (Sec. 4.3) and its top-down map. When a CoW's confidence exceeds a threshold, it plans to the location of the goal and issues the STOP action. We now describe the ingredients used to make the CoWs evaluated in our experiments (Sec. 6). + +# 4.1. Depth-based Mapping + +As a CoW moves, it constructs a top-down map using input depth, pose estimates, and known agent height. The map is initialized using known camera intrinsics and the first depth image. Since a CoW knows the intended consequences of its actions (e.g., MOVEFORWARD should result in a $0.25\mathrm{m}$ translation), each action is represented as a + +![](images/744a07bb4f5244059a6ff11e7c1a5757595ad2d06b97fe38d12f86a9ccb77226.jpg) + +![](images/b530d1ea053a38abe377f49c828f5ba002be7ddfec13acf8b51e25e253fbd124.jpg) + +pose delta transform to approximate a transition. To deal with noise associated with actuation or depth, a CoW maintains a map at $0.125\mathrm{m}$ resolution. To improve map accuracy, a CoW checks for failed actions by comparing successive depth frames for movements (see Appx. A for details). Using known agent height $(0.9\mathrm{m})$ , map cells are projected to the ground plane to maintain a top-down representation of the world, which suffices for most navigation applications. Cells close to the floor are considered free space (white points in Fig. 3 (a)), while other cells are considered occupied (blue points in Fig. 3 (a)). + +# 4.2. Exploration + +Exploration generates diverse egocentric views so a CoW is more likely to view the language-specified target object. We consider two exploration methods, frontier-based and learning-based. + +Frontier based exploration (FBE) [82]. Using the top-down map discussed in Sec. 4.1, a CoW can navigate using a simple exploration heuristic: move to the frontier between free and unknown space to discover new regions. Once the navigator reaches a frontier (visualized as purple points in Fig. 3 (a)), it moves greedily to the next closest frontier. Since the map is updated at every timestep, a noisy pose estimate can contribute to inaccuracies. For example, narrow passages may collapse in the map due to pose drift. To circumvent such problems, we reinitialize the map when no paths exist to any frontiers in the map. + +Learnable exploration. In addition to FBE, we consider learnable alternatives, which may explore more intelligently but incur substantial training costs. We investigate an architecture and reward structure similar to prior work in embodied AI (e.g., [25, 38, 75]). Specifically, we adopt a frozen CLIP backbone with a trainable GRU [16] and linear heads for the actor and critic networks. We train agents independently in HABITAT [67] and ROBOTHOR [18] simulation environments for 60M steps each, using DD-PPO [69, 77] in the AllenAct [76] framework. We employ a simple count-based reward [73]. All training scenes are disjoint from downstream navigation test scenes. For details on reward, hyperparameters, and training, see Appx. B. + +# 4.3. Object Localization + +Successful navigation depends on object localization: the ability to tell if and where an object is in an image. Regions of high object relevance, extracted from 2D images, are projected to the depth-based map (Fig. 3 (b)) where they serve as natural navigation targets. To determine if and when a target is in an image, we consider the following object localization modules, used in our experiments (Sec. 6). For more details see Appx. C. + +Adapting open-vocabulary classifiers. We experiment with three strategies to turn CLIP [61] models into object localizers. First, we utilize the CLIP text encoder to embed $k$ referring expressions, which specify regions where the target object may appear in the image. For example, "a plant in the top left of the image." We then match the language embeddings against a CLIP visual embedding for the current observation. We compute similarity between the image and text features to determine relevance scores over the regions. Second, we discretize the image into $k$ smaller patches and obtain CLIP patch embeddings. We then convolve each patch embedding with a CLIP text embedding for the target object. If the object is in a patch, the relevance score for that patch should be high. Third, we modify an interpretability method [13, 70] designed to extract object relevancy from vision transformers (ViTs) [24]. Using a target CLIP text embedding and gradient information accumulated through the CLIP vision encoder, we construct a relevance map over input pixels, which qualitatively segments the target when it is in view. + +Adapting open-vocabulary detectors and segmentors. In addition to CLIP-based methods, we consider two additional open-vocabulary models for object localization. First, the MDETR segmentation model [36], which extends the DETR detector [8] to take arbitrary text and images as input and output box detections. The base model is fine-tuned on PhraseCut [80], a dataset of paired masks and attribute descriptions, to support segmentation. Second, we consider the OWL-ViT detector [49], which uses a set prediction fine-tuning recipe to turn CLIP-like models into object detectors. We use this MDETR and OWL-ViT models to directly query images for targets. + +Post-processing. The aforementioned models give continuous valued predictions. However, we are interested in binary masks giving if and where objects are in images. Hence, we threshold predictions for each model (see Appx. C for details). We further investigate two strategies for using the masks downstream: (1) using the entire mask or (2) using the center pixel. The second strategy is reasonable because only part of an object needs to be detected for successful navigation. + +Target driven planning. Recall, CoWs have depth sensors. We back-project object relevance from 2D images into the + +![](images/028d123bae99555101968f4ed45ab80b17727dbff4968acaa1239af1a3cf8653.jpg) +Figure 4. Uncommon objects in PASTURE. + +depth-based map (Sec. 4.1). We keep only the max relevance for each map cell (Fig. 3 (b)). CoWs can then plan to high relevance areas in the map. See Appx. D for additional method visualization. + +Incorporating object priors. Since CoW does not train or fine-tune on navigation datasets, we investigate alternative approaches to inject object-level priors into the model. For each target object, we prompt GPT-3.5 [55] to generate rooms where the target objects are likely to be found. For example, GPT-3.5 states that apples are likely to be found in "kitchen" or "dining room" scenes. Following this prior, a GPT-3.5 enabled CoW first uses its object localization module to localize a kitchen or a dining room, and then looks for an apple. This straightforward extension, demonstrates how outside information can be incorporated into CoW. + +# 5. The PASTURE Benchmark + +To evaluate CoW baselines and future methods on L-ZSON, we introduce the PASTURE evaluation benchmark. PASTURE builds on ROBOTHER validation scenes, which have parallel environments in the real-world. We target ROBOTHER to facilitate future real-world benchmarking. PASTURE probes for seven capabilities explained in Sec. 5.1. We provide dataset statistics in Sec. 5.2. + +# 5.1. PASTURE Tasks + +PASTURE evaluates seven core L-ZSON capabilities. + +Uncommon objects. Traditional benchmarks (e.g., ROBOTHOR and HABITAT MP3D) evaluate agents on common object categories like TVs; however, given the rich diversity of objects in homes, we would like to understand navigation performance on uncommon objects. Hence we add 12 new objects to each room. We use names shown in Fig. 4 as instance labels, which are minimal descriptions to identify each object. Some identifiers refer to text in images (e.g., "whiteboard saying CVPR") or to appearance attributes (e.g., "wooden toy airplane"). Other objects are less common in North America, like "mate", which is a popular Argentinian drink. + +Appearance descriptions. To evaluate if baselines can take advantage of visual attributes, we introduce descriptions of the form: “ $\{size\}$ , $\{color\}$ , $\{material\}$ $\{object\}$ . For + +example: "small, red apple", "orange basketball", "small, black, metallic alarm clock". Objects are considered small if their 3D bounding box diagonal is below a threshold. We determine color and materials by inspection. + +Spatial descriptions. To test if agents can leverage spatial information in navigation, we introduce descriptions: “{object} on top of $\{x\}$ , near $\{y\}$ , $\{z\}$ , ...”. For example, “house plant on a dresser near a spray bottle”. To determine [on top of] relations, we use THOR metadata and to determine [nearness] we use a distance threshold between pairs of objects. We inspect all descriptions for correctness. + +Appearance descriptions with distractors. To probe if appearance attributes better facilitate finding objects in the presence of distractors, we reuse the appearance captions from before, but evaluate on an modified environment with two visually distinct instances of each ROBOTHOR object category. For example, for the task of finding a "red apple", we have both a red apple and a green apple in the room. A navigator must leverage appearance information—and not just class information—to successfully complete the task. Distractor objects are sufficiently far from the target objects so that finding a distractor cannot count as success. + +Spatial descriptions with distractors. This capability is similar to the one above; however, we evaluate with spatial descriptions in the presence of distractor objects. + +Hidden object descriptions. An ideal object navigator should find objects, even when they are hidden. Hence, we introduce descriptions: “{object} under/in {x}”. For example, “basketball in the dresser drawers” or “vase under the sofa”. We sample large objects (e.g., beds, sofas, dressers) in each scene to determine [under/in] relations. Additionally we remove visible instances of {object} from the room. + +Hidden object descriptions with distractors. We use the hidden object descriptions from before, but reintroduce visible instances of $\{object\}$ to serve as distractors. Consider finding a "mug under the bed". A distractor mug will also appear in the scene making the task more challenging. + +# 5.2. Dataset Creation and Statistics + +PASTURE contains three variations for each of the original 15 validation ROBOTHOR rooms: uncommon objects added, additional object instances added, and target objects removed. For each of the seven settings presented above, we evaluate over 12 object instances in 15 rooms with two initial agent starting locations. Hence PASTURE consists of 2,520 tasks, which is a similar order of magnitude to ROBOTHOR (1,800) and HABITAT MP3D (2,195) validation sets. For appearance attributes, $47\%$ of the objects are considered "small". Each object gets an average of 1.2 color descriptors out of 22 possible choices, and 0.6 material descriptors out of 5 possible choices. Similarly, for spatial attributes, each object gets one object it is on top of or in (e.g., + +“vase in a shelving unit”) and an average of 1.9 objects it is near. For a sample of appearance and spatial attributes see Fig. 1. For more dataset details and statistics see Appx. E. + +# 6. Experiments + +We first present our experimental setup, including the datasets, metrics, embodiment, and baselines considered in our study (Sec. 6.1). Then we present results on PASTURE, thereby elucidating the strengths and weakness of CoW baselines for L-ZSON (Sec. 6.2). Finally, we compare to prior ZSON art in ROBOTHER and HABITAT (MP3D) environments (Sec. 6.3). + +# 6.1. Experimental setup + +**Environments.** We consider PASTURE (Sec. 5), ROBOTHER [18], and HABITAT (MP3D) [67] validation sets as our test sets. We utilize validation sets for testing because official test set ground-truth is not publicly available. Domains are setup with noise that is faithful to their original challenge settings. For ROBOTHER—and by extension PASTURE—this means actuation noise but no depth noise. For HABITAT this means considerable depth noise and reconstruction artifacts, but no actuation noise. + +Navigation Metrics. We adopt standard object navigation metrics to measure performance: + +- SUCCESS (SR): the fraction of episodes where the agent executes STOP within $1.0\mathrm{m}$ of the target object. +- Success weighted by inverse path length (SPL): Success weighted by the oracle shortest path length and normalized by the actual path length [5]. This metric points to the success efficiency of the agent. + +In ROBOTHOR and PASTURE, the target must additionally be visible for the episode to be a success, which this is not the case in HABITAT—as specified in their 2021 challenge. + +Embodiment. The agent is a LoCoBot [29]. All agents have discrete actions: {MOVEFORWARD, ROTATERIGHT, ROTATELEFT, STOP}. The move action advances the agent by $0.25\mathrm{m}$ , while rotation actions pivot the camera by $30^{\circ}$ . + +CoW Baselines. For exploration we consider policies presented in Sec. 4.2: FBE heuristic exploration, learnable exploration optimized on HABITAT (MP3D) train scenes, and learned exploration optimized on ROBOTHOR train scenes. Learned exploration requires training in simulation—which is counter to our zero-shot goals; nonetheless, we ablate these explorers to contextualize their performance within the CoW framework. FBE is the default CoW exploration strategy. + +For object localization, we consider: + +- CLIP with $k = 9$ referring expressions (CLIP-Ref.) +- CLIP with $k = 9$ patches (CLIP-Patch) +- CLIP with gradient relevance (CLIP-Grad.) + +
IDCoW breedsPASTUREROBOTHOR
Uncom.Appear.SpaceAppear. distract SRSpace distract SRHid.Hid. distract SRAvg.
LocalizerArch.SRSRSRSRSRSRSRSPLSRSPLSR
▲CLIP-Ref.B/323.60.61.70.61.72.22.50.91.81.01.8
■CLIP-Ref.B/161.42.82.83.13.31.71.91.72.42.12.7
▲CLIP-PatchB/3218.113.313.38.610.817.517.89.014.210.620.3
■CLIP-PatchB/1610.611.47.810.88.116.415.67.711.59.715.7
▲CLIP-Grad.B/3216.111.911.79.710.314.416.19.212.99.715.2
■CLIP-Grad.B/168.110.88.68.66.711.111.46.79.38.611.6
◇MDETRB33.17.25.06.94.78.18.95.46.38.49.9
▲OWLB/3232.826.419.419.416.119.214.412.621.116.926.7
■OWLB/1631.926.918.919.414.718.115.812.620.817.227.5
ProcTHOR fine-tune (supervised) [19]n/an/an/an/an/an/an/an/an/a27.466.4
+ +Table 1. Benchmarking CoWs on PASTURE for L-ZSON. On PASTURE we identify several key takeaways. (1) Average success on PASTURE is lower than on ROBOTHOR; however, CoWs are surprisingly good at finding uncommon objects (Uncom.), often finding them at higher rates than more common ROBOTHOR objects. (2) Comparing square (■) vs. triangle (▲) IDs, we see that architectures (Arch.) using more compute (i.e., ViT-B/16) often perform comparably or worse than their competitors (i.e., ViT-B/32). This is especially true for CLIP [61] models (indicated in pink, orange, and purple). (3) Blue OWL-ViT [49] models perform best. (4) PASTURE tasks with distractor objects (distract) hurt performance and natural language specification is not sufficient to mitigate against the added difficulties in these tasks. (5) A supervised baseline shown in gray significantly outperforms CoWs on ROBOTHOR; however, it is unable to support PASTURE tasks out-of-the-box. + +- MDETR segmentation model (MDETR) +- OWL-ViT detector (OWL) + +Descriptions of these models are in Sec. 4.3 and additional details are in Appx. C. All models are open-vocabulary. No models are fine-tuned on navigation, and hence we consider their inference zero-shot on our tasks.2 We also consider various backbone architectures: + +- A vision transformer [24], ViT-B/32 ( $\triangle$ B/32) +- ViT-B/16 (■ B/16), which uses a smaller patch size of 16x16 and hence more compute. +- EfficientNet B3 ( $\spadesuit$ B3), which is convolutional and similar in compute requirements to a ViT/B32. + +For every model we evaluate with post-processing as the default setting, where only the center pixel of detections is registered in the top-down CoW map. Recall, this is a sensible strategy as only some part of the object needs to be found for an episode to be successful. We find that this decision improves performance on our best performing models from 0.1 to 6.0 percentage points on ROBOTHOR SUCCESS. For a full comparison between models with and without post-processing see Appx. F. For details on hyperparameters, learned agents, object localization threshold tuning, and CLIP prompt-tuning, see Appx. C, G. + +End-to-end learnable baselines. We also compare against methods that are trained in simulation for millions of steps: + +- EmbCLIP-ZSON [38] trains on eight ROBOTHR categories, using CLIP language embeddings to specify the goal objects. At test time, the model is evaluated on four held-out object categories, which are also specified CLIP language embeddings for the target category names. +- SemanticNav-ZSON [46] trains models separately, one for each dataset, for image goal navigation. Image goals are specified with CLIP visual embeddings. At test time, image goals are swapped for CLIP language embeddings for the object goals. We compare to the MP3D model. + +Both EmbCLIP-ZSON and SemanticNav-ZSON leverage multi-modal CLIP visual and language embeddings in learnable frameworks that require simulation training. + +# 6.2. CoWs on PASTURE + +Tab. 1 shows our main results of CoWs evaluated on PASTURE. For category-level results see Appx. H. We now discuss several salient questions. + +How well can CoWs find common objects vs. uncommon objects? Comparing ROBOTHOR and uncommon (Uncom.) PASTURE success rate (SR) in Tab. 1—first and last columns—we notice that CoWs often find uncommon objects at higher rates than common ROBOTHOR objects (e.g., by $\sim 6$ percentage points SUCCESS for the OWL ViT-B/32 CoW ( $\triangle$ )). We hypothesize that though uncommon objects are less prevalent in daily life, they are still represented in open-vocabulary datasets and hence recognizable for the object localization modules. We further ex + +![](images/c08efe085810ec7489087ad49b99ca9d5562e6dc03738f53c5d2af944044b086.jpg) +(a) Attribute object navigation + +![](images/e5c82a109c048ad9062ef6ee47c8ab41ed2c0d4d029e09f19abae3c746c5c1fe.jpg) +(b) Attribute object navigation with distractors +Figure 5. PASTURE object navigation with descriptions. In general object navigation with descriptions is more challenging than the ROBOTOR object navigation task, as indicated by trend lines lying below the $y = x$ line. (a) Appearance descriptions are more helpful than spatial descriptions. (b) Performance further drops when distractor objects are introduced to the environment. However, CoWs are still able to make better use of appearance description than spatial descriptions. (c) Models in the lower success regime (<15% ROBOTOR SUCCESS) perform comparably on finding hidden objects. However, this trend plateaus for higher success models. (d) Trends are similar when distractor objects are introduced for hidden object navigation. + +![](images/e8ea830b0c8da8f02a872487755d76c6447c60a7110e3b42e1f8f0862616823c.jpg) +(c) Hidden object navigation + +![](images/cb2d52d7231379da66db7359b05d1718824a1723620903f966b7f16b8b2de868.jpg) +(d) Hidden object navigation with distractors + +plore this hypothesis in Appx. E by visualizing CLIP retrieval results on LAION-5B [68] for the uncommon object categories. The relatively high performance on uncommon objects speaks to the flexibility of CoW baselines and their ability to inherit desirable properties from the open-vocabulary models that they are constructed from. + +Can CoWs utilize appearance and spatial descriptions? Looking at Fig. 5 (a) we see that neither appearance nor spatial descriptions improve CoW performance compared to their ROBOTHOR baseline performance (i.e., most points lie under the $y = x$ line). However, CoW is able to take better advantage of appearance descriptions than spatial descriptions. These results motivate future investigation on open-vocabulary object localization with a greater focus on textual object attributes. + +Can CoWs find visible objects in the presence of distractors? In Fig. 5 (b) we see that CoWs experience performance degradation when compared to Fig. 5 (a). We conclude that appearance and spatial attributes added as language input are not sufficient to deal with the complexity of distractors given current open-vocabulary models. + +Can CoWs find hidden objects? Looking at Fig. 5 (c) we notice that models in the lower success regime (less than $15\%$ SUCCESS on ROBOTHOR) are able to find hidden objects at about the same rate as ROBOTHOR objects (i.e., they lie on the $y = x$ line). OWL models in the higher success regime $(>15\%)$ do not continue this trend; however, they do achieve higher absolute accuracy as seen in Tab. 1. Dealing with occlusion is a longstanding problem in computer vision, and these results provide a foundation upon which future hidden object navigation work can improve. + +Can CoWs find hidden objects in the presence of distractors? Comparing Figs. 5 (c) and (d), we notice similar trends lines, with the best models performing worse with distractors. This suggests that distractors are less of a con + +
IDLoc.CoW breedsPASTURE (Avg.)ROBOTHOR
Arch.Exp. StrategySPLSRSPLSR
▲OWLB/32ROBOTHOR learn.10.217.313.120.9
▲OWLB/32HABITAT learn.8.619.49.820.4
▲OWLB/32FBE12.621.116.926.7
+ +Table 2. Exploration ablation. For a fixed object localizer (OWL-ViT B/32 with post processing), we ablate over different choices of exploration policy: the FBE heuristic, agents trained in ROBOTHOR, and HABITAT (MP3D). We find that FBE outperforms learnable alternatives on both PASTURE and ROBOTHOR. HABITAT learnable model perform worst, but are not trained on any PASTURE or ROBOTHOR data. + +
IDCoW breedsObj. PriorPASTURE Uncom.ROBOTHOR
Loc.Arch.SPLSRSPLSR
▲OWLB/32None20.532.816.826.7
▲OWLB/32GPT-3.522.236.917.027.5
+ +Table 3. CoW with GPT-3.5 priors. Leveraging GPT-3.5, we generate priors for objects (e.g., apples are likely to be in dining room scenes). Instead of directly searching for the target object, CoW first searches for the scenes, which boosts performance. + +cern in the case of hidden objects than for visible object targets. In light of the fact that detection methods generally work better on larger objects, we hypothesize this effect is because distractor objects are smaller (e.g., apples, vases, basketballs) than objects used to conceal target categories (e.g., beds, sofas, etc.). + +What exploration method performs best? We ablate the decision to use FBE for most experiments by fixing an object localizer (OWL, B/32 $(\triangle))$ and comparing against ROBOTHR learnable exploration and HABITAT learnable exploration in Tab. 2. We notice that FBE performs best in all cases; however, learnable exploration still performs well suggesting that these models do learn useful strategies for the downstream tasks. + +![](images/5beab85ea8a0266fdfe3a1171921383455df55bd1b781209469696e873de385b.jpg) +Figure 6. Failure analysis for OWL, B/32 (▲). Exploration and object localization errors occur at similar ratios, with increased localization failures in the presence of distractors. + +Can CoW incorporate object priors? Examining Tab. 3, we see that incorporating GPT-3.5 object-level priors improves performance on both PASTURE uncommon objects and ROBOTOR. These initial results suggest positive trends for incorporating outside knowledge into CoW. Future work may consider more sophisticated methods for injecting priors to steer navigation. + +How do CoWs fail? We identify three high-level failure modes. (1) Exploration fail: the target is never seen. (2) Object localization fail: the target is seen but the localizer never fires. (3) Planning fail: the target is seen and the localizer fires, but planning fails due to inaccuracy in the map representation (Sec. 4.2). Looking at Fig. 6, we notice a large fraction of failures are due to exploration and object localization. This suggests CoWs may continue to improve as research in these fields progress. In Fig. 6 we also see that in cases where distractors are present a higher fraction of object localization failures occur, further suggesting that open-vocabulary models currently struggle to make full use of attribute prompts. See Appx. I for more failure analysis. + +# 6.3. Comparison to Prior Art + +We primarily evaluate CoWs in general L-ZSON settings; however, we further evaluate CoWs on ZSON benchmarks to establish them as a strong baseline for these tasks. Recall, ZSON can be seen as a case of L-ZSON where only object goals are specified (no attributes). + +In Tab. 4, we see there exists a CoW that outperforms the end-to-end baselines in all cases except SUCCESS on HABITAT (MP3D). For instance, the CLIP-Grad., B/32 ( $\triangleleft$ ) matches the SemanticNav-ZSON model on HABITAT (MP3D) SPL: 4.9 for CoW v.s. 4.8 for the competitor, while improving over EmbCLIP-ZSON ROBOTHOR by 15.6 percentage points. To contextualize this result, CoWs train for $O$ navigation steps, while SemanticNav-ZSON and EmbCLIP-ZSON train in the target evaluation simulators for 500M and 60M steps respectively. + +
IDCoW breedsHABITAT (MP3D)ROBOTHON (subset)ROBOTHON (full)Nav. training steps
Loc.Arch.SPLSRSPLSRSPLSR
▲CLIP-Grad.B/324.99.215.023.79.715.20
▲OWLB/323.77.420.832.516.926.70
EmbCLIP-ZSON [38]---8.1-14.0*60M
SemanticNav-ZSON [46]4.815.3----500M
+ +Table 4. Comparison to prior art on existing ZSON benchmarks. CoWs are able to match or out-compete existing methods that leverage millions of steps of navigation training in the evaluation simulator. *indicates a result from prior work that includes, non-zero-shot evaluation. Specifically, only 1/4 of the evaluations are zero-shot on ROBOTOR (subset) and the remaining 3/4 on categories seen during training. + +The superior performance of SemanticNav-ZSON in terms of MP3D SUCCESS indicates that there can be benefits to in-domain learning. Future work may consider unifying the benefits of CoW-like models and fine-tuned models. + +# 7. Limitations and Conclusion + +Limitations. While our evaluation of CoWs on HABITAT, ROBOTHER, and PASTURE is a step towards assessing their performance in different domains, ultimately, real-world performance matters most. Hence, the biggest limitation of our work is the lack of large-scale, real-world benchmarking—which is also missing in much of the related literature. Additionally, CoW inherits the metalimitations of the object localization and exploration methods considered. For example, object localizers require tuning a confidence threshold to balance precision and recall. Finally, we do not consider different agent embodiment or continuous action spaces. This is a pertinent investigation given recent findings of Pratt et al. [59] that agent morphology can be a big determinant of downstream performance. + +Conclusion. This paper introduces the PASTURE benchmark for language-driven zero-shot object navigation and several CLIP on Wheels baselines, translating the successes of existing zero-shot models to an embodied task. We view CoW as an instance of using open-vocabulary models, with text-based interfaces, to tackle robotics tasks in more flexible settings. We hope that the baselines and the proposed benchmark will spur the field to explore broader and more powerful forms of zero-shot embodied AI. + +Acknowledgement. We would like to thank Jessie Chapman, Cheng Chi, Huy Ha, Zeyi Liu, Sachit Menon, and Sarah Pratt for valuable feedback. This work was supported in part by NSF CMMI-2037101, NSF IIS-2132519, and an Amazon Research Award. SYG is supported by a NSF Graduate Research Fellowship. The views and conclusions contained herein are those of the authors and should not be interpreted as necessarily representing the official policies, either expressed or implied, of the sponsors. + +# References + +[1] Ziad Al-Halah, Santhosh K. Ramakrishnan, and Kristen Grauman. Zero experience required: Plug & play modular transfer learning for semantic visual navigation. arXiv, 2022. 2 +[2] Peter Anderson, Qi Wu, Damien Teney, Jake Bruce, Mark Johnson, Niko Sunderhauf, Ian Reid, Stephen Gould, and Anton Van Den Hengel. Vision-and-language navigation: Interpreting visually-grounded navigation instructions in real environments. CVPR, 2018. 2 +[3] Donghyeon Baek, Youngmin Oh, and Bumsub Ham. Exploiting a joint embedding space for generalized zero-shot semantic segmentation. ICCV, 2021. 1 +[4] Ankan Bansal, Karan Sikka, Gaurav Sharma, Rama Chellappa, and Ajay Divakaran. Zero-shot object detection. ECCV, 2018. 1 +[5] Dhruv Batra, Aaron Gokaslan, Aniruddha Kembhavi, Oleksandr Maksymets, Roozbeh Mottaghi, Manolis Savva, Alexander Toshev, and Erik Wijmans. Objectnav revisited: On evaluation of embodied agents navigating to objects. arXiv, 2020. 1, 3, 5 +[6] Maxime Bucher, Tuan-Hung VU, Matthieu Cord, and Patrick Pérez. Zero-shot semantic segmentation. NeurIPS, 2019. 1 +[7] Yuri Burda, Harrison Edwards, Amos J. Storkey, and Oleg Klimov. Exploration by random network distillation. *ICLR*, 2018. 2 +[8] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. ECCV, 2020. 4, 14 +[9] Matthew Chang, Arjun Gupta, and Saurabh Gupta. Semantic visual navigation by watching youtube videos. NeurIPS, 2020. 2 +[10] Devendra Singh Chaplot, Dhiraj Gandhi, Abhinav Gupta, and Ruslan Salakhutdinov. Object goal navigation using goal-oriented semantic exploration. NeurIPS, 2020. 2 +[11] Devendra Singh Chaplot, Dhiraj Gandhi, Saurabh Gupta, Abhinav Kumar Gupta, and Ruslan Salakhutdinov. Learning to explore using active neural slam. ICLR, 2020. 2 +[12] Prithvijit Chattopadhyay, Judy Hoffman, Roozbeh Mottaghi, and Aniruddha Kembhavi. Robustnav: Towards benchmarking robustness in embodied navigation. ICCV, 2021. 2 +[13] Hila Chefer, Shir Gur, and Lior Wolf. Transformer interpretability beyond attention visualization. CVPR, 2021. 4, 13, 14 +[14] Tao Chen, Saurabh Gupta, and Abhinav Gupta. Learning exploration policies for navigation. *ICLR*, 2019. 2 +[15] Jiaxin Cheng, Soumyaroop Nandi, Prem Natarajan, and Wael Abd-Almageed. Sign: Spatial-information incorporated generative network for generalized zero-shot semantic segmentation. ICCV, 2021. 1 +[16] Kyunghyun Cho, Bart van Merrienboer, Dzmitry Bahdanau, and Yoshua Bengio. On the properties of neural machine translation: Encoder-decoder approaches. In SSST@EMNLP, 2014. 3 +[17] Andrew J Davison and David W Murray. Mobile robot localisation using active vision. ECCV, 1998. 2 + +[18] Matt Deitke, Winson Han, Alvaro Herrasti, Aniruddha Kembhavi, Eric Kolve, Roozbeh Mottaghi, Jordi Salvador, Dustin Schwenk, Eli VanderBilt, Matthew Wallingford, Luca Weihs, Mark Yatskar, and Ali Farhadi. Robothor: An open simulation-to-real embodied ai platform. CVPR, 2020. 1, 3, 5 +[19] Matt Deitke, Eli VanderBilt, Alvaro Herrasti, Luca Weihs, Jordi Salvador, Kiana Ehsani, Winson Han, Eric Kolve, Ali Farhadi, Aniruddha Kembhavi, and Roozbeh Mottaghi. Proctor: Large-scale embodied ai using procedural generation. NeurIPS, 2022. 2, 6 +[20] Frank Dellaert, Dieter Fox, Wolfram Burgard, and Sebastian Thrun. Monte carlo localization for mobile robots. ICRA, 1999. 2 +[21] Berkan Demirel, Ramazan Gokberk Cinbis, and Nazli Ikizler-Cinbis. Zero-shot object detection by hybrid region embedding. BMVC, 2018. 1 +[22] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, K. Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. CVPR, 2009. 1 +[23] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv, 2020. 13 +[24] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. ICLR, 2020. 4, 6 +[25] Samir Yitzhak Gadre, Kiana Ehsani, Shuran Song, and Roozbeh Mottaghi. Continuous scene representations for embodied ai. CVPR, 2022. 2, 3 +[26] Daniel Gordon, Abhishek Kadian, Devi Parikh, Judy Hoffman, and Dhruv Batra. Splitnet: Sim2sim and task2task transfer for embodied visual navigation. CVPR, 2019. 2 +[27] Xiuye Gu, Tsung-Yi Lin, Weicheng Kuo, and Yin Cui. Zeroshot detection via vision and language knowledge distillation. arXiv, 2021. 1 +[28] Agrim Gupta, Piotr Dólar, and Ross B. Girshick. Lvis: A dataset for large vocabulary instance segmentation. CVPR, 2019. 1 +[29] Abhinav Kumar Gupta, Adithyavairavan Murali, Dhiraj Gandhi, and Terrel Pinto. Robot learning in homes: Improving generalization and reducing dataset bias. NeurIPS, 2018. 5 +[30] Saurabh Gupta, James Davidson, Sergey Levine, Rahul Sukthankar, and Jitendra Malik. Cognitive mapping and planning for visual navigation. CVPR, pages 2616-2625, 2017. 2 +[31] Meera Hahn, Devendra Singh Chaplot, and Shubham Tulsiani. No rl, no simulation: Learning to navigate without navigating. NeurIPS, 2021. 2 +[32] Peter Henry, Michael Krainin, Evan Herbst, Xiaofeng Ren, and Dieter Fox. RGB-d mapping: Using depth cameras for dense 3d modeling of indoor environments. Experimental robotics, 2014. 2 + +[33] Ping Hu, Stan Sclaroff, and Kate Saenko. Uncertainty-aware learning for zero-shot semantic segmentation. NeurIPS, 2020. 1 +[34] Chenguang Huang, Oier Mees, Andy Zeng, and Wolfram Burgard. Visual language maps for robot navigation. arXiv preprint arXiv:2210.05714, 2022. 2 +[35] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc V Le, Yunhsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. ICML, 2021. 1, 13 +[36] Aishwarya Kamath, Mannat Singh, Yann LeCun, Ishan Misra, Gabriel Synnaeve, and Nicolas Carion. Mdetr - modulated detection for end-to-end multi-modal understanding. ICCV, 2021. 1, 4, 13, 14 +[37] Naoki Kato, Toshihiko Yamasaki, and Kiyoharu Aizawa. Zero-shot semantic segmentation via variational mapping. ICCV-W, 2019. 1 +[38] Apoorv Khandelwal, Luca Weihs, Roozbeh Mottaghi, and Aniruddha Kembhavi. Simple but effective: Clip embeddings for embodied ai. arXiv, 2021. 1, 2, 3, 6, 8 +[39] Jacob Krantz, Erik Wijmans, Arjun Majumdar, Dhruv Batra, and Stefan Lee. Beyond the nav-graph: Vision-and-language navigation in continuous environments. In European Conference on Computer Vision, pages 104-120. Springer, 2020. 2 +[40] Alexander Ku, Peter Anderson, Roma Patel, Eugene Ie, and Jason Baldridge. Room-across-room: Multilingual vision-and-language navigation with dense spatiotemporal grounding. arXiv preprint arXiv:2010.07954, 2020. 2 +[41] Benjamin Kuipers and Yung-Tai Byun. A robot exploration and mapping strategy based on a semantic hierarchy of spatial representations. Robotics and autonomous systems, 1991. 2 +[42] Xinghang Li, Di Guo, Huaping Liu, and Fuchun Sun. Revece: Remote embodied visual referring expression in continuous environment. RA-L, 2022. 2 +[43] Zhihui Li, Lina Yao, Xiaqin Zhang, Xianzhi Wang, Salil S. Kanhere, and Huaxiang Zhang. Zero-shot object detection with textual descriptions. AAAI, 2019. 1 +[44] Yiqing Liang, Boyuan Chen, and Shuran Song. Sscnav: Confidence-aware semantic scene completion for visual semantic navigation. ICRA, 2021. 2 +[45] Tsung-Yi Lin, Michael Maire, Serge Belongie, Lubomir Bourdev, Ross Girshick, James Hays, Pietro Perona, Deva Ramanan, C. Lawrence Zitnick, and Piotr Dólar. Microsoft coco: Common objects in context. ECCV, 2014. 1 +[46] Arjun Majumdar, Gunjan Aggarwal, Bhavika Devnani, Judy Hoffman, and Dhruv Batra. Zson: Zero-shot object-goal navigation using multimodal goal embeddings. arXiv preprint arXiv:2206.12403, 2022. 1, 2, 3, 6, 8 +[47] Qiaomei Mao, Chong Wang, Sheng Yu, Ye Zheng, and Yuqi Li. Zero-shot object detection with attributes-based category similarity. IEEE Transactions on Circuits and Systems II: Express Briefs, 2020. 1 +[48] Lina Mezghani, Sainbayar Sukhbaatar, Thibaut Lavril, Oleksandr Maksymets, Dhruv Batra, Piotr Bojanowski, and Alahari Karteek. Memory-augmented reinforcement learning for image-goal navigation. arXiv, 2021. 2 + +[49] Matthias Minderer, Alexey A. Gritsenko, Austin Stone, Maxim Neumann, Dirk Weissenborn, Alexey Dosovitskiy, Aravindh Mahendran, Anurag Arnab, Mostafa Dehghani, Zhuoran Shen, Xiao Wang, Xiaohua Zhai, Thomas Kipf, and Neil Houlsby. Simple open-vocabulary object detection with vision transformers. ECCV, 2022. 1, 4, 6, 13, 14 +[50] Grégoire Montavon, Sebastian Lapuschkin, Alexander Binder, Wojciech Samek, and Klaus-Robert Müller. Explaining nonlinear classification decisions with deep taylor decomposition. Pattern recognition, 65:211-222, 2017. 13 +[51] Hans Moravec and Alberto Elfes. High resolution maps from wide angle sonar. ICRA, 1985. 2 +[52] Richard A Newcombe, Shahram Izadi, Otmar Hilliges, David Molyneaux, David Kim, Andrew J Davison, Pushmeet Kohi, Jamie Shotton, Steve Hodges, and Andrew Fitzgibbon. Kinectfusion: Real-time dense surface mapping and tracking. ISMAR, 2011. 2 +[53] Lachlan Nicholson, Michael Milford, and Niko Sunderhauf. Quadricslam: Dual quadrics from object detections as landmarks in object-oriented slam. RA-L, 2019. 2 +[54] Clark F Olson and Larry H Matthies. Maximum likelihood rover localization by matching range maps. ICRA, 1998. 2 +[55] Long Ouyang, Jeff Wu, Xu Jiang, Diogo Almeida, Carroll L Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, et al. Training language models to follow instructions with human feedback. NeurIPS, 2022. 4, 15 +[56] Simone Parisi, Victoria Dean, Deepak Pathak, and Abhinav Kumar Gupta. Interesting object, curious agent: Learning task-agnostic exploration. NeurIPS, 2021. 2 +[57] Deepak Pathak, Pulkit Agrawal, Alexei A. Efros, and Trevor Darrell. Curiosity-driven exploration by self-supervised prediction. ICML, 2017. 2 +[58] Hieu Pham, Zihang Dai, Golnaz Ghiasi, Hanxiao Liu, Adams Wei Yu, Minh-Thang Luong, Mingxing Tan, and Quoc V. Le. Combined scaling for zero-shot transfer learning. arXiv, 2021. 1, 13 +[59] Sarah Pratt, Luca Weihs, and Ali Farhadi. The introspective agent: Interdependence of strategy, physiology, and sensing for embodied agents. arXiv, 2022. 8 +[60] Yuankai Qi, Qi Wu, Peter Anderson, Xin Wang, William Yang Wang, Chunhua Shen, and Anton van den Hengel. Reverie: Remote embodied visual referring expression in real indoor environments. CVPR, 2020. 2 +[61] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. ICML, 2021. 1, 2, 4, 6, 13 +[62] Shafin Rahman, Salman Hameed Khan, and Fatih Murat Porkikli. Zero-shot object detection: Learning to simultaneously recognize and localize novel concepts. ACCV, 2018. 1 +[63] Roberta Raileanu and Tim Rocktäschel. Ride: Rewarding impact-driven exploration for procedurally-generated environments. *ICLR*, 2020. 2 + +[64] Santhosh K. Ramakrishnan, Devendra Singh Chaplot, Ziad Al-Halah, Jitendra Malik, and Kristen Grauman. Poni: Potential functions for objectgoal navigation with interaction-free learning. CVPR, 2022. 2 +[65] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, Alexander C. Berg, and Li Fei-Fei. Imagenet large scale visual recognition challenge. IJCV, 2015. 1 +[66] Manolis Savva, Angel X Chang, Alexey Dosovitskiy, Thomas Funkhouser, and Vladlen Koltun. Minos: Multimodal indoor simulator for navigation in complex environments. arXiv, 2017. 2 +[67] Manolis Savva, Abhishek Kadian, Oleksandr Maksymets, Yili Zhao, Erik Wijmans, Bhavana Jain, Julian Straub, Jia Liu, Vladlen Koltun, Jitendra Malik, Devi Parikh, and Dhruv Batra. Habitat: A Platform for Embodied AI Research. ICCV, 2019. 1, 3, 5 +[68] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. NeurIPS, 2022. 7, 16, 18 +[69] John Schulman, Filip Wolski, Prafulla Dhariwal, Alec Radford, and Oleg Klimov. Proximal policy optimization algorithms. arXiv, 2017. 3 +[70] Ramprasaath R. Selvaraju, Abhishek Das, Ramakrishna Vedantam, Michael Cogswell, Devi Parikh, and Dhruv Batra. Grad-cam: Visual explanations from deep networks via gradient-based localization. IJCV, 2019. 4, 13, 14 +[71] Mohit Shridhar, Jesse Thomason, Daniel Gordon, Yonatan Bisk, Winson Han, Roozbeh Mottaghi, Luke Zettlemoyer, and Dieter Fox. Alfred: A benchmark for interpreting grounded instructions for everyday tasks. CVPR, 2020. 2 +[72] Shuran Song, Linguang Zhang, and Jianxiong Xiao. Robot in a room: Toward perfect object recognition in closed environments. arXiv, 2015. 2 +[73] Alexander L. Strehl and Michael L. Littman. An analysis of model-based interval estimation for markov decision processes. J. Comput. Syst. Sci., 2008, 2, 3 +[74] Saim Wani, Shivansh Patel, Unnat Jain, Angel X. Chang, and Manolis Savva. Mution: Benchmarking semantic map memory using multi-object navigation. NeurIPS, 2020. 2 +[75] Luca Weihs, Matt Deitke, Aniruddha Kembhavi, and Roozbeh Mottaghi. Visual room rearrangement. CVPR, 2021. 2, 3 +[76] Luca Weihs, Jordi Salvador, Klemen Kotar, Unnat Jain, KuoHao Zeng, Roozbeh Mottaghi, and Aniruddha Kembhavi. Allenact: A framework for embodied ai research. arXiv, 2020. 3, 12 +[77] Erik Wijmans, Abhishek Kadian, Ari S. Morcos, Stefan Lee, Irfan Essa, D. Parikh, Manolis Savva, and Dhruv Batra. Dppo: Learning near-perfect pointgoal navigators from 2.5 billion frames. *ICLR*, 2019. 2, 3 +[78] Brian H Wilcox. Robotic vehicles for planetary exploration. Applied Intelligence, 1992. 2 + +[79] Mitchell Wortsman, Kiana Ehsani, Mohammad Rastegari, Ali Farhadi, and Roozbeh Mottaghi. Learning to learn how to learn: Self-adaptive visual navigation using meta-learning. CVPR, 2019. 1, 2 +[80] Chenyun Wu, Zhe Lin, Scott D. Cohen, Trung Bui, and Subhransu Maji. Phrasecut: Language-based image segmentation in the wild. CVPR, 2020. 4, 14 +[81] Fei Xia, Amir R Zamir, Zhiyang He, Alexander Sax, Jitendra Malik, and Silvio Savarese. Gibson env: Real-world perception for embodied agents. CVPR, 2018. 2 +[82] Brian Yamauchi. A frontier-based approach for autonomous exploration. CIRA, 1997. 3 +[83] Claudia Yan, Dipendra Misra, Andrew Bennnett, Aaron Walsman, Yonatan Bisk, and Yoav Artzi. Chalet: Cornell house agent learning environment. arXiv, 2018. 2 +[84] Wei Yang, Xiaolong Wang, Ali Farhadi, Abhinav Gupta, and Roozbeh Mottaghi. Visual semantic navigation using scene priors. arXiv, 2018. 2 +[85] Xiaohua Zhai, Xiao Wang, Basil Mustafa, Andreas Steiner, Daniel Keysers, Alexander Kolesnikov, and Lucas Beyer. Lit: Zero-shot transfer with locked-image text tuning. CVPR, 2022. 14 +[86] Chong Zhou, Chen Change Loy, and Bo Dai. Denseclip: Extract free dense labels from clip. arXiv, 2021. 1 +[87] Fengda Zhu, Xiwen Liang, Yi Zhu, Xiaojun Chang, and Xiaodan Liang. Soon: Scenario oriented object navigation with graph-based exploration. CVPR, 2021. 2 +[88] Pengkai Zhu, Hanxiao Wang, and Venkatesh Saligrama. Zero shot detection. IEEE Transactions on Circuits and Systems for Video Technology, 2020. 1 +[89] Yuke Zhu, Roozbeh Mottaghi, Eric Kolve, Joseph J Lim, Abhinav Gupta, Li Fei-Fei, and Ali Farhadi. Target-driven visual navigation in indoor scenes using deep reinforcement learning. ICRA, 2017. 1, 2 \ No newline at end of file diff --git a/2023/CoWs on Pasture_ Baselines and Benchmarks for Language-Driven Zero-Shot Object Navigation/images.zip b/2023/CoWs on Pasture_ Baselines and Benchmarks for Language-Driven Zero-Shot Object Navigation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..a043ab5939203a8cb3922eefc2501ddbb8be203a --- /dev/null +++ b/2023/CoWs on Pasture_ Baselines and Benchmarks for Language-Driven Zero-Shot Object Navigation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2f768b212cd59b09ae311880fb2fb4aa9618cf96b17b3b7c06cc1405a91528fb +size 393055 diff --git a/2023/CoWs on Pasture_ Baselines and Benchmarks for Language-Driven Zero-Shot Object Navigation/layout.json b/2023/CoWs on Pasture_ Baselines and Benchmarks for Language-Driven Zero-Shot Object Navigation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..4491af8848dd1b23c23764ec9b641691ee9724f2 --- /dev/null +++ b/2023/CoWs on Pasture_ Baselines and Benchmarks for Language-Driven Zero-Shot Object Navigation/layout.json @@ -0,0 +1,9481 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 140, + 103, + 455, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 140, + 103, + 455, + 140 + ], + "spans": [ + { + "bbox": [ + 140, + 103, + 455, + 140 + ], + "type": "text", + "content": "CoWs on PASTURE: Baselines and Benchmarks for Language-Driven Zero-Shot Object Navigation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 160, + 536, + 177 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 160, + 536, + 177 + ], + "spans": [ + { + "bbox": [ + 54, + 160, + 536, + 177 + ], + "type": "text", + "content": "Samir Yitzhak Gadre" + }, + { + "bbox": [ + 54, + 160, + 536, + 177 + ], + "type": "inline_equation", + "content": "^{\\diamond}" + }, + { + "bbox": [ + 54, + 160, + 536, + 177 + ], + "type": "text", + "content": " Mitchell Wortsman" + }, + { + "bbox": [ + 54, + 160, + 536, + 177 + ], + "type": "inline_equation", + "content": "^{\\dagger}" + }, + { + "bbox": [ + 54, + 160, + 536, + 177 + ], + "type": "text", + "content": " Gabriel Ilharco" + }, + { + "bbox": [ + 54, + 160, + 536, + 177 + ], + "type": "inline_equation", + "content": "^{\\dagger}" + }, + { + "bbox": [ + 54, + 160, + 536, + 177 + ], + "type": "text", + "content": " Ludwig Schmidt" + }, + { + "bbox": [ + 54, + 160, + 536, + 177 + ], + "type": "inline_equation", + "content": "^{\\dagger}" + }, + { + "bbox": [ + 54, + 160, + 536, + 177 + ], + "type": "text", + "content": " Shuran Song" + }, + { + "bbox": [ + 54, + 160, + 536, + 177 + ], + "type": "inline_equation", + "content": "^{\\diamond}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 143, + 202, + 192, + 216 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 202, + 192, + 216 + ], + "spans": [ + { + "bbox": [ + 143, + 202, + 192, + 216 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 228, + 290, + 528 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 228, + 290, + 528 + ], + "spans": [ + { + "bbox": [ + 46, + 228, + 290, + 528 + ], + "type": "text", + "content": "For robots to be generally useful, they must be able to find arbitrary objects described by people (i.e., be language-driven) even without expensive navigation training on in-domain data (i.e., perform zero-shot inference). We explore these capabilities in a unified setting: language-driven zero-shot object navigation (L-ZSON). Inspired by the recent success of open-vocabulary models for image classification, we investigate a straightforward framework, CLIP on Wheels (CoW), to adapt open-vocabulary models to this task without fine-tuning. To better evaluate L-ZSON, we introduce the PASTURE benchmark, which considers finding uncommon objects, objects described by spatial and appearance attributes, and hidden objects described relative to visible objects. We conduct an in-depth empirical study by directly deploying 22 CoW baselines across HABITAT, ROBOTHOR, and PASTURE. In total, we evaluate over 90k navigation episodes and find that (1) CoW baselines often struggle to leverage language descriptions but are proficient at finding uncommon objects. (2) A simple CoW, with CLIP-based object localization and classical exploration—and no additional training—matches the navigation efficiency of a state-of-the-art ZSON method trained for 500M steps on HABITAT MP3D data. This same CoW provides a 15.6 percentage point improvement in success over a state-of-the-art ROBOTHOR ZSON model." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 548, + 128, + 562 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 548, + 128, + 562 + ], + "spans": [ + { + "bbox": [ + 47, + 548, + 128, + 562 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 569, + 287, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 569, + 287, + 677 + ], + "spans": [ + { + "bbox": [ + 46, + 569, + 287, + 677 + ], + "type": "text", + "content": "To be more widely applicable, robots should be language-driven: able to deduce goals based on arbitrary text input instead of being constrained to a fixed set of object categories. While existing image classification, semantic segmentation, and object navigation benchmarks like ImageNet-1k [65], ImageNet-21k [22], MS-COCO [45], LVIS [28], HABITAT [67], and ROBOTHOR [18] include a vast array of everyday items, they do not capture all objects that matter to people. For instance, a lost \"toy airplane\" may" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 310, + 201, + 526, + 211 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 310, + 201, + 526, + 211 + ], + "spans": [ + { + "bbox": [ + 310, + 201, + 526, + 211 + ], + "type": "text", + "content": "Sample tasks Top-down visualization Egocentric Observations" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 308, + 213, + 386, + 221 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 213, + 386, + 221 + ], + "spans": [ + { + "bbox": [ + 308, + 213, + 386, + 221 + ], + "type": "text", + "content": "(a) Finding uncommon objects" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 309, + 222, + 544, + 262 + ], + "blocks": [ + { + "bbox": [ + 309, + 222, + 544, + 262 + ], + "lines": [ + { + "bbox": [ + 309, + 222, + 544, + 262 + ], + "spans": [ + { + "bbox": [ + 309, + 222, + 544, + 262 + ], + "type": "image", + "image_path": "16652846a9fca43c2db6cb634fa5e8d0256f61a82637d3a954319e5ebb771b2f.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 308, + 264, + 479, + 273 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 264, + 479, + 273 + ], + "spans": [ + { + "bbox": [ + 308, + 264, + 479, + 273 + ], + "type": "text", + "content": "(b) Finding objects based on attributes in the presence of distractors" + } + ] + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 309, + 274, + 544, + 313 + ], + "blocks": [ + { + "bbox": [ + 309, + 274, + 544, + 313 + ], + "lines": [ + { + "bbox": [ + 309, + 274, + 544, + 313 + ], + "spans": [ + { + "bbox": [ + 309, + 274, + 544, + 313 + ], + "type": "image", + "image_path": "31f0ab0fe5e9a07a5095a87d64338440d54a55360dd693f2b489171f040d34fa.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "bbox": [ + 309, + 315, + 450, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 309, + 315, + 450, + 323 + ], + "spans": [ + { + "bbox": [ + 309, + 315, + 450, + 323 + ], + "type": "text", + "content": "(c) Finding hidden objects in the presence of distractors" + } + ] + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 309, + 323, + 544, + 364 + ], + "blocks": [ + { + "bbox": [ + 309, + 323, + 544, + 364 + ], + "lines": [ + { + "bbox": [ + 309, + 323, + 544, + 364 + ], + "spans": [ + { + "bbox": [ + 309, + 323, + 544, + 364 + ], + "type": "image", + "image_path": "866fb55efa25e9630e6e03edbb445623423d7bcd17120fc94c04cd76ad751fdc.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 372, + 547, + 449 + ], + "lines": [ + { + "bbox": [ + 304, + 372, + 547, + 449 + ], + "spans": [ + { + "bbox": [ + 304, + 372, + 547, + 449 + ], + "type": "text", + "content": "Figure 1. The PASTURE benchmark for L-ZSON. Text specifies navigation goal objects. Agents do not train on these tasks, making the evaluation protocol zero-shot. (a) Uncommon object goals like \"llama wicker basket\", not found in existing navigation benchmarks. (b) Appearance, spatial descriptions, which are necessary to find the correct object. (c) Hidden object descriptions, which localize objects that are not visible." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 459, + 547, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 459, + 547, + 483 + ], + "spans": [ + { + "bbox": [ + 304, + 459, + 547, + 483 + ], + "type": "text", + "content": "become relevant in a kindergarten classroom, but this object is not annotated in any of the aforementioned datasets." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 484, + 547, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 484, + 547, + 628 + ], + "spans": [ + { + "bbox": [ + 304, + 484, + 547, + 628 + ], + "type": "text", + "content": "In this paper, we study Language-driven zero-shot object navigation (L-ZSON), a more challenging but also more applicable version of object navigation [5, 18, 67, 79, 89] and ZSON [38, 46] tasks. In L-ZSON, an agent must find an object based on a description, which may contain different levels of granularity (e.g., \"toy airplane\", \"toy airplane under the bed\", or \"wooden toy airplane\"). L-ZSON encompasses ZSON, which specifies only the target category [38, 46]. Since L-ZSON is \"zero-shot\", we consider agents without access to navigation training on the target objects or domains. This reflects realistic application scenarios, where the environment and object set may not be known a priori." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 630, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 630, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 630, + 547, + 715 + ], + "type": "text", + "content": "Performing L-ZSON in any environment with unstructured language input is challenging; however, recent advances in open-vocabulary models for image classification [35, 58, 61], object detection [4, 21, 27, 36, 43, 47, 49, 62, 88], and semantic segmentation [3, 6, 15, 33, 36, 37, 86] present a promising foundation. These models provide an interface where one specifies—in text—the arbitrary ob" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 683, + 287, + 702 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 683, + 287, + 702 + ], + "spans": [ + { + "bbox": [ + 47, + 683, + 287, + 702 + ], + "type": "inline_equation", + "content": "^{\\text{©}}" + }, + { + "bbox": [ + 47, + 683, + 287, + 702 + ], + "type": "text", + "content": " Columbia University, " + }, + { + "bbox": [ + 47, + 683, + 287, + 702 + ], + "type": "inline_equation", + "content": "^{\\dagger}" + }, + { + "bbox": [ + 47, + 683, + 287, + 702 + ], + "type": "text", + "content": "University of Washington. Correspondence to sy@cs.columbia.edu." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 59, + 703, + 260, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 703, + 260, + 712 + ], + "spans": [ + { + "bbox": [ + 59, + 703, + 260, + 712 + ], + "type": "text", + "content": "1For code, data, and videos, see cow.cs.columbia.edu/" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "23171" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 167 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 167 + ], + "type": "text", + "content": "jects they wish to classify, detect, or segment. For example, CLIP [61] open-vocabulary classifiers compute similarity scores between an input image and a set of user-specified captions (e.g., \"a photo of a toy airplane.\", ..., selecting the caption with the highest score to determine the image classification label. Given the flexibility of these models, we would like to understand their capability to execute embodied tasks even without additional training." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 168, + 287, + 194 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 168, + 287, + 194 + ], + "spans": [ + { + "bbox": [ + 47, + 168, + 287, + 194 + ], + "type": "text", + "content": "To this end, we present baselines and benchmarks for L-ZSON. More specifically:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 198, + 289, + 466 + ], + "type": "list", + "angle": 0, + "index": 4, + "blocks": [ + { + "bbox": [ + 47, + 198, + 289, + 342 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 198, + 289, + 342 + ], + "spans": [ + { + "bbox": [ + 47, + 198, + 289, + 342 + ], + "type": "text", + "content": "- A collection of baseline algorithms, CLIP on Wheels (CoW), which adapt open-vocabulary models to the task of L-ZSON. CoW takes inspiration from the semantic mapping line of work [11, 41, 53], and decomposes the navigation task into exploration when the language target is not confidently localized, and target-driven planning otherwise. CoW retains the textual user interface of the original open-vocabulary model and does not require any navigation training. We evaluate 22 CoWs, ablating over many open-vocabulary models, exploration policies, backbones, prompting strategies, and post-processing strategies." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 346, + 289, + 466 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 346, + 289, + 466 + ], + "spans": [ + { + "bbox": [ + 47, + 346, + 289, + 466 + ], + "type": "text", + "content": "- A new benchmark, PASTURE, to evaluate CoW and future methods on L-ZSON. We design PASTURE, visualized in Fig. 1, to study capabilities that traditional object navigation agents, which are trained on a fixed set of categories, do not possess. We consider the ability to find (1) uncommon objects (e.g., \"tie-dye surfboard\"), (2) objects by their spatial and appearance attributes in the presence of distractor objects (e.g., \"green apple\" vs. \"red apple\"), and (3) objects that cannot be visually observed (e.g., \"mug under the bed\")." + } + ] + } + ], + "index": 3 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 46, + 473, + 288, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 473, + 288, + 641 + ], + "spans": [ + { + "bbox": [ + 46, + 473, + 288, + 641 + ], + "type": "text", + "content": "Together the CoW baselines and PASTURE benchmark allow us to conduct extensive studies on the capabilities of open-vocabulary models in the context of L-ZSON embodied tasks. Our experiments demonstrate CoW's potential on uncommon objects and limitations in taking full advantage of language descriptions—thereby providing empirical motivation for future studies. To contextualize CoW relative to prior zero-shot methods, we additionally evaluate on the HABITAT MP3D dataset. We find that our best CoW achieves navigation efficiency (SPL) that matches a state-of-the-art ZSON method [46] that trains on MP3D training data for 500M steps. On a ROBOTHOR object subset, considered in prior work, the same CoW beats the leading method [38] by 15.6 percentage points in task success." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 654, + 134, + 666 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 654, + 134, + 666 + ], + "spans": [ + { + "bbox": [ + 47, + 654, + 134, + 666 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "text", + "content": "Mapping and exploration. Exploring effectively with a mobile robot is a long-standing problem in vision and robotics. Classical methods often decompose the task" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 72, + 545, + 192 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 192 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 192 + ], + "type": "text", + "content": "into map reconstruction [30, 32, 51, 52, 72], agent localization [17, 20, 54], and planning [41, 78]. Recent work investigates learned alternatives for exploration [7, 14, 56, 57, 63]. Here, agents are often trained end-to-end with self-supervised rewards (e.g., curiosity [57]) or supervised rewards (e.g., state visitation counts [25, 73, 75]). Learning-based methods typically need less hand-tuning, but require millions of training steps and reward engineering. We test both classical and learnable exploration strategies in the context of CoW to study their applicability to L-ZSON." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 195, + 545, + 278 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 195, + 545, + 278 + ], + "spans": [ + { + "bbox": [ + 304, + 195, + 545, + 278 + ], + "type": "text", + "content": "Goal-conditioned navigation. Apart from open-ended exploration, many navigation tasks are goal-conditioned, where the agent needs to navigate to a specified position (i.e., point goal [11, 12, 26, 31, 66, 77, 81, 83]), view of the environment (i.e., image goal [48, 64, 89]), or object category (i.e., object goal [1, 9, 10, 12, 19, 44, 74, 79, 84]). We consider an object goal navigation task." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 282, + 545, + 462 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 282, + 545, + 462 + ], + "spans": [ + { + "bbox": [ + 304, + 282, + 545, + 462 + ], + "type": "text", + "content": "Vision-Language Navigation. Prior work investigates language-based navigation, where language provides step-by-step instructions for the task [2, 34, 39, 40, 71]. This line of work demonstrates the benefits of additional language input for robot navigation, especially for long-horizon tasks (e.g., room-to-room navigation [40]). However, providing detailed step-by-step instructions (e.g., move 3 meters south [34]) could be challenging and time-consuming. In our L-ZSON task, an algorithm gets natural language as the goal description instead of low-level instructions. Prior work also investigates navigation with target descriptions in supervised settings [42, 60, 87]. In contrast, we explore a zero-shot evaluation protocol and consider finding hidden objects (e.g., \"mug under bed\") and uncommon objects (e.g., \"tie-dye surfboard\")." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 465, + 545, + 644 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 465, + 545, + 644 + ], + "spans": [ + { + "bbox": [ + 304, + 465, + 545, + 644 + ], + "type": "text", + "content": "Zero-shot object navigation (ZSON). Recent work studies object navigation in zero-shot settings, where agents are evaluated on object categories that they are not explicitly trained on [38, 46]. Our task encompasses ZSON; however it also considers cases where more information—object attributes or hidden objects descriptions—is specified. Khandelwal et al. [38] train on a subset of ROBOTHOR categories and evaluate on a held-out set. In concurrent work, Majumdar et al. [46] train on an image goal navigation task and evaluate on object navigation downstream by leveraging CLIP multi-modal embeddings. Both algorithms necessitate navigation training for millions of steps and train separate models for each simulation domain. In contrast, CoW baselines do not necessitate any simulation training and can be deployed in multiple environments." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 656, + 416, + 669 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 656, + 416, + 669 + ], + "spans": [ + { + "bbox": [ + 306, + 656, + 416, + 669 + ], + "type": "text", + "content": "3. The L-ZSON Task" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "content": "Language-driven zero-shot object navigation (L-ZSON) involves navigating to goal objects, specified in language, without explicit training to do so. Let " + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{O}" + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "content": " denote a set of" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "23172" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 49, + 70, + 309, + 170 + ], + "blocks": [ + { + "bbox": [ + 49, + 70, + 309, + 170 + ], + "lines": [ + { + "bbox": [ + 49, + 70, + 309, + 170 + ], + "spans": [ + { + "bbox": [ + 49, + 70, + 309, + 170 + ], + "type": "image", + "image_path": "432331edfae158dc49e0cd33234c1db9de98c336e4e79390c16442d512f8dfb2.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 177, + 287, + 211 + ], + "lines": [ + { + "bbox": [ + 46, + 177, + 287, + 211 + ], + "spans": [ + { + "bbox": [ + 46, + 177, + 287, + 211 + ], + "type": "text", + "content": "Figure 2. CLIP on Wheels (CoW) overview. A CoW uses a policy to explore and an object localizer (e.g., an open-vocabulary object detector) to determine if an object goal is in view." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 304, + 152, + 545, + 208 + ], + "lines": [ + { + "bbox": [ + 304, + 152, + 545, + 208 + ], + "spans": [ + { + "bbox": [ + 304, + 152, + 545, + 208 + ], + "type": "text", + "content": "Figure 3. Mapping. Top-down map created from egocentric depth observations as a CoW roams a space. (a) Frontier Based Exploration [82] showing a planned path exploration path to the next frontier. (b) Back-projected object relevance scores provide object goal targets when a CoW has found an object." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 216, + 290, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 216, + 290, + 396 + ], + "spans": [ + { + "bbox": [ + 46, + 216, + 290, + 396 + ], + "type": "text", + "content": "natural language descriptions of target objects with potentially many attributes (e.g., \"plant\", \"snake plant\", \"plant under the bed\", etc.). This contrasts with definitions studied in prior object navigation [5, 18] and ZSON [38, 46] tasks, which focus on high-level categories like \"plant\". Let " + }, + { + "bbox": [ + 46, + 216, + 290, + 396 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 46, + 216, + 290, + 396 + ], + "type": "text", + "content": " denote the set of navigation scenes. Let " + }, + { + "bbox": [ + 46, + 216, + 290, + 396 + ], + "type": "inline_equation", + "content": "p_0" + }, + { + "bbox": [ + 46, + 216, + 290, + 396 + ], + "type": "text", + "content": " describe the initial pose of an agent. A navigation episode " + }, + { + "bbox": [ + 46, + 216, + 290, + 396 + ], + "type": "inline_equation", + "content": "\\tau \\in \\mathcal{T}" + }, + { + "bbox": [ + 46, + 216, + 290, + 396 + ], + "type": "text", + "content": " is written as a tuple " + }, + { + "bbox": [ + 46, + 216, + 290, + 396 + ], + "type": "inline_equation", + "content": "\\tau = (s, o, p_0)" + }, + { + "bbox": [ + 46, + 216, + 290, + 396 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 216, + 290, + 396 + ], + "type": "inline_equation", + "content": "s \\in S" + }, + { + "bbox": [ + 46, + 216, + 290, + 396 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 216, + 290, + 396 + ], + "type": "inline_equation", + "content": "o \\in \\mathcal{O}" + }, + { + "bbox": [ + 46, + 216, + 290, + 396 + ], + "type": "text", + "content": ". Each " + }, + { + "bbox": [ + 46, + 216, + 290, + 396 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 46, + 216, + 290, + 396 + ], + "type": "text", + "content": " is a zero-shot task as tuples of this form are not seen during training. Starting at " + }, + { + "bbox": [ + 46, + 216, + 290, + 396 + ], + "type": "inline_equation", + "content": "p_0" + }, + { + "bbox": [ + 46, + 216, + 290, + 396 + ], + "type": "text", + "content": ", an embodied agent's goal is to find " + }, + { + "bbox": [ + 46, + 216, + 290, + 396 + ], + "type": "inline_equation", + "content": "o" + }, + { + "bbox": [ + 46, + 216, + 290, + 396 + ], + "type": "text", + "content": ". The agent receives observations and sensor readings " + }, + { + "bbox": [ + 46, + 216, + 290, + 396 + ], + "type": "inline_equation", + "content": "I_t" + }, + { + "bbox": [ + 46, + 216, + 290, + 396 + ], + "type": "text", + "content": " (e.g., RGB-D images). At each timestep " + }, + { + "bbox": [ + 46, + 216, + 290, + 396 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 216, + 290, + 396 + ], + "type": "text", + "content": ", the agent executes a navigation action " + }, + { + "bbox": [ + 46, + 216, + 290, + 396 + ], + "type": "inline_equation", + "content": "a \\in \\mathcal{A}" + }, + { + "bbox": [ + 46, + 216, + 290, + 396 + ], + "type": "text", + "content": ". A special action " + }, + { + "bbox": [ + 46, + 216, + 290, + 396 + ], + "type": "inline_equation", + "content": "\\text{STOP} \\in \\mathcal{A}" + }, + { + "bbox": [ + 46, + 216, + 290, + 396 + ], + "type": "text", + "content": " terminates the episode. If the agent is within " + }, + { + "bbox": [ + 46, + 216, + 290, + 396 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 46, + 216, + 290, + 396 + ], + "type": "text", + "content": " units of " + }, + { + "bbox": [ + 46, + 216, + 290, + 396 + ], + "type": "inline_equation", + "content": "o" + }, + { + "bbox": [ + 46, + 216, + 290, + 396 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 216, + 290, + 396 + ], + "type": "inline_equation", + "content": "o" + }, + { + "bbox": [ + 46, + 216, + 290, + 396 + ], + "type": "text", + "content": " meets a visibility criteria, the episode is successful." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 405, + 234, + 418 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 405, + 234, + 418 + ], + "spans": [ + { + "bbox": [ + 47, + 405, + 234, + 418 + ], + "type": "text", + "content": "4. CLIP on Wheels (CoW) Baselines" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 425, + 288, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 425, + 288, + 617 + ], + "spans": [ + { + "bbox": [ + 46, + 425, + 288, + 617 + ], + "type": "text", + "content": "To address L-ZSON, we investigate a simple baseline approach, CoW, which adapts open-vocabulary models like CLIP to make them suitable for the task. A CoW takes as input an egocentric RGB-D image and an object goal specified in language. As a CoW moves, it updates a top-down map of the world created using RGB-D observations and pose estimates (Sec. 4.1). Each CoW gets an exploration policy and a zero-shot object localization module as seen in Fig. 2. To observe diverse views of the scene, a CoW explores using a policy (Sec. 4.2). As the CoW roams, it keeps track of its confidence about the target object's location using an object localization module (Sec. 4.3) and its top-down map. When a CoW's confidence exceeds a threshold, it plans to the location of the goal and issues the STOP action. We now describe the ingredients used to make the CoWs evaluated in our experiments (Sec. 6)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 623, + 174, + 636 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 623, + 174, + 636 + ], + "spans": [ + { + "bbox": [ + 47, + 623, + 174, + 636 + ], + "type": "text", + "content": "4.1. Depth-based Mapping" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 642, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 642, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 642, + 287, + 713 + ], + "type": "text", + "content": "As a CoW moves, it constructs a top-down map using input depth, pose estimates, and known agent height. The map is initialized using known camera intrinsics and the first depth image. Since a CoW knows the intended consequences of its actions (e.g., MOVEFORWARD should result in a " + }, + { + "bbox": [ + 46, + 642, + 287, + 713 + ], + "type": "inline_equation", + "content": "0.25\\mathrm{m}" + }, + { + "bbox": [ + 46, + 642, + 287, + 713 + ], + "type": "text", + "content": " translation), each action is represented as a" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 309, + 72, + 425, + 145 + ], + "blocks": [ + { + "bbox": [ + 309, + 72, + 425, + 145 + ], + "lines": [ + { + "bbox": [ + 309, + 72, + 425, + 145 + ], + "spans": [ + { + "bbox": [ + 309, + 72, + 425, + 145 + ], + "type": "image", + "image_path": "744a07bb4f5244059a6ff11e7c1a5757595ad2d06b97fe38d12f86a9ccb77226.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 427, + 72, + 543, + 144 + ], + "blocks": [ + { + "bbox": [ + 427, + 72, + 543, + 144 + ], + "lines": [ + { + "bbox": [ + 427, + 72, + 543, + 144 + ], + "spans": [ + { + "bbox": [ + 427, + 72, + 543, + 144 + ], + "type": "image", + "image_path": "b530d1ea053a38abe377f49c828f5ba002be7ddfec13acf8b51e25e253fbd124.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 215, + 546, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 215, + 546, + 346 + ], + "spans": [ + { + "bbox": [ + 304, + 215, + 546, + 346 + ], + "type": "text", + "content": "pose delta transform to approximate a transition. To deal with noise associated with actuation or depth, a CoW maintains a map at " + }, + { + "bbox": [ + 304, + 215, + 546, + 346 + ], + "type": "inline_equation", + "content": "0.125\\mathrm{m}" + }, + { + "bbox": [ + 304, + 215, + 546, + 346 + ], + "type": "text", + "content": " resolution. To improve map accuracy, a CoW checks for failed actions by comparing successive depth frames for movements (see Appx. A for details). Using known agent height " + }, + { + "bbox": [ + 304, + 215, + 546, + 346 + ], + "type": "inline_equation", + "content": "(0.9\\mathrm{m})" + }, + { + "bbox": [ + 304, + 215, + 546, + 346 + ], + "type": "text", + "content": ", map cells are projected to the ground plane to maintain a top-down representation of the world, which suffices for most navigation applications. Cells close to the floor are considered free space (white points in Fig. 3 (a)), while other cells are considered occupied (blue points in Fig. 3 (a))." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 354, + 384, + 367 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 354, + 384, + 367 + ], + "spans": [ + { + "bbox": [ + 306, + 354, + 384, + 367 + ], + "type": "text", + "content": "4.2. Exploration" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 373, + 545, + 421 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 373, + 545, + 421 + ], + "spans": [ + { + "bbox": [ + 304, + 373, + 545, + 421 + ], + "type": "text", + "content": "Exploration generates diverse egocentric views so a CoW is more likely to view the language-specified target object. We consider two exploration methods, frontier-based and learning-based." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 423, + 546, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 423, + 546, + 555 + ], + "spans": [ + { + "bbox": [ + 304, + 423, + 546, + 555 + ], + "type": "text", + "content": "Frontier based exploration (FBE) [82]. Using the top-down map discussed in Sec. 4.1, a CoW can navigate using a simple exploration heuristic: move to the frontier between free and unknown space to discover new regions. Once the navigator reaches a frontier (visualized as purple points in Fig. 3 (a)), it moves greedily to the next closest frontier. Since the map is updated at every timestep, a noisy pose estimate can contribute to inaccuracies. For example, narrow passages may collapse in the map due to pose drift. To circumvent such problems, we reinitialize the map when no paths exist to any frontiers in the map." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 558, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 558, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 558, + 545, + 713 + ], + "type": "text", + "content": "Learnable exploration. In addition to FBE, we consider learnable alternatives, which may explore more intelligently but incur substantial training costs. We investigate an architecture and reward structure similar to prior work in embodied AI (e.g., [25, 38, 75]). Specifically, we adopt a frozen CLIP backbone with a trainable GRU [16] and linear heads for the actor and critic networks. We train agents independently in HABITAT [67] and ROBOTHOR [18] simulation environments for 60M steps each, using DD-PPO [69, 77] in the AllenAct [76] framework. We employ a simple count-based reward [73]. All training scenes are disjoint from downstream navigation test scenes. For details on reward, hyperparameters, and training, see Appx. B." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "23173" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 163, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 163, + 85 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 163, + 85 + ], + "type": "text", + "content": "4.3. Object Localization" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 91, + 287, + 187 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 91, + 287, + 187 + ], + "spans": [ + { + "bbox": [ + 46, + 91, + 287, + 187 + ], + "type": "text", + "content": "Successful navigation depends on object localization: the ability to tell if and where an object is in an image. Regions of high object relevance, extracted from 2D images, are projected to the depth-based map (Fig. 3 (b)) where they serve as natural navigation targets. To determine if and when a target is in an image, we consider the following object localization modules, used in our experiments (Sec. 6). For more details see Appx. C." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 190, + 289, + 429 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 190, + 289, + 429 + ], + "spans": [ + { + "bbox": [ + 46, + 190, + 289, + 429 + ], + "type": "text", + "content": "Adapting open-vocabulary classifiers. We experiment with three strategies to turn CLIP [61] models into object localizers. First, we utilize the CLIP text encoder to embed " + }, + { + "bbox": [ + 46, + 190, + 289, + 429 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 190, + 289, + 429 + ], + "type": "text", + "content": " referring expressions, which specify regions where the target object may appear in the image. For example, \"a plant in the top left of the image.\" We then match the language embeddings against a CLIP visual embedding for the current observation. We compute similarity between the image and text features to determine relevance scores over the regions. Second, we discretize the image into " + }, + { + "bbox": [ + 46, + 190, + 289, + 429 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 190, + 289, + 429 + ], + "type": "text", + "content": " smaller patches and obtain CLIP patch embeddings. We then convolve each patch embedding with a CLIP text embedding for the target object. If the object is in a patch, the relevance score for that patch should be high. Third, we modify an interpretability method [13, 70] designed to extract object relevancy from vision transformers (ViTs) [24]. Using a target CLIP text embedding and gradient information accumulated through the CLIP vision encoder, we construct a relevance map over input pixels, which qualitatively segments the target when it is in view." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 432, + 287, + 576 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 432, + 287, + 576 + ], + "spans": [ + { + "bbox": [ + 46, + 432, + 287, + 576 + ], + "type": "text", + "content": "Adapting open-vocabulary detectors and segmentors. In addition to CLIP-based methods, we consider two additional open-vocabulary models for object localization. First, the MDETR segmentation model [36], which extends the DETR detector [8] to take arbitrary text and images as input and output box detections. The base model is fine-tuned on PhraseCut [80], a dataset of paired masks and attribute descriptions, to support segmentation. Second, we consider the OWL-ViT detector [49], which uses a set prediction fine-tuning recipe to turn CLIP-like models into object detectors. We use this MDETR and OWL-ViT models to directly query images for targets." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 578, + 287, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 578, + 287, + 687 + ], + "spans": [ + { + "bbox": [ + 46, + 578, + 287, + 687 + ], + "type": "text", + "content": "Post-processing. The aforementioned models give continuous valued predictions. However, we are interested in binary masks giving if and where objects are in images. Hence, we threshold predictions for each model (see Appx. C for details). We further investigate two strategies for using the masks downstream: (1) using the entire mask or (2) using the center pixel. The second strategy is reasonable because only part of an object needs to be detected for successful navigation." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 689, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 287, + 715 + ], + "type": "text", + "content": "Target driven planning. Recall, CoWs have depth sensors. We back-project object relevance from 2D images into the" + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 308, + 72, + 545, + 169 + ], + "blocks": [ + { + "bbox": [ + 308, + 72, + 545, + 169 + ], + "lines": [ + { + "bbox": [ + 308, + 72, + 545, + 169 + ], + "spans": [ + { + "bbox": [ + 308, + 72, + 545, + 169 + ], + "type": "image", + "image_path": "028d123bae99555101968f4ed45ab80b17727dbff4968acaa1239af1a3cf8653.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 345, + 176, + 506, + 188 + ], + "lines": [ + { + "bbox": [ + 345, + 176, + 506, + 188 + ], + "spans": [ + { + "bbox": [ + 345, + 176, + 506, + 188 + ], + "type": "text", + "content": "Figure 4. Uncommon objects in PASTURE." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 194, + 546, + 242 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 194, + 546, + 242 + ], + "spans": [ + { + "bbox": [ + 304, + 194, + 546, + 242 + ], + "type": "text", + "content": "depth-based map (Sec. 4.1). We keep only the max relevance for each map cell (Fig. 3 (b)). CoWs can then plan to high relevance areas in the map. See Appx. D for additional method visualization." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 245, + 546, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 245, + 546, + 376 + ], + "spans": [ + { + "bbox": [ + 304, + 245, + 546, + 376 + ], + "type": "text", + "content": "Incorporating object priors. Since CoW does not train or fine-tune on navigation datasets, we investigate alternative approaches to inject object-level priors into the model. For each target object, we prompt GPT-3.5 [55] to generate rooms where the target objects are likely to be found. For example, GPT-3.5 states that apples are likely to be found in \"kitchen\" or \"dining room\" scenes. Following this prior, a GPT-3.5 enabled CoW first uses its object localization module to localize a kitchen or a dining room, and then looks for an apple. This straightforward extension, demonstrates how outside information can be incorporated into CoW." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 387, + 457, + 399 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 387, + 457, + 399 + ], + "spans": [ + { + "bbox": [ + 305, + 387, + 457, + 399 + ], + "type": "text", + "content": "5. The PASTURE Benchmark" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 407, + 545, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 407, + 545, + 491 + ], + "spans": [ + { + "bbox": [ + 304, + 407, + 545, + 491 + ], + "type": "text", + "content": "To evaluate CoW baselines and future methods on L-ZSON, we introduce the PASTURE evaluation benchmark. PASTURE builds on ROBOTHER validation scenes, which have parallel environments in the real-world. We target ROBOTHER to facilitate future real-world benchmarking. PASTURE probes for seven capabilities explained in Sec. 5.1. We provide dataset statistics in Sec. 5.2." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 498, + 400, + 510 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 498, + 400, + 510 + ], + "spans": [ + { + "bbox": [ + 306, + 498, + 400, + 510 + ], + "type": "text", + "content": "5.1. PASTURE Tasks" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 317, + 516, + 533, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 516, + 533, + 529 + ], + "spans": [ + { + "bbox": [ + 317, + 516, + 533, + 529 + ], + "type": "text", + "content": "PASTURE evaluates seven core L-ZSON capabilities." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 531, + 545, + 674 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 531, + 545, + 674 + ], + "spans": [ + { + "bbox": [ + 304, + 531, + 545, + 674 + ], + "type": "text", + "content": "Uncommon objects. Traditional benchmarks (e.g., ROBOTHOR and HABITAT MP3D) evaluate agents on common object categories like TVs; however, given the rich diversity of objects in homes, we would like to understand navigation performance on uncommon objects. Hence we add 12 new objects to each room. We use names shown in Fig. 4 as instance labels, which are minimal descriptions to identify each object. Some identifiers refer to text in images (e.g., \"whiteboard saying CVPR\") or to appearance attributes (e.g., \"wooden toy airplane\"). Other objects are less common in North America, like \"mate\", which is a popular Argentinian drink." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 677, + 546, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 546, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 546, + 715 + ], + "type": "text", + "content": "Appearance descriptions. To evaluate if baselines can take advantage of visual attributes, we introduce descriptions of the form: “" + }, + { + "bbox": [ + 304, + 677, + 546, + 715 + ], + "type": "inline_equation", + "content": "\\{size\\}" + }, + { + "bbox": [ + 304, + 677, + 546, + 715 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 677, + 546, + 715 + ], + "type": "inline_equation", + "content": "\\{color\\}" + }, + { + "bbox": [ + 304, + 677, + 546, + 715 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 677, + 546, + 715 + ], + "type": "inline_equation", + "content": "\\{material\\}" + }, + { + "bbox": [ + 304, + 677, + 546, + 715 + ], + "type": "inline_equation", + "content": "\\{object\\}" + }, + { + "bbox": [ + 304, + 677, + 546, + 715 + ], + "type": "text", + "content": ". For" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "23174" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "text", + "content": "example: \"small, red apple\", \"orange basketball\", \"small, black, metallic alarm clock\". Objects are considered small if their 3D bounding box diagonal is below a threshold. We determine color and materials by inspection." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 123, + 286, + 207 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 123, + 286, + 207 + ], + "spans": [ + { + "bbox": [ + 46, + 123, + 286, + 207 + ], + "type": "text", + "content": "Spatial descriptions. To test if agents can leverage spatial information in navigation, we introduce descriptions: “{object} on top of " + }, + { + "bbox": [ + 46, + 123, + 286, + 207 + ], + "type": "inline_equation", + "content": "\\{x\\}" + }, + { + "bbox": [ + 46, + 123, + 286, + 207 + ], + "type": "text", + "content": ", near " + }, + { + "bbox": [ + 46, + 123, + 286, + 207 + ], + "type": "inline_equation", + "content": "\\{y\\}" + }, + { + "bbox": [ + 46, + 123, + 286, + 207 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 123, + 286, + 207 + ], + "type": "inline_equation", + "content": "\\{z\\}" + }, + { + "bbox": [ + 46, + 123, + 286, + 207 + ], + "type": "text", + "content": ", ...”. For example, “house plant on a dresser near a spray bottle”. To determine [on top of] relations, we use THOR metadata and to determine [nearness] we use a distance threshold between pairs of objects. We inspect all descriptions for correctness." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 210, + 286, + 342 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 210, + 286, + 342 + ], + "spans": [ + { + "bbox": [ + 46, + 210, + 286, + 342 + ], + "type": "text", + "content": "Appearance descriptions with distractors. To probe if appearance attributes better facilitate finding objects in the presence of distractors, we reuse the appearance captions from before, but evaluate on an modified environment with two visually distinct instances of each ROBOTHOR object category. For example, for the task of finding a \"red apple\", we have both a red apple and a green apple in the room. A navigator must leverage appearance information—and not just class information—to successfully complete the task. Distractor objects are sufficiently far from the target objects so that finding a distractor cannot count as success." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 345, + 286, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 345, + 286, + 380 + ], + "spans": [ + { + "bbox": [ + 46, + 345, + 286, + 380 + ], + "type": "text", + "content": "Spatial descriptions with distractors. This capability is similar to the one above; however, we evaluate with spatial descriptions in the presence of distractor objects." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 384, + 286, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 384, + 286, + 468 + ], + "spans": [ + { + "bbox": [ + 46, + 384, + 286, + 468 + ], + "type": "text", + "content": "Hidden object descriptions. An ideal object navigator should find objects, even when they are hidden. Hence, we introduce descriptions: “{object} under/in {x}”. For example, “basketball in the dresser drawers” or “vase under the sofa”. We sample large objects (e.g., beds, sofas, dressers) in each scene to determine [under/in] relations. Additionally we remove visible instances of {object} from the room." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 471, + 286, + 530 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 471, + 286, + 530 + ], + "spans": [ + { + "bbox": [ + 46, + 471, + 286, + 530 + ], + "type": "text", + "content": "Hidden object descriptions with distractors. We use the hidden object descriptions from before, but reintroduce visible instances of " + }, + { + "bbox": [ + 46, + 471, + 286, + 530 + ], + "type": "inline_equation", + "content": "\\{object\\}" + }, + { + "bbox": [ + 46, + 471, + 286, + 530 + ], + "type": "text", + "content": " to serve as distractors. Consider finding a \"mug under the bed\". A distractor mug will also appear in the scene making the task more challenging." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 539, + 214, + 551 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 539, + 214, + 551 + ], + "spans": [ + { + "bbox": [ + 47, + 539, + 214, + 551 + ], + "type": "text", + "content": "5.2. Dataset Creation and Statistics" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 558, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 558, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 558, + 286, + 713 + ], + "type": "text", + "content": "PASTURE contains three variations for each of the original 15 validation ROBOTHOR rooms: uncommon objects added, additional object instances added, and target objects removed. For each of the seven settings presented above, we evaluate over 12 object instances in 15 rooms with two initial agent starting locations. Hence PASTURE consists of 2,520 tasks, which is a similar order of magnitude to ROBOTHOR (1,800) and HABITAT MP3D (2,195) validation sets. For appearance attributes, " + }, + { + "bbox": [ + 46, + 558, + 286, + 713 + ], + "type": "inline_equation", + "content": "47\\%" + }, + { + "bbox": [ + 46, + 558, + 286, + 713 + ], + "type": "text", + "content": " of the objects are considered \"small\". Each object gets an average of 1.2 color descriptors out of 22 possible choices, and 0.6 material descriptors out of 5 possible choices. Similarly, for spatial attributes, each object gets one object it is on top of or in (e.g.," + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 72, + 545, + 109 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 545, + 109 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 545, + 109 + ], + "type": "text", + "content": "“vase in a shelving unit”) and an average of 1.9 objects it is near. For a sample of appearance and spatial attributes see Fig. 1. For more dataset details and statistics see Appx. E." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 118, + 387, + 132 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 118, + 387, + 132 + ], + "spans": [ + { + "bbox": [ + 306, + 118, + 387, + 132 + ], + "type": "text", + "content": "6. Experiments" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 139, + 545, + 222 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 139, + 545, + 222 + ], + "spans": [ + { + "bbox": [ + 304, + 139, + 545, + 222 + ], + "type": "text", + "content": "We first present our experimental setup, including the datasets, metrics, embodiment, and baselines considered in our study (Sec. 6.1). Then we present results on PASTURE, thereby elucidating the strengths and weakness of CoW baselines for L-ZSON (Sec. 6.2). Finally, we compare to prior ZSON art in ROBOTHER and HABITAT (MP3D) environments (Sec. 6.3)." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 230, + 420, + 243 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 230, + 420, + 243 + ], + "spans": [ + { + "bbox": [ + 306, + 230, + 420, + 243 + ], + "type": "text", + "content": "6.1. Experimental setup" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 251, + 545, + 358 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 251, + 545, + 358 + ], + "spans": [ + { + "bbox": [ + 304, + 251, + 545, + 358 + ], + "type": "text", + "content": "**Environments.** We consider PASTURE (Sec. 5), ROBOTHER [18], and HABITAT (MP3D) [67] validation sets as our test sets. We utilize validation sets for testing because official test set ground-truth is not publicly available. Domains are setup with noise that is faithful to their original challenge settings. For ROBOTHER—and by extension PASTURE—this means actuation noise but no depth noise. For HABITAT this means considerable depth noise and reconstruction artifacts, but no actuation noise." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 361, + 545, + 385 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 361, + 545, + 385 + ], + "spans": [ + { + "bbox": [ + 305, + 361, + 545, + 385 + ], + "type": "text", + "content": "Navigation Metrics. We adopt standard object navigation metrics to measure performance:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 388, + 545, + 460 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 306, + 388, + 545, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 388, + 545, + 411 + ], + "spans": [ + { + "bbox": [ + 306, + 388, + 545, + 411 + ], + "type": "text", + "content": "- SUCCESS (SR): the fraction of episodes where the agent executes STOP within " + }, + { + "bbox": [ + 306, + 388, + 545, + 411 + ], + "type": "inline_equation", + "content": "1.0\\mathrm{m}" + }, + { + "bbox": [ + 306, + 388, + 545, + 411 + ], + "type": "text", + "content": " of the target object." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 413, + 545, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 413, + 545, + 460 + ], + "spans": [ + { + "bbox": [ + 306, + 413, + 545, + 460 + ], + "type": "text", + "content": "- Success weighted by inverse path length (SPL): Success weighted by the oracle shortest path length and normalized by the actual path length [5]. This metric points to the success efficiency of the agent." + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 304, + 466, + 545, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 466, + 545, + 502 + ], + "spans": [ + { + "bbox": [ + 304, + 466, + 545, + 502 + ], + "type": "text", + "content": "In ROBOTHOR and PASTURE, the target must additionally be visible for the episode to be a success, which this is not the case in HABITAT—as specified in their 2021 challenge." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 505, + 545, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 505, + 545, + 552 + ], + "spans": [ + { + "bbox": [ + 304, + 505, + 545, + 552 + ], + "type": "text", + "content": "Embodiment. The agent is a LoCoBot [29]. All agents have discrete actions: {MOVEFORWARD, ROTATERIGHT, ROTATELEFT, STOP}. The move action advances the agent by " + }, + { + "bbox": [ + 304, + 505, + 545, + 552 + ], + "type": "inline_equation", + "content": "0.25\\mathrm{m}" + }, + { + "bbox": [ + 304, + 505, + 545, + 552 + ], + "type": "text", + "content": ", while rotation actions pivot the camera by " + }, + { + "bbox": [ + 304, + 505, + 545, + 552 + ], + "type": "inline_equation", + "content": "30^{\\circ}" + }, + { + "bbox": [ + 304, + 505, + 545, + 552 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 555, + 545, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 555, + 545, + 662 + ], + "spans": [ + { + "bbox": [ + 304, + 555, + 545, + 662 + ], + "type": "text", + "content": "CoW Baselines. For exploration we consider policies presented in Sec. 4.2: FBE heuristic exploration, learnable exploration optimized on HABITAT (MP3D) train scenes, and learned exploration optimized on ROBOTHOR train scenes. Learned exploration requires training in simulation—which is counter to our zero-shot goals; nonetheless, we ablate these explorers to contextualize their performance within the CoW framework. FBE is the default CoW exploration strategy." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 663, + 466, + 674 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 663, + 466, + 674 + ], + "spans": [ + { + "bbox": [ + 317, + 663, + 466, + 674 + ], + "type": "text", + "content": "For object localization, we consider:" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 306, + 677, + 522, + 716 + ], + "type": "list", + "angle": 0, + "index": 24, + "blocks": [ + { + "bbox": [ + 306, + 677, + 522, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 677, + 522, + 689 + ], + "spans": [ + { + "bbox": [ + 306, + 677, + 522, + 689 + ], + "type": "text", + "content": "- CLIP with " + }, + { + "bbox": [ + 306, + 677, + 522, + 689 + ], + "type": "inline_equation", + "content": "k = 9" + }, + { + "bbox": [ + 306, + 677, + 522, + 689 + ], + "type": "text", + "content": " referring expressions (CLIP-Ref.)" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 306, + 691, + 474, + 703 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 691, + 474, + 703 + ], + "spans": [ + { + "bbox": [ + 306, + 691, + 474, + 703 + ], + "type": "text", + "content": "- CLIP with " + }, + { + "bbox": [ + 306, + 691, + 474, + 703 + ], + "type": "inline_equation", + "content": "k = 9" + }, + { + "bbox": [ + 306, + 691, + 474, + 703 + ], + "type": "text", + "content": " patches (CLIP-Patch)" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 306, + 705, + 490, + 716 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 705, + 490, + 716 + ], + "spans": [ + { + "bbox": [ + 306, + 705, + 490, + 716 + ], + "type": "text", + "content": "- CLIP with gradient relevance (CLIP-Grad.)" + } + ] + } + ], + "index": 23 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "23175" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 52, + 70, + 541, + 262 + ], + "blocks": [ + { + "bbox": [ + 52, + 70, + 541, + 262 + ], + "lines": [ + { + "bbox": [ + 52, + 70, + 541, + 262 + ], + "spans": [ + { + "bbox": [ + 52, + 70, + 541, + 262 + ], + "type": "table", + "html": "
IDCoW breedsPASTUREROBOTHOR
Uncom.Appear.SpaceAppear. distract SRSpace distract SRHid.Hid. distract SRAvg.
LocalizerArch.SRSRSRSRSRSRSRSPLSRSPLSR
▲CLIP-Ref.B/323.60.61.70.61.72.22.50.91.81.01.8
■CLIP-Ref.B/161.42.82.83.13.31.71.91.72.42.12.7
▲CLIP-PatchB/3218.113.313.38.610.817.517.89.014.210.620.3
■CLIP-PatchB/1610.611.47.810.88.116.415.67.711.59.715.7
▲CLIP-Grad.B/3216.111.911.79.710.314.416.19.212.99.715.2
■CLIP-Grad.B/168.110.88.68.66.711.111.46.79.38.611.6
◇MDETRB33.17.25.06.94.78.18.95.46.38.49.9
▲OWLB/3232.826.419.419.416.119.214.412.621.116.926.7
■OWLB/1631.926.918.919.414.718.115.812.620.817.227.5
ProcTHOR fine-tune (supervised) [19]n/an/an/an/an/an/an/an/an/a27.466.4
", + "image_path": "09f95fb87e8e3586cd849b725aebe4948ce6be28c2b80c8886917a367e39b204.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 267, + 547, + 354 + ], + "lines": [ + { + "bbox": [ + 46, + 267, + 547, + 354 + ], + "spans": [ + { + "bbox": [ + 46, + 267, + 547, + 354 + ], + "type": "text", + "content": "Table 1. Benchmarking CoWs on PASTURE for L-ZSON. On PASTURE we identify several key takeaways. (1) Average success on PASTURE is lower than on ROBOTHOR; however, CoWs are surprisingly good at finding uncommon objects (Uncom.), often finding them at higher rates than more common ROBOTHOR objects. (2) Comparing square (■) vs. triangle (▲) IDs, we see that architectures (Arch.) using more compute (i.e., ViT-B/16) often perform comparably or worse than their competitors (i.e., ViT-B/32). This is especially true for CLIP [61] models (indicated in pink, orange, and purple). (3) Blue OWL-ViT [49] models perform best. (4) PASTURE tasks with distractor objects (distract) hurt performance and natural language specification is not sufficient to mitigate against the added difficulties in these tasks. (5) A supervised baseline shown in gray significantly outperforms CoWs on ROBOTHOR; however, it is unable to support PASTURE tasks out-of-the-box." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 47, + 361, + 220, + 389 + ], + "type": "list", + "angle": 0, + "index": 4, + "blocks": [ + { + "bbox": [ + 47, + 361, + 220, + 374 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 361, + 220, + 374 + ], + "spans": [ + { + "bbox": [ + 47, + 361, + 220, + 374 + ], + "type": "text", + "content": "- MDETR segmentation model (MDETR)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 378, + 167, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 378, + 167, + 389 + ], + "spans": [ + { + "bbox": [ + 47, + 378, + 167, + 389 + ], + "type": "text", + "content": "- OWL-ViT detector (OWL)" + } + ] + } + ], + "index": 3 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 46, + 392, + 287, + 452 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 392, + 287, + 452 + ], + "spans": [ + { + "bbox": [ + 46, + 392, + 287, + 452 + ], + "type": "text", + "content": "Descriptions of these models are in Sec. 4.3 and additional details are in Appx. C. All models are open-vocabulary. No models are fine-tuned on navigation, and hence we consider their inference zero-shot on our tasks.2 We also consider various backbone architectures:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 456, + 287, + 522 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 47, + 456, + 241, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 456, + 241, + 468 + ], + "spans": [ + { + "bbox": [ + 47, + 456, + 241, + 468 + ], + "type": "text", + "content": "- A vision transformer [24], ViT-B/32 (" + }, + { + "bbox": [ + 47, + 456, + 241, + 468 + ], + "type": "inline_equation", + "content": "\\triangle" + }, + { + "bbox": [ + 47, + 456, + 241, + 468 + ], + "type": "text", + "content": " B/32)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 471, + 287, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 471, + 287, + 495 + ], + "spans": [ + { + "bbox": [ + 47, + 471, + 287, + 495 + ], + "type": "text", + "content": "- ViT-B/16 (■ B/16), which uses a smaller patch size of 16x16 and hence more compute." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 498, + 287, + 522 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 498, + 287, + 522 + ], + "spans": [ + { + "bbox": [ + 47, + 498, + 287, + 522 + ], + "type": "text", + "content": "- EfficientNet B3 (" + }, + { + "bbox": [ + 47, + 498, + 287, + 522 + ], + "type": "inline_equation", + "content": "\\spadesuit" + }, + { + "bbox": [ + 47, + 498, + 287, + 522 + ], + "type": "text", + "content": "B3), which is convolutional and similar in compute requirements to a ViT/B32." + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 46, + 526, + 287, + 658 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 526, + 287, + 658 + ], + "spans": [ + { + "bbox": [ + 46, + 526, + 287, + 658 + ], + "type": "text", + "content": "For every model we evaluate with post-processing as the default setting, where only the center pixel of detections is registered in the top-down CoW map. Recall, this is a sensible strategy as only some part of the object needs to be found for an episode to be successful. We find that this decision improves performance on our best performing models from 0.1 to 6.0 percentage points on ROBOTHOR SUCCESS. For a full comparison between models with and without post-processing see Appx. F. For details on hyperparameters, learned agents, object localization threshold tuning, and CLIP prompt-tuning, see Appx. C, G." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 661, + 287, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 661, + 287, + 685 + ], + "spans": [ + { + "bbox": [ + 47, + 661, + 287, + 685 + ], + "type": "text", + "content": "End-to-end learnable baselines. We also compare against methods that are trained in simulation for millions of steps:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 361, + 545, + 487 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 306, + 361, + 545, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 361, + 545, + 422 + ], + "spans": [ + { + "bbox": [ + 306, + 361, + 545, + 422 + ], + "type": "text", + "content": "- EmbCLIP-ZSON [38] trains on eight ROBOTHR categories, using CLIP language embeddings to specify the goal objects. At test time, the model is evaluated on four held-out object categories, which are also specified CLIP language embeddings for the target category names." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 426, + 545, + 487 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 426, + 545, + 487 + ], + "spans": [ + { + "bbox": [ + 306, + 426, + 545, + 487 + ], + "type": "text", + "content": "- SemanticNav-ZSON [46] trains models separately, one for each dataset, for image goal navigation. Image goals are specified with CLIP visual embeddings. At test time, image goals are swapped for CLIP language embeddings for the object goals. We compare to the MP3D model." + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 305, + 490, + 545, + 526 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 490, + 545, + 526 + ], + "spans": [ + { + "bbox": [ + 305, + 490, + 545, + 526 + ], + "type": "text", + "content": "Both EmbCLIP-ZSON and SemanticNav-ZSON leverage multi-modal CLIP visual and language embeddings in learnable frameworks that require simulation training." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 536, + 415, + 547 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 536, + 415, + 547 + ], + "spans": [ + { + "bbox": [ + 306, + 536, + 415, + 547 + ], + "type": "text", + "content": "6.2. CoWs on PASTURE" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 305, + 554, + 545, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 554, + 545, + 590 + ], + "spans": [ + { + "bbox": [ + 305, + 554, + 545, + 590 + ], + "type": "text", + "content": "Tab. 1 shows our main results of CoWs evaluated on PASTURE. For category-level results see Appx. H. We now discuss several salient questions." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "text", + "content": "How well can CoWs find common objects vs. uncommon objects? Comparing ROBOTHOR and uncommon (Uncom.) PASTURE success rate (SR) in Tab. 1—first and last columns—we notice that CoWs often find uncommon objects at higher rates than common ROBOTHOR objects (e.g., by " + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\sim 6" + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "text", + "content": " percentage points SUCCESS for the OWL ViT-B/32 CoW (" + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\triangle" + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "text", + "content": ")). We hypothesize that though uncommon objects are less prevalent in daily life, they are still represented in open-vocabulary datasets and hence recognizable for the object localization modules. We further ex" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 47, + 693, + 287, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 693, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 693, + 287, + 713 + ], + "type": "text", + "content": "2The claim is not that these models have never seen any synthetic data in their large-scale training sets, only that they are not trained to navigate." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "23176" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 87, + 174, + 201 + ], + "blocks": [ + { + "bbox": [ + 80, + 76, + 162, + 84 + ], + "lines": [ + { + "bbox": [ + 80, + 76, + 162, + 84 + ], + "spans": [ + { + "bbox": [ + 80, + 76, + 162, + 84 + ], + "type": "text", + "content": "(a) Attribute object navigation" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 56, + 87, + 174, + 201 + ], + "lines": [ + { + "bbox": [ + 56, + 87, + 174, + 201 + ], + "spans": [ + { + "bbox": [ + 56, + 87, + 174, + 201 + ], + "type": "image", + "image_path": "c08efe085810ec7489087ad49b99ca9d5562e6dc03738f53c5d2af944044b086.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 178, + 89, + 293, + 200 + ], + "blocks": [ + { + "bbox": [ + 200, + 76, + 282, + 91 + ], + "lines": [ + { + "bbox": [ + 200, + 76, + 282, + 91 + ], + "spans": [ + { + "bbox": [ + 200, + 76, + 282, + 91 + ], + "type": "text", + "content": "(b) Attribute object navigation with distractors" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 178, + 89, + 293, + 200 + ], + "lines": [ + { + "bbox": [ + 178, + 89, + 293, + 200 + ], + "spans": [ + { + "bbox": [ + 178, + 89, + 293, + 200 + ], + "type": "image", + "image_path": "e5c82a109c048ad9062ef6ee47c8ab41ed2c0d4d029e09f19abae3c746c5c1fe.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 204, + 545, + 270 + ], + "lines": [ + { + "bbox": [ + 46, + 204, + 545, + 270 + ], + "spans": [ + { + "bbox": [ + 46, + 204, + 545, + 270 + ], + "type": "text", + "content": "Figure 5. PASTURE object navigation with descriptions. In general object navigation with descriptions is more challenging than the ROBOTOR object navigation task, as indicated by trend lines lying below the " + }, + { + "bbox": [ + 46, + 204, + 545, + 270 + ], + "type": "inline_equation", + "content": "y = x" + }, + { + "bbox": [ + 46, + 204, + 545, + 270 + ], + "type": "text", + "content": " line. (a) Appearance descriptions are more helpful than spatial descriptions. (b) Performance further drops when distractor objects are introduced to the environment. However, CoWs are still able to make better use of appearance description than spatial descriptions. (c) Models in the lower success regime (<15% ROBOTOR SUCCESS) perform comparably on finding hidden objects. However, this trend plateaus for higher success models. (d) Trends are similar when distractor objects are introduced for hidden object navigation." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 296, + 89, + 413, + 200 + ], + "blocks": [ + { + "bbox": [ + 323, + 76, + 400, + 85 + ], + "lines": [ + { + "bbox": [ + 323, + 76, + 400, + 85 + ], + "spans": [ + { + "bbox": [ + 323, + 76, + 400, + 85 + ], + "type": "text", + "content": "(c) Hidden object navigation" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 296, + 89, + 413, + 200 + ], + "lines": [ + { + "bbox": [ + 296, + 89, + 413, + 200 + ], + "spans": [ + { + "bbox": [ + 296, + 89, + 413, + 200 + ], + "type": "image", + "image_path": "e8ea830b0c8da8f02a872487755d76c6447c60a7110e3b42e1f8f0862616823c.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 417, + 89, + 533, + 200 + ], + "blocks": [ + { + "bbox": [ + 443, + 76, + 520, + 91 + ], + "lines": [ + { + "bbox": [ + 443, + 76, + 520, + 91 + ], + "spans": [ + { + "bbox": [ + 443, + 76, + 520, + 91 + ], + "type": "text", + "content": "(d) Hidden object navigation with distractors" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 417, + 89, + 533, + 200 + ], + "lines": [ + { + "bbox": [ + 417, + 89, + 533, + 200 + ], + "spans": [ + { + "bbox": [ + 417, + 89, + 533, + 200 + ], + "type": "image", + "image_path": "cb2d52d7231379da66db7359b05d1718824a1723620903f966b7f16b8b2de868.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 283, + 286, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 283, + 286, + 354 + ], + "spans": [ + { + "bbox": [ + 46, + 283, + 286, + 354 + ], + "type": "text", + "content": "plore this hypothesis in Appx. E by visualizing CLIP retrieval results on LAION-5B [68] for the uncommon object categories. The relatively high performance on uncommon objects speaks to the flexibility of CoW baselines and their ability to inherit desirable properties from the open-vocabulary models that they are constructed from." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 357, + 286, + 464 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 357, + 286, + 464 + ], + "spans": [ + { + "bbox": [ + 46, + 357, + 286, + 464 + ], + "type": "text", + "content": "Can CoWs utilize appearance and spatial descriptions? Looking at Fig. 5 (a) we see that neither appearance nor spatial descriptions improve CoW performance compared to their ROBOTHOR baseline performance (i.e., most points lie under the " + }, + { + "bbox": [ + 46, + 357, + 286, + 464 + ], + "type": "inline_equation", + "content": "y = x" + }, + { + "bbox": [ + 46, + 357, + 286, + 464 + ], + "type": "text", + "content": " line). However, CoW is able to take better advantage of appearance descriptions than spatial descriptions. These results motivate future investigation on open-vocabulary object localization with a greater focus on textual object attributes." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 468, + 286, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 468, + 286, + 540 + ], + "spans": [ + { + "bbox": [ + 46, + 468, + 286, + 540 + ], + "type": "text", + "content": "Can CoWs find visible objects in the presence of distractors? In Fig. 5 (b) we see that CoWs experience performance degradation when compared to Fig. 5 (a). We conclude that appearance and spatial attributes added as language input are not sufficient to deal with the complexity of distractors given current open-vocabulary models." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 543, + 286, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 543, + 286, + 662 + ], + "spans": [ + { + "bbox": [ + 46, + 543, + 286, + 662 + ], + "type": "text", + "content": "Can CoWs find hidden objects? Looking at Fig. 5 (c) we notice that models in the lower success regime (less than " + }, + { + "bbox": [ + 46, + 543, + 286, + 662 + ], + "type": "inline_equation", + "content": "15\\%" + }, + { + "bbox": [ + 46, + 543, + 286, + 662 + ], + "type": "text", + "content": " SUCCESS on ROBOTHOR) are able to find hidden objects at about the same rate as ROBOTHOR objects (i.e., they lie on the " + }, + { + "bbox": [ + 46, + 543, + 286, + 662 + ], + "type": "inline_equation", + "content": "y = x" + }, + { + "bbox": [ + 46, + 543, + 286, + 662 + ], + "type": "text", + "content": " line). OWL models in the higher success regime " + }, + { + "bbox": [ + 46, + 543, + 286, + 662 + ], + "type": "inline_equation", + "content": "(>15\\%)" + }, + { + "bbox": [ + 46, + 543, + 286, + 662 + ], + "type": "text", + "content": " do not continue this trend; however, they do achieve higher absolute accuracy as seen in Tab. 1. Dealing with occlusion is a longstanding problem in computer vision, and these results provide a foundation upon which future hidden object navigation work can improve." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 46, + 665, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 286, + 713 + ], + "type": "text", + "content": "Can CoWs find hidden objects in the presence of distractors? Comparing Figs. 5 (c) and (d), we notice similar trends lines, with the best models performing worse with distractors. This suggests that distractors are less of a con" + } + ] + } + ], + "index": 13 + }, + { + "type": "table", + "bbox": [ + 309, + 281, + 541, + 340 + ], + "blocks": [ + { + "bbox": [ + 309, + 281, + 541, + 340 + ], + "lines": [ + { + "bbox": [ + 309, + 281, + 541, + 340 + ], + "spans": [ + { + "bbox": [ + 309, + 281, + 541, + 340 + ], + "type": "table", + "html": "
IDLoc.CoW breedsPASTURE (Avg.)ROBOTHOR
Arch.Exp. StrategySPLSRSPLSR
▲OWLB/32ROBOTHOR learn.10.217.313.120.9
▲OWLB/32HABITAT learn.8.619.49.820.4
▲OWLB/32FBE12.621.116.926.7
", + "image_path": "27e9702207cfe9fe9ea99b438d70b63e83f7f51d807f8aeb889883027058be5f.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "table_body" + } + ], + "index": 14 + }, + { + "type": "table", + "bbox": [ + 310, + 427, + 541, + 477 + ], + "blocks": [ + { + "bbox": [ + 305, + 346, + 545, + 422 + ], + "lines": [ + { + "bbox": [ + 305, + 346, + 545, + 422 + ], + "spans": [ + { + "bbox": [ + 305, + 346, + 545, + 422 + ], + "type": "text", + "content": "Table 2. Exploration ablation. For a fixed object localizer (OWL-ViT B/32 with post processing), we ablate over different choices of exploration policy: the FBE heuristic, agents trained in ROBOTHOR, and HABITAT (MP3D). We find that FBE outperforms learnable alternatives on both PASTURE and ROBOTHOR. HABITAT learnable model perform worst, but are not trained on any PASTURE or ROBOTHOR data." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 310, + 427, + 541, + 477 + ], + "lines": [ + { + "bbox": [ + 310, + 427, + 541, + 477 + ], + "spans": [ + { + "bbox": [ + 310, + 427, + 541, + 477 + ], + "type": "table", + "html": "
IDCoW breedsObj. PriorPASTURE Uncom.ROBOTHOR
Loc.Arch.SPLSRSPLSR
▲OWLB/32None20.532.816.826.7
▲OWLB/32GPT-3.522.236.917.027.5
", + "image_path": "21c0200d8363d30e2cc9b0663730b24c573ff41844f76c01d42cfc9936bd4b24.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "table_body" + } + ], + "index": 16 + }, + { + "bbox": [ + 305, + 482, + 545, + 527 + ], + "lines": [ + { + "bbox": [ + 305, + 482, + 545, + 527 + ], + "spans": [ + { + "bbox": [ + 305, + 482, + 545, + 527 + ], + "type": "text", + "content": "Table 3. CoW with GPT-3.5 priors. Leveraging GPT-3.5, we generate priors for objects (e.g., apples are likely to be in dining room scenes). Instead of directly searching for the target object, CoW first searches for the scenes, which boosts performance." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 540, + 545, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 540, + 545, + 613 + ], + "spans": [ + { + "bbox": [ + 304, + 540, + 545, + 613 + ], + "type": "text", + "content": "cern in the case of hidden objects than for visible object targets. In light of the fact that detection methods generally work better on larger objects, we hypothesize this effect is because distractor objects are smaller (e.g., apples, vases, basketballs) than objects used to conceal target categories (e.g., beds, sofas, etc.)." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 617, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 617, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 617, + 545, + 712 + ], + "type": "text", + "content": "What exploration method performs best? We ablate the decision to use FBE for most experiments by fixing an object localizer (OWL, B/32 " + }, + { + "bbox": [ + 304, + 617, + 545, + 712 + ], + "type": "inline_equation", + "content": "(\\triangle))" + }, + { + "bbox": [ + 304, + 617, + 545, + 712 + ], + "type": "text", + "content": " and comparing against ROBOTHR learnable exploration and HABITAT learnable exploration in Tab. 2. We notice that FBE performs best in all cases; however, learnable exploration still performs well suggesting that these models do learn useful strategies for the downstream tasks." + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "23177" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 76, + 76, + 258, + 207 + ], + "blocks": [ + { + "bbox": [ + 76, + 76, + 258, + 207 + ], + "lines": [ + { + "bbox": [ + 76, + 76, + 258, + 207 + ], + "spans": [ + { + "bbox": [ + 76, + 76, + 258, + 207 + ], + "type": "image", + "image_path": "5beab85ea8a0266fdfe3a1171921383455df55bd1b781209469696e873de385b.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 217, + 289, + 251 + ], + "lines": [ + { + "bbox": [ + 46, + 217, + 289, + 251 + ], + "spans": [ + { + "bbox": [ + 46, + 217, + 289, + 251 + ], + "type": "text", + "content": "Figure 6. Failure analysis for OWL, B/32 (▲). Exploration and object localization errors occur at similar ratios, with increased localization failures in the presence of distractors." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 261, + 287, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 261, + 287, + 346 + ], + "spans": [ + { + "bbox": [ + 46, + 261, + 287, + 346 + ], + "type": "text", + "content": "Can CoW incorporate object priors? Examining Tab. 3, we see that incorporating GPT-3.5 object-level priors improves performance on both PASTURE uncommon objects and ROBOTOR. These initial results suggest positive trends for incorporating outside knowledge into CoW. Future work may consider more sophisticated methods for injecting priors to steer navigation." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 348, + 288, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 348, + 288, + 506 + ], + "spans": [ + { + "bbox": [ + 46, + 348, + 288, + 506 + ], + "type": "text", + "content": "How do CoWs fail? We identify three high-level failure modes. (1) Exploration fail: the target is never seen. (2) Object localization fail: the target is seen but the localizer never fires. (3) Planning fail: the target is seen and the localizer fires, but planning fails due to inaccuracy in the map representation (Sec. 4.2). Looking at Fig. 6, we notice a large fraction of failures are due to exploration and object localization. This suggests CoWs may continue to improve as research in these fields progress. In Fig. 6 we also see that in cases where distractors are present a higher fraction of object localization failures occur, further suggesting that open-vocabulary models currently struggle to make full use of attribute prompts. See Appx. I for more failure analysis." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 514, + 187, + 528 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 514, + 187, + 528 + ], + "spans": [ + { + "bbox": [ + 47, + 514, + 187, + 528 + ], + "type": "text", + "content": "6.3. Comparison to Prior Art" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 533, + 287, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 533, + 287, + 593 + ], + "spans": [ + { + "bbox": [ + 46, + 533, + 287, + 593 + ], + "type": "text", + "content": "We primarily evaluate CoWs in general L-ZSON settings; however, we further evaluate CoWs on ZSON benchmarks to establish them as a strong baseline for these tasks. Recall, ZSON can be seen as a case of L-ZSON where only object goals are specified (no attributes)." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "text", + "content": "In Tab. 4, we see there exists a CoW that outperforms the end-to-end baselines in all cases except SUCCESS on HABITAT (MP3D). For instance, the CLIP-Grad., B/32 (" + }, + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\triangleleft" + }, + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "text", + "content": ") matches the SemanticNav-ZSON model on HABITAT (MP3D) SPL: 4.9 for CoW v.s. 4.8 for the competitor, while improving over EmbCLIP-ZSON ROBOTHOR by 15.6 percentage points. To contextualize this result, CoWs train for " + }, + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "inline_equation", + "content": "O" + }, + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "text", + "content": " navigation steps, while SemanticNav-ZSON and EmbCLIP-ZSON train in the target evaluation simulators for 500M and 60M steps respectively." + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 315, + 71, + 538, + 155 + ], + "blocks": [ + { + "bbox": [ + 315, + 71, + 538, + 155 + ], + "lines": [ + { + "bbox": [ + 315, + 71, + 538, + 155 + ], + "spans": [ + { + "bbox": [ + 315, + 71, + 538, + 155 + ], + "type": "table", + "html": "
IDCoW breedsHABITAT (MP3D)ROBOTHON (subset)ROBOTHON (full)Nav. training steps
Loc.Arch.SPLSRSPLSRSPLSR
▲CLIP-Grad.B/324.99.215.023.79.715.20
▲OWLB/323.77.420.832.516.926.70
EmbCLIP-ZSON [38]---8.1-14.0*60M
SemanticNav-ZSON [46]4.815.3----500M
", + "image_path": "9900d387144252b9e5c8d9ea3b0beb45ffe14a4ce3364e08d002845713818384.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 162, + 545, + 240 + ], + "lines": [ + { + "bbox": [ + 304, + 162, + 545, + 240 + ], + "spans": [ + { + "bbox": [ + 304, + 162, + 545, + 240 + ], + "type": "text", + "content": "Table 4. Comparison to prior art on existing ZSON benchmarks. CoWs are able to match or out-compete existing methods that leverage millions of steps of navigation training in the evaluation simulator. *indicates a result from prior work that includes, non-zero-shot evaluation. Specifically, only 1/4 of the evaluations are zero-shot on ROBOTOR (subset) and the remaining 3/4 on categories seen during training." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 251, + 545, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 251, + 545, + 300 + ], + "spans": [ + { + "bbox": [ + 304, + 251, + 545, + 300 + ], + "type": "text", + "content": "The superior performance of SemanticNav-ZSON in terms of MP3D SUCCESS indicates that there can be benefits to in-domain learning. Future work may consider unifying the benefits of CoW-like models and fine-tuned models." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 311, + 463, + 323 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 311, + 463, + 323 + ], + "spans": [ + { + "bbox": [ + 306, + 311, + 463, + 323 + ], + "type": "text", + "content": "7. Limitations and Conclusion" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 333, + 545, + 501 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 333, + 545, + 501 + ], + "spans": [ + { + "bbox": [ + 304, + 333, + 545, + 501 + ], + "type": "text", + "content": "Limitations. While our evaluation of CoWs on HABITAT, ROBOTHER, and PASTURE is a step towards assessing their performance in different domains, ultimately, real-world performance matters most. Hence, the biggest limitation of our work is the lack of large-scale, real-world benchmarking—which is also missing in much of the related literature. Additionally, CoW inherits the metalimitations of the object localization and exploration methods considered. For example, object localizers require tuning a confidence threshold to balance precision and recall. Finally, we do not consider different agent embodiment or continuous action spaces. This is a pertinent investigation given recent findings of Pratt et al. [59] that agent morphology can be a big determinant of downstream performance." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 503, + 545, + 612 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 503, + 545, + 612 + ], + "spans": [ + { + "bbox": [ + 304, + 503, + 545, + 612 + ], + "type": "text", + "content": "Conclusion. This paper introduces the PASTURE benchmark for language-driven zero-shot object navigation and several CLIP on Wheels baselines, translating the successes of existing zero-shot models to an embodied task. We view CoW as an instance of using open-vocabulary models, with text-based interfaces, to tackle robotics tasks in more flexible settings. We hope that the baselines and the proposed benchmark will spur the field to explore broader and more powerful forms of zero-shot embodied AI." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 615, + 545, + 710 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 615, + 545, + 710 + ], + "spans": [ + { + "bbox": [ + 304, + 615, + 545, + 710 + ], + "type": "text", + "content": "Acknowledgement. We would like to thank Jessie Chapman, Cheng Chi, Huy Ha, Zeyi Liu, Sachit Menon, and Sarah Pratt for valuable feedback. This work was supported in part by NSF CMMI-2037101, NSF IIS-2132519, and an Amazon Research Award. SYG is supported by a NSF Graduate Research Fellowship. The views and conclusions contained herein are those of the authors and should not be interpreted as necessarily representing the official policies, either expressed or implied, of the sponsors." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "23178" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 133 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 133 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 133 + ], + "type": "text", + "content": "[1] Ziad Al-Halah, Santhosh K. Ramakrishnan, and Kristen Grauman. Zero experience required: Plug & play modular transfer learning for semantic visual navigation. arXiv, 2022. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 136, + 288, + 190 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 136, + 288, + 190 + ], + "spans": [ + { + "bbox": [ + 53, + 136, + 288, + 190 + ], + "type": "text", + "content": "[2] Peter Anderson, Qi Wu, Damien Teney, Jake Bruce, Mark Johnson, Niko Sunderhauf, Ian Reid, Stephen Gould, and Anton Van Den Hengel. Vision-and-language navigation: Interpreting visually-grounded navigation instructions in real environments. CVPR, 2018. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 191, + 288, + 223 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 191, + 288, + 223 + ], + "spans": [ + { + "bbox": [ + 53, + 191, + 288, + 223 + ], + "type": "text", + "content": "[3] Donghyeon Baek, Youngmin Oh, and Bumsub Ham. Exploiting a joint embedding space for generalized zero-shot semantic segmentation. ICCV, 2021. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 224, + 288, + 256 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 224, + 288, + 256 + ], + "spans": [ + { + "bbox": [ + 53, + 224, + 288, + 256 + ], + "type": "text", + "content": "[4] Ankan Bansal, Karan Sikka, Gaurav Sharma, Rama Chellappa, and Ajay Divakaran. Zero-shot object detection. ECCV, 2018. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 258, + 287, + 312 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 258, + 287, + 312 + ], + "spans": [ + { + "bbox": [ + 53, + 258, + 287, + 312 + ], + "type": "text", + "content": "[5] Dhruv Batra, Aaron Gokaslan, Aniruddha Kembhavi, Oleksandr Maksymets, Roozbeh Mottaghi, Manolis Savva, Alexander Toshev, and Erik Wijmans. Objectnav revisited: On evaluation of embodied agents navigating to objects. arXiv, 2020. 1, 3, 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 313, + 288, + 345 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 313, + 288, + 345 + ], + "spans": [ + { + "bbox": [ + 53, + 313, + 288, + 345 + ], + "type": "text", + "content": "[6] Maxime Bucher, Tuan-Hung VU, Matthieu Cord, and Patrick Pérez. Zero-shot semantic segmentation. NeurIPS, 2019. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 346, + 287, + 378 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 346, + 287, + 378 + ], + "spans": [ + { + "bbox": [ + 53, + 346, + 287, + 378 + ], + "type": "text", + "content": "[7] Yuri Burda, Harrison Edwards, Amos J. Storkey, and Oleg Klimov. Exploration by random network distillation. *ICLR*, 2018. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 380, + 288, + 412 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 380, + 288, + 412 + ], + "spans": [ + { + "bbox": [ + 53, + 380, + 288, + 412 + ], + "type": "text", + "content": "[8] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. ECCV, 2020. 4, 14" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 414, + 288, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 414, + 288, + 445 + ], + "spans": [ + { + "bbox": [ + 53, + 414, + 288, + 445 + ], + "type": "text", + "content": "[9] Matthew Chang, Arjun Gupta, and Saurabh Gupta. Semantic visual navigation by watching youtube videos. NeurIPS, 2020. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 447, + 287, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 447, + 287, + 479 + ], + "spans": [ + { + "bbox": [ + 48, + 447, + 287, + 479 + ], + "type": "text", + "content": "[10] Devendra Singh Chaplot, Dhiraj Gandhi, Abhinav Gupta, and Ruslan Salakhutdinov. Object goal navigation using goal-oriented semantic exploration. NeurIPS, 2020. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 480, + 287, + 512 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 480, + 287, + 512 + ], + "spans": [ + { + "bbox": [ + 48, + 480, + 287, + 512 + ], + "type": "text", + "content": "[11] Devendra Singh Chaplot, Dhiraj Gandhi, Saurabh Gupta, Abhinav Kumar Gupta, and Ruslan Salakhutdinov. Learning to explore using active neural slam. ICLR, 2020. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 514, + 287, + 546 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 514, + 287, + 546 + ], + "spans": [ + { + "bbox": [ + 48, + 514, + 287, + 546 + ], + "type": "text", + "content": "[12] Prithvijit Chattopadhyay, Judy Hoffman, Roozbeh Mottaghi, and Aniruddha Kembhavi. Robustnav: Towards benchmarking robustness in embodied navigation. ICCV, 2021. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 547, + 287, + 579 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 547, + 287, + 579 + ], + "spans": [ + { + "bbox": [ + 48, + 547, + 287, + 579 + ], + "type": "text", + "content": "[13] Hila Chefer, Shir Gur, and Lior Wolf. Transformer interpretability beyond attention visualization. CVPR, 2021. 4, 13, 14" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 581, + 287, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 581, + 287, + 602 + ], + "spans": [ + { + "bbox": [ + 48, + 581, + 287, + 602 + ], + "type": "text", + "content": "[14] Tao Chen, Saurabh Gupta, and Abhinav Gupta. Learning exploration policies for navigation. *ICLR*, 2019. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 603, + 287, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 603, + 287, + 645 + ], + "spans": [ + { + "bbox": [ + 48, + 603, + 287, + 645 + ], + "type": "text", + "content": "[15] Jiaxin Cheng, Soumyaroop Nandi, Prem Natarajan, and Wael Abd-Almageed. Sign: Spatial-information incorporated generative network for generalized zero-shot semantic segmentation. ICCV, 2021. 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 48, + 647, + 287, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 647, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 647, + 287, + 689 + ], + "type": "text", + "content": "[16] Kyunghyun Cho, Bart van Merrienboer, Dzmitry Bahdanau, and Yoshua Bengio. On the properties of neural machine translation: Encoder-decoder approaches. In SSST@EMNLP, 2014. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "type": "text", + "content": "[17] Andrew J Davison and David W Murray. Mobile robot localisation using active vision. ECCV, 1998. 2" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 34, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 137 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 137 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 137 + ], + "type": "text", + "content": "[18] Matt Deitke, Winson Han, Alvaro Herrasti, Aniruddha Kembhavi, Eric Kolve, Roozbeh Mottaghi, Jordi Salvador, Dustin Schwenk, Eli VanderBilt, Matthew Wallingford, Luca Weihs, Mark Yatskar, and Ali Farhadi. Robothor: An open simulation-to-real embodied ai platform. CVPR, 2020. 1, 3, 5" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 139, + 545, + 193 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 139, + 545, + 193 + ], + "spans": [ + { + "bbox": [ + 307, + 139, + 545, + 193 + ], + "type": "text", + "content": "[19] Matt Deitke, Eli VanderBilt, Alvaro Herrasti, Luca Weihs, Jordi Salvador, Kiana Ehsani, Winson Han, Eric Kolve, Ali Farhadi, Aniruddha Kembhavi, and Roozbeh Mottaghi. Proctor: Large-scale embodied ai using procedural generation. NeurIPS, 2022. 2, 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 194, + 545, + 225 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 194, + 545, + 225 + ], + "spans": [ + { + "bbox": [ + 307, + 194, + 545, + 225 + ], + "type": "text", + "content": "[20] Frank Dellaert, Dieter Fox, Wolfram Burgard, and Sebastian Thrun. Monte carlo localization for mobile robots. ICRA, 1999. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 228, + 545, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 228, + 545, + 259 + ], + "spans": [ + { + "bbox": [ + 307, + 228, + 545, + 259 + ], + "type": "text", + "content": "[21] Berkan Demirel, Ramazan Gokberk Cinbis, and Nazli Ikizler-Cinbis. Zero-shot object detection by hybrid region embedding. BMVC, 2018. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 261, + 545, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 261, + 545, + 293 + ], + "spans": [ + { + "bbox": [ + 307, + 261, + 545, + 293 + ], + "type": "text", + "content": "[22] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, K. Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. CVPR, 2009. 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 294, + 545, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 294, + 545, + 348 + ], + "spans": [ + { + "bbox": [ + 307, + 294, + 545, + 348 + ], + "type": "text", + "content": "[23] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv, 2020. 13" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 350, + 545, + 414 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 350, + 545, + 414 + ], + "spans": [ + { + "bbox": [ + 307, + 350, + 545, + 414 + ], + "type": "text", + "content": "[24] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. ICLR, 2020. 4, 6" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 415, + 545, + 447 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 415, + 545, + 447 + ], + "spans": [ + { + "bbox": [ + 307, + 415, + 545, + 447 + ], + "type": "text", + "content": "[25] Samir Yitzhak Gadre, Kiana Ehsani, Shuran Song, and Roozbeh Mottaghi. Continuous scene representations for embodied ai. CVPR, 2022. 2, 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 449, + 545, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 449, + 545, + 480 + ], + "spans": [ + { + "bbox": [ + 307, + 449, + 545, + 480 + ], + "type": "text", + "content": "[26] Daniel Gordon, Abhishek Kadian, Devi Parikh, Judy Hoffman, and Dhruv Batra. Splitnet: Sim2sim and task2task transfer for embodied visual navigation. CVPR, 2019. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 482, + 545, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 482, + 545, + 514 + ], + "spans": [ + { + "bbox": [ + 307, + 482, + 545, + 514 + ], + "type": "text", + "content": "[27] Xiuye Gu, Tsung-Yi Lin, Weicheng Kuo, and Yin Cui. Zeroshot detection via vision and language knowledge distillation. arXiv, 2021. 1" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 515, + 545, + 546 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 515, + 545, + 546 + ], + "spans": [ + { + "bbox": [ + 307, + 515, + 545, + 546 + ], + "type": "text", + "content": "[28] Agrim Gupta, Piotr Dólar, and Ross B. Girshick. Lvis: A dataset for large vocabulary instance segmentation. CVPR, 2019. 1" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 307, + 548, + 545, + 591 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 548, + 545, + 591 + ], + "spans": [ + { + "bbox": [ + 307, + 548, + 545, + 591 + ], + "type": "text", + "content": "[29] Abhinav Kumar Gupta, Adithyavairavan Murali, Dhiraj Gandhi, and Terrel Pinto. Robot learning in homes: Improving generalization and reducing dataset bias. NeurIPS, 2018. 5" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 307, + 593, + 545, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 593, + 545, + 634 + ], + "spans": [ + { + "bbox": [ + 307, + 593, + 545, + 634 + ], + "type": "text", + "content": "[30] Saurabh Gupta, James Davidson, Sergey Levine, Rahul Sukthankar, and Jitendra Malik. Cognitive mapping and planning for visual navigation. CVPR, pages 2616-2625, 2017. 2" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 307, + 636, + 545, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 636, + 545, + 669 + ], + "spans": [ + { + "bbox": [ + 307, + 636, + 545, + 669 + ], + "type": "text", + "content": "[31] Meera Hahn, Devendra Singh Chaplot, and Shubham Tulsiani. No rl, no simulation: Learning to navigate without navigating. NeurIPS, 2021. 2" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 307, + 670, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 670, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 307, + 670, + 545, + 712 + ], + "type": "text", + "content": "[32] Peter Henry, Michael Krainin, Evan Herbst, Xiaofeng Ren, and Dieter Fox. RGB-d mapping: Using depth cameras for dense 3d modeling of indoor environments. Experimental robotics, 2014. 2" + } + ] + } + ], + "index": 33 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "23179" + } + ] + } + ], + "index": 35 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 105 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 105 + ], + "type": "text", + "content": "[33] Ping Hu, Stan Sclaroff, and Kate Saenko. Uncertainty-aware learning for zero-shot semantic segmentation. NeurIPS, 2020. 1" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 107, + 287, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 107, + 287, + 140 + ], + "spans": [ + { + "bbox": [ + 49, + 107, + 287, + 140 + ], + "type": "text", + "content": "[34] Chenguang Huang, Oier Mees, Andy Zeng, and Wolfram Burgard. Visual language maps for robot navigation. arXiv preprint arXiv:2210.05714, 2022. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 141, + 287, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 141, + 287, + 185 + ], + "spans": [ + { + "bbox": [ + 49, + 141, + 287, + 185 + ], + "type": "text", + "content": "[35] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc V Le, Yunhsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. ICML, 2021. 1, 13" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 186, + 287, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 186, + 287, + 228 + ], + "spans": [ + { + "bbox": [ + 49, + 186, + 287, + 228 + ], + "type": "text", + "content": "[36] Aishwarya Kamath, Mannat Singh, Yann LeCun, Ishan Misra, Gabriel Synnaeve, and Nicolas Carion. Mdetr - modulated detection for end-to-end multi-modal understanding. ICCV, 2021. 1, 4, 13, 14" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 49, + 231, + 286, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 231, + 286, + 262 + ], + "spans": [ + { + "bbox": [ + 49, + 231, + 286, + 262 + ], + "type": "text", + "content": "[37] Naoki Kato, Toshihiko Yamasaki, and Kiyoharu Aizawa. Zero-shot semantic segmentation via variational mapping. ICCV-W, 2019. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 49, + 264, + 287, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 264, + 287, + 297 + ], + "spans": [ + { + "bbox": [ + 49, + 264, + 287, + 297 + ], + "type": "text", + "content": "[38] Apoorv Khandelwal, Luca Weihs, Roozbeh Mottaghi, and Aniruddha Kembhavi. Simple but effective: Clip embeddings for embodied ai. arXiv, 2021. 1, 2, 3, 6, 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 49, + 299, + 287, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 299, + 287, + 342 + ], + "spans": [ + { + "bbox": [ + 49, + 299, + 287, + 342 + ], + "type": "text", + "content": "[39] Jacob Krantz, Erik Wijmans, Arjun Majumdar, Dhruv Batra, and Stefan Lee. Beyond the nav-graph: Vision-and-language navigation in continuous environments. In European Conference on Computer Vision, pages 104-120. Springer, 2020. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 49, + 344, + 287, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 344, + 287, + 387 + ], + "spans": [ + { + "bbox": [ + 49, + 344, + 287, + 387 + ], + "type": "text", + "content": "[40] Alexander Ku, Peter Anderson, Roma Patel, Eugene Ie, and Jason Baldridge. Room-across-room: Multilingual vision-and-language navigation with dense spatiotemporal grounding. arXiv preprint arXiv:2010.07954, 2020. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 49, + 388, + 287, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 388, + 287, + 430 + ], + "spans": [ + { + "bbox": [ + 49, + 388, + 287, + 430 + ], + "type": "text", + "content": "[41] Benjamin Kuipers and Yung-Tai Byun. A robot exploration and mapping strategy based on a semantic hierarchy of spatial representations. Robotics and autonomous systems, 1991. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 49, + 433, + 287, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 433, + 287, + 464 + ], + "spans": [ + { + "bbox": [ + 49, + 433, + 287, + 464 + ], + "type": "text", + "content": "[42] Xinghang Li, Di Guo, Huaping Liu, and Fuchun Sun. Revece: Remote embodied visual referring expression in continuous environment. RA-L, 2022. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 49, + 467, + 287, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 467, + 287, + 499 + ], + "spans": [ + { + "bbox": [ + 49, + 467, + 287, + 499 + ], + "type": "text", + "content": "[43] Zhihui Li, Lina Yao, Xiaqin Zhang, Xianzhi Wang, Salil S. Kanhere, and Huaxiang Zhang. Zero-shot object detection with textual descriptions. AAAI, 2019. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 49, + 501, + 287, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 501, + 287, + 533 + ], + "spans": [ + { + "bbox": [ + 49, + 501, + 287, + 533 + ], + "type": "text", + "content": "[44] Yiqing Liang, Boyuan Chen, and Shuran Song. Sscnav: Confidence-aware semantic scene completion for visual semantic navigation. ICRA, 2021. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 49, + 535, + 287, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 535, + 287, + 578 + ], + "spans": [ + { + "bbox": [ + 49, + 535, + 287, + 578 + ], + "type": "text", + "content": "[45] Tsung-Yi Lin, Michael Maire, Serge Belongie, Lubomir Bourdev, Ross Girshick, James Hays, Pietro Perona, Deva Ramanan, C. Lawrence Zitnick, and Piotr Dólar. Microsoft coco: Common objects in context. ECCV, 2014. 1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 49, + 580, + 287, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 580, + 287, + 624 + ], + "spans": [ + { + "bbox": [ + 49, + 580, + 287, + 624 + ], + "type": "text", + "content": "[46] Arjun Majumdar, Gunjan Aggarwal, Bhavika Devnani, Judy Hoffman, and Dhruv Batra. Zson: Zero-shot object-goal navigation using multimodal goal embeddings. arXiv preprint arXiv:2206.12403, 2022. 1, 2, 3, 6, 8" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 49, + 624, + 287, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 624, + 287, + 668 + ], + "spans": [ + { + "bbox": [ + 49, + 624, + 287, + 668 + ], + "type": "text", + "content": "[47] Qiaomei Mao, Chong Wang, Sheng Yu, Ye Zheng, and Yuqi Li. Zero-shot object detection with attributes-based category similarity. IEEE Transactions on Circuits and Systems II: Express Briefs, 2020. 1" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 49, + 670, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 670, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 49, + 670, + 287, + 713 + ], + "type": "text", + "content": "[48] Lina Mezghani, Sainbayar Sukhbaatar, Thibaut Lavril, Oleksandr Maksymets, Dhruv Batra, Piotr Bojanowski, and Alahari Karteek. Memory-augmented reinforcement learning for image-goal navigation. arXiv, 2021. 2" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 32, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 138 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 138 + ], + "type": "text", + "content": "[49] Matthias Minderer, Alexey A. Gritsenko, Austin Stone, Maxim Neumann, Dirk Weissenborn, Alexey Dosovitskiy, Aravindh Mahendran, Anurag Arnab, Mostafa Dehghani, Zhuoran Shen, Xiao Wang, Xiaohua Zhai, Thomas Kipf, and Neil Houlsby. Simple open-vocabulary object detection with vision transformers. ECCV, 2022. 1, 4, 6, 13, 14" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 140, + 545, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 140, + 545, + 184 + ], + "spans": [ + { + "bbox": [ + 307, + 140, + 545, + 184 + ], + "type": "text", + "content": "[50] Grégoire Montavon, Sebastian Lapuschkin, Alexander Binder, Wojciech Samek, and Klaus-Robert Müller. Explaining nonlinear classification decisions with deep taylor decomposition. Pattern recognition, 65:211-222, 2017. 13" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 186, + 545, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 186, + 545, + 206 + ], + "spans": [ + { + "bbox": [ + 308, + 186, + 545, + 206 + ], + "type": "text", + "content": "[51] Hans Moravec and Alberto Elfes. High resolution maps from wide angle sonar. ICRA, 1985. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 208, + 545, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 208, + 545, + 262 + ], + "spans": [ + { + "bbox": [ + 308, + 208, + 545, + 262 + ], + "type": "text", + "content": "[52] Richard A Newcombe, Shahram Izadi, Otmar Hilliges, David Molyneaux, David Kim, Andrew J Davison, Pushmeet Kohi, Jamie Shotton, Steve Hodges, and Andrew Fitzgibbon. Kinectfusion: Real-time dense surface mapping and tracking. ISMAR, 2011. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 264, + 545, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 264, + 545, + 296 + ], + "spans": [ + { + "bbox": [ + 308, + 264, + 545, + 296 + ], + "type": "text", + "content": "[53] Lachlan Nicholson, Michael Milford, and Niko Sunderhauf. Quadricslam: Dual quadrics from object detections as landmarks in object-oriented slam. RA-L, 2019. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 298, + 545, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 298, + 545, + 319 + ], + "spans": [ + { + "bbox": [ + 308, + 298, + 545, + 319 + ], + "type": "text", + "content": "[54] Clark F Olson and Larry H Matthies. Maximum likelihood rover localization by matching range maps. ICRA, 1998. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 321, + 545, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 321, + 545, + 374 + ], + "spans": [ + { + "bbox": [ + 308, + 321, + 545, + 374 + ], + "type": "text", + "content": "[55] Long Ouyang, Jeff Wu, Xu Jiang, Diogo Almeida, Carroll L Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, et al. Training language models to follow instructions with human feedback. NeurIPS, 2022. 4, 15" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 376, + 545, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 376, + 545, + 409 + ], + "spans": [ + { + "bbox": [ + 308, + 376, + 545, + 409 + ], + "type": "text", + "content": "[56] Simone Parisi, Victoria Dean, Deepak Pathak, and Abhinav Kumar Gupta. Interesting object, curious agent: Learning task-agnostic exploration. NeurIPS, 2021. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 411, + 545, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 411, + 545, + 442 + ], + "spans": [ + { + "bbox": [ + 308, + 411, + 545, + 442 + ], + "type": "text", + "content": "[57] Deepak Pathak, Pulkit Agrawal, Alexei A. Efros, and Trevor Darrell. Curiosity-driven exploration by self-supervised prediction. ICML, 2017. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 445, + 545, + 488 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 445, + 545, + 488 + ], + "spans": [ + { + "bbox": [ + 308, + 445, + 545, + 488 + ], + "type": "text", + "content": "[58] Hieu Pham, Zihang Dai, Golnaz Ghiasi, Hanxiao Liu, Adams Wei Yu, Minh-Thang Luong, Mingxing Tan, and Quoc V. Le. Combined scaling for zero-shot transfer learning. arXiv, 2021. 1, 13" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 490, + 545, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 490, + 545, + 521 + ], + "spans": [ + { + "bbox": [ + 308, + 490, + 545, + 521 + ], + "type": "text", + "content": "[59] Sarah Pratt, Luca Weihs, and Ali Farhadi. The introspective agent: Interdependence of strategy, physiology, and sensing for embodied agents. arXiv, 2022. 8" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 523, + 545, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 523, + 545, + 567 + ], + "spans": [ + { + "bbox": [ + 308, + 523, + 545, + 567 + ], + "type": "text", + "content": "[60] Yuankai Qi, Qi Wu, Peter Anderson, Xin Wang, William Yang Wang, Chunhua Shen, and Anton van den Hengel. Reverie: Remote embodied visual referring expression in real indoor environments. CVPR, 2020. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 308, + 568, + 545, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 568, + 545, + 633 + ], + "spans": [ + { + "bbox": [ + 308, + 568, + 545, + 633 + ], + "type": "text", + "content": "[61] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. ICML, 2021. 1, 2, 4, 6, 13" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 308, + 635, + 545, + 677 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 635, + 545, + 677 + ], + "spans": [ + { + "bbox": [ + 308, + 635, + 545, + 677 + ], + "type": "text", + "content": "[62] Shafin Rahman, Salman Hameed Khan, and Fatih Murat Porkikli. Zero-shot object detection: Learning to simultaneously recognize and localize novel concepts. ACCV, 2018. 1" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 308, + 680, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 680, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 308, + 680, + 545, + 712 + ], + "type": "text", + "content": "[63] Roberta Raileanu and Tim Rocktäschel. Ride: Rewarding impact-driven exploration for procedurally-generated environments. *ICLR*, 2020. 2" + } + ] + } + ], + "index": 31 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "23180" + } + ] + } + ], + "index": 33 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "text", + "content": "[64] Santhosh K. Ramakrishnan, Devendra Singh Chaplot, Ziad Al-Halah, Jitendra Malik, and Kristen Grauman. Poni: Potential functions for objectgoal navigation with interaction-free learning. CVPR, 2022. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 119, + 287, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 119, + 287, + 173 + ], + "spans": [ + { + "bbox": [ + 48, + 119, + 287, + 173 + ], + "type": "text", + "content": "[65] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, Alexander C. Berg, and Li Fei-Fei. Imagenet large scale visual recognition challenge. IJCV, 2015. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 175, + 287, + 218 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 175, + 287, + 218 + ], + "spans": [ + { + "bbox": [ + 48, + 175, + 287, + 218 + ], + "type": "text", + "content": "[66] Manolis Savva, Angel X Chang, Alexey Dosovitskiy, Thomas Funkhouser, and Vladlen Koltun. Minos: Multimodal indoor simulator for navigation in complex environments. arXiv, 2017. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 221, + 287, + 275 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 221, + 287, + 275 + ], + "spans": [ + { + "bbox": [ + 48, + 221, + 287, + 275 + ], + "type": "text", + "content": "[67] Manolis Savva, Abhishek Kadian, Oleksandr Maksymets, Yili Zhao, Erik Wijmans, Bhavana Jain, Julian Straub, Jia Liu, Vladlen Koltun, Jitendra Malik, Devi Parikh, and Dhruv Batra. Habitat: A Platform for Embodied AI Research. ICCV, 2019. 1, 3, 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 277, + 287, + 332 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 277, + 287, + 332 + ], + "spans": [ + { + "bbox": [ + 48, + 277, + 287, + 332 + ], + "type": "text", + "content": "[68] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. NeurIPS, 2022. 7, 16, 18" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 335, + 287, + 367 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 335, + 287, + 367 + ], + "spans": [ + { + "bbox": [ + 48, + 335, + 287, + 367 + ], + "type": "text", + "content": "[69] John Schulman, Filip Wolski, Prafulla Dhariwal, Alec Radford, and Oleg Klimov. Proximal policy optimization algorithms. arXiv, 2017. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 369, + 287, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 369, + 287, + 413 + ], + "spans": [ + { + "bbox": [ + 48, + 369, + 287, + 413 + ], + "type": "text", + "content": "[70] Ramprasaath R. Selvaraju, Abhishek Das, Ramakrishna Vedantam, Michael Cogswell, Devi Parikh, and Dhruv Batra. Grad-cam: Visual explanations from deep networks via gradient-based localization. IJCV, 2019. 4, 13, 14" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 415, + 287, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 415, + 287, + 459 + ], + "spans": [ + { + "bbox": [ + 48, + 415, + 287, + 459 + ], + "type": "text", + "content": "[71] Mohit Shridhar, Jesse Thomason, Daniel Gordon, Yonatan Bisk, Winson Han, Roozbeh Mottaghi, Luke Zettlemoyer, and Dieter Fox. Alfred: A benchmark for interpreting grounded instructions for everyday tasks. CVPR, 2020. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 460, + 287, + 493 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 460, + 287, + 493 + ], + "spans": [ + { + "bbox": [ + 48, + 460, + 287, + 493 + ], + "type": "text", + "content": "[72] Shuran Song, Linguang Zhang, and Jianxiong Xiao. Robot in a room: Toward perfect object recognition in closed environments. arXiv, 2015. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 495, + 287, + 528 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 495, + 287, + 528 + ], + "spans": [ + { + "bbox": [ + 48, + 495, + 287, + 528 + ], + "type": "text", + "content": "[73] Alexander L. Strehl and Michael L. Littman. An analysis of model-based interval estimation for markov decision processes. J. Comput. Syst. Sci., 2008, 2, 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 530, + 287, + 563 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 530, + 287, + 563 + ], + "spans": [ + { + "bbox": [ + 48, + 530, + 287, + 563 + ], + "type": "text", + "content": "[74] Saim Wani, Shivansh Patel, Unnat Jain, Angel X. Chang, and Manolis Savva. Mution: Benchmarking semantic map memory using multi-object navigation. NeurIPS, 2020. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 565, + 287, + 597 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 565, + 287, + 597 + ], + "spans": [ + { + "bbox": [ + 48, + 565, + 287, + 597 + ], + "type": "text", + "content": "[75] Luca Weihs, Matt Deitke, Aniruddha Kembhavi, and Roozbeh Mottaghi. Visual room rearrangement. CVPR, 2021. 2, 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 600, + 287, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 600, + 287, + 643 + ], + "spans": [ + { + "bbox": [ + 48, + 600, + 287, + 643 + ], + "type": "text", + "content": "[76] Luca Weihs, Jordi Salvador, Klemen Kotar, Unnat Jain, KuoHao Zeng, Roozbeh Mottaghi, and Aniruddha Kembhavi. Allenact: A framework for embodied ai research. arXiv, 2020. 3, 12" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 646, + 287, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 646, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 646, + 287, + 689 + ], + "type": "text", + "content": "[77] Erik Wijmans, Abhishek Kadian, Ari S. Morcos, Stefan Lee, Irfan Essa, D. Parikh, Manolis Savva, and Dhruv Batra. Dppo: Learning near-perfect pointgoal navigators from 2.5 billion frames. *ICLR*, 2019. 2, 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 691, + 286, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 286, + 713 + ], + "type": "text", + "content": "[78] Brian H Wilcox. Robotic vehicles for planetary exploration. Applied Intelligence, 1992. 2" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 455 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "text", + "content": "[79] Mitchell Wortsman, Kiana Ehsani, Mohammad Rastegari, Ali Farhadi, and Roozbeh Mottaghi. Learning to learn how to learn: Self-adaptive visual navigation using meta-learning. CVPR, 2019. 1, 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 118, + 545, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 545, + 150 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 545, + 150 + ], + "type": "text", + "content": "[80] Chenyun Wu, Zhe Lin, Scott D. Cohen, Trung Bui, and Subhransu Maji. Phrasecut: Language-based image segmentation in the wild. CVPR, 2020. 4, 14" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 152, + 545, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 152, + 545, + 184 + ], + "spans": [ + { + "bbox": [ + 307, + 152, + 545, + 184 + ], + "type": "text", + "content": "[81] Fei Xia, Amir R Zamir, Zhiyang He, Alexander Sax, Jitendra Malik, and Silvio Savarese. Gibson env: Real-world perception for embodied agents. CVPR, 2018. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 186, + 545, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 186, + 545, + 207 + ], + "spans": [ + { + "bbox": [ + 307, + 186, + 545, + 207 + ], + "type": "text", + "content": "[82] Brian Yamauchi. A frontier-based approach for autonomous exploration. CIRA, 1997. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 209, + 545, + 241 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 209, + 545, + 241 + ], + "spans": [ + { + "bbox": [ + 307, + 209, + 545, + 241 + ], + "type": "text", + "content": "[83] Claudia Yan, Dipendra Misra, Andrew Bennnett, Aaron Walsman, Yonatan Bisk, and Yoav Artzi. Chalet: Cornell house agent learning environment. arXiv, 2018. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 243, + 545, + 275 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 243, + 545, + 275 + ], + "spans": [ + { + "bbox": [ + 307, + 243, + 545, + 275 + ], + "type": "text", + "content": "[84] Wei Yang, Xiaolong Wang, Ali Farhadi, Abhinav Gupta, and Roozbeh Mottaghi. Visual semantic navigation using scene priors. arXiv, 2018. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 277, + 545, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 277, + 545, + 319 + ], + "spans": [ + { + "bbox": [ + 307, + 277, + 545, + 319 + ], + "type": "text", + "content": "[85] Xiaohua Zhai, Xiao Wang, Basil Mustafa, Andreas Steiner, Daniel Keysers, Alexander Kolesnikov, and Lucas Beyer. Lit: Zero-shot transfer with locked-image text tuning. CVPR, 2022. 14" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 321, + 545, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 321, + 545, + 342 + ], + "spans": [ + { + "bbox": [ + 307, + 321, + 545, + 342 + ], + "type": "text", + "content": "[86] Chong Zhou, Chen Change Loy, and Bo Dai. Denseclip: Extract free dense labels from clip. arXiv, 2021. 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 344, + 545, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 344, + 545, + 376 + ], + "spans": [ + { + "bbox": [ + 307, + 344, + 545, + 376 + ], + "type": "text", + "content": "[87] Fengda Zhu, Xiwen Liang, Yi Zhu, Xiaojun Chang, and Xiaodan Liang. Soon: Scenario oriented object navigation with graph-based exploration. CVPR, 2021. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 378, + 545, + 411 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 378, + 545, + 411 + ], + "spans": [ + { + "bbox": [ + 307, + 378, + 545, + 411 + ], + "type": "text", + "content": "[88] Pengkai Zhu, Hanxiao Wang, and Venkatesh Saligrama. Zero shot detection. IEEE Transactions on Circuits and Systems for Video Technology, 2020. 1" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 412, + 545, + 455 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 412, + 545, + 455 + ], + "spans": [ + { + "bbox": [ + 307, + 412, + 545, + 455 + ], + "type": "text", + "content": "[89] Yuke Zhu, Roozbeh Mottaghi, Eric Kolve, Joseph J Lim, Abhinav Gupta, Li Fei-Fei, and Ali Farhadi. Target-driven visual navigation in indoor scenes using deep reinforcement learning. ICRA, 2017. 1, 2" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "23181" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Coaching a Teachable Student/67002b28-2cb3-47c6-8f0a-06c8bb56b862_content_list.json b/2023/Coaching a Teachable Student/67002b28-2cb3-47c6-8f0a-06c8bb56b862_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..b687631f90c508404da0285ad26b903e7af2db02 --- /dev/null +++ b/2023/Coaching a Teachable Student/67002b28-2cb3-47c6-8f0a-06c8bb56b862_content_list.json @@ -0,0 +1,1406 @@ +[ + { + "type": "text", + "text": "Coaching a Teachable Student", + "text_level": 1, + "bbox": [ + 330, + 130, + 640, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Jimuyang Zhang Zanming Huang Eshed Ohn-Bar Boston University", + "bbox": [ + 267, + 180, + 702, + 217 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{zhangjim, huangtom, eohnbar}@bu.edu", + "bbox": [ + 323, + 219, + 643, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 268, + 313, + 284 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We propose a novel knowledge distillation framework for effectively teaching a sensorimotor student agent to drive from the supervision of a privileged teacher agent. Current distillation for sensorimotor agents methods tend to result in suboptimal learned driving behavior by the student, which we hypothesize is due to inherent differences between the input, modeling capacity, and optimization processes of the two agents. We develop a novel distillation scheme that can address these limitations and close the gap between the sensorimotor agent and its privileged teacher. Our key insight is to design a student which learns to align their input features with the teacher's privileged Bird's Eye View (BEV) space. The student then can benefit from direct supervision by the teacher over the internal representation learning. To scaffold the difficult sensorimotor learning task, the student model is optimized via a student-paced coaching mechanism with various auxiliary supervision. We further propose a high-capacity imitation learned privileged agent that surpasses prior privileged agents in CARLA and ensures the student learns safe driving behavior. Our proposed sensorimotor agent results in a robust image-based behavior cloning agent in CARLA, improving over current models by over $20.6\\%$ in driving score without requiring LiDAR, historical observations, ensemble of models, on-policy data aggregation or reinforcement learning.", + "bbox": [ + 75, + 300, + 473, + 680 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 693, + 209, + 709 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Learning internal representations for making intricate driving decisions from images involves a complex optimization task [4, 43, 54]. The inherent challenge for end-to-end training of driving agents lies in the immense complexity of learning to map high-dimensional visual observations into general and safe navigational decisions [13, 20, 70]. Even given millions of training examples [4], today's agents still fail to reliably learn an internal representation that can be used for robust processing of complex visual scenarios (e.g., dense urban settings with intricate layouts and dynamic obstacles) in a safe and task-driven manner [13, 20, 70].", + "bbox": [ + 75, + 718, + 468, + 885 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To ease the challenging sensorimotor training task, re", + "bbox": [ + 96, + 885, + 468, + 901 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/c07ac891b0e814efddc31764fff5c6d21d4fb1cd5b91050d00569bfe4c084176.jpg", + "image_caption": [ + "Figure 1. Effective Knowledge Distillation for Sensorimotor Agents. Our proposed CaT (Coaching a Teachable student) framework enables highly effective knowledge transfer between a privileged teacher and a sensorimotor (i.e., image-based) student. An alignment module learns to transform image-based features to the teacher's BEV feature space, where the student can then leverage extensive and direct supervision on its learned intermediate representations. The student model is optimized via a coaching mechanism with extensive auxiliary supervision in order to further scaffold the difficult sensorimotor learning task." + ], + "image_footnote": [], + "bbox": [ + 504, + 268, + 890, + 431 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "cent approaches decompose the task into stages, e.g., by first training a high-capacity privileged network with complete knowledge of the world and distilling its knowledge into a less capable vision-based student network [11, 13, 23, 47, 79]. However, due to the inherent differences between the inputs and architectures of the two agents, current methods rely on limited supervisory mechanisms from the teacher, i.e., exclusively through the teacher's output [11, 13] or knowledge distillation of a single final fully-connected layer [71, 79, 81]. Moreover, the privileged teacher's demonstration targets may be noisy or difficult for the student to imitate, given the limited perspective [27]. In this work, we sought to develop a more effective knowledge distillation paradigm for training a sensorimotor agent to drive. Our key insight is to enable more extensive supervision from the teacher by reducing the gap between internal modeling and learning capabilities between the two agents.", + "bbox": [ + 496, + 594, + 892, + 851 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Our proposed approach for holistic knowledge distillation is informed by human instruction, which often involves structured supervision in addition to high-level demonstra", + "bbox": [ + 498, + 854, + 893, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "7805", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "tions, e.g., providing various hints to scaffold information in a way that the student can better understand [30]. When teaching others new and challenging skills, i.e., where a student may not be able to replicate the demonstration such as riding a bicycle or driving a vehicle, teachers may provide additional supervision regarding the underlying task structure and their own internal reasoning [64]. Analogously to our settings, the privileged teaching agent can potentially provide richer supervision when teaching a limited capacity sensorimotor student, i.e., through more careful and direct guidance of the underlying representation learning.", + "bbox": [ + 75, + 90, + 470, + 257 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In our work, we introduce CaT, a novel method for teaching a sensorimotor student to drive using supervision from a privileged teacher. Our key insights are threefold: 1) **Effective Teacher:** We propose to incorporate explicit safety-aware cues into the BEV space that facilitate a surprisingly effective teacher agent design. While prior privileged agents struggle to learn to drive in complex urban driving scenes, we demonstrate our learned agent to match expert-level decision-making. 2) **Teachable Student via Alignment:** An IPM-based transformer alignment module can facilitate direct distillation of most of the teacher's features and better guide the student learning process. 3) **Student-paced Coaching:** A coaching mechanism for managing difficult samples can scaffold knowledge and lead to improved model optimization by better considering the ability of the student. By holistically tackling the complex knowledge distillation task with extensive teacher and auxiliary supervision, we are able to train a state-of-the-art image-based agent in CARLA [21]. Through ablation studies, input design, and interpretable feature analysis, we also provide critical insights into current limitations in learning robust and generalized representations for driving agents.", + "bbox": [ + 75, + 257, + 470, + 590 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 76, + 601, + 218, + 617 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Knowledge Distillation: We study knowledge distillation methods originally developed for model compression and acceleration [12,17,29,31,35,36,52,55,74,75] in the context of training sensorimotor. While approaches for feature distillation have been minimally explored in this context, we are motivated by their success in other domains, including image classification [2,38,40,42,60], object detection [67], semantic segmentation [46, 78], and natural language processing [3,68,73]. Yet, applying such techniques is not trivial given the drastically differing inputs between the privileged agent and student and the overwhelming sensorimotor task. Consequently, driving policy distillation methods only provide supervision either from the teacher's output [11,13] or the features of a single fully-connected layer [71,79,81], which (we hypothesize) does not provide sufficient hints to guide the student training. In this work, we provide novel mechanisms to close this gap and enable extensive supervision through deep feature distillation.", + "bbox": [ + 75, + 628, + 468, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Imitation Learning to Drive: Recent approaches in imitation learning (IL) to drive can be traced to Pomerleau [54]. Recently, more elaborate IL-based approaches for driving have emerged [9, 10, 25, 45, 50, 76, 77, 80, 83]. Specifically, decomposing the imitation learning task into two stages, i.e., by first learning a privileged agent through behavior cloning [11, 13] or reinforcement learning [8, 62, 71, 79] and then training the sensorimotor agent to mimic the output of the privileged agent. Our study is motivated by such approaches, yet we explore the benefits of increased supervision by the privileged agent. Moreover, prior privileged agents produce noisy and sub-optimal demonstration supervision in complex urban scenarios (even when optimized with RL, as demonstrated by our analysis). While most current studies employ the RL-based teacher of [79], this can be problematic and inefficient in safety-critical conditions in the real-world. In our work, we introduce a novel privileged teacher which can greatly surpass prior agents while only relying on offline behavior cloning.", + "bbox": [ + 496, + 90, + 893, + 378 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Intermediate Representation for Driving: Learning an effective 3D scene representation is crucial for safe autonomous driving. Researchers may obtain such representations by lifting the image to 3D using estimated depth and projecting the frustums into a BEV grid [32, 53]. Alternatively, transformer-based architectures can also enable mapping a camera image to the BEV space [14, 82], i.e., by attending to image-based information when populating a BEV-sampled grid. Related to our work is the study of Chen et al. [11], which learns a BEV representation from RGB image and LiDAR input. A motion planner is then used to generate future waypoints from the BEV. As this introduces significant challenges, we propose to leverage an IPM-based alignment module which can better structure to image-to-BEV projection task.", + "bbox": [ + 496, + 388, + 893, + 618 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Curriculum and Self-paced Learning: Approaches for curriculum learning methods often structure learning from various difficulty samples [5,24,37,39,44,59,61,63,69,84]. To better consider the challenging sensorimotor task and limited capacity of the student, we adopt a curriculum (i.e., an imitation coach), with progressively more challenging samples. He et al. [27] also proposes training a coach across iterations of on-policy data collection [56]. In contrast, we implement and demonstrate the benefit of a coach over training iterations without a data aggregation part. This can facilitate more efficient model training. The approach is motivated by self-paced learning [41, 72], where the model selects easy samples dynamically at each iteration based on a defined loss. In contrast to such studies, we only smooth the targets of the hard samples (based on the student), instead of discarding them. We empirically find this to result in improved coaching, potentially due to improved scaffolding of the difficult samples.", + "bbox": [ + 496, + 628, + 893, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "7806", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 76, + 89, + 168, + 104 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our goal is to decompose the challenging sensorimotor learning task and ease model optimization through effective supervision from a teacher agent. We first formulate the problem of privileged agent distillation in Sec. 3.1. Next, we address current limitations in privileged agent design to train a robust imitation learned teacher in Sec. 3.2. We then propose a transformer-based student architecture to resolve image-to-BEV feature alignment (Sec. 3.3). Finally, we train the sensorimotor student via extensive supervision and knowledge distillation (Sec. 3.4) via a progressive (i.e., coaching) mechanism (Sec. 3.5).", + "bbox": [ + 75, + 114, + 472, + 282 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Formulation", + "text_level": 1, + "bbox": [ + 76, + 290, + 210, + 305 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Driving can be formulated as a sequential decision-making problem over a set of sensory observations $\\mathcal{X}^s$ , a set of actions (motor commands) $\\mathcal{V}$ , and a policy function $f_{\\theta}^{s} \\colon \\mathcal{X}^{s} \\to \\mathcal{V}$ for mapping observations to actions at each time step [27,50,56]. We consider the task of learning a sensorimotor student agent via knowledge distillation [13, 23] from a teacher agent $f_{\\psi}^{t} \\colon \\mathcal{X}^{t} \\to \\mathcal{V}$ with privileged access to environmental information, e.g., ground-truth lanes, obstacles, and traffic light states. We parameterize the two agents as neural networks using weights $[\\pmb{\\theta}, \\pmb{\\psi}] \\in \\mathbb{R}^d$ , and denote $\\mathcal{F}^s(\\cdot; \\pmb{\\theta})$ and $\\mathcal{F}^t(\\cdot; \\pmb{\\psi})$ to be the feature maps for the sensorimotor student network and teacher network, respectively. Given a dataset $\\mathcal{D}$ comprising sensory and privileged observations and a loss function $\\mathcal{L}$ , the student can be optimized from the teacher using", + "bbox": [ + 75, + 313, + 472, + 541 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\underset {\\boldsymbol {\\theta}} {\\operatorname {a r g m i n}} \\mathbb {E} _ {(\\mathbf {x} ^ {s}, \\mathbf {x} ^ {t}) \\sim \\mathcal {D}} [ \\mathcal {L} (\\mathcal {F} ^ {s} (\\mathbf {x} ^ {s}; \\boldsymbol {\\theta}), \\mathcal {F} ^ {t} (\\mathbf {x} ^ {t}; \\boldsymbol {\\psi})) ] \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 99, + 549, + 468, + 574 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where the loss may be defined over a final layer, i.e., to match the output of the teacher [11, 13], or over internal layers as well, i.e., feature distillation [71, 79, 81]. Given the challenging end-to-end sensorimotor learning task, the role of the teacher is to provide effective supervision to the student, i.e., informative targets in Eqn. 1. However, as the two agents operate from drastically differing inputs, the knowledge transfer from the teacher to the student can become ineffective. Specifically, the task of the student is significantly more challenging than that of the teacher, i.e., due to inherent noise and uncertainty. Indeed, there is currently a substantial gap between the performance of current sensorimotor learning agents and their teaching agents, even assuming access to expensive and high-quality 3D observations such as LiDAR [11, 13, 16] (not used in our work). The teaching process should also ideally consider any limited capacity of the student to facilitate effective optimization and learning [27]. Moreover, learning an optimal and robust teacher in itself can be a challenging task [13], even assuming the simplified learning task such as solved perception, as we demonstrate in Sec. 4 and addressed in Sec. 3.2.", + "bbox": [ + 75, + 583, + 472, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Problem Setup and Agent Observations: We develop our CaT framework in the context of the CARLA simulator [21]. Our objective is to train a goal-conditional sensorimotor agent [16, 18] for mapping image and goal observations to vehicle throttle and steering control [13, 48]. We follow standard conditional imitation learning [11, 16] and assume access to three (non-overlapping) RGB camera views $\\mathbf{I} = [\\mathbf{I}_0, \\mathbf{I}_1, \\mathbf{I}_2] \\in \\mathbb{R}^{W \\times H \\times 3}$ , a categorical navigational command $c \\in \\{1, \\dots, 6\\}$ (i.e., turn left, turn right, follow, forward, left lane changing and right lane changing) and an intermediate noisy goal $\\mathbf{g} \\in \\mathbb{R}^2$ sampled from a GNSS (Global Navigation Satellite System) [11]. We note that both the command and GNSS observations can be easily obtained by today's vehicles, i.e., through generation of an $\\mathrm{A}^*$ plan [26] from a static map and coarse positioning system. The GNSS goals are sampled every 50-100 meters and reflect real-world measurement errors [11]. We compute a BEV $\\mathbf{B} \\in \\{0, 1\\}^{W_B \\times H_B \\times C_B}$ by rendering privileged (i.e., ground-truth) information from the underlying simulation, including 3D location and state of lanes, pedestrians, vehicles, and traffic lights [79]. While prior work has generally leveraged a standardized BEV representation [13, 32, 34, 57, 76], we find its design to be crucial to training an optimal teacher policy as will be discussed in Sec 3.2. We let $\\mathbf{x}^s = (\\mathbf{I}, \\mathbf{g}, c) \\in \\mathcal{X}^s$ and $\\mathbf{x}^t = (\\mathbf{B}, \\mathbf{g}, c) \\in \\mathcal{X}^t$ be the student and teacher observations, respectively. Given these observations at each time step, the agent learns to predict 10 future 2D waypoints in top-down vehicle coordinates for the next 2.5 seconds of driving. These are then given to a lateral and longitudinal PID controllers [13, 48, 66] to generate the final action.", + "bbox": [ + 496, + 90, + 893, + 559 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Given this formulation, we next discuss our teacher agent design and training process. In particular, we aim to learn a privileged agent which can not only effectively solve the goal-oriented navigation task, but also facilitate effective distillation and coaching of a sensorimotor student policy, as will be discussed in Sec. 3.4.", + "bbox": [ + 496, + 560, + 893, + 650 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Learning an Effective Teacher", + "text_level": 1, + "bbox": [ + 500, + 664, + 769, + 680 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In our formulation, learning a student agent begins with training an effective privileged agent $f_{\\psi}^{t}$ . However, we find this itself to be a complex task. Even with complete knowledge of perfect BEV perception, current privileged agents are suboptimal, significantly under-performing CARLA's built-in autopilot [13, 79]. In this section, we uncover the underlying reason for this under-performance through the BEV design. Informed by this key finding, we train a highly effective imitation-learned teacher that surpasses prior privileged agents and matches expert-level performance in CARLA without requiring extensive data collection [56] or reinforcement learning [79]. Our proposed privileged agent can then enable effective supervision of the student in Sec. 3.3.", + "bbox": [ + 496, + 688, + 893, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "7807", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/5c949646a75d96ca06d760734237b42d48f99536cc947e87bd81106f55f5696c.jpg", + "image_caption": [ + "Figure 2. Overview of Architecture and Distillation Process of CaT. We first train a high-capacity privileged teacher agent (colored in red) which takes an augmented BEV representation as input, and produces a set of future waypoints conditioned on command and short-term goal. Our teachable student agent (colored in blue) maps the image features to the BEV space using an alignment module, which is trained by distillation of intermediate features (arrows colored in blue) and final output (arrow colored in green). Additionally, we leverage supervision from auxiliary segmentation (arrow colored in red) and command prediction tasks (arrow colored in purple) interlaced throughout the network to facilitate the learning of task-relevant features." + ], + "image_footnote": [], + "bbox": [ + 81, + 89, + 893, + 311 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Teacher Training via Direct Expert Imitation: In imitation learning, an oracle (i.e., expert driver) $f^{*} \\colon \\mathcal{X}^{t} \\to \\mathcal{Y}$ is defined for demonstrating optimal actions $\\mathbf{y}^{*}$ . In CARLA [21], the expert is defined using longitudinal and lateral PID controllers which are carefully tuned and augmented by a collection of manually specified rules for handling diverse scenarios (e.g., sudden braking) over the underlying simulation state. The designed expert can then be used to generate trajectories as supervised data for learning the privileged policy function $f_{\\psi}^{t}$ with behavior cloning [18, 19, 48]. Our privileged agent is a ResNet-18 model [28], as shown in Fig. 2 (further increasing modeling capacity was not found to be beneficial). The ResNet model is followed by a GRU-based conditional waypoint predictor [11, 16]. Due to the complexity of our dense urban navigation task, e.g., with multi-lane roads, intersections, and merging, we incrementally refine the waypoints using two GRUs within each conditional branch. The first GRU regresses the preliminary set of waypoints via sequential waypoints directly from the embedded features. Subsequently, the second GRU takes the predicted waypoints together with the short-term goal and embedded features, to produce a refined set of waypoint targets.", + "bbox": [ + 75, + 433, + 472, + 782 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Even assuming perfect perception, we find it difficult to train privileged agents that exhibit robust planning behavior in complex urban driving scenarios. Specifically, when using the most challenging CARLA benchmark (Longest6 [16], also detailed in Sec 4) using the standard BEV representation of Chen et al. [13] leads to $26\\%$ driving score compared to $72\\%$ by the rule-based expert. The rein", + "bbox": [ + 75, + 794, + 468, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "forcement learning agent of Zhang et al. [79] achieves $60\\%$ on this task. We sought to explore the limits of behavior cloning for this task, as a noisy and sub-optimal teacher can hinder the training of the student model. In particular, we hypothesize that the reason for the poor performance lies in the increased complexity of the BEV state representation in dense scenarios. In such scenarios, learning to extract task-relevant details can become more challenging. To address this issue, we propose to introduce additional channels (akin to affordances [9,22,58]) that can more easily translate into safety-critical decision-making, simplify the learning task, and provide an expert-level teaching agent.", + "bbox": [ + 496, + 433, + 893, + 617 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "BEV with Safety Hints: Our BEV encodes state information into distinct channels, including drivable road regions, the desired route based on the command and layout, lane markings, and dynamic obstacles (see Fig. 3). Given the low performance of prior privileged agents, in particular around dynamic objects, the agent could benefit from more explicit safety-driven cues in the BEV. We propose to add two types of channels of (1) predicted agents' future and (2) entity attention. First, we utilize a kinematics bicycle model [16] in order to efficiently predict future trajectories of dynamic objects. This enables us to iteratively predict and represent short-term future position, orientation, and speed of agents (our supplementary contains additional information regarding this process). Secondly, we encode an explicit attention channel for highlighting potential future infractions, i.e., based on the kinematics bicycle model applied to the ego-vehicle. While this assumes that the current vehicle throttle and steering remain unchanged, the entity", + "bbox": [ + 496, + 628, + 895, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "7808", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/73eeaae3b8b28898a1237ee7e1faff2380488617b79ab9d151c46b90ce0bf569.jpg", + "image_caption": [ + "(a) Baseline [79]" + ], + "image_footnote": [], + "bbox": [ + 81, + 106, + 233, + 224 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/fc806bd673925eeb5ae8eb6eb6e21ed36a083ed94b230d75d84906c3936f1bac.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 235, + 107, + 387, + 224 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/fd79ed0d0f73eb8661bc18da399b93b74e16fcf3bb024c507ad7c3d9b8878c47.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 398, + 109, + 485, + 155 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/cd4190bd9b1aaec56b3c88657eececf62a2e7490db02bd9413bb5a621e807bf5.jpg", + "image_caption": [ + "(b) With Safety Hints", + "(c) Road", + "(g) Pedestrians", + "Figure 3. Visualization of Our Proposed BEV with Safety Hints. (a) and (b) visualize the baseline and our proposed BEV by encoding different channels into colors. Roads are shown in dark grey, Desired Route is shown in light grey, Lane Marks are shown in magenta, Vehicles are shown in blue with past trajectories shown in decreasing saturation with time into the past, Pedestrians are shown in light blue, Traffic Lights are shown in red or green according to its state, Agent Forecasts are shown in purple with decreasing saturation with time into the future, Entity Attention is shown in orange. (c) - (j) shows the different channels separately." + ], + "image_footnote": [], + "bbox": [ + 398, + 172, + 485, + 239 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/b4024c7fd22bea493e03de456ef1978ad97234ae0eae693bc839cc22b7dc9ec7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 511, + 89, + 599, + 156 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/bbc8e8a2da2d118bb459310c0bd2db244beb0838983a1409943cfb677f16caf6.jpg", + "image_caption": [ + "(d) Desired Route", + "(h) Traffic Lights" + ], + "image_footnote": [], + "bbox": [ + 511, + 172, + 596, + 239 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/36f96d7fddefc58add70214537bc7032084d60e96b581ce078d55d2d5a9c8867.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 640, + 89, + 727, + 155 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/02093be44a605a22045abe59041659a6bae0e47327c3c8ca5ca8c9f9b6a5ca10.jpg", + "image_caption": [ + "(e) Lane Marks", + "(i) Agent Forecast" + ], + "image_footnote": [], + "bbox": [ + 640, + 172, + 727, + 239 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/0389cadfe8fd59778bcc6fc88d06cf550ebdb972cd0a5ba6593af2cb91120a97.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 777, + 90, + 864, + 155 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/e2ac63a43c8db4a86b4ff9dabc0741bf9773f5b501f128474f7a30ef018291a3.jpg", + "image_caption": [ + "(f) Vehicles", + "(j)", + "Entity Attention" + ], + "image_footnote": [], + "bbox": [ + 777, + 172, + 864, + 239 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "attention channel helps the agent better reason over safety while correlating states with actions. Surprisingly, the additions of the safety hints channels provide a strong teacher with offline behavior cloning alone, even improving over the rule-based expert (Table 1, $73\\%$ vs. $72\\%$ ). We note that many of the remaining failures are due to simulation timeout in crowded intersections. Our boosted teacher agent will be critical for effective distillation in Sec. 3.4.", + "bbox": [ + 75, + 349, + 468, + 468 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Learning a Teachable Student", + "text_level": 1, + "bbox": [ + 76, + 496, + 341, + 511 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Due to differing inputs and modeling capacities, it can be difficult to align the image-based student features and output with the BEV-based privileged teacher. We propose to alleviate misalignment by designing a teachable student network, i.e., a model that matches as much of the teacher architecture as possible. Specifically, we propose to improve teacher-student model alignment by leveraging an internal image-to-BEV alignment module. As shown in Fig. 2, the three residual blocks following the alignment module can consequently facilitate knowledge transfer via direct distillation of most of the teacher's features in Sec. 3.4.", + "bbox": [ + 75, + 525, + 468, + 691 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "IPM-based Transformer Alignment Module: Explicitly mapping images to BEV space for downstream tasks can be challenging. While this can be achieved via depth estimation and lifting [32,53], we sought to incorporate an efficient and differentiable module which can effectively attend and reason over various features while mapping from front view to BEV. A transformer-based architecture [14,65] provides a natural choice as it can be optimized end-to-end with respect to the driving task while also carefully aligning features among spaces with arbitrary dimensionalities. As shown in Fig. 2, we sample queries $\\mathbf{Q}_{init} \\in \\mathbb{R}^{L_H \\times L_W \\times L_C}$ using a spatial parameterization of the BEV space and process them using a self-attention module [14,65]. The output", + "bbox": [ + 75, + 703, + 468, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "of self-attention module $Q$ can be formulated as", + "bbox": [ + 500, + 349, + 816, + 364 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {Q} = \\operatorname {s o f t m a x} \\left(\\frac {\\mathbf {Q} _ {\\text {i n i t}} \\mathbf {K} _ {\\text {i n i t}} ^ {T}}{\\sqrt {d}}\\right) \\mathbf {V} _ {\\text {i n i t}} \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 575, + 368, + 890, + 402 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathbf{K}_{init}$ , $\\mathbf{V}_{init} \\in \\mathbb{R}^{L_H \\times L_W \\times L_C}$ provide the key-value pairs and $d$ is the dimension of the query. Subsequently, each query point $\\mathbf{q} \\in \\mathbf{Q}$ is mapped to its corresponding reference point $\\mathbf{p} \\in \\mathbb{R}^3$ (homogeneous coordinates) in the image features $\\mathcal{F}_{RGB} \\in \\mathbb{R}^{Z_H \\times Z_W \\times Z_C}$ using an Inverse Perspective Mapping (IPM)", + "bbox": [ + 496, + 407, + 890, + 500 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {p} = s \\mathbf {P} _ {k} \\mathbf {R} _ {k} (\\mathbf {q} - \\mathbf {t} _ {k}) \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 620, + 502, + 888, + 517 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $s$ is a scale factor, $\\mathbf{P}_k\\in \\mathbb{R}^{3\\times 3}$ , $\\mathbf{R}_k\\in \\mathbb{R}^{3\\times 3}$ , and $\\mathbf{t}_k\\in \\mathbb{R}^3$ are the $k^{th}$ camera intrinsics, rotation and translation relative to the center of the ego-vehicle, respectively [6]. Note that we employ a different IPM mapping for each camera view to perform the projection.", + "bbox": [ + 496, + 518, + 890, + 595 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To populate the student's BEV features $\\mathcal{F}_{BEV} \\in \\mathbb{R}^{H_B \\times W_B \\times C_B}$ (i.e., the output of the alignment module in Fig. 2), we leverage a deformable cross-attention mechanism based on deformable DETR [85]. This enables the network to attend to multiple regions around the reference point in the image features when transforming to BEV feature space. The reference point $\\mathbf{p}$ and the learned deformable offsets can then be used as keys $\\mathbf{K}$ , with their corresponding features from the extracted images features the values $\\mathbf{V}$ , for a cross-attention module", + "bbox": [ + 496, + 595, + 890, + 744 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {F} _ {B E V} = \\operatorname {D e f o r m A t t n} (\\mathbf {Q}, \\mathcal {F} _ {R G B}, \\mathbf {H}) \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 568, + 752, + 890, + 768 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathbf{H} = \\{\\mathbf{P}_k, \\mathbf{R}_k, t_k\\}_{k=1}^3$ contains the combined multiview IPM parameters from Eqn. 3. In addition to the proposed alignment and teachable modules, we also incorporate prediction modules for auxiliary tasks (i.e., BEV prediction [16]), discussed in Sec. 3.4.", + "bbox": [ + 496, + 773, + 890, + 849 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "GRU-based Waypoint Predictor: The transformed imaged-to-BEV features are inputted to three residual blocks and a GRU-based conditional branch based on the command [11].", + "bbox": [ + 496, + 854, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "7809", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Student Loss", + "text_level": 1, + "bbox": [ + 76, + 90, + 215, + 104 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Even with our carefully designed architecture, the student's image-based learning task remains significantly more challenging than the teacher's. Thus, both optimization and alignment with the teacher can benefit from extensive supervision and gradual student instruction, i.e., to scaffold supervision and better match the student's ability. Our distillation process provides ample beneficial supervision for the student through three holistic mechanisms. First, we incorporate deep distillation losses, i.e., both over output (i.e., waypoint regression task) and feature maps, as shown in Fig. 2, that can more directly supervise internal layers. Second, we incorporate additional auxiliary losses that regularize the optimization process. This includes ground-truth supervised losses, e.g., defined over segmentation and control command, and a geometric loss defined over the post-alignment features that supervises for correct BEV structure. Finally, we leverage a student-paced coaching mechanism for gradually increasing the difficulty of the waypoint prediction task throughout the iterations in Sec. 3.5.", + "bbox": [ + 75, + 114, + 472, + 402 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Loss Functions: Our optimization objective for guiding the distillation process is a weighted sum over both distillation and auxiliary tasks:", + "bbox": [ + 75, + 409, + 470, + 455 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {C a T} = \\mathcal {L} _ {\\text {o u t}} + \\mathcal {L} _ {\\text {f e a t}} + \\mathcal {L} _ {\\text {s e g}} + \\mathcal {L} _ {\\text {c m d}} \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 140, + 469, + 468, + 487 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The various loss terms enable guiding holistic aspects of student learning.", + "bbox": [ + 76, + 500, + 470, + 531 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Output Distillation: We leverage an $L_{1}$ loss computed over all of the conditional branches for each instance", + "bbox": [ + 76, + 537, + 468, + 569 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {o u t} = \\sum_ {c = 1} ^ {C} \\| f _ {\\boldsymbol {\\theta}} ^ {s} \\left(\\mathbf {x} ^ {s}, c\\right) - f _ {\\boldsymbol {\\psi}} ^ {t} \\left(\\mathbf {x} ^ {t}, c\\right) \\| _ {1} \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 148, + 582, + 468, + 625 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "This can be done by sampling from the teacher with different commands $c$ (following Chen et al. [13]).", + "bbox": [ + 76, + 635, + 468, + 667 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Feature Distillation Loss: Our proposed student architecture provides BEV-space feature alignment, i.e., for direct internal feature matching between the student and the teacher (the first term in our loss in Eqn. 7). This is in contrast to current methods which distill the output or a final fully-connected layer [11, 13, 71, 79]. However, while this can provide rich supervision for the student, the two models should not match entirely, i.e., due to differences in information processing for a perception-based planning task. Moreover, the BEV-based features can be highly sparse and structured, with slight feature offsets being less meaningful than obtaining a task and scene-relevant representation. To flexibly account for the student-based task and effectively represent structural information in BEV space, our feature loss is computed as", + "bbox": [ + 75, + 672, + 472, + 901 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} _ {f e a t} = \\sum_ {i = 1} ^ {3} \\left[ \\| \\mathcal {F} _ {i} ^ {s} (\\mathbf {x} ^ {s}) - \\mathcal {F} _ {i} ^ {t} (\\mathbf {x} ^ {t}) \\| _ {2} + \\right. \\\\ \\left\\| T _ {i} ^ {s} \\left(\\mathcal {F} _ {i} ^ {s} \\left(\\mathbf {x} ^ {s}\\right)\\right) - T _ {i} ^ {t} \\left(\\mathcal {F} _ {i} ^ {t} \\left(\\mathbf {x} ^ {t}\\right)\\right) \\right\\| _ {2} + \\tag {7} \\\\ \\lambda_ {C D} \\left\\| \\mathcal {F} _ {i} ^ {s} (\\mathbf {x} ^ {s}) - \\mathcal {F} _ {i} ^ {t} (\\mathbf {x} ^ {t})\\right) \\| _ {C D} \\Bigg ] \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 555, + 88, + 890, + 178 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where we use $\\mathcal{F}_i$ to denote layers within the networks (as shown in Fig. 2, we directly distill three layers), $T_{i}$ indicates the features following several convolutional layers (trained jointly with the student network model), and $CD$ stands for a Chamfer Distance [51] (we set $\\lambda_{CD} = 0.1$ ). The final term is computed over thresholded activations and a spatial soft-argmax [13] from each feature map.", + "bbox": [ + 496, + 186, + 893, + 295 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Task-oriented Auxiliary Tasks: To ensure essential task-relevant information is preserved throughout the distillation process, we leverage regularizing supervision in the form of segmentation and command-based auxiliary tasks interlaced throughout the network [16, 49]. Our segmentation loss $\\mathcal{L}_{seg}$ is a cross-entropy loss, computed over the ground-truth BEV and averaged across the segmentation prediction heads. Our command prediction loss $\\mathcal{L}_{cmd}$ is binary cross-entropy supervised by the ground-truth command $c$ , which facilitates learning task-relevant features [11, 19].", + "bbox": [ + 496, + 300, + 890, + 452 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.5. Student-paced Coaching", + "text_level": 1, + "bbox": [ + 500, + 460, + 725, + 478 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "While Sec. 3.4 provides ample supervision on the student's internal representations, the teacher's targets may still be difficult to imitate for the student. To better consider the learning ability of the student, we propose to leverage a student-paced training mechanism which can gradually coach the student, i.e., through increasingly difficult samples. When computing the loss in Eqn. 5, we define a coach,", + "bbox": [ + 496, + 484, + 890, + 590 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {F} ^ {t} \\leftarrow \\lambda_ {i} \\mathcal {F} ^ {s} + \\left(1 - \\lambda_ {i}\\right) \\mathcal {F} ^ {t}, \\text {i f} \\mathcal {L} _ {C a T} > \\tau_ {i} \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 555, + 595, + 890, + 613 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "which interpolates the teaching targets with the student's predictions and features. $\\lambda_{i}$ is a hyperparameter which is linearly decreased to 0 over training iteration $i$ . The modified targets are only computed over the hard samples at each batch. Within each batch, $\\tau_{i}$ is a threshold defining the lowest $50\\%$ of the scored samples for the loss with respect to the privileged teacher. While smoothing the targets in this manner may seem counterintuitive, i.e., compared to techniques which mine hard examples, the coaching mechanism aims to stabilize training by reducing the difficulty when the student is unable to perform the optimal action. Eqn. 8 enables adjusting the learning rate in a sample-selective manner (initially suppressing difficult samples).", + "bbox": [ + 496, + 619, + 890, + 816 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 500, + 830, + 633, + 847 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We use the CARLA simulator (version 0.9.10.1) [21] for data generation and closed-loop evaluation of the proposed CaT framework. We also use open-loop evaluation", + "bbox": [ + 496, + 854, + 890, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "7810", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/72c197e27232dbec756df79c290655552364a92f5996bd6ebb089498291960f7.jpg", + "table_caption": [ + "Table 1. Quantitative Evaluation on the Longest6 Benchmark. Comparison of CaT with prior methods in terms of Driving Score (DS), Route Completion (RC), and Infraction Score (IS). Additional infraction metrics (Pedestrian Collisions (Ped), Vehicle Collisions (Veh), Layout Collisions (LC), Red Light Violations (Red), Off-road Infraction (OR), Route deviation (Dev), Route Timouts (TO), Agent Blocked (Blk)) are shown. FD refers to Feature Distillation, SH refers to Safety Hints. *re-trained by us using the publicly available code. Mean and standard deviation are computed over three runs. CaT outperforms the state-of-the-art by a large margin in terms of DS." + ], + "table_footnote": [], + "table_body": "
MethodRGBLiDARDS↑RC↑IS↑Ped↓Veh↓LC↓Red↓OR↓Dev↓TO↓Blk↓
LAV [11]✓✓48.41±3.4080.71±0.840.60±0.040.000.500.190.070.200.070.010.29
TransFuser [16]✓✓46.20±2.5783.61±1.160.57±0.000.290.380.280.040.070.070.000.32
WOR [10]✓X17.36±2.9543.46±2.990.54±0.060.050.640.150.840.150.890.040.45
NEAT [15]✓X24.08±3.3059.94±0.500.49±0.020.010.710.210.180.180.000.022.83
TCP* [71]✓X42.86±0.6361.83±4.190.71±0.040.010.420.110.040.080.010.010.75
CaT (w/o Alignment, Coaching, FD)✓X39.48±0.6760.96±1.650.68±0.010.031.290.200.080.840.000.021.49
CaT (w/o Alignment, Coaching)✓X40.64±0.9862.45±0.460.67±0.010.021.070.330.140.380.000.021.11
CaT (w/o Coaching, FD, SH)✓X44.10±0.4065.84±5.550.72±0.030.010.260.050.110.150.000.000.66
CaT (w/o Coaching, SH)✓X49.69±2.2881.10±0.580.64±0.020.010.780.030.080.200.000.020.51
CaT (w/o Coaching)✓X55.55±1.4181.97±2.340.68±0.010.020.300.050.070.060.000.020.35
CaT✓X58.36±2.2478.79±1.500.77±0.020.010.200.020.050.300.000.040.44
Privileged Agents:
RL Expert (Roach) [79]--60.14±2.4085.83±0.600.69±0.030.060.220.000.010.000.000.040.07
Rule-based Expert--71.96±2.1377.46±3.110.91±0.000.010.060.000.000.000.000.030.41
Basic BEV Agent [13]--24.08±2.8373.36±1.080.31±0.060.011.452.450.090.270.110.010.24
+ History and Desired Path--52.81±1.7979.34±3.650.71±0.060.020.300.000.000.000.000.020.53
+ Agent Forecast--65.73±0.9383.50±1.180.79±0.020.010.210.000.020.000.000.010.28
+ Entity Attention--73.30±1.0787.44±0.280.83±0.020.050.110.000.020.000.000.020.12
", + "bbox": [ + 76, + 162, + 890, + 375 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "on nuScenes [7, 33, 77] (ADE, FDE, and collision rate). In Carla, we leverage the Longest6 Benchmark [16], which uses the six longest routes of each town (Town01 - Town06) from the set of 76 routes provided by the official CARLA leaderboard [1] (a total of 36 routes). We note in addition to the standard cross weather and time of day generalization evaluation, the long and dense traffic conditions in Longest6 present the most challenging settings across the various CARLA-based benchmarks. To evaluate our models, we follow standard metrics and report Route Completion (RC, in terms of completed route percentage), Infraction Score (IS, a penalty factor over infractions), and Driving Score (DS, computed from the prior two) [1, 11, 16]).", + "bbox": [ + 75, + 386, + 472, + 583 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.1. Comparison with Prior Methods", + "text_level": 1, + "bbox": [ + 76, + 604, + 364, + 619 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "CARLA Results: As shown in Table 1, using feature distillation, safety hints-based BEV for the teacher, and the proposed coaching mechanism, CaT is able to obtain a $58.36\\%$ DS. Specifically, we achieve state-of-the-art performance among all prior agents, including LiDAR-based approaches [11, 16] (by $20.6\\%$ in terms of DS, from 48.41 to 58.36, and $28.3\\%$ in terms of IS). We also note that such approaches provide strong baselines, due to various LiDAR-based safety checks that the agents perform on top of the trained policy, which we do not employ. Moreover, CaT outperforms the prior RGB-only state-of-the-art agent TCP [71] by $36.16\\%$ DS. To further validate the generalization of CaT, we discuss an additional benchmark split from TCP [71] and LAV [11] in the supplementary. We also find that removal of the alignment module degrades DS from $44.10\\%$ to $39.48\\%$ , and is shown to only slightly benefit from feature distillation ( $40.64\\%$ DS) indicating the effectiveness of the proposed architecture.", + "bbox": [ + 75, + 628, + 472, + 902 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/9bb5654956a20a02dab4ee9460c67a51131b440c31eb7dfbb81d3d24b11eb68f.jpg", + "table_caption": [ + "Table 2. Open-Loop Evaluation on nuScenes." + ], + "table_footnote": [], + "table_body": "
MethodADE (m) ↓FDE (m) ↓Coll. (%) ↓
BEV Agent0.330.520.49
CaT (w/o Coaching, FD, SH)0.480.430.68
CaT0.410.360.27
", + "bbox": [ + 506, + 400, + 885, + 460 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Real-World Evaluation: To further analyse the benefits of CaT in realistic driving settings, Table 2 shows open-loop evaluation (ADE, FDE, and Collision rate [77]) for nuScenes [7]. On the official validation split, the privileged agent performs best with an ADE of 0.33. CaT achieves 0.41 ADE, improving by $14.6\\%$ gain over an agent without distillation and coaching. Moreover, we find a significant reduction in the collision rate by $60.3\\%$ compared to the baseline. Additional details regarding nuScenes evaluation can be found in the supplementary.", + "bbox": [ + 496, + 474, + 890, + 627 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. Ablation Studies", + "text_level": 1, + "bbox": [ + 500, + 638, + 663, + 654 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Teacher Comparison: Table 1 depicts our teacher model ablation. Specifically, we find the proposed BEV channels to drastically improve the privileged agent's performance by simplifying the learning task (73.30% DS, surpassing the built-in rule-based expert of 71.96%). We note that our behavior cloning agent also significantly outperforms an RL-based expert [79] (60.14%), simply through an effective BEV design. Thus, our findings apply to real-world scenarios where interactive agent training may be unsafe. We also observe that the improved teacher benefits the student agent. Specifically, incorporating the proposed BEV-based safety hints results in student performance gains, from 49.69% to 55.55%.", + "bbox": [ + 496, + 665, + 890, + 861 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Impact of Feature Distillation: Table 3 presents the benefits of our proposed multi-layer feature distillation frame", + "bbox": [ + 498, + 869, + 890, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7811", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/be6e47ce4a4140cd30f20b8992747d2d209dbabb2f92c0ab24a41a7f3729d30a.jpg", + "image_caption": [ + "Figure 4. Visualization of Model Input and Features. We show the BEV ground truth (BEV) and RGB image ground truth (Image) on the top row. Next, we visualize the intermediate layer activations of a sensorimotor agent without the proposed alignment module [71] (No Alignment Module), privileged teacher agent (Privileged), student agent trained with output distillation only (Output Distillation), and the proposed agent (CaT). We demonstrate that our teacher agent with proposed knowledge distillation is able to guide the student agent in learning scene-relevant representations and attend to safety-critical entities (circled in red)." + ], + "image_footnote": [], + "bbox": [ + 156, + 88, + 810, + 366 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/0f8afabc7b8a470317888a0db8ca37762ec1efbe6c9ff97c01af83b76dea9d6a.jpg", + "table_caption": [ + "Table 3. Ablation Study on Feature Distillation Layers. All loss terms in Eqn. 7 benefit training with deep feature distillation." + ], + "table_footnote": [], + "table_body": "
MethodDS ↑RC ↑IS ↑
No Distillation44.1065.840.72
One Layer [71,79]45.2369.330.68
Three Layers L249.3166.920.78
Three Layers L2 + LCD51.9562.820.87
Three Layers Lfeat55.5581.970.68
", + "bbox": [ + 99, + 481, + 446, + 588 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "work. We note that prior methods only distill one final fully connected layer (before the prediction head [71, 79], referred to as 'one layer' in Table 3). We note that such approaches result in minimal gains, whereas CaT is able to benefit from the extensive feature distillation (improving from $45.23\\%$ to $55.55\\%$ ). Incorporating multiple types of feature distillation losses also benefits performance.", + "bbox": [ + 75, + 611, + 468, + 717 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.3. Feature Visualization", + "text_level": 1, + "bbox": [ + 76, + 726, + 277, + 742 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Fig. 4 depicts the learned features of the initial feature distilled layer (out of the three in Fig. 2). We visualize activations from two channels, to better understand the learned representation by the agents. Interestingly, the visualization shows how traditionally trained sensorimotor agents (i.e., without the proposed alignment module) mostly operate in image-space features in this layer despite the BEV waypoint prediction task. This validates our main hypothesis, that such agents have limited ability to learn from a BEV-based agent, even with BEV-based auxiliary prediction tasks. We", + "bbox": [ + 75, + 750, + 470, + 902 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "also find that output distillation provides a weak supervision for critical components of the BEV, including layout and object features. After feature distillation, the features of our CaT model are able to not only keep the driving-related semantic information i.e., drivable road and route, but also focus on safety-critical entities. More visualization examples are provided in the supplementary.", + "bbox": [ + 496, + 450, + 893, + 558 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 577, + 617, + 592 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We present CaT, a novel knowledge distillation method for effectively training a sensorimotor student agent using the supervision from a privileged agent. We leverage an alignment module to better map image features to the BEV space, thus enabling extensive supervision from a BEV-based teacher over the intermediate feature learning. As both agents can be trained offline via imitation learning, our findings are directly relevant to real-world settings where on-policy interactions may be unsafe to perform. Finally, we carefully account for inherent differences between the student and the teacher using a student-paced coaching mechanism with various auxiliary supervision tasks. Through the improved knowledge distillation process, our experiments result in a state-of-the-art camera-based agent in CARLA. Our ablation into input and model design guides future directions through interpretable analysis into current limitations of end-to-end driving models.", + "bbox": [ + 496, + 604, + 893, + 861 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments: We thank the Red Hat Collaboratory for supporting this research.", + "bbox": [ + 498, + 869, + 890, + 901 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "7812", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Carla autonomous driving leaderboard. https://leaderboard.carla.org/, 2022.7", + "[2] Yonathan Aflalo, Asaf Noy, Ming Lin, Itamar Friedman, and Lihi Zelnik. Knapsack pruning with inner distillation. arXiv preprint arXiv:2002.08258, 2020. 2", + "[3] Gustavo Aguilar, Yuan Ling, Yu Zhang, Benjamin Yao, Xing Fan, and Chenlei Guo. Knowledge distillation from internal representations. In AAAI, 2020. 2", + "[4] Mayank Bansal, Alex Krizhevsky, and Abhijit Ogale. ChauffeurNet: Learning to drive by imitating the best and synthesizing the worst. In RSS, 2019. 1", + "[5] Yoshua Bengio, Jérôme Louradour, Ronan Collobert, and Jason Weston. Curriculum learning. In ICML, 2009. 2", + "[6] Massimo Bertozzi, Alberto Broggi, and Alessandra Fascioli. An extension to the inverse perspective mapping to handle non-flat roads. In IV, 1998. 5", + "[7] Holger Caesar, Varun Bankiti, Alex H Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuscenes: A multimodal dataset for autonomous driving. In CVPR, 2020. 7", + "[8] Raphael Chekroun, Marin Toromanoff, Sascha Hornauer, and Fabien Moutarde. Gri: General reinforced imitation and its application to vision-based autonomous driving. arXiv preprint arXiv:2111.08575, 2021. 2", + "[9] Chenyi Chen, Ari Seff, Alain L. Kornhauser, and Jianxiong Xiao. DeepDriving: Learning affordance for direct perception in autonomous driving. In ICCV, 2015. 2, 4", + "[10] Dian Chen, Vladlen Koltun, and Philipp Krahenbuhl. Learning to drive from a world on rails. In ICCV, 2021. 2, 7", + "[11] Dian Chen and Philipp Krahenbuhl. Learning from all vehicles. In CVPR, 2022. 1, 2, 3, 4, 5, 6, 7", + "[12] Defang Chen, Jian-Ping Mei, Yuan Zhang, Can Wang, Zhe Wang, Yan Feng, and Chun Chen. Cross-layer distillation with semantic calibration. In AAAI, 2021. 2", + "[13] Dian Chen, Brady Zhou, Vladlen Koltun, and Philipp Krahenbuhl. Learning by cheating. In CoRL, 2020. 1, 2, 3, 4, 6, 7", + "[14] Li Chen, Chonghao Sima, Yang Li, Zehan Zheng, Jiajie Xu, Xiangwei Geng, Hongyang Li, Conghui He, Jianping Shi, Yu Qiao, et al. Persformer: 3d lane detection via perspective transformer and the openlane benchmark. arXiv preprint arXiv:2203.11089, 2022. 2, 5", + "[15] Kashyap Chitta, Aditya Prakash, and Andreas Geiger. Neat: Neural attention fields for end-to-end autonomous driving. In ICCV, 2021. 7", + "[16] Kashyap Chitta, Aditya Prakash, Bernhard Jaeger, Zehao Yu, Katrin Renz, and Andreas Geiger. Transfuser: Imitation with transformer-based sensor fusion for autonomous driving. arXiv preprint arXiv:2205.15997, 2022. 3, 4, 5, 6, 7", + "[17] Inseop Chung, SeongUk Park, Jangho Kim, and Nojun Kwak. Feature-map-level online adversarial knowledge distillation. In ICML, 2020. 2", + "[18] Felipe Codevilla, Matthias Muiller, Antonio López, Vladlen Koltun, and Alexey Dosovitskiy. End-to-end driving via conditional imitation learning. In ICRA, 2018. 3, 4" + ], + "bbox": [ + 78, + 114, + 468, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[19] Felipe Codevilla, Eder Santana, Antonio M López, and Adrien Gaidon. Exploring the limitations of behavior cloning for autonomous driving. In ICCV, 2019. 4, 6", + "[20] Pim de Haan, Dinesh Jayaraman, and Sergey Levine. Causal confusion in imitation learning. In NeurIPS, 2019. 1", + "[21] Alexey Dosovitskiy, German Ros, Felipe Codevilla, Antonio Lopez, and Vladlen Koltun. CARLA: An open urban driving simulator. In CoRL, 2017. 2, 3, 4, 6", + "[22] James J Gibson. The ecological approach to visual perception: classic edition. Psychology press, 2014. 4", + "[23] Jianping Gou, Baosheng Yu, Stephen J Maybank, and Dacheng Tao. Knowledge distillation: A survey. In IJCV, 2021. 1, 3", + "[24] Sheng Guo, Weilin Huang, Haozhi Zhang, Chenfan Zhuang, Dengke Dong, Matthew R Scott, and Dinglong Huang. Curriculumnet: Weakly supervised learning from large-scale web images. In ECCV, 2018. 2", + "[25] Saurabh Gupta, James Davidson, Sergey Levine, Rahul Sukthankar, and Jitendra Malik. Cognitive mapping and planning for visual navigation. In CVPR, 2017. 2", + "[26] Peter E Hart, Nils J Nilsson, and Bertram Raphael. A formal basis for the heuristic determination of minimum cost paths. IEEE transactions on Systems Science and Cybernetics, 1968, 3", + "[27] He He, Jason Eisner, and Hal Daume. Imitation learning by coaching. In NeurIPS, 2012. 1, 2, 3", + "[28] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, 2016. 4", + "[29] Byeongho Heo, Jeesoo Kim, Sangwoo Yun, Hyojin Park, No-jun Kwak, and Jin Young Choi. A comprehensive overhaul of feature distillation. In ICCV, 2019. 2", + "[30] Roger Hiemstra and Burton Sisco. Individualizing Instruction. Making Learning Personal, Empowering, and Successful. 1990. 2", + "[31] Geoffrey Hinton, Oriol Vinyals, Jeff Dean, et al. Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531, 2(7), 2015. 2", + "[32] Anthony Hu, Zak Murez, Nikhil Mohan, Sofia Dudas, Jeffrey Hawke, Vijay Badrinarayanan, Roberto Cipolla, and Alex Kendall. Fiery: Future instance prediction in bird's-eye view from surround monocular cameras. In ICCV, 2021. 2, 3, 5", + "[33] Peiyun Hu, Aaron Huang, John Dolan, David Held, and Deva Ramanan. Safe local motion planning with self-supervised freespace forecasting. In CVPR, 2021. 7", + "[34] Zanming Huang, Zhongkai Shangguan, Jimuyang Zhang, Gilad Bar, Matthew Boyd, and Eshed Ohn-Bar. ASSISTER: Assistive navigation via conditional instruction generation. In ECCV, 2022. 3", + "[35] Zehao Huang and Naiyan Wang. Like what you like: Knowledge distill via neuron selectivity transfer. arXiv preprint arXiv:1707.01219, 2017. 2", + "[36] Mingi Ji, Byeongho Heo, and Sungrae Park. Show, attend and distill: Knowledge distillation via attention-based feature matching. In AAAI, 2021. 2" + ], + "bbox": [ + 501, + 92, + 890, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "7813", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[37] Lu Jiang, Zhengyuan Zhou, Thomas Leung, Li-Jia Li, and Li Fei-Fei. Mentornet: Learning data-driven curriculum for very deep neural networks on corrupted labels. In ICML, 2018. 2", + "[38] Jangho Kim, SeongUk Park, and Nojun Kwak. Paraphrasing complex network: Network compression via factor transfer. In NeurIPS, 2018. 2", + "[39] Tom Kocmi and Ondrej Bojar. Curriculum learning and minibatch bucketing in neural machine translation. arXiv preprint arXiv:1707.09533, 2017. 2", + "[40] Animesh Koratana, Daniel Kang, Peter Bailis, and Matei Zaharia. Lit: Learned intermediate representation training for model compression. In ICML, 2019. 2", + "[41] M Kumar, Benjamin Packer, and Daphne Koller. Self-paced learning for latent variable models. In NeurIPS, 2010. 2", + "[42] Ravi Kumar Kushawaha, Saurabh Kumar, Biplab Banerjee, and Rajbabu Velmurugan. Distilling spikes: Knowledge distillation in spiking neural networks. In ICPR, 2021. 2", + "[43] Sergey Levine, Chelsea Finn, Trevor Darrell, and Pieter Abbeel. End-to-end training of deep visuomotor policies. The Journal of Machine Learning Research, 17(1):1334-1373, 2016. 1", + "[44] Dong Li, Jia-Bin Huang, Yali Li, Shengjin Wang, and Ming-Hsuan Yang. Weakly supervised object localization with progressive domain adaptation. In CVPR, 2016. 2", + "[45] Guohao Li, Matthias Mueller, Vincent Casser, Neil Smith, Dominik L Michels, and Bernard Ghanem. Oil: Observational imitation learning. In RSS, 2019. 2", + "[46] Yang Liu, Wei Zhang, and Jun Wang. Source-free domain adaptation for semantic segmentation. In CVPR, 2021. 2", + "[47] Takahiro Miki, Joonho Lee, Jemin Hwangbo, Lorenz Wellhausen, Vladlen Koltun, and Marco Hutter. Learning robust perceptive locomotion for quadrupedal robots in the wild. Science Robotics, 7(62), 2022. 1", + "[48] Matthias Müller, Alexey Dosovitskiy, Bernard Ghanem, and Vladlen Koltun. Driving policy transfer via modularity and abstraction. In CoRL, 2018. 3, 4", + "[49] Eshed Ohn-Bar, Aditya Prakash, Aseem Behl, Kashyap Chitta, and Andreas Geiger. Learning situational driving. In CVPR, 2020. 6", + "[50] Takayuki Osa, Joni Pajarinen, Gerhard Neumann, J Andrew Bagnell, Pieter Abbeel, Jan Peters, et al. An algorithmic perspective on imitation learning. Foundations and Trends® in Robotics, 2018. 2, 3", + "[51] Despoina Paschalidou, Angelos Katharopoulos, Andreas Geiger, and Sanja Fidler. Neural parts: Learning expressive 3D shape abstractions with invertible neural networks. In CVPR, 2021. 6", + "[52] Nikolaos Passalis and Anastasios Tefas. Learning deep representations with probabilistic knowledge transfer. In ECCV, 2018. 2", + "[53] Jonah Philion and Sanja Fidler. Lift, splat, shoot: Encoding images from arbitrary camera rigs by implicitly unprojecting to 3D. In ECCV, 2020. 2, 5", + "[54] Dean A Pomerleau. ALVINN: An autonomous land vehicle in a neural network. In NeurIPS, 1989. 1, 2" + ], + "bbox": [ + 78, + 90, + 468, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[55] Adriana Romero, Nicolas Ballas, Samira Ebrahimi Kahou, Antoine Chassang, Carlo Gatta, and Yoshua Bengio. Fitnets: Hints for thin deep nets. arXiv preprint arXiv:1412.6550, 2014. 2", + "[56] Stéphane Ross, Geoffrey Gordon, and Drew Bagnell. A reduction of imitation learning and structured prediction to no-regret online learning. In AISTATS, 2011. 2, 3", + "[57] Avishkar Saha, Oscar Mendez, Chris Russell, and Richard Bowden. Translating images into maps. In ICRA, 2022. 3", + "[58] Axel Sauer, Nikolay Savinov, and Andreas Geiger. Conditional affordance learning for driving in urban environments. In CoRL, 2018. 4", + "[59] Miaojing Shi and Vittorio Ferrari. Weakly supervised object localization using size estimates. In ECCV, 2016. 2", + "[60] Suraj Srinivas and Francois Fleuret. Knowledge transfer with Jacobian matching. In ICML, 2018. 2", + "[61] Yuxing Tang, Xiaosong Wang, Adam P Harrison, Le Lu, Jing Xiao, and Ronald M Summers. Attention-guided curriculum learning for weakly supervised classification and localization of thoracic diseases on chest radiographs. In MLMI, 2018. 2", + "[62] Marin Toromanoff, Emilie Wirbel, and Fabien Moutarde. End-to-end model-free reinforcement learning for urban driving using implicit affordances. In CVPR, 2020. 2", + "[63] Radu Tudor Ionescu, Bogdan Alexe, Marius Leordeanu, Marius Popescu, Dim P Papadopoulos, and Vittorio Ferrari. How hard can it be? estimating the difficulty of visual search in an image. In CVPR, 2016. 2", + "[64] Rachel R Van Der Stuyf. Scaffolding as a teaching strategy. Adolescent learning and development, 2002. 2", + "[65] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In NeurIPS, 2017. 5", + "[66] Antonio Visioli. Practical PID control. Springer Science & Business Media, 2006. 3", + "[67] Tao Wang, Li Yuan, Xiaopeng Zhang, and Jiashi Feng. Distilling object detectors with fine-grained feature imitation. In CVPR, 2019. 2", + "[68] Wenhui Wang, Hangbo Bao, Shaohan Huang, Li Dong, and Furu Wei. Minilmv2: Multi-head self-attention relation distillation for compressing pretrained transformers. arXiv preprint arXiv:2012.15828, 2020. 2", + "[69] Wei Wang, Isaac Caswell, and Ciprian Chelba. Dynamically composing domain-data selection with clean-data selection by \"co-curricular learning\" for neural machine translation. arXiv preprint arXiv:1906.01130, 2019. 2", + "[70] Chuan Wen, Jierui Lin, Trevor Darrell, Dinesh Jayaraman, and Yang Gao. Fighting copycat agents in behavioral cloning from observation histories. In NeurIPS, 2020. 1", + "[71] Penghao Wu, Xiaosong Jia, Li Chen, Junchi Yan, Hongyang Li, and Yu Qiao. Trajectory-guided control prediction for end-to-end autonomous driving: A simple yet strong baseline. arXiv preprint arXiv:2206.08129, 2022. 1, 2, 3, 6, 7, 8", + "[72] Liuyu Xiang, Guiguang Ding, and Jungong Han. Learning from multiple experts: Self-paced knowledge distillation for long-tailed classification. In ECCV, 2020. 2" + ], + "bbox": [ + 501, + 92, + 890, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "7814", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[73] Canwen Xu, Wangchunshu Zhou, Tao Ge, Furu Wei, and Ming Zhou. Bert-of-theseus: Compressing bert by progressive module replacing. arXiv preprint arXiv:2002.02925, 2020. 2", + "[74] Jing Yang, Brais Martinez, Adrian Bulat, and Georgios Tzimiropoulos. Knowledge distillation via adaptive instance normalization. arXiv preprint arXiv:2003.04289, 2020. 2", + "[75] Junho Yim, Donggyu Joo, Jihoon Bae, and Junmo Kim. A gift from knowledge distillation: Fast optimization, network minimization and transfer learning. In CVPR, 2017. 2", + "[76] Jimuyang Zhang and Eshed Ohn-Bar. Learning by watching. In CVPR, 2021. 2, 3", + "[77] Jimuyang Zhang, Ruizhao Zhu, and Eshed Ohn-Bar. SelfD: Self-learning large-scale driving policies from the web. In CVPR, 2022. 2, 7", + "[78] Pan Zhang, Bo Zhang, Ting Zhang, Dong Chen, and Fang Wen. Robust mutual learning for semi-supervised semantic segmentation. arXiv preprint arXiv:2106.00609, 2021. 2", + "[79] Zhejun Zhang, Alexander Liniger, Dengxin Dai, Fisher Yu, and Luc Van Gool. End-to-end urban driving by imitating a reinforcement learning coach. In ICCV, 2021. 1, 2, 3, 4, 5, 6, 7, 8", + "[80] Albert Zhao, Tong He, Yitao Liang, Haibin Huang, Guy Van den Broeck, and Stefano Soatto. Lates: Latent space distillation for teacher-student driving policy learning. arXiv preprint arXiv:1912.02973, 2019. 2", + "[81] Albert Zhao, Tong He, Yitao Liang, Haibin Huang, Guy Van den Broeck, and Stefano Soatto. Sam: Squeeze-and-mimic networks for conditional visual driving policy learning. CoRL, 2020. 1, 2, 3", + "[82] Brady Zhou and Philipp Krahenbuhl. Cross-view transformers for real-time map-view semantic segmentation. In CVPR, 2022. 2", + "[83] Brady Zhou, Philipp Krähenbuhl, and Vladlen Koltun. Does computer vision matter for action? Science Robotics, 4(30), 2019. 2", + "[84] Yikai Zhou, Baosong Yang, Derek F Wong, Yu Wan, and Lidia S Chao. Uncertainty-aware curriculum learning for neural machine translation. In ACL, 2020. 2", + "[85] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable detr: Deformable transformers for end-to-end object detection. arXiv preprint arXiv:2010.04159, 2020. 5" + ], + "bbox": [ + 78, + 90, + 468, + 700 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "7815", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2023/Coaching a Teachable Student/67002b28-2cb3-47c6-8f0a-06c8bb56b862_model.json b/2023/Coaching a Teachable Student/67002b28-2cb3-47c6-8f0a-06c8bb56b862_model.json new file mode 100644 index 0000000000000000000000000000000000000000..c85377cf2eed849a70cb9e29071ea551f2ed56bc --- /dev/null +++ b/2023/Coaching a Teachable Student/67002b28-2cb3-47c6-8f0a-06c8bb56b862_model.json @@ -0,0 +1,2378 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.331, + 0.131, + 0.642, + 0.154 + ], + "angle": 0, + "content": "Coaching a Teachable Student" + }, + { + "type": "text", + "bbox": [ + 0.268, + 0.181, + 0.704, + 0.218 + ], + "angle": 0, + "content": "Jimuyang Zhang Zanming Huang Eshed Ohn-Bar Boston University" + }, + { + "type": "text", + "bbox": [ + 0.324, + 0.22, + 0.644, + 0.234 + ], + "angle": 0, + "content": "{zhangjim, huangtom, eohnbar}@bu.edu" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.269, + 0.314, + 0.285 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.301, + 0.474, + 0.681 + ], + "angle": 0, + "content": "We propose a novel knowledge distillation framework for effectively teaching a sensorimotor student agent to drive from the supervision of a privileged teacher agent. Current distillation for sensorimotor agents methods tend to result in suboptimal learned driving behavior by the student, which we hypothesize is due to inherent differences between the input, modeling capacity, and optimization processes of the two agents. We develop a novel distillation scheme that can address these limitations and close the gap between the sensorimotor agent and its privileged teacher. Our key insight is to design a student which learns to align their input features with the teacher's privileged Bird's Eye View (BEV) space. The student then can benefit from direct supervision by the teacher over the internal representation learning. To scaffold the difficult sensorimotor learning task, the student model is optimized via a student-paced coaching mechanism with various auxiliary supervision. We further propose a high-capacity imitation learned privileged agent that surpasses prior privileged agents in CARLA and ensures the student learns safe driving behavior. Our proposed sensorimotor agent results in a robust image-based behavior cloning agent in CARLA, improving over current models by over \\(20.6\\%\\) in driving score without requiring LiDAR, historical observations, ensemble of models, on-policy data aggregation or reinforcement learning." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.694, + 0.21, + 0.71 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.719, + 0.47, + 0.886 + ], + "angle": 0, + "content": "Learning internal representations for making intricate driving decisions from images involves a complex optimization task [4, 43, 54]. The inherent challenge for end-to-end training of driving agents lies in the immense complexity of learning to map high-dimensional visual observations into general and safe navigational decisions [13, 20, 70]. Even given millions of training examples [4], today's agents still fail to reliably learn an internal representation that can be used for robust processing of complex visual scenarios (e.g., dense urban settings with intricate layouts and dynamic obstacles) in a safe and task-driven manner [13, 20, 70]." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.886, + 0.47, + 0.902 + ], + "angle": 0, + "content": "To ease the challenging sensorimotor training task, re" + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.27, + 0.891, + 0.432 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.438, + 0.895, + 0.577 + ], + "angle": 0, + "content": "Figure 1. Effective Knowledge Distillation for Sensorimotor Agents. Our proposed CaT (Coaching a Teachable student) framework enables highly effective knowledge transfer between a privileged teacher and a sensorimotor (i.e., image-based) student. An alignment module learns to transform image-based features to the teacher's BEV feature space, where the student can then leverage extensive and direct supervision on its learned intermediate representations. The student model is optimized via a coaching mechanism with extensive auxiliary supervision in order to further scaffold the difficult sensorimotor learning task." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.595, + 0.893, + 0.852 + ], + "angle": 0, + "content": "cent approaches decompose the task into stages, e.g., by first training a high-capacity privileged network with complete knowledge of the world and distilling its knowledge into a less capable vision-based student network [11, 13, 23, 47, 79]. However, due to the inherent differences between the inputs and architectures of the two agents, current methods rely on limited supervisory mechanisms from the teacher, i.e., exclusively through the teacher's output [11, 13] or knowledge distillation of a single final fully-connected layer [71, 79, 81]. Moreover, the privileged teacher's demonstration targets may be noisy or difficult for the student to imitate, given the limited perspective [27]. In this work, we sought to develop a more effective knowledge distillation paradigm for training a sensorimotor agent to drive. Our key insight is to enable more extensive supervision from the teacher by reducing the gap between internal modeling and learning capabilities between the two agents." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.856, + 0.894, + 0.901 + ], + "angle": 0, + "content": "Our proposed approach for holistic knowledge distillation is informed by human instruction, which often involves structured supervision in addition to high-level demonstra" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7805" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.471, + 0.258 + ], + "angle": 0, + "content": "tions, e.g., providing various hints to scaffold information in a way that the student can better understand [30]. When teaching others new and challenging skills, i.e., where a student may not be able to replicate the demonstration such as riding a bicycle or driving a vehicle, teachers may provide additional supervision regarding the underlying task structure and their own internal reasoning [64]. Analogously to our settings, the privileged teaching agent can potentially provide richer supervision when teaching a limited capacity sensorimotor student, i.e., through more careful and direct guidance of the underlying representation learning." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.258, + 0.471, + 0.591 + ], + "angle": 0, + "content": "In our work, we introduce CaT, a novel method for teaching a sensorimotor student to drive using supervision from a privileged teacher. Our key insights are threefold: 1) **Effective Teacher:** We propose to incorporate explicit safety-aware cues into the BEV space that facilitate a surprisingly effective teacher agent design. While prior privileged agents struggle to learn to drive in complex urban driving scenes, we demonstrate our learned agent to match expert-level decision-making. 2) **Teachable Student via Alignment:** An IPM-based transformer alignment module can facilitate direct distillation of most of the teacher's features and better guide the student learning process. 3) **Student-paced Coaching:** A coaching mechanism for managing difficult samples can scaffold knowledge and lead to improved model optimization by better considering the ability of the student. By holistically tackling the complex knowledge distillation task with extensive teacher and auxiliary supervision, we are able to train a state-of-the-art image-based agent in CARLA [21]. Through ablation studies, input design, and interpretable feature analysis, we also provide critical insights into current limitations in learning robust and generalized representations for driving agents." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.602, + 0.22, + 0.618 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.629, + 0.47, + 0.901 + ], + "angle": 0, + "content": "Knowledge Distillation: We study knowledge distillation methods originally developed for model compression and acceleration [12,17,29,31,35,36,52,55,74,75] in the context of training sensorimotor. While approaches for feature distillation have been minimally explored in this context, we are motivated by their success in other domains, including image classification [2,38,40,42,60], object detection [67], semantic segmentation [46, 78], and natural language processing [3,68,73]. Yet, applying such techniques is not trivial given the drastically differing inputs between the privileged agent and student and the overwhelming sensorimotor task. Consequently, driving policy distillation methods only provide supervision either from the teacher's output [11,13] or the features of a single fully-connected layer [71,79,81], which (we hypothesize) does not provide sufficient hints to guide the student training. In this work, we provide novel mechanisms to close this gap and enable extensive supervision through deep feature distillation." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.895, + 0.38 + ], + "angle": 0, + "content": "Imitation Learning to Drive: Recent approaches in imitation learning (IL) to drive can be traced to Pomerleau [54]. Recently, more elaborate IL-based approaches for driving have emerged [9, 10, 25, 45, 50, 76, 77, 80, 83]. Specifically, decomposing the imitation learning task into two stages, i.e., by first learning a privileged agent through behavior cloning [11, 13] or reinforcement learning [8, 62, 71, 79] and then training the sensorimotor agent to mimic the output of the privileged agent. Our study is motivated by such approaches, yet we explore the benefits of increased supervision by the privileged agent. Moreover, prior privileged agents produce noisy and sub-optimal demonstration supervision in complex urban scenarios (even when optimized with RL, as demonstrated by our analysis). While most current studies employ the RL-based teacher of [79], this can be problematic and inefficient in safety-critical conditions in the real-world. In our work, we introduce a novel privileged teacher which can greatly surpass prior agents while only relying on offline behavior cloning." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.39, + 0.895, + 0.619 + ], + "angle": 0, + "content": "Intermediate Representation for Driving: Learning an effective 3D scene representation is crucial for safe autonomous driving. Researchers may obtain such representations by lifting the image to 3D using estimated depth and projecting the frustums into a BEV grid [32, 53]. Alternatively, transformer-based architectures can also enable mapping a camera image to the BEV space [14, 82], i.e., by attending to image-based information when populating a BEV-sampled grid. Related to our work is the study of Chen et al. [11], which learns a BEV representation from RGB image and LiDAR input. A motion planner is then used to generate future waypoints from the BEV. As this introduces significant challenges, we propose to leverage an IPM-based alignment module which can better structure to image-to-BEV projection task." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.629, + 0.895, + 0.901 + ], + "angle": 0, + "content": "Curriculum and Self-paced Learning: Approaches for curriculum learning methods often structure learning from various difficulty samples [5,24,37,39,44,59,61,63,69,84]. To better consider the challenging sensorimotor task and limited capacity of the student, we adopt a curriculum (i.e., an imitation coach), with progressively more challenging samples. He et al. [27] also proposes training a coach across iterations of on-policy data collection [56]. In contrast, we implement and demonstrate the benefit of a coach over training iterations without a data aggregation part. This can facilitate more efficient model training. The approach is motivated by self-paced learning [41, 72], where the model selects easy samples dynamically at each iteration based on a defined loss. In contrast to such studies, we only smooth the targets of the hard samples (based on the student), instead of discarding them. We empirically find this to result in improved coaching, potentially due to improved scaffolding of the difficult samples." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.956 + ], + "angle": 0, + "content": "7806" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.078, + 0.09, + 0.169, + 0.105 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.115, + 0.473, + 0.283 + ], + "angle": 0, + "content": "Our goal is to decompose the challenging sensorimotor learning task and ease model optimization through effective supervision from a teacher agent. We first formulate the problem of privileged agent distillation in Sec. 3.1. Next, we address current limitations in privileged agent design to train a robust imitation learned teacher in Sec. 3.2. We then propose a transformer-based student architecture to resolve image-to-BEV feature alignment (Sec. 3.3). Finally, we train the sensorimotor student via extensive supervision and knowledge distillation (Sec. 3.4) via a progressive (i.e., coaching) mechanism (Sec. 3.5)." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.291, + 0.212, + 0.306 + ], + "angle": 0, + "content": "3.1. Formulation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.314, + 0.473, + 0.542 + ], + "angle": 0, + "content": "Driving can be formulated as a sequential decision-making problem over a set of sensory observations \\(\\mathcal{X}^s\\), a set of actions (motor commands) \\(\\mathcal{V}\\), and a policy function \\(f_{\\theta}^{s} \\colon \\mathcal{X}^{s} \\to \\mathcal{V}\\) for mapping observations to actions at each time step [27,50,56]. We consider the task of learning a sensorimotor student agent via knowledge distillation [13, 23] from a teacher agent \\(f_{\\psi}^{t} \\colon \\mathcal{X}^{t} \\to \\mathcal{V}\\) with privileged access to environmental information, e.g., ground-truth lanes, obstacles, and traffic light states. We parameterize the two agents as neural networks using weights \\([\\pmb{\\theta}, \\pmb{\\psi}] \\in \\mathbb{R}^d\\), and denote \\(\\mathcal{F}^s(\\cdot; \\pmb{\\theta})\\) and \\(\\mathcal{F}^t(\\cdot; \\pmb{\\psi})\\) to be the feature maps for the sensorimotor student network and teacher network, respectively. Given a dataset \\(\\mathcal{D}\\) comprising sensory and privileged observations and a loss function \\(\\mathcal{L}\\), the student can be optimized from the teacher using" + }, + { + "type": "equation", + "bbox": [ + 0.1, + 0.55, + 0.47, + 0.575 + ], + "angle": 0, + "content": "\\[\n\\underset {\\boldsymbol {\\theta}} {\\operatorname {a r g m i n}} \\mathbb {E} _ {(\\mathbf {x} ^ {s}, \\mathbf {x} ^ {t}) \\sim \\mathcal {D}} [ \\mathcal {L} (\\mathcal {F} ^ {s} (\\mathbf {x} ^ {s}; \\boldsymbol {\\theta}), \\mathcal {F} ^ {t} (\\mathbf {x} ^ {t}; \\boldsymbol {\\psi})) ] \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.584, + 0.473, + 0.902 + ], + "angle": 0, + "content": "where the loss may be defined over a final layer, i.e., to match the output of the teacher [11, 13], or over internal layers as well, i.e., feature distillation [71, 79, 81]. Given the challenging end-to-end sensorimotor learning task, the role of the teacher is to provide effective supervision to the student, i.e., informative targets in Eqn. 1. However, as the two agents operate from drastically differing inputs, the knowledge transfer from the teacher to the student can become ineffective. Specifically, the task of the student is significantly more challenging than that of the teacher, i.e., due to inherent noise and uncertainty. Indeed, there is currently a substantial gap between the performance of current sensorimotor learning agents and their teaching agents, even assuming access to expensive and high-quality 3D observations such as LiDAR [11, 13, 16] (not used in our work). The teaching process should also ideally consider any limited capacity of the student to facilitate effective optimization and learning [27]. Moreover, learning an optimal and robust teacher in itself can be a challenging task [13], even assuming the simplified learning task such as solved perception, as we demonstrate in Sec. 4 and addressed in Sec. 3.2." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.895, + 0.56 + ], + "angle": 0, + "content": "Problem Setup and Agent Observations: We develop our CaT framework in the context of the CARLA simulator [21]. Our objective is to train a goal-conditional sensorimotor agent [16, 18] for mapping image and goal observations to vehicle throttle and steering control [13, 48]. We follow standard conditional imitation learning [11, 16] and assume access to three (non-overlapping) RGB camera views \\(\\mathbf{I} = [\\mathbf{I}_0, \\mathbf{I}_1, \\mathbf{I}_2] \\in \\mathbb{R}^{W \\times H \\times 3}\\), a categorical navigational command \\(c \\in \\{1, \\dots, 6\\}\\) (i.e., turn left, turn right, follow, forward, left lane changing and right lane changing) and an intermediate noisy goal \\(\\mathbf{g} \\in \\mathbb{R}^2\\) sampled from a GNSS (Global Navigation Satellite System) [11]. We note that both the command and GNSS observations can be easily obtained by today's vehicles, i.e., through generation of an \\(\\mathrm{A}^*\\) plan [26] from a static map and coarse positioning system. The GNSS goals are sampled every 50-100 meters and reflect real-world measurement errors [11]. We compute a BEV \\(\\mathbf{B} \\in \\{0, 1\\}^{W_B \\times H_B \\times C_B}\\) by rendering privileged (i.e., ground-truth) information from the underlying simulation, including 3D location and state of lanes, pedestrians, vehicles, and traffic lights [79]. While prior work has generally leveraged a standardized BEV representation [13, 32, 34, 57, 76], we find its design to be crucial to training an optimal teacher policy as will be discussed in Sec 3.2. We let \\(\\mathbf{x}^s = (\\mathbf{I}, \\mathbf{g}, c) \\in \\mathcal{X}^s\\) and \\(\\mathbf{x}^t = (\\mathbf{B}, \\mathbf{g}, c) \\in \\mathcal{X}^t\\) be the student and teacher observations, respectively. Given these observations at each time step, the agent learns to predict 10 future 2D waypoints in top-down vehicle coordinates for the next 2.5 seconds of driving. These are then given to a lateral and longitudinal PID controllers [13, 48, 66] to generate the final action." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.561, + 0.895, + 0.651 + ], + "angle": 0, + "content": "Given this formulation, we next discuss our teacher agent design and training process. In particular, we aim to learn a privileged agent which can not only effectively solve the goal-oriented navigation task, but also facilitate effective distillation and coaching of a sensorimotor student policy, as will be discussed in Sec. 3.4." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.665, + 0.771, + 0.681 + ], + "angle": 0, + "content": "3.2. Learning an Effective Teacher" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.689, + 0.895, + 0.901 + ], + "angle": 0, + "content": "In our formulation, learning a student agent begins with training an effective privileged agent \\( f_{\\psi}^{t} \\). However, we find this itself to be a complex task. Even with complete knowledge of perfect BEV perception, current privileged agents are suboptimal, significantly under-performing CARLA's built-in autopilot [13, 79]. In this section, we uncover the underlying reason for this under-performance through the BEV design. Informed by this key finding, we train a highly effective imitation-learned teacher that surpasses prior privileged agents and matches expert-level performance in CARLA without requiring extensive data collection [56] or reinforcement learning [79]. Our proposed privileged agent can then enable effective supervision of the student in Sec. 3.3." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7807" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.09, + 0.895, + 0.313 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.324, + 0.896, + 0.409 + ], + "angle": 0, + "content": "Figure 2. Overview of Architecture and Distillation Process of CaT. We first train a high-capacity privileged teacher agent (colored in red) which takes an augmented BEV representation as input, and produces a set of future waypoints conditioned on command and short-term goal. Our teachable student agent (colored in blue) maps the image features to the BEV space using an alignment module, which is trained by distillation of intermediate features (arrows colored in blue) and final output (arrow colored in green). Additionally, we leverage supervision from auxiliary segmentation (arrow colored in red) and command prediction tasks (arrow colored in purple) interlaced throughout the network to facilitate the learning of task-relevant features." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.434, + 0.473, + 0.784 + ], + "angle": 0, + "content": "Teacher Training via Direct Expert Imitation: In imitation learning, an oracle (i.e., expert driver) \\( f^{*} \\colon \\mathcal{X}^{t} \\to \\mathcal{Y} \\) is defined for demonstrating optimal actions \\( \\mathbf{y}^{*} \\). In CARLA [21], the expert is defined using longitudinal and lateral PID controllers which are carefully tuned and augmented by a collection of manually specified rules for handling diverse scenarios (e.g., sudden braking) over the underlying simulation state. The designed expert can then be used to generate trajectories as supervised data for learning the privileged policy function \\( f_{\\psi}^{t} \\) with behavior cloning [18, 19, 48]. Our privileged agent is a ResNet-18 model [28], as shown in Fig. 2 (further increasing modeling capacity was not found to be beneficial). The ResNet model is followed by a GRU-based conditional waypoint predictor [11, 16]. Due to the complexity of our dense urban navigation task, e.g., with multi-lane roads, intersections, and merging, we incrementally refine the waypoints using two GRUs within each conditional branch. The first GRU regresses the preliminary set of waypoints via sequential waypoints directly from the embedded features. Subsequently, the second GRU takes the predicted waypoints together with the short-term goal and embedded features, to produce a refined set of waypoint targets." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.795, + 0.47, + 0.903 + ], + "angle": 0, + "content": "Even assuming perfect perception, we find it difficult to train privileged agents that exhibit robust planning behavior in complex urban driving scenarios. Specifically, when using the most challenging CARLA benchmark (Longest6 [16], also detailed in Sec 4) using the standard BEV representation of Chen et al. [13] leads to \\(26\\%\\) driving score compared to \\(72\\%\\) by the rule-based expert. The rein" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.434, + 0.895, + 0.618 + ], + "angle": 0, + "content": "forcement learning agent of Zhang et al. [79] achieves \\(60\\%\\) on this task. We sought to explore the limits of behavior cloning for this task, as a noisy and sub-optimal teacher can hinder the training of the student model. In particular, we hypothesize that the reason for the poor performance lies in the increased complexity of the BEV state representation in dense scenarios. In such scenarios, learning to extract task-relevant details can become more challenging. To address this issue, we propose to introduce additional channels (akin to affordances [9,22,58]) that can more easily translate into safety-critical decision-making, simplify the learning task, and provide an expert-level teaching agent." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.629, + 0.896, + 0.903 + ], + "angle": 0, + "content": "BEV with Safety Hints: Our BEV encodes state information into distinct channels, including drivable road regions, the desired route based on the command and layout, lane markings, and dynamic obstacles (see Fig. 3). Given the low performance of prior privileged agents, in particular around dynamic objects, the agent could benefit from more explicit safety-driven cues in the BEV. We propose to add two types of channels of (1) predicted agents' future and (2) entity attention. First, we utilize a kinematics bicycle model [16] in order to efficiently predict future trajectories of dynamic objects. This enables us to iteratively predict and represent short-term future position, orientation, and speed of agents (our supplementary contains additional information regarding this process). Secondly, we encode an explicit attention channel for highlighting potential future infractions, i.e., based on the kinematics bicycle model applied to the ego-vehicle. While this assumes that the current vehicle throttle and steering remain unchanged, the entity" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "7808" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.107, + 0.235, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.104, + 0.246, + 0.216, + 0.261 + ], + "angle": 0, + "content": "(a) Baseline [79]" + }, + { + "type": "image", + "bbox": [ + 0.236, + 0.108, + 0.388, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.242, + 0.246, + 0.384, + 0.26 + ], + "angle": 0, + "content": "(b) With Safety Hints" + }, + { + "type": "image", + "bbox": [ + 0.399, + 0.111, + 0.486, + 0.156 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.413, + 0.16, + 0.473, + 0.173 + ], + "angle": 0, + "content": "(c) Road" + }, + { + "type": "image", + "bbox": [ + 0.399, + 0.174, + 0.486, + 0.241 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.393, + 0.246, + 0.492, + 0.261 + ], + "angle": 0, + "content": "(g) Pedestrians" + }, + { + "type": "image", + "bbox": [ + 0.512, + 0.09, + 0.6, + 0.157 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.497, + 0.16, + 0.615, + 0.173 + ], + "angle": 0, + "content": "(d) Desired Route" + }, + { + "type": "image", + "bbox": [ + 0.512, + 0.174, + 0.598, + 0.241 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.246, + 0.613, + 0.261 + ], + "angle": 0, + "content": "(h) Traffic Lights" + }, + { + "type": "image", + "bbox": [ + 0.641, + 0.09, + 0.728, + 0.156 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.634, + 0.16, + 0.735, + 0.173 + ], + "angle": 0, + "content": "(e) Lane Marks" + }, + { + "type": "image", + "bbox": [ + 0.641, + 0.174, + 0.728, + 0.241 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.619, + 0.246, + 0.746, + 0.261 + ], + "angle": 0, + "content": "(i) Agent Forecast" + }, + { + "type": "image", + "bbox": [ + 0.779, + 0.091, + 0.866, + 0.156 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.783, + 0.159, + 0.861, + 0.173 + ], + "angle": 0, + "content": "(f) Vehicles" + }, + { + "type": "image", + "bbox": [ + 0.779, + 0.174, + 0.866, + 0.241 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.753, + 0.246, + 0.772, + 0.261 + ], + "angle": 0, + "content": "(j)" + }, + { + "type": "image_caption", + "bbox": [ + 0.779, + 0.246, + 0.887, + 0.261 + ], + "angle": 0, + "content": "Entity Attention" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.267, + 0.893, + 0.337 + ], + "angle": 0, + "content": "Figure 3. Visualization of Our Proposed BEV with Safety Hints. (a) and (b) visualize the baseline and our proposed BEV by encoding different channels into colors. Roads are shown in dark grey, Desired Route is shown in light grey, Lane Marks are shown in magenta, Vehicles are shown in blue with past trajectories shown in decreasing saturation with time into the past, Pedestrians are shown in light blue, Traffic Lights are shown in red or green according to its state, Agent Forecasts are shown in purple with decreasing saturation with time into the future, Entity Attention is shown in orange. (c) - (j) shows the different channels separately." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.35, + 0.469, + 0.469 + ], + "angle": 0, + "content": "attention channel helps the agent better reason over safety while correlating states with actions. Surprisingly, the additions of the safety hints channels provide a strong teacher with offline behavior cloning alone, even improving over the rule-based expert (Table 1, \\(73\\%\\) vs. \\(72\\%\\)). We note that many of the remaining failures are due to simulation timeout in crowded intersections. Our boosted teacher agent will be critical for effective distillation in Sec. 3.4." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.497, + 0.343, + 0.512 + ], + "angle": 0, + "content": "3.3. Learning a Teachable Student" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.526, + 0.469, + 0.692 + ], + "angle": 0, + "content": "Due to differing inputs and modeling capacities, it can be difficult to align the image-based student features and output with the BEV-based privileged teacher. We propose to alleviate misalignment by designing a teachable student network, i.e., a model that matches as much of the teacher architecture as possible. Specifically, we propose to improve teacher-student model alignment by leveraging an internal image-to-BEV alignment module. As shown in Fig. 2, the three residual blocks following the alignment module can consequently facilitate knowledge transfer via direct distillation of most of the teacher's features in Sec. 3.4." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.704, + 0.469, + 0.901 + ], + "angle": 0, + "content": "IPM-based Transformer Alignment Module: Explicitly mapping images to BEV space for downstream tasks can be challenging. While this can be achieved via depth estimation and lifting [32,53], we sought to incorporate an efficient and differentiable module which can effectively attend and reason over various features while mapping from front view to BEV. A transformer-based architecture [14,65] provides a natural choice as it can be optimized end-to-end with respect to the driving task while also carefully aligning features among spaces with arbitrary dimensionalities. As shown in Fig. 2, we sample queries \\(\\mathbf{Q}_{init} \\in \\mathbb{R}^{L_H \\times L_W \\times L_C}\\) using a spatial parameterization of the BEV space and process them using a self-attention module [14,65]. The output" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.35, + 0.818, + 0.365 + ], + "angle": 0, + "content": "of self-attention module \\(Q\\) can be formulated as" + }, + { + "type": "equation", + "bbox": [ + 0.576, + 0.369, + 0.892, + 0.403 + ], + "angle": 0, + "content": "\\[\n\\mathbf {Q} = \\operatorname {s o f t m a x} \\left(\\frac {\\mathbf {Q} _ {\\text {i n i t}} \\mathbf {K} _ {\\text {i n i t}} ^ {T}}{\\sqrt {d}}\\right) \\mathbf {V} _ {\\text {i n i t}} \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.408, + 0.892, + 0.5 + ], + "angle": 0, + "content": "where \\(\\mathbf{K}_{init}\\), \\(\\mathbf{V}_{init} \\in \\mathbb{R}^{L_H \\times L_W \\times L_C}\\) provide the key-value pairs and \\(d\\) is the dimension of the query. Subsequently, each query point \\(\\mathbf{q} \\in \\mathbf{Q}\\) is mapped to its corresponding reference point \\(\\mathbf{p} \\in \\mathbb{R}^3\\) (homogeneous coordinates) in the image features \\(\\mathcal{F}_{RGB} \\in \\mathbb{R}^{Z_H \\times Z_W \\times Z_C}\\) using an Inverse Perspective Mapping (IPM)" + }, + { + "type": "equation", + "bbox": [ + 0.622, + 0.503, + 0.89, + 0.518 + ], + "angle": 0, + "content": "\\[\n\\mathbf {p} = s \\mathbf {P} _ {k} \\mathbf {R} _ {k} (\\mathbf {q} - \\mathbf {t} _ {k}) \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.52, + 0.892, + 0.596 + ], + "angle": 0, + "content": "where \\(s\\) is a scale factor, \\(\\mathbf{P}_k\\in \\mathbb{R}^{3\\times 3}\\), \\(\\mathbf{R}_k\\in \\mathbb{R}^{3\\times 3}\\), and \\(\\mathbf{t}_k\\in \\mathbb{R}^3\\) are the \\(k^{th}\\) camera intrinsics, rotation and translation relative to the center of the ego-vehicle, respectively [6]. Note that we employ a different IPM mapping for each camera view to perform the projection." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.596, + 0.892, + 0.746 + ], + "angle": 0, + "content": "To populate the student's BEV features \\(\\mathcal{F}_{BEV} \\in \\mathbb{R}^{H_B \\times W_B \\times C_B}\\) (i.e., the output of the alignment module in Fig. 2), we leverage a deformable cross-attention mechanism based on deformable DETR [85]. This enables the network to attend to multiple regions around the reference point in the image features when transforming to BEV feature space. The reference point \\(\\mathbf{p}\\) and the learned deformable offsets can then be used as keys \\(\\mathbf{K}\\), with their corresponding features from the extracted images features the values \\(\\mathbf{V}\\), for a cross-attention module" + }, + { + "type": "equation", + "bbox": [ + 0.57, + 0.753, + 0.891, + 0.77 + ], + "angle": 0, + "content": "\\[\n\\mathcal {F} _ {B E V} = \\operatorname {D e f o r m A t t n} (\\mathbf {Q}, \\mathcal {F} _ {R G B}, \\mathbf {H}) \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.775, + 0.891, + 0.85 + ], + "angle": 0, + "content": "where \\(\\mathbf{H} = \\{\\mathbf{P}_k, \\mathbf{R}_k, t_k\\}_{k=1}^3\\) contains the combined multiview IPM parameters from Eqn. 3. In addition to the proposed alignment and teachable modules, we also incorporate prediction modules for auxiliary tasks (i.e., BEV prediction [16]), discussed in Sec. 3.4." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.892, + 0.901 + ], + "angle": 0, + "content": "GRU-based Waypoint Predictor: The transformed imaged-to-BEV features are inputted to three residual blocks and a GRU-based conditional branch based on the command [11]." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7809" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.078, + 0.091, + 0.216, + 0.106 + ], + "angle": 0, + "content": "3.4. Student Loss" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.115, + 0.473, + 0.403 + ], + "angle": 0, + "content": "Even with our carefully designed architecture, the student's image-based learning task remains significantly more challenging than the teacher's. Thus, both optimization and alignment with the teacher can benefit from extensive supervision and gradual student instruction, i.e., to scaffold supervision and better match the student's ability. Our distillation process provides ample beneficial supervision for the student through three holistic mechanisms. First, we incorporate deep distillation losses, i.e., both over output (i.e., waypoint regression task) and feature maps, as shown in Fig. 2, that can more directly supervise internal layers. Second, we incorporate additional auxiliary losses that regularize the optimization process. This includes ground-truth supervised losses, e.g., defined over segmentation and control command, and a geometric loss defined over the post-alignment features that supervises for correct BEV structure. Finally, we leverage a student-paced coaching mechanism for gradually increasing the difficulty of the waypoint prediction task throughout the iterations in Sec. 3.5." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.41, + 0.471, + 0.456 + ], + "angle": 0, + "content": "Loss Functions: Our optimization objective for guiding the distillation process is a weighted sum over both distillation and auxiliary tasks:" + }, + { + "type": "equation", + "bbox": [ + 0.142, + 0.47, + 0.47, + 0.488 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {C a T} = \\mathcal {L} _ {\\text {o u t}} + \\mathcal {L} _ {\\text {f e a t}} + \\mathcal {L} _ {\\text {s e g}} + \\mathcal {L} _ {\\text {c m d}} \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.501, + 0.471, + 0.532 + ], + "angle": 0, + "content": "The various loss terms enable guiding holistic aspects of student learning." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.539, + 0.47, + 0.57 + ], + "angle": 0, + "content": "Output Distillation: We leverage an \\( L_{1} \\) loss computed over all of the conditional branches for each instance" + }, + { + "type": "equation", + "bbox": [ + 0.149, + 0.583, + 0.47, + 0.625 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {o u t} = \\sum_ {c = 1} ^ {C} \\| f _ {\\boldsymbol {\\theta}} ^ {s} \\left(\\mathbf {x} ^ {s}, c\\right) - f _ {\\boldsymbol {\\psi}} ^ {t} \\left(\\mathbf {x} ^ {t}, c\\right) \\| _ {1} \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.636, + 0.469, + 0.668 + ], + "angle": 0, + "content": "This can be done by sampling from the teacher with different commands \\( c \\) (following Chen et al. [13])." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.674, + 0.473, + 0.902 + ], + "angle": 0, + "content": "Feature Distillation Loss: Our proposed student architecture provides BEV-space feature alignment, i.e., for direct internal feature matching between the student and the teacher (the first term in our loss in Eqn. 7). This is in contrast to current methods which distill the output or a final fully-connected layer [11, 13, 71, 79]. However, while this can provide rich supervision for the student, the two models should not match entirely, i.e., due to differences in information processing for a perception-based planning task. Moreover, the BEV-based features can be highly sparse and structured, with slight feature offsets being less meaningful than obtaining a task and scene-relevant representation. To flexibly account for the student-based task and effectively represent structural information in BEV space, our feature loss is computed as" + }, + { + "type": "equation", + "bbox": [ + 0.557, + 0.089, + 0.891, + 0.179 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} _ {f e a t} = \\sum_ {i = 1} ^ {3} \\left[ \\| \\mathcal {F} _ {i} ^ {s} (\\mathbf {x} ^ {s}) - \\mathcal {F} _ {i} ^ {t} (\\mathbf {x} ^ {t}) \\| _ {2} + \\right. \\\\ \\left\\| T _ {i} ^ {s} \\left(\\mathcal {F} _ {i} ^ {s} \\left(\\mathbf {x} ^ {s}\\right)\\right) - T _ {i} ^ {t} \\left(\\mathcal {F} _ {i} ^ {t} \\left(\\mathbf {x} ^ {t}\\right)\\right) \\right\\| _ {2} + \\tag {7} \\\\ \\lambda_ {C D} \\left\\| \\mathcal {F} _ {i} ^ {s} (\\mathbf {x} ^ {s}) - \\mathcal {F} _ {i} ^ {t} (\\mathbf {x} ^ {t})\\right) \\| _ {C D} \\Bigg ] \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.188, + 0.894, + 0.296 + ], + "angle": 0, + "content": "where we use \\(\\mathcal{F}_i\\) to denote layers within the networks (as shown in Fig. 2, we directly distill three layers), \\(T_{i}\\) indicates the features following several convolutional layers (trained jointly with the student network model), and \\(CD\\) stands for a Chamfer Distance [51] (we set \\(\\lambda_{CD} = 0.1\\)). The final term is computed over thresholded activations and a spatial soft-argmax [13] from each feature map." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.301, + 0.892, + 0.453 + ], + "angle": 0, + "content": "Task-oriented Auxiliary Tasks: To ensure essential task-relevant information is preserved throughout the distillation process, we leverage regularizing supervision in the form of segmentation and command-based auxiliary tasks interlaced throughout the network [16, 49]. Our segmentation loss \\(\\mathcal{L}_{seg}\\) is a cross-entropy loss, computed over the ground-truth BEV and averaged across the segmentation prediction heads. Our command prediction loss \\(\\mathcal{L}_{cmd}\\) is binary cross-entropy supervised by the ground-truth command \\(c\\), which facilitates learning task-relevant features [11, 19]." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.462, + 0.726, + 0.479 + ], + "angle": 0, + "content": "3.5. Student-paced Coaching" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.485, + 0.892, + 0.592 + ], + "angle": 0, + "content": "While Sec. 3.4 provides ample supervision on the student's internal representations, the teacher's targets may still be difficult to imitate for the student. To better consider the learning ability of the student, we propose to leverage a student-paced training mechanism which can gradually coach the student, i.e., through increasingly difficult samples. When computing the loss in Eqn. 5, we define a coach," + }, + { + "type": "equation", + "bbox": [ + 0.557, + 0.596, + 0.891, + 0.614 + ], + "angle": 0, + "content": "\\[\n\\mathcal {F} ^ {t} \\leftarrow \\lambda_ {i} \\mathcal {F} ^ {s} + \\left(1 - \\lambda_ {i}\\right) \\mathcal {F} ^ {t}, \\text {i f} \\mathcal {L} _ {C a T} > \\tau_ {i} \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.62, + 0.892, + 0.817 + ], + "angle": 0, + "content": "which interpolates the teaching targets with the student's predictions and features. \\(\\lambda_{i}\\) is a hyperparameter which is linearly decreased to 0 over training iteration \\(i\\). The modified targets are only computed over the hard samples at each batch. Within each batch, \\(\\tau_{i}\\) is a threshold defining the lowest \\(50\\%\\) of the scored samples for the loss with respect to the privileged teacher. While smoothing the targets in this manner may seem counterintuitive, i.e., compared to techniques which mine hard examples, the coaching mechanism aims to stabilize training by reducing the difficulty when the student is unable to perform the optimal action. Eqn. 8 enables adjusting the learning rate in a sample-selective manner (initially suppressing difficult samples)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.831, + 0.634, + 0.848 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.892, + 0.902 + ], + "angle": 0, + "content": "We use the CARLA simulator (version 0.9.10.1) [21] for data generation and closed-loop evaluation of the proposed CaT framework. We also use open-loop evaluation" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7810" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.089, + 0.895, + 0.161 + ], + "angle": 0, + "content": "Table 1. Quantitative Evaluation on the Longest6 Benchmark. Comparison of CaT with prior methods in terms of Driving Score (DS), Route Completion (RC), and Infraction Score (IS). Additional infraction metrics (Pedestrian Collisions (Ped), Vehicle Collisions (Veh), Layout Collisions (LC), Red Light Violations (Red), Off-road Infraction (OR), Route deviation (Dev), Route Timouts (TO), Agent Blocked (Blk)) are shown. FD refers to Feature Distillation, SH refers to Safety Hints. *re-trained by us using the publicly available code. Mean and standard deviation are computed over three runs. CaT outperforms the state-of-the-art by a large margin in terms of DS." + }, + { + "type": "table", + "bbox": [ + 0.078, + 0.163, + 0.892, + 0.375 + ], + "angle": 0, + "content": "
MethodRGBLiDARDS↑RC↑IS↑Ped↓Veh↓LC↓Red↓OR↓Dev↓TO↓Blk↓
LAV [11]✓✓48.41±3.4080.71±0.840.60±0.040.000.500.190.070.200.070.010.29
TransFuser [16]✓✓46.20±2.5783.61±1.160.57±0.000.290.380.280.040.070.070.000.32
WOR [10]✓X17.36±2.9543.46±2.990.54±0.060.050.640.150.840.150.890.040.45
NEAT [15]✓X24.08±3.3059.94±0.500.49±0.020.010.710.210.180.180.000.022.83
TCP* [71]✓X42.86±0.6361.83±4.190.71±0.040.010.420.110.040.080.010.010.75
CaT (w/o Alignment, Coaching, FD)✓X39.48±0.6760.96±1.650.68±0.010.031.290.200.080.840.000.021.49
CaT (w/o Alignment, Coaching)✓X40.64±0.9862.45±0.460.67±0.010.021.070.330.140.380.000.021.11
CaT (w/o Coaching, FD, SH)✓X44.10±0.4065.84±5.550.72±0.030.010.260.050.110.150.000.000.66
CaT (w/o Coaching, SH)✓X49.69±2.2881.10±0.580.64±0.020.010.780.030.080.200.000.020.51
CaT (w/o Coaching)✓X55.55±1.4181.97±2.340.68±0.010.020.300.050.070.060.000.020.35
CaT✓X58.36±2.2478.79±1.500.77±0.020.010.200.020.050.300.000.040.44
Privileged Agents:
RL Expert (Roach) [79]--60.14±2.4085.83±0.600.69±0.030.060.220.000.010.000.000.040.07
Rule-based Expert--71.96±2.1377.46±3.110.91±0.000.010.060.000.000.000.000.030.41
Basic BEV Agent [13]--24.08±2.8373.36±1.080.31±0.060.011.452.450.090.270.110.010.24
+ History and Desired Path--52.81±1.7979.34±3.650.71±0.060.020.300.000.000.000.000.020.53
+ Agent Forecast--65.73±0.9383.50±1.180.79±0.020.010.210.000.020.000.000.010.28
+ Entity Attention--73.30±1.0787.44±0.280.83±0.020.050.110.000.020.000.000.020.12
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.387, + 0.473, + 0.584 + ], + "angle": 0, + "content": "on nuScenes [7, 33, 77] (ADE, FDE, and collision rate). In Carla, we leverage the Longest6 Benchmark [16], which uses the six longest routes of each town (Town01 - Town06) from the set of 76 routes provided by the official CARLA leaderboard [1] (a total of 36 routes). We note in addition to the standard cross weather and time of day generalization evaluation, the long and dense traffic conditions in Longest6 present the most challenging settings across the various CARLA-based benchmarks. To evaluate our models, we follow standard metrics and report Route Completion (RC, in terms of completed route percentage), Infraction Score (IS, a penalty factor over infractions), and Driving Score (DS, computed from the prior two) [1, 11, 16])." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.605, + 0.365, + 0.62 + ], + "angle": 0, + "content": "4.1. Comparison with Prior Methods" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.629, + 0.473, + 0.903 + ], + "angle": 0, + "content": "CARLA Results: As shown in Table 1, using feature distillation, safety hints-based BEV for the teacher, and the proposed coaching mechanism, CaT is able to obtain a \\(58.36\\%\\) DS. Specifically, we achieve state-of-the-art performance among all prior agents, including LiDAR-based approaches [11, 16] (by \\(20.6\\%\\) in terms of DS, from 48.41 to 58.36, and \\(28.3\\%\\) in terms of IS). We also note that such approaches provide strong baselines, due to various LiDAR-based safety checks that the agents perform on top of the trained policy, which we do not employ. Moreover, CaT outperforms the prior RGB-only state-of-the-art agent TCP [71] by \\(36.16\\%\\) DS. To further validate the generalization of CaT, we discuss an additional benchmark split from TCP [71] and LAV [11] in the supplementary. We also find that removal of the alignment module degrades DS from \\(44.10\\%\\) to \\(39.48\\%\\), and is shown to only slightly benefit from feature distillation (\\(40.64\\%\\) DS) indicating the effectiveness of the proposed architecture." + }, + { + "type": "table_caption", + "bbox": [ + 0.553, + 0.386, + 0.838, + 0.4 + ], + "angle": 0, + "content": "Table 2. Open-Loop Evaluation on nuScenes." + }, + { + "type": "table", + "bbox": [ + 0.508, + 0.401, + 0.887, + 0.462 + ], + "angle": 0, + "content": "
MethodADE (m) ↓FDE (m) ↓Coll. (%) ↓
BEV Agent0.330.520.49
CaT (w/o Coaching, FD, SH)0.480.430.68
CaT0.410.360.27
" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.476, + 0.892, + 0.628 + ], + "angle": 0, + "content": "Real-World Evaluation: To further analyse the benefits of CaT in realistic driving settings, Table 2 shows open-loop evaluation (ADE, FDE, and Collision rate [77]) for nuScenes [7]. On the official validation split, the privileged agent performs best with an ADE of 0.33. CaT achieves 0.41 ADE, improving by \\(14.6\\%\\) gain over an agent without distillation and coaching. Moreover, we find a significant reduction in the collision rate by \\(60.3\\%\\) compared to the baseline. Additional details regarding nuScenes evaluation can be found in the supplementary." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.64, + 0.665, + 0.655 + ], + "angle": 0, + "content": "4.2. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.666, + 0.892, + 0.862 + ], + "angle": 0, + "content": "Teacher Comparison: Table 1 depicts our teacher model ablation. Specifically, we find the proposed BEV channels to drastically improve the privileged agent's performance by simplifying the learning task (73.30% DS, surpassing the built-in rule-based expert of 71.96%). We note that our behavior cloning agent also significantly outperforms an RL-based expert [79] (60.14%), simply through an effective BEV design. Thus, our findings apply to real-world scenarios where interactive agent training may be unsafe. We also observe that the improved teacher benefits the student agent. Specifically, incorporating the proposed BEV-based safety hints results in student performance gains, from 49.69% to 55.55%." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.871, + 0.892, + 0.903 + ], + "angle": 0, + "content": "Impact of Feature Distillation: Table 3 presents the benefits of our proposed multi-layer feature distillation frame" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "7811" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.158, + 0.089, + 0.812, + 0.367 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.368, + 0.895, + 0.44 + ], + "angle": 0, + "content": "Figure 4. Visualization of Model Input and Features. We show the BEV ground truth (BEV) and RGB image ground truth (Image) on the top row. Next, we visualize the intermediate layer activations of a sensorimotor agent without the proposed alignment module [71] (No Alignment Module), privileged teacher agent (Privileged), student agent trained with output distillation only (Output Distillation), and the proposed agent (CaT). We demonstrate that our teacher agent with proposed knowledge distillation is able to guide the student agent in learning scene-relevant representations and attend to safety-critical entities (circled in red)." + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.45, + 0.47, + 0.478 + ], + "angle": 0, + "content": "Table 3. Ablation Study on Feature Distillation Layers. All loss terms in Eqn. 7 benefit training with deep feature distillation." + }, + { + "type": "table", + "bbox": [ + 0.101, + 0.482, + 0.447, + 0.589 + ], + "angle": 0, + "content": "
MethodDS ↑RC ↑IS ↑
No Distillation44.1065.840.72
One Layer [71,79]45.2369.330.68
Three Layers L249.3166.920.78
Three Layers L2 + LCD51.9562.820.87
Three Layers Lfeat55.5581.970.68
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.612, + 0.47, + 0.718 + ], + "angle": 0, + "content": "work. We note that prior methods only distill one final fully connected layer (before the prediction head [71, 79], referred to as 'one layer' in Table 3). We note that such approaches result in minimal gains, whereas CaT is able to benefit from the extensive feature distillation (improving from \\(45.23\\%\\) to \\(55.55\\%\\)). Incorporating multiple types of feature distillation losses also benefits performance." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.727, + 0.279, + 0.743 + ], + "angle": 0, + "content": "4.3. Feature Visualization" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Fig. 4 depicts the learned features of the initial feature distilled layer (out of the three in Fig. 2). We visualize activations from two channels, to better understand the learned representation by the agents. Interestingly, the visualization shows how traditionally trained sensorimotor agents (i.e., without the proposed alignment module) mostly operate in image-space features in this layer despite the BEV waypoint prediction task. This validates our main hypothesis, that such agents have limited ability to learn from a BEV-based agent, even with BEV-based auxiliary prediction tasks. We" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.452, + 0.895, + 0.559 + ], + "angle": 0, + "content": "also find that output distillation provides a weak supervision for critical components of the BEV, including layout and object features. After feature distillation, the features of our CaT model are able to not only keep the driving-related semantic information i.e., drivable road and route, but also focus on safety-critical entities. More visualization examples are provided in the supplementary." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.578, + 0.619, + 0.593 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.605, + 0.895, + 0.862 + ], + "angle": 0, + "content": "We present CaT, a novel knowledge distillation method for effectively training a sensorimotor student agent using the supervision from a privileged agent. We leverage an alignment module to better map image features to the BEV space, thus enabling extensive supervision from a BEV-based teacher over the intermediate feature learning. As both agents can be trained offline via imitation learning, our findings are directly relevant to real-world settings where on-policy interactions may be unsafe to perform. Finally, we carefully account for inherent differences between the student and the teacher using a student-paced coaching mechanism with various auxiliary supervision tasks. Through the improved knowledge distillation process, our experiments result in a state-of-the-art camera-based agent in CARLA. Our ablation into input and model design guides future directions through interpretable analysis into current limitations of end-to-end driving models." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.871, + 0.892, + 0.902 + ], + "angle": 0, + "content": "Acknowledgments: We thank the Red Hat Collaboratory for supporting this research." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7812" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.116, + 0.468, + 0.143 + ], + "angle": 0, + "content": "[1] Carla autonomous driving leaderboard. https://leaderboard.carla.org/, 2022.7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.145, + 0.469, + 0.186 + ], + "angle": 0, + "content": "[2] Yonathan Aflalo, Asaf Noy, Ming Lin, Itamar Friedman, and Lihi Zelnik. Knapsack pruning with inner distillation. arXiv preprint arXiv:2002.08258, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.188, + 0.469, + 0.228 + ], + "angle": 0, + "content": "[3] Gustavo Aguilar, Yuan Ling, Yu Zhang, Benjamin Yao, Xing Fan, and Chenlei Guo. Knowledge distillation from internal representations. In AAAI, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.231, + 0.469, + 0.272 + ], + "angle": 0, + "content": "[4] Mayank Bansal, Alex Krizhevsky, and Abhijit Ogale. ChauffeurNet: Learning to drive by imitating the best and synthesizing the worst. In RSS, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.274, + 0.469, + 0.3 + ], + "angle": 0, + "content": "[5] Yoshua Bengio, Jérôme Louradour, Ronan Collobert, and Jason Weston. Curriculum learning. In ICML, 2009. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.303, + 0.469, + 0.343 + ], + "angle": 0, + "content": "[6] Massimo Bertozzi, Alberto Broggi, and Alessandra Fascioli. An extension to the inverse perspective mapping to handle non-flat roads. In IV, 1998. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.346, + 0.469, + 0.401 + ], + "angle": 0, + "content": "[7] Holger Caesar, Varun Bankiti, Alex H Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuscenes: A multimodal dataset for autonomous driving. In CVPR, 2020. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.403, + 0.469, + 0.457 + ], + "angle": 0, + "content": "[8] Raphael Chekroun, Marin Toromanoff, Sascha Hornauer, and Fabien Moutarde. Gri: General reinforced imitation and its application to vision-based autonomous driving. arXiv preprint arXiv:2111.08575, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.46, + 0.469, + 0.5 + ], + "angle": 0, + "content": "[9] Chenyi Chen, Ari Seff, Alain L. Kornhauser, and Jianxiong Xiao. DeepDriving: Learning affordance for direct perception in autonomous driving. In ICCV, 2015. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.502, + 0.469, + 0.529 + ], + "angle": 0, + "content": "[10] Dian Chen, Vladlen Koltun, and Philipp Krahenbuhl. Learning to drive from a world on rails. In ICCV, 2021. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.532, + 0.469, + 0.558 + ], + "angle": 0, + "content": "[11] Dian Chen and Philipp Krahenbuhl. Learning from all vehicles. In CVPR, 2022. 1, 2, 3, 4, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.561, + 0.469, + 0.601 + ], + "angle": 0, + "content": "[12] Defang Chen, Jian-Ping Mei, Yuan Zhang, Can Wang, Zhe Wang, Yan Feng, and Chun Chen. Cross-layer distillation with semantic calibration. In AAAI, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.603, + 0.469, + 0.643 + ], + "angle": 0, + "content": "[13] Dian Chen, Brady Zhou, Vladlen Koltun, and Philipp Krahenbuhl. Learning by cheating. In CoRL, 2020. 1, 2, 3, 4, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.646, + 0.469, + 0.714 + ], + "angle": 0, + "content": "[14] Li Chen, Chonghao Sima, Yang Li, Zehan Zheng, Jiajie Xu, Xiangwei Geng, Hongyang Li, Conghui He, Jianping Shi, Yu Qiao, et al. Persformer: 3d lane detection via perspective transformer and the openlane benchmark. arXiv preprint arXiv:2203.11089, 2022. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.717, + 0.469, + 0.757 + ], + "angle": 0, + "content": "[15] Kashyap Chitta, Aditya Prakash, and Andreas Geiger. Neat: Neural attention fields for end-to-end autonomous driving. In ICCV, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.76, + 0.469, + 0.815 + ], + "angle": 0, + "content": "[16] Kashyap Chitta, Aditya Prakash, Bernhard Jaeger, Zehao Yu, Katrin Renz, and Andreas Geiger. Transfuser: Imitation with transformer-based sensor fusion for autonomous driving. arXiv preprint arXiv:2205.15997, 2022. 3, 4, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.817, + 0.469, + 0.857 + ], + "angle": 0, + "content": "[17] Inseop Chung, SeongUk Park, Jangho Kim, and Nojun Kwak. Feature-map-level online adversarial knowledge distillation. In ICML, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.859, + 0.469, + 0.9 + ], + "angle": 0, + "content": "[18] Felipe Codevilla, Matthias Muiller, Antonio López, Vladlen Koltun, and Alexey Dosovitskiy. End-to-end driving via conditional imitation learning. In ICRA, 2018. 3, 4" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.116, + 0.469, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.134 + ], + "angle": 0, + "content": "[19] Felipe Codevilla, Eder Santana, Antonio M López, and Adrien Gaidon. Exploring the limitations of behavior cloning for autonomous driving. In ICCV, 2019. 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.137, + 0.892, + 0.164 + ], + "angle": 0, + "content": "[20] Pim de Haan, Dinesh Jayaraman, and Sergey Levine. Causal confusion in imitation learning. In NeurIPS, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.166, + 0.892, + 0.207 + ], + "angle": 0, + "content": "[21] Alexey Dosovitskiy, German Ros, Felipe Codevilla, Antonio Lopez, and Vladlen Koltun. CARLA: An open urban driving simulator. In CoRL, 2017. 2, 3, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.21, + 0.892, + 0.237 + ], + "angle": 0, + "content": "[22] James J Gibson. The ecological approach to visual perception: classic edition. Psychology press, 2014. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.239, + 0.892, + 0.279 + ], + "angle": 0, + "content": "[23] Jianping Gou, Baosheng Yu, Stephen J Maybank, and Dacheng Tao. Knowledge distillation: A survey. In IJCV, 2021. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.282, + 0.892, + 0.337 + ], + "angle": 0, + "content": "[24] Sheng Guo, Weilin Huang, Haozhi Zhang, Chenfan Zhuang, Dengke Dong, Matthew R Scott, and Dinglong Huang. Curriculumnet: Weakly supervised learning from large-scale web images. In ECCV, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.34, + 0.892, + 0.381 + ], + "angle": 0, + "content": "[25] Saurabh Gupta, James Davidson, Sergey Levine, Rahul Sukthankar, and Jitendra Malik. Cognitive mapping and planning for visual navigation. In CVPR, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.384, + 0.892, + 0.436 + ], + "angle": 0, + "content": "[26] Peter E Hart, Nils J Nilsson, and Bertram Raphael. A formal basis for the heuristic determination of minimum cost paths. IEEE transactions on Systems Science and Cybernetics, 1968, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.441, + 0.892, + 0.467 + ], + "angle": 0, + "content": "[27] He He, Jason Eisner, and Hal Daume. Imitation learning by coaching. In NeurIPS, 2012. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.471, + 0.892, + 0.51 + ], + "angle": 0, + "content": "[28] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, 2016. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.514, + 0.892, + 0.554 + ], + "angle": 0, + "content": "[29] Byeongho Heo, Jeesoo Kim, Sangwoo Yun, Hyojin Park, No-jun Kwak, and Jin Young Choi. A comprehensive overhaul of feature distillation. In ICCV, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.557, + 0.892, + 0.598 + ], + "angle": 0, + "content": "[30] Roger Hiemstra and Burton Sisco. Individualizing Instruction. Making Learning Personal, Empowering, and Successful. 1990. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.601, + 0.892, + 0.641 + ], + "angle": 0, + "content": "[31] Geoffrey Hinton, Oriol Vinyals, Jeff Dean, et al. Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531, 2(7), 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.644, + 0.892, + 0.712 + ], + "angle": 0, + "content": "[32] Anthony Hu, Zak Murez, Nikhil Mohan, Sofia Dudas, Jeffrey Hawke, Vijay Badrinarayanan, Roberto Cipolla, and Alex Kendall. Fiery: Future instance prediction in bird's-eye view from surround monocular cameras. In ICCV, 2021. 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.716, + 0.892, + 0.757 + ], + "angle": 0, + "content": "[33] Peiyun Hu, Aaron Huang, John Dolan, David Held, and Deva Ramanan. Safe local motion planning with self-supervised freespace forecasting. In CVPR, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.759, + 0.892, + 0.813 + ], + "angle": 0, + "content": "[34] Zanming Huang, Zhongkai Shangguan, Jimuyang Zhang, Gilad Bar, Matthew Boyd, and Eshed Ohn-Bar. ASSISTER: Assistive navigation via conditional instruction generation. In ECCV, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.816, + 0.892, + 0.856 + ], + "angle": 0, + "content": "[35] Zehao Huang and Naiyan Wang. Like what you like: Knowledge distill via neuron selectivity transfer. arXiv preprint arXiv:1707.01219, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.859, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[36] Mingi Ji, Byeongho Heo, and Sungrae Park. Show, attend and distill: Knowledge distillation via attention-based feature matching. In AAAI, 2021. 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7813" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.147 + ], + "angle": 0, + "content": "[37] Lu Jiang, Zhengyuan Zhou, Thomas Leung, Li-Jia Li, and Li Fei-Fei. Mentornet: Learning data-driven curriculum for very deep neural networks on corrupted labels. In ICML, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.15, + 0.47, + 0.191 + ], + "angle": 0, + "content": "[38] Jangho Kim, SeongUk Park, and Nojun Kwak. Paraphrasing complex network: Network compression via factor transfer. In NeurIPS, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.193, + 0.47, + 0.234 + ], + "angle": 0, + "content": "[39] Tom Kocmi and Ondrej Bojar. Curriculum learning and minibatch bucketing in neural machine translation. arXiv preprint arXiv:1707.09533, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.237, + 0.47, + 0.278 + ], + "angle": 0, + "content": "[40] Animesh Koratana, Daniel Kang, Peter Bailis, and Matei Zaharia. Lit: Learned intermediate representation training for model compression. In ICML, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.281, + 0.469, + 0.308 + ], + "angle": 0, + "content": "[41] M Kumar, Benjamin Packer, and Daphne Koller. Self-paced learning for latent variable models. In NeurIPS, 2010. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.31, + 0.469, + 0.352 + ], + "angle": 0, + "content": "[42] Ravi Kumar Kushawaha, Saurabh Kumar, Biplab Banerjee, and Rajbabu Velmurugan. Distilling spikes: Knowledge distillation in spiking neural networks. In ICPR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.354, + 0.469, + 0.408 + ], + "angle": 0, + "content": "[43] Sergey Levine, Chelsea Finn, Trevor Darrell, and Pieter Abbeel. End-to-end training of deep visuomotor policies. The Journal of Machine Learning Research, 17(1):1334-1373, 2016. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.411, + 0.469, + 0.452 + ], + "angle": 0, + "content": "[44] Dong Li, Jia-Bin Huang, Yali Li, Shengjin Wang, and Ming-Hsuan Yang. Weakly supervised object localization with progressive domain adaptation. In CVPR, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.454, + 0.469, + 0.495 + ], + "angle": 0, + "content": "[45] Guohao Li, Matthias Mueller, Vincent Casser, Neil Smith, Dominik L Michels, and Bernard Ghanem. Oil: Observational imitation learning. In RSS, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.498, + 0.469, + 0.525 + ], + "angle": 0, + "content": "[46] Yang Liu, Wei Zhang, and Jun Wang. Source-free domain adaptation for semantic segmentation. In CVPR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.527, + 0.469, + 0.581 + ], + "angle": 0, + "content": "[47] Takahiro Miki, Joonho Lee, Jemin Hwangbo, Lorenz Wellhausen, Vladlen Koltun, and Marco Hutter. Learning robust perceptive locomotion for quadrupedal robots in the wild. Science Robotics, 7(62), 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.584, + 0.469, + 0.625 + ], + "angle": 0, + "content": "[48] Matthias Müller, Alexey Dosovitskiy, Bernard Ghanem, and Vladlen Koltun. Driving policy transfer via modularity and abstraction. In CoRL, 2018. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.628, + 0.469, + 0.668 + ], + "angle": 0, + "content": "[49] Eshed Ohn-Bar, Aditya Prakash, Aseem Behl, Kashyap Chitta, and Andreas Geiger. Learning situational driving. In CVPR, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.671, + 0.469, + 0.726 + ], + "angle": 0, + "content": "[50] Takayuki Osa, Joni Pajarinen, Gerhard Neumann, J Andrew Bagnell, Pieter Abbeel, Jan Peters, et al. An algorithmic perspective on imitation learning. Foundations and Trends® in Robotics, 2018. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.729, + 0.469, + 0.783 + ], + "angle": 0, + "content": "[51] Despoina Paschalidou, Angelos Katharopoulos, Andreas Geiger, and Sanja Fidler. Neural parts: Learning expressive 3D shape abstractions with invertible neural networks. In CVPR, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.786, + 0.469, + 0.826 + ], + "angle": 0, + "content": "[52] Nikolaos Passalis and Anastasios Tefas. Learning deep representations with probabilistic knowledge transfer. In ECCV, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.829, + 0.469, + 0.87 + ], + "angle": 0, + "content": "[53] Jonah Philion and Sanja Fidler. Lift, splat, shoot: Encoding images from arbitrary camera rigs by implicitly unprojecting to 3D. In ECCV, 2020. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.469, + 0.9 + ], + "angle": 0, + "content": "[54] Dean A Pomerleau. ALVINN: An autonomous land vehicle in a neural network. In NeurIPS, 1989. 1, 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.147 + ], + "angle": 0, + "content": "[55] Adriana Romero, Nicolas Ballas, Samira Ebrahimi Kahou, Antoine Chassang, Carlo Gatta, and Yoshua Bengio. Fitnets: Hints for thin deep nets. arXiv preprint arXiv:1412.6550, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.15, + 0.892, + 0.191 + ], + "angle": 0, + "content": "[56] Stéphane Ross, Geoffrey Gordon, and Drew Bagnell. A reduction of imitation learning and structured prediction to no-regret online learning. In AISTATS, 2011. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.193, + 0.892, + 0.222 + ], + "angle": 0, + "content": "[57] Avishkar Saha, Oscar Mendez, Chris Russell, and Richard Bowden. Translating images into maps. In ICRA, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.224, + 0.892, + 0.263 + ], + "angle": 0, + "content": "[58] Axel Sauer, Nikolay Savinov, and Andreas Geiger. Conditional affordance learning for driving in urban environments. In CoRL, 2018. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.266, + 0.892, + 0.294 + ], + "angle": 0, + "content": "[59] Miaojing Shi and Vittorio Ferrari. Weakly supervised object localization using size estimates. In ECCV, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.296, + 0.892, + 0.324 + ], + "angle": 0, + "content": "[60] Suraj Srinivas and Francois Fleuret. Knowledge transfer with Jacobian matching. In ICML, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.327, + 0.892, + 0.381 + ], + "angle": 0, + "content": "[61] Yuxing Tang, Xiaosong Wang, Adam P Harrison, Le Lu, Jing Xiao, and Ronald M Summers. Attention-guided curriculum learning for weakly supervised classification and localization of thoracic diseases on chest radiographs. In MLMI, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.383, + 0.892, + 0.424 + ], + "angle": 0, + "content": "[62] Marin Toromanoff, Emilie Wirbel, and Fabien Moutarde. End-to-end model-free reinforcement learning for urban driving using implicit affordances. In CVPR, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.426, + 0.892, + 0.482 + ], + "angle": 0, + "content": "[63] Radu Tudor Ionescu, Bogdan Alexe, Marius Leordeanu, Marius Popescu, Dim P Papadopoulos, and Vittorio Ferrari. How hard can it be? estimating the difficulty of visual search in an image. In CVPR, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.484, + 0.892, + 0.512 + ], + "angle": 0, + "content": "[64] Rachel R Van Der Stuyf. Scaffolding as a teaching strategy. Adolescent learning and development, 2002. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.514, + 0.892, + 0.555 + ], + "angle": 0, + "content": "[65] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In NeurIPS, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.557, + 0.892, + 0.584 + ], + "angle": 0, + "content": "[66] Antonio Visioli. Practical PID control. Springer Science & Business Media, 2006. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.586, + 0.892, + 0.627 + ], + "angle": 0, + "content": "[67] Tao Wang, Li Yuan, Xiaopeng Zhang, and Jiashi Feng. Distilling object detectors with fine-grained feature imitation. In CVPR, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.63, + 0.892, + 0.685 + ], + "angle": 0, + "content": "[68] Wenhui Wang, Hangbo Bao, Shaohan Huang, Li Dong, and Furu Wei. Minilmv2: Multi-head self-attention relation distillation for compressing pretrained transformers. arXiv preprint arXiv:2012.15828, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.687, + 0.892, + 0.742 + ], + "angle": 0, + "content": "[69] Wei Wang, Isaac Caswell, and Ciprian Chelba. Dynamically composing domain-data selection with clean-data selection by \"co-curricular learning\" for neural machine translation. arXiv preprint arXiv:1906.01130, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.744, + 0.892, + 0.786 + ], + "angle": 0, + "content": "[70] Chuan Wen, Jierui Lin, Trevor Darrell, Dinesh Jayaraman, and Yang Gao. Fighting copycat agents in behavioral cloning from observation histories. In NeurIPS, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.788, + 0.892, + 0.855 + ], + "angle": 0, + "content": "[71] Penghao Wu, Xiaosong Jia, Li Chen, Junchi Yan, Hongyang Li, and Yu Qiao. Trajectory-guided control prediction for end-to-end autonomous driving: A simple yet strong baseline. arXiv preprint arXiv:2206.08129, 2022. 1, 2, 3, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.858, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[72] Liuyu Xiang, Guiguang Ding, and Jungong Han. Learning from multiple experts: Self-paced knowledge distillation for long-tailed classification. In ECCV, 2020. 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "7814" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.147 + ], + "angle": 0, + "content": "[73] Canwen Xu, Wangchunshu Zhou, Tao Ge, Furu Wei, and Ming Zhou. Bert-of-theseus: Compressing bert by progressive module replacing. arXiv preprint arXiv:2002.02925, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.149, + 0.47, + 0.19 + ], + "angle": 0, + "content": "[74] Jing Yang, Brais Martinez, Adrian Bulat, and Georgios Tzimiropoulos. Knowledge distillation via adaptive instance normalization. arXiv preprint arXiv:2003.04289, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.192, + 0.469, + 0.234 + ], + "angle": 0, + "content": "[75] Junho Yim, Donggyu Joo, Jihoon Bae, and Junmo Kim. A gift from knowledge distillation: Fast optimization, network minimization and transfer learning. In CVPR, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.235, + 0.469, + 0.262 + ], + "angle": 0, + "content": "[76] Jimuyang Zhang and Eshed Ohn-Bar. Learning by watching. In CVPR, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.264, + 0.469, + 0.304 + ], + "angle": 0, + "content": "[77] Jimuyang Zhang, Ruizhao Zhu, and Eshed Ohn-Bar. SelfD: Self-learning large-scale driving policies from the web. In CVPR, 2022. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.307, + 0.469, + 0.348 + ], + "angle": 0, + "content": "[78] Pan Zhang, Bo Zhang, Ting Zhang, Dong Chen, and Fang Wen. Robust mutual learning for semi-supervised semantic segmentation. arXiv preprint arXiv:2106.00609, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.35, + 0.469, + 0.404 + ], + "angle": 0, + "content": "[79] Zhejun Zhang, Alexander Liniger, Dengxin Dai, Fisher Yu, and Luc Van Gool. End-to-end urban driving by imitating a reinforcement learning coach. In ICCV, 2021. 1, 2, 3, 4, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.406, + 0.469, + 0.461 + ], + "angle": 0, + "content": "[80] Albert Zhao, Tong He, Yitao Liang, Haibin Huang, Guy Van den Broeck, and Stefano Soatto. Lates: Latent space distillation for teacher-student driving policy learning. arXiv preprint arXiv:1912.02973, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.463, + 0.469, + 0.517 + ], + "angle": 0, + "content": "[81] Albert Zhao, Tong He, Yitao Liang, Haibin Huang, Guy Van den Broeck, and Stefano Soatto. Sam: Squeeze-and-mimic networks for conditional visual driving policy learning. CoRL, 2020. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.519, + 0.469, + 0.559 + ], + "angle": 0, + "content": "[82] Brady Zhou and Philipp Krahenbuhl. Cross-view transformers for real-time map-view semantic segmentation. In CVPR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.562, + 0.469, + 0.602 + ], + "angle": 0, + "content": "[83] Brady Zhou, Philipp Krähenbuhl, and Vladlen Koltun. Does computer vision matter for action? Science Robotics, 4(30), 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.604, + 0.469, + 0.646 + ], + "angle": 0, + "content": "[84] Yikai Zhou, Baosong Yang, Derek F Wong, Yu Wan, and Lidia S Chao. Uncertainty-aware curriculum learning for neural machine translation. In ACL, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.647, + 0.469, + 0.702 + ], + "angle": 0, + "content": "[85] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable detr: Deformable transformers for end-to-end object detection. arXiv preprint arXiv:2010.04159, 2020. 5" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.702 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "7815" + } + ] +] \ No newline at end of file diff --git a/2023/Coaching a Teachable Student/67002b28-2cb3-47c6-8f0a-06c8bb56b862_origin.pdf b/2023/Coaching a Teachable Student/67002b28-2cb3-47c6-8f0a-06c8bb56b862_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..83fff3f3a15c0df1d3fad2c4325c3709d7f89a61 --- /dev/null +++ b/2023/Coaching a Teachable Student/67002b28-2cb3-47c6-8f0a-06c8bb56b862_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:51fa889799c3dcf00e1e47591cbcbcefc6df0edc4b52fdc3a3c9113554e46c02 +size 1097881 diff --git a/2023/Coaching a Teachable Student/full.md b/2023/Coaching a Teachable Student/full.md new file mode 100644 index 0000000000000000000000000000000000000000..4b3596d08a43e84b5ae65a69da3773bdef213676 --- /dev/null +++ b/2023/Coaching a Teachable Student/full.md @@ -0,0 +1,312 @@ +# Coaching a Teachable Student + +Jimuyang Zhang Zanming Huang Eshed Ohn-Bar Boston University + +{zhangjim, huangtom, eohnbar}@bu.edu + +# Abstract + +We propose a novel knowledge distillation framework for effectively teaching a sensorimotor student agent to drive from the supervision of a privileged teacher agent. Current distillation for sensorimotor agents methods tend to result in suboptimal learned driving behavior by the student, which we hypothesize is due to inherent differences between the input, modeling capacity, and optimization processes of the two agents. We develop a novel distillation scheme that can address these limitations and close the gap between the sensorimotor agent and its privileged teacher. Our key insight is to design a student which learns to align their input features with the teacher's privileged Bird's Eye View (BEV) space. The student then can benefit from direct supervision by the teacher over the internal representation learning. To scaffold the difficult sensorimotor learning task, the student model is optimized via a student-paced coaching mechanism with various auxiliary supervision. We further propose a high-capacity imitation learned privileged agent that surpasses prior privileged agents in CARLA and ensures the student learns safe driving behavior. Our proposed sensorimotor agent results in a robust image-based behavior cloning agent in CARLA, improving over current models by over $20.6\%$ in driving score without requiring LiDAR, historical observations, ensemble of models, on-policy data aggregation or reinforcement learning. + +# 1. Introduction + +Learning internal representations for making intricate driving decisions from images involves a complex optimization task [4, 43, 54]. The inherent challenge for end-to-end training of driving agents lies in the immense complexity of learning to map high-dimensional visual observations into general and safe navigational decisions [13, 20, 70]. Even given millions of training examples [4], today's agents still fail to reliably learn an internal representation that can be used for robust processing of complex visual scenarios (e.g., dense urban settings with intricate layouts and dynamic obstacles) in a safe and task-driven manner [13, 20, 70]. + +To ease the challenging sensorimotor training task, re + +![](images/c07ac891b0e814efddc31764fff5c6d21d4fb1cd5b91050d00569bfe4c084176.jpg) +Figure 1. Effective Knowledge Distillation for Sensorimotor Agents. Our proposed CaT (Coaching a Teachable student) framework enables highly effective knowledge transfer between a privileged teacher and a sensorimotor (i.e., image-based) student. An alignment module learns to transform image-based features to the teacher's BEV feature space, where the student can then leverage extensive and direct supervision on its learned intermediate representations. The student model is optimized via a coaching mechanism with extensive auxiliary supervision in order to further scaffold the difficult sensorimotor learning task. + +cent approaches decompose the task into stages, e.g., by first training a high-capacity privileged network with complete knowledge of the world and distilling its knowledge into a less capable vision-based student network [11, 13, 23, 47, 79]. However, due to the inherent differences between the inputs and architectures of the two agents, current methods rely on limited supervisory mechanisms from the teacher, i.e., exclusively through the teacher's output [11, 13] or knowledge distillation of a single final fully-connected layer [71, 79, 81]. Moreover, the privileged teacher's demonstration targets may be noisy or difficult for the student to imitate, given the limited perspective [27]. In this work, we sought to develop a more effective knowledge distillation paradigm for training a sensorimotor agent to drive. Our key insight is to enable more extensive supervision from the teacher by reducing the gap between internal modeling and learning capabilities between the two agents. + +Our proposed approach for holistic knowledge distillation is informed by human instruction, which often involves structured supervision in addition to high-level demonstra + +tions, e.g., providing various hints to scaffold information in a way that the student can better understand [30]. When teaching others new and challenging skills, i.e., where a student may not be able to replicate the demonstration such as riding a bicycle or driving a vehicle, teachers may provide additional supervision regarding the underlying task structure and their own internal reasoning [64]. Analogously to our settings, the privileged teaching agent can potentially provide richer supervision when teaching a limited capacity sensorimotor student, i.e., through more careful and direct guidance of the underlying representation learning. + +In our work, we introduce CaT, a novel method for teaching a sensorimotor student to drive using supervision from a privileged teacher. Our key insights are threefold: 1) **Effective Teacher:** We propose to incorporate explicit safety-aware cues into the BEV space that facilitate a surprisingly effective teacher agent design. While prior privileged agents struggle to learn to drive in complex urban driving scenes, we demonstrate our learned agent to match expert-level decision-making. 2) **Teachable Student via Alignment:** An IPM-based transformer alignment module can facilitate direct distillation of most of the teacher's features and better guide the student learning process. 3) **Student-paced Coaching:** A coaching mechanism for managing difficult samples can scaffold knowledge and lead to improved model optimization by better considering the ability of the student. By holistically tackling the complex knowledge distillation task with extensive teacher and auxiliary supervision, we are able to train a state-of-the-art image-based agent in CARLA [21]. Through ablation studies, input design, and interpretable feature analysis, we also provide critical insights into current limitations in learning robust and generalized representations for driving agents. + +# 2. Related Work + +Knowledge Distillation: We study knowledge distillation methods originally developed for model compression and acceleration [12,17,29,31,35,36,52,55,74,75] in the context of training sensorimotor. While approaches for feature distillation have been minimally explored in this context, we are motivated by their success in other domains, including image classification [2,38,40,42,60], object detection [67], semantic segmentation [46, 78], and natural language processing [3,68,73]. Yet, applying such techniques is not trivial given the drastically differing inputs between the privileged agent and student and the overwhelming sensorimotor task. Consequently, driving policy distillation methods only provide supervision either from the teacher's output [11,13] or the features of a single fully-connected layer [71,79,81], which (we hypothesize) does not provide sufficient hints to guide the student training. In this work, we provide novel mechanisms to close this gap and enable extensive supervision through deep feature distillation. + +Imitation Learning to Drive: Recent approaches in imitation learning (IL) to drive can be traced to Pomerleau [54]. Recently, more elaborate IL-based approaches for driving have emerged [9, 10, 25, 45, 50, 76, 77, 80, 83]. Specifically, decomposing the imitation learning task into two stages, i.e., by first learning a privileged agent through behavior cloning [11, 13] or reinforcement learning [8, 62, 71, 79] and then training the sensorimotor agent to mimic the output of the privileged agent. Our study is motivated by such approaches, yet we explore the benefits of increased supervision by the privileged agent. Moreover, prior privileged agents produce noisy and sub-optimal demonstration supervision in complex urban scenarios (even when optimized with RL, as demonstrated by our analysis). While most current studies employ the RL-based teacher of [79], this can be problematic and inefficient in safety-critical conditions in the real-world. In our work, we introduce a novel privileged teacher which can greatly surpass prior agents while only relying on offline behavior cloning. + +Intermediate Representation for Driving: Learning an effective 3D scene representation is crucial for safe autonomous driving. Researchers may obtain such representations by lifting the image to 3D using estimated depth and projecting the frustums into a BEV grid [32, 53]. Alternatively, transformer-based architectures can also enable mapping a camera image to the BEV space [14, 82], i.e., by attending to image-based information when populating a BEV-sampled grid. Related to our work is the study of Chen et al. [11], which learns a BEV representation from RGB image and LiDAR input. A motion planner is then used to generate future waypoints from the BEV. As this introduces significant challenges, we propose to leverage an IPM-based alignment module which can better structure to image-to-BEV projection task. + +Curriculum and Self-paced Learning: Approaches for curriculum learning methods often structure learning from various difficulty samples [5,24,37,39,44,59,61,63,69,84]. To better consider the challenging sensorimotor task and limited capacity of the student, we adopt a curriculum (i.e., an imitation coach), with progressively more challenging samples. He et al. [27] also proposes training a coach across iterations of on-policy data collection [56]. In contrast, we implement and demonstrate the benefit of a coach over training iterations without a data aggregation part. This can facilitate more efficient model training. The approach is motivated by self-paced learning [41, 72], where the model selects easy samples dynamically at each iteration based on a defined loss. In contrast to such studies, we only smooth the targets of the hard samples (based on the student), instead of discarding them. We empirically find this to result in improved coaching, potentially due to improved scaffolding of the difficult samples. + +# 3. Method + +Our goal is to decompose the challenging sensorimotor learning task and ease model optimization through effective supervision from a teacher agent. We first formulate the problem of privileged agent distillation in Sec. 3.1. Next, we address current limitations in privileged agent design to train a robust imitation learned teacher in Sec. 3.2. We then propose a transformer-based student architecture to resolve image-to-BEV feature alignment (Sec. 3.3). Finally, we train the sensorimotor student via extensive supervision and knowledge distillation (Sec. 3.4) via a progressive (i.e., coaching) mechanism (Sec. 3.5). + +# 3.1. Formulation + +Driving can be formulated as a sequential decision-making problem over a set of sensory observations $\mathcal{X}^s$ , a set of actions (motor commands) $\mathcal{V}$ , and a policy function $f_{\theta}^{s} \colon \mathcal{X}^{s} \to \mathcal{V}$ for mapping observations to actions at each time step [27,50,56]. We consider the task of learning a sensorimotor student agent via knowledge distillation [13, 23] from a teacher agent $f_{\psi}^{t} \colon \mathcal{X}^{t} \to \mathcal{V}$ with privileged access to environmental information, e.g., ground-truth lanes, obstacles, and traffic light states. We parameterize the two agents as neural networks using weights $[\pmb{\theta}, \pmb{\psi}] \in \mathbb{R}^d$ , and denote $\mathcal{F}^s(\cdot; \pmb{\theta})$ and $\mathcal{F}^t(\cdot; \pmb{\psi})$ to be the feature maps for the sensorimotor student network and teacher network, respectively. Given a dataset $\mathcal{D}$ comprising sensory and privileged observations and a loss function $\mathcal{L}$ , the student can be optimized from the teacher using + +$$ +\underset {\boldsymbol {\theta}} {\operatorname {a r g m i n}} \mathbb {E} _ {(\mathbf {x} ^ {s}, \mathbf {x} ^ {t}) \sim \mathcal {D}} [ \mathcal {L} (\mathcal {F} ^ {s} (\mathbf {x} ^ {s}; \boldsymbol {\theta}), \mathcal {F} ^ {t} (\mathbf {x} ^ {t}; \boldsymbol {\psi})) ] \tag {1} +$$ + +where the loss may be defined over a final layer, i.e., to match the output of the teacher [11, 13], or over internal layers as well, i.e., feature distillation [71, 79, 81]. Given the challenging end-to-end sensorimotor learning task, the role of the teacher is to provide effective supervision to the student, i.e., informative targets in Eqn. 1. However, as the two agents operate from drastically differing inputs, the knowledge transfer from the teacher to the student can become ineffective. Specifically, the task of the student is significantly more challenging than that of the teacher, i.e., due to inherent noise and uncertainty. Indeed, there is currently a substantial gap between the performance of current sensorimotor learning agents and their teaching agents, even assuming access to expensive and high-quality 3D observations such as LiDAR [11, 13, 16] (not used in our work). The teaching process should also ideally consider any limited capacity of the student to facilitate effective optimization and learning [27]. Moreover, learning an optimal and robust teacher in itself can be a challenging task [13], even assuming the simplified learning task such as solved perception, as we demonstrate in Sec. 4 and addressed in Sec. 3.2. + +Problem Setup and Agent Observations: We develop our CaT framework in the context of the CARLA simulator [21]. Our objective is to train a goal-conditional sensorimotor agent [16, 18] for mapping image and goal observations to vehicle throttle and steering control [13, 48]. We follow standard conditional imitation learning [11, 16] and assume access to three (non-overlapping) RGB camera views $\mathbf{I} = [\mathbf{I}_0, \mathbf{I}_1, \mathbf{I}_2] \in \mathbb{R}^{W \times H \times 3}$ , a categorical navigational command $c \in \{1, \dots, 6\}$ (i.e., turn left, turn right, follow, forward, left lane changing and right lane changing) and an intermediate noisy goal $\mathbf{g} \in \mathbb{R}^2$ sampled from a GNSS (Global Navigation Satellite System) [11]. We note that both the command and GNSS observations can be easily obtained by today's vehicles, i.e., through generation of an $\mathrm{A}^*$ plan [26] from a static map and coarse positioning system. The GNSS goals are sampled every 50-100 meters and reflect real-world measurement errors [11]. We compute a BEV $\mathbf{B} \in \{0, 1\}^{W_B \times H_B \times C_B}$ by rendering privileged (i.e., ground-truth) information from the underlying simulation, including 3D location and state of lanes, pedestrians, vehicles, and traffic lights [79]. While prior work has generally leveraged a standardized BEV representation [13, 32, 34, 57, 76], we find its design to be crucial to training an optimal teacher policy as will be discussed in Sec 3.2. We let $\mathbf{x}^s = (\mathbf{I}, \mathbf{g}, c) \in \mathcal{X}^s$ and $\mathbf{x}^t = (\mathbf{B}, \mathbf{g}, c) \in \mathcal{X}^t$ be the student and teacher observations, respectively. Given these observations at each time step, the agent learns to predict 10 future 2D waypoints in top-down vehicle coordinates for the next 2.5 seconds of driving. These are then given to a lateral and longitudinal PID controllers [13, 48, 66] to generate the final action. + +Given this formulation, we next discuss our teacher agent design and training process. In particular, we aim to learn a privileged agent which can not only effectively solve the goal-oriented navigation task, but also facilitate effective distillation and coaching of a sensorimotor student policy, as will be discussed in Sec. 3.4. + +# 3.2. Learning an Effective Teacher + +In our formulation, learning a student agent begins with training an effective privileged agent $f_{\psi}^{t}$ . However, we find this itself to be a complex task. Even with complete knowledge of perfect BEV perception, current privileged agents are suboptimal, significantly under-performing CARLA's built-in autopilot [13, 79]. In this section, we uncover the underlying reason for this under-performance through the BEV design. Informed by this key finding, we train a highly effective imitation-learned teacher that surpasses prior privileged agents and matches expert-level performance in CARLA without requiring extensive data collection [56] or reinforcement learning [79]. Our proposed privileged agent can then enable effective supervision of the student in Sec. 3.3. + +![](images/5c949646a75d96ca06d760734237b42d48f99536cc947e87bd81106f55f5696c.jpg) +Figure 2. Overview of Architecture and Distillation Process of CaT. We first train a high-capacity privileged teacher agent (colored in red) which takes an augmented BEV representation as input, and produces a set of future waypoints conditioned on command and short-term goal. Our teachable student agent (colored in blue) maps the image features to the BEV space using an alignment module, which is trained by distillation of intermediate features (arrows colored in blue) and final output (arrow colored in green). Additionally, we leverage supervision from auxiliary segmentation (arrow colored in red) and command prediction tasks (arrow colored in purple) interlaced throughout the network to facilitate the learning of task-relevant features. + +Teacher Training via Direct Expert Imitation: In imitation learning, an oracle (i.e., expert driver) $f^{*} \colon \mathcal{X}^{t} \to \mathcal{Y}$ is defined for demonstrating optimal actions $\mathbf{y}^{*}$ . In CARLA [21], the expert is defined using longitudinal and lateral PID controllers which are carefully tuned and augmented by a collection of manually specified rules for handling diverse scenarios (e.g., sudden braking) over the underlying simulation state. The designed expert can then be used to generate trajectories as supervised data for learning the privileged policy function $f_{\psi}^{t}$ with behavior cloning [18, 19, 48]. Our privileged agent is a ResNet-18 model [28], as shown in Fig. 2 (further increasing modeling capacity was not found to be beneficial). The ResNet model is followed by a GRU-based conditional waypoint predictor [11, 16]. Due to the complexity of our dense urban navigation task, e.g., with multi-lane roads, intersections, and merging, we incrementally refine the waypoints using two GRUs within each conditional branch. The first GRU regresses the preliminary set of waypoints via sequential waypoints directly from the embedded features. Subsequently, the second GRU takes the predicted waypoints together with the short-term goal and embedded features, to produce a refined set of waypoint targets. + +Even assuming perfect perception, we find it difficult to train privileged agents that exhibit robust planning behavior in complex urban driving scenarios. Specifically, when using the most challenging CARLA benchmark (Longest6 [16], also detailed in Sec 4) using the standard BEV representation of Chen et al. [13] leads to $26\%$ driving score compared to $72\%$ by the rule-based expert. The rein + +forcement learning agent of Zhang et al. [79] achieves $60\%$ on this task. We sought to explore the limits of behavior cloning for this task, as a noisy and sub-optimal teacher can hinder the training of the student model. In particular, we hypothesize that the reason for the poor performance lies in the increased complexity of the BEV state representation in dense scenarios. In such scenarios, learning to extract task-relevant details can become more challenging. To address this issue, we propose to introduce additional channels (akin to affordances [9,22,58]) that can more easily translate into safety-critical decision-making, simplify the learning task, and provide an expert-level teaching agent. + +BEV with Safety Hints: Our BEV encodes state information into distinct channels, including drivable road regions, the desired route based on the command and layout, lane markings, and dynamic obstacles (see Fig. 3). Given the low performance of prior privileged agents, in particular around dynamic objects, the agent could benefit from more explicit safety-driven cues in the BEV. We propose to add two types of channels of (1) predicted agents' future and (2) entity attention. First, we utilize a kinematics bicycle model [16] in order to efficiently predict future trajectories of dynamic objects. This enables us to iteratively predict and represent short-term future position, orientation, and speed of agents (our supplementary contains additional information regarding this process). Secondly, we encode an explicit attention channel for highlighting potential future infractions, i.e., based on the kinematics bicycle model applied to the ego-vehicle. While this assumes that the current vehicle throttle and steering remain unchanged, the entity + +![](images/73eeaae3b8b28898a1237ee7e1faff2380488617b79ab9d151c46b90ce0bf569.jpg) +(a) Baseline [79] + +![](images/fc806bd673925eeb5ae8eb6eb6e21ed36a083ed94b230d75d84906c3936f1bac.jpg) + +![](images/fd79ed0d0f73eb8661bc18da399b93b74e16fcf3bb024c507ad7c3d9b8878c47.jpg) + +![](images/cd4190bd9b1aaec56b3c88657eececf62a2e7490db02bd9413bb5a621e807bf5.jpg) +(b) With Safety Hints +(c) Road +(g) Pedestrians +Figure 3. Visualization of Our Proposed BEV with Safety Hints. (a) and (b) visualize the baseline and our proposed BEV by encoding different channels into colors. Roads are shown in dark grey, Desired Route is shown in light grey, Lane Marks are shown in magenta, Vehicles are shown in blue with past trajectories shown in decreasing saturation with time into the past, Pedestrians are shown in light blue, Traffic Lights are shown in red or green according to its state, Agent Forecasts are shown in purple with decreasing saturation with time into the future, Entity Attention is shown in orange. (c) - (j) shows the different channels separately. + +![](images/b4024c7fd22bea493e03de456ef1978ad97234ae0eae693bc839cc22b7dc9ec7.jpg) + +![](images/bbc8e8a2da2d118bb459310c0bd2db244beb0838983a1409943cfb677f16caf6.jpg) +(d) Desired Route +(h) Traffic Lights + +![](images/36f96d7fddefc58add70214537bc7032084d60e96b581ce078d55d2d5a9c8867.jpg) + +![](images/02093be44a605a22045abe59041659a6bae0e47327c3c8ca5ca8c9f9b6a5ca10.jpg) +(e) Lane Marks +(i) Agent Forecast + +![](images/0389cadfe8fd59778bcc6fc88d06cf550ebdb972cd0a5ba6593af2cb91120a97.jpg) + +![](images/e2ac63a43c8db4a86b4ff9dabc0741bf9773f5b501f128474f7a30ef018291a3.jpg) +(f) Vehicles +(j) +Entity Attention + +attention channel helps the agent better reason over safety while correlating states with actions. Surprisingly, the additions of the safety hints channels provide a strong teacher with offline behavior cloning alone, even improving over the rule-based expert (Table 1, $73\%$ vs. $72\%$ ). We note that many of the remaining failures are due to simulation timeout in crowded intersections. Our boosted teacher agent will be critical for effective distillation in Sec. 3.4. + +# 3.3. Learning a Teachable Student + +Due to differing inputs and modeling capacities, it can be difficult to align the image-based student features and output with the BEV-based privileged teacher. We propose to alleviate misalignment by designing a teachable student network, i.e., a model that matches as much of the teacher architecture as possible. Specifically, we propose to improve teacher-student model alignment by leveraging an internal image-to-BEV alignment module. As shown in Fig. 2, the three residual blocks following the alignment module can consequently facilitate knowledge transfer via direct distillation of most of the teacher's features in Sec. 3.4. + +IPM-based Transformer Alignment Module: Explicitly mapping images to BEV space for downstream tasks can be challenging. While this can be achieved via depth estimation and lifting [32,53], we sought to incorporate an efficient and differentiable module which can effectively attend and reason over various features while mapping from front view to BEV. A transformer-based architecture [14,65] provides a natural choice as it can be optimized end-to-end with respect to the driving task while also carefully aligning features among spaces with arbitrary dimensionalities. As shown in Fig. 2, we sample queries $\mathbf{Q}_{init} \in \mathbb{R}^{L_H \times L_W \times L_C}$ using a spatial parameterization of the BEV space and process them using a self-attention module [14,65]. The output + +of self-attention module $Q$ can be formulated as + +$$ +\mathbf {Q} = \operatorname {s o f t m a x} \left(\frac {\mathbf {Q} _ {\text {i n i t}} \mathbf {K} _ {\text {i n i t}} ^ {T}}{\sqrt {d}}\right) \mathbf {V} _ {\text {i n i t}} \tag {2} +$$ + +where $\mathbf{K}_{init}$ , $\mathbf{V}_{init} \in \mathbb{R}^{L_H \times L_W \times L_C}$ provide the key-value pairs and $d$ is the dimension of the query. Subsequently, each query point $\mathbf{q} \in \mathbf{Q}$ is mapped to its corresponding reference point $\mathbf{p} \in \mathbb{R}^3$ (homogeneous coordinates) in the image features $\mathcal{F}_{RGB} \in \mathbb{R}^{Z_H \times Z_W \times Z_C}$ using an Inverse Perspective Mapping (IPM) + +$$ +\mathbf {p} = s \mathbf {P} _ {k} \mathbf {R} _ {k} (\mathbf {q} - \mathbf {t} _ {k}) \tag {3} +$$ + +where $s$ is a scale factor, $\mathbf{P}_k\in \mathbb{R}^{3\times 3}$ , $\mathbf{R}_k\in \mathbb{R}^{3\times 3}$ , and $\mathbf{t}_k\in \mathbb{R}^3$ are the $k^{th}$ camera intrinsics, rotation and translation relative to the center of the ego-vehicle, respectively [6]. Note that we employ a different IPM mapping for each camera view to perform the projection. + +To populate the student's BEV features $\mathcal{F}_{BEV} \in \mathbb{R}^{H_B \times W_B \times C_B}$ (i.e., the output of the alignment module in Fig. 2), we leverage a deformable cross-attention mechanism based on deformable DETR [85]. This enables the network to attend to multiple regions around the reference point in the image features when transforming to BEV feature space. The reference point $\mathbf{p}$ and the learned deformable offsets can then be used as keys $\mathbf{K}$ , with their corresponding features from the extracted images features the values $\mathbf{V}$ , for a cross-attention module + +$$ +\mathcal {F} _ {B E V} = \operatorname {D e f o r m A t t n} (\mathbf {Q}, \mathcal {F} _ {R G B}, \mathbf {H}) \tag {4} +$$ + +where $\mathbf{H} = \{\mathbf{P}_k, \mathbf{R}_k, t_k\}_{k=1}^3$ contains the combined multiview IPM parameters from Eqn. 3. In addition to the proposed alignment and teachable modules, we also incorporate prediction modules for auxiliary tasks (i.e., BEV prediction [16]), discussed in Sec. 3.4. + +GRU-based Waypoint Predictor: The transformed imaged-to-BEV features are inputted to three residual blocks and a GRU-based conditional branch based on the command [11]. + +# 3.4. Student Loss + +Even with our carefully designed architecture, the student's image-based learning task remains significantly more challenging than the teacher's. Thus, both optimization and alignment with the teacher can benefit from extensive supervision and gradual student instruction, i.e., to scaffold supervision and better match the student's ability. Our distillation process provides ample beneficial supervision for the student through three holistic mechanisms. First, we incorporate deep distillation losses, i.e., both over output (i.e., waypoint regression task) and feature maps, as shown in Fig. 2, that can more directly supervise internal layers. Second, we incorporate additional auxiliary losses that regularize the optimization process. This includes ground-truth supervised losses, e.g., defined over segmentation and control command, and a geometric loss defined over the post-alignment features that supervises for correct BEV structure. Finally, we leverage a student-paced coaching mechanism for gradually increasing the difficulty of the waypoint prediction task throughout the iterations in Sec. 3.5. + +Loss Functions: Our optimization objective for guiding the distillation process is a weighted sum over both distillation and auxiliary tasks: + +$$ +\mathcal {L} _ {C a T} = \mathcal {L} _ {\text {o u t}} + \mathcal {L} _ {\text {f e a t}} + \mathcal {L} _ {\text {s e g}} + \mathcal {L} _ {\text {c m d}} \tag {5} +$$ + +The various loss terms enable guiding holistic aspects of student learning. + +Output Distillation: We leverage an $L_{1}$ loss computed over all of the conditional branches for each instance + +$$ +\mathcal {L} _ {o u t} = \sum_ {c = 1} ^ {C} \| f _ {\boldsymbol {\theta}} ^ {s} \left(\mathbf {x} ^ {s}, c\right) - f _ {\boldsymbol {\psi}} ^ {t} \left(\mathbf {x} ^ {t}, c\right) \| _ {1} \tag {6} +$$ + +This can be done by sampling from the teacher with different commands $c$ (following Chen et al. [13]). + +Feature Distillation Loss: Our proposed student architecture provides BEV-space feature alignment, i.e., for direct internal feature matching between the student and the teacher (the first term in our loss in Eqn. 7). This is in contrast to current methods which distill the output or a final fully-connected layer [11, 13, 71, 79]. However, while this can provide rich supervision for the student, the two models should not match entirely, i.e., due to differences in information processing for a perception-based planning task. Moreover, the BEV-based features can be highly sparse and structured, with slight feature offsets being less meaningful than obtaining a task and scene-relevant representation. To flexibly account for the student-based task and effectively represent structural information in BEV space, our feature loss is computed as + +$$ +\begin{array}{l} \mathcal {L} _ {f e a t} = \sum_ {i = 1} ^ {3} \left[ \| \mathcal {F} _ {i} ^ {s} (\mathbf {x} ^ {s}) - \mathcal {F} _ {i} ^ {t} (\mathbf {x} ^ {t}) \| _ {2} + \right. \\ \left\| T _ {i} ^ {s} \left(\mathcal {F} _ {i} ^ {s} \left(\mathbf {x} ^ {s}\right)\right) - T _ {i} ^ {t} \left(\mathcal {F} _ {i} ^ {t} \left(\mathbf {x} ^ {t}\right)\right) \right\| _ {2} + \tag {7} \\ \lambda_ {C D} \left\| \mathcal {F} _ {i} ^ {s} (\mathbf {x} ^ {s}) - \mathcal {F} _ {i} ^ {t} (\mathbf {x} ^ {t})\right) \| _ {C D} \Bigg ] \\ \end{array} +$$ + +where we use $\mathcal{F}_i$ to denote layers within the networks (as shown in Fig. 2, we directly distill three layers), $T_{i}$ indicates the features following several convolutional layers (trained jointly with the student network model), and $CD$ stands for a Chamfer Distance [51] (we set $\lambda_{CD} = 0.1$ ). The final term is computed over thresholded activations and a spatial soft-argmax [13] from each feature map. + +Task-oriented Auxiliary Tasks: To ensure essential task-relevant information is preserved throughout the distillation process, we leverage regularizing supervision in the form of segmentation and command-based auxiliary tasks interlaced throughout the network [16, 49]. Our segmentation loss $\mathcal{L}_{seg}$ is a cross-entropy loss, computed over the ground-truth BEV and averaged across the segmentation prediction heads. Our command prediction loss $\mathcal{L}_{cmd}$ is binary cross-entropy supervised by the ground-truth command $c$ , which facilitates learning task-relevant features [11, 19]. + +# 3.5. Student-paced Coaching + +While Sec. 3.4 provides ample supervision on the student's internal representations, the teacher's targets may still be difficult to imitate for the student. To better consider the learning ability of the student, we propose to leverage a student-paced training mechanism which can gradually coach the student, i.e., through increasingly difficult samples. When computing the loss in Eqn. 5, we define a coach, + +$$ +\mathcal {F} ^ {t} \leftarrow \lambda_ {i} \mathcal {F} ^ {s} + \left(1 - \lambda_ {i}\right) \mathcal {F} ^ {t}, \text {i f} \mathcal {L} _ {C a T} > \tau_ {i} \tag {8} +$$ + +which interpolates the teaching targets with the student's predictions and features. $\lambda_{i}$ is a hyperparameter which is linearly decreased to 0 over training iteration $i$ . The modified targets are only computed over the hard samples at each batch. Within each batch, $\tau_{i}$ is a threshold defining the lowest $50\%$ of the scored samples for the loss with respect to the privileged teacher. While smoothing the targets in this manner may seem counterintuitive, i.e., compared to techniques which mine hard examples, the coaching mechanism aims to stabilize training by reducing the difficulty when the student is unable to perform the optimal action. Eqn. 8 enables adjusting the learning rate in a sample-selective manner (initially suppressing difficult samples). + +# 4. Experiments + +We use the CARLA simulator (version 0.9.10.1) [21] for data generation and closed-loop evaluation of the proposed CaT framework. We also use open-loop evaluation + +Table 1. Quantitative Evaluation on the Longest6 Benchmark. Comparison of CaT with prior methods in terms of Driving Score (DS), Route Completion (RC), and Infraction Score (IS). Additional infraction metrics (Pedestrian Collisions (Ped), Vehicle Collisions (Veh), Layout Collisions (LC), Red Light Violations (Red), Off-road Infraction (OR), Route deviation (Dev), Route Timouts (TO), Agent Blocked (Blk)) are shown. FD refers to Feature Distillation, SH refers to Safety Hints. *re-trained by us using the publicly available code. Mean and standard deviation are computed over three runs. CaT outperforms the state-of-the-art by a large margin in terms of DS. + +
MethodRGBLiDARDS↑RC↑IS↑Ped↓Veh↓LC↓Red↓OR↓Dev↓TO↓Blk↓
LAV [11]✓✓48.41±3.4080.71±0.840.60±0.040.000.500.190.070.200.070.010.29
TransFuser [16]✓✓46.20±2.5783.61±1.160.57±0.000.290.380.280.040.070.070.000.32
WOR [10]✓X17.36±2.9543.46±2.990.54±0.060.050.640.150.840.150.890.040.45
NEAT [15]✓X24.08±3.3059.94±0.500.49±0.020.010.710.210.180.180.000.022.83
TCP* [71]✓X42.86±0.6361.83±4.190.71±0.040.010.420.110.040.080.010.010.75
CaT (w/o Alignment, Coaching, FD)✓X39.48±0.6760.96±1.650.68±0.010.031.290.200.080.840.000.021.49
CaT (w/o Alignment, Coaching)✓X40.64±0.9862.45±0.460.67±0.010.021.070.330.140.380.000.021.11
CaT (w/o Coaching, FD, SH)✓X44.10±0.4065.84±5.550.72±0.030.010.260.050.110.150.000.000.66
CaT (w/o Coaching, SH)✓X49.69±2.2881.10±0.580.64±0.020.010.780.030.080.200.000.020.51
CaT (w/o Coaching)✓X55.55±1.4181.97±2.340.68±0.010.020.300.050.070.060.000.020.35
CaT✓X58.36±2.2478.79±1.500.77±0.020.010.200.020.050.300.000.040.44
Privileged Agents:
RL Expert (Roach) [79]--60.14±2.4085.83±0.600.69±0.030.060.220.000.010.000.000.040.07
Rule-based Expert--71.96±2.1377.46±3.110.91±0.000.010.060.000.000.000.000.030.41
Basic BEV Agent [13]--24.08±2.8373.36±1.080.31±0.060.011.452.450.090.270.110.010.24
+ History and Desired Path--52.81±1.7979.34±3.650.71±0.060.020.300.000.000.000.000.020.53
+ Agent Forecast--65.73±0.9383.50±1.180.79±0.020.010.210.000.020.000.000.010.28
+ Entity Attention--73.30±1.0787.44±0.280.83±0.020.050.110.000.020.000.000.020.12
+ +on nuScenes [7, 33, 77] (ADE, FDE, and collision rate). In Carla, we leverage the Longest6 Benchmark [16], which uses the six longest routes of each town (Town01 - Town06) from the set of 76 routes provided by the official CARLA leaderboard [1] (a total of 36 routes). We note in addition to the standard cross weather and time of day generalization evaluation, the long and dense traffic conditions in Longest6 present the most challenging settings across the various CARLA-based benchmarks. To evaluate our models, we follow standard metrics and report Route Completion (RC, in terms of completed route percentage), Infraction Score (IS, a penalty factor over infractions), and Driving Score (DS, computed from the prior two) [1, 11, 16]). + +# 4.1. Comparison with Prior Methods + +CARLA Results: As shown in Table 1, using feature distillation, safety hints-based BEV for the teacher, and the proposed coaching mechanism, CaT is able to obtain a $58.36\%$ DS. Specifically, we achieve state-of-the-art performance among all prior agents, including LiDAR-based approaches [11, 16] (by $20.6\%$ in terms of DS, from 48.41 to 58.36, and $28.3\%$ in terms of IS). We also note that such approaches provide strong baselines, due to various LiDAR-based safety checks that the agents perform on top of the trained policy, which we do not employ. Moreover, CaT outperforms the prior RGB-only state-of-the-art agent TCP [71] by $36.16\%$ DS. To further validate the generalization of CaT, we discuss an additional benchmark split from TCP [71] and LAV [11] in the supplementary. We also find that removal of the alignment module degrades DS from $44.10\%$ to $39.48\%$ , and is shown to only slightly benefit from feature distillation ( $40.64\%$ DS) indicating the effectiveness of the proposed architecture. + +Table 2. Open-Loop Evaluation on nuScenes. + +
MethodADE (m) ↓FDE (m) ↓Coll. (%) ↓
BEV Agent0.330.520.49
CaT (w/o Coaching, FD, SH)0.480.430.68
CaT0.410.360.27
+ +Real-World Evaluation: To further analyse the benefits of CaT in realistic driving settings, Table 2 shows open-loop evaluation (ADE, FDE, and Collision rate [77]) for nuScenes [7]. On the official validation split, the privileged agent performs best with an ADE of 0.33. CaT achieves 0.41 ADE, improving by $14.6\%$ gain over an agent without distillation and coaching. Moreover, we find a significant reduction in the collision rate by $60.3\%$ compared to the baseline. Additional details regarding nuScenes evaluation can be found in the supplementary. + +# 4.2. Ablation Studies + +Teacher Comparison: Table 1 depicts our teacher model ablation. Specifically, we find the proposed BEV channels to drastically improve the privileged agent's performance by simplifying the learning task (73.30% DS, surpassing the built-in rule-based expert of 71.96%). We note that our behavior cloning agent also significantly outperforms an RL-based expert [79] (60.14%), simply through an effective BEV design. Thus, our findings apply to real-world scenarios where interactive agent training may be unsafe. We also observe that the improved teacher benefits the student agent. Specifically, incorporating the proposed BEV-based safety hints results in student performance gains, from 49.69% to 55.55%. + +Impact of Feature Distillation: Table 3 presents the benefits of our proposed multi-layer feature distillation frame + +![](images/be6e47ce4a4140cd30f20b8992747d2d209dbabb2f92c0ab24a41a7f3729d30a.jpg) +Figure 4. Visualization of Model Input and Features. We show the BEV ground truth (BEV) and RGB image ground truth (Image) on the top row. Next, we visualize the intermediate layer activations of a sensorimotor agent without the proposed alignment module [71] (No Alignment Module), privileged teacher agent (Privileged), student agent trained with output distillation only (Output Distillation), and the proposed agent (CaT). We demonstrate that our teacher agent with proposed knowledge distillation is able to guide the student agent in learning scene-relevant representations and attend to safety-critical entities (circled in red). + +Table 3. Ablation Study on Feature Distillation Layers. All loss terms in Eqn. 7 benefit training with deep feature distillation. + +
MethodDS ↑RC ↑IS ↑
No Distillation44.1065.840.72
One Layer [71,79]45.2369.330.68
Three Layers L249.3166.920.78
Three Layers L2 + LCD51.9562.820.87
Three Layers Lfeat55.5581.970.68
+ +work. We note that prior methods only distill one final fully connected layer (before the prediction head [71, 79], referred to as 'one layer' in Table 3). We note that such approaches result in minimal gains, whereas CaT is able to benefit from the extensive feature distillation (improving from $45.23\%$ to $55.55\%$ ). Incorporating multiple types of feature distillation losses also benefits performance. + +# 4.3. Feature Visualization + +Fig. 4 depicts the learned features of the initial feature distilled layer (out of the three in Fig. 2). We visualize activations from two channels, to better understand the learned representation by the agents. Interestingly, the visualization shows how traditionally trained sensorimotor agents (i.e., without the proposed alignment module) mostly operate in image-space features in this layer despite the BEV waypoint prediction task. This validates our main hypothesis, that such agents have limited ability to learn from a BEV-based agent, even with BEV-based auxiliary prediction tasks. We + +also find that output distillation provides a weak supervision for critical components of the BEV, including layout and object features. After feature distillation, the features of our CaT model are able to not only keep the driving-related semantic information i.e., drivable road and route, but also focus on safety-critical entities. More visualization examples are provided in the supplementary. + +# 5. Conclusion + +We present CaT, a novel knowledge distillation method for effectively training a sensorimotor student agent using the supervision from a privileged agent. We leverage an alignment module to better map image features to the BEV space, thus enabling extensive supervision from a BEV-based teacher over the intermediate feature learning. As both agents can be trained offline via imitation learning, our findings are directly relevant to real-world settings where on-policy interactions may be unsafe to perform. Finally, we carefully account for inherent differences between the student and the teacher using a student-paced coaching mechanism with various auxiliary supervision tasks. Through the improved knowledge distillation process, our experiments result in a state-of-the-art camera-based agent in CARLA. Our ablation into input and model design guides future directions through interpretable analysis into current limitations of end-to-end driving models. + +Acknowledgments: We thank the Red Hat Collaboratory for supporting this research. + +# References + +[1] Carla autonomous driving leaderboard. https://leaderboard.carla.org/, 2022.7 +[2] Yonathan Aflalo, Asaf Noy, Ming Lin, Itamar Friedman, and Lihi Zelnik. Knapsack pruning with inner distillation. arXiv preprint arXiv:2002.08258, 2020. 2 +[3] Gustavo Aguilar, Yuan Ling, Yu Zhang, Benjamin Yao, Xing Fan, and Chenlei Guo. Knowledge distillation from internal representations. In AAAI, 2020. 2 +[4] Mayank Bansal, Alex Krizhevsky, and Abhijit Ogale. ChauffeurNet: Learning to drive by imitating the best and synthesizing the worst. In RSS, 2019. 1 +[5] Yoshua Bengio, Jérôme Louradour, Ronan Collobert, and Jason Weston. Curriculum learning. In ICML, 2009. 2 +[6] Massimo Bertozzi, Alberto Broggi, and Alessandra Fascioli. An extension to the inverse perspective mapping to handle non-flat roads. In IV, 1998. 5 +[7] Holger Caesar, Varun Bankiti, Alex H Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuscenes: A multimodal dataset for autonomous driving. In CVPR, 2020. 7 +[8] Raphael Chekroun, Marin Toromanoff, Sascha Hornauer, and Fabien Moutarde. Gri: General reinforced imitation and its application to vision-based autonomous driving. arXiv preprint arXiv:2111.08575, 2021. 2 +[9] Chenyi Chen, Ari Seff, Alain L. Kornhauser, and Jianxiong Xiao. DeepDriving: Learning affordance for direct perception in autonomous driving. In ICCV, 2015. 2, 4 +[10] Dian Chen, Vladlen Koltun, and Philipp Krahenbuhl. Learning to drive from a world on rails. In ICCV, 2021. 2, 7 +[11] Dian Chen and Philipp Krahenbuhl. Learning from all vehicles. In CVPR, 2022. 1, 2, 3, 4, 5, 6, 7 +[12] Defang Chen, Jian-Ping Mei, Yuan Zhang, Can Wang, Zhe Wang, Yan Feng, and Chun Chen. Cross-layer distillation with semantic calibration. In AAAI, 2021. 2 +[13] Dian Chen, Brady Zhou, Vladlen Koltun, and Philipp Krahenbuhl. Learning by cheating. In CoRL, 2020. 1, 2, 3, 4, 6, 7 +[14] Li Chen, Chonghao Sima, Yang Li, Zehan Zheng, Jiajie Xu, Xiangwei Geng, Hongyang Li, Conghui He, Jianping Shi, Yu Qiao, et al. Persformer: 3d lane detection via perspective transformer and the openlane benchmark. arXiv preprint arXiv:2203.11089, 2022. 2, 5 +[15] Kashyap Chitta, Aditya Prakash, and Andreas Geiger. Neat: Neural attention fields for end-to-end autonomous driving. In ICCV, 2021. 7 +[16] Kashyap Chitta, Aditya Prakash, Bernhard Jaeger, Zehao Yu, Katrin Renz, and Andreas Geiger. Transfuser: Imitation with transformer-based sensor fusion for autonomous driving. arXiv preprint arXiv:2205.15997, 2022. 3, 4, 5, 6, 7 +[17] Inseop Chung, SeongUk Park, Jangho Kim, and Nojun Kwak. Feature-map-level online adversarial knowledge distillation. In ICML, 2020. 2 +[18] Felipe Codevilla, Matthias Muiller, Antonio López, Vladlen Koltun, and Alexey Dosovitskiy. End-to-end driving via conditional imitation learning. In ICRA, 2018. 3, 4 + +[19] Felipe Codevilla, Eder Santana, Antonio M López, and Adrien Gaidon. Exploring the limitations of behavior cloning for autonomous driving. In ICCV, 2019. 4, 6 +[20] Pim de Haan, Dinesh Jayaraman, and Sergey Levine. Causal confusion in imitation learning. In NeurIPS, 2019. 1 +[21] Alexey Dosovitskiy, German Ros, Felipe Codevilla, Antonio Lopez, and Vladlen Koltun. CARLA: An open urban driving simulator. In CoRL, 2017. 2, 3, 4, 6 +[22] James J Gibson. The ecological approach to visual perception: classic edition. Psychology press, 2014. 4 +[23] Jianping Gou, Baosheng Yu, Stephen J Maybank, and Dacheng Tao. Knowledge distillation: A survey. In IJCV, 2021. 1, 3 +[24] Sheng Guo, Weilin Huang, Haozhi Zhang, Chenfan Zhuang, Dengke Dong, Matthew R Scott, and Dinglong Huang. Curriculumnet: Weakly supervised learning from large-scale web images. In ECCV, 2018. 2 +[25] Saurabh Gupta, James Davidson, Sergey Levine, Rahul Sukthankar, and Jitendra Malik. Cognitive mapping and planning for visual navigation. In CVPR, 2017. 2 +[26] Peter E Hart, Nils J Nilsson, and Bertram Raphael. A formal basis for the heuristic determination of minimum cost paths. IEEE transactions on Systems Science and Cybernetics, 1968, 3 +[27] He He, Jason Eisner, and Hal Daume. Imitation learning by coaching. In NeurIPS, 2012. 1, 2, 3 +[28] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, 2016. 4 +[29] Byeongho Heo, Jeesoo Kim, Sangwoo Yun, Hyojin Park, No-jun Kwak, and Jin Young Choi. A comprehensive overhaul of feature distillation. In ICCV, 2019. 2 +[30] Roger Hiemstra and Burton Sisco. Individualizing Instruction. Making Learning Personal, Empowering, and Successful. 1990. 2 +[31] Geoffrey Hinton, Oriol Vinyals, Jeff Dean, et al. Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531, 2(7), 2015. 2 +[32] Anthony Hu, Zak Murez, Nikhil Mohan, Sofia Dudas, Jeffrey Hawke, Vijay Badrinarayanan, Roberto Cipolla, and Alex Kendall. Fiery: Future instance prediction in bird's-eye view from surround monocular cameras. In ICCV, 2021. 2, 3, 5 +[33] Peiyun Hu, Aaron Huang, John Dolan, David Held, and Deva Ramanan. Safe local motion planning with self-supervised freespace forecasting. In CVPR, 2021. 7 +[34] Zanming Huang, Zhongkai Shangguan, Jimuyang Zhang, Gilad Bar, Matthew Boyd, and Eshed Ohn-Bar. ASSISTER: Assistive navigation via conditional instruction generation. In ECCV, 2022. 3 +[35] Zehao Huang and Naiyan Wang. Like what you like: Knowledge distill via neuron selectivity transfer. arXiv preprint arXiv:1707.01219, 2017. 2 +[36] Mingi Ji, Byeongho Heo, and Sungrae Park. Show, attend and distill: Knowledge distillation via attention-based feature matching. In AAAI, 2021. 2 + +[37] Lu Jiang, Zhengyuan Zhou, Thomas Leung, Li-Jia Li, and Li Fei-Fei. Mentornet: Learning data-driven curriculum for very deep neural networks on corrupted labels. In ICML, 2018. 2 +[38] Jangho Kim, SeongUk Park, and Nojun Kwak. Paraphrasing complex network: Network compression via factor transfer. In NeurIPS, 2018. 2 +[39] Tom Kocmi and Ondrej Bojar. Curriculum learning and minibatch bucketing in neural machine translation. arXiv preprint arXiv:1707.09533, 2017. 2 +[40] Animesh Koratana, Daniel Kang, Peter Bailis, and Matei Zaharia. Lit: Learned intermediate representation training for model compression. In ICML, 2019. 2 +[41] M Kumar, Benjamin Packer, and Daphne Koller. Self-paced learning for latent variable models. In NeurIPS, 2010. 2 +[42] Ravi Kumar Kushawaha, Saurabh Kumar, Biplab Banerjee, and Rajbabu Velmurugan. Distilling spikes: Knowledge distillation in spiking neural networks. In ICPR, 2021. 2 +[43] Sergey Levine, Chelsea Finn, Trevor Darrell, and Pieter Abbeel. End-to-end training of deep visuomotor policies. The Journal of Machine Learning Research, 17(1):1334-1373, 2016. 1 +[44] Dong Li, Jia-Bin Huang, Yali Li, Shengjin Wang, and Ming-Hsuan Yang. Weakly supervised object localization with progressive domain adaptation. In CVPR, 2016. 2 +[45] Guohao Li, Matthias Mueller, Vincent Casser, Neil Smith, Dominik L Michels, and Bernard Ghanem. Oil: Observational imitation learning. In RSS, 2019. 2 +[46] Yang Liu, Wei Zhang, and Jun Wang. Source-free domain adaptation for semantic segmentation. In CVPR, 2021. 2 +[47] Takahiro Miki, Joonho Lee, Jemin Hwangbo, Lorenz Wellhausen, Vladlen Koltun, and Marco Hutter. Learning robust perceptive locomotion for quadrupedal robots in the wild. Science Robotics, 7(62), 2022. 1 +[48] Matthias Müller, Alexey Dosovitskiy, Bernard Ghanem, and Vladlen Koltun. Driving policy transfer via modularity and abstraction. In CoRL, 2018. 3, 4 +[49] Eshed Ohn-Bar, Aditya Prakash, Aseem Behl, Kashyap Chitta, and Andreas Geiger. Learning situational driving. In CVPR, 2020. 6 +[50] Takayuki Osa, Joni Pajarinen, Gerhard Neumann, J Andrew Bagnell, Pieter Abbeel, Jan Peters, et al. An algorithmic perspective on imitation learning. Foundations and Trends® in Robotics, 2018. 2, 3 +[51] Despoina Paschalidou, Angelos Katharopoulos, Andreas Geiger, and Sanja Fidler. Neural parts: Learning expressive 3D shape abstractions with invertible neural networks. In CVPR, 2021. 6 +[52] Nikolaos Passalis and Anastasios Tefas. Learning deep representations with probabilistic knowledge transfer. In ECCV, 2018. 2 +[53] Jonah Philion and Sanja Fidler. Lift, splat, shoot: Encoding images from arbitrary camera rigs by implicitly unprojecting to 3D. In ECCV, 2020. 2, 5 +[54] Dean A Pomerleau. ALVINN: An autonomous land vehicle in a neural network. In NeurIPS, 1989. 1, 2 + +[55] Adriana Romero, Nicolas Ballas, Samira Ebrahimi Kahou, Antoine Chassang, Carlo Gatta, and Yoshua Bengio. Fitnets: Hints for thin deep nets. arXiv preprint arXiv:1412.6550, 2014. 2 +[56] Stéphane Ross, Geoffrey Gordon, and Drew Bagnell. A reduction of imitation learning and structured prediction to no-regret online learning. In AISTATS, 2011. 2, 3 +[57] Avishkar Saha, Oscar Mendez, Chris Russell, and Richard Bowden. Translating images into maps. In ICRA, 2022. 3 +[58] Axel Sauer, Nikolay Savinov, and Andreas Geiger. Conditional affordance learning for driving in urban environments. In CoRL, 2018. 4 +[59] Miaojing Shi and Vittorio Ferrari. Weakly supervised object localization using size estimates. In ECCV, 2016. 2 +[60] Suraj Srinivas and Francois Fleuret. Knowledge transfer with Jacobian matching. In ICML, 2018. 2 +[61] Yuxing Tang, Xiaosong Wang, Adam P Harrison, Le Lu, Jing Xiao, and Ronald M Summers. Attention-guided curriculum learning for weakly supervised classification and localization of thoracic diseases on chest radiographs. In MLMI, 2018. 2 +[62] Marin Toromanoff, Emilie Wirbel, and Fabien Moutarde. End-to-end model-free reinforcement learning for urban driving using implicit affordances. In CVPR, 2020. 2 +[63] Radu Tudor Ionescu, Bogdan Alexe, Marius Leordeanu, Marius Popescu, Dim P Papadopoulos, and Vittorio Ferrari. How hard can it be? estimating the difficulty of visual search in an image. In CVPR, 2016. 2 +[64] Rachel R Van Der Stuyf. Scaffolding as a teaching strategy. Adolescent learning and development, 2002. 2 +[65] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In NeurIPS, 2017. 5 +[66] Antonio Visioli. Practical PID control. Springer Science & Business Media, 2006. 3 +[67] Tao Wang, Li Yuan, Xiaopeng Zhang, and Jiashi Feng. Distilling object detectors with fine-grained feature imitation. In CVPR, 2019. 2 +[68] Wenhui Wang, Hangbo Bao, Shaohan Huang, Li Dong, and Furu Wei. Minilmv2: Multi-head self-attention relation distillation for compressing pretrained transformers. arXiv preprint arXiv:2012.15828, 2020. 2 +[69] Wei Wang, Isaac Caswell, and Ciprian Chelba. Dynamically composing domain-data selection with clean-data selection by "co-curricular learning" for neural machine translation. arXiv preprint arXiv:1906.01130, 2019. 2 +[70] Chuan Wen, Jierui Lin, Trevor Darrell, Dinesh Jayaraman, and Yang Gao. Fighting copycat agents in behavioral cloning from observation histories. In NeurIPS, 2020. 1 +[71] Penghao Wu, Xiaosong Jia, Li Chen, Junchi Yan, Hongyang Li, and Yu Qiao. Trajectory-guided control prediction for end-to-end autonomous driving: A simple yet strong baseline. arXiv preprint arXiv:2206.08129, 2022. 1, 2, 3, 6, 7, 8 +[72] Liuyu Xiang, Guiguang Ding, and Jungong Han. Learning from multiple experts: Self-paced knowledge distillation for long-tailed classification. In ECCV, 2020. 2 + +[73] Canwen Xu, Wangchunshu Zhou, Tao Ge, Furu Wei, and Ming Zhou. Bert-of-theseus: Compressing bert by progressive module replacing. arXiv preprint arXiv:2002.02925, 2020. 2 +[74] Jing Yang, Brais Martinez, Adrian Bulat, and Georgios Tzimiropoulos. Knowledge distillation via adaptive instance normalization. arXiv preprint arXiv:2003.04289, 2020. 2 +[75] Junho Yim, Donggyu Joo, Jihoon Bae, and Junmo Kim. A gift from knowledge distillation: Fast optimization, network minimization and transfer learning. In CVPR, 2017. 2 +[76] Jimuyang Zhang and Eshed Ohn-Bar. Learning by watching. In CVPR, 2021. 2, 3 +[77] Jimuyang Zhang, Ruizhao Zhu, and Eshed Ohn-Bar. SelfD: Self-learning large-scale driving policies from the web. In CVPR, 2022. 2, 7 +[78] Pan Zhang, Bo Zhang, Ting Zhang, Dong Chen, and Fang Wen. Robust mutual learning for semi-supervised semantic segmentation. arXiv preprint arXiv:2106.00609, 2021. 2 +[79] Zhejun Zhang, Alexander Liniger, Dengxin Dai, Fisher Yu, and Luc Van Gool. End-to-end urban driving by imitating a reinforcement learning coach. In ICCV, 2021. 1, 2, 3, 4, 5, 6, 7, 8 +[80] Albert Zhao, Tong He, Yitao Liang, Haibin Huang, Guy Van den Broeck, and Stefano Soatto. Lates: Latent space distillation for teacher-student driving policy learning. arXiv preprint arXiv:1912.02973, 2019. 2 +[81] Albert Zhao, Tong He, Yitao Liang, Haibin Huang, Guy Van den Broeck, and Stefano Soatto. Sam: Squeeze-and-mimic networks for conditional visual driving policy learning. CoRL, 2020. 1, 2, 3 +[82] Brady Zhou and Philipp Krahenbuhl. Cross-view transformers for real-time map-view semantic segmentation. In CVPR, 2022. 2 +[83] Brady Zhou, Philipp Krähenbuhl, and Vladlen Koltun. Does computer vision matter for action? Science Robotics, 4(30), 2019. 2 +[84] Yikai Zhou, Baosong Yang, Derek F Wong, Yu Wan, and Lidia S Chao. Uncertainty-aware curriculum learning for neural machine translation. In ACL, 2020. 2 +[85] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable detr: Deformable transformers for end-to-end object detection. arXiv preprint arXiv:2010.04159, 2020. 5 \ No newline at end of file diff --git a/2023/Coaching a Teachable Student/images.zip b/2023/Coaching a Teachable Student/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..313834adb4b862b338d55ce8fdb8ead9ea2a04db --- /dev/null +++ b/2023/Coaching a Teachable Student/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:22cea570b647a756e9a7c6af32a6cde384f5f317fcc81586c6251a3597bd278b +size 466225 diff --git a/2023/Coaching a Teachable Student/layout.json b/2023/Coaching a Teachable Student/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..2cd4dce42b84f9b991ce9847250437d1e052f3fd --- /dev/null +++ b/2023/Coaching a Teachable Student/layout.json @@ -0,0 +1,8657 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 202, + 103, + 392, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 202, + 103, + 392, + 121 + ], + "spans": [ + { + "bbox": [ + 202, + 103, + 392, + 121 + ], + "type": "text", + "content": "Coaching a Teachable Student" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 164, + 143, + 430, + 172 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 164, + 143, + 430, + 172 + ], + "spans": [ + { + "bbox": [ + 164, + 143, + 430, + 172 + ], + "type": "text", + "content": "Jimuyang Zhang Zanming Huang Eshed Ohn-Bar Boston University" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 198, + 174, + 394, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 198, + 174, + 394, + 185 + ], + "spans": [ + { + "bbox": [ + 198, + 174, + 394, + 185 + ], + "type": "text", + "content": "{zhangjim, huangtom, eohnbar}@bu.edu" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "spans": [ + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 238, + 290, + 539 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 238, + 290, + 539 + ], + "spans": [ + { + "bbox": [ + 46, + 238, + 290, + 539 + ], + "type": "text", + "content": "We propose a novel knowledge distillation framework for effectively teaching a sensorimotor student agent to drive from the supervision of a privileged teacher agent. Current distillation for sensorimotor agents methods tend to result in suboptimal learned driving behavior by the student, which we hypothesize is due to inherent differences between the input, modeling capacity, and optimization processes of the two agents. We develop a novel distillation scheme that can address these limitations and close the gap between the sensorimotor agent and its privileged teacher. Our key insight is to design a student which learns to align their input features with the teacher's privileged Bird's Eye View (BEV) space. The student then can benefit from direct supervision by the teacher over the internal representation learning. To scaffold the difficult sensorimotor learning task, the student model is optimized via a student-paced coaching mechanism with various auxiliary supervision. We further propose a high-capacity imitation learned privileged agent that surpasses prior privileged agents in CARLA and ensures the student learns safe driving behavior. Our proposed sensorimotor agent results in a robust image-based behavior cloning agent in CARLA, improving over current models by over " + }, + { + "bbox": [ + 46, + 238, + 290, + 539 + ], + "type": "inline_equation", + "content": "20.6\\%" + }, + { + "bbox": [ + 46, + 238, + 290, + 539 + ], + "type": "text", + "content": " in driving score without requiring LiDAR, historical observations, ensemble of models, on-policy data aggregation or reinforcement learning." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 549, + 128, + 562 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 549, + 128, + 562 + ], + "spans": [ + { + "bbox": [ + 47, + 549, + 128, + 562 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 569, + 287, + 701 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 569, + 287, + 701 + ], + "spans": [ + { + "bbox": [ + 46, + 569, + 287, + 701 + ], + "type": "text", + "content": "Learning internal representations for making intricate driving decisions from images involves a complex optimization task [4, 43, 54]. The inherent challenge for end-to-end training of driving agents lies in the immense complexity of learning to map high-dimensional visual observations into general and safe navigational decisions [13, 20, 70]. Even given millions of training examples [4], today's agents still fail to reliably learn an internal representation that can be used for robust processing of complex visual scenarios (e.g., dense urban settings with intricate layouts and dynamic obstacles) in a safe and task-driven manner [13, 20, 70]." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 59, + 701, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 701, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 59, + 701, + 287, + 714 + ], + "type": "text", + "content": "To ease the challenging sensorimotor training task, re" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 309, + 213, + 545, + 342 + ], + "blocks": [ + { + "bbox": [ + 309, + 213, + 545, + 342 + ], + "lines": [ + { + "bbox": [ + 309, + 213, + 545, + 342 + ], + "spans": [ + { + "bbox": [ + 309, + 213, + 545, + 342 + ], + "type": "image", + "image_path": "c07ac891b0e814efddc31764fff5c6d21d4fb1cd5b91050d00569bfe4c084176.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 346, + 547, + 456 + ], + "lines": [ + { + "bbox": [ + 305, + 346, + 547, + 456 + ], + "spans": [ + { + "bbox": [ + 305, + 346, + 547, + 456 + ], + "type": "text", + "content": "Figure 1. Effective Knowledge Distillation for Sensorimotor Agents. Our proposed CaT (Coaching a Teachable student) framework enables highly effective knowledge transfer between a privileged teacher and a sensorimotor (i.e., image-based) student. An alignment module learns to transform image-based features to the teacher's BEV feature space, where the student can then leverage extensive and direct supervision on its learned intermediate representations. The student model is optimized via a coaching mechanism with extensive auxiliary supervision in order to further scaffold the difficult sensorimotor learning task." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 471, + 546, + 674 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 471, + 546, + 674 + ], + "spans": [ + { + "bbox": [ + 304, + 471, + 546, + 674 + ], + "type": "text", + "content": "cent approaches decompose the task into stages, e.g., by first training a high-capacity privileged network with complete knowledge of the world and distilling its knowledge into a less capable vision-based student network [11, 13, 23, 47, 79]. However, due to the inherent differences between the inputs and architectures of the two agents, current methods rely on limited supervisory mechanisms from the teacher, i.e., exclusively through the teacher's output [11, 13] or knowledge distillation of a single final fully-connected layer [71, 79, 81]. Moreover, the privileged teacher's demonstration targets may be noisy or difficult for the student to imitate, given the limited perspective [27]. In this work, we sought to develop a more effective knowledge distillation paradigm for training a sensorimotor agent to drive. Our key insight is to enable more extensive supervision from the teacher by reducing the gap between internal modeling and learning capabilities between the two agents." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 677, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 677, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 677, + 547, + 713 + ], + "type": "text", + "content": "Our proposed approach for holistic knowledge distillation is informed by human instruction, which often involves structured supervision in addition to high-level demonstra" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7805" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 288, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 288, + 204 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 288, + 204 + ], + "type": "text", + "content": "tions, e.g., providing various hints to scaffold information in a way that the student can better understand [30]. When teaching others new and challenging skills, i.e., where a student may not be able to replicate the demonstration such as riding a bicycle or driving a vehicle, teachers may provide additional supervision regarding the underlying task structure and their own internal reasoning [64]. Analogously to our settings, the privileged teaching agent can potentially provide richer supervision when teaching a limited capacity sensorimotor student, i.e., through more careful and direct guidance of the underlying representation learning." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 204, + 288, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 204, + 288, + 468 + ], + "spans": [ + { + "bbox": [ + 46, + 204, + 288, + 468 + ], + "type": "text", + "content": "In our work, we introduce CaT, a novel method for teaching a sensorimotor student to drive using supervision from a privileged teacher. Our key insights are threefold: 1) **Effective Teacher:** We propose to incorporate explicit safety-aware cues into the BEV space that facilitate a surprisingly effective teacher agent design. While prior privileged agents struggle to learn to drive in complex urban driving scenes, we demonstrate our learned agent to match expert-level decision-making. 2) **Teachable Student via Alignment:** An IPM-based transformer alignment module can facilitate direct distillation of most of the teacher's features and better guide the student learning process. 3) **Student-paced Coaching:** A coaching mechanism for managing difficult samples can scaffold knowledge and lead to improved model optimization by better considering the ability of the student. By holistically tackling the complex knowledge distillation task with extensive teacher and auxiliary supervision, we are able to train a state-of-the-art image-based agent in CARLA [21]. Through ablation studies, input design, and interpretable feature analysis, we also provide critical insights into current limitations in learning robust and generalized representations for driving agents." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 476, + 134, + 489 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 476, + 134, + 489 + ], + "spans": [ + { + "bbox": [ + 47, + 476, + 134, + 489 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 498, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 498, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 498, + 287, + 713 + ], + "type": "text", + "content": "Knowledge Distillation: We study knowledge distillation methods originally developed for model compression and acceleration [12,17,29,31,35,36,52,55,74,75] in the context of training sensorimotor. While approaches for feature distillation have been minimally explored in this context, we are motivated by their success in other domains, including image classification [2,38,40,42,60], object detection [67], semantic segmentation [46, 78], and natural language processing [3,68,73]. Yet, applying such techniques is not trivial given the drastically differing inputs between the privileged agent and student and the overwhelming sensorimotor task. Consequently, driving policy distillation methods only provide supervision either from the teacher's output [11,13] or the features of a single fully-connected layer [71,79,81], which (we hypothesize) does not provide sufficient hints to guide the student training. In this work, we provide novel mechanisms to close this gap and enable extensive supervision through deep feature distillation." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 304, + 72, + 547, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 300 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 300 + ], + "type": "text", + "content": "Imitation Learning to Drive: Recent approaches in imitation learning (IL) to drive can be traced to Pomerleau [54]. Recently, more elaborate IL-based approaches for driving have emerged [9, 10, 25, 45, 50, 76, 77, 80, 83]. Specifically, decomposing the imitation learning task into two stages, i.e., by first learning a privileged agent through behavior cloning [11, 13] or reinforcement learning [8, 62, 71, 79] and then training the sensorimotor agent to mimic the output of the privileged agent. Our study is motivated by such approaches, yet we explore the benefits of increased supervision by the privileged agent. Moreover, prior privileged agents produce noisy and sub-optimal demonstration supervision in complex urban scenarios (even when optimized with RL, as demonstrated by our analysis). While most current studies employ the RL-based teacher of [79], this can be problematic and inefficient in safety-critical conditions in the real-world. In our work, we introduce a novel privileged teacher which can greatly surpass prior agents while only relying on offline behavior cloning." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 308, + 547, + 490 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 308, + 547, + 490 + ], + "spans": [ + { + "bbox": [ + 304, + 308, + 547, + 490 + ], + "type": "text", + "content": "Intermediate Representation for Driving: Learning an effective 3D scene representation is crucial for safe autonomous driving. Researchers may obtain such representations by lifting the image to 3D using estimated depth and projecting the frustums into a BEV grid [32, 53]. Alternatively, transformer-based architectures can also enable mapping a camera image to the BEV space [14, 82], i.e., by attending to image-based information when populating a BEV-sampled grid. Related to our work is the study of Chen et al. [11], which learns a BEV representation from RGB image and LiDAR input. A motion planner is then used to generate future waypoints from the BEV. As this introduces significant challenges, we propose to leverage an IPM-based alignment module which can better structure to image-to-BEV projection task." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 498, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 498, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 498, + 547, + 713 + ], + "type": "text", + "content": "Curriculum and Self-paced Learning: Approaches for curriculum learning methods often structure learning from various difficulty samples [5,24,37,39,44,59,61,63,69,84]. To better consider the challenging sensorimotor task and limited capacity of the student, we adopt a curriculum (i.e., an imitation coach), with progressively more challenging samples. He et al. [27] also proposes training a coach across iterations of on-policy data collection [56]. In contrast, we implement and demonstrate the benefit of a coach over training iterations without a data aggregation part. This can facilitate more efficient model training. The approach is motivated by self-paced learning [41, 72], where the model selects easy samples dynamically at each iteration based on a defined loss. In contrast to such studies, we only smooth the targets of the hard samples (based on the student), instead of discarding them. We empirically find this to result in improved coaching, potentially due to improved scaffolding of the difficult samples." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "7806" + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 71, + 103, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 71, + 103, + 83 + ], + "spans": [ + { + "bbox": [ + 47, + 71, + 103, + 83 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 91, + 289, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 91, + 289, + 224 + ], + "spans": [ + { + "bbox": [ + 46, + 91, + 289, + 224 + ], + "type": "text", + "content": "Our goal is to decompose the challenging sensorimotor learning task and ease model optimization through effective supervision from a teacher agent. We first formulate the problem of privileged agent distillation in Sec. 3.1. Next, we address current limitations in privileged agent design to train a robust imitation learned teacher in Sec. 3.2. We then propose a transformer-based student architecture to resolve image-to-BEV feature alignment (Sec. 3.3). Finally, we train the sensorimotor student via extensive supervision and knowledge distillation (Sec. 3.4) via a progressive (i.e., coaching) mechanism (Sec. 3.5)." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 230, + 129, + 242 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 230, + 129, + 242 + ], + "spans": [ + { + "bbox": [ + 47, + 230, + 129, + 242 + ], + "type": "text", + "content": "3.1. Formulation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 248, + 289, + 429 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 248, + 289, + 429 + ], + "spans": [ + { + "bbox": [ + 46, + 248, + 289, + 429 + ], + "type": "text", + "content": "Driving can be formulated as a sequential decision-making problem over a set of sensory observations " + }, + { + "bbox": [ + 46, + 248, + 289, + 429 + ], + "type": "inline_equation", + "content": "\\mathcal{X}^s" + }, + { + "bbox": [ + 46, + 248, + 289, + 429 + ], + "type": "text", + "content": ", a set of actions (motor commands) " + }, + { + "bbox": [ + 46, + 248, + 289, + 429 + ], + "type": "inline_equation", + "content": "\\mathcal{V}" + }, + { + "bbox": [ + 46, + 248, + 289, + 429 + ], + "type": "text", + "content": ", and a policy function " + }, + { + "bbox": [ + 46, + 248, + 289, + 429 + ], + "type": "inline_equation", + "content": "f_{\\theta}^{s} \\colon \\mathcal{X}^{s} \\to \\mathcal{V}" + }, + { + "bbox": [ + 46, + 248, + 289, + 429 + ], + "type": "text", + "content": " for mapping observations to actions at each time step [27,50,56]. We consider the task of learning a sensorimotor student agent via knowledge distillation [13, 23] from a teacher agent " + }, + { + "bbox": [ + 46, + 248, + 289, + 429 + ], + "type": "inline_equation", + "content": "f_{\\psi}^{t} \\colon \\mathcal{X}^{t} \\to \\mathcal{V}" + }, + { + "bbox": [ + 46, + 248, + 289, + 429 + ], + "type": "text", + "content": " with privileged access to environmental information, e.g., ground-truth lanes, obstacles, and traffic light states. We parameterize the two agents as neural networks using weights " + }, + { + "bbox": [ + 46, + 248, + 289, + 429 + ], + "type": "inline_equation", + "content": "[\\pmb{\\theta}, \\pmb{\\psi}] \\in \\mathbb{R}^d" + }, + { + "bbox": [ + 46, + 248, + 289, + 429 + ], + "type": "text", + "content": ", and denote " + }, + { + "bbox": [ + 46, + 248, + 289, + 429 + ], + "type": "inline_equation", + "content": "\\mathcal{F}^s(\\cdot; \\pmb{\\theta})" + }, + { + "bbox": [ + 46, + 248, + 289, + 429 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 248, + 289, + 429 + ], + "type": "inline_equation", + "content": "\\mathcal{F}^t(\\cdot; \\pmb{\\psi})" + }, + { + "bbox": [ + 46, + 248, + 289, + 429 + ], + "type": "text", + "content": " to be the feature maps for the sensorimotor student network and teacher network, respectively. Given a dataset " + }, + { + "bbox": [ + 46, + 248, + 289, + 429 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 46, + 248, + 289, + 429 + ], + "type": "text", + "content": " comprising sensory and privileged observations and a loss function " + }, + { + "bbox": [ + 46, + 248, + 289, + 429 + ], + "type": "inline_equation", + "content": "\\mathcal{L}" + }, + { + "bbox": [ + 46, + 248, + 289, + 429 + ], + "type": "text", + "content": ", the student can be optimized from the teacher using" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 435, + 287, + 455 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 435, + 287, + 455 + ], + "spans": [ + { + "bbox": [ + 61, + 435, + 287, + 455 + ], + "type": "interline_equation", + "content": "\\underset {\\boldsymbol {\\theta}} {\\operatorname {a r g m i n}} \\mathbb {E} _ {(\\mathbf {x} ^ {s}, \\mathbf {x} ^ {t}) \\sim \\mathcal {D}} [ \\mathcal {L} (\\mathcal {F} ^ {s} (\\mathbf {x} ^ {s}; \\boldsymbol {\\theta}), \\mathcal {F} ^ {t} (\\mathbf {x} ^ {t}; \\boldsymbol {\\psi})) ] \\tag {1}", + "image_path": "fe166d54731d4166a68ff7d53cb14d1c372bfda1d04321153d8a478f66c700d5.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 462, + 289, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 462, + 289, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 462, + 289, + 714 + ], + "type": "text", + "content": "where the loss may be defined over a final layer, i.e., to match the output of the teacher [11, 13], or over internal layers as well, i.e., feature distillation [71, 79, 81]. Given the challenging end-to-end sensorimotor learning task, the role of the teacher is to provide effective supervision to the student, i.e., informative targets in Eqn. 1. However, as the two agents operate from drastically differing inputs, the knowledge transfer from the teacher to the student can become ineffective. Specifically, the task of the student is significantly more challenging than that of the teacher, i.e., due to inherent noise and uncertainty. Indeed, there is currently a substantial gap between the performance of current sensorimotor learning agents and their teaching agents, even assuming access to expensive and high-quality 3D observations such as LiDAR [11, 13, 16] (not used in our work). The teaching process should also ideally consider any limited capacity of the student to facilitate effective optimization and learning [27]. Moreover, learning an optimal and robust teacher in itself can be a challenging task [13], even assuming the simplified learning task such as solved perception, as we demonstrate in Sec. 4 and addressed in Sec. 3.2." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 72, + 547, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 443 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 443 + ], + "type": "text", + "content": "Problem Setup and Agent Observations: We develop our CaT framework in the context of the CARLA simulator [21]. Our objective is to train a goal-conditional sensorimotor agent [16, 18] for mapping image and goal observations to vehicle throttle and steering control [13, 48]. We follow standard conditional imitation learning [11, 16] and assume access to three (non-overlapping) RGB camera views " + }, + { + "bbox": [ + 304, + 72, + 547, + 443 + ], + "type": "inline_equation", + "content": "\\mathbf{I} = [\\mathbf{I}_0, \\mathbf{I}_1, \\mathbf{I}_2] \\in \\mathbb{R}^{W \\times H \\times 3}" + }, + { + "bbox": [ + 304, + 72, + 547, + 443 + ], + "type": "text", + "content": ", a categorical navigational command " + }, + { + "bbox": [ + 304, + 72, + 547, + 443 + ], + "type": "inline_equation", + "content": "c \\in \\{1, \\dots, 6\\}" + }, + { + "bbox": [ + 304, + 72, + 547, + 443 + ], + "type": "text", + "content": " (i.e., turn left, turn right, follow, forward, left lane changing and right lane changing) and an intermediate noisy goal " + }, + { + "bbox": [ + 304, + 72, + 547, + 443 + ], + "type": "inline_equation", + "content": "\\mathbf{g} \\in \\mathbb{R}^2" + }, + { + "bbox": [ + 304, + 72, + 547, + 443 + ], + "type": "text", + "content": " sampled from a GNSS (Global Navigation Satellite System) [11]. We note that both the command and GNSS observations can be easily obtained by today's vehicles, i.e., through generation of an " + }, + { + "bbox": [ + 304, + 72, + 547, + 443 + ], + "type": "inline_equation", + "content": "\\mathrm{A}^*" + }, + { + "bbox": [ + 304, + 72, + 547, + 443 + ], + "type": "text", + "content": " plan [26] from a static map and coarse positioning system. The GNSS goals are sampled every 50-100 meters and reflect real-world measurement errors [11]. We compute a BEV " + }, + { + "bbox": [ + 304, + 72, + 547, + 443 + ], + "type": "inline_equation", + "content": "\\mathbf{B} \\in \\{0, 1\\}^{W_B \\times H_B \\times C_B}" + }, + { + "bbox": [ + 304, + 72, + 547, + 443 + ], + "type": "text", + "content": " by rendering privileged (i.e., ground-truth) information from the underlying simulation, including 3D location and state of lanes, pedestrians, vehicles, and traffic lights [79]. While prior work has generally leveraged a standardized BEV representation [13, 32, 34, 57, 76], we find its design to be crucial to training an optimal teacher policy as will be discussed in Sec 3.2. We let " + }, + { + "bbox": [ + 304, + 72, + 547, + 443 + ], + "type": "inline_equation", + "content": "\\mathbf{x}^s = (\\mathbf{I}, \\mathbf{g}, c) \\in \\mathcal{X}^s" + }, + { + "bbox": [ + 304, + 72, + 547, + 443 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 72, + 547, + 443 + ], + "type": "inline_equation", + "content": "\\mathbf{x}^t = (\\mathbf{B}, \\mathbf{g}, c) \\in \\mathcal{X}^t" + }, + { + "bbox": [ + 304, + 72, + 547, + 443 + ], + "type": "text", + "content": " be the student and teacher observations, respectively. Given these observations at each time step, the agent learns to predict 10 future 2D waypoints in top-down vehicle coordinates for the next 2.5 seconds of driving. These are then given to a lateral and longitudinal PID controllers [13, 48, 66] to generate the final action." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 444, + 547, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 444, + 547, + 515 + ], + "spans": [ + { + "bbox": [ + 304, + 444, + 547, + 515 + ], + "type": "text", + "content": "Given this formulation, we next discuss our teacher agent design and training process. In particular, we aim to learn a privileged agent which can not only effectively solve the goal-oriented navigation task, but also facilitate effective distillation and coaching of a sensorimotor student policy, as will be discussed in Sec. 3.4." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 526, + 471, + 539 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 526, + 471, + 539 + ], + "spans": [ + { + "bbox": [ + 306, + 526, + 471, + 539 + ], + "type": "text", + "content": "3.2. Learning an Effective Teacher" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 545, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 545, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 545, + 547, + 713 + ], + "type": "text", + "content": "In our formulation, learning a student agent begins with training an effective privileged agent " + }, + { + "bbox": [ + 304, + 545, + 547, + 713 + ], + "type": "inline_equation", + "content": "f_{\\psi}^{t}" + }, + { + "bbox": [ + 304, + 545, + 547, + 713 + ], + "type": "text", + "content": ". However, we find this itself to be a complex task. Even with complete knowledge of perfect BEV perception, current privileged agents are suboptimal, significantly under-performing CARLA's built-in autopilot [13, 79]. In this section, we uncover the underlying reason for this under-performance through the BEV design. Informed by this key finding, we train a highly effective imitation-learned teacher that surpasses prior privileged agents and matches expert-level performance in CARLA without requiring extensive data collection [56] or reinforcement learning [79]. Our proposed privileged agent can then enable effective supervision of the student in Sec. 3.3." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7807" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 71, + 547, + 247 + ], + "blocks": [ + { + "bbox": [ + 50, + 71, + 547, + 247 + ], + "lines": [ + { + "bbox": [ + 50, + 71, + 547, + 247 + ], + "spans": [ + { + "bbox": [ + 50, + 71, + 547, + 247 + ], + "type": "image", + "image_path": "5c949646a75d96ca06d760734237b42d48f99536cc947e87bd81106f55f5696c.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 256, + 548, + 323 + ], + "lines": [ + { + "bbox": [ + 46, + 256, + 548, + 323 + ], + "spans": [ + { + "bbox": [ + 46, + 256, + 548, + 323 + ], + "type": "text", + "content": "Figure 2. Overview of Architecture and Distillation Process of CaT. We first train a high-capacity privileged teacher agent (colored in red) which takes an augmented BEV representation as input, and produces a set of future waypoints conditioned on command and short-term goal. Our teachable student agent (colored in blue) maps the image features to the BEV space using an alignment module, which is trained by distillation of intermediate features (arrows colored in blue) and final output (arrow colored in green). Additionally, we leverage supervision from auxiliary segmentation (arrow colored in red) and command prediction tasks (arrow colored in purple) interlaced throughout the network to facilitate the learning of task-relevant features." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 343, + 289, + 620 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 343, + 289, + 620 + ], + "spans": [ + { + "bbox": [ + 46, + 343, + 289, + 620 + ], + "type": "text", + "content": "Teacher Training via Direct Expert Imitation: In imitation learning, an oracle (i.e., expert driver) " + }, + { + "bbox": [ + 46, + 343, + 289, + 620 + ], + "type": "inline_equation", + "content": "f^{*} \\colon \\mathcal{X}^{t} \\to \\mathcal{Y}" + }, + { + "bbox": [ + 46, + 343, + 289, + 620 + ], + "type": "text", + "content": " is defined for demonstrating optimal actions " + }, + { + "bbox": [ + 46, + 343, + 289, + 620 + ], + "type": "inline_equation", + "content": "\\mathbf{y}^{*}" + }, + { + "bbox": [ + 46, + 343, + 289, + 620 + ], + "type": "text", + "content": ". In CARLA [21], the expert is defined using longitudinal and lateral PID controllers which are carefully tuned and augmented by a collection of manually specified rules for handling diverse scenarios (e.g., sudden braking) over the underlying simulation state. The designed expert can then be used to generate trajectories as supervised data for learning the privileged policy function " + }, + { + "bbox": [ + 46, + 343, + 289, + 620 + ], + "type": "inline_equation", + "content": "f_{\\psi}^{t}" + }, + { + "bbox": [ + 46, + 343, + 289, + 620 + ], + "type": "text", + "content": " with behavior cloning [18, 19, 48]. Our privileged agent is a ResNet-18 model [28], as shown in Fig. 2 (further increasing modeling capacity was not found to be beneficial). The ResNet model is followed by a GRU-based conditional waypoint predictor [11, 16]. Due to the complexity of our dense urban navigation task, e.g., with multi-lane roads, intersections, and merging, we incrementally refine the waypoints using two GRUs within each conditional branch. The first GRU regresses the preliminary set of waypoints via sequential waypoints directly from the embedded features. Subsequently, the second GRU takes the predicted waypoints together with the short-term goal and embedded features, to produce a refined set of waypoint targets." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 629, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 629, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 629, + 287, + 715 + ], + "type": "text", + "content": "Even assuming perfect perception, we find it difficult to train privileged agents that exhibit robust planning behavior in complex urban driving scenarios. Specifically, when using the most challenging CARLA benchmark (Longest6 [16], also detailed in Sec 4) using the standard BEV representation of Chen et al. [13] leads to " + }, + { + "bbox": [ + 46, + 629, + 287, + 715 + ], + "type": "inline_equation", + "content": "26\\%" + }, + { + "bbox": [ + 46, + 629, + 287, + 715 + ], + "type": "text", + "content": " driving score compared to " + }, + { + "bbox": [ + 46, + 629, + 287, + 715 + ], + "type": "inline_equation", + "content": "72\\%" + }, + { + "bbox": [ + 46, + 629, + 287, + 715 + ], + "type": "text", + "content": " by the rule-based expert. The rein" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 304, + 343, + 547, + 489 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 343, + 547, + 489 + ], + "spans": [ + { + "bbox": [ + 304, + 343, + 547, + 489 + ], + "type": "text", + "content": "forcement learning agent of Zhang et al. [79] achieves " + }, + { + "bbox": [ + 304, + 343, + 547, + 489 + ], + "type": "inline_equation", + "content": "60\\%" + }, + { + "bbox": [ + 304, + 343, + 547, + 489 + ], + "type": "text", + "content": " on this task. We sought to explore the limits of behavior cloning for this task, as a noisy and sub-optimal teacher can hinder the training of the student model. In particular, we hypothesize that the reason for the poor performance lies in the increased complexity of the BEV state representation in dense scenarios. In such scenarios, learning to extract task-relevant details can become more challenging. To address this issue, we propose to introduce additional channels (akin to affordances [9,22,58]) that can more easily translate into safety-critical decision-making, simplify the learning task, and provide an expert-level teaching agent." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 498, + 548, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 498, + 548, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 498, + 548, + 715 + ], + "type": "text", + "content": "BEV with Safety Hints: Our BEV encodes state information into distinct channels, including drivable road regions, the desired route based on the command and layout, lane markings, and dynamic obstacles (see Fig. 3). Given the low performance of prior privileged agents, in particular around dynamic objects, the agent could benefit from more explicit safety-driven cues in the BEV. We propose to add two types of channels of (1) predicted agents' future and (2) entity attention. First, we utilize a kinematics bicycle model [16] in order to efficiently predict future trajectories of dynamic objects. This enables us to iteratively predict and represent short-term future position, orientation, and speed of agents (our supplementary contains additional information regarding this process). Secondly, we encode an explicit attention channel for highlighting potential future infractions, i.e., based on the kinematics bicycle model applied to the ego-vehicle. While this assumes that the current vehicle throttle and steering remain unchanged, the entity" + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7808" + } + ] + } + ], + "index": 6 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 84, + 143, + 178 + ], + "blocks": [ + { + "bbox": [ + 50, + 84, + 143, + 178 + ], + "lines": [ + { + "bbox": [ + 50, + 84, + 143, + 178 + ], + "spans": [ + { + "bbox": [ + 50, + 84, + 143, + 178 + ], + "type": "image", + "image_path": "73eeaae3b8b28898a1237ee7e1faff2380488617b79ab9d151c46b90ce0bf569.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 63, + 194, + 132, + 206 + ], + "lines": [ + { + "bbox": [ + 63, + 194, + 132, + 206 + ], + "spans": [ + { + "bbox": [ + 63, + 194, + 132, + 206 + ], + "type": "text", + "content": "(a) Baseline [79]" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 144, + 85, + 237, + 178 + ], + "blocks": [ + { + "bbox": [ + 144, + 85, + 237, + 178 + ], + "lines": [ + { + "bbox": [ + 144, + 85, + 237, + 178 + ], + "spans": [ + { + "bbox": [ + 144, + 85, + 237, + 178 + ], + "type": "image", + "image_path": "fc806bd673925eeb5ae8eb6eb6e21ed36a083ed94b230d75d84906c3936f1bac.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 244, + 87, + 297, + 123 + ], + "blocks": [ + { + "bbox": [ + 244, + 87, + 297, + 123 + ], + "lines": [ + { + "bbox": [ + 244, + 87, + 297, + 123 + ], + "spans": [ + { + "bbox": [ + 244, + 87, + 297, + 123 + ], + "type": "image", + "image_path": "fd79ed0d0f73eb8661bc18da399b93b74e16fcf3bb024c507ad7c3d9b8878c47.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 244, + 137, + 297, + 190 + ], + "blocks": [ + { + "bbox": [ + 148, + 194, + 235, + 205 + ], + "lines": [ + { + "bbox": [ + 148, + 194, + 235, + 205 + ], + "spans": [ + { + "bbox": [ + 148, + 194, + 235, + 205 + ], + "type": "text", + "content": "(b) With Safety Hints" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 252, + 126, + 289, + 137 + ], + "lines": [ + { + "bbox": [ + 252, + 126, + 289, + 137 + ], + "spans": [ + { + "bbox": [ + 252, + 126, + 289, + 137 + ], + "type": "text", + "content": "(c) Road" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 244, + 137, + 297, + 190 + ], + "lines": [ + { + "bbox": [ + 244, + 137, + 297, + 190 + ], + "spans": [ + { + "bbox": [ + 244, + 137, + 297, + 190 + ], + "type": "image", + "image_path": "cd4190bd9b1aaec56b3c88657eececf62a2e7490db02bd9413bb5a621e807bf5.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 240, + 194, + 301, + 206 + ], + "lines": [ + { + "bbox": [ + 240, + 194, + 301, + 206 + ], + "spans": [ + { + "bbox": [ + 240, + 194, + 301, + 206 + ], + "type": "text", + "content": "(g) Pedestrians" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 211, + 546, + 266 + ], + "lines": [ + { + "bbox": [ + 46, + 211, + 546, + 266 + ], + "spans": [ + { + "bbox": [ + 46, + 211, + 546, + 266 + ], + "type": "text", + "content": "Figure 3. Visualization of Our Proposed BEV with Safety Hints. (a) and (b) visualize the baseline and our proposed BEV by encoding different channels into colors. Roads are shown in dark grey, Desired Route is shown in light grey, Lane Marks are shown in magenta, Vehicles are shown in blue with past trajectories shown in decreasing saturation with time into the past, Pedestrians are shown in light blue, Traffic Lights are shown in red or green according to its state, Agent Forecasts are shown in purple with decreasing saturation with time into the future, Entity Attention is shown in orange. (c) - (j) shows the different channels separately." + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 313, + 71, + 367, + 124 + ], + "blocks": [ + { + "bbox": [ + 313, + 71, + 367, + 124 + ], + "lines": [ + { + "bbox": [ + 313, + 71, + 367, + 124 + ], + "spans": [ + { + "bbox": [ + 313, + 71, + 367, + 124 + ], + "type": "image", + "image_path": "b4024c7fd22bea493e03de456ef1978ad97234ae0eae693bc839cc22b7dc9ec7.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 313, + 137, + 365, + 190 + ], + "blocks": [ + { + "bbox": [ + 304, + 126, + 376, + 137 + ], + "lines": [ + { + "bbox": [ + 304, + 126, + 376, + 137 + ], + "spans": [ + { + "bbox": [ + 304, + 126, + 376, + 137 + ], + "type": "text", + "content": "(d) Desired Route" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 137, + 365, + 190 + ], + "lines": [ + { + "bbox": [ + 313, + 137, + 365, + 190 + ], + "spans": [ + { + "bbox": [ + 313, + 137, + 365, + 190 + ], + "type": "image", + "image_path": "bbc8e8a2da2d118bb459310c0bd2db244beb0838983a1409943cfb677f16caf6.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 194, + 375, + 206 + ], + "lines": [ + { + "bbox": [ + 304, + 194, + 375, + 206 + ], + "spans": [ + { + "bbox": [ + 304, + 194, + 375, + 206 + ], + "type": "text", + "content": "(h) Traffic Lights" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 392, + 71, + 445, + 123 + ], + "blocks": [ + { + "bbox": [ + 392, + 71, + 445, + 123 + ], + "lines": [ + { + "bbox": [ + 392, + 71, + 445, + 123 + ], + "spans": [ + { + "bbox": [ + 392, + 71, + 445, + 123 + ], + "type": "image", + "image_path": "36f96d7fddefc58add70214537bc7032084d60e96b581ce078d55d2d5a9c8867.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 392, + 137, + 445, + 190 + ], + "blocks": [ + { + "bbox": [ + 388, + 126, + 449, + 137 + ], + "lines": [ + { + "bbox": [ + 388, + 126, + 449, + 137 + ], + "spans": [ + { + "bbox": [ + 388, + 126, + 449, + 137 + ], + "type": "text", + "content": "(e) Lane Marks" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 392, + 137, + 445, + 190 + ], + "lines": [ + { + "bbox": [ + 392, + 137, + 445, + 190 + ], + "spans": [ + { + "bbox": [ + 392, + 137, + 445, + 190 + ], + "type": "image", + "image_path": "02093be44a605a22045abe59041659a6bae0e47327c3c8ca5ca8c9f9b6a5ca10.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 378, + 194, + 456, + 206 + ], + "lines": [ + { + "bbox": [ + 378, + 194, + 456, + 206 + ], + "spans": [ + { + "bbox": [ + 378, + 194, + 456, + 206 + ], + "type": "text", + "content": "(i) Agent Forecast" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 476, + 72, + 529, + 123 + ], + "blocks": [ + { + "bbox": [ + 476, + 72, + 529, + 123 + ], + "lines": [ + { + "bbox": [ + 476, + 72, + 529, + 123 + ], + "spans": [ + { + "bbox": [ + 476, + 72, + 529, + 123 + ], + "type": "image", + "image_path": "0389cadfe8fd59778bcc6fc88d06cf550ebdb972cd0a5ba6593af2cb91120a97.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 476, + 137, + 529, + 190 + ], + "blocks": [ + { + "bbox": [ + 479, + 125, + 526, + 137 + ], + "lines": [ + { + "bbox": [ + 479, + 125, + 526, + 137 + ], + "spans": [ + { + "bbox": [ + 479, + 125, + 526, + 137 + ], + "type": "text", + "content": "(f) Vehicles" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 476, + 137, + 529, + 190 + ], + "lines": [ + { + "bbox": [ + 476, + 137, + 529, + 190 + ], + "spans": [ + { + "bbox": [ + 476, + 137, + 529, + 190 + ], + "type": "image", + "image_path": "e2ac63a43c8db4a86b4ff9dabc0741bf9773f5b501f128474f7a30ef018291a3.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 460, + 194, + 472, + 206 + ], + "lines": [ + { + "bbox": [ + 460, + 194, + 472, + 206 + ], + "spans": [ + { + "bbox": [ + 460, + 194, + 472, + 206 + ], + "type": "text", + "content": "(j)" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 476, + 194, + 542, + 206 + ], + "lines": [ + { + "bbox": [ + 476, + 194, + 542, + 206 + ], + "spans": [ + { + "bbox": [ + 476, + 194, + 542, + 206 + ], + "type": "text", + "content": "Entity Attention" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "bbox": [ + 46, + 277, + 287, + 371 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 277, + 287, + 371 + ], + "spans": [ + { + "bbox": [ + 46, + 277, + 287, + 371 + ], + "type": "text", + "content": "attention channel helps the agent better reason over safety while correlating states with actions. Surprisingly, the additions of the safety hints channels provide a strong teacher with offline behavior cloning alone, even improving over the rule-based expert (Table 1, " + }, + { + "bbox": [ + 46, + 277, + 287, + 371 + ], + "type": "inline_equation", + "content": "73\\%" + }, + { + "bbox": [ + 46, + 277, + 287, + 371 + ], + "type": "text", + "content": " vs. " + }, + { + "bbox": [ + 46, + 277, + 287, + 371 + ], + "type": "inline_equation", + "content": "72\\%" + }, + { + "bbox": [ + 46, + 277, + 287, + 371 + ], + "type": "text", + "content": "). We note that many of the remaining failures are due to simulation timeout in crowded intersections. Our boosted teacher agent will be critical for effective distillation in Sec. 3.4." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 47, + 393, + 209, + 405 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 393, + 209, + 405 + ], + "spans": [ + { + "bbox": [ + 47, + 393, + 209, + 405 + ], + "type": "text", + "content": "3.3. Learning a Teachable Student" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 46, + 416, + 287, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 416, + 287, + 548 + ], + "spans": [ + { + "bbox": [ + 46, + 416, + 287, + 548 + ], + "type": "text", + "content": "Due to differing inputs and modeling capacities, it can be difficult to align the image-based student features and output with the BEV-based privileged teacher. We propose to alleviate misalignment by designing a teachable student network, i.e., a model that matches as much of the teacher architecture as possible. Specifically, we propose to improve teacher-student model alignment by leveraging an internal image-to-BEV alignment module. As shown in Fig. 2, the three residual blocks following the alignment module can consequently facilitate knowledge transfer via direct distillation of most of the teacher's features in Sec. 3.4." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 46, + 557, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 557, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 557, + 287, + 713 + ], + "type": "text", + "content": "IPM-based Transformer Alignment Module: Explicitly mapping images to BEV space for downstream tasks can be challenging. While this can be achieved via depth estimation and lifting [32,53], we sought to incorporate an efficient and differentiable module which can effectively attend and reason over various features while mapping from front view to BEV. A transformer-based architecture [14,65] provides a natural choice as it can be optimized end-to-end with respect to the driving task while also carefully aligning features among spaces with arbitrary dimensionalities. As shown in Fig. 2, we sample queries " + }, + { + "bbox": [ + 46, + 557, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}_{init} \\in \\mathbb{R}^{L_H \\times L_W \\times L_C}" + }, + { + "bbox": [ + 46, + 557, + 287, + 713 + ], + "type": "text", + "content": " using a spatial parameterization of the BEV space and process them using a self-attention module [14,65]. The output" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 306, + 277, + 500, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 277, + 500, + 289 + ], + "spans": [ + { + "bbox": [ + 306, + 277, + 500, + 289 + ], + "type": "text", + "content": "of self-attention module " + }, + { + "bbox": [ + 306, + 277, + 500, + 289 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 306, + 277, + 500, + 289 + ], + "type": "text", + "content": " can be formulated as" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 352, + 292, + 545, + 319 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 352, + 292, + 545, + 319 + ], + "spans": [ + { + "bbox": [ + 352, + 292, + 545, + 319 + ], + "type": "interline_equation", + "content": "\\mathbf {Q} = \\operatorname {s o f t m a x} \\left(\\frac {\\mathbf {Q} _ {\\text {i n i t}} \\mathbf {K} _ {\\text {i n i t}} ^ {T}}{\\sqrt {d}}\\right) \\mathbf {V} _ {\\text {i n i t}} \\tag {2}", + "image_path": "a56c9525cd2f59c445087afef7512b77b7cf65a30bf0f9bc5c4e9eb8c48f66d8.jpg" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 304, + 323, + 545, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 323, + 545, + 396 + ], + "spans": [ + { + "bbox": [ + 304, + 323, + 545, + 396 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 323, + 545, + 396 + ], + "type": "inline_equation", + "content": "\\mathbf{K}_{init}" + }, + { + "bbox": [ + 304, + 323, + 545, + 396 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 323, + 545, + 396 + ], + "type": "inline_equation", + "content": "\\mathbf{V}_{init} \\in \\mathbb{R}^{L_H \\times L_W \\times L_C}" + }, + { + "bbox": [ + 304, + 323, + 545, + 396 + ], + "type": "text", + "content": " provide the key-value pairs and " + }, + { + "bbox": [ + 304, + 323, + 545, + 396 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 304, + 323, + 545, + 396 + ], + "type": "text", + "content": " is the dimension of the query. Subsequently, each query point " + }, + { + "bbox": [ + 304, + 323, + 545, + 396 + ], + "type": "inline_equation", + "content": "\\mathbf{q} \\in \\mathbf{Q}" + }, + { + "bbox": [ + 304, + 323, + 545, + 396 + ], + "type": "text", + "content": " is mapped to its corresponding reference point " + }, + { + "bbox": [ + 304, + 323, + 545, + 396 + ], + "type": "inline_equation", + "content": "\\mathbf{p} \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 304, + 323, + 545, + 396 + ], + "type": "text", + "content": " (homogeneous coordinates) in the image features " + }, + { + "bbox": [ + 304, + 323, + 545, + 396 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{RGB} \\in \\mathbb{R}^{Z_H \\times Z_W \\times Z_C}" + }, + { + "bbox": [ + 304, + 323, + 545, + 396 + ], + "type": "text", + "content": " using an Inverse Perspective Mapping (IPM)" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 380, + 398, + 544, + 410 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 380, + 398, + 544, + 410 + ], + "spans": [ + { + "bbox": [ + 380, + 398, + 544, + 410 + ], + "type": "interline_equation", + "content": "\\mathbf {p} = s \\mathbf {P} _ {k} \\mathbf {R} _ {k} (\\mathbf {q} - \\mathbf {t} _ {k}) \\tag {3}", + "image_path": "61d739706cd8a8770fc7c59e5c5fe8cb32b170fbbfee50ddf083964a71f69946.jpg" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 304, + 411, + 545, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 411, + 545, + 472 + ], + "spans": [ + { + "bbox": [ + 304, + 411, + 545, + 472 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 411, + 545, + 472 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 304, + 411, + 545, + 472 + ], + "type": "text", + "content": " is a scale factor, " + }, + { + "bbox": [ + 304, + 411, + 545, + 472 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_k\\in \\mathbb{R}^{3\\times 3}" + }, + { + "bbox": [ + 304, + 411, + 545, + 472 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 411, + 545, + 472 + ], + "type": "inline_equation", + "content": "\\mathbf{R}_k\\in \\mathbb{R}^{3\\times 3}" + }, + { + "bbox": [ + 304, + 411, + 545, + 472 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 411, + 545, + 472 + ], + "type": "inline_equation", + "content": "\\mathbf{t}_k\\in \\mathbb{R}^3" + }, + { + "bbox": [ + 304, + 411, + 545, + 472 + ], + "type": "text", + "content": " are the " + }, + { + "bbox": [ + 304, + 411, + 545, + 472 + ], + "type": "inline_equation", + "content": "k^{th}" + }, + { + "bbox": [ + 304, + 411, + 545, + 472 + ], + "type": "text", + "content": " camera intrinsics, rotation and translation relative to the center of the ego-vehicle, respectively [6]. Note that we employ a different IPM mapping for each camera view to perform the projection." + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 304, + 472, + 545, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 472, + 545, + 590 + ], + "spans": [ + { + "bbox": [ + 304, + 472, + 545, + 590 + ], + "type": "text", + "content": "To populate the student's BEV features " + }, + { + "bbox": [ + 304, + 472, + 545, + 590 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{BEV} \\in \\mathbb{R}^{H_B \\times W_B \\times C_B}" + }, + { + "bbox": [ + 304, + 472, + 545, + 590 + ], + "type": "text", + "content": " (i.e., the output of the alignment module in Fig. 2), we leverage a deformable cross-attention mechanism based on deformable DETR [85]. This enables the network to attend to multiple regions around the reference point in the image features when transforming to BEV feature space. The reference point " + }, + { + "bbox": [ + 304, + 472, + 545, + 590 + ], + "type": "inline_equation", + "content": "\\mathbf{p}" + }, + { + "bbox": [ + 304, + 472, + 545, + 590 + ], + "type": "text", + "content": " and the learned deformable offsets can then be used as keys " + }, + { + "bbox": [ + 304, + 472, + 545, + 590 + ], + "type": "inline_equation", + "content": "\\mathbf{K}" + }, + { + "bbox": [ + 304, + 472, + 545, + 590 + ], + "type": "text", + "content": ", with their corresponding features from the extracted images features the values " + }, + { + "bbox": [ + 304, + 472, + 545, + 590 + ], + "type": "inline_equation", + "content": "\\mathbf{V}" + }, + { + "bbox": [ + 304, + 472, + 545, + 590 + ], + "type": "text", + "content": ", for a cross-attention module" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 348, + 596, + 545, + 609 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 348, + 596, + 545, + 609 + ], + "spans": [ + { + "bbox": [ + 348, + 596, + 545, + 609 + ], + "type": "interline_equation", + "content": "\\mathcal {F} _ {B E V} = \\operatorname {D e f o r m A t t n} (\\mathbf {Q}, \\mathcal {F} _ {R G B}, \\mathbf {H}) \\tag {4}", + "image_path": "dc7349fac6515920841b8dca96c397f866852978401c6f65d00dee4765ed3b0f.jpg" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 304, + 613, + 545, + 673 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 613, + 545, + 673 + ], + "spans": [ + { + "bbox": [ + 304, + 613, + 545, + 673 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 613, + 545, + 673 + ], + "type": "inline_equation", + "content": "\\mathbf{H} = \\{\\mathbf{P}_k, \\mathbf{R}_k, t_k\\}_{k=1}^3" + }, + { + "bbox": [ + 304, + 613, + 545, + 673 + ], + "type": "text", + "content": " contains the combined multiview IPM parameters from Eqn. 3. In addition to the proposed alignment and teachable modules, we also incorporate prediction modules for auxiliary tasks (i.e., BEV prediction [16]), discussed in Sec. 3.4." + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "content": "GRU-based Waypoint Predictor: The transformed imaged-to-BEV features are inputted to three residual blocks and a GRU-based conditional branch based on the command [11]." + } + ] + } + ], + "index": 34 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7809" + } + ] + } + ], + "index": 35 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 132, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 132, + 83 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 132, + 83 + ], + "type": "text", + "content": "3.4. Student Loss" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 91, + 289, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 91, + 289, + 319 + ], + "spans": [ + { + "bbox": [ + 46, + 91, + 289, + 319 + ], + "type": "text", + "content": "Even with our carefully designed architecture, the student's image-based learning task remains significantly more challenging than the teacher's. Thus, both optimization and alignment with the teacher can benefit from extensive supervision and gradual student instruction, i.e., to scaffold supervision and better match the student's ability. Our distillation process provides ample beneficial supervision for the student through three holistic mechanisms. First, we incorporate deep distillation losses, i.e., both over output (i.e., waypoint regression task) and feature maps, as shown in Fig. 2, that can more directly supervise internal layers. Second, we incorporate additional auxiliary losses that regularize the optimization process. This includes ground-truth supervised losses, e.g., defined over segmentation and control command, and a geometric loss defined over the post-alignment features that supervises for correct BEV structure. Finally, we leverage a student-paced coaching mechanism for gradually increasing the difficulty of the waypoint prediction task throughout the iterations in Sec. 3.5." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 324, + 288, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 324, + 288, + 361 + ], + "spans": [ + { + "bbox": [ + 46, + 324, + 288, + 361 + ], + "type": "text", + "content": "Loss Functions: Our optimization objective for guiding the distillation process is a weighted sum over both distillation and auxiliary tasks:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 86, + 372, + 287, + 386 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 86, + 372, + 287, + 386 + ], + "spans": [ + { + "bbox": [ + 86, + 372, + 287, + 386 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {C a T} = \\mathcal {L} _ {\\text {o u t}} + \\mathcal {L} _ {\\text {f e a t}} + \\mathcal {L} _ {\\text {s e g}} + \\mathcal {L} _ {\\text {c m d}} \\tag {5}", + "image_path": "c0c547300e57b2f2f7a7ccb46fd6cc6905b621e4f9298a2b66ff0ba2f1bff20c.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 396, + 288, + 421 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 396, + 288, + 421 + ], + "spans": [ + { + "bbox": [ + 47, + 396, + 288, + 421 + ], + "type": "text", + "content": "The various loss terms enable guiding holistic aspects of student learning." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 426, + 287, + 451 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 426, + 287, + 451 + ], + "spans": [ + { + "bbox": [ + 47, + 426, + 287, + 451 + ], + "type": "text", + "content": "Output Distillation: We leverage an " + }, + { + "bbox": [ + 47, + 426, + 287, + 451 + ], + "type": "inline_equation", + "content": "L_{1}" + }, + { + "bbox": [ + 47, + 426, + 287, + 451 + ], + "type": "text", + "content": " loss computed over all of the conditional branches for each instance" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 91, + 461, + 287, + 495 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 461, + 287, + 495 + ], + "spans": [ + { + "bbox": [ + 91, + 461, + 287, + 495 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {o u t} = \\sum_ {c = 1} ^ {C} \\| f _ {\\boldsymbol {\\theta}} ^ {s} \\left(\\mathbf {x} ^ {s}, c\\right) - f _ {\\boldsymbol {\\psi}} ^ {t} \\left(\\mathbf {x} ^ {t}, c\\right) \\| _ {1} \\tag {6}", + "image_path": "e9455cf31fb6eeda21fa1c0a9df97afba69d5437f679bfc2ae9c252c12a1e1dc.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 503, + 287, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 503, + 287, + 529 + ], + "spans": [ + { + "bbox": [ + 47, + 503, + 287, + 529 + ], + "type": "text", + "content": "This can be done by sampling from the teacher with different commands " + }, + { + "bbox": [ + 47, + 503, + 287, + 529 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 47, + 503, + 287, + 529 + ], + "type": "text", + "content": " (following Chen et al. [13])." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 533, + 289, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 533, + 289, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 533, + 289, + 714 + ], + "type": "text", + "content": "Feature Distillation Loss: Our proposed student architecture provides BEV-space feature alignment, i.e., for direct internal feature matching between the student and the teacher (the first term in our loss in Eqn. 7). This is in contrast to current methods which distill the output or a final fully-connected layer [11, 13, 71, 79]. However, while this can provide rich supervision for the student, the two models should not match entirely, i.e., due to differences in information processing for a perception-based planning task. Moreover, the BEV-based features can be highly sparse and structured, with slight feature offsets being less meaningful than obtaining a task and scene-relevant representation. To flexibly account for the student-based task and effectively represent structural information in BEV space, our feature loss is computed as" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 340, + 70, + 545, + 141 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 340, + 70, + 545, + 141 + ], + "spans": [ + { + "bbox": [ + 340, + 70, + 545, + 141 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} _ {f e a t} = \\sum_ {i = 1} ^ {3} \\left[ \\| \\mathcal {F} _ {i} ^ {s} (\\mathbf {x} ^ {s}) - \\mathcal {F} _ {i} ^ {t} (\\mathbf {x} ^ {t}) \\| _ {2} + \\right. \\\\ \\left\\| T _ {i} ^ {s} \\left(\\mathcal {F} _ {i} ^ {s} \\left(\\mathbf {x} ^ {s}\\right)\\right) - T _ {i} ^ {t} \\left(\\mathcal {F} _ {i} ^ {t} \\left(\\mathbf {x} ^ {t}\\right)\\right) \\right\\| _ {2} + \\tag {7} \\\\ \\lambda_ {C D} \\left\\| \\mathcal {F} _ {i} ^ {s} (\\mathbf {x} ^ {s}) - \\mathcal {F} _ {i} ^ {t} (\\mathbf {x} ^ {t})\\right) \\| _ {C D} \\Bigg ] \\\\ \\end{array}", + "image_path": "36508ca4c333c3c3b2b248eba7a8ab2b9d3c286fba99146cc48c188f857dd7df.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 148, + 547, + 234 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 148, + 547, + 234 + ], + "spans": [ + { + "bbox": [ + 304, + 148, + 547, + 234 + ], + "type": "text", + "content": "where we use " + }, + { + "bbox": [ + 304, + 148, + 547, + 234 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_i" + }, + { + "bbox": [ + 304, + 148, + 547, + 234 + ], + "type": "text", + "content": " to denote layers within the networks (as shown in Fig. 2, we directly distill three layers), " + }, + { + "bbox": [ + 304, + 148, + 547, + 234 + ], + "type": "inline_equation", + "content": "T_{i}" + }, + { + "bbox": [ + 304, + 148, + 547, + 234 + ], + "type": "text", + "content": " indicates the features following several convolutional layers (trained jointly with the student network model), and " + }, + { + "bbox": [ + 304, + 148, + 547, + 234 + ], + "type": "inline_equation", + "content": "CD" + }, + { + "bbox": [ + 304, + 148, + 547, + 234 + ], + "type": "text", + "content": " stands for a Chamfer Distance [51] (we set " + }, + { + "bbox": [ + 304, + 148, + 547, + 234 + ], + "type": "inline_equation", + "content": "\\lambda_{CD} = 0.1" + }, + { + "bbox": [ + 304, + 148, + 547, + 234 + ], + "type": "text", + "content": "). The final term is computed over thresholded activations and a spatial soft-argmax [13] from each feature map." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 238, + 545, + 358 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 238, + 545, + 358 + ], + "spans": [ + { + "bbox": [ + 304, + 238, + 545, + 358 + ], + "type": "text", + "content": "Task-oriented Auxiliary Tasks: To ensure essential task-relevant information is preserved throughout the distillation process, we leverage regularizing supervision in the form of segmentation and command-based auxiliary tasks interlaced throughout the network [16, 49]. Our segmentation loss " + }, + { + "bbox": [ + 304, + 238, + 545, + 358 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{seg}" + }, + { + "bbox": [ + 304, + 238, + 545, + 358 + ], + "type": "text", + "content": " is a cross-entropy loss, computed over the ground-truth BEV and averaged across the segmentation prediction heads. Our command prediction loss " + }, + { + "bbox": [ + 304, + 238, + 545, + 358 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{cmd}" + }, + { + "bbox": [ + 304, + 238, + 545, + 358 + ], + "type": "text", + "content": " is binary cross-entropy supervised by the ground-truth command " + }, + { + "bbox": [ + 304, + 238, + 545, + 358 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 304, + 238, + 545, + 358 + ], + "type": "text", + "content": ", which facilitates learning task-relevant features [11, 19]." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 365, + 444, + 379 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 365, + 444, + 379 + ], + "spans": [ + { + "bbox": [ + 306, + 365, + 444, + 379 + ], + "type": "text", + "content": "3.5. Student-paced Coaching" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 384, + 545, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 384, + 545, + 468 + ], + "spans": [ + { + "bbox": [ + 304, + 384, + 545, + 468 + ], + "type": "text", + "content": "While Sec. 3.4 provides ample supervision on the student's internal representations, the teacher's targets may still be difficult to imitate for the student. To better consider the learning ability of the student, we propose to leverage a student-paced training mechanism which can gradually coach the student, i.e., through increasingly difficult samples. When computing the loss in Eqn. 5, we define a coach," + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 340, + 472, + 545, + 486 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 340, + 472, + 545, + 486 + ], + "spans": [ + { + "bbox": [ + 340, + 472, + 545, + 486 + ], + "type": "interline_equation", + "content": "\\mathcal {F} ^ {t} \\leftarrow \\lambda_ {i} \\mathcal {F} ^ {s} + \\left(1 - \\lambda_ {i}\\right) \\mathcal {F} ^ {t}, \\text {i f} \\mathcal {L} _ {C a T} > \\tau_ {i} \\tag {8}", + "image_path": "07283ac2dcf1d584050b89f4f79add88e5d7e31a928e37e89c8064e1e4582f93.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 491, + 545, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 491, + 545, + 647 + ], + "spans": [ + { + "bbox": [ + 304, + 491, + 545, + 647 + ], + "type": "text", + "content": "which interpolates the teaching targets with the student's predictions and features. " + }, + { + "bbox": [ + 304, + 491, + 545, + 647 + ], + "type": "inline_equation", + "content": "\\lambda_{i}" + }, + { + "bbox": [ + 304, + 491, + 545, + 647 + ], + "type": "text", + "content": " is a hyperparameter which is linearly decreased to 0 over training iteration " + }, + { + "bbox": [ + 304, + 491, + 545, + 647 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 491, + 545, + 647 + ], + "type": "text", + "content": ". The modified targets are only computed over the hard samples at each batch. Within each batch, " + }, + { + "bbox": [ + 304, + 491, + 545, + 647 + ], + "type": "inline_equation", + "content": "\\tau_{i}" + }, + { + "bbox": [ + 304, + 491, + 545, + 647 + ], + "type": "text", + "content": " is a threshold defining the lowest " + }, + { + "bbox": [ + 304, + 491, + 545, + 647 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 304, + 491, + 545, + 647 + ], + "type": "text", + "content": " of the scored samples for the loss with respect to the privileged teacher. While smoothing the targets in this manner may seem counterintuitive, i.e., compared to techniques which mine hard examples, the coaching mechanism aims to stabilize training by reducing the difficulty when the student is unable to perform the optimal action. Eqn. 8 enables adjusting the learning rate in a sample-selective manner (initially suppressing difficult samples)." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 658, + 388, + 671 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 658, + 388, + 671 + ], + "spans": [ + { + "bbox": [ + 306, + 658, + 388, + 671 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 677, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 545, + 714 + ], + "type": "text", + "content": "We use the CARLA simulator (version 0.9.10.1) [21] for data generation and closed-loop evaluation of the proposed CaT framework. We also use open-loop evaluation" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7810" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 47, + 129, + 545, + 297 + ], + "blocks": [ + { + "bbox": [ + 46, + 70, + 547, + 127 + ], + "lines": [ + { + "bbox": [ + 46, + 70, + 547, + 127 + ], + "spans": [ + { + "bbox": [ + 46, + 70, + 547, + 127 + ], + "type": "text", + "content": "Table 1. Quantitative Evaluation on the Longest6 Benchmark. Comparison of CaT with prior methods in terms of Driving Score (DS), Route Completion (RC), and Infraction Score (IS). Additional infraction metrics (Pedestrian Collisions (Ped), Vehicle Collisions (Veh), Layout Collisions (LC), Red Light Violations (Red), Off-road Infraction (OR), Route deviation (Dev), Route Timouts (TO), Agent Blocked (Blk)) are shown. FD refers to Feature Distillation, SH refers to Safety Hints. *re-trained by us using the publicly available code. Mean and standard deviation are computed over three runs. CaT outperforms the state-of-the-art by a large margin in terms of DS." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 47, + 129, + 545, + 297 + ], + "lines": [ + { + "bbox": [ + 47, + 129, + 545, + 297 + ], + "spans": [ + { + "bbox": [ + 47, + 129, + 545, + 297 + ], + "type": "table", + "html": "
MethodRGBLiDARDS↑RC↑IS↑Ped↓Veh↓LC↓Red↓OR↓Dev↓TO↓Blk↓
LAV [11]✓✓48.41±3.4080.71±0.840.60±0.040.000.500.190.070.200.070.010.29
TransFuser [16]✓✓46.20±2.5783.61±1.160.57±0.000.290.380.280.040.070.070.000.32
WOR [10]✓X17.36±2.9543.46±2.990.54±0.060.050.640.150.840.150.890.040.45
NEAT [15]✓X24.08±3.3059.94±0.500.49±0.020.010.710.210.180.180.000.022.83
TCP* [71]✓X42.86±0.6361.83±4.190.71±0.040.010.420.110.040.080.010.010.75
CaT (w/o Alignment, Coaching, FD)✓X39.48±0.6760.96±1.650.68±0.010.031.290.200.080.840.000.021.49
CaT (w/o Alignment, Coaching)✓X40.64±0.9862.45±0.460.67±0.010.021.070.330.140.380.000.021.11
CaT (w/o Coaching, FD, SH)✓X44.10±0.4065.84±5.550.72±0.030.010.260.050.110.150.000.000.66
CaT (w/o Coaching, SH)✓X49.69±2.2881.10±0.580.64±0.020.010.780.030.080.200.000.020.51
CaT (w/o Coaching)✓X55.55±1.4181.97±2.340.68±0.010.020.300.050.070.060.000.020.35
CaT✓X58.36±2.2478.79±1.500.77±0.020.010.200.020.050.300.000.040.44
Privileged Agents:
RL Expert (Roach) [79]--60.14±2.4085.83±0.600.69±0.030.060.220.000.010.000.000.040.07
Rule-based Expert--71.96±2.1377.46±3.110.91±0.000.010.060.000.000.000.000.030.41
Basic BEV Agent [13]--24.08±2.8373.36±1.080.31±0.060.011.452.450.090.270.110.010.24
+ History and Desired Path--52.81±1.7979.34±3.650.71±0.060.020.300.000.000.000.000.020.53
+ Agent Forecast--65.73±0.9383.50±1.180.79±0.020.010.210.000.020.000.000.010.28
+ Entity Attention--73.30±1.0787.44±0.280.83±0.020.050.110.000.020.000.000.020.12
", + "image_path": "72c197e27232dbec756df79c290655552364a92f5996bd6ebb089498291960f7.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 306, + 289, + 462 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 306, + 289, + 462 + ], + "spans": [ + { + "bbox": [ + 46, + 306, + 289, + 462 + ], + "type": "text", + "content": "on nuScenes [7, 33, 77] (ADE, FDE, and collision rate). In Carla, we leverage the Longest6 Benchmark [16], which uses the six longest routes of each town (Town01 - Town06) from the set of 76 routes provided by the official CARLA leaderboard [1] (a total of 36 routes). We note in addition to the standard cross weather and time of day generalization evaluation, the long and dense traffic conditions in Longest6 present the most challenging settings across the various CARLA-based benchmarks. To evaluate our models, we follow standard metrics and report Route Completion (RC, in terms of completed route percentage), Infraction Score (IS, a penalty factor over infractions), and Driving Score (DS, computed from the prior two) [1, 11, 16])." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 479, + 223, + 491 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 479, + 223, + 491 + ], + "spans": [ + { + "bbox": [ + 47, + 479, + 223, + 491 + ], + "type": "text", + "content": "4.1. Comparison with Prior Methods" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 498, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 498, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 498, + 289, + 715 + ], + "type": "text", + "content": "CARLA Results: As shown in Table 1, using feature distillation, safety hints-based BEV for the teacher, and the proposed coaching mechanism, CaT is able to obtain a " + }, + { + "bbox": [ + 46, + 498, + 289, + 715 + ], + "type": "inline_equation", + "content": "58.36\\%" + }, + { + "bbox": [ + 46, + 498, + 289, + 715 + ], + "type": "text", + "content": " DS. Specifically, we achieve state-of-the-art performance among all prior agents, including LiDAR-based approaches [11, 16] (by " + }, + { + "bbox": [ + 46, + 498, + 289, + 715 + ], + "type": "inline_equation", + "content": "20.6\\%" + }, + { + "bbox": [ + 46, + 498, + 289, + 715 + ], + "type": "text", + "content": " in terms of DS, from 48.41 to 58.36, and " + }, + { + "bbox": [ + 46, + 498, + 289, + 715 + ], + "type": "inline_equation", + "content": "28.3\\%" + }, + { + "bbox": [ + 46, + 498, + 289, + 715 + ], + "type": "text", + "content": " in terms of IS). We also note that such approaches provide strong baselines, due to various LiDAR-based safety checks that the agents perform on top of the trained policy, which we do not employ. Moreover, CaT outperforms the prior RGB-only state-of-the-art agent TCP [71] by " + }, + { + "bbox": [ + 46, + 498, + 289, + 715 + ], + "type": "inline_equation", + "content": "36.16\\%" + }, + { + "bbox": [ + 46, + 498, + 289, + 715 + ], + "type": "text", + "content": " DS. To further validate the generalization of CaT, we discuss an additional benchmark split from TCP [71] and LAV [11] in the supplementary. We also find that removal of the alignment module degrades DS from " + }, + { + "bbox": [ + 46, + 498, + 289, + 715 + ], + "type": "inline_equation", + "content": "44.10\\%" + }, + { + "bbox": [ + 46, + 498, + 289, + 715 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 46, + 498, + 289, + 715 + ], + "type": "inline_equation", + "content": "39.48\\%" + }, + { + "bbox": [ + 46, + 498, + 289, + 715 + ], + "type": "text", + "content": ", and is shown to only slightly benefit from feature distillation (" + }, + { + "bbox": [ + 46, + 498, + 289, + 715 + ], + "type": "inline_equation", + "content": "40.64\\%" + }, + { + "bbox": [ + 46, + 498, + 289, + 715 + ], + "type": "text", + "content": " DS) indicating the effectiveness of the proposed architecture." + } + ] + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 310, + 317, + 542, + 365 + ], + "blocks": [ + { + "bbox": [ + 338, + 305, + 512, + 316 + ], + "lines": [ + { + "bbox": [ + 338, + 305, + 512, + 316 + ], + "spans": [ + { + "bbox": [ + 338, + 305, + 512, + 316 + ], + "type": "text", + "content": "Table 2. Open-Loop Evaluation on nuScenes." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 310, + 317, + 542, + 365 + ], + "lines": [ + { + "bbox": [ + 310, + 317, + 542, + 365 + ], + "spans": [ + { + "bbox": [ + 310, + 317, + 542, + 365 + ], + "type": "table", + "html": "
MethodADE (m) ↓FDE (m) ↓Coll. (%) ↓
BEV Agent0.330.520.49
CaT (w/o Coaching, FD, SH)0.480.430.68
CaT0.410.360.27
", + "image_path": "9bb5654956a20a02dab4ee9460c67a51131b440c31eb7dfbb81d3d24b11eb68f.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 376, + 545, + 497 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 376, + 545, + 497 + ], + "spans": [ + { + "bbox": [ + 304, + 376, + 545, + 497 + ], + "type": "text", + "content": "Real-World Evaluation: To further analyse the benefits of CaT in realistic driving settings, Table 2 shows open-loop evaluation (ADE, FDE, and Collision rate [77]) for nuScenes [7]. On the official validation split, the privileged agent performs best with an ADE of 0.33. CaT achieves 0.41 ADE, improving by " + }, + { + "bbox": [ + 304, + 376, + 545, + 497 + ], + "type": "inline_equation", + "content": "14.6\\%" + }, + { + "bbox": [ + 304, + 376, + 545, + 497 + ], + "type": "text", + "content": " gain over an agent without distillation and coaching. Moreover, we find a significant reduction in the collision rate by " + }, + { + "bbox": [ + 304, + 376, + 545, + 497 + ], + "type": "inline_equation", + "content": "60.3\\%" + }, + { + "bbox": [ + 304, + 376, + 545, + 497 + ], + "type": "text", + "content": " compared to the baseline. Additional details regarding nuScenes evaluation can be found in the supplementary." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 506, + 406, + 518 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 506, + 406, + 518 + ], + "spans": [ + { + "bbox": [ + 306, + 506, + 406, + 518 + ], + "type": "text", + "content": "4.2. Ablation Studies" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 527, + 545, + 682 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 527, + 545, + 682 + ], + "spans": [ + { + "bbox": [ + 304, + 527, + 545, + 682 + ], + "type": "text", + "content": "Teacher Comparison: Table 1 depicts our teacher model ablation. Specifically, we find the proposed BEV channels to drastically improve the privileged agent's performance by simplifying the learning task (73.30% DS, surpassing the built-in rule-based expert of 71.96%). We note that our behavior cloning agent also significantly outperforms an RL-based expert [79] (60.14%), simply through an effective BEV design. Thus, our findings apply to real-world scenarios where interactive agent training may be unsafe. We also observe that the improved teacher benefits the student agent. Specifically, incorporating the proposed BEV-based safety hints results in student performance gains, from 49.69% to 55.55%." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 689, + 545, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 689, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 305, + 689, + 545, + 715 + ], + "type": "text", + "content": "Impact of Feature Distillation: Table 3 presents the benefits of our proposed multi-layer feature distillation frame" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "7811" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 96, + 70, + 496, + 290 + ], + "blocks": [ + { + "bbox": [ + 96, + 70, + 496, + 290 + ], + "lines": [ + { + "bbox": [ + 96, + 70, + 496, + 290 + ], + "spans": [ + { + "bbox": [ + 96, + 70, + 496, + 290 + ], + "type": "image", + "image_path": "be6e47ce4a4140cd30f20b8992747d2d209dbabb2f92c0ab24a41a7f3729d30a.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 291, + 547, + 348 + ], + "lines": [ + { + "bbox": [ + 46, + 291, + 547, + 348 + ], + "spans": [ + { + "bbox": [ + 46, + 291, + 547, + 348 + ], + "type": "text", + "content": "Figure 4. Visualization of Model Input and Features. We show the BEV ground truth (BEV) and RGB image ground truth (Image) on the top row. Next, we visualize the intermediate layer activations of a sensorimotor agent without the proposed alignment module [71] (No Alignment Module), privileged teacher agent (Privileged), student agent trained with output distillation only (Output Distillation), and the proposed agent (CaT). We demonstrate that our teacher agent with proposed knowledge distillation is able to guide the student agent in learning scene-relevant representations and attend to safety-critical entities (circled in red)." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 61, + 381, + 273, + 466 + ], + "blocks": [ + { + "bbox": [ + 47, + 356, + 287, + 378 + ], + "lines": [ + { + "bbox": [ + 47, + 356, + 287, + 378 + ], + "spans": [ + { + "bbox": [ + 47, + 356, + 287, + 378 + ], + "type": "text", + "content": "Table 3. Ablation Study on Feature Distillation Layers. All loss terms in Eqn. 7 benefit training with deep feature distillation." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 61, + 381, + 273, + 466 + ], + "lines": [ + { + "bbox": [ + 61, + 381, + 273, + 466 + ], + "spans": [ + { + "bbox": [ + 61, + 381, + 273, + 466 + ], + "type": "table", + "html": "
MethodDS ↑RC ↑IS ↑
No Distillation44.1065.840.72
One Layer [71,79]45.2369.330.68
Three Layers L249.3166.920.78
Three Layers L2 + LCD51.9562.820.87
Three Layers Lfeat55.5581.970.68
", + "image_path": "0f8afabc7b8a470317888a0db8ca37762ec1efbe6c9ff97c01af83b76dea9d6a.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 484, + 287, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 484, + 287, + 568 + ], + "spans": [ + { + "bbox": [ + 46, + 484, + 287, + 568 + ], + "type": "text", + "content": "work. We note that prior methods only distill one final fully connected layer (before the prediction head [71, 79], referred to as 'one layer' in Table 3). We note that such approaches result in minimal gains, whereas CaT is able to benefit from the extensive feature distillation (improving from " + }, + { + "bbox": [ + 46, + 484, + 287, + 568 + ], + "type": "inline_equation", + "content": "45.23\\%" + }, + { + "bbox": [ + 46, + 484, + 287, + 568 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 46, + 484, + 287, + 568 + ], + "type": "inline_equation", + "content": "55.55\\%" + }, + { + "bbox": [ + 46, + 484, + 287, + 568 + ], + "type": "text", + "content": "). Incorporating multiple types of feature distillation losses also benefits performance." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 575, + 170, + 588 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 575, + 170, + 588 + ], + "spans": [ + { + "bbox": [ + 47, + 575, + 170, + 588 + ], + "type": "text", + "content": "4.3. Feature Visualization" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 594, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 288, + 715 + ], + "type": "text", + "content": "Fig. 4 depicts the learned features of the initial feature distilled layer (out of the three in Fig. 2). We visualize activations from two channels, to better understand the learned representation by the agents. Interestingly, the visualization shows how traditionally trained sensorimotor agents (i.e., without the proposed alignment module) mostly operate in image-space features in this layer despite the BEV waypoint prediction task. This validates our main hypothesis, that such agents have limited ability to learn from a BEV-based agent, even with BEV-based auxiliary prediction tasks. We" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 357, + 547, + 442 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 357, + 547, + 442 + ], + "spans": [ + { + "bbox": [ + 304, + 357, + 547, + 442 + ], + "type": "text", + "content": "also find that output distillation provides a weak supervision for critical components of the BEV, including layout and object features. After feature distillation, the features of our CaT model are able to not only keep the driving-related semantic information i.e., drivable road and route, but also focus on safety-critical entities. More visualization examples are provided in the supplementary." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 457, + 378, + 469 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 457, + 378, + 469 + ], + "spans": [ + { + "bbox": [ + 306, + 457, + 378, + 469 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 479, + 547, + 682 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 479, + 547, + 682 + ], + "spans": [ + { + "bbox": [ + 304, + 479, + 547, + 682 + ], + "type": "text", + "content": "We present CaT, a novel knowledge distillation method for effectively training a sensorimotor student agent using the supervision from a privileged agent. We leverage an alignment module to better map image features to the BEV space, thus enabling extensive supervision from a BEV-based teacher over the intermediate feature learning. As both agents can be trained offline via imitation learning, our findings are directly relevant to real-world settings where on-policy interactions may be unsafe to perform. Finally, we carefully account for inherent differences between the student and the teacher using a student-paced coaching mechanism with various auxiliary supervision tasks. Through the improved knowledge distillation process, our experiments result in a state-of-the-art camera-based agent in CARLA. Our ablation into input and model design guides future directions through interpretable analysis into current limitations of end-to-end driving models." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 689, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 689, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 305, + 689, + 545, + 714 + ], + "type": "text", + "content": "Acknowledgments: We thank the Red Hat Collaboratory for supporting this research." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7812" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 53, + 91, + 286, + 113 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 286, + 113 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 286, + 113 + ], + "type": "text", + "content": "[1] Carla autonomous driving leaderboard. https://leaderboard.carla.org/, 2022.7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 114, + 287, + 147 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 114, + 287, + 147 + ], + "spans": [ + { + "bbox": [ + 53, + 114, + 287, + 147 + ], + "type": "text", + "content": "[2] Yonathan Aflalo, Asaf Noy, Ming Lin, Itamar Friedman, and Lihi Zelnik. Knapsack pruning with inner distillation. arXiv preprint arXiv:2002.08258, 2020. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 148, + 287, + 180 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 148, + 287, + 180 + ], + "spans": [ + { + "bbox": [ + 53, + 148, + 287, + 180 + ], + "type": "text", + "content": "[3] Gustavo Aguilar, Yuan Ling, Yu Zhang, Benjamin Yao, Xing Fan, and Chenlei Guo. Knowledge distillation from internal representations. In AAAI, 2020. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 182, + 287, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 182, + 287, + 215 + ], + "spans": [ + { + "bbox": [ + 53, + 182, + 287, + 215 + ], + "type": "text", + "content": "[4] Mayank Bansal, Alex Krizhevsky, and Abhijit Ogale. ChauffeurNet: Learning to drive by imitating the best and synthesizing the worst. In RSS, 2019. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 217, + 287, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 217, + 287, + 237 + ], + "spans": [ + { + "bbox": [ + 53, + 217, + 287, + 237 + ], + "type": "text", + "content": "[5] Yoshua Bengio, Jérôme Louradour, Ronan Collobert, and Jason Weston. Curriculum learning. In ICML, 2009. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 239, + 287, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 239, + 287, + 271 + ], + "spans": [ + { + "bbox": [ + 53, + 239, + 287, + 271 + ], + "type": "text", + "content": "[6] Massimo Bertozzi, Alberto Broggi, and Alessandra Fascioli. An extension to the inverse perspective mapping to handle non-flat roads. In IV, 1998. 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 274, + 287, + 317 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 274, + 287, + 317 + ], + "spans": [ + { + "bbox": [ + 53, + 274, + 287, + 317 + ], + "type": "text", + "content": "[7] Holger Caesar, Varun Bankiti, Alex H Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuscenes: A multimodal dataset for autonomous driving. In CVPR, 2020. 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 319, + 287, + 361 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 319, + 287, + 361 + ], + "spans": [ + { + "bbox": [ + 53, + 319, + 287, + 361 + ], + "type": "text", + "content": "[8] Raphael Chekroun, Marin Toromanoff, Sascha Hornauer, and Fabien Moutarde. Gri: General reinforced imitation and its application to vision-based autonomous driving. arXiv preprint arXiv:2111.08575, 2021. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 364, + 287, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 364, + 287, + 396 + ], + "spans": [ + { + "bbox": [ + 53, + 364, + 287, + 396 + ], + "type": "text", + "content": "[9] Chenyi Chen, Ari Seff, Alain L. Kornhauser, and Jianxiong Xiao. DeepDriving: Learning affordance for direct perception in autonomous driving. In ICCV, 2015. 2, 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 397, + 287, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 397, + 287, + 418 + ], + "spans": [ + { + "bbox": [ + 48, + 397, + 287, + 418 + ], + "type": "text", + "content": "[10] Dian Chen, Vladlen Koltun, and Philipp Krahenbuhl. Learning to drive from a world on rails. In ICCV, 2021. 2, 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 421, + 287, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 421, + 287, + 441 + ], + "spans": [ + { + "bbox": [ + 48, + 421, + 287, + 441 + ], + "type": "text", + "content": "[11] Dian Chen and Philipp Krahenbuhl. Learning from all vehicles. In CVPR, 2022. 1, 2, 3, 4, 5, 6, 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 444, + 287, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 444, + 287, + 475 + ], + "spans": [ + { + "bbox": [ + 48, + 444, + 287, + 475 + ], + "type": "text", + "content": "[12] Defang Chen, Jian-Ping Mei, Yuan Zhang, Can Wang, Zhe Wang, Yan Feng, and Chun Chen. Cross-layer distillation with semantic calibration. In AAAI, 2021. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 477, + 287, + 509 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 477, + 287, + 509 + ], + "spans": [ + { + "bbox": [ + 48, + 477, + 287, + 509 + ], + "type": "text", + "content": "[13] Dian Chen, Brady Zhou, Vladlen Koltun, and Philipp Krahenbuhl. Learning by cheating. In CoRL, 2020. 1, 2, 3, 4, 6, 7" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 511, + 287, + 565 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 511, + 287, + 565 + ], + "spans": [ + { + "bbox": [ + 48, + 511, + 287, + 565 + ], + "type": "text", + "content": "[14] Li Chen, Chonghao Sima, Yang Li, Zehan Zheng, Jiajie Xu, Xiangwei Geng, Hongyang Li, Conghui He, Jianping Shi, Yu Qiao, et al. Persformer: 3d lane detection via perspective transformer and the openlane benchmark. arXiv preprint arXiv:2203.11089, 2022. 2, 5" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 567, + 287, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 567, + 287, + 599 + ], + "spans": [ + { + "bbox": [ + 48, + 567, + 287, + 599 + ], + "type": "text", + "content": "[15] Kashyap Chitta, Aditya Prakash, and Andreas Geiger. Neat: Neural attention fields for end-to-end autonomous driving. In ICCV, 2021. 7" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 48, + 601, + 287, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 601, + 287, + 645 + ], + "spans": [ + { + "bbox": [ + 48, + 601, + 287, + 645 + ], + "type": "text", + "content": "[16] Kashyap Chitta, Aditya Prakash, Bernhard Jaeger, Zehao Yu, Katrin Renz, and Andreas Geiger. Transfuser: Imitation with transformer-based sensor fusion for autonomous driving. arXiv preprint arXiv:2205.15997, 2022. 3, 4, 5, 6, 7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 48, + 647, + 287, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 647, + 287, + 678 + ], + "spans": [ + { + "bbox": [ + 48, + 647, + 287, + 678 + ], + "type": "text", + "content": "[17] Inseop Chung, SeongUk Park, Jangho Kim, and Nojun Kwak. Feature-map-level online adversarial knowledge distillation. In ICML, 2020. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 48, + 680, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 680, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 680, + 287, + 712 + ], + "type": "text", + "content": "[18] Felipe Codevilla, Matthias Muiller, Antonio López, Vladlen Koltun, and Alexey Dosovitskiy. End-to-end driving via conditional imitation learning. In ICRA, 2018. 3, 4" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 38, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "text", + "content": "[19] Felipe Codevilla, Eder Santana, Antonio M López, and Adrien Gaidon. Exploring the limitations of behavior cloning for autonomous driving. In ICCV, 2019. 4, 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 108, + 545, + 129 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 108, + 545, + 129 + ], + "spans": [ + { + "bbox": [ + 307, + 108, + 545, + 129 + ], + "type": "text", + "content": "[20] Pim de Haan, Dinesh Jayaraman, and Sergey Levine. Causal confusion in imitation learning. In NeurIPS, 2019. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 131, + 545, + 163 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 131, + 545, + 163 + ], + "spans": [ + { + "bbox": [ + 307, + 131, + 545, + 163 + ], + "type": "text", + "content": "[21] Alexey Dosovitskiy, German Ros, Felipe Codevilla, Antonio Lopez, and Vladlen Koltun. CARLA: An open urban driving simulator. In CoRL, 2017. 2, 3, 4, 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 166, + 545, + 187 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 166, + 545, + 187 + ], + "spans": [ + { + "bbox": [ + 307, + 166, + 545, + 187 + ], + "type": "text", + "content": "[22] James J Gibson. The ecological approach to visual perception: classic edition. Psychology press, 2014. 4" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 189, + 545, + 220 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 189, + 545, + 220 + ], + "spans": [ + { + "bbox": [ + 307, + 189, + 545, + 220 + ], + "type": "text", + "content": "[23] Jianping Gou, Baosheng Yu, Stephen J Maybank, and Dacheng Tao. Knowledge distillation: A survey. In IJCV, 2021. 1, 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 223, + 545, + 266 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 223, + 545, + 266 + ], + "spans": [ + { + "bbox": [ + 307, + 223, + 545, + 266 + ], + "type": "text", + "content": "[24] Sheng Guo, Weilin Huang, Haozhi Zhang, Chenfan Zhuang, Dengke Dong, Matthew R Scott, and Dinglong Huang. Curriculumnet: Weakly supervised learning from large-scale web images. In ECCV, 2018. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 269, + 545, + 301 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 269, + 545, + 301 + ], + "spans": [ + { + "bbox": [ + 307, + 269, + 545, + 301 + ], + "type": "text", + "content": "[25] Saurabh Gupta, James Davidson, Sergey Levine, Rahul Sukthankar, and Jitendra Malik. Cognitive mapping and planning for visual navigation. In CVPR, 2017. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 304, + 545, + 345 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 304, + 545, + 345 + ], + "spans": [ + { + "bbox": [ + 307, + 304, + 545, + 345 + ], + "type": "text", + "content": "[26] Peter E Hart, Nils J Nilsson, and Bertram Raphael. A formal basis for the heuristic determination of minimum cost paths. IEEE transactions on Systems Science and Cybernetics, 1968, 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 349, + 545, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 349, + 545, + 369 + ], + "spans": [ + { + "bbox": [ + 307, + 349, + 545, + 369 + ], + "type": "text", + "content": "[27] He He, Jason Eisner, and Hal Daume. Imitation learning by coaching. In NeurIPS, 2012. 1, 2, 3" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 373, + 545, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 373, + 545, + 403 + ], + "spans": [ + { + "bbox": [ + 307, + 373, + 545, + 403 + ], + "type": "text", + "content": "[28] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, 2016. 4" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 307, + 407, + 545, + 438 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 407, + 545, + 438 + ], + "spans": [ + { + "bbox": [ + 307, + 407, + 545, + 438 + ], + "type": "text", + "content": "[29] Byeongho Heo, Jeesoo Kim, Sangwoo Yun, Hyojin Park, No-jun Kwak, and Jin Young Choi. A comprehensive overhaul of feature distillation. In ICCV, 2019. 2" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 307, + 441, + 545, + 473 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 441, + 545, + 473 + ], + "spans": [ + { + "bbox": [ + 307, + 441, + 545, + 473 + ], + "type": "text", + "content": "[30] Roger Hiemstra and Burton Sisco. Individualizing Instruction. Making Learning Personal, Empowering, and Successful. 1990. 2" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 307, + 475, + 545, + 507 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 475, + 545, + 507 + ], + "spans": [ + { + "bbox": [ + 307, + 475, + 545, + 507 + ], + "type": "text", + "content": "[31] Geoffrey Hinton, Oriol Vinyals, Jeff Dean, et al. Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531, 2(7), 2015. 2" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 307, + 510, + 545, + 563 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 510, + 545, + 563 + ], + "spans": [ + { + "bbox": [ + 307, + 510, + 545, + 563 + ], + "type": "text", + "content": "[32] Anthony Hu, Zak Murez, Nikhil Mohan, Sofia Dudas, Jeffrey Hawke, Vijay Badrinarayanan, Roberto Cipolla, and Alex Kendall. Fiery: Future instance prediction in bird's-eye view from surround monocular cameras. In ICCV, 2021. 2, 3, 5" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 307, + 567, + 545, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 567, + 545, + 599 + ], + "spans": [ + { + "bbox": [ + 307, + 567, + 545, + 599 + ], + "type": "text", + "content": "[33] Peiyun Hu, Aaron Huang, John Dolan, David Held, and Deva Ramanan. Safe local motion planning with self-supervised freespace forecasting. In CVPR, 2021. 7" + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 307, + 601, + 545, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 601, + 545, + 643 + ], + "spans": [ + { + "bbox": [ + 307, + 601, + 545, + 643 + ], + "type": "text", + "content": "[34] Zanming Huang, Zhongkai Shangguan, Jimuyang Zhang, Gilad Bar, Matthew Boyd, and Eshed Ohn-Bar. ASSISTER: Assistive navigation via conditional instruction generation. In ECCV, 2022. 3" + } + ] + } + ], + "index": 35 + }, + { + "bbox": [ + 307, + 646, + 545, + 677 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 646, + 545, + 677 + ], + "spans": [ + { + "bbox": [ + 307, + 646, + 545, + 677 + ], + "type": "text", + "content": "[35] Zehao Huang and Naiyan Wang. Like what you like: Knowledge distill via neuron selectivity transfer. arXiv preprint arXiv:1707.01219, 2017. 2" + } + ] + } + ], + "index": 36 + }, + { + "bbox": [ + 307, + 680, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 680, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 307, + 680, + 545, + 712 + ], + "type": "text", + "content": "[36] Mingi Ji, Byeongho Heo, and Sungrae Park. Show, attend and distill: Knowledge distillation via attention-based feature matching. In AAAI, 2021. 2" + } + ] + } + ], + "index": 37 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7813" + } + ] + } + ], + "index": 39 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "text", + "content": "[37] Lu Jiang, Zhengyuan Zhou, Thomas Leung, Li-Jia Li, and Li Fei-Fei. Mentornet: Learning data-driven curriculum for very deep neural networks on corrupted labels. In ICML, 2018. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 118, + 287, + 151 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 118, + 287, + 151 + ], + "spans": [ + { + "bbox": [ + 48, + 118, + 287, + 151 + ], + "type": "text", + "content": "[38] Jangho Kim, SeongUk Park, and Nojun Kwak. Paraphrasing complex network: Network compression via factor transfer. In NeurIPS, 2018. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 152, + 287, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 152, + 287, + 185 + ], + "spans": [ + { + "bbox": [ + 48, + 152, + 287, + 185 + ], + "type": "text", + "content": "[39] Tom Kocmi and Ondrej Bojar. Curriculum learning and minibatch bucketing in neural machine translation. arXiv preprint arXiv:1707.09533, 2017. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 187, + 287, + 220 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 187, + 287, + 220 + ], + "spans": [ + { + "bbox": [ + 48, + 187, + 287, + 220 + ], + "type": "text", + "content": "[40] Animesh Koratana, Daniel Kang, Peter Bailis, and Matei Zaharia. Lit: Learned intermediate representation training for model compression. In ICML, 2019. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 222, + 287, + 243 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 222, + 287, + 243 + ], + "spans": [ + { + "bbox": [ + 48, + 222, + 287, + 243 + ], + "type": "text", + "content": "[41] M Kumar, Benjamin Packer, and Daphne Koller. Self-paced learning for latent variable models. In NeurIPS, 2010. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 245, + 287, + 278 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 245, + 287, + 278 + ], + "spans": [ + { + "bbox": [ + 48, + 245, + 287, + 278 + ], + "type": "text", + "content": "[42] Ravi Kumar Kushawaha, Saurabh Kumar, Biplab Banerjee, and Rajbabu Velmurugan. Distilling spikes: Knowledge distillation in spiking neural networks. In ICPR, 2021. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 280, + 287, + 323 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 280, + 287, + 323 + ], + "spans": [ + { + "bbox": [ + 48, + 280, + 287, + 323 + ], + "type": "text", + "content": "[43] Sergey Levine, Chelsea Finn, Trevor Darrell, and Pieter Abbeel. End-to-end training of deep visuomotor policies. The Journal of Machine Learning Research, 17(1):1334-1373, 2016. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 325, + 287, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 325, + 287, + 357 + ], + "spans": [ + { + "bbox": [ + 48, + 325, + 287, + 357 + ], + "type": "text", + "content": "[44] Dong Li, Jia-Bin Huang, Yali Li, Shengjin Wang, and Ming-Hsuan Yang. Weakly supervised object localization with progressive domain adaptation. In CVPR, 2016. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 359, + 287, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 359, + 287, + 392 + ], + "spans": [ + { + "bbox": [ + 48, + 359, + 287, + 392 + ], + "type": "text", + "content": "[45] Guohao Li, Matthias Mueller, Vincent Casser, Neil Smith, Dominik L Michels, and Bernard Ghanem. Oil: Observational imitation learning. In RSS, 2019. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 394, + 287, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 394, + 287, + 415 + ], + "spans": [ + { + "bbox": [ + 48, + 394, + 287, + 415 + ], + "type": "text", + "content": "[46] Yang Liu, Wei Zhang, and Jun Wang. Source-free domain adaptation for semantic segmentation. In CVPR, 2021. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 417, + 287, + 460 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 417, + 287, + 460 + ], + "spans": [ + { + "bbox": [ + 48, + 417, + 287, + 460 + ], + "type": "text", + "content": "[47] Takahiro Miki, Joonho Lee, Jemin Hwangbo, Lorenz Wellhausen, Vladlen Koltun, and Marco Hutter. Learning robust perceptive locomotion for quadrupedal robots in the wild. Science Robotics, 7(62), 2022. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 462, + 287, + 495 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 462, + 287, + 495 + ], + "spans": [ + { + "bbox": [ + 48, + 462, + 287, + 495 + ], + "type": "text", + "content": "[48] Matthias Müller, Alexey Dosovitskiy, Bernard Ghanem, and Vladlen Koltun. Driving policy transfer via modularity and abstraction. In CoRL, 2018. 3, 4" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 497, + 287, + 529 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 497, + 287, + 529 + ], + "spans": [ + { + "bbox": [ + 48, + 497, + 287, + 529 + ], + "type": "text", + "content": "[49] Eshed Ohn-Bar, Aditya Prakash, Aseem Behl, Kashyap Chitta, and Andreas Geiger. Learning situational driving. In CVPR, 2020. 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 531, + 287, + 574 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 531, + 287, + 574 + ], + "spans": [ + { + "bbox": [ + 48, + 531, + 287, + 574 + ], + "type": "text", + "content": "[50] Takayuki Osa, Joni Pajarinen, Gerhard Neumann, J Andrew Bagnell, Pieter Abbeel, Jan Peters, et al. An algorithmic perspective on imitation learning. Foundations and Trends® in Robotics, 2018. 2, 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 577, + 287, + 620 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 577, + 287, + 620 + ], + "spans": [ + { + "bbox": [ + 48, + 577, + 287, + 620 + ], + "type": "text", + "content": "[51] Despoina Paschalidou, Angelos Katharopoulos, Andreas Geiger, and Sanja Fidler. Neural parts: Learning expressive 3D shape abstractions with invertible neural networks. In CVPR, 2021. 6" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 622, + 287, + 654 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 622, + 287, + 654 + ], + "spans": [ + { + "bbox": [ + 48, + 622, + 287, + 654 + ], + "type": "text", + "content": "[52] Nikolaos Passalis and Anastasios Tefas. Learning deep representations with probabilistic knowledge transfer. In ECCV, 2018. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 48, + 656, + 287, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 656, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 656, + 287, + 689 + ], + "type": "text", + "content": "[53] Jonah Philion and Sanja Fidler. Lift, splat, shoot: Encoding images from arbitrary camera rigs by implicitly unprojecting to 3D. In ECCV, 2020. 2, 5" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 48, + 691, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 287, + 712 + ], + "type": "text", + "content": "[54] Dean A Pomerleau. ALVINN: An autonomous land vehicle in a neural network. In NeurIPS, 1989. 1, 2" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 37, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "text", + "content": "[55] Adriana Romero, Nicolas Ballas, Samira Ebrahimi Kahou, Antoine Chassang, Carlo Gatta, and Yoshua Bengio. Fitnets: Hints for thin deep nets. arXiv preprint arXiv:1412.6550, 2014. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 118, + 545, + 151 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 545, + 151 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 545, + 151 + ], + "type": "text", + "content": "[56] Stéphane Ross, Geoffrey Gordon, and Drew Bagnell. A reduction of imitation learning and structured prediction to no-regret online learning. In AISTATS, 2011. 2, 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 152, + 545, + 175 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 152, + 545, + 175 + ], + "spans": [ + { + "bbox": [ + 307, + 152, + 545, + 175 + ], + "type": "text", + "content": "[57] Avishkar Saha, Oscar Mendez, Chris Russell, and Richard Bowden. Translating images into maps. In ICRA, 2022. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 177, + 545, + 208 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 177, + 545, + 208 + ], + "spans": [ + { + "bbox": [ + 307, + 177, + 545, + 208 + ], + "type": "text", + "content": "[58] Axel Sauer, Nikolay Savinov, and Andreas Geiger. Conditional affordance learning for driving in urban environments. In CoRL, 2018. 4" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 210, + 545, + 232 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 210, + 545, + 232 + ], + "spans": [ + { + "bbox": [ + 307, + 210, + 545, + 232 + ], + "type": "text", + "content": "[59] Miaojing Shi and Vittorio Ferrari. Weakly supervised object localization using size estimates. In ECCV, 2016. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 234, + 545, + 256 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 234, + 545, + 256 + ], + "spans": [ + { + "bbox": [ + 307, + 234, + 545, + 256 + ], + "type": "text", + "content": "[60] Suraj Srinivas and Francois Fleuret. Knowledge transfer with Jacobian matching. In ICML, 2018. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 258, + 545, + 301 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 258, + 545, + 301 + ], + "spans": [ + { + "bbox": [ + 307, + 258, + 545, + 301 + ], + "type": "text", + "content": "[61] Yuxing Tang, Xiaosong Wang, Adam P Harrison, Le Lu, Jing Xiao, and Ronald M Summers. Attention-guided curriculum learning for weakly supervised classification and localization of thoracic diseases on chest radiographs. In MLMI, 2018. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 303, + 545, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 303, + 545, + 335 + ], + "spans": [ + { + "bbox": [ + 307, + 303, + 545, + 335 + ], + "type": "text", + "content": "[62] Marin Toromanoff, Emilie Wirbel, and Fabien Moutarde. End-to-end model-free reinforcement learning for urban driving using implicit affordances. In CVPR, 2020. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 337, + 545, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 337, + 545, + 381 + ], + "spans": [ + { + "bbox": [ + 307, + 337, + 545, + 381 + ], + "type": "text", + "content": "[63] Radu Tudor Ionescu, Bogdan Alexe, Marius Leordeanu, Marius Popescu, Dim P Papadopoulos, and Vittorio Ferrari. How hard can it be? estimating the difficulty of visual search in an image. In CVPR, 2016. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 383, + 545, + 405 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 383, + 545, + 405 + ], + "spans": [ + { + "bbox": [ + 307, + 383, + 545, + 405 + ], + "type": "text", + "content": "[64] Rachel R Van Der Stuyf. Scaffolding as a teaching strategy. Adolescent learning and development, 2002. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 407, + 545, + 439 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 407, + 545, + 439 + ], + "spans": [ + { + "bbox": [ + 307, + 407, + 545, + 439 + ], + "type": "text", + "content": "[65] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In NeurIPS, 2017. 5" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 307, + 441, + 545, + 462 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 441, + 545, + 462 + ], + "spans": [ + { + "bbox": [ + 307, + 441, + 545, + 462 + ], + "type": "text", + "content": "[66] Antonio Visioli. Practical PID control. Springer Science & Business Media, 2006. 3" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 307, + 464, + 545, + 496 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 464, + 545, + 496 + ], + "spans": [ + { + "bbox": [ + 307, + 464, + 545, + 496 + ], + "type": "text", + "content": "[67] Tao Wang, Li Yuan, Xiaopeng Zhang, and Jiashi Feng. Distilling object detectors with fine-grained feature imitation. In CVPR, 2019. 2" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 307, + 498, + 545, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 498, + 545, + 542 + ], + "spans": [ + { + "bbox": [ + 307, + 498, + 545, + 542 + ], + "type": "text", + "content": "[68] Wenhui Wang, Hangbo Bao, Shaohan Huang, Li Dong, and Furu Wei. Minilmv2: Multi-head self-attention relation distillation for compressing pretrained transformers. arXiv preprint arXiv:2012.15828, 2020. 2" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 307, + 544, + 545, + 587 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 544, + 545, + 587 + ], + "spans": [ + { + "bbox": [ + 307, + 544, + 545, + 587 + ], + "type": "text", + "content": "[69] Wei Wang, Isaac Caswell, and Ciprian Chelba. Dynamically composing domain-data selection with clean-data selection by \"co-curricular learning\" for neural machine translation. arXiv preprint arXiv:1906.01130, 2019. 2" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 307, + 589, + 545, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 589, + 545, + 622 + ], + "spans": [ + { + "bbox": [ + 307, + 589, + 545, + 622 + ], + "type": "text", + "content": "[70] Chuan Wen, Jierui Lin, Trevor Darrell, Dinesh Jayaraman, and Yang Gao. Fighting copycat agents in behavioral cloning from observation histories. In NeurIPS, 2020. 1" + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 307, + 624, + 545, + 677 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 624, + 545, + 677 + ], + "spans": [ + { + "bbox": [ + 307, + 624, + 545, + 677 + ], + "type": "text", + "content": "[71] Penghao Wu, Xiaosong Jia, Li Chen, Junchi Yan, Hongyang Li, and Yu Qiao. Trajectory-guided control prediction for end-to-end autonomous driving: A simple yet strong baseline. arXiv preprint arXiv:2206.08129, 2022. 1, 2, 3, 6, 7, 8" + } + ] + } + ], + "index": 35 + }, + { + "bbox": [ + 307, + 679, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 679, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 307, + 679, + 545, + 712 + ], + "type": "text", + "content": "[72] Liuyu Xiang, Guiguang Ding, and Jungong Han. Learning from multiple experts: Self-paced knowledge distillation for long-tailed classification. In ECCV, 2020. 2" + } + ] + } + ], + "index": 36 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7814" + } + ] + } + ], + "index": 38 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 555 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "text", + "content": "[73] Canwen Xu, Wangchunshu Zhou, Tao Ge, Furu Wei, and Ming Zhou. Bert-of-theseus: Compressing bert by progressive module replacing. arXiv preprint arXiv:2002.02925, 2020. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 118, + 287, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 118, + 287, + 150 + ], + "spans": [ + { + "bbox": [ + 49, + 118, + 287, + 150 + ], + "type": "text", + "content": "[74] Jing Yang, Brais Martinez, Adrian Bulat, and Georgios Tzimiropoulos. Knowledge distillation via adaptive instance normalization. arXiv preprint arXiv:2003.04289, 2020. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 152, + 287, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 152, + 287, + 185 + ], + "spans": [ + { + "bbox": [ + 48, + 152, + 287, + 185 + ], + "type": "text", + "content": "[75] Junho Yim, Donggyu Joo, Jihoon Bae, and Junmo Kim. A gift from knowledge distillation: Fast optimization, network minimization and transfer learning. In CVPR, 2017. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 186, + 287, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 186, + 287, + 207 + ], + "spans": [ + { + "bbox": [ + 48, + 186, + 287, + 207 + ], + "type": "text", + "content": "[76] Jimuyang Zhang and Eshed Ohn-Bar. Learning by watching. In CVPR, 2021. 2, 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 209, + 287, + 240 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 209, + 287, + 240 + ], + "spans": [ + { + "bbox": [ + 48, + 209, + 287, + 240 + ], + "type": "text", + "content": "[77] Jimuyang Zhang, Ruizhao Zhu, and Eshed Ohn-Bar. SelfD: Self-learning large-scale driving policies from the web. In CVPR, 2022. 2, 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 243, + 287, + 275 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 243, + 287, + 275 + ], + "spans": [ + { + "bbox": [ + 48, + 243, + 287, + 275 + ], + "type": "text", + "content": "[78] Pan Zhang, Bo Zhang, Ting Zhang, Dong Chen, and Fang Wen. Robust mutual learning for semi-supervised semantic segmentation. arXiv preprint arXiv:2106.00609, 2021. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 277, + 287, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 277, + 287, + 319 + ], + "spans": [ + { + "bbox": [ + 48, + 277, + 287, + 319 + ], + "type": "text", + "content": "[79] Zhejun Zhang, Alexander Liniger, Dengxin Dai, Fisher Yu, and Luc Van Gool. End-to-end urban driving by imitating a reinforcement learning coach. In ICCV, 2021. 1, 2, 3, 4, 5, 6, 7, 8" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 321, + 287, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 321, + 287, + 365 + ], + "spans": [ + { + "bbox": [ + 48, + 321, + 287, + 365 + ], + "type": "text", + "content": "[80] Albert Zhao, Tong He, Yitao Liang, Haibin Huang, Guy Van den Broeck, and Stefano Soatto. Lates: Latent space distillation for teacher-student driving policy learning. arXiv preprint arXiv:1912.02973, 2019. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 366, + 287, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 366, + 287, + 409 + ], + "spans": [ + { + "bbox": [ + 48, + 366, + 287, + 409 + ], + "type": "text", + "content": "[81] Albert Zhao, Tong He, Yitao Liang, Haibin Huang, Guy Van den Broeck, and Stefano Soatto. Sam: Squeeze-and-mimic networks for conditional visual driving policy learning. CoRL, 2020. 1, 2, 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 411, + 287, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 411, + 287, + 442 + ], + "spans": [ + { + "bbox": [ + 48, + 411, + 287, + 442 + ], + "type": "text", + "content": "[82] Brady Zhou and Philipp Krahenbuhl. Cross-view transformers for real-time map-view semantic segmentation. In CVPR, 2022. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 445, + 287, + 476 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 445, + 287, + 476 + ], + "spans": [ + { + "bbox": [ + 48, + 445, + 287, + 476 + ], + "type": "text", + "content": "[83] Brady Zhou, Philipp Krähenbuhl, and Vladlen Koltun. Does computer vision matter for action? Science Robotics, 4(30), 2019. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 478, + 287, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 478, + 287, + 511 + ], + "spans": [ + { + "bbox": [ + 48, + 478, + 287, + 511 + ], + "type": "text", + "content": "[84] Yikai Zhou, Baosong Yang, Derek F Wong, Yu Wan, and Lidia S Chao. Uncertainty-aware curriculum learning for neural machine translation. In ACL, 2020. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 512, + 287, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 512, + 287, + 555 + ], + "spans": [ + { + "bbox": [ + 48, + 512, + 287, + 555 + ], + "type": "text", + "content": "[85] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable detr: Deformable transformers for end-to-end object detection. arXiv preprint arXiv:2010.04159, 2020. 5" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7815" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/CodeTalker_ Speech-Driven 3D Facial Animation With Discrete Motion Prior/103a044f-c0a4-40fe-af3f-96d0d65c126f_content_list.json b/2023/CodeTalker_ Speech-Driven 3D Facial Animation With Discrete Motion Prior/103a044f-c0a4-40fe-af3f-96d0d65c126f_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..0fd2b159eaad1ac4889cba29b5a2f5707cf66f91 --- /dev/null +++ b/2023/CodeTalker_ Speech-Driven 3D Facial Animation With Discrete Motion Prior/103a044f-c0a4-40fe-af3f-96d0d65c126f_content_list.json @@ -0,0 +1,1465 @@ +[ + { + "type": "text", + "text": "CodeTalker: Speech-Driven 3D Facial Animation with Discrete Motion Prior", + "text_level": 1, + "bbox": [ + 98, + 130, + 870, + 151 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Jinbo Xing $^{1}$ Menghan Xia $^{2,*}$ Yuechen Zhang $^{1}$ Xiaodong Cun $^{2}$ Jue Wang $^{2}$ Tien-Tsin Wong $^{1}$ $^{1}$ The Chinese University of Hong Kong \n $^{2}$ Tencent AI Lab", + "bbox": [ + 99, + 179, + 867, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{jbxing,yczhang21,ttwong}@cse.cuhk.edu.hk", + "bbox": [ + 109, + 218, + 472, + 234 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{menghanxyz,vinthony,arphid}@gmail.com", + "bbox": [ + 521, + 219, + 857, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 268, + 313, + 285 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Speech-driven 3D facial animation has been widely studied, yet there is still a gap to achieving realism and vividness due to the highly ill-posed nature and scarcity of audiovisual data. Existing works typically formulate the cross-modal mapping into a regression task, which suffers from the regression-to-mean problem leading to over-smoothed facial motions. In this paper, we propose to cast speech-driven facial animation as a code query task in a finite proxy space of the learned codebook, which effectively promotes the vividness of the generated motions by reducing the cross-modal mapping uncertainty. The codebook is learned by self-reconstruction over real facial motions and thus embedded with realistic facial motion priors. Over the discrete motion space, a temporal autoregressive model is employed to sequentially synthesize facial motions from the input speech signal, which guarantees lip-sync as well as plausible facial expressions. We demonstrate that our approach outperforms current state-of-the-art methods both qualitatively and quantitatively. Also, a user study further justifies our superiority in perceptual quality. Code and video demo are available at https://doubiu.github.io/projects/codetalker.", + "bbox": [ + 75, + 301, + 472, + 633 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 667, + 209, + 683 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3D facial animation has been an active research topic for decades, as attributed to its broad applications in virtual reality, film production, and games. The high correlation between speech and facial gestures (especially lip movements) makes it possible to drive the facial animation with a speech signal. Early attempts are mainly made to build the complex mapping rules between phonemes and their visual counterpart, which usually have limited performance [53,63]. With the advances in deep learning, recent speech-driven facial animation techniques push forward the state-of-the-art significantly. However, it still remains challenging to generate human-like motions.", + "bbox": [ + 75, + 693, + 470, + 875 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "As an ill-posed problem, speech-driven facial animation generally has multiple plausible outputs for every input. Such ambiguity tends to cause over-smoothed results. Anyhow, person-specific approaches [29, 49] can usually obtain decent facial motions because of the relatively consistent talking style, but have low scalability to general applications. Recently, VOCA [10] extends these methods to generalize across different identities, however, they generally exhibit mild or static upper face expressions. This is because VOCA formulates the speech-to-motion mapping as a regression task, which encourages averaged motions, especially in the upper face that is only weakly or even uncorrelated to the speech signal. To reduce the uncertainty, FaceFormer [16] utilizes long-term audio context through a transformer-based model and synthesizes the sequential motions in an autoregressive manner. Although it gains important performance promotion, it still inherits the weakness of one-to-one mapping formulation and suffers from a lack of subtle high-frequency motions. Differently, MeshTalk [50] models a categorical latent space for facial animation that disentangles audio-correlated and audio-uncorrelated information so that both aspects could be well handled. Anyway, the employed quantization and categorical latent space representation are not well-suited for motion prior learning, rendering the training tricky and consequently hindering its performance.", + "bbox": [ + 500, + 270, + 893, + 662 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We get inspiration from 3D Face Morphable Model (3DMM) [35], where general facial expressions are represented in a low-dimensional space. Accordingly, we propose to formulate speech-driven facial animation as a code query task in a finite proxy space of the learned discrete codebook prior. The codebook is learned by self-reconstruction over real facial motions using a vector-quantized autoencoder (VQ-VAE) [57], which along with the decoder stores the realistic facial motion priors. In contrast to the continuous linear space of 3DMM, combinations of codebook items form a discrete prior space with only finite cardinality. Still, in the context of the decoder, the code representation possesses high expressiveness. Through mapping the speech to the finite proxy space, the uncertainty of the speech-to-motion mapping is signif", + "bbox": [ + 496, + 672, + 893, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Corresponding Author.", + "bbox": [ + 99, + 886, + 235, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "12780", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "icantly attenuated and hence promotes the quality of motion synthesis. Conceptually, the proxy space approximates the facial motion space, where the learned codebook items serve as discrete motion primitives.", + "bbox": [ + 78, + 90, + 467, + 150 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Based on the learned discrete codebook, we propose a code-query-based temporal autoregressive model for speech-conditioned facial motion synthesis, called CodeTalker. Specifically, taking a speech signal as input, our model predicts the motion feature tokens in a temporal recursive manner. Then, the feature tokens are used to query the code sequence in the discrete space, followed by facial motion reconstruction. Thanks to the contextual modeling over history motions and cross-modal alignment, the proposed CodeTalker shows the advantages of achieving accurate lip motions and natural expressions. Extensive experiments show that the proposed CodeTalker demonstrates superior performance on existing datasets. Systematic studies and experiments are conducted to demonstrate the merits of our method over previous works. The contributions of our work are as follows:", + "bbox": [ + 78, + 152, + 467, + 391 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We model the facial motion space with discrete primitives in a novel way, which offers advantages to promote motion synthesis realism against cross-modal uncertainty.", + "- We propose a discrete motion prior based temporal autoregressive model for speech-driven facial animation, which outperforms existing state-of-the-art methods." + ], + "bbox": [ + 96, + 404, + 467, + 518 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 78, + 534, + 225, + 550 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Speech-driven 3D Facial Animation", + "text_level": 1, + "bbox": [ + 78, + 559, + 383, + 574 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Computer facial animation is a long-standing task [44] and has attracted rapidly increased interest over the past decades [5, 20, 30, 32, 34, 54, 62, 70]. As a branch, speech-driven facial animation is to reenact a person in sync with input speech sequences. While extensive literature in this field works on 2D talking heads [1, 7-9, 11, 23, 26, 27, 36, 37, 43, 47, 51, 59, 61, 65-67], we focus on facial animation on 3D models in this work, which can be roughly categorized into linguistics-based and learning-based methods.", + "bbox": [ + 78, + 583, + 467, + 717 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Linguistics-based methods. Typically, linguistics-based methods [12, 40, 53, 63] establish a set of complex mapping rules between phonemes and their visual counterparts, i.e., visemes [19, 33, 41]. For example, the dominance function [40] is to determine the influence of phonemes on the respective facial animation control parameters. Xu et al. [63] defines animation curves for a constructed canonical set of visemes to generate synchronized mouth movements. There are also some methods considering the many-to-many mapping between phonemes and visemes, as demonstrated in the dynamic visemes model [53] and, more recently, the JALI [12]. Based on psycholinguistic consid", + "bbox": [ + 78, + 719, + 467, + 898 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "erations and built upon the Facial Action Coding System (FACS) [13], JALI factors mouth movements into lip and jaw rig animation and generate compelling co-articulation results. Although these methods have explicit control over the animation, they have complex procedures and lack a principled way to animate the entire face.", + "bbox": [ + 501, + 90, + 890, + 180 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Learning-based methods. Learning-based methods [6, 10, 16, 17, 24, 29, 38, 46, 52, 60] resort to a data-driven framework. Cao et al. [6] achieve emotional lip sync by the proposed constrained search and Anime Graph structure. Recently, Taylor et al. [52] propose a deep-learning-based model utilizing a sliding window approach on the transcribed phoneme sequences input. Karras et al. [29] propose a convolution-based network with a learnable emotion database to animate a speech-driven 3D mesh. More recently, VisemeNet [69] employ a three-stage Long Short-Term Memory (LSTM) network to predict the animation curve for a lower face lip model.", + "bbox": [ + 503, + 181, + 890, + 363 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We review the most related works more concretely here as they have the same setting as this work, i.e., training on high-resolution paired audio-mesh data and speaker-independently animating entire face meshes in vertex space. MeshTalk [50] successfully disentangles audio-correlated and uncorrelated facial information with a categorical latent space. However, the latent space adopted is not optimal with limited expressiveness, thus the animation quality is not stable when applied in a data-scarcity setting. VOCA [10] employs powerful audio feature extraction models and can generate facial animation with different speaking styles. Furthermore, FaceFormer [16] considers long-term audio context with transformer [58] rendering temporally stable animations. Despite the appealing animations, both suffer from the over-smoothing problem, as they directly regress the facial motion in the highly ill-posed audio-visual mapping with large uncertainty and ambiguity.", + "bbox": [ + 503, + 364, + 890, + 621 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Discrete Prior Learning", + "text_level": 1, + "bbox": [ + 503, + 633, + 718, + 648 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In the last decades, discrete prior representation with learned dictionaries has demonstrated its superiority in image restoration tasks [14, 22, 28, 55, 56], since clear image details are well-preserved in the dictionaries. This line of techniques further inspires the high-capacity and high-compressed discrete prior learning. VQ-VAE [57] first presents to learn discrete representations (codebook) of images and autoregressively model their distribution for image synthesis. The follow-up works, VQ-VAE2 [48] and VQGAN [15] further improve the quality of high-resolution image synthesis. Recently, discrete prior learning has been exploited for image colorization [25], inpainting [45], blind face restoration [68], text-to-image synthesis [21], etc.", + "bbox": [ + 503, + 657, + 890, + 853 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In addition to image modality, most recent works also explore the power of discrete prior learning in tasks with other modalities, such as dyadic face motion generation [42], co", + "bbox": [ + 503, + 854, + 890, + 898 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "12781", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "speech gesture synthesis [2], speech enhancement [64]. Inspired by codebook learning, this work investigates to learn discrete motion prior for speech-driven 3D facial animation. Different from [42], we exploit the discrete motion primitives for facial motion representation in a context-rich manner, which is more effective to learn general priors.", + "bbox": [ + 75, + 90, + 472, + 183 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 76, + 191, + 168, + 207 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We aim to synthesize sequential 3D facial motions from a speech signal, so that any neutral face mesh could be animated as a lip-synchronized talking face. However, this is an ill-posed problem since one speech could be matched by multiple potential facial animations. Such ambiguity tends to make cross-modal learning suffer from averaged motions and lack of subtle variations. To bypass this barrier, we propose to first model the facial motion space with the learned discrete motion prior, and then learn a speech-conditioned temporal autoregressive model over this space, which promotes robustness against the cross-modal uncertainty.", + "bbox": [ + 75, + 214, + 468, + 381 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Formulation. Let $\\mathbf{M}_{1:T} = (\\mathbf{m}_1, \\dots, \\mathbf{m}_T)$ be a sequence of facial motions, where each frame $\\mathbf{m}_t \\in \\mathbb{R}^{V \\times 3}$ denotes the 3D movement of $V$ vertices over a neutral-face mesh template $\\mathbf{h} \\in \\mathbb{R}^{V \\times 3}$ . Let further $\\mathbf{A}_{1:T} = (\\mathbf{a}_1, \\dots, \\mathbf{a}_T)$ be a sequence of speech snippets, each of which $\\mathbf{a}_t \\in \\mathbb{R}^d$ has $d$ samples to align with the corresponding (visual) frame $\\mathbf{m}_t$ . Then, our goal is to sequentially synthesize $\\mathbf{M}_{1:T}$ from $\\mathbf{A}_{1:T}$ so that an arbitrary neutral facial template $\\mathbf{f}$ could be animated as $\\mathbf{H}_{1:T} = \\{\\mathbf{m}_1 + \\mathbf{h}, \\dots, \\mathbf{m}_T + \\mathbf{h}\\}$ .", + "bbox": [ + 75, + 383, + 470, + 521 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Discrete Facial Motion Space", + "text_level": 1, + "bbox": [ + 76, + 527, + 336, + 542 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Visual realistic facial animations should present accurate lip motions and natural expressions. To achieve this from speech signals, extra motion priors are required to reduce the uncertainty and complement realistic motion components. As witnessed by the recent image restoration task [68], discrete codebook prior [57] demonstrates advantages in guaranteeing high-fidelity results even from a severely degraded input. Inspired by this, we propose to model the facial motion space as a discrete codebook by learning from tracked real-world facial motions.", + "bbox": [ + 75, + 550, + 468, + 700 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Codebook of motion primitives. We manage to learn a codebook $\\mathcal{Z} = \\{\\mathbf{z}_k\\in \\mathbb{R}^C\\}_{k = 1}^N$ that allows any facial motion $\\mathbf{m}_t$ to be represented by a group of allocated items $\\{\\mathbf{z}_k\\}_{k\\in S}$ , where $S$ denotes the selected index set through Eq. 1. Conceptually, the codebook items serve as the motion primitives of a facial motion space. To this end, we pre-train a transformer-based VQ-VAE that consists of an encoder $E$ , a decoder $D$ , and a context-rich codebook $\\mathcal{Z}$ , under the self-reconstruction of realistic facial motions. As shown in Figure 1, the facial motions $\\mathbf{M}_{1:T}$ is first embedded as a temporal feature $\\hat{\\mathbf{Z}} = E(\\mathbf{M}_{1:T})\\in \\mathbb{R}^{T'\\times H\\times C}$ , where $H$ is the number of face components and $T^{\\prime}$ denotes the number of encoded temporal units ( $P = \\frac{T}{T'}$ frames). Then, we", + "bbox": [ + 75, + 704, + 470, + 902 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/59d87b9d9bb2ae868b0ffb210539c242e9a29f751d08f3952822e16248b290c5.jpg", + "image_caption": [ + "Figure 1. Learning framework of facial motion space. The learned motion primitives, as embedded in the codebook, serve to represent the facial motions in a spatial and temporal manner." + ], + "image_footnote": [], + "bbox": [ + 501, + 87, + 893, + 233 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "obtain the quantized motion sequence $\\mathbf{Z}_{\\mathbf{q}} \\in \\mathbb{R}^{T' \\times H \\times C}$ via an element-wise quantization function $Q(\\cdot)$ that maps each item in $\\hat{\\mathbf{Z}}$ to its nearest entry in codebook $\\mathcal{Z}$ :", + "bbox": [ + 498, + 291, + 892, + 340 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {Z} _ {\\mathbf {q}} = Q (\\hat {\\mathbf {Z}}) := \\underset {\\mathbf {z} _ {k} \\in \\mathcal {Z}} {\\arg \\min } \\| \\hat {\\mathbf {z}} _ {t} - \\mathbf {z} _ {k} \\| _ {2}. \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 571, + 353, + 890, + 378 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Then, the self-reconstruction is given by:", + "bbox": [ + 500, + 385, + 772, + 401 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathbf {M}} _ {1: T} = D \\left(\\mathbf {Z} _ {\\mathbf {q}}\\right) = D \\left(Q \\left(E \\left(\\mathbf {M} _ {1: T}\\right)\\right)\\right). \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 566, + 411, + 890, + 429 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Note that, the discrete facial motion space reduces the mapping ambiguity with the finite cardinality, but never sacrifices its expressiveness thanks to its context-rich representation as a latent space.", + "bbox": [ + 498, + 440, + 890, + 501 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Training objectives. Similar to [57], to supervise the quantized autoencoder training, we adopt a motion-level loss and two intermediate code-level losses:", + "bbox": [ + 498, + 503, + 892, + 550 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} _ {\\mathrm {V Q}} = \\left\\| \\mathbf {M} _ {1: T} - \\hat {\\mathbf {M}} _ {1: T} \\right\\| _ {1} \\tag {3} \\\\ + \\left\\| \\operatorname {s g} (\\hat {\\mathbf {Z}}) - \\mathbf {Z} _ {\\mathbf {q}} \\right\\| _ {2} ^ {2} + \\beta \\left\\| \\hat {\\mathbf {Z}} - \\operatorname {s g} (\\mathbf {Z} _ {\\mathbf {q}}) \\right\\| _ {2} ^ {2}, \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 539, + 559, + 890, + 599 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where the first term is a reconstruction loss, the latter two are adopted to update the codebook items by reducing the distance between the codebook $\\mathcal{Z}$ and embedded features $\\hat{\\mathbf{Z}}$ . $\\mathrm{sg}(\\cdot)$ stands for a stop-gradient operation and $\\beta$ is a weighting factor controlling the update rate of the codebook and encoder. Since the quantization function (Eq. 1) is not differentiable, the straight-through gradient estimator [4,57] is employed to copy the gradients from the decoder input to the encoder output.", + "bbox": [ + 498, + 609, + 893, + 744 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Discussion. Recently, Learn2Listen [42] has applied VQ-VAE for facial expression synthesis in response to a given talking head harnessing 2D monocular videos to obtain 3DMM coefficients. In addition to distinct applications, here we would like to emphasize our major differences. First, Learn2Listen constructs speaker-specific codebooks while ours uses a generic codebook that is feasible to represent arbitrary facial motions. Since cross-character motions are absorbed, our codebook is naturally embedded with more plentiful priors. Second, Learn2Listen utilizes", + "bbox": [ + 496, + 750, + 893, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "12782", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/bb53616013f399ae8117acaf337e90ee822fc0adab0aa30f92e5648d4075cea1.jpg", + "image_caption": [ + "Figure 2. Diagram of our speech-driven motion synthesis model. Given the speech $\\mathbf{A}_{1:T}$ and style vector $\\mathbf{s}$ as input, the model learn to recursively generate a sequence of facial motions by predicting the motion codes. As embedded with well-learned motion priors, the pre-trained codebook and decoder are frozen during training." + ], + "image_footnote": [], + "bbox": [ + 78, + 89, + 890, + 281 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/2b293390f136ca97ceaf31c7bd7d135b7650aaffaea82fa6fa1abae0ffc961ed.jpg", + "image_caption": [ + "Figure 3. Concept comparison with Learn2Listen [42]. (Top) The speaker-specific facial expression coefficient prior [42], in which each code represents a sequence of facial expression coefficients. (Bottom) Our speaker-agnostic generic motion prior, in which each code represents the motion primitive of face components. The blue dotted boxes indicate what information each code may represent conceptually." + ], + "image_footnote": [], + "bbox": [ + 76, + 335, + 467, + 468 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "the codebook to represent common sequences of facial expressions by the way of 3DMM coefficients, i.e., each code represents a sequence (8 frames) of facial expressions. Differently, our codebook is formulated to represent the vertex-based facial motion space, where the codes are embedded with per-vertex motions of facial components and represent the facial motion (within a temporal unit) in a context-rich manner. As compared in Figure 3, the codebook of Learn2Listen is learned to memorize typical sequential facial expressions of a specific speaker within 3DMM space, which cannot synthesize realistic facial motions with subtle details due to the limited expressiveness of 3DMM and is bounded by the accuracy of 3D reconstruction techniques [42]. As the first attempt, our codebook is learned to represent the generic facial motion space with motion primitives for captured facial mesh data, which is more effective to embed general priors preserving vivid facial details.", + "bbox": [ + 75, + 582, + 468, + 838 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We further discuss the hyper-parameters of the codebook. First, the length of the temporal unit $P$ and the number of face components $H$ determine the complexity of the motion primitives in temporal and spatial aspects", + "bbox": [ + 75, + 839, + 468, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "respectively. Generally, complex motion primitives cause low flexibility and reusability and thus hinder representation effectiveness. On the opposite, overly simple motion primitives challenge motion prediction due to the lack of semantics. Besides, the codebook size $N$ and the feature dimension $C$ determine the representation capability, which should be defined according to the complexity of the dataset and in cooperation with $P$ and $H$ . In our experiment, we set $N = 256$ , $P = 1$ , $H = 8$ or $H = 16$ , and $C = 64$ or $C = 128$ depending on the dataset, which lead to high-quality results as justified by the ablation studies in Section 4.5. More details can be found in the Supplement.", + "bbox": [ + 496, + 335, + 890, + 517 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Speech-Driven Motion Synthesis", + "text_level": 1, + "bbox": [ + 500, + 527, + 784, + 544 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "With the learned discrete motion prior, we can build a cross-modal mapping from the input speech to the target motion codes that could be further decoded into realistic facial motions. Along with the speech, we further adopt a control on the talking styles as input, i.e., a style vector $\\mathbf{s} \\in \\mathbb{R}_+^M \\cup \\{0\\}$ , where $M$ is the dimension of the learned style space (see Eq. 4). Conditioning on the speech $\\mathbf{A}_{1:T}$ and the style vector $\\mathbf{s}$ , a temporal autoregressive model, composed of a speech encoder $E_{\\mathrm{speech}}$ and a cross-modal decoder $D_{\\mathrm{cross-modal}}$ , is employed to learn over the facial motion space, as depicted in Figure 2.", + "bbox": [ + 496, + 551, + 890, + 718 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Following FaceFormer [16], our speech encoder adopts the architecture of the state-of-the-art self-supervised pretrained speech model, wav2vec 2.0 [3], which consists of an audio feature extractor and a multi-layer transformer encoder. The audio feature extractor converts the speech of raw waveform into feature vectors through a temporal convolutions network (TCN). Benefiting from the effective attention scheme, the transformer encoder converts the audio features into contextualized speech representations. Apart from the pre-trained codebook and VQ-VAE decoder, our cross-modal decoder contains an embedding block and a multi-layer transformer decoder with causal self-attention.", + "bbox": [ + 496, + 719, + 892, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "12783", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The embedding block combines the past facial motions and the style embedding via:", + "bbox": [ + 76, + 90, + 468, + 121 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {F} _ {\\mathrm {e m b}} ^ {1: t - 1} = \\mathcal {P} _ {\\theta} \\left(\\hat {\\mathbf {M}} _ {1: t - 1}\\right) + \\mathbf {B} \\cdot \\frac {\\mathbf {S}}{\\| \\mathbf {S} \\| _ {1}}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 153, + 132, + 468, + 161 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathcal{P}_{\\theta}$ is a linear projection layer, and $\\mathbf{B} = [\\mathbf{b}_1,\\dots,\\mathbf{b}_M]\\in \\mathbb{R}^{C\\times M}$ denotes the $M$ learnable basis vectors that span the style space linearly. Alike to Face-Former [16], we equip the transformer decoder with causal self-attention to learn the dependencies between each frame in the context of the past facial motion sequence, and with cross-modal attention to align the audio and motion modalities. The output features $\\hat{\\mathbf{Z}}^{1:t}$ is further quantized into $\\mathbf{Z}_{\\mathbf{q}}^{1:t}$ via Eq. 1 and decoded by the pre-trained VQ-VAE decoder. The newly predicted motion $\\hat{\\mathbf{m}}_t$ is used to update the past motions as $\\hat{\\mathbf{M}}_{1:t}$ , in preparation for the next prediction. Formally, this recursive process can be written as:", + "bbox": [ + 76, + 171, + 468, + 353 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathbf {m}} _ {t} = D _ {\\text {c r o s s - m o d a l}} \\left(E _ {\\text {s p e e c h}} \\left(\\mathbf {A} _ {1: T}\\right), \\mathbf {s}, \\tilde {\\mathbf {M}} _ {1: t - 1}\\right). \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 122, + 364, + 468, + 383 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Training objectives. We train the transformer encoder, decoder and the embedding block for cross-modality mapping, while keeping the codebook $\\mathcal{Z}$ and motion decoder $D$ frozen. To benefit from the speech representation learning from large-scale corpora, we initialize the TCN and transformer encoder with the pre-trained wav2vec 2.0 weights. Overall, the autoregressive model is trained in a teaching-forcing scheme, under the constraint of two loss terms: (i) feature regularity loss $\\mathcal{L}_{\\mathrm{reg}}$ measuring the deviation between the predicted motion feature $\\hat{\\mathbf{Z}}^{1:T}$ and the quantized feature $\\mathbf{Z}_{\\mathbf{q}}^{1:T}$ from codebook, and (ii) motion loss $\\mathcal{L}_{\\mathrm{motion}}$ measuring the difference between the predicted motions $\\hat{\\mathbf{M}}_{1:T}$ and the ground-truth motions $\\mathbf{M}_{1:T}$ , which plays an important role to stabilize the training process. The final loss function is:", + "bbox": [ + 76, + 397, + 468, + 612 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} _ {\\text {s y n}} = \\mathcal {L} _ {\\text {r e g}} + \\mathcal {L} _ {\\text {m o t i o n}} \\\\ = \\| \\hat {\\mathbf {Z}} ^ {1: T} - \\operatorname {s g} \\left(\\mathbf {Z} _ {\\mathbf {q}} ^ {1: T}\\right) \\| _ {2} ^ {2} + \\| \\hat {\\mathbf {M}} _ {1: T} - \\mathbf {M} _ {1: T} \\| _ {2} ^ {2}. \\tag {6} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 96, + 625, + 467, + 662 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Training Details", + "text_level": 1, + "bbox": [ + 76, + 674, + 236, + 690 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "At stage one, we train the VQ-VAE model (Figure 1) on a single NVIDIA V100 for 200 epochs ( $\\sim$ 2 hours) with the AdamW [39] optimizer ( $\\beta_{1} = 0.9$ , $\\beta_{2} = 0.999$ and $\\epsilon = 1e - 8$ ), where the learning rate is initialized as $10^{-4}$ , and the mini-batch size is set as 1. At stage two, we train the temporal autoregressive model with the Adam optimizer [31]. The training duration is 100 epochs ( $\\sim$ 3 hours) and other hyper-parameters remain unchanged as stage one.", + "bbox": [ + 75, + 698, + 468, + 819 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Style embedding space. The style embedding space is linearly spanned by $M$ learned basis vectors, where each style is represented by a style vector $\\mathbf{s}$ that serves as the linear combination coefficients or a coordinate. In VOCA [10], it assigns each speaker with a category-like", + "bbox": [ + 75, + 825, + 468, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "one-hot style vector. Instead, we propose a concept to formulate a uniform style space via some learnable basis vectors (Eq. 4), where the style vector is no longer bound with speaker ID but associated with every talking sample. During training, considering the limited style diversity of training datasets, we assign each speaker (e.g. no. $i$ ) with a standard unit vector $\\mathbf{e}_i$ as a style vector, under the assumption that each speaker is associated with a unique and consistent style. Anyway, arbitrary style vectors are allowed to interpolate new talking styles during inference.", + "bbox": [ + 496, + 90, + 890, + 242 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 498, + 248, + 633, + 265 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Datasets and Implementations", + "text_level": 1, + "bbox": [ + 498, + 273, + 767, + 289 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We employ two widely used datasets, BIWI [18] and VOCASET [10], to train and test different methods in our experiments. Both datasets contain 4D face scans together with utterances spoken in English. BIWI contains 40 unique sentences shared across all speakers in the dataset, while VOCASET contains 255 unique sentences, which are partially shared among different speakers.", + "bbox": [ + 496, + 296, + 890, + 402 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "BIWI dataset. BIWI is a 3D audio-visual corpus of affective speech and facial expression in the form of dense dynamic 3D face geometries, which is originally proposed to study affective communication. There is a total of 40 sentences uttered by 14 subjects, eight females and six males. Each sentence was recorded twice: with and without emotion. On average, each sentence is 4.67 seconds long. The 3D face dynamics are captured at 25fps, each with 23370 vertices and registered topology. We follow the data splits in [16] and use the emotional subset. Specifically, the training set (BIWI-Train) contains 192 sentences, while the validation set (BIWI-Val) contains 24 sentences. There are two testing sets, in which BIWI-Test-A includes 24 sentences spoken by six seen subjects and BIWI-Test-B contains 32 sentences spoken by eight unseen subjects. BIWI-Test-A can be used for both quantitative and qualitative evaluation due to the seen subjects during training, while BIWI-Test-B is more suitable for qualitative evaluation.", + "bbox": [ + 496, + 402, + 890, + 672 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "VOCASET dataset. VOCASET is comprised of 480 paired audio-visual sequences recorded from 12 subjects. The facial motion is captured at 60fps and is about 4 seconds long. Different from BIWI, each 3D face mesh is registered to the FLAME [35] topology with 5023 vertices. We adopt the same training (VOCA-Train), validation (VOCA-Val), and testing (VOCA-Test) splits as VOCA [10] and FaceFormer for fair comparisons.", + "bbox": [ + 496, + 674, + 890, + 794 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Implementations. We compare our work with three state-of-the-art methods: VOCA [10], MeshTalk [50] and FaceFormer [16]. We train and test VOCA on BIWI using the official codebase, while directly testing the released model that was trained on VOCASET. For MeshTalk, we train and test it using the official implementation on the two datasets. To compare with FaceFormer, we conduct test-", + "bbox": [ + 496, + 795, + 890, + 898 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "12784", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/34021c8088ba812203796b2d428ee9175c8cf2d72f1678aa6d012111702556d0.jpg", + "table_caption": [ + "Table 1. Quantitative evaluation on BIWI-Test-A. Lower means better for both metrics." + ], + "table_footnote": [], + "table_body": "
MethodLip Vertex Error (×10-4mm)FDD (×10-5mm)
VOCA6.55638.1816
MeshTalk5.91815.1025
FaceFormer5.30774.6408
CodeTalker (Ours)4.79144.1170
", + "bbox": [ + 84, + 122, + 460, + 223 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "ing directly using the pre-trained weights. Among the four methods, VOCA, FaceFormer and our CodeTalker require conditioning on a training speaking style during testing. For unseen subjects, we generate facial animations conditioned on all training styles. More details about the implementation details can be found in the Supplementary Material.", + "bbox": [ + 75, + 234, + 467, + 325 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Quantitative Evaluation", + "text_level": 1, + "bbox": [ + 76, + 334, + 297, + 349 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Following MeshTalk [50] and FaceFormer [16], we adopt the lip vertex error to measure the lip synchronization, which is the only publicly proposed metric for speech-driven facial animation evaluation, to our best knowledge. As a complement, we introduce a new quantitative measurement, i.e., upper-face motion statistics, to evaluate the overall facial dynamics.", + "bbox": [ + 75, + 357, + 467, + 462 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Lip vertex error. It measures the lip deviation of a sequence with respect to the ground truth, i.e., calculating the maximal L2 error of all lip vertices for each frame and takes the average over all frames.", + "bbox": [ + 75, + 463, + 468, + 523 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Upper-face dynamics deviation. The upper-face expression is just loosely correlated with the speech, depending on personal talking styles and the semantics of speech content. With this belief, we propose to measure the variation of facial dynamics for a motion sequence in comparison with that of the ground truth. Specifically, the upper-face dynamics deviation (FDD) is calculated by:", + "bbox": [ + 75, + 523, + 468, + 630 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {F D D} \\left(\\mathbf {M} _ {1: T}, \\hat {\\mathbf {M}} _ {1: T}\\right) = \\frac {\\sum_ {v \\in \\mathcal {S} _ {U}} \\left(\\operatorname {d y n} \\left(\\mathbf {M} _ {1 : T} ^ {v}\\right) - \\operatorname {d y n} \\left(\\hat {\\mathbf {M}} _ {1 : T} ^ {v}\\right)\\right)}{\\left| \\mathcal {S} _ {U} \\right|}, \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 76, + 641, + 475, + 688 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\mathbf{M}_{1:T}^v\\in \\mathbb{R}^{3\\times T}$ denotes the motions of the $v$ -th vertex, and $S_U$ is the index set of upper-face vertices. $\\mathrm{dyn}(\\cdot)$ denotes the standard deviation of the element-wise L2 norm along the temporal axis.", + "bbox": [ + 75, + 688, + 467, + 750 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We calculate the lip vertex error and upper-face dynamics deviation (FDD) over all sequences in BIWI-Test-A and take the average for comparison. According to Table 1, the proposed CodeTalker achieves lower error than the existing state-of-the-arts, suggesting that it produces more accurate lip-synchronized movements. Besides, Table 1 shows that our CodeTalker achieves the best performance in terms of FDD. It indicates the high consistency between the predicted upper-face expressions together with the trend of facial dynamics (conditioned on the speech and talking styles)", + "bbox": [ + 75, + 750, + 468, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "and those of the ground truth.", + "bbox": [ + 500, + 90, + 696, + 106 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Qualitative Evaluation", + "text_level": 1, + "bbox": [ + 500, + 118, + 709, + 133 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We visually compare our method with other competitors in Figure 4. For fair comparison, we assign the same talking style to VOCA, FaceFormer and our CodeTalker as conditional input, which is sampled at random. To check the lip synchronization performance, we illustrate three typical frames of synthesized facial animations that speak at specific syllables, as compared in the upper partition in Figure 4. We can observe that compared with the competitors, the lip movements produced by our CodeTalker are more accurately articulated with the speech signals and also more consistent with those of the reference. For example, CodeTalker produces better lip sync with proper mouth closures when pronouncing bilabial consonant /b/ (i.e., \"bedside\" in the upper-right case of Figure 4), compared to VOCA and MeshTalk; for the even challenging speech parts \"waterproof\" and \"shaving\" that need to pout, CodeTalker can produce accurate lip shapes while other methods suffer from the over-smoothing problem and fail to lip-sync correctly (Zoom in for better inspection).", + "bbox": [ + 496, + 143, + 890, + 430 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Different from lip movements, facial expressions only have weak correlations with the speech signal, which tends to be static in front of cross-modal mapping ambiguity. To visualize the facial motion dynamics, we calculate the temporal statistics of adjacent-frame facial motions within a sequence. Specifically, we first calculate the inter-frame motion L2 distance and then compute the mean and standard deviation (std) across the sequence at each vertex. The higher mean value indicates stronger facial movements, while the higher std value suggests richer variations of facial dynamics. Two examples are visualized in the last two rows of Figure 4, evidencing that our method outperforms others in achieving both stronger facial movements and a broader range of dynamics. It is mainly attributed to the superiority of the discrete facial motion space, which promotes the robustness to cross-modal uncertainty effectively. Readers are recommended to watch the animation comparisons in the Supplemental Video.", + "bbox": [ + 496, + 431, + 890, + 703 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Talking style interpolation. Our model can synthesize new speaking styles from the learned style embedding space. To inspect the effects on VOCA-Test, we select two speaking style vectors, i.e., $\\mathbf{e}_i$ and $\\mathbf{e}_j$ , which correspond to large and slight lip articulations respectively, and interpolate new talking style vectors $\\mathbf{s}_{\\mathrm{new}} = \\mathbf{B} \\cdot [\\omega \\mathbf{e}_i + (1 - \\omega) \\mathbf{e}_j]$ with a linear coefficient $\\omega$ . For the synthesized 3D animations of a sampled sequence, we plot the lower-upper lip distances across frames for each style in Figure 5, from which we observe the smooth transition of mouth amplitudes between the two typical styles. It is not only useful to synthesize new talking styles, but also practical to match a specific speaking performance of an unseen subject during training.", + "bbox": [ + 496, + 704, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "12785", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/affde4db5730ff6cffbd1477c7f35d573f095ba6793e02af018a4ecc9ab39454.jpg", + "image_caption": [ + "Figure 4. Visual comparisons of sampled facial motions animated by different methods on VOCA-Test (left) and BIWI-Test-B (right). The upper partition shows the facial animation conditioned on different speech parts, while the lower depicts the temporal statistics (mean and standard deviation) of adjacent-frame motion variations within a sequence." + ], + "image_footnote": [], + "bbox": [ + 78, + 85, + 501, + 435 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/8c26e7797a3bfe4ec4341c4bacd29f14d0d1a53b8b130bc088cb5d32ddf9fedf.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 501, + 85, + 893, + 435 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/4b04dc2ffc959720fb6cf189f9e133b644298f5d20e524c9b1c62ec97ba874a5.jpg", + "image_caption": [ + "Figure 5. Distance between lower and upper lip for our predictions within a sequence conditioned on different weighted linear combinations of two style vectors." + ], + "image_footnote": [], + "bbox": [ + 78, + 496, + 472, + 642 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. User Study", + "text_level": 1, + "bbox": [ + 76, + 696, + 199, + 712 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The human perception system has been evolutionarily adapted to understanding subtle facial motions and capturing lip synchronization. Thus, it is still the most reliable measure in the speech-driven facial animation task. We conduct a user study to evaluate the quality of animated faces in perceptual lip synchronization and realism, compared with VOCA, MeshTalk, FaceFormer and the ground truth. We adopt A/B tests for each comparison, i.e., ours vs. competitor, in terms of realistic facial animation and lip sync. For BIWI, we obtain the results of four kinds of comparisons by randomly selecting 30 samples from BIWI-Test-B, respectively. To achieve the most variations in terms of speaking", + "bbox": [ + 75, + 719, + 470, + 902 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/9ee9e9789b0556ed0fa1c16cb2099f45fd94f59d3f12765b51af9ef63d589c46.jpg", + "table_caption": [ + "Table 2. User study results on BIWI-Test-B and VOCA-Test. We adopt A/B testing and report the percentage of answers where A is preferred over B." + ], + "table_footnote": [], + "table_body": "
CompetitorsBIWI-Test-BVOCA-Test
Lip SyncRealismLip SyncRealism
Ours vs. VOCA92.4789.2586.0284.95
Ours vs. MeshTalk80.6582.8095.7092.47
Ours vs. FaceFormer53.7656.9970.9769.89
Ours vs. GT43.0149.4643.0143.01
", + "bbox": [ + 503, + 542, + 890, + 648 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "styles, we ensure the sampling results can fairly cover all conditioning styles. Thus, $120\\mathrm{A}$ vs. B pairs (30 samples $\\times 4$ comparisons) are created for BIWI-Test-B. Each pair is judged by at least 3 different participants separately, and finally, 372 entries are collected in total. For the user study on VOCASET, we apply the same setting as that on the BIWI dataset, i.e., another $120\\mathrm{A}$ vs. B pairs from VOCA-Test set, finally yielding 372 entries as well. In this study, 31 participants with good vision and hearing ability complete the evaluation successfully. Moreover, each participant is involved in all 8 kinds of comparisons to make better exposure and cover the diversity of favorability.", + "bbox": [ + 496, + 657, + 890, + 839 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The percentage of A/B testing in terms of lip sync and realism on BIWI-Test-B is tabulated in Table 2, which shows that participants favor CodeTalker over competitors. Based on the visual analysis in Section 4.3, we attribute this to the", + "bbox": [ + 496, + 839, + 893, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "12786", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/0046cc2055500718bc9f967b44586cc7e21613cf2c49093fdb435c8a9453048b.jpg", + "table_caption": [ + "Table 3. Ablation study on the representation space of codebook. The performance is measured by the reconstruction error (i.e., L2 error) and lip vertex error on VOCA-Test and BIWI-Test-A." + ], + "table_footnote": [], + "table_body": "
VariantsVOCA-TestBIWI-Test-A
Rec. Error (×10-5mm)Rec. Error (×10-5mm)Lip Vertex Error (×10-4mm)
Shape-ent. codebook2.754.076.41
Motion codebook (Ours)0.082.834.79
Speech-driven synthesis"finding Nan's cameo""an excellent room"
Shape-ent. codebookMotion (Ours) codebookShape-ent. codebookMotion (Ours) codebook
", + "bbox": [ + 78, + 135, + 468, + 388 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Figure 6. Visual comparisons of reconstruction and speech-driven motion synthesis results with different representation spaces on VOCA-Test (left) and BIWI-Test-A (right).", + "bbox": [ + 76, + 396, + 468, + 438 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "facial animation synthesized by CodeTalker having more expressive facial motions, accurate lip shape, and well-synchronized mouth movements. For VOCA-Test, which has a nature of fewer upper-face motions, a similar favorability can still be observed in Table 2. We believe the reasons are at least three-fold: subtle motions around the eyes, more accurate lip movements and expressive motions in the lower face. Although be aware of a gap between our predictions and the recorded performance (ground truth), we surprisingly get over $40\\%$ preference still. Overall, the user study justifies that the facial animations produced by CodeTalker have superior perceptual quality.", + "bbox": [ + 75, + 445, + 468, + 627 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.5. Ablation Studies", + "text_level": 1, + "bbox": [ + 76, + 635, + 240, + 648 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We study several key designs of our proposed method in this section, including the representation space and the hyper-parameters of the codebook construction.", + "bbox": [ + 75, + 659, + 468, + 703 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Representation space. To study the superiority of our motion-based representation, we construct a baseline that learns a shape-entangled codebook, i.e., the codes represent shapes $(\\mathbf{m_t} + \\mathbf{h})$ instead of motions $(\\mathbf{m_t})$ . As shown in Table 3, the baseline decreases the reconstruction accuracy significantly, which is evidenced by the visualized examples in Figure 6. It is mainly because the shape-entangled sequence contains more speaker-specific information that hinders the reusability of the codes. A direct weakness is the poor generalization for self-reconstruction, which further impedes cross-modal mapping correctness. In contrast, our proposed speaker-agnostic motion representation is more effective to represent generic motion priors shared", + "bbox": [ + 75, + 704, + 468, + 900 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/4ed4c8fd48debac4bccf63b2118f8fae2e9993cf3b6daa8e533b8e563687963c.jpg", + "image_caption": [ + "Figure 7. Model performance comparisons with different hyperparameters of the codebook, i.e., the length of temporal unit $P$ and the number of face components $H$ . We measure the reconstruction error and lip vertex error on BIWI-Test-A." + ], + "image_footnote": [], + "bbox": [ + 500, + 88, + 694, + 223 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/e0030efa375101435fceb90c2527f23521101929501c92ed6062e7bc3d39dedf.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 699, + 88, + 893, + 223 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "across individuals, and hence promotes the quality of both self-reconstruction and speech-driven motion synthesis.", + "bbox": [ + 496, + 297, + 890, + 328 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Codebook construction. We further study the hyperparameters used for codebook construction. We evaluate the performance of different settings $\\langle P, H \\rangle$ by measuring their reconstruction accuracy and cross-modal mapping accuracy (namely lip vertex error). First, we evaluate the reconstruction accuracy as shown in Figure 7(a). On one hand, increasing $P$ degrades the reconstruction accuracy, which could be explained by the increased complexity of the motion to be represented. On the other hand, increasing $H$ eases the reconstruction but risks over-fitting, which explains the general benefits $(H < 8)$ but inferior performance when $H \\geq 8$ . Notably, a similar trend could be found in the cross-modal mapping performance, as shown in Figure 7(b). We conjecture that complex motion primitives cause lower reusability and higher redundancy, resulting in ambiguity in the cross-modal code query process.", + "bbox": [ + 496, + 329, + 892, + 570 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Discussion and Conclusion", + "text_level": 1, + "bbox": [ + 500, + 584, + 746, + 599 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We demonstrated the advantages of casting speech-driven facial animation as a code query task in the discrete space, which notably promotes the motion synthesis quality against cross-modal ambiguity. By comparing to the existing state-of-the-arts, our proposed method shows superiority in achieving accurate lip sync and vivid facial expressions. However, we still follow the assumption that facial motions are independent of shapes, whose rationality may deserve further studies. Also, the overall perceptual quality still lags behind the ground truth, primarily due to the scarcity of paired audio-visual data. Additionally, the acquired generic motion prior adheres to the motion distribution delineated by the training set, which may deviate from real-world facial motions. As a future work, it is intriguing to guide the 3D facial animation by utilizing priors from large-scale available talking head videos.", + "bbox": [ + 496, + 609, + 890, + 849 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements. This project is partially supported by Hong Kong Innovation and Technology Fund (ITF) (ref: ITS/307/20FP and ITS/313/20).", + "bbox": [ + 496, + 854, + 890, + 898 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "12787", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 174, + 106 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Mohammed M Alghamdi, He Wang, Andrew J Bulpitt, and David C Hogg. Talking head from speech audio using a pretrained image generator. In ACM International Conference on Multimedia (MM), 2022. 2", + "[2] Tenglong Ao, Qingzhe Gao, Yuke Lou, Baoquan Chen, and Libin Liu. Rhythmic gesticulator: Rhythm-aware co-speech gesture synthesis with hierarchical neural embeddings. ACM Transactions on Graphics (TOG), pages 1-19, 2022. 3", + "[3] Alexei Baevski, Yuhao Zhou, Abdelrahman Mohamed, and Michael Auli. wav2vec 2.0: A framework for self-supervised learning of speech representations. In Conference on Neural Information Processing Systems (NeurIPS), 2020. 4", + "[4] Yoshua Bengio, Nicholas Léonard, and Aaron Courville. Estimating or propagating gradients through stochastic neurons for conditional computation. arXiv preprint arXiv:1308.3432, 2013. 3", + "[5] Chen Cao, Hongzhi Wu, Yanlin Weng, Tianjia Shao, and Kun Zhou. Real-time facial animation with image-based dynamic avatars. ACM Transactions on Graphics (TOG), 35(4):1-12, 2016. 2", + "[6] Yong Cao, Wen C Tien, Petros Faloutsos, and Frédéric Pighin. Expressive speech-driven facial animation. ACM Transactions on Graphics (TOG), 24(4):1283-1302, 2005. 2", + "[7] Lele Chen, Guofeng Cui, Celong Liu, Zhong Li, Ziyi Kou, Yi Xu, and Chenliang Xu. Talking-head generation with rhythmic head motion. In European Conference on Computer Vision (ECCV), 2020. 2", + "[8] Lele Chen, Zhiheng Li, Ross K Maddox, Zhiyao Duan, and Chenliang Xu. Lip movements generation at a glance. In European Conference on Computer Vision (ECCV), 2018. 2", + "[9] Joon Son Chung and Andrew Zisserman. Out of time: automated lip sync in the wild. In Asian Conference on Computer Vision (ACCV), 2016. 2", + "[10] Daniel Cudeiro, Timo Bolkart, Cassidy Laidlaw, Anurag Ranjan, and Michael J Black. Capture, learning, and synthesis of 3d speaking styles. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 1, 2, 5", + "[11] Dipanjan Das, Sandika Biswas, Sanjana Sinha, and Brojeshwar Bhowmick. Speech-driven facial animation using cascaded gans for learning of motion and texture. In European Conference on Computer Vision (ECCV), 2020. 2", + "[12] Pif Edwards, Chris Landreth, Eugene Fume, and Karan Singh. Jali: an animator-centric viseme model for expressive lip synchronization. ACM Transactions on Graphics (TOG), 35(4):1-11, 2016. 2", + "[13] Paul Ekman and Wallace V Friesen. Facial action coding system. Environmental Psychology & Nonverbal Behavior, 1978. 2", + "[14] Michael Elad and Michal Aharon. Image denoising via learned dictionaries and sparse representation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2006. 2", + "[15] Patrick Esser, Robin Rombach, and Bjorn Ommer. Taming transformers for high-resolution image synthesis. In" + ], + "bbox": [ + 78, + 114, + 470, + 902 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 2", + "[16] Yingruo Fan, Zhaojiang Lin, Jun Saito, Wenping Wang, and Taku Komura. Faceformer: Speech-driven 3d facial animation with transformers. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 1, 2, 4, 5, 6", + "[17] Yingruo Fan, Zhaojiang Lin, Jun Saito, Wenping Wang, and Taku Komura. Joint audio-text model for expressive speech-driven 3d facial animation. In ACM Symposium on Interactive 3D Graphics and Games (i3D), 2022. 2", + "[18] Gabriele Fanelli, Juergen Gall, Harald Romsdorfer, Thibaut Weise, and Luc Van Gool. A 3-d audio-visual corpus of affective communication. IEEE Transactions on Multimedia (TMM), 12(6):591-598, 2010. 5", + "[19] Cletus G Fisher. Confusions among visually perceived consonants. Journal of speech and hearing research, 11(4):796-804, 1968. 2", + "[20] Ohad Fried, Ayush Tewari, Michael Zollhöfer, Adam Finkelstein, Eli Shechtman, Dan B Goldman, Kyle Genova, Zeyu Jin, Christian Theobalt, and Maneesh Agrawala. Text-based editing of talking-head video. ACM Transactions on Graphics (TOG), 38(4):1-14, 2019. 2", + "[21] Shuyang Gu, Dong Chen, Jianmin Bao, Fang Wen, Bo Zhang, Dongdong Chen, Lu Yuan, and Baining Guo. Vector quantized diffusion model for text-to-image synthesis. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2", + "[22] Shuhang Gu, Wangmeng Zuo, Qi Xie, Deyu Meng, Xiangchu Feng, and Lei Zhang. Convolutional sparse coding for image super-resolution. In IEEE/CVF International Conference on Computer Vision (ICCV), 2015. 2", + "[23] Yudong Guo, Keyu Chen, Sen Liang, Yong-Jin Liu, Hujun Bao, and Juyong Zhang. Ad-nerf: Audio driven neural radiance fields for talking head synthesis. In IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 2", + "[24] Ikhsanul Habibie, Weipeng Xu, Dushyant Mehta, Lingjie Liu, Hans-Peter Seidel, Gerard Pons-Moll, Mohamed Elgharib, and Christian Theobalt. Learning speech-driven 3d conversational gestures from video. In ACM International Conference on Intelligent Virtual Agents (IVA), 2021. 2", + "[25] Zhitong Huang, Nanxuan Zhao, and Jing Liao. Unicolor: A unified framework for multi-modal colorization with transformer. ACM Transactions on Graphics (TOG), pages 1-16, 2022. 2", + "[26] Xinya Ji, Hang Zhou, Kaisiyuan Wang, Qianyi Wu, Wayne Wu, Feng Xu, and Xun Cao. Eamm: One-shot emotional talking face via audio-based emotion-aware motion model. In ACM SIGGRAPH Conference, 2022. 2", + "[27] Xinya Ji, Hang Zhou, Kaisiyuan Wang, Wayne Wu, Chen Change Loy, Xun Cao, and Feng Xu. Audio-driven emotional video portraits. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 2", + "[28] Younghyun Jo and Seon Joo Kim. Practical single-image super-resolution using look-up table. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 2" + ], + "bbox": [ + 501, + 92, + 893, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "12788", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[29] Tero Karras, Timo Aila, Samuli Laine, Antti Herva, and Jaakko Lehtinen. Audio-driven facial animation by joint end-to-end learning of pose and emotion. ACM Transactions on Graphics (TOG), 36(4):1-12, 2017. 1, 2", + "[30] Hyeongwoo Kim, Pablo Garrido, Ayush Tewari, Weipeng Xu, Justus Thies, Matthias Niessner, Patrick Pérez, Christian Richardt, Michael Zollhöfer, and Christian Theobalt. Deep video portraits. ACM Transactions on Graphics (TOG), 37(4):1-14, 2018. 2", + "[31] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In International Conference on Learning Representations (ICLR), 2014. 5", + "[32] Avisek Lahiri, Vivek Kwatra, Christian Frueh, John Lewis, and Chris Bregler. Lipsync3d: Data-efficient learning of personalized 3d talking faces from video using pose and lighting normalization. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 2", + "[33] John Lewis. Automated lip-sync: Background and techniques. The Journal of Visualization and Computer Animation, 2(4):118-122, 1991. 2", + "[34] Hao Li, Jihun Yu, Yuting Ye, and Chris Bregler. Realtime facial animation with on-the-fly correctives. ACM Transactions on Graphics (TOG), 32(4):42-1, 2013. 2", + "[35] Tianye Li, Timo Bolkart, Michael J Black, Hao Li, and Javier Romero. Learning a model of facial shape and expression from 4d scans. ACM Transactions on Graphics (TOG), 36(6):194-1, 2017. 1, 5", + "[36] Borong Liang, Yan Pan, Zhizhi Guo, Hang Zhou, Zhibin Hong, Xiaoguang Han, Junyu Han, Jingtuo Liu, Errui Ding, and Jingdong Wang. Expressive talking head generation with granular audio-visual control. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2", + "[37] Xian Liu, Yinghao Xu, Qianyi Wu, Hang Zhou, Wayne Wu, and Bolei Zhou. Semantic-aware implicit neural auditory-driven video portrait generation. In European Conference on Computer Vision (ECCV), 2022. 2", + "[38] Yilong Liu, Feng Xu, Jinxiang Chai, Xin Tong, Lijuan Wang, and Qiang Huo. Video-audio driven real-time facial animation. ACM Transactions on Graphics (TOG), 34(6):1-10, 2015. 2", + "[39] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In International Conference on Learning Representations (ICLR), 2019. 5", + "[40] DW Massaro, MM Cohen, M Tabain, J Beskow, and R Clark. Animated speech: research progress and applications. Audiovisual Speech Processing, pages 309-345, 2012. 2", + "[41] Wesley Mattheyses and Werner Verhelst. Audiovisual speech synthesis: An overview of the state-of-the-art. Speech Communication, 66:182-217, 2015. 2", + "[42] Evonne Ng, Hanbyul Joo, Liwen Hu, Hao Li, Trevor Darrell, Angjoo Kanazawa, and Shiry Ginosar. Learning to listen: Modeling non-deterministic dyadic facial motion. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2, 3, 4", + "[43] Youxin Pang, Yong Zhang, Weize Quan, Yanbo Fan, Xiaodong Cun, Ying Shan, and Dong-ming Yan. Dpe: Disentanglement of pose and expression for general video portrait editing. arXiv preprint arXiv:2301.06281, 2023. 2" + ], + "bbox": [ + 78, + 90, + 468, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[44] Frederic I Parke and Keith Waters. Computer facial animation. CRC press, 2008. 2", + "[45] Jialun Peng, Dong Liu, Songcen Xu, and Houqiang Li. Generating diverse structure for image inpainting with hierarchical vq-vae. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 2", + "[46] Hai Xuan Pham, Yuting Wang, and Vladimir Pavlovic. End-to-end learning for 3d facial animation from speech. In ACM International Conference on Multimodal Interaction (ICMI), 2018. 2", + "[47] KR Prajwal, Rudrabha Mukhopadhyay, Vinay P Namboodiri, and CV Jawahar. A lip sync expert is all you need for speech to lip generation in the wild. In ACM International Conference on Multimedia (MM), 2020. 2", + "[48] Ali Razavi, Aaron Van den Oord, and Oriol Vinyals. Generating diverse high-fidelity images with vq-vae-2. In Conference on Neural Information Processing Systems (NeurIPS), 2019. 2", + "[49] Alexander Richard, Colin Lea, Shugao Ma, Jurgen Gall, Fernando De la Torre, and Yaser Sheikh. Audio-and gaze-driven facial animation of codec avatars. In IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), 2021. 1", + "[50] Alexander Richard, Michael Zollhöfer, Yandong Wen, Fernando De la Torre, and Yaser Sheikh. Meshtalk: 3d face animation from speech using cross-modality disentanglement. In IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 1, 2, 5, 6", + "[51] Shuai Shen, Wanhua Li, Zheng Zhu, Yueqi Duan, Jie Zhou, and Jiwen Lu. Learning dynamic facial radiance fields for few-shot talking head synthesis. In European Conference on Computer Vision (ECCV), 2022. 2", + "[52] Sarah Taylor, Taehwan Kim, Yisong Yue, Moshe Mahler, James Krahe, Anastasio Garcia Rodriguez, Jessica Hodgins, and Iain Matthews. A deep learning approach for generalized speech animation. ACM Transactions on Graphics (TOG), 36(4):1-11, 2017. 2", + "[53] Sarah L Taylor, Moshe Mahler, Barry-John Theobald, and Iain Matthews. Dynamic units of visual speech. In ACM/Eurographics Symposium on Computer Animation (SCA), 2012. 1, 2", + "[54] Justus Thies, Mohamed Elgharib, Ayush Tewari, Christian Theobalt, and Matthias Nießner. Neural voice puppetry: Audio-driven facial reenactment. In European Conference on Computer Vision (ECCV), 2020. 2", + "[55] Radu Timofte, Vincent De Smet, and Luc Van Gool. Anchored neighborhood regression for fast example-based super-resolution. In IEEE/CVF International Conference on Computer Vision (ICCV), 2013. 2", + "[56] Radu Timofte, Vincent De Smet, and Luc Van Gool. A+: Adjusted anchored neighborhood regression for fast super-resolution. In Asian Conference on Computer Vision (ACCV), 2014. 2", + "[57] Aaron Van Den Oord, Oriol Vinyals, et al. Neural discrete representation learning. Conference on Neural Information Processing Systems (NeurIPS), 2017. 1, 2, 3", + "[58] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia" + ], + "bbox": [ + 501, + 92, + 890, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "12789", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Polosukhin. Attention is all you need. Conference on Neural Information Processing Systems (NeurIPS), 2017. 2", + "[59] Konstantinos Vougioukas, Stavros Petridis, and Maja Pantic. Realistic speech-driven facial animation with gans. International Journal of Computer Vision (IJCV), 128(5):1398-1413, 2020. 2", + "[60] Qianyun Wang, Zhenfeng Fan, and Shihong Xia. 3d-talkemo: Learning to synthesize 3d emotional talking head. arXiv preprint arXiv:2104.12051, 2021. 2", + "[61] Suzhen Wang, Lincheng Li, Yu Ding, and Xin Yu. One-shot talking face generation from single-speaker audio-visual correlation learning. In AAAI Conference on Artificial Intelligence (AAAI), 2022. 2", + "[62] Thibaut Weise, Sofien Bouaziz, Hao Li, and Mark Pauly. Realtime performance-based facial animation. ACM Transactions on Graphics (TOG), 30(4):1-10, 2011. 2", + "[63] Yuyu Xu, Andrew W Feng, Stacy Marsella, and Ari Shapiro. A practical and configurable lip sync method for games. In Proceedings of Motion on Games, 2013. 1, 2", + "[64] Karren Yang, Dejan Markovic, Steven Krenn, Vasu Agrawal, and Alexander Richard. Audio-visual speech codes: Rethinking audio-visual speech enhancement by re-synthesis. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 3", + "[65] Ran Yi, Zipeng Ye, Juyong Zhang, Hujun Bao, and Yong-Jin Liu. Audio-driven talking face video generation with learning-based personalized head pose. arXiv preprint arXiv:2002.10137, 2020. 2", + "[66] Bowen Zhang, Chenyang Qi, Pan Zhang, Bo Zhang, Hsiang-Tao Wu, Dong Chen, Qifeng Chen, Yong Wang, and Fang Wen. Metaportrait: Identity-preserving talking head generation with fast personalized adaptation. arXiv preprint arXiv:2212.08062, 2022. 2", + "[67] Hang Zhou, Yasheng Sun, Wayne Wu, Chen Change Loy, Xiaogang Wang, and Ziwei Liu. Pose-controllable talking face generation by implicitly modularized audio-visual representation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 2", + "[68] Shangchen Zhou, Kelvin CK Chan, Chongyi Li, and Chen Change Loy. Towards robust blind face restoration with codebook lookup transformer. Conference on Neural Information Processing Systems (NeurIPS), 2022. 2, 3", + "[69] Yang Zhou, Zhan Xu, Chris Landreth, Evangelos Kalogerakis, Subhransu Maji, and Karan Singh. Visemenet: Audio-driven animator-centric speech animation. ACM Transactions on Graphics (TOG), 37(4):1-10, 2018. 2", + "[70] Michael Zollhöfer, Justus Thies, Pablo Garrido, Derek Bradley, Thabo Beeler, Patrick Pérez, Marc Stamminger, Matthias Nießner, and Christian Theobalt. State of the art on monocular 3d face reconstruction, tracking, and applications. Computer Graphics Forum (CGF), 37(2):523-550, 2018. 2" + ], + "bbox": [ + 78, + 90, + 468, + 811 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "12790", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2023/CodeTalker_ Speech-Driven 3D Facial Animation With Discrete Motion Prior/103a044f-c0a4-40fe-af3f-96d0d65c126f_model.json b/2023/CodeTalker_ Speech-Driven 3D Facial Animation With Discrete Motion Prior/103a044f-c0a4-40fe-af3f-96d0d65c126f_model.json new file mode 100644 index 0000000000000000000000000000000000000000..77f99bea5adef60abe99e7993ae955b0f737f3e8 --- /dev/null +++ b/2023/CodeTalker_ Speech-Driven 3D Facial Animation With Discrete Motion Prior/103a044f-c0a4-40fe-af3f-96d0d65c126f_model.json @@ -0,0 +1,2246 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.099, + 0.131, + 0.872, + 0.152 + ], + "angle": 0, + "content": "CodeTalker: Speech-Driven 3D Facial Animation with Discrete Motion Prior" + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.18, + 0.868, + 0.217 + ], + "angle": 0, + "content": "Jinbo Xing\\(^{1}\\) Menghan Xia\\(^{2,*}\\) Yuechen Zhang\\(^{1}\\) Xiaodong Cun\\(^{2}\\) Jue Wang\\(^{2}\\) Tien-Tsin Wong\\(^{1}\\) \n\\(^{1}\\)The Chinese University of Hong Kong \n\\(^{2}\\)Tencent AI Lab" + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.219, + 0.473, + 0.235 + ], + "angle": 0, + "content": "{jbxing,yczhang21,ttwong}@cse.cuhk.edu.hk" + }, + { + "type": "text", + "bbox": [ + 0.522, + 0.22, + 0.858, + 0.234 + ], + "angle": 0, + "content": "{menghanxyz,vinthony,arphid}@gmail.com" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.269, + 0.314, + 0.286 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.302, + 0.473, + 0.635 + ], + "angle": 0, + "content": "Speech-driven 3D facial animation has been widely studied, yet there is still a gap to achieving realism and vividness due to the highly ill-posed nature and scarcity of audiovisual data. Existing works typically formulate the cross-modal mapping into a regression task, which suffers from the regression-to-mean problem leading to over-smoothed facial motions. In this paper, we propose to cast speech-driven facial animation as a code query task in a finite proxy space of the learned codebook, which effectively promotes the vividness of the generated motions by reducing the cross-modal mapping uncertainty. The codebook is learned by self-reconstruction over real facial motions and thus embedded with realistic facial motion priors. Over the discrete motion space, a temporal autoregressive model is employed to sequentially synthesize facial motions from the input speech signal, which guarantees lip-sync as well as plausible facial expressions. We demonstrate that our approach outperforms current state-of-the-art methods both qualitatively and quantitatively. Also, a user study further justifies our superiority in perceptual quality. Code and video demo are available at https://doubiu.github.io/projects/codetalker." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.668, + 0.21, + 0.684 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.694, + 0.471, + 0.875 + ], + "angle": 0, + "content": "3D facial animation has been an active research topic for decades, as attributed to its broad applications in virtual reality, film production, and games. The high correlation between speech and facial gestures (especially lip movements) makes it possible to drive the facial animation with a speech signal. Early attempts are mainly made to build the complex mapping rules between phonemes and their visual counterpart, which usually have limited performance [53,63]. With the advances in deep learning, recent speech-driven facial animation techniques push forward the state-of-the-art significantly. However, it still remains challenging to generate human-like motions." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.271, + 0.895, + 0.664 + ], + "angle": 0, + "content": "As an ill-posed problem, speech-driven facial animation generally has multiple plausible outputs for every input. Such ambiguity tends to cause over-smoothed results. Anyhow, person-specific approaches [29, 49] can usually obtain decent facial motions because of the relatively consistent talking style, but have low scalability to general applications. Recently, VOCA [10] extends these methods to generalize across different identities, however, they generally exhibit mild or static upper face expressions. This is because VOCA formulates the speech-to-motion mapping as a regression task, which encourages averaged motions, especially in the upper face that is only weakly or even uncorrelated to the speech signal. To reduce the uncertainty, FaceFormer [16] utilizes long-term audio context through a transformer-based model and synthesizes the sequential motions in an autoregressive manner. Although it gains important performance promotion, it still inherits the weakness of one-to-one mapping formulation and suffers from a lack of subtle high-frequency motions. Differently, MeshTalk [50] models a categorical latent space for facial animation that disentangles audio-correlated and audio-uncorrelated information so that both aspects could be well handled. Anyway, the employed quantization and categorical latent space representation are not well-suited for motion prior learning, rendering the training tricky and consequently hindering its performance." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.674, + 0.895, + 0.903 + ], + "angle": 0, + "content": "We get inspiration from 3D Face Morphable Model (3DMM) [35], where general facial expressions are represented in a low-dimensional space. Accordingly, we propose to formulate speech-driven facial animation as a code query task in a finite proxy space of the learned discrete codebook prior. The codebook is learned by self-reconstruction over real facial motions using a vector-quantized autoencoder (VQ-VAE) [57], which along with the decoder stores the realistic facial motion priors. In contrast to the continuous linear space of 3DMM, combinations of codebook items form a discrete prior space with only finite cardinality. Still, in the context of the decoder, the code representation possesses high expressiveness. Through mapping the speech to the finite proxy space, the uncertainty of the speech-to-motion mapping is signif" + }, + { + "type": "page_footnote", + "bbox": [ + 0.101, + 0.887, + 0.236, + 0.901 + ], + "angle": 0, + "content": "* Corresponding Author." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "12780" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.08, + 0.092, + 0.468, + 0.151 + ], + "angle": 0, + "content": "icantly attenuated and hence promotes the quality of motion synthesis. Conceptually, the proxy space approximates the facial motion space, where the learned codebook items serve as discrete motion primitives." + }, + { + "type": "text", + "bbox": [ + 0.08, + 0.153, + 0.468, + 0.392 + ], + "angle": 0, + "content": "Based on the learned discrete codebook, we propose a code-query-based temporal autoregressive model for speech-conditioned facial motion synthesis, called CodeTalker. Specifically, taking a speech signal as input, our model predicts the motion feature tokens in a temporal recursive manner. Then, the feature tokens are used to query the code sequence in the discrete space, followed by facial motion reconstruction. Thanks to the contextual modeling over history motions and cross-modal alignment, the proposed CodeTalker shows the advantages of achieving accurate lip motions and natural expressions. Extensive experiments show that the proposed CodeTalker demonstrates superior performance on existing datasets. Systematic studies and experiments are conducted to demonstrate the merits of our method over previous works. The contributions of our work are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.405, + 0.468, + 0.464 + ], + "angle": 0, + "content": "- We model the facial motion space with discrete primitives in a novel way, which offers advantages to promote motion synthesis realism against cross-modal uncertainty." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.476, + 0.468, + 0.519 + ], + "angle": 0, + "content": "- We propose a discrete motion prior based temporal autoregressive model for speech-driven facial animation, which outperforms existing state-of-the-art methods." + }, + { + "type": "list", + "bbox": [ + 0.097, + 0.405, + 0.468, + 0.519 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.08, + 0.535, + 0.226, + 0.551 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "title", + "bbox": [ + 0.08, + 0.56, + 0.385, + 0.575 + ], + "angle": 0, + "content": "2.1. Speech-driven 3D Facial Animation" + }, + { + "type": "text", + "bbox": [ + 0.08, + 0.584, + 0.468, + 0.718 + ], + "angle": 0, + "content": "Computer facial animation is a long-standing task [44] and has attracted rapidly increased interest over the past decades [5, 20, 30, 32, 34, 54, 62, 70]. As a branch, speech-driven facial animation is to reenact a person in sync with input speech sequences. While extensive literature in this field works on 2D talking heads [1, 7-9, 11, 23, 26, 27, 36, 37, 43, 47, 51, 59, 61, 65-67], we focus on facial animation on 3D models in this work, which can be roughly categorized into linguistics-based and learning-based methods." + }, + { + "type": "text", + "bbox": [ + 0.08, + 0.72, + 0.468, + 0.9 + ], + "angle": 0, + "content": "Linguistics-based methods. Typically, linguistics-based methods [12, 40, 53, 63] establish a set of complex mapping rules between phonemes and their visual counterparts, i.e., visemes [19, 33, 41]. For example, the dominance function [40] is to determine the influence of phonemes on the respective facial animation control parameters. Xu et al. [63] defines animation curves for a constructed canonical set of visemes to generate synchronized mouth movements. There are also some methods considering the many-to-many mapping between phonemes and visemes, as demonstrated in the dynamic visemes model [53] and, more recently, the JALI [12]. Based on psycholinguistic consid" + }, + { + "type": "text", + "bbox": [ + 0.503, + 0.092, + 0.891, + 0.181 + ], + "angle": 0, + "content": "erations and built upon the Facial Action Coding System (FACS) [13], JALI factors mouth movements into lip and jaw rig animation and generate compelling co-articulation results. Although these methods have explicit control over the animation, they have complex procedures and lack a principled way to animate the entire face." + }, + { + "type": "text", + "bbox": [ + 0.504, + 0.183, + 0.891, + 0.364 + ], + "angle": 0, + "content": "Learning-based methods. Learning-based methods [6, 10, 16, 17, 24, 29, 38, 46, 52, 60] resort to a data-driven framework. Cao et al. [6] achieve emotional lip sync by the proposed constrained search and Anime Graph structure. Recently, Taylor et al. [52] propose a deep-learning-based model utilizing a sliding window approach on the transcribed phoneme sequences input. Karras et al. [29] propose a convolution-based network with a learnable emotion database to animate a speech-driven 3D mesh. More recently, VisemeNet [69] employ a three-stage Long Short-Term Memory (LSTM) network to predict the animation curve for a lower face lip model." + }, + { + "type": "text", + "bbox": [ + 0.504, + 0.366, + 0.891, + 0.622 + ], + "angle": 0, + "content": "We review the most related works more concretely here as they have the same setting as this work, i.e., training on high-resolution paired audio-mesh data and speaker-independently animating entire face meshes in vertex space. MeshTalk [50] successfully disentangles audio-correlated and uncorrelated facial information with a categorical latent space. However, the latent space adopted is not optimal with limited expressiveness, thus the animation quality is not stable when applied in a data-scarcity setting. VOCA [10] employs powerful audio feature extraction models and can generate facial animation with different speaking styles. Furthermore, FaceFormer [16] considers long-term audio context with transformer [58] rendering temporally stable animations. Despite the appealing animations, both suffer from the over-smoothing problem, as they directly regress the facial motion in the highly ill-posed audio-visual mapping with large uncertainty and ambiguity." + }, + { + "type": "title", + "bbox": [ + 0.504, + 0.634, + 0.719, + 0.65 + ], + "angle": 0, + "content": "2.2. Discrete Prior Learning" + }, + { + "type": "text", + "bbox": [ + 0.504, + 0.659, + 0.891, + 0.854 + ], + "angle": 0, + "content": "In the last decades, discrete prior representation with learned dictionaries has demonstrated its superiority in image restoration tasks [14, 22, 28, 55, 56], since clear image details are well-preserved in the dictionaries. This line of techniques further inspires the high-capacity and high-compressed discrete prior learning. VQ-VAE [57] first presents to learn discrete representations (codebook) of images and autoregressively model their distribution for image synthesis. The follow-up works, VQ-VAE2 [48] and VQGAN [15] further improve the quality of high-resolution image synthesis. Recently, discrete prior learning has been exploited for image colorization [25], inpainting [45], blind face restoration [68], text-to-image synthesis [21], etc." + }, + { + "type": "text", + "bbox": [ + 0.504, + 0.856, + 0.891, + 0.9 + ], + "angle": 0, + "content": "In addition to image modality, most recent works also explore the power of discrete prior learning in tasks with other modalities, such as dyadic face motion generation [42], co" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.518, + 0.957 + ], + "angle": 0, + "content": "12781" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.184 + ], + "angle": 0, + "content": "speech gesture synthesis [2], speech enhancement [64]. Inspired by codebook learning, this work investigates to learn discrete motion prior for speech-driven 3D facial animation. Different from [42], we exploit the discrete motion primitives for facial motion representation in a context-rich manner, which is more effective to learn general priors." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.192, + 0.169, + 0.208 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.215, + 0.47, + 0.382 + ], + "angle": 0, + "content": "We aim to synthesize sequential 3D facial motions from a speech signal, so that any neutral face mesh could be animated as a lip-synchronized talking face. However, this is an ill-posed problem since one speech could be matched by multiple potential facial animations. Such ambiguity tends to make cross-modal learning suffer from averaged motions and lack of subtle variations. To bypass this barrier, we propose to first model the facial motion space with the learned discrete motion prior, and then learn a speech-conditioned temporal autoregressive model over this space, which promotes robustness against the cross-modal uncertainty." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.385, + 0.471, + 0.522 + ], + "angle": 0, + "content": "Formulation. Let \\(\\mathbf{M}_{1:T} = (\\mathbf{m}_1, \\dots, \\mathbf{m}_T)\\) be a sequence of facial motions, where each frame \\(\\mathbf{m}_t \\in \\mathbb{R}^{V \\times 3}\\) denotes the 3D movement of \\(V\\) vertices over a neutral-face mesh template \\(\\mathbf{h} \\in \\mathbb{R}^{V \\times 3}\\). Let further \\(\\mathbf{A}_{1:T} = (\\mathbf{a}_1, \\dots, \\mathbf{a}_T)\\) be a sequence of speech snippets, each of which \\(\\mathbf{a}_t \\in \\mathbb{R}^d\\) has \\(d\\) samples to align with the corresponding (visual) frame \\(\\mathbf{m}_t\\). Then, our goal is to sequentially synthesize \\(\\mathbf{M}_{1:T}\\) from \\(\\mathbf{A}_{1:T}\\) so that an arbitrary neutral facial template \\(\\mathbf{f}\\) could be animated as \\(\\mathbf{H}_{1:T} = \\{\\mathbf{m}_1 + \\mathbf{h}, \\dots, \\mathbf{m}_T + \\mathbf{h}\\}\\)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.528, + 0.338, + 0.544 + ], + "angle": 0, + "content": "3.1. Discrete Facial Motion Space" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.551, + 0.47, + 0.702 + ], + "angle": 0, + "content": "Visual realistic facial animations should present accurate lip motions and natural expressions. To achieve this from speech signals, extra motion priors are required to reduce the uncertainty and complement realistic motion components. As witnessed by the recent image restoration task [68], discrete codebook prior [57] demonstrates advantages in guaranteeing high-fidelity results even from a severely degraded input. Inspired by this, we propose to model the facial motion space as a discrete codebook by learning from tracked real-world facial motions." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.705, + 0.471, + 0.904 + ], + "angle": 0, + "content": "Codebook of motion primitives. We manage to learn a codebook \\(\\mathcal{Z} = \\{\\mathbf{z}_k\\in \\mathbb{R}^C\\}_{k = 1}^N\\) that allows any facial motion \\(\\mathbf{m}_t\\) to be represented by a group of allocated items \\(\\{\\mathbf{z}_k\\}_{k\\in S}\\), where \\(S\\) denotes the selected index set through Eq. 1. Conceptually, the codebook items serve as the motion primitives of a facial motion space. To this end, we pre-train a transformer-based VQ-VAE that consists of an encoder \\(E\\), a decoder \\(D\\), and a context-rich codebook \\(\\mathcal{Z}\\), under the self-reconstruction of realistic facial motions. As shown in Figure 1, the facial motions \\(\\mathbf{M}_{1:T}\\) is first embedded as a temporal feature \\(\\hat{\\mathbf{Z}} = E(\\mathbf{M}_{1:T})\\in \\mathbb{R}^{T'\\times H\\times C}\\), where \\(H\\) is the number of face components and \\(T^{\\prime}\\) denotes the number of encoded temporal units (\\(P = \\frac{T}{T'}\\) frames). Then, we" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.088, + 0.895, + 0.234 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.243, + 0.894, + 0.286 + ], + "angle": 0, + "content": "Figure 1. Learning framework of facial motion space. The learned motion primitives, as embedded in the codebook, serve to represent the facial motions in a spatial and temporal manner." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.292, + 0.893, + 0.341 + ], + "angle": 0, + "content": "obtain the quantized motion sequence \\(\\mathbf{Z}_{\\mathbf{q}} \\in \\mathbb{R}^{T' \\times H \\times C}\\) via an element-wise quantization function \\(Q(\\cdot)\\) that maps each item in \\(\\hat{\\mathbf{Z}}\\) to its nearest entry in codebook \\(\\mathcal{Z}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.573, + 0.354, + 0.892, + 0.38 + ], + "angle": 0, + "content": "\\[\n\\mathbf {Z} _ {\\mathbf {q}} = Q (\\hat {\\mathbf {Z}}) := \\underset {\\mathbf {z} _ {k} \\in \\mathcal {Z}} {\\arg \\min } \\| \\hat {\\mathbf {z}} _ {t} - \\mathbf {z} _ {k} \\| _ {2}. \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.386, + 0.774, + 0.402 + ], + "angle": 0, + "content": "Then, the self-reconstruction is given by:" + }, + { + "type": "equation", + "bbox": [ + 0.568, + 0.412, + 0.892, + 0.43 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathbf {M}} _ {1: T} = D \\left(\\mathbf {Z} _ {\\mathbf {q}}\\right) = D \\left(Q \\left(E \\left(\\mathbf {M} _ {1: T}\\right)\\right)\\right). \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.441, + 0.892, + 0.502 + ], + "angle": 0, + "content": "Note that, the discrete facial motion space reduces the mapping ambiguity with the finite cardinality, but never sacrifices its expressiveness thanks to its context-rich representation as a latent space." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.505, + 0.893, + 0.551 + ], + "angle": 0, + "content": "Training objectives. Similar to [57], to supervise the quantized autoencoder training, we adopt a motion-level loss and two intermediate code-level losses:" + }, + { + "type": "equation", + "bbox": [ + 0.54, + 0.56, + 0.892, + 0.6 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} _ {\\mathrm {V Q}} = \\left\\| \\mathbf {M} _ {1: T} - \\hat {\\mathbf {M}} _ {1: T} \\right\\| _ {1} \\tag {3} \\\\ + \\left\\| \\operatorname {s g} (\\hat {\\mathbf {Z}}) - \\mathbf {Z} _ {\\mathbf {q}} \\right\\| _ {2} ^ {2} + \\beta \\left\\| \\hat {\\mathbf {Z}} - \\operatorname {s g} (\\mathbf {Z} _ {\\mathbf {q}}) \\right\\| _ {2} ^ {2}, \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.61, + 0.894, + 0.746 + ], + "angle": 0, + "content": "where the first term is a reconstruction loss, the latter two are adopted to update the codebook items by reducing the distance between the codebook \\(\\mathcal{Z}\\) and embedded features \\(\\hat{\\mathbf{Z}}\\). \\(\\mathrm{sg}(\\cdot)\\) stands for a stop-gradient operation and \\(\\beta\\) is a weighting factor controlling the update rate of the codebook and encoder. Since the quantization function (Eq. 1) is not differentiable, the straight-through gradient estimator [4,57] is employed to copy the gradients from the decoder input to the encoder output." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.895, + 0.902 + ], + "angle": 0, + "content": "Discussion. Recently, Learn2Listen [42] has applied VQ-VAE for facial expression synthesis in response to a given talking head harnessing 2D monocular videos to obtain 3DMM coefficients. In addition to distinct applications, here we would like to emphasize our major differences. First, Learn2Listen constructs speaker-specific codebooks while ours uses a generic codebook that is feasible to represent arbitrary facial motions. Since cross-character motions are absorbed, our codebook is naturally embedded with more plentiful priors. Second, Learn2Listen utilizes" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12782" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.08, + 0.09, + 0.892, + 0.282 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.287, + 0.893, + 0.33 + ], + "angle": 0, + "content": "Figure 2. Diagram of our speech-driven motion synthesis model. Given the speech \\(\\mathbf{A}_{1:T}\\) and style vector \\(\\mathbf{s}\\) as input, the model learn to recursively generate a sequence of facial motions by predicting the motion codes. As embedded with well-learned motion priors, the pre-trained codebook and decoder are frozen during training." + }, + { + "type": "image", + "bbox": [ + 0.078, + 0.336, + 0.468, + 0.469 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.479, + 0.47, + 0.576 + ], + "angle": 0, + "content": "Figure 3. Concept comparison with Learn2Listen [42]. (Top) The speaker-specific facial expression coefficient prior [42], in which each code represents a sequence of facial expression coefficients. (Bottom) Our speaker-agnostic generic motion prior, in which each code represents the motion primitive of face components. The blue dotted boxes indicate what information each code may represent conceptually." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.583, + 0.47, + 0.839 + ], + "angle": 0, + "content": "the codebook to represent common sequences of facial expressions by the way of 3DMM coefficients, i.e., each code represents a sequence (8 frames) of facial expressions. Differently, our codebook is formulated to represent the vertex-based facial motion space, where the codes are embedded with per-vertex motions of facial components and represent the facial motion (within a temporal unit) in a context-rich manner. As compared in Figure 3, the codebook of Learn2Listen is learned to memorize typical sequential facial expressions of a specific speaker within 3DMM space, which cannot synthesize realistic facial motions with subtle details due to the limited expressiveness of 3DMM and is bounded by the accuracy of 3D reconstruction techniques [42]. As the first attempt, our codebook is learned to represent the generic facial motion space with motion primitives for captured facial mesh data, which is more effective to embed general priors preserving vivid facial details." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.469, + 0.901 + ], + "angle": 0, + "content": "We further discuss the hyper-parameters of the codebook. First, the length of the temporal unit \\( P \\) and the number of face components \\( H \\) determine the complexity of the motion primitives in temporal and spatial aspects" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.337, + 0.892, + 0.518 + ], + "angle": 0, + "content": "respectively. Generally, complex motion primitives cause low flexibility and reusability and thus hinder representation effectiveness. On the opposite, overly simple motion primitives challenge motion prediction due to the lack of semantics. Besides, the codebook size \\( N \\) and the feature dimension \\( C \\) determine the representation capability, which should be defined according to the complexity of the dataset and in cooperation with \\( P \\) and \\( H \\). In our experiment, we set \\( N = 256 \\), \\( P = 1 \\), \\( H = 8 \\) or \\( H = 16 \\), and \\( C = 64 \\) or \\( C = 128 \\) depending on the dataset, which lead to high-quality results as justified by the ablation studies in Section 4.5. More details can be found in the Supplement." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.529, + 0.785, + 0.545 + ], + "angle": 0, + "content": "3.2. Speech-Driven Motion Synthesis" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.553, + 0.892, + 0.719 + ], + "angle": 0, + "content": "With the learned discrete motion prior, we can build a cross-modal mapping from the input speech to the target motion codes that could be further decoded into realistic facial motions. Along with the speech, we further adopt a control on the talking styles as input, i.e., a style vector \\(\\mathbf{s} \\in \\mathbb{R}_+^M \\cup \\{0\\}\\), where \\(M\\) is the dimension of the learned style space (see Eq. 4). Conditioning on the speech \\(\\mathbf{A}_{1:T}\\) and the style vector \\(\\mathbf{s}\\), a temporal autoregressive model, composed of a speech encoder \\(E_{\\mathrm{speech}}\\) and a cross-modal decoder \\(D_{\\mathrm{cross-modal}}\\), is employed to learn over the facial motion space, as depicted in Figure 2." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.893, + 0.901 + ], + "angle": 0, + "content": "Following FaceFormer [16], our speech encoder adopts the architecture of the state-of-the-art self-supervised pretrained speech model, wav2vec 2.0 [3], which consists of an audio feature extractor and a multi-layer transformer encoder. The audio feature extractor converts the speech of raw waveform into feature vectors through a temporal convolutions network (TCN). Benefiting from the effective attention scheme, the transformer encoder converts the audio features into contextualized speech representations. Apart from the pre-trained codebook and VQ-VAE decoder, our cross-modal decoder contains an embedding block and a multi-layer transformer decoder with causal self-attention." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "12783" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.47, + 0.122 + ], + "angle": 0, + "content": "The embedding block combines the past facial motions and the style embedding via:" + }, + { + "type": "equation", + "bbox": [ + 0.154, + 0.133, + 0.469, + 0.162 + ], + "angle": 0, + "content": "\\[\n\\mathbf {F} _ {\\mathrm {e m b}} ^ {1: t - 1} = \\mathcal {P} _ {\\theta} \\left(\\hat {\\mathbf {M}} _ {1: t - 1}\\right) + \\mathbf {B} \\cdot \\frac {\\mathbf {S}}{\\| \\mathbf {S} \\| _ {1}}, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.172, + 0.47, + 0.354 + ], + "angle": 0, + "content": "where \\(\\mathcal{P}_{\\theta}\\) is a linear projection layer, and \\(\\mathbf{B} = [\\mathbf{b}_1,\\dots,\\mathbf{b}_M]\\in \\mathbb{R}^{C\\times M}\\) denotes the \\(M\\) learnable basis vectors that span the style space linearly. Alike to Face-Former [16], we equip the transformer decoder with causal self-attention to learn the dependencies between each frame in the context of the past facial motion sequence, and with cross-modal attention to align the audio and motion modalities. The output features \\(\\hat{\\mathbf{Z}}^{1:t}\\) is further quantized into \\(\\mathbf{Z}_{\\mathbf{q}}^{1:t}\\) via Eq. 1 and decoded by the pre-trained VQ-VAE decoder. The newly predicted motion \\(\\hat{\\mathbf{m}}_t\\) is used to update the past motions as \\(\\hat{\\mathbf{M}}_{1:t}\\), in preparation for the next prediction. Formally, this recursive process can be written as:" + }, + { + "type": "equation", + "bbox": [ + 0.124, + 0.366, + 0.469, + 0.384 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathbf {m}} _ {t} = D _ {\\text {c r o s s - m o d a l}} \\left(E _ {\\text {s p e e c h}} \\left(\\mathbf {A} _ {1: T}\\right), \\mathbf {s}, \\tilde {\\mathbf {M}} _ {1: t - 1}\\right). \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.398, + 0.469, + 0.613 + ], + "angle": 0, + "content": "Training objectives. We train the transformer encoder, decoder and the embedding block for cross-modality mapping, while keeping the codebook \\(\\mathcal{Z}\\) and motion decoder \\(D\\) frozen. To benefit from the speech representation learning from large-scale corpora, we initialize the TCN and transformer encoder with the pre-trained wav2vec 2.0 weights. Overall, the autoregressive model is trained in a teaching-forcing scheme, under the constraint of two loss terms: (i) feature regularity loss \\(\\mathcal{L}_{\\mathrm{reg}}\\) measuring the deviation between the predicted motion feature \\(\\hat{\\mathbf{Z}}^{1:T}\\) and the quantized feature \\(\\mathbf{Z}_{\\mathbf{q}}^{1:T}\\) from codebook, and (ii) motion loss \\(\\mathcal{L}_{\\mathrm{motion}}\\) measuring the difference between the predicted motions \\(\\hat{\\mathbf{M}}_{1:T}\\) and the ground-truth motions \\(\\mathbf{M}_{1:T}\\), which plays an important role to stabilize the training process. The final loss function is:" + }, + { + "type": "equation", + "bbox": [ + 0.097, + 0.626, + 0.468, + 0.664 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} _ {\\text {s y n}} = \\mathcal {L} _ {\\text {r e g}} + \\mathcal {L} _ {\\text {m o t i o n}} \\\\ = \\| \\hat {\\mathbf {Z}} ^ {1: T} - \\operatorname {s g} \\left(\\mathbf {Z} _ {\\mathbf {q}} ^ {1: T}\\right) \\| _ {2} ^ {2} + \\| \\hat {\\mathbf {M}} _ {1: T} - \\mathbf {M} _ {1: T} \\| _ {2} ^ {2}. \\tag {6} \\\\ \\end{array}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.675, + 0.238, + 0.691 + ], + "angle": 0, + "content": "3.3. Training Details" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.699, + 0.469, + 0.82 + ], + "angle": 0, + "content": "At stage one, we train the VQ-VAE model (Figure 1) on a single NVIDIA V100 for 200 epochs (\\(\\sim\\)2 hours) with the AdamW [39] optimizer (\\(\\beta_{1} = 0.9\\), \\(\\beta_{2} = 0.999\\) and \\(\\epsilon = 1e - 8\\)), where the learning rate is initialized as \\(10^{-4}\\), and the mini-batch size is set as 1. At stage two, we train the temporal autoregressive model with the Adam optimizer [31]. The training duration is 100 epochs (\\(\\sim\\)3 hours) and other hyper-parameters remain unchanged as stage one." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.826, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Style embedding space. The style embedding space is linearly spanned by \\( M \\) learned basis vectors, where each style is represented by a style vector \\( \\mathbf{s} \\) that serves as the linear combination coefficients or a coordinate. In VOCA [10], it assigns each speaker with a category-like" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.243 + ], + "angle": 0, + "content": "one-hot style vector. Instead, we propose a concept to formulate a uniform style space via some learnable basis vectors (Eq. 4), where the style vector is no longer bound with speaker ID but associated with every talking sample. During training, considering the limited style diversity of training datasets, we assign each speaker (e.g. no. \\(i\\)) with a standard unit vector \\(\\mathbf{e}_i\\) as a style vector, under the assumption that each speaker is associated with a unique and consistent style. Anyway, arbitrary style vectors are allowed to interpolate new talking styles during inference." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.249, + 0.634, + 0.266 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.274, + 0.769, + 0.29 + ], + "angle": 0, + "content": "4.1. Datasets and Implementations" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.297, + 0.892, + 0.403 + ], + "angle": 0, + "content": "We employ two widely used datasets, BIWI [18] and VOCASET [10], to train and test different methods in our experiments. Both datasets contain 4D face scans together with utterances spoken in English. BIWI contains 40 unique sentences shared across all speakers in the dataset, while VOCASET contains 255 unique sentences, which are partially shared among different speakers." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.403, + 0.892, + 0.674 + ], + "angle": 0, + "content": "BIWI dataset. BIWI is a 3D audio-visual corpus of affective speech and facial expression in the form of dense dynamic 3D face geometries, which is originally proposed to study affective communication. There is a total of 40 sentences uttered by 14 subjects, eight females and six males. Each sentence was recorded twice: with and without emotion. On average, each sentence is 4.67 seconds long. The 3D face dynamics are captured at 25fps, each with 23370 vertices and registered topology. We follow the data splits in [16] and use the emotional subset. Specifically, the training set (BIWI-Train) contains 192 sentences, while the validation set (BIWI-Val) contains 24 sentences. There are two testing sets, in which BIWI-Test-A includes 24 sentences spoken by six seen subjects and BIWI-Test-B contains 32 sentences spoken by eight unseen subjects. BIWI-Test-A can be used for both quantitative and qualitative evaluation due to the seen subjects during training, while BIWI-Test-B is more suitable for qualitative evaluation." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.675, + 0.892, + 0.795 + ], + "angle": 0, + "content": "VOCASET dataset. VOCASET is comprised of 480 paired audio-visual sequences recorded from 12 subjects. The facial motion is captured at 60fps and is about 4 seconds long. Different from BIWI, each 3D face mesh is registered to the FLAME [35] topology with 5023 vertices. We adopt the same training (VOCA-Train), validation (VOCA-Val), and testing (VOCA-Test) splits as VOCA [10] and FaceFormer for fair comparisons." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.796, + 0.892, + 0.9 + ], + "angle": 0, + "content": "Implementations. We compare our work with three state-of-the-art methods: VOCA [10], MeshTalk [50] and FaceFormer [16]. We train and test VOCA on BIWI using the official codebase, while directly testing the released model that was trained on VOCASET. For MeshTalk, we train and test it using the official implementation on the two datasets. To compare with FaceFormer, we conduct test-" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12784" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.09, + 0.47, + 0.117 + ], + "angle": 0, + "content": "Table 1. Quantitative evaluation on BIWI-Test-A. Lower means better for both metrics." + }, + { + "type": "table", + "bbox": [ + 0.085, + 0.123, + 0.462, + 0.224 + ], + "angle": 0, + "content": "
MethodLip Vertex Error (×10-4mm)FDD (×10-5mm)
VOCA6.55638.1816
MeshTalk5.91815.1025
FaceFormer5.30774.6408
CodeTalker (Ours)4.79144.1170
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.235, + 0.468, + 0.326 + ], + "angle": 0, + "content": "ing directly using the pre-trained weights. Among the four methods, VOCA, FaceFormer and our CodeTalker require conditioning on a training speaking style during testing. For unseen subjects, we generate facial animations conditioned on all training styles. More details about the implementation details can be found in the Supplementary Material." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.335, + 0.298, + 0.351 + ], + "angle": 0, + "content": "4.2. Quantitative Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.358, + 0.468, + 0.463 + ], + "angle": 0, + "content": "Following MeshTalk [50] and FaceFormer [16], we adopt the lip vertex error to measure the lip synchronization, which is the only publicly proposed metric for speech-driven facial animation evaluation, to our best knowledge. As a complement, we introduce a new quantitative measurement, i.e., upper-face motion statistics, to evaluate the overall facial dynamics." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.464, + 0.469, + 0.524 + ], + "angle": 0, + "content": "Lip vertex error. It measures the lip deviation of a sequence with respect to the ground truth, i.e., calculating the maximal L2 error of all lip vertices for each frame and takes the average over all frames." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.525, + 0.469, + 0.631 + ], + "angle": 0, + "content": "Upper-face dynamics deviation. The upper-face expression is just loosely correlated with the speech, depending on personal talking styles and the semantics of speech content. With this belief, we propose to measure the variation of facial dynamics for a motion sequence in comparison with that of the ground truth. Specifically, the upper-face dynamics deviation (FDD) is calculated by:" + }, + { + "type": "equation", + "bbox": [ + 0.077, + 0.642, + 0.476, + 0.689 + ], + "angle": 0, + "content": "\\[\n\\operatorname {F D D} \\left(\\mathbf {M} _ {1: T}, \\hat {\\mathbf {M}} _ {1: T}\\right) = \\frac {\\sum_ {v \\in \\mathcal {S} _ {U}} \\left(\\operatorname {d y n} \\left(\\mathbf {M} _ {1 : T} ^ {v}\\right) - \\operatorname {d y n} \\left(\\hat {\\mathbf {M}} _ {1 : T} ^ {v}\\right)\\right)}{\\left| \\mathcal {S} _ {U} \\right|}, \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.689, + 0.468, + 0.75 + ], + "angle": 0, + "content": "where \\(\\mathbf{M}_{1:T}^v\\in \\mathbb{R}^{3\\times T}\\) denotes the motions of the \\(v\\) -th vertex, and \\(S_U\\) is the index set of upper-face vertices. \\(\\mathrm{dyn}(\\cdot)\\) denotes the standard deviation of the element-wise L2 norm along the temporal axis." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.469, + 0.902 + ], + "angle": 0, + "content": "We calculate the lip vertex error and upper-face dynamics deviation (FDD) over all sequences in BIWI-Test-A and take the average for comparison. According to Table 1, the proposed CodeTalker achieves lower error than the existing state-of-the-arts, suggesting that it produces more accurate lip-synchronized movements. Besides, Table 1 shows that our CodeTalker achieves the best performance in terms of FDD. It indicates the high consistency between the predicted upper-face expressions together with the trend of facial dynamics (conditioned on the speech and talking styles)" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.092, + 0.697, + 0.107 + ], + "angle": 0, + "content": "and those of the ground truth." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.119, + 0.71, + 0.135 + ], + "angle": 0, + "content": "4.3. Qualitative Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.144, + 0.892, + 0.431 + ], + "angle": 0, + "content": "We visually compare our method with other competitors in Figure 4. For fair comparison, we assign the same talking style to VOCA, FaceFormer and our CodeTalker as conditional input, which is sampled at random. To check the lip synchronization performance, we illustrate three typical frames of synthesized facial animations that speak at specific syllables, as compared in the upper partition in Figure 4. We can observe that compared with the competitors, the lip movements produced by our CodeTalker are more accurately articulated with the speech signals and also more consistent with those of the reference. For example, CodeTalker produces better lip sync with proper mouth closures when pronouncing bilabial consonant /b/ (i.e., \"bedside\" in the upper-right case of Figure 4), compared to VOCA and MeshTalk; for the even challenging speech parts \"waterproof\" and \"shaving\" that need to pout, CodeTalker can produce accurate lip shapes while other methods suffer from the over-smoothing problem and fail to lip-sync correctly (Zoom in for better inspection)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.432, + 0.892, + 0.704 + ], + "angle": 0, + "content": "Different from lip movements, facial expressions only have weak correlations with the speech signal, which tends to be static in front of cross-modal mapping ambiguity. To visualize the facial motion dynamics, we calculate the temporal statistics of adjacent-frame facial motions within a sequence. Specifically, we first calculate the inter-frame motion L2 distance and then compute the mean and standard deviation (std) across the sequence at each vertex. The higher mean value indicates stronger facial movements, while the higher std value suggests richer variations of facial dynamics. Two examples are visualized in the last two rows of Figure 4, evidencing that our method outperforms others in achieving both stronger facial movements and a broader range of dynamics. It is mainly attributed to the superiority of the discrete facial motion space, which promotes the robustness to cross-modal uncertainty effectively. Readers are recommended to watch the animation comparisons in the Supplemental Video." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.705, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Talking style interpolation. Our model can synthesize new speaking styles from the learned style embedding space. To inspect the effects on VOCA-Test, we select two speaking style vectors, i.e., \\(\\mathbf{e}_i\\) and \\(\\mathbf{e}_j\\), which correspond to large and slight lip articulations respectively, and interpolate new talking style vectors \\(\\mathbf{s}_{\\mathrm{new}} = \\mathbf{B} \\cdot [\\omega \\mathbf{e}_i + (1 - \\omega) \\mathbf{e}_j]\\) with a linear coefficient \\(\\omega\\). For the synthesized 3D animations of a sampled sequence, we plot the lower-upper lip distances across frames for each style in Figure 5, from which we observe the smooth transition of mouth amplitudes between the two typical styles. It is not only useful to synthesize new talking styles, but also practical to match a specific speaking performance of an unseen subject during training." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "12785" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.08, + 0.086, + 0.502, + 0.436 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.087, + 0.895, + 0.436 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.445, + 0.895, + 0.488 + ], + "angle": 0, + "content": "Figure 4. Visual comparisons of sampled facial motions animated by different methods on VOCA-Test (left) and BIWI-Test-B (right). The upper partition shows the facial animation conditioned on different speech parts, while the lower depicts the temporal statistics (mean and standard deviation) of adjacent-frame motion variations within a sequence." + }, + { + "type": "image", + "bbox": [ + 0.079, + 0.497, + 0.473, + 0.643 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.646, + 0.47, + 0.688 + ], + "angle": 0, + "content": "Figure 5. Distance between lower and upper lip for our predictions within a sequence conditioned on different weighted linear combinations of two style vectors." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.697, + 0.2, + 0.713 + ], + "angle": 0, + "content": "4.4. User Study" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.72, + 0.471, + 0.903 + ], + "angle": 0, + "content": "The human perception system has been evolutionarily adapted to understanding subtle facial motions and capturing lip synchronization. Thus, it is still the most reliable measure in the speech-driven facial animation task. We conduct a user study to evaluate the quality of animated faces in perceptual lip synchronization and realism, compared with VOCA, MeshTalk, FaceFormer and the ground truth. We adopt A/B tests for each comparison, i.e., ours vs. competitor, in terms of realistic facial animation and lip sync. For BIWI, we obtain the results of four kinds of comparisons by randomly selecting 30 samples from BIWI-Test-B, respectively. To achieve the most variations in terms of speaking" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.498, + 0.892, + 0.54 + ], + "angle": 0, + "content": "Table 2. User study results on BIWI-Test-B and VOCA-Test. We adopt A/B testing and report the percentage of answers where A is preferred over B." + }, + { + "type": "table", + "bbox": [ + 0.504, + 0.544, + 0.892, + 0.649 + ], + "angle": 0, + "content": "
CompetitorsBIWI-Test-BVOCA-Test
Lip SyncRealismLip SyncRealism
Ours vs. VOCA92.4789.2586.0284.95
Ours vs. MeshTalk80.6582.8095.7092.47
Ours vs. FaceFormer53.7656.9970.9769.89
Ours vs. GT43.0149.4643.0143.01
" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.658, + 0.892, + 0.84 + ], + "angle": 0, + "content": "styles, we ensure the sampling results can fairly cover all conditioning styles. Thus, \\(120\\mathrm{A}\\) vs. B pairs (30 samples \\(\\times 4\\) comparisons) are created for BIWI-Test-B. Each pair is judged by at least 3 different participants separately, and finally, 372 entries are collected in total. For the user study on VOCASET, we apply the same setting as that on the BIWI dataset, i.e., another \\(120\\mathrm{A}\\) vs. B pairs from VOCA-Test set, finally yielding 372 entries as well. In this study, 31 participants with good vision and hearing ability complete the evaluation successfully. Moreover, each participant is involved in all 8 kinds of comparisons to make better exposure and cover the diversity of favorability." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.894, + 0.901 + ], + "angle": 0, + "content": "The percentage of A/B testing in terms of lip sync and realism on BIWI-Test-B is tabulated in Table 2, which shows that participants favor CodeTalker over competitors. Based on the visual analysis in Section 4.3, we attribute this to the" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "12786" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.09, + 0.47, + 0.131 + ], + "angle": 0, + "content": "Table 3. Ablation study on the representation space of codebook. The performance is measured by the reconstruction error (i.e., L2 error) and lip vertex error on VOCA-Test and BIWI-Test-A." + }, + { + "type": "table", + "bbox": [ + 0.08, + 0.136, + 0.469, + 0.389 + ], + "angle": 0, + "content": "
VariantsVOCA-TestBIWI-Test-A
Rec. Error (×10-5mm)Rec. Error (×10-5mm)Lip Vertex Error (×10-4mm)
Shape-ent. codebook2.754.076.41
Motion codebook (Ours)0.082.834.79
Speech-driven synthesis"finding Nan's cameo""an excellent room"
Shape-ent. codebookMotion (Ours) codebookShape-ent. codebookMotion (Ours) codebook
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.397, + 0.47, + 0.439 + ], + "angle": 0, + "content": "Figure 6. Visual comparisons of reconstruction and speech-driven motion synthesis results with different representation spaces on VOCA-Test (left) and BIWI-Test-A (right)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.446, + 0.47, + 0.628 + ], + "angle": 0, + "content": "facial animation synthesized by CodeTalker having more expressive facial motions, accurate lip shape, and well-synchronized mouth movements. For VOCA-Test, which has a nature of fewer upper-face motions, a similar favorability can still be observed in Table 2. We believe the reasons are at least three-fold: subtle motions around the eyes, more accurate lip movements and expressive motions in the lower face. Although be aware of a gap between our predictions and the recorded performance (ground truth), we surprisingly get over \\(40\\%\\) preference still. Overall, the user study justifies that the facial animations produced by CodeTalker have superior perceptual quality." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.636, + 0.241, + 0.65 + ], + "angle": 0, + "content": "4.5. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.66, + 0.469, + 0.704 + ], + "angle": 0, + "content": "We study several key designs of our proposed method in this section, including the representation space and the hyper-parameters of the codebook construction." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.705, + 0.47, + 0.901 + ], + "angle": 0, + "content": "Representation space. To study the superiority of our motion-based representation, we construct a baseline that learns a shape-entangled codebook, i.e., the codes represent shapes \\((\\mathbf{m_t} + \\mathbf{h})\\) instead of motions \\((\\mathbf{m_t})\\). As shown in Table 3, the baseline decreases the reconstruction accuracy significantly, which is evidenced by the visualized examples in Figure 6. It is mainly because the shape-entangled sequence contains more speaker-specific information that hinders the reusability of the codes. A direct weakness is the poor generalization for self-reconstruction, which further impedes cross-modal mapping correctness. In contrast, our proposed speaker-agnostic motion representation is more effective to represent generic motion priors shared" + }, + { + "type": "image", + "bbox": [ + 0.501, + 0.089, + 0.695, + 0.224 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.7, + 0.089, + 0.895, + 0.224 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.232, + 0.892, + 0.288 + ], + "angle": 0, + "content": "Figure 7. Model performance comparisons with different hyperparameters of the codebook, i.e., the length of temporal unit \\( P \\) and the number of face components \\( H \\). We measure the reconstruction error and lip vertex error on BIWI-Test-A." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.299, + 0.892, + 0.329 + ], + "angle": 0, + "content": "across individuals, and hence promotes the quality of both self-reconstruction and speech-driven motion synthesis." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.33, + 0.893, + 0.571 + ], + "angle": 0, + "content": "Codebook construction. We further study the hyperparameters used for codebook construction. We evaluate the performance of different settings \\(\\langle P, H \\rangle\\) by measuring their reconstruction accuracy and cross-modal mapping accuracy (namely lip vertex error). First, we evaluate the reconstruction accuracy as shown in Figure 7(a). On one hand, increasing \\(P\\) degrades the reconstruction accuracy, which could be explained by the increased complexity of the motion to be represented. On the other hand, increasing \\(H\\) eases the reconstruction but risks over-fitting, which explains the general benefits \\((H < 8)\\) but inferior performance when \\(H \\geq 8\\). Notably, a similar trend could be found in the cross-modal mapping performance, as shown in Figure 7(b). We conjecture that complex motion primitives cause lower reusability and higher redundancy, resulting in ambiguity in the cross-modal code query process." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.585, + 0.747, + 0.6 + ], + "angle": 0, + "content": "5. Discussion and Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.61, + 0.892, + 0.851 + ], + "angle": 0, + "content": "We demonstrated the advantages of casting speech-driven facial animation as a code query task in the discrete space, which notably promotes the motion synthesis quality against cross-modal ambiguity. By comparing to the existing state-of-the-arts, our proposed method shows superiority in achieving accurate lip sync and vivid facial expressions. However, we still follow the assumption that facial motions are independent of shapes, whose rationality may deserve further studies. Also, the overall perceptual quality still lags behind the ground truth, primarily due to the scarcity of paired audio-visual data. Additionally, the acquired generic motion prior adheres to the motion distribution delineated by the training set, which may deviate from real-world facial motions. As a future work, it is intriguing to guide the 3D facial animation by utilizing priors from large-scale available talking head videos." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.855, + 0.892, + 0.899 + ], + "angle": 0, + "content": "Acknowledgements. This project is partially supported by Hong Kong Innovation and Technology Fund (ITF) (ref: ITS/307/20FP and ITS/313/20)." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "12787" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.079, + 0.09, + 0.176, + 0.107 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.116, + 0.47, + 0.171 + ], + "angle": 0, + "content": "[1] Mohammed M Alghamdi, He Wang, Andrew J Bulpitt, and David C Hogg. Talking head from speech audio using a pretrained image generator. In ACM International Conference on Multimedia (MM), 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.173, + 0.472, + 0.228 + ], + "angle": 0, + "content": "[2] Tenglong Ao, Qingzhe Gao, Yuke Lou, Baoquan Chen, and Libin Liu. Rhythmic gesticulator: Rhythm-aware co-speech gesture synthesis with hierarchical neural embeddings. ACM Transactions on Graphics (TOG), pages 1-19, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.23, + 0.471, + 0.286 + ], + "angle": 0, + "content": "[3] Alexei Baevski, Yuhao Zhou, Abdelrahman Mohamed, and Michael Auli. wav2vec 2.0: A framework for self-supervised learning of speech representations. In Conference on Neural Information Processing Systems (NeurIPS), 2020. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.288, + 0.471, + 0.342 + ], + "angle": 0, + "content": "[4] Yoshua Bengio, Nicholas Léonard, and Aaron Courville. Estimating or propagating gradients through stochastic neurons for conditional computation. arXiv preprint arXiv:1308.3432, 2013. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.344, + 0.471, + 0.399 + ], + "angle": 0, + "content": "[5] Chen Cao, Hongzhi Wu, Yanlin Weng, Tianjia Shao, and Kun Zhou. Real-time facial animation with image-based dynamic avatars. ACM Transactions on Graphics (TOG), 35(4):1-12, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.402, + 0.471, + 0.442 + ], + "angle": 0, + "content": "[6] Yong Cao, Wen C Tien, Petros Faloutsos, and Frédéric Pighin. Expressive speech-driven facial animation. ACM Transactions on Graphics (TOG), 24(4):1283-1302, 2005. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.444, + 0.471, + 0.5 + ], + "angle": 0, + "content": "[7] Lele Chen, Guofeng Cui, Celong Liu, Zhong Li, Ziyi Kou, Yi Xu, and Chenliang Xu. Talking-head generation with rhythmic head motion. In European Conference on Computer Vision (ECCV), 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.502, + 0.471, + 0.543 + ], + "angle": 0, + "content": "[8] Lele Chen, Zhiheng Li, Ross K Maddox, Zhiyao Duan, and Chenliang Xu. Lip movements generation at a glance. In European Conference on Computer Vision (ECCV), 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.545, + 0.471, + 0.585 + ], + "angle": 0, + "content": "[9] Joon Son Chung and Andrew Zisserman. Out of time: automated lip sync in the wild. In Asian Conference on Computer Vision (ACCV), 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.588, + 0.471, + 0.656 + ], + "angle": 0, + "content": "[10] Daniel Cudeiro, Timo Bolkart, Cassidy Laidlaw, Anurag Ranjan, and Michael J Black. Capture, learning, and synthesis of 3d speaking styles. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 1, 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.659, + 0.471, + 0.714 + ], + "angle": 0, + "content": "[11] Dipanjan Das, Sandika Biswas, Sanjana Sinha, and Brojeshwar Bhowmick. Speech-driven facial animation using cascaded gans for learning of motion and texture. In European Conference on Computer Vision (ECCV), 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.716, + 0.471, + 0.77 + ], + "angle": 0, + "content": "[12] Pif Edwards, Chris Landreth, Eugene Fume, and Karan Singh. Jali: an animator-centric viseme model for expressive lip synchronization. ACM Transactions on Graphics (TOG), 35(4):1-11, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.773, + 0.471, + 0.813 + ], + "angle": 0, + "content": "[13] Paul Ekman and Wallace V Friesen. Facial action coding system. Environmental Psychology & Nonverbal Behavior, 1978. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.816, + 0.471, + 0.871 + ], + "angle": 0, + "content": "[14] Michael Elad and Michal Aharon. Image denoising via learned dictionaries and sparse representation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2006. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.873, + 0.471, + 0.903 + ], + "angle": 0, + "content": "[15] Patrick Esser, Robin Rombach, and Bjorn Ommer. Taming transformers for high-resolution image synthesis. In" + }, + { + "type": "list", + "bbox": [ + 0.079, + 0.116, + 0.472, + 0.903 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.122, + 0.894, + 0.19 + ], + "angle": 0, + "content": "[16] Yingruo Fan, Zhaojiang Lin, Jun Saito, Wenping Wang, and Taku Komura. Faceformer: Speech-driven 3d facial animation with transformers. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 1, 2, 4, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.193, + 0.892, + 0.248 + ], + "angle": 0, + "content": "[17] Yingruo Fan, Zhaojiang Lin, Jun Saito, Wenping Wang, and Taku Komura. Joint audio-text model for expressive speech-driven 3d facial animation. In ACM Symposium on Interactive 3D Graphics and Games (i3D), 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.25, + 0.892, + 0.305 + ], + "angle": 0, + "content": "[18] Gabriele Fanelli, Juergen Gall, Harald Romsdorfer, Thibaut Weise, and Luc Van Gool. A 3-d audio-visual corpus of affective communication. IEEE Transactions on Multimedia (TMM), 12(6):591-598, 2010. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.307, + 0.892, + 0.347 + ], + "angle": 0, + "content": "[19] Cletus G Fisher. Confusions among visually perceived consonants. Journal of speech and hearing research, 11(4):796-804, 1968. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.349, + 0.892, + 0.419 + ], + "angle": 0, + "content": "[20] Ohad Fried, Ayush Tewari, Michael Zollhöfer, Adam Finkelstein, Eli Shechtman, Dan B Goldman, Kyle Genova, Zeyu Jin, Christian Theobalt, and Maneesh Agrawala. Text-based editing of talking-head video. ACM Transactions on Graphics (TOG), 38(4):1-14, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.421, + 0.892, + 0.489 + ], + "angle": 0, + "content": "[21] Shuyang Gu, Dong Chen, Jianmin Bao, Fang Wen, Bo Zhang, Dongdong Chen, Lu Yuan, and Baining Guo. Vector quantized diffusion model for text-to-image synthesis. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.491, + 0.892, + 0.546 + ], + "angle": 0, + "content": "[22] Shuhang Gu, Wangmeng Zuo, Qi Xie, Deyu Meng, Xiangchu Feng, and Lei Zhang. Convolutional sparse coding for image super-resolution. In IEEE/CVF International Conference on Computer Vision (ICCV), 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.548, + 0.892, + 0.603 + ], + "angle": 0, + "content": "[23] Yudong Guo, Keyu Chen, Sen Liang, Yong-Jin Liu, Hujun Bao, and Juyong Zhang. Ad-nerf: Audio driven neural radiance fields for talking head synthesis. In IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.605, + 0.892, + 0.673 + ], + "angle": 0, + "content": "[24] Ikhsanul Habibie, Weipeng Xu, Dushyant Mehta, Lingjie Liu, Hans-Peter Seidel, Gerard Pons-Moll, Mohamed Elgharib, and Christian Theobalt. Learning speech-driven 3d conversational gestures from video. In ACM International Conference on Intelligent Virtual Agents (IVA), 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.676, + 0.892, + 0.73 + ], + "angle": 0, + "content": "[25] Zhitong Huang, Nanxuan Zhao, and Jing Liao. Unicolor: A unified framework for multi-modal colorization with transformer. ACM Transactions on Graphics (TOG), pages 1-16, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.733, + 0.892, + 0.787 + ], + "angle": 0, + "content": "[26] Xinya Ji, Hang Zhou, Kaisiyuan Wang, Qianyi Wu, Wayne Wu, Feng Xu, and Xun Cao. Eamm: One-shot emotional talking face via audio-based emotion-aware motion model. In ACM SIGGRAPH Conference, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.789, + 0.892, + 0.844 + ], + "angle": 0, + "content": "[27] Xinya Ji, Hang Zhou, Kaisiyuan Wang, Wayne Wu, Chen Change Loy, Xun Cao, and Feng Xu. Audio-driven emotional video portraits. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.846, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[28] Younghyun Jo and Seon Joo Kim. Practical single-image super-resolution using look-up table. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "12788" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.147 + ], + "angle": 0, + "content": "[29] Tero Karras, Timo Aila, Samuli Laine, Antti Herva, and Jaakko Lehtinen. Audio-driven facial animation by joint end-to-end learning of pose and emotion. ACM Transactions on Graphics (TOG), 36(4):1-12, 2017. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.148, + 0.47, + 0.217 + ], + "angle": 0, + "content": "[30] Hyeongwoo Kim, Pablo Garrido, Ayush Tewari, Weipeng Xu, Justus Thies, Matthias Niessner, Patrick Pérez, Christian Richardt, Michael Zollhöfer, and Christian Theobalt. Deep video portraits. ACM Transactions on Graphics (TOG), 37(4):1-14, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.218, + 0.47, + 0.259 + ], + "angle": 0, + "content": "[31] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In International Conference on Learning Representations (ICLR), 2014. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.26, + 0.47, + 0.328 + ], + "angle": 0, + "content": "[32] Avisek Lahiri, Vivek Kwatra, Christian Frueh, John Lewis, and Chris Bregler. Lipsync3d: Data-efficient learning of personalized 3d talking faces from video using pose and lighting normalization. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.33, + 0.47, + 0.371 + ], + "angle": 0, + "content": "[33] John Lewis. Automated lip-sync: Background and techniques. The Journal of Visualization and Computer Animation, 2(4):118-122, 1991. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.372, + 0.47, + 0.412 + ], + "angle": 0, + "content": "[34] Hao Li, Jihun Yu, Yuting Ye, and Chris Bregler. Realtime facial animation with on-the-fly correctives. ACM Transactions on Graphics (TOG), 32(4):42-1, 2013. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.413, + 0.47, + 0.467 + ], + "angle": 0, + "content": "[35] Tianye Li, Timo Bolkart, Michael J Black, Hao Li, and Javier Romero. Learning a model of facial shape and expression from 4d scans. ACM Transactions on Graphics (TOG), 36(6):194-1, 2017. 1, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.469, + 0.47, + 0.538 + ], + "angle": 0, + "content": "[36] Borong Liang, Yan Pan, Zhizhi Guo, Hang Zhou, Zhibin Hong, Xiaoguang Han, Junyu Han, Jingtuo Liu, Errui Ding, and Jingdong Wang. Expressive talking head generation with granular audio-visual control. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.539, + 0.47, + 0.593 + ], + "angle": 0, + "content": "[37] Xian Liu, Yinghao Xu, Qianyi Wu, Hang Zhou, Wayne Wu, and Bolei Zhou. Semantic-aware implicit neural auditory-driven video portrait generation. In European Conference on Computer Vision (ECCV), 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.595, + 0.47, + 0.648 + ], + "angle": 0, + "content": "[38] Yilong Liu, Feng Xu, Jinxiang Chai, Xin Tong, Lijuan Wang, and Qiang Huo. Video-audio driven real-time facial animation. ACM Transactions on Graphics (TOG), 34(6):1-10, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.65, + 0.47, + 0.691 + ], + "angle": 0, + "content": "[39] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In International Conference on Learning Representations (ICLR), 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.692, + 0.47, + 0.733 + ], + "angle": 0, + "content": "[40] DW Massaro, MM Cohen, M Tabain, J Beskow, and R Clark. Animated speech: research progress and applications. Audiovisual Speech Processing, pages 309-345, 2012. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.734, + 0.47, + 0.774 + ], + "angle": 0, + "content": "[41] Wesley Mattheyses and Werner Verhelst. Audiovisual speech synthesis: An overview of the state-of-the-art. Speech Communication, 66:182-217, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.776, + 0.47, + 0.844 + ], + "angle": 0, + "content": "[42] Evonne Ng, Hanbyul Joo, Liwen Hu, Hao Li, Trevor Darrell, Angjoo Kanazawa, and Shiry Ginosar. Learning to listen: Modeling non-deterministic dyadic facial motion. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.846, + 0.47, + 0.9 + ], + "angle": 0, + "content": "[43] Youxin Pang, Yong Zhang, Weize Quan, Yanbo Fan, Xiaodong Cun, Ying Shan, and Dong-ming Yan. Dpe: Disentanglement of pose and expression for general video portrait editing. arXiv preprint arXiv:2301.06281, 2023. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "[44] Frederic I Parke and Keith Waters. Computer facial animation. CRC press, 2008. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.121, + 0.892, + 0.176 + ], + "angle": 0, + "content": "[45] Jialun Peng, Dong Liu, Songcen Xu, and Houqiang Li. Generating diverse structure for image inpainting with hierarchical vq-vae. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.177, + 0.892, + 0.23 + ], + "angle": 0, + "content": "[46] Hai Xuan Pham, Yuting Wang, and Vladimir Pavlovic. End-to-end learning for 3d facial animation from speech. In ACM International Conference on Multimodal Interaction (ICMI), 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.232, + 0.892, + 0.287 + ], + "angle": 0, + "content": "[47] KR Prajwal, Rudrabha Mukhopadhyay, Vinay P Namboodiri, and CV Jawahar. A lip sync expert is all you need for speech to lip generation in the wild. In ACM International Conference on Multimedia (MM), 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.288, + 0.892, + 0.342 + ], + "angle": 0, + "content": "[48] Ali Razavi, Aaron Van den Oord, and Oriol Vinyals. Generating diverse high-fidelity images with vq-vae-2. In Conference on Neural Information Processing Systems (NeurIPS), 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.343, + 0.892, + 0.411 + ], + "angle": 0, + "content": "[49] Alexander Richard, Colin Lea, Shugao Ma, Jurgen Gall, Fernando De la Torre, and Yaser Sheikh. Audio-and gaze-driven facial animation of codec avatars. In IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.413, + 0.892, + 0.482 + ], + "angle": 0, + "content": "[50] Alexander Richard, Michael Zollhöfer, Yandong Wen, Fernando De la Torre, and Yaser Sheikh. Meshtalk: 3d face animation from speech using cross-modality disentanglement. In IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 1, 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.483, + 0.892, + 0.538 + ], + "angle": 0, + "content": "[51] Shuai Shen, Wanhua Li, Zheng Zhu, Yueqi Duan, Jie Zhou, and Jiwen Lu. Learning dynamic facial radiance fields for few-shot talking head synthesis. In European Conference on Computer Vision (ECCV), 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.539, + 0.892, + 0.607 + ], + "angle": 0, + "content": "[52] Sarah Taylor, Taehwan Kim, Yisong Yue, Moshe Mahler, James Krahe, Anastasio Garcia Rodriguez, Jessica Hodgins, and Iain Matthews. A deep learning approach for generalized speech animation. ACM Transactions on Graphics (TOG), 36(4):1-11, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.608, + 0.892, + 0.663 + ], + "angle": 0, + "content": "[53] Sarah L Taylor, Moshe Mahler, Barry-John Theobald, and Iain Matthews. Dynamic units of visual speech. In ACM/Eurographics Symposium on Computer Animation (SCA), 2012. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.664, + 0.892, + 0.719 + ], + "angle": 0, + "content": "[54] Justus Thies, Mohamed Elgharib, Ayush Tewari, Christian Theobalt, and Matthias Nießner. Neural voice puppetry: Audio-driven facial reenactment. In European Conference on Computer Vision (ECCV), 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.72, + 0.892, + 0.774 + ], + "angle": 0, + "content": "[55] Radu Timofte, Vincent De Smet, and Luc Van Gool. Anchored neighborhood regression for fast example-based super-resolution. In IEEE/CVF International Conference on Computer Vision (ICCV), 2013. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.775, + 0.892, + 0.83 + ], + "angle": 0, + "content": "[56] Radu Timofte, Vincent De Smet, and Luc Van Gool. A+: Adjusted anchored neighborhood regression for fast super-resolution. In Asian Conference on Computer Vision (ACCV), 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.831, + 0.892, + 0.872 + ], + "angle": 0, + "content": "[57] Aaron Van Den Oord, Oriol Vinyals, et al. Neural discrete representation learning. Conference on Neural Information Processing Systems (NeurIPS), 2017. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.873, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[58] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "12789" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.092, + 0.47, + 0.12 + ], + "angle": 0, + "content": "Polosukhin. Attention is all you need. Conference on Neural Information Processing Systems (NeurIPS), 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.122, + 0.47, + 0.176 + ], + "angle": 0, + "content": "[59] Konstantinos Vougioukas, Stavros Petridis, and Maja Pantic. Realistic speech-driven facial animation with gans. International Journal of Computer Vision (IJCV), 128(5):1398-1413, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.179, + 0.469, + 0.219 + ], + "angle": 0, + "content": "[60] Qianyun Wang, Zhenfeng Fan, and Shihong Xia. 3d-talkemo: Learning to synthesize 3d emotional talking head. arXiv preprint arXiv:2104.12051, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.222, + 0.469, + 0.276 + ], + "angle": 0, + "content": "[61] Suzhen Wang, Lincheng Li, Yu Ding, and Xin Yu. One-shot talking face generation from single-speaker audio-visual correlation learning. In AAAI Conference on Artificial Intelligence (AAAI), 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.278, + 0.469, + 0.319 + ], + "angle": 0, + "content": "[62] Thibaut Weise, Sofien Bouaziz, Hao Li, and Mark Pauly. Realtime performance-based facial animation. ACM Transactions on Graphics (TOG), 30(4):1-10, 2011. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.321, + 0.469, + 0.361 + ], + "angle": 0, + "content": "[63] Yuyu Xu, Andrew W Feng, Stacy Marsella, and Ari Shapiro. A practical and configurable lip sync method for games. In Proceedings of Motion on Games, 2013. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.364, + 0.469, + 0.432 + ], + "angle": 0, + "content": "[64] Karren Yang, Dejan Markovic, Steven Krenn, Vasu Agrawal, and Alexander Richard. Audio-visual speech codes: Rethinking audio-visual speech enhancement by re-synthesis. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.434, + 0.469, + 0.488 + ], + "angle": 0, + "content": "[65] Ran Yi, Zipeng Ye, Juyong Zhang, Hujun Bao, and Yong-Jin Liu. Audio-driven talking face video generation with learning-based personalized head pose. arXiv preprint arXiv:2002.10137, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.491, + 0.469, + 0.559 + ], + "angle": 0, + "content": "[66] Bowen Zhang, Chenyang Qi, Pan Zhang, Bo Zhang, Hsiang-Tao Wu, Dong Chen, Qifeng Chen, Yong Wang, and Fang Wen. Metaportrait: Identity-preserving talking head generation with fast personalized adaptation. arXiv preprint arXiv:2212.08062, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.562, + 0.469, + 0.63 + ], + "angle": 0, + "content": "[67] Hang Zhou, Yasheng Sun, Wayne Wu, Chen Change Loy, Xiaogang Wang, and Ziwei Liu. Pose-controllable talking face generation by implicitly modularized audio-visual representation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.632, + 0.469, + 0.687 + ], + "angle": 0, + "content": "[68] Shangchen Zhou, Kelvin CK Chan, Chongyi Li, and Chen Change Loy. Towards robust blind face restoration with codebook lookup transformer. Conference on Neural Information Processing Systems (NeurIPS), 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.689, + 0.469, + 0.743 + ], + "angle": 0, + "content": "[69] Yang Zhou, Zhan Xu, Chris Landreth, Evangelos Kalogerakis, Subhransu Maji, and Karan Singh. Visemenet: Audio-driven animator-centric speech animation. ACM Transactions on Graphics (TOG), 37(4):1-10, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.745, + 0.469, + 0.813 + ], + "angle": 0, + "content": "[70] Michael Zollhöfer, Justus Thies, Pablo Garrido, Derek Bradley, Thabo Beeler, Patrick Pérez, Marc Stamminger, Matthias Nießner, and Christian Theobalt. State of the art on monocular 3d face reconstruction, tracking, and applications. Computer Graphics Forum (CGF), 37(2):523-550, 2018. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.813 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "12790" + } + ] +] \ No newline at end of file diff --git a/2023/CodeTalker_ Speech-Driven 3D Facial Animation With Discrete Motion Prior/103a044f-c0a4-40fe-af3f-96d0d65c126f_origin.pdf b/2023/CodeTalker_ Speech-Driven 3D Facial Animation With Discrete Motion Prior/103a044f-c0a4-40fe-af3f-96d0d65c126f_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..612b1b7d8aaa184b8de0ac6e7905f1cd07796495 --- /dev/null +++ b/2023/CodeTalker_ Speech-Driven 3D Facial Animation With Discrete Motion Prior/103a044f-c0a4-40fe-af3f-96d0d65c126f_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:76e6b16cc0840d1123bc5d9a8e2ac2b1394166b9e95f92e0c21113b66fd68539 +size 3250798 diff --git a/2023/CodeTalker_ Speech-Driven 3D Facial Animation With Discrete Motion Prior/full.md b/2023/CodeTalker_ Speech-Driven 3D Facial Animation With Discrete Motion Prior/full.md new file mode 100644 index 0000000000000000000000000000000000000000..eb6221e1818102de6babc3effe38cea4ffb8a996 --- /dev/null +++ b/2023/CodeTalker_ Speech-Driven 3D Facial Animation With Discrete Motion Prior/full.md @@ -0,0 +1,304 @@ +# CodeTalker: Speech-Driven 3D Facial Animation with Discrete Motion Prior + +Jinbo Xing $^{1}$ Menghan Xia $^{2,*}$ Yuechen Zhang $^{1}$ Xiaodong Cun $^{2}$ Jue Wang $^{2}$ Tien-Tsin Wong $^{1}$ $^{1}$ The Chinese University of Hong Kong + $^{2}$ Tencent AI Lab + +{jbxing,yczhang21,ttwong}@cse.cuhk.edu.hk + +{menghanxyz,vinthony,arphid}@gmail.com + +# Abstract + +Speech-driven 3D facial animation has been widely studied, yet there is still a gap to achieving realism and vividness due to the highly ill-posed nature and scarcity of audiovisual data. Existing works typically formulate the cross-modal mapping into a regression task, which suffers from the regression-to-mean problem leading to over-smoothed facial motions. In this paper, we propose to cast speech-driven facial animation as a code query task in a finite proxy space of the learned codebook, which effectively promotes the vividness of the generated motions by reducing the cross-modal mapping uncertainty. The codebook is learned by self-reconstruction over real facial motions and thus embedded with realistic facial motion priors. Over the discrete motion space, a temporal autoregressive model is employed to sequentially synthesize facial motions from the input speech signal, which guarantees lip-sync as well as plausible facial expressions. We demonstrate that our approach outperforms current state-of-the-art methods both qualitatively and quantitatively. Also, a user study further justifies our superiority in perceptual quality. Code and video demo are available at https://doubiu.github.io/projects/codetalker. + +# 1. Introduction + +3D facial animation has been an active research topic for decades, as attributed to its broad applications in virtual reality, film production, and games. The high correlation between speech and facial gestures (especially lip movements) makes it possible to drive the facial animation with a speech signal. Early attempts are mainly made to build the complex mapping rules between phonemes and their visual counterpart, which usually have limited performance [53,63]. With the advances in deep learning, recent speech-driven facial animation techniques push forward the state-of-the-art significantly. However, it still remains challenging to generate human-like motions. + +As an ill-posed problem, speech-driven facial animation generally has multiple plausible outputs for every input. Such ambiguity tends to cause over-smoothed results. Anyhow, person-specific approaches [29, 49] can usually obtain decent facial motions because of the relatively consistent talking style, but have low scalability to general applications. Recently, VOCA [10] extends these methods to generalize across different identities, however, they generally exhibit mild or static upper face expressions. This is because VOCA formulates the speech-to-motion mapping as a regression task, which encourages averaged motions, especially in the upper face that is only weakly or even uncorrelated to the speech signal. To reduce the uncertainty, FaceFormer [16] utilizes long-term audio context through a transformer-based model and synthesizes the sequential motions in an autoregressive manner. Although it gains important performance promotion, it still inherits the weakness of one-to-one mapping formulation and suffers from a lack of subtle high-frequency motions. Differently, MeshTalk [50] models a categorical latent space for facial animation that disentangles audio-correlated and audio-uncorrelated information so that both aspects could be well handled. Anyway, the employed quantization and categorical latent space representation are not well-suited for motion prior learning, rendering the training tricky and consequently hindering its performance. + +We get inspiration from 3D Face Morphable Model (3DMM) [35], where general facial expressions are represented in a low-dimensional space. Accordingly, we propose to formulate speech-driven facial animation as a code query task in a finite proxy space of the learned discrete codebook prior. The codebook is learned by self-reconstruction over real facial motions using a vector-quantized autoencoder (VQ-VAE) [57], which along with the decoder stores the realistic facial motion priors. In contrast to the continuous linear space of 3DMM, combinations of codebook items form a discrete prior space with only finite cardinality. Still, in the context of the decoder, the code representation possesses high expressiveness. Through mapping the speech to the finite proxy space, the uncertainty of the speech-to-motion mapping is signif + +icantly attenuated and hence promotes the quality of motion synthesis. Conceptually, the proxy space approximates the facial motion space, where the learned codebook items serve as discrete motion primitives. + +Based on the learned discrete codebook, we propose a code-query-based temporal autoregressive model for speech-conditioned facial motion synthesis, called CodeTalker. Specifically, taking a speech signal as input, our model predicts the motion feature tokens in a temporal recursive manner. Then, the feature tokens are used to query the code sequence in the discrete space, followed by facial motion reconstruction. Thanks to the contextual modeling over history motions and cross-modal alignment, the proposed CodeTalker shows the advantages of achieving accurate lip motions and natural expressions. Extensive experiments show that the proposed CodeTalker demonstrates superior performance on existing datasets. Systematic studies and experiments are conducted to demonstrate the merits of our method over previous works. The contributions of our work are as follows: + +- We model the facial motion space with discrete primitives in a novel way, which offers advantages to promote motion synthesis realism against cross-modal uncertainty. +- We propose a discrete motion prior based temporal autoregressive model for speech-driven facial animation, which outperforms existing state-of-the-art methods. + +# 2. Related Works + +# 2.1. Speech-driven 3D Facial Animation + +Computer facial animation is a long-standing task [44] and has attracted rapidly increased interest over the past decades [5, 20, 30, 32, 34, 54, 62, 70]. As a branch, speech-driven facial animation is to reenact a person in sync with input speech sequences. While extensive literature in this field works on 2D talking heads [1, 7-9, 11, 23, 26, 27, 36, 37, 43, 47, 51, 59, 61, 65-67], we focus on facial animation on 3D models in this work, which can be roughly categorized into linguistics-based and learning-based methods. + +Linguistics-based methods. Typically, linguistics-based methods [12, 40, 53, 63] establish a set of complex mapping rules between phonemes and their visual counterparts, i.e., visemes [19, 33, 41]. For example, the dominance function [40] is to determine the influence of phonemes on the respective facial animation control parameters. Xu et al. [63] defines animation curves for a constructed canonical set of visemes to generate synchronized mouth movements. There are also some methods considering the many-to-many mapping between phonemes and visemes, as demonstrated in the dynamic visemes model [53] and, more recently, the JALI [12]. Based on psycholinguistic consid + +erations and built upon the Facial Action Coding System (FACS) [13], JALI factors mouth movements into lip and jaw rig animation and generate compelling co-articulation results. Although these methods have explicit control over the animation, they have complex procedures and lack a principled way to animate the entire face. + +Learning-based methods. Learning-based methods [6, 10, 16, 17, 24, 29, 38, 46, 52, 60] resort to a data-driven framework. Cao et al. [6] achieve emotional lip sync by the proposed constrained search and Anime Graph structure. Recently, Taylor et al. [52] propose a deep-learning-based model utilizing a sliding window approach on the transcribed phoneme sequences input. Karras et al. [29] propose a convolution-based network with a learnable emotion database to animate a speech-driven 3D mesh. More recently, VisemeNet [69] employ a three-stage Long Short-Term Memory (LSTM) network to predict the animation curve for a lower face lip model. + +We review the most related works more concretely here as they have the same setting as this work, i.e., training on high-resolution paired audio-mesh data and speaker-independently animating entire face meshes in vertex space. MeshTalk [50] successfully disentangles audio-correlated and uncorrelated facial information with a categorical latent space. However, the latent space adopted is not optimal with limited expressiveness, thus the animation quality is not stable when applied in a data-scarcity setting. VOCA [10] employs powerful audio feature extraction models and can generate facial animation with different speaking styles. Furthermore, FaceFormer [16] considers long-term audio context with transformer [58] rendering temporally stable animations. Despite the appealing animations, both suffer from the over-smoothing problem, as they directly regress the facial motion in the highly ill-posed audio-visual mapping with large uncertainty and ambiguity. + +# 2.2. Discrete Prior Learning + +In the last decades, discrete prior representation with learned dictionaries has demonstrated its superiority in image restoration tasks [14, 22, 28, 55, 56], since clear image details are well-preserved in the dictionaries. This line of techniques further inspires the high-capacity and high-compressed discrete prior learning. VQ-VAE [57] first presents to learn discrete representations (codebook) of images and autoregressively model their distribution for image synthesis. The follow-up works, VQ-VAE2 [48] and VQGAN [15] further improve the quality of high-resolution image synthesis. Recently, discrete prior learning has been exploited for image colorization [25], inpainting [45], blind face restoration [68], text-to-image synthesis [21], etc. + +In addition to image modality, most recent works also explore the power of discrete prior learning in tasks with other modalities, such as dyadic face motion generation [42], co + +speech gesture synthesis [2], speech enhancement [64]. Inspired by codebook learning, this work investigates to learn discrete motion prior for speech-driven 3D facial animation. Different from [42], we exploit the discrete motion primitives for facial motion representation in a context-rich manner, which is more effective to learn general priors. + +# 3. Method + +We aim to synthesize sequential 3D facial motions from a speech signal, so that any neutral face mesh could be animated as a lip-synchronized talking face. However, this is an ill-posed problem since one speech could be matched by multiple potential facial animations. Such ambiguity tends to make cross-modal learning suffer from averaged motions and lack of subtle variations. To bypass this barrier, we propose to first model the facial motion space with the learned discrete motion prior, and then learn a speech-conditioned temporal autoregressive model over this space, which promotes robustness against the cross-modal uncertainty. + +Formulation. Let $\mathbf{M}_{1:T} = (\mathbf{m}_1, \dots, \mathbf{m}_T)$ be a sequence of facial motions, where each frame $\mathbf{m}_t \in \mathbb{R}^{V \times 3}$ denotes the 3D movement of $V$ vertices over a neutral-face mesh template $\mathbf{h} \in \mathbb{R}^{V \times 3}$ . Let further $\mathbf{A}_{1:T} = (\mathbf{a}_1, \dots, \mathbf{a}_T)$ be a sequence of speech snippets, each of which $\mathbf{a}_t \in \mathbb{R}^d$ has $d$ samples to align with the corresponding (visual) frame $\mathbf{m}_t$ . Then, our goal is to sequentially synthesize $\mathbf{M}_{1:T}$ from $\mathbf{A}_{1:T}$ so that an arbitrary neutral facial template $\mathbf{f}$ could be animated as $\mathbf{H}_{1:T} = \{\mathbf{m}_1 + \mathbf{h}, \dots, \mathbf{m}_T + \mathbf{h}\}$ . + +# 3.1. Discrete Facial Motion Space + +Visual realistic facial animations should present accurate lip motions and natural expressions. To achieve this from speech signals, extra motion priors are required to reduce the uncertainty and complement realistic motion components. As witnessed by the recent image restoration task [68], discrete codebook prior [57] demonstrates advantages in guaranteeing high-fidelity results even from a severely degraded input. Inspired by this, we propose to model the facial motion space as a discrete codebook by learning from tracked real-world facial motions. + +Codebook of motion primitives. We manage to learn a codebook $\mathcal{Z} = \{\mathbf{z}_k\in \mathbb{R}^C\}_{k = 1}^N$ that allows any facial motion $\mathbf{m}_t$ to be represented by a group of allocated items $\{\mathbf{z}_k\}_{k\in S}$ , where $S$ denotes the selected index set through Eq. 1. Conceptually, the codebook items serve as the motion primitives of a facial motion space. To this end, we pre-train a transformer-based VQ-VAE that consists of an encoder $E$ , a decoder $D$ , and a context-rich codebook $\mathcal{Z}$ , under the self-reconstruction of realistic facial motions. As shown in Figure 1, the facial motions $\mathbf{M}_{1:T}$ is first embedded as a temporal feature $\hat{\mathbf{Z}} = E(\mathbf{M}_{1:T})\in \mathbb{R}^{T'\times H\times C}$ , where $H$ is the number of face components and $T^{\prime}$ denotes the number of encoded temporal units ( $P = \frac{T}{T'}$ frames). Then, we + +![](images/59d87b9d9bb2ae868b0ffb210539c242e9a29f751d08f3952822e16248b290c5.jpg) +Figure 1. Learning framework of facial motion space. The learned motion primitives, as embedded in the codebook, serve to represent the facial motions in a spatial and temporal manner. + +obtain the quantized motion sequence $\mathbf{Z}_{\mathbf{q}} \in \mathbb{R}^{T' \times H \times C}$ via an element-wise quantization function $Q(\cdot)$ that maps each item in $\hat{\mathbf{Z}}$ to its nearest entry in codebook $\mathcal{Z}$ : + +$$ +\mathbf {Z} _ {\mathbf {q}} = Q (\hat {\mathbf {Z}}) := \underset {\mathbf {z} _ {k} \in \mathcal {Z}} {\arg \min } \| \hat {\mathbf {z}} _ {t} - \mathbf {z} _ {k} \| _ {2}. \tag {1} +$$ + +Then, the self-reconstruction is given by: + +$$ +\hat {\mathbf {M}} _ {1: T} = D \left(\mathbf {Z} _ {\mathbf {q}}\right) = D \left(Q \left(E \left(\mathbf {M} _ {1: T}\right)\right)\right). \tag {2} +$$ + +Note that, the discrete facial motion space reduces the mapping ambiguity with the finite cardinality, but never sacrifices its expressiveness thanks to its context-rich representation as a latent space. + +Training objectives. Similar to [57], to supervise the quantized autoencoder training, we adopt a motion-level loss and two intermediate code-level losses: + +$$ +\begin{array}{l} \mathcal {L} _ {\mathrm {V Q}} = \left\| \mathbf {M} _ {1: T} - \hat {\mathbf {M}} _ {1: T} \right\| _ {1} \tag {3} \\ + \left\| \operatorname {s g} (\hat {\mathbf {Z}}) - \mathbf {Z} _ {\mathbf {q}} \right\| _ {2} ^ {2} + \beta \left\| \hat {\mathbf {Z}} - \operatorname {s g} (\mathbf {Z} _ {\mathbf {q}}) \right\| _ {2} ^ {2}, \\ \end{array} +$$ + +where the first term is a reconstruction loss, the latter two are adopted to update the codebook items by reducing the distance between the codebook $\mathcal{Z}$ and embedded features $\hat{\mathbf{Z}}$ . $\mathrm{sg}(\cdot)$ stands for a stop-gradient operation and $\beta$ is a weighting factor controlling the update rate of the codebook and encoder. Since the quantization function (Eq. 1) is not differentiable, the straight-through gradient estimator [4,57] is employed to copy the gradients from the decoder input to the encoder output. + +Discussion. Recently, Learn2Listen [42] has applied VQ-VAE for facial expression synthesis in response to a given talking head harnessing 2D monocular videos to obtain 3DMM coefficients. In addition to distinct applications, here we would like to emphasize our major differences. First, Learn2Listen constructs speaker-specific codebooks while ours uses a generic codebook that is feasible to represent arbitrary facial motions. Since cross-character motions are absorbed, our codebook is naturally embedded with more plentiful priors. Second, Learn2Listen utilizes + +![](images/bb53616013f399ae8117acaf337e90ee822fc0adab0aa30f92e5648d4075cea1.jpg) +Figure 2. Diagram of our speech-driven motion synthesis model. Given the speech $\mathbf{A}_{1:T}$ and style vector $\mathbf{s}$ as input, the model learn to recursively generate a sequence of facial motions by predicting the motion codes. As embedded with well-learned motion priors, the pre-trained codebook and decoder are frozen during training. + +![](images/2b293390f136ca97ceaf31c7bd7d135b7650aaffaea82fa6fa1abae0ffc961ed.jpg) +Figure 3. Concept comparison with Learn2Listen [42]. (Top) The speaker-specific facial expression coefficient prior [42], in which each code represents a sequence of facial expression coefficients. (Bottom) Our speaker-agnostic generic motion prior, in which each code represents the motion primitive of face components. The blue dotted boxes indicate what information each code may represent conceptually. + +the codebook to represent common sequences of facial expressions by the way of 3DMM coefficients, i.e., each code represents a sequence (8 frames) of facial expressions. Differently, our codebook is formulated to represent the vertex-based facial motion space, where the codes are embedded with per-vertex motions of facial components and represent the facial motion (within a temporal unit) in a context-rich manner. As compared in Figure 3, the codebook of Learn2Listen is learned to memorize typical sequential facial expressions of a specific speaker within 3DMM space, which cannot synthesize realistic facial motions with subtle details due to the limited expressiveness of 3DMM and is bounded by the accuracy of 3D reconstruction techniques [42]. As the first attempt, our codebook is learned to represent the generic facial motion space with motion primitives for captured facial mesh data, which is more effective to embed general priors preserving vivid facial details. + +We further discuss the hyper-parameters of the codebook. First, the length of the temporal unit $P$ and the number of face components $H$ determine the complexity of the motion primitives in temporal and spatial aspects + +respectively. Generally, complex motion primitives cause low flexibility and reusability and thus hinder representation effectiveness. On the opposite, overly simple motion primitives challenge motion prediction due to the lack of semantics. Besides, the codebook size $N$ and the feature dimension $C$ determine the representation capability, which should be defined according to the complexity of the dataset and in cooperation with $P$ and $H$ . In our experiment, we set $N = 256$ , $P = 1$ , $H = 8$ or $H = 16$ , and $C = 64$ or $C = 128$ depending on the dataset, which lead to high-quality results as justified by the ablation studies in Section 4.5. More details can be found in the Supplement. + +# 3.2. Speech-Driven Motion Synthesis + +With the learned discrete motion prior, we can build a cross-modal mapping from the input speech to the target motion codes that could be further decoded into realistic facial motions. Along with the speech, we further adopt a control on the talking styles as input, i.e., a style vector $\mathbf{s} \in \mathbb{R}_+^M \cup \{0\}$ , where $M$ is the dimension of the learned style space (see Eq. 4). Conditioning on the speech $\mathbf{A}_{1:T}$ and the style vector $\mathbf{s}$ , a temporal autoregressive model, composed of a speech encoder $E_{\mathrm{speech}}$ and a cross-modal decoder $D_{\mathrm{cross-modal}}$ , is employed to learn over the facial motion space, as depicted in Figure 2. + +Following FaceFormer [16], our speech encoder adopts the architecture of the state-of-the-art self-supervised pretrained speech model, wav2vec 2.0 [3], which consists of an audio feature extractor and a multi-layer transformer encoder. The audio feature extractor converts the speech of raw waveform into feature vectors through a temporal convolutions network (TCN). Benefiting from the effective attention scheme, the transformer encoder converts the audio features into contextualized speech representations. Apart from the pre-trained codebook and VQ-VAE decoder, our cross-modal decoder contains an embedding block and a multi-layer transformer decoder with causal self-attention. + +The embedding block combines the past facial motions and the style embedding via: + +$$ +\mathbf {F} _ {\mathrm {e m b}} ^ {1: t - 1} = \mathcal {P} _ {\theta} \left(\hat {\mathbf {M}} _ {1: t - 1}\right) + \mathbf {B} \cdot \frac {\mathbf {S}}{\| \mathbf {S} \| _ {1}}, \tag {4} +$$ + +where $\mathcal{P}_{\theta}$ is a linear projection layer, and $\mathbf{B} = [\mathbf{b}_1,\dots,\mathbf{b}_M]\in \mathbb{R}^{C\times M}$ denotes the $M$ learnable basis vectors that span the style space linearly. Alike to Face-Former [16], we equip the transformer decoder with causal self-attention to learn the dependencies between each frame in the context of the past facial motion sequence, and with cross-modal attention to align the audio and motion modalities. The output features $\hat{\mathbf{Z}}^{1:t}$ is further quantized into $\mathbf{Z}_{\mathbf{q}}^{1:t}$ via Eq. 1 and decoded by the pre-trained VQ-VAE decoder. The newly predicted motion $\hat{\mathbf{m}}_t$ is used to update the past motions as $\hat{\mathbf{M}}_{1:t}$ , in preparation for the next prediction. Formally, this recursive process can be written as: + +$$ +\hat {\mathbf {m}} _ {t} = D _ {\text {c r o s s - m o d a l}} \left(E _ {\text {s p e e c h}} \left(\mathbf {A} _ {1: T}\right), \mathbf {s}, \tilde {\mathbf {M}} _ {1: t - 1}\right). \tag {5} +$$ + +Training objectives. We train the transformer encoder, decoder and the embedding block for cross-modality mapping, while keeping the codebook $\mathcal{Z}$ and motion decoder $D$ frozen. To benefit from the speech representation learning from large-scale corpora, we initialize the TCN and transformer encoder with the pre-trained wav2vec 2.0 weights. Overall, the autoregressive model is trained in a teaching-forcing scheme, under the constraint of two loss terms: (i) feature regularity loss $\mathcal{L}_{\mathrm{reg}}$ measuring the deviation between the predicted motion feature $\hat{\mathbf{Z}}^{1:T}$ and the quantized feature $\mathbf{Z}_{\mathbf{q}}^{1:T}$ from codebook, and (ii) motion loss $\mathcal{L}_{\mathrm{motion}}$ measuring the difference between the predicted motions $\hat{\mathbf{M}}_{1:T}$ and the ground-truth motions $\mathbf{M}_{1:T}$ , which plays an important role to stabilize the training process. The final loss function is: + +$$ +\begin{array}{l} \mathcal {L} _ {\text {s y n}} = \mathcal {L} _ {\text {r e g}} + \mathcal {L} _ {\text {m o t i o n}} \\ = \| \hat {\mathbf {Z}} ^ {1: T} - \operatorname {s g} \left(\mathbf {Z} _ {\mathbf {q}} ^ {1: T}\right) \| _ {2} ^ {2} + \| \hat {\mathbf {M}} _ {1: T} - \mathbf {M} _ {1: T} \| _ {2} ^ {2}. \tag {6} \\ \end{array} +$$ + +# 3.3. Training Details + +At stage one, we train the VQ-VAE model (Figure 1) on a single NVIDIA V100 for 200 epochs ( $\sim$ 2 hours) with the AdamW [39] optimizer ( $\beta_{1} = 0.9$ , $\beta_{2} = 0.999$ and $\epsilon = 1e - 8$ ), where the learning rate is initialized as $10^{-4}$ , and the mini-batch size is set as 1. At stage two, we train the temporal autoregressive model with the Adam optimizer [31]. The training duration is 100 epochs ( $\sim$ 3 hours) and other hyper-parameters remain unchanged as stage one. + +Style embedding space. The style embedding space is linearly spanned by $M$ learned basis vectors, where each style is represented by a style vector $\mathbf{s}$ that serves as the linear combination coefficients or a coordinate. In VOCA [10], it assigns each speaker with a category-like + +one-hot style vector. Instead, we propose a concept to formulate a uniform style space via some learnable basis vectors (Eq. 4), where the style vector is no longer bound with speaker ID but associated with every talking sample. During training, considering the limited style diversity of training datasets, we assign each speaker (e.g. no. $i$ ) with a standard unit vector $\mathbf{e}_i$ as a style vector, under the assumption that each speaker is associated with a unique and consistent style. Anyway, arbitrary style vectors are allowed to interpolate new talking styles during inference. + +# 4. Experiments + +# 4.1. Datasets and Implementations + +We employ two widely used datasets, BIWI [18] and VOCASET [10], to train and test different methods in our experiments. Both datasets contain 4D face scans together with utterances spoken in English. BIWI contains 40 unique sentences shared across all speakers in the dataset, while VOCASET contains 255 unique sentences, which are partially shared among different speakers. + +BIWI dataset. BIWI is a 3D audio-visual corpus of affective speech and facial expression in the form of dense dynamic 3D face geometries, which is originally proposed to study affective communication. There is a total of 40 sentences uttered by 14 subjects, eight females and six males. Each sentence was recorded twice: with and without emotion. On average, each sentence is 4.67 seconds long. The 3D face dynamics are captured at 25fps, each with 23370 vertices and registered topology. We follow the data splits in [16] and use the emotional subset. Specifically, the training set (BIWI-Train) contains 192 sentences, while the validation set (BIWI-Val) contains 24 sentences. There are two testing sets, in which BIWI-Test-A includes 24 sentences spoken by six seen subjects and BIWI-Test-B contains 32 sentences spoken by eight unseen subjects. BIWI-Test-A can be used for both quantitative and qualitative evaluation due to the seen subjects during training, while BIWI-Test-B is more suitable for qualitative evaluation. + +VOCASET dataset. VOCASET is comprised of 480 paired audio-visual sequences recorded from 12 subjects. The facial motion is captured at 60fps and is about 4 seconds long. Different from BIWI, each 3D face mesh is registered to the FLAME [35] topology with 5023 vertices. We adopt the same training (VOCA-Train), validation (VOCA-Val), and testing (VOCA-Test) splits as VOCA [10] and FaceFormer for fair comparisons. + +Implementations. We compare our work with three state-of-the-art methods: VOCA [10], MeshTalk [50] and FaceFormer [16]. We train and test VOCA on BIWI using the official codebase, while directly testing the released model that was trained on VOCASET. For MeshTalk, we train and test it using the official implementation on the two datasets. To compare with FaceFormer, we conduct test- + +Table 1. Quantitative evaluation on BIWI-Test-A. Lower means better for both metrics. + +
MethodLip Vertex Error (×10-4mm)FDD (×10-5mm)
VOCA6.55638.1816
MeshTalk5.91815.1025
FaceFormer5.30774.6408
CodeTalker (Ours)4.79144.1170
+ +ing directly using the pre-trained weights. Among the four methods, VOCA, FaceFormer and our CodeTalker require conditioning on a training speaking style during testing. For unseen subjects, we generate facial animations conditioned on all training styles. More details about the implementation details can be found in the Supplementary Material. + +# 4.2. Quantitative Evaluation + +Following MeshTalk [50] and FaceFormer [16], we adopt the lip vertex error to measure the lip synchronization, which is the only publicly proposed metric for speech-driven facial animation evaluation, to our best knowledge. As a complement, we introduce a new quantitative measurement, i.e., upper-face motion statistics, to evaluate the overall facial dynamics. + +Lip vertex error. It measures the lip deviation of a sequence with respect to the ground truth, i.e., calculating the maximal L2 error of all lip vertices for each frame and takes the average over all frames. + +Upper-face dynamics deviation. The upper-face expression is just loosely correlated with the speech, depending on personal talking styles and the semantics of speech content. With this belief, we propose to measure the variation of facial dynamics for a motion sequence in comparison with that of the ground truth. Specifically, the upper-face dynamics deviation (FDD) is calculated by: + +$$ +\operatorname {F D D} \left(\mathbf {M} _ {1: T}, \hat {\mathbf {M}} _ {1: T}\right) = \frac {\sum_ {v \in \mathcal {S} _ {U}} \left(\operatorname {d y n} \left(\mathbf {M} _ {1 : T} ^ {v}\right) - \operatorname {d y n} \left(\hat {\mathbf {M}} _ {1 : T} ^ {v}\right)\right)}{\left| \mathcal {S} _ {U} \right|}, \tag {7} +$$ + +where $\mathbf{M}_{1:T}^v\in \mathbb{R}^{3\times T}$ denotes the motions of the $v$ -th vertex, and $S_U$ is the index set of upper-face vertices. $\mathrm{dyn}(\cdot)$ denotes the standard deviation of the element-wise L2 norm along the temporal axis. + +We calculate the lip vertex error and upper-face dynamics deviation (FDD) over all sequences in BIWI-Test-A and take the average for comparison. According to Table 1, the proposed CodeTalker achieves lower error than the existing state-of-the-arts, suggesting that it produces more accurate lip-synchronized movements. Besides, Table 1 shows that our CodeTalker achieves the best performance in terms of FDD. It indicates the high consistency between the predicted upper-face expressions together with the trend of facial dynamics (conditioned on the speech and talking styles) + +and those of the ground truth. + +# 4.3. Qualitative Evaluation + +We visually compare our method with other competitors in Figure 4. For fair comparison, we assign the same talking style to VOCA, FaceFormer and our CodeTalker as conditional input, which is sampled at random. To check the lip synchronization performance, we illustrate three typical frames of synthesized facial animations that speak at specific syllables, as compared in the upper partition in Figure 4. We can observe that compared with the competitors, the lip movements produced by our CodeTalker are more accurately articulated with the speech signals and also more consistent with those of the reference. For example, CodeTalker produces better lip sync with proper mouth closures when pronouncing bilabial consonant /b/ (i.e., "bedside" in the upper-right case of Figure 4), compared to VOCA and MeshTalk; for the even challenging speech parts "waterproof" and "shaving" that need to pout, CodeTalker can produce accurate lip shapes while other methods suffer from the over-smoothing problem and fail to lip-sync correctly (Zoom in for better inspection). + +Different from lip movements, facial expressions only have weak correlations with the speech signal, which tends to be static in front of cross-modal mapping ambiguity. To visualize the facial motion dynamics, we calculate the temporal statistics of adjacent-frame facial motions within a sequence. Specifically, we first calculate the inter-frame motion L2 distance and then compute the mean and standard deviation (std) across the sequence at each vertex. The higher mean value indicates stronger facial movements, while the higher std value suggests richer variations of facial dynamics. Two examples are visualized in the last two rows of Figure 4, evidencing that our method outperforms others in achieving both stronger facial movements and a broader range of dynamics. It is mainly attributed to the superiority of the discrete facial motion space, which promotes the robustness to cross-modal uncertainty effectively. Readers are recommended to watch the animation comparisons in the Supplemental Video. + +Talking style interpolation. Our model can synthesize new speaking styles from the learned style embedding space. To inspect the effects on VOCA-Test, we select two speaking style vectors, i.e., $\mathbf{e}_i$ and $\mathbf{e}_j$ , which correspond to large and slight lip articulations respectively, and interpolate new talking style vectors $\mathbf{s}_{\mathrm{new}} = \mathbf{B} \cdot [\omega \mathbf{e}_i + (1 - \omega) \mathbf{e}_j]$ with a linear coefficient $\omega$ . For the synthesized 3D animations of a sampled sequence, we plot the lower-upper lip distances across frames for each style in Figure 5, from which we observe the smooth transition of mouth amplitudes between the two typical styles. It is not only useful to synthesize new talking styles, but also practical to match a specific speaking performance of an unseen subject during training. + +![](images/affde4db5730ff6cffbd1477c7f35d573f095ba6793e02af018a4ecc9ab39454.jpg) +Figure 4. Visual comparisons of sampled facial motions animated by different methods on VOCA-Test (left) and BIWI-Test-B (right). The upper partition shows the facial animation conditioned on different speech parts, while the lower depicts the temporal statistics (mean and standard deviation) of adjacent-frame motion variations within a sequence. + +![](images/8c26e7797a3bfe4ec4341c4bacd29f14d0d1a53b8b130bc088cb5d32ddf9fedf.jpg) + +![](images/4b04dc2ffc959720fb6cf189f9e133b644298f5d20e524c9b1c62ec97ba874a5.jpg) +Figure 5. Distance between lower and upper lip for our predictions within a sequence conditioned on different weighted linear combinations of two style vectors. + +# 4.4. User Study + +The human perception system has been evolutionarily adapted to understanding subtle facial motions and capturing lip synchronization. Thus, it is still the most reliable measure in the speech-driven facial animation task. We conduct a user study to evaluate the quality of animated faces in perceptual lip synchronization and realism, compared with VOCA, MeshTalk, FaceFormer and the ground truth. We adopt A/B tests for each comparison, i.e., ours vs. competitor, in terms of realistic facial animation and lip sync. For BIWI, we obtain the results of four kinds of comparisons by randomly selecting 30 samples from BIWI-Test-B, respectively. To achieve the most variations in terms of speaking + +Table 2. User study results on BIWI-Test-B and VOCA-Test. We adopt A/B testing and report the percentage of answers where A is preferred over B. + +
CompetitorsBIWI-Test-BVOCA-Test
Lip SyncRealismLip SyncRealism
Ours vs. VOCA92.4789.2586.0284.95
Ours vs. MeshTalk80.6582.8095.7092.47
Ours vs. FaceFormer53.7656.9970.9769.89
Ours vs. GT43.0149.4643.0143.01
+ +styles, we ensure the sampling results can fairly cover all conditioning styles. Thus, $120\mathrm{A}$ vs. B pairs (30 samples $\times 4$ comparisons) are created for BIWI-Test-B. Each pair is judged by at least 3 different participants separately, and finally, 372 entries are collected in total. For the user study on VOCASET, we apply the same setting as that on the BIWI dataset, i.e., another $120\mathrm{A}$ vs. B pairs from VOCA-Test set, finally yielding 372 entries as well. In this study, 31 participants with good vision and hearing ability complete the evaluation successfully. Moreover, each participant is involved in all 8 kinds of comparisons to make better exposure and cover the diversity of favorability. + +The percentage of A/B testing in terms of lip sync and realism on BIWI-Test-B is tabulated in Table 2, which shows that participants favor CodeTalker over competitors. Based on the visual analysis in Section 4.3, we attribute this to the + +Table 3. Ablation study on the representation space of codebook. The performance is measured by the reconstruction error (i.e., L2 error) and lip vertex error on VOCA-Test and BIWI-Test-A. + +
VariantsVOCA-TestBIWI-Test-A
Rec. Error (×10-5mm)Rec. Error (×10-5mm)Lip Vertex Error (×10-4mm)
Shape-ent. codebook2.754.076.41
Motion codebook (Ours)0.082.834.79
Speech-driven synthesis"finding Nan's cameo""an excellent room"
Shape-ent. codebookMotion (Ours) codebookShape-ent. codebookMotion (Ours) codebook
+ +Figure 6. Visual comparisons of reconstruction and speech-driven motion synthesis results with different representation spaces on VOCA-Test (left) and BIWI-Test-A (right). + +facial animation synthesized by CodeTalker having more expressive facial motions, accurate lip shape, and well-synchronized mouth movements. For VOCA-Test, which has a nature of fewer upper-face motions, a similar favorability can still be observed in Table 2. We believe the reasons are at least three-fold: subtle motions around the eyes, more accurate lip movements and expressive motions in the lower face. Although be aware of a gap between our predictions and the recorded performance (ground truth), we surprisingly get over $40\%$ preference still. Overall, the user study justifies that the facial animations produced by CodeTalker have superior perceptual quality. + +# 4.5. Ablation Studies + +We study several key designs of our proposed method in this section, including the representation space and the hyper-parameters of the codebook construction. + +Representation space. To study the superiority of our motion-based representation, we construct a baseline that learns a shape-entangled codebook, i.e., the codes represent shapes $(\mathbf{m_t} + \mathbf{h})$ instead of motions $(\mathbf{m_t})$ . As shown in Table 3, the baseline decreases the reconstruction accuracy significantly, which is evidenced by the visualized examples in Figure 6. It is mainly because the shape-entangled sequence contains more speaker-specific information that hinders the reusability of the codes. A direct weakness is the poor generalization for self-reconstruction, which further impedes cross-modal mapping correctness. In contrast, our proposed speaker-agnostic motion representation is more effective to represent generic motion priors shared + +![](images/4ed4c8fd48debac4bccf63b2118f8fae2e9993cf3b6daa8e533b8e563687963c.jpg) +Figure 7. Model performance comparisons with different hyperparameters of the codebook, i.e., the length of temporal unit $P$ and the number of face components $H$ . We measure the reconstruction error and lip vertex error on BIWI-Test-A. + +![](images/e0030efa375101435fceb90c2527f23521101929501c92ed6062e7bc3d39dedf.jpg) + +across individuals, and hence promotes the quality of both self-reconstruction and speech-driven motion synthesis. + +Codebook construction. We further study the hyperparameters used for codebook construction. We evaluate the performance of different settings $\langle P, H \rangle$ by measuring their reconstruction accuracy and cross-modal mapping accuracy (namely lip vertex error). First, we evaluate the reconstruction accuracy as shown in Figure 7(a). On one hand, increasing $P$ degrades the reconstruction accuracy, which could be explained by the increased complexity of the motion to be represented. On the other hand, increasing $H$ eases the reconstruction but risks over-fitting, which explains the general benefits $(H < 8)$ but inferior performance when $H \geq 8$ . Notably, a similar trend could be found in the cross-modal mapping performance, as shown in Figure 7(b). We conjecture that complex motion primitives cause lower reusability and higher redundancy, resulting in ambiguity in the cross-modal code query process. + +# 5. Discussion and Conclusion + +We demonstrated the advantages of casting speech-driven facial animation as a code query task in the discrete space, which notably promotes the motion synthesis quality against cross-modal ambiguity. By comparing to the existing state-of-the-arts, our proposed method shows superiority in achieving accurate lip sync and vivid facial expressions. However, we still follow the assumption that facial motions are independent of shapes, whose rationality may deserve further studies. Also, the overall perceptual quality still lags behind the ground truth, primarily due to the scarcity of paired audio-visual data. Additionally, the acquired generic motion prior adheres to the motion distribution delineated by the training set, which may deviate from real-world facial motions. As a future work, it is intriguing to guide the 3D facial animation by utilizing priors from large-scale available talking head videos. + +Acknowledgements. This project is partially supported by Hong Kong Innovation and Technology Fund (ITF) (ref: ITS/307/20FP and ITS/313/20). + +# References + +[1] Mohammed M Alghamdi, He Wang, Andrew J Bulpitt, and David C Hogg. Talking head from speech audio using a pretrained image generator. In ACM International Conference on Multimedia (MM), 2022. 2 +[2] Tenglong Ao, Qingzhe Gao, Yuke Lou, Baoquan Chen, and Libin Liu. Rhythmic gesticulator: Rhythm-aware co-speech gesture synthesis with hierarchical neural embeddings. ACM Transactions on Graphics (TOG), pages 1-19, 2022. 3 +[3] Alexei Baevski, Yuhao Zhou, Abdelrahman Mohamed, and Michael Auli. wav2vec 2.0: A framework for self-supervised learning of speech representations. In Conference on Neural Information Processing Systems (NeurIPS), 2020. 4 +[4] Yoshua Bengio, Nicholas Léonard, and Aaron Courville. Estimating or propagating gradients through stochastic neurons for conditional computation. arXiv preprint arXiv:1308.3432, 2013. 3 +[5] Chen Cao, Hongzhi Wu, Yanlin Weng, Tianjia Shao, and Kun Zhou. Real-time facial animation with image-based dynamic avatars. ACM Transactions on Graphics (TOG), 35(4):1-12, 2016. 2 +[6] Yong Cao, Wen C Tien, Petros Faloutsos, and Frédéric Pighin. Expressive speech-driven facial animation. ACM Transactions on Graphics (TOG), 24(4):1283-1302, 2005. 2 +[7] Lele Chen, Guofeng Cui, Celong Liu, Zhong Li, Ziyi Kou, Yi Xu, and Chenliang Xu. Talking-head generation with rhythmic head motion. In European Conference on Computer Vision (ECCV), 2020. 2 +[8] Lele Chen, Zhiheng Li, Ross K Maddox, Zhiyao Duan, and Chenliang Xu. Lip movements generation at a glance. In European Conference on Computer Vision (ECCV), 2018. 2 +[9] Joon Son Chung and Andrew Zisserman. Out of time: automated lip sync in the wild. In Asian Conference on Computer Vision (ACCV), 2016. 2 +[10] Daniel Cudeiro, Timo Bolkart, Cassidy Laidlaw, Anurag Ranjan, and Michael J Black. Capture, learning, and synthesis of 3d speaking styles. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 1, 2, 5 +[11] Dipanjan Das, Sandika Biswas, Sanjana Sinha, and Brojeshwar Bhowmick. Speech-driven facial animation using cascaded gans for learning of motion and texture. In European Conference on Computer Vision (ECCV), 2020. 2 +[12] Pif Edwards, Chris Landreth, Eugene Fume, and Karan Singh. Jali: an animator-centric viseme model for expressive lip synchronization. ACM Transactions on Graphics (TOG), 35(4):1-11, 2016. 2 +[13] Paul Ekman and Wallace V Friesen. Facial action coding system. Environmental Psychology & Nonverbal Behavior, 1978. 2 +[14] Michael Elad and Michal Aharon. Image denoising via learned dictionaries and sparse representation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2006. 2 +[15] Patrick Esser, Robin Rombach, and Bjorn Ommer. Taming transformers for high-resolution image synthesis. In + +IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 2 +[16] Yingruo Fan, Zhaojiang Lin, Jun Saito, Wenping Wang, and Taku Komura. Faceformer: Speech-driven 3d facial animation with transformers. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 1, 2, 4, 5, 6 +[17] Yingruo Fan, Zhaojiang Lin, Jun Saito, Wenping Wang, and Taku Komura. Joint audio-text model for expressive speech-driven 3d facial animation. In ACM Symposium on Interactive 3D Graphics and Games (i3D), 2022. 2 +[18] Gabriele Fanelli, Juergen Gall, Harald Romsdorfer, Thibaut Weise, and Luc Van Gool. A 3-d audio-visual corpus of affective communication. IEEE Transactions on Multimedia (TMM), 12(6):591-598, 2010. 5 +[19] Cletus G Fisher. Confusions among visually perceived consonants. Journal of speech and hearing research, 11(4):796-804, 1968. 2 +[20] Ohad Fried, Ayush Tewari, Michael Zollhöfer, Adam Finkelstein, Eli Shechtman, Dan B Goldman, Kyle Genova, Zeyu Jin, Christian Theobalt, and Maneesh Agrawala. Text-based editing of talking-head video. ACM Transactions on Graphics (TOG), 38(4):1-14, 2019. 2 +[21] Shuyang Gu, Dong Chen, Jianmin Bao, Fang Wen, Bo Zhang, Dongdong Chen, Lu Yuan, and Baining Guo. Vector quantized diffusion model for text-to-image synthesis. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2 +[22] Shuhang Gu, Wangmeng Zuo, Qi Xie, Deyu Meng, Xiangchu Feng, and Lei Zhang. Convolutional sparse coding for image super-resolution. In IEEE/CVF International Conference on Computer Vision (ICCV), 2015. 2 +[23] Yudong Guo, Keyu Chen, Sen Liang, Yong-Jin Liu, Hujun Bao, and Juyong Zhang. Ad-nerf: Audio driven neural radiance fields for talking head synthesis. In IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 2 +[24] Ikhsanul Habibie, Weipeng Xu, Dushyant Mehta, Lingjie Liu, Hans-Peter Seidel, Gerard Pons-Moll, Mohamed Elgharib, and Christian Theobalt. Learning speech-driven 3d conversational gestures from video. In ACM International Conference on Intelligent Virtual Agents (IVA), 2021. 2 +[25] Zhitong Huang, Nanxuan Zhao, and Jing Liao. Unicolor: A unified framework for multi-modal colorization with transformer. ACM Transactions on Graphics (TOG), pages 1-16, 2022. 2 +[26] Xinya Ji, Hang Zhou, Kaisiyuan Wang, Qianyi Wu, Wayne Wu, Feng Xu, and Xun Cao. Eamm: One-shot emotional talking face via audio-based emotion-aware motion model. In ACM SIGGRAPH Conference, 2022. 2 +[27] Xinya Ji, Hang Zhou, Kaisiyuan Wang, Wayne Wu, Chen Change Loy, Xun Cao, and Feng Xu. Audio-driven emotional video portraits. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 2 +[28] Younghyun Jo and Seon Joo Kim. Practical single-image super-resolution using look-up table. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 2 + +[29] Tero Karras, Timo Aila, Samuli Laine, Antti Herva, and Jaakko Lehtinen. Audio-driven facial animation by joint end-to-end learning of pose and emotion. ACM Transactions on Graphics (TOG), 36(4):1-12, 2017. 1, 2 +[30] Hyeongwoo Kim, Pablo Garrido, Ayush Tewari, Weipeng Xu, Justus Thies, Matthias Niessner, Patrick Pérez, Christian Richardt, Michael Zollhöfer, and Christian Theobalt. Deep video portraits. ACM Transactions on Graphics (TOG), 37(4):1-14, 2018. 2 +[31] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In International Conference on Learning Representations (ICLR), 2014. 5 +[32] Avisek Lahiri, Vivek Kwatra, Christian Frueh, John Lewis, and Chris Bregler. Lipsync3d: Data-efficient learning of personalized 3d talking faces from video using pose and lighting normalization. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 2 +[33] John Lewis. Automated lip-sync: Background and techniques. The Journal of Visualization and Computer Animation, 2(4):118-122, 1991. 2 +[34] Hao Li, Jihun Yu, Yuting Ye, and Chris Bregler. Realtime facial animation with on-the-fly correctives. ACM Transactions on Graphics (TOG), 32(4):42-1, 2013. 2 +[35] Tianye Li, Timo Bolkart, Michael J Black, Hao Li, and Javier Romero. Learning a model of facial shape and expression from 4d scans. ACM Transactions on Graphics (TOG), 36(6):194-1, 2017. 1, 5 +[36] Borong Liang, Yan Pan, Zhizhi Guo, Hang Zhou, Zhibin Hong, Xiaoguang Han, Junyu Han, Jingtuo Liu, Errui Ding, and Jingdong Wang. Expressive talking head generation with granular audio-visual control. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2 +[37] Xian Liu, Yinghao Xu, Qianyi Wu, Hang Zhou, Wayne Wu, and Bolei Zhou. Semantic-aware implicit neural auditory-driven video portrait generation. In European Conference on Computer Vision (ECCV), 2022. 2 +[38] Yilong Liu, Feng Xu, Jinxiang Chai, Xin Tong, Lijuan Wang, and Qiang Huo. Video-audio driven real-time facial animation. ACM Transactions on Graphics (TOG), 34(6):1-10, 2015. 2 +[39] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In International Conference on Learning Representations (ICLR), 2019. 5 +[40] DW Massaro, MM Cohen, M Tabain, J Beskow, and R Clark. Animated speech: research progress and applications. Audiovisual Speech Processing, pages 309-345, 2012. 2 +[41] Wesley Mattheyses and Werner Verhelst. Audiovisual speech synthesis: An overview of the state-of-the-art. Speech Communication, 66:182-217, 2015. 2 +[42] Evonne Ng, Hanbyul Joo, Liwen Hu, Hao Li, Trevor Darrell, Angjoo Kanazawa, and Shiry Ginosar. Learning to listen: Modeling non-deterministic dyadic facial motion. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2, 3, 4 +[43] Youxin Pang, Yong Zhang, Weize Quan, Yanbo Fan, Xiaodong Cun, Ying Shan, and Dong-ming Yan. Dpe: Disentanglement of pose and expression for general video portrait editing. arXiv preprint arXiv:2301.06281, 2023. 2 + +[44] Frederic I Parke and Keith Waters. Computer facial animation. CRC press, 2008. 2 +[45] Jialun Peng, Dong Liu, Songcen Xu, and Houqiang Li. Generating diverse structure for image inpainting with hierarchical vq-vae. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 2 +[46] Hai Xuan Pham, Yuting Wang, and Vladimir Pavlovic. End-to-end learning for 3d facial animation from speech. In ACM International Conference on Multimodal Interaction (ICMI), 2018. 2 +[47] KR Prajwal, Rudrabha Mukhopadhyay, Vinay P Namboodiri, and CV Jawahar. A lip sync expert is all you need for speech to lip generation in the wild. In ACM International Conference on Multimedia (MM), 2020. 2 +[48] Ali Razavi, Aaron Van den Oord, and Oriol Vinyals. Generating diverse high-fidelity images with vq-vae-2. In Conference on Neural Information Processing Systems (NeurIPS), 2019. 2 +[49] Alexander Richard, Colin Lea, Shugao Ma, Jurgen Gall, Fernando De la Torre, and Yaser Sheikh. Audio-and gaze-driven facial animation of codec avatars. In IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), 2021. 1 +[50] Alexander Richard, Michael Zollhöfer, Yandong Wen, Fernando De la Torre, and Yaser Sheikh. Meshtalk: 3d face animation from speech using cross-modality disentanglement. In IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 1, 2, 5, 6 +[51] Shuai Shen, Wanhua Li, Zheng Zhu, Yueqi Duan, Jie Zhou, and Jiwen Lu. Learning dynamic facial radiance fields for few-shot talking head synthesis. In European Conference on Computer Vision (ECCV), 2022. 2 +[52] Sarah Taylor, Taehwan Kim, Yisong Yue, Moshe Mahler, James Krahe, Anastasio Garcia Rodriguez, Jessica Hodgins, and Iain Matthews. A deep learning approach for generalized speech animation. ACM Transactions on Graphics (TOG), 36(4):1-11, 2017. 2 +[53] Sarah L Taylor, Moshe Mahler, Barry-John Theobald, and Iain Matthews. Dynamic units of visual speech. In ACM/Eurographics Symposium on Computer Animation (SCA), 2012. 1, 2 +[54] Justus Thies, Mohamed Elgharib, Ayush Tewari, Christian Theobalt, and Matthias Nießner. Neural voice puppetry: Audio-driven facial reenactment. In European Conference on Computer Vision (ECCV), 2020. 2 +[55] Radu Timofte, Vincent De Smet, and Luc Van Gool. Anchored neighborhood regression for fast example-based super-resolution. In IEEE/CVF International Conference on Computer Vision (ICCV), 2013. 2 +[56] Radu Timofte, Vincent De Smet, and Luc Van Gool. A+: Adjusted anchored neighborhood regression for fast super-resolution. In Asian Conference on Computer Vision (ACCV), 2014. 2 +[57] Aaron Van Den Oord, Oriol Vinyals, et al. Neural discrete representation learning. Conference on Neural Information Processing Systems (NeurIPS), 2017. 1, 2, 3 +[58] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia + +Polosukhin. Attention is all you need. Conference on Neural Information Processing Systems (NeurIPS), 2017. 2 +[59] Konstantinos Vougioukas, Stavros Petridis, and Maja Pantic. Realistic speech-driven facial animation with gans. International Journal of Computer Vision (IJCV), 128(5):1398-1413, 2020. 2 +[60] Qianyun Wang, Zhenfeng Fan, and Shihong Xia. 3d-talkemo: Learning to synthesize 3d emotional talking head. arXiv preprint arXiv:2104.12051, 2021. 2 +[61] Suzhen Wang, Lincheng Li, Yu Ding, and Xin Yu. One-shot talking face generation from single-speaker audio-visual correlation learning. In AAAI Conference on Artificial Intelligence (AAAI), 2022. 2 +[62] Thibaut Weise, Sofien Bouaziz, Hao Li, and Mark Pauly. Realtime performance-based facial animation. ACM Transactions on Graphics (TOG), 30(4):1-10, 2011. 2 +[63] Yuyu Xu, Andrew W Feng, Stacy Marsella, and Ari Shapiro. A practical and configurable lip sync method for games. In Proceedings of Motion on Games, 2013. 1, 2 +[64] Karren Yang, Dejan Markovic, Steven Krenn, Vasu Agrawal, and Alexander Richard. Audio-visual speech codes: Rethinking audio-visual speech enhancement by re-synthesis. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 3 +[65] Ran Yi, Zipeng Ye, Juyong Zhang, Hujun Bao, and Yong-Jin Liu. Audio-driven talking face video generation with learning-based personalized head pose. arXiv preprint arXiv:2002.10137, 2020. 2 +[66] Bowen Zhang, Chenyang Qi, Pan Zhang, Bo Zhang, Hsiang-Tao Wu, Dong Chen, Qifeng Chen, Yong Wang, and Fang Wen. Metaportrait: Identity-preserving talking head generation with fast personalized adaptation. arXiv preprint arXiv:2212.08062, 2022. 2 +[67] Hang Zhou, Yasheng Sun, Wayne Wu, Chen Change Loy, Xiaogang Wang, and Ziwei Liu. Pose-controllable talking face generation by implicitly modularized audio-visual representation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 2 +[68] Shangchen Zhou, Kelvin CK Chan, Chongyi Li, and Chen Change Loy. Towards robust blind face restoration with codebook lookup transformer. Conference on Neural Information Processing Systems (NeurIPS), 2022. 2, 3 +[69] Yang Zhou, Zhan Xu, Chris Landreth, Evangelos Kalogerakis, Subhransu Maji, and Karan Singh. Visemenet: Audio-driven animator-centric speech animation. ACM Transactions on Graphics (TOG), 37(4):1-10, 2018. 2 +[70] Michael Zollhöfer, Justus Thies, Pablo Garrido, Derek Bradley, Thabo Beeler, Patrick Pérez, Marc Stamminger, Matthias Nießner, and Christian Theobalt. State of the art on monocular 3d face reconstruction, tracking, and applications. Computer Graphics Forum (CGF), 37(2):523-550, 2018. 2 \ No newline at end of file diff --git a/2023/CodeTalker_ Speech-Driven 3D Facial Animation With Discrete Motion Prior/images.zip b/2023/CodeTalker_ Speech-Driven 3D Facial Animation With Discrete Motion Prior/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..f0c9bbd05bb0f65eee0962348746e9352db7052d --- /dev/null +++ b/2023/CodeTalker_ Speech-Driven 3D Facial Animation With Discrete Motion Prior/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:46c4a3a8fe3e25ae899fef0be30f4bb20fe842bc9024f42ac7cf9cc325e48822 +size 427221 diff --git a/2023/CodeTalker_ Speech-Driven 3D Facial Animation With Discrete Motion Prior/layout.json b/2023/CodeTalker_ Speech-Driven 3D Facial Animation With Discrete Motion Prior/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..6989d9f088c6bcc8dba8b11f79b14d9c996adabb --- /dev/null +++ b/2023/CodeTalker_ Speech-Driven 3D Facial Animation With Discrete Motion Prior/layout.json @@ -0,0 +1,8931 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 60, + 103, + 533, + 120 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 60, + 103, + 533, + 120 + ], + "spans": [ + { + "bbox": [ + 60, + 103, + 533, + 120 + ], + "type": "text", + "content": "CodeTalker: Speech-Driven 3D Facial Animation with Discrete Motion Prior" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 142, + 531, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 142, + 531, + 171 + ], + "spans": [ + { + "bbox": [ + 61, + 142, + 531, + 171 + ], + "type": "text", + "content": "Jinbo Xing" + }, + { + "bbox": [ + 61, + 142, + 531, + 171 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 61, + 142, + 531, + 171 + ], + "type": "text", + "content": " Menghan Xia" + }, + { + "bbox": [ + 61, + 142, + 531, + 171 + ], + "type": "inline_equation", + "content": "^{2,*}" + }, + { + "bbox": [ + 61, + 142, + 531, + 171 + ], + "type": "text", + "content": " Yuechen Zhang" + }, + { + "bbox": [ + 61, + 142, + 531, + 171 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 61, + 142, + 531, + 171 + ], + "type": "text", + "content": " Xiaodong Cun" + }, + { + "bbox": [ + 61, + 142, + 531, + 171 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 61, + 142, + 531, + 171 + ], + "type": "text", + "content": " Jue Wang" + }, + { + "bbox": [ + 61, + 142, + 531, + 171 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 61, + 142, + 531, + 171 + ], + "type": "text", + "content": " Tien-Tsin Wong" + }, + { + "bbox": [ + 61, + 142, + 531, + 171 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 61, + 142, + 531, + 171 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 61, + 142, + 531, + 171 + ], + "type": "text", + "content": "The Chinese University of Hong Kong \n" + }, + { + "bbox": [ + 61, + 142, + 531, + 171 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 61, + 142, + 531, + 171 + ], + "type": "text", + "content": "Tencent AI Lab" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 67, + 173, + 289, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 173, + 289, + 186 + ], + "spans": [ + { + "bbox": [ + 67, + 173, + 289, + 186 + ], + "type": "text", + "content": "{jbxing,yczhang21,ttwong}@cse.cuhk.edu.hk" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 319, + 174, + 525, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 174, + 525, + 185 + ], + "spans": [ + { + "bbox": [ + 319, + 174, + 525, + 185 + ], + "type": "text", + "content": "{menghanxyz,vinthony,arphid}@gmail.com" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 143, + 213, + 192, + 226 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 213, + 192, + 226 + ], + "spans": [ + { + "bbox": [ + 143, + 213, + 192, + 226 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 239, + 289, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 239, + 289, + 502 + ], + "spans": [ + { + "bbox": [ + 46, + 239, + 289, + 502 + ], + "type": "text", + "content": "Speech-driven 3D facial animation has been widely studied, yet there is still a gap to achieving realism and vividness due to the highly ill-posed nature and scarcity of audiovisual data. Existing works typically formulate the cross-modal mapping into a regression task, which suffers from the regression-to-mean problem leading to over-smoothed facial motions. In this paper, we propose to cast speech-driven facial animation as a code query task in a finite proxy space of the learned codebook, which effectively promotes the vividness of the generated motions by reducing the cross-modal mapping uncertainty. The codebook is learned by self-reconstruction over real facial motions and thus embedded with realistic facial motion priors. Over the discrete motion space, a temporal autoregressive model is employed to sequentially synthesize facial motions from the input speech signal, which guarantees lip-sync as well as plausible facial expressions. We demonstrate that our approach outperforms current state-of-the-art methods both qualitatively and quantitatively. Also, a user study further justifies our superiority in perceptual quality. Code and video demo are available at https://doubiu.github.io/projects/codetalker." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 529, + 128, + 541 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 529, + 128, + 541 + ], + "spans": [ + { + "bbox": [ + 47, + 529, + 128, + 541 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 549, + 288, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 549, + 288, + 693 + ], + "spans": [ + { + "bbox": [ + 46, + 549, + 288, + 693 + ], + "type": "text", + "content": "3D facial animation has been an active research topic for decades, as attributed to its broad applications in virtual reality, film production, and games. The high correlation between speech and facial gestures (especially lip movements) makes it possible to drive the facial animation with a speech signal. Early attempts are mainly made to build the complex mapping rules between phonemes and their visual counterpart, which usually have limited performance [53,63]. With the advances in deep learning, recent speech-driven facial animation techniques push forward the state-of-the-art significantly. However, it still remains challenging to generate human-like motions." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 214, + 547, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 214, + 547, + 525 + ], + "spans": [ + { + "bbox": [ + 306, + 214, + 547, + 525 + ], + "type": "text", + "content": "As an ill-posed problem, speech-driven facial animation generally has multiple plausible outputs for every input. Such ambiguity tends to cause over-smoothed results. Anyhow, person-specific approaches [29, 49] can usually obtain decent facial motions because of the relatively consistent talking style, but have low scalability to general applications. Recently, VOCA [10] extends these methods to generalize across different identities, however, they generally exhibit mild or static upper face expressions. This is because VOCA formulates the speech-to-motion mapping as a regression task, which encourages averaged motions, especially in the upper face that is only weakly or even uncorrelated to the speech signal. To reduce the uncertainty, FaceFormer [16] utilizes long-term audio context through a transformer-based model and synthesizes the sequential motions in an autoregressive manner. Although it gains important performance promotion, it still inherits the weakness of one-to-one mapping formulation and suffers from a lack of subtle high-frequency motions. Differently, MeshTalk [50] models a categorical latent space for facial animation that disentangles audio-correlated and audio-uncorrelated information so that both aspects could be well handled. Anyway, the employed quantization and categorical latent space representation are not well-suited for motion prior learning, rendering the training tricky and consequently hindering its performance." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 533, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 533, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 533, + 547, + 715 + ], + "type": "text", + "content": "We get inspiration from 3D Face Morphable Model (3DMM) [35], where general facial expressions are represented in a low-dimensional space. Accordingly, we propose to formulate speech-driven facial animation as a code query task in a finite proxy space of the learned discrete codebook prior. The codebook is learned by self-reconstruction over real facial motions using a vector-quantized autoencoder (VQ-VAE) [57], which along with the decoder stores the realistic facial motion priors. In contrast to the continuous linear space of 3DMM, combinations of codebook items form a discrete prior space with only finite cardinality. Still, in the context of the decoder, the code representation possesses high expressiveness. Through mapping the speech to the finite proxy space, the uncertainty of the speech-to-motion mapping is signif" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 702, + 144, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 702, + 144, + 713 + ], + "spans": [ + { + "bbox": [ + 61, + 702, + 144, + 713 + ], + "type": "text", + "content": "* Corresponding Author." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "12780" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 286, + 119 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 286, + 119 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 286, + 119 + ], + "type": "text", + "content": "icantly attenuated and hence promotes the quality of motion synthesis. Conceptually, the proxy space approximates the facial motion space, where the learned codebook items serve as discrete motion primitives." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 121, + 286, + 310 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 121, + 286, + 310 + ], + "spans": [ + { + "bbox": [ + 48, + 121, + 286, + 310 + ], + "type": "text", + "content": "Based on the learned discrete codebook, we propose a code-query-based temporal autoregressive model for speech-conditioned facial motion synthesis, called CodeTalker. Specifically, taking a speech signal as input, our model predicts the motion feature tokens in a temporal recursive manner. Then, the feature tokens are used to query the code sequence in the discrete space, followed by facial motion reconstruction. Thanks to the contextual modeling over history motions and cross-modal alignment, the proposed CodeTalker shows the advantages of achieving accurate lip motions and natural expressions. Extensive experiments show that the proposed CodeTalker demonstrates superior performance on existing datasets. Systematic studies and experiments are conducted to demonstrate the merits of our method over previous works. The contributions of our work are as follows:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 59, + 320, + 286, + 411 + ], + "type": "list", + "angle": 0, + "index": 4, + "blocks": [ + { + "bbox": [ + 59, + 320, + 286, + 367 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 320, + 286, + 367 + ], + "spans": [ + { + "bbox": [ + 59, + 320, + 286, + 367 + ], + "type": "text", + "content": "- We model the facial motion space with discrete primitives in a novel way, which offers advantages to promote motion synthesis realism against cross-modal uncertainty." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 59, + 376, + 286, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 376, + 286, + 411 + ], + "spans": [ + { + "bbox": [ + 59, + 376, + 286, + 411 + ], + "type": "text", + "content": "- We propose a discrete motion prior based temporal autoregressive model for speech-driven facial animation, which outperforms existing state-of-the-art methods." + } + ] + } + ], + "index": 3 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 48, + 423, + 138, + 436 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 423, + 138, + 436 + ], + "spans": [ + { + "bbox": [ + 48, + 423, + 138, + 436 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 443, + 235, + 455 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 443, + 235, + 455 + ], + "spans": [ + { + "bbox": [ + 48, + 443, + 235, + 455 + ], + "type": "text", + "content": "2.1. Speech-driven 3D Facial Animation" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 462, + 286, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 462, + 286, + 568 + ], + "spans": [ + { + "bbox": [ + 48, + 462, + 286, + 568 + ], + "type": "text", + "content": "Computer facial animation is a long-standing task [44] and has attracted rapidly increased interest over the past decades [5, 20, 30, 32, 34, 54, 62, 70]. As a branch, speech-driven facial animation is to reenact a person in sync with input speech sequences. While extensive literature in this field works on 2D talking heads [1, 7-9, 11, 23, 26, 27, 36, 37, 43, 47, 51, 59, 61, 65-67], we focus on facial animation on 3D models in this work, which can be roughly categorized into linguistics-based and learning-based methods." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 570, + 286, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 570, + 286, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 570, + 286, + 712 + ], + "type": "text", + "content": "Linguistics-based methods. Typically, linguistics-based methods [12, 40, 53, 63] establish a set of complex mapping rules between phonemes and their visual counterparts, i.e., visemes [19, 33, 41]. For example, the dominance function [40] is to determine the influence of phonemes on the respective facial animation control parameters. Xu et al. [63] defines animation curves for a constructed canonical set of visemes to generate synchronized mouth movements. There are also some methods considering the many-to-many mapping between phonemes and visemes, as demonstrated in the dynamic visemes model [53] and, more recently, the JALI [12]. Based on psycholinguistic consid" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 307, + 72, + 545, + 143 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 72, + 545, + 143 + ], + "spans": [ + { + "bbox": [ + 307, + 72, + 545, + 143 + ], + "type": "text", + "content": "erations and built upon the Facial Action Coding System (FACS) [13], JALI factors mouth movements into lip and jaw rig animation and generate compelling co-articulation results. Although these methods have explicit control over the animation, they have complex procedures and lack a principled way to animate the entire face." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 308, + 144, + 545, + 288 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 144, + 545, + 288 + ], + "spans": [ + { + "bbox": [ + 308, + 144, + 545, + 288 + ], + "type": "text", + "content": "Learning-based methods. Learning-based methods [6, 10, 16, 17, 24, 29, 38, 46, 52, 60] resort to a data-driven framework. Cao et al. [6] achieve emotional lip sync by the proposed constrained search and Anime Graph structure. Recently, Taylor et al. [52] propose a deep-learning-based model utilizing a sliding window approach on the transcribed phoneme sequences input. Karras et al. [29] propose a convolution-based network with a learnable emotion database to animate a speech-driven 3D mesh. More recently, VisemeNet [69] employ a three-stage Long Short-Term Memory (LSTM) network to predict the animation curve for a lower face lip model." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 308, + 289, + 545, + 492 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 289, + 545, + 492 + ], + "spans": [ + { + "bbox": [ + 308, + 289, + 545, + 492 + ], + "type": "text", + "content": "We review the most related works more concretely here as they have the same setting as this work, i.e., training on high-resolution paired audio-mesh data and speaker-independently animating entire face meshes in vertex space. MeshTalk [50] successfully disentangles audio-correlated and uncorrelated facial information with a categorical latent space. However, the latent space adopted is not optimal with limited expressiveness, thus the animation quality is not stable when applied in a data-scarcity setting. VOCA [10] employs powerful audio feature extraction models and can generate facial animation with different speaking styles. Furthermore, FaceFormer [16] considers long-term audio context with transformer [58] rendering temporally stable animations. Despite the appealing animations, both suffer from the over-smoothing problem, as they directly regress the facial motion in the highly ill-posed audio-visual mapping with large uncertainty and ambiguity." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 308, + 502, + 440, + 514 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 502, + 440, + 514 + ], + "spans": [ + { + "bbox": [ + 308, + 502, + 440, + 514 + ], + "type": "text", + "content": "2.2. Discrete Prior Learning" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 308, + 521, + 545, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 521, + 545, + 676 + ], + "spans": [ + { + "bbox": [ + 308, + 521, + 545, + 676 + ], + "type": "text", + "content": "In the last decades, discrete prior representation with learned dictionaries has demonstrated its superiority in image restoration tasks [14, 22, 28, 55, 56], since clear image details are well-preserved in the dictionaries. This line of techniques further inspires the high-capacity and high-compressed discrete prior learning. VQ-VAE [57] first presents to learn discrete representations (codebook) of images and autoregressively model their distribution for image synthesis. The follow-up works, VQ-VAE2 [48] and VQGAN [15] further improve the quality of high-resolution image synthesis. Recently, discrete prior learning has been exploited for image colorization [25], inpainting [45], blind face restoration [68], text-to-image synthesis [21], etc." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 308, + 677, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 677, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 308, + 677, + 545, + 712 + ], + "type": "text", + "content": "In addition to image modality, most recent works also explore the power of discrete prior learning in tasks with other modalities, such as dyadic face motion generation [42], co" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "12781" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 145 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 145 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 145 + ], + "type": "text", + "content": "speech gesture synthesis [2], speech enhancement [64]. Inspired by codebook learning, this work investigates to learn discrete motion prior for speech-driven 3D facial animation. Different from [42], we exploit the discrete motion primitives for facial motion representation in a context-rich manner, which is more effective to learn general priors." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 152, + 103, + 164 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 152, + 103, + 164 + ], + "spans": [ + { + "bbox": [ + 47, + 152, + 103, + 164 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 170, + 287, + 302 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 170, + 287, + 302 + ], + "spans": [ + { + "bbox": [ + 46, + 170, + 287, + 302 + ], + "type": "text", + "content": "We aim to synthesize sequential 3D facial motions from a speech signal, so that any neutral face mesh could be animated as a lip-synchronized talking face. However, this is an ill-posed problem since one speech could be matched by multiple potential facial animations. Such ambiguity tends to make cross-modal learning suffer from averaged motions and lack of subtle variations. To bypass this barrier, we propose to first model the facial motion space with the learned discrete motion prior, and then learn a speech-conditioned temporal autoregressive model over this space, which promotes robustness against the cross-modal uncertainty." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 304, + 288, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 304, + 288, + 413 + ], + "spans": [ + { + "bbox": [ + 46, + 304, + 288, + 413 + ], + "type": "text", + "content": "Formulation. Let " + }, + { + "bbox": [ + 46, + 304, + 288, + 413 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_{1:T} = (\\mathbf{m}_1, \\dots, \\mathbf{m}_T)" + }, + { + "bbox": [ + 46, + 304, + 288, + 413 + ], + "type": "text", + "content": " be a sequence of facial motions, where each frame " + }, + { + "bbox": [ + 46, + 304, + 288, + 413 + ], + "type": "inline_equation", + "content": "\\mathbf{m}_t \\in \\mathbb{R}^{V \\times 3}" + }, + { + "bbox": [ + 46, + 304, + 288, + 413 + ], + "type": "text", + "content": " denotes the 3D movement of " + }, + { + "bbox": [ + 46, + 304, + 288, + 413 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 46, + 304, + 288, + 413 + ], + "type": "text", + "content": " vertices over a neutral-face mesh template " + }, + { + "bbox": [ + 46, + 304, + 288, + 413 + ], + "type": "inline_equation", + "content": "\\mathbf{h} \\in \\mathbb{R}^{V \\times 3}" + }, + { + "bbox": [ + 46, + 304, + 288, + 413 + ], + "type": "text", + "content": ". Let further " + }, + { + "bbox": [ + 46, + 304, + 288, + 413 + ], + "type": "inline_equation", + "content": "\\mathbf{A}_{1:T} = (\\mathbf{a}_1, \\dots, \\mathbf{a}_T)" + }, + { + "bbox": [ + 46, + 304, + 288, + 413 + ], + "type": "text", + "content": " be a sequence of speech snippets, each of which " + }, + { + "bbox": [ + 46, + 304, + 288, + 413 + ], + "type": "inline_equation", + "content": "\\mathbf{a}_t \\in \\mathbb{R}^d" + }, + { + "bbox": [ + 46, + 304, + 288, + 413 + ], + "type": "text", + "content": " has " + }, + { + "bbox": [ + 46, + 304, + 288, + 413 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 46, + 304, + 288, + 413 + ], + "type": "text", + "content": " samples to align with the corresponding (visual) frame " + }, + { + "bbox": [ + 46, + 304, + 288, + 413 + ], + "type": "inline_equation", + "content": "\\mathbf{m}_t" + }, + { + "bbox": [ + 46, + 304, + 288, + 413 + ], + "type": "text", + "content": ". Then, our goal is to sequentially synthesize " + }, + { + "bbox": [ + 46, + 304, + 288, + 413 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_{1:T}" + }, + { + "bbox": [ + 46, + 304, + 288, + 413 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 46, + 304, + 288, + 413 + ], + "type": "inline_equation", + "content": "\\mathbf{A}_{1:T}" + }, + { + "bbox": [ + 46, + 304, + 288, + 413 + ], + "type": "text", + "content": " so that an arbitrary neutral facial template " + }, + { + "bbox": [ + 46, + 304, + 288, + 413 + ], + "type": "inline_equation", + "content": "\\mathbf{f}" + }, + { + "bbox": [ + 46, + 304, + 288, + 413 + ], + "type": "text", + "content": " could be animated as " + }, + { + "bbox": [ + 46, + 304, + 288, + 413 + ], + "type": "inline_equation", + "content": "\\mathbf{H}_{1:T} = \\{\\mathbf{m}_1 + \\mathbf{h}, \\dots, \\mathbf{m}_T + \\mathbf{h}\\}" + }, + { + "bbox": [ + 46, + 304, + 288, + 413 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 418, + 206, + 430 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 418, + 206, + 430 + ], + "spans": [ + { + "bbox": [ + 47, + 418, + 206, + 430 + ], + "type": "text", + "content": "3.1. Discrete Facial Motion Space" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 436, + 287, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 436, + 287, + 555 + ], + "spans": [ + { + "bbox": [ + 46, + 436, + 287, + 555 + ], + "type": "text", + "content": "Visual realistic facial animations should present accurate lip motions and natural expressions. To achieve this from speech signals, extra motion priors are required to reduce the uncertainty and complement realistic motion components. As witnessed by the recent image restoration task [68], discrete codebook prior [57] demonstrates advantages in guaranteeing high-fidelity results even from a severely degraded input. Inspired by this, we propose to model the facial motion space as a discrete codebook by learning from tracked real-world facial motions." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 558, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 558, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 558, + 288, + 715 + ], + "type": "text", + "content": "Codebook of motion primitives. We manage to learn a codebook " + }, + { + "bbox": [ + 46, + 558, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{Z} = \\{\\mathbf{z}_k\\in \\mathbb{R}^C\\}_{k = 1}^N" + }, + { + "bbox": [ + 46, + 558, + 288, + 715 + ], + "type": "text", + "content": " that allows any facial motion " + }, + { + "bbox": [ + 46, + 558, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{m}_t" + }, + { + "bbox": [ + 46, + 558, + 288, + 715 + ], + "type": "text", + "content": " to be represented by a group of allocated items " + }, + { + "bbox": [ + 46, + 558, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{z}_k\\}_{k\\in S}" + }, + { + "bbox": [ + 46, + 558, + 288, + 715 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 558, + 288, + 715 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 46, + 558, + 288, + 715 + ], + "type": "text", + "content": " denotes the selected index set through Eq. 1. Conceptually, the codebook items serve as the motion primitives of a facial motion space. To this end, we pre-train a transformer-based VQ-VAE that consists of an encoder " + }, + { + "bbox": [ + 46, + 558, + 288, + 715 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 46, + 558, + 288, + 715 + ], + "type": "text", + "content": ", a decoder " + }, + { + "bbox": [ + 46, + 558, + 288, + 715 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 46, + 558, + 288, + 715 + ], + "type": "text", + "content": ", and a context-rich codebook " + }, + { + "bbox": [ + 46, + 558, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{Z}" + }, + { + "bbox": [ + 46, + 558, + 288, + 715 + ], + "type": "text", + "content": ", under the self-reconstruction of realistic facial motions. As shown in Figure 1, the facial motions " + }, + { + "bbox": [ + 46, + 558, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_{1:T}" + }, + { + "bbox": [ + 46, + 558, + 288, + 715 + ], + "type": "text", + "content": " is first embedded as a temporal feature " + }, + { + "bbox": [ + 46, + 558, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{Z}} = E(\\mathbf{M}_{1:T})\\in \\mathbb{R}^{T'\\times H\\times C}" + }, + { + "bbox": [ + 46, + 558, + 288, + 715 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 558, + 288, + 715 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 46, + 558, + 288, + 715 + ], + "type": "text", + "content": " is the number of face components and " + }, + { + "bbox": [ + 46, + 558, + 288, + 715 + ], + "type": "inline_equation", + "content": "T^{\\prime}" + }, + { + "bbox": [ + 46, + 558, + 288, + 715 + ], + "type": "text", + "content": " denotes the number of encoded temporal units (" + }, + { + "bbox": [ + 46, + 558, + 288, + 715 + ], + "type": "inline_equation", + "content": "P = \\frac{T}{T'}" + }, + { + "bbox": [ + 46, + 558, + 288, + 715 + ], + "type": "text", + "content": " frames). Then, we" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 307, + 69, + 547, + 185 + ], + "blocks": [ + { + "bbox": [ + 307, + 69, + 547, + 185 + ], + "lines": [ + { + "bbox": [ + 307, + 69, + 547, + 185 + ], + "spans": [ + { + "bbox": [ + 307, + 69, + 547, + 185 + ], + "type": "image", + "image_path": "59d87b9d9bb2ae868b0ffb210539c242e9a29f751d08f3952822e16248b290c5.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 192, + 547, + 226 + ], + "lines": [ + { + "bbox": [ + 305, + 192, + 547, + 226 + ], + "spans": [ + { + "bbox": [ + 305, + 192, + 547, + 226 + ], + "type": "text", + "content": "Figure 1. Learning framework of facial motion space. The learned motion primitives, as embedded in the codebook, serve to represent the facial motions in a spatial and temporal manner." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 231, + 546, + 270 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 231, + 546, + 270 + ], + "spans": [ + { + "bbox": [ + 305, + 231, + 546, + 270 + ], + "type": "text", + "content": "obtain the quantized motion sequence " + }, + { + "bbox": [ + 305, + 231, + 546, + 270 + ], + "type": "inline_equation", + "content": "\\mathbf{Z}_{\\mathbf{q}} \\in \\mathbb{R}^{T' \\times H \\times C}" + }, + { + "bbox": [ + 305, + 231, + 546, + 270 + ], + "type": "text", + "content": " via an element-wise quantization function " + }, + { + "bbox": [ + 305, + 231, + 546, + 270 + ], + "type": "inline_equation", + "content": "Q(\\cdot)" + }, + { + "bbox": [ + 305, + 231, + 546, + 270 + ], + "type": "text", + "content": " that maps each item in " + }, + { + "bbox": [ + 305, + 231, + 546, + 270 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{Z}}" + }, + { + "bbox": [ + 305, + 231, + 546, + 270 + ], + "type": "text", + "content": " to its nearest entry in codebook " + }, + { + "bbox": [ + 305, + 231, + 546, + 270 + ], + "type": "inline_equation", + "content": "\\mathcal{Z}" + }, + { + "bbox": [ + 305, + 231, + 546, + 270 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 350, + 280, + 545, + 300 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 350, + 280, + 545, + 300 + ], + "spans": [ + { + "bbox": [ + 350, + 280, + 545, + 300 + ], + "type": "interline_equation", + "content": "\\mathbf {Z} _ {\\mathbf {q}} = Q (\\hat {\\mathbf {Z}}) := \\underset {\\mathbf {z} _ {k} \\in \\mathcal {Z}} {\\arg \\min } \\| \\hat {\\mathbf {z}} _ {t} - \\mathbf {z} _ {k} \\| _ {2}. \\tag {1}", + "image_path": "a45f74fb33ad25df4c9101a404e6dca3dba8a97cd94f80d1141b503b41c8b02f.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 305, + 473, + 318 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 305, + 473, + 318 + ], + "spans": [ + { + "bbox": [ + 306, + 305, + 473, + 318 + ], + "type": "text", + "content": "Then, the self-reconstruction is given by:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 347, + 326, + 545, + 340 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 347, + 326, + 545, + 340 + ], + "spans": [ + { + "bbox": [ + 347, + 326, + 545, + 340 + ], + "type": "interline_equation", + "content": "\\hat {\\mathbf {M}} _ {1: T} = D \\left(\\mathbf {Z} _ {\\mathbf {q}}\\right) = D \\left(Q \\left(E \\left(\\mathbf {M} _ {1: T}\\right)\\right)\\right). \\tag {2}", + "image_path": "4378bfc5d1fff2cd708a9da28fe25f451c77813afb27d24f3c3c0a158e8199c3.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 349, + 545, + 397 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 349, + 545, + 397 + ], + "spans": [ + { + "bbox": [ + 305, + 349, + 545, + 397 + ], + "type": "text", + "content": "Note that, the discrete facial motion space reduces the mapping ambiguity with the finite cardinality, but never sacrifices its expressiveness thanks to its context-rich representation as a latent space." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 399, + 546, + 436 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 399, + 546, + 436 + ], + "spans": [ + { + "bbox": [ + 305, + 399, + 546, + 436 + ], + "type": "text", + "content": "Training objectives. Similar to [57], to supervise the quantized autoencoder training, we adopt a motion-level loss and two intermediate code-level losses:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 330, + 443, + 545, + 475 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 330, + 443, + 545, + 475 + ], + "spans": [ + { + "bbox": [ + 330, + 443, + 545, + 475 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} _ {\\mathrm {V Q}} = \\left\\| \\mathbf {M} _ {1: T} - \\hat {\\mathbf {M}} _ {1: T} \\right\\| _ {1} \\tag {3} \\\\ + \\left\\| \\operatorname {s g} (\\hat {\\mathbf {Z}}) - \\mathbf {Z} _ {\\mathbf {q}} \\right\\| _ {2} ^ {2} + \\beta \\left\\| \\hat {\\mathbf {Z}} - \\operatorname {s g} (\\mathbf {Z} _ {\\mathbf {q}}) \\right\\| _ {2} ^ {2}, \\\\ \\end{array}", + "image_path": "218835fe9c19d284cac527773659d52e3e579b0a072eb759ecc3a80f1aea0a08.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 305, + 483, + 547, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 483, + 547, + 590 + ], + "spans": [ + { + "bbox": [ + 305, + 483, + 547, + 590 + ], + "type": "text", + "content": "where the first term is a reconstruction loss, the latter two are adopted to update the codebook items by reducing the distance between the codebook " + }, + { + "bbox": [ + 305, + 483, + 547, + 590 + ], + "type": "inline_equation", + "content": "\\mathcal{Z}" + }, + { + "bbox": [ + 305, + 483, + 547, + 590 + ], + "type": "text", + "content": " and embedded features " + }, + { + "bbox": [ + 305, + 483, + 547, + 590 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{Z}}" + }, + { + "bbox": [ + 305, + 483, + 547, + 590 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 305, + 483, + 547, + 590 + ], + "type": "inline_equation", + "content": "\\mathrm{sg}(\\cdot)" + }, + { + "bbox": [ + 305, + 483, + 547, + 590 + ], + "type": "text", + "content": " stands for a stop-gradient operation and " + }, + { + "bbox": [ + 305, + 483, + 547, + 590 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 305, + 483, + 547, + 590 + ], + "type": "text", + "content": " is a weighting factor controlling the update rate of the codebook and encoder. Since the quantization function (Eq. 1) is not differentiable, the straight-through gradient estimator [4,57] is employed to copy the gradients from the decoder input to the encoder output." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 594, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 547, + 714 + ], + "type": "text", + "content": "Discussion. Recently, Learn2Listen [42] has applied VQ-VAE for facial expression synthesis in response to a given talking head harnessing 2D monocular videos to obtain 3DMM coefficients. In addition to distinct applications, here we would like to emphasize our major differences. First, Learn2Listen constructs speaker-specific codebooks while ours uses a generic codebook that is feasible to represent arbitrary facial motions. Since cross-character motions are absorbed, our codebook is naturally embedded with more plentiful priors. Second, Learn2Listen utilizes" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12782" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 48, + 71, + 545, + 223 + ], + "blocks": [ + { + "bbox": [ + 48, + 71, + 545, + 223 + ], + "lines": [ + { + "bbox": [ + 48, + 71, + 545, + 223 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 545, + 223 + ], + "type": "image", + "image_path": "bb53616013f399ae8117acaf337e90ee822fc0adab0aa30f92e5648d4075cea1.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 227, + 546, + 261 + ], + "lines": [ + { + "bbox": [ + 46, + 227, + 546, + 261 + ], + "spans": [ + { + "bbox": [ + 46, + 227, + 546, + 261 + ], + "type": "text", + "content": "Figure 2. Diagram of our speech-driven motion synthesis model. Given the speech " + }, + { + "bbox": [ + 46, + 227, + 546, + 261 + ], + "type": "inline_equation", + "content": "\\mathbf{A}_{1:T}" + }, + { + "bbox": [ + 46, + 227, + 546, + 261 + ], + "type": "text", + "content": " and style vector " + }, + { + "bbox": [ + 46, + 227, + 546, + 261 + ], + "type": "inline_equation", + "content": "\\mathbf{s}" + }, + { + "bbox": [ + 46, + 227, + 546, + 261 + ], + "type": "text", + "content": " as input, the model learn to recursively generate a sequence of facial motions by predicting the motion codes. As embedded with well-learned motion priors, the pre-trained codebook and decoder are frozen during training." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 47, + 266, + 286, + 371 + ], + "blocks": [ + { + "bbox": [ + 47, + 266, + 286, + 371 + ], + "lines": [ + { + "bbox": [ + 47, + 266, + 286, + 371 + ], + "spans": [ + { + "bbox": [ + 47, + 266, + 286, + 371 + ], + "type": "image", + "image_path": "2b293390f136ca97ceaf31c7bd7d135b7650aaffaea82fa6fa1abae0ffc961ed.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 379, + 287, + 456 + ], + "lines": [ + { + "bbox": [ + 46, + 379, + 287, + 456 + ], + "spans": [ + { + "bbox": [ + 46, + 379, + 287, + 456 + ], + "type": "text", + "content": "Figure 3. Concept comparison with Learn2Listen [42]. (Top) The speaker-specific facial expression coefficient prior [42], in which each code represents a sequence of facial expression coefficients. (Bottom) Our speaker-agnostic generic motion prior, in which each code represents the motion primitive of face components. The blue dotted boxes indicate what information each code may represent conceptually." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 461, + 287, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 461, + 287, + 664 + ], + "spans": [ + { + "bbox": [ + 46, + 461, + 287, + 664 + ], + "type": "text", + "content": "the codebook to represent common sequences of facial expressions by the way of 3DMM coefficients, i.e., each code represents a sequence (8 frames) of facial expressions. Differently, our codebook is formulated to represent the vertex-based facial motion space, where the codes are embedded with per-vertex motions of facial components and represent the facial motion (within a temporal unit) in a context-rich manner. As compared in Figure 3, the codebook of Learn2Listen is learned to memorize typical sequential facial expressions of a specific speaker within 3DMM space, which cannot synthesize realistic facial motions with subtle details due to the limited expressiveness of 3DMM and is bounded by the accuracy of 3D reconstruction techniques [42]. As the first attempt, our codebook is learned to represent the generic facial motion space with motion primitives for captured facial mesh data, which is more effective to embed general priors preserving vivid facial details." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 665, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 287, + 713 + ], + "type": "text", + "content": "We further discuss the hyper-parameters of the codebook. First, the length of the temporal unit " + }, + { + "bbox": [ + 46, + 665, + 287, + 713 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 46, + 665, + 287, + 713 + ], + "type": "text", + "content": " and the number of face components " + }, + { + "bbox": [ + 46, + 665, + 287, + 713 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 46, + 665, + 287, + 713 + ], + "type": "text", + "content": " determine the complexity of the motion primitives in temporal and spatial aspects" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 266, + 545, + 410 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 266, + 545, + 410 + ], + "spans": [ + { + "bbox": [ + 304, + 266, + 545, + 410 + ], + "type": "text", + "content": "respectively. Generally, complex motion primitives cause low flexibility and reusability and thus hinder representation effectiveness. On the opposite, overly simple motion primitives challenge motion prediction due to the lack of semantics. Besides, the codebook size " + }, + { + "bbox": [ + 304, + 266, + 545, + 410 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 266, + 545, + 410 + ], + "type": "text", + "content": " and the feature dimension " + }, + { + "bbox": [ + 304, + 266, + 545, + 410 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 304, + 266, + 545, + 410 + ], + "type": "text", + "content": " determine the representation capability, which should be defined according to the complexity of the dataset and in cooperation with " + }, + { + "bbox": [ + 304, + 266, + 545, + 410 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 304, + 266, + 545, + 410 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 266, + 545, + 410 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 304, + 266, + 545, + 410 + ], + "type": "text", + "content": ". In our experiment, we set " + }, + { + "bbox": [ + 304, + 266, + 545, + 410 + ], + "type": "inline_equation", + "content": "N = 256" + }, + { + "bbox": [ + 304, + 266, + 545, + 410 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 266, + 545, + 410 + ], + "type": "inline_equation", + "content": "P = 1" + }, + { + "bbox": [ + 304, + 266, + 545, + 410 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 266, + 545, + 410 + ], + "type": "inline_equation", + "content": "H = 8" + }, + { + "bbox": [ + 304, + 266, + 545, + 410 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 304, + 266, + 545, + 410 + ], + "type": "inline_equation", + "content": "H = 16" + }, + { + "bbox": [ + 304, + 266, + 545, + 410 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 266, + 545, + 410 + ], + "type": "inline_equation", + "content": "C = 64" + }, + { + "bbox": [ + 304, + 266, + 545, + 410 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 304, + 266, + 545, + 410 + ], + "type": "inline_equation", + "content": "C = 128" + }, + { + "bbox": [ + 304, + 266, + 545, + 410 + ], + "type": "text", + "content": " depending on the dataset, which lead to high-quality results as justified by the ablation studies in Section 4.5. More details can be found in the Supplement." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 418, + 480, + 431 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 418, + 480, + 431 + ], + "spans": [ + { + "bbox": [ + 306, + 418, + 480, + 431 + ], + "type": "text", + "content": "3.2. Speech-Driven Motion Synthesis" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 437, + 545, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 437, + 545, + 569 + ], + "spans": [ + { + "bbox": [ + 304, + 437, + 545, + 569 + ], + "type": "text", + "content": "With the learned discrete motion prior, we can build a cross-modal mapping from the input speech to the target motion codes that could be further decoded into realistic facial motions. Along with the speech, we further adopt a control on the talking styles as input, i.e., a style vector " + }, + { + "bbox": [ + 304, + 437, + 545, + 569 + ], + "type": "inline_equation", + "content": "\\mathbf{s} \\in \\mathbb{R}_+^M \\cup \\{0\\}" + }, + { + "bbox": [ + 304, + 437, + 545, + 569 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 437, + 545, + 569 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 304, + 437, + 545, + 569 + ], + "type": "text", + "content": " is the dimension of the learned style space (see Eq. 4). Conditioning on the speech " + }, + { + "bbox": [ + 304, + 437, + 545, + 569 + ], + "type": "inline_equation", + "content": "\\mathbf{A}_{1:T}" + }, + { + "bbox": [ + 304, + 437, + 545, + 569 + ], + "type": "text", + "content": " and the style vector " + }, + { + "bbox": [ + 304, + 437, + 545, + 569 + ], + "type": "inline_equation", + "content": "\\mathbf{s}" + }, + { + "bbox": [ + 304, + 437, + 545, + 569 + ], + "type": "text", + "content": ", a temporal autoregressive model, composed of a speech encoder " + }, + { + "bbox": [ + 304, + 437, + 545, + 569 + ], + "type": "inline_equation", + "content": "E_{\\mathrm{speech}}" + }, + { + "bbox": [ + 304, + 437, + 545, + 569 + ], + "type": "text", + "content": " and a cross-modal decoder " + }, + { + "bbox": [ + 304, + 437, + 545, + 569 + ], + "type": "inline_equation", + "content": "D_{\\mathrm{cross-modal}}" + }, + { + "bbox": [ + 304, + 437, + 545, + 569 + ], + "type": "text", + "content": ", is employed to learn over the facial motion space, as depicted in Figure 2." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 570, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 546, + 713 + ], + "type": "text", + "content": "Following FaceFormer [16], our speech encoder adopts the architecture of the state-of-the-art self-supervised pretrained speech model, wav2vec 2.0 [3], which consists of an audio feature extractor and a multi-layer transformer encoder. The audio feature extractor converts the speech of raw waveform into feature vectors through a temporal convolutions network (TCN). Benefiting from the effective attention scheme, the transformer encoder converts the audio features into contextualized speech representations. Apart from the pre-trained codebook and VQ-VAE decoder, our cross-modal decoder contains an embedding block and a multi-layer transformer decoder with causal self-attention." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12783" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 287, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 287, + 96 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 287, + 96 + ], + "type": "text", + "content": "The embedding block combines the past facial motions and the style embedding via:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 94, + 105, + 287, + 128 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 94, + 105, + 287, + 128 + ], + "spans": [ + { + "bbox": [ + 94, + 105, + 287, + 128 + ], + "type": "interline_equation", + "content": "\\mathbf {F} _ {\\mathrm {e m b}} ^ {1: t - 1} = \\mathcal {P} _ {\\theta} \\left(\\hat {\\mathbf {M}} _ {1: t - 1}\\right) + \\mathbf {B} \\cdot \\frac {\\mathbf {S}}{\\| \\mathbf {S} \\| _ {1}}, \\tag {4}", + "image_path": "63b833a5652700c5e1ebcc2197f67c2e5eaecf8590e19ec2f22fd076951edc50.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 136, + 287, + 280 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 136, + 287, + 280 + ], + "spans": [ + { + "bbox": [ + 47, + 136, + 287, + 280 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 136, + 287, + 280 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_{\\theta}" + }, + { + "bbox": [ + 47, + 136, + 287, + 280 + ], + "type": "text", + "content": " is a linear projection layer, and " + }, + { + "bbox": [ + 47, + 136, + 287, + 280 + ], + "type": "inline_equation", + "content": "\\mathbf{B} = [\\mathbf{b}_1,\\dots,\\mathbf{b}_M]\\in \\mathbb{R}^{C\\times M}" + }, + { + "bbox": [ + 47, + 136, + 287, + 280 + ], + "type": "text", + "content": " denotes the " + }, + { + "bbox": [ + 47, + 136, + 287, + 280 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 47, + 136, + 287, + 280 + ], + "type": "text", + "content": " learnable basis vectors that span the style space linearly. Alike to Face-Former [16], we equip the transformer decoder with causal self-attention to learn the dependencies between each frame in the context of the past facial motion sequence, and with cross-modal attention to align the audio and motion modalities. The output features " + }, + { + "bbox": [ + 47, + 136, + 287, + 280 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{Z}}^{1:t}" + }, + { + "bbox": [ + 47, + 136, + 287, + 280 + ], + "type": "text", + "content": " is further quantized into " + }, + { + "bbox": [ + 47, + 136, + 287, + 280 + ], + "type": "inline_equation", + "content": "\\mathbf{Z}_{\\mathbf{q}}^{1:t}" + }, + { + "bbox": [ + 47, + 136, + 287, + 280 + ], + "type": "text", + "content": " via Eq. 1 and decoded by the pre-trained VQ-VAE decoder. The newly predicted motion " + }, + { + "bbox": [ + 47, + 136, + 287, + 280 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{m}}_t" + }, + { + "bbox": [ + 47, + 136, + 287, + 280 + ], + "type": "text", + "content": " is used to update the past motions as " + }, + { + "bbox": [ + 47, + 136, + 287, + 280 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{M}}_{1:t}" + }, + { + "bbox": [ + 47, + 136, + 287, + 280 + ], + "type": "text", + "content": ", in preparation for the next prediction. Formally, this recursive process can be written as:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 75, + 289, + 287, + 304 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 289, + 287, + 304 + ], + "spans": [ + { + "bbox": [ + 75, + 289, + 287, + 304 + ], + "type": "interline_equation", + "content": "\\hat {\\mathbf {m}} _ {t} = D _ {\\text {c r o s s - m o d a l}} \\left(E _ {\\text {s p e e c h}} \\left(\\mathbf {A} _ {1: T}\\right), \\mathbf {s}, \\tilde {\\mathbf {M}} _ {1: t - 1}\\right). \\tag {5}", + "image_path": "7147736868f32b0cfcca6b532d084c4d5309eb8ff0082fad3763da54bfe7c2d7.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 315, + 287, + 485 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 315, + 287, + 485 + ], + "spans": [ + { + "bbox": [ + 47, + 315, + 287, + 485 + ], + "type": "text", + "content": "Training objectives. We train the transformer encoder, decoder and the embedding block for cross-modality mapping, while keeping the codebook " + }, + { + "bbox": [ + 47, + 315, + 287, + 485 + ], + "type": "inline_equation", + "content": "\\mathcal{Z}" + }, + { + "bbox": [ + 47, + 315, + 287, + 485 + ], + "type": "text", + "content": " and motion decoder " + }, + { + "bbox": [ + 47, + 315, + 287, + 485 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 47, + 315, + 287, + 485 + ], + "type": "text", + "content": " frozen. To benefit from the speech representation learning from large-scale corpora, we initialize the TCN and transformer encoder with the pre-trained wav2vec 2.0 weights. Overall, the autoregressive model is trained in a teaching-forcing scheme, under the constraint of two loss terms: (i) feature regularity loss " + }, + { + "bbox": [ + 47, + 315, + 287, + 485 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{reg}}" + }, + { + "bbox": [ + 47, + 315, + 287, + 485 + ], + "type": "text", + "content": " measuring the deviation between the predicted motion feature " + }, + { + "bbox": [ + 47, + 315, + 287, + 485 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{Z}}^{1:T}" + }, + { + "bbox": [ + 47, + 315, + 287, + 485 + ], + "type": "text", + "content": " and the quantized feature " + }, + { + "bbox": [ + 47, + 315, + 287, + 485 + ], + "type": "inline_equation", + "content": "\\mathbf{Z}_{\\mathbf{q}}^{1:T}" + }, + { + "bbox": [ + 47, + 315, + 287, + 485 + ], + "type": "text", + "content": " from codebook, and (ii) motion loss " + }, + { + "bbox": [ + 47, + 315, + 287, + 485 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{motion}}" + }, + { + "bbox": [ + 47, + 315, + 287, + 485 + ], + "type": "text", + "content": " measuring the difference between the predicted motions " + }, + { + "bbox": [ + 47, + 315, + 287, + 485 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{M}}_{1:T}" + }, + { + "bbox": [ + 47, + 315, + 287, + 485 + ], + "type": "text", + "content": " and the ground-truth motions " + }, + { + "bbox": [ + 47, + 315, + 287, + 485 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_{1:T}" + }, + { + "bbox": [ + 47, + 315, + 287, + 485 + ], + "type": "text", + "content": ", which plays an important role to stabilize the training process. The final loss function is:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 59, + 495, + 286, + 525 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 495, + 286, + 525 + ], + "spans": [ + { + "bbox": [ + 59, + 495, + 286, + 525 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} _ {\\text {s y n}} = \\mathcal {L} _ {\\text {r e g}} + \\mathcal {L} _ {\\text {m o t i o n}} \\\\ = \\| \\hat {\\mathbf {Z}} ^ {1: T} - \\operatorname {s g} \\left(\\mathbf {Z} _ {\\mathbf {q}} ^ {1: T}\\right) \\| _ {2} ^ {2} + \\| \\hat {\\mathbf {M}} _ {1: T} - \\mathbf {M} _ {1: T} \\| _ {2} ^ {2}. \\tag {6} \\\\ \\end{array}", + "image_path": "1fe7f6f3a686e7b0dfb37a1c63e543606875a768cc1d62f18d36570ab20b36c8.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 534, + 145, + 547 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 534, + 145, + 547 + ], + "spans": [ + { + "bbox": [ + 47, + 534, + 145, + 547 + ], + "type": "text", + "content": "3.3. Training Details" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 553, + 287, + 649 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 553, + 287, + 649 + ], + "spans": [ + { + "bbox": [ + 46, + 553, + 287, + 649 + ], + "type": "text", + "content": "At stage one, we train the VQ-VAE model (Figure 1) on a single NVIDIA V100 for 200 epochs (" + }, + { + "bbox": [ + 46, + 553, + 287, + 649 + ], + "type": "inline_equation", + "content": "\\sim" + }, + { + "bbox": [ + 46, + 553, + 287, + 649 + ], + "type": "text", + "content": "2 hours) with the AdamW [39] optimizer (" + }, + { + "bbox": [ + 46, + 553, + 287, + 649 + ], + "type": "inline_equation", + "content": "\\beta_{1} = 0.9" + }, + { + "bbox": [ + 46, + 553, + 287, + 649 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 553, + 287, + 649 + ], + "type": "inline_equation", + "content": "\\beta_{2} = 0.999" + }, + { + "bbox": [ + 46, + 553, + 287, + 649 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 553, + 287, + 649 + ], + "type": "inline_equation", + "content": "\\epsilon = 1e - 8" + }, + { + "bbox": [ + 46, + 553, + 287, + 649 + ], + "type": "text", + "content": "), where the learning rate is initialized as " + }, + { + "bbox": [ + 46, + 553, + 287, + 649 + ], + "type": "inline_equation", + "content": "10^{-4}" + }, + { + "bbox": [ + 46, + 553, + 287, + 649 + ], + "type": "text", + "content": ", and the mini-batch size is set as 1. At stage two, we train the temporal autoregressive model with the Adam optimizer [31]. The training duration is 100 epochs (" + }, + { + "bbox": [ + 46, + 553, + 287, + 649 + ], + "type": "inline_equation", + "content": "\\sim" + }, + { + "bbox": [ + 46, + 553, + 287, + 649 + ], + "type": "text", + "content": "3 hours) and other hyper-parameters remain unchanged as stage one." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 654, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 654, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 654, + 287, + 713 + ], + "type": "text", + "content": "Style embedding space. The style embedding space is linearly spanned by " + }, + { + "bbox": [ + 46, + 654, + 287, + 713 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 46, + 654, + 287, + 713 + ], + "type": "text", + "content": " learned basis vectors, where each style is represented by a style vector " + }, + { + "bbox": [ + 46, + 654, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{s}" + }, + { + "bbox": [ + 46, + 654, + 287, + 713 + ], + "type": "text", + "content": " that serves as the linear combination coefficients or a coordinate. In VOCA [10], it assigns each speaker with a category-like" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 72, + 545, + 192 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 192 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 192 + ], + "type": "text", + "content": "one-hot style vector. Instead, we propose a concept to formulate a uniform style space via some learnable basis vectors (Eq. 4), where the style vector is no longer bound with speaker ID but associated with every talking sample. During training, considering the limited style diversity of training datasets, we assign each speaker (e.g. no. " + }, + { + "bbox": [ + 304, + 72, + 545, + 192 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 72, + 545, + 192 + ], + "type": "text", + "content": ") with a standard unit vector " + }, + { + "bbox": [ + 304, + 72, + 545, + 192 + ], + "type": "inline_equation", + "content": "\\mathbf{e}_i" + }, + { + "bbox": [ + 304, + 72, + 545, + 192 + ], + "type": "text", + "content": " as a style vector, under the assumption that each speaker is associated with a unique and consistent style. Anyway, arbitrary style vectors are allowed to interpolate new talking styles during inference." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 197, + 388, + 210 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 197, + 388, + 210 + ], + "spans": [ + { + "bbox": [ + 305, + 197, + 388, + 210 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 217, + 470, + 229 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 217, + 470, + 229 + ], + "spans": [ + { + "bbox": [ + 305, + 217, + 470, + 229 + ], + "type": "text", + "content": "4.1. Datasets and Implementations" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 235, + 545, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 235, + 545, + 319 + ], + "spans": [ + { + "bbox": [ + 304, + 235, + 545, + 319 + ], + "type": "text", + "content": "We employ two widely used datasets, BIWI [18] and VOCASET [10], to train and test different methods in our experiments. Both datasets contain 4D face scans together with utterances spoken in English. BIWI contains 40 unique sentences shared across all speakers in the dataset, while VOCASET contains 255 unique sentences, which are partially shared among different speakers." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 319, + 545, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 319, + 545, + 533 + ], + "spans": [ + { + "bbox": [ + 304, + 319, + 545, + 533 + ], + "type": "text", + "content": "BIWI dataset. BIWI is a 3D audio-visual corpus of affective speech and facial expression in the form of dense dynamic 3D face geometries, which is originally proposed to study affective communication. There is a total of 40 sentences uttered by 14 subjects, eight females and six males. Each sentence was recorded twice: with and without emotion. On average, each sentence is 4.67 seconds long. The 3D face dynamics are captured at 25fps, each with 23370 vertices and registered topology. We follow the data splits in [16] and use the emotional subset. Specifically, the training set (BIWI-Train) contains 192 sentences, while the validation set (BIWI-Val) contains 24 sentences. There are two testing sets, in which BIWI-Test-A includes 24 sentences spoken by six seen subjects and BIWI-Test-B contains 32 sentences spoken by eight unseen subjects. BIWI-Test-A can be used for both quantitative and qualitative evaluation due to the seen subjects during training, while BIWI-Test-B is more suitable for qualitative evaluation." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 534, + 545, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 534, + 545, + 629 + ], + "spans": [ + { + "bbox": [ + 304, + 534, + 545, + 629 + ], + "type": "text", + "content": "VOCASET dataset. VOCASET is comprised of 480 paired audio-visual sequences recorded from 12 subjects. The facial motion is captured at 60fps and is about 4 seconds long. Different from BIWI, each 3D face mesh is registered to the FLAME [35] topology with 5023 vertices. We adopt the same training (VOCA-Train), validation (VOCA-Val), and testing (VOCA-Test) splits as VOCA [10] and FaceFormer for fair comparisons." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 630, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 630, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 630, + 545, + 712 + ], + "type": "text", + "content": "Implementations. We compare our work with three state-of-the-art methods: VOCA [10], MeshTalk [50] and FaceFormer [16]. We train and test VOCA on BIWI using the official codebase, while directly testing the released model that was trained on VOCASET. For MeshTalk, we train and test it using the official implementation on the two datasets. To compare with FaceFormer, we conduct test-" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12784" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 52, + 97, + 282, + 177 + ], + "blocks": [ + { + "bbox": [ + 47, + 71, + 287, + 92 + ], + "lines": [ + { + "bbox": [ + 47, + 71, + 287, + 92 + ], + "spans": [ + { + "bbox": [ + 47, + 71, + 287, + 92 + ], + "type": "text", + "content": "Table 1. Quantitative evaluation on BIWI-Test-A. Lower means better for both metrics." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 52, + 97, + 282, + 177 + ], + "lines": [ + { + "bbox": [ + 52, + 97, + 282, + 177 + ], + "spans": [ + { + "bbox": [ + 52, + 97, + 282, + 177 + ], + "type": "table", + "html": "
MethodLip Vertex Error (×10-4mm)FDD (×10-5mm)
VOCA6.55638.1816
MeshTalk5.91815.1025
FaceFormer5.30774.6408
CodeTalker (Ours)4.79144.1170
", + "image_path": "34021c8088ba812203796b2d428ee9175c8cf2d72f1678aa6d012111702556d0.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 186, + 286, + 258 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 186, + 286, + 258 + ], + "spans": [ + { + "bbox": [ + 46, + 186, + 286, + 258 + ], + "type": "text", + "content": "ing directly using the pre-trained weights. Among the four methods, VOCA, FaceFormer and our CodeTalker require conditioning on a training speaking style during testing. For unseen subjects, we generate facial animations conditioned on all training styles. More details about the implementation details can be found in the Supplementary Material." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 265, + 182, + 277 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 265, + 182, + 277 + ], + "spans": [ + { + "bbox": [ + 47, + 265, + 182, + 277 + ], + "type": "text", + "content": "4.2. Quantitative Evaluation" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 283, + 286, + 366 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 283, + 286, + 366 + ], + "spans": [ + { + "bbox": [ + 46, + 283, + 286, + 366 + ], + "type": "text", + "content": "Following MeshTalk [50] and FaceFormer [16], we adopt the lip vertex error to measure the lip synchronization, which is the only publicly proposed metric for speech-driven facial animation evaluation, to our best knowledge. As a complement, we introduce a new quantitative measurement, i.e., upper-face motion statistics, to evaluate the overall facial dynamics." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 367, + 287, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 367, + 287, + 415 + ], + "spans": [ + { + "bbox": [ + 46, + 367, + 287, + 415 + ], + "type": "text", + "content": "Lip vertex error. It measures the lip deviation of a sequence with respect to the ground truth, i.e., calculating the maximal L2 error of all lip vertices for each frame and takes the average over all frames." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 415, + 287, + 499 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 415, + 287, + 499 + ], + "spans": [ + { + "bbox": [ + 46, + 415, + 287, + 499 + ], + "type": "text", + "content": "Upper-face dynamics deviation. The upper-face expression is just loosely correlated with the speech, depending on personal talking styles and the semantics of speech content. With this belief, we propose to measure the variation of facial dynamics for a motion sequence in comparison with that of the ground truth. Specifically, the upper-face dynamics deviation (FDD) is calculated by:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 508, + 291, + 545 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 508, + 291, + 545 + ], + "spans": [ + { + "bbox": [ + 47, + 508, + 291, + 545 + ], + "type": "interline_equation", + "content": "\\operatorname {F D D} \\left(\\mathbf {M} _ {1: T}, \\hat {\\mathbf {M}} _ {1: T}\\right) = \\frac {\\sum_ {v \\in \\mathcal {S} _ {U}} \\left(\\operatorname {d y n} \\left(\\mathbf {M} _ {1 : T} ^ {v}\\right) - \\operatorname {d y n} \\left(\\hat {\\mathbf {M}} _ {1 : T} ^ {v}\\right)\\right)}{\\left| \\mathcal {S} _ {U} \\right|}, \\tag {7}", + "image_path": "9f2c4f8ab82b7ae0c38d95f20240e1241d91dd65b4950828f2ebf8a964a09344.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 545, + 286, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 545, + 286, + 594 + ], + "spans": [ + { + "bbox": [ + 46, + 545, + 286, + 594 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 545, + 286, + 594 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_{1:T}^v\\in \\mathbb{R}^{3\\times T}" + }, + { + "bbox": [ + 46, + 545, + 286, + 594 + ], + "type": "text", + "content": " denotes the motions of the " + }, + { + "bbox": [ + 46, + 545, + 286, + 594 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 46, + 545, + 286, + 594 + ], + "type": "text", + "content": " -th vertex, and " + }, + { + "bbox": [ + 46, + 545, + 286, + 594 + ], + "type": "inline_equation", + "content": "S_U" + }, + { + "bbox": [ + 46, + 545, + 286, + 594 + ], + "type": "text", + "content": " is the index set of upper-face vertices. " + }, + { + "bbox": [ + 46, + 545, + 286, + 594 + ], + "type": "inline_equation", + "content": "\\mathrm{dyn}(\\cdot)" + }, + { + "bbox": [ + 46, + 545, + 286, + 594 + ], + "type": "text", + "content": " denotes the standard deviation of the element-wise L2 norm along the temporal axis." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 594, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 287, + 714 + ], + "type": "text", + "content": "We calculate the lip vertex error and upper-face dynamics deviation (FDD) over all sequences in BIWI-Test-A and take the average for comparison. According to Table 1, the proposed CodeTalker achieves lower error than the existing state-of-the-arts, suggesting that it produces more accurate lip-synchronized movements. Besides, Table 1 shows that our CodeTalker achieves the best performance in terms of FDD. It indicates the high consistency between the predicted upper-face expressions together with the trend of facial dynamics (conditioned on the speech and talking styles)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 72, + 426, + 84 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 426, + 84 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 426, + 84 + ], + "type": "text", + "content": "and those of the ground truth." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 94, + 434, + 106 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 94, + 434, + 106 + ], + "spans": [ + { + "bbox": [ + 306, + 94, + 434, + 106 + ], + "type": "text", + "content": "4.3. Qualitative Evaluation" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 114, + 545, + 341 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 114, + 545, + 341 + ], + "spans": [ + { + "bbox": [ + 304, + 114, + 545, + 341 + ], + "type": "text", + "content": "We visually compare our method with other competitors in Figure 4. For fair comparison, we assign the same talking style to VOCA, FaceFormer and our CodeTalker as conditional input, which is sampled at random. To check the lip synchronization performance, we illustrate three typical frames of synthesized facial animations that speak at specific syllables, as compared in the upper partition in Figure 4. We can observe that compared with the competitors, the lip movements produced by our CodeTalker are more accurately articulated with the speech signals and also more consistent with those of the reference. For example, CodeTalker produces better lip sync with proper mouth closures when pronouncing bilabial consonant /b/ (i.e., \"bedside\" in the upper-right case of Figure 4), compared to VOCA and MeshTalk; for the even challenging speech parts \"waterproof\" and \"shaving\" that need to pout, CodeTalker can produce accurate lip shapes while other methods suffer from the over-smoothing problem and fail to lip-sync correctly (Zoom in for better inspection)." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 342, + 545, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 342, + 545, + 557 + ], + "spans": [ + { + "bbox": [ + 304, + 342, + 545, + 557 + ], + "type": "text", + "content": "Different from lip movements, facial expressions only have weak correlations with the speech signal, which tends to be static in front of cross-modal mapping ambiguity. To visualize the facial motion dynamics, we calculate the temporal statistics of adjacent-frame facial motions within a sequence. Specifically, we first calculate the inter-frame motion L2 distance and then compute the mean and standard deviation (std) across the sequence at each vertex. The higher mean value indicates stronger facial movements, while the higher std value suggests richer variations of facial dynamics. Two examples are visualized in the last two rows of Figure 4, evidencing that our method outperforms others in achieving both stronger facial movements and a broader range of dynamics. It is mainly attributed to the superiority of the discrete facial motion space, which promotes the robustness to cross-modal uncertainty effectively. Readers are recommended to watch the animation comparisons in the Supplemental Video." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 558, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 558, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 558, + 545, + 713 + ], + "type": "text", + "content": "Talking style interpolation. Our model can synthesize new speaking styles from the learned style embedding space. To inspect the effects on VOCA-Test, we select two speaking style vectors, i.e., " + }, + { + "bbox": [ + 304, + 558, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{e}_i" + }, + { + "bbox": [ + 304, + 558, + 545, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 558, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{e}_j" + }, + { + "bbox": [ + 304, + 558, + 545, + 713 + ], + "type": "text", + "content": ", which correspond to large and slight lip articulations respectively, and interpolate new talking style vectors " + }, + { + "bbox": [ + 304, + 558, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{s}_{\\mathrm{new}} = \\mathbf{B} \\cdot [\\omega \\mathbf{e}_i + (1 - \\omega) \\mathbf{e}_j]" + }, + { + "bbox": [ + 304, + 558, + 545, + 713 + ], + "type": "text", + "content": " with a linear coefficient " + }, + { + "bbox": [ + 304, + 558, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\omega" + }, + { + "bbox": [ + 304, + 558, + 545, + 713 + ], + "type": "text", + "content": ". For the synthesized 3D animations of a sampled sequence, we plot the lower-upper lip distances across frames for each style in Figure 5, from which we observe the smooth transition of mouth amplitudes between the two typical styles. It is not only useful to synthesize new talking styles, but also practical to match a specific speaking performance of an unseen subject during training." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12785" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 48, + 68, + 307, + 345 + ], + "blocks": [ + { + "bbox": [ + 48, + 68, + 307, + 345 + ], + "lines": [ + { + "bbox": [ + 48, + 68, + 307, + 345 + ], + "spans": [ + { + "bbox": [ + 48, + 68, + 307, + 345 + ], + "type": "image", + "image_path": "affde4db5730ff6cffbd1477c7f35d573f095ba6793e02af018a4ecc9ab39454.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 352, + 547, + 386 + ], + "lines": [ + { + "bbox": [ + 46, + 352, + 547, + 386 + ], + "spans": [ + { + "bbox": [ + 46, + 352, + 547, + 386 + ], + "type": "text", + "content": "Figure 4. Visual comparisons of sampled facial motions animated by different methods on VOCA-Test (left) and BIWI-Test-B (right). The upper partition shows the facial animation conditioned on different speech parts, while the lower depicts the temporal statistics (mean and standard deviation) of adjacent-frame motion variations within a sequence." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 307, + 68, + 547, + 345 + ], + "blocks": [ + { + "bbox": [ + 307, + 68, + 547, + 345 + ], + "lines": [ + { + "bbox": [ + 307, + 68, + 547, + 345 + ], + "spans": [ + { + "bbox": [ + 307, + 68, + 547, + 345 + ], + "type": "image", + "image_path": "8c26e7797a3bfe4ec4341c4bacd29f14d0d1a53b8b130bc088cb5d32ddf9fedf.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 48, + 393, + 289, + 509 + ], + "blocks": [ + { + "bbox": [ + 48, + 393, + 289, + 509 + ], + "lines": [ + { + "bbox": [ + 48, + 393, + 289, + 509 + ], + "spans": [ + { + "bbox": [ + 48, + 393, + 289, + 509 + ], + "type": "image", + "image_path": "4b04dc2ffc959720fb6cf189f9e133b644298f5d20e524c9b1c62ec97ba874a5.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 511, + 287, + 544 + ], + "lines": [ + { + "bbox": [ + 46, + 511, + 287, + 544 + ], + "spans": [ + { + "bbox": [ + 46, + 511, + 287, + 544 + ], + "type": "text", + "content": "Figure 5. Distance between lower and upper lip for our predictions within a sequence conditioned on different weighted linear combinations of two style vectors." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 552, + 122, + 564 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 552, + 122, + 564 + ], + "spans": [ + { + "bbox": [ + 47, + 552, + 122, + 564 + ], + "type": "text", + "content": "4.4. User Study" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 570, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 570, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 570, + 288, + 715 + ], + "type": "text", + "content": "The human perception system has been evolutionarily adapted to understanding subtle facial motions and capturing lip synchronization. Thus, it is still the most reliable measure in the speech-driven facial animation task. We conduct a user study to evaluate the quality of animated faces in perceptual lip synchronization and realism, compared with VOCA, MeshTalk, FaceFormer and the ground truth. We adopt A/B tests for each comparison, i.e., ours vs. competitor, in terms of realistic facial animation and lip sync. For BIWI, we obtain the results of four kinds of comparisons by randomly selecting 30 samples from BIWI-Test-B, respectively. To achieve the most variations in terms of speaking" + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 308, + 430, + 545, + 514 + ], + "blocks": [ + { + "bbox": [ + 305, + 394, + 545, + 427 + ], + "lines": [ + { + "bbox": [ + 305, + 394, + 545, + 427 + ], + "spans": [ + { + "bbox": [ + 305, + 394, + 545, + 427 + ], + "type": "text", + "content": "Table 2. User study results on BIWI-Test-B and VOCA-Test. We adopt A/B testing and report the percentage of answers where A is preferred over B." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 308, + 430, + 545, + 514 + ], + "lines": [ + { + "bbox": [ + 308, + 430, + 545, + 514 + ], + "spans": [ + { + "bbox": [ + 308, + 430, + 545, + 514 + ], + "type": "table", + "html": "
CompetitorsBIWI-Test-BVOCA-Test
Lip SyncRealismLip SyncRealism
Ours vs. VOCA92.4789.2586.0284.95
Ours vs. MeshTalk80.6582.8095.7092.47
Ours vs. FaceFormer53.7656.9970.9769.89
Ours vs. GT43.0149.4643.0143.01
", + "image_path": "9ee9e9789b0556ed0fa1c16cb2099f45fd94f59d3f12765b51af9ef63d589c46.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 521, + 545, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 521, + 545, + 665 + ], + "spans": [ + { + "bbox": [ + 304, + 521, + 545, + 665 + ], + "type": "text", + "content": "styles, we ensure the sampling results can fairly cover all conditioning styles. Thus, " + }, + { + "bbox": [ + 304, + 521, + 545, + 665 + ], + "type": "inline_equation", + "content": "120\\mathrm{A}" + }, + { + "bbox": [ + 304, + 521, + 545, + 665 + ], + "type": "text", + "content": " vs. B pairs (30 samples " + }, + { + "bbox": [ + 304, + 521, + 545, + 665 + ], + "type": "inline_equation", + "content": "\\times 4" + }, + { + "bbox": [ + 304, + 521, + 545, + 665 + ], + "type": "text", + "content": " comparisons) are created for BIWI-Test-B. Each pair is judged by at least 3 different participants separately, and finally, 372 entries are collected in total. For the user study on VOCASET, we apply the same setting as that on the BIWI dataset, i.e., another " + }, + { + "bbox": [ + 304, + 521, + 545, + 665 + ], + "type": "inline_equation", + "content": "120\\mathrm{A}" + }, + { + "bbox": [ + 304, + 521, + 545, + 665 + ], + "type": "text", + "content": " vs. B pairs from VOCA-Test set, finally yielding 372 entries as well. In this study, 31 participants with good vision and hearing ability complete the evaluation successfully. Moreover, each participant is involved in all 8 kinds of comparisons to make better exposure and cover the diversity of favorability." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "type": "text", + "content": "The percentage of A/B testing in terms of lip sync and realism on BIWI-Test-B is tabulated in Table 2, which shows that participants favor CodeTalker over competitors. Based on the visual analysis in Section 4.3, we attribute this to the" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "12786" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 48, + 107, + 287, + 308 + ], + "blocks": [ + { + "bbox": [ + 47, + 71, + 287, + 103 + ], + "lines": [ + { + "bbox": [ + 47, + 71, + 287, + 103 + ], + "spans": [ + { + "bbox": [ + 47, + 71, + 287, + 103 + ], + "type": "text", + "content": "Table 3. Ablation study on the representation space of codebook. The performance is measured by the reconstruction error (i.e., L2 error) and lip vertex error on VOCA-Test and BIWI-Test-A." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 48, + 107, + 287, + 308 + ], + "lines": [ + { + "bbox": [ + 48, + 107, + 287, + 308 + ], + "spans": [ + { + "bbox": [ + 48, + 107, + 287, + 308 + ], + "type": "table", + "html": "
VariantsVOCA-TestBIWI-Test-A
Rec. Error (×10-5mm)Rec. Error (×10-5mm)Lip Vertex Error (×10-4mm)
Shape-ent. codebook2.754.076.41
Motion codebook (Ours)0.082.834.79
Speech-driven synthesis"finding Nan's cameo""an excellent room"
Shape-ent. codebookMotion (Ours) codebookShape-ent. codebookMotion (Ours) codebook
", + "image_path": "0046cc2055500718bc9f967b44586cc7e21613cf2c49093fdb435c8a9453048b.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 314, + 287, + 347 + ], + "lines": [ + { + "bbox": [ + 47, + 314, + 287, + 347 + ], + "spans": [ + { + "bbox": [ + 47, + 314, + 287, + 347 + ], + "type": "text", + "content": "Figure 6. Visual comparisons of reconstruction and speech-driven motion synthesis results with different representation spaces on VOCA-Test (left) and BIWI-Test-A (right)." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 353, + 287, + 497 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 353, + 287, + 497 + ], + "spans": [ + { + "bbox": [ + 46, + 353, + 287, + 497 + ], + "type": "text", + "content": "facial animation synthesized by CodeTalker having more expressive facial motions, accurate lip shape, and well-synchronized mouth movements. For VOCA-Test, which has a nature of fewer upper-face motions, a similar favorability can still be observed in Table 2. We believe the reasons are at least three-fold: subtle motions around the eyes, more accurate lip movements and expressive motions in the lower face. Although be aware of a gap between our predictions and the recorded performance (ground truth), we surprisingly get over " + }, + { + "bbox": [ + 46, + 353, + 287, + 497 + ], + "type": "inline_equation", + "content": "40\\%" + }, + { + "bbox": [ + 46, + 353, + 287, + 497 + ], + "type": "text", + "content": " preference still. Overall, the user study justifies that the facial animations produced by CodeTalker have superior perceptual quality." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 503, + 147, + 514 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 503, + 147, + 514 + ], + "spans": [ + { + "bbox": [ + 47, + 503, + 147, + 514 + ], + "type": "text", + "content": "4.5. Ablation Studies" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 522, + 287, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 522, + 287, + 557 + ], + "spans": [ + { + "bbox": [ + 46, + 522, + 287, + 557 + ], + "type": "text", + "content": "We study several key designs of our proposed method in this section, including the representation space and the hyper-parameters of the codebook construction." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 558, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 558, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 558, + 287, + 713 + ], + "type": "text", + "content": "Representation space. To study the superiority of our motion-based representation, we construct a baseline that learns a shape-entangled codebook, i.e., the codes represent shapes " + }, + { + "bbox": [ + 46, + 558, + 287, + 713 + ], + "type": "inline_equation", + "content": "(\\mathbf{m_t} + \\mathbf{h})" + }, + { + "bbox": [ + 46, + 558, + 287, + 713 + ], + "type": "text", + "content": " instead of motions " + }, + { + "bbox": [ + 46, + 558, + 287, + 713 + ], + "type": "inline_equation", + "content": "(\\mathbf{m_t})" + }, + { + "bbox": [ + 46, + 558, + 287, + 713 + ], + "type": "text", + "content": ". As shown in Table 3, the baseline decreases the reconstruction accuracy significantly, which is evidenced by the visualized examples in Figure 6. It is mainly because the shape-entangled sequence contains more speaker-specific information that hinders the reusability of the codes. A direct weakness is the poor generalization for self-reconstruction, which further impedes cross-modal mapping correctness. In contrast, our proposed speaker-agnostic motion representation is more effective to represent generic motion priors shared" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 306, + 70, + 425, + 177 + ], + "blocks": [ + { + "bbox": [ + 306, + 70, + 425, + 177 + ], + "lines": [ + { + "bbox": [ + 306, + 70, + 425, + 177 + ], + "spans": [ + { + "bbox": [ + 306, + 70, + 425, + 177 + ], + "type": "image", + "image_path": "4ed4c8fd48debac4bccf63b2118f8fae2e9993cf3b6daa8e533b8e563687963c.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 183, + 545, + 228 + ], + "lines": [ + { + "bbox": [ + 304, + 183, + 545, + 228 + ], + "spans": [ + { + "bbox": [ + 304, + 183, + 545, + 228 + ], + "type": "text", + "content": "Figure 7. Model performance comparisons with different hyperparameters of the codebook, i.e., the length of temporal unit " + }, + { + "bbox": [ + 304, + 183, + 545, + 228 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 304, + 183, + 545, + 228 + ], + "type": "text", + "content": " and the number of face components " + }, + { + "bbox": [ + 304, + 183, + 545, + 228 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 304, + 183, + 545, + 228 + ], + "type": "text", + "content": ". We measure the reconstruction error and lip vertex error on BIWI-Test-A." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 428, + 70, + 547, + 177 + ], + "blocks": [ + { + "bbox": [ + 428, + 70, + 547, + 177 + ], + "lines": [ + { + "bbox": [ + 428, + 70, + 547, + 177 + ], + "spans": [ + { + "bbox": [ + 428, + 70, + 547, + 177 + ], + "type": "image", + "image_path": "e0030efa375101435fceb90c2527f23521101929501c92ed6062e7bc3d39dedf.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 236, + 545, + 260 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 236, + 545, + 260 + ], + "spans": [ + { + "bbox": [ + 304, + 236, + 545, + 260 + ], + "type": "text", + "content": "across individuals, and hence promotes the quality of both self-reconstruction and speech-driven motion synthesis." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 261, + 546, + 452 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 261, + 546, + 452 + ], + "spans": [ + { + "bbox": [ + 304, + 261, + 546, + 452 + ], + "type": "text", + "content": "Codebook construction. We further study the hyperparameters used for codebook construction. We evaluate the performance of different settings " + }, + { + "bbox": [ + 304, + 261, + 546, + 452 + ], + "type": "inline_equation", + "content": "\\langle P, H \\rangle" + }, + { + "bbox": [ + 304, + 261, + 546, + 452 + ], + "type": "text", + "content": " by measuring their reconstruction accuracy and cross-modal mapping accuracy (namely lip vertex error). First, we evaluate the reconstruction accuracy as shown in Figure 7(a). On one hand, increasing " + }, + { + "bbox": [ + 304, + 261, + 546, + 452 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 304, + 261, + 546, + 452 + ], + "type": "text", + "content": " degrades the reconstruction accuracy, which could be explained by the increased complexity of the motion to be represented. On the other hand, increasing " + }, + { + "bbox": [ + 304, + 261, + 546, + 452 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 304, + 261, + 546, + 452 + ], + "type": "text", + "content": " eases the reconstruction but risks over-fitting, which explains the general benefits " + }, + { + "bbox": [ + 304, + 261, + 546, + 452 + ], + "type": "inline_equation", + "content": "(H < 8)" + }, + { + "bbox": [ + 304, + 261, + 546, + 452 + ], + "type": "text", + "content": " but inferior performance when " + }, + { + "bbox": [ + 304, + 261, + 546, + 452 + ], + "type": "inline_equation", + "content": "H \\geq 8" + }, + { + "bbox": [ + 304, + 261, + 546, + 452 + ], + "type": "text", + "content": ". Notably, a similar trend could be found in the cross-modal mapping performance, as shown in Figure 7(b). We conjecture that complex motion primitives cause lower reusability and higher redundancy, resulting in ambiguity in the cross-modal code query process." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 463, + 457, + 475 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 463, + 457, + 475 + ], + "spans": [ + { + "bbox": [ + 306, + 463, + 457, + 475 + ], + "type": "text", + "content": "5. Discussion and Conclusion" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 483, + 545, + 673 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 483, + 545, + 673 + ], + "spans": [ + { + "bbox": [ + 304, + 483, + 545, + 673 + ], + "type": "text", + "content": "We demonstrated the advantages of casting speech-driven facial animation as a code query task in the discrete space, which notably promotes the motion synthesis quality against cross-modal ambiguity. By comparing to the existing state-of-the-arts, our proposed method shows superiority in achieving accurate lip sync and vivid facial expressions. However, we still follow the assumption that facial motions are independent of shapes, whose rationality may deserve further studies. Also, the overall perceptual quality still lags behind the ground truth, primarily due to the scarcity of paired audio-visual data. Additionally, the acquired generic motion prior adheres to the motion distribution delineated by the training set, which may deviate from real-world facial motions. As a future work, it is intriguing to guide the 3D facial animation by utilizing priors from large-scale available talking head videos." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 677, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 545, + 712 + ], + "type": "text", + "content": "Acknowledgements. This project is partially supported by Hong Kong Innovation and Technology Fund (ITF) (ref: ITS/307/20FP and ITS/313/20)." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12787" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 715 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 135 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 135 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 135 + ], + "type": "text", + "content": "[1] Mohammed M Alghamdi, He Wang, Andrew J Bulpitt, and David C Hogg. Talking head from speech audio using a pretrained image generator. In ACM International Conference on Multimedia (MM), 2022. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 137, + 288, + 180 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 137, + 288, + 180 + ], + "spans": [ + { + "bbox": [ + 53, + 137, + 288, + 180 + ], + "type": "text", + "content": "[2] Tenglong Ao, Qingzhe Gao, Yuke Lou, Baoquan Chen, and Libin Liu. Rhythmic gesticulator: Rhythm-aware co-speech gesture synthesis with hierarchical neural embeddings. ACM Transactions on Graphics (TOG), pages 1-19, 2022. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 182, + 288, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 182, + 288, + 226 + ], + "spans": [ + { + "bbox": [ + 53, + 182, + 288, + 226 + ], + "type": "text", + "content": "[3] Alexei Baevski, Yuhao Zhou, Abdelrahman Mohamed, and Michael Auli. wav2vec 2.0: A framework for self-supervised learning of speech representations. In Conference on Neural Information Processing Systems (NeurIPS), 2020. 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 228, + 288, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 228, + 288, + 270 + ], + "spans": [ + { + "bbox": [ + 53, + 228, + 288, + 270 + ], + "type": "text", + "content": "[4] Yoshua Bengio, Nicholas Léonard, and Aaron Courville. Estimating or propagating gradients through stochastic neurons for conditional computation. arXiv preprint arXiv:1308.3432, 2013. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 272, + 288, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 272, + 288, + 316 + ], + "spans": [ + { + "bbox": [ + 53, + 272, + 288, + 316 + ], + "type": "text", + "content": "[5] Chen Cao, Hongzhi Wu, Yanlin Weng, Tianjia Shao, and Kun Zhou. Real-time facial animation with image-based dynamic avatars. ACM Transactions on Graphics (TOG), 35(4):1-12, 2016. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 318, + 288, + 350 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 318, + 288, + 350 + ], + "spans": [ + { + "bbox": [ + 53, + 318, + 288, + 350 + ], + "type": "text", + "content": "[6] Yong Cao, Wen C Tien, Petros Faloutsos, and Frédéric Pighin. Expressive speech-driven facial animation. ACM Transactions on Graphics (TOG), 24(4):1283-1302, 2005. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 351, + 288, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 351, + 288, + 396 + ], + "spans": [ + { + "bbox": [ + 53, + 351, + 288, + 396 + ], + "type": "text", + "content": "[7] Lele Chen, Guofeng Cui, Celong Liu, Zhong Li, Ziyi Kou, Yi Xu, and Chenliang Xu. Talking-head generation with rhythmic head motion. In European Conference on Computer Vision (ECCV), 2020. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 397, + 288, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 397, + 288, + 430 + ], + "spans": [ + { + "bbox": [ + 53, + 397, + 288, + 430 + ], + "type": "text", + "content": "[8] Lele Chen, Zhiheng Li, Ross K Maddox, Zhiyao Duan, and Chenliang Xu. Lip movements generation at a glance. In European Conference on Computer Vision (ECCV), 2018. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 431, + 288, + 463 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 431, + 288, + 463 + ], + "spans": [ + { + "bbox": [ + 53, + 431, + 288, + 463 + ], + "type": "text", + "content": "[9] Joon Son Chung and Andrew Zisserman. Out of time: automated lip sync in the wild. In Asian Conference on Computer Vision (ACCV), 2016. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 465, + 288, + 519 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 465, + 288, + 519 + ], + "spans": [ + { + "bbox": [ + 48, + 465, + 288, + 519 + ], + "type": "text", + "content": "[10] Daniel Cudeiro, Timo Bolkart, Cassidy Laidlaw, Anurag Ranjan, and Michael J Black. Capture, learning, and synthesis of 3d speaking styles. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 1, 2, 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 521, + 288, + 565 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 521, + 288, + 565 + ], + "spans": [ + { + "bbox": [ + 48, + 521, + 288, + 565 + ], + "type": "text", + "content": "[11] Dipanjan Das, Sandika Biswas, Sanjana Sinha, and Brojeshwar Bhowmick. Speech-driven facial animation using cascaded gans for learning of motion and texture. In European Conference on Computer Vision (ECCV), 2020. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 567, + 288, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 567, + 288, + 609 + ], + "spans": [ + { + "bbox": [ + 48, + 567, + 288, + 609 + ], + "type": "text", + "content": "[12] Pif Edwards, Chris Landreth, Eugene Fume, and Karan Singh. Jali: an animator-centric viseme model for expressive lip synchronization. ACM Transactions on Graphics (TOG), 35(4):1-11, 2016. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 612, + 288, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 612, + 288, + 643 + ], + "spans": [ + { + "bbox": [ + 48, + 612, + 288, + 643 + ], + "type": "text", + "content": "[13] Paul Ekman and Wallace V Friesen. Facial action coding system. Environmental Psychology & Nonverbal Behavior, 1978. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 646, + 288, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 646, + 288, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 646, + 288, + 689 + ], + "type": "text", + "content": "[14] Michael Elad and Michal Aharon. Image denoising via learned dictionaries and sparse representation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2006. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 691, + 288, + 715 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 288, + 715 + ], + "type": "text", + "content": "[15] Patrick Esser, Robin Rombach, and Bjorn Ommer. Taming transformers for high-resolution image synthesis. In" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 712 + ], + "type": "list", + "angle": 0, + "index": 31, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "text", + "content": "IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 96, + 547, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 96, + 547, + 150 + ], + "spans": [ + { + "bbox": [ + 307, + 96, + 547, + 150 + ], + "type": "text", + "content": "[16] Yingruo Fan, Zhaojiang Lin, Jun Saito, Wenping Wang, and Taku Komura. Faceformer: Speech-driven 3d facial animation with transformers. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 1, 2, 4, 5, 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 152, + 545, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 152, + 545, + 196 + ], + "spans": [ + { + "bbox": [ + 307, + 152, + 545, + 196 + ], + "type": "text", + "content": "[17] Yingruo Fan, Zhaojiang Lin, Jun Saito, Wenping Wang, and Taku Komura. Joint audio-text model for expressive speech-driven 3d facial animation. In ACM Symposium on Interactive 3D Graphics and Games (i3D), 2022. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 198, + 545, + 241 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 198, + 545, + 241 + ], + "spans": [ + { + "bbox": [ + 307, + 198, + 545, + 241 + ], + "type": "text", + "content": "[18] Gabriele Fanelli, Juergen Gall, Harald Romsdorfer, Thibaut Weise, and Luc Van Gool. A 3-d audio-visual corpus of affective communication. IEEE Transactions on Multimedia (TMM), 12(6):591-598, 2010. 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 243, + 545, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 243, + 545, + 274 + ], + "spans": [ + { + "bbox": [ + 307, + 243, + 545, + 274 + ], + "type": "text", + "content": "[19] Cletus G Fisher. Confusions among visually perceived consonants. Journal of speech and hearing research, 11(4):796-804, 1968. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 276, + 545, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 276, + 545, + 331 + ], + "spans": [ + { + "bbox": [ + 307, + 276, + 545, + 331 + ], + "type": "text", + "content": "[20] Ohad Fried, Ayush Tewari, Michael Zollhöfer, Adam Finkelstein, Eli Shechtman, Dan B Goldman, Kyle Genova, Zeyu Jin, Christian Theobalt, and Maneesh Agrawala. Text-based editing of talking-head video. ACM Transactions on Graphics (TOG), 38(4):1-14, 2019. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 333, + 545, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 333, + 545, + 387 + ], + "spans": [ + { + "bbox": [ + 307, + 333, + 545, + 387 + ], + "type": "text", + "content": "[21] Shuyang Gu, Dong Chen, Jianmin Bao, Fang Wen, Bo Zhang, Dongdong Chen, Lu Yuan, and Baining Guo. Vector quantized diffusion model for text-to-image synthesis. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 388, + 545, + 432 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 388, + 545, + 432 + ], + "spans": [ + { + "bbox": [ + 307, + 388, + 545, + 432 + ], + "type": "text", + "content": "[22] Shuhang Gu, Wangmeng Zuo, Qi Xie, Deyu Meng, Xiangchu Feng, and Lei Zhang. Convolutional sparse coding for image super-resolution. In IEEE/CVF International Conference on Computer Vision (ICCV), 2015. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 434, + 545, + 477 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 434, + 545, + 477 + ], + "spans": [ + { + "bbox": [ + 307, + 434, + 545, + 477 + ], + "type": "text", + "content": "[23] Yudong Guo, Keyu Chen, Sen Liang, Yong-Jin Liu, Hujun Bao, and Juyong Zhang. Ad-nerf: Audio driven neural radiance fields for talking head synthesis. In IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 479, + 545, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 479, + 545, + 533 + ], + "spans": [ + { + "bbox": [ + 307, + 479, + 545, + 533 + ], + "type": "text", + "content": "[24] Ikhsanul Habibie, Weipeng Xu, Dushyant Mehta, Lingjie Liu, Hans-Peter Seidel, Gerard Pons-Moll, Mohamed Elgharib, and Christian Theobalt. Learning speech-driven 3d conversational gestures from video. In ACM International Conference on Intelligent Virtual Agents (IVA), 2021. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 535, + 545, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 535, + 545, + 578 + ], + "spans": [ + { + "bbox": [ + 307, + 535, + 545, + 578 + ], + "type": "text", + "content": "[25] Zhitong Huang, Nanxuan Zhao, and Jing Liao. Unicolor: A unified framework for multi-modal colorization with transformer. ACM Transactions on Graphics (TOG), pages 1-16, 2022. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 580, + 545, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 580, + 545, + 623 + ], + "spans": [ + { + "bbox": [ + 307, + 580, + 545, + 623 + ], + "type": "text", + "content": "[26] Xinya Ji, Hang Zhou, Kaisiyuan Wang, Qianyi Wu, Wayne Wu, Feng Xu, and Xun Cao. Eamm: One-shot emotional talking face via audio-based emotion-aware motion model. In ACM SIGGRAPH Conference, 2022. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 624, + 545, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 624, + 545, + 668 + ], + "spans": [ + { + "bbox": [ + 307, + 624, + 545, + 668 + ], + "type": "text", + "content": "[27] Xinya Ji, Hang Zhou, Kaisiyuan Wang, Wayne Wu, Chen Change Loy, Xun Cao, and Feng Xu. Audio-driven emotional video portraits. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 2" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 307, + 670, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 670, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 307, + 670, + 545, + 712 + ], + "type": "text", + "content": "[28] Younghyun Jo and Seon Joo Kim. Practical single-image super-resolution using look-up table. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 2" + } + ] + } + ], + "index": 30 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "12788" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "text", + "content": "[29] Tero Karras, Timo Aila, Samuli Laine, Antti Herva, and Jaakko Lehtinen. Audio-driven facial animation by joint end-to-end learning of pose and emotion. ACM Transactions on Graphics (TOG), 36(4):1-12, 2017. 1, 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 117, + 287, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 117, + 287, + 171 + ], + "spans": [ + { + "bbox": [ + 48, + 117, + 287, + 171 + ], + "type": "text", + "content": "[30] Hyeongwoo Kim, Pablo Garrido, Ayush Tewari, Weipeng Xu, Justus Thies, Matthias Niessner, Patrick Pérez, Christian Richardt, Michael Zollhöfer, and Christian Theobalt. Deep video portraits. ACM Transactions on Graphics (TOG), 37(4):1-14, 2018. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 172, + 287, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 172, + 287, + 205 + ], + "spans": [ + { + "bbox": [ + 48, + 172, + 287, + 205 + ], + "type": "text", + "content": "[31] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In International Conference on Learning Representations (ICLR), 2014. 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 205, + 287, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 205, + 287, + 259 + ], + "spans": [ + { + "bbox": [ + 48, + 205, + 287, + 259 + ], + "type": "text", + "content": "[32] Avisek Lahiri, Vivek Kwatra, Christian Frueh, John Lewis, and Chris Bregler. Lipsync3d: Data-efficient learning of personalized 3d talking faces from video using pose and lighting normalization. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 261, + 287, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 261, + 287, + 293 + ], + "spans": [ + { + "bbox": [ + 48, + 261, + 287, + 293 + ], + "type": "text", + "content": "[33] John Lewis. Automated lip-sync: Background and techniques. The Journal of Visualization and Computer Animation, 2(4):118-122, 1991. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 294, + 287, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 294, + 287, + 326 + ], + "spans": [ + { + "bbox": [ + 48, + 294, + 287, + 326 + ], + "type": "text", + "content": "[34] Hao Li, Jihun Yu, Yuting Ye, and Chris Bregler. Realtime facial animation with on-the-fly correctives. ACM Transactions on Graphics (TOG), 32(4):42-1, 2013. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 327, + 287, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 327, + 287, + 369 + ], + "spans": [ + { + "bbox": [ + 48, + 327, + 287, + 369 + ], + "type": "text", + "content": "[35] Tianye Li, Timo Bolkart, Michael J Black, Hao Li, and Javier Romero. Learning a model of facial shape and expression from 4d scans. ACM Transactions on Graphics (TOG), 36(6):194-1, 2017. 1, 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 371, + 287, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 371, + 287, + 426 + ], + "spans": [ + { + "bbox": [ + 48, + 371, + 287, + 426 + ], + "type": "text", + "content": "[36] Borong Liang, Yan Pan, Zhizhi Guo, Hang Zhou, Zhibin Hong, Xiaoguang Han, Junyu Han, Jingtuo Liu, Errui Ding, and Jingdong Wang. Expressive talking head generation with granular audio-visual control. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 426, + 287, + 469 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 426, + 287, + 469 + ], + "spans": [ + { + "bbox": [ + 48, + 426, + 287, + 469 + ], + "type": "text", + "content": "[37] Xian Liu, Yinghao Xu, Qianyi Wu, Hang Zhou, Wayne Wu, and Bolei Zhou. Semantic-aware implicit neural auditory-driven video portrait generation. In European Conference on Computer Vision (ECCV), 2022. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 471, + 287, + 513 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 471, + 287, + 513 + ], + "spans": [ + { + "bbox": [ + 48, + 471, + 287, + 513 + ], + "type": "text", + "content": "[38] Yilong Liu, Feng Xu, Jinxiang Chai, Xin Tong, Lijuan Wang, and Qiang Huo. Video-audio driven real-time facial animation. ACM Transactions on Graphics (TOG), 34(6):1-10, 2015. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 514, + 287, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 514, + 287, + 547 + ], + "spans": [ + { + "bbox": [ + 48, + 514, + 287, + 547 + ], + "type": "text", + "content": "[39] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In International Conference on Learning Representations (ICLR), 2019. 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 548, + 287, + 580 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 548, + 287, + 580 + ], + "spans": [ + { + "bbox": [ + 48, + 548, + 287, + 580 + ], + "type": "text", + "content": "[40] DW Massaro, MM Cohen, M Tabain, J Beskow, and R Clark. Animated speech: research progress and applications. Audiovisual Speech Processing, pages 309-345, 2012. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 581, + 287, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 581, + 287, + 613 + ], + "spans": [ + { + "bbox": [ + 48, + 581, + 287, + 613 + ], + "type": "text", + "content": "[41] Wesley Mattheyses and Werner Verhelst. Audiovisual speech synthesis: An overview of the state-of-the-art. Speech Communication, 66:182-217, 2015. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 614, + 287, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 614, + 287, + 668 + ], + "spans": [ + { + "bbox": [ + 48, + 614, + 287, + 668 + ], + "type": "text", + "content": "[42] Evonne Ng, Hanbyul Joo, Liwen Hu, Hao Li, Trevor Darrell, Angjoo Kanazawa, and Shiry Ginosar. Learning to listen: Modeling non-deterministic dyadic facial motion. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2, 3, 4" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "type": "text", + "content": "[43] Youxin Pang, Yong Zhang, Weize Quan, Yanbo Fan, Xiaodong Cun, Ying Shan, and Dong-ming Yan. Dpe: Disentanglement of pose and expression for general video portrait editing. arXiv preprint arXiv:2301.06281, 2023. 2" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 31, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 95 + ], + "type": "text", + "content": "[44] Frederic I Parke and Keith Waters. Computer facial animation. CRC press, 2008. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 95, + 545, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 95, + 545, + 139 + ], + "spans": [ + { + "bbox": [ + 307, + 95, + 545, + 139 + ], + "type": "text", + "content": "[45] Jialun Peng, Dong Liu, Songcen Xu, and Houqiang Li. Generating diverse structure for image inpainting with hierarchical vq-vae. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 140, + 545, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 140, + 545, + 182 + ], + "spans": [ + { + "bbox": [ + 307, + 140, + 545, + 182 + ], + "type": "text", + "content": "[46] Hai Xuan Pham, Yuting Wang, and Vladimir Pavlovic. End-to-end learning for 3d facial animation from speech. In ACM International Conference on Multimodal Interaction (ICMI), 2018. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 183, + 545, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 183, + 545, + 227 + ], + "spans": [ + { + "bbox": [ + 307, + 183, + 545, + 227 + ], + "type": "text", + "content": "[47] KR Prajwal, Rudrabha Mukhopadhyay, Vinay P Namboodiri, and CV Jawahar. A lip sync expert is all you need for speech to lip generation in the wild. In ACM International Conference on Multimedia (MM), 2020. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 228, + 545, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 228, + 545, + 270 + ], + "spans": [ + { + "bbox": [ + 307, + 228, + 545, + 270 + ], + "type": "text", + "content": "[48] Ali Razavi, Aaron Van den Oord, and Oriol Vinyals. Generating diverse high-fidelity images with vq-vae-2. In Conference on Neural Information Processing Systems (NeurIPS), 2019. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 271, + 545, + 325 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 271, + 545, + 325 + ], + "spans": [ + { + "bbox": [ + 307, + 271, + 545, + 325 + ], + "type": "text", + "content": "[49] Alexander Richard, Colin Lea, Shugao Ma, Jurgen Gall, Fernando De la Torre, and Yaser Sheikh. Audio-and gaze-driven facial animation of codec avatars. In IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), 2021. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 327, + 545, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 327, + 545, + 381 + ], + "spans": [ + { + "bbox": [ + 307, + 327, + 545, + 381 + ], + "type": "text", + "content": "[50] Alexander Richard, Michael Zollhöfer, Yandong Wen, Fernando De la Torre, and Yaser Sheikh. Meshtalk: 3d face animation from speech using cross-modality disentanglement. In IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 1, 2, 5, 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 382, + 545, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 382, + 545, + 426 + ], + "spans": [ + { + "bbox": [ + 307, + 382, + 545, + 426 + ], + "type": "text", + "content": "[51] Shuai Shen, Wanhua Li, Zheng Zhu, Yueqi Duan, Jie Zhou, and Jiwen Lu. Learning dynamic facial radiance fields for few-shot talking head synthesis. In European Conference on Computer Vision (ECCV), 2022. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 426, + 545, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 426, + 545, + 480 + ], + "spans": [ + { + "bbox": [ + 307, + 426, + 545, + 480 + ], + "type": "text", + "content": "[52] Sarah Taylor, Taehwan Kim, Yisong Yue, Moshe Mahler, James Krahe, Anastasio Garcia Rodriguez, Jessica Hodgins, and Iain Matthews. A deep learning approach for generalized speech animation. ACM Transactions on Graphics (TOG), 36(4):1-11, 2017. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 481, + 545, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 481, + 545, + 525 + ], + "spans": [ + { + "bbox": [ + 307, + 481, + 545, + 525 + ], + "type": "text", + "content": "[53] Sarah L Taylor, Moshe Mahler, Barry-John Theobald, and Iain Matthews. Dynamic units of visual speech. In ACM/Eurographics Symposium on Computer Animation (SCA), 2012. 1, 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 525, + 545, + 569 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 525, + 545, + 569 + ], + "spans": [ + { + "bbox": [ + 307, + 525, + 545, + 569 + ], + "type": "text", + "content": "[54] Justus Thies, Mohamed Elgharib, Ayush Tewari, Christian Theobalt, and Matthias Nießner. Neural voice puppetry: Audio-driven facial reenactment. In European Conference on Computer Vision (ECCV), 2020. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 570, + 545, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 570, + 545, + 613 + ], + "spans": [ + { + "bbox": [ + 307, + 570, + 545, + 613 + ], + "type": "text", + "content": "[55] Radu Timofte, Vincent De Smet, and Luc Van Gool. Anchored neighborhood regression for fast example-based super-resolution. In IEEE/CVF International Conference on Computer Vision (ICCV), 2013. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 613, + 545, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 613, + 545, + 657 + ], + "spans": [ + { + "bbox": [ + 307, + 613, + 545, + 657 + ], + "type": "text", + "content": "[56] Radu Timofte, Vincent De Smet, and Luc Van Gool. A+: Adjusted anchored neighborhood regression for fast super-resolution. In Asian Conference on Computer Vision (ACCV), 2014. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 658, + 545, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 658, + 545, + 690 + ], + "spans": [ + { + "bbox": [ + 307, + 658, + 545, + 690 + ], + "type": "text", + "content": "[57] Aaron Van Den Oord, Oriol Vinyals, et al. Neural discrete representation learning. Conference on Neural Information Processing Systems (NeurIPS), 2017. 1, 2, 3" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 307, + 691, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 691, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 307, + 691, + 545, + 712 + ], + "type": "text", + "content": "[58] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia" + } + ] + } + ], + "index": 30 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "12789" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 643 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 66, + 72, + 287, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 72, + 287, + 95 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 287, + 95 + ], + "type": "text", + "content": "Polosukhin. Attention is all you need. Conference on Neural Information Processing Systems (NeurIPS), 2017. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 96, + 287, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 96, + 287, + 139 + ], + "spans": [ + { + "bbox": [ + 48, + 96, + 287, + 139 + ], + "type": "text", + "content": "[59] Konstantinos Vougioukas, Stavros Petridis, and Maja Pantic. Realistic speech-driven facial animation with gans. International Journal of Computer Vision (IJCV), 128(5):1398-1413, 2020. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 141, + 287, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 141, + 287, + 173 + ], + "spans": [ + { + "bbox": [ + 48, + 141, + 287, + 173 + ], + "type": "text", + "content": "[60] Qianyun Wang, Zhenfeng Fan, and Shihong Xia. 3d-talkemo: Learning to synthesize 3d emotional talking head. arXiv preprint arXiv:2104.12051, 2021. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 175, + 287, + 218 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 175, + 287, + 218 + ], + "spans": [ + { + "bbox": [ + 48, + 175, + 287, + 218 + ], + "type": "text", + "content": "[61] Suzhen Wang, Lincheng Li, Yu Ding, and Xin Yu. One-shot talking face generation from single-speaker audio-visual correlation learning. In AAAI Conference on Artificial Intelligence (AAAI), 2022. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 220, + 287, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 220, + 287, + 252 + ], + "spans": [ + { + "bbox": [ + 48, + 220, + 287, + 252 + ], + "type": "text", + "content": "[62] Thibaut Weise, Sofien Bouaziz, Hao Li, and Mark Pauly. Realtime performance-based facial animation. ACM Transactions on Graphics (TOG), 30(4):1-10, 2011. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 254, + 287, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 254, + 287, + 285 + ], + "spans": [ + { + "bbox": [ + 48, + 254, + 287, + 285 + ], + "type": "text", + "content": "[63] Yuyu Xu, Andrew W Feng, Stacy Marsella, and Ari Shapiro. A practical and configurable lip sync method for games. In Proceedings of Motion on Games, 2013. 1, 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 288, + 287, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 288, + 287, + 342 + ], + "spans": [ + { + "bbox": [ + 48, + 288, + 287, + 342 + ], + "type": "text", + "content": "[64] Karren Yang, Dejan Markovic, Steven Krenn, Vasu Agrawal, and Alexander Richard. Audio-visual speech codes: Rethinking audio-visual speech enhancement by re-synthesis. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 343, + 287, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 343, + 287, + 386 + ], + "spans": [ + { + "bbox": [ + 48, + 343, + 287, + 386 + ], + "type": "text", + "content": "[65] Ran Yi, Zipeng Ye, Juyong Zhang, Hujun Bao, and Yong-Jin Liu. Audio-driven talking face video generation with learning-based personalized head pose. arXiv preprint arXiv:2002.10137, 2020. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 388, + 287, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 388, + 287, + 442 + ], + "spans": [ + { + "bbox": [ + 48, + 388, + 287, + 442 + ], + "type": "text", + "content": "[66] Bowen Zhang, Chenyang Qi, Pan Zhang, Bo Zhang, Hsiang-Tao Wu, Dong Chen, Qifeng Chen, Yong Wang, and Fang Wen. Metaportrait: Identity-preserving talking head generation with fast personalized adaptation. arXiv preprint arXiv:2212.08062, 2022. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 445, + 287, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 445, + 287, + 498 + ], + "spans": [ + { + "bbox": [ + 48, + 445, + 287, + 498 + ], + "type": "text", + "content": "[67] Hang Zhou, Yasheng Sun, Wayne Wu, Chen Change Loy, Xiaogang Wang, and Ziwei Liu. Pose-controllable talking face generation by implicitly modularized audio-visual representation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 500, + 287, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 500, + 287, + 544 + ], + "spans": [ + { + "bbox": [ + 48, + 500, + 287, + 544 + ], + "type": "text", + "content": "[68] Shangchen Zhou, Kelvin CK Chan, Chongyi Li, and Chen Change Loy. Towards robust blind face restoration with codebook lookup transformer. Conference on Neural Information Processing Systems (NeurIPS), 2022. 2, 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 545, + 287, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 545, + 287, + 588 + ], + "spans": [ + { + "bbox": [ + 48, + 545, + 287, + 588 + ], + "type": "text", + "content": "[69] Yang Zhou, Zhan Xu, Chris Landreth, Evangelos Kalogerakis, Subhransu Maji, and Karan Singh. Visemenet: Audio-driven animator-centric speech animation. ACM Transactions on Graphics (TOG), 37(4):1-10, 2018. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 590, + 287, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 590, + 287, + 643 + ], + "spans": [ + { + "bbox": [ + 48, + 590, + 287, + 643 + ], + "type": "text", + "content": "[70] Michael Zollhöfer, Justus Thies, Pablo Garrido, Derek Bradley, Thabo Beeler, Patrick Pérez, Marc Stamminger, Matthias Nießner, and Christian Theobalt. State of the art on monocular 3d face reconstruction, tracking, and applications. Computer Graphics Forum (CGF), 37(2):523-550, 2018. 2" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "12790" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Collaboration Helps Camera Overtake LiDAR in 3D Detection/0d6cf930-2724-4762-9022-d5c25917e98a_content_list.json b/2023/Collaboration Helps Camera Overtake LiDAR in 3D Detection/0d6cf930-2724-4762-9022-d5c25917e98a_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..c85aeadfb4c8e93d31f56c873518ea7c8ff99710 --- /dev/null +++ b/2023/Collaboration Helps Camera Overtake LiDAR in 3D Detection/0d6cf930-2724-4762-9022-d5c25917e98a_content_list.json @@ -0,0 +1,1904 @@ +[ + { + "type": "text", + "text": "Collaboration Helps Camera Overtake LiDAR in 3D Detection", + "text_level": 1, + "bbox": [ + 168, + 130, + 802, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yue Hu $^{1}$ Yifan Lu $^{1}$ Runsheng Xu $^{2}$ Weidi Xie $^{1,3}$ Siheng Chen $^{1,3*}$ Yanfeng Wang $^{3,1}$", + "bbox": [ + 114, + 178, + 852, + 198 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Cooperative Medianet Innovation Center, Shanghai Jiao Tong University $^{2}$ University of California, Los Angeles $^{3}$ Shanghai AI Laboratory", + "bbox": [ + 194, + 198, + 774, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1{18671129361, yifan.lu, weidi, sihengc, wangyanfeng}@sjtu.edu.cn 2{rxx3386}@ucla.edu", + "bbox": [ + 104, + 233, + 862, + 252 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 286, + 312, + 301 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Camera-only 3D detection provides an economical solution with a simple configuration for localizing objects in 3D space compared to LiDAR-based detection systems. However, a major challenge lies in precise depth estimation due to the lack of direct 3D measurements in the input. Many previous methods attempt to improve depth estimation through network designs, e.g., deformable layers and larger receptive fields. This work proposes an orthogonal direction, improving the camera-only 3D detection by introducing multi-agent collaborations. Our proposed collaborative camera-only 3D detection (CoCa3D) enables agents to share complementary information with each other through communication. Meanwhile, we optimize communication efficiency by selecting the most informative cues. The shared messages from multiple viewpoints disambiguate the single-agent estimated depth and complement the occluded and long-range regions in the single-agent view. We evaluate CoCa3D in one real-world dataset and two new simulation datasets. Results show that CoCa3D improves previous SOTA performances by $44.21\\%$ on DAIR-V2X, $30.60\\%$ on OPV2V+, $12.59\\%$ on CoPerception-UAVs+ for AP@70. Our preliminary results show a potential that with sufficient collaboration, the camera might overtake LiDAR in some practical scenarios. We released the dataset and code.", + "bbox": [ + 75, + 318, + 473, + 696 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 715, + 209, + 732 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "As a fundamental task of computer vision, 3D object detection aims to localize objects in the 3D physical space given an agent's real-time sensor inputs. It is crucial in a wide range of applications, including autonomous driving [10,33,43], surveillance systems [39], robotics [12] and unmanned aerial vehicles [7]. Depending on sensor setups, there are multiple technical solutions to realize 3D object detection. In this spectrum, one extreme emphasizes raising the upper bound of detection performance, which uses", + "bbox": [ + 75, + 738, + 468, + 875 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/533b5c4540de8a20efd2ffd45831fc9ec38d15407d6ad624c14b541c08725489.jpg", + "image_caption": [ + "(a) Scene" + ], + "image_footnote": [], + "bbox": [ + 506, + 277, + 727, + 327 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/61367b8982a22dd1c653adceee4bfd0bb0fd5e2078a2c9edfb71b5211cbecb66.jpg", + "image_caption": [ + "(b) Single camera confidence" + ], + "image_footnote": [], + "bbox": [ + 727, + 277, + 880, + 327 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/a35e09bd0d3ef860530a57319a3a54db804f1935d8f6f5b30f1aa058fc56a776.jpg", + "image_caption": [ + "(c) Collaborative cameras" + ], + "image_footnote": [], + "bbox": [ + 506, + 335, + 725, + 385 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/214c684466d2a3206bfb5805f21a810ec1793cc90cacdd835bd79e982f9f8ef0.jpg", + "image_caption": [ + "(d) Collaborative cameras confidence" + ], + "image_footnote": [], + "bbox": [ + 727, + 335, + 880, + 386 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/c1562dc636730206503729f7ec8a58be3630d47f54b8ec8f717805d3f2a66a46.jpg", + "image_caption": [ + "(e) LiDAR (Ego)", + "Figure 1. Collaborative camera-only 3D detection can disambiguate the single-view estimated depth, address the long-range and occlusion issues, and approach LiDAR in 3D detection." + ], + "image_footnote": [], + "bbox": [ + 504, + 393, + 725, + 443 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/5f3093ac9f3cd9f0cd7378b907f7f3e9bc7b7f9b3d60a7fc8569266b9e7daea3.jpg", + "image_caption": [ + "(f) LiDAR confidence" + ], + "image_footnote": [], + "bbox": [ + 727, + 393, + 880, + 443 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "high-end LiDAR sensor [2, 3, 10, 21, 28, 42, 45] to collect precise 3D measurements. However, this approach is too expensive to scale up. The other extreme solution emphasizes cost effectiveness, which tries to use thrifty sensor setups, e.g. only using cameras to detect 3D objects in real-time [1, 8, 9, 23, 25, 26, 30, 32, 41, 43]. However, camera-only 3D detection is significantly and consistently worse than LiDAR-based detection in most scenarios [19].", + "bbox": [ + 496, + 505, + 892, + 627 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this paper, we propose an orthogonal direction for improving camera-only 3D detection performances by introducing multi-agent collaborations. Hypothetically, empowered by advanced communication systems, multiple agents equipped only with cameras could share visual information with each other. This would bring three outstanding benefits. First, different viewpoints from multiple agents can largely resolve the depth ambiguity issue in camera-only 3D detection, bridging the gap with expensive LiDARs on depth estimation. Second, multi-agent collaboration avoids inevitable limitations in single-agent 3D detection, such as occlusion and long-range issues, and potentially enables more holistic 3D detection; that is, detecting all the objects existed in the 3D scene, including those beyond visual range. Since LiDAR also suffers from limited field of view, this potentially enables collaborative cameras to outperform LiDAR. Third, the total expense of a large fleet of vehicles is significantly reduced as cameras are much", + "bbox": [ + 496, + 628, + 893, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author.", + "bbox": [ + 94, + 887, + 220, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "9243", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "cheaper than LiDAR. However, multi-agent collaboration also brings new challenges. Different from many multiview geometry problems, here we also have to concern communication bandwidth constraints. Thus, each agent needs to select the most informative cues to share.", + "bbox": [ + 75, + 90, + 468, + 165 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Following this design rationale, we propose a novel collaborative camera-only 3D detection framework CoCa3D. It includes three parts: i) single-agent camera-only 3D detection, which achieves basic depth estimation and 3D detection for each agent; ii) collaborative depth estimation, which disambiguates the estimated depths by promoting spatial consistency across multiple agents' viewpoints; and iii) collaborative detection feature learning, which complements detection features by sharing key detection messages with each other. Compared to recent collaborative perception methods [11,14] that are dealing with LiDAR, CoCa3D specifically designs novel collaborative depth estimation to customize the task of camera-only 3D detection.", + "bbox": [ + 75, + 167, + 468, + 362 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To evaluate CoCa3D, we conduct comprehensive experiments on one real-world dataset, DAIR-V2X [40], and two new simulation datasets, OPV2V+ and CoPerception-UAVs+, which are extended based on original OPV2V [37] and CoPerception-UAVs [6] with more collaborative agents that cover three types of agents (cars, infrastructures and drones). Our results show that i) with 10 collaborative agents, CoCa3D enables camera-only detectors to overtake LiDAR-based detectors on OPV2V+; and ii) CoCa3D consistently outperforms previous works in the performance-bandwidth trade-off across multiple datasets by a large margin, improving the previous SOTA performances by $30.60\\%$ on OPV2V+, $12.59\\%$ on CoPerception-UAVs+, $44.21\\%$ on DAIR-V2X for AP@70. To sum up, our contributions are:", + "bbox": [ + 75, + 363, + 468, + 574 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose a novel collaborative camera-only 3D detection framework CoCa3D, which improves the detection ability of cameras with multi-agent collaboration, promoting more holistic 3D detection.", + "- We propose core communication-efficient collaboration techniques, which explore the spatially sparse yet critical depth messages and tackle the depth ambiguity, occlusion, and long-range issues by fusing complementary information from different viewpoints, achieving more accurate and complete 3D representation.", + "- We expand two previous collaborative datasets with more agents, and conduct extensive experiments, validating that i) CoCa3D significantly bridges the performance gap between camera and LiDAR on OPV2V+ and DAIR-V2X; and ii) CoCa3D achieves the state-of-the-art performance-bandwidth trade-off." + ], + "bbox": [ + 75, + 574, + 468, + 814 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 76, + 818, + 218, + 834 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Camera-only 3D object detection. Camera-only 3D detection aims to detect objects in the 3D space given the 2D image through explicit or implicit depth estimation [7, 9, 23, 26, 30, 41]. Recently, bird's-eye-view (BEV) repre", + "bbox": [ + 75, + 839, + 468, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "sentations are widely used for their computation efficiency and comparable performance compared to the 3D voxel features [10]. To get the BEV features, there are two types of methods. The depth-based methods [8, 23, 25] first estimate the depth distribution, and then attentively project the 2D image features along the projection ray to get the 3D voxel features and collapse it to BEV feature. The query-based methods [16, 22, 43] first initialize queries for each BEV grid and then leverage transformer-architecture-based cross-attention to query image features with camera-aware positional embedding. Here, our single-agent camera-only detector follows the simple-yet-effective CaDDN [25].", + "bbox": [ + 496, + 90, + 890, + 272 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "LiDAR 3D object detection. LiDAR-based 3D detection achieves excellent performance due to the precise 3D measurements of the input data. Two well-known ways to encode LiDAR points include voxel-based [10, 42, 45] and point-based [28,29,38]. Voxel-based methods divide the 3D space into regular voxels [45] or pillars [10], and encode the point inside into feature representations. Point-based methods are usually based on the PointNet [24] series to aggregate the feature of points. Then the point features will be used for the proposal refinement. LiDAR-based 3D detection performs well, but high-quality LiDAR is hard to be adopted in a large scale due to the expensive cost. Here we propose an economic camera-only solution by introducing multi-agent collaboration, whose performance can catch up with LiDAR given a sufficient number of agents.", + "bbox": [ + 496, + 273, + 890, + 501 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Collaborative perception. Collaborative perception is an emerging application of multi-agent systems, which promotes perception performance by enabling agents to share information with other agents through communication. In this emerging research field, a surge of high-quality datasets, (V2X-SIM [13], OPV2V [37], DAIR-V2X [40], CoPerception-UAVs [6]), and collaborative methods [5,6,13-15,17,18,20,31,35-37,40] aimed for better performance-bandwidth trade-off have been proposed. Collaborative methods have three types: early, late and intermediate fusion. Compared with early fusion(i.e. transmitting raw data) and late fusion (i.e. transmitting outputs), intermediate fusion(i.e. transmitting and fusing features encoded by deep networks) achieves a better performance-bandwidth trade-off. DiscoNet [14] adopts knowledge distillation to take the advantage of both early and intermediate collaboration. Where2comm [6] optimizes communication efficiency by sharing spatially sparse yet perceptually critical information. However, previous works mainly focus on LiDAR-based 3D object detection, while we investigate collaborative camera-only detection and specifically enhance depth estimation via multi-agent collaboration.", + "bbox": [ + 496, + 502, + 890, + 834 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3. Collaborative Camera-Only 3D Detection", + "text_level": 1, + "bbox": [ + 498, + 845, + 870, + 864 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "This section presents CoCa3D, collaborative camera-only 3D detection framework, which enables multiple", + "bbox": [ + 498, + 869, + 890, + 901 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "9244", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/ba90d29575aba984d580ef376a6ed1027979f5229376939c1bd197e709656b79.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 81, + 87, + 478, + 178 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/b813dba96905d50e5b54918027331eabb7e082785cbf05699461eb85e4259a2c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 483, + 88, + 880, + 178 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/f4b230329c19b0c42f6a9374a9326edf0826269ac4099d74eddd26d4d45ced13.jpg", + "image_caption": [ + "Figure 2. System overview. CoCa3D is a camera-only 3D detector integrated with two collaboration modules. Collaborative depth estimation (Co-Depth) enhances the single-agent estimated depth to achieve more accurate 3D feature. Collaborative detection feature learning (Co-FL) complements the single-agent 3D feature to achieve more holistic 3D detection." + ], + "image_footnote": [], + "bbox": [ + 83, + 180, + 875, + 268 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "agents to share visual information with each other, promoting more holistic 3D detection.", + "bbox": [ + 75, + 324, + 470, + 354 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Problem Formulation", + "text_level": 1, + "bbox": [ + 76, + 363, + 279, + 377 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Consider $N$ agents in the scene, let $\\mathcal{X}_i$ be the RGB image collected by the $i$ th agent and $\\mathcal{O}_i^0$ be the corresponding ground-truth detection. The objective of CoCa3D is to maximize detection performances of all agents given certain communication budget $B$ ; that is,", + "bbox": [ + 75, + 381, + 468, + 452 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\max _ {\\theta , \\mathcal {P}} \\sum_ {i} g \\left(\\Phi_ {\\theta} \\left(\\mathcal {X} _ {i}, \\{\\mathcal {P} _ {i \\rightarrow j} \\} _ {j = 1} ^ {N}\\right), \\mathcal {O} _ {i} ^ {0}\\right), \\mathrm {s . t .} \\sum_ {i} | \\mathcal {P} _ {i \\rightarrow j} | \\leq B,\n$$\n", + "text_format": "latex", + "bbox": [ + 76, + 459, + 467, + 489 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $g(\\cdot, \\cdot)$ is the detection evaluation metric, $\\Phi(\\cdot)$ is a detection model with trainable parameter $\\theta$ , and $\\mathcal{P}_{i \\to j}$ is the message transmitted from the $i$ th agent to the $j$ th agent. The key challenge is to determine the messages $\\mathcal{P}_{i \\to j}$ , which should be both informative and compact. Our design rational comes from two aspects: First, since the major gap between camera and LiDAR is the depth, the message should include depth information. This would allow different viewpoints from multiple agents to disambiguate the infinite depth possibilities and localize the correct depth candidate. Second, the message should include detection clues to provide complementary detection information, which can fundamentally overcome inevitable limitations of single-agent detection, such as occlusion and long-range issues.", + "bbox": [ + 75, + 497, + 468, + 709 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Based on the above intuition, the proposed CoCa3D includes a) single-agent camera-only 3D detection, which achieves basic depth estimation and detection ability, and b) multi-agent collaboration, which shares both estimated depth information and detection features to improve 3D representation and detection performances; see Fig. 2.", + "bbox": [ + 75, + 710, + 468, + 801 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Single-agent camera-only 3D detection", + "text_level": 1, + "bbox": [ + 76, + 805, + 410, + 821 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Single-agent camera-only 3D object detection network learns to detect 3D objects in the physical space based on the 2D camera inputs. Here we adopt CaDNN [25], an off-the-shelf architecture. The main idea is to project the flat image feature into all the possible depths in the 3D space,", + "bbox": [ + 75, + 825, + 470, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "and then attentively aggregate the 3D voxel feature and collapse them into the BEV feature. To mention that, the BEV feature is used because it largely reduces the computation cost, while performing similarly to 3D voxel features [10]. The architecture consists of five modules as detailed below.", + "bbox": [ + 496, + 324, + 890, + 398 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Encoder. The encoder extracts features from the input images. For agent $i$ , given its input $\\mathcal{X}_i$ , the output feature map of encoder is $\\mathcal{F}_i = \\Phi_{\\mathrm{enc}}(\\mathcal{X}_i) \\in \\mathbb{R}^{H \\times W \\times C}$ , where $H, W, C$ are the height, width and channel of the image feature.", + "bbox": [ + 496, + 402, + 890, + 463 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Depth estimation. The depth estimation module predicts pixel-wise categorical depth distribution over a set of predefined depth bins to accurately locate image information in 3D space. Treating the estimation as a classification problem can capture the inherent depth estimation uncertainty to reduce the impact of erroneous depth estimates, as stated in CaDNN [25]. To achieve this, a parametric depth estimation network $\\Phi_{\\mathrm{Depth}}(\\cdot)$ is used, and the pixel-wise categorical depth distribution is obtained as", + "bbox": [ + 496, + 465, + 890, + 601 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {D} _ {i} = \\Phi_ {\\text {D e p t h}} \\left(\\mathcal {F} _ {i}\\right) \\in \\mathbb {R} ^ {H \\times W \\times D}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 586, + 602, + 890, + 619 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "whose $(h,w)$ th element $\\mathcal{D}_i(h,w)\\in \\mathbb{R}^D$ refers to the estimated depth distribution of the $(h,w)$ th image pixel with $D$ as the number of discretized depth bins.", + "bbox": [ + 496, + 621, + 893, + 667 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Voxel transformation. Voxel transformation projects the image feature to the 3D space, given all the possible depth candidates and the known image calibration matrix, resulting in a 3D voxel representation $\\widehat{\\mathcal{V}}_i\\in \\mathbb{R}^{X\\times Y\\times Z\\times C}$ . The camera projection matrix $\\mathbf{P}\\in \\mathbb{R}^{3\\times 4}$ defines the mapping between the 3D voxel coordinate to the image pixel coordinate, this is, $[u,v,d] = \\mathbf{P}\\cdot [x,y,z,1]$ . Similarly, the depth probabilities can be transformed to the 3D voxel space, resulting in $\\widehat{\\mathcal{D}}_i\\in \\mathbb{R}^{X\\times Y\\times Z}$ , where each element indicates the confidence that the feature pixel belongs to the voxel.", + "bbox": [ + 496, + 670, + 890, + 821 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Collapse. The voxel feature is collapsed to a single height plane to generate the bird's-eye-view (BEV) feature as $\\overline{B}_i = \\Phi_{\\mathrm{clp}}(\\widehat{\\mathcal{D}}_i \\odot \\widehat{\\mathcal{V}}_i) \\in \\mathbb{R}^{X \\times Y \\times C}$ , where $\\odot$ is the element-wise multiplication. Symbol $\\cdot / \\widehat{\\cdot} / \\widehat{\\cdot}$ reflects 2D image / 3D voxel / BEV spaces, respectively. The collapse network $\\Phi_{\\mathrm{clp}}(\\cdot)$", + "bbox": [ + 496, + 825, + 893, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "9245", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "flattens the 3D voxel feature along the $Z$ -axis and then applies $1 \\times 1$ convolution to reduce the channel dimension.", + "bbox": [ + 75, + 90, + 467, + 119 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Decoder. The decoder takes the BEV feature as input, and outputs the objects including class and regression. Here, we adopt CenterNet [44], an off-the-shelf detector. Given the BEV feature $\\overline{B}_i$ , the detection decoder $\\Phi_{\\mathrm{dec}}(\\cdot)$ generate the dense heatmap of $i$ th agent by $\\overline{\\mathcal{O}}_i = \\Phi_{\\mathrm{dec}}(\\overline{B}_i) \\in \\mathbb{R}^{X \\times Y \\times 7}$ , where each location of $\\overline{\\mathcal{O}}_i$ represents a rotated box with class $(c, x, y, h, w, \\cos \\alpha, \\sin \\alpha)$ , denoting class confidence, position residual, size and angle. Non-maximum suppression (NMS) is applied to the dense predictions and generate the final sparse output of the 3D detection system.", + "bbox": [ + 75, + 121, + 467, + 272 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Discussion. Admittedly, single-agent camera-only 3D detection makes localizing objects in the 3D space given 2D image input possible through numerous ingenious algorithm designs. The 3D detection performance is still far from perfect due to the ambiguous single-agent depth estimation and limited visual range. Tackling these two issues are challenging for a single agent, but from a perspective of multi-agent collaboration, sharing complementary visual clues makes depth estimation and visual range enlargement much more natural. This motivates our following designs.", + "bbox": [ + 75, + 275, + 467, + 425 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Multi-agent collaboration", + "text_level": 1, + "bbox": [ + 76, + 431, + 310, + 446 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Multi-agent collaboration includes two parts: i) collaborative depth estimation, which enables the sharing of depth information estimated by individual agents to produce more accurate 3D representation and less biased BEV feature; and ii) collaborative detection feature learning, which enables the sharing of detection features of individual agents to achieve more holistic 3D detection.", + "bbox": [ + 75, + 450, + 467, + 555 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3.1 Collaborative depth estimation", + "text_level": 1, + "bbox": [ + 76, + 561, + 346, + 575 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Collaborative depth estimation (Co-Depth) targets to disambiguate infinite depth possibilities in single-agent camera-only depth estimation and localize the correct depth candidate through multi-view consistency. The intuition is, for a correct depth candidate, its corresponding 3D location should be spatially consistent from multiple agents' viewpoints. To achieve this, each agent can exchange the depth information through communication. Meanwhile, we promote communication efficiency by selecting the most critical and unambiguous depth information. Accordingly, Co-Depth includes: a) depth uncertainty aware message packing, which packs compact messages with unambiguous depth information; and b) depth message fusion, which enhances depth estimation with the received depth messages.", + "bbox": [ + 75, + 580, + 467, + 792 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Depth uncertainty aware message packing. Depth uncertainty-aware (DUA) message packing packs the most critical depth information used for multi-view consistency into the to-be-sent message based on the depth uncertainty. The depth message includes: i) the voxel feature, which is used for the multi-view visual similarity measurement; and ii) the depth probability, which indicates the confidence that", + "bbox": [ + 75, + 795, + 467, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "the feature pixel belongs to the voxel and is used for the multi-view candidate selection.", + "bbox": [ + 498, + 90, + 890, + 119 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The spatially sparse features of $i$ th agent are selected based on its spatial depth uncertainty map in the 2D image space $\\mathbf{U}_i\\in \\mathbb{R}^{H\\times W}$ , that can be represented by the pixelwise entropy from corresponding depth distribution,", + "bbox": [ + 496, + 121, + 890, + 181 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {U} _ {i} (h, w) = \\mathrm {H} (\\mathcal {D} _ {i} (h, w)) \\in \\mathbb {R},\n$$\n", + "text_format": "latex", + "bbox": [ + 589, + 185, + 799, + 200 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathrm{H}(\\cdot)$ is the entropy function and $\\mathcal{D}_i(h,w)$ is the depth distribution at the $(h,w)$ th image pixel, following from (1). The intuition is, for the locations with low uncertainty score, an agent is confident about which depth bin the feature pixel belongs to. Thus, sharing information at these locations can help with depth estimation accuracy. On the contrary, for the locations with high uncertainty score, an agent is hard to tell which depth bin the feature pixel belongs to. Transmitting these spatial features would introduce ambiguity, and even degrade the single-agent depth estimation.", + "bbox": [ + 496, + 204, + 890, + 354 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Specifically, a binary selection matrix is used to represent whether each location is selected or not, where 1 denotes selected, and 0 elsewhere. Let $\\mathbf{M}_{i\\rightarrow j}^{d}\\in \\{0,1\\}^{H\\times W}$ be a selection matrix to determine the depth message sent from agent $i$ to $j$ . Its $(h,w)$ th element is", + "bbox": [ + 496, + 354, + 890, + 431 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {M} _ {i \\rightarrow j} ^ {d} (h, w) = I (\\mathbf {U} _ {i} (h, w) < u _ {\\text {t h r e}}),\n$$\n", + "text_format": "latex", + "bbox": [ + 568, + 433, + 820, + 450 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $I(\\cdot)$ is an indicator function and $u_{\\mathrm{thre}}$ is a predefined threshold. This reflects that a specific spatial area will be selected when its depth uncertainty is below a threshold.", + "bbox": [ + 496, + 453, + 890, + 497 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The selection matrix can be projected in the 3D voxel space, resulting in $\\widehat{\\mathbf{M}}_{i\\rightarrow j}^{d}\\in \\mathbb{R}^{X\\times Y\\times Z}$ . Then, the selected voxel feature map $\\widehat{\\mathcal{Z}}_{i\\rightarrow j}$ and the selected depth distribution $\\widehat{\\mathcal{E}}_{i\\rightarrow j}$ sent from the $i$ th agent to the $j$ th agent are obtained as", + "bbox": [ + 496, + 498, + 890, + 566 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\widehat {\\mathcal {Z}} _ {i \\rightarrow j} = \\widehat {\\mathbf {M}} _ {i \\rightarrow j} ^ {d} \\odot \\widehat {\\mathcal {V}} _ {i} \\in \\mathbb {R} ^ {X \\times Y \\times Z \\times C}, \\\\ \\widehat {\\mathcal {E}} _ {i \\rightarrow j} = \\widehat {\\mathbf {M}} _ {i \\rightarrow j} ^ {d} \\odot \\widehat {\\mathcal {D}} _ {i} \\in \\mathbb {R} ^ {X \\times Y \\times Z}, \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 570, + 568, + 818, + 611 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\widehat{\\mathcal{V}}_i$ and $\\widehat{\\mathcal{D}}_i$ are the full 3D voxel feature map and depth distribution. Overall, the depth message sent from the $i$ th agent to the $j$ th agent is $\\mathcal{P}_{i\\rightarrow j}^{d} = (\\widehat{\\mathcal{E}}_{i\\rightarrow j},\\widehat{\\mathcal{Z}}_{i\\rightarrow j})$ , where both entries are spatially sparse and communication efficient.", + "bbox": [ + 496, + 614, + 890, + 676 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Depth message fusion. Depth message fusion targets to enhance the depth estimation given the received depth messages through different viewpoints of multiple agents. The intuition is that for a correct depth candidate, the visual features at the same 3D point observed by multiple agents should be similar. To achieve this, we introduce multiview depth consistency weighting. Let $\\widehat{S}_i(x,y,z) \\in \\mathbb{R}$ be the matching score between agent $i$ and its neighbors $\\mathcal{N}_i$ , those who share messages with current agent, at coordinate $(x,y,z)$ . The multi-view matching score is obtained as", + "bbox": [ + 496, + 678, + 890, + 820 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\widehat {\\mathcal {W}} _ {j \\rightarrow i} (x, y, z) = I (\\widehat {\\mathcal {E}} _ {j \\rightarrow i} (x, y, z) > p _ {\\text {t h r e}}) \\in \\{0, 1 \\},\n$$\n", + "text_format": "latex", + "bbox": [ + 500, + 824, + 821, + 842 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\widehat {\\mathcal {S}} _ {i} (x, y, z) = \\sum_ {j \\in \\mathcal {N} _ {i}} \\widehat {\\mathcal {W}} _ {j \\rightarrow i} (x, y, z) \\left\\langle \\widehat {\\mathcal {V}} _ {i} (x, y, z), \\widehat {\\mathcal {Z}} _ {j \\rightarrow i} (x, y, z) \\right\\rangle ,\n$$\n", + "text_format": "latex", + "bbox": [ + 501, + 845, + 888, + 876 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\widehat{\\mathcal{W}}_{j\\rightarrow i}(x,y,z)$ is a binary weight used to filter out the", + "bbox": [ + 501, + 882, + 890, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "9246", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "neighboring view that has a low depth confidence, below a threshold $p_{\\mathrm{thre}}$ and $\\langle \\cdot ,\\cdot \\rangle$ is the inner product that measures the visual feature consistency across two views. This means that either the depth candidate is wrong or it is not visible in that view (e.g. due to occlusion), where the multi-view matching score should not be considered.", + "bbox": [ + 75, + 90, + 468, + 181 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Given the multi-view matching score, the collaborative depth distribution is obtained as", + "bbox": [ + 76, + 181, + 468, + 212 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\widehat {\\mathcal {D}} _ {i} ^ {\\prime} = \\Phi_ {\\mathrm {C o D e p t h}} \\left(\\left[ \\widehat {\\mathcal {D}} _ {i}; \\widehat {\\mathcal {S}} _ {i} \\right]\\right) \\in \\mathbb {R} ^ {X \\times Y \\times Z},\n$$\n", + "text_format": "latex", + "bbox": [ + 133, + 212, + 411, + 238 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where the $[\\cdot ]$ denotes concatenation, $\\Phi_{\\mathrm{CoDepth}}(\\cdot)$ is the collaborative depth estimation network implemented with $1\\times 1$ convolutions followed by a sigmoid activation. Each element of the collaborative depth distribution $\\widehat{D}_i^{\\prime}$ reflects the confidence of the feature locates in the corresponding voxel. Note that i) the collaboratively updated depth probability is decided on both the single-view confidence and the multiview consistency; ii) the depth consistency weighting discards the candidates with low single-view depth probability to avoid ambiguity and make the estimation more robust.", + "bbox": [ + 75, + 237, + 468, + 387 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Given this collaborative depth distribution, each agent can weigh the 3D voxel feature and then collapse into a BEV feature with a collapse network $\\Phi_{\\mathrm{clp}}(\\cdot)$ , this is, $\\overline{B}_i = \\Phi_{\\mathrm{clp}}(\\widehat{\\mathcal{D}}_i' \\odot \\widehat{\\mathcal{V}}_i) \\in \\mathbb{R}^{X \\times Y \\times C}$ . The collapse network is implemented with $1 \\times 1$ convolution to reduce channel dimension.", + "bbox": [ + 76, + 388, + 468, + 465 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3.2 Collaborative detection feature learning", + "text_level": 1, + "bbox": [ + 76, + 470, + 408, + 486 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Collaborative depth estimation carefully refines the depth and promotes more accurate 3D representations for each single agent. However, the physical limitations of single agent, such as, limited field of view, occlusion, and long-range issues still remain. To achieve more holistic 3D detection, each agent should be able to exchange 3D detection feature and make use of the complementary information. Meanwhile, we promote communication efficiency by selecting the most perceptually critical information. Therefore, collaborative detection feature learning (Co-FL) includes: a) detection confidence aware message packing, which packs spatially sparse yet perceptually critical 3D features with the guidance of the detection confidence; and ii) detection message fusion, which enhances the 3D feature with the received detection messages.", + "bbox": [ + 75, + 489, + 468, + 715 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Detection confidence aware message packing. The detection confidence aware (DCA) message packing targets to pack the complementary perceptual information into a compact message. The core idea is to explore the spatial heterogeneity of perceptual information. The intuition is that the areas that contain objects are more critical than background areas. During collaboration, areas with objects could help recover the miss-detected objects due to the limited view; and background areas could be omitted to save the precious bandwidth. To achieve this, we implement the spatial confidence map with the detection confidence map by", + "bbox": [ + 75, + 718, + 468, + 883 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\overline {{\\mathbf {C}}} _ {i} = \\Phi_ {\\mathrm {c l s}} (\\overline {{\\mathcal {B}}} _ {i}) \\in \\mathbb {R} ^ {X \\times Y},\n$$\n", + "text_format": "latex", + "bbox": [ + 184, + 883, + 357, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\Phi_{\\mathrm{cls}}(\\cdot)$ denotes the standard classification network in the single-agent decoder $\\Phi_{\\mathrm{det}}(\\cdot), \\overline{B}_i$ is the BEV feature generated with the collaboratively estimated depth. Based on this map, agents decide which spatial area to communicate; that is, each agent offers spatially sparse, yet critical features to support other agents.", + "bbox": [ + 496, + 90, + 890, + 181 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Specifically, a binary selection matrix is used to represent each location is selected or not. Let $\\overline{\\mathbf{M}}_{i\\rightarrow j}^{f}\\in$ $\\{0,1\\}^{X\\times Y}$ be a selection matrix to determine the detection message sent from agent $i$ to $j$ . Its $(x,y)$ th element is", + "bbox": [ + 496, + 181, + 890, + 246 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\overline {{\\mathbf {M}}} _ {i \\rightarrow j} ^ {f} (x, y) = I (\\overline {{\\mathbf {C}}} _ {i} (x, y) > c _ {\\mathrm {t h r e}}),\n$$\n", + "text_format": "latex", + "bbox": [ + 573, + 247, + 815, + 268 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $I(\\cdot)$ is an indicator function and $c_{\\mathrm{thre}}$ is a predefined threshold. This reflects that a specific spatial area will be selected when its detection confidence is above a threshold. Overall, the detection message sent from the $i$ th agent to the $j$ th agent is $\\mathcal{P}_{i\\rightarrow j}^{f} = \\overline{\\mathbf{M}}_{i\\rightarrow j}^{f}\\odot \\overline{\\mathcal{B}_{i}}$ . Only selected features and their indices are packed, and they can be recovered at the receiver. This greatly reduces communication cost.", + "bbox": [ + 496, + 268, + 890, + 377 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Detection message fusion. Here we augment the detection feature of each agent by aggregating the received detection messages from other agents. We implement this with simple but effective non-parametric point-wise maximum fusion. Specifically, for the $i$ th agent, after receiving the $j$ th agent's message $\\mathcal{P}_{j\\rightarrow i}^{f}$ . We also include the ego feature map in fusion and denote $\\mathcal{P}_{i\\rightarrow i}^{f} = \\overline{\\mathcal{B}}_{i}$ to make the formulation simple and consistent, where $\\mathcal{P}_{i\\rightarrow i}^{f}$ might not be sparse. The fused BEV feature is obtained as", + "bbox": [ + 496, + 378, + 890, + 518 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\overline {{\\mathcal {B}}} _ {i} ^ {\\prime} = \\max _ {j \\in \\mathcal {N} _ {i} \\cup \\{i \\}} (\\mathcal {P} _ {j \\rightarrow i} ^ {f}) \\in \\mathbb {R} ^ {X \\times Y \\times C},\n$$\n", + "text_format": "latex", + "bbox": [ + 575, + 518, + 813, + 545 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\max (\\cdot)$ maximizes the corresponding features from multiple agents at each individual spatial location. Note that attention fusion is not permutation invariant, as attention weights vary with the ordering of key and query. Here we simply use the max operator to avoid this permutation variant issue. The fused BEV feature is output to the decoder to generate the final detection.", + "bbox": [ + 496, + 546, + 890, + 651 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Losses", + "text_level": 1, + "bbox": [ + 500, + 655, + 586, + 667 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To train the overall system, we supervise two tasks: categorical depth estimation and 3D object detection. The overall loss is $L = L_{\\mathrm{dep}}(\\mathcal{D}_i,\\mathcal{D}_i^0) + L_{\\mathrm{det}}(\\overline{\\mathcal{O}}_i,\\mathcal{O}_i^0)$ , where $\\mathcal{D}_i^0$ and $\\mathcal{O}_i^0$ is the $i$ th agent's ground-truth depth category and objects, respectively, $L_{\\mathrm{dep}}$ and $L_{\\mathrm{det}}$ is the depth classification loss [25] and detection loss [44], respectively.", + "bbox": [ + 496, + 670, + 890, + 762 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experimental Results", + "text_level": 1, + "bbox": [ + 498, + 768, + 705, + 785 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Datasets and experimental settings", + "text_level": 1, + "bbox": [ + 500, + 790, + 802, + 806 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We cover three datasets, both real-world and simulation scenarios, and multiple agent types. Metric Average Precision (AP) at Intersection-over-Union (IoU) threshold of 0.30, 0.50, 0.70, and 0.80 are used. Metric multi-class classification accuracy is used for depth accuracy. The communication volume follows the standard setting as [6, 36, 37]", + "bbox": [ + 496, + 810, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "9247", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/a2f60f240805255f8f1dc709609a6dca3813614148c32d2799170a0fe999670a.jpg", + "image_caption": [ + "(a) OPV2V+" + ], + "image_footnote": [], + "bbox": [ + 86, + 89, + 315, + 226 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/f3107b653a6f4b821e989cb70177cb3af91bab71ec7271131755d18c51d9fe1e.jpg", + "image_caption": [ + "(b) DAIR-V2X", + "Figure 3. CoCa3D steadily improves 3D detection performance as the number of agents grows. In OPV2V+, around 10 collaborative agents enable collaborative cameras to catch up with LiDAR." + ], + "image_footnote": [], + "bbox": [ + 351, + 90, + 576, + 227 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/154529910b4e8f787067760ef14e99b5a99367e2c998e75248a8825d5add2866.jpg", + "image_caption": [ + "(c) CoPerception-UAVs+" + ], + "image_footnote": [], + "bbox": [ + 612, + 90, + 836, + 227 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/9bcb5a5e415b0c5aedb5056b36f10734438dd1105e1056b99f41f3a25dc574e6.jpg", + "table_caption": [ + "Table 1. CoCa3D significantly outperforms the previous SOTAs, e.g. improve by ${30.60}\\%$ on OPV2V, ${12.59}\\%$ on CoPerception-UAVs, 44.21% on DAIR-V2X for metric AP@70." + ], + "table_footnote": [], + "table_body": "
MethodOPV2V+CoPerception-UAVs+DAIR-V2X
AP@30AP@50AP@70AP@50AP@70AP@80AP@30AP@50AP@70
No Collaboration0.27480.20410.08530.69560.49000.23090.09770.05240.0305
Late Fusion0.65010.61980.51090.72060.53720.25970.20600.10780.0455
When2com (CVPR'20)0.48530.42110.37370.82190.67050.41020.19570.09840.0459
V2VNet (ECCV'20)0.62460.50420.38520.90930.71770.38040.16400.08470.0512
DiscoNet (NeurIPS'21)0.73000.60090.41790.90540.70790.35640.18360.12620.0683
V2X-ViT (ECCV'22)0.83460.66590.39460.90940.71430.35250.18620.10750.0490
Where2comm (NeurIPS'22)0.81910.70890.47410.91020.73830.36760.17540.10250.0547
CoCa3D0.86420.82600.66750.94970.85020.58350.35220.22600.0985
", + "bbox": [ + 78, + 299, + 906, + 454 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "that counts the message size by byte in log scale with base 2. To compare communication results straightforwardly and fair, we do not consider any extra compression.", + "bbox": [ + 75, + 459, + 468, + 505 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "OPV2V+. The original OPV2V [37] is a large-scale vehicle-to-vehicle collaborative perception dataset, cosimulated by OpenCDA [34] and CARLA [4]. Here we introduce an extended version, OPV2V+, which includes more collaborative agents ( $\\approx 10$ ). Each agent is equipt with 4 cameras and 4 depth sensors with resolution $600 \\times 800$ . The detection range is $281.6\\mathrm{m} \\times 80\\mathrm{m}$ . Our single-agent camera-only detector follows CaDDN [25] with 50 depth categories in linear-increasing spacing mode.", + "bbox": [ + 75, + 522, + 468, + 659 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "DAIR-V2X. DAIR-V2X [40] is the only public real-world collaborative perception dataset. It contains two agents: a vehicle and a road-side-unit with image resolution $1080 \\times 1920$ . The perception range is $100\\mathrm{m} \\times 79\\mathrm{m}$ . Our single-agent camera-only detector also follows CaDDN [25].", + "bbox": [ + 75, + 676, + 468, + 752 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "CoPerception-UAVs+. The original CoPerception-UAVs [6] is a large-scale UAV-based collaborative perception dataset, co-simulated by AirSim [27] and CARLA [4]. Here we introduce an extended version, CoPerception-UAVs+, which includes more collaborative agents ( $\\approx 10$ ). Each agent equips with 1 camera and 1 depth sensor with resolution $450 \\times 800$ . The detection range is $192\\mathrm{m} \\times 96\\mathrm{m}$ . Our single-agent camera-only detector follows DVDET [?] with 10 depth categories in uniform spacing mode.", + "bbox": [ + 75, + 768, + 468, + 905 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Quantitative evaluation", + "text_level": 1, + "bbox": [ + 500, + 459, + 718, + 474 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Collaborative camera-only 3D detection overtakes Li-DAR. Fig. 3 investigates how the collaborative camera-only 3D detection performance changes with the number of collaborative agents on multiple datasets. To set up a goal, we consider a LiDAR-based detector implemented by the widely used PointPillar [10]. Fig. 3 (a) shows that: i) with 10 collaborative agents, CoCa3D outperforms the LiDAR 3D detection at both AP@0.5/0.7 on OPV2V+! ii) the detection performance positively increases as the number of collaborative agents. And we then use the slope of OPV2V+ to fit the function of collaborative detection performance with agent number on DAIR-V2X as it only has two agents available. Fig. 3 (b) shows that camera-only 3D detection is expected to outperform LiDAR 3D detection with 7 collaborative vehicles in the real scenario. As LiDAR is not a commonly available equipment for drones, we set a high-quality RGB-D camera instead. Fig. 3 (c) shows that with 3 collaborative drones, camera-only 3D detection outperforms ground-truth depth-based 3D detection. Furthermore, the steadily increasing detection performance with the number of collaborative agents encourages the agents to actively collaborate and achieve consistent improvement.", + "bbox": [ + 496, + 484, + 890, + 818 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Benchmark comparison. Tab. 1 compares the proposed CoCa3D with previous collaborative methods. We consider single-agent detection without collaboration (No collaboration), When2com [17], V2VNet [31], DiscoNet [14], V2XViT [36], Where2comm [6] and late fusion, where agents", + "bbox": [ + 496, + 825, + 892, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "9248", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/574773ea293cbae2c7aa1e882b024d3a824a3bd276803cc97f71241ea290af98.jpg", + "table_caption": [ + "Table 2. CoCa3D outperforms single-agent camera-only 3D detection with GT depth on OPV2V+ and CoPerception-UAVs+." + ], + "table_footnote": [], + "table_body": "
Co -DepthCo -FLOPV2V+CoPerception-UAVs+
AP@30AP@50AP@70AP@50AP@70AP@80
--0.27480.20410.08530.72130.54210.2846
GT-0.34540.25530.09730.83470.67640.4120
-✓0.82010.71910.47560.90840.72560.4028
GT✓0.91200.88050.74340.95050.83980.5504
✓✓0.86420.82600.66750.94950.85180.5849
", + "bbox": [ + 76, + 119, + 478, + 223 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "directly exchange the detected 3D boxes. We see that the proposed CoCa3D significantly outperforms previous state-of-the-arts, improves the SOTA performance by $30.60\\%$ on OPV2V+, $12.59\\%$ on CoPerception-UAVs+, $44.21\\%$ on DAIR-V2X for AP@70. The reason is that collaborative depth estimation promotes more accurate 3D features per single agent, and the enhanced single agent features further facilitate collaborative detection feature learning, while previous collaborative methods do not specifically consider depth ambiguity and collaborate over single-agent features.", + "bbox": [ + 75, + 239, + 468, + 391 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Multi-agent collaboration evaluation. Tab. 2 assess the effectiveness of the proposed multi-agent collaboration on OPV2V+ and CoPerception-UAVs+ datasets. Ground-truth (GT) depth is included to provide the upper bound of collaborative camera-only 3D detection. We see that: i) both collaborations can consistently improve the performance; ii) camera-only 3D detection with collaboratively estimated depth significantly outperforms that with single-agent estimated depth and even approaches the upper bound with ground truth depth.", + "bbox": [ + 75, + 393, + 468, + 544 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Qualitative evaluation", + "text_level": 1, + "bbox": [ + 76, + 556, + 282, + 571 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Visualization of depth and uncertainty. Fig. 5 shows that collaborative depth estimation outperforms single-agent depth estimation and approaches the ground truth depth. We see that: i) the single-agent depth estimation can estimate the relative depth while not able to precisely localize the depth candidate, for example, the vehicles are higher than the plane it sits while this plane is not grounded to the right category; ii) by introducing multi-view geometry, the collaboratively estimated depth can smoothly and accurately ground the plane; iii) the depth uncertainty is larger for the long-range and background areas. The reason is that the distant regions are hard to localize as they occupy too few image pixels, and the background regions are hard to localize due to the texture-less surfaces.", + "bbox": [ + 75, + 582, + 468, + 792 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Visualization of detection results. Fig. 6 shows that compared to single LiDAR 3D detection, cameras with collaborations (Co-Depth and Co-FL) are able to achieve holistic and accurate detection results. The reason is that the single agent has some fundamental physical limitations such as long-range and occlusion issues. Fig. 6(a-c) shows that augmenting a single agent sensor going from just a cam", + "bbox": [ + 75, + 795, + 468, + 901 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/7777d30890ed305547aa55805548455d59034ed95194532e5a35d0e5ef057308.jpg", + "image_caption": [ + "(a) OPV2V+", + "Figure 4. CoCa3D achieves superior detection performance and communication cost trade-off over various bandwidths." + ], + "image_footnote": [], + "bbox": [ + 500, + 88, + 694, + 227 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/1efed90c2d1174b3e405e6cdf2e94a7b0bc072f0c7b1c2f2cef2af5e12efc3f6.jpg", + "image_caption": [ + "(b) CoPerception-UAVs+" + ], + "image_footnote": [], + "bbox": [ + 694, + 89, + 885, + 227 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/62d0dea4fa7874fb7307635b70843e8d2e34a53bd310416a6ae123446825300f.jpg", + "table_caption": [ + "Table 3. Co-Depth significantly improves depth accuracy." + ], + "table_footnote": [], + "table_body": "
Full planeForeground
SingleCollaborationGainSingleCollaborationGain
0.61670.778126.17% ↑0.75860.81337.21% ↑
", + "bbox": [ + 500, + 296, + 890, + 342 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/a930e6a802daf33f229a965a9d75cdc29f045c19d499cc53549b75a6b9f7f407.jpg", + "table_caption": [ + "Table 4. Ablation of collaborative depth estimation on CoPerception-UAVs+. Uniform/Linear denotes uniform/linear-increasing spacing. Dense and sparse supervision denote applying depth supervision over the full image plane and the object regions." + ], + "table_footnote": [], + "table_body": "
SpacingSupervisionAP@50AP@70AP@80
UniformDense0.95020.84440.5746
UniformSparse0.93300.80420.4829
LinearDense0.94840.83650.5580
LinearSparse0.94950.85180.5849
", + "bbox": [ + 529, + 411, + 859, + 492 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "era to including a depth sensor to LiDAR can improve the detection but still not able to achieve holistic detection. Fig. 6(d)/(b) and (a)/(e) compares the detection results with and without collaborative feature learning, we see that collaboration can help detect lots of the missed objects in the single-agent detection. Fig. 6(d,e,f) show that: i) both collaborations, Co-FL and Co-Depth, consistently improve the detection performance; ii) with collaboration, collaborative camera-only agents can outperform LiDAR in 3D detection.", + "bbox": [ + 496, + 508, + 890, + 643 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Ablation studies", + "text_level": 1, + "bbox": [ + 500, + 652, + 660, + 667 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Trade-off between detection performance and communication cost. Fig. 4 compares the proposed CoCa3D with the previous communication-efficient solution Where2comm [6] in terms of the trade-off between detection performance (AP@IoU=50/70) and communication bandwidth. We see that: i) CoCa3D can adapt to varying bandwidths by adjusting the depth uncertainty threshold and the detection confidence threshold; ii) CoCa3D achieves superior detection performance and communication cost trade-off to Where2comm over varying communication bandwidth. The gain mainly comes from Co-Depth's improvement in depth estimation, which improves single-agent BEV features. With enhanced single BEV features, the Co-FL can generate better-augmented BEV features, resulting in improved detection performance.", + "bbox": [ + 496, + 674, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "9249", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/49de5e98a900e8abb045db6097aa3993efcc7c528e71b8f4b90dcf6750adbe55.jpg", + "image_caption": [ + "(a) Image (RV)" + ], + "image_footnote": [], + "bbox": [ + 83, + 111, + 232, + 170 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/3fc22249c1fb463024213745e99f77430f96b8744179701430810a2aca1e2fee.jpg", + "image_caption": [ + "(b) Single (RV)" + ], + "image_footnote": [], + "bbox": [ + 245, + 98, + 383, + 183 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/d746ffc473f0c0594175c27fed6605a5b25699246da947bda10e7d22896aef5d.jpg", + "image_caption": [ + "(c)Collaboration (RV)" + ], + "image_footnote": [], + "bbox": [ + 408, + 98, + 549, + 183 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/4019488b619a7896b4c0c67e40c332734d8f2995f3050e79fa04461aa19a7502.jpg", + "image_caption": [ + "(d) Ground-truth (RV)" + ], + "image_footnote": [], + "bbox": [ + 573, + 98, + 712, + 183 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/aa6856ffbdc0eeea699590483143f3d64ca6696b8600dfa42d15a6686aed6910.jpg", + "image_caption": [ + "(e) Uncertainty (RV)" + ], + "image_footnote": [], + "bbox": [ + 736, + 98, + 880, + 184 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/73672d6740f21637bb69e8ea5101c0fb0aa14f4de5ff80c10b926cd92fc006cd.jpg", + "image_caption": [ + "(f) Image (BEV)" + ], + "image_footnote": [], + "bbox": [ + 83, + 217, + 232, + 277 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/e6291efcddbb4e23c71b20dfe063e7e8970380bdd63dc36cd19e54b3479b9c38.jpg", + "image_caption": [ + "(g) Single (BEV)" + ], + "image_footnote": [], + "bbox": [ + 245, + 204, + 383, + 287 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/4f1144a47c1d8e63a11af41658ae55578cb7fafc149bdc84ba85a13e5c349c11.jpg", + "image_caption": [ + "(h) Collaboration (BEV)" + ], + "image_footnote": [], + "bbox": [ + 408, + 205, + 547, + 287 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/de294883f4a13cb7c4b41047ad8ae3ba2eba029777cac267532ca40dedffa61f.jpg", + "image_caption": [ + "(i) Ground-truth (BEV)" + ], + "image_footnote": [], + "bbox": [ + 570, + 205, + 710, + 287 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/4fe9ba96c50e05d317ae2fab159eeda1156aff9ea17b892d99293ef81a1c6764.jpg", + "image_caption": [ + "(j) Uncertainty (BEV)" + ], + "image_footnote": [], + "bbox": [ + 733, + 205, + 879, + 289 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/c2bb5012670046e2025e884429e6cf073212ba3da5ea009e528deea4069c0f0f.jpg", + "image_caption": [ + "Figure 5. Visualization of the depth and uncertainty in image range-view (RV) and bird's eye view (BEV) on CoPerception-UAVs+. Collaboratively estimated depth improves the single-agent estimated depth and approaches ground-truth depth." + ], + "image_footnote": [], + "bbox": [ + 78, + 343, + 341, + 425 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/ce287aecd05eb7304316a8822cb786807001242aeda91ee6e0a410610d4acbed.jpg", + "image_caption": [ + "(b) RGB-D" + ], + "image_footnote": [], + "bbox": [ + 352, + 343, + 614, + 425 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/81df8c719b4e4cc6d7e17501777c517f84540672e18322fc4ad5defc851ca1e6.jpg", + "image_caption": [ + "(c) LiDAR" + ], + "image_footnote": [], + "bbox": [ + 627, + 343, + 890, + 425 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/c1b171869492ab1e4755a397d005d66d30997625f0473ecdb2f25f353a344e20.jpg", + "image_caption": [ + "(a) Camera", + "(d) Camera with Co-FL", + "Figure 6. CoCa3D outperforms LiDAR detection on OPV2V+ with 10 agents. Green and red boxes denote GT and detection respectively." + ], + "image_footnote": [], + "bbox": [ + 78, + 441, + 339, + 523 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/8e6cd7224aaea17acc63f899b115ab304f18b4fc78917d02a20460a5064e3d1b.jpg", + "image_caption": [ + "(e) RGB-D with Co-FL" + ], + "image_footnote": [], + "bbox": [ + 351, + 441, + 612, + 523 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/a9af25cbb9ff2f40e35bed9a78b359d4dee4cb964cd8837399df92d88a24c7a5.jpg", + "image_caption": [ + "(f) Camera with Co-Depth and Co-FL (CoCa3D)" + ], + "image_footnote": [], + "bbox": [ + 624, + 441, + 887, + 523 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effect of components in collaborative depth estimation. Tab. 3 assesses the gain of collaborative depth estimation (Co-Depth) over single-agent depth estimation (S-Depth) in the depth accuracy metric. We see that: i) Co-Depth steadily improves S-Depth over the full plane; ii) the depth accuracy of the foreground objects is higher than the full plane, as the texture-less backgrounds have fewer cues to localize. Tab. 4 assesses the effects of the depth spacing and supervision choices. We see that: i) overall, the Co-Depth is robust to the different spacing and supervision choices on AP@50/70; ii) linear-increasing spacing is robust to the various supervision choices that uniform spacing which fails the stricter metric AP@80 without dense supervision. The reason is that the linearly increasing spacing takes into account the prior depth distribution and assigns more depth candidates to frequently occurring depth ranges, which contributes to more balanced and easier depth learning.", + "bbox": [ + 75, + 561, + 468, + 819 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion and Limitation", + "text_level": 1, + "bbox": [ + 76, + 840, + 326, + 857 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We propose CoCa3D, a novel collaborative camera-only 3D detection that approaches holistic 3D detection. The", + "bbox": [ + 76, + 869, + 468, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "core idea is to introduce multi-agent collaboration to improve the detection ability of cameras. Meanwhile, the communication cost is optimized, and each agent carefully selects the spatially sparse yet critical depth messages to share. Extensive experiments covering both real-world and simulation scenarios, and multi-type agents (cars, drones, and infrastructures) show that CoCa3D not only achieves state-of-the-art perception-bandwidth trade-off, but overtakes LiDAR-based detectors with a sufficient number of collaborative agents on OPV2V+.", + "bbox": [ + 496, + 561, + 890, + 712 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Limitation and future work. It is expensive to collect a real-world multi-agent perception dataset. So far, DAIR-V2X is the sole public real-world dataset, which only has one vehicle and one roadside unit. This work mainly leverages simulation data to validate the proposed novel methods and sketch a promising research direction. We advocate more resources for real-world data collection.", + "bbox": [ + 496, + 713, + 890, + 816 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgment. This research is supported by the National Key RD Program of China under Grant 2021ZD0112801, NSFC under Grant 62171276 and the Science and Technology Commission of Shanghai Municipal under Grant 21511100900 and 22DZ2229005.", + "bbox": [ + 496, + 818, + 890, + 893 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "9250", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 174, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In European conference on computer vision, pages 213-229. Springer, 2020. 1", + "[2] Siheng Chen, Baoan Liu, Chen Feng, Carlos Vallespi-Gonzalez, and Carl K. Wellington. 3d point cloud processing and learning for autonomous driving: Impacting map creation, localization, and perception. IEEE Signal Processing Magazine, 38:68-86, 2021. 1", + "[3] Jiajun Deng, Shaoshuai Shi, Peiwei Li, Wengang Zhou, Yanyong Zhang, and Houqiang Li. Voxel r-cnn: Towards high performance voxel-based 3d object detection. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 35, pages 1201-1209, 2021. 1", + "[4] Alexey Dosovitskiy, German Ros, Felipe Codevilla, Antonio Lopez, and Vladlen Koltun. Carla: An open urban driving simulator. In Conference on robot learning, pages 1-16. PMLR, 2017. 6", + "[5] Peng Gao, Rui Guo, Hongsheng Lu, and Hao Zhang. Regularized graph matching for correspondence identification under uncertainty in collaborative perception. Robotics: Science and Systems XVI, 2020. 2", + "[6] Yue Hu, Shaoheng Fang, Zixing Lei, Yiqi Zhong, and Siheng Chen. Where2comm: Communication-efficient collaborative perception via spatial confidence maps. Advances in Neural Information Processing Systems, 2022. 2, 5, 6, 7", + "[7] Yue Hu, Shaoheng Fang, Weidi Xie, and Siheng Chen. Aerial monocular 3d object detection. IEEE Robotics and Automation Letters, 8:1959-1966, 2022. 1, 2", + "[8] Junjie Huang and Guan Huang. Bevdet4d: Exploit temporal cues in multi-camera 3d object detection. arXiv preprint arXiv:2203.17054, 2022. 1, 2", + "[9] Junjie Huang, Guan Huang, Zheng Zhu, and Dalong Du. Bevdet: High-performance multi-camera 3d object detection in bird-eye-view. arXiv preprint arXiv:2112.11790, 2021.1, 2", + "[10] Alex H Lang, Sourabh Vora, Holger Caesar, Lubing Zhou, Jiong Yang, and Oscar Beijbom. Pointpillars: Fast encoders for object detection from point clouds. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12697-12705, 2019. 1, 2, 3, 6", + "[11] Zixing Lei, Shunli Ren, Yue Hu, Wenjun Zhang, and Siheng Chen. Latency-aware collaborative perception. ECCV, 2022. 2", + "[12] Jing Li, Rui Li, Jiehao Li, Junzheng Wang, Qingbin Wu, and Xu Liu. Dual-view 3d object recognition and detection via lidar point cloud and camera image. Robotics Auton. Syst., 150:103999, 2022. 1", + "[13] Yiming Li, Dekun Ma, Ziyan An, Zixun Wang, Yiqi Zhong, Siheng Chen, and Chen Feng. V2x-sim: Multi-agent collaborative perception dataset and benchmark for autonomous driving. IEEE Robotics and Automation Letters, 2022. 2", + "[14] Yiming Li, Shunli Ren, Pengxiang Wu, Siheng Chen, Chen Feng, and Wenjun Zhang. Learning distilled collaboration graph for multi-agent perception. Advances in Neural Information Processing Systems, 34:29541-29552, 2021. 2, 6" + ], + "bbox": [ + 78, + 114, + 470, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[15] Yiming Li, Juexiao Zhang, Dekun Ma, Yue Wang, and Chen Feng. Multi-robot scene completion: Towards task-agnostic collaborative perception. In Conference on Robot Learning, 2022. 2", + "[16] Zhiqi Li, Wenhai Wang, Hongyang Li, Enze Xie, Chonghao Sima, Tong Lu, Qiao Yu, and Jifeng Dai. Bevformer: Learning bird's-eye-view representation from multi-camera images via spatiotemporal transformers. arXiv preprint arXiv:2203.17270, 2022. 2", + "[17] Yen-Cheng Liu, Junjiao Tian, Nathaniel Glaser, and Zsolt Kira. When2com: Multi-agent perception via communication graph grouping. In Proceedings of the IEEE/CVF Conference on computer vision and pattern recognition, pages 4106-4115, 2020. 2, 6", + "[18] Yen-Cheng Liu, Junjiao Tian, Chih-Yao Ma, Nathan Glaser, Chia-Wen Kuo, and Zsolt Kira. Who2com: Collaborative perception via learnable handshake communication. In 2020 IEEE International Conference on Robotics and Automation (ICRA), pages 6876-6883. IEEE, 2020. 2", + "[19] Zhijian Liu, Haotian Tang, Alexander Amini, Xinyu Yang, Huizi Mao, Daniela Rus, and Song Han. Bevfusion: Multi-task multi-sensor fusion with unified bird's-eye view representation. ArXiv, abs/2205.13542, 2022. 1", + "[20] Yifan Lu, Quanhao Li, Baoan Liu, Mehrdad Dianati, Chen Feng, Siheng Chen, and Yanfeng Wang. Robust collaborative 3d object detection in presence of pose errors. ICRA, 2023. 2", + "[21] Jiageng Mao, Minzhe Niu, Haoyue Bai, Xiaodan Liang, Hang Xu, and Chunjing Xu. Pyramid r-cnn: Towards better performance and adaptability for 3d object detection. ICCV, 2021. 1", + "[22] Lang Peng, Zhirong Chen, Zhangjie Fu, Pengpeng Liang, and Erkang Cheng. Bevsegformer: Bird's eye view semantic segmentation from arbitrary camera rigs. arXiv preprint arXiv:2203.04050, 2022. 2", + "[23] Jonah Philion and Sanja Fidler. Lift, splat, shoot: Encoding images from arbitrary camera rigs by implicitly unprojecting to 3d. In European Conference on Computer Vision, pages 194-210. Springer, 2020. 1, 2", + "[24] Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. Pointnet: Deep learning on point sets for 3d classification and segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 652-660, 2017. 2", + "[25] Cody Reading, Ali Harakeh, Julia Chae, and Steven L. Waslander. Categorical depth distribution network for monocular 3d object detection. CVPR, 2021. 1, 2, 3, 5, 6", + "[26] Thomas Roddick, Alex Kendall, and Roberto Cipolla. Orthographic feature transform for monocular 3d object detection. arXiv preprint arXiv:1811.08188, 2018. 1, 2", + "[27] Shital Shah, Debadeepta Dey, Chris Lovett, and Ashish Kapoor. Airsim: High-fidelity visual and physical simulation for autonomous vehicles. In Field and service robotics, pages 621-635. Springer, 2018. 6", + "[28] Shaoshuai Shi, Chaoxu Guo, Li Jiang, Zhe Wang, Jianping Shi, Xiaogang Wang, and Hongsheng Li. Pv-rcnn: Point" + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9251", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "voxel feature set abstraction for 3d object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10529-10538, 2020. 1, 2", + "[29] Shaoshuai Shi, Li Jiang, Jiajun Deng, Zhe Wang, Chaoxu Guo, Jianping Shi, Xiaogang Wang, and Hongsheng Li. Pvrconn++: Point-voxel feature set abstraction with local vector representation for 3d object detection. arXiv preprint arXiv:2102.00463, 2021. 2", + "[30] Tai Wang, Jiangmiao Pang, and Dahua Lin. Monocular 3d object detection with depth from motion. In European Conference on Computer Vision (ECCV), 2022. 1, 2", + "[31] Tsun-Hsuan Wang, Sivabalan Manivasagam, Ming Liang, Bin Yang, Wenyuan Zeng, and Raquel Urtasun. V2vnet: Vehicle-to-vehicle communication for joint perception and prediction. In European Conference on Computer Vision, pages 605–621. Springer, 2020. 2, 6", + "[32] Yue Wang, Vitor Campagnolo Guizilini, Tianyuan Zhang, Yilun Wang, Hang Zhao, and Justin Solomon. Detr3d: 3d object detection from multi-view images via 3d-to-2d queries. In Conference on Robot Learning, pages 180–191. PMLR, 2022. 1", + "[33] Pengxiang Wu, Siheng Chen, and Dimitris N. Metaxas. Motionnet: Joint perception and motion prediction for autonomous driving based on bird's eye view maps. 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11382-11392, 2020. 1", + "[34] Runsheng Xu, Yi Guo, Xu Han, Xin Xia, Hao Xiang, and Jiaqi Ma. Opencda: an open cooperative driving automation framework integrated with co-simulation. In 2021 IEEE International Intelligent Transportation Systems Conference (ITSC), pages 1155-1162. IEEE, 2021. 6", + "[35] Runsheng Xu, Zhengzhong Tu, Hao Xiang, Wei Shao, Bolei Zhou, and Jiaqi Ma. CoBEVT: Cooperative bird's eye view semantic segmentation with sparse transformers. CoRL, 2022. 2", + "[36] Runsheng Xu, Hao Xiang, Zhengzhong Tu, Xin Xia, Ming-Hsuan Yang, and Jiaqi Ma. V2x-vit: Vehicle-to-everything cooperative perception with vision transformer. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXXIX, pages 107-124. Springer, 2022. 2, 5, 6", + "[37] Runsheng Xu, Hao Xiang, Xin Xia, Xu Han, Jinlong Li, and Jiaqi Ma. Opv2v: An open benchmark dataset and fusion pipeline for perception with vehicle-to-vehicle communication. In 2022 International Conference on Robotics and Automation (ICRA), pages 2583–2589. IEEE, 2022. 2, 5, 6", + "[38] Zetong Yang, Yanan Sun, Shu Liu, and Jiaya Jia. 3dssd: Point-based 3d single stage object detector. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11040-11048, 2020. 2", + "[39] Tianwei Yin, Xingyi Zhou, and Philipp Krahenbuhl. Center-based 3d object detection and tracking. 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11779-11788, 2021. 1", + "[40] Haibao Yu, Yizhen Luo, Mao Shu, Yiyi Huo, Zebang Yang, Yifeng Shi, Zhenglong Guo, Hanyu Li, Xing Hu, Jirui Yuan, and Zaiqing Nie. Dair-v2x: A large-scale dataset for vehicleinfrastructure cooperative 3d object detection. In IEEE/CVF" + ], + "bbox": [ + 78, + 90, + 468, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Conf. on Computer Vision and Pattern Recognition (CVPR), June 2022. 2, 6", + "[41] Yunpeng Zhang, Zheng Zhu, Wenzhao Zheng, Junjie Huang, Guan Huang, Jie Zhou, and Jiwen Lu. Beverse: Unified perception and prediction in birds-eye-view for vision-centric autonomous driving. arXiv preprint arXiv:2205.09743, 2022. 1, 2", + "[42] Wu Zheng, Weiliang Tang, Sijin Chen, Li Jiang, and Chi-Wing Fu. Cia-ssd: Confident iou-aware single-stage object detector from point cloud. In Proceedings of the AAAI conference on artificial intelligence, volume 35, pages 3555–3562, 2021. 1, 2", + "[43] Brady Zhou and Philipp Kähenbuhl. Cross-view transformers for real-time map-view semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13760-13769, 2022. 1, 2", + "[44] Xingyi Zhou, Dequan Wang, and Philipp Krahenbuhl. Objects as points. In arXiv preprint arXiv:1904.07850, 2019. 4, 5", + "[45] Yin Zhou and Oncel Tuzel. Voxelnet: End-to-end learning for point cloud based 3d object detection. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4490-4499, 2018. 1, 2" + ], + "bbox": [ + 501, + 90, + 893, + 416 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "9252", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2023/Collaboration Helps Camera Overtake LiDAR in 3D Detection/0d6cf930-2724-4762-9022-d5c25917e98a_model.json b/2023/Collaboration Helps Camera Overtake LiDAR in 3D Detection/0d6cf930-2724-4762-9022-d5c25917e98a_model.json new file mode 100644 index 0000000000000000000000000000000000000000..35a74a079c00422d969783a36b67e051ff837c60 --- /dev/null +++ b/2023/Collaboration Helps Camera Overtake LiDAR in 3D Detection/0d6cf930-2724-4762-9022-d5c25917e98a_model.json @@ -0,0 +1,2651 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.169, + 0.131, + 0.803, + 0.154 + ], + "angle": 0, + "content": "Collaboration Helps Camera Overtake LiDAR in 3D Detection" + }, + { + "type": "text", + "bbox": [ + 0.116, + 0.179, + 0.854, + 0.199 + ], + "angle": 0, + "content": "Yue Hu\\(^{1}\\) Yifan Lu\\(^{1}\\) Runsheng Xu\\(^{2}\\) Weidi Xie\\(^{1,3}\\) Siheng Chen\\(^{1,3*}\\) Yanfeng Wang\\(^{3,1}\\)" + }, + { + "type": "text", + "bbox": [ + 0.195, + 0.199, + 0.775, + 0.234 + ], + "angle": 0, + "content": "\\(^{1}\\)Cooperative Medianet Innovation Center, Shanghai Jiao Tong University \\(^{2}\\)University of California, Los Angeles \\(^{3}\\)Shanghai AI Laboratory" + }, + { + "type": "text", + "bbox": [ + 0.106, + 0.234, + 0.863, + 0.253 + ], + "angle": 0, + "content": "1{18671129361, yifan.lu, weidi, sihengc, wangyanfeng}@sjtu.edu.cn 2{rxx3386}@ucla.edu" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.287, + 0.313, + 0.303 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.319, + 0.474, + 0.697 + ], + "angle": 0, + "content": "Camera-only 3D detection provides an economical solution with a simple configuration for localizing objects in 3D space compared to LiDAR-based detection systems. However, a major challenge lies in precise depth estimation due to the lack of direct 3D measurements in the input. Many previous methods attempt to improve depth estimation through network designs, e.g., deformable layers and larger receptive fields. This work proposes an orthogonal direction, improving the camera-only 3D detection by introducing multi-agent collaborations. Our proposed collaborative camera-only 3D detection (CoCa3D) enables agents to share complementary information with each other through communication. Meanwhile, we optimize communication efficiency by selecting the most informative cues. The shared messages from multiple viewpoints disambiguate the single-agent estimated depth and complement the occluded and long-range regions in the single-agent view. We evaluate CoCa3D in one real-world dataset and two new simulation datasets. Results show that CoCa3D improves previous SOTA performances by \\(44.21\\%\\) on DAIR-V2X, \\(30.60\\%\\) on OPV2V+, \\(12.59\\%\\) on CoPerception-UAVs+ for AP@70. Our preliminary results show a potential that with sufficient collaboration, the camera might overtake LiDAR in some practical scenarios. We released the dataset and code." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.717, + 0.21, + 0.733 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.739, + 0.47, + 0.876 + ], + "angle": 0, + "content": "As a fundamental task of computer vision, 3D object detection aims to localize objects in the 3D physical space given an agent's real-time sensor inputs. It is crucial in a wide range of applications, including autonomous driving [10,33,43], surveillance systems [39], robotics [12] and unmanned aerial vehicles [7]. Depending on sensor setups, there are multiple technical solutions to realize 3D object detection. In this spectrum, one extreme emphasizes raising the upper bound of detection performance, which uses" + }, + { + "type": "image", + "bbox": [ + 0.507, + 0.279, + 0.728, + 0.328 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.602, + 0.328, + 0.64, + 0.337 + ], + "angle": 0, + "content": "(a) Scene" + }, + { + "type": "image", + "bbox": [ + 0.728, + 0.279, + 0.882, + 0.328 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.748, + 0.328, + 0.865, + 0.337 + ], + "angle": 0, + "content": "(b) Single camera confidence" + }, + { + "type": "image", + "bbox": [ + 0.507, + 0.337, + 0.727, + 0.386 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.568, + 0.386, + 0.671, + 0.394 + ], + "angle": 0, + "content": "(c) Collaborative cameras" + }, + { + "type": "image", + "bbox": [ + 0.728, + 0.337, + 0.882, + 0.387 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.731, + 0.387, + 0.881, + 0.394 + ], + "angle": 0, + "content": "(d) Collaborative cameras confidence" + }, + { + "type": "image", + "bbox": [ + 0.506, + 0.395, + 0.727, + 0.444 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.587, + 0.444, + 0.654, + 0.455 + ], + "angle": 0, + "content": "(e) LiDAR (Ego)" + }, + { + "type": "image", + "bbox": [ + 0.728, + 0.395, + 0.882, + 0.444 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.762, + 0.445, + 0.849, + 0.454 + ], + "angle": 0, + "content": "(f) LiDAR confidence" + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.456, + 0.892, + 0.498 + ], + "angle": 0, + "content": "Figure 1. Collaborative camera-only 3D detection can disambiguate the single-view estimated depth, address the long-range and occlusion issues, and approach LiDAR in 3D detection." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.506, + 0.893, + 0.628 + ], + "angle": 0, + "content": "high-end LiDAR sensor [2, 3, 10, 21, 28, 42, 45] to collect precise 3D measurements. However, this approach is too expensive to scale up. The other extreme solution emphasizes cost effectiveness, which tries to use thrifty sensor setups, e.g. only using cameras to detect 3D objects in real-time [1, 8, 9, 23, 25, 26, 30, 32, 41, 43]. However, camera-only 3D detection is significantly and consistently worse than LiDAR-based detection in most scenarios [19]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.629, + 0.895, + 0.902 + ], + "angle": 0, + "content": "In this paper, we propose an orthogonal direction for improving camera-only 3D detection performances by introducing multi-agent collaborations. Hypothetically, empowered by advanced communication systems, multiple agents equipped only with cameras could share visual information with each other. This would bring three outstanding benefits. First, different viewpoints from multiple agents can largely resolve the depth ambiguity issue in camera-only 3D detection, bridging the gap with expensive LiDARs on depth estimation. Second, multi-agent collaboration avoids inevitable limitations in single-agent 3D detection, such as occlusion and long-range issues, and potentially enables more holistic 3D detection; that is, detecting all the objects existed in the 3D scene, including those beyond visual range. Since LiDAR also suffers from limited field of view, this potentially enables collaborative cameras to outperform LiDAR. Third, the total expense of a large fleet of vehicles is significantly reduced as cameras are much" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.888, + 0.222, + 0.9 + ], + "angle": 0, + "content": "*Corresponding author." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9243" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.166 + ], + "angle": 0, + "content": "cheaper than LiDAR. However, multi-agent collaboration also brings new challenges. Different from many multiview geometry problems, here we also have to concern communication bandwidth constraints. Thus, each agent needs to select the most informative cues to share." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.168, + 0.47, + 0.363 + ], + "angle": 0, + "content": "Following this design rationale, we propose a novel collaborative camera-only 3D detection framework CoCa3D. It includes three parts: i) single-agent camera-only 3D detection, which achieves basic depth estimation and 3D detection for each agent; ii) collaborative depth estimation, which disambiguates the estimated depths by promoting spatial consistency across multiple agents' viewpoints; and iii) collaborative detection feature learning, which complements detection features by sharing key detection messages with each other. Compared to recent collaborative perception methods [11,14] that are dealing with LiDAR, CoCa3D specifically designs novel collaborative depth estimation to customize the task of camera-only 3D detection." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.364, + 0.47, + 0.575 + ], + "angle": 0, + "content": "To evaluate CoCa3D, we conduct comprehensive experiments on one real-world dataset, DAIR-V2X [40], and two new simulation datasets, OPV2V+ and CoPerception-UAVs+, which are extended based on original OPV2V [37] and CoPerception-UAVs [6] with more collaborative agents that cover three types of agents (cars, infrastructures and drones). Our results show that i) with 10 collaborative agents, CoCa3D enables camera-only detectors to overtake LiDAR-based detectors on OPV2V+; and ii) CoCa3D consistently outperforms previous works in the performance-bandwidth trade-off across multiple datasets by a large margin, improving the previous SOTA performances by \\(30.60\\%\\) on OPV2V+, \\(12.59\\%\\) on CoPerception-UAVs+, \\(44.21\\%\\) on DAIR-V2X for AP@70. To sum up, our contributions are:" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.575, + 0.469, + 0.635 + ], + "angle": 0, + "content": "- We propose a novel collaborative camera-only 3D detection framework CoCa3D, which improves the detection ability of cameras with multi-agent collaboration, promoting more holistic 3D detection." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.636, + 0.469, + 0.725 + ], + "angle": 0, + "content": "- We propose core communication-efficient collaboration techniques, which explore the spatially sparse yet critical depth messages and tackle the depth ambiguity, occlusion, and long-range issues by fusing complementary information from different viewpoints, achieving more accurate and complete 3D representation." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.726, + 0.469, + 0.815 + ], + "angle": 0, + "content": "- We expand two previous collaborative datasets with more agents, and conduct extensive experiments, validating that i) CoCa3D significantly bridges the performance gap between camera and LiDAR on OPV2V+ and DAIR-V2X; and ii) CoCa3D achieves the state-of-the-art performance-bandwidth trade-off." + }, + { + "type": "list", + "bbox": [ + 0.076, + 0.575, + 0.469, + 0.815 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.819, + 0.219, + 0.835 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.469, + 0.902 + ], + "angle": 0, + "content": "Camera-only 3D object detection. Camera-only 3D detection aims to detect objects in the 3D space given the 2D image through explicit or implicit depth estimation [7, 9, 23, 26, 30, 41]. Recently, bird's-eye-view (BEV) repre" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.273 + ], + "angle": 0, + "content": "sentations are widely used for their computation efficiency and comparable performance compared to the 3D voxel features [10]. To get the BEV features, there are two types of methods. The depth-based methods [8, 23, 25] first estimate the depth distribution, and then attentively project the 2D image features along the projection ray to get the 3D voxel features and collapse it to BEV feature. The query-based methods [16, 22, 43] first initialize queries for each BEV grid and then leverage transformer-architecture-based cross-attention to query image features with camera-aware positional embedding. Here, our single-agent camera-only detector follows the simple-yet-effective CaDDN [25]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.274, + 0.892, + 0.502 + ], + "angle": 0, + "content": "LiDAR 3D object detection. LiDAR-based 3D detection achieves excellent performance due to the precise 3D measurements of the input data. Two well-known ways to encode LiDAR points include voxel-based [10, 42, 45] and point-based [28,29,38]. Voxel-based methods divide the 3D space into regular voxels [45] or pillars [10], and encode the point inside into feature representations. Point-based methods are usually based on the PointNet [24] series to aggregate the feature of points. Then the point features will be used for the proposal refinement. LiDAR-based 3D detection performs well, but high-quality LiDAR is hard to be adopted in a large scale due to the expensive cost. Here we propose an economic camera-only solution by introducing multi-agent collaboration, whose performance can catch up with LiDAR given a sufficient number of agents." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.503, + 0.892, + 0.835 + ], + "angle": 0, + "content": "Collaborative perception. Collaborative perception is an emerging application of multi-agent systems, which promotes perception performance by enabling agents to share information with other agents through communication. In this emerging research field, a surge of high-quality datasets, (V2X-SIM [13], OPV2V [37], DAIR-V2X [40], CoPerception-UAVs [6]), and collaborative methods [5,6,13-15,17,18,20,31,35-37,40] aimed for better performance-bandwidth trade-off have been proposed. Collaborative methods have three types: early, late and intermediate fusion. Compared with early fusion(i.e. transmitting raw data) and late fusion (i.e. transmitting outputs), intermediate fusion(i.e. transmitting and fusing features encoded by deep networks) achieves a better performance-bandwidth trade-off. DiscoNet [14] adopts knowledge distillation to take the advantage of both early and intermediate collaboration. Where2comm [6] optimizes communication efficiency by sharing spatially sparse yet perceptually critical information. However, previous works mainly focus on LiDAR-based 3D object detection, while we investigate collaborative camera-only detection and specifically enhance depth estimation via multi-agent collaboration." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.847, + 0.871, + 0.865 + ], + "angle": 0, + "content": "3. Collaborative Camera-Only 3D Detection" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.871, + 0.892, + 0.902 + ], + "angle": 0, + "content": "This section presents CoCa3D, collaborative camera-only 3D detection framework, which enables multiple" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9244" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.088, + 0.48, + 0.179 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.484, + 0.089, + 0.882, + 0.179 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.084, + 0.181, + 0.877, + 0.269 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.274, + 0.895, + 0.318 + ], + "angle": 0, + "content": "Figure 2. System overview. CoCa3D is a camera-only 3D detector integrated with two collaboration modules. Collaborative depth estimation (Co-Depth) enhances the single-agent estimated depth to achieve more accurate 3D feature. Collaborative detection feature learning (Co-FL) complements the single-agent 3D feature to achieve more holistic 3D detection." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.325, + 0.471, + 0.356 + ], + "angle": 0, + "content": "agents to share visual information with each other, promoting more holistic 3D detection." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.364, + 0.281, + 0.378 + ], + "angle": 0, + "content": "3.1. Problem Formulation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.382, + 0.47, + 0.453 + ], + "angle": 0, + "content": "Consider \\(N\\) agents in the scene, let \\(\\mathcal{X}_i\\) be the RGB image collected by the \\(i\\)th agent and \\(\\mathcal{O}_i^0\\) be the corresponding ground-truth detection. The objective of CoCa3D is to maximize detection performances of all agents given certain communication budget \\(B\\); that is," + }, + { + "type": "equation", + "bbox": [ + 0.077, + 0.46, + 0.468, + 0.49 + ], + "angle": 0, + "content": "\\[\n\\max _ {\\theta , \\mathcal {P}} \\sum_ {i} g \\left(\\Phi_ {\\theta} \\left(\\mathcal {X} _ {i}, \\{\\mathcal {P} _ {i \\rightarrow j} \\} _ {j = 1} ^ {N}\\right), \\mathcal {O} _ {i} ^ {0}\\right), \\mathrm {s . t .} \\sum_ {i} | \\mathcal {P} _ {i \\rightarrow j} | \\leq B,\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.498, + 0.47, + 0.71 + ], + "angle": 0, + "content": "where \\( g(\\cdot, \\cdot) \\) is the detection evaluation metric, \\( \\Phi(\\cdot) \\) is a detection model with trainable parameter \\( \\theta \\), and \\( \\mathcal{P}_{i \\to j} \\) is the message transmitted from the \\( i \\)th agent to the \\( j \\)th agent. The key challenge is to determine the messages \\( \\mathcal{P}_{i \\to j} \\), which should be both informative and compact. Our design rational comes from two aspects: First, since the major gap between camera and LiDAR is the depth, the message should include depth information. This would allow different viewpoints from multiple agents to disambiguate the infinite depth possibilities and localize the correct depth candidate. Second, the message should include detection clues to provide complementary detection information, which can fundamentally overcome inevitable limitations of single-agent detection, such as occlusion and long-range issues." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.711, + 0.47, + 0.802 + ], + "angle": 0, + "content": "Based on the above intuition, the proposed CoCa3D includes a) single-agent camera-only 3D detection, which achieves basic depth estimation and detection ability, and b) multi-agent collaboration, which shares both estimated depth information and detection features to improve 3D representation and detection performances; see Fig. 2." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.806, + 0.411, + 0.822 + ], + "angle": 0, + "content": "3.2. Single-agent camera-only 3D detection" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.826, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Single-agent camera-only 3D object detection network learns to detect 3D objects in the physical space based on the 2D camera inputs. Here we adopt CaDNN [25], an off-the-shelf architecture. The main idea is to project the flat image feature into all the possible depths in the 3D space," + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.325, + 0.892, + 0.4 + ], + "angle": 0, + "content": "and then attentively aggregate the 3D voxel feature and collapse them into the BEV feature. To mention that, the BEV feature is used because it largely reduces the computation cost, while performing similarly to 3D voxel features [10]. The architecture consists of five modules as detailed below." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.403, + 0.892, + 0.464 + ], + "angle": 0, + "content": "Encoder. The encoder extracts features from the input images. For agent \\(i\\), given its input \\(\\mathcal{X}_i\\), the output feature map of encoder is \\(\\mathcal{F}_i = \\Phi_{\\mathrm{enc}}(\\mathcal{X}_i) \\in \\mathbb{R}^{H \\times W \\times C}\\), where \\(H, W, C\\) are the height, width and channel of the image feature." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.467, + 0.892, + 0.602 + ], + "angle": 0, + "content": "Depth estimation. The depth estimation module predicts pixel-wise categorical depth distribution over a set of predefined depth bins to accurately locate image information in 3D space. Treating the estimation as a classification problem can capture the inherent depth estimation uncertainty to reduce the impact of erroneous depth estimates, as stated in CaDNN [25]. To achieve this, a parametric depth estimation network \\(\\Phi_{\\mathrm{Depth}}(\\cdot)\\) is used, and the pixel-wise categorical depth distribution is obtained as" + }, + { + "type": "equation", + "bbox": [ + 0.587, + 0.603, + 0.891, + 0.621 + ], + "angle": 0, + "content": "\\[\n\\mathcal {D} _ {i} = \\Phi_ {\\text {D e p t h}} \\left(\\mathcal {F} _ {i}\\right) \\in \\mathbb {R} ^ {H \\times W \\times D}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.622, + 0.894, + 0.669 + ], + "angle": 0, + "content": "whose \\((h,w)\\) th element \\(\\mathcal{D}_i(h,w)\\in \\mathbb{R}^D\\) refers to the estimated depth distribution of the \\((h,w)\\) th image pixel with \\(D\\) as the number of discretized depth bins." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.671, + 0.892, + 0.822 + ], + "angle": 0, + "content": "Voxel transformation. Voxel transformation projects the image feature to the 3D space, given all the possible depth candidates and the known image calibration matrix, resulting in a 3D voxel representation \\(\\widehat{\\mathcal{V}}_i\\in \\mathbb{R}^{X\\times Y\\times Z\\times C}\\). The camera projection matrix \\(\\mathbf{P}\\in \\mathbb{R}^{3\\times 4}\\) defines the mapping between the 3D voxel coordinate to the image pixel coordinate, this is, \\([u,v,d] = \\mathbf{P}\\cdot [x,y,z,1]\\). Similarly, the depth probabilities can be transformed to the 3D voxel space, resulting in \\(\\widehat{\\mathcal{D}}_i\\in \\mathbb{R}^{X\\times Y\\times Z}\\), where each element indicates the confidence that the feature pixel belongs to the voxel." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.826, + 0.894, + 0.903 + ], + "angle": 0, + "content": "Collapse. The voxel feature is collapsed to a single height plane to generate the bird's-eye-view (BEV) feature as \\(\\overline{B}_i = \\Phi_{\\mathrm{clp}}(\\widehat{\\mathcal{D}}_i \\odot \\widehat{\\mathcal{V}}_i) \\in \\mathbb{R}^{X \\times Y \\times C}\\), where \\(\\odot\\) is the element-wise multiplication. Symbol \\(\\cdot / \\widehat{\\cdot} / \\widehat{\\cdot}\\) reflects 2D image / 3D voxel / BEV spaces, respectively. The collapse network \\(\\Phi_{\\mathrm{clp}}(\\cdot)\\)" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9245" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.468, + 0.12 + ], + "angle": 0, + "content": "flattens the 3D voxel feature along the \\(Z\\)-axis and then applies \\(1 \\times 1\\) convolution to reduce the channel dimension." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.122, + 0.468, + 0.273 + ], + "angle": 0, + "content": "Decoder. The decoder takes the BEV feature as input, and outputs the objects including class and regression. Here, we adopt CenterNet [44], an off-the-shelf detector. Given the BEV feature \\(\\overline{B}_i\\), the detection decoder \\(\\Phi_{\\mathrm{dec}}(\\cdot)\\) generate the dense heatmap of \\(i\\)th agent by \\(\\overline{\\mathcal{O}}_i = \\Phi_{\\mathrm{dec}}(\\overline{B}_i) \\in \\mathbb{R}^{X \\times Y \\times 7}\\), where each location of \\(\\overline{\\mathcal{O}}_i\\) represents a rotated box with class \\((c, x, y, h, w, \\cos \\alpha, \\sin \\alpha)\\), denoting class confidence, position residual, size and angle. Non-maximum suppression (NMS) is applied to the dense predictions and generate the final sparse output of the 3D detection system." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.276, + 0.468, + 0.426 + ], + "angle": 0, + "content": "Discussion. Admittedly, single-agent camera-only 3D detection makes localizing objects in the 3D space given 2D image input possible through numerous ingenious algorithm designs. The 3D detection performance is still far from perfect due to the ambiguous single-agent depth estimation and limited visual range. Tackling these two issues are challenging for a single agent, but from a perspective of multi-agent collaboration, sharing complementary visual clues makes depth estimation and visual range enlargement much more natural. This motivates our following designs." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.432, + 0.312, + 0.448 + ], + "angle": 0, + "content": "3.3. Multi-agent collaboration" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.452, + 0.468, + 0.556 + ], + "angle": 0, + "content": "Multi-agent collaboration includes two parts: i) collaborative depth estimation, which enables the sharing of depth information estimated by individual agents to produce more accurate 3D representation and less biased BEV feature; and ii) collaborative detection feature learning, which enables the sharing of detection features of individual agents to achieve more holistic 3D detection." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.562, + 0.347, + 0.576 + ], + "angle": 0, + "content": "3.3.1 Collaborative depth estimation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.581, + 0.468, + 0.793 + ], + "angle": 0, + "content": "Collaborative depth estimation (Co-Depth) targets to disambiguate infinite depth possibilities in single-agent camera-only depth estimation and localize the correct depth candidate through multi-view consistency. The intuition is, for a correct depth candidate, its corresponding 3D location should be spatially consistent from multiple agents' viewpoints. To achieve this, each agent can exchange the depth information through communication. Meanwhile, we promote communication efficiency by selecting the most critical and unambiguous depth information. Accordingly, Co-Depth includes: a) depth uncertainty aware message packing, which packs compact messages with unambiguous depth information; and b) depth message fusion, which enhances depth estimation with the received depth messages." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.796, + 0.468, + 0.901 + ], + "angle": 0, + "content": "Depth uncertainty aware message packing. Depth uncertainty-aware (DUA) message packing packs the most critical depth information used for multi-view consistency into the to-be-sent message based on the depth uncertainty. The depth message includes: i) the voxel feature, which is used for the multi-view visual similarity measurement; and ii) the depth probability, which indicates the confidence that" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.891, + 0.12 + ], + "angle": 0, + "content": "the feature pixel belongs to the voxel and is used for the multi-view candidate selection." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.122, + 0.892, + 0.182 + ], + "angle": 0, + "content": "The spatially sparse features of \\(i\\)th agent are selected based on its spatial depth uncertainty map in the 2D image space \\(\\mathbf{U}_i\\in \\mathbb{R}^{H\\times W}\\), that can be represented by the pixelwise entropy from corresponding depth distribution," + }, + { + "type": "equation", + "bbox": [ + 0.591, + 0.186, + 0.8, + 0.201 + ], + "angle": 0, + "content": "\\[\n\\mathbf {U} _ {i} (h, w) = \\mathrm {H} (\\mathcal {D} _ {i} (h, w)) \\in \\mathbb {R},\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.205, + 0.892, + 0.355 + ], + "angle": 0, + "content": "where \\(\\mathrm{H}(\\cdot)\\) is the entropy function and \\(\\mathcal{D}_i(h,w)\\) is the depth distribution at the \\((h,w)\\)th image pixel, following from (1). The intuition is, for the locations with low uncertainty score, an agent is confident about which depth bin the feature pixel belongs to. Thus, sharing information at these locations can help with depth estimation accuracy. On the contrary, for the locations with high uncertainty score, an agent is hard to tell which depth bin the feature pixel belongs to. Transmitting these spatial features would introduce ambiguity, and even degrade the single-agent depth estimation." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.356, + 0.892, + 0.432 + ], + "angle": 0, + "content": "Specifically, a binary selection matrix is used to represent whether each location is selected or not, where 1 denotes selected, and 0 elsewhere. Let \\(\\mathbf{M}_{i\\rightarrow j}^{d}\\in \\{0,1\\}^{H\\times W}\\) be a selection matrix to determine the depth message sent from agent \\(i\\) to \\(j\\). Its \\((h,w)\\)th element is" + }, + { + "type": "equation", + "bbox": [ + 0.57, + 0.434, + 0.821, + 0.452 + ], + "angle": 0, + "content": "\\[\n\\mathbf {M} _ {i \\rightarrow j} ^ {d} (h, w) = I (\\mathbf {U} _ {i} (h, w) < u _ {\\text {t h r e}}),\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.454, + 0.892, + 0.498 + ], + "angle": 0, + "content": "where \\(I(\\cdot)\\) is an indicator function and \\(u_{\\mathrm{thre}}\\) is a predefined threshold. This reflects that a specific spatial area will be selected when its depth uncertainty is below a threshold." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.499, + 0.892, + 0.567 + ], + "angle": 0, + "content": "The selection matrix can be projected in the 3D voxel space, resulting in \\(\\widehat{\\mathbf{M}}_{i\\rightarrow j}^{d}\\in \\mathbb{R}^{X\\times Y\\times Z}\\). Then, the selected voxel feature map \\(\\widehat{\\mathcal{Z}}_{i\\rightarrow j}\\) and the selected depth distribution \\(\\widehat{\\mathcal{E}}_{i\\rightarrow j}\\) sent from the \\(i\\)th agent to the \\(j\\)th agent are obtained as" + }, + { + "type": "equation", + "bbox": [ + 0.571, + 0.569, + 0.82, + 0.612 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\widehat {\\mathcal {Z}} _ {i \\rightarrow j} = \\widehat {\\mathbf {M}} _ {i \\rightarrow j} ^ {d} \\odot \\widehat {\\mathcal {V}} _ {i} \\in \\mathbb {R} ^ {X \\times Y \\times Z \\times C}, \\\\ \\widehat {\\mathcal {E}} _ {i \\rightarrow j} = \\widehat {\\mathbf {M}} _ {i \\rightarrow j} ^ {d} \\odot \\widehat {\\mathcal {D}} _ {i} \\in \\mathbb {R} ^ {X \\times Y \\times Z}, \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.616, + 0.892, + 0.678 + ], + "angle": 0, + "content": "where \\(\\widehat{\\mathcal{V}}_i\\) and \\(\\widehat{\\mathcal{D}}_i\\) are the full 3D voxel feature map and depth distribution. Overall, the depth message sent from the \\(i\\)th agent to the \\(j\\)th agent is \\(\\mathcal{P}_{i\\rightarrow j}^{d} = (\\widehat{\\mathcal{E}}_{i\\rightarrow j},\\widehat{\\mathcal{Z}}_{i\\rightarrow j})\\), where both entries are spatially sparse and communication efficient." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.679, + 0.892, + 0.821 + ], + "angle": 0, + "content": "Depth message fusion. Depth message fusion targets to enhance the depth estimation given the received depth messages through different viewpoints of multiple agents. The intuition is that for a correct depth candidate, the visual features at the same 3D point observed by multiple agents should be similar. To achieve this, we introduce multiview depth consistency weighting. Let \\(\\widehat{S}_i(x,y,z) \\in \\mathbb{R}\\) be the matching score between agent \\(i\\) and its neighbors \\(\\mathcal{N}_i\\), those who share messages with current agent, at coordinate \\((x,y,z)\\). The multi-view matching score is obtained as" + }, + { + "type": "equation", + "bbox": [ + 0.501, + 0.825, + 0.822, + 0.843 + ], + "angle": 0, + "content": "\\[\n\\widehat {\\mathcal {W}} _ {j \\rightarrow i} (x, y, z) = I (\\widehat {\\mathcal {E}} _ {j \\rightarrow i} (x, y, z) > p _ {\\text {t h r e}}) \\in \\{0, 1 \\},\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.502, + 0.846, + 0.889, + 0.877 + ], + "angle": 0, + "content": "\\[\n\\widehat {\\mathcal {S}} _ {i} (x, y, z) = \\sum_ {j \\in \\mathcal {N} _ {i}} \\widehat {\\mathcal {W}} _ {j \\rightarrow i} (x, y, z) \\left\\langle \\widehat {\\mathcal {V}} _ {i} (x, y, z), \\widehat {\\mathcal {Z}} _ {j \\rightarrow i} (x, y, z) \\right\\rangle ,\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.503, + 0.883, + 0.891, + 0.902 + ], + "angle": 0, + "content": "where \\(\\widehat{\\mathcal{W}}_{j\\rightarrow i}(x,y,z)\\) is a binary weight used to filter out the" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "9246" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.182 + ], + "angle": 0, + "content": "neighboring view that has a low depth confidence, below a threshold \\( p_{\\mathrm{thre}} \\) and \\( \\langle \\cdot ,\\cdot \\rangle \\) is the inner product that measures the visual feature consistency across two views. This means that either the depth candidate is wrong or it is not visible in that view (e.g. due to occlusion), where the multi-view matching score should not be considered." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.183, + 0.47, + 0.213 + ], + "angle": 0, + "content": "Given the multi-view matching score, the collaborative depth distribution is obtained as" + }, + { + "type": "equation", + "bbox": [ + 0.135, + 0.213, + 0.412, + 0.239 + ], + "angle": 0, + "content": "\\[\n\\widehat {\\mathcal {D}} _ {i} ^ {\\prime} = \\Phi_ {\\mathrm {C o D e p t h}} \\left(\\left[ \\widehat {\\mathcal {D}} _ {i}; \\widehat {\\mathcal {S}} _ {i} \\right]\\right) \\in \\mathbb {R} ^ {X \\times Y \\times Z},\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.238, + 0.469, + 0.388 + ], + "angle": 0, + "content": "where the \\([\\cdot ]\\) denotes concatenation, \\(\\Phi_{\\mathrm{CoDepth}}(\\cdot)\\) is the collaborative depth estimation network implemented with \\(1\\times 1\\) convolutions followed by a sigmoid activation. Each element of the collaborative depth distribution \\(\\widehat{D}_i^{\\prime}\\) reflects the confidence of the feature locates in the corresponding voxel. Note that i) the collaboratively updated depth probability is decided on both the single-view confidence and the multiview consistency; ii) the depth consistency weighting discards the candidates with low single-view depth probability to avoid ambiguity and make the estimation more robust." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.389, + 0.469, + 0.466 + ], + "angle": 0, + "content": "Given this collaborative depth distribution, each agent can weigh the 3D voxel feature and then collapse into a BEV feature with a collapse network \\(\\Phi_{\\mathrm{clp}}(\\cdot)\\), this is, \\(\\overline{B}_i = \\Phi_{\\mathrm{clp}}(\\widehat{\\mathcal{D}}_i' \\odot \\widehat{\\mathcal{V}}_i) \\in \\mathbb{R}^{X \\times Y \\times C}\\). The collapse network is implemented with \\(1 \\times 1\\) convolution to reduce channel dimension." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.472, + 0.41, + 0.487 + ], + "angle": 0, + "content": "3.3.2 Collaborative detection feature learning" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.49, + 0.469, + 0.716 + ], + "angle": 0, + "content": "Collaborative depth estimation carefully refines the depth and promotes more accurate 3D representations for each single agent. However, the physical limitations of single agent, such as, limited field of view, occlusion, and long-range issues still remain. To achieve more holistic 3D detection, each agent should be able to exchange 3D detection feature and make use of the complementary information. Meanwhile, we promote communication efficiency by selecting the most perceptually critical information. Therefore, collaborative detection feature learning (Co-FL) includes: a) detection confidence aware message packing, which packs spatially sparse yet perceptually critical 3D features with the guidance of the detection confidence; and ii) detection message fusion, which enhances the 3D feature with the received detection messages." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.719, + 0.469, + 0.884 + ], + "angle": 0, + "content": "Detection confidence aware message packing. The detection confidence aware (DCA) message packing targets to pack the complementary perceptual information into a compact message. The core idea is to explore the spatial heterogeneity of perceptual information. The intuition is that the areas that contain objects are more critical than background areas. During collaboration, areas with objects could help recover the miss-detected objects due to the limited view; and background areas could be omitted to save the precious bandwidth. To achieve this, we implement the spatial confidence map with the detection confidence map by" + }, + { + "type": "equation", + "bbox": [ + 0.186, + 0.885, + 0.359, + 0.902 + ], + "angle": 0, + "content": "\\[\n\\overline {{\\mathbf {C}}} _ {i} = \\Phi_ {\\mathrm {c l s}} (\\overline {{\\mathcal {B}}} _ {i}) \\in \\mathbb {R} ^ {X \\times Y},\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.182 + ], + "angle": 0, + "content": "where \\(\\Phi_{\\mathrm{cls}}(\\cdot)\\) denotes the standard classification network in the single-agent decoder \\(\\Phi_{\\mathrm{det}}(\\cdot), \\overline{B}_i\\) is the BEV feature generated with the collaboratively estimated depth. Based on this map, agents decide which spatial area to communicate; that is, each agent offers spatially sparse, yet critical features to support other agents." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.183, + 0.892, + 0.247 + ], + "angle": 0, + "content": "Specifically, a binary selection matrix is used to represent each location is selected or not. Let \\(\\overline{\\mathbf{M}}_{i\\rightarrow j}^{f}\\in\\) \\(\\{0,1\\}^{X\\times Y}\\) be a selection matrix to determine the detection message sent from agent \\(i\\) to \\(j\\). Its \\((x,y)\\)th element is" + }, + { + "type": "equation", + "bbox": [ + 0.574, + 0.248, + 0.816, + 0.269 + ], + "angle": 0, + "content": "\\[\n\\overline {{\\mathbf {M}}} _ {i \\rightarrow j} ^ {f} (x, y) = I (\\overline {{\\mathbf {C}}} _ {i} (x, y) > c _ {\\mathrm {t h r e}}),\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.269, + 0.892, + 0.378 + ], + "angle": 0, + "content": "where \\(I(\\cdot)\\) is an indicator function and \\(c_{\\mathrm{thre}}\\) is a predefined threshold. This reflects that a specific spatial area will be selected when its detection confidence is above a threshold. Overall, the detection message sent from the \\(i\\)th agent to the \\(j\\)th agent is \\(\\mathcal{P}_{i\\rightarrow j}^{f} = \\overline{\\mathbf{M}}_{i\\rightarrow j}^{f}\\odot \\overline{\\mathcal{B}_{i}}\\). Only selected features and their indices are packed, and they can be recovered at the receiver. This greatly reduces communication cost." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.38, + 0.892, + 0.519 + ], + "angle": 0, + "content": "Detection message fusion. Here we augment the detection feature of each agent by aggregating the received detection messages from other agents. We implement this with simple but effective non-parametric point-wise maximum fusion. Specifically, for the \\(i\\)th agent, after receiving the \\(j\\)th agent's message \\(\\mathcal{P}_{j\\rightarrow i}^{f}\\). We also include the ego feature map in fusion and denote \\(\\mathcal{P}_{i\\rightarrow i}^{f} = \\overline{\\mathcal{B}}_{i}\\) to make the formulation simple and consistent, where \\(\\mathcal{P}_{i\\rightarrow i}^{f}\\) might not be sparse. The fused BEV feature is obtained as" + }, + { + "type": "equation", + "bbox": [ + 0.576, + 0.52, + 0.815, + 0.546 + ], + "angle": 0, + "content": "\\[\n\\overline {{\\mathcal {B}}} _ {i} ^ {\\prime} = \\max _ {j \\in \\mathcal {N} _ {i} \\cup \\{i \\}} (\\mathcal {P} _ {j \\rightarrow i} ^ {f}) \\in \\mathbb {R} ^ {X \\times Y \\times C},\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.547, + 0.892, + 0.652 + ], + "angle": 0, + "content": "where \\(\\max (\\cdot)\\) maximizes the corresponding features from multiple agents at each individual spatial location. Note that attention fusion is not permutation invariant, as attention weights vary with the ordering of key and query. Here we simply use the max operator to avoid this permutation variant issue. The fused BEV feature is output to the decoder to generate the final detection." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.656, + 0.587, + 0.669 + ], + "angle": 0, + "content": "3.4. Losses" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.671, + 0.892, + 0.763 + ], + "angle": 0, + "content": "To train the overall system, we supervise two tasks: categorical depth estimation and 3D object detection. The overall loss is \\( L = L_{\\mathrm{dep}}(\\mathcal{D}_i,\\mathcal{D}_i^0) + L_{\\mathrm{det}}(\\overline{\\mathcal{O}}_i,\\mathcal{O}_i^0) \\), where \\( \\mathcal{D}_i^0 \\) and \\( \\mathcal{O}_i^0 \\) is the \\( i \\)th agent's ground-truth depth category and objects, respectively, \\( L_{\\mathrm{dep}} \\) and \\( L_{\\mathrm{det}} \\) is the depth classification loss [25] and detection loss [44], respectively." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.769, + 0.706, + 0.786 + ], + "angle": 0, + "content": "4. Experimental Results" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.791, + 0.803, + 0.807 + ], + "angle": 0, + "content": "4.1. Datasets and experimental settings" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.811, + 0.892, + 0.901 + ], + "angle": 0, + "content": "We cover three datasets, both real-world and simulation scenarios, and multiple agent types. Metric Average Precision (AP) at Intersection-over-Union (IoU) threshold of 0.30, 0.50, 0.70, and 0.80 are used. Metric multi-class classification accuracy is used for depth accuracy. The communication volume follows the standard setting as [6, 36, 37]" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9247" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.088, + 0.09, + 0.316, + 0.227 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.185, + 0.229, + 0.255, + 0.241 + ], + "angle": 0, + "content": "(a) OPV2V+" + }, + { + "type": "image", + "bbox": [ + 0.352, + 0.091, + 0.578, + 0.228 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.442, + 0.229, + 0.525, + 0.241 + ], + "angle": 0, + "content": "(b) DAIR-V2X" + }, + { + "type": "image", + "bbox": [ + 0.614, + 0.092, + 0.837, + 0.228 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.679, + 0.229, + 0.813, + 0.242 + ], + "angle": 0, + "content": "(c) CoPerception-UAVs+" + }, + { + "type": "image_caption", + "bbox": [ + 0.075, + 0.243, + 0.89, + 0.271 + ], + "angle": 0, + "content": "Figure 3. CoCa3D steadily improves 3D detection performance as the number of agents grows. In OPV2V+, around 10 collaborative agents enable collaborative cameras to catch up with LiDAR." + }, + { + "type": "table_caption", + "bbox": [ + 0.075, + 0.271, + 0.89, + 0.299 + ], + "angle": 0, + "content": "Table 1. CoCa3D significantly outperforms the previous SOTAs, e.g. improve by \\( {30.60}\\% \\) on OPV2V, \\( {12.59}\\% \\) on CoPerception-UAVs, 44.21% on DAIR-V2X for metric AP@70." + }, + { + "type": "table", + "bbox": [ + 0.079, + 0.3, + 0.908, + 0.455 + ], + "angle": 0, + "content": "
MethodOPV2V+CoPerception-UAVs+DAIR-V2X
AP@30AP@50AP@70AP@50AP@70AP@80AP@30AP@50AP@70
No Collaboration0.27480.20410.08530.69560.49000.23090.09770.05240.0305
Late Fusion0.65010.61980.51090.72060.53720.25970.20600.10780.0455
When2com (CVPR'20)0.48530.42110.37370.82190.67050.41020.19570.09840.0459
V2VNet (ECCV'20)0.62460.50420.38520.90930.71770.38040.16400.08470.0512
DiscoNet (NeurIPS'21)0.73000.60090.41790.90540.70790.35640.18360.12620.0683
V2X-ViT (ECCV'22)0.83460.66590.39460.90940.71430.35250.18620.10750.0490
Where2comm (NeurIPS'22)0.81910.70890.47410.91020.73830.36760.17540.10250.0547
CoCa3D0.86420.82600.66750.94970.85020.58350.35220.22600.0985
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.46, + 0.47, + 0.506 + ], + "angle": 0, + "content": "that counts the message size by byte in log scale with base 2. To compare communication results straightforwardly and fair, we do not consider any extra compression." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.523, + 0.47, + 0.66 + ], + "angle": 0, + "content": "OPV2V+. The original OPV2V [37] is a large-scale vehicle-to-vehicle collaborative perception dataset, cosimulated by OpenCDA [34] and CARLA [4]. Here we introduce an extended version, OPV2V+, which includes more collaborative agents (\\(\\approx 10\\)). Each agent is equipt with 4 cameras and 4 depth sensors with resolution \\(600 \\times 800\\). The detection range is \\(281.6\\mathrm{m} \\times 80\\mathrm{m}\\). Our single-agent camera-only detector follows CaDDN [25] with 50 depth categories in linear-increasing spacing mode." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.677, + 0.47, + 0.753 + ], + "angle": 0, + "content": "DAIR-V2X. DAIR-V2X [40] is the only public real-world collaborative perception dataset. It contains two agents: a vehicle and a road-side-unit with image resolution \\(1080 \\times 1920\\). The perception range is \\(100\\mathrm{m} \\times 79\\mathrm{m}\\). Our single-agent camera-only detector also follows CaDDN [25]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.769, + 0.47, + 0.906 + ], + "angle": 0, + "content": "CoPerception-UAVs+. The original CoPerception-UAVs [6] is a large-scale UAV-based collaborative perception dataset, co-simulated by AirSim [27] and CARLA [4]. Here we introduce an extended version, CoPerception-UAVs+, which includes more collaborative agents (\\(\\approx 10\\)). Each agent equips with 1 camera and 1 depth sensor with resolution \\(450 \\times 800\\). The detection range is \\(192\\mathrm{m} \\times 96\\mathrm{m}\\). Our single-agent camera-only detector follows DVDET [?] with 10 depth categories in uniform spacing mode." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.46, + 0.719, + 0.475 + ], + "angle": 0, + "content": "4.2. Quantitative evaluation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.486, + 0.892, + 0.819 + ], + "angle": 0, + "content": "Collaborative camera-only 3D detection overtakes Li-DAR. Fig. 3 investigates how the collaborative camera-only 3D detection performance changes with the number of collaborative agents on multiple datasets. To set up a goal, we consider a LiDAR-based detector implemented by the widely used PointPillar [10]. Fig. 3 (a) shows that: i) with 10 collaborative agents, CoCa3D outperforms the LiDAR 3D detection at both AP@0.5/0.7 on OPV2V+! ii) the detection performance positively increases as the number of collaborative agents. And we then use the slope of OPV2V+ to fit the function of collaborative detection performance with agent number on DAIR-V2X as it only has two agents available. Fig. 3 (b) shows that camera-only 3D detection is expected to outperform LiDAR 3D detection with 7 collaborative vehicles in the real scenario. As LiDAR is not a commonly available equipment for drones, we set a high-quality RGB-D camera instead. Fig. 3 (c) shows that with 3 collaborative drones, camera-only 3D detection outperforms ground-truth depth-based 3D detection. Furthermore, the steadily increasing detection performance with the number of collaborative agents encourages the agents to actively collaborate and achieve consistent improvement." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.826, + 0.893, + 0.901 + ], + "angle": 0, + "content": "Benchmark comparison. Tab. 1 compares the proposed CoCa3D with previous collaborative methods. We consider single-agent detection without collaboration (No collaboration), When2com [17], V2VNet [31], DiscoNet [14], V2XViT [36], Where2comm [6] and late fusion, where agents" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "9248" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.09, + 0.47, + 0.117 + ], + "angle": 0, + "content": "Table 2. CoCa3D outperforms single-agent camera-only 3D detection with GT depth on OPV2V+ and CoPerception-UAVs+." + }, + { + "type": "table", + "bbox": [ + 0.078, + 0.121, + 0.48, + 0.224 + ], + "angle": 0, + "content": "
Co -DepthCo -FLOPV2V+CoPerception-UAVs+
AP@30AP@50AP@70AP@50AP@70AP@80
--0.27480.20410.08530.72130.54210.2846
GT-0.34540.25530.09730.83470.67640.4120
-✓0.82010.71910.47560.90840.72560.4028
GT✓0.91200.88050.74340.95050.83980.5504
✓✓0.86420.82600.66750.94950.85180.5849
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.241, + 0.47, + 0.392 + ], + "angle": 0, + "content": "directly exchange the detected 3D boxes. We see that the proposed CoCa3D significantly outperforms previous state-of-the-arts, improves the SOTA performance by \\(30.60\\%\\) on OPV2V+, \\(12.59\\%\\) on CoPerception-UAVs+, \\(44.21\\%\\) on DAIR-V2X for AP@70. The reason is that collaborative depth estimation promotes more accurate 3D features per single agent, and the enhanced single agent features further facilitate collaborative detection feature learning, while previous collaborative methods do not specifically consider depth ambiguity and collaborate over single-agent features." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.394, + 0.47, + 0.545 + ], + "angle": 0, + "content": "Multi-agent collaboration evaluation. Tab. 2 assess the effectiveness of the proposed multi-agent collaboration on OPV2V+ and CoPerception-UAVs+ datasets. Ground-truth (GT) depth is included to provide the upper bound of collaborative camera-only 3D detection. We see that: i) both collaborations can consistently improve the performance; ii) camera-only 3D detection with collaboratively estimated depth significantly outperforms that with single-agent estimated depth and even approaches the upper bound with ground truth depth." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.558, + 0.283, + 0.573 + ], + "angle": 0, + "content": "4.3. Qualitative evaluation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.583, + 0.47, + 0.793 + ], + "angle": 0, + "content": "Visualization of depth and uncertainty. Fig. 5 shows that collaborative depth estimation outperforms single-agent depth estimation and approaches the ground truth depth. We see that: i) the single-agent depth estimation can estimate the relative depth while not able to precisely localize the depth candidate, for example, the vehicles are higher than the plane it sits while this plane is not grounded to the right category; ii) by introducing multi-view geometry, the collaboratively estimated depth can smoothly and accurately ground the plane; iii) the depth uncertainty is larger for the long-range and background areas. The reason is that the distant regions are hard to localize as they occupy too few image pixels, and the background regions are hard to localize due to the texture-less surfaces." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.796, + 0.47, + 0.902 + ], + "angle": 0, + "content": "Visualization of detection results. Fig. 6 shows that compared to single LiDAR 3D detection, cameras with collaborations (Co-Depth and Co-FL) are able to achieve holistic and accurate detection results. The reason is that the single agent has some fundamental physical limitations such as long-range and occlusion issues. Fig. 6(a-c) shows that augmenting a single agent sensor going from just a cam" + }, + { + "type": "image", + "bbox": [ + 0.501, + 0.089, + 0.695, + 0.228 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.564, + 0.231, + 0.634, + 0.243 + ], + "angle": 0, + "content": "(a) OPV2V+" + }, + { + "type": "image", + "bbox": [ + 0.696, + 0.09, + 0.886, + 0.228 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.725, + 0.231, + 0.859, + 0.243 + ], + "angle": 0, + "content": "(b) CoPerception-UAVs+" + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.245, + 0.892, + 0.272 + ], + "angle": 0, + "content": "Figure 4. CoCa3D achieves superior detection performance and communication cost trade-off over various bandwidths." + }, + { + "type": "table_caption", + "bbox": [ + 0.524, + 0.279, + 0.867, + 0.294 + ], + "angle": 0, + "content": "Table 3. Co-Depth significantly improves depth accuracy." + }, + { + "type": "table", + "bbox": [ + 0.501, + 0.297, + 0.892, + 0.343 + ], + "angle": 0, + "content": "
Full planeForeground
SingleCollaborationGainSingleCollaborationGain
0.61670.778126.17% ↑0.75860.81337.21% ↑
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.353, + 0.892, + 0.408 + ], + "angle": 0, + "content": "Table 4. Ablation of collaborative depth estimation on CoPerception-UAVs+. Uniform/Linear denotes uniform/linear-increasing spacing. Dense and sparse supervision denote applying depth supervision over the full image plane and the object regions." + }, + { + "type": "table", + "bbox": [ + 0.53, + 0.412, + 0.86, + 0.493 + ], + "angle": 0, + "content": "
SpacingSupervisionAP@50AP@70AP@80
UniformDense0.95020.84440.5746
UniformSparse0.93300.80420.4829
LinearDense0.94840.83650.5580
LinearSparse0.94950.85180.5849
" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.509, + 0.892, + 0.645 + ], + "angle": 0, + "content": "era to including a depth sensor to LiDAR can improve the detection but still not able to achieve holistic detection. Fig. 6(d)/(b) and (a)/(e) compares the detection results with and without collaborative feature learning, we see that collaboration can help detect lots of the missed objects in the single-agent detection. Fig. 6(d,e,f) show that: i) both collaborations, Co-FL and Co-Depth, consistently improve the detection performance; ii) with collaboration, collaborative camera-only agents can outperform LiDAR in 3D detection." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.654, + 0.661, + 0.669 + ], + "angle": 0, + "content": "4.4. Ablation studies" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.675, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Trade-off between detection performance and communication cost. Fig. 4 compares the proposed CoCa3D with the previous communication-efficient solution Where2comm [6] in terms of the trade-off between detection performance (AP@IoU=50/70) and communication bandwidth. We see that: i) CoCa3D can adapt to varying bandwidths by adjusting the depth uncertainty threshold and the detection confidence threshold; ii) CoCa3D achieves superior detection performance and communication cost trade-off to Where2comm over varying communication bandwidth. The gain mainly comes from Co-Depth's improvement in depth estimation, which improves single-agent BEV features. With enhanced single BEV features, the Co-FL can generate better-augmented BEV features, resulting in improved detection performance." + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "9249" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.084, + 0.112, + 0.233, + 0.171 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.117, + 0.189, + 0.199, + 0.202 + ], + "angle": 0, + "content": "(a) Image (RV)" + }, + { + "type": "image", + "bbox": [ + 0.246, + 0.099, + 0.385, + 0.184 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.28, + 0.189, + 0.364, + 0.202 + ], + "angle": 0, + "content": "(b) Single (RV)" + }, + { + "type": "image", + "bbox": [ + 0.41, + 0.099, + 0.55, + 0.184 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.425, + 0.189, + 0.546, + 0.201 + ], + "angle": 0, + "content": "(c)Collaboration (RV)" + }, + { + "type": "image", + "bbox": [ + 0.574, + 0.099, + 0.714, + 0.184 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.591, + 0.189, + 0.708, + 0.201 + ], + "angle": 0, + "content": "(d) Ground-truth (RV)" + }, + { + "type": "image", + "bbox": [ + 0.738, + 0.099, + 0.882, + 0.185 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.758, + 0.189, + 0.869, + 0.201 + ], + "angle": 0, + "content": "(e) Uncertainty (RV)" + }, + { + "type": "image", + "bbox": [ + 0.084, + 0.218, + 0.233, + 0.278 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.113, + 0.296, + 0.203, + 0.308 + ], + "angle": 0, + "content": "(f) Image (BEV)" + }, + { + "type": "image", + "bbox": [ + 0.246, + 0.205, + 0.385, + 0.289 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.275, + 0.296, + 0.367, + 0.308 + ], + "angle": 0, + "content": "(g) Single (BEV)" + }, + { + "type": "image", + "bbox": [ + 0.409, + 0.206, + 0.549, + 0.289 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.418, + 0.296, + 0.548, + 0.308 + ], + "angle": 0, + "content": "(h) Collaboration (BEV)" + }, + { + "type": "image", + "bbox": [ + 0.571, + 0.206, + 0.711, + 0.289 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.584, + 0.296, + 0.708, + 0.308 + ], + "angle": 0, + "content": "(i) Ground-truth (BEV)" + }, + { + "type": "image", + "bbox": [ + 0.734, + 0.206, + 0.88, + 0.29 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.75, + 0.296, + 0.868, + 0.308 + ], + "angle": 0, + "content": "(j) Uncertainty (BEV)" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.313, + 0.89, + 0.341 + ], + "angle": 0, + "content": "Figure 5. Visualization of the depth and uncertainty in image range-view (RV) and bird's eye view (BEV) on CoPerception-UAVs+. Collaboratively estimated depth improves the single-agent estimated depth and approaches ground-truth depth." + }, + { + "type": "image", + "bbox": [ + 0.079, + 0.344, + 0.342, + 0.426 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.18, + 0.429, + 0.241, + 0.44 + ], + "angle": 0, + "content": "(a) Camera" + }, + { + "type": "image", + "bbox": [ + 0.353, + 0.344, + 0.616, + 0.426 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.455, + 0.429, + 0.517, + 0.44 + ], + "angle": 0, + "content": "(b) RGB-D" + }, + { + "type": "image", + "bbox": [ + 0.629, + 0.344, + 0.892, + 0.426 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.732, + 0.429, + 0.79, + 0.44 + ], + "angle": 0, + "content": "(c) LiDAR" + }, + { + "type": "image", + "bbox": [ + 0.079, + 0.442, + 0.341, + 0.524 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.148, + 0.526, + 0.274, + 0.537 + ], + "angle": 0, + "content": "(d) Camera with Co-FL" + }, + { + "type": "image", + "bbox": [ + 0.352, + 0.442, + 0.614, + 0.524 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.421, + 0.526, + 0.546, + 0.538 + ], + "angle": 0, + "content": "(e) RGB-D with Co-FL" + }, + { + "type": "image", + "bbox": [ + 0.625, + 0.442, + 0.888, + 0.524 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.626, + 0.526, + 0.886, + 0.539 + ], + "angle": 0, + "content": "(f) Camera with Co-Depth and Co-FL (CoCa3D)" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.543, + 0.89, + 0.558 + ], + "angle": 0, + "content": "Figure 6. CoCa3D outperforms LiDAR detection on OPV2V+ with 10 agents. Green and red boxes denote GT and detection respectively." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.563, + 0.47, + 0.82 + ], + "angle": 0, + "content": "Effect of components in collaborative depth estimation. Tab. 3 assesses the gain of collaborative depth estimation (Co-Depth) over single-agent depth estimation (S-Depth) in the depth accuracy metric. We see that: i) Co-Depth steadily improves S-Depth over the full plane; ii) the depth accuracy of the foreground objects is higher than the full plane, as the texture-less backgrounds have fewer cues to localize. Tab. 4 assesses the effects of the depth spacing and supervision choices. We see that: i) overall, the Co-Depth is robust to the different spacing and supervision choices on AP@50/70; ii) linear-increasing spacing is robust to the various supervision choices that uniform spacing which fails the stricter metric AP@80 without dense supervision. The reason is that the linearly increasing spacing takes into account the prior depth distribution and assigns more depth candidates to frequently occurring depth ranges, which contributes to more balanced and easier depth learning." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.842, + 0.327, + 0.858 + ], + "angle": 0, + "content": "5. Conclusion and Limitation" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.47, + 0.901 + ], + "angle": 0, + "content": "We propose CoCa3D, a novel collaborative camera-only 3D detection that approaches holistic 3D detection. The" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.563, + 0.892, + 0.713 + ], + "angle": 0, + "content": "core idea is to introduce multi-agent collaboration to improve the detection ability of cameras. Meanwhile, the communication cost is optimized, and each agent carefully selects the spatially sparse yet critical depth messages to share. Extensive experiments covering both real-world and simulation scenarios, and multi-type agents (cars, drones, and infrastructures) show that CoCa3D not only achieves state-of-the-art perception-bandwidth trade-off, but overtakes LiDAR-based detectors with a sufficient number of collaborative agents on OPV2V+." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.714, + 0.892, + 0.818 + ], + "angle": 0, + "content": "Limitation and future work. It is expensive to collect a real-world multi-agent perception dataset. So far, DAIR-V2X is the sole public real-world dataset, which only has one vehicle and one roadside unit. This work mainly leverages simulation data to validate the proposed novel methods and sketch a promising research direction. We advocate more resources for real-world data collection." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.819, + 0.892, + 0.895 + ], + "angle": 0, + "content": "Acknowledgment. This research is supported by the National Key RD Program of China under Grant 2021ZD0112801, NSFC under Grant 62171276 and the Science and Technology Commission of Shanghai Municipal under Grant 21511100900 and 22DZ2229005." + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9250" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.175, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.171 + ], + "angle": 0, + "content": "[1] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In European conference on computer vision, pages 213-229. Springer, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.172, + 0.472, + 0.24 + ], + "angle": 0, + "content": "[2] Siheng Chen, Baoan Liu, Chen Feng, Carlos Vallespi-Gonzalez, and Carl K. Wellington. 3d point cloud processing and learning for autonomous driving: Impacting map creation, localization, and perception. IEEE Signal Processing Magazine, 38:68-86, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.242, + 0.47, + 0.311 + ], + "angle": 0, + "content": "[3] Jiajun Deng, Shaoshuai Shi, Peiwei Li, Wengang Zhou, Yanyong Zhang, and Houqiang Li. Voxel r-cnn: Towards high performance voxel-based 3d object detection. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 35, pages 1201-1209, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.312, + 0.47, + 0.366 + ], + "angle": 0, + "content": "[4] Alexey Dosovitskiy, German Ros, Felipe Codevilla, Antonio Lopez, and Vladlen Koltun. Carla: An open urban driving simulator. In Conference on robot learning, pages 1-16. PMLR, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.368, + 0.47, + 0.422 + ], + "angle": 0, + "content": "[5] Peng Gao, Rui Guo, Hongsheng Lu, and Hao Zhang. Regularized graph matching for correspondence identification under uncertainty in collaborative perception. Robotics: Science and Systems XVI, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.424, + 0.47, + 0.478 + ], + "angle": 0, + "content": "[6] Yue Hu, Shaoheng Fang, Zixing Lei, Yiqi Zhong, and Siheng Chen. Where2comm: Communication-efficient collaborative perception via spatial confidence maps. Advances in Neural Information Processing Systems, 2022. 2, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.48, + 0.47, + 0.521 + ], + "angle": 0, + "content": "[7] Yue Hu, Shaoheng Fang, Weidi Xie, and Siheng Chen. Aerial monocular 3d object detection. IEEE Robotics and Automation Letters, 8:1959-1966, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.523, + 0.47, + 0.563 + ], + "angle": 0, + "content": "[8] Junjie Huang and Guan Huang. Bevdet4d: Exploit temporal cues in multi-camera 3d object detection. arXiv preprint arXiv:2203.17054, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.565, + 0.47, + 0.618 + ], + "angle": 0, + "content": "[9] Junjie Huang, Guan Huang, Zheng Zhu, and Dalong Du. Bevdet: High-performance multi-camera 3d object detection in bird-eye-view. arXiv preprint arXiv:2112.11790, 2021.1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.621, + 0.47, + 0.689 + ], + "angle": 0, + "content": "[10] Alex H Lang, Sourabh Vora, Holger Caesar, Lubing Zhou, Jiong Yang, and Oscar Beijbom. Pointpillars: Fast encoders for object detection from point clouds. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12697-12705, 2019. 1, 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.691, + 0.469, + 0.731 + ], + "angle": 0, + "content": "[11] Zixing Lei, Shunli Ren, Yue Hu, Wenjun Zhang, and Siheng Chen. Latency-aware collaborative perception. ECCV, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.733, + 0.469, + 0.787 + ], + "angle": 0, + "content": "[12] Jing Li, Rui Li, Jiehao Li, Junzheng Wang, Qingbin Wu, and Xu Liu. Dual-view 3d object recognition and detection via lidar point cloud and camera image. Robotics Auton. Syst., 150:103999, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.789, + 0.469, + 0.844 + ], + "angle": 0, + "content": "[13] Yiming Li, Dekun Ma, Ziyan An, Zixun Wang, Yiqi Zhong, Siheng Chen, and Chen Feng. V2x-sim: Multi-agent collaborative perception dataset and benchmark for autonomous driving. IEEE Robotics and Automation Letters, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.846, + 0.469, + 0.9 + ], + "angle": 0, + "content": "[14] Yiming Li, Shunli Ren, Pengxiang Wu, Siheng Chen, Chen Feng, and Wenjun Zhang. Learning distilled collaboration graph for multi-agent perception. Advances in Neural Information Processing Systems, 34:29541-29552, 2021. 2, 6" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.472, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.147 + ], + "angle": 0, + "content": "[15] Yiming Li, Juexiao Zhang, Dekun Ma, Yue Wang, and Chen Feng. Multi-robot scene completion: Towards task-agnostic collaborative perception. In Conference on Robot Learning, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.15, + 0.892, + 0.218 + ], + "angle": 0, + "content": "[16] Zhiqi Li, Wenhai Wang, Hongyang Li, Enze Xie, Chonghao Sima, Tong Lu, Qiao Yu, and Jifeng Dai. Bevformer: Learning bird's-eye-view representation from multi-camera images via spatiotemporal transformers. arXiv preprint arXiv:2203.17270, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.221, + 0.892, + 0.289 + ], + "angle": 0, + "content": "[17] Yen-Cheng Liu, Junjiao Tian, Nathaniel Glaser, and Zsolt Kira. When2com: Multi-agent perception via communication graph grouping. In Proceedings of the IEEE/CVF Conference on computer vision and pattern recognition, pages 4106-4115, 2020. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.291, + 0.892, + 0.36 + ], + "angle": 0, + "content": "[18] Yen-Cheng Liu, Junjiao Tian, Chih-Yao Ma, Nathan Glaser, Chia-Wen Kuo, and Zsolt Kira. Who2com: Collaborative perception via learnable handshake communication. In 2020 IEEE International Conference on Robotics and Automation (ICRA), pages 6876-6883. IEEE, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.362, + 0.892, + 0.416 + ], + "angle": 0, + "content": "[19] Zhijian Liu, Haotian Tang, Alexander Amini, Xinyu Yang, Huizi Mao, Daniela Rus, and Song Han. Bevfusion: Multi-task multi-sensor fusion with unified bird's-eye view representation. ArXiv, abs/2205.13542, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.418, + 0.892, + 0.472 + ], + "angle": 0, + "content": "[20] Yifan Lu, Quanhao Li, Baoan Liu, Mehrdad Dianati, Chen Feng, Siheng Chen, and Yanfeng Wang. Robust collaborative 3d object detection in presence of pose errors. ICRA, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.475, + 0.892, + 0.53 + ], + "angle": 0, + "content": "[21] Jiageng Mao, Minzhe Niu, Haoyue Bai, Xiaodan Liang, Hang Xu, and Chunjing Xu. Pyramid r-cnn: Towards better performance and adaptability for 3d object detection. ICCV, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.532, + 0.892, + 0.587 + ], + "angle": 0, + "content": "[22] Lang Peng, Zhirong Chen, Zhangjie Fu, Pengpeng Liang, and Erkang Cheng. Bevsegformer: Bird's eye view semantic segmentation from arbitrary camera rigs. arXiv preprint arXiv:2203.04050, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.589, + 0.892, + 0.644 + ], + "angle": 0, + "content": "[23] Jonah Philion and Sanja Fidler. Lift, splat, shoot: Encoding images from arbitrary camera rigs by implicitly unprojecting to 3d. In European Conference on Computer Vision, pages 194-210. Springer, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.646, + 0.892, + 0.713 + ], + "angle": 0, + "content": "[24] Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. Pointnet: Deep learning on point sets for 3d classification and segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 652-660, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.716, + 0.892, + 0.771 + ], + "angle": 0, + "content": "[25] Cody Reading, Ali Harakeh, Julia Chae, and Steven L. Waslander. Categorical depth distribution network for monocular 3d object detection. CVPR, 2021. 1, 2, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.773, + 0.892, + 0.814 + ], + "angle": 0, + "content": "[26] Thomas Roddick, Alex Kendall, and Roberto Cipolla. Orthographic feature transform for monocular 3d object detection. arXiv preprint arXiv:1811.08188, 2018. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.817, + 0.892, + 0.872 + ], + "angle": 0, + "content": "[27] Shital Shah, Debadeepta Dey, Chris Lovett, and Ashish Kapoor. Airsim: High-fidelity visual and physical simulation for autonomous vehicles. In Field and service robotics, pages 621-635. Springer, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.874, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[28] Shaoshuai Shi, Chaoxu Guo, Li Jiang, Zhe Wang, Jianping Shi, Xiaogang Wang, and Hongsheng Li. Pv-rcnn: Point" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "9251" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.092, + 0.47, + 0.134 + ], + "angle": 0, + "content": "voxel feature set abstraction for 3d object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10529-10538, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.135, + 0.47, + 0.203 + ], + "angle": 0, + "content": "[29] Shaoshuai Shi, Li Jiang, Jiajun Deng, Zhe Wang, Chaoxu Guo, Jianping Shi, Xiaogang Wang, and Hongsheng Li. Pvrconn++: Point-voxel feature set abstraction with local vector representation for 3d object detection. arXiv preprint arXiv:2102.00463, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.204, + 0.47, + 0.247 + ], + "angle": 0, + "content": "[30] Tai Wang, Jiangmiao Pang, and Dahua Lin. Monocular 3d object detection with depth from motion. In European Conference on Computer Vision (ECCV), 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.247, + 0.47, + 0.316 + ], + "angle": 0, + "content": "[31] Tsun-Hsuan Wang, Sivabalan Manivasagam, Ming Liang, Bin Yang, Wenyuan Zeng, and Raquel Urtasun. V2vnet: Vehicle-to-vehicle communication for joint perception and prediction. In European Conference on Computer Vision, pages 605–621. Springer, 2020. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.316, + 0.47, + 0.384 + ], + "angle": 0, + "content": "[32] Yue Wang, Vitor Campagnolo Guizilini, Tianyuan Zhang, Yilun Wang, Hang Zhao, and Justin Solomon. Detr3d: 3d object detection from multi-view images via 3d-to-2d queries. In Conference on Robot Learning, pages 180–191. PMLR, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.385, + 0.47, + 0.455 + ], + "angle": 0, + "content": "[33] Pengxiang Wu, Siheng Chen, and Dimitris N. Metaxas. Motionnet: Joint perception and motion prediction for autonomous driving based on bird's eye view maps. 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11382-11392, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.455, + 0.47, + 0.524 + ], + "angle": 0, + "content": "[34] Runsheng Xu, Yi Guo, Xu Han, Xin Xia, Hao Xiang, and Jiaqi Ma. Opencda: an open cooperative driving automation framework integrated with co-simulation. In 2021 IEEE International Intelligent Transportation Systems Conference (ITSC), pages 1155-1162. IEEE, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.525, + 0.47, + 0.579 + ], + "angle": 0, + "content": "[35] Runsheng Xu, Zhengzhong Tu, Hao Xiang, Wei Shao, Bolei Zhou, and Jiaqi Ma. CoBEVT: Cooperative bird's eye view semantic segmentation with sparse transformers. CoRL, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.58, + 0.47, + 0.663 + ], + "angle": 0, + "content": "[36] Runsheng Xu, Hao Xiang, Zhengzhong Tu, Xin Xia, Ming-Hsuan Yang, and Jiaqi Ma. V2x-vit: Vehicle-to-everything cooperative perception with vision transformer. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXXIX, pages 107-124. Springer, 2022. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.664, + 0.47, + 0.734 + ], + "angle": 0, + "content": "[37] Runsheng Xu, Hao Xiang, Xin Xia, Xu Han, Jinlong Li, and Jiaqi Ma. Opv2v: An open benchmark dataset and fusion pipeline for perception with vehicle-to-vehicle communication. In 2022 International Conference on Robotics and Automation (ICRA), pages 2583–2589. IEEE, 2022. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.734, + 0.47, + 0.789 + ], + "angle": 0, + "content": "[38] Zetong Yang, Yanan Sun, Shu Liu, and Jiaya Jia. 3dssd: Point-based 3d single stage object detector. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11040-11048, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.79, + 0.47, + 0.845 + ], + "angle": 0, + "content": "[39] Tianwei Yin, Xingyi Zhou, and Philipp Krahenbuhl. Center-based 3d object detection and tracking. 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11779-11788, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.846, + 0.47, + 0.901 + ], + "angle": 0, + "content": "[40] Haibao Yu, Yizhen Luo, Mao Shu, Yiyi Huo, Zebang Yang, Yifeng Shi, Zhenglong Guo, Hanyu Li, Xing Hu, Jirui Yuan, and Zaiqing Nie. Dair-v2x: A large-scale dataset for vehicleinfrastructure cooperative 3d object detection. In IEEE/CVF" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.092, + 0.892, + 0.12 + ], + "angle": 0, + "content": "Conf. on Computer Vision and Pattern Recognition (CVPR), June 2022. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.121, + 0.892, + 0.189 + ], + "angle": 0, + "content": "[41] Yunpeng Zhang, Zheng Zhu, Wenzhao Zheng, Junjie Huang, Guan Huang, Jie Zhou, and Jiwen Lu. Beverse: Unified perception and prediction in birds-eye-view for vision-centric autonomous driving. arXiv preprint arXiv:2205.09743, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.19, + 0.892, + 0.261 + ], + "angle": 0, + "content": "[42] Wu Zheng, Weiliang Tang, Sijin Chen, Li Jiang, and Chi-Wing Fu. Cia-ssd: Confident iou-aware single-stage object detector from point cloud. In Proceedings of the AAAI conference on artificial intelligence, volume 35, pages 3555–3562, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.262, + 0.892, + 0.318 + ], + "angle": 0, + "content": "[43] Brady Zhou and Philipp Kähenbuhl. Cross-view transformers for real-time map-view semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13760-13769, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.319, + 0.892, + 0.359 + ], + "angle": 0, + "content": "[44] Xingyi Zhou, Dequan Wang, and Philipp Krahenbuhl. Objects as points. In arXiv preprint arXiv:1904.07850, 2019. 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.36, + 0.895, + 0.417 + ], + "angle": 0, + "content": "[45] Yin Zhou and Oncel Tuzel. Voxelnet: End-to-end learning for point cloud based 3d object detection. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4490-4499, 2018. 1, 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.092, + 0.895, + 0.417 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "9252" + } + ] +] \ No newline at end of file diff --git a/2023/Collaboration Helps Camera Overtake LiDAR in 3D Detection/0d6cf930-2724-4762-9022-d5c25917e98a_origin.pdf b/2023/Collaboration Helps Camera Overtake LiDAR in 3D Detection/0d6cf930-2724-4762-9022-d5c25917e98a_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..cade307358aedc049e5d3b33bdccdb64bcc68483 --- /dev/null +++ b/2023/Collaboration Helps Camera Overtake LiDAR in 3D Detection/0d6cf930-2724-4762-9022-d5c25917e98a_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bab7815f87d5ac0578d9574e52a62589a0d6e539410a7e74c30303c340de41d8 +size 5560194 diff --git a/2023/Collaboration Helps Camera Overtake LiDAR in 3D Detection/full.md b/2023/Collaboration Helps Camera Overtake LiDAR in 3D Detection/full.md new file mode 100644 index 0000000000000000000000000000000000000000..ee08081403b7a81c4dfe731c7176fb91fbaf8901 --- /dev/null +++ b/2023/Collaboration Helps Camera Overtake LiDAR in 3D Detection/full.md @@ -0,0 +1,385 @@ +# Collaboration Helps Camera Overtake LiDAR in 3D Detection + +Yue Hu $^{1}$ Yifan Lu $^{1}$ Runsheng Xu $^{2}$ Weidi Xie $^{1,3}$ Siheng Chen $^{1,3*}$ Yanfeng Wang $^{3,1}$ + +$^{1}$ Cooperative Medianet Innovation Center, Shanghai Jiao Tong University $^{2}$ University of California, Los Angeles $^{3}$ Shanghai AI Laboratory + +1{18671129361, yifan.lu, weidi, sihengc, wangyanfeng}@sjtu.edu.cn 2{rxx3386}@ucla.edu + +# Abstract + +Camera-only 3D detection provides an economical solution with a simple configuration for localizing objects in 3D space compared to LiDAR-based detection systems. However, a major challenge lies in precise depth estimation due to the lack of direct 3D measurements in the input. Many previous methods attempt to improve depth estimation through network designs, e.g., deformable layers and larger receptive fields. This work proposes an orthogonal direction, improving the camera-only 3D detection by introducing multi-agent collaborations. Our proposed collaborative camera-only 3D detection (CoCa3D) enables agents to share complementary information with each other through communication. Meanwhile, we optimize communication efficiency by selecting the most informative cues. The shared messages from multiple viewpoints disambiguate the single-agent estimated depth and complement the occluded and long-range regions in the single-agent view. We evaluate CoCa3D in one real-world dataset and two new simulation datasets. Results show that CoCa3D improves previous SOTA performances by $44.21\%$ on DAIR-V2X, $30.60\%$ on OPV2V+, $12.59\%$ on CoPerception-UAVs+ for AP@70. Our preliminary results show a potential that with sufficient collaboration, the camera might overtake LiDAR in some practical scenarios. We released the dataset and code. + +# 1. Introduction + +As a fundamental task of computer vision, 3D object detection aims to localize objects in the 3D physical space given an agent's real-time sensor inputs. It is crucial in a wide range of applications, including autonomous driving [10,33,43], surveillance systems [39], robotics [12] and unmanned aerial vehicles [7]. Depending on sensor setups, there are multiple technical solutions to realize 3D object detection. In this spectrum, one extreme emphasizes raising the upper bound of detection performance, which uses + +![](images/533b5c4540de8a20efd2ffd45831fc9ec38d15407d6ad624c14b541c08725489.jpg) +(a) Scene + +![](images/61367b8982a22dd1c653adceee4bfd0bb0fd5e2078a2c9edfb71b5211cbecb66.jpg) +(b) Single camera confidence + +![](images/a35e09bd0d3ef860530a57319a3a54db804f1935d8f6f5b30f1aa058fc56a776.jpg) +(c) Collaborative cameras + +![](images/214c684466d2a3206bfb5805f21a810ec1793cc90cacdd835bd79e982f9f8ef0.jpg) +(d) Collaborative cameras confidence + +![](images/c1562dc636730206503729f7ec8a58be3630d47f54b8ec8f717805d3f2a66a46.jpg) +(e) LiDAR (Ego) +Figure 1. Collaborative camera-only 3D detection can disambiguate the single-view estimated depth, address the long-range and occlusion issues, and approach LiDAR in 3D detection. + +![](images/5f3093ac9f3cd9f0cd7378b907f7f3e9bc7b7f9b3d60a7fc8569266b9e7daea3.jpg) +(f) LiDAR confidence + +high-end LiDAR sensor [2, 3, 10, 21, 28, 42, 45] to collect precise 3D measurements. However, this approach is too expensive to scale up. The other extreme solution emphasizes cost effectiveness, which tries to use thrifty sensor setups, e.g. only using cameras to detect 3D objects in real-time [1, 8, 9, 23, 25, 26, 30, 32, 41, 43]. However, camera-only 3D detection is significantly and consistently worse than LiDAR-based detection in most scenarios [19]. + +In this paper, we propose an orthogonal direction for improving camera-only 3D detection performances by introducing multi-agent collaborations. Hypothetically, empowered by advanced communication systems, multiple agents equipped only with cameras could share visual information with each other. This would bring three outstanding benefits. First, different viewpoints from multiple agents can largely resolve the depth ambiguity issue in camera-only 3D detection, bridging the gap with expensive LiDARs on depth estimation. Second, multi-agent collaboration avoids inevitable limitations in single-agent 3D detection, such as occlusion and long-range issues, and potentially enables more holistic 3D detection; that is, detecting all the objects existed in the 3D scene, including those beyond visual range. Since LiDAR also suffers from limited field of view, this potentially enables collaborative cameras to outperform LiDAR. Third, the total expense of a large fleet of vehicles is significantly reduced as cameras are much + +cheaper than LiDAR. However, multi-agent collaboration also brings new challenges. Different from many multiview geometry problems, here we also have to concern communication bandwidth constraints. Thus, each agent needs to select the most informative cues to share. + +Following this design rationale, we propose a novel collaborative camera-only 3D detection framework CoCa3D. It includes three parts: i) single-agent camera-only 3D detection, which achieves basic depth estimation and 3D detection for each agent; ii) collaborative depth estimation, which disambiguates the estimated depths by promoting spatial consistency across multiple agents' viewpoints; and iii) collaborative detection feature learning, which complements detection features by sharing key detection messages with each other. Compared to recent collaborative perception methods [11,14] that are dealing with LiDAR, CoCa3D specifically designs novel collaborative depth estimation to customize the task of camera-only 3D detection. + +To evaluate CoCa3D, we conduct comprehensive experiments on one real-world dataset, DAIR-V2X [40], and two new simulation datasets, OPV2V+ and CoPerception-UAVs+, which are extended based on original OPV2V [37] and CoPerception-UAVs [6] with more collaborative agents that cover three types of agents (cars, infrastructures and drones). Our results show that i) with 10 collaborative agents, CoCa3D enables camera-only detectors to overtake LiDAR-based detectors on OPV2V+; and ii) CoCa3D consistently outperforms previous works in the performance-bandwidth trade-off across multiple datasets by a large margin, improving the previous SOTA performances by $30.60\%$ on OPV2V+, $12.59\%$ on CoPerception-UAVs+, $44.21\%$ on DAIR-V2X for AP@70. To sum up, our contributions are: + +- We propose a novel collaborative camera-only 3D detection framework CoCa3D, which improves the detection ability of cameras with multi-agent collaboration, promoting more holistic 3D detection. +- We propose core communication-efficient collaboration techniques, which explore the spatially sparse yet critical depth messages and tackle the depth ambiguity, occlusion, and long-range issues by fusing complementary information from different viewpoints, achieving more accurate and complete 3D representation. +- We expand two previous collaborative datasets with more agents, and conduct extensive experiments, validating that i) CoCa3D significantly bridges the performance gap between camera and LiDAR on OPV2V+ and DAIR-V2X; and ii) CoCa3D achieves the state-of-the-art performance-bandwidth trade-off. + +# 2. Related Work + +Camera-only 3D object detection. Camera-only 3D detection aims to detect objects in the 3D space given the 2D image through explicit or implicit depth estimation [7, 9, 23, 26, 30, 41]. Recently, bird's-eye-view (BEV) repre + +sentations are widely used for their computation efficiency and comparable performance compared to the 3D voxel features [10]. To get the BEV features, there are two types of methods. The depth-based methods [8, 23, 25] first estimate the depth distribution, and then attentively project the 2D image features along the projection ray to get the 3D voxel features and collapse it to BEV feature. The query-based methods [16, 22, 43] first initialize queries for each BEV grid and then leverage transformer-architecture-based cross-attention to query image features with camera-aware positional embedding. Here, our single-agent camera-only detector follows the simple-yet-effective CaDDN [25]. + +LiDAR 3D object detection. LiDAR-based 3D detection achieves excellent performance due to the precise 3D measurements of the input data. Two well-known ways to encode LiDAR points include voxel-based [10, 42, 45] and point-based [28,29,38]. Voxel-based methods divide the 3D space into regular voxels [45] or pillars [10], and encode the point inside into feature representations. Point-based methods are usually based on the PointNet [24] series to aggregate the feature of points. Then the point features will be used for the proposal refinement. LiDAR-based 3D detection performs well, but high-quality LiDAR is hard to be adopted in a large scale due to the expensive cost. Here we propose an economic camera-only solution by introducing multi-agent collaboration, whose performance can catch up with LiDAR given a sufficient number of agents. + +Collaborative perception. Collaborative perception is an emerging application of multi-agent systems, which promotes perception performance by enabling agents to share information with other agents through communication. In this emerging research field, a surge of high-quality datasets, (V2X-SIM [13], OPV2V [37], DAIR-V2X [40], CoPerception-UAVs [6]), and collaborative methods [5,6,13-15,17,18,20,31,35-37,40] aimed for better performance-bandwidth trade-off have been proposed. Collaborative methods have three types: early, late and intermediate fusion. Compared with early fusion(i.e. transmitting raw data) and late fusion (i.e. transmitting outputs), intermediate fusion(i.e. transmitting and fusing features encoded by deep networks) achieves a better performance-bandwidth trade-off. DiscoNet [14] adopts knowledge distillation to take the advantage of both early and intermediate collaboration. Where2comm [6] optimizes communication efficiency by sharing spatially sparse yet perceptually critical information. However, previous works mainly focus on LiDAR-based 3D object detection, while we investigate collaborative camera-only detection and specifically enhance depth estimation via multi-agent collaboration. + +# 3. Collaborative Camera-Only 3D Detection + +This section presents CoCa3D, collaborative camera-only 3D detection framework, which enables multiple + +![](images/ba90d29575aba984d580ef376a6ed1027979f5229376939c1bd197e709656b79.jpg) + +![](images/b813dba96905d50e5b54918027331eabb7e082785cbf05699461eb85e4259a2c.jpg) + +![](images/f4b230329c19b0c42f6a9374a9326edf0826269ac4099d74eddd26d4d45ced13.jpg) +Figure 2. System overview. CoCa3D is a camera-only 3D detector integrated with two collaboration modules. Collaborative depth estimation (Co-Depth) enhances the single-agent estimated depth to achieve more accurate 3D feature. Collaborative detection feature learning (Co-FL) complements the single-agent 3D feature to achieve more holistic 3D detection. + +agents to share visual information with each other, promoting more holistic 3D detection. + +# 3.1. Problem Formulation + +Consider $N$ agents in the scene, let $\mathcal{X}_i$ be the RGB image collected by the $i$ th agent and $\mathcal{O}_i^0$ be the corresponding ground-truth detection. The objective of CoCa3D is to maximize detection performances of all agents given certain communication budget $B$ ; that is, + +$$ +\max _ {\theta , \mathcal {P}} \sum_ {i} g \left(\Phi_ {\theta} \left(\mathcal {X} _ {i}, \{\mathcal {P} _ {i \rightarrow j} \} _ {j = 1} ^ {N}\right), \mathcal {O} _ {i} ^ {0}\right), \mathrm {s . t .} \sum_ {i} | \mathcal {P} _ {i \rightarrow j} | \leq B, +$$ + +where $g(\cdot, \cdot)$ is the detection evaluation metric, $\Phi(\cdot)$ is a detection model with trainable parameter $\theta$ , and $\mathcal{P}_{i \to j}$ is the message transmitted from the $i$ th agent to the $j$ th agent. The key challenge is to determine the messages $\mathcal{P}_{i \to j}$ , which should be both informative and compact. Our design rational comes from two aspects: First, since the major gap between camera and LiDAR is the depth, the message should include depth information. This would allow different viewpoints from multiple agents to disambiguate the infinite depth possibilities and localize the correct depth candidate. Second, the message should include detection clues to provide complementary detection information, which can fundamentally overcome inevitable limitations of single-agent detection, such as occlusion and long-range issues. + +Based on the above intuition, the proposed CoCa3D includes a) single-agent camera-only 3D detection, which achieves basic depth estimation and detection ability, and b) multi-agent collaboration, which shares both estimated depth information and detection features to improve 3D representation and detection performances; see Fig. 2. + +# 3.2. Single-agent camera-only 3D detection + +Single-agent camera-only 3D object detection network learns to detect 3D objects in the physical space based on the 2D camera inputs. Here we adopt CaDNN [25], an off-the-shelf architecture. The main idea is to project the flat image feature into all the possible depths in the 3D space, + +and then attentively aggregate the 3D voxel feature and collapse them into the BEV feature. To mention that, the BEV feature is used because it largely reduces the computation cost, while performing similarly to 3D voxel features [10]. The architecture consists of five modules as detailed below. + +Encoder. The encoder extracts features from the input images. For agent $i$ , given its input $\mathcal{X}_i$ , the output feature map of encoder is $\mathcal{F}_i = \Phi_{\mathrm{enc}}(\mathcal{X}_i) \in \mathbb{R}^{H \times W \times C}$ , where $H, W, C$ are the height, width and channel of the image feature. + +Depth estimation. The depth estimation module predicts pixel-wise categorical depth distribution over a set of predefined depth bins to accurately locate image information in 3D space. Treating the estimation as a classification problem can capture the inherent depth estimation uncertainty to reduce the impact of erroneous depth estimates, as stated in CaDNN [25]. To achieve this, a parametric depth estimation network $\Phi_{\mathrm{Depth}}(\cdot)$ is used, and the pixel-wise categorical depth distribution is obtained as + +$$ +\mathcal {D} _ {i} = \Phi_ {\text {D e p t h}} \left(\mathcal {F} _ {i}\right) \in \mathbb {R} ^ {H \times W \times D}, \tag {1} +$$ + +whose $(h,w)$ th element $\mathcal{D}_i(h,w)\in \mathbb{R}^D$ refers to the estimated depth distribution of the $(h,w)$ th image pixel with $D$ as the number of discretized depth bins. + +Voxel transformation. Voxel transformation projects the image feature to the 3D space, given all the possible depth candidates and the known image calibration matrix, resulting in a 3D voxel representation $\widehat{\mathcal{V}}_i\in \mathbb{R}^{X\times Y\times Z\times C}$ . The camera projection matrix $\mathbf{P}\in \mathbb{R}^{3\times 4}$ defines the mapping between the 3D voxel coordinate to the image pixel coordinate, this is, $[u,v,d] = \mathbf{P}\cdot [x,y,z,1]$ . Similarly, the depth probabilities can be transformed to the 3D voxel space, resulting in $\widehat{\mathcal{D}}_i\in \mathbb{R}^{X\times Y\times Z}$ , where each element indicates the confidence that the feature pixel belongs to the voxel. + +Collapse. The voxel feature is collapsed to a single height plane to generate the bird's-eye-view (BEV) feature as $\overline{B}_i = \Phi_{\mathrm{clp}}(\widehat{\mathcal{D}}_i \odot \widehat{\mathcal{V}}_i) \in \mathbb{R}^{X \times Y \times C}$ , where $\odot$ is the element-wise multiplication. Symbol $\cdot / \widehat{\cdot} / \widehat{\cdot}$ reflects 2D image / 3D voxel / BEV spaces, respectively. The collapse network $\Phi_{\mathrm{clp}}(\cdot)$ + +flattens the 3D voxel feature along the $Z$ -axis and then applies $1 \times 1$ convolution to reduce the channel dimension. + +Decoder. The decoder takes the BEV feature as input, and outputs the objects including class and regression. Here, we adopt CenterNet [44], an off-the-shelf detector. Given the BEV feature $\overline{B}_i$ , the detection decoder $\Phi_{\mathrm{dec}}(\cdot)$ generate the dense heatmap of $i$ th agent by $\overline{\mathcal{O}}_i = \Phi_{\mathrm{dec}}(\overline{B}_i) \in \mathbb{R}^{X \times Y \times 7}$ , where each location of $\overline{\mathcal{O}}_i$ represents a rotated box with class $(c, x, y, h, w, \cos \alpha, \sin \alpha)$ , denoting class confidence, position residual, size and angle. Non-maximum suppression (NMS) is applied to the dense predictions and generate the final sparse output of the 3D detection system. + +Discussion. Admittedly, single-agent camera-only 3D detection makes localizing objects in the 3D space given 2D image input possible through numerous ingenious algorithm designs. The 3D detection performance is still far from perfect due to the ambiguous single-agent depth estimation and limited visual range. Tackling these two issues are challenging for a single agent, but from a perspective of multi-agent collaboration, sharing complementary visual clues makes depth estimation and visual range enlargement much more natural. This motivates our following designs. + +# 3.3. Multi-agent collaboration + +Multi-agent collaboration includes two parts: i) collaborative depth estimation, which enables the sharing of depth information estimated by individual agents to produce more accurate 3D representation and less biased BEV feature; and ii) collaborative detection feature learning, which enables the sharing of detection features of individual agents to achieve more holistic 3D detection. + +# 3.3.1 Collaborative depth estimation + +Collaborative depth estimation (Co-Depth) targets to disambiguate infinite depth possibilities in single-agent camera-only depth estimation and localize the correct depth candidate through multi-view consistency. The intuition is, for a correct depth candidate, its corresponding 3D location should be spatially consistent from multiple agents' viewpoints. To achieve this, each agent can exchange the depth information through communication. Meanwhile, we promote communication efficiency by selecting the most critical and unambiguous depth information. Accordingly, Co-Depth includes: a) depth uncertainty aware message packing, which packs compact messages with unambiguous depth information; and b) depth message fusion, which enhances depth estimation with the received depth messages. + +Depth uncertainty aware message packing. Depth uncertainty-aware (DUA) message packing packs the most critical depth information used for multi-view consistency into the to-be-sent message based on the depth uncertainty. The depth message includes: i) the voxel feature, which is used for the multi-view visual similarity measurement; and ii) the depth probability, which indicates the confidence that + +the feature pixel belongs to the voxel and is used for the multi-view candidate selection. + +The spatially sparse features of $i$ th agent are selected based on its spatial depth uncertainty map in the 2D image space $\mathbf{U}_i\in \mathbb{R}^{H\times W}$ , that can be represented by the pixelwise entropy from corresponding depth distribution, + +$$ +\mathbf {U} _ {i} (h, w) = \mathrm {H} (\mathcal {D} _ {i} (h, w)) \in \mathbb {R}, +$$ + +where $\mathrm{H}(\cdot)$ is the entropy function and $\mathcal{D}_i(h,w)$ is the depth distribution at the $(h,w)$ th image pixel, following from (1). The intuition is, for the locations with low uncertainty score, an agent is confident about which depth bin the feature pixel belongs to. Thus, sharing information at these locations can help with depth estimation accuracy. On the contrary, for the locations with high uncertainty score, an agent is hard to tell which depth bin the feature pixel belongs to. Transmitting these spatial features would introduce ambiguity, and even degrade the single-agent depth estimation. + +Specifically, a binary selection matrix is used to represent whether each location is selected or not, where 1 denotes selected, and 0 elsewhere. Let $\mathbf{M}_{i\rightarrow j}^{d}\in \{0,1\}^{H\times W}$ be a selection matrix to determine the depth message sent from agent $i$ to $j$ . Its $(h,w)$ th element is + +$$ +\mathbf {M} _ {i \rightarrow j} ^ {d} (h, w) = I (\mathbf {U} _ {i} (h, w) < u _ {\text {t h r e}}), +$$ + +where $I(\cdot)$ is an indicator function and $u_{\mathrm{thre}}$ is a predefined threshold. This reflects that a specific spatial area will be selected when its depth uncertainty is below a threshold. + +The selection matrix can be projected in the 3D voxel space, resulting in $\widehat{\mathbf{M}}_{i\rightarrow j}^{d}\in \mathbb{R}^{X\times Y\times Z}$ . Then, the selected voxel feature map $\widehat{\mathcal{Z}}_{i\rightarrow j}$ and the selected depth distribution $\widehat{\mathcal{E}}_{i\rightarrow j}$ sent from the $i$ th agent to the $j$ th agent are obtained as + +$$ +\begin{array}{l} \widehat {\mathcal {Z}} _ {i \rightarrow j} = \widehat {\mathbf {M}} _ {i \rightarrow j} ^ {d} \odot \widehat {\mathcal {V}} _ {i} \in \mathbb {R} ^ {X \times Y \times Z \times C}, \\ \widehat {\mathcal {E}} _ {i \rightarrow j} = \widehat {\mathbf {M}} _ {i \rightarrow j} ^ {d} \odot \widehat {\mathcal {D}} _ {i} \in \mathbb {R} ^ {X \times Y \times Z}, \\ \end{array} +$$ + +where $\widehat{\mathcal{V}}_i$ and $\widehat{\mathcal{D}}_i$ are the full 3D voxel feature map and depth distribution. Overall, the depth message sent from the $i$ th agent to the $j$ th agent is $\mathcal{P}_{i\rightarrow j}^{d} = (\widehat{\mathcal{E}}_{i\rightarrow j},\widehat{\mathcal{Z}}_{i\rightarrow j})$ , where both entries are spatially sparse and communication efficient. + +Depth message fusion. Depth message fusion targets to enhance the depth estimation given the received depth messages through different viewpoints of multiple agents. The intuition is that for a correct depth candidate, the visual features at the same 3D point observed by multiple agents should be similar. To achieve this, we introduce multiview depth consistency weighting. Let $\widehat{S}_i(x,y,z) \in \mathbb{R}$ be the matching score between agent $i$ and its neighbors $\mathcal{N}_i$ , those who share messages with current agent, at coordinate $(x,y,z)$ . The multi-view matching score is obtained as + +$$ +\widehat {\mathcal {W}} _ {j \rightarrow i} (x, y, z) = I (\widehat {\mathcal {E}} _ {j \rightarrow i} (x, y, z) > p _ {\text {t h r e}}) \in \{0, 1 \}, +$$ + +$$ +\widehat {\mathcal {S}} _ {i} (x, y, z) = \sum_ {j \in \mathcal {N} _ {i}} \widehat {\mathcal {W}} _ {j \rightarrow i} (x, y, z) \left\langle \widehat {\mathcal {V}} _ {i} (x, y, z), \widehat {\mathcal {Z}} _ {j \rightarrow i} (x, y, z) \right\rangle , +$$ + +where $\widehat{\mathcal{W}}_{j\rightarrow i}(x,y,z)$ is a binary weight used to filter out the + +neighboring view that has a low depth confidence, below a threshold $p_{\mathrm{thre}}$ and $\langle \cdot ,\cdot \rangle$ is the inner product that measures the visual feature consistency across two views. This means that either the depth candidate is wrong or it is not visible in that view (e.g. due to occlusion), where the multi-view matching score should not be considered. + +Given the multi-view matching score, the collaborative depth distribution is obtained as + +$$ +\widehat {\mathcal {D}} _ {i} ^ {\prime} = \Phi_ {\mathrm {C o D e p t h}} \left(\left[ \widehat {\mathcal {D}} _ {i}; \widehat {\mathcal {S}} _ {i} \right]\right) \in \mathbb {R} ^ {X \times Y \times Z}, +$$ + +where the $[\cdot ]$ denotes concatenation, $\Phi_{\mathrm{CoDepth}}(\cdot)$ is the collaborative depth estimation network implemented with $1\times 1$ convolutions followed by a sigmoid activation. Each element of the collaborative depth distribution $\widehat{D}_i^{\prime}$ reflects the confidence of the feature locates in the corresponding voxel. Note that i) the collaboratively updated depth probability is decided on both the single-view confidence and the multiview consistency; ii) the depth consistency weighting discards the candidates with low single-view depth probability to avoid ambiguity and make the estimation more robust. + +Given this collaborative depth distribution, each agent can weigh the 3D voxel feature and then collapse into a BEV feature with a collapse network $\Phi_{\mathrm{clp}}(\cdot)$ , this is, $\overline{B}_i = \Phi_{\mathrm{clp}}(\widehat{\mathcal{D}}_i' \odot \widehat{\mathcal{V}}_i) \in \mathbb{R}^{X \times Y \times C}$ . The collapse network is implemented with $1 \times 1$ convolution to reduce channel dimension. + +# 3.3.2 Collaborative detection feature learning + +Collaborative depth estimation carefully refines the depth and promotes more accurate 3D representations for each single agent. However, the physical limitations of single agent, such as, limited field of view, occlusion, and long-range issues still remain. To achieve more holistic 3D detection, each agent should be able to exchange 3D detection feature and make use of the complementary information. Meanwhile, we promote communication efficiency by selecting the most perceptually critical information. Therefore, collaborative detection feature learning (Co-FL) includes: a) detection confidence aware message packing, which packs spatially sparse yet perceptually critical 3D features with the guidance of the detection confidence; and ii) detection message fusion, which enhances the 3D feature with the received detection messages. + +Detection confidence aware message packing. The detection confidence aware (DCA) message packing targets to pack the complementary perceptual information into a compact message. The core idea is to explore the spatial heterogeneity of perceptual information. The intuition is that the areas that contain objects are more critical than background areas. During collaboration, areas with objects could help recover the miss-detected objects due to the limited view; and background areas could be omitted to save the precious bandwidth. To achieve this, we implement the spatial confidence map with the detection confidence map by + +$$ +\overline {{\mathbf {C}}} _ {i} = \Phi_ {\mathrm {c l s}} (\overline {{\mathcal {B}}} _ {i}) \in \mathbb {R} ^ {X \times Y}, +$$ + +where $\Phi_{\mathrm{cls}}(\cdot)$ denotes the standard classification network in the single-agent decoder $\Phi_{\mathrm{det}}(\cdot), \overline{B}_i$ is the BEV feature generated with the collaboratively estimated depth. Based on this map, agents decide which spatial area to communicate; that is, each agent offers spatially sparse, yet critical features to support other agents. + +Specifically, a binary selection matrix is used to represent each location is selected or not. Let $\overline{\mathbf{M}}_{i\rightarrow j}^{f}\in$ $\{0,1\}^{X\times Y}$ be a selection matrix to determine the detection message sent from agent $i$ to $j$ . Its $(x,y)$ th element is + +$$ +\overline {{\mathbf {M}}} _ {i \rightarrow j} ^ {f} (x, y) = I (\overline {{\mathbf {C}}} _ {i} (x, y) > c _ {\mathrm {t h r e}}), +$$ + +where $I(\cdot)$ is an indicator function and $c_{\mathrm{thre}}$ is a predefined threshold. This reflects that a specific spatial area will be selected when its detection confidence is above a threshold. Overall, the detection message sent from the $i$ th agent to the $j$ th agent is $\mathcal{P}_{i\rightarrow j}^{f} = \overline{\mathbf{M}}_{i\rightarrow j}^{f}\odot \overline{\mathcal{B}_{i}}$ . Only selected features and their indices are packed, and they can be recovered at the receiver. This greatly reduces communication cost. + +Detection message fusion. Here we augment the detection feature of each agent by aggregating the received detection messages from other agents. We implement this with simple but effective non-parametric point-wise maximum fusion. Specifically, for the $i$ th agent, after receiving the $j$ th agent's message $\mathcal{P}_{j\rightarrow i}^{f}$ . We also include the ego feature map in fusion and denote $\mathcal{P}_{i\rightarrow i}^{f} = \overline{\mathcal{B}}_{i}$ to make the formulation simple and consistent, where $\mathcal{P}_{i\rightarrow i}^{f}$ might not be sparse. The fused BEV feature is obtained as + +$$ +\overline {{\mathcal {B}}} _ {i} ^ {\prime} = \max _ {j \in \mathcal {N} _ {i} \cup \{i \}} (\mathcal {P} _ {j \rightarrow i} ^ {f}) \in \mathbb {R} ^ {X \times Y \times C}, +$$ + +where $\max (\cdot)$ maximizes the corresponding features from multiple agents at each individual spatial location. Note that attention fusion is not permutation invariant, as attention weights vary with the ordering of key and query. Here we simply use the max operator to avoid this permutation variant issue. The fused BEV feature is output to the decoder to generate the final detection. + +# 3.4. Losses + +To train the overall system, we supervise two tasks: categorical depth estimation and 3D object detection. The overall loss is $L = L_{\mathrm{dep}}(\mathcal{D}_i,\mathcal{D}_i^0) + L_{\mathrm{det}}(\overline{\mathcal{O}}_i,\mathcal{O}_i^0)$ , where $\mathcal{D}_i^0$ and $\mathcal{O}_i^0$ is the $i$ th agent's ground-truth depth category and objects, respectively, $L_{\mathrm{dep}}$ and $L_{\mathrm{det}}$ is the depth classification loss [25] and detection loss [44], respectively. + +# 4. Experimental Results + +# 4.1. Datasets and experimental settings + +We cover three datasets, both real-world and simulation scenarios, and multiple agent types. Metric Average Precision (AP) at Intersection-over-Union (IoU) threshold of 0.30, 0.50, 0.70, and 0.80 are used. Metric multi-class classification accuracy is used for depth accuracy. The communication volume follows the standard setting as [6, 36, 37] + +![](images/a2f60f240805255f8f1dc709609a6dca3813614148c32d2799170a0fe999670a.jpg) +(a) OPV2V+ + +![](images/f3107b653a6f4b821e989cb70177cb3af91bab71ec7271131755d18c51d9fe1e.jpg) +(b) DAIR-V2X +Figure 3. CoCa3D steadily improves 3D detection performance as the number of agents grows. In OPV2V+, around 10 collaborative agents enable collaborative cameras to catch up with LiDAR. + +![](images/154529910b4e8f787067760ef14e99b5a99367e2c998e75248a8825d5add2866.jpg) +(c) CoPerception-UAVs+ + +Table 1. CoCa3D significantly outperforms the previous SOTAs, e.g. improve by ${30.60}\%$ on OPV2V, ${12.59}\%$ on CoPerception-UAVs, 44.21% on DAIR-V2X for metric AP@70. + +
MethodOPV2V+CoPerception-UAVs+DAIR-V2X
AP@30AP@50AP@70AP@50AP@70AP@80AP@30AP@50AP@70
No Collaboration0.27480.20410.08530.69560.49000.23090.09770.05240.0305
Late Fusion0.65010.61980.51090.72060.53720.25970.20600.10780.0455
When2com (CVPR'20)0.48530.42110.37370.82190.67050.41020.19570.09840.0459
V2VNet (ECCV'20)0.62460.50420.38520.90930.71770.38040.16400.08470.0512
DiscoNet (NeurIPS'21)0.73000.60090.41790.90540.70790.35640.18360.12620.0683
V2X-ViT (ECCV'22)0.83460.66590.39460.90940.71430.35250.18620.10750.0490
Where2comm (NeurIPS'22)0.81910.70890.47410.91020.73830.36760.17540.10250.0547
CoCa3D0.86420.82600.66750.94970.85020.58350.35220.22600.0985
+ +that counts the message size by byte in log scale with base 2. To compare communication results straightforwardly and fair, we do not consider any extra compression. + +OPV2V+. The original OPV2V [37] is a large-scale vehicle-to-vehicle collaborative perception dataset, cosimulated by OpenCDA [34] and CARLA [4]. Here we introduce an extended version, OPV2V+, which includes more collaborative agents ( $\approx 10$ ). Each agent is equipt with 4 cameras and 4 depth sensors with resolution $600 \times 800$ . The detection range is $281.6\mathrm{m} \times 80\mathrm{m}$ . Our single-agent camera-only detector follows CaDDN [25] with 50 depth categories in linear-increasing spacing mode. + +DAIR-V2X. DAIR-V2X [40] is the only public real-world collaborative perception dataset. It contains two agents: a vehicle and a road-side-unit with image resolution $1080 \times 1920$ . The perception range is $100\mathrm{m} \times 79\mathrm{m}$ . Our single-agent camera-only detector also follows CaDDN [25]. + +CoPerception-UAVs+. The original CoPerception-UAVs [6] is a large-scale UAV-based collaborative perception dataset, co-simulated by AirSim [27] and CARLA [4]. Here we introduce an extended version, CoPerception-UAVs+, which includes more collaborative agents ( $\approx 10$ ). Each agent equips with 1 camera and 1 depth sensor with resolution $450 \times 800$ . The detection range is $192\mathrm{m} \times 96\mathrm{m}$ . Our single-agent camera-only detector follows DVDET [?] with 10 depth categories in uniform spacing mode. + +# 4.2. Quantitative evaluation + +Collaborative camera-only 3D detection overtakes Li-DAR. Fig. 3 investigates how the collaborative camera-only 3D detection performance changes with the number of collaborative agents on multiple datasets. To set up a goal, we consider a LiDAR-based detector implemented by the widely used PointPillar [10]. Fig. 3 (a) shows that: i) with 10 collaborative agents, CoCa3D outperforms the LiDAR 3D detection at both AP@0.5/0.7 on OPV2V+! ii) the detection performance positively increases as the number of collaborative agents. And we then use the slope of OPV2V+ to fit the function of collaborative detection performance with agent number on DAIR-V2X as it only has two agents available. Fig. 3 (b) shows that camera-only 3D detection is expected to outperform LiDAR 3D detection with 7 collaborative vehicles in the real scenario. As LiDAR is not a commonly available equipment for drones, we set a high-quality RGB-D camera instead. Fig. 3 (c) shows that with 3 collaborative drones, camera-only 3D detection outperforms ground-truth depth-based 3D detection. Furthermore, the steadily increasing detection performance with the number of collaborative agents encourages the agents to actively collaborate and achieve consistent improvement. + +Benchmark comparison. Tab. 1 compares the proposed CoCa3D with previous collaborative methods. We consider single-agent detection without collaboration (No collaboration), When2com [17], V2VNet [31], DiscoNet [14], V2XViT [36], Where2comm [6] and late fusion, where agents + +Table 2. CoCa3D outperforms single-agent camera-only 3D detection with GT depth on OPV2V+ and CoPerception-UAVs+. + +
Co -DepthCo -FLOPV2V+CoPerception-UAVs+
AP@30AP@50AP@70AP@50AP@70AP@80
--0.27480.20410.08530.72130.54210.2846
GT-0.34540.25530.09730.83470.67640.4120
-✓0.82010.71910.47560.90840.72560.4028
GT✓0.91200.88050.74340.95050.83980.5504
✓✓0.86420.82600.66750.94950.85180.5849
+ +directly exchange the detected 3D boxes. We see that the proposed CoCa3D significantly outperforms previous state-of-the-arts, improves the SOTA performance by $30.60\%$ on OPV2V+, $12.59\%$ on CoPerception-UAVs+, $44.21\%$ on DAIR-V2X for AP@70. The reason is that collaborative depth estimation promotes more accurate 3D features per single agent, and the enhanced single agent features further facilitate collaborative detection feature learning, while previous collaborative methods do not specifically consider depth ambiguity and collaborate over single-agent features. + +Multi-agent collaboration evaluation. Tab. 2 assess the effectiveness of the proposed multi-agent collaboration on OPV2V+ and CoPerception-UAVs+ datasets. Ground-truth (GT) depth is included to provide the upper bound of collaborative camera-only 3D detection. We see that: i) both collaborations can consistently improve the performance; ii) camera-only 3D detection with collaboratively estimated depth significantly outperforms that with single-agent estimated depth and even approaches the upper bound with ground truth depth. + +# 4.3. Qualitative evaluation + +Visualization of depth and uncertainty. Fig. 5 shows that collaborative depth estimation outperforms single-agent depth estimation and approaches the ground truth depth. We see that: i) the single-agent depth estimation can estimate the relative depth while not able to precisely localize the depth candidate, for example, the vehicles are higher than the plane it sits while this plane is not grounded to the right category; ii) by introducing multi-view geometry, the collaboratively estimated depth can smoothly and accurately ground the plane; iii) the depth uncertainty is larger for the long-range and background areas. The reason is that the distant regions are hard to localize as they occupy too few image pixels, and the background regions are hard to localize due to the texture-less surfaces. + +Visualization of detection results. Fig. 6 shows that compared to single LiDAR 3D detection, cameras with collaborations (Co-Depth and Co-FL) are able to achieve holistic and accurate detection results. The reason is that the single agent has some fundamental physical limitations such as long-range and occlusion issues. Fig. 6(a-c) shows that augmenting a single agent sensor going from just a cam + +![](images/7777d30890ed305547aa55805548455d59034ed95194532e5a35d0e5ef057308.jpg) +(a) OPV2V+ +Figure 4. CoCa3D achieves superior detection performance and communication cost trade-off over various bandwidths. + +![](images/1efed90c2d1174b3e405e6cdf2e94a7b0bc072f0c7b1c2f2cef2af5e12efc3f6.jpg) +(b) CoPerception-UAVs+ + +Table 3. Co-Depth significantly improves depth accuracy. + +
Full planeForeground
SingleCollaborationGainSingleCollaborationGain
0.61670.778126.17% ↑0.75860.81337.21% ↑
+ +Table 4. Ablation of collaborative depth estimation on CoPerception-UAVs+. Uniform/Linear denotes uniform/linear-increasing spacing. Dense and sparse supervision denote applying depth supervision over the full image plane and the object regions. + +
SpacingSupervisionAP@50AP@70AP@80
UniformDense0.95020.84440.5746
UniformSparse0.93300.80420.4829
LinearDense0.94840.83650.5580
LinearSparse0.94950.85180.5849
+ +era to including a depth sensor to LiDAR can improve the detection but still not able to achieve holistic detection. Fig. 6(d)/(b) and (a)/(e) compares the detection results with and without collaborative feature learning, we see that collaboration can help detect lots of the missed objects in the single-agent detection. Fig. 6(d,e,f) show that: i) both collaborations, Co-FL and Co-Depth, consistently improve the detection performance; ii) with collaboration, collaborative camera-only agents can outperform LiDAR in 3D detection. + +# 4.4. Ablation studies + +Trade-off between detection performance and communication cost. Fig. 4 compares the proposed CoCa3D with the previous communication-efficient solution Where2comm [6] in terms of the trade-off between detection performance (AP@IoU=50/70) and communication bandwidth. We see that: i) CoCa3D can adapt to varying bandwidths by adjusting the depth uncertainty threshold and the detection confidence threshold; ii) CoCa3D achieves superior detection performance and communication cost trade-off to Where2comm over varying communication bandwidth. The gain mainly comes from Co-Depth's improvement in depth estimation, which improves single-agent BEV features. With enhanced single BEV features, the Co-FL can generate better-augmented BEV features, resulting in improved detection performance. + +![](images/49de5e98a900e8abb045db6097aa3993efcc7c528e71b8f4b90dcf6750adbe55.jpg) +(a) Image (RV) + +![](images/3fc22249c1fb463024213745e99f77430f96b8744179701430810a2aca1e2fee.jpg) +(b) Single (RV) + +![](images/d746ffc473f0c0594175c27fed6605a5b25699246da947bda10e7d22896aef5d.jpg) +(c)Collaboration (RV) + +![](images/4019488b619a7896b4c0c67e40c332734d8f2995f3050e79fa04461aa19a7502.jpg) +(d) Ground-truth (RV) + +![](images/aa6856ffbdc0eeea699590483143f3d64ca6696b8600dfa42d15a6686aed6910.jpg) +(e) Uncertainty (RV) + +![](images/73672d6740f21637bb69e8ea5101c0fb0aa14f4de5ff80c10b926cd92fc006cd.jpg) +(f) Image (BEV) + +![](images/e6291efcddbb4e23c71b20dfe063e7e8970380bdd63dc36cd19e54b3479b9c38.jpg) +(g) Single (BEV) + +![](images/4f1144a47c1d8e63a11af41658ae55578cb7fafc149bdc84ba85a13e5c349c11.jpg) +(h) Collaboration (BEV) + +![](images/de294883f4a13cb7c4b41047ad8ae3ba2eba029777cac267532ca40dedffa61f.jpg) +(i) Ground-truth (BEV) + +![](images/4fe9ba96c50e05d317ae2fab159eeda1156aff9ea17b892d99293ef81a1c6764.jpg) +(j) Uncertainty (BEV) + +![](images/c2bb5012670046e2025e884429e6cf073212ba3da5ea009e528deea4069c0f0f.jpg) +Figure 5. Visualization of the depth and uncertainty in image range-view (RV) and bird's eye view (BEV) on CoPerception-UAVs+. Collaboratively estimated depth improves the single-agent estimated depth and approaches ground-truth depth. + +![](images/ce287aecd05eb7304316a8822cb786807001242aeda91ee6e0a410610d4acbed.jpg) +(b) RGB-D + +![](images/81df8c719b4e4cc6d7e17501777c517f84540672e18322fc4ad5defc851ca1e6.jpg) +(c) LiDAR + +![](images/c1b171869492ab1e4755a397d005d66d30997625f0473ecdb2f25f353a344e20.jpg) +(a) Camera +(d) Camera with Co-FL +Figure 6. CoCa3D outperforms LiDAR detection on OPV2V+ with 10 agents. Green and red boxes denote GT and detection respectively. + +![](images/8e6cd7224aaea17acc63f899b115ab304f18b4fc78917d02a20460a5064e3d1b.jpg) +(e) RGB-D with Co-FL + +![](images/a9af25cbb9ff2f40e35bed9a78b359d4dee4cb964cd8837399df92d88a24c7a5.jpg) +(f) Camera with Co-Depth and Co-FL (CoCa3D) + +Effect of components in collaborative depth estimation. Tab. 3 assesses the gain of collaborative depth estimation (Co-Depth) over single-agent depth estimation (S-Depth) in the depth accuracy metric. We see that: i) Co-Depth steadily improves S-Depth over the full plane; ii) the depth accuracy of the foreground objects is higher than the full plane, as the texture-less backgrounds have fewer cues to localize. Tab. 4 assesses the effects of the depth spacing and supervision choices. We see that: i) overall, the Co-Depth is robust to the different spacing and supervision choices on AP@50/70; ii) linear-increasing spacing is robust to the various supervision choices that uniform spacing which fails the stricter metric AP@80 without dense supervision. The reason is that the linearly increasing spacing takes into account the prior depth distribution and assigns more depth candidates to frequently occurring depth ranges, which contributes to more balanced and easier depth learning. + +# 5. Conclusion and Limitation + +We propose CoCa3D, a novel collaborative camera-only 3D detection that approaches holistic 3D detection. The + +core idea is to introduce multi-agent collaboration to improve the detection ability of cameras. Meanwhile, the communication cost is optimized, and each agent carefully selects the spatially sparse yet critical depth messages to share. Extensive experiments covering both real-world and simulation scenarios, and multi-type agents (cars, drones, and infrastructures) show that CoCa3D not only achieves state-of-the-art perception-bandwidth trade-off, but overtakes LiDAR-based detectors with a sufficient number of collaborative agents on OPV2V+. + +Limitation and future work. It is expensive to collect a real-world multi-agent perception dataset. So far, DAIR-V2X is the sole public real-world dataset, which only has one vehicle and one roadside unit. This work mainly leverages simulation data to validate the proposed novel methods and sketch a promising research direction. We advocate more resources for real-world data collection. + +Acknowledgment. This research is supported by the National Key RD Program of China under Grant 2021ZD0112801, NSFC under Grant 62171276 and the Science and Technology Commission of Shanghai Municipal under Grant 21511100900 and 22DZ2229005. + +# References + +[1] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In European conference on computer vision, pages 213-229. Springer, 2020. 1 +[2] Siheng Chen, Baoan Liu, Chen Feng, Carlos Vallespi-Gonzalez, and Carl K. Wellington. 3d point cloud processing and learning for autonomous driving: Impacting map creation, localization, and perception. IEEE Signal Processing Magazine, 38:68-86, 2021. 1 +[3] Jiajun Deng, Shaoshuai Shi, Peiwei Li, Wengang Zhou, Yanyong Zhang, and Houqiang Li. Voxel r-cnn: Towards high performance voxel-based 3d object detection. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 35, pages 1201-1209, 2021. 1 +[4] Alexey Dosovitskiy, German Ros, Felipe Codevilla, Antonio Lopez, and Vladlen Koltun. Carla: An open urban driving simulator. In Conference on robot learning, pages 1-16. PMLR, 2017. 6 +[5] Peng Gao, Rui Guo, Hongsheng Lu, and Hao Zhang. Regularized graph matching for correspondence identification under uncertainty in collaborative perception. Robotics: Science and Systems XVI, 2020. 2 +[6] Yue Hu, Shaoheng Fang, Zixing Lei, Yiqi Zhong, and Siheng Chen. Where2comm: Communication-efficient collaborative perception via spatial confidence maps. Advances in Neural Information Processing Systems, 2022. 2, 5, 6, 7 +[7] Yue Hu, Shaoheng Fang, Weidi Xie, and Siheng Chen. Aerial monocular 3d object detection. IEEE Robotics and Automation Letters, 8:1959-1966, 2022. 1, 2 +[8] Junjie Huang and Guan Huang. Bevdet4d: Exploit temporal cues in multi-camera 3d object detection. arXiv preprint arXiv:2203.17054, 2022. 1, 2 +[9] Junjie Huang, Guan Huang, Zheng Zhu, and Dalong Du. Bevdet: High-performance multi-camera 3d object detection in bird-eye-view. arXiv preprint arXiv:2112.11790, 2021.1, 2 +[10] Alex H Lang, Sourabh Vora, Holger Caesar, Lubing Zhou, Jiong Yang, and Oscar Beijbom. Pointpillars: Fast encoders for object detection from point clouds. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12697-12705, 2019. 1, 2, 3, 6 +[11] Zixing Lei, Shunli Ren, Yue Hu, Wenjun Zhang, and Siheng Chen. Latency-aware collaborative perception. ECCV, 2022. 2 +[12] Jing Li, Rui Li, Jiehao Li, Junzheng Wang, Qingbin Wu, and Xu Liu. Dual-view 3d object recognition and detection via lidar point cloud and camera image. Robotics Auton. Syst., 150:103999, 2022. 1 +[13] Yiming Li, Dekun Ma, Ziyan An, Zixun Wang, Yiqi Zhong, Siheng Chen, and Chen Feng. V2x-sim: Multi-agent collaborative perception dataset and benchmark for autonomous driving. IEEE Robotics and Automation Letters, 2022. 2 +[14] Yiming Li, Shunli Ren, Pengxiang Wu, Siheng Chen, Chen Feng, and Wenjun Zhang. Learning distilled collaboration graph for multi-agent perception. Advances in Neural Information Processing Systems, 34:29541-29552, 2021. 2, 6 + +[15] Yiming Li, Juexiao Zhang, Dekun Ma, Yue Wang, and Chen Feng. Multi-robot scene completion: Towards task-agnostic collaborative perception. In Conference on Robot Learning, 2022. 2 +[16] Zhiqi Li, Wenhai Wang, Hongyang Li, Enze Xie, Chonghao Sima, Tong Lu, Qiao Yu, and Jifeng Dai. Bevformer: Learning bird's-eye-view representation from multi-camera images via spatiotemporal transformers. arXiv preprint arXiv:2203.17270, 2022. 2 +[17] Yen-Cheng Liu, Junjiao Tian, Nathaniel Glaser, and Zsolt Kira. When2com: Multi-agent perception via communication graph grouping. In Proceedings of the IEEE/CVF Conference on computer vision and pattern recognition, pages 4106-4115, 2020. 2, 6 +[18] Yen-Cheng Liu, Junjiao Tian, Chih-Yao Ma, Nathan Glaser, Chia-Wen Kuo, and Zsolt Kira. Who2com: Collaborative perception via learnable handshake communication. In 2020 IEEE International Conference on Robotics and Automation (ICRA), pages 6876-6883. IEEE, 2020. 2 +[19] Zhijian Liu, Haotian Tang, Alexander Amini, Xinyu Yang, Huizi Mao, Daniela Rus, and Song Han. Bevfusion: Multi-task multi-sensor fusion with unified bird's-eye view representation. ArXiv, abs/2205.13542, 2022. 1 +[20] Yifan Lu, Quanhao Li, Baoan Liu, Mehrdad Dianati, Chen Feng, Siheng Chen, and Yanfeng Wang. Robust collaborative 3d object detection in presence of pose errors. ICRA, 2023. 2 +[21] Jiageng Mao, Minzhe Niu, Haoyue Bai, Xiaodan Liang, Hang Xu, and Chunjing Xu. Pyramid r-cnn: Towards better performance and adaptability for 3d object detection. ICCV, 2021. 1 +[22] Lang Peng, Zhirong Chen, Zhangjie Fu, Pengpeng Liang, and Erkang Cheng. Bevsegformer: Bird's eye view semantic segmentation from arbitrary camera rigs. arXiv preprint arXiv:2203.04050, 2022. 2 +[23] Jonah Philion and Sanja Fidler. Lift, splat, shoot: Encoding images from arbitrary camera rigs by implicitly unprojecting to 3d. In European Conference on Computer Vision, pages 194-210. Springer, 2020. 1, 2 +[24] Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. Pointnet: Deep learning on point sets for 3d classification and segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 652-660, 2017. 2 +[25] Cody Reading, Ali Harakeh, Julia Chae, and Steven L. Waslander. Categorical depth distribution network for monocular 3d object detection. CVPR, 2021. 1, 2, 3, 5, 6 +[26] Thomas Roddick, Alex Kendall, and Roberto Cipolla. Orthographic feature transform for monocular 3d object detection. arXiv preprint arXiv:1811.08188, 2018. 1, 2 +[27] Shital Shah, Debadeepta Dey, Chris Lovett, and Ashish Kapoor. Airsim: High-fidelity visual and physical simulation for autonomous vehicles. In Field and service robotics, pages 621-635. Springer, 2018. 6 +[28] Shaoshuai Shi, Chaoxu Guo, Li Jiang, Zhe Wang, Jianping Shi, Xiaogang Wang, and Hongsheng Li. Pv-rcnn: Point + +voxel feature set abstraction for 3d object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10529-10538, 2020. 1, 2 +[29] Shaoshuai Shi, Li Jiang, Jiajun Deng, Zhe Wang, Chaoxu Guo, Jianping Shi, Xiaogang Wang, and Hongsheng Li. Pvrconn++: Point-voxel feature set abstraction with local vector representation for 3d object detection. arXiv preprint arXiv:2102.00463, 2021. 2 +[30] Tai Wang, Jiangmiao Pang, and Dahua Lin. Monocular 3d object detection with depth from motion. In European Conference on Computer Vision (ECCV), 2022. 1, 2 +[31] Tsun-Hsuan Wang, Sivabalan Manivasagam, Ming Liang, Bin Yang, Wenyuan Zeng, and Raquel Urtasun. V2vnet: Vehicle-to-vehicle communication for joint perception and prediction. In European Conference on Computer Vision, pages 605–621. Springer, 2020. 2, 6 +[32] Yue Wang, Vitor Campagnolo Guizilini, Tianyuan Zhang, Yilun Wang, Hang Zhao, and Justin Solomon. Detr3d: 3d object detection from multi-view images via 3d-to-2d queries. In Conference on Robot Learning, pages 180–191. PMLR, 2022. 1 +[33] Pengxiang Wu, Siheng Chen, and Dimitris N. Metaxas. Motionnet: Joint perception and motion prediction for autonomous driving based on bird's eye view maps. 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11382-11392, 2020. 1 +[34] Runsheng Xu, Yi Guo, Xu Han, Xin Xia, Hao Xiang, and Jiaqi Ma. Opencda: an open cooperative driving automation framework integrated with co-simulation. In 2021 IEEE International Intelligent Transportation Systems Conference (ITSC), pages 1155-1162. IEEE, 2021. 6 +[35] Runsheng Xu, Zhengzhong Tu, Hao Xiang, Wei Shao, Bolei Zhou, and Jiaqi Ma. CoBEVT: Cooperative bird's eye view semantic segmentation with sparse transformers. CoRL, 2022. 2 +[36] Runsheng Xu, Hao Xiang, Zhengzhong Tu, Xin Xia, Ming-Hsuan Yang, and Jiaqi Ma. V2x-vit: Vehicle-to-everything cooperative perception with vision transformer. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXXIX, pages 107-124. Springer, 2022. 2, 5, 6 +[37] Runsheng Xu, Hao Xiang, Xin Xia, Xu Han, Jinlong Li, and Jiaqi Ma. Opv2v: An open benchmark dataset and fusion pipeline for perception with vehicle-to-vehicle communication. In 2022 International Conference on Robotics and Automation (ICRA), pages 2583–2589. IEEE, 2022. 2, 5, 6 +[38] Zetong Yang, Yanan Sun, Shu Liu, and Jiaya Jia. 3dssd: Point-based 3d single stage object detector. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11040-11048, 2020. 2 +[39] Tianwei Yin, Xingyi Zhou, and Philipp Krahenbuhl. Center-based 3d object detection and tracking. 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11779-11788, 2021. 1 +[40] Haibao Yu, Yizhen Luo, Mao Shu, Yiyi Huo, Zebang Yang, Yifeng Shi, Zhenglong Guo, Hanyu Li, Xing Hu, Jirui Yuan, and Zaiqing Nie. Dair-v2x: A large-scale dataset for vehicleinfrastructure cooperative 3d object detection. In IEEE/CVF + +Conf. on Computer Vision and Pattern Recognition (CVPR), June 2022. 2, 6 +[41] Yunpeng Zhang, Zheng Zhu, Wenzhao Zheng, Junjie Huang, Guan Huang, Jie Zhou, and Jiwen Lu. Beverse: Unified perception and prediction in birds-eye-view for vision-centric autonomous driving. arXiv preprint arXiv:2205.09743, 2022. 1, 2 +[42] Wu Zheng, Weiliang Tang, Sijin Chen, Li Jiang, and Chi-Wing Fu. Cia-ssd: Confident iou-aware single-stage object detector from point cloud. In Proceedings of the AAAI conference on artificial intelligence, volume 35, pages 3555–3562, 2021. 1, 2 +[43] Brady Zhou and Philipp Kähenbuhl. Cross-view transformers for real-time map-view semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13760-13769, 2022. 1, 2 +[44] Xingyi Zhou, Dequan Wang, and Philipp Krahenbuhl. Objects as points. In arXiv preprint arXiv:1904.07850, 2019. 4, 5 +[45] Yin Zhou and Oncel Tuzel. Voxelnet: End-to-end learning for point cloud based 3d object detection. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4490-4499, 2018. 1, 2 \ No newline at end of file diff --git a/2023/Collaboration Helps Camera Overtake LiDAR in 3D Detection/images.zip b/2023/Collaboration Helps Camera Overtake LiDAR in 3D Detection/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..a0d1c3949b3290217aed34d9253be0c7c8f228b1 --- /dev/null +++ b/2023/Collaboration Helps Camera Overtake LiDAR in 3D Detection/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a2d2bd70a81ce66f67ca4bcecf9e9b1d8856c8346101a2036ca1f1aa65176364 +size 576169 diff --git a/2023/Collaboration Helps Camera Overtake LiDAR in 3D Detection/layout.json b/2023/Collaboration Helps Camera Overtake LiDAR in 3D Detection/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..a100b7115ca22638d8dd42ad7f0fb7cc0fe18fdf --- /dev/null +++ b/2023/Collaboration Helps Camera Overtake LiDAR in 3D Detection/layout.json @@ -0,0 +1,10771 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 103, + 103, + 491, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 103, + 103, + 491, + 121 + ], + "spans": [ + { + "bbox": [ + 103, + 103, + 491, + 121 + ], + "type": "text", + "content": "Collaboration Helps Camera Overtake LiDAR in 3D Detection" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 70, + 141, + 522, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 141, + 522, + 157 + ], + "spans": [ + { + "bbox": [ + 70, + 141, + 522, + 157 + ], + "type": "text", + "content": "Yue Hu" + }, + { + "bbox": [ + 70, + 141, + 522, + 157 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 70, + 141, + 522, + 157 + ], + "type": "text", + "content": " Yifan Lu" + }, + { + "bbox": [ + 70, + 141, + 522, + 157 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 70, + 141, + 522, + 157 + ], + "type": "text", + "content": " Runsheng Xu" + }, + { + "bbox": [ + 70, + 141, + 522, + 157 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 70, + 141, + 522, + 157 + ], + "type": "text", + "content": " Weidi Xie" + }, + { + "bbox": [ + 70, + 141, + 522, + 157 + ], + "type": "inline_equation", + "content": "^{1,3}" + }, + { + "bbox": [ + 70, + 141, + 522, + 157 + ], + "type": "text", + "content": " Siheng Chen" + }, + { + "bbox": [ + 70, + 141, + 522, + 157 + ], + "type": "inline_equation", + "content": "^{1,3*}" + }, + { + "bbox": [ + 70, + 141, + 522, + 157 + ], + "type": "text", + "content": " Yanfeng Wang" + }, + { + "bbox": [ + 70, + 141, + 522, + 157 + ], + "type": "inline_equation", + "content": "^{3,1}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 119, + 157, + 474, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 119, + 157, + 474, + 185 + ], + "spans": [ + { + "bbox": [ + 119, + 157, + 474, + 185 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 119, + 157, + 474, + 185 + ], + "type": "text", + "content": "Cooperative Medianet Innovation Center, Shanghai Jiao Tong University " + }, + { + "bbox": [ + 119, + 157, + 474, + 185 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 119, + 157, + 474, + 185 + ], + "type": "text", + "content": "University of California, Los Angeles " + }, + { + "bbox": [ + 119, + 157, + 474, + 185 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 119, + 157, + 474, + 185 + ], + "type": "text", + "content": "Shanghai AI Laboratory" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 64, + 185, + 528, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 185, + 528, + 200 + ], + "spans": [ + { + "bbox": [ + 64, + 185, + 528, + 200 + ], + "type": "text", + "content": "1{18671129361, yifan.lu, weidi, sihengc, wangyanfeng}@sjtu.edu.cn 2{rxx3386}@ucla.edu" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 143, + 227, + 191, + 239 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 227, + 191, + 239 + ], + "spans": [ + { + "bbox": [ + 143, + 227, + 191, + 239 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 252, + 290, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 252, + 290, + 552 + ], + "spans": [ + { + "bbox": [ + 46, + 252, + 290, + 552 + ], + "type": "text", + "content": "Camera-only 3D detection provides an economical solution with a simple configuration for localizing objects in 3D space compared to LiDAR-based detection systems. However, a major challenge lies in precise depth estimation due to the lack of direct 3D measurements in the input. Many previous methods attempt to improve depth estimation through network designs, e.g., deformable layers and larger receptive fields. This work proposes an orthogonal direction, improving the camera-only 3D detection by introducing multi-agent collaborations. Our proposed collaborative camera-only 3D detection (CoCa3D) enables agents to share complementary information with each other through communication. Meanwhile, we optimize communication efficiency by selecting the most informative cues. The shared messages from multiple viewpoints disambiguate the single-agent estimated depth and complement the occluded and long-range regions in the single-agent view. We evaluate CoCa3D in one real-world dataset and two new simulation datasets. Results show that CoCa3D improves previous SOTA performances by " + }, + { + "bbox": [ + 46, + 252, + 290, + 552 + ], + "type": "inline_equation", + "content": "44.21\\%" + }, + { + "bbox": [ + 46, + 252, + 290, + 552 + ], + "type": "text", + "content": " on DAIR-V2X, " + }, + { + "bbox": [ + 46, + 252, + 290, + 552 + ], + "type": "inline_equation", + "content": "30.60\\%" + }, + { + "bbox": [ + 46, + 252, + 290, + 552 + ], + "type": "text", + "content": " on OPV2V+, " + }, + { + "bbox": [ + 46, + 252, + 290, + 552 + ], + "type": "inline_equation", + "content": "12.59\\%" + }, + { + "bbox": [ + 46, + 252, + 290, + 552 + ], + "type": "text", + "content": " on CoPerception-UAVs+ for AP@70. Our preliminary results show a potential that with sufficient collaboration, the camera might overtake LiDAR in some practical scenarios. We released the dataset and code." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 567, + 128, + 580 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 567, + 128, + 580 + ], + "spans": [ + { + "bbox": [ + 47, + 567, + 128, + 580 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 585, + 287, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 585, + 287, + 693 + ], + "spans": [ + { + "bbox": [ + 46, + 585, + 287, + 693 + ], + "type": "text", + "content": "As a fundamental task of computer vision, 3D object detection aims to localize objects in the 3D physical space given an agent's real-time sensor inputs. It is crucial in a wide range of applications, including autonomous driving [10,33,43], surveillance systems [39], robotics [12] and unmanned aerial vehicles [7]. Depending on sensor setups, there are multiple technical solutions to realize 3D object detection. In this spectrum, one extreme emphasizes raising the upper bound of detection performance, which uses" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 310, + 220, + 445, + 259 + ], + "blocks": [ + { + "bbox": [ + 310, + 220, + 445, + 259 + ], + "lines": [ + { + "bbox": [ + 310, + 220, + 445, + 259 + ], + "spans": [ + { + "bbox": [ + 310, + 220, + 445, + 259 + ], + "type": "image", + "image_path": "533b5c4540de8a20efd2ffd45831fc9ec38d15407d6ad624c14b541c08725489.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 368, + 259, + 391, + 266 + ], + "lines": [ + { + "bbox": [ + 368, + 259, + 391, + 266 + ], + "spans": [ + { + "bbox": [ + 368, + 259, + 391, + 266 + ], + "type": "text", + "content": "(a) Scene" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 445, + 220, + 539, + 259 + ], + "blocks": [ + { + "bbox": [ + 445, + 220, + 539, + 259 + ], + "lines": [ + { + "bbox": [ + 445, + 220, + 539, + 259 + ], + "spans": [ + { + "bbox": [ + 445, + 220, + 539, + 259 + ], + "type": "image", + "image_path": "61367b8982a22dd1c653adceee4bfd0bb0fd5e2078a2c9edfb71b5211cbecb66.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 457, + 259, + 529, + 266 + ], + "lines": [ + { + "bbox": [ + 457, + 259, + 529, + 266 + ], + "spans": [ + { + "bbox": [ + 457, + 259, + 529, + 266 + ], + "type": "text", + "content": "(b) Single camera confidence" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 310, + 266, + 444, + 305 + ], + "blocks": [ + { + "bbox": [ + 310, + 266, + 444, + 305 + ], + "lines": [ + { + "bbox": [ + 310, + 266, + 444, + 305 + ], + "spans": [ + { + "bbox": [ + 310, + 266, + 444, + 305 + ], + "type": "image", + "image_path": "a35e09bd0d3ef860530a57319a3a54db804f1935d8f6f5b30f1aa058fc56a776.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 347, + 305, + 410, + 312 + ], + "lines": [ + { + "bbox": [ + 347, + 305, + 410, + 312 + ], + "spans": [ + { + "bbox": [ + 347, + 305, + 410, + 312 + ], + "type": "text", + "content": "(c) Collaborative cameras" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 445, + 266, + 539, + 306 + ], + "blocks": [ + { + "bbox": [ + 445, + 266, + 539, + 306 + ], + "lines": [ + { + "bbox": [ + 445, + 266, + 539, + 306 + ], + "spans": [ + { + "bbox": [ + 445, + 266, + 539, + 306 + ], + "type": "image", + "image_path": "214c684466d2a3206bfb5805f21a810ec1793cc90cacdd835bd79e982f9f8ef0.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 447, + 306, + 539, + 312 + ], + "lines": [ + { + "bbox": [ + 447, + 306, + 539, + 312 + ], + "spans": [ + { + "bbox": [ + 447, + 306, + 539, + 312 + ], + "type": "text", + "content": "(d) Collaborative cameras confidence" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 309, + 312, + 444, + 351 + ], + "blocks": [ + { + "bbox": [ + 309, + 312, + 444, + 351 + ], + "lines": [ + { + "bbox": [ + 309, + 312, + 444, + 351 + ], + "spans": [ + { + "bbox": [ + 309, + 312, + 444, + 351 + ], + "type": "image", + "image_path": "c1562dc636730206503729f7ec8a58be3630d47f54b8ec8f717805d3f2a66a46.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 359, + 351, + 400, + 360 + ], + "lines": [ + { + "bbox": [ + 359, + 351, + 400, + 360 + ], + "spans": [ + { + "bbox": [ + 359, + 351, + 400, + 360 + ], + "type": "text", + "content": "(e) LiDAR (Ego)" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 305, + 361, + 545, + 394 + ], + "lines": [ + { + "bbox": [ + 305, + 361, + 545, + 394 + ], + "spans": [ + { + "bbox": [ + 305, + 361, + 545, + 394 + ], + "type": "text", + "content": "Figure 1. Collaborative camera-only 3D detection can disambiguate the single-view estimated depth, address the long-range and occlusion issues, and approach LiDAR in 3D detection." + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 445, + 312, + 539, + 351 + ], + "blocks": [ + { + "bbox": [ + 445, + 312, + 539, + 351 + ], + "lines": [ + { + "bbox": [ + 445, + 312, + 539, + 351 + ], + "spans": [ + { + "bbox": [ + 445, + 312, + 539, + 351 + ], + "type": "image", + "image_path": "5f3093ac9f3cd9f0cd7378b907f7f3e9bc7b7f9b3d60a7fc8569266b9e7daea3.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 466, + 352, + 519, + 359 + ], + "lines": [ + { + "bbox": [ + 466, + 352, + 519, + 359 + ], + "spans": [ + { + "bbox": [ + 466, + 352, + 519, + 359 + ], + "type": "text", + "content": "(f) LiDAR confidence" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 400, + 546, + 497 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 400, + 546, + 497 + ], + "spans": [ + { + "bbox": [ + 304, + 400, + 546, + 497 + ], + "type": "text", + "content": "high-end LiDAR sensor [2, 3, 10, 21, 28, 42, 45] to collect precise 3D measurements. However, this approach is too expensive to scale up. The other extreme solution emphasizes cost effectiveness, which tries to use thrifty sensor setups, e.g. only using cameras to detect 3D objects in real-time [1, 8, 9, 23, 25, 26, 30, 32, 41, 43]. However, camera-only 3D detection is significantly and consistently worse than LiDAR-based detection in most scenarios [19]." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 304, + 498, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 498, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 498, + 547, + 714 + ], + "type": "text", + "content": "In this paper, we propose an orthogonal direction for improving camera-only 3D detection performances by introducing multi-agent collaborations. Hypothetically, empowered by advanced communication systems, multiple agents equipped only with cameras could share visual information with each other. This would bring three outstanding benefits. First, different viewpoints from multiple agents can largely resolve the depth ambiguity issue in camera-only 3D detection, bridging the gap with expensive LiDARs on depth estimation. Second, multi-agent collaboration avoids inevitable limitations in single-agent 3D detection, such as occlusion and long-range issues, and potentially enables more holistic 3D detection; that is, detecting all the objects existed in the 3D scene, including those beyond visual range. Since LiDAR also suffers from limited field of view, this potentially enables collaborative cameras to outperform LiDAR. Third, the total expense of a large fleet of vehicles is significantly reduced as cameras are much" + } + ] + } + ], + "index": 24 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 703, + 135, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 703, + 135, + 712 + ], + "spans": [ + { + "bbox": [ + 58, + 703, + 135, + 712 + ], + "type": "text", + "content": "*Corresponding author." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "9243" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 131 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 131 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 131 + ], + "type": "text", + "content": "cheaper than LiDAR. However, multi-agent collaboration also brings new challenges. Different from many multiview geometry problems, here we also have to concern communication bandwidth constraints. Thus, each agent needs to select the most informative cues to share." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 133, + 287, + 287 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 133, + 287, + 287 + ], + "spans": [ + { + "bbox": [ + 46, + 133, + 287, + 287 + ], + "type": "text", + "content": "Following this design rationale, we propose a novel collaborative camera-only 3D detection framework CoCa3D. It includes three parts: i) single-agent camera-only 3D detection, which achieves basic depth estimation and 3D detection for each agent; ii) collaborative depth estimation, which disambiguates the estimated depths by promoting spatial consistency across multiple agents' viewpoints; and iii) collaborative detection feature learning, which complements detection features by sharing key detection messages with each other. Compared to recent collaborative perception methods [11,14] that are dealing with LiDAR, CoCa3D specifically designs novel collaborative depth estimation to customize the task of camera-only 3D detection." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 288, + 287, + 455 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 288, + 287, + 455 + ], + "spans": [ + { + "bbox": [ + 46, + 288, + 287, + 455 + ], + "type": "text", + "content": "To evaluate CoCa3D, we conduct comprehensive experiments on one real-world dataset, DAIR-V2X [40], and two new simulation datasets, OPV2V+ and CoPerception-UAVs+, which are extended based on original OPV2V [37] and CoPerception-UAVs [6] with more collaborative agents that cover three types of agents (cars, infrastructures and drones). Our results show that i) with 10 collaborative agents, CoCa3D enables camera-only detectors to overtake LiDAR-based detectors on OPV2V+; and ii) CoCa3D consistently outperforms previous works in the performance-bandwidth trade-off across multiple datasets by a large margin, improving the previous SOTA performances by " + }, + { + "bbox": [ + 46, + 288, + 287, + 455 + ], + "type": "inline_equation", + "content": "30.60\\%" + }, + { + "bbox": [ + 46, + 288, + 287, + 455 + ], + "type": "text", + "content": " on OPV2V+, " + }, + { + "bbox": [ + 46, + 288, + 287, + 455 + ], + "type": "inline_equation", + "content": "12.59\\%" + }, + { + "bbox": [ + 46, + 288, + 287, + 455 + ], + "type": "text", + "content": " on CoPerception-UAVs+, " + }, + { + "bbox": [ + 46, + 288, + 287, + 455 + ], + "type": "inline_equation", + "content": "44.21\\%" + }, + { + "bbox": [ + 46, + 288, + 287, + 455 + ], + "type": "text", + "content": " on DAIR-V2X for AP@70. To sum up, our contributions are:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 455, + 287, + 645 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 46, + 455, + 287, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 455, + 287, + 502 + ], + "spans": [ + { + "bbox": [ + 46, + 455, + 287, + 502 + ], + "type": "text", + "content": "- We propose a novel collaborative camera-only 3D detection framework CoCa3D, which improves the detection ability of cameras with multi-agent collaboration, promoting more holistic 3D detection." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 503, + 287, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 503, + 287, + 574 + ], + "spans": [ + { + "bbox": [ + 46, + 503, + 287, + 574 + ], + "type": "text", + "content": "- We propose core communication-efficient collaboration techniques, which explore the spatially sparse yet critical depth messages and tackle the depth ambiguity, occlusion, and long-range issues by fusing complementary information from different viewpoints, achieving more accurate and complete 3D representation." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 574, + 287, + 645 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 574, + 287, + 645 + ], + "spans": [ + { + "bbox": [ + 46, + 574, + 287, + 645 + ], + "type": "text", + "content": "- We expand two previous collaborative datasets with more agents, and conduct extensive experiments, validating that i) CoCa3D significantly bridges the performance gap between camera and LiDAR on OPV2V+ and DAIR-V2X; and ii) CoCa3D achieves the state-of-the-art performance-bandwidth trade-off." + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 47, + 648, + 134, + 661 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 648, + 134, + 661 + ], + "spans": [ + { + "bbox": [ + 47, + 648, + 134, + 661 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "text", + "content": "Camera-only 3D object detection. Camera-only 3D detection aims to detect objects in the 3D space given the 2D image through explicit or implicit depth estimation [7, 9, 23, 26, 30, 41]. Recently, bird's-eye-view (BEV) repre" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 72, + 545, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 216 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 216 + ], + "type": "text", + "content": "sentations are widely used for their computation efficiency and comparable performance compared to the 3D voxel features [10]. To get the BEV features, there are two types of methods. The depth-based methods [8, 23, 25] first estimate the depth distribution, and then attentively project the 2D image features along the projection ray to get the 3D voxel features and collapse it to BEV feature. The query-based methods [16, 22, 43] first initialize queries for each BEV grid and then leverage transformer-architecture-based cross-attention to query image features with camera-aware positional embedding. Here, our single-agent camera-only detector follows the simple-yet-effective CaDDN [25]." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 217, + 545, + 397 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 217, + 545, + 397 + ], + "spans": [ + { + "bbox": [ + 304, + 217, + 545, + 397 + ], + "type": "text", + "content": "LiDAR 3D object detection. LiDAR-based 3D detection achieves excellent performance due to the precise 3D measurements of the input data. Two well-known ways to encode LiDAR points include voxel-based [10, 42, 45] and point-based [28,29,38]. Voxel-based methods divide the 3D space into regular voxels [45] or pillars [10], and encode the point inside into feature representations. Point-based methods are usually based on the PointNet [24] series to aggregate the feature of points. Then the point features will be used for the proposal refinement. LiDAR-based 3D detection performs well, but high-quality LiDAR is hard to be adopted in a large scale due to the expensive cost. Here we propose an economic camera-only solution by introducing multi-agent collaboration, whose performance can catch up with LiDAR given a sufficient number of agents." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 398, + 545, + 661 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 398, + 545, + 661 + ], + "spans": [ + { + "bbox": [ + 304, + 398, + 545, + 661 + ], + "type": "text", + "content": "Collaborative perception. Collaborative perception is an emerging application of multi-agent systems, which promotes perception performance by enabling agents to share information with other agents through communication. In this emerging research field, a surge of high-quality datasets, (V2X-SIM [13], OPV2V [37], DAIR-V2X [40], CoPerception-UAVs [6]), and collaborative methods [5,6,13-15,17,18,20,31,35-37,40] aimed for better performance-bandwidth trade-off have been proposed. Collaborative methods have three types: early, late and intermediate fusion. Compared with early fusion(i.e. transmitting raw data) and late fusion (i.e. transmitting outputs), intermediate fusion(i.e. transmitting and fusing features encoded by deep networks) achieves a better performance-bandwidth trade-off. DiscoNet [14] adopts knowledge distillation to take the advantage of both early and intermediate collaboration. Where2comm [6] optimizes communication efficiency by sharing spatially sparse yet perceptually critical information. However, previous works mainly focus on LiDAR-based 3D object detection, while we investigate collaborative camera-only detection and specifically enhance depth estimation via multi-agent collaboration." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 670, + 533, + 685 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 670, + 533, + 685 + ], + "spans": [ + { + "bbox": [ + 305, + 670, + 533, + 685 + ], + "type": "text", + "content": "3. Collaborative Camera-Only 3D Detection" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 689, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 689, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 305, + 689, + 545, + 714 + ], + "type": "text", + "content": "This section presents CoCa3D, collaborative camera-only 3D detection framework, which enables multiple" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "9244" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 69, + 293, + 141 + ], + "blocks": [ + { + "bbox": [ + 50, + 69, + 293, + 141 + ], + "lines": [ + { + "bbox": [ + 50, + 69, + 293, + 141 + ], + "spans": [ + { + "bbox": [ + 50, + 69, + 293, + 141 + ], + "type": "image", + "image_path": "ba90d29575aba984d580ef376a6ed1027979f5229376939c1bd197e709656b79.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 296, + 70, + 539, + 141 + ], + "blocks": [ + { + "bbox": [ + 296, + 70, + 539, + 141 + ], + "lines": [ + { + "bbox": [ + 296, + 70, + 539, + 141 + ], + "spans": [ + { + "bbox": [ + 296, + 70, + 539, + 141 + ], + "type": "image", + "image_path": "b813dba96905d50e5b54918027331eabb7e082785cbf05699461eb85e4259a2c.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 51, + 143, + 536, + 213 + ], + "blocks": [ + { + "bbox": [ + 51, + 143, + 536, + 213 + ], + "lines": [ + { + "bbox": [ + 51, + 143, + 536, + 213 + ], + "spans": [ + { + "bbox": [ + 51, + 143, + 536, + 213 + ], + "type": "image", + "image_path": "f4b230329c19b0c42f6a9374a9326edf0826269ac4099d74eddd26d4d45ced13.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 217, + 547, + 251 + ], + "lines": [ + { + "bbox": [ + 46, + 217, + 547, + 251 + ], + "spans": [ + { + "bbox": [ + 46, + 217, + 547, + 251 + ], + "type": "text", + "content": "Figure 2. System overview. CoCa3D is a camera-only 3D detector integrated with two collaboration modules. Collaborative depth estimation (Co-Depth) enhances the single-agent estimated depth to achieve more accurate 3D feature. Collaborative detection feature learning (Co-FL) complements the single-agent 3D feature to achieve more holistic 3D detection." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 257, + 288, + 281 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 257, + 288, + 281 + ], + "spans": [ + { + "bbox": [ + 46, + 257, + 288, + 281 + ], + "type": "text", + "content": "agents to share visual information with each other, promoting more holistic 3D detection." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 288, + 171, + 299 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 288, + 171, + 299 + ], + "spans": [ + { + "bbox": [ + 47, + 288, + 171, + 299 + ], + "type": "text", + "content": "3.1. Problem Formulation" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 302, + 287, + 358 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 302, + 287, + 358 + ], + "spans": [ + { + "bbox": [ + 46, + 302, + 287, + 358 + ], + "type": "text", + "content": "Consider " + }, + { + "bbox": [ + 46, + 302, + 287, + 358 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 302, + 287, + 358 + ], + "type": "text", + "content": " agents in the scene, let " + }, + { + "bbox": [ + 46, + 302, + 287, + 358 + ], + "type": "inline_equation", + "content": "\\mathcal{X}_i" + }, + { + "bbox": [ + 46, + 302, + 287, + 358 + ], + "type": "text", + "content": " be the RGB image collected by the " + }, + { + "bbox": [ + 46, + 302, + 287, + 358 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 302, + 287, + 358 + ], + "type": "text", + "content": "th agent and " + }, + { + "bbox": [ + 46, + 302, + 287, + 358 + ], + "type": "inline_equation", + "content": "\\mathcal{O}_i^0" + }, + { + "bbox": [ + 46, + 302, + 287, + 358 + ], + "type": "text", + "content": " be the corresponding ground-truth detection. The objective of CoCa3D is to maximize detection performances of all agents given certain communication budget " + }, + { + "bbox": [ + 46, + 302, + 287, + 358 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 46, + 302, + 287, + 358 + ], + "type": "text", + "content": "; that is," + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 364, + 286, + 388 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 364, + 286, + 388 + ], + "spans": [ + { + "bbox": [ + 47, + 364, + 286, + 388 + ], + "type": "interline_equation", + "content": "\\max _ {\\theta , \\mathcal {P}} \\sum_ {i} g \\left(\\Phi_ {\\theta} \\left(\\mathcal {X} _ {i}, \\{\\mathcal {P} _ {i \\rightarrow j} \\} _ {j = 1} ^ {N}\\right), \\mathcal {O} _ {i} ^ {0}\\right), \\mathrm {s . t .} \\sum_ {i} | \\mathcal {P} _ {i \\rightarrow j} | \\leq B,", + "image_path": "5594bae0040a2a2c3bf000b819194fbeb09ac5b90f48e2ca87fc2a207b1e3ffa.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 394, + 287, + 562 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 394, + 287, + 562 + ], + "spans": [ + { + "bbox": [ + 46, + 394, + 287, + 562 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 394, + 287, + 562 + ], + "type": "inline_equation", + "content": "g(\\cdot, \\cdot)" + }, + { + "bbox": [ + 46, + 394, + 287, + 562 + ], + "type": "text", + "content": " is the detection evaluation metric, " + }, + { + "bbox": [ + 46, + 394, + 287, + 562 + ], + "type": "inline_equation", + "content": "\\Phi(\\cdot)" + }, + { + "bbox": [ + 46, + 394, + 287, + 562 + ], + "type": "text", + "content": " is a detection model with trainable parameter " + }, + { + "bbox": [ + 46, + 394, + 287, + 562 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 46, + 394, + 287, + 562 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 394, + 287, + 562 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_{i \\to j}" + }, + { + "bbox": [ + 46, + 394, + 287, + 562 + ], + "type": "text", + "content": " is the message transmitted from the " + }, + { + "bbox": [ + 46, + 394, + 287, + 562 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 394, + 287, + 562 + ], + "type": "text", + "content": "th agent to the " + }, + { + "bbox": [ + 46, + 394, + 287, + 562 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 46, + 394, + 287, + 562 + ], + "type": "text", + "content": "th agent. The key challenge is to determine the messages " + }, + { + "bbox": [ + 46, + 394, + 287, + 562 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_{i \\to j}" + }, + { + "bbox": [ + 46, + 394, + 287, + 562 + ], + "type": "text", + "content": ", which should be both informative and compact. Our design rational comes from two aspects: First, since the major gap between camera and LiDAR is the depth, the message should include depth information. This would allow different viewpoints from multiple agents to disambiguate the infinite depth possibilities and localize the correct depth candidate. Second, the message should include detection clues to provide complementary detection information, which can fundamentally overcome inevitable limitations of single-agent detection, such as occlusion and long-range issues." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 563, + 287, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 563, + 287, + 635 + ], + "spans": [ + { + "bbox": [ + 46, + 563, + 287, + 635 + ], + "type": "text", + "content": "Based on the above intuition, the proposed CoCa3D includes a) single-agent camera-only 3D detection, which achieves basic depth estimation and detection ability, and b) multi-agent collaboration, which shares both estimated depth information and detection features to improve 3D representation and detection performances; see Fig. 2." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 638, + 251, + 651 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 638, + 251, + 651 + ], + "spans": [ + { + "bbox": [ + 47, + 638, + 251, + 651 + ], + "type": "text", + "content": "3.2. Single-agent camera-only 3D detection" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 654, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 654, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 654, + 288, + 715 + ], + "type": "text", + "content": "Single-agent camera-only 3D object detection network learns to detect 3D objects in the physical space based on the 2D camera inputs. Here we adopt CaDNN [25], an off-the-shelf architecture. The main idea is to project the flat image feature into all the possible depths in the 3D space," + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 257, + 545, + 316 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 257, + 545, + 316 + ], + "spans": [ + { + "bbox": [ + 304, + 257, + 545, + 316 + ], + "type": "text", + "content": "and then attentively aggregate the 3D voxel feature and collapse them into the BEV feature. To mention that, the BEV feature is used because it largely reduces the computation cost, while performing similarly to 3D voxel features [10]. The architecture consists of five modules as detailed below." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 319, + 545, + 367 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 319, + 545, + 367 + ], + "spans": [ + { + "bbox": [ + 304, + 319, + 545, + 367 + ], + "type": "text", + "content": "Encoder. The encoder extracts features from the input images. For agent " + }, + { + "bbox": [ + 304, + 319, + 545, + 367 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 319, + 545, + 367 + ], + "type": "text", + "content": ", given its input " + }, + { + "bbox": [ + 304, + 319, + 545, + 367 + ], + "type": "inline_equation", + "content": "\\mathcal{X}_i" + }, + { + "bbox": [ + 304, + 319, + 545, + 367 + ], + "type": "text", + "content": ", the output feature map of encoder is " + }, + { + "bbox": [ + 304, + 319, + 545, + 367 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_i = \\Phi_{\\mathrm{enc}}(\\mathcal{X}_i) \\in \\mathbb{R}^{H \\times W \\times C}" + }, + { + "bbox": [ + 304, + 319, + 545, + 367 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 319, + 545, + 367 + ], + "type": "inline_equation", + "content": "H, W, C" + }, + { + "bbox": [ + 304, + 319, + 545, + 367 + ], + "type": "text", + "content": " are the height, width and channel of the image feature." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 369, + 545, + 476 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 369, + 545, + 476 + ], + "spans": [ + { + "bbox": [ + 304, + 369, + 545, + 476 + ], + "type": "text", + "content": "Depth estimation. The depth estimation module predicts pixel-wise categorical depth distribution over a set of predefined depth bins to accurately locate image information in 3D space. Treating the estimation as a classification problem can capture the inherent depth estimation uncertainty to reduce the impact of erroneous depth estimates, as stated in CaDNN [25]. To achieve this, a parametric depth estimation network " + }, + { + "bbox": [ + 304, + 369, + 545, + 476 + ], + "type": "inline_equation", + "content": "\\Phi_{\\mathrm{Depth}}(\\cdot)" + }, + { + "bbox": [ + 304, + 369, + 545, + 476 + ], + "type": "text", + "content": " is used, and the pixel-wise categorical depth distribution is obtained as" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 359, + 477, + 545, + 491 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 359, + 477, + 545, + 491 + ], + "spans": [ + { + "bbox": [ + 359, + 477, + 545, + 491 + ], + "type": "interline_equation", + "content": "\\mathcal {D} _ {i} = \\Phi_ {\\text {D e p t h}} \\left(\\mathcal {F} _ {i}\\right) \\in \\mathbb {R} ^ {H \\times W \\times D}, \\tag {1}", + "image_path": "358c4ccf7523993fa1440faca80401c25772e584e2ac7ead996e984808367ca6.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 492, + 547, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 492, + 547, + 529 + ], + "spans": [ + { + "bbox": [ + 304, + 492, + 547, + 529 + ], + "type": "text", + "content": "whose " + }, + { + "bbox": [ + 304, + 492, + 547, + 529 + ], + "type": "inline_equation", + "content": "(h,w)" + }, + { + "bbox": [ + 304, + 492, + 547, + 529 + ], + "type": "text", + "content": " th element " + }, + { + "bbox": [ + 304, + 492, + 547, + 529 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_i(h,w)\\in \\mathbb{R}^D" + }, + { + "bbox": [ + 304, + 492, + 547, + 529 + ], + "type": "text", + "content": " refers to the estimated depth distribution of the " + }, + { + "bbox": [ + 304, + 492, + 547, + 529 + ], + "type": "inline_equation", + "content": "(h,w)" + }, + { + "bbox": [ + 304, + 492, + 547, + 529 + ], + "type": "text", + "content": " th image pixel with " + }, + { + "bbox": [ + 304, + 492, + 547, + 529 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 304, + 492, + 547, + 529 + ], + "type": "text", + "content": " as the number of discretized depth bins." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 531, + 545, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 531, + 545, + 651 + ], + "spans": [ + { + "bbox": [ + 304, + 531, + 545, + 651 + ], + "type": "text", + "content": "Voxel transformation. Voxel transformation projects the image feature to the 3D space, given all the possible depth candidates and the known image calibration matrix, resulting in a 3D voxel representation " + }, + { + "bbox": [ + 304, + 531, + 545, + 651 + ], + "type": "inline_equation", + "content": "\\widehat{\\mathcal{V}}_i\\in \\mathbb{R}^{X\\times Y\\times Z\\times C}" + }, + { + "bbox": [ + 304, + 531, + 545, + 651 + ], + "type": "text", + "content": ". The camera projection matrix " + }, + { + "bbox": [ + 304, + 531, + 545, + 651 + ], + "type": "inline_equation", + "content": "\\mathbf{P}\\in \\mathbb{R}^{3\\times 4}" + }, + { + "bbox": [ + 304, + 531, + 545, + 651 + ], + "type": "text", + "content": " defines the mapping between the 3D voxel coordinate to the image pixel coordinate, this is, " + }, + { + "bbox": [ + 304, + 531, + 545, + 651 + ], + "type": "inline_equation", + "content": "[u,v,d] = \\mathbf{P}\\cdot [x,y,z,1]" + }, + { + "bbox": [ + 304, + 531, + 545, + 651 + ], + "type": "text", + "content": ". Similarly, the depth probabilities can be transformed to the 3D voxel space, resulting in " + }, + { + "bbox": [ + 304, + 531, + 545, + 651 + ], + "type": "inline_equation", + "content": "\\widehat{\\mathcal{D}}_i\\in \\mathbb{R}^{X\\times Y\\times Z}" + }, + { + "bbox": [ + 304, + 531, + 545, + 651 + ], + "type": "text", + "content": ", where each element indicates the confidence that the feature pixel belongs to the voxel." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 654, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 547, + 715 + ], + "type": "text", + "content": "Collapse. The voxel feature is collapsed to a single height plane to generate the bird's-eye-view (BEV) feature as " + }, + { + "bbox": [ + 304, + 654, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\overline{B}_i = \\Phi_{\\mathrm{clp}}(\\widehat{\\mathcal{D}}_i \\odot \\widehat{\\mathcal{V}}_i) \\in \\mathbb{R}^{X \\times Y \\times C}" + }, + { + "bbox": [ + 304, + 654, + 547, + 715 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 654, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\odot" + }, + { + "bbox": [ + 304, + 654, + 547, + 715 + ], + "type": "text", + "content": " is the element-wise multiplication. Symbol " + }, + { + "bbox": [ + 304, + 654, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\cdot / \\widehat{\\cdot} / \\widehat{\\cdot}" + }, + { + "bbox": [ + 304, + 654, + 547, + 715 + ], + "type": "text", + "content": " reflects 2D image / 3D voxel / BEV spaces, respectively. The collapse network " + }, + { + "bbox": [ + 304, + 654, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\Phi_{\\mathrm{clp}}(\\cdot)" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "9245" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 286, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 286, + 95 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 286, + 95 + ], + "type": "text", + "content": "flattens the 3D voxel feature along the " + }, + { + "bbox": [ + 46, + 72, + 286, + 95 + ], + "type": "inline_equation", + "content": "Z" + }, + { + "bbox": [ + 46, + 72, + 286, + 95 + ], + "type": "text", + "content": "-axis and then applies " + }, + { + "bbox": [ + 46, + 72, + 286, + 95 + ], + "type": "inline_equation", + "content": "1 \\times 1" + }, + { + "bbox": [ + 46, + 72, + 286, + 95 + ], + "type": "text", + "content": " convolution to reduce the channel dimension." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 96, + 286, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 96, + 286, + 216 + ], + "spans": [ + { + "bbox": [ + 46, + 96, + 286, + 216 + ], + "type": "text", + "content": "Decoder. The decoder takes the BEV feature as input, and outputs the objects including class and regression. Here, we adopt CenterNet [44], an off-the-shelf detector. Given the BEV feature " + }, + { + "bbox": [ + 46, + 96, + 286, + 216 + ], + "type": "inline_equation", + "content": "\\overline{B}_i" + }, + { + "bbox": [ + 46, + 96, + 286, + 216 + ], + "type": "text", + "content": ", the detection decoder " + }, + { + "bbox": [ + 46, + 96, + 286, + 216 + ], + "type": "inline_equation", + "content": "\\Phi_{\\mathrm{dec}}(\\cdot)" + }, + { + "bbox": [ + 46, + 96, + 286, + 216 + ], + "type": "text", + "content": " generate the dense heatmap of " + }, + { + "bbox": [ + 46, + 96, + 286, + 216 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 96, + 286, + 216 + ], + "type": "text", + "content": "th agent by " + }, + { + "bbox": [ + 46, + 96, + 286, + 216 + ], + "type": "inline_equation", + "content": "\\overline{\\mathcal{O}}_i = \\Phi_{\\mathrm{dec}}(\\overline{B}_i) \\in \\mathbb{R}^{X \\times Y \\times 7}" + }, + { + "bbox": [ + 46, + 96, + 286, + 216 + ], + "type": "text", + "content": ", where each location of " + }, + { + "bbox": [ + 46, + 96, + 286, + 216 + ], + "type": "inline_equation", + "content": "\\overline{\\mathcal{O}}_i" + }, + { + "bbox": [ + 46, + 96, + 286, + 216 + ], + "type": "text", + "content": " represents a rotated box with class " + }, + { + "bbox": [ + 46, + 96, + 286, + 216 + ], + "type": "inline_equation", + "content": "(c, x, y, h, w, \\cos \\alpha, \\sin \\alpha)" + }, + { + "bbox": [ + 46, + 96, + 286, + 216 + ], + "type": "text", + "content": ", denoting class confidence, position residual, size and angle. Non-maximum suppression (NMS) is applied to the dense predictions and generate the final sparse output of the 3D detection system." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 218, + 286, + 337 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 218, + 286, + 337 + ], + "spans": [ + { + "bbox": [ + 46, + 218, + 286, + 337 + ], + "type": "text", + "content": "Discussion. Admittedly, single-agent camera-only 3D detection makes localizing objects in the 3D space given 2D image input possible through numerous ingenious algorithm designs. The 3D detection performance is still far from perfect due to the ambiguous single-agent depth estimation and limited visual range. Tackling these two issues are challenging for a single agent, but from a perspective of multi-agent collaboration, sharing complementary visual clues makes depth estimation and visual range enlargement much more natural. This motivates our following designs." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 342, + 190, + 354 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 342, + 190, + 354 + ], + "spans": [ + { + "bbox": [ + 47, + 342, + 190, + 354 + ], + "type": "text", + "content": "3.3. Multi-agent collaboration" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 357, + 286, + 440 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 357, + 286, + 440 + ], + "spans": [ + { + "bbox": [ + 46, + 357, + 286, + 440 + ], + "type": "text", + "content": "Multi-agent collaboration includes two parts: i) collaborative depth estimation, which enables the sharing of depth information estimated by individual agents to produce more accurate 3D representation and less biased BEV feature; and ii) collaborative detection feature learning, which enables the sharing of detection features of individual agents to achieve more holistic 3D detection." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 445, + 212, + 456 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 445, + 212, + 456 + ], + "spans": [ + { + "bbox": [ + 47, + 445, + 212, + 456 + ], + "type": "text", + "content": "3.3.1 Collaborative depth estimation" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 460, + 286, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 460, + 286, + 628 + ], + "spans": [ + { + "bbox": [ + 46, + 460, + 286, + 628 + ], + "type": "text", + "content": "Collaborative depth estimation (Co-Depth) targets to disambiguate infinite depth possibilities in single-agent camera-only depth estimation and localize the correct depth candidate through multi-view consistency. The intuition is, for a correct depth candidate, its corresponding 3D location should be spatially consistent from multiple agents' viewpoints. To achieve this, each agent can exchange the depth information through communication. Meanwhile, we promote communication efficiency by selecting the most critical and unambiguous depth information. Accordingly, Co-Depth includes: a) depth uncertainty aware message packing, which packs compact messages with unambiguous depth information; and b) depth message fusion, which enhances depth estimation with the received depth messages." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 630, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 630, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 630, + 286, + 713 + ], + "type": "text", + "content": "Depth uncertainty aware message packing. Depth uncertainty-aware (DUA) message packing packs the most critical depth information used for multi-view consistency into the to-be-sent message based on the depth uncertainty. The depth message includes: i) the voxel feature, which is used for the multi-view visual similarity measurement; and ii) the depth probability, which indicates the confidence that" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 72, + 545, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 545, + 95 + ], + "type": "text", + "content": "the feature pixel belongs to the voxel and is used for the multi-view candidate selection." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 96, + 545, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 96, + 545, + 144 + ], + "spans": [ + { + "bbox": [ + 304, + 96, + 545, + 144 + ], + "type": "text", + "content": "The spatially sparse features of " + }, + { + "bbox": [ + 304, + 96, + 545, + 144 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 96, + 545, + 144 + ], + "type": "text", + "content": "th agent are selected based on its spatial depth uncertainty map in the 2D image space " + }, + { + "bbox": [ + 304, + 96, + 545, + 144 + ], + "type": "inline_equation", + "content": "\\mathbf{U}_i\\in \\mathbb{R}^{H\\times W}" + }, + { + "bbox": [ + 304, + 96, + 545, + 144 + ], + "type": "text", + "content": ", that can be represented by the pixelwise entropy from corresponding depth distribution," + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 361, + 147, + 489, + 159 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 361, + 147, + 489, + 159 + ], + "spans": [ + { + "bbox": [ + 361, + 147, + 489, + 159 + ], + "type": "interline_equation", + "content": "\\mathbf {U} _ {i} (h, w) = \\mathrm {H} (\\mathcal {D} _ {i} (h, w)) \\in \\mathbb {R},", + "image_path": "702dc0ea18fac64807c81ecccbbd3843a90478feff3742607f987095eebe0477.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 162, + 545, + 281 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 162, + 545, + 281 + ], + "spans": [ + { + "bbox": [ + 304, + 162, + 545, + 281 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 162, + 545, + 281 + ], + "type": "inline_equation", + "content": "\\mathrm{H}(\\cdot)" + }, + { + "bbox": [ + 304, + 162, + 545, + 281 + ], + "type": "text", + "content": " is the entropy function and " + }, + { + "bbox": [ + 304, + 162, + 545, + 281 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_i(h,w)" + }, + { + "bbox": [ + 304, + 162, + 545, + 281 + ], + "type": "text", + "content": " is the depth distribution at the " + }, + { + "bbox": [ + 304, + 162, + 545, + 281 + ], + "type": "inline_equation", + "content": "(h,w)" + }, + { + "bbox": [ + 304, + 162, + 545, + 281 + ], + "type": "text", + "content": "th image pixel, following from (1). The intuition is, for the locations with low uncertainty score, an agent is confident about which depth bin the feature pixel belongs to. Thus, sharing information at these locations can help with depth estimation accuracy. On the contrary, for the locations with high uncertainty score, an agent is hard to tell which depth bin the feature pixel belongs to. Transmitting these spatial features would introduce ambiguity, and even degrade the single-agent depth estimation." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 281, + 545, + 342 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 281, + 545, + 342 + ], + "spans": [ + { + "bbox": [ + 304, + 281, + 545, + 342 + ], + "type": "text", + "content": "Specifically, a binary selection matrix is used to represent whether each location is selected or not, where 1 denotes selected, and 0 elsewhere. Let " + }, + { + "bbox": [ + 304, + 281, + 545, + 342 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_{i\\rightarrow j}^{d}\\in \\{0,1\\}^{H\\times W}" + }, + { + "bbox": [ + 304, + 281, + 545, + 342 + ], + "type": "text", + "content": " be a selection matrix to determine the depth message sent from agent " + }, + { + "bbox": [ + 304, + 281, + 545, + 342 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 281, + 545, + 342 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 304, + 281, + 545, + 342 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 304, + 281, + 545, + 342 + ], + "type": "text", + "content": ". Its " + }, + { + "bbox": [ + 304, + 281, + 545, + 342 + ], + "type": "inline_equation", + "content": "(h,w)" + }, + { + "bbox": [ + 304, + 281, + 545, + 342 + ], + "type": "text", + "content": "th element is" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 348, + 343, + 502, + 357 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 348, + 343, + 502, + 357 + ], + "spans": [ + { + "bbox": [ + 348, + 343, + 502, + 357 + ], + "type": "interline_equation", + "content": "\\mathbf {M} _ {i \\rightarrow j} ^ {d} (h, w) = I (\\mathbf {U} _ {i} (h, w) < u _ {\\text {t h r e}}),", + "image_path": "9f5248450497b896ac08593cd3706e17d99bdbf352aa907f621b037efd7903fb.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 359, + 545, + 394 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 359, + 545, + 394 + ], + "spans": [ + { + "bbox": [ + 304, + 359, + 545, + 394 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 359, + 545, + 394 + ], + "type": "inline_equation", + "content": "I(\\cdot)" + }, + { + "bbox": [ + 304, + 359, + 545, + 394 + ], + "type": "text", + "content": " is an indicator function and " + }, + { + "bbox": [ + 304, + 359, + 545, + 394 + ], + "type": "inline_equation", + "content": "u_{\\mathrm{thre}}" + }, + { + "bbox": [ + 304, + 359, + 545, + 394 + ], + "type": "text", + "content": " is a predefined threshold. This reflects that a specific spatial area will be selected when its depth uncertainty is below a threshold." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 395, + 545, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 395, + 545, + 449 + ], + "spans": [ + { + "bbox": [ + 304, + 395, + 545, + 449 + ], + "type": "text", + "content": "The selection matrix can be projected in the 3D voxel space, resulting in " + }, + { + "bbox": [ + 304, + 395, + 545, + 449 + ], + "type": "inline_equation", + "content": "\\widehat{\\mathbf{M}}_{i\\rightarrow j}^{d}\\in \\mathbb{R}^{X\\times Y\\times Z}" + }, + { + "bbox": [ + 304, + 395, + 545, + 449 + ], + "type": "text", + "content": ". Then, the selected voxel feature map " + }, + { + "bbox": [ + 304, + 395, + 545, + 449 + ], + "type": "inline_equation", + "content": "\\widehat{\\mathcal{Z}}_{i\\rightarrow j}" + }, + { + "bbox": [ + 304, + 395, + 545, + 449 + ], + "type": "text", + "content": " and the selected depth distribution " + }, + { + "bbox": [ + 304, + 395, + 545, + 449 + ], + "type": "inline_equation", + "content": "\\widehat{\\mathcal{E}}_{i\\rightarrow j}" + }, + { + "bbox": [ + 304, + 395, + 545, + 449 + ], + "type": "text", + "content": " sent from the " + }, + { + "bbox": [ + 304, + 395, + 545, + 449 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 395, + 545, + 449 + ], + "type": "text", + "content": "th agent to the " + }, + { + "bbox": [ + 304, + 395, + 545, + 449 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 304, + 395, + 545, + 449 + ], + "type": "text", + "content": "th agent are obtained as" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 349, + 450, + 501, + 484 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 349, + 450, + 501, + 484 + ], + "spans": [ + { + "bbox": [ + 349, + 450, + 501, + 484 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\widehat {\\mathcal {Z}} _ {i \\rightarrow j} = \\widehat {\\mathbf {M}} _ {i \\rightarrow j} ^ {d} \\odot \\widehat {\\mathcal {V}} _ {i} \\in \\mathbb {R} ^ {X \\times Y \\times Z \\times C}, \\\\ \\widehat {\\mathcal {E}} _ {i \\rightarrow j} = \\widehat {\\mathbf {M}} _ {i \\rightarrow j} ^ {d} \\odot \\widehat {\\mathcal {D}} _ {i} \\in \\mathbb {R} ^ {X \\times Y \\times Z}, \\\\ \\end{array}", + "image_path": "b1c4ec9da3e0f611bd372ceccf37e5882f8d1f65a471ac1ab25367b61bb02c28.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 487, + 545, + 536 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 487, + 545, + 536 + ], + "spans": [ + { + "bbox": [ + 304, + 487, + 545, + 536 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 487, + 545, + 536 + ], + "type": "inline_equation", + "content": "\\widehat{\\mathcal{V}}_i" + }, + { + "bbox": [ + 304, + 487, + 545, + 536 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 487, + 545, + 536 + ], + "type": "inline_equation", + "content": "\\widehat{\\mathcal{D}}_i" + }, + { + "bbox": [ + 304, + 487, + 545, + 536 + ], + "type": "text", + "content": " are the full 3D voxel feature map and depth distribution. Overall, the depth message sent from the " + }, + { + "bbox": [ + 304, + 487, + 545, + 536 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 487, + 545, + 536 + ], + "type": "text", + "content": "th agent to the " + }, + { + "bbox": [ + 304, + 487, + 545, + 536 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 304, + 487, + 545, + 536 + ], + "type": "text", + "content": "th agent is " + }, + { + "bbox": [ + 304, + 487, + 545, + 536 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_{i\\rightarrow j}^{d} = (\\widehat{\\mathcal{E}}_{i\\rightarrow j},\\widehat{\\mathcal{Z}}_{i\\rightarrow j})" + }, + { + "bbox": [ + 304, + 487, + 545, + 536 + ], + "type": "text", + "content": ", where both entries are spatially sparse and communication efficient." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 537, + 545, + 650 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 537, + 545, + 650 + ], + "spans": [ + { + "bbox": [ + 304, + 537, + 545, + 650 + ], + "type": "text", + "content": "Depth message fusion. Depth message fusion targets to enhance the depth estimation given the received depth messages through different viewpoints of multiple agents. The intuition is that for a correct depth candidate, the visual features at the same 3D point observed by multiple agents should be similar. To achieve this, we introduce multiview depth consistency weighting. Let " + }, + { + "bbox": [ + 304, + 537, + 545, + 650 + ], + "type": "inline_equation", + "content": "\\widehat{S}_i(x,y,z) \\in \\mathbb{R}" + }, + { + "bbox": [ + 304, + 537, + 545, + 650 + ], + "type": "text", + "content": " be the matching score between agent " + }, + { + "bbox": [ + 304, + 537, + 545, + 650 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 537, + 545, + 650 + ], + "type": "text", + "content": " and its neighbors " + }, + { + "bbox": [ + 304, + 537, + 545, + 650 + ], + "type": "inline_equation", + "content": "\\mathcal{N}_i" + }, + { + "bbox": [ + 304, + 537, + 545, + 650 + ], + "type": "text", + "content": ", those who share messages with current agent, at coordinate " + }, + { + "bbox": [ + 304, + 537, + 545, + 650 + ], + "type": "inline_equation", + "content": "(x,y,z)" + }, + { + "bbox": [ + 304, + 537, + 545, + 650 + ], + "type": "text", + "content": ". The multi-view matching score is obtained as" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 306, + 653, + 503, + 667 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 653, + 503, + 667 + ], + "spans": [ + { + "bbox": [ + 306, + 653, + 503, + 667 + ], + "type": "interline_equation", + "content": "\\widehat {\\mathcal {W}} _ {j \\rightarrow i} (x, y, z) = I (\\widehat {\\mathcal {E}} _ {j \\rightarrow i} (x, y, z) > p _ {\\text {t h r e}}) \\in \\{0, 1 \\},", + "image_path": "17bdfb488a6f0f8fedc94a4bf25bde81de3cd485058c3a1c151034f1ec9a455f.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 670, + 544, + 694 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 670, + 544, + 694 + ], + "spans": [ + { + "bbox": [ + 307, + 670, + 544, + 694 + ], + "type": "interline_equation", + "content": "\\widehat {\\mathcal {S}} _ {i} (x, y, z) = \\sum_ {j \\in \\mathcal {N} _ {i}} \\widehat {\\mathcal {W}} _ {j \\rightarrow i} (x, y, z) \\left\\langle \\widehat {\\mathcal {V}} _ {i} (x, y, z), \\widehat {\\mathcal {Z}} _ {j \\rightarrow i} (x, y, z) \\right\\rangle ,", + "image_path": "4d87713fb5baf5bc74a39331db25fb28759379655f96ecfdfea4bcfeb9679085.jpg" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 699, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 699, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 307, + 699, + 545, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 307, + 699, + 545, + 714 + ], + "type": "inline_equation", + "content": "\\widehat{\\mathcal{W}}_{j\\rightarrow i}(x,y,z)" + }, + { + "bbox": [ + 307, + 699, + 545, + 714 + ], + "type": "text", + "content": " is a binary weight used to filter out the" + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "9246" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 144 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 144 + ], + "type": "text", + "content": "neighboring view that has a low depth confidence, below a threshold " + }, + { + "bbox": [ + 46, + 72, + 287, + 144 + ], + "type": "inline_equation", + "content": "p_{\\mathrm{thre}}" + }, + { + "bbox": [ + 46, + 72, + 287, + 144 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 72, + 287, + 144 + ], + "type": "inline_equation", + "content": "\\langle \\cdot ,\\cdot \\rangle" + }, + { + "bbox": [ + 46, + 72, + 287, + 144 + ], + "type": "text", + "content": " is the inner product that measures the visual feature consistency across two views. This means that either the depth candidate is wrong or it is not visible in that view (e.g. due to occlusion), where the multi-view matching score should not be considered." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 144, + 287, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 144, + 287, + 168 + ], + "spans": [ + { + "bbox": [ + 47, + 144, + 287, + 168 + ], + "type": "text", + "content": "Given the multi-view matching score, the collaborative depth distribution is obtained as" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 82, + 168, + 252, + 189 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 82, + 168, + 252, + 189 + ], + "spans": [ + { + "bbox": [ + 82, + 168, + 252, + 189 + ], + "type": "interline_equation", + "content": "\\widehat {\\mathcal {D}} _ {i} ^ {\\prime} = \\Phi_ {\\mathrm {C o D e p t h}} \\left(\\left[ \\widehat {\\mathcal {D}} _ {i}; \\widehat {\\mathcal {S}} _ {i} \\right]\\right) \\in \\mathbb {R} ^ {X \\times Y \\times Z},", + "image_path": "5b759f068f674bfa9c407857212fe869f6ba4772e8371fa8dccaae1e8e8f9ef2.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 188, + 287, + 307 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 188, + 287, + 307 + ], + "spans": [ + { + "bbox": [ + 46, + 188, + 287, + 307 + ], + "type": "text", + "content": "where the " + }, + { + "bbox": [ + 46, + 188, + 287, + 307 + ], + "type": "inline_equation", + "content": "[\\cdot ]" + }, + { + "bbox": [ + 46, + 188, + 287, + 307 + ], + "type": "text", + "content": " denotes concatenation, " + }, + { + "bbox": [ + 46, + 188, + 287, + 307 + ], + "type": "inline_equation", + "content": "\\Phi_{\\mathrm{CoDepth}}(\\cdot)" + }, + { + "bbox": [ + 46, + 188, + 287, + 307 + ], + "type": "text", + "content": " is the collaborative depth estimation network implemented with " + }, + { + "bbox": [ + 46, + 188, + 287, + 307 + ], + "type": "inline_equation", + "content": "1\\times 1" + }, + { + "bbox": [ + 46, + 188, + 287, + 307 + ], + "type": "text", + "content": " convolutions followed by a sigmoid activation. Each element of the collaborative depth distribution " + }, + { + "bbox": [ + 46, + 188, + 287, + 307 + ], + "type": "inline_equation", + "content": "\\widehat{D}_i^{\\prime}" + }, + { + "bbox": [ + 46, + 188, + 287, + 307 + ], + "type": "text", + "content": " reflects the confidence of the feature locates in the corresponding voxel. Note that i) the collaboratively updated depth probability is decided on both the single-view confidence and the multiview consistency; ii) the depth consistency weighting discards the candidates with low single-view depth probability to avoid ambiguity and make the estimation more robust." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 308, + 287, + 369 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 308, + 287, + 369 + ], + "spans": [ + { + "bbox": [ + 47, + 308, + 287, + 369 + ], + "type": "text", + "content": "Given this collaborative depth distribution, each agent can weigh the 3D voxel feature and then collapse into a BEV feature with a collapse network " + }, + { + "bbox": [ + 47, + 308, + 287, + 369 + ], + "type": "inline_equation", + "content": "\\Phi_{\\mathrm{clp}}(\\cdot)" + }, + { + "bbox": [ + 47, + 308, + 287, + 369 + ], + "type": "text", + "content": ", this is, " + }, + { + "bbox": [ + 47, + 308, + 287, + 369 + ], + "type": "inline_equation", + "content": "\\overline{B}_i = \\Phi_{\\mathrm{clp}}(\\widehat{\\mathcal{D}}_i' \\odot \\widehat{\\mathcal{V}}_i) \\in \\mathbb{R}^{X \\times Y \\times C}" + }, + { + "bbox": [ + 47, + 308, + 287, + 369 + ], + "type": "text", + "content": ". The collapse network is implemented with " + }, + { + "bbox": [ + 47, + 308, + 287, + 369 + ], + "type": "inline_equation", + "content": "1 \\times 1" + }, + { + "bbox": [ + 47, + 308, + 287, + 369 + ], + "type": "text", + "content": " convolution to reduce channel dimension." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 373, + 250, + 385 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 373, + 250, + 385 + ], + "spans": [ + { + "bbox": [ + 47, + 373, + 250, + 385 + ], + "type": "text", + "content": "3.3.2 Collaborative detection feature learning" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 388, + 287, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 388, + 287, + 567 + ], + "spans": [ + { + "bbox": [ + 46, + 388, + 287, + 567 + ], + "type": "text", + "content": "Collaborative depth estimation carefully refines the depth and promotes more accurate 3D representations for each single agent. However, the physical limitations of single agent, such as, limited field of view, occlusion, and long-range issues still remain. To achieve more holistic 3D detection, each agent should be able to exchange 3D detection feature and make use of the complementary information. Meanwhile, we promote communication efficiency by selecting the most perceptually critical information. Therefore, collaborative detection feature learning (Co-FL) includes: a) detection confidence aware message packing, which packs spatially sparse yet perceptually critical 3D features with the guidance of the detection confidence; and ii) detection message fusion, which enhances the 3D feature with the received detection messages." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 569, + 287, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 569, + 287, + 700 + ], + "spans": [ + { + "bbox": [ + 46, + 569, + 287, + 700 + ], + "type": "text", + "content": "Detection confidence aware message packing. The detection confidence aware (DCA) message packing targets to pack the complementary perceptual information into a compact message. The core idea is to explore the spatial heterogeneity of perceptual information. The intuition is that the areas that contain objects are more critical than background areas. During collaboration, areas with objects could help recover the miss-detected objects due to the limited view; and background areas could be omitted to save the precious bandwidth. To achieve this, we implement the spatial confidence map with the detection confidence map by" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 113, + 700, + 219, + 714 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 113, + 700, + 219, + 714 + ], + "spans": [ + { + "bbox": [ + 113, + 700, + 219, + 714 + ], + "type": "interline_equation", + "content": "\\overline {{\\mathbf {C}}} _ {i} = \\Phi_ {\\mathrm {c l s}} (\\overline {{\\mathcal {B}}} _ {i}) \\in \\mathbb {R} ^ {X \\times Y},", + "image_path": "369508b2e5d0cd5e6287b542b84455f9c577d8989ccf34fd8cb0dc1af56f56ba.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "inline_equation", + "content": "\\Phi_{\\mathrm{cls}}(\\cdot)" + }, + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "text", + "content": " denotes the standard classification network in the single-agent decoder " + }, + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "inline_equation", + "content": "\\Phi_{\\mathrm{det}}(\\cdot), \\overline{B}_i" + }, + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "text", + "content": " is the BEV feature generated with the collaboratively estimated depth. Based on this map, agents decide which spatial area to communicate; that is, each agent offers spatially sparse, yet critical features to support other agents." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 144, + 545, + 195 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 144, + 545, + 195 + ], + "spans": [ + { + "bbox": [ + 304, + 144, + 545, + 195 + ], + "type": "text", + "content": "Specifically, a binary selection matrix is used to represent each location is selected or not. Let " + }, + { + "bbox": [ + 304, + 144, + 545, + 195 + ], + "type": "inline_equation", + "content": "\\overline{\\mathbf{M}}_{i\\rightarrow j}^{f}\\in" + }, + { + "bbox": [ + 304, + 144, + 545, + 195 + ], + "type": "inline_equation", + "content": "\\{0,1\\}^{X\\times Y}" + }, + { + "bbox": [ + 304, + 144, + 545, + 195 + ], + "type": "text", + "content": " be a selection matrix to determine the detection message sent from agent " + }, + { + "bbox": [ + 304, + 144, + 545, + 195 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 144, + 545, + 195 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 304, + 144, + 545, + 195 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 304, + 144, + 545, + 195 + ], + "type": "text", + "content": ". Its " + }, + { + "bbox": [ + 304, + 144, + 545, + 195 + ], + "type": "inline_equation", + "content": "(x,y)" + }, + { + "bbox": [ + 304, + 144, + 545, + 195 + ], + "type": "text", + "content": "th element is" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 351, + 196, + 499, + 213 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 351, + 196, + 499, + 213 + ], + "spans": [ + { + "bbox": [ + 351, + 196, + 499, + 213 + ], + "type": "interline_equation", + "content": "\\overline {{\\mathbf {M}}} _ {i \\rightarrow j} ^ {f} (x, y) = I (\\overline {{\\mathbf {C}}} _ {i} (x, y) > c _ {\\mathrm {t h r e}}),", + "image_path": "4d72e50627d9f64072e3a13bd4c058ac4cdf83cd06cde4385e977cc4fc2b6af9.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 213, + 545, + 299 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 213, + 545, + 299 + ], + "spans": [ + { + "bbox": [ + 304, + 213, + 545, + 299 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 213, + 545, + 299 + ], + "type": "inline_equation", + "content": "I(\\cdot)" + }, + { + "bbox": [ + 304, + 213, + 545, + 299 + ], + "type": "text", + "content": " is an indicator function and " + }, + { + "bbox": [ + 304, + 213, + 545, + 299 + ], + "type": "inline_equation", + "content": "c_{\\mathrm{thre}}" + }, + { + "bbox": [ + 304, + 213, + 545, + 299 + ], + "type": "text", + "content": " is a predefined threshold. This reflects that a specific spatial area will be selected when its detection confidence is above a threshold. Overall, the detection message sent from the " + }, + { + "bbox": [ + 304, + 213, + 545, + 299 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 213, + 545, + 299 + ], + "type": "text", + "content": "th agent to the " + }, + { + "bbox": [ + 304, + 213, + 545, + 299 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 304, + 213, + 545, + 299 + ], + "type": "text", + "content": "th agent is " + }, + { + "bbox": [ + 304, + 213, + 545, + 299 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_{i\\rightarrow j}^{f} = \\overline{\\mathbf{M}}_{i\\rightarrow j}^{f}\\odot \\overline{\\mathcal{B}_{i}}" + }, + { + "bbox": [ + 304, + 213, + 545, + 299 + ], + "type": "text", + "content": ". Only selected features and their indices are packed, and they can be recovered at the receiver. This greatly reduces communication cost." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 300, + 545, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 300, + 545, + 411 + ], + "spans": [ + { + "bbox": [ + 304, + 300, + 545, + 411 + ], + "type": "text", + "content": "Detection message fusion. Here we augment the detection feature of each agent by aggregating the received detection messages from other agents. We implement this with simple but effective non-parametric point-wise maximum fusion. Specifically, for the " + }, + { + "bbox": [ + 304, + 300, + 545, + 411 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 300, + 545, + 411 + ], + "type": "text", + "content": "th agent, after receiving the " + }, + { + "bbox": [ + 304, + 300, + 545, + 411 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 304, + 300, + 545, + 411 + ], + "type": "text", + "content": "th agent's message " + }, + { + "bbox": [ + 304, + 300, + 545, + 411 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_{j\\rightarrow i}^{f}" + }, + { + "bbox": [ + 304, + 300, + 545, + 411 + ], + "type": "text", + "content": ". We also include the ego feature map in fusion and denote " + }, + { + "bbox": [ + 304, + 300, + 545, + 411 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_{i\\rightarrow i}^{f} = \\overline{\\mathcal{B}}_{i}" + }, + { + "bbox": [ + 304, + 300, + 545, + 411 + ], + "type": "text", + "content": " to make the formulation simple and consistent, where " + }, + { + "bbox": [ + 304, + 300, + 545, + 411 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_{i\\rightarrow i}^{f}" + }, + { + "bbox": [ + 304, + 300, + 545, + 411 + ], + "type": "text", + "content": " might not be sparse. The fused BEV feature is obtained as" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 352, + 411, + 498, + 432 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 352, + 411, + 498, + 432 + ], + "spans": [ + { + "bbox": [ + 352, + 411, + 498, + 432 + ], + "type": "interline_equation", + "content": "\\overline {{\\mathcal {B}}} _ {i} ^ {\\prime} = \\max _ {j \\in \\mathcal {N} _ {i} \\cup \\{i \\}} (\\mathcal {P} _ {j \\rightarrow i} ^ {f}) \\in \\mathbb {R} ^ {X \\times Y \\times C},", + "image_path": "9a155017c61e09599289accb50830374a67cc93fbde48b04eaacdb94f481e232.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 433, + 545, + 516 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 433, + 545, + 516 + ], + "spans": [ + { + "bbox": [ + 304, + 433, + 545, + 516 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 433, + 545, + 516 + ], + "type": "inline_equation", + "content": "\\max (\\cdot)" + }, + { + "bbox": [ + 304, + 433, + 545, + 516 + ], + "type": "text", + "content": " maximizes the corresponding features from multiple agents at each individual spatial location. Note that attention fusion is not permutation invariant, as attention weights vary with the ordering of key and query. Here we simply use the max operator to avoid this permutation variant issue. The fused BEV feature is output to the decoder to generate the final detection." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 519, + 359, + 529 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 519, + 359, + 529 + ], + "spans": [ + { + "bbox": [ + 306, + 519, + 359, + 529 + ], + "type": "text", + "content": "3.4. Losses" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 531, + 545, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 531, + 545, + 604 + ], + "spans": [ + { + "bbox": [ + 304, + 531, + 545, + 604 + ], + "type": "text", + "content": "To train the overall system, we supervise two tasks: categorical depth estimation and 3D object detection. The overall loss is " + }, + { + "bbox": [ + 304, + 531, + 545, + 604 + ], + "type": "inline_equation", + "content": "L = L_{\\mathrm{dep}}(\\mathcal{D}_i,\\mathcal{D}_i^0) + L_{\\mathrm{det}}(\\overline{\\mathcal{O}}_i,\\mathcal{O}_i^0)" + }, + { + "bbox": [ + 304, + 531, + 545, + 604 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 531, + 545, + 604 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_i^0" + }, + { + "bbox": [ + 304, + 531, + 545, + 604 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 531, + 545, + 604 + ], + "type": "inline_equation", + "content": "\\mathcal{O}_i^0" + }, + { + "bbox": [ + 304, + 531, + 545, + 604 + ], + "type": "text", + "content": " is the " + }, + { + "bbox": [ + 304, + 531, + 545, + 604 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 531, + 545, + 604 + ], + "type": "text", + "content": "th agent's ground-truth depth category and objects, respectively, " + }, + { + "bbox": [ + 304, + 531, + 545, + 604 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{dep}}" + }, + { + "bbox": [ + 304, + 531, + 545, + 604 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 531, + 545, + 604 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{det}}" + }, + { + "bbox": [ + 304, + 531, + 545, + 604 + ], + "type": "text", + "content": " is the depth classification loss [25] and detection loss [44], respectively." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 305, + 609, + 432, + 622 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 609, + 432, + 622 + ], + "spans": [ + { + "bbox": [ + 305, + 609, + 432, + 622 + ], + "type": "text", + "content": "4. Experimental Results" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 306, + 626, + 491, + 639 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 626, + 491, + 639 + ], + "spans": [ + { + "bbox": [ + 306, + 626, + 491, + 639 + ], + "type": "text", + "content": "4.1. Datasets and experimental settings" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "type": "text", + "content": "We cover three datasets, both real-world and simulation scenarios, and multiple agent types. Metric Average Precision (AP) at Intersection-over-Union (IoU) threshold of 0.30, 0.50, 0.70, and 0.80 are used. Metric multi-class classification accuracy is used for depth accuracy. The communication volume follows the standard setting as [6, 36, 37]" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "9247" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 53, + 71, + 193, + 179 + ], + "blocks": [ + { + "bbox": [ + 53, + 71, + 193, + 179 + ], + "lines": [ + { + "bbox": [ + 53, + 71, + 193, + 179 + ], + "spans": [ + { + "bbox": [ + 53, + 71, + 193, + 179 + ], + "type": "image", + "image_path": "a2f60f240805255f8f1dc709609a6dca3813614148c32d2799170a0fe999670a.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 113, + 181, + 156, + 190 + ], + "lines": [ + { + "bbox": [ + 113, + 181, + 156, + 190 + ], + "spans": [ + { + "bbox": [ + 113, + 181, + 156, + 190 + ], + "type": "text", + "content": "(a) OPV2V+" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 215, + 72, + 353, + 180 + ], + "blocks": [ + { + "bbox": [ + 215, + 72, + 353, + 180 + ], + "lines": [ + { + "bbox": [ + 215, + 72, + 353, + 180 + ], + "spans": [ + { + "bbox": [ + 215, + 72, + 353, + 180 + ], + "type": "image", + "image_path": "f3107b653a6f4b821e989cb70177cb3af91bab71ec7271131755d18c51d9fe1e.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 270, + 181, + 321, + 190 + ], + "lines": [ + { + "bbox": [ + 270, + 181, + 321, + 190 + ], + "spans": [ + { + "bbox": [ + 270, + 181, + 321, + 190 + ], + "type": "text", + "content": "(b) DAIR-V2X" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 45, + 192, + 544, + 214 + ], + "lines": [ + { + "bbox": [ + 45, + 192, + 544, + 214 + ], + "spans": [ + { + "bbox": [ + 45, + 192, + 544, + 214 + ], + "type": "text", + "content": "Figure 3. CoCa3D steadily improves 3D detection performance as the number of agents grows. In OPV2V+, around 10 collaborative agents enable collaborative cameras to catch up with LiDAR." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 375, + 72, + 512, + 180 + ], + "blocks": [ + { + "bbox": [ + 375, + 72, + 512, + 180 + ], + "lines": [ + { + "bbox": [ + 375, + 72, + 512, + 180 + ], + "spans": [ + { + "bbox": [ + 375, + 72, + 512, + 180 + ], + "type": "image", + "image_path": "154529910b4e8f787067760ef14e99b5a99367e2c998e75248a8825d5add2866.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 415, + 181, + 497, + 191 + ], + "lines": [ + { + "bbox": [ + 415, + 181, + 497, + 191 + ], + "spans": [ + { + "bbox": [ + 415, + 181, + 497, + 191 + ], + "type": "text", + "content": "(c) CoPerception-UAVs+" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 48, + 237, + 555, + 360 + ], + "blocks": [ + { + "bbox": [ + 45, + 214, + 544, + 236 + ], + "lines": [ + { + "bbox": [ + 45, + 214, + 544, + 236 + ], + "spans": [ + { + "bbox": [ + 45, + 214, + 544, + 236 + ], + "type": "text", + "content": "Table 1. CoCa3D significantly outperforms the previous SOTAs, e.g. improve by " + }, + { + "bbox": [ + 45, + 214, + 544, + 236 + ], + "type": "inline_equation", + "content": "{30.60}\\%" + }, + { + "bbox": [ + 45, + 214, + 544, + 236 + ], + "type": "text", + "content": " on OPV2V, " + }, + { + "bbox": [ + 45, + 214, + 544, + 236 + ], + "type": "inline_equation", + "content": "{12.59}\\%" + }, + { + "bbox": [ + 45, + 214, + 544, + 236 + ], + "type": "text", + "content": " on CoPerception-UAVs, 44.21% on DAIR-V2X for metric AP@70." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 48, + 237, + 555, + 360 + ], + "lines": [ + { + "bbox": [ + 48, + 237, + 555, + 360 + ], + "spans": [ + { + "bbox": [ + 48, + 237, + 555, + 360 + ], + "type": "table", + "html": "
MethodOPV2V+CoPerception-UAVs+DAIR-V2X
AP@30AP@50AP@70AP@50AP@70AP@80AP@30AP@50AP@70
No Collaboration0.27480.20410.08530.69560.49000.23090.09770.05240.0305
Late Fusion0.65010.61980.51090.72060.53720.25970.20600.10780.0455
When2com (CVPR'20)0.48530.42110.37370.82190.67050.41020.19570.09840.0459
V2VNet (ECCV'20)0.62460.50420.38520.90930.71770.38040.16400.08470.0512
DiscoNet (NeurIPS'21)0.73000.60090.41790.90540.70790.35640.18360.12620.0683
V2X-ViT (ECCV'22)0.83460.66590.39460.90940.71430.35250.18620.10750.0490
Where2comm (NeurIPS'22)0.81910.70890.47410.91020.73830.36760.17540.10250.0547
CoCa3D0.86420.82600.66750.94970.85020.58350.35220.22600.0985
", + "image_path": "9bcb5a5e415b0c5aedb5056b36f10734438dd1105e1056b99f41f3a25dc574e6.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 364, + 287, + 400 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 364, + 287, + 400 + ], + "spans": [ + { + "bbox": [ + 46, + 364, + 287, + 400 + ], + "type": "text", + "content": "that counts the message size by byte in log scale with base 2. To compare communication results straightforwardly and fair, we do not consider any extra compression." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 414, + 287, + 522 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 414, + 287, + 522 + ], + "spans": [ + { + "bbox": [ + 46, + 414, + 287, + 522 + ], + "type": "text", + "content": "OPV2V+. The original OPV2V [37] is a large-scale vehicle-to-vehicle collaborative perception dataset, cosimulated by OpenCDA [34] and CARLA [4]. Here we introduce an extended version, OPV2V+, which includes more collaborative agents (" + }, + { + "bbox": [ + 46, + 414, + 287, + 522 + ], + "type": "inline_equation", + "content": "\\approx 10" + }, + { + "bbox": [ + 46, + 414, + 287, + 522 + ], + "type": "text", + "content": "). Each agent is equipt with 4 cameras and 4 depth sensors with resolution " + }, + { + "bbox": [ + 46, + 414, + 287, + 522 + ], + "type": "inline_equation", + "content": "600 \\times 800" + }, + { + "bbox": [ + 46, + 414, + 287, + 522 + ], + "type": "text", + "content": ". The detection range is " + }, + { + "bbox": [ + 46, + 414, + 287, + 522 + ], + "type": "inline_equation", + "content": "281.6\\mathrm{m} \\times 80\\mathrm{m}" + }, + { + "bbox": [ + 46, + 414, + 287, + 522 + ], + "type": "text", + "content": ". Our single-agent camera-only detector follows CaDDN [25] with 50 depth categories in linear-increasing spacing mode." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 536, + 287, + 596 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 536, + 287, + 596 + ], + "spans": [ + { + "bbox": [ + 46, + 536, + 287, + 596 + ], + "type": "text", + "content": "DAIR-V2X. DAIR-V2X [40] is the only public real-world collaborative perception dataset. It contains two agents: a vehicle and a road-side-unit with image resolution " + }, + { + "bbox": [ + 46, + 536, + 287, + 596 + ], + "type": "inline_equation", + "content": "1080 \\times 1920" + }, + { + "bbox": [ + 46, + 536, + 287, + 596 + ], + "type": "text", + "content": ". The perception range is " + }, + { + "bbox": [ + 46, + 536, + 287, + 596 + ], + "type": "inline_equation", + "content": "100\\mathrm{m} \\times 79\\mathrm{m}" + }, + { + "bbox": [ + 46, + 536, + 287, + 596 + ], + "type": "text", + "content": ". Our single-agent camera-only detector also follows CaDDN [25]." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 609, + 287, + 717 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 609, + 287, + 717 + ], + "spans": [ + { + "bbox": [ + 46, + 609, + 287, + 717 + ], + "type": "text", + "content": "CoPerception-UAVs+. The original CoPerception-UAVs [6] is a large-scale UAV-based collaborative perception dataset, co-simulated by AirSim [27] and CARLA [4]. Here we introduce an extended version, CoPerception-UAVs+, which includes more collaborative agents (" + }, + { + "bbox": [ + 46, + 609, + 287, + 717 + ], + "type": "inline_equation", + "content": "\\approx 10" + }, + { + "bbox": [ + 46, + 609, + 287, + 717 + ], + "type": "text", + "content": "). Each agent equips with 1 camera and 1 depth sensor with resolution " + }, + { + "bbox": [ + 46, + 609, + 287, + 717 + ], + "type": "inline_equation", + "content": "450 \\times 800" + }, + { + "bbox": [ + 46, + 609, + 287, + 717 + ], + "type": "text", + "content": ". The detection range is " + }, + { + "bbox": [ + 46, + 609, + 287, + 717 + ], + "type": "inline_equation", + "content": "192\\mathrm{m} \\times 96\\mathrm{m}" + }, + { + "bbox": [ + 46, + 609, + 287, + 717 + ], + "type": "text", + "content": ". Our single-agent camera-only detector follows DVDET [?] with 10 depth categories in uniform spacing mode." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 364, + 440, + 376 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 364, + 440, + 376 + ], + "spans": [ + { + "bbox": [ + 306, + 364, + 440, + 376 + ], + "type": "text", + "content": "4.2. Quantitative evaluation" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 384, + 545, + 648 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 384, + 545, + 648 + ], + "spans": [ + { + "bbox": [ + 304, + 384, + 545, + 648 + ], + "type": "text", + "content": "Collaborative camera-only 3D detection overtakes Li-DAR. Fig. 3 investigates how the collaborative camera-only 3D detection performance changes with the number of collaborative agents on multiple datasets. To set up a goal, we consider a LiDAR-based detector implemented by the widely used PointPillar [10]. Fig. 3 (a) shows that: i) with 10 collaborative agents, CoCa3D outperforms the LiDAR 3D detection at both AP@0.5/0.7 on OPV2V+! ii) the detection performance positively increases as the number of collaborative agents. And we then use the slope of OPV2V+ to fit the function of collaborative detection performance with agent number on DAIR-V2X as it only has two agents available. Fig. 3 (b) shows that camera-only 3D detection is expected to outperform LiDAR 3D detection with 7 collaborative vehicles in the real scenario. As LiDAR is not a commonly available equipment for drones, we set a high-quality RGB-D camera instead. Fig. 3 (c) shows that with 3 collaborative drones, camera-only 3D detection outperforms ground-truth depth-based 3D detection. Furthermore, the steadily increasing detection performance with the number of collaborative agents encourages the agents to actively collaborate and achieve consistent improvement." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 654, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 546, + 713 + ], + "type": "text", + "content": "Benchmark comparison. Tab. 1 compares the proposed CoCa3D with previous collaborative methods. We consider single-agent detection without collaboration (No collaboration), When2com [17], V2VNet [31], DiscoNet [14], V2XViT [36], Where2comm [6] and late fusion, where agents" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "9248" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 47, + 95, + 293, + 177 + ], + "blocks": [ + { + "bbox": [ + 47, + 71, + 287, + 92 + ], + "lines": [ + { + "bbox": [ + 47, + 71, + 287, + 92 + ], + "spans": [ + { + "bbox": [ + 47, + 71, + 287, + 92 + ], + "type": "text", + "content": "Table 2. CoCa3D outperforms single-agent camera-only 3D detection with GT depth on OPV2V+ and CoPerception-UAVs+." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 47, + 95, + 293, + 177 + ], + "lines": [ + { + "bbox": [ + 47, + 95, + 293, + 177 + ], + "spans": [ + { + "bbox": [ + 47, + 95, + 293, + 177 + ], + "type": "table", + "html": "
Co -DepthCo -FLOPV2V+CoPerception-UAVs+
AP@30AP@50AP@70AP@50AP@70AP@80
--0.27480.20410.08530.72130.54210.2846
GT-0.34540.25530.09730.83470.67640.4120
-✓0.82010.71910.47560.90840.72560.4028
GT✓0.91200.88050.74340.95050.83980.5504
✓✓0.86420.82600.66750.94950.85180.5849
", + "image_path": "574773ea293cbae2c7aa1e882b024d3a824a3bd276803cc97f71241ea290af98.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 190, + 287, + 310 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 190, + 287, + 310 + ], + "spans": [ + { + "bbox": [ + 46, + 190, + 287, + 310 + ], + "type": "text", + "content": "directly exchange the detected 3D boxes. We see that the proposed CoCa3D significantly outperforms previous state-of-the-arts, improves the SOTA performance by " + }, + { + "bbox": [ + 46, + 190, + 287, + 310 + ], + "type": "inline_equation", + "content": "30.60\\%" + }, + { + "bbox": [ + 46, + 190, + 287, + 310 + ], + "type": "text", + "content": " on OPV2V+, " + }, + { + "bbox": [ + 46, + 190, + 287, + 310 + ], + "type": "inline_equation", + "content": "12.59\\%" + }, + { + "bbox": [ + 46, + 190, + 287, + 310 + ], + "type": "text", + "content": " on CoPerception-UAVs+, " + }, + { + "bbox": [ + 46, + 190, + 287, + 310 + ], + "type": "inline_equation", + "content": "44.21\\%" + }, + { + "bbox": [ + 46, + 190, + 287, + 310 + ], + "type": "text", + "content": " on DAIR-V2X for AP@70. The reason is that collaborative depth estimation promotes more accurate 3D features per single agent, and the enhanced single agent features further facilitate collaborative detection feature learning, while previous collaborative methods do not specifically consider depth ambiguity and collaborate over single-agent features." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 312, + 287, + 431 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 312, + 287, + 431 + ], + "spans": [ + { + "bbox": [ + 46, + 312, + 287, + 431 + ], + "type": "text", + "content": "Multi-agent collaboration evaluation. Tab. 2 assess the effectiveness of the proposed multi-agent collaboration on OPV2V+ and CoPerception-UAVs+ datasets. Ground-truth (GT) depth is included to provide the upper bound of collaborative camera-only 3D detection. We see that: i) both collaborations can consistently improve the performance; ii) camera-only 3D detection with collaboratively estimated depth significantly outperforms that with single-agent estimated depth and even approaches the upper bound with ground truth depth." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 441, + 173, + 453 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 441, + 173, + 453 + ], + "spans": [ + { + "bbox": [ + 47, + 441, + 173, + 453 + ], + "type": "text", + "content": "4.3. Qualitative evaluation" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 461, + 287, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 461, + 287, + 628 + ], + "spans": [ + { + "bbox": [ + 46, + 461, + 287, + 628 + ], + "type": "text", + "content": "Visualization of depth and uncertainty. Fig. 5 shows that collaborative depth estimation outperforms single-agent depth estimation and approaches the ground truth depth. We see that: i) the single-agent depth estimation can estimate the relative depth while not able to precisely localize the depth candidate, for example, the vehicles are higher than the plane it sits while this plane is not grounded to the right category; ii) by introducing multi-view geometry, the collaboratively estimated depth can smoothly and accurately ground the plane; iii) the depth uncertainty is larger for the long-range and background areas. The reason is that the distant regions are hard to localize as they occupy too few image pixels, and the background regions are hard to localize due to the texture-less surfaces." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 630, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 630, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 630, + 287, + 714 + ], + "type": "text", + "content": "Visualization of detection results. Fig. 6 shows that compared to single LiDAR 3D detection, cameras with collaborations (Co-Depth and Co-FL) are able to achieve holistic and accurate detection results. The reason is that the single agent has some fundamental physical limitations such as long-range and occlusion issues. Fig. 6(a-c) shows that augmenting a single agent sensor going from just a cam" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 306, + 70, + 425, + 180 + ], + "blocks": [ + { + "bbox": [ + 306, + 70, + 425, + 180 + ], + "lines": [ + { + "bbox": [ + 306, + 70, + 425, + 180 + ], + "spans": [ + { + "bbox": [ + 306, + 70, + 425, + 180 + ], + "type": "image", + "image_path": "7777d30890ed305547aa55805548455d59034ed95194532e5a35d0e5ef057308.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 345, + 182, + 388, + 192 + ], + "lines": [ + { + "bbox": [ + 345, + 182, + 388, + 192 + ], + "spans": [ + { + "bbox": [ + 345, + 182, + 388, + 192 + ], + "type": "text", + "content": "(a) OPV2V+" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 305, + 194, + 545, + 215 + ], + "lines": [ + { + "bbox": [ + 305, + 194, + 545, + 215 + ], + "spans": [ + { + "bbox": [ + 305, + 194, + 545, + 215 + ], + "type": "text", + "content": "Figure 4. CoCa3D achieves superior detection performance and communication cost trade-off over various bandwidths." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 425, + 71, + 542, + 180 + ], + "blocks": [ + { + "bbox": [ + 425, + 71, + 542, + 180 + ], + "lines": [ + { + "bbox": [ + 425, + 71, + 542, + 180 + ], + "spans": [ + { + "bbox": [ + 425, + 71, + 542, + 180 + ], + "type": "image", + "image_path": "1efed90c2d1174b3e405e6cdf2e94a7b0bc072f0c7b1c2f2cef2af5e12efc3f6.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 443, + 182, + 525, + 192 + ], + "lines": [ + { + "bbox": [ + 443, + 182, + 525, + 192 + ], + "spans": [ + { + "bbox": [ + 443, + 182, + 525, + 192 + ], + "type": "text", + "content": "(b) CoPerception-UAVs+" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 306, + 235, + 545, + 271 + ], + "blocks": [ + { + "bbox": [ + 320, + 220, + 530, + 232 + ], + "lines": [ + { + "bbox": [ + 320, + 220, + 530, + 232 + ], + "spans": [ + { + "bbox": [ + 320, + 220, + 530, + 232 + ], + "type": "text", + "content": "Table 3. Co-Depth significantly improves depth accuracy." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 306, + 235, + 545, + 271 + ], + "lines": [ + { + "bbox": [ + 306, + 235, + 545, + 271 + ], + "spans": [ + { + "bbox": [ + 306, + 235, + 545, + 271 + ], + "type": "table", + "html": "
Full planeForeground
SingleCollaborationGainSingleCollaborationGain
0.61670.778126.17% ↑0.75860.81337.21% ↑
", + "image_path": "62d0dea4fa7874fb7307635b70843e8d2e34a53bd310416a6ae123446825300f.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_body" + } + ], + "index": 13 + }, + { + "type": "table", + "bbox": [ + 324, + 326, + 526, + 390 + ], + "blocks": [ + { + "bbox": [ + 305, + 279, + 545, + 323 + ], + "lines": [ + { + "bbox": [ + 305, + 279, + 545, + 323 + ], + "spans": [ + { + "bbox": [ + 305, + 279, + 545, + 323 + ], + "type": "text", + "content": "Table 4. Ablation of collaborative depth estimation on CoPerception-UAVs+. Uniform/Linear denotes uniform/linear-increasing spacing. Dense and sparse supervision denote applying depth supervision over the full image plane and the object regions." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 324, + 326, + 526, + 390 + ], + "lines": [ + { + "bbox": [ + 324, + 326, + 526, + 390 + ], + "spans": [ + { + "bbox": [ + 324, + 326, + 526, + 390 + ], + "type": "table", + "html": "
SpacingSupervisionAP@50AP@70AP@80
UniformDense0.95020.84440.5746
UniformSparse0.93300.80420.4829
LinearDense0.94840.83650.5580
LinearSparse0.94950.85180.5849
", + "image_path": "a930e6a802daf33f229a965a9d75cdc29f045c19d499cc53549b75a6b9f7f407.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "table_body" + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 403, + 545, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 403, + 545, + 510 + ], + "spans": [ + { + "bbox": [ + 304, + 403, + 545, + 510 + ], + "type": "text", + "content": "era to including a depth sensor to LiDAR can improve the detection but still not able to achieve holistic detection. Fig. 6(d)/(b) and (a)/(e) compares the detection results with and without collaborative feature learning, we see that collaboration can help detect lots of the missed objects in the single-agent detection. Fig. 6(d,e,f) show that: i) both collaborations, Co-FL and Co-Depth, consistently improve the detection performance; ii) with collaboration, collaborative camera-only agents can outperform LiDAR in 3D detection." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 517, + 404, + 529 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 517, + 404, + 529 + ], + "spans": [ + { + "bbox": [ + 306, + 517, + 404, + 529 + ], + "type": "text", + "content": "4.4. Ablation studies" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 534, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 534, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 534, + 545, + 713 + ], + "type": "text", + "content": "Trade-off between detection performance and communication cost. Fig. 4 compares the proposed CoCa3D with the previous communication-efficient solution Where2comm [6] in terms of the trade-off between detection performance (AP@IoU=50/70) and communication bandwidth. We see that: i) CoCa3D can adapt to varying bandwidths by adjusting the depth uncertainty threshold and the detection confidence threshold; ii) CoCa3D achieves superior detection performance and communication cost trade-off to Where2comm over varying communication bandwidth. The gain mainly comes from Co-Depth's improvement in depth estimation, which improves single-agent BEV features. With enhanced single BEV features, the Co-FL can generate better-augmented BEV features, resulting in improved detection performance." + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "9249" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 51, + 88, + 142, + 135 + ], + "blocks": [ + { + "bbox": [ + 51, + 88, + 142, + 135 + ], + "lines": [ + { + "bbox": [ + 51, + 88, + 142, + 135 + ], + "spans": [ + { + "bbox": [ + 51, + 88, + 142, + 135 + ], + "type": "image", + "image_path": "49de5e98a900e8abb045db6097aa3993efcc7c528e71b8f4b90dcf6750adbe55.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 71, + 149, + 121, + 159 + ], + "lines": [ + { + "bbox": [ + 71, + 149, + 121, + 159 + ], + "spans": [ + { + "bbox": [ + 71, + 149, + 121, + 159 + ], + "type": "text", + "content": "(a) Image (RV)" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 150, + 78, + 235, + 145 + ], + "blocks": [ + { + "bbox": [ + 150, + 78, + 235, + 145 + ], + "lines": [ + { + "bbox": [ + 150, + 78, + 235, + 145 + ], + "spans": [ + { + "bbox": [ + 150, + 78, + 235, + 145 + ], + "type": "image", + "image_path": "3fc22249c1fb463024213745e99f77430f96b8744179701430810a2aca1e2fee.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 171, + 149, + 222, + 159 + ], + "lines": [ + { + "bbox": [ + 171, + 149, + 222, + 159 + ], + "spans": [ + { + "bbox": [ + 171, + 149, + 222, + 159 + ], + "type": "text", + "content": "(b) Single (RV)" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 250, + 78, + 336, + 145 + ], + "blocks": [ + { + "bbox": [ + 250, + 78, + 336, + 145 + ], + "lines": [ + { + "bbox": [ + 250, + 78, + 336, + 145 + ], + "spans": [ + { + "bbox": [ + 250, + 78, + 336, + 145 + ], + "type": "image", + "image_path": "d746ffc473f0c0594175c27fed6605a5b25699246da947bda10e7d22896aef5d.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 260, + 149, + 334, + 159 + ], + "lines": [ + { + "bbox": [ + 260, + 149, + 334, + 159 + ], + "spans": [ + { + "bbox": [ + 260, + 149, + 334, + 159 + ], + "type": "text", + "content": "(c)Collaboration (RV)" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 351, + 78, + 436, + 145 + ], + "blocks": [ + { + "bbox": [ + 351, + 78, + 436, + 145 + ], + "lines": [ + { + "bbox": [ + 351, + 78, + 436, + 145 + ], + "spans": [ + { + "bbox": [ + 351, + 78, + 436, + 145 + ], + "type": "image", + "image_path": "4019488b619a7896b4c0c67e40c332734d8f2995f3050e79fa04461aa19a7502.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 361, + 149, + 433, + 159 + ], + "lines": [ + { + "bbox": [ + 361, + 149, + 433, + 159 + ], + "spans": [ + { + "bbox": [ + 361, + 149, + 433, + 159 + ], + "type": "text", + "content": "(d) Ground-truth (RV)" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 451, + 78, + 539, + 146 + ], + "blocks": [ + { + "bbox": [ + 451, + 78, + 539, + 146 + ], + "lines": [ + { + "bbox": [ + 451, + 78, + 539, + 146 + ], + "spans": [ + { + "bbox": [ + 451, + 78, + 539, + 146 + ], + "type": "image", + "image_path": "aa6856ffbdc0eeea699590483143f3d64ca6696b8600dfa42d15a6686aed6910.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 463, + 149, + 531, + 159 + ], + "lines": [ + { + "bbox": [ + 463, + 149, + 531, + 159 + ], + "spans": [ + { + "bbox": [ + 463, + 149, + 531, + 159 + ], + "type": "text", + "content": "(e) Uncertainty (RV)" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 51, + 172, + 142, + 220 + ], + "blocks": [ + { + "bbox": [ + 51, + 172, + 142, + 220 + ], + "lines": [ + { + "bbox": [ + 51, + 172, + 142, + 220 + ], + "spans": [ + { + "bbox": [ + 51, + 172, + 142, + 220 + ], + "type": "image", + "image_path": "73672d6740f21637bb69e8ea5101c0fb0aa14f4de5ff80c10b926cd92fc006cd.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 69, + 234, + 124, + 243 + ], + "lines": [ + { + "bbox": [ + 69, + 234, + 124, + 243 + ], + "spans": [ + { + "bbox": [ + 69, + 234, + 124, + 243 + ], + "type": "text", + "content": "(f) Image (BEV)" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 150, + 162, + 235, + 228 + ], + "blocks": [ + { + "bbox": [ + 150, + 162, + 235, + 228 + ], + "lines": [ + { + "bbox": [ + 150, + 162, + 235, + 228 + ], + "spans": [ + { + "bbox": [ + 150, + 162, + 235, + 228 + ], + "type": "image", + "image_path": "e6291efcddbb4e23c71b20dfe063e7e8970380bdd63dc36cd19e54b3479b9c38.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 168, + 234, + 224, + 243 + ], + "lines": [ + { + "bbox": [ + 168, + 234, + 224, + 243 + ], + "spans": [ + { + "bbox": [ + 168, + 234, + 224, + 243 + ], + "type": "text", + "content": "(g) Single (BEV)" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 250, + 163, + 335, + 228 + ], + "blocks": [ + { + "bbox": [ + 250, + 163, + 335, + 228 + ], + "lines": [ + { + "bbox": [ + 250, + 163, + 335, + 228 + ], + "spans": [ + { + "bbox": [ + 250, + 163, + 335, + 228 + ], + "type": "image", + "image_path": "4f1144a47c1d8e63a11af41658ae55578cb7fafc149bdc84ba85a13e5c349c11.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 255, + 234, + 335, + 243 + ], + "lines": [ + { + "bbox": [ + 255, + 234, + 335, + 243 + ], + "spans": [ + { + "bbox": [ + 255, + 234, + 335, + 243 + ], + "type": "text", + "content": "(h) Collaboration (BEV)" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 349, + 163, + 435, + 228 + ], + "blocks": [ + { + "bbox": [ + 349, + 163, + 435, + 228 + ], + "lines": [ + { + "bbox": [ + 349, + 163, + 435, + 228 + ], + "spans": [ + { + "bbox": [ + 349, + 163, + 435, + 228 + ], + "type": "image", + "image_path": "de294883f4a13cb7c4b41047ad8ae3ba2eba029777cac267532ca40dedffa61f.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 357, + 234, + 433, + 243 + ], + "lines": [ + { + "bbox": [ + 357, + 234, + 433, + 243 + ], + "spans": [ + { + "bbox": [ + 357, + 234, + 433, + 243 + ], + "type": "text", + "content": "(i) Ground-truth (BEV)" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 449, + 163, + 538, + 229 + ], + "blocks": [ + { + "bbox": [ + 449, + 163, + 538, + 229 + ], + "lines": [ + { + "bbox": [ + 449, + 163, + 538, + 229 + ], + "spans": [ + { + "bbox": [ + 449, + 163, + 538, + 229 + ], + "type": "image", + "image_path": "4fe9ba96c50e05d317ae2fab159eeda1156aff9ea17b892d99293ef81a1c6764.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 459, + 234, + 531, + 243 + ], + "lines": [ + { + "bbox": [ + 459, + 234, + 531, + 243 + ], + "spans": [ + { + "bbox": [ + 459, + 234, + 531, + 243 + ], + "type": "text", + "content": "(j) Uncertainty (BEV)" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 48, + 272, + 209, + 337 + ], + "blocks": [ + { + "bbox": [ + 46, + 247, + 544, + 270 + ], + "lines": [ + { + "bbox": [ + 46, + 247, + 544, + 270 + ], + "spans": [ + { + "bbox": [ + 46, + 247, + 544, + 270 + ], + "type": "text", + "content": "Figure 5. Visualization of the depth and uncertainty in image range-view (RV) and bird's eye view (BEV) on CoPerception-UAVs+. Collaboratively estimated depth improves the single-agent estimated depth and approaches ground-truth depth." + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 48, + 272, + 209, + 337 + ], + "lines": [ + { + "bbox": [ + 48, + 272, + 209, + 337 + ], + "spans": [ + { + "bbox": [ + 48, + 272, + 209, + 337 + ], + "type": "image", + "image_path": "c2bb5012670046e2025e884429e6cf073212ba3da5ea009e528deea4069c0f0f.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 216, + 272, + 376, + 337 + ], + "blocks": [ + { + "bbox": [ + 216, + 272, + 376, + 337 + ], + "lines": [ + { + "bbox": [ + 216, + 272, + 376, + 337 + ], + "spans": [ + { + "bbox": [ + 216, + 272, + 376, + 337 + ], + "type": "image", + "image_path": "ce287aecd05eb7304316a8822cb786807001242aeda91ee6e0a410610d4acbed.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 278, + 339, + 316, + 348 + ], + "lines": [ + { + "bbox": [ + 278, + 339, + 316, + 348 + ], + "spans": [ + { + "bbox": [ + 278, + 339, + 316, + 348 + ], + "type": "text", + "content": "(b) RGB-D" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 384, + 272, + 545, + 337 + ], + "blocks": [ + { + "bbox": [ + 384, + 272, + 545, + 337 + ], + "lines": [ + { + "bbox": [ + 384, + 272, + 545, + 337 + ], + "spans": [ + { + "bbox": [ + 384, + 272, + 545, + 337 + ], + "type": "image", + "image_path": "81df8c719b4e4cc6d7e17501777c517f84540672e18322fc4ad5defc851ca1e6.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 447, + 339, + 483, + 348 + ], + "lines": [ + { + "bbox": [ + 447, + 339, + 483, + 348 + ], + "spans": [ + { + "bbox": [ + 447, + 339, + 483, + 348 + ], + "type": "text", + "content": "(c) LiDAR" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_caption" + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 48, + 350, + 208, + 415 + ], + "blocks": [ + { + "bbox": [ + 110, + 339, + 147, + 348 + ], + "lines": [ + { + "bbox": [ + 110, + 339, + 147, + 348 + ], + "spans": [ + { + "bbox": [ + 110, + 339, + 147, + 348 + ], + "type": "text", + "content": "(a) Camera" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 48, + 350, + 208, + 415 + ], + "lines": [ + { + "bbox": [ + 48, + 350, + 208, + 415 + ], + "spans": [ + { + "bbox": [ + 48, + 350, + 208, + 415 + ], + "type": "image", + "image_path": "c1b171869492ab1e4755a397d005d66d30997625f0473ecdb2f25f353a344e20.jpg" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 90, + 416, + 167, + 425 + ], + "lines": [ + { + "bbox": [ + 90, + 416, + 167, + 425 + ], + "spans": [ + { + "bbox": [ + 90, + 416, + 167, + 425 + ], + "type": "text", + "content": "(d) Camera with Co-FL" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 430, + 544, + 441 + ], + "lines": [ + { + "bbox": [ + 46, + 430, + 544, + 441 + ], + "spans": [ + { + "bbox": [ + 46, + 430, + 544, + 441 + ], + "type": "text", + "content": "Figure 6. CoCa3D outperforms LiDAR detection on OPV2V+ with 10 agents. Green and red boxes denote GT and detection respectively." + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_caption" + } + ], + "index": 27 + }, + { + "type": "image", + "bbox": [ + 215, + 350, + 375, + 415 + ], + "blocks": [ + { + "bbox": [ + 215, + 350, + 375, + 415 + ], + "lines": [ + { + "bbox": [ + 215, + 350, + 375, + 415 + ], + "spans": [ + { + "bbox": [ + 215, + 350, + 375, + 415 + ], + "type": "image", + "image_path": "8e6cd7224aaea17acc63f899b115ab304f18b4fc78917d02a20460a5064e3d1b.jpg" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 257, + 416, + 334, + 426 + ], + "lines": [ + { + "bbox": [ + 257, + 416, + 334, + 426 + ], + "spans": [ + { + "bbox": [ + 257, + 416, + 334, + 426 + ], + "type": "text", + "content": "(e) RGB-D with Co-FL" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_caption" + } + ], + "index": 29 + }, + { + "type": "image", + "bbox": [ + 382, + 350, + 543, + 415 + ], + "blocks": [ + { + "bbox": [ + 382, + 350, + 543, + 415 + ], + "lines": [ + { + "bbox": [ + 382, + 350, + 543, + 415 + ], + "spans": [ + { + "bbox": [ + 382, + 350, + 543, + 415 + ], + "type": "image", + "image_path": "a9af25cbb9ff2f40e35bed9a78b359d4dee4cb964cd8837399df92d88a24c7a5.jpg" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 383, + 416, + 542, + 426 + ], + "lines": [ + { + "bbox": [ + 383, + 416, + 542, + 426 + ], + "spans": [ + { + "bbox": [ + 383, + 416, + 542, + 426 + ], + "type": "text", + "content": "(f) Camera with Co-Depth and Co-FL (CoCa3D)" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_caption" + } + ], + "index": 31 + }, + { + "bbox": [ + 46, + 445, + 287, + 649 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 445, + 287, + 649 + ], + "spans": [ + { + "bbox": [ + 46, + 445, + 287, + 649 + ], + "type": "text", + "content": "Effect of components in collaborative depth estimation. Tab. 3 assesses the gain of collaborative depth estimation (Co-Depth) over single-agent depth estimation (S-Depth) in the depth accuracy metric. We see that: i) Co-Depth steadily improves S-Depth over the full plane; ii) the depth accuracy of the foreground objects is higher than the full plane, as the texture-less backgrounds have fewer cues to localize. Tab. 4 assesses the effects of the depth spacing and supervision choices. We see that: i) overall, the Co-Depth is robust to the different spacing and supervision choices on AP@50/70; ii) linear-increasing spacing is robust to the various supervision choices that uniform spacing which fails the stricter metric AP@80 without dense supervision. The reason is that the linearly increasing spacing takes into account the prior depth distribution and assigns more depth candidates to frequently occurring depth ranges, which contributes to more balanced and easier depth learning." + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 47, + 666, + 200, + 679 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 666, + 200, + 679 + ], + "spans": [ + { + "bbox": [ + 47, + 666, + 200, + 679 + ], + "type": "text", + "content": "5. Conclusion and Limitation" + } + ] + } + ], + "index": 35 + }, + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "content": "We propose CoCa3D, a novel collaborative camera-only 3D detection that approaches holistic 3D detection. The" + } + ] + } + ], + "index": 36 + }, + { + "bbox": [ + 304, + 445, + 545, + 564 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 445, + 545, + 564 + ], + "spans": [ + { + "bbox": [ + 304, + 445, + 545, + 564 + ], + "type": "text", + "content": "core idea is to introduce multi-agent collaboration to improve the detection ability of cameras. Meanwhile, the communication cost is optimized, and each agent carefully selects the spatially sparse yet critical depth messages to share. Extensive experiments covering both real-world and simulation scenarios, and multi-type agents (cars, drones, and infrastructures) show that CoCa3D not only achieves state-of-the-art perception-bandwidth trade-off, but overtakes LiDAR-based detectors with a sufficient number of collaborative agents on OPV2V+." + } + ] + } + ], + "index": 37 + }, + { + "bbox": [ + 304, + 565, + 545, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 565, + 545, + 647 + ], + "spans": [ + { + "bbox": [ + 304, + 565, + 545, + 647 + ], + "type": "text", + "content": "Limitation and future work. It is expensive to collect a real-world multi-agent perception dataset. So far, DAIR-V2X is the sole public real-world dataset, which only has one vehicle and one roadside unit. This work mainly leverages simulation data to validate the proposed novel methods and sketch a promising research direction. We advocate more resources for real-world data collection." + } + ] + } + ], + "index": 38 + }, + { + "bbox": [ + 304, + 648, + 545, + 708 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 648, + 545, + 708 + ], + "spans": [ + { + "bbox": [ + 304, + 648, + 545, + 708 + ], + "type": "text", + "content": "Acknowledgment. This research is supported by the National Key RD Program of China under Grant 2021ZD0112801, NSFC under Grant 62171276 and the Science and Technology Commission of Shanghai Municipal under Grant 21511100900 and 22DZ2229005." + } + ] + } + ], + "index": 39 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "9250" + } + ] + } + ], + "index": 40 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 135 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 135 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 135 + ], + "type": "text", + "content": "[1] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In European conference on computer vision, pages 213-229. Springer, 2020. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 136, + 288, + 190 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 136, + 288, + 190 + ], + "spans": [ + { + "bbox": [ + 53, + 136, + 288, + 190 + ], + "type": "text", + "content": "[2] Siheng Chen, Baoan Liu, Chen Feng, Carlos Vallespi-Gonzalez, and Carl K. Wellington. 3d point cloud processing and learning for autonomous driving: Impacting map creation, localization, and perception. IEEE Signal Processing Magazine, 38:68-86, 2021. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 191, + 287, + 246 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 191, + 287, + 246 + ], + "spans": [ + { + "bbox": [ + 53, + 191, + 287, + 246 + ], + "type": "text", + "content": "[3] Jiajun Deng, Shaoshuai Shi, Peiwei Li, Wengang Zhou, Yanyong Zhang, and Houqiang Li. Voxel r-cnn: Towards high performance voxel-based 3d object detection. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 35, pages 1201-1209, 2021. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 247, + 287, + 289 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 247, + 287, + 289 + ], + "spans": [ + { + "bbox": [ + 53, + 247, + 287, + 289 + ], + "type": "text", + "content": "[4] Alexey Dosovitskiy, German Ros, Felipe Codevilla, Antonio Lopez, and Vladlen Koltun. Carla: An open urban driving simulator. In Conference on robot learning, pages 1-16. PMLR, 2017. 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 291, + 287, + 334 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 291, + 287, + 334 + ], + "spans": [ + { + "bbox": [ + 53, + 291, + 287, + 334 + ], + "type": "text", + "content": "[5] Peng Gao, Rui Guo, Hongsheng Lu, and Hao Zhang. Regularized graph matching for correspondence identification under uncertainty in collaborative perception. Robotics: Science and Systems XVI, 2020. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 335, + 287, + 378 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 335, + 287, + 378 + ], + "spans": [ + { + "bbox": [ + 53, + 335, + 287, + 378 + ], + "type": "text", + "content": "[6] Yue Hu, Shaoheng Fang, Zixing Lei, Yiqi Zhong, and Siheng Chen. Where2comm: Communication-efficient collaborative perception via spatial confidence maps. Advances in Neural Information Processing Systems, 2022. 2, 5, 6, 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 380, + 287, + 412 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 380, + 287, + 412 + ], + "spans": [ + { + "bbox": [ + 53, + 380, + 287, + 412 + ], + "type": "text", + "content": "[7] Yue Hu, Shaoheng Fang, Weidi Xie, and Siheng Chen. Aerial monocular 3d object detection. IEEE Robotics and Automation Letters, 8:1959-1966, 2022. 1, 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 414, + 287, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 414, + 287, + 445 + ], + "spans": [ + { + "bbox": [ + 53, + 414, + 287, + 445 + ], + "type": "text", + "content": "[8] Junjie Huang and Guan Huang. Bevdet4d: Exploit temporal cues in multi-camera 3d object detection. arXiv preprint arXiv:2203.17054, 2022. 1, 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 447, + 287, + 489 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 447, + 287, + 489 + ], + "spans": [ + { + "bbox": [ + 53, + 447, + 287, + 489 + ], + "type": "text", + "content": "[9] Junjie Huang, Guan Huang, Zheng Zhu, and Dalong Du. Bevdet: High-performance multi-camera 3d object detection in bird-eye-view. arXiv preprint arXiv:2112.11790, 2021.1, 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 491, + 287, + 545 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 491, + 287, + 545 + ], + "spans": [ + { + "bbox": [ + 48, + 491, + 287, + 545 + ], + "type": "text", + "content": "[10] Alex H Lang, Sourabh Vora, Holger Caesar, Lubing Zhou, Jiong Yang, and Oscar Beijbom. Pointpillars: Fast encoders for object detection from point clouds. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12697-12705, 2019. 1, 2, 3, 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 547, + 287, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 547, + 287, + 578 + ], + "spans": [ + { + "bbox": [ + 48, + 547, + 287, + 578 + ], + "type": "text", + "content": "[11] Zixing Lei, Shunli Ren, Yue Hu, Wenjun Zhang, and Siheng Chen. Latency-aware collaborative perception. ECCV, 2022. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 580, + 287, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 580, + 287, + 623 + ], + "spans": [ + { + "bbox": [ + 48, + 580, + 287, + 623 + ], + "type": "text", + "content": "[12] Jing Li, Rui Li, Jiehao Li, Junzheng Wang, Qingbin Wu, and Xu Liu. Dual-view 3d object recognition and detection via lidar point cloud and camera image. Robotics Auton. Syst., 150:103999, 2022. 1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 624, + 287, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 624, + 287, + 668 + ], + "spans": [ + { + "bbox": [ + 48, + 624, + 287, + 668 + ], + "type": "text", + "content": "[13] Yiming Li, Dekun Ma, Ziyan An, Zixun Wang, Yiqi Zhong, Siheng Chen, and Chen Feng. V2x-sim: Multi-agent collaborative perception dataset and benchmark for autonomous driving. IEEE Robotics and Automation Letters, 2022. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "type": "text", + "content": "[14] Yiming Li, Shunli Ren, Pengxiang Wu, Siheng Chen, Chen Feng, and Wenjun Zhang. Learning distilled collaboration graph for multi-agent perception. Advances in Neural Information Processing Systems, 34:29541-29552, 2021. 2, 6" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "text", + "content": "[15] Yiming Li, Juexiao Zhang, Dekun Ma, Yue Wang, and Chen Feng. Multi-robot scene completion: Towards task-agnostic collaborative perception. In Conference on Robot Learning, 2022. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 118, + 545, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 545, + 172 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 545, + 172 + ], + "type": "text", + "content": "[16] Zhiqi Li, Wenhai Wang, Hongyang Li, Enze Xie, Chonghao Sima, Tong Lu, Qiao Yu, and Jifeng Dai. Bevformer: Learning bird's-eye-view representation from multi-camera images via spatiotemporal transformers. arXiv preprint arXiv:2203.17270, 2022. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 175, + 545, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 175, + 545, + 228 + ], + "spans": [ + { + "bbox": [ + 307, + 175, + 545, + 228 + ], + "type": "text", + "content": "[17] Yen-Cheng Liu, Junjiao Tian, Nathaniel Glaser, and Zsolt Kira. When2com: Multi-agent perception via communication graph grouping. In Proceedings of the IEEE/CVF Conference on computer vision and pattern recognition, pages 4106-4115, 2020. 2, 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 230, + 545, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 230, + 545, + 285 + ], + "spans": [ + { + "bbox": [ + 308, + 230, + 545, + 285 + ], + "type": "text", + "content": "[18] Yen-Cheng Liu, Junjiao Tian, Chih-Yao Ma, Nathan Glaser, Chia-Wen Kuo, and Zsolt Kira. Who2com: Collaborative perception via learnable handshake communication. In 2020 IEEE International Conference on Robotics and Automation (ICRA), pages 6876-6883. IEEE, 2020. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 286, + 545, + 329 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 286, + 545, + 329 + ], + "spans": [ + { + "bbox": [ + 308, + 286, + 545, + 329 + ], + "type": "text", + "content": "[19] Zhijian Liu, Haotian Tang, Alexander Amini, Xinyu Yang, Huizi Mao, Daniela Rus, and Song Han. Bevfusion: Multi-task multi-sensor fusion with unified bird's-eye view representation. ArXiv, abs/2205.13542, 2022. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 331, + 545, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 331, + 545, + 373 + ], + "spans": [ + { + "bbox": [ + 308, + 331, + 545, + 373 + ], + "type": "text", + "content": "[20] Yifan Lu, Quanhao Li, Baoan Liu, Mehrdad Dianati, Chen Feng, Siheng Chen, and Yanfeng Wang. Robust collaborative 3d object detection in presence of pose errors. ICRA, 2023. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 376, + 545, + 419 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 376, + 545, + 419 + ], + "spans": [ + { + "bbox": [ + 308, + 376, + 545, + 419 + ], + "type": "text", + "content": "[21] Jiageng Mao, Minzhe Niu, Haoyue Bai, Xiaodan Liang, Hang Xu, and Chunjing Xu. Pyramid r-cnn: Towards better performance and adaptability for 3d object detection. ICCV, 2021. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 421, + 545, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 421, + 545, + 464 + ], + "spans": [ + { + "bbox": [ + 308, + 421, + 545, + 464 + ], + "type": "text", + "content": "[22] Lang Peng, Zhirong Chen, Zhangjie Fu, Pengpeng Liang, and Erkang Cheng. Bevsegformer: Bird's eye view semantic segmentation from arbitrary camera rigs. arXiv preprint arXiv:2203.04050, 2022. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 466, + 545, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 466, + 545, + 510 + ], + "spans": [ + { + "bbox": [ + 308, + 466, + 545, + 510 + ], + "type": "text", + "content": "[23] Jonah Philion and Sanja Fidler. Lift, splat, shoot: Encoding images from arbitrary camera rigs by implicitly unprojecting to 3d. In European Conference on Computer Vision, pages 194-210. Springer, 2020. 1, 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 511, + 545, + 564 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 511, + 545, + 564 + ], + "spans": [ + { + "bbox": [ + 308, + 511, + 545, + 564 + ], + "type": "text", + "content": "[24] Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. Pointnet: Deep learning on point sets for 3d classification and segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 652-660, 2017. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 567, + 545, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 567, + 545, + 610 + ], + "spans": [ + { + "bbox": [ + 308, + 567, + 545, + 610 + ], + "type": "text", + "content": "[25] Cody Reading, Ali Harakeh, Julia Chae, and Steven L. Waslander. Categorical depth distribution network for monocular 3d object detection. CVPR, 2021. 1, 2, 3, 5, 6" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 612, + 545, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 612, + 545, + 644 + ], + "spans": [ + { + "bbox": [ + 308, + 612, + 545, + 644 + ], + "type": "text", + "content": "[26] Thomas Roddick, Alex Kendall, and Roberto Cipolla. Orthographic feature transform for monocular 3d object detection. arXiv preprint arXiv:1811.08188, 2018. 1, 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 647, + 545, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 647, + 545, + 690 + ], + "spans": [ + { + "bbox": [ + 308, + 647, + 545, + 690 + ], + "type": "text", + "content": "[27] Shital Shah, Debadeepta Dey, Chris Lovett, and Ashish Kapoor. Airsim: High-fidelity visual and physical simulation for autonomous vehicles. In Field and service robotics, pages 621-635. Springer, 2018. 6" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 308, + 692, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 692, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 692, + 545, + 713 + ], + "type": "text", + "content": "[28] Shaoshuai Shi, Chaoxu Guo, Li Jiang, Zhe Wang, Jianping Shi, Xiaogang Wang, and Hongsheng Li. Pv-rcnn: Point" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "9251" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 66, + 72, + 287, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 72, + 287, + 106 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 287, + 106 + ], + "type": "text", + "content": "voxel feature set abstraction for 3d object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10529-10538, 2020. 1, 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 106, + 287, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 106, + 287, + 160 + ], + "spans": [ + { + "bbox": [ + 48, + 106, + 287, + 160 + ], + "type": "text", + "content": "[29] Shaoshuai Shi, Li Jiang, Jiajun Deng, Zhe Wang, Chaoxu Guo, Jianping Shi, Xiaogang Wang, and Hongsheng Li. Pvrconn++: Point-voxel feature set abstraction with local vector representation for 3d object detection. arXiv preprint arXiv:2102.00463, 2021. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 161, + 287, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 161, + 287, + 195 + ], + "spans": [ + { + "bbox": [ + 48, + 161, + 287, + 195 + ], + "type": "text", + "content": "[30] Tai Wang, Jiangmiao Pang, and Dahua Lin. Monocular 3d object detection with depth from motion. In European Conference on Computer Vision (ECCV), 2022. 1, 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 195, + 287, + 250 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 195, + 287, + 250 + ], + "spans": [ + { + "bbox": [ + 48, + 195, + 287, + 250 + ], + "type": "text", + "content": "[31] Tsun-Hsuan Wang, Sivabalan Manivasagam, Ming Liang, Bin Yang, Wenyuan Zeng, and Raquel Urtasun. V2vnet: Vehicle-to-vehicle communication for joint perception and prediction. In European Conference on Computer Vision, pages 605–621. Springer, 2020. 2, 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 250, + 287, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 250, + 287, + 304 + ], + "spans": [ + { + "bbox": [ + 48, + 250, + 287, + 304 + ], + "type": "text", + "content": "[32] Yue Wang, Vitor Campagnolo Guizilini, Tianyuan Zhang, Yilun Wang, Hang Zhao, and Justin Solomon. Detr3d: 3d object detection from multi-view images via 3d-to-2d queries. In Conference on Robot Learning, pages 180–191. PMLR, 2022. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 304, + 287, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 304, + 287, + 360 + ], + "spans": [ + { + "bbox": [ + 48, + 304, + 287, + 360 + ], + "type": "text", + "content": "[33] Pengxiang Wu, Siheng Chen, and Dimitris N. Metaxas. Motionnet: Joint perception and motion prediction for autonomous driving based on bird's eye view maps. 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11382-11392, 2020. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 360, + 287, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 360, + 287, + 415 + ], + "spans": [ + { + "bbox": [ + 48, + 360, + 287, + 415 + ], + "type": "text", + "content": "[34] Runsheng Xu, Yi Guo, Xu Han, Xin Xia, Hao Xiang, and Jiaqi Ma. Opencda: an open cooperative driving automation framework integrated with co-simulation. In 2021 IEEE International Intelligent Transportation Systems Conference (ITSC), pages 1155-1162. IEEE, 2021. 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 415, + 287, + 458 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 415, + 287, + 458 + ], + "spans": [ + { + "bbox": [ + 48, + 415, + 287, + 458 + ], + "type": "text", + "content": "[35] Runsheng Xu, Zhengzhong Tu, Hao Xiang, Wei Shao, Bolei Zhou, and Jiaqi Ma. CoBEVT: Cooperative bird's eye view semantic segmentation with sparse transformers. CoRL, 2022. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 459, + 287, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 459, + 287, + 525 + ], + "spans": [ + { + "bbox": [ + 48, + 459, + 287, + 525 + ], + "type": "text", + "content": "[36] Runsheng Xu, Hao Xiang, Zhengzhong Tu, Xin Xia, Ming-Hsuan Yang, and Jiaqi Ma. V2x-vit: Vehicle-to-everything cooperative perception with vision transformer. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXXIX, pages 107-124. Springer, 2022. 2, 5, 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 525, + 287, + 581 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 525, + 287, + 581 + ], + "spans": [ + { + "bbox": [ + 48, + 525, + 287, + 581 + ], + "type": "text", + "content": "[37] Runsheng Xu, Hao Xiang, Xin Xia, Xu Han, Jinlong Li, and Jiaqi Ma. Opv2v: An open benchmark dataset and fusion pipeline for perception with vehicle-to-vehicle communication. In 2022 International Conference on Robotics and Automation (ICRA), pages 2583–2589. IEEE, 2022. 2, 5, 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 581, + 287, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 581, + 287, + 624 + ], + "spans": [ + { + "bbox": [ + 48, + 581, + 287, + 624 + ], + "type": "text", + "content": "[38] Zetong Yang, Yanan Sun, Shu Liu, and Jiaya Jia. 3dssd: Point-based 3d single stage object detector. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11040-11048, 2020. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 625, + 287, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 625, + 287, + 669 + ], + "spans": [ + { + "bbox": [ + 48, + 625, + 287, + 669 + ], + "type": "text", + "content": "[39] Tianwei Yin, Xingyi Zhou, and Philipp Krahenbuhl. Center-based 3d object detection and tracking. 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11779-11788, 2021. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 670, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 670, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 670, + 287, + 713 + ], + "type": "text", + "content": "[40] Haibao Yu, Yizhen Luo, Mao Shu, Yiyi Huo, Zebang Yang, Yifeng Shi, Zhenglong Guo, Hanyu Li, Xing Hu, Jirui Yuan, and Zaiqing Nie. Dair-v2x: A large-scale dataset for vehicleinfrastructure cooperative 3d object detection. In IEEE/CVF" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 72, + 547, + 330 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 326, + 72, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 72, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 326, + 72, + 545, + 95 + ], + "type": "text", + "content": "Conf. on Computer Vision and Pattern Recognition (CVPR), June 2022. 2, 6" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 95, + 545, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 95, + 545, + 149 + ], + "spans": [ + { + "bbox": [ + 307, + 95, + 545, + 149 + ], + "type": "text", + "content": "[41] Yunpeng Zhang, Zheng Zhu, Wenzhao Zheng, Junjie Huang, Guan Huang, Jie Zhou, and Jiwen Lu. Beverse: Unified perception and prediction in birds-eye-view for vision-centric autonomous driving. arXiv preprint arXiv:2205.09743, 2022. 1, 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 150, + 545, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 150, + 545, + 206 + ], + "spans": [ + { + "bbox": [ + 307, + 150, + 545, + 206 + ], + "type": "text", + "content": "[42] Wu Zheng, Weiliang Tang, Sijin Chen, Li Jiang, and Chi-Wing Fu. Cia-ssd: Confident iou-aware single-stage object detector from point cloud. In Proceedings of the AAAI conference on artificial intelligence, volume 35, pages 3555–3562, 2021. 1, 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 207, + 545, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 207, + 545, + 251 + ], + "spans": [ + { + "bbox": [ + 307, + 207, + 545, + 251 + ], + "type": "text", + "content": "[43] Brady Zhou and Philipp Kähenbuhl. Cross-view transformers for real-time map-view semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13760-13769, 2022. 1, 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 252, + 545, + 284 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 252, + 545, + 284 + ], + "spans": [ + { + "bbox": [ + 307, + 252, + 545, + 284 + ], + "type": "text", + "content": "[44] Xingyi Zhou, Dequan Wang, and Philipp Krahenbuhl. Objects as points. In arXiv preprint arXiv:1904.07850, 2019. 4, 5" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 285, + 547, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 285, + 547, + 330 + ], + "spans": [ + { + "bbox": [ + 308, + 285, + 547, + 330 + ], + "type": "text", + "content": "[45] Yin Zhou and Oncel Tuzel. Voxelnet: End-to-end learning for point cloud based 3d object detection. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4490-4499, 2018. 1, 2" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "9252" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Collaborative Diffusion for Multi-Modal Face Generation and Editing/3b7ffa39-2244-45b9-9833-f768a1b5b8c6_content_list.json b/2023/Collaborative Diffusion for Multi-Modal Face Generation and Editing/3b7ffa39-2244-45b9-9833-f768a1b5b8c6_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..2d961cfd1f906c26f9781d3274d043b77f14ccde --- /dev/null +++ b/2023/Collaborative Diffusion for Multi-Modal Face Generation and Editing/3b7ffa39-2244-45b9-9833-f768a1b5b8c6_content_list.json @@ -0,0 +1,2078 @@ +[ + { + "type": "text", + "text": "Collaborative Diffusion for Multi-Modal Face Generation and Editing", + "text_level": 1, + "bbox": [ + 132, + 92, + 836, + 114 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Ziqi Huang Kelvin C.K. Chan Yuming Jiang Ziwei Liu S-Lab, Nanyang Technological University", + "bbox": [ + 230, + 143, + 736, + 178 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{ziqi002, chan0899, yuming002, ziwei.liu}@ntu.edu.sg", + "bbox": [ + 254, + 180, + 712, + 196 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/3bf3231c7177e7c670a9268704a60932cf22a3678547ecc6519ec94df8742f14.jpg", + "image_caption": [ + "Figure 1. We propose Collaborative Diffusion, where users can use multiple modalities to control face generation and editing. (a) Face Generation. Given multi-modal controls, our framework synthesizes high-quality images consistent with the input conditions. (b) Face Editing. Collaborative Diffusion also supports multi-modal editing of real images with promising identity preservation capability." + ], + "image_footnote": [], + "bbox": [ + 81, + 229, + 890, + 494 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 565, + 313, + 580 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Diffusion models arise as a powerful generative tool recently. Despite the great progress, existing diffusion models mainly focus on uni-modal control, i.e., the diffusion process is driven by only one modality of condition. To further unleash the users' creativity, it is desirable for the model to be controllable by multiple modalities simultaneously, e.g. generating and editing faces by describing the age (text-driven) while drawing the face shape (mask-driven).", + "bbox": [ + 75, + 589, + 470, + 710 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this work, we present Collaborative Diffusion, where pre-trained uni-modal diffusion models collaborate to achieve multi-modal face generation and editing without re-training. Our key insight is that diffusion models driven by different modalities are inherently complementary regarding the latent denoising steps, where bilateral connections can be established upon. Specifically, we propose dynamic diffuser, a meta-network that adaptively hallucinates multi-modal denoising steps by predicting the spatial-temporal", + "bbox": [ + 75, + 710, + 470, + 848 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "influence functions for each pre-trained uni-modal model. Collaborative Diffusion not only collaborates generation capabilities from uni-modal diffusion models, but also integrates multiple uni-modal manipulations to perform multimodal editing. Extensive qualitative and quantitative experiments demonstrate the superiority of our framework in both image quality and condition consistency.", + "bbox": [ + 500, + 566, + 893, + 672 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 501, + 705, + 630, + 720 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recent years have witnessed substantial progress in image synthesis and editing with the surge of diffusion models [9, 20, 56, 58]. In addition to the remarkable synthesis quality, one appealing property of diffusion models is the flexibility of conditioning on various modalities, such as texts [3, 16, 29, 30, 39, 48, 51], segmentation masks [48, 65, 66], and sketches [7, 65]. However, existing explorations are largely confined to the use of a single modality at a time. The exploitation of multiple conditions remains under-explored. As a generative tool, its controllability is still limited.", + "bbox": [ + 496, + 732, + 893, + 883 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To unleash users' creativity, it is desirable that the model", + "bbox": [ + 519, + 885, + 890, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 808, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "Corresponding author. \nProject page: https://ziqihuangg.github.io/projects/collaborative-diffusion.html \nCode: https://github.com/ziqihuangg/Collaborative-Diffusion", + "bbox": [ + 81, + 864, + 447, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "6080", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "is simultaneously controllable by multiple modalities. While it is trivial to extend the current supervised framework with multiple modalities, training a large-scale model from scratch is computationally expensive, especially when extensive hyper-parameter tuning and delicate architecture designs are needed. More importantly, each trained model could only accept a fixed combination of modalities, and hence re-training is necessary when a subset of modalities are absent, or when additional modalities become available. The above demonstrates the necessity of a unified framework that effectively exploits pre-trained models and integrate them for multi-modal synthesis and editing.", + "bbox": [ + 75, + 90, + 472, + 271 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we propose Collaborative Diffusion, a framework that synergizes pre-trained uni-modal diffusion models for multi-modal face generation and editing without the need of re-training. Motivated by the fact that different modalities are complementary to each other (e.g., text for age and mask for hair shape), we explore the possibility of establishing lateral connections between models driven by different modalities. We propose dynamic diffuser to adaptively predict the spatial-temporal influence function for each pre-trained model. The dynamic diffuser dynamically determines spatial-varying and temporal-varying influences of each model, suppressing contributions from irrelevant modalities while enhancing contributions from admissible modalities. In addition to multi-modal synthesis, the simplicity and flexibility of our framework enable extension to multi-modal face editing with minimal modifications. In particular, the dynamic diffuser is first trained for collaborative synthesis. It is then fixed and combined with existing face editing approaches [29, 50] for multi-modal editing. It is worth-mentioning that users can select the best editing approaches based on their needs without the need of altering the dynamic diffusers.", + "bbox": [ + 75, + 272, + 472, + 604 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We demonstrate both qualitatively and quantitatively that our method achieves superior image quality and condition consistency in both synthesis and editing tasks. Our contributions can be summarized as follows:", + "bbox": [ + 75, + 604, + 470, + 662 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We introduce Collaborative Diffusion, which exploits pre-trained uni-modal diffusion models for multi-modal controls without re-training. Our approach is the first attempt towards flexible integration of uni-modal diffusion models into a single collaborative framework.", + "Tailored for the iterative property of diffusion models, we propose dynamic diffuser, which predicts the spatial-varying and temporal-varying influence functions to selectively enhance or suppress the contributions of the given modalities at each iterative step.", + "- We demonstrate the flexibility of our framework by extending it to face editing driven by multiple modalities. Both quantitative and qualitative results demonstrate the superiority of Collaborative Diffusion in multi-modal face generation and editing." + ], + "bbox": [ + 96, + 666, + 468, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 89, + 640, + 106 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Diffusion Models. Diffusion models [20, 56, 58] have recently become a mainstream approach for image synthesis [9, 11, 38] apart from Generative Adversarial Networks (GANs) [14], and success has also been found in various domains including video generation [17, 19, 55, 63], image restoration [21, 52], semantic segmentation [1, 4, 15], and natural language processing [2]. In the diffusion-based framework, models are trained with score-matching objectives [22, 64] at various noise levels, and sampling is done via iterative denoising. Existing works focus on improving the performance and efficiency of diffusion models through enhanced architecture designs [16, 48] and sampling schemes [57]. In contrast, this work focuses on exploiting existing models, and providing a succinct framework for multi-modal synthesis and editing without large-scale re-training of models.", + "bbox": [ + 496, + 114, + 893, + 354 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Face Generation. Existing face generation approaches can be divided into three main directions. Following the GAN paradigm, the StyleGAN series [26-28] boost the quality of facial synthesis, and provide an interpretable latent space for steerable style controls and manipulations. The vector-quantized approaches [12, 61] learn a discrete codebook by mapping the input images into a low-dimensional discrete feature space. The learned codebook is then sampled, either sequentially [12, 61] or parallelly [5, 6, 16], for synthesis. In contrast to the previous two approaches, diffusion models are trained with a stationary objective, without the need of optimizing complex losses (e.g., adversarial loss) or balancing multiple objectives (e.g., codebook loss versus reconstruction loss). With training simplicity as a merit, diffusion-based approaches have become increasingly popular in recent years. Our framework falls in the diffusion-based paradigm. In particular, we leverage pre-trained diffusion models for multi-modal generation and editing.", + "bbox": [ + 496, + 357, + 895, + 628 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Conditional Face Generation and Editing. Conditional generation [10, 11, 13, 31, 34, 37, 40, 43-46, 51, 65, 67-71] and editing [8, 32, 41, 53, 54, 67, 68] is an active line of research focusing on conditioning generative models on different modalities, such as texts [24, 41, 67, 68], segmentation masks [32, 33, 40, 47], and audios [59]. For example, StyleCLIP [41], DiffusionCLIP [30], and many others [35, 60] have demonstrated remarkable performance in text-guided face generation and editing. However, most existing models do not support simultaneous conditioning on multiple modalities (e.g., text and mask at the same time), and supporting additional modalities often requires time-consuming model re-training and extensive hyper-parameter tuning, which are not preferable in general. In this work, we propose Collaborative Diffusion to exploit pre-trained uni-modal diffusion models [48] (e.g., text-driven and mask-driven models) to achieve multi-modal conditioning without model retraining.", + "bbox": [ + 496, + 628, + 895, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "6081", + "bbox": [ + 482, + 945, + 513, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/dbf0ef99a8c6ed2c77ed88a1f99b3f247180cfd54c544982d71242804f33497c.jpg", + "image_caption": [ + "Figure 2. Overview of Collaborative Diffusion. We use pre-trained uni-modal diffusion models to perform multi-modal guided face generation and editing. At each step of the reverse process (i.e., from timestep $t$ to $t - 1$ ), the dynamic diffuser predicts the spatial-varying and temporal-varying influence function to selectively enhance or suppress the contributions of the given modality." + ], + "image_footnote": [], + "bbox": [ + 78, + 88, + 898, + 354 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Collaborative Diffusion", + "text_level": 1, + "bbox": [ + 76, + 441, + 299, + 458 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We propose Collaborative Diffusion, which exploits multiple pre-trained uni-modal diffusion models (Section 3.1) for multi-modal generation and editing. The key of our framework is the dynamic diffuser, which adaptively predicts the influence functions to enhance or suppress the contributions of the pre-trained models based on the spatial-temporal influences of the modalities. Our framework is compatible with most existing approaches for both multi-modal guided synthesis (Section 3.2) and multi-modal editing (Section 3.3).", + "bbox": [ + 75, + 470, + 473, + 609 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Uni-Modal Conditional Diffusion Models", + "text_level": 1, + "bbox": [ + 76, + 631, + 429, + 647 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Diffusion models are a class of generative models that model the data distribution in the form of $p_{\\theta}(\\mathbf{x}_0) \\coloneqq \\int p_{\\theta}(\\mathbf{x}_{0:T}) d\\mathbf{x}_{1:T}$ . The diffusion process (a.k.a. forward process) gradually adds Gaussian noise to the data and eventually corrupts the data $\\mathbf{x}_0$ into an approximately pure Gaussian noise $\\mathbf{x}_T$ using a variance schedule $\\beta_1, \\ldots, \\beta_T$ :", + "bbox": [ + 75, + 660, + 472, + 752 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nq \\left(\\mathbf {x} _ {1: T} \\mid \\mathbf {x} _ {0}\\right) := \\prod_ {t = 1} ^ {T} q \\left(\\mathbf {x} _ {t} \\mid \\mathbf {x} _ {t - 1}\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 130, + 772, + 470, + 813 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nq (\\mathbf {x} _ {t} | \\mathbf {x} _ {t - 1}) := \\mathcal {N} (\\mathbf {x} _ {t}; \\sqrt {1 - \\beta_ {t}} \\mathbf {x} _ {t - 1}, \\beta_ {t} \\mathbf {I}).\n$$\n", + "text_format": "latex", + "bbox": [ + 132, + 816, + 413, + 835 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Reversing the forward process allows sampling new data $\\mathbf{x}_0$ by starting from $p(\\mathbf{x}_T) = \\mathcal{N}(\\mathbf{x}_T; \\mathbf{0}, \\mathbf{I})$ . The reverse process is defined as a Markov chain where each step is a learned", + "bbox": [ + 75, + 854, + 470, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Gaussian transition $(\\pmb{\\mu}_{\\theta},\\pmb{\\Sigma}_{\\theta})$", + "bbox": [ + 500, + 441, + 699, + 459 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\np _ {\\theta} \\left(\\mathbf {x} _ {0: T}\\right) := p \\left(\\mathbf {x} _ {T}\\right) \\prod_ {t = 1} ^ {T} p _ {\\theta} \\left(\\mathbf {x} _ {t - 1} \\mid \\mathbf {x} _ {t}\\right), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 535, + 465, + 892, + 506 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\np _ {\\theta} \\left(\\mathbf {x} _ {t - 1} \\mid \\mathbf {x} _ {t}\\right) := \\mathcal {N} \\left(\\mathbf {x} _ {t - 1}; \\boldsymbol {\\mu} _ {\\theta} \\left(\\mathbf {x} _ {t}, t\\right), \\boldsymbol {\\Sigma} _ {\\theta} \\left(\\mathbf {x} _ {t}, t\\right)\\right).\n$$\n", + "text_format": "latex", + "bbox": [ + 537, + 508, + 851, + 526 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Training diffusion models relies on minimizing the variational bound on $p(x)$ 's negative log-likelihood. The commonly used optimization objective $L_{\\mathrm{DM}}$ [20] reparameterizes the learnable Gaussian transition as $\\epsilon_{\\theta}(\\cdot)$ , and temporally reweights the variational bound to trade for better sample quality:", + "bbox": [ + 496, + 530, + 895, + 621 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nL _ {\\mathrm {D M}} (\\theta) := \\mathbb {E} _ {t, \\mathbf {x} _ {0}, \\boldsymbol {\\epsilon} \\sim \\mathcal {N} (\\mathbf {0}, \\mathbf {I})} \\left[ \\| \\boldsymbol {\\epsilon} - \\boldsymbol {\\epsilon} _ {\\theta} (\\mathbf {x} _ {t}, t) \\| ^ {2} \\right], \\quad (3)\n$$\n", + "text_format": "latex", + "bbox": [ + 540, + 626, + 892, + 655 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mathbf{x}_t$ can be directly approximated by $\\mathbf{x}_t = \\sqrt{\\bar{\\alpha}_t}\\mathbf{x}_0 + \\sqrt{1 - \\bar{\\alpha}_t}\\pmb {\\epsilon}$ , with $\\bar{\\alpha}_{t}:= \\prod_{s = 1}^{t}\\alpha_{s}$ and $\\alpha_{t}:= 1 - \\beta_{t}$ .", + "bbox": [ + 498, + 657, + 893, + 689 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To sample data $\\mathbf{x}_0$ from a trained diffusion model $\\epsilon_{\\theta}(\\cdot)$ , we iteratively denoise $\\mathbf{x}_t$ from $t = T$ to $t = 1$ with noise $\\mathbf{z}$ :", + "bbox": [ + 498, + 689, + 893, + 719 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {x} _ {t - 1} = \\frac {1}{\\sqrt {\\alpha_ {t}}} \\left(\\mathbf {x} _ {t} - \\frac {1 - \\alpha_ {t}}{\\sqrt {1 - \\bar {\\alpha} _ {t}}} \\boldsymbol {\\epsilon} _ {\\theta} (\\mathbf {x} _ {t}, t)\\right) + \\sigma_ {t} \\mathbf {z} \\quad (4)\n$$\n", + "text_format": "latex", + "bbox": [ + 535, + 724, + 892, + 758 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The unconditional diffusion models can be extended to model conditional distributions $p_{\\theta}(\\mathbf{x}_0|c)$ , where $\\mathbf{x}_0$ is the image corresponding to the condition $c$ such as class labels, segmentation masks, and text descriptions. The conditional diffusion model receives an additional input $\\tau(c)$ and is trained by minimizing $\\| \\boldsymbol{\\epsilon} - \\boldsymbol{\\epsilon}_{\\theta}(\\mathbf{x}_t, t, \\tau(c)) \\|^2$ , where $\\tau(\\cdot)$ is an encoder that projects the condition $c$ to an embedding $\\tau(c)$ . For brevity, we will use $c$ to represent $\\tau(c)$ in our subsequent discussions.", + "bbox": [ + 496, + 762, + 893, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "6082", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "code", + "sub_type": "algorithm", + "code_caption": [ + "Algorithm 1 Dynamic Diffuser Training" + ], + "code_body": "1: repeat \n2: $\\mathbf{x}_0, c_1, c_2, \\ldots, c_M \\sim q(\\mathbf{x}_0, c_1, c_2, \\ldots, c_M)$ \n3: $t \\sim \\mathrm{Uniform}(\\{1, \\ldots, T\\})$ \n4: $\\epsilon \\sim \\mathcal{N}(\\mathbf{0}, \\mathbf{I})$ \n5: for $m = 1, \\ldots, M$ do \n6: $\\epsilon_{pred,m,t} = \\epsilon_{\\theta_m}(\\sqrt{\\bar{\\alpha}_t}\\mathbf{x}_0 + \\sqrt{1 - \\bar{\\alpha}_t}\\boldsymbol{\\epsilon}, t, c_m)$ \n7: $\\mathbf{I}_{m,t} = \\mathbf{D}_{\\phi_m}(\\sqrt{\\bar{\\alpha}_t}\\mathbf{x}_0 + \\sqrt{1 - \\bar{\\alpha}_t}\\boldsymbol{\\epsilon}, t, c_m)$ \n8: end for \n9: $\\hat{\\mathbf{I}}_{m,t,p} = \\frac{\\exp(\\mathbf{I}_{m,t,p})}{\\sum_{j=1}^{M}\\exp(\\mathbf{I}_{j,t,p})}$ , softmax at each pixel $p$ \n10: $\\epsilon_{pred,t} = \\sum_{m=1}^{M}\\hat{\\mathbf{I}}_{m,t} \\odot \\epsilon_{pred,m,t}$ \n11: Take gradient descent step on $\\nabla_{\\phi} \\| \\boldsymbol{\\epsilon} - \\boldsymbol{\\epsilon}_{pred,t} \\|^2$ where $\\phi = \\{\\phi_m | m = 1, \\ldots, M\\}$ \n12: until converged", + "bbox": [ + 83, + 125, + 444, + 330 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/f73bd6938ea41ca9e485b986ac7e1896cd841e0c6dee6210685de9919a86cf41.jpg", + "image_caption": [ + "Figure 3. Dynamic Diffuser. It predicts the influence function that represents the desired level of contribution from each pre-trained diffusion model. The influence function varies at different timestep $t$ and different spatial locations, and is dependent on the state of diffusion $\\mathbf{x}_t$ and the condition input $c_m$ ." + ], + "image_footnote": [], + "bbox": [ + 135, + 344, + 408, + 465 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Multi-Modal Collaborative Synthesis", + "text_level": 1, + "bbox": [ + 76, + 584, + 397, + 599 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In our Collaborative Diffusion, multiple uni-modal diffusion models collaborate at each step of the denoising process for multi-modal guided synthesis. The core of our framework is the dynamic diffuser, which determines the extent of contribution from each collaborator by predicting the spatial-temporal influence functions.", + "bbox": [ + 75, + 614, + 470, + 705 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Problem Formulation. Given $M$ pre-trained uni-modal conditional diffusion models $\\{\\epsilon_{\\theta_m}\\}$ (each models the distribution $p(\\mathbf{x}_0|c_m)$ ), where the modality index $m = 1, \\ldots, M$ , our objective is to sample from $p(\\mathbf{x}_0|\\mathbf{c})$ , where $\\mathbf{c} = \\{c_1, c_2, \\dots, c_M\\}$ , without altering pre-trained models.", + "bbox": [ + 76, + 712, + 470, + 789 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Dynamic Diffuser. In diffusion models, each step of the reverse process requires predicting the noise $\\epsilon$ . With multiple diffusion models collaborating, we need to carefully determine when, where, and how much each diffusion model contributes to the prediction $\\epsilon$ . At each diffusion timestep $t = T, \\dots, 1$ , the influence $\\mathbf{I}_{m,t}$ from each pre-trained diffusion model $\\epsilon_{\\theta_m}$ is adaptively determined by a dynamic", + "bbox": [ + 75, + 794, + 472, + 902 + ], + "page_idx": 3 + }, + { + "type": "code", + "sub_type": "algorithm", + "code_caption": [ + "Algorithm 2 Collaborative Sampling" + ], + "code_body": "1: $\\mathbf{x}_T\\sim \\mathcal{N}(\\mathbf{0},\\mathbf{I})$ \n2: for $t = T,\\dots ,1$ do \n3: $\\mathbf{z}\\sim \\mathcal{N}(\\mathbf{0},\\mathbf{I})$ if $t > 1$ , else $\\mathbf{z} = \\mathbf{0}$ \n4: for $m = 1,\\ldots ,M$ do \n5: $\\epsilon_{pred,m,t} = \\epsilon_{\\theta_m}(\\mathbf{x}_t,t,c_m)$ \n6: $\\mathbf{I}_{m,t} = \\mathbf{D}_{\\phi_m}(\\mathbf{x}_t,t,c_m)$ \n7: end for \n8: $\\hat{\\mathbf{I}}_{m,t,p} = \\frac{\\exp(\\mathbf{I}_{m,t,p})}{\\sum_{j = 1}^{M}\\exp(\\mathbf{I}_{j,t,p})}$ , softmax at each pixel $p$ \n9: $\\epsilon_{pred,t} = \\sum_{m = 1}^{M}\\hat{\\mathbf{I}}_{m,t}\\odot \\epsilon_{pred,m,t}$ \n10: $\\mathbf{x}_{t - 1} = \\frac{1}{\\sqrt{\\alpha_t}}\\left(\\mathbf{x}_t - \\frac{1 - \\alpha_t}{\\sqrt{1 - \\bar{\\alpha}_t}}\\pmb{\\epsilon}_{pred,t}\\right) + \\sigma_t\\mathbf{z}$ \n11: end for \n12: return ${\\bf x}_0$", + "bbox": [ + 491, + 128, + 813, + 327 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "diffuser $\\mathbf{D}_{\\phi_m}$", + "bbox": [ + 500, + 347, + 594, + 364 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {I} _ {m, t} = \\mathbf {D} _ {\\phi_ {m}} (\\mathbf {x} _ {t}, t, c _ {m}), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 614, + 371, + 890, + 387 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $m = 1,\\dots,M$ is the index of the modalities, $\\mathbf{I}_{m,t}\\in$ $\\mathbb{R}^{h\\times w}$ $\\mathbf{x}_t$ is the noisy image at time $t$ , $c_{m}$ is the condition of the $m^{th}$ modality, and $\\mathbf{D}_{\\phi_m}$ is the dynamic diffuser implemented by a UNet [49]. To regularize the overall influence strength, we perform softmax across all modalities' $\\mathbf{I}_{m,t}$ at each pixel $p$ to obtain the final influence function $\\hat{\\mathbf{I}}_{m,t}$ :", + "bbox": [ + 496, + 395, + 893, + 489 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathbf {I}} _ {m, t, p} = \\frac {\\exp \\left(\\mathbf {I} _ {m , t , p}\\right)}{\\sum_ {j = 1} ^ {M} \\exp \\left(\\mathbf {I} _ {j , t , p}\\right)}. \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 601, + 496, + 890, + 534 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Multi-Modal Collaboration. We use the $M$ learned influence functions $\\hat{\\mathbf{I}}_{m,t}$ to control the contribution from each pre-trained diffusion model at each denoising step:", + "bbox": [ + 496, + 540, + 893, + 585 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {\\epsilon} _ {p r e d, t} = \\sum_ {m = 1} ^ {M} \\hat {\\mathbf {I}} _ {m, t} \\odot \\boldsymbol {\\epsilon} _ {\\theta_ {m}} \\left(x _ {t}, t, c _ {m}\\right) \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 573, + 593, + 890, + 635 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\epsilon_{\\theta_m}$ is the $m^{th}$ collaborator, and $\\odot$ denotes pixel-wise multiplication. The complete training procedure of dynamic diffusers and the image sampling strategy in our Collaborative Diffusion framework are detailed in Algorithm 1 and Algorithm 2, respectively.", + "bbox": [ + 496, + 643, + 893, + 719 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Relation with Composable Diffusion. Composable Diffusion [36] combines two instances of the same text-to-image diffusion model to achieve compositional visual generation via intermediate results addition. Our framework is related to Composable Diffusion in terms of composing diffusion models for image synthesis, but is substantially different in terms of the task nature and methodology. Composable Diffusion aims to decompose the text condition into elementary segments to factorize the conditional synthesis problem, while we aim to integrate uni-modal collaborators to achieve multi-modal controls. Furthermore, different from Composable Diffusion, which can only compose instances of the", + "bbox": [ + 496, + 719, + 893, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "6083", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Algorithm 3 Collaborative Editing", + "text_level": 1, + "bbox": [ + 78, + 104, + 308, + 121 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Require:", + "text_level": 1, + "bbox": [ + 80, + 128, + 135, + 141 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "input image $\\mathbf{x}_{input}$ , target conditions $c_{m,target}$ , diffusion models $\\epsilon_{\\theta_m}$ , dynamic diffusers $\\mathbf{D}_{\\phi_m}$ , $(m = 1, \\dots, M)$ , interpolation scale $\\alpha$", + "bbox": [ + 102, + 143, + 480, + 196 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1: for $m = 1,\\ldots ,M$ do $\\triangleright$ Uni-Modal Editing", + "2: $c_{m} = c_{m, \\text{target}}$", + "3: $\\mathbf{x}_t = \\sqrt{\\bar{\\alpha}_t}\\mathbf{x}_{input} + \\sqrt{1 - \\bar{\\alpha}_t}\\pmb{\\epsilon}$", + "4: $c_{m,opt} = \\mathrm{argmin}_{c_m}\\mathbb{E}_{\\epsilon ,t}\\left\\| \\boldsymbol {\\epsilon} - \\boldsymbol{\\epsilon}_{\\theta_m}(\\mathbf{x}_t,t,c_m)\\right\\| ^2$", + "5: $\\theta_{m,opt} = \\operatorname*{argmin}_{\\theta_m}\\mathbb{E}_{\\epsilon ,t}\\| \\pmb {\\epsilon} - \\pmb{\\epsilon}_{\\theta_m}(\\mathbf{x}_t,t,c_{m,opt})\\|^2$", + "6: $c_{m,int} = \\alpha \\cdot c_{m,target} + (1 - \\alpha) \\cdot c_{m,opt}$", + "7: end for" + ], + "bbox": [ + 86, + 200, + 480, + 299 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "8: $\\mathbf{x}_T\\sim \\mathcal{N}(\\mathbf{0},\\mathbf{I})$ ▷Collaborate the Uni-Modal Edits", + "9: for $t = T, \\dots, 1$ do", + "10: $\\mathbf{z} \\sim \\mathcal{N}(\\mathbf{0}, \\mathbf{I})$ if $t > 1$ , else $\\mathbf{z} = \\mathbf{0}$", + "11: for $m = 0,\\dots,M$ do", + "12: $\\pmb{\\epsilon}_{pred,m,t} = \\pmb{\\epsilon}_{\\theta_{m,opt}}(\\mathbf{x}_t,t,c_{m,int})$", + "13: $\\mathbf{I}_{m,t} = \\mathbf{D}_{\\phi_m}(\\mathbf{x}_t,t,c_{m,int})$", + "14: end for", + "15: $\\hat{\\mathbf{I}}_{m,t,p} = \\frac{\\exp(\\mathbf{I}_{m,t,p})}{\\sum_{j = 1}^{M}\\exp(\\mathbf{I}_{j,t,p})}$ , softmax at each pixel $p$", + "16: $\\pmb{\\epsilon}_{pred,t} = \\sum_{m=1}^{M} \\hat{\\mathbf{I}}_{m,t} \\odot \\pmb{\\epsilon}_{pred,m,t}$", + "17: $\\mathbf{x}_{t - 1} = \\frac{1}{\\sqrt{\\alpha_t}}\\left(\\mathbf{x}_t - \\frac{1 - \\alpha_t}{\\sqrt{1 - \\bar{\\alpha}_t}}\\pmb {\\epsilon}_{pred,t}\\right) + \\sigma_t\\mathbf{z}$", + "18: end for", + "19: return $\\mathbf{x}_0$" + ], + "bbox": [ + 81, + 303, + 480, + 500 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "same text-based synthesis model (i.e., the weights of the diffusion models being composed are the same), our framework possesses the flexibility in integrating models with different weights, architectures, and modalities through the learned dynamic diffusers.", + "bbox": [ + 75, + 526, + 470, + 602 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Multi-Modal Collaborative Editing", + "text_level": 1, + "bbox": [ + 76, + 617, + 382, + 633 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In addition to face synthesis, our framework is also capable of combining multiple facial manipulations, each of which is guided by a different modality, into a collaborative edit. Our collaborative framework in theory can integrate any existing uni-modal diffusion-based editing approach for collaborative editing.", + "bbox": [ + 75, + 642, + 468, + 733 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this work, we demonstrate such possibility by extending Imagic [29] to the multi-modal paradigm. We first follow Imagic to fine-tune the embeddings and models to better capture the input face identity during editing. The trained dynamic diffusers discussed in Section 3.2 are then used to combine the fine-tuned models. Algorithm 3 displays the complete procedure of collaborative editing. Note that finetuning the pre-trained models is for identity preservation proposed in Imagic, which is independent to our framework. The extension of our framework to editing requires no further training of dynamic diffusers.", + "bbox": [ + 75, + 734, + 470, + 901 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/9afea71ee8fa0a294a9108b1256c1416ff8d90f970bc96f75d04ce55f06121c3.jpg", + "image_caption": [ + "Multi-Modal Conditions" + ], + "image_footnote": [ + "He is a young adult. This man has a rough growth of stubble." + ], + "bbox": [ + 503, + 99, + 573, + 152 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/e3fc0353b8f8566b51e014a0a17c5231e48c82e30bb0a033d501cc873fe91127.jpg", + "image_caption": [ + "Figure 4. Qualitative Comparison of Face Generation. In the second example, TediGAN and Composable fails to follow mask, while ours generates results highly consistent to both conditions." + ], + "image_footnote": [ + "There is no beard. This woman looks extremely young." + ], + "bbox": [ + 503, + 210, + 573, + 263 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/ff5c22629236d074c61a2c66f9f4c231185a5cb50d3ba7088c0e44b72b0fbb7d.jpg", + "image_caption": [ + "Generated Images" + ], + "image_footnote": [], + "bbox": [ + 671, + 99, + 741, + 152 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/7a8b8046e2bd97ef9d3789d697463be08dc7d8d48d5b8265b3370f2e3bd3bb59.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 745, + 99, + 813, + 152 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/d2afee1f2fad5f00c14f368195b69901dedc916a6cfed21b2dfaec8b755c0cef.jpg", + "image_caption": [], + "image_footnote": [ + "His face is covered with short beard. \nThis man is in the thirties." + ], + "bbox": [ + 671, + 154, + 741, + 209 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/f4e49f4c2256072aa21fbebb01b7f6367d2e8fba4ac00d4904ea181897c1881c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 816, + 99, + 885, + 152 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/237f4c875b4324a67acb1696a3f2b9296622fc37220d16b2426bea81ace0dfe5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 745, + 154, + 813, + 208 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/244c8f0d25ffe5a9aef5bf2ce195a999ebf28fb483a0b9d07a0e04c3a03bfac7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 816, + 155, + 883, + 209 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/ca7caebd14fc05883ffa33da300861836f97f6a52567ead526753691c97f72a9.jpg", + "image_caption": [], + "image_footnote": [ + "TediGAN Composable" + ], + "bbox": [ + 671, + 210, + 741, + 263 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/b46c9420b617f1d743b96885c32a4fd385a3bf2cc7829a5ca925862a8f5c04f5.jpg", + "image_caption": [], + "image_footnote": [ + "Ours" + ], + "bbox": [ + 816, + 210, + 885, + 263 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/7174d87008736e4024c323dc9a414840ad5a48a8aa63f2401450cb976bb55686.jpg", + "image_caption": [ + "Input Image" + ], + "image_footnote": [], + "bbox": [ + 503, + 337, + 563, + 382 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/e8b4d0e2ed709a0cfb1a85472acf26cba3d8b7fc7133dd44fb4a22889b715598.jpg", + "image_caption": [ + "Multi-Modal Target Conditions" + ], + "image_footnote": [ + "He is a teen. The face is covered with short pointed beard." + ], + "bbox": [ + 566, + 337, + 625, + 382 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/3520e2288f41faa1c38bc6c6e2d5a215550965b1cb2f5c5221d6aa674f478faf.jpg", + "image_caption": [ + "Figure 5. Qualitative Comparison of Face Editing. While TediGAN's hair shape is inconsistent with mask, and Composable fails to generate beard according to text, our framework produces results highly consistent with both conditions while maintaining identity." + ], + "image_footnote": [], + "bbox": [ + 503, + 383, + 562, + 428 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/b9bcb15e50f9ce26476b5d0cdbe998e1d22705c07102469e0a265325bb676f28.jpg", + "image_caption": [], + "image_footnote": [ + "This man has \nheard of medium \nlength. He is in his \nthirties." + ], + "bbox": [ + 566, + 383, + 625, + 428 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/1b5fbed737513bb8fb7e433512dd81419bad04689ea696f726badb309451bc68.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 705, + 337, + 767, + 382 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/e4f47afe8cd0521502406ea665e0807f2b350d035dc45346fb3e3bbb226cb3a5.jpg", + "image_caption": [ + "Edited Images" + ], + "image_footnote": [], + "bbox": [ + 767, + 337, + 823, + 382 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/6d67adf7d26a0464be27cd72f9160cafa604671f957b1a2eb987a89652a6fab6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 828, + 337, + 885, + 382 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/52d0b0cdf1f55777c71752121eec113d75a6668d743f0fd746646696290b6fa7.jpg", + "image_caption": [ + "TediGAN" + ], + "image_footnote": [], + "bbox": [ + 707, + 383, + 767, + 428 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/59bec688eed7b918af67b7643a945ec0611547dda362e23b37878875dd8f68f8.jpg", + "image_caption": [ + "Composable" + ], + "image_footnote": [], + "bbox": [ + 769, + 383, + 826, + 428 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/860490e9e49f19e2c9acbeb7b30ef2470cca211c67d3c1870bdb1c4734d8df04.jpg", + "image_caption": [], + "image_footnote": [ + "Ours" + ], + "bbox": [ + 828, + 383, + 885, + 428 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 500, + 511, + 630, + 529 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Experimental Setup", + "text_level": 1, + "bbox": [ + 500, + 537, + 689, + 553 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Datasets. CelebAMask-HQ [32] consists of manually annotated segmentation masks for each of the 30,000 images in the CelebA-HQ dataset [25]. Each mask has up to 19 classes, including the main facial components such as hair, skin, eyes, and nose, and accessories such as eyeglasses and cloth. CelebA-Dialog [23] provides rich and fine-grained natural language descriptions for the images in CelebA-HQ. In this work, we train the mask-driven diffusion models on CelebAMask-HQ, and the text-driven models on CelebA-Dialog. For dynamic diffuser, we simply combine the mask and text for each image without further processing.", + "bbox": [ + 496, + 561, + 893, + 727 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Implementation Details. We adopt LDM [48] for our unimodal diffusion models since it achieves a good balance between quality and speed. Our dynamic diffuser is a time-conditional UNet [49], with conditional embeddings injected via cross-attention [62]. The dynamic diffuser is $\\times 30$ smaller than the pre-trained diffusion model. The detailed architecture and settings are provided in the supplementary material.", + "bbox": [ + 496, + 728, + 893, + 834 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Comparison Methods", + "text_level": 1, + "bbox": [ + 500, + 845, + 700, + 861 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "TediGAN [67, 68] is a StyleGAN-based face synthesis and manipulation method. It projects both text and mask condi", + "bbox": [ + 500, + 869, + 893, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "6084", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/a33367654f1906533af4ab8fb28400b3d09af7ec3b3a6b192cfc9a2ed3b675af.jpg", + "image_caption": [ + "Multi-Modal Conditions", + "His face is covered with short beard. He is a young adult." + ], + "image_footnote": [], + "bbox": [ + 83, + 102, + 151, + 155 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/bb022db63d2e0adda2410a35ba889f6ac9bd6f3162bb0874c76831e972c85b6b.jpg", + "image_caption": [ + "She looks very young.", + "Figure 6. Diversity of Face Generation. Given input conditions, our method demonstrates promising synthesis diversity in facial attributes that are not constrained by a specific combination of multi-modal conditions, such as hair color and skin tone." + ], + "image_footnote": [], + "bbox": [ + 84, + 159, + 151, + 212 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/c9264fd09f2d4328f6efb8749e2dc8ee0e2d5b23ecadadc7c17a8ced9229def2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 246, + 101, + 316, + 154 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/d6bdd71d45001981bf303c32e376a35b43cdb345ed551ce66b1759ae390ba564.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 318, + 101, + 385, + 154 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/7c6e471ca7ab0380639a28fb4222c278a84db6fca4f27f478c051160ce6283c1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 390, + 101, + 457, + 154 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/d1108f04cff15f86373edadb91d4f5b793e4a515af9316e627e781f69920e8b5.jpg", + "image_caption": [ + "Generated Images" + ], + "image_footnote": [], + "bbox": [ + 460, + 101, + 527, + 154 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/b85a133ac7ded7c791d388dbfa9f44caa22fd7677041c0a7b3ed5d350c70aa8c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 531, + 102, + 599, + 154 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/46dad76165ae071f27551e8d0dec4b66c0d989c27f8a5df9269f1031e9285611.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 604, + 101, + 669, + 154 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/afdddf07b2a9fd2b480f5c17adad67e692fe0a1b3dd69a58926a66ff2833e285.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 673, + 101, + 740, + 154 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/67e6e83f4b5991a0da8c9fc8a071b75a44625193cb63e67c7f0ab1da68894c0e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 746, + 101, + 812, + 154 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/fef8961c5e3fca360eaa8f50a9b596e7a12d577ddf0310ac8d1e04b23721cdbe.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 816, + 101, + 883, + 154 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/357fb1a75d26255bfc3c028468306284eae275f3d7cc38eb3a4580533b7a1f36.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 246, + 157, + 313, + 210 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/cbf82fbd3a6cf6eb8ad9a14c73c25df05e5811b405e6a94342def08bd6f47b7a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 316, + 157, + 385, + 210 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/0289ddbb97f925db0dfc73a83794eccfaca1873b93a33184ffa6e2b0fcb46318.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 390, + 157, + 457, + 210 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/040d206d3991b0e0e22dd2f0addd5bccadac9e3fa335e7ead1143191b1d397d7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 460, + 157, + 527, + 210 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/41d152e219859cd14bd89155bccd6c42522d5d3c5664a81541b0d96a3550c19c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 531, + 157, + 598, + 210 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/a0cd8a2c011fb9b0d48c4f97a0a8b714872c57904954fd15e57152671df61a00.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 602, + 157, + 669, + 210 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/a5585141c1f994de37bfc736e02d3d1993d86ff76f9877870f1e103d454f991e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 674, + 157, + 741, + 210 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/ad75a40e67e59702546d2b24179c997660cde30ca8fe0e82eceb8e3d2695aab0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 746, + 157, + 813, + 210 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/c7b47b8845ca6d621f22a0b1fa5705cfb6b11abd21bb38cbd37d1494e088674c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 816, + 157, + 883, + 212 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "tions into StyleGAN's $\\mathcal{W}+$ latent space, and performs style mixing to achieve multi-modal control.", + "bbox": [ + 75, + 281, + 467, + 311 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Composable Diffusion [36] parallelly implements two instances of the same text-to-image diffusion model to perform compositional scene generation. We generalize it by replacing the two text-to-image model instances by the text-driven model and the mask-driven model trained using LDM [48]. The two models are composed by averaging the intermediate diffusion results at each step of the reverse process.", + "bbox": [ + 75, + 313, + 470, + 419 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Evaluation Metrics", + "text_level": 1, + "bbox": [ + 76, + 431, + 261, + 448 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "FID. We use Frechet Inception Distance (FID) [18] to measure the quality of images synthesized by different methods. FID computes the feature representation's distance between generated images and real images. A lower FID implies better sample quality.", + "bbox": [ + 75, + 457, + 470, + 532 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "CLIP Score. CLIP [42] is a vision-language model trained on large-scale datasets. It uses an image encoder and a text encoder to project images and texts to a common feature space, respectively. The CLIP score is computed as the cosine similarity between the normalized image and text embeddings. A higher score usually indicates higher consistency between the output image and the text caption.", + "bbox": [ + 75, + 535, + 468, + 641 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Mask Accuracy. For each output image, we predict the segmentation mask using the face parsing network provided by CelebAMask-HQ [32]. Mask accuracy is the pixel-wise accuracy against the ground-truth segmentation. A higher average accuracy indicates better consistency between the output image and the segmentation mask.", + "bbox": [ + 75, + 642, + 468, + 733 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "User Study. We conduct user study with 25 human evaluators to measure the effectiveness of the methods perceptually. For face generation, we randomly sample multi-modal conditions in the validation split of CelebA-HQ Dataset, and then synthesize output images given the conditions. Evaluators are provided with the input conditions and the output images, and they are asked to choose the best image based on 1) image photo-realism, 2) image-text consistency, 3) image-mask consistency. User study is conducted similarly for face editing, except that 1) evaluators are also provided with the input image for editing, and 2) they are also asked", + "bbox": [ + 75, + 734, + 470, + 901 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/ebe4090b3693feca9cdad83e22a49e40c1565aca502eff8fb6eb05615f7cafd8.jpg", + "table_caption": [ + "Table 1. Quantitative Results of Face Generation. Compared with TediGAN and Composable Diffusion, our method synthesizes images with better quality (lower FID), and higher consistency with the text and mask conditions." + ], + "table_footnote": [], + "table_body": "
MethodFID ↓Text (%) ↑Mask (%) ↑
TediGAN [67,68]157.8124.2772.19
Composable [36]124.6223.9476.11
Ours111.3624.5180.25
", + "bbox": [ + 511, + 345, + 880, + 407 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/3d7719ce0812780236e50b03b2ace2e8717640c2d109eddd3949b7ae090b9495.jpg", + "image_caption": [ + "(a) Face Generation", + "Figure 7. User Study. Among the three methods, the majority of users vote for our results as the best for both generation (69.40%) and editing (84.37%), in terms of image quality, condition consistency, and identity preservation." + ], + "image_footnote": [], + "bbox": [ + 522, + 422, + 647, + 507 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/d7f016d4115b42bf87fe163aa49feb0ab08580371afafecedc03ea9308f1dcc8.jpg", + "image_caption": [ + "(b) Face Editing" + ], + "image_footnote": [ + "TediGAN", + "Composable", + "Ours" + ], + "bbox": [ + 655, + 422, + 781, + 507 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "to assess the level of identity preservation between the edited image and input image.", + "bbox": [ + 498, + 619, + 890, + 650 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.4. Comparative Study", + "text_level": 1, + "bbox": [ + 500, + 662, + 686, + 679 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Qualitative Comparison. In Figure 8, we provide the generation results of our Collaborative Diffusion under different combinations of texts and masks. It is observed that our framework is capable of synthesizing realistic outputs consistent with the multi-modal inputs, even for relatively rare conditions, such as a man with long hair. We then compare our method with TediGAN [67, 68] and Composable Diffusion [36] in Figure 4. For example, in the second example, TediGAN and Composable Diffusion's results are inconsistent with the mask, while ours is able to maintain both mask and text consistency. We also show in Figure 6 that our method synthesizes images with high diversity without losing condition consistency.", + "bbox": [ + 496, + 686, + 893, + 883 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In addition to image generation, it is also observed in", + "bbox": [ + 517, + 885, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6085", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/6ae37ff451b6a6749125bc47f529b2c1727b26ec0107413352237b6e84634ed1.jpg", + "image_caption": [ + "Figure 8. Face Generation Results. Our method generates realistic images under different combinations of multi-modal conditions, even for relatively rare combinations in the training distribution, such as a man with long hair." + ], + "image_footnote": [], + "bbox": [ + 83, + 88, + 885, + 474 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Figure 9 that our framework is able to edit images based on multi-modal conditions while preserving identity. The comparison to existing works in Figure 5 also verifies our effectiveness. For example, in the second example, while TediGAN is unable to synthesize hair consistent to the mask and Composable Diffusion fails to generate beard according to the text, our framework is able to generate results highly consistent to both conditions while maintaining identity.", + "bbox": [ + 75, + 542, + 468, + 662 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Quantitative Comparison. As shown in Table 1, our Collaborate Diffusion outperforms TediGAN and Composable Diffusion in all three objective metrics. In addition, as depicted in Figure 7, our framework achieves the best result in $69.40\\%$ and $84.37\\%$ of the time for face generation and editing, respectively. These results verify our effectiveness in both image quality and image-condition consistency.", + "bbox": [ + 75, + 670, + 470, + 776 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.5. Ablation Study", + "text_level": 1, + "bbox": [ + 76, + 809, + 228, + 825 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In this section, we visualize the influence functions, and show that it is necessary for the influence function to be both spatially and temporally varying to facilitate effective collaboration.", + "bbox": [ + 75, + 839, + 468, + 898 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.5.1 Spatial Variations of Influence Functions", + "text_level": 1, + "bbox": [ + 500, + 542, + 836, + 556 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "As shown in Figure 10, the influence functions behave differently at different spatial regions. For instance, the influence for the mask-driven model mainly lies on the contours of facial regions, such as the outline of hair, face, and eyes, as these regions are crucial in defining facial layout. In contrast, the influence for the text-driven model is stronger at skin regions including cheeks and chin. This is because the attributes related skin texture, such as age and beard length, are better described by text.", + "bbox": [ + 496, + 566, + 893, + 700 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To verify its necessity, we remove the spatial variation of influence functions. From Table 2 we observe that removing the spatial variance results in deterioration in both output quality and condition consistency. This corroborates our hypothesis that it is important to assign different weights to different modalities at different spatial regions.", + "bbox": [ + 496, + 703, + 893, + 792 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.5.2 Temporal Variations of Influence Functions", + "text_level": 1, + "bbox": [ + 500, + 815, + 856, + 830 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In addition to the spatial variation, it is also observed in Figure 10 that the influence from the mask-driven model is stronger at earlier diffusion stages (i.e., larger $t$ ), since early stages focus on initializing the facial layout using the mask-", + "bbox": [ + 496, + 839, + 893, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "6086", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/57068522a380006889e26d06ddfccbfc8f02c36f0ff3ee1f9979eb113625785d.jpg", + "image_caption": [ + "Figure 9. Face Editing Results. Given the input real image and target conditions, we display the edited image using our method." + ], + "image_footnote": [], + "bbox": [ + 99, + 88, + 872, + 369 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/1c37642a1c99a2a65435f81a57bac20a5395e200b8834a938f2dfcb76a5fbe5f.jpg", + "image_caption": [ + "Figure 10. Visualization of Influence Functions. The influence functions vary spatially at different face regions, and temporally at different diffusion timesteps." + ], + "image_footnote": [], + "bbox": [ + 83, + 422, + 465, + 563 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/d29a506eca7161418ee3121e0d9e05b0c0e233c304addcc480b0208676d573d0.jpg", + "table_caption": [ + "Table 2. Ablation Study. Temporal or spatial suppression in influence variation introduces performance drops, which shows the necessity of influence functions' spatial-temporal adaptivity." + ], + "table_footnote": [], + "table_body": "
MethodFID ↓Text (%) ↑Mask (%) ↑
Ours w/o Spatial117.8124.3680.08
Ours w/o Temporal117.3424.4877.07
Ours111.3624.5180.25
", + "bbox": [ + 81, + 681, + 460, + 744 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "driven model's predictions. At later stages, the influence from the text-driven model increases as the textural details (e.g., skin wrinkles and beard length) are instantiated using information from the text.", + "bbox": [ + 75, + 773, + 468, + 835 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We remove the temporal variation in the influence functions to demonstrate its importance. It is shown in Table 2 that both the quality and condition consistency drop without the temporal variation.", + "bbox": [ + 75, + 839, + 470, + 902 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 422, + 619, + 436 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "With generative AI gaining increasing attention, multimodal conditioning becomes an indispensable direction to unleash creativity and enable comprehensive controls from human creators. In this work, we take the first step forward and propose Collaborative Diffusion, where pre-trained uni-modal diffusion models collaboratively achieve multi-modal face generation and editing without being re-trained. With our dynamic diffuser, this framework could be used to extend arbitrary uni-modal approach to the multi-modal paradigm through predicting the relative influence of different modalities. We believe our idea of synergizing uni-modal models for multi-modal tasks would be a good inspiration for future works in various domains such as motion and 3D generation.", + "bbox": [ + 496, + 454, + 893, + 652 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Limitations and Future Work. Since our framework focuses on exploiting pre-trained diffusion models, our performance is dependent on the capability of each model. An orthogonal direction is to train each collaborator on large-scale datasets for performance gain.", + "bbox": [ + 496, + 657, + 893, + 736 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Potential Negative Societal Impacts. The facial manipulation capabilities of Collaborative Diffusion could be applied maliciously on real human faces. We advise users to apply it only for proper recreational purposes.", + "bbox": [ + 496, + 741, + 893, + 803 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgement. We would like to thank Liangyu Chen and Bo Li for proofreading. This study is supported by NTU NAP, MOE AcRF Tier 1 (2021-T1-001-088), and under the RIE2020 Industry Alignment Fund - Industry Collaboration Projects (IAF-ICP) Funding Initiative, as well as cash and in-kind contribution from the industry partner(s).", + "bbox": [ + 496, + 809, + 893, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "6087", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Tomer Amit, Eliya Nachmani, Tal Shaharbany, and Lior Wolf. SegDiff: Image segmentation with diffusion probabilistic models. arXiv preprint arXiv:2112.00390, 2021. 2", + "[2] Jacob Austin, Daniel D Johnson, Jonathan Ho, Daniel Tarlow, and Rianne van den Berg. Structured denoising diffusion models in discrete state-spaces. In NeurIPS, 2021. 2", + "[3] Omri Avrahami, Dani Lischinski, and Ohad Fried. Blended diffusion for text-driven editing of natural images. In CVPR, 2022. 1", + "[4] Dmitry Baranchuk, Ivan Rubachev, Andrey Voynov, Valentin Khrulkov, and Artem Babenko. Label-efficient semantic segmentation with diffusion models. In ICLR, 2022. 2", + "[5] Sam Bond-Taylor, Peter Hessey, Hiroshi Sasaki, Toby P Breckon, and Chris G Willcocks. Unleashing transformers: parallel token prediction with discrete absorbing diffusion for fast high-resolution image generation from vector-quantized codes. In ECCV, 2022. 2", + "[6] Huiwen Chang, Han Zhang, Lu Jiang, Ce Liu, and William T Freeman. MaskGIT: Masked generative image transformer. In CVPR, 2022. 2", + "[7] Shin-I Cheng, Yu-Jie Chen, Wei-Chen Chiu, Hung-Yu Tseng, and Hsin-Ying Lee. Adaptively-realistic image generation from stroke and sketch with diffusion model. In WACV, 2023. 1", + "[8] Guillaume Couairon, Jakob Verbeek, Holger Schwenk, and Matthieu Cord. Diffedit: Diffusion-based semantic image editing with mask guidance. arXiv preprint arXiv:2210.11427, 2022. 2", + "[9] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat GANs on image synthesis. In NeurIPS, 2021. 1, 2", + "[10] Ming Ding, Zhuoyi Yang, Wenyi Hong, Wendi Zheng, Chang Zhou, Da Yin, Junyang Lin, Xu Zou, Zhou Shao, Hongxia Yang, et al. CogView: Mastering text-to-image generation via transformers. In NeurIPS, 2021. 2", + "[11] Patrick Esser, Robin Rombach, Andreas Blattmann, and Bjorn Ommer. ImageBART: Bidirectional context with multinomial diffusion for autoregressive image synthesis. In NeurIPS, 2021. 2", + "[12] Patrick Esser, Robin Rombach, and Bjorn Ommer. Taming transformers for high-resolution image synthesis. In CVPR, 2021. 2", + "[13] Oran Gafni, Adam Polyak, Oron Ashual, Shelly Sheynin, Devi Parikh, and Yaniv Taigman. Make-a-scene: Scene-based text-to-image generation with human priors. arXiv preprint arXiv:2203.13131, 2022. 2", + "[14] Ian J Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron C Courville, and Yoshua Bengio. Generative adversarial nets. In NeurIPS, 2014. 2", + "[15] Alexandros Graikos, Nikolay Malkin, Nebojsa Jojic, and Dimitris Samaras. Diffusion models as plug-and-play priors. In NeurIPS, 2022. 2", + "[16] Shuyang Gu, Dong Chen, Jianmin Bao, Fang Wen, Bo Zhang, Dongdong Chen, Lu Yuan, and Baining Guo. Vector quantized diffusion model for text-to-image synthesis. In CVPR, 2022. 1, 2" + ], + "bbox": [ + 78, + 114, + 470, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[17] William Harvey, Saeid Naderiparizi, Vaden Masrani, Christian Weilbach, and Frank Wood. Flexible diffusion modeling of long videos. arXiv preprint arXiv:2205.11495, 2022. 2", + "[18] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. GANs trained by a two time-scale update rule converge to a local nash equilibrium. In NeurIPS, 2017. 6", + "[19] Jonathan Ho, William Chan, Chitwan Sahara, Jay Whang, Ruiqi Gao, Alexey Gritsanko, Diederik P Kingma, Ben Poole, Mohammad Norouzi, David J Fleet, et al. Imagen video: High definition video generation with diffusion models. arXiv preprint arXiv:2210.02303, 2022. 2", + "[20] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In NeurIPS, 2020. 1, 2, 3", + "[21] Jonathan Ho, Chitwan Sahara, William Chan, David J Fleet, Mohammad Norouzi, and Tim Salimans. Cascaded diffusion models for high fidelity image generation. JMLR, 2022. 2", + "[22] Aapo Hyvarinen and Peter Dayan. Estimation of non-normalized statistical models by score matching. JMLR, 2005. 2", + "[23] Yuming Jiang, Ziqi Huang, Xingang Pan, Chen Change Loy, and Ziwei Liu. Talk-to-Edit: Fine-grained facial editing via dialog. In ICCV, 2021. 5", + "[24] Yuming Jiang, Shuai Yang, Haonan Qju, Wayne Wu, Chen Change Loy, and Ziwei Liu. Text2human: Text-driven controllable human image generation. ACM TOG, 2022. 2", + "[25] Tero Karras, Timo Aila, Samuli Laine, and Jaakko Lehtinen. Progressive growing of GANs for improved quality, stability, and variation. In ICLR, 2018. 5", + "[26] Tero Karras, Miika Aittala, Samuli Laine, Erik Härkönen, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Alias-free generative adversarial networks. In NeurIPS, 2021. 2", + "[27] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In CVPR, 2019. 2", + "[28] Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of StyleGAN. In CVPR, 2020. 2", + "[29] Bahjat Kawar, Shiran Zada, Oran Lang, Omer Tov, Huiwen Chang, Tali Dekel, Inbar Mosseri, and Michal Irani. Imagic: Text-based real image editing with diffusion models. arXiv preprint arXiv:2210.09276, 2022. 1, 2, 5", + "[30] Gwanghyun Kim, Taesung Kwon, and Jong Chul Ye. DiffusionCLIP: Text-guided diffusion models for robust image manipulation. In CVPR, 2022. 1, 2", + "[31] Jing Yu Koh, Jason Baldridge, Honglak Lee, and Yinfei Yang. Text-to-image generation grounded by fine-grained user attention. In WACV, 2021. 2", + "[32] Cheng-Han Lee, Ziwei Liu, Lingyun Wu, and Ping Luo. MaskGAN: Towards diverse and interactive facial image manipulation. In CVPR, 2020. 2, 5, 6", + "[33] Bowen Li, Xiaojuan Qi, Thomas Lukasiewicz, and Philip HS Torr. ManiGAN: Text-guided image manipulation. In CVPR, 2020. 2", + "[34] Wenbo Li, Pengchuan Zhang, Lei Zhang, Qiuyuan Huang, Xiaodong He, Siwei Lyu, and Jianfeng Gao. Object-driven text-to-image synthesis via adversarial training. In CVPR, 2019. 2" + ], + "bbox": [ + 501, + 92, + 893, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "6088", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[35] Zhiheng Li, Martin Renqiang Min, Kai Li, and Chenliang Xu. StyleT2I: Toward compositional and high-fidelity text-to-image synthesis. In CVPR, 2022. 2", + "[36] Nan Liu, Shuang Li, Yilun Du, Antonio Torralba, and Joshua B Tenenbaum. Compositional visual generation with composable diffusion models. arXiv preprint arXiv:2206.01714, 2022. 4, 6", + "[37] Xihui Liu, Dong Huk Park, Samaneh Azadi, Gong Zhang, Arman Chopikyan, Yuxiao Hu, Humphrey Shi, Anna Rohrbach, and Trevor Darrell. More control for free! image synthesis with semantic diffusion guidance. In WACV, 2023. 2", + "[38] Chenlin Meng, Yutong He, Yang Song, Jiaming Song, Jiajun Wu, Jun-Yan Zhu, and Stefano Ermon. SDEdit: Guided image synthesis and editing with stochastic differential equations. In ICLR, 2022. 2", + "[39] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. GLIDE: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741, 2021. 1", + "[40] Taesung Park, Ming-Yu Liu, Ting-Chun Wang, and Jun-Yan Zhu. Semantic image synthesis with spatially-adaptive normalization. In CVPR, 2019. 2", + "[41] Or Patashnik, Zongze Wu, Eli Shechtman, Daniel Cohen-Or, and Dani Lischinski. StyleCLIP: Text-driven manipulation of stylegan imagery. In ICCV, 2021. 2", + "[42] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, 2021. 6", + "[43] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with CLIP latents. arXiv preprint arXiv:2204.06125, 2022. 2", + "[44] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In ICML, 2021. 2", + "[45] Scott Reed, Zeynep Akata, Xinchen Yan, Lajanugen Logeswaran, Bernt Schiele, and Honglak Lee. Generative adversarial text to image synthesis. In ICML, 2016. 2", + "[46] Scott E Reed, Zeynep Akata, Santosh Mohan, Samuel Tenka, Bernt Schiele, and Honglak Lee. Learning what and where to draw. In NeurIPS, 2016. 2", + "[47] Elad Richardson, Yuval Alaluf, Or Patashnik, Yotam Nitzan, Yaniv Azar, Stav Shapiro, and Daniel Cohen-Or. Encoding in style: a StyleGAN encoder for image-to-image translation. In CVPR, 2021. 2", + "[48] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 1, 2, 5, 6", + "[49] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-Net: Convolutional networks for biomedical image segmentation. In MICCAI, 2015. 4, 5", + "[50] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine" + ], + "bbox": [ + 78, + 90, + 470, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "tuning text-to-image diffusion models for subject-driven generation. arXiv preprint arXiv:2208.12242, 2022. 2", + "[51] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily Denton, Seyed Kamyar Seyed Ghasemipour, Burcu Karagol Ayan, S Sara Mahdavi, Rapha Gontijo Lopes, et al. Photorealistic text-to-image diffusion models with deep language understanding. arXiv preprint arXiv:2205.11487, 2022. 1, 2", + "[52] Chitwan Sahara, Jonathan Ho, William Chan, Tim Salimans, David J Fleet, and Mohammad Norouzi. Image superresolution via iterative refinement. IEEE TPAMI, 2022. 2", + "[53] Yujun Shen, Jinjin Gu, Xiaou Tang, and Bolei Zhou. Interpreting the latent space of GANs for semantic face editing. In CVPR, 2020. 2", + "[54] Yujun Shen, Ceyuan Yang, Xiaou Tang, and Bolei Zhou. InterFaceGAN: Interpreting the disentangled face representation learned by GANs. IEEE TPAMI, 2020. 2", + "[55] Uriel Singer, Adam Polyak, Thomas Hayes, Xi Yin, Jie An, Songyang Zhang, Qiyuan Hu, Harry Yang, Oron Ashual, Oran Gafni, et al. Make-a-video: Text-to-video generation without text-video data. arXiv preprint arXiv:2209.14792, 2022. 2", + "[56] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In ICML, 2015. 1, 2", + "[57] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In ICLR, 2021. 2", + "[58] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. In ICLR, 2021. 1, 2", + "[59] Yang Song, Jingwen Zhu, Dawei Li, Xiaolong Wang, and Hairong Qi. Talking face generation by conditional recurrent adversarial network. *IJCAI*, 2019. 2", + "[60] Jianxin Sun, Qiyao Deng, Qi Li, Muyi Sun, Min Ren, and Zhenan Sun. AnyFace: Free-style text-to-face synthesis and manipulation. In CVPR, 2022. 2", + "[61] Aaron Van Den Oord, Oriol Vinyals, et al. Neural discrete representation learning. In NeurIPS, 2017. 2", + "[62] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In NeurIPS, 2017. 5", + "[63] Ruben Villegas, Mohammad Babaeizadeh, Pieter-Jan Kindermans, Hernan Moraldo, Han Zhang, Mohammad Taghi Saffar, Santiago Castro, Julius Kunze, and Dumitru Erhan. Phenaki: Variable length video generation from open domain textual description. arXiv preprint arXiv:2210.02399, 2022. 2", + "[64] Pascal Vincent. A connection between score matching and denoising autoencoders. Neural Computation, 2011. 2", + "[65] Tengfei Wang, Ting Zhang, Bo Zhang, Hao Ouyang, Dong Chen, Qifeng Chen, and Fang Wen. Pretraining is all you need for image-to-image translation. arXiv preprint arXiv:2205.12952, 2022. 1, 2", + "[66] Weilun Wang, Jianmin Bao, Wengang Zhou, Dongdong Chen, Dong Chen, Lu Yuan, and Houqiang Li. Semantic image synthesis via diffusion models. arXiv preprint arXiv:2207.00050, 2022. 1" + ], + "bbox": [ + 501, + 92, + 893, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "6089", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[67] Weihao Xia, Yujiu Yang, Jing-Hao Xue, and Baoyuan Wu. TediGAN: Text-guided diverse face image generation and manipulation. In CVPR, 2021. 2, 5, 6", + "[68] Weihao Xia, Yujiu Yang, Jing-Hao Xue, and Baoyuan Wu. Towards open-world text-guided face image generation and manipulation. arXiv preprint arXiv:2104.08910, 2021. 2, 5, 6", + "[69] Tao Xu, Pengchuan Zhang, Qiuyuan Huang, Han Zhang, Zhe Gan, Xiaolei Huang, and Xiaodong He. AttnGAN: Fine-grained text to image generation with attentional generative adversarial networks. In CVPR, 2018. 2", + "[70] Han Zhang, Tao Xu, Hongsheng Li, Shaoting Zhang, Xiaogang Wang, Xiaolei Huang, and Dimitris N Metaxas. StackGAN: Text to photo-realistic image synthesis with stacked generative adversarial networks. In ICCV, 2017. 2", + "[71] Han Zhang, Tao Xu, Hongsheng Li, Shaoting Zhang, Xiaogang Wang, Xiaolei Huang, and Dimitris N Metaxas. StackGAN++: Realistic image synthesis with stacked generative adversarial networks. IEEE TPAMI, 2018. 2" + ], + "bbox": [ + 78, + 90, + 470, + 345 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "6090", + "bbox": [ + 482, + 945, + 516, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2023/Collaborative Diffusion for Multi-Modal Face Generation and Editing/3b7ffa39-2244-45b9-9833-f768a1b5b8c6_model.json b/2023/Collaborative Diffusion for Multi-Modal Face Generation and Editing/3b7ffa39-2244-45b9-9833-f768a1b5b8c6_model.json new file mode 100644 index 0000000000000000000000000000000000000000..1ece78e710e0403c8c9f7c280dbcab5739296e87 --- /dev/null +++ b/2023/Collaborative Diffusion for Multi-Modal Face Generation and Editing/3b7ffa39-2244-45b9-9833-f768a1b5b8c6_model.json @@ -0,0 +1,3258 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.809, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.133, + 0.093, + 0.838, + 0.116 + ], + "angle": 0, + "content": "Collaborative Diffusion for Multi-Modal Face Generation and Editing" + }, + { + "type": "text", + "bbox": [ + 0.231, + 0.144, + 0.737, + 0.179 + ], + "angle": 0, + "content": "Ziqi Huang Kelvin C.K. Chan Yuming Jiang Ziwei Liu S-Lab, Nanyang Technological University" + }, + { + "type": "text", + "bbox": [ + 0.256, + 0.181, + 0.714, + 0.197 + ], + "angle": 0, + "content": "{ziqi002, chan0899, yuming002, ziwei.liu}@ntu.edu.sg" + }, + { + "type": "image", + "bbox": [ + 0.083, + 0.23, + 0.891, + 0.495 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.505, + 0.893, + 0.549 + ], + "angle": 0, + "content": "Figure 1. We propose Collaborative Diffusion, where users can use multiple modalities to control face generation and editing. (a) Face Generation. Given multi-modal controls, our framework synthesizes high-quality images consistent with the input conditions. (b) Face Editing. Collaborative Diffusion also supports multi-modal editing of real images with promising identity preservation capability." + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.566, + 0.314, + 0.582 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.59, + 0.471, + 0.711 + ], + "angle": 0, + "content": "Diffusion models arise as a powerful generative tool recently. Despite the great progress, existing diffusion models mainly focus on uni-modal control, i.e., the diffusion process is driven by only one modality of condition. To further unleash the users' creativity, it is desirable for the model to be controllable by multiple modalities simultaneously, e.g. generating and editing faces by describing the age (text-driven) while drawing the face shape (mask-driven)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.712, + 0.471, + 0.849 + ], + "angle": 0, + "content": "In this work, we present Collaborative Diffusion, where pre-trained uni-modal diffusion models collaborate to achieve multi-modal face generation and editing without re-training. Our key insight is that diffusion models driven by different modalities are inherently complementary regarding the latent denoising steps, where bilateral connections can be established upon. Specifically, we propose dynamic diffuser, a meta-network that adaptively hallucinates multi-modal denoising steps by predicting the spatial-temporal" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.568, + 0.894, + 0.674 + ], + "angle": 0, + "content": "influence functions for each pre-trained uni-modal model. Collaborative Diffusion not only collaborates generation capabilities from uni-modal diffusion models, but also integrates multiple uni-modal manipulations to perform multimodal editing. Extensive qualitative and quantitative experiments demonstrate the superiority of our framework in both image quality and condition consistency." + }, + { + "type": "title", + "bbox": [ + 0.502, + 0.707, + 0.631, + 0.722 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.733, + 0.894, + 0.884 + ], + "angle": 0, + "content": "Recent years have witnessed substantial progress in image synthesis and editing with the surge of diffusion models [9, 20, 56, 58]. In addition to the remarkable synthesis quality, one appealing property of diffusion models is the flexibility of conditioning on various modalities, such as texts [3, 16, 29, 30, 39, 48, 51], segmentation masks [48, 65, 66], and sketches [7, 65]. However, existing explorations are largely confined to the use of a single modality at a time. The exploitation of multiple conditions remains under-explored. As a generative tool, its controllability is still limited." + }, + { + "type": "text", + "bbox": [ + 0.52, + 0.886, + 0.892, + 0.901 + ], + "angle": 0, + "content": "To unleash users' creativity, it is desirable that the model" + }, + { + "type": "page_footnote", + "bbox": [ + 0.083, + 0.865, + 0.449, + 0.901 + ], + "angle": 0, + "content": "Corresponding author. \nProject page: https://ziqihuangg.github.io/projects/collaborative-diffusion.html \nCode: https://github.com/ziqihuangg/Collaborative-Diffusion" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6080" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.272 + ], + "angle": 0, + "content": "is simultaneously controllable by multiple modalities. While it is trivial to extend the current supervised framework with multiple modalities, training a large-scale model from scratch is computationally expensive, especially when extensive hyper-parameter tuning and delicate architecture designs are needed. More importantly, each trained model could only accept a fixed combination of modalities, and hence re-training is necessary when a subset of modalities are absent, or when additional modalities become available. The above demonstrates the necessity of a unified framework that effectively exploits pre-trained models and integrate them for multi-modal synthesis and editing." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.273, + 0.473, + 0.605 + ], + "angle": 0, + "content": "In this paper, we propose Collaborative Diffusion, a framework that synergizes pre-trained uni-modal diffusion models for multi-modal face generation and editing without the need of re-training. Motivated by the fact that different modalities are complementary to each other (e.g., text for age and mask for hair shape), we explore the possibility of establishing lateral connections between models driven by different modalities. We propose dynamic diffuser to adaptively predict the spatial-temporal influence function for each pre-trained model. The dynamic diffuser dynamically determines spatial-varying and temporal-varying influences of each model, suppressing contributions from irrelevant modalities while enhancing contributions from admissible modalities. In addition to multi-modal synthesis, the simplicity and flexibility of our framework enable extension to multi-modal face editing with minimal modifications. In particular, the dynamic diffuser is first trained for collaborative synthesis. It is then fixed and combined with existing face editing approaches [29, 50] for multi-modal editing. It is worth-mentioning that users can select the best editing approaches based on their needs without the need of altering the dynamic diffusers." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.606, + 0.472, + 0.664 + ], + "angle": 0, + "content": "We demonstrate both qualitatively and quantitatively that our method achieves superior image quality and condition consistency in both synthesis and editing tasks. Our contributions can be summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.667, + 0.47, + 0.742 + ], + "angle": 0, + "content": "- We introduce Collaborative Diffusion, which exploits pre-trained uni-modal diffusion models for multi-modal controls without re-training. Our approach is the first attempt towards flexible integration of uni-modal diffusion models into a single collaborative framework." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.746, + 0.47, + 0.821 + ], + "angle": 0, + "content": "Tailored for the iterative property of diffusion models, we propose dynamic diffuser, which predicts the spatial-varying and temporal-varying influence functions to selectively enhance or suppress the contributions of the given modalities at each iterative step." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.826, + 0.47, + 0.901 + ], + "angle": 0, + "content": "- We demonstrate the flexibility of our framework by extending it to face editing driven by multiple modalities. Both quantitative and qualitative results demonstrate the superiority of Collaborative Diffusion in multi-modal face generation and editing." + }, + { + "type": "list", + "bbox": [ + 0.097, + 0.667, + 0.47, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.09, + 0.642, + 0.107 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.115, + 0.895, + 0.356 + ], + "angle": 0, + "content": "Diffusion Models. Diffusion models [20, 56, 58] have recently become a mainstream approach for image synthesis [9, 11, 38] apart from Generative Adversarial Networks (GANs) [14], and success has also been found in various domains including video generation [17, 19, 55, 63], image restoration [21, 52], semantic segmentation [1, 4, 15], and natural language processing [2]. In the diffusion-based framework, models are trained with score-matching objectives [22, 64] at various noise levels, and sampling is done via iterative denoising. Existing works focus on improving the performance and efficiency of diffusion models through enhanced architecture designs [16, 48] and sampling schemes [57]. In contrast, this work focuses on exploiting existing models, and providing a succinct framework for multi-modal synthesis and editing without large-scale re-training of models." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.358, + 0.897, + 0.629 + ], + "angle": 0, + "content": "Face Generation. Existing face generation approaches can be divided into three main directions. Following the GAN paradigm, the StyleGAN series [26-28] boost the quality of facial synthesis, and provide an interpretable latent space for steerable style controls and manipulations. The vector-quantized approaches [12, 61] learn a discrete codebook by mapping the input images into a low-dimensional discrete feature space. The learned codebook is then sampled, either sequentially [12, 61] or parallelly [5, 6, 16], for synthesis. In contrast to the previous two approaches, diffusion models are trained with a stationary objective, without the need of optimizing complex losses (e.g., adversarial loss) or balancing multiple objectives (e.g., codebook loss versus reconstruction loss). With training simplicity as a merit, diffusion-based approaches have become increasingly popular in recent years. Our framework falls in the diffusion-based paradigm. In particular, we leverage pre-trained diffusion models for multi-modal generation and editing." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.63, + 0.897, + 0.901 + ], + "angle": 0, + "content": "Conditional Face Generation and Editing. Conditional generation [10, 11, 13, 31, 34, 37, 40, 43-46, 51, 65, 67-71] and editing [8, 32, 41, 53, 54, 67, 68] is an active line of research focusing on conditioning generative models on different modalities, such as texts [24, 41, 67, 68], segmentation masks [32, 33, 40, 47], and audios [59]. For example, StyleCLIP [41], DiffusionCLIP [30], and many others [35, 60] have demonstrated remarkable performance in text-guided face generation and editing. However, most existing models do not support simultaneous conditioning on multiple modalities (e.g., text and mask at the same time), and supporting additional modalities often requires time-consuming model re-training and extensive hyper-parameter tuning, which are not preferable in general. In this work, we propose Collaborative Diffusion to exploit pre-trained uni-modal diffusion models [48] (e.g., text-driven and mask-driven models) to achieve multi-modal conditioning without model retraining." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.514, + 0.957 + ], + "angle": 0, + "content": "6081" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.08, + 0.089, + 0.899, + 0.355 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.362, + 0.895, + 0.406 + ], + "angle": 0, + "content": "Figure 2. Overview of Collaborative Diffusion. We use pre-trained uni-modal diffusion models to perform multi-modal guided face generation and editing. At each step of the reverse process (i.e., from timestep \\( t \\) to \\( t - 1 \\)), the dynamic diffuser predicts the spatial-varying and temporal-varying influence function to selectively enhance or suppress the contributions of the given modality." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.442, + 0.3, + 0.459 + ], + "angle": 0, + "content": "3. Collaborative Diffusion" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.472, + 0.475, + 0.61 + ], + "angle": 0, + "content": "We propose Collaborative Diffusion, which exploits multiple pre-trained uni-modal diffusion models (Section 3.1) for multi-modal generation and editing. The key of our framework is the dynamic diffuser, which adaptively predicts the influence functions to enhance or suppress the contributions of the pre-trained models based on the spatial-temporal influences of the modalities. Our framework is compatible with most existing approaches for both multi-modal guided synthesis (Section 3.2) and multi-modal editing (Section 3.3)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.632, + 0.43, + 0.648 + ], + "angle": 0, + "content": "3.1. Uni-Modal Conditional Diffusion Models" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.661, + 0.473, + 0.753 + ], + "angle": 0, + "content": "Diffusion models are a class of generative models that model the data distribution in the form of \\( p_{\\theta}(\\mathbf{x}_0) \\coloneqq \\int p_{\\theta}(\\mathbf{x}_{0:T}) d\\mathbf{x}_{1:T} \\). The diffusion process (a.k.a. forward process) gradually adds Gaussian noise to the data and eventually corrupts the data \\( \\mathbf{x}_0 \\) into an approximately pure Gaussian noise \\( \\mathbf{x}_T \\) using a variance schedule \\( \\beta_1, \\ldots, \\beta_T \\):" + }, + { + "type": "equation", + "bbox": [ + 0.132, + 0.773, + 0.471, + 0.814 + ], + "angle": 0, + "content": "\\[\nq \\left(\\mathbf {x} _ {1: T} \\mid \\mathbf {x} _ {0}\\right) := \\prod_ {t = 1} ^ {T} q \\left(\\mathbf {x} _ {t} \\mid \\mathbf {x} _ {t - 1}\\right), \\tag {1}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.133, + 0.817, + 0.414, + 0.837 + ], + "angle": 0, + "content": "\\[\nq (\\mathbf {x} _ {t} | \\mathbf {x} _ {t - 1}) := \\mathcal {N} (\\mathbf {x} _ {t}; \\sqrt {1 - \\beta_ {t}} \\mathbf {x} _ {t - 1}, \\beta_ {t} \\mathbf {I}).\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Reversing the forward process allows sampling new data \\(\\mathbf{x}_0\\) by starting from \\(p(\\mathbf{x}_T) = \\mathcal{N}(\\mathbf{x}_T; \\mathbf{0}, \\mathbf{I})\\). The reverse process is defined as a Markov chain where each step is a learned" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.443, + 0.7, + 0.46 + ], + "angle": 0, + "content": "Gaussian transition \\((\\pmb{\\mu}_{\\theta},\\pmb{\\Sigma}_{\\theta})\\)" + }, + { + "type": "equation", + "bbox": [ + 0.537, + 0.466, + 0.893, + 0.507 + ], + "angle": 0, + "content": "\\[\np _ {\\theta} \\left(\\mathbf {x} _ {0: T}\\right) := p \\left(\\mathbf {x} _ {T}\\right) \\prod_ {t = 1} ^ {T} p _ {\\theta} \\left(\\mathbf {x} _ {t - 1} \\mid \\mathbf {x} _ {t}\\right), \\tag {2}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.539, + 0.51, + 0.852, + 0.527 + ], + "angle": 0, + "content": "\\[\np _ {\\theta} \\left(\\mathbf {x} _ {t - 1} \\mid \\mathbf {x} _ {t}\\right) := \\mathcal {N} \\left(\\mathbf {x} _ {t - 1}; \\boldsymbol {\\mu} _ {\\theta} \\left(\\mathbf {x} _ {t}, t\\right), \\boldsymbol {\\Sigma} _ {\\theta} \\left(\\mathbf {x} _ {t}, t\\right)\\right).\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.531, + 0.896, + 0.622 + ], + "angle": 0, + "content": "Training diffusion models relies on minimizing the variational bound on \\( p(x) \\)'s negative log-likelihood. The commonly used optimization objective \\( L_{\\mathrm{DM}} \\) [20] reparameterizes the learnable Gaussian transition as \\( \\epsilon_{\\theta}(\\cdot) \\), and temporally reweights the variational bound to trade for better sample quality:" + }, + { + "type": "equation", + "bbox": [ + 0.541, + 0.627, + 0.893, + 0.656 + ], + "angle": 0, + "content": "\\[\nL _ {\\mathrm {D M}} (\\theta) := \\mathbb {E} _ {t, \\mathbf {x} _ {0}, \\boldsymbol {\\epsilon} \\sim \\mathcal {N} (\\mathbf {0}, \\mathbf {I})} \\left[ \\| \\boldsymbol {\\epsilon} - \\boldsymbol {\\epsilon} _ {\\theta} (\\mathbf {x} _ {t}, t) \\| ^ {2} \\right], \\quad (3)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.659, + 0.895, + 0.69 + ], + "angle": 0, + "content": "where \\(\\mathbf{x}_t\\) can be directly approximated by \\(\\mathbf{x}_t = \\sqrt{\\bar{\\alpha}_t}\\mathbf{x}_0 + \\sqrt{1 - \\bar{\\alpha}_t}\\pmb {\\epsilon}\\), with \\(\\bar{\\alpha}_{t}:= \\prod_{s = 1}^{t}\\alpha_{s}\\) and \\(\\alpha_{t}:= 1 - \\beta_{t}\\)." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.69, + 0.894, + 0.72 + ], + "angle": 0, + "content": "To sample data \\(\\mathbf{x}_0\\) from a trained diffusion model \\(\\epsilon_{\\theta}(\\cdot)\\), we iteratively denoise \\(\\mathbf{x}_t\\) from \\(t = T\\) to \\(t = 1\\) with noise \\(\\mathbf{z}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.536, + 0.725, + 0.893, + 0.76 + ], + "angle": 0, + "content": "\\[\n\\mathbf {x} _ {t - 1} = \\frac {1}{\\sqrt {\\alpha_ {t}}} \\left(\\mathbf {x} _ {t} - \\frac {1 - \\alpha_ {t}}{\\sqrt {1 - \\bar {\\alpha} _ {t}}} \\boldsymbol {\\epsilon} _ {\\theta} (\\mathbf {x} _ {t}, t)\\right) + \\sigma_ {t} \\mathbf {z} \\quad (4)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.763, + 0.895, + 0.902 + ], + "angle": 0, + "content": "The unconditional diffusion models can be extended to model conditional distributions \\( p_{\\theta}(\\mathbf{x}_0|c) \\), where \\( \\mathbf{x}_0 \\) is the image corresponding to the condition \\( c \\) such as class labels, segmentation masks, and text descriptions. The conditional diffusion model receives an additional input \\( \\tau(c) \\) and is trained by minimizing \\( \\| \\boldsymbol{\\epsilon} - \\boldsymbol{\\epsilon}_{\\theta}(\\mathbf{x}_t, t, \\tau(c)) \\|^2 \\), where \\( \\tau(\\cdot) \\) is an encoder that projects the condition \\( c \\) to an embedding \\( \\tau(c) \\). For brevity, we will use \\( c \\) to represent \\( \\tau(c) \\) in our subsequent discussions." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6082" + } + ], + [ + { + "type": "code_caption", + "bbox": [ + 0.079, + 0.106, + 0.348, + 0.122 + ], + "angle": 0, + "content": "Algorithm 1 Dynamic Diffuser Training" + }, + { + "type": "algorithm", + "bbox": [ + 0.084, + 0.126, + 0.445, + 0.331 + ], + "angle": 0, + "content": "1: repeat \n2: \\(\\mathbf{x}_0, c_1, c_2, \\ldots, c_M \\sim q(\\mathbf{x}_0, c_1, c_2, \\ldots, c_M)\\) \n3: \\(t \\sim \\mathrm{Uniform}(\\{1, \\ldots, T\\})\\) \n4: \\(\\epsilon \\sim \\mathcal{N}(\\mathbf{0}, \\mathbf{I})\\) \n5: for \\(m = 1, \\ldots, M\\) do \n6: \\(\\epsilon_{pred,m,t} = \\epsilon_{\\theta_m}(\\sqrt{\\bar{\\alpha}_t}\\mathbf{x}_0 + \\sqrt{1 - \\bar{\\alpha}_t}\\boldsymbol{\\epsilon}, t, c_m)\\) \n7: \\(\\mathbf{I}_{m,t} = \\mathbf{D}_{\\phi_m}(\\sqrt{\\bar{\\alpha}_t}\\mathbf{x}_0 + \\sqrt{1 - \\bar{\\alpha}_t}\\boldsymbol{\\epsilon}, t, c_m)\\) \n8: end for \n9: \\(\\hat{\\mathbf{I}}_{m,t,p} = \\frac{\\exp(\\mathbf{I}_{m,t,p})}{\\sum_{j=1}^{M}\\exp(\\mathbf{I}_{j,t,p})}\\), softmax at each pixel \\(p\\) \n10: \\(\\epsilon_{pred,t} = \\sum_{m=1}^{M}\\hat{\\mathbf{I}}_{m,t} \\odot \\epsilon_{pred,m,t}\\) \n11: Take gradient descent step on \\(\\nabla_{\\phi} \\| \\boldsymbol{\\epsilon} - \\boldsymbol{\\epsilon}_{pred,t} \\|^2\\) where \\(\\phi = \\{\\phi_m | m = 1, \\ldots, M\\}\\) \n12: until converged" + }, + { + "type": "image", + "bbox": [ + 0.137, + 0.345, + 0.41, + 0.466 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.476, + 0.471, + 0.547 + ], + "angle": 0, + "content": "Figure 3. Dynamic Diffuser. It predicts the influence function that represents the desired level of contribution from each pre-trained diffusion model. The influence function varies at different timestep \\( t \\) and different spatial locations, and is dependent on the state of diffusion \\( \\mathbf{x}_t \\) and the condition input \\( c_m \\)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.585, + 0.398, + 0.601 + ], + "angle": 0, + "content": "3.2. Multi-Modal Collaborative Synthesis" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.615, + 0.471, + 0.707 + ], + "angle": 0, + "content": "In our Collaborative Diffusion, multiple uni-modal diffusion models collaborate at each step of the denoising process for multi-modal guided synthesis. The core of our framework is the dynamic diffuser, which determines the extent of contribution from each collaborator by predicting the spatial-temporal influence functions." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.713, + 0.471, + 0.79 + ], + "angle": 0, + "content": "Problem Formulation. Given \\(M\\) pre-trained uni-modal conditional diffusion models \\(\\{\\epsilon_{\\theta_m}\\}\\) (each models the distribution \\(p(\\mathbf{x}_0|c_m)\\)), where the modality index \\(m = 1, \\ldots, M\\), our objective is to sample from \\(p(\\mathbf{x}_0|\\mathbf{c})\\), where \\(\\mathbf{c} = \\{c_1, c_2, \\dots, c_M\\}\\), without altering pre-trained models." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.795, + 0.473, + 0.903 + ], + "angle": 0, + "content": "Dynamic Diffuser. In diffusion models, each step of the reverse process requires predicting the noise \\(\\epsilon\\). With multiple diffusion models collaborating, we need to carefully determine when, where, and how much each diffusion model contributes to the prediction \\(\\epsilon\\). At each diffusion timestep \\(t = T, \\dots, 1\\), the influence \\(\\mathbf{I}_{m,t}\\) from each pre-trained diffusion model \\(\\epsilon_{\\theta_m}\\) is adaptively determined by a dynamic" + }, + { + "type": "code_caption", + "bbox": [ + 0.485, + 0.106, + 0.733, + 0.122 + ], + "angle": 0, + "content": "Algorithm 2 Collaborative Sampling" + }, + { + "type": "algorithm", + "bbox": [ + 0.492, + 0.129, + 0.815, + 0.328 + ], + "angle": 0, + "content": "1: \\(\\mathbf{x}_T\\sim \\mathcal{N}(\\mathbf{0},\\mathbf{I})\\) \n2: for \\(t = T,\\dots ,1\\) do \n3: \\(\\mathbf{z}\\sim \\mathcal{N}(\\mathbf{0},\\mathbf{I})\\) if \\(t > 1\\) , else \\(\\mathbf{z} = \\mathbf{0}\\) \n4: for \\(m = 1,\\ldots ,M\\) do \n5: \\(\\epsilon_{pred,m,t} = \\epsilon_{\\theta_m}(\\mathbf{x}_t,t,c_m)\\) \n6: \\(\\mathbf{I}_{m,t} = \\mathbf{D}_{\\phi_m}(\\mathbf{x}_t,t,c_m)\\) \n7: end for \n8: \\(\\hat{\\mathbf{I}}_{m,t,p} = \\frac{\\exp(\\mathbf{I}_{m,t,p})}{\\sum_{j = 1}^{M}\\exp(\\mathbf{I}_{j,t,p})}\\) , softmax at each pixel \\(p\\) \n9: \\(\\epsilon_{pred,t} = \\sum_{m = 1}^{M}\\hat{\\mathbf{I}}_{m,t}\\odot \\epsilon_{pred,m,t}\\) \n10: \\(\\mathbf{x}_{t - 1} = \\frac{1}{\\sqrt{\\alpha_t}}\\left(\\mathbf{x}_t - \\frac{1 - \\alpha_t}{\\sqrt{1 - \\bar{\\alpha}_t}}\\pmb{\\epsilon}_{pred,t}\\right) + \\sigma_t\\mathbf{z}\\) \n11: end for \n12: return \\({\\bf x}_0\\)" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.348, + 0.596, + 0.365 + ], + "angle": 0, + "content": "diffuser \\(\\mathbf{D}_{\\phi_m}\\)" + }, + { + "type": "equation", + "bbox": [ + 0.616, + 0.372, + 0.892, + 0.388 + ], + "angle": 0, + "content": "\\[\n\\mathbf {I} _ {m, t} = \\mathbf {D} _ {\\phi_ {m}} (\\mathbf {x} _ {t}, t, c _ {m}), \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.396, + 0.895, + 0.49 + ], + "angle": 0, + "content": "where \\(m = 1,\\dots,M\\) is the index of the modalities, \\(\\mathbf{I}_{m,t}\\in\\) \\(\\mathbb{R}^{h\\times w}\\) \\(\\mathbf{x}_t\\) is the noisy image at time \\(t\\), \\(c_{m}\\) is the condition of the \\(m^{th}\\) modality, and \\(\\mathbf{D}_{\\phi_m}\\) is the dynamic diffuser implemented by a UNet [49]. To regularize the overall influence strength, we perform softmax across all modalities' \\(\\mathbf{I}_{m,t}\\) at each pixel \\(p\\) to obtain the final influence function \\(\\hat{\\mathbf{I}}_{m,t}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.602, + 0.497, + 0.892, + 0.535 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathbf {I}} _ {m, t, p} = \\frac {\\exp \\left(\\mathbf {I} _ {m , t , p}\\right)}{\\sum_ {j = 1} ^ {M} \\exp \\left(\\mathbf {I} _ {j , t , p}\\right)}. \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.541, + 0.895, + 0.587 + ], + "angle": 0, + "content": "Multi-Modal Collaboration. We use the \\(M\\) learned influence functions \\(\\hat{\\mathbf{I}}_{m,t}\\) to control the contribution from each pre-trained diffusion model at each denoising step:" + }, + { + "type": "equation", + "bbox": [ + 0.574, + 0.594, + 0.892, + 0.636 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {\\epsilon} _ {p r e d, t} = \\sum_ {m = 1} ^ {M} \\hat {\\mathbf {I}} _ {m, t} \\odot \\boldsymbol {\\epsilon} _ {\\theta_ {m}} \\left(x _ {t}, t, c _ {m}\\right) \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.644, + 0.894, + 0.72 + ], + "angle": 0, + "content": "where \\(\\epsilon_{\\theta_m}\\) is the \\(m^{th}\\) collaborator, and \\(\\odot\\) denotes pixel-wise multiplication. The complete training procedure of dynamic diffusers and the image sampling strategy in our Collaborative Diffusion framework are detailed in Algorithm 1 and Algorithm 2, respectively." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.895, + 0.901 + ], + "angle": 0, + "content": "Relation with Composable Diffusion. Composable Diffusion [36] combines two instances of the same text-to-image diffusion model to achieve compositional visual generation via intermediate results addition. Our framework is related to Composable Diffusion in terms of composing diffusion models for image synthesis, but is substantially different in terms of the task nature and methodology. Composable Diffusion aims to decompose the text condition into elementary segments to factorize the conditional synthesis problem, while we aim to integrate uni-modal collaborators to achieve multi-modal controls. Furthermore, different from Composable Diffusion, which can only compose instances of the" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6083" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.079, + 0.106, + 0.31, + 0.122 + ], + "angle": 0, + "content": "Algorithm 3 Collaborative Editing" + }, + { + "type": "title", + "bbox": [ + 0.081, + 0.13, + 0.137, + 0.142 + ], + "angle": 0, + "content": "Require:" + }, + { + "type": "text", + "bbox": [ + 0.103, + 0.144, + 0.482, + 0.198 + ], + "angle": 0, + "content": "input image \\(\\mathbf{x}_{input}\\), target conditions \\(c_{m,target}\\), diffusion models \\(\\epsilon_{\\theta_m}\\), dynamic diffusers \\(\\mathbf{D}_{\\phi_m}\\), \\((m = 1, \\dots, M)\\), interpolation scale \\(\\alpha\\)" + }, + { + "type": "text", + "bbox": [ + 0.087, + 0.202, + 0.482, + 0.216 + ], + "angle": 0, + "content": "1: for \\(m = 1,\\ldots ,M\\) do \\(\\triangleright\\) Uni-Modal Editing" + }, + { + "type": "text", + "bbox": [ + 0.087, + 0.218, + 0.208, + 0.23 + ], + "angle": 0, + "content": "2: \\( c_{m} = c_{m, \\text{target}} \\)" + }, + { + "type": "text", + "bbox": [ + 0.087, + 0.23, + 0.297, + 0.243 + ], + "angle": 0, + "content": "3: \\(\\mathbf{x}_t = \\sqrt{\\bar{\\alpha}_t}\\mathbf{x}_{input} + \\sqrt{1 - \\bar{\\alpha}_t}\\pmb{\\epsilon}\\)" + }, + { + "type": "text", + "bbox": [ + 0.087, + 0.243, + 0.402, + 0.258 + ], + "angle": 0, + "content": "4: \\(c_{m,opt} = \\mathrm{argmin}_{c_m}\\mathbb{E}_{\\epsilon ,t}\\left\\| \\boldsymbol {\\epsilon} - \\boldsymbol{\\epsilon}_{\\theta_m}(\\mathbf{x}_t,t,c_m)\\right\\| ^2\\)" + }, + { + "type": "text", + "bbox": [ + 0.087, + 0.258, + 0.423, + 0.274 + ], + "angle": 0, + "content": "5: \\(\\theta_{m,opt} = \\operatorname*{argmin}_{\\theta_m}\\mathbb{E}_{\\epsilon ,t}\\| \\pmb {\\epsilon} - \\pmb{\\epsilon}_{\\theta_m}(\\mathbf{x}_t,t,c_{m,opt})\\|^2\\)" + }, + { + "type": "text", + "bbox": [ + 0.087, + 0.274, + 0.364, + 0.287 + ], + "angle": 0, + "content": "6: \\(c_{m,int} = \\alpha \\cdot c_{m,target} + (1 - \\alpha) \\cdot c_{m,opt}\\)" + }, + { + "type": "text", + "bbox": [ + 0.087, + 0.288, + 0.153, + 0.3 + ], + "angle": 0, + "content": "7: end for" + }, + { + "type": "list", + "bbox": [ + 0.087, + 0.202, + 0.482, + 0.3 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.087, + 0.304, + 0.482, + 0.318 + ], + "angle": 0, + "content": "8: \\(\\mathbf{x}_T\\sim \\mathcal{N}(\\mathbf{0},\\mathbf{I})\\) ▷Collaborate the Uni-Modal Edits" + }, + { + "type": "text", + "bbox": [ + 0.087, + 0.319, + 0.226, + 0.331 + ], + "angle": 0, + "content": "9: for \\( t = T, \\dots, 1 \\) do" + }, + { + "type": "text", + "bbox": [ + 0.082, + 0.332, + 0.31, + 0.345 + ], + "angle": 0, + "content": "10: \\(\\mathbf{z} \\sim \\mathcal{N}(\\mathbf{0}, \\mathbf{I})\\) if \\(t > 1\\), else \\(\\mathbf{z} = \\mathbf{0}\\)" + }, + { + "type": "text", + "bbox": [ + 0.082, + 0.346, + 0.239, + 0.358 + ], + "angle": 0, + "content": "11: for \\(m = 0,\\dots,M\\) do" + }, + { + "type": "text", + "bbox": [ + 0.082, + 0.36, + 0.34, + 0.375 + ], + "angle": 0, + "content": "12: \\(\\pmb{\\epsilon}_{pred,m,t} = \\pmb{\\epsilon}_{\\theta_{m,opt}}(\\mathbf{x}_t,t,c_{m,int})\\)" + }, + { + "type": "text", + "bbox": [ + 0.082, + 0.377, + 0.294, + 0.392 + ], + "angle": 0, + "content": "13: \\(\\mathbf{I}_{m,t} = \\mathbf{D}_{\\phi_m}(\\mathbf{x}_t,t,c_{m,int})\\)" + }, + { + "type": "text", + "bbox": [ + 0.082, + 0.395, + 0.16, + 0.407 + ], + "angle": 0, + "content": "14: end for" + }, + { + "type": "text", + "bbox": [ + 0.082, + 0.409, + 0.407, + 0.432 + ], + "angle": 0, + "content": "15: \\(\\hat{\\mathbf{I}}_{m,t,p} = \\frac{\\exp(\\mathbf{I}_{m,t,p})}{\\sum_{j = 1}^{M}\\exp(\\mathbf{I}_{j,t,p})}\\), softmax at each pixel \\(p\\)" + }, + { + "type": "text", + "bbox": [ + 0.082, + 0.435, + 0.319, + 0.452 + ], + "angle": 0, + "content": "16: \\(\\pmb{\\epsilon}_{pred,t} = \\sum_{m=1}^{M} \\hat{\\mathbf{I}}_{m,t} \\odot \\pmb{\\epsilon}_{pred,m,t}\\)" + }, + { + "type": "text", + "bbox": [ + 0.082, + 0.455, + 0.369, + 0.476 + ], + "angle": 0, + "content": "17: \\(\\mathbf{x}_{t - 1} = \\frac{1}{\\sqrt{\\alpha_t}}\\left(\\mathbf{x}_t - \\frac{1 - \\alpha_t}{\\sqrt{1 - \\bar{\\alpha}_t}}\\pmb {\\epsilon}_{pred,t}\\right) + \\sigma_t\\mathbf{z}\\)" + }, + { + "type": "text", + "bbox": [ + 0.082, + 0.476, + 0.153, + 0.487 + ], + "angle": 0, + "content": "18: end for" + }, + { + "type": "text", + "bbox": [ + 0.082, + 0.489, + 0.167, + 0.501 + ], + "angle": 0, + "content": "19: return \\(\\mathbf{x}_0\\)" + }, + { + "type": "list", + "bbox": [ + 0.082, + 0.304, + 0.482, + 0.501 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.527, + 0.472, + 0.603 + ], + "angle": 0, + "content": "same text-based synthesis model (i.e., the weights of the diffusion models being composed are the same), our framework possesses the flexibility in integrating models with different weights, architectures, and modalities through the learned dynamic diffusers." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.618, + 0.383, + 0.634 + ], + "angle": 0, + "content": "3.3. Multi-Modal Collaborative Editing" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.643, + 0.47, + 0.734 + ], + "angle": 0, + "content": "In addition to face synthesis, our framework is also capable of combining multiple facial manipulations, each of which is guided by a different modality, into a collaborative edit. Our collaborative framework in theory can integrate any existing uni-modal diffusion-based editing approach for collaborative editing." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.735, + 0.471, + 0.902 + ], + "angle": 0, + "content": "In this work, we demonstrate such possibility by extending Imagic [29] to the multi-modal paradigm. We first follow Imagic to fine-tune the embeddings and models to better capture the input face identity during editing. The trained dynamic diffusers discussed in Section 3.2 are then used to combine the fine-tuned models. Algorithm 3 displays the complete procedure of collaborative editing. Note that finetuning the pre-trained models is for identity preservation proposed in Imagic, which is independent to our framework. The extension of our framework to editing requires no further training of dynamic diffusers." + }, + { + "type": "image_caption", + "bbox": [ + 0.553, + 0.09, + 0.647, + 0.099 + ], + "angle": 0, + "content": "Multi-Modal Conditions" + }, + { + "type": "image", + "bbox": [ + 0.504, + 0.1, + 0.574, + 0.154 + ], + "angle": 0, + "content": null + }, + { + "type": "image_footnote", + "bbox": [ + 0.58, + 0.11, + 0.666, + 0.145 + ], + "angle": 0, + "content": "He is a young adult. This man has a rough growth of stubble." + }, + { + "type": "image_footnote", + "bbox": [ + 0.581, + 0.166, + 0.663, + 0.201 + ], + "angle": 0, + "content": "His face is covered with short beard. \nThis man is in the thirties." + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.1, + 0.666, + 0.201 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.504, + 0.211, + 0.574, + 0.265 + ], + "angle": 0, + "content": null + }, + { + "type": "image_footnote", + "bbox": [ + 0.58, + 0.225, + 0.659, + 0.252 + ], + "angle": 0, + "content": "There is no beard. This woman looks extremely young." + }, + { + "type": "image_caption", + "bbox": [ + 0.744, + 0.09, + 0.815, + 0.099 + ], + "angle": 0, + "content": "Generated Images" + }, + { + "type": "image", + "bbox": [ + 0.672, + 0.1, + 0.743, + 0.154 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.746, + 0.101, + 0.814, + 0.154 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.672, + 0.155, + 0.743, + 0.21 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.817, + 0.101, + 0.886, + 0.154 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.746, + 0.155, + 0.815, + 0.209 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.817, + 0.156, + 0.885, + 0.21 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.672, + 0.211, + 0.743, + 0.265 + ], + "angle": 0, + "content": null + }, + { + "type": "image_footnote", + "bbox": [ + 0.69, + 0.266, + 0.806, + 0.275 + ], + "angle": 0, + "content": "TediGAN Composable" + }, + { + "type": "image", + "bbox": [ + 0.817, + 0.211, + 0.886, + 0.265 + ], + "angle": 0, + "content": null + }, + { + "type": "image_footnote", + "bbox": [ + 0.841, + 0.266, + 0.862, + 0.275 + ], + "angle": 0, + "content": "Ours" + }, + { + "type": "image_caption", + "bbox": [ + 0.5, + 0.277, + 0.893, + 0.318 + ], + "angle": 0, + "content": "Figure 4. Qualitative Comparison of Face Generation. In the second example, TediGAN and Composable fails to follow mask, while ours generates results highly consistent to both conditions." + }, + { + "type": "image_caption", + "bbox": [ + 0.511, + 0.328, + 0.553, + 0.337 + ], + "angle": 0, + "content": "Input Image" + }, + { + "type": "image", + "bbox": [ + 0.504, + 0.338, + 0.564, + 0.383 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.586, + 0.328, + 0.688, + 0.337 + ], + "angle": 0, + "content": "Multi-Modal Target Conditions" + }, + { + "type": "image", + "bbox": [ + 0.568, + 0.338, + 0.626, + 0.383 + ], + "angle": 0, + "content": null + }, + { + "type": "image_footnote", + "bbox": [ + 0.631, + 0.346, + 0.702, + 0.376 + ], + "angle": 0, + "content": "He is a teen. The face is covered with short pointed beard." + }, + { + "type": "image", + "bbox": [ + 0.504, + 0.384, + 0.563, + 0.429 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.568, + 0.384, + 0.626, + 0.429 + ], + "angle": 0, + "content": null + }, + { + "type": "image_footnote", + "bbox": [ + 0.631, + 0.393, + 0.698, + 0.422 + ], + "angle": 0, + "content": "This man has \nheard of medium \nlength. He is in his \nthirties." + }, + { + "type": "image_caption", + "bbox": [ + 0.775, + 0.328, + 0.822, + 0.337 + ], + "angle": 0, + "content": "Edited Images" + }, + { + "type": "image", + "bbox": [ + 0.707, + 0.338, + 0.768, + 0.383 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.768, + 0.338, + 0.825, + 0.383 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.829, + 0.338, + 0.886, + 0.383 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.709, + 0.384, + 0.768, + 0.429 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.723, + 0.43, + 0.754, + 0.438 + ], + "angle": 0, + "content": "TediGAN" + }, + { + "type": "image", + "bbox": [ + 0.77, + 0.384, + 0.827, + 0.429 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.777, + 0.43, + 0.818, + 0.438 + ], + "angle": 0, + "content": "Composable" + }, + { + "type": "image", + "bbox": [ + 0.829, + 0.384, + 0.886, + 0.429 + ], + "angle": 0, + "content": null + }, + { + "type": "image_footnote", + "bbox": [ + 0.848, + 0.43, + 0.868, + 0.438 + ], + "angle": 0, + "content": "Ours" + }, + { + "type": "image_caption", + "bbox": [ + 0.5, + 0.439, + 0.893, + 0.495 + ], + "angle": 0, + "content": "Figure 5. Qualitative Comparison of Face Editing. While TediGAN's hair shape is inconsistent with mask, and Composable fails to generate beard according to text, our framework produces results highly consistent with both conditions while maintaining identity." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.512, + 0.632, + 0.53 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.538, + 0.69, + 0.554 + ], + "angle": 0, + "content": "4.1. Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.562, + 0.894, + 0.728 + ], + "angle": 0, + "content": "Datasets. CelebAMask-HQ [32] consists of manually annotated segmentation masks for each of the 30,000 images in the CelebA-HQ dataset [25]. Each mask has up to 19 classes, including the main facial components such as hair, skin, eyes, and nose, and accessories such as eyeglasses and cloth. CelebA-Dialog [23] provides rich and fine-grained natural language descriptions for the images in CelebA-HQ. In this work, we train the mask-driven diffusion models on CelebAMask-HQ, and the text-driven models on CelebA-Dialog. For dynamic diffuser, we simply combine the mask and text for each image without further processing." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.729, + 0.894, + 0.835 + ], + "angle": 0, + "content": "Implementation Details. We adopt LDM [48] for our unimodal diffusion models since it achieves a good balance between quality and speed. Our dynamic diffuser is a time-conditional UNet [49], with conditional embeddings injected via cross-attention [62]. The dynamic diffuser is \\(\\times 30\\) smaller than the pre-trained diffusion model. The detailed architecture and settings are provided in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.847, + 0.702, + 0.862 + ], + "angle": 0, + "content": "4.2. Comparison Methods" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.894, + 0.901 + ], + "angle": 0, + "content": "TediGAN [67, 68] is a StyleGAN-based face synthesis and manipulation method. It projects both text and mask condi" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6084" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.12, + 0.089, + 0.216, + 0.099 + ], + "angle": 0, + "content": "Multi-Modal Conditions" + }, + { + "type": "image", + "bbox": [ + 0.084, + 0.103, + 0.152, + 0.156 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.155, + 0.11, + 0.236, + 0.15 + ], + "angle": 0, + "content": "His face is covered with short beard. He is a young adult." + }, + { + "type": "image", + "bbox": [ + 0.085, + 0.16, + 0.153, + 0.213 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.155, + 0.178, + 0.22, + 0.195 + ], + "angle": 0, + "content": "She looks very young." + }, + { + "type": "image", + "bbox": [ + 0.248, + 0.102, + 0.317, + 0.155 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.319, + 0.102, + 0.387, + 0.155 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.391, + 0.102, + 0.459, + 0.155 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.462, + 0.102, + 0.529, + 0.155 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.532, + 0.091, + 0.602, + 0.099 + ], + "angle": 0, + "content": "Generated Images" + }, + { + "type": "image", + "bbox": [ + 0.532, + 0.103, + 0.6, + 0.155 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.605, + 0.102, + 0.671, + 0.155 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.674, + 0.102, + 0.741, + 0.155 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.747, + 0.102, + 0.813, + 0.155 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.818, + 0.102, + 0.885, + 0.155 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.248, + 0.159, + 0.315, + 0.212 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.318, + 0.159, + 0.387, + 0.212 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.391, + 0.159, + 0.459, + 0.212 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.462, + 0.159, + 0.528, + 0.212 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.532, + 0.159, + 0.599, + 0.212 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.604, + 0.159, + 0.671, + 0.212 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.675, + 0.159, + 0.743, + 0.212 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.747, + 0.159, + 0.814, + 0.212 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.818, + 0.159, + 0.885, + 0.213 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.227, + 0.892, + 0.255 + ], + "angle": 0, + "content": "Figure 6. Diversity of Face Generation. Given input conditions, our method demonstrates promising synthesis diversity in facial attributes that are not constrained by a specific combination of multi-modal conditions, such as hair color and skin tone." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.282, + 0.468, + 0.312 + ], + "angle": 0, + "content": "tions into StyleGAN's \\(\\mathcal{W}+\\) latent space, and performs style mixing to achieve multi-modal control." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.314, + 0.471, + 0.42 + ], + "angle": 0, + "content": "Composable Diffusion [36] parallelly implements two instances of the same text-to-image diffusion model to perform compositional scene generation. We generalize it by replacing the two text-to-image model instances by the text-driven model and the mask-driven model trained using LDM [48]. The two models are composed by averaging the intermediate diffusion results at each step of the reverse process." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.433, + 0.262, + 0.449 + ], + "angle": 0, + "content": "4.3. Evaluation Metrics" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.458, + 0.471, + 0.534 + ], + "angle": 0, + "content": "FID. We use Frechet Inception Distance (FID) [18] to measure the quality of images synthesized by different methods. FID computes the feature representation's distance between generated images and real images. A lower FID implies better sample quality." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.536, + 0.47, + 0.642 + ], + "angle": 0, + "content": "CLIP Score. CLIP [42] is a vision-language model trained on large-scale datasets. It uses an image encoder and a text encoder to project images and texts to a common feature space, respectively. The CLIP score is computed as the cosine similarity between the normalized image and text embeddings. A higher score usually indicates higher consistency between the output image and the text caption." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.643, + 0.47, + 0.734 + ], + "angle": 0, + "content": "Mask Accuracy. For each output image, we predict the segmentation mask using the face parsing network provided by CelebAMask-HQ [32]. Mask accuracy is the pixel-wise accuracy against the ground-truth segmentation. A higher average accuracy indicates better consistency between the output image and the segmentation mask." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.735, + 0.471, + 0.902 + ], + "angle": 0, + "content": "User Study. We conduct user study with 25 human evaluators to measure the effectiveness of the methods perceptually. For face generation, we randomly sample multi-modal conditions in the validation split of CelebA-HQ Dataset, and then synthesize output images given the conditions. Evaluators are provided with the input conditions and the output images, and they are asked to choose the best image based on 1) image photo-realism, 2) image-text consistency, 3) image-mask consistency. User study is conducted similarly for face editing, except that 1) evaluators are also provided with the input image for editing, and 2) they are also asked" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.28, + 0.892, + 0.335 + ], + "angle": 0, + "content": "Table 1. Quantitative Results of Face Generation. Compared with TediGAN and Composable Diffusion, our method synthesizes images with better quality (lower FID), and higher consistency with the text and mask conditions." + }, + { + "type": "table", + "bbox": [ + 0.512, + 0.347, + 0.882, + 0.409 + ], + "angle": 0, + "content": "
MethodFID ↓Text (%) ↑Mask (%) ↑
TediGAN [67,68]157.8124.2772.19
Composable [36]124.6223.9476.11
Ours111.3624.5180.25
" + }, + { + "type": "image", + "bbox": [ + 0.523, + 0.424, + 0.648, + 0.508 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.549, + 0.512, + 0.637, + 0.522 + ], + "angle": 0, + "content": "(a) Face Generation" + }, + { + "type": "image", + "bbox": [ + 0.656, + 0.424, + 0.782, + 0.508 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.692, + 0.512, + 0.765, + 0.523 + ], + "angle": 0, + "content": "(b) Face Editing" + }, + { + "type": "image_footnote", + "bbox": [ + 0.801, + 0.448, + 0.85, + 0.456 + ], + "angle": 0, + "content": "TediGAN" + }, + { + "type": "image_footnote", + "bbox": [ + 0.804, + 0.461, + 0.863, + 0.469 + ], + "angle": 0, + "content": "Composable" + }, + { + "type": "image_footnote", + "bbox": [ + 0.804, + 0.473, + 0.834, + 0.481 + ], + "angle": 0, + "content": "Ours" + }, + { + "type": "list", + "bbox": [ + 0.801, + 0.448, + 0.863, + 0.481 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.536, + 0.894, + 0.591 + ], + "angle": 0, + "content": "Figure 7. User Study. Among the three methods, the majority of users vote for our results as the best for both generation (69.40%) and editing (84.37%), in terms of image quality, condition consistency, and identity preservation." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.621, + 0.892, + 0.651 + ], + "angle": 0, + "content": "to assess the level of identity preservation between the edited image and input image." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.664, + 0.687, + 0.68 + ], + "angle": 0, + "content": "4.4. Comparative Study" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.688, + 0.895, + 0.884 + ], + "angle": 0, + "content": "Qualitative Comparison. In Figure 8, we provide the generation results of our Collaborative Diffusion under different combinations of texts and masks. It is observed that our framework is capable of synthesizing realistic outputs consistent with the multi-modal inputs, even for relatively rare conditions, such as a man with long hair. We then compare our method with TediGAN [67, 68] and Composable Diffusion [36] in Figure 4. For example, in the second example, TediGAN and Composable Diffusion's results are inconsistent with the mask, while ours is able to maintain both mask and text consistency. We also show in Figure 6 that our method synthesizes images with high diversity without losing condition consistency." + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.886, + 0.892, + 0.901 + ], + "angle": 0, + "content": "In addition to image generation, it is also observed in" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6085" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.084, + 0.089, + 0.887, + 0.476 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.488, + 0.893, + 0.517 + ], + "angle": 0, + "content": "Figure 8. Face Generation Results. Our method generates realistic images under different combinations of multi-modal conditions, even for relatively rare combinations in the training distribution, such as a man with long hair." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.543, + 0.47, + 0.663 + ], + "angle": 0, + "content": "Figure 9 that our framework is able to edit images based on multi-modal conditions while preserving identity. The comparison to existing works in Figure 5 also verifies our effectiveness. For example, in the second example, while TediGAN is unable to synthesize hair consistent to the mask and Composable Diffusion fails to generate beard according to the text, our framework is able to generate results highly consistent to both conditions while maintaining identity." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.671, + 0.471, + 0.777 + ], + "angle": 0, + "content": "Quantitative Comparison. As shown in Table 1, our Collaborate Diffusion outperforms TediGAN and Composable Diffusion in all three objective metrics. In addition, as depicted in Figure 7, our framework achieves the best result in \\(69.40\\%\\) and \\(84.37\\%\\) of the time for face generation and editing, respectively. These results verify our effectiveness in both image quality and image-condition consistency." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.81, + 0.229, + 0.826 + ], + "angle": 0, + "content": "4.5. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.47, + 0.9 + ], + "angle": 0, + "content": "In this section, we visualize the influence functions, and show that it is necessary for the influence function to be both spatially and temporally varying to facilitate effective collaboration." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.543, + 0.838, + 0.558 + ], + "angle": 0, + "content": "4.5.1 Spatial Variations of Influence Functions" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.567, + 0.894, + 0.702 + ], + "angle": 0, + "content": "As shown in Figure 10, the influence functions behave differently at different spatial regions. For instance, the influence for the mask-driven model mainly lies on the contours of facial regions, such as the outline of hair, face, and eyes, as these regions are crucial in defining facial layout. In contrast, the influence for the text-driven model is stronger at skin regions including cheeks and chin. This is because the attributes related skin texture, such as age and beard length, are better described by text." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.704, + 0.894, + 0.794 + ], + "angle": 0, + "content": "To verify its necessity, we remove the spatial variation of influence functions. From Table 2 we observe that removing the spatial variance results in deterioration in both output quality and condition consistency. This corroborates our hypothesis that it is important to assign different weights to different modalities at different spatial regions." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.816, + 0.857, + 0.831 + ], + "angle": 0, + "content": "4.5.2 Temporal Variations of Influence Functions" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.894, + 0.901 + ], + "angle": 0, + "content": "In addition to the spatial variation, it is also observed in Figure 10 that the influence from the mask-driven model is stronger at earlier diffusion stages (i.e., larger \\(t\\)), since early stages focus on initializing the facial layout using the mask-" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.956 + ], + "angle": 0, + "content": "6086" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.1, + 0.089, + 0.873, + 0.371 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.103, + 0.383, + 0.867, + 0.399 + ], + "angle": 0, + "content": "Figure 9. Face Editing Results. Given the input real image and target conditions, we display the edited image using our method." + }, + { + "type": "image", + "bbox": [ + 0.084, + 0.423, + 0.466, + 0.564 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.574, + 0.47, + 0.618 + ], + "angle": 0, + "content": "Figure 10. Visualization of Influence Functions. The influence functions vary spatially at different face regions, and temporally at different diffusion timesteps." + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.636, + 0.47, + 0.679 + ], + "angle": 0, + "content": "Table 2. Ablation Study. Temporal or spatial suppression in influence variation introduces performance drops, which shows the necessity of influence functions' spatial-temporal adaptivity." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.683, + 0.462, + 0.745 + ], + "angle": 0, + "content": "
MethodFID ↓Text (%) ↑Mask (%) ↑
Ours w/o Spatial117.8124.3680.08
Ours w/o Temporal117.3424.4877.07
Ours111.3624.5180.25
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.775, + 0.47, + 0.837 + ], + "angle": 0, + "content": "driven model's predictions. At later stages, the influence from the text-driven model increases as the textural details (e.g., skin wrinkles and beard length) are instantiated using information from the text." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.471, + 0.903 + ], + "angle": 0, + "content": "We remove the temporal variation in the influence functions to demonstrate its importance. It is shown in Table 2 that both the quality and condition consistency drop without the temporal variation." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.423, + 0.62, + 0.438 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.455, + 0.895, + 0.653 + ], + "angle": 0, + "content": "With generative AI gaining increasing attention, multimodal conditioning becomes an indispensable direction to unleash creativity and enable comprehensive controls from human creators. In this work, we take the first step forward and propose Collaborative Diffusion, where pre-trained uni-modal diffusion models collaboratively achieve multi-modal face generation and editing without being re-trained. With our dynamic diffuser, this framework could be used to extend arbitrary uni-modal approach to the multi-modal paradigm through predicting the relative influence of different modalities. We believe our idea of synergizing uni-modal models for multi-modal tasks would be a good inspiration for future works in various domains such as motion and 3D generation." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.659, + 0.895, + 0.737 + ], + "angle": 0, + "content": "Limitations and Future Work. Since our framework focuses on exploiting pre-trained diffusion models, our performance is dependent on the capability of each model. An orthogonal direction is to train each collaborator on large-scale datasets for performance gain." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.742, + 0.895, + 0.804 + ], + "angle": 0, + "content": "Potential Negative Societal Impacts. The facial manipulation capabilities of Collaborative Diffusion could be applied maliciously on real human faces. We advise users to apply it only for proper recreational purposes." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.81, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Acknowledgement. We would like to thank Liangyu Chen and Bo Li for proofreading. This study is supported by NTU NAP, MOE AcRF Tier 1 (2021-T1-001-088), and under the RIE2020 Industry Alignment Fund - Industry Collaboration Projects (IAF-ICP) Funding Initiative, as well as cash and in-kind contribution from the industry partner(s)." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6087" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.156 + ], + "angle": 0, + "content": "[1] Tomer Amit, Eliya Nachmani, Tal Shaharbany, and Lior Wolf. SegDiff: Image segmentation with diffusion probabilistic models. arXiv preprint arXiv:2112.00390, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.157, + 0.471, + 0.199 + ], + "angle": 0, + "content": "[2] Jacob Austin, Daniel D Johnson, Jonathan Ho, Daniel Tarlow, and Rianne van den Berg. Structured denoising diffusion models in discrete state-spaces. In NeurIPS, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.2, + 0.471, + 0.24 + ], + "angle": 0, + "content": "[3] Omri Avrahami, Dani Lischinski, and Ohad Fried. Blended diffusion for text-driven editing of natural images. In CVPR, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.242, + 0.471, + 0.282 + ], + "angle": 0, + "content": "[4] Dmitry Baranchuk, Ivan Rubachev, Andrey Voynov, Valentin Khrulkov, and Artem Babenko. Label-efficient semantic segmentation with diffusion models. In ICLR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.284, + 0.471, + 0.352 + ], + "angle": 0, + "content": "[5] Sam Bond-Taylor, Peter Hessey, Hiroshi Sasaki, Toby P Breckon, and Chris G Willcocks. Unleashing transformers: parallel token prediction with discrete absorbing diffusion for fast high-resolution image generation from vector-quantized codes. In ECCV, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.354, + 0.471, + 0.394 + ], + "angle": 0, + "content": "[6] Huiwen Chang, Han Zhang, Lu Jiang, Ce Liu, and William T Freeman. MaskGIT: Masked generative image transformer. In CVPR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.397, + 0.471, + 0.45 + ], + "angle": 0, + "content": "[7] Shin-I Cheng, Yu-Jie Chen, Wei-Chen Chiu, Hung-Yu Tseng, and Hsin-Ying Lee. Adaptively-realistic image generation from stroke and sketch with diffusion model. In WACV, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.452, + 0.471, + 0.507 + ], + "angle": 0, + "content": "[8] Guillaume Couairon, Jakob Verbeek, Holger Schwenk, and Matthieu Cord. Diffedit: Diffusion-based semantic image editing with mask guidance. arXiv preprint arXiv:2210.11427, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.509, + 0.471, + 0.536 + ], + "angle": 0, + "content": "[9] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat GANs on image synthesis. In NeurIPS, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.537, + 0.47, + 0.591 + ], + "angle": 0, + "content": "[10] Ming Ding, Zhuoyi Yang, Wenyi Hong, Wendi Zheng, Chang Zhou, Da Yin, Junyang Lin, Xu Zou, Zhou Shao, Hongxia Yang, et al. CogView: Mastering text-to-image generation via transformers. In NeurIPS, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.593, + 0.47, + 0.647 + ], + "angle": 0, + "content": "[11] Patrick Esser, Robin Rombach, Andreas Blattmann, and Bjorn Ommer. ImageBART: Bidirectional context with multinomial diffusion for autoregressive image synthesis. In NeurIPS, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.649, + 0.47, + 0.689 + ], + "angle": 0, + "content": "[12] Patrick Esser, Robin Rombach, and Bjorn Ommer. Taming transformers for high-resolution image synthesis. In CVPR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.691, + 0.47, + 0.745 + ], + "angle": 0, + "content": "[13] Oran Gafni, Adam Polyak, Oron Ashual, Shelly Sheynin, Devi Parikh, and Yaniv Taigman. Make-a-scene: Scene-based text-to-image generation with human priors. arXiv preprint arXiv:2203.13131, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.747, + 0.47, + 0.801 + ], + "angle": 0, + "content": "[14] Ian J Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron C Courville, and Yoshua Bengio. Generative adversarial nets. In NeurIPS, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.803, + 0.47, + 0.844 + ], + "angle": 0, + "content": "[15] Alexandros Graikos, Nikolay Malkin, Nebojsa Jojic, and Dimitris Samaras. Diffusion models as plug-and-play priors. In NeurIPS, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.846, + 0.47, + 0.9 + ], + "angle": 0, + "content": "[16] Shuyang Gu, Dong Chen, Jianmin Bao, Fang Wen, Bo Zhang, Dongdong Chen, Lu Yuan, and Baining Guo. Vector quantized diffusion model for text-to-image synthesis. In CVPR, 2022. 1, 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.471, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.134 + ], + "angle": 0, + "content": "[17] William Harvey, Saeid Naderiparizi, Vaden Masrani, Christian Weilbach, and Frank Wood. Flexible diffusion modeling of long videos. arXiv preprint arXiv:2205.11495, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.136, + 0.894, + 0.188 + ], + "angle": 0, + "content": "[18] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. GANs trained by a two time-scale update rule converge to a local nash equilibrium. In NeurIPS, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.19, + 0.894, + 0.259 + ], + "angle": 0, + "content": "[19] Jonathan Ho, William Chan, Chitwan Sahara, Jay Whang, Ruiqi Gao, Alexey Gritsanko, Diederik P Kingma, Ben Poole, Mohammad Norouzi, David J Fleet, et al. Imagen video: High definition video generation with diffusion models. arXiv preprint arXiv:2210.02303, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.261, + 0.894, + 0.287 + ], + "angle": 0, + "content": "[20] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In NeurIPS, 2020. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.288, + 0.894, + 0.329 + ], + "angle": 0, + "content": "[21] Jonathan Ho, Chitwan Sahara, William Chan, David J Fleet, Mohammad Norouzi, and Tim Salimans. Cascaded diffusion models for high fidelity image generation. JMLR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.33, + 0.894, + 0.37 + ], + "angle": 0, + "content": "[22] Aapo Hyvarinen and Peter Dayan. Estimation of non-normalized statistical models by score matching. JMLR, 2005. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.372, + 0.894, + 0.412 + ], + "angle": 0, + "content": "[23] Yuming Jiang, Ziqi Huang, Xingang Pan, Chen Change Loy, and Ziwei Liu. Talk-to-Edit: Fine-grained facial editing via dialog. In ICCV, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.414, + 0.894, + 0.455 + ], + "angle": 0, + "content": "[24] Yuming Jiang, Shuai Yang, Haonan Qju, Wayne Wu, Chen Change Loy, and Ziwei Liu. Text2human: Text-driven controllable human image generation. ACM TOG, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.456, + 0.894, + 0.495 + ], + "angle": 0, + "content": "[25] Tero Karras, Timo Aila, Samuli Laine, and Jaakko Lehtinen. Progressive growing of GANs for improved quality, stability, and variation. In ICLR, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.497, + 0.894, + 0.538 + ], + "angle": 0, + "content": "[26] Tero Karras, Miika Aittala, Samuli Laine, Erik Härkönen, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Alias-free generative adversarial networks. In NeurIPS, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.539, + 0.894, + 0.579 + ], + "angle": 0, + "content": "[27] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In CVPR, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.581, + 0.894, + 0.622 + ], + "angle": 0, + "content": "[28] Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of StyleGAN. In CVPR, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.623, + 0.894, + 0.678 + ], + "angle": 0, + "content": "[29] Bahjat Kawar, Shiran Zada, Oran Lang, Omer Tov, Huiwen Chang, Tali Dekel, Inbar Mosseri, and Michal Irani. Imagic: Text-based real image editing with diffusion models. arXiv preprint arXiv:2210.09276, 2022. 1, 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.679, + 0.894, + 0.718 + ], + "angle": 0, + "content": "[30] Gwanghyun Kim, Taesung Kwon, and Jong Chul Ye. DiffusionCLIP: Text-guided diffusion models for robust image manipulation. In CVPR, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.72, + 0.894, + 0.761 + ], + "angle": 0, + "content": "[31] Jing Yu Koh, Jason Baldridge, Honglak Lee, and Yinfei Yang. Text-to-image generation grounded by fine-grained user attention. In WACV, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.762, + 0.894, + 0.802 + ], + "angle": 0, + "content": "[32] Cheng-Han Lee, Ziwei Liu, Lingyun Wu, and Ping Luo. MaskGAN: Towards diverse and interactive facial image manipulation. In CVPR, 2020. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.804, + 0.894, + 0.844 + ], + "angle": 0, + "content": "[33] Bowen Li, Xiaojuan Qi, Thomas Lukasiewicz, and Philip HS Torr. ManiGAN: Text-guided image manipulation. In CVPR, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.845, + 0.894, + 0.899 + ], + "angle": 0, + "content": "[34] Wenbo Li, Pengchuan Zhang, Lei Zhang, Qiuyuan Huang, Xiaodong He, Siwei Lyu, and Jianfeng Gao. Object-driven text-to-image synthesis via adversarial training. In CVPR, 2019. 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.956 + ], + "angle": 0, + "content": "6088" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.134 + ], + "angle": 0, + "content": "[35] Zhiheng Li, Martin Renqiang Min, Kai Li, and Chenliang Xu. StyleT2I: Toward compositional and high-fidelity text-to-image synthesis. In CVPR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.136, + 0.472, + 0.19 + ], + "angle": 0, + "content": "[36] Nan Liu, Shuang Li, Yilun Du, Antonio Torralba, and Joshua B Tenenbaum. Compositional visual generation with composable diffusion models. arXiv preprint arXiv:2206.01714, 2022. 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.192, + 0.472, + 0.247 + ], + "angle": 0, + "content": "[37] Xihui Liu, Dong Huk Park, Samaneh Azadi, Gong Zhang, Arman Chopikyan, Yuxiao Hu, Humphrey Shi, Anna Rohrbach, and Trevor Darrell. More control for free! image synthesis with semantic diffusion guidance. In WACV, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.249, + 0.472, + 0.303 + ], + "angle": 0, + "content": "[38] Chenlin Meng, Yutong He, Yang Song, Jiaming Song, Jiajun Wu, Jun-Yan Zhu, and Stefano Ermon. SDEdit: Guided image synthesis and editing with stochastic differential equations. In ICLR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.305, + 0.472, + 0.374 + ], + "angle": 0, + "content": "[39] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. GLIDE: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.376, + 0.472, + 0.417 + ], + "angle": 0, + "content": "[40] Taesung Park, Ming-Yu Liu, Ting-Chun Wang, and Jun-Yan Zhu. Semantic image synthesis with spatially-adaptive normalization. In CVPR, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.419, + 0.472, + 0.46 + ], + "angle": 0, + "content": "[41] Or Patashnik, Zongze Wu, Eli Shechtman, Daniel Cohen-Or, and Dani Lischinski. StyleCLIP: Text-driven manipulation of stylegan imagery. In ICCV, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.462, + 0.472, + 0.53 + ], + "angle": 0, + "content": "[42] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.532, + 0.472, + 0.586 + ], + "angle": 0, + "content": "[43] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with CLIP latents. arXiv preprint arXiv:2204.06125, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.589, + 0.472, + 0.63 + ], + "angle": 0, + "content": "[44] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In ICML, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.632, + 0.472, + 0.673 + ], + "angle": 0, + "content": "[45] Scott Reed, Zeynep Akata, Xinchen Yan, Lajanugen Logeswaran, Bernt Schiele, and Honglak Lee. Generative adversarial text to image synthesis. In ICML, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.675, + 0.472, + 0.715 + ], + "angle": 0, + "content": "[46] Scott E Reed, Zeynep Akata, Santosh Mohan, Samuel Tenka, Bernt Schiele, and Honglak Lee. Learning what and where to draw. In NeurIPS, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.717, + 0.472, + 0.771 + ], + "angle": 0, + "content": "[47] Elad Richardson, Yuval Alaluf, Or Patashnik, Yotam Nitzan, Yaniv Azar, Stav Shapiro, and Daniel Cohen-Or. Encoding in style: a StyleGAN encoder for image-to-image translation. In CVPR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.774, + 0.472, + 0.828 + ], + "angle": 0, + "content": "[48] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 1, 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.83, + 0.472, + 0.871 + ], + "angle": 0, + "content": "[49] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-Net: Convolutional networks for biomedical image segmentation. In MICCAI, 2015. 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.472, + 0.9 + ], + "angle": 0, + "content": "[50] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.894, + 0.12 + ], + "angle": 0, + "content": "tuning text-to-image diffusion models for subject-driven generation. arXiv preprint arXiv:2208.12242, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.122, + 0.894, + 0.203 + ], + "angle": 0, + "content": "[51] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily Denton, Seyed Kamyar Seyed Ghasemipour, Burcu Karagol Ayan, S Sara Mahdavi, Rapha Gontijo Lopes, et al. Photorealistic text-to-image diffusion models with deep language understanding. arXiv preprint arXiv:2205.11487, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.206, + 0.894, + 0.247 + ], + "angle": 0, + "content": "[52] Chitwan Sahara, Jonathan Ho, William Chan, Tim Salimans, David J Fleet, and Mohammad Norouzi. Image superresolution via iterative refinement. IEEE TPAMI, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.249, + 0.894, + 0.289 + ], + "angle": 0, + "content": "[53] Yujun Shen, Jinjin Gu, Xiaou Tang, and Bolei Zhou. Interpreting the latent space of GANs for semantic face editing. In CVPR, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.291, + 0.894, + 0.332 + ], + "angle": 0, + "content": "[54] Yujun Shen, Ceyuan Yang, Xiaou Tang, and Bolei Zhou. InterFaceGAN: Interpreting the disentangled face representation learned by GANs. IEEE TPAMI, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.334, + 0.894, + 0.401 + ], + "angle": 0, + "content": "[55] Uriel Singer, Adam Polyak, Thomas Hayes, Xi Yin, Jie An, Songyang Zhang, Qiyuan Hu, Harry Yang, Oron Ashual, Oran Gafni, et al. Make-a-video: Text-to-video generation without text-video data. arXiv preprint arXiv:2209.14792, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.404, + 0.894, + 0.446 + ], + "angle": 0, + "content": "[56] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In ICML, 2015. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.447, + 0.894, + 0.474 + ], + "angle": 0, + "content": "[57] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In ICLR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.476, + 0.894, + 0.53 + ], + "angle": 0, + "content": "[58] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. In ICLR, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.532, + 0.894, + 0.573 + ], + "angle": 0, + "content": "[59] Yang Song, Jingwen Zhu, Dawei Li, Xiaolong Wang, and Hairong Qi. Talking face generation by conditional recurrent adversarial network. *IJCAI*, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.575, + 0.894, + 0.616 + ], + "angle": 0, + "content": "[60] Jianxin Sun, Qiyao Deng, Qi Li, Muyi Sun, Min Ren, and Zhenan Sun. AnyFace: Free-style text-to-face synthesis and manipulation. In CVPR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.618, + 0.894, + 0.646 + ], + "angle": 0, + "content": "[61] Aaron Van Den Oord, Oriol Vinyals, et al. Neural discrete representation learning. In NeurIPS, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.647, + 0.894, + 0.688 + ], + "angle": 0, + "content": "[62] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In NeurIPS, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.689, + 0.894, + 0.759 + ], + "angle": 0, + "content": "[63] Ruben Villegas, Mohammad Babaeizadeh, Pieter-Jan Kindermans, Hernan Moraldo, Han Zhang, Mohammad Taghi Saffar, Santiago Castro, Julius Kunze, and Dumitru Erhan. Phenaki: Variable length video generation from open domain textual description. arXiv preprint arXiv:2210.02399, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.76, + 0.894, + 0.788 + ], + "angle": 0, + "content": "[64] Pascal Vincent. A connection between score matching and denoising autoencoders. Neural Computation, 2011. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.789, + 0.894, + 0.843 + ], + "angle": 0, + "content": "[65] Tengfei Wang, Ting Zhang, Bo Zhang, Hao Ouyang, Dong Chen, Qifeng Chen, and Fang Wen. Pretraining is all you need for image-to-image translation. arXiv preprint arXiv:2205.12952, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.845, + 0.894, + 0.899 + ], + "angle": 0, + "content": "[66] Weilun Wang, Jianmin Bao, Wengang Zhou, Dongdong Chen, Dong Chen, Lu Yuan, and Houqiang Li. Semantic image synthesis via diffusion models. arXiv preprint arXiv:2207.00050, 2022. 1" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6089" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.134 + ], + "angle": 0, + "content": "[67] Weihao Xia, Yujiu Yang, Jing-Hao Xue, and Baoyuan Wu. TediGAN: Text-guided diverse face image generation and manipulation. In CVPR, 2021. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.136, + 0.471, + 0.177 + ], + "angle": 0, + "content": "[68] Weihao Xia, Yujiu Yang, Jing-Hao Xue, and Baoyuan Wu. Towards open-world text-guided face image generation and manipulation. arXiv preprint arXiv:2104.08910, 2021. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.179, + 0.471, + 0.233 + ], + "angle": 0, + "content": "[69] Tao Xu, Pengchuan Zhang, Qiuyuan Huang, Han Zhang, Zhe Gan, Xiaolei Huang, and Xiaodong He. AttnGAN: Fine-grained text to image generation with attentional generative adversarial networks. In CVPR, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.235, + 0.471, + 0.29 + ], + "angle": 0, + "content": "[70] Han Zhang, Tao Xu, Hongsheng Li, Shaoting Zhang, Xiaogang Wang, Xiaolei Huang, and Dimitris N Metaxas. StackGAN: Text to photo-realistic image synthesis with stacked generative adversarial networks. In ICCV, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.292, + 0.471, + 0.346 + ], + "angle": 0, + "content": "[71] Han Zhang, Tao Xu, Hongsheng Li, Shaoting Zhang, Xiaogang Wang, Xiaolei Huang, and Dimitris N Metaxas. StackGAN++: Realistic image synthesis with stacked generative adversarial networks. IEEE TPAMI, 2018. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.471, + 0.346 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.517, + 0.956 + ], + "angle": 0, + "content": "6090" + } + ] +] \ No newline at end of file diff --git a/2023/Collaborative Diffusion for Multi-Modal Face Generation and Editing/3b7ffa39-2244-45b9-9833-f768a1b5b8c6_origin.pdf b/2023/Collaborative Diffusion for Multi-Modal Face Generation and Editing/3b7ffa39-2244-45b9-9833-f768a1b5b8c6_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..62f9a4342068fe8d6e85930d4ec62f6af769a666 --- /dev/null +++ b/2023/Collaborative Diffusion for Multi-Modal Face Generation and Editing/3b7ffa39-2244-45b9-9833-f768a1b5b8c6_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4fea92124215a9cfa6cb83c15525c0a1c54b612da88c60d6bb4c7a2adcb6ddee +size 3961188 diff --git a/2023/Collaborative Diffusion for Multi-Modal Face Generation and Editing/full.md b/2023/Collaborative Diffusion for Multi-Modal Face Generation and Editing/full.md new file mode 100644 index 0000000000000000000000000000000000000000..e8cd5a3c379b9d10c8cb29a41c7c97b9d6c5d9ca --- /dev/null +++ b/2023/Collaborative Diffusion for Multi-Modal Face Generation and Editing/full.md @@ -0,0 +1,461 @@ +# Collaborative Diffusion for Multi-Modal Face Generation and Editing + +Ziqi Huang Kelvin C.K. Chan Yuming Jiang Ziwei Liu S-Lab, Nanyang Technological University + +{ziqi002, chan0899, yuming002, ziwei.liu}@ntu.edu.sg + +![](images/3bf3231c7177e7c670a9268704a60932cf22a3678547ecc6519ec94df8742f14.jpg) +Figure 1. We propose Collaborative Diffusion, where users can use multiple modalities to control face generation and editing. (a) Face Generation. Given multi-modal controls, our framework synthesizes high-quality images consistent with the input conditions. (b) Face Editing. Collaborative Diffusion also supports multi-modal editing of real images with promising identity preservation capability. + +# Abstract + +Diffusion models arise as a powerful generative tool recently. Despite the great progress, existing diffusion models mainly focus on uni-modal control, i.e., the diffusion process is driven by only one modality of condition. To further unleash the users' creativity, it is desirable for the model to be controllable by multiple modalities simultaneously, e.g. generating and editing faces by describing the age (text-driven) while drawing the face shape (mask-driven). + +In this work, we present Collaborative Diffusion, where pre-trained uni-modal diffusion models collaborate to achieve multi-modal face generation and editing without re-training. Our key insight is that diffusion models driven by different modalities are inherently complementary regarding the latent denoising steps, where bilateral connections can be established upon. Specifically, we propose dynamic diffuser, a meta-network that adaptively hallucinates multi-modal denoising steps by predicting the spatial-temporal + +influence functions for each pre-trained uni-modal model. Collaborative Diffusion not only collaborates generation capabilities from uni-modal diffusion models, but also integrates multiple uni-modal manipulations to perform multimodal editing. Extensive qualitative and quantitative experiments demonstrate the superiority of our framework in both image quality and condition consistency. + +# 1. Introduction + +Recent years have witnessed substantial progress in image synthesis and editing with the surge of diffusion models [9, 20, 56, 58]. In addition to the remarkable synthesis quality, one appealing property of diffusion models is the flexibility of conditioning on various modalities, such as texts [3, 16, 29, 30, 39, 48, 51], segmentation masks [48, 65, 66], and sketches [7, 65]. However, existing explorations are largely confined to the use of a single modality at a time. The exploitation of multiple conditions remains under-explored. As a generative tool, its controllability is still limited. + +To unleash users' creativity, it is desirable that the model + +is simultaneously controllable by multiple modalities. While it is trivial to extend the current supervised framework with multiple modalities, training a large-scale model from scratch is computationally expensive, especially when extensive hyper-parameter tuning and delicate architecture designs are needed. More importantly, each trained model could only accept a fixed combination of modalities, and hence re-training is necessary when a subset of modalities are absent, or when additional modalities become available. The above demonstrates the necessity of a unified framework that effectively exploits pre-trained models and integrate them for multi-modal synthesis and editing. + +In this paper, we propose Collaborative Diffusion, a framework that synergizes pre-trained uni-modal diffusion models for multi-modal face generation and editing without the need of re-training. Motivated by the fact that different modalities are complementary to each other (e.g., text for age and mask for hair shape), we explore the possibility of establishing lateral connections between models driven by different modalities. We propose dynamic diffuser to adaptively predict the spatial-temporal influence function for each pre-trained model. The dynamic diffuser dynamically determines spatial-varying and temporal-varying influences of each model, suppressing contributions from irrelevant modalities while enhancing contributions from admissible modalities. In addition to multi-modal synthesis, the simplicity and flexibility of our framework enable extension to multi-modal face editing with minimal modifications. In particular, the dynamic diffuser is first trained for collaborative synthesis. It is then fixed and combined with existing face editing approaches [29, 50] for multi-modal editing. It is worth-mentioning that users can select the best editing approaches based on their needs without the need of altering the dynamic diffusers. + +We demonstrate both qualitatively and quantitatively that our method achieves superior image quality and condition consistency in both synthesis and editing tasks. Our contributions can be summarized as follows: + +- We introduce Collaborative Diffusion, which exploits pre-trained uni-modal diffusion models for multi-modal controls without re-training. Our approach is the first attempt towards flexible integration of uni-modal diffusion models into a single collaborative framework. +Tailored for the iterative property of diffusion models, we propose dynamic diffuser, which predicts the spatial-varying and temporal-varying influence functions to selectively enhance or suppress the contributions of the given modalities at each iterative step. +- We demonstrate the flexibility of our framework by extending it to face editing driven by multiple modalities. Both quantitative and qualitative results demonstrate the superiority of Collaborative Diffusion in multi-modal face generation and editing. + +# 2. Related Work + +Diffusion Models. Diffusion models [20, 56, 58] have recently become a mainstream approach for image synthesis [9, 11, 38] apart from Generative Adversarial Networks (GANs) [14], and success has also been found in various domains including video generation [17, 19, 55, 63], image restoration [21, 52], semantic segmentation [1, 4, 15], and natural language processing [2]. In the diffusion-based framework, models are trained with score-matching objectives [22, 64] at various noise levels, and sampling is done via iterative denoising. Existing works focus on improving the performance and efficiency of diffusion models through enhanced architecture designs [16, 48] and sampling schemes [57]. In contrast, this work focuses on exploiting existing models, and providing a succinct framework for multi-modal synthesis and editing without large-scale re-training of models. + +Face Generation. Existing face generation approaches can be divided into three main directions. Following the GAN paradigm, the StyleGAN series [26-28] boost the quality of facial synthesis, and provide an interpretable latent space for steerable style controls and manipulations. The vector-quantized approaches [12, 61] learn a discrete codebook by mapping the input images into a low-dimensional discrete feature space. The learned codebook is then sampled, either sequentially [12, 61] or parallelly [5, 6, 16], for synthesis. In contrast to the previous two approaches, diffusion models are trained with a stationary objective, without the need of optimizing complex losses (e.g., adversarial loss) or balancing multiple objectives (e.g., codebook loss versus reconstruction loss). With training simplicity as a merit, diffusion-based approaches have become increasingly popular in recent years. Our framework falls in the diffusion-based paradigm. In particular, we leverage pre-trained diffusion models for multi-modal generation and editing. + +Conditional Face Generation and Editing. Conditional generation [10, 11, 13, 31, 34, 37, 40, 43-46, 51, 65, 67-71] and editing [8, 32, 41, 53, 54, 67, 68] is an active line of research focusing on conditioning generative models on different modalities, such as texts [24, 41, 67, 68], segmentation masks [32, 33, 40, 47], and audios [59]. For example, StyleCLIP [41], DiffusionCLIP [30], and many others [35, 60] have demonstrated remarkable performance in text-guided face generation and editing. However, most existing models do not support simultaneous conditioning on multiple modalities (e.g., text and mask at the same time), and supporting additional modalities often requires time-consuming model re-training and extensive hyper-parameter tuning, which are not preferable in general. In this work, we propose Collaborative Diffusion to exploit pre-trained uni-modal diffusion models [48] (e.g., text-driven and mask-driven models) to achieve multi-modal conditioning without model retraining. + +![](images/dbf0ef99a8c6ed2c77ed88a1f99b3f247180cfd54c544982d71242804f33497c.jpg) +Figure 2. Overview of Collaborative Diffusion. We use pre-trained uni-modal diffusion models to perform multi-modal guided face generation and editing. At each step of the reverse process (i.e., from timestep $t$ to $t - 1$ ), the dynamic diffuser predicts the spatial-varying and temporal-varying influence function to selectively enhance or suppress the contributions of the given modality. + +# 3. Collaborative Diffusion + +We propose Collaborative Diffusion, which exploits multiple pre-trained uni-modal diffusion models (Section 3.1) for multi-modal generation and editing. The key of our framework is the dynamic diffuser, which adaptively predicts the influence functions to enhance or suppress the contributions of the pre-trained models based on the spatial-temporal influences of the modalities. Our framework is compatible with most existing approaches for both multi-modal guided synthesis (Section 3.2) and multi-modal editing (Section 3.3). + +# 3.1. Uni-Modal Conditional Diffusion Models + +Diffusion models are a class of generative models that model the data distribution in the form of $p_{\theta}(\mathbf{x}_0) \coloneqq \int p_{\theta}(\mathbf{x}_{0:T}) d\mathbf{x}_{1:T}$ . The diffusion process (a.k.a. forward process) gradually adds Gaussian noise to the data and eventually corrupts the data $\mathbf{x}_0$ into an approximately pure Gaussian noise $\mathbf{x}_T$ using a variance schedule $\beta_1, \ldots, \beta_T$ : + +$$ +q \left(\mathbf {x} _ {1: T} \mid \mathbf {x} _ {0}\right) := \prod_ {t = 1} ^ {T} q \left(\mathbf {x} _ {t} \mid \mathbf {x} _ {t - 1}\right), \tag {1} +$$ + +$$ +q (\mathbf {x} _ {t} | \mathbf {x} _ {t - 1}) := \mathcal {N} (\mathbf {x} _ {t}; \sqrt {1 - \beta_ {t}} \mathbf {x} _ {t - 1}, \beta_ {t} \mathbf {I}). +$$ + +Reversing the forward process allows sampling new data $\mathbf{x}_0$ by starting from $p(\mathbf{x}_T) = \mathcal{N}(\mathbf{x}_T; \mathbf{0}, \mathbf{I})$ . The reverse process is defined as a Markov chain where each step is a learned + +Gaussian transition $(\pmb{\mu}_{\theta},\pmb{\Sigma}_{\theta})$ + +$$ +p _ {\theta} \left(\mathbf {x} _ {0: T}\right) := p \left(\mathbf {x} _ {T}\right) \prod_ {t = 1} ^ {T} p _ {\theta} \left(\mathbf {x} _ {t - 1} \mid \mathbf {x} _ {t}\right), \tag {2} +$$ + +$$ +p _ {\theta} \left(\mathbf {x} _ {t - 1} \mid \mathbf {x} _ {t}\right) := \mathcal {N} \left(\mathbf {x} _ {t - 1}; \boldsymbol {\mu} _ {\theta} \left(\mathbf {x} _ {t}, t\right), \boldsymbol {\Sigma} _ {\theta} \left(\mathbf {x} _ {t}, t\right)\right). +$$ + +Training diffusion models relies on minimizing the variational bound on $p(x)$ 's negative log-likelihood. The commonly used optimization objective $L_{\mathrm{DM}}$ [20] reparameterizes the learnable Gaussian transition as $\epsilon_{\theta}(\cdot)$ , and temporally reweights the variational bound to trade for better sample quality: + +$$ +L _ {\mathrm {D M}} (\theta) := \mathbb {E} _ {t, \mathbf {x} _ {0}, \boldsymbol {\epsilon} \sim \mathcal {N} (\mathbf {0}, \mathbf {I})} \left[ \| \boldsymbol {\epsilon} - \boldsymbol {\epsilon} _ {\theta} (\mathbf {x} _ {t}, t) \| ^ {2} \right], \quad (3) +$$ + +where $\mathbf{x}_t$ can be directly approximated by $\mathbf{x}_t = \sqrt{\bar{\alpha}_t}\mathbf{x}_0 + \sqrt{1 - \bar{\alpha}_t}\pmb {\epsilon}$ , with $\bar{\alpha}_{t}:= \prod_{s = 1}^{t}\alpha_{s}$ and $\alpha_{t}:= 1 - \beta_{t}$ . + +To sample data $\mathbf{x}_0$ from a trained diffusion model $\epsilon_{\theta}(\cdot)$ , we iteratively denoise $\mathbf{x}_t$ from $t = T$ to $t = 1$ with noise $\mathbf{z}$ : + +$$ +\mathbf {x} _ {t - 1} = \frac {1}{\sqrt {\alpha_ {t}}} \left(\mathbf {x} _ {t} - \frac {1 - \alpha_ {t}}{\sqrt {1 - \bar {\alpha} _ {t}}} \boldsymbol {\epsilon} _ {\theta} (\mathbf {x} _ {t}, t)\right) + \sigma_ {t} \mathbf {z} \quad (4) +$$ + +The unconditional diffusion models can be extended to model conditional distributions $p_{\theta}(\mathbf{x}_0|c)$ , where $\mathbf{x}_0$ is the image corresponding to the condition $c$ such as class labels, segmentation masks, and text descriptions. The conditional diffusion model receives an additional input $\tau(c)$ and is trained by minimizing $\| \boldsymbol{\epsilon} - \boldsymbol{\epsilon}_{\theta}(\mathbf{x}_t, t, \tau(c)) \|^2$ , where $\tau(\cdot)$ is an encoder that projects the condition $c$ to an embedding $\tau(c)$ . For brevity, we will use $c$ to represent $\tau(c)$ in our subsequent discussions. + +Algorithm 1 Dynamic Diffuser Training +1: repeat +2: $\mathbf{x}_0, c_1, c_2, \ldots, c_M \sim q(\mathbf{x}_0, c_1, c_2, \ldots, c_M)$ +3: $t \sim \mathrm{Uniform}(\{1, \ldots, T\})$ +4: $\epsilon \sim \mathcal{N}(\mathbf{0}, \mathbf{I})$ +5: for $m = 1, \ldots, M$ do +6: $\epsilon_{pred,m,t} = \epsilon_{\theta_m}(\sqrt{\bar{\alpha}_t}\mathbf{x}_0 + \sqrt{1 - \bar{\alpha}_t}\boldsymbol{\epsilon}, t, c_m)$ +7: $\mathbf{I}_{m,t} = \mathbf{D}_{\phi_m}(\sqrt{\bar{\alpha}_t}\mathbf{x}_0 + \sqrt{1 - \bar{\alpha}_t}\boldsymbol{\epsilon}, t, c_m)$ +8: end for +9: $\hat{\mathbf{I}}_{m,t,p} = \frac{\exp(\mathbf{I}_{m,t,p})}{\sum_{j=1}^{M}\exp(\mathbf{I}_{j,t,p})}$ , softmax at each pixel $p$ +10: $\epsilon_{pred,t} = \sum_{m=1}^{M}\hat{\mathbf{I}}_{m,t} \odot \epsilon_{pred,m,t}$ +11: Take gradient descent step on $\nabla_{\phi} \| \boldsymbol{\epsilon} - \boldsymbol{\epsilon}_{pred,t} \|^2$ where $\phi = \{\phi_m | m = 1, \ldots, M\}$ +12: until converged + +![](images/f73bd6938ea41ca9e485b986ac7e1896cd841e0c6dee6210685de9919a86cf41.jpg) +Figure 3. Dynamic Diffuser. It predicts the influence function that represents the desired level of contribution from each pre-trained diffusion model. The influence function varies at different timestep $t$ and different spatial locations, and is dependent on the state of diffusion $\mathbf{x}_t$ and the condition input $c_m$ . + +# 3.2. Multi-Modal Collaborative Synthesis + +In our Collaborative Diffusion, multiple uni-modal diffusion models collaborate at each step of the denoising process for multi-modal guided synthesis. The core of our framework is the dynamic diffuser, which determines the extent of contribution from each collaborator by predicting the spatial-temporal influence functions. + +Problem Formulation. Given $M$ pre-trained uni-modal conditional diffusion models $\{\epsilon_{\theta_m}\}$ (each models the distribution $p(\mathbf{x}_0|c_m)$ ), where the modality index $m = 1, \ldots, M$ , our objective is to sample from $p(\mathbf{x}_0|\mathbf{c})$ , where $\mathbf{c} = \{c_1, c_2, \dots, c_M\}$ , without altering pre-trained models. + +Dynamic Diffuser. In diffusion models, each step of the reverse process requires predicting the noise $\epsilon$ . With multiple diffusion models collaborating, we need to carefully determine when, where, and how much each diffusion model contributes to the prediction $\epsilon$ . At each diffusion timestep $t = T, \dots, 1$ , the influence $\mathbf{I}_{m,t}$ from each pre-trained diffusion model $\epsilon_{\theta_m}$ is adaptively determined by a dynamic + +Algorithm 2 Collaborative Sampling +1: $\mathbf{x}_T\sim \mathcal{N}(\mathbf{0},\mathbf{I})$ +2: for $t = T,\dots ,1$ do +3: $\mathbf{z}\sim \mathcal{N}(\mathbf{0},\mathbf{I})$ if $t > 1$ , else $\mathbf{z} = \mathbf{0}$ +4: for $m = 1,\ldots ,M$ do +5: $\epsilon_{pred,m,t} = \epsilon_{\theta_m}(\mathbf{x}_t,t,c_m)$ +6: $\mathbf{I}_{m,t} = \mathbf{D}_{\phi_m}(\mathbf{x}_t,t,c_m)$ +7: end for +8: $\hat{\mathbf{I}}_{m,t,p} = \frac{\exp(\mathbf{I}_{m,t,p})}{\sum_{j = 1}^{M}\exp(\mathbf{I}_{j,t,p})}$ , softmax at each pixel $p$ +9: $\epsilon_{pred,t} = \sum_{m = 1}^{M}\hat{\mathbf{I}}_{m,t}\odot \epsilon_{pred,m,t}$ +10: $\mathbf{x}_{t - 1} = \frac{1}{\sqrt{\alpha_t}}\left(\mathbf{x}_t - \frac{1 - \alpha_t}{\sqrt{1 - \bar{\alpha}_t}}\pmb{\epsilon}_{pred,t}\right) + \sigma_t\mathbf{z}$ +11: end for +12: return ${\bf x}_0$ + +diffuser $\mathbf{D}_{\phi_m}$ + +$$ +\mathbf {I} _ {m, t} = \mathbf {D} _ {\phi_ {m}} (\mathbf {x} _ {t}, t, c _ {m}), \tag {5} +$$ + +where $m = 1,\dots,M$ is the index of the modalities, $\mathbf{I}_{m,t}\in$ $\mathbb{R}^{h\times w}$ $\mathbf{x}_t$ is the noisy image at time $t$ , $c_{m}$ is the condition of the $m^{th}$ modality, and $\mathbf{D}_{\phi_m}$ is the dynamic diffuser implemented by a UNet [49]. To regularize the overall influence strength, we perform softmax across all modalities' $\mathbf{I}_{m,t}$ at each pixel $p$ to obtain the final influence function $\hat{\mathbf{I}}_{m,t}$ : + +$$ +\hat {\mathbf {I}} _ {m, t, p} = \frac {\exp \left(\mathbf {I} _ {m , t , p}\right)}{\sum_ {j = 1} ^ {M} \exp \left(\mathbf {I} _ {j , t , p}\right)}. \tag {6} +$$ + +Multi-Modal Collaboration. We use the $M$ learned influence functions $\hat{\mathbf{I}}_{m,t}$ to control the contribution from each pre-trained diffusion model at each denoising step: + +$$ +\boldsymbol {\epsilon} _ {p r e d, t} = \sum_ {m = 1} ^ {M} \hat {\mathbf {I}} _ {m, t} \odot \boldsymbol {\epsilon} _ {\theta_ {m}} \left(x _ {t}, t, c _ {m}\right) \tag {7} +$$ + +where $\epsilon_{\theta_m}$ is the $m^{th}$ collaborator, and $\odot$ denotes pixel-wise multiplication. The complete training procedure of dynamic diffusers and the image sampling strategy in our Collaborative Diffusion framework are detailed in Algorithm 1 and Algorithm 2, respectively. + +Relation with Composable Diffusion. Composable Diffusion [36] combines two instances of the same text-to-image diffusion model to achieve compositional visual generation via intermediate results addition. Our framework is related to Composable Diffusion in terms of composing diffusion models for image synthesis, but is substantially different in terms of the task nature and methodology. Composable Diffusion aims to decompose the text condition into elementary segments to factorize the conditional synthesis problem, while we aim to integrate uni-modal collaborators to achieve multi-modal controls. Furthermore, different from Composable Diffusion, which can only compose instances of the + +# Algorithm 3 Collaborative Editing + +# Require: + +input image $\mathbf{x}_{input}$ , target conditions $c_{m,target}$ , diffusion models $\epsilon_{\theta_m}$ , dynamic diffusers $\mathbf{D}_{\phi_m}$ , $(m = 1, \dots, M)$ , interpolation scale $\alpha$ + +1: for $m = 1,\ldots ,M$ do $\triangleright$ Uni-Modal Editing +2: $c_{m} = c_{m, \text{target}}$ +3: $\mathbf{x}_t = \sqrt{\bar{\alpha}_t}\mathbf{x}_{input} + \sqrt{1 - \bar{\alpha}_t}\pmb{\epsilon}$ +4: $c_{m,opt} = \mathrm{argmin}_{c_m}\mathbb{E}_{\epsilon ,t}\left\| \boldsymbol {\epsilon} - \boldsymbol{\epsilon}_{\theta_m}(\mathbf{x}_t,t,c_m)\right\| ^2$ +5: $\theta_{m,opt} = \operatorname*{argmin}_{\theta_m}\mathbb{E}_{\epsilon ,t}\| \pmb {\epsilon} - \pmb{\epsilon}_{\theta_m}(\mathbf{x}_t,t,c_{m,opt})\|^2$ +6: $c_{m,int} = \alpha \cdot c_{m,target} + (1 - \alpha) \cdot c_{m,opt}$ +7: end for + +8: $\mathbf{x}_T\sim \mathcal{N}(\mathbf{0},\mathbf{I})$ ▷Collaborate the Uni-Modal Edits +9: for $t = T, \dots, 1$ do +10: $\mathbf{z} \sim \mathcal{N}(\mathbf{0}, \mathbf{I})$ if $t > 1$ , else $\mathbf{z} = \mathbf{0}$ +11: for $m = 0,\dots,M$ do +12: $\pmb{\epsilon}_{pred,m,t} = \pmb{\epsilon}_{\theta_{m,opt}}(\mathbf{x}_t,t,c_{m,int})$ +13: $\mathbf{I}_{m,t} = \mathbf{D}_{\phi_m}(\mathbf{x}_t,t,c_{m,int})$ +14: end for +15: $\hat{\mathbf{I}}_{m,t,p} = \frac{\exp(\mathbf{I}_{m,t,p})}{\sum_{j = 1}^{M}\exp(\mathbf{I}_{j,t,p})}$ , softmax at each pixel $p$ +16: $\pmb{\epsilon}_{pred,t} = \sum_{m=1}^{M} \hat{\mathbf{I}}_{m,t} \odot \pmb{\epsilon}_{pred,m,t}$ +17: $\mathbf{x}_{t - 1} = \frac{1}{\sqrt{\alpha_t}}\left(\mathbf{x}_t - \frac{1 - \alpha_t}{\sqrt{1 - \bar{\alpha}_t}}\pmb {\epsilon}_{pred,t}\right) + \sigma_t\mathbf{z}$ +18: end for +19: return $\mathbf{x}_0$ + +same text-based synthesis model (i.e., the weights of the diffusion models being composed are the same), our framework possesses the flexibility in integrating models with different weights, architectures, and modalities through the learned dynamic diffusers. + +# 3.3. Multi-Modal Collaborative Editing + +In addition to face synthesis, our framework is also capable of combining multiple facial manipulations, each of which is guided by a different modality, into a collaborative edit. Our collaborative framework in theory can integrate any existing uni-modal diffusion-based editing approach for collaborative editing. + +In this work, we demonstrate such possibility by extending Imagic [29] to the multi-modal paradigm. We first follow Imagic to fine-tune the embeddings and models to better capture the input face identity during editing. The trained dynamic diffusers discussed in Section 3.2 are then used to combine the fine-tuned models. Algorithm 3 displays the complete procedure of collaborative editing. Note that finetuning the pre-trained models is for identity preservation proposed in Imagic, which is independent to our framework. The extension of our framework to editing requires no further training of dynamic diffusers. + +Multi-Modal Conditions +![](images/9afea71ee8fa0a294a9108b1256c1416ff8d90f970bc96f75d04ce55f06121c3.jpg) +He is a young adult. This man has a rough growth of stubble. + +Figure 4. Qualitative Comparison of Face Generation. In the second example, TediGAN and Composable fails to follow mask, while ours generates results highly consistent to both conditions. +![](images/e3fc0353b8f8566b51e014a0a17c5231e48c82e30bb0a033d501cc873fe91127.jpg) +There is no beard. This woman looks extremely young. + +![](images/ff5c22629236d074c61a2c66f9f4c231185a5cb50d3ba7088c0e44b72b0fbb7d.jpg) +Generated Images + +![](images/7a8b8046e2bd97ef9d3789d697463be08dc7d8d48d5b8265b3370f2e3bd3bb59.jpg) + +![](images/d2afee1f2fad5f00c14f368195b69901dedc916a6cfed21b2dfaec8b755c0cef.jpg) +His face is covered with short beard. +This man is in the thirties. + +![](images/f4e49f4c2256072aa21fbebb01b7f6367d2e8fba4ac00d4904ea181897c1881c.jpg) + +![](images/237f4c875b4324a67acb1696a3f2b9296622fc37220d16b2426bea81ace0dfe5.jpg) + +![](images/244c8f0d25ffe5a9aef5bf2ce195a999ebf28fb483a0b9d07a0e04c3a03bfac7.jpg) + +![](images/ca7caebd14fc05883ffa33da300861836f97f6a52567ead526753691c97f72a9.jpg) +TediGAN Composable + +![](images/b46c9420b617f1d743b96885c32a4fd385a3bf2cc7829a5ca925862a8f5c04f5.jpg) +Ours + +![](images/7174d87008736e4024c323dc9a414840ad5a48a8aa63f2401450cb976bb55686.jpg) +Input Image + +Multi-Modal Target Conditions +![](images/e8b4d0e2ed709a0cfb1a85472acf26cba3d8b7fc7133dd44fb4a22889b715598.jpg) +He is a teen. The face is covered with short pointed beard. + +![](images/3520e2288f41faa1c38bc6c6e2d5a215550965b1cb2f5c5221d6aa674f478faf.jpg) +Figure 5. Qualitative Comparison of Face Editing. While TediGAN's hair shape is inconsistent with mask, and Composable fails to generate beard according to text, our framework produces results highly consistent with both conditions while maintaining identity. + +![](images/b9bcb15e50f9ce26476b5d0cdbe998e1d22705c07102469e0a265325bb676f28.jpg) +This man has +heard of medium +length. He is in his +thirties. + +![](images/1b5fbed737513bb8fb7e433512dd81419bad04689ea696f726badb309451bc68.jpg) + +![](images/e4f47afe8cd0521502406ea665e0807f2b350d035dc45346fb3e3bbb226cb3a5.jpg) +Edited Images + +![](images/6d67adf7d26a0464be27cd72f9160cafa604671f957b1a2eb987a89652a6fab6.jpg) + +![](images/52d0b0cdf1f55777c71752121eec113d75a6668d743f0fd746646696290b6fa7.jpg) +TediGAN + +![](images/59bec688eed7b918af67b7643a945ec0611547dda362e23b37878875dd8f68f8.jpg) +Composable + +![](images/860490e9e49f19e2c9acbeb7b30ef2470cca211c67d3c1870bdb1c4734d8df04.jpg) +Ours + +# 4. Experiments + +# 4.1. Experimental Setup + +Datasets. CelebAMask-HQ [32] consists of manually annotated segmentation masks for each of the 30,000 images in the CelebA-HQ dataset [25]. Each mask has up to 19 classes, including the main facial components such as hair, skin, eyes, and nose, and accessories such as eyeglasses and cloth. CelebA-Dialog [23] provides rich and fine-grained natural language descriptions for the images in CelebA-HQ. In this work, we train the mask-driven diffusion models on CelebAMask-HQ, and the text-driven models on CelebA-Dialog. For dynamic diffuser, we simply combine the mask and text for each image without further processing. + +Implementation Details. We adopt LDM [48] for our unimodal diffusion models since it achieves a good balance between quality and speed. Our dynamic diffuser is a time-conditional UNet [49], with conditional embeddings injected via cross-attention [62]. The dynamic diffuser is $\times 30$ smaller than the pre-trained diffusion model. The detailed architecture and settings are provided in the supplementary material. + +# 4.2. Comparison Methods + +TediGAN [67, 68] is a StyleGAN-based face synthesis and manipulation method. It projects both text and mask condi + +![](images/a33367654f1906533af4ab8fb28400b3d09af7ec3b3a6b192cfc9a2ed3b675af.jpg) +Multi-Modal Conditions +His face is covered with short beard. He is a young adult. + +![](images/bb022db63d2e0adda2410a35ba889f6ac9bd6f3162bb0874c76831e972c85b6b.jpg) +She looks very young. +Figure 6. Diversity of Face Generation. Given input conditions, our method demonstrates promising synthesis diversity in facial attributes that are not constrained by a specific combination of multi-modal conditions, such as hair color and skin tone. + +![](images/c9264fd09f2d4328f6efb8749e2dc8ee0e2d5b23ecadadc7c17a8ced9229def2.jpg) + +![](images/d6bdd71d45001981bf303c32e376a35b43cdb345ed551ce66b1759ae390ba564.jpg) + +![](images/7c6e471ca7ab0380639a28fb4222c278a84db6fca4f27f478c051160ce6283c1.jpg) + +![](images/d1108f04cff15f86373edadb91d4f5b793e4a515af9316e627e781f69920e8b5.jpg) +Generated Images + +![](images/b85a133ac7ded7c791d388dbfa9f44caa22fd7677041c0a7b3ed5d350c70aa8c.jpg) + +![](images/46dad76165ae071f27551e8d0dec4b66c0d989c27f8a5df9269f1031e9285611.jpg) + +![](images/afdddf07b2a9fd2b480f5c17adad67e692fe0a1b3dd69a58926a66ff2833e285.jpg) + +![](images/67e6e83f4b5991a0da8c9fc8a071b75a44625193cb63e67c7f0ab1da68894c0e.jpg) + +![](images/fef8961c5e3fca360eaa8f50a9b596e7a12d577ddf0310ac8d1e04b23721cdbe.jpg) + +![](images/357fb1a75d26255bfc3c028468306284eae275f3d7cc38eb3a4580533b7a1f36.jpg) + +![](images/cbf82fbd3a6cf6eb8ad9a14c73c25df05e5811b405e6a94342def08bd6f47b7a.jpg) + +![](images/0289ddbb97f925db0dfc73a83794eccfaca1873b93a33184ffa6e2b0fcb46318.jpg) + +![](images/040d206d3991b0e0e22dd2f0addd5bccadac9e3fa335e7ead1143191b1d397d7.jpg) + +![](images/41d152e219859cd14bd89155bccd6c42522d5d3c5664a81541b0d96a3550c19c.jpg) + +![](images/a0cd8a2c011fb9b0d48c4f97a0a8b714872c57904954fd15e57152671df61a00.jpg) + +![](images/a5585141c1f994de37bfc736e02d3d1993d86ff76f9877870f1e103d454f991e.jpg) + +![](images/ad75a40e67e59702546d2b24179c997660cde30ca8fe0e82eceb8e3d2695aab0.jpg) + +![](images/c7b47b8845ca6d621f22a0b1fa5705cfb6b11abd21bb38cbd37d1494e088674c.jpg) + +tions into StyleGAN's $\mathcal{W}+$ latent space, and performs style mixing to achieve multi-modal control. + +Composable Diffusion [36] parallelly implements two instances of the same text-to-image diffusion model to perform compositional scene generation. We generalize it by replacing the two text-to-image model instances by the text-driven model and the mask-driven model trained using LDM [48]. The two models are composed by averaging the intermediate diffusion results at each step of the reverse process. + +# 4.3. Evaluation Metrics + +FID. We use Frechet Inception Distance (FID) [18] to measure the quality of images synthesized by different methods. FID computes the feature representation's distance between generated images and real images. A lower FID implies better sample quality. + +CLIP Score. CLIP [42] is a vision-language model trained on large-scale datasets. It uses an image encoder and a text encoder to project images and texts to a common feature space, respectively. The CLIP score is computed as the cosine similarity between the normalized image and text embeddings. A higher score usually indicates higher consistency between the output image and the text caption. + +Mask Accuracy. For each output image, we predict the segmentation mask using the face parsing network provided by CelebAMask-HQ [32]. Mask accuracy is the pixel-wise accuracy against the ground-truth segmentation. A higher average accuracy indicates better consistency between the output image and the segmentation mask. + +User Study. We conduct user study with 25 human evaluators to measure the effectiveness of the methods perceptually. For face generation, we randomly sample multi-modal conditions in the validation split of CelebA-HQ Dataset, and then synthesize output images given the conditions. Evaluators are provided with the input conditions and the output images, and they are asked to choose the best image based on 1) image photo-realism, 2) image-text consistency, 3) image-mask consistency. User study is conducted similarly for face editing, except that 1) evaluators are also provided with the input image for editing, and 2) they are also asked + +Table 1. Quantitative Results of Face Generation. Compared with TediGAN and Composable Diffusion, our method synthesizes images with better quality (lower FID), and higher consistency with the text and mask conditions. + +
MethodFID ↓Text (%) ↑Mask (%) ↑
TediGAN [67,68]157.8124.2772.19
Composable [36]124.6223.9476.11
Ours111.3624.5180.25
+ +![](images/3d7719ce0812780236e50b03b2ace2e8717640c2d109eddd3949b7ae090b9495.jpg) +(a) Face Generation +Figure 7. User Study. Among the three methods, the majority of users vote for our results as the best for both generation (69.40%) and editing (84.37%), in terms of image quality, condition consistency, and identity preservation. + +(b) Face Editing +![](images/d7f016d4115b42bf87fe163aa49feb0ab08580371afafecedc03ea9308f1dcc8.jpg) +TediGAN +Composable +Ours + +to assess the level of identity preservation between the edited image and input image. + +# 4.4. Comparative Study + +Qualitative Comparison. In Figure 8, we provide the generation results of our Collaborative Diffusion under different combinations of texts and masks. It is observed that our framework is capable of synthesizing realistic outputs consistent with the multi-modal inputs, even for relatively rare conditions, such as a man with long hair. We then compare our method with TediGAN [67, 68] and Composable Diffusion [36] in Figure 4. For example, in the second example, TediGAN and Composable Diffusion's results are inconsistent with the mask, while ours is able to maintain both mask and text consistency. We also show in Figure 6 that our method synthesizes images with high diversity without losing condition consistency. + +In addition to image generation, it is also observed in + +![](images/6ae37ff451b6a6749125bc47f529b2c1727b26ec0107413352237b6e84634ed1.jpg) +Figure 8. Face Generation Results. Our method generates realistic images under different combinations of multi-modal conditions, even for relatively rare combinations in the training distribution, such as a man with long hair. + +Figure 9 that our framework is able to edit images based on multi-modal conditions while preserving identity. The comparison to existing works in Figure 5 also verifies our effectiveness. For example, in the second example, while TediGAN is unable to synthesize hair consistent to the mask and Composable Diffusion fails to generate beard according to the text, our framework is able to generate results highly consistent to both conditions while maintaining identity. + +Quantitative Comparison. As shown in Table 1, our Collaborate Diffusion outperforms TediGAN and Composable Diffusion in all three objective metrics. In addition, as depicted in Figure 7, our framework achieves the best result in $69.40\%$ and $84.37\%$ of the time for face generation and editing, respectively. These results verify our effectiveness in both image quality and image-condition consistency. + +# 4.5. Ablation Study + +In this section, we visualize the influence functions, and show that it is necessary for the influence function to be both spatially and temporally varying to facilitate effective collaboration. + +# 4.5.1 Spatial Variations of Influence Functions + +As shown in Figure 10, the influence functions behave differently at different spatial regions. For instance, the influence for the mask-driven model mainly lies on the contours of facial regions, such as the outline of hair, face, and eyes, as these regions are crucial in defining facial layout. In contrast, the influence for the text-driven model is stronger at skin regions including cheeks and chin. This is because the attributes related skin texture, such as age and beard length, are better described by text. + +To verify its necessity, we remove the spatial variation of influence functions. From Table 2 we observe that removing the spatial variance results in deterioration in both output quality and condition consistency. This corroborates our hypothesis that it is important to assign different weights to different modalities at different spatial regions. + +# 4.5.2 Temporal Variations of Influence Functions + +In addition to the spatial variation, it is also observed in Figure 10 that the influence from the mask-driven model is stronger at earlier diffusion stages (i.e., larger $t$ ), since early stages focus on initializing the facial layout using the mask- + +![](images/57068522a380006889e26d06ddfccbfc8f02c36f0ff3ee1f9979eb113625785d.jpg) +Figure 9. Face Editing Results. Given the input real image and target conditions, we display the edited image using our method. + +![](images/1c37642a1c99a2a65435f81a57bac20a5395e200b8834a938f2dfcb76a5fbe5f.jpg) +Figure 10. Visualization of Influence Functions. The influence functions vary spatially at different face regions, and temporally at different diffusion timesteps. + +Table 2. Ablation Study. Temporal or spatial suppression in influence variation introduces performance drops, which shows the necessity of influence functions' spatial-temporal adaptivity. + +
MethodFID ↓Text (%) ↑Mask (%) ↑
Ours w/o Spatial117.8124.3680.08
Ours w/o Temporal117.3424.4877.07
Ours111.3624.5180.25
+ +driven model's predictions. At later stages, the influence from the text-driven model increases as the textural details (e.g., skin wrinkles and beard length) are instantiated using information from the text. + +We remove the temporal variation in the influence functions to demonstrate its importance. It is shown in Table 2 that both the quality and condition consistency drop without the temporal variation. + +# 5. Conclusion + +With generative AI gaining increasing attention, multimodal conditioning becomes an indispensable direction to unleash creativity and enable comprehensive controls from human creators. In this work, we take the first step forward and propose Collaborative Diffusion, where pre-trained uni-modal diffusion models collaboratively achieve multi-modal face generation and editing without being re-trained. With our dynamic diffuser, this framework could be used to extend arbitrary uni-modal approach to the multi-modal paradigm through predicting the relative influence of different modalities. We believe our idea of synergizing uni-modal models for multi-modal tasks would be a good inspiration for future works in various domains such as motion and 3D generation. + +Limitations and Future Work. Since our framework focuses on exploiting pre-trained diffusion models, our performance is dependent on the capability of each model. An orthogonal direction is to train each collaborator on large-scale datasets for performance gain. + +Potential Negative Societal Impacts. The facial manipulation capabilities of Collaborative Diffusion could be applied maliciously on real human faces. We advise users to apply it only for proper recreational purposes. + +Acknowledgement. We would like to thank Liangyu Chen and Bo Li for proofreading. This study is supported by NTU NAP, MOE AcRF Tier 1 (2021-T1-001-088), and under the RIE2020 Industry Alignment Fund - Industry Collaboration Projects (IAF-ICP) Funding Initiative, as well as cash and in-kind contribution from the industry partner(s). + +# References + +[1] Tomer Amit, Eliya Nachmani, Tal Shaharbany, and Lior Wolf. SegDiff: Image segmentation with diffusion probabilistic models. arXiv preprint arXiv:2112.00390, 2021. 2 +[2] Jacob Austin, Daniel D Johnson, Jonathan Ho, Daniel Tarlow, and Rianne van den Berg. Structured denoising diffusion models in discrete state-spaces. In NeurIPS, 2021. 2 +[3] Omri Avrahami, Dani Lischinski, and Ohad Fried. Blended diffusion for text-driven editing of natural images. In CVPR, 2022. 1 +[4] Dmitry Baranchuk, Ivan Rubachev, Andrey Voynov, Valentin Khrulkov, and Artem Babenko. Label-efficient semantic segmentation with diffusion models. In ICLR, 2022. 2 +[5] Sam Bond-Taylor, Peter Hessey, Hiroshi Sasaki, Toby P Breckon, and Chris G Willcocks. Unleashing transformers: parallel token prediction with discrete absorbing diffusion for fast high-resolution image generation from vector-quantized codes. In ECCV, 2022. 2 +[6] Huiwen Chang, Han Zhang, Lu Jiang, Ce Liu, and William T Freeman. MaskGIT: Masked generative image transformer. In CVPR, 2022. 2 +[7] Shin-I Cheng, Yu-Jie Chen, Wei-Chen Chiu, Hung-Yu Tseng, and Hsin-Ying Lee. Adaptively-realistic image generation from stroke and sketch with diffusion model. In WACV, 2023. 1 +[8] Guillaume Couairon, Jakob Verbeek, Holger Schwenk, and Matthieu Cord. Diffedit: Diffusion-based semantic image editing with mask guidance. arXiv preprint arXiv:2210.11427, 2022. 2 +[9] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat GANs on image synthesis. In NeurIPS, 2021. 1, 2 +[10] Ming Ding, Zhuoyi Yang, Wenyi Hong, Wendi Zheng, Chang Zhou, Da Yin, Junyang Lin, Xu Zou, Zhou Shao, Hongxia Yang, et al. CogView: Mastering text-to-image generation via transformers. In NeurIPS, 2021. 2 +[11] Patrick Esser, Robin Rombach, Andreas Blattmann, and Bjorn Ommer. ImageBART: Bidirectional context with multinomial diffusion for autoregressive image synthesis. In NeurIPS, 2021. 2 +[12] Patrick Esser, Robin Rombach, and Bjorn Ommer. Taming transformers for high-resolution image synthesis. In CVPR, 2021. 2 +[13] Oran Gafni, Adam Polyak, Oron Ashual, Shelly Sheynin, Devi Parikh, and Yaniv Taigman. Make-a-scene: Scene-based text-to-image generation with human priors. arXiv preprint arXiv:2203.13131, 2022. 2 +[14] Ian J Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron C Courville, and Yoshua Bengio. Generative adversarial nets. In NeurIPS, 2014. 2 +[15] Alexandros Graikos, Nikolay Malkin, Nebojsa Jojic, and Dimitris Samaras. Diffusion models as plug-and-play priors. In NeurIPS, 2022. 2 +[16] Shuyang Gu, Dong Chen, Jianmin Bao, Fang Wen, Bo Zhang, Dongdong Chen, Lu Yuan, and Baining Guo. Vector quantized diffusion model for text-to-image synthesis. In CVPR, 2022. 1, 2 + +[17] William Harvey, Saeid Naderiparizi, Vaden Masrani, Christian Weilbach, and Frank Wood. Flexible diffusion modeling of long videos. arXiv preprint arXiv:2205.11495, 2022. 2 +[18] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. GANs trained by a two time-scale update rule converge to a local nash equilibrium. In NeurIPS, 2017. 6 +[19] Jonathan Ho, William Chan, Chitwan Sahara, Jay Whang, Ruiqi Gao, Alexey Gritsanko, Diederik P Kingma, Ben Poole, Mohammad Norouzi, David J Fleet, et al. Imagen video: High definition video generation with diffusion models. arXiv preprint arXiv:2210.02303, 2022. 2 +[20] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In NeurIPS, 2020. 1, 2, 3 +[21] Jonathan Ho, Chitwan Sahara, William Chan, David J Fleet, Mohammad Norouzi, and Tim Salimans. Cascaded diffusion models for high fidelity image generation. JMLR, 2022. 2 +[22] Aapo Hyvarinen and Peter Dayan. Estimation of non-normalized statistical models by score matching. JMLR, 2005. 2 +[23] Yuming Jiang, Ziqi Huang, Xingang Pan, Chen Change Loy, and Ziwei Liu. Talk-to-Edit: Fine-grained facial editing via dialog. In ICCV, 2021. 5 +[24] Yuming Jiang, Shuai Yang, Haonan Qju, Wayne Wu, Chen Change Loy, and Ziwei Liu. Text2human: Text-driven controllable human image generation. ACM TOG, 2022. 2 +[25] Tero Karras, Timo Aila, Samuli Laine, and Jaakko Lehtinen. Progressive growing of GANs for improved quality, stability, and variation. In ICLR, 2018. 5 +[26] Tero Karras, Miika Aittala, Samuli Laine, Erik Härkönen, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Alias-free generative adversarial networks. In NeurIPS, 2021. 2 +[27] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In CVPR, 2019. 2 +[28] Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of StyleGAN. In CVPR, 2020. 2 +[29] Bahjat Kawar, Shiran Zada, Oran Lang, Omer Tov, Huiwen Chang, Tali Dekel, Inbar Mosseri, and Michal Irani. Imagic: Text-based real image editing with diffusion models. arXiv preprint arXiv:2210.09276, 2022. 1, 2, 5 +[30] Gwanghyun Kim, Taesung Kwon, and Jong Chul Ye. DiffusionCLIP: Text-guided diffusion models for robust image manipulation. In CVPR, 2022. 1, 2 +[31] Jing Yu Koh, Jason Baldridge, Honglak Lee, and Yinfei Yang. Text-to-image generation grounded by fine-grained user attention. In WACV, 2021. 2 +[32] Cheng-Han Lee, Ziwei Liu, Lingyun Wu, and Ping Luo. MaskGAN: Towards diverse and interactive facial image manipulation. In CVPR, 2020. 2, 5, 6 +[33] Bowen Li, Xiaojuan Qi, Thomas Lukasiewicz, and Philip HS Torr. ManiGAN: Text-guided image manipulation. In CVPR, 2020. 2 +[34] Wenbo Li, Pengchuan Zhang, Lei Zhang, Qiuyuan Huang, Xiaodong He, Siwei Lyu, and Jianfeng Gao. Object-driven text-to-image synthesis via adversarial training. In CVPR, 2019. 2 + +[35] Zhiheng Li, Martin Renqiang Min, Kai Li, and Chenliang Xu. StyleT2I: Toward compositional and high-fidelity text-to-image synthesis. In CVPR, 2022. 2 +[36] Nan Liu, Shuang Li, Yilun Du, Antonio Torralba, and Joshua B Tenenbaum. Compositional visual generation with composable diffusion models. arXiv preprint arXiv:2206.01714, 2022. 4, 6 +[37] Xihui Liu, Dong Huk Park, Samaneh Azadi, Gong Zhang, Arman Chopikyan, Yuxiao Hu, Humphrey Shi, Anna Rohrbach, and Trevor Darrell. More control for free! image synthesis with semantic diffusion guidance. In WACV, 2023. 2 +[38] Chenlin Meng, Yutong He, Yang Song, Jiaming Song, Jiajun Wu, Jun-Yan Zhu, and Stefano Ermon. SDEdit: Guided image synthesis and editing with stochastic differential equations. In ICLR, 2022. 2 +[39] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. GLIDE: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741, 2021. 1 +[40] Taesung Park, Ming-Yu Liu, Ting-Chun Wang, and Jun-Yan Zhu. Semantic image synthesis with spatially-adaptive normalization. In CVPR, 2019. 2 +[41] Or Patashnik, Zongze Wu, Eli Shechtman, Daniel Cohen-Or, and Dani Lischinski. StyleCLIP: Text-driven manipulation of stylegan imagery. In ICCV, 2021. 2 +[42] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, 2021. 6 +[43] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with CLIP latents. arXiv preprint arXiv:2204.06125, 2022. 2 +[44] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In ICML, 2021. 2 +[45] Scott Reed, Zeynep Akata, Xinchen Yan, Lajanugen Logeswaran, Bernt Schiele, and Honglak Lee. Generative adversarial text to image synthesis. In ICML, 2016. 2 +[46] Scott E Reed, Zeynep Akata, Santosh Mohan, Samuel Tenka, Bernt Schiele, and Honglak Lee. Learning what and where to draw. In NeurIPS, 2016. 2 +[47] Elad Richardson, Yuval Alaluf, Or Patashnik, Yotam Nitzan, Yaniv Azar, Stav Shapiro, and Daniel Cohen-Or. Encoding in style: a StyleGAN encoder for image-to-image translation. In CVPR, 2021. 2 +[48] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 1, 2, 5, 6 +[49] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-Net: Convolutional networks for biomedical image segmentation. In MICCAI, 2015. 4, 5 +[50] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine + +tuning text-to-image diffusion models for subject-driven generation. arXiv preprint arXiv:2208.12242, 2022. 2 +[51] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily Denton, Seyed Kamyar Seyed Ghasemipour, Burcu Karagol Ayan, S Sara Mahdavi, Rapha Gontijo Lopes, et al. Photorealistic text-to-image diffusion models with deep language understanding. arXiv preprint arXiv:2205.11487, 2022. 1, 2 +[52] Chitwan Sahara, Jonathan Ho, William Chan, Tim Salimans, David J Fleet, and Mohammad Norouzi. Image superresolution via iterative refinement. IEEE TPAMI, 2022. 2 +[53] Yujun Shen, Jinjin Gu, Xiaou Tang, and Bolei Zhou. Interpreting the latent space of GANs for semantic face editing. In CVPR, 2020. 2 +[54] Yujun Shen, Ceyuan Yang, Xiaou Tang, and Bolei Zhou. InterFaceGAN: Interpreting the disentangled face representation learned by GANs. IEEE TPAMI, 2020. 2 +[55] Uriel Singer, Adam Polyak, Thomas Hayes, Xi Yin, Jie An, Songyang Zhang, Qiyuan Hu, Harry Yang, Oron Ashual, Oran Gafni, et al. Make-a-video: Text-to-video generation without text-video data. arXiv preprint arXiv:2209.14792, 2022. 2 +[56] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In ICML, 2015. 1, 2 +[57] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In ICLR, 2021. 2 +[58] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. In ICLR, 2021. 1, 2 +[59] Yang Song, Jingwen Zhu, Dawei Li, Xiaolong Wang, and Hairong Qi. Talking face generation by conditional recurrent adversarial network. *IJCAI*, 2019. 2 +[60] Jianxin Sun, Qiyao Deng, Qi Li, Muyi Sun, Min Ren, and Zhenan Sun. AnyFace: Free-style text-to-face synthesis and manipulation. In CVPR, 2022. 2 +[61] Aaron Van Den Oord, Oriol Vinyals, et al. Neural discrete representation learning. In NeurIPS, 2017. 2 +[62] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In NeurIPS, 2017. 5 +[63] Ruben Villegas, Mohammad Babaeizadeh, Pieter-Jan Kindermans, Hernan Moraldo, Han Zhang, Mohammad Taghi Saffar, Santiago Castro, Julius Kunze, and Dumitru Erhan. Phenaki: Variable length video generation from open domain textual description. arXiv preprint arXiv:2210.02399, 2022. 2 +[64] Pascal Vincent. A connection between score matching and denoising autoencoders. Neural Computation, 2011. 2 +[65] Tengfei Wang, Ting Zhang, Bo Zhang, Hao Ouyang, Dong Chen, Qifeng Chen, and Fang Wen. Pretraining is all you need for image-to-image translation. arXiv preprint arXiv:2205.12952, 2022. 1, 2 +[66] Weilun Wang, Jianmin Bao, Wengang Zhou, Dongdong Chen, Dong Chen, Lu Yuan, and Houqiang Li. Semantic image synthesis via diffusion models. arXiv preprint arXiv:2207.00050, 2022. 1 + +[67] Weihao Xia, Yujiu Yang, Jing-Hao Xue, and Baoyuan Wu. TediGAN: Text-guided diverse face image generation and manipulation. In CVPR, 2021. 2, 5, 6 +[68] Weihao Xia, Yujiu Yang, Jing-Hao Xue, and Baoyuan Wu. Towards open-world text-guided face image generation and manipulation. arXiv preprint arXiv:2104.08910, 2021. 2, 5, 6 +[69] Tao Xu, Pengchuan Zhang, Qiuyuan Huang, Han Zhang, Zhe Gan, Xiaolei Huang, and Xiaodong He. AttnGAN: Fine-grained text to image generation with attentional generative adversarial networks. In CVPR, 2018. 2 +[70] Han Zhang, Tao Xu, Hongsheng Li, Shaoting Zhang, Xiaogang Wang, Xiaolei Huang, and Dimitris N Metaxas. StackGAN: Text to photo-realistic image synthesis with stacked generative adversarial networks. In ICCV, 2017. 2 +[71] Han Zhang, Tao Xu, Hongsheng Li, Shaoting Zhang, Xiaogang Wang, Xiaolei Huang, and Dimitris N Metaxas. StackGAN++: Realistic image synthesis with stacked generative adversarial networks. IEEE TPAMI, 2018. 2 \ No newline at end of file diff --git a/2023/Collaborative Diffusion for Multi-Modal Face Generation and Editing/images.zip b/2023/Collaborative Diffusion for Multi-Modal Face Generation and Editing/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..fb4f442006367560359062e0961f6a194e5bfab2 --- /dev/null +++ b/2023/Collaborative Diffusion for Multi-Modal Face Generation and Editing/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:450f4c0bbd67353b7de80ea4979c821bcc9ae48002f6fdd378a94eed13c80538 +size 866844 diff --git a/2023/Collaborative Diffusion for Multi-Modal Face Generation and Editing/layout.json b/2023/Collaborative Diffusion for Multi-Modal Face Generation and Editing/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..12518a254c95422e6eb140118a023b87f92da00b --- /dev/null +++ b/2023/Collaborative Diffusion for Multi-Modal Face Generation and Editing/layout.json @@ -0,0 +1,12428 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 81, + 73, + 512, + 91 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 81, + 73, + 512, + 91 + ], + "spans": [ + { + "bbox": [ + 81, + 73, + 512, + 91 + ], + "type": "text", + "content": "Collaborative Diffusion for Multi-Modal Face Generation and Editing" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 141, + 114, + 451, + 141 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 114, + 451, + 141 + ], + "spans": [ + { + "bbox": [ + 141, + 114, + 451, + 141 + ], + "type": "text", + "content": "Ziqi Huang Kelvin C.K. Chan Yuming Jiang Ziwei Liu S-Lab, Nanyang Technological University" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 156, + 143, + 436, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 156, + 143, + 436, + 156 + ], + "spans": [ + { + "bbox": [ + 156, + 143, + 436, + 156 + ], + "type": "text", + "content": "{ziqi002, chan0899, yuming002, ziwei.liu}@ntu.edu.sg" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 50, + 182, + 545, + 392 + ], + "blocks": [ + { + "bbox": [ + 50, + 182, + 545, + 392 + ], + "lines": [ + { + "bbox": [ + 50, + 182, + 545, + 392 + ], + "spans": [ + { + "bbox": [ + 50, + 182, + 545, + 392 + ], + "type": "image", + "image_path": "3bf3231c7177e7c670a9268704a60932cf22a3678547ecc6519ec94df8742f14.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 399, + 546, + 434 + ], + "lines": [ + { + "bbox": [ + 46, + 399, + 546, + 434 + ], + "spans": [ + { + "bbox": [ + 46, + 399, + 546, + 434 + ], + "type": "text", + "content": "Figure 1. We propose Collaborative Diffusion, where users can use multiple modalities to control face generation and editing. (a) Face Generation. Given multi-modal controls, our framework synthesizes high-quality images consistent with the input conditions. (b) Face Editing. Collaborative Diffusion also supports multi-modal editing of real images with promising identity preservation capability." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 143, + 448, + 192, + 460 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 448, + 192, + 460 + ], + "spans": [ + { + "bbox": [ + 143, + 448, + 192, + 460 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 467, + 288, + 563 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 467, + 288, + 563 + ], + "spans": [ + { + "bbox": [ + 46, + 467, + 288, + 563 + ], + "type": "text", + "content": "Diffusion models arise as a powerful generative tool recently. Despite the great progress, existing diffusion models mainly focus on uni-modal control, i.e., the diffusion process is driven by only one modality of condition. To further unleash the users' creativity, it is desirable for the model to be controllable by multiple modalities simultaneously, e.g. generating and editing faces by describing the age (text-driven) while drawing the face shape (mask-driven)." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 563, + 288, + 672 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 563, + 288, + 672 + ], + "spans": [ + { + "bbox": [ + 46, + 563, + 288, + 672 + ], + "type": "text", + "content": "In this work, we present Collaborative Diffusion, where pre-trained uni-modal diffusion models collaborate to achieve multi-modal face generation and editing without re-training. Our key insight is that diffusion models driven by different modalities are inherently complementary regarding the latent denoising steps, where bilateral connections can be established upon. Specifically, we propose dynamic diffuser, a meta-network that adaptively hallucinates multi-modal denoising steps by predicting the spatial-temporal" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 449, + 547, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 449, + 547, + 533 + ], + "spans": [ + { + "bbox": [ + 306, + 449, + 547, + 533 + ], + "type": "text", + "content": "influence functions for each pre-trained uni-modal model. Collaborative Diffusion not only collaborates generation capabilities from uni-modal diffusion models, but also integrates multiple uni-modal manipulations to perform multimodal editing. Extensive qualitative and quantitative experiments demonstrate the superiority of our framework in both image quality and condition consistency." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 307, + 559, + 386, + 571 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 559, + 386, + 571 + ], + "spans": [ + { + "bbox": [ + 307, + 559, + 386, + 571 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 580, + 547, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 580, + 547, + 700 + ], + "spans": [ + { + "bbox": [ + 304, + 580, + 547, + 700 + ], + "type": "text", + "content": "Recent years have witnessed substantial progress in image synthesis and editing with the surge of diffusion models [9, 20, 56, 58]. In addition to the remarkable synthesis quality, one appealing property of diffusion models is the flexibility of conditioning on various modalities, such as texts [3, 16, 29, 30, 39, 48, 51], segmentation masks [48, 65, 66], and sketches [7, 65]. However, existing explorations are largely confined to the use of a single modality at a time. The exploitation of multiple conditions remains under-explored. As a generative tool, its controllability is still limited." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 318, + 701, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 701, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 318, + 701, + 545, + 713 + ], + "type": "text", + "content": "To unleash users' creativity, it is desirable that the model" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 495, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 495, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 495, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 50, + 685, + 274, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 50, + 685, + 274, + 713 + ], + "spans": [ + { + "bbox": [ + 50, + 685, + 274, + 713 + ], + "type": "text", + "content": "Corresponding author. \nProject page: https://ziqihuangg.github.io/projects/collaborative-diffusion.html \nCode: https://github.com/ziqihuangg/Collaborative-Diffusion" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6080" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 215 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 215 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 215 + ], + "type": "text", + "content": "is simultaneously controllable by multiple modalities. While it is trivial to extend the current supervised framework with multiple modalities, training a large-scale model from scratch is computationally expensive, especially when extensive hyper-parameter tuning and delicate architecture designs are needed. More importantly, each trained model could only accept a fixed combination of modalities, and hence re-training is necessary when a subset of modalities are absent, or when additional modalities become available. The above demonstrates the necessity of a unified framework that effectively exploits pre-trained models and integrate them for multi-modal synthesis and editing." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 216, + 289, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 216, + 289, + 479 + ], + "spans": [ + { + "bbox": [ + 46, + 216, + 289, + 479 + ], + "type": "text", + "content": "In this paper, we propose Collaborative Diffusion, a framework that synergizes pre-trained uni-modal diffusion models for multi-modal face generation and editing without the need of re-training. Motivated by the fact that different modalities are complementary to each other (e.g., text for age and mask for hair shape), we explore the possibility of establishing lateral connections between models driven by different modalities. We propose dynamic diffuser to adaptively predict the spatial-temporal influence function for each pre-trained model. The dynamic diffuser dynamically determines spatial-varying and temporal-varying influences of each model, suppressing contributions from irrelevant modalities while enhancing contributions from admissible modalities. In addition to multi-modal synthesis, the simplicity and flexibility of our framework enable extension to multi-modal face editing with minimal modifications. In particular, the dynamic diffuser is first trained for collaborative synthesis. It is then fixed and combined with existing face editing approaches [29, 50] for multi-modal editing. It is worth-mentioning that users can select the best editing approaches based on their needs without the need of altering the dynamic diffusers." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 479, + 288, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 479, + 288, + 525 + ], + "spans": [ + { + "bbox": [ + 46, + 479, + 288, + 525 + ], + "type": "text", + "content": "We demonstrate both qualitatively and quantitatively that our method achieves superior image quality and condition consistency in both synthesis and editing tasks. Our contributions can be summarized as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 59, + 528, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 59, + 528, + 287, + 587 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 528, + 287, + 587 + ], + "spans": [ + { + "bbox": [ + 59, + 528, + 287, + 587 + ], + "type": "text", + "content": "- We introduce Collaborative Diffusion, which exploits pre-trained uni-modal diffusion models for multi-modal controls without re-training. Our approach is the first attempt towards flexible integration of uni-modal diffusion models into a single collaborative framework." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 59, + 590, + 287, + 650 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 590, + 287, + 650 + ], + "spans": [ + { + "bbox": [ + 59, + 590, + 287, + 650 + ], + "type": "text", + "content": "Tailored for the iterative property of diffusion models, we propose dynamic diffuser, which predicts the spatial-varying and temporal-varying influence functions to selectively enhance or suppress the contributions of the given modalities at each iterative step." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 59, + 654, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 654, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 59, + 654, + 287, + 713 + ], + "type": "text", + "content": "- We demonstrate the flexibility of our framework by extending it to face editing driven by multiple modalities. Both quantitative and qualitative results demonstrate the superiority of Collaborative Diffusion in multi-modal face generation and editing." + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 71, + 392, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 71, + 392, + 84 + ], + "spans": [ + { + "bbox": [ + 306, + 71, + 392, + 84 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 91, + 547, + 281 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 91, + 547, + 281 + ], + "spans": [ + { + "bbox": [ + 304, + 91, + 547, + 281 + ], + "type": "text", + "content": "Diffusion Models. Diffusion models [20, 56, 58] have recently become a mainstream approach for image synthesis [9, 11, 38] apart from Generative Adversarial Networks (GANs) [14], and success has also been found in various domains including video generation [17, 19, 55, 63], image restoration [21, 52], semantic segmentation [1, 4, 15], and natural language processing [2]. In the diffusion-based framework, models are trained with score-matching objectives [22, 64] at various noise levels, and sampling is done via iterative denoising. Existing works focus on improving the performance and efficiency of diffusion models through enhanced architecture designs [16, 48] and sampling schemes [57]. In contrast, this work focuses on exploiting existing models, and providing a succinct framework for multi-modal synthesis and editing without large-scale re-training of models." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 283, + 548, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 283, + 548, + 498 + ], + "spans": [ + { + "bbox": [ + 304, + 283, + 548, + 498 + ], + "type": "text", + "content": "Face Generation. Existing face generation approaches can be divided into three main directions. Following the GAN paradigm, the StyleGAN series [26-28] boost the quality of facial synthesis, and provide an interpretable latent space for steerable style controls and manipulations. The vector-quantized approaches [12, 61] learn a discrete codebook by mapping the input images into a low-dimensional discrete feature space. The learned codebook is then sampled, either sequentially [12, 61] or parallelly [5, 6, 16], for synthesis. In contrast to the previous two approaches, diffusion models are trained with a stationary objective, without the need of optimizing complex losses (e.g., adversarial loss) or balancing multiple objectives (e.g., codebook loss versus reconstruction loss). With training simplicity as a merit, diffusion-based approaches have become increasingly popular in recent years. Our framework falls in the diffusion-based paradigm. In particular, we leverage pre-trained diffusion models for multi-modal generation and editing." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 498, + 548, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 498, + 548, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 498, + 548, + 713 + ], + "type": "text", + "content": "Conditional Face Generation and Editing. Conditional generation [10, 11, 13, 31, 34, 37, 40, 43-46, 51, 65, 67-71] and editing [8, 32, 41, 53, 54, 67, 68] is an active line of research focusing on conditioning generative models on different modalities, such as texts [24, 41, 67, 68], segmentation masks [32, 33, 40, 47], and audios [59]. For example, StyleCLIP [41], DiffusionCLIP [30], and many others [35, 60] have demonstrated remarkable performance in text-guided face generation and editing. However, most existing models do not support simultaneous conditioning on multiple modalities (e.g., text and mask at the same time), and supporting additional modalities often requires time-consuming model re-training and extensive hyper-parameter tuning, which are not preferable in general. In this work, we propose Collaborative Diffusion to exploit pre-trained uni-modal diffusion models [48] (e.g., text-driven and mask-driven models) to achieve multi-modal conditioning without model retraining." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 314, + 757 + ], + "type": "text", + "content": "6081" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 48, + 70, + 550, + 281 + ], + "blocks": [ + { + "bbox": [ + 48, + 70, + 550, + 281 + ], + "lines": [ + { + "bbox": [ + 48, + 70, + 550, + 281 + ], + "spans": [ + { + "bbox": [ + 48, + 70, + 550, + 281 + ], + "type": "image", + "image_path": "dbf0ef99a8c6ed2c77ed88a1f99b3f247180cfd54c544982d71242804f33497c.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 286, + 547, + 321 + ], + "lines": [ + { + "bbox": [ + 46, + 286, + 547, + 321 + ], + "spans": [ + { + "bbox": [ + 46, + 286, + 547, + 321 + ], + "type": "text", + "content": "Figure 2. Overview of Collaborative Diffusion. We use pre-trained uni-modal diffusion models to perform multi-modal guided face generation and editing. At each step of the reverse process (i.e., from timestep " + }, + { + "bbox": [ + 46, + 286, + 547, + 321 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 286, + 547, + 321 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 46, + 286, + 547, + 321 + ], + "type": "inline_equation", + "content": "t - 1" + }, + { + "bbox": [ + 46, + 286, + 547, + 321 + ], + "type": "text", + "content": "), the dynamic diffuser predicts the spatial-varying and temporal-varying influence function to selectively enhance or suppress the contributions of the given modality." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 350, + 183, + 363 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 350, + 183, + 363 + ], + "spans": [ + { + "bbox": [ + 47, + 350, + 183, + 363 + ], + "type": "text", + "content": "3. Collaborative Diffusion" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 373, + 290, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 373, + 290, + 483 + ], + "spans": [ + { + "bbox": [ + 46, + 373, + 290, + 483 + ], + "type": "text", + "content": "We propose Collaborative Diffusion, which exploits multiple pre-trained uni-modal diffusion models (Section 3.1) for multi-modal generation and editing. The key of our framework is the dynamic diffuser, which adaptively predicts the influence functions to enhance or suppress the contributions of the pre-trained models based on the spatial-temporal influences of the modalities. Our framework is compatible with most existing approaches for both multi-modal guided synthesis (Section 3.2) and multi-modal editing (Section 3.3)." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 500, + 263, + 513 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 500, + 263, + 513 + ], + "spans": [ + { + "bbox": [ + 47, + 500, + 263, + 513 + ], + "type": "text", + "content": "3.1. Uni-Modal Conditional Diffusion Models" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 523, + 289, + 596 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 523, + 289, + 596 + ], + "spans": [ + { + "bbox": [ + 46, + 523, + 289, + 596 + ], + "type": "text", + "content": "Diffusion models are a class of generative models that model the data distribution in the form of " + }, + { + "bbox": [ + 46, + 523, + 289, + 596 + ], + "type": "inline_equation", + "content": "p_{\\theta}(\\mathbf{x}_0) \\coloneqq \\int p_{\\theta}(\\mathbf{x}_{0:T}) d\\mathbf{x}_{1:T}" + }, + { + "bbox": [ + 46, + 523, + 289, + 596 + ], + "type": "text", + "content": ". The diffusion process (a.k.a. forward process) gradually adds Gaussian noise to the data and eventually corrupts the data " + }, + { + "bbox": [ + 46, + 523, + 289, + 596 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_0" + }, + { + "bbox": [ + 46, + 523, + 289, + 596 + ], + "type": "text", + "content": " into an approximately pure Gaussian noise " + }, + { + "bbox": [ + 46, + 523, + 289, + 596 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_T" + }, + { + "bbox": [ + 46, + 523, + 289, + 596 + ], + "type": "text", + "content": " using a variance schedule " + }, + { + "bbox": [ + 46, + 523, + 289, + 596 + ], + "type": "inline_equation", + "content": "\\beta_1, \\ldots, \\beta_T" + }, + { + "bbox": [ + 46, + 523, + 289, + 596 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 80, + 612, + 288, + 644 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 80, + 612, + 288, + 644 + ], + "spans": [ + { + "bbox": [ + 80, + 612, + 288, + 644 + ], + "type": "interline_equation", + "content": "q \\left(\\mathbf {x} _ {1: T} \\mid \\mathbf {x} _ {0}\\right) := \\prod_ {t = 1} ^ {T} q \\left(\\mathbf {x} _ {t} \\mid \\mathbf {x} _ {t - 1}\\right), \\tag {1}", + "image_path": "88d89dc411754a89c19e9087aaffc63968eb00b1a7f920a9810e66fd9b46e619.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 81, + 647, + 253, + 662 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 81, + 647, + 253, + 662 + ], + "spans": [ + { + "bbox": [ + 81, + 647, + 253, + 662 + ], + "type": "interline_equation", + "content": "q (\\mathbf {x} _ {t} | \\mathbf {x} _ {t - 1}) := \\mathcal {N} (\\mathbf {x} _ {t}; \\sqrt {1 - \\beta_ {t}} \\mathbf {x} _ {t - 1}, \\beta_ {t} \\mathbf {I}).", + "image_path": "952cf86815435864c9fca6b68b5381f3ea9a86fb0229198cdb5d64fe5425de3b.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 677, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 288, + 714 + ], + "type": "text", + "content": "Reversing the forward process allows sampling new data " + }, + { + "bbox": [ + 46, + 677, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_0" + }, + { + "bbox": [ + 46, + 677, + 288, + 714 + ], + "type": "text", + "content": " by starting from " + }, + { + "bbox": [ + 46, + 677, + 288, + 714 + ], + "type": "inline_equation", + "content": "p(\\mathbf{x}_T) = \\mathcal{N}(\\mathbf{x}_T; \\mathbf{0}, \\mathbf{I})" + }, + { + "bbox": [ + 46, + 677, + 288, + 714 + ], + "type": "text", + "content": ". The reverse process is defined as a Markov chain where each step is a learned" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 350, + 428, + 364 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 350, + 428, + 364 + ], + "spans": [ + { + "bbox": [ + 306, + 350, + 428, + 364 + ], + "type": "text", + "content": "Gaussian transition " + }, + { + "bbox": [ + 306, + 350, + 428, + 364 + ], + "type": "inline_equation", + "content": "(\\pmb{\\mu}_{\\theta},\\pmb{\\Sigma}_{\\theta})" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 328, + 369, + 546, + 401 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 328, + 369, + 546, + 401 + ], + "spans": [ + { + "bbox": [ + 328, + 369, + 546, + 401 + ], + "type": "interline_equation", + "content": "p _ {\\theta} \\left(\\mathbf {x} _ {0: T}\\right) := p \\left(\\mathbf {x} _ {T}\\right) \\prod_ {t = 1} ^ {T} p _ {\\theta} \\left(\\mathbf {x} _ {t - 1} \\mid \\mathbf {x} _ {t}\\right), \\tag {2}", + "image_path": "b979b87037395b0ccb1ef94d382028205e30e0c807f01619eb63c2782ca4c592.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 329, + 403, + 521, + 417 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 329, + 403, + 521, + 417 + ], + "spans": [ + { + "bbox": [ + 329, + 403, + 521, + 417 + ], + "type": "interline_equation", + "content": "p _ {\\theta} \\left(\\mathbf {x} _ {t - 1} \\mid \\mathbf {x} _ {t}\\right) := \\mathcal {N} \\left(\\mathbf {x} _ {t - 1}; \\boldsymbol {\\mu} _ {\\theta} \\left(\\mathbf {x} _ {t}, t\\right), \\boldsymbol {\\Sigma} _ {\\theta} \\left(\\mathbf {x} _ {t}, t\\right)\\right).", + "image_path": "75235f9db648bbb604d4bfdba82a58a50aa438d32d28994c5723fd42bf1051b2.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 420, + 548, + 492 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 420, + 548, + 492 + ], + "spans": [ + { + "bbox": [ + 304, + 420, + 548, + 492 + ], + "type": "text", + "content": "Training diffusion models relies on minimizing the variational bound on " + }, + { + "bbox": [ + 304, + 420, + 548, + 492 + ], + "type": "inline_equation", + "content": "p(x)" + }, + { + "bbox": [ + 304, + 420, + 548, + 492 + ], + "type": "text", + "content": "'s negative log-likelihood. The commonly used optimization objective " + }, + { + "bbox": [ + 304, + 420, + 548, + 492 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{DM}}" + }, + { + "bbox": [ + 304, + 420, + 548, + 492 + ], + "type": "text", + "content": " [20] reparameterizes the learnable Gaussian transition as " + }, + { + "bbox": [ + 304, + 420, + 548, + 492 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}(\\cdot)" + }, + { + "bbox": [ + 304, + 420, + 548, + 492 + ], + "type": "text", + "content": ", and temporally reweights the variational bound to trade for better sample quality:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 331, + 496, + 546, + 519 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 331, + 496, + 546, + 519 + ], + "spans": [ + { + "bbox": [ + 331, + 496, + 546, + 519 + ], + "type": "interline_equation", + "content": "L _ {\\mathrm {D M}} (\\theta) := \\mathbb {E} _ {t, \\mathbf {x} _ {0}, \\boldsymbol {\\epsilon} \\sim \\mathcal {N} (\\mathbf {0}, \\mathbf {I})} \\left[ \\| \\boldsymbol {\\epsilon} - \\boldsymbol {\\epsilon} _ {\\theta} (\\mathbf {x} _ {t}, t) \\| ^ {2} \\right], \\quad (3)", + "image_path": "517720367f5b3a9013b14737accba488bd795ee7579bddb7e729ae5095ed0d2e.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 521, + 547, + 546 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 521, + 547, + 546 + ], + "spans": [ + { + "bbox": [ + 305, + 521, + 547, + 546 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 521, + 547, + 546 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_t" + }, + { + "bbox": [ + 305, + 521, + 547, + 546 + ], + "type": "text", + "content": " can be directly approximated by " + }, + { + "bbox": [ + 305, + 521, + 547, + 546 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_t = \\sqrt{\\bar{\\alpha}_t}\\mathbf{x}_0 + \\sqrt{1 - \\bar{\\alpha}_t}\\pmb {\\epsilon}" + }, + { + "bbox": [ + 305, + 521, + 547, + 546 + ], + "type": "text", + "content": ", with " + }, + { + "bbox": [ + 305, + 521, + 547, + 546 + ], + "type": "inline_equation", + "content": "\\bar{\\alpha}_{t}:= \\prod_{s = 1}^{t}\\alpha_{s}" + }, + { + "bbox": [ + 305, + 521, + 547, + 546 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 521, + 547, + 546 + ], + "type": "inline_equation", + "content": "\\alpha_{t}:= 1 - \\beta_{t}" + }, + { + "bbox": [ + 305, + 521, + 547, + 546 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 546, + 547, + 570 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 546, + 547, + 570 + ], + "spans": [ + { + "bbox": [ + 305, + 546, + 547, + 570 + ], + "type": "text", + "content": "To sample data " + }, + { + "bbox": [ + 305, + 546, + 547, + 570 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_0" + }, + { + "bbox": [ + 305, + 546, + 547, + 570 + ], + "type": "text", + "content": " from a trained diffusion model " + }, + { + "bbox": [ + 305, + 546, + 547, + 570 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}(\\cdot)" + }, + { + "bbox": [ + 305, + 546, + 547, + 570 + ], + "type": "text", + "content": ", we iteratively denoise " + }, + { + "bbox": [ + 305, + 546, + 547, + 570 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_t" + }, + { + "bbox": [ + 305, + 546, + 547, + 570 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 305, + 546, + 547, + 570 + ], + "type": "inline_equation", + "content": "t = T" + }, + { + "bbox": [ + 305, + 546, + 547, + 570 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 305, + 546, + 547, + 570 + ], + "type": "inline_equation", + "content": "t = 1" + }, + { + "bbox": [ + 305, + 546, + 547, + 570 + ], + "type": "text", + "content": " with noise " + }, + { + "bbox": [ + 305, + 546, + 547, + 570 + ], + "type": "inline_equation", + "content": "\\mathbf{z}" + }, + { + "bbox": [ + 305, + 546, + 547, + 570 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 328, + 574, + 546, + 601 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 328, + 574, + 546, + 601 + ], + "spans": [ + { + "bbox": [ + 328, + 574, + 546, + 601 + ], + "type": "interline_equation", + "content": "\\mathbf {x} _ {t - 1} = \\frac {1}{\\sqrt {\\alpha_ {t}}} \\left(\\mathbf {x} _ {t} - \\frac {1 - \\alpha_ {t}}{\\sqrt {1 - \\bar {\\alpha} _ {t}}} \\boldsymbol {\\epsilon} _ {\\theta} (\\mathbf {x} _ {t}, t)\\right) + \\sigma_ {t} \\mathbf {z} \\quad (4)", + "image_path": "1ddc739f893d48d190afe0adb2381b760eab07878b58093a686a324f3aa0b46f.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 604, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 604, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 604, + 547, + 714 + ], + "type": "text", + "content": "The unconditional diffusion models can be extended to model conditional distributions " + }, + { + "bbox": [ + 304, + 604, + 547, + 714 + ], + "type": "inline_equation", + "content": "p_{\\theta}(\\mathbf{x}_0|c)" + }, + { + "bbox": [ + 304, + 604, + 547, + 714 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 604, + 547, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_0" + }, + { + "bbox": [ + 304, + 604, + 547, + 714 + ], + "type": "text", + "content": " is the image corresponding to the condition " + }, + { + "bbox": [ + 304, + 604, + 547, + 714 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 304, + 604, + 547, + 714 + ], + "type": "text", + "content": " such as class labels, segmentation masks, and text descriptions. The conditional diffusion model receives an additional input " + }, + { + "bbox": [ + 304, + 604, + 547, + 714 + ], + "type": "inline_equation", + "content": "\\tau(c)" + }, + { + "bbox": [ + 304, + 604, + 547, + 714 + ], + "type": "text", + "content": " and is trained by minimizing " + }, + { + "bbox": [ + 304, + 604, + 547, + 714 + ], + "type": "inline_equation", + "content": "\\| \\boldsymbol{\\epsilon} - \\boldsymbol{\\epsilon}_{\\theta}(\\mathbf{x}_t, t, \\tau(c)) \\|^2" + }, + { + "bbox": [ + 304, + 604, + 547, + 714 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 604, + 547, + 714 + ], + "type": "inline_equation", + "content": "\\tau(\\cdot)" + }, + { + "bbox": [ + 304, + 604, + 547, + 714 + ], + "type": "text", + "content": " is an encoder that projects the condition " + }, + { + "bbox": [ + 304, + 604, + 547, + 714 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 304, + 604, + 547, + 714 + ], + "type": "text", + "content": " to an embedding " + }, + { + "bbox": [ + 304, + 604, + 547, + 714 + ], + "type": "inline_equation", + "content": "\\tau(c)" + }, + { + "bbox": [ + 304, + 604, + 547, + 714 + ], + "type": "text", + "content": ". For brevity, we will use " + }, + { + "bbox": [ + 304, + 604, + 547, + 714 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 304, + 604, + 547, + 714 + ], + "type": "text", + "content": " to represent " + }, + { + "bbox": [ + 304, + 604, + 547, + 714 + ], + "type": "inline_equation", + "content": "\\tau(c)" + }, + { + "bbox": [ + 304, + 604, + 547, + 714 + ], + "type": "text", + "content": " in our subsequent discussions." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6082" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "code", + "bbox": [ + 51, + 99, + 272, + 262 + ], + "blocks": [ + { + "bbox": [ + 48, + 83, + 212, + 96 + ], + "lines": [ + { + "bbox": [ + 48, + 83, + 212, + 96 + ], + "spans": [ + { + "bbox": [ + 48, + 83, + 212, + 96 + ], + "type": "text", + "content": "Algorithm 1 Dynamic Diffuser Training" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "code_caption" + }, + { + "bbox": [ + 51, + 99, + 272, + 262 + ], + "lines": [ + { + "bbox": [ + 51, + 99, + 272, + 262 + ], + "spans": [ + { + "bbox": [ + 51, + 99, + 272, + 262 + ], + "type": "text", + "content": "1: repeat \n2: " + }, + { + "bbox": [ + 51, + 99, + 272, + 262 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_0, c_1, c_2, \\ldots, c_M \\sim q(\\mathbf{x}_0, c_1, c_2, \\ldots, c_M)" + }, + { + "bbox": [ + 51, + 99, + 272, + 262 + ], + "type": "text", + "content": " \n3: " + }, + { + "bbox": [ + 51, + 99, + 272, + 262 + ], + "type": "inline_equation", + "content": "t \\sim \\mathrm{Uniform}(\\{1, \\ldots, T\\})" + }, + { + "bbox": [ + 51, + 99, + 272, + 262 + ], + "type": "text", + "content": " \n4: " + }, + { + "bbox": [ + 51, + 99, + 272, + 262 + ], + "type": "inline_equation", + "content": "\\epsilon \\sim \\mathcal{N}(\\mathbf{0}, \\mathbf{I})" + }, + { + "bbox": [ + 51, + 99, + 272, + 262 + ], + "type": "text", + "content": " \n5: for " + }, + { + "bbox": [ + 51, + 99, + 272, + 262 + ], + "type": "inline_equation", + "content": "m = 1, \\ldots, M" + }, + { + "bbox": [ + 51, + 99, + 272, + 262 + ], + "type": "text", + "content": " do \n6: " + }, + { + "bbox": [ + 51, + 99, + 272, + 262 + ], + "type": "inline_equation", + "content": "\\epsilon_{pred,m,t} = \\epsilon_{\\theta_m}(\\sqrt{\\bar{\\alpha}_t}\\mathbf{x}_0 + \\sqrt{1 - \\bar{\\alpha}_t}\\boldsymbol{\\epsilon}, t, c_m)" + }, + { + "bbox": [ + 51, + 99, + 272, + 262 + ], + "type": "text", + "content": " \n7: " + }, + { + "bbox": [ + 51, + 99, + 272, + 262 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_{m,t} = \\mathbf{D}_{\\phi_m}(\\sqrt{\\bar{\\alpha}_t}\\mathbf{x}_0 + \\sqrt{1 - \\bar{\\alpha}_t}\\boldsymbol{\\epsilon}, t, c_m)" + }, + { + "bbox": [ + 51, + 99, + 272, + 262 + ], + "type": "text", + "content": " \n8: end for \n9: " + }, + { + "bbox": [ + 51, + 99, + 272, + 262 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{I}}_{m,t,p} = \\frac{\\exp(\\mathbf{I}_{m,t,p})}{\\sum_{j=1}^{M}\\exp(\\mathbf{I}_{j,t,p})}" + }, + { + "bbox": [ + 51, + 99, + 272, + 262 + ], + "type": "text", + "content": ", softmax at each pixel " + }, + { + "bbox": [ + 51, + 99, + 272, + 262 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 51, + 99, + 272, + 262 + ], + "type": "text", + "content": " \n10: " + }, + { + "bbox": [ + 51, + 99, + 272, + 262 + ], + "type": "inline_equation", + "content": "\\epsilon_{pred,t} = \\sum_{m=1}^{M}\\hat{\\mathbf{I}}_{m,t} \\odot \\epsilon_{pred,m,t}" + }, + { + "bbox": [ + 51, + 99, + 272, + 262 + ], + "type": "text", + "content": " \n11: Take gradient descent step on " + }, + { + "bbox": [ + 51, + 99, + 272, + 262 + ], + "type": "inline_equation", + "content": "\\nabla_{\\phi} \\| \\boldsymbol{\\epsilon} - \\boldsymbol{\\epsilon}_{pred,t} \\|^2" + }, + { + "bbox": [ + 51, + 99, + 272, + 262 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 51, + 99, + 272, + 262 + ], + "type": "inline_equation", + "content": "\\phi = \\{\\phi_m | m = 1, \\ldots, M\\}" + }, + { + "bbox": [ + 51, + 99, + 272, + 262 + ], + "type": "text", + "content": " \n12: until converged" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "code_body" + } + ], + "index": 1, + "sub_type": "algorithm" + }, + { + "type": "image", + "bbox": [ + 83, + 273, + 250, + 369 + ], + "blocks": [ + { + "bbox": [ + 83, + 273, + 250, + 369 + ], + "lines": [ + { + "bbox": [ + 83, + 273, + 250, + 369 + ], + "spans": [ + { + "bbox": [ + 83, + 273, + 250, + 369 + ], + "type": "image", + "image_path": "f73bd6938ea41ca9e485b986ac7e1896cd841e0c6dee6210685de9919a86cf41.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 376, + 288, + 433 + ], + "lines": [ + { + "bbox": [ + 46, + 376, + 288, + 433 + ], + "spans": [ + { + "bbox": [ + 46, + 376, + 288, + 433 + ], + "type": "text", + "content": "Figure 3. Dynamic Diffuser. It predicts the influence function that represents the desired level of contribution from each pre-trained diffusion model. The influence function varies at different timestep " + }, + { + "bbox": [ + 46, + 376, + 288, + 433 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 376, + 288, + 433 + ], + "type": "text", + "content": " and different spatial locations, and is dependent on the state of diffusion " + }, + { + "bbox": [ + 46, + 376, + 288, + 433 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_t" + }, + { + "bbox": [ + 46, + 376, + 288, + 433 + ], + "type": "text", + "content": " and the condition input " + }, + { + "bbox": [ + 46, + 376, + 288, + 433 + ], + "type": "inline_equation", + "content": "c_m" + }, + { + "bbox": [ + 46, + 376, + 288, + 433 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 463, + 243, + 475 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 463, + 243, + 475 + ], + "spans": [ + { + "bbox": [ + 47, + 463, + 243, + 475 + ], + "type": "text", + "content": "3.2. Multi-Modal Collaborative Synthesis" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 487, + 288, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 487, + 288, + 559 + ], + "spans": [ + { + "bbox": [ + 46, + 487, + 288, + 559 + ], + "type": "text", + "content": "In our Collaborative Diffusion, multiple uni-modal diffusion models collaborate at each step of the denoising process for multi-modal guided synthesis. The core of our framework is the dynamic diffuser, which determines the extent of contribution from each collaborator by predicting the spatial-temporal influence functions." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 564, + 288, + 625 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 564, + 288, + 625 + ], + "spans": [ + { + "bbox": [ + 47, + 564, + 288, + 625 + ], + "type": "text", + "content": "Problem Formulation. Given " + }, + { + "bbox": [ + 47, + 564, + 288, + 625 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 47, + 564, + 288, + 625 + ], + "type": "text", + "content": " pre-trained uni-modal conditional diffusion models " + }, + { + "bbox": [ + 47, + 564, + 288, + 625 + ], + "type": "inline_equation", + "content": "\\{\\epsilon_{\\theta_m}\\}" + }, + { + "bbox": [ + 47, + 564, + 288, + 625 + ], + "type": "text", + "content": " (each models the distribution " + }, + { + "bbox": [ + 47, + 564, + 288, + 625 + ], + "type": "inline_equation", + "content": "p(\\mathbf{x}_0|c_m)" + }, + { + "bbox": [ + 47, + 564, + 288, + 625 + ], + "type": "text", + "content": "), where the modality index " + }, + { + "bbox": [ + 47, + 564, + 288, + 625 + ], + "type": "inline_equation", + "content": "m = 1, \\ldots, M" + }, + { + "bbox": [ + 47, + 564, + 288, + 625 + ], + "type": "text", + "content": ", our objective is to sample from " + }, + { + "bbox": [ + 47, + 564, + 288, + 625 + ], + "type": "inline_equation", + "content": "p(\\mathbf{x}_0|\\mathbf{c})" + }, + { + "bbox": [ + 47, + 564, + 288, + 625 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 47, + 564, + 288, + 625 + ], + "type": "inline_equation", + "content": "\\mathbf{c} = \\{c_1, c_2, \\dots, c_M\\}" + }, + { + "bbox": [ + 47, + 564, + 288, + 625 + ], + "type": "text", + "content": ", without altering pre-trained models." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 629, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 629, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 629, + 289, + 715 + ], + "type": "text", + "content": "Dynamic Diffuser. In diffusion models, each step of the reverse process requires predicting the noise " + }, + { + "bbox": [ + 46, + 629, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 46, + 629, + 289, + 715 + ], + "type": "text", + "content": ". With multiple diffusion models collaborating, we need to carefully determine when, where, and how much each diffusion model contributes to the prediction " + }, + { + "bbox": [ + 46, + 629, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 46, + 629, + 289, + 715 + ], + "type": "text", + "content": ". At each diffusion timestep " + }, + { + "bbox": [ + 46, + 629, + 289, + 715 + ], + "type": "inline_equation", + "content": "t = T, \\dots, 1" + }, + { + "bbox": [ + 46, + 629, + 289, + 715 + ], + "type": "text", + "content": ", the influence " + }, + { + "bbox": [ + 46, + 629, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_{m,t}" + }, + { + "bbox": [ + 46, + 629, + 289, + 715 + ], + "type": "text", + "content": " from each pre-trained diffusion model " + }, + { + "bbox": [ + 46, + 629, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta_m}" + }, + { + "bbox": [ + 46, + 629, + 289, + 715 + ], + "type": "text", + "content": " is adaptively determined by a dynamic" + } + ] + } + ], + "index": 7 + }, + { + "type": "code", + "bbox": [ + 301, + 102, + 498, + 259 + ], + "blocks": [ + { + "bbox": [ + 296, + 83, + 448, + 96 + ], + "lines": [ + { + "bbox": [ + 296, + 83, + 448, + 96 + ], + "spans": [ + { + "bbox": [ + 296, + 83, + 448, + 96 + ], + "type": "text", + "content": "Algorithm 2 Collaborative Sampling" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "code_caption" + }, + { + "bbox": [ + 301, + 102, + 498, + 259 + ], + "lines": [ + { + "bbox": [ + 301, + 102, + 498, + 259 + ], + "spans": [ + { + "bbox": [ + 301, + 102, + 498, + 259 + ], + "type": "text", + "content": "1: " + }, + { + "bbox": [ + 301, + 102, + 498, + 259 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_T\\sim \\mathcal{N}(\\mathbf{0},\\mathbf{I})" + }, + { + "bbox": [ + 301, + 102, + 498, + 259 + ], + "type": "text", + "content": " \n2: for " + }, + { + "bbox": [ + 301, + 102, + 498, + 259 + ], + "type": "inline_equation", + "content": "t = T,\\dots ,1" + }, + { + "bbox": [ + 301, + 102, + 498, + 259 + ], + "type": "text", + "content": " do \n3: " + }, + { + "bbox": [ + 301, + 102, + 498, + 259 + ], + "type": "inline_equation", + "content": "\\mathbf{z}\\sim \\mathcal{N}(\\mathbf{0},\\mathbf{I})" + }, + { + "bbox": [ + 301, + 102, + 498, + 259 + ], + "type": "text", + "content": " if " + }, + { + "bbox": [ + 301, + 102, + 498, + 259 + ], + "type": "inline_equation", + "content": "t > 1" + }, + { + "bbox": [ + 301, + 102, + 498, + 259 + ], + "type": "text", + "content": " , else " + }, + { + "bbox": [ + 301, + 102, + 498, + 259 + ], + "type": "inline_equation", + "content": "\\mathbf{z} = \\mathbf{0}" + }, + { + "bbox": [ + 301, + 102, + 498, + 259 + ], + "type": "text", + "content": " \n4: for " + }, + { + "bbox": [ + 301, + 102, + 498, + 259 + ], + "type": "inline_equation", + "content": "m = 1,\\ldots ,M" + }, + { + "bbox": [ + 301, + 102, + 498, + 259 + ], + "type": "text", + "content": " do \n5: " + }, + { + "bbox": [ + 301, + 102, + 498, + 259 + ], + "type": "inline_equation", + "content": "\\epsilon_{pred,m,t} = \\epsilon_{\\theta_m}(\\mathbf{x}_t,t,c_m)" + }, + { + "bbox": [ + 301, + 102, + 498, + 259 + ], + "type": "text", + "content": " \n6: " + }, + { + "bbox": [ + 301, + 102, + 498, + 259 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_{m,t} = \\mathbf{D}_{\\phi_m}(\\mathbf{x}_t,t,c_m)" + }, + { + "bbox": [ + 301, + 102, + 498, + 259 + ], + "type": "text", + "content": " \n7: end for \n8: " + }, + { + "bbox": [ + 301, + 102, + 498, + 259 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{I}}_{m,t,p} = \\frac{\\exp(\\mathbf{I}_{m,t,p})}{\\sum_{j = 1}^{M}\\exp(\\mathbf{I}_{j,t,p})}" + }, + { + "bbox": [ + 301, + 102, + 498, + 259 + ], + "type": "text", + "content": " , softmax at each pixel " + }, + { + "bbox": [ + 301, + 102, + 498, + 259 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 301, + 102, + 498, + 259 + ], + "type": "text", + "content": " \n9: " + }, + { + "bbox": [ + 301, + 102, + 498, + 259 + ], + "type": "inline_equation", + "content": "\\epsilon_{pred,t} = \\sum_{m = 1}^{M}\\hat{\\mathbf{I}}_{m,t}\\odot \\epsilon_{pred,m,t}" + }, + { + "bbox": [ + 301, + 102, + 498, + 259 + ], + "type": "text", + "content": " \n10: " + }, + { + "bbox": [ + 301, + 102, + 498, + 259 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{t - 1} = \\frac{1}{\\sqrt{\\alpha_t}}\\left(\\mathbf{x}_t - \\frac{1 - \\alpha_t}{\\sqrt{1 - \\bar{\\alpha}_t}}\\pmb{\\epsilon}_{pred,t}\\right) + \\sigma_t\\mathbf{z}" + }, + { + "bbox": [ + 301, + 102, + 498, + 259 + ], + "type": "text", + "content": " \n11: end for \n12: return " + }, + { + "bbox": [ + 301, + 102, + 498, + 259 + ], + "type": "inline_equation", + "content": "{\\bf x}_0" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "code_body" + } + ], + "index": 9, + "sub_type": "algorithm" + }, + { + "bbox": [ + 306, + 275, + 364, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 275, + 364, + 289 + ], + "spans": [ + { + "bbox": [ + 306, + 275, + 364, + 289 + ], + "type": "text", + "content": "diffuser " + }, + { + "bbox": [ + 306, + 275, + 364, + 289 + ], + "type": "inline_equation", + "content": "\\mathbf{D}_{\\phi_m}" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 376, + 294, + 545, + 307 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 376, + 294, + 545, + 307 + ], + "spans": [ + { + "bbox": [ + 376, + 294, + 545, + 307 + ], + "type": "interline_equation", + "content": "\\mathbf {I} _ {m, t} = \\mathbf {D} _ {\\phi_ {m}} (\\mathbf {x} _ {t}, t, c _ {m}), \\tag {5}", + "image_path": "6b0afc388daf2f53572049392035d21006995a7cd754e3a7e1e313c0a5499bca.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 313, + 547, + 388 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 313, + 547, + 388 + ], + "spans": [ + { + "bbox": [ + 304, + 313, + 547, + 388 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 313, + 547, + 388 + ], + "type": "inline_equation", + "content": "m = 1,\\dots,M" + }, + { + "bbox": [ + 304, + 313, + 547, + 388 + ], + "type": "text", + "content": " is the index of the modalities, " + }, + { + "bbox": [ + 304, + 313, + 547, + 388 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_{m,t}\\in" + }, + { + "bbox": [ + 304, + 313, + 547, + 388 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^{h\\times w}" + }, + { + "bbox": [ + 304, + 313, + 547, + 388 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_t" + }, + { + "bbox": [ + 304, + 313, + 547, + 388 + ], + "type": "text", + "content": " is the noisy image at time " + }, + { + "bbox": [ + 304, + 313, + 547, + 388 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 313, + 547, + 388 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 313, + 547, + 388 + ], + "type": "inline_equation", + "content": "c_{m}" + }, + { + "bbox": [ + 304, + 313, + 547, + 388 + ], + "type": "text", + "content": " is the condition of the " + }, + { + "bbox": [ + 304, + 313, + 547, + 388 + ], + "type": "inline_equation", + "content": "m^{th}" + }, + { + "bbox": [ + 304, + 313, + 547, + 388 + ], + "type": "text", + "content": " modality, and " + }, + { + "bbox": [ + 304, + 313, + 547, + 388 + ], + "type": "inline_equation", + "content": "\\mathbf{D}_{\\phi_m}" + }, + { + "bbox": [ + 304, + 313, + 547, + 388 + ], + "type": "text", + "content": " is the dynamic diffuser implemented by a UNet [49]. To regularize the overall influence strength, we perform softmax across all modalities' " + }, + { + "bbox": [ + 304, + 313, + 547, + 388 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_{m,t}" + }, + { + "bbox": [ + 304, + 313, + 547, + 388 + ], + "type": "text", + "content": " at each pixel " + }, + { + "bbox": [ + 304, + 313, + 547, + 388 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 304, + 313, + 547, + 388 + ], + "type": "text", + "content": " to obtain the final influence function " + }, + { + "bbox": [ + 304, + 313, + 547, + 388 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{I}}_{m,t}" + }, + { + "bbox": [ + 304, + 313, + 547, + 388 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 368, + 393, + 545, + 423 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 368, + 393, + 545, + 423 + ], + "spans": [ + { + "bbox": [ + 368, + 393, + 545, + 423 + ], + "type": "interline_equation", + "content": "\\hat {\\mathbf {I}} _ {m, t, p} = \\frac {\\exp \\left(\\mathbf {I} _ {m , t , p}\\right)}{\\sum_ {j = 1} ^ {M} \\exp \\left(\\mathbf {I} _ {j , t , p}\\right)}. \\tag {6}", + "image_path": "086e610e15237a3a053ef9829746828d5837d3b36697d21b47651c74004a3c22.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 428, + 547, + 464 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 428, + 547, + 464 + ], + "spans": [ + { + "bbox": [ + 304, + 428, + 547, + 464 + ], + "type": "text", + "content": "Multi-Modal Collaboration. We use the " + }, + { + "bbox": [ + 304, + 428, + 547, + 464 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 304, + 428, + 547, + 464 + ], + "type": "text", + "content": " learned influence functions " + }, + { + "bbox": [ + 304, + 428, + 547, + 464 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{I}}_{m,t}" + }, + { + "bbox": [ + 304, + 428, + 547, + 464 + ], + "type": "text", + "content": " to control the contribution from each pre-trained diffusion model at each denoising step:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 351, + 470, + 545, + 503 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 351, + 470, + 545, + 503 + ], + "spans": [ + { + "bbox": [ + 351, + 470, + 545, + 503 + ], + "type": "interline_equation", + "content": "\\boldsymbol {\\epsilon} _ {p r e d, t} = \\sum_ {m = 1} ^ {M} \\hat {\\mathbf {I}} _ {m, t} \\odot \\boldsymbol {\\epsilon} _ {\\theta_ {m}} \\left(x _ {t}, t, c _ {m}\\right) \\tag {7}", + "image_path": "ad07ba0a06e866862352661d6052b924059bda9c5f6214b3443f77ed39334caa.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 510, + 547, + 570 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 510, + 547, + 570 + ], + "spans": [ + { + "bbox": [ + 304, + 510, + 547, + 570 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 510, + 547, + 570 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta_m}" + }, + { + "bbox": [ + 304, + 510, + 547, + 570 + ], + "type": "text", + "content": " is the " + }, + { + "bbox": [ + 304, + 510, + 547, + 570 + ], + "type": "inline_equation", + "content": "m^{th}" + }, + { + "bbox": [ + 304, + 510, + 547, + 570 + ], + "type": "text", + "content": " collaborator, and " + }, + { + "bbox": [ + 304, + 510, + 547, + 570 + ], + "type": "inline_equation", + "content": "\\odot" + }, + { + "bbox": [ + 304, + 510, + 547, + 570 + ], + "type": "text", + "content": " denotes pixel-wise multiplication. The complete training procedure of dynamic diffusers and the image sampling strategy in our Collaborative Diffusion framework are detailed in Algorithm 1 and Algorithm 2, respectively." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 570, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 547, + 713 + ], + "type": "text", + "content": "Relation with Composable Diffusion. Composable Diffusion [36] combines two instances of the same text-to-image diffusion model to achieve compositional visual generation via intermediate results addition. Our framework is related to Composable Diffusion in terms of composing diffusion models for image synthesis, but is substantially different in terms of the task nature and methodology. Composable Diffusion aims to decompose the text condition into elementary segments to factorize the conditional synthesis problem, while we aim to integrate uni-modal collaborators to achieve multi-modal controls. Furthermore, different from Composable Diffusion, which can only compose instances of the" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6083" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 83, + 189, + 96 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 83, + 189, + 96 + ], + "spans": [ + { + "bbox": [ + 48, + 83, + 189, + 96 + ], + "type": "text", + "content": "Algorithm 3 Collaborative Editing" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 102, + 83, + 112 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 102, + 83, + 112 + ], + "spans": [ + { + "bbox": [ + 49, + 102, + 83, + 112 + ], + "type": "text", + "content": "Require:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 63, + 114, + 294, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 114, + 294, + 156 + ], + "spans": [ + { + "bbox": [ + 63, + 114, + 294, + 156 + ], + "type": "text", + "content": "input image " + }, + { + "bbox": [ + 63, + 114, + 294, + 156 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{input}" + }, + { + "bbox": [ + 63, + 114, + 294, + 156 + ], + "type": "text", + "content": ", target conditions " + }, + { + "bbox": [ + 63, + 114, + 294, + 156 + ], + "type": "inline_equation", + "content": "c_{m,target}" + }, + { + "bbox": [ + 63, + 114, + 294, + 156 + ], + "type": "text", + "content": ", diffusion models " + }, + { + "bbox": [ + 63, + 114, + 294, + 156 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta_m}" + }, + { + "bbox": [ + 63, + 114, + 294, + 156 + ], + "type": "text", + "content": ", dynamic diffusers " + }, + { + "bbox": [ + 63, + 114, + 294, + 156 + ], + "type": "inline_equation", + "content": "\\mathbf{D}_{\\phi_m}" + }, + { + "bbox": [ + 63, + 114, + 294, + 156 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 63, + 114, + 294, + 156 + ], + "type": "inline_equation", + "content": "(m = 1, \\dots, M)" + }, + { + "bbox": [ + 63, + 114, + 294, + 156 + ], + "type": "text", + "content": ", interpolation scale " + }, + { + "bbox": [ + 63, + 114, + 294, + 156 + ], + "type": "inline_equation", + "content": "\\alpha" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 159, + 294, + 237 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 53, + 159, + 294, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 159, + 294, + 171 + ], + "spans": [ + { + "bbox": [ + 53, + 159, + 294, + 171 + ], + "type": "text", + "content": "1: for " + }, + { + "bbox": [ + 53, + 159, + 294, + 171 + ], + "type": "inline_equation", + "content": "m = 1,\\ldots ,M" + }, + { + "bbox": [ + 53, + 159, + 294, + 171 + ], + "type": "text", + "content": " do " + }, + { + "bbox": [ + 53, + 159, + 294, + 171 + ], + "type": "inline_equation", + "content": "\\triangleright" + }, + { + "bbox": [ + 53, + 159, + 294, + 171 + ], + "type": "text", + "content": " Uni-Modal Editing" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 172, + 127, + 182 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 172, + 127, + 182 + ], + "spans": [ + { + "bbox": [ + 53, + 172, + 127, + 182 + ], + "type": "text", + "content": "2: " + }, + { + "bbox": [ + 53, + 172, + 127, + 182 + ], + "type": "inline_equation", + "content": "c_{m} = c_{m, \\text{target}}" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 182, + 181, + 192 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 182, + 181, + 192 + ], + "spans": [ + { + "bbox": [ + 53, + 182, + 181, + 192 + ], + "type": "text", + "content": "3: " + }, + { + "bbox": [ + 53, + 182, + 181, + 192 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_t = \\sqrt{\\bar{\\alpha}_t}\\mathbf{x}_{input} + \\sqrt{1 - \\bar{\\alpha}_t}\\pmb{\\epsilon}" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 192, + 246, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 192, + 246, + 204 + ], + "spans": [ + { + "bbox": [ + 53, + 192, + 246, + 204 + ], + "type": "text", + "content": "4: " + }, + { + "bbox": [ + 53, + 192, + 246, + 204 + ], + "type": "inline_equation", + "content": "c_{m,opt} = \\mathrm{argmin}_{c_m}\\mathbb{E}_{\\epsilon ,t}\\left\\| \\boldsymbol {\\epsilon} - \\boldsymbol{\\epsilon}_{\\theta_m}(\\mathbf{x}_t,t,c_m)\\right\\| ^2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 204, + 258, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 204, + 258, + 217 + ], + "spans": [ + { + "bbox": [ + 53, + 204, + 258, + 217 + ], + "type": "text", + "content": "5: " + }, + { + "bbox": [ + 53, + 204, + 258, + 217 + ], + "type": "inline_equation", + "content": "\\theta_{m,opt} = \\operatorname*{argmin}_{\\theta_m}\\mathbb{E}_{\\epsilon ,t}\\| \\pmb {\\epsilon} - \\pmb{\\epsilon}_{\\theta_m}(\\mathbf{x}_t,t,c_{m,opt})\\|^2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 217, + 222, + 227 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 217, + 222, + 227 + ], + "spans": [ + { + "bbox": [ + 53, + 217, + 222, + 227 + ], + "type": "text", + "content": "6: " + }, + { + "bbox": [ + 53, + 217, + 222, + 227 + ], + "type": "inline_equation", + "content": "c_{m,int} = \\alpha \\cdot c_{m,target} + (1 - \\alpha) \\cdot c_{m,opt}" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 228, + 93, + 237 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 228, + 93, + 237 + ], + "spans": [ + { + "bbox": [ + 53, + 228, + 93, + 237 + ], + "type": "text", + "content": "7: end for" + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 50, + 240, + 294, + 396 + ], + "type": "list", + "angle": 0, + "index": 23, + "blocks": [ + { + "bbox": [ + 53, + 240, + 294, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 240, + 294, + 251 + ], + "spans": [ + { + "bbox": [ + 53, + 240, + 294, + 251 + ], + "type": "text", + "content": "8: " + }, + { + "bbox": [ + 53, + 240, + 294, + 251 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_T\\sim \\mathcal{N}(\\mathbf{0},\\mathbf{I})" + }, + { + "bbox": [ + 53, + 240, + 294, + 251 + ], + "type": "text", + "content": " ▷Collaborate the Uni-Modal Edits" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 53, + 252, + 138, + 262 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 252, + 138, + 262 + ], + "spans": [ + { + "bbox": [ + 53, + 252, + 138, + 262 + ], + "type": "text", + "content": "9: for " + }, + { + "bbox": [ + 53, + 252, + 138, + 262 + ], + "type": "inline_equation", + "content": "t = T, \\dots, 1" + }, + { + "bbox": [ + 53, + 252, + 138, + 262 + ], + "type": "text", + "content": " do" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 50, + 262, + 189, + 273 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 50, + 262, + 189, + 273 + ], + "spans": [ + { + "bbox": [ + 50, + 262, + 189, + 273 + ], + "type": "text", + "content": "10: " + }, + { + "bbox": [ + 50, + 262, + 189, + 273 + ], + "type": "inline_equation", + "content": "\\mathbf{z} \\sim \\mathcal{N}(\\mathbf{0}, \\mathbf{I})" + }, + { + "bbox": [ + 50, + 262, + 189, + 273 + ], + "type": "text", + "content": " if " + }, + { + "bbox": [ + 50, + 262, + 189, + 273 + ], + "type": "inline_equation", + "content": "t > 1" + }, + { + "bbox": [ + 50, + 262, + 189, + 273 + ], + "type": "text", + "content": ", else " + }, + { + "bbox": [ + 50, + 262, + 189, + 273 + ], + "type": "inline_equation", + "content": "\\mathbf{z} = \\mathbf{0}" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 50, + 274, + 146, + 283 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 50, + 274, + 146, + 283 + ], + "spans": [ + { + "bbox": [ + 50, + 274, + 146, + 283 + ], + "type": "text", + "content": "11: for " + }, + { + "bbox": [ + 50, + 274, + 146, + 283 + ], + "type": "inline_equation", + "content": "m = 0,\\dots,M" + }, + { + "bbox": [ + 50, + 274, + 146, + 283 + ], + "type": "text", + "content": " do" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 50, + 285, + 208, + 297 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 50, + 285, + 208, + 297 + ], + "spans": [ + { + "bbox": [ + 50, + 285, + 208, + 297 + ], + "type": "text", + "content": "12: " + }, + { + "bbox": [ + 50, + 285, + 208, + 297 + ], + "type": "inline_equation", + "content": "\\pmb{\\epsilon}_{pred,m,t} = \\pmb{\\epsilon}_{\\theta_{m,opt}}(\\mathbf{x}_t,t,c_{m,int})" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 50, + 298, + 179, + 310 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 50, + 298, + 179, + 310 + ], + "spans": [ + { + "bbox": [ + 50, + 298, + 179, + 310 + ], + "type": "text", + "content": "13: " + }, + { + "bbox": [ + 50, + 298, + 179, + 310 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_{m,t} = \\mathbf{D}_{\\phi_m}(\\mathbf{x}_t,t,c_{m,int})" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 50, + 312, + 97, + 322 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 50, + 312, + 97, + 322 + ], + "spans": [ + { + "bbox": [ + 50, + 312, + 97, + 322 + ], + "type": "text", + "content": "14: end for" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 50, + 323, + 249, + 342 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 50, + 323, + 249, + 342 + ], + "spans": [ + { + "bbox": [ + 50, + 323, + 249, + 342 + ], + "type": "text", + "content": "15: " + }, + { + "bbox": [ + 50, + 323, + 249, + 342 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{I}}_{m,t,p} = \\frac{\\exp(\\mathbf{I}_{m,t,p})}{\\sum_{j = 1}^{M}\\exp(\\mathbf{I}_{j,t,p})}" + }, + { + "bbox": [ + 50, + 323, + 249, + 342 + ], + "type": "text", + "content": ", softmax at each pixel " + }, + { + "bbox": [ + 50, + 323, + 249, + 342 + ], + "type": "inline_equation", + "content": "p" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 50, + 344, + 195, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 50, + 344, + 195, + 357 + ], + "spans": [ + { + "bbox": [ + 50, + 344, + 195, + 357 + ], + "type": "text", + "content": "16: " + }, + { + "bbox": [ + 50, + 344, + 195, + 357 + ], + "type": "inline_equation", + "content": "\\pmb{\\epsilon}_{pred,t} = \\sum_{m=1}^{M} \\hat{\\mathbf{I}}_{m,t} \\odot \\pmb{\\epsilon}_{pred,m,t}" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 50, + 360, + 225, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 50, + 360, + 225, + 376 + ], + "spans": [ + { + "bbox": [ + 50, + 360, + 225, + 376 + ], + "type": "text", + "content": "17: " + }, + { + "bbox": [ + 50, + 360, + 225, + 376 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{t - 1} = \\frac{1}{\\sqrt{\\alpha_t}}\\left(\\mathbf{x}_t - \\frac{1 - \\alpha_t}{\\sqrt{1 - \\bar{\\alpha}_t}}\\pmb {\\epsilon}_{pred,t}\\right) + \\sigma_t\\mathbf{z}" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 50, + 376, + 93, + 385 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 50, + 376, + 93, + 385 + ], + "spans": [ + { + "bbox": [ + 50, + 376, + 93, + 385 + ], + "type": "text", + "content": "18: end for" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 50, + 387, + 102, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 50, + 387, + 102, + 396 + ], + "spans": [ + { + "bbox": [ + 50, + 387, + 102, + 396 + ], + "type": "text", + "content": "19: return " + }, + { + "bbox": [ + 50, + 387, + 102, + 396 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_0" + } + ] + } + ], + "index": 22 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 46, + 417, + 288, + 477 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 417, + 288, + 477 + ], + "spans": [ + { + "bbox": [ + 46, + 417, + 288, + 477 + ], + "type": "text", + "content": "same text-based synthesis model (i.e., the weights of the diffusion models being composed are the same), our framework possesses the flexibility in integrating models with different weights, architectures, and modalities through the learned dynamic diffusers." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 47, + 489, + 234, + 502 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 489, + 234, + 502 + ], + "spans": [ + { + "bbox": [ + 47, + 489, + 234, + 502 + ], + "type": "text", + "content": "3.3. Multi-Modal Collaborative Editing" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 46, + 509, + 287, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 509, + 287, + 581 + ], + "spans": [ + { + "bbox": [ + 46, + 509, + 287, + 581 + ], + "type": "text", + "content": "In addition to face synthesis, our framework is also capable of combining multiple facial manipulations, each of which is guided by a different modality, into a collaborative edit. Our collaborative framework in theory can integrate any existing uni-modal diffusion-based editing approach for collaborative editing." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 46, + 582, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 582, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 582, + 288, + 714 + ], + "type": "text", + "content": "In this work, we demonstrate such possibility by extending Imagic [29] to the multi-modal paradigm. We first follow Imagic to fine-tune the embeddings and models to better capture the input face identity during editing. The trained dynamic diffusers discussed in Section 3.2 are then used to combine the fine-tuned models. Algorithm 3 displays the complete procedure of collaborative editing. Note that finetuning the pre-trained models is for identity preservation proposed in Imagic, which is independent to our framework. The extension of our framework to editing requires no further training of dynamic diffusers." + } + ] + } + ], + "index": 27 + }, + { + "type": "image", + "bbox": [ + 308, + 79, + 351, + 121 + ], + "blocks": [ + { + "bbox": [ + 338, + 71, + 395, + 78 + ], + "lines": [ + { + "bbox": [ + 338, + 71, + 395, + 78 + ], + "spans": [ + { + "bbox": [ + 338, + 71, + 395, + 78 + ], + "type": "text", + "content": "Multi-Modal Conditions" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 308, + 79, + 351, + 121 + ], + "lines": [ + { + "bbox": [ + 308, + 79, + 351, + 121 + ], + "spans": [ + { + "bbox": [ + 308, + 79, + 351, + 121 + ], + "type": "image", + "image_path": "9afea71ee8fa0a294a9108b1256c1416ff8d90f970bc96f75d04ce55f06121c3.jpg" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 354, + 87, + 407, + 114 + ], + "lines": [ + { + "bbox": [ + 354, + 87, + 407, + 114 + ], + "spans": [ + { + "bbox": [ + 354, + 87, + 407, + 114 + ], + "type": "text", + "content": "He is a young adult. This man has a rough growth of stubble." + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_footnote" + } + ], + "index": 29 + }, + { + "type": "image", + "bbox": [ + 308, + 167, + 351, + 209 + ], + "blocks": [ + { + "bbox": [ + 308, + 167, + 351, + 209 + ], + "lines": [ + { + "bbox": [ + 308, + 167, + 351, + 209 + ], + "spans": [ + { + "bbox": [ + 308, + 167, + 351, + 209 + ], + "type": "image", + "image_path": "e3fc0353b8f8566b51e014a0a17c5231e48c82e30bb0a033d501cc873fe91127.jpg" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 354, + 178, + 403, + 199 + ], + "lines": [ + { + "bbox": [ + 354, + 178, + 403, + 199 + ], + "spans": [ + { + "bbox": [ + 354, + 178, + 403, + 199 + ], + "type": "text", + "content": "There is no beard. This woman looks extremely young." + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 306, + 219, + 546, + 251 + ], + "lines": [ + { + "bbox": [ + 306, + 219, + 546, + 251 + ], + "spans": [ + { + "bbox": [ + 306, + 219, + 546, + 251 + ], + "type": "text", + "content": "Figure 4. Qualitative Comparison of Face Generation. In the second example, TediGAN and Composable fails to follow mask, while ours generates results highly consistent to both conditions." + } + ] + } + ], + "index": 46, + "angle": 0, + "type": "image_caption" + } + ], + "index": 33 + }, + { + "type": "image", + "bbox": [ + 411, + 79, + 454, + 121 + ], + "blocks": [ + { + "bbox": [ + 455, + 71, + 498, + 78 + ], + "lines": [ + { + "bbox": [ + 455, + 71, + 498, + 78 + ], + "spans": [ + { + "bbox": [ + 455, + 71, + 498, + 78 + ], + "type": "text", + "content": "Generated Images" + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 411, + 79, + 454, + 121 + ], + "lines": [ + { + "bbox": [ + 411, + 79, + 454, + 121 + ], + "spans": [ + { + "bbox": [ + 411, + 79, + 454, + 121 + ], + "type": "image", + "image_path": "ff5c22629236d074c61a2c66f9f4c231185a5cb50d3ba7088c0e44b72b0fbb7d.jpg" + } + ] + } + ], + "index": 36, + "angle": 0, + "type": "image_body" + } + ], + "index": 36 + }, + { + "type": "image", + "bbox": [ + 456, + 79, + 498, + 121 + ], + "blocks": [ + { + "bbox": [ + 456, + 79, + 498, + 121 + ], + "lines": [ + { + "bbox": [ + 456, + 79, + 498, + 121 + ], + "spans": [ + { + "bbox": [ + 456, + 79, + 498, + 121 + ], + "type": "image", + "image_path": "7a8b8046e2bd97ef9d3789d697463be08dc7d8d48d5b8265b3370f2e3bd3bb59.jpg" + } + ] + } + ], + "index": 37, + "angle": 0, + "type": "image_body" + } + ], + "index": 37 + }, + { + "type": "image", + "bbox": [ + 411, + 122, + 454, + 166 + ], + "blocks": [ + { + "bbox": [ + 355, + 131, + 405, + 159 + ], + "lines": [ + { + "bbox": [ + 355, + 131, + 405, + 159 + ], + "spans": [ + { + "bbox": [ + 355, + 131, + 405, + 159 + ], + "type": "text", + "content": "His face is covered with short beard. \nThis man is in the thirties." + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 411, + 122, + 454, + 166 + ], + "lines": [ + { + "bbox": [ + 411, + 122, + 454, + 166 + ], + "spans": [ + { + "bbox": [ + 411, + 122, + 454, + 166 + ], + "type": "image", + "image_path": "d2afee1f2fad5f00c14f368195b69901dedc916a6cfed21b2dfaec8b755c0cef.jpg" + } + ] + } + ], + "index": 38, + "angle": 0, + "type": "image_body" + } + ], + "index": 38 + }, + { + "type": "image", + "bbox": [ + 500, + 79, + 542, + 121 + ], + "blocks": [ + { + "bbox": [ + 500, + 79, + 542, + 121 + ], + "lines": [ + { + "bbox": [ + 500, + 79, + 542, + 121 + ], + "spans": [ + { + "bbox": [ + 500, + 79, + 542, + 121 + ], + "type": "image", + "image_path": "f4e49f4c2256072aa21fbebb01b7f6367d2e8fba4ac00d4904ea181897c1881c.jpg" + } + ] + } + ], + "index": 39, + "angle": 0, + "type": "image_body" + } + ], + "index": 39 + }, + { + "type": "image", + "bbox": [ + 456, + 122, + 498, + 165 + ], + "blocks": [ + { + "bbox": [ + 456, + 122, + 498, + 165 + ], + "lines": [ + { + "bbox": [ + 456, + 122, + 498, + 165 + ], + "spans": [ + { + "bbox": [ + 456, + 122, + 498, + 165 + ], + "type": "image", + "image_path": "237f4c875b4324a67acb1696a3f2b9296622fc37220d16b2426bea81ace0dfe5.jpg" + } + ] + } + ], + "index": 40, + "angle": 0, + "type": "image_body" + } + ], + "index": 40 + }, + { + "type": "image", + "bbox": [ + 500, + 123, + 541, + 166 + ], + "blocks": [ + { + "bbox": [ + 500, + 123, + 541, + 166 + ], + "lines": [ + { + "bbox": [ + 500, + 123, + 541, + 166 + ], + "spans": [ + { + "bbox": [ + 500, + 123, + 541, + 166 + ], + "type": "image", + "image_path": "244c8f0d25ffe5a9aef5bf2ce195a999ebf28fb483a0b9d07a0e04c3a03bfac7.jpg" + } + ] + } + ], + "index": 41, + "angle": 0, + "type": "image_body" + } + ], + "index": 41 + }, + { + "type": "image", + "bbox": [ + 411, + 167, + 454, + 209 + ], + "blocks": [ + { + "bbox": [ + 411, + 167, + 454, + 209 + ], + "lines": [ + { + "bbox": [ + 411, + 167, + 454, + 209 + ], + "spans": [ + { + "bbox": [ + 411, + 167, + 454, + 209 + ], + "type": "image", + "image_path": "ca7caebd14fc05883ffa33da300861836f97f6a52567ead526753691c97f72a9.jpg" + } + ] + } + ], + "index": 42, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 422, + 210, + 493, + 217 + ], + "lines": [ + { + "bbox": [ + 422, + 210, + 493, + 217 + ], + "spans": [ + { + "bbox": [ + 422, + 210, + 493, + 217 + ], + "type": "text", + "content": "TediGAN Composable" + } + ] + } + ], + "index": 43, + "angle": 0, + "type": "image_footnote" + } + ], + "index": 42 + }, + { + "type": "image", + "bbox": [ + 500, + 167, + 542, + 209 + ], + "blocks": [ + { + "bbox": [ + 500, + 167, + 542, + 209 + ], + "lines": [ + { + "bbox": [ + 500, + 167, + 542, + 209 + ], + "spans": [ + { + "bbox": [ + 500, + 167, + 542, + 209 + ], + "type": "image", + "image_path": "b46c9420b617f1d743b96885c32a4fd385a3bf2cc7829a5ca925862a8f5c04f5.jpg" + } + ] + } + ], + "index": 44, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 514, + 210, + 527, + 217 + ], + "lines": [ + { + "bbox": [ + 514, + 210, + 527, + 217 + ], + "spans": [ + { + "bbox": [ + 514, + 210, + 527, + 217 + ], + "type": "text", + "content": "Ours" + } + ] + } + ], + "index": 45, + "angle": 0, + "type": "image_footnote" + } + ], + "index": 44 + }, + { + "type": "image", + "bbox": [ + 308, + 267, + 345, + 303 + ], + "blocks": [ + { + "bbox": [ + 312, + 259, + 338, + 266 + ], + "lines": [ + { + "bbox": [ + 312, + 259, + 338, + 266 + ], + "spans": [ + { + "bbox": [ + 312, + 259, + 338, + 266 + ], + "type": "text", + "content": "Input Image" + } + ] + } + ], + "index": 47, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 308, + 267, + 345, + 303 + ], + "lines": [ + { + "bbox": [ + 308, + 267, + 345, + 303 + ], + "spans": [ + { + "bbox": [ + 308, + 267, + 345, + 303 + ], + "type": "image", + "image_path": "7174d87008736e4024c323dc9a414840ad5a48a8aa63f2401450cb976bb55686.jpg" + } + ] + } + ], + "index": 48, + "angle": 0, + "type": "image_body" + } + ], + "index": 48 + }, + { + "type": "image", + "bbox": [ + 347, + 267, + 383, + 303 + ], + "blocks": [ + { + "bbox": [ + 358, + 259, + 421, + 266 + ], + "lines": [ + { + "bbox": [ + 358, + 259, + 421, + 266 + ], + "spans": [ + { + "bbox": [ + 358, + 259, + 421, + 266 + ], + "type": "text", + "content": "Multi-Modal Target Conditions" + } + ] + } + ], + "index": 49, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 347, + 267, + 383, + 303 + ], + "lines": [ + { + "bbox": [ + 347, + 267, + 383, + 303 + ], + "spans": [ + { + "bbox": [ + 347, + 267, + 383, + 303 + ], + "type": "image", + "image_path": "e8b4d0e2ed709a0cfb1a85472acf26cba3d8b7fc7133dd44fb4a22889b715598.jpg" + } + ] + } + ], + "index": 50, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 386, + 274, + 429, + 297 + ], + "lines": [ + { + "bbox": [ + 386, + 274, + 429, + 297 + ], + "spans": [ + { + "bbox": [ + 386, + 274, + 429, + 297 + ], + "type": "text", + "content": "He is a teen. The face is covered with short pointed beard." + } + ] + } + ], + "index": 51, + "angle": 0, + "type": "image_footnote" + } + ], + "index": 50 + }, + { + "type": "image", + "bbox": [ + 308, + 304, + 344, + 339 + ], + "blocks": [ + { + "bbox": [ + 308, + 304, + 344, + 339 + ], + "lines": [ + { + "bbox": [ + 308, + 304, + 344, + 339 + ], + "spans": [ + { + "bbox": [ + 308, + 304, + 344, + 339 + ], + "type": "image", + "image_path": "3520e2288f41faa1c38bc6c6e2d5a215550965b1cb2f5c5221d6aa674f478faf.jpg" + } + ] + } + ], + "index": 52, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 306, + 347, + 546, + 392 + ], + "lines": [ + { + "bbox": [ + 306, + 347, + 546, + 392 + ], + "spans": [ + { + "bbox": [ + 306, + 347, + 546, + 392 + ], + "type": "text", + "content": "Figure 5. Qualitative Comparison of Face Editing. While TediGAN's hair shape is inconsistent with mask, and Composable fails to generate beard according to text, our framework produces results highly consistent with both conditions while maintaining identity." + } + ] + } + ], + "index": 65, + "angle": 0, + "type": "image_caption" + } + ], + "index": 52 + }, + { + "type": "image", + "bbox": [ + 347, + 304, + 383, + 339 + ], + "blocks": [ + { + "bbox": [ + 347, + 304, + 383, + 339 + ], + "lines": [ + { + "bbox": [ + 347, + 304, + 383, + 339 + ], + "spans": [ + { + "bbox": [ + 347, + 304, + 383, + 339 + ], + "type": "image", + "image_path": "b9bcb15e50f9ce26476b5d0cdbe998e1d22705c07102469e0a265325bb676f28.jpg" + } + ] + } + ], + "index": 53, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 386, + 311, + 427, + 334 + ], + "lines": [ + { + "bbox": [ + 386, + 311, + 427, + 334 + ], + "spans": [ + { + "bbox": [ + 386, + 311, + 427, + 334 + ], + "type": "text", + "content": "This man has \nheard of medium \nlength. He is in his \nthirties." + } + ] + } + ], + "index": 54, + "angle": 0, + "type": "image_footnote" + } + ], + "index": 53 + }, + { + "type": "image", + "bbox": [ + 432, + 267, + 470, + 303 + ], + "blocks": [ + { + "bbox": [ + 432, + 267, + 470, + 303 + ], + "lines": [ + { + "bbox": [ + 432, + 267, + 470, + 303 + ], + "spans": [ + { + "bbox": [ + 432, + 267, + 470, + 303 + ], + "type": "image", + "image_path": "1b5fbed737513bb8fb7e433512dd81419bad04689ea696f726badb309451bc68.jpg" + } + ] + } + ], + "index": 56, + "angle": 0, + "type": "image_body" + } + ], + "index": 56 + }, + { + "type": "image", + "bbox": [ + 470, + 267, + 504, + 303 + ], + "blocks": [ + { + "bbox": [ + 474, + 259, + 503, + 266 + ], + "lines": [ + { + "bbox": [ + 474, + 259, + 503, + 266 + ], + "spans": [ + { + "bbox": [ + 474, + 259, + 503, + 266 + ], + "type": "text", + "content": "Edited Images" + } + ] + } + ], + "index": 55, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 470, + 267, + 504, + 303 + ], + "lines": [ + { + "bbox": [ + 470, + 267, + 504, + 303 + ], + "spans": [ + { + "bbox": [ + 470, + 267, + 504, + 303 + ], + "type": "image", + "image_path": "e4f47afe8cd0521502406ea665e0807f2b350d035dc45346fb3e3bbb226cb3a5.jpg" + } + ] + } + ], + "index": 57, + "angle": 0, + "type": "image_body" + } + ], + "index": 57 + }, + { + "type": "image", + "bbox": [ + 507, + 267, + 542, + 303 + ], + "blocks": [ + { + "bbox": [ + 507, + 267, + 542, + 303 + ], + "lines": [ + { + "bbox": [ + 507, + 267, + 542, + 303 + ], + "spans": [ + { + "bbox": [ + 507, + 267, + 542, + 303 + ], + "type": "image", + "image_path": "6d67adf7d26a0464be27cd72f9160cafa604671f957b1a2eb987a89652a6fab6.jpg" + } + ] + } + ], + "index": 58, + "angle": 0, + "type": "image_body" + } + ], + "index": 58 + }, + { + "type": "image", + "bbox": [ + 433, + 304, + 470, + 339 + ], + "blocks": [ + { + "bbox": [ + 433, + 304, + 470, + 339 + ], + "lines": [ + { + "bbox": [ + 433, + 304, + 470, + 339 + ], + "spans": [ + { + "bbox": [ + 433, + 304, + 470, + 339 + ], + "type": "image", + "image_path": "52d0b0cdf1f55777c71752121eec113d75a6668d743f0fd746646696290b6fa7.jpg" + } + ] + } + ], + "index": 59, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 442, + 340, + 461, + 346 + ], + "lines": [ + { + "bbox": [ + 442, + 340, + 461, + 346 + ], + "spans": [ + { + "bbox": [ + 442, + 340, + 461, + 346 + ], + "type": "text", + "content": "TediGAN" + } + ] + } + ], + "index": 60, + "angle": 0, + "type": "image_caption" + } + ], + "index": 59 + }, + { + "type": "image", + "bbox": [ + 471, + 304, + 506, + 339 + ], + "blocks": [ + { + "bbox": [ + 471, + 304, + 506, + 339 + ], + "lines": [ + { + "bbox": [ + 471, + 304, + 506, + 339 + ], + "spans": [ + { + "bbox": [ + 471, + 304, + 506, + 339 + ], + "type": "image", + "image_path": "59bec688eed7b918af67b7643a945ec0611547dda362e23b37878875dd8f68f8.jpg" + } + ] + } + ], + "index": 61, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 475, + 340, + 500, + 346 + ], + "lines": [ + { + "bbox": [ + 475, + 340, + 500, + 346 + ], + "spans": [ + { + "bbox": [ + 475, + 340, + 500, + 346 + ], + "type": "text", + "content": "Composable" + } + ] + } + ], + "index": 62, + "angle": 0, + "type": "image_caption" + } + ], + "index": 61 + }, + { + "type": "image", + "bbox": [ + 507, + 304, + 542, + 339 + ], + "blocks": [ + { + "bbox": [ + 507, + 304, + 542, + 339 + ], + "lines": [ + { + "bbox": [ + 507, + 304, + 542, + 339 + ], + "spans": [ + { + "bbox": [ + 507, + 304, + 542, + 339 + ], + "type": "image", + "image_path": "860490e9e49f19e2c9acbeb7b30ef2470cca211c67d3c1870bdb1c4734d8df04.jpg" + } + ] + } + ], + "index": 63, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 518, + 340, + 531, + 346 + ], + "lines": [ + { + "bbox": [ + 518, + 340, + 531, + 346 + ], + "spans": [ + { + "bbox": [ + 518, + 340, + 531, + 346 + ], + "type": "text", + "content": "Ours" + } + ] + } + ], + "index": 64, + "angle": 0, + "type": "image_footnote" + } + ], + "index": 63 + }, + { + "bbox": [ + 306, + 405, + 386, + 419 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 405, + 386, + 419 + ], + "spans": [ + { + "bbox": [ + 306, + 405, + 386, + 419 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 66 + }, + { + "bbox": [ + 306, + 426, + 422, + 438 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 426, + 422, + 438 + ], + "spans": [ + { + "bbox": [ + 306, + 426, + 422, + 438 + ], + "type": "text", + "content": "4.1. Experimental Setup" + } + ] + } + ], + "index": 67 + }, + { + "bbox": [ + 304, + 445, + 547, + 576 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 445, + 547, + 576 + ], + "spans": [ + { + "bbox": [ + 304, + 445, + 547, + 576 + ], + "type": "text", + "content": "Datasets. CelebAMask-HQ [32] consists of manually annotated segmentation masks for each of the 30,000 images in the CelebA-HQ dataset [25]. Each mask has up to 19 classes, including the main facial components such as hair, skin, eyes, and nose, and accessories such as eyeglasses and cloth. CelebA-Dialog [23] provides rich and fine-grained natural language descriptions for the images in CelebA-HQ. In this work, we train the mask-driven diffusion models on CelebAMask-HQ, and the text-driven models on CelebA-Dialog. For dynamic diffuser, we simply combine the mask and text for each image without further processing." + } + ] + } + ], + "index": 68 + }, + { + "bbox": [ + 304, + 577, + 547, + 661 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 577, + 547, + 661 + ], + "spans": [ + { + "bbox": [ + 304, + 577, + 547, + 661 + ], + "type": "text", + "content": "Implementation Details. We adopt LDM [48] for our unimodal diffusion models since it achieves a good balance between quality and speed. Our dynamic diffuser is a time-conditional UNet [49], with conditional embeddings injected via cross-attention [62]. The dynamic diffuser is " + }, + { + "bbox": [ + 304, + 577, + 547, + 661 + ], + "type": "inline_equation", + "content": "\\times 30" + }, + { + "bbox": [ + 304, + 577, + 547, + 661 + ], + "type": "text", + "content": " smaller than the pre-trained diffusion model. The detailed architecture and settings are provided in the supplementary material." + } + ] + } + ], + "index": 69 + }, + { + "bbox": [ + 306, + 670, + 429, + 682 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 670, + 429, + 682 + ], + "spans": [ + { + "bbox": [ + 306, + 670, + 429, + 682 + ], + "type": "text", + "content": "4.2. Comparison Methods" + } + ] + } + ], + "index": 70 + }, + { + "bbox": [ + 306, + 689, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 547, + 713 + ], + "type": "text", + "content": "TediGAN [67, 68] is a StyleGAN-based face synthesis and manipulation method. It projects both text and mask condi" + } + ] + } + ], + "index": 71 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6084" + } + ] + } + ], + "index": 72 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 51, + 81, + 93, + 123 + ], + "blocks": [ + { + "bbox": [ + 73, + 70, + 132, + 78 + ], + "lines": [ + { + "bbox": [ + 73, + 70, + 132, + 78 + ], + "spans": [ + { + "bbox": [ + 73, + 70, + 132, + 78 + ], + "type": "text", + "content": "Multi-Modal Conditions" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 51, + 81, + 93, + 123 + ], + "lines": [ + { + "bbox": [ + 51, + 81, + 93, + 123 + ], + "spans": [ + { + "bbox": [ + 51, + 81, + 93, + 123 + ], + "type": "image", + "image_path": "a33367654f1906533af4ab8fb28400b3d09af7ec3b3a6b192cfc9a2ed3b675af.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 94, + 87, + 144, + 118 + ], + "lines": [ + { + "bbox": [ + 94, + 87, + 144, + 118 + ], + "spans": [ + { + "bbox": [ + 94, + 87, + 144, + 118 + ], + "type": "text", + "content": "His face is covered with short beard. He is a young adult." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 52, + 126, + 93, + 168 + ], + "blocks": [ + { + "bbox": [ + 52, + 126, + 93, + 168 + ], + "lines": [ + { + "bbox": [ + 52, + 126, + 93, + 168 + ], + "spans": [ + { + "bbox": [ + 52, + 126, + 93, + 168 + ], + "type": "image", + "image_path": "bb022db63d2e0adda2410a35ba889f6ac9bd6f3162bb0874c76831e972c85b6b.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 94, + 140, + 134, + 154 + ], + "lines": [ + { + "bbox": [ + 94, + 140, + 134, + 154 + ], + "spans": [ + { + "bbox": [ + 94, + 140, + 134, + 154 + ], + "type": "text", + "content": "She looks very young." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 179, + 545, + 201 + ], + "lines": [ + { + "bbox": [ + 46, + 179, + 545, + 201 + ], + "spans": [ + { + "bbox": [ + 46, + 179, + 545, + 201 + ], + "type": "text", + "content": "Figure 6. Diversity of Face Generation. Given input conditions, our method demonstrates promising synthesis diversity in facial attributes that are not constrained by a specific combination of multi-modal conditions, such as hair color and skin tone." + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 151, + 80, + 194, + 122 + ], + "blocks": [ + { + "bbox": [ + 151, + 80, + 194, + 122 + ], + "lines": [ + { + "bbox": [ + 151, + 80, + 194, + 122 + ], + "spans": [ + { + "bbox": [ + 151, + 80, + 194, + 122 + ], + "type": "image", + "image_path": "c9264fd09f2d4328f6efb8749e2dc8ee0e2d5b23ecadadc7c17a8ced9229def2.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 195, + 80, + 236, + 122 + ], + "blocks": [ + { + "bbox": [ + 195, + 80, + 236, + 122 + ], + "lines": [ + { + "bbox": [ + 195, + 80, + 236, + 122 + ], + "spans": [ + { + "bbox": [ + 195, + 80, + 236, + 122 + ], + "type": "image", + "image_path": "d6bdd71d45001981bf303c32e376a35b43cdb345ed551ce66b1759ae390ba564.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 239, + 80, + 280, + 122 + ], + "blocks": [ + { + "bbox": [ + 239, + 80, + 280, + 122 + ], + "lines": [ + { + "bbox": [ + 239, + 80, + 280, + 122 + ], + "spans": [ + { + "bbox": [ + 239, + 80, + 280, + 122 + ], + "type": "image", + "image_path": "7c6e471ca7ab0380639a28fb4222c278a84db6fca4f27f478c051160ce6283c1.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 282, + 80, + 323, + 122 + ], + "blocks": [ + { + "bbox": [ + 282, + 80, + 323, + 122 + ], + "lines": [ + { + "bbox": [ + 282, + 80, + 323, + 122 + ], + "spans": [ + { + "bbox": [ + 282, + 80, + 323, + 122 + ], + "type": "image", + "image_path": "d1108f04cff15f86373edadb91d4f5b793e4a515af9316e627e781f69920e8b5.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 325, + 72, + 368, + 78 + ], + "lines": [ + { + "bbox": [ + 325, + 72, + 368, + 78 + ], + "spans": [ + { + "bbox": [ + 325, + 72, + 368, + 78 + ], + "type": "text", + "content": "Generated Images" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 325, + 81, + 367, + 122 + ], + "blocks": [ + { + "bbox": [ + 325, + 81, + 367, + 122 + ], + "lines": [ + { + "bbox": [ + 325, + 81, + 367, + 122 + ], + "spans": [ + { + "bbox": [ + 325, + 81, + 367, + 122 + ], + "type": "image", + "image_path": "b85a133ac7ded7c791d388dbfa9f44caa22fd7677041c0a7b3ed5d350c70aa8c.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 370, + 80, + 410, + 122 + ], + "blocks": [ + { + "bbox": [ + 370, + 80, + 410, + 122 + ], + "lines": [ + { + "bbox": [ + 370, + 80, + 410, + 122 + ], + "spans": [ + { + "bbox": [ + 370, + 80, + 410, + 122 + ], + "type": "image", + "image_path": "46dad76165ae071f27551e8d0dec4b66c0d989c27f8a5df9269f1031e9285611.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 412, + 80, + 453, + 122 + ], + "blocks": [ + { + "bbox": [ + 412, + 80, + 453, + 122 + ], + "lines": [ + { + "bbox": [ + 412, + 80, + 453, + 122 + ], + "spans": [ + { + "bbox": [ + 412, + 80, + 453, + 122 + ], + "type": "image", + "image_path": "afdddf07b2a9fd2b480f5c17adad67e692fe0a1b3dd69a58926a66ff2833e285.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 457, + 80, + 497, + 122 + ], + "blocks": [ + { + "bbox": [ + 457, + 80, + 497, + 122 + ], + "lines": [ + { + "bbox": [ + 457, + 80, + 497, + 122 + ], + "spans": [ + { + "bbox": [ + 457, + 80, + 497, + 122 + ], + "type": "image", + "image_path": "67e6e83f4b5991a0da8c9fc8a071b75a44625193cb63e67c7f0ab1da68894c0e.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 500, + 80, + 541, + 122 + ], + "blocks": [ + { + "bbox": [ + 500, + 80, + 541, + 122 + ], + "lines": [ + { + "bbox": [ + 500, + 80, + 541, + 122 + ], + "spans": [ + { + "bbox": [ + 500, + 80, + 541, + 122 + ], + "type": "image", + "image_path": "fef8961c5e3fca360eaa8f50a9b596e7a12d577ddf0310ac8d1e04b23721cdbe.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 151, + 125, + 192, + 167 + ], + "blocks": [ + { + "bbox": [ + 151, + 125, + 192, + 167 + ], + "lines": [ + { + "bbox": [ + 151, + 125, + 192, + 167 + ], + "spans": [ + { + "bbox": [ + 151, + 125, + 192, + 167 + ], + "type": "image", + "image_path": "357fb1a75d26255bfc3c028468306284eae275f3d7cc38eb3a4580533b7a1f36.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 194, + 125, + 236, + 167 + ], + "blocks": [ + { + "bbox": [ + 194, + 125, + 236, + 167 + ], + "lines": [ + { + "bbox": [ + 194, + 125, + 236, + 167 + ], + "spans": [ + { + "bbox": [ + 194, + 125, + 236, + 167 + ], + "type": "image", + "image_path": "cbf82fbd3a6cf6eb8ad9a14c73c25df05e5811b405e6a94342def08bd6f47b7a.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 239, + 125, + 280, + 167 + ], + "blocks": [ + { + "bbox": [ + 239, + 125, + 280, + 167 + ], + "lines": [ + { + "bbox": [ + 239, + 125, + 280, + 167 + ], + "spans": [ + { + "bbox": [ + 239, + 125, + 280, + 167 + ], + "type": "image", + "image_path": "0289ddbb97f925db0dfc73a83794eccfaca1873b93a33184ffa6e2b0fcb46318.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 282, + 125, + 323, + 167 + ], + "blocks": [ + { + "bbox": [ + 282, + 125, + 323, + 167 + ], + "lines": [ + { + "bbox": [ + 282, + 125, + 323, + 167 + ], + "spans": [ + { + "bbox": [ + 282, + 125, + 323, + 167 + ], + "type": "image", + "image_path": "040d206d3991b0e0e22dd2f0addd5bccadac9e3fa335e7ead1143191b1d397d7.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 325, + 125, + 366, + 167 + ], + "blocks": [ + { + "bbox": [ + 325, + 125, + 366, + 167 + ], + "lines": [ + { + "bbox": [ + 325, + 125, + 366, + 167 + ], + "spans": [ + { + "bbox": [ + 325, + 125, + 366, + 167 + ], + "type": "image", + "image_path": "41d152e219859cd14bd89155bccd6c42522d5d3c5664a81541b0d96a3550c19c.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 369, + 125, + 410, + 167 + ], + "blocks": [ + { + "bbox": [ + 369, + 125, + 410, + 167 + ], + "lines": [ + { + "bbox": [ + 369, + 125, + 410, + 167 + ], + "spans": [ + { + "bbox": [ + 369, + 125, + 410, + 167 + ], + "type": "image", + "image_path": "a0cd8a2c011fb9b0d48c4f97a0a8b714872c57904954fd15e57152671df61a00.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 413, + 125, + 454, + 167 + ], + "blocks": [ + { + "bbox": [ + 413, + 125, + 454, + 167 + ], + "lines": [ + { + "bbox": [ + 413, + 125, + 454, + 167 + ], + "spans": [ + { + "bbox": [ + 413, + 125, + 454, + 167 + ], + "type": "image", + "image_path": "a5585141c1f994de37bfc736e02d3d1993d86ff76f9877870f1e103d454f991e.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 457, + 125, + 498, + 167 + ], + "blocks": [ + { + "bbox": [ + 457, + 125, + 498, + 167 + ], + "lines": [ + { + "bbox": [ + 457, + 125, + 498, + 167 + ], + "spans": [ + { + "bbox": [ + 457, + 125, + 498, + 167 + ], + "type": "image", + "image_path": "ad75a40e67e59702546d2b24179c997660cde30ca8fe0e82eceb8e3d2695aab0.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 500, + 125, + 541, + 168 + ], + "blocks": [ + { + "bbox": [ + 500, + 125, + 541, + 168 + ], + "lines": [ + { + "bbox": [ + 500, + 125, + 541, + 168 + ], + "spans": [ + { + "bbox": [ + 500, + 125, + 541, + 168 + ], + "type": "image", + "image_path": "c7b47b8845ca6d621f22a0b1fa5705cfb6b11abd21bb38cbd37d1494e088674c.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + } + ], + "index": 23 + }, + { + "bbox": [ + 46, + 223, + 286, + 247 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 223, + 286, + 247 + ], + "spans": [ + { + "bbox": [ + 46, + 223, + 286, + 247 + ], + "type": "text", + "content": "tions into StyleGAN's " + }, + { + "bbox": [ + 46, + 223, + 286, + 247 + ], + "type": "inline_equation", + "content": "\\mathcal{W}+" + }, + { + "bbox": [ + 46, + 223, + 286, + 247 + ], + "type": "text", + "content": " latent space, and performs style mixing to achieve multi-modal control." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 46, + 248, + 288, + 332 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 248, + 288, + 332 + ], + "spans": [ + { + "bbox": [ + 46, + 248, + 288, + 332 + ], + "type": "text", + "content": "Composable Diffusion [36] parallelly implements two instances of the same text-to-image diffusion model to perform compositional scene generation. We generalize it by replacing the two text-to-image model instances by the text-driven model and the mask-driven model trained using LDM [48]. The two models are composed by averaging the intermediate diffusion results at each step of the reverse process." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 47, + 342, + 160, + 355 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 342, + 160, + 355 + ], + "spans": [ + { + "bbox": [ + 47, + 342, + 160, + 355 + ], + "type": "text", + "content": "4.3. Evaluation Metrics" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 46, + 362, + 288, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 362, + 288, + 422 + ], + "spans": [ + { + "bbox": [ + 46, + 362, + 288, + 422 + ], + "type": "text", + "content": "FID. We use Frechet Inception Distance (FID) [18] to measure the quality of images synthesized by different methods. FID computes the feature representation's distance between generated images and real images. A lower FID implies better sample quality." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 46, + 424, + 287, + 508 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 424, + 287, + 508 + ], + "spans": [ + { + "bbox": [ + 46, + 424, + 287, + 508 + ], + "type": "text", + "content": "CLIP Score. CLIP [42] is a vision-language model trained on large-scale datasets. It uses an image encoder and a text encoder to project images and texts to a common feature space, respectively. The CLIP score is computed as the cosine similarity between the normalized image and text embeddings. A higher score usually indicates higher consistency between the output image and the text caption." + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 46, + 509, + 287, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 509, + 287, + 581 + ], + "spans": [ + { + "bbox": [ + 46, + 509, + 287, + 581 + ], + "type": "text", + "content": "Mask Accuracy. For each output image, we predict the segmentation mask using the face parsing network provided by CelebAMask-HQ [32]. Mask accuracy is the pixel-wise accuracy against the ground-truth segmentation. A higher average accuracy indicates better consistency between the output image and the segmentation mask." + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 46, + 582, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 582, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 582, + 288, + 714 + ], + "type": "text", + "content": "User Study. We conduct user study with 25 human evaluators to measure the effectiveness of the methods perceptually. For face generation, we randomly sample multi-modal conditions in the validation split of CelebA-HQ Dataset, and then synthesize output images given the conditions. Evaluators are provided with the input conditions and the output images, and they are asked to choose the best image based on 1) image photo-realism, 2) image-text consistency, 3) image-mask consistency. User study is conducted similarly for face editing, except that 1) evaluators are also provided with the input image for editing, and 2) they are also asked" + } + ] + } + ], + "index": 31 + }, + { + "type": "table", + "bbox": [ + 313, + 274, + 539, + 323 + ], + "blocks": [ + { + "bbox": [ + 305, + 221, + 545, + 265 + ], + "lines": [ + { + "bbox": [ + 305, + 221, + 545, + 265 + ], + "spans": [ + { + "bbox": [ + 305, + 221, + 545, + 265 + ], + "type": "text", + "content": "Table 1. Quantitative Results of Face Generation. Compared with TediGAN and Composable Diffusion, our method synthesizes images with better quality (lower FID), and higher consistency with the text and mask conditions." + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 313, + 274, + 539, + 323 + ], + "lines": [ + { + "bbox": [ + 313, + 274, + 539, + 323 + ], + "spans": [ + { + "bbox": [ + 313, + 274, + 539, + 323 + ], + "type": "table", + "html": "
MethodFID ↓Text (%) ↑Mask (%) ↑
TediGAN [67,68]157.8124.2772.19
Composable [36]124.6223.9476.11
Ours111.3624.5180.25
", + "image_path": "ebe4090b3693feca9cdad83e22a49e40c1565aca502eff8fb6eb05615f7cafd8.jpg" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "table_body" + } + ], + "index": 33 + }, + { + "type": "image", + "bbox": [ + 320, + 335, + 396, + 402 + ], + "blocks": [ + { + "bbox": [ + 320, + 335, + 396, + 402 + ], + "lines": [ + { + "bbox": [ + 320, + 335, + 396, + 402 + ], + "spans": [ + { + "bbox": [ + 320, + 335, + 396, + 402 + ], + "type": "image", + "image_path": "3d7719ce0812780236e50b03b2ace2e8717640c2d109eddd3949b7ae090b9495.jpg" + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 335, + 405, + 389, + 413 + ], + "lines": [ + { + "bbox": [ + 335, + 405, + 389, + 413 + ], + "spans": [ + { + "bbox": [ + 335, + 405, + 389, + 413 + ], + "type": "text", + "content": "(a) Face Generation" + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 305, + 424, + 547, + 468 + ], + "lines": [ + { + "bbox": [ + 305, + 424, + 547, + 468 + ], + "spans": [ + { + "bbox": [ + 305, + 424, + 547, + 468 + ], + "type": "text", + "content": "Figure 7. User Study. Among the three methods, the majority of users vote for our results as the best for both generation (69.40%) and editing (84.37%), in terms of image quality, condition consistency, and identity preservation." + } + ] + } + ], + "index": 42, + "angle": 0, + "type": "image_caption" + } + ], + "index": 34 + }, + { + "type": "image", + "bbox": [ + 401, + 335, + 478, + 402 + ], + "blocks": [ + { + "bbox": [ + 401, + 335, + 478, + 402 + ], + "lines": [ + { + "bbox": [ + 401, + 335, + 478, + 402 + ], + "spans": [ + { + "bbox": [ + 401, + 335, + 478, + 402 + ], + "type": "image", + "image_path": "d7f016d4115b42bf87fe163aa49feb0ab08580371afafecedc03ea9308f1dcc8.jpg" + } + ] + } + ], + "index": 36, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 423, + 405, + 468, + 414 + ], + "lines": [ + { + "bbox": [ + 423, + 405, + 468, + 414 + ], + "spans": [ + { + "bbox": [ + 423, + 405, + 468, + 414 + ], + "type": "text", + "content": "(b) Face Editing" + } + ] + } + ], + "index": 37, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 490, + 354, + 520, + 361 + ], + "lines": [ + { + "bbox": [ + 490, + 354, + 520, + 361 + ], + "spans": [ + { + "bbox": [ + 490, + 354, + 520, + 361 + ], + "type": "text", + "content": "TediGAN" + } + ] + } + ], + "index": 38, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 492, + 365, + 528, + 371 + ], + "lines": [ + { + "bbox": [ + 492, + 365, + 528, + 371 + ], + "spans": [ + { + "bbox": [ + 492, + 365, + 528, + 371 + ], + "type": "text", + "content": "Composable" + } + ] + } + ], + "index": 39, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 492, + 374, + 510, + 380 + ], + "lines": [ + { + "bbox": [ + 492, + 374, + 510, + 380 + ], + "spans": [ + { + "bbox": [ + 492, + 374, + 510, + 380 + ], + "type": "text", + "content": "Ours" + } + ] + } + ], + "index": 40, + "angle": 0, + "type": "image_footnote" + } + ], + "index": 36 + }, + { + "bbox": [ + 305, + 491, + 545, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 491, + 545, + 515 + ], + "spans": [ + { + "bbox": [ + 305, + 491, + 545, + 515 + ], + "type": "text", + "content": "to assess the level of identity preservation between the edited image and input image." + } + ] + } + ], + "index": 43 + }, + { + "bbox": [ + 306, + 525, + 420, + 538 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 525, + 420, + 538 + ], + "spans": [ + { + "bbox": [ + 306, + 525, + 420, + 538 + ], + "type": "text", + "content": "4.4. Comparative Study" + } + ] + } + ], + "index": 44 + }, + { + "bbox": [ + 304, + 544, + 547, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 544, + 547, + 700 + ], + "spans": [ + { + "bbox": [ + 304, + 544, + 547, + 700 + ], + "type": "text", + "content": "Qualitative Comparison. In Figure 8, we provide the generation results of our Collaborative Diffusion under different combinations of texts and masks. It is observed that our framework is capable of synthesizing realistic outputs consistent with the multi-modal inputs, even for relatively rare conditions, such as a man with long hair. We then compare our method with TediGAN [67, 68] and Composable Diffusion [36] in Figure 4. For example, in the second example, TediGAN and Composable Diffusion's results are inconsistent with the mask, while ours is able to maintain both mask and text consistency. We also show in Figure 6 that our method synthesizes images with high diversity without losing condition consistency." + } + ] + } + ], + "index": 45 + }, + { + "bbox": [ + 317, + 701, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 701, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 317, + 701, + 545, + 713 + ], + "type": "text", + "content": "In addition to image generation, it is also observed in" + } + ] + } + ], + "index": 46 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6085" + } + ] + } + ], + "index": 47 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 51, + 70, + 542, + 376 + ], + "blocks": [ + { + "bbox": [ + 51, + 70, + 542, + 376 + ], + "lines": [ + { + "bbox": [ + 51, + 70, + 542, + 376 + ], + "spans": [ + { + "bbox": [ + 51, + 70, + 542, + 376 + ], + "type": "image", + "image_path": "6ae37ff451b6a6749125bc47f529b2c1727b26ec0107413352237b6e84634ed1.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 386, + 546, + 409 + ], + "lines": [ + { + "bbox": [ + 46, + 386, + 546, + 409 + ], + "spans": [ + { + "bbox": [ + 46, + 386, + 546, + 409 + ], + "type": "text", + "content": "Figure 8. Face Generation Results. Our method generates realistic images under different combinations of multi-modal conditions, even for relatively rare combinations in the training distribution, such as a man with long hair." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 430, + 287, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 430, + 287, + 525 + ], + "spans": [ + { + "bbox": [ + 46, + 430, + 287, + 525 + ], + "type": "text", + "content": "Figure 9 that our framework is able to edit images based on multi-modal conditions while preserving identity. The comparison to existing works in Figure 5 also verifies our effectiveness. For example, in the second example, while TediGAN is unable to synthesize hair consistent to the mask and Composable Diffusion fails to generate beard according to the text, our framework is able to generate results highly consistent to both conditions while maintaining identity." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 531, + 288, + 615 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 531, + 288, + 615 + ], + "spans": [ + { + "bbox": [ + 46, + 531, + 288, + 615 + ], + "type": "text", + "content": "Quantitative Comparison. As shown in Table 1, our Collaborate Diffusion outperforms TediGAN and Composable Diffusion in all three objective metrics. In addition, as depicted in Figure 7, our framework achieves the best result in " + }, + { + "bbox": [ + 46, + 531, + 288, + 615 + ], + "type": "inline_equation", + "content": "69.40\\%" + }, + { + "bbox": [ + 46, + 531, + 288, + 615 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 531, + 288, + 615 + ], + "type": "inline_equation", + "content": "84.37\\%" + }, + { + "bbox": [ + 46, + 531, + 288, + 615 + ], + "type": "text", + "content": " of the time for face generation and editing, respectively. These results verify our effectiveness in both image quality and image-condition consistency." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 641, + 140, + 654 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 641, + 140, + 654 + ], + "spans": [ + { + "bbox": [ + 47, + 641, + 140, + 654 + ], + "type": "text", + "content": "4.5. Ablation Study" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 665, + 287, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 287, + 712 + ], + "type": "text", + "content": "In this section, we visualize the influence functions, and show that it is necessary for the influence function to be both spatially and temporally varying to facilitate effective collaboration." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 430, + 512, + 441 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 430, + 512, + 441 + ], + "spans": [ + { + "bbox": [ + 306, + 430, + 512, + 441 + ], + "type": "text", + "content": "4.5.1 Spatial Variations of Influence Functions" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 449, + 547, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 449, + 547, + 555 + ], + "spans": [ + { + "bbox": [ + 304, + 449, + 547, + 555 + ], + "type": "text", + "content": "As shown in Figure 10, the influence functions behave differently at different spatial regions. For instance, the influence for the mask-driven model mainly lies on the contours of facial regions, such as the outline of hair, face, and eyes, as these regions are crucial in defining facial layout. In contrast, the influence for the text-driven model is stronger at skin regions including cheeks and chin. This is because the attributes related skin texture, such as age and beard length, are better described by text." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 557, + 547, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 557, + 547, + 628 + ], + "spans": [ + { + "bbox": [ + 304, + 557, + 547, + 628 + ], + "type": "text", + "content": "To verify its necessity, we remove the spatial variation of influence functions. From Table 2 we observe that removing the spatial variance results in deterioration in both output quality and condition consistency. This corroborates our hypothesis that it is important to assign different weights to different modalities at different spatial regions." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 646, + 524, + 658 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 646, + 524, + 658 + ], + "spans": [ + { + "bbox": [ + 306, + 646, + 524, + 658 + ], + "type": "text", + "content": "4.5.2 Temporal Variations of Influence Functions" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "type": "text", + "content": "In addition to the spatial variation, it is also observed in Figure 10 that the influence from the mask-driven model is stronger at earlier diffusion stages (i.e., larger " + }, + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "type": "text", + "content": "), since early stages focus on initializing the facial layout using the mask-" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "6086" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 61, + 70, + 534, + 293 + ], + "blocks": [ + { + "bbox": [ + 61, + 70, + 534, + 293 + ], + "lines": [ + { + "bbox": [ + 61, + 70, + 534, + 293 + ], + "spans": [ + { + "bbox": [ + 61, + 70, + 534, + 293 + ], + "type": "image", + "image_path": "57068522a380006889e26d06ddfccbfc8f02c36f0ff3ee1f9979eb113625785d.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 63, + 303, + 530, + 316 + ], + "lines": [ + { + "bbox": [ + 63, + 303, + 530, + 316 + ], + "spans": [ + { + "bbox": [ + 63, + 303, + 530, + 316 + ], + "type": "text", + "content": "Figure 9. Face Editing Results. Given the input real image and target conditions, we display the edited image using our method." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 51, + 335, + 285, + 446 + ], + "blocks": [ + { + "bbox": [ + 51, + 335, + 285, + 446 + ], + "lines": [ + { + "bbox": [ + 51, + 335, + 285, + 446 + ], + "spans": [ + { + "bbox": [ + 51, + 335, + 285, + 446 + ], + "type": "image", + "image_path": "1c37642a1c99a2a65435f81a57bac20a5395e200b8834a938f2dfcb76a5fbe5f.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 454, + 287, + 489 + ], + "lines": [ + { + "bbox": [ + 46, + 454, + 287, + 489 + ], + "spans": [ + { + "bbox": [ + 46, + 454, + 287, + 489 + ], + "type": "text", + "content": "Figure 10. Visualization of Influence Functions. The influence functions vary spatially at different face regions, and temporally at different diffusion timesteps." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 50, + 540, + 282, + 590 + ], + "blocks": [ + { + "bbox": [ + 47, + 503, + 287, + 537 + ], + "lines": [ + { + "bbox": [ + 47, + 503, + 287, + 537 + ], + "spans": [ + { + "bbox": [ + 47, + 503, + 287, + 537 + ], + "type": "text", + "content": "Table 2. Ablation Study. Temporal or spatial suppression in influence variation introduces performance drops, which shows the necessity of influence functions' spatial-temporal adaptivity." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 50, + 540, + 282, + 590 + ], + "lines": [ + { + "bbox": [ + 50, + 540, + 282, + 590 + ], + "spans": [ + { + "bbox": [ + 50, + 540, + 282, + 590 + ], + "type": "table", + "html": "
MethodFID ↓Text (%) ↑Mask (%) ↑
Ours w/o Spatial117.8124.3680.08
Ours w/o Temporal117.3424.4877.07
Ours111.3624.5180.25
", + "image_path": "d29a506eca7161418ee3121e0d9e05b0c0e233c304addcc480b0208676d573d0.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 613, + 287, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 613, + 287, + 662 + ], + "spans": [ + { + "bbox": [ + 46, + 613, + 287, + 662 + ], + "type": "text", + "content": "driven model's predictions. At later stages, the influence from the text-driven model increases as the textural details (e.g., skin wrinkles and beard length) are instantiated using information from the text." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "type": "text", + "content": "We remove the temporal variation in the influence functions to demonstrate its importance. It is shown in Table 2 that both the quality and condition consistency drop without the temporal variation." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 335, + 379, + 346 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 335, + 379, + 346 + ], + "spans": [ + { + "bbox": [ + 306, + 335, + 379, + 346 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 360, + 547, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 360, + 547, + 517 + ], + "spans": [ + { + "bbox": [ + 304, + 360, + 547, + 517 + ], + "type": "text", + "content": "With generative AI gaining increasing attention, multimodal conditioning becomes an indispensable direction to unleash creativity and enable comprehensive controls from human creators. In this work, we take the first step forward and propose Collaborative Diffusion, where pre-trained uni-modal diffusion models collaboratively achieve multi-modal face generation and editing without being re-trained. With our dynamic diffuser, this framework could be used to extend arbitrary uni-modal approach to the multi-modal paradigm through predicting the relative influence of different modalities. We believe our idea of synergizing uni-modal models for multi-modal tasks would be a good inspiration for future works in various domains such as motion and 3D generation." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 521, + 547, + 583 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 521, + 547, + 583 + ], + "spans": [ + { + "bbox": [ + 304, + 521, + 547, + 583 + ], + "type": "text", + "content": "Limitations and Future Work. Since our framework focuses on exploiting pre-trained diffusion models, our performance is dependent on the capability of each model. An orthogonal direction is to train each collaborator on large-scale datasets for performance gain." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 587, + 547, + 636 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 587, + 547, + 636 + ], + "spans": [ + { + "bbox": [ + 304, + 587, + 547, + 636 + ], + "type": "text", + "content": "Potential Negative Societal Impacts. The facial manipulation capabilities of Collaborative Diffusion could be applied maliciously on real human faces. We advise users to apply it only for proper recreational purposes." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 641, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 641, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 641, + 547, + 715 + ], + "type": "text", + "content": "Acknowledgement. We would like to thank Liangyu Chen and Bo Li for proofreading. This study is supported by NTU NAP, MOE AcRF Tier 1 (2021-T1-001-088), and under the RIE2020 Industry Alignment Fund - Industry Collaboration Projects (IAF-ICP) Funding Initiative, as well as cash and in-kind contribution from the industry partner(s)." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6087" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 123 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 123 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 123 + ], + "type": "text", + "content": "[1] Tomer Amit, Eliya Nachmani, Tal Shaharbany, and Lior Wolf. SegDiff: Image segmentation with diffusion probabilistic models. arXiv preprint arXiv:2112.00390, 2021. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 124, + 288, + 157 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 124, + 288, + 157 + ], + "spans": [ + { + "bbox": [ + 53, + 124, + 288, + 157 + ], + "type": "text", + "content": "[2] Jacob Austin, Daniel D Johnson, Jonathan Ho, Daniel Tarlow, and Rianne van den Berg. Structured denoising diffusion models in discrete state-spaces. In NeurIPS, 2021. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 158, + 288, + 190 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 158, + 288, + 190 + ], + "spans": [ + { + "bbox": [ + 53, + 158, + 288, + 190 + ], + "type": "text", + "content": "[3] Omri Avrahami, Dani Lischinski, and Ohad Fried. Blended diffusion for text-driven editing of natural images. In CVPR, 2022. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 191, + 288, + 223 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 191, + 288, + 223 + ], + "spans": [ + { + "bbox": [ + 53, + 191, + 288, + 223 + ], + "type": "text", + "content": "[4] Dmitry Baranchuk, Ivan Rubachev, Andrey Voynov, Valentin Khrulkov, and Artem Babenko. Label-efficient semantic segmentation with diffusion models. In ICLR, 2022. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 224, + 288, + 278 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 224, + 288, + 278 + ], + "spans": [ + { + "bbox": [ + 53, + 224, + 288, + 278 + ], + "type": "text", + "content": "[5] Sam Bond-Taylor, Peter Hessey, Hiroshi Sasaki, Toby P Breckon, and Chris G Willcocks. Unleashing transformers: parallel token prediction with discrete absorbing diffusion for fast high-resolution image generation from vector-quantized codes. In ECCV, 2022. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 280, + 288, + 312 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 280, + 288, + 312 + ], + "spans": [ + { + "bbox": [ + 53, + 280, + 288, + 312 + ], + "type": "text", + "content": "[6] Huiwen Chang, Han Zhang, Lu Jiang, Ce Liu, and William T Freeman. MaskGIT: Masked generative image transformer. In CVPR, 2022. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 314, + 288, + 356 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 314, + 288, + 356 + ], + "spans": [ + { + "bbox": [ + 53, + 314, + 288, + 356 + ], + "type": "text", + "content": "[7] Shin-I Cheng, Yu-Jie Chen, Wei-Chen Chiu, Hung-Yu Tseng, and Hsin-Ying Lee. Adaptively-realistic image generation from stroke and sketch with diffusion model. In WACV, 2023. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 357, + 288, + 401 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 357, + 288, + 401 + ], + "spans": [ + { + "bbox": [ + 53, + 357, + 288, + 401 + ], + "type": "text", + "content": "[8] Guillaume Couairon, Jakob Verbeek, Holger Schwenk, and Matthieu Cord. Diffedit: Diffusion-based semantic image editing with mask guidance. arXiv preprint arXiv:2210.11427, 2022. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 403, + 288, + 424 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 403, + 288, + 424 + ], + "spans": [ + { + "bbox": [ + 53, + 403, + 288, + 424 + ], + "type": "text", + "content": "[9] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat GANs on image synthesis. In NeurIPS, 2021. 1, 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 425, + 287, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 425, + 287, + 468 + ], + "spans": [ + { + "bbox": [ + 48, + 425, + 287, + 468 + ], + "type": "text", + "content": "[10] Ming Ding, Zhuoyi Yang, Wenyi Hong, Wendi Zheng, Chang Zhou, Da Yin, Junyang Lin, Xu Zou, Zhou Shao, Hongxia Yang, et al. CogView: Mastering text-to-image generation via transformers. In NeurIPS, 2021. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 469, + 287, + 512 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 469, + 287, + 512 + ], + "spans": [ + { + "bbox": [ + 48, + 469, + 287, + 512 + ], + "type": "text", + "content": "[11] Patrick Esser, Robin Rombach, Andreas Blattmann, and Bjorn Ommer. ImageBART: Bidirectional context with multinomial diffusion for autoregressive image synthesis. In NeurIPS, 2021. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 514, + 287, + 545 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 514, + 287, + 545 + ], + "spans": [ + { + "bbox": [ + 48, + 514, + 287, + 545 + ], + "type": "text", + "content": "[12] Patrick Esser, Robin Rombach, and Bjorn Ommer. Taming transformers for high-resolution image synthesis. In CVPR, 2021. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 547, + 287, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 547, + 287, + 590 + ], + "spans": [ + { + "bbox": [ + 48, + 547, + 287, + 590 + ], + "type": "text", + "content": "[13] Oran Gafni, Adam Polyak, Oron Ashual, Shelly Sheynin, Devi Parikh, and Yaniv Taigman. Make-a-scene: Scene-based text-to-image generation with human priors. arXiv preprint arXiv:2203.13131, 2022. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 591, + 287, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 591, + 287, + 634 + ], + "spans": [ + { + "bbox": [ + 48, + 591, + 287, + 634 + ], + "type": "text", + "content": "[14] Ian J Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron C Courville, and Yoshua Bengio. Generative adversarial nets. In NeurIPS, 2014. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 635, + 287, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 635, + 287, + 668 + ], + "spans": [ + { + "bbox": [ + 48, + 635, + 287, + 668 + ], + "type": "text", + "content": "[15] Alexandros Graikos, Nikolay Malkin, Nebojsa Jojic, and Dimitris Samaras. Diffusion models as plug-and-play priors. In NeurIPS, 2022. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "type": "text", + "content": "[16] Shuyang Gu, Dong Chen, Jianmin Bao, Fang Wen, Bo Zhang, Dongdong Chen, Lu Yuan, and Baining Guo. Vector quantized diffusion model for text-to-image synthesis. In CVPR, 2022. 1, 2" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 712 + ], + "type": "list", + "angle": 0, + "index": 36, + "blocks": [ + { + "bbox": [ + 307, + 73, + 547, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 547, + 106 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 547, + 106 + ], + "type": "text", + "content": "[17] William Harvey, Saeid Naderiparizi, Vaden Masrani, Christian Weilbach, and Frank Wood. Flexible diffusion modeling of long videos. arXiv preprint arXiv:2205.11495, 2022. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 107, + 547, + 148 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 107, + 547, + 148 + ], + "spans": [ + { + "bbox": [ + 307, + 107, + 547, + 148 + ], + "type": "text", + "content": "[18] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. GANs trained by a two time-scale update rule converge to a local nash equilibrium. In NeurIPS, 2017. 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 150, + 547, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 150, + 547, + 205 + ], + "spans": [ + { + "bbox": [ + 308, + 150, + 547, + 205 + ], + "type": "text", + "content": "[19] Jonathan Ho, William Chan, Chitwan Sahara, Jay Whang, Ruiqi Gao, Alexey Gritsanko, Diederik P Kingma, Ben Poole, Mohammad Norouzi, David J Fleet, et al. Imagen video: High definition video generation with diffusion models. arXiv preprint arXiv:2210.02303, 2022. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 206, + 547, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 206, + 547, + 227 + ], + "spans": [ + { + "bbox": [ + 308, + 206, + 547, + 227 + ], + "type": "text", + "content": "[20] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In NeurIPS, 2020. 1, 2, 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 228, + 547, + 260 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 228, + 547, + 260 + ], + "spans": [ + { + "bbox": [ + 308, + 228, + 547, + 260 + ], + "type": "text", + "content": "[21] Jonathan Ho, Chitwan Sahara, William Chan, David J Fleet, Mohammad Norouzi, and Tim Salimans. Cascaded diffusion models for high fidelity image generation. JMLR, 2022. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 261, + 547, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 261, + 547, + 293 + ], + "spans": [ + { + "bbox": [ + 308, + 261, + 547, + 293 + ], + "type": "text", + "content": "[22] Aapo Hyvarinen and Peter Dayan. Estimation of non-normalized statistical models by score matching. JMLR, 2005. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 294, + 547, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 294, + 547, + 326 + ], + "spans": [ + { + "bbox": [ + 308, + 294, + 547, + 326 + ], + "type": "text", + "content": "[23] Yuming Jiang, Ziqi Huang, Xingang Pan, Chen Change Loy, and Ziwei Liu. Talk-to-Edit: Fine-grained facial editing via dialog. In ICCV, 2021. 5" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 327, + 547, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 327, + 547, + 360 + ], + "spans": [ + { + "bbox": [ + 308, + 327, + 547, + 360 + ], + "type": "text", + "content": "[24] Yuming Jiang, Shuai Yang, Haonan Qju, Wayne Wu, Chen Change Loy, and Ziwei Liu. Text2human: Text-driven controllable human image generation. ACM TOG, 2022. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 361, + 547, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 361, + 547, + 392 + ], + "spans": [ + { + "bbox": [ + 308, + 361, + 547, + 392 + ], + "type": "text", + "content": "[25] Tero Karras, Timo Aila, Samuli Laine, and Jaakko Lehtinen. Progressive growing of GANs for improved quality, stability, and variation. In ICLR, 2018. 5" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 393, + 547, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 393, + 547, + 426 + ], + "spans": [ + { + "bbox": [ + 308, + 393, + 547, + 426 + ], + "type": "text", + "content": "[26] Tero Karras, Miika Aittala, Samuli Laine, Erik Härkönen, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Alias-free generative adversarial networks. In NeurIPS, 2021. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 426, + 547, + 458 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 426, + 547, + 458 + ], + "spans": [ + { + "bbox": [ + 308, + 426, + 547, + 458 + ], + "type": "text", + "content": "[27] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In CVPR, 2019. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 308, + 460, + 547, + 492 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 460, + 547, + 492 + ], + "spans": [ + { + "bbox": [ + 308, + 460, + 547, + 492 + ], + "type": "text", + "content": "[28] Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of StyleGAN. In CVPR, 2020. 2" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 308, + 493, + 547, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 493, + 547, + 536 + ], + "spans": [ + { + "bbox": [ + 308, + 493, + 547, + 536 + ], + "type": "text", + "content": "[29] Bahjat Kawar, Shiran Zada, Oran Lang, Omer Tov, Huiwen Chang, Tali Dekel, Inbar Mosseri, and Michal Irani. Imagic: Text-based real image editing with diffusion models. arXiv preprint arXiv:2210.09276, 2022. 1, 2, 5" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 308, + 537, + 547, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 537, + 547, + 568 + ], + "spans": [ + { + "bbox": [ + 308, + 537, + 547, + 568 + ], + "type": "text", + "content": "[30] Gwanghyun Kim, Taesung Kwon, and Jong Chul Ye. DiffusionCLIP: Text-guided diffusion models for robust image manipulation. In CVPR, 2022. 1, 2" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 308, + 570, + 547, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 570, + 547, + 602 + ], + "spans": [ + { + "bbox": [ + 308, + 570, + 547, + 602 + ], + "type": "text", + "content": "[31] Jing Yu Koh, Jason Baldridge, Honglak Lee, and Yinfei Yang. Text-to-image generation grounded by fine-grained user attention. In WACV, 2021. 2" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 308, + 603, + 547, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 603, + 547, + 635 + ], + "spans": [ + { + "bbox": [ + 308, + 603, + 547, + 635 + ], + "type": "text", + "content": "[32] Cheng-Han Lee, Ziwei Liu, Lingyun Wu, and Ping Luo. MaskGAN: Towards diverse and interactive facial image manipulation. In CVPR, 2020. 2, 5, 6" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 308, + 636, + 547, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 636, + 547, + 668 + ], + "spans": [ + { + "bbox": [ + 308, + 636, + 547, + 668 + ], + "type": "text", + "content": "[33] Bowen Li, Xiaojuan Qi, Thomas Lukasiewicz, and Philip HS Torr. ManiGAN: Text-guided image manipulation. In CVPR, 2020. 2" + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 308, + 669, + 547, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 669, + 547, + 712 + ], + "spans": [ + { + "bbox": [ + 308, + 669, + 547, + 712 + ], + "type": "text", + "content": "[34] Wenbo Li, Pengchuan Zhang, Lei Zhang, Qiuyuan Huang, Xiaodong He, Siwei Lyu, and Jianfeng Gao. Object-driven text-to-image synthesis via adversarial training. In CVPR, 2019. 2" + } + ] + } + ], + "index": 35 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "6088" + } + ] + } + ], + "index": 37 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 288, + 106 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 288, + 106 + ], + "type": "text", + "content": "[35] Zhiheng Li, Martin Renqiang Min, Kai Li, and Chenliang Xu. StyleT2I: Toward compositional and high-fidelity text-to-image synthesis. In CVPR, 2022. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 107, + 288, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 107, + 288, + 150 + ], + "spans": [ + { + "bbox": [ + 48, + 107, + 288, + 150 + ], + "type": "text", + "content": "[36] Nan Liu, Shuang Li, Yilun Du, Antonio Torralba, and Joshua B Tenenbaum. Compositional visual generation with composable diffusion models. arXiv preprint arXiv:2206.01714, 2022. 4, 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 152, + 288, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 152, + 288, + 195 + ], + "spans": [ + { + "bbox": [ + 48, + 152, + 288, + 195 + ], + "type": "text", + "content": "[37] Xihui Liu, Dong Huk Park, Samaneh Azadi, Gong Zhang, Arman Chopikyan, Yuxiao Hu, Humphrey Shi, Anna Rohrbach, and Trevor Darrell. More control for free! image synthesis with semantic diffusion guidance. In WACV, 2023. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 197, + 288, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 197, + 288, + 239 + ], + "spans": [ + { + "bbox": [ + 48, + 197, + 288, + 239 + ], + "type": "text", + "content": "[38] Chenlin Meng, Yutong He, Yang Song, Jiaming Song, Jiajun Wu, Jun-Yan Zhu, and Stefano Ermon. SDEdit: Guided image synthesis and editing with stochastic differential equations. In ICLR, 2022. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 241, + 288, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 241, + 288, + 296 + ], + "spans": [ + { + "bbox": [ + 48, + 241, + 288, + 296 + ], + "type": "text", + "content": "[39] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. GLIDE: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741, 2021. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 297, + 288, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 297, + 288, + 330 + ], + "spans": [ + { + "bbox": [ + 48, + 297, + 288, + 330 + ], + "type": "text", + "content": "[40] Taesung Park, Ming-Yu Liu, Ting-Chun Wang, and Jun-Yan Zhu. Semantic image synthesis with spatially-adaptive normalization. In CVPR, 2019. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 331, + 288, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 331, + 288, + 364 + ], + "spans": [ + { + "bbox": [ + 48, + 331, + 288, + 364 + ], + "type": "text", + "content": "[41] Or Patashnik, Zongze Wu, Eli Shechtman, Daniel Cohen-Or, and Dani Lischinski. StyleCLIP: Text-driven manipulation of stylegan imagery. In ICCV, 2021. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 365, + 288, + 419 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 365, + 288, + 419 + ], + "spans": [ + { + "bbox": [ + 48, + 365, + 288, + 419 + ], + "type": "text", + "content": "[42] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, 2021. 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 421, + 288, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 421, + 288, + 464 + ], + "spans": [ + { + "bbox": [ + 48, + 421, + 288, + 464 + ], + "type": "text", + "content": "[43] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with CLIP latents. arXiv preprint arXiv:2204.06125, 2022. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 466, + 288, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 466, + 288, + 498 + ], + "spans": [ + { + "bbox": [ + 48, + 466, + 288, + 498 + ], + "type": "text", + "content": "[44] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In ICML, 2021. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 500, + 288, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 500, + 288, + 533 + ], + "spans": [ + { + "bbox": [ + 48, + 500, + 288, + 533 + ], + "type": "text", + "content": "[45] Scott Reed, Zeynep Akata, Xinchen Yan, Lajanugen Logeswaran, Bernt Schiele, and Honglak Lee. Generative adversarial text to image synthesis. In ICML, 2016. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 534, + 288, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 534, + 288, + 566 + ], + "spans": [ + { + "bbox": [ + 48, + 534, + 288, + 566 + ], + "type": "text", + "content": "[46] Scott E Reed, Zeynep Akata, Santosh Mohan, Samuel Tenka, Bernt Schiele, and Honglak Lee. Learning what and where to draw. In NeurIPS, 2016. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 567, + 288, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 567, + 288, + 610 + ], + "spans": [ + { + "bbox": [ + 48, + 567, + 288, + 610 + ], + "type": "text", + "content": "[47] Elad Richardson, Yuval Alaluf, Or Patashnik, Yotam Nitzan, Yaniv Azar, Stav Shapiro, and Daniel Cohen-Or. Encoding in style: a StyleGAN encoder for image-to-image translation. In CVPR, 2021. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 613, + 288, + 655 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 613, + 288, + 655 + ], + "spans": [ + { + "bbox": [ + 48, + 613, + 288, + 655 + ], + "type": "text", + "content": "[48] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 1, 2, 5, 6" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 657, + 288, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 657, + 288, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 657, + 288, + 689 + ], + "type": "text", + "content": "[49] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-Net: Convolutional networks for biomedical image segmentation. In MICCAI, 2015. 4, 5" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 691, + 288, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 288, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 288, + 712 + ], + "type": "text", + "content": "[50] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 712 + ], + "type": "list", + "angle": 0, + "index": 34, + "blocks": [ + { + "bbox": [ + 327, + 73, + 547, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 547, + 95 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 547, + 95 + ], + "type": "text", + "content": "tuning text-to-image diffusion models for subject-driven generation. arXiv preprint arXiv:2208.12242, 2022. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 96, + 547, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 96, + 547, + 160 + ], + "spans": [ + { + "bbox": [ + 307, + 96, + 547, + 160 + ], + "type": "text", + "content": "[51] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily Denton, Seyed Kamyar Seyed Ghasemipour, Burcu Karagol Ayan, S Sara Mahdavi, Rapha Gontijo Lopes, et al. Photorealistic text-to-image diffusion models with deep language understanding. arXiv preprint arXiv:2205.11487, 2022. 1, 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 163, + 547, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 163, + 547, + 195 + ], + "spans": [ + { + "bbox": [ + 307, + 163, + 547, + 195 + ], + "type": "text", + "content": "[52] Chitwan Sahara, Jonathan Ho, William Chan, Tim Salimans, David J Fleet, and Mohammad Norouzi. Image superresolution via iterative refinement. IEEE TPAMI, 2022. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 197, + 547, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 197, + 547, + 228 + ], + "spans": [ + { + "bbox": [ + 307, + 197, + 547, + 228 + ], + "type": "text", + "content": "[53] Yujun Shen, Jinjin Gu, Xiaou Tang, and Bolei Zhou. Interpreting the latent space of GANs for semantic face editing. In CVPR, 2020. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 230, + 547, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 230, + 547, + 262 + ], + "spans": [ + { + "bbox": [ + 307, + 230, + 547, + 262 + ], + "type": "text", + "content": "[54] Yujun Shen, Ceyuan Yang, Xiaou Tang, and Bolei Zhou. InterFaceGAN: Interpreting the disentangled face representation learned by GANs. IEEE TPAMI, 2020. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 264, + 547, + 317 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 264, + 547, + 317 + ], + "spans": [ + { + "bbox": [ + 307, + 264, + 547, + 317 + ], + "type": "text", + "content": "[55] Uriel Singer, Adam Polyak, Thomas Hayes, Xi Yin, Jie An, Songyang Zhang, Qiyuan Hu, Harry Yang, Oron Ashual, Oran Gafni, et al. Make-a-video: Text-to-video generation without text-video data. arXiv preprint arXiv:2209.14792, 2022. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 319, + 547, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 319, + 547, + 353 + ], + "spans": [ + { + "bbox": [ + 307, + 319, + 547, + 353 + ], + "type": "text", + "content": "[56] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In ICML, 2015. 1, 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 354, + 547, + 375 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 354, + 547, + 375 + ], + "spans": [ + { + "bbox": [ + 307, + 354, + 547, + 375 + ], + "type": "text", + "content": "[57] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In ICLR, 2021. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 376, + 547, + 419 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 376, + 547, + 419 + ], + "spans": [ + { + "bbox": [ + 307, + 376, + 547, + 419 + ], + "type": "text", + "content": "[58] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. In ICLR, 2021. 1, 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 421, + 547, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 421, + 547, + 453 + ], + "spans": [ + { + "bbox": [ + 307, + 421, + 547, + 453 + ], + "type": "text", + "content": "[59] Yang Song, Jingwen Zhu, Dawei Li, Xiaolong Wang, and Hairong Qi. Talking face generation by conditional recurrent adversarial network. *IJCAI*, 2019. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 455, + 547, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 455, + 547, + 487 + ], + "spans": [ + { + "bbox": [ + 307, + 455, + 547, + 487 + ], + "type": "text", + "content": "[60] Jianxin Sun, Qiyao Deng, Qi Li, Muyi Sun, Min Ren, and Zhenan Sun. AnyFace: Free-style text-to-face synthesis and manipulation. In CVPR, 2022. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 489, + 547, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 489, + 547, + 511 + ], + "spans": [ + { + "bbox": [ + 307, + 489, + 547, + 511 + ], + "type": "text", + "content": "[61] Aaron Van Den Oord, Oriol Vinyals, et al. Neural discrete representation learning. In NeurIPS, 2017. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 512, + 547, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 512, + 547, + 544 + ], + "spans": [ + { + "bbox": [ + 307, + 512, + 547, + 544 + ], + "type": "text", + "content": "[62] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In NeurIPS, 2017. 5" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 307, + 545, + 547, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 545, + 547, + 601 + ], + "spans": [ + { + "bbox": [ + 307, + 545, + 547, + 601 + ], + "type": "text", + "content": "[63] Ruben Villegas, Mohammad Babaeizadeh, Pieter-Jan Kindermans, Hernan Moraldo, Han Zhang, Mohammad Taghi Saffar, Santiago Castro, Julius Kunze, and Dumitru Erhan. Phenaki: Variable length video generation from open domain textual description. arXiv preprint arXiv:2210.02399, 2022. 2" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 307, + 601, + 547, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 601, + 547, + 624 + ], + "spans": [ + { + "bbox": [ + 307, + 601, + 547, + 624 + ], + "type": "text", + "content": "[64] Pascal Vincent. A connection between score matching and denoising autoencoders. Neural Computation, 2011. 2" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 307, + 624, + 547, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 624, + 547, + 667 + ], + "spans": [ + { + "bbox": [ + 307, + 624, + 547, + 667 + ], + "type": "text", + "content": "[65] Tengfei Wang, Ting Zhang, Bo Zhang, Hao Ouyang, Dong Chen, Qifeng Chen, and Fang Wen. Pretraining is all you need for image-to-image translation. arXiv preprint arXiv:2205.12952, 2022. 1, 2" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 307, + 669, + 547, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 669, + 547, + 712 + ], + "spans": [ + { + "bbox": [ + 307, + 669, + 547, + 712 + ], + "type": "text", + "content": "[66] Weilun Wang, Jianmin Bao, Wengang Zhou, Dongdong Chen, Dong Chen, Lu Yuan, and Houqiang Li. Semantic image synthesis via diffusion models. arXiv preprint arXiv:2207.00050, 2022. 1" + } + ] + } + ], + "index": 33 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "6089" + } + ] + } + ], + "index": 35 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 274 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "type": "text", + "content": "[67] Weihao Xia, Yujiu Yang, Jing-Hao Xue, and Baoyuan Wu. TediGAN: Text-guided diverse face image generation and manipulation. In CVPR, 2021. 2, 5, 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 107, + 288, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 107, + 288, + 140 + ], + "spans": [ + { + "bbox": [ + 49, + 107, + 288, + 140 + ], + "type": "text", + "content": "[68] Weihao Xia, Yujiu Yang, Jing-Hao Xue, and Baoyuan Wu. Towards open-world text-guided face image generation and manipulation. arXiv preprint arXiv:2104.08910, 2021. 2, 5, 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 141, + 288, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 141, + 288, + 184 + ], + "spans": [ + { + "bbox": [ + 49, + 141, + 288, + 184 + ], + "type": "text", + "content": "[69] Tao Xu, Pengchuan Zhang, Qiuyuan Huang, Han Zhang, Zhe Gan, Xiaolei Huang, and Xiaodong He. AttnGAN: Fine-grained text to image generation with attentional generative adversarial networks. In CVPR, 2018. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 186, + 288, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 186, + 288, + 229 + ], + "spans": [ + { + "bbox": [ + 49, + 186, + 288, + 229 + ], + "type": "text", + "content": "[70] Han Zhang, Tao Xu, Hongsheng Li, Shaoting Zhang, Xiaogang Wang, Xiaolei Huang, and Dimitris N Metaxas. StackGAN: Text to photo-realistic image synthesis with stacked generative adversarial networks. In ICCV, 2017. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 49, + 231, + 288, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 231, + 288, + 274 + ], + "spans": [ + { + "bbox": [ + 49, + 231, + 288, + 274 + ], + "type": "text", + "content": "[71] Han Zhang, Tao Xu, Hongsheng Li, Shaoting Zhang, Xiaogang Wang, Xiaolei Huang, and Dimitris N Metaxas. StackGAN++: Realistic image synthesis with stacked generative adversarial networks. IEEE TPAMI, 2018. 2" + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 316, + 757 + ], + "type": "text", + "content": "6090" + } + ] + } + ], + "index": 6 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Collaborative Noisy Label Cleaner_ Learning Scene-Aware Trailers for Multi-Modal Highlight Detection in Movies/b5b9b0f2-713f-4176-89e7-acec845e8cbb_content_list.json b/2023/Collaborative Noisy Label Cleaner_ Learning Scene-Aware Trailers for Multi-Modal Highlight Detection in Movies/b5b9b0f2-713f-4176-89e7-acec845e8cbb_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..b2f023f2358ce41579a2a6c72d0e71e875544d1a --- /dev/null +++ b/2023/Collaborative Noisy Label Cleaner_ Learning Scene-Aware Trailers for Multi-Modal Highlight Detection in Movies/b5b9b0f2-713f-4176-89e7-acec845e8cbb_content_list.json @@ -0,0 +1,1618 @@ +[ + { + "type": "text", + "text": "Collaborative Noisy Label Cleaner: Learning Scene-aware Trailers for Multi-modal Highlight Detection in Movies", + "text_level": 1, + "bbox": [ + 129, + 130, + 841, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Bei Gan Xiujun Shu* Ruizhi Qiao* Haoqian Wu Keyu Chen Hanjun Li Bo Ren Tencent YouTu Lab {stylegan, xiujunshu, ruizhiqiao, linuswu, yolochen, hanjunli, timren}@tencent.com", + "bbox": [ + 114, + 203, + 854, + 258 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/fbfc481364a10593e809c90f8d3dda2dee5a61a400850c6328c9389a0281ca03.jpg", + "image_caption": [ + "Figure 1. As the preview of the movie, trailers are selected by professionals to grab an audience's attention. However, trailers are usually composed with shots sparsely selected from movies to avoid spoilers, and the audience cannot get complete highlight information. Some trailer clips convey the artistic style of the film only and lack movie storylines, disturbing the audience's impressions. In addition, different audiences may be interested in different styles of clips, which makes it challenging to learn highlights from them." + ], + "image_footnote": [], + "bbox": [ + 120, + 284, + 844, + 532 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 612, + 313, + 628 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Movie highlights stand out of the screenplay for efficient browsing and play a crucial role on social media platforms. Based on existing efforts, this work has two observations: (1) For different annotators, labeling highlight has uncertainty, which leads to inaccurate and time-consuming annotations. (2) Besides previous supervised or unsupervised settings, some existing video corpora can be useful, e.g., trailers, but they are often noisy and incomplete to cover the full highlights. In this work, we study a more practical and promising setting, i.e., reformulating highlight detection as \"learning with noisy labels\". This setting does not require time-consuming manual annotations and can fully utilize existing abundant video corpora. First, based on movie trailers, we leverage scene segmentation to obtain complete shots, which are regarded as noisy labels. Then, we propose a Collaborative noisy Label", + "bbox": [ + 75, + 636, + 473, + 878 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Cleaner (CLC) framework to learn from noisy highlight moments. CLC consists of two modules: augmented crosspropagation (ACP) and multi-modality cleaning (MMC). The former aims to exploit the closely related audio-visual signals and fuse them to learn unified multi-modal representations. The latter aims to achieve cleaner highlight labels by observing the changes in losses among different modalities. To verify the effectiveness of CLC, we further collect a large-scale highlight dataset named MovieLights. Comprehensive experiments on MovieLights and YouTube Highlights datasets demonstrate the effectiveness of our approach. Code has been made available at: https://github.com/TencentYoutuResearch/HighlightDetection-CLC.", + "bbox": [ + 500, + 613, + 893, + 824 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 500, + 844, + 632, + 859 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "With the growing number of new publications of movies in theaters and streaming media, audiences become even", + "bbox": [ + 500, + 869, + 893, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author.", + "bbox": [ + 94, + 887, + 220, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "18898", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "harder to choose their favorite one to enjoy for the next two hours. An effective solution is to watch the movie trailers before choosing the right movie. This is because trailers are generally carefully edited by filmmakers and contain the most prominent clips from the original movies. As a condensed version of full-length movies, trailers are elaborately made with highlight moments to impress the audiences. Consequently, they are high potential in serving as supervision sources to train automatic video highlight detection algorithms and facilitating the mass production of derivative works for video creators in online video platforms, e.g., YouTube and TikTok.", + "bbox": [ + 75, + 90, + 472, + 272 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Existing video highlight detection (VHD) approaches are generally trained with annotated key moments of long-form videos. However, they are not suitable to tackle the movie highlight detection task by directly learning from trailers. The edited shots in trailers are not equivalent to ground-truth highlight annotations in movies. Although a previous work [43] leverages the officially-released trailers as the weak supervision to train a highlight detector, the highlighted ness of trailer shots is extremely noisy and varies with the preference of audiences, as shown in Fig. 1. On one hand, trailers tend to be purposefully edited to avoid spoilers, thus missing key moments of the storylines. On the other hand, some less important moments in the original movies are over-emphasized in the trailers because of some artistic or commercial factors. The subjective nature of trailer shots makes them noisy for the VHD task, which is ignored by existing VHD approaches.", + "bbox": [ + 75, + 275, + 472, + 532 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To alleviate the issue, we reformulate the highlight detection task as \"learning with noisy labels\". Specifically, we first leverage a scene-segmentation model to obtain the movie scene boundaries. The clips containing trailers and clips from the same scenes as the trailers provide more complete storylines. They have a higher probability of being highlight moments but still contain some noisy moments. Subsequently, we introduce a framework named Collaborative noisy Label Cleaner (CLC) to learn from these pseudo-noisy labels. The framework firstly enhances the modality perceptual consistency via the augmented crosspropagation (ACP) module, which exploits closely related audio-visual signals during training. In addition, a multimodality cleaning (MMC) mechanism is designed to filter out noisy and incomplete labels.", + "bbox": [ + 75, + 535, + 472, + 762 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To support this study and facilitate benchmarking existing methods in this direction, we construct MovieLights, a Movie Highlight Detection Dataset. MovieLights contains 174 movies and the highlight moments are all from officially released trailers. The total length of these videos is over 370 hours. We conduct extensive experiments on MovieLights, in which our CLC exhibits promising results. We also demonstrate that our proposed CLC achieves significant performance-boosting over the state-of-the-art on", + "bbox": [ + 75, + 763, + 472, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "the public VHD benchmarks.", + "bbox": [ + 498, + 90, + 696, + 104 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In summary, our major contributions are as follows:", + "bbox": [ + 517, + 106, + 861, + 121 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We introduce a scene-aware paradigm to learn high-light moments in movies without any manual annotation. To the best of our knowledge, this is the first time that highlights detection is regarded as learning with noisy labels.", + "- We present an augmented cross-propagation to capture the interactions across modalities and a consistency loss to maximize the agreement between the different modalities.", + "- We incorporate a multi-modality noisy label cleaner to tackle label noise, which further improves the robustness of networks to annotation noise.", + "- Experiments on movie datasets and benchmark datasets validate the effectiveness of our framework." + ], + "bbox": [ + 517, + 128, + 890, + 364 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 500, + 377, + 650, + 393 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Video Highlight Detection. This task aims to identifying the interesting moments from untrimmed videos. In recent years, the videos studied for this task extend from domain-specific sport videos [41] to general videos such as social media videos [37], news [35], first-person videos [53] and vlog [25]. Most of previous works [14,21,53] interpret the video highlight detection task as a segment-level ranking problem. They compare pairwise segments from same domain video in order to learn a model that assigns highlight scores to these segments where the highlight segments receive higher scores than the non-highlight segments. MINI-Net [16] proposes to cast highlight detection as multiple instance ranking network learning. SL-Module [50] explores the highlight detection problem through Unsupervised Domain Adaptation (UDA) [32]. UMT [29] integrates highlight detection and moment retrieval into a unified framework and conduct joint optimization. PLD [45] models the video highlight detection into a pixel-level distinction estimation task. In this work, we regard highlight detection as learning with noisy labels. Joint-VA [2] also considers video highlight detection from the perspective of noise. However, it focuses on noise in features, such as videos having noisy audio when the microphone constantly has water splashing against it. We focus on specific annotation noise in video highlight detection.", + "bbox": [ + 496, + 402, + 890, + 779 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Studies on Movies and Trailers. Studies on movies and trailers have received increased attention in research. [18] introduces a comprehensive dataset for movie understanding. [4, 42] try to model the relationships among the movie characters. [7, 34, 47] focus on breaking the storylines of movies into semantically cohesive parts. Besides the studies on movies, efforts have been made to develop trailer understanding. [57] presents a movie summarization system", + "bbox": [ + 496, + 779, + 890, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "18899", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "and composes movie summaries in terms of user experience evaluation. [9] designs a movie trailer dataset for the evaluation of video-based recommender systems. [19] is the first approach that bridges trailers and movies and allows knowledge learned from trailers to be transferred to full movie analysis. In [43], the visual module and the temporal analysis module are respectively trained on trailers and movies. Because of the inaccessibility of public trailer-related benchmarks, we construct a new dataset (MovieLights) to detect the highlight moments in movies.", + "bbox": [ + 75, + 90, + 472, + 242 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Learning with Noisy Labels. Learning with noisy labels has been a long-standing problem in computer vision. There are three kinds of approaches to this problem. One of the most common strategies for tackling label noise is to capture the transition probabilities between noisy labels and clean labels [31, 38, 48, 49, 52, 56]. Another solution is to design robust loss functions for model training against noisy labels [8, 26, 28, 30, 44, 55]. A popular method is to design a mechanism to select clean samples or give lower weight for noise samples in the training set to reduce impact of noise [15, 17, 20, 39, 46, 51]. In this paper, we attempt to solve the problem by exploiting multi-modalities nature of movies.", + "bbox": [ + 75, + 243, + 472, + 439 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Movie Highlight Dataset", + "text_level": 1, + "bbox": [ + 76, + 449, + 307, + 467 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this paper, we aim to detect the highlight moments in movies by learning from easily accessible trailers as the noisy supervision. However, the existing movie and trailer-related benchmarks [3, 9, 18] lack sufficient functions for this task, such as the absence of full-length movies and ground-truth highlight annotations. Huang et al. [19] propose to respectively learn visual representations from trailers and temporal structure from full-length movies in their constructed Large-Scale Movie and Trailer Dataset (LSMTD). Wang [43] constructs a Trailer Moment Detection Dataset (TMDD) for detecting trailer moments from full-length movies without explicit human annotation. Both LSMTD and TMDD are not publicly available, while TMDD only contains three movie genres.", + "bbox": [ + 75, + 476, + 468, + 686 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The inaccessibility of public benchmarks motivates us to construct a new dataset, named Movie Highlight Detection Dataset (MovieLights). In particular, we purchase a set of movies from commercial channels and collect their corresponding trailers from streaming platforms such as YouTube, covering at least 25 genres to ensure content diversity. The movies and trailers are then prepossessed by shot segmentation [36] and scene segmentation [47], respectively. The resulting shots are a series of consecutive frames taken by the camera until a physical interruption, and the scenes are consecutive shots that share a semantically related theme.", + "bbox": [ + 75, + 688, + 468, + 868 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "As seen in Fig. 2, to build the ground truth, we conduct Faiss [22] to obtain visual similarity matching between", + "bbox": [ + 76, + 869, + 468, + 901 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/2d233ba27a1ebdcf65e78f232b433d38bef4dbaef03c4f3eba3e9bf801fb0603.jpg", + "image_caption": [ + "Figure 2. The labeling process of MovieLights. For the training set, we introduce a scene-aware paradigm to obtain labels automatically. For the testing set, we collect 2 sets of labels from different annotators." + ], + "image_footnote": [], + "bbox": [ + 501, + 88, + 893, + 223 + ], + "page_idx": 2 + }, + { + "type": "table", + "img_path": "images/3a9e4ae078c76ef97f82a76549573ac1c657cc744ae8f91001f44d7cdf31a421.jpg", + "table_caption": [ + "Table 1. The basic statistics of MovieLights." + ], + "table_footnote": [], + "table_body": "
TrainTest
Movie Number14430
Avg Durations per Movie2.19h2.14h
Avg Shot Number per Movie18521940
Avg Scene Number per Movie207193
Annotator1 Positive sample Proportion-0.27
Annotator2 Positive sample Proportion-0.30
Positive sample proportion0.350.21
", + "bbox": [ + 516, + 325, + 875, + 446 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "trailer frames and movie frames. We locate trailer moments in the movie and align them with the movie shots as annotation references. For the testing set, we collect 2 sets of moments for each movie from different workers, and these moments are annotated by different annotators independently. To ensure the consistency of results from different annotations, during the annotation procedure, all highlight moments must be related to the annotation references. Though all selected shots are relevant to the trailers, as highlight moments can be subjective, they may still vary in their saliency and time span. We calculate the intersection between every pair of moments annotated as the ground truth. However, the vast diversity of movie storylines makes the annotation challenging as it is time-consuming and requires annotators to be familiar with the movie. To collect a large amount of training data efficiently, we introduce a scene-aware paradigm to obtain the highlight moments label without any manual annotation. Specifically, we expand the shot-level annotation references to the scene span as positive samples automatically. It will capture the complete scene context of the trailer shot with movie storylines. Since the trailer shots may contain some less important moments, the acquired highlight labels are still noisy.", + "bbox": [ + 496, + 476, + 890, + 823 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "As seen in Tab. 1, MovieLights contains 174 movies in full length with their official trailers and it is split into a training set with 144 movies, and a testing set with 30 movies. The content diversity is ensured by the rich domain informations (more than 25 genres) and abundant segments", + "bbox": [ + 496, + 825, + 893, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "18900", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "(325k shots and 36k scenes). Most movies in our dataset are between 90 to 150 minutes, and the length of the annotated moments varies from tens of seconds to several minutes. As the acquired positive highlight moments take up $35\\%$ in the training set while the annotated true positives take up $21\\%$ in the testing set, the difference tells the obvious existence of label noise.", + "bbox": [ + 75, + 90, + 468, + 195 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We plan to release the dataset publicly to promote further study of movie analysis. Due to copyright issues, trailers and movies will be released in the form of extracted features in visual and audio modalities.", + "bbox": [ + 75, + 196, + 470, + 257 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Approach", + "text_level": 1, + "bbox": [ + 76, + 268, + 184, + 284 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1. Overview", + "text_level": 1, + "bbox": [ + 76, + 292, + 186, + 306 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The overall architecture of our Collaborative noisy Label Cleaner (CLC) framework is illustrated in Fig. 3, which includes three modules: feature extraction, augmented crosspropagation (ACP), and multi-modal cleaning (MMC).", + "bbox": [ + 75, + 316, + 468, + 378 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In our framework, both the visual and audio modalities are utilized. For feature extraction and encoding, we first split the video $V$ into $T$ shots. We characterize the $i^{th}$ shots by two vectors, i.e., $\\mathbf{v}_i$ for the visual features, and $\\hat{\\mathbf{a}}_{\\mathbf{i}}$ for the audio features, where $i = 1,2,\\dots,T$ . These features are extracted using pre-trained visual [11] and audio feature extractors [24]. The parameters of the two extractors are frozen during training.", + "bbox": [ + 75, + 378, + 468, + 500 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The ACP and MMC are the core components of our framework. Since the visual-audio signals in videos are closely related but do not always contribute to highlight detection, the ACP module exploits the relationship via unimodal and cross-modal interactions. Then, it learns unified multi-modal representations for highlight detection. As the obtained highlight moments after scene segmentation are noisy, the MMC module firstly observes the changes in unimodal losses, then filters the noisy labels and utilizes the clean ones for multi-modal supervised learning. Next, we will introduce details for each component.", + "bbox": [ + 75, + 500, + 468, + 667 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.2. Augmented Cross-Propagation", + "text_level": 1, + "bbox": [ + 76, + 679, + 349, + 695 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Cross-Propagation. To predict the highlights, the model needs to understand the storylines of the movie. Meanwhile, visual and audio inputs do not always contribute to accurate prediction. Therefore, temporal modeling and modality interaction are the keys to achieving successful highlight detection. To achieve this goal, we design our ACP module to fuse the visual-audio modalities. It involves three steps in total.", + "bbox": [ + 75, + 703, + 468, + 821 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "First, in order to align the multi-modal features, we introduce $h$ which is fully-connected (FC) layer with ReLU to the $\\hat{\\mathbf{a}}_{i}$ such that it has same dimension as the $\\mathbf{v}_i$ .", + "bbox": [ + 75, + 825, + 468, + 869 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {a} _ {i} = h \\left(\\hat {\\mathbf {a}} _ {\\mathrm {i}}\\right). \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 230, + 885, + 468, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Movies are composed of consecutive audio-visual clips. Therefore, to measure the highlighted ness of a given shot, one must consider the relationship of the shot with its adjacent shots. The self-attention mechanism has shown effectiveness in capturing the long-term dependencies in previous works. We leverage self-attention to capture the temporal relationship for $\\mathbf{v}_i$ and $\\mathbf{a}_i$ via Eq. 2 and Eq. 3, respectively.", + "bbox": [ + 496, + 90, + 890, + 210 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {v} _ {i} ^ {s} = \\operatorname {s o f t m a x} \\left(\\frac {\\left(\\mathbf {v} _ {i} W _ {1} ^ {v}\\right) \\left(\\mathbf {v} W _ {2} ^ {v}\\right) ^ {\\top}}{\\sqrt {d}}\\right) \\left(\\mathbf {v} W _ {3} ^ {v}\\right), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 542, + 219, + 890, + 255 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {a} _ {i} ^ {s} = \\operatorname {s o f t m a x} \\left(\\frac {\\left(\\mathbf {a} _ {i} W _ {1} ^ {a}\\right) \\left(\\mathbf {a} W _ {2} ^ {a}\\right) ^ {\\top}}{\\sqrt {d}}\\right) \\left(\\mathbf {a} W _ {3} ^ {a}\\right), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 544, + 262, + 890, + 297 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathbf{v} = [\\mathbf{v}_1; \\mathbf{v}_2; \\ldots; \\mathbf{v}_T]$ and $\\mathbf{a} = [\\mathbf{a}_1; \\mathbf{a}_2; \\ldots; \\mathbf{a}_T]$ ; the scaling factor $d$ is equal to the visual/audio feature dimension and $(*)^{\\top}$ denotes the transpose operator; $W^{v}$ and $W^{a}$ are learnable matrices of two modalities, which are implemented by a linear layer. Uni-modal self-attention can well capture uni-modal temporal contexts and enhance clip features within the same modality.", + "bbox": [ + 496, + 300, + 890, + 406 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Despite the above self-attention capturing the clip interactions within the uni-modality, it is critical to capture the interactions across modalities. To capture semantic associations based on multi-modal signals, we introduce cross-attention to update the features of each modality.", + "bbox": [ + 496, + 407, + 890, + 482 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {v} _ {i} ^ {c} = \\operatorname {s o f t m a x} \\left(\\frac {\\left(\\mathbf {v} _ {i} W _ {4} ^ {v}\\right) \\left(\\mathbf {a} W _ {4} ^ {a}\\right) ^ {\\top}}{\\sqrt {d}}\\right) \\left(\\mathbf {a} W _ {5} ^ {a}\\right), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 542, + 489, + 890, + 525 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {a} _ {i} ^ {c} = \\operatorname {s o f t m a x} \\left(\\frac {\\left(\\mathbf {a} _ {i} W _ {6} ^ {a}\\right) \\left(\\mathbf {v} W _ {5} ^ {v}\\right) ^ {\\top}}{\\sqrt {d}}\\right) \\left(\\mathbf {v} W _ {6} ^ {v}\\right), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 542, + 532, + 890, + 568 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Through cross-attention, the information from two modalities are connected. However, considering audiovisual temporal asynchrony, it is necessary to select effective information from multi-modality. We augment the relevant positive connections and dampen the irrelevant connections. The strength of these connections is measured by the cross-correlation matrix, computed by,", + "bbox": [ + 496, + 571, + 890, + 676 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {c} ^ {v} = R e L U \\left(\\frac {\\mathbf {v a} ^ {\\top}}{\\sqrt {d}}\\right), \\mathbf {c} ^ {a} = R e L U \\left(\\frac {\\mathbf {a v} ^ {\\top}}{\\sqrt {d}}\\right). \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 542, + 684, + 890, + 720 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For each modality, the cross-correlation matrix $\\mathbf{c}$ is used to re-weight the cross-attention features. Finally, we obtain updated visual features $\\bar{\\mathbf{v}}_{\\mathrm{i}}$ and audio features $\\bar{\\mathbf{a}}_{\\mathrm{i}}$ by fusion of the original features, enhanced uni-modal features, and cross-modal features.", + "bbox": [ + 496, + 726, + 890, + 801 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\bar {\\mathbf {v}} _ {\\mathbf {i}} = f \\left(\\mathbf {v} _ {i}, \\mathbf {v} _ {i} ^ {s}, \\left(\\sum_ {j = 1} ^ {T} \\mathbf {c} _ {i j} ^ {v}\\right) * \\mathbf {v} _ {i} ^ {c}\\right), \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 591, + 808, + 890, + 849 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\overline {{\\mathbf {a}}} _ {\\mathrm {i}} = f \\left(\\mathbf {a} _ {i}, \\mathbf {a} _ {i} ^ {s}, \\left(\\sum_ {j = 1} ^ {T} \\mathbf {c} _ {i j} ^ {a}\\right) * \\mathbf {a} _ {i} ^ {c}\\right), \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 593, + 861, + 890, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "18901", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/8d3f2c2c28d1eaf9ba040dba693a9e53fc538b71c299fb54f14b73a8cd5f1b5a.jpg", + "image_caption": [ + "Figure 3. Overview of the proposed CLC. It includes three modules: feature extraction, augmented cross-propagation (ACP), and multi-modal cleaning (MMC). The visual and audio modalities of the input video are represented as vectors by the feature extraction module. Then the features are augmented by ACP module to capture semantic associations across modalities. MMC is used to filter outs noisy and incomplete labeling with additional uni-modal branches. During inference, we remove the uni-modal branches and only rely on the prediction of multi-modal branch. More details of the CLC are shown in Sec. 4.1" + ], + "image_footnote": [], + "bbox": [ + 83, + 101, + 883, + 342 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $f$ is the fusion function consisting of FC layers and ReLU to further project the features.", + "bbox": [ + 76, + 438, + 470, + 468 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Clearly, the cross-correlation matrix will assign large weights to clips that are relevant to the other modality. In addition, the ReLU activation in Eq. 6 cuts off connections with negative similarity values and only relevant positive connections would be preserved. By the above operations, the original feature will be infused with richer information. Consistency Loss. Multi-modal inputs help to comprehensively learn by integrating different aspects and boosting model performance. However, they are not fully exploited because some modality-specific features may be weakened in the fusion even when the multi-modal model outperforms its uni-modal counterpart. In this work, we provide parallel branches for each modality separately to obtain uni-modal prediction during training. As seen in Fig. 3, all branches share the same clip feature extracted from ACP and the feature is fed into different branches independently. In each branch, we develop a temporal model $G$ to obtain its prediction score $y_{i}$ of the $i^{th}$ clip being a highlight as follows:", + "bbox": [ + 75, + 469, + 468, + 741 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {y} _ {i} ^ {M M} = G \\left(\\bar {\\mathbf {v}} _ {\\mathbf {i}}, \\bar {\\mathbf {a}} _ {\\mathbf {i}}\\right), \\mathbf {y} _ {i} ^ {U M _ {v}} = G \\left(\\bar {\\mathbf {v}} _ {\\mathbf {i}}\\right), \\mathbf {y} _ {i} ^ {U M _ {a}} = G \\left(\\bar {\\mathbf {a}} _ {\\mathbf {i}}\\right). \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 86, + 750, + 468, + 768 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $MM$ is multi-modal branch; $UM_{v}$ is visual branch and $UM_{a}$ is audio branch.", + "bbox": [ + 76, + 779, + 468, + 808 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "However, the gap between the different modalities would bring instability in the joint optimization process. We employ the auxiliary consistency loss to guarantee consistency between the different modalities. Given the multi-modal features of a clip, they are consistent if they share the same prediction. Specifically, the consistency loss is defined as", + "bbox": [ + 75, + 810, + 468, + 898 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "the cross-entropy between the multi-modal prediction probability $\\mathbf{y}_i^{MM}$ and uni-modal prediction probability $\\mathbf{y}_i^{UM}$ :", + "bbox": [ + 498, + 438, + 890, + 470 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {c o n s}} = - \\left(\\sum_ {i = 1} ^ {N} \\mathbf {y} _ {i} ^ {U M _ {v}} \\log \\mathbf {y} _ {i} ^ {M M} + \\sum_ {i = 1} ^ {N} \\mathbf {y} _ {i} ^ {U M _ {a}} \\log \\mathbf {y} _ {i} ^ {M M}\\right), \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 509, + 476, + 890, + 511 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $N$ is the number of samples in a batch.", + "bbox": [ + 500, + 517, + 802, + 532 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "This consistency loss not only implicitly enhances uni-modal information, but also explicitly guides the multimodal branch to robust supervision. The added uni-modal branches are only utilized in the training phase and are disabled during the inference stage.", + "bbox": [ + 496, + 532, + 890, + 609 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.3. Multi-modal Cleaning", + "text_level": 1, + "bbox": [ + 500, + 619, + 707, + 635 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Multi-modality Sample Cleaning. Annotation noise is inevitable in VHD due to subjectivity depending on the users and annotators. In this paper, we argue that highlight detection should be regarded as Learning with Noisy Labels. To alleviate the performance drop caused by noisy labels, we adopt a multi-modality collaborative cleaning to adaptively filter noisy samples with noisy modality information.", + "bbox": [ + 496, + 643, + 890, + 748 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Firstly, we maintain multiple outputs simultaneously which are predicted by different branches. The uni-modal branches independently select clean samples based on the low-loss criterion in which instances with lower losses are treated clean samples. Contrary to existing noisy sample selection methods [12, 54], which directly discard high-loss samples, we keep all samples to train the uni-modal branches. Then, we update the multi-modal branch using only clean samples selected by both uni-modal branches in the back-propagation. The samples for the multi-modal", + "bbox": [ + 496, + 750, + 892, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "18902", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "branch training are selected dynamically while all samples participate in the uni-mode branches training. In this way, MMC obtains information on all samples to avoid the model defecting to favoring easy samples. More details of MMC are presented in the supplementary material.", + "bbox": [ + 75, + 90, + 468, + 166 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Each branch has its own loss function. For the uni-modal branch, we employ the cross-entropy loss with all samples as follows:", + "bbox": [ + 75, + 167, + 468, + 210 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} ^ {U M _ {v}} = \\mathcal {L} _ {\\mathrm {c e}} ^ {U M _ {v}} = - \\sum_ {i = 1} ^ {N} \\mathbf {g} _ {i} \\log \\mathbf {y} _ {i} ^ {U M _ {v}}, \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 142, + 219, + 468, + 258 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} ^ {U M _ {a}} = \\mathcal {L} _ {\\mathrm {c e}} ^ {U M _ {a}} = - \\sum_ {i = 1} ^ {N} \\mathbf {g} _ {i} \\log \\mathbf {y} _ {i} ^ {U M _ {a}}. \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 142, + 270, + 468, + 311 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For the multi-modal branch, its cross-entropy loss and consistency loss are updated with a re-weighting scheme with clean samples.", + "bbox": [ + 76, + 315, + 468, + 361 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {c e}} ^ {\\mathrm {M M}} = - \\sum_ {i = 1} ^ {N ^ {\\prime}} \\mathbf {g} _ {i} \\log \\mathbf {y} _ {i} ^ {M M}, \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 178, + 371, + 468, + 412 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {c o n s}} ^ {M M} = - \\left(\\sum_ {i = 1} ^ {N ^ {\\prime}} \\mathbf {y} _ {i} ^ {U M _ {v}} \\log \\mathbf {y} _ {i} ^ {M M} + \\sum_ {i = 1} ^ {N ^ {\\prime}} \\mathbf {y} _ {i} ^ {U M _ {a}} \\log \\mathbf {y} _ {i} ^ {M M}\\right), \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 84, + 422, + 468, + 460 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} ^ {\\mathrm {M M}} = \\mathcal {L} _ {\\mathrm {c e}} ^ {\\mathrm {M M}} + \\beta \\mathcal {L} _ {\\mathrm {c o n s}} ^ {M M}, \\tag {15}\n$$\n", + "text_format": "latex", + "bbox": [ + 187, + 462, + 468, + 481 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\mathbf{g}_i$ and $\\mathbf{y}_i^*$ denote the ground-truth and predicted probability of the $i^{th}$ clip, respectively; $N$ is the number of samples in a batch and $N'$ is the number of samples selected by MMC; and $\\beta$ are designed to balance different loss terms.", + "bbox": [ + 75, + 486, + 468, + 559 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Post processing. In experiments we observe that noise makes jitter prediction curves along the temporal dimension, which may cause discontinuous thresholds for highlight selection. Therefore, we apply a median filter to smooth the prediction curves. Supposing $\\mathbf{y} = [\\mathbf{y}_1; \\mathbf{y}_2; \\dots; \\mathbf{y}_T]$ is the original curve predicted by the CLC, the smoothed curve $\\mathbf{s} = [\\mathbf{s}_1; \\mathbf{s}_2; \\dots; \\mathbf{s}_T]$ is given by:", + "bbox": [ + 75, + 561, + 468, + 667 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {s} _ {i} = \\left\\{ \\begin{array}{l l} \\operatorname {M e d} \\left(\\mathbf {y} _ {i - k}, \\mathbf {y} _ {i + k}\\right), & k < i \\leq T - k \\\\ \\mathbf {y} _ {i}, & \\text {o t h e r w i s e} \\end{array} \\right. \\tag {16}\n$$\n", + "text_format": "latex", + "bbox": [ + 109, + 676, + 468, + 718 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $k$ is the window size, and \"Med\" denotes the median filter.", + "bbox": [ + 75, + 726, + 468, + 753 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Experiment", + "text_level": 1, + "bbox": [ + 76, + 762, + 200, + 777 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1. Datasets and Experimental Settings", + "text_level": 1, + "bbox": [ + 76, + 786, + 385, + 801 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Datasets. We evaluate our CLC on the constructed dataset MovieLights and public YouTube Highlights dataset [37]. MovieLights is split into training and testing sets, each containing 144 and 30 movies respectively.", + "bbox": [ + 75, + 809, + 468, + 869 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We split movies into shots. The movie features are represented at shot-level. We use the middle frame of each", + "bbox": [ + 75, + 869, + 468, + 898 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "shot to extract its visual feature with ViT [10] pre-trained by CLIP [33]. We align timestamps of audio clips with the visual shots, and sample the audio clip of each with $16K$ Hz sampling rate and 512 windowed signal length. The resulted shot-level audio features are obtained with the PANN audio network [24] pretrained on AudioSet [13].", + "bbox": [ + 496, + 90, + 890, + 181 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The YouTube Highlights contains six distinct categories with a total 422 videos currently available. Following the practice of prior efforts, we train a highlight detector for each category. YouTube Highlights provides two annotations: Harvested Highlight and Mturk Highlight. In the Harvested annotation, the match label specifies if each clip is matched in the edited video, where 1 denotes matched, -1 denotes unmatched and 0 denotes the borderline cases. In the Mturk annotation, the highlight labels are marked by multiple turkers of different styles, making it noisier than the Harvested annotation.", + "bbox": [ + 496, + 183, + 890, + 348 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "On YouTube Highlights, we use the same protocol and data preprocessing as [29]. It obtain clip-level visual features and optical flow features using I3D [6] pre-trained on Kinetics400 [23]. It use a PANN audio network [24] pretrained on AudioSet [13] to obtain audio features that align with the visual clips. Frame-level features are average-pooled within each clip for both audio and visual features to generate a clip-level feature. Since each feature vector spans 32 consecutive frames, we follow [29] and consider the feature vector corresponded to a clip if their overlap is more than $50\\%$ .", + "bbox": [ + 496, + 352, + 890, + 517 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Benchmarks. To better inspect the robustness of our CLC against noisy labels, we also apply label perturbations in training set of YouTube Highlights while keeping the ground-truths in the testing set unchanged. YouTube Highlights has two benchmarks for comparison. 1) Harvested with matched: we regard the clips labeled with matched as the highlighted clips and this is the same setting as in previous works [29, 50]. 2) Harvested with matched and borderline: clips labeled with matched and borderline are treated as the highlight moments. In this benchmark, the training set contains some highlighted clips with weak confidence. We select the clips whose mturk-label is over score 1 as the highlighted clips, which means that at least one turker selects the clip as a highlight. There are labeled by different types of annotators between Mturk and the test set and bring greater noise. We select the clips whose mturk-label is over score 1 as the highlighted clips, which means that at least one turker selects the clip as a highlight. Due the labeling gap between the Mturk annotation in the training set and the clean testing set, this benchmark is even noisier.", + "bbox": [ + 496, + 520, + 890, + 821 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Baselines. We introduce CLC-, the degenerated version of CLC, as a baseline. Similar to CLC, CLC- is a Bi-LSTM-based model and takes the temporal sequence of shot features as input but lacks the modules of augmented cross-propagation and multi-modality noisy label cleaner.", + "bbox": [ + 496, + 824, + 890, + 898 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "18903", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/2640eb5038901fc2617578ee8e6c3106976f0f26758fc1c48f77a3649dfbb8a1.jpg", + "table_caption": [ + "Table 2. Results on MovieLights." + ], + "table_footnote": [], + "table_body": "
MethodsModalitymAP
GIFs [14]V25.48
SL-Module [50]V32.34
SL-Module [50]VA34.27
UMT [29]VA38.7
CLC-VA39.65
CLC- w/ SCE [44]VA39.83
CLC- w/ LS [40]VA40.49
CLCVA43.88
", + "bbox": [ + 114, + 103, + 433, + 239 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Evaluation Metric. We adopt mean Average Precision (mAP) as the evaluation metric for MovieLights and YouTube Highlights. Considering that a highlighted moment in one video is not necessarily more interesting than non-highlight moments in other videos, we evaluate on each test video independently and report the averaged results.", + "bbox": [ + 75, + 263, + 468, + 354 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Implementation Details. On the MovieLights, we train our model using SGD, with a learning rate of 0.01. We train for 50 epochs. Before the cross-attention modules, we project each modality into a vector of 512 dimension. The key, query, and value vectors all share the same dimension. Weight $\\beta$ in Eq. 15 is empirically set to 0.1 and window size $k$ in Eq. 16 is set to 9.", + "bbox": [ + 75, + 354, + 468, + 460 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.2. Results on MovieLights", + "text_level": 1, + "bbox": [ + 76, + 470, + 294, + 487 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "On MovieLights, we train our model with the noisy pseudo labels. To compare with previous state-of-the-art highlight detection works, we train UMT [29] and SL-Module [50] using the same protocol and data preprocessing as in CLC. We also compare with Video2GIF [14] using its off-the-shelf tool1. The upper part of Tab. 2 illustrates the significant performance gain of CLC.", + "bbox": [ + 75, + 494, + 468, + 599 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To demonstrate the advantages of CLC in learning with noisy labels, we make comparisons with two main-stream label noise approaches: Label Smoothing [40] and SCE loss [44]. Specifically, we insert Label Smoothing or SCE into our CLC- framework to create two baseline VHD methods to tackle label noise. The bottom part of Tab. 2 shows that CLC outperforms the two baseline methods by a notable margin, indicating that our augmented cross-propagation and multi-modality noisy label cleaner are more effective than vanilla label noise approaches in VHD tasks.", + "bbox": [ + 75, + 599, + 470, + 765 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.3. Ablation Study", + "text_level": 1, + "bbox": [ + 76, + 776, + 230, + 792 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In this experiment, we analyze the impact of each module. The results are summarized in Tab. 3.", + "bbox": [ + 75, + 800, + 468, + 829 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Multi-modality Sample Cleaning. We first inspect the impact of the multi-modality sample filter module because we are primary concerned with learning with noisy labels", + "bbox": [ + 75, + 830, + 468, + 876 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/e99e3d7373ea63ca6643dd2458f8ea84b24bd15ff0455dffdf1675ca23628d9d.jpg", + "table_caption": [ + "Table 3. Ablation results of MoiveLightts." + ], + "table_footnote": [], + "table_body": "
MMSCCPCLPPmAP
××××39.65
✓×××41.69
✓✓××42.79
✓✓✓×43.22
✓✓✓✓43.88
", + "bbox": [ + 535, + 103, + 857, + 196 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/af631e9394c246580adc4784f623a084b52e73b65d9b6aa0b3f266905ba70e43.jpg", + "table_caption": [ + "Table 4. Results on YouTube Highlights." + ], + "table_footnote": [], + "table_body": "
Methodsdoggym.park.ska.ski.surf.Avg.
GIFs [14]30.833.55455.432.854.146.4
LSVM [37]60.041.061.062.036.061.053.6
HighlightMe [5]63737264526264
MINI-Net [16]58.261.770.272.258.760.164.4
CHD [1]60.671.174.249.868.268.565.4
Trail [43]63.382.562.352.974.579.369.1
SL-Module [50]70.853.277.272.566.176.269.3
Joint-VA [2]64.571.980.86273.278.371.8
PLD [45]74.970.277.957.570.77973
CO-AV [27]60.9668974.16981.174.7
UMT [29]65.975.281.671.872.382.774.9
CLC(ours)70.579.483.983.579.583.680.1
", + "bbox": [ + 501, + 224, + 897, + 414 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "in VHD. The $2\\%$ performance gain from the module over baseline shows the importance of filtering noisy sample. Based on the this observation, the subsequent ablation experiments are conducted under the setting of noise filtering. Cross-Propagation. We then evaluate the impact of the feature augmentation module. Compared to naive feature concatenation, the models with augmented features show superior performance. The results validate that feature augmentation module can better explore the complementary information from different modalities and suppress the mutual disturbance of desynchronized uni-modal information.", + "bbox": [ + 496, + 439, + 890, + 606 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Consistency Loss. We examine the contribution of the consistency loss. The results demonstrate that the employment of the auxiliary multi-modal constraint further increases the model robustness. This consistency loss not only implicitly enhances uni-modal information, but also explicitly guides the multi-modal branch to better learning.", + "bbox": [ + 496, + 607, + 890, + 698 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Post Processing. We compare the highlight prediction curves with and without the post processing median filter in Fig. 4(a,b), which prevents disruptive prediction variation and improves the overall performance.", + "bbox": [ + 496, + 699, + 890, + 758 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.4. Results on YouTube Highlights", + "text_level": 1, + "bbox": [ + 500, + 770, + 772, + 786 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We conduct experiments on the public video highlight detection benchmarks YouTube Highlights, including six domain video datasets, to verify the generalization ability of CLC.", + "bbox": [ + 496, + 792, + 890, + 853 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The setting of Harvested Highlight with matched is consistent with the previous work [29]. As shown in Tab. 4, CLC achieves state-of-the-art performance on the YouTube", + "bbox": [ + 496, + 854, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_footnote", + "text": "1https://github.com/gyglim/video2gif_code", + "bbox": [ + 94, + 886, + 323, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "18904", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/41c5bd013edb20170b65a2247d3c6c90f2c95d9a7e47fd267d16acb828ca2197.jpg", + "image_caption": [ + "Figure 4. (a, b) Comparison of original prediction curve with filtered prediction curve. (c) Results on YouTube Highlights with noisy label." + ], + "image_footnote": [], + "bbox": [ + 81, + 88, + 346, + 250 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/08a751add36e3c82c96298e24a9351831f506cc8c00ebff9c438a4a59c78f4b9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 348, + 88, + 612, + 250 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/7b8c3bf7b26235ea5d7dec63ccf8e99eb352033c68bd6c1fa05cd3bc0aa7317c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 614, + 89, + 890, + 244 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/4558dd337bc734db059b19638b79f66ec78f5787ec839649a3def075860ccac5.jpg", + "table_caption": [ + "Table 5. Results on YouTube Highlights with Noisy Label." + ], + "table_footnote": [], + "table_body": "
AnnotationNoiseMethodsdoggym.park.ska.ski.surf.Avg.
Harvested matchedcleanUMT [29]65.9075.2081.6071.8072.3082.7074.90
Harvested matchedcleanCLC70.5179.4383.8583.5179.4683.5680.05 (↑ 5.15)
Harvested borderlineslight noiseUMT [29]65.9374.3181.5871.8470.2482.4674.39
Harvested borderlineslight noiseCLC69.4180.7378.5085.3681.1183.1679.71 (↑ 5.32)
Mturksevere noiseUMT [29]63.7876.1675.0273.6269.9981.5973.36
Mturksevere noiseCLC66.9280.4485.9282.3378.0581.7279.22 (↑ 5.86)
", + "bbox": [ + 114, + 292, + 854, + 401 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/6edce3294648c6dd7f15fb78235913315f4342ffd14e1622ebcd97ae6035f2fd.jpg", + "image_caption": [ + "Figure 5. The highlight moments selected by our CLC from Skyfall and Sherlock Holmes. Top: Bond gives chase to a professional hitman by car to find a classified hard drive, and then a firefight erupted in the market. Bottom: Holmes and Blackwood are facing off, and then Blackwood reaches for a weapon to kill Holmes, but accidentally trips off a scaffolding and falls to his death." + ], + "image_footnote": [], + "bbox": [ + 84, + 415, + 883, + 529 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Highlights, outperforming the existing multi-modal highlight detection methods in the average metric across all categories. Specifically, CLC achieves best performance in three out of the six categories, while maintaining reasonably competitive performance in the other three categories. These results support our claim that highlight detection should be regarded as learning with noisy labels.", + "bbox": [ + 75, + 590, + 468, + 696 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To quantify how CLC is robust to different levels of label noise, we inspect the settings of Harvested Highlight with matched and borderline and Mturk Highlight, which are perturbed by varying degrees of label noise in YouTube Highlights.", + "bbox": [ + 75, + 707, + 468, + 784 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Tab. 5 exhibits the performances of CLC and UMT [29] at different noise levels. As the noise level increases, the VHD task becomes more difficult, but the performance superiority of our CLC over UMT becomes even more obvious. It is illustrated in Fig. 4(c) that compared with the most recent state-of-the-art UMT, our model can achieve better performance even when disturbed by severe label noise.", + "bbox": [ + 75, + 794, + 468, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.5. Visualization", + "text_level": 1, + "bbox": [ + 500, + 589, + 635, + 604 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "As shown in Fig. 5, we present some visualization examples of the detected highlight clips in MovieLights by CLC. The examples clearly show that the prediction of CLC is in accordance with user expectation. We will provide more examples in the supplementary material.", + "bbox": [ + 496, + 614, + 890, + 690 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 700, + 617, + 715 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this study, we present Collaborative noisy Label Cleaner (CLC), a novel framework to handle noisy labels in video highlight detection. We make use of the augmented cross-propagation module to better enhance network robustness and multi-modality cleaning to achieve cleaner highlight labels by observing the loss changes of different modalities. We demonstrate the state-of-the-art performance of our method with extensive experiments on MovieLights and YouTube Highlights datasets. In future work, we are interested in extending the proposed mechanisms to other video-understanding tasks such as scene segmentation and video temporal grounding.", + "bbox": [ + 496, + 727, + 890, + 909 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "18905", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Taivanbat Badamdorj, Mrigank Rochan, Yang Wang, and Li Cheng. Contrastive learning for unsupervised video highlight detection. In CVPR, 2022. 7", + "[2] Taivanbat Badamdorj, Mrigank Rochan, Yang Wang, Jiaze Wang, and Li Cheng. Joint visual and audio learning for video highlight detection. In ICCV, 2021. 2, 7", + "[3] Max Bain, Arsha Nagrani, Andrew Brown, and Andrew Zisserman. Condensed movies: Story based retrieval with contextual embeddings. In ACCV, 2020. 3", + "[4] David Bamman, Brendan O'Connor Noah, and A. Smith. Learning latent personas of film characters. 2014. 2", + "[5] Uttaran Bhattacharya, Gang Wu, Stefano Petrangeli, Viswanathan Swaminathan, and Dinesh Manocha. Highlighte: Detecting highlights from human-centric videos. In ICCV, 2021. 7", + "[6] Joao Carreira and Andrew Zisserman. Quo vadis, action recognition? a new model and the kinetics dataset. 2017. 6", + "[7] Shixing Chen, Xiaohan Nie, David Fan, Dongqing Zhang, Vimal Bhat, and Raffay Hamid. Shot contrastive self-supervised learning for scene boundary detection. 2021. 2", + "[8] Hao Cheng, Zhaowei Zhu, Xingyu Li, Yifei Gong, Xing Sun, and Yang Liu. Learning with instance-dependent label noise: A sample sieve approach. In ICLR, 2021. 3", + "[9] Yashar Deldjoo, Mihai Gabriel Constantin, and Bogdan Ionescu. Mmtf-14k: A multifaceted movie trailer feature dataset for recommendation and retrieval. In ACMMM, 2018. 3", + "[10] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Xiaohua Zhai Dirk Weissenborn, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. 2021. 6", + "[11] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. 2021. 4", + "[12] Curtis G, Northcutt, Lu Jiang, and Isaac L. Chuang. Confident learning: Estimating uncertainty in dataset labels. 2021. 5", + "[13] Jort F. Gemmeke, Daniel P. W. Ellis, Dylan Freedman, Aren Jansen, Wade Lawrence, R. Channing Moore, Manoj Plakal, and Marvin Ritter. Audio set: An ontology and human-labeled dataset for audio events. 2017. 6", + "[14] Michael Gygli, Yale Song, and Liangliang Cao. Video2gif: Automatic generation of animated gifs from video. In CVPR, 2016. 2, 7", + "[15] Bo Han, Quanming Yao, Xingrui Yu, Gang Niu, Miao Xu, Weihua Hu, Ivor Tsang, and Masashi Sugiyama. Co-teaching: Robust training of deep neural networks with extremely noisy labels. In NeurIPS, 2018. 3", + "[16] Fa-Ting Hong, Xuanteng Huang, Wei-Hong Li, and WeiShi Zheng. Mini-net: Multiple instance ranking network for video highlight detection. In ECCV, 2020. 2, 7" + ], + "bbox": [ + 78, + 114, + 468, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[17] Jinchi Huang, Lie Qu, Rongfei Jia, and Binqiang Zhao. O2unet: A simple noisy label detection approach for deep neural network. In ICCV, 2019. 3", + "[18] Qingqiu Huang, Yu Xiong, Anyi Rao, Jiaze Wang, and Dahua Lin. Movienet: A holistic dataset for movie understanding. In ECCV, 2020. 2, 3", + "[19] Qingqiu Huang, Yuanjun Xiong, Yu Xiong, Yuqi Zhang, and Dahua Lin. From trailers to storylines: An efficient way to learn from movies. arXiv preprint arXiv:1806.05341, 2018. 3", + "[20] Lu Jiang, Zhengyuan Zhou, Thomas Leung, Li-Jia Li, and Li Fei-Fei. Mentornet: Learning data-driven curriculum for very deep neural networks on corrupted labels. In ICML, 2018. 3", + "[21] Yifan Jiao, Zhetao Li, Shucheng Huang, Xiaoshan Yang, Bin Liu, and Tianzhu Zhang. Three-dimensional attention-based deep ranking model for video highlight detection. In IEEE Transactions on Multimedia, page 2693-2705, 2018. 2", + "[22] Jeff Johnson, Matthijs Douze, and Hervé Jégou. Billionscale similarity search with gpus. IEEE Transactions on Big Data, page 1-1, 2019. 3", + "[23] Will Kay, Joao Carreira, Karen Simonyan, Brian Zhang, Chloe Hillier, Sudheendra Vijayanarasimhan, Fabio Viola, Tim Green, Trevor Back, Paul Natsev, Mustafa Suleyman, and Andrew Zisserman. The kinetics human action video dataset. arXiv preprint arXiv:1705.06950, 2017. 6", + "[24] Qiuqiang Kong, Yin Cao, Turab Iqbal, Yuxuan Wang, Wenwu Wang, and Mark D Plumbley. Panns: Large-scale pretrained audio neural networks for audio pattern recognition. IEEE/ACM Transactions on Audio, Speech, and Language Processing, 28:2880-2894, 2020. 4, 6", + "[25] Jie Lei, Tamara L. Berg, and Mohit Bansal. Qvhighlights: Detecting moments and highlights in videos via natural language queries. In NeurIPS, 2021. 2", + "[26] Junnan Li, Richard Socher, and Steven C.H. Hoi. Dividemix: Learning with noisy labels as semi-supervised learning. 2020. 3", + "[27] Shuaicheng Li, Feng Zhang, Kunlin Yang, Lingbo Liu, Shinan Liu, Jun Hou, and Shuai Yi. Probing visual-audio representation for video highlight detection via hard-pairs guided contrastive learning. arXiv preprint arXiv:2206.10157, 2022. 7", + "[28] Yang Liu and Hongyi Guo. Peer loss functions: Learning from noisy labels without knowing noise rates. 2020. 3", + "[29] Ye Liu, Siyuan Li, Yang Wu, Chang Wen Chen, Ying Shan, and Xiaohu Qie. Umt: Unified multi-modal transformers for joint video moment retrieval and highlight detection. In CVPR, 2022. 2, 6, 7, 8", + "[30] Xingjun Ma, Hanxun Huang, Yisen Wang, Simone Romano, Sarah Erfani, and James Bailey. Normalized loss functions for deep learning with noisy labels. 2020. 3", + "[31] Curtis G. Northcutt, Lu Jiang, and Isaac L. Chuang. Confident learning: Estimating uncertainty in dataset labels. In JAIR, 2019. 3", + "[32] Sinno Jialin Pan and Qiang Yang. A survey on transfer learning. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, page 1345-1359, 2010. 2" + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "18906", + "bbox": [ + 480, + 945, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[33] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Sandhini Agarwal Gabriel Goh, Girish Sastry, Amanda Askell, Pamela Mishkin, and et al. Jack Clark. Learning transferable visual models from natural language supervision. arXiv preprint arXiv:2103.00020, 2021. 6", + "[34] Anyi Rao, Linning Xu, Yu Xiong, Guodong Xu, Qingqiu Huang, Bolei Zhou, and Dahua Lin. A local-to-global approach to multi-modal movie scene segmentation. 2020. 2", + "[35] Yale Song, Yahoo Labs, New York, Jordi Vallmitjana, Amanda Stent, and Alejandro Jaime. Tvsum: Summarizing web videos using titles. In CVPR, 2015. 2", + "[36] Tomáš Souček and Jakub Lokoc. Transnet v2: An effective deep network architecture for fast shot transition detection. arXiv preprint arXiv:2008.04838, 2020. 3", + "[37] Min Sun, Ali Farhadi, and Steve Seitz. Ranking domain-specific highlights by analyzing edited videos. In ECCV, 2014. 2, 6, 7", + "[38] Zeren Sun, Fumin Shen, Dan Huang, Qiong Wang, Xiangbo Shu, Yazhou Yao, and Jinhui Tang. Pnp: Robust learning from noisy labels by probabilistic noise prediction. In CVPR, 2022. 3", + "[39] Zeren Sun, Yazhou Yao, Xiu-Shen Wei, Yongshun Zhang, Fumin Shen, Jianxin Wu, Jian Zhang, and Heng-Tao Shen. Webly supervised fine-grained recognition-benchmark datasets and an approach. In ICCV, 2021. 3", + "[40] Christian Szegedy, Vincent Vanhoucke, Sergey Ioffe, Jonathon Shlens, and Zbigniew Wojna. Rethinking the inception architecture for computer vision. 2016. 7", + "[41] Hao Tang, Vivek Kwatra, Mehmet Emre Sargin, and Ullas Gargi. Detecting highlights in sports videos: Cricket as a test case. In ICME, 2011. 2", + "[42] Paul Vicol, Makarand Tapaswi, Lluis Castrejon, and Sanja Fidler. Towards understanding human-centric situations from videos. 2018. 2", + "[43] Lezi Wang, Dong Liu, Rohit Puri, and Dimitris N. Metaxas. Learning trailer moments in full-length movies with co-contrastive attention. In ECCV, 2020. 2, 3, 7", + "[44] Yisen Wang, Xingjun Ma, Zaiyi Chen, Yuan Luo, Jinfeng Yi, and James Bailey. Symmetric cross entropy for robust learning with noisy labels. 2019. 3, 7", + "[45] Fanyue Wei, Biao Wang, Tiezheng Ge, Yuning Jiang, Wen Li, and Lixin Duan. Learning pixel-level distinctions for video highlight detection. In CVPR, 2022. 2, 7", + "[46] Hongxin Wei, Lei Feng, Xiangyu Chen, and Bo An. Combating noisy labels by agreement: A joint training method with co-regularization. In CVPR, 2020. 3", + "[47] Haoqian Wu, Keyu Chen, Yanan Luo, Ruizhi Qiao, Bo Ren, Haozhe Liu, Weicheng Xie, and Linlin Shen. Scene consistency representation learning for video scene segmentation. In CVPR, 2022. 2, 3", + "[48] Xiaobo Xia, Tongliang Liu, Bo Han, Nannan Wang, Mingming Gong, Haifeng Liu, Gang Niu, Dacheng Tao, and Masashi Sugiyama. Part-dependent label noise: Towards instance-dependent label noise. In NeurIPS, 2020. 3", + "[49] Xiaobo Xia, Tongliang Liu, Nannan Wang, Bo Han, Chen Gong, Gang Niu, and Masashi Sugiyama. Are anchor points really indispensable in label-noise learning? In NeurIPS, 2019. 3" + ], + "bbox": [ + 78, + 90, + 468, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[50] Minghao Xu, Hang Wang, Bingbing Ni, Riheng Zhu, Zhenbang Sun, and Changhu Wang. Cross-category video highlight detection via set-based learning. In ICCV, 2021. 2, 6, 7", + "[51] Fengxiang Yang, Ke Li, Zhun Zhong, Zhiming Luo, Xing Sun, Hao Cheng, Xiaowei Guo, Feiyue Huang, Rongrong Ji, and Shaozi Li. Asymmetric co-teaching for unsupervised cross domain person re-identification. In AAAI, 2020. 3", + "[52] Shuo Yang, Erkun Yang, Bo Han, Yang Liu, Min Xu, Gang Niu, and Tongliang Liu. Estimating instance-dependent bayes-label transition matrix using a deep neural network. In CVPR, 2022. 3", + "[53] Ting Yao, Tao Mei, and Yong Rui. Highlight detection with pairwise deep ranking for first-person video summarization. In CVPR, 2016. 2", + "[54] Chuanyi Zhang, Yazhou Yao*, Xing Xu, Jie Shao, Jingkuan Song, Zechao Li, and Zhenmin Tang. Extracting useful knowledge from noisy web images via data purification for fine-grained recognition. In ACMMM, 2021. 5", + "[55] Zhilu Zhang and Mert R. Sabuncu. Generalized cross entropy loss for training deep neural networks with noisy labels. In NeurIPS, 2018. 3", + "[56] Guoqing Zheng, Ahmed Hassan Awadallah, and Susan Dumas. Meta label correction for noisy label learning. In AAAI, 2021. 3", + "[57] A. Zlatintsi, P. Koutras, N. Efthymiou, P. Maragos, A. Potamianos, and K. Pastra. Quality evaluation of computational models for movie summarization. 2015. 2" + ], + "bbox": [ + 501, + 92, + 890, + 487 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "18907", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2023/Collaborative Noisy Label Cleaner_ Learning Scene-Aware Trailers for Multi-Modal Highlight Detection in Movies/b5b9b0f2-713f-4176-89e7-acec845e8cbb_model.json b/2023/Collaborative Noisy Label Cleaner_ Learning Scene-Aware Trailers for Multi-Modal Highlight Detection in Movies/b5b9b0f2-713f-4176-89e7-acec845e8cbb_model.json new file mode 100644 index 0000000000000000000000000000000000000000..d703dfd42f951c0da66580a5fdb2b5c055a9d3ce --- /dev/null +++ b/2023/Collaborative Noisy Label Cleaner_ Learning Scene-Aware Trailers for Multi-Modal Highlight Detection in Movies/b5b9b0f2-713f-4176-89e7-acec845e8cbb_model.json @@ -0,0 +1,2266 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.13, + 0.131, + 0.842, + 0.177 + ], + "angle": 0, + "content": "Collaborative Noisy Label Cleaner: Learning Scene-aware Trailers for Multi-modal Highlight Detection in Movies" + }, + { + "type": "text", + "bbox": [ + 0.115, + 0.204, + 0.856, + 0.26 + ], + "angle": 0, + "content": "Bei Gan Xiujun Shu* Ruizhi Qiao* Haoqian Wu Keyu Chen Hanjun Li Bo Ren Tencent YouTu Lab {stylegan, xiujunshu, ruizhiqiao, linuswu, yolochen, hanjunli, timren}@tencent.com" + }, + { + "type": "image", + "bbox": [ + 0.122, + 0.285, + 0.846, + 0.533 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.075, + 0.542, + 0.895, + 0.598 + ], + "angle": 0, + "content": "Figure 1. As the preview of the movie, trailers are selected by professionals to grab an audience's attention. However, trailers are usually composed with shots sparsely selected from movies to avoid spoilers, and the audience cannot get complete highlight information. Some trailer clips convey the artistic style of the film only and lack movie storylines, disturbing the audience's impressions. In addition, different audiences may be interested in different styles of clips, which makes it challenging to learn highlights from them." + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.613, + 0.314, + 0.629 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.637, + 0.474, + 0.879 + ], + "angle": 0, + "content": "Movie highlights stand out of the screenplay for efficient browsing and play a crucial role on social media platforms. Based on existing efforts, this work has two observations: (1) For different annotators, labeling highlight has uncertainty, which leads to inaccurate and time-consuming annotations. (2) Besides previous supervised or unsupervised settings, some existing video corpora can be useful, e.g., trailers, but they are often noisy and incomplete to cover the full highlights. In this work, we study a more practical and promising setting, i.e., reformulating highlight detection as \"learning with noisy labels\". This setting does not require time-consuming manual annotations and can fully utilize existing abundant video corpora. First, based on movie trailers, we leverage scene segmentation to obtain complete shots, which are regarded as noisy labels. Then, we propose a Collaborative noisy Label" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.614, + 0.895, + 0.825 + ], + "angle": 0, + "content": "Cleaner (CLC) framework to learn from noisy highlight moments. CLC consists of two modules: augmented crosspropagation (ACP) and multi-modality cleaning (MMC). The former aims to exploit the closely related audio-visual signals and fuse them to learn unified multi-modal representations. The latter aims to achieve cleaner highlight labels by observing the changes in losses among different modalities. To verify the effectiveness of CLC, we further collect a large-scale highlight dataset named MovieLights. Comprehensive experiments on MovieLights and YouTube Highlights datasets demonstrate the effectiveness of our approach. Code has been made available at: https://github.com/TencentYoutuResearch/HighlightDetection-CLC." + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.845, + 0.633, + 0.861 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.894, + 0.901 + ], + "angle": 0, + "content": "With the growing number of new publications of movies in theaters and streaming media, audiences become even" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.888, + 0.222, + 0.9 + ], + "angle": 0, + "content": "*Corresponding author." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18898" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.273 + ], + "angle": 0, + "content": "harder to choose their favorite one to enjoy for the next two hours. An effective solution is to watch the movie trailers before choosing the right movie. This is because trailers are generally carefully edited by filmmakers and contain the most prominent clips from the original movies. As a condensed version of full-length movies, trailers are elaborately made with highlight moments to impress the audiences. Consequently, they are high potential in serving as supervision sources to train automatic video highlight detection algorithms and facilitating the mass production of derivative works for video creators in online video platforms, e.g., YouTube and TikTok." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.276, + 0.473, + 0.534 + ], + "angle": 0, + "content": "Existing video highlight detection (VHD) approaches are generally trained with annotated key moments of long-form videos. However, they are not suitable to tackle the movie highlight detection task by directly learning from trailers. The edited shots in trailers are not equivalent to ground-truth highlight annotations in movies. Although a previous work [43] leverages the officially-released trailers as the weak supervision to train a highlight detector, the highlighted ness of trailer shots is extremely noisy and varies with the preference of audiences, as shown in Fig. 1. On one hand, trailers tend to be purposefully edited to avoid spoilers, thus missing key moments of the storylines. On the other hand, some less important moments in the original movies are over-emphasized in the trailers because of some artistic or commercial factors. The subjective nature of trailer shots makes them noisy for the VHD task, which is ignored by existing VHD approaches." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.536, + 0.473, + 0.763 + ], + "angle": 0, + "content": "To alleviate the issue, we reformulate the highlight detection task as \"learning with noisy labels\". Specifically, we first leverage a scene-segmentation model to obtain the movie scene boundaries. The clips containing trailers and clips from the same scenes as the trailers provide more complete storylines. They have a higher probability of being highlight moments but still contain some noisy moments. Subsequently, we introduce a framework named Collaborative noisy Label Cleaner (CLC) to learn from these pseudo-noisy labels. The framework firstly enhances the modality perceptual consistency via the augmented crosspropagation (ACP) module, which exploits closely related audio-visual signals during training. In addition, a multimodality cleaning (MMC) mechanism is designed to filter out noisy and incomplete labels." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.473, + 0.903 + ], + "angle": 0, + "content": "To support this study and facilitate benchmarking existing methods in this direction, we construct MovieLights, a Movie Highlight Detection Dataset. MovieLights contains 174 movies and the highlight moments are all from officially released trailers. The total length of these videos is over 370 hours. We conduct extensive experiments on MovieLights, in which our CLC exhibits promising results. We also demonstrate that our proposed CLC achieves significant performance-boosting over the state-of-the-art on" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.697, + 0.106 + ], + "angle": 0, + "content": "the public VHD benchmarks." + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.107, + 0.862, + 0.122 + ], + "angle": 0, + "content": "In summary, our major contributions are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.129, + 0.892, + 0.205 + ], + "angle": 0, + "content": "- We introduce a scene-aware paradigm to learn high-light moments in movies without any manual annotation. To the best of our knowledge, this is the first time that highlights detection is regarded as learning with noisy labels." + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.213, + 0.892, + 0.272 + ], + "angle": 0, + "content": "- We present an augmented cross-propagation to capture the interactions across modalities and a consistency loss to maximize the agreement between the different modalities." + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.282, + 0.892, + 0.327 + ], + "angle": 0, + "content": "- We incorporate a multi-modality noisy label cleaner to tackle label noise, which further improves the robustness of networks to annotation noise." + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.336, + 0.892, + 0.365 + ], + "angle": 0, + "content": "- Experiments on movie datasets and benchmark datasets validate the effectiveness of our framework." + }, + { + "type": "list", + "bbox": [ + 0.519, + 0.129, + 0.892, + 0.365 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.378, + 0.651, + 0.394 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.403, + 0.892, + 0.78 + ], + "angle": 0, + "content": "Video Highlight Detection. This task aims to identifying the interesting moments from untrimmed videos. In recent years, the videos studied for this task extend from domain-specific sport videos [41] to general videos such as social media videos [37], news [35], first-person videos [53] and vlog [25]. Most of previous works [14,21,53] interpret the video highlight detection task as a segment-level ranking problem. They compare pairwise segments from same domain video in order to learn a model that assigns highlight scores to these segments where the highlight segments receive higher scores than the non-highlight segments. MINI-Net [16] proposes to cast highlight detection as multiple instance ranking network learning. SL-Module [50] explores the highlight detection problem through Unsupervised Domain Adaptation (UDA) [32]. UMT [29] integrates highlight detection and moment retrieval into a unified framework and conduct joint optimization. PLD [45] models the video highlight detection into a pixel-level distinction estimation task. In this work, we regard highlight detection as learning with noisy labels. Joint-VA [2] also considers video highlight detection from the perspective of noise. However, it focuses on noise in features, such as videos having noisy audio when the microphone constantly has water splashing against it. We focus on specific annotation noise in video highlight detection." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.78, + 0.892, + 0.903 + ], + "angle": 0, + "content": "Studies on Movies and Trailers. Studies on movies and trailers have received increased attention in research. [18] introduces a comprehensive dataset for movie understanding. [4, 42] try to model the relationships among the movie characters. [7, 34, 47] focus on breaking the storylines of movies into semantically cohesive parts. Besides the studies on movies, efforts have been made to develop trailer understanding. [57] presents a movie summarization system" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "18899" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.243 + ], + "angle": 0, + "content": "and composes movie summaries in terms of user experience evaluation. [9] designs a movie trailer dataset for the evaluation of video-based recommender systems. [19] is the first approach that bridges trailers and movies and allows knowledge learned from trailers to be transferred to full movie analysis. In [43], the visual module and the temporal analysis module are respectively trained on trailers and movies. Because of the inaccessibility of public trailer-related benchmarks, we construct a new dataset (MovieLights) to detect the highlight moments in movies." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.244, + 0.473, + 0.44 + ], + "angle": 0, + "content": "Learning with Noisy Labels. Learning with noisy labels has been a long-standing problem in computer vision. There are three kinds of approaches to this problem. One of the most common strategies for tackling label noise is to capture the transition probabilities between noisy labels and clean labels [31, 38, 48, 49, 52, 56]. Another solution is to design robust loss functions for model training against noisy labels [8, 26, 28, 30, 44, 55]. A popular method is to design a mechanism to select clean samples or give lower weight for noise samples in the training set to reduce impact of noise [15, 17, 20, 39, 46, 51]. In this paper, we attempt to solve the problem by exploiting multi-modalities nature of movies." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.45, + 0.308, + 0.468 + ], + "angle": 0, + "content": "3. Movie Highlight Dataset" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.477, + 0.47, + 0.688 + ], + "angle": 0, + "content": "In this paper, we aim to detect the highlight moments in movies by learning from easily accessible trailers as the noisy supervision. However, the existing movie and trailer-related benchmarks [3, 9, 18] lack sufficient functions for this task, such as the absence of full-length movies and ground-truth highlight annotations. Huang et al. [19] propose to respectively learn visual representations from trailers and temporal structure from full-length movies in their constructed Large-Scale Movie and Trailer Dataset (LSMTD). Wang [43] constructs a Trailer Moment Detection Dataset (TMDD) for detecting trailer moments from full-length movies without explicit human annotation. Both LSMTD and TMDD are not publicly available, while TMDD only contains three movie genres." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.689, + 0.47, + 0.869 + ], + "angle": 0, + "content": "The inaccessibility of public benchmarks motivates us to construct a new dataset, named Movie Highlight Detection Dataset (MovieLights). In particular, we purchase a set of movies from commercial channels and collect their corresponding trailers from streaming platforms such as YouTube, covering at least 25 genres to ensure content diversity. The movies and trailers are then prepossessed by shot segmentation [36] and scene segmentation [47], respectively. The resulting shots are a series of consecutive frames taken by the camera until a physical interruption, and the scenes are consecutive shots that share a semantically related theme." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.47, + 0.902 + ], + "angle": 0, + "content": "As seen in Fig. 2, to build the ground truth, we conduct Faiss [22] to obtain visual similarity matching between" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.089, + 0.895, + 0.224 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.237, + 0.895, + 0.294 + ], + "angle": 0, + "content": "Figure 2. The labeling process of MovieLights. For the training set, we introduce a scene-aware paradigm to obtain labels automatically. For the testing set, we collect 2 sets of labels from different annotators." + }, + { + "type": "table_caption", + "bbox": [ + 0.563, + 0.313, + 0.829, + 0.326 + ], + "angle": 0, + "content": "Table 1. The basic statistics of MovieLights." + }, + { + "type": "table", + "bbox": [ + 0.517, + 0.327, + 0.877, + 0.448 + ], + "angle": 0, + "content": "
TrainTest
Movie Number14430
Avg Durations per Movie2.19h2.14h
Avg Shot Number per Movie18521940
Avg Scene Number per Movie207193
Annotator1 Positive sample Proportion-0.27
Annotator2 Positive sample Proportion-0.30
Positive sample proportion0.350.21
" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.477, + 0.892, + 0.824 + ], + "angle": 0, + "content": "trailer frames and movie frames. We locate trailer moments in the movie and align them with the movie shots as annotation references. For the testing set, we collect 2 sets of moments for each movie from different workers, and these moments are annotated by different annotators independently. To ensure the consistency of results from different annotations, during the annotation procedure, all highlight moments must be related to the annotation references. Though all selected shots are relevant to the trailers, as highlight moments can be subjective, they may still vary in their saliency and time span. We calculate the intersection between every pair of moments annotated as the ground truth. However, the vast diversity of movie storylines makes the annotation challenging as it is time-consuming and requires annotators to be familiar with the movie. To collect a large amount of training data efficiently, we introduce a scene-aware paradigm to obtain the highlight moments label without any manual annotation. Specifically, we expand the shot-level annotation references to the scene span as positive samples automatically. It will capture the complete scene context of the trailer shot with movie storylines. Since the trailer shots may contain some less important moments, the acquired highlight labels are still noisy." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.826, + 0.894, + 0.903 + ], + "angle": 0, + "content": "As seen in Tab. 1, MovieLights contains 174 movies in full length with their official trailers and it is split into a training set with 144 movies, and a testing set with 30 movies. The content diversity is ensured by the rich domain informations (more than 25 genres) and abundant segments" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "18900" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.196 + ], + "angle": 0, + "content": "(325k shots and 36k scenes). Most movies in our dataset are between 90 to 150 minutes, and the length of the annotated moments varies from tens of seconds to several minutes. As the acquired positive highlight moments take up \\(35\\%\\) in the training set while the annotated true positives take up \\(21\\%\\) in the testing set, the difference tells the obvious existence of label noise." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.198, + 0.471, + 0.258 + ], + "angle": 0, + "content": "We plan to release the dataset publicly to promote further study of movie analysis. Due to copyright issues, trailers and movies will be released in the form of extracted features in visual and audio modalities." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.269, + 0.186, + 0.285 + ], + "angle": 0, + "content": "4. Approach" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.294, + 0.187, + 0.308 + ], + "angle": 0, + "content": "4.1. Overview" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.318, + 0.469, + 0.379 + ], + "angle": 0, + "content": "The overall architecture of our Collaborative noisy Label Cleaner (CLC) framework is illustrated in Fig. 3, which includes three modules: feature extraction, augmented crosspropagation (ACP), and multi-modal cleaning (MMC)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.38, + 0.469, + 0.501 + ], + "angle": 0, + "content": "In our framework, both the visual and audio modalities are utilized. For feature extraction and encoding, we first split the video \\(V\\) into \\(T\\) shots. We characterize the \\(i^{th}\\) shots by two vectors, i.e., \\(\\mathbf{v}_i\\) for the visual features, and \\(\\hat{\\mathbf{a}}_{\\mathbf{i}}\\) for the audio features, where \\(i = 1,2,\\dots,T\\). These features are extracted using pre-trained visual [11] and audio feature extractors [24]. The parameters of the two extractors are frozen during training." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.501, + 0.47, + 0.668 + ], + "angle": 0, + "content": "The ACP and MMC are the core components of our framework. Since the visual-audio signals in videos are closely related but do not always contribute to highlight detection, the ACP module exploits the relationship via unimodal and cross-modal interactions. Then, it learns unified multi-modal representations for highlight detection. As the obtained highlight moments after scene segmentation are noisy, the MMC module firstly observes the changes in unimodal losses, then filters the noisy labels and utilizes the clean ones for multi-modal supervised learning. Next, we will introduce details for each component." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.68, + 0.35, + 0.696 + ], + "angle": 0, + "content": "4.2. Augmented Cross-Propagation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.704, + 0.469, + 0.822 + ], + "angle": 0, + "content": "Cross-Propagation. To predict the highlights, the model needs to understand the storylines of the movie. Meanwhile, visual and audio inputs do not always contribute to accurate prediction. Therefore, temporal modeling and modality interaction are the keys to achieving successful highlight detection. To achieve this goal, we design our ACP module to fuse the visual-audio modalities. It involves three steps in total." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.826, + 0.469, + 0.871 + ], + "angle": 0, + "content": "First, in order to align the multi-modal features, we introduce \\( h \\) which is fully-connected (FC) layer with ReLU to the \\( \\hat{\\mathbf{a}}_{i} \\) such that it has same dimension as the \\( \\mathbf{v}_i \\)." + }, + { + "type": "equation", + "bbox": [ + 0.232, + 0.886, + 0.469, + 0.902 + ], + "angle": 0, + "content": "\\[\n\\mathbf {a} _ {i} = h \\left(\\hat {\\mathbf {a}} _ {\\mathrm {i}}\\right). \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.212 + ], + "angle": 0, + "content": "Movies are composed of consecutive audio-visual clips. Therefore, to measure the highlighted ness of a given shot, one must consider the relationship of the shot with its adjacent shots. The self-attention mechanism has shown effectiveness in capturing the long-term dependencies in previous works. We leverage self-attention to capture the temporal relationship for \\(\\mathbf{v}_i\\) and \\(\\mathbf{a}_i\\) via Eq. 2 and Eq. 3, respectively." + }, + { + "type": "equation", + "bbox": [ + 0.543, + 0.22, + 0.892, + 0.256 + ], + "angle": 0, + "content": "\\[\n\\mathbf {v} _ {i} ^ {s} = \\operatorname {s o f t m a x} \\left(\\frac {\\left(\\mathbf {v} _ {i} W _ {1} ^ {v}\\right) \\left(\\mathbf {v} W _ {2} ^ {v}\\right) ^ {\\top}}{\\sqrt {d}}\\right) \\left(\\mathbf {v} W _ {3} ^ {v}\\right), \\tag {2}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.545, + 0.263, + 0.892, + 0.298 + ], + "angle": 0, + "content": "\\[\n\\mathbf {a} _ {i} ^ {s} = \\operatorname {s o f t m a x} \\left(\\frac {\\left(\\mathbf {a} _ {i} W _ {1} ^ {a}\\right) \\left(\\mathbf {a} W _ {2} ^ {a}\\right) ^ {\\top}}{\\sqrt {d}}\\right) \\left(\\mathbf {a} W _ {3} ^ {a}\\right), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.301, + 0.892, + 0.407 + ], + "angle": 0, + "content": "where \\(\\mathbf{v} = [\\mathbf{v}_1; \\mathbf{v}_2; \\ldots; \\mathbf{v}_T]\\) and \\(\\mathbf{a} = [\\mathbf{a}_1; \\mathbf{a}_2; \\ldots; \\mathbf{a}_T]\\); the scaling factor \\(d\\) is equal to the visual/audio feature dimension and \\((*)^{\\top}\\) denotes the transpose operator; \\(W^{v}\\) and \\(W^{a}\\) are learnable matrices of two modalities, which are implemented by a linear layer. Uni-modal self-attention can well capture uni-modal temporal contexts and enhance clip features within the same modality." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.408, + 0.892, + 0.483 + ], + "angle": 0, + "content": "Despite the above self-attention capturing the clip interactions within the uni-modality, it is critical to capture the interactions across modalities. To capture semantic associations based on multi-modal signals, we introduce cross-attention to update the features of each modality." + }, + { + "type": "equation", + "bbox": [ + 0.543, + 0.491, + 0.892, + 0.526 + ], + "angle": 0, + "content": "\\[\n\\mathbf {v} _ {i} ^ {c} = \\operatorname {s o f t m a x} \\left(\\frac {\\left(\\mathbf {v} _ {i} W _ {4} ^ {v}\\right) \\left(\\mathbf {a} W _ {4} ^ {a}\\right) ^ {\\top}}{\\sqrt {d}}\\right) \\left(\\mathbf {a} W _ {5} ^ {a}\\right), \\tag {4}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.543, + 0.533, + 0.892, + 0.569 + ], + "angle": 0, + "content": "\\[\n\\mathbf {a} _ {i} ^ {c} = \\operatorname {s o f t m a x} \\left(\\frac {\\left(\\mathbf {a} _ {i} W _ {6} ^ {a}\\right) \\left(\\mathbf {v} W _ {5} ^ {v}\\right) ^ {\\top}}{\\sqrt {d}}\\right) \\left(\\mathbf {v} W _ {6} ^ {v}\\right), \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.572, + 0.892, + 0.678 + ], + "angle": 0, + "content": "Through cross-attention, the information from two modalities are connected. However, considering audiovisual temporal asynchrony, it is necessary to select effective information from multi-modality. We augment the relevant positive connections and dampen the irrelevant connections. The strength of these connections is measured by the cross-correlation matrix, computed by," + }, + { + "type": "equation", + "bbox": [ + 0.543, + 0.685, + 0.892, + 0.721 + ], + "angle": 0, + "content": "\\[\n\\mathbf {c} ^ {v} = R e L U \\left(\\frac {\\mathbf {v a} ^ {\\top}}{\\sqrt {d}}\\right), \\mathbf {c} ^ {a} = R e L U \\left(\\frac {\\mathbf {a v} ^ {\\top}}{\\sqrt {d}}\\right). \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.727, + 0.892, + 0.802 + ], + "angle": 0, + "content": "For each modality, the cross-correlation matrix \\(\\mathbf{c}\\) is used to re-weight the cross-attention features. Finally, we obtain updated visual features \\(\\bar{\\mathbf{v}}_{\\mathrm{i}}\\) and audio features \\(\\bar{\\mathbf{a}}_{\\mathrm{i}}\\) by fusion of the original features, enhanced uni-modal features, and cross-modal features." + }, + { + "type": "equation", + "bbox": [ + 0.592, + 0.809, + 0.892, + 0.851 + ], + "angle": 0, + "content": "\\[\n\\bar {\\mathbf {v}} _ {\\mathbf {i}} = f \\left(\\mathbf {v} _ {i}, \\mathbf {v} _ {i} ^ {s}, \\left(\\sum_ {j = 1} ^ {T} \\mathbf {c} _ {i j} ^ {v}\\right) * \\mathbf {v} _ {i} ^ {c}\\right), \\tag {7}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.594, + 0.862, + 0.892, + 0.904 + ], + "angle": 0, + "content": "\\[\n\\overline {{\\mathbf {a}}} _ {\\mathrm {i}} = f \\left(\\mathbf {a} _ {i}, \\mathbf {a} _ {i} ^ {s}, \\left(\\sum_ {j = 1} ^ {T} \\mathbf {c} _ {i j} ^ {a}\\right) * \\mathbf {a} _ {i} ^ {c}\\right), \\tag {8}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "18901" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.084, + 0.102, + 0.885, + 0.343 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.361, + 0.893, + 0.431 + ], + "angle": 0, + "content": "Figure 3. Overview of the proposed CLC. It includes three modules: feature extraction, augmented cross-propagation (ACP), and multi-modal cleaning (MMC). The visual and audio modalities of the input video are represented as vectors by the feature extraction module. Then the features are augmented by ACP module to capture semantic associations across modalities. MMC is used to filter outs noisy and incomplete labeling with additional uni-modal branches. During inference, we remove the uni-modal branches and only rely on the prediction of multi-modal branch. More details of the CLC are shown in Sec. 4.1" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.439, + 0.471, + 0.469 + ], + "angle": 0, + "content": "where \\( f \\) is the fusion function consisting of FC layers and ReLU to further project the features." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.47, + 0.47, + 0.742 + ], + "angle": 0, + "content": "Clearly, the cross-correlation matrix will assign large weights to clips that are relevant to the other modality. In addition, the ReLU activation in Eq. 6 cuts off connections with negative similarity values and only relevant positive connections would be preserved. By the above operations, the original feature will be infused with richer information. Consistency Loss. Multi-modal inputs help to comprehensively learn by integrating different aspects and boosting model performance. However, they are not fully exploited because some modality-specific features may be weakened in the fusion even when the multi-modal model outperforms its uni-modal counterpart. In this work, we provide parallel branches for each modality separately to obtain uni-modal prediction during training. As seen in Fig. 3, all branches share the same clip feature extracted from ACP and the feature is fed into different branches independently. In each branch, we develop a temporal model \\( G \\) to obtain its prediction score \\( y_{i} \\) of the \\( i^{th} \\) clip being a highlight as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.088, + 0.751, + 0.469, + 0.77 + ], + "angle": 0, + "content": "\\[\n\\mathbf {y} _ {i} ^ {M M} = G \\left(\\bar {\\mathbf {v}} _ {\\mathbf {i}}, \\bar {\\mathbf {a}} _ {\\mathbf {i}}\\right), \\mathbf {y} _ {i} ^ {U M _ {v}} = G \\left(\\bar {\\mathbf {v}} _ {\\mathbf {i}}\\right), \\mathbf {y} _ {i} ^ {U M _ {a}} = G \\left(\\bar {\\mathbf {a}} _ {\\mathbf {i}}\\right). \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.78, + 0.469, + 0.809 + ], + "angle": 0, + "content": "where \\(MM\\) is multi-modal branch; \\(UM_{v}\\) is visual branch and \\(UM_{a}\\) is audio branch." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.811, + 0.469, + 0.9 + ], + "angle": 0, + "content": "However, the gap between the different modalities would bring instability in the joint optimization process. We employ the auxiliary consistency loss to guarantee consistency between the different modalities. Given the multi-modal features of a clip, they are consistent if they share the same prediction. Specifically, the consistency loss is defined as" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.439, + 0.892, + 0.471 + ], + "angle": 0, + "content": "the cross-entropy between the multi-modal prediction probability \\(\\mathbf{y}_i^{MM}\\) and uni-modal prediction probability \\(\\mathbf{y}_i^{UM}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.51, + 0.477, + 0.892, + 0.512 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {c o n s}} = - \\left(\\sum_ {i = 1} ^ {N} \\mathbf {y} _ {i} ^ {U M _ {v}} \\log \\mathbf {y} _ {i} ^ {M M} + \\sum_ {i = 1} ^ {N} \\mathbf {y} _ {i} ^ {U M _ {a}} \\log \\mathbf {y} _ {i} ^ {M M}\\right), \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.518, + 0.803, + 0.533 + ], + "angle": 0, + "content": "where \\(N\\) is the number of samples in a batch." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.534, + 0.892, + 0.61 + ], + "angle": 0, + "content": "This consistency loss not only implicitly enhances uni-modal information, but also explicitly guides the multimodal branch to robust supervision. The added uni-modal branches are only utilized in the training phase and are disabled during the inference stage." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.62, + 0.708, + 0.636 + ], + "angle": 0, + "content": "4.3. Multi-modal Cleaning" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.644, + 0.892, + 0.749 + ], + "angle": 0, + "content": "Multi-modality Sample Cleaning. Annotation noise is inevitable in VHD due to subjectivity depending on the users and annotators. In this paper, we argue that highlight detection should be regarded as Learning with Noisy Labels. To alleviate the performance drop caused by noisy labels, we adopt a multi-modality collaborative cleaning to adaptively filter noisy samples with noisy modality information." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.893, + 0.901 + ], + "angle": 0, + "content": "Firstly, we maintain multiple outputs simultaneously which are predicted by different branches. The uni-modal branches independently select clean samples based on the low-loss criterion in which instances with lower losses are treated clean samples. Contrary to existing noisy sample selection methods [12, 54], which directly discard high-loss samples, we keep all samples to train the uni-modal branches. Then, we update the multi-modal branch using only clean samples selected by both uni-modal branches in the back-propagation. The samples for the multi-modal" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18902" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.167 + ], + "angle": 0, + "content": "branch training are selected dynamically while all samples participate in the uni-mode branches training. In this way, MMC obtains information on all samples to avoid the model defecting to favoring easy samples. More details of MMC are presented in the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.168, + 0.47, + 0.211 + ], + "angle": 0, + "content": "Each branch has its own loss function. For the uni-modal branch, we employ the cross-entropy loss with all samples as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.143, + 0.22, + 0.469, + 0.26 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} ^ {U M _ {v}} = \\mathcal {L} _ {\\mathrm {c e}} ^ {U M _ {v}} = - \\sum_ {i = 1} ^ {N} \\mathbf {g} _ {i} \\log \\mathbf {y} _ {i} ^ {U M _ {v}}, \\tag {11}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.143, + 0.271, + 0.469, + 0.312 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} ^ {U M _ {a}} = \\mathcal {L} _ {\\mathrm {c e}} ^ {U M _ {a}} = - \\sum_ {i = 1} ^ {N} \\mathbf {g} _ {i} \\log \\mathbf {y} _ {i} ^ {U M _ {a}}. \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.316, + 0.47, + 0.362 + ], + "angle": 0, + "content": "For the multi-modal branch, its cross-entropy loss and consistency loss are updated with a re-weighting scheme with clean samples." + }, + { + "type": "equation", + "bbox": [ + 0.179, + 0.372, + 0.469, + 0.413 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {c e}} ^ {\\mathrm {M M}} = - \\sum_ {i = 1} ^ {N ^ {\\prime}} \\mathbf {g} _ {i} \\log \\mathbf {y} _ {i} ^ {M M}, \\tag {13}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.086, + 0.423, + 0.469, + 0.462 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {c o n s}} ^ {M M} = - \\left(\\sum_ {i = 1} ^ {N ^ {\\prime}} \\mathbf {y} _ {i} ^ {U M _ {v}} \\log \\mathbf {y} _ {i} ^ {M M} + \\sum_ {i = 1} ^ {N ^ {\\prime}} \\mathbf {y} _ {i} ^ {U M _ {a}} \\log \\mathbf {y} _ {i} ^ {M M}\\right), \\tag {14}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.189, + 0.463, + 0.469, + 0.482 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} ^ {\\mathrm {M M}} = \\mathcal {L} _ {\\mathrm {c e}} ^ {\\mathrm {M M}} + \\beta \\mathcal {L} _ {\\mathrm {c o n s}} ^ {M M}, \\tag {15}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.487, + 0.469, + 0.56 + ], + "angle": 0, + "content": "where \\(\\mathbf{g}_i\\) and \\(\\mathbf{y}_i^*\\) denote the ground-truth and predicted probability of the \\(i^{th}\\) clip, respectively; \\(N\\) is the number of samples in a batch and \\(N'\\) is the number of samples selected by MMC; and \\(\\beta\\) are designed to balance different loss terms." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.563, + 0.469, + 0.669 + ], + "angle": 0, + "content": "Post processing. In experiments we observe that noise makes jitter prediction curves along the temporal dimension, which may cause discontinuous thresholds for highlight selection. Therefore, we apply a median filter to smooth the prediction curves. Supposing \\(\\mathbf{y} = [\\mathbf{y}_1; \\mathbf{y}_2; \\dots; \\mathbf{y}_T]\\) is the original curve predicted by the CLC, the smoothed curve \\(\\mathbf{s} = [\\mathbf{s}_1; \\mathbf{s}_2; \\dots; \\mathbf{s}_T]\\) is given by:" + }, + { + "type": "equation", + "bbox": [ + 0.11, + 0.678, + 0.469, + 0.719 + ], + "angle": 0, + "content": "\\[\n\\mathbf {s} _ {i} = \\left\\{ \\begin{array}{l l} \\operatorname {M e d} \\left(\\mathbf {y} _ {i - k}, \\mathbf {y} _ {i + k}\\right), & k < i \\leq T - k \\\\ \\mathbf {y} _ {i}, & \\text {o t h e r w i s e} \\end{array} \\right. \\tag {16}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.727, + 0.469, + 0.755 + ], + "angle": 0, + "content": "where \\( k \\) is the window size, and \"Med\" denotes the median filter." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.763, + 0.201, + 0.779 + ], + "angle": 0, + "content": "5. Experiment" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.787, + 0.387, + 0.803 + ], + "angle": 0, + "content": "5.1. Datasets and Experimental Settings" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.81, + 0.469, + 0.871 + ], + "angle": 0, + "content": "Datasets. We evaluate our CLC on the constructed dataset MovieLights and public YouTube Highlights dataset [37]. MovieLights is split into training and testing sets, each containing 144 and 30 movies respectively." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.871, + 0.469, + 0.9 + ], + "angle": 0, + "content": "We split movies into shots. The movie features are represented at shot-level. We use the middle frame of each" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.182 + ], + "angle": 0, + "content": "shot to extract its visual feature with ViT [10] pre-trained by CLIP [33]. We align timestamps of audio clips with the visual shots, and sample the audio clip of each with \\(16K\\) Hz sampling rate and 512 windowed signal length. The resulted shot-level audio features are obtained with the PANN audio network [24] pretrained on AudioSet [13]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.184, + 0.892, + 0.349 + ], + "angle": 0, + "content": "The YouTube Highlights contains six distinct categories with a total 422 videos currently available. Following the practice of prior efforts, we train a highlight detector for each category. YouTube Highlights provides two annotations: Harvested Highlight and Mturk Highlight. In the Harvested annotation, the match label specifies if each clip is matched in the edited video, where 1 denotes matched, -1 denotes unmatched and 0 denotes the borderline cases. In the Mturk annotation, the highlight labels are marked by multiple turkers of different styles, making it noisier than the Harvested annotation." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.353, + 0.892, + 0.518 + ], + "angle": 0, + "content": "On YouTube Highlights, we use the same protocol and data preprocessing as [29]. It obtain clip-level visual features and optical flow features using I3D [6] pre-trained on Kinetics400 [23]. It use a PANN audio network [24] pretrained on AudioSet [13] to obtain audio features that align with the visual clips. Frame-level features are average-pooled within each clip for both audio and visual features to generate a clip-level feature. Since each feature vector spans 32 consecutive frames, we follow [29] and consider the feature vector corresponded to a clip if their overlap is more than \\(50\\%\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.521, + 0.892, + 0.822 + ], + "angle": 0, + "content": "Benchmarks. To better inspect the robustness of our CLC against noisy labels, we also apply label perturbations in training set of YouTube Highlights while keeping the ground-truths in the testing set unchanged. YouTube Highlights has two benchmarks for comparison. 1) Harvested with matched: we regard the clips labeled with matched as the highlighted clips and this is the same setting as in previous works [29, 50]. 2) Harvested with matched and borderline: clips labeled with matched and borderline are treated as the highlight moments. In this benchmark, the training set contains some highlighted clips with weak confidence. We select the clips whose mturk-label is over score 1 as the highlighted clips, which means that at least one turker selects the clip as a highlight. There are labeled by different types of annotators between Mturk and the test set and bring greater noise. We select the clips whose mturk-label is over score 1 as the highlighted clips, which means that at least one turker selects the clip as a highlight. Due the labeling gap between the Mturk annotation in the training set and the clean testing set, this benchmark is even noisier." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.825, + 0.892, + 0.9 + ], + "angle": 0, + "content": "Baselines. We introduce CLC-, the degenerated version of CLC, as a baseline. Similar to CLC, CLC- is a Bi-LSTM-based model and takes the temporal sequence of shot features as input but lacks the modules of augmented cross-propagation and multi-modality noisy label cleaner." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18903" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.173, + 0.09, + 0.373, + 0.103 + ], + "angle": 0, + "content": "Table 2. Results on MovieLights." + }, + { + "type": "table", + "bbox": [ + 0.115, + 0.104, + 0.434, + 0.24 + ], + "angle": 0, + "content": "
MethodsModalitymAP
GIFs [14]V25.48
SL-Module [50]V32.34
SL-Module [50]VA34.27
UMT [29]VA38.7
CLC-VA39.65
CLC- w/ SCE [44]VA39.83
CLC- w/ LS [40]VA40.49
CLCVA43.88
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.265, + 0.47, + 0.356 + ], + "angle": 0, + "content": "Evaluation Metric. We adopt mean Average Precision (mAP) as the evaluation metric for MovieLights and YouTube Highlights. Considering that a highlighted moment in one video is not necessarily more interesting than non-highlight moments in other videos, we evaluate on each test video independently and report the averaged results." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.356, + 0.47, + 0.462 + ], + "angle": 0, + "content": "Implementation Details. On the MovieLights, we train our model using SGD, with a learning rate of 0.01. We train for 50 epochs. Before the cross-attention modules, we project each modality into a vector of 512 dimension. The key, query, and value vectors all share the same dimension. Weight \\(\\beta\\) in Eq. 15 is empirically set to 0.1 and window size \\(k\\) in Eq. 16 is set to 9." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.471, + 0.295, + 0.488 + ], + "angle": 0, + "content": "5.2. Results on MovieLights" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.495, + 0.47, + 0.601 + ], + "angle": 0, + "content": "On MovieLights, we train our model with the noisy pseudo labels. To compare with previous state-of-the-art highlight detection works, we train UMT [29] and SL-Module [50] using the same protocol and data preprocessing as in CLC. We also compare with Video2GIF [14] using its off-the-shelf tool1. The upper part of Tab. 2 illustrates the significant performance gain of CLC." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.601, + 0.471, + 0.766 + ], + "angle": 0, + "content": "To demonstrate the advantages of CLC in learning with noisy labels, we make comparisons with two main-stream label noise approaches: Label Smoothing [40] and SCE loss [44]. Specifically, we insert Label Smoothing or SCE into our CLC- framework to create two baseline VHD methods to tackle label noise. The bottom part of Tab. 2 shows that CLC outperforms the two baseline methods by a notable margin, indicating that our augmented cross-propagation and multi-modality noisy label cleaner are more effective than vanilla label noise approaches in VHD tasks." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.777, + 0.231, + 0.794 + ], + "angle": 0, + "content": "5.3. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.801, + 0.469, + 0.83 + ], + "angle": 0, + "content": "In this experiment, we analyze the impact of each module. The results are summarized in Tab. 3." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.832, + 0.469, + 0.877 + ], + "angle": 0, + "content": "Multi-modality Sample Cleaning. We first inspect the impact of the multi-modality sample filter module because we are primary concerned with learning with noisy labels" + }, + { + "type": "table_caption", + "bbox": [ + 0.568, + 0.09, + 0.823, + 0.103 + ], + "angle": 0, + "content": "Table 3. Ablation results of MoiveLightts." + }, + { + "type": "table", + "bbox": [ + 0.537, + 0.104, + 0.858, + 0.197 + ], + "angle": 0, + "content": "
MMSCCPCLPPmAP
××××39.65
✓×××41.69
✓✓××42.79
✓✓✓×43.22
✓✓✓✓43.88
" + }, + { + "type": "table_caption", + "bbox": [ + 0.574, + 0.211, + 0.816, + 0.224 + ], + "angle": 0, + "content": "Table 4. Results on YouTube Highlights." + }, + { + "type": "table", + "bbox": [ + 0.502, + 0.226, + 0.898, + 0.415 + ], + "angle": 0, + "content": "
Methodsdoggym.park.ska.ski.surf.Avg.
GIFs [14]30.833.55455.432.854.146.4
LSVM [37]60.041.061.062.036.061.053.6
HighlightMe [5]63737264526264
MINI-Net [16]58.261.770.272.258.760.164.4
CHD [1]60.671.174.249.868.268.565.4
Trail [43]63.382.562.352.974.579.369.1
SL-Module [50]70.853.277.272.566.176.269.3
Joint-VA [2]64.571.980.86273.278.371.8
PLD [45]74.970.277.957.570.77973
CO-AV [27]60.9668974.16981.174.7
UMT [29]65.975.281.671.872.382.774.9
CLC(ours)70.579.483.983.579.583.680.1
" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.44, + 0.892, + 0.607 + ], + "angle": 0, + "content": "in VHD. The \\(2\\%\\) performance gain from the module over baseline shows the importance of filtering noisy sample. Based on the this observation, the subsequent ablation experiments are conducted under the setting of noise filtering. Cross-Propagation. We then evaluate the impact of the feature augmentation module. Compared to naive feature concatenation, the models with augmented features show superior performance. The results validate that feature augmentation module can better explore the complementary information from different modalities and suppress the mutual disturbance of desynchronized uni-modal information." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.608, + 0.892, + 0.699 + ], + "angle": 0, + "content": "Consistency Loss. We examine the contribution of the consistency loss. The results demonstrate that the employment of the auxiliary multi-modal constraint further increases the model robustness. This consistency loss not only implicitly enhances uni-modal information, but also explicitly guides the multi-modal branch to better learning." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.7, + 0.892, + 0.76 + ], + "angle": 0, + "content": "Post Processing. We compare the highlight prediction curves with and without the post processing median filter in Fig. 4(a,b), which prevents disruptive prediction variation and improves the overall performance." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.771, + 0.774, + 0.787 + ], + "angle": 0, + "content": "5.4. Results on YouTube Highlights" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.794, + 0.892, + 0.854 + ], + "angle": 0, + "content": "We conduct experiments on the public video highlight detection benchmarks YouTube Highlights, including six domain video datasets, to verify the generalization ability of CLC." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.892, + 0.901 + ], + "angle": 0, + "content": "The setting of Harvested Highlight with matched is consistent with the previous work [29]. As shown in Tab. 4, CLC achieves state-of-the-art performance on the YouTube" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.887, + 0.325, + 0.901 + ], + "angle": 0, + "content": "1https://github.com/gyglim/video2gif_code" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18904" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.089, + 0.347, + 0.25 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.349, + 0.089, + 0.614, + 0.25 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.616, + 0.09, + 0.891, + 0.245 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.261, + 0.891, + 0.277 + ], + "angle": 0, + "content": "Figure 4. (a, b) Comparison of original prediction curve with filtered prediction curve. (c) Results on YouTube Highlights with noisy label." + }, + { + "type": "table_caption", + "bbox": [ + 0.31, + 0.279, + 0.658, + 0.293 + ], + "angle": 0, + "content": "Table 5. Results on YouTube Highlights with Noisy Label." + }, + { + "type": "table", + "bbox": [ + 0.115, + 0.294, + 0.856, + 0.402 + ], + "angle": 0, + "content": "
AnnotationNoiseMethodsdoggym.park.ska.ski.surf.Avg.
Harvested matchedcleanUMT [29]65.9075.2081.6071.8072.3082.7074.90
Harvested matchedcleanCLC70.5179.4383.8583.5179.4683.5680.05 (↑ 5.15)
Harvested borderlineslight noiseUMT [29]65.9374.3181.5871.8470.2482.4674.39
Harvested borderlineslight noiseCLC69.4180.7378.5085.3681.1183.1679.71 (↑ 5.32)
Mturksevere noiseUMT [29]63.7876.1675.0273.6269.9981.5973.36
Mturksevere noiseCLC66.9280.4485.9282.3378.0581.7279.22 (↑ 5.86)
" + }, + { + "type": "image", + "bbox": [ + 0.086, + 0.416, + 0.885, + 0.53 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.54, + 0.893, + 0.584 + ], + "angle": 0, + "content": "Figure 5. The highlight moments selected by our CLC from Skyfall and Sherlock Holmes. Top: Bond gives chase to a professional hitman by car to find a classified hard drive, and then a firefight erupted in the market. Bottom: Holmes and Blackwood are facing off, and then Blackwood reaches for a weapon to kill Holmes, but accidentally trips off a scaffolding and falls to his death." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.591, + 0.469, + 0.698 + ], + "angle": 0, + "content": "Highlights, outperforming the existing multi-modal highlight detection methods in the average metric across all categories. Specifically, CLC achieves best performance in three out of the six categories, while maintaining reasonably competitive performance in the other three categories. These results support our claim that highlight detection should be regarded as learning with noisy labels." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.708, + 0.469, + 0.785 + ], + "angle": 0, + "content": "To quantify how CLC is robust to different levels of label noise, we inspect the settings of Harvested Highlight with matched and borderline and Mturk Highlight, which are perturbed by varying degrees of label noise in YouTube Highlights." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.795, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Tab. 5 exhibits the performances of CLC and UMT [29] at different noise levels. As the noise level increases, the VHD task becomes more difficult, but the performance superiority of our CLC over UMT becomes even more obvious. It is illustrated in Fig. 4(c) that compared with the most recent state-of-the-art UMT, our model can achieve better performance even when disturbed by severe label noise." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.59, + 0.637, + 0.605 + ], + "angle": 0, + "content": "5.5. Visualization" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.615, + 0.892, + 0.691 + ], + "angle": 0, + "content": "As shown in Fig. 5, we present some visualization examples of the detected highlight clips in MovieLights by CLC. The examples clearly show that the prediction of CLC is in accordance with user expectation. We will provide more examples in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.701, + 0.619, + 0.717 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.728, + 0.892, + 0.91 + ], + "angle": 0, + "content": "In this study, we present Collaborative noisy Label Cleaner (CLC), a novel framework to handle noisy labels in video highlight detection. We make use of the augmented cross-propagation module to better enhance network robustness and multi-modality cleaning to achieve cleaner highlight labels by observing the loss changes of different modalities. We demonstrate the state-of-the-art performance of our method with extensive experiments on MovieLights and YouTube Highlights datasets. In future work, we are interested in extending the proposed mechanisms to other video-understanding tasks such as scene segmentation and video temporal grounding." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "18905" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.468, + 0.156 + ], + "angle": 0, + "content": "[1] Taivanbat Badamdorj, Mrigank Rochan, Yang Wang, and Li Cheng. Contrastive learning for unsupervised video highlight detection. In CVPR, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.157, + 0.469, + 0.199 + ], + "angle": 0, + "content": "[2] Taivanbat Badamdorj, Mrigank Rochan, Yang Wang, Jiaze Wang, and Li Cheng. Joint visual and audio learning for video highlight detection. In ICCV, 2021. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.2, + 0.469, + 0.242 + ], + "angle": 0, + "content": "[3] Max Bain, Arsha Nagrani, Andrew Brown, and Andrew Zisserman. Condensed movies: Story based retrieval with contextual embeddings. In ACCV, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.243, + 0.468, + 0.27 + ], + "angle": 0, + "content": "[4] David Bamman, Brendan O'Connor Noah, and A. Smith. Learning latent personas of film characters. 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.271, + 0.469, + 0.325 + ], + "angle": 0, + "content": "[5] Uttaran Bhattacharya, Gang Wu, Stefano Petrangeli, Viswanathan Swaminathan, and Dinesh Manocha. Highlighte: Detecting highlights from human-centric videos. In ICCV, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.327, + 0.468, + 0.367 + ], + "angle": 0, + "content": "[6] Joao Carreira and Andrew Zisserman. Quo vadis, action recognition? a new model and the kinetics dataset. 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.368, + 0.469, + 0.41 + ], + "angle": 0, + "content": "[7] Shixing Chen, Xiaohan Nie, David Fan, Dongqing Zhang, Vimal Bhat, and Raffay Hamid. Shot contrastive self-supervised learning for scene boundary detection. 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.411, + 0.468, + 0.452 + ], + "angle": 0, + "content": "[8] Hao Cheng, Zhaowei Zhu, Xingyu Li, Yifei Gong, Xing Sun, and Yang Liu. Learning with instance-dependent label noise: A sample sieve approach. In ICLR, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.453, + 0.469, + 0.507 + ], + "angle": 0, + "content": "[9] Yashar Deldjoo, Mihai Gabriel Constantin, and Bogdan Ionescu. Mmtf-14k: A multifaceted movie trailer feature dataset for recommendation and retrieval. In ACMMM, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.509, + 0.469, + 0.591 + ], + "angle": 0, + "content": "[10] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Xiaohua Zhai Dirk Weissenborn, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.593, + 0.469, + 0.662 + ], + "angle": 0, + "content": "[11] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.663, + 0.469, + 0.703 + ], + "angle": 0, + "content": "[12] Curtis G, Northcutt, Lu Jiang, and Isaac L. Chuang. Confident learning: Estimating uncertainty in dataset labels. 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.705, + 0.469, + 0.76 + ], + "angle": 0, + "content": "[13] Jort F. Gemmeke, Daniel P. W. Ellis, Dylan Freedman, Aren Jansen, Wade Lawrence, R. Channing Moore, Manoj Plakal, and Marvin Ritter. Audio set: An ontology and human-labeled dataset for audio events. 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.761, + 0.469, + 0.802 + ], + "angle": 0, + "content": "[14] Michael Gygli, Yale Song, and Liangliang Cao. Video2gif: Automatic generation of animated gifs from video. In CVPR, 2016. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.803, + 0.469, + 0.858 + ], + "angle": 0, + "content": "[15] Bo Han, Quanming Yao, Xingrui Yu, Gang Niu, Miao Xu, Weihua Hu, Ivor Tsang, and Masashi Sugiyama. Co-teaching: Robust training of deep neural networks with extremely noisy labels. In NeurIPS, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.859, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[16] Fa-Ting Hong, Xuanteng Huang, Wei-Hong Li, and WeiShi Zheng. Mini-net: Multiple instance ranking network for video highlight detection. In ECCV, 2020. 2, 7" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.469, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.134 + ], + "angle": 0, + "content": "[17] Jinchi Huang, Lie Qu, Rongfei Jia, and Binqiang Zhao. O2unet: A simple noisy label detection approach for deep neural network. In ICCV, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.136, + 0.892, + 0.177 + ], + "angle": 0, + "content": "[18] Qingqiu Huang, Yu Xiong, Anyi Rao, Jiaze Wang, and Dahua Lin. Movienet: A holistic dataset for movie understanding. In ECCV, 2020. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.179, + 0.892, + 0.233 + ], + "angle": 0, + "content": "[19] Qingqiu Huang, Yuanjun Xiong, Yu Xiong, Yuqi Zhang, and Dahua Lin. From trailers to storylines: An efficient way to learn from movies. arXiv preprint arXiv:1806.05341, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.235, + 0.892, + 0.289 + ], + "angle": 0, + "content": "[20] Lu Jiang, Zhengyuan Zhou, Thomas Leung, Li-Jia Li, and Li Fei-Fei. Mentornet: Learning data-driven curriculum for very deep neural networks on corrupted labels. In ICML, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.291, + 0.892, + 0.347 + ], + "angle": 0, + "content": "[21] Yifan Jiao, Zhetao Li, Shucheng Huang, Xiaoshan Yang, Bin Liu, and Tianzhu Zhang. Three-dimensional attention-based deep ranking model for video highlight detection. In IEEE Transactions on Multimedia, page 2693-2705, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.349, + 0.892, + 0.39 + ], + "angle": 0, + "content": "[22] Jeff Johnson, Matthijs Douze, and Hervé Jégou. Billionscale similarity search with gpus. IEEE Transactions on Big Data, page 1-1, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.391, + 0.892, + 0.46 + ], + "angle": 0, + "content": "[23] Will Kay, Joao Carreira, Karen Simonyan, Brian Zhang, Chloe Hillier, Sudheendra Vijayanarasimhan, Fabio Viola, Tim Green, Trevor Back, Paul Natsev, Mustafa Suleyman, and Andrew Zisserman. The kinetics human action video dataset. arXiv preprint arXiv:1705.06950, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.461, + 0.892, + 0.531 + ], + "angle": 0, + "content": "[24] Qiuqiang Kong, Yin Cao, Turab Iqbal, Yuxuan Wang, Wenwu Wang, and Mark D Plumbley. Panns: Large-scale pretrained audio neural networks for audio pattern recognition. IEEE/ACM Transactions on Audio, Speech, and Language Processing, 28:2880-2894, 2020. 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.532, + 0.892, + 0.573 + ], + "angle": 0, + "content": "[25] Jie Lei, Tamara L. Berg, and Mohit Bansal. Qvhighlights: Detecting moments and highlights in videos via natural language queries. In NeurIPS, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.575, + 0.892, + 0.616 + ], + "angle": 0, + "content": "[26] Junnan Li, Richard Socher, and Steven C.H. Hoi. Dividemix: Learning with noisy labels as semi-supervised learning. 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.617, + 0.892, + 0.685 + ], + "angle": 0, + "content": "[27] Shuaicheng Li, Feng Zhang, Kunlin Yang, Lingbo Liu, Shinan Liu, Jun Hou, and Shuai Yi. Probing visual-audio representation for video highlight detection via hard-pairs guided contrastive learning. arXiv preprint arXiv:2206.10157, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.688, + 0.892, + 0.716 + ], + "angle": 0, + "content": "[28] Yang Liu and Hongyi Guo. Peer loss functions: Learning from noisy labels without knowing noise rates. 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.717, + 0.892, + 0.772 + ], + "angle": 0, + "content": "[29] Ye Liu, Siyuan Li, Yang Wu, Chang Wen Chen, Ying Shan, and Xiaohu Qie. Umt: Unified multi-modal transformers for joint video moment retrieval and highlight detection. In CVPR, 2022. 2, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.774, + 0.892, + 0.815 + ], + "angle": 0, + "content": "[30] Xingjun Ma, Hanxun Huang, Yisen Wang, Simone Romano, Sarah Erfani, and James Bailey. Normalized loss functions for deep learning with noisy labels. 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.817, + 0.892, + 0.857 + ], + "angle": 0, + "content": "[31] Curtis G. Northcutt, Lu Jiang, and Isaac L. Chuang. Confident learning: Estimating uncertainty in dataset labels. In JAIR, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.859, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[32] Sinno Jialin Pan and Qiang Yang. A survey on transfer learning. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, page 1345-1359, 2010. 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "18906" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.161 + ], + "angle": 0, + "content": "[33] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Sandhini Agarwal Gabriel Goh, Girish Sastry, Amanda Askell, Pamela Mishkin, and et al. Jack Clark. Learning transferable visual models from natural language supervision. arXiv preprint arXiv:2103.00020, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.162, + 0.47, + 0.204 + ], + "angle": 0, + "content": "[34] Anyi Rao, Linning Xu, Yu Xiong, Guodong Xu, Qingqiu Huang, Bolei Zhou, and Dahua Lin. A local-to-global approach to multi-modal movie scene segmentation. 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.205, + 0.469, + 0.245 + ], + "angle": 0, + "content": "[35] Yale Song, Yahoo Labs, New York, Jordi Vallmitjana, Amanda Stent, and Alejandro Jaime. Tvsum: Summarizing web videos using titles. In CVPR, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.247, + 0.469, + 0.287 + ], + "angle": 0, + "content": "[36] Tomáš Souček and Jakub Lokoc. Transnet v2: An effective deep network architecture for fast shot transition detection. arXiv preprint arXiv:2008.04838, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.288, + 0.469, + 0.328 + ], + "angle": 0, + "content": "[37] Min Sun, Ali Farhadi, and Steve Seitz. Ranking domain-specific highlights by analyzing edited videos. In ECCV, 2014. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.33, + 0.469, + 0.383 + ], + "angle": 0, + "content": "[38] Zeren Sun, Fumin Shen, Dan Huang, Qiong Wang, Xiangbo Shu, Yazhou Yao, and Jinhui Tang. Pnp: Robust learning from noisy labels by probabilistic noise prediction. In CVPR, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.385, + 0.469, + 0.44 + ], + "angle": 0, + "content": "[39] Zeren Sun, Yazhou Yao, Xiu-Shen Wei, Yongshun Zhang, Fumin Shen, Jianxin Wu, Jian Zhang, and Heng-Tao Shen. Webly supervised fine-grained recognition-benchmark datasets and an approach. In ICCV, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.442, + 0.469, + 0.482 + ], + "angle": 0, + "content": "[40] Christian Szegedy, Vincent Vanhoucke, Sergey Ioffe, Jonathon Shlens, and Zbigniew Wojna. Rethinking the inception architecture for computer vision. 2016. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.484, + 0.469, + 0.523 + ], + "angle": 0, + "content": "[41] Hao Tang, Vivek Kwatra, Mehmet Emre Sargin, and Ullas Gargi. Detecting highlights in sports videos: Cricket as a test case. In ICME, 2011. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.525, + 0.469, + 0.565 + ], + "angle": 0, + "content": "[42] Paul Vicol, Makarand Tapaswi, Lluis Castrejon, and Sanja Fidler. Towards understanding human-centric situations from videos. 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.567, + 0.469, + 0.607 + ], + "angle": 0, + "content": "[43] Lezi Wang, Dong Liu, Rohit Puri, and Dimitris N. Metaxas. Learning trailer moments in full-length movies with co-contrastive attention. In ECCV, 2020. 2, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.609, + 0.469, + 0.65 + ], + "angle": 0, + "content": "[44] Yisen Wang, Xingjun Ma, Zaiyi Chen, Yuan Luo, Jinfeng Yi, and James Bailey. Symmetric cross entropy for robust learning with noisy labels. 2019. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.651, + 0.469, + 0.691 + ], + "angle": 0, + "content": "[45] Fanyue Wei, Biao Wang, Tiezheng Ge, Yuning Jiang, Wen Li, and Lixin Duan. Learning pixel-level distinctions for video highlight detection. In CVPR, 2022. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.693, + 0.469, + 0.733 + ], + "angle": 0, + "content": "[46] Hongxin Wei, Lei Feng, Xiangyu Chen, and Bo An. Combating noisy labels by agreement: A joint training method with co-regularization. In CVPR, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.735, + 0.469, + 0.788 + ], + "angle": 0, + "content": "[47] Haoqian Wu, Keyu Chen, Yanan Luo, Ruizhi Qiao, Bo Ren, Haozhe Liu, Weicheng Xie, and Linlin Shen. Scene consistency representation learning for video scene segmentation. In CVPR, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.79, + 0.469, + 0.844 + ], + "angle": 0, + "content": "[48] Xiaobo Xia, Tongliang Liu, Bo Han, Nannan Wang, Mingming Gong, Haifeng Liu, Gang Niu, Dacheng Tao, and Masashi Sugiyama. Part-dependent label noise: Towards instance-dependent label noise. In NeurIPS, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.846, + 0.469, + 0.899 + ], + "angle": 0, + "content": "[49] Xiaobo Xia, Tongliang Liu, Nannan Wang, Bo Han, Chen Gong, Gang Niu, and Masashi Sugiyama. Are anchor points really indispensable in label-noise learning? In NeurIPS, 2019. 3" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.146 + ], + "angle": 0, + "content": "[50] Minghao Xu, Hang Wang, Bingbing Ni, Riheng Zhu, Zhenbang Sun, and Changhu Wang. Cross-category video highlight detection via set-based learning. In ICCV, 2021. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.149, + 0.892, + 0.205 + ], + "angle": 0, + "content": "[51] Fengxiang Yang, Ke Li, Zhun Zhong, Zhiming Luo, Xing Sun, Hao Cheng, Xiaowei Guo, Feiyue Huang, Rongrong Ji, and Shaozi Li. Asymmetric co-teaching for unsupervised cross domain person re-identification. In AAAI, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.206, + 0.892, + 0.26 + ], + "angle": 0, + "content": "[52] Shuo Yang, Erkun Yang, Bo Han, Yang Liu, Min Xu, Gang Niu, and Tongliang Liu. Estimating instance-dependent bayes-label transition matrix using a deep neural network. In CVPR, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.262, + 0.892, + 0.303 + ], + "angle": 0, + "content": "[53] Ting Yao, Tao Mei, and Yong Rui. Highlight detection with pairwise deep ranking for first-person video summarization. In CVPR, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.305, + 0.892, + 0.36 + ], + "angle": 0, + "content": "[54] Chuanyi Zhang, Yazhou Yao*, Xing Xu, Jie Shao, Jingkuan Song, Zechao Li, and Zhenmin Tang. Extracting useful knowledge from noisy web images via data purification for fine-grained recognition. In ACMMM, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.362, + 0.892, + 0.403 + ], + "angle": 0, + "content": "[55] Zhilu Zhang and Mert R. Sabuncu. Generalized cross entropy loss for training deep neural networks with noisy labels. In NeurIPS, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.405, + 0.892, + 0.445 + ], + "angle": 0, + "content": "[56] Guoqing Zheng, Ahmed Hassan Awadallah, and Susan Dumas. Meta label correction for noisy label learning. In AAAI, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.447, + 0.892, + 0.488 + ], + "angle": 0, + "content": "[57] A. Zlatintsi, P. Koutras, N. Efthymiou, P. Maragos, A. Potamianos, and K. Pastra. Quality evaluation of computational models for movie summarization. 2015. 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.488 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.956 + ], + "angle": 0, + "content": "18907" + } + ] +] \ No newline at end of file diff --git a/2023/Collaborative Noisy Label Cleaner_ Learning Scene-Aware Trailers for Multi-Modal Highlight Detection in Movies/b5b9b0f2-713f-4176-89e7-acec845e8cbb_origin.pdf b/2023/Collaborative Noisy Label Cleaner_ Learning Scene-Aware Trailers for Multi-Modal Highlight Detection in Movies/b5b9b0f2-713f-4176-89e7-acec845e8cbb_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..52693ab2123498e41a1b140ff3b09f733bc5241d --- /dev/null +++ b/2023/Collaborative Noisy Label Cleaner_ Learning Scene-Aware Trailers for Multi-Modal Highlight Detection in Movies/b5b9b0f2-713f-4176-89e7-acec845e8cbb_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7e50d98e7e294f7c4af10010df0fc691b8789f627dd002bfcc0c206d215f8101 +size 3020135 diff --git a/2023/Collaborative Noisy Label Cleaner_ Learning Scene-Aware Trailers for Multi-Modal Highlight Detection in Movies/full.md b/2023/Collaborative Noisy Label Cleaner_ Learning Scene-Aware Trailers for Multi-Modal Highlight Detection in Movies/full.md new file mode 100644 index 0000000000000000000000000000000000000000..7eed8ef6761b7ed4112fce8c4c9607fec58e48cd --- /dev/null +++ b/2023/Collaborative Noisy Label Cleaner_ Learning Scene-Aware Trailers for Multi-Modal Highlight Detection in Movies/full.md @@ -0,0 +1,340 @@ +# Collaborative Noisy Label Cleaner: Learning Scene-aware Trailers for Multi-modal Highlight Detection in Movies + +Bei Gan Xiujun Shu* Ruizhi Qiao* Haoqian Wu Keyu Chen Hanjun Li Bo Ren Tencent YouTu Lab {stylegan, xiujunshu, ruizhiqiao, linuswu, yolochen, hanjunli, timren}@tencent.com + +![](images/fbfc481364a10593e809c90f8d3dda2dee5a61a400850c6328c9389a0281ca03.jpg) +Figure 1. As the preview of the movie, trailers are selected by professionals to grab an audience's attention. However, trailers are usually composed with shots sparsely selected from movies to avoid spoilers, and the audience cannot get complete highlight information. Some trailer clips convey the artistic style of the film only and lack movie storylines, disturbing the audience's impressions. In addition, different audiences may be interested in different styles of clips, which makes it challenging to learn highlights from them. + +# Abstract + +Movie highlights stand out of the screenplay for efficient browsing and play a crucial role on social media platforms. Based on existing efforts, this work has two observations: (1) For different annotators, labeling highlight has uncertainty, which leads to inaccurate and time-consuming annotations. (2) Besides previous supervised or unsupervised settings, some existing video corpora can be useful, e.g., trailers, but they are often noisy and incomplete to cover the full highlights. In this work, we study a more practical and promising setting, i.e., reformulating highlight detection as "learning with noisy labels". This setting does not require time-consuming manual annotations and can fully utilize existing abundant video corpora. First, based on movie trailers, we leverage scene segmentation to obtain complete shots, which are regarded as noisy labels. Then, we propose a Collaborative noisy Label + +Cleaner (CLC) framework to learn from noisy highlight moments. CLC consists of two modules: augmented crosspropagation (ACP) and multi-modality cleaning (MMC). The former aims to exploit the closely related audio-visual signals and fuse them to learn unified multi-modal representations. The latter aims to achieve cleaner highlight labels by observing the changes in losses among different modalities. To verify the effectiveness of CLC, we further collect a large-scale highlight dataset named MovieLights. Comprehensive experiments on MovieLights and YouTube Highlights datasets demonstrate the effectiveness of our approach. Code has been made available at: https://github.com/TencentYoutuResearch/HighlightDetection-CLC. + +# 1. Introduction + +With the growing number of new publications of movies in theaters and streaming media, audiences become even + +harder to choose their favorite one to enjoy for the next two hours. An effective solution is to watch the movie trailers before choosing the right movie. This is because trailers are generally carefully edited by filmmakers and contain the most prominent clips from the original movies. As a condensed version of full-length movies, trailers are elaborately made with highlight moments to impress the audiences. Consequently, they are high potential in serving as supervision sources to train automatic video highlight detection algorithms and facilitating the mass production of derivative works for video creators in online video platforms, e.g., YouTube and TikTok. + +Existing video highlight detection (VHD) approaches are generally trained with annotated key moments of long-form videos. However, they are not suitable to tackle the movie highlight detection task by directly learning from trailers. The edited shots in trailers are not equivalent to ground-truth highlight annotations in movies. Although a previous work [43] leverages the officially-released trailers as the weak supervision to train a highlight detector, the highlighted ness of trailer shots is extremely noisy and varies with the preference of audiences, as shown in Fig. 1. On one hand, trailers tend to be purposefully edited to avoid spoilers, thus missing key moments of the storylines. On the other hand, some less important moments in the original movies are over-emphasized in the trailers because of some artistic or commercial factors. The subjective nature of trailer shots makes them noisy for the VHD task, which is ignored by existing VHD approaches. + +To alleviate the issue, we reformulate the highlight detection task as "learning with noisy labels". Specifically, we first leverage a scene-segmentation model to obtain the movie scene boundaries. The clips containing trailers and clips from the same scenes as the trailers provide more complete storylines. They have a higher probability of being highlight moments but still contain some noisy moments. Subsequently, we introduce a framework named Collaborative noisy Label Cleaner (CLC) to learn from these pseudo-noisy labels. The framework firstly enhances the modality perceptual consistency via the augmented crosspropagation (ACP) module, which exploits closely related audio-visual signals during training. In addition, a multimodality cleaning (MMC) mechanism is designed to filter out noisy and incomplete labels. + +To support this study and facilitate benchmarking existing methods in this direction, we construct MovieLights, a Movie Highlight Detection Dataset. MovieLights contains 174 movies and the highlight moments are all from officially released trailers. The total length of these videos is over 370 hours. We conduct extensive experiments on MovieLights, in which our CLC exhibits promising results. We also demonstrate that our proposed CLC achieves significant performance-boosting over the state-of-the-art on + +the public VHD benchmarks. + +In summary, our major contributions are as follows: + +- We introduce a scene-aware paradigm to learn high-light moments in movies without any manual annotation. To the best of our knowledge, this is the first time that highlights detection is regarded as learning with noisy labels. +- We present an augmented cross-propagation to capture the interactions across modalities and a consistency loss to maximize the agreement between the different modalities. +- We incorporate a multi-modality noisy label cleaner to tackle label noise, which further improves the robustness of networks to annotation noise. +- Experiments on movie datasets and benchmark datasets validate the effectiveness of our framework. + +# 2. Related Works + +Video Highlight Detection. This task aims to identifying the interesting moments from untrimmed videos. In recent years, the videos studied for this task extend from domain-specific sport videos [41] to general videos such as social media videos [37], news [35], first-person videos [53] and vlog [25]. Most of previous works [14,21,53] interpret the video highlight detection task as a segment-level ranking problem. They compare pairwise segments from same domain video in order to learn a model that assigns highlight scores to these segments where the highlight segments receive higher scores than the non-highlight segments. MINI-Net [16] proposes to cast highlight detection as multiple instance ranking network learning. SL-Module [50] explores the highlight detection problem through Unsupervised Domain Adaptation (UDA) [32]. UMT [29] integrates highlight detection and moment retrieval into a unified framework and conduct joint optimization. PLD [45] models the video highlight detection into a pixel-level distinction estimation task. In this work, we regard highlight detection as learning with noisy labels. Joint-VA [2] also considers video highlight detection from the perspective of noise. However, it focuses on noise in features, such as videos having noisy audio when the microphone constantly has water splashing against it. We focus on specific annotation noise in video highlight detection. + +Studies on Movies and Trailers. Studies on movies and trailers have received increased attention in research. [18] introduces a comprehensive dataset for movie understanding. [4, 42] try to model the relationships among the movie characters. [7, 34, 47] focus on breaking the storylines of movies into semantically cohesive parts. Besides the studies on movies, efforts have been made to develop trailer understanding. [57] presents a movie summarization system + +and composes movie summaries in terms of user experience evaluation. [9] designs a movie trailer dataset for the evaluation of video-based recommender systems. [19] is the first approach that bridges trailers and movies and allows knowledge learned from trailers to be transferred to full movie analysis. In [43], the visual module and the temporal analysis module are respectively trained on trailers and movies. Because of the inaccessibility of public trailer-related benchmarks, we construct a new dataset (MovieLights) to detect the highlight moments in movies. + +Learning with Noisy Labels. Learning with noisy labels has been a long-standing problem in computer vision. There are three kinds of approaches to this problem. One of the most common strategies for tackling label noise is to capture the transition probabilities between noisy labels and clean labels [31, 38, 48, 49, 52, 56]. Another solution is to design robust loss functions for model training against noisy labels [8, 26, 28, 30, 44, 55]. A popular method is to design a mechanism to select clean samples or give lower weight for noise samples in the training set to reduce impact of noise [15, 17, 20, 39, 46, 51]. In this paper, we attempt to solve the problem by exploiting multi-modalities nature of movies. + +# 3. Movie Highlight Dataset + +In this paper, we aim to detect the highlight moments in movies by learning from easily accessible trailers as the noisy supervision. However, the existing movie and trailer-related benchmarks [3, 9, 18] lack sufficient functions for this task, such as the absence of full-length movies and ground-truth highlight annotations. Huang et al. [19] propose to respectively learn visual representations from trailers and temporal structure from full-length movies in their constructed Large-Scale Movie and Trailer Dataset (LSMTD). Wang [43] constructs a Trailer Moment Detection Dataset (TMDD) for detecting trailer moments from full-length movies without explicit human annotation. Both LSMTD and TMDD are not publicly available, while TMDD only contains three movie genres. + +The inaccessibility of public benchmarks motivates us to construct a new dataset, named Movie Highlight Detection Dataset (MovieLights). In particular, we purchase a set of movies from commercial channels and collect their corresponding trailers from streaming platforms such as YouTube, covering at least 25 genres to ensure content diversity. The movies and trailers are then prepossessed by shot segmentation [36] and scene segmentation [47], respectively. The resulting shots are a series of consecutive frames taken by the camera until a physical interruption, and the scenes are consecutive shots that share a semantically related theme. + +As seen in Fig. 2, to build the ground truth, we conduct Faiss [22] to obtain visual similarity matching between + +![](images/2d233ba27a1ebdcf65e78f232b433d38bef4dbaef03c4f3eba3e9bf801fb0603.jpg) +Figure 2. The labeling process of MovieLights. For the training set, we introduce a scene-aware paradigm to obtain labels automatically. For the testing set, we collect 2 sets of labels from different annotators. + +Table 1. The basic statistics of MovieLights. + +
TrainTest
Movie Number14430
Avg Durations per Movie2.19h2.14h
Avg Shot Number per Movie18521940
Avg Scene Number per Movie207193
Annotator1 Positive sample Proportion-0.27
Annotator2 Positive sample Proportion-0.30
Positive sample proportion0.350.21
+ +trailer frames and movie frames. We locate trailer moments in the movie and align them with the movie shots as annotation references. For the testing set, we collect 2 sets of moments for each movie from different workers, and these moments are annotated by different annotators independently. To ensure the consistency of results from different annotations, during the annotation procedure, all highlight moments must be related to the annotation references. Though all selected shots are relevant to the trailers, as highlight moments can be subjective, they may still vary in their saliency and time span. We calculate the intersection between every pair of moments annotated as the ground truth. However, the vast diversity of movie storylines makes the annotation challenging as it is time-consuming and requires annotators to be familiar with the movie. To collect a large amount of training data efficiently, we introduce a scene-aware paradigm to obtain the highlight moments label without any manual annotation. Specifically, we expand the shot-level annotation references to the scene span as positive samples automatically. It will capture the complete scene context of the trailer shot with movie storylines. Since the trailer shots may contain some less important moments, the acquired highlight labels are still noisy. + +As seen in Tab. 1, MovieLights contains 174 movies in full length with their official trailers and it is split into a training set with 144 movies, and a testing set with 30 movies. The content diversity is ensured by the rich domain informations (more than 25 genres) and abundant segments + +(325k shots and 36k scenes). Most movies in our dataset are between 90 to 150 minutes, and the length of the annotated moments varies from tens of seconds to several minutes. As the acquired positive highlight moments take up $35\%$ in the training set while the annotated true positives take up $21\%$ in the testing set, the difference tells the obvious existence of label noise. + +We plan to release the dataset publicly to promote further study of movie analysis. Due to copyright issues, trailers and movies will be released in the form of extracted features in visual and audio modalities. + +# 4. Approach + +# 4.1. Overview + +The overall architecture of our Collaborative noisy Label Cleaner (CLC) framework is illustrated in Fig. 3, which includes three modules: feature extraction, augmented crosspropagation (ACP), and multi-modal cleaning (MMC). + +In our framework, both the visual and audio modalities are utilized. For feature extraction and encoding, we first split the video $V$ into $T$ shots. We characterize the $i^{th}$ shots by two vectors, i.e., $\mathbf{v}_i$ for the visual features, and $\hat{\mathbf{a}}_{\mathbf{i}}$ for the audio features, where $i = 1,2,\dots,T$ . These features are extracted using pre-trained visual [11] and audio feature extractors [24]. The parameters of the two extractors are frozen during training. + +The ACP and MMC are the core components of our framework. Since the visual-audio signals in videos are closely related but do not always contribute to highlight detection, the ACP module exploits the relationship via unimodal and cross-modal interactions. Then, it learns unified multi-modal representations for highlight detection. As the obtained highlight moments after scene segmentation are noisy, the MMC module firstly observes the changes in unimodal losses, then filters the noisy labels and utilizes the clean ones for multi-modal supervised learning. Next, we will introduce details for each component. + +# 4.2. Augmented Cross-Propagation + +Cross-Propagation. To predict the highlights, the model needs to understand the storylines of the movie. Meanwhile, visual and audio inputs do not always contribute to accurate prediction. Therefore, temporal modeling and modality interaction are the keys to achieving successful highlight detection. To achieve this goal, we design our ACP module to fuse the visual-audio modalities. It involves three steps in total. + +First, in order to align the multi-modal features, we introduce $h$ which is fully-connected (FC) layer with ReLU to the $\hat{\mathbf{a}}_{i}$ such that it has same dimension as the $\mathbf{v}_i$ . + +$$ +\mathbf {a} _ {i} = h \left(\hat {\mathbf {a}} _ {\mathrm {i}}\right). \tag {1} +$$ + +Movies are composed of consecutive audio-visual clips. Therefore, to measure the highlighted ness of a given shot, one must consider the relationship of the shot with its adjacent shots. The self-attention mechanism has shown effectiveness in capturing the long-term dependencies in previous works. We leverage self-attention to capture the temporal relationship for $\mathbf{v}_i$ and $\mathbf{a}_i$ via Eq. 2 and Eq. 3, respectively. + +$$ +\mathbf {v} _ {i} ^ {s} = \operatorname {s o f t m a x} \left(\frac {\left(\mathbf {v} _ {i} W _ {1} ^ {v}\right) \left(\mathbf {v} W _ {2} ^ {v}\right) ^ {\top}}{\sqrt {d}}\right) \left(\mathbf {v} W _ {3} ^ {v}\right), \tag {2} +$$ + +$$ +\mathbf {a} _ {i} ^ {s} = \operatorname {s o f t m a x} \left(\frac {\left(\mathbf {a} _ {i} W _ {1} ^ {a}\right) \left(\mathbf {a} W _ {2} ^ {a}\right) ^ {\top}}{\sqrt {d}}\right) \left(\mathbf {a} W _ {3} ^ {a}\right), \tag {3} +$$ + +where $\mathbf{v} = [\mathbf{v}_1; \mathbf{v}_2; \ldots; \mathbf{v}_T]$ and $\mathbf{a} = [\mathbf{a}_1; \mathbf{a}_2; \ldots; \mathbf{a}_T]$ ; the scaling factor $d$ is equal to the visual/audio feature dimension and $(*)^{\top}$ denotes the transpose operator; $W^{v}$ and $W^{a}$ are learnable matrices of two modalities, which are implemented by a linear layer. Uni-modal self-attention can well capture uni-modal temporal contexts and enhance clip features within the same modality. + +Despite the above self-attention capturing the clip interactions within the uni-modality, it is critical to capture the interactions across modalities. To capture semantic associations based on multi-modal signals, we introduce cross-attention to update the features of each modality. + +$$ +\mathbf {v} _ {i} ^ {c} = \operatorname {s o f t m a x} \left(\frac {\left(\mathbf {v} _ {i} W _ {4} ^ {v}\right) \left(\mathbf {a} W _ {4} ^ {a}\right) ^ {\top}}{\sqrt {d}}\right) \left(\mathbf {a} W _ {5} ^ {a}\right), \tag {4} +$$ + +$$ +\mathbf {a} _ {i} ^ {c} = \operatorname {s o f t m a x} \left(\frac {\left(\mathbf {a} _ {i} W _ {6} ^ {a}\right) \left(\mathbf {v} W _ {5} ^ {v}\right) ^ {\top}}{\sqrt {d}}\right) \left(\mathbf {v} W _ {6} ^ {v}\right), \tag {5} +$$ + +Through cross-attention, the information from two modalities are connected. However, considering audiovisual temporal asynchrony, it is necessary to select effective information from multi-modality. We augment the relevant positive connections and dampen the irrelevant connections. The strength of these connections is measured by the cross-correlation matrix, computed by, + +$$ +\mathbf {c} ^ {v} = R e L U \left(\frac {\mathbf {v a} ^ {\top}}{\sqrt {d}}\right), \mathbf {c} ^ {a} = R e L U \left(\frac {\mathbf {a v} ^ {\top}}{\sqrt {d}}\right). \tag {6} +$$ + +For each modality, the cross-correlation matrix $\mathbf{c}$ is used to re-weight the cross-attention features. Finally, we obtain updated visual features $\bar{\mathbf{v}}_{\mathrm{i}}$ and audio features $\bar{\mathbf{a}}_{\mathrm{i}}$ by fusion of the original features, enhanced uni-modal features, and cross-modal features. + +$$ +\bar {\mathbf {v}} _ {\mathbf {i}} = f \left(\mathbf {v} _ {i}, \mathbf {v} _ {i} ^ {s}, \left(\sum_ {j = 1} ^ {T} \mathbf {c} _ {i j} ^ {v}\right) * \mathbf {v} _ {i} ^ {c}\right), \tag {7} +$$ + +$$ +\overline {{\mathbf {a}}} _ {\mathrm {i}} = f \left(\mathbf {a} _ {i}, \mathbf {a} _ {i} ^ {s}, \left(\sum_ {j = 1} ^ {T} \mathbf {c} _ {i j} ^ {a}\right) * \mathbf {a} _ {i} ^ {c}\right), \tag {8} +$$ + +![](images/8d3f2c2c28d1eaf9ba040dba693a9e53fc538b71c299fb54f14b73a8cd5f1b5a.jpg) +Figure 3. Overview of the proposed CLC. It includes three modules: feature extraction, augmented cross-propagation (ACP), and multi-modal cleaning (MMC). The visual and audio modalities of the input video are represented as vectors by the feature extraction module. Then the features are augmented by ACP module to capture semantic associations across modalities. MMC is used to filter outs noisy and incomplete labeling with additional uni-modal branches. During inference, we remove the uni-modal branches and only rely on the prediction of multi-modal branch. More details of the CLC are shown in Sec. 4.1 + +where $f$ is the fusion function consisting of FC layers and ReLU to further project the features. + +Clearly, the cross-correlation matrix will assign large weights to clips that are relevant to the other modality. In addition, the ReLU activation in Eq. 6 cuts off connections with negative similarity values and only relevant positive connections would be preserved. By the above operations, the original feature will be infused with richer information. Consistency Loss. Multi-modal inputs help to comprehensively learn by integrating different aspects and boosting model performance. However, they are not fully exploited because some modality-specific features may be weakened in the fusion even when the multi-modal model outperforms its uni-modal counterpart. In this work, we provide parallel branches for each modality separately to obtain uni-modal prediction during training. As seen in Fig. 3, all branches share the same clip feature extracted from ACP and the feature is fed into different branches independently. In each branch, we develop a temporal model $G$ to obtain its prediction score $y_{i}$ of the $i^{th}$ clip being a highlight as follows: + +$$ +\mathbf {y} _ {i} ^ {M M} = G \left(\bar {\mathbf {v}} _ {\mathbf {i}}, \bar {\mathbf {a}} _ {\mathbf {i}}\right), \mathbf {y} _ {i} ^ {U M _ {v}} = G \left(\bar {\mathbf {v}} _ {\mathbf {i}}\right), \mathbf {y} _ {i} ^ {U M _ {a}} = G \left(\bar {\mathbf {a}} _ {\mathbf {i}}\right). \tag {9} +$$ + +where $MM$ is multi-modal branch; $UM_{v}$ is visual branch and $UM_{a}$ is audio branch. + +However, the gap between the different modalities would bring instability in the joint optimization process. We employ the auxiliary consistency loss to guarantee consistency between the different modalities. Given the multi-modal features of a clip, they are consistent if they share the same prediction. Specifically, the consistency loss is defined as + +the cross-entropy between the multi-modal prediction probability $\mathbf{y}_i^{MM}$ and uni-modal prediction probability $\mathbf{y}_i^{UM}$ : + +$$ +\mathcal {L} _ {\text {c o n s}} = - \left(\sum_ {i = 1} ^ {N} \mathbf {y} _ {i} ^ {U M _ {v}} \log \mathbf {y} _ {i} ^ {M M} + \sum_ {i = 1} ^ {N} \mathbf {y} _ {i} ^ {U M _ {a}} \log \mathbf {y} _ {i} ^ {M M}\right), \tag {10} +$$ + +where $N$ is the number of samples in a batch. + +This consistency loss not only implicitly enhances uni-modal information, but also explicitly guides the multimodal branch to robust supervision. The added uni-modal branches are only utilized in the training phase and are disabled during the inference stage. + +# 4.3. Multi-modal Cleaning + +Multi-modality Sample Cleaning. Annotation noise is inevitable in VHD due to subjectivity depending on the users and annotators. In this paper, we argue that highlight detection should be regarded as Learning with Noisy Labels. To alleviate the performance drop caused by noisy labels, we adopt a multi-modality collaborative cleaning to adaptively filter noisy samples with noisy modality information. + +Firstly, we maintain multiple outputs simultaneously which are predicted by different branches. The uni-modal branches independently select clean samples based on the low-loss criterion in which instances with lower losses are treated clean samples. Contrary to existing noisy sample selection methods [12, 54], which directly discard high-loss samples, we keep all samples to train the uni-modal branches. Then, we update the multi-modal branch using only clean samples selected by both uni-modal branches in the back-propagation. The samples for the multi-modal + +branch training are selected dynamically while all samples participate in the uni-mode branches training. In this way, MMC obtains information on all samples to avoid the model defecting to favoring easy samples. More details of MMC are presented in the supplementary material. + +Each branch has its own loss function. For the uni-modal branch, we employ the cross-entropy loss with all samples as follows: + +$$ +\mathcal {L} ^ {U M _ {v}} = \mathcal {L} _ {\mathrm {c e}} ^ {U M _ {v}} = - \sum_ {i = 1} ^ {N} \mathbf {g} _ {i} \log \mathbf {y} _ {i} ^ {U M _ {v}}, \tag {11} +$$ + +$$ +\mathcal {L} ^ {U M _ {a}} = \mathcal {L} _ {\mathrm {c e}} ^ {U M _ {a}} = - \sum_ {i = 1} ^ {N} \mathbf {g} _ {i} \log \mathbf {y} _ {i} ^ {U M _ {a}}. \tag {12} +$$ + +For the multi-modal branch, its cross-entropy loss and consistency loss are updated with a re-weighting scheme with clean samples. + +$$ +\mathcal {L} _ {\mathrm {c e}} ^ {\mathrm {M M}} = - \sum_ {i = 1} ^ {N ^ {\prime}} \mathbf {g} _ {i} \log \mathbf {y} _ {i} ^ {M M}, \tag {13} +$$ + +$$ +\mathcal {L} _ {\text {c o n s}} ^ {M M} = - \left(\sum_ {i = 1} ^ {N ^ {\prime}} \mathbf {y} _ {i} ^ {U M _ {v}} \log \mathbf {y} _ {i} ^ {M M} + \sum_ {i = 1} ^ {N ^ {\prime}} \mathbf {y} _ {i} ^ {U M _ {a}} \log \mathbf {y} _ {i} ^ {M M}\right), \tag {14} +$$ + +$$ +\mathcal {L} ^ {\mathrm {M M}} = \mathcal {L} _ {\mathrm {c e}} ^ {\mathrm {M M}} + \beta \mathcal {L} _ {\mathrm {c o n s}} ^ {M M}, \tag {15} +$$ + +where $\mathbf{g}_i$ and $\mathbf{y}_i^*$ denote the ground-truth and predicted probability of the $i^{th}$ clip, respectively; $N$ is the number of samples in a batch and $N'$ is the number of samples selected by MMC; and $\beta$ are designed to balance different loss terms. + +Post processing. In experiments we observe that noise makes jitter prediction curves along the temporal dimension, which may cause discontinuous thresholds for highlight selection. Therefore, we apply a median filter to smooth the prediction curves. Supposing $\mathbf{y} = [\mathbf{y}_1; \mathbf{y}_2; \dots; \mathbf{y}_T]$ is the original curve predicted by the CLC, the smoothed curve $\mathbf{s} = [\mathbf{s}_1; \mathbf{s}_2; \dots; \mathbf{s}_T]$ is given by: + +$$ +\mathbf {s} _ {i} = \left\{ \begin{array}{l l} \operatorname {M e d} \left(\mathbf {y} _ {i - k}, \mathbf {y} _ {i + k}\right), & k < i \leq T - k \\ \mathbf {y} _ {i}, & \text {o t h e r w i s e} \end{array} \right. \tag {16} +$$ + +where $k$ is the window size, and "Med" denotes the median filter. + +# 5. Experiment + +# 5.1. Datasets and Experimental Settings + +Datasets. We evaluate our CLC on the constructed dataset MovieLights and public YouTube Highlights dataset [37]. MovieLights is split into training and testing sets, each containing 144 and 30 movies respectively. + +We split movies into shots. The movie features are represented at shot-level. We use the middle frame of each + +shot to extract its visual feature with ViT [10] pre-trained by CLIP [33]. We align timestamps of audio clips with the visual shots, and sample the audio clip of each with $16K$ Hz sampling rate and 512 windowed signal length. The resulted shot-level audio features are obtained with the PANN audio network [24] pretrained on AudioSet [13]. + +The YouTube Highlights contains six distinct categories with a total 422 videos currently available. Following the practice of prior efforts, we train a highlight detector for each category. YouTube Highlights provides two annotations: Harvested Highlight and Mturk Highlight. In the Harvested annotation, the match label specifies if each clip is matched in the edited video, where 1 denotes matched, -1 denotes unmatched and 0 denotes the borderline cases. In the Mturk annotation, the highlight labels are marked by multiple turkers of different styles, making it noisier than the Harvested annotation. + +On YouTube Highlights, we use the same protocol and data preprocessing as [29]. It obtain clip-level visual features and optical flow features using I3D [6] pre-trained on Kinetics400 [23]. It use a PANN audio network [24] pretrained on AudioSet [13] to obtain audio features that align with the visual clips. Frame-level features are average-pooled within each clip for both audio and visual features to generate a clip-level feature. Since each feature vector spans 32 consecutive frames, we follow [29] and consider the feature vector corresponded to a clip if their overlap is more than $50\%$ . + +Benchmarks. To better inspect the robustness of our CLC against noisy labels, we also apply label perturbations in training set of YouTube Highlights while keeping the ground-truths in the testing set unchanged. YouTube Highlights has two benchmarks for comparison. 1) Harvested with matched: we regard the clips labeled with matched as the highlighted clips and this is the same setting as in previous works [29, 50]. 2) Harvested with matched and borderline: clips labeled with matched and borderline are treated as the highlight moments. In this benchmark, the training set contains some highlighted clips with weak confidence. We select the clips whose mturk-label is over score 1 as the highlighted clips, which means that at least one turker selects the clip as a highlight. There are labeled by different types of annotators between Mturk and the test set and bring greater noise. We select the clips whose mturk-label is over score 1 as the highlighted clips, which means that at least one turker selects the clip as a highlight. Due the labeling gap between the Mturk annotation in the training set and the clean testing set, this benchmark is even noisier. + +Baselines. We introduce CLC-, the degenerated version of CLC, as a baseline. Similar to CLC, CLC- is a Bi-LSTM-based model and takes the temporal sequence of shot features as input but lacks the modules of augmented cross-propagation and multi-modality noisy label cleaner. + +Table 2. Results on MovieLights. + +
MethodsModalitymAP
GIFs [14]V25.48
SL-Module [50]V32.34
SL-Module [50]VA34.27
UMT [29]VA38.7
CLC-VA39.65
CLC- w/ SCE [44]VA39.83
CLC- w/ LS [40]VA40.49
CLCVA43.88
+ +Evaluation Metric. We adopt mean Average Precision (mAP) as the evaluation metric for MovieLights and YouTube Highlights. Considering that a highlighted moment in one video is not necessarily more interesting than non-highlight moments in other videos, we evaluate on each test video independently and report the averaged results. + +Implementation Details. On the MovieLights, we train our model using SGD, with a learning rate of 0.01. We train for 50 epochs. Before the cross-attention modules, we project each modality into a vector of 512 dimension. The key, query, and value vectors all share the same dimension. Weight $\beta$ in Eq. 15 is empirically set to 0.1 and window size $k$ in Eq. 16 is set to 9. + +# 5.2. Results on MovieLights + +On MovieLights, we train our model with the noisy pseudo labels. To compare with previous state-of-the-art highlight detection works, we train UMT [29] and SL-Module [50] using the same protocol and data preprocessing as in CLC. We also compare with Video2GIF [14] using its off-the-shelf tool1. The upper part of Tab. 2 illustrates the significant performance gain of CLC. + +To demonstrate the advantages of CLC in learning with noisy labels, we make comparisons with two main-stream label noise approaches: Label Smoothing [40] and SCE loss [44]. Specifically, we insert Label Smoothing or SCE into our CLC- framework to create two baseline VHD methods to tackle label noise. The bottom part of Tab. 2 shows that CLC outperforms the two baseline methods by a notable margin, indicating that our augmented cross-propagation and multi-modality noisy label cleaner are more effective than vanilla label noise approaches in VHD tasks. + +# 5.3. Ablation Study + +In this experiment, we analyze the impact of each module. The results are summarized in Tab. 3. + +Multi-modality Sample Cleaning. We first inspect the impact of the multi-modality sample filter module because we are primary concerned with learning with noisy labels + +Table 3. Ablation results of MoiveLightts. + +
MMSCCPCLPPmAP
××××39.65
✓×××41.69
✓✓××42.79
✓✓✓×43.22
✓✓✓✓43.88
+ +Table 4. Results on YouTube Highlights. + +
Methodsdoggym.park.ska.ski.surf.Avg.
GIFs [14]30.833.55455.432.854.146.4
LSVM [37]60.041.061.062.036.061.053.6
HighlightMe [5]63737264526264
MINI-Net [16]58.261.770.272.258.760.164.4
CHD [1]60.671.174.249.868.268.565.4
Trail [43]63.382.562.352.974.579.369.1
SL-Module [50]70.853.277.272.566.176.269.3
Joint-VA [2]64.571.980.86273.278.371.8
PLD [45]74.970.277.957.570.77973
CO-AV [27]60.9668974.16981.174.7
UMT [29]65.975.281.671.872.382.774.9
CLC(ours)70.579.483.983.579.583.680.1
+ +in VHD. The $2\%$ performance gain from the module over baseline shows the importance of filtering noisy sample. Based on the this observation, the subsequent ablation experiments are conducted under the setting of noise filtering. Cross-Propagation. We then evaluate the impact of the feature augmentation module. Compared to naive feature concatenation, the models with augmented features show superior performance. The results validate that feature augmentation module can better explore the complementary information from different modalities and suppress the mutual disturbance of desynchronized uni-modal information. + +Consistency Loss. We examine the contribution of the consistency loss. The results demonstrate that the employment of the auxiliary multi-modal constraint further increases the model robustness. This consistency loss not only implicitly enhances uni-modal information, but also explicitly guides the multi-modal branch to better learning. + +Post Processing. We compare the highlight prediction curves with and without the post processing median filter in Fig. 4(a,b), which prevents disruptive prediction variation and improves the overall performance. + +# 5.4. Results on YouTube Highlights + +We conduct experiments on the public video highlight detection benchmarks YouTube Highlights, including six domain video datasets, to verify the generalization ability of CLC. + +The setting of Harvested Highlight with matched is consistent with the previous work [29]. As shown in Tab. 4, CLC achieves state-of-the-art performance on the YouTube + +![](images/41c5bd013edb20170b65a2247d3c6c90f2c95d9a7e47fd267d16acb828ca2197.jpg) +Figure 4. (a, b) Comparison of original prediction curve with filtered prediction curve. (c) Results on YouTube Highlights with noisy label. + +![](images/08a751add36e3c82c96298e24a9351831f506cc8c00ebff9c438a4a59c78f4b9.jpg) + +![](images/7b8c3bf7b26235ea5d7dec63ccf8e99eb352033c68bd6c1fa05cd3bc0aa7317c.jpg) + +Table 5. Results on YouTube Highlights with Noisy Label. + +
AnnotationNoiseMethodsdoggym.park.ska.ski.surf.Avg.
Harvested matchedcleanUMT [29]65.9075.2081.6071.8072.3082.7074.90
Harvested matchedcleanCLC70.5179.4383.8583.5179.4683.5680.05 (↑ 5.15)
Harvested borderlineslight noiseUMT [29]65.9374.3181.5871.8470.2482.4674.39
Harvested borderlineslight noiseCLC69.4180.7378.5085.3681.1183.1679.71 (↑ 5.32)
Mturksevere noiseUMT [29]63.7876.1675.0273.6269.9981.5973.36
Mturksevere noiseCLC66.9280.4485.9282.3378.0581.7279.22 (↑ 5.86)
+ +![](images/6edce3294648c6dd7f15fb78235913315f4342ffd14e1622ebcd97ae6035f2fd.jpg) +Figure 5. The highlight moments selected by our CLC from Skyfall and Sherlock Holmes. Top: Bond gives chase to a professional hitman by car to find a classified hard drive, and then a firefight erupted in the market. Bottom: Holmes and Blackwood are facing off, and then Blackwood reaches for a weapon to kill Holmes, but accidentally trips off a scaffolding and falls to his death. + +Highlights, outperforming the existing multi-modal highlight detection methods in the average metric across all categories. Specifically, CLC achieves best performance in three out of the six categories, while maintaining reasonably competitive performance in the other three categories. These results support our claim that highlight detection should be regarded as learning with noisy labels. + +To quantify how CLC is robust to different levels of label noise, we inspect the settings of Harvested Highlight with matched and borderline and Mturk Highlight, which are perturbed by varying degrees of label noise in YouTube Highlights. + +Tab. 5 exhibits the performances of CLC and UMT [29] at different noise levels. As the noise level increases, the VHD task becomes more difficult, but the performance superiority of our CLC over UMT becomes even more obvious. It is illustrated in Fig. 4(c) that compared with the most recent state-of-the-art UMT, our model can achieve better performance even when disturbed by severe label noise. + +# 5.5. Visualization + +As shown in Fig. 5, we present some visualization examples of the detected highlight clips in MovieLights by CLC. The examples clearly show that the prediction of CLC is in accordance with user expectation. We will provide more examples in the supplementary material. + +# 6. Conclusion + +In this study, we present Collaborative noisy Label Cleaner (CLC), a novel framework to handle noisy labels in video highlight detection. We make use of the augmented cross-propagation module to better enhance network robustness and multi-modality cleaning to achieve cleaner highlight labels by observing the loss changes of different modalities. We demonstrate the state-of-the-art performance of our method with extensive experiments on MovieLights and YouTube Highlights datasets. In future work, we are interested in extending the proposed mechanisms to other video-understanding tasks such as scene segmentation and video temporal grounding. + +# References + +[1] Taivanbat Badamdorj, Mrigank Rochan, Yang Wang, and Li Cheng. Contrastive learning for unsupervised video highlight detection. In CVPR, 2022. 7 +[2] Taivanbat Badamdorj, Mrigank Rochan, Yang Wang, Jiaze Wang, and Li Cheng. Joint visual and audio learning for video highlight detection. In ICCV, 2021. 2, 7 +[3] Max Bain, Arsha Nagrani, Andrew Brown, and Andrew Zisserman. Condensed movies: Story based retrieval with contextual embeddings. In ACCV, 2020. 3 +[4] David Bamman, Brendan O'Connor Noah, and A. Smith. Learning latent personas of film characters. 2014. 2 +[5] Uttaran Bhattacharya, Gang Wu, Stefano Petrangeli, Viswanathan Swaminathan, and Dinesh Manocha. Highlighte: Detecting highlights from human-centric videos. In ICCV, 2021. 7 +[6] Joao Carreira and Andrew Zisserman. Quo vadis, action recognition? a new model and the kinetics dataset. 2017. 6 +[7] Shixing Chen, Xiaohan Nie, David Fan, Dongqing Zhang, Vimal Bhat, and Raffay Hamid. Shot contrastive self-supervised learning for scene boundary detection. 2021. 2 +[8] Hao Cheng, Zhaowei Zhu, Xingyu Li, Yifei Gong, Xing Sun, and Yang Liu. Learning with instance-dependent label noise: A sample sieve approach. In ICLR, 2021. 3 +[9] Yashar Deldjoo, Mihai Gabriel Constantin, and Bogdan Ionescu. Mmtf-14k: A multifaceted movie trailer feature dataset for recommendation and retrieval. In ACMMM, 2018. 3 +[10] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Xiaohua Zhai Dirk Weissenborn, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. 2021. 6 +[11] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. 2021. 4 +[12] Curtis G, Northcutt, Lu Jiang, and Isaac L. Chuang. Confident learning: Estimating uncertainty in dataset labels. 2021. 5 +[13] Jort F. Gemmeke, Daniel P. W. Ellis, Dylan Freedman, Aren Jansen, Wade Lawrence, R. Channing Moore, Manoj Plakal, and Marvin Ritter. Audio set: An ontology and human-labeled dataset for audio events. 2017. 6 +[14] Michael Gygli, Yale Song, and Liangliang Cao. Video2gif: Automatic generation of animated gifs from video. In CVPR, 2016. 2, 7 +[15] Bo Han, Quanming Yao, Xingrui Yu, Gang Niu, Miao Xu, Weihua Hu, Ivor Tsang, and Masashi Sugiyama. Co-teaching: Robust training of deep neural networks with extremely noisy labels. In NeurIPS, 2018. 3 +[16] Fa-Ting Hong, Xuanteng Huang, Wei-Hong Li, and WeiShi Zheng. Mini-net: Multiple instance ranking network for video highlight detection. In ECCV, 2020. 2, 7 + +[17] Jinchi Huang, Lie Qu, Rongfei Jia, and Binqiang Zhao. O2unet: A simple noisy label detection approach for deep neural network. In ICCV, 2019. 3 +[18] Qingqiu Huang, Yu Xiong, Anyi Rao, Jiaze Wang, and Dahua Lin. Movienet: A holistic dataset for movie understanding. In ECCV, 2020. 2, 3 +[19] Qingqiu Huang, Yuanjun Xiong, Yu Xiong, Yuqi Zhang, and Dahua Lin. From trailers to storylines: An efficient way to learn from movies. arXiv preprint arXiv:1806.05341, 2018. 3 +[20] Lu Jiang, Zhengyuan Zhou, Thomas Leung, Li-Jia Li, and Li Fei-Fei. Mentornet: Learning data-driven curriculum for very deep neural networks on corrupted labels. In ICML, 2018. 3 +[21] Yifan Jiao, Zhetao Li, Shucheng Huang, Xiaoshan Yang, Bin Liu, and Tianzhu Zhang. Three-dimensional attention-based deep ranking model for video highlight detection. In IEEE Transactions on Multimedia, page 2693-2705, 2018. 2 +[22] Jeff Johnson, Matthijs Douze, and Hervé Jégou. Billionscale similarity search with gpus. IEEE Transactions on Big Data, page 1-1, 2019. 3 +[23] Will Kay, Joao Carreira, Karen Simonyan, Brian Zhang, Chloe Hillier, Sudheendra Vijayanarasimhan, Fabio Viola, Tim Green, Trevor Back, Paul Natsev, Mustafa Suleyman, and Andrew Zisserman. The kinetics human action video dataset. arXiv preprint arXiv:1705.06950, 2017. 6 +[24] Qiuqiang Kong, Yin Cao, Turab Iqbal, Yuxuan Wang, Wenwu Wang, and Mark D Plumbley. Panns: Large-scale pretrained audio neural networks for audio pattern recognition. IEEE/ACM Transactions on Audio, Speech, and Language Processing, 28:2880-2894, 2020. 4, 6 +[25] Jie Lei, Tamara L. Berg, and Mohit Bansal. Qvhighlights: Detecting moments and highlights in videos via natural language queries. In NeurIPS, 2021. 2 +[26] Junnan Li, Richard Socher, and Steven C.H. Hoi. Dividemix: Learning with noisy labels as semi-supervised learning. 2020. 3 +[27] Shuaicheng Li, Feng Zhang, Kunlin Yang, Lingbo Liu, Shinan Liu, Jun Hou, and Shuai Yi. Probing visual-audio representation for video highlight detection via hard-pairs guided contrastive learning. arXiv preprint arXiv:2206.10157, 2022. 7 +[28] Yang Liu and Hongyi Guo. Peer loss functions: Learning from noisy labels without knowing noise rates. 2020. 3 +[29] Ye Liu, Siyuan Li, Yang Wu, Chang Wen Chen, Ying Shan, and Xiaohu Qie. Umt: Unified multi-modal transformers for joint video moment retrieval and highlight detection. In CVPR, 2022. 2, 6, 7, 8 +[30] Xingjun Ma, Hanxun Huang, Yisen Wang, Simone Romano, Sarah Erfani, and James Bailey. Normalized loss functions for deep learning with noisy labels. 2020. 3 +[31] Curtis G. Northcutt, Lu Jiang, and Isaac L. Chuang. Confident learning: Estimating uncertainty in dataset labels. In JAIR, 2019. 3 +[32] Sinno Jialin Pan and Qiang Yang. A survey on transfer learning. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, page 1345-1359, 2010. 2 + +[33] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Sandhini Agarwal Gabriel Goh, Girish Sastry, Amanda Askell, Pamela Mishkin, and et al. Jack Clark. Learning transferable visual models from natural language supervision. arXiv preprint arXiv:2103.00020, 2021. 6 +[34] Anyi Rao, Linning Xu, Yu Xiong, Guodong Xu, Qingqiu Huang, Bolei Zhou, and Dahua Lin. A local-to-global approach to multi-modal movie scene segmentation. 2020. 2 +[35] Yale Song, Yahoo Labs, New York, Jordi Vallmitjana, Amanda Stent, and Alejandro Jaime. Tvsum: Summarizing web videos using titles. In CVPR, 2015. 2 +[36] Tomáš Souček and Jakub Lokoc. Transnet v2: An effective deep network architecture for fast shot transition detection. arXiv preprint arXiv:2008.04838, 2020. 3 +[37] Min Sun, Ali Farhadi, and Steve Seitz. Ranking domain-specific highlights by analyzing edited videos. In ECCV, 2014. 2, 6, 7 +[38] Zeren Sun, Fumin Shen, Dan Huang, Qiong Wang, Xiangbo Shu, Yazhou Yao, and Jinhui Tang. Pnp: Robust learning from noisy labels by probabilistic noise prediction. In CVPR, 2022. 3 +[39] Zeren Sun, Yazhou Yao, Xiu-Shen Wei, Yongshun Zhang, Fumin Shen, Jianxin Wu, Jian Zhang, and Heng-Tao Shen. Webly supervised fine-grained recognition-benchmark datasets and an approach. In ICCV, 2021. 3 +[40] Christian Szegedy, Vincent Vanhoucke, Sergey Ioffe, Jonathon Shlens, and Zbigniew Wojna. Rethinking the inception architecture for computer vision. 2016. 7 +[41] Hao Tang, Vivek Kwatra, Mehmet Emre Sargin, and Ullas Gargi. Detecting highlights in sports videos: Cricket as a test case. In ICME, 2011. 2 +[42] Paul Vicol, Makarand Tapaswi, Lluis Castrejon, and Sanja Fidler. Towards understanding human-centric situations from videos. 2018. 2 +[43] Lezi Wang, Dong Liu, Rohit Puri, and Dimitris N. Metaxas. Learning trailer moments in full-length movies with co-contrastive attention. In ECCV, 2020. 2, 3, 7 +[44] Yisen Wang, Xingjun Ma, Zaiyi Chen, Yuan Luo, Jinfeng Yi, and James Bailey. Symmetric cross entropy for robust learning with noisy labels. 2019. 3, 7 +[45] Fanyue Wei, Biao Wang, Tiezheng Ge, Yuning Jiang, Wen Li, and Lixin Duan. Learning pixel-level distinctions for video highlight detection. In CVPR, 2022. 2, 7 +[46] Hongxin Wei, Lei Feng, Xiangyu Chen, and Bo An. Combating noisy labels by agreement: A joint training method with co-regularization. In CVPR, 2020. 3 +[47] Haoqian Wu, Keyu Chen, Yanan Luo, Ruizhi Qiao, Bo Ren, Haozhe Liu, Weicheng Xie, and Linlin Shen. Scene consistency representation learning for video scene segmentation. In CVPR, 2022. 2, 3 +[48] Xiaobo Xia, Tongliang Liu, Bo Han, Nannan Wang, Mingming Gong, Haifeng Liu, Gang Niu, Dacheng Tao, and Masashi Sugiyama. Part-dependent label noise: Towards instance-dependent label noise. In NeurIPS, 2020. 3 +[49] Xiaobo Xia, Tongliang Liu, Nannan Wang, Bo Han, Chen Gong, Gang Niu, and Masashi Sugiyama. Are anchor points really indispensable in label-noise learning? In NeurIPS, 2019. 3 + +[50] Minghao Xu, Hang Wang, Bingbing Ni, Riheng Zhu, Zhenbang Sun, and Changhu Wang. Cross-category video highlight detection via set-based learning. In ICCV, 2021. 2, 6, 7 +[51] Fengxiang Yang, Ke Li, Zhun Zhong, Zhiming Luo, Xing Sun, Hao Cheng, Xiaowei Guo, Feiyue Huang, Rongrong Ji, and Shaozi Li. Asymmetric co-teaching for unsupervised cross domain person re-identification. In AAAI, 2020. 3 +[52] Shuo Yang, Erkun Yang, Bo Han, Yang Liu, Min Xu, Gang Niu, and Tongliang Liu. Estimating instance-dependent bayes-label transition matrix using a deep neural network. In CVPR, 2022. 3 +[53] Ting Yao, Tao Mei, and Yong Rui. Highlight detection with pairwise deep ranking for first-person video summarization. In CVPR, 2016. 2 +[54] Chuanyi Zhang, Yazhou Yao*, Xing Xu, Jie Shao, Jingkuan Song, Zechao Li, and Zhenmin Tang. Extracting useful knowledge from noisy web images via data purification for fine-grained recognition. In ACMMM, 2021. 5 +[55] Zhilu Zhang and Mert R. Sabuncu. Generalized cross entropy loss for training deep neural networks with noisy labels. In NeurIPS, 2018. 3 +[56] Guoqing Zheng, Ahmed Hassan Awadallah, and Susan Dumas. Meta label correction for noisy label learning. In AAAI, 2021. 3 +[57] A. Zlatintsi, P. Koutras, N. Efthymiou, P. Maragos, A. Potamianos, and K. Pastra. Quality evaluation of computational models for movie summarization. 2015. 2 \ No newline at end of file diff --git a/2023/Collaborative Noisy Label Cleaner_ Learning Scene-Aware Trailers for Multi-Modal Highlight Detection in Movies/images.zip b/2023/Collaborative Noisy Label Cleaner_ Learning Scene-Aware Trailers for Multi-Modal Highlight Detection in Movies/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..937a0dd3668a23cdd5d164e0786afcbec8c53486 --- /dev/null +++ b/2023/Collaborative Noisy Label Cleaner_ Learning Scene-Aware Trailers for Multi-Modal Highlight Detection in Movies/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bb5bce177bc871111bf16cdeadda70b9d8e969b1430b8abe361a6151c723f3c4 +size 613634 diff --git a/2023/Collaborative Noisy Label Cleaner_ Learning Scene-Aware Trailers for Multi-Modal Highlight Detection in Movies/layout.json b/2023/Collaborative Noisy Label Cleaner_ Learning Scene-Aware Trailers for Multi-Modal Highlight Detection in Movies/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..4dfa6baed5a518a3c3b00a7041b38c2da59e09df --- /dev/null +++ b/2023/Collaborative Noisy Label Cleaner_ Learning Scene-Aware Trailers for Multi-Modal Highlight Detection in Movies/layout.json @@ -0,0 +1,7838 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 79, + 103, + 515, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 79, + 103, + 515, + 140 + ], + "spans": [ + { + "bbox": [ + 79, + 103, + 515, + 140 + ], + "type": "text", + "content": "Collaborative Noisy Label Cleaner: Learning Scene-aware Trailers for Multi-modal Highlight Detection in Movies" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 70, + 161, + 523, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 161, + 523, + 205 + ], + "spans": [ + { + "bbox": [ + 70, + 161, + 523, + 205 + ], + "type": "text", + "content": "Bei Gan Xiujun Shu* Ruizhi Qiao* Haoqian Wu Keyu Chen Hanjun Li Bo Ren Tencent YouTu Lab {stylegan, xiujunshu, ruizhiqiao, linuswu, yolochen, hanjunli, timren}@tencent.com" + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 74, + 225, + 517, + 422 + ], + "blocks": [ + { + "bbox": [ + 74, + 225, + 517, + 422 + ], + "lines": [ + { + "bbox": [ + 74, + 225, + 517, + 422 + ], + "spans": [ + { + "bbox": [ + 74, + 225, + 517, + 422 + ], + "type": "image", + "image_path": "fbfc481364a10593e809c90f8d3dda2dee5a61a400850c6328c9389a0281ca03.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 45, + 429, + 547, + 473 + ], + "lines": [ + { + "bbox": [ + 45, + 429, + 547, + 473 + ], + "spans": [ + { + "bbox": [ + 45, + 429, + 547, + 473 + ], + "type": "text", + "content": "Figure 1. As the preview of the movie, trailers are selected by professionals to grab an audience's attention. However, trailers are usually composed with shots sparsely selected from movies to avoid spoilers, and the audience cannot get complete highlight information. Some trailer clips convey the artistic style of the film only and lack movie storylines, disturbing the audience's impressions. In addition, different audiences may be interested in different styles of clips, which makes it challenging to learn highlights from them." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 143, + 485, + 192, + 498 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 485, + 192, + 498 + ], + "spans": [ + { + "bbox": [ + 143, + 485, + 192, + 498 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 504, + 290, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 504, + 290, + 696 + ], + "spans": [ + { + "bbox": [ + 46, + 504, + 290, + 696 + ], + "type": "text", + "content": "Movie highlights stand out of the screenplay for efficient browsing and play a crucial role on social media platforms. Based on existing efforts, this work has two observations: (1) For different annotators, labeling highlight has uncertainty, which leads to inaccurate and time-consuming annotations. (2) Besides previous supervised or unsupervised settings, some existing video corpora can be useful, e.g., trailers, but they are often noisy and incomplete to cover the full highlights. In this work, we study a more practical and promising setting, i.e., reformulating highlight detection as \"learning with noisy labels\". This setting does not require time-consuming manual annotations and can fully utilize existing abundant video corpora. First, based on movie trailers, we leverage scene segmentation to obtain complete shots, which are regarded as noisy labels. Then, we propose a Collaborative noisy Label" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 486, + 547, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 486, + 547, + 653 + ], + "spans": [ + { + "bbox": [ + 306, + 486, + 547, + 653 + ], + "type": "text", + "content": "Cleaner (CLC) framework to learn from noisy highlight moments. CLC consists of two modules: augmented crosspropagation (ACP) and multi-modality cleaning (MMC). The former aims to exploit the closely related audio-visual signals and fuse them to learn unified multi-modal representations. The latter aims to achieve cleaner highlight labels by observing the changes in losses among different modalities. To verify the effectiveness of CLC, we further collect a large-scale highlight dataset named MovieLights. Comprehensive experiments on MovieLights and YouTube Highlights datasets demonstrate the effectiveness of our approach. Code has been made available at: https://github.com/TencentYoutuResearch/HighlightDetection-CLC." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 669, + 387, + 681 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 669, + 387, + 681 + ], + "spans": [ + { + "bbox": [ + 306, + 669, + 387, + 681 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 689, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 547, + 713 + ], + "type": "text", + "content": "With the growing number of new publications of movies in theaters and streaming media, audiences become even" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 703, + 135, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 703, + 135, + 712 + ], + "spans": [ + { + "bbox": [ + 58, + 703, + 135, + 712 + ], + "type": "text", + "content": "*Corresponding author." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18898" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 216 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 216 + ], + "type": "text", + "content": "harder to choose their favorite one to enjoy for the next two hours. An effective solution is to watch the movie trailers before choosing the right movie. This is because trailers are generally carefully edited by filmmakers and contain the most prominent clips from the original movies. As a condensed version of full-length movies, trailers are elaborately made with highlight moments to impress the audiences. Consequently, they are high potential in serving as supervision sources to train automatic video highlight detection algorithms and facilitating the mass production of derivative works for video creators in online video platforms, e.g., YouTube and TikTok." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 218, + 289, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 218, + 289, + 422 + ], + "spans": [ + { + "bbox": [ + 46, + 218, + 289, + 422 + ], + "type": "text", + "content": "Existing video highlight detection (VHD) approaches are generally trained with annotated key moments of long-form videos. However, they are not suitable to tackle the movie highlight detection task by directly learning from trailers. The edited shots in trailers are not equivalent to ground-truth highlight annotations in movies. Although a previous work [43] leverages the officially-released trailers as the weak supervision to train a highlight detector, the highlighted ness of trailer shots is extremely noisy and varies with the preference of audiences, as shown in Fig. 1. On one hand, trailers tend to be purposefully edited to avoid spoilers, thus missing key moments of the storylines. On the other hand, some less important moments in the original movies are over-emphasized in the trailers because of some artistic or commercial factors. The subjective nature of trailer shots makes them noisy for the VHD task, which is ignored by existing VHD approaches." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 424, + 289, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 424, + 289, + 604 + ], + "spans": [ + { + "bbox": [ + 46, + 424, + 289, + 604 + ], + "type": "text", + "content": "To alleviate the issue, we reformulate the highlight detection task as \"learning with noisy labels\". Specifically, we first leverage a scene-segmentation model to obtain the movie scene boundaries. The clips containing trailers and clips from the same scenes as the trailers provide more complete storylines. They have a higher probability of being highlight moments but still contain some noisy moments. Subsequently, we introduce a framework named Collaborative noisy Label Cleaner (CLC) to learn from these pseudo-noisy labels. The framework firstly enhances the modality perceptual consistency via the augmented crosspropagation (ACP) module, which exploits closely related audio-visual signals during training. In addition, a multimodality cleaning (MMC) mechanism is designed to filter out noisy and incomplete labels." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 605, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 289, + 715 + ], + "type": "text", + "content": "To support this study and facilitate benchmarking existing methods in this direction, we construct MovieLights, a Movie Highlight Detection Dataset. MovieLights contains 174 movies and the highlight moments are all from officially released trailers. The total length of these videos is over 370 hours. We conduct extensive experiments on MovieLights, in which our CLC exhibits promising results. We also demonstrate that our proposed CLC achieves significant performance-boosting over the state-of-the-art on" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 305, + 72, + 426, + 83 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 426, + 83 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 426, + 83 + ], + "type": "text", + "content": "the public VHD benchmarks." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 317, + 84, + 527, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 84, + 527, + 96 + ], + "spans": [ + { + "bbox": [ + 317, + 84, + 527, + 96 + ], + "type": "text", + "content": "In summary, our major contributions are as follows:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 317, + 102, + 545, + 289 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 317, + 102, + 545, + 162 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 102, + 545, + 162 + ], + "spans": [ + { + "bbox": [ + 317, + 102, + 545, + 162 + ], + "type": "text", + "content": "- We introduce a scene-aware paradigm to learn high-light moments in movies without any manual annotation. To the best of our knowledge, this is the first time that highlights detection is regarded as learning with noisy labels." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 317, + 168, + 545, + 215 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 168, + 545, + 215 + ], + "spans": [ + { + "bbox": [ + 317, + 168, + 545, + 215 + ], + "type": "text", + "content": "- We present an augmented cross-propagation to capture the interactions across modalities and a consistency loss to maximize the agreement between the different modalities." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 317, + 223, + 545, + 258 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 223, + 545, + 258 + ], + "spans": [ + { + "bbox": [ + 317, + 223, + 545, + 258 + ], + "type": "text", + "content": "- We incorporate a multi-modality noisy label cleaner to tackle label noise, which further improves the robustness of networks to annotation noise." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 317, + 266, + 545, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 266, + 545, + 289 + ], + "spans": [ + { + "bbox": [ + 317, + 266, + 545, + 289 + ], + "type": "text", + "content": "- Experiments on movie datasets and benchmark datasets validate the effectiveness of our framework." + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 299, + 398, + 312 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 299, + 398, + 312 + ], + "spans": [ + { + "bbox": [ + 306, + 299, + 398, + 312 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 319, + 545, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 319, + 545, + 617 + ], + "spans": [ + { + "bbox": [ + 304, + 319, + 545, + 617 + ], + "type": "text", + "content": "Video Highlight Detection. This task aims to identifying the interesting moments from untrimmed videos. In recent years, the videos studied for this task extend from domain-specific sport videos [41] to general videos such as social media videos [37], news [35], first-person videos [53] and vlog [25]. Most of previous works [14,21,53] interpret the video highlight detection task as a segment-level ranking problem. They compare pairwise segments from same domain video in order to learn a model that assigns highlight scores to these segments where the highlight segments receive higher scores than the non-highlight segments. MINI-Net [16] proposes to cast highlight detection as multiple instance ranking network learning. SL-Module [50] explores the highlight detection problem through Unsupervised Domain Adaptation (UDA) [32]. UMT [29] integrates highlight detection and moment retrieval into a unified framework and conduct joint optimization. PLD [45] models the video highlight detection into a pixel-level distinction estimation task. In this work, we regard highlight detection as learning with noisy labels. Joint-VA [2] also considers video highlight detection from the perspective of noise. However, it focuses on noise in features, such as videos having noisy audio when the microphone constantly has water splashing against it. We focus on specific annotation noise in video highlight detection." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 617, + 545, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 617, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 617, + 545, + 715 + ], + "type": "text", + "content": "Studies on Movies and Trailers. Studies on movies and trailers have received increased attention in research. [18] introduces a comprehensive dataset for movie understanding. [4, 42] try to model the relationships among the movie characters. [7, 34, 47] focus on breaking the storylines of movies into semantically cohesive parts. Besides the studies on movies, efforts have been made to develop trailer understanding. [57] presents a movie summarization system" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "18899" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 192 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 192 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 192 + ], + "type": "text", + "content": "and composes movie summaries in terms of user experience evaluation. [9] designs a movie trailer dataset for the evaluation of video-based recommender systems. [19] is the first approach that bridges trailers and movies and allows knowledge learned from trailers to be transferred to full movie analysis. In [43], the visual module and the temporal analysis module are respectively trained on trailers and movies. Because of the inaccessibility of public trailer-related benchmarks, we construct a new dataset (MovieLights) to detect the highlight moments in movies." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 193, + 289, + 348 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 193, + 289, + 348 + ], + "spans": [ + { + "bbox": [ + 46, + 193, + 289, + 348 + ], + "type": "text", + "content": "Learning with Noisy Labels. Learning with noisy labels has been a long-standing problem in computer vision. There are three kinds of approaches to this problem. One of the most common strategies for tackling label noise is to capture the transition probabilities between noisy labels and clean labels [31, 38, 48, 49, 52, 56]. Another solution is to design robust loss functions for model training against noisy labels [8, 26, 28, 30, 44, 55]. A popular method is to design a mechanism to select clean samples or give lower weight for noise samples in the training set to reduce impact of noise [15, 17, 20, 39, 46, 51]. In this paper, we attempt to solve the problem by exploiting multi-modalities nature of movies." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 356, + 188, + 370 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 356, + 188, + 370 + ], + "spans": [ + { + "bbox": [ + 47, + 356, + 188, + 370 + ], + "type": "text", + "content": "3. Movie Highlight Dataset" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 377, + 287, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 377, + 287, + 544 + ], + "spans": [ + { + "bbox": [ + 46, + 377, + 287, + 544 + ], + "type": "text", + "content": "In this paper, we aim to detect the highlight moments in movies by learning from easily accessible trailers as the noisy supervision. However, the existing movie and trailer-related benchmarks [3, 9, 18] lack sufficient functions for this task, such as the absence of full-length movies and ground-truth highlight annotations. Huang et al. [19] propose to respectively learn visual representations from trailers and temporal structure from full-length movies in their constructed Large-Scale Movie and Trailer Dataset (LSMTD). Wang [43] constructs a Trailer Moment Detection Dataset (TMDD) for detecting trailer moments from full-length movies without explicit human annotation. Both LSMTD and TMDD are not publicly available, while TMDD only contains three movie genres." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 545, + 287, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 545, + 287, + 688 + ], + "spans": [ + { + "bbox": [ + 46, + 545, + 287, + 688 + ], + "type": "text", + "content": "The inaccessibility of public benchmarks motivates us to construct a new dataset, named Movie Highlight Detection Dataset (MovieLights). In particular, we purchase a set of movies from commercial channels and collect their corresponding trailers from streaming platforms such as YouTube, covering at least 25 genres to ensure content diversity. The movies and trailers are then prepossessed by shot segmentation [36] and scene segmentation [47], respectively. The resulting shots are a series of consecutive frames taken by the camera until a physical interruption, and the scenes are consecutive shots that share a semantically related theme." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "type": "text", + "content": "As seen in Fig. 2, to build the ground truth, we conduct Faiss [22] to obtain visual similarity matching between" + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 307, + 70, + 547, + 177 + ], + "blocks": [ + { + "bbox": [ + 307, + 70, + 547, + 177 + ], + "lines": [ + { + "bbox": [ + 307, + 70, + 547, + 177 + ], + "spans": [ + { + "bbox": [ + 307, + 70, + 547, + 177 + ], + "type": "image", + "image_path": "2d233ba27a1ebdcf65e78f232b433d38bef4dbaef03c4f3eba3e9bf801fb0603.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 187, + 547, + 232 + ], + "lines": [ + { + "bbox": [ + 305, + 187, + 547, + 232 + ], + "spans": [ + { + "bbox": [ + 305, + 187, + 547, + 232 + ], + "type": "text", + "content": "Figure 2. The labeling process of MovieLights. For the training set, we introduce a scene-aware paradigm to obtain labels automatically. For the testing set, we collect 2 sets of labels from different annotators." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 316, + 258, + 536, + 354 + ], + "blocks": [ + { + "bbox": [ + 344, + 247, + 507, + 258 + ], + "lines": [ + { + "bbox": [ + 344, + 247, + 507, + 258 + ], + "spans": [ + { + "bbox": [ + 344, + 247, + 507, + 258 + ], + "type": "text", + "content": "Table 1. The basic statistics of MovieLights." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 316, + 258, + 536, + 354 + ], + "lines": [ + { + "bbox": [ + 316, + 258, + 536, + 354 + ], + "spans": [ + { + "bbox": [ + 316, + 258, + 536, + 354 + ], + "type": "table", + "html": "
TrainTest
Movie Number14430
Avg Durations per Movie2.19h2.14h
Avg Shot Number per Movie18521940
Avg Scene Number per Movie207193
Annotator1 Positive sample Proportion-0.27
Annotator2 Positive sample Proportion-0.30
Positive sample proportion0.350.21
", + "image_path": "3a9e4ae078c76ef97f82a76549573ac1c657cc744ae8f91001f44d7cdf31a421.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 377, + 545, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 377, + 545, + 652 + ], + "spans": [ + { + "bbox": [ + 304, + 377, + 545, + 652 + ], + "type": "text", + "content": "trailer frames and movie frames. We locate trailer moments in the movie and align them with the movie shots as annotation references. For the testing set, we collect 2 sets of moments for each movie from different workers, and these moments are annotated by different annotators independently. To ensure the consistency of results from different annotations, during the annotation procedure, all highlight moments must be related to the annotation references. Though all selected shots are relevant to the trailers, as highlight moments can be subjective, they may still vary in their saliency and time span. We calculate the intersection between every pair of moments annotated as the ground truth. However, the vast diversity of movie storylines makes the annotation challenging as it is time-consuming and requires annotators to be familiar with the movie. To collect a large amount of training data efficiently, we introduce a scene-aware paradigm to obtain the highlight moments label without any manual annotation. Specifically, we expand the shot-level annotation references to the scene span as positive samples automatically. It will capture the complete scene context of the trailer shot with movie storylines. Since the trailer shots may contain some less important moments, the acquired highlight labels are still noisy." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 654, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 547, + 715 + ], + "type": "text", + "content": "As seen in Tab. 1, MovieLights contains 174 movies in full length with their official trailers and it is split into a training set with 144 movies, and a testing set with 30 movies. The content diversity is ensured by the rich domain informations (more than 25 genres) and abundant segments" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "18900" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 155 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 155 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 155 + ], + "type": "text", + "content": "(325k shots and 36k scenes). Most movies in our dataset are between 90 to 150 minutes, and the length of the annotated moments varies from tens of seconds to several minutes. As the acquired positive highlight moments take up " + }, + { + "bbox": [ + 46, + 72, + 287, + 155 + ], + "type": "inline_equation", + "content": "35\\%" + }, + { + "bbox": [ + 46, + 72, + 287, + 155 + ], + "type": "text", + "content": " in the training set while the annotated true positives take up " + }, + { + "bbox": [ + 46, + 72, + 287, + 155 + ], + "type": "inline_equation", + "content": "21\\%" + }, + { + "bbox": [ + 46, + 72, + 287, + 155 + ], + "type": "text", + "content": " in the testing set, the difference tells the obvious existence of label noise." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 156, + 288, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 156, + 288, + 204 + ], + "spans": [ + { + "bbox": [ + 46, + 156, + 288, + 204 + ], + "type": "text", + "content": "We plan to release the dataset publicly to promote further study of movie analysis. Due to copyright issues, trailers and movies will be released in the form of extracted features in visual and audio modalities." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 213, + 113, + 225 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 213, + 113, + 225 + ], + "spans": [ + { + "bbox": [ + 47, + 213, + 113, + 225 + ], + "type": "text", + "content": "4. Approach" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 232, + 114, + 243 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 232, + 114, + 243 + ], + "spans": [ + { + "bbox": [ + 47, + 232, + 114, + 243 + ], + "type": "text", + "content": "4.1. Overview" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 251, + 287, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 251, + 287, + 300 + ], + "spans": [ + { + "bbox": [ + 46, + 251, + 287, + 300 + ], + "type": "text", + "content": "The overall architecture of our Collaborative noisy Label Cleaner (CLC) framework is illustrated in Fig. 3, which includes three modules: feature extraction, augmented crosspropagation (ACP), and multi-modal cleaning (MMC)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 300, + 287, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 300, + 287, + 396 + ], + "spans": [ + { + "bbox": [ + 46, + 300, + 287, + 396 + ], + "type": "text", + "content": "In our framework, both the visual and audio modalities are utilized. For feature extraction and encoding, we first split the video " + }, + { + "bbox": [ + 46, + 300, + 287, + 396 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 46, + 300, + 287, + 396 + ], + "type": "text", + "content": " into " + }, + { + "bbox": [ + 46, + 300, + 287, + 396 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 46, + 300, + 287, + 396 + ], + "type": "text", + "content": " shots. We characterize the " + }, + { + "bbox": [ + 46, + 300, + 287, + 396 + ], + "type": "inline_equation", + "content": "i^{th}" + }, + { + "bbox": [ + 46, + 300, + 287, + 396 + ], + "type": "text", + "content": " shots by two vectors, i.e., " + }, + { + "bbox": [ + 46, + 300, + 287, + 396 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_i" + }, + { + "bbox": [ + 46, + 300, + 287, + 396 + ], + "type": "text", + "content": " for the visual features, and " + }, + { + "bbox": [ + 46, + 300, + 287, + 396 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{a}}_{\\mathbf{i}}" + }, + { + "bbox": [ + 46, + 300, + 287, + 396 + ], + "type": "text", + "content": " for the audio features, where " + }, + { + "bbox": [ + 46, + 300, + 287, + 396 + ], + "type": "inline_equation", + "content": "i = 1,2,\\dots,T" + }, + { + "bbox": [ + 46, + 300, + 287, + 396 + ], + "type": "text", + "content": ". These features are extracted using pre-trained visual [11] and audio feature extractors [24]. The parameters of the two extractors are frozen during training." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 396, + 287, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 396, + 287, + 529 + ], + "spans": [ + { + "bbox": [ + 46, + 396, + 287, + 529 + ], + "type": "text", + "content": "The ACP and MMC are the core components of our framework. Since the visual-audio signals in videos are closely related but do not always contribute to highlight detection, the ACP module exploits the relationship via unimodal and cross-modal interactions. Then, it learns unified multi-modal representations for highlight detection. As the obtained highlight moments after scene segmentation are noisy, the MMC module firstly observes the changes in unimodal losses, then filters the noisy labels and utilizes the clean ones for multi-modal supervised learning. Next, we will introduce details for each component." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 538, + 214, + 551 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 538, + 214, + 551 + ], + "spans": [ + { + "bbox": [ + 47, + 538, + 214, + 551 + ], + "type": "text", + "content": "4.2. Augmented Cross-Propagation" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 557, + 287, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 557, + 287, + 651 + ], + "spans": [ + { + "bbox": [ + 46, + 557, + 287, + 651 + ], + "type": "text", + "content": "Cross-Propagation. To predict the highlights, the model needs to understand the storylines of the movie. Meanwhile, visual and audio inputs do not always contribute to accurate prediction. Therefore, temporal modeling and modality interaction are the keys to achieving successful highlight detection. To achieve this goal, we design our ACP module to fuse the visual-audio modalities. It involves three steps in total." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 654, + 287, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 654, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 46, + 654, + 287, + 689 + ], + "type": "text", + "content": "First, in order to align the multi-modal features, we introduce " + }, + { + "bbox": [ + 46, + 654, + 287, + 689 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 46, + 654, + 287, + 689 + ], + "type": "text", + "content": " which is fully-connected (FC) layer with ReLU to the " + }, + { + "bbox": [ + 46, + 654, + 287, + 689 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{a}}_{i}" + }, + { + "bbox": [ + 46, + 654, + 287, + 689 + ], + "type": "text", + "content": " such that it has same dimension as the " + }, + { + "bbox": [ + 46, + 654, + 287, + 689 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_i" + }, + { + "bbox": [ + 46, + 654, + 287, + 689 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 141, + 701, + 287, + 714 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 701, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 141, + 701, + 287, + 714 + ], + "type": "interline_equation", + "content": "\\mathbf {a} _ {i} = h \\left(\\hat {\\mathbf {a}} _ {\\mathrm {i}}\\right). \\tag {1}", + "image_path": "b7e90c641ec4bcebbafa387aaaf8536c97bf10f8c505de26546f3f9ed87358f1.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 72, + 545, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 167 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 167 + ], + "type": "text", + "content": "Movies are composed of consecutive audio-visual clips. Therefore, to measure the highlighted ness of a given shot, one must consider the relationship of the shot with its adjacent shots. The self-attention mechanism has shown effectiveness in capturing the long-term dependencies in previous works. We leverage self-attention to capture the temporal relationship for " + }, + { + "bbox": [ + 304, + 72, + 545, + 167 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_i" + }, + { + "bbox": [ + 304, + 72, + 545, + 167 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 72, + 545, + 167 + ], + "type": "inline_equation", + "content": "\\mathbf{a}_i" + }, + { + "bbox": [ + 304, + 72, + 545, + 167 + ], + "type": "text", + "content": " via Eq. 2 and Eq. 3, respectively." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 332, + 174, + 545, + 202 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 332, + 174, + 545, + 202 + ], + "spans": [ + { + "bbox": [ + 332, + 174, + 545, + 202 + ], + "type": "interline_equation", + "content": "\\mathbf {v} _ {i} ^ {s} = \\operatorname {s o f t m a x} \\left(\\frac {\\left(\\mathbf {v} _ {i} W _ {1} ^ {v}\\right) \\left(\\mathbf {v} W _ {2} ^ {v}\\right) ^ {\\top}}{\\sqrt {d}}\\right) \\left(\\mathbf {v} W _ {3} ^ {v}\\right), \\tag {2}", + "image_path": "aa46eed88f95f404666dc135e094549d6d400b1eb01510ed0c3388d4e41bcf70.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 333, + 208, + 545, + 236 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 208, + 545, + 236 + ], + "spans": [ + { + "bbox": [ + 333, + 208, + 545, + 236 + ], + "type": "interline_equation", + "content": "\\mathbf {a} _ {i} ^ {s} = \\operatorname {s o f t m a x} \\left(\\frac {\\left(\\mathbf {a} _ {i} W _ {1} ^ {a}\\right) \\left(\\mathbf {a} W _ {2} ^ {a}\\right) ^ {\\top}}{\\sqrt {d}}\\right) \\left(\\mathbf {a} W _ {3} ^ {a}\\right), \\tag {3}", + "image_path": "e7e071ca74e0377b8777d6cffa91e030f5b8384e0ec60449782378616b9877f9.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 238, + 545, + 322 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 238, + 545, + 322 + ], + "spans": [ + { + "bbox": [ + 304, + 238, + 545, + 322 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 238, + 545, + 322 + ], + "type": "inline_equation", + "content": "\\mathbf{v} = [\\mathbf{v}_1; \\mathbf{v}_2; \\ldots; \\mathbf{v}_T]" + }, + { + "bbox": [ + 304, + 238, + 545, + 322 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 238, + 545, + 322 + ], + "type": "inline_equation", + "content": "\\mathbf{a} = [\\mathbf{a}_1; \\mathbf{a}_2; \\ldots; \\mathbf{a}_T]" + }, + { + "bbox": [ + 304, + 238, + 545, + 322 + ], + "type": "text", + "content": "; the scaling factor " + }, + { + "bbox": [ + 304, + 238, + 545, + 322 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 304, + 238, + 545, + 322 + ], + "type": "text", + "content": " is equal to the visual/audio feature dimension and " + }, + { + "bbox": [ + 304, + 238, + 545, + 322 + ], + "type": "inline_equation", + "content": "(*)^{\\top}" + }, + { + "bbox": [ + 304, + 238, + 545, + 322 + ], + "type": "text", + "content": " denotes the transpose operator; " + }, + { + "bbox": [ + 304, + 238, + 545, + 322 + ], + "type": "inline_equation", + "content": "W^{v}" + }, + { + "bbox": [ + 304, + 238, + 545, + 322 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 238, + 545, + 322 + ], + "type": "inline_equation", + "content": "W^{a}" + }, + { + "bbox": [ + 304, + 238, + 545, + 322 + ], + "type": "text", + "content": " are learnable matrices of two modalities, which are implemented by a linear layer. Uni-modal self-attention can well capture uni-modal temporal contexts and enhance clip features within the same modality." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 323, + 545, + 382 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 323, + 545, + 382 + ], + "spans": [ + { + "bbox": [ + 304, + 323, + 545, + 382 + ], + "type": "text", + "content": "Despite the above self-attention capturing the clip interactions within the uni-modality, it is critical to capture the interactions across modalities. To capture semantic associations based on multi-modal signals, we introduce cross-attention to update the features of each modality." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 332, + 388, + 545, + 416 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 332, + 388, + 545, + 416 + ], + "spans": [ + { + "bbox": [ + 332, + 388, + 545, + 416 + ], + "type": "interline_equation", + "content": "\\mathbf {v} _ {i} ^ {c} = \\operatorname {s o f t m a x} \\left(\\frac {\\left(\\mathbf {v} _ {i} W _ {4} ^ {v}\\right) \\left(\\mathbf {a} W _ {4} ^ {a}\\right) ^ {\\top}}{\\sqrt {d}}\\right) \\left(\\mathbf {a} W _ {5} ^ {a}\\right), \\tag {4}", + "image_path": "b852663ddb7d15500c977c41e042b15ae9fd03f254017f1491de3cbf49fa11d4.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 332, + 422, + 545, + 450 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 332, + 422, + 545, + 450 + ], + "spans": [ + { + "bbox": [ + 332, + 422, + 545, + 450 + ], + "type": "interline_equation", + "content": "\\mathbf {a} _ {i} ^ {c} = \\operatorname {s o f t m a x} \\left(\\frac {\\left(\\mathbf {a} _ {i} W _ {6} ^ {a}\\right) \\left(\\mathbf {v} W _ {5} ^ {v}\\right) ^ {\\top}}{\\sqrt {d}}\\right) \\left(\\mathbf {v} W _ {6} ^ {v}\\right), \\tag {5}", + "image_path": "ce8d0db1debc231c83148ec25306beea9cc6faff3dc76b2716b85d166774806a.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 453, + 545, + 536 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 453, + 545, + 536 + ], + "spans": [ + { + "bbox": [ + 304, + 453, + 545, + 536 + ], + "type": "text", + "content": "Through cross-attention, the information from two modalities are connected. However, considering audiovisual temporal asynchrony, it is necessary to select effective information from multi-modality. We augment the relevant positive connections and dampen the irrelevant connections. The strength of these connections is measured by the cross-correlation matrix, computed by," + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 332, + 542, + 545, + 571 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 332, + 542, + 545, + 571 + ], + "spans": [ + { + "bbox": [ + 332, + 542, + 545, + 571 + ], + "type": "interline_equation", + "content": "\\mathbf {c} ^ {v} = R e L U \\left(\\frac {\\mathbf {v a} ^ {\\top}}{\\sqrt {d}}\\right), \\mathbf {c} ^ {a} = R e L U \\left(\\frac {\\mathbf {a v} ^ {\\top}}{\\sqrt {d}}\\right). \\tag {6}", + "image_path": "6682d8ecfded7c0330ed11d525f18acb620012c5809a989b4e6365c1bb691a97.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 575, + 545, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 575, + 545, + 635 + ], + "spans": [ + { + "bbox": [ + 304, + 575, + 545, + 635 + ], + "type": "text", + "content": "For each modality, the cross-correlation matrix " + }, + { + "bbox": [ + 304, + 575, + 545, + 635 + ], + "type": "inline_equation", + "content": "\\mathbf{c}" + }, + { + "bbox": [ + 304, + 575, + 545, + 635 + ], + "type": "text", + "content": " is used to re-weight the cross-attention features. Finally, we obtain updated visual features " + }, + { + "bbox": [ + 304, + 575, + 545, + 635 + ], + "type": "inline_equation", + "content": "\\bar{\\mathbf{v}}_{\\mathrm{i}}" + }, + { + "bbox": [ + 304, + 575, + 545, + 635 + ], + "type": "text", + "content": " and audio features " + }, + { + "bbox": [ + 304, + 575, + 545, + 635 + ], + "type": "inline_equation", + "content": "\\bar{\\mathbf{a}}_{\\mathrm{i}}" + }, + { + "bbox": [ + 304, + 575, + 545, + 635 + ], + "type": "text", + "content": " by fusion of the original features, enhanced uni-modal features, and cross-modal features." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 362, + 640, + 545, + 673 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 362, + 640, + 545, + 673 + ], + "spans": [ + { + "bbox": [ + 362, + 640, + 545, + 673 + ], + "type": "interline_equation", + "content": "\\bar {\\mathbf {v}} _ {\\mathbf {i}} = f \\left(\\mathbf {v} _ {i}, \\mathbf {v} _ {i} ^ {s}, \\left(\\sum_ {j = 1} ^ {T} \\mathbf {c} _ {i j} ^ {v}\\right) * \\mathbf {v} _ {i} ^ {c}\\right), \\tag {7}", + "image_path": "4ccd6848934c423259e980245a8d7a701d93d85fcf341c7b261c654c756b067f.jpg" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 363, + 682, + 545, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 363, + 682, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 363, + 682, + 545, + 715 + ], + "type": "interline_equation", + "content": "\\overline {{\\mathbf {a}}} _ {\\mathrm {i}} = f \\left(\\mathbf {a} _ {i}, \\mathbf {a} _ {i} ^ {s}, \\left(\\sum_ {j = 1} ^ {T} \\mathbf {c} _ {i j} ^ {a}\\right) * \\mathbf {a} _ {i} ^ {c}\\right), \\tag {8}", + "image_path": "6b95790cf2a5a7705d9949ea029ffc3c6bf06cc827014d711d2639b91ecf1d01.jpg" + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18901" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 51, + 80, + 541, + 271 + ], + "blocks": [ + { + "bbox": [ + 51, + 80, + 541, + 271 + ], + "lines": [ + { + "bbox": [ + 51, + 80, + 541, + 271 + ], + "spans": [ + { + "bbox": [ + 51, + 80, + 541, + 271 + ], + "type": "image", + "image_path": "8d3f2c2c28d1eaf9ba040dba693a9e53fc538b71c299fb54f14b73a8cd5f1b5a.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 285, + 546, + 341 + ], + "lines": [ + { + "bbox": [ + 46, + 285, + 546, + 341 + ], + "spans": [ + { + "bbox": [ + 46, + 285, + 546, + 341 + ], + "type": "text", + "content": "Figure 3. Overview of the proposed CLC. It includes three modules: feature extraction, augmented cross-propagation (ACP), and multi-modal cleaning (MMC). The visual and audio modalities of the input video are represented as vectors by the feature extraction module. Then the features are augmented by ACP module to capture semantic associations across modalities. MMC is used to filter outs noisy and incomplete labeling with additional uni-modal branches. During inference, we remove the uni-modal branches and only rely on the prediction of multi-modal branch. More details of the CLC are shown in Sec. 4.1" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 347, + 288, + 371 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 347, + 288, + 371 + ], + "spans": [ + { + "bbox": [ + 47, + 347, + 288, + 371 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 347, + 288, + 371 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 47, + 347, + 288, + 371 + ], + "type": "text", + "content": " is the fusion function consisting of FC layers and ReLU to further project the features." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 372, + 287, + 587 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 372, + 287, + 587 + ], + "spans": [ + { + "bbox": [ + 46, + 372, + 287, + 587 + ], + "type": "text", + "content": "Clearly, the cross-correlation matrix will assign large weights to clips that are relevant to the other modality. In addition, the ReLU activation in Eq. 6 cuts off connections with negative similarity values and only relevant positive connections would be preserved. By the above operations, the original feature will be infused with richer information. Consistency Loss. Multi-modal inputs help to comprehensively learn by integrating different aspects and boosting model performance. However, they are not fully exploited because some modality-specific features may be weakened in the fusion even when the multi-modal model outperforms its uni-modal counterpart. In this work, we provide parallel branches for each modality separately to obtain uni-modal prediction during training. As seen in Fig. 3, all branches share the same clip feature extracted from ACP and the feature is fed into different branches independently. In each branch, we develop a temporal model " + }, + { + "bbox": [ + 46, + 372, + 287, + 587 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 46, + 372, + 287, + 587 + ], + "type": "text", + "content": " to obtain its prediction score " + }, + { + "bbox": [ + 46, + 372, + 287, + 587 + ], + "type": "inline_equation", + "content": "y_{i}" + }, + { + "bbox": [ + 46, + 372, + 287, + 587 + ], + "type": "text", + "content": " of the " + }, + { + "bbox": [ + 46, + 372, + 287, + 587 + ], + "type": "inline_equation", + "content": "i^{th}" + }, + { + "bbox": [ + 46, + 372, + 287, + 587 + ], + "type": "text", + "content": " clip being a highlight as follows:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 594, + 287, + 609 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 594, + 287, + 609 + ], + "spans": [ + { + "bbox": [ + 53, + 594, + 287, + 609 + ], + "type": "interline_equation", + "content": "\\mathbf {y} _ {i} ^ {M M} = G \\left(\\bar {\\mathbf {v}} _ {\\mathbf {i}}, \\bar {\\mathbf {a}} _ {\\mathbf {i}}\\right), \\mathbf {y} _ {i} ^ {U M _ {v}} = G \\left(\\bar {\\mathbf {v}} _ {\\mathbf {i}}\\right), \\mathbf {y} _ {i} ^ {U M _ {a}} = G \\left(\\bar {\\mathbf {a}} _ {\\mathbf {i}}\\right). \\tag {9}", + "image_path": "4757674fcbbdf02511b3bf54a19ab33a5f6b00abab09ea7f4df8699e6ff893a8.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 617, + 287, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 617, + 287, + 640 + ], + "spans": [ + { + "bbox": [ + 47, + 617, + 287, + 640 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 617, + 287, + 640 + ], + "type": "inline_equation", + "content": "MM" + }, + { + "bbox": [ + 47, + 617, + 287, + 640 + ], + "type": "text", + "content": " is multi-modal branch; " + }, + { + "bbox": [ + 47, + 617, + 287, + 640 + ], + "type": "inline_equation", + "content": "UM_{v}" + }, + { + "bbox": [ + 47, + 617, + 287, + 640 + ], + "type": "text", + "content": " is visual branch and " + }, + { + "bbox": [ + 47, + 617, + 287, + 640 + ], + "type": "inline_equation", + "content": "UM_{a}" + }, + { + "bbox": [ + 47, + 617, + 287, + 640 + ], + "type": "text", + "content": " is audio branch." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 642, + 287, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 642, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 46, + 642, + 287, + 712 + ], + "type": "text", + "content": "However, the gap between the different modalities would bring instability in the joint optimization process. We employ the auxiliary consistency loss to guarantee consistency between the different modalities. Given the multi-modal features of a clip, they are consistent if they share the same prediction. Specifically, the consistency loss is defined as" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 347, + 545, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 347, + 545, + 373 + ], + "spans": [ + { + "bbox": [ + 305, + 347, + 545, + 373 + ], + "type": "text", + "content": "the cross-entropy between the multi-modal prediction probability " + }, + { + "bbox": [ + 305, + 347, + 545, + 373 + ], + "type": "inline_equation", + "content": "\\mathbf{y}_i^{MM}" + }, + { + "bbox": [ + 305, + 347, + 545, + 373 + ], + "type": "text", + "content": " and uni-modal prediction probability " + }, + { + "bbox": [ + 305, + 347, + 545, + 373 + ], + "type": "inline_equation", + "content": "\\mathbf{y}_i^{UM}" + }, + { + "bbox": [ + 305, + 347, + 545, + 373 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 312, + 377, + 545, + 405 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 377, + 545, + 405 + ], + "spans": [ + { + "bbox": [ + 312, + 377, + 545, + 405 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {c o n s}} = - \\left(\\sum_ {i = 1} ^ {N} \\mathbf {y} _ {i} ^ {U M _ {v}} \\log \\mathbf {y} _ {i} ^ {M M} + \\sum_ {i = 1} ^ {N} \\mathbf {y} _ {i} ^ {U M _ {a}} \\log \\mathbf {y} _ {i} ^ {M M}\\right), \\tag {10}", + "image_path": "888dc0573f631c5b9b15e067c801150295722f7677fba27679b7553dca0d9d45.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 410, + 491, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 410, + 491, + 422 + ], + "spans": [ + { + "bbox": [ + 306, + 410, + 491, + 422 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 306, + 410, + 491, + 422 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 306, + 410, + 491, + 422 + ], + "type": "text", + "content": " is the number of samples in a batch." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 422, + 545, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 422, + 545, + 483 + ], + "spans": [ + { + "bbox": [ + 304, + 422, + 545, + 483 + ], + "type": "text", + "content": "This consistency loss not only implicitly enhances uni-modal information, but also explicitly guides the multimodal branch to robust supervision. The added uni-modal branches are only utilized in the training phase and are disabled during the inference stage." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 491, + 433, + 503 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 491, + 433, + 503 + ], + "spans": [ + { + "bbox": [ + 306, + 491, + 433, + 503 + ], + "type": "text", + "content": "4.3. Multi-modal Cleaning" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 510, + 545, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 510, + 545, + 593 + ], + "spans": [ + { + "bbox": [ + 304, + 510, + 545, + 593 + ], + "type": "text", + "content": "Multi-modality Sample Cleaning. Annotation noise is inevitable in VHD due to subjectivity depending on the users and annotators. In this paper, we argue that highlight detection should be regarded as Learning with Noisy Labels. To alleviate the performance drop caused by noisy labels, we adopt a multi-modality collaborative cleaning to adaptively filter noisy samples with noisy modality information." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 594, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 546, + 713 + ], + "type": "text", + "content": "Firstly, we maintain multiple outputs simultaneously which are predicted by different branches. The uni-modal branches independently select clean samples based on the low-loss criterion in which instances with lower losses are treated clean samples. Contrary to existing noisy sample selection methods [12, 54], which directly discard high-loss samples, we keep all samples to train the uni-modal branches. Then, we update the multi-modal branch using only clean samples selected by both uni-modal branches in the back-propagation. The samples for the multi-modal" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18902" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 132 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 132 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 132 + ], + "type": "text", + "content": "branch training are selected dynamically while all samples participate in the uni-mode branches training. In this way, MMC obtains information on all samples to avoid the model defecting to favoring easy samples. More details of MMC are presented in the supplementary material." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 133, + 287, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 133, + 287, + 167 + ], + "spans": [ + { + "bbox": [ + 46, + 133, + 287, + 167 + ], + "type": "text", + "content": "Each branch has its own loss function. For the uni-modal branch, we employ the cross-entropy loss with all samples as follows:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 87, + 174, + 287, + 205 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 87, + 174, + 287, + 205 + ], + "spans": [ + { + "bbox": [ + 87, + 174, + 287, + 205 + ], + "type": "interline_equation", + "content": "\\mathcal {L} ^ {U M _ {v}} = \\mathcal {L} _ {\\mathrm {c e}} ^ {U M _ {v}} = - \\sum_ {i = 1} ^ {N} \\mathbf {g} _ {i} \\log \\mathbf {y} _ {i} ^ {U M _ {v}}, \\tag {11}", + "image_path": "b22b113f8d3457c25835da8ff62662b382ad3154ddcbb945496899da4699c8d3.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 87, + 214, + 287, + 247 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 87, + 214, + 287, + 247 + ], + "spans": [ + { + "bbox": [ + 87, + 214, + 287, + 247 + ], + "type": "interline_equation", + "content": "\\mathcal {L} ^ {U M _ {a}} = \\mathcal {L} _ {\\mathrm {c e}} ^ {U M _ {a}} = - \\sum_ {i = 1} ^ {N} \\mathbf {g} _ {i} \\log \\mathbf {y} _ {i} ^ {U M _ {a}}. \\tag {12}", + "image_path": "16826924d87f2dac3e0af5ceefde1ef978b0adc52ead78882c667cc30584f603.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 250, + 287, + 286 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 250, + 287, + 286 + ], + "spans": [ + { + "bbox": [ + 47, + 250, + 287, + 286 + ], + "type": "text", + "content": "For the multi-modal branch, its cross-entropy loss and consistency loss are updated with a re-weighting scheme with clean samples." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 109, + 294, + 287, + 327 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 109, + 294, + 287, + 327 + ], + "spans": [ + { + "bbox": [ + 109, + 294, + 287, + 327 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {c e}} ^ {\\mathrm {M M}} = - \\sum_ {i = 1} ^ {N ^ {\\prime}} \\mathbf {g} _ {i} \\log \\mathbf {y} _ {i} ^ {M M}, \\tag {13}", + "image_path": "2c4417c13cb932c27060f8ae357b1b3572b5388890018e568860db3cb2387c59.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 52, + 335, + 287, + 365 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 52, + 335, + 287, + 365 + ], + "spans": [ + { + "bbox": [ + 52, + 335, + 287, + 365 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {c o n s}} ^ {M M} = - \\left(\\sum_ {i = 1} ^ {N ^ {\\prime}} \\mathbf {y} _ {i} ^ {U M _ {v}} \\log \\mathbf {y} _ {i} ^ {M M} + \\sum_ {i = 1} ^ {N ^ {\\prime}} \\mathbf {y} _ {i} ^ {U M _ {a}} \\log \\mathbf {y} _ {i} ^ {M M}\\right), \\tag {14}", + "image_path": "5669eca71bf056aec3afd23cdcfa778658e0f898e172e8f933ee0cfe9a36af9e.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 115, + 366, + 287, + 381 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 115, + 366, + 287, + 381 + ], + "spans": [ + { + "bbox": [ + 115, + 366, + 287, + 381 + ], + "type": "interline_equation", + "content": "\\mathcal {L} ^ {\\mathrm {M M}} = \\mathcal {L} _ {\\mathrm {c e}} ^ {\\mathrm {M M}} + \\beta \\mathcal {L} _ {\\mathrm {c o n s}} ^ {M M}, \\tag {15}", + "image_path": "a5ece9e1454c80cee48efe3b7608e7ce3f1ddc84a90791040c2b5c8868b546f8.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 385, + 287, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 385, + 287, + 443 + ], + "spans": [ + { + "bbox": [ + 46, + 385, + 287, + 443 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 385, + 287, + 443 + ], + "type": "inline_equation", + "content": "\\mathbf{g}_i" + }, + { + "bbox": [ + 46, + 385, + 287, + 443 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 385, + 287, + 443 + ], + "type": "inline_equation", + "content": "\\mathbf{y}_i^*" + }, + { + "bbox": [ + 46, + 385, + 287, + 443 + ], + "type": "text", + "content": " denote the ground-truth and predicted probability of the " + }, + { + "bbox": [ + 46, + 385, + 287, + 443 + ], + "type": "inline_equation", + "content": "i^{th}" + }, + { + "bbox": [ + 46, + 385, + 287, + 443 + ], + "type": "text", + "content": " clip, respectively; " + }, + { + "bbox": [ + 46, + 385, + 287, + 443 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 385, + 287, + 443 + ], + "type": "text", + "content": " is the number of samples in a batch and " + }, + { + "bbox": [ + 46, + 385, + 287, + 443 + ], + "type": "inline_equation", + "content": "N'" + }, + { + "bbox": [ + 46, + 385, + 287, + 443 + ], + "type": "text", + "content": " is the number of samples selected by MMC; and " + }, + { + "bbox": [ + 46, + 385, + 287, + 443 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 46, + 385, + 287, + 443 + ], + "type": "text", + "content": " are designed to balance different loss terms." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 445, + 287, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 445, + 287, + 529 + ], + "spans": [ + { + "bbox": [ + 46, + 445, + 287, + 529 + ], + "type": "text", + "content": "Post processing. In experiments we observe that noise makes jitter prediction curves along the temporal dimension, which may cause discontinuous thresholds for highlight selection. Therefore, we apply a median filter to smooth the prediction curves. Supposing " + }, + { + "bbox": [ + 46, + 445, + 287, + 529 + ], + "type": "inline_equation", + "content": "\\mathbf{y} = [\\mathbf{y}_1; \\mathbf{y}_2; \\dots; \\mathbf{y}_T]" + }, + { + "bbox": [ + 46, + 445, + 287, + 529 + ], + "type": "text", + "content": " is the original curve predicted by the CLC, the smoothed curve " + }, + { + "bbox": [ + 46, + 445, + 287, + 529 + ], + "type": "inline_equation", + "content": "\\mathbf{s} = [\\mathbf{s}_1; \\mathbf{s}_2; \\dots; \\mathbf{s}_T]" + }, + { + "bbox": [ + 46, + 445, + 287, + 529 + ], + "type": "text", + "content": " is given by:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 67, + 536, + 287, + 569 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 536, + 287, + 569 + ], + "spans": [ + { + "bbox": [ + 67, + 536, + 287, + 569 + ], + "type": "interline_equation", + "content": "\\mathbf {s} _ {i} = \\left\\{ \\begin{array}{l l} \\operatorname {M e d} \\left(\\mathbf {y} _ {i - k}, \\mathbf {y} _ {i + k}\\right), & k < i \\leq T - k \\\\ \\mathbf {y} _ {i}, & \\text {o t h e r w i s e} \\end{array} \\right. \\tag {16}", + "image_path": "618344de3965cc6415e00010540b7c32ea1fe4ea321e1d6699972e414c65c626.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 575, + 287, + 597 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 575, + 287, + 597 + ], + "spans": [ + { + "bbox": [ + 46, + 575, + 287, + 597 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 575, + 287, + 597 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 575, + 287, + 597 + ], + "type": "text", + "content": " is the window size, and \"Med\" denotes the median filter." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 47, + 604, + 123, + 616 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 604, + 123, + 616 + ], + "spans": [ + { + "bbox": [ + 47, + 604, + 123, + 616 + ], + "type": "text", + "content": "5. Experiment" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 47, + 623, + 236, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 623, + 236, + 635 + ], + "spans": [ + { + "bbox": [ + 47, + 623, + 236, + 635 + ], + "type": "text", + "content": "5.1. Datasets and Experimental Settings" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 46, + 641, + 287, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 641, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 46, + 641, + 287, + 689 + ], + "type": "text", + "content": "Datasets. We evaluate our CLC on the constructed dataset MovieLights and public YouTube Highlights dataset [37]. MovieLights is split into training and testing sets, each containing 144 and 30 movies respectively." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 46, + 689, + 287, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 689, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 46, + 689, + 287, + 712 + ], + "type": "text", + "content": "We split movies into shots. The movie features are represented at shot-level. We use the middle frame of each" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "text", + "content": "shot to extract its visual feature with ViT [10] pre-trained by CLIP [33]. We align timestamps of audio clips with the visual shots, and sample the audio clip of each with " + }, + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "inline_equation", + "content": "16K" + }, + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "text", + "content": " Hz sampling rate and 512 windowed signal length. The resulted shot-level audio features are obtained with the PANN audio network [24] pretrained on AudioSet [13]." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 145, + 545, + 276 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 145, + 545, + 276 + ], + "spans": [ + { + "bbox": [ + 304, + 145, + 545, + 276 + ], + "type": "text", + "content": "The YouTube Highlights contains six distinct categories with a total 422 videos currently available. Following the practice of prior efforts, we train a highlight detector for each category. YouTube Highlights provides two annotations: Harvested Highlight and Mturk Highlight. In the Harvested annotation, the match label specifies if each clip is matched in the edited video, where 1 denotes matched, -1 denotes unmatched and 0 denotes the borderline cases. In the Mturk annotation, the highlight labels are marked by multiple turkers of different styles, making it noisier than the Harvested annotation." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 279, + 545, + 410 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 279, + 545, + 410 + ], + "spans": [ + { + "bbox": [ + 304, + 279, + 545, + 410 + ], + "type": "text", + "content": "On YouTube Highlights, we use the same protocol and data preprocessing as [29]. It obtain clip-level visual features and optical flow features using I3D [6] pre-trained on Kinetics400 [23]. It use a PANN audio network [24] pretrained on AudioSet [13] to obtain audio features that align with the visual clips. Frame-level features are average-pooled within each clip for both audio and visual features to generate a clip-level feature. Since each feature vector spans 32 consecutive frames, we follow [29] and consider the feature vector corresponded to a clip if their overlap is more than " + }, + { + "bbox": [ + 304, + 279, + 545, + 410 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 304, + 279, + 545, + 410 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 412, + 545, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 412, + 545, + 651 + ], + "spans": [ + { + "bbox": [ + 304, + 412, + 545, + 651 + ], + "type": "text", + "content": "Benchmarks. To better inspect the robustness of our CLC against noisy labels, we also apply label perturbations in training set of YouTube Highlights while keeping the ground-truths in the testing set unchanged. YouTube Highlights has two benchmarks for comparison. 1) Harvested with matched: we regard the clips labeled with matched as the highlighted clips and this is the same setting as in previous works [29, 50]. 2) Harvested with matched and borderline: clips labeled with matched and borderline are treated as the highlight moments. In this benchmark, the training set contains some highlighted clips with weak confidence. We select the clips whose mturk-label is over score 1 as the highlighted clips, which means that at least one turker selects the clip as a highlight. There are labeled by different types of annotators between Mturk and the test set and bring greater noise. We select the clips whose mturk-label is over score 1 as the highlighted clips, which means that at least one turker selects the clip as a highlight. Due the labeling gap between the Mturk annotation in the training set and the clean testing set, this benchmark is even noisier." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 653, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 653, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 653, + 545, + 712 + ], + "type": "text", + "content": "Baselines. We introduce CLC-, the degenerated version of CLC, as a baseline. Similar to CLC, CLC- is a Bi-LSTM-based model and takes the temporal sequence of shot features as input but lacks the modules of augmented cross-propagation and multi-modality noisy label cleaner." + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18903" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 70, + 82, + 265, + 190 + ], + "blocks": [ + { + "bbox": [ + 105, + 71, + 228, + 81 + ], + "lines": [ + { + "bbox": [ + 105, + 71, + 228, + 81 + ], + "spans": [ + { + "bbox": [ + 105, + 71, + 228, + 81 + ], + "type": "text", + "content": "Table 2. Results on MovieLights." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 70, + 82, + 265, + 190 + ], + "lines": [ + { + "bbox": [ + 70, + 82, + 265, + 190 + ], + "spans": [ + { + "bbox": [ + 70, + 82, + 265, + 190 + ], + "type": "table", + "html": "
MethodsModalitymAP
GIFs [14]V25.48
SL-Module [50]V32.34
SL-Module [50]VA34.27
UMT [29]VA38.7
CLC-VA39.65
CLC- w/ SCE [44]VA39.83
CLC- w/ LS [40]VA40.49
CLCVA43.88
", + "image_path": "2640eb5038901fc2617578ee8e6c3106976f0f26758fc1c48f77a3649dfbb8a1.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 209, + 287, + 281 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 209, + 287, + 281 + ], + "spans": [ + { + "bbox": [ + 46, + 209, + 287, + 281 + ], + "type": "text", + "content": "Evaluation Metric. We adopt mean Average Precision (mAP) as the evaluation metric for MovieLights and YouTube Highlights. Considering that a highlighted moment in one video is not necessarily more interesting than non-highlight moments in other videos, we evaluate on each test video independently and report the averaged results." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 281, + 287, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 281, + 287, + 365 + ], + "spans": [ + { + "bbox": [ + 46, + 281, + 287, + 365 + ], + "type": "text", + "content": "Implementation Details. On the MovieLights, we train our model using SGD, with a learning rate of 0.01. We train for 50 epochs. Before the cross-attention modules, we project each modality into a vector of 512 dimension. The key, query, and value vectors all share the same dimension. Weight " + }, + { + "bbox": [ + 46, + 281, + 287, + 365 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 46, + 281, + 287, + 365 + ], + "type": "text", + "content": " in Eq. 15 is empirically set to 0.1 and window size " + }, + { + "bbox": [ + 46, + 281, + 287, + 365 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 281, + 287, + 365 + ], + "type": "text", + "content": " in Eq. 16 is set to 9." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 373, + 180, + 386 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 373, + 180, + 386 + ], + "spans": [ + { + "bbox": [ + 47, + 373, + 180, + 386 + ], + "type": "text", + "content": "5.2. Results on MovieLights" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 392, + 287, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 392, + 287, + 475 + ], + "spans": [ + { + "bbox": [ + 46, + 392, + 287, + 475 + ], + "type": "text", + "content": "On MovieLights, we train our model with the noisy pseudo labels. To compare with previous state-of-the-art highlight detection works, we train UMT [29] and SL-Module [50] using the same protocol and data preprocessing as in CLC. We also compare with Video2GIF [14] using its off-the-shelf tool1. The upper part of Tab. 2 illustrates the significant performance gain of CLC." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 475, + 288, + 606 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 475, + 288, + 606 + ], + "spans": [ + { + "bbox": [ + 46, + 475, + 288, + 606 + ], + "type": "text", + "content": "To demonstrate the advantages of CLC in learning with noisy labels, we make comparisons with two main-stream label noise approaches: Label Smoothing [40] and SCE loss [44]. Specifically, we insert Label Smoothing or SCE into our CLC- framework to create two baseline VHD methods to tackle label noise. The bottom part of Tab. 2 shows that CLC outperforms the two baseline methods by a notable margin, indicating that our augmented cross-propagation and multi-modality noisy label cleaner are more effective than vanilla label noise approaches in VHD tasks." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 615, + 141, + 628 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 615, + 141, + 628 + ], + "spans": [ + { + "bbox": [ + 47, + 615, + 141, + 628 + ], + "type": "text", + "content": "5.3. Ablation Study" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 634, + 287, + 657 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 634, + 287, + 657 + ], + "spans": [ + { + "bbox": [ + 46, + 634, + 287, + 657 + ], + "type": "text", + "content": "In this experiment, we analyze the impact of each module. The results are summarized in Tab. 3." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 658, + 287, + 694 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 658, + 287, + 694 + ], + "spans": [ + { + "bbox": [ + 46, + 658, + 287, + 694 + ], + "type": "text", + "content": "Multi-modality Sample Cleaning. We first inspect the impact of the multi-modality sample filter module because we are primary concerned with learning with noisy labels" + } + ] + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 328, + 82, + 525, + 156 + ], + "blocks": [ + { + "bbox": [ + 347, + 71, + 503, + 81 + ], + "lines": [ + { + "bbox": [ + 347, + 71, + 503, + 81 + ], + "spans": [ + { + "bbox": [ + 347, + 71, + 503, + 81 + ], + "type": "text", + "content": "Table 3. Ablation results of MoiveLightts." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 328, + 82, + 525, + 156 + ], + "lines": [ + { + "bbox": [ + 328, + 82, + 525, + 156 + ], + "spans": [ + { + "bbox": [ + 328, + 82, + 525, + 156 + ], + "type": "table", + "html": "
MMSCCPCLPPmAP
××××39.65
✓×××41.69
✓✓××42.79
✓✓✓×43.22
✓✓✓✓43.88
", + "image_path": "e99e3d7373ea63ca6643dd2458f8ea84b24bd15ff0455dffdf1675ca23628d9d.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "type": "table", + "bbox": [ + 307, + 178, + 549, + 328 + ], + "blocks": [ + { + "bbox": [ + 351, + 167, + 499, + 177 + ], + "lines": [ + { + "bbox": [ + 351, + 167, + 499, + 177 + ], + "spans": [ + { + "bbox": [ + 351, + 167, + 499, + 177 + ], + "type": "text", + "content": "Table 4. Results on YouTube Highlights." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 307, + 178, + 549, + 328 + ], + "lines": [ + { + "bbox": [ + 307, + 178, + 549, + 328 + ], + "spans": [ + { + "bbox": [ + 307, + 178, + 549, + 328 + ], + "type": "table", + "html": "
Methodsdoggym.park.ska.ski.surf.Avg.
GIFs [14]30.833.55455.432.854.146.4
LSVM [37]60.041.061.062.036.061.053.6
HighlightMe [5]63737264526264
MINI-Net [16]58.261.770.272.258.760.164.4
CHD [1]60.671.174.249.868.268.565.4
Trail [43]63.382.562.352.974.579.369.1
SL-Module [50]70.853.277.272.566.176.269.3
Joint-VA [2]64.571.980.86273.278.371.8
PLD [45]74.970.277.957.570.77973
CO-AV [27]60.9668974.16981.174.7
UMT [29]65.975.281.671.872.382.774.9
CLC(ours)70.579.483.983.579.583.680.1
", + "image_path": "af631e9394c246580adc4784f623a084b52e73b65d9b6aa0b3f266905ba70e43.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_body" + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 348, + 545, + 480 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 348, + 545, + 480 + ], + "spans": [ + { + "bbox": [ + 304, + 348, + 545, + 480 + ], + "type": "text", + "content": "in VHD. The " + }, + { + "bbox": [ + 304, + 348, + 545, + 480 + ], + "type": "inline_equation", + "content": "2\\%" + }, + { + "bbox": [ + 304, + 348, + 545, + 480 + ], + "type": "text", + "content": " performance gain from the module over baseline shows the importance of filtering noisy sample. Based on the this observation, the subsequent ablation experiments are conducted under the setting of noise filtering. Cross-Propagation. We then evaluate the impact of the feature augmentation module. Compared to naive feature concatenation, the models with augmented features show superior performance. The results validate that feature augmentation module can better explore the complementary information from different modalities and suppress the mutual disturbance of desynchronized uni-modal information." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 481, + 545, + 553 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 481, + 545, + 553 + ], + "spans": [ + { + "bbox": [ + 304, + 481, + 545, + 553 + ], + "type": "text", + "content": "Consistency Loss. We examine the contribution of the consistency loss. The results demonstrate that the employment of the auxiliary multi-modal constraint further increases the model robustness. This consistency loss not only implicitly enhances uni-modal information, but also explicitly guides the multi-modal branch to better learning." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 554, + 545, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 554, + 545, + 601 + ], + "spans": [ + { + "bbox": [ + 304, + 554, + 545, + 601 + ], + "type": "text", + "content": "Post Processing. We compare the highlight prediction curves with and without the post processing median filter in Fig. 4(a,b), which prevents disruptive prediction variation and improves the overall performance." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 610, + 473, + 623 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 610, + 473, + 623 + ], + "spans": [ + { + "bbox": [ + 306, + 610, + 473, + 623 + ], + "type": "text", + "content": "5.4. Results on YouTube Highlights" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 628, + 545, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 628, + 545, + 676 + ], + "spans": [ + { + "bbox": [ + 304, + 628, + 545, + 676 + ], + "type": "text", + "content": "We conduct experiments on the public video highlight detection benchmarks YouTube Highlights, including six domain video datasets, to verify the generalization ability of CLC." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "content": "The setting of Harvested Highlight with matched is consistent with the previous work [29]. As shown in Tab. 4, CLC achieves state-of-the-art performance on the YouTube" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 58, + 702, + 198, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 702, + 198, + 713 + ], + "spans": [ + { + "bbox": [ + 58, + 702, + 198, + 713 + ], + "type": "text", + "content": "1https://github.com/gyglim/video2gif_code" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18904" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 70, + 212, + 198 + ], + "blocks": [ + { + "bbox": [ + 50, + 70, + 212, + 198 + ], + "lines": [ + { + "bbox": [ + 50, + 70, + 212, + 198 + ], + "spans": [ + { + "bbox": [ + 50, + 70, + 212, + 198 + ], + "type": "image", + "image_path": "41c5bd013edb20170b65a2247d3c6c90f2c95d9a7e47fd267d16acb828ca2197.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 47, + 206, + 545, + 219 + ], + "lines": [ + { + "bbox": [ + 47, + 206, + 545, + 219 + ], + "spans": [ + { + "bbox": [ + 47, + 206, + 545, + 219 + ], + "type": "text", + "content": "Figure 4. (a, b) Comparison of original prediction curve with filtered prediction curve. (c) Results on YouTube Highlights with noisy label." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 213, + 70, + 375, + 198 + ], + "blocks": [ + { + "bbox": [ + 213, + 70, + 375, + 198 + ], + "lines": [ + { + "bbox": [ + 213, + 70, + 375, + 198 + ], + "spans": [ + { + "bbox": [ + 213, + 70, + 375, + 198 + ], + "type": "image", + "image_path": "08a751add36e3c82c96298e24a9351831f506cc8c00ebff9c438a4a59c78f4b9.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 376, + 71, + 545, + 194 + ], + "blocks": [ + { + "bbox": [ + 376, + 71, + 545, + 194 + ], + "lines": [ + { + "bbox": [ + 376, + 71, + 545, + 194 + ], + "spans": [ + { + "bbox": [ + 376, + 71, + 545, + 194 + ], + "type": "image", + "image_path": "7b8c3bf7b26235ea5d7dec63ccf8e99eb352033c68bd6c1fa05cd3bc0aa7317c.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 70, + 232, + 523, + 318 + ], + "blocks": [ + { + "bbox": [ + 189, + 220, + 402, + 232 + ], + "lines": [ + { + "bbox": [ + 189, + 220, + 402, + 232 + ], + "spans": [ + { + "bbox": [ + 189, + 220, + 402, + 232 + ], + "type": "text", + "content": "Table 5. Results on YouTube Highlights with Noisy Label." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 70, + 232, + 523, + 318 + ], + "lines": [ + { + "bbox": [ + 70, + 232, + 523, + 318 + ], + "spans": [ + { + "bbox": [ + 70, + 232, + 523, + 318 + ], + "type": "table", + "html": "
AnnotationNoiseMethodsdoggym.park.ska.ski.surf.Avg.
Harvested matchedcleanUMT [29]65.9075.2081.6071.8072.3082.7074.90
Harvested matchedcleanCLC70.5179.4383.8583.5179.4683.5680.05 (↑ 5.15)
Harvested borderlineslight noiseUMT [29]65.9374.3181.5871.8470.2482.4674.39
Harvested borderlineslight noiseCLC69.4180.7378.5085.3681.1183.1679.71 (↑ 5.32)
Mturksevere noiseUMT [29]63.7876.1675.0273.6269.9981.5973.36
Mturksevere noiseCLC66.9280.4485.9282.3378.0581.7279.22 (↑ 5.86)
", + "image_path": "4558dd337bc734db059b19638b79f66ec78f5787ec839649a3def075860ccac5.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 52, + 329, + 541, + 419 + ], + "blocks": [ + { + "bbox": [ + 52, + 329, + 541, + 419 + ], + "lines": [ + { + "bbox": [ + 52, + 329, + 541, + 419 + ], + "spans": [ + { + "bbox": [ + 52, + 329, + 541, + 419 + ], + "type": "image", + "image_path": "6edce3294648c6dd7f15fb78235913315f4342ffd14e1622ebcd97ae6035f2fd.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 427, + 546, + 462 + ], + "lines": [ + { + "bbox": [ + 46, + 427, + 546, + 462 + ], + "spans": [ + { + "bbox": [ + 46, + 427, + 546, + 462 + ], + "type": "text", + "content": "Figure 5. The highlight moments selected by our CLC from Skyfall and Sherlock Holmes. Top: Bond gives chase to a professional hitman by car to find a classified hard drive, and then a firefight erupted in the market. Bottom: Holmes and Blackwood are facing off, and then Blackwood reaches for a weapon to kill Holmes, but accidentally trips off a scaffolding and falls to his death." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 468, + 287, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 468, + 287, + 552 + ], + "spans": [ + { + "bbox": [ + 46, + 468, + 287, + 552 + ], + "type": "text", + "content": "Highlights, outperforming the existing multi-modal highlight detection methods in the average metric across all categories. Specifically, CLC achieves best performance in three out of the six categories, while maintaining reasonably competitive performance in the other three categories. These results support our claim that highlight detection should be regarded as learning with noisy labels." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 560, + 287, + 621 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 560, + 287, + 621 + ], + "spans": [ + { + "bbox": [ + 46, + 560, + 287, + 621 + ], + "type": "text", + "content": "To quantify how CLC is robust to different levels of label noise, we inspect the settings of Harvested Highlight with matched and borderline and Mturk Highlight, which are perturbed by varying degrees of label noise in YouTube Highlights." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 629, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 629, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 629, + 287, + 713 + ], + "type": "text", + "content": "Tab. 5 exhibits the performances of CLC and UMT [29] at different noise levels. As the noise level increases, the VHD task becomes more difficult, but the performance superiority of our CLC over UMT becomes even more obvious. It is illustrated in Fig. 4(c) that compared with the most recent state-of-the-art UMT, our model can achieve better performance even when disturbed by severe label noise." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 467, + 389, + 479 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 467, + 389, + 479 + ], + "spans": [ + { + "bbox": [ + 306, + 467, + 389, + 479 + ], + "type": "text", + "content": "5.5. Visualization" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 487, + 545, + 547 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 487, + 545, + 547 + ], + "spans": [ + { + "bbox": [ + 304, + 487, + 545, + 547 + ], + "type": "text", + "content": "As shown in Fig. 5, we present some visualization examples of the detected highlight clips in MovieLights by CLC. The examples clearly show that the prediction of CLC is in accordance with user expectation. We will provide more examples in the supplementary material." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 555, + 378, + 567 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 555, + 378, + 567 + ], + "spans": [ + { + "bbox": [ + 306, + 555, + 378, + 567 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 576, + 545, + 720 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 576, + 545, + 720 + ], + "spans": [ + { + "bbox": [ + 304, + 576, + 545, + 720 + ], + "type": "text", + "content": "In this study, we present Collaborative noisy Label Cleaner (CLC), a novel framework to handle noisy labels in video highlight detection. We make use of the augmented cross-propagation module to better enhance network robustness and multi-modality cleaning to achieve cleaner highlight labels by observing the loss changes of different modalities. We demonstrate the state-of-the-art performance of our method with extensive experiments on MovieLights and YouTube Highlights datasets. In future work, we are interested in extending the proposed mechanisms to other video-understanding tasks such as scene segmentation and video temporal grounding." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18905" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 53, + 91, + 286, + 123 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 286, + 123 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 286, + 123 + ], + "type": "text", + "content": "[1] Taivanbat Badamdorj, Mrigank Rochan, Yang Wang, and Li Cheng. Contrastive learning for unsupervised video highlight detection. In CVPR, 2022. 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 124, + 287, + 157 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 124, + 287, + 157 + ], + "spans": [ + { + "bbox": [ + 53, + 124, + 287, + 157 + ], + "type": "text", + "content": "[2] Taivanbat Badamdorj, Mrigank Rochan, Yang Wang, Jiaze Wang, and Li Cheng. Joint visual and audio learning for video highlight detection. In ICCV, 2021. 2, 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 158, + 287, + 191 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 158, + 287, + 191 + ], + "spans": [ + { + "bbox": [ + 53, + 158, + 287, + 191 + ], + "type": "text", + "content": "[3] Max Bain, Arsha Nagrani, Andrew Brown, and Andrew Zisserman. Condensed movies: Story based retrieval with contextual embeddings. In ACCV, 2020. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 192, + 286, + 213 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 192, + 286, + 213 + ], + "spans": [ + { + "bbox": [ + 53, + 192, + 286, + 213 + ], + "type": "text", + "content": "[4] David Bamman, Brendan O'Connor Noah, and A. Smith. Learning latent personas of film characters. 2014. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 214, + 287, + 257 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 214, + 287, + 257 + ], + "spans": [ + { + "bbox": [ + 53, + 214, + 287, + 257 + ], + "type": "text", + "content": "[5] Uttaran Bhattacharya, Gang Wu, Stefano Petrangeli, Viswanathan Swaminathan, and Dinesh Manocha. Highlighte: Detecting highlights from human-centric videos. In ICCV, 2021. 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 258, + 286, + 290 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 258, + 286, + 290 + ], + "spans": [ + { + "bbox": [ + 53, + 258, + 286, + 290 + ], + "type": "text", + "content": "[6] Joao Carreira and Andrew Zisserman. Quo vadis, action recognition? a new model and the kinetics dataset. 2017. 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 291, + 287, + 324 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 291, + 287, + 324 + ], + "spans": [ + { + "bbox": [ + 53, + 291, + 287, + 324 + ], + "type": "text", + "content": "[7] Shixing Chen, Xiaohan Nie, David Fan, Dongqing Zhang, Vimal Bhat, and Raffay Hamid. Shot contrastive self-supervised learning for scene boundary detection. 2021. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 325, + 286, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 325, + 286, + 357 + ], + "spans": [ + { + "bbox": [ + 53, + 325, + 286, + 357 + ], + "type": "text", + "content": "[8] Hao Cheng, Zhaowei Zhu, Xingyu Li, Yifei Gong, Xing Sun, and Yang Liu. Learning with instance-dependent label noise: A sample sieve approach. In ICLR, 2021. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 358, + 287, + 401 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 358, + 287, + 401 + ], + "spans": [ + { + "bbox": [ + 53, + 358, + 287, + 401 + ], + "type": "text", + "content": "[9] Yashar Deldjoo, Mihai Gabriel Constantin, and Bogdan Ionescu. Mmtf-14k: A multifaceted movie trailer feature dataset for recommendation and retrieval. In ACMMM, 2018. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 403, + 287, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 403, + 287, + 468 + ], + "spans": [ + { + "bbox": [ + 48, + 403, + 287, + 468 + ], + "type": "text", + "content": "[10] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Xiaohua Zhai Dirk Weissenborn, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. 2021. 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 469, + 287, + 524 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 469, + 287, + 524 + ], + "spans": [ + { + "bbox": [ + 48, + 469, + 287, + 524 + ], + "type": "text", + "content": "[11] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. 2021. 4" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 525, + 287, + 556 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 525, + 287, + 556 + ], + "spans": [ + { + "bbox": [ + 48, + 525, + 287, + 556 + ], + "type": "text", + "content": "[12] Curtis G, Northcutt, Lu Jiang, and Isaac L. Chuang. Confident learning: Estimating uncertainty in dataset labels. 2021. 5" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 558, + 287, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 558, + 287, + 601 + ], + "spans": [ + { + "bbox": [ + 48, + 558, + 287, + 601 + ], + "type": "text", + "content": "[13] Jort F. Gemmeke, Daniel P. W. Ellis, Dylan Freedman, Aren Jansen, Wade Lawrence, R. Channing Moore, Manoj Plakal, and Marvin Ritter. Audio set: An ontology and human-labeled dataset for audio events. 2017. 6" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 602, + 287, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 602, + 287, + 635 + ], + "spans": [ + { + "bbox": [ + 48, + 602, + 287, + 635 + ], + "type": "text", + "content": "[14] Michael Gygli, Yale Song, and Liangliang Cao. Video2gif: Automatic generation of animated gifs from video. In CVPR, 2016. 2, 7" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 635, + 287, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 635, + 287, + 679 + ], + "spans": [ + { + "bbox": [ + 48, + 635, + 287, + 679 + ], + "type": "text", + "content": "[15] Bo Han, Quanming Yao, Xingrui Yu, Gang Niu, Miao Xu, Weihua Hu, Ivor Tsang, and Masashi Sugiyama. Co-teaching: Robust training of deep neural networks with extremely noisy labels. In NeurIPS, 2018. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "type": "text", + "content": "[16] Fa-Ting Hong, Xuanteng Huang, Wei-Hong Li, and WeiShi Zheng. Mini-net: Multiple instance ranking network for video highlight detection. In ECCV, 2020. 2, 7" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 34, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "text", + "content": "[17] Jinchi Huang, Lie Qu, Rongfei Jia, and Binqiang Zhao. O2unet: A simple noisy label detection approach for deep neural network. In ICCV, 2019. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 107, + 545, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 107, + 545, + 140 + ], + "spans": [ + { + "bbox": [ + 307, + 107, + 545, + 140 + ], + "type": "text", + "content": "[18] Qingqiu Huang, Yu Xiong, Anyi Rao, Jiaze Wang, and Dahua Lin. Movienet: A holistic dataset for movie understanding. In ECCV, 2020. 2, 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 141, + 545, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 141, + 545, + 184 + ], + "spans": [ + { + "bbox": [ + 307, + 141, + 545, + 184 + ], + "type": "text", + "content": "[19] Qingqiu Huang, Yuanjun Xiong, Yu Xiong, Yuqi Zhang, and Dahua Lin. From trailers to storylines: An efficient way to learn from movies. arXiv preprint arXiv:1806.05341, 2018. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 186, + 545, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 186, + 545, + 228 + ], + "spans": [ + { + "bbox": [ + 307, + 186, + 545, + 228 + ], + "type": "text", + "content": "[20] Lu Jiang, Zhengyuan Zhou, Thomas Leung, Li-Jia Li, and Li Fei-Fei. Mentornet: Learning data-driven curriculum for very deep neural networks on corrupted labels. In ICML, 2018. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 230, + 545, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 230, + 545, + 274 + ], + "spans": [ + { + "bbox": [ + 307, + 230, + 545, + 274 + ], + "type": "text", + "content": "[21] Yifan Jiao, Zhetao Li, Shucheng Huang, Xiaoshan Yang, Bin Liu, and Tianzhu Zhang. Three-dimensional attention-based deep ranking model for video highlight detection. In IEEE Transactions on Multimedia, page 2693-2705, 2018. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 276, + 545, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 276, + 545, + 308 + ], + "spans": [ + { + "bbox": [ + 307, + 276, + 545, + 308 + ], + "type": "text", + "content": "[22] Jeff Johnson, Matthijs Douze, and Hervé Jégou. Billionscale similarity search with gpus. IEEE Transactions on Big Data, page 1-1, 2019. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 309, + 545, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 309, + 545, + 364 + ], + "spans": [ + { + "bbox": [ + 307, + 309, + 545, + 364 + ], + "type": "text", + "content": "[23] Will Kay, Joao Carreira, Karen Simonyan, Brian Zhang, Chloe Hillier, Sudheendra Vijayanarasimhan, Fabio Viola, Tim Green, Trevor Back, Paul Natsev, Mustafa Suleyman, and Andrew Zisserman. The kinetics human action video dataset. arXiv preprint arXiv:1705.06950, 2017. 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 365, + 545, + 420 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 365, + 545, + 420 + ], + "spans": [ + { + "bbox": [ + 307, + 365, + 545, + 420 + ], + "type": "text", + "content": "[24] Qiuqiang Kong, Yin Cao, Turab Iqbal, Yuxuan Wang, Wenwu Wang, and Mark D Plumbley. Panns: Large-scale pretrained audio neural networks for audio pattern recognition. IEEE/ACM Transactions on Audio, Speech, and Language Processing, 28:2880-2894, 2020. 4, 6" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 421, + 545, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 421, + 545, + 453 + ], + "spans": [ + { + "bbox": [ + 307, + 421, + 545, + 453 + ], + "type": "text", + "content": "[25] Jie Lei, Tamara L. Berg, and Mohit Bansal. Qvhighlights: Detecting moments and highlights in videos via natural language queries. In NeurIPS, 2021. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 455, + 545, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 455, + 545, + 487 + ], + "spans": [ + { + "bbox": [ + 307, + 455, + 545, + 487 + ], + "type": "text", + "content": "[26] Junnan Li, Richard Socher, and Steven C.H. Hoi. Dividemix: Learning with noisy labels as semi-supervised learning. 2020. 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 488, + 545, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 488, + 545, + 542 + ], + "spans": [ + { + "bbox": [ + 307, + 488, + 545, + 542 + ], + "type": "text", + "content": "[27] Shuaicheng Li, Feng Zhang, Kunlin Yang, Lingbo Liu, Shinan Liu, Jun Hou, and Shuai Yi. Probing visual-audio representation for video highlight detection via hard-pairs guided contrastive learning. arXiv preprint arXiv:2206.10157, 2022. 7" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 544, + 545, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 544, + 545, + 567 + ], + "spans": [ + { + "bbox": [ + 307, + 544, + 545, + 567 + ], + "type": "text", + "content": "[28] Yang Liu and Hongyi Guo. Peer loss functions: Learning from noisy labels without knowing noise rates. 2020. 3" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 307, + 567, + 545, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 567, + 545, + 611 + ], + "spans": [ + { + "bbox": [ + 307, + 567, + 545, + 611 + ], + "type": "text", + "content": "[29] Ye Liu, Siyuan Li, Yang Wu, Chang Wen Chen, Ying Shan, and Xiaohu Qie. Umt: Unified multi-modal transformers for joint video moment retrieval and highlight detection. In CVPR, 2022. 2, 6, 7, 8" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 307, + 613, + 545, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 613, + 545, + 645 + ], + "spans": [ + { + "bbox": [ + 307, + 613, + 545, + 645 + ], + "type": "text", + "content": "[30] Xingjun Ma, Hanxun Huang, Yisen Wang, Simone Romano, Sarah Erfani, and James Bailey. Normalized loss functions for deep learning with noisy labels. 2020. 3" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 307, + 647, + 545, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 647, + 545, + 678 + ], + "spans": [ + { + "bbox": [ + 307, + 647, + 545, + 678 + ], + "type": "text", + "content": "[31] Curtis G. Northcutt, Lu Jiang, and Isaac L. Chuang. Confident learning: Estimating uncertainty in dataset labels. In JAIR, 2019. 3" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 307, + 680, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 680, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 680, + 545, + 713 + ], + "type": "text", + "content": "[32] Sinno Jialin Pan and Qiang Yang. A survey on transfer learning. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, page 1345-1359, 2010. 2" + } + ] + } + ], + "index": 33 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "text", + "content": "18906" + } + ] + } + ], + "index": 35 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "text", + "content": "[33] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Sandhini Agarwal Gabriel Goh, Girish Sastry, Amanda Askell, Pamela Mishkin, and et al. Jack Clark. Learning transferable visual models from natural language supervision. arXiv preprint arXiv:2103.00020, 2021. 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 128, + 287, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 128, + 287, + 161 + ], + "spans": [ + { + "bbox": [ + 48, + 128, + 287, + 161 + ], + "type": "text", + "content": "[34] Anyi Rao, Linning Xu, Yu Xiong, Guodong Xu, Qingqiu Huang, Bolei Zhou, and Dahua Lin. A local-to-global approach to multi-modal movie scene segmentation. 2020. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 162, + 287, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 162, + 287, + 194 + ], + "spans": [ + { + "bbox": [ + 48, + 162, + 287, + 194 + ], + "type": "text", + "content": "[35] Yale Song, Yahoo Labs, New York, Jordi Vallmitjana, Amanda Stent, and Alejandro Jaime. Tvsum: Summarizing web videos using titles. In CVPR, 2015. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 195, + 287, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 195, + 287, + 227 + ], + "spans": [ + { + "bbox": [ + 48, + 195, + 287, + 227 + ], + "type": "text", + "content": "[36] Tomáš Souček and Jakub Lokoc. Transnet v2: An effective deep network architecture for fast shot transition detection. arXiv preprint arXiv:2008.04838, 2020. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 228, + 287, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 228, + 287, + 259 + ], + "spans": [ + { + "bbox": [ + 48, + 228, + 287, + 259 + ], + "type": "text", + "content": "[37] Min Sun, Ali Farhadi, and Steve Seitz. Ranking domain-specific highlights by analyzing edited videos. In ECCV, 2014. 2, 6, 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 261, + 287, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 261, + 287, + 303 + ], + "spans": [ + { + "bbox": [ + 48, + 261, + 287, + 303 + ], + "type": "text", + "content": "[38] Zeren Sun, Fumin Shen, Dan Huang, Qiong Wang, Xiangbo Shu, Yazhou Yao, and Jinhui Tang. Pnp: Robust learning from noisy labels by probabilistic noise prediction. In CVPR, 2022. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 304, + 287, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 304, + 287, + 348 + ], + "spans": [ + { + "bbox": [ + 48, + 304, + 287, + 348 + ], + "type": "text", + "content": "[39] Zeren Sun, Yazhou Yao, Xiu-Shen Wei, Yongshun Zhang, Fumin Shen, Jianxin Wu, Jian Zhang, and Heng-Tao Shen. Webly supervised fine-grained recognition-benchmark datasets and an approach. In ICCV, 2021. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 350, + 287, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 350, + 287, + 381 + ], + "spans": [ + { + "bbox": [ + 48, + 350, + 287, + 381 + ], + "type": "text", + "content": "[40] Christian Szegedy, Vincent Vanhoucke, Sergey Ioffe, Jonathon Shlens, and Zbigniew Wojna. Rethinking the inception architecture for computer vision. 2016. 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 383, + 287, + 414 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 383, + 287, + 414 + ], + "spans": [ + { + "bbox": [ + 48, + 383, + 287, + 414 + ], + "type": "text", + "content": "[41] Hao Tang, Vivek Kwatra, Mehmet Emre Sargin, and Ullas Gargi. Detecting highlights in sports videos: Cricket as a test case. In ICME, 2011. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 415, + 287, + 447 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 415, + 287, + 447 + ], + "spans": [ + { + "bbox": [ + 48, + 415, + 287, + 447 + ], + "type": "text", + "content": "[42] Paul Vicol, Makarand Tapaswi, Lluis Castrejon, and Sanja Fidler. Towards understanding human-centric situations from videos. 2018. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 449, + 287, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 449, + 287, + 480 + ], + "spans": [ + { + "bbox": [ + 48, + 449, + 287, + 480 + ], + "type": "text", + "content": "[43] Lezi Wang, Dong Liu, Rohit Puri, and Dimitris N. Metaxas. Learning trailer moments in full-length movies with co-contrastive attention. In ECCV, 2020. 2, 3, 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 482, + 287, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 482, + 287, + 514 + ], + "spans": [ + { + "bbox": [ + 48, + 482, + 287, + 514 + ], + "type": "text", + "content": "[44] Yisen Wang, Xingjun Ma, Zaiyi Chen, Yuan Luo, Jinfeng Yi, and James Bailey. Symmetric cross entropy for robust learning with noisy labels. 2019. 3, 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 515, + 287, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 515, + 287, + 547 + ], + "spans": [ + { + "bbox": [ + 48, + 515, + 287, + 547 + ], + "type": "text", + "content": "[45] Fanyue Wei, Biao Wang, Tiezheng Ge, Yuning Jiang, Wen Li, and Lixin Duan. Learning pixel-level distinctions for video highlight detection. In CVPR, 2022. 2, 7" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 548, + 287, + 580 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 548, + 287, + 580 + ], + "spans": [ + { + "bbox": [ + 48, + 548, + 287, + 580 + ], + "type": "text", + "content": "[46] Hongxin Wei, Lei Feng, Xiangyu Chen, and Bo An. Combating noisy labels by agreement: A joint training method with co-regularization. In CVPR, 2020. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 582, + 287, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 582, + 287, + 624 + ], + "spans": [ + { + "bbox": [ + 48, + 582, + 287, + 624 + ], + "type": "text", + "content": "[47] Haoqian Wu, Keyu Chen, Yanan Luo, Ruizhi Qiao, Bo Ren, Haozhe Liu, Weicheng Xie, and Linlin Shen. Scene consistency representation learning for video scene segmentation. In CVPR, 2022. 2, 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 625, + 287, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 625, + 287, + 668 + ], + "spans": [ + { + "bbox": [ + 48, + 625, + 287, + 668 + ], + "type": "text", + "content": "[48] Xiaobo Xia, Tongliang Liu, Bo Han, Nannan Wang, Mingming Gong, Haifeng Liu, Gang Niu, Dacheng Tao, and Masashi Sugiyama. Part-dependent label noise: Towards instance-dependent label noise. In NeurIPS, 2020. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "type": "text", + "content": "[49] Xiaobo Xia, Tongliang Liu, Nannan Wang, Bo Han, Chen Gong, Gang Niu, and Masashi Sugiyama. Are anchor points really indispensable in label-noise learning? In NeurIPS, 2019. 3" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 386 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 115 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 115 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 115 + ], + "type": "text", + "content": "[50] Minghao Xu, Hang Wang, Bingbing Ni, Riheng Zhu, Zhenbang Sun, and Changhu Wang. Cross-category video highlight detection via set-based learning. In ICCV, 2021. 2, 6, 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 118, + 545, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 545, + 162 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 545, + 162 + ], + "type": "text", + "content": "[51] Fengxiang Yang, Ke Li, Zhun Zhong, Zhiming Luo, Xing Sun, Hao Cheng, Xiaowei Guo, Feiyue Huang, Rongrong Ji, and Shaozi Li. Asymmetric co-teaching for unsupervised cross domain person re-identification. In AAAI, 2020. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 163, + 545, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 163, + 545, + 205 + ], + "spans": [ + { + "bbox": [ + 307, + 163, + 545, + 205 + ], + "type": "text", + "content": "[52] Shuo Yang, Erkun Yang, Bo Han, Yang Liu, Min Xu, Gang Niu, and Tongliang Liu. Estimating instance-dependent bayes-label transition matrix using a deep neural network. In CVPR, 2022. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 207, + 545, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 207, + 545, + 239 + ], + "spans": [ + { + "bbox": [ + 307, + 207, + 545, + 239 + ], + "type": "text", + "content": "[53] Ting Yao, Tao Mei, and Yong Rui. Highlight detection with pairwise deep ranking for first-person video summarization. In CVPR, 2016. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 241, + 545, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 241, + 545, + 285 + ], + "spans": [ + { + "bbox": [ + 307, + 241, + 545, + 285 + ], + "type": "text", + "content": "[54] Chuanyi Zhang, Yazhou Yao*, Xing Xu, Jie Shao, Jingkuan Song, Zechao Li, and Zhenmin Tang. Extracting useful knowledge from noisy web images via data purification for fine-grained recognition. In ACMMM, 2021. 5" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 286, + 545, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 286, + 545, + 319 + ], + "spans": [ + { + "bbox": [ + 307, + 286, + 545, + 319 + ], + "type": "text", + "content": "[55] Zhilu Zhang and Mert R. Sabuncu. Generalized cross entropy loss for training deep neural networks with noisy labels. In NeurIPS, 2018. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 320, + 545, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 320, + 545, + 352 + ], + "spans": [ + { + "bbox": [ + 307, + 320, + 545, + 352 + ], + "type": "text", + "content": "[56] Guoqing Zheng, Ahmed Hassan Awadallah, and Susan Dumas. Meta label correction for noisy label learning. In AAAI, 2021. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 354, + 545, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 354, + 545, + 386 + ], + "spans": [ + { + "bbox": [ + 307, + 354, + 545, + 386 + ], + "type": "text", + "content": "[57] A. Zlatintsi, P. Koutras, N. Efthymiou, P. Maragos, A. Potamianos, and K. Pastra. Quality evaluation of computational models for movie summarization. 2015. 2" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18907" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Collaborative Static and Dynamic Vision-Language Streams for Spatio-Temporal Video Grounding/40d01bed-f6c0-4114-9925-d43e564f9389_content_list.json b/2023/Collaborative Static and Dynamic Vision-Language Streams for Spatio-Temporal Video Grounding/40d01bed-f6c0-4114-9925-d43e564f9389_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..2205e7637d78ebdc6a81bba2eed2f15de9bdc66d --- /dev/null +++ b/2023/Collaborative Static and Dynamic Vision-Language Streams for Spatio-Temporal Video Grounding/40d01bed-f6c0-4114-9925-d43e564f9389_content_list.json @@ -0,0 +1,1417 @@ +[ + { + "type": "text", + "text": "Collaborative Static and Dynamic Vision-Language Streams for Spatio-Temporal Video Grounding", + "text_level": 1, + "bbox": [ + 181, + 130, + 789, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Zihang Lin1, Chaolei Tan1, Jian-Fang Hu1,3,4*, Zhi Jin1, Tiancai Ye2, Wei-Shi Zheng1,3,4, Sun Yat-sen University, China", + "bbox": [ + 142, + 202, + 825, + 237 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3Guangdong Province Key Laboratory of Information Security Technology, China \n4 Key Laboratory of Machine Intelligence and Advanced Computing, Ministry of Education, China", + "bbox": [ + 93, + 239, + 874, + 273 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{1inzh59, tanchlei}@mail2.sysu.edu.cn, {hujf5, jinzh26}@mail.sysu.edu.cn, tiancaiye@tencent.com, wszheng@ieee.org", + "bbox": [ + 163, + 276, + 803, + 309 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 344, + 313, + 359 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Spatio-Temporal Video Grounding (STVG) aims to localize the target object spatially and temporally according to the given language query. It is a challenging task in which the model should well understand dynamic visual cues (e.g., motions) and static visual cues (e.g., object appearances) in the language description, which requires effective joint modeling of spatio-temporal visual-linguistic dependencies. In this work, we propose a novel framework in which a static vision-language stream and a dynamic vision-language stream are developed to collaboratively reason the target tube. The static stream performs cross-modal understanding in a single frame and learns to attend to the target object spatially according to intra-frame visual cues like object appearances. The dynamic stream models visual-linguistic dependencies across multiple consecutive frames to capture dynamic cues like motions. We further design a novel cross-stream collaborative block between the two streams, which enables the static and dynamic streams to transfer useful and complementary information from each other to achieve collaborative reasoning. Experimental results show the effectiveness of the collaboration of the two streams and our overall framework achieves new state-of-the-art performance on both HCSTVG and VidSTG datasets.", + "bbox": [ + 75, + 377, + 473, + 739 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 771, + 209, + 787 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Vision-language cross-modal understanding is a challenging yet important research problem that bridges the communication between humans and artificial intelligence systems. It has attracted increasing attention and many vision-language tasks were studied in recent years, like vi", + "bbox": [ + 75, + 797, + 468, + 875 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "sual grounding [13, 32], VQA [7, 31], image/video captioning [15, 40], etc. In this work, we focus on a challenging vision-language task named Spatio-Temporal Video Grounding (STVG) which was recently proposed in [42]. Given a language query indicating an object (as shown in Figure 1), STVG aims to localize the target object spatially and temporally in the video. In this task, the input language query may express different kinds of visual concepts, thus the model requires to well capture and understand these concepts in both vision and language modalities.", + "bbox": [ + 496, + 345, + 893, + 496 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In STVG task, dynamic visual concepts like human motion and static visual concepts like object appearance are both important for distinguishing the target object from other objects that occurred in the same video. For example, in the first sample in Figure 1, the two men are dressed alike (i.e., their static appearance cues are similar), and we can only distinguish them by motion. In the second sample, the two women perform similar actions, they both stand up (i.e., they have similar dynamic motion cues), here, we can only distinguish them by their clothes. The above examples show that static or dynamic visual cues alone cannot solve the STVG task well. And it implies that modeling static and dynamic visual-linguistic dependencies and collaboratively utilizing them are important for addressing STVG task.", + "bbox": [ + 496, + 502, + 893, + 715 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Humans treat static and dynamic cues differently [4, 21], but this was overlooked in previous STVG works [22, 24, 30]. Taking the first query in Figure 1 as an example, a human would randomly or evenly click on some locations on the video's progress bar to find candidate frames containing a man in blue coat. And he will play the video around that frame and attend on the candidate man to check whether he performs the action described in the text (i.e., \"turns around and stops by the stone\"). In the above process, the human understands static and dynamic cues in different ways (i.e., view a single frame and watch the video clip, respectively) and determines the target object by jointly considering the", + "bbox": [ + 496, + 719, + 895, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author.", + "bbox": [ + 94, + 886, + 220, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "23100", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Query: The man in the blue coat turns around and stops by the stone.", + "bbox": [ + 133, + 93, + 589, + 107 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/03aab8328a8d944b4c530af2fe68465433d7aee9e43f259850bd582dda0c5472.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 130, + 112, + 834, + 179 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Query: The woman in light gray stands up and turns.", + "bbox": [ + 133, + 181, + 478, + 196 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/da337362c1f2cc61ea9ebfa9a52fba8ebcd14cbbf6c017d6e80b21fbf4192c51.jpg", + "image_caption": [ + "Figure 1. Two examples for Spatio-Temporal Video Grounding task. In the first case, the two men are both dressed alike in blue, thus understanding the action described in the query sentence is essential to recognize the person of interest. In the second case, both of the two women stand up, we can only distinguish them by their clothes. (Best viewed zoomed in on screen.)" + ], + "image_footnote": [], + "bbox": [ + 130, + 200, + 834, + 268 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "static and dynamic cues in an \"attend-and-check\" process. Inspired by the above observations, we propose a framework that consists of a static and a dynamic vision-language (VL) streams to model static and dynamic cues, respectively. And we design a novel cross-stream collaboration block between the two streams to simulate the \"attend-and-check\" process, which exchanges useful and complementary information learned in each stream and enables the two streams to collaboratively reason the target object.", + "bbox": [ + 75, + 330, + 468, + 467 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Specifically, in this work, our static vision-language (VL) stream learns to attend to some candidate regions according to static visual cues like appearance, while the dynamic VL stream learns to understand the dynamic visual cues like action described in the text query. Then, in the collaboration block, we guide the dynamic stream to only focus and attend on the motion of the candidate objects by using the learned attended region in the static stream. And we transfer the text-motion matching information learned in the dynamic stream to the static stream, to help it further check and determine the target object and predict a more consistent tube. With the above cross-stream collaboration blocks, both the static and dynamic vision-language streams can learn reciprocal information from the other stream, which is effective for achieving more accurate spatio-temporal grounding predictions. We conduct experiments on HCSTVG [24] and VidSTG [42] datasets and our approach outperforms previous approaches by a considerable margin. Ablation studies demonstrate the effectiveness of each component in our proposed cross-stream collaboration block and show its superiority over commonly used counterparts in video understanding works [4, 21].", + "bbox": [ + 75, + 472, + 470, + 805 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In summary, our contributions are: 1), we develop an effective framework that contains two parallel streams to model static-dynamic visual-linguistic dependencies for complete cross-modal understanding; 2), we propose a novel cross-stream collaboration block between the two streams to exchange reciprocal information for each other", + "bbox": [ + 75, + 810, + 470, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "and enable collaborative reasoning of the target object; 3). Our overall framework achieves state-of-the-art performances on HCSTVG [24] and VidSTG datasets [42].", + "bbox": [ + 498, + 330, + 890, + 378 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 391, + 640, + 406 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Spatio-Temporal Video Grounding. Spatio-temporal video grounding aims to localize the target object spatially and temporally according to a language query. Early works [24, 41, 42] employ pre-trained detectors like Faster R-CNN [18] to detect objects in each frame, and build their model upon the detection features of each object. STGRN [42] and OMRN [41] learn object relations with spatio-temporal graphs and multi-branch relation networks, respectively. STGVT [24] learns cross-modal representations with visual transformers for video-sentence matching and temporal localization. However, these methods are limited to the pre-detected bounding boxes and cannot localize object categories that are not defined in the pretraining dataset. Recent works [22, 30] design their framework based on strong pre-trained vision-language models and train their models end-to-end. These approaches do not require pre-detected bounding boxes and they perform much better. However, they directly collapse some dimensions (i.e., spatial or temporal) in cross-modal attention calculation to reach an acceptable computation cost, which results in large information loss, and thus they can not well capture the important static-dynamic cross-modal contextual relationship. To address this weakness, we propose to model static and dynamic visual-linguistic correspondence with parallel streams and learn to reason the target object in a collaborative framework.", + "bbox": [ + 496, + 417, + 890, + 808 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Temporal Grounding. Given a language query, temporal grounding aims to predict the time span in a video specified by the language query [1, 6]. Existing works [11, 16, 17, 29, 35, 36, 38, 39] can be divided into proposal-based and proposal-free approaches. Proposal-based approaches generate some proposals (manually designed [29, 39] or", + "bbox": [ + 496, + 810, + 893, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "23101", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/dbdfe90608f64b0d4cad33eee37bb20dccab73e0ad187dbee22bc5ab28785bda.jpg", + "image_caption": [ + "Figure 2. An overview of the proposed framework. Our framework mainly consists of a static VL( vision-language) stream and a dynamic VL stream. The static stream learns to attend to the spatial locations of the target object according to static cues like object appearance. The dynamic stream learns motion-text correspondence according to dynamic cues like human action. We further devise a cross-stream collaboration block that enables the two streams to query useful and complementary information from the other stream." + ], + "image_footnote": [], + "bbox": [ + 106, + 85, + 867, + 295 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "predicted by the model [35]) and predict the IoU scores between each proposal and the ground truth. Proposal-free approaches address moment localization by directly regressing the starting/ending points of the target time spans [11, 16] or predicting the probability of being a starting/ending point for each clip and choosing the time span with the highest joint probability as the prediction [38]. In this work, we follow [23, 39] to employ a proposal-based grounding head. And in contrast to most temporal grounding works that collapse the spatial dimension, we also incorporate spatial cues to enhance temporal localization.", + "bbox": [ + 75, + 369, + 468, + 536 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Text-conditioned Object Detection. Given an image, text-conditioned object detection aims to localize the target object indicated by a sentence (also named referring image expression task) or all objects described in a sentence (also named phrase grounding task). Most existing works [32,37] develop their models based on off-the-shelf detection models and match the detected objects with the sentence query. Recently, MDETR [10] and GLIP [12] unified detection and phrase grounding tasks, and they trained their models with large-scale data. These models can achieve excellent performance on text-conditioned object detection and are used to initialize model weights for STVG task [23, 30]. However, these models lack the ability of modeling dynamic cues. In this work, we design a novel collaboration framework to capture both static and dynamic cues for STVG task.", + "bbox": [ + 75, + 537, + 468, + 762 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Two-stream Models. Considering that static and dynamic visual cues are quite different but complementary in video understanding, a number of two-stream models [4,5,21,25, 26, 28] are proposed to capture these two kinds of visual cues. For example, Two-Stream ConvNet [21] introduces an optical flow stream to capture object motion. SlowFast [4] develops slow and fast pathways with different temporal resolutions to capture static and dynamic visual cues. However, these works typically fuse different streams equally", + "bbox": [ + 75, + 763, + 468, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "via some simple symmetric operations, e.g., assembling the output, sum/concat/FC/temporal-convolution on the features. In contrast, we propose a novel asymmetrical cross-stream collaboration block to exploit better collaboration between the two streams considering their different abilities.", + "bbox": [ + 498, + 369, + 890, + 446 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Collaborative Static-Dynamic VL Streams", + "text_level": 1, + "bbox": [ + 498, + 465, + 877, + 483 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We illustrate our framework in Figure 2. Our framework mainly consists of a static and a dynamic VL (vision-language) stream to model static and dynamic visual-linguistic dependencies for complete cross-modal understanding. The static stream performs cross-modal understanding for static contextual information, i.e., finding the object that matches the query text in a still frame according to static visual cues, like object appearance which is important for achieving accurate spatial grounding. The dynamic stream performs cross-modal understanding for dynamic contextual information, i.e., finding the temporal moment that best matches the query sentence according to dynamic visual cues, like action which is important for achieving accurate temporal grounding. Both the static and dynamic streams learn cross-modal correspondence with a transformer-like architecture that contains multiple layers. To enable information transmission between the two streams, we further design a novel cross-stream collaboration block which is placed after each transformer layer in the two streams. In this way, both the static and dynamic VL streams can absorb complementary information from each other and achieve collaborative inferring, which can greatly reduce the uncertainty in ambiguous and hard cases where different objects have similar appearances or perform similar actions. Finally, we employ the prediction head to predict a spatio-temporal tube. In the following, we will introduce each component in detail.", + "bbox": [ + 496, + 493, + 892, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "23102", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/99dd6c752da3f4540f0da8e13041971ab5b12d44fb8da44ebe6bc5754c2af8ce.jpg", + "image_caption": [ + "(a)" + ], + "image_footnote": [], + "bbox": [ + 125, + 90, + 361, + 244 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/385c90247b5e7890bb1ce7fb01fa08a070f95f4c94f1a31e4300c9cad16b4863.jpg", + "image_caption": [ + "(b)", + "Figure 3. The detailed architectures of the proposed framework. (a): The cross-modal transformer in the static VL stream. (b): The architecture of each space-time cross-modal transformer layer in the dynamic VL stream." + ], + "image_footnote": [], + "bbox": [ + 385, + 90, + 836, + 247 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1. Static VL Stream", + "text_level": 1, + "bbox": [ + 76, + 306, + 246, + 321 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The static VL stream is employed to perform cross-modal static context understanding and it processes each frame independently. Inspired by DETR [3], we design our static VL stream as a stack of $N$ cross-modal transformer encoder layers and $N$ cross-modal transformer decoder layers, as presented in Figure 2 and 3 (a). For the $t$ -th frame, the inputs of our cross-modal transformer encoder are the concatenation of $\\mathbb{R}^{HW\\times d}$ -sized visual features and $\\mathbb{R}^{L\\times d}$ -sized language features, where $H,W$ are the resolution of the visual features extracted from the static image frame, and $L$ is the number of text tokens in the input text query. These visual and language features are obtained by a pretrained image encoder [9] and language encoder [14], respectively. We employ an FC layer behind each feature extractor to project the visual and language features to a shared embedding space. The outputs of the last cross-modal transformer encoder layer form a $\\mathbb{R}^{(HW + L)\\times d}$ -sized cross-modal memory $\\mathcal{M}_t$ , which captures rich interactions between intra-frame static visual cues and linguistic descriptions. Then in the decoder, a learnable object query vector $O_{t}^{0}\\in \\mathbb{R}^{d}$ is inputted to repeatedly query object appearance and location information from the memory $\\mathcal{M}_t$ via a cross-attention layer. Here, the object query gradually learns to attend to the object that matched the text query.", + "bbox": [ + 75, + 330, + 473, + 695 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Dynamic VL Stream", + "text_level": 1, + "bbox": [ + 76, + 709, + 272, + 724 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The dynamic VL stream performs cross-modal understanding for dynamic contextual information. As illustrated in Figure 2, we first extract clip-level visual features $F_{clip} \\in \\mathbb{R}^{T \\times H \\times W \\times c}$ from $T$ uniformly sampled video clips by a pre-trained 3D-CNN [4]. Then we employ an FC layer to project the channel dimension from $c$ to $d$ and obtain $F_v^0$ . $F_v^0$ is fed into a Space-Time Cross-Modal Transformer (STCMT) which consists of $N$ layers to model the visual-linguistic dependencies from a dynamic perspective. The detailed architecture of each layer in STCMT is illustrated in Figure 3 (b). In each layer, we first per", + "bbox": [ + 75, + 734, + 470, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "form intra-modality self-attention for the dynamic visual features and linguistic features. For visual features, in order to reduce computation cost, we follow TimesFormer [2] to split the spatio-temporal attention into separate attentions. Denoting the visual features after self-attention as $\\mathcal{H}_v\\in \\mathbb{R}^{T\\times H\\times W\\times d}$ and the linguistic features after self-attention as $\\mathcal{H}_l\\in \\mathbb{R}^{L\\times d}$ . We then perform cross-attention between $\\mathcal{H}_v$ and $\\mathcal{H}_l$ as:", + "bbox": [ + 496, + 306, + 893, + 429 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} Q _ {v} ^ {(h, w)} = W _ {v} ^ {q} \\mathcal {H} _ {v} ^ {(h, w)}, K _ {v} = W _ {v} ^ {k} \\overline {{\\mathcal {H}}} _ {v}, V _ {v} = W _ {v} ^ {v} \\overline {{\\mathcal {H}}} _ {v}, \\\\ Q _ {l} = W _ {l} ^ {v} \\mathcal {H} _ {l}, K _ {l} = W _ {l} ^ {k} \\mathcal {H} _ {l}, V _ {l} = W _ {l} ^ {v} \\mathcal {H} _ {l}, \\\\ \\widetilde {\\mathcal {H}} _ {v} ^ {(h, w)} = \\mathcal {H} _ {v} ^ {(h, w)} + \\text {A t t e n t i o n} \\left(Q _ {v} ^ {(h, w)}, K _ {l}, V _ {l}\\right), \\tag {1} \\\\ \\widetilde {\\mathcal {H}} _ {l} = \\mathcal {H} _ {l} + \\operatorname {A t t e n t i o n} \\left(Q _ {l}, K _ {v}, V _ {v}\\right), \\\\ \\operatorname {A t t e n t i o n} (Q, K, V) = \\operatorname {s o f t m a x} \\left(\\frac {Q K ^ {T}}{\\sqrt {d _ {k}}}\\right) V, \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 514, + 434, + 890, + 555 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $W_{v}^{q}, W_{v}^{k}, W_{v}^{v}, W_{l}^{q}, W_{l}^{k}, W_{l}^{v}$ are learnable weighted matrices for computing queries, keys and values for the attention mechanism. $\\overline{\\mathcal{H}}_v \\in \\mathbb{R}^{T \\times d}$ is obtained by conducting mean pooling on $\\mathcal{H}_v$ along the spatial dimensions. $d_{k}$ is the dimension of the queries and keys. $\\mathcal{H}_{v}^{(h,w)} \\in \\mathbb{R}^{T \\times d}$ indicates the visual feature in $\\mathcal{H}_v$ at spatial position $(h,w)$ . $\\widetilde{\\mathcal{H}}_v, \\widetilde{\\mathcal{H}}_l$ are the output visual and linguistic features after cross-attention, respectively. The cross-attention operation is designed to explore rich interactions between the visual and linguistic features, which enables the model to learn a powerful cross-modal representation of the depicted dynamic cues. This is essential for grounding temporal moments. After computing the cross-attention between the visual features and linguistic features, we finally employ Feed-Forward Network (FFN) to process both features. In our dynamic VL stream, rich context is learned from the two modalities, and the visual dynamic features and the linguistic features are fused well.", + "bbox": [ + 496, + 561, + 892, + 839 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Cross-Stream Collaboration", + "text_level": 1, + "bbox": [ + 500, + 847, + 754, + 861 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To enable the static and the dynamic VL streams to learn complementary information from each other and to perform", + "bbox": [ + 498, + 869, + 893, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "23103", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/fd3fa8a7f1b0b333cb0fc5874d730b2f0015b15612af26863f8a3957bc4e91b1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 89, + 85, + 467, + 213 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/6c37b60705fa816cbaa988687d898d56feb542f3dd93943eb7403dcfadbea6d4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 89, + 213, + 467, + 321 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/339803897c87cc76b22403108c1a22ed1a942e2496deac32ae83e5af9f680cf5.jpg", + "image_caption": [ + "Figure 4. The architectures of our cross-stream collaboration." + ], + "image_footnote": [], + "bbox": [ + 89, + 321, + 465, + 449 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "collaborative reasoning, we design a novel cross-stream collaboration block between the two streams. As shown in Figure 4, this block is placed after each decoder layer $\\{\\mathcal{B}_s^i\\}_{i=1}^N$ in static stream and each layer $\\{\\mathcal{B}_a^i\\}_{i=1}^N$ in dynamic stream. It consists of a static-to-dynamic and a dynamic-to-static information transmission block. They are designed asymmetrically to exploit better collaboration between the two streams considering their different abilities. In the following, we will introduce the detailed designs.", + "bbox": [ + 75, + 489, + 468, + 632 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The static-to-dynamic information transmission block is designed to guide the dynamic VL stream to attend to the spatial region that is highly related to the objects depicted in the query text, by utilizing the cross-attention weights learned in the decoder layers of the static VL stream. Concretely, it is formulated as follows:", + "bbox": [ + 75, + 632, + 468, + 723 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\widetilde {F} _ {v} ^ {i} = \\operatorname {L a y e r N o r m} \\left(F _ {v} ^ {i} + A ^ {i} \\odot \\mathrm {F C} \\left(F _ {v} ^ {i}\\right)\\right), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 140, + 731, + 468, + 752 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $F_{v}^{i}\\in \\mathbb{R}^{T\\times HW\\times d}$ is the output visual feature of the $i$ -th layer in dynamic VL stream and $A^i\\in \\mathbb{R}^{T\\times HW\\times d}$ is the cross-attention weights (replicate $d$ times to have $d$ channels at the last dimension) calculated between the object queries $\\left\\{O_t^i\\right\\}_{t = 1}^T$ and corresponding encoded memories $\\{M_t\\}_{t = 1}^T$ in $\\mathcal{B}_s^i$ . $\\odot$ indicates hadamard product. The intuition behind this design is that the cross-attention weights learned in the static VL stream can attend to regions matching the language description, and this can serve as a strong guidance", + "bbox": [ + 76, + 758, + 470, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Query: The tall boy comes to the dining table and sits down.", + "bbox": [ + 504, + 90, + 674, + 114 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/178f155f96040b622396c008ac822160ae693dfa61a481bd9af0019d031f549b.jpg", + "image_caption": [ + "Figure 5. Visualization of the learned attention map." + ], + "image_footnote": [], + "bbox": [ + 504, + 117, + 697, + 215 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Query: The man in the red stripe goes to the sofa and turns.", + "bbox": [ + 697, + 90, + 870, + 114 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/efdbc4396c7fdd480f50bc14d109099cdfce2d7c5b37360176086b3772243c7c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 699, + 116, + 792, + 165 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/61c9a497fe65480b659b6bbb1a7ae2771b667ec8d266b07c9f7bfb3b603d64d3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 794, + 117, + 887, + 165 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/939e3cf9fd1c63ff29799aaaa52e4c99e99f4b4df82d61bbd1901cd35d5cd658.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 700, + 166, + 792, + 215 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/b4abc11a50067098853f1e8d350f971f7749281da6bcbb45ddae82ca51bc1aff.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 794, + 166, + 887, + 215 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "to help the dynamic VL stream focus more on the dynamic variations around the object-related regions. We visualize the attention maps in Figure 5. In the first sample, there are multiple boys and the attention map has a high score at the best-matched person (i.e., the tallest boy in each frame). In the second sample, we observe that when the man in red stripe is absent, the attention weights are relatively smooth. However, the attention weights become sharp once the target man appeared. These samples intuitively show that our static-to-dynamic information transmission block can guide the dynamic stream to focus on motions of the target object.", + "bbox": [ + 496, + 258, + 890, + 425 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The dynamic-to-static information transmission block is designed to transfer the learned motion-text correspondence to the static VL stream. It enhances $O_{t}^{i}$ to $\\widetilde{O}_{t}^{i}$ with a cross-attention and a query mixing operation as depicted in Figure 4(c). In each block, the object query $O_{t}^{i}$ in the static stream first queries some dynamic information from $F_{v}^{i}[t]$ (the $t$ -th feature in $F_{v}^{i}$ ) via a cross-attention module, this operation enhances the object query representations with dynamic information learned in the dynamic stream. Then we employ a query mixing operation (implemented as temporal self-attention) to query and mix information from object queries of other frames, which further injects cross-frame information to object queries. With the proposed cross-stream collaboration block, both the static and dynamic VL streams can learn complementary information from each other in an effective way. And the enhanced representations $\\widetilde{F}_{v}^{i}, \\widetilde{O}_{t}^{i}$ are inputted to the next transformer layer in the static and dynamic VL streams, respectively.", + "bbox": [ + 496, + 426, + 892, + 698 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Prediction Heads", + "text_level": 1, + "bbox": [ + 500, + 709, + 668, + 723 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this section, we introduce our prediction heads which include a spatial and a temporal prediction head.", + "bbox": [ + 496, + 732, + 890, + 762 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Spatial Prediction Head. The spatial prediction head predicts the location of the target object. For the $t$ -th frame, the input is the object query representation $\\widetilde{O}_t^N \\in \\mathbb{R}^d$ outputted by the last collaboration block. And we implement a 3-layer MLP to predict the bounding box location (represented by center coordinates and size) $\\tilde{b}_t \\in \\mathbb{R}^4$ of the target object.", + "bbox": [ + 496, + 763, + 890, + 854 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Temporal Prediction Head. The temporal prediction head predicts the time span of the target temporal moment. Specifically, we first perform spatial mean pooling on $\\widetilde{F}_v^N$", + "bbox": [ + 496, + 854, + 890, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "23104", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/8661c7db511363ea09ac4d23e6146d65b818a12e112ae8ba05d65e362d857126.jpg", + "table_caption": [ + "Table 1. Comparison results with previous works on HCSTVG-v1 test set and HCSTVG-v2 val set [24]." + ], + "table_footnote": [], + "table_body": "
DatasetMethodm_vIoUvIoU@0.3vIoU@0.5
HCSTVG-v1STGVT [24]18.226.89.5
STVGBert [22]20.429.411.3
TubeDETR [30]32.449.823.5
Ours36.962.234.8
HCSTVG-v2Yu et al. [33]30.0--
MMN [27]30.349.025.6
Aug. 2D-TAN [23]30.450.418.8
TubeDETR [30]36.458.830.6
Ours38.765.533.8
", + "bbox": [ + 91, + 128, + 415, + 281 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/c2574383d49ef2e51014539b9eb34687da3e850f6c4c46906f1cbe71d985a48e.jpg", + "table_caption": [ + "Table 2. Comparison results with state-of-the-art approaches on VidSTG [42] test set." + ], + "table_footnote": [], + "table_body": "
MethodDeclarative SentencesInterrogative Sentences
m_vIoUvIoU@0.3vIoU@0.5m_vIoUvIoU@0.3vIoU@0.5
STGRN [42]19.825.814.618.321.112.8
STGVT [24]21.629.818.9———
OMRN [41]23.132.616.420.628.414.1
STVGBert [22]24.030.918.422.526.016.0
TubeDETR [30]30.442.528.225.735.723.2
Ours33.747.232.828.539.926.2
", + "bbox": [ + 446, + 128, + 875, + 280 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "to obtain a temporal feature $F_{d} \\in \\mathbb{R}^{T \\times d}$ , and we use an FC layer to adjust its dimension to $d_{m}$ . Then we follow 2D-TAN [39] to construct a 2D-proposal map $M \\in \\mathbb{R}^{T \\times T \\times d_{m}}$ :", + "bbox": [ + 75, + 304, + 468, + 354 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nM _ {i j} = \\left\\{ \\begin{array}{l l} \\mathbf {M P} ([ F _ {d} [ i ], F _ {d} [ i + 1 ], \\dots , F _ {d} [ j ]) & i \\leq j \\\\ \\mathbf {0} & i > j \\end{array} \\right. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 96, + 363, + 468, + 405 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\mathrm{MP}(\\cdot)$ is mean pooling and $M_{ij}$ indicates the feature of temporal moment proposal $C_{ij}$ with $t_i, t_{j+1}$ as start and end time stamps, respectively. Here $t_i = \\frac{i}{T} \\cdot T_{video}$ and $T_{video}$ is the duration of the input video. $F_d[i]$ is the $i$ -th feature in $F_d$ . We employ several convolutional layers to transform the 2D map $M$ into a score map $\\hat{S} \\in \\mathbb{R}^{T \\times T \\times 1}$ , where $\\hat{S}_{ij}$ represents the matching score of temporal proposal $C_{ij}$ . For inference, we take the proposal with the highest score as the prediction of the target temporal time span. And we take the predicted bounding boxes from the spatial prediction head for frames within the predicted temporal span to form the final predicted tube.", + "bbox": [ + 75, + 417, + 472, + 599 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.5. Model Training", + "text_level": 1, + "bbox": [ + 76, + 611, + 235, + 628 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We train our model with loss $L = L_{s} + L_{t}$ , where $L_{s}$ and $L_{t}$ are spatial and temporal localization losses, respectively. Specifically, they are defined as follows:", + "bbox": [ + 76, + 636, + 468, + 681 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nL _ {s} = \\lambda_ {1} L _ {l 1} (\\hat {b}, b) + \\lambda_ {2} L _ {G I o U} (\\hat {b}, b), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 153, + 694, + 468, + 712 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nL _ {t} = \\lambda_ {3} L _ {t g} (\\hat {S}, S) + \\lambda_ {4} L _ {t a}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 156, + 715, + 468, + 734 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $L_{l1}$ and $L_{GIoU}$ are L1 loss and GIoU loss [19] on the predicted bounding boxes, respectively. The temporal grounding loss $L_{tg}$ is defined as a binary cross-entropy loss (as done in [23, 39]) between the predicted score map $\\hat{S}$ and the ground truth map $S$ , where each element in $S$ represents the IoU between the corresponding proposal and the ground truth temporal moment. $L_{ta}$ is a commonly used temporal attentive loss [34] for accelerating convergence of $L_{tg}$ , and it encourages the model to predict a high matching score for those frames/clips inside the target temporal span.", + "bbox": [ + 75, + 747, + 472, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 500, + 304, + 633, + 321 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Experimental Settings", + "text_level": 1, + "bbox": [ + 500, + 330, + 705, + 347 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Datasets. We evaluate our method on HCSTVG-v1 dataset, HCSTVG-v2 dataset [24] and VidSTG [42] dataset. HCSTVG datasets are collected from movie scenes and the duration of each video clip is around 20 seconds. This set is quite challenging for spatio-temporal grounding as some video clips contain many persons conducting similar actions. HCSTVG-v1 dataset consists of 4500 and 1160 video-text pairs for training and testing, respectively. HCSTVG-v2 dataset expanded HCSTVG-v1 dataset and improved the annotation quality. It contains 10131, 2000, 4413 samples for training, validation and testing, respectively. Since the annotations of the test set are not publicly available in HCSTVG-v2, we report results on the validation set. VidSTG dataset [42] is constructed based on VidOR [20] dataset according to the object relation annotations. It contains 99943 video-text pairs, including 44808 declarative sentence queries and 55135 interrogative sentence queries. The training, validation and test set consist of 80684, 8956 and 10303 sentences respectively, and 5436, 602 and 732 videos respectively. However, since VidSTG [42] is annotated based on VidOR [20] dataset, the text queries are limited to describe the pre-defined object/relation categories in VidOR [20].", + "bbox": [ + 496, + 354, + 890, + 700 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation Metrics. We follow previous works [22, 23, 30, 42] to use mean vIoU as the main evaluation metric. vIoU is defined as $\\frac{1}{|T_u|}\\sum_{t\\in T_i}I_oU(\\hat{b}_t,b_t)$ , where $T_{i}$ and $T_{u}$ indicate the intersection and union between the time intervals obtained from ground truth annotation and model prediction, respectively. $\\hat{b}_t,b_t$ are the predicted bounding box and ground truth bounding box for the $t$ -th frame, respectively. We average the vIoU score over all the samples to obtain mean vIoU (termed m_vIoU). We also report vIoU@R which indicates the proportion of samples with vIoU higher than $R$ . In ablation studies, we follow TubeDETR [30] to also report the sIoU and tIoU metrics that evaluate the spatial and temporal grounding performances, respectively.", + "bbox": [ + 496, + 702, + 892, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "23105", + "bbox": [ + 478, + 944, + 517, + 957 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "They are defined as $sIoU = \\frac{1}{|T_{gt}|}\\sum_{t\\in T_i}IoU(\\hat{b_t},b_t)$ and $tIoU = \\frac{|T_i|}{|T_u|}$ , where $T_{gt}$ indicates the set of frames inside the ground truth time interval.", + "bbox": [ + 75, + 89, + 468, + 143 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Implementation Details. We use ResNet101 [9] as our image encoder and Roberta-base [14] as the text encoder to extract image visual features and language features. And we use the Slowfast [4] model pre-trained on AVA [8] as the video encoder. We initialize part of the weights of our static VL stream using pre-trained MDETR [10] as done in [23, 30]. The number of transformer layers in both streams is set as $N = 6$ . The loss weights are set as $\\lambda_1 = 5, \\lambda_2 = 2, \\lambda_3 = 5, \\lambda_4 = 1$ . We train our model on VidSTG [42], HCSTVG-v1, HCSTVG-v2 [24] datasets for 10, 10, 4 epochs, respectively. For most other hyperparameters, we keep them consistent with previous works [10, 30]. Further implementation details can be found in the supplementary material.", + "bbox": [ + 75, + 143, + 470, + 354 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. Experimental Results", + "text_level": 1, + "bbox": [ + 76, + 366, + 277, + 382 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In this section, we compare the performance of our approach with previous methods on HCSTVG dataset [24] and VidSTG [42] dataset.", + "bbox": [ + 75, + 388, + 468, + 434 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "HCSTVG datasets. We first compare our method with state-of-the-arts on HCSTVG datasets [24] and results are shown in Table 1. On HCSTVG-v1 test set, our approach outperforms previous state-of-the-art method TubeDETR [30], by a considerable margin. We achieve $36.9\\%$ m_vIoU, which is $4.5\\%$ higher than TubeDETR. On HCSTVG-v2 validation set [24], we compare our method with TubeDETR [30] and some winners from HCVG challenge2021. As shown in Table 1, we also outperform TubeDETR by $2.3\\%$ m_vIoU. The above results demonstrate that our proposed model with two collaborative vision-language streams can effectively capture static-dynamic cross-modal dependencies for addressing STVG task.", + "bbox": [ + 75, + 435, + 470, + 631 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "VidSTG dataset. We also report results on VidSTG dataset [42]. As listed in Table 2, we also achieve state-of-the-art performance with a m_vIoU of $33.7\\%$ on declarative sentences and $28.5\\%$ on interrogative sentences, respectively. This further shows the effectiveness of the proposed collaborative two-stream model.", + "bbox": [ + 75, + 631, + 468, + 722 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Ablation Study", + "text_level": 1, + "bbox": [ + 76, + 734, + 230, + 750 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In this section, we conduct ablation experiments on HCSTVG-v2 dataset [24] to verify the effectiveness of some important components in the proposed method.", + "bbox": [ + 75, + 757, + 468, + 801 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Quantitative evaluation on the Cross-Stream Collaboration Block. We first verify the effectiveness of static-to-dynamic and dynamic-to-static information transmission blocks by removing one of them from our developed frame", + "bbox": [ + 75, + 803, + 468, + 863 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/b8c9463e349954991a3b15ae4c46409258158c2ef420c094e0cdaa685061e757.jpg", + "table_caption": [ + "Table 3. Evaluation on the cross-stream collaboration mechanism." + ], + "table_footnote": [], + "table_body": "
s-to-dd-to-sm_vIoUvIoU@0.3vIoU@0.5m_tIoUm_sIoU
36.461.229.556.164.9
✓37.563.631.457.464.6
✓37.162.330.256.265.7
✓✓38.765.533.858.165.7
", + "bbox": [ + 526, + 108, + 859, + 184 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "work. We present the evaluation results in Table 3. Compared to the baseline in which the whole cross-stream collaboration block is removed (the first line in Table 3), it is clear that individually adopting static-to-dynamic information transmission (termed \"s-to-d\") or dynamic-to-static information transmission (termed \"d-to-s\") can both improve the grounding performance on vIoU metrics. We also observe that \"s-to-d\" mainly improves the temporal grounding accuracy while \"d-to-s\" improves the spatial grounding performance. And our full model that jointly employs the whole cross-stream collaboration block achieves the best performance of $38.7\\%$ m_vIoU. The above results show that both the static-to-dynamic and the dynamic-to-static information transmission blocks are effective for providing complementary information to the other stream, and thus employing them greatly improves the performance.", + "bbox": [ + 496, + 203, + 890, + 445 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Qualitative evaluation on the Cross-Stream Collaboration Block. To better understand how and why our cross-stream collaboration block significantly improves the grounding accuracy, we further provide 2 qualitative samples in Figure 6. In the first sample, there are three boys coming to the dining table. In this case, our baseline model without cross-stream collaboration failed to predict correct temporal grounding results since it is confused by the similar action performed by the three boys. With our proposed cross-stream collaboration block, the dynamic VL stream can distinguish which boy is the \"tall boy\" by absorbing knowledge from the static VL stream, then it can correctly predict the target temporal time span. For the second sample, there are two women both dressed in a skirt, thus in our baseline where the cross-stream collaboration is removed, the static stream failed to capture and understand the action \"let goes of ... and turns into the room\", thus it cannot distinguish which is the target woman. By employing the proposed cross-stream collaboration block, the motion information learned in the dynamic VL stream can be transmitted to the static stream to make it aware of the motion information which can help distinguish the target person. The above two samples clearly show that the proposed cross-stream collaboration block can effectively exchange complementary information between the static and the dynamic VL streams, which can greatly reduce the uncertainty in the hard and ambiguous cases where some objects have a similar appearance or motion so that the model can produce a more accurate prediction.", + "bbox": [ + 496, + 446, + 892, + 883 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Spatial information in the dynamic VL stream. We", + "bbox": [ + 500, + 885, + 890, + 901 + ], + "page_idx": 6 + }, + { + "type": "page_footnote", + "text": "1In this paper, we report the corrected TubeDETR [30] performance updated by TubeDETR authors at https://github.com/antoyang/TubeDETR.", + "bbox": [ + 75, + 875, + 468, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "23106", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Query: The tall boy comes to the dining table and sits down.", + "bbox": [ + 91, + 88, + 491, + 102 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/174757091560f1f4ae73952479f40a22c5b1c7d72424054e7b0e31655e0c4d6e.jpg", + "image_caption": [ + "Figure 6. Visualization of the predicted tube of our approach (yellow), our approach without cross-stream collaboration (red), and ground truth (green). In the first sample, our approach without cross-stream collaboration predicts a wrong temporal time span. In the second sample, ours w/o the collaboration block predicts wrong spatial bounding boxes. (Best viewed zoomed in on screen.)" + ], + "image_footnote": [], + "bbox": [ + 91, + 103, + 883, + 282 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/7b1ffdebe30f26865890cd9dd860507537fe9c08280fad732676b6d32d758b95.jpg", + "table_caption": [ + "Table 4. Evaluation on designs of the dynamic VL stream." + ], + "table_footnote": [], + "table_body": "
Method|Ours|Ours w/o spatial|Ours w/o mean pool
m_tIoU|56.1|54.7|56.2
", + "bbox": [ + 129, + 377, + 411, + 416 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/e557b6538dfae8b1cc4017d550b05f1455d36cdf315c7b0fe4b222324896fa82.jpg", + "table_caption": [ + "Table 5. Comparison between different cross-stream collaboration block designs." + ], + "table_footnote": [], + "table_body": "
Methodm_vIoUvIoU@0.3vIoU@0.5m_tIoUm_sIoU
Ours38.765.533.858.165.7
None36.461.229.556.164.9
sum36.662.530.055.664.4
FC + sum36.862.830.557.363.4
concat + FC37.163.230.356.963.9
", + "bbox": [ + 106, + 457, + 434, + 545 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "present ablation studies on different design choices of the dynamic VL stream and results are presented in Table 4. Here, we discard the static stream to eliminate its influence and we report temporal grounding accuracy using m_tIoU metric. By mean pooling at the spatial dimension of the video feature inputted to the dynamic VL stream (termed \"Ours w/o spatial\"), it results in a drop of more than $1\\%$ m_tIoU which shows the importance of learning spatial visual interaction for modeling dynamic cues. And since we develop the dynamic VL stream to mainly learn dynamic cross-modal correspondence, we employ spatial mean pooling on $\\mathcal{H}_v$ to obtain $\\overline{\\mathcal{H}}_v$ so that the text-to-visual cross-attention is performed only on the temporal dimension which can greatly reduce computation cost. By removing this mean pooling operation (termed \"Ours w/o mean pool\"), the text-to-visual cross-attention is performed among all spatio-temporal tokens, but it has very limited improvement on the performance, which validates the rationality of our design.", + "bbox": [ + 75, + 566, + 468, + 852 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Further study of the collaboration block. To further verify the effectiveness of the proposed collaboration block, we replace it with some symmetric operations, including", + "bbox": [ + 75, + 854, + 468, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "\"sum\", \"FC + sum\" and \"concat + FC\", which are some operations commonly used in the community [4, 21]. As shown in Table 5, \"concat + FC\" achieves the best performance among these baselines, but it only achieves a limited improvement of $0.7\\%$ m_vIoU, which indicates that simply merging the information from the two streams can not fully exploit their complementary abilities. In contrast, \"Ours\" outperforms \"concat + FC\" by a considerable margin of more than $3\\%$ in terms of vIoU@0.5, demonstrating that our asymmetric cross-stream collaboration block enables the two streams to effectively exchange complementary information and to collaboratively reason the target object.", + "bbox": [ + 496, + 359, + 890, + 542 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 561, + 617, + 577 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this work, we propose a novel framework for Spatio-Temporal Video Grounding task. Our framework mainly consists of a static VL(vision-language) stream which captures static visual cues like object appearance, and a dynamic VL stream which understands dynamic visual cues like action. We elaborately design the framework as a collaborative system in which the two streams collaborative localize the target object via a novel cross-stream collaboration block, which is proven to be effective. We also present visualization results which provide some interesting insights and show how the collaboration of the two streams helps improve the prediction. Our approach significantly outperforms previous methods on VidSTG [42] and HC-STVG dataset [24], which demonstrates its effectiveness.", + "bbox": [ + 496, + 590, + 890, + 800 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements. This work was supported partially by the NSFC (U21A20471, U22A2095, 62076260, 61772570), Guangdong Natural Science Funds Project (2020B1515120085, 2023B1515040025), Guangdong NSF for Distinguished Young Scholar (2022B1515020009, 2018B030306025), and the Key-Area Research and Development Program of Guangzhou (202007030004). We thank Bing Shuai for the helpful discussions.", + "bbox": [ + 496, + 801, + 890, + 898 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "23107", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Lisa Anne Hendricks, Oliver Wang, Eli Shechtman, Josef Sivic, Trevor Darrell, and Bryan Russell. Localizing moments in video with natural language. In Proceedings of the IEEE international conference on computer vision, pages 5803-5812, 2017. 2", + "[2] Gedas Bertasius, Heng Wang, and Lorenzo Torresani. Is space-time attention all you need for video understanding? In Proceedings of the International Conference on Machine Learning (ICML), July 2021. 4", + "[3] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In European conference on computer vision, pages 213-229. Springer, 2020. 4", + "[4] Christoph Feichtenhofer, Haoqi Fan, Jitendra Malik, and Kaiming He. Slowfast networks for video recognition. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6202-6211, 2019. 1, 2, 3, 4, 7, 8", + "[5] Christoph Feichtenhofer, Axel Pinz, and Richard P Wildes. Spatiotemporal multiplier networks for video action recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4768-4777, 2017. 3", + "[6] Jiyang Gao, Chen Sun, Zhenheng Yang, and Ram Nevatia. Tall: Temporal activity localization via language query. In Proceedings of the IEEE international conference on computer vision, pages 5267-5275, 2017. 2", + "[7] Yash Goyal, Tejas Khot, Douglas Summers-Stay, Dhruv Bartra, and Devi Parikh. Making the v in vqa matter: Elevating the role of image understanding in visual question answering. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6904-6913, 2017. 1", + "[8] Chunhui Gu, Chen Sun, David A Ross, Carl Vondrick, Caroline Pantofaru, Yeqing Li, Sudheendra Vijayanarasimhan, George Toderici, Susanna Ricco, Rahul Sukthankar, et al. Ava: A video dataset of spatio-temporally localized atomic visual actions. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 6047-6056, 2018. 7", + "[9] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 4, 7", + "[10] Aishwarya Kamath, Mannat Singh, Yann LeCun, Gabriel Synnaeve, Ishan Misra, and Nicolas Carion. Mdetr-modulated detection for end-to-end multi-modal understanding. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1780-1790, 2021. 3, 7", + "[11] Kun Li, Dan Guo, and Meng Wang. Proposal-free video grounding with contextual pyramid network. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 35, pages 1902-1910, 2021. 2, 3", + "[12] Liunian Harold Li*, Pengchuan Zhang*, Haotian Zhang*, Jianwei Yang, Chunyuan Li, Yiwu Zhong, Lijuan Wang, Lu Yuan, Lei Zhang, Jenq-Neng Hwang, Kai-Wei Chang, and Jianfeng Gao. Grounded language-image pre-training. In CVPR, 2022. 3" + ], + "bbox": [ + 78, + 114, + 470, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[13] Yue Liao, Si Liu, Guanbin Li, Fei Wang, Yanjie Chen, Chen Qian, and Bo Li. A real-time cross-modality correlation filtering method for referring expression comprehension. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10880-10889, 2020. 1", + "[14] Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. Roberta: A robustly optimized bert pretraining approach. arXiv preprint arXiv:1907.11692, 2019. 4, 7", + "[15] Jiasen Lu, Dhruv Batra, Devi Parikh, and Stefan Lee. Vilbert: Pretraining task-agnostic visiolinguistic representations for vision-and-language tasks. Advances in neural information processing systems, 32, 2019. 1", + "[16] Jonghwan Mun, Minsu Cho, and Bohyung Han. Local-global video-text interactions for temporal grounding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10810-10819, 2020. 2, 3", + "[17] Guoshun Nan, Rui Qiao, Yao Xiao, Jun Liu, Sicong Leng, Hao Zhang, and Wei Lu. Interventional video grounding with dual contrastive learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2765-2775, 2021. 2", + "[18] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. Advances in neural information processing systems, 28, 2015. 2", + "[19] Hamid Rezatofighi, Nathan Tsoi, JunYoung Gwak, Amir Sadeghian, Ian Reid, and Silvio Savarese. Generalized intersection over union: A metric and a loss for bounding box regression. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 658-666, 2019. 6", + "[20] Xindi Shang, Donglin Di, Junbin Xiao, Yu Cao, Xun Yang, and Tat-Seng Chua. Annotating objects and relations in user-generated videos. In Proceedings of the 2019 on International Conference on Multimedia Retrieval, pages 279-287. ACM, 2019. 6", + "[21] Karen Simonyan and Andrew Zisserman. Two-stream convolutional networks for action recognition in videos. Advances in neural information processing systems, 27, 2014. 1, 2, 3, 8", + "[22] Rui Su, Qian Yu, and Dong Xu. Stvgbert: A visual-linguistic transformer based framework for spatio-temporal video grounding. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 1533–1542, October 2021. 1, 2, 6", + "[23] Chaolei Tan, Zihang Lin, Jian-Fang Hu, Xiang Li, and Wei-Shi Zheng. Augmented 2d-tan: A two-stage approach for human-centric spatio-temporal video grounding. arXiv preprint arXiv:2106.10634, 2021. 3, 6, 7", + "[24] Zongheng Tang, Yue Liao, Si Liu, Guanbin Li, Xiaojie Jin, Hongxu Jiang, Qian Yu, and Dong Xu. Human-centric spatio-temporal video grounding with visual transformers. IEEE Transactions on Circuits and Systems for Video Technology, pages 1-1, 2021. 1, 2, 6, 7, 8" + ], + "bbox": [ + 501, + 92, + 890, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "23108", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[25] Limin Wang, Yuanjun Xiong, Zhe Wang, and Yu Qiao. Towards good practices for very deep two-stream convnets. arXiv preprint arXiv:1507.02159, 2015. 3", + "[26] Xuanhan Wang, Lianli Gao, Peng Wang, Xiaoshuai Sun, and Xianglong Liu. Two-stream 3-d convnet fusion for action recognition in videos with arbitrary size and length. IEEE Transactions on Multimedia, 20(3):634-644, 2017. 3", + "[27] Zhenzhi Wang, Limin Wang, Tao Wu, Tianhao Li, and Gangshan Wu. Negative sample matters: A renaissance of metric learning for temporal grounding. CoRR, abs/2109.04872, 2021. 6", + "[28] Zuxuan Wu, Yu-Gang Jiang, Xi Wang, Hao Ye, and Xi-angyang Xue. Multi-stream multi-class fusion of deep networks for video classification. In Proceedings of the 24th ACM international conference on Multimedia, pages 791–800, 2016. 3", + "[29] Shaoning Xiao, Long Chen, Songyang Zhang, Wei Ji, Jian Shao, Lu Ye, and Jun Xiao. Boundary proposal network for two-stage natural language video localization. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 35, pages 2986-2994, 2021. 2", + "[30] Antoine Yang, Antoine Miech, Josef Sivic, Ivan Laptev, and Cordelia Schmid. Tubedetr: Spatio-temporal video grounding with transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 1, 2, 3, 6, 7", + "[31] Kexin Yi, Jiajun Wu, Chuang Gan, Antonio Torralba, Pushmeet Kohli, and Josh Tenenbaum. Neural-symbolic vqa: Disentangling reasoning from vision and language understanding. Advances in neural information processing systems, 31, 2018. 1", + "[32] Licheng Yu, Zhe Lin, Xiaohui Shen, Jimei Yang, Xin Lu, Mohit Bansal, and Tamara L Berg. Mattnet: Modular attention network for referring expression comprehension. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 1307-1315, 2018. 1, 3", + "[33] Yi Yu, Xinying Wang, Wei Hu, Xun Luo, and Cheng Li. 2rd place solutions in the hc-stvg track of person in context challenge 2021. arXiv preprint arXiv:2106.07166, 2021. 6", + "[34] Yitian Yuan, Tao Mei, and Wenwu Zhu. To find where you talk: Temporal sentence localization in video with attention based location regression. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 33, pages 9159-9166, 2019. 6", + "[35] Runhao Zeng, Haoming Xu, Wenbing Huang, Peihao Chen, Mingkui Tan, and Chuang Gan. Dense regression network for video grounding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10287-10296, 2020. 2, 3", + "[36] Da Zhang, Xiyang Dai, Xin Wang, Yuan-Fang Wang, and Larry S Davis. Man: Moment alignment network for natural language moment retrieval via iterative graph adjustment. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1247-1257, 2019. 2", + "[37] Hanwang Zhang, Yulei Niu, and Shih-Fu Chang. Grounding referring expressions in images by variational context. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 4158-4166, 2018. 3" + ], + "bbox": [ + 78, + 90, + 467, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[38] Hao Zhang, Aixin Sun, Wei Jing, Liangli Zhen, Joey Tianyi Zhou, and Rick Siow Mong Goh. Natural language video localization: A revisit in span-based question answering framework. IEEE transactions on pattern analysis and machine intelligence, 2021. 2, 3", + "[39] Songyang Zhang, Houwen Peng, Jianlong Fu, and Jiebo Luo. Learning 2d temporal adjacent networks for moment localization with natural language. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 34, pages 12870-12877, 2020. 2, 3, 6", + "[40] Zhiwang Zhang, Dong Xu, Wanli Ouyang, and Chuanqi Tan. Show, tell and summarize: Dense video captioning using visual cue aided sentence summarization. IEEE Transactions on Circuits and Systems for Video Technology, 30(9):3130-3139, 2019. 1", + "[41] Zhu Zhang, Zhou Zhao, Zhijie Lin, Baoxing Huai, and Jing Yuan. Object-aware multi-branch relation networks for spatio-temporal video grounding. In Proceedings of the Twenty-Ninth International Joint Conference on Artificial Intelligence, IJCAI'20, 2021. 2, 6", + "[42] Zhu Zhang, Zhou Zhao, Yang Zhao, Qi Wang, Huasheng Liu, and Lianli Gao. Where does it exist: Spatio-temporal video grounding for multi-form sentences. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2020. 1, 2, 6, 7, 8" + ], + "bbox": [ + 501, + 90, + 890, + 441 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "23109", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2023/Collaborative Static and Dynamic Vision-Language Streams for Spatio-Temporal Video Grounding/40d01bed-f6c0-4114-9925-d43e564f9389_model.json b/2023/Collaborative Static and Dynamic Vision-Language Streams for Spatio-Temporal Video Grounding/40d01bed-f6c0-4114-9925-d43e564f9389_model.json new file mode 100644 index 0000000000000000000000000000000000000000..57c11725df9ec5a58562f6a919f2532bb7026dd5 --- /dev/null +++ b/2023/Collaborative Static and Dynamic Vision-Language Streams for Spatio-Temporal Video Grounding/40d01bed-f6c0-4114-9925-d43e564f9389_model.json @@ -0,0 +1,1903 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.182, + 0.131, + 0.79, + 0.177 + ], + "angle": 0, + "content": "Collaborative Static and Dynamic Vision-Language Streams for Spatio-Temporal Video Grounding" + }, + { + "type": "text", + "bbox": [ + 0.143, + 0.203, + 0.826, + 0.238 + ], + "angle": 0, + "content": "Zihang Lin1, Chaolei Tan1, Jian-Fang Hu1,3,4*, Zhi Jin1, Tiancai Ye2, Wei-Shi Zheng1,3,4, Sun Yat-sen University, China" + }, + { + "type": "text", + "bbox": [ + 0.094, + 0.24, + 0.875, + 0.275 + ], + "angle": 0, + "content": "3Guangdong Province Key Laboratory of Information Security Technology, China \n4 Key Laboratory of Machine Intelligence and Advanced Computing, Ministry of Education, China" + }, + { + "type": "text", + "bbox": [ + 0.165, + 0.277, + 0.804, + 0.31 + ], + "angle": 0, + "content": "{1inzh59, tanchlei}@mail2.sysu.edu.cn, {hujf5, jinzh26}@mail.sysu.edu.cn, tiancaiye@tencent.com, wszheng@ieee.org" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.345, + 0.314, + 0.361 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.378, + 0.474, + 0.74 + ], + "angle": 0, + "content": "Spatio-Temporal Video Grounding (STVG) aims to localize the target object spatially and temporally according to the given language query. It is a challenging task in which the model should well understand dynamic visual cues (e.g., motions) and static visual cues (e.g., object appearances) in the language description, which requires effective joint modeling of spatio-temporal visual-linguistic dependencies. In this work, we propose a novel framework in which a static vision-language stream and a dynamic vision-language stream are developed to collaboratively reason the target tube. The static stream performs cross-modal understanding in a single frame and learns to attend to the target object spatially according to intra-frame visual cues like object appearances. The dynamic stream models visual-linguistic dependencies across multiple consecutive frames to capture dynamic cues like motions. We further design a novel cross-stream collaborative block between the two streams, which enables the static and dynamic streams to transfer useful and complementary information from each other to achieve collaborative reasoning. Experimental results show the effectiveness of the collaboration of the two streams and our overall framework achieves new state-of-the-art performance on both HCSTVG and VidSTG datasets." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.772, + 0.21, + 0.789 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.798, + 0.47, + 0.875 + ], + "angle": 0, + "content": "Vision-language cross-modal understanding is a challenging yet important research problem that bridges the communication between humans and artificial intelligence systems. It has attracted increasing attention and many vision-language tasks were studied in recent years, like vi" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.346, + 0.895, + 0.497 + ], + "angle": 0, + "content": "sual grounding [13, 32], VQA [7, 31], image/video captioning [15, 40], etc. In this work, we focus on a challenging vision-language task named Spatio-Temporal Video Grounding (STVG) which was recently proposed in [42]. Given a language query indicating an object (as shown in Figure 1), STVG aims to localize the target object spatially and temporally in the video. In this task, the input language query may express different kinds of visual concepts, thus the model requires to well capture and understand these concepts in both vision and language modalities." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.503, + 0.895, + 0.716 + ], + "angle": 0, + "content": "In STVG task, dynamic visual concepts like human motion and static visual concepts like object appearance are both important for distinguishing the target object from other objects that occurred in the same video. For example, in the first sample in Figure 1, the two men are dressed alike (i.e., their static appearance cues are similar), and we can only distinguish them by motion. In the second sample, the two women perform similar actions, they both stand up (i.e., they have similar dynamic motion cues), here, we can only distinguish them by their clothes. The above examples show that static or dynamic visual cues alone cannot solve the STVG task well. And it implies that modeling static and dynamic visual-linguistic dependencies and collaboratively utilizing them are important for addressing STVG task." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.896, + 0.902 + ], + "angle": 0, + "content": "Humans treat static and dynamic cues differently [4, 21], but this was overlooked in previous STVG works [22, 24, 30]. Taking the first query in Figure 1 as an example, a human would randomly or evenly click on some locations on the video's progress bar to find candidate frames containing a man in blue coat. And he will play the video around that frame and attend on the candidate man to check whether he performs the action described in the text (i.e., \"turns around and stops by the stone\"). In the above process, the human understands static and dynamic cues in different ways (i.e., view a single frame and watch the video clip, respectively) and determines the target object by jointly considering the" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.887, + 0.222, + 0.9 + ], + "angle": 0, + "content": "*Corresponding author." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "23100" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.135, + 0.094, + 0.59, + 0.108 + ], + "angle": 0, + "content": "Query: The man in the blue coat turns around and stops by the stone." + }, + { + "type": "image", + "bbox": [ + 0.132, + 0.113, + 0.835, + 0.18 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.135, + 0.183, + 0.479, + 0.197 + ], + "angle": 0, + "content": "Query: The woman in light gray stands up and turns." + }, + { + "type": "image", + "bbox": [ + 0.132, + 0.202, + 0.835, + 0.269 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.275, + 0.895, + 0.318 + ], + "angle": 0, + "content": "Figure 1. Two examples for Spatio-Temporal Video Grounding task. In the first case, the two men are both dressed alike in blue, thus understanding the action described in the query sentence is essential to recognize the person of interest. In the second case, both of the two women stand up, we can only distinguish them by their clothes. (Best viewed zoomed in on screen.)" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.332, + 0.47, + 0.468 + ], + "angle": 0, + "content": "static and dynamic cues in an \"attend-and-check\" process. Inspired by the above observations, we propose a framework that consists of a static and a dynamic vision-language (VL) streams to model static and dynamic cues, respectively. And we design a novel cross-stream collaboration block between the two streams to simulate the \"attend-and-check\" process, which exchanges useful and complementary information learned in each stream and enables the two streams to collaboratively reason the target object." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.473, + 0.471, + 0.806 + ], + "angle": 0, + "content": "Specifically, in this work, our static vision-language (VL) stream learns to attend to some candidate regions according to static visual cues like appearance, while the dynamic VL stream learns to understand the dynamic visual cues like action described in the text query. Then, in the collaboration block, we guide the dynamic stream to only focus and attend on the motion of the candidate objects by using the learned attended region in the static stream. And we transfer the text-motion matching information learned in the dynamic stream to the static stream, to help it further check and determine the target object and predict a more consistent tube. With the above cross-stream collaboration blocks, both the static and dynamic vision-language streams can learn reciprocal information from the other stream, which is effective for achieving more accurate spatio-temporal grounding predictions. We conduct experiments on HCSTVG [24] and VidSTG [42] datasets and our approach outperforms previous approaches by a considerable margin. Ablation studies demonstrate the effectiveness of each component in our proposed cross-stream collaboration block and show its superiority over commonly used counterparts in video understanding works [4, 21]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.811, + 0.471, + 0.902 + ], + "angle": 0, + "content": "In summary, our contributions are: 1), we develop an effective framework that contains two parallel streams to model static-dynamic visual-linguistic dependencies for complete cross-modal understanding; 2), we propose a novel cross-stream collaboration block between the two streams to exchange reciprocal information for each other" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.332, + 0.892, + 0.379 + ], + "angle": 0, + "content": "and enable collaborative reasoning of the target object; 3). Our overall framework achieves state-of-the-art performances on HCSTVG [24] and VidSTG datasets [42]." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.392, + 0.642, + 0.407 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.418, + 0.892, + 0.809 + ], + "angle": 0, + "content": "Spatio-Temporal Video Grounding. Spatio-temporal video grounding aims to localize the target object spatially and temporally according to a language query. Early works [24, 41, 42] employ pre-trained detectors like Faster R-CNN [18] to detect objects in each frame, and build their model upon the detection features of each object. STGRN [42] and OMRN [41] learn object relations with spatio-temporal graphs and multi-branch relation networks, respectively. STGVT [24] learns cross-modal representations with visual transformers for video-sentence matching and temporal localization. However, these methods are limited to the pre-detected bounding boxes and cannot localize object categories that are not defined in the pretraining dataset. Recent works [22, 30] design their framework based on strong pre-trained vision-language models and train their models end-to-end. These approaches do not require pre-detected bounding boxes and they perform much better. However, they directly collapse some dimensions (i.e., spatial or temporal) in cross-modal attention calculation to reach an acceptable computation cost, which results in large information loss, and thus they can not well capture the important static-dynamic cross-modal contextual relationship. To address this weakness, we propose to model static and dynamic visual-linguistic correspondence with parallel streams and learn to reason the target object in a collaborative framework." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.811, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Temporal Grounding. Given a language query, temporal grounding aims to predict the time span in a video specified by the language query [1, 6]. Existing works [11, 16, 17, 29, 35, 36, 38, 39] can be divided into proposal-based and proposal-free approaches. Proposal-based approaches generate some proposals (manually designed [29, 39] or" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "23101" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.107, + 0.087, + 0.868, + 0.296 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.3, + 0.895, + 0.357 + ], + "angle": 0, + "content": "Figure 2. An overview of the proposed framework. Our framework mainly consists of a static VL( vision-language) stream and a dynamic VL stream. The static stream learns to attend to the spatial locations of the target object according to static cues like object appearance. The dynamic stream learns motion-text correspondence according to dynamic cues like human action. We further devise a cross-stream collaboration block that enables the two streams to query useful and complementary information from the other stream." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.371, + 0.47, + 0.537 + ], + "angle": 0, + "content": "predicted by the model [35]) and predict the IoU scores between each proposal and the ground truth. Proposal-free approaches address moment localization by directly regressing the starting/ending points of the target time spans [11, 16] or predicting the probability of being a starting/ending point for each clip and choosing the time span with the highest joint probability as the prediction [38]. In this work, we follow [23, 39] to employ a proposal-based grounding head. And in contrast to most temporal grounding works that collapse the spatial dimension, we also incorporate spatial cues to enhance temporal localization." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.538, + 0.47, + 0.763 + ], + "angle": 0, + "content": "Text-conditioned Object Detection. Given an image, text-conditioned object detection aims to localize the target object indicated by a sentence (also named referring image expression task) or all objects described in a sentence (also named phrase grounding task). Most existing works [32,37] develop their models based on off-the-shelf detection models and match the detected objects with the sentence query. Recently, MDETR [10] and GLIP [12] unified detection and phrase grounding tasks, and they trained their models with large-scale data. These models can achieve excellent performance on text-conditioned object detection and are used to initialize model weights for STVG task [23, 30]. However, these models lack the ability of modeling dynamic cues. In this work, we design a novel collaboration framework to capture both static and dynamic cues for STVG task." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.47, + 0.903 + ], + "angle": 0, + "content": "Two-stream Models. Considering that static and dynamic visual cues are quite different but complementary in video understanding, a number of two-stream models [4,5,21,25, 26, 28] are proposed to capture these two kinds of visual cues. For example, Two-Stream ConvNet [21] introduces an optical flow stream to capture object motion. SlowFast [4] develops slow and fast pathways with different temporal resolutions to capture static and dynamic visual cues. However, these works typically fuse different streams equally" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.371, + 0.892, + 0.448 + ], + "angle": 0, + "content": "via some simple symmetric operations, e.g., assembling the output, sum/concat/FC/temporal-convolution on the features. In contrast, we propose a novel asymmetrical cross-stream collaboration block to exploit better collaboration between the two streams considering their different abilities." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.466, + 0.879, + 0.484 + ], + "angle": 0, + "content": "3. Collaborative Static-Dynamic VL Streams" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.494, + 0.893, + 0.901 + ], + "angle": 0, + "content": "We illustrate our framework in Figure 2. Our framework mainly consists of a static and a dynamic VL (vision-language) stream to model static and dynamic visual-linguistic dependencies for complete cross-modal understanding. The static stream performs cross-modal understanding for static contextual information, i.e., finding the object that matches the query text in a still frame according to static visual cues, like object appearance which is important for achieving accurate spatial grounding. The dynamic stream performs cross-modal understanding for dynamic contextual information, i.e., finding the temporal moment that best matches the query sentence according to dynamic visual cues, like action which is important for achieving accurate temporal grounding. Both the static and dynamic streams learn cross-modal correspondence with a transformer-like architecture that contains multiple layers. To enable information transmission between the two streams, we further design a novel cross-stream collaboration block which is placed after each transformer layer in the two streams. In this way, both the static and dynamic VL streams can absorb complementary information from each other and achieve collaborative inferring, which can greatly reduce the uncertainty in ambiguous and hard cases where different objects have similar appearances or perform similar actions. Finally, we employ the prediction head to predict a spatio-temporal tube. In the following, we will introduce each component in detail." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "23102" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.127, + 0.092, + 0.362, + 0.246 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.241, + 0.248, + 0.257, + 0.26 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.387, + 0.092, + 0.838, + 0.248 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.613, + 0.247, + 0.629, + 0.258 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.263, + 0.895, + 0.293 + ], + "angle": 0, + "content": "Figure 3. The detailed architectures of the proposed framework. (a): The cross-modal transformer in the static VL stream. (b): The architecture of each space-time cross-modal transformer layer in the dynamic VL stream." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.307, + 0.248, + 0.323 + ], + "angle": 0, + "content": "3.1. Static VL Stream" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.332, + 0.474, + 0.696 + ], + "angle": 0, + "content": "The static VL stream is employed to perform cross-modal static context understanding and it processes each frame independently. Inspired by DETR [3], we design our static VL stream as a stack of \\(N\\) cross-modal transformer encoder layers and \\(N\\) cross-modal transformer decoder layers, as presented in Figure 2 and 3 (a). For the \\(t\\)-th frame, the inputs of our cross-modal transformer encoder are the concatenation of \\(\\mathbb{R}^{HW\\times d}\\)-sized visual features and \\(\\mathbb{R}^{L\\times d}\\)-sized language features, where \\(H,W\\) are the resolution of the visual features extracted from the static image frame, and \\(L\\) is the number of text tokens in the input text query. These visual and language features are obtained by a pretrained image encoder [9] and language encoder [14], respectively. We employ an FC layer behind each feature extractor to project the visual and language features to a shared embedding space. The outputs of the last cross-modal transformer encoder layer form a \\(\\mathbb{R}^{(HW + L)\\times d}\\)-sized cross-modal memory \\(\\mathcal{M}_t\\), which captures rich interactions between intra-frame static visual cues and linguistic descriptions. Then in the decoder, a learnable object query vector \\(O_{t}^{0}\\in \\mathbb{R}^{d}\\) is inputted to repeatedly query object appearance and location information from the memory \\(\\mathcal{M}_t\\) via a cross-attention layer. Here, the object query gradually learns to attend to the object that matched the text query." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.71, + 0.273, + 0.726 + ], + "angle": 0, + "content": "3.2. Dynamic VL Stream" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.735, + 0.471, + 0.902 + ], + "angle": 0, + "content": "The dynamic VL stream performs cross-modal understanding for dynamic contextual information. As illustrated in Figure 2, we first extract clip-level visual features \\( F_{clip} \\in \\mathbb{R}^{T \\times H \\times W \\times c} \\) from \\( T \\) uniformly sampled video clips by a pre-trained 3D-CNN [4]. Then we employ an FC layer to project the channel dimension from \\( c \\) to \\( d \\) and obtain \\( F_v^0 \\). \\( F_v^0 \\) is fed into a Space-Time Cross-Modal Transformer (STCMT) which consists of \\( N \\) layers to model the visual-linguistic dependencies from a dynamic perspective. The detailed architecture of each layer in STCMT is illustrated in Figure 3 (b). In each layer, we first per" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.308, + 0.895, + 0.43 + ], + "angle": 0, + "content": "form intra-modality self-attention for the dynamic visual features and linguistic features. For visual features, in order to reduce computation cost, we follow TimesFormer [2] to split the spatio-temporal attention into separate attentions. Denoting the visual features after self-attention as \\(\\mathcal{H}_v\\in \\mathbb{R}^{T\\times H\\times W\\times d}\\) and the linguistic features after self-attention as \\(\\mathcal{H}_l\\in \\mathbb{R}^{L\\times d}\\). We then perform cross-attention between \\(\\mathcal{H}_v\\) and \\(\\mathcal{H}_l\\) as:" + }, + { + "type": "equation", + "bbox": [ + 0.516, + 0.435, + 0.892, + 0.556 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} Q _ {v} ^ {(h, w)} = W _ {v} ^ {q} \\mathcal {H} _ {v} ^ {(h, w)}, K _ {v} = W _ {v} ^ {k} \\overline {{\\mathcal {H}}} _ {v}, V _ {v} = W _ {v} ^ {v} \\overline {{\\mathcal {H}}} _ {v}, \\\\ Q _ {l} = W _ {l} ^ {v} \\mathcal {H} _ {l}, K _ {l} = W _ {l} ^ {k} \\mathcal {H} _ {l}, V _ {l} = W _ {l} ^ {v} \\mathcal {H} _ {l}, \\\\ \\widetilde {\\mathcal {H}} _ {v} ^ {(h, w)} = \\mathcal {H} _ {v} ^ {(h, w)} + \\text {A t t e n t i o n} \\left(Q _ {v} ^ {(h, w)}, K _ {l}, V _ {l}\\right), \\tag {1} \\\\ \\widetilde {\\mathcal {H}} _ {l} = \\mathcal {H} _ {l} + \\operatorname {A t t e n t i o n} \\left(Q _ {l}, K _ {v}, V _ {v}\\right), \\\\ \\operatorname {A t t e n t i o n} (Q, K, V) = \\operatorname {s o f t m a x} \\left(\\frac {Q K ^ {T}}{\\sqrt {d _ {k}}}\\right) V, \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.563, + 0.893, + 0.84 + ], + "angle": 0, + "content": "where \\( W_{v}^{q}, W_{v}^{k}, W_{v}^{v}, W_{l}^{q}, W_{l}^{k}, W_{l}^{v} \\) are learnable weighted matrices for computing queries, keys and values for the attention mechanism. \\( \\overline{\\mathcal{H}}_v \\in \\mathbb{R}^{T \\times d} \\) is obtained by conducting mean pooling on \\( \\mathcal{H}_v \\) along the spatial dimensions. \\( d_{k} \\) is the dimension of the queries and keys. \\( \\mathcal{H}_{v}^{(h,w)} \\in \\mathbb{R}^{T \\times d} \\) indicates the visual feature in \\( \\mathcal{H}_v \\) at spatial position \\( (h,w) \\). \\( \\widetilde{\\mathcal{H}}_v, \\widetilde{\\mathcal{H}}_l \\) are the output visual and linguistic features after cross-attention, respectively. The cross-attention operation is designed to explore rich interactions between the visual and linguistic features, which enables the model to learn a powerful cross-modal representation of the depicted dynamic cues. This is essential for grounding temporal moments. After computing the cross-attention between the visual features and linguistic features, we finally employ Feed-Forward Network (FFN) to process both features. In our dynamic VL stream, rich context is learned from the two modalities, and the visual dynamic features and the linguistic features are fused well." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.848, + 0.755, + 0.862 + ], + "angle": 0, + "content": "3.3. Cross-Stream Collaboration" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.871, + 0.894, + 0.902 + ], + "angle": 0, + "content": "To enable the static and the dynamic VL streams to learn complementary information from each other and to perform" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "23103" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.091, + 0.087, + 0.468, + 0.214 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.09, + 0.214, + 0.468, + 0.322 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.09, + 0.323, + 0.466, + 0.45 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.461, + 0.456, + 0.475 + ], + "angle": 0, + "content": "Figure 4. The architectures of our cross-stream collaboration." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.49, + 0.47, + 0.633 + ], + "angle": 0, + "content": "collaborative reasoning, we design a novel cross-stream collaboration block between the two streams. As shown in Figure 4, this block is placed after each decoder layer \\(\\{\\mathcal{B}_s^i\\}_{i=1}^N\\) in static stream and each layer \\(\\{\\mathcal{B}_a^i\\}_{i=1}^N\\) in dynamic stream. It consists of a static-to-dynamic and a dynamic-to-static information transmission block. They are designed asymmetrically to exploit better collaboration between the two streams considering their different abilities. In the following, we will introduce the detailed designs." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.633, + 0.47, + 0.724 + ], + "angle": 0, + "content": "The static-to-dynamic information transmission block is designed to guide the dynamic VL stream to attend to the spatial region that is highly related to the objects depicted in the query text, by utilizing the cross-attention weights learned in the decoder layers of the static VL stream. Concretely, it is formulated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.141, + 0.732, + 0.47, + 0.753 + ], + "angle": 0, + "content": "\\[\n\\widetilde {F} _ {v} ^ {i} = \\operatorname {L a y e r N o r m} \\left(F _ {v} ^ {i} + A ^ {i} \\odot \\mathrm {F C} \\left(F _ {v} ^ {i}\\right)\\right), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.759, + 0.471, + 0.902 + ], + "angle": 0, + "content": "where \\(F_{v}^{i}\\in \\mathbb{R}^{T\\times HW\\times d}\\) is the output visual feature of the \\(i\\)-th layer in dynamic VL stream and \\(A^i\\in \\mathbb{R}^{T\\times HW\\times d}\\) is the cross-attention weights (replicate \\(d\\) times to have \\(d\\) channels at the last dimension) calculated between the object queries \\(\\left\\{O_t^i\\right\\}_{t = 1}^T\\) and corresponding encoded memories \\(\\{M_t\\}_{t = 1}^T\\) in \\(\\mathcal{B}_s^i\\). \\(\\odot\\) indicates hadamard product. The intuition behind this design is that the cross-attention weights learned in the static VL stream can attend to regions matching the language description, and this can serve as a strong guidance" + }, + { + "type": "text", + "bbox": [ + 0.505, + 0.092, + 0.676, + 0.115 + ], + "angle": 0, + "content": "Query: The tall boy comes to the dining table and sits down." + }, + { + "type": "image", + "bbox": [ + 0.506, + 0.118, + 0.698, + 0.216 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.699, + 0.092, + 0.871, + 0.115 + ], + "angle": 0, + "content": "Query: The man in the red stripe goes to the sofa and turns." + }, + { + "type": "image", + "bbox": [ + 0.7, + 0.117, + 0.794, + 0.166 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.795, + 0.118, + 0.888, + 0.166 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.701, + 0.167, + 0.794, + 0.216 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.795, + 0.167, + 0.888, + 0.216 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.54, + 0.228, + 0.851, + 0.242 + ], + "angle": 0, + "content": "Figure 5. Visualization of the learned attention map." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.259, + 0.892, + 0.426 + ], + "angle": 0, + "content": "to help the dynamic VL stream focus more on the dynamic variations around the object-related regions. We visualize the attention maps in Figure 5. In the first sample, there are multiple boys and the attention map has a high score at the best-matched person (i.e., the tallest boy in each frame). In the second sample, we observe that when the man in red stripe is absent, the attention weights are relatively smooth. However, the attention weights become sharp once the target man appeared. These samples intuitively show that our static-to-dynamic information transmission block can guide the dynamic stream to focus on motions of the target object." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.427, + 0.893, + 0.699 + ], + "angle": 0, + "content": "The dynamic-to-static information transmission block is designed to transfer the learned motion-text correspondence to the static VL stream. It enhances \\( O_{t}^{i} \\) to \\( \\widetilde{O}_{t}^{i} \\) with a cross-attention and a query mixing operation as depicted in Figure 4(c). In each block, the object query \\( O_{t}^{i} \\) in the static stream first queries some dynamic information from \\( F_{v}^{i}[t] \\) (the \\( t \\)-th feature in \\( F_{v}^{i} \\)) via a cross-attention module, this operation enhances the object query representations with dynamic information learned in the dynamic stream. Then we employ a query mixing operation (implemented as temporal self-attention) to query and mix information from object queries of other frames, which further injects cross-frame information to object queries. With the proposed cross-stream collaboration block, both the static and dynamic VL streams can learn complementary information from each other in an effective way. And the enhanced representations \\( \\widetilde{F}_{v}^{i}, \\widetilde{O}_{t}^{i} \\) are inputted to the next transformer layer in the static and dynamic VL streams, respectively." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.71, + 0.669, + 0.724 + ], + "angle": 0, + "content": "3.4. Prediction Heads" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.733, + 0.892, + 0.763 + ], + "angle": 0, + "content": "In this section, we introduce our prediction heads which include a spatial and a temporal prediction head." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.892, + 0.856 + ], + "angle": 0, + "content": "Spatial Prediction Head. The spatial prediction head predicts the location of the target object. For the \\(t\\)-th frame, the input is the object query representation \\(\\widetilde{O}_t^N \\in \\mathbb{R}^d\\) outputted by the last collaboration block. And we implement a 3-layer MLP to predict the bounding box location (represented by center coordinates and size) \\(\\tilde{b}_t \\in \\mathbb{R}^4\\) of the target object." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.892, + 0.902 + ], + "angle": 0, + "content": "Temporal Prediction Head. The temporal prediction head predicts the time span of the target temporal moment. Specifically, we first perform spatial mean pooling on \\(\\widetilde{F}_v^N\\)" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "23104" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.089, + 0.089, + 0.419, + 0.118 + ], + "angle": 0, + "content": "Table 1. Comparison results with previous works on HCSTVG-v1 test set and HCSTVG-v2 val set [24]." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.129, + 0.416, + 0.282 + ], + "angle": 0, + "content": "
DatasetMethodm_vIoUvIoU@0.3vIoU@0.5
HCSTVG-v1STGVT [24]18.226.89.5
STVGBert [22]20.429.411.3
TubeDETR [30]32.449.823.5
Ours36.962.234.8
HCSTVG-v2Yu et al. [33]30.0--
MMN [27]30.349.025.6
Aug. 2D-TAN [23]30.450.418.8
TubeDETR [30]36.458.830.6
Ours38.765.533.8
" + }, + { + "type": "table_caption", + "bbox": [ + 0.433, + 0.089, + 0.892, + 0.118 + ], + "angle": 0, + "content": "Table 2. Comparison results with state-of-the-art approaches on VidSTG [42] test set." + }, + { + "type": "table", + "bbox": [ + 0.447, + 0.129, + 0.877, + 0.281 + ], + "angle": 0, + "content": "
MethodDeclarative SentencesInterrogative Sentences
m_vIoUvIoU@0.3vIoU@0.5m_vIoUvIoU@0.3vIoU@0.5
STGRN [42]19.825.814.618.321.112.8
STGVT [24]21.629.818.9———
OMRN [41]23.132.616.420.628.414.1
STVGBert [22]24.030.918.422.526.016.0
TubeDETR [30]30.442.528.225.735.723.2
Ours33.747.232.828.539.926.2
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.305, + 0.47, + 0.355 + ], + "angle": 0, + "content": "to obtain a temporal feature \\( F_{d} \\in \\mathbb{R}^{T \\times d} \\), and we use an FC layer to adjust its dimension to \\( d_{m} \\). Then we follow 2D-TAN [39] to construct a 2D-proposal map \\( M \\in \\mathbb{R}^{T \\times T \\times d_{m}} \\):" + }, + { + "type": "equation", + "bbox": [ + 0.098, + 0.364, + 0.47, + 0.406 + ], + "angle": 0, + "content": "\\[\nM _ {i j} = \\left\\{ \\begin{array}{l l} \\mathbf {M P} ([ F _ {d} [ i ], F _ {d} [ i + 1 ], \\dots , F _ {d} [ j ]) & i \\leq j \\\\ \\mathbf {0} & i > j \\end{array} \\right. \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.418, + 0.473, + 0.601 + ], + "angle": 0, + "content": "where \\(\\mathrm{MP}(\\cdot)\\) is mean pooling and \\(M_{ij}\\) indicates the feature of temporal moment proposal \\(C_{ij}\\) with \\(t_i, t_{j+1}\\) as start and end time stamps, respectively. Here \\(t_i = \\frac{i}{T} \\cdot T_{video}\\) and \\(T_{video}\\) is the duration of the input video. \\(F_d[i]\\) is the \\(i\\)-th feature in \\(F_d\\). We employ several convolutional layers to transform the 2D map \\(M\\) into a score map \\(\\hat{S} \\in \\mathbb{R}^{T \\times T \\times 1}\\), where \\(\\hat{S}_{ij}\\) represents the matching score of temporal proposal \\(C_{ij}\\). For inference, we take the proposal with the highest score as the prediction of the target temporal time span. And we take the predicted bounding boxes from the spatial prediction head for frames within the predicted temporal span to form the final predicted tube." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.612, + 0.236, + 0.629 + ], + "angle": 0, + "content": "3.5. Model Training" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.637, + 0.47, + 0.683 + ], + "angle": 0, + "content": "We train our model with loss \\( L = L_{s} + L_{t} \\), where \\( L_{s} \\) and \\( L_{t} \\) are spatial and temporal localization losses, respectively. Specifically, they are defined as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.154, + 0.695, + 0.469, + 0.713 + ], + "angle": 0, + "content": "\\[\nL _ {s} = \\lambda_ {1} L _ {l 1} (\\hat {b}, b) + \\lambda_ {2} L _ {G I o U} (\\hat {b}, b), \\tag {4}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.157, + 0.717, + 0.469, + 0.735 + ], + "angle": 0, + "content": "\\[\nL _ {t} = \\lambda_ {3} L _ {t g} (\\hat {S}, S) + \\lambda_ {4} L _ {t a}, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.748, + 0.473, + 0.903 + ], + "angle": 0, + "content": "where \\( L_{l1} \\) and \\( L_{GIoU} \\) are L1 loss and GIoU loss [19] on the predicted bounding boxes, respectively. The temporal grounding loss \\( L_{tg} \\) is defined as a binary cross-entropy loss (as done in [23, 39]) between the predicted score map \\( \\hat{S} \\) and the ground truth map \\( S \\), where each element in \\( S \\) represents the IoU between the corresponding proposal and the ground truth temporal moment. \\( L_{ta} \\) is a commonly used temporal attentive loss [34] for accelerating convergence of \\( L_{tg} \\), and it encourages the model to predict a high matching score for those frames/clips inside the target temporal span." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.305, + 0.634, + 0.323 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.331, + 0.707, + 0.348 + ], + "angle": 0, + "content": "4.1. Experimental Settings" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.355, + 0.892, + 0.702 + ], + "angle": 0, + "content": "Datasets. We evaluate our method on HCSTVG-v1 dataset, HCSTVG-v2 dataset [24] and VidSTG [42] dataset. HCSTVG datasets are collected from movie scenes and the duration of each video clip is around 20 seconds. This set is quite challenging for spatio-temporal grounding as some video clips contain many persons conducting similar actions. HCSTVG-v1 dataset consists of 4500 and 1160 video-text pairs for training and testing, respectively. HCSTVG-v2 dataset expanded HCSTVG-v1 dataset and improved the annotation quality. It contains 10131, 2000, 4413 samples for training, validation and testing, respectively. Since the annotations of the test set are not publicly available in HCSTVG-v2, we report results on the validation set. VidSTG dataset [42] is constructed based on VidOR [20] dataset according to the object relation annotations. It contains 99943 video-text pairs, including 44808 declarative sentence queries and 55135 interrogative sentence queries. The training, validation and test set consist of 80684, 8956 and 10303 sentences respectively, and 5436, 602 and 732 videos respectively. However, since VidSTG [42] is annotated based on VidOR [20] dataset, the text queries are limited to describe the pre-defined object/relation categories in VidOR [20]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.703, + 0.893, + 0.903 + ], + "angle": 0, + "content": "Evaluation Metrics. We follow previous works [22, 23, 30, 42] to use mean vIoU as the main evaluation metric. vIoU is defined as \\(\\frac{1}{|T_u|}\\sum_{t\\in T_i}I_oU(\\hat{b}_t,b_t)\\), where \\(T_{i}\\) and \\(T_{u}\\) indicate the intersection and union between the time intervals obtained from ground truth annotation and model prediction, respectively. \\(\\hat{b}_t,b_t\\) are the predicted bounding box and ground truth bounding box for the \\(t\\)-th frame, respectively. We average the vIoU score over all the samples to obtain mean vIoU (termed m_vIoU). We also report vIoU@R which indicates the proportion of samples with vIoU higher than \\(R\\). In ablation studies, we follow TubeDETR [30] to also report the sIoU and tIoU metrics that evaluate the spatial and temporal grounding performances, respectively." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.958 + ], + "angle": 0, + "content": "23105" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.09, + 0.47, + 0.144 + ], + "angle": 0, + "content": "They are defined as \\( sIoU = \\frac{1}{|T_{gt}|}\\sum_{t\\in T_i}IoU(\\hat{b_t},b_t) \\) and \\( tIoU = \\frac{|T_i|}{|T_u|} \\), where \\( T_{gt} \\) indicates the set of frames inside the ground truth time interval." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.144, + 0.471, + 0.356 + ], + "angle": 0, + "content": "Implementation Details. We use ResNet101 [9] as our image encoder and Roberta-base [14] as the text encoder to extract image visual features and language features. And we use the Slowfast [4] model pre-trained on AVA [8] as the video encoder. We initialize part of the weights of our static VL stream using pre-trained MDETR [10] as done in [23, 30]. The number of transformer layers in both streams is set as \\( N = 6 \\). The loss weights are set as \\( \\lambda_1 = 5, \\lambda_2 = 2, \\lambda_3 = 5, \\lambda_4 = 1 \\). We train our model on VidSTG [42], HCSTVG-v1, HCSTVG-v2 [24] datasets for 10, 10, 4 epochs, respectively. For most other hyperparameters, we keep them consistent with previous works [10, 30]. Further implementation details can be found in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.367, + 0.279, + 0.383 + ], + "angle": 0, + "content": "4.2. Experimental Results" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.39, + 0.47, + 0.435 + ], + "angle": 0, + "content": "In this section, we compare the performance of our approach with previous methods on HCSTVG dataset [24] and VidSTG [42] dataset." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.436, + 0.471, + 0.632 + ], + "angle": 0, + "content": "HCSTVG datasets. We first compare our method with state-of-the-arts on HCSTVG datasets [24] and results are shown in Table 1. On HCSTVG-v1 test set, our approach outperforms previous state-of-the-art method TubeDETR [30], by a considerable margin. We achieve \\(36.9\\%\\) m_vIoU, which is \\(4.5\\%\\) higher than TubeDETR. On HCSTVG-v2 validation set [24], we compare our method with TubeDETR [30] and some winners from HCVG challenge2021. As shown in Table 1, we also outperform TubeDETR by \\(2.3\\%\\) m_vIoU. The above results demonstrate that our proposed model with two collaborative vision-language streams can effectively capture static-dynamic cross-modal dependencies for addressing STVG task." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.632, + 0.47, + 0.723 + ], + "angle": 0, + "content": "VidSTG dataset. We also report results on VidSTG dataset [42]. As listed in Table 2, we also achieve state-of-the-art performance with a m_vIoU of \\(33.7\\%\\) on declarative sentences and \\(28.5\\%\\) on interrogative sentences, respectively. This further shows the effectiveness of the proposed collaborative two-stream model." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.735, + 0.231, + 0.751 + ], + "angle": 0, + "content": "4.3. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.758, + 0.47, + 0.803 + ], + "angle": 0, + "content": "In this section, we conduct ablation experiments on HCSTVG-v2 dataset [24] to verify the effectiveness of some important components in the proposed method." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.804, + 0.469, + 0.864 + ], + "angle": 0, + "content": "Quantitative evaluation on the Cross-Stream Collaboration Block. We first verify the effectiveness of static-to-dynamic and dynamic-to-static information transmission blocks by removing one of them from our developed frame" + }, + { + "type": "table_caption", + "bbox": [ + 0.5, + 0.09, + 0.89, + 0.104 + ], + "angle": 0, + "content": "Table 3. Evaluation on the cross-stream collaboration mechanism." + }, + { + "type": "table", + "bbox": [ + 0.527, + 0.109, + 0.86, + 0.185 + ], + "angle": 0, + "content": "
s-to-dd-to-sm_vIoUvIoU@0.3vIoU@0.5m_tIoUm_sIoU
36.461.229.556.164.9
✓37.563.631.457.464.6
✓37.162.330.256.265.7
✓✓38.765.533.858.165.7
" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.204, + 0.892, + 0.446 + ], + "angle": 0, + "content": "work. We present the evaluation results in Table 3. Compared to the baseline in which the whole cross-stream collaboration block is removed (the first line in Table 3), it is clear that individually adopting static-to-dynamic information transmission (termed \"s-to-d\") or dynamic-to-static information transmission (termed \"d-to-s\") can both improve the grounding performance on vIoU metrics. We also observe that \"s-to-d\" mainly improves the temporal grounding accuracy while \"d-to-s\" improves the spatial grounding performance. And our full model that jointly employs the whole cross-stream collaboration block achieves the best performance of \\(38.7\\%\\) m_vIoU. The above results show that both the static-to-dynamic and the dynamic-to-static information transmission blocks are effective for providing complementary information to the other stream, and thus employing them greatly improves the performance." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.448, + 0.893, + 0.884 + ], + "angle": 0, + "content": "Qualitative evaluation on the Cross-Stream Collaboration Block. To better understand how and why our cross-stream collaboration block significantly improves the grounding accuracy, we further provide 2 qualitative samples in Figure 6. In the first sample, there are three boys coming to the dining table. In this case, our baseline model without cross-stream collaboration failed to predict correct temporal grounding results since it is confused by the similar action performed by the three boys. With our proposed cross-stream collaboration block, the dynamic VL stream can distinguish which boy is the \"tall boy\" by absorbing knowledge from the static VL stream, then it can correctly predict the target temporal time span. For the second sample, there are two women both dressed in a skirt, thus in our baseline where the cross-stream collaboration is removed, the static stream failed to capture and understand the action \"let goes of ... and turns into the room\", thus it cannot distinguish which is the target woman. By employing the proposed cross-stream collaboration block, the motion information learned in the dynamic VL stream can be transmitted to the static stream to make it aware of the motion information which can help distinguish the target person. The above two samples clearly show that the proposed cross-stream collaboration block can effectively exchange complementary information between the static and the dynamic VL streams, which can greatly reduce the uncertainty in the hard and ambiguous cases where some objects have a similar appearance or motion so that the model can produce a more accurate prediction." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.886, + 0.892, + 0.902 + ], + "angle": 0, + "content": "Spatial information in the dynamic VL stream. We" + }, + { + "type": "page_footnote", + "bbox": [ + 0.076, + 0.875, + 0.469, + 0.901 + ], + "angle": 0, + "content": "1In this paper, we report the corrected TubeDETR [30] performance updated by TubeDETR authors at https://github.com/antoyang/TubeDETR." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "23106" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.093, + 0.089, + 0.493, + 0.103 + ], + "angle": 0, + "content": "Query: The tall boy comes to the dining table and sits down." + }, + { + "type": "image", + "bbox": [ + 0.093, + 0.104, + 0.885, + 0.284 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.3, + 0.893, + 0.342 + ], + "angle": 0, + "content": "Figure 6. Visualization of the predicted tube of our approach (yellow), our approach without cross-stream collaboration (red), and ground truth (green). In the first sample, our approach without cross-stream collaboration predicts a wrong temporal time span. In the second sample, ours w/o the collaboration block predicts wrong spatial bounding boxes. (Best viewed zoomed in on screen.)" + }, + { + "type": "table_caption", + "bbox": [ + 0.1, + 0.36, + 0.446, + 0.374 + ], + "angle": 0, + "content": "Table 4. Evaluation on designs of the dynamic VL stream." + }, + { + "type": "table", + "bbox": [ + 0.13, + 0.378, + 0.412, + 0.417 + ], + "angle": 0, + "content": "
Method|Ours|Ours w/o spatial|Ours w/o mean pool
m_tIoU|56.1|54.7|56.2
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.426, + 0.468, + 0.454 + ], + "angle": 0, + "content": "Table 5. Comparison between different cross-stream collaboration block designs." + }, + { + "type": "table", + "bbox": [ + 0.107, + 0.458, + 0.436, + 0.546 + ], + "angle": 0, + "content": "
Methodm_vIoUvIoU@0.3vIoU@0.5m_tIoUm_sIoU
Ours38.765.533.858.165.7
None36.461.229.556.164.9
sum36.662.530.055.664.4
FC + sum36.862.830.557.363.4
concat + FC37.163.230.356.963.9
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.567, + 0.47, + 0.853 + ], + "angle": 0, + "content": "present ablation studies on different design choices of the dynamic VL stream and results are presented in Table 4. Here, we discard the static stream to eliminate its influence and we report temporal grounding accuracy using m_tIoU metric. By mean pooling at the spatial dimension of the video feature inputted to the dynamic VL stream (termed \"Ours w/o spatial\"), it results in a drop of more than \\(1\\%\\) m_tIoU which shows the importance of learning spatial visual interaction for modeling dynamic cues. And since we develop the dynamic VL stream to mainly learn dynamic cross-modal correspondence, we employ spatial mean pooling on \\(\\mathcal{H}_v\\) to obtain \\(\\overline{\\mathcal{H}}_v\\) so that the text-to-visual cross-attention is performed only on the temporal dimension which can greatly reduce computation cost. By removing this mean pooling operation (termed \"Ours w/o mean pool\"), the text-to-visual cross-attention is performed among all spatio-temporal tokens, but it has very limited improvement on the performance, which validates the rationality of our design." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Further study of the collaboration block. To further verify the effectiveness of the proposed collaboration block, we replace it with some symmetric operations, including" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.361, + 0.892, + 0.543 + ], + "angle": 0, + "content": "\"sum\", \"FC + sum\" and \"concat + FC\", which are some operations commonly used in the community [4, 21]. As shown in Table 5, \"concat + FC\" achieves the best performance among these baselines, but it only achieves a limited improvement of \\(0.7\\%\\) m_vIoU, which indicates that simply merging the information from the two streams can not fully exploit their complementary abilities. In contrast, \"Ours\" outperforms \"concat + FC\" by a considerable margin of more than \\(3\\%\\) in terms of vIoU@0.5, demonstrating that our asymmetric cross-stream collaboration block enables the two streams to effectively exchange complementary information and to collaboratively reason the target object." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.563, + 0.618, + 0.578 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.591, + 0.892, + 0.801 + ], + "angle": 0, + "content": "In this work, we propose a novel framework for Spatio-Temporal Video Grounding task. Our framework mainly consists of a static VL(vision-language) stream which captures static visual cues like object appearance, and a dynamic VL stream which understands dynamic visual cues like action. We elaborately design the framework as a collaborative system in which the two streams collaborative localize the target object via a novel cross-stream collaboration block, which is proven to be effective. We also present visualization results which provide some interesting insights and show how the collaboration of the two streams helps improve the prediction. Our approach significantly outperforms previous methods on VidSTG [42] and HC-STVG dataset [24], which demonstrates its effectiveness." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.803, + 0.892, + 0.9 + ], + "angle": 0, + "content": "Acknowledgements. This work was supported partially by the NSFC (U21A20471, U22A2095, 62076260, 61772570), Guangdong Natural Science Funds Project (2020B1515120085, 2023B1515040025), Guangdong NSF for Distinguished Young Scholar (2022B1515020009, 2018B030306025), and the Key-Area Research and Development Program of Guangzhou (202007030004). We thank Bing Shuai for the helpful discussions." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "23107" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.116, + 0.47, + 0.185 + ], + "angle": 0, + "content": "[1] Lisa Anne Hendricks, Oliver Wang, Eli Shechtman, Josef Sivic, Trevor Darrell, and Bryan Russell. Localizing moments in video with natural language. In Proceedings of the IEEE international conference on computer vision, pages 5803-5812, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.187, + 0.472, + 0.243 + ], + "angle": 0, + "content": "[2] Gedas Bertasius, Heng Wang, and Lorenzo Torresani. Is space-time attention all you need for video understanding? In Proceedings of the International Conference on Machine Learning (ICML), July 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.246, + 0.471, + 0.302 + ], + "angle": 0, + "content": "[3] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In European conference on computer vision, pages 213-229. Springer, 2020. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.303, + 0.471, + 0.358 + ], + "angle": 0, + "content": "[4] Christoph Feichtenhofer, Haoqi Fan, Jitendra Malik, and Kaiming He. Slowfast networks for video recognition. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6202-6211, 2019. 1, 2, 3, 4, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.36, + 0.47, + 0.416 + ], + "angle": 0, + "content": "[5] Christoph Feichtenhofer, Axel Pinz, and Richard P Wildes. Spatiotemporal multiplier networks for video action recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4768-4777, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.418, + 0.47, + 0.473 + ], + "angle": 0, + "content": "[6] Jiyang Gao, Chen Sun, Zhenheng Yang, and Ram Nevatia. Tall: Temporal activity localization via language query. In Proceedings of the IEEE international conference on computer vision, pages 5267-5275, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.476, + 0.47, + 0.545 + ], + "angle": 0, + "content": "[7] Yash Goyal, Tejas Khot, Douglas Summers-Stay, Dhruv Bartra, and Devi Parikh. Making the v in vqa matter: Elevating the role of image understanding in visual question answering. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6904-6913, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.547, + 0.47, + 0.643 + ], + "angle": 0, + "content": "[8] Chunhui Gu, Chen Sun, David A Ross, Carl Vondrick, Caroline Pantofaru, Yeqing Li, Sudheendra Vijayanarasimhan, George Toderici, Susanna Ricco, Rahul Sukthankar, et al. Ava: A video dataset of spatio-temporally localized atomic visual actions. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 6047-6056, 2018. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.645, + 0.47, + 0.701 + ], + "angle": 0, + "content": "[9] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 4, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.703, + 0.47, + 0.773 + ], + "angle": 0, + "content": "[10] Aishwarya Kamath, Mannat Singh, Yann LeCun, Gabriel Synnaeve, Ishan Misra, and Nicolas Carion. Mdetr-modulated detection for end-to-end multi-modal understanding. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1780-1790, 2021. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.775, + 0.47, + 0.83 + ], + "angle": 0, + "content": "[11] Kun Li, Dan Guo, and Meng Wang. Proposal-free video grounding with contextual pyramid network. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 35, pages 1902-1910, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.832, + 0.47, + 0.9 + ], + "angle": 0, + "content": "[12] Liunian Harold Li*, Pengchuan Zhang*, Haotian Zhang*, Jianwei Yang, Chunyuan Li, Yiwu Zhong, Lijuan Wang, Lu Yuan, Lei Zhang, Jenq-Neng Hwang, Kai-Wei Chang, and Jianfeng Gao. Grounded language-image pre-training. In CVPR, 2022. 3" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.116, + 0.472, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.163 + ], + "angle": 0, + "content": "[13] Yue Liao, Si Liu, Guanbin Li, Fei Wang, Yanjie Chen, Chen Qian, and Bo Li. A real-time cross-modality correlation filtering method for referring expression comprehension. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10880-10889, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.164, + 0.892, + 0.233 + ], + "angle": 0, + "content": "[14] Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. Roberta: A robustly optimized bert pretraining approach. arXiv preprint arXiv:1907.11692, 2019. 4, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.235, + 0.892, + 0.29 + ], + "angle": 0, + "content": "[15] Jiasen Lu, Dhruv Batra, Devi Parikh, and Stefan Lee. Vilbert: Pretraining task-agnostic visiolinguistic representations for vision-and-language tasks. Advances in neural information processing systems, 32, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.292, + 0.892, + 0.359 + ], + "angle": 0, + "content": "[16] Jonghwan Mun, Minsu Cho, and Bohyung Han. Local-global video-text interactions for temporal grounding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10810-10819, 2020. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.362, + 0.892, + 0.432 + ], + "angle": 0, + "content": "[17] Guoshun Nan, Rui Qiao, Yao Xiao, Jun Liu, Sicong Leng, Hao Zhang, and Wei Lu. Interventional video grounding with dual contrastive learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2765-2775, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.434, + 0.892, + 0.489 + ], + "angle": 0, + "content": "[18] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. Advances in neural information processing systems, 28, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.491, + 0.892, + 0.572 + ], + "angle": 0, + "content": "[19] Hamid Rezatofighi, Nathan Tsoi, JunYoung Gwak, Amir Sadeghian, Ian Reid, and Silvio Savarese. Generalized intersection over union: A metric and a loss for bounding box regression. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 658-666, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.576, + 0.892, + 0.644 + ], + "angle": 0, + "content": "[20] Xindi Shang, Donglin Di, Junbin Xiao, Yu Cao, Xun Yang, and Tat-Seng Chua. Annotating objects and relations in user-generated videos. In Proceedings of the 2019 on International Conference on Multimedia Retrieval, pages 279-287. ACM, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.647, + 0.892, + 0.702 + ], + "angle": 0, + "content": "[21] Karen Simonyan and Andrew Zisserman. Two-stream convolutional networks for action recognition in videos. Advances in neural information processing systems, 27, 2014. 1, 2, 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.704, + 0.892, + 0.772 + ], + "angle": 0, + "content": "[22] Rui Su, Qian Yu, and Dong Xu. Stvgbert: A visual-linguistic transformer based framework for spatio-temporal video grounding. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 1533–1542, October 2021. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.775, + 0.892, + 0.83 + ], + "angle": 0, + "content": "[23] Chaolei Tan, Zihang Lin, Jian-Fang Hu, Xiang Li, and Wei-Shi Zheng. Augmented 2d-tan: A two-stage approach for human-centric spatio-temporal video grounding. arXiv preprint arXiv:2106.10634, 2021. 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.832, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[24] Zongheng Tang, Yue Liao, Si Liu, Guanbin Li, Xiaojie Jin, Hongxu Jiang, Qian Yu, and Dong Xu. Human-centric spatio-temporal video grounding with visual transformers. IEEE Transactions on Circuits and Systems for Video Technology, pages 1-1, 2021. 1, 2, 6, 7, 8" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "23108" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.468, + 0.134 + ], + "angle": 0, + "content": "[25] Limin Wang, Yuanjun Xiong, Zhe Wang, and Yu Qiao. Towards good practices for very deep two-stream convnets. arXiv preprint arXiv:1507.02159, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.134, + 0.468, + 0.189 + ], + "angle": 0, + "content": "[26] Xuanhan Wang, Lianli Gao, Peng Wang, Xiaoshuai Sun, and Xianglong Liu. Two-stream 3-d convnet fusion for action recognition in videos with arbitrary size and length. IEEE Transactions on Multimedia, 20(3):634-644, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.19, + 0.468, + 0.244 + ], + "angle": 0, + "content": "[27] Zhenzhi Wang, Limin Wang, Tao Wu, Tianhao Li, and Gangshan Wu. Negative sample matters: A renaissance of metric learning for temporal grounding. CoRR, abs/2109.04872, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.246, + 0.468, + 0.314 + ], + "angle": 0, + "content": "[28] Zuxuan Wu, Yu-Gang Jiang, Xi Wang, Hao Ye, and Xi-angyang Xue. Multi-stream multi-class fusion of deep networks for video classification. In Proceedings of the 24th ACM international conference on Multimedia, pages 791–800, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.316, + 0.468, + 0.384 + ], + "angle": 0, + "content": "[29] Shaoning Xiao, Long Chen, Songyang Zhang, Wei Ji, Jian Shao, Lu Ye, and Jun Xiao. Boundary proposal network for two-stage natural language video localization. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 35, pages 2986-2994, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.385, + 0.468, + 0.454 + ], + "angle": 0, + "content": "[30] Antoine Yang, Antoine Miech, Josef Sivic, Ivan Laptev, and Cordelia Schmid. Tubedetr: Spatio-temporal video grounding with transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 1, 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.455, + 0.468, + 0.523 + ], + "angle": 0, + "content": "[31] Kexin Yi, Jiajun Wu, Chuang Gan, Antonio Torralba, Pushmeet Kohli, and Josh Tenenbaum. Neural-symbolic vqa: Disentangling reasoning from vision and language understanding. Advances in neural information processing systems, 31, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.525, + 0.468, + 0.594 + ], + "angle": 0, + "content": "[32] Licheng Yu, Zhe Lin, Xiaohui Shen, Jimei Yang, Xin Lu, Mohit Bansal, and Tamara L Berg. Mattnet: Modular attention network for referring expression comprehension. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 1307-1315, 2018. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.595, + 0.468, + 0.636 + ], + "angle": 0, + "content": "[33] Yi Yu, Xinying Wang, Wei Hu, Xun Luo, and Cheng Li. 2rd place solutions in the hc-stvg track of person in context challenge 2021. arXiv preprint arXiv:2106.07166, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.637, + 0.468, + 0.704 + ], + "angle": 0, + "content": "[34] Yitian Yuan, Tao Mei, and Wenwu Zhu. To find where you talk: Temporal sentence localization in video with attention based location regression. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 33, pages 9159-9166, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.706, + 0.468, + 0.773 + ], + "angle": 0, + "content": "[35] Runhao Zeng, Haoming Xu, Wenbing Huang, Peihao Chen, Mingkui Tan, and Chuang Gan. Dense regression network for video grounding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10287-10296, 2020. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.776, + 0.468, + 0.845 + ], + "angle": 0, + "content": "[36] Da Zhang, Xiyang Dai, Xin Wang, Yuan-Fang Wang, and Larry S Davis. Man: Moment alignment network for natural language moment retrieval via iterative graph adjustment. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1247-1257, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.846, + 0.468, + 0.9 + ], + "angle": 0, + "content": "[37] Hanwang Zhang, Yulei Niu, and Shih-Fu Chang. Grounding referring expressions in images by variational context. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 4158-4166, 2018. 3" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.468, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.092, + 0.892, + 0.161 + ], + "angle": 0, + "content": "[38] Hao Zhang, Aixin Sun, Wei Jing, Liangli Zhen, Joey Tianyi Zhou, and Rick Siow Mong Goh. Natural language video localization: A revisit in span-based question answering framework. IEEE transactions on pattern analysis and machine intelligence, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.163, + 0.892, + 0.231 + ], + "angle": 0, + "content": "[39] Songyang Zhang, Houwen Peng, Jianlong Fu, and Jiebo Luo. Learning 2d temporal adjacent networks for moment localization with natural language. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 34, pages 12870-12877, 2020. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.233, + 0.892, + 0.301 + ], + "angle": 0, + "content": "[40] Zhiwang Zhang, Dong Xu, Wanli Ouyang, and Chuanqi Tan. Show, tell and summarize: Dense video captioning using visual cue aided sentence summarization. IEEE Transactions on Circuits and Systems for Video Technology, 30(9):3130-3139, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.304, + 0.892, + 0.373 + ], + "angle": 0, + "content": "[41] Zhu Zhang, Zhou Zhao, Zhijie Lin, Baoxing Huai, and Jing Yuan. Object-aware multi-branch relation networks for spatio-temporal video grounding. In Proceedings of the Twenty-Ninth International Joint Conference on Artificial Intelligence, IJCAI'20, 2021. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.374, + 0.892, + 0.443 + ], + "angle": 0, + "content": "[42] Zhu Zhang, Zhou Zhao, Yang Zhao, Qi Wang, Huasheng Liu, and Lianli Gao. Where does it exist: Spatio-temporal video grounding for multi-form sentences. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2020. 1, 2, 6, 7, 8" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.092, + 0.892, + 0.443 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "23109" + } + ] +] \ No newline at end of file diff --git a/2023/Collaborative Static and Dynamic Vision-Language Streams for Spatio-Temporal Video Grounding/40d01bed-f6c0-4114-9925-d43e564f9389_origin.pdf b/2023/Collaborative Static and Dynamic Vision-Language Streams for Spatio-Temporal Video Grounding/40d01bed-f6c0-4114-9925-d43e564f9389_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..06fb932c20e9c4f7f9d1d3d68dba154ce7faa5f4 --- /dev/null +++ b/2023/Collaborative Static and Dynamic Vision-Language Streams for Spatio-Temporal Video Grounding/40d01bed-f6c0-4114-9925-d43e564f9389_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c901111faedad375f464d5c49d2cd50f78c5aae3933b9b8d7a10c44073a3ae9f +size 1862156 diff --git a/2023/Collaborative Static and Dynamic Vision-Language Streams for Spatio-Temporal Video Grounding/full.md b/2023/Collaborative Static and Dynamic Vision-Language Streams for Spatio-Temporal Video Grounding/full.md new file mode 100644 index 0000000000000000000000000000000000000000..05463a7cf0e932bc3eb6335fc8e34d95173c4d7b --- /dev/null +++ b/2023/Collaborative Static and Dynamic Vision-Language Streams for Spatio-Temporal Video Grounding/full.md @@ -0,0 +1,270 @@ +# Collaborative Static and Dynamic Vision-Language Streams for Spatio-Temporal Video Grounding + +Zihang Lin1, Chaolei Tan1, Jian-Fang Hu1,3,4*, Zhi Jin1, Tiancai Ye2, Wei-Shi Zheng1,3,4, Sun Yat-sen University, China + +3Guangdong Province Key Laboratory of Information Security Technology, China +4 Key Laboratory of Machine Intelligence and Advanced Computing, Ministry of Education, China + +{1inzh59, tanchlei}@mail2.sysu.edu.cn, {hujf5, jinzh26}@mail.sysu.edu.cn, tiancaiye@tencent.com, wszheng@ieee.org + +# Abstract + +Spatio-Temporal Video Grounding (STVG) aims to localize the target object spatially and temporally according to the given language query. It is a challenging task in which the model should well understand dynamic visual cues (e.g., motions) and static visual cues (e.g., object appearances) in the language description, which requires effective joint modeling of spatio-temporal visual-linguistic dependencies. In this work, we propose a novel framework in which a static vision-language stream and a dynamic vision-language stream are developed to collaboratively reason the target tube. The static stream performs cross-modal understanding in a single frame and learns to attend to the target object spatially according to intra-frame visual cues like object appearances. The dynamic stream models visual-linguistic dependencies across multiple consecutive frames to capture dynamic cues like motions. We further design a novel cross-stream collaborative block between the two streams, which enables the static and dynamic streams to transfer useful and complementary information from each other to achieve collaborative reasoning. Experimental results show the effectiveness of the collaboration of the two streams and our overall framework achieves new state-of-the-art performance on both HCSTVG and VidSTG datasets. + +# 1. Introduction + +Vision-language cross-modal understanding is a challenging yet important research problem that bridges the communication between humans and artificial intelligence systems. It has attracted increasing attention and many vision-language tasks were studied in recent years, like vi + +sual grounding [13, 32], VQA [7, 31], image/video captioning [15, 40], etc. In this work, we focus on a challenging vision-language task named Spatio-Temporal Video Grounding (STVG) which was recently proposed in [42]. Given a language query indicating an object (as shown in Figure 1), STVG aims to localize the target object spatially and temporally in the video. In this task, the input language query may express different kinds of visual concepts, thus the model requires to well capture and understand these concepts in both vision and language modalities. + +In STVG task, dynamic visual concepts like human motion and static visual concepts like object appearance are both important for distinguishing the target object from other objects that occurred in the same video. For example, in the first sample in Figure 1, the two men are dressed alike (i.e., their static appearance cues are similar), and we can only distinguish them by motion. In the second sample, the two women perform similar actions, they both stand up (i.e., they have similar dynamic motion cues), here, we can only distinguish them by their clothes. The above examples show that static or dynamic visual cues alone cannot solve the STVG task well. And it implies that modeling static and dynamic visual-linguistic dependencies and collaboratively utilizing them are important for addressing STVG task. + +Humans treat static and dynamic cues differently [4, 21], but this was overlooked in previous STVG works [22, 24, 30]. Taking the first query in Figure 1 as an example, a human would randomly or evenly click on some locations on the video's progress bar to find candidate frames containing a man in blue coat. And he will play the video around that frame and attend on the candidate man to check whether he performs the action described in the text (i.e., "turns around and stops by the stone"). In the above process, the human understands static and dynamic cues in different ways (i.e., view a single frame and watch the video clip, respectively) and determines the target object by jointly considering the + +Query: The man in the blue coat turns around and stops by the stone. + +![](images/03aab8328a8d944b4c530af2fe68465433d7aee9e43f259850bd582dda0c5472.jpg) + +Query: The woman in light gray stands up and turns. + +![](images/da337362c1f2cc61ea9ebfa9a52fba8ebcd14cbbf6c017d6e80b21fbf4192c51.jpg) +Figure 1. Two examples for Spatio-Temporal Video Grounding task. In the first case, the two men are both dressed alike in blue, thus understanding the action described in the query sentence is essential to recognize the person of interest. In the second case, both of the two women stand up, we can only distinguish them by their clothes. (Best viewed zoomed in on screen.) + +static and dynamic cues in an "attend-and-check" process. Inspired by the above observations, we propose a framework that consists of a static and a dynamic vision-language (VL) streams to model static and dynamic cues, respectively. And we design a novel cross-stream collaboration block between the two streams to simulate the "attend-and-check" process, which exchanges useful and complementary information learned in each stream and enables the two streams to collaboratively reason the target object. + +Specifically, in this work, our static vision-language (VL) stream learns to attend to some candidate regions according to static visual cues like appearance, while the dynamic VL stream learns to understand the dynamic visual cues like action described in the text query. Then, in the collaboration block, we guide the dynamic stream to only focus and attend on the motion of the candidate objects by using the learned attended region in the static stream. And we transfer the text-motion matching information learned in the dynamic stream to the static stream, to help it further check and determine the target object and predict a more consistent tube. With the above cross-stream collaboration blocks, both the static and dynamic vision-language streams can learn reciprocal information from the other stream, which is effective for achieving more accurate spatio-temporal grounding predictions. We conduct experiments on HCSTVG [24] and VidSTG [42] datasets and our approach outperforms previous approaches by a considerable margin. Ablation studies demonstrate the effectiveness of each component in our proposed cross-stream collaboration block and show its superiority over commonly used counterparts in video understanding works [4, 21]. + +In summary, our contributions are: 1), we develop an effective framework that contains two parallel streams to model static-dynamic visual-linguistic dependencies for complete cross-modal understanding; 2), we propose a novel cross-stream collaboration block between the two streams to exchange reciprocal information for each other + +and enable collaborative reasoning of the target object; 3). Our overall framework achieves state-of-the-art performances on HCSTVG [24] and VidSTG datasets [42]. + +# 2. Related Work + +Spatio-Temporal Video Grounding. Spatio-temporal video grounding aims to localize the target object spatially and temporally according to a language query. Early works [24, 41, 42] employ pre-trained detectors like Faster R-CNN [18] to detect objects in each frame, and build their model upon the detection features of each object. STGRN [42] and OMRN [41] learn object relations with spatio-temporal graphs and multi-branch relation networks, respectively. STGVT [24] learns cross-modal representations with visual transformers for video-sentence matching and temporal localization. However, these methods are limited to the pre-detected bounding boxes and cannot localize object categories that are not defined in the pretraining dataset. Recent works [22, 30] design their framework based on strong pre-trained vision-language models and train their models end-to-end. These approaches do not require pre-detected bounding boxes and they perform much better. However, they directly collapse some dimensions (i.e., spatial or temporal) in cross-modal attention calculation to reach an acceptable computation cost, which results in large information loss, and thus they can not well capture the important static-dynamic cross-modal contextual relationship. To address this weakness, we propose to model static and dynamic visual-linguistic correspondence with parallel streams and learn to reason the target object in a collaborative framework. + +Temporal Grounding. Given a language query, temporal grounding aims to predict the time span in a video specified by the language query [1, 6]. Existing works [11, 16, 17, 29, 35, 36, 38, 39] can be divided into proposal-based and proposal-free approaches. Proposal-based approaches generate some proposals (manually designed [29, 39] or + +![](images/dbdfe90608f64b0d4cad33eee37bb20dccab73e0ad187dbee22bc5ab28785bda.jpg) +Figure 2. An overview of the proposed framework. Our framework mainly consists of a static VL( vision-language) stream and a dynamic VL stream. The static stream learns to attend to the spatial locations of the target object according to static cues like object appearance. The dynamic stream learns motion-text correspondence according to dynamic cues like human action. We further devise a cross-stream collaboration block that enables the two streams to query useful and complementary information from the other stream. + +predicted by the model [35]) and predict the IoU scores between each proposal and the ground truth. Proposal-free approaches address moment localization by directly regressing the starting/ending points of the target time spans [11, 16] or predicting the probability of being a starting/ending point for each clip and choosing the time span with the highest joint probability as the prediction [38]. In this work, we follow [23, 39] to employ a proposal-based grounding head. And in contrast to most temporal grounding works that collapse the spatial dimension, we also incorporate spatial cues to enhance temporal localization. + +Text-conditioned Object Detection. Given an image, text-conditioned object detection aims to localize the target object indicated by a sentence (also named referring image expression task) or all objects described in a sentence (also named phrase grounding task). Most existing works [32,37] develop their models based on off-the-shelf detection models and match the detected objects with the sentence query. Recently, MDETR [10] and GLIP [12] unified detection and phrase grounding tasks, and they trained their models with large-scale data. These models can achieve excellent performance on text-conditioned object detection and are used to initialize model weights for STVG task [23, 30]. However, these models lack the ability of modeling dynamic cues. In this work, we design a novel collaboration framework to capture both static and dynamic cues for STVG task. + +Two-stream Models. Considering that static and dynamic visual cues are quite different but complementary in video understanding, a number of two-stream models [4,5,21,25, 26, 28] are proposed to capture these two kinds of visual cues. For example, Two-Stream ConvNet [21] introduces an optical flow stream to capture object motion. SlowFast [4] develops slow and fast pathways with different temporal resolutions to capture static and dynamic visual cues. However, these works typically fuse different streams equally + +via some simple symmetric operations, e.g., assembling the output, sum/concat/FC/temporal-convolution on the features. In contrast, we propose a novel asymmetrical cross-stream collaboration block to exploit better collaboration between the two streams considering their different abilities. + +# 3. Collaborative Static-Dynamic VL Streams + +We illustrate our framework in Figure 2. Our framework mainly consists of a static and a dynamic VL (vision-language) stream to model static and dynamic visual-linguistic dependencies for complete cross-modal understanding. The static stream performs cross-modal understanding for static contextual information, i.e., finding the object that matches the query text in a still frame according to static visual cues, like object appearance which is important for achieving accurate spatial grounding. The dynamic stream performs cross-modal understanding for dynamic contextual information, i.e., finding the temporal moment that best matches the query sentence according to dynamic visual cues, like action which is important for achieving accurate temporal grounding. Both the static and dynamic streams learn cross-modal correspondence with a transformer-like architecture that contains multiple layers. To enable information transmission between the two streams, we further design a novel cross-stream collaboration block which is placed after each transformer layer in the two streams. In this way, both the static and dynamic VL streams can absorb complementary information from each other and achieve collaborative inferring, which can greatly reduce the uncertainty in ambiguous and hard cases where different objects have similar appearances or perform similar actions. Finally, we employ the prediction head to predict a spatio-temporal tube. In the following, we will introduce each component in detail. + +![](images/99dd6c752da3f4540f0da8e13041971ab5b12d44fb8da44ebe6bc5754c2af8ce.jpg) +(a) + +![](images/385c90247b5e7890bb1ce7fb01fa08a070f95f4c94f1a31e4300c9cad16b4863.jpg) +(b) +Figure 3. The detailed architectures of the proposed framework. (a): The cross-modal transformer in the static VL stream. (b): The architecture of each space-time cross-modal transformer layer in the dynamic VL stream. + +# 3.1. Static VL Stream + +The static VL stream is employed to perform cross-modal static context understanding and it processes each frame independently. Inspired by DETR [3], we design our static VL stream as a stack of $N$ cross-modal transformer encoder layers and $N$ cross-modal transformer decoder layers, as presented in Figure 2 and 3 (a). For the $t$ -th frame, the inputs of our cross-modal transformer encoder are the concatenation of $\mathbb{R}^{HW\times d}$ -sized visual features and $\mathbb{R}^{L\times d}$ -sized language features, where $H,W$ are the resolution of the visual features extracted from the static image frame, and $L$ is the number of text tokens in the input text query. These visual and language features are obtained by a pretrained image encoder [9] and language encoder [14], respectively. We employ an FC layer behind each feature extractor to project the visual and language features to a shared embedding space. The outputs of the last cross-modal transformer encoder layer form a $\mathbb{R}^{(HW + L)\times d}$ -sized cross-modal memory $\mathcal{M}_t$ , which captures rich interactions between intra-frame static visual cues and linguistic descriptions. Then in the decoder, a learnable object query vector $O_{t}^{0}\in \mathbb{R}^{d}$ is inputted to repeatedly query object appearance and location information from the memory $\mathcal{M}_t$ via a cross-attention layer. Here, the object query gradually learns to attend to the object that matched the text query. + +# 3.2. Dynamic VL Stream + +The dynamic VL stream performs cross-modal understanding for dynamic contextual information. As illustrated in Figure 2, we first extract clip-level visual features $F_{clip} \in \mathbb{R}^{T \times H \times W \times c}$ from $T$ uniformly sampled video clips by a pre-trained 3D-CNN [4]. Then we employ an FC layer to project the channel dimension from $c$ to $d$ and obtain $F_v^0$ . $F_v^0$ is fed into a Space-Time Cross-Modal Transformer (STCMT) which consists of $N$ layers to model the visual-linguistic dependencies from a dynamic perspective. The detailed architecture of each layer in STCMT is illustrated in Figure 3 (b). In each layer, we first per + +form intra-modality self-attention for the dynamic visual features and linguistic features. For visual features, in order to reduce computation cost, we follow TimesFormer [2] to split the spatio-temporal attention into separate attentions. Denoting the visual features after self-attention as $\mathcal{H}_v\in \mathbb{R}^{T\times H\times W\times d}$ and the linguistic features after self-attention as $\mathcal{H}_l\in \mathbb{R}^{L\times d}$ . We then perform cross-attention between $\mathcal{H}_v$ and $\mathcal{H}_l$ as: + +$$ +\begin{array}{l} Q _ {v} ^ {(h, w)} = W _ {v} ^ {q} \mathcal {H} _ {v} ^ {(h, w)}, K _ {v} = W _ {v} ^ {k} \overline {{\mathcal {H}}} _ {v}, V _ {v} = W _ {v} ^ {v} \overline {{\mathcal {H}}} _ {v}, \\ Q _ {l} = W _ {l} ^ {v} \mathcal {H} _ {l}, K _ {l} = W _ {l} ^ {k} \mathcal {H} _ {l}, V _ {l} = W _ {l} ^ {v} \mathcal {H} _ {l}, \\ \widetilde {\mathcal {H}} _ {v} ^ {(h, w)} = \mathcal {H} _ {v} ^ {(h, w)} + \text {A t t e n t i o n} \left(Q _ {v} ^ {(h, w)}, K _ {l}, V _ {l}\right), \tag {1} \\ \widetilde {\mathcal {H}} _ {l} = \mathcal {H} _ {l} + \operatorname {A t t e n t i o n} \left(Q _ {l}, K _ {v}, V _ {v}\right), \\ \operatorname {A t t e n t i o n} (Q, K, V) = \operatorname {s o f t m a x} \left(\frac {Q K ^ {T}}{\sqrt {d _ {k}}}\right) V, \\ \end{array} +$$ + +where $W_{v}^{q}, W_{v}^{k}, W_{v}^{v}, W_{l}^{q}, W_{l}^{k}, W_{l}^{v}$ are learnable weighted matrices for computing queries, keys and values for the attention mechanism. $\overline{\mathcal{H}}_v \in \mathbb{R}^{T \times d}$ is obtained by conducting mean pooling on $\mathcal{H}_v$ along the spatial dimensions. $d_{k}$ is the dimension of the queries and keys. $\mathcal{H}_{v}^{(h,w)} \in \mathbb{R}^{T \times d}$ indicates the visual feature in $\mathcal{H}_v$ at spatial position $(h,w)$ . $\widetilde{\mathcal{H}}_v, \widetilde{\mathcal{H}}_l$ are the output visual and linguistic features after cross-attention, respectively. The cross-attention operation is designed to explore rich interactions between the visual and linguistic features, which enables the model to learn a powerful cross-modal representation of the depicted dynamic cues. This is essential for grounding temporal moments. After computing the cross-attention between the visual features and linguistic features, we finally employ Feed-Forward Network (FFN) to process both features. In our dynamic VL stream, rich context is learned from the two modalities, and the visual dynamic features and the linguistic features are fused well. + +# 3.3. Cross-Stream Collaboration + +To enable the static and the dynamic VL streams to learn complementary information from each other and to perform + +![](images/fd3fa8a7f1b0b333cb0fc5874d730b2f0015b15612af26863f8a3957bc4e91b1.jpg) + +![](images/6c37b60705fa816cbaa988687d898d56feb542f3dd93943eb7403dcfadbea6d4.jpg) + +![](images/339803897c87cc76b22403108c1a22ed1a942e2496deac32ae83e5af9f680cf5.jpg) +Figure 4. The architectures of our cross-stream collaboration. + +collaborative reasoning, we design a novel cross-stream collaboration block between the two streams. As shown in Figure 4, this block is placed after each decoder layer $\{\mathcal{B}_s^i\}_{i=1}^N$ in static stream and each layer $\{\mathcal{B}_a^i\}_{i=1}^N$ in dynamic stream. It consists of a static-to-dynamic and a dynamic-to-static information transmission block. They are designed asymmetrically to exploit better collaboration between the two streams considering their different abilities. In the following, we will introduce the detailed designs. + +The static-to-dynamic information transmission block is designed to guide the dynamic VL stream to attend to the spatial region that is highly related to the objects depicted in the query text, by utilizing the cross-attention weights learned in the decoder layers of the static VL stream. Concretely, it is formulated as follows: + +$$ +\widetilde {F} _ {v} ^ {i} = \operatorname {L a y e r N o r m} \left(F _ {v} ^ {i} + A ^ {i} \odot \mathrm {F C} \left(F _ {v} ^ {i}\right)\right), \tag {2} +$$ + +where $F_{v}^{i}\in \mathbb{R}^{T\times HW\times d}$ is the output visual feature of the $i$ -th layer in dynamic VL stream and $A^i\in \mathbb{R}^{T\times HW\times d}$ is the cross-attention weights (replicate $d$ times to have $d$ channels at the last dimension) calculated between the object queries $\left\{O_t^i\right\}_{t = 1}^T$ and corresponding encoded memories $\{M_t\}_{t = 1}^T$ in $\mathcal{B}_s^i$ . $\odot$ indicates hadamard product. The intuition behind this design is that the cross-attention weights learned in the static VL stream can attend to regions matching the language description, and this can serve as a strong guidance + +Query: The tall boy comes to the dining table and sits down. + +![](images/178f155f96040b622396c008ac822160ae693dfa61a481bd9af0019d031f549b.jpg) +Figure 5. Visualization of the learned attention map. + +Query: The man in the red stripe goes to the sofa and turns. + +![](images/efdbc4396c7fdd480f50bc14d109099cdfce2d7c5b37360176086b3772243c7c.jpg) + +![](images/61c9a497fe65480b659b6bbb1a7ae2771b667ec8d266b07c9f7bfb3b603d64d3.jpg) + +![](images/939e3cf9fd1c63ff29799aaaa52e4c99e99f4b4df82d61bbd1901cd35d5cd658.jpg) + +![](images/b4abc11a50067098853f1e8d350f971f7749281da6bcbb45ddae82ca51bc1aff.jpg) + +to help the dynamic VL stream focus more on the dynamic variations around the object-related regions. We visualize the attention maps in Figure 5. In the first sample, there are multiple boys and the attention map has a high score at the best-matched person (i.e., the tallest boy in each frame). In the second sample, we observe that when the man in red stripe is absent, the attention weights are relatively smooth. However, the attention weights become sharp once the target man appeared. These samples intuitively show that our static-to-dynamic information transmission block can guide the dynamic stream to focus on motions of the target object. + +The dynamic-to-static information transmission block is designed to transfer the learned motion-text correspondence to the static VL stream. It enhances $O_{t}^{i}$ to $\widetilde{O}_{t}^{i}$ with a cross-attention and a query mixing operation as depicted in Figure 4(c). In each block, the object query $O_{t}^{i}$ in the static stream first queries some dynamic information from $F_{v}^{i}[t]$ (the $t$ -th feature in $F_{v}^{i}$ ) via a cross-attention module, this operation enhances the object query representations with dynamic information learned in the dynamic stream. Then we employ a query mixing operation (implemented as temporal self-attention) to query and mix information from object queries of other frames, which further injects cross-frame information to object queries. With the proposed cross-stream collaboration block, both the static and dynamic VL streams can learn complementary information from each other in an effective way. And the enhanced representations $\widetilde{F}_{v}^{i}, \widetilde{O}_{t}^{i}$ are inputted to the next transformer layer in the static and dynamic VL streams, respectively. + +# 3.4. Prediction Heads + +In this section, we introduce our prediction heads which include a spatial and a temporal prediction head. + +Spatial Prediction Head. The spatial prediction head predicts the location of the target object. For the $t$ -th frame, the input is the object query representation $\widetilde{O}_t^N \in \mathbb{R}^d$ outputted by the last collaboration block. And we implement a 3-layer MLP to predict the bounding box location (represented by center coordinates and size) $\tilde{b}_t \in \mathbb{R}^4$ of the target object. + +Temporal Prediction Head. The temporal prediction head predicts the time span of the target temporal moment. Specifically, we first perform spatial mean pooling on $\widetilde{F}_v^N$ + +Table 1. Comparison results with previous works on HCSTVG-v1 test set and HCSTVG-v2 val set [24]. + +
DatasetMethodm_vIoUvIoU@0.3vIoU@0.5
HCSTVG-v1STGVT [24]18.226.89.5
STVGBert [22]20.429.411.3
TubeDETR [30]32.449.823.5
Ours36.962.234.8
HCSTVG-v2Yu et al. [33]30.0--
MMN [27]30.349.025.6
Aug. 2D-TAN [23]30.450.418.8
TubeDETR [30]36.458.830.6
Ours38.765.533.8
+ +Table 2. Comparison results with state-of-the-art approaches on VidSTG [42] test set. + +
MethodDeclarative SentencesInterrogative Sentences
m_vIoUvIoU@0.3vIoU@0.5m_vIoUvIoU@0.3vIoU@0.5
STGRN [42]19.825.814.618.321.112.8
STGVT [24]21.629.818.9———
OMRN [41]23.132.616.420.628.414.1
STVGBert [22]24.030.918.422.526.016.0
TubeDETR [30]30.442.528.225.735.723.2
Ours33.747.232.828.539.926.2
+ +to obtain a temporal feature $F_{d} \in \mathbb{R}^{T \times d}$ , and we use an FC layer to adjust its dimension to $d_{m}$ . Then we follow 2D-TAN [39] to construct a 2D-proposal map $M \in \mathbb{R}^{T \times T \times d_{m}}$ : + +$$ +M _ {i j} = \left\{ \begin{array}{l l} \mathbf {M P} ([ F _ {d} [ i ], F _ {d} [ i + 1 ], \dots , F _ {d} [ j ]) & i \leq j \\ \mathbf {0} & i > j \end{array} \right. \tag {3} +$$ + +where $\mathrm{MP}(\cdot)$ is mean pooling and $M_{ij}$ indicates the feature of temporal moment proposal $C_{ij}$ with $t_i, t_{j+1}$ as start and end time stamps, respectively. Here $t_i = \frac{i}{T} \cdot T_{video}$ and $T_{video}$ is the duration of the input video. $F_d[i]$ is the $i$ -th feature in $F_d$ . We employ several convolutional layers to transform the 2D map $M$ into a score map $\hat{S} \in \mathbb{R}^{T \times T \times 1}$ , where $\hat{S}_{ij}$ represents the matching score of temporal proposal $C_{ij}$ . For inference, we take the proposal with the highest score as the prediction of the target temporal time span. And we take the predicted bounding boxes from the spatial prediction head for frames within the predicted temporal span to form the final predicted tube. + +# 3.5. Model Training + +We train our model with loss $L = L_{s} + L_{t}$ , where $L_{s}$ and $L_{t}$ are spatial and temporal localization losses, respectively. Specifically, they are defined as follows: + +$$ +L _ {s} = \lambda_ {1} L _ {l 1} (\hat {b}, b) + \lambda_ {2} L _ {G I o U} (\hat {b}, b), \tag {4} +$$ + +$$ +L _ {t} = \lambda_ {3} L _ {t g} (\hat {S}, S) + \lambda_ {4} L _ {t a}, \tag {5} +$$ + +where $L_{l1}$ and $L_{GIoU}$ are L1 loss and GIoU loss [19] on the predicted bounding boxes, respectively. The temporal grounding loss $L_{tg}$ is defined as a binary cross-entropy loss (as done in [23, 39]) between the predicted score map $\hat{S}$ and the ground truth map $S$ , where each element in $S$ represents the IoU between the corresponding proposal and the ground truth temporal moment. $L_{ta}$ is a commonly used temporal attentive loss [34] for accelerating convergence of $L_{tg}$ , and it encourages the model to predict a high matching score for those frames/clips inside the target temporal span. + +# 4. Experiments + +# 4.1. Experimental Settings + +Datasets. We evaluate our method on HCSTVG-v1 dataset, HCSTVG-v2 dataset [24] and VidSTG [42] dataset. HCSTVG datasets are collected from movie scenes and the duration of each video clip is around 20 seconds. This set is quite challenging for spatio-temporal grounding as some video clips contain many persons conducting similar actions. HCSTVG-v1 dataset consists of 4500 and 1160 video-text pairs for training and testing, respectively. HCSTVG-v2 dataset expanded HCSTVG-v1 dataset and improved the annotation quality. It contains 10131, 2000, 4413 samples for training, validation and testing, respectively. Since the annotations of the test set are not publicly available in HCSTVG-v2, we report results on the validation set. VidSTG dataset [42] is constructed based on VidOR [20] dataset according to the object relation annotations. It contains 99943 video-text pairs, including 44808 declarative sentence queries and 55135 interrogative sentence queries. The training, validation and test set consist of 80684, 8956 and 10303 sentences respectively, and 5436, 602 and 732 videos respectively. However, since VidSTG [42] is annotated based on VidOR [20] dataset, the text queries are limited to describe the pre-defined object/relation categories in VidOR [20]. + +Evaluation Metrics. We follow previous works [22, 23, 30, 42] to use mean vIoU as the main evaluation metric. vIoU is defined as $\frac{1}{|T_u|}\sum_{t\in T_i}I_oU(\hat{b}_t,b_t)$ , where $T_{i}$ and $T_{u}$ indicate the intersection and union between the time intervals obtained from ground truth annotation and model prediction, respectively. $\hat{b}_t,b_t$ are the predicted bounding box and ground truth bounding box for the $t$ -th frame, respectively. We average the vIoU score over all the samples to obtain mean vIoU (termed m_vIoU). We also report vIoU@R which indicates the proportion of samples with vIoU higher than $R$ . In ablation studies, we follow TubeDETR [30] to also report the sIoU and tIoU metrics that evaluate the spatial and temporal grounding performances, respectively. + +They are defined as $sIoU = \frac{1}{|T_{gt}|}\sum_{t\in T_i}IoU(\hat{b_t},b_t)$ and $tIoU = \frac{|T_i|}{|T_u|}$ , where $T_{gt}$ indicates the set of frames inside the ground truth time interval. + +Implementation Details. We use ResNet101 [9] as our image encoder and Roberta-base [14] as the text encoder to extract image visual features and language features. And we use the Slowfast [4] model pre-trained on AVA [8] as the video encoder. We initialize part of the weights of our static VL stream using pre-trained MDETR [10] as done in [23, 30]. The number of transformer layers in both streams is set as $N = 6$ . The loss weights are set as $\lambda_1 = 5, \lambda_2 = 2, \lambda_3 = 5, \lambda_4 = 1$ . We train our model on VidSTG [42], HCSTVG-v1, HCSTVG-v2 [24] datasets for 10, 10, 4 epochs, respectively. For most other hyperparameters, we keep them consistent with previous works [10, 30]. Further implementation details can be found in the supplementary material. + +# 4.2. Experimental Results + +In this section, we compare the performance of our approach with previous methods on HCSTVG dataset [24] and VidSTG [42] dataset. + +HCSTVG datasets. We first compare our method with state-of-the-arts on HCSTVG datasets [24] and results are shown in Table 1. On HCSTVG-v1 test set, our approach outperforms previous state-of-the-art method TubeDETR [30], by a considerable margin. We achieve $36.9\%$ m_vIoU, which is $4.5\%$ higher than TubeDETR. On HCSTVG-v2 validation set [24], we compare our method with TubeDETR [30] and some winners from HCVG challenge2021. As shown in Table 1, we also outperform TubeDETR by $2.3\%$ m_vIoU. The above results demonstrate that our proposed model with two collaborative vision-language streams can effectively capture static-dynamic cross-modal dependencies for addressing STVG task. + +VidSTG dataset. We also report results on VidSTG dataset [42]. As listed in Table 2, we also achieve state-of-the-art performance with a m_vIoU of $33.7\%$ on declarative sentences and $28.5\%$ on interrogative sentences, respectively. This further shows the effectiveness of the proposed collaborative two-stream model. + +# 4.3. Ablation Study + +In this section, we conduct ablation experiments on HCSTVG-v2 dataset [24] to verify the effectiveness of some important components in the proposed method. + +Quantitative evaluation on the Cross-Stream Collaboration Block. We first verify the effectiveness of static-to-dynamic and dynamic-to-static information transmission blocks by removing one of them from our developed frame + +Table 3. Evaluation on the cross-stream collaboration mechanism. + +
s-to-dd-to-sm_vIoUvIoU@0.3vIoU@0.5m_tIoUm_sIoU
36.461.229.556.164.9
✓37.563.631.457.464.6
✓37.162.330.256.265.7
✓✓38.765.533.858.165.7
+ +work. We present the evaluation results in Table 3. Compared to the baseline in which the whole cross-stream collaboration block is removed (the first line in Table 3), it is clear that individually adopting static-to-dynamic information transmission (termed "s-to-d") or dynamic-to-static information transmission (termed "d-to-s") can both improve the grounding performance on vIoU metrics. We also observe that "s-to-d" mainly improves the temporal grounding accuracy while "d-to-s" improves the spatial grounding performance. And our full model that jointly employs the whole cross-stream collaboration block achieves the best performance of $38.7\%$ m_vIoU. The above results show that both the static-to-dynamic and the dynamic-to-static information transmission blocks are effective for providing complementary information to the other stream, and thus employing them greatly improves the performance. + +Qualitative evaluation on the Cross-Stream Collaboration Block. To better understand how and why our cross-stream collaboration block significantly improves the grounding accuracy, we further provide 2 qualitative samples in Figure 6. In the first sample, there are three boys coming to the dining table. In this case, our baseline model without cross-stream collaboration failed to predict correct temporal grounding results since it is confused by the similar action performed by the three boys. With our proposed cross-stream collaboration block, the dynamic VL stream can distinguish which boy is the "tall boy" by absorbing knowledge from the static VL stream, then it can correctly predict the target temporal time span. For the second sample, there are two women both dressed in a skirt, thus in our baseline where the cross-stream collaboration is removed, the static stream failed to capture and understand the action "let goes of ... and turns into the room", thus it cannot distinguish which is the target woman. By employing the proposed cross-stream collaboration block, the motion information learned in the dynamic VL stream can be transmitted to the static stream to make it aware of the motion information which can help distinguish the target person. The above two samples clearly show that the proposed cross-stream collaboration block can effectively exchange complementary information between the static and the dynamic VL streams, which can greatly reduce the uncertainty in the hard and ambiguous cases where some objects have a similar appearance or motion so that the model can produce a more accurate prediction. + +Spatial information in the dynamic VL stream. We + +Query: The tall boy comes to the dining table and sits down. + +![](images/174757091560f1f4ae73952479f40a22c5b1c7d72424054e7b0e31655e0c4d6e.jpg) +Figure 6. Visualization of the predicted tube of our approach (yellow), our approach without cross-stream collaboration (red), and ground truth (green). In the first sample, our approach without cross-stream collaboration predicts a wrong temporal time span. In the second sample, ours w/o the collaboration block predicts wrong spatial bounding boxes. (Best viewed zoomed in on screen.) + +Table 4. Evaluation on designs of the dynamic VL stream. + +
Method|Ours|Ours w/o spatial|Ours w/o mean pool
m_tIoU|56.1|54.7|56.2
+ +Table 5. Comparison between different cross-stream collaboration block designs. + +
Methodm_vIoUvIoU@0.3vIoU@0.5m_tIoUm_sIoU
Ours38.765.533.858.165.7
None36.461.229.556.164.9
sum36.662.530.055.664.4
FC + sum36.862.830.557.363.4
concat + FC37.163.230.356.963.9
+ +present ablation studies on different design choices of the dynamic VL stream and results are presented in Table 4. Here, we discard the static stream to eliminate its influence and we report temporal grounding accuracy using m_tIoU metric. By mean pooling at the spatial dimension of the video feature inputted to the dynamic VL stream (termed "Ours w/o spatial"), it results in a drop of more than $1\%$ m_tIoU which shows the importance of learning spatial visual interaction for modeling dynamic cues. And since we develop the dynamic VL stream to mainly learn dynamic cross-modal correspondence, we employ spatial mean pooling on $\mathcal{H}_v$ to obtain $\overline{\mathcal{H}}_v$ so that the text-to-visual cross-attention is performed only on the temporal dimension which can greatly reduce computation cost. By removing this mean pooling operation (termed "Ours w/o mean pool"), the text-to-visual cross-attention is performed among all spatio-temporal tokens, but it has very limited improvement on the performance, which validates the rationality of our design. + +Further study of the collaboration block. To further verify the effectiveness of the proposed collaboration block, we replace it with some symmetric operations, including + +"sum", "FC + sum" and "concat + FC", which are some operations commonly used in the community [4, 21]. As shown in Table 5, "concat + FC" achieves the best performance among these baselines, but it only achieves a limited improvement of $0.7\%$ m_vIoU, which indicates that simply merging the information from the two streams can not fully exploit their complementary abilities. In contrast, "Ours" outperforms "concat + FC" by a considerable margin of more than $3\%$ in terms of vIoU@0.5, demonstrating that our asymmetric cross-stream collaboration block enables the two streams to effectively exchange complementary information and to collaboratively reason the target object. + +# 5. Conclusion + +In this work, we propose a novel framework for Spatio-Temporal Video Grounding task. Our framework mainly consists of a static VL(vision-language) stream which captures static visual cues like object appearance, and a dynamic VL stream which understands dynamic visual cues like action. We elaborately design the framework as a collaborative system in which the two streams collaborative localize the target object via a novel cross-stream collaboration block, which is proven to be effective. We also present visualization results which provide some interesting insights and show how the collaboration of the two streams helps improve the prediction. Our approach significantly outperforms previous methods on VidSTG [42] and HC-STVG dataset [24], which demonstrates its effectiveness. + +Acknowledgements. This work was supported partially by the NSFC (U21A20471, U22A2095, 62076260, 61772570), Guangdong Natural Science Funds Project (2020B1515120085, 2023B1515040025), Guangdong NSF for Distinguished Young Scholar (2022B1515020009, 2018B030306025), and the Key-Area Research and Development Program of Guangzhou (202007030004). We thank Bing Shuai for the helpful discussions. + +# References + +[1] Lisa Anne Hendricks, Oliver Wang, Eli Shechtman, Josef Sivic, Trevor Darrell, and Bryan Russell. Localizing moments in video with natural language. In Proceedings of the IEEE international conference on computer vision, pages 5803-5812, 2017. 2 +[2] Gedas Bertasius, Heng Wang, and Lorenzo Torresani. Is space-time attention all you need for video understanding? In Proceedings of the International Conference on Machine Learning (ICML), July 2021. 4 +[3] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In European conference on computer vision, pages 213-229. Springer, 2020. 4 +[4] Christoph Feichtenhofer, Haoqi Fan, Jitendra Malik, and Kaiming He. Slowfast networks for video recognition. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6202-6211, 2019. 1, 2, 3, 4, 7, 8 +[5] Christoph Feichtenhofer, Axel Pinz, and Richard P Wildes. Spatiotemporal multiplier networks for video action recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4768-4777, 2017. 3 +[6] Jiyang Gao, Chen Sun, Zhenheng Yang, and Ram Nevatia. Tall: Temporal activity localization via language query. In Proceedings of the IEEE international conference on computer vision, pages 5267-5275, 2017. 2 +[7] Yash Goyal, Tejas Khot, Douglas Summers-Stay, Dhruv Bartra, and Devi Parikh. Making the v in vqa matter: Elevating the role of image understanding in visual question answering. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6904-6913, 2017. 1 +[8] Chunhui Gu, Chen Sun, David A Ross, Carl Vondrick, Caroline Pantofaru, Yeqing Li, Sudheendra Vijayanarasimhan, George Toderici, Susanna Ricco, Rahul Sukthankar, et al. Ava: A video dataset of spatio-temporally localized atomic visual actions. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 6047-6056, 2018. 7 +[9] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 4, 7 +[10] Aishwarya Kamath, Mannat Singh, Yann LeCun, Gabriel Synnaeve, Ishan Misra, and Nicolas Carion. Mdetr-modulated detection for end-to-end multi-modal understanding. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1780-1790, 2021. 3, 7 +[11] Kun Li, Dan Guo, and Meng Wang. Proposal-free video grounding with contextual pyramid network. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 35, pages 1902-1910, 2021. 2, 3 +[12] Liunian Harold Li*, Pengchuan Zhang*, Haotian Zhang*, Jianwei Yang, Chunyuan Li, Yiwu Zhong, Lijuan Wang, Lu Yuan, Lei Zhang, Jenq-Neng Hwang, Kai-Wei Chang, and Jianfeng Gao. Grounded language-image pre-training. In CVPR, 2022. 3 + +[13] Yue Liao, Si Liu, Guanbin Li, Fei Wang, Yanjie Chen, Chen Qian, and Bo Li. A real-time cross-modality correlation filtering method for referring expression comprehension. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10880-10889, 2020. 1 +[14] Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. Roberta: A robustly optimized bert pretraining approach. arXiv preprint arXiv:1907.11692, 2019. 4, 7 +[15] Jiasen Lu, Dhruv Batra, Devi Parikh, and Stefan Lee. Vilbert: Pretraining task-agnostic visiolinguistic representations for vision-and-language tasks. Advances in neural information processing systems, 32, 2019. 1 +[16] Jonghwan Mun, Minsu Cho, and Bohyung Han. Local-global video-text interactions for temporal grounding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10810-10819, 2020. 2, 3 +[17] Guoshun Nan, Rui Qiao, Yao Xiao, Jun Liu, Sicong Leng, Hao Zhang, and Wei Lu. Interventional video grounding with dual contrastive learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2765-2775, 2021. 2 +[18] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. Advances in neural information processing systems, 28, 2015. 2 +[19] Hamid Rezatofighi, Nathan Tsoi, JunYoung Gwak, Amir Sadeghian, Ian Reid, and Silvio Savarese. Generalized intersection over union: A metric and a loss for bounding box regression. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 658-666, 2019. 6 +[20] Xindi Shang, Donglin Di, Junbin Xiao, Yu Cao, Xun Yang, and Tat-Seng Chua. Annotating objects and relations in user-generated videos. In Proceedings of the 2019 on International Conference on Multimedia Retrieval, pages 279-287. ACM, 2019. 6 +[21] Karen Simonyan and Andrew Zisserman. Two-stream convolutional networks for action recognition in videos. Advances in neural information processing systems, 27, 2014. 1, 2, 3, 8 +[22] Rui Su, Qian Yu, and Dong Xu. Stvgbert: A visual-linguistic transformer based framework for spatio-temporal video grounding. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 1533–1542, October 2021. 1, 2, 6 +[23] Chaolei Tan, Zihang Lin, Jian-Fang Hu, Xiang Li, and Wei-Shi Zheng. Augmented 2d-tan: A two-stage approach for human-centric spatio-temporal video grounding. arXiv preprint arXiv:2106.10634, 2021. 3, 6, 7 +[24] Zongheng Tang, Yue Liao, Si Liu, Guanbin Li, Xiaojie Jin, Hongxu Jiang, Qian Yu, and Dong Xu. Human-centric spatio-temporal video grounding with visual transformers. IEEE Transactions on Circuits and Systems for Video Technology, pages 1-1, 2021. 1, 2, 6, 7, 8 + +[25] Limin Wang, Yuanjun Xiong, Zhe Wang, and Yu Qiao. Towards good practices for very deep two-stream convnets. arXiv preprint arXiv:1507.02159, 2015. 3 +[26] Xuanhan Wang, Lianli Gao, Peng Wang, Xiaoshuai Sun, and Xianglong Liu. Two-stream 3-d convnet fusion for action recognition in videos with arbitrary size and length. IEEE Transactions on Multimedia, 20(3):634-644, 2017. 3 +[27] Zhenzhi Wang, Limin Wang, Tao Wu, Tianhao Li, and Gangshan Wu. Negative sample matters: A renaissance of metric learning for temporal grounding. CoRR, abs/2109.04872, 2021. 6 +[28] Zuxuan Wu, Yu-Gang Jiang, Xi Wang, Hao Ye, and Xi-angyang Xue. Multi-stream multi-class fusion of deep networks for video classification. In Proceedings of the 24th ACM international conference on Multimedia, pages 791–800, 2016. 3 +[29] Shaoning Xiao, Long Chen, Songyang Zhang, Wei Ji, Jian Shao, Lu Ye, and Jun Xiao. Boundary proposal network for two-stage natural language video localization. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 35, pages 2986-2994, 2021. 2 +[30] Antoine Yang, Antoine Miech, Josef Sivic, Ivan Laptev, and Cordelia Schmid. Tubedetr: Spatio-temporal video grounding with transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 1, 2, 3, 6, 7 +[31] Kexin Yi, Jiajun Wu, Chuang Gan, Antonio Torralba, Pushmeet Kohli, and Josh Tenenbaum. Neural-symbolic vqa: Disentangling reasoning from vision and language understanding. Advances in neural information processing systems, 31, 2018. 1 +[32] Licheng Yu, Zhe Lin, Xiaohui Shen, Jimei Yang, Xin Lu, Mohit Bansal, and Tamara L Berg. Mattnet: Modular attention network for referring expression comprehension. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 1307-1315, 2018. 1, 3 +[33] Yi Yu, Xinying Wang, Wei Hu, Xun Luo, and Cheng Li. 2rd place solutions in the hc-stvg track of person in context challenge 2021. arXiv preprint arXiv:2106.07166, 2021. 6 +[34] Yitian Yuan, Tao Mei, and Wenwu Zhu. To find where you talk: Temporal sentence localization in video with attention based location regression. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 33, pages 9159-9166, 2019. 6 +[35] Runhao Zeng, Haoming Xu, Wenbing Huang, Peihao Chen, Mingkui Tan, and Chuang Gan. Dense regression network for video grounding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10287-10296, 2020. 2, 3 +[36] Da Zhang, Xiyang Dai, Xin Wang, Yuan-Fang Wang, and Larry S Davis. Man: Moment alignment network for natural language moment retrieval via iterative graph adjustment. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1247-1257, 2019. 2 +[37] Hanwang Zhang, Yulei Niu, and Shih-Fu Chang. Grounding referring expressions in images by variational context. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 4158-4166, 2018. 3 + +[38] Hao Zhang, Aixin Sun, Wei Jing, Liangli Zhen, Joey Tianyi Zhou, and Rick Siow Mong Goh. Natural language video localization: A revisit in span-based question answering framework. IEEE transactions on pattern analysis and machine intelligence, 2021. 2, 3 +[39] Songyang Zhang, Houwen Peng, Jianlong Fu, and Jiebo Luo. Learning 2d temporal adjacent networks for moment localization with natural language. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 34, pages 12870-12877, 2020. 2, 3, 6 +[40] Zhiwang Zhang, Dong Xu, Wanli Ouyang, and Chuanqi Tan. Show, tell and summarize: Dense video captioning using visual cue aided sentence summarization. IEEE Transactions on Circuits and Systems for Video Technology, 30(9):3130-3139, 2019. 1 +[41] Zhu Zhang, Zhou Zhao, Zhijie Lin, Baoxing Huai, and Jing Yuan. Object-aware multi-branch relation networks for spatio-temporal video grounding. In Proceedings of the Twenty-Ninth International Joint Conference on Artificial Intelligence, IJCAI'20, 2021. 2, 6 +[42] Zhu Zhang, Zhou Zhao, Yang Zhao, Qi Wang, Huasheng Liu, and Lianli Gao. Where does it exist: Spatio-temporal video grounding for multi-form sentences. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2020. 1, 2, 6, 7, 8 \ No newline at end of file diff --git a/2023/Collaborative Static and Dynamic Vision-Language Streams for Spatio-Temporal Video Grounding/images.zip b/2023/Collaborative Static and Dynamic Vision-Language Streams for Spatio-Temporal Video Grounding/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..ebd6bb21eca25648e9525d6844c56287c6bdac72 --- /dev/null +++ b/2023/Collaborative Static and Dynamic Vision-Language Streams for Spatio-Temporal Video Grounding/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:df2145d16db3add3dd3e8f64aed6a2d4ad46031fa02aa179c6c11533956dfd71 +size 590607 diff --git a/2023/Collaborative Static and Dynamic Vision-Language Streams for Spatio-Temporal Video Grounding/layout.json b/2023/Collaborative Static and Dynamic Vision-Language Streams for Spatio-Temporal Video Grounding/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..a6bf7485c1c9b0cab201268ffb5ef2f45bb7844f --- /dev/null +++ b/2023/Collaborative Static and Dynamic Vision-Language Streams for Spatio-Temporal Video Grounding/layout.json @@ -0,0 +1,7931 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 111, + 103, + 483, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 111, + 103, + 483, + 140 + ], + "spans": [ + { + "bbox": [ + 111, + 103, + 483, + 140 + ], + "type": "text", + "content": "Collaborative Static and Dynamic Vision-Language Streams for Spatio-Temporal Video Grounding" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 87, + 160, + 505, + 188 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 87, + 160, + 505, + 188 + ], + "spans": [ + { + "bbox": [ + 87, + 160, + 505, + 188 + ], + "type": "text", + "content": "Zihang Lin1, Chaolei Tan1, Jian-Fang Hu1,3,4*, Zhi Jin1, Tiancai Ye2, Wei-Shi Zheng1,3,4, Sun Yat-sen University, China" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 57, + 190, + 535, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 190, + 535, + 217 + ], + "spans": [ + { + "bbox": [ + 57, + 190, + 535, + 217 + ], + "type": "text", + "content": "3Guangdong Province Key Laboratory of Information Security Technology, China \n4 Key Laboratory of Machine Intelligence and Advanced Computing, Ministry of Education, China" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 100, + 219, + 492, + 245 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 100, + 219, + 492, + 245 + ], + "spans": [ + { + "bbox": [ + 100, + 219, + 492, + 245 + ], + "type": "text", + "content": "{1inzh59, tanchlei}@mail2.sysu.edu.cn, {hujf5, jinzh26}@mail.sysu.edu.cn, tiancaiye@tencent.com, wszheng@ieee.org" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 143, + 273, + 192, + 285 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 273, + 192, + 285 + ], + "spans": [ + { + "bbox": [ + 143, + 273, + 192, + 285 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 299, + 290, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 299, + 290, + 586 + ], + "spans": [ + { + "bbox": [ + 46, + 299, + 290, + 586 + ], + "type": "text", + "content": "Spatio-Temporal Video Grounding (STVG) aims to localize the target object spatially and temporally according to the given language query. It is a challenging task in which the model should well understand dynamic visual cues (e.g., motions) and static visual cues (e.g., object appearances) in the language description, which requires effective joint modeling of spatio-temporal visual-linguistic dependencies. In this work, we propose a novel framework in which a static vision-language stream and a dynamic vision-language stream are developed to collaboratively reason the target tube. The static stream performs cross-modal understanding in a single frame and learns to attend to the target object spatially according to intra-frame visual cues like object appearances. The dynamic stream models visual-linguistic dependencies across multiple consecutive frames to capture dynamic cues like motions. We further design a novel cross-stream collaborative block between the two streams, which enables the static and dynamic streams to transfer useful and complementary information from each other to achieve collaborative reasoning. Experimental results show the effectiveness of the collaboration of the two streams and our overall framework achieves new state-of-the-art performance on both HCSTVG and VidSTG datasets." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 611, + 128, + 624 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 611, + 128, + 624 + ], + "spans": [ + { + "bbox": [ + 47, + 611, + 128, + 624 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 632, + 287, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 632, + 287, + 693 + ], + "spans": [ + { + "bbox": [ + 46, + 632, + 287, + 693 + ], + "type": "text", + "content": "Vision-language cross-modal understanding is a challenging yet important research problem that bridges the communication between humans and artificial intelligence systems. It has attracted increasing attention and many vision-language tasks were studied in recent years, like vi" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 274, + 547, + 393 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 274, + 547, + 393 + ], + "spans": [ + { + "bbox": [ + 304, + 274, + 547, + 393 + ], + "type": "text", + "content": "sual grounding [13, 32], VQA [7, 31], image/video captioning [15, 40], etc. In this work, we focus on a challenging vision-language task named Spatio-Temporal Video Grounding (STVG) which was recently proposed in [42]. Given a language query indicating an object (as shown in Figure 1), STVG aims to localize the target object spatially and temporally in the video. In this task, the input language query may express different kinds of visual concepts, thus the model requires to well capture and understand these concepts in both vision and language modalities." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 398, + 547, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 398, + 547, + 567 + ], + "spans": [ + { + "bbox": [ + 304, + 398, + 547, + 567 + ], + "type": "text", + "content": "In STVG task, dynamic visual concepts like human motion and static visual concepts like object appearance are both important for distinguishing the target object from other objects that occurred in the same video. For example, in the first sample in Figure 1, the two men are dressed alike (i.e., their static appearance cues are similar), and we can only distinguish them by motion. In the second sample, the two women perform similar actions, they both stand up (i.e., they have similar dynamic motion cues), here, we can only distinguish them by their clothes. The above examples show that static or dynamic visual cues alone cannot solve the STVG task well. And it implies that modeling static and dynamic visual-linguistic dependencies and collaboratively utilizing them are important for addressing STVG task." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 570, + 548, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 548, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 548, + 714 + ], + "type": "text", + "content": "Humans treat static and dynamic cues differently [4, 21], but this was overlooked in previous STVG works [22, 24, 30]. Taking the first query in Figure 1 as an example, a human would randomly or evenly click on some locations on the video's progress bar to find candidate frames containing a man in blue coat. And he will play the video around that frame and attend on the candidate man to check whether he performs the action described in the text (i.e., \"turns around and stops by the stone\"). In the above process, the human understands static and dynamic cues in different ways (i.e., view a single frame and watch the video clip, respectively) and determines the target object by jointly considering the" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 702, + 135, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 702, + 135, + 712 + ], + "spans": [ + { + "bbox": [ + 58, + 702, + 135, + 712 + ], + "type": "text", + "content": "*Corresponding author." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "23100" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 82, + 74, + 361, + 85 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 82, + 74, + 361, + 85 + ], + "spans": [ + { + "bbox": [ + 82, + 74, + 361, + 85 + ], + "type": "text", + "content": "Query: The man in the blue coat turns around and stops by the stone." + } + ] + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 80, + 89, + 511, + 142 + ], + "blocks": [ + { + "bbox": [ + 80, + 89, + 511, + 142 + ], + "lines": [ + { + "bbox": [ + 80, + 89, + 511, + 142 + ], + "spans": [ + { + "bbox": [ + 80, + 89, + 511, + 142 + ], + "type": "image", + "image_path": "03aab8328a8d944b4c530af2fe68465433d7aee9e43f259850bd582dda0c5472.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 82, + 144, + 293, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 82, + 144, + 293, + 156 + ], + "spans": [ + { + "bbox": [ + 82, + 144, + 293, + 156 + ], + "type": "text", + "content": "Query: The woman in light gray stands up and turns." + } + ] + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 80, + 159, + 511, + 213 + ], + "blocks": [ + { + "bbox": [ + 80, + 159, + 511, + 213 + ], + "lines": [ + { + "bbox": [ + 80, + 159, + 511, + 213 + ], + "spans": [ + { + "bbox": [ + 80, + 159, + 511, + 213 + ], + "type": "image", + "image_path": "da337362c1f2cc61ea9ebfa9a52fba8ebcd14cbbf6c017d6e80b21fbf4192c51.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 217, + 547, + 251 + ], + "lines": [ + { + "bbox": [ + 46, + 217, + 547, + 251 + ], + "spans": [ + { + "bbox": [ + 46, + 217, + 547, + 251 + ], + "type": "text", + "content": "Figure 1. Two examples for Spatio-Temporal Video Grounding task. In the first case, the two men are both dressed alike in blue, thus understanding the action described in the query sentence is essential to recognize the person of interest. In the second case, both of the two women stand up, we can only distinguish them by their clothes. (Best viewed zoomed in on screen.)" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 262, + 287, + 370 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 262, + 287, + 370 + ], + "spans": [ + { + "bbox": [ + 46, + 262, + 287, + 370 + ], + "type": "text", + "content": "static and dynamic cues in an \"attend-and-check\" process. Inspired by the above observations, we propose a framework that consists of a static and a dynamic vision-language (VL) streams to model static and dynamic cues, respectively. And we design a novel cross-stream collaboration block between the two streams to simulate the \"attend-and-check\" process, which exchanges useful and complementary information learned in each stream and enables the two streams to collaboratively reason the target object." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 374, + 288, + 638 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 374, + 288, + 638 + ], + "spans": [ + { + "bbox": [ + 46, + 374, + 288, + 638 + ], + "type": "text", + "content": "Specifically, in this work, our static vision-language (VL) stream learns to attend to some candidate regions according to static visual cues like appearance, while the dynamic VL stream learns to understand the dynamic visual cues like action described in the text query. Then, in the collaboration block, we guide the dynamic stream to only focus and attend on the motion of the candidate objects by using the learned attended region in the static stream. And we transfer the text-motion matching information learned in the dynamic stream to the static stream, to help it further check and determine the target object and predict a more consistent tube. With the above cross-stream collaboration blocks, both the static and dynamic vision-language streams can learn reciprocal information from the other stream, which is effective for achieving more accurate spatio-temporal grounding predictions. We conduct experiments on HCSTVG [24] and VidSTG [42] datasets and our approach outperforms previous approaches by a considerable margin. Ablation studies demonstrate the effectiveness of each component in our proposed cross-stream collaboration block and show its superiority over commonly used counterparts in video understanding works [4, 21]." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 642, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 642, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 642, + 288, + 714 + ], + "type": "text", + "content": "In summary, our contributions are: 1), we develop an effective framework that contains two parallel streams to model static-dynamic visual-linguistic dependencies for complete cross-modal understanding; 2), we propose a novel cross-stream collaboration block between the two streams to exchange reciprocal information for each other" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 262, + 545, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 262, + 545, + 300 + ], + "spans": [ + { + "bbox": [ + 305, + 262, + 545, + 300 + ], + "type": "text", + "content": "and enable collaborative reasoning of the target object; 3). Our overall framework achieves state-of-the-art performances on HCSTVG [24] and VidSTG datasets [42]." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 310, + 392, + 322 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 310, + 392, + 322 + ], + "spans": [ + { + "bbox": [ + 306, + 310, + 392, + 322 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 331, + 545, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 331, + 545, + 640 + ], + "spans": [ + { + "bbox": [ + 304, + 331, + 545, + 640 + ], + "type": "text", + "content": "Spatio-Temporal Video Grounding. Spatio-temporal video grounding aims to localize the target object spatially and temporally according to a language query. Early works [24, 41, 42] employ pre-trained detectors like Faster R-CNN [18] to detect objects in each frame, and build their model upon the detection features of each object. STGRN [42] and OMRN [41] learn object relations with spatio-temporal graphs and multi-branch relation networks, respectively. STGVT [24] learns cross-modal representations with visual transformers for video-sentence matching and temporal localization. However, these methods are limited to the pre-detected bounding boxes and cannot localize object categories that are not defined in the pretraining dataset. Recent works [22, 30] design their framework based on strong pre-trained vision-language models and train their models end-to-end. These approaches do not require pre-detected bounding boxes and they perform much better. However, they directly collapse some dimensions (i.e., spatial or temporal) in cross-modal attention calculation to reach an acceptable computation cost, which results in large information loss, and thus they can not well capture the important static-dynamic cross-modal contextual relationship. To address this weakness, we propose to model static and dynamic visual-linguistic correspondence with parallel streams and learn to reason the target object in a collaborative framework." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 642, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 642, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 642, + 547, + 715 + ], + "type": "text", + "content": "Temporal Grounding. Given a language query, temporal grounding aims to predict the time span in a video specified by the language query [1, 6]. Existing works [11, 16, 17, 29, 35, 36, 38, 39] can be divided into proposal-based and proposal-free approaches. Proposal-based approaches generate some proposals (manually designed [29, 39] or" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "23101" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 65, + 68, + 531, + 234 + ], + "blocks": [ + { + "bbox": [ + 65, + 68, + 531, + 234 + ], + "lines": [ + { + "bbox": [ + 65, + 68, + 531, + 234 + ], + "spans": [ + { + "bbox": [ + 65, + 68, + 531, + 234 + ], + "type": "image", + "image_path": "dbdfe90608f64b0d4cad33eee37bb20dccab73e0ad187dbee22bc5ab28785bda.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 237, + 547, + 282 + ], + "lines": [ + { + "bbox": [ + 46, + 237, + 547, + 282 + ], + "spans": [ + { + "bbox": [ + 46, + 237, + 547, + 282 + ], + "type": "text", + "content": "Figure 2. An overview of the proposed framework. Our framework mainly consists of a static VL( vision-language) stream and a dynamic VL stream. The static stream learns to attend to the spatial locations of the target object according to static cues like object appearance. The dynamic stream learns motion-text correspondence according to dynamic cues like human action. We further devise a cross-stream collaboration block that enables the two streams to query useful and complementary information from the other stream." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 293, + 287, + 425 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 293, + 287, + 425 + ], + "spans": [ + { + "bbox": [ + 46, + 293, + 287, + 425 + ], + "type": "text", + "content": "predicted by the model [35]) and predict the IoU scores between each proposal and the ground truth. Proposal-free approaches address moment localization by directly regressing the starting/ending points of the target time spans [11, 16] or predicting the probability of being a starting/ending point for each clip and choosing the time span with the highest joint probability as the prediction [38]. In this work, we follow [23, 39] to employ a proposal-based grounding head. And in contrast to most temporal grounding works that collapse the spatial dimension, we also incorporate spatial cues to enhance temporal localization." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 426, + 287, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 426, + 287, + 604 + ], + "spans": [ + { + "bbox": [ + 46, + 426, + 287, + 604 + ], + "type": "text", + "content": "Text-conditioned Object Detection. Given an image, text-conditioned object detection aims to localize the target object indicated by a sentence (also named referring image expression task) or all objects described in a sentence (also named phrase grounding task). Most existing works [32,37] develop their models based on off-the-shelf detection models and match the detected objects with the sentence query. Recently, MDETR [10] and GLIP [12] unified detection and phrase grounding tasks, and they trained their models with large-scale data. These models can achieve excellent performance on text-conditioned object detection and are used to initialize model weights for STVG task [23, 30]. However, these models lack the ability of modeling dynamic cues. In this work, we design a novel collaboration framework to capture both static and dynamic cues for STVG task." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 605, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 287, + 715 + ], + "type": "text", + "content": "Two-stream Models. Considering that static and dynamic visual cues are quite different but complementary in video understanding, a number of two-stream models [4,5,21,25, 26, 28] are proposed to capture these two kinds of visual cues. For example, Two-Stream ConvNet [21] introduces an optical flow stream to capture object motion. SlowFast [4] develops slow and fast pathways with different temporal resolutions to capture static and dynamic visual cues. However, these works typically fuse different streams equally" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 305, + 293, + 545, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 293, + 545, + 354 + ], + "spans": [ + { + "bbox": [ + 305, + 293, + 545, + 354 + ], + "type": "text", + "content": "via some simple symmetric operations, e.g., assembling the output, sum/concat/FC/temporal-convolution on the features. In contrast, we propose a novel asymmetrical cross-stream collaboration block to exploit better collaboration between the two streams considering their different abilities." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 305, + 369, + 537, + 383 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 369, + 537, + 383 + ], + "spans": [ + { + "bbox": [ + 305, + 369, + 537, + 383 + ], + "type": "text", + "content": "3. Collaborative Static-Dynamic VL Streams" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 391, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 391, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 391, + 546, + 713 + ], + "type": "text", + "content": "We illustrate our framework in Figure 2. Our framework mainly consists of a static and a dynamic VL (vision-language) stream to model static and dynamic visual-linguistic dependencies for complete cross-modal understanding. The static stream performs cross-modal understanding for static contextual information, i.e., finding the object that matches the query text in a still frame according to static visual cues, like object appearance which is important for achieving accurate spatial grounding. The dynamic stream performs cross-modal understanding for dynamic contextual information, i.e., finding the temporal moment that best matches the query sentence according to dynamic visual cues, like action which is important for achieving accurate temporal grounding. Both the static and dynamic streams learn cross-modal correspondence with a transformer-like architecture that contains multiple layers. To enable information transmission between the two streams, we further design a novel cross-stream collaboration block which is placed after each transformer layer in the two streams. In this way, both the static and dynamic VL streams can absorb complementary information from each other and achieve collaborative inferring, which can greatly reduce the uncertainty in ambiguous and hard cases where different objects have similar appearances or perform similar actions. Finally, we employ the prediction head to predict a spatio-temporal tube. In the following, we will introduce each component in detail." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "23102" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 77, + 72, + 221, + 194 + ], + "blocks": [ + { + "bbox": [ + 77, + 72, + 221, + 194 + ], + "lines": [ + { + "bbox": [ + 77, + 72, + 221, + 194 + ], + "spans": [ + { + "bbox": [ + 77, + 72, + 221, + 194 + ], + "type": "image", + "image_path": "99dd6c752da3f4540f0da8e13041971ab5b12d44fb8da44ebe6bc5754c2af8ce.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 147, + 196, + 157, + 205 + ], + "lines": [ + { + "bbox": [ + 147, + 196, + 157, + 205 + ], + "spans": [ + { + "bbox": [ + 147, + 196, + 157, + 205 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 236, + 72, + 512, + 196 + ], + "blocks": [ + { + "bbox": [ + 236, + 72, + 512, + 196 + ], + "lines": [ + { + "bbox": [ + 236, + 72, + 512, + 196 + ], + "spans": [ + { + "bbox": [ + 236, + 72, + 512, + 196 + ], + "type": "image", + "image_path": "385c90247b5e7890bb1ce7fb01fa08a070f95f4c94f1a31e4300c9cad16b4863.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 375, + 195, + 384, + 204 + ], + "lines": [ + { + "bbox": [ + 375, + 195, + 384, + 204 + ], + "spans": [ + { + "bbox": [ + 375, + 195, + 384, + 204 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 208, + 547, + 232 + ], + "lines": [ + { + "bbox": [ + 46, + 208, + 547, + 232 + ], + "spans": [ + { + "bbox": [ + 46, + 208, + 547, + 232 + ], + "type": "text", + "content": "Figure 3. The detailed architectures of the proposed framework. (a): The cross-modal transformer in the static VL stream. (b): The architecture of each space-time cross-modal transformer layer in the dynamic VL stream." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 243, + 151, + 255 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 243, + 151, + 255 + ], + "spans": [ + { + "bbox": [ + 47, + 243, + 151, + 255 + ], + "type": "text", + "content": "3.1. Static VL Stream" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 262, + 290, + 551 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 262, + 290, + 551 + ], + "spans": [ + { + "bbox": [ + 46, + 262, + 290, + 551 + ], + "type": "text", + "content": "The static VL stream is employed to perform cross-modal static context understanding and it processes each frame independently. Inspired by DETR [3], we design our static VL stream as a stack of " + }, + { + "bbox": [ + 46, + 262, + 290, + 551 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 262, + 290, + 551 + ], + "type": "text", + "content": " cross-modal transformer encoder layers and " + }, + { + "bbox": [ + 46, + 262, + 290, + 551 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 262, + 290, + 551 + ], + "type": "text", + "content": " cross-modal transformer decoder layers, as presented in Figure 2 and 3 (a). For the " + }, + { + "bbox": [ + 46, + 262, + 290, + 551 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 262, + 290, + 551 + ], + "type": "text", + "content": "-th frame, the inputs of our cross-modal transformer encoder are the concatenation of " + }, + { + "bbox": [ + 46, + 262, + 290, + 551 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^{HW\\times d}" + }, + { + "bbox": [ + 46, + 262, + 290, + 551 + ], + "type": "text", + "content": "-sized visual features and " + }, + { + "bbox": [ + 46, + 262, + 290, + 551 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^{L\\times d}" + }, + { + "bbox": [ + 46, + 262, + 290, + 551 + ], + "type": "text", + "content": "-sized language features, where " + }, + { + "bbox": [ + 46, + 262, + 290, + 551 + ], + "type": "inline_equation", + "content": "H,W" + }, + { + "bbox": [ + 46, + 262, + 290, + 551 + ], + "type": "text", + "content": " are the resolution of the visual features extracted from the static image frame, and " + }, + { + "bbox": [ + 46, + 262, + 290, + 551 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 46, + 262, + 290, + 551 + ], + "type": "text", + "content": " is the number of text tokens in the input text query. These visual and language features are obtained by a pretrained image encoder [9] and language encoder [14], respectively. We employ an FC layer behind each feature extractor to project the visual and language features to a shared embedding space. The outputs of the last cross-modal transformer encoder layer form a " + }, + { + "bbox": [ + 46, + 262, + 290, + 551 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^{(HW + L)\\times d}" + }, + { + "bbox": [ + 46, + 262, + 290, + 551 + ], + "type": "text", + "content": "-sized cross-modal memory " + }, + { + "bbox": [ + 46, + 262, + 290, + 551 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_t" + }, + { + "bbox": [ + 46, + 262, + 290, + 551 + ], + "type": "text", + "content": ", which captures rich interactions between intra-frame static visual cues and linguistic descriptions. Then in the decoder, a learnable object query vector " + }, + { + "bbox": [ + 46, + 262, + 290, + 551 + ], + "type": "inline_equation", + "content": "O_{t}^{0}\\in \\mathbb{R}^{d}" + }, + { + "bbox": [ + 46, + 262, + 290, + 551 + ], + "type": "text", + "content": " is inputted to repeatedly query object appearance and location information from the memory " + }, + { + "bbox": [ + 46, + 262, + 290, + 551 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_t" + }, + { + "bbox": [ + 46, + 262, + 290, + 551 + ], + "type": "text", + "content": " via a cross-attention layer. Here, the object query gradually learns to attend to the object that matched the text query." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 562, + 167, + 574 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 562, + 167, + 574 + ], + "spans": [ + { + "bbox": [ + 47, + 562, + 167, + 574 + ], + "type": "text", + "content": "3.2. Dynamic VL Stream" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 582, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 582, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 582, + 288, + 714 + ], + "type": "text", + "content": "The dynamic VL stream performs cross-modal understanding for dynamic contextual information. As illustrated in Figure 2, we first extract clip-level visual features " + }, + { + "bbox": [ + 46, + 582, + 288, + 714 + ], + "type": "inline_equation", + "content": "F_{clip} \\in \\mathbb{R}^{T \\times H \\times W \\times c}" + }, + { + "bbox": [ + 46, + 582, + 288, + 714 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 46, + 582, + 288, + 714 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 46, + 582, + 288, + 714 + ], + "type": "text", + "content": " uniformly sampled video clips by a pre-trained 3D-CNN [4]. Then we employ an FC layer to project the channel dimension from " + }, + { + "bbox": [ + 46, + 582, + 288, + 714 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 46, + 582, + 288, + 714 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 46, + 582, + 288, + 714 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 46, + 582, + 288, + 714 + ], + "type": "text", + "content": " and obtain " + }, + { + "bbox": [ + 46, + 582, + 288, + 714 + ], + "type": "inline_equation", + "content": "F_v^0" + }, + { + "bbox": [ + 46, + 582, + 288, + 714 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 46, + 582, + 288, + 714 + ], + "type": "inline_equation", + "content": "F_v^0" + }, + { + "bbox": [ + 46, + 582, + 288, + 714 + ], + "type": "text", + "content": " is fed into a Space-Time Cross-Modal Transformer (STCMT) which consists of " + }, + { + "bbox": [ + 46, + 582, + 288, + 714 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 582, + 288, + 714 + ], + "type": "text", + "content": " layers to model the visual-linguistic dependencies from a dynamic perspective. The detailed architecture of each layer in STCMT is illustrated in Figure 3 (b). In each layer, we first per" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 243, + 547, + 340 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 243, + 547, + 340 + ], + "spans": [ + { + "bbox": [ + 304, + 243, + 547, + 340 + ], + "type": "text", + "content": "form intra-modality self-attention for the dynamic visual features and linguistic features. For visual features, in order to reduce computation cost, we follow TimesFormer [2] to split the spatio-temporal attention into separate attentions. Denoting the visual features after self-attention as " + }, + { + "bbox": [ + 304, + 243, + 547, + 340 + ], + "type": "inline_equation", + "content": "\\mathcal{H}_v\\in \\mathbb{R}^{T\\times H\\times W\\times d}" + }, + { + "bbox": [ + 304, + 243, + 547, + 340 + ], + "type": "text", + "content": " and the linguistic features after self-attention as " + }, + { + "bbox": [ + 304, + 243, + 547, + 340 + ], + "type": "inline_equation", + "content": "\\mathcal{H}_l\\in \\mathbb{R}^{L\\times d}" + }, + { + "bbox": [ + 304, + 243, + 547, + 340 + ], + "type": "text", + "content": ". We then perform cross-attention between " + }, + { + "bbox": [ + 304, + 243, + 547, + 340 + ], + "type": "inline_equation", + "content": "\\mathcal{H}_v" + }, + { + "bbox": [ + 304, + 243, + 547, + 340 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 243, + 547, + 340 + ], + "type": "inline_equation", + "content": "\\mathcal{H}_l" + }, + { + "bbox": [ + 304, + 243, + 547, + 340 + ], + "type": "text", + "content": " as:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 315, + 344, + 545, + 440 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 344, + 545, + 440 + ], + "spans": [ + { + "bbox": [ + 315, + 344, + 545, + 440 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} Q _ {v} ^ {(h, w)} = W _ {v} ^ {q} \\mathcal {H} _ {v} ^ {(h, w)}, K _ {v} = W _ {v} ^ {k} \\overline {{\\mathcal {H}}} _ {v}, V _ {v} = W _ {v} ^ {v} \\overline {{\\mathcal {H}}} _ {v}, \\\\ Q _ {l} = W _ {l} ^ {v} \\mathcal {H} _ {l}, K _ {l} = W _ {l} ^ {k} \\mathcal {H} _ {l}, V _ {l} = W _ {l} ^ {v} \\mathcal {H} _ {l}, \\\\ \\widetilde {\\mathcal {H}} _ {v} ^ {(h, w)} = \\mathcal {H} _ {v} ^ {(h, w)} + \\text {A t t e n t i o n} \\left(Q _ {v} ^ {(h, w)}, K _ {l}, V _ {l}\\right), \\tag {1} \\\\ \\widetilde {\\mathcal {H}} _ {l} = \\mathcal {H} _ {l} + \\operatorname {A t t e n t i o n} \\left(Q _ {l}, K _ {v}, V _ {v}\\right), \\\\ \\operatorname {A t t e n t i o n} (Q, K, V) = \\operatorname {s o f t m a x} \\left(\\frac {Q K ^ {T}}{\\sqrt {d _ {k}}}\\right) V, \\\\ \\end{array}", + "image_path": "244b55e7213b018e42234fc01f983dacef7ed389493fc284f662546980765b33.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 445, + 546, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 445, + 546, + 665 + ], + "spans": [ + { + "bbox": [ + 304, + 445, + 546, + 665 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 445, + 546, + 665 + ], + "type": "inline_equation", + "content": "W_{v}^{q}, W_{v}^{k}, W_{v}^{v}, W_{l}^{q}, W_{l}^{k}, W_{l}^{v}" + }, + { + "bbox": [ + 304, + 445, + 546, + 665 + ], + "type": "text", + "content": " are learnable weighted matrices for computing queries, keys and values for the attention mechanism. " + }, + { + "bbox": [ + 304, + 445, + 546, + 665 + ], + "type": "inline_equation", + "content": "\\overline{\\mathcal{H}}_v \\in \\mathbb{R}^{T \\times d}" + }, + { + "bbox": [ + 304, + 445, + 546, + 665 + ], + "type": "text", + "content": " is obtained by conducting mean pooling on " + }, + { + "bbox": [ + 304, + 445, + 546, + 665 + ], + "type": "inline_equation", + "content": "\\mathcal{H}_v" + }, + { + "bbox": [ + 304, + 445, + 546, + 665 + ], + "type": "text", + "content": " along the spatial dimensions. " + }, + { + "bbox": [ + 304, + 445, + 546, + 665 + ], + "type": "inline_equation", + "content": "d_{k}" + }, + { + "bbox": [ + 304, + 445, + 546, + 665 + ], + "type": "text", + "content": " is the dimension of the queries and keys. " + }, + { + "bbox": [ + 304, + 445, + 546, + 665 + ], + "type": "inline_equation", + "content": "\\mathcal{H}_{v}^{(h,w)} \\in \\mathbb{R}^{T \\times d}" + }, + { + "bbox": [ + 304, + 445, + 546, + 665 + ], + "type": "text", + "content": " indicates the visual feature in " + }, + { + "bbox": [ + 304, + 445, + 546, + 665 + ], + "type": "inline_equation", + "content": "\\mathcal{H}_v" + }, + { + "bbox": [ + 304, + 445, + 546, + 665 + ], + "type": "text", + "content": " at spatial position " + }, + { + "bbox": [ + 304, + 445, + 546, + 665 + ], + "type": "inline_equation", + "content": "(h,w)" + }, + { + "bbox": [ + 304, + 445, + 546, + 665 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 304, + 445, + 546, + 665 + ], + "type": "inline_equation", + "content": "\\widetilde{\\mathcal{H}}_v, \\widetilde{\\mathcal{H}}_l" + }, + { + "bbox": [ + 304, + 445, + 546, + 665 + ], + "type": "text", + "content": " are the output visual and linguistic features after cross-attention, respectively. The cross-attention operation is designed to explore rich interactions between the visual and linguistic features, which enables the model to learn a powerful cross-modal representation of the depicted dynamic cues. This is essential for grounding temporal moments. After computing the cross-attention between the visual features and linguistic features, we finally employ Feed-Forward Network (FFN) to process both features. In our dynamic VL stream, rich context is learned from the two modalities, and the visual dynamic features and the linguistic features are fused well." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 671, + 462, + 682 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 671, + 462, + 682 + ], + "spans": [ + { + "bbox": [ + 306, + 671, + 462, + 682 + ], + "type": "text", + "content": "3.3. Cross-Stream Collaboration" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 689, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 689, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 305, + 689, + 547, + 714 + ], + "type": "text", + "content": "To enable the static and the dynamic VL streams to learn complementary information from each other and to perform" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "23103" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 55, + 68, + 286, + 169 + ], + "blocks": [ + { + "bbox": [ + 55, + 68, + 286, + 169 + ], + "lines": [ + { + "bbox": [ + 55, + 68, + 286, + 169 + ], + "spans": [ + { + "bbox": [ + 55, + 68, + 286, + 169 + ], + "type": "image", + "image_path": "fd3fa8a7f1b0b333cb0fc5874d730b2f0015b15612af26863f8a3957bc4e91b1.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 55, + 169, + 286, + 255 + ], + "blocks": [ + { + "bbox": [ + 55, + 169, + 286, + 255 + ], + "lines": [ + { + "bbox": [ + 55, + 169, + 286, + 255 + ], + "spans": [ + { + "bbox": [ + 55, + 169, + 286, + 255 + ], + "type": "image", + "image_path": "6c37b60705fa816cbaa988687d898d56feb542f3dd93943eb7403dcfadbea6d4.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 55, + 255, + 285, + 356 + ], + "blocks": [ + { + "bbox": [ + 55, + 255, + 285, + 356 + ], + "lines": [ + { + "bbox": [ + 55, + 255, + 285, + 356 + ], + "spans": [ + { + "bbox": [ + 55, + 255, + 285, + 356 + ], + "type": "image", + "image_path": "339803897c87cc76b22403108c1a22ed1a942e2496deac32ae83e5af9f680cf5.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 365, + 279, + 376 + ], + "lines": [ + { + "bbox": [ + 55, + 365, + 279, + 376 + ], + "spans": [ + { + "bbox": [ + 55, + 365, + 279, + 376 + ], + "type": "text", + "content": "Figure 4. The architectures of our cross-stream collaboration." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 388, + 287, + 501 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 388, + 287, + 501 + ], + "spans": [ + { + "bbox": [ + 46, + 388, + 287, + 501 + ], + "type": "text", + "content": "collaborative reasoning, we design a novel cross-stream collaboration block between the two streams. As shown in Figure 4, this block is placed after each decoder layer " + }, + { + "bbox": [ + 46, + 388, + 287, + 501 + ], + "type": "inline_equation", + "content": "\\{\\mathcal{B}_s^i\\}_{i=1}^N" + }, + { + "bbox": [ + 46, + 388, + 287, + 501 + ], + "type": "text", + "content": " in static stream and each layer " + }, + { + "bbox": [ + 46, + 388, + 287, + 501 + ], + "type": "inline_equation", + "content": "\\{\\mathcal{B}_a^i\\}_{i=1}^N" + }, + { + "bbox": [ + 46, + 388, + 287, + 501 + ], + "type": "text", + "content": " in dynamic stream. It consists of a static-to-dynamic and a dynamic-to-static information transmission block. They are designed asymmetrically to exploit better collaboration between the two streams considering their different abilities. In the following, we will introduce the detailed designs." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 501, + 287, + 573 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 501, + 287, + 573 + ], + "spans": [ + { + "bbox": [ + 46, + 501, + 287, + 573 + ], + "type": "text", + "content": "The static-to-dynamic information transmission block is designed to guide the dynamic VL stream to attend to the spatial region that is highly related to the objects depicted in the query text, by utilizing the cross-attention weights learned in the decoder layers of the static VL stream. Concretely, it is formulated as follows:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 86, + 579, + 287, + 596 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 86, + 579, + 287, + 596 + ], + "spans": [ + { + "bbox": [ + 86, + 579, + 287, + 596 + ], + "type": "interline_equation", + "content": "\\widetilde {F} _ {v} ^ {i} = \\operatorname {L a y e r N o r m} \\left(F _ {v} ^ {i} + A ^ {i} \\odot \\mathrm {F C} \\left(F _ {v} ^ {i}\\right)\\right), \\tag {2}", + "image_path": "eda9923ada076a796bc00fb75649ea35e8b34ef4dab17a21d47d90ac7de49561.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 601, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 601, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 601, + 288, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 601, + 288, + 714 + ], + "type": "inline_equation", + "content": "F_{v}^{i}\\in \\mathbb{R}^{T\\times HW\\times d}" + }, + { + "bbox": [ + 47, + 601, + 288, + 714 + ], + "type": "text", + "content": " is the output visual feature of the " + }, + { + "bbox": [ + 47, + 601, + 288, + 714 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 47, + 601, + 288, + 714 + ], + "type": "text", + "content": "-th layer in dynamic VL stream and " + }, + { + "bbox": [ + 47, + 601, + 288, + 714 + ], + "type": "inline_equation", + "content": "A^i\\in \\mathbb{R}^{T\\times HW\\times d}" + }, + { + "bbox": [ + 47, + 601, + 288, + 714 + ], + "type": "text", + "content": " is the cross-attention weights (replicate " + }, + { + "bbox": [ + 47, + 601, + 288, + 714 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 47, + 601, + 288, + 714 + ], + "type": "text", + "content": " times to have " + }, + { + "bbox": [ + 47, + 601, + 288, + 714 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 47, + 601, + 288, + 714 + ], + "type": "text", + "content": " channels at the last dimension) calculated between the object queries " + }, + { + "bbox": [ + 47, + 601, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\left\\{O_t^i\\right\\}_{t = 1}^T" + }, + { + "bbox": [ + 47, + 601, + 288, + 714 + ], + "type": "text", + "content": " and corresponding encoded memories " + }, + { + "bbox": [ + 47, + 601, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\{M_t\\}_{t = 1}^T" + }, + { + "bbox": [ + 47, + 601, + 288, + 714 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 47, + 601, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{B}_s^i" + }, + { + "bbox": [ + 47, + 601, + 288, + 714 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 47, + 601, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\odot" + }, + { + "bbox": [ + 47, + 601, + 288, + 714 + ], + "type": "text", + "content": " indicates hadamard product. The intuition behind this design is that the cross-attention weights learned in the static VL stream can attend to regions matching the language description, and this can serve as a strong guidance" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 309, + 72, + 413, + 91 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 309, + 72, + 413, + 91 + ], + "spans": [ + { + "bbox": [ + 309, + 72, + 413, + 91 + ], + "type": "text", + "content": "Query: The tall boy comes to the dining table and sits down." + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 309, + 93, + 427, + 171 + ], + "blocks": [ + { + "bbox": [ + 309, + 93, + 427, + 171 + ], + "lines": [ + { + "bbox": [ + 309, + 93, + 427, + 171 + ], + "spans": [ + { + "bbox": [ + 309, + 93, + 427, + 171 + ], + "type": "image", + "image_path": "178f155f96040b622396c008ac822160ae693dfa61a481bd9af0019d031f549b.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 330, + 180, + 520, + 191 + ], + "lines": [ + { + "bbox": [ + 330, + 180, + 520, + 191 + ], + "spans": [ + { + "bbox": [ + 330, + 180, + 520, + 191 + ], + "type": "text", + "content": "Figure 5. Visualization of the learned attention map." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 427, + 72, + 533, + 91 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 427, + 72, + 533, + 91 + ], + "spans": [ + { + "bbox": [ + 427, + 72, + 533, + 91 + ], + "type": "text", + "content": "Query: The man in the red stripe goes to the sofa and turns." + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 428, + 92, + 485, + 131 + ], + "blocks": [ + { + "bbox": [ + 428, + 92, + 485, + 131 + ], + "lines": [ + { + "bbox": [ + 428, + 92, + 485, + 131 + ], + "spans": [ + { + "bbox": [ + 428, + 92, + 485, + 131 + ], + "type": "image", + "image_path": "efdbc4396c7fdd480f50bc14d109099cdfce2d7c5b37360176086b3772243c7c.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 486, + 93, + 543, + 131 + ], + "blocks": [ + { + "bbox": [ + 486, + 93, + 543, + 131 + ], + "lines": [ + { + "bbox": [ + 486, + 93, + 543, + 131 + ], + "spans": [ + { + "bbox": [ + 486, + 93, + 543, + 131 + ], + "type": "image", + "image_path": "61c9a497fe65480b659b6bbb1a7ae2771b667ec8d266b07c9f7bfb3b603d64d3.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 429, + 132, + 485, + 171 + ], + "blocks": [ + { + "bbox": [ + 429, + 132, + 485, + 171 + ], + "lines": [ + { + "bbox": [ + 429, + 132, + 485, + 171 + ], + "spans": [ + { + "bbox": [ + 429, + 132, + 485, + 171 + ], + "type": "image", + "image_path": "939e3cf9fd1c63ff29799aaaa52e4c99e99f4b4df82d61bbd1901cd35d5cd658.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 486, + 132, + 543, + 171 + ], + "blocks": [ + { + "bbox": [ + 486, + 132, + 543, + 171 + ], + "lines": [ + { + "bbox": [ + 486, + 132, + 543, + 171 + ], + "spans": [ + { + "bbox": [ + 486, + 132, + 543, + 171 + ], + "type": "image", + "image_path": "b4abc11a50067098853f1e8d350f971f7749281da6bcbb45ddae82ca51bc1aff.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 205, + 545, + 337 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 205, + 545, + 337 + ], + "spans": [ + { + "bbox": [ + 304, + 205, + 545, + 337 + ], + "type": "text", + "content": "to help the dynamic VL stream focus more on the dynamic variations around the object-related regions. We visualize the attention maps in Figure 5. In the first sample, there are multiple boys and the attention map has a high score at the best-matched person (i.e., the tallest boy in each frame). In the second sample, we observe that when the man in red stripe is absent, the attention weights are relatively smooth. However, the attention weights become sharp once the target man appeared. These samples intuitively show that our static-to-dynamic information transmission block can guide the dynamic stream to focus on motions of the target object." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 338, + 546, + 553 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 338, + 546, + 553 + ], + "spans": [ + { + "bbox": [ + 304, + 338, + 546, + 553 + ], + "type": "text", + "content": "The dynamic-to-static information transmission block is designed to transfer the learned motion-text correspondence to the static VL stream. It enhances " + }, + { + "bbox": [ + 304, + 338, + 546, + 553 + ], + "type": "inline_equation", + "content": "O_{t}^{i}" + }, + { + "bbox": [ + 304, + 338, + 546, + 553 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 304, + 338, + 546, + 553 + ], + "type": "inline_equation", + "content": "\\widetilde{O}_{t}^{i}" + }, + { + "bbox": [ + 304, + 338, + 546, + 553 + ], + "type": "text", + "content": " with a cross-attention and a query mixing operation as depicted in Figure 4(c). In each block, the object query " + }, + { + "bbox": [ + 304, + 338, + 546, + 553 + ], + "type": "inline_equation", + "content": "O_{t}^{i}" + }, + { + "bbox": [ + 304, + 338, + 546, + 553 + ], + "type": "text", + "content": " in the static stream first queries some dynamic information from " + }, + { + "bbox": [ + 304, + 338, + 546, + 553 + ], + "type": "inline_equation", + "content": "F_{v}^{i}[t]" + }, + { + "bbox": [ + 304, + 338, + 546, + 553 + ], + "type": "text", + "content": " (the " + }, + { + "bbox": [ + 304, + 338, + 546, + 553 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 338, + 546, + 553 + ], + "type": "text", + "content": "-th feature in " + }, + { + "bbox": [ + 304, + 338, + 546, + 553 + ], + "type": "inline_equation", + "content": "F_{v}^{i}" + }, + { + "bbox": [ + 304, + 338, + 546, + 553 + ], + "type": "text", + "content": ") via a cross-attention module, this operation enhances the object query representations with dynamic information learned in the dynamic stream. Then we employ a query mixing operation (implemented as temporal self-attention) to query and mix information from object queries of other frames, which further injects cross-frame information to object queries. With the proposed cross-stream collaboration block, both the static and dynamic VL streams can learn complementary information from each other in an effective way. And the enhanced representations " + }, + { + "bbox": [ + 304, + 338, + 546, + 553 + ], + "type": "inline_equation", + "content": "\\widetilde{F}_{v}^{i}, \\widetilde{O}_{t}^{i}" + }, + { + "bbox": [ + 304, + 338, + 546, + 553 + ], + "type": "text", + "content": " are inputted to the next transformer layer in the static and dynamic VL streams, respectively." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 306, + 562, + 409, + 573 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 562, + 409, + 573 + ], + "spans": [ + { + "bbox": [ + 306, + 562, + 409, + 573 + ], + "type": "text", + "content": "3.4. Prediction Heads" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 580, + 545, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 580, + 545, + 604 + ], + "spans": [ + { + "bbox": [ + 304, + 580, + 545, + 604 + ], + "type": "text", + "content": "In this section, we introduce our prediction heads which include a spatial and a temporal prediction head." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 605, + 545, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 545, + 677 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 545, + 677 + ], + "type": "text", + "content": "Spatial Prediction Head. The spatial prediction head predicts the location of the target object. For the " + }, + { + "bbox": [ + 304, + 605, + 545, + 677 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 605, + 545, + 677 + ], + "type": "text", + "content": "-th frame, the input is the object query representation " + }, + { + "bbox": [ + 304, + 605, + 545, + 677 + ], + "type": "inline_equation", + "content": "\\widetilde{O}_t^N \\in \\mathbb{R}^d" + }, + { + "bbox": [ + 304, + 605, + 545, + 677 + ], + "type": "text", + "content": " outputted by the last collaboration block. And we implement a 3-layer MLP to predict the bounding box location (represented by center coordinates and size) " + }, + { + "bbox": [ + 304, + 605, + 545, + 677 + ], + "type": "inline_equation", + "content": "\\tilde{b}_t \\in \\mathbb{R}^4" + }, + { + "bbox": [ + 304, + 605, + 545, + 677 + ], + "type": "text", + "content": " of the target object." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 677, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 545, + 714 + ], + "type": "text", + "content": "Temporal Prediction Head. The temporal prediction head predicts the time span of the target temporal moment. Specifically, we first perform spatial mean pooling on " + }, + { + "bbox": [ + 304, + 677, + 545, + 714 + ], + "type": "inline_equation", + "content": "\\widetilde{F}_v^N" + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "23104" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 102, + 254, + 223 + ], + "blocks": [ + { + "bbox": [ + 54, + 70, + 256, + 93 + ], + "lines": [ + { + "bbox": [ + 54, + 70, + 256, + 93 + ], + "spans": [ + { + "bbox": [ + 54, + 70, + 256, + 93 + ], + "type": "text", + "content": "Table 1. Comparison results with previous works on HCSTVG-v1 test set and HCSTVG-v2 val set [24]." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 102, + 254, + 223 + ], + "lines": [ + { + "bbox": [ + 56, + 102, + 254, + 223 + ], + "spans": [ + { + "bbox": [ + 56, + 102, + 254, + 223 + ], + "type": "table", + "html": "
DatasetMethodm_vIoUvIoU@0.3vIoU@0.5
HCSTVG-v1STGVT [24]18.226.89.5
STVGBert [22]20.429.411.3
TubeDETR [30]32.449.823.5
Ours36.962.234.8
HCSTVG-v2Yu et al. [33]30.0--
MMN [27]30.349.025.6
Aug. 2D-TAN [23]30.450.418.8
TubeDETR [30]36.458.830.6
Ours38.765.533.8
", + "image_path": "8661c7db511363ea09ac4d23e6146d65b818a12e112ae8ba05d65e362d857126.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 273, + 102, + 536, + 222 + ], + "blocks": [ + { + "bbox": [ + 264, + 70, + 545, + 93 + ], + "lines": [ + { + "bbox": [ + 264, + 70, + 545, + 93 + ], + "spans": [ + { + "bbox": [ + 264, + 70, + 545, + 93 + ], + "type": "text", + "content": "Table 2. Comparison results with state-of-the-art approaches on VidSTG [42] test set." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 273, + 102, + 536, + 222 + ], + "lines": [ + { + "bbox": [ + 273, + 102, + 536, + 222 + ], + "spans": [ + { + "bbox": [ + 273, + 102, + 536, + 222 + ], + "type": "table", + "html": "
MethodDeclarative SentencesInterrogative Sentences
m_vIoUvIoU@0.3vIoU@0.5m_vIoUvIoU@0.3vIoU@0.5
STGRN [42]19.825.814.618.321.112.8
STGVT [24]21.629.818.9———
OMRN [41]23.132.616.420.628.414.1
STVGBert [22]24.030.918.422.526.016.0
TubeDETR [30]30.442.528.225.735.723.2
Ours33.747.232.828.539.926.2
", + "image_path": "c2574383d49ef2e51014539b9eb34687da3e850f6c4c46906f1cbe71d985a48e.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 241, + 287, + 281 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 241, + 287, + 281 + ], + "spans": [ + { + "bbox": [ + 46, + 241, + 287, + 281 + ], + "type": "text", + "content": "to obtain a temporal feature " + }, + { + "bbox": [ + 46, + 241, + 287, + 281 + ], + "type": "inline_equation", + "content": "F_{d} \\in \\mathbb{R}^{T \\times d}" + }, + { + "bbox": [ + 46, + 241, + 287, + 281 + ], + "type": "text", + "content": ", and we use an FC layer to adjust its dimension to " + }, + { + "bbox": [ + 46, + 241, + 287, + 281 + ], + "type": "inline_equation", + "content": "d_{m}" + }, + { + "bbox": [ + 46, + 241, + 287, + 281 + ], + "type": "text", + "content": ". Then we follow 2D-TAN [39] to construct a 2D-proposal map " + }, + { + "bbox": [ + 46, + 241, + 287, + 281 + ], + "type": "inline_equation", + "content": "M \\in \\mathbb{R}^{T \\times T \\times d_{m}}" + }, + { + "bbox": [ + 46, + 241, + 287, + 281 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 59, + 288, + 287, + 321 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 288, + 287, + 321 + ], + "spans": [ + { + "bbox": [ + 59, + 288, + 287, + 321 + ], + "type": "interline_equation", + "content": "M _ {i j} = \\left\\{ \\begin{array}{l l} \\mathbf {M P} ([ F _ {d} [ i ], F _ {d} [ i + 1 ], \\dots , F _ {d} [ j ]) & i \\leq j \\\\ \\mathbf {0} & i > j \\end{array} \\right. \\tag {3}", + "image_path": "df8f0dd9947be701fde0c68d108539411ee5227d06b7343d77d7711a8e195db4.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 331, + 289, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 331, + 289, + 475 + ], + "spans": [ + { + "bbox": [ + 46, + 331, + 289, + 475 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 331, + 289, + 475 + ], + "type": "inline_equation", + "content": "\\mathrm{MP}(\\cdot)" + }, + { + "bbox": [ + 46, + 331, + 289, + 475 + ], + "type": "text", + "content": " is mean pooling and " + }, + { + "bbox": [ + 46, + 331, + 289, + 475 + ], + "type": "inline_equation", + "content": "M_{ij}" + }, + { + "bbox": [ + 46, + 331, + 289, + 475 + ], + "type": "text", + "content": " indicates the feature of temporal moment proposal " + }, + { + "bbox": [ + 46, + 331, + 289, + 475 + ], + "type": "inline_equation", + "content": "C_{ij}" + }, + { + "bbox": [ + 46, + 331, + 289, + 475 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 46, + 331, + 289, + 475 + ], + "type": "inline_equation", + "content": "t_i, t_{j+1}" + }, + { + "bbox": [ + 46, + 331, + 289, + 475 + ], + "type": "text", + "content": " as start and end time stamps, respectively. Here " + }, + { + "bbox": [ + 46, + 331, + 289, + 475 + ], + "type": "inline_equation", + "content": "t_i = \\frac{i}{T} \\cdot T_{video}" + }, + { + "bbox": [ + 46, + 331, + 289, + 475 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 331, + 289, + 475 + ], + "type": "inline_equation", + "content": "T_{video}" + }, + { + "bbox": [ + 46, + 331, + 289, + 475 + ], + "type": "text", + "content": " is the duration of the input video. " + }, + { + "bbox": [ + 46, + 331, + 289, + 475 + ], + "type": "inline_equation", + "content": "F_d[i]" + }, + { + "bbox": [ + 46, + 331, + 289, + 475 + ], + "type": "text", + "content": " is the " + }, + { + "bbox": [ + 46, + 331, + 289, + 475 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 331, + 289, + 475 + ], + "type": "text", + "content": "-th feature in " + }, + { + "bbox": [ + 46, + 331, + 289, + 475 + ], + "type": "inline_equation", + "content": "F_d" + }, + { + "bbox": [ + 46, + 331, + 289, + 475 + ], + "type": "text", + "content": ". We employ several convolutional layers to transform the 2D map " + }, + { + "bbox": [ + 46, + 331, + 289, + 475 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 46, + 331, + 289, + 475 + ], + "type": "text", + "content": " into a score map " + }, + { + "bbox": [ + 46, + 331, + 289, + 475 + ], + "type": "inline_equation", + "content": "\\hat{S} \\in \\mathbb{R}^{T \\times T \\times 1}" + }, + { + "bbox": [ + 46, + 331, + 289, + 475 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 331, + 289, + 475 + ], + "type": "inline_equation", + "content": "\\hat{S}_{ij}" + }, + { + "bbox": [ + 46, + 331, + 289, + 475 + ], + "type": "text", + "content": " represents the matching score of temporal proposal " + }, + { + "bbox": [ + 46, + 331, + 289, + 475 + ], + "type": "inline_equation", + "content": "C_{ij}" + }, + { + "bbox": [ + 46, + 331, + 289, + 475 + ], + "type": "text", + "content": ". For inference, we take the proposal with the highest score as the prediction of the target temporal time span. And we take the predicted bounding boxes from the spatial prediction head for frames within the predicted temporal span to form the final predicted tube." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 484, + 144, + 498 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 484, + 144, + 498 + ], + "spans": [ + { + "bbox": [ + 47, + 484, + 144, + 498 + ], + "type": "text", + "content": "3.5. Model Training" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 504, + 287, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 504, + 287, + 540 + ], + "spans": [ + { + "bbox": [ + 47, + 504, + 287, + 540 + ], + "type": "text", + "content": "We train our model with loss " + }, + { + "bbox": [ + 47, + 504, + 287, + 540 + ], + "type": "inline_equation", + "content": "L = L_{s} + L_{t}" + }, + { + "bbox": [ + 47, + 504, + 287, + 540 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 47, + 504, + 287, + 540 + ], + "type": "inline_equation", + "content": "L_{s}" + }, + { + "bbox": [ + 47, + 504, + 287, + 540 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 504, + 287, + 540 + ], + "type": "inline_equation", + "content": "L_{t}" + }, + { + "bbox": [ + 47, + 504, + 287, + 540 + ], + "type": "text", + "content": " are spatial and temporal localization losses, respectively. Specifically, they are defined as follows:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 94, + 550, + 287, + 564 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 94, + 550, + 287, + 564 + ], + "spans": [ + { + "bbox": [ + 94, + 550, + 287, + 564 + ], + "type": "interline_equation", + "content": "L _ {s} = \\lambda_ {1} L _ {l 1} (\\hat {b}, b) + \\lambda_ {2} L _ {G I o U} (\\hat {b}, b), \\tag {4}", + "image_path": "f2c41b7c4daa2b53f4ded3b5d392ed60a1828ebc28119efeee399ac2f9f9100d.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 96, + 567, + 287, + 582 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 96, + 567, + 287, + 582 + ], + "spans": [ + { + "bbox": [ + 96, + 567, + 287, + 582 + ], + "type": "interline_equation", + "content": "L _ {t} = \\lambda_ {3} L _ {t g} (\\hat {S}, S) + \\lambda_ {4} L _ {t a}, \\tag {5}", + "image_path": "295426df65132f65b41c2db2474b259030ac0604f0cee858d019ba582c5db621.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 592, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 592, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 592, + 289, + 715 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 592, + 289, + 715 + ], + "type": "inline_equation", + "content": "L_{l1}" + }, + { + "bbox": [ + 46, + 592, + 289, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 592, + 289, + 715 + ], + "type": "inline_equation", + "content": "L_{GIoU}" + }, + { + "bbox": [ + 46, + 592, + 289, + 715 + ], + "type": "text", + "content": " are L1 loss and GIoU loss [19] on the predicted bounding boxes, respectively. The temporal grounding loss " + }, + { + "bbox": [ + 46, + 592, + 289, + 715 + ], + "type": "inline_equation", + "content": "L_{tg}" + }, + { + "bbox": [ + 46, + 592, + 289, + 715 + ], + "type": "text", + "content": " is defined as a binary cross-entropy loss (as done in [23, 39]) between the predicted score map " + }, + { + "bbox": [ + 46, + 592, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\hat{S}" + }, + { + "bbox": [ + 46, + 592, + 289, + 715 + ], + "type": "text", + "content": " and the ground truth map " + }, + { + "bbox": [ + 46, + 592, + 289, + 715 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 46, + 592, + 289, + 715 + ], + "type": "text", + "content": ", where each element in " + }, + { + "bbox": [ + 46, + 592, + 289, + 715 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 46, + 592, + 289, + 715 + ], + "type": "text", + "content": " represents the IoU between the corresponding proposal and the ground truth temporal moment. " + }, + { + "bbox": [ + 46, + 592, + 289, + 715 + ], + "type": "inline_equation", + "content": "L_{ta}" + }, + { + "bbox": [ + 46, + 592, + 289, + 715 + ], + "type": "text", + "content": " is a commonly used temporal attentive loss [34] for accelerating convergence of " + }, + { + "bbox": [ + 46, + 592, + 289, + 715 + ], + "type": "inline_equation", + "content": "L_{tg}" + }, + { + "bbox": [ + 46, + 592, + 289, + 715 + ], + "type": "text", + "content": ", and it encourages the model to predict a high matching score for those frames/clips inside the target temporal span." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 241, + 388, + 255 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 241, + 388, + 255 + ], + "spans": [ + { + "bbox": [ + 306, + 241, + 388, + 255 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 262, + 432, + 275 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 262, + 432, + 275 + ], + "spans": [ + { + "bbox": [ + 306, + 262, + 432, + 275 + ], + "type": "text", + "content": "4.1. Experimental Settings" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 281, + 545, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 281, + 545, + 555 + ], + "spans": [ + { + "bbox": [ + 304, + 281, + 545, + 555 + ], + "type": "text", + "content": "Datasets. We evaluate our method on HCSTVG-v1 dataset, HCSTVG-v2 dataset [24] and VidSTG [42] dataset. HCSTVG datasets are collected from movie scenes and the duration of each video clip is around 20 seconds. This set is quite challenging for spatio-temporal grounding as some video clips contain many persons conducting similar actions. HCSTVG-v1 dataset consists of 4500 and 1160 video-text pairs for training and testing, respectively. HCSTVG-v2 dataset expanded HCSTVG-v1 dataset and improved the annotation quality. It contains 10131, 2000, 4413 samples for training, validation and testing, respectively. Since the annotations of the test set are not publicly available in HCSTVG-v2, we report results on the validation set. VidSTG dataset [42] is constructed based on VidOR [20] dataset according to the object relation annotations. It contains 99943 video-text pairs, including 44808 declarative sentence queries and 55135 interrogative sentence queries. The training, validation and test set consist of 80684, 8956 and 10303 sentences respectively, and 5436, 602 and 732 videos respectively. However, since VidSTG [42] is annotated based on VidOR [20] dataset, the text queries are limited to describe the pre-defined object/relation categories in VidOR [20]." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 556, + 546, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 556, + 546, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 556, + 546, + 715 + ], + "type": "text", + "content": "Evaluation Metrics. We follow previous works [22, 23, 30, 42] to use mean vIoU as the main evaluation metric. vIoU is defined as " + }, + { + "bbox": [ + 304, + 556, + 546, + 715 + ], + "type": "inline_equation", + "content": "\\frac{1}{|T_u|}\\sum_{t\\in T_i}I_oU(\\hat{b}_t,b_t)" + }, + { + "bbox": [ + 304, + 556, + 546, + 715 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 556, + 546, + 715 + ], + "type": "inline_equation", + "content": "T_{i}" + }, + { + "bbox": [ + 304, + 556, + 546, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 556, + 546, + 715 + ], + "type": "inline_equation", + "content": "T_{u}" + }, + { + "bbox": [ + 304, + 556, + 546, + 715 + ], + "type": "text", + "content": " indicate the intersection and union between the time intervals obtained from ground truth annotation and model prediction, respectively. " + }, + { + "bbox": [ + 304, + 556, + 546, + 715 + ], + "type": "inline_equation", + "content": "\\hat{b}_t,b_t" + }, + { + "bbox": [ + 304, + 556, + 546, + 715 + ], + "type": "text", + "content": " are the predicted bounding box and ground truth bounding box for the " + }, + { + "bbox": [ + 304, + 556, + 546, + 715 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 556, + 546, + 715 + ], + "type": "text", + "content": "-th frame, respectively. We average the vIoU score over all the samples to obtain mean vIoU (termed m_vIoU). We also report vIoU@R which indicates the proportion of samples with vIoU higher than " + }, + { + "bbox": [ + 304, + 556, + 546, + 715 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 304, + 556, + 546, + 715 + ], + "type": "text", + "content": ". In ablation studies, we follow TubeDETR [30] to also report the sIoU and tIoU metrics that evaluate the spatial and temporal grounding performances, respectively." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "type": "text", + "content": "23105" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 71, + 287, + 114 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 71, + 287, + 114 + ], + "spans": [ + { + "bbox": [ + 46, + 71, + 287, + 114 + ], + "type": "text", + "content": "They are defined as " + }, + { + "bbox": [ + 46, + 71, + 287, + 114 + ], + "type": "inline_equation", + "content": "sIoU = \\frac{1}{|T_{gt}|}\\sum_{t\\in T_i}IoU(\\hat{b_t},b_t)" + }, + { + "bbox": [ + 46, + 71, + 287, + 114 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 71, + 287, + 114 + ], + "type": "inline_equation", + "content": "tIoU = \\frac{|T_i|}{|T_u|}" + }, + { + "bbox": [ + 46, + 71, + 287, + 114 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 71, + 287, + 114 + ], + "type": "inline_equation", + "content": "T_{gt}" + }, + { + "bbox": [ + 46, + 71, + 287, + 114 + ], + "type": "text", + "content": " indicates the set of frames inside the ground truth time interval." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 114, + 288, + 281 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 114, + 288, + 281 + ], + "spans": [ + { + "bbox": [ + 46, + 114, + 288, + 281 + ], + "type": "text", + "content": "Implementation Details. We use ResNet101 [9] as our image encoder and Roberta-base [14] as the text encoder to extract image visual features and language features. And we use the Slowfast [4] model pre-trained on AVA [8] as the video encoder. We initialize part of the weights of our static VL stream using pre-trained MDETR [10] as done in [23, 30]. The number of transformer layers in both streams is set as " + }, + { + "bbox": [ + 46, + 114, + 288, + 281 + ], + "type": "inline_equation", + "content": "N = 6" + }, + { + "bbox": [ + 46, + 114, + 288, + 281 + ], + "type": "text", + "content": ". The loss weights are set as " + }, + { + "bbox": [ + 46, + 114, + 288, + 281 + ], + "type": "inline_equation", + "content": "\\lambda_1 = 5, \\lambda_2 = 2, \\lambda_3 = 5, \\lambda_4 = 1" + }, + { + "bbox": [ + 46, + 114, + 288, + 281 + ], + "type": "text", + "content": ". We train our model on VidSTG [42], HCSTVG-v1, HCSTVG-v2 [24] datasets for 10, 10, 4 epochs, respectively. For most other hyperparameters, we keep them consistent with previous works [10, 30]. Further implementation details can be found in the supplementary material." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 290, + 170, + 303 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 290, + 170, + 303 + ], + "spans": [ + { + "bbox": [ + 47, + 290, + 170, + 303 + ], + "type": "text", + "content": "4.2. Experimental Results" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 308, + 287, + 344 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 308, + 287, + 344 + ], + "spans": [ + { + "bbox": [ + 46, + 308, + 287, + 344 + ], + "type": "text", + "content": "In this section, we compare the performance of our approach with previous methods on HCSTVG dataset [24] and VidSTG [42] dataset." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 345, + 288, + 500 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 345, + 288, + 500 + ], + "spans": [ + { + "bbox": [ + 46, + 345, + 288, + 500 + ], + "type": "text", + "content": "HCSTVG datasets. We first compare our method with state-of-the-arts on HCSTVG datasets [24] and results are shown in Table 1. On HCSTVG-v1 test set, our approach outperforms previous state-of-the-art method TubeDETR [30], by a considerable margin. We achieve " + }, + { + "bbox": [ + 46, + 345, + 288, + 500 + ], + "type": "inline_equation", + "content": "36.9\\%" + }, + { + "bbox": [ + 46, + 345, + 288, + 500 + ], + "type": "text", + "content": " m_vIoU, which is " + }, + { + "bbox": [ + 46, + 345, + 288, + 500 + ], + "type": "inline_equation", + "content": "4.5\\%" + }, + { + "bbox": [ + 46, + 345, + 288, + 500 + ], + "type": "text", + "content": " higher than TubeDETR. On HCSTVG-v2 validation set [24], we compare our method with TubeDETR [30] and some winners from HCVG challenge2021. As shown in Table 1, we also outperform TubeDETR by " + }, + { + "bbox": [ + 46, + 345, + 288, + 500 + ], + "type": "inline_equation", + "content": "2.3\\%" + }, + { + "bbox": [ + 46, + 345, + 288, + 500 + ], + "type": "text", + "content": " m_vIoU. The above results demonstrate that our proposed model with two collaborative vision-language streams can effectively capture static-dynamic cross-modal dependencies for addressing STVG task." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 500, + 287, + 572 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 500, + 287, + 572 + ], + "spans": [ + { + "bbox": [ + 46, + 500, + 287, + 572 + ], + "type": "text", + "content": "VidSTG dataset. We also report results on VidSTG dataset [42]. As listed in Table 2, we also achieve state-of-the-art performance with a m_vIoU of " + }, + { + "bbox": [ + 46, + 500, + 287, + 572 + ], + "type": "inline_equation", + "content": "33.7\\%" + }, + { + "bbox": [ + 46, + 500, + 287, + 572 + ], + "type": "text", + "content": " on declarative sentences and " + }, + { + "bbox": [ + 46, + 500, + 287, + 572 + ], + "type": "inline_equation", + "content": "28.5\\%" + }, + { + "bbox": [ + 46, + 500, + 287, + 572 + ], + "type": "text", + "content": " on interrogative sentences, respectively. This further shows the effectiveness of the proposed collaborative two-stream model." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 582, + 141, + 594 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 582, + 141, + 594 + ], + "spans": [ + { + "bbox": [ + 47, + 582, + 141, + 594 + ], + "type": "text", + "content": "4.3. Ablation Study" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 600, + 287, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 600, + 287, + 635 + ], + "spans": [ + { + "bbox": [ + 46, + 600, + 287, + 635 + ], + "type": "text", + "content": "In this section, we conduct ablation experiments on HCSTVG-v2 dataset [24] to verify the effectiveness of some important components in the proposed method." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 636, + 287, + 684 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 636, + 287, + 684 + ], + "spans": [ + { + "bbox": [ + 46, + 636, + 287, + 684 + ], + "type": "text", + "content": "Quantitative evaluation on the Cross-Stream Collaboration Block. We first verify the effectiveness of static-to-dynamic and dynamic-to-static information transmission blocks by removing one of them from our developed frame" + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 322, + 86, + 526, + 146 + ], + "blocks": [ + { + "bbox": [ + 306, + 71, + 544, + 82 + ], + "lines": [ + { + "bbox": [ + 306, + 71, + 544, + 82 + ], + "spans": [ + { + "bbox": [ + 306, + 71, + 544, + 82 + ], + "type": "text", + "content": "Table 3. Evaluation on the cross-stream collaboration mechanism." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 322, + 86, + 526, + 146 + ], + "lines": [ + { + "bbox": [ + 322, + 86, + 526, + 146 + ], + "spans": [ + { + "bbox": [ + 322, + 86, + 526, + 146 + ], + "type": "table", + "html": "
s-to-dd-to-sm_vIoUvIoU@0.3vIoU@0.5m_tIoUm_sIoU
36.461.229.556.164.9
✓37.563.631.457.464.6
✓37.162.330.256.265.7
✓✓38.765.533.858.165.7
", + "image_path": "b8c9463e349954991a3b15ae4c46409258158c2ef420c094e0cdaa685061e757.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 161, + 545, + 353 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 161, + 545, + 353 + ], + "spans": [ + { + "bbox": [ + 304, + 161, + 545, + 353 + ], + "type": "text", + "content": "work. We present the evaluation results in Table 3. Compared to the baseline in which the whole cross-stream collaboration block is removed (the first line in Table 3), it is clear that individually adopting static-to-dynamic information transmission (termed \"s-to-d\") or dynamic-to-static information transmission (termed \"d-to-s\") can both improve the grounding performance on vIoU metrics. We also observe that \"s-to-d\" mainly improves the temporal grounding accuracy while \"d-to-s\" improves the spatial grounding performance. And our full model that jointly employs the whole cross-stream collaboration block achieves the best performance of " + }, + { + "bbox": [ + 304, + 161, + 545, + 353 + ], + "type": "inline_equation", + "content": "38.7\\%" + }, + { + "bbox": [ + 304, + 161, + 545, + 353 + ], + "type": "text", + "content": " m_vIoU. The above results show that both the static-to-dynamic and the dynamic-to-static information transmission blocks are effective for providing complementary information to the other stream, and thus employing them greatly improves the performance." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 354, + 546, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 354, + 546, + 700 + ], + "spans": [ + { + "bbox": [ + 304, + 354, + 546, + 700 + ], + "type": "text", + "content": "Qualitative evaluation on the Cross-Stream Collaboration Block. To better understand how and why our cross-stream collaboration block significantly improves the grounding accuracy, we further provide 2 qualitative samples in Figure 6. In the first sample, there are three boys coming to the dining table. In this case, our baseline model without cross-stream collaboration failed to predict correct temporal grounding results since it is confused by the similar action performed by the three boys. With our proposed cross-stream collaboration block, the dynamic VL stream can distinguish which boy is the \"tall boy\" by absorbing knowledge from the static VL stream, then it can correctly predict the target temporal time span. For the second sample, there are two women both dressed in a skirt, thus in our baseline where the cross-stream collaboration is removed, the static stream failed to capture and understand the action \"let goes of ... and turns into the room\", thus it cannot distinguish which is the target woman. By employing the proposed cross-stream collaboration block, the motion information learned in the dynamic VL stream can be transmitted to the static stream to make it aware of the motion information which can help distinguish the target person. The above two samples clearly show that the proposed cross-stream collaboration block can effectively exchange complementary information between the static and the dynamic VL streams, which can greatly reduce the uncertainty in the hard and ambiguous cases where some objects have a similar appearance or motion so that the model can produce a more accurate prediction." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 701, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 701, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 306, + 701, + 545, + 714 + ], + "type": "text", + "content": "Spatial information in the dynamic VL stream. We" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 46, + 693, + 287, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 693, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 693, + 287, + 713 + ], + "type": "text", + "content": "1In this paper, we report the corrected TubeDETR [30] performance updated by TubeDETR authors at https://github.com/antoyang/TubeDETR." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "23106" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 70, + 301, + 81 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 70, + 301, + 81 + ], + "spans": [ + { + "bbox": [ + 56, + 70, + 301, + 81 + ], + "type": "text", + "content": "Query: The tall boy comes to the dining table and sits down." + } + ] + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 56, + 82, + 541, + 224 + ], + "blocks": [ + { + "bbox": [ + 56, + 82, + 541, + 224 + ], + "lines": [ + { + "bbox": [ + 56, + 82, + 541, + 224 + ], + "spans": [ + { + "bbox": [ + 56, + 82, + 541, + 224 + ], + "type": "image", + "image_path": "174757091560f1f4ae73952479f40a22c5b1c7d72424054e7b0e31655e0c4d6e.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 237, + 546, + 270 + ], + "lines": [ + { + "bbox": [ + 46, + 237, + 546, + 270 + ], + "spans": [ + { + "bbox": [ + 46, + 237, + 546, + 270 + ], + "type": "text", + "content": "Figure 6. Visualization of the predicted tube of our approach (yellow), our approach without cross-stream collaboration (red), and ground truth (green). In the first sample, our approach without cross-stream collaboration predicts a wrong temporal time span. In the second sample, ours w/o the collaboration block predicts wrong spatial bounding boxes. (Best viewed zoomed in on screen.)" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 79, + 299, + 252, + 330 + ], + "blocks": [ + { + "bbox": [ + 61, + 285, + 272, + 296 + ], + "lines": [ + { + "bbox": [ + 61, + 285, + 272, + 296 + ], + "spans": [ + { + "bbox": [ + 61, + 285, + 272, + 296 + ], + "type": "text", + "content": "Table 4. Evaluation on designs of the dynamic VL stream." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 79, + 299, + 252, + 330 + ], + "lines": [ + { + "bbox": [ + 79, + 299, + 252, + 330 + ], + "spans": [ + { + "bbox": [ + 79, + 299, + 252, + 330 + ], + "type": "table", + "html": "
Method|Ours|Ours w/o spatial|Ours w/o mean pool
m_tIoU|56.1|54.7|56.2
", + "image_path": "7b1ffdebe30f26865890cd9dd860507537fe9c08280fad732676b6d32d758b95.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 65, + 362, + 266, + 432 + ], + "blocks": [ + { + "bbox": [ + 47, + 337, + 286, + 359 + ], + "lines": [ + { + "bbox": [ + 47, + 337, + 286, + 359 + ], + "spans": [ + { + "bbox": [ + 47, + 337, + 286, + 359 + ], + "type": "text", + "content": "Table 5. Comparison between different cross-stream collaboration block designs." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 65, + 362, + 266, + 432 + ], + "lines": [ + { + "bbox": [ + 65, + 362, + 266, + 432 + ], + "spans": [ + { + "bbox": [ + 65, + 362, + 266, + 432 + ], + "type": "table", + "html": "
Methodm_vIoUvIoU@0.3vIoU@0.5m_tIoUm_sIoU
Ours38.765.533.858.165.7
None36.461.229.556.164.9
sum36.662.530.055.664.4
FC + sum36.862.830.557.363.4
concat + FC37.163.230.356.963.9
", + "image_path": "e557b6538dfae8b1cc4017d550b05f1455d36cdf315c7b0fe4b222324896fa82.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 449, + 287, + 675 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 449, + 287, + 675 + ], + "spans": [ + { + "bbox": [ + 46, + 449, + 287, + 675 + ], + "type": "text", + "content": "present ablation studies on different design choices of the dynamic VL stream and results are presented in Table 4. Here, we discard the static stream to eliminate its influence and we report temporal grounding accuracy using m_tIoU metric. By mean pooling at the spatial dimension of the video feature inputted to the dynamic VL stream (termed \"Ours w/o spatial\"), it results in a drop of more than " + }, + { + "bbox": [ + 46, + 449, + 287, + 675 + ], + "type": "inline_equation", + "content": "1\\%" + }, + { + "bbox": [ + 46, + 449, + 287, + 675 + ], + "type": "text", + "content": " m_tIoU which shows the importance of learning spatial visual interaction for modeling dynamic cues. And since we develop the dynamic VL stream to mainly learn dynamic cross-modal correspondence, we employ spatial mean pooling on " + }, + { + "bbox": [ + 46, + 449, + 287, + 675 + ], + "type": "inline_equation", + "content": "\\mathcal{H}_v" + }, + { + "bbox": [ + 46, + 449, + 287, + 675 + ], + "type": "text", + "content": " to obtain " + }, + { + "bbox": [ + 46, + 449, + 287, + 675 + ], + "type": "inline_equation", + "content": "\\overline{\\mathcal{H}}_v" + }, + { + "bbox": [ + 46, + 449, + 287, + 675 + ], + "type": "text", + "content": " so that the text-to-visual cross-attention is performed only on the temporal dimension which can greatly reduce computation cost. By removing this mean pooling operation (termed \"Ours w/o mean pool\"), the text-to-visual cross-attention is performed among all spatio-temporal tokens, but it has very limited improvement on the performance, which validates the rationality of our design." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "text", + "content": "Further study of the collaboration block. To further verify the effectiveness of the proposed collaboration block, we replace it with some symmetric operations, including" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 285, + 545, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 285, + 545, + 430 + ], + "spans": [ + { + "bbox": [ + 304, + 285, + 545, + 430 + ], + "type": "text", + "content": "\"sum\", \"FC + sum\" and \"concat + FC\", which are some operations commonly used in the community [4, 21]. As shown in Table 5, \"concat + FC\" achieves the best performance among these baselines, but it only achieves a limited improvement of " + }, + { + "bbox": [ + 304, + 285, + 545, + 430 + ], + "type": "inline_equation", + "content": "0.7\\%" + }, + { + "bbox": [ + 304, + 285, + 545, + 430 + ], + "type": "text", + "content": " m_vIoU, which indicates that simply merging the information from the two streams can not fully exploit their complementary abilities. In contrast, \"Ours\" outperforms \"concat + FC\" by a considerable margin of more than " + }, + { + "bbox": [ + 304, + 285, + 545, + 430 + ], + "type": "inline_equation", + "content": "3\\%" + }, + { + "bbox": [ + 304, + 285, + 545, + 430 + ], + "type": "text", + "content": " in terms of vIoU@0.5, demonstrating that our asymmetric cross-stream collaboration block enables the two streams to effectively exchange complementary information and to collaboratively reason the target object." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 445, + 378, + 457 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 445, + 378, + 457 + ], + "spans": [ + { + "bbox": [ + 306, + 445, + 378, + 457 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 468, + 545, + 634 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 468, + 545, + 634 + ], + "spans": [ + { + "bbox": [ + 304, + 468, + 545, + 634 + ], + "type": "text", + "content": "In this work, we propose a novel framework for Spatio-Temporal Video Grounding task. Our framework mainly consists of a static VL(vision-language) stream which captures static visual cues like object appearance, and a dynamic VL stream which understands dynamic visual cues like action. We elaborately design the framework as a collaborative system in which the two streams collaborative localize the target object via a novel cross-stream collaboration block, which is proven to be effective. We also present visualization results which provide some interesting insights and show how the collaboration of the two streams helps improve the prediction. Our approach significantly outperforms previous methods on VidSTG [42] and HC-STVG dataset [24], which demonstrates its effectiveness." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 635, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 635, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 635, + 545, + 712 + ], + "type": "text", + "content": "Acknowledgements. This work was supported partially by the NSFC (U21A20471, U22A2095, 62076260, 61772570), Guangdong Natural Science Funds Project (2020B1515120085, 2023B1515040025), Guangdong NSF for Distinguished Young Scholar (2022B1515020009, 2018B030306025), and the Key-Area Research and Development Program of Guangzhou (202007030004). We thank Bing Shuai for the helpful discussions." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "23107" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 146 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 146 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 146 + ], + "type": "text", + "content": "[1] Lisa Anne Hendricks, Oliver Wang, Eli Shechtman, Josef Sivic, Trevor Darrell, and Bryan Russell. Localizing moments in video with natural language. In Proceedings of the IEEE international conference on computer vision, pages 5803-5812, 2017. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 148, + 288, + 192 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 148, + 288, + 192 + ], + "spans": [ + { + "bbox": [ + 53, + 148, + 288, + 192 + ], + "type": "text", + "content": "[2] Gedas Bertasius, Heng Wang, and Lorenzo Torresani. Is space-time attention all you need for video understanding? In Proceedings of the International Conference on Machine Learning (ICML), July 2021. 4" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 194, + 288, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 194, + 288, + 239 + ], + "spans": [ + { + "bbox": [ + 53, + 194, + 288, + 239 + ], + "type": "text", + "content": "[3] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In European conference on computer vision, pages 213-229. Springer, 2020. 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 239, + 288, + 283 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 239, + 288, + 283 + ], + "spans": [ + { + "bbox": [ + 53, + 239, + 288, + 283 + ], + "type": "text", + "content": "[4] Christoph Feichtenhofer, Haoqi Fan, Jitendra Malik, and Kaiming He. Slowfast networks for video recognition. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6202-6211, 2019. 1, 2, 3, 4, 7, 8" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 285, + 287, + 329 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 285, + 287, + 329 + ], + "spans": [ + { + "bbox": [ + 53, + 285, + 287, + 329 + ], + "type": "text", + "content": "[5] Christoph Feichtenhofer, Axel Pinz, and Richard P Wildes. Spatiotemporal multiplier networks for video action recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4768-4777, 2017. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 331, + 287, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 331, + 287, + 374 + ], + "spans": [ + { + "bbox": [ + 53, + 331, + 287, + 374 + ], + "type": "text", + "content": "[6] Jiyang Gao, Chen Sun, Zhenheng Yang, and Ram Nevatia. Tall: Temporal activity localization via language query. In Proceedings of the IEEE international conference on computer vision, pages 5267-5275, 2017. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 376, + 287, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 376, + 287, + 431 + ], + "spans": [ + { + "bbox": [ + 53, + 376, + 287, + 431 + ], + "type": "text", + "content": "[7] Yash Goyal, Tejas Khot, Douglas Summers-Stay, Dhruv Bartra, and Devi Parikh. Making the v in vqa matter: Elevating the role of image understanding in visual question answering. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6904-6913, 2017. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 433, + 287, + 509 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 433, + 287, + 509 + ], + "spans": [ + { + "bbox": [ + 53, + 433, + 287, + 509 + ], + "type": "text", + "content": "[8] Chunhui Gu, Chen Sun, David A Ross, Carl Vondrick, Caroline Pantofaru, Yeqing Li, Sudheendra Vijayanarasimhan, George Toderici, Susanna Ricco, Rahul Sukthankar, et al. Ava: A video dataset of spatio-temporally localized atomic visual actions. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 6047-6056, 2018. 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 510, + 287, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 510, + 287, + 555 + ], + "spans": [ + { + "bbox": [ + 53, + 510, + 287, + 555 + ], + "type": "text", + "content": "[9] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 4, 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 556, + 287, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 556, + 287, + 612 + ], + "spans": [ + { + "bbox": [ + 48, + 556, + 287, + 612 + ], + "type": "text", + "content": "[10] Aishwarya Kamath, Mannat Singh, Yann LeCun, Gabriel Synnaeve, Ishan Misra, and Nicolas Carion. Mdetr-modulated detection for end-to-end multi-modal understanding. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1780-1790, 2021. 3, 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 613, + 287, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 613, + 287, + 657 + ], + "spans": [ + { + "bbox": [ + 48, + 613, + 287, + 657 + ], + "type": "text", + "content": "[11] Kun Li, Dan Guo, and Meng Wang. Proposal-free video grounding with contextual pyramid network. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 35, pages 1902-1910, 2021. 2, 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 658, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 287, + 712 + ], + "type": "text", + "content": "[12] Liunian Harold Li*, Pengchuan Zhang*, Haotian Zhang*, Jianwei Yang, Chunyuan Li, Yiwu Zhong, Lijuan Wang, Lu Yuan, Lei Zhang, Jenq-Neng Hwang, Kai-Wei Chang, and Jianfeng Gao. Grounded language-image pre-training. In CVPR, 2022. 3" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 129 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 129 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 129 + ], + "type": "text", + "content": "[13] Yue Liao, Si Liu, Guanbin Li, Fei Wang, Yanjie Chen, Chen Qian, and Bo Li. A real-time cross-modality correlation filtering method for referring expression comprehension. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10880-10889, 2020. 1" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 129, + 545, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 129, + 545, + 184 + ], + "spans": [ + { + "bbox": [ + 307, + 129, + 545, + 184 + ], + "type": "text", + "content": "[14] Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. Roberta: A robustly optimized bert pretraining approach. arXiv preprint arXiv:1907.11692, 2019. 4, 7" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 186, + 545, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 186, + 545, + 229 + ], + "spans": [ + { + "bbox": [ + 308, + 186, + 545, + 229 + ], + "type": "text", + "content": "[15] Jiasen Lu, Dhruv Batra, Devi Parikh, and Stefan Lee. Vilbert: Pretraining task-agnostic visiolinguistic representations for vision-and-language tasks. Advances in neural information processing systems, 32, 2019. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 231, + 545, + 284 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 231, + 545, + 284 + ], + "spans": [ + { + "bbox": [ + 308, + 231, + 545, + 284 + ], + "type": "text", + "content": "[16] Jonghwan Mun, Minsu Cho, and Bohyung Han. Local-global video-text interactions for temporal grounding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10810-10819, 2020. 2, 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 286, + 545, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 286, + 545, + 342 + ], + "spans": [ + { + "bbox": [ + 308, + 286, + 545, + 342 + ], + "type": "text", + "content": "[17] Guoshun Nan, Rui Qiao, Yao Xiao, Jun Liu, Sicong Leng, Hao Zhang, and Wei Lu. Interventional video grounding with dual contrastive learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2765-2775, 2021. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 343, + 545, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 343, + 545, + 387 + ], + "spans": [ + { + "bbox": [ + 308, + 343, + 545, + 387 + ], + "type": "text", + "content": "[18] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. Advances in neural information processing systems, 28, 2015. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 388, + 545, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 388, + 545, + 453 + ], + "spans": [ + { + "bbox": [ + 308, + 388, + 545, + 453 + ], + "type": "text", + "content": "[19] Hamid Rezatofighi, Nathan Tsoi, JunYoung Gwak, Amir Sadeghian, Ian Reid, and Silvio Savarese. Generalized intersection over union: A metric and a loss for bounding box regression. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 658-666, 2019. 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 456, + 545, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 456, + 545, + 510 + ], + "spans": [ + { + "bbox": [ + 308, + 456, + 545, + 510 + ], + "type": "text", + "content": "[20] Xindi Shang, Donglin Di, Junbin Xiao, Yu Cao, Xun Yang, and Tat-Seng Chua. Annotating objects and relations in user-generated videos. In Proceedings of the 2019 on International Conference on Multimedia Retrieval, pages 279-287. ACM, 2019. 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 512, + 545, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 512, + 545, + 555 + ], + "spans": [ + { + "bbox": [ + 308, + 512, + 545, + 555 + ], + "type": "text", + "content": "[21] Karen Simonyan and Andrew Zisserman. Two-stream convolutional networks for action recognition in videos. Advances in neural information processing systems, 27, 2014. 1, 2, 3, 8" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 557, + 545, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 557, + 545, + 611 + ], + "spans": [ + { + "bbox": [ + 308, + 557, + 545, + 611 + ], + "type": "text", + "content": "[22] Rui Su, Qian Yu, and Dong Xu. Stvgbert: A visual-linguistic transformer based framework for spatio-temporal video grounding. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 1533–1542, October 2021. 1, 2, 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 613, + 545, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 613, + 545, + 657 + ], + "spans": [ + { + "bbox": [ + 308, + 613, + 545, + 657 + ], + "type": "text", + "content": "[23] Chaolei Tan, Zihang Lin, Jian-Fang Hu, Xiang Li, and Wei-Shi Zheng. Augmented 2d-tan: A two-stage approach for human-centric spatio-temporal video grounding. arXiv preprint arXiv:2106.10634, 2021. 3, 6, 7" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 658, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 658, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 308, + 658, + 545, + 712 + ], + "type": "text", + "content": "[24] Zongheng Tang, Yue Liao, Si Liu, Guanbin Li, Xiaojie Jin, Hongxu Jiang, Qian Yu, and Dong Xu. Human-centric spatio-temporal video grounding with visual transformers. IEEE Transactions on Circuits and Systems for Video Technology, pages 1-1, 2021. 1, 2, 6, 7, 8" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "23108" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 286, + 712 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 48, + 72, + 286, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 286, + 106 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 286, + 106 + ], + "type": "text", + "content": "[25] Limin Wang, Yuanjun Xiong, Zhe Wang, and Yu Qiao. Towards good practices for very deep two-stream convnets. arXiv preprint arXiv:1507.02159, 2015. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 106, + 286, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 106, + 286, + 149 + ], + "spans": [ + { + "bbox": [ + 48, + 106, + 286, + 149 + ], + "type": "text", + "content": "[26] Xuanhan Wang, Lianli Gao, Peng Wang, Xiaoshuai Sun, and Xianglong Liu. Two-stream 3-d convnet fusion for action recognition in videos with arbitrary size and length. IEEE Transactions on Multimedia, 20(3):634-644, 2017. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 150, + 286, + 193 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 150, + 286, + 193 + ], + "spans": [ + { + "bbox": [ + 48, + 150, + 286, + 193 + ], + "type": "text", + "content": "[27] Zhenzhi Wang, Limin Wang, Tao Wu, Tianhao Li, and Gangshan Wu. Negative sample matters: A renaissance of metric learning for temporal grounding. CoRR, abs/2109.04872, 2021. 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 194, + 286, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 194, + 286, + 248 + ], + "spans": [ + { + "bbox": [ + 48, + 194, + 286, + 248 + ], + "type": "text", + "content": "[28] Zuxuan Wu, Yu-Gang Jiang, Xi Wang, Hao Ye, and Xi-angyang Xue. Multi-stream multi-class fusion of deep networks for video classification. In Proceedings of the 24th ACM international conference on Multimedia, pages 791–800, 2016. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 250, + 286, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 250, + 286, + 304 + ], + "spans": [ + { + "bbox": [ + 48, + 250, + 286, + 304 + ], + "type": "text", + "content": "[29] Shaoning Xiao, Long Chen, Songyang Zhang, Wei Ji, Jian Shao, Lu Ye, and Jun Xiao. Boundary proposal network for two-stage natural language video localization. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 35, pages 2986-2994, 2021. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 304, + 286, + 359 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 304, + 286, + 359 + ], + "spans": [ + { + "bbox": [ + 48, + 304, + 286, + 359 + ], + "type": "text", + "content": "[30] Antoine Yang, Antoine Miech, Josef Sivic, Ivan Laptev, and Cordelia Schmid. Tubedetr: Spatio-temporal video grounding with transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 1, 2, 3, 6, 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 360, + 286, + 414 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 360, + 286, + 414 + ], + "spans": [ + { + "bbox": [ + 48, + 360, + 286, + 414 + ], + "type": "text", + "content": "[31] Kexin Yi, Jiajun Wu, Chuang Gan, Antonio Torralba, Pushmeet Kohli, and Josh Tenenbaum. Neural-symbolic vqa: Disentangling reasoning from vision and language understanding. Advances in neural information processing systems, 31, 2018. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 415, + 286, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 415, + 286, + 470 + ], + "spans": [ + { + "bbox": [ + 48, + 415, + 286, + 470 + ], + "type": "text", + "content": "[32] Licheng Yu, Zhe Lin, Xiaohui Shen, Jimei Yang, Xin Lu, Mohit Bansal, and Tamara L Berg. Mattnet: Modular attention network for referring expression comprehension. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 1307-1315, 2018. 1, 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 471, + 286, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 471, + 286, + 503 + ], + "spans": [ + { + "bbox": [ + 48, + 471, + 286, + 503 + ], + "type": "text", + "content": "[33] Yi Yu, Xinying Wang, Wei Hu, Xun Luo, and Cheng Li. 2rd place solutions in the hc-stvg track of person in context challenge 2021. arXiv preprint arXiv:2106.07166, 2021. 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 504, + 286, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 504, + 286, + 557 + ], + "spans": [ + { + "bbox": [ + 48, + 504, + 286, + 557 + ], + "type": "text", + "content": "[34] Yitian Yuan, Tao Mei, and Wenwu Zhu. To find where you talk: Temporal sentence localization in video with attention based location regression. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 33, pages 9159-9166, 2019. 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 559, + 286, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 559, + 286, + 612 + ], + "spans": [ + { + "bbox": [ + 48, + 559, + 286, + 612 + ], + "type": "text", + "content": "[35] Runhao Zeng, Haoming Xu, Wenbing Huang, Peihao Chen, Mingkui Tan, and Chuang Gan. Dense regression network for video grounding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10287-10296, 2020. 2, 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 614, + 286, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 614, + 286, + 669 + ], + "spans": [ + { + "bbox": [ + 48, + 614, + 286, + 669 + ], + "type": "text", + "content": "[36] Da Zhang, Xiyang Dai, Xin Wang, Yuan-Fang Wang, and Larry S Davis. Man: Moment alignment network for natural language moment retrieval via iterative graph adjustment. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1247-1257, 2019. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 670, + 286, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 670, + 286, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 670, + 286, + 712 + ], + "type": "text", + "content": "[37] Hanwang Zhang, Yulei Niu, and Shih-Fu Chang. Grounding referring expressions in images by variational context. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 4158-4166, 2018. 3" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 72, + 545, + 350 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 307, + 72, + 545, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 72, + 545, + 127 + ], + "spans": [ + { + "bbox": [ + 307, + 72, + 545, + 127 + ], + "type": "text", + "content": "[38] Hao Zhang, Aixin Sun, Wei Jing, Liangli Zhen, Joey Tianyi Zhou, and Rick Siow Mong Goh. Natural language video localization: A revisit in span-based question answering framework. IEEE transactions on pattern analysis and machine intelligence, 2021. 2, 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 129, + 545, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 129, + 545, + 182 + ], + "spans": [ + { + "bbox": [ + 307, + 129, + 545, + 182 + ], + "type": "text", + "content": "[39] Songyang Zhang, Houwen Peng, Jianlong Fu, and Jiebo Luo. Learning 2d temporal adjacent networks for moment localization with natural language. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 34, pages 12870-12877, 2020. 2, 3, 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 184, + 545, + 238 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 184, + 545, + 238 + ], + "spans": [ + { + "bbox": [ + 307, + 184, + 545, + 238 + ], + "type": "text", + "content": "[40] Zhiwang Zhang, Dong Xu, Wanli Ouyang, and Chuanqi Tan. Show, tell and summarize: Dense video captioning using visual cue aided sentence summarization. IEEE Transactions on Circuits and Systems for Video Technology, 30(9):3130-3139, 2019. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 240, + 545, + 295 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 240, + 545, + 295 + ], + "spans": [ + { + "bbox": [ + 307, + 240, + 545, + 295 + ], + "type": "text", + "content": "[41] Zhu Zhang, Zhou Zhao, Zhijie Lin, Baoxing Huai, and Jing Yuan. Object-aware multi-branch relation networks for spatio-temporal video grounding. In Proceedings of the Twenty-Ninth International Joint Conference on Artificial Intelligence, IJCAI'20, 2021. 2, 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 296, + 545, + 350 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 296, + 545, + 350 + ], + "spans": [ + { + "bbox": [ + 307, + 296, + 545, + 350 + ], + "type": "text", + "content": "[42] Zhu Zhang, Zhou Zhao, Yang Zhao, Qi Wang, Huasheng Liu, and Lianli Gao. Where does it exist: Spatio-temporal video grounding for multi-form sentences. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2020. 1, 2, 6, 7, 8" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "23109" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file